diff --git a/.gitattributes b/.gitattributes index c5c4465c05b2ae53431cc503a448961b7fa62b53..9d4e12e5f25e7d50960636215d37d5a5cbca555a 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1127,3 +1127,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/Higher-Order[[:space:]]Ratio[[:space:]]Cycles[[:space:]]for[[:space:]]Fast[[:space:]]and[[:space:]]Globally[[:space:]]Optimal[[:space:]]Shape[[:space:]]Matching/7b6743f1-f248-4de2-8fe9-b4d5688fc8ca_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/HistoFS_[[:space:]]Non-IID[[:space:]]Histopathologic[[:space:]]Whole[[:space:]]Slide[[:space:]]Image[[:space:]]Classification[[:space:]]via[[:space:]]Federated[[:space:]]Style[[:space:]]Transfer[[:space:]]with[[:space:]]RoI-Preserving/752a608c-dc4e-4d64-86f3-2a1f5341e163_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/HoGS_[[:space:]]Unified[[:space:]]Near[[:space:]]and[[:space:]]Far[[:space:]]Object[[:space:]]Reconstruction[[:space:]]via[[:space:]]Homogeneous[[:space:]]Gaussian[[:space:]]Splatting/be0a5430-87f7-4bd3-9631-e2ed5908577c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Image[[:space:]]Quality[[:space:]]Assessment_[[:space:]]From[[:space:]]Human[[:space:]]to[[:space:]]Machine[[:space:]]Preference/9c43c7f6-7667-4c47-8440-32e15bc27376_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Image[[:space:]]Quality[[:space:]]Assessment_[[:space:]]Investigating[[:space:]]Causal[[:space:]]Perceptual[[:space:]]Effects[[:space:]]with[[:space:]]Abductive[[:space:]]Counterfactual[[:space:]]Inference/5e0e415c-9773-49a3-97ea-205f1760b866_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Image[[:space:]]Reconstruction[[:space:]]from[[:space:]]Readout-Multiplexed[[:space:]]Single-Photon[[:space:]]Detector[[:space:]]Arrays/960040c5-ed98-43eb-9b66-fa83e44db35d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Image[[:space:]]Referenced[[:space:]]Sketch[[:space:]]Colorization[[:space:]]Based[[:space:]]on[[:space:]]Animation[[:space:]]Creation[[:space:]]Workflow/9a0c23ba-5904-4499-9efc-651298f2730b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Image[[:space:]]is[[:space:]]All[[:space:]]You[[:space:]]Need[[:space:]]to[[:space:]]Empower[[:space:]]Large-scale[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]In-Domain[[:space:]]Generation/e387f9f2-242c-46fd-9d3c-463d01e38cba_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Imagine[[:space:]]and[[:space:]]Seek_[[:space:]]Improving[[:space:]]Composed[[:space:]]Image[[:space:]]Retrieval[[:space:]]with[[:space:]]an[[:space:]]Imagined[[:space:]]Proxy/f227006c-210b-4aac-822f-93d2a98acc36_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ImagineFSL_[[:space:]]Self-Supervised[[:space:]]Pretraining[[:space:]]Matters[[:space:]]on[[:space:]]Imagined[[:space:]]Base[[:space:]]Set[[:space:]]for[[:space:]]VLM-based[[:space:]]Few-shot[[:space:]]Learning/b527a5f4-addd-424b-8f6f-3155526e27dd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Img-Diff_[[:space:]]Contrastive[[:space:]]Data[[:space:]]Synthesis[[:space:]]for[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models/e6b1bb30-e9d5-4708-960e-31ff7e5e9a5b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Immune_[[:space:]]Improving[[:space:]]Safety[[:space:]]Against[[:space:]]Jailbreaks[[:space:]]in[[:space:]]Multi-modal[[:space:]]LLMs[[:space:]]via[[:space:]]Inference-Time[[:space:]]Alignment/66fe35a0-e089-4c9c-afba-719727b0b2df_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Implicit[[:space:]]Bias[[:space:]]Injection[[:space:]]Attacks[[:space:]]against[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models/5a713a88-9e78-41c3-b9a4-4778994a6f99_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Implicit[[:space:]]Correspondence[[:space:]]Learning[[:space:]]for[[:space:]]Image-to-Point[[:space:]]Cloud[[:space:]]Registration/3b6935ad-40bc-4903-80f6-f62148a4f3d3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improve[[:space:]]Representation[[:space:]]for[[:space:]]Imbalanced[[:space:]]Regression[[:space:]]through[[:space:]]Geometric[[:space:]]Constraints/e4bdbcad-b68b-4255-a63d-f2d51b9212bd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improved[[:space:]]Monocular[[:space:]]Depth[[:space:]]Prediction[[:space:]]Using[[:space:]]Distance[[:space:]]Transform[[:space:]]Over[[:space:]]Pre-semantic[[:space:]]Contours[[:space:]]with[[:space:]]Self-supervised[[:space:]]Neural[[:space:]]Networks/18d0caf1-d799-41ef-95e4-0a0e2105060c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improved[[:space:]]Video[[:space:]]VAE[[:space:]]for[[:space:]]Latent[[:space:]]Video[[:space:]]Diffusion[[:space:]]Model/d6e6ed64-039e-4e0b-a0ce-69a2cf999362_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improving[[:space:]]Accuracy[[:space:]]and[[:space:]]Calibration[[:space:]]via[[:space:]]Differentiated[[:space:]]Deep[[:space:]]Mutual[[:space:]]Learning/6a2c48a3-aafc-4ad2-9262-8700829f154e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improving[[:space:]]Adversarial[[:space:]]Transferability[[:space:]]on[[:space:]]Vision[[:space:]]Transformers[[:space:]]via[[:space:]]Forward[[:space:]]Propagation[[:space:]]Refinement/c0d9e11b-700c-40d7-8dca-46321613cb22_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improving[[:space:]]Autoregressive[[:space:]]Visual[[:space:]]Generation[[:space:]]with[[:space:]]Cluster-Oriented[[:space:]]Token[[:space:]]Prediction/2639ed20-911e-44d7-9142-a8fcfe0f6597_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improving[[:space:]]Diffusion[[:space:]]Inverse[[:space:]]Problem[[:space:]]Solving[[:space:]]with[[:space:]]Decoupled[[:space:]]Noise[[:space:]]Annealing/d107e339-bf08-4731-b903-e7f3360d585d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improving[[:space:]]Editability[[:space:]]in[[:space:]]Image[[:space:]]Generation[[:space:]]with[[:space:]]Layer-wise[[:space:]]Memory/5ef2c05f-1b0c-4899-806c-c1cd90e73e44_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improving[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]with[[:space:]]Localized[[:space:]]Points[[:space:]]Management/6436626a-e39e-4d87-9a2d-0ebf83ce94c9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improving[[:space:]]Personalized[[:space:]]Search[[:space:]]with[[:space:]]Regularized[[:space:]]Low-Rank[[:space:]]Parameter[[:space:]]Updates/673cbe39-bd0b-47dd-99b2-3c870c8defd5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improving[[:space:]]Semi-Supervised[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]with[[:space:]]Sliced-Wasserstein[[:space:]]Feature[[:space:]]Alignment[[:space:]]and[[:space:]]Uniformity/ca5cab4a-bdaf-48a0-9763-5efeeaa13237_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improving[[:space:]]Sound[[:space:]]Source[[:space:]]Localization[[:space:]]with[[:space:]]Joint[[:space:]]Slot[[:space:]]Attention[[:space:]]on[[:space:]]Image[[:space:]]and[[:space:]]Audio/feaa7f56-7770-404f-a305-b468e86df1e5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improving[[:space:]]Transferable[[:space:]]Targeted[[:space:]]Attacks[[:space:]]with[[:space:]]Feature[[:space:]]Tuning[[:space:]]Mixup/5356bb72-29a6-442b-984c-adce3c009504_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improving[[:space:]]Visual[[:space:]]and[[:space:]]Downstream[[:space:]]Performance[[:space:]]of[[:space:]]Low-Light[[:space:]]Enhancer[[:space:]]with[[:space:]]Vision[[:space:]]Foundation[[:space:]]Models[[:space:]]Collaboration/0e31d729-0f45-4249-a70f-af6d4389a18f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improving[[:space:]]the[[:space:]]Training[[:space:]]of[[:space:]]Data-Efficient[[:space:]]GANs[[:space:]]via[[:space:]]Quality[[:space:]]Aware[[:space:]]Dynamic[[:space:]]Discriminator[[:space:]]Rejection[[:space:]]Sampling/9b5acaea-53bf-4055-87a9-1a5bf68ccd00_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Improving[[:space:]]the[[:space:]]Transferability[[:space:]]of[[:space:]]Adversarial[[:space:]]Attacks[[:space:]]on[[:space:]]Face[[:space:]]Recognition[[:space:]]with[[:space:]]Diverse[[:space:]]Parameters[[:space:]]Augmentation/aa0474ee-612f-4595-88a1-c6ba9ec15b27_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Imputation-free[[:space:]]and[[:space:]]Alignment-free_[[:space:]]Incomplete[[:space:]]Multi-view[[:space:]]Clustering[[:space:]]Driven[[:space:]]by[[:space:]]Consensus[[:space:]]Semantic[[:space:]]Learning/380dd383-b00d-4370-9caa-3f9395391322_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/InPO_[[:space:]]Inversion[[:space:]]Preference[[:space:]]Optimization[[:space:]]with[[:space:]]Reparametrized[[:space:]]DDIM[[:space:]]for[[:space:]]Efficient[[:space:]]Diffusion[[:space:]]Model[[:space:]]Alignment/14854e80-d5af-4094-8f56-d57259d0057b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/IncEventGS_[[:space:]]Pose-Free[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Event[[:space:]]Camera/c5e31ebc-2f5f-4333-a4a5-98739dcdbce4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Incomplete[[:space:]]Multi-View[[:space:]]Multi-label[[:space:]]Learning[[:space:]]via[[:space:]]Disentangled[[:space:]]Representation[[:space:]]and[[:space:]]Label[[:space:]]Semantic[[:space:]]Embedding/419f8c28-377b-4f29-8c8b-a2cba093a85b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Incomplete[[:space:]]Multi-modal[[:space:]]Brain[[:space:]]Tumor[[:space:]]Segmentation[[:space:]]via[[:space:]]Learnable[[:space:]]Sorting[[:space:]]State[[:space:]]Space[[:space:]]Model/58ab88d4-b760-40c9-8213-6c50aa676cf3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Incorporating[[:space:]]Dense[[:space:]]Knowledge[[:space:]]Alignment[[:space:]]into[[:space:]]Unified[[:space:]]Multimodal[[:space:]]Representation[[:space:]]Models/f0305c72-f982-4ece-ab22-ea79e8e3955f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Incremental[[:space:]]Object[[:space:]]Keypoint[[:space:]]Learning/d3b12539-b476-45cb-a752-bae4990c3622_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/IndoorGS_[[:space:]]Geometric[[:space:]]Cues[[:space:]]Guided[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Indoor[[:space:]]Scene[[:space:]]Reconstruction/848b3be8-e7e7-4fb9-82b5-d4028969c01d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Inference-Scale[[:space:]]Complexity[[:space:]]in[[:space:]]ANN-SNN[[:space:]]Conversion[[:space:]]for[[:space:]]High-Performance[[:space:]]and[[:space:]]Low-Power[[:space:]]Applications/6b151e98-3001-4e0f-b77a-ea1941036ad2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Infighting[[:space:]]in[[:space:]]the[[:space:]]Dark_[[:space:]]Multi-Label[[:space:]]Backdoor[[:space:]]Attack[[:space:]]in[[:space:]]Federated[[:space:]]Learning/bab57aa6-5f36-4253-ae3a-1e5e1f4632d7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Infinity_[[:space:]]Scaling[[:space:]]Bitwise[[:space:]]AutoRegressive[[:space:]]Modeling[[:space:]]for[[:space:]]High-Resolution[[:space:]]Image[[:space:]]Synthesis/d1db9220-1c65-401b-9202-6e181bfed6b4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/InsTaG_[[:space:]]Learning[[:space:]]Personalized[[:space:]]3D[[:space:]]Talking[[:space:]]Head[[:space:]]from[[:space:]]Few-Second[[:space:]]Video/1014c4e5-3a69-4fa3-98b3-a879e2f45651_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Insight-V_[[:space:]]Exploring[[:space:]]Long-Chain[[:space:]]Visual[[:space:]]Reasoning[[:space:]]with[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models/cb8122c4-a3cd-447b-9677-ef086abe8658_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/InsightEdit_[[:space:]]Towards[[:space:]]Better[[:space:]]Instruction[[:space:]]Following[[:space:]]for[[:space:]]Image[[:space:]]Editing/fb96c75a-1b6e-4864-bf83-3b9ecc089e1a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Insightful[[:space:]]Instance[[:space:]]Features[[:space:]]for[[:space:]]3D[[:space:]]Instance[[:space:]]Segmentation/275c7e3b-7085-4434-b01d-e129139897fa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Inst3D-LMM_[[:space:]]Instance-Aware[[:space:]]3D[[:space:]]Scene[[:space:]]Understanding[[:space:]]with[[:space:]]Multi-modal[[:space:]]Instruction[[:space:]]Tuning/ada2a42d-606c-400d-85d5-a68b46a5f62b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Instance-wise[[:space:]]Supervision-level[[:space:]]Optimization[[:space:]]in[[:space:]]Active[[:space:]]Learning/6ab274ce-c563-45f8-9f14-82dfa36b95f4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/InstanceCap_[[:space:]]Improving[[:space:]]Text-to-Video[[:space:]]Generation[[:space:]]via[[:space:]]Instance-aware[[:space:]]Structured[[:space:]]Caption/63b6efae-0d11-487a-b293-b1b579ef5966_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/InstanceGaussian_[[:space:]]Appearance-Semantic[[:space:]]Joint[[:space:]]Gaussian[[:space:]]Representation[[:space:]]for[[:space:]]3D[[:space:]]Instance-Level[[:space:]]Perception/e6f2e48d-8efc-4d47-81e4-de4b789d0ecb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Instant[[:space:]]Adversarial[[:space:]]Purification[[:space:]]with[[:space:]]Adversarial[[:space:]]Consistency[[:space:]]Distillation/661ff35f-282c-423b-9185-bb6d8ff5da4b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Instant[[:space:]]Gaussian[[:space:]]Stream_[[:space:]]Fast[[:space:]]and[[:space:]]Generalizable[[:space:]]Streaming[[:space:]]of[[:space:]]Dynamic[[:space:]]Scene[[:space:]]Reconstruction[[:space:]]via[[:space:]]Gaussian[[:space:]]Splatting/c9f2cc0e-b9b5-4129-a1a3-c1fc202a4666_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Instant3dit_[[:space:]]Multiview[[:space:]]Inpainting[[:space:]]for[[:space:]]Fast[[:space:]]Editing[[:space:]]of[[:space:]]3D[[:space:]]Objects/671f9dcf-f80a-4cc0-9e9d-72da3c00b249_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Instruct-CLIP_[[:space:]]Improving[[:space:]]Instruction-Guided[[:space:]]Image[[:space:]]Editing[[:space:]]with[[:space:]]Automated[[:space:]]Data[[:space:]]Refinement[[:space:]]Using[[:space:]]Contrastive[[:space:]]Learning/a04fb8db-d9e4-4e5e-a8c0-00327c1bd422_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Instruction-based[[:space:]]Image[[:space:]]Manipulation[[:space:]]by[[:space:]]Watching[[:space:]]How[[:space:]]Things[[:space:]]Move/6db53ffc-36dc-4f4f-8a58-b9c5bb9ce24b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Integral[[:space:]]Fast[[:space:]]Fourier[[:space:]]Color[[:space:]]Constancy/9fd420e6-b4fa-48bd-940e-f8dcc99e9a7d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/InterAct_[[:space:]]Advancing[[:space:]]Large-Scale[[:space:]]Versatile[[:space:]]3D[[:space:]]Human-Object[[:space:]]Interaction[[:space:]]Generation/4511c287-17f9-4b70-a6d5-d77c48a5192a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/InterDyn_[[:space:]]Controllable[[:space:]]Interactive[[:space:]]Dynamics[[:space:]]with[[:space:]]Video[[:space:]]Diffusion[[:space:]]Models/1657842e-f939-4237-87d3-49f995d53e68_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/InterMimic_[[:space:]]Towards[[:space:]]Universal[[:space:]]Whole-Body[[:space:]]Control[[:space:]]for[[:space:]]Physics-Based[[:space:]]Human-Object[[:space:]]Interactions/c17df58e-4f4a-4403-b950-698942e97f1f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/InteractAnything_[[:space:]]Zero-shot[[:space:]]Human[[:space:]]Object[[:space:]]Interaction[[:space:]]Synthesis[[:space:]]via[[:space:]]LLM[[:space:]]Feedback[[:space:]]and[[:space:]]Object[[:space:]]Affordance[[:space:]]Parsing/bba3c5bb-a0b8-4050-9fca-51eeaf9696e0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/InteractVLM_[[:space:]]3D[[:space:]]Interaction[[:space:]]Reasoning[[:space:]]from[[:space:]]2D[[:space:]]Foundational[[:space:]]Models/9f03277a-37b8-49e7-9fad-13e2e5f171b2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/InteractionMap_[[:space:]]Improving[[:space:]]Online[[:space:]]Vectorized[[:space:]]HDMap[[:space:]]Construction[[:space:]]with[[:space:]]Interaction/447efac7-303b-4880-a09c-b0e941c7cbb1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Interactive[[:space:]]Medical[[:space:]]Image[[:space:]]Analysis[[:space:]]with[[:space:]]Concept-based[[:space:]]Similarity[[:space:]]Reasoning/a8ba1cd8-3d09-4bb8-8912-43eb50d5dbf1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Interactive[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation_[[:space:]]A[[:space:]]Benchmark[[:space:]]Dataset[[:space:]]and[[:space:]]Baseline/72ec2f30-d729-4ccc-bc2d-90a1d1bd3d41_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Interleaved-Modal[[:space:]]Chain-of-Thought/3edf2a1c-2168-41b7-a536-b8e17bc35145_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Interpretable[[:space:]]Generative[[:space:]]Models[[:space:]]through[[:space:]]Post-hoc[[:space:]]Concept[[:space:]]Bottlenecks/86d682d5-19b8-4583-bdec-4fcac9608ef2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Interpretable[[:space:]]Image[[:space:]]Classification[[:space:]]via[[:space:]]Non-parametric[[:space:]]Part[[:space:]]Prototype[[:space:]]Learning/d624a8da-e662-4173-965d-8a4cb70d4b42_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Interpreting[[:space:]]Object-level[[:space:]]Foundation[[:space:]]Models[[:space:]]via[[:space:]]Visual[[:space:]]Precision[[:space:]]Search/396da971-babe-4c65-8e62-f7e2e9a705a9_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/Image Quality Assessment_ From Human to Machine Preference/9c43c7f6-7667-4c47-8440-32e15bc27376_content_list.json b/2025/Image Quality Assessment_ From Human to Machine Preference/9c43c7f6-7667-4c47-8440-32e15bc27376_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c0fe5932ed7547297f7116c967ed6123ddd9735d --- /dev/null +++ b/2025/Image Quality Assessment_ From Human to Machine Preference/9c43c7f6-7667-4c47-8440-32e15bc27376_content_list.json @@ -0,0 +1,1948 @@ +[ + { + "type": "text", + "text": "Image Quality Assessment: From Human to Machine Preference", + "text_level": 1, + "bbox": [ + 171, + 130, + 826, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chunyi Li $^{1,2,4*}$ , Yuan Tian $^{1,2*}$ , Xiaoyue Ling $^{1*}$ , Zicheng Zhang $^{1,2}$ , Haodong Duan $^{2}$ , Haoning Wu $^{3,4}$ , Ziheng Jia $^{1}$ , Xiaohong Liu $^{1}$ , Xiongkuo Min $^{1}$ , Guo Lu $^{1\\dagger}$ , Weisi Lin $^{4}$ , Guangtao Zhai $^{1,2\\dagger}$ \nShanghai Jiao Tong University $^{1}$ , Shanghai AI Lab $^{2}$ , Moonshot AI $^{3}$ , Nanyang Technological University $^{4}$", + "bbox": [ + 99, + 179, + 916, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d346aa8b3722dceff3a924054764b758fe8d0bd085c64b4c82ffc951c515063.jpg", + "image_caption": [ + "Figure 1. The significant gap between the well-explored Human Vision System (HVS) and the emerging Machine Vision System (MVS). Their different perception mechanisms leading to images that are subjectively satisfactory to humans not being applicable to the machines downstream tasks such as detection and question answering, vice versa." + ], + "image_footnote": [], + "bbox": [ + 101, + 243, + 903, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 517, + 326, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image Quality Assessment (IQA) based on human subjective preferences has undergone extensive research in the past decades. However, with the development of communication protocols, the visual data consumption volume of machines has gradually surpassed that of humans. For machines, the preference depends on downstream tasks such as segmentation and detection, rather than visual appeal. Considering the huge gap between human and machine visual systems, this paper proposes the topic: Image Quality Assessment for Machine Vision for the first time. Specifically, we (1) defined the subjective preferences of machines, including downstream tasks, test models, and evaluation metrics; (2) established the Machine Preference Database (MPD), which contains 2.25M fine-grained annotations and 30k reference/distorted image pair instances; (3) verified the performance of mainstream IQA algorithms on MPD. Experiments show that current IQA metrics are human-centric and cannot accurately characterize machine preferences. We sincerely hope that MPD can promote the evolution of IQA from human to machine preferences. Project page is on: https://github.com/lcyszxdxc/MPD.", + "bbox": [ + 88, + 550, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 517, + 643, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid ascend of smart cities [20] and the flourishment of the Internet of Things (IoT) [86] over the past decade have fundamentally transformed the framework of the application end. According to the Cisco [13] white paper, the number of Machine-to-Machine (M2M) connections first exceeded that of Machine-to-Human (M2H) in 2023, reaching 147 billion. Machines have gradually replaced humans and become the primary consumers of image/video data. In the field of image processing, the primary goal is to enhance the quality of processed images, ensuring they conform to human perception. Previously, the performance of a compression/reconstruction/generation algorithm [39-41] is often judged by its visual appeal to humans. This technology, known as Image Quality Assessment (IQA), comprehensively models the Human Visual System (HVS) to predict human subjective preferences. Consequently, with the evolution of application ends, IQA today has to transcend HVS and serve Machine Vision Systems (MVS) more effectively.", + "bbox": [ + 511, + 547, + 906, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, there are significant disparities in the perceptual mechanisms between HVS and MVS. This discrepancy arises because HVS prioritizes the similarity between reference and distorted images, such as texture, structure, and color, whereas machines focus on the consistency of results", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 114, + 895, + 228, + 909 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 245, + 895, + 379, + 909 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7570", + "bbox": [ + 482, + 948, + 514, + 960 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/8598cbef269baa30bcd78548b52fdbb96c4674d82aed4ae20f7bebafd2047a9f.jpg", + "table_caption": [ + "Table 1. Comparison of MPD with other IQA databases. As the first machine preference-oriented library, MPD has a larger image scale, a wider range of corruption scenarios, and more complete annotation content." + ], + "table_footnote": [], + "table_body": "
DatabaseImageCorruptionAnnotation
ReferenceDistortedResolutionTypesStrengthNumDimensionLabeling Criteria
LIVE[57]297797685525k1Human preference on NSIs
TID2013[61]253k512245514k1Human preference on NSIs
KADID-10K[47]8110k1k255304k1Human preference on NSIs
CLIC2021[6]5853k1k103484k1Human preference on compressed images
NTIRE2022[24]25029k2884051.13m1Human preference on super-resolution images
KonIQ-10K[29]-10k1k--1.21m5Human preference on NSIs
AGIQA-3K[35]-3k1k--125k2Human preference on AIGIs
NTIRE2024[38]-20k1k--420k2Human preference on AIGIs
MPD (ours)1k30k1k3052.25m5Machine preference on NSIs, SCIs, and AIGIs
", + "bbox": [ + 114, + 122, + 885, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in downstream tasks. As illustrated in Figure 1, for distortions resulting from factors like encoding/quantization, HVSs perceive a noticeable degradation in quality, yet the performance of tasks like segmentation/detection by machines remains unaffected. Conversely, some minor disturbances that the human eye barely notices can significantly disrupt the output of MVSs, making it difficult to migrate their evaluation algorithms.", + "bbox": [ + 88, + 318, + 480, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "What images do machines prefer? This remains an open question. For HVS, over the past two decades, IQA has witnessed significant development. Given the variations in image preferences among individuals, particularly across dimensions such as brightness and chroma, the scoring dimensions, the annotations from multiple subjects, and the integration of these annotations into global subjective preference labels have been well established. Hundreds of fine-grained databases can drive IQA algorithms that closely align with human perception. Conversely, MVS development is much slower, regardless of an end-to-end IQA algorithm, there is even no database that annotates images with an overall preference score of machines. This hinders the comprehensive evaluation of image processing metrics designed for machine vision, limiting the experience of application ends. Considering these issues, we first attempt to generalize the concept of IQA from HVS to MVS. Our contributions can be summarized as follows:", + "bbox": [ + 91, + 441, + 482, + 712 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A innovative definition of the subjective preferences of machines. We propose a standardized process for collecting subjective labels from machines. Specifically, we identify the downstream tasks that determine the quality of machine perception, including four multimodal tasks related to Large Multimodal Models (LMMs) and three classic Computer Vision (CV) tasks. By referencing the multi-dimensional scoring mechanism of humans, we comprehensively evaluate the performance of machines in these tasks to obtain ground truth labels.", + "- A large-scale database named the Machine Preference Database (MPD). Using 15 general LMMs and 15 spe" + ], + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cialized CV models, we collected 2.25 million preference data points on 30,000 reference/distorted image pairs, establishing the first IQA database for machine vision. In particular, we conducted an in-depth analysis of each participating machine to explore the internal patterns of MVS perception and their differences with HVS.", + "bbox": [ + 524, + 318, + 906, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- A complex benchmark test for existing IQA algorithms. Experiments on the MPD demonstrate that these algorithms designed for HVS cannot fully model MVS, whose predictions do not match machine ratings. After finetuning based on the MPD, their performance improves slightly, but still cannot characterize machine preferences, especially when facing mild distortion.", + "bbox": [ + 513, + 409, + 906, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 530, + 663, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Quality Assessment Database", + "text_level": 1, + "bbox": [ + 511, + 556, + 828, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "IQA datasets include Full Reference (FR) and No Reference (NR). FR datasets manually add corruption to reference images, whereas NR directly capture distorted images. Typically, these datasets focus on a single image type, such as Natural Scenes Images (NSIs) [6, 24, 47, 57, 61], Screen Content Images (SCIs) [21, 56, 58], and AI-Generated Images (AIGIs) [35, 38, 51, 91], with quality labels derived solely from human preferences as listed in Table 1. On contrast, in the era where machines surpass humans as application ends, MPD stands out as the first machine preference-oriented library. It offers a larger image scale, a broader range of corruption scenarios, and a more comprehensive annotation context than other IQA databases.", + "bbox": [ + 509, + 578, + 906, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Machine Vision Downstream Task", + "text_level": 1, + "bbox": [ + 511, + 787, + 812, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the absence of a unified quality indicator, image processing algorithms, taking compression and restoration metrics for machine vision for example, are often in a fragmented state, with the performance of each algorithm being validated by only one or two downstream tasks. [93, 94] These tasks vary, as do the models used to perform them, mak", + "bbox": [ + 511, + 810, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7571", + "bbox": [ + 482, + 948, + 513, + 960 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/673ed7353ff32666bd066c02e6968121cce7d918dd5ab57864b91b5f7406ca12.jpg", + "image_caption": [ + "Figure 2. Overview of Machine Preference Database (MPD), with 30k large-scale reference/distorted image pairs, meticulously annotated with 2.25M fine-grained preference score from the mean opinion of 15 mainstream LMMs and 15 specific CV models. It provides a multi-dimensional fidelity evaluation of various downstream machine vision tasks." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 906, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing the evaluation criteria disparate and incomplete. For compression algorithms, validation indicators include the classification [30, 60], image captioning [12], object detection [80], or semantic segmentation [31] result before/after compression, with each indicator differing and lacking standardization or completeness. Restoration algorithms fare slightly better, with most consistently relying on object detection [48, 83, 84]. Nevertheless, some employ traditional Fast-RCNN [65], while others utilize more advanced models like ViTDet[34]. The variation in model bases further compromises the comprehensiveness and objectivity of the evaluation outcomes. Thus, there is an urgent need for a unified representation of machine preferences.", + "bbox": [ + 88, + 536, + 485, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Database Construction", + "text_level": 1, + "bbox": [ + 89, + 747, + 308, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Reference & Distorted Image Collection", + "text_level": 1, + "bbox": [ + 89, + 772, + 434, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to the previous IQA database, we first selected 1,000 high-quality reference images and then artificially degraded their quality to obtain 30,000 pairs of reference and distorted images. To comprehensively represent the common content on the internet, we referred to the CMC-Bench [42] and selected the three most common image categories: Natural Scene Images (NSIs), including", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "User/Professional Generated Content (UGC/PGC) [6, 46]; Screen Content Images (SCIs), including webpages [67], games [92], and movies [56]; and AI-Generated Images (AIGIs), generated by six representative text-to-image models [8, 25, 28, 43, 64, 66]. The collection process is shown in Figure 2. Both human viewers and IQA models have verified this image sequence to ensure that they are undistorted as reference images.", + "bbox": [ + 511, + 536, + 906, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the distorted images, we modeled the corruption caused by the current communication protocols, which were divided into 30 scenarios. Specifically, these corruptions were classified into seven categories: Blur, various types of unclear image; Luminance, global brightness changes; Chrominance, global color changes; Contrast, global changes in contrast/saturation; Noise, random noise of different distributions; Compression, codec algorithm like JPEG; and Spatial, local pixel-level changes. For each corruption, we defined five levels of intensity, from mild to severe. To fairly explore the impact of each corruption, we artificially control the distortion strength to ensure that quality degradation perceived by the human eye is", + "bbox": [ + 511, + 659, + 908, + 857 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1For the definition of all these 30 corruptions, specific details, and examples, please refer to the appendix.", + "bbox": [ + 511, + 875, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7572", + "bbox": [ + 482, + 948, + 514, + 960 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "aligned at the same level. Thus, for each reference image, we randomly selected the intensity to add all the corruptions mentioned above, resulting in 30,000 distorted images. The reference/distorted image pairs will then be annotated by machine subjects.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Different Task Annotation", + "text_level": 1, + "bbox": [ + 89, + 174, + 330, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the past five years, the evaluation of image processing algorithms for machine vision has primarily focused on seven tasks: Regression, Classification, Visual Question Answering (VQA), Caption (CAP), Segmentation (SEG), Detection (DET), and Retrieval (RET). Among these, the first four tasks can be accomplished by general LMMs, whereas the latter three tasks necessitate specialized models due to their unique requirements. Note that the regression and classification tasks in this context are analogous to Yes-or-No (YoN) and Multiple-Choice Questions (MCQ) within LMMs. To this end, we enlisted five experienced human experts to design YoN/VQA/CAP questions and answers for 1,000 reference images. These questions are designed to be challenging, as illustrated in the example provided in Figure 2, especially in the MCQ sets where confusion options are introduced. This approach enables us to accurately measure the degree of quality degradation perceived by the machine.", + "bbox": [ + 89, + 196, + 483, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We apply these seven tasks to both the reference and distorted images, measure the similarity of the output results, and derive scores for each of the seven sub-dimensions. Specifically, for $Y_{oN}$ tasks, we employ a logit-level evaluation approach based on softmax similarity, as previous LMM benchmarks [50, 78] do. This entails LMMs confirming or denying the questions designed for each reference/distorted image pair. The confidence difference in the next token output, 'Yes' or 'No', reflects the difference in the perceptual quality degradation from the perspective of machines, namely a score $S_{Y_{oN}}$ :", + "bbox": [ + 89, + 453, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {Y o N} = \\left| \\sigma \\left(P _ {\\text {d i s}} (\\text {Y e s}, \\text {N o})\\right) - \\sigma \\left(P _ {\\text {r e f}} (\\text {Y e s}, \\text {N o})\\right) \\right|, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 626, + 480, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{dis,ref}(\\cdot)$ denotes the output probability of certain logits, normalized by the softmax function $\\sigma(\\cdot)$ . MCQ tasks follow a similar evaluation paradigm. We set a correct option and three confusing options (the options are usually no more than 4, see [18, 50, 79]), and the model outputs the probability of each option as a four-element vector. The distance between the two vectors in space can reflect the perceptual quality score $S_{MCQ}$ for machines:", + "bbox": [ + 89, + 648, + 483, + 772 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {M C Q} = \\cos \\left(P _ {d i s} (\\mathrm {A}, \\mathrm {B}, \\mathrm {C}, \\mathrm {D}), P _ {r e f} (\\mathrm {A}, \\mathrm {B}, \\mathrm {C}, \\mathrm {D})\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 777, + 482, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the distance is measured by cosine similarity. For $VQA$ tasks, we conduct sentence-level evaluation beyond the logits. LMMs will be required to answer a certain question in about 5 words, and the semantic difference between two output sentences will be the score $S_{VQA}$ :", + "bbox": [ + 89, + 801, + 483, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {V Q A} = \\operatorname {C L I P} \\left(T _ {\\text {d i s}}, T _ {\\text {r e f}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 885, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d2bae0fc7fe3d3f3f89cd0af85c937ba37dd6748e5e2e3a17fc4a817173b4015.jpg", + "image_caption": [ + "(a) Correlation matrix" + ], + "image_footnote": [], + "bbox": [ + 524, + 90, + 743, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e7c910a6015712bc19d0552851df8d958166593387a3df4d2849170438e5d109.jpg", + "image_caption": [ + "(b) Score distribution", + "Figure 3. Correlation between the general preference score and the score in seven different downstream tasks." + ], + "image_footnote": [], + "bbox": [ + 751, + 97, + 890, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T_{dis,ref}$ represents the whole output text, measured by the CLIP text encoder. $CAP$ tasks require a longer output, by describing the main subjects, actions, settings, and any notable objects or features of the image in about 40 words. Considering the remarkable text length, we use indicators dedicated to image captioning to analyze the differences between the two long sentences in $S_{CAP}$ :", + "bbox": [ + 511, + 329, + 905, + 435 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {C A P} = \\sum_ {\\text {e v a l}} \\operatorname {e v a l} \\left(T _ {d i s}, T _ {r e f}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 446, + 905, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where eval $\\in$ (BLEU, CIDEr, SPICE) [4, 59, 72], three most commonly-used captioning indicators. Then, for SEG tasks, referring to their common evaluation criteria [26], we calculate the Intersection-over-Union (IoU) ratio in $S_{SEG}$ :", + "bbox": [ + 511, + 489, + 905, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {S E G} = \\operatorname {I o U} \\left(M _ {d i s}, M _ {r e f}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 563, + 903, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_{dis,ref}$ stands for binarized segmentation masks. DET tasks will be more complex, whose results include both spatial information and category labels. Thus, we consider the detection failed when the categories are different, and only further calculate the IoU when they are consistent [96], as described in $S_{DET}$ :", + "bbox": [ + 511, + 592, + 905, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {D E T} = \\operatorname {A c c} _ {1} \\left(C _ {d i s}, C _ {r e f}\\right) \\cdot \\operatorname {I o U} \\left(B _ {d i s}, B _ {r e f}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 695, + 903, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $C_{dis,ref}$ denotes the category sequence ordered by confidence, and $B_{dis,ref}$ is the detected bounding box. Finally for RET tasks, we merge three common Image-to-Text retrieval indicators [7], namely Top-1/5/10 accuracy, to evaluate the retrieved sequence as $S_{RET}$ :", + "bbox": [ + 511, + 723, + 905, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {R E T} = \\sum_ {\\mathrm {i} \\in 1, 5, 1 0} \\operatorname {A c c} _ {\\mathrm {i}} \\left(C _ {d i s}, C _ {r e f}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 811, + 903, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the above process, we can obtain the preferences of LMMs in the first four dimensions, and the preferences of each specific CV model on its target task.", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7573", + "bbox": [ + 482, + 948, + 514, + 960 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e8d65818a113bf9db39eef65672b57531532cab7b2e8e46415876d5b43f55b1c.jpg", + "image_caption": [ + "Gaussian...filter" + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 228, + 133 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/db3f628068b730c7a69e833d98f294b9ef40d8ff65226b69281711602df07bf2.jpg", + "image_caption": [ + "Lens...blur" + ], + "image_footnote": [], + "bbox": [ + 230, + 90, + 364, + 136 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1c212d6e249c3417c8c55869afbaffc0a924ffc9efcaf1480208a7103bebac37.jpg", + "image_caption": [ + "Motion...blur" + ], + "image_footnote": [], + "bbox": [ + 366, + 90, + 498, + 136 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bd6925fd0641dc3a5421fe1ffac387c86e3452f6a9ad279cac5e2c8287428640.jpg", + "image_caption": [ + "Color...diffusion" + ], + "image_footnote": [], + "bbox": [ + 500, + 90, + 632, + 136 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/75fb09b407bf5b3246d0c0e2b871aa824ec78fb3ed8f886e7a36db5f249742bd.jpg", + "image_caption": [ + "Color...shift" + ], + "image_footnote": [], + "bbox": [ + 633, + 90, + 766, + 136 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/22d3543ebc8db02e13ff242573fe268d52b9a499019c9ec30007e2d4691e093f.jpg", + "image_caption": [ + "Color...quantization" + ], + "image_footnote": [], + "bbox": [ + 767, + 90, + 900, + 136 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/59b4dc0f28f900c096aeb50c01cd4f8718782cb4eed09ebc56c9e59faa57ea79.jpg", + "image_caption": [ + "HSV...saturation" + ], + "image_footnote": [], + "bbox": [ + 94, + 146, + 228, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1257e89ecb593dbf6afb89b344e384087939a7eaa029a06c58634d50f739606d.jpg", + "image_caption": [ + "Lab...saturation" + ], + "image_footnote": [], + "bbox": [ + 230, + 146, + 362, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/35a7e7be1a60ea062193f22f600003864dc75c9eb912ed965e1a3f0a9c23fd1f.jpg", + "image_caption": [ + "JP2K...compression" + ], + "image_footnote": [], + "bbox": [ + 366, + 146, + 498, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/efc4e2b5af31b2fdc7949355d8cbafb7e0f1247b9740dc049f96061792a2b7d7.jpg", + "image_caption": [ + "JPEG...compression" + ], + "image_footnote": [], + "bbox": [ + 500, + 146, + 632, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2f3d141369cb920ba485bbed96c11e515aafdb552d3e36e3ae9198ccf7c6ee4b.jpg", + "image_caption": [ + "WEBP...compression" + ], + "image_footnote": [], + "bbox": [ + 633, + 146, + 766, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8ed2cc6cf417c7c524641651d9bf03aebb8c47de42cd9b4c57ff61924ff65e9e.jpg", + "image_caption": [ + "White...noise" + ], + "image_footnote": [], + "bbox": [ + 767, + 146, + 900, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/88b66cc9a3e3ad8fc3e335d902dc36683ab9662c74ce131e3b2c728d7a4ba7dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 202, + 228, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/46d1e2daed45ff8ee19f0338adbe1808d8bb4c7d62a97ab5ed4c7d1fb180ddb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 202, + 362, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/18bbb6e8651196f9254edffb7e7e13b0593dc09b471110380e0f06a99fc395ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 202, + 498, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fc8f917b543e54f7e2078c9d335041f5c37de5627322668de6861dbdd2f33359.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 202, + 632, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee80b749df35adb05c60d05e59c0f80059074846d042f29655f28c1e2ac3fc6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 202, + 766, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5f13ce1bcbf5f3674405705c9ed11f5d1f241f864ab65cb4493a3ac7388ecf8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 202, + 900, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/300f69721f4929300ec94ec5f5b58890bd59b38f4e855e79a9421808136aa2d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 258, + 228, + 313 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1f6804de1385154997b507f3fe6ca1762d1a03936cc7ff72d166b3eeabe4e081.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 258, + 362, + 313 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1ec7fc27464c5cd3b12af9ad3bd3fc1ff28905c4066271f8bf046f838aa4314a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 258, + 498, + 313 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f54ae830e0f90b05116bd9a1eedde014f1bda51a373d2bad8b209c5522858c59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 258, + 632, + 313 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af3155f66e6d9ccf282555c882731163e7ba691bd666d4139753f64baef0a55c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 258, + 766, + 313 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/77acb24a647dd17b6f22829f6f62d22bff02f79238b27cb36798741526756d4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 258, + 900, + 313 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/33c98ebaee779c2ba8f22326c350fc20bef7b2eb6de7ac20a837919011a33ad5.jpg", + "image_caption": [ + "Figure 4. MOS score of MPD, visualized in 30 corruption subsets. Different color denotes corruption strength Level 1-Level 2-Level 3-Level 4-Level 5. Results show the sensitivity of the machines to each corruption varied significantly." + ], + "image_footnote": [], + "bbox": [ + 94, + 315, + 228, + 369 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5027069540ea3bccbeb6aa54dde68289b5297373efb222dd005c36037a48aa6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 315, + 362, + 369 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b3513eb939171d21592e79de294567b69057083d38c250082d139c15e01748a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 315, + 498, + 369 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0c4211e552dcb5536a75c26c01be1e2c7f939c858c6776df8b06538a6803c345.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 315, + 632, + 369 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e73fa53bc956bc2ceac7b635b046354e1742cc5cdc74a388c8f9d7411dad0655.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 315, + 766, + 369 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4b8229b3b26febf67320c8bb388a72ad7b5cc44843f5f1bdb41ed485a79c43b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 315, + 900, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Preference Score Collection", + "text_level": 1, + "bbox": [ + 89, + 422, + 339, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the ITU human subjective annotation standards [71], we adopt the same settings to obtain the machine Mean Opinion Score (MOS). Specifically, for YoN, MCQ, VQA, and CAP tasks, we use 15 general LMMs for inference, namely DeepseekVL [53], InstructBLIP [14], InternVL2 [11], InternLM-XComposer2 [16], LLaVA1.5 [49], LLaVANext [44], LLaVA-OneVision [37], Mantis [32], MiniInternVL [22], MPlugOwl3 [81], Ovis [54], Phi3.5 [1], Pangea [85], Qwen2-VL [75], and Yi1.5 [52]. We set all temperature parameters as 0 to avoid unstable output to ensure all perceptual quality degradations come from distortion. Comprehensively considering popular trends and performance, we implement 5 models for each specific CV tasks, namely SEG: BoxInst [70], ConvNext [77], Mask-RCNN [27], SCNet [73], and RTMDet (segmentation mode) [55]; DET: Dino[89], Mask-RCNN [65], RTMDet (detection mode) [55], ViTDet [34], and Yolo-X [23]; RET: DIME [63], InternVL (retrieval-based tuning) [10], LAVIS (retrieval-based tuning) [36], NAAF [90], and $\\mathrm{VSE}++$ [19]. After obtaining scores for these 7 tasks, the evaluation can be divided into five dimensions, one dimension for each first four LMM tasks, and the last three tasks are unified into the last dimension, ensuring 15 subjects in each $(0,1)$ score for MOS. Thus merging all five dimensions into the $(0,5)$ MOS. Note that we normalize the score of each dimension to avoid where one dimension has a higher score while another is lower (e.g. VQA/CAP), which would lead to unfairness in the weight of the overall MOS.", + "bbox": [ + 89, + 446, + 483, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, considering the rapid development of LMM, we also apply 3 advanced LMMs, GPT4o [2], GPT4o-", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mini [2], InternLM-XComposer2d5 [16]; and 4 early-stage LMMs, QWen1.5-VL [5], VisualGLM [17], Monkey [45], MPlugOwl2 [82] on certain tasks. We compared these results with the above-mentioned data in MPD, to prove the long-term usability of MPD through their consistency.", + "bbox": [ + 511, + 424, + 905, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Database Validation", + "text_level": 1, + "bbox": [ + 511, + 512, + 705, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section analyzes the preference data characteristics of MPD, to prove the usability of this database, including two perspectives: (1) the mean opinion of all machines; (2) the difference between each machine as a subject.", + "bbox": [ + 511, + 537, + 906, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall: The MOS of MPD and the sub-scores of the seven tasks are illustrated in Figure 3, with their correlation quantified by Spearman Rank-order Correlation Coefficient (SRCC). Notably, the inter-task correlation is relatively low, particularly for VQA and the three tasks that rely on specialized CV models. Given the current state of image processing algorithms for machine vision, each employing different tasks as performance metrics, the universality and robustness of such evaluation are inherently constrained. The data distribution further substantiates this argument, with YoN/MCQ/CAP being more evenly distributed, VQA exhibiting a higher concentration of high scores, and SEG/DET/RET clustering at lower scores. Moreover, the strong correlation between the total score and the other tasks is evident, with all SRCC surpassing 0.5, underscoring the reliability of the MPD, for various downstream tasks. The MOS distribution across 30 types of corruption and 5 strength levels is depicted in Figure 4. Results show for corruption scenarios that appear similar to HVS, the MVS perception mechanism diverges significantly. For", + "bbox": [ + 511, + 599, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7574", + "bbox": [ + 482, + 948, + 514, + 959 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4f12ad9cbaa53718cbafceebbae4c914164a862f9f64ee4ad18ed46e2d673540.jpg", + "image_caption": [ + "Strength 1" + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 248, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5b11c5e7eec650aee15617ad440f1c9bb290e943aff8f9bd2863db85f73ad7cd.jpg", + "image_caption": [ + "Strength 2" + ], + "image_footnote": [], + "bbox": [ + 251, + 88, + 380, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/04c7b38e9fd0a7ae5802d6da455caae6608bf09d0e953e03999c63dea6459707.jpg", + "image_caption": [ + "Strength 3" + ], + "image_footnote": [], + "bbox": [ + 383, + 88, + 516, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d85c61f1fe6090e5b519a1b70d3a366d2f9eeb11e54ced17648f6de9039ba905.jpg", + "image_caption": [ + "Strength", + "Strength 5" + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ed53d883119b1d3dc2dfba449f9a515f21358656635a7e9ad312172a021c09c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 98, + 872, + 166 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6dbf5dbe7941a6a52246d6add2bd797c351bd85a158f7c01cf24da41b779df44.jpg", + "image_caption": [ + "Figure 5. Annotations from different LMMs subjects in MPD under 5 corruption strengths, including $\\mathcal{P}$ content types and 4 tasks.", + "Human (0.76)" + ], + "image_footnote": [], + "bbox": [ + 117, + 218, + 220, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a2077a8a6ae67ae80c8a8e5ea4c34aa8fc825150b8990d843f6dfe6c5c65827e.jpg", + "image_caption": [ + "MPD (0.62)" + ], + "image_footnote": [], + "bbox": [ + 225, + 218, + 330, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1226e1f518c84a851518fec60497b2036e020ed06340953e21b336885e151261.jpg", + "image_caption": [ + "YoN (0.50)", + "Figure 6. Correlation matrix of the proposed MPD in each dimension. A darker color denotes a higher SRCC, with mean SRCC attached below each matrix. The index of each matrix is arranged in the $\\uparrow$ , $\\rightarrow$ directions, in the same order as Section 3.3. The correlation between human subjects is greater than machines, and the correlation of the overall MOS is greater than the score in each dimension." + ], + "image_footnote": [], + "bbox": [ + 334, + 218, + 439, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f31c9e67faf073c4d6d83c985ad4d30df5f6383f47800d69ae3c4f71fcea8fd5.jpg", + "image_caption": [ + "MCQ (0.55)" + ], + "image_footnote": [], + "bbox": [ + 444, + 218, + 549, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b4bba999e196b1fd78bce3175088bc56aa8e86744b26a67cbff4bc268cdd3482.jpg", + "image_caption": [ + "VQA (0.37)" + ], + "image_footnote": [], + "bbox": [ + 553, + 218, + 658, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/065a8dd2b564dee173e7eb415ac9212b4d34a113c7b6e7b27a58916c7f032769.jpg", + "image_caption": [ + "CAP (0.39)" + ], + "image_footnote": [], + "bbox": [ + 663, + 218, + 767, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1cf7815dedc0891a3bed14374d3c1aa24ef4e9583ed20a6635c5937a87fb5276.jpg", + "image_caption": [ + "Other (0.53)" + ], + "image_footnote": [], + "bbox": [ + 774, + 218, + 875, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "instance, the machine is highly sensitive to 'Lens blur' but rarely affected by 'Mean brighten'. A typical example is the four Macro Block-related corruptions, which are nearly indistinguishable for HVS but exhibit significantly lower perception quality for MVS when encountering 'Block exchange' compared to the other three. Furthermore, the impact of distortion intensity on the MVS perceptual quality varies markedly. For example, as the intensity level increases, the 'Lab saturation' quality diminishes, while the 'HSV saturation' quality remains unchanged. These findings underscore the differences between MVS and HVS, reinforcing the significance of establishing such a unified quality representation in MPD.", + "bbox": [ + 88, + 388, + 483, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Each machine: We visualize the benchmark performance of six typical LMMs (as space limited) under different corruption strengths, as shown in Figure 5. For different content types, the scores of various machines on NSI/SCI/AIGI dimensions do not vary significantly. For different tasks, advanced models such as InternVL2 excel in dealing with distorted images in YoN, while traditional models like LLaVA1.5 can provide relatively stable results in CAP. As the corruption strength increases by one level, the scores of each dimension decrease by 0.01-0.02. Finally, for all machine subjects, we explored their internal differences in Figure 6. We manually conducted human annotation for 500 instances (the same 15 subjects) and compared the preference differences between humans and machines. The index of each matrix (i.e., the machine it represents) is arranged in the $\\uparrow$ and $\\rightarrow$ directions, in the same order as in Section 3.3. Overall, the SRCC between humans is 0.76, while that of machines is only 0.62, indicating that the divergence in machine image preferences is greater than that of humans. Further exploration of the preference scores", + "bbox": [ + 89, + 598, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for each dimension shows that the relatedness of machines in VQA and CAP tasks is the lowest. Additionally, there are a few subjects whose assessments differ significantly from other machines $^2$ , such as InstructBlip in YoN, Mantis in MCQ, Pangea in CAP, and DIME/NAAF in RET. This is often due to special token outputs or multilingual mechanisms, which lead to insufficient robustness, resulting in a higher proportion of errors in distorted images compared to other models, thus leading to inconsistent results. In summary, the larger preference differences make IQA for MVS a more difficult problem than traditional IQA, but also a more valuable and promising research topic.", + "bbox": [ + 511, + 388, + 906, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 511, + 585, + 637, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Setups", + "text_level": 1, + "bbox": [ + 511, + 612, + 694, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We randomly partitioned the MPD into the train/val set, with 24,000 and 6,000 reference and distorted image pairs, respectively, according to an 8:2 ratio. For MVS, Figure 4 illustrates significant variations in the perceptual quality of different corruptions. Given that machines exhibit a larger Just Noticeable Difference (JND) than humans, our corruption strength exceeds that of previous IQA databases for humans3. To conduct a realistic and reliable evaluation of machine preference, we designate the val set as 'Severe Distortion' and select a subset of 'Mild Distortion' to represent the corruption strength in the real world. The specific splitting", + "bbox": [ + 511, + 635, + 906, + 801 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2The ITU human subjective annotation standard [71] for humans will exclude data with weak correlation with other subjects, but this is to prevent eye fatigue or malicious attacks during annotation. Given that the machine's inference process is completely objective, we will refer to the outputs of each machine rather than excluding them.", + "bbox": [ + 511, + 814, + 903, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3See appendix for visualization in 5 corruption strength, where the Level 1 of our MPD can even ensemble Level 4 in KADID-10K[47].", + "bbox": [ + 511, + 875, + 903, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7575", + "bbox": [ + 482, + 948, + 514, + 959 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/183c79a084ea8884ac7c6793625a5bdb42be1f4607764605b736223c569c9d50.jpg", + "table_caption": [ + "Table 2. Using IQA metrics to predict machine preference, including baseline, FR, and NR algorithm, human visual system-based metrics are marked in $\\bullet$ . The first row represents two distortion intensities and three image content subsets, and the second row represents five scoring dimensions. [Keys: Best; Second best; Baseline (BL); Inferior to the baseline.]" + ], + "table_footnote": [], + "table_body": "
TypeMetricHVSSevere DistortionMild DistortionNSISCIAIGI
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
BLPSNR0.38730.26720.46640.30970.21270.61970.41160.28200.49450.31350.21460.33290.52250.36840.5985
SSIM[76]0.59690.41570.55370.26720.18280.28180.62880.44660.59320.54150.37480.53930.59800.42780.5687
FRAHIQ[33]0.80610.61530.81830.53040.36850.51110.84470.66010.85020.79620.60400.84700.79960.60820.7966
CKDN[95]✔0.74840.55600.71740.06040.04050.10610.81000.59880.76700.72030.51710.76610.69920.51020.6869
DISTS[15]0.73810.54430.73720.46160.32320.44860.79120.60600.80460.72750.53840.74290.72190.53650.7292
LPIPS[87]✔0.62530.44550.56010.00790.00920.00560.69650.51320.64780.56180.39200.56860.64290.45180.5481
PieAPP[62]✔0.47370.01630.23770.30300.21000.23490.45360.11870.27290.51730.09540.21350.46140.20090.3168
TOPIQ-FR[9]✔0.71780.52520.71820.42480.29440.48410.75060.56140.75850.72430.53220.71860.69430.50630.7061
NRARNIQA[3]0.83370.64260.81120.23880.16080.27540.87040.66900.84780.82200.62170.84200.84490.65780.7954
CLIPQA[74]✔0.64980.46620.59010.18220.12100.17970.76700.57710.71410.66900.47510.62150.68770.49340.5977
DBCNN[88]✔0.56520.39080.56960.13340.08940.15940.59780.41810.62190.56920.39740.60920.54220.38100.5613
HyperIQA[68]0.81050.61890.80120.53880.37740.47120.89080.70740.86910.82730.64030.86010.82120.62420.8076
NIMA[69]0.75620.56570.78580.45700.31710.43420.81530.62030.82240.78000.58810.80920.78460.59530.7622
TOPIQ-NR[9]✔0.65620.46160.63580.19610.13180.19890.70110.48230.66130.66450.46240.63530.54610.34220.5249
TypeMetricHVSYoNMCQVQACAPOthers
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
BLPSNR0.27530.18920.34670.32730.22390.43720.20670.13780.21200.43190.29930.49890.37980.26170.4683
SSIM[76]0.39600.27280.38040.50890.34990.48800.29600.19880.27560.55500.38610.54470.50720.34910.4766
FRAHIQ[33]0.60170.42830.62120.73140.53740.74660.42440.29240.45830.72200.53410.75460.61840.43960.6311
CKDN[95]✔0.54910.37990.54820.68660.48630.65960.40450.27630.39000.70440.51360.69910.54340.38220.5313
DISTS[15]0.52670.36950.53940.64860.46250.65490.35120.23950.36760.66490.48330.68120.62030.44190.6332
LPIPS[87]✔0.46770.32520.42450.55170.39120.51690.32160.21690.29350.56200.39960.54730.49270.33870.4242
PieAPP[62]✔0.33080.01730.16630.43100.19000.20870.23850.01320.10320.42200.01520.21360.37350.02090.1747
TOPIQ-FR[9]✔0.50110.34800.51180.60380.42320.61650.37000.25080.38990.63270.45450.64790.61990.43830.6281
NRARNIQA[3]0.62280.43940.61720.73420.53510.72420.44090.30260.44540.76560.56930.76640.65600.45660.6309
CLIPQA[74]✔0.46130.31670.43860.50900.35560.51170.29710.20280.32040.63790.45510.56730.57510.40480.4747
DBCNN[88]✔0.37980.25670.39690.47930.32580.49510.25560.17190.27830.52400.36330.53880.50850.34650.5093
HyperIQA[68]0.60100.42780.60660.70660.51550.71460.43220.29960.44190.71530.52340.74430.64610.45980.6329
NIMA[69]0.55440.39040.58870.68930.50020.70360.44600.30740.47750.63130.45310.68420.58100.41060.6181
TOPIQ-NR[9]✔0.49390.33580.47970.54690.37130.53700.35780.24370.37870.51120.33750.50890.57690.39340.5564
", + "bbox": [ + 93, + 133, + 906, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mechanism is detailed in the Appendix. To benchmark the performance of quality metrics, three global indicators were employed: SRCC, Kendall Rank-order Correlation Coefficient (KRCC), and Pearson Linear Correlation Coefficient (PLCC), to evaluate the consistency between the objective quality score and the subjective MOS. Among these, SRCC and KRCC represent the prediction monotonicity, while PLCC measures the accuracy.", + "bbox": [ + 89, + 537, + 482, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "12 mainstream IQA metrics are implemented for comparison, which have all shown satisfying results in previous human-centric IQA tasks, including 6 FR metrics: AHIQ [33], CKDN [95], DISTS [15], LPIPS [87], PieAPP [62], TOPIQ-FR [9]; and 6 NR metrics: ARNIQA [3], CLIPIQA [74], DBCNN [88], HyperIAQ [68], NIMA [69], TOPIQN [9]. Some methods specifically model the HVS, such as DBCNN, which has a dedicated stage for simulating human perception; others rely on the remarkable fitting ability of large-scale parameters, like the ViT in AHIQ. We have marked all HVS-based metrics as $\\bullet$ . Moreover, two naive quality indicators, PSNR and SSIM [76] are set as the baseline. We train the above 12 metrics on MPD with the learning rate as $10^{-5}$ for 50 epochs, under the default settings in pyiqa [9] toolbox and evaluate the performance on", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the two subsets above, three content types, and five scoring dimensions. We collect MPD preference annotations on $16 \\times$ NVIDIA A800 SXM4 80GB GPUs, and then conduct IQA training/validation on one GPU above.", + "bbox": [ + 511, + 539, + 906, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Experiment Result and Discussion", + "text_level": 1, + "bbox": [ + 511, + 606, + 815, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 presents the performance of current advanced IQA models on the MPD. Overall, regardless of FR or NR, no metric is highly consistent with machine preference, although they are all good at predicting human preference. An interesting finding is that the more accurately and completely a method models the HVS, the more difficult it is to evaluate the perceptual quality of MVS. Considering that most of the current IQA works emphasize human vision, while machines have already surpassed humans to become the main consumer of visual signals, future IQA needs to find a way to balance these two goals.", + "bbox": [ + 509, + 628, + 906, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the two distortion subsets, fitting-based methods such as AHIQ/ARNIQA/HyperIQA can achieve SRCC/PLCC values exceeding 0.8 on Severe Distortion. However, this is not a challenging task. Under the same distortion conditions, they have even exceeded 0.95 on the HVS dataset [33, 61]. The performance of Mild Distortion is even worse, and no method can even outperform the", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7576", + "bbox": [ + 482, + 948, + 514, + 960 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d456787f1c5c24f9c1ea6beb4560d6b0a1d3e10b9653be9744a9a23fc6c6bafc.jpg", + "table_caption": [ + "Table 3. Cross validation between human and machine preference database. For IQA metrics, training on machine preference data can reach a good result in evaluating human preference; but training on human preference is unacceptable for MPD. [Keys: Best; Lower than 0.4.]" + ], + "table_footnote": [], + "table_body": "
Val-setKADID-10KTID2013MPD (Mild)MPD (Severe)
MetricSRCC↑PLCC↑SRCC↑PLCC↑SRCC↑PLCC↑SRCC↑PLCC↑
AHIQ[33]0.79500.77320.76510.78840.15140.18460.47530.3518
CKDN[95]0.85990.72820.69890.60290.17130.20550.34940.3228
DISTS[15]0.82390.81900.72310.77590.41050.40770.46160.4486
LPIPS[87]0.61840.54510.70540.63260.23910.31180.46830.3711
PieAPP[62]0.47830.34680.68320.45430.12850.06780.17910.1235
TOPIQ-FR[9]0.67680.68440.77820.74730.51120.54950.62230.6685
ARNIQA[3]0.76840.70230.77010.72050.48290.45820.55970.4849
CLIPIQA[62]0.62860.57290.52370.39100.13670.14780.32010.3203
DBCNN[88]0.51750.53510.44860.57860.13590.15920.40240.3877
HyperIQA[68]0.66500.65470.70240.73130.08200.10740.45190.4521
NIMA[69]0.47950.49150.56940.64230.16290.17750.47090.4845
TOPIQ-NR[9]0.52320.50770.54550.61740.23880.27540.48290.4582
", + "bbox": [ + 91, + 160, + 483, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "simple PSNR, which proves that there is still a lot of room for optimization of IQA for machine preference. For the three contents, after verifying them separately, the performance has slightly increased, which is caused by the internal quality differences of the contents. In general, predicting NSI quality is relatively simple, while SCI and AIGI are more difficult. This may be due to the fact that the initial weights of the IQA model have rich NSI priors. For the five dimensions, we can see that the results of MCQ and CAP are most consistent with machine preference; YoN and Others (three common CV tasks) are second; VQA is the worst. This is because VQA problems are usually changes in image details (such as the color of a button on a piece of clothing), which are more difficult for machines to detect than other tasks. In short, IQA for Machine preference still has great research potential, and it is necessary to develop more accurate and effective evaluation mechanisms in the future.", + "bbox": [ + 89, + 340, + 483, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Cross Database Validation", + "text_level": 1, + "bbox": [ + 89, + 606, + 331, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further analyze the differences between MVS and HVS, we conducted cross-validation using MPD and two of the most commonly used human-oriented datasets, TID2013 [61] and KADID-10K [47], employing the same training mechanism as in the previous chapter. We observed that after training on MPD, most IQA metrics still possess a certain ability to predict human preferences, with SRCC values around 0.7 to 0.8; conversely, after training on the other two human-centric databases, IQA metrics struggle to predict machine preferences, with results on MPD not even reaching 0.4. Machine preference data will not affect human tasks, but the opposite will. This finding holds certain guiding significance for future IQA, particularly for a human-machine-friendly paradigm.", + "bbox": [ + 89, + 628, + 485, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Long-term Availability Validation", + "text_level": 1, + "bbox": [ + 89, + 848, + 387, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The machine subjects in MPD are categorized into two groups: common CV models and LMMs. The former mod", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/16331855e00e8ac7e144badfc6023cbe011b025c5ff62f43c9c61ac100922c21.jpg", + "image_caption": [ + "Figure 7. The long-term usability of MPD. The preference scores of MPD are consistent with four early-stage LMMs and three advanced LMMs (marked as Traditional/Advanced Models in the background). Their high SRCC proves that MPD can still characterize the perceptual quality of LMMs in the near future." + ], + "image_footnote": [], + "bbox": [ + 519, + 80, + 929, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "els, such as the separation/detection benchmark method, have been largely finalized and generally recognized. In contrast, LMMs are still under rapid evolvement. It is conceivable that MPD may not be able to capture the LMMs that will emerge in the future, or with the update of MPD, the preference of previous LMMs may be sacrificed. Thus, we include 3 advanced and 4 early-stage LMMs as Sec 3.3 listed, integrating them into two databases similar to MPD. Given the inconsistencies in LMM preference on VQA and CAP tasks, as illustrated in Figure 6, we have verified the consistency of MPD ground truth labels with advanced/early-stage labels. For each image, the relationship between these two labels is depicted in Figure 7. The findings indicate that MPD exhibits a high correlation with early models on VQA and with advanced models on CAP. Generally, the SRCC is around 0.6, which is considered acceptable. Considering the GPT-4o surpasses open-source models, we anticipate that MPD is able to characterize LMMs preferences for images for at least a year. In the future, We will continue to update MPD by periodicity introducing the latest LMMs to avoid this limitation.", + "bbox": [ + 511, + 330, + 908, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper extends the application of IQA from human to machine vision for the first time and establishes the first large-scale, fine-grained, multi-dimensional quality database for machines, namely MPD. We select 15 LMMs and 15 common CV models as subjects, show them $30\\mathrm{k}$ reference/distorted image pairs, and summarize their performance on 7 downstream tasks as quality scores. In-depth analysis of the MPD content and IQA benchmark experiments have proved the huge gap between machine and human visual systems, and current human-centric IQA methods cannot accurately predict machine preferences. As the visual data consumption volume of machines has already surpassed humans, we hope that MPD can inspire effective IQA metrics to serve the upcoming machine-centric era.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7577", + "bbox": [ + 482, + 948, + 514, + 960 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work was supported by the National Natural Science Foundation of China under Grant 62225112, 62301310, 623B2073, 62471290; and in part by Sichuan Science and Technology Program under Grant 2024NSFSC1426.", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 188, + 187, + 204 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 5", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5", + "[3] Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. Arniqa: Learning distortion manifold for image quality assessment. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 188-197, 2024. 7, 8", + "[4] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 382-398. Springer, 2016. 4", + "[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 5", + "[6] Johannes Balle, Philip A Chou, David Minnen, Saurabh Singh, Nick Johnston, Eirikur Agustsson, Sung Jin Hwang, and George Toderici. Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing, 15(2):339-353, 2020. 2, 3", + "[7] Min Cao, Shiping Li, Juntao Li, Liqiang Nie, and Min Zhang. Image-text retrieval: A survey on recent research and development. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 5410-5417. International Joint Conferences on Artificial Intelligence Organization, 2022. Survey Track. 4", + "[8] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 3", + "[9] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 33:2404-2418, 2024. 7, 8", + "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing" + ], + "bbox": [ + 93, + 213, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5", + "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24185–24198, 2024. 5", + "[12] Jinyoung Choi and Bohyung Han. Task-aware quantization network forJPEG image compression. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 309-324. Springer, 2020. 3", + "[13] U Cisco. Cisco annual internet report (2018-2023) white paper. Cisco: San Jose, CA, USA, 10(1):1-35, 2020. 1", + "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 5", + "[15] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 7, 8", + "[16] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Ji-aqi Wang. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 5", + "[17] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, 2022. 5", + "[18] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, Dahua Lin, and Kai Chen. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models, 2024. 4", + "[19] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 5", + "[20] Wen Gao, Siwei Ma, Lingyu Duan, Yonghong Tian, Peiyin Xing, Yaowei Wang, Shanshe Wang, Huizhu Jia, and Tiejun Huang. Digital retina: A way to make the city brain more efficient by visual coding. IEEE Transactions on Circuits and Systems for Video Technology, 31(11):4147-4161, 2021. 1", + "[21] Yixuan Gao, Yuqin Cao, Tengchuan Kou, Wei Sun, Yunlong Dong, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Vdpve: Vqa dataset for perceptual video enhancement. In Proceedings of the IEEE/CVF Conference on Computer" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7578", + "bbox": [ + 482, + 948, + 514, + 959 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition (CVPR) Workshops, pages 1474-1483, 2023. 2", + "[22] Zhangwei Gao, Zhe Chen, Erfei Cui, Yiming Ren, Weiyun Wang, Jinguo Zhu, Hao Tian, Shenglong Ye, Junjun He, Xizhou Zhu, Lewei Lu, Tong Lu, Yu Qiao, Jifeng Dai, and Wenhai Wang. Mini-internvl: A flexible-transfer pocket multimodal model with 55", + "[23] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 5", + "[24] Jinjin Gu, Haoming Cai, Chao Dong, Jimmy S Ren, Radu Timofte, Yuan Gong, Shanshan Lao, Shuwei Shi, Jiahao Wang, Sidi Yang, et al. Ntire 2022 challenge on perceptual image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 951-967, 2022. 2", + "[25] Yatharth Gupta, Vishnu V. Jaddipal, Harish Prabhala, Sayak Paul, and Patrick Von Platen. Progressive knowledge distillation of stable diffusion x1 using layer level loss, 2024. 3", + "[26] Shijie Hao, Yuan Zhou, and Yanrong Guo. A brief survey on semantic segmentation with deep learning. Neurocomputing, 406:302-321, 2020. 4", + "[27] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 5", + "[28] David Holz. Midjourney. https://www.midjourney.com, 2023. 3", + "[29] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 2", + "[30] Yunzhong Hou, Liang Zheng, and Stephen Gould. Learning to structure an image with few colors. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 10116-10125, 2020. 3", + "[31] Zhimeng Huang, Chuanmin Jia, Shanshe Wang, and Siwei Ma. Visual analysis motivated rate-distortion model for image coding. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 3", + "[32] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning, 2024. 5", + "[33] Shanshan Lao, Yuan Gong, Shuwei Shi, Sidi Yang, Tianhe Wu, Jiahao Wang, Weihao Xia, and Yujiu Yang. Attentionshelp cnns see better: Attention-based hybrid image quality assessment network, 2022. 7, 8", + "[34] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 3, 5", + "[35] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiqa-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2", + "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5", + "[37] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024.5", + "[38] Chunyi Li, Tengchuan Kou, Yixuan Gao, Yuqin Cao, Wei Sun, Zicheng Zhang, Yingjie Zhou, Zhichao Zhang, Weixia Zhang, Haoning Wu, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Aigia-20k: A large database for a-generated image quality assessment, 2024. 2", + "[39] Chunyi Li, Guo Lu, Donghui Feng, Haoning Wu, Zicheng Zhang, Xiaohong Liu, Guangtao Zhai, Weisi Lin, and Wenjun Zhang. Misc: Ultra-low bitrate image semantic compression driven by large multimodal model, 2024. 1", + "[40] Chunyi Li, Haoning Wu, Hongkun Hao, Zicheng Zhang, Tengchuan Kou, Chaofeng Chen, Lei Bai, Xiaohong Liu, Weisi Lin, and Guangtao Zhai. G-refine: A general quality refiner for text-to-image generation. In Proceedings of the 32nd ACM International Conference on Multimedia, page 7375-7384, New York, NY, USA, 2024. Association for Computing Machinery.", + "[41] Chunyi Li, Haoning Wu, Zicheng Zhang, Hongkun Hao, Kaiwei Zhang, Lei Bai, Xiaohong Liu, Xiongkuo Min, Weisi Lin, and Guangtao Zhai. Q-refine: A perceptual quality refiner for ai-generated image. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2024. 1", + "[42] Chunyi Li, Xiele Wu, Haoning Wu, Donghui Feng, Zicheng Zhang, Guo Lu, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Cmc-bench: Towards a new paradigm of visual signal compression, 2024. 3", + "[43] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024. 3", + "[44] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 5", + "[45] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 5", + "[46] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3", + "[47] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Tenth International Conference on Quality of Multimedia Experience (QoMEX), pages 1-3. IEEE, 2019. 2, 6, 8" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7579", + "bbox": [ + 482, + 948, + 514, + 959 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Risheng Liu, Long Ma, Tengyu Ma, Xin Fan, and Zhongxuan Luo. Learning with nested scene modeling and cooperative architecture search for low-light vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5953-5969, 2022. 3", + "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 5", + "[50] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 4", + "[51] Xiaohong Liu, Xiongkuo Min, Guangtao Zhai, Chunyi Li, Tengchuan Kou, et al. Ntire 2024 quality assessment of ai-generated content challenge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 6337-6362, 2024. 2", + "[52] Yuan Liu, Le Tian, Xiao Zhou, and Jie Zhou. Rethinking overlooked aspects in vision-language models. arXiv preprint arXiv:22405.11850, 2024. 5", + "[53] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Yaofeng Sun, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 5", + "[54] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model, 2024. 5", + "[55] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 5", + "[56] Xiongkuo Min, Kede Ma, Ke Gu, Guangtao Zhai, Zhou Wang, and Weisi Lin. Unified blind quality assessment of compressed natural, graphic, and screen content images. IEEE Transactions on Image Processing, 26(11):5462-5474, 2017. 2, 3", + "[57] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE transactions on Image Processing, 20 (12):3350-3364, 2011. 2", + "[58] Zhangkai Ni, Lin Ma, Huanqiang Zeng, Jing Chen, Canhui Cai, and Kai-Kuang Ma. Esim: Edge similarity for screen content image quality assessment. IEEE Transactions on Image Processing, 26(10):4818-4831, 2017. 2", + "[59] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 4", + "[60] Neel Patwa, Nilesh Ahuja, Srinivasa Somayazulu, Omesh Tickoo, Srenivas Varadarajan, and Shashidhar Koolagudi. Semantic-preserving image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 1281-1285. IEEE, 2020. 3" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, et al. Image database tid2013: Peculiarities, results and perspectives. Signal processing: Image communication, 30:57-77, 2015. 2, 7, 8", + "[62] Ekta Prashnani, Hong Cai, Yasamin Mostofi, and Pradeep Sen. Pieapp: Perceptual image-error assessment through pairwise preference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1808-1817, 2018. 7, 8", + "[63] Leigang Qu, Meng Liu, Jianlong Wu, Zan Gao, and Liqiang Nie. Dynamic modality interaction modeling for image-text retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1104-1113, 2021. 5", + "[64] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 3", + "[65] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 3, 5", + "[66] Robin Rombach, Andreas Blattmann, and Björn Ommer. Text-guided synthesis of artistic images with retrieval-augmented diffusion models, 2022. 3", + "[67] Chengyao Shen and Qi Zhao. Webpage saliency. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13, pages 33-46. Springer, 2014. 3", + "[68] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2020. 7, 8", + "[69] Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment. IEEE Transactions on Image Processing, 27(8): 3998-4011, 2018. 7, 8", + "[70] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 5", + "[71] I. T. Union. Methodology for the subjective assessment of the quality of television pictures. ITU-R Recommendation BT. 500-11, 2002. 5, 6", + "[72] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 4", + "[73] Thang Vu, Haeyong Kang, and Chang D Yoo. Scnet: Training inference sample consistency for instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2701-2709, 2021. 5", + "[74] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Pro" + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7580", + "bbox": [ + 482, + 948, + 514, + 959 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 7", + "[75] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5", + "[76] Zhou Wang. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7", + "[77] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 5", + "[78] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-bench: A benchmark for general-purpose foundation models on low-level vision. arXiv preprint arXiv:2309.14181, 2023. 4", + "[79] XAI. grok-1.5v, 2024. 4", + "[80] Zhaohui Yang, Yunhe Wang, Chang Xu, Peng Du, Chao Xu, Chunjing Xu, and Qi Tian. Discernible image compression. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1561-1569, 2020. 3", + "[81] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024.5", + "[82] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13040–13051, 2024. 5", + "[83] Masakazu Yoshimura, Junji Otsuka, Atsushi Irie, and Takeshi Ohashi. Dynamicisp: dynamically controlled image signal processor for image recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12866-12876, 2023. 3", + "[84] Ke Yu, Zexian Li, Yue Peng, Chen Change Loy, and Jinwei Gu. Reconfigisp: Reconfigurable camera image processing pipeline. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4248-4257, 2021. 3", + "[85] Xiang Yue, Yueqi Song, Akari Asai, Seungone Kim, Jean de Dieu Nyandwi, Simran Khanuja, Anjali Kantharuban, Lintang Sutawika, Sathyanarayanan Ramamoorthy, and Graham Neubig. Pangea: A fully open multilingual multimodal llm for 39 languages, 2024. 5", + "[86] Jing Zhang and Dacheng Tao. Empowering things with intelligence: a survey of the progress, challenges, and opportunities in artificial intelligence of things. IEEE Internet of Things Journal, 8(10):7789-7817, 2020. 1", + "[87] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7, 8", + "[88] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2018. 7, 8", + "[89] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 5", + "[90] Kun Zhang, Zhendong Mao, Quan Wang, and Yongdong Zhang. Negative-aware attention framework for image-text matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15661-15670, 2022. 5", + "[91] Zicheng Zhang, Chunyi Li, Wei Sun, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. A perceptual quality assessment exploration for aigc images. In IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 440-445, 2023. 2", + "[92] Zicheng Zhang, Wei Sun, Yingjie Zhou, Jun Jia, Zhichao Zhang, Jing Liu, Xiongkuo Min, and Guangtao Zhai. Subjective and objective quality assessment for in-the-wild computer graphics images. ACM Transactions on Multimedia Computing, Communications and Applications, 20(4):1-22, 2023. 3", + "[93] Qi Zhang, Shanshe Wang, Xinfeng Zhang, Chuanmin Jia, Zhao Wang, Siwei Ma, and Wen Gao. Perceptual video coding for machines via satisfied machine ratio modeling, 2024. 2", + "[94] Qi Zhang, Shanshe Wang, Xinfeng Zhang, Siwei Ma, Jingshan Pan, and Wen Gao. Predicting satisfied user and machine ratio for compressed images: A unified approach, 2024. 2", + "[95] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10242-10251, 2021. 7, 8", + "[96] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 4" + ], + "bbox": [ + 516, + 92, + 903, + 712 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "7581", + "bbox": [ + 482, + 948, + 513, + 959 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Image Quality Assessment_ From Human to Machine Preference/9c43c7f6-7667-4c47-8440-32e15bc27376_model.json b/2025/Image Quality Assessment_ From Human to Machine Preference/9c43c7f6-7667-4c47-8440-32e15bc27376_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f5aa6d1fb263d7b1e0db9e695082aa0fe8435dbe --- /dev/null +++ b/2025/Image Quality Assessment_ From Human to Machine Preference/9c43c7f6-7667-4c47-8440-32e15bc27376_model.json @@ -0,0 +1,3194 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.172, + 0.131, + 0.828, + 0.154 + ], + "angle": 0, + "content": "Image Quality Assessment: From Human to Machine Preference" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.18, + 0.917, + 0.236 + ], + "angle": 0, + "content": "Chunyi Li\\(^{1,2,4*}\\), Yuan Tian\\(^{1,2*}\\), Xiaoyue Ling\\(^{1*}\\), Zicheng Zhang\\(^{1,2}\\), Haodong Duan\\(^{2}\\), Haoning Wu\\(^{3,4}\\), Ziheng Jia\\(^{1}\\), Xiaohong Liu\\(^{1}\\), Xiongkuo Min\\(^{1}\\), Guo Lu\\(^{1\\dagger}\\), Weisi Lin\\(^{4}\\), Guangtao Zhai\\(^{1,2\\dagger}\\) \nShanghai Jiao Tong University\\(^{1}\\), Shanghai AI Lab\\(^{2}\\), Moonshot AI\\(^{3}\\), Nanyang Technological University\\(^{4}\\)" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.244, + 0.905, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.463, + 0.907, + 0.506 + ], + "angle": 0, + "content": "Figure 1. The significant gap between the well-explored Human Vision System (HVS) and the emerging Machine Vision System (MVS). Their different perception mechanisms leading to images that are subjectively satisfactory to humans not being applicable to the machines downstream tasks such as detection and question answering, vice versa." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.518, + 0.327, + 0.535 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.551, + 0.486, + 0.87 + ], + "angle": 0, + "content": "Image Quality Assessment (IQA) based on human subjective preferences has undergone extensive research in the past decades. However, with the development of communication protocols, the visual data consumption volume of machines has gradually surpassed that of humans. For machines, the preference depends on downstream tasks such as segmentation and detection, rather than visual appeal. Considering the huge gap between human and machine visual systems, this paper proposes the topic: Image Quality Assessment for Machine Vision for the first time. Specifically, we (1) defined the subjective preferences of machines, including downstream tasks, test models, and evaluation metrics; (2) established the Machine Preference Database (MPD), which contains 2.25M fine-grained annotations and 30k reference/distorted image pair instances; (3) verified the performance of mainstream IQA algorithms on MPD. Experiments show that current IQA metrics are human-centric and cannot accurately characterize machine preferences. We sincerely hope that MPD can promote the evolution of IQA from human to machine preferences. Project page is on: https://github.com/lcyszxdxc/MPD." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.518, + 0.645, + 0.533 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.907, + 0.821 + ], + "angle": 0, + "content": "The rapid ascend of smart cities [20] and the flourishment of the Internet of Things (IoT) [86] over the past decade have fundamentally transformed the framework of the application end. According to the Cisco [13] white paper, the number of Machine-to-Machine (M2M) connections first exceeded that of Machine-to-Human (M2H) in 2023, reaching 147 billion. Machines have gradually replaced humans and become the primary consumers of image/video data. In the field of image processing, the primary goal is to enhance the quality of processed images, ensuring they conform to human perception. Previously, the performance of a compression/reconstruction/generation algorithm [39-41] is often judged by its visual appeal to humans. This technology, known as Image Quality Assessment (IQA), comprehensively models the Human Visual System (HVS) to predict human subjective preferences. Consequently, with the evolution of application ends, IQA today has to transcend HVS and serve Machine Vision Systems (MVS) more effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "However, there are significant disparities in the perceptual mechanisms between HVS and MVS. This discrepancy arises because HVS prioritizes the similarity between reference and distorted images, such as texture, structure, and color, whereas machines focus on the consistency of results" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.896, + 0.23, + 0.91 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.246, + 0.896, + 0.38, + 0.91 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.949, + 0.516, + 0.961 + ], + "angle": 0, + "content": "7570" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison of MPD with other IQA databases. As the first machine preference-oriented library, MPD has a larger image scale, a wider range of corruption scenarios, and more complete annotation content." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.123, + 0.887, + 0.294 + ], + "angle": 0, + "content": "
DatabaseImageCorruptionAnnotation
ReferenceDistortedResolutionTypesStrengthNumDimensionLabeling Criteria
LIVE[57]297797685525k1Human preference on NSIs
TID2013[61]253k512245514k1Human preference on NSIs
KADID-10K[47]8110k1k255304k1Human preference on NSIs
CLIC2021[6]5853k1k103484k1Human preference on compressed images
NTIRE2022[24]25029k2884051.13m1Human preference on super-resolution images
KonIQ-10K[29]-10k1k--1.21m5Human preference on NSIs
AGIQA-3K[35]-3k1k--125k2Human preference on AIGIs
NTIRE2024[38]-20k1k--420k2Human preference on AIGIs
MPD (ours)1k30k1k3052.25m5Machine preference on NSIs, SCIs, and AIGIs
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.319, + 0.482, + 0.44 + ], + "angle": 0, + "content": "in downstream tasks. As illustrated in Figure 1, for distortions resulting from factors like encoding/quantization, HVSs perceive a noticeable degradation in quality, yet the performance of tasks like segmentation/detection by machines remains unaffected. Conversely, some minor disturbances that the human eye barely notices can significantly disrupt the output of MVSs, making it difficult to migrate their evaluation algorithms." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.442, + 0.483, + 0.713 + ], + "angle": 0, + "content": "What images do machines prefer? This remains an open question. For HVS, over the past two decades, IQA has witnessed significant development. Given the variations in image preferences among individuals, particularly across dimensions such as brightness and chroma, the scoring dimensions, the annotations from multiple subjects, and the integration of these annotations into global subjective preference labels have been well established. Hundreds of fine-grained databases can drive IQA algorithms that closely align with human perception. Conversely, MVS development is much slower, regardless of an end-to-end IQA algorithm, there is even no database that annotates images with an overall preference score of machines. This hinders the comprehensive evaluation of image processing metrics designed for machine vision, limiting the experience of application ends. Considering these issues, we first attempt to generalize the concept of IQA from HVS to MVS. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.72, + 0.483, + 0.87 + ], + "angle": 0, + "content": "- A innovative definition of the subjective preferences of machines. We propose a standardized process for collecting subjective labels from machines. Specifically, we identify the downstream tasks that determine the quality of machine perception, including four multimodal tasks related to Large Multimodal Models (LMMs) and three classic Computer Vision (CV) tasks. By referencing the multi-dimensional scoring mechanism of humans, we comprehensively evaluate the performance of machines in these tasks to obtain ground truth labels." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.482, + 0.902 + ], + "angle": 0, + "content": "- A large-scale database named the Machine Preference Database (MPD). Using 15 general LMMs and 15 spe" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.319, + 0.907, + 0.409 + ], + "angle": 0, + "content": "cialized CV models, we collected 2.25 million preference data points on 30,000 reference/distorted image pairs, establishing the first IQA database for machine vision. In particular, we conducted an in-depth analysis of each participating machine to explore the internal patterns of MVS perception and their differences with HVS." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.41, + 0.907, + 0.516 + ], + "angle": 0, + "content": "- A complex benchmark test for existing IQA algorithms. Experiments on the MPD demonstrate that these algorithms designed for HVS cannot fully model MVS, whose predictions do not match machine ratings. After finetuning based on the MPD, their performance improves slightly, but still cannot characterize machine preferences, especially when facing mild distortion." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.531, + 0.665, + 0.547 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.557, + 0.829, + 0.573 + ], + "angle": 0, + "content": "2.1. Image Quality Assessment Database" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.579, + 0.907, + 0.777 + ], + "angle": 0, + "content": "IQA datasets include Full Reference (FR) and No Reference (NR). FR datasets manually add corruption to reference images, whereas NR directly capture distorted images. Typically, these datasets focus on a single image type, such as Natural Scenes Images (NSIs) [6, 24, 47, 57, 61], Screen Content Images (SCIs) [21, 56, 58], and AI-Generated Images (AIGIs) [35, 38, 51, 91], with quality labels derived solely from human preferences as listed in Table 1. On contrast, in the era where machines surpass humans as application ends, MPD stands out as the first machine preference-oriented library. It offers a larger image scale, a broader range of corruption scenarios, and a more comprehensive annotation context than other IQA databases." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.813, + 0.803 + ], + "angle": 0, + "content": "2.2. Machine Vision Downstream Task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In the absence of a unified quality indicator, image processing algorithms, taking compression and restoration metrics for machine vision for example, are often in a fragmented state, with the performance of each algorithm being validated by only one or two downstream tasks. [93, 94] These tasks vary, as do the models used to perform them, mak" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.949, + 0.514, + 0.961 + ], + "angle": 0, + "content": "7571" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.907, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.478, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Figure 2. Overview of Machine Preference Database (MPD), with 30k large-scale reference/distorted image pairs, meticulously annotated with 2.25M fine-grained preference score from the mean opinion of 15 mainstream LMMs and 15 specific CV models. It provides a multi-dimensional fidelity evaluation of various downstream machine vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.537, + 0.486, + 0.734 + ], + "angle": 0, + "content": "ing the evaluation criteria disparate and incomplete. For compression algorithms, validation indicators include the classification [30, 60], image captioning [12], object detection [80], or semantic segmentation [31] result before/after compression, with each indicator differing and lacking standardization or completeness. Restoration algorithms fare slightly better, with most consistently relying on object detection [48, 83, 84]. Nevertheless, some employ traditional Fast-RCNN [65], while others utilize more advanced models like ViTDet[34]. The variation in model bases further compromises the comprehensiveness and objectivity of the evaluation outcomes. Thus, there is an urgent need for a unified representation of machine preferences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.748, + 0.31, + 0.763 + ], + "angle": 0, + "content": "3. Database Construction" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.435, + 0.789 + ], + "angle": 0, + "content": "3.1. Reference & Distorted Image Collection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Similar to the previous IQA database, we first selected 1,000 high-quality reference images and then artificially degraded their quality to obtain 30,000 pairs of reference and distorted images. To comprehensively represent the common content on the internet, we referred to the CMC-Bench [42] and selected the three most common image categories: Natural Scene Images (NSIs), including" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.657 + ], + "angle": 0, + "content": "User/Professional Generated Content (UGC/PGC) [6, 46]; Screen Content Images (SCIs), including webpages [67], games [92], and movies [56]; and AI-Generated Images (AIGIs), generated by six representative text-to-image models [8, 25, 28, 43, 64, 66]. The collection process is shown in Figure 2. Both human viewers and IQA models have verified this image sequence to ensure that they are undistorted as reference images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.858 + ], + "angle": 0, + "content": "For the distorted images, we modeled the corruption caused by the current communication protocols, which were divided into 30 scenarios. Specifically, these corruptions were classified into seven categories: Blur, various types of unclear image; Luminance, global brightness changes; Chrominance, global color changes; Contrast, global changes in contrast/saturation; Noise, random noise of different distributions; Compression, codec algorithm like JPEG; and Spatial, local pixel-level changes. For each corruption, we defined five levels of intensity, from mild to severe. To fairly explore the impact of each corruption, we artificially control the distortion strength to ensure that quality degradation perceived by the human eye is" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.907, + 0.902 + ], + "angle": 0, + "content": "1For the definition of all these 30 corruptions, specific details, and examples, please refer to the appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.949, + 0.516, + 0.961 + ], + "angle": 0, + "content": "7572" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "aligned at the same level. Thus, for each reference image, we randomly selected the intensity to add all the corruptions mentioned above, resulting in 30,000 distorted images. The reference/distorted image pairs will then be annotated by machine subjects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.175, + 0.331, + 0.189 + ], + "angle": 0, + "content": "3.2. Different Task Annotation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.197, + 0.484, + 0.452 + ], + "angle": 0, + "content": "In the past five years, the evaluation of image processing algorithms for machine vision has primarily focused on seven tasks: Regression, Classification, Visual Question Answering (VQA), Caption (CAP), Segmentation (SEG), Detection (DET), and Retrieval (RET). Among these, the first four tasks can be accomplished by general LMMs, whereas the latter three tasks necessitate specialized models due to their unique requirements. Note that the regression and classification tasks in this context are analogous to Yes-or-No (YoN) and Multiple-Choice Questions (MCQ) within LMMs. To this end, we enlisted five experienced human experts to design YoN/VQA/CAP questions and answers for 1,000 reference images. These questions are designed to be challenging, as illustrated in the example provided in Figure 2, especially in the MCQ sets where confusion options are introduced. This approach enables us to accurately measure the degree of quality degradation perceived by the machine." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.484, + 0.62 + ], + "angle": 0, + "content": "We apply these seven tasks to both the reference and distorted images, measure the similarity of the output results, and derive scores for each of the seven sub-dimensions. Specifically, for \\( Y_{oN} \\) tasks, we employ a logit-level evaluation approach based on softmax similarity, as previous LMM benchmarks [50, 78] do. This entails LMMs confirming or denying the questions designed for each reference/distorted image pair. The confidence difference in the next token output, 'Yes' or 'No', reflects the difference in the perceptual quality degradation from the perspective of machines, namely a score \\( S_{Y_{oN}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.627, + 0.482, + 0.645 + ], + "angle": 0, + "content": "\\[\nS _ {Y o N} = \\left| \\sigma \\left(P _ {\\text {d i s}} (\\text {Y e s}, \\text {N o})\\right) - \\sigma \\left(P _ {\\text {r e f}} (\\text {Y e s}, \\text {N o})\\right) \\right|, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.773 + ], + "angle": 0, + "content": "where \\( P_{dis,ref}(\\cdot) \\) denotes the output probability of certain logits, normalized by the softmax function \\( \\sigma(\\cdot) \\). MCQ tasks follow a similar evaluation paradigm. We set a correct option and three confusing options (the options are usually no more than 4, see [18, 50, 79]), and the model outputs the probability of each option as a four-element vector. The distance between the two vectors in space can reflect the perceptual quality score \\( S_{MCQ} \\) for machines:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.779, + 0.483, + 0.796 + ], + "angle": 0, + "content": "\\[\nS _ {M C Q} = \\cos \\left(P _ {d i s} (\\mathrm {A}, \\mathrm {B}, \\mathrm {C}, \\mathrm {D}), P _ {r e f} (\\mathrm {A}, \\mathrm {B}, \\mathrm {C}, \\mathrm {D})\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.803, + 0.484, + 0.879 + ], + "angle": 0, + "content": "where the distance is measured by cosine similarity. For \\(VQA\\) tasks, we conduct sentence-level evaluation beyond the logits. LMMs will be required to answer a certain question in about 5 words, and the semantic difference between two output sentences will be the score \\(S_{VQA}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.886, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nS _ {V Q A} = \\operatorname {C L I P} \\left(T _ {\\text {d i s}}, T _ {\\text {r e f}}\\right), \\tag {3}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.092, + 0.745, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.242, + 0.708, + 0.256 + ], + "angle": 0, + "content": "(a) Correlation matrix" + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.098, + 0.892, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.241, + 0.893, + 0.255 + ], + "angle": 0, + "content": "(b) Score distribution" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.275, + 0.906, + 0.302 + ], + "angle": 0, + "content": "Figure 3. Correlation between the general preference score and the score in seven different downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.906, + 0.436 + ], + "angle": 0, + "content": "where \\( T_{dis,ref} \\) represents the whole output text, measured by the CLIP text encoder. \\( CAP \\) tasks require a longer output, by describing the main subjects, actions, settings, and any notable objects or features of the image in about 40 words. Considering the remarkable text length, we use indicators dedicated to image captioning to analyze the differences between the two long sentences in \\( S_{CAP} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.447, + 0.906, + 0.48 + ], + "angle": 0, + "content": "\\[\nS _ {C A P} = \\sum_ {\\text {e v a l}} \\operatorname {e v a l} \\left(T _ {d i s}, T _ {r e f}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.49, + 0.906, + 0.552 + ], + "angle": 0, + "content": "where eval \\(\\in\\) (BLEU, CIDEr, SPICE) [4, 59, 72], three most commonly-used captioning indicators. Then, for SEG tasks, referring to their common evaluation criteria [26], we calculate the Intersection-over-Union (IoU) ratio in \\(S_{SEG}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.564, + 0.905, + 0.581 + ], + "angle": 0, + "content": "\\[\nS _ {S E G} = \\operatorname {I o U} \\left(M _ {d i s}, M _ {r e f}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.906, + 0.683 + ], + "angle": 0, + "content": "where \\( M_{dis,ref} \\) stands for binarized segmentation masks. DET tasks will be more complex, whose results include both spatial information and category labels. Thus, we consider the detection failed when the categories are different, and only further calculate the IoU when they are consistent [96], as described in \\( S_{DET} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.696, + 0.905, + 0.713 + ], + "angle": 0, + "content": "\\[\nS _ {D E T} = \\operatorname {A c c} _ {1} \\left(C _ {d i s}, C _ {r e f}\\right) \\cdot \\operatorname {I o U} \\left(B _ {d i s}, B _ {r e f}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.724, + 0.906, + 0.8 + ], + "angle": 0, + "content": "where \\( C_{dis,ref} \\) denotes the category sequence ordered by confidence, and \\( B_{dis,ref} \\) is the detected bounding box. Finally for RET tasks, we merge three common Image-to-Text retrieval indicators [7], namely Top-1/5/10 accuracy, to evaluate the retrieved sequence as \\( S_{RET} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.812, + 0.905, + 0.845 + ], + "angle": 0, + "content": "\\[\nS _ {R E T} = \\sum_ {\\mathrm {i} \\in 1, 5, 1 0} \\operatorname {A c c} _ {\\mathrm {i}} \\left(C _ {d i s}, C _ {r e f}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "After the above process, we can obtain the preferences of LMMs in the first four dimensions, and the preferences of each specific CV model on its target task." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.949, + 0.516, + 0.961 + ], + "angle": 0, + "content": "7573" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.23, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.137, + 0.207, + 0.145 + ], + "angle": 0, + "content": "Gaussian...filter" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.091, + 0.365, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.137, + 0.332, + 0.145 + ], + "angle": 0, + "content": "Lens...blur" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.091, + 0.499, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.137, + 0.471, + 0.145 + ], + "angle": 0, + "content": "Motion...blur" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.091, + 0.633, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.137, + 0.612, + 0.145 + ], + "angle": 0, + "content": "Color...diffusion" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.091, + 0.767, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.137, + 0.736, + 0.145 + ], + "angle": 0, + "content": "Color...shift" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.091, + 0.901, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.137, + 0.888, + 0.146 + ], + "angle": 0, + "content": "Color...quantization" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.147, + 0.23, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.193, + 0.209, + 0.202 + ], + "angle": 0, + "content": "HSV...saturation" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.147, + 0.364, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.193, + 0.341, + 0.201 + ], + "angle": 0, + "content": "Lab...saturation" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.147, + 0.499, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.193, + 0.484, + 0.202 + ], + "angle": 0, + "content": "JP2K...compression" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.147, + 0.633, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.193, + 0.619, + 0.202 + ], + "angle": 0, + "content": "JPEG...compression" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.147, + 0.767, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.193, + 0.756, + 0.202 + ], + "angle": 0, + "content": "WEBP...compression" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.147, + 0.901, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.193, + 0.874, + 0.201 + ], + "angle": 0, + "content": "White...noise" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.203, + 0.23, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.203, + 0.364, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.203, + 0.499, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.203, + 0.633, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.203, + 0.767, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.203, + 0.901, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.26, + 0.23, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.26, + 0.364, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.26, + 0.499, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.26, + 0.633, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.26, + 0.767, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.26, + 0.901, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.316, + 0.23, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.316, + 0.364, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.316, + 0.499, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.316, + 0.633, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.316, + 0.767, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.316, + 0.901, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.385, + 0.907, + 0.413 + ], + "angle": 0, + "content": "Figure 4. MOS score of MPD, visualized in 30 corruption subsets. Different color denotes corruption strength Level 1-Level 2-Level 3-Level 4-Level 5. Results show the sensitivity of the machines to each corruption varied significantly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.424, + 0.34, + 0.439 + ], + "angle": 0, + "content": "3.3. Preference Score Collection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.869 + ], + "angle": 0, + "content": "Following the ITU human subjective annotation standards [71], we adopt the same settings to obtain the machine Mean Opinion Score (MOS). Specifically, for YoN, MCQ, VQA, and CAP tasks, we use 15 general LMMs for inference, namely DeepseekVL [53], InstructBLIP [14], InternVL2 [11], InternLM-XComposer2 [16], LLaVA1.5 [49], LLaVANext [44], LLaVA-OneVision [37], Mantis [32], MiniInternVL [22], MPlugOwl3 [81], Ovis [54], Phi3.5 [1], Pangea [85], Qwen2-VL [75], and Yi1.5 [52]. We set all temperature parameters as 0 to avoid unstable output to ensure all perceptual quality degradations come from distortion. Comprehensively considering popular trends and performance, we implement 5 models for each specific CV tasks, namely SEG: BoxInst [70], ConvNext [77], Mask-RCNN [27], SCNet [73], and RTMDet (segmentation mode) [55]; DET: Dino[89], Mask-RCNN [65], RTMDet (detection mode) [55], ViTDet [34], and Yolo-X [23]; RET: DIME [63], InternVL (retrieval-based tuning) [10], LAVIS (retrieval-based tuning) [36], NAAF [90], and \\(\\mathrm{VSE}++\\) [19]. After obtaining scores for these 7 tasks, the evaluation can be divided into five dimensions, one dimension for each first four LMM tasks, and the last three tasks are unified into the last dimension, ensuring 15 subjects in each \\((0,1)\\) score for MOS. Thus merging all five dimensions into the \\((0,5)\\) MOS. Note that we normalize the score of each dimension to avoid where one dimension has a higher score while another is lower (e.g. VQA/CAP), which would lead to unfairness in the weight of the overall MOS." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In addition, considering the rapid development of LMM, we also apply 3 advanced LMMs, GPT4o [2], GPT4o-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.906, + 0.501 + ], + "angle": 0, + "content": "mini [2], InternLM-XComposer2d5 [16]; and 4 early-stage LMMs, QWen1.5-VL [5], VisualGLM [17], Monkey [45], MPlugOwl2 [82] on certain tasks. We compared these results with the above-mentioned data in MPD, to prove the long-term usability of MPD through their consistency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.513, + 0.707, + 0.529 + ], + "angle": 0, + "content": "4. Database Validation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.599 + ], + "angle": 0, + "content": "This section analyzes the preference data characteristics of MPD, to prove the usability of this database, including two perspectives: (1) the mean opinion of all machines; (2) the difference between each machine as a subject." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Overall: The MOS of MPD and the sub-scores of the seven tasks are illustrated in Figure 3, with their correlation quantified by Spearman Rank-order Correlation Coefficient (SRCC). Notably, the inter-task correlation is relatively low, particularly for VQA and the three tasks that rely on specialized CV models. Given the current state of image processing algorithms for machine vision, each employing different tasks as performance metrics, the universality and robustness of such evaluation are inherently constrained. The data distribution further substantiates this argument, with YoN/MCQ/CAP being more evenly distributed, VQA exhibiting a higher concentration of high scores, and SEG/DET/RET clustering at lower scores. Moreover, the strong correlation between the total score and the other tasks is evident, with all SRCC surpassing 0.5, underscoring the reliability of the MPD, for various downstream tasks. The MOS distribution across 30 types of corruption and 5 strength levels is depicted in Figure 4. Results show for corruption scenarios that appear similar to HVS, the MVS perception mechanism diverges significantly. For" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.949, + 0.516, + 0.96 + ], + "angle": 0, + "content": "7574" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.249, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.178, + 0.218, + 0.192 + ], + "angle": 0, + "content": "Strength 1" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.089, + 0.382, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.178, + 0.352, + 0.192 + ], + "angle": 0, + "content": "Strength 2" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.089, + 0.517, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.178, + 0.486, + 0.192 + ], + "angle": 0, + "content": "Strength 3" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.089, + 0.783, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.178, + 0.617, + 0.192 + ], + "angle": 0, + "content": "Strength" + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.179, + 0.753, + 0.192 + ], + "angle": 0, + "content": "Strength 5" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.099, + 0.874, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.196, + 0.879, + 0.218 + ], + "angle": 0, + "content": "Figure 5. Annotations from different LMMs subjects in MPD under 5 corruption strengths, including \\(\\mathcal{P}\\) content types and 4 tasks." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.219, + 0.222, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.302, + 0.218, + 0.316 + ], + "angle": 0, + "content": "Human (0.76)" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.219, + 0.331, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.302, + 0.321, + 0.316 + ], + "angle": 0, + "content": "MPD (0.62)" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.219, + 0.44, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.302, + 0.425, + 0.316 + ], + "angle": 0, + "content": "YoN (0.50)" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.219, + 0.55, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.302, + 0.54, + 0.316 + ], + "angle": 0, + "content": "MCQ (0.55)" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.219, + 0.659, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.302, + 0.647, + 0.316 + ], + "angle": 0, + "content": "VQA (0.37)" + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.219, + 0.768, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.302, + 0.755, + 0.316 + ], + "angle": 0, + "content": "CAP (0.39)" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.219, + 0.877, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.302, + 0.866, + 0.316 + ], + "angle": 0, + "content": "Other (0.53)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.32, + 0.907, + 0.362 + ], + "angle": 0, + "content": "Figure 6. Correlation matrix of the proposed MPD in each dimension. A darker color denotes a higher SRCC, with mean SRCC attached below each matrix. The index of each matrix is arranged in the \\(\\uparrow\\), \\(\\rightarrow\\) directions, in the same order as Section 3.3. The correlation between human subjects is greater than machines, and the correlation of the overall MOS is greater than the score in each dimension." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.389, + 0.485, + 0.587 + ], + "angle": 0, + "content": "instance, the machine is highly sensitive to 'Lens blur' but rarely affected by 'Mean brighten'. A typical example is the four Macro Block-related corruptions, which are nearly indistinguishable for HVS but exhibit significantly lower perception quality for MVS when encountering 'Block exchange' compared to the other three. Furthermore, the impact of distortion intensity on the MVS perceptual quality varies markedly. For example, as the intensity level increases, the 'Lab saturation' quality diminishes, while the 'HSV saturation' quality remains unchanged. These findings underscore the differences between MVS and HVS, reinforcing the significance of establishing such a unified quality representation in MPD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Each machine: We visualize the benchmark performance of six typical LMMs (as space limited) under different corruption strengths, as shown in Figure 5. For different content types, the scores of various machines on NSI/SCI/AIGI dimensions do not vary significantly. For different tasks, advanced models such as InternVL2 excel in dealing with distorted images in YoN, while traditional models like LLaVA1.5 can provide relatively stable results in CAP. As the corruption strength increases by one level, the scores of each dimension decrease by 0.01-0.02. Finally, for all machine subjects, we explored their internal differences in Figure 6. We manually conducted human annotation for 500 instances (the same 15 subjects) and compared the preference differences between humans and machines. The index of each matrix (i.e., the machine it represents) is arranged in the \\(\\uparrow\\) and \\(\\rightarrow\\) directions, in the same order as in Section 3.3. Overall, the SRCC between humans is 0.76, while that of machines is only 0.62, indicating that the divergence in machine image preferences is greater than that of humans. Further exploration of the preference scores" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.908, + 0.571 + ], + "angle": 0, + "content": "for each dimension shows that the relatedness of machines in VQA and CAP tasks is the lowest. Additionally, there are a few subjects whose assessments differ significantly from other machines\\(^2\\), such as InstructBlip in YoN, Mantis in MCQ, Pangea in CAP, and DIME/NAAF in RET. This is often due to special token outputs or multilingual mechanisms, which lead to insufficient robustness, resulting in a higher proportion of errors in distorted images compared to other models, thus leading to inconsistent results. In summary, the larger preference differences make IQA for MVS a more difficult problem than traditional IQA, but also a more valuable and promising research topic." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.587, + 0.638, + 0.604 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.613, + 0.696, + 0.63 + ], + "angle": 0, + "content": "5.1. Experiment Setups" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.907, + 0.803 + ], + "angle": 0, + "content": "We randomly partitioned the MPD into the train/val set, with 24,000 and 6,000 reference and distorted image pairs, respectively, according to an 8:2 ratio. For MVS, Figure 4 illustrates significant variations in the perceptual quality of different corruptions. Given that machines exhibit a larger Just Noticeable Difference (JND) than humans, our corruption strength exceeds that of previous IQA databases for humans3. To conduct a realistic and reliable evaluation of machine preference, we designate the val set as 'Severe Distortion' and select a subset of 'Mild Distortion' to represent the corruption strength in the real world. The specific splitting" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.815, + 0.905, + 0.876 + ], + "angle": 0, + "content": "2The ITU human subjective annotation standard [71] for humans will exclude data with weak correlation with other subjects, but this is to prevent eye fatigue or malicious attacks during annotation. Given that the machine's inference process is completely objective, we will refer to the outputs of each machine rather than excluding them." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.876, + 0.905, + 0.9 + ], + "angle": 0, + "content": "3See appendix for visualization in 5 corruption strength, where the Level 1 of our MPD can even ensemble Level 4 in KADID-10K[47]." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.815, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.949, + 0.516, + 0.96 + ], + "angle": 0, + "content": "7575" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 2. Using IQA metrics to predict machine preference, including baseline, FR, and NR algorithm, human visual system-based metrics are marked in \\(\\bullet\\). The first row represents two distortion intensities and three image content subsets, and the second row represents five scoring dimensions. [Keys: Best; Second best; Baseline (BL); Inferior to the baseline.]" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.135, + 0.907, + 0.523 + ], + "angle": 0, + "content": "
TypeMetricHVSSevere DistortionMild DistortionNSISCIAIGI
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
BLPSNR0.38730.26720.46640.30970.21270.61970.41160.28200.49450.31350.21460.33290.52250.36840.5985
SSIM[76]0.59690.41570.55370.26720.18280.28180.62880.44660.59320.54150.37480.53930.59800.42780.5687
FRAHIQ[33]0.80610.61530.81830.53040.36850.51110.84470.66010.85020.79620.60400.84700.79960.60820.7966
CKDN[95]✔0.74840.55600.71740.06040.04050.10610.81000.59880.76700.72030.51710.76610.69920.51020.6869
DISTS[15]0.73810.54430.73720.46160.32320.44860.79120.60600.80460.72750.53840.74290.72190.53650.7292
LPIPS[87]✔0.62530.44550.56010.00790.00920.00560.69650.51320.64780.56180.39200.56860.64290.45180.5481
PieAPP[62]✔0.47370.01630.23770.30300.21000.23490.45360.11870.27290.51730.09540.21350.46140.20090.3168
TOPIQ-FR[9]✔0.71780.52520.71820.42480.29440.48410.75060.56140.75850.72430.53220.71860.69430.50630.7061
NRARNIQA[3]0.83370.64260.81120.23880.16080.27540.87040.66900.84780.82200.62170.84200.84490.65780.7954
CLIPQA[74]✔0.64980.46620.59010.18220.12100.17970.76700.57710.71410.66900.47510.62150.68770.49340.5977
DBCNN[88]✔0.56520.39080.56960.13340.08940.15940.59780.41810.62190.56920.39740.60920.54220.38100.5613
HyperIQA[68]0.81050.61890.80120.53880.37740.47120.89080.70740.86910.82730.64030.86010.82120.62420.8076
NIMA[69]0.75620.56570.78580.45700.31710.43420.81530.62030.82240.78000.58810.80920.78460.59530.7622
TOPIQ-NR[9]✔0.65620.46160.63580.19610.13180.19890.70110.48230.66130.66450.46240.63530.54610.34220.5249
TypeMetricHVSYoNMCQVQACAPOthers
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
BLPSNR0.27530.18920.34670.32730.22390.43720.20670.13780.21200.43190.29930.49890.37980.26170.4683
SSIM[76]0.39600.27280.38040.50890.34990.48800.29600.19880.27560.55500.38610.54470.50720.34910.4766
FRAHIQ[33]0.60170.42830.62120.73140.53740.74660.42440.29240.45830.72200.53410.75460.61840.43960.6311
CKDN[95]✔0.54910.37990.54820.68660.48630.65960.40450.27630.39000.70440.51360.69910.54340.38220.5313
DISTS[15]0.52670.36950.53940.64860.46250.65490.35120.23950.36760.66490.48330.68120.62030.44190.6332
LPIPS[87]✔0.46770.32520.42450.55170.39120.51690.32160.21690.29350.56200.39960.54730.49270.33870.4242
PieAPP[62]✔0.33080.01730.16630.43100.19000.20870.23850.01320.10320.42200.01520.21360.37350.02090.1747
TOPIQ-FR[9]✔0.50110.34800.51180.60380.42320.61650.37000.25080.38990.63270.45450.64790.61990.43830.6281
NRARNIQA[3]0.62280.43940.61720.73420.53510.72420.44090.30260.44540.76560.56930.76640.65600.45660.6309
CLIPQA[74]✔0.46130.31670.43860.50900.35560.51170.29710.20280.32040.63790.45510.56730.57510.40480.4747
DBCNN[88]✔0.37980.25670.39690.47930.32580.49510.25560.17190.27830.52400.36330.53880.50850.34650.5093
HyperIQA[68]0.60100.42780.60660.70660.51550.71460.43220.29960.44190.71530.52340.74430.64610.45980.6329
NIMA[69]0.55440.39040.58870.68930.50020.70360.44600.30740.47750.63130.45310.68420.58100.41060.6181
TOPIQ-NR[9]✔0.49390.33580.47970.54690.37130.53700.35780.24370.37870.51120.33750.50890.57690.39340.5564
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.66 + ], + "angle": 0, + "content": "mechanism is detailed in the Appendix. To benchmark the performance of quality metrics, three global indicators were employed: SRCC, Kendall Rank-order Correlation Coefficient (KRCC), and Pearson Linear Correlation Coefficient (PLCC), to evaluate the consistency between the objective quality score and the subjective MOS. Among these, SRCC and KRCC represent the prediction monotonicity, while PLCC measures the accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "12 mainstream IQA metrics are implemented for comparison, which have all shown satisfying results in previous human-centric IQA tasks, including 6 FR metrics: AHIQ [33], CKDN [95], DISTS [15], LPIPS [87], PieAPP [62], TOPIQ-FR [9]; and 6 NR metrics: ARNIQA [3], CLIPIQA [74], DBCNN [88], HyperIAQ [68], NIMA [69], TOPIQN [9]. Some methods specifically model the HVS, such as DBCNN, which has a dedicated stage for simulating human perception; others rely on the remarkable fitting ability of large-scale parameters, like the ViT in AHIQ. We have marked all HVS-based metrics as \\(\\bullet\\). Moreover, two naive quality indicators, PSNR and SSIM [76] are set as the baseline. We train the above 12 metrics on MPD with the learning rate as \\(10^{-5}\\) for 50 epochs, under the default settings in pyiqa [9] toolbox and evaluate the performance on" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.54, + 0.907, + 0.6 + ], + "angle": 0, + "content": "the two subsets above, three content types, and five scoring dimensions. We collect MPD preference annotations on \\(16 \\times\\) NVIDIA A800 SXM4 80GB GPUs, and then conduct IQA training/validation on one GPU above." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.816, + 0.623 + ], + "angle": 0, + "content": "5.2. Experiment Result and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.629, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Table 2 presents the performance of current advanced IQA models on the MPD. Overall, regardless of FR or NR, no metric is highly consistent with machine preference, although they are all good at predicting human preference. An interesting finding is that the more accurately and completely a method models the HVS, the more difficult it is to evaluate the perceptual quality of MVS. Considering that most of the current IQA works emphasize human vision, while machines have already surpassed humans to become the main consumer of visual signals, future IQA needs to find a way to balance these two goals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For the two distortion subsets, fitting-based methods such as AHIQ/ARNIQA/HyperIQA can achieve SRCC/PLCC values exceeding 0.8 on Severe Distortion. However, this is not a challenging task. Under the same distortion conditions, they have even exceeded 0.95 on the HVS dataset [33, 61]. The performance of Mild Distortion is even worse, and no method can even outperform the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.949, + 0.516, + 0.961 + ], + "angle": 0, + "content": "7576" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.159 + ], + "angle": 0, + "content": "Table 3. Cross validation between human and machine preference database. For IQA metrics, training on machine preference data can reach a good result in evaluating human preference; but training on human preference is unacceptable for MPD. [Keys: Best; Lower than 0.4.]" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.161, + 0.485, + 0.328 + ], + "angle": 0, + "content": "
Val-setKADID-10KTID2013MPD (Mild)MPD (Severe)
MetricSRCC↑PLCC↑SRCC↑PLCC↑SRCC↑PLCC↑SRCC↑PLCC↑
AHIQ[33]0.79500.77320.76510.78840.15140.18460.47530.3518
CKDN[95]0.85990.72820.69890.60290.17130.20550.34940.3228
DISTS[15]0.82390.81900.72310.77590.41050.40770.46160.4486
LPIPS[87]0.61840.54510.70540.63260.23910.31180.46830.3711
PieAPP[62]0.47830.34680.68320.45430.12850.06780.17910.1235
TOPIQ-FR[9]0.67680.68440.77820.74730.51120.54950.62230.6685
ARNIQA[3]0.76840.70230.77010.72050.48290.45820.55970.4849
CLIPIQA[62]0.62860.57290.52370.39100.13670.14780.32010.3203
DBCNN[88]0.51750.53510.44860.57860.13590.15920.40240.3877
HyperIQA[68]0.66500.65470.70240.73130.08200.10740.45190.4521
NIMA[69]0.47950.49150.56940.64230.16290.17750.47090.4845
TOPIQ-NR[9]0.52320.50770.54550.61740.23880.27540.48290.4582
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.485, + 0.598 + ], + "angle": 0, + "content": "simple PSNR, which proves that there is still a lot of room for optimization of IQA for machine preference. For the three contents, after verifying them separately, the performance has slightly increased, which is caused by the internal quality differences of the contents. In general, predicting NSI quality is relatively simple, while SCI and AIGI are more difficult. This may be due to the fact that the initial weights of the IQA model have rich NSI priors. For the five dimensions, we can see that the results of MCQ and CAP are most consistent with machine preference; YoN and Others (three common CV tasks) are second; VQA is the worst. This is because VQA problems are usually changes in image details (such as the color of a button on a piece of clothing), which are more difficult for machines to detect than other tasks. In short, IQA for Machine preference still has great research potential, and it is necessary to develop more accurate and effective evaluation mechanisms in the future." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.332, + 0.622 + ], + "angle": 0, + "content": "5.3. Cross Database Validation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.486, + 0.841 + ], + "angle": 0, + "content": "To further analyze the differences between MVS and HVS, we conducted cross-validation using MPD and two of the most commonly used human-oriented datasets, TID2013 [61] and KADID-10K [47], employing the same training mechanism as in the previous chapter. We observed that after training on MPD, most IQA metrics still possess a certain ability to predict human preferences, with SRCC values around 0.7 to 0.8; conversely, after training on the other two human-centric databases, IQA metrics struggle to predict machine preferences, with results on MPD not even reaching 0.4. Machine preference data will not affect human tasks, but the opposite will. This finding holds certain guiding significance for future IQA, particularly for a human-machine-friendly paradigm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.388, + 0.866 + ], + "angle": 0, + "content": "5.4. Long-term Availability Validation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "The machine subjects in MPD are categorized into two groups: common CV models and LMMs. The former mod" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.081, + 0.931, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.233, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 7. The long-term usability of MPD. The preference scores of MPD are consistent with four early-stage LMMs and three advanced LMMs (marked as Traditional/Advanced Models in the background). Their high SRCC proves that MPD can still characterize the perceptual quality of LMMs in the near future." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.331, + 0.909, + 0.649 + ], + "angle": 0, + "content": "els, such as the separation/detection benchmark method, have been largely finalized and generally recognized. In contrast, LMMs are still under rapid evolvement. It is conceivable that MPD may not be able to capture the LMMs that will emerge in the future, or with the update of MPD, the preference of previous LMMs may be sacrificed. Thus, we include 3 advanced and 4 early-stage LMMs as Sec 3.3 listed, integrating them into two databases similar to MPD. Given the inconsistencies in LMM preference on VQA and CAP tasks, as illustrated in Figure 6, we have verified the consistency of MPD ground truth labels with advanced/early-stage labels. For each image, the relationship between these two labels is depicted in Figure 7. The findings indicate that MPD exhibits a high correlation with early models on VQA and with advanced models on CAP. Generally, the SRCC is around 0.6, which is considered acceptable. Considering the GPT-4o surpasses open-source models, we anticipate that MPD is able to characterize LMMs preferences for images for at least a year. In the future, We will continue to update MPD by periodicity introducing the latest LMMs to avoid this limitation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This paper extends the application of IQA from human to machine vision for the first time and establishes the first large-scale, fine-grained, multi-dimensional quality database for machines, namely MPD. We select 15 LMMs and 15 common CV models as subjects, show them \\(30\\mathrm{k}\\) reference/distorted image pairs, and summarize their performance on 7 downstream tasks as quality scores. In-depth analysis of the MPD content and IQA benchmark experiments have proved the huge gap between machine and human visual systems, and current human-centric IQA methods cannot accurately predict machine preferences. As the visual data consumption volume of machines has already surpassed humans, we hope that MPD can inspire effective IQA metrics to serve the upcoming machine-centric era." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.949, + 0.516, + 0.961 + ], + "angle": 0, + "content": "7577" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.177 + ], + "angle": 0, + "content": "The work was supported by the National Natural Science Foundation of China under Grant 62225112, 62301310, 623B2073, 62471290; and in part by Sichuan Science and Technology Program under Grant 2024NSFSC1426." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.189, + 0.188, + 0.205 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[3] Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. Arniqa: Learning distortion manifold for image quality assessment. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 188-197, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.484, + 0.508 + ], + "angle": 0, + "content": "[4] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 382-398. Springer, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[6] Johannes Balle, Philip A Chou, David Minnen, Saurabh Singh, Nick Johnston, Eirikur Agustsson, Sung Jin Hwang, and George Toderici. Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing, 15(2):339-353, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[7] Min Cao, Shiping Li, Juntao Li, Liqiang Nie, and Min Zhang. Image-text retrieval: A survey on recent research and development. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 5410-5417. International Joint Conferences on Artificial Intelligence Organization, 2022. Survey Track. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.719, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[8] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[9] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 33:2404-2418, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.214, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24185–24198, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[12] Jinyoung Choi and Bohyung Han. Task-aware quantization network forJPEG image compression. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 309-324. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[13] U Cisco. Cisco annual internet report (2018-2023) white paper. Cisco: San Jose, CA, USA, 10(1):1-35, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.448 + ], + "angle": 0, + "content": "[15] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[16] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Ji-aqi Wang. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[17] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.646, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[18] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, Dahua Lin, and Kai Chen. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[19] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[20] Wen Gao, Siwei Ma, Lingyu Duan, Yonghong Tian, Peiyin Xing, Yaowei Wang, Shanshe Wang, Huizhu Jia, and Tiejun Huang. Digital retina: A way to make the city brain more efficient by visual coding. IEEE Transactions on Circuits and Systems for Video Technology, 31(11):4147-4161, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[21] Yixuan Gao, Yuqin Cao, Tengchuan Kou, Wei Sun, Yunlong Dong, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Vdpve: Vqa dataset for perceptual video enhancement. In Proceedings of the IEEE/CVF Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.949, + 0.516, + 0.96 + ], + "angle": 0, + "content": "7578" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Vision and Pattern Recognition (CVPR) Workshops, pages 1474-1483, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[22] Zhangwei Gao, Zhe Chen, Erfei Cui, Yiming Ren, Weiyun Wang, Jinguo Zhu, Hao Tian, Shenglong Ye, Junjun He, Xizhou Zhu, Lewei Lu, Tong Lu, Yu Qiao, Jifeng Dai, and Wenhai Wang. Mini-internvl: A flexible-transfer pocket multimodal model with 55" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[23] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.318 + ], + "angle": 0, + "content": "[24] Jinjin Gu, Haoming Cai, Chao Dong, Jimmy S Ren, Radu Timofte, Yuan Gong, Shanshan Lao, Shuwei Shi, Jiahao Wang, Sidi Yang, et al. Ntire 2022 challenge on perceptual image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 951-967, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.361 + ], + "angle": 0, + "content": "[25] Yatharth Gupta, Vishnu V. Jaddipal, Harish Prabhala, Sayak Paul, and Patrick Von Platen. Progressive knowledge distillation of stable diffusion x1 using layer level loss, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[26] Shijie Hao, Yuan Zhou, and Yanrong Guo. A brief survey on semantic segmentation with deep learning. Neurocomputing, 406:302-321, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[27] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.481, + 0.475 + ], + "angle": 0, + "content": "[28] David Holz. Midjourney. https://www.midjourney.com, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[29] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[30] Yunzhong Hou, Liang Zheng, and Stephen Gould. Learning to structure an image with few colors. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 10116-10125, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[31] Zhimeng Huang, Chuanmin Jia, Shanshe Wang, and Siwei Ma. Visual analysis motivated rate-distortion model for image coding. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[32] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[33] Shanshan Lao, Yuan Gong, Shuwei Shi, Sidi Yang, Tianhe Wu, Jiahao Wang, Weihao Xia, and Yujiu Yang. Attentionshelp cnns see better: Attention-based hybrid image quality assessment network, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[34] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[35] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiqa-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[37] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[38] Chunyi Li, Tengchuan Kou, Yixuan Gao, Yuqin Cao, Wei Sun, Zicheng Zhang, Yingjie Zhou, Zhichao Zhang, Weixia Zhang, Haoning Wu, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Aigia-20k: A large database for a-generated image quality assessment, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[39] Chunyi Li, Guo Lu, Donghui Feng, Haoning Wu, Zicheng Zhang, Xiaohong Liu, Guangtao Zhai, Weisi Lin, and Wenjun Zhang. Misc: Ultra-low bitrate image semantic compression driven by large multimodal model, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[40] Chunyi Li, Haoning Wu, Hongkun Hao, Zicheng Zhang, Tengchuan Kou, Chaofeng Chen, Lei Bai, Xiaohong Liu, Weisi Lin, and Guangtao Zhai. G-refine: A general quality refiner for text-to-image generation. In Proceedings of the 32nd ACM International Conference on Multimedia, page 7375-7384, New York, NY, USA, 2024. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[41] Chunyi Li, Haoning Wu, Zicheng Zhang, Hongkun Hao, Kaiwei Zhang, Lei Bai, Xiaohong Liu, Xiongkuo Min, Weisi Lin, and Guangtao Zhai. Q-refine: A perceptual quality refiner for ai-generated image. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[42] Chunyi Li, Xiele Wu, Haoning Wu, Donghui Feng, Zicheng Zhang, Guo Lu, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Cmc-bench: Towards a new paradigm of visual signal compression, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[43] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[44] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[45] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[46] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[47] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Tenth International Conference on Quality of Multimedia Experience (QoMEX), pages 1-3. IEEE, 2019. 2, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.949, + 0.516, + 0.96 + ], + "angle": 0, + "content": "7579" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[48] Risheng Liu, Long Ma, Tengyu Ma, Xin Fan, and Zhongxuan Luo. Learning with nested scene modeling and cooperative architecture search for low-light vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5953-5969, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[50] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[51] Xiaohong Liu, Xiongkuo Min, Guangtao Zhai, Chunyi Li, Tengchuan Kou, et al. Ntire 2024 quality assessment of ai-generated content challenge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 6337-6362, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[52] Yuan Liu, Le Tian, Xiao Zhou, and Jie Zhou. Rethinking overlooked aspects in vision-language models. arXiv preprint arXiv:22405.11850, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[53] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Yaofeng Sun, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[54] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[55] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[56] Xiongkuo Min, Kede Ma, Ke Gu, Guangtao Zhai, Zhou Wang, and Weisi Lin. Unified blind quality assessment of compressed natural, graphic, and screen content images. IEEE Transactions on Image Processing, 26(11):5462-5474, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[57] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE transactions on Image Processing, 20 (12):3350-3364, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[58] Zhangkai Ni, Lin Ma, Huanqiang Zeng, Jing Chen, Canhui Cai, and Kai-Kuang Ma. Esim: Edge similarity for screen content image quality assessment. IEEE Transactions on Image Processing, 26(10):4818-4831, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[59] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[60] Neel Patwa, Nilesh Ahuja, Srinivasa Somayazulu, Omesh Tickoo, Srenivas Varadarajan, and Shashidhar Koolagudi. Semantic-preserving image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 1281-1285. IEEE, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[61] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, et al. Image database tid2013: Peculiarities, results and perspectives. Signal processing: Image communication, 30:57-77, 2015. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[62] Ekta Prashnani, Hong Cai, Yasamin Mostofi, and Pradeep Sen. Pieapp: Perceptual image-error assessment through pairwise preference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1808-1817, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[63] Leigang Qu, Meng Liu, Jianlong Wu, Zan Gao, and Liqiang Nie. Dynamic modality interaction modeling for image-text retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1104-1113, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[64] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[65] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[66] Robin Rombach, Andreas Blattmann, and Björn Ommer. Text-guided synthesis of artistic images with retrieval-augmented diffusion models, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[67] Chengyao Shen and Qi Zhao. Webpage saliency. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13, pages 33-46. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[68] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[69] Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment. IEEE Transactions on Image Processing, 27(8): 3998-4011, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[70] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[71] I. T. Union. Methodology for the subjective assessment of the quality of television pictures. ITU-R Recommendation BT. 500-11, 2002. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[72] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[73] Thang Vu, Haeyong Kang, and Chang D Yoo. Scnet: Training inference sample consistency for instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2701-2709, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[74] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.949, + 0.516, + 0.96 + ], + "angle": 0, + "content": "7580" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "ceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[75] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[76] Zhou Wang. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.317 + ], + "angle": 0, + "content": "[77] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[78] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-bench: A benchmark for general-purpose foundation models on low-level vision. arXiv preprint arXiv:2309.14181, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.391, + 0.274, + 0.404 + ], + "angle": 0, + "content": "[79] XAI. grok-1.5v, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[80] Zhaohui Yang, Yunhe Wang, Chang Xu, Peng Du, Chao Xu, Chunjing Xu, and Qi Tian. Discernible image compression. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1561-1569, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[81] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[82] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13040–13051, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[83] Masakazu Yoshimura, Junji Otsuka, Atsushi Irie, and Takeshi Ohashi. Dynamicisp: dynamically controlled image signal processor for image recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12866-12876, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[84] Ke Yu, Zexian Li, Yue Peng, Chen Change Loy, and Jinwei Gu. Reconfigisp: Reconfigurable camera image processing pipeline. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4248-4257, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[85] Xiang Yue, Yueqi Song, Akari Asai, Seungone Kim, Jean de Dieu Nyandwi, Simran Khanuja, Anjali Kantharuban, Lintang Sutawika, Sathyanarayanan Ramamoorthy, and Graham Neubig. Pangea: A fully open multilingual multimodal llm for 39 languages, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[86] Jing Zhang and Dacheng Tao. Empowering things with intelligence: a survey of the progress, challenges, and opportunities in artificial intelligence of things. IEEE Internet of Things Journal, 8(10):7789-7817, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[87] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[88] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[89] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[90] Kun Zhang, Zhendong Mao, Quan Wang, and Yongdong Zhang. Negative-aware attention framework for image-text matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15661-15670, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[91] Zicheng Zhang, Chunyi Li, Wei Sun, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. A perceptual quality assessment exploration for aigc images. In IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 440-445, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.485 + ], + "angle": 0, + "content": "[92] Zicheng Zhang, Wei Sun, Yingjie Zhou, Jun Jia, Zhichao Zhang, Jing Liu, Xiongkuo Min, and Guangtao Zhai. Subjective and objective quality assessment for in-the-wild computer graphics images. ACM Transactions on Multimedia Computing, Communications and Applications, 20(4):1-22, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.542 + ], + "angle": 0, + "content": "[93] Qi Zhang, Shanshe Wang, Xinfeng Zhang, Chuanmin Jia, Zhao Wang, Siwei Ma, and Wen Gao. Perceptual video coding for machines via satisfied machine ratio modeling, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.544, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[94] Qi Zhang, Shanshe Wang, Xinfeng Zhang, Siwei Ma, Jingshan Pan, and Wen Gao. Predicting satisfied user and machine ratio for compressed images: A unified approach, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[95] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10242-10251, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[96] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.949, + 0.514, + 0.96 + ], + "angle": 0, + "content": "7581" + } + ] +] \ No newline at end of file diff --git a/2025/Image Quality Assessment_ From Human to Machine Preference/9c43c7f6-7667-4c47-8440-32e15bc27376_origin.pdf b/2025/Image Quality Assessment_ From Human to Machine Preference/9c43c7f6-7667-4c47-8440-32e15bc27376_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..51fcbe1aeca463b16078bd3a69764f9415a2e27f --- /dev/null +++ b/2025/Image Quality Assessment_ From Human to Machine Preference/9c43c7f6-7667-4c47-8440-32e15bc27376_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:026e21a990f146cf93f34e72dab58fe5cb823707863427f38ff9cbd01042a88c +size 1374479 diff --git a/2025/Image Quality Assessment_ From Human to Machine Preference/full.md b/2025/Image Quality Assessment_ From Human to Machine Preference/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5ebd97811fe678eb834f1c79ea33d017a22ca538 --- /dev/null +++ b/2025/Image Quality Assessment_ From Human to Machine Preference/full.md @@ -0,0 +1,405 @@ +# Image Quality Assessment: From Human to Machine Preference + +Chunyi Li $^{1,2,4*}$ , Yuan Tian $^{1,2*}$ , Xiaoyue Ling $^{1*}$ , Zicheng Zhang $^{1,2}$ , Haodong Duan $^{2}$ , Haoning Wu $^{3,4}$ , Ziheng Jia $^{1}$ , Xiaohong Liu $^{1}$ , Xiongkuo Min $^{1}$ , Guo Lu $^{1\dagger}$ , Weisi Lin $^{4}$ , Guangtao Zhai $^{1,2\dagger}$ +Shanghai Jiao Tong University $^{1}$ , Shanghai AI Lab $^{2}$ , Moonshot AI $^{3}$ , Nanyang Technological University $^{4}$ + +![](images/6d346aa8b3722dceff3a924054764b758fe8d0bd085c64b4c82ffc951c515063.jpg) +Figure 1. The significant gap between the well-explored Human Vision System (HVS) and the emerging Machine Vision System (MVS). Their different perception mechanisms leading to images that are subjectively satisfactory to humans not being applicable to the machines downstream tasks such as detection and question answering, vice versa. + +# Abstract + +Image Quality Assessment (IQA) based on human subjective preferences has undergone extensive research in the past decades. However, with the development of communication protocols, the visual data consumption volume of machines has gradually surpassed that of humans. For machines, the preference depends on downstream tasks such as segmentation and detection, rather than visual appeal. Considering the huge gap between human and machine visual systems, this paper proposes the topic: Image Quality Assessment for Machine Vision for the first time. Specifically, we (1) defined the subjective preferences of machines, including downstream tasks, test models, and evaluation metrics; (2) established the Machine Preference Database (MPD), which contains 2.25M fine-grained annotations and 30k reference/distorted image pair instances; (3) verified the performance of mainstream IQA algorithms on MPD. Experiments show that current IQA metrics are human-centric and cannot accurately characterize machine preferences. We sincerely hope that MPD can promote the evolution of IQA from human to machine preferences. Project page is on: https://github.com/lcyszxdxc/MPD. + +# 1. Introduction + +The rapid ascend of smart cities [20] and the flourishment of the Internet of Things (IoT) [86] over the past decade have fundamentally transformed the framework of the application end. According to the Cisco [13] white paper, the number of Machine-to-Machine (M2M) connections first exceeded that of Machine-to-Human (M2H) in 2023, reaching 147 billion. Machines have gradually replaced humans and become the primary consumers of image/video data. In the field of image processing, the primary goal is to enhance the quality of processed images, ensuring they conform to human perception. Previously, the performance of a compression/reconstruction/generation algorithm [39-41] is often judged by its visual appeal to humans. This technology, known as Image Quality Assessment (IQA), comprehensively models the Human Visual System (HVS) to predict human subjective preferences. Consequently, with the evolution of application ends, IQA today has to transcend HVS and serve Machine Vision Systems (MVS) more effectively. + +However, there are significant disparities in the perceptual mechanisms between HVS and MVS. This discrepancy arises because HVS prioritizes the similarity between reference and distorted images, such as texture, structure, and color, whereas machines focus on the consistency of results + +Table 1. Comparison of MPD with other IQA databases. As the first machine preference-oriented library, MPD has a larger image scale, a wider range of corruption scenarios, and more complete annotation content. + +
DatabaseImageCorruptionAnnotation
ReferenceDistortedResolutionTypesStrengthNumDimensionLabeling Criteria
LIVE[57]297797685525k1Human preference on NSIs
TID2013[61]253k512245514k1Human preference on NSIs
KADID-10K[47]8110k1k255304k1Human preference on NSIs
CLIC2021[6]5853k1k103484k1Human preference on compressed images
NTIRE2022[24]25029k2884051.13m1Human preference on super-resolution images
KonIQ-10K[29]-10k1k--1.21m5Human preference on NSIs
AGIQA-3K[35]-3k1k--125k2Human preference on AIGIs
NTIRE2024[38]-20k1k--420k2Human preference on AIGIs
MPD (ours)1k30k1k3052.25m5Machine preference on NSIs, SCIs, and AIGIs
+ +in downstream tasks. As illustrated in Figure 1, for distortions resulting from factors like encoding/quantization, HVSs perceive a noticeable degradation in quality, yet the performance of tasks like segmentation/detection by machines remains unaffected. Conversely, some minor disturbances that the human eye barely notices can significantly disrupt the output of MVSs, making it difficult to migrate their evaluation algorithms. + +What images do machines prefer? This remains an open question. For HVS, over the past two decades, IQA has witnessed significant development. Given the variations in image preferences among individuals, particularly across dimensions such as brightness and chroma, the scoring dimensions, the annotations from multiple subjects, and the integration of these annotations into global subjective preference labels have been well established. Hundreds of fine-grained databases can drive IQA algorithms that closely align with human perception. Conversely, MVS development is much slower, regardless of an end-to-end IQA algorithm, there is even no database that annotates images with an overall preference score of machines. This hinders the comprehensive evaluation of image processing metrics designed for machine vision, limiting the experience of application ends. Considering these issues, we first attempt to generalize the concept of IQA from HVS to MVS. Our contributions can be summarized as follows: + +- A innovative definition of the subjective preferences of machines. We propose a standardized process for collecting subjective labels from machines. Specifically, we identify the downstream tasks that determine the quality of machine perception, including four multimodal tasks related to Large Multimodal Models (LMMs) and three classic Computer Vision (CV) tasks. By referencing the multi-dimensional scoring mechanism of humans, we comprehensively evaluate the performance of machines in these tasks to obtain ground truth labels. +- A large-scale database named the Machine Preference Database (MPD). Using 15 general LMMs and 15 spe + +cialized CV models, we collected 2.25 million preference data points on 30,000 reference/distorted image pairs, establishing the first IQA database for machine vision. In particular, we conducted an in-depth analysis of each participating machine to explore the internal patterns of MVS perception and their differences with HVS. + +- A complex benchmark test for existing IQA algorithms. Experiments on the MPD demonstrate that these algorithms designed for HVS cannot fully model MVS, whose predictions do not match machine ratings. After finetuning based on the MPD, their performance improves slightly, but still cannot characterize machine preferences, especially when facing mild distortion. + +# 2. Related Works + +# 2.1. Image Quality Assessment Database + +IQA datasets include Full Reference (FR) and No Reference (NR). FR datasets manually add corruption to reference images, whereas NR directly capture distorted images. Typically, these datasets focus on a single image type, such as Natural Scenes Images (NSIs) [6, 24, 47, 57, 61], Screen Content Images (SCIs) [21, 56, 58], and AI-Generated Images (AIGIs) [35, 38, 51, 91], with quality labels derived solely from human preferences as listed in Table 1. On contrast, in the era where machines surpass humans as application ends, MPD stands out as the first machine preference-oriented library. It offers a larger image scale, a broader range of corruption scenarios, and a more comprehensive annotation context than other IQA databases. + +# 2.2. Machine Vision Downstream Task + +In the absence of a unified quality indicator, image processing algorithms, taking compression and restoration metrics for machine vision for example, are often in a fragmented state, with the performance of each algorithm being validated by only one or two downstream tasks. [93, 94] These tasks vary, as do the models used to perform them, mak + +![](images/673ed7353ff32666bd066c02e6968121cce7d918dd5ab57864b91b5f7406ca12.jpg) +Figure 2. Overview of Machine Preference Database (MPD), with 30k large-scale reference/distorted image pairs, meticulously annotated with 2.25M fine-grained preference score from the mean opinion of 15 mainstream LMMs and 15 specific CV models. It provides a multi-dimensional fidelity evaluation of various downstream machine vision tasks. + +ing the evaluation criteria disparate and incomplete. For compression algorithms, validation indicators include the classification [30, 60], image captioning [12], object detection [80], or semantic segmentation [31] result before/after compression, with each indicator differing and lacking standardization or completeness. Restoration algorithms fare slightly better, with most consistently relying on object detection [48, 83, 84]. Nevertheless, some employ traditional Fast-RCNN [65], while others utilize more advanced models like ViTDet[34]. The variation in model bases further compromises the comprehensiveness and objectivity of the evaluation outcomes. Thus, there is an urgent need for a unified representation of machine preferences. + +# 3. Database Construction + +# 3.1. Reference & Distorted Image Collection + +Similar to the previous IQA database, we first selected 1,000 high-quality reference images and then artificially degraded their quality to obtain 30,000 pairs of reference and distorted images. To comprehensively represent the common content on the internet, we referred to the CMC-Bench [42] and selected the three most common image categories: Natural Scene Images (NSIs), including + +User/Professional Generated Content (UGC/PGC) [6, 46]; Screen Content Images (SCIs), including webpages [67], games [92], and movies [56]; and AI-Generated Images (AIGIs), generated by six representative text-to-image models [8, 25, 28, 43, 64, 66]. The collection process is shown in Figure 2. Both human viewers and IQA models have verified this image sequence to ensure that they are undistorted as reference images. + +For the distorted images, we modeled the corruption caused by the current communication protocols, which were divided into 30 scenarios. Specifically, these corruptions were classified into seven categories: Blur, various types of unclear image; Luminance, global brightness changes; Chrominance, global color changes; Contrast, global changes in contrast/saturation; Noise, random noise of different distributions; Compression, codec algorithm like JPEG; and Spatial, local pixel-level changes. For each corruption, we defined five levels of intensity, from mild to severe. To fairly explore the impact of each corruption, we artificially control the distortion strength to ensure that quality degradation perceived by the human eye is + +aligned at the same level. Thus, for each reference image, we randomly selected the intensity to add all the corruptions mentioned above, resulting in 30,000 distorted images. The reference/distorted image pairs will then be annotated by machine subjects. + +# 3.2. Different Task Annotation + +In the past five years, the evaluation of image processing algorithms for machine vision has primarily focused on seven tasks: Regression, Classification, Visual Question Answering (VQA), Caption (CAP), Segmentation (SEG), Detection (DET), and Retrieval (RET). Among these, the first four tasks can be accomplished by general LMMs, whereas the latter three tasks necessitate specialized models due to their unique requirements. Note that the regression and classification tasks in this context are analogous to Yes-or-No (YoN) and Multiple-Choice Questions (MCQ) within LMMs. To this end, we enlisted five experienced human experts to design YoN/VQA/CAP questions and answers for 1,000 reference images. These questions are designed to be challenging, as illustrated in the example provided in Figure 2, especially in the MCQ sets where confusion options are introduced. This approach enables us to accurately measure the degree of quality degradation perceived by the machine. + +We apply these seven tasks to both the reference and distorted images, measure the similarity of the output results, and derive scores for each of the seven sub-dimensions. Specifically, for $Y_{oN}$ tasks, we employ a logit-level evaluation approach based on softmax similarity, as previous LMM benchmarks [50, 78] do. This entails LMMs confirming or denying the questions designed for each reference/distorted image pair. The confidence difference in the next token output, 'Yes' or 'No', reflects the difference in the perceptual quality degradation from the perspective of machines, namely a score $S_{Y_{oN}}$ : + +$$ +S _ {Y o N} = \left| \sigma \left(P _ {\text {d i s}} (\text {Y e s}, \text {N o})\right) - \sigma \left(P _ {\text {r e f}} (\text {Y e s}, \text {N o})\right) \right|, \tag {1} +$$ + +where $P_{dis,ref}(\cdot)$ denotes the output probability of certain logits, normalized by the softmax function $\sigma(\cdot)$ . MCQ tasks follow a similar evaluation paradigm. We set a correct option and three confusing options (the options are usually no more than 4, see [18, 50, 79]), and the model outputs the probability of each option as a four-element vector. The distance between the two vectors in space can reflect the perceptual quality score $S_{MCQ}$ for machines: + +$$ +S _ {M C Q} = \cos \left(P _ {d i s} (\mathrm {A}, \mathrm {B}, \mathrm {C}, \mathrm {D}), P _ {r e f} (\mathrm {A}, \mathrm {B}, \mathrm {C}, \mathrm {D})\right), \tag {2} +$$ + +where the distance is measured by cosine similarity. For $VQA$ tasks, we conduct sentence-level evaluation beyond the logits. LMMs will be required to answer a certain question in about 5 words, and the semantic difference between two output sentences will be the score $S_{VQA}$ : + +$$ +S _ {V Q A} = \operatorname {C L I P} \left(T _ {\text {d i s}}, T _ {\text {r e f}}\right), \tag {3} +$$ + +![](images/d2bae0fc7fe3d3f3f89cd0af85c937ba37dd6748e5e2e3a17fc4a817173b4015.jpg) +(a) Correlation matrix + +![](images/e7c910a6015712bc19d0552851df8d958166593387a3df4d2849170438e5d109.jpg) +(b) Score distribution +Figure 3. Correlation between the general preference score and the score in seven different downstream tasks. + +where $T_{dis,ref}$ represents the whole output text, measured by the CLIP text encoder. $CAP$ tasks require a longer output, by describing the main subjects, actions, settings, and any notable objects or features of the image in about 40 words. Considering the remarkable text length, we use indicators dedicated to image captioning to analyze the differences between the two long sentences in $S_{CAP}$ : + +$$ +S _ {C A P} = \sum_ {\text {e v a l}} \operatorname {e v a l} \left(T _ {d i s}, T _ {r e f}\right), \tag {4} +$$ + +where eval $\in$ (BLEU, CIDEr, SPICE) [4, 59, 72], three most commonly-used captioning indicators. Then, for SEG tasks, referring to their common evaluation criteria [26], we calculate the Intersection-over-Union (IoU) ratio in $S_{SEG}$ : + +$$ +S _ {S E G} = \operatorname {I o U} \left(M _ {d i s}, M _ {r e f}\right), \tag {5} +$$ + +where $M_{dis,ref}$ stands for binarized segmentation masks. DET tasks will be more complex, whose results include both spatial information and category labels. Thus, we consider the detection failed when the categories are different, and only further calculate the IoU when they are consistent [96], as described in $S_{DET}$ : + +$$ +S _ {D E T} = \operatorname {A c c} _ {1} \left(C _ {d i s}, C _ {r e f}\right) \cdot \operatorname {I o U} \left(B _ {d i s}, B _ {r e f}\right), \tag {6} +$$ + +where $C_{dis,ref}$ denotes the category sequence ordered by confidence, and $B_{dis,ref}$ is the detected bounding box. Finally for RET tasks, we merge three common Image-to-Text retrieval indicators [7], namely Top-1/5/10 accuracy, to evaluate the retrieved sequence as $S_{RET}$ : + +$$ +S _ {R E T} = \sum_ {\mathrm {i} \in 1, 5, 1 0} \operatorname {A c c} _ {\mathrm {i}} \left(C _ {d i s}, C _ {r e f}\right). \tag {7} +$$ + +After the above process, we can obtain the preferences of LMMs in the first four dimensions, and the preferences of each specific CV model on its target task. + +![](images/e8d65818a113bf9db39eef65672b57531532cab7b2e8e46415876d5b43f55b1c.jpg) +Gaussian...filter + +![](images/db3f628068b730c7a69e833d98f294b9ef40d8ff65226b69281711602df07bf2.jpg) +Lens...blur + +![](images/1c212d6e249c3417c8c55869afbaffc0a924ffc9efcaf1480208a7103bebac37.jpg) +Motion...blur + +![](images/bd6925fd0641dc3a5421fe1ffac387c86e3452f6a9ad279cac5e2c8287428640.jpg) +Color...diffusion + +![](images/75fb09b407bf5b3246d0c0e2b871aa824ec78fb3ed8f886e7a36db5f249742bd.jpg) +Color...shift + +![](images/22d3543ebc8db02e13ff242573fe268d52b9a499019c9ec30007e2d4691e093f.jpg) +Color...quantization + +![](images/59b4dc0f28f900c096aeb50c01cd4f8718782cb4eed09ebc56c9e59faa57ea79.jpg) +HSV...saturation + +![](images/1257e89ecb593dbf6afb89b344e384087939a7eaa029a06c58634d50f739606d.jpg) +Lab...saturation + +![](images/35a7e7be1a60ea062193f22f600003864dc75c9eb912ed965e1a3f0a9c23fd1f.jpg) +JP2K...compression + +![](images/efc4e2b5af31b2fdc7949355d8cbafb7e0f1247b9740dc049f96061792a2b7d7.jpg) +JPEG...compression + +![](images/2f3d141369cb920ba485bbed96c11e515aafdb552d3e36e3ae9198ccf7c6ee4b.jpg) +WEBP...compression + +![](images/8ed2cc6cf417c7c524641651d9bf03aebb8c47de42cd9b4c57ff61924ff65e9e.jpg) +White...noise + +![](images/88b66cc9a3e3ad8fc3e335d902dc36683ab9662c74ce131e3b2c728d7a4ba7dd.jpg) + +![](images/46d1e2daed45ff8ee19f0338adbe1808d8bb4c7d62a97ab5ed4c7d1fb180ddb2.jpg) + +![](images/18bbb6e8651196f9254edffb7e7e13b0593dc09b471110380e0f06a99fc395ed.jpg) + +![](images/fc8f917b543e54f7e2078c9d335041f5c37de5627322668de6861dbdd2f33359.jpg) + +![](images/ee80b749df35adb05c60d05e59c0f80059074846d042f29655f28c1e2ac3fc6a.jpg) + +![](images/5f13ce1bcbf5f3674405705c9ed11f5d1f241f864ab65cb4493a3ac7388ecf8e.jpg) + +![](images/300f69721f4929300ec94ec5f5b58890bd59b38f4e855e79a9421808136aa2d2.jpg) + +![](images/1f6804de1385154997b507f3fe6ca1762d1a03936cc7ff72d166b3eeabe4e081.jpg) + +![](images/1ec7fc27464c5cd3b12af9ad3bd3fc1ff28905c4066271f8bf046f838aa4314a.jpg) + +![](images/f54ae830e0f90b05116bd9a1eedde014f1bda51a373d2bad8b209c5522858c59.jpg) + +![](images/af3155f66e6d9ccf282555c882731163e7ba691bd666d4139753f64baef0a55c.jpg) + +![](images/77acb24a647dd17b6f22829f6f62d22bff02f79238b27cb36798741526756d4b.jpg) + +![](images/33c98ebaee779c2ba8f22326c350fc20bef7b2eb6de7ac20a837919011a33ad5.jpg) +Figure 4. MOS score of MPD, visualized in 30 corruption subsets. Different color denotes corruption strength Level 1-Level 2-Level 3-Level 4-Level 5. Results show the sensitivity of the machines to each corruption varied significantly. + +![](images/5027069540ea3bccbeb6aa54dde68289b5297373efb222dd005c36037a48aa6e.jpg) + +![](images/b3513eb939171d21592e79de294567b69057083d38c250082d139c15e01748a2.jpg) + +![](images/0c4211e552dcb5536a75c26c01be1e2c7f939c858c6776df8b06538a6803c345.jpg) + +![](images/e73fa53bc956bc2ceac7b635b046354e1742cc5cdc74a388c8f9d7411dad0655.jpg) + +![](images/4b8229b3b26febf67320c8bb388a72ad7b5cc44843f5f1bdb41ed485a79c43b0.jpg) + +# 3.3. Preference Score Collection + +Following the ITU human subjective annotation standards [71], we adopt the same settings to obtain the machine Mean Opinion Score (MOS). Specifically, for YoN, MCQ, VQA, and CAP tasks, we use 15 general LMMs for inference, namely DeepseekVL [53], InstructBLIP [14], InternVL2 [11], InternLM-XComposer2 [16], LLaVA1.5 [49], LLaVANext [44], LLaVA-OneVision [37], Mantis [32], MiniInternVL [22], MPlugOwl3 [81], Ovis [54], Phi3.5 [1], Pangea [85], Qwen2-VL [75], and Yi1.5 [52]. We set all temperature parameters as 0 to avoid unstable output to ensure all perceptual quality degradations come from distortion. Comprehensively considering popular trends and performance, we implement 5 models for each specific CV tasks, namely SEG: BoxInst [70], ConvNext [77], Mask-RCNN [27], SCNet [73], and RTMDet (segmentation mode) [55]; DET: Dino[89], Mask-RCNN [65], RTMDet (detection mode) [55], ViTDet [34], and Yolo-X [23]; RET: DIME [63], InternVL (retrieval-based tuning) [10], LAVIS (retrieval-based tuning) [36], NAAF [90], and $\mathrm{VSE}++$ [19]. After obtaining scores for these 7 tasks, the evaluation can be divided into five dimensions, one dimension for each first four LMM tasks, and the last three tasks are unified into the last dimension, ensuring 15 subjects in each $(0,1)$ score for MOS. Thus merging all five dimensions into the $(0,5)$ MOS. Note that we normalize the score of each dimension to avoid where one dimension has a higher score while another is lower (e.g. VQA/CAP), which would lead to unfairness in the weight of the overall MOS. + +In addition, considering the rapid development of LMM, we also apply 3 advanced LMMs, GPT4o [2], GPT4o- + +mini [2], InternLM-XComposer2d5 [16]; and 4 early-stage LMMs, QWen1.5-VL [5], VisualGLM [17], Monkey [45], MPlugOwl2 [82] on certain tasks. We compared these results with the above-mentioned data in MPD, to prove the long-term usability of MPD through their consistency. + +# 4. Database Validation + +This section analyzes the preference data characteristics of MPD, to prove the usability of this database, including two perspectives: (1) the mean opinion of all machines; (2) the difference between each machine as a subject. + +Overall: The MOS of MPD and the sub-scores of the seven tasks are illustrated in Figure 3, with their correlation quantified by Spearman Rank-order Correlation Coefficient (SRCC). Notably, the inter-task correlation is relatively low, particularly for VQA and the three tasks that rely on specialized CV models. Given the current state of image processing algorithms for machine vision, each employing different tasks as performance metrics, the universality and robustness of such evaluation are inherently constrained. The data distribution further substantiates this argument, with YoN/MCQ/CAP being more evenly distributed, VQA exhibiting a higher concentration of high scores, and SEG/DET/RET clustering at lower scores. Moreover, the strong correlation between the total score and the other tasks is evident, with all SRCC surpassing 0.5, underscoring the reliability of the MPD, for various downstream tasks. The MOS distribution across 30 types of corruption and 5 strength levels is depicted in Figure 4. Results show for corruption scenarios that appear similar to HVS, the MVS perception mechanism diverges significantly. For + +![](images/4f12ad9cbaa53718cbafceebbae4c914164a862f9f64ee4ad18ed46e2d673540.jpg) +Strength 1 + +![](images/5b11c5e7eec650aee15617ad440f1c9bb290e943aff8f9bd2863db85f73ad7cd.jpg) +Strength 2 + +![](images/04c7b38e9fd0a7ae5802d6da455caae6608bf09d0e953e03999c63dea6459707.jpg) +Strength 3 + +![](images/d85c61f1fe6090e5b519a1b70d3a366d2f9eeb11e54ced17648f6de9039ba905.jpg) +Strength +Strength 5 + +![](images/3ed53d883119b1d3dc2dfba449f9a515f21358656635a7e9ad312172a021c09c.jpg) + +![](images/6dbf5dbe7941a6a52246d6add2bd797c351bd85a158f7c01cf24da41b779df44.jpg) +Figure 5. Annotations from different LMMs subjects in MPD under 5 corruption strengths, including $\mathcal{P}$ content types and 4 tasks. +Human (0.76) + +![](images/a2077a8a6ae67ae80c8a8e5ea4c34aa8fc825150b8990d843f6dfe6c5c65827e.jpg) +MPD (0.62) + +![](images/1226e1f518c84a851518fec60497b2036e020ed06340953e21b336885e151261.jpg) +YoN (0.50) +Figure 6. Correlation matrix of the proposed MPD in each dimension. A darker color denotes a higher SRCC, with mean SRCC attached below each matrix. The index of each matrix is arranged in the $\uparrow$ , $\rightarrow$ directions, in the same order as Section 3.3. The correlation between human subjects is greater than machines, and the correlation of the overall MOS is greater than the score in each dimension. + +![](images/f31c9e67faf073c4d6d83c985ad4d30df5f6383f47800d69ae3c4f71fcea8fd5.jpg) +MCQ (0.55) + +![](images/b4bba999e196b1fd78bce3175088bc56aa8e86744b26a67cbff4bc268cdd3482.jpg) +VQA (0.37) + +![](images/065a8dd2b564dee173e7eb415ac9212b4d34a113c7b6e7b27a58916c7f032769.jpg) +CAP (0.39) + +![](images/1cf7815dedc0891a3bed14374d3c1aa24ef4e9583ed20a6635c5937a87fb5276.jpg) +Other (0.53) + +instance, the machine is highly sensitive to 'Lens blur' but rarely affected by 'Mean brighten'. A typical example is the four Macro Block-related corruptions, which are nearly indistinguishable for HVS but exhibit significantly lower perception quality for MVS when encountering 'Block exchange' compared to the other three. Furthermore, the impact of distortion intensity on the MVS perceptual quality varies markedly. For example, as the intensity level increases, the 'Lab saturation' quality diminishes, while the 'HSV saturation' quality remains unchanged. These findings underscore the differences between MVS and HVS, reinforcing the significance of establishing such a unified quality representation in MPD. + +Each machine: We visualize the benchmark performance of six typical LMMs (as space limited) under different corruption strengths, as shown in Figure 5. For different content types, the scores of various machines on NSI/SCI/AIGI dimensions do not vary significantly. For different tasks, advanced models such as InternVL2 excel in dealing with distorted images in YoN, while traditional models like LLaVA1.5 can provide relatively stable results in CAP. As the corruption strength increases by one level, the scores of each dimension decrease by 0.01-0.02. Finally, for all machine subjects, we explored their internal differences in Figure 6. We manually conducted human annotation for 500 instances (the same 15 subjects) and compared the preference differences between humans and machines. The index of each matrix (i.e., the machine it represents) is arranged in the $\uparrow$ and $\rightarrow$ directions, in the same order as in Section 3.3. Overall, the SRCC between humans is 0.76, while that of machines is only 0.62, indicating that the divergence in machine image preferences is greater than that of humans. Further exploration of the preference scores + +for each dimension shows that the relatedness of machines in VQA and CAP tasks is the lowest. Additionally, there are a few subjects whose assessments differ significantly from other machines $^2$ , such as InstructBlip in YoN, Mantis in MCQ, Pangea in CAP, and DIME/NAAF in RET. This is often due to special token outputs or multilingual mechanisms, which lead to insufficient robustness, resulting in a higher proportion of errors in distorted images compared to other models, thus leading to inconsistent results. In summary, the larger preference differences make IQA for MVS a more difficult problem than traditional IQA, but also a more valuable and promising research topic. + +# 5. Experiment + +# 5.1. Experiment Setups + +We randomly partitioned the MPD into the train/val set, with 24,000 and 6,000 reference and distorted image pairs, respectively, according to an 8:2 ratio. For MVS, Figure 4 illustrates significant variations in the perceptual quality of different corruptions. Given that machines exhibit a larger Just Noticeable Difference (JND) than humans, our corruption strength exceeds that of previous IQA databases for humans3. To conduct a realistic and reliable evaluation of machine preference, we designate the val set as 'Severe Distortion' and select a subset of 'Mild Distortion' to represent the corruption strength in the real world. The specific splitting + +Table 2. Using IQA metrics to predict machine preference, including baseline, FR, and NR algorithm, human visual system-based metrics are marked in $\bullet$ . The first row represents two distortion intensities and three image content subsets, and the second row represents five scoring dimensions. [Keys: Best; Second best; Baseline (BL); Inferior to the baseline.] + +
TypeMetricHVSSevere DistortionMild DistortionNSISCIAIGI
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
BLPSNR0.38730.26720.46640.30970.21270.61970.41160.28200.49450.31350.21460.33290.52250.36840.5985
SSIM[76]0.59690.41570.55370.26720.18280.28180.62880.44660.59320.54150.37480.53930.59800.42780.5687
FRAHIQ[33]0.80610.61530.81830.53040.36850.51110.84470.66010.85020.79620.60400.84700.79960.60820.7966
CKDN[95]✔0.74840.55600.71740.06040.04050.10610.81000.59880.76700.72030.51710.76610.69920.51020.6869
DISTS[15]0.73810.54430.73720.46160.32320.44860.79120.60600.80460.72750.53840.74290.72190.53650.7292
LPIPS[87]✔0.62530.44550.56010.00790.00920.00560.69650.51320.64780.56180.39200.56860.64290.45180.5481
PieAPP[62]✔0.47370.01630.23770.30300.21000.23490.45360.11870.27290.51730.09540.21350.46140.20090.3168
TOPIQ-FR[9]✔0.71780.52520.71820.42480.29440.48410.75060.56140.75850.72430.53220.71860.69430.50630.7061
NRARNIQA[3]0.83370.64260.81120.23880.16080.27540.87040.66900.84780.82200.62170.84200.84490.65780.7954
CLIPQA[74]✔0.64980.46620.59010.18220.12100.17970.76700.57710.71410.66900.47510.62150.68770.49340.5977
DBCNN[88]✔0.56520.39080.56960.13340.08940.15940.59780.41810.62190.56920.39740.60920.54220.38100.5613
HyperIQA[68]0.81050.61890.80120.53880.37740.47120.89080.70740.86910.82730.64030.86010.82120.62420.8076
NIMA[69]0.75620.56570.78580.45700.31710.43420.81530.62030.82240.78000.58810.80920.78460.59530.7622
TOPIQ-NR[9]✔0.65620.46160.63580.19610.13180.19890.70110.48230.66130.66450.46240.63530.54610.34220.5249
TypeMetricHVSYoNMCQVQACAPOthers
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
BLPSNR0.27530.18920.34670.32730.22390.43720.20670.13780.21200.43190.29930.49890.37980.26170.4683
SSIM[76]0.39600.27280.38040.50890.34990.48800.29600.19880.27560.55500.38610.54470.50720.34910.4766
FRAHIQ[33]0.60170.42830.62120.73140.53740.74660.42440.29240.45830.72200.53410.75460.61840.43960.6311
CKDN[95]✔0.54910.37990.54820.68660.48630.65960.40450.27630.39000.70440.51360.69910.54340.38220.5313
DISTS[15]0.52670.36950.53940.64860.46250.65490.35120.23950.36760.66490.48330.68120.62030.44190.6332
LPIPS[87]✔0.46770.32520.42450.55170.39120.51690.32160.21690.29350.56200.39960.54730.49270.33870.4242
PieAPP[62]✔0.33080.01730.16630.43100.19000.20870.23850.01320.10320.42200.01520.21360.37350.02090.1747
TOPIQ-FR[9]✔0.50110.34800.51180.60380.42320.61650.37000.25080.38990.63270.45450.64790.61990.43830.6281
NRARNIQA[3]0.62280.43940.61720.73420.53510.72420.44090.30260.44540.76560.56930.76640.65600.45660.6309
CLIPQA[74]✔0.46130.31670.43860.50900.35560.51170.29710.20280.32040.63790.45510.56730.57510.40480.4747
DBCNN[88]✔0.37980.25670.39690.47930.32580.49510.25560.17190.27830.52400.36330.53880.50850.34650.5093
HyperIQA[68]0.60100.42780.60660.70660.51550.71460.43220.29960.44190.71530.52340.74430.64610.45980.6329
NIMA[69]0.55440.39040.58870.68930.50020.70360.44600.30740.47750.63130.45310.68420.58100.41060.6181
TOPIQ-NR[9]✔0.49390.33580.47970.54690.37130.53700.35780.24370.37870.51120.33750.50890.57690.39340.5564
+ +mechanism is detailed in the Appendix. To benchmark the performance of quality metrics, three global indicators were employed: SRCC, Kendall Rank-order Correlation Coefficient (KRCC), and Pearson Linear Correlation Coefficient (PLCC), to evaluate the consistency between the objective quality score and the subjective MOS. Among these, SRCC and KRCC represent the prediction monotonicity, while PLCC measures the accuracy. + +12 mainstream IQA metrics are implemented for comparison, which have all shown satisfying results in previous human-centric IQA tasks, including 6 FR metrics: AHIQ [33], CKDN [95], DISTS [15], LPIPS [87], PieAPP [62], TOPIQ-FR [9]; and 6 NR metrics: ARNIQA [3], CLIPIQA [74], DBCNN [88], HyperIAQ [68], NIMA [69], TOPIQN [9]. Some methods specifically model the HVS, such as DBCNN, which has a dedicated stage for simulating human perception; others rely on the remarkable fitting ability of large-scale parameters, like the ViT in AHIQ. We have marked all HVS-based metrics as $\bullet$ . Moreover, two naive quality indicators, PSNR and SSIM [76] are set as the baseline. We train the above 12 metrics on MPD with the learning rate as $10^{-5}$ for 50 epochs, under the default settings in pyiqa [9] toolbox and evaluate the performance on + +the two subsets above, three content types, and five scoring dimensions. We collect MPD preference annotations on $16 \times$ NVIDIA A800 SXM4 80GB GPUs, and then conduct IQA training/validation on one GPU above. + +# 5.2. Experiment Result and Discussion + +Table 2 presents the performance of current advanced IQA models on the MPD. Overall, regardless of FR or NR, no metric is highly consistent with machine preference, although they are all good at predicting human preference. An interesting finding is that the more accurately and completely a method models the HVS, the more difficult it is to evaluate the perceptual quality of MVS. Considering that most of the current IQA works emphasize human vision, while machines have already surpassed humans to become the main consumer of visual signals, future IQA needs to find a way to balance these two goals. + +For the two distortion subsets, fitting-based methods such as AHIQ/ARNIQA/HyperIQA can achieve SRCC/PLCC values exceeding 0.8 on Severe Distortion. However, this is not a challenging task. Under the same distortion conditions, they have even exceeded 0.95 on the HVS dataset [33, 61]. The performance of Mild Distortion is even worse, and no method can even outperform the + +Table 3. Cross validation between human and machine preference database. For IQA metrics, training on machine preference data can reach a good result in evaluating human preference; but training on human preference is unacceptable for MPD. [Keys: Best; Lower than 0.4.] + +
Val-setKADID-10KTID2013MPD (Mild)MPD (Severe)
MetricSRCC↑PLCC↑SRCC↑PLCC↑SRCC↑PLCC↑SRCC↑PLCC↑
AHIQ[33]0.79500.77320.76510.78840.15140.18460.47530.3518
CKDN[95]0.85990.72820.69890.60290.17130.20550.34940.3228
DISTS[15]0.82390.81900.72310.77590.41050.40770.46160.4486
LPIPS[87]0.61840.54510.70540.63260.23910.31180.46830.3711
PieAPP[62]0.47830.34680.68320.45430.12850.06780.17910.1235
TOPIQ-FR[9]0.67680.68440.77820.74730.51120.54950.62230.6685
ARNIQA[3]0.76840.70230.77010.72050.48290.45820.55970.4849
CLIPIQA[62]0.62860.57290.52370.39100.13670.14780.32010.3203
DBCNN[88]0.51750.53510.44860.57860.13590.15920.40240.3877
HyperIQA[68]0.66500.65470.70240.73130.08200.10740.45190.4521
NIMA[69]0.47950.49150.56940.64230.16290.17750.47090.4845
TOPIQ-NR[9]0.52320.50770.54550.61740.23880.27540.48290.4582
+ +simple PSNR, which proves that there is still a lot of room for optimization of IQA for machine preference. For the three contents, after verifying them separately, the performance has slightly increased, which is caused by the internal quality differences of the contents. In general, predicting NSI quality is relatively simple, while SCI and AIGI are more difficult. This may be due to the fact that the initial weights of the IQA model have rich NSI priors. For the five dimensions, we can see that the results of MCQ and CAP are most consistent with machine preference; YoN and Others (three common CV tasks) are second; VQA is the worst. This is because VQA problems are usually changes in image details (such as the color of a button on a piece of clothing), which are more difficult for machines to detect than other tasks. In short, IQA for Machine preference still has great research potential, and it is necessary to develop more accurate and effective evaluation mechanisms in the future. + +# 5.3. Cross Database Validation + +To further analyze the differences between MVS and HVS, we conducted cross-validation using MPD and two of the most commonly used human-oriented datasets, TID2013 [61] and KADID-10K [47], employing the same training mechanism as in the previous chapter. We observed that after training on MPD, most IQA metrics still possess a certain ability to predict human preferences, with SRCC values around 0.7 to 0.8; conversely, after training on the other two human-centric databases, IQA metrics struggle to predict machine preferences, with results on MPD not even reaching 0.4. Machine preference data will not affect human tasks, but the opposite will. This finding holds certain guiding significance for future IQA, particularly for a human-machine-friendly paradigm. + +# 5.4. Long-term Availability Validation + +The machine subjects in MPD are categorized into two groups: common CV models and LMMs. The former mod + +![](images/16331855e00e8ac7e144badfc6023cbe011b025c5ff62f43c9c61ac100922c21.jpg) +Figure 7. The long-term usability of MPD. The preference scores of MPD are consistent with four early-stage LMMs and three advanced LMMs (marked as Traditional/Advanced Models in the background). Their high SRCC proves that MPD can still characterize the perceptual quality of LMMs in the near future. + +els, such as the separation/detection benchmark method, have been largely finalized and generally recognized. In contrast, LMMs are still under rapid evolvement. It is conceivable that MPD may not be able to capture the LMMs that will emerge in the future, or with the update of MPD, the preference of previous LMMs may be sacrificed. Thus, we include 3 advanced and 4 early-stage LMMs as Sec 3.3 listed, integrating them into two databases similar to MPD. Given the inconsistencies in LMM preference on VQA and CAP tasks, as illustrated in Figure 6, we have verified the consistency of MPD ground truth labels with advanced/early-stage labels. For each image, the relationship between these two labels is depicted in Figure 7. The findings indicate that MPD exhibits a high correlation with early models on VQA and with advanced models on CAP. Generally, the SRCC is around 0.6, which is considered acceptable. Considering the GPT-4o surpasses open-source models, we anticipate that MPD is able to characterize LMMs preferences for images for at least a year. In the future, We will continue to update MPD by periodicity introducing the latest LMMs to avoid this limitation. + +# 6. Conclusion + +This paper extends the application of IQA from human to machine vision for the first time and establishes the first large-scale, fine-grained, multi-dimensional quality database for machines, namely MPD. We select 15 LMMs and 15 common CV models as subjects, show them $30\mathrm{k}$ reference/distorted image pairs, and summarize their performance on 7 downstream tasks as quality scores. In-depth analysis of the MPD content and IQA benchmark experiments have proved the huge gap between machine and human visual systems, and current human-centric IQA methods cannot accurately predict machine preferences. As the visual data consumption volume of machines has already surpassed humans, we hope that MPD can inspire effective IQA metrics to serve the upcoming machine-centric era. + +# 7. Acknowledgment + +The work was supported by the National Natural Science Foundation of China under Grant 62225112, 62301310, 623B2073, 62471290; and in part by Sichuan Science and Technology Program under Grant 2024NSFSC1426. + +# References + +[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 5 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5 +[3] Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. Arniqa: Learning distortion manifold for image quality assessment. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 188-197, 2024. 7, 8 +[4] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 382-398. Springer, 2016. 4 +[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 5 +[6] Johannes Balle, Philip A Chou, David Minnen, Saurabh Singh, Nick Johnston, Eirikur Agustsson, Sung Jin Hwang, and George Toderici. Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing, 15(2):339-353, 2020. 2, 3 +[7] Min Cao, Shiping Li, Juntao Li, Liqiang Nie, and Min Zhang. Image-text retrieval: A survey on recent research and development. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 5410-5417. International Joint Conferences on Artificial Intelligence Organization, 2022. Survey Track. 4 +[8] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 3 +[9] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 33:2404-2418, 2024. 7, 8 +[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing + +the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5 +[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24185–24198, 2024. 5 +[12] Jinyoung Choi and Bohyung Han. Task-aware quantization network forJPEG image compression. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 309-324. Springer, 2020. 3 +[13] U Cisco. Cisco annual internet report (2018-2023) white paper. Cisco: San Jose, CA, USA, 10(1):1-35, 2020. 1 +[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 5 +[15] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 7, 8 +[16] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Ji-aqi Wang. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 5 +[17] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, 2022. 5 +[18] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, Dahua Lin, and Kai Chen. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models, 2024. 4 +[19] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 5 +[20] Wen Gao, Siwei Ma, Lingyu Duan, Yonghong Tian, Peiyin Xing, Yaowei Wang, Shanshe Wang, Huizhu Jia, and Tiejun Huang. Digital retina: A way to make the city brain more efficient by visual coding. IEEE Transactions on Circuits and Systems for Video Technology, 31(11):4147-4161, 2021. 1 +[21] Yixuan Gao, Yuqin Cao, Tengchuan Kou, Wei Sun, Yunlong Dong, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Vdpve: Vqa dataset for perceptual video enhancement. In Proceedings of the IEEE/CVF Conference on Computer + +Vision and Pattern Recognition (CVPR) Workshops, pages 1474-1483, 2023. 2 +[22] Zhangwei Gao, Zhe Chen, Erfei Cui, Yiming Ren, Weiyun Wang, Jinguo Zhu, Hao Tian, Shenglong Ye, Junjun He, Xizhou Zhu, Lewei Lu, Tong Lu, Yu Qiao, Jifeng Dai, and Wenhai Wang. Mini-internvl: A flexible-transfer pocket multimodal model with 55 +[23] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 5 +[24] Jinjin Gu, Haoming Cai, Chao Dong, Jimmy S Ren, Radu Timofte, Yuan Gong, Shanshan Lao, Shuwei Shi, Jiahao Wang, Sidi Yang, et al. Ntire 2022 challenge on perceptual image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 951-967, 2022. 2 +[25] Yatharth Gupta, Vishnu V. Jaddipal, Harish Prabhala, Sayak Paul, and Patrick Von Platen. Progressive knowledge distillation of stable diffusion x1 using layer level loss, 2024. 3 +[26] Shijie Hao, Yuan Zhou, and Yanrong Guo. A brief survey on semantic segmentation with deep learning. Neurocomputing, 406:302-321, 2020. 4 +[27] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 5 +[28] David Holz. Midjourney. https://www.midjourney.com, 2023. 3 +[29] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 2 +[30] Yunzhong Hou, Liang Zheng, and Stephen Gould. Learning to structure an image with few colors. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 10116-10125, 2020. 3 +[31] Zhimeng Huang, Chuanmin Jia, Shanshe Wang, and Siwei Ma. Visual analysis motivated rate-distortion model for image coding. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 3 +[32] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning, 2024. 5 +[33] Shanshan Lao, Yuan Gong, Shuwei Shi, Sidi Yang, Tianhe Wu, Jiahao Wang, Weihao Xia, and Yujiu Yang. Attentionshelp cnns see better: Attention-based hybrid image quality assessment network, 2022. 7, 8 +[34] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 3, 5 +[35] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiqa-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2 +[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with + +frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5 +[37] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024.5 +[38] Chunyi Li, Tengchuan Kou, Yixuan Gao, Yuqin Cao, Wei Sun, Zicheng Zhang, Yingjie Zhou, Zhichao Zhang, Weixia Zhang, Haoning Wu, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Aigia-20k: A large database for a-generated image quality assessment, 2024. 2 +[39] Chunyi Li, Guo Lu, Donghui Feng, Haoning Wu, Zicheng Zhang, Xiaohong Liu, Guangtao Zhai, Weisi Lin, and Wenjun Zhang. Misc: Ultra-low bitrate image semantic compression driven by large multimodal model, 2024. 1 +[40] Chunyi Li, Haoning Wu, Hongkun Hao, Zicheng Zhang, Tengchuan Kou, Chaofeng Chen, Lei Bai, Xiaohong Liu, Weisi Lin, and Guangtao Zhai. G-refine: A general quality refiner for text-to-image generation. In Proceedings of the 32nd ACM International Conference on Multimedia, page 7375-7384, New York, NY, USA, 2024. Association for Computing Machinery. +[41] Chunyi Li, Haoning Wu, Zicheng Zhang, Hongkun Hao, Kaiwei Zhang, Lei Bai, Xiaohong Liu, Xiongkuo Min, Weisi Lin, and Guangtao Zhai. Q-refine: A perceptual quality refiner for ai-generated image. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2024. 1 +[42] Chunyi Li, Xiele Wu, Haoning Wu, Donghui Feng, Zicheng Zhang, Guo Lu, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Cmc-bench: Towards a new paradigm of visual signal compression, 2024. 3 +[43] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024. 3 +[44] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 5 +[45] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 5 +[46] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3 +[47] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Tenth International Conference on Quality of Multimedia Experience (QoMEX), pages 1-3. IEEE, 2019. 2, 6, 8 + +[48] Risheng Liu, Long Ma, Tengyu Ma, Xin Fan, and Zhongxuan Luo. Learning with nested scene modeling and cooperative architecture search for low-light vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5953-5969, 2022. 3 +[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 5 +[50] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 4 +[51] Xiaohong Liu, Xiongkuo Min, Guangtao Zhai, Chunyi Li, Tengchuan Kou, et al. Ntire 2024 quality assessment of ai-generated content challenge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 6337-6362, 2024. 2 +[52] Yuan Liu, Le Tian, Xiao Zhou, and Jie Zhou. Rethinking overlooked aspects in vision-language models. arXiv preprint arXiv:22405.11850, 2024. 5 +[53] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Yaofeng Sun, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 5 +[54] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model, 2024. 5 +[55] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 5 +[56] Xiongkuo Min, Kede Ma, Ke Gu, Guangtao Zhai, Zhou Wang, and Weisi Lin. Unified blind quality assessment of compressed natural, graphic, and screen content images. IEEE Transactions on Image Processing, 26(11):5462-5474, 2017. 2, 3 +[57] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE transactions on Image Processing, 20 (12):3350-3364, 2011. 2 +[58] Zhangkai Ni, Lin Ma, Huanqiang Zeng, Jing Chen, Canhui Cai, and Kai-Kuang Ma. Esim: Edge similarity for screen content image quality assessment. IEEE Transactions on Image Processing, 26(10):4818-4831, 2017. 2 +[59] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 4 +[60] Neel Patwa, Nilesh Ahuja, Srinivasa Somayazulu, Omesh Tickoo, Srenivas Varadarajan, and Shashidhar Koolagudi. Semantic-preserving image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 1281-1285. IEEE, 2020. 3 + +[61] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, et al. Image database tid2013: Peculiarities, results and perspectives. Signal processing: Image communication, 30:57-77, 2015. 2, 7, 8 +[62] Ekta Prashnani, Hong Cai, Yasamin Mostofi, and Pradeep Sen. Pieapp: Perceptual image-error assessment through pairwise preference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1808-1817, 2018. 7, 8 +[63] Leigang Qu, Meng Liu, Jianlong Wu, Zan Gao, and Liqiang Nie. Dynamic modality interaction modeling for image-text retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1104-1113, 2021. 5 +[64] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 3 +[65] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 3, 5 +[66] Robin Rombach, Andreas Blattmann, and Björn Ommer. Text-guided synthesis of artistic images with retrieval-augmented diffusion models, 2022. 3 +[67] Chengyao Shen and Qi Zhao. Webpage saliency. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13, pages 33-46. Springer, 2014. 3 +[68] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2020. 7, 8 +[69] Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment. IEEE Transactions on Image Processing, 27(8): 3998-4011, 2018. 7, 8 +[70] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 5 +[71] I. T. Union. Methodology for the subjective assessment of the quality of television pictures. ITU-R Recommendation BT. 500-11, 2002. 5, 6 +[72] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 4 +[73] Thang Vu, Haeyong Kang, and Chang D Yoo. Scnet: Training inference sample consistency for instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2701-2709, 2021. 5 +[74] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Pro + +ceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 7 +[75] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5 +[76] Zhou Wang. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7 +[77] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 5 +[78] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-bench: A benchmark for general-purpose foundation models on low-level vision. arXiv preprint arXiv:2309.14181, 2023. 4 +[79] XAI. grok-1.5v, 2024. 4 +[80] Zhaohui Yang, Yunhe Wang, Chang Xu, Peng Du, Chao Xu, Chunjing Xu, and Qi Tian. Discernible image compression. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1561-1569, 2020. 3 +[81] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024.5 +[82] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13040–13051, 2024. 5 +[83] Masakazu Yoshimura, Junji Otsuka, Atsushi Irie, and Takeshi Ohashi. Dynamicisp: dynamically controlled image signal processor for image recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12866-12876, 2023. 3 +[84] Ke Yu, Zexian Li, Yue Peng, Chen Change Loy, and Jinwei Gu. Reconfigisp: Reconfigurable camera image processing pipeline. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4248-4257, 2021. 3 +[85] Xiang Yue, Yueqi Song, Akari Asai, Seungone Kim, Jean de Dieu Nyandwi, Simran Khanuja, Anjali Kantharuban, Lintang Sutawika, Sathyanarayanan Ramamoorthy, and Graham Neubig. Pangea: A fully open multilingual multimodal llm for 39 languages, 2024. 5 +[86] Jing Zhang and Dacheng Tao. Empowering things with intelligence: a survey of the progress, challenges, and opportunities in artificial intelligence of things. IEEE Internet of Things Journal, 8(10):7789-7817, 2020. 1 +[87] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of + +deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7, 8 +[88] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2018. 7, 8 +[89] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 5 +[90] Kun Zhang, Zhendong Mao, Quan Wang, and Yongdong Zhang. Negative-aware attention framework for image-text matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15661-15670, 2022. 5 +[91] Zicheng Zhang, Chunyi Li, Wei Sun, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. A perceptual quality assessment exploration for aigc images. In IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 440-445, 2023. 2 +[92] Zicheng Zhang, Wei Sun, Yingjie Zhou, Jun Jia, Zhichao Zhang, Jing Liu, Xiongkuo Min, and Guangtao Zhai. Subjective and objective quality assessment for in-the-wild computer graphics images. ACM Transactions on Multimedia Computing, Communications and Applications, 20(4):1-22, 2023. 3 +[93] Qi Zhang, Shanshe Wang, Xinfeng Zhang, Chuanmin Jia, Zhao Wang, Siwei Ma, and Wen Gao. Perceptual video coding for machines via satisfied machine ratio modeling, 2024. 2 +[94] Qi Zhang, Shanshe Wang, Xinfeng Zhang, Siwei Ma, Jingshan Pan, and Wen Gao. Predicting satisfied user and machine ratio for compressed images: A unified approach, 2024. 2 +[95] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10242-10251, 2021. 7, 8 +[96] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 4 \ No newline at end of file diff --git a/2025/Image Quality Assessment_ From Human to Machine Preference/images.zip b/2025/Image Quality Assessment_ From Human to Machine Preference/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b4787d76e4d5d67b81b4c4deb7c3a93861efd20d --- /dev/null +++ b/2025/Image Quality Assessment_ From Human to Machine Preference/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:454093551a3fd1a94f59cb81e329c0958dafb95ce92a0575fab755bc303c2fd5 +size 1173467 diff --git a/2025/Image Quality Assessment_ From Human to Machine Preference/layout.json b/2025/Image Quality Assessment_ From Human to Machine Preference/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0490603582542971c83427518a5bffd68c4ab2c8 --- /dev/null +++ b/2025/Image Quality Assessment_ From Human to Machine Preference/layout.json @@ -0,0 +1,10978 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 105, + 103, + 506, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 103, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 121 + ], + "type": "text", + "content": "Image Quality Assessment: From Human to Machine Preference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": "Chunyi Li" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{1,2,4*}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Yuan Tian" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Xiaoyue Ling" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Zicheng Zhang" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Haodong Duan" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Haoning Wu" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Ziheng Jia" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Xiaohong Liu" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Xiongkuo Min" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Guo Lu" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Weisi Lin" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Guangtao Zhai" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": " \nShanghai Jiao Tong University" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Shanghai AI Lab" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Moonshot AI" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "text", + "content": ", Nanyang Technological University" + }, + { + "bbox": [ + 61, + 142, + 561, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 193, + 553, + 365 + ], + "blocks": [ + { + "bbox": [ + 62, + 193, + 553, + 365 + ], + "lines": [ + { + "bbox": [ + 62, + 193, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 193, + 553, + 365 + ], + "type": "image", + "image_path": "6d346aa8b3722dceff3a924054764b758fe8d0bd085c64b4c82ffc951c515063.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 366, + 555, + 400 + ], + "lines": [ + { + "bbox": [ + 55, + 366, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 555, + 400 + ], + "type": "text", + "content": "Figure 1. The significant gap between the well-explored Human Vision System (HVS) and the emerging Machine Vision System (MVS). Their different perception mechanisms leading to images that are subjectively satisfactory to humans not being applicable to the machines downstream tasks such as detection and question answering, vice versa." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 410, + 200, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 410, + 200, + 423 + ], + "spans": [ + { + "bbox": [ + 152, + 410, + 200, + 423 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 436, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 436, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 436, + 297, + 689 + ], + "type": "text", + "content": "Image Quality Assessment (IQA) based on human subjective preferences has undergone extensive research in the past decades. However, with the development of communication protocols, the visual data consumption volume of machines has gradually surpassed that of humans. For machines, the preference depends on downstream tasks such as segmentation and detection, rather than visual appeal. Considering the huge gap between human and machine visual systems, this paper proposes the topic: Image Quality Assessment for Machine Vision for the first time. Specifically, we (1) defined the subjective preferences of machines, including downstream tasks, test models, and evaluation metrics; (2) established the Machine Preference Database (MPD), which contains 2.25M fine-grained annotations and 30k reference/distorted image pair instances; (3) verified the performance of mainstream IQA algorithms on MPD. Experiments show that current IQA metrics are human-centric and cannot accurately characterize machine preferences. We sincerely hope that MPD can promote the evolution of IQA from human to machine preferences. Project page is on: https://github.com/lcyszxdxc/MPD." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 410, + 394, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 410, + 394, + 422 + ], + "spans": [ + { + "bbox": [ + 315, + 410, + 394, + 422 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 434, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 650 + ], + "type": "text", + "content": "The rapid ascend of smart cities [20] and the flourishment of the Internet of Things (IoT) [86] over the past decade have fundamentally transformed the framework of the application end. According to the Cisco [13] white paper, the number of Machine-to-Machine (M2M) connections first exceeded that of Machine-to-Human (M2H) in 2023, reaching 147 billion. Machines have gradually replaced humans and become the primary consumers of image/video data. In the field of image processing, the primary goal is to enhance the quality of processed images, ensuring they conform to human perception. Previously, the performance of a compression/reconstruction/generation algorithm [39-41] is often judged by its visual appeal to humans. This technology, known as Image Quality Assessment (IQA), comprehensively models the Human Visual System (HVS) to predict human subjective preferences. Consequently, with the evolution of application ends, IQA today has to transcend HVS and serve Machine Vision Systems (MVS) more effectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "However, there are significant disparities in the perceptual mechanisms between HVS and MVS. This discrepancy arises because HVS prioritizes the similarity between reference and distorted images, such as texture, structure, and color, whereas machines focus on the consistency of results" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 709, + 140, + 720 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 709, + 140, + 720 + ], + "spans": [ + { + "bbox": [ + 70, + 709, + 140, + 720 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 150, + 709, + 232, + 720 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 709, + 232, + 720 + ], + "spans": [ + { + "bbox": [ + 150, + 709, + 232, + 720 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "type": "text", + "content": "7570" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 97, + 542, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Comparison of MPD with other IQA databases. As the first machine preference-oriented library, MPD has a larger image scale, a wider range of corruption scenarios, and more complete annotation content." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 97, + 542, + 232 + ], + "lines": [ + { + "bbox": [ + 70, + 97, + 542, + 232 + ], + "spans": [ + { + "bbox": [ + 70, + 97, + 542, + 232 + ], + "type": "table", + "html": "
DatabaseImageCorruptionAnnotation
ReferenceDistortedResolutionTypesStrengthNumDimensionLabeling Criteria
LIVE[57]297797685525k1Human preference on NSIs
TID2013[61]253k512245514k1Human preference on NSIs
KADID-10K[47]8110k1k255304k1Human preference on NSIs
CLIC2021[6]5853k1k103484k1Human preference on compressed images
NTIRE2022[24]25029k2884051.13m1Human preference on super-resolution images
KonIQ-10K[29]-10k1k--1.21m5Human preference on NSIs
AGIQA-3K[35]-3k1k--125k2Human preference on AIGIs
NTIRE2024[38]-20k1k--420k2Human preference on AIGIs
MPD (ours)1k30k1k3052.25m5Machine preference on NSIs, SCIs, and AIGIs
", + "image_path": "8598cbef269baa30bcd78548b52fdbb96c4674d82aed4ae20f7bebafd2047a9f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 252, + 294, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 252, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 294, + 348 + ], + "type": "text", + "content": "in downstream tasks. As illustrated in Figure 1, for distortions resulting from factors like encoding/quantization, HVSs perceive a noticeable degradation in quality, yet the performance of tasks like segmentation/detection by machines remains unaffected. Conversely, some minor disturbances that the human eye barely notices can significantly disrupt the output of MVSs, making it difficult to migrate their evaluation algorithms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 350, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 564 + ], + "type": "text", + "content": "What images do machines prefer? This remains an open question. For HVS, over the past two decades, IQA has witnessed significant development. Given the variations in image preferences among individuals, particularly across dimensions such as brightness and chroma, the scoring dimensions, the annotations from multiple subjects, and the integration of these annotations into global subjective preference labels have been well established. Hundreds of fine-grained databases can drive IQA algorithms that closely align with human perception. Conversely, MVS development is much slower, regardless of an end-to-end IQA algorithm, there is even no database that annotates images with an overall preference score of machines. This hinders the comprehensive evaluation of image processing metrics designed for machine vision, limiting the experience of application ends. Considering these issues, we first attempt to generalize the concept of IQA from HVS to MVS. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "content": "- A innovative definition of the subjective preferences of machines. We propose a standardized process for collecting subjective labels from machines. Specifically, we identify the downstream tasks that determine the quality of machine perception, including four multimodal tasks related to Large Multimodal Models (LMMs) and three classic Computer Vision (CV) tasks. By referencing the multi-dimensional scoring mechanism of humans, we comprehensively evaluate the performance of machines in these tasks to obtain ground truth labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "content": "- A large-scale database named the Machine Preference Database (MPD). Using 15 general LMMs and 15 spe" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 252, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 252, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 321, + 252, + 555, + 323 + ], + "type": "text", + "content": "cialized CV models, we collected 2.25 million preference data points on 30,000 reference/distorted image pairs, establishing the first IQA database for machine vision. In particular, we conducted an in-depth analysis of each participating machine to explore the internal patterns of MVS perception and their differences with HVS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 324, + 555, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 324, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 314, + 324, + 555, + 408 + ], + "type": "text", + "content": "- A complex benchmark test for existing IQA algorithms. Experiments on the MPD demonstrate that these algorithms designed for HVS cannot fully model MVS, whose predictions do not match machine ratings. After finetuning based on the MPD, their performance improves slightly, but still cannot characterize machine preferences, especially when facing mild distortion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 420, + 406, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 420, + 406, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 420, + 406, + 433 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 441, + 507, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 507, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 507, + 453 + ], + "type": "text", + "content": "2.1. Image Quality Assessment Database" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 458, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 458, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 312, + 458, + 555, + 615 + ], + "type": "text", + "content": "IQA datasets include Full Reference (FR) and No Reference (NR). FR datasets manually add corruption to reference images, whereas NR directly capture distorted images. Typically, these datasets focus on a single image type, such as Natural Scenes Images (NSIs) [6, 24, 47, 57, 61], Screen Content Images (SCIs) [21, 56, 58], and AI-Generated Images (AIGIs) [35, 38, 51, 91], with quality labels derived solely from human preferences as listed in Table 1. On contrast, in the era where machines surpass humans as application ends, MPD stands out as the first machine preference-oriented library. It offers a larger image scale, a broader range of corruption scenarios, and a more comprehensive annotation context than other IQA databases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 624, + 497, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 497, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 497, + 635 + ], + "type": "text", + "content": "2.2. Machine Vision Downstream Task" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": "In the absence of a unified quality indicator, image processing algorithms, taking compression and restoration metrics for machine vision for example, are often in a fragmented state, with the performance of each algorithm being validated by only one or two downstream tasks. [93, 94] These tasks vary, as do the models used to perform them, mak" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 751, + 314, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 751, + 314, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 751, + 314, + 761 + ], + "type": "text", + "content": "7571" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 555, + 370 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 555, + 370 + ], + "type": "image", + "image_path": "673ed7353ff32666bd066c02e6968121cce7d918dd5ab57864b91b5f7406ca12.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 378, + 555, + 411 + ], + "lines": [ + { + "bbox": [ + 55, + 378, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 555, + 411 + ], + "type": "text", + "content": "Figure 2. Overview of Machine Preference Database (MPD), with 30k large-scale reference/distorted image pairs, meticulously annotated with 2.25M fine-grained preference score from the mean opinion of 15 mainstream LMMs and 15 specific CV models. It provides a multi-dimensional fidelity evaluation of various downstream machine vision tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 425, + 297, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 425, + 297, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 297, + 581 + ], + "type": "text", + "content": "ing the evaluation criteria disparate and incomplete. For compression algorithms, validation indicators include the classification [30, 60], image captioning [12], object detection [80], or semantic segmentation [31] result before/after compression, with each indicator differing and lacking standardization or completeness. Restoration algorithms fare slightly better, with most consistently relying on object detection [48, 83, 84]. Nevertheless, some employ traditional Fast-RCNN [65], while others utilize more advanced models like ViTDet[34]. The variation in model bases further compromises the comprehensiveness and objectivity of the evaluation outcomes. Thus, there is an urgent need for a unified representation of machine preferences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 592, + 189, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 189, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 189, + 604 + ], + "type": "text", + "content": "3. Database Construction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 612, + 266, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 266, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 266, + 624 + ], + "type": "text", + "content": "3.1. Reference & Distorted Image Collection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Similar to the previous IQA database, we first selected 1,000 high-quality reference images and then artificially degraded their quality to obtain 30,000 pairs of reference and distorted images. To comprehensively represent the common content on the internet, we referred to the CMC-Bench [42] and selected the three most common image categories: Natural Scene Images (NSIs), including" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 425, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 520 + ], + "type": "text", + "content": "User/Professional Generated Content (UGC/PGC) [6, 46]; Screen Content Images (SCIs), including webpages [67], games [92], and movies [56]; and AI-Generated Images (AIGIs), generated by six representative text-to-image models [8, 25, 28, 43, 64, 66]. The collection process is shown in Figure 2. Both human viewers and IQA models have verified this image sequence to ensure that they are undistorted as reference images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 522, + 556, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 679 + ], + "type": "text", + "content": "For the distorted images, we modeled the corruption caused by the current communication protocols, which were divided into 30 scenarios. Specifically, these corruptions were classified into seven categories: Blur, various types of unclear image; Luminance, global brightness changes; Chrominance, global color changes; Contrast, global changes in contrast/saturation; Noise, random noise of different distributions; Compression, codec algorithm like JPEG; and Spatial, local pixel-level changes. For each corruption, we defined five levels of intensity, from mild to severe. To fairly explore the impact of each corruption, we artificially control the distortion strength to ensure that quality degradation perceived by the human eye is" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 714 + ], + "type": "text", + "content": "1For the definition of all these 30 corruptions, specific details, and examples, please refer to the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "type": "text", + "content": "7572" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "aligned at the same level. Thus, for each reference image, we randomly selected the intensity to add all the corruptions mentioned above, resulting in 30,000 distorted images. The reference/distorted image pairs will then be annotated by machine subjects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 138, + 202, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 202, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 202, + 149 + ], + "type": "text", + "content": "3.2. Different Task Annotation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 296, + 357 + ], + "type": "text", + "content": "In the past five years, the evaluation of image processing algorithms for machine vision has primarily focused on seven tasks: Regression, Classification, Visual Question Answering (VQA), Caption (CAP), Segmentation (SEG), Detection (DET), and Retrieval (RET). Among these, the first four tasks can be accomplished by general LMMs, whereas the latter three tasks necessitate specialized models due to their unique requirements. Note that the regression and classification tasks in this context are analogous to Yes-or-No (YoN) and Multiple-Choice Questions (MCQ) within LMMs. To this end, we enlisted five experienced human experts to design YoN/VQA/CAP questions and answers for 1,000 reference images. These questions are designed to be challenging, as illustrated in the example provided in Figure 2, especially in the MCQ sets where confusion options are introduced. This approach enables us to accurately measure the degree of quality degradation perceived by the machine." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 359, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 296, + 491 + ], + "type": "text", + "content": "We apply these seven tasks to both the reference and distorted images, measure the similarity of the output results, and derive scores for each of the seven sub-dimensions. Specifically, for " + }, + { + "bbox": [ + 55, + 359, + 296, + 491 + ], + "type": "inline_equation", + "content": "Y_{oN}" + }, + { + "bbox": [ + 55, + 359, + 296, + 491 + ], + "type": "text", + "content": " tasks, we employ a logit-level evaluation approach based on softmax similarity, as previous LMM benchmarks [50, 78] do. This entails LMMs confirming or denying the questions designed for each reference/distorted image pair. The confidence difference in the next token output, 'Yes' or 'No', reflects the difference in the perceptual quality degradation from the perspective of machines, namely a score " + }, + { + "bbox": [ + 55, + 359, + 296, + 491 + ], + "type": "inline_equation", + "content": "S_{Y_{oN}}" + }, + { + "bbox": [ + 55, + 359, + 296, + 491 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 496, + 294, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 496, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 67, + 496, + 294, + 510 + ], + "type": "interline_equation", + "content": "S _ {Y o N} = \\left| \\sigma \\left(P _ {\\text {d i s}} (\\text {Y e s}, \\text {N o})\\right) - \\sigma \\left(P _ {\\text {r e f}} (\\text {Y e s}, \\text {N o})\\right) \\right|, \\tag {1}", + "image_path": "43f1c440545431db1a077f481857591ec322bb89f8f007115acc74eed9f1f9ee.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "inline_equation", + "content": "P_{dis,ref}(\\cdot)" + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "text", + "content": " denotes the output probability of certain logits, normalized by the softmax function " + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "text", + "content": ". MCQ tasks follow a similar evaluation paradigm. We set a correct option and three confusing options (the options are usually no more than 4, see [18, 50, 79]), and the model outputs the probability of each option as a four-element vector. The distance between the two vectors in space can reflect the perceptual quality score " + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "inline_equation", + "content": "S_{MCQ}" + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "text", + "content": " for machines:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 616, + 295, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 616, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 64, + 616, + 295, + 630 + ], + "type": "interline_equation", + "content": "S _ {M C Q} = \\cos \\left(P _ {d i s} (\\mathrm {A}, \\mathrm {B}, \\mathrm {C}, \\mathrm {D}), P _ {r e f} (\\mathrm {A}, \\mathrm {B}, \\mathrm {C}, \\mathrm {D})\\right), \\tag {2}", + "image_path": "819bdf212126070646f8e794bcd323517625fa6591c060397736568c74370907.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": "where the distance is measured by cosine similarity. For " + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "inline_equation", + "content": "VQA" + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": " tasks, we conduct sentence-level evaluation beyond the logits. LMMs will be required to answer a certain question in about 5 words, and the semantic difference between two output sentences will be the score " + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "inline_equation", + "content": "S_{VQA}" + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 701, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 701, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 117, + 701, + 295, + 715 + ], + "type": "interline_equation", + "content": "S _ {V Q A} = \\operatorname {C L I P} \\left(T _ {\\text {d i s}}, T _ {\\text {r e f}}\\right), \\tag {3}", + "image_path": "f0d1e240fd179c0d62fb33bd444a70f948dcdf44e573d063cf2f0898098d7541.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 455, + 186 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 455, + 186 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 455, + 186 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 455, + 186 + ], + "type": "image", + "image_path": "d2bae0fc7fe3d3f3f89cd0af85c937ba37dd6748e5e2e3a17fc4a817173b4015.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 191, + 433, + 202 + ], + "lines": [ + { + "bbox": [ + 343, + 191, + 433, + 202 + ], + "spans": [ + { + "bbox": [ + 343, + 191, + 433, + 202 + ], + "type": "text", + "content": "(a) Correlation matrix" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 460, + 77, + 545, + 186 + ], + "blocks": [ + { + "bbox": [ + 460, + 77, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 460, + 77, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 460, + 77, + 545, + 186 + ], + "type": "image", + "image_path": "e7c910a6015712bc19d0552851df8d958166593387a3df4d2849170438e5d109.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 190, + 546, + 201 + ], + "lines": [ + { + "bbox": [ + 460, + 190, + 546, + 201 + ], + "spans": [ + { + "bbox": [ + 460, + 190, + 546, + 201 + ], + "type": "text", + "content": "(b) Score distribution" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 217, + 554, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 554, + 239 + ], + "type": "text", + "content": "Figure 3. Correlation between the general preference score and the score in seven different downstream tasks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 261, + 554, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 554, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 554, + 345 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 261, + 554, + 345 + ], + "type": "inline_equation", + "content": "T_{dis,ref}" + }, + { + "bbox": [ + 313, + 261, + 554, + 345 + ], + "type": "text", + "content": " represents the whole output text, measured by the CLIP text encoder. " + }, + { + "bbox": [ + 313, + 261, + 554, + 345 + ], + "type": "inline_equation", + "content": "CAP" + }, + { + "bbox": [ + 313, + 261, + 554, + 345 + ], + "type": "text", + "content": " tasks require a longer output, by describing the main subjects, actions, settings, and any notable objects or features of the image in about 40 words. Considering the remarkable text length, we use indicators dedicated to image captioning to analyze the differences between the two long sentences in " + }, + { + "bbox": [ + 313, + 261, + 554, + 345 + ], + "type": "inline_equation", + "content": "S_{CAP}" + }, + { + "bbox": [ + 313, + 261, + 554, + 345 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 371, + 354, + 554, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 354, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 371, + 354, + 554, + 380 + ], + "type": "interline_equation", + "content": "S _ {C A P} = \\sum_ {\\text {e v a l}} \\operatorname {e v a l} \\left(T _ {d i s}, T _ {r e f}\\right), \\tag {4}", + "image_path": "d494d4ec189fd8d5cde949ff4820ae8ca7b964b8bec93a09c90dbdbb19c1ea32.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "text", + "content": "where eval " + }, + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\in" + }, + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "text", + "content": " (BLEU, CIDEr, SPICE) [4, 59, 72], three most commonly-used captioning indicators. Then, for SEG tasks, referring to their common evaluation criteria [26], we calculate the Intersection-over-Union (IoU) ratio in " + }, + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "inline_equation", + "content": "S_{SEG}" + }, + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 376, + 446, + 553, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 446, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 376, + 446, + 553, + 460 + ], + "type": "interline_equation", + "content": "S _ {S E G} = \\operatorname {I o U} \\left(M _ {d i s}, M _ {r e f}\\right), \\tag {5}", + "image_path": "92a131c051a49b6715720ff6540c844775842c5d5be904cb871598d2f6545204.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 469, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 554, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 469, + 554, + 540 + ], + "type": "inline_equation", + "content": "M_{dis,ref}" + }, + { + "bbox": [ + 313, + 469, + 554, + 540 + ], + "type": "text", + "content": " stands for binarized segmentation masks. DET tasks will be more complex, whose results include both spatial information and category labels. Thus, we consider the detection failed when the categories are different, and only further calculate the IoU when they are consistent [96], as described in " + }, + { + "bbox": [ + 313, + 469, + 554, + 540 + ], + "type": "inline_equation", + "content": "S_{DET}" + }, + { + "bbox": [ + 313, + 469, + 554, + 540 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 339, + 551, + 553, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 551, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 339, + 551, + 553, + 564 + ], + "type": "interline_equation", + "content": "S _ {D E T} = \\operatorname {A c c} _ {1} \\left(C _ {d i s}, C _ {r e f}\\right) \\cdot \\operatorname {I o U} \\left(B _ {d i s}, B _ {r e f}\\right), \\tag {6}", + "image_path": "44122ac93f3c00d2234fadb7d35128db6f8fabdbc01f0a6a1c6895204d076b8e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 573, + 554, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 554, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 573, + 554, + 633 + ], + "type": "inline_equation", + "content": "C_{dis,ref}" + }, + { + "bbox": [ + 313, + 573, + 554, + 633 + ], + "type": "text", + "content": " denotes the category sequence ordered by confidence, and " + }, + { + "bbox": [ + 313, + 573, + 554, + 633 + ], + "type": "inline_equation", + "content": "B_{dis,ref}" + }, + { + "bbox": [ + 313, + 573, + 554, + 633 + ], + "type": "text", + "content": " is the detected bounding box. Finally for RET tasks, we merge three common Image-to-Text retrieval indicators [7], namely Top-1/5/10 accuracy, to evaluate the retrieved sequence as " + }, + { + "bbox": [ + 313, + 573, + 554, + 633 + ], + "type": "inline_equation", + "content": "S_{RET}" + }, + { + "bbox": [ + 313, + 573, + 554, + 633 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 362, + 643, + 553, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 643, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 362, + 643, + 553, + 669 + ], + "type": "interline_equation", + "content": "S _ {R E T} = \\sum_ {\\mathrm {i} \\in 1, 5, 1 0} \\operatorname {A c c} _ {\\mathrm {i}} \\left(C _ {d i s}, C _ {r e f}\\right). \\tag {7}", + "image_path": "f4a4cba0f6890b6cea037963f1f5eaa4317d1d35b5b97d6d56c1a3760a356b4a.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "After the above process, we can obtain the preferences of LMMs in the first four dimensions, and the preferences of each specific CV model on its target task." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "type": "text", + "content": "7573" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 140, + 106 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 140, + 106 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 140, + 106 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 140, + 106 + ], + "type": "image", + "image_path": "e8d65818a113bf9db39eef65672b57531532cab7b2e8e46415876d5b43f55b1c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 108, + 126, + 114 + ], + "lines": [ + { + "bbox": [ + 83, + 108, + 126, + 114 + ], + "spans": [ + { + "bbox": [ + 83, + 108, + 126, + 114 + ], + "type": "text", + "content": "Gaussian...filter" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 141, + 72, + 223, + 108 + ], + "blocks": [ + { + "bbox": [ + 141, + 72, + 223, + 108 + ], + "lines": [ + { + "bbox": [ + 141, + 72, + 223, + 108 + ], + "spans": [ + { + "bbox": [ + 141, + 72, + 223, + 108 + ], + "type": "image", + "image_path": "db3f628068b730c7a69e833d98f294b9ef40d8ff65226b69281711602df07bf2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 108, + 203, + 114 + ], + "lines": [ + { + "bbox": [ + 172, + 108, + 203, + 114 + ], + "spans": [ + { + "bbox": [ + 172, + 108, + 203, + 114 + ], + "type": "text", + "content": "Lens...blur" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 224, + 72, + 305, + 108 + ], + "blocks": [ + { + "bbox": [ + 224, + 72, + 305, + 108 + ], + "lines": [ + { + "bbox": [ + 224, + 72, + 305, + 108 + ], + "spans": [ + { + "bbox": [ + 224, + 72, + 305, + 108 + ], + "type": "image", + "image_path": "1c212d6e249c3417c8c55869afbaffc0a924ffc9efcaf1480208a7103bebac37.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 108, + 288, + 114 + ], + "lines": [ + { + "bbox": [ + 251, + 108, + 288, + 114 + ], + "spans": [ + { + "bbox": [ + 251, + 108, + 288, + 114 + ], + "type": "text", + "content": "Motion...blur" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 387, + 108 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 387, + 108 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 387, + 108 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 387, + 108 + ], + "type": "image", + "image_path": "bd6925fd0641dc3a5421fe1ffac387c86e3452f6a9ad279cac5e2c8287428640.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 108, + 374, + 114 + ], + "lines": [ + { + "bbox": [ + 329, + 108, + 374, + 114 + ], + "spans": [ + { + "bbox": [ + 329, + 108, + 374, + 114 + ], + "type": "text", + "content": "Color...diffusion" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 388, + 72, + 469, + 108 + ], + "blocks": [ + { + "bbox": [ + 388, + 72, + 469, + 108 + ], + "lines": [ + { + "bbox": [ + 388, + 72, + 469, + 108 + ], + "spans": [ + { + "bbox": [ + 388, + 72, + 469, + 108 + ], + "type": "image", + "image_path": "75fb09b407bf5b3246d0c0e2b871aa824ec78fb3ed8f886e7a36db5f249742bd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 108, + 450, + 114 + ], + "lines": [ + { + "bbox": [ + 417, + 108, + 450, + 114 + ], + "spans": [ + { + "bbox": [ + 417, + 108, + 450, + 114 + ], + "type": "text", + "content": "Color...shift" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 470, + 72, + 551, + 108 + ], + "blocks": [ + { + "bbox": [ + 470, + 72, + 551, + 108 + ], + "lines": [ + { + "bbox": [ + 470, + 72, + 551, + 108 + ], + "spans": [ + { + "bbox": [ + 470, + 72, + 551, + 108 + ], + "type": "image", + "image_path": "22d3543ebc8db02e13ff242573fe268d52b9a499019c9ec30007e2d4691e093f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 108, + 543, + 115 + ], + "lines": [ + { + "bbox": [ + 490, + 108, + 543, + 115 + ], + "spans": [ + { + "bbox": [ + 490, + 108, + 543, + 115 + ], + "type": "text", + "content": "Color...quantization" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 58, + 116, + 140, + 152 + ], + "blocks": [ + { + "bbox": [ + 58, + 116, + 140, + 152 + ], + "lines": [ + { + "bbox": [ + 58, + 116, + 140, + 152 + ], + "spans": [ + { + "bbox": [ + 58, + 116, + 140, + 152 + ], + "type": "image", + "image_path": "59b4dc0f28f900c096aeb50c01cd4f8718782cb4eed09ebc56c9e59faa57ea79.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 152, + 127, + 159 + ], + "lines": [ + { + "bbox": [ + 82, + 152, + 127, + 159 + ], + "spans": [ + { + "bbox": [ + 82, + 152, + 127, + 159 + ], + "type": "text", + "content": "HSV...saturation" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 141, + 116, + 222, + 152 + ], + "blocks": [ + { + "bbox": [ + 141, + 116, + 222, + 152 + ], + "lines": [ + { + "bbox": [ + 141, + 116, + 222, + 152 + ], + "spans": [ + { + "bbox": [ + 141, + 116, + 222, + 152 + ], + "type": "image", + "image_path": "1257e89ecb593dbf6afb89b344e384087939a7eaa029a06c58634d50f739606d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 152, + 208, + 159 + ], + "lines": [ + { + "bbox": [ + 166, + 152, + 208, + 159 + ], + "spans": [ + { + "bbox": [ + 166, + 152, + 208, + 159 + ], + "type": "text", + "content": "Lab...saturation" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 224, + 116, + 305, + 152 + ], + "blocks": [ + { + "bbox": [ + 224, + 116, + 305, + 152 + ], + "lines": [ + { + "bbox": [ + 224, + 116, + 305, + 152 + ], + "spans": [ + { + "bbox": [ + 224, + 116, + 305, + 152 + ], + "type": "image", + "image_path": "35a7e7be1a60ea062193f22f600003864dc75c9eb912ed965e1a3f0a9c23fd1f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 152, + 296, + 159 + ], + "lines": [ + { + "bbox": [ + 243, + 152, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 243, + 152, + 296, + 159 + ], + "type": "text", + "content": "JP2K...compression" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 306, + 116, + 387, + 152 + ], + "blocks": [ + { + "bbox": [ + 306, + 116, + 387, + 152 + ], + "lines": [ + { + "bbox": [ + 306, + 116, + 387, + 152 + ], + "spans": [ + { + "bbox": [ + 306, + 116, + 387, + 152 + ], + "type": "image", + "image_path": "efc4e2b5af31b2fdc7949355d8cbafb7e0f1247b9740dc049f96061792a2b7d7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 152, + 378, + 159 + ], + "lines": [ + { + "bbox": [ + 325, + 152, + 378, + 159 + ], + "spans": [ + { + "bbox": [ + 325, + 152, + 378, + 159 + ], + "type": "text", + "content": "JPEG...compression" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 388, + 116, + 469, + 152 + ], + "blocks": [ + { + "bbox": [ + 388, + 116, + 469, + 152 + ], + "lines": [ + { + "bbox": [ + 388, + 116, + 469, + 152 + ], + "spans": [ + { + "bbox": [ + 388, + 116, + 469, + 152 + ], + "type": "image", + "image_path": "2f3d141369cb920ba485bbed96c11e515aafdb552d3e36e3ae9198ccf7c6ee4b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 152, + 462, + 159 + ], + "lines": [ + { + "bbox": [ + 406, + 152, + 462, + 159 + ], + "spans": [ + { + "bbox": [ + 406, + 152, + 462, + 159 + ], + "type": "text", + "content": "WEBP...compression" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 470, + 116, + 551, + 152 + ], + "blocks": [ + { + "bbox": [ + 470, + 116, + 551, + 152 + ], + "lines": [ + { + "bbox": [ + 470, + 116, + 551, + 152 + ], + "spans": [ + { + "bbox": [ + 470, + 116, + 551, + 152 + ], + "type": "image", + "image_path": "8ed2cc6cf417c7c524641651d9bf03aebb8c47de42cd9b4c57ff61924ff65e9e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 152, + 534, + 159 + ], + "lines": [ + { + "bbox": [ + 499, + 152, + 534, + 159 + ], + "spans": [ + { + "bbox": [ + 499, + 152, + 534, + 159 + ], + "type": "text", + "content": "White...noise" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 58, + 160, + 140, + 204 + ], + "blocks": [ + { + "bbox": [ + 58, + 160, + 140, + 204 + ], + "lines": [ + { + "bbox": [ + 58, + 160, + 140, + 204 + ], + "spans": [ + { + "bbox": [ + 58, + 160, + 140, + 204 + ], + "type": "image", + "image_path": "88b66cc9a3e3ad8fc3e335d902dc36683ab9662c74ce131e3b2c728d7a4ba7dd.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 141, + 160, + 222, + 204 + ], + "blocks": [ + { + "bbox": [ + 141, + 160, + 222, + 204 + ], + "lines": [ + { + "bbox": [ + 141, + 160, + 222, + 204 + ], + "spans": [ + { + "bbox": [ + 141, + 160, + 222, + 204 + ], + "type": "image", + "image_path": "46d1e2daed45ff8ee19f0338adbe1808d8bb4c7d62a97ab5ed4c7d1fb180ddb2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 224, + 160, + 305, + 204 + ], + "blocks": [ + { + "bbox": [ + 224, + 160, + 305, + 204 + ], + "lines": [ + { + "bbox": [ + 224, + 160, + 305, + 204 + ], + "spans": [ + { + "bbox": [ + 224, + 160, + 305, + 204 + ], + "type": "image", + "image_path": "18bbb6e8651196f9254edffb7e7e13b0593dc09b471110380e0f06a99fc395ed.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 306, + 160, + 387, + 204 + ], + "blocks": [ + { + "bbox": [ + 306, + 160, + 387, + 204 + ], + "lines": [ + { + "bbox": [ + 306, + 160, + 387, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 160, + 387, + 204 + ], + "type": "image", + "image_path": "fc8f917b543e54f7e2078c9d335041f5c37de5627322668de6861dbdd2f33359.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 388, + 160, + 469, + 204 + ], + "blocks": [ + { + "bbox": [ + 388, + 160, + 469, + 204 + ], + "lines": [ + { + "bbox": [ + 388, + 160, + 469, + 204 + ], + "spans": [ + { + "bbox": [ + 388, + 160, + 469, + 204 + ], + "type": "image", + "image_path": "ee80b749df35adb05c60d05e59c0f80059074846d042f29655f28c1e2ac3fc6a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 470, + 160, + 551, + 204 + ], + "blocks": [ + { + "bbox": [ + 470, + 160, + 551, + 204 + ], + "lines": [ + { + "bbox": [ + 470, + 160, + 551, + 204 + ], + "spans": [ + { + "bbox": [ + 470, + 160, + 551, + 204 + ], + "type": "image", + "image_path": "5f13ce1bcbf5f3674405705c9ed11f5d1f241f864ab65cb4493a3ac7388ecf8e.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 58, + 205, + 140, + 248 + ], + "blocks": [ + { + "bbox": [ + 58, + 205, + 140, + 248 + ], + "lines": [ + { + "bbox": [ + 58, + 205, + 140, + 248 + ], + "spans": [ + { + "bbox": [ + 58, + 205, + 140, + 248 + ], + "type": "image", + "image_path": "300f69721f4929300ec94ec5f5b58890bd59b38f4e855e79a9421808136aa2d2.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 141, + 205, + 222, + 248 + ], + "blocks": [ + { + "bbox": [ + 141, + 205, + 222, + 248 + ], + "lines": [ + { + "bbox": [ + 141, + 205, + 222, + 248 + ], + "spans": [ + { + "bbox": [ + 141, + 205, + 222, + 248 + ], + "type": "image", + "image_path": "1f6804de1385154997b507f3fe6ca1762d1a03936cc7ff72d166b3eeabe4e081.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 224, + 205, + 305, + 248 + ], + "blocks": [ + { + "bbox": [ + 224, + 205, + 305, + 248 + ], + "lines": [ + { + "bbox": [ + 224, + 205, + 305, + 248 + ], + "spans": [ + { + "bbox": [ + 224, + 205, + 305, + 248 + ], + "type": "image", + "image_path": "1ec7fc27464c5cd3b12af9ad3bd3fc1ff28905c4066271f8bf046f838aa4314a.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 306, + 205, + 387, + 248 + ], + "blocks": [ + { + "bbox": [ + 306, + 205, + 387, + 248 + ], + "lines": [ + { + "bbox": [ + 306, + 205, + 387, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 387, + 248 + ], + "type": "image", + "image_path": "f54ae830e0f90b05116bd9a1eedde014f1bda51a373d2bad8b209c5522858c59.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 388, + 205, + 469, + 248 + ], + "blocks": [ + { + "bbox": [ + 388, + 205, + 469, + 248 + ], + "lines": [ + { + "bbox": [ + 388, + 205, + 469, + 248 + ], + "spans": [ + { + "bbox": [ + 388, + 205, + 469, + 248 + ], + "type": "image", + "image_path": "af3155f66e6d9ccf282555c882731163e7ba691bd666d4139753f64baef0a55c.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 470, + 205, + 551, + 248 + ], + "blocks": [ + { + "bbox": [ + 470, + 205, + 551, + 248 + ], + "lines": [ + { + "bbox": [ + 470, + 205, + 551, + 248 + ], + "spans": [ + { + "bbox": [ + 470, + 205, + 551, + 248 + ], + "type": "image", + "image_path": "77acb24a647dd17b6f22829f6f62d22bff02f79238b27cb36798741526756d4b.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 58, + 250, + 140, + 293 + ], + "blocks": [ + { + "bbox": [ + 58, + 250, + 140, + 293 + ], + "lines": [ + { + "bbox": [ + 58, + 250, + 140, + 293 + ], + "spans": [ + { + "bbox": [ + 58, + 250, + 140, + 293 + ], + "type": "image", + "image_path": "33c98ebaee779c2ba8f22326c350fc20bef7b2eb6de7ac20a837919011a33ad5.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 304, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 304, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 555, + 327 + ], + "type": "text", + "content": "Figure 4. MOS score of MPD, visualized in 30 corruption subsets. Different color denotes corruption strength Level 1-Level 2-Level 3-Level 4-Level 5. Results show the sensitivity of the machines to each corruption varied significantly." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 141, + 250, + 222, + 293 + ], + "blocks": [ + { + "bbox": [ + 141, + 250, + 222, + 293 + ], + "lines": [ + { + "bbox": [ + 141, + 250, + 222, + 293 + ], + "spans": [ + { + "bbox": [ + 141, + 250, + 222, + 293 + ], + "type": "image", + "image_path": "5027069540ea3bccbeb6aa54dde68289b5297373efb222dd005c36037a48aa6e.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 224, + 250, + 305, + 293 + ], + "blocks": [ + { + "bbox": [ + 224, + 250, + 305, + 293 + ], + "lines": [ + { + "bbox": [ + 224, + 250, + 305, + 293 + ], + "spans": [ + { + "bbox": [ + 224, + 250, + 305, + 293 + ], + "type": "image", + "image_path": "b3513eb939171d21592e79de294567b69057083d38c250082d139c15e01748a2.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 306, + 250, + 387, + 293 + ], + "blocks": [ + { + "bbox": [ + 306, + 250, + 387, + 293 + ], + "lines": [ + { + "bbox": [ + 306, + 250, + 387, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 250, + 387, + 293 + ], + "type": "image", + "image_path": "0c4211e552dcb5536a75c26c01be1e2c7f939c858c6776df8b06538a6803c345.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 388, + 250, + 469, + 293 + ], + "blocks": [ + { + "bbox": [ + 388, + 250, + 469, + 293 + ], + "lines": [ + { + "bbox": [ + 388, + 250, + 469, + 293 + ], + "spans": [ + { + "bbox": [ + 388, + 250, + 469, + 293 + ], + "type": "image", + "image_path": "e73fa53bc956bc2ceac7b635b046354e1742cc5cdc74a388c8f9d7411dad0655.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 470, + 250, + 551, + 293 + ], + "blocks": [ + { + "bbox": [ + 470, + 250, + 551, + 293 + ], + "lines": [ + { + "bbox": [ + 470, + 250, + 551, + 293 + ], + "spans": [ + { + "bbox": [ + 470, + 250, + 551, + 293 + ], + "type": "image", + "image_path": "4b8229b3b26febf67320c8bb388a72ad7b5cc44843f5f1bdb41ed485a79c43b0.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "bbox": [ + 55, + 335, + 208, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 208, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 208, + 347 + ], + "type": "text", + "content": "3.3. Preference Score Collection" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 55, + 354, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 688 + ], + "type": "text", + "content": "Following the ITU human subjective annotation standards [71], we adopt the same settings to obtain the machine Mean Opinion Score (MOS). Specifically, for YoN, MCQ, VQA, and CAP tasks, we use 15 general LMMs for inference, namely DeepseekVL [53], InstructBLIP [14], InternVL2 [11], InternLM-XComposer2 [16], LLaVA1.5 [49], LLaVANext [44], LLaVA-OneVision [37], Mantis [32], MiniInternVL [22], MPlugOwl3 [81], Ovis [54], Phi3.5 [1], Pangea [85], Qwen2-VL [75], and Yi1.5 [52]. We set all temperature parameters as 0 to avoid unstable output to ensure all perceptual quality degradations come from distortion. Comprehensively considering popular trends and performance, we implement 5 models for each specific CV tasks, namely SEG: BoxInst [70], ConvNext [77], Mask-RCNN [27], SCNet [73], and RTMDet (segmentation mode) [55]; DET: Dino[89], Mask-RCNN [65], RTMDet (detection mode) [55], ViTDet [34], and Yolo-X [23]; RET: DIME [63], InternVL (retrieval-based tuning) [10], LAVIS (retrieval-based tuning) [36], NAAF [90], and " + }, + { + "bbox": [ + 55, + 354, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{VSE}++" + }, + { + "bbox": [ + 55, + 354, + 296, + 688 + ], + "type": "text", + "content": " [19]. After obtaining scores for these 7 tasks, the evaluation can be divided into five dimensions, one dimension for each first four LMM tasks, and the last three tasks are unified into the last dimension, ensuring 15 subjects in each " + }, + { + "bbox": [ + 55, + 354, + 296, + 688 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 55, + 354, + 296, + 688 + ], + "type": "text", + "content": " score for MOS. Thus merging all five dimensions into the " + }, + { + "bbox": [ + 55, + 354, + 296, + 688 + ], + "type": "inline_equation", + "content": "(0,5)" + }, + { + "bbox": [ + 55, + 354, + 296, + 688 + ], + "type": "text", + "content": " MOS. Note that we normalize the score of each dimension to avoid where one dimension has a higher score while another is lower (e.g. VQA/CAP), which would lead to unfairness in the weight of the overall MOS." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "In addition, considering the rapid development of LMM, we also apply 3 advanced LMMs, GPT4o [2], GPT4o-" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 313, + 336, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 396 + ], + "type": "text", + "content": "mini [2], InternLM-XComposer2d5 [16]; and 4 early-stage LMMs, QWen1.5-VL [5], VisualGLM [17], Monkey [45], MPlugOwl2 [82] on certain tasks. We compared these results with the above-mentioned data in MPD, to prove the long-term usability of MPD through their consistency." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 406, + 432, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 432, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 432, + 418 + ], + "type": "text", + "content": "4. Database Validation" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 426, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 474 + ], + "type": "text", + "content": "This section analyzes the preference data characteristics of MPD, to prove the usability of this database, including two perspectives: (1) the mean opinion of all machines; (2) the difference between each machine as a subject." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 475, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 713 + ], + "type": "text", + "content": "Overall: The MOS of MPD and the sub-scores of the seven tasks are illustrated in Figure 3, with their correlation quantified by Spearman Rank-order Correlation Coefficient (SRCC). Notably, the inter-task correlation is relatively low, particularly for VQA and the three tasks that rely on specialized CV models. Given the current state of image processing algorithms for machine vision, each employing different tasks as performance metrics, the universality and robustness of such evaluation are inherently constrained. The data distribution further substantiates this argument, with YoN/MCQ/CAP being more evenly distributed, VQA exhibiting a higher concentration of high scores, and SEG/DET/RET clustering at lower scores. Moreover, the strong correlation between the total score and the other tasks is evident, with all SRCC surpassing 0.5, underscoring the reliability of the MPD, for various downstream tasks. The MOS distribution across 30 types of corruption and 5 strength levels is depicted in Figure 4. Results show for corruption scenarios that appear similar to HVS, the MVS perception mechanism diverges significantly. For" + } + ] + } + ], + "index": 49 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "spans": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "type": "text", + "content": "7574" + } + ] + } + ], + "index": 50 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 152, + 137 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 152, + 137 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 152, + 137 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 152, + 137 + ], + "type": "image", + "image_path": "4f12ad9cbaa53718cbafceebbae4c914164a862f9f64ee4ad18ed46e2d673540.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 140, + 133, + 152 + ], + "lines": [ + { + "bbox": [ + 90, + 140, + 133, + 152 + ], + "spans": [ + { + "bbox": [ + 90, + 140, + 133, + 152 + ], + "type": "text", + "content": "Strength 1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 154, + 70, + 233, + 137 + ], + "blocks": [ + { + "bbox": [ + 154, + 70, + 233, + 137 + ], + "lines": [ + { + "bbox": [ + 154, + 70, + 233, + 137 + ], + "spans": [ + { + "bbox": [ + 154, + 70, + 233, + 137 + ], + "type": "image", + "image_path": "5b11c5e7eec650aee15617ad440f1c9bb290e943aff8f9bd2863db85f73ad7cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 140, + 215, + 152 + ], + "lines": [ + { + "bbox": [ + 172, + 140, + 215, + 152 + ], + "spans": [ + { + "bbox": [ + 172, + 140, + 215, + 152 + ], + "type": "text", + "content": "Strength 2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 70, + 316, + 137 + ], + "blocks": [ + { + "bbox": [ + 235, + 70, + 316, + 137 + ], + "lines": [ + { + "bbox": [ + 235, + 70, + 316, + 137 + ], + "spans": [ + { + "bbox": [ + 235, + 70, + 316, + 137 + ], + "type": "image", + "image_path": "04c7b38e9fd0a7ae5802d6da455caae6608bf09d0e953e03999c63dea6459707.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 140, + 297, + 152 + ], + "lines": [ + { + "bbox": [ + 254, + 140, + 297, + 152 + ], + "spans": [ + { + "bbox": [ + 254, + 140, + 297, + 152 + ], + "type": "text", + "content": "Strength 3" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 479, + 140 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 479, + 140 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 479, + 140 + ], + "type": "image", + "image_path": "d85c61f1fe6090e5b519a1b70d3a366d2f9eeb11e54ced17648f6de9039ba905.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 140, + 377, + 152 + ], + "lines": [ + { + "bbox": [ + 335, + 140, + 377, + 152 + ], + "spans": [ + { + "bbox": [ + 335, + 140, + 377, + 152 + ], + "type": "text", + "content": "Strength" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 141, + 460, + 152 + ], + "lines": [ + { + "bbox": [ + 417, + 141, + 460, + 152 + ], + "spans": [ + { + "bbox": [ + 417, + 141, + 460, + 152 + ], + "type": "text", + "content": "Strength 5" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 481, + 78, + 534, + 132 + ], + "blocks": [ + { + "bbox": [ + 481, + 78, + 534, + 132 + ], + "lines": [ + { + "bbox": [ + 481, + 78, + 534, + 132 + ], + "spans": [ + { + "bbox": [ + 481, + 78, + 534, + 132 + ], + "type": "image", + "image_path": "3ed53d883119b1d3dc2dfba449f9a515f21358656635a7e9ad312172a021c09c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 72, + 173, + 135, + 235 + ], + "blocks": [ + { + "bbox": [ + 70, + 155, + 537, + 172 + ], + "lines": [ + { + "bbox": [ + 70, + 155, + 537, + 172 + ], + "spans": [ + { + "bbox": [ + 70, + 155, + 537, + 172 + ], + "type": "text", + "content": "Figure 5. Annotations from different LMMs subjects in MPD under 5 corruption strengths, including " + }, + { + "bbox": [ + 70, + 155, + 537, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 70, + 155, + 537, + 172 + ], + "type": "text", + "content": " content types and 4 tasks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 173, + 135, + 235 + ], + "lines": [ + { + "bbox": [ + 72, + 173, + 135, + 235 + ], + "spans": [ + { + "bbox": [ + 72, + 173, + 135, + 235 + ], + "type": "image", + "image_path": "6dbf5dbe7941a6a52246d6add2bd797c351bd85a158f7c01cf24da41b779df44.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 239, + 133, + 250 + ], + "lines": [ + { + "bbox": [ + 73, + 239, + 133, + 250 + ], + "spans": [ + { + "bbox": [ + 73, + 239, + 133, + 250 + ], + "type": "text", + "content": "Human (0.76)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 138, + 173, + 202, + 235 + ], + "blocks": [ + { + "bbox": [ + 138, + 173, + 202, + 235 + ], + "lines": [ + { + "bbox": [ + 138, + 173, + 202, + 235 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 202, + 235 + ], + "type": "image", + "image_path": "a2077a8a6ae67ae80c8a8e5ea4c34aa8fc825150b8990d843f6dfe6c5c65827e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 239, + 196, + 250 + ], + "lines": [ + { + "bbox": [ + 145, + 239, + 196, + 250 + ], + "spans": [ + { + "bbox": [ + 145, + 239, + 196, + 250 + ], + "type": "text", + "content": "MPD (0.62)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 205, + 173, + 269, + 236 + ], + "blocks": [ + { + "bbox": [ + 205, + 173, + 269, + 236 + ], + "lines": [ + { + "bbox": [ + 205, + 173, + 269, + 236 + ], + "spans": [ + { + "bbox": [ + 205, + 173, + 269, + 236 + ], + "type": "image", + "image_path": "1226e1f518c84a851518fec60497b2036e020ed06340953e21b336885e151261.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 239, + 260, + 250 + ], + "lines": [ + { + "bbox": [ + 214, + 239, + 260, + 250 + ], + "spans": [ + { + "bbox": [ + 214, + 239, + 260, + 250 + ], + "type": "text", + "content": "YoN (0.50)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 253, + 555, + 286 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 555, + 286 + ], + "type": "text", + "content": "Figure 6. Correlation matrix of the proposed MPD in each dimension. A darker color denotes a higher SRCC, with mean SRCC attached below each matrix. The index of each matrix is arranged in the " + }, + { + "bbox": [ + 55, + 253, + 555, + 286 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 253, + 555, + 286 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 253, + 555, + 286 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 253, + 555, + 286 + ], + "type": "text", + "content": " directions, in the same order as Section 3.3. The correlation between human subjects is greater than machines, and the correlation of the overall MOS is greater than the score in each dimension." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 272, + 173, + 336, + 236 + ], + "blocks": [ + { + "bbox": [ + 272, + 173, + 336, + 236 + ], + "lines": [ + { + "bbox": [ + 272, + 173, + 336, + 236 + ], + "spans": [ + { + "bbox": [ + 272, + 173, + 336, + 236 + ], + "type": "image", + "image_path": "f31c9e67faf073c4d6d83c985ad4d30df5f6383f47800d69ae3c4f71fcea8fd5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 239, + 330, + 250 + ], + "lines": [ + { + "bbox": [ + 279, + 239, + 330, + 250 + ], + "spans": [ + { + "bbox": [ + 279, + 239, + 330, + 250 + ], + "type": "text", + "content": "MCQ (0.55)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 339, + 173, + 403, + 236 + ], + "blocks": [ + { + "bbox": [ + 339, + 173, + 403, + 236 + ], + "lines": [ + { + "bbox": [ + 339, + 173, + 403, + 236 + ], + "spans": [ + { + "bbox": [ + 339, + 173, + 403, + 236 + ], + "type": "image", + "image_path": "b4bba999e196b1fd78bce3175088bc56aa8e86744b26a67cbff4bc268cdd3482.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 239, + 395, + 250 + ], + "lines": [ + { + "bbox": [ + 347, + 239, + 395, + 250 + ], + "spans": [ + { + "bbox": [ + 347, + 239, + 395, + 250 + ], + "type": "text", + "content": "VQA (0.37)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 406, + 173, + 470, + 236 + ], + "blocks": [ + { + "bbox": [ + 406, + 173, + 470, + 236 + ], + "lines": [ + { + "bbox": [ + 406, + 173, + 470, + 236 + ], + "spans": [ + { + "bbox": [ + 406, + 173, + 470, + 236 + ], + "type": "image", + "image_path": "065a8dd2b564dee173e7eb415ac9212b4d34a113c7b6e7b27a58916c7f032769.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 239, + 462, + 250 + ], + "lines": [ + { + "bbox": [ + 414, + 239, + 462, + 250 + ], + "spans": [ + { + "bbox": [ + 414, + 239, + 462, + 250 + ], + "type": "text", + "content": "CAP (0.39)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 474, + 173, + 536, + 236 + ], + "blocks": [ + { + "bbox": [ + 474, + 173, + 536, + 236 + ], + "lines": [ + { + "bbox": [ + 474, + 173, + 536, + 236 + ], + "spans": [ + { + "bbox": [ + 474, + 173, + 536, + 236 + ], + "type": "image", + "image_path": "1cf7815dedc0891a3bed14374d3c1aa24ef4e9583ed20a6635c5937a87fb5276.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 239, + 529, + 250 + ], + "lines": [ + { + "bbox": [ + 479, + 239, + 529, + 250 + ], + "spans": [ + { + "bbox": [ + 479, + 239, + 529, + 250 + ], + "type": "text", + "content": "Other (0.53)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 54, + 308, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 308, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 296, + 464 + ], + "type": "text", + "content": "instance, the machine is highly sensitive to 'Lens blur' but rarely affected by 'Mean brighten'. A typical example is the four Macro Block-related corruptions, which are nearly indistinguishable for HVS but exhibit significantly lower perception quality for MVS when encountering 'Block exchange' compared to the other three. Furthermore, the impact of distortion intensity on the MVS perceptual quality varies markedly. For example, as the intensity level increases, the 'Lab saturation' quality diminishes, while the 'HSV saturation' quality remains unchanged. These findings underscore the differences between MVS and HVS, reinforcing the significance of establishing such a unified quality representation in MPD." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": "Each machine: We visualize the benchmark performance of six typical LMMs (as space limited) under different corruption strengths, as shown in Figure 5. For different content types, the scores of various machines on NSI/SCI/AIGI dimensions do not vary significantly. For different tasks, advanced models such as InternVL2 excel in dealing with distorted images in YoN, while traditional models like LLaVA1.5 can provide relatively stable results in CAP. As the corruption strength increases by one level, the scores of each dimension decrease by 0.01-0.02. Finally, for all machine subjects, we explored their internal differences in Figure 6. We manually conducted human annotation for 500 instances (the same 15 subjects) and compared the preference differences between humans and machines. The index of each matrix (i.e., the machine it represents) is arranged in the " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " directions, in the same order as in Section 3.3. Overall, the SRCC between humans is 0.76, while that of machines is only 0.62, indicating that the divergence in machine image preferences is greater than that of humans. Further exploration of the preference scores" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 308, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 452 + ], + "type": "text", + "content": "for each dimension shows that the relatedness of machines in VQA and CAP tasks is the lowest. Additionally, there are a few subjects whose assessments differ significantly from other machines" + }, + { + "bbox": [ + 313, + 308, + 555, + 452 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 308, + 555, + 452 + ], + "type": "text", + "content": ", such as InstructBlip in YoN, Mantis in MCQ, Pangea in CAP, and DIME/NAAF in RET. This is often due to special token outputs or multilingual mechanisms, which lead to insufficient robustness, resulting in a higher proportion of errors in distorted images compared to other models, thus leading to inconsistent results. In summary, the larger preference differences make IQA for MVS a more difficult problem than traditional IQA, but also a more valuable and promising research topic." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 464, + 390, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 390, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 390, + 478 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 485, + 425, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 425, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 425, + 498 + ], + "type": "text", + "content": "5.1. Experiment Setups" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 503, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 555, + 635 + ], + "type": "text", + "content": "We randomly partitioned the MPD into the train/val set, with 24,000 and 6,000 reference and distorted image pairs, respectively, according to an 8:2 ratio. For MVS, Figure 4 illustrates significant variations in the perceptual quality of different corruptions. Given that machines exhibit a larger Just Noticeable Difference (JND) than humans, our corruption strength exceeds that of previous IQA databases for humans3. To conduct a realistic and reliable evaluation of machine preference, we designate the val set as 'Severe Distortion' and select a subset of 'Mild Distortion' to represent the corruption strength in the real world. The specific splitting" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 645, + 553, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 553, + 693 + ], + "type": "text", + "content": "2The ITU human subjective annotation standard [71] for humans will exclude data with weak correlation with other subjects, but this is to prevent eye fatigue or malicious attacks during annotation. Given that the machine's inference process is completely objective, we will refer to the outputs of each machine rather than excluding them." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "text", + "content": "3See appendix for visualization in 5 corruption strength, where the Level 1 of our MPD can even ensemble Level 4 in KADID-10K[47]." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "spans": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "type": "text", + "content": "7575" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 106, + 555, + 414 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 2. Using IQA metrics to predict machine preference, including baseline, FR, and NR algorithm, human visual system-based metrics are marked in " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": ". The first row represents two distortion intensities and three image content subsets, and the second row represents five scoring dimensions. [Keys: Best; Second best; Baseline (BL); Inferior to the baseline.]" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 106, + 555, + 414 + ], + "lines": [ + { + "bbox": [ + 57, + 106, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 57, + 106, + 555, + 414 + ], + "type": "table", + "html": "
TypeMetricHVSSevere DistortionMild DistortionNSISCIAIGI
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
BLPSNR0.38730.26720.46640.30970.21270.61970.41160.28200.49450.31350.21460.33290.52250.36840.5985
SSIM[76]0.59690.41570.55370.26720.18280.28180.62880.44660.59320.54150.37480.53930.59800.42780.5687
FRAHIQ[33]0.80610.61530.81830.53040.36850.51110.84470.66010.85020.79620.60400.84700.79960.60820.7966
CKDN[95]✔0.74840.55600.71740.06040.04050.10610.81000.59880.76700.72030.51710.76610.69920.51020.6869
DISTS[15]0.73810.54430.73720.46160.32320.44860.79120.60600.80460.72750.53840.74290.72190.53650.7292
LPIPS[87]✔0.62530.44550.56010.00790.00920.00560.69650.51320.64780.56180.39200.56860.64290.45180.5481
PieAPP[62]✔0.47370.01630.23770.30300.21000.23490.45360.11870.27290.51730.09540.21350.46140.20090.3168
TOPIQ-FR[9]✔0.71780.52520.71820.42480.29440.48410.75060.56140.75850.72430.53220.71860.69430.50630.7061
NRARNIQA[3]0.83370.64260.81120.23880.16080.27540.87040.66900.84780.82200.62170.84200.84490.65780.7954
CLIPQA[74]✔0.64980.46620.59010.18220.12100.17970.76700.57710.71410.66900.47510.62150.68770.49340.5977
DBCNN[88]✔0.56520.39080.56960.13340.08940.15940.59780.41810.62190.56920.39740.60920.54220.38100.5613
HyperIQA[68]0.81050.61890.80120.53880.37740.47120.89080.70740.86910.82730.64030.86010.82120.62420.8076
NIMA[69]0.75620.56570.78580.45700.31710.43420.81530.62030.82240.78000.58810.80920.78460.59530.7622
TOPIQ-NR[9]✔0.65620.46160.63580.19610.13180.19890.70110.48230.66130.66450.46240.63530.54610.34220.5249
TypeMetricHVSYoNMCQVQACAPOthers
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
BLPSNR0.27530.18920.34670.32730.22390.43720.20670.13780.21200.43190.29930.49890.37980.26170.4683
SSIM[76]0.39600.27280.38040.50890.34990.48800.29600.19880.27560.55500.38610.54470.50720.34910.4766
FRAHIQ[33]0.60170.42830.62120.73140.53740.74660.42440.29240.45830.72200.53410.75460.61840.43960.6311
CKDN[95]✔0.54910.37990.54820.68660.48630.65960.40450.27630.39000.70440.51360.69910.54340.38220.5313
DISTS[15]0.52670.36950.53940.64860.46250.65490.35120.23950.36760.66490.48330.68120.62030.44190.6332
LPIPS[87]✔0.46770.32520.42450.55170.39120.51690.32160.21690.29350.56200.39960.54730.49270.33870.4242
PieAPP[62]✔0.33080.01730.16630.43100.19000.20870.23850.01320.10320.42200.01520.21360.37350.02090.1747
TOPIQ-FR[9]✔0.50110.34800.51180.60380.42320.61650.37000.25080.38990.63270.45450.64790.61990.43830.6281
NRARNIQA[3]0.62280.43940.61720.73420.53510.72420.44090.30260.44540.76560.56930.76640.65600.45660.6309
CLIPQA[74]✔0.46130.31670.43860.50900.35560.51170.29710.20280.32040.63790.45510.56730.57510.40480.4747
DBCNN[88]✔0.37980.25670.39690.47930.32580.49510.25560.17190.27830.52400.36330.53880.50850.34650.5093
HyperIQA[68]0.60100.42780.60660.70660.51550.71460.43220.29960.44190.71530.52340.74430.64610.45980.6329
NIMA[69]0.55440.39040.58870.68930.50020.70360.44600.30740.47750.63130.45310.68420.58100.41060.6181
TOPIQ-NR[9]✔0.49390.33580.47970.54690.37130.53700.35780.24370.37870.51120.33750.50890.57690.39340.5564
", + "image_path": "183c79a084ea8884ac7c6793625a5bdb42be1f4607764605b736223c569c9d50.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 426, + 295, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 522 + ], + "type": "text", + "content": "mechanism is detailed in the Appendix. To benchmark the performance of quality metrics, three global indicators were employed: SRCC, Kendall Rank-order Correlation Coefficient (KRCC), and Pearson Linear Correlation Coefficient (PLCC), to evaluate the consistency between the objective quality score and the subjective MOS. Among these, SRCC and KRCC represent the prediction monotonicity, while PLCC measures the accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "12 mainstream IQA metrics are implemented for comparison, which have all shown satisfying results in previous human-centric IQA tasks, including 6 FR metrics: AHIQ [33], CKDN [95], DISTS [15], LPIPS [87], PieAPP [62], TOPIQ-FR [9]; and 6 NR metrics: ARNIQA [3], CLIPIQA [74], DBCNN [88], HyperIAQ [68], NIMA [69], TOPIQN [9]. Some methods specifically model the HVS, such as DBCNN, which has a dedicated stage for simulating human perception; others rely on the remarkable fitting ability of large-scale parameters, like the ViT in AHIQ. We have marked all HVS-based metrics as " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": ". Moreover, two naive quality indicators, PSNR and SSIM [76] are set as the baseline. We train the above 12 metrics on MPD with the learning rate as " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " for 50 epochs, under the default settings in pyiqa [9] toolbox and evaluate the performance on" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 427, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 475 + ], + "type": "text", + "content": "the two subsets above, three content types, and five scoring dimensions. We collect MPD preference annotations on " + }, + { + "bbox": [ + 313, + 427, + 555, + 475 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 313, + 427, + 555, + 475 + ], + "type": "text", + "content": " NVIDIA A800 SXM4 80GB GPUs, and then conduct IQA training/validation on one GPU above." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 480, + 499, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 499, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 499, + 493 + ], + "type": "text", + "content": "5.2. Experiment Result and Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 498, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 555, + 629 + ], + "type": "text", + "content": "Table 2 presents the performance of current advanced IQA models on the MPD. Overall, regardless of FR or NR, no metric is highly consistent with machine preference, although they are all good at predicting human preference. An interesting finding is that the more accurately and completely a method models the HVS, the more difficult it is to evaluate the perceptual quality of MVS. Considering that most of the current IQA works emphasize human vision, while machines have already surpassed humans to become the main consumer of visual signals, future IQA needs to find a way to balance these two goals." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "For the two distortion subsets, fitting-based methods such as AHIQ/ARNIQA/HyperIQA can achieve SRCC/PLCC values exceeding 0.8 on Severe Distortion. However, this is not a challenging task. Under the same distortion conditions, they have even exceeded 0.95 on the HVS dataset [33, 61]. The performance of Mild Distortion is even worse, and no method can even outperform the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "type": "text", + "content": "7576" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 127, + 296, + 259 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 125 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 125 + ], + "type": "text", + "content": "Table 3. Cross validation between human and machine preference database. For IQA metrics, training on machine preference data can reach a good result in evaluating human preference; but training on human preference is unacceptable for MPD. [Keys: Best; Lower than 0.4.]" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 127, + 296, + 259 + ], + "lines": [ + { + "bbox": [ + 56, + 127, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 127, + 296, + 259 + ], + "type": "table", + "html": "
Val-setKADID-10KTID2013MPD (Mild)MPD (Severe)
MetricSRCC↑PLCC↑SRCC↑PLCC↑SRCC↑PLCC↑SRCC↑PLCC↑
AHIQ[33]0.79500.77320.76510.78840.15140.18460.47530.3518
CKDN[95]0.85990.72820.69890.60290.17130.20550.34940.3228
DISTS[15]0.82390.81900.72310.77590.41050.40770.46160.4486
LPIPS[87]0.61840.54510.70540.63260.23910.31180.46830.3711
PieAPP[62]0.47830.34680.68320.45430.12850.06780.17910.1235
TOPIQ-FR[9]0.67680.68440.77820.74730.51120.54950.62230.6685
ARNIQA[3]0.76840.70230.77010.72050.48290.45820.55970.4849
CLIPIQA[62]0.62860.57290.52370.39100.13670.14780.32010.3203
DBCNN[88]0.51750.53510.44860.57860.13590.15920.40240.3877
HyperIQA[68]0.66500.65470.70240.73130.08200.10740.45190.4521
NIMA[69]0.47950.49150.56940.64230.16290.17750.47090.4845
TOPIQ-NR[9]0.52320.50770.54550.61740.23880.27540.48290.4582
", + "image_path": "d456787f1c5c24f9c1ea6beb4560d6b0a1d3e10b9653be9744a9a23fc6c6bafc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 270, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 473 + ], + "type": "text", + "content": "simple PSNR, which proves that there is still a lot of room for optimization of IQA for machine preference. For the three contents, after verifying them separately, the performance has slightly increased, which is caused by the internal quality differences of the contents. In general, predicting NSI quality is relatively simple, while SCI and AIGI are more difficult. This may be due to the fact that the initial weights of the IQA model have rich NSI priors. For the five dimensions, we can see that the results of MCQ and CAP are most consistent with machine preference; YoN and Others (three common CV tasks) are second; VQA is the worst. This is because VQA problems are usually changes in image details (such as the color of a button on a piece of clothing), which are more difficult for machines to detect than other tasks. In short, IQA for Machine preference still has great research potential, and it is necessary to develop more accurate and effective evaluation mechanisms in the future." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 480, + 203, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 203, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 203, + 492 + ], + "type": "text", + "content": "5.3. Cross Database Validation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 297, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 297, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 297, + 666 + ], + "type": "text", + "content": "To further analyze the differences between MVS and HVS, we conducted cross-validation using MPD and two of the most commonly used human-oriented datasets, TID2013 [61] and KADID-10K [47], employing the same training mechanism as in the previous chapter. We observed that after training on MPD, most IQA metrics still possess a certain ability to predict human preferences, with SRCC values around 0.7 to 0.8; conversely, after training on the other two human-centric databases, IQA metrics struggle to predict machine preferences, with results on MPD not even reaching 0.4. Machine preference data will not affect human tasks, but the opposite will. This finding holds certain guiding significance for future IQA, particularly for a human-machine-friendly paradigm." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 237, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 237, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 237, + 685 + ], + "type": "text", + "content": "5.4. Long-term Availability Validation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "The machine subjects in MPD are categorized into two groups: common CV models and LMMs. The former mod" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 64, + 569, + 175 + ], + "blocks": [ + { + "bbox": [ + 318, + 64, + 569, + 175 + ], + "lines": [ + { + "bbox": [ + 318, + 64, + 569, + 175 + ], + "spans": [ + { + "bbox": [ + 318, + 64, + 569, + 175 + ], + "type": "image", + "image_path": "16331855e00e8ac7e144badfc6023cbe011b025c5ff62f43c9c61ac100922c21.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 184, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 313, + 184, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 555, + 240 + ], + "type": "text", + "content": "Figure 7. The long-term usability of MPD. The preference scores of MPD are consistent with four early-stage LMMs and three advanced LMMs (marked as Traditional/Advanced Models in the background). Their high SRCC proves that MPD can still characterize the perceptual quality of LMMs in the near future." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 262, + 556, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 556, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 556, + 514 + ], + "type": "text", + "content": "els, such as the separation/detection benchmark method, have been largely finalized and generally recognized. In contrast, LMMs are still under rapid evolvement. It is conceivable that MPD may not be able to capture the LMMs that will emerge in the future, or with the update of MPD, the preference of previous LMMs may be sacrificed. Thus, we include 3 advanced and 4 early-stage LMMs as Sec 3.3 listed, integrating them into two databases similar to MPD. Given the inconsistencies in LMM preference on VQA and CAP tasks, as illustrated in Figure 6, we have verified the consistency of MPD ground truth labels with advanced/early-stage labels. For each image, the relationship between these two labels is depicted in Figure 7. The findings indicate that MPD exhibits a high correlation with early models on VQA and with advanced models on CAP. Generally, the SRCC is around 0.6, which is considered acceptable. Considering the GPT-4o surpasses open-source models, we anticipate that MPD is able to characterize LMMs preferences for images for at least a year. In the future, We will continue to update MPD by periodicity introducing the latest LMMs to avoid this limitation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 388, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "This paper extends the application of IQA from human to machine vision for the first time and establishes the first large-scale, fine-grained, multi-dimensional quality database for machines, namely MPD. We select 15 LMMs and 15 common CV models as subjects, show them " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "30\\mathrm{k}" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " reference/distorted image pairs, and summarize their performance on 7 downstream tasks as quality scores. In-depth analysis of the MPD content and IQA benchmark experiments have proved the huge gap between machine and human visual systems, and current human-centric IQA methods cannot accurately predict machine preferences. As the visual data consumption volume of machines has already surpassed humans, we hope that MPD can inspire effective IQA metrics to serve the upcoming machine-centric era." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 751, + 315, + 761 + ], + "type": "text", + "content": "7577" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "text", + "content": "7. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "The work was supported by the National Natural Science Foundation of China under Grant 62225112, 62301310, 623B2073, 62471290; and in part by Sichuan Science and Technology Program under Grant 2024NSFSC1426." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 169, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 169, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 296, + 224 + ], + "type": "text", + "content": "[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 225, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 296, + 278 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 296, + 335 + ], + "type": "text", + "content": "[3] Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. Arniqa: Learning distortion manifold for image quality assessment. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 188-197, 2024. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 336, + 296, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 296, + 402 + ], + "type": "text", + "content": "[4] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 382-398. Springer, 2016. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 446 + ], + "type": "text", + "content": "[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 501 + ], + "type": "text", + "content": "[6] Johannes Balle, Philip A Chou, David Minnen, Saurabh Singh, Nick Johnston, Eirikur Agustsson, Sung Jin Hwang, and George Toderici. Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing, 15(2):339-353, 2020. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 568 + ], + "type": "text", + "content": "[7] Min Cao, Shiping Li, Juntao Li, Liqiang Nie, and Min Zhang. Image-text retrieval: A survey on recent research and development. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 5410-5417. International Joint Conferences on Artificial Intelligence Organization, 2022. Survey Track. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 569, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 296, + 623 + ], + "type": "text", + "content": "[8] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "text", + "content": "[9] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 33:2404-2418, 2024. 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "text", + "content": "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 173 + ], + "type": "text", + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24185–24198, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "text", + "content": "[12] Jinyoung Choi and Bohyung Han. Task-aware quantization network forJPEG image compression. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 309-324. Springer, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 253 + ], + "type": "text", + "content": "[13] U Cisco. Cisco annual internet report (2018-2023) white paper. Cisco: San Jose, CA, USA, 10(1):1-35, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "type": "text", + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "type": "text", + "content": "[15] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 355, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 443 + ], + "type": "text", + "content": "[16] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Ji-aqi Wang. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 444, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 510 + ], + "type": "text", + "content": "[17] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 511, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 511, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 511, + 553, + 566 + ], + "type": "text", + "content": "[18] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, Dahua Lin, and Kai Chen. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models, 2024. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 567, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 601 + ], + "type": "text", + "content": "[19] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 601, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 666 + ], + "type": "text", + "content": "[20] Wen Gao, Siwei Ma, Lingyu Duan, Yonghong Tian, Peiyin Xing, Yaowei Wang, Shanshe Wang, Huizhu Jia, and Tiejun Huang. Digital retina: A way to make the city brain more efficient by visual coding. IEEE Transactions on Circuits and Systems for Video Technology, 31(11):4147-4161, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 669, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 714 + ], + "type": "text", + "content": "[21] Yixuan Gao, Yuqin Cao, Tengchuan Kou, Wei Sun, Yunlong Dong, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Vdpve: Vqa dataset for perceptual video enhancement. In Proceedings of the IEEE/CVF Conference on Computer" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "spans": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "type": "text", + "content": "7578" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "text", + "content": "Vision and Pattern Recognition (CVPR) Workshops, pages 1474-1483, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[22] Zhangwei Gao, Zhe Chen, Erfei Cui, Yiming Ren, Weiyun Wang, Jinguo Zhu, Hao Tian, Shenglong Ye, Junjun He, Xizhou Zhu, Lewei Lu, Tong Lu, Yu Qiao, Jifeng Dai, and Wenhai Wang. Mini-internvl: A flexible-transfer pocket multimodal model with 55" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[23] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "type": "text", + "content": "[24] Jinjin Gu, Haoming Cai, Chao Dong, Jimmy S Ren, Radu Timofte, Yuan Gong, Shanshan Lao, Shuwei Shi, Jiahao Wang, Sidi Yang, et al. Ntire 2022 challenge on perceptual image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 951-967, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 285 + ], + "type": "text", + "content": "[25] Yatharth Gupta, Vishnu V. Jaddipal, Harish Prabhala, Sayak Paul, and Patrick Von Platen. Progressive knowledge distillation of stable diffusion x1 using layer level loss, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "text", + "content": "[26] Shijie Hao, Yuan Zhou, and Yanrong Guo. A brief survey on semantic segmentation with deep learning. Neurocomputing, 406:302-321, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "text", + "content": "[27] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 376 + ], + "type": "text", + "content": "[28] David Holz. Midjourney. https://www.midjourney.com, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 377, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 421 + ], + "type": "text", + "content": "[29] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "type": "text", + "content": "[30] Yunzhong Hou, Liang Zheng, and Stephen Gould. Learning to structure an image with few colors. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 10116-10125, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 467, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 511 + ], + "type": "text", + "content": "[31] Zhimeng Huang, Chuanmin Jia, Shanshe Wang, and Siwei Ma. Visual analysis motivated rate-distortion model for image coding. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "text", + "content": "[32] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "text", + "content": "[33] Shanshan Lao, Yuan Gong, Shuwei Shi, Sidi Yang, Tianhe Wu, Jiahao Wang, Weihao Xia, and Yujiu Yang. Attentionshelp cnns see better: Attention-based hybrid image quality assessment network, 2022. 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 590, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 635 + ], + "type": "text", + "content": "[34] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[35] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiqa-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "type": "text", + "content": "[37] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024.5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "type": "text", + "content": "[38] Chunyi Li, Tengchuan Kou, Yixuan Gao, Yuqin Cao, Wei Sun, Zicheng Zhang, Yingjie Zhou, Zhichao Zhang, Weixia Zhang, Haoning Wu, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Aigia-20k: A large database for a-generated image quality assessment, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[39] Chunyi Li, Guo Lu, Donghui Feng, Haoning Wu, Zicheng Zhang, Xiaohong Liu, Guangtao Zhai, Weisi Lin, and Wenjun Zhang. Misc: Ultra-low bitrate image semantic compression driven by large multimodal model, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 254, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 331 + ], + "type": "text", + "content": "[40] Chunyi Li, Haoning Wu, Hongkun Hao, Zicheng Zhang, Tengchuan Kou, Chaofeng Chen, Lei Bai, Xiaohong Liu, Weisi Lin, and Guangtao Zhai. G-refine: A general quality refiner for text-to-image generation. In Proceedings of the 32nd ACM International Conference on Multimedia, page 7375-7384, New York, NY, USA, 2024. Association for Computing Machinery." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 397 + ], + "type": "text", + "content": "[41] Chunyi Li, Haoning Wu, Zicheng Zhang, Hongkun Hao, Kaiwei Zhang, Lei Bai, Xiaohong Liu, Xiongkuo Min, Weisi Lin, and Guangtao Zhai. Q-refine: A perceptual quality refiner for ai-generated image. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[42] Chunyi Li, Xiele Wu, Haoning Wu, Donghui Feng, Zicheng Zhang, Guo Lu, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Cmc-bench: Towards a new paradigm of visual signal compression, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 487 + ], + "type": "text", + "content": "[43] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "type": "text", + "content": "[44] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 535, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 600 + ], + "type": "text", + "content": "[45] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 602, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 668 + ], + "type": "text", + "content": "[46] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "text", + "content": "[47] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Tenth International Conference on Quality of Multimedia Experience (QoMEX), pages 1-3. IEEE, 2019. 2, 6, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "spans": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "type": "text", + "content": "7579" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[48] Risheng Liu, Long Ma, Tengyu Ma, Xin Fan, and Zhongxuan Luo. Learning with nested scene modeling and cooperative architecture search for low-light vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5953-5969, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[50] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "type": "text", + "content": "[51] Xiaohong Liu, Xiongkuo Min, Guangtao Zhai, Chunyi Li, Tengchuan Kou, et al. Ntire 2024 quality assessment of ai-generated content challenge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 6337-6362, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "text", + "content": "[52] Yuan Liu, Le Tian, Xiao Zhou, and Jie Zhou. Rethinking overlooked aspects in vision-language models. arXiv preprint arXiv:22405.11850, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "text", + "content": "[53] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Yaofeng Sun, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 410 + ], + "type": "text", + "content": "[54] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "text", + "content": "[55] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "text", + "content": "[56] Xiongkuo Min, Kede Ma, Ke Gu, Guangtao Zhai, Zhou Wang, and Weisi Lin. Unified blind quality assessment of compressed natural, graphic, and screen content images. IEEE Transactions on Image Processing, 26(11):5462-5474, 2017. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "text", + "content": "[57] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE transactions on Image Processing, 20 (12):3350-3364, 2011. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "text", + "content": "[58] Zhangkai Ni, Lin Ma, Huanqiang Zeng, Jing Chen, Canhui Cai, and Kai-Kuang Ma. Esim: Edge similarity for screen content image quality assessment. IEEE Transactions on Image Processing, 26(10):4818-4831, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "type": "text", + "content": "[59] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[60] Neel Patwa, Nilesh Ahuja, Srinivasa Somayazulu, Omesh Tickoo, Srenivas Varadarajan, and Shashidhar Koolagudi. Semantic-preserving image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 1281-1285. IEEE, 2020. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "text", + "content": "[61] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, et al. Image database tid2013: Peculiarities, results and perspectives. Signal processing: Image communication, 30:57-77, 2015. 2, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[62] Ekta Prashnani, Hong Cai, Yasamin Mostofi, and Pradeep Sen. Pieapp: Perceptual image-error assessment through pairwise preference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1808-1817, 2018. 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 196, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 251 + ], + "type": "text", + "content": "[63] Leigang Qu, Meng Liu, Jianlong Wu, Zan Gao, and Liqiang Nie. Dynamic modality interaction modeling for image-text retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1104-1113, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "text", + "content": "[64] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[65] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "type": "text", + "content": "[66] Robin Rombach, Andreas Blattmann, and Björn Ommer. Text-guided synthesis of artistic images with retrieval-augmented diffusion models, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "text", + "content": "[67] Chengyao Shen and Qi Zhao. Webpage saliency. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13, pages 33-46. Springer, 2014. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 474 + ], + "type": "text", + "content": "[68] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2020. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 477, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 553, + 510 + ], + "type": "text", + "content": "[69] Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment. IEEE Transactions on Image Processing, 27(8): 3998-4011, 2018. 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 511, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 565 + ], + "type": "text", + "content": "[70] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "text", + "content": "[71] I. T. Union. Methodology for the subjective assessment of the quality of television pictures. ITU-R Recommendation BT. 500-11, 2002. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[72] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[73] Thang Vu, Haeyong Kang, and Chang D Yoo. Scnet: Training inference sample consistency for instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2701-2709, 2021. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "text", + "content": "[74] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Pro" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "spans": [ + { + "bbox": [ + 295, + 751, + 315, + 760 + ], + "type": "text", + "content": "7580" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "ceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[75] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[76] Zhou Wang. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "type": "text", + "content": "[77] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 308 + ], + "type": "text", + "content": "[78] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-bench: A benchmark for general-purpose foundation models on low-level vision. arXiv preprint arXiv:2309.14181, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 309, + 167, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 309, + 167, + 319 + ], + "spans": [ + { + "bbox": [ + 57, + 309, + 167, + 319 + ], + "type": "text", + "content": "[79] XAI. grok-1.5v, 2024. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "text", + "content": "[80] Zhaohui Yang, Yunhe Wang, Chang Xu, Peng Du, Chao Xu, Chunjing Xu, and Qi Tian. Discernible image compression. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1561-1569, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "type": "text", + "content": "[81] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024.5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 488 + ], + "type": "text", + "content": "[82] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13040–13051, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "text", + "content": "[83] Masakazu Yoshimura, Junji Otsuka, Atsushi Irie, and Takeshi Ohashi. Dynamicisp: dynamically controlled image signal processor for image recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12866-12876, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "text", + "content": "[84] Ke Yu, Zexian Li, Yue Peng, Chen Change Loy, and Jinwei Gu. Reconfigisp: Reconfigurable camera image processing pipeline. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4248-4257, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 645 + ], + "type": "text", + "content": "[85] Xiang Yue, Yueqi Song, Akari Asai, Seungone Kim, Jean de Dieu Nyandwi, Simran Khanuja, Anjali Kantharuban, Lintang Sutawika, Sathyanarayanan Ramamoorthy, and Graham Neubig. Pangea: A fully open multilingual multimodal llm for 39 languages, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "text", + "content": "[86] Jing Zhang and Dacheng Tao. Empowering things with intelligence: a survey of the progress, challenges, and opportunities in artificial intelligence of things. IEEE Internet of Things Journal, 8(10):7789-7817, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[87] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 564 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[88] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2018. 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[89] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[90] Kun Zhang, Zhendong Mao, Quan Wang, and Yongdong Zhang. Negative-aware attention framework for image-text matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15661-15670, 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 263, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 319 + ], + "type": "text", + "content": "[91] Zicheng Zhang, Chunyi Li, Wei Sun, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. A perceptual quality assessment exploration for aigc images. In IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 440-445, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 319, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 384 + ], + "type": "text", + "content": "[92] Zicheng Zhang, Wei Sun, Yingjie Zhou, Jun Jia, Zhichao Zhang, Jing Liu, Xiongkuo Min, and Guangtao Zhai. Subjective and objective quality assessment for in-the-wild computer graphics images. ACM Transactions on Multimedia Computing, Communications and Applications, 20(4):1-22, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 386, + 553, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 429 + ], + "type": "text", + "content": "[93] Qi Zhang, Shanshe Wang, Xinfeng Zhang, Chuanmin Jia, Zhao Wang, Siwei Ma, and Wen Gao. Perceptual video coding for machines via satisfied machine ratio modeling, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 430, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 473 + ], + "type": "text", + "content": "[94] Qi Zhang, Shanshe Wang, Xinfeng Zhang, Siwei Ma, Jingshan Pan, and Wen Gao. Predicting satisfied user and machine ratio for compressed images: A unified approach, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 475, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 531 + ], + "type": "text", + "content": "[95] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10242-10251, 2021. 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 532, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 553, + 564 + ], + "type": "text", + "content": "[96] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 4" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 751, + 314, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 751, + 314, + 760 + ], + "spans": [ + { + "bbox": [ + 295, + 751, + 314, + 760 + ], + "type": "text", + "content": "7581" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/5e0e415c-9773-49a3-97ea-205f1760b866_content_list.json b/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/5e0e415c-9773-49a3-97ea-205f1760b866_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4fe539797bcd8d2d61f30dd8956a243f38a01f2b --- /dev/null +++ b/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/5e0e415c-9773-49a3-97ea-205f1760b866_content_list.json @@ -0,0 +1,1716 @@ +[ + { + "type": "text", + "text": "Image Quality Assessment: Investigating Causal Perceptual Effects with Abductive Counterfactual Inference", + "text_level": 1, + "bbox": [ + 135, + 128, + 861, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenhao Shen $^{1}$ Mingliang Zhou $^{1*}$ Yu Chen $^{1}$ Xuekai Wei $^{1}$ Yong Feng $^{1}$ Huayan Pu $^{2}$ Weijia Jia $^{3}$ \\\n $^{1}$ College of Computer Science, Chongqing University\\\n $^{2}$ State Key Laboratory of Mechanical Transmissions, Chongqing University\\\n $^{3}$ BNU-UIC Institute of Artificial Intelligence and Future Networks, Beijing Normal University", + "bbox": [ + 104, + 202, + 906, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shenwenhao163@163.com mingliangzhou@cqu.edu.cn yuchen@stu.cqu.edu.cn xuekaiwei2-c@my.cityu.edu.hk fengyong@cqu.edu.cn phygood_2001@shu.edu.cn jiawj@bnu.edu.cn", + "bbox": [ + 101, + 277, + 906, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 337, + 326, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing full-reference image quality assessment (FR-IQA) methods often fail to capture the complex causal mechanisms that underlie human perceptual responses to image distortions, limiting their ability to generalize across diverse scenarios. In this paper, we propose an FR-IQA method based on abductive counterfactual inference to investigate the causal relationships between deep network features and perceptual distortions. First, we explore the causal effects of deep features on perception and integrate causal reasoning with feature comparison, constructing a model that effectively handles complex distortion types across different IQA scenarios. Second, the analysis of the perceptual causal correlations of our proposed method is independent of the backbone architecture and thus can be applied to a variety of deep networks. Through abductive counterfactual experiments, we validate the proposed causal relationships, confirming the model's superior perceptual relevance and interpretability of quality scores. The experimental results demonstrate the robustness and effectiveness of the method, providing competitive quality predictions across multiple benchmarks. The source code is available at https://anonymous.4open.science/r/DeepCausalQuality-25BC.", + "bbox": [ + 89, + 375, + 485, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 337, + 643, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image quality assessment (IQA) is a fundamental area within image processing that is dedicated to evaluating the perceptual quality of images. The primary objective of IQA is to quantify how closely an image aligns with the human visual system (HVS), particularly when subjected to various distortions such as compression, noise, or blurring. Given its importance, IQA is integral to numerous applications, including image compression, transmission, and enhancement [3, 6, 40], where preserving visual fidelity is critical. IQA methods are generally classified into three main categories: full-reference, reduced-reference, and no-reference approaches, all of which can be considered typical regression problems [11, 13, 41].", + "bbox": [ + 511, + 367, + 906, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In full-reference image quality assessment (FR-IQA), the evaluation process is typically divided into three key stages: feature decomposition, feature comparison, and perceptual score mapping. Recent advancements have highlighted the pivotal role that deep learning methods play in aligning IQA predictions with human subjective scores [7, 14]. Notably, deep representations of images have been shown to strongly correlate with perceptual relevance, emphasizing their significance in quality assessment [4, 28, 39]. In deep networks, different channels capture distinct features of original and distorted images, collectively encoding perceptual information. When quality degradation occurs, these features in the distorted image become contaminated to varying degrees, resulting in altered distinctions from those of the reference image [18]. However, the mere alteration of features is insufficient to fully capture the perceptual quality differences between the original and distorted images, as changes in correlation alone do not adequately explain the complex, nonlinear effects of distortions on human visual perception. Furthermore, while existing methods are effective in capturing statistical dependencies, they fail to reflect the intrinsic mecha", + "bbox": [ + 511, + 568, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Mingliang Zhou.", + "bbox": [ + 107, + 779, + 331, + 791 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62176027, in part by Chongqing Talent under Grant cstc2024ycjh-bgzxm0082, in part by Central University Operating Expenses under Grant 2024CDJGF-044, in part by Chongqing new YC project under Grant CSTB2024YCJH-KYXM0126, in part by the Key Project of Science and Technology Research Program of Chongqing Education Commission of China under Grant No. KJZD-K202402501, and in part by the National Nature Science Foundation of China under Grant No. 62262006.", + "bbox": [ + 89, + 791, + 480, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17990", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nisms that drive changes in perceived quality. Specifically, these methods only quantify how distortions affect the similarity between feature representations and cannot elucidate how these distortions influence human perception. Consequently, they are unable to differentiate between distortions that may exhibit similar feature distances yet have significantly different perceptual impacts. This limitation restricts their capacity to provide accurate quality assessments. Additionally, no existing theoretical framework fully explains why IQA remains such a challenging task or why certain methods succeed in some contexts and fail in others. This lack of clarity impedes advancements within the field.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we define IQA as a counterfactual inference problem, offering a perspective that bridges existing theoretical gaps [25]. Counterfactual reasoning provides a formalized structure for defining the minimal perceptual cost, enabling a more precise understanding of how distortions affect perceived quality. Given that the reference image serves as the baseline for the original visual content, an effective IQA method should be able to predict scores on the basis of the perceptual distortions present in degraded images. However, the manner in which distortions are perceived is essentially probabilistic and is influenced by a variety of contextual factors and prior experience. While deep learning-based methods are adept at capturing complex feature interactions, they often fail to incorporate the probabilistic relationships that underlie perceptual judgments. By constructing IQA from a causal reasoning perspective, our approach aims to quantify perceptual differences in a way that reflects these inherent uncertainties and complexities. This causal perspective provides us with not only a measure of feature similarity but also a more nuanced understanding of how different distortions affect human perception. The main contributions of our study are as follows:", + "bbox": [ + 91, + 277, + 483, + 625 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We investigate the causal effects of deep network features on perception, effectively integrating causal relationships with feature comparisons to construct a deep network-based FR-IQA model. By employing distance comparisons among the causal relationships of deep features, the model can accurately address complex distortion types encountered across various image quality assessment scenarios, resulting in competitive quality prediction outcomes.", + "- The proposed method exhibits significant generality, as the analysis of perceptual causal correlations within deep features is independent of the backbone architecture. This characteristic enables the implementation of the approach across different deep backbone networks, thereby enhancing its versatility and applicability in diverse contexts of image quality assessment.", + "- The validity of the proposed causal relationships is substantiated through experimental evidence derived from" + ], + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "abductive counterfactual inference, which demonstrates superior perceptual relevance and score interpretability and establishes its robustness and practicality.", + "bbox": [ + 526, + 90, + 906, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 152, + 656, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Deep Network-based FR-IQA", + "text_level": 1, + "bbox": [ + 511, + 176, + 777, + 193 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep networks have been employed to model the complex relationships between reference and distorted images in FR-IQA. In [11], convolutional operations on both reference and distorted images are utilized to obtain sensitivity maps, which are then applied to the difference map for quality estimation. Gao et al. [7] measured the local similarity between features extracted from deep neural networks (DNNs) for both reference and distorted images. The local quality indices are aggregated to produce the overall quality score. Subsequent studies have shown that VGG network features serve as effective training losses for tasks such as image synthesis. Notably, the study of [39] compares deep features across various architectures and tasks, showing that perceptual similarity is an emerging property shared by deep visual representations. Hence, some advancements focus on perceptual similarity or differences in deep feature space. For example, Ding et al. [4] employed the structural similarity index (SSIM) between deep features to measure perceptual texture similarity, offering a unified deep image structure and texture similarity (DISTS) framework, and the adaptive DISTS (ADISTS) [5] improved it by using a dispersion index to identify texture regions and adaptively aggregate structure and texture measurements. In contrast, Liao et al. [17] projected image degradations into perceptual space and computed the Wasserstein distance in deep feature space, highlighting the importance of distributional differences in deep feature representations for quality assessment.", + "bbox": [ + 511, + 200, + 906, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methods are limited in capturing perceptual quality differences because they rely on feature alterations and statistical dependencies and fail to model the nonlinear effects of distortions on human perception. They quantify feature similarity without revealing how distortions impact perception, leading to difficulties in distinguishing distortions with similar feature distances but different perceptual effects. Additionally, the absence of a robust theoretical framework to explain why some methods succeed or fail in different scenarios limits progress in IQA.", + "bbox": [ + 511, + 625, + 908, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Causal Inference in Deep Learning", + "text_level": 1, + "bbox": [ + 511, + 787, + 818, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In deep learning, with the presence of confounding factors, the network can easily capture spurious correlations between inputs and their outputs. Therefore, the application of causal inference integration in deep learning has received increasing attention. Causal inference provides a theoretical foundation for disentangling correlation from", + "bbox": [ + 511, + 810, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17991", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "causation, which is particularly useful for improving model generalizability, robustness, and interpretability [12] [35]. In deep learning, causal approaches have been explored to mitigate issues such as bias, confounding, and spurious correlations that often arise from data-driven methods [1, 30]. Yue et al. [36] proposed an interventional few-shot learning method that applies backdoor adjustment to mitigate the confounding effect of pretrained knowledge. Liu et al. [20] addressed both visual confounders in the object detection stage and linguistic confounders in the decoding process for image captioning tasks via causal intervention, leading to substantial improvements in generalizability across different datasets. Hu et al. [9] modelled the causal relationship between image-text matching and fake news detection in multimodal settings and used causal inference to eliminate confounding effects during training and leverage matching bias to improve detection accuracy.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the context of image quality assessment, causal reasoning offers a promising avenue for improving the accuracy and consistency of quality prediction. By focusing on the causal relationship between image features and perceived quality, it can explore how causal models mitigate the problems of spurious correlation and overfitting to better align with human perceptual judgments and improve the reliability of deep learning models in real-world applications.", + "bbox": [ + 89, + 348, + 480, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 498, + 295, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The general objective in image quality assessment is to minimize the discrepancy between the predicted quality score and the subjective mean opinion score (MOS), which can be formalized as:", + "bbox": [ + 89, + 523, + 482, + 583 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\zeta , \\omega , \\theta} \\mathbb {E} [ l (f _ {\\zeta} (\\operatorname {D i s} _ {\\omega} (\\phi_ {\\theta} (I), \\phi_ {\\theta} (D))), M O S) ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 595, + 482, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ represents the parameters of model $\\phi$ , $\\mathrm{Dis}(\\phi_{\\theta}(I),\\phi_{\\theta}(D))$ denotes the distance between the feature representations of the original image $I$ and the distorted image $D$ in the deep feature space $\\phi (\\cdot)$ , and $l(\\cdot)$ is the loss function that quantifies the difference between the predicted score $f_{\\zeta}(\\mathrm{Dis}_{\\omega}(\\phi (I),\\phi (D)))$ and the subjective quality. The expectation $\\mathbb{E}$ is taken over the data distribution.", + "bbox": [ + 89, + 632, + 482, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For full-reference image quality assessment, we consider the causal relationship between the original image $I$ and the distorted image $D$ . The relationship can be expressed as follows:", + "bbox": [ + 89, + 753, + 482, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} _ {S} = m \\left(\\phi_ {\\gamma} (I, D)\\right), \\quad \\eta \\perp \\gamma \\text {a n d} \\gamma , \\eta \\sim F _ {\\theta} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 825, + 480, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $m(\\cdot)$ represents a function that maps the joint feature representation $\\phi_{\\gamma}(I,D)$ of the original and distorted images to the subjective quality $\\mathcal{Q}_S$ . The parameter $\\gamma$ is", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pivotal in this formulation, as it represents the features with causal implications for image quality, thereby influencing the perceptual evaluation. In contrast, $\\eta$ signifies the noise components that are independent of the deep representation parameter $\\gamma$ , suggesting that these components do not contribute to the quality assessment. Both $\\gamma$ and $\\eta$ are derived from the distribution $F_{\\theta}$ , indicating their dependence on the model's parameters, $\\theta$ . Specifically, $\\theta$ represents the foundational parameters of the deep feature representation, which gives rise to two distinct components: $\\gamma$ , the relevant features that causally affect image quality, and $\\eta$ , the extraneous noise that does not contribute to the quality assessment. This causal relationship isolates the effect of distortions from external factors, ensuring that the predicted quality score is influenced primarily by distortions between the reference and distorted images, whereas other factors are captured by $\\eta$ .", + "bbox": [ + 511, + 90, + 903, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In light of the established causal relationship, the subsequent phase focuses on prioritizing causal features $\\gamma$ while effectively disregarding the noise components $\\eta$ . The goal is to adjust the model parameters $\\theta$ to better capture the relevant causal features $\\gamma$ that significantly influence the subjective quality score $Q_{S}$ . The optimization process is designed to minimize the prediction error across all conceivable data distributions, thereby ensuring that the model remains robust against varying inputs and distortions. This objective can be formally articulated as follows:", + "bbox": [ + 511, + 348, + 903, + 500 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\zeta , \\omega , \\theta \\rightarrow \\gamma} \\sup _ {P \\in \\mathcal {P}} \\mathbb {E} _ {P} \\left[ l \\left(f _ {\\zeta} \\left(\\operatorname {D i s} _ {\\omega} \\left(\\phi_ {\\theta} (I), \\phi_ {\\theta} (D)\\right)\\right), M O S\\right)\\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 512, + 903, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the term $\\min_{\\theta \\to \\gamma}$ signifies the optimization of model parameters $\\theta$ to enhance the representation of the causal parameters $\\gamma$ , $\\mathcal{P}$ represents the set of all possible distributions that are consistent with the structural causal model, and the supremum operator $\\sup_{P\\in \\mathcal{P}}$ accounts for the worst-case scenario over these distributions.", + "bbox": [ + 511, + 549, + 903, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our study, instead of training the model from scratch, we define $\\theta$ as the pretrained model weights. We subsequently employ variable interventions to validate the autoregressive properties of the model, facilitating the filtration of extraneous information. This strategy aims to consider only causal representation parameters $\\gamma$ in the quality prediction process while effectively ignoring the noise component $\\eta$ .", + "bbox": [ + 511, + 641, + 905, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 761, + 648, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Framework", + "text_level": 1, + "bbox": [ + 511, + 787, + 640, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 1 (a), the proposed framework integrates multiple stages, starting with the input of reference and distorted images. These images are first processed in the deep domain, where deep learning models extract high-level feature representations. These feature representations serve as the foundation for the subsequent causal inference", + "bbox": [ + 511, + 810, + 903, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17992", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f9874e4eccd13350e1503e0f940418f97216307d9f0b82dd670817893fb4e3ce.jpg", + "image_caption": [ + "(a) Framework", + "(a(c) Structural Causal Model)" + ], + "image_footnote": [], + "bbox": [ + 94, + 111, + 898, + 305 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1b1accfd3c380f58e206fafb862b7deb58fad0c9823136471836aaaaea9b7581.jpg", + "image_caption": [ + "(b) Confounder Dictionary Embedding", + "Figure 1. The framework of our proposed method that separates and leverages causal features for perceptual quality assessment." + ], + "image_footnote": [], + "bbox": [ + 99, + 333, + 651, + 500 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/471c9eff8eb9f353447c3fc51cb950a20bfc64d0cf5e1aa3289cb5085d11abbf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 329, + 866, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "stage. In this stage, the relationships between the extracted features are analysed to identify causal variables that are robust to distortions and potential confounders. Finally, the quality assessment step uses these identified causal variables to evaluate the perceptual quality of the distorted image, producing a quality score that reflects the minimal perceptual cost on the basis of the underlying causal structure.", + "bbox": [ + 89, + 555, + 482, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To explain the causal relationships considered in our approach, we employ a structural causal model (SCM) for the full-reference image quality assessment. The SCM represents variables as nodes and relationships as edges in a graphical structure, using directed arrows to indicate causal relationships between two variables. In our model, we incorporate the reference image $I$ , the distorted image $D$ , the representations of both images $\\phi_{\\theta}(I)$ and $\\phi_{\\theta}(D)$ , and the image quality score $Q$ , as depicted in Figure 1 (c). We assume that the images are characterized by factors that have causal relationships with quality, along with elements that are unrelated to quality. In this framework, image quality is directly determined by factors that causally influence it. Simply using a training process or pretrained weights from other fields to assess image quality typically yields repre", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sentations that are correlated with perceptual quality but do not capture causal relationships. Therefore, we introduce an unknown exogenous variable $U$ to account for uncertainties and identify the representations that are causally related to quality.", + "bbox": [ + 511, + 555, + 906, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Deep Causal Measurement", + "text_level": 1, + "bbox": [ + 511, + 641, + 756, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our approach, to verify causal relationships in perceptual quality assessment, we design a causal validation process using an SCM with an exogenous variable $U$ . This process begins with a causal intervention to test the effect of image characteristics on perceived quality, using the operator $\\mathrm{do}(\\cdot)$ to map both reference and distorted images onto a shared causal feature space. The intervention can be formulated as follows:", + "bbox": [ + 511, + 662, + 906, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {I} = \\phi_ {\\theta} (\\mathrm {d o} (I = i)), \\quad \\mathbf {f} _ {D} = \\phi_ {\\theta} (\\mathrm {d o} (D = d)) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 796, + 906, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi_{\\theta}$ is the feature extraction function parameterized by $\\theta$ , which maps the images into the deep feature space, and $\\mathbf{f}$ denotes the deep features. To substantiate the causal nature of the features, we introduce the exogenous variable $U$ and perform simultaneous interventions $v$ on both the original", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17993", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "reference image features $\\phi_{\\theta}(I)$ and distorted image features $\\phi_{\\theta}(D)$ to generate the intervened causal representations $\\mathbf{f}'$ . This intervention yields representations formulated as follows:", + "bbox": [ + 89, + 90, + 480, + 148 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {I} ^ {\\prime} = \\phi_ {\\theta} (\\mathrm {d o} (I, v)), \\quad \\mathbf {f} _ {D} ^ {\\prime} = \\phi_ {\\theta} (\\mathrm {d o} (D, v)) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 161, + 482, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we calculate the perceptual distance $\\mathrm{Dis}(\\cdot)$ before and after the intervention to assess the invariance of the causal representation. If the perceptual distance differs significantly between the pre- and postintervention states, then the feature representation can be deemed causal, as it is sensitive to distortions influencing perceptual quality. This invariance can be evaluated as follows:", + "bbox": [ + 89, + 186, + 483, + 292 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta = \\operatorname {D i s} \\left(\\mathbf {f} _ {I}, \\mathbf {f} _ {D}\\right) - \\operatorname {D i s} \\left(\\mathbf {f} _ {I} ^ {\\prime}, \\mathbf {f} _ {D} ^ {\\prime}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 301, + 482, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If $\\Delta \\neq 0$ , it confirms that the causal factors within the feature representations significantly affect perceptual quality. In practice, we vary the disturbance intensity to determine the strength of the perceptual causal relationship within the feature representations, and we use a confounder dictionary to record the representations that maintain causality at all intensities, denoted $\\Gamma (\\mathbf{f}_I,\\mathbf{f}_D)$ . The causal influence on perceptual quality is then quantified by observing how the perceptual distance changes with varying intervention intensities.", + "bbox": [ + 89, + 328, + 483, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The regression invariance is validated if the perceptual distance $\\mathrm{Dis}(\\mathbf{f}_I',\\mathbf{f}_D')$ yields a consistent quality score $Q$ under intervention, where $\\mathbf{f}$ is parameterized by $\\gamma$ , such that:", + "bbox": [ + 89, + 479, + 483, + 525 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\mathbb {E} [ Q | \\gamma , v ] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 534, + 482, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This indicates that the quality prediction model relies solely on the causal features $\\mathbf{f}$ , with $\\gamma$ representing the parameter of $\\phi$ in this context, excluding the influence of noise $\\eta$ . The SCM, therefore, ensures that the quality assessment is both robust to noise and aligned with causal perceptual features.", + "bbox": [ + 89, + 560, + 482, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, on the basis of the SCM, the perceptual differences between the reference image and the distorted image are quantified via a distance metric. Our method is predicated on the concept of perceptual cost, facilitating a quantitative analysis of the differences in perceived quality between reference and distorted images. Specifically, we define the perceptual cost between two distributions, $P_{X}$ and $P_{Y}$ , utilizing the causal transport cost, which is formally expressed as:", + "bbox": [ + 89, + 636, + 483, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC O T \\left(P _ {X}, P _ {Y}\\right) = \\inf _ {g \\in G \\left(P _ {X}, P _ {Y}\\right)} \\int \\Gamma (x, y) c (x, y) d g (x, y) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 780, + 482, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $G(P_X, P_Y)$ represents the joint distribution of the reference and distorted images, $c(x, y)$ is the 12 norm of the difference between the distributions, and $\\Gamma(x, y)$ denotes the causal confounder dictionary between the two distributions.", + "bbox": [ + 89, + 825, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Connection to Existing Methods and HVS", + "text_level": 1, + "bbox": [ + 511, + 90, + 870, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing deep learning-based methods [4, 17, 39] leverage high-level feature representations from deep neural networks to improve perceptual alignment by statistically correlating feature distances with quality scores. LPIPS [39] measures the distance between deep network features extracted from image patches, whereas DeepWSD [17] leverages the Wasserstein distance within the deep feature space to evaluate perceptual similarity. DISTS [4], on the other hand, assesses the similarity between image structures and textures by comparing deep feature representations. However, they still rely primarily on correlations and statistical dependencies rather than causal mechanisms, which can limit their ability to generalize effectively across diverse datasets and distortion types. In contrast, our approach introduces a causal framework that aims to reveal the underlying mechanisms driving perceptual differences, providing a more robust and theoretically grounded approach to IQA.", + "bbox": [ + 511, + 112, + 903, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HVS is known to be sensitive to specific perceptual features (e.g., contrast, texture, and structural integrity) that are often distorted by various image processing operations. Previous studies have attempted to model HVS by designing perceptual metrics or using deep learning to automatically learn these perceptual features. However, these models usually fail to distinguish between causal features that actually affect perceptual quality and incidental noise that may be captured by deep features. By introducing a causal model that uses an intervention-based framework, we provide a clearer path for identifying causal features that determine perceptual quality, providing insight into the relationship between image content distortion and perceptual quality in a way that is more consistent with the mechanisms of HVS.", + "bbox": [ + 511, + 369, + 906, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 611, + 718, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 636, + 720, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The study employs six widely used IQA databases to evaluate the effectiveness of the proposed method. These databases include the LIVE IQA [27] database, CSIQ [16], TID2008 [21], TID2013 [22], KADID [19], and PIPAL [10], each containing different types of image distortions and varying numbers of reference images. LIVE [27] comprises 29 reference images and 779 distorted images across five types of distortions. It provides quality scores on a scale of [0, 100]. Containing 30 reference images and 866 distorted images, CSIQ [16] covers six distortion types, with subjective quality scores normalized to the range [0, 1]. TID2008 [21] and TID2013 [22] include 25 reference images. TID2008 has 1,700 distorted images with 17 types of distortions, whereas TID2013 expands to 3,000 distorted images and 24 distortion types. Both databases use a quality score range of [0, 9]. KADID [19] provides a large-scale", + "bbox": [ + 511, + 657, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17994", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0956c9045f3f85dc9e2ac8a888dd80fd28bb19deb0fb37cddc9d840314ce74f3.jpg", + "table_caption": [ + "Table 1. Results of the performance comparison conducted on benchmark datasets." + ], + "table_footnote": [ + "Notably, DISTS and TOPIQ-FR were trained on the KADID [19] and PIPAL [10], respectively, which are marked with an asterisk (*) in the table for these methods. PieAPP and LPIPS were individually trained on their proposed datasets [23] and [39], respectively." + ], + "table_body": "
MethodLIVE [27]CSIQ [16]TID2008 [21]TID2013 [22]KADID [19]PIPAL [10]
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
PSNR0.7810.8010.7920.8070.5070.5250.6640.6870.6700.6760.3980.392
SSIM [33]0.8470.8510.8100.8330.6250.6240.6650.6270.6100.6190.4890.486
MS-SSIM [32]0.8860.9030.8750.8790.8420.8540.8310.7860.8240.8260.5710.545
VIF [26]0.9490.9530.8990.8990.7980.7490.7710.6770.6850.6790.5720.545
MAD [16]0.9040.9070.9220.9220.6810.7080.7370.7430.7170.7260.6140.591
FSIM [37]0.9100.9200.9020.9150.8750.8840.8760.8510.8520.8540.5970.573
VSI [38]0.8770.8990.9120.9290.8710.8950.8980.8950.8770.8780.5480.526
GMSD [34]0.9090.9100.9380.9390.8780.8910.8580.8040.8470.8470.6140.569
NLPD [15]0.8820.8890.9130.9260.8660.8770.8320.7990.8090.8120.4890.464
PieAPP [23]0.8660.8650.8640.8830.7520.7970.8090.8440.8570.8650.7020.701
LPIPS [39]0.8660.8630.8910.8950.7220.7180.7130.7130.8380.8370.5840.584
DISTS [4]0.9240.9250.9190.9200.8290.8140.8540.8300.886*0.886*0.6450.627
DeepWSD [17]0.9040.9250.9410.9500.9000.9040.8940.8740.8870.8880.5030.500
TOPIQ-FR [2]0.8820.8870.8940.8940.8730.8720.8540.8200.8960.8950.837*0.809*
Our-VGG0.9290.9320.9490.9520.9050.8970.9090.8840.8980.8990.6750.657
Our-ResNet0.9140.9210.9160.9230.8590.8680.8620.8520.8880.8900.5380.524
Our-EffNet0.9270.9320.9330.9380.9030.8990.8990.8790.9050.9070.6940.656
", + "bbox": [ + 96, + 114, + 901, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "collection of 81 reference images and 10,125 distorted images, spanning 25 different distortion types. The quality scores fall within the interval [1, 5]. As one of the most extensive datasets, PIPAL [10] includes 250 reference images and 25,850 distorted images across 40 distortion types. The subjective scores range from 868-1857. Notably, we used only 24200 images from the training and validation sets since the labels of the test set were not available. Given the diversity in score ranges across these databases, a linear normalization process is applied to standardize the MOS values, ensuring consistency in the evaluation process. This step ensures that a fair comparison of the quality of predictions can be made across different datasets.", + "bbox": [ + 88, + 507, + 482, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this study, we evaluate a variety of representative and state-of-the-art image quality assessment techniques, including both feature-based and deep learning approaches, including the PSNR, SSIM [33], MS-SSIM [32], VIF [26], MAD [16], FSIM [37], VSI [38], GMSD [34], NLPD [15], PieAPP [23], LPIPS [39], DISTS [4], DeepWSD [17], and TOPIQ-FR [2]. All the comparative methods were re-evaluated in our study. Specifically, for the deep learning-based approaches, we utilized the weight files provided by the authors of these methods to assess their performance through cross-dataset predictions. The training datasets used for these weights are indicated in the table. For the deep domain methods that do not require training, we fol", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "lowed the same protocol as traditional methods and conducted tests directly on the entire dataset. To evaluate the performance of these methods, we employ the Pearson linear correlation coefficient (PLCC) and the Spearman rank correlation coefficient (SRCC) as the key metrics. The PLCC is used to assess the accuracy of the predictions, whereas the SRCC evaluates the monotonic relationship by examining the rank correlation between the predicted quality scores and the corresponding MOS values.", + "bbox": [ + 511, + 507, + 903, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed method is training-free and can be seamlessly applied to various deep learning models for image quality assessment. Specifically, we utilize VGG [29], ResNet [8] and EfficientNet [31] as backbone networks to extract deep feature representations. These networks are pretrained on large-scale datasets; thus, the features can be utilized to assess image quality without additional training.", + "bbox": [ + 511, + 643, + 903, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Quality Prediction Performance", + "text_level": 1, + "bbox": [ + 511, + 757, + 795, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct quality prediction experiments to evaluate the performance and generalizability of our model in comparison with several state-of-the-art methods. The methods we compare can be categorized into three main groups: traditional image quality assessment methods, end-to-end deep learning models, and deep feature-based similarity (or difference) metrics that do not require training. As presented in Table 1 and visualized in the scatter plot in Fig-", + "bbox": [ + 511, + 780, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17995", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e591eb43172702c48b101e02e126cb0979ac1b437fbcea0281c7ed736dce9ed4.jpg", + "image_caption": [ + "(1) SSIM / LIVE" + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 367, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17df6d65834efb588cbfb14267323c0a6a28b8917bc9a4f981171bc0f8c157f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 88, + 506, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c11132faf853ec91786e90e7e767be9de3092af51dd840676ced79b4b119236.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 89, + 638, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/764be3e05b40fb3618cb5b05975f7758f67ef4b4c4381e7c66a4b4b61f53082c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 89, + 777, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a2d13c6f6eb0de52d0114df707227e30553092f7f0970ebe4cbee74c0ddfa53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 89, + 900, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1617eff18b50cafa62090a2e61215db5bfea5cb0963cfe8912e00db29bfe6613.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 185, + 243, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8b67a9889bc3e4f04fb5cf9f485a32c8f23d58d5d8fd60fb632dace9eb9a085.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 185, + 362, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2ca4f68a256d68ae790399d8d2b073f5e7024121d22a441aa0973a10549cfc0.jpg", + "image_caption": [ + "(3) SSIM/TID2008" + ], + "image_footnote": [], + "bbox": [ + 362, + 185, + 500, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6b710b2da393771d294bc5679dd563804825a31f2a6f6aca9401467a42a713f.jpg", + "image_caption": [ + "(4) SSIM / TID2013" + ], + "image_footnote": [], + "bbox": [ + 500, + 185, + 638, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a0fb8bfd8f56f13a3916b5f303ba03c7e6586e6a54dad0dccc2ad0396d31a5a.jpg", + "image_caption": [ + "(5) SSIM / KADID" + ], + "image_footnote": [], + "bbox": [ + 638, + 185, + 776, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db9d15600bdabdde2a46486d159807d8a6983f5656985209a7512845c66df966.jpg", + "image_caption": [ + "(6) SSIM / PIPAL" + ], + "image_footnote": [], + "bbox": [ + 776, + 185, + 900, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a3f6e054e38771ef3c81967d2cdb75eeef73f205fdcd56391a2fb77f0c315d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 272, + 498, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e44869b1229331edce472255c8515b94a65abd3d7247e2b665226329dd70680.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 272, + 750, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0326af9f7f5cd3f3299f1c5563883bf5c67de507bfee7338a96a31f9132b1153.jpg", + "image_caption": [ + "(7) DISTS/LIVE" + ], + "image_footnote": [], + "bbox": [ + 94, + 282, + 243, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d973d0691146ea437da3e6536d76325f95834f014072876cde781f803650d9e.jpg", + "image_caption": [ + "(8) DISTS / CSIQ" + ], + "image_footnote": [], + "bbox": [ + 243, + 282, + 362, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7348994fc4b8b3c7c4cddcb6fbbc39e343407ef98ae6bf1299f3d1e8f73efe46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 282, + 500, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b6838f96eafc679ebafef7a3eac0cac3c5b53c10e6aa0683101a8b425498a438.jpg", + "image_caption": [ + "(10) DISTS/TID2013" + ], + "image_footnote": [], + "bbox": [ + 500, + 282, + 638, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b593655c55c30bb10f4a7dc81877a2942d1241fc7959152027ffcb930ea8075.jpg", + "image_caption": [ + "(12) DISTS/PIPAL" + ], + "image_footnote": [], + "bbox": [ + 638, + 282, + 777, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a401c25bce7c3881a1869ddce272250dada48416a17aee89fa1df3592f4fb223.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 282, + 900, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b900a28943f0da369d53a2fc3df3542851c6efdf5e05adf7d133deeadb4481c2.jpg", + "image_caption": [ + "(13) DeepWSD / LIVE", + "(14) DeepWSD / CSIQ", + "(19) TOPIQ-FR / LIVE", + "(20) TOPIQ-FR / CSIQ" + ], + "image_footnote": [], + "bbox": [ + 94, + 380, + 243, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/706a2e441a5bf10849bd36bbb4fd390b685aff8f21d484440ace5c74aa10c9e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 380, + 362, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c47e13b27de9dd281d0e1528df25961554ac75b039be765957d9322e5b1a030.jpg", + "image_caption": [ + "(15) DeepWSD/TID2008", + "(21) TOPIQ-FR/TID2008" + ], + "image_footnote": [], + "bbox": [ + 362, + 380, + 500, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/303d77f2e2897a3b1b2ee9d137819a61ff28ad68cd8840c2a02a1c9f06d8a10f.jpg", + "image_caption": [ + "(16) DeepWSD / TID2013", + "(17) DeepWSD / KADID", + "(22) TOPIQ-FR / TID2013", + "(23) TOPIQ-FR / KADID" + ], + "image_footnote": [], + "bbox": [ + 500, + 380, + 638, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b228b9b374d8fbb1727eb9da1f440c9f5745c7b5da85eb32d68e555dcd846393.jpg", + "image_caption": [ + "(18) DeepWSD / PIPAL", + "(24) TOPIQ-FR / PIPAL" + ], + "image_footnote": [], + "bbox": [ + 638, + 380, + 777, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0577efd4b72d56d47be0cc0fd87c04c3062e970468112163f68d78e68a64b002.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 380, + 900, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a6c09933cdf42aaf0e35ac47beee01c6b8cfdb1341d60dcb9bb8741bc48ca3e.jpg", + "image_caption": [ + "(25) Ours / LIVE", + "Figure 2. Scatterplot of prediction results for different methods across databases. We performed normalization on both labels and predictions and 4-parameter logistic function is utilized before comparison, as suggested in [24], and the values of the lower-better methods are reversed before regression." + ], + "image_footnote": [], + "bbox": [ + 94, + 478, + 243, + 563 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ddd7987f48ee6dcb5dc254dcd964fcac42e811ef6348967e350e3e024022048.jpg", + "image_caption": [ + "(26) Ours / CSIQ" + ], + "image_footnote": [], + "bbox": [ + 243, + 478, + 362, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/95faa752d5380ef5f74664b1073e6cf952c847ca9cc066dc3fce7906ebb54a68.jpg", + "image_caption": [ + "(27) Ours / TID2008" + ], + "image_footnote": [], + "bbox": [ + 370, + 478, + 500, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7d6d28b46cdb7486d390798fbd6997c0cc0ea1de326e11e9415bb13742135db.jpg", + "image_caption": [ + "(28) Ours / TID2013" + ], + "image_footnote": [], + "bbox": [ + 500, + 478, + 638, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8ed3e67a7cdb4f223f3afda41fcd30c4200e05cd3d66b3b809003d5bbae307c8.jpg", + "image_caption": [ + "(29) Ours / KADID" + ], + "image_footnote": [], + "bbox": [ + 638, + 478, + 777, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29fde11d053dc9d29212ff6e0bb2cb826ec244a4c54c58a8baad9f2c2daaa352.jpg", + "image_caption": [ + "(30) Ours / PIPAL" + ], + "image_footnote": [], + "bbox": [ + 777, + 478, + 900, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ure 2, our method demonstrates superior overall performance and consistently aligns closely with actual scores across datasets. Traditional techniques, which are typically based on handcrafted features, often struggle to capture the complex, nonlinear relationships between image distortions and perceptual quality, resulting in relatively lower accuracy across diverse datasets. End-to-end deep learning methods, while showing a marked improvement over traditional methods, exhibit notable limitations in terms of generalizability. These methods are typically trained on specific datasets, and as a result, their performance tends to degrade when evaluated on datasets that differ from the training set, struggling to maintain consistency and accuracy when applied to unseen data. Furthermore, the deep feature-based similarity methods that do not rely on training also perform well, although they tend to be less precise", + "bbox": [ + 89, + 657, + 483, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "than our model is. These methods capture image distortions via pretrained features, but their inability to explicitly model causal relationships results in slightly less accurate predictions, particularly when the perceptual impact of distortions varies significantly across datasets. In contrast, our approach, which leverages causal reasoning through an SCM and does not require training, excels in maintaining consistent performance across various datasets. The results demonstrate that, by focusing on causally relevant features, our model is less sensitive to the idiosyncrasies of individual datasets and is capable of producing accurate predictions for unseen images. This robustness in generalization suggests that the causal framework used in our model provides a more reliable representation of perceptual quality, independent of dataset-specific biases.", + "bbox": [ + 511, + 657, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17996", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b8eb0078e516f455cb7b3a9bcc7529c0f9aa8f4b6afc1a984896c167df147f9f.jpg", + "table_caption": [ + "Table 2. Performance comparison results of the ablation experiments." + ], + "table_footnote": [], + "table_body": "
ModelLIVE [27]CSIQ [16]TID2013 [22]
ParameterPLCCSRCCPLCCSRCCPLCCSRCC
φ = φθ0.9010.9150.9130.9160.8840.867
φ = φγ0.9290.9320.9490.9520.9090.884
φ = φη0.8430.8660.8030.8310.7860.789
", + "bbox": [ + 96, + 128, + 488, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance comparison results of the ablation experiments conducted on three benchmark datasets, where we use OurVGG as the default architecture. The models evaluated in this study correspond to different configurations of weight assignments for $\\phi$ , where its values $\\phi_{\\theta}$ , $\\phi_{\\gamma}$ , and $\\phi_{\\eta}$ represent the original pretrained weights, the causal effect weights, and the residual component excluding the causal effect, respectively.", + "bbox": [ + 109, + 229, + 482, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Analysis", + "text_level": 1, + "bbox": [ + 89, + 349, + 264, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we present the abductive counterfactual experiments to evaluate the impact of different weight configurations in our model. Specifically, we analyse the results when varying the weights of $\\theta$ , $\\gamma$ , and $\\eta$ . Here, $\\theta$ represents the original pretrained weights, $\\gamma$ corresponds to the weight associated with the causal effect, and $\\eta$ denotes the remaining part of the model, which is the portion of $\\theta$ excluding $\\gamma$ . The ablation analysis is conducted by systematically adjusting each of these weights and observing their influence on the model's performance. The results are summarized in Table 2, which indicate that although noncausal representations also exhibit some perceptual relevance, their lack of causal relationships limits their effectiveness. When these noncausal representations are removed from the original feature set, the model's predictive performance improves. This suggests that incorporating those features that capture causal effects can enhance the model's ability to make more accurate and reliable quality predictions.", + "bbox": [ + 88, + 371, + 482, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 89, + 652, + 302, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we visually demonstrate the effect of causal intervention on different channels. As shown in Figure 3, we demonstrate the visualization effects of causal intervention between the same reference image and different distorted images (blurred and additive Gaussian noise) via the Our-VGG architecture. These visualizations reveal the varying levels of sensitivity each perceptual channel exhibits to different types of distortions in the pretrained model, thus providing a clear indication of how such distortions alter the perceptual characteristics of the image. The horizontal axis in Figure 3 represents the channels, and the color intensity corresponds to the strength of the causal relationship, with brighter colors indicating a stronger causal relationship. Causal features reflect attributes that are crucial for human perception of image quality, as they directly", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99d0e3ba122d38ba067257c08ae2064ded16a448839924f0781da2a8bf3f37a1.jpg", + "image_caption": [ + "Blur" + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 622, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/209574be8c01c266f39f3bf848e90c30162269ad8877df937bdd5037de674695.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 88, + 919, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/604315eedc6e3a5437f5e12e003c149b714e9ac3a6a7555b0f8efb36ffc5806c.jpg", + "image_caption": [ + "Gaussian Noise" + ], + "image_footnote": [], + "bbox": [ + 517, + 202, + 622, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fabbb6897c6c1d84f8cbc3c14b41be19899015eec8290a7faa8f40c55852ba0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 199, + 919, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6518ea22d993259ab7cc85da313a2258635a6defa5d1b953dd7b477d39dc8ce3.jpg", + "image_caption": [ + "Figure 3. Effects of causal interventions on the perceptual representations of images under various distortions. The displayed channels, from top to bottom, correspond to the second, fourth, and fifth stages of the VGG-16 architecture." + ], + "image_footnote": [], + "bbox": [ + 635, + 263, + 916, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "influence perceptual judgments. In contrast, noncausal features typically correspond to higher-level semantic aspects of the image, such as object recognition, which are less relevant for perceptual quality but play a significant role in high level vision tasks. This distinction is supported by the observation that deeper layers of the network are more attuned to classification tasks, while quality-related features are predominantly represented in the shallower layers, underscoring the channel-specific nature of perceptual changes. By isolating the causal effects, we can identify specific features or regions of the image that are most sensitive to certain types of distortions, thus further elucidating how different distortions affect overall image perception.", + "bbox": [ + 511, + 393, + 906, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 603, + 633, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, we proposed an FR-IQA method grounded in abductive counterfactual inference to analyse the causal relationships between deep network features and perceptual distortions. We investigated the causal effects of deep features on human perception and incorporated causal reasoning with feature comparisons, constructing a model capable of addressing various complex distortions in different IQA scenarios. The proposed approach demonstrated strong generalizability, as the analysis of perceptual causal correlations was shown to be independent of the backbone architecture, making it adaptable across diverse deep networks. Additionally, abductive counterfactual experiments validated the effectiveness of the proposed causal framework, confirming the model's superior perceptual relevance and interpretability of quality scores. The experimental results reinforced the robustness and competitiveness of the method, which achieved consistent performance across multiple IQA benchmarks.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17997", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hengrui Cai, Yixin Wang, Michael Jordan, and Rui Song. On learning necessary and sufficient causal graphs. In Advances in Neural Information Processing Systems, pages 42148-42160. Curran Associates, Inc., 2023. 3", + "[2] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 33:2404-2418, 2024. 6", + "[3] Li-Heng Chen, Christos G. Bampis, Zhi Li, Andrey Norkin, and Alan C. Bovik. Proxiqa: A proxy approach to perceptual optimization of learned image compression. IEEE Transactions on Image Processing, 30:360-373, 2021. 1", + "[4] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 1, 2, 5, 6", + "[5] Keyan Ding, Yi Liu, Xueyi Zou, Shiqi Wang, and Kede Ma. Locally adaptive structure and texture similarity for image quality assessment. In Proceedings of the 29th ACM International Conference on Multimedia, page 2483-2491, New York, NY, USA, 2021. Association for Computing Machinery. 2", + "[6] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Comparison of full-reference image quality models for optimization of image processing systems. International Journal of Computer Vision, 129(4):1258-1281, 2021. 1", + "[7] Fei Gao, Yi Wang, Panpeng Li, Min Tan, Jun Yu, and Yani Zhu. Deepsim: Deep similarity for image quality assessment. Neurocomputing, 257:104-114, 2017. Machine Learning and Signal Processing for Big Multimedia Analysis. 1, 2", + "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6", + "[9] Linmei Hu, Ziwei Chen, Ziwang Zhao, Jianhua Yin, and Liqiang Nie. Causal inference for leveraging image-text matching bias in multi-modal fake news detection. IEEE Transactions on Knowledge and Data Engineering, 35(11): 11141-11152, 2023. 3", + "[10] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S. Ren, and Dong Chao. Pipal: A large-scale image quality assessment dataset for perceptual image restoration. In Computer Vision - ECCV 2020, pages 633-651, Cham, 2020. Springer International Publishing. 5, 6", + "[11] Jongyoo Kim and Sanghoon Lee. Deep learning of human visual sensitivity in image quality assessment framework. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1676-1684, 2017. 1, 2", + "[12] Junho Kim, Byung-Kwan Lee, and Yong Man Ro. Demystifying causal features on adversarial examples and causal inoculation for robust network by adversarial instrumental variable regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12302-12312, 2023. 3" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Xuting Lan, Mingliang Zhou, Xueyong Xu, Xuekai Wei, Xingran Liao, Huayan Pu, Jun Luo, Tao Xiang, Bin Fang, and Zhaowei Shang. Multilevel feature fusion for end-to-end blind image quality assessment. IEEE Transactions on Broadcasting, 69(3):801-811, 2023. 1", + "[14] Shanshan Lao, Yuan Gong, Shuwei Shi, Sidi Yang, Tianhe Wu, Jiahao Wang, Weihao Xia, and Yujiu Yang. Attentions help cnns see better: Attention-based hybrid image quality assessment network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1140-1149, 2022. 1", + "[15] Valero Laparra, Johannes Balle, Alexander Berardino, and Eero P. Simoncelii. Perceptual image quality assessment using a normalized laplacian pyramid. In Human Vision and Electronic Imaging, pages 43-48, United States, 2016. Society for Imaging Science and Technology. 6", + "[16] Eric Cooper Larson and Damon Michael Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of Electronic Imaging, 19 (1):011006, 2010. 5, 6, 8", + "[17] Xingran Liao, Baoliang Chen, Hanwei Zhu, Shiqi Wang, Mingliang Zhou, and Sam Kwong. Deepwsd: Projecting degradations in perceptual space to wasserstein distance in deep feature space. In Proceedings of the 30th ACM International Conference on Multimedia, page 970-978, New York, NY, USA, 2022. Association for Computing Machinery. 2, 5, 6", + "[18] Xingran Liao, Xuekai Wei, Mingliang Zhou, Zhengguo Li, and Sam Kwong. Image quality assessment: Measuring perceptual degradation via distribution measures in deep feature spaces. IEEE Transactions on Image Processing, 33:4044-4059, 2024. 1", + "[19] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience, pages 1–3, 2019. 5, 6", + "[20] Bing Liu, Dong Wang, Xu Yang, Yong Zhou, Rui Yao, Zhiwen Shao, and Jiaqi Zhao. Show, deconfound and tell: Image captioning with causal inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18041-18050, 2022. 3", + "[21] Nikolay Ponomarenko, Vladimir Lukin, Alexander Zelensky, Karen Egiazarian, Marco Carli, and Federica Battisti. Tid2008-a database for evaluation of full-reference visual quality assessment metrics. Advances of Modern Radioelectronics, 10(4):30-45, 2009. 5, 6", + "[22] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, and C.-C. Jay Kuo. Image database tid2013: Peculiarities, results and perspectives. Signal Processing: Image Communication, 30: 57-77, 2015. 5, 6, 8", + "[23] Ekta Prashnani, Hong Cai, Yasamin Mostofi, and Pradeep Sen. Pieapp: Perceptual image-error assessment through pairwise preference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1808-1817, 2018. 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17998", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Ann Marie Rohaly, Philip J. Corriveau, John M. Libert, Arthur A. Webster, Vittorio Baroncini, John Beerends, Jean-Louis Blin, Laura Contin, Takahiro Hamada, David Harrison, Andries P. Hekstra, Jeffrey Lubin, Yukihiro Nishida, Ricardo Nishihara, John C. Pearson, Antonio Franca Pessoa, Neil Pickford, Alexander Schertz, Massimo Visca, Andrew B. Watson, and Stefan Winkler. Video Quality Experts Group: current results and future directions. In Visual Communications and Image Processing 2000, pages 742 - 753. International Society for Optics and Photonics, SPIE, 2000.", + "[25] Ladan Shams and Ulrik R. Beierholm. Causal inference in perception. Trends in Cognitive Sciences, 14(9):425-432, 2010. 2", + "[26] H.R. Sheikh and A.C. Bovik. Image information and visual quality. IEEE Transactions on Image Processing, 15(2):430-444, 2006. 6", + "[27] H.R. Sheikh, M.F. Sabir, and A.C. Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on Image Processing, 15(11): 3440-3451, 2006. 5, 6, 8", + "[28] Wenhao Shen, Mingliang Zhou, Jun Luo, Zhengguo Li, and Sam Kwong. Graph-represented distribution similarity index for full-reference image quality assessment. IEEE Transactions on Image Processing, 33:3075-3089, 2024. 1", + "[29] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, page 1-14, 2015. 6", + "[30] Xue Song, Jiequan Cui, Hanwang Zhang, Jingjing Chen, Richang Hong, and Yu-Gang Jiang. Doubly abductive counterfactual inference for text-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9162-9171, 2024. 3", + "[31] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 6", + "[32] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems and Computers, 2003, pages 1398-1402. IEEE, 2003. 6", + "[33] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6", + "[34] Wufeng Xue, Lei Zhang, Xuanqin Mou, and Alan C. Bovik. Gradient magnitude similarity deviation: A highly efficient perceptual image quality index. IEEE Transactions on Image Processing, 23(2):684-695, 2014. 6", + "[35] Mengyue Yang, Xinyu Cai, Furui Liu, Weinan Zhang, and Jun Wang. Specify robust causal representation from mixed observations. New York, NY, USA, 2023. Association for Computing Machinery. 3", + "[36] Zhongqi Yue, Hanwang Zhang, Qianru Sun, and Xian-Sheng Hua. Interventional few-shot learning. In Advances in Neural Information Processing Systems, pages 2734-2746. Curran Associates, Inc., 2020. 3" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. Fsim: A feature similarity index for image quality assessment. IEEE Transactions on Image Processing, 20(8):2378-2386, 2011. 6", + "[38] Lin Zhang, Ying Shen, and Hongyu Li. Vsi: A visual saliency-induced index for perceptual image quality assessment. IEEE Transactions on Image Processing, 23(10): 4270-4281, 2014. 6", + "[39] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 1, 2, 5, 6", + "[40] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10242-10251, 2021. 1", + "[41] Mingliang Zhou, Shujun Lang, Taiping Zhang, Xingran Liao, Zhaowei Shang, Tao Xiang, and Bin Fang. Attentional feature fusion for end-to-end blind image quality assessment. IEEE Transactions on Broadcasting, 69(1):144-152, 2022. 1" + ], + "bbox": [ + 516, + 92, + 903, + 401 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17999", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/5e0e415c-9773-49a3-97ea-205f1760b866_model.json b/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/5e0e415c-9773-49a3-97ea-205f1760b866_model.json new file mode 100644 index 0000000000000000000000000000000000000000..da69c67acbf122e7fa92b67e26dc942cfc5a509c --- /dev/null +++ b/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/5e0e415c-9773-49a3-97ea-205f1760b866_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.13, + 0.862, + 0.175 + ], + "angle": 0, + "content": "Image Quality Assessment: Investigating Causal Perceptual Effects with Abductive Counterfactual Inference" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.203, + 0.908, + 0.277 + ], + "angle": 0, + "content": "Wenhao Shen\\(^{1}\\) Mingliang Zhou\\(^{1*}\\) Yu Chen\\(^{1}\\) Xuekai Wei\\(^{1}\\) Yong Feng\\(^{1}\\) Huayan Pu\\(^{2}\\) Weijia Jia\\(^{3}\\)\\\n\\(^{1}\\)College of Computer Science, Chongqing University\\\n\\(^{2}\\)State Key Laboratory of Mechanical Transmissions, Chongqing University\\\n\\(^{3}\\)BNU-UIC Institute of Artificial Intelligence and Future Networks, Beijing Normal University" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.278, + 0.908, + 0.304 + ], + "angle": 0, + "content": "shenwenhao163@163.com mingliangzhou@cqu.edu.cn yuchen@stu.cqu.edu.cn xuekaiwei2-c@my.cityu.edu.hk fengyong@cqu.edu.cn phygood_2001@shu.edu.cn jiawj@bnu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.338, + 0.327, + 0.354 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.486, + 0.737 + ], + "angle": 0, + "content": "Existing full-reference image quality assessment (FR-IQA) methods often fail to capture the complex causal mechanisms that underlie human perceptual responses to image distortions, limiting their ability to generalize across diverse scenarios. In this paper, we propose an FR-IQA method based on abductive counterfactual inference to investigate the causal relationships between deep network features and perceptual distortions. First, we explore the causal effects of deep features on perception and integrate causal reasoning with feature comparison, constructing a model that effectively handles complex distortion types across different IQA scenarios. Second, the analysis of the perceptual causal correlations of our proposed method is independent of the backbone architecture and thus can be applied to a variety of deep networks. Through abductive counterfactual experiments, we validate the proposed causal relationships, confirming the model's superior perceptual relevance and interpretability of quality scores. The experimental results demonstrate the robustness and effectiveness of the method, providing competitive quality predictions across multiple benchmarks. The source code is available at https://anonymous.4open.science/r/DeepCausalQuality-25BC." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.338, + 0.645, + 0.354 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.907, + 0.565 + ], + "angle": 0, + "content": "Image quality assessment (IQA) is a fundamental area within image processing that is dedicated to evaluating the perceptual quality of images. The primary objective of IQA is to quantify how closely an image aligns with the human visual system (HVS), particularly when subjected to various distortions such as compression, noise, or blurring. Given its importance, IQA is integral to numerous applications, including image compression, transmission, and enhancement [3, 6, 40], where preserving visual fidelity is critical. IQA methods are generally classified into three main categories: full-reference, reduced-reference, and no-reference approaches, all of which can be considered typical regression problems [11, 13, 41]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In full-reference image quality assessment (FR-IQA), the evaluation process is typically divided into three key stages: feature decomposition, feature comparison, and perceptual score mapping. Recent advancements have highlighted the pivotal role that deep learning methods play in aligning IQA predictions with human subjective scores [7, 14]. Notably, deep representations of images have been shown to strongly correlate with perceptual relevance, emphasizing their significance in quality assessment [4, 28, 39]. In deep networks, different channels capture distinct features of original and distorted images, collectively encoding perceptual information. When quality degradation occurs, these features in the distorted image become contaminated to varying degrees, resulting in altered distinctions from those of the reference image [18]. However, the mere alteration of features is insufficient to fully capture the perceptual quality differences between the original and distorted images, as changes in correlation alone do not adequately explain the complex, nonlinear effects of distortions on human visual perception. Furthermore, while existing methods are effective in capturing statistical dependencies, they fail to reflect the intrinsic mecha" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.78, + 0.332, + 0.792 + ], + "angle": 0, + "content": "*Corresponding author: Mingliang Zhou." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.792, + 0.482, + 0.9 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62176027, in part by Chongqing Talent under Grant cstc2024ycjh-bgzxm0082, in part by Central University Operating Expenses under Grant 2024CDJGF-044, in part by Chongqing new YC project under Grant CSTB2024YCJH-KYXM0126, in part by the Key Project of Science and Technology Research Program of Chongqing Education Commission of China under Grant No. KJZD-K202402501, and in part by the National Nature Science Foundation of China under Grant No. 62262006." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.78, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17990" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "nisms that drive changes in perceived quality. Specifically, these methods only quantify how distortions affect the similarity between feature representations and cannot elucidate how these distortions influence human perception. Consequently, they are unable to differentiate between distortions that may exhibit similar feature distances yet have significantly different perceptual impacts. This limitation restricts their capacity to provide accurate quality assessments. Additionally, no existing theoretical framework fully explains why IQA remains such a challenging task or why certain methods succeed in some contexts and fail in others. This lack of clarity impedes advancements within the field." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.278, + 0.485, + 0.625 + ], + "angle": 0, + "content": "To address these challenges, we define IQA as a counterfactual inference problem, offering a perspective that bridges existing theoretical gaps [25]. Counterfactual reasoning provides a formalized structure for defining the minimal perceptual cost, enabling a more precise understanding of how distortions affect perceived quality. Given that the reference image serves as the baseline for the original visual content, an effective IQA method should be able to predict scores on the basis of the perceptual distortions present in degraded images. However, the manner in which distortions are perceived is essentially probabilistic and is influenced by a variety of contextual factors and prior experience. While deep learning-based methods are adept at capturing complex feature interactions, they often fail to incorporate the probabilistic relationships that underlie perceptual judgments. By constructing IQA from a causal reasoning perspective, our approach aims to quantify perceptual differences in a way that reflects these inherent uncertainties and complexities. This causal perspective provides us with not only a measure of feature similarity but also a more nuanced understanding of how different distortions affect human perception. The main contributions of our study are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.629, + 0.483, + 0.763 + ], + "angle": 0, + "content": "- We investigate the causal effects of deep network features on perception, effectively integrating causal relationships with feature comparisons to construct a deep network-based FR-IQA model. By employing distance comparisons among the causal relationships of deep features, the model can accurately address complex distortion types encountered across various image quality assessment scenarios, resulting in competitive quality prediction outcomes." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.87 + ], + "angle": 0, + "content": "- The proposed method exhibits significant generality, as the analysis of perceptual causal correlations within deep features is independent of the backbone architecture. This characteristic enables the implementation of the approach across different deep backbone networks, thereby enhancing its versatility and applicability in diverse contexts of image quality assessment." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- The validity of the proposed causal relationships is substantiated through experimental evidence derived from" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.629, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "abductive counterfactual inference, which demonstrates superior perceptual relevance and score interpretability and establishes its robustness and practicality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.153, + 0.658, + 0.168 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.178, + 0.779, + 0.194 + ], + "angle": 0, + "content": "2.1. Deep Network-based FR-IQA" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.201, + 0.907, + 0.623 + ], + "angle": 0, + "content": "Deep networks have been employed to model the complex relationships between reference and distorted images in FR-IQA. In [11], convolutional operations on both reference and distorted images are utilized to obtain sensitivity maps, which are then applied to the difference map for quality estimation. Gao et al. [7] measured the local similarity between features extracted from deep neural networks (DNNs) for both reference and distorted images. The local quality indices are aggregated to produce the overall quality score. Subsequent studies have shown that VGG network features serve as effective training losses for tasks such as image synthesis. Notably, the study of [39] compares deep features across various architectures and tasks, showing that perceptual similarity is an emerging property shared by deep visual representations. Hence, some advancements focus on perceptual similarity or differences in deep feature space. For example, Ding et al. [4] employed the structural similarity index (SSIM) between deep features to measure perceptual texture similarity, offering a unified deep image structure and texture similarity (DISTS) framework, and the adaptive DISTS (ADISTS) [5] improved it by using a dispersion index to identify texture regions and adaptively aggregate structure and texture measurements. In contrast, Liao et al. [17] projected image degradations into perceptual space and computed the Wasserstein distance in deep feature space, highlighting the importance of distributional differences in deep feature representations for quality assessment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.909, + 0.777 + ], + "angle": 0, + "content": "Existing methods are limited in capturing perceptual quality differences because they rely on feature alterations and statistical dependencies and fail to model the nonlinear effects of distortions on human perception. They quantify feature similarity without revealing how distortions impact perception, leading to difficulties in distinguishing distortions with similar feature distances but different perceptual effects. Additionally, the absence of a robust theoretical framework to explain why some methods succeed or fail in different scenarios limits progress in IQA." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.819, + 0.804 + ], + "angle": 0, + "content": "2.2. Causal Inference in Deep Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In deep learning, with the presence of confounding factors, the network can easily capture spurious correlations between inputs and their outputs. Therefore, the application of causal inference integration in deep learning has received increasing attention. Causal inference provides a theoretical foundation for disentangling correlation from" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17991" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "causation, which is particularly useful for improving model generalizability, robustness, and interpretability [12] [35]. In deep learning, causal approaches have been explored to mitigate issues such as bias, confounding, and spurious correlations that often arise from data-driven methods [1, 30]. Yue et al. [36] proposed an interventional few-shot learning method that applies backdoor adjustment to mitigate the confounding effect of pretrained knowledge. Liu et al. [20] addressed both visual confounders in the object detection stage and linguistic confounders in the decoding process for image captioning tasks via causal intervention, leading to substantial improvements in generalizability across different datasets. Hu et al. [9] modelled the causal relationship between image-text matching and fake news detection in multimodal settings and used causal inference to eliminate confounding effects during training and leverage matching bias to improve detection accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.482, + 0.483 + ], + "angle": 0, + "content": "In the context of image quality assessment, causal reasoning offers a promising avenue for improving the accuracy and consistency of quality prediction. By focusing on the causal relationship between image features and perceived quality, it can explore how causal models mitigate the problems of spurious correlation and overfitting to better align with human perceptual judgments and improve the reliability of deep learning models in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.499, + 0.296, + 0.515 + ], + "angle": 0, + "content": "3. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.483, + 0.584 + ], + "angle": 0, + "content": "The general objective in image quality assessment is to minimize the discrepancy between the predicted quality score and the subjective mean opinion score (MOS), which can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.597, + 0.483, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\zeta , \\omega , \\theta} \\mathbb {E} [ l (f _ {\\zeta} (\\operatorname {D i s} _ {\\omega} (\\phi_ {\\theta} (I), \\phi_ {\\theta} (D))), M O S) ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.483, + 0.752 + ], + "angle": 0, + "content": "where \\(\\theta\\) represents the parameters of model \\(\\phi\\), \\(\\mathrm{Dis}(\\phi_{\\theta}(I),\\phi_{\\theta}(D))\\) denotes the distance between the feature representations of the original image \\(I\\) and the distorted image \\(D\\) in the deep feature space \\(\\phi (\\cdot)\\), and \\(l(\\cdot)\\) is the loss function that quantifies the difference between the predicted score \\(f_{\\zeta}(\\mathrm{Dis}_{\\omega}(\\phi (I),\\phi (D)))\\) and the subjective quality. The expectation \\(\\mathbb{E}\\) is taken over the data distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.754, + 0.483, + 0.813 + ], + "angle": 0, + "content": "For full-reference image quality assessment, we consider the causal relationship between the original image \\(I\\) and the distorted image \\(D\\). The relationship can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.827, + 0.482, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} _ {S} = m \\left(\\phi_ {\\gamma} (I, D)\\right), \\quad \\eta \\perp \\gamma \\text {a n d} \\gamma , \\eta \\sim F _ {\\theta} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Here, \\(m(\\cdot)\\) represents a function that maps the joint feature representation \\(\\phi_{\\gamma}(I,D)\\) of the original and distorted images to the subjective quality \\(\\mathcal{Q}_S\\). The parameter \\(\\gamma\\) is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.349 + ], + "angle": 0, + "content": "pivotal in this formulation, as it represents the features with causal implications for image quality, thereby influencing the perceptual evaluation. In contrast, \\(\\eta\\) signifies the noise components that are independent of the deep representation parameter \\(\\gamma\\), suggesting that these components do not contribute to the quality assessment. Both \\(\\gamma\\) and \\(\\eta\\) are derived from the distribution \\(F_{\\theta}\\), indicating their dependence on the model's parameters, \\(\\theta\\). Specifically, \\(\\theta\\) represents the foundational parameters of the deep feature representation, which gives rise to two distinct components: \\(\\gamma\\), the relevant features that causally affect image quality, and \\(\\eta\\), the extraneous noise that does not contribute to the quality assessment. This causal relationship isolates the effect of distortions from external factors, ensuring that the predicted quality score is influenced primarily by distortions between the reference and distorted images, whereas other factors are captured by \\(\\eta\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.905, + 0.501 + ], + "angle": 0, + "content": "In light of the established causal relationship, the subsequent phase focuses on prioritizing causal features \\(\\gamma\\) while effectively disregarding the noise components \\(\\eta\\). The goal is to adjust the model parameters \\(\\theta\\) to better capture the relevant causal features \\(\\gamma\\) that significantly influence the subjective quality score \\(Q_{S}\\). The optimization process is designed to minimize the prediction error across all conceivable data distributions, thereby ensuring that the model remains robust against varying inputs and distortions. This objective can be formally articulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.513, + 0.905, + 0.538 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\zeta , \\omega , \\theta \\rightarrow \\gamma} \\sup _ {P \\in \\mathcal {P}} \\mathbb {E} _ {P} \\left[ l \\left(f _ {\\zeta} \\left(\\operatorname {D i s} _ {\\omega} \\left(\\phi_ {\\theta} (I), \\phi_ {\\theta} (D)\\right)\\right), M O S\\right)\\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.905, + 0.639 + ], + "angle": 0, + "content": "where the term \\(\\min_{\\theta \\to \\gamma}\\) signifies the optimization of model parameters \\(\\theta\\) to enhance the representation of the causal parameters \\(\\gamma\\), \\(\\mathcal{P}\\) represents the set of all possible distributions that are consistent with the structural causal model, and the supremum operator \\(\\sup_{P\\in \\mathcal{P}}\\) accounts for the worst-case scenario over these distributions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.906, + 0.747 + ], + "angle": 0, + "content": "In our study, instead of training the model from scratch, we define \\(\\theta\\) as the pretrained model weights. We subsequently employ variable interventions to validate the autoregressive properties of the model, facilitating the filtration of extraneous information. This strategy aims to consider only causal representation parameters \\(\\gamma\\) in the quality prediction process while effectively ignoring the noise component \\(\\eta\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.762, + 0.649, + 0.779 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.641, + 0.802 + ], + "angle": 0, + "content": "4.1. Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.9 + ], + "angle": 0, + "content": "As illustrated in Figure 1 (a), the proposed framework integrates multiple stages, starting with the input of reference and distorted images. These images are first processed in the deep domain, where deep learning models extract high-level feature representations. These feature representations serve as the foundation for the subsequent causal inference" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17992" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.095, + 0.56, + 0.112 + ], + "angle": 0, + "content": "(a) Framework" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.112, + 0.899, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.311, + 0.537, + 0.328 + ], + "angle": 0, + "content": "(b) Confounder Dictionary Embedding" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.334, + 0.653, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.31, + 0.897, + 0.327 + ], + "angle": 0, + "content": "(a(c) Structural Causal Model)" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.33, + 0.867, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.515, + 0.875, + 0.529 + ], + "angle": 0, + "content": "Figure 1. The framework of our proposed method that separates and leverages causal features for perceptual quality assessment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.483, + 0.662 + ], + "angle": 0, + "content": "stage. In this stage, the relationships between the extracted features are analysed to identify causal variables that are robust to distortions and potential confounders. Finally, the quality assessment step uses these identified causal variables to evaluate the perceptual quality of the distorted image, producing a quality score that reflects the minimal perceptual cost on the basis of the underlying causal structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To explain the causal relationships considered in our approach, we employ a structural causal model (SCM) for the full-reference image quality assessment. The SCM represents variables as nodes and relationships as edges in a graphical structure, using directed arrows to indicate causal relationships between two variables. In our model, we incorporate the reference image \\(I\\), the distorted image \\(D\\), the representations of both images \\(\\phi_{\\theta}(I)\\) and \\(\\phi_{\\theta}(D)\\), and the image quality score \\(Q\\), as depicted in Figure 1 (c). We assume that the images are characterized by factors that have causal relationships with quality, along with elements that are unrelated to quality. In this framework, image quality is directly determined by factors that causally influence it. Simply using a training process or pretrained weights from other fields to assess image quality typically yields repre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.907, + 0.631 + ], + "angle": 0, + "content": "sentations that are correlated with perceptual quality but do not capture causal relationships. Therefore, we introduce an unknown exogenous variable \\( U \\) to account for uncertainties and identify the representations that are causally related to quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.642, + 0.757, + 0.658 + ], + "angle": 0, + "content": "4.2. Deep Causal Measurement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.907, + 0.783 + ], + "angle": 0, + "content": "In our approach, to verify causal relationships in perceptual quality assessment, we design a causal validation process using an SCM with an exogenous variable \\( U \\). This process begins with a causal intervention to test the effect of image characteristics on perceived quality, using the operator \\( \\mathrm{do}(\\cdot) \\) to map both reference and distorted images onto a shared causal feature space. The intervention can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.797, + 0.907, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {I} = \\phi_ {\\theta} (\\mathrm {d o} (I = i)), \\quad \\mathbf {f} _ {D} = \\phi_ {\\theta} (\\mathrm {d o} (D = d)) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\phi_{\\theta}\\) is the feature extraction function parameterized by \\(\\theta\\), which maps the images into the deep feature space, and \\(\\mathbf{f}\\) denotes the deep features. To substantiate the causal nature of the features, we introduce the exogenous variable \\(U\\) and perform simultaneous interventions \\(v\\) on both the original" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17993" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.15 + ], + "angle": 0, + "content": "reference image features \\(\\phi_{\\theta}(I)\\) and distorted image features \\(\\phi_{\\theta}(D)\\) to generate the intervened causal representations \\(\\mathbf{f}'\\). This intervention yields representations formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.162, + 0.483, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {I} ^ {\\prime} = \\phi_ {\\theta} (\\mathrm {d o} (I, v)), \\quad \\mathbf {f} _ {D} ^ {\\prime} = \\phi_ {\\theta} (\\mathrm {d o} (D, v)) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.188, + 0.484, + 0.293 + ], + "angle": 0, + "content": "Next, we calculate the perceptual distance \\(\\mathrm{Dis}(\\cdot)\\) before and after the intervention to assess the invariance of the causal representation. If the perceptual distance differs significantly between the pre- and postintervention states, then the feature representation can be deemed causal, as it is sensitive to distortions influencing perceptual quality. This invariance can be evaluated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.303, + 0.483, + 0.321 + ], + "angle": 0, + "content": "\\[\n\\Delta = \\operatorname {D i s} \\left(\\mathbf {f} _ {I}, \\mathbf {f} _ {D}\\right) - \\operatorname {D i s} \\left(\\mathbf {f} _ {I} ^ {\\prime}, \\mathbf {f} _ {D} ^ {\\prime}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.329, + 0.484, + 0.478 + ], + "angle": 0, + "content": "If \\(\\Delta \\neq 0\\), it confirms that the causal factors within the feature representations significantly affect perceptual quality. In practice, we vary the disturbance intensity to determine the strength of the perceptual causal relationship within the feature representations, and we use a confounder dictionary to record the representations that maintain causality at all intensities, denoted \\(\\Gamma (\\mathbf{f}_I,\\mathbf{f}_D)\\). The causal influence on perceptual quality is then quantified by observing how the perceptual distance changes with varying intervention intensities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.484, + 0.526 + ], + "angle": 0, + "content": "The regression invariance is validated if the perceptual distance \\(\\mathrm{Dis}(\\mathbf{f}_I',\\mathbf{f}_D')\\) yields a consistent quality score \\(Q\\) under intervention, where \\(\\mathbf{f}\\) is parameterized by \\(\\gamma\\), such that:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.535, + 0.483, + 0.553 + ], + "angle": 0, + "content": "\\[\nQ = \\mathbb {E} [ Q | \\gamma , v ] \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.637 + ], + "angle": 0, + "content": "This indicates that the quality prediction model relies solely on the causal features \\( \\mathbf{f} \\), with \\( \\gamma \\) representing the parameter of \\( \\phi \\) in this context, excluding the influence of noise \\( \\eta \\). The SCM, therefore, ensures that the quality assessment is both robust to noise and aligned with causal perceptual features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.484, + 0.773 + ], + "angle": 0, + "content": "Finally, on the basis of the SCM, the perceptual differences between the reference image and the distorted image are quantified via a distance metric. Our method is predicated on the concept of perceptual cost, facilitating a quantitative analysis of the differences in perceived quality between reference and distorted images. Specifically, we define the perceptual cost between two distributions, \\( P_{X} \\) and \\( P_{Y} \\), utilizing the causal transport cost, which is formally expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.781, + 0.483, + 0.825 + ], + "angle": 0, + "content": "\\[\nC O T \\left(P _ {X}, P _ {Y}\\right) = \\inf _ {g \\in G \\left(P _ {X}, P _ {Y}\\right)} \\int \\Gamma (x, y) c (x, y) d g (x, y) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.9 + ], + "angle": 0, + "content": "where \\( G(P_X, P_Y) \\) represents the joint distribution of the reference and distorted images, \\( c(x, y) \\) is the 12 norm of the difference between the distributions, and \\( \\Gamma(x, y) \\) denotes the causal confounder dictionary between the two distributions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.872, + 0.107 + ], + "angle": 0, + "content": "4.3. Connection to Existing Methods and HVS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.905, + 0.37 + ], + "angle": 0, + "content": "Existing deep learning-based methods [4, 17, 39] leverage high-level feature representations from deep neural networks to improve perceptual alignment by statistically correlating feature distances with quality scores. LPIPS [39] measures the distance between deep network features extracted from image patches, whereas DeepWSD [17] leverages the Wasserstein distance within the deep feature space to evaluate perceptual similarity. DISTS [4], on the other hand, assesses the similarity between image structures and textures by comparing deep feature representations. However, they still rely primarily on correlations and statistical dependencies rather than causal mechanisms, which can limit their ability to generalize effectively across diverse datasets and distortion types. In contrast, our approach introduces a causal framework that aims to reveal the underlying mechanisms driving perceptual differences, providing a more robust and theoretically grounded approach to IQA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.907, + 0.596 + ], + "angle": 0, + "content": "HVS is known to be sensitive to specific perceptual features (e.g., contrast, texture, and structural integrity) that are often distorted by various image processing operations. Previous studies have attempted to model HVS by designing perceptual metrics or using deep learning to automatically learn these perceptual features. However, these models usually fail to distinguish between causal features that actually affect perceptual quality and incidental noise that may be captured by deep features. By introducing a causal model that uses an intervention-based framework, we provide a clearer path for identifying causal features that determine perceptual quality, providing insight into the relationship between image content distortion and perceptual quality in a way that is more consistent with the mechanisms of HVS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.612, + 0.72, + 0.629 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.722, + 0.654 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The study employs six widely used IQA databases to evaluate the effectiveness of the proposed method. These databases include the LIVE IQA [27] database, CSIQ [16], TID2008 [21], TID2013 [22], KADID [19], and PIPAL [10], each containing different types of image distortions and varying numbers of reference images. LIVE [27] comprises 29 reference images and 779 distorted images across five types of distortions. It provides quality scores on a scale of [0, 100]. Containing 30 reference images and 866 distorted images, CSIQ [16] covers six distortion types, with subjective quality scores normalized to the range [0, 1]. TID2008 [21] and TID2013 [22] include 25 reference images. TID2008 has 1,700 distorted images with 17 types of distortions, whereas TID2013 expands to 3,000 distorted images and 24 distortion types. Both databases use a quality score range of [0, 9]. KADID [19] provides a large-scale" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17994" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.253, + 0.09, + 0.744, + 0.104 + ], + "angle": 0, + "content": "Table 1. Results of the performance comparison conducted on benchmark datasets." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.116, + 0.903, + 0.434 + ], + "angle": 0, + "content": "
MethodLIVE [27]CSIQ [16]TID2008 [21]TID2013 [22]KADID [19]PIPAL [10]
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
PSNR0.7810.8010.7920.8070.5070.5250.6640.6870.6700.6760.3980.392
SSIM [33]0.8470.8510.8100.8330.6250.6240.6650.6270.6100.6190.4890.486
MS-SSIM [32]0.8860.9030.8750.8790.8420.8540.8310.7860.8240.8260.5710.545
VIF [26]0.9490.9530.8990.8990.7980.7490.7710.6770.6850.6790.5720.545
MAD [16]0.9040.9070.9220.9220.6810.7080.7370.7430.7170.7260.6140.591
FSIM [37]0.9100.9200.9020.9150.8750.8840.8760.8510.8520.8540.5970.573
VSI [38]0.8770.8990.9120.9290.8710.8950.8980.8950.8770.8780.5480.526
GMSD [34]0.9090.9100.9380.9390.8780.8910.8580.8040.8470.8470.6140.569
NLPD [15]0.8820.8890.9130.9260.8660.8770.8320.7990.8090.8120.4890.464
PieAPP [23]0.8660.8650.8640.8830.7520.7970.8090.8440.8570.8650.7020.701
LPIPS [39]0.8660.8630.8910.8950.7220.7180.7130.7130.8380.8370.5840.584
DISTS [4]0.9240.9250.9190.9200.8290.8140.8540.8300.886*0.886*0.6450.627
DeepWSD [17]0.9040.9250.9410.9500.9000.9040.8940.8740.8870.8880.5030.500
TOPIQ-FR [2]0.8820.8870.8940.8940.8730.8720.8540.8200.8960.8950.837*0.809*
Our-VGG0.9290.9320.9490.9520.9050.8970.9090.8840.8980.8990.6750.657
Our-ResNet0.9140.9210.9160.9230.8590.8680.8620.8520.8880.8900.5380.524
Our-EffNet0.9270.9320.9330.9380.9030.8990.8990.8790.9050.9070.6940.656
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.114, + 0.437, + 0.907, + 0.484 + ], + "angle": 0, + "content": "Notably, DISTS and TOPIQ-FR were trained on the KADID [19] and PIPAL [10], respectively, which are marked with an asterisk (*) in the table for these methods. PieAPP and LPIPS were individually trained on their proposed datasets [23] and [39], respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.483, + 0.703 + ], + "angle": 0, + "content": "collection of 81 reference images and 10,125 distorted images, spanning 25 different distortion types. The quality scores fall within the interval [1, 5]. As one of the most extensive datasets, PIPAL [10] includes 250 reference images and 25,850 distorted images across 40 distortion types. The subjective scores range from 868-1857. Notably, we used only 24200 images from the training and validation sets since the labels of the test set were not available. Given the diversity in score ranges across these databases, a linear normalization process is applied to standardize the MOS values, ensuring consistency in the evaluation process. This step ensures that a fair comparison of the quality of predictions can be made across different datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this study, we evaluate a variety of representative and state-of-the-art image quality assessment techniques, including both feature-based and deep learning approaches, including the PSNR, SSIM [33], MS-SSIM [32], VIF [26], MAD [16], FSIM [37], VSI [38], GMSD [34], NLPD [15], PieAPP [23], LPIPS [39], DISTS [4], DeepWSD [17], and TOPIQ-FR [2]. All the comparative methods were re-evaluated in our study. Specifically, for the deep learning-based approaches, we utilized the weight files provided by the authors of these methods to assess their performance through cross-dataset predictions. The training datasets used for these weights are indicated in the table. For the deep domain methods that do not require training, we fol" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.905, + 0.644 + ], + "angle": 0, + "content": "lowed the same protocol as traditional methods and conducted tests directly on the entire dataset. To evaluate the performance of these methods, we employ the Pearson linear correlation coefficient (PLCC) and the Spearman rank correlation coefficient (SRCC) as the key metrics. The PLCC is used to assess the accuracy of the predictions, whereas the SRCC evaluates the monotonic relationship by examining the rank correlation between the predicted quality scores and the corresponding MOS values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.905, + 0.751 + ], + "angle": 0, + "content": "Our proposed method is training-free and can be seamlessly applied to various deep learning models for image quality assessment. Specifically, we utilize VGG [29], ResNet [8] and EfficientNet [31] as backbone networks to extract deep feature representations. These networks are pretrained on large-scale datasets; thus, the features can be utilized to assess image quality without additional training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.796, + 0.773 + ], + "angle": 0, + "content": "5.2. Quality Prediction Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We conduct quality prediction experiments to evaluate the performance and generalizability of our model in comparison with several state-of-the-art methods. The methods we compare can be categorized into three main groups: traditional image quality assessment methods, end-to-end deep learning models, and deep feature-based similarity (or difference) metrics that do not require training. As presented in Table 1 and visualized in the scatter plot in Fig-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17995" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.368, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.175, + 0.207, + 0.185 + ], + "angle": 0, + "content": "(1) SSIM / LIVE" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.089, + 0.508, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.175, + 0.481, + 0.186 + ], + "angle": 0, + "content": "(3) SSIM/TID2008" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.09, + 0.64, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.175, + 0.613, + 0.186 + ], + "angle": 0, + "content": "(4) SSIM / TID2013" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.09, + 0.779, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.175, + 0.745, + 0.186 + ], + "angle": 0, + "content": "(5) SSIM / KADID" + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.09, + 0.901, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.175, + 0.875, + 0.186 + ], + "angle": 0, + "content": "(6) SSIM / PIPAL" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.186, + 0.244, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.186, + 0.364, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.186, + 0.501, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.186, + 0.639, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.186, + 0.777, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.186, + 0.901, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.273, + 0.209, + 0.283 + ], + "angle": 0, + "content": "(7) DISTS/LIVE" + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.273, + 0.341, + 0.283 + ], + "angle": 0, + "content": "(8) DISTS / CSIQ" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.273, + 0.499, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.273, + 0.619, + 0.283 + ], + "angle": 0, + "content": "(10) DISTS/TID2013" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.273, + 0.751, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.759, + 0.273, + 0.88, + 0.283 + ], + "angle": 0, + "content": "(12) DISTS/PIPAL" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.284, + 0.244, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.284, + 0.364, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.284, + 0.501, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.284, + 0.639, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.284, + 0.778, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.284, + 0.901, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.371, + 0.221, + 0.381 + ], + "angle": 0, + "content": "(13) DeepWSD / LIVE" + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.371, + 0.353, + 0.381 + ], + "angle": 0, + "content": "(14) DeepWSD / CSIQ" + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.371, + 0.495, + 0.381 + ], + "angle": 0, + "content": "(15) DeepWSD/TID2008" + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.371, + 0.631, + 0.381 + ], + "angle": 0, + "content": "(16) DeepWSD / TID2013" + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.371, + 0.759, + 0.381 + ], + "angle": 0, + "content": "(17) DeepWSD / KADID" + }, + { + "type": "image_caption", + "bbox": [ + 0.778, + 0.371, + 0.89, + 0.381 + ], + "angle": 0, + "content": "(18) DeepWSD / PIPAL" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.381, + 0.244, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.381, + 0.364, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.381, + 0.501, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.381, + 0.639, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.381, + 0.778, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.381, + 0.901, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.468, + 0.221, + 0.478 + ], + "angle": 0, + "content": "(19) TOPIQ-FR / LIVE" + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.468, + 0.353, + 0.478 + ], + "angle": 0, + "content": "(20) TOPIQ-FR / CSIQ" + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.468, + 0.495, + 0.478 + ], + "angle": 0, + "content": "(21) TOPIQ-FR/TID2008" + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.468, + 0.631, + 0.478 + ], + "angle": 0, + "content": "(22) TOPIQ-FR / TID2013" + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.468, + 0.759, + 0.478 + ], + "angle": 0, + "content": "(23) TOPIQ-FR / KADID" + }, + { + "type": "image_caption", + "bbox": [ + 0.778, + 0.468, + 0.89, + 0.478 + ], + "angle": 0, + "content": "(24) TOPIQ-FR / PIPAL" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.479, + 0.244, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.566, + 0.207, + 0.576 + ], + "angle": 0, + "content": "(25) Ours / LIVE" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.479, + 0.364, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.566, + 0.34, + 0.576 + ], + "angle": 0, + "content": "(26) Ours / CSIQ" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.479, + 0.501, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.566, + 0.481, + 0.576 + ], + "angle": 0, + "content": "(27) Ours / TID2008" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.479, + 0.639, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.566, + 0.614, + 0.576 + ], + "angle": 0, + "content": "(28) Ours / TID2013" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.479, + 0.778, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.566, + 0.745, + 0.576 + ], + "angle": 0, + "content": "(29) Ours / KADID" + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.479, + 0.901, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.566, + 0.876, + 0.576 + ], + "angle": 0, + "content": "(30) Ours / PIPAL" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.589, + 0.908, + 0.631 + ], + "angle": 0, + "content": "Figure 2. Scatterplot of prediction results for different methods across databases. We performed normalization on both labels and predictions and 4-parameter logistic function is utilized before comparison, as suggested in [24], and the values of the lower-better methods are reversed before regression." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.485, + 0.9 + ], + "angle": 0, + "content": "ure 2, our method demonstrates superior overall performance and consistently aligns closely with actual scores across datasets. Traditional techniques, which are typically based on handcrafted features, often struggle to capture the complex, nonlinear relationships between image distortions and perceptual quality, resulting in relatively lower accuracy across diverse datasets. End-to-end deep learning methods, while showing a marked improvement over traditional methods, exhibit notable limitations in terms of generalizability. These methods are typically trained on specific datasets, and as a result, their performance tends to degrade when evaluated on datasets that differ from the training set, struggling to maintain consistency and accuracy when applied to unseen data. Furthermore, the deep feature-based similarity methods that do not rely on training also perform well, although they tend to be less precise" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.885 + ], + "angle": 0, + "content": "than our model is. These methods capture image distortions via pretrained features, but their inability to explicitly model causal relationships results in slightly less accurate predictions, particularly when the perceptual impact of distortions varies significantly across datasets. In contrast, our approach, which leverages causal reasoning through an SCM and does not require training, excels in maintaining consistent performance across various datasets. The results demonstrate that, by focusing on causally relevant features, our model is less sensitive to the idiosyncrasies of individual datasets and is capable of producing accurate predictions for unseen images. This robustness in generalization suggests that the causal framework used in our model provides a more reliable representation of perceptual quality, independent of dataset-specific biases." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17996" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 2. Performance comparison results of the ablation experiments." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.13, + 0.489, + 0.228 + ], + "angle": 0, + "content": "
ModelLIVE [27]CSIQ [16]TID2013 [22]
ParameterPLCCSRCCPLCCSRCCPLCCSRCC
φ = φθ0.9010.9150.9130.9160.8840.867
φ = φγ0.9290.9320.9490.9520.9090.884
φ = φη0.8430.8660.8030.8310.7860.789
" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.231, + 0.483, + 0.327 + ], + "angle": 0, + "content": "Performance comparison results of the ablation experiments conducted on three benchmark datasets, where we use OurVGG as the default architecture. The models evaluated in this study correspond to different configurations of weight assignments for \\(\\phi\\), where its values \\(\\phi_{\\theta}\\), \\(\\phi_{\\gamma}\\), and \\(\\phi_{\\eta}\\) represent the original pretrained weights, the causal effect weights, and the residual component excluding the causal effect, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.35, + 0.265, + 0.367 + ], + "angle": 0, + "content": "5.3. Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.372, + 0.483, + 0.645 + ], + "angle": 0, + "content": "In this section, we present the abductive counterfactual experiments to evaluate the impact of different weight configurations in our model. Specifically, we analyse the results when varying the weights of \\(\\theta\\), \\(\\gamma\\), and \\(\\eta\\). Here, \\(\\theta\\) represents the original pretrained weights, \\(\\gamma\\) corresponds to the weight associated with the causal effect, and \\(\\eta\\) denotes the remaining part of the model, which is the portion of \\(\\theta\\) excluding \\(\\gamma\\). The ablation analysis is conducted by systematically adjusting each of these weights and observing their influence on the model's performance. The results are summarized in Table 2, which indicate that although noncausal representations also exhibit some perceptual relevance, their lack of causal relationships limits their effectiveness. When these noncausal representations are removed from the original feature set, the model's predictive performance improves. This suggests that incorporating those features that capture causal effects can enhance the model's ability to make more accurate and reliable quality predictions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.303, + 0.669 + ], + "angle": 0, + "content": "5.4. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we visually demonstrate the effect of causal intervention on different channels. As shown in Figure 3, we demonstrate the visualization effects of causal intervention between the same reference image and different distorted images (blurred and additive Gaussian noise) via the Our-VGG architecture. These visualizations reveal the varying levels of sensitivity each perceptual channel exhibits to different types of distortions in the pretrained model, thus providing a clear indication of how such distortions alter the perceptual characteristics of the image. The horizontal axis in Figure 3 represents the channels, and the color intensity corresponds to the strength of the causal relationship, with brighter colors indicating a stronger causal relationship. Causal features reflect attributes that are crucial for human perception of image quality, as they directly" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.623, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.175, + 0.582, + 0.184 + ], + "angle": 0, + "content": "Blur" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.92, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.203, + 0.623, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.287, + 0.605, + 0.296 + ], + "angle": 0, + "content": "Gaussian Noise" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.2, + 0.92, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.264, + 0.918, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.312, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Effects of causal interventions on the perceptual representations of images under various distortions. The displayed channels, from top to bottom, correspond to the second, fourth, and fifth stages of the VGG-16 architecture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.907, + 0.591 + ], + "angle": 0, + "content": "influence perceptual judgments. In contrast, noncausal features typically correspond to higher-level semantic aspects of the image, such as object recognition, which are less relevant for perceptual quality but play a significant role in high level vision tasks. This distinction is supported by the observation that deeper layers of the network are more attuned to classification tasks, while quality-related features are predominantly represented in the shallower layers, underscoring the channel-specific nature of perceptual changes. By isolating the causal effects, we can identify specific features or regions of the image that are most sensitive to certain types of distortions, thus further elucidating how different distortions affect overall image perception." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.634, + 0.62 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In conclusion, we proposed an FR-IQA method grounded in abductive counterfactual inference to analyse the causal relationships between deep network features and perceptual distortions. We investigated the causal effects of deep features on human perception and incorporated causal reasoning with feature comparisons, constructing a model capable of addressing various complex distortions in different IQA scenarios. The proposed approach demonstrated strong generalizability, as the analysis of perceptual causal correlations was shown to be independent of the backbone architecture, making it adaptable across diverse deep networks. Additionally, abductive counterfactual experiments validated the effectiveness of the proposed causal framework, confirming the model's superior perceptual relevance and interpretability of quality scores. The experimental results reinforced the robustness and competitiveness of the method, which achieved consistent performance across multiple IQA benchmarks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17997" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Hengrui Cai, Yixin Wang, Michael Jordan, and Rui Song. On learning necessary and sufficient causal graphs. In Advances in Neural Information Processing Systems, pages 42148-42160. Curran Associates, Inc., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 33:2404-2418, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Li-Heng Chen, Christos G. Bampis, Zhi Li, Andrey Norkin, and Alan C. Bovik. Proxiqa: A proxy approach to perceptual optimization of learned image compression. IEEE Transactions on Image Processing, 30:360-373, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[4] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[5] Keyan Ding, Yi Liu, Xueyi Zou, Shiqi Wang, and Kede Ma. Locally adaptive structure and texture similarity for image quality assessment. In Proceedings of the 29th ACM International Conference on Multimedia, page 2483-2491, New York, NY, USA, 2021. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[6] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Comparison of full-reference image quality models for optimization of image processing systems. International Journal of Computer Vision, 129(4):1258-1281, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.483, + 0.563 + ], + "angle": 0, + "content": "[7] Fei Gao, Yi Wang, Panpeng Li, Min Tan, Jun Yu, and Yani Zhu. Deepsim: Deep similarity for image quality assessment. Neurocomputing, 257:104-114, 2017. Machine Learning and Signal Processing for Big Multimedia Analysis. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[9] Linmei Hu, Ziwei Chen, Ziwang Zhao, Jianhua Yin, and Liqiang Nie. Causal inference for leveraging image-text matching bias in multi-modal fake news detection. IEEE Transactions on Knowledge and Data Engineering, 35(11): 11141-11152, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[10] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S. Ren, and Dong Chao. Pipal: A large-scale image quality assessment dataset for perceptual image restoration. In Computer Vision - ECCV 2020, pages 633-651, Cham, 2020. Springer International Publishing. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[11] Jongyoo Kim and Sanghoon Lee. Deep learning of human visual sensitivity in image quality assessment framework. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1676-1684, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Junho Kim, Byung-Kwan Lee, and Yong Man Ro. Demystifying causal features on adversarial examples and causal inoculation for robust network by adversarial instrumental variable regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12302-12312, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[13] Xuting Lan, Mingliang Zhou, Xueyong Xu, Xuekai Wei, Xingran Liao, Huayan Pu, Jun Luo, Tao Xiang, Bin Fang, and Zhaowei Shang. Multilevel feature fusion for end-to-end blind image quality assessment. IEEE Transactions on Broadcasting, 69(3):801-811, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[14] Shanshan Lao, Yuan Gong, Shuwei Shi, Sidi Yang, Tianhe Wu, Jiahao Wang, Weihao Xia, and Yujiu Yang. Attentions help cnns see better: Attention-based hybrid image quality assessment network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1140-1149, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[15] Valero Laparra, Johannes Balle, Alexander Berardino, and Eero P. Simoncelii. Perceptual image quality assessment using a normalized laplacian pyramid. In Human Vision and Electronic Imaging, pages 43-48, United States, 2016. Society for Imaging Science and Technology. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[16] Eric Cooper Larson and Damon Michael Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of Electronic Imaging, 19 (1):011006, 2010. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[17] Xingran Liao, Baoliang Chen, Hanwei Zhu, Shiqi Wang, Mingliang Zhou, and Sam Kwong. Deepwsd: Projecting degradations in perceptual space to wasserstein distance in deep feature space. In Proceedings of the 30th ACM International Conference on Multimedia, page 970-978, New York, NY, USA, 2022. Association for Computing Machinery. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[18] Xingran Liao, Xuekai Wei, Mingliang Zhou, Zhengguo Li, and Sam Kwong. Image quality assessment: Measuring perceptual degradation via distribution measures in deep feature spaces. IEEE Transactions on Image Processing, 33:4044-4059, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[19] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience, pages 1–3, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[20] Bing Liu, Dong Wang, Xu Yang, Yong Zhou, Rui Yao, Zhiwen Shao, and Jiaqi Zhao. Show, deconfound and tell: Image captioning with causal inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18041-18050, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[21] Nikolay Ponomarenko, Vladimir Lukin, Alexander Zelensky, Karen Egiazarian, Marco Carli, and Federica Battisti. Tid2008-a database for evaluation of full-reference visual quality assessment metrics. Advances of Modern Radioelectronics, 10(4):30-45, 2009. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[22] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, and C.-C. Jay Kuo. Image database tid2013: Peculiarities, results and perspectives. Signal Processing: Image Communication, 30: 57-77, 2015. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[23] Ekta Prashnani, Hong Cai, Yasamin Mostofi, and Pradeep Sen. Pieapp: Perceptual image-error assessment through pairwise preference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1808-1817, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17998" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.243 + ], + "angle": 0, + "content": "[24] Ann Marie Rohaly, Philip J. Corriveau, John M. Libert, Arthur A. Webster, Vittorio Baroncini, John Beerends, Jean-Louis Blin, Laura Contin, Takahiro Hamada, David Harrison, Andries P. Hekstra, Jeffrey Lubin, Yukihiro Nishida, Ricardo Nishihara, John C. Pearson, Antonio Franca Pessoa, Neil Pickford, Alexander Schertz, Massimo Visca, Andrew B. Watson, and Stefan Winkler. Video Quality Experts Group: current results and future directions. In Visual Communications and Image Processing 2000, pages 742 - 753. International Society for Optics and Photonics, SPIE, 2000." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.285 + ], + "angle": 0, + "content": "[25] Ladan Shams and Ulrik R. Beierholm. Causal inference in perception. Trends in Cognitive Sciences, 14(9):425-432, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[26] H.R. Sheikh and A.C. Bovik. Image information and visual quality. IEEE Transactions on Image Processing, 15(2):430-444, 2006. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.329, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[27] H.R. Sheikh, M.F. Sabir, and A.C. Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on Image Processing, 15(11): 3440-3451, 2006. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[28] Wenhao Shen, Mingliang Zhou, Jun Luo, Zhengguo Li, and Sam Kwong. Graph-represented distribution similarity index for full-reference image quality assessment. IEEE Transactions on Image Processing, 33:3075-3089, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[29] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, page 1-14, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[30] Xue Song, Jiequan Cui, Hanwang Zhang, Jingjing Chen, Richang Hong, and Yu-Gang Jiang. Doubly abductive counterfactual inference for text-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9162-9171, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[31] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[32] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems and Computers, 2003, pages 1398-1402. IEEE, 2003. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[33] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[34] Wufeng Xue, Lei Zhang, Xuanqin Mou, and Alan C. Bovik. Gradient magnitude similarity deviation: A highly efficient perceptual image quality index. IEEE Transactions on Image Processing, 23(2):684-695, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[35] Mengyue Yang, Xinyu Cai, Furui Liu, Weinan Zhang, and Jun Wang. Specify robust causal representation from mixed observations. New York, NY, USA, 2023. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Zhongqi Yue, Hanwang Zhang, Qianru Sun, and Xian-Sheng Hua. Interventional few-shot learning. In Advances in Neural Information Processing Systems, pages 2734-2746. Curran Associates, Inc., 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[37] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. Fsim: A feature similarity index for image quality assessment. IEEE Transactions on Image Processing, 20(8):2378-2386, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[38] Lin Zhang, Ying Shen, and Hongyu Li. Vsi: A visual saliency-induced index for perceptual image quality assessment. IEEE Transactions on Image Processing, 23(10): 4270-4281, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[39] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[40] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10242-10251, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[41] Mingliang Zhou, Shujun Lang, Taiping Zhang, Xingran Liao, Zhaowei Shang, Tao Xiang, and Bin Fang. Attentional feature fusion for end-to-end blind image quality assessment. IEEE Transactions on Broadcasting, 69(1):144-152, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17999" + } + ] +] \ No newline at end of file diff --git a/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/5e0e415c-9773-49a3-97ea-205f1760b866_origin.pdf b/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/5e0e415c-9773-49a3-97ea-205f1760b866_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..94701d16f26c77a6cff462bce31b58d7e7abd8fa --- /dev/null +++ b/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/5e0e415c-9773-49a3-97ea-205f1760b866_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6dcd809f88569a4efb54321ba7c173b951a05975fcfebfb415b9788d085fab25 +size 5677399 diff --git a/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/full.md b/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a5f7f58c2c1b257aa7b5dec1fa69448858b26270 --- /dev/null +++ b/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/full.md @@ -0,0 +1,340 @@ +# Image Quality Assessment: Investigating Causal Perceptual Effects with Abductive Counterfactual Inference + +Wenhao Shen $^{1}$ Mingliang Zhou $^{1*}$ Yu Chen $^{1}$ Xuekai Wei $^{1}$ Yong Feng $^{1}$ Huayan Pu $^{2}$ Weijia Jia $^{3}$ \ + $^{1}$ College of Computer Science, Chongqing University\ + $^{2}$ State Key Laboratory of Mechanical Transmissions, Chongqing University\ + $^{3}$ BNU-UIC Institute of Artificial Intelligence and Future Networks, Beijing Normal University + +shenwenhao163@163.com mingliangzhou@cqu.edu.cn yuchen@stu.cqu.edu.cn xuekaiwei2-c@my.cityu.edu.hk fengyong@cqu.edu.cn phygood_2001@shu.edu.cn jiawj@bnu.edu.cn + +# Abstract + +Existing full-reference image quality assessment (FR-IQA) methods often fail to capture the complex causal mechanisms that underlie human perceptual responses to image distortions, limiting their ability to generalize across diverse scenarios. In this paper, we propose an FR-IQA method based on abductive counterfactual inference to investigate the causal relationships between deep network features and perceptual distortions. First, we explore the causal effects of deep features on perception and integrate causal reasoning with feature comparison, constructing a model that effectively handles complex distortion types across different IQA scenarios. Second, the analysis of the perceptual causal correlations of our proposed method is independent of the backbone architecture and thus can be applied to a variety of deep networks. Through abductive counterfactual experiments, we validate the proposed causal relationships, confirming the model's superior perceptual relevance and interpretability of quality scores. The experimental results demonstrate the robustness and effectiveness of the method, providing competitive quality predictions across multiple benchmarks. The source code is available at https://anonymous.4open.science/r/DeepCausalQuality-25BC. + +# 1. Introduction + +Image quality assessment (IQA) is a fundamental area within image processing that is dedicated to evaluating the perceptual quality of images. The primary objective of IQA is to quantify how closely an image aligns with the human visual system (HVS), particularly when subjected to various distortions such as compression, noise, or blurring. Given its importance, IQA is integral to numerous applications, including image compression, transmission, and enhancement [3, 6, 40], where preserving visual fidelity is critical. IQA methods are generally classified into three main categories: full-reference, reduced-reference, and no-reference approaches, all of which can be considered typical regression problems [11, 13, 41]. + +In full-reference image quality assessment (FR-IQA), the evaluation process is typically divided into three key stages: feature decomposition, feature comparison, and perceptual score mapping. Recent advancements have highlighted the pivotal role that deep learning methods play in aligning IQA predictions with human subjective scores [7, 14]. Notably, deep representations of images have been shown to strongly correlate with perceptual relevance, emphasizing their significance in quality assessment [4, 28, 39]. In deep networks, different channels capture distinct features of original and distorted images, collectively encoding perceptual information. When quality degradation occurs, these features in the distorted image become contaminated to varying degrees, resulting in altered distinctions from those of the reference image [18]. However, the mere alteration of features is insufficient to fully capture the perceptual quality differences between the original and distorted images, as changes in correlation alone do not adequately explain the complex, nonlinear effects of distortions on human visual perception. Furthermore, while existing methods are effective in capturing statistical dependencies, they fail to reflect the intrinsic mecha + +nisms that drive changes in perceived quality. Specifically, these methods only quantify how distortions affect the similarity between feature representations and cannot elucidate how these distortions influence human perception. Consequently, they are unable to differentiate between distortions that may exhibit similar feature distances yet have significantly different perceptual impacts. This limitation restricts their capacity to provide accurate quality assessments. Additionally, no existing theoretical framework fully explains why IQA remains such a challenging task or why certain methods succeed in some contexts and fail in others. This lack of clarity impedes advancements within the field. + +To address these challenges, we define IQA as a counterfactual inference problem, offering a perspective that bridges existing theoretical gaps [25]. Counterfactual reasoning provides a formalized structure for defining the minimal perceptual cost, enabling a more precise understanding of how distortions affect perceived quality. Given that the reference image serves as the baseline for the original visual content, an effective IQA method should be able to predict scores on the basis of the perceptual distortions present in degraded images. However, the manner in which distortions are perceived is essentially probabilistic and is influenced by a variety of contextual factors and prior experience. While deep learning-based methods are adept at capturing complex feature interactions, they often fail to incorporate the probabilistic relationships that underlie perceptual judgments. By constructing IQA from a causal reasoning perspective, our approach aims to quantify perceptual differences in a way that reflects these inherent uncertainties and complexities. This causal perspective provides us with not only a measure of feature similarity but also a more nuanced understanding of how different distortions affect human perception. The main contributions of our study are as follows: + +- We investigate the causal effects of deep network features on perception, effectively integrating causal relationships with feature comparisons to construct a deep network-based FR-IQA model. By employing distance comparisons among the causal relationships of deep features, the model can accurately address complex distortion types encountered across various image quality assessment scenarios, resulting in competitive quality prediction outcomes. +- The proposed method exhibits significant generality, as the analysis of perceptual causal correlations within deep features is independent of the backbone architecture. This characteristic enables the implementation of the approach across different deep backbone networks, thereby enhancing its versatility and applicability in diverse contexts of image quality assessment. +- The validity of the proposed causal relationships is substantiated through experimental evidence derived from + +abductive counterfactual inference, which demonstrates superior perceptual relevance and score interpretability and establishes its robustness and practicality. + +# 2. Related Work + +# 2.1. Deep Network-based FR-IQA + +Deep networks have been employed to model the complex relationships between reference and distorted images in FR-IQA. In [11], convolutional operations on both reference and distorted images are utilized to obtain sensitivity maps, which are then applied to the difference map for quality estimation. Gao et al. [7] measured the local similarity between features extracted from deep neural networks (DNNs) for both reference and distorted images. The local quality indices are aggregated to produce the overall quality score. Subsequent studies have shown that VGG network features serve as effective training losses for tasks such as image synthesis. Notably, the study of [39] compares deep features across various architectures and tasks, showing that perceptual similarity is an emerging property shared by deep visual representations. Hence, some advancements focus on perceptual similarity or differences in deep feature space. For example, Ding et al. [4] employed the structural similarity index (SSIM) between deep features to measure perceptual texture similarity, offering a unified deep image structure and texture similarity (DISTS) framework, and the adaptive DISTS (ADISTS) [5] improved it by using a dispersion index to identify texture regions and adaptively aggregate structure and texture measurements. In contrast, Liao et al. [17] projected image degradations into perceptual space and computed the Wasserstein distance in deep feature space, highlighting the importance of distributional differences in deep feature representations for quality assessment. + +Existing methods are limited in capturing perceptual quality differences because they rely on feature alterations and statistical dependencies and fail to model the nonlinear effects of distortions on human perception. They quantify feature similarity without revealing how distortions impact perception, leading to difficulties in distinguishing distortions with similar feature distances but different perceptual effects. Additionally, the absence of a robust theoretical framework to explain why some methods succeed or fail in different scenarios limits progress in IQA. + +# 2.2. Causal Inference in Deep Learning + +In deep learning, with the presence of confounding factors, the network can easily capture spurious correlations between inputs and their outputs. Therefore, the application of causal inference integration in deep learning has received increasing attention. Causal inference provides a theoretical foundation for disentangling correlation from + +causation, which is particularly useful for improving model generalizability, robustness, and interpretability [12] [35]. In deep learning, causal approaches have been explored to mitigate issues such as bias, confounding, and spurious correlations that often arise from data-driven methods [1, 30]. Yue et al. [36] proposed an interventional few-shot learning method that applies backdoor adjustment to mitigate the confounding effect of pretrained knowledge. Liu et al. [20] addressed both visual confounders in the object detection stage and linguistic confounders in the decoding process for image captioning tasks via causal intervention, leading to substantial improvements in generalizability across different datasets. Hu et al. [9] modelled the causal relationship between image-text matching and fake news detection in multimodal settings and used causal inference to eliminate confounding effects during training and leverage matching bias to improve detection accuracy. + +In the context of image quality assessment, causal reasoning offers a promising avenue for improving the accuracy and consistency of quality prediction. By focusing on the causal relationship between image features and perceived quality, it can explore how causal models mitigate the problems of spurious correlation and overfitting to better align with human perceptual judgments and improve the reliability of deep learning models in real-world applications. + +# 3. Problem Formulation + +The general objective in image quality assessment is to minimize the discrepancy between the predicted quality score and the subjective mean opinion score (MOS), which can be formalized as: + +$$ +\min _ {\zeta , \omega , \theta} \mathbb {E} [ l (f _ {\zeta} (\operatorname {D i s} _ {\omega} (\phi_ {\theta} (I), \phi_ {\theta} (D))), M O S) ] \tag {1} +$$ + +where $\theta$ represents the parameters of model $\phi$ , $\mathrm{Dis}(\phi_{\theta}(I),\phi_{\theta}(D))$ denotes the distance between the feature representations of the original image $I$ and the distorted image $D$ in the deep feature space $\phi (\cdot)$ , and $l(\cdot)$ is the loss function that quantifies the difference between the predicted score $f_{\zeta}(\mathrm{Dis}_{\omega}(\phi (I),\phi (D)))$ and the subjective quality. The expectation $\mathbb{E}$ is taken over the data distribution. + +For full-reference image quality assessment, we consider the causal relationship between the original image $I$ and the distorted image $D$ . The relationship can be expressed as follows: + +$$ +\mathcal {Q} _ {S} = m \left(\phi_ {\gamma} (I, D)\right), \quad \eta \perp \gamma \text {a n d} \gamma , \eta \sim F _ {\theta} \tag {2} +$$ + +Here, $m(\cdot)$ represents a function that maps the joint feature representation $\phi_{\gamma}(I,D)$ of the original and distorted images to the subjective quality $\mathcal{Q}_S$ . The parameter $\gamma$ is + +pivotal in this formulation, as it represents the features with causal implications for image quality, thereby influencing the perceptual evaluation. In contrast, $\eta$ signifies the noise components that are independent of the deep representation parameter $\gamma$ , suggesting that these components do not contribute to the quality assessment. Both $\gamma$ and $\eta$ are derived from the distribution $F_{\theta}$ , indicating their dependence on the model's parameters, $\theta$ . Specifically, $\theta$ represents the foundational parameters of the deep feature representation, which gives rise to two distinct components: $\gamma$ , the relevant features that causally affect image quality, and $\eta$ , the extraneous noise that does not contribute to the quality assessment. This causal relationship isolates the effect of distortions from external factors, ensuring that the predicted quality score is influenced primarily by distortions between the reference and distorted images, whereas other factors are captured by $\eta$ . + +In light of the established causal relationship, the subsequent phase focuses on prioritizing causal features $\gamma$ while effectively disregarding the noise components $\eta$ . The goal is to adjust the model parameters $\theta$ to better capture the relevant causal features $\gamma$ that significantly influence the subjective quality score $Q_{S}$ . The optimization process is designed to minimize the prediction error across all conceivable data distributions, thereby ensuring that the model remains robust against varying inputs and distortions. This objective can be formally articulated as follows: + +$$ +\min _ {\zeta , \omega , \theta \rightarrow \gamma} \sup _ {P \in \mathcal {P}} \mathbb {E} _ {P} \left[ l \left(f _ {\zeta} \left(\operatorname {D i s} _ {\omega} \left(\phi_ {\theta} (I), \phi_ {\theta} (D)\right)\right), M O S\right)\right] \tag {3} +$$ + +where the term $\min_{\theta \to \gamma}$ signifies the optimization of model parameters $\theta$ to enhance the representation of the causal parameters $\gamma$ , $\mathcal{P}$ represents the set of all possible distributions that are consistent with the structural causal model, and the supremum operator $\sup_{P\in \mathcal{P}}$ accounts for the worst-case scenario over these distributions. + +In our study, instead of training the model from scratch, we define $\theta$ as the pretrained model weights. We subsequently employ variable interventions to validate the autoregressive properties of the model, facilitating the filtration of extraneous information. This strategy aims to consider only causal representation parameters $\gamma$ in the quality prediction process while effectively ignoring the noise component $\eta$ . + +# 4. Methodology + +# 4.1. Framework + +As illustrated in Figure 1 (a), the proposed framework integrates multiple stages, starting with the input of reference and distorted images. These images are first processed in the deep domain, where deep learning models extract high-level feature representations. These feature representations serve as the foundation for the subsequent causal inference + +![](images/f9874e4eccd13350e1503e0f940418f97216307d9f0b82dd670817893fb4e3ce.jpg) +(a) Framework +(a(c) Structural Causal Model) + +![](images/1b1accfd3c380f58e206fafb862b7deb58fad0c9823136471836aaaaea9b7581.jpg) +(b) Confounder Dictionary Embedding +Figure 1. The framework of our proposed method that separates and leverages causal features for perceptual quality assessment. + +![](images/471c9eff8eb9f353447c3fc51cb950a20bfc64d0cf5e1aa3289cb5085d11abbf.jpg) + +stage. In this stage, the relationships between the extracted features are analysed to identify causal variables that are robust to distortions and potential confounders. Finally, the quality assessment step uses these identified causal variables to evaluate the perceptual quality of the distorted image, producing a quality score that reflects the minimal perceptual cost on the basis of the underlying causal structure. + +To explain the causal relationships considered in our approach, we employ a structural causal model (SCM) for the full-reference image quality assessment. The SCM represents variables as nodes and relationships as edges in a graphical structure, using directed arrows to indicate causal relationships between two variables. In our model, we incorporate the reference image $I$ , the distorted image $D$ , the representations of both images $\phi_{\theta}(I)$ and $\phi_{\theta}(D)$ , and the image quality score $Q$ , as depicted in Figure 1 (c). We assume that the images are characterized by factors that have causal relationships with quality, along with elements that are unrelated to quality. In this framework, image quality is directly determined by factors that causally influence it. Simply using a training process or pretrained weights from other fields to assess image quality typically yields repre + +sentations that are correlated with perceptual quality but do not capture causal relationships. Therefore, we introduce an unknown exogenous variable $U$ to account for uncertainties and identify the representations that are causally related to quality. + +# 4.2. Deep Causal Measurement + +In our approach, to verify causal relationships in perceptual quality assessment, we design a causal validation process using an SCM with an exogenous variable $U$ . This process begins with a causal intervention to test the effect of image characteristics on perceived quality, using the operator $\mathrm{do}(\cdot)$ to map both reference and distorted images onto a shared causal feature space. The intervention can be formulated as follows: + +$$ +\mathbf {f} _ {I} = \phi_ {\theta} (\mathrm {d o} (I = i)), \quad \mathbf {f} _ {D} = \phi_ {\theta} (\mathrm {d o} (D = d)) \tag {4} +$$ + +where $\phi_{\theta}$ is the feature extraction function parameterized by $\theta$ , which maps the images into the deep feature space, and $\mathbf{f}$ denotes the deep features. To substantiate the causal nature of the features, we introduce the exogenous variable $U$ and perform simultaneous interventions $v$ on both the original + +reference image features $\phi_{\theta}(I)$ and distorted image features $\phi_{\theta}(D)$ to generate the intervened causal representations $\mathbf{f}'$ . This intervention yields representations formulated as follows: + +$$ +\mathbf {f} _ {I} ^ {\prime} = \phi_ {\theta} (\mathrm {d o} (I, v)), \quad \mathbf {f} _ {D} ^ {\prime} = \phi_ {\theta} (\mathrm {d o} (D, v)) \tag {5} +$$ + +Next, we calculate the perceptual distance $\mathrm{Dis}(\cdot)$ before and after the intervention to assess the invariance of the causal representation. If the perceptual distance differs significantly between the pre- and postintervention states, then the feature representation can be deemed causal, as it is sensitive to distortions influencing perceptual quality. This invariance can be evaluated as follows: + +$$ +\Delta = \operatorname {D i s} \left(\mathbf {f} _ {I}, \mathbf {f} _ {D}\right) - \operatorname {D i s} \left(\mathbf {f} _ {I} ^ {\prime}, \mathbf {f} _ {D} ^ {\prime}\right) \tag {6} +$$ + +If $\Delta \neq 0$ , it confirms that the causal factors within the feature representations significantly affect perceptual quality. In practice, we vary the disturbance intensity to determine the strength of the perceptual causal relationship within the feature representations, and we use a confounder dictionary to record the representations that maintain causality at all intensities, denoted $\Gamma (\mathbf{f}_I,\mathbf{f}_D)$ . The causal influence on perceptual quality is then quantified by observing how the perceptual distance changes with varying intervention intensities. + +The regression invariance is validated if the perceptual distance $\mathrm{Dis}(\mathbf{f}_I',\mathbf{f}_D')$ yields a consistent quality score $Q$ under intervention, where $\mathbf{f}$ is parameterized by $\gamma$ , such that: + +$$ +Q = \mathbb {E} [ Q | \gamma , v ] \tag {7} +$$ + +This indicates that the quality prediction model relies solely on the causal features $\mathbf{f}$ , with $\gamma$ representing the parameter of $\phi$ in this context, excluding the influence of noise $\eta$ . The SCM, therefore, ensures that the quality assessment is both robust to noise and aligned with causal perceptual features. + +Finally, on the basis of the SCM, the perceptual differences between the reference image and the distorted image are quantified via a distance metric. Our method is predicated on the concept of perceptual cost, facilitating a quantitative analysis of the differences in perceived quality between reference and distorted images. Specifically, we define the perceptual cost between two distributions, $P_{X}$ and $P_{Y}$ , utilizing the causal transport cost, which is formally expressed as: + +$$ +C O T \left(P _ {X}, P _ {Y}\right) = \inf _ {g \in G \left(P _ {X}, P _ {Y}\right)} \int \Gamma (x, y) c (x, y) d g (x, y) \tag {8} +$$ + +where $G(P_X, P_Y)$ represents the joint distribution of the reference and distorted images, $c(x, y)$ is the 12 norm of the difference between the distributions, and $\Gamma(x, y)$ denotes the causal confounder dictionary between the two distributions. + +# 4.3. Connection to Existing Methods and HVS + +Existing deep learning-based methods [4, 17, 39] leverage high-level feature representations from deep neural networks to improve perceptual alignment by statistically correlating feature distances with quality scores. LPIPS [39] measures the distance between deep network features extracted from image patches, whereas DeepWSD [17] leverages the Wasserstein distance within the deep feature space to evaluate perceptual similarity. DISTS [4], on the other hand, assesses the similarity between image structures and textures by comparing deep feature representations. However, they still rely primarily on correlations and statistical dependencies rather than causal mechanisms, which can limit their ability to generalize effectively across diverse datasets and distortion types. In contrast, our approach introduces a causal framework that aims to reveal the underlying mechanisms driving perceptual differences, providing a more robust and theoretically grounded approach to IQA. + +HVS is known to be sensitive to specific perceptual features (e.g., contrast, texture, and structural integrity) that are often distorted by various image processing operations. Previous studies have attempted to model HVS by designing perceptual metrics or using deep learning to automatically learn these perceptual features. However, these models usually fail to distinguish between causal features that actually affect perceptual quality and incidental noise that may be captured by deep features. By introducing a causal model that uses an intervention-based framework, we provide a clearer path for identifying causal features that determine perceptual quality, providing insight into the relationship between image content distortion and perceptual quality in a way that is more consistent with the mechanisms of HVS. + +# 5. Experimental Results + +# 5.1. Experimental Settings + +The study employs six widely used IQA databases to evaluate the effectiveness of the proposed method. These databases include the LIVE IQA [27] database, CSIQ [16], TID2008 [21], TID2013 [22], KADID [19], and PIPAL [10], each containing different types of image distortions and varying numbers of reference images. LIVE [27] comprises 29 reference images and 779 distorted images across five types of distortions. It provides quality scores on a scale of [0, 100]. Containing 30 reference images and 866 distorted images, CSIQ [16] covers six distortion types, with subjective quality scores normalized to the range [0, 1]. TID2008 [21] and TID2013 [22] include 25 reference images. TID2008 has 1,700 distorted images with 17 types of distortions, whereas TID2013 expands to 3,000 distorted images and 24 distortion types. Both databases use a quality score range of [0, 9]. KADID [19] provides a large-scale + +Table 1. Results of the performance comparison conducted on benchmark datasets. + +
MethodLIVE [27]CSIQ [16]TID2008 [21]TID2013 [22]KADID [19]PIPAL [10]
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
PSNR0.7810.8010.7920.8070.5070.5250.6640.6870.6700.6760.3980.392
SSIM [33]0.8470.8510.8100.8330.6250.6240.6650.6270.6100.6190.4890.486
MS-SSIM [32]0.8860.9030.8750.8790.8420.8540.8310.7860.8240.8260.5710.545
VIF [26]0.9490.9530.8990.8990.7980.7490.7710.6770.6850.6790.5720.545
MAD [16]0.9040.9070.9220.9220.6810.7080.7370.7430.7170.7260.6140.591
FSIM [37]0.9100.9200.9020.9150.8750.8840.8760.8510.8520.8540.5970.573
VSI [38]0.8770.8990.9120.9290.8710.8950.8980.8950.8770.8780.5480.526
GMSD [34]0.9090.9100.9380.9390.8780.8910.8580.8040.8470.8470.6140.569
NLPD [15]0.8820.8890.9130.9260.8660.8770.8320.7990.8090.8120.4890.464
PieAPP [23]0.8660.8650.8640.8830.7520.7970.8090.8440.8570.8650.7020.701
LPIPS [39]0.8660.8630.8910.8950.7220.7180.7130.7130.8380.8370.5840.584
DISTS [4]0.9240.9250.9190.9200.8290.8140.8540.8300.886*0.886*0.6450.627
DeepWSD [17]0.9040.9250.9410.9500.9000.9040.8940.8740.8870.8880.5030.500
TOPIQ-FR [2]0.8820.8870.8940.8940.8730.8720.8540.8200.8960.8950.837*0.809*
Our-VGG0.9290.9320.9490.9520.9050.8970.9090.8840.8980.8990.6750.657
Our-ResNet0.9140.9210.9160.9230.8590.8680.8620.8520.8880.8900.5380.524
Our-EffNet0.9270.9320.9330.9380.9030.8990.8990.8790.9050.9070.6940.656
+ +Notably, DISTS and TOPIQ-FR were trained on the KADID [19] and PIPAL [10], respectively, which are marked with an asterisk (*) in the table for these methods. PieAPP and LPIPS were individually trained on their proposed datasets [23] and [39], respectively. + +collection of 81 reference images and 10,125 distorted images, spanning 25 different distortion types. The quality scores fall within the interval [1, 5]. As one of the most extensive datasets, PIPAL [10] includes 250 reference images and 25,850 distorted images across 40 distortion types. The subjective scores range from 868-1857. Notably, we used only 24200 images from the training and validation sets since the labels of the test set were not available. Given the diversity in score ranges across these databases, a linear normalization process is applied to standardize the MOS values, ensuring consistency in the evaluation process. This step ensures that a fair comparison of the quality of predictions can be made across different datasets. + +In this study, we evaluate a variety of representative and state-of-the-art image quality assessment techniques, including both feature-based and deep learning approaches, including the PSNR, SSIM [33], MS-SSIM [32], VIF [26], MAD [16], FSIM [37], VSI [38], GMSD [34], NLPD [15], PieAPP [23], LPIPS [39], DISTS [4], DeepWSD [17], and TOPIQ-FR [2]. All the comparative methods were re-evaluated in our study. Specifically, for the deep learning-based approaches, we utilized the weight files provided by the authors of these methods to assess their performance through cross-dataset predictions. The training datasets used for these weights are indicated in the table. For the deep domain methods that do not require training, we fol + +lowed the same protocol as traditional methods and conducted tests directly on the entire dataset. To evaluate the performance of these methods, we employ the Pearson linear correlation coefficient (PLCC) and the Spearman rank correlation coefficient (SRCC) as the key metrics. The PLCC is used to assess the accuracy of the predictions, whereas the SRCC evaluates the monotonic relationship by examining the rank correlation between the predicted quality scores and the corresponding MOS values. + +Our proposed method is training-free and can be seamlessly applied to various deep learning models for image quality assessment. Specifically, we utilize VGG [29], ResNet [8] and EfficientNet [31] as backbone networks to extract deep feature representations. These networks are pretrained on large-scale datasets; thus, the features can be utilized to assess image quality without additional training. + +# 5.2. Quality Prediction Performance + +We conduct quality prediction experiments to evaluate the performance and generalizability of our model in comparison with several state-of-the-art methods. The methods we compare can be categorized into three main groups: traditional image quality assessment methods, end-to-end deep learning models, and deep feature-based similarity (or difference) metrics that do not require training. As presented in Table 1 and visualized in the scatter plot in Fig- + +![](images/e591eb43172702c48b101e02e126cb0979ac1b437fbcea0281c7ed736dce9ed4.jpg) +(1) SSIM / LIVE + +![](images/17df6d65834efb588cbfb14267323c0a6a28b8917bc9a4f981171bc0f8c157f0.jpg) + +![](images/3c11132faf853ec91786e90e7e767be9de3092af51dd840676ced79b4b119236.jpg) + +![](images/764be3e05b40fb3618cb5b05975f7758f67ef4b4c4381e7c66a4b4b61f53082c.jpg) + +![](images/7a2d13c6f6eb0de52d0114df707227e30553092f7f0970ebe4cbee74c0ddfa53.jpg) + +![](images/1617eff18b50cafa62090a2e61215db5bfea5cb0963cfe8912e00db29bfe6613.jpg) + +![](images/f8b67a9889bc3e4f04fb5cf9f485a32c8f23d58d5d8fd60fb632dace9eb9a085.jpg) + +![](images/a2ca4f68a256d68ae790399d8d2b073f5e7024121d22a441aa0973a10549cfc0.jpg) +(3) SSIM/TID2008 + +![](images/e6b710b2da393771d294bc5679dd563804825a31f2a6f6aca9401467a42a713f.jpg) +(4) SSIM / TID2013 + +![](images/6a0fb8bfd8f56f13a3916b5f303ba03c7e6586e6a54dad0dccc2ad0396d31a5a.jpg) +(5) SSIM / KADID + +![](images/db9d15600bdabdde2a46486d159807d8a6983f5656985209a7512845c66df966.jpg) +(6) SSIM / PIPAL + +![](images/6a3f6e054e38771ef3c81967d2cdb75eeef73f205fdcd56391a2fb77f0c315d8.jpg) + +![](images/0e44869b1229331edce472255c8515b94a65abd3d7247e2b665226329dd70680.jpg) + +![](images/0326af9f7f5cd3f3299f1c5563883bf5c67de507bfee7338a96a31f9132b1153.jpg) +(7) DISTS/LIVE + +![](images/2d973d0691146ea437da3e6536d76325f95834f014072876cde781f803650d9e.jpg) +(8) DISTS / CSIQ + +![](images/7348994fc4b8b3c7c4cddcb6fbbc39e343407ef98ae6bf1299f3d1e8f73efe46.jpg) + +![](images/b6838f96eafc679ebafef7a3eac0cac3c5b53c10e6aa0683101a8b425498a438.jpg) +(10) DISTS/TID2013 + +![](images/1b593655c55c30bb10f4a7dc81877a2942d1241fc7959152027ffcb930ea8075.jpg) +(12) DISTS/PIPAL + +![](images/a401c25bce7c3881a1869ddce272250dada48416a17aee89fa1df3592f4fb223.jpg) + +![](images/b900a28943f0da369d53a2fc3df3542851c6efdf5e05adf7d133deeadb4481c2.jpg) +(13) DeepWSD / LIVE +(14) DeepWSD / CSIQ +(19) TOPIQ-FR / LIVE +(20) TOPIQ-FR / CSIQ + +![](images/706a2e441a5bf10849bd36bbb4fd390b685aff8f21d484440ace5c74aa10c9e2.jpg) + +![](images/1c47e13b27de9dd281d0e1528df25961554ac75b039be765957d9322e5b1a030.jpg) +(15) DeepWSD/TID2008 +(21) TOPIQ-FR/TID2008 + +![](images/303d77f2e2897a3b1b2ee9d137819a61ff28ad68cd8840c2a02a1c9f06d8a10f.jpg) +(16) DeepWSD / TID2013 +(17) DeepWSD / KADID +(22) TOPIQ-FR / TID2013 +(23) TOPIQ-FR / KADID + +![](images/b228b9b374d8fbb1727eb9da1f440c9f5745c7b5da85eb32d68e555dcd846393.jpg) +(18) DeepWSD / PIPAL +(24) TOPIQ-FR / PIPAL + +![](images/0577efd4b72d56d47be0cc0fd87c04c3062e970468112163f68d78e68a64b002.jpg) + +![](images/1a6c09933cdf42aaf0e35ac47beee01c6b8cfdb1341d60dcb9bb8741bc48ca3e.jpg) +(25) Ours / LIVE +Figure 2. Scatterplot of prediction results for different methods across databases. We performed normalization on both labels and predictions and 4-parameter logistic function is utilized before comparison, as suggested in [24], and the values of the lower-better methods are reversed before regression. + +![](images/9ddd7987f48ee6dcb5dc254dcd964fcac42e811ef6348967e350e3e024022048.jpg) +(26) Ours / CSIQ + +![](images/95faa752d5380ef5f74664b1073e6cf952c847ca9cc066dc3fce7906ebb54a68.jpg) +(27) Ours / TID2008 + +![](images/e7d6d28b46cdb7486d390798fbd6997c0cc0ea1de326e11e9415bb13742135db.jpg) +(28) Ours / TID2013 + +![](images/8ed3e67a7cdb4f223f3afda41fcd30c4200e05cd3d66b3b809003d5bbae307c8.jpg) +(29) Ours / KADID + +![](images/29fde11d053dc9d29212ff6e0bb2cb826ec244a4c54c58a8baad9f2c2daaa352.jpg) +(30) Ours / PIPAL + +ure 2, our method demonstrates superior overall performance and consistently aligns closely with actual scores across datasets. Traditional techniques, which are typically based on handcrafted features, often struggle to capture the complex, nonlinear relationships between image distortions and perceptual quality, resulting in relatively lower accuracy across diverse datasets. End-to-end deep learning methods, while showing a marked improvement over traditional methods, exhibit notable limitations in terms of generalizability. These methods are typically trained on specific datasets, and as a result, their performance tends to degrade when evaluated on datasets that differ from the training set, struggling to maintain consistency and accuracy when applied to unseen data. Furthermore, the deep feature-based similarity methods that do not rely on training also perform well, although they tend to be less precise + +than our model is. These methods capture image distortions via pretrained features, but their inability to explicitly model causal relationships results in slightly less accurate predictions, particularly when the perceptual impact of distortions varies significantly across datasets. In contrast, our approach, which leverages causal reasoning through an SCM and does not require training, excels in maintaining consistent performance across various datasets. The results demonstrate that, by focusing on causally relevant features, our model is less sensitive to the idiosyncrasies of individual datasets and is capable of producing accurate predictions for unseen images. This robustness in generalization suggests that the causal framework used in our model provides a more reliable representation of perceptual quality, independent of dataset-specific biases. + +Table 2. Performance comparison results of the ablation experiments. + +
ModelLIVE [27]CSIQ [16]TID2013 [22]
ParameterPLCCSRCCPLCCSRCCPLCCSRCC
φ = φθ0.9010.9150.9130.9160.8840.867
φ = φγ0.9290.9320.9490.9520.9090.884
φ = φη0.8430.8660.8030.8310.7860.789
+ +Performance comparison results of the ablation experiments conducted on three benchmark datasets, where we use OurVGG as the default architecture. The models evaluated in this study correspond to different configurations of weight assignments for $\phi$ , where its values $\phi_{\theta}$ , $\phi_{\gamma}$ , and $\phi_{\eta}$ represent the original pretrained weights, the causal effect weights, and the residual component excluding the causal effect, respectively. + +# 5.3. Ablation Analysis + +In this section, we present the abductive counterfactual experiments to evaluate the impact of different weight configurations in our model. Specifically, we analyse the results when varying the weights of $\theta$ , $\gamma$ , and $\eta$ . Here, $\theta$ represents the original pretrained weights, $\gamma$ corresponds to the weight associated with the causal effect, and $\eta$ denotes the remaining part of the model, which is the portion of $\theta$ excluding $\gamma$ . The ablation analysis is conducted by systematically adjusting each of these weights and observing their influence on the model's performance. The results are summarized in Table 2, which indicate that although noncausal representations also exhibit some perceptual relevance, their lack of causal relationships limits their effectiveness. When these noncausal representations are removed from the original feature set, the model's predictive performance improves. This suggests that incorporating those features that capture causal effects can enhance the model's ability to make more accurate and reliable quality predictions. + +# 5.4. Qualitative Evaluation + +In this section, we visually demonstrate the effect of causal intervention on different channels. As shown in Figure 3, we demonstrate the visualization effects of causal intervention between the same reference image and different distorted images (blurred and additive Gaussian noise) via the Our-VGG architecture. These visualizations reveal the varying levels of sensitivity each perceptual channel exhibits to different types of distortions in the pretrained model, thus providing a clear indication of how such distortions alter the perceptual characteristics of the image. The horizontal axis in Figure 3 represents the channels, and the color intensity corresponds to the strength of the causal relationship, with brighter colors indicating a stronger causal relationship. Causal features reflect attributes that are crucial for human perception of image quality, as they directly + +![](images/99d0e3ba122d38ba067257c08ae2064ded16a448839924f0781da2a8bf3f37a1.jpg) +Blur + +![](images/209574be8c01c266f39f3bf848e90c30162269ad8877df937bdd5037de674695.jpg) + +![](images/604315eedc6e3a5437f5e12e003c149b714e9ac3a6a7555b0f8efb36ffc5806c.jpg) +Gaussian Noise + +![](images/fabbb6897c6c1d84f8cbc3c14b41be19899015eec8290a7faa8f40c55852ba0a.jpg) + +![](images/6518ea22d993259ab7cc85da313a2258635a6defa5d1b953dd7b477d39dc8ce3.jpg) +Figure 3. Effects of causal interventions on the perceptual representations of images under various distortions. The displayed channels, from top to bottom, correspond to the second, fourth, and fifth stages of the VGG-16 architecture. + +influence perceptual judgments. In contrast, noncausal features typically correspond to higher-level semantic aspects of the image, such as object recognition, which are less relevant for perceptual quality but play a significant role in high level vision tasks. This distinction is supported by the observation that deeper layers of the network are more attuned to classification tasks, while quality-related features are predominantly represented in the shallower layers, underscoring the channel-specific nature of perceptual changes. By isolating the causal effects, we can identify specific features or regions of the image that are most sensitive to certain types of distortions, thus further elucidating how different distortions affect overall image perception. + +# 6. Conclusion + +In conclusion, we proposed an FR-IQA method grounded in abductive counterfactual inference to analyse the causal relationships between deep network features and perceptual distortions. We investigated the causal effects of deep features on human perception and incorporated causal reasoning with feature comparisons, constructing a model capable of addressing various complex distortions in different IQA scenarios. The proposed approach demonstrated strong generalizability, as the analysis of perceptual causal correlations was shown to be independent of the backbone architecture, making it adaptable across diverse deep networks. Additionally, abductive counterfactual experiments validated the effectiveness of the proposed causal framework, confirming the model's superior perceptual relevance and interpretability of quality scores. The experimental results reinforced the robustness and competitiveness of the method, which achieved consistent performance across multiple IQA benchmarks. + +# References + +[1] Hengrui Cai, Yixin Wang, Michael Jordan, and Rui Song. On learning necessary and sufficient causal graphs. In Advances in Neural Information Processing Systems, pages 42148-42160. Curran Associates, Inc., 2023. 3 +[2] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 33:2404-2418, 2024. 6 +[3] Li-Heng Chen, Christos G. Bampis, Zhi Li, Andrey Norkin, and Alan C. Bovik. Proxiqa: A proxy approach to perceptual optimization of learned image compression. IEEE Transactions on Image Processing, 30:360-373, 2021. 1 +[4] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 1, 2, 5, 6 +[5] Keyan Ding, Yi Liu, Xueyi Zou, Shiqi Wang, and Kede Ma. Locally adaptive structure and texture similarity for image quality assessment. In Proceedings of the 29th ACM International Conference on Multimedia, page 2483-2491, New York, NY, USA, 2021. Association for Computing Machinery. 2 +[6] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Comparison of full-reference image quality models for optimization of image processing systems. International Journal of Computer Vision, 129(4):1258-1281, 2021. 1 +[7] Fei Gao, Yi Wang, Panpeng Li, Min Tan, Jun Yu, and Yani Zhu. Deepsim: Deep similarity for image quality assessment. Neurocomputing, 257:104-114, 2017. Machine Learning and Signal Processing for Big Multimedia Analysis. 1, 2 +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6 +[9] Linmei Hu, Ziwei Chen, Ziwang Zhao, Jianhua Yin, and Liqiang Nie. Causal inference for leveraging image-text matching bias in multi-modal fake news detection. IEEE Transactions on Knowledge and Data Engineering, 35(11): 11141-11152, 2023. 3 +[10] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S. Ren, and Dong Chao. Pipal: A large-scale image quality assessment dataset for perceptual image restoration. In Computer Vision - ECCV 2020, pages 633-651, Cham, 2020. Springer International Publishing. 5, 6 +[11] Jongyoo Kim and Sanghoon Lee. Deep learning of human visual sensitivity in image quality assessment framework. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1676-1684, 2017. 1, 2 +[12] Junho Kim, Byung-Kwan Lee, and Yong Man Ro. Demystifying causal features on adversarial examples and causal inoculation for robust network by adversarial instrumental variable regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12302-12312, 2023. 3 + +[13] Xuting Lan, Mingliang Zhou, Xueyong Xu, Xuekai Wei, Xingran Liao, Huayan Pu, Jun Luo, Tao Xiang, Bin Fang, and Zhaowei Shang. Multilevel feature fusion for end-to-end blind image quality assessment. IEEE Transactions on Broadcasting, 69(3):801-811, 2023. 1 +[14] Shanshan Lao, Yuan Gong, Shuwei Shi, Sidi Yang, Tianhe Wu, Jiahao Wang, Weihao Xia, and Yujiu Yang. Attentions help cnns see better: Attention-based hybrid image quality assessment network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1140-1149, 2022. 1 +[15] Valero Laparra, Johannes Balle, Alexander Berardino, and Eero P. Simoncelii. Perceptual image quality assessment using a normalized laplacian pyramid. In Human Vision and Electronic Imaging, pages 43-48, United States, 2016. Society for Imaging Science and Technology. 6 +[16] Eric Cooper Larson and Damon Michael Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of Electronic Imaging, 19 (1):011006, 2010. 5, 6, 8 +[17] Xingran Liao, Baoliang Chen, Hanwei Zhu, Shiqi Wang, Mingliang Zhou, and Sam Kwong. Deepwsd: Projecting degradations in perceptual space to wasserstein distance in deep feature space. In Proceedings of the 30th ACM International Conference on Multimedia, page 970-978, New York, NY, USA, 2022. Association for Computing Machinery. 2, 5, 6 +[18] Xingran Liao, Xuekai Wei, Mingliang Zhou, Zhengguo Li, and Sam Kwong. Image quality assessment: Measuring perceptual degradation via distribution measures in deep feature spaces. IEEE Transactions on Image Processing, 33:4044-4059, 2024. 1 +[19] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience, pages 1–3, 2019. 5, 6 +[20] Bing Liu, Dong Wang, Xu Yang, Yong Zhou, Rui Yao, Zhiwen Shao, and Jiaqi Zhao. Show, deconfound and tell: Image captioning with causal inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18041-18050, 2022. 3 +[21] Nikolay Ponomarenko, Vladimir Lukin, Alexander Zelensky, Karen Egiazarian, Marco Carli, and Federica Battisti. Tid2008-a database for evaluation of full-reference visual quality assessment metrics. Advances of Modern Radioelectronics, 10(4):30-45, 2009. 5, 6 +[22] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, and C.-C. Jay Kuo. Image database tid2013: Peculiarities, results and perspectives. Signal Processing: Image Communication, 30: 57-77, 2015. 5, 6, 8 +[23] Ekta Prashnani, Hong Cai, Yasamin Mostofi, and Pradeep Sen. Pieapp: Perceptual image-error assessment through pairwise preference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1808-1817, 2018. 6 + +[24] Ann Marie Rohaly, Philip J. Corriveau, John M. Libert, Arthur A. Webster, Vittorio Baroncini, John Beerends, Jean-Louis Blin, Laura Contin, Takahiro Hamada, David Harrison, Andries P. Hekstra, Jeffrey Lubin, Yukihiro Nishida, Ricardo Nishihara, John C. Pearson, Antonio Franca Pessoa, Neil Pickford, Alexander Schertz, Massimo Visca, Andrew B. Watson, and Stefan Winkler. Video Quality Experts Group: current results and future directions. In Visual Communications and Image Processing 2000, pages 742 - 753. International Society for Optics and Photonics, SPIE, 2000. +[25] Ladan Shams and Ulrik R. Beierholm. Causal inference in perception. Trends in Cognitive Sciences, 14(9):425-432, 2010. 2 +[26] H.R. Sheikh and A.C. Bovik. Image information and visual quality. IEEE Transactions on Image Processing, 15(2):430-444, 2006. 6 +[27] H.R. Sheikh, M.F. Sabir, and A.C. Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on Image Processing, 15(11): 3440-3451, 2006. 5, 6, 8 +[28] Wenhao Shen, Mingliang Zhou, Jun Luo, Zhengguo Li, and Sam Kwong. Graph-represented distribution similarity index for full-reference image quality assessment. IEEE Transactions on Image Processing, 33:3075-3089, 2024. 1 +[29] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, page 1-14, 2015. 6 +[30] Xue Song, Jiequan Cui, Hanwang Zhang, Jingjing Chen, Richang Hong, and Yu-Gang Jiang. Doubly abductive counterfactual inference for text-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9162-9171, 2024. 3 +[31] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 6 +[32] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems and Computers, 2003, pages 1398-1402. IEEE, 2003. 6 +[33] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6 +[34] Wufeng Xue, Lei Zhang, Xuanqin Mou, and Alan C. Bovik. Gradient magnitude similarity deviation: A highly efficient perceptual image quality index. IEEE Transactions on Image Processing, 23(2):684-695, 2014. 6 +[35] Mengyue Yang, Xinyu Cai, Furui Liu, Weinan Zhang, and Jun Wang. Specify robust causal representation from mixed observations. New York, NY, USA, 2023. Association for Computing Machinery. 3 +[36] Zhongqi Yue, Hanwang Zhang, Qianru Sun, and Xian-Sheng Hua. Interventional few-shot learning. In Advances in Neural Information Processing Systems, pages 2734-2746. Curran Associates, Inc., 2020. 3 + +[37] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. Fsim: A feature similarity index for image quality assessment. IEEE Transactions on Image Processing, 20(8):2378-2386, 2011. 6 +[38] Lin Zhang, Ying Shen, and Hongyu Li. Vsi: A visual saliency-induced index for perceptual image quality assessment. IEEE Transactions on Image Processing, 23(10): 4270-4281, 2014. 6 +[39] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 1, 2, 5, 6 +[40] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10242-10251, 2021. 1 +[41] Mingliang Zhou, Shujun Lang, Taiping Zhang, Xingran Liao, Zhaowei Shang, Tao Xiang, and Bin Fang. Attentional feature fusion for end-to-end blind image quality assessment. IEEE Transactions on Broadcasting, 69(1):144-152, 2022. 1 \ No newline at end of file diff --git a/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/images.zip b/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..10258e9d603610770520f5706863149002c1e115 --- /dev/null +++ b/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:602ac800cc3e61697c60afcd4943f074c4acc69b0c679e7cd72b5acd4a95b9ad +size 725778 diff --git a/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/layout.json b/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..37e77745de823b9bcabf0d5229844abe8ff950ea --- /dev/null +++ b/2025/Image Quality Assessment_ Investigating Causal Perceptual Effects with Abductive Counterfactual Inference/layout.json @@ -0,0 +1,9524 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 102, + 527, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 102, + 527, + 138 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 527, + 138 + ], + "type": "text", + "content": "Image Quality Assessment: Investigating Causal Perceptual Effects with Abductive Counterfactual Inference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "text", + "content": "Wenhao Shen" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "text", + "content": " Mingliang Zhou" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "text", + "content": " Yu Chen" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "text", + "content": " Xuekai Wei" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "text", + "content": " Yong Feng" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "text", + "content": " Huayan Pu" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "text", + "content": " Weijia Jia" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "text", + "content": "\\\n" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "text", + "content": "College of Computer Science, Chongqing University\\\n" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "text", + "content": "State Key Laboratory of Mechanical Transmissions, Chongqing University\\\n" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 160, + 555, + 219 + ], + "type": "text", + "content": "BNU-UIC Institute of Artificial Intelligence and Future Networks, Beijing Normal University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 220, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 555, + 240 + ], + "type": "text", + "content": "shenwenhao163@163.com mingliangzhou@cqu.edu.cn yuchen@stu.cqu.edu.cn xuekaiwei2-c@my.cityu.edu.hk fengyong@cqu.edu.cn phygood_2001@shu.edu.cn jiawj@bnu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 267, + 200, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 267, + 200, + 280 + ], + "spans": [ + { + "bbox": [ + 151, + 267, + 200, + 280 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 297, + 297, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 297, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 297, + 583 + ], + "type": "text", + "content": "Existing full-reference image quality assessment (FR-IQA) methods often fail to capture the complex causal mechanisms that underlie human perceptual responses to image distortions, limiting their ability to generalize across diverse scenarios. In this paper, we propose an FR-IQA method based on abductive counterfactual inference to investigate the causal relationships between deep network features and perceptual distortions. First, we explore the causal effects of deep features on perception and integrate causal reasoning with feature comparison, constructing a model that effectively handles complex distortion types across different IQA scenarios. Second, the analysis of the perceptual causal correlations of our proposed method is independent of the backbone architecture and thus can be applied to a variety of deep networks. Through abductive counterfactual experiments, we validate the proposed causal relationships, confirming the model's superior perceptual relevance and interpretability of quality scores. The experimental results demonstrate the robustness and effectiveness of the method, providing competitive quality predictions across multiple benchmarks. The source code is available at https://anonymous.4open.science/r/DeepCausalQuality-25BC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 267, + 394, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 267, + 394, + 280 + ], + "spans": [ + { + "bbox": [ + 315, + 267, + 394, + 280 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 291, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 447 + ], + "type": "text", + "content": "Image quality assessment (IQA) is a fundamental area within image processing that is dedicated to evaluating the perceptual quality of images. The primary objective of IQA is to quantify how closely an image aligns with the human visual system (HVS), particularly when subjected to various distortions such as compression, noise, or blurring. Given its importance, IQA is integral to numerous applications, including image compression, transmission, and enhancement [3, 6, 40], where preserving visual fidelity is critical. IQA methods are generally classified into three main categories: full-reference, reduced-reference, and no-reference approaches, all of which can be considered typical regression problems [11, 13, 41]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": "In full-reference image quality assessment (FR-IQA), the evaluation process is typically divided into three key stages: feature decomposition, feature comparison, and perceptual score mapping. Recent advancements have highlighted the pivotal role that deep learning methods play in aligning IQA predictions with human subjective scores [7, 14]. Notably, deep representations of images have been shown to strongly correlate with perceptual relevance, emphasizing their significance in quality assessment [4, 28, 39]. In deep networks, different channels capture distinct features of original and distorted images, collectively encoding perceptual information. When quality degradation occurs, these features in the distorted image become contaminated to varying degrees, resulting in altered distinctions from those of the reference image [18]. However, the mere alteration of features is insufficient to fully capture the perceptual quality differences between the original and distorted images, as changes in correlation alone do not adequately explain the complex, nonlinear effects of distortions on human visual perception. Furthermore, while existing methods are effective in capturing statistical dependencies, they fail to reflect the intrinsic mecha" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 617, + 203, + 627 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 617, + 203, + 627 + ], + "spans": [ + { + "bbox": [ + 66, + 617, + 203, + 627 + ], + "type": "text", + "content": "*Corresponding author: Mingliang Zhou." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 627, + 294, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 294, + 712 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62176027, in part by Chongqing Talent under Grant cstc2024ycjh-bgzxm0082, in part by Central University Operating Expenses under Grant 2024CDJGF-044, in part by Chongqing new YC project under Grant CSTB2024YCJH-KYXM0126, in part by the Key Project of Science and Technology Research Program of Chongqing Education Commission of China under Grant No. KJZD-K202402501, and in part by the National Nature Science Foundation of China under Grant No. 62262006." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17990" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "nisms that drive changes in perceived quality. Specifically, these methods only quantify how distortions affect the similarity between feature representations and cannot elucidate how these distortions influence human perception. Consequently, they are unable to differentiate between distortions that may exhibit similar feature distances yet have significantly different perceptual impacts. This limitation restricts their capacity to provide accurate quality assessments. Additionally, no existing theoretical framework fully explains why IQA remains such a challenging task or why certain methods succeed in some contexts and fail in others. This lack of clarity impedes advancements within the field." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": "To address these challenges, we define IQA as a counterfactual inference problem, offering a perspective that bridges existing theoretical gaps [25]. Counterfactual reasoning provides a formalized structure for defining the minimal perceptual cost, enabling a more precise understanding of how distortions affect perceived quality. Given that the reference image serves as the baseline for the original visual content, an effective IQA method should be able to predict scores on the basis of the perceptual distortions present in degraded images. However, the manner in which distortions are perceived is essentially probabilistic and is influenced by a variety of contextual factors and prior experience. While deep learning-based methods are adept at capturing complex feature interactions, they often fail to incorporate the probabilistic relationships that underlie perceptual judgments. By constructing IQA from a causal reasoning perspective, our approach aims to quantify perceptual differences in a way that reflects these inherent uncertainties and complexities. This causal perspective provides us with not only a measure of feature similarity but also a more nuanced understanding of how different distortions affect human perception. The main contributions of our study are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "text", + "content": "- We investigate the causal effects of deep network features on perception, effectively integrating causal relationships with feature comparisons to construct a deep network-based FR-IQA model. By employing distance comparisons among the causal relationships of deep features, the model can accurately address complex distortion types encountered across various image quality assessment scenarios, resulting in competitive quality prediction outcomes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "- The proposed method exhibits significant generality, as the analysis of perceptual causal correlations within deep features is independent of the backbone architecture. This characteristic enables the implementation of the approach across different deep backbone networks, thereby enhancing its versatility and applicability in diverse contexts of image quality assessment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "- The validity of the proposed causal relationships is substantiated through experimental evidence derived from" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 555, + 109 + ], + "type": "text", + "content": "abductive counterfactual inference, which demonstrates superior perceptual relevance and score interpretability and establishes its robustness and practicality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 121, + 402, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 402, + 133 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 402, + 133 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 140, + 476, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 476, + 153 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 476, + 153 + ], + "type": "text", + "content": "2.1. Deep Network-based FR-IQA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 159, + 555, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 555, + 493 + ], + "type": "text", + "content": "Deep networks have been employed to model the complex relationships between reference and distorted images in FR-IQA. In [11], convolutional operations on both reference and distorted images are utilized to obtain sensitivity maps, which are then applied to the difference map for quality estimation. Gao et al. [7] measured the local similarity between features extracted from deep neural networks (DNNs) for both reference and distorted images. The local quality indices are aggregated to produce the overall quality score. Subsequent studies have shown that VGG network features serve as effective training losses for tasks such as image synthesis. Notably, the study of [39] compares deep features across various architectures and tasks, showing that perceptual similarity is an emerging property shared by deep visual representations. Hence, some advancements focus on perceptual similarity or differences in deep feature space. For example, Ding et al. [4] employed the structural similarity index (SSIM) between deep features to measure perceptual texture similarity, offering a unified deep image structure and texture similarity (DISTS) framework, and the adaptive DISTS (ADISTS) [5] improved it by using a dispersion index to identify texture regions and adaptively aggregate structure and texture measurements. In contrast, Liao et al. [17] projected image degradations into perceptual space and computed the Wasserstein distance in deep feature space, highlighting the importance of distributional differences in deep feature representations for quality assessment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 495, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 556, + 615 + ], + "type": "text", + "content": "Existing methods are limited in capturing perceptual quality differences because they rely on feature alterations and statistical dependencies and fail to model the nonlinear effects of distortions on human perception. They quantify feature similarity without revealing how distortions impact perception, leading to difficulties in distinguishing distortions with similar feature distances but different perceptual effects. Additionally, the absence of a robust theoretical framework to explain why some methods succeed or fail in different scenarios limits progress in IQA." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 624, + 501, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 501, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 501, + 636 + ], + "type": "text", + "content": "2.2. Causal Inference in Deep Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": "In deep learning, with the presence of confounding factors, the network can easily capture spurious correlations between inputs and their outputs. Therefore, the application of causal inference integration in deep learning has received increasing attention. Causal inference provides a theoretical foundation for disentangling correlation from" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17991" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "causation, which is particularly useful for improving model generalizability, robustness, and interpretability [12] [35]. In deep learning, causal approaches have been explored to mitigate issues such as bias, confounding, and spurious correlations that often arise from data-driven methods [1, 30]. Yue et al. [36] proposed an interventional few-shot learning method that applies backdoor adjustment to mitigate the confounding effect of pretrained knowledge. Liu et al. [20] addressed both visual confounders in the object detection stage and linguistic confounders in the decoding process for image captioning tasks via causal intervention, leading to substantial improvements in generalizability across different datasets. Hu et al. [9] modelled the causal relationship between image-text matching and fake news detection in multimodal settings and used causal inference to eliminate confounding effects during training and leverage matching bias to improve detection accuracy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 276, + 294, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 294, + 382 + ], + "type": "text", + "content": "In the context of image quality assessment, causal reasoning offers a promising avenue for improving the accuracy and consistency of quality prediction. By focusing on the causal relationship between image features and perceived quality, it can explore how causal models mitigate the problems of spurious correlation and overfitting to better align with human perceptual judgments and improve the reliability of deep learning models in real-world applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 395, + 181, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 181, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 181, + 407 + ], + "type": "text", + "content": "3. Problem Formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 462 + ], + "type": "text", + "content": "The general objective in image quality assessment is to minimize the discrepancy between the predicted quality score and the subjective mean opinion score (MOS), which can be formalized as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 472, + 295, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 472, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 87, + 472, + 295, + 491 + ], + "type": "interline_equation", + "content": "\\min _ {\\zeta , \\omega , \\theta} \\mathbb {E} [ l (f _ {\\zeta} (\\operatorname {D i s} _ {\\omega} (\\phi_ {\\theta} (I), \\phi_ {\\theta} (D))), M O S) ] \\tag {1}", + "image_path": "a229ef9656a9156c1379c7f9cef3fa3c5979fe8b1b4471a4460b39f70eb8e414.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "text", + "content": " represents the parameters of model " + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "inline_equation", + "content": "\\mathrm{Dis}(\\phi_{\\theta}(I),\\phi_{\\theta}(D))" + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "text", + "content": " denotes the distance between the feature representations of the original image " + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "text", + "content": " and the distorted image " + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "text", + "content": " in the deep feature space " + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "inline_equation", + "content": "l(\\cdot)" + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "text", + "content": " is the loss function that quantifies the difference between the predicted score " + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "inline_equation", + "content": "f_{\\zeta}(\\mathrm{Dis}_{\\omega}(\\phi (I),\\phi (D)))" + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "text", + "content": " and the subjective quality. The expectation " + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "inline_equation", + "content": "\\mathbb{E}" + }, + { + "bbox": [ + 55, + 501, + 295, + 595 + ], + "type": "text", + "content": " is taken over the data distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 597, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 643 + ], + "type": "text", + "content": "For full-reference image quality assessment, we consider the causal relationship between the original image " + }, + { + "bbox": [ + 55, + 597, + 295, + 643 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 597, + 295, + 643 + ], + "type": "text", + "content": " and the distorted image " + }, + { + "bbox": [ + 55, + 597, + 295, + 643 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 597, + 295, + 643 + ], + "type": "text", + "content": ". The relationship can be expressed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 654, + 294, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 654, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 85, + 654, + 294, + 668 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} _ {S} = m \\left(\\phi_ {\\gamma} (I, D)\\right), \\quad \\eta \\perp \\gamma \\text {a n d} \\gamma , \\eta \\sim F _ {\\theta} \\tag {2}", + "image_path": "be52a8e1eeb1ccedba715087ea98824455e199ff393804db2b3570f60ac66828.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "inline_equation", + "content": "m(\\cdot)" + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": " represents a function that maps the joint feature representation " + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{\\gamma}(I,D)" + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": " of the original and distorted images to the subjective quality " + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_S" + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": ". The parameter " + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": "pivotal in this formulation, as it represents the features with causal implications for image quality, thereby influencing the perceptual evaluation. In contrast, " + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": " signifies the noise components that are independent of the deep representation parameter " + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": ", suggesting that these components do not contribute to the quality assessment. Both " + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": " are derived from the distribution " + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": ", indicating their dependence on the model's parameters, " + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": " represents the foundational parameters of the deep feature representation, which gives rise to two distinct components: " + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": ", the relevant features that causally affect image quality, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": ", the extraneous noise that does not contribute to the quality assessment. This causal relationship isolates the effect of distortions from external factors, ensuring that the predicted quality score is influenced primarily by distortions between the reference and distorted images, whereas other factors are captured by " + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 276, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 396 + ], + "type": "text", + "content": "In light of the established causal relationship, the subsequent phase focuses on prioritizing causal features " + }, + { + "bbox": [ + 313, + 276, + 553, + 396 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 276, + 553, + 396 + ], + "type": "text", + "content": " while effectively disregarding the noise components " + }, + { + "bbox": [ + 313, + 276, + 553, + 396 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 276, + 553, + 396 + ], + "type": "text", + "content": ". The goal is to adjust the model parameters " + }, + { + "bbox": [ + 313, + 276, + 553, + 396 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 276, + 553, + 396 + ], + "type": "text", + "content": " to better capture the relevant causal features " + }, + { + "bbox": [ + 313, + 276, + 553, + 396 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 276, + 553, + 396 + ], + "type": "text", + "content": " that significantly influence the subjective quality score " + }, + { + "bbox": [ + 313, + 276, + 553, + 396 + ], + "type": "inline_equation", + "content": "Q_{S}" + }, + { + "bbox": [ + 313, + 276, + 553, + 396 + ], + "type": "text", + "content": ". The optimization process is designed to minimize the prediction error across all conceivable data distributions, thereby ensuring that the model remains robust against varying inputs and distortions. This objective can be formally articulated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 406, + 553, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 406, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 321, + 406, + 553, + 426 + ], + "type": "interline_equation", + "content": "\\min _ {\\zeta , \\omega , \\theta \\rightarrow \\gamma} \\sup _ {P \\in \\mathcal {P}} \\mathbb {E} _ {P} \\left[ l \\left(f _ {\\zeta} \\left(\\operatorname {D i s} _ {\\omega} \\left(\\phi_ {\\theta} (I), \\phi_ {\\theta} (D)\\right)\\right), M O S\\right)\\right] \\tag {3}", + "image_path": "ad4f117da3f12a6bf802f7203a681502f30db70d76d87dc66c4f62b698c25f8a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 435, + 553, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 553, + 506 + ], + "type": "text", + "content": "where the term " + }, + { + "bbox": [ + 313, + 435, + 553, + 506 + ], + "type": "inline_equation", + "content": "\\min_{\\theta \\to \\gamma}" + }, + { + "bbox": [ + 313, + 435, + 553, + 506 + ], + "type": "text", + "content": " signifies the optimization of model parameters " + }, + { + "bbox": [ + 313, + 435, + 553, + 506 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 435, + 553, + 506 + ], + "type": "text", + "content": " to enhance the representation of the causal parameters " + }, + { + "bbox": [ + 313, + 435, + 553, + 506 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 435, + 553, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 435, + 553, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 435, + 553, + 506 + ], + "type": "text", + "content": " represents the set of all possible distributions that are consistent with the structural causal model, and the supremum operator " + }, + { + "bbox": [ + 313, + 435, + 553, + 506 + ], + "type": "inline_equation", + "content": "\\sup_{P\\in \\mathcal{P}}" + }, + { + "bbox": [ + 313, + 435, + 553, + 506 + ], + "type": "text", + "content": " accounts for the worst-case scenario over these distributions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": "In our study, instead of training the model from scratch, we define " + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": " as the pretrained model weights. We subsequently employ variable interventions to validate the autoregressive properties of the model, facilitating the filtration of extraneous information. This strategy aims to consider only causal representation parameters " + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": " in the quality prediction process while effectively ignoring the noise component " + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 603, + 397, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 397, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 397, + 616 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 624, + 392, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 392, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 392, + 635 + ], + "type": "text", + "content": "4.1. Framework" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 712 + ], + "type": "text", + "content": "As illustrated in Figure 1 (a), the proposed framework integrates multiple stages, starting with the input of reference and distorted images. These images are first processed in the deep domain, where deep learning models extract high-level feature representations. These feature representations serve as the foundation for the subsequent causal inference" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17992" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 88, + 550, + 242 + ], + "blocks": [ + { + "bbox": [ + 265, + 75, + 342, + 88 + ], + "lines": [ + { + "bbox": [ + 265, + 75, + 342, + 88 + ], + "spans": [ + { + "bbox": [ + 265, + 75, + 342, + 88 + ], + "type": "text", + "content": "(a) Framework" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 88, + 550, + 242 + ], + "lines": [ + { + "bbox": [ + 58, + 88, + 550, + 242 + ], + "spans": [ + { + "bbox": [ + 58, + 88, + 550, + 242 + ], + "type": "image", + "image_path": "f9874e4eccd13350e1503e0f940418f97216307d9f0b82dd670817893fb4e3ce.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 245, + 548, + 258 + ], + "lines": [ + { + "bbox": [ + 405, + 245, + 548, + 258 + ], + "spans": [ + { + "bbox": [ + 405, + 245, + 548, + 258 + ], + "type": "text", + "content": "(a(c) Structural Causal Model)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 264, + 399, + 396 + ], + "blocks": [ + { + "bbox": [ + 132, + 246, + 328, + 259 + ], + "lines": [ + { + "bbox": [ + 132, + 246, + 328, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 328, + 259 + ], + "type": "text", + "content": "(b) Confounder Dictionary Embedding" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 264, + 399, + 396 + ], + "lines": [ + { + "bbox": [ + 61, + 264, + 399, + 396 + ], + "spans": [ + { + "bbox": [ + 61, + 264, + 399, + 396 + ], + "type": "image", + "image_path": "1b1accfd3c380f58e206fafb862b7deb58fad0c9823136471836aaaaea9b7581.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 407, + 535, + 418 + ], + "lines": [ + { + "bbox": [ + 73, + 407, + 535, + 418 + ], + "spans": [ + { + "bbox": [ + 73, + 407, + 535, + 418 + ], + "type": "text", + "content": "Figure 1. The framework of our proposed method that separates and leverages causal features for perceptual quality assessment." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 427, + 261, + 530, + 392 + ], + "blocks": [ + { + "bbox": [ + 427, + 261, + 530, + 392 + ], + "lines": [ + { + "bbox": [ + 427, + 261, + 530, + 392 + ], + "spans": [ + { + "bbox": [ + 427, + 261, + 530, + 392 + ], + "type": "image", + "image_path": "471c9eff8eb9f353447c3fc51cb950a20bfc64d0cf5e1aa3289cb5085d11abbf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 440, + 295, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 295, + 524 + ], + "type": "text", + "content": "stage. In this stage, the relationships between the extracted features are analysed to identify causal variables that are robust to distortions and potential confounders. Finally, the quality assessment step uses these identified causal variables to evaluate the perceptual quality of the distorted image, producing a quality score that reflects the minimal perceptual cost on the basis of the underlying causal structure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "To explain the causal relationships considered in our approach, we employ a structural causal model (SCM) for the full-reference image quality assessment. The SCM represents variables as nodes and relationships as edges in a graphical structure, using directed arrows to indicate causal relationships between two variables. In our model, we incorporate the reference image " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": ", the distorted image " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": ", the representations of both images " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\phi_{\\theta}(I)" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\phi_{\\theta}(D)" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": ", and the image quality score " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": ", as depicted in Figure 1 (c). We assume that the images are characterized by factors that have causal relationships with quality, along with elements that are unrelated to quality. In this framework, image quality is directly determined by factors that causally influence it. Simply using a training process or pretrained weights from other fields to assess image quality typically yields repre" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 440, + 555, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 555, + 499 + ], + "type": "text", + "content": "sentations that are correlated with perceptual quality but do not capture causal relationships. Therefore, we introduce an unknown exogenous variable " + }, + { + "bbox": [ + 313, + 440, + 555, + 499 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 440, + 555, + 499 + ], + "type": "text", + "content": " to account for uncertainties and identify the representations that are causally related to quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 508, + 463, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 463, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 463, + 521 + ], + "type": "text", + "content": "4.2. Deep Causal Measurement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 525, + 555, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 620 + ], + "type": "text", + "content": "In our approach, to verify causal relationships in perceptual quality assessment, we design a causal validation process using an SCM with an exogenous variable " + }, + { + "bbox": [ + 313, + 525, + 555, + 620 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 525, + 555, + 620 + ], + "type": "text", + "content": ". This process begins with a causal intervention to test the effect of image characteristics on perceived quality, using the operator " + }, + { + "bbox": [ + 313, + 525, + 555, + 620 + ], + "type": "inline_equation", + "content": "\\mathrm{do}(\\cdot)" + }, + { + "bbox": [ + 313, + 525, + 555, + 620 + ], + "type": "text", + "content": " to map both reference and distorted images onto a shared causal feature space. The intervention can be formulated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 631, + 555, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 631, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 341, + 631, + 555, + 643 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {I} = \\phi_ {\\theta} (\\mathrm {d o} (I = i)), \\quad \\mathbf {f} _ {D} = \\phi_ {\\theta} (\\mathrm {d o} (D = d)) \\tag {4}", + "image_path": "2b6a5d5dc68acf13f5b7c1a9122b749e599a0fafbe7923e8bc5ee123147447bf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\phi_{\\theta}" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " is the feature extraction function parameterized by " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": ", which maps the images into the deep feature space, and " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " denotes the deep features. To substantiate the causal nature of the features, we introduce the exogenous variable " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " and perform simultaneous interventions " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " on both the original" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17993" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 118 + ], + "type": "text", + "content": "reference image features " + }, + { + "bbox": [ + 55, + 72, + 294, + 118 + ], + "type": "inline_equation", + "content": "\\phi_{\\theta}(I)" + }, + { + "bbox": [ + 55, + 72, + 294, + 118 + ], + "type": "text", + "content": " and distorted image features " + }, + { + "bbox": [ + 55, + 72, + 294, + 118 + ], + "type": "inline_equation", + "content": "\\phi_{\\theta}(D)" + }, + { + "bbox": [ + 55, + 72, + 294, + 118 + ], + "type": "text", + "content": " to generate the intervened causal representations " + }, + { + "bbox": [ + 55, + 72, + 294, + 118 + ], + "type": "inline_equation", + "content": "\\mathbf{f}'" + }, + { + "bbox": [ + 55, + 72, + 294, + 118 + ], + "type": "text", + "content": ". This intervention yields representations formulated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 128, + 295, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 128, + 295, + 142 + ], + "spans": [ + { + "bbox": [ + 91, + 128, + 295, + 142 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {I} ^ {\\prime} = \\phi_ {\\theta} (\\mathrm {d o} (I, v)), \\quad \\mathbf {f} _ {D} ^ {\\prime} = \\phi_ {\\theta} (\\mathrm {d o} (D, v)) \\tag {5}", + "image_path": "dce0e53457a0f0f2b4cbd77848501aa75dff8dc8ae2828e7be3364d794a51dfe.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 148, + 296, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 296, + 232 + ], + "type": "text", + "content": "Next, we calculate the perceptual distance " + }, + { + "bbox": [ + 55, + 148, + 296, + 232 + ], + "type": "inline_equation", + "content": "\\mathrm{Dis}(\\cdot)" + }, + { + "bbox": [ + 55, + 148, + 296, + 232 + ], + "type": "text", + "content": " before and after the intervention to assess the invariance of the causal representation. If the perceptual distance differs significantly between the pre- and postintervention states, then the feature representation can be deemed causal, as it is sensitive to distortions influencing perceptual quality. This invariance can be evaluated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 239, + 295, + 254 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 239, + 295, + 254 + ], + "spans": [ + { + "bbox": [ + 111, + 239, + 295, + 254 + ], + "type": "interline_equation", + "content": "\\Delta = \\operatorname {D i s} \\left(\\mathbf {f} _ {I}, \\mathbf {f} _ {D}\\right) - \\operatorname {D i s} \\left(\\mathbf {f} _ {I} ^ {\\prime}, \\mathbf {f} _ {D} ^ {\\prime}\\right) \\tag {6}", + "image_path": "37b5d8ba86fab4b01858d997a57c6e58d162f97fb5c9eb4b38e1ef3428624614.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 260, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 296, + 378 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 55, + 260, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\Delta \\neq 0" + }, + { + "bbox": [ + 55, + 260, + 296, + 378 + ], + "type": "text", + "content": ", it confirms that the causal factors within the feature representations significantly affect perceptual quality. In practice, we vary the disturbance intensity to determine the strength of the perceptual causal relationship within the feature representations, and we use a confounder dictionary to record the representations that maintain causality at all intensities, denoted " + }, + { + "bbox": [ + 55, + 260, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\Gamma (\\mathbf{f}_I,\\mathbf{f}_D)" + }, + { + "bbox": [ + 55, + 260, + 296, + 378 + ], + "type": "text", + "content": ". The causal influence on perceptual quality is then quantified by observing how the perceptual distance changes with varying intervention intensities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "content": "The regression invariance is validated if the perceptual distance " + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\mathrm{Dis}(\\mathbf{f}_I',\\mathbf{f}_D')" + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "content": " yields a consistent quality score " + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "content": " under intervention, where " + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "content": " is parameterized by " + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "content": ", such that:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 423, + 295, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 423, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 143, + 423, + 295, + 437 + ], + "type": "interline_equation", + "content": "Q = \\mathbb {E} [ Q | \\gamma , v ] \\tag {7}", + "image_path": "58a9396dced0c9227ef1559681e1f09385d815a58de6d51d550a2d8ce559501f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 444, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 504 + ], + "type": "text", + "content": "This indicates that the quality prediction model relies solely on the causal features " + }, + { + "bbox": [ + 55, + 444, + 295, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 55, + 444, + 295, + 504 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 444, + 295, + 504 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 444, + 295, + 504 + ], + "type": "text", + "content": " representing the parameter of " + }, + { + "bbox": [ + 55, + 444, + 295, + 504 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 444, + 295, + 504 + ], + "type": "text", + "content": " in this context, excluding the influence of noise " + }, + { + "bbox": [ + 55, + 444, + 295, + 504 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 444, + 295, + 504 + ], + "type": "text", + "content": ". The SCM, therefore, ensures that the quality assessment is both robust to noise and aligned with causal perceptual features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 504, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 296, + 612 + ], + "type": "text", + "content": "Finally, on the basis of the SCM, the perceptual differences between the reference image and the distorted image are quantified via a distance metric. Our method is predicated on the concept of perceptual cost, facilitating a quantitative analysis of the differences in perceived quality between reference and distorted images. Specifically, we define the perceptual cost between two distributions, " + }, + { + "bbox": [ + 55, + 504, + 296, + 612 + ], + "type": "inline_equation", + "content": "P_{X}" + }, + { + "bbox": [ + 55, + 504, + 296, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 504, + 296, + 612 + ], + "type": "inline_equation", + "content": "P_{Y}" + }, + { + "bbox": [ + 55, + 504, + 296, + 612 + ], + "type": "text", + "content": ", utilizing the causal transport cost, which is formally expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 618, + 295, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 618, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 58, + 618, + 295, + 653 + ], + "type": "interline_equation", + "content": "C O T \\left(P _ {X}, P _ {Y}\\right) = \\inf _ {g \\in G \\left(P _ {X}, P _ {Y}\\right)} \\int \\Gamma (x, y) c (x, y) d g (x, y) \\tag {8}", + "image_path": "8d2f62340cdeefeb6db33be31069bf9bf7b0831f81028fc62ed2126b8564da44.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "G(P_X, P_Y)" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " represents the joint distribution of the reference and distorted images, " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "c(x, y)" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " is the 12 norm of the difference between the distributions, and " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\Gamma(x, y)" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " denotes the causal confounder dictionary between the two distributions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 533, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 533, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 533, + 84 + ], + "type": "text", + "content": "4.3. Connection to Existing Methods and HVS" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 89, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 293 + ], + "type": "text", + "content": "Existing deep learning-based methods [4, 17, 39] leverage high-level feature representations from deep neural networks to improve perceptual alignment by statistically correlating feature distances with quality scores. LPIPS [39] measures the distance between deep network features extracted from image patches, whereas DeepWSD [17] leverages the Wasserstein distance within the deep feature space to evaluate perceptual similarity. DISTS [4], on the other hand, assesses the similarity between image structures and textures by comparing deep feature representations. However, they still rely primarily on correlations and statistical dependencies rather than causal mechanisms, which can limit their ability to generalize effectively across diverse datasets and distortion types. In contrast, our approach introduces a causal framework that aims to reveal the underlying mechanisms driving perceptual differences, providing a more robust and theoretically grounded approach to IQA." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 293, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 472 + ], + "type": "text", + "content": "HVS is known to be sensitive to specific perceptual features (e.g., contrast, texture, and structural integrity) that are often distorted by various image processing operations. Previous studies have attempted to model HVS by designing perceptual metrics or using deep learning to automatically learn these perceptual features. However, these models usually fail to distinguish between causal features that actually affect perceptual quality and incidental noise that may be captured by deep features. By introducing a causal model that uses an intervention-based framework, we provide a clearer path for identifying causal features that determine perceptual quality, providing insight into the relationship between image content distortion and perceptual quality in a way that is more consistent with the mechanisms of HVS." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 484, + 440, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 440, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 440, + 498 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 504, + 441, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 441, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 441, + 517 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": "The study employs six widely used IQA databases to evaluate the effectiveness of the proposed method. These databases include the LIVE IQA [27] database, CSIQ [16], TID2008 [21], TID2013 [22], KADID [19], and PIPAL [10], each containing different types of image distortions and varying numbers of reference images. LIVE [27] comprises 29 reference images and 779 distorted images across five types of distortions. It provides quality scores on a scale of [0, 100]. Containing 30 reference images and 866 distorted images, CSIQ [16] covers six distortion types, with subjective quality scores normalized to the range [0, 1]. TID2008 [21] and TID2013 [22] include 25 reference images. TID2008 has 1,700 distorted images with 17 types of distortions, whereas TID2013 expands to 3,000 distorted images and 24 distortion types. Both databases use a quality score range of [0, 9]. KADID [19] provides a large-scale" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17994" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 91, + 552, + 343 + ], + "blocks": [ + { + "bbox": [ + 154, + 71, + 455, + 82 + ], + "lines": [ + { + "bbox": [ + 154, + 71, + 455, + 82 + ], + "spans": [ + { + "bbox": [ + 154, + 71, + 455, + 82 + ], + "type": "text", + "content": "Table 1. Results of the performance comparison conducted on benchmark datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 91, + 552, + 343 + ], + "lines": [ + { + "bbox": [ + 59, + 91, + 552, + 343 + ], + "spans": [ + { + "bbox": [ + 59, + 91, + 552, + 343 + ], + "type": "table", + "html": "
MethodLIVE [27]CSIQ [16]TID2008 [21]TID2013 [22]KADID [19]PIPAL [10]
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
PSNR0.7810.8010.7920.8070.5070.5250.6640.6870.6700.6760.3980.392
SSIM [33]0.8470.8510.8100.8330.6250.6240.6650.6270.6100.6190.4890.486
MS-SSIM [32]0.8860.9030.8750.8790.8420.8540.8310.7860.8240.8260.5710.545
VIF [26]0.9490.9530.8990.8990.7980.7490.7710.6770.6850.6790.5720.545
MAD [16]0.9040.9070.9220.9220.6810.7080.7370.7430.7170.7260.6140.591
FSIM [37]0.9100.9200.9020.9150.8750.8840.8760.8510.8520.8540.5970.573
VSI [38]0.8770.8990.9120.9290.8710.8950.8980.8950.8770.8780.5480.526
GMSD [34]0.9090.9100.9380.9390.8780.8910.8580.8040.8470.8470.6140.569
NLPD [15]0.8820.8890.9130.9260.8660.8770.8320.7990.8090.8120.4890.464
PieAPP [23]0.8660.8650.8640.8830.7520.7970.8090.8440.8570.8650.7020.701
LPIPS [39]0.8660.8630.8910.8950.7220.7180.7130.7130.8380.8370.5840.584
DISTS [4]0.9240.9250.9190.9200.8290.8140.8540.8300.886*0.886*0.6450.627
DeepWSD [17]0.9040.9250.9410.9500.9000.9040.8940.8740.8870.8880.5030.500
TOPIQ-FR [2]0.8820.8870.8940.8940.8730.8720.8540.8200.8960.8950.837*0.809*
Our-VGG0.9290.9320.9490.9520.9050.8970.9090.8840.8980.8990.6750.657
Our-ResNet0.9140.9210.9160.9230.8590.8680.8620.8520.8880.8900.5380.524
Our-EffNet0.9270.9320.9330.9380.9030.8990.8990.8790.9050.9070.6940.656
", + "image_path": "0956c9045f3f85dc9e2ac8a888dd80fd28bb19deb0fb37cddc9d840314ce74f3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 69, + 346, + 555, + 383 + ], + "lines": [ + { + "bbox": [ + 69, + 346, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 69, + 346, + 555, + 383 + ], + "type": "text", + "content": "Notably, DISTS and TOPIQ-FR were trained on the KADID [19] and PIPAL [10], respectively, which are marked with an asterisk (*) in the table for these methods. PieAPP and LPIPS were individually trained on their proposed datasets [23] and [39], respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 402, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 295, + 556 + ], + "type": "text", + "content": "collection of 81 reference images and 10,125 distorted images, spanning 25 different distortion types. The quality scores fall within the interval [1, 5]. As one of the most extensive datasets, PIPAL [10] includes 250 reference images and 25,850 distorted images across 40 distortion types. The subjective scores range from 868-1857. Notably, we used only 24200 images from the training and validation sets since the labels of the test set were not available. Given the diversity in score ranges across these databases, a linear normalization process is applied to standardize the MOS values, ensuring consistency in the evaluation process. This step ensures that a fair comparison of the quality of predictions can be made across different datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "In this study, we evaluate a variety of representative and state-of-the-art image quality assessment techniques, including both feature-based and deep learning approaches, including the PSNR, SSIM [33], MS-SSIM [32], VIF [26], MAD [16], FSIM [37], VSI [38], GMSD [34], NLPD [15], PieAPP [23], LPIPS [39], DISTS [4], DeepWSD [17], and TOPIQ-FR [2]. All the comparative methods were re-evaluated in our study. Specifically, for the deep learning-based approaches, we utilized the weight files provided by the authors of these methods to assess their performance through cross-dataset predictions. The training datasets used for these weights are indicated in the table. For the deep domain methods that do not require training, we fol" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 402, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 553, + 510 + ], + "type": "text", + "content": "lowed the same protocol as traditional methods and conducted tests directly on the entire dataset. To evaluate the performance of these methods, we employ the Pearson linear correlation coefficient (PLCC) and the Spearman rank correlation coefficient (SRCC) as the key metrics. The PLCC is used to assess the accuracy of the predictions, whereas the SRCC evaluates the monotonic relationship by examining the rank correlation between the predicted quality scores and the corresponding MOS values." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 510, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 594 + ], + "type": "text", + "content": "Our proposed method is training-free and can be seamlessly applied to various deep learning models for image quality assessment. Specifically, we utilize VGG [29], ResNet [8] and EfficientNet [31] as backbone networks to extract deep feature representations. These networks are pretrained on large-scale datasets; thus, the features can be utilized to assess image quality without additional training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 600, + 487, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 487, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 487, + 612 + ], + "type": "text", + "content": "5.2. Quality Prediction Performance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": "We conduct quality prediction experiments to evaluate the performance and generalizability of our model in comparison with several state-of-the-art methods. The methods we compare can be categorized into three main groups: traditional image quality assessment methods, end-to-end deep learning models, and deep feature-based similarity (or difference) metrics that do not require training. As presented in Table 1 and visualized in the scatter plot in Fig-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17995" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 225, + 137 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 225, + 137 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 225, + 137 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 225, + 137 + ], + "type": "image", + "image_path": "e591eb43172702c48b101e02e126cb0979ac1b437fbcea0281c7ed736dce9ed4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 138, + 126, + 146 + ], + "lines": [ + { + "bbox": [ + 77, + 138, + 126, + 146 + ], + "spans": [ + { + "bbox": [ + 77, + 138, + 126, + 146 + ], + "type": "text", + "content": "(1) SSIM / LIVE" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 70, + 310, + 137 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 310, + 137 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 310, + 137 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 310, + 137 + ], + "type": "image", + "image_path": "17df6d65834efb588cbfb14267323c0a6a28b8917bc9a4f981171bc0f8c157f0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 391, + 137 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 391, + 137 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 391, + 137 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 391, + 137 + ], + "type": "image", + "image_path": "3c11132faf853ec91786e90e7e767be9de3092af51dd840676ced79b4b119236.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 391, + 71, + 476, + 137 + ], + "blocks": [ + { + "bbox": [ + 391, + 71, + 476, + 137 + ], + "lines": [ + { + "bbox": [ + 391, + 71, + 476, + 137 + ], + "spans": [ + { + "bbox": [ + 391, + 71, + 476, + 137 + ], + "type": "image", + "image_path": "764be3e05b40fb3618cb5b05975f7758f67ef4b4c4381e7c66a4b4b61f53082c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 477, + 71, + 551, + 137 + ], + "blocks": [ + { + "bbox": [ + 477, + 71, + 551, + 137 + ], + "lines": [ + { + "bbox": [ + 477, + 71, + 551, + 137 + ], + "spans": [ + { + "bbox": [ + 477, + 71, + 551, + 137 + ], + "type": "image", + "image_path": "7a2d13c6f6eb0de52d0114df707227e30553092f7f0970ebe4cbee74c0ddfa53.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 58, + 147, + 149, + 214 + ], + "blocks": [ + { + "bbox": [ + 58, + 147, + 149, + 214 + ], + "lines": [ + { + "bbox": [ + 58, + 147, + 149, + 214 + ], + "spans": [ + { + "bbox": [ + 58, + 147, + 149, + 214 + ], + "type": "image", + "image_path": "1617eff18b50cafa62090a2e61215db5bfea5cb0963cfe8912e00db29bfe6613.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 149, + 147, + 222, + 214 + ], + "blocks": [ + { + "bbox": [ + 149, + 147, + 222, + 214 + ], + "lines": [ + { + "bbox": [ + 149, + 147, + 222, + 214 + ], + "spans": [ + { + "bbox": [ + 149, + 147, + 222, + 214 + ], + "type": "image", + "image_path": "f8b67a9889bc3e4f04fb5cf9f485a32c8f23d58d5d8fd60fb632dace9eb9a085.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 222, + 147, + 306, + 214 + ], + "blocks": [ + { + "bbox": [ + 235, + 138, + 294, + 147 + ], + "lines": [ + { + "bbox": [ + 235, + 138, + 294, + 147 + ], + "spans": [ + { + "bbox": [ + 235, + 138, + 294, + 147 + ], + "type": "text", + "content": "(3) SSIM/TID2008" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 222, + 147, + 306, + 214 + ], + "lines": [ + { + "bbox": [ + 222, + 147, + 306, + 214 + ], + "spans": [ + { + "bbox": [ + 222, + 147, + 306, + 214 + ], + "type": "image", + "image_path": "a2ca4f68a256d68ae790399d8d2b073f5e7024121d22a441aa0973a10549cfc0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 147, + 391, + 214 + ], + "blocks": [ + { + "bbox": [ + 317, + 138, + 375, + 147 + ], + "lines": [ + { + "bbox": [ + 317, + 138, + 375, + 147 + ], + "spans": [ + { + "bbox": [ + 317, + 138, + 375, + 147 + ], + "type": "text", + "content": "(4) SSIM / TID2013" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 147, + 391, + 214 + ], + "lines": [ + { + "bbox": [ + 306, + 147, + 391, + 214 + ], + "spans": [ + { + "bbox": [ + 306, + 147, + 391, + 214 + ], + "type": "image", + "image_path": "e6b710b2da393771d294bc5679dd563804825a31f2a6f6aca9401467a42a713f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 391, + 147, + 475, + 214 + ], + "blocks": [ + { + "bbox": [ + 400, + 138, + 455, + 147 + ], + "lines": [ + { + "bbox": [ + 400, + 138, + 455, + 147 + ], + "spans": [ + { + "bbox": [ + 400, + 138, + 455, + 147 + ], + "type": "text", + "content": "(5) SSIM / KADID" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 147, + 475, + 214 + ], + "lines": [ + { + "bbox": [ + 391, + 147, + 475, + 214 + ], + "spans": [ + { + "bbox": [ + 391, + 147, + 475, + 214 + ], + "type": "image", + "image_path": "6a0fb8bfd8f56f13a3916b5f303ba03c7e6586e6a54dad0dccc2ad0396d31a5a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 475, + 147, + 551, + 214 + ], + "blocks": [ + { + "bbox": [ + 484, + 138, + 535, + 147 + ], + "lines": [ + { + "bbox": [ + 484, + 138, + 535, + 147 + ], + "spans": [ + { + "bbox": [ + 484, + 138, + 535, + 147 + ], + "type": "text", + "content": "(6) SSIM / PIPAL" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 475, + 147, + 551, + 214 + ], + "lines": [ + { + "bbox": [ + 475, + 147, + 551, + 214 + ], + "spans": [ + { + "bbox": [ + 475, + 147, + 551, + 214 + ], + "type": "image", + "image_path": "db9d15600bdabdde2a46486d159807d8a6983f5656985209a7512845c66df966.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 220, + 216, + 305, + 291 + ], + "blocks": [ + { + "bbox": [ + 220, + 216, + 305, + 291 + ], + "lines": [ + { + "bbox": [ + 220, + 216, + 305, + 291 + ], + "spans": [ + { + "bbox": [ + 220, + 216, + 305, + 291 + ], + "type": "image", + "image_path": "6a3f6e054e38771ef3c81967d2cdb75eeef73f205fdcd56391a2fb77f0c315d8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 396, + 216, + 459, + 291 + ], + "blocks": [ + { + "bbox": [ + 396, + 216, + 459, + 291 + ], + "lines": [ + { + "bbox": [ + 396, + 216, + 459, + 291 + ], + "spans": [ + { + "bbox": [ + 396, + 216, + 459, + 291 + ], + "type": "image", + "image_path": "0e44869b1229331edce472255c8515b94a65abd3d7247e2b665226329dd70680.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 58, + 224, + 149, + 291 + ], + "blocks": [ + { + "bbox": [ + 75, + 216, + 127, + 224 + ], + "lines": [ + { + "bbox": [ + 75, + 216, + 127, + 224 + ], + "spans": [ + { + "bbox": [ + 75, + 216, + 127, + 224 + ], + "type": "text", + "content": "(7) DISTS/LIVE" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 224, + 149, + 291 + ], + "lines": [ + { + "bbox": [ + 58, + 224, + 149, + 291 + ], + "spans": [ + { + "bbox": [ + 58, + 224, + 149, + 291 + ], + "type": "image", + "image_path": "0326af9f7f5cd3f3299f1c5563883bf5c67de507bfee7338a96a31f9132b1153.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 149, + 224, + 222, + 291 + ], + "blocks": [ + { + "bbox": [ + 157, + 216, + 208, + 224 + ], + "lines": [ + { + "bbox": [ + 157, + 216, + 208, + 224 + ], + "spans": [ + { + "bbox": [ + 157, + 216, + 208, + 224 + ], + "type": "text", + "content": "(8) DISTS / CSIQ" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 224, + 222, + 291 + ], + "lines": [ + { + "bbox": [ + 149, + 224, + 222, + 291 + ], + "spans": [ + { + "bbox": [ + 149, + 224, + 222, + 291 + ], + "type": "image", + "image_path": "2d973d0691146ea437da3e6536d76325f95834f014072876cde781f803650d9e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 222, + 224, + 306, + 291 + ], + "blocks": [ + { + "bbox": [ + 222, + 224, + 306, + 291 + ], + "lines": [ + { + "bbox": [ + 222, + 224, + 306, + 291 + ], + "spans": [ + { + "bbox": [ + 222, + 224, + 306, + 291 + ], + "type": "image", + "image_path": "7348994fc4b8b3c7c4cddcb6fbbc39e343407ef98ae6bf1299f3d1e8f73efe46.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 306, + 224, + 391, + 291 + ], + "blocks": [ + { + "bbox": [ + 313, + 216, + 378, + 224 + ], + "lines": [ + { + "bbox": [ + 313, + 216, + 378, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 378, + 224 + ], + "type": "text", + "content": "(10) DISTS/TID2013" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 224, + 391, + 291 + ], + "lines": [ + { + "bbox": [ + 306, + 224, + 391, + 291 + ], + "spans": [ + { + "bbox": [ + 306, + 224, + 391, + 291 + ], + "type": "image", + "image_path": "b6838f96eafc679ebafef7a3eac0cac3c5b53c10e6aa0683101a8b425498a438.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 391, + 224, + 476, + 291 + ], + "blocks": [ + { + "bbox": [ + 464, + 216, + 538, + 224 + ], + "lines": [ + { + "bbox": [ + 464, + 216, + 538, + 224 + ], + "spans": [ + { + "bbox": [ + 464, + 216, + 538, + 224 + ], + "type": "text", + "content": "(12) DISTS/PIPAL" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 224, + 476, + 291 + ], + "lines": [ + { + "bbox": [ + 391, + 224, + 476, + 291 + ], + "spans": [ + { + "bbox": [ + 391, + 224, + 476, + 291 + ], + "type": "image", + "image_path": "1b593655c55c30bb10f4a7dc81877a2942d1241fc7959152027ffcb930ea8075.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 476, + 224, + 551, + 291 + ], + "blocks": [ + { + "bbox": [ + 476, + 224, + 551, + 291 + ], + "lines": [ + { + "bbox": [ + 476, + 224, + 551, + 291 + ], + "spans": [ + { + "bbox": [ + 476, + 224, + 551, + 291 + ], + "type": "image", + "image_path": "a401c25bce7c3881a1869ddce272250dada48416a17aee89fa1df3592f4fb223.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 58, + 301, + 149, + 369 + ], + "blocks": [ + { + "bbox": [ + 68, + 293, + 135, + 301 + ], + "lines": [ + { + "bbox": [ + 68, + 293, + 135, + 301 + ], + "spans": [ + { + "bbox": [ + 68, + 293, + 135, + 301 + ], + "type": "text", + "content": "(13) DeepWSD / LIVE" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 293, + 216, + 301 + ], + "lines": [ + { + "bbox": [ + 149, + 293, + 216, + 301 + ], + "spans": [ + { + "bbox": [ + 149, + 293, + 216, + 301 + ], + "type": "text", + "content": "(14) DeepWSD / CSIQ" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 301, + 149, + 369 + ], + "lines": [ + { + "bbox": [ + 58, + 301, + 149, + 369 + ], + "spans": [ + { + "bbox": [ + 58, + 301, + 149, + 369 + ], + "type": "image", + "image_path": "b900a28943f0da369d53a2fc3df3542851c6efdf5e05adf7d133deeadb4481c2.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 370, + 135, + 378 + ], + "lines": [ + { + "bbox": [ + 67, + 370, + 135, + 378 + ], + "spans": [ + { + "bbox": [ + 67, + 370, + 135, + 378 + ], + "type": "text", + "content": "(19) TOPIQ-FR / LIVE" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 370, + 216, + 378 + ], + "lines": [ + { + "bbox": [ + 149, + 370, + 216, + 378 + ], + "spans": [ + { + "bbox": [ + 149, + 370, + 216, + 378 + ], + "type": "text", + "content": "(20) TOPIQ-FR / CSIQ" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 149, + 301, + 222, + 369 + ], + "blocks": [ + { + "bbox": [ + 149, + 301, + 222, + 369 + ], + "lines": [ + { + "bbox": [ + 149, + 301, + 222, + 369 + ], + "spans": [ + { + "bbox": [ + 149, + 301, + 222, + 369 + ], + "type": "image", + "image_path": "706a2e441a5bf10849bd36bbb4fd390b685aff8f21d484440ace5c74aa10c9e2.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 222, + 301, + 306, + 369 + ], + "blocks": [ + { + "bbox": [ + 227, + 293, + 302, + 301 + ], + "lines": [ + { + "bbox": [ + 227, + 293, + 302, + 301 + ], + "spans": [ + { + "bbox": [ + 227, + 293, + 302, + 301 + ], + "type": "text", + "content": "(15) DeepWSD/TID2008" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 222, + 301, + 306, + 369 + ], + "lines": [ + { + "bbox": [ + 222, + 301, + 306, + 369 + ], + "spans": [ + { + "bbox": [ + 222, + 301, + 306, + 369 + ], + "type": "image", + "image_path": "1c47e13b27de9dd281d0e1528df25961554ac75b039be765957d9322e5b1a030.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 370, + 302, + 378 + ], + "lines": [ + { + "bbox": [ + 227, + 370, + 302, + 378 + ], + "spans": [ + { + "bbox": [ + 227, + 370, + 302, + 378 + ], + "type": "text", + "content": "(21) TOPIQ-FR/TID2008" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 306, + 301, + 391, + 369 + ], + "blocks": [ + { + "bbox": [ + 307, + 293, + 386, + 301 + ], + "lines": [ + { + "bbox": [ + 307, + 293, + 386, + 301 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 386, + 301 + ], + "type": "text", + "content": "(16) DeepWSD / TID2013" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 293, + 464, + 301 + ], + "lines": [ + { + "bbox": [ + 391, + 293, + 464, + 301 + ], + "spans": [ + { + "bbox": [ + 391, + 293, + 464, + 301 + ], + "type": "text", + "content": "(17) DeepWSD / KADID" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 301, + 391, + 369 + ], + "lines": [ + { + "bbox": [ + 306, + 301, + 391, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 301, + 391, + 369 + ], + "type": "image", + "image_path": "303d77f2e2897a3b1b2ee9d137819a61ff28ad68cd8840c2a02a1c9f06d8a10f.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 370, + 386, + 378 + ], + "lines": [ + { + "bbox": [ + 307, + 370, + 386, + 378 + ], + "spans": [ + { + "bbox": [ + 307, + 370, + 386, + 378 + ], + "type": "text", + "content": "(22) TOPIQ-FR / TID2013" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 370, + 464, + 378 + ], + "lines": [ + { + "bbox": [ + 391, + 370, + 464, + 378 + ], + "spans": [ + { + "bbox": [ + 391, + 370, + 464, + 378 + ], + "type": "text", + "content": "(23) TOPIQ-FR / KADID" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 391, + 301, + 476, + 369 + ], + "blocks": [ + { + "bbox": [ + 476, + 293, + 544, + 301 + ], + "lines": [ + { + "bbox": [ + 476, + 293, + 544, + 301 + ], + "spans": [ + { + "bbox": [ + 476, + 293, + 544, + 301 + ], + "type": "text", + "content": "(18) DeepWSD / PIPAL" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 301, + 476, + 369 + ], + "lines": [ + { + "bbox": [ + 391, + 301, + 476, + 369 + ], + "spans": [ + { + "bbox": [ + 391, + 301, + 476, + 369 + ], + "type": "image", + "image_path": "b228b9b374d8fbb1727eb9da1f440c9f5745c7b5da85eb32d68e555dcd846393.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 370, + 544, + 378 + ], + "lines": [ + { + "bbox": [ + 476, + 370, + 544, + 378 + ], + "spans": [ + { + "bbox": [ + 476, + 370, + 544, + 378 + ], + "type": "text", + "content": "(24) TOPIQ-FR / PIPAL" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 476, + 301, + 551, + 369 + ], + "blocks": [ + { + "bbox": [ + 476, + 301, + 551, + 369 + ], + "lines": [ + { + "bbox": [ + 476, + 301, + 551, + 369 + ], + "spans": [ + { + "bbox": [ + 476, + 301, + 551, + 369 + ], + "type": "image", + "image_path": "0577efd4b72d56d47be0cc0fd87c04c3062e970468112163f68d78e68a64b002.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 58, + 379, + 149, + 446 + ], + "blocks": [ + { + "bbox": [ + 58, + 379, + 149, + 446 + ], + "lines": [ + { + "bbox": [ + 58, + 379, + 149, + 446 + ], + "spans": [ + { + "bbox": [ + 58, + 379, + 149, + 446 + ], + "type": "image", + "image_path": "1a6c09933cdf42aaf0e35ac47beee01c6b8cfdb1341d60dcb9bb8741bc48ca3e.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 448, + 126, + 456 + ], + "lines": [ + { + "bbox": [ + 77, + 448, + 126, + 456 + ], + "spans": [ + { + "bbox": [ + 77, + 448, + 126, + 456 + ], + "type": "text", + "content": "(25) Ours / LIVE" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 466, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 55, + 466, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 555, + 499 + ], + "type": "text", + "content": "Figure 2. Scatterplot of prediction results for different methods across databases. We performed normalization on both labels and predictions and 4-parameter logistic function is utilized before comparison, as suggested in [24], and the values of the lower-better methods are reversed before regression." + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 149, + 379, + 222, + 445 + ], + "blocks": [ + { + "bbox": [ + 149, + 379, + 222, + 445 + ], + "lines": [ + { + "bbox": [ + 149, + 379, + 222, + 445 + ], + "spans": [ + { + "bbox": [ + 149, + 379, + 222, + 445 + ], + "type": "image", + "image_path": "9ddd7987f48ee6dcb5dc254dcd964fcac42e811ef6348967e350e3e024022048.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 448, + 208, + 456 + ], + "lines": [ + { + "bbox": [ + 157, + 448, + 208, + 456 + ], + "spans": [ + { + "bbox": [ + 157, + 448, + 208, + 456 + ], + "type": "text", + "content": "(26) Ours / CSIQ" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 227, + 379, + 306, + 445 + ], + "blocks": [ + { + "bbox": [ + 227, + 379, + 306, + 445 + ], + "lines": [ + { + "bbox": [ + 227, + 379, + 306, + 445 + ], + "spans": [ + { + "bbox": [ + 227, + 379, + 306, + 445 + ], + "type": "image", + "image_path": "95faa752d5380ef5f74664b1073e6cf952c847ca9cc066dc3fce7906ebb54a68.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 448, + 294, + 456 + ], + "lines": [ + { + "bbox": [ + 235, + 448, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 235, + 448, + 294, + 456 + ], + "type": "text", + "content": "(27) Ours / TID2008" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 306, + 379, + 391, + 445 + ], + "blocks": [ + { + "bbox": [ + 306, + 379, + 391, + 445 + ], + "lines": [ + { + "bbox": [ + 306, + 379, + 391, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 379, + 391, + 445 + ], + "type": "image", + "image_path": "e7d6d28b46cdb7486d390798fbd6997c0cc0ea1de326e11e9415bb13742135db.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 448, + 375, + 456 + ], + "lines": [ + { + "bbox": [ + 317, + 448, + 375, + 456 + ], + "spans": [ + { + "bbox": [ + 317, + 448, + 375, + 456 + ], + "type": "text", + "content": "(28) Ours / TID2013" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 391, + 379, + 476, + 445 + ], + "blocks": [ + { + "bbox": [ + 391, + 379, + 476, + 445 + ], + "lines": [ + { + "bbox": [ + 391, + 379, + 476, + 445 + ], + "spans": [ + { + "bbox": [ + 391, + 379, + 476, + 445 + ], + "type": "image", + "image_path": "8ed3e67a7cdb4f223f3afda41fcd30c4200e05cd3d66b3b809003d5bbae307c8.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 448, + 455, + 456 + ], + "lines": [ + { + "bbox": [ + 400, + 448, + 455, + 456 + ], + "spans": [ + { + "bbox": [ + 400, + 448, + 455, + 456 + ], + "type": "text", + "content": "(29) Ours / KADID" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 476, + 379, + 551, + 445 + ], + "blocks": [ + { + "bbox": [ + 476, + 379, + 551, + 445 + ], + "lines": [ + { + "bbox": [ + 476, + 379, + 551, + 445 + ], + "spans": [ + { + "bbox": [ + 476, + 379, + 551, + 445 + ], + "type": "image", + "image_path": "29fde11d053dc9d29212ff6e0bb2cb826ec244a4c54c58a8baad9f2c2daaa352.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 448, + 536, + 456 + ], + "lines": [ + { + "bbox": [ + 484, + 448, + 536, + 456 + ], + "spans": [ + { + "bbox": [ + 484, + 448, + 536, + 456 + ], + "type": "text", + "content": "(30) Ours / PIPAL" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 56 + }, + { + "bbox": [ + 55, + 521, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 712 + ], + "type": "text", + "content": "ure 2, our method demonstrates superior overall performance and consistently aligns closely with actual scores across datasets. Traditional techniques, which are typically based on handcrafted features, often struggle to capture the complex, nonlinear relationships between image distortions and perceptual quality, resulting in relatively lower accuracy across diverse datasets. End-to-end deep learning methods, while showing a marked improvement over traditional methods, exhibit notable limitations in terms of generalizability. These methods are typically trained on specific datasets, and as a result, their performance tends to degrade when evaluated on datasets that differ from the training set, struggling to maintain consistency and accuracy when applied to unseen data. Furthermore, the deep feature-based similarity methods that do not rely on training also perform well, although they tend to be less precise" + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 313, + 521, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 700 + ], + "type": "text", + "content": "than our model is. These methods capture image distortions via pretrained features, but their inability to explicitly model causal relationships results in slightly less accurate predictions, particularly when the perceptual impact of distortions varies significantly across datasets. In contrast, our approach, which leverages causal reasoning through an SCM and does not require training, excels in maintaining consistent performance across various datasets. The results demonstrate that, by focusing on causally relevant features, our model is less sensitive to the idiosyncrasies of individual datasets and is capable of producing accurate predictions for unseen images. This robustness in generalization suggests that the causal framework used in our model provides a more reliable representation of perceptual quality, independent of dataset-specific biases." + } + ] + } + ], + "index": 60 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17996" + } + ] + } + ], + "index": 61 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 102, + 299, + 180 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 2. Performance comparison results of the ablation experiments." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 102, + 299, + 180 + ], + "lines": [ + { + "bbox": [ + 59, + 102, + 299, + 180 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 299, + 180 + ], + "type": "table", + "html": "
ModelLIVE [27]CSIQ [16]TID2013 [22]
ParameterPLCCSRCCPLCCSRCCPLCCSRCC
φ = φθ0.9010.9150.9130.9160.8840.867
φ = φγ0.9290.9320.9490.9520.9090.884
φ = φη0.8430.8660.8030.8310.7860.789
", + "image_path": "b8eb0078e516f455cb7b3a9bcc7529c0f9aa8f4b6afc1a984896c167df147f9f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 182, + 295, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 182, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 67, + 182, + 295, + 258 + ], + "type": "text", + "content": "Performance comparison results of the ablation experiments conducted on three benchmark datasets, where we use OurVGG as the default architecture. The models evaluated in this study correspond to different configurations of weight assignments for " + }, + { + "bbox": [ + 67, + 182, + 295, + 258 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 67, + 182, + 295, + 258 + ], + "type": "text", + "content": ", where its values " + }, + { + "bbox": [ + 67, + 182, + 295, + 258 + ], + "type": "inline_equation", + "content": "\\phi_{\\theta}" + }, + { + "bbox": [ + 67, + 182, + 295, + 258 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 182, + 295, + 258 + ], + "type": "inline_equation", + "content": "\\phi_{\\gamma}" + }, + { + "bbox": [ + 67, + 182, + 295, + 258 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 67, + 182, + 295, + 258 + ], + "type": "inline_equation", + "content": "\\phi_{\\eta}" + }, + { + "bbox": [ + 67, + 182, + 295, + 258 + ], + "type": "text", + "content": " represent the original pretrained weights, the causal effect weights, and the residual component excluding the causal effect, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 162, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 162, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 162, + 290 + ], + "type": "text", + "content": "5.3. Ablation Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "text", + "content": "In this section, we present the abductive counterfactual experiments to evaluate the impact of different weight configurations in our model. Specifically, we analyse the results when varying the weights of " + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "text", + "content": " represents the original pretrained weights, " + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "text", + "content": " corresponds to the weight associated with the causal effect, and " + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "text", + "content": " denotes the remaining part of the model, which is the portion of " + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "text", + "content": " excluding " + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 294, + 295, + 510 + ], + "type": "text", + "content": ". The ablation analysis is conducted by systematically adjusting each of these weights and observing their influence on the model's performance. The results are summarized in Table 2, which indicate that although noncausal representations also exhibit some perceptual relevance, their lack of causal relationships limits their effectiveness. When these noncausal representations are removed from the original feature set, the model's predictive performance improves. This suggests that incorporating those features that capture causal effects can enhance the model's ability to make more accurate and reliable quality predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 517, + 185, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 185, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 185, + 529 + ], + "type": "text", + "content": "5.4. Qualitative Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "In this section, we visually demonstrate the effect of causal intervention on different channels. As shown in Figure 3, we demonstrate the visualization effects of causal intervention between the same reference image and different distorted images (blurred and additive Gaussian noise) via the Our-VGG architecture. These visualizations reveal the varying levels of sensitivity each perceptual channel exhibits to different types of distortions in the pretrained model, thus providing a clear indication of how such distortions alter the perceptual characteristics of the image. The horizontal axis in Figure 3 represents the channels, and the color intensity corresponds to the strength of the causal relationship, with brighter colors indicating a stronger causal relationship. Causal features reflect attributes that are crucial for human perception of image quality, as they directly" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 381, + 137 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 381, + 137 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 381, + 137 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 381, + 137 + ], + "type": "image", + "image_path": "99d0e3ba122d38ba067257c08ae2064ded16a448839924f0781da2a8bf3f37a1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 138, + 356, + 145 + ], + "lines": [ + { + "bbox": [ + 342, + 138, + 356, + 145 + ], + "spans": [ + { + "bbox": [ + 342, + 138, + 356, + 145 + ], + "type": "text", + "content": "Blur" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 563, + 146 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 563, + 146 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 563, + 146 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 563, + 146 + ], + "type": "image", + "image_path": "209574be8c01c266f39f3bf848e90c30162269ad8877df937bdd5037de674695.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 160, + 381, + 225 + ], + "blocks": [ + { + "bbox": [ + 317, + 160, + 381, + 225 + ], + "lines": [ + { + "bbox": [ + 317, + 160, + 381, + 225 + ], + "spans": [ + { + "bbox": [ + 317, + 160, + 381, + 225 + ], + "type": "image", + "image_path": "604315eedc6e3a5437f5e12e003c149b714e9ac3a6a7555b0f8efb36ffc5806c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 227, + 370, + 234 + ], + "lines": [ + { + "bbox": [ + 328, + 227, + 370, + 234 + ], + "spans": [ + { + "bbox": [ + 328, + 227, + 370, + 234 + ], + "type": "text", + "content": "Gaussian Noise" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 389, + 158, + 563, + 208 + ], + "blocks": [ + { + "bbox": [ + 389, + 158, + 563, + 208 + ], + "lines": [ + { + "bbox": [ + 389, + 158, + 563, + 208 + ], + "spans": [ + { + "bbox": [ + 389, + 158, + 563, + 208 + ], + "type": "image", + "image_path": "fabbb6897c6c1d84f8cbc3c14b41be19899015eec8290a7faa8f40c55852ba0a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 389, + 209, + 561, + 236 + ], + "blocks": [ + { + "bbox": [ + 389, + 209, + 561, + 236 + ], + "lines": [ + { + "bbox": [ + 389, + 209, + 561, + 236 + ], + "spans": [ + { + "bbox": [ + 389, + 209, + 561, + 236 + ], + "type": "image", + "image_path": "6518ea22d993259ab7cc85da313a2258635a6defa5d1b953dd7b477d39dc8ce3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 247, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 291 + ], + "type": "text", + "content": "Figure 3. Effects of causal interventions on the perceptual representations of images under various distortions. The displayed channels, from top to bottom, correspond to the second, fourth, and fifth stages of the VGG-16 architecture." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 312, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 468 + ], + "type": "text", + "content": "influence perceptual judgments. In contrast, noncausal features typically correspond to higher-level semantic aspects of the image, such as object recognition, which are less relevant for perceptual quality but play a significant role in high level vision tasks. This distinction is supported by the observation that deeper layers of the network are more attuned to classification tasks, while quality-related features are predominantly represented in the shallower layers, underscoring the channel-specific nature of perceptual changes. By isolating the causal effects, we can identify specific features or regions of the image that are most sensitive to certain types of distortions, thus further elucidating how different distortions affect overall image perception." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "In conclusion, we proposed an FR-IQA method grounded in abductive counterfactual inference to analyse the causal relationships between deep network features and perceptual distortions. We investigated the causal effects of deep features on human perception and incorporated causal reasoning with feature comparisons, constructing a model capable of addressing various complex distortions in different IQA scenarios. The proposed approach demonstrated strong generalizability, as the analysis of perceptual causal correlations was shown to be independent of the backbone architecture, making it adaptable across diverse deep networks. Additionally, abductive counterfactual experiments validated the effectiveness of the proposed causal framework, confirming the model's superior perceptual relevance and interpretability of quality scores. The experimental results reinforced the robustness and competitiveness of the method, which achieved consistent performance across multiple IQA benchmarks." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17997" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Hengrui Cai, Yixin Wang, Michael Jordan, and Rui Song. On learning necessary and sufficient causal graphs. In Advances in Neural Information Processing Systems, pages 42148-42160. Curran Associates, Inc., 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 190 + ], + "type": "text", + "content": "[2] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 33:2404-2418, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "text", + "content": "[3] Li-Heng Chen, Christos G. Bampis, Zhi Li, Andrey Norkin, and Alan C. Bovik. Proxiqa: A proxy approach to perceptual optimization of learned image compression. IEEE Transactions on Image Processing, 30:360-373, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "type": "text", + "content": "[4] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 295, + 346 + ], + "type": "text", + "content": "[5] Keyan Ding, Yi Liu, Xueyi Zou, Shiqi Wang, and Kede Ma. Locally adaptive structure and texture similarity for image quality assessment. In Proceedings of the 29th ACM International Conference on Multimedia, page 2483-2491, New York, NY, USA, 2021. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "type": "text", + "content": "[6] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Comparison of full-reference image quality models for optimization of image processing systems. International Journal of Computer Vision, 129(4):1258-1281, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 445 + ], + "type": "text", + "content": "[7] Fei Gao, Yi Wang, Panpeng Li, Min Tan, Jun Yu, and Yani Zhu. Deepsim: Deep similarity for image quality assessment. Neurocomputing, 257:104-114, 2017. Machine Learning and Signal Processing for Big Multimedia Analysis. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "text", + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "type": "text", + "content": "[9] Linmei Hu, Ziwei Chen, Ziwang Zhao, Jianhua Yin, and Liqiang Nie. Causal inference for leveraging image-text matching bias in multi-modal fake news detection. IEEE Transactions on Knowledge and Data Engineering, 35(11): 11141-11152, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 602 + ], + "type": "text", + "content": "[10] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S. Ren, and Dong Chao. Pipal: A large-scale image quality assessment dataset for perceptual image restoration. In Computer Vision - ECCV 2020, pages 633-651, Cham, 2020. Springer International Publishing. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 647 + ], + "type": "text", + "content": "[11] Jongyoo Kim and Sanghoon Lee. Deep learning of human visual sensitivity in image quality assessment framework. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1676-1684, 2017. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 714 + ], + "type": "text", + "content": "[12] Junho Kim, Byung-Kwan Lee, and Yong Man Ro. Demystifying causal features on adversarial examples and causal inoculation for robust network by adversarial instrumental variable regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12302-12312, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[13] Xuting Lan, Mingliang Zhou, Xueyong Xu, Xuekai Wei, Xingran Liao, Huayan Pu, Jun Luo, Tao Xiang, Bin Fang, and Zhaowei Shang. Multilevel feature fusion for end-to-end blind image quality assessment. IEEE Transactions on Broadcasting, 69(3):801-811, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "text", + "content": "[14] Shanshan Lao, Yuan Gong, Shuwei Shi, Sidi Yang, Tianhe Wu, Jiahao Wang, Weihao Xia, and Yujiu Yang. Attentions help cnns see better: Attention-based hybrid image quality assessment network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1140-1149, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "text", + "content": "[15] Valero Laparra, Johannes Balle, Alexander Berardino, and Eero P. Simoncelii. Perceptual image quality assessment using a normalized laplacian pyramid. In Human Vision and Electronic Imaging, pages 43-48, United States, 2016. Society for Imaging Science and Technology. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[16] Eric Cooper Larson and Damon Michael Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of Electronic Imaging, 19 (1):011006, 2010. 5, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 298, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 553, + 374 + ], + "type": "text", + "content": "[17] Xingran Liao, Baoliang Chen, Hanwei Zhu, Shiqi Wang, Mingliang Zhou, and Sam Kwong. Deepwsd: Projecting degradations in perceptual space to wasserstein distance in deep feature space. In Proceedings of the 30th ACM International Conference on Multimedia, page 970-978, New York, NY, USA, 2022. Association for Computing Machinery. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 376, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 431 + ], + "type": "text", + "content": "[18] Xingran Liao, Xuekai Wei, Mingliang Zhou, Zhengguo Li, and Sam Kwong. Image quality assessment: Measuring perceptual degradation via distribution measures in deep feature spaces. IEEE Transactions on Image Processing, 33:4044-4059, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "text", + "content": "[19] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience, pages 1–3, 2019. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "text", + "content": "[20] Bing Liu, Dong Wang, Xu Yang, Yong Zhou, Rui Yao, Zhiwen Shao, and Jiaqi Zhao. Show, deconfound and tell: Image captioning with causal inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18041-18050, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 535, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 590 + ], + "type": "text", + "content": "[21] Nikolay Ponomarenko, Vladimir Lukin, Alexander Zelensky, Karen Egiazarian, Marco Carli, and Federica Battisti. Tid2008-a database for evaluation of full-reference visual quality assessment metrics. Advances of Modern Radioelectronics, 10(4):30-45, 2009. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 590, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 656 + ], + "type": "text", + "content": "[22] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, and C.-C. Jay Kuo. Image database tid2013: Peculiarities, results and perspectives. Signal Processing: Image Communication, 30: 57-77, 2015. 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[23] Ekta Prashnani, Hong Cai, Yasamin Mostofi, and Pradeep Sen. Pieapp: Perceptual image-error assessment through pairwise preference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1808-1817, 2018. 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17998" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 192 + ], + "type": "text", + "content": "[24] Ann Marie Rohaly, Philip J. Corriveau, John M. Libert, Arthur A. Webster, Vittorio Baroncini, John Beerends, Jean-Louis Blin, Laura Contin, Takahiro Hamada, David Harrison, Andries P. Hekstra, Jeffrey Lubin, Yukihiro Nishida, Ricardo Nishihara, John C. Pearson, Antonio Franca Pessoa, Neil Pickford, Alexander Schertz, Massimo Visca, Andrew B. Watson, and Stefan Winkler. Video Quality Experts Group: current results and future directions. In Visual Communications and Image Processing 2000, pages 742 - 753. International Society for Optics and Photonics, SPIE, 2000." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 194, + 294, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 225 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 225 + ], + "type": "text", + "content": "[25] Ladan Shams and Ulrik R. Beierholm. Causal inference in perception. Trends in Cognitive Sciences, 14(9):425-432, 2010. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 228, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 258 + ], + "type": "text", + "content": "[26] H.R. Sheikh and A.C. Bovik. Image information and visual quality. IEEE Transactions on Image Processing, 15(2):430-444, 2006. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 260, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 260, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 260, + 294, + 303 + ], + "type": "text", + "content": "[27] H.R. Sheikh, M.F. Sabir, and A.C. Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on Image Processing, 15(11): 3440-3451, 2006. 5, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "type": "text", + "content": "[28] Wenhao Shen, Mingliang Zhou, Jun Luo, Zhengguo Li, and Sam Kwong. Graph-represented distribution similarity index for full-reference image quality assessment. IEEE Transactions on Image Processing, 33:3075-3089, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[29] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, page 1-14, 2015. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 448 + ], + "type": "text", + "content": "[30] Xue Song, Jiequan Cui, Hanwang Zhang, Jingjing Chen, Richang Hong, and Yu-Gang Jiang. Doubly abductive counterfactual inference for text-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9162-9171, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "text", + "content": "[31] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "text", + "content": "[32] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems and Computers, 2003, pages 1398-1402. IEEE, 2003. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "type": "text", + "content": "[33] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[34] Wufeng Xue, Lei Zhang, Xuanqin Mou, and Alan C. Bovik. Gradient magnitude similarity deviation: A highly efficient perceptual image quality index. IEEE Transactions on Image Processing, 23(2):684-695, 2014. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "text", + "content": "[35] Mengyue Yang, Xinyu Cai, Furui Liu, Weinan Zhang, and Jun Wang. Specify robust causal representation from mixed observations. New York, NY, USA, 2023. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[36] Zhongqi Yue, Hanwang Zhang, Qianru Sun, and Xian-Sheng Hua. Interventional few-shot learning. In Advances in Neural Information Processing Systems, pages 2734-2746. Curran Associates, Inc., 2020. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 318 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[37] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. Fsim: A feature similarity index for image quality assessment. IEEE Transactions on Image Processing, 20(8):2378-2386, 2011. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[38] Lin Zhang, Ying Shen, and Hongyu Li. Vsi: A visual saliency-induced index for perceptual image quality assessment. IEEE Transactions on Image Processing, 23(10): 4270-4281, 2014. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[39] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "text", + "content": "[40] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10242-10251, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 318 + ], + "type": "text", + "content": "[41] Mingliang Zhou, Shujun Lang, Taiping Zhang, Xingran Liao, Zhaowei Shang, Tao Xiang, and Bin Fang. Attentional feature fusion for end-to-end blind image quality assessment. IEEE Transactions on Broadcasting, 69(1):144-152, 2022. 1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17999" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/960040c5-ed98-43eb-9b66-fa83e44db35d_content_list.json b/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/960040c5-ed98-43eb-9b66-fa83e44db35d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b7469fa00c18cdaeba3a4d62c67ba7d8fe6f324c --- /dev/null +++ b/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/960040c5-ed98-43eb-9b66-fa83e44db35d_content_list.json @@ -0,0 +1,1812 @@ +[ + { + "type": "text", + "text": "Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shashwath Bharadwaj, Ruangrawee Kitichotkul, Akshay Agarwal and Vivek K Goyal Boston University", + "bbox": [ + 158, + 180, + 836, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shash@bu.edu", + "bbox": [ + 441, + 219, + 550, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Readout multiplexing is a promising solution to overcome hardware limitations and data bottlenecks in imaging with single-photon detectors. Conventional multiplexed readout processing creates an upper bound on photon counts at a very fine time scale, where frames with multiple detected photons must either be discarded or allowed to introduce significant bias. We formulate multiphoton coincidence resolution as an inverse imaging problem and introduce a solution framework to probabilistically resolve the spatial locations of photon incidences. Specifically, we develop a theoretical abstraction of row-column multiplexing and a model of photon events that make readouts ambiguous. Using this, we propose a novel estimator that spatially resolves up to four coincident photons. Monte Carlo simulations show that our proposed method increases the peak signal-to-noise ratio (PSNR) of reconstruction by 3 to 4 dB compared to conventional methods under optimal incident flux conditions. Additionally, this method reduces the required number of readout frames to achieve the same mean-squared error as other methods by a factor of $\\sim 4$ . Finally, our method matches the Cramér-Rao bound for detection probability estimation for a wider range of incident flux values compared to conventional methods. While demonstrated for a specific detector type and readout architecture, this method can be extended to more general multiplexing with different detector models.", + "bbox": [ + 91, + 301, + 483, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 723, + 220, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single-photon detectors have been widely studied for use in applications like biological imaging [7, 34], lidar [15, 29], and quantum optics [9] due to properties such as single-photon sensitivity and high-precision time-tagging of photon arrivals. Single-photon avalanche diodes (SPADs) have been of particular interest due to compatibility with CMOS manufacturing techniques and room temperature operation [11]. However, they have low detection efficiencies (especially at mid infrared wavelengths), which means that only a small fraction of incident light is detected as pho", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tons. Superconducting nanowire single-photon detectors (SNSPDs) are emerging devices that offer near-unity quantum efficiency across a range of wavelengths, low dark counts, and low jitter, making them compelling alternatives to SPADs [38]. However, the requirement of cryogenic cooling to maintain superconductivity has largely restricted their use to laboratory experiments.", + "bbox": [ + 511, + 270, + 906, + 377 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single-photon imaging in typical commercial applications requires arrays of single-photon detectors [28]. While megapixel SPAD arrays have already been deployed in mass consumer products [3, 25], high data transfer rates from these arrays for further processing remain a bottleneck. For instance, typical room-scale ranging experiments with SPAD array-based single-photon lidar can have data generation rates of the order of gigabits per second when the array resolution approaches the megapixel scale. It is attractive to compress the readout to less than a stream of all digitized detection times. Methods like in-pixel histogramming [18], histogram compression through sketching [14], equidepth histogramming [19, 32], and differential readouts [36] have been proposed to address high data transfer rates. These approaches to avoid the cost of high temporal resolution are complementary to our emphasis on transverse spatial dimensions in this work.", + "bbox": [ + 509, + 381, + 906, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Implementing high spatial resolution SNSPD arrays is a significant challenge. As the number of pixels in the array grows, the heat load introduced by reading out each pixel individually becomes incompatible with the cryogenics of the system [12]. To mitigate this, recent efforts have focused on addressing groups of pixels in the readout, starting with Allman et al. [1] who introduced row-column readout multiplexing. In this mechanism, readouts indicate photon detection at each row and each column of the array instead of every pixel, which reduces the required number of readout lines for an $n \\times n$ array from $n^2$ to $2n$ . Wollman et al. [37] extended this scheme to demonstrate the first kilopixel SNSPD array. McCaughan et al. [24] introduced the thermally coupled imager with time-of-flight multiplexing, where separate readout lines for each row and column are replaced with a single bus for all rows and all columns. This makes the required number of readouts independent of", + "bbox": [ + 509, + 643, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11406", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the array size. Recently, Oripov et al. [27] leveraged this method to demonstrate a 400,000 pixel superconducting camera, showing the viability of this scheme to approach the sizes of commercial imaging arrays.", + "bbox": [ + 93, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the gain in scaling array sizes, readout multiplexing requires the incident photon flux to be very low for unambiguous signal reconstruction, limiting the array's photon count rate. When two or more photons are incident on the array within the period of a single acquisition, row-column readouts can give an ambiguous set of candidate locations for their incidences. Traditional reconstruction techniques [37] do not distinguish between these locations and assume equal probabilities of detection at each candidate pixel. This results in the misattribution of photon counts to locations where no photon incidences occurred. Alternatively, since readouts with a single detected photon provide the exact row and column indices where the photon was incident on the array, ambiguous readouts with multiple detections can be discarded for signal reconstruction. However, this leads to high variance in the reconstruction (especially at high incident fluxes) and underutilization of the spatial dimensions of the array by requiring most readouts to contain detections at a single pixel. Resolving the spatial locations of multiphoton coincidences using multiplexed readouts is an ill-posed inverse problem.", + "bbox": [ + 93, + 152, + 480, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we propose a solution framework to leverage most of the measured data to reduce the mean-squared error of image reconstruction. Specifically, we develop a multiphoton estimator that redistributes photon counts from ambiguous readouts to candidate pixels such that an approximate likelihood of observations is maximized. A key choice for our algorithm design is the omission of spatial priors in order to not obscure the source of performance gain. Spatial priors can be used in conjunction with the developed method for further improvements. Monte Carlo simulations show that our proposed method increases the peak signal-to-noise ratio (PSNR) of reconstruction by 3 to $4\\mathrm{dB}$ compared to conventional methods under optimal incident flux conditions. Further, the optimal incident photon flux for our method is $\\sim 1.4$ photons per readout which is 0.5 to 1.1 photons per readout higher than conventional methods. We also demonstrate that our estimator matches the Cramér-Rao Bound (CRB) for the estimation of photon detection probabilities at each pixel of an array across a range of incident photon flux values. Finally, our method achieves a factor of $\\sim 4$ reduction in the required number of readout frames for the same mean-squared error of image reconstruction. These results demonstrate that our proposed estimator is well-suited for high-flux, low-latency imaging.", + "bbox": [ + 93, + 469, + 480, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 94, + 845, + 230, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image reconstruction from multiplexed readouts.", + "bbox": [ + 94, + 872, + 480, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compressed sensing methods have been explored for the", + "bbox": [ + 93, + 887, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "design of multiplexed readout architectures in areas like positron emission tomography (PET) [10, 13, 22, 26], biological imaging [6, 30] and gamma-ray imaging [4, 5]. While readout multiplexing for specific imaging applications has been extensively studied [8, 20], the focus of these works is on the modeling of optical and electrical properties of the imaging system such as the shape of the wavefront of the detected light [2, 31], coded apertures for rejecting multipixel events [5] and IC design for resistor multiplexing [13]. The explicit study of individual photon events measured by single-photon detectors and their impact on the multiplexed readouts does not yet exist. Our work introduces a theoretical study of the readouts based on the combinatorics of photon events and an approximate maximum likelihood estimation of photon detection probabilities at each pixel. We also analyze the performance of a novel multiphoton estimator with variations in imaging parameters. Our work is closest to the analysis presented by van den Berg et al. [33], where group-testing inspired surface codes are studied for deterministic multiphoton coincidence resolution in time-to-digital converter readouts from SPAD arrays. However, our work specifically addresses intensity imaging with SNSPD arrays where a small probability of error is tolerable. We provide a probabilistic model of the readouts and solutions for image reconstruction from a specific device architecture for up to four coincident photons.", + "bbox": [ + 516, + 90, + 903, + 497 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Imaging with SNSPD arrays. While large-scale SNSPD array development is still in the nascent stages, several works have conducted proof-of-principle imaging experiments with prototypes. Zhao et al. introduced a delay-line-based superconducting imager with an effective pixel count of $\\sim 590$ [39]. This method inspired the thermally coupled imager [24], which in turn resulted in the 400,000-pixel camera by Oripov et al. [27]. Subsequent efforts to use SNSPD arrays for imaging [16, 21, 35] involve intricate hardware control such as bias-current sweeping, use of idle pixels, and geometric design of the nanowire structure. Our solution is purely computational and does not require additional hardware modifications. In addition, the modeling of incident photon events and their multiplexed readouts can be extended to several detector models in high-energy physics or particle detection. While the method described in this work provides a framework to resolve an arbitrarily large number of coincident photons, we limit the discussion in the next sections to up to four coincident photons to make the demonstrated solutions computationally tractable and easier to illustrate and analyze.", + "bbox": [ + 516, + 500, + 903, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Measurement and Readout Model", + "text_level": 1, + "bbox": [ + 516, + 830, + 821, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Figure 1 illustrates imaging a scene using an $n \\times n$ SNSPD array with row-column readouts. Measurements are read out after integrating the incident signal for a fixed time in", + "bbox": [ + 516, + 856, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11407", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/52a45b326e446bd427013d7350541df707f25a8777cfe0c29caa2082ec7d5307.jpg", + "image_caption": [ + "Figure 1. Schematic of image reconstruction with row-column readouts. (Left) Multiple frames of a ground truth image are measured using an SNSPD array to give row-column readout vectors. (Right) Comparison of conventional and proposed reconstruction techniques to estimate the ground truth." + ], + "image_footnote": [], + "bbox": [ + 96, + 104, + 911, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "terval. Let $\\Lambda \\in \\mathbb{R}^{n\\times n}$ be the flux of an $n\\times n$ ground truth image in that integration time. The number of photons arriving at pixel $(i,j)$ at a discrete time index $t$ is modeled as", + "bbox": [ + 88, + 415, + 480, + 474 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nX _ {i j} ^ {t} \\sim \\operatorname {P o i s s o n} (\\Lambda_ {i j}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 474, + 482, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We assume pixel saturation with the detection of a single photon, i.e., the detectors are not photon-number resolving [23]. Thus, we define binary-valued photon-incidence indicator $Y^{t} \\in \\{0,1\\}^{n \\times n}$ by $Y_{ij}^{t} = \\mathbf{1}[X_{ij}^{t} > 0]$ , where $\\mathbf{1}[\\cdot]$ is the indicator function. Then, $Y^{t}$ is a matrix where each entry is an independent Bernoulli random variable $Y_{ij}^{t} \\sim \\mathrm{Bernoulli}(q_{ij})$ , where", + "bbox": [ + 89, + 498, + 483, + 604 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq _ {i j} = 1 - e ^ {- \\Lambda_ {i j}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 616, + 482, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is the probability of photon detection at pixel $(i,j)$ .", + "bbox": [ + 89, + 643, + 429, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each measurement consists of a row readout $R^t \\in \\{0,1\\}^n$ and a column readout $C^t \\in \\{0,1\\}^n$ :", + "bbox": [ + 89, + 659, + 483, + 691 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {i} ^ {t} = \\mathbf {1} \\left[ X _ {i j} ^ {t} > 0 \\text {f o r a n y} j \\in \\{1, \\dots , n \\} \\right], \\tag {3a}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 700, + 482, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC _ {j} ^ {t} = \\mathbf {1} \\left[ X _ {i j} ^ {t} > 0 \\text {f o r a n y} i \\in \\{1, \\dots , n \\} \\right]. \\tag {3b}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 720, + 482, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We then define a readout frame at time $t$ as $(R^t, C^t)$ . When multiple photons are incident on the array within a single integration period, the problem of resolving the spatial locations of their incidence is ill-posed. As an example, consider a $2 \\times 2$ array with pixels numbered from 1 to 4 as shown in event $E_0$ of Fig. 2. The frame ([1,1],[1,1]) could have resulted from photons detected at pixels $\\{1,4\\}$ or $\\{2,3\\}$ or any of: $\\{1,2,3\\}$ , $\\{1,2,4\\}$ , $\\{1,3,4\\}$ , $\\{2,3,4\\}$ , or $\\{1,2,3,4\\}$ . More generally, a readout frame is unambiguous if $\\sum_{i=1}^{n} R_i^t = 1$ or $\\sum_{j=1}^{n} C_j^t = 1$ ; otherwise, it is", + "bbox": [ + 89, + 748, + 483, + 904 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ambiguous. Our problem is to estimate $\\Lambda$ from measurements in $T$ frames $\\{(R^t,C^t)\\}_{t = 1}^T$ . In particular, we are interested in exploiting the information contained in ambiguous readout frames.", + "bbox": [ + 511, + 415, + 906, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Estimators", + "text_level": 1, + "bbox": [ + 511, + 498, + 630, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the three estimators compared in this study, namely the naive, single-photon, and multiphoton estimators. Each is formulated first as an estimator $\\widehat{q}_{ij}$ for the probability $q_{ij}$ defined in (2). The incident flux at each pixel can then be estimated elementwise as", + "bbox": [ + 511, + 526, + 906, + 602 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\Lambda} _ {i j} = - \\log \\left(1 - e ^ {- \\widehat {q} _ {i j}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 618, + 906, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Naive Estimator", + "text_level": 1, + "bbox": [ + 511, + 679, + 674, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The naive estimator (NE) assumes that a photon is detected at each of the candidate pixels in ambiguous readout frames. Thus, the probability of detecting a photon at pixel $(i,j)$ is estimated as", + "bbox": [ + 511, + 704, + 906, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {q} _ {i j, \\mathrm {n a i v e}} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} R _ {i} ^ {t} C _ {j} ^ {t}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 768, + 906, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The naive estimator overestimates $q_{ij}$ because it misattributes photon counts in ambiguous frames to pixels where no photon incidence occurred. The extra counts imputed by this scheme produce ghost spots in the reconstructed image, as depicted in Fig. 1.", + "bbox": [ + 511, + 824, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11408", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/47c47aa4a7e1833e48837c1ee4d78c86baf585f5be0b9ff544630fed751bbf23.jpg", + "image_caption": [ + "Figure 2. Row-column readout frames for a $2 \\times 2$ array. Yellow spots indicate incident photon locations. Accounting for pixel saturation, 16 photon detection events are possible with this array. Row-column readout results in 9 unambiguous events $E_0, E_1, \\ldots, E_8$ and 7 events $E_9, E_{10}, \\ldots, E_{15}$ that are ambiguous because they all result in $R^t = [1,1]$ and $C^t = [1,1]$ . Pixel numbers are indicated in event $E_0$ , which corresponds to an empty readout." + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 480, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The positive bias of the naive estimator is", + "bbox": [ + 89, + 422, + 367, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {E} \\left[ \\widehat {q} _ {i j, \\text {n a i v e}} \\right] - q _ {i j} \\\\ = \\left(1 - q _ {i j}\\right) \\left(1 - \\prod_ {k \\in R _ {i} ^ {\\prime}} \\left(1 - q _ {k j}\\right)\\right) \\left(1 - \\prod_ {\\ell \\in C _ {j} ^ {\\prime}} \\left(1 - q _ {i \\ell}\\right)\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 446, + 483, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R_{i}^{\\prime}$ is the set of rows except row $i$ and $C_j^\\prime$ is the set of columns except column $j$ . A derivation is given in the supplement.", + "bbox": [ + 89, + 540, + 483, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Single-Photon Estimator", + "text_level": 1, + "bbox": [ + 89, + 594, + 316, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A simple way to prevent misattributions and thus reduce bias is to only use frames with a single detected photon in the reconstruction since these unambiguously give the spatial locations of photon incidence. We thus define the single-photon estimator (SPE):", + "bbox": [ + 89, + 617, + 483, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {q} _ {i j, \\text {s i n g l e}} = \\left\\{ \\begin{array}{l l} \\frac {N _ {i j}}{N _ {i j} + N _ {0}}, & \\text {i f} N _ {i j} + N _ {0} > 0; \\\\ 0, & \\text {i f} N _ {i j} + N _ {0} = 0, \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 700, + 480, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{ij}$ is the number of frames where a photon is detected only at pixel $(i,j)$ , and $N_0$ is the number of frames without any detected photons across the whole array. The SPE is unbiased when $N_{i,j} + N_0 > 0$ , as shown in the supplement. However, the discarding of frames with multiple detected photons results in high variance.", + "bbox": [ + 89, + 750, + 483, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Multiphoton Estimator", + "text_level": 1, + "bbox": [ + 89, + 848, + 305, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now, we develop a multiphoton estimator (ME) that uses photon counts from ambiguous readouts for reconstruction.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For simplicity, here we derive expressions for detection probability estimates at each pixel of a $2 \\times 2$ array and illustrate how this estimator maximizes an approximate likelihood of observations. The general form of the ME is presented in the supplement. If each pixel were read out individually, the likelihood of observations could be written as", + "bbox": [ + 511, + 90, + 905, + 194 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\prod_ {i = 1} ^ {2} \\prod_ {j = 1} ^ {2} (1 - q _ {i j}) ^ {v _ {i j}} \\left(q _ {i j}\\right) ^ {f _ {i j}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 194, + 903, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $v_{ij}$ is the total number of frames with no detections at pixel $(i,j)$ and $f_{ij}$ is the total number of frames with a detection at pixel $(i,j)$ . Then, the maximum likelihood estimate $\\widehat{q}_{ij}^{\\mathrm{ML}}$ could be computed as", + "bbox": [ + 511, + 239, + 905, + 304 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {q} _ {i j} ^ {\\mathrm {M L}} = \\frac {f _ {i j}}{T}, \\quad i = 1, 2, j = 1, 2, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 314, + 903, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ is the total number of measured frames.", + "bbox": [ + 511, + 353, + 834, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With row-column readouts, our knowledge of locations of photon detections is reduced. We define $v_{ij}$ and $f_{ij}$ from unambiguous frames only and find that the likelihood of observations becomes", + "bbox": [ + 511, + 369, + 905, + 426 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\left\\{\\left(R ^ {t}, C ^ {t}\\right) \\right\\} _ {t = 1} ^ {T}; q\\right) = U (q) A (q), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 439, + 903, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with the factor", + "bbox": [ + 513, + 469, + 614, + 482 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nU (q) = \\prod_ {i = 1} ^ {2} \\prod_ {j = 1} ^ {2} \\left(1 - q _ {i j}\\right) ^ {v _ {i j}} \\left(q _ {i j}\\right) ^ {f _ {i j}} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 493, + 903, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from unambiguous frames and the factor", + "bbox": [ + 511, + 546, + 784, + 560 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA (q) = \\left(q _ {1 1} q _ {2 2} + q _ {1 2} q _ {2 1} - q _ {1 1} q _ {1 2} q _ {2 1} q _ {2 2}\\right) ^ {M _ {9}} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 571, + 903, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from ambiguous frames. Here, $M_9$ is the total number of ambiguous readout frames due to events $E_9,\\ldots ,E_{15}$ shown in Fig. 2.", + "bbox": [ + 511, + 601, + 905, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The factor in (12) results from the sum of probabilities $\\sum_{k=9}^{15} \\mathbb{P}(E_k)$ . This makes it challenging to find analytical expressions for the maximum likelihood estimates of detection probabilities. However, if we can estimate the fractions of the $M_9$ frames resulting from each of the events $E_9, \\ldots, E_{15}$ , we can re-write (12) as a product of probabilities of events $E_9, \\ldots, E_{15}$ to express (10) similarly to (8) and obtain analytical expressions for approximate maximum likelihood estimates $\\widehat{q}_{ij}^{\\mathrm{a}}$ . We first find expressions for the conditional probabilities $g_9, \\ldots, g_{15}$ of events $E_9, \\ldots, E_{15}$ , given that an ambiguous frame was read out. For example, $g_9$ is the conditional probability of event $E_9$ given that a frame $R^t = [1,1]$ , $C^t = [1,1]$ was measured and is given by", + "bbox": [ + 511, + 647, + 906, + 858 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {9} = \\frac {\\mathbb {P} \\left(E _ {9}\\right)}{\\sum_ {k = 9} ^ {1 5} \\mathbb {P} \\left(E _ {k}\\right)}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 868, + 903, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11409", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the inherent limitation of having a finite amount of data, computing $\\mathbb{P}(E_9),\\ldots ,\\mathbb{P}(E_{15})$ is hindered by ambiguities. Thus, we estimate these probabilities using the measured single-photon frames, which are unambiguous and unbiased. For example, an estimate of $\\mathbb{P}(E_9)$ can be calculated as", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathbb {P}} (E _ {9}) = \\widehat {q} _ {1 1} ^ {\\mathrm {s}} \\widehat {q} _ {2 2} ^ {\\mathrm {s}} (1 - \\widehat {q} _ {1 2} ^ {\\mathrm {s}}) (1 - \\widehat {q} _ {2 1} ^ {\\mathrm {s}}). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 186, + 482, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar expressions can be derived for $\\widehat{\\mathbb{P}}(E_{10}),\\ldots,\\widehat{\\mathbb{P}}(E_{15})$ which are then substituted in (13) in place of $\\mathbb{P}(E_9),\\dots ,\\mathbb{P}(E_{15})$ to get estimates $\\widehat{g}_9,\\dots ,\\widehat{g}_{15}$ of conditional probabilities $g_{9},\\ldots ,g_{15}$ . The fractions of the $M_9$ frames resulting from events $E_{9},\\ldots E_{15}$ can then be estimated as $\\widehat{g}_9M_9,\\ldots ,\\widehat{g}_{15}M_9$ . Finally, these estimates can be used to simplify (12) to a product of probabilities and obtain the approximate maximum likelihood estimates", + "bbox": [ + 89, + 212, + 483, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widehat {q} _ {1 1} ^ {\\mathrm {a}} = \\frac {M _ {1} + M _ {5} + M _ {6} + M _ {9} (\\widehat {g} _ {9} + \\widehat {g} _ {1 1} + \\widehat {g} _ {1 2} + \\widehat {g} _ {1 3} + \\widehat {g} _ {1 5})}{T}, \\\\ \\widehat {q} _ {1 2} ^ {\\mathrm {a}} = \\frac {M _ {2} + M _ {5} + M _ {7} + M _ {9} (\\widehat {g} _ {1 0} + \\widehat {g} _ {1 1} + \\widehat {g} _ {1 2} + \\widehat {g} _ {1 4} + \\widehat {g} _ {1 5})}{T}, \\\\ \\widehat {q} _ {2 1} ^ {\\mathrm {a}} = \\frac {M _ {3} + M _ {6} + M _ {8} + M _ {9} (\\widehat {g} _ {1 0} + \\widehat {g} _ {1 2} + \\widehat {g} _ {1 3} + \\widehat {g} _ {1 4} + \\widehat {g} _ {1 5})}{T}, \\\\ \\hat {q} _ {2 2} ^ {\\mathrm {a}} = \\frac {M _ {4} + M _ {7} + M _ {8} + M _ {9} (\\widehat {g} _ {9} + \\widehat {g} _ {1 1} + \\widehat {g} _ {1 3} + \\widehat {g} _ {1 4} + \\widehat {g} _ {1 5})}{T} \\tag {15} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 340, + 500, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_1, \\ldots, M_8$ are the numbers of unambiguous readout frames corresponding to events $E_1, \\ldots, E_8$ .", + "bbox": [ + 89, + 479, + 482, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While computing the bias of the ME analytically is complex, numerical simulations show that it has a low bias. Furthermore, since the ME depends on the SPE for an initial estimate of the detection probabilities, when the SPE is poor due to few measured single-photon frames, the reduction in MSE from the ME is relatively lesser than in cases when the SPE has low variance.", + "bbox": [ + 89, + 508, + 483, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Scaling to Higher Dimensional Arrays", + "text_level": 1, + "bbox": [ + 89, + 622, + 419, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In a $2 \\times 2$ array, there is a single type of ambiguous readout which indicates photon incidences at both rows and both columns of the array. As the spatial dimension of the array increases, there are more types of ambiguous readouts. For instance, with a $3 \\times 3$ array, we can have the following ambiguous readouts:", + "bbox": [ + 89, + 643, + 482, + 733 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- 2 rows and 2 columns fired - 4 candidate pixel locations;", + "- 2 rows and 3 columns fired - 6 candidate pixel locations;", + "- 3 rows and 2 columns fired - 6 candidate pixel locations;", + "- 3 rows and 3 columns fired -9 candidate pixel locations. Resolving these additional ambiguous readouts requires the modification of the expressions for conditional probabilities $g$ . As the number of rows and columns with photon coincidences increases, the number of unique terms in the expressions for $g$ grows rapidly. Hence, we restrict our analysis and simulations to at most 4-photon coincidences and ignore terms resulting from 5-photon coincidences and above." + ], + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We expect that this assumption introduces a small bias in the ME, especially at higher incident fluxes. Nevertheless, the improvement over handling only single-photon frames can be substantial.", + "bbox": [ + 511, + 90, + 905, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 513, + 167, + 601, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Imaging Arbitrary Scenes", + "text_level": 1, + "bbox": [ + 511, + 193, + 750, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3 shows the Monte Carlo simulation results of reconstructing $32 \\times 32$ images from 100,000 measured frames using each of the estimators. Note that these are estimates of the flux $\\hat{\\Lambda}_{ij}$ at each pixel, computed using the detection probability estimate $\\hat{q}_{ij}$ from the three estimators followed by the transformation in (4). Reconstructed in rows (a), (b), (c), and (d) are at a scaling of 3 mean photons per frame (PPF), while row (e) is at a scaling of 4 mean PPF. Mean PPF is the average number of detected photons in each measured frame. Each experiment is repeated 100 times for the computation of PSNR.", + "bbox": [ + 511, + 215, + 905, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In rows (a), (b), and (c), it can be seen that the NE introduces horizontal and vertical streaks due to misattributions from multiphoton coincidences. For example, in row (a) the SPE achieves $\\sim 5$ dB PSNR improvement over the NE and does not show the same artifacts, although the petals of the flower appear noisy.", + "bbox": [ + 511, + 382, + 905, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed ME outperforms the naive estimator by $\\sim 11\\mathrm{dB}$ while its improvement over the single-photon estimator is $\\sim 6\\mathrm{dB}$ . Further, we note that the ME preserves features in the ground truth image better than both of the other estimators, in addition to eliminating misattribution artifacts. The petals of the flower in row (a) are more clearly visible, while these are missing in the other reconstructions. The improvement in the reconstruction depends on some features of the ground truth image. For example, in row (d) we consider a rotated version of the image in row (a). It can be seen that the misattributions in the NE are more concentrated towards the center-left of the image as compared to row (a) where the horizontal and vertical streaks cut across the image. The PSNR of the naive reconstruction is lower than that in row (a), however, the multiphoton reconstruction still achieves $\\sim 5\\mathrm{dB}$ improvement. In row (e) we consider the image from row (b) with a PPF scaling of 4. We note that the PSNR of the NE reduces by $\\sim 0.6$ dB while the PSNR of the SPE and ME reduce by $\\sim 3$ dB each. This behavior is expected as an increase in the average PPF causes the number of frames with a single detected photon to decrease, which reduces the PSNR of the SPE and consequently the ME. However, the performance of the ME can be improved by considering additional multiphoton events.", + "bbox": [ + 511, + 474, + 906, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Optimal Photons Per Frame", + "text_level": 1, + "bbox": [ + 511, + 847, + 767, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In single-photon imaging, high signal strength is often beneficial for parameter estimation. However, for readout", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11410", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0854ebe751fc7ecad720bc03f241ecdb7133a1020552c100d4a7b1d284e4c97d.jpg", + "image_caption": [ + "(a)", + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 112, + 114, + 187, + 170 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4cb4c05d0831fd24ffe7b79b7bf39bdb983a53fc3b874b821c1a8ddc4ee36fbe.jpg", + "image_caption": [ + "Naive", + "16.39 dB" + ], + "image_footnote": [], + "bbox": [ + 205, + 114, + 279, + 170 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a109cc8f01f6d863c5c459a35bf5c37e836c4604cd807d19fcb114d4a9707773.jpg", + "image_caption": [ + "17.05 dB" + ], + "image_footnote": [], + "bbox": [ + 207, + 186, + 279, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d2fb0d1ccf9d1b77694f31e870c4d679ddba6ce9a7349b645e853b6a326be5af.jpg", + "image_caption": [ + "13.17 dB" + ], + "image_footnote": [], + "bbox": [ + 207, + 258, + 279, + 314 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/524604aa7eb71ab13b7c2cad14cc99a59791853cfc1370e1e7e3831fd50c5c59.jpg", + "image_caption": [ + "16.15 dB" + ], + "image_footnote": [], + "bbox": [ + 207, + 329, + 279, + 385 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aee72b6f18da28f8708d513436647c4a478c143cc03330ded32191f5c5c73fc0.jpg", + "image_caption": [ + "16.44 dB", + "(e)", + "Figure 3. Reconstruction of arbitrary $32 \\times 32$ images from 100 000 measured frames using the naive, single, and multiphoton estimators. Each PSNR value is from averaging over 100 trials. Rows (a)-(c) show that the multiphoton estimator is 4 to 6 dB better than the single-photon estimator and around 6 to 11 dB better than the naive estimator. Row (d) shows the impact of rotating the ground truth image on the reconstruction. Row (e) shows the impact of increasing the average number of photons per frame from 3 to 4. The SPE rejects $\\sim 85\\%$ of measured frames at a PPF of 3." + ], + "image_footnote": [], + "bbox": [ + 207, + 401, + 279, + 457 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0945a320a4448bf3e9b86e026f7201c8cf477356d8d58644c5b43fb1a234bcd.jpg", + "image_caption": [ + "Single-Photon" + ], + "image_footnote": [], + "bbox": [ + 297, + 114, + 370, + 170 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b581499b19afb24926b662f06da2ecf86795de9e932322970d6c0fabb929a862.jpg", + "image_caption": [ + "21.55 dB", + "22.43 dB" + ], + "image_footnote": [], + "bbox": [ + 299, + 186, + 370, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e38a06ce716abb9928099add0c336cd3d3d1a5c710d3b9fa41083b48619531f8.jpg", + "image_caption": [ + "16.63 dB" + ], + "image_footnote": [], + "bbox": [ + 299, + 258, + 370, + 314 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/13fe92ea23853e0cdcf7125dc8e641d83e61ce2e61d6bf02097d74b56f957f43.jpg", + "image_caption": [ + "16.60 dB" + ], + "image_footnote": [], + "bbox": [ + 297, + 329, + 370, + 385 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3685f1a69eb6a9aff78ae6421d09b6d02b44daac875fc6dcc6d16081ca2d80c3.jpg", + "image_caption": [ + "19.59 dB" + ], + "image_footnote": [], + "bbox": [ + 297, + 401, + 370, + 457 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/52f61186c13fb31a4dbc1e1e70d5a8ee66880619c966d46d9491172ddc8d846f.jpg", + "image_caption": [ + "Multiphoton" + ], + "image_footnote": [], + "bbox": [ + 388, + 114, + 462, + 170 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/900b2d151395b3d981d6078d37968f6287fcea5dbd29746f6d0859ba4c11c879.jpg", + "image_caption": [ + "27.22 dB", + "28.41 dB" + ], + "image_footnote": [], + "bbox": [ + 390, + 186, + 462, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0cd7e4665127c5cefd2b44478cd620961d26a680e0c2382f33afc5589fe7bfc6.jpg", + "image_caption": [ + "22.54 dB" + ], + "image_footnote": [], + "bbox": [ + 390, + 258, + 462, + 314 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0a0e13c9e88ce93373e11745863a9f0b481daeea4b9394bf5d1332740573f451.jpg", + "image_caption": [ + "21.79 dB" + ], + "image_footnote": [], + "bbox": [ + 390, + 329, + 462, + 385 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b0fc083ccf464eb0bab494a1756cbe1191a4c4cad3edb2142172afcd7cd5aa30.jpg", + "image_caption": [ + "25.98 dB" + ], + "image_footnote": [], + "bbox": [ + 390, + 401, + 462, + 457 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "multiplexed arrays, high flux also leads to more ambiguous readout frames, which makes image reconstruction challenging. We therefore expect that there exists an optimal incident photon flux level where each of the estimators achieves the minimum mean-square error. The following results, as shown in Fig. 4, validate this hypothesis.", + "bbox": [ + 89, + 638, + 482, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A ground truth image can be scaled in experimental settings by selecting a desired integration time or attenuating the incident photon flux using a neutral density filter. To study the dependence of estimator performance on the incident photon flux, we perform a simulation with a high emitted photon flux from a source that is attenuated before being incident on the array. We then study the change in mean-squared error per pixel $(\\mathrm{MSE} = \\| \\widehat{\\Lambda} - \\Lambda \\|_2^2 / n^2)$ of the reconstructed image with varying strength of the attenuating filter. The results of this simulation are shown in Fig. 4 for the reconstruction of a $32 \\times 32$ ground truth image in", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/532c68806317831e870b561a4699782325bb469862effa92b5e38aea42f74030.jpg", + "image_caption": [ + "Figure 4. Mean-squared errors of the estimators studied as functions of mean photons per frame. (Ground truth is image in Fig. 3(c).)" + ], + "image_footnote": [], + "bbox": [ + 519, + 104, + 885, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "row (c) of Fig. 3.", + "text_level": 1, + "bbox": [ + 511, + 411, + 629, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that as the mean PPF increases, the MSEs of all the estimators initially decrease, reach a minimum, and then increase. High MSE at low PPF values is due to the lack of detected photons for accurate reconstruction. The increase in MSE at high PPF values is expected due to the high variance of SPE and the high probability of multiphoton events not modeled by the ME. It can also be seen that the minimum MSE of the multiphoton estimator is lower than that of the single-photon estimator, and both of these are much lower than the naive estimator. Furthermore, the minimum value of MSE for the ME occurs at a PPF of 1.4, which is higher than that for the SPE at 0.83, and that for NE at 0.45. We expect these incidence flux values to be realistic in conditions similar to imaging under moonlight, where high-efficiency detectors like SNSPDs are typically used. Thus, the ME produces a better reconstruction compared to both the NE and the SPE when multiple photons are detected simultaneously at the array. This results in an increased photon count rate and a better utilization of the spatial dimensions of the array.", + "bbox": [ + 511, + 428, + 906, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. 2-, 3- and 4-Photon Estimators", + "text_level": 1, + "bbox": [ + 511, + 741, + 784, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As explained in Sec. 4.4, our estimator resolves up to 4 simultaneously incident photons. The conditional probability expressions (g's) contain terms corresponding to every possible photon event that can result in the measured readout. However, we can simplify these expressions by restricting the number of modeled events. For instance, if we include only 2-photon coincidences in the case of a $2 \\times 2$ array, only conditional probabilities of events $E_{9}$ and $E_{10}$ would be summed in (13). Fig. 4 shows the performance of the ME", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11411", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df831896f988a54927e707630b3dc510a5f5a3d08e013165a1b7c5ae8da31c23.jpg", + "image_caption": [ + "Naive", + "22.59 dB" + ], + "image_footnote": [], + "bbox": [ + 187, + 122, + 287, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/90a7dcbeb8831e0466399d52e5fdb40c87bf4774d5d2ead4e860d9340e06ba45.jpg", + "image_caption": [ + "Single-Photon", + "23.64 dB" + ], + "image_footnote": [], + "bbox": [ + 318, + 122, + 418, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85129537de147b3bad9295e0a352d1c5c7ec72326b5e0c6e92cb7984a733408a.jpg", + "image_caption": [ + "2-photon", + "23.83 dB", + "Figure 5. Image reconstruction by operating each of the studied estimators at mean PPF values corresponding to their lowest MSE's. Indicated values are PSNR of reconstructions." + ], + "image_footnote": [], + "bbox": [ + 93, + 257, + 194, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc699e795f461035376998688a59582f49fdf3c763c7537ad8ec5598bbb09635.jpg", + "image_caption": [ + "3-photon", + "24.60 dB" + ], + "image_footnote": [], + "bbox": [ + 225, + 257, + 330, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2d585d3963615f4cc791b4628d550342dd3927fcd4a6fb3028853de30c8b71f.jpg", + "image_caption": [ + "4-photon", + "26.74 dB" + ], + "image_footnote": [], + "bbox": [ + 362, + 257, + 467, + 337 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "when 2-, 3- and 4-photon are modeled in $g$ . It can be seen that the MSE of reconstruction is successively reduced with each additional photon considered in the estimator. The 2-photon estimator achieves its lowest MSE at a mean PPF of 1, while the 3- and 4-photon estimators achieve their lowest MSEs at 1.2 and 1.4 mean PPF, respectively.", + "bbox": [ + 89, + 455, + 483, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image reconstruction using each of the estimators in Fig. 4 at their optimal mean PPF values is summarized in Fig. 5. We see that the 4-photon estimator achieves the highest PSNR, which is 3 to 4 dB higher than the NE and the SPE.", + "bbox": [ + 89, + 553, + 483, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Reduction in Total Integration Time", + "text_level": 1, + "bbox": [ + 89, + 659, + 405, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since the multiphoton estimator enables the accurate interpretation of photon coincidences, it can be leveraged to reduce integration times in imaging with single-photon detectors. The performance of the three estimators in achieving a target reconstruction MSE is shown in Fig. 6. It can be seen that to achieve an MSE of 0.01, the ME requires 25,000 measured frames, and the SPE requires 100,000 measured frames, while the NE does not achieve this MSE within the range of frame values studied. This shows that the ME achieves a factor $\\sim 4$ reduction in the integration time for the same reconstruction quality as the other estimators. In comparison, a fully-addressed readout scheme achieves the target MSE with only 5000 frames, but with the increased complexity of readout.", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d1906ced67ab22223014d577a43c985cfd0395d033c73bb214d5b1ace7985df.jpg", + "image_caption": [ + "Figure 6. Reduction in the required number of frames to achieve the same target MSE by the NE, SPE,ME, and a fully-addressed readout scheme." + ], + "image_footnote": [], + "bbox": [ + 517, + 102, + 897, + 332 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Cramér-Rao Bound", + "text_level": 1, + "bbox": [ + 511, + 411, + 705, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Cramér-Rao bound (CRB) provides the lowest possible variance of an unbiased estimator for parameter estimation. Here, we compare the performance of the NE, SPE, 2-photon, 3-photon, and 4-photon estimators on the estimation of a $2 \\times 2$ ground truth, against CRB values at different incident photon fluxes. We first compute the Fisher information matrix (FIM) using the likelihood expression for row-column readouts from a $2 \\times 2$ array. The complete derivation of expressions for elements of the FIM are shown in the supplement. The element $(1,1)$ of the FIM for a single measurement is", + "bbox": [ + 511, + 433, + 906, + 601 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {I} _ {1 1} = \\frac {(1 - q _ {4}) (1 - q _ {2} q _ {3})}{q _ {1}} + \\frac {(1 - q _ {2} q _ {3})}{1 - q _ {1}} \\\\ + \\frac {q _ {4} ^ {2} \\left(1 - q _ {2} q _ {3}\\right) ^ {2}}{q _ {1} q _ {4} + q _ {2} q _ {3} - q _ {1} q _ {2} q _ {3} q _ {4}}. \\tag {16} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 625, + 903, + 693 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Other diagonal elements have similar forms as shown in the supplement. The CRB is then computed as the mean of the diagonal elements in $\\mathcal{I}^{-1}$ . The variation of CRB values, as well as MSEs of the studied estimators as a function of incident photon flux, is shown in Fig. 7.", + "bbox": [ + 511, + 703, + 905, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It can be seen that the ME curve matches the CRB curve for a range of mean PPF values. At high fluxes, the MSE of the ME increases due to the high variance of the single-photon estimates, which are used to calculate $g$ in (13). In the general case with bright spots at all four pixels of the array, the 4-photon estimator most closely matches the CRB across the range of PPF values studied. The 2- and 3-photon estimators deviate from the CRB due to the modeling of", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11412", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9adc2f33e01880b126796b9a8a0089858f74bddfcd55e80e8c50eaf6cd623f9e.jpg", + "image_caption": [ + "Figure 7. Comparison of estimator performance with the Cramér-Rao Bound. The ME matches the CRB for a range of incident photon fluxes and deviates from it at high fluxes. With 4 bright spots, model mismatch causes 2- and 3-photon estimators to become biased and deviate from the CRB at lower incident photon fluxes compared to the 4-photon estimator." + ], + "image_footnote": [], + "bbox": [ + 101, + 95, + 450, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "only a subset of events that can cause ambiguities and omission of terms in the expression for conditional probabilities in (13). The dependence of the estimator performance on the ground truth image is shown in the Supplement.", + "bbox": [ + 89, + 458, + 485, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Increasing array resolution", + "text_level": 1, + "bbox": [ + 89, + 547, + 336, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our analysis of the multiphoton estimator's performance so far has been limited to $32 \\times 32$ arrays to match the sensor developed in [37]. Here, we study its use in arrays with higher pixel counts and finer spatial resolutions. We emulate a fixed sensor area with a varying number of pixels and fixed maximum photon flux. This results in photons per pixel being inversely proportional to the number of pixels. We include attenuation optimization for each reconstruction method, similar to Fig. 4.", + "bbox": [ + 89, + 575, + 483, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results for row 1 of Fig. 3 as the ground truth are shown in Fig. 8 Performance improvement from the ME increases with array size (see Fig. 8(a)). This shows that our method is applicable to array sizes encountered in commercial imaging use cases. Nevertheless, we expect coincidences not modeled to be a limiting factor. Further, computation time increases linearly with an increase in array size. Thus, for a practical implementation of our algorithm, it could be beneficial to divide the sensor area into smaller blocks, the readouts from each of which can be processed using the ME. Such a processing approach would be particularly beneficial in sensor architectures similar to [17].", + "bbox": [ + 89, + 719, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/985c34be740c8b7943127fc80199ba14d2221c3d731d53f413155aff16ffd5cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 99, + 870, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41357b09a5d95f2bb815f4f53d986396596d07a0ed67d0c6a502b2263b31e885.jpg", + "image_caption": [ + "Figure 8. Influence of array size. (a) MSEs at optimum attenuation for each estimator. (b) Optimum photons per frame for each estimator." + ], + "image_footnote": [], + "bbox": [ + 537, + 330, + 870, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions and Future Works", + "text_level": 1, + "bbox": [ + 511, + 652, + 799, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We formulate multiphoton coincidence resolution in readout multiplexed single-photon detector arrays as an inverse imaging problem and propose a novel estimator to resolve up to four photon coincidences with row-column multiplexed readouts. Our method enables high-flux, low-latency, high-resolution image reconstructions that significantly improve upon conventional methods. Future work can extend this method to resolve an abitrarily large number of coincident photon events or other types of multiplexed readouts. The use of spatial priors and deep learning methods can be explored to improve the presented solutions. This line of work could be especially useful for the use of SNSPD arrays in photon-starved imaging applications like deep space imaging or biological imaging.", + "bbox": [ + 511, + 688, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11413", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the U.S. National Science Foundation under Grant 2039762 and in part by a gift from Dr. John Zheng Sun. We thank Prof. Hoover Rueda-Chacon for helpful discussions.", + "bbox": [ + 89, + 114, + 485, + 185 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 202, + 187, + 217 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] M. S. Allman, V. B. Verma, M. Stevens, T. Gerrits, R. D. Horansky, A. E. Lita, F. Marsili, A. Beyer, M. D. Shaw, D. Kumor, R. Mirin, and S. W. Nam. A near-infrared 64-pixel superconducting nanowire single photon detector array with integrated multiplexed readout. Applied Physics Letters, 106 (19), 2015. 1", + "[2] Ibrahim Ben Daya. Systematic characterization framework of row-column ultrasound imaging systems. Master's thesis, University of Waterloo, 2015. 2", + "[3] PL Bonanno. Physical and cost comparison of smartphone laser autofocus solutions. In Proc. Int. SPAD Sensor Workshop, 2024. 1", + "[4] Daniel AB Bonifacio and Maurizio Moralles. Modeling of 3d gamma interaction position in a monolithic scintillator block with a row-column summing readout. In 2012 IEEE Nuclear Science Symposium and Medical Imaging Conference Record (NSS/MIC), pages 2606-2613. IEEE, 2012. 2", + "[5] Jihwan Boo, Mark D Hammig, and Manhee Jeong. Row-column readout method to mitigate radiographic-image blurring from multipixel events in a coded-aperture imaging system. IEEE Transactions on Nuclear Science, 68(5):1175-1183, 2021. 2", + "[6] Caroline K Brennan, Zi Yao, Anastasia A Ionkina, Colin M Rathbun, Buvaneshwari Sathishkumar, and Jennifer A Prescher. Multiplexed bioluminescence imaging with a substrate unmixing platform. Cell Chemical Biology, 29(11): 1649-1660, 2022. 2", + "[7] Claudio Bruschini, Harald Homulle, Ivan Michel Antolovic, Samuel Burri, and Edoardo Charbon. Single-photon avalanche diode imagers in biophotonics: review and outlook. Light: Science & Applications, 8(1):87, 2019. 1", + "[8] Nicolas Callens and Georges GE Gielen. Analysis and comparison of readout architectures and analog-to-digital converters for 3d-stacked cmos image sensors. IEEE Transactions on Circuits and Systems I: Regular Papers, 68(8):3117-3130, 2021. 2", + "[9] Francesco Ceccarelli, Giulia Acconcia, Angelo Gulinatti, Massimo Ghioni, Ivan Rech, and Roberto Osellame. Recent advances and future perspectives of single-photon avalanche diodes for quantum photonics applications. Advanced Quantum Technologies, 4(2):2000102, 2021. 1", + "[10] Garry Chinn, Peter D. Olcott, and Craig S. Levin. Sparse signal recovery methods for multiplexing pet detector readout. IEEE Transactions on Medical Imaging, 32(5):932-942, 2013. 2", + "[11] Iris Cusini, Davide Berretta, Enrico Conca, Alfonso Incoronato, Francesca Madonini, Arianna Adelaide Maurina," + ], + "bbox": [ + 93, + 226, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chiara Nonne, Simone Riccardo, and Federica Villa. Historical perspectives, state of art and research trends of SPAD arrays and their applications (Part II: SPAD arrays). Frontiers in Physics, 10:906671, 2022. 1", + "[12] Iman Esmaeil Zadeh, J Chang, Johannes WN Los, Samuel Gyger, Ali W Elshaari, Stephan Steinhauer, Sander N Dorenbos, and Val Zwiller. Superconducting nanowire single-photon detectors: A perspective on evolution, state-of-the-art, future developments, and applications. Applied Physics Letters, 118(19), 2021. 1", + "[13] Andrew L Goertzen, Xuezhu Zhang, Megan M McClarty, Eric J Berg, Chen-Yi Liu, Piotr Kozlowski, Fabrice Retiere, Lawrence Ryner, Vesna Sossi, Greg Stortz, et al. Design and performance of a resistor multiplexing readout circuit for a sipm detector. IEEE Transactions on Nuclear Science, 60 (3):1541-1549, 2013. 2", + "[14] Rémi Gribonval, Antoine Chatalic, Nicolas Keriven, Vincent Schellekens, Laurent Jacques, and Philip Schniter. Sketching data sets for large-scale learning. IEEE Signal Process. Mag., 38(5):12-36, 2021. 1", + "[15] Yanqiu Guan, Haochen Li, Li Xue, Rui Yin, Labao Zhang, Hao Wang, Guanghao Zhu, Lin Kang, Jian Chen, and Peiheng Wu. Lidar with superconducting nanowire single-photon detectors: Recent advances and developments. Optics and Lasers in Engineering, 156:107102, 2022. 1", + "[16] Yanqiu Guan, Haochen Li, Labao Zhang, Hao Wang, Guanglong He, Biao Zhang, Yue Fei, Jiayu Lv, Xiao Zhang, Rui Yin, et al. SNSPD array with single-channel readout based on compressive sensing. ACS Photonics, 9(9):3102-3109, 2022. 2", + "[17] Hao Hao, Qing-Yuan Zhao, Yang-Hui Huang, Jie Deng, Fan Yang, Sai-Ying Ru, Zhen Liu, Chao Wan, Hao Liu, Zhi-Jian Li, et al. A compact multi-pixel superconducting nanowire single-photon detector array supporting gigabit space-to-ground communications. Light: Science & Applications, 13(1):25, 2024. 8", + "[18] Sam W. Hutchings, Nick Johnston, Istvan Gyongy, Tarek Al Abbas, Neale A. W. Dutton, Max Tyler, Susan Chan, Jonathan Leach, and Robert K. Henderson. A reconfigurable 3-D-stacked SPAD imager with in-pixel histogramming for flash LIDAR or high-speed time-of-flight imaging. IEEE J. Solid-State Circuits, 54(11):2947-2956, 2019. 1", + "[19] Atul Ingle and David Maier. Count-free single-photon 3d imaging with race logic. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-12, 2023. 1", + "[20] Jørgen Arendt Jensen, Mikkel Schou, Lasse Thurmann Jørgensen, Borislav G Tomov, Matthias Bo Stuart, Marie Sand Traberg, Iman Taghavi, Sigrid Huesebø Øygaard, Martin Lind Ommen, Kitty Steenberg, et al. Anatomic and functional imaging using row-column arrays. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 69(10):2722-2738, 2022. 2", + "[21] Ling-Dong Kong, Hui Wang, Qing-Yuan Zhao, Jia-Wei Guo, Yang-Hui Huang, Hao Hao, Shi Chen, Xue-Cou Tu, LaBao Zhang, Xiao-Qing Jia, et al. Readout-efficient superconducting nanowire single-photon imager with orthogonal time-amplitude multiplexing by hotspot quantization. Nature Photonics, 17(1):65-72, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11414", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Sangwon Lee, Yong Choi, Jihoon Kang, and Jin Ho Jung. Development of a multiplexed readout with high position resolution for positron emission tomography. *Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment*, 850:42-47, 2017. 2", + "[23] Feiyan Li, Hang Han, Qi Chen, Biao Zhang, Han Bao, Yue Dai, Rui Ge, Shuya Guo, Guanglong He, Yue Fei, et al. Saturation efficiency for detecting $1550\\mathrm{nm}$ photons with a $2\\times 2$ array of $\\mathrm{Mo}_{0.8}\\mathrm{Si}_{0.2}$ nanowires at $2.2\\mathrm{K}$ . Photonics Research, 9(3):389-394, 2021. 3", + "[24] Adam N McCaughan, Yao Zhai, Boris Korzh, Jason P Allmaras, Bakhrom G Oripov, Matthew D Shaw, and Sae Woo Nam. The thermally coupled imager: A scalable readout architecture for superconducting nanowire single photon detectors. Applied Physics Letters, 121(10), 2022. 1, 2", + "[25] Kazuhiro Morimoto. Megapixel spad cameras for time-resolved applications. Technical report, EPFL, 2021. 1", + "[26] Peter D Olcott, Garry Chinn, and Craig S Levin. Compressed sensing for the multiplexing of pet detectors. In 2011 IEEE Nuclear Science Symposium Conference Record, pages 3224-3226. IEEE, 2011. 2", + "[27] B. G. Oripov, D. S. Rampini, J. Allmaras, M. D. Shaw, S. W. Nam, B. Korzh, and A. N. McCaughan. A superconducting nanowire single-photon camera with 400,000 pixels. Nature, 622(7984):730-734, 2023. 2", + "[28] François Piron, Daniel Morrison, Mehmet Rasit Yuce, and Jean-Michel Redoute. A review of single-photon avalanche diode time-of-flight imaging sensor arrays. IEEE Sensors Journal, 21(11):12654-12666, 2020. 1", + "[29] Joshua Rapp, Julian Tachella, Yoann Altmann, Stephen McLaughlin, and Vivek K Goyal. Advances in single-photon lidar for autonomous vehicles: Working principles, challenges, and recent advances. IEEE Signal Processing Magazine, 37(4):62-71, 2020. 1", + "[30] Morten Fischer Rasmussen and Jørgen Arendt Jensen. 3-d ultrasound imaging performance of a row-column addressed 2-d array transducer: A measurement study. In 2013 IEEE International Ultrasonics Symposium (IUS), pages 1460-1463. IEEE, 2013. 2", + "[31] Morten Fischer Rasmussen, Thomas Lehrmann Christiansen, Erik Vilain Thomsen, and Jørgen Arendt Jensen. 3-d imaging using row-column-addressed arrays with integrated apodization-part i: apodization design and line element beamforming. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 62(5):947-958, 2015. 2", + "[32] Kaustubh Sadekar, David Maier, and Atul Ingle. Single-photon 3d imaging with equi-depth photon histograms. In European Conference on Computer Vision, pages 381-398. Springer, 2024. 1", + "[33] Ewout van den Berg, Emmanuel Candes, Garry Chinn, Craig Levin, Peter Demetri Olcott, and Carlos Sing-Long. Single-photon sampling architecture for solid-state imaging sensors. Proceedings of the National Academy of Sciences, 110(30): E2752-E2761, 2013. 2", + "[34] Hailu Wang, Jiaxiang Guo, Jinshui Miao, Wenjin Luo, Yue Gu, Runzhang Xie, Fang Wang, Lili Zhang, Peng Wang, and" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Weida Hu. Emerging single-photon detectors based on low-dimensional materials. Small, 18(5):2103963, 2022. 1", + "[35] Hui Wang, Zhi-jian Li, Xue-Mei Hu, Hao Hao, Jia-wei Guo, Yang-hui Huang, Hao Liu, Chao Wan, Xue-cou Tu, Xiaoqing Jia, et al. Image distortion by ambiguous multiple-photon detections in a superconducting nanowire single-photon imager and the correction method. Optics Express, 31(14):23579-23588, 2023. 2", + "[36] Mel White, Shahaboddin Ghajari, Tianyi Zhang, Akshit Dave, Ashok Veeraraghavan, and Alyosha Molnar. A differential SPAD array architecture in $0.18\\mu \\mathrm{m}$ CMOS for HDR imaging. In Proc. IEEE Int. Symp. Circuits Syst., pages 292-296, 2022. 1", + "[37] Emma E Wollman, Varun B Verma, Adriana E Lita, William H Farr, Matthew D Shaw, Richard P Mirin, and Sae Woo Nam. Kilopixel array of superconducting nanowire single-photon detectors. Optics express, 27(24):35279-35289, 2019. 1, 2, 8", + "[38] Lixing You. Superconducting nanowire single-photon detectors for quantum information. Nanophotonics, 9(9):2673-2692, 2020. 1", + "[39] Qing-Yuan Zhao, Di Zhu, Niccolò Calandri, Andrew E Dane, Adam N McCaughan, Francesco Bellei, Hao-Zhu Wang, Daniel F Santavicca, and Karl K Berggren. Single-photon imager based on a superconducting nanowire delay line. Nature Photonics, 11(4):247-251, 2017. 2" + ], + "bbox": [ + 516, + 92, + 903, + 457 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11415", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/960040c5-ed98-43eb-9b66-fa83e44db35d_model.json b/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/960040c5-ed98-43eb-9b66-fa83e44db35d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..07f89522723fcabf8425548141b83629fd000a4e --- /dev/null +++ b/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/960040c5-ed98-43eb-9b66-fa83e44db35d_model.json @@ -0,0 +1,2541 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.908, + 0.153 + ], + "angle": 0, + "content": "Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.181, + 0.838, + 0.217 + ], + "angle": 0, + "content": "Shashwath Bharadwaj, Ruangrawee Kitichotkul, Akshay Agarwal and Vivek K Goyal Boston University" + }, + { + "type": "text", + "bbox": [ + 0.442, + 0.22, + 0.551, + 0.233 + ], + "angle": 0, + "content": "shash@bu.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.302, + 0.485, + 0.694 + ], + "angle": 0, + "content": "Readout multiplexing is a promising solution to overcome hardware limitations and data bottlenecks in imaging with single-photon detectors. Conventional multiplexed readout processing creates an upper bound on photon counts at a very fine time scale, where frames with multiple detected photons must either be discarded or allowed to introduce significant bias. We formulate multiphoton coincidence resolution as an inverse imaging problem and introduce a solution framework to probabilistically resolve the spatial locations of photon incidences. Specifically, we develop a theoretical abstraction of row-column multiplexing and a model of photon events that make readouts ambiguous. Using this, we propose a novel estimator that spatially resolves up to four coincident photons. Monte Carlo simulations show that our proposed method increases the peak signal-to-noise ratio (PSNR) of reconstruction by 3 to 4 dB compared to conventional methods under optimal incident flux conditions. Additionally, this method reduces the required number of readout frames to achieve the same mean-squared error as other methods by a factor of \\(\\sim 4\\). Finally, our method matches the Cramér-Rao bound for detection probability estimation for a wider range of incident flux values compared to conventional methods. While demonstrated for a specific detector type and readout architecture, this method can be extended to more general multiplexing with different detector models." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.724, + 0.222, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Single-photon detectors have been widely studied for use in applications like biological imaging [7, 34], lidar [15, 29], and quantum optics [9] due to properties such as single-photon sensitivity and high-precision time-tagging of photon arrivals. Single-photon avalanche diodes (SPADs) have been of particular interest due to compatibility with CMOS manufacturing techniques and room temperature operation [11]. However, they have low detection efficiencies (especially at mid infrared wavelengths), which means that only a small fraction of incident light is detected as pho" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.271, + 0.907, + 0.378 + ], + "angle": 0, + "content": "tons. Superconducting nanowire single-photon detectors (SNSPDs) are emerging devices that offer near-unity quantum efficiency across a range of wavelengths, low dark counts, and low jitter, making them compelling alternatives to SPADs [38]. However, the requirement of cryogenic cooling to maintain superconductivity has largely restricted their use to laboratory experiments." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.382, + 0.907, + 0.639 + ], + "angle": 0, + "content": "Single-photon imaging in typical commercial applications requires arrays of single-photon detectors [28]. While megapixel SPAD arrays have already been deployed in mass consumer products [3, 25], high data transfer rates from these arrays for further processing remain a bottleneck. For instance, typical room-scale ranging experiments with SPAD array-based single-photon lidar can have data generation rates of the order of gigabits per second when the array resolution approaches the megapixel scale. It is attractive to compress the readout to less than a stream of all digitized detection times. Methods like in-pixel histogramming [18], histogram compression through sketching [14], equidepth histogramming [19, 32], and differential readouts [36] have been proposed to address high data transfer rates. These approaches to avoid the cost of high temporal resolution are complementary to our emphasis on transverse spatial dimensions in this work." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Implementing high spatial resolution SNSPD arrays is a significant challenge. As the number of pixels in the array grows, the heat load introduced by reading out each pixel individually becomes incompatible with the cryogenics of the system [12]. To mitigate this, recent efforts have focused on addressing groups of pixels in the readout, starting with Allman et al. [1] who introduced row-column readout multiplexing. In this mechanism, readouts indicate photon detection at each row and each column of the array instead of every pixel, which reduces the required number of readout lines for an \\( n \\times n \\) array from \\( n^2 \\) to \\( 2n \\). Wollman et al. [37] extended this scheme to demonstrate the first kilopixel SNSPD array. McCaughan et al. [24] introduced the thermally coupled imager with time-of-flight multiplexing, where separate readout lines for each row and column are replaced with a single bus for all rows and all columns. This makes the required number of readouts independent of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "11406" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.481, + 0.152 + ], + "angle": 0, + "content": "the array size. Recently, Oripov et al. [27] leveraged this method to demonstrate a 400,000 pixel superconducting camera, showing the viability of this scheme to approach the sizes of commercial imaging arrays." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.153, + 0.482, + 0.468 + ], + "angle": 0, + "content": "Despite the gain in scaling array sizes, readout multiplexing requires the incident photon flux to be very low for unambiguous signal reconstruction, limiting the array's photon count rate. When two or more photons are incident on the array within the period of a single acquisition, row-column readouts can give an ambiguous set of candidate locations for their incidences. Traditional reconstruction techniques [37] do not distinguish between these locations and assume equal probabilities of detection at each candidate pixel. This results in the misattribution of photon counts to locations where no photon incidences occurred. Alternatively, since readouts with a single detected photon provide the exact row and column indices where the photon was incident on the array, ambiguous readouts with multiple detections can be discarded for signal reconstruction. However, this leads to high variance in the reconstruction (especially at high incident fluxes) and underutilization of the spatial dimensions of the array by requiring most readouts to contain detections at a single pixel. Resolving the spatial locations of multiphoton coincidences using multiplexed readouts is an ill-posed inverse problem." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.47, + 0.482, + 0.831 + ], + "angle": 0, + "content": "Here, we propose a solution framework to leverage most of the measured data to reduce the mean-squared error of image reconstruction. Specifically, we develop a multiphoton estimator that redistributes photon counts from ambiguous readouts to candidate pixels such that an approximate likelihood of observations is maximized. A key choice for our algorithm design is the omission of spatial priors in order to not obscure the source of performance gain. Spatial priors can be used in conjunction with the developed method for further improvements. Monte Carlo simulations show that our proposed method increases the peak signal-to-noise ratio (PSNR) of reconstruction by 3 to \\(4\\mathrm{dB}\\) compared to conventional methods under optimal incident flux conditions. Further, the optimal incident photon flux for our method is \\(\\sim 1.4\\) photons per readout which is 0.5 to 1.1 photons per readout higher than conventional methods. We also demonstrate that our estimator matches the Cramér-Rao Bound (CRB) for the estimation of photon detection probabilities at each pixel of an array across a range of incident photon flux values. Finally, our method achieves a factor of \\(\\sim 4\\) reduction in the required number of readout frames for the same mean-squared error of image reconstruction. These results demonstrate that our proposed estimator is well-suited for high-flux, low-latency imaging." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.846, + 0.231, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.873, + 0.481, + 0.885 + ], + "angle": 0, + "content": "Image reconstruction from multiplexed readouts." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.888, + 0.481, + 0.9 + ], + "angle": 0, + "content": "Compressed sensing methods have been explored for the" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.498 + ], + "angle": 0, + "content": "design of multiplexed readout architectures in areas like positron emission tomography (PET) [10, 13, 22, 26], biological imaging [6, 30] and gamma-ray imaging [4, 5]. While readout multiplexing for specific imaging applications has been extensively studied [8, 20], the focus of these works is on the modeling of optical and electrical properties of the imaging system such as the shape of the wavefront of the detected light [2, 31], coded apertures for rejecting multipixel events [5] and IC design for resistor multiplexing [13]. The explicit study of individual photon events measured by single-photon detectors and their impact on the multiplexed readouts does not yet exist. Our work introduces a theoretical study of the readouts based on the combinatorics of photon events and an approximate maximum likelihood estimation of photon detection probabilities at each pixel. We also analyze the performance of a novel multiphoton estimator with variations in imaging parameters. Our work is closest to the analysis presented by van den Berg et al. [33], where group-testing inspired surface codes are studied for deterministic multiphoton coincidence resolution in time-to-digital converter readouts from SPAD arrays. However, our work specifically addresses intensity imaging with SNSPD arrays where a small probability of error is tolerable. We provide a probabilistic model of the readouts and solutions for image reconstruction from a specific device architecture for up to four coincident photons." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.501, + 0.905, + 0.817 + ], + "angle": 0, + "content": "Imaging with SNSPD arrays. While large-scale SNSPD array development is still in the nascent stages, several works have conducted proof-of-principle imaging experiments with prototypes. Zhao et al. introduced a delay-line-based superconducting imager with an effective pixel count of \\(\\sim 590\\) [39]. This method inspired the thermally coupled imager [24], which in turn resulted in the 400,000-pixel camera by Oripov et al. [27]. Subsequent efforts to use SNSPD arrays for imaging [16, 21, 35] involve intricate hardware control such as bias-current sweeping, use of idle pixels, and geometric design of the nanowire structure. Our solution is purely computational and does not require additional hardware modifications. In addition, the modeling of incident photon events and their multiplexed readouts can be extended to several detector models in high-energy physics or particle detection. While the method described in this work provides a framework to resolve an arbitrarily large number of coincident photons, we limit the discussion in the next sections to up to four coincident photons to make the demonstrated solutions computationally tractable and easier to illustrate and analyze." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.832, + 0.822, + 0.845 + ], + "angle": 0, + "content": "3. Measurement and Readout Model" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.857, + 0.904, + 0.9 + ], + "angle": 0, + "content": "Figure 1 illustrates imaging a scene using an \\(n \\times n\\) SNSPD array with row-column readouts. Measurements are read out after integrating the incident signal for a fixed time in" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11407" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.106, + 0.913, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.347, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 1. Schematic of image reconstruction with row-column readouts. (Left) Multiple frames of a ground truth image are measured using an SNSPD array to give row-column readout vectors. (Right) Comparison of conventional and proposed reconstruction techniques to estimate the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.482, + 0.475 + ], + "angle": 0, + "content": "terval. Let \\(\\Lambda \\in \\mathbb{R}^{n\\times n}\\) be the flux of an \\(n\\times n\\) ground truth image in that integration time. The number of photons arriving at pixel \\((i,j)\\) at a discrete time index \\(t\\) is modeled as" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.476, + 0.483, + 0.495 + ], + "angle": 0, + "content": "\\[\nX _ {i j} ^ {t} \\sim \\operatorname {P o i s s o n} (\\Lambda_ {i j}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.484, + 0.606 + ], + "angle": 0, + "content": "We assume pixel saturation with the detection of a single photon, i.e., the detectors are not photon-number resolving [23]. Thus, we define binary-valued photon-incidence indicator \\( Y^{t} \\in \\{0,1\\}^{n \\times n} \\) by \\( Y_{ij}^{t} = \\mathbf{1}[X_{ij}^{t} > 0] \\), where \\( \\mathbf{1}[\\cdot] \\) is the indicator function. Then, \\( Y^{t} \\) is a matrix where each entry is an independent Bernoulli random variable \\( Y_{ij}^{t} \\sim \\mathrm{Bernoulli}(q_{ij}) \\), where" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.617, + 0.483, + 0.637 + ], + "angle": 0, + "content": "\\[\nq _ {i j} = 1 - e ^ {- \\Lambda_ {i j}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.431, + 0.66 + ], + "angle": 0, + "content": "is the probability of photon detection at pixel \\((i,j)\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.66, + 0.484, + 0.692 + ], + "angle": 0, + "content": "Each measurement consists of a row readout \\( R^t \\in \\{0,1\\}^n \\) and a column readout \\( C^t \\in \\{0,1\\}^n \\):" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.701, + 0.483, + 0.72 + ], + "angle": 0, + "content": "\\[\nR _ {i} ^ {t} = \\mathbf {1} \\left[ X _ {i j} ^ {t} > 0 \\text {f o r a n y} j \\in \\{1, \\dots , n \\} \\right], \\tag {3a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.722, + 0.483, + 0.741 + ], + "angle": 0, + "content": "\\[\nC _ {j} ^ {t} = \\mathbf {1} \\left[ X _ {i j} ^ {t} > 0 \\text {f o r a n y} i \\in \\{1, \\dots , n \\} \\right]. \\tag {3b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.485, + 0.905 + ], + "angle": 0, + "content": "We then define a readout frame at time \\( t \\) as \\( (R^t, C^t) \\). When multiple photons are incident on the array within a single integration period, the problem of resolving the spatial locations of their incidence is ill-posed. As an example, consider a \\( 2 \\times 2 \\) array with pixels numbered from 1 to 4 as shown in event \\( E_0 \\) of Fig. 2. The frame ([1,1],[1,1]) could have resulted from photons detected at pixels \\( \\{1,4\\} \\) or \\( \\{2,3\\} \\) or any of: \\( \\{1,2,3\\} \\), \\( \\{1,2,4\\} \\), \\( \\{1,3,4\\} \\), \\( \\{2,3,4\\} \\), or \\( \\{1,2,3,4\\} \\). More generally, a readout frame is unambiguous if \\( \\sum_{i=1}^{n} R_i^t = 1 \\) or \\( \\sum_{j=1}^{n} C_j^t = 1 \\); otherwise, it is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.907, + 0.476 + ], + "angle": 0, + "content": "ambiguous. Our problem is to estimate \\(\\Lambda\\) from measurements in \\(T\\) frames \\(\\{(R^t,C^t)\\}_{t = 1}^T\\). In particular, we are interested in exploiting the information contained in ambiguous readout frames." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.499, + 0.631, + 0.515 + ], + "angle": 0, + "content": "4. Estimators" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.603 + ], + "angle": 0, + "content": "In this section, we introduce the three estimators compared in this study, namely the naive, single-photon, and multiphoton estimators. Each is formulated first as an estimator \\(\\widehat{q}_{ij}\\) for the probability \\(q_{ij}\\) defined in (2). The incident flux at each pixel can then be estimated elementwise as" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.619, + 0.907, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\Lambda} _ {i j} = - \\log \\left(1 - e ^ {- \\widehat {q} _ {i j}}\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.676, + 0.695 + ], + "angle": 0, + "content": "4.1. Naive Estimator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.764 + ], + "angle": 0, + "content": "The naive estimator (NE) assumes that a photon is detected at each of the candidate pixels in ambiguous readout frames. Thus, the probability of detecting a photon at pixel \\((i,j)\\) is estimated as" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.77, + 0.907, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\widehat {q} _ {i j, \\mathrm {n a i v e}} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} R _ {i} ^ {t} C _ {j} ^ {t}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The naive estimator overestimates \\( q_{ij} \\) because it misattributes photon counts in ambiguous frames to pixels where no photon incidence occurred. The extra counts imputed by this scheme produce ghost spots in the reconstructed image, as depicted in Fig. 1." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11408" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.482, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.286, + 0.485, + 0.397 + ], + "angle": 0, + "content": "Figure 2. Row-column readout frames for a \\(2 \\times 2\\) array. Yellow spots indicate incident photon locations. Accounting for pixel saturation, 16 photon detection events are possible with this array. Row-column readout results in 9 unambiguous events \\(E_0, E_1, \\ldots, E_8\\) and 7 events \\(E_9, E_{10}, \\ldots, E_{15}\\) that are ambiguous because they all result in \\(R^t = [1,1]\\) and \\(C^t = [1,1]\\). Pixel numbers are indicated in event \\(E_0\\), which corresponds to an empty readout." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.423, + 0.368, + 0.437 + ], + "angle": 0, + "content": "The positive bias of the naive estimator is" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.447, + 0.484, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {E} \\left[ \\widehat {q} _ {i j, \\text {n a i v e}} \\right] - q _ {i j} \\\\ = \\left(1 - q _ {i j}\\right) \\left(1 - \\prod_ {k \\in R _ {i} ^ {\\prime}} \\left(1 - q _ {k j}\\right)\\right) \\left(1 - \\prod_ {\\ell \\in C _ {j} ^ {\\prime}} \\left(1 - q _ {i \\ell}\\right)\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.541, + 0.484, + 0.587 + ], + "angle": 0, + "content": "where \\( R_{i}^{\\prime} \\) is the set of rows except row \\( i \\) and \\( C_j^\\prime \\) is the set of columns except column \\( j \\). A derivation is given in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.595, + 0.317, + 0.611 + ], + "angle": 0, + "content": "4.2. Single-Photon Estimator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.618, + 0.484, + 0.693 + ], + "angle": 0, + "content": "A simple way to prevent misattributions and thus reduce bias is to only use frames with a single detected photon in the reconstruction since these unambiguously give the spatial locations of photon incidence. We thus define the single-photon estimator (SPE):" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.701, + 0.482, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\widehat {q} _ {i j, \\text {s i n g l e}} = \\left\\{ \\begin{array}{l l} \\frac {N _ {i j}}{N _ {i j} + N _ {0}}, & \\text {i f} N _ {i j} + N _ {0} > 0; \\\\ 0, & \\text {i f} N _ {i j} + N _ {0} = 0, \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.84 + ], + "angle": 0, + "content": "where \\(N_{ij}\\) is the number of frames where a photon is detected only at pixel \\((i,j)\\), and \\(N_0\\) is the number of frames without any detected photons across the whole array. The SPE is unbiased when \\(N_{i,j} + N_0 > 0\\), as shown in the supplement. However, the discarding of frames with multiple detected photons results in high variance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.306, + 0.865 + ], + "angle": 0, + "content": "4.3. Multiphoton Estimator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Now, we develop a multiphoton estimator (ME) that uses photon counts from ambiguous readouts for reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.195 + ], + "angle": 0, + "content": "For simplicity, here we derive expressions for detection probability estimates at each pixel of a \\(2 \\times 2\\) array and illustrate how this estimator maximizes an approximate likelihood of observations. The general form of the ME is presented in the supplement. If each pixel were read out individually, the likelihood of observations could be written as" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.195, + 0.905, + 0.236 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\prod_ {i = 1} ^ {2} \\prod_ {j = 1} ^ {2} (1 - q _ {i j}) ^ {v _ {i j}} \\left(q _ {i j}\\right) ^ {f _ {i j}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.241, + 0.906, + 0.305 + ], + "angle": 0, + "content": "where \\( v_{ij} \\) is the total number of frames with no detections at pixel \\( (i,j) \\) and \\( f_{ij} \\) is the total number of frames with a detection at pixel \\( (i,j) \\). Then, the maximum likelihood estimate \\( \\widehat{q}_{ij}^{\\mathrm{ML}} \\) could be computed as" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.315, + 0.905, + 0.345 + ], + "angle": 0, + "content": "\\[\n\\hat {q} _ {i j} ^ {\\mathrm {M L}} = \\frac {f _ {i j}}{T}, \\quad i = 1, 2, j = 1, 2, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.354, + 0.836, + 0.368 + ], + "angle": 0, + "content": "where \\(T\\) is the total number of measured frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.906, + 0.428 + ], + "angle": 0, + "content": "With row-column readouts, our knowledge of locations of photon detections is reduced. We define \\( v_{ij} \\) and \\( f_{ij} \\) from unambiguous frames only and find that the likelihood of observations becomes" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.44, + 0.905, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\left\\{\\left(R ^ {t}, C ^ {t}\\right) \\right\\} _ {t = 1} ^ {T}; q\\right) = U (q) A (q), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.47, + 0.615, + 0.483 + ], + "angle": 0, + "content": "with the factor" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.494, + 0.905, + 0.535 + ], + "angle": 0, + "content": "\\[\nU (q) = \\prod_ {i = 1} ^ {2} \\prod_ {j = 1} ^ {2} \\left(1 - q _ {i j}\\right) ^ {v _ {i j}} \\left(q _ {i j}\\right) ^ {f _ {i j}} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.547, + 0.785, + 0.561 + ], + "angle": 0, + "content": "from unambiguous frames and the factor" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.573, + 0.905, + 0.591 + ], + "angle": 0, + "content": "\\[\nA (q) = \\left(q _ {1 1} q _ {2 2} + q _ {1 2} q _ {2 1} - q _ {1 1} q _ {1 2} q _ {2 1} q _ {2 2}\\right) ^ {M _ {9}} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.906, + 0.647 + ], + "angle": 0, + "content": "from ambiguous frames. Here, \\(M_9\\) is the total number of ambiguous readout frames due to events \\(E_9,\\ldots ,E_{15}\\) shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.648, + 0.907, + 0.859 + ], + "angle": 0, + "content": "The factor in (12) results from the sum of probabilities \\(\\sum_{k=9}^{15} \\mathbb{P}(E_k)\\). This makes it challenging to find analytical expressions for the maximum likelihood estimates of detection probabilities. However, if we can estimate the fractions of the \\(M_9\\) frames resulting from each of the events \\(E_9, \\ldots, E_{15}\\), we can re-write (12) as a product of probabilities of events \\(E_9, \\ldots, E_{15}\\) to express (10) similarly to (8) and obtain analytical expressions for approximate maximum likelihood estimates \\(\\widehat{q}_{ij}^{\\mathrm{a}}\\). We first find expressions for the conditional probabilities \\(g_9, \\ldots, g_{15}\\) of events \\(E_9, \\ldots, E_{15}\\), given that an ambiguous frame was read out. For example, \\(g_9\\) is the conditional probability of event \\(E_9\\) given that a frame \\(R^t = [1,1]\\), \\(C^t = [1,1]\\) was measured and is given by" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.869, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\ng _ {9} = \\frac {\\mathbb {P} \\left(E _ {9}\\right)}{\\sum_ {k = 9} ^ {1 5} \\mathbb {P} \\left(E _ {k}\\right)}. \\tag {13}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11409" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "In addition to the inherent limitation of having a finite amount of data, computing \\(\\mathbb{P}(E_9),\\ldots ,\\mathbb{P}(E_{15})\\) is hindered by ambiguities. Thus, we estimate these probabilities using the measured single-photon frames, which are unambiguous and unbiased. For example, an estimate of \\(\\mathbb{P}(E_9)\\) can be calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.188, + 0.483, + 0.208 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathbb {P}} (E _ {9}) = \\widehat {q} _ {1 1} ^ {\\mathrm {s}} \\widehat {q} _ {2 2} ^ {\\mathrm {s}} (1 - \\widehat {q} _ {1 2} ^ {\\mathrm {s}}) (1 - \\widehat {q} _ {2 1} ^ {\\mathrm {s}}). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.485, + 0.336 + ], + "angle": 0, + "content": "Similar expressions can be derived for \\(\\widehat{\\mathbb{P}}(E_{10}),\\ldots,\\widehat{\\mathbb{P}}(E_{15})\\) which are then substituted in (13) in place of \\(\\mathbb{P}(E_9),\\dots ,\\mathbb{P}(E_{15})\\) to get estimates \\(\\widehat{g}_9,\\dots ,\\widehat{g}_{15}\\) of conditional probabilities \\(g_{9},\\ldots ,g_{15}\\). The fractions of the \\(M_9\\) frames resulting from events \\(E_{9},\\ldots E_{15}\\) can then be estimated as \\(\\widehat{g}_9M_9,\\ldots ,\\widehat{g}_{15}M_9\\). Finally, these estimates can be used to simplify (12) to a product of probabilities and obtain the approximate maximum likelihood estimates" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.342, + 0.5, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widehat {q} _ {1 1} ^ {\\mathrm {a}} = \\frac {M _ {1} + M _ {5} + M _ {6} + M _ {9} (\\widehat {g} _ {9} + \\widehat {g} _ {1 1} + \\widehat {g} _ {1 2} + \\widehat {g} _ {1 3} + \\widehat {g} _ {1 5})}{T}, \\\\ \\widehat {q} _ {1 2} ^ {\\mathrm {a}} = \\frac {M _ {2} + M _ {5} + M _ {7} + M _ {9} (\\widehat {g} _ {1 0} + \\widehat {g} _ {1 1} + \\widehat {g} _ {1 2} + \\widehat {g} _ {1 4} + \\widehat {g} _ {1 5})}{T}, \\\\ \\widehat {q} _ {2 1} ^ {\\mathrm {a}} = \\frac {M _ {3} + M _ {6} + M _ {8} + M _ {9} (\\widehat {g} _ {1 0} + \\widehat {g} _ {1 2} + \\widehat {g} _ {1 3} + \\widehat {g} _ {1 4} + \\widehat {g} _ {1 5})}{T}, \\\\ \\hat {q} _ {2 2} ^ {\\mathrm {a}} = \\frac {M _ {4} + M _ {7} + M _ {8} + M _ {9} (\\widehat {g} _ {9} + \\widehat {g} _ {1 1} + \\widehat {g} _ {1 3} + \\widehat {g} _ {1 4} + \\widehat {g} _ {1 5})}{T} \\tag {15} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.483, + 0.509 + ], + "angle": 0, + "content": "where \\(M_1, \\ldots, M_8\\) are the numbers of unambiguous readout frames corresponding to events \\(E_1, \\ldots, E_8\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.484, + 0.614 + ], + "angle": 0, + "content": "While computing the bias of the ME analytically is complex, numerical simulations show that it has a low bias. Furthermore, since the ME depends on the SPE for an initial estimate of the detection probabilities, when the SPE is poor due to few measured single-photon frames, the reduction in MSE from the ME is relatively lesser than in cases when the SPE has low variance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.623, + 0.421, + 0.639 + ], + "angle": 0, + "content": "4.4. Scaling to Higher Dimensional Arrays" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.734 + ], + "angle": 0, + "content": "In a \\(2 \\times 2\\) array, there is a single type of ambiguous readout which indicates photon incidences at both rows and both columns of the array. As the spatial dimension of the array increases, there are more types of ambiguous readouts. For instance, with a \\(3 \\times 3\\) array, we can have the following ambiguous readouts:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.735, + 0.483, + 0.75 + ], + "angle": 0, + "content": "- 2 rows and 2 columns fired - 4 candidate pixel locations;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.75, + 0.482, + 0.765 + ], + "angle": 0, + "content": "- 2 rows and 3 columns fired - 6 candidate pixel locations;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.766, + 0.482, + 0.78 + ], + "angle": 0, + "content": "- 3 rows and 2 columns fired - 6 candidate pixel locations;" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- 3 rows and 3 columns fired -9 candidate pixel locations. Resolving these additional ambiguous readouts requires the modification of the expressions for conditional probabilities \\( g \\). As the number of rows and columns with photon coincidences increases, the number of unique terms in the expressions for \\( g \\) grows rapidly. Hence, we restrict our analysis and simulations to at most 4-photon coincidences and ignore terms resulting from 5-photon coincidences and above." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.151 + ], + "angle": 0, + "content": "We expect that this assumption introduces a small bias in the ME, especially at higher incident fluxes. Nevertheless, the improvement over handling only single-photon frames can be substantial." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.168, + 0.602, + 0.184 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.194, + 0.75, + 0.21 + ], + "angle": 0, + "content": "5.1. Imaging Arbitrary Scenes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.216, + 0.906, + 0.381 + ], + "angle": 0, + "content": "Figure 3 shows the Monte Carlo simulation results of reconstructing \\(32 \\times 32\\) images from 100,000 measured frames using each of the estimators. Note that these are estimates of the flux \\(\\hat{\\Lambda}_{ij}\\) at each pixel, computed using the detection probability estimate \\(\\hat{q}_{ij}\\) from the three estimators followed by the transformation in (4). Reconstructed in rows (a), (b), (c), and (d) are at a scaling of 3 mean photons per frame (PPF), while row (e) is at a scaling of 4 mean PPF. Mean PPF is the average number of detected photons in each measured frame. Each experiment is repeated 100 times for the computation of PSNR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.906, + 0.474 + ], + "angle": 0, + "content": "In rows (a), (b), and (c), it can be seen that the NE introduces horizontal and vertical streaks due to misattributions from multiphoton coincidences. For example, in row (a) the SPE achieves \\(\\sim 5\\) dB PSNR improvement over the NE and does not show the same artifacts, although the petals of the flower appear noisy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.837 + ], + "angle": 0, + "content": "Our proposed ME outperforms the naive estimator by \\(\\sim 11\\mathrm{dB}\\) while its improvement over the single-photon estimator is \\(\\sim 6\\mathrm{dB}\\). Further, we note that the ME preserves features in the ground truth image better than both of the other estimators, in addition to eliminating misattribution artifacts. The petals of the flower in row (a) are more clearly visible, while these are missing in the other reconstructions. The improvement in the reconstruction depends on some features of the ground truth image. For example, in row (d) we consider a rotated version of the image in row (a). It can be seen that the misattributions in the NE are more concentrated towards the center-left of the image as compared to row (a) where the horizontal and vertical streaks cut across the image. The PSNR of the naive reconstruction is lower than that in row (a), however, the multiphoton reconstruction still achieves \\(\\sim 5\\mathrm{dB}\\) improvement. In row (e) we consider the image from row (b) with a PPF scaling of 4. We note that the PSNR of the NE reduces by \\(\\sim 0.6\\) dB while the PSNR of the SPE and ME reduce by \\(\\sim 3\\) dB each. This behavior is expected as an increase in the average PPF causes the number of frames with a single detected photon to decrease, which reduces the PSNR of the SPE and consequently the ME. However, the performance of the ME can be improved by considering additional multiphoton events." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.768, + 0.864 + ], + "angle": 0, + "content": "5.2. Optimal Photons Per Frame" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In single-photon imaging, high signal strength is often beneficial for parameter estimation. However, for readout" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11410" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.137, + 0.109, + 0.15 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.101, + 0.191, + 0.113 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.115, + 0.188, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.102, + 0.259, + 0.113 + ], + "angle": 0, + "content": "Naive" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.115, + 0.28, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.174, + 0.272, + 0.185 + ], + "angle": 0, + "content": "16.39 dB" + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.187, + 0.28, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.244, + 0.272, + 0.255 + ], + "angle": 0, + "content": "17.05 dB" + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.259, + 0.28, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.316, + 0.272, + 0.326 + ], + "angle": 0, + "content": "13.17 dB" + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.33, + 0.28, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.387, + 0.272, + 0.397 + ], + "angle": 0, + "content": "16.15 dB" + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.402, + 0.28, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.459, + 0.272, + 0.47 + ], + "angle": 0, + "content": "16.44 dB" + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.101, + 0.378, + 0.113 + ], + "angle": 0, + "content": "Single-Photon" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.115, + 0.372, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.175, + 0.362, + 0.185 + ], + "angle": 0, + "content": "21.55 dB" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.187, + 0.372, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.244, + 0.362, + 0.255 + ], + "angle": 0, + "content": "22.43 dB" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.259, + 0.372, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.316, + 0.363, + 0.326 + ], + "angle": 0, + "content": "16.63 dB" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.33, + 0.372, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.387, + 0.362, + 0.397 + ], + "angle": 0, + "content": "16.60 dB" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.402, + 0.372, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.459, + 0.362, + 0.47 + ], + "angle": 0, + "content": "19.59 dB" + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.101, + 0.465, + 0.113 + ], + "angle": 0, + "content": "Multiphoton" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.115, + 0.464, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.175, + 0.455, + 0.185 + ], + "angle": 0, + "content": "27.22 dB" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.187, + 0.464, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.244, + 0.455, + 0.255 + ], + "angle": 0, + "content": "28.41 dB" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.259, + 0.464, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.316, + 0.455, + 0.326 + ], + "angle": 0, + "content": "22.54 dB" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.33, + 0.464, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.387, + 0.455, + 0.397 + ], + "angle": 0, + "content": "21.79 dB" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.402, + 0.464, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.459, + 0.455, + 0.47 + ], + "angle": 0, + "content": "25.98 dB" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.425, + 0.109, + 0.437 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.481, + 0.484, + 0.606 + ], + "angle": 0, + "content": "Figure 3. Reconstruction of arbitrary \\(32 \\times 32\\) images from 100 000 measured frames using the naive, single, and multiphoton estimators. Each PSNR value is from averaging over 100 trials. Rows (a)-(c) show that the multiphoton estimator is 4 to 6 dB better than the single-photon estimator and around 6 to 11 dB better than the naive estimator. Row (d) shows the impact of rotating the ground truth image on the reconstruction. Row (e) shows the impact of increasing the average number of photons per frame from 3 to 4. The SPE rejects \\(\\sim 85\\%\\) of measured frames at a PPF of 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.483, + 0.732 + ], + "angle": 0, + "content": "multiplexed arrays, high flux also leads to more ambiguous readout frames, which makes image reconstruction challenging. We therefore expect that there exists an optimal incident photon flux level where each of the estimators achieves the minimum mean-square error. The following results, as shown in Fig. 4, validate this hypothesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "A ground truth image can be scaled in experimental settings by selecting a desired integration time or attenuating the incident photon flux using a neutral density filter. To study the dependence of estimator performance on the incident photon flux, we perform a simulation with a high emitted photon flux from a source that is attenuated before being incident on the array. We then study the change in mean-squared error per pixel \\((\\mathrm{MSE} = \\| \\widehat{\\Lambda} - \\Lambda \\|_2^2 / n^2)\\) of the reconstructed image with varying strength of the attenuating filter. The results of this simulation are shown in Fig. 4 for the reconstruction of a \\(32 \\times 32\\) ground truth image in" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.105, + 0.887, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.341, + 0.906, + 0.384 + ], + "angle": 0, + "content": "Figure 4. Mean-squared errors of the estimators studied as functions of mean photons per frame. (Ground truth is image in Fig. 3(c).)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.412, + 0.63, + 0.427 + ], + "angle": 0, + "content": "row (c) of Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.73 + ], + "angle": 0, + "content": "Note that as the mean PPF increases, the MSEs of all the estimators initially decrease, reach a minimum, and then increase. High MSE at low PPF values is due to the lack of detected photons for accurate reconstruction. The increase in MSE at high PPF values is expected due to the high variance of SPE and the high probability of multiphoton events not modeled by the ME. It can also be seen that the minimum MSE of the multiphoton estimator is lower than that of the single-photon estimator, and both of these are much lower than the naive estimator. Furthermore, the minimum value of MSE for the ME occurs at a PPF of 1.4, which is higher than that for the SPE at 0.83, and that for NE at 0.45. We expect these incidence flux values to be realistic in conditions similar to imaging under moonlight, where high-efficiency detectors like SNSPDs are typically used. Thus, the ME produces a better reconstruction compared to both the NE and the SPE when multiple photons are detected simultaneously at the array. This results in an increased photon count rate and a better utilization of the spatial dimensions of the array." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.785, + 0.757 + ], + "angle": 0, + "content": "5.3. 2-, 3- and 4-Photon Estimators" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As explained in Sec. 4.4, our estimator resolves up to 4 simultaneously incident photons. The conditional probability expressions (g's) contain terms corresponding to every possible photon event that can result in the measured readout. However, we can simplify these expressions by restricting the number of modeled events. For instance, if we include only 2-photon coincidences in the case of a \\(2 \\times 2\\) array, only conditional probabilities of events \\(E_{9}\\) and \\(E_{10}\\) would be summed in (13). Fig. 4 shows the performance of the ME" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11411" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.104, + 0.256, + 0.117 + ], + "angle": 0, + "content": "Naive" + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.123, + 0.288, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.207, + 0.272, + 0.22 + ], + "angle": 0, + "content": "22.59 dB" + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.104, + 0.415, + 0.117 + ], + "angle": 0, + "content": "Single-Photon" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.123, + 0.419, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.207, + 0.402, + 0.22 + ], + "angle": 0, + "content": "23.64 dB" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.241, + 0.174, + 0.255 + ], + "angle": 0, + "content": "2-photon" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.258, + 0.195, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.343, + 0.179, + 0.356 + ], + "angle": 0, + "content": "23.83 dB" + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.242, + 0.306, + 0.256 + ], + "angle": 0, + "content": "3-photon" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.258, + 0.331, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.343, + 0.311, + 0.356 + ], + "angle": 0, + "content": "24.60 dB" + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.242, + 0.446, + 0.256 + ], + "angle": 0, + "content": "4-photon" + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.258, + 0.468, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.343, + 0.45, + 0.356 + ], + "angle": 0, + "content": "26.74 dB" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.373, + 0.483, + 0.415 + ], + "angle": 0, + "content": "Figure 5. Image reconstruction by operating each of the studied estimators at mean PPF values corresponding to their lowest MSE's. Indicated values are PSNR of reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.547 + ], + "angle": 0, + "content": "when 2-, 3- and 4-photon are modeled in \\( g \\). It can be seen that the MSE of reconstruction is successively reduced with each additional photon considered in the estimator. The 2-photon estimator achieves its lowest MSE at a mean PPF of 1, while the 3- and 4-photon estimators achieve their lowest MSEs at 1.2 and 1.4 mean PPF, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.629 + ], + "angle": 0, + "content": "Image reconstruction using each of the estimators in Fig. 4 at their optimal mean PPF values is summarized in Fig. 5. We see that the 4-photon estimator achieves the highest PSNR, which is 3 to 4 dB higher than the NE and the SPE." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.66, + 0.406, + 0.677 + ], + "angle": 0, + "content": "5.4. Reduction in Total Integration Time" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Since the multiphoton estimator enables the accurate interpretation of photon coincidences, it can be leveraged to reduce integration times in imaging with single-photon detectors. The performance of the three estimators in achieving a target reconstruction MSE is shown in Fig. 6. It can be seen that to achieve an MSE of 0.01, the ME requires 25,000 measured frames, and the SPE requires 100,000 measured frames, while the NE does not achieve this MSE within the range of frame values studied. This shows that the ME achieves a factor \\(\\sim 4\\) reduction in the integration time for the same reconstruction quality as the other estimators. In comparison, a fully-addressed readout scheme achieves the target MSE with only 5000 frames, but with the increased complexity of readout." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.103, + 0.898, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.343, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Figure 6. Reduction in the required number of frames to achieve the same target MSE by the NE, SPE,ME, and a fully-addressed readout scheme." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.412, + 0.706, + 0.427 + ], + "angle": 0, + "content": "5.5. Cramér-Rao Bound" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.602 + ], + "angle": 0, + "content": "The Cramér-Rao bound (CRB) provides the lowest possible variance of an unbiased estimator for parameter estimation. Here, we compare the performance of the NE, SPE, 2-photon, 3-photon, and 4-photon estimators on the estimation of a \\(2 \\times 2\\) ground truth, against CRB values at different incident photon fluxes. We first compute the Fisher information matrix (FIM) using the likelihood expression for row-column readouts from a \\(2 \\times 2\\) array. The complete derivation of expressions for elements of the FIM are shown in the supplement. The element \\((1,1)\\) of the FIM for a single measurement is" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.626, + 0.905, + 0.694 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {I} _ {1 1} = \\frac {(1 - q _ {4}) (1 - q _ {2} q _ {3})}{q _ {1}} + \\frac {(1 - q _ {2} q _ {3})}{1 - q _ {1}} \\\\ + \\frac {q _ {4} ^ {2} \\left(1 - q _ {2} q _ {3}\\right) ^ {2}}{q _ {1} q _ {4} + q _ {2} q _ {3} - q _ {1} q _ {2} q _ {3} q _ {4}}. \\tag {16} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Other diagonal elements have similar forms as shown in the supplement. The CRB is then computed as the mean of the diagonal elements in \\(\\mathcal{I}^{-1}\\). The variation of CRB values, as well as MSEs of the studied estimators as a function of incident photon flux, is shown in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "It can be seen that the ME curve matches the CRB curve for a range of mean PPF values. At high fluxes, the MSE of the ME increases due to the high variance of the single-photon estimates, which are used to calculate \\( g \\) in (13). In the general case with bright spots at all four pixels of the array, the 4-photon estimator most closely matches the CRB across the range of PPF values studied. The 2- and 3-photon estimators deviate from the CRB due to the modeling of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11412" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.097, + 0.452, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.334, + 0.488, + 0.419 + ], + "angle": 0, + "content": "Figure 7. Comparison of estimator performance with the Cramér-Rao Bound. The ME matches the CRB for a range of incident photon fluxes and deviates from it at high fluxes. With 4 bright spots, model mismatch causes 2- and 3-photon estimators to become biased and deviate from the CRB at lower incident photon fluxes compared to the 4-photon estimator." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.486, + 0.52 + ], + "angle": 0, + "content": "only a subset of events that can cause ambiguities and omission of terms in the expression for conditional probabilities in (13). The dependence of the estimator performance on the ground truth image is shown in the Supplement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.548, + 0.338, + 0.566 + ], + "angle": 0, + "content": "5.6. Increasing array resolution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.713 + ], + "angle": 0, + "content": "Our analysis of the multiphoton estimator's performance so far has been limited to \\(32 \\times 32\\) arrays to match the sensor developed in [37]. Here, we study its use in arrays with higher pixel counts and finer spatial resolutions. We emulate a fixed sensor area with a varying number of pixels and fixed maximum photon flux. This results in photons per pixel being inversely proportional to the number of pixels. We include attenuation optimization for each reconstruction method, similar to Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Results for row 1 of Fig. 3 as the ground truth are shown in Fig. 8 Performance improvement from the ME increases with array size (see Fig. 8(a)). This shows that our method is applicable to array sizes encountered in commercial imaging use cases. Nevertheless, we expect coincidences not modeled to be a limiting factor. Further, computation time increases linearly with an increase in array size. Thus, for a practical implementation of our algorithm, it could be beneficial to divide the sensor area into smaller blocks, the readouts from each of which can be processed using the ME. Such a processing approach would be particularly beneficial in sensor architectures similar to [17]." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.1, + 0.872, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.332, + 0.872, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.566, + 0.909, + 0.608 + ], + "angle": 0, + "content": "Figure 8. Influence of array size. (a) MSEs at optimum attenuation for each estimator. (b) Optimum photons per frame for each estimator." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.654, + 0.8, + 0.67 + ], + "angle": 0, + "content": "6. Conclusions and Future Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We formulate multiphoton coincidence resolution in readout multiplexed single-photon detector arrays as an inverse imaging problem and propose a novel estimator to resolve up to four photon coincidences with row-column multiplexed readouts. Our method enables high-flux, low-latency, high-resolution image reconstructions that significantly improve upon conventional methods. Future work can extend this method to resolve an abitrarily large number of coincident photon events or other types of multiplexed readouts. The use of spatial priors and deep learning methods can be explored to improve the presented solutions. This line of work could be especially useful for the use of SNSPD arrays in photon-starved imaging applications like deep space imaging or biological imaging." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11413" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.186 + ], + "angle": 0, + "content": "This work was supported in part by the U.S. National Science Foundation under Grant 2039762 and in part by a gift from Dr. John Zheng Sun. We thank Prof. Hoover Rueda-Chacon for helpful discussions." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.203, + 0.188, + 0.218 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.227, + 0.482, + 0.31 + ], + "angle": 0, + "content": "[1] M. S. Allman, V. B. Verma, M. Stevens, T. Gerrits, R. D. Horansky, A. E. Lita, F. Marsili, A. Beyer, M. D. Shaw, D. Kumor, R. Mirin, and S. W. Nam. A near-infrared 64-pixel superconducting nanowire single photon detector array with integrated multiplexed readout. Applied Physics Letters, 106 (19), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.482, + 0.354 + ], + "angle": 0, + "content": "[2] Ibrahim Ben Daya. Systematic characterization framework of row-column ultrasound imaging systems. Master's thesis, University of Waterloo, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.357, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[3] PL Bonanno. Physical and cost comparison of smartphone laser autofocus solutions. In Proc. Int. SPAD Sensor Workshop, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.482, + 0.47 + ], + "angle": 0, + "content": "[4] Daniel AB Bonifacio and Maurizio Moralles. Modeling of 3d gamma interaction position in a monolithic scintillator block with a row-column summing readout. In 2012 IEEE Nuclear Science Symposium and Medical Imaging Conference Record (NSS/MIC), pages 2606-2613. IEEE, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.472, + 0.482, + 0.54 + ], + "angle": 0, + "content": "[5] Jihwan Boo, Mark D Hammig, and Manhee Jeong. Row-column readout method to mitigate radiographic-image blurring from multipixel events in a coded-aperture imaging system. IEEE Transactions on Nuclear Science, 68(5):1175-1183, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.543, + 0.482, + 0.611 + ], + "angle": 0, + "content": "[6] Caroline K Brennan, Zi Yao, Anastasia A Ionkina, Colin M Rathbun, Buvaneshwari Sathishkumar, and Jennifer A Prescher. Multiplexed bioluminescence imaging with a substrate unmixing platform. Cell Chemical Biology, 29(11): 1649-1660, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.615, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[7] Claudio Bruschini, Harald Homulle, Ivan Michel Antolovic, Samuel Burri, and Edoardo Charbon. Single-photon avalanche diode imagers in biophotonics: review and outlook. Light: Science & Applications, 8(1):87, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[8] Nicolas Callens and Georges GE Gielen. Analysis and comparison of readout architectures and analog-to-digital converters for 3d-stacked cmos image sensors. IEEE Transactions on Circuits and Systems I: Regular Papers, 68(8):3117-3130, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.744, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[9] Francesco Ceccarelli, Giulia Acconcia, Angelo Gulinatti, Massimo Ghioni, Ivan Rech, and Roberto Osellame. Recent advances and future perspectives of single-photon avalanche diodes for quantum photonics applications. Advanced Quantum Technologies, 4(2):2000102, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.482, + 0.869 + ], + "angle": 0, + "content": "[10] Garry Chinn, Peter D. Olcott, and Craig S. Levin. Sparse signal recovery methods for multiplexing pet detector readout. IEEE Transactions on Medical Imaging, 32(5):932-942, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Iris Cusini, Davide Berretta, Enrico Conca, Alfonso Incoronato, Francesca Madonini, Arianna Adelaide Maurina," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.227, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Chiara Nonne, Simone Riccardo, and Federica Villa. Historical perspectives, state of art and research trends of SPAD arrays and their applications (Part II: SPAD arrays). Frontiers in Physics, 10:906671, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[12] Iman Esmaeil Zadeh, J Chang, Johannes WN Los, Samuel Gyger, Ali W Elshaari, Stephan Steinhauer, Sander N Dorenbos, and Val Zwiller. Superconducting nanowire single-photon detectors: A perspective on evolution, state-of-the-art, future developments, and applications. Applied Physics Letters, 118(19), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[13] Andrew L Goertzen, Xuezhu Zhang, Megan M McClarty, Eric J Berg, Chen-Yi Liu, Piotr Kozlowski, Fabrice Retiere, Lawrence Ryner, Vesna Sossi, Greg Stortz, et al. Design and performance of a resistor multiplexing readout circuit for a sipm detector. IEEE Transactions on Nuclear Science, 60 (3):1541-1549, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.316, + 0.904, + 0.37 + ], + "angle": 0, + "content": "[14] Rémi Gribonval, Antoine Chatalic, Nicolas Keriven, Vincent Schellekens, Laurent Jacques, and Philip Schniter. Sketching data sets for large-scale learning. IEEE Signal Process. Mag., 38(5):12-36, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.904, + 0.44 + ], + "angle": 0, + "content": "[15] Yanqiu Guan, Haochen Li, Li Xue, Rui Yin, Labao Zhang, Hao Wang, Guanghao Zhu, Lin Kang, Jian Chen, and Peiheng Wu. Lidar with superconducting nanowire single-photon detectors: Recent advances and developments. Optics and Lasers in Engineering, 156:107102, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[16] Yanqiu Guan, Haochen Li, Labao Zhang, Hao Wang, Guanglong He, Biao Zhang, Yue Fei, Jiayu Lv, Xiao Zhang, Rui Yin, et al. SNSPD array with single-channel readout based on compressive sensing. ACS Photonics, 9(9):3102-3109, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[17] Hao Hao, Qing-Yuan Zhao, Yang-Hui Huang, Jie Deng, Fan Yang, Sai-Ying Ru, Zhen Liu, Chao Wan, Hao Liu, Zhi-Jian Li, et al. A compact multi-pixel superconducting nanowire single-photon detector array supporting gigabit space-to-ground communications. Light: Science & Applications, 13(1):25, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[18] Sam W. Hutchings, Nick Johnston, Istvan Gyongy, Tarek Al Abbas, Neale A. W. Dutton, Max Tyler, Susan Chan, Jonathan Leach, and Robert K. Henderson. A reconfigurable 3-D-stacked SPAD imager with in-pixel histogramming for flash LIDAR or high-speed time-of-flight imaging. IEEE J. Solid-State Circuits, 54(11):2947-2956, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[19] Atul Ingle and David Maier. Count-free single-photon 3d imaging with race logic. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-12, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[20] Jørgen Arendt Jensen, Mikkel Schou, Lasse Thurmann Jørgensen, Borislav G Tomov, Matthias Bo Stuart, Marie Sand Traberg, Iman Taghavi, Sigrid Huesebø Øygaard, Martin Lind Ommen, Kitty Steenberg, et al. Anatomic and functional imaging using row-column arrays. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 69(10):2722-2738, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[21] Ling-Dong Kong, Hui Wang, Qing-Yuan Zhao, Jia-Wei Guo, Yang-Hui Huang, Hao Hao, Shi Chen, Xue-Cou Tu, LaBao Zhang, Xiao-Qing Jia, et al. Readout-efficient superconducting nanowire single-photon imager with orthogonal time-amplitude multiplexing by hotspot quantization. Nature Photonics, 17(1):65-72, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11414" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[22] Sangwon Lee, Yong Choi, Jihoon Kang, and Jin Ho Jung. Development of a multiplexed readout with high position resolution for positron emission tomography. *Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment*, 850:42-47, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[23] Feiyan Li, Hang Han, Qi Chen, Biao Zhang, Han Bao, Yue Dai, Rui Ge, Shuya Guo, Guanglong He, Yue Fei, et al. Saturation efficiency for detecting \\(1550\\mathrm{nm}\\) photons with a \\(2\\times 2\\) array of \\(\\mathrm{Mo}_{0.8}\\mathrm{Si}_{0.2}\\) nanowires at \\(2.2\\mathrm{K}\\). Photonics Research, 9(3):389-394, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[24] Adam N McCaughan, Yao Zhai, Boris Korzh, Jason P Allmaras, Bakhrom G Oripov, Matthew D Shaw, and Sae Woo Nam. The thermally coupled imager: A scalable readout architecture for superconducting nanowire single photon detectors. Applied Physics Letters, 121(10), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[25] Kazuhiro Morimoto. Megapixel spad cameras for time-resolved applications. Technical report, EPFL, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[26] Peter D Olcott, Garry Chinn, and Craig S Levin. Compressed sensing for the multiplexing of pet detectors. In 2011 IEEE Nuclear Science Symposium Conference Record, pages 3224-3226. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[27] B. G. Oripov, D. S. Rampini, J. Allmaras, M. D. Shaw, S. W. Nam, B. Korzh, and A. N. McCaughan. A superconducting nanowire single-photon camera with 400,000 pixels. Nature, 622(7984):730-734, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[28] François Piron, Daniel Morrison, Mehmet Rasit Yuce, and Jean-Michel Redoute. A review of single-photon avalanche diode time-of-flight imaging sensor arrays. IEEE Sensors Journal, 21(11):12654-12666, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[29] Joshua Rapp, Julian Tachella, Yoann Altmann, Stephen McLaughlin, and Vivek K Goyal. Advances in single-photon lidar for autonomous vehicles: Working principles, challenges, and recent advances. IEEE Signal Processing Magazine, 37(4):62-71, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[30] Morten Fischer Rasmussen and Jørgen Arendt Jensen. 3-d ultrasound imaging performance of a row-column addressed 2-d array transducer: A measurement study. In 2013 IEEE International Ultrasonics Symposium (IUS), pages 1460-1463. IEEE, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[31] Morten Fischer Rasmussen, Thomas Lehrmann Christiansen, Erik Vilain Thomsen, and Jørgen Arendt Jensen. 3-d imaging using row-column-addressed arrays with integrated apodization-part i: apodization design and line element beamforming. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 62(5):947-958, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[32] Kaustubh Sadekar, David Maier, and Atul Ingle. Single-photon 3d imaging with equi-depth photon histograms. In European Conference on Computer Vision, pages 381-398. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[33] Ewout van den Berg, Emmanuel Candes, Garry Chinn, Craig Levin, Peter Demetri Olcott, and Carlos Sing-Long. Single-photon sampling architecture for solid-state imaging sensors. Proceedings of the National Academy of Sciences, 110(30): E2752-E2761, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[34] Hailu Wang, Jiaxiang Guo, Jinshui Miao, Wenjin Luo, Yue Gu, Runzhang Xie, Fang Wang, Lili Zhang, Peng Wang, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Weida Hu. Emerging single-photon detectors based on low-dimensional materials. Small, 18(5):2103963, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[35] Hui Wang, Zhi-jian Li, Xue-Mei Hu, Hao Hao, Jia-wei Guo, Yang-hui Huang, Hao Liu, Chao Wan, Xue-cou Tu, Xiaoqing Jia, et al. Image distortion by ambiguous multiple-photon detections in a superconducting nanowire single-photon imager and the correction method. Optics Express, 31(14):23579-23588, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[36] Mel White, Shahaboddin Ghajari, Tianyi Zhang, Akshit Dave, Ashok Veeraraghavan, and Alyosha Molnar. A differential SPAD array architecture in \\(0.18\\mu \\mathrm{m}\\) CMOS for HDR imaging. In Proc. IEEE Int. Symp. Circuits Syst., pages 292-296, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[37] Emma E Wollman, Varun B Verma, Adriana E Lita, William H Farr, Matthew D Shaw, Richard P Mirin, and Sae Woo Nam. Kilopixel array of superconducting nanowire single-photon detectors. Optics express, 27(24):35279-35289, 2019. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[38] Lixing You. Superconducting nanowire single-photon detectors for quantum information. Nanophotonics, 9(9):2673-2692, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.458 + ], + "angle": 0, + "content": "[39] Qing-Yuan Zhao, Di Zhu, Niccolò Calandri, Andrew E Dane, Adam N McCaughan, Francesco Bellei, Hao-Zhu Wang, Daniel F Santavicca, and Karl K Berggren. Single-photon imager based on a superconducting nanowire delay line. Nature Photonics, 11(4):247-251, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11415" + } + ] +] \ No newline at end of file diff --git a/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/960040c5-ed98-43eb-9b66-fa83e44db35d_origin.pdf b/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/960040c5-ed98-43eb-9b66-fa83e44db35d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..abf062bbb6e664fbd7dfa32ac1d37aeaae71d901 --- /dev/null +++ b/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/960040c5-ed98-43eb-9b66-fa83e44db35d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17e6850ebb344596e1fba649112437391237369b20e5dfa3a6975346ba4692d5 +size 1800763 diff --git a/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/full.md b/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7036da60f5e9c3067a97ade93b7d8b719623f91a --- /dev/null +++ b/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/full.md @@ -0,0 +1,378 @@ +# Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays + +Shashwath Bharadwaj, Ruangrawee Kitichotkul, Akshay Agarwal and Vivek K Goyal Boston University + +shash@bu.edu + +# Abstract + +Readout multiplexing is a promising solution to overcome hardware limitations and data bottlenecks in imaging with single-photon detectors. Conventional multiplexed readout processing creates an upper bound on photon counts at a very fine time scale, where frames with multiple detected photons must either be discarded or allowed to introduce significant bias. We formulate multiphoton coincidence resolution as an inverse imaging problem and introduce a solution framework to probabilistically resolve the spatial locations of photon incidences. Specifically, we develop a theoretical abstraction of row-column multiplexing and a model of photon events that make readouts ambiguous. Using this, we propose a novel estimator that spatially resolves up to four coincident photons. Monte Carlo simulations show that our proposed method increases the peak signal-to-noise ratio (PSNR) of reconstruction by 3 to 4 dB compared to conventional methods under optimal incident flux conditions. Additionally, this method reduces the required number of readout frames to achieve the same mean-squared error as other methods by a factor of $\sim 4$ . Finally, our method matches the Cramér-Rao bound for detection probability estimation for a wider range of incident flux values compared to conventional methods. While demonstrated for a specific detector type and readout architecture, this method can be extended to more general multiplexing with different detector models. + +# 1. Introduction + +Single-photon detectors have been widely studied for use in applications like biological imaging [7, 34], lidar [15, 29], and quantum optics [9] due to properties such as single-photon sensitivity and high-precision time-tagging of photon arrivals. Single-photon avalanche diodes (SPADs) have been of particular interest due to compatibility with CMOS manufacturing techniques and room temperature operation [11]. However, they have low detection efficiencies (especially at mid infrared wavelengths), which means that only a small fraction of incident light is detected as pho + +tons. Superconducting nanowire single-photon detectors (SNSPDs) are emerging devices that offer near-unity quantum efficiency across a range of wavelengths, low dark counts, and low jitter, making them compelling alternatives to SPADs [38]. However, the requirement of cryogenic cooling to maintain superconductivity has largely restricted their use to laboratory experiments. + +Single-photon imaging in typical commercial applications requires arrays of single-photon detectors [28]. While megapixel SPAD arrays have already been deployed in mass consumer products [3, 25], high data transfer rates from these arrays for further processing remain a bottleneck. For instance, typical room-scale ranging experiments with SPAD array-based single-photon lidar can have data generation rates of the order of gigabits per second when the array resolution approaches the megapixel scale. It is attractive to compress the readout to less than a stream of all digitized detection times. Methods like in-pixel histogramming [18], histogram compression through sketching [14], equidepth histogramming [19, 32], and differential readouts [36] have been proposed to address high data transfer rates. These approaches to avoid the cost of high temporal resolution are complementary to our emphasis on transverse spatial dimensions in this work. + +Implementing high spatial resolution SNSPD arrays is a significant challenge. As the number of pixels in the array grows, the heat load introduced by reading out each pixel individually becomes incompatible with the cryogenics of the system [12]. To mitigate this, recent efforts have focused on addressing groups of pixels in the readout, starting with Allman et al. [1] who introduced row-column readout multiplexing. In this mechanism, readouts indicate photon detection at each row and each column of the array instead of every pixel, which reduces the required number of readout lines for an $n \times n$ array from $n^2$ to $2n$ . Wollman et al. [37] extended this scheme to demonstrate the first kilopixel SNSPD array. McCaughan et al. [24] introduced the thermally coupled imager with time-of-flight multiplexing, where separate readout lines for each row and column are replaced with a single bus for all rows and all columns. This makes the required number of readouts independent of + +the array size. Recently, Oripov et al. [27] leveraged this method to demonstrate a 400,000 pixel superconducting camera, showing the viability of this scheme to approach the sizes of commercial imaging arrays. + +Despite the gain in scaling array sizes, readout multiplexing requires the incident photon flux to be very low for unambiguous signal reconstruction, limiting the array's photon count rate. When two or more photons are incident on the array within the period of a single acquisition, row-column readouts can give an ambiguous set of candidate locations for their incidences. Traditional reconstruction techniques [37] do not distinguish between these locations and assume equal probabilities of detection at each candidate pixel. This results in the misattribution of photon counts to locations where no photon incidences occurred. Alternatively, since readouts with a single detected photon provide the exact row and column indices where the photon was incident on the array, ambiguous readouts with multiple detections can be discarded for signal reconstruction. However, this leads to high variance in the reconstruction (especially at high incident fluxes) and underutilization of the spatial dimensions of the array by requiring most readouts to contain detections at a single pixel. Resolving the spatial locations of multiphoton coincidences using multiplexed readouts is an ill-posed inverse problem. + +Here, we propose a solution framework to leverage most of the measured data to reduce the mean-squared error of image reconstruction. Specifically, we develop a multiphoton estimator that redistributes photon counts from ambiguous readouts to candidate pixels such that an approximate likelihood of observations is maximized. A key choice for our algorithm design is the omission of spatial priors in order to not obscure the source of performance gain. Spatial priors can be used in conjunction with the developed method for further improvements. Monte Carlo simulations show that our proposed method increases the peak signal-to-noise ratio (PSNR) of reconstruction by 3 to $4\mathrm{dB}$ compared to conventional methods under optimal incident flux conditions. Further, the optimal incident photon flux for our method is $\sim 1.4$ photons per readout which is 0.5 to 1.1 photons per readout higher than conventional methods. We also demonstrate that our estimator matches the Cramér-Rao Bound (CRB) for the estimation of photon detection probabilities at each pixel of an array across a range of incident photon flux values. Finally, our method achieves a factor of $\sim 4$ reduction in the required number of readout frames for the same mean-squared error of image reconstruction. These results demonstrate that our proposed estimator is well-suited for high-flux, low-latency imaging. + +# 2. Related Work + +Image reconstruction from multiplexed readouts. + +Compressed sensing methods have been explored for the + +design of multiplexed readout architectures in areas like positron emission tomography (PET) [10, 13, 22, 26], biological imaging [6, 30] and gamma-ray imaging [4, 5]. While readout multiplexing for specific imaging applications has been extensively studied [8, 20], the focus of these works is on the modeling of optical and electrical properties of the imaging system such as the shape of the wavefront of the detected light [2, 31], coded apertures for rejecting multipixel events [5] and IC design for resistor multiplexing [13]. The explicit study of individual photon events measured by single-photon detectors and their impact on the multiplexed readouts does not yet exist. Our work introduces a theoretical study of the readouts based on the combinatorics of photon events and an approximate maximum likelihood estimation of photon detection probabilities at each pixel. We also analyze the performance of a novel multiphoton estimator with variations in imaging parameters. Our work is closest to the analysis presented by van den Berg et al. [33], where group-testing inspired surface codes are studied for deterministic multiphoton coincidence resolution in time-to-digital converter readouts from SPAD arrays. However, our work specifically addresses intensity imaging with SNSPD arrays where a small probability of error is tolerable. We provide a probabilistic model of the readouts and solutions for image reconstruction from a specific device architecture for up to four coincident photons. + +Imaging with SNSPD arrays. While large-scale SNSPD array development is still in the nascent stages, several works have conducted proof-of-principle imaging experiments with prototypes. Zhao et al. introduced a delay-line-based superconducting imager with an effective pixel count of $\sim 590$ [39]. This method inspired the thermally coupled imager [24], which in turn resulted in the 400,000-pixel camera by Oripov et al. [27]. Subsequent efforts to use SNSPD arrays for imaging [16, 21, 35] involve intricate hardware control such as bias-current sweeping, use of idle pixels, and geometric design of the nanowire structure. Our solution is purely computational and does not require additional hardware modifications. In addition, the modeling of incident photon events and their multiplexed readouts can be extended to several detector models in high-energy physics or particle detection. While the method described in this work provides a framework to resolve an arbitrarily large number of coincident photons, we limit the discussion in the next sections to up to four coincident photons to make the demonstrated solutions computationally tractable and easier to illustrate and analyze. + +# 3. Measurement and Readout Model + +Figure 1 illustrates imaging a scene using an $n \times n$ SNSPD array with row-column readouts. Measurements are read out after integrating the incident signal for a fixed time in + +![](images/52a45b326e446bd427013d7350541df707f25a8777cfe0c29caa2082ec7d5307.jpg) +Figure 1. Schematic of image reconstruction with row-column readouts. (Left) Multiple frames of a ground truth image are measured using an SNSPD array to give row-column readout vectors. (Right) Comparison of conventional and proposed reconstruction techniques to estimate the ground truth. + +terval. Let $\Lambda \in \mathbb{R}^{n\times n}$ be the flux of an $n\times n$ ground truth image in that integration time. The number of photons arriving at pixel $(i,j)$ at a discrete time index $t$ is modeled as + +$$ +X _ {i j} ^ {t} \sim \operatorname {P o i s s o n} (\Lambda_ {i j}). \tag {1} +$$ + +We assume pixel saturation with the detection of a single photon, i.e., the detectors are not photon-number resolving [23]. Thus, we define binary-valued photon-incidence indicator $Y^{t} \in \{0,1\}^{n \times n}$ by $Y_{ij}^{t} = \mathbf{1}[X_{ij}^{t} > 0]$ , where $\mathbf{1}[\cdot]$ is the indicator function. Then, $Y^{t}$ is a matrix where each entry is an independent Bernoulli random variable $Y_{ij}^{t} \sim \mathrm{Bernoulli}(q_{ij})$ , where + +$$ +q _ {i j} = 1 - e ^ {- \Lambda_ {i j}} \tag {2} +$$ + +is the probability of photon detection at pixel $(i,j)$ . + +Each measurement consists of a row readout $R^t \in \{0,1\}^n$ and a column readout $C^t \in \{0,1\}^n$ : + +$$ +R _ {i} ^ {t} = \mathbf {1} \left[ X _ {i j} ^ {t} > 0 \text {f o r a n y} j \in \{1, \dots , n \} \right], \tag {3a} +$$ + +$$ +C _ {j} ^ {t} = \mathbf {1} \left[ X _ {i j} ^ {t} > 0 \text {f o r a n y} i \in \{1, \dots , n \} \right]. \tag {3b} +$$ + +We then define a readout frame at time $t$ as $(R^t, C^t)$ . When multiple photons are incident on the array within a single integration period, the problem of resolving the spatial locations of their incidence is ill-posed. As an example, consider a $2 \times 2$ array with pixels numbered from 1 to 4 as shown in event $E_0$ of Fig. 2. The frame ([1,1],[1,1]) could have resulted from photons detected at pixels $\{1,4\}$ or $\{2,3\}$ or any of: $\{1,2,3\}$ , $\{1,2,4\}$ , $\{1,3,4\}$ , $\{2,3,4\}$ , or $\{1,2,3,4\}$ . More generally, a readout frame is unambiguous if $\sum_{i=1}^{n} R_i^t = 1$ or $\sum_{j=1}^{n} C_j^t = 1$ ; otherwise, it is + +ambiguous. Our problem is to estimate $\Lambda$ from measurements in $T$ frames $\{(R^t,C^t)\}_{t = 1}^T$ . In particular, we are interested in exploiting the information contained in ambiguous readout frames. + +# 4. Estimators + +In this section, we introduce the three estimators compared in this study, namely the naive, single-photon, and multiphoton estimators. Each is formulated first as an estimator $\widehat{q}_{ij}$ for the probability $q_{ij}$ defined in (2). The incident flux at each pixel can then be estimated elementwise as + +$$ +\widehat {\Lambda} _ {i j} = - \log \left(1 - e ^ {- \widehat {q} _ {i j}}\right). \tag {4} +$$ + +# 4.1. Naive Estimator + +The naive estimator (NE) assumes that a photon is detected at each of the candidate pixels in ambiguous readout frames. Thus, the probability of detecting a photon at pixel $(i,j)$ is estimated as + +$$ +\widehat {q} _ {i j, \mathrm {n a i v e}} = \frac {1}{T} \sum_ {t = 1} ^ {T} R _ {i} ^ {t} C _ {j} ^ {t}. \tag {5} +$$ + +The naive estimator overestimates $q_{ij}$ because it misattributes photon counts in ambiguous frames to pixels where no photon incidence occurred. The extra counts imputed by this scheme produce ghost spots in the reconstructed image, as depicted in Fig. 1. + +![](images/47c47aa4a7e1833e48837c1ee4d78c86baf585f5be0b9ff544630fed751bbf23.jpg) +Figure 2. Row-column readout frames for a $2 \times 2$ array. Yellow spots indicate incident photon locations. Accounting for pixel saturation, 16 photon detection events are possible with this array. Row-column readout results in 9 unambiguous events $E_0, E_1, \ldots, E_8$ and 7 events $E_9, E_{10}, \ldots, E_{15}$ that are ambiguous because they all result in $R^t = [1,1]$ and $C^t = [1,1]$ . Pixel numbers are indicated in event $E_0$ , which corresponds to an empty readout. + +The positive bias of the naive estimator is + +$$ +\begin{array}{l} \mathbb {E} \left[ \widehat {q} _ {i j, \text {n a i v e}} \right] - q _ {i j} \\ = \left(1 - q _ {i j}\right) \left(1 - \prod_ {k \in R _ {i} ^ {\prime}} \left(1 - q _ {k j}\right)\right) \left(1 - \prod_ {\ell \in C _ {j} ^ {\prime}} \left(1 - q _ {i \ell}\right)\right), \tag {6} \\ \end{array} +$$ + +where $R_{i}^{\prime}$ is the set of rows except row $i$ and $C_j^\prime$ is the set of columns except column $j$ . A derivation is given in the supplement. + +# 4.2. Single-Photon Estimator + +A simple way to prevent misattributions and thus reduce bias is to only use frames with a single detected photon in the reconstruction since these unambiguously give the spatial locations of photon incidence. We thus define the single-photon estimator (SPE): + +$$ +\widehat {q} _ {i j, \text {s i n g l e}} = \left\{ \begin{array}{l l} \frac {N _ {i j}}{N _ {i j} + N _ {0}}, & \text {i f} N _ {i j} + N _ {0} > 0; \\ 0, & \text {i f} N _ {i j} + N _ {0} = 0, \end{array} \right. \tag {7} +$$ + +where $N_{ij}$ is the number of frames where a photon is detected only at pixel $(i,j)$ , and $N_0$ is the number of frames without any detected photons across the whole array. The SPE is unbiased when $N_{i,j} + N_0 > 0$ , as shown in the supplement. However, the discarding of frames with multiple detected photons results in high variance. + +# 4.3. Multiphoton Estimator + +Now, we develop a multiphoton estimator (ME) that uses photon counts from ambiguous readouts for reconstruction. + +For simplicity, here we derive expressions for detection probability estimates at each pixel of a $2 \times 2$ array and illustrate how this estimator maximizes an approximate likelihood of observations. The general form of the ME is presented in the supplement. If each pixel were read out individually, the likelihood of observations could be written as + +$$ +\mathcal {L} _ {p} = \prod_ {i = 1} ^ {2} \prod_ {j = 1} ^ {2} (1 - q _ {i j}) ^ {v _ {i j}} \left(q _ {i j}\right) ^ {f _ {i j}}, \tag {8} +$$ + +where $v_{ij}$ is the total number of frames with no detections at pixel $(i,j)$ and $f_{ij}$ is the total number of frames with a detection at pixel $(i,j)$ . Then, the maximum likelihood estimate $\widehat{q}_{ij}^{\mathrm{ML}}$ could be computed as + +$$ +\hat {q} _ {i j} ^ {\mathrm {M L}} = \frac {f _ {i j}}{T}, \quad i = 1, 2, j = 1, 2, \tag {9} +$$ + +where $T$ is the total number of measured frames. + +With row-column readouts, our knowledge of locations of photon detections is reduced. We define $v_{ij}$ and $f_{ij}$ from unambiguous frames only and find that the likelihood of observations becomes + +$$ +\mathcal {L} \left(\left\{\left(R ^ {t}, C ^ {t}\right) \right\} _ {t = 1} ^ {T}; q\right) = U (q) A (q), \tag {10} +$$ + +with the factor + +$$ +U (q) = \prod_ {i = 1} ^ {2} \prod_ {j = 1} ^ {2} \left(1 - q _ {i j}\right) ^ {v _ {i j}} \left(q _ {i j}\right) ^ {f _ {i j}} \tag {11} +$$ + +from unambiguous frames and the factor + +$$ +A (q) = \left(q _ {1 1} q _ {2 2} + q _ {1 2} q _ {2 1} - q _ {1 1} q _ {1 2} q _ {2 1} q _ {2 2}\right) ^ {M _ {9}} \tag {12} +$$ + +from ambiguous frames. Here, $M_9$ is the total number of ambiguous readout frames due to events $E_9,\ldots ,E_{15}$ shown in Fig. 2. + +The factor in (12) results from the sum of probabilities $\sum_{k=9}^{15} \mathbb{P}(E_k)$ . This makes it challenging to find analytical expressions for the maximum likelihood estimates of detection probabilities. However, if we can estimate the fractions of the $M_9$ frames resulting from each of the events $E_9, \ldots, E_{15}$ , we can re-write (12) as a product of probabilities of events $E_9, \ldots, E_{15}$ to express (10) similarly to (8) and obtain analytical expressions for approximate maximum likelihood estimates $\widehat{q}_{ij}^{\mathrm{a}}$ . We first find expressions for the conditional probabilities $g_9, \ldots, g_{15}$ of events $E_9, \ldots, E_{15}$ , given that an ambiguous frame was read out. For example, $g_9$ is the conditional probability of event $E_9$ given that a frame $R^t = [1,1]$ , $C^t = [1,1]$ was measured and is given by + +$$ +g _ {9} = \frac {\mathbb {P} \left(E _ {9}\right)}{\sum_ {k = 9} ^ {1 5} \mathbb {P} \left(E _ {k}\right)}. \tag {13} +$$ + +In addition to the inherent limitation of having a finite amount of data, computing $\mathbb{P}(E_9),\ldots ,\mathbb{P}(E_{15})$ is hindered by ambiguities. Thus, we estimate these probabilities using the measured single-photon frames, which are unambiguous and unbiased. For example, an estimate of $\mathbb{P}(E_9)$ can be calculated as + +$$ +\widehat {\mathbb {P}} (E _ {9}) = \widehat {q} _ {1 1} ^ {\mathrm {s}} \widehat {q} _ {2 2} ^ {\mathrm {s}} (1 - \widehat {q} _ {1 2} ^ {\mathrm {s}}) (1 - \widehat {q} _ {2 1} ^ {\mathrm {s}}). \tag {14} +$$ + +Similar expressions can be derived for $\widehat{\mathbb{P}}(E_{10}),\ldots,\widehat{\mathbb{P}}(E_{15})$ which are then substituted in (13) in place of $\mathbb{P}(E_9),\dots ,\mathbb{P}(E_{15})$ to get estimates $\widehat{g}_9,\dots ,\widehat{g}_{15}$ of conditional probabilities $g_{9},\ldots ,g_{15}$ . The fractions of the $M_9$ frames resulting from events $E_{9},\ldots E_{15}$ can then be estimated as $\widehat{g}_9M_9,\ldots ,\widehat{g}_{15}M_9$ . Finally, these estimates can be used to simplify (12) to a product of probabilities and obtain the approximate maximum likelihood estimates + +$$ +\begin{array}{l} \widehat {q} _ {1 1} ^ {\mathrm {a}} = \frac {M _ {1} + M _ {5} + M _ {6} + M _ {9} (\widehat {g} _ {9} + \widehat {g} _ {1 1} + \widehat {g} _ {1 2} + \widehat {g} _ {1 3} + \widehat {g} _ {1 5})}{T}, \\ \widehat {q} _ {1 2} ^ {\mathrm {a}} = \frac {M _ {2} + M _ {5} + M _ {7} + M _ {9} (\widehat {g} _ {1 0} + \widehat {g} _ {1 1} + \widehat {g} _ {1 2} + \widehat {g} _ {1 4} + \widehat {g} _ {1 5})}{T}, \\ \widehat {q} _ {2 1} ^ {\mathrm {a}} = \frac {M _ {3} + M _ {6} + M _ {8} + M _ {9} (\widehat {g} _ {1 0} + \widehat {g} _ {1 2} + \widehat {g} _ {1 3} + \widehat {g} _ {1 4} + \widehat {g} _ {1 5})}{T}, \\ \hat {q} _ {2 2} ^ {\mathrm {a}} = \frac {M _ {4} + M _ {7} + M _ {8} + M _ {9} (\widehat {g} _ {9} + \widehat {g} _ {1 1} + \widehat {g} _ {1 3} + \widehat {g} _ {1 4} + \widehat {g} _ {1 5})}{T} \tag {15} \\ \end{array} +$$ + +where $M_1, \ldots, M_8$ are the numbers of unambiguous readout frames corresponding to events $E_1, \ldots, E_8$ . + +While computing the bias of the ME analytically is complex, numerical simulations show that it has a low bias. Furthermore, since the ME depends on the SPE for an initial estimate of the detection probabilities, when the SPE is poor due to few measured single-photon frames, the reduction in MSE from the ME is relatively lesser than in cases when the SPE has low variance. + +# 4.4. Scaling to Higher Dimensional Arrays + +In a $2 \times 2$ array, there is a single type of ambiguous readout which indicates photon incidences at both rows and both columns of the array. As the spatial dimension of the array increases, there are more types of ambiguous readouts. For instance, with a $3 \times 3$ array, we can have the following ambiguous readouts: + +- 2 rows and 2 columns fired - 4 candidate pixel locations; +- 2 rows and 3 columns fired - 6 candidate pixel locations; +- 3 rows and 2 columns fired - 6 candidate pixel locations; +- 3 rows and 3 columns fired -9 candidate pixel locations. Resolving these additional ambiguous readouts requires the modification of the expressions for conditional probabilities $g$ . As the number of rows and columns with photon coincidences increases, the number of unique terms in the expressions for $g$ grows rapidly. Hence, we restrict our analysis and simulations to at most 4-photon coincidences and ignore terms resulting from 5-photon coincidences and above. + +We expect that this assumption introduces a small bias in the ME, especially at higher incident fluxes. Nevertheless, the improvement over handling only single-photon frames can be substantial. + +# 5. Results + +# 5.1. Imaging Arbitrary Scenes + +Figure 3 shows the Monte Carlo simulation results of reconstructing $32 \times 32$ images from 100,000 measured frames using each of the estimators. Note that these are estimates of the flux $\hat{\Lambda}_{ij}$ at each pixel, computed using the detection probability estimate $\hat{q}_{ij}$ from the three estimators followed by the transformation in (4). Reconstructed in rows (a), (b), (c), and (d) are at a scaling of 3 mean photons per frame (PPF), while row (e) is at a scaling of 4 mean PPF. Mean PPF is the average number of detected photons in each measured frame. Each experiment is repeated 100 times for the computation of PSNR. + +In rows (a), (b), and (c), it can be seen that the NE introduces horizontal and vertical streaks due to misattributions from multiphoton coincidences. For example, in row (a) the SPE achieves $\sim 5$ dB PSNR improvement over the NE and does not show the same artifacts, although the petals of the flower appear noisy. + +Our proposed ME outperforms the naive estimator by $\sim 11\mathrm{dB}$ while its improvement over the single-photon estimator is $\sim 6\mathrm{dB}$ . Further, we note that the ME preserves features in the ground truth image better than both of the other estimators, in addition to eliminating misattribution artifacts. The petals of the flower in row (a) are more clearly visible, while these are missing in the other reconstructions. The improvement in the reconstruction depends on some features of the ground truth image. For example, in row (d) we consider a rotated version of the image in row (a). It can be seen that the misattributions in the NE are more concentrated towards the center-left of the image as compared to row (a) where the horizontal and vertical streaks cut across the image. The PSNR of the naive reconstruction is lower than that in row (a), however, the multiphoton reconstruction still achieves $\sim 5\mathrm{dB}$ improvement. In row (e) we consider the image from row (b) with a PPF scaling of 4. We note that the PSNR of the NE reduces by $\sim 0.6$ dB while the PSNR of the SPE and ME reduce by $\sim 3$ dB each. This behavior is expected as an increase in the average PPF causes the number of frames with a single detected photon to decrease, which reduces the PSNR of the SPE and consequently the ME. However, the performance of the ME can be improved by considering additional multiphoton events. + +# 5.2. Optimal Photons Per Frame + +In single-photon imaging, high signal strength is often beneficial for parameter estimation. However, for readout + +![](images/0854ebe751fc7ecad720bc03f241ecdb7133a1020552c100d4a7b1d284e4c97d.jpg) +(a) +Ground Truth + +![](images/4cb4c05d0831fd24ffe7b79b7bf39bdb983a53fc3b874b821c1a8ddc4ee36fbe.jpg) +Naive +16.39 dB + +![](images/a109cc8f01f6d863c5c459a35bf5c37e836c4604cd807d19fcb114d4a9707773.jpg) +17.05 dB + +![](images/d2fb0d1ccf9d1b77694f31e870c4d679ddba6ce9a7349b645e853b6a326be5af.jpg) +13.17 dB + +![](images/524604aa7eb71ab13b7c2cad14cc99a59791853cfc1370e1e7e3831fd50c5c59.jpg) +16.15 dB + +![](images/aee72b6f18da28f8708d513436647c4a478c143cc03330ded32191f5c5c73fc0.jpg) +16.44 dB +(e) +Figure 3. Reconstruction of arbitrary $32 \times 32$ images from 100 000 measured frames using the naive, single, and multiphoton estimators. Each PSNR value is from averaging over 100 trials. Rows (a)-(c) show that the multiphoton estimator is 4 to 6 dB better than the single-photon estimator and around 6 to 11 dB better than the naive estimator. Row (d) shows the impact of rotating the ground truth image on the reconstruction. Row (e) shows the impact of increasing the average number of photons per frame from 3 to 4. The SPE rejects $\sim 85\%$ of measured frames at a PPF of 3. + +![](images/c0945a320a4448bf3e9b86e026f7201c8cf477356d8d58644c5b43fb1a234bcd.jpg) +Single-Photon + +![](images/b581499b19afb24926b662f06da2ecf86795de9e932322970d6c0fabb929a862.jpg) +21.55 dB +22.43 dB + +![](images/e38a06ce716abb9928099add0c336cd3d3d1a5c710d3b9fa41083b48619531f8.jpg) +16.63 dB + +![](images/13fe92ea23853e0cdcf7125dc8e641d83e61ce2e61d6bf02097d74b56f957f43.jpg) +16.60 dB + +![](images/3685f1a69eb6a9aff78ae6421d09b6d02b44daac875fc6dcc6d16081ca2d80c3.jpg) +19.59 dB + +![](images/52f61186c13fb31a4dbc1e1e70d5a8ee66880619c966d46d9491172ddc8d846f.jpg) +Multiphoton + +![](images/900b2d151395b3d981d6078d37968f6287fcea5dbd29746f6d0859ba4c11c879.jpg) +27.22 dB +28.41 dB + +![](images/0cd7e4665127c5cefd2b44478cd620961d26a680e0c2382f33afc5589fe7bfc6.jpg) +22.54 dB + +![](images/0a0e13c9e88ce93373e11745863a9f0b481daeea4b9394bf5d1332740573f451.jpg) +21.79 dB + +![](images/b0fc083ccf464eb0bab494a1756cbe1191a4c4cad3edb2142172afcd7cd5aa30.jpg) +25.98 dB + +multiplexed arrays, high flux also leads to more ambiguous readout frames, which makes image reconstruction challenging. We therefore expect that there exists an optimal incident photon flux level where each of the estimators achieves the minimum mean-square error. The following results, as shown in Fig. 4, validate this hypothesis. + +A ground truth image can be scaled in experimental settings by selecting a desired integration time or attenuating the incident photon flux using a neutral density filter. To study the dependence of estimator performance on the incident photon flux, we perform a simulation with a high emitted photon flux from a source that is attenuated before being incident on the array. We then study the change in mean-squared error per pixel $(\mathrm{MSE} = \| \widehat{\Lambda} - \Lambda \|_2^2 / n^2)$ of the reconstructed image with varying strength of the attenuating filter. The results of this simulation are shown in Fig. 4 for the reconstruction of a $32 \times 32$ ground truth image in + +![](images/532c68806317831e870b561a4699782325bb469862effa92b5e38aea42f74030.jpg) +Figure 4. Mean-squared errors of the estimators studied as functions of mean photons per frame. (Ground truth is image in Fig. 3(c).) + +# row (c) of Fig. 3. + +Note that as the mean PPF increases, the MSEs of all the estimators initially decrease, reach a minimum, and then increase. High MSE at low PPF values is due to the lack of detected photons for accurate reconstruction. The increase in MSE at high PPF values is expected due to the high variance of SPE and the high probability of multiphoton events not modeled by the ME. It can also be seen that the minimum MSE of the multiphoton estimator is lower than that of the single-photon estimator, and both of these are much lower than the naive estimator. Furthermore, the minimum value of MSE for the ME occurs at a PPF of 1.4, which is higher than that for the SPE at 0.83, and that for NE at 0.45. We expect these incidence flux values to be realistic in conditions similar to imaging under moonlight, where high-efficiency detectors like SNSPDs are typically used. Thus, the ME produces a better reconstruction compared to both the NE and the SPE when multiple photons are detected simultaneously at the array. This results in an increased photon count rate and a better utilization of the spatial dimensions of the array. + +# 5.3. 2-, 3- and 4-Photon Estimators + +As explained in Sec. 4.4, our estimator resolves up to 4 simultaneously incident photons. The conditional probability expressions (g's) contain terms corresponding to every possible photon event that can result in the measured readout. However, we can simplify these expressions by restricting the number of modeled events. For instance, if we include only 2-photon coincidences in the case of a $2 \times 2$ array, only conditional probabilities of events $E_{9}$ and $E_{10}$ would be summed in (13). Fig. 4 shows the performance of the ME + +![](images/df831896f988a54927e707630b3dc510a5f5a3d08e013165a1b7c5ae8da31c23.jpg) +Naive +22.59 dB + +![](images/90a7dcbeb8831e0466399d52e5fdb40c87bf4774d5d2ead4e860d9340e06ba45.jpg) +Single-Photon +23.64 dB + +![](images/85129537de147b3bad9295e0a352d1c5c7ec72326b5e0c6e92cb7984a733408a.jpg) +2-photon +23.83 dB +Figure 5. Image reconstruction by operating each of the studied estimators at mean PPF values corresponding to their lowest MSE's. Indicated values are PSNR of reconstructions. + +![](images/bc699e795f461035376998688a59582f49fdf3c763c7537ad8ec5598bbb09635.jpg) +3-photon +24.60 dB + +![](images/a2d585d3963615f4cc791b4628d550342dd3927fcd4a6fb3028853de30c8b71f.jpg) +4-photon +26.74 dB + +when 2-, 3- and 4-photon are modeled in $g$ . It can be seen that the MSE of reconstruction is successively reduced with each additional photon considered in the estimator. The 2-photon estimator achieves its lowest MSE at a mean PPF of 1, while the 3- and 4-photon estimators achieve their lowest MSEs at 1.2 and 1.4 mean PPF, respectively. + +Image reconstruction using each of the estimators in Fig. 4 at their optimal mean PPF values is summarized in Fig. 5. We see that the 4-photon estimator achieves the highest PSNR, which is 3 to 4 dB higher than the NE and the SPE. + +# 5.4. Reduction in Total Integration Time + +Since the multiphoton estimator enables the accurate interpretation of photon coincidences, it can be leveraged to reduce integration times in imaging with single-photon detectors. The performance of the three estimators in achieving a target reconstruction MSE is shown in Fig. 6. It can be seen that to achieve an MSE of 0.01, the ME requires 25,000 measured frames, and the SPE requires 100,000 measured frames, while the NE does not achieve this MSE within the range of frame values studied. This shows that the ME achieves a factor $\sim 4$ reduction in the integration time for the same reconstruction quality as the other estimators. In comparison, a fully-addressed readout scheme achieves the target MSE with only 5000 frames, but with the increased complexity of readout. + +![](images/7d1906ced67ab22223014d577a43c985cfd0395d033c73bb214d5b1ace7985df.jpg) +Figure 6. Reduction in the required number of frames to achieve the same target MSE by the NE, SPE,ME, and a fully-addressed readout scheme. + +# 5.5. Cramér-Rao Bound + +The Cramér-Rao bound (CRB) provides the lowest possible variance of an unbiased estimator for parameter estimation. Here, we compare the performance of the NE, SPE, 2-photon, 3-photon, and 4-photon estimators on the estimation of a $2 \times 2$ ground truth, against CRB values at different incident photon fluxes. We first compute the Fisher information matrix (FIM) using the likelihood expression for row-column readouts from a $2 \times 2$ array. The complete derivation of expressions for elements of the FIM are shown in the supplement. The element $(1,1)$ of the FIM for a single measurement is + +$$ +\begin{array}{l} \mathcal {I} _ {1 1} = \frac {(1 - q _ {4}) (1 - q _ {2} q _ {3})}{q _ {1}} + \frac {(1 - q _ {2} q _ {3})}{1 - q _ {1}} \\ + \frac {q _ {4} ^ {2} \left(1 - q _ {2} q _ {3}\right) ^ {2}}{q _ {1} q _ {4} + q _ {2} q _ {3} - q _ {1} q _ {2} q _ {3} q _ {4}}. \tag {16} \\ \end{array} +$$ + +Other diagonal elements have similar forms as shown in the supplement. The CRB is then computed as the mean of the diagonal elements in $\mathcal{I}^{-1}$ . The variation of CRB values, as well as MSEs of the studied estimators as a function of incident photon flux, is shown in Fig. 7. + +It can be seen that the ME curve matches the CRB curve for a range of mean PPF values. At high fluxes, the MSE of the ME increases due to the high variance of the single-photon estimates, which are used to calculate $g$ in (13). In the general case with bright spots at all four pixels of the array, the 4-photon estimator most closely matches the CRB across the range of PPF values studied. The 2- and 3-photon estimators deviate from the CRB due to the modeling of + +![](images/9adc2f33e01880b126796b9a8a0089858f74bddfcd55e80e8c50eaf6cd623f9e.jpg) +Figure 7. Comparison of estimator performance with the Cramér-Rao Bound. The ME matches the CRB for a range of incident photon fluxes and deviates from it at high fluxes. With 4 bright spots, model mismatch causes 2- and 3-photon estimators to become biased and deviate from the CRB at lower incident photon fluxes compared to the 4-photon estimator. + +only a subset of events that can cause ambiguities and omission of terms in the expression for conditional probabilities in (13). The dependence of the estimator performance on the ground truth image is shown in the Supplement. + +# 5.6. Increasing array resolution + +Our analysis of the multiphoton estimator's performance so far has been limited to $32 \times 32$ arrays to match the sensor developed in [37]. Here, we study its use in arrays with higher pixel counts and finer spatial resolutions. We emulate a fixed sensor area with a varying number of pixels and fixed maximum photon flux. This results in photons per pixel being inversely proportional to the number of pixels. We include attenuation optimization for each reconstruction method, similar to Fig. 4. + +Results for row 1 of Fig. 3 as the ground truth are shown in Fig. 8 Performance improvement from the ME increases with array size (see Fig. 8(a)). This shows that our method is applicable to array sizes encountered in commercial imaging use cases. Nevertheless, we expect coincidences not modeled to be a limiting factor. Further, computation time increases linearly with an increase in array size. Thus, for a practical implementation of our algorithm, it could be beneficial to divide the sensor area into smaller blocks, the readouts from each of which can be processed using the ME. Such a processing approach would be particularly beneficial in sensor architectures similar to [17]. + +![](images/985c34be740c8b7943127fc80199ba14d2221c3d731d53f413155aff16ffd5cd.jpg) + +![](images/41357b09a5d95f2bb815f4f53d986396596d07a0ed67d0c6a502b2263b31e885.jpg) +Figure 8. Influence of array size. (a) MSEs at optimum attenuation for each estimator. (b) Optimum photons per frame for each estimator. + +# 6. Conclusions and Future Works + +We formulate multiphoton coincidence resolution in readout multiplexed single-photon detector arrays as an inverse imaging problem and propose a novel estimator to resolve up to four photon coincidences with row-column multiplexed readouts. Our method enables high-flux, low-latency, high-resolution image reconstructions that significantly improve upon conventional methods. Future work can extend this method to resolve an abitrarily large number of coincident photon events or other types of multiplexed readouts. The use of spatial priors and deep learning methods can be explored to improve the presented solutions. This line of work could be especially useful for the use of SNSPD arrays in photon-starved imaging applications like deep space imaging or biological imaging. + +# Acknowledgements + +This work was supported in part by the U.S. National Science Foundation under Grant 2039762 and in part by a gift from Dr. John Zheng Sun. We thank Prof. Hoover Rueda-Chacon for helpful discussions. + +# References + +[1] M. S. Allman, V. B. Verma, M. Stevens, T. Gerrits, R. D. Horansky, A. E. Lita, F. Marsili, A. Beyer, M. D. Shaw, D. Kumor, R. Mirin, and S. W. Nam. A near-infrared 64-pixel superconducting nanowire single photon detector array with integrated multiplexed readout. Applied Physics Letters, 106 (19), 2015. 1 +[2] Ibrahim Ben Daya. Systematic characterization framework of row-column ultrasound imaging systems. Master's thesis, University of Waterloo, 2015. 2 +[3] PL Bonanno. Physical and cost comparison of smartphone laser autofocus solutions. In Proc. Int. SPAD Sensor Workshop, 2024. 1 +[4] Daniel AB Bonifacio and Maurizio Moralles. Modeling of 3d gamma interaction position in a monolithic scintillator block with a row-column summing readout. In 2012 IEEE Nuclear Science Symposium and Medical Imaging Conference Record (NSS/MIC), pages 2606-2613. IEEE, 2012. 2 +[5] Jihwan Boo, Mark D Hammig, and Manhee Jeong. Row-column readout method to mitigate radiographic-image blurring from multipixel events in a coded-aperture imaging system. IEEE Transactions on Nuclear Science, 68(5):1175-1183, 2021. 2 +[6] Caroline K Brennan, Zi Yao, Anastasia A Ionkina, Colin M Rathbun, Buvaneshwari Sathishkumar, and Jennifer A Prescher. Multiplexed bioluminescence imaging with a substrate unmixing platform. Cell Chemical Biology, 29(11): 1649-1660, 2022. 2 +[7] Claudio Bruschini, Harald Homulle, Ivan Michel Antolovic, Samuel Burri, and Edoardo Charbon. Single-photon avalanche diode imagers in biophotonics: review and outlook. Light: Science & Applications, 8(1):87, 2019. 1 +[8] Nicolas Callens and Georges GE Gielen. Analysis and comparison of readout architectures and analog-to-digital converters for 3d-stacked cmos image sensors. IEEE Transactions on Circuits and Systems I: Regular Papers, 68(8):3117-3130, 2021. 2 +[9] Francesco Ceccarelli, Giulia Acconcia, Angelo Gulinatti, Massimo Ghioni, Ivan Rech, and Roberto Osellame. Recent advances and future perspectives of single-photon avalanche diodes for quantum photonics applications. Advanced Quantum Technologies, 4(2):2000102, 2021. 1 +[10] Garry Chinn, Peter D. Olcott, and Craig S. Levin. Sparse signal recovery methods for multiplexing pet detector readout. IEEE Transactions on Medical Imaging, 32(5):932-942, 2013. 2 +[11] Iris Cusini, Davide Berretta, Enrico Conca, Alfonso Incoronato, Francesca Madonini, Arianna Adelaide Maurina, + +Chiara Nonne, Simone Riccardo, and Federica Villa. Historical perspectives, state of art and research trends of SPAD arrays and their applications (Part II: SPAD arrays). Frontiers in Physics, 10:906671, 2022. 1 +[12] Iman Esmaeil Zadeh, J Chang, Johannes WN Los, Samuel Gyger, Ali W Elshaari, Stephan Steinhauer, Sander N Dorenbos, and Val Zwiller. Superconducting nanowire single-photon detectors: A perspective on evolution, state-of-the-art, future developments, and applications. Applied Physics Letters, 118(19), 2021. 1 +[13] Andrew L Goertzen, Xuezhu Zhang, Megan M McClarty, Eric J Berg, Chen-Yi Liu, Piotr Kozlowski, Fabrice Retiere, Lawrence Ryner, Vesna Sossi, Greg Stortz, et al. Design and performance of a resistor multiplexing readout circuit for a sipm detector. IEEE Transactions on Nuclear Science, 60 (3):1541-1549, 2013. 2 +[14] Rémi Gribonval, Antoine Chatalic, Nicolas Keriven, Vincent Schellekens, Laurent Jacques, and Philip Schniter. Sketching data sets for large-scale learning. IEEE Signal Process. Mag., 38(5):12-36, 2021. 1 +[15] Yanqiu Guan, Haochen Li, Li Xue, Rui Yin, Labao Zhang, Hao Wang, Guanghao Zhu, Lin Kang, Jian Chen, and Peiheng Wu. Lidar with superconducting nanowire single-photon detectors: Recent advances and developments. Optics and Lasers in Engineering, 156:107102, 2022. 1 +[16] Yanqiu Guan, Haochen Li, Labao Zhang, Hao Wang, Guanglong He, Biao Zhang, Yue Fei, Jiayu Lv, Xiao Zhang, Rui Yin, et al. SNSPD array with single-channel readout based on compressive sensing. ACS Photonics, 9(9):3102-3109, 2022. 2 +[17] Hao Hao, Qing-Yuan Zhao, Yang-Hui Huang, Jie Deng, Fan Yang, Sai-Ying Ru, Zhen Liu, Chao Wan, Hao Liu, Zhi-Jian Li, et al. A compact multi-pixel superconducting nanowire single-photon detector array supporting gigabit space-to-ground communications. Light: Science & Applications, 13(1):25, 2024. 8 +[18] Sam W. Hutchings, Nick Johnston, Istvan Gyongy, Tarek Al Abbas, Neale A. W. Dutton, Max Tyler, Susan Chan, Jonathan Leach, and Robert K. Henderson. A reconfigurable 3-D-stacked SPAD imager with in-pixel histogramming for flash LIDAR or high-speed time-of-flight imaging. IEEE J. Solid-State Circuits, 54(11):2947-2956, 2019. 1 +[19] Atul Ingle and David Maier. Count-free single-photon 3d imaging with race logic. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-12, 2023. 1 +[20] Jørgen Arendt Jensen, Mikkel Schou, Lasse Thurmann Jørgensen, Borislav G Tomov, Matthias Bo Stuart, Marie Sand Traberg, Iman Taghavi, Sigrid Huesebø Øygaard, Martin Lind Ommen, Kitty Steenberg, et al. Anatomic and functional imaging using row-column arrays. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 69(10):2722-2738, 2022. 2 +[21] Ling-Dong Kong, Hui Wang, Qing-Yuan Zhao, Jia-Wei Guo, Yang-Hui Huang, Hao Hao, Shi Chen, Xue-Cou Tu, LaBao Zhang, Xiao-Qing Jia, et al. Readout-efficient superconducting nanowire single-photon imager with orthogonal time-amplitude multiplexing by hotspot quantization. Nature Photonics, 17(1):65-72, 2023. 2 + +[22] Sangwon Lee, Yong Choi, Jihoon Kang, and Jin Ho Jung. Development of a multiplexed readout with high position resolution for positron emission tomography. *Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment*, 850:42-47, 2017. 2 +[23] Feiyan Li, Hang Han, Qi Chen, Biao Zhang, Han Bao, Yue Dai, Rui Ge, Shuya Guo, Guanglong He, Yue Fei, et al. Saturation efficiency for detecting $1550\mathrm{nm}$ photons with a $2\times 2$ array of $\mathrm{Mo}_{0.8}\mathrm{Si}_{0.2}$ nanowires at $2.2\mathrm{K}$ . Photonics Research, 9(3):389-394, 2021. 3 +[24] Adam N McCaughan, Yao Zhai, Boris Korzh, Jason P Allmaras, Bakhrom G Oripov, Matthew D Shaw, and Sae Woo Nam. The thermally coupled imager: A scalable readout architecture for superconducting nanowire single photon detectors. Applied Physics Letters, 121(10), 2022. 1, 2 +[25] Kazuhiro Morimoto. Megapixel spad cameras for time-resolved applications. Technical report, EPFL, 2021. 1 +[26] Peter D Olcott, Garry Chinn, and Craig S Levin. Compressed sensing for the multiplexing of pet detectors. In 2011 IEEE Nuclear Science Symposium Conference Record, pages 3224-3226. IEEE, 2011. 2 +[27] B. G. Oripov, D. S. Rampini, J. Allmaras, M. D. Shaw, S. W. Nam, B. Korzh, and A. N. McCaughan. A superconducting nanowire single-photon camera with 400,000 pixels. Nature, 622(7984):730-734, 2023. 2 +[28] François Piron, Daniel Morrison, Mehmet Rasit Yuce, and Jean-Michel Redoute. A review of single-photon avalanche diode time-of-flight imaging sensor arrays. IEEE Sensors Journal, 21(11):12654-12666, 2020. 1 +[29] Joshua Rapp, Julian Tachella, Yoann Altmann, Stephen McLaughlin, and Vivek K Goyal. Advances in single-photon lidar for autonomous vehicles: Working principles, challenges, and recent advances. IEEE Signal Processing Magazine, 37(4):62-71, 2020. 1 +[30] Morten Fischer Rasmussen and Jørgen Arendt Jensen. 3-d ultrasound imaging performance of a row-column addressed 2-d array transducer: A measurement study. In 2013 IEEE International Ultrasonics Symposium (IUS), pages 1460-1463. IEEE, 2013. 2 +[31] Morten Fischer Rasmussen, Thomas Lehrmann Christiansen, Erik Vilain Thomsen, and Jørgen Arendt Jensen. 3-d imaging using row-column-addressed arrays with integrated apodization-part i: apodization design and line element beamforming. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 62(5):947-958, 2015. 2 +[32] Kaustubh Sadekar, David Maier, and Atul Ingle. Single-photon 3d imaging with equi-depth photon histograms. In European Conference on Computer Vision, pages 381-398. Springer, 2024. 1 +[33] Ewout van den Berg, Emmanuel Candes, Garry Chinn, Craig Levin, Peter Demetri Olcott, and Carlos Sing-Long. Single-photon sampling architecture for solid-state imaging sensors. Proceedings of the National Academy of Sciences, 110(30): E2752-E2761, 2013. 2 +[34] Hailu Wang, Jiaxiang Guo, Jinshui Miao, Wenjin Luo, Yue Gu, Runzhang Xie, Fang Wang, Lili Zhang, Peng Wang, and + +Weida Hu. Emerging single-photon detectors based on low-dimensional materials. Small, 18(5):2103963, 2022. 1 +[35] Hui Wang, Zhi-jian Li, Xue-Mei Hu, Hao Hao, Jia-wei Guo, Yang-hui Huang, Hao Liu, Chao Wan, Xue-cou Tu, Xiaoqing Jia, et al. Image distortion by ambiguous multiple-photon detections in a superconducting nanowire single-photon imager and the correction method. Optics Express, 31(14):23579-23588, 2023. 2 +[36] Mel White, Shahaboddin Ghajari, Tianyi Zhang, Akshit Dave, Ashok Veeraraghavan, and Alyosha Molnar. A differential SPAD array architecture in $0.18\mu \mathrm{m}$ CMOS for HDR imaging. In Proc. IEEE Int. Symp. Circuits Syst., pages 292-296, 2022. 1 +[37] Emma E Wollman, Varun B Verma, Adriana E Lita, William H Farr, Matthew D Shaw, Richard P Mirin, and Sae Woo Nam. Kilopixel array of superconducting nanowire single-photon detectors. Optics express, 27(24):35279-35289, 2019. 1, 2, 8 +[38] Lixing You. Superconducting nanowire single-photon detectors for quantum information. Nanophotonics, 9(9):2673-2692, 2020. 1 +[39] Qing-Yuan Zhao, Di Zhu, Niccolò Calandri, Andrew E Dane, Adam N McCaughan, Francesco Bellei, Hao-Zhu Wang, Daniel F Santavicca, and Karl K Berggren. Single-photon imager based on a superconducting nanowire delay line. Nature Photonics, 11(4):247-251, 2017. 2 \ No newline at end of file diff --git a/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/images.zip b/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6f7a961a10cea258ae9fcbc9ab3e691c56612083 --- /dev/null +++ b/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec6a5830e08f2eddc7d3622e8a644dcd3e8a43afe38e9b0677876cf36ab13e17 +size 475352 diff --git a/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/layout.json b/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..17836e94b6ed522d8b31e2dc68f9539d64db1380 --- /dev/null +++ b/2025/Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays/layout.json @@ -0,0 +1,10351 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "type": "text", + "content": "Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 143, + 512, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 143, + 512, + 171 + ], + "spans": [ + { + "bbox": [ + 97, + 143, + 512, + 171 + ], + "type": "text", + "content": "Shashwath Bharadwaj, Ruangrawee Kitichotkul, Akshay Agarwal and Vivek K Goyal Boston University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 270, + 174, + 337, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 174, + 337, + 184 + ], + "spans": [ + { + "bbox": [ + 270, + 174, + 337, + 184 + ], + "type": "text", + "content": "shash@bu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 239, + 296, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 549 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 549 + ], + "type": "text", + "content": "Readout multiplexing is a promising solution to overcome hardware limitations and data bottlenecks in imaging with single-photon detectors. Conventional multiplexed readout processing creates an upper bound on photon counts at a very fine time scale, where frames with multiple detected photons must either be discarded or allowed to introduce significant bias. We formulate multiphoton coincidence resolution as an inverse imaging problem and introduce a solution framework to probabilistically resolve the spatial locations of photon incidences. Specifically, we develop a theoretical abstraction of row-column multiplexing and a model of photon events that make readouts ambiguous. Using this, we propose a novel estimator that spatially resolves up to four coincident photons. Monte Carlo simulations show that our proposed method increases the peak signal-to-noise ratio (PSNR) of reconstruction by 3 to 4 dB compared to conventional methods under optimal incident flux conditions. Additionally, this method reduces the required number of readout frames to achieve the same mean-squared error as other methods by a factor of " + }, + { + "bbox": [ + 56, + 239, + 296, + 549 + ], + "type": "inline_equation", + "content": "\\sim 4" + }, + { + "bbox": [ + 56, + 239, + 296, + 549 + ], + "type": "text", + "content": ". Finally, our method matches the Cramér-Rao bound for detection probability estimation for a wider range of incident flux values compared to conventional methods. While demonstrated for a specific detector type and readout architecture, this method can be extended to more general multiplexing with different detector models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 573, + 135, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 573, + 135, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 573, + 135, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Single-photon detectors have been widely studied for use in applications like biological imaging [7, 34], lidar [15, 29], and quantum optics [9] due to properties such as single-photon sensitivity and high-precision time-tagging of photon arrivals. Single-photon avalanche diodes (SPADs) have been of particular interest due to compatibility with CMOS manufacturing techniques and room temperature operation [11]. However, they have low detection efficiencies (especially at mid infrared wavelengths), which means that only a small fraction of incident light is detected as pho" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 214, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 299 + ], + "type": "text", + "content": "tons. Superconducting nanowire single-photon detectors (SNSPDs) are emerging devices that offer near-unity quantum efficiency across a range of wavelengths, low dark counts, and low jitter, making them compelling alternatives to SPADs [38]. However, the requirement of cryogenic cooling to maintain superconductivity has largely restricted their use to laboratory experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 302, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 302, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 312, + 302, + 555, + 506 + ], + "type": "text", + "content": "Single-photon imaging in typical commercial applications requires arrays of single-photon detectors [28]. While megapixel SPAD arrays have already been deployed in mass consumer products [3, 25], high data transfer rates from these arrays for further processing remain a bottleneck. For instance, typical room-scale ranging experiments with SPAD array-based single-photon lidar can have data generation rates of the order of gigabits per second when the array resolution approaches the megapixel scale. It is attractive to compress the readout to less than a stream of all digitized detection times. Methods like in-pixel histogramming [18], histogram compression through sketching [14], equidepth histogramming [19, 32], and differential readouts [36] have been proposed to address high data transfer rates. These approaches to avoid the cost of high temporal resolution are complementary to our emphasis on transverse spatial dimensions in this work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 556, + 714 + ], + "type": "text", + "content": "Implementing high spatial resolution SNSPD arrays is a significant challenge. As the number of pixels in the array grows, the heat load introduced by reading out each pixel individually becomes incompatible with the cryogenics of the system [12]. To mitigate this, recent efforts have focused on addressing groups of pixels in the readout, starting with Allman et al. [1] who introduced row-column readout multiplexing. In this mechanism, readouts indicate photon detection at each row and each column of the array instead of every pixel, which reduces the required number of readout lines for an " + }, + { + "bbox": [ + 312, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 312, + 510, + 556, + 714 + ], + "type": "text", + "content": " array from " + }, + { + "bbox": [ + 312, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "n^2" + }, + { + "bbox": [ + 312, + 510, + 556, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 312, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "2n" + }, + { + "bbox": [ + 312, + 510, + 556, + 714 + ], + "type": "text", + "content": ". Wollman et al. [37] extended this scheme to demonstrate the first kilopixel SNSPD array. McCaughan et al. [24] introduced the thermally coupled imager with time-of-flight multiplexing, where separate readout lines for each row and column are replaced with a single bus for all rows and all columns. This makes the required number of readouts independent of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11406" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 120 + ], + "type": "text", + "content": "the array size. Recently, Oripov et al. [27] leveraged this method to demonstrate a 400,000 pixel superconducting camera, showing the viability of this scheme to approach the sizes of commercial imaging arrays." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 121, + 294, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 121, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 57, + 121, + 294, + 370 + ], + "type": "text", + "content": "Despite the gain in scaling array sizes, readout multiplexing requires the incident photon flux to be very low for unambiguous signal reconstruction, limiting the array's photon count rate. When two or more photons are incident on the array within the period of a single acquisition, row-column readouts can give an ambiguous set of candidate locations for their incidences. Traditional reconstruction techniques [37] do not distinguish between these locations and assume equal probabilities of detection at each candidate pixel. This results in the misattribution of photon counts to locations where no photon incidences occurred. Alternatively, since readouts with a single detected photon provide the exact row and column indices where the photon was incident on the array, ambiguous readouts with multiple detections can be discarded for signal reconstruction. However, this leads to high variance in the reconstruction (especially at high incident fluxes) and underutilization of the spatial dimensions of the array by requiring most readouts to contain detections at a single pixel. Resolving the spatial locations of multiphoton coincidences using multiplexed readouts is an ill-posed inverse problem." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 372, + 294, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 372, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 372, + 294, + 658 + ], + "type": "text", + "content": "Here, we propose a solution framework to leverage most of the measured data to reduce the mean-squared error of image reconstruction. Specifically, we develop a multiphoton estimator that redistributes photon counts from ambiguous readouts to candidate pixels such that an approximate likelihood of observations is maximized. A key choice for our algorithm design is the omission of spatial priors in order to not obscure the source of performance gain. Spatial priors can be used in conjunction with the developed method for further improvements. Monte Carlo simulations show that our proposed method increases the peak signal-to-noise ratio (PSNR) of reconstruction by 3 to " + }, + { + "bbox": [ + 57, + 372, + 294, + 658 + ], + "type": "inline_equation", + "content": "4\\mathrm{dB}" + }, + { + "bbox": [ + 57, + 372, + 294, + 658 + ], + "type": "text", + "content": " compared to conventional methods under optimal incident flux conditions. Further, the optimal incident photon flux for our method is " + }, + { + "bbox": [ + 57, + 372, + 294, + 658 + ], + "type": "inline_equation", + "content": "\\sim 1.4" + }, + { + "bbox": [ + 57, + 372, + 294, + 658 + ], + "type": "text", + "content": " photons per readout which is 0.5 to 1.1 photons per readout higher than conventional methods. We also demonstrate that our estimator matches the Cramér-Rao Bound (CRB) for the estimation of photon detection probabilities at each pixel of an array across a range of incident photon flux values. Finally, our method achieves a factor of " + }, + { + "bbox": [ + 57, + 372, + 294, + 658 + ], + "type": "inline_equation", + "content": "\\sim 4" + }, + { + "bbox": [ + 57, + 372, + 294, + 658 + ], + "type": "text", + "content": " reduction in the required number of readout frames for the same mean-squared error of image reconstruction. These results demonstrate that our proposed estimator is well-suited for high-flux, low-latency imaging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 670, + 141, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 670, + 141, + 681 + ], + "spans": [ + { + "bbox": [ + 58, + 670, + 141, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 691, + 294, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 691, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 58, + 691, + 294, + 700 + ], + "type": "text", + "content": "Image reconstruction from multiplexed readouts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 703, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 703, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 703, + 294, + 712 + ], + "type": "text", + "content": "Compressed sensing methods have been explored for the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 72, + 553, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 394 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 394 + ], + "type": "text", + "content": "design of multiplexed readout architectures in areas like positron emission tomography (PET) [10, 13, 22, 26], biological imaging [6, 30] and gamma-ray imaging [4, 5]. While readout multiplexing for specific imaging applications has been extensively studied [8, 20], the focus of these works is on the modeling of optical and electrical properties of the imaging system such as the shape of the wavefront of the detected light [2, 31], coded apertures for rejecting multipixel events [5] and IC design for resistor multiplexing [13]. The explicit study of individual photon events measured by single-photon detectors and their impact on the multiplexed readouts does not yet exist. Our work introduces a theoretical study of the readouts based on the combinatorics of photon events and an approximate maximum likelihood estimation of photon detection probabilities at each pixel. We also analyze the performance of a novel multiphoton estimator with variations in imaging parameters. Our work is closest to the analysis presented by van den Berg et al. [33], where group-testing inspired surface codes are studied for deterministic multiphoton coincidence resolution in time-to-digital converter readouts from SPAD arrays. However, our work specifically addresses intensity imaging with SNSPD arrays where a small probability of error is tolerable. We provide a probabilistic model of the readouts and solutions for image reconstruction from a specific device architecture for up to four coincident photons." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 396, + 553, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 396, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 553, + 647 + ], + "type": "text", + "content": "Imaging with SNSPD arrays. While large-scale SNSPD array development is still in the nascent stages, several works have conducted proof-of-principle imaging experiments with prototypes. Zhao et al. introduced a delay-line-based superconducting imager with an effective pixel count of " + }, + { + "bbox": [ + 316, + 396, + 553, + 647 + ], + "type": "inline_equation", + "content": "\\sim 590" + }, + { + "bbox": [ + 316, + 396, + 553, + 647 + ], + "type": "text", + "content": " [39]. This method inspired the thermally coupled imager [24], which in turn resulted in the 400,000-pixel camera by Oripov et al. [27]. Subsequent efforts to use SNSPD arrays for imaging [16, 21, 35] involve intricate hardware control such as bias-current sweeping, use of idle pixels, and geometric design of the nanowire structure. Our solution is purely computational and does not require additional hardware modifications. In addition, the modeling of incident photon events and their multiplexed readouts can be extended to several detector models in high-energy physics or particle detection. While the method described in this work provides a framework to resolve an arbitrarily large number of coincident photons, we limit the discussion in the next sections to up to four coincident photons to make the demonstrated solutions computationally tractable and easier to illustrate and analyze." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 658, + 503, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 503, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 503, + 669 + ], + "type": "text", + "content": "3. Measurement and Readout Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 678, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 678, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 678, + 553, + 712 + ], + "type": "text", + "content": "Figure 1 illustrates imaging a scene using an " + }, + { + "bbox": [ + 316, + 678, + 553, + 712 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 316, + 678, + 553, + 712 + ], + "type": "text", + "content": " SNSPD array with row-column readouts. Measurements are read out after integrating the incident signal for a fixed time in" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "11407" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 83, + 558, + 262 + ], + "blocks": [ + { + "bbox": [ + 59, + 83, + 558, + 262 + ], + "lines": [ + { + "bbox": [ + 59, + 83, + 558, + 262 + ], + "spans": [ + { + "bbox": [ + 59, + 83, + 558, + 262 + ], + "type": "image", + "image_path": "52a45b326e446bd427013d7350541df707f25a8777cfe0c29caa2082ec7d5307.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 274, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 54, + 274, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 555, + 309 + ], + "type": "text", + "content": "Figure 1. Schematic of image reconstruction with row-column readouts. (Left) Multiple frames of a ground truth image are measured using an SNSPD array to give row-column readout vectors. (Right) Comparison of conventional and proposed reconstruction techniques to estimate the ground truth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 329, + 294, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 294, + 376 + ], + "type": "text", + "content": "terval. Let " + }, + { + "bbox": [ + 54, + 329, + 294, + 376 + ], + "type": "inline_equation", + "content": "\\Lambda \\in \\mathbb{R}^{n\\times n}" + }, + { + "bbox": [ + 54, + 329, + 294, + 376 + ], + "type": "text", + "content": " be the flux of an " + }, + { + "bbox": [ + 54, + 329, + 294, + 376 + ], + "type": "inline_equation", + "content": "n\\times n" + }, + { + "bbox": [ + 54, + 329, + 294, + 376 + ], + "type": "text", + "content": " ground truth image in that integration time. The number of photons arriving at pixel " + }, + { + "bbox": [ + 54, + 329, + 294, + 376 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 54, + 329, + 294, + 376 + ], + "type": "text", + "content": " at a discrete time index " + }, + { + "bbox": [ + 54, + 329, + 294, + 376 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 329, + 294, + 376 + ], + "type": "text", + "content": " is modeled as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 376, + 295, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 295, + 392 + ], + "type": "interline_equation", + "content": "X _ {i j} ^ {t} \\sim \\operatorname {P o i s s o n} (\\Lambda_ {i j}). \\tag {1}", + "image_path": "4c0225527aa470da81d3f127bbaa7dadadeb107f38d60657eba564db4e41ed04.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "content": "We assume pixel saturation with the detection of a single photon, i.e., the detectors are not photon-number resolving [23]. Thus, we define binary-valued photon-incidence indicator " + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "inline_equation", + "content": "Y^{t} \\in \\{0,1\\}^{n \\times n}" + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "inline_equation", + "content": "Y_{ij}^{t} = \\mathbf{1}[X_{ij}^{t} > 0]" + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{1}[\\cdot]" + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "content": " is the indicator function. Then, " + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "inline_equation", + "content": "Y^{t}" + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "content": " is a matrix where each entry is an independent Bernoulli random variable " + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "inline_equation", + "content": "Y_{ij}^{t} \\sim \\mathrm{Bernoulli}(q_{ij})" + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 488, + 295, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 488, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 140, + 488, + 295, + 504 + ], + "type": "interline_equation", + "content": "q _ {i j} = 1 - e ^ {- \\Lambda_ {i j}} \\tag {2}", + "image_path": "1d0175eadb9e610edbc0245e51d272cb59955605eb7b66d3c5bbd6b4a72eeef1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 263, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 263, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 263, + 522 + ], + "type": "text", + "content": "is the probability of photon detection at pixel " + }, + { + "bbox": [ + 55, + 510, + 263, + 522 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 510, + 263, + 522 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 548 + ], + "type": "text", + "content": "Each measurement consists of a row readout " + }, + { + "bbox": [ + 55, + 522, + 296, + 548 + ], + "type": "inline_equation", + "content": "R^t \\in \\{0,1\\}^n" + }, + { + "bbox": [ + 55, + 522, + 296, + 548 + ], + "type": "text", + "content": " and a column readout " + }, + { + "bbox": [ + 55, + 522, + 296, + 548 + ], + "type": "inline_equation", + "content": "C^t \\in \\{0,1\\}^n" + }, + { + "bbox": [ + 55, + 522, + 296, + 548 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 555, + 295, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 555, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 88, + 555, + 295, + 570 + ], + "type": "interline_equation", + "content": "R _ {i} ^ {t} = \\mathbf {1} \\left[ X _ {i j} ^ {t} > 0 \\text {f o r a n y} j \\in \\{1, \\dots , n \\} \\right], \\tag {3a}", + "image_path": "025aa01f9268a7ad4790a91257d07b255626bb147b4b65a8613511760a340184.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 571, + 295, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 571, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 89, + 571, + 295, + 586 + ], + "type": "interline_equation", + "content": "C _ {j} ^ {t} = \\mathbf {1} \\left[ X _ {i j} ^ {t} > 0 \\text {f o r a n y} i \\in \\{1, \\dots , n \\} \\right]. \\tag {3b}", + "image_path": "466a1756ab31577876a864c0e6dc819c167791f10b2381842579325ea02aaa70.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": "We then define a readout frame at time " + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "inline_equation", + "content": "(R^t, C^t)" + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": ". When multiple photons are incident on the array within a single integration period, the problem of resolving the spatial locations of their incidence is ill-posed. As an example, consider a " + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": " array with pixels numbered from 1 to 4 as shown in event " + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "inline_equation", + "content": "E_0" + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": " of Fig. 2. The frame ([1,1],[1,1]) could have resulted from photons detected at pixels " + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\{1,4\\}" + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\{2,3\\}" + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": " or any of: " + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\{1,2,3\\}" + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\{1,2,4\\}" + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\{1,3,4\\}" + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\{2,3,4\\}" + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\{1,2,3,4\\}" + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": ". More generally, a readout frame is unambiguous if " + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{n} R_i^t = 1" + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\sum_{j=1}^{n} C_j^t = 1" + }, + { + "bbox": [ + 55, + 593, + 296, + 716 + ], + "type": "text", + "content": "; otherwise, it is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 329, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 376 + ], + "type": "text", + "content": "ambiguous. Our problem is to estimate " + }, + { + "bbox": [ + 313, + 329, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 313, + 329, + 555, + 376 + ], + "type": "text", + "content": " from measurements in " + }, + { + "bbox": [ + 313, + 329, + 555, + 376 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 329, + 555, + 376 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 313, + 329, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\{(R^t,C^t)\\}_{t = 1}^T" + }, + { + "bbox": [ + 313, + 329, + 555, + 376 + ], + "type": "text", + "content": ". In particular, we are interested in exploiting the information contained in ambiguous readout frames." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 395, + 386, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 386, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 386, + 407 + ], + "type": "text", + "content": "4. Estimators" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 417, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 477 + ], + "type": "text", + "content": "In this section, we introduce the three estimators compared in this study, namely the naive, single-photon, and multiphoton estimators. Each is formulated first as an estimator " + }, + { + "bbox": [ + 313, + 417, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\widehat{q}_{ij}" + }, + { + "bbox": [ + 313, + 417, + 555, + 477 + ], + "type": "text", + "content": " for the probability " + }, + { + "bbox": [ + 313, + 417, + 555, + 477 + ], + "type": "inline_equation", + "content": "q_{ij}" + }, + { + "bbox": [ + 313, + 417, + 555, + 477 + ], + "type": "text", + "content": " defined in (2). The incident flux at each pixel can then be estimated elementwise as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 380, + 490, + 555, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 490, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 380, + 490, + 555, + 512 + ], + "type": "interline_equation", + "content": "\\widehat {\\Lambda} _ {i j} = - \\log \\left(1 - e ^ {- \\widehat {q} _ {i j}}\\right). \\tag {4}", + "image_path": "de49694a1d3f8d1547dfd6a5b7b1fc1b6058df535626ba0fbcbc7036aab45467.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 538, + 413, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 413, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 413, + 550 + ], + "type": "text", + "content": "4.1. Naive Estimator" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 605 + ], + "type": "text", + "content": "The naive estimator (NE) assumes that a photon is detected at each of the candidate pixels in ambiguous readout frames. Thus, the probability of detecting a photon at pixel " + }, + { + "bbox": [ + 313, + 558, + 555, + 605 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 558, + 555, + 605 + ], + "type": "text", + "content": " is estimated as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 383, + 609, + 555, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 609, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 383, + 609, + 555, + 643 + ], + "type": "interline_equation", + "content": "\\widehat {q} _ {i j, \\mathrm {n a i v e}} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} R _ {i} ^ {t} C _ {j} ^ {t}. \\tag {5}", + "image_path": "53813063829335acf906929a204b2d1b033ca9fb4f36cfd1106531f0cf76a7e9.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": "The naive estimator overestimates " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "q_{ij}" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " because it misattributes photon counts in ambiguous frames to pixels where no photon incidence occurred. The extra counts imputed by this scheme produce ghost spots in the reconstructed image, as depicted in Fig. 1." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11408" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 294, + 213 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 294, + 213 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 294, + 213 + ], + "type": "image", + "image_path": "47c47aa4a7e1833e48837c1ee4d78c86baf585f5be0b9ff544630fed751bbf23.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "type": "text", + "content": "Figure 2. Row-column readout frames for a " + }, + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "type": "text", + "content": " array. Yellow spots indicate incident photon locations. Accounting for pixel saturation, 16 photon detection events are possible with this array. Row-column readout results in 9 unambiguous events " + }, + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "type": "inline_equation", + "content": "E_0, E_1, \\ldots, E_8" + }, + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "type": "text", + "content": " and 7 events " + }, + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "type": "inline_equation", + "content": "E_9, E_{10}, \\ldots, E_{15}" + }, + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "type": "text", + "content": " that are ambiguous because they all result in " + }, + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "type": "inline_equation", + "content": "R^t = [1,1]" + }, + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "type": "inline_equation", + "content": "C^t = [1,1]" + }, + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "type": "text", + "content": ". Pixel numbers are indicated in event " + }, + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "type": "inline_equation", + "content": "E_0" + }, + { + "bbox": [ + 55, + 226, + 296, + 314 + ], + "type": "text", + "content": ", which corresponds to an empty readout." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 335, + 225, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 225, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 225, + 346 + ], + "type": "text", + "content": "The positive bias of the naive estimator is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 296, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 420 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {E} \\left[ \\widehat {q} _ {i j, \\text {n a i v e}} \\right] - q _ {i j} \\\\ = \\left(1 - q _ {i j}\\right) \\left(1 - \\prod_ {k \\in R _ {i} ^ {\\prime}} \\left(1 - q _ {k j}\\right)\\right) \\left(1 - \\prod_ {\\ell \\in C _ {j} ^ {\\prime}} \\left(1 - q _ {i \\ell}\\right)\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "6540ae6f21dd9d38075b125aefe1078ae378b55d72e2c8d632a7b57934b3b74a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "inline_equation", + "content": "R_{i}^{\\prime}" + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "text", + "content": " is the set of rows except row " + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "inline_equation", + "content": "C_j^\\prime" + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "text", + "content": " is the set of columns except column " + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "text", + "content": ". A derivation is given in the supplement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 471, + 194, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 194, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 194, + 483 + ], + "type": "text", + "content": "4.2. Single-Photon Estimator" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 489, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 296, + 548 + ], + "type": "text", + "content": "A simple way to prevent misattributions and thus reduce bias is to only use frames with a single detected photon in the reconstruction since these unambiguously give the spatial locations of photon incidence. We thus define the single-photon estimator (SPE):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 87, + 555, + 294, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 555, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 87, + 555, + 294, + 587 + ], + "type": "interline_equation", + "content": "\\widehat {q} _ {i j, \\text {s i n g l e}} = \\left\\{ \\begin{array}{l l} \\frac {N _ {i j}}{N _ {i j} + N _ {0}}, & \\text {i f} N _ {i j} + N _ {0} > 0; \\\\ 0, & \\text {i f} N _ {i j} + N _ {0} = 0, \\end{array} \\right. \\tag {7}", + "image_path": "b571be718687aa208fa4b89a8088653ea31426f86f186829896f25c7009d0d3b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 296, + 665 + ], + "type": "inline_equation", + "content": "N_{ij}" + }, + { + "bbox": [ + 55, + 594, + 296, + 665 + ], + "type": "text", + "content": " is the number of frames where a photon is detected only at pixel " + }, + { + "bbox": [ + 55, + 594, + 296, + 665 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 594, + 296, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 594, + 296, + 665 + ], + "type": "inline_equation", + "content": "N_0" + }, + { + "bbox": [ + 55, + 594, + 296, + 665 + ], + "type": "text", + "content": " is the number of frames without any detected photons across the whole array. The SPE is unbiased when " + }, + { + "bbox": [ + 55, + 594, + 296, + 665 + ], + "type": "inline_equation", + "content": "N_{i,j} + N_0 > 0" + }, + { + "bbox": [ + 55, + 594, + 296, + 665 + ], + "type": "text", + "content": ", as shown in the supplement. However, the discarding of frames with multiple detected photons results in high variance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 672, + 187, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 187, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 187, + 685 + ], + "type": "text", + "content": "4.3. Multiphoton Estimator" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Now, we develop a multiphoton estimator (ME) that uses photon counts from ambiguous readouts for reconstruction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 154 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 154 + ], + "type": "text", + "content": "For simplicity, here we derive expressions for detection probability estimates at each pixel of a " + }, + { + "bbox": [ + 313, + 72, + 554, + 154 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 72, + 554, + 154 + ], + "type": "text", + "content": " array and illustrate how this estimator maximizes an approximate likelihood of observations. The general form of the ME is presented in the supplement. If each pixel were read out individually, the likelihood of observations could be written as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 365, + 154, + 553, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 154, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 365, + 154, + 553, + 186 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\prod_ {i = 1} ^ {2} \\prod_ {j = 1} ^ {2} (1 - q _ {i j}) ^ {v _ {i j}} \\left(q _ {i j}\\right) ^ {f _ {i j}}, \\tag {8}", + "image_path": "17fbf3af67a5b00cc0aadadd0f4d01da9a1899a7b179c7d17d209cfa90528036.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 190, + 554, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 554, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 554, + 241 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 190, + 554, + 241 + ], + "type": "inline_equation", + "content": "v_{ij}" + }, + { + "bbox": [ + 313, + 190, + 554, + 241 + ], + "type": "text", + "content": " is the total number of frames with no detections at pixel " + }, + { + "bbox": [ + 313, + 190, + 554, + 241 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 190, + 554, + 241 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 190, + 554, + 241 + ], + "type": "inline_equation", + "content": "f_{ij}" + }, + { + "bbox": [ + 313, + 190, + 554, + 241 + ], + "type": "text", + "content": " is the total number of frames with a detection at pixel " + }, + { + "bbox": [ + 313, + 190, + 554, + 241 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 190, + 554, + 241 + ], + "type": "text", + "content": ". Then, the maximum likelihood estimate " + }, + { + "bbox": [ + 313, + 190, + 554, + 241 + ], + "type": "inline_equation", + "content": "\\widehat{q}_{ij}^{\\mathrm{ML}}" + }, + { + "bbox": [ + 313, + 190, + 554, + 241 + ], + "type": "text", + "content": " could be computed as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 249, + 553, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 249, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 361, + 249, + 553, + 273 + ], + "type": "interline_equation", + "content": "\\hat {q} _ {i j} ^ {\\mathrm {M L}} = \\frac {f _ {i j}}{T}, \\quad i = 1, 2, j = 1, 2, \\tag {9}", + "image_path": "6ce25b04b8ad494dfcaa64bcb086eae7b7a9a989b46334c8465af64302cb16be.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 280, + 511, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 511, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 511, + 291 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 280, + 511, + 291 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 280, + 511, + 291 + ], + "type": "text", + "content": " is the total number of measured frames." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 293, + 554, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 554, + 338 + ], + "type": "text", + "content": "With row-column readouts, our knowledge of locations of photon detections is reduced. We define " + }, + { + "bbox": [ + 313, + 293, + 554, + 338 + ], + "type": "inline_equation", + "content": "v_{ij}" + }, + { + "bbox": [ + 313, + 293, + 554, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 293, + 554, + 338 + ], + "type": "inline_equation", + "content": "f_{ij}" + }, + { + "bbox": [ + 313, + 293, + 554, + 338 + ], + "type": "text", + "content": " from unambiguous frames only and find that the likelihood of observations becomes" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 363, + 348, + 553, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 348, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 363, + 348, + 553, + 363 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\left\\{\\left(R ^ {t}, C ^ {t}\\right) \\right\\} _ {t = 1} ^ {T}; q\\right) = U (q) A (q), \\tag {10}", + "image_path": "4d48da5617f982df2351ecfe79ec672b325f0cbfdbf0d65f550427274ed59e5a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 372, + 376, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 372, + 376, + 382 + ], + "spans": [ + { + "bbox": [ + 314, + 372, + 376, + 382 + ], + "type": "text", + "content": "with the factor" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 362, + 391, + 553, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 391, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 362, + 391, + 553, + 423 + ], + "type": "interline_equation", + "content": "U (q) = \\prod_ {i = 1} ^ {2} \\prod_ {j = 1} ^ {2} \\left(1 - q _ {i j}\\right) ^ {v _ {i j}} \\left(q _ {i j}\\right) ^ {f _ {i j}} \\tag {11}", + "image_path": "52cb806e2f2f30f8424be5c45badfad9ec02e6682f1c017eb5cba1f5e7869716.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 433, + 480, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 480, + 444 + ], + "type": "text", + "content": "from unambiguous frames and the factor" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 334, + 453, + 553, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 453, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 334, + 453, + 553, + 468 + ], + "type": "interline_equation", + "content": "A (q) = \\left(q _ {1 1} q _ {2 2} + q _ {1 2} q _ {2 1} - q _ {1 1} q _ {1 2} q _ {2 1} q _ {2 2}\\right) ^ {M _ {9}} \\tag {12}", + "image_path": "3e069e38bd4cfe1caf75d437c83c2aaba4192a7f91cd473a91445f3f0dcfcc2f.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 476, + 554, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 554, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 554, + 512 + ], + "type": "text", + "content": "from ambiguous frames. Here, " + }, + { + "bbox": [ + 313, + 476, + 554, + 512 + ], + "type": "inline_equation", + "content": "M_9" + }, + { + "bbox": [ + 313, + 476, + 554, + 512 + ], + "type": "text", + "content": " is the total number of ambiguous readout frames due to events " + }, + { + "bbox": [ + 313, + 476, + 554, + 512 + ], + "type": "inline_equation", + "content": "E_9,\\ldots ,E_{15}" + }, + { + "bbox": [ + 313, + 476, + 554, + 512 + ], + "type": "text", + "content": " shown in Fig. 2." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "text", + "content": "The factor in (12) results from the sum of probabilities " + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "inline_equation", + "content": "\\sum_{k=9}^{15} \\mathbb{P}(E_k)" + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "text", + "content": ". This makes it challenging to find analytical expressions for the maximum likelihood estimates of detection probabilities. However, if we can estimate the fractions of the " + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "inline_equation", + "content": "M_9" + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "text", + "content": " frames resulting from each of the events " + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "inline_equation", + "content": "E_9, \\ldots, E_{15}" + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "text", + "content": ", we can re-write (12) as a product of probabilities of events " + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "inline_equation", + "content": "E_9, \\ldots, E_{15}" + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "text", + "content": " to express (10) similarly to (8) and obtain analytical expressions for approximate maximum likelihood estimates " + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "inline_equation", + "content": "\\widehat{q}_{ij}^{\\mathrm{a}}" + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "text", + "content": ". We first find expressions for the conditional probabilities " + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "inline_equation", + "content": "g_9, \\ldots, g_{15}" + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "text", + "content": " of events " + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "inline_equation", + "content": "E_9, \\ldots, E_{15}" + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "text", + "content": ", given that an ambiguous frame was read out. For example, " + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "inline_equation", + "content": "g_9" + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "text", + "content": " is the conditional probability of event " + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "inline_equation", + "content": "E_9" + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "text", + "content": " given that a frame " + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "inline_equation", + "content": "R^t = [1,1]" + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "inline_equation", + "content": "C^t = [1,1]" + }, + { + "bbox": [ + 313, + 513, + 555, + 680 + ], + "type": "text", + "content": " was measured and is given by" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 391, + 688, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 688, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 391, + 688, + 553, + 717 + ], + "type": "interline_equation", + "content": "g _ {9} = \\frac {\\mathbb {P} \\left(E _ {9}\\right)}{\\sum_ {k = 9} ^ {1 5} \\mathbb {P} \\left(E _ {k}\\right)}. \\tag {13}", + "image_path": "3539b7f105cd4ad24917625bf6dcfe721478457fb991cfd5adcef258a72cd4c5.jpg" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11409" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "In addition to the inherent limitation of having a finite amount of data, computing " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(E_9),\\ldots ,\\mathbb{P}(E_{15})" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is hindered by ambiguities. Thus, we estimate these probabilities using the measured single-photon frames, which are unambiguous and unbiased. For example, an estimate of " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(E_9)" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " can be calculated as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 148, + 295, + 164 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 148, + 295, + 164 + ], + "spans": [ + { + "bbox": [ + 102, + 148, + 295, + 164 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathbb {P}} (E _ {9}) = \\widehat {q} _ {1 1} ^ {\\mathrm {s}} \\widehat {q} _ {2 2} ^ {\\mathrm {s}} (1 - \\widehat {q} _ {1 2} ^ {\\mathrm {s}}) (1 - \\widehat {q} _ {2 1} ^ {\\mathrm {s}}). \\tag {14}", + "image_path": "e525f93dbe97d46e51a225bfc9d06492da67ef5ef18c0b6c54936d8a095aec1c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "text", + "content": "Similar expressions can be derived for " + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbb{P}}(E_{10}),\\ldots,\\widehat{\\mathbb{P}}(E_{15})" + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "text", + "content": " which are then substituted in (13) in place of " + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(E_9),\\dots ,\\mathbb{P}(E_{15})" + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "text", + "content": " to get estimates " + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\widehat{g}_9,\\dots ,\\widehat{g}_{15}" + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "text", + "content": " of conditional probabilities " + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "inline_equation", + "content": "g_{9},\\ldots ,g_{15}" + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "text", + "content": ". The fractions of the " + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "inline_equation", + "content": "M_9" + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "text", + "content": " frames resulting from events " + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "inline_equation", + "content": "E_{9},\\ldots E_{15}" + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "text", + "content": " can then be estimated as " + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\widehat{g}_9M_9,\\ldots ,\\widehat{g}_{15}M_9" + }, + { + "bbox": [ + 55, + 168, + 296, + 266 + ], + "type": "text", + "content": ". Finally, these estimates can be used to simplify (12) to a product of probabilities and obtain the approximate maximum likelihood estimates" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 306, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 306, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 306, + 378 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widehat {q} _ {1 1} ^ {\\mathrm {a}} = \\frac {M _ {1} + M _ {5} + M _ {6} + M _ {9} (\\widehat {g} _ {9} + \\widehat {g} _ {1 1} + \\widehat {g} _ {1 2} + \\widehat {g} _ {1 3} + \\widehat {g} _ {1 5})}{T}, \\\\ \\widehat {q} _ {1 2} ^ {\\mathrm {a}} = \\frac {M _ {2} + M _ {5} + M _ {7} + M _ {9} (\\widehat {g} _ {1 0} + \\widehat {g} _ {1 1} + \\widehat {g} _ {1 2} + \\widehat {g} _ {1 4} + \\widehat {g} _ {1 5})}{T}, \\\\ \\widehat {q} _ {2 1} ^ {\\mathrm {a}} = \\frac {M _ {3} + M _ {6} + M _ {8} + M _ {9} (\\widehat {g} _ {1 0} + \\widehat {g} _ {1 2} + \\widehat {g} _ {1 3} + \\widehat {g} _ {1 4} + \\widehat {g} _ {1 5})}{T}, \\\\ \\hat {q} _ {2 2} ^ {\\mathrm {a}} = \\frac {M _ {4} + M _ {7} + M _ {8} + M _ {9} (\\widehat {g} _ {9} + \\widehat {g} _ {1 1} + \\widehat {g} _ {1 3} + \\widehat {g} _ {1 4} + \\widehat {g} _ {1 5})}{T} \\tag {15} \\\\ \\end{array}", + "image_path": "62e9bf62dc89623232f8595af6a59507652b2d38fe2aefbc7d1c9ba41be74823.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "inline_equation", + "content": "M_1, \\ldots, M_8" + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "content": " are the numbers of unambiguous readout frames corresponding to events " + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "inline_equation", + "content": "E_1, \\ldots, E_8" + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 486 + ], + "type": "text", + "content": "While computing the bias of the ME analytically is complex, numerical simulations show that it has a low bias. Furthermore, since the ME depends on the SPE for an initial estimate of the detection probabilities, when the SPE is poor due to few measured single-photon frames, the reduction in MSE from the ME is relatively lesser than in cases when the SPE has low variance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 493, + 257, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 257, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 257, + 506 + ], + "type": "text", + "content": "4.4. Scaling to Higher Dimensional Arrays" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "content": "In a " + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "content": " array, there is a single type of ambiguous readout which indicates photon incidences at both rows and both columns of the array. As the spatial dimension of the array increases, there are more types of ambiguous readouts. For instance, with a " + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "content": " array, we can have the following ambiguous readouts:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 582, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 594 + ], + "type": "text", + "content": "- 2 rows and 2 columns fired - 4 candidate pixel locations;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 294, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 294, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 294, + 605 + ], + "type": "text", + "content": "- 2 rows and 3 columns fired - 6 candidate pixel locations;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 606, + 294, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 294, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 294, + 617 + ], + "type": "text", + "content": "- 3 rows and 2 columns fired - 6 candidate pixel locations;" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "- 3 rows and 3 columns fired -9 candidate pixel locations. Resolving these additional ambiguous readouts requires the modification of the expressions for conditional probabilities " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": ". As the number of rows and columns with photon coincidences increases, the number of unique terms in the expressions for " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": " grows rapidly. Hence, we restrict our analysis and simulations to at most 4-photon coincidences and ignore terms resulting from 5-photon coincidences and above." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": "We expect that this assumption introduces a small bias in the ME, especially at higher incident fluxes. Nevertheless, the improvement over handling only single-photon frames can be substantial." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 133, + 368, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 368, + 145 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 368, + 145 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 153, + 459, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 153, + 459, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 459, + 166 + ], + "type": "text", + "content": "5.1. Imaging Arbitrary Scenes" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 171, + 554, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 554, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 554, + 301 + ], + "type": "text", + "content": "Figure 3 shows the Monte Carlo simulation results of reconstructing " + }, + { + "bbox": [ + 313, + 171, + 554, + 301 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 171, + 554, + 301 + ], + "type": "text", + "content": " images from 100,000 measured frames using each of the estimators. Note that these are estimates of the flux " + }, + { + "bbox": [ + 313, + 171, + 554, + 301 + ], + "type": "inline_equation", + "content": "\\hat{\\Lambda}_{ij}" + }, + { + "bbox": [ + 313, + 171, + 554, + 301 + ], + "type": "text", + "content": " at each pixel, computed using the detection probability estimate " + }, + { + "bbox": [ + 313, + 171, + 554, + 301 + ], + "type": "inline_equation", + "content": "\\hat{q}_{ij}" + }, + { + "bbox": [ + 313, + 171, + 554, + 301 + ], + "type": "text", + "content": " from the three estimators followed by the transformation in (4). Reconstructed in rows (a), (b), (c), and (d) are at a scaling of 3 mean photons per frame (PPF), while row (e) is at a scaling of 4 mean PPF. Mean PPF is the average number of detected photons in each measured frame. Each experiment is repeated 100 times for the computation of PSNR." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 303, + 554, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 554, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 554, + 375 + ], + "type": "text", + "content": "In rows (a), (b), and (c), it can be seen that the NE introduces horizontal and vertical streaks due to misattributions from multiphoton coincidences. For example, in row (a) the SPE achieves " + }, + { + "bbox": [ + 313, + 303, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\sim 5" + }, + { + "bbox": [ + 313, + 303, + 554, + 375 + ], + "type": "text", + "content": " dB PSNR improvement over the NE and does not show the same artifacts, although the petals of the flower appear noisy." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 376, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 662 + ], + "type": "text", + "content": "Our proposed ME outperforms the naive estimator by " + }, + { + "bbox": [ + 313, + 376, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\sim 11\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 376, + 555, + 662 + ], + "type": "text", + "content": " while its improvement over the single-photon estimator is " + }, + { + "bbox": [ + 313, + 376, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\sim 6\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 376, + 555, + 662 + ], + "type": "text", + "content": ". Further, we note that the ME preserves features in the ground truth image better than both of the other estimators, in addition to eliminating misattribution artifacts. The petals of the flower in row (a) are more clearly visible, while these are missing in the other reconstructions. The improvement in the reconstruction depends on some features of the ground truth image. For example, in row (d) we consider a rotated version of the image in row (a). It can be seen that the misattributions in the NE are more concentrated towards the center-left of the image as compared to row (a) where the horizontal and vertical streaks cut across the image. The PSNR of the naive reconstruction is lower than that in row (a), however, the multiphoton reconstruction still achieves " + }, + { + "bbox": [ + 313, + 376, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\sim 5\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 376, + 555, + 662 + ], + "type": "text", + "content": " improvement. In row (e) we consider the image from row (b) with a PPF scaling of 4. We note that the PSNR of the NE reduces by " + }, + { + "bbox": [ + 313, + 376, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\sim 0.6" + }, + { + "bbox": [ + 313, + 376, + 555, + 662 + ], + "type": "text", + "content": " dB while the PSNR of the SPE and ME reduce by " + }, + { + "bbox": [ + 313, + 376, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\sim 3" + }, + { + "bbox": [ + 313, + 376, + 555, + 662 + ], + "type": "text", + "content": " dB each. This behavior is expected as an increase in the average PPF causes the number of frames with a single detected photon to decrease, which reduces the PSNR of the SPE and consequently the ME. However, the performance of the ME can be improved by considering additional multiphoton events." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 671, + 470, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 470, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 470, + 684 + ], + "type": "text", + "content": "5.2. Optimal Photons Per Frame" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "In single-photon imaging, high signal strength is often beneficial for parameter estimation. However, for readout" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11410" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 91, + 115, + 135 + ], + "blocks": [ + { + "bbox": [ + 53, + 108, + 66, + 118 + ], + "lines": [ + { + "bbox": [ + 53, + 108, + 66, + 118 + ], + "spans": [ + { + "bbox": [ + 53, + 108, + 66, + 118 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 79, + 116, + 89 + ], + "lines": [ + { + "bbox": [ + 66, + 79, + 116, + 89 + ], + "spans": [ + { + "bbox": [ + 66, + 79, + 116, + 89 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 91, + 115, + 135 + ], + "lines": [ + { + "bbox": [ + 69, + 91, + 115, + 135 + ], + "spans": [ + { + "bbox": [ + 69, + 91, + 115, + 135 + ], + "type": "image", + "image_path": "0854ebe751fc7ecad720bc03f241ecdb7133a1020552c100d4a7b1d284e4c97d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 126, + 91, + 171, + 135 + ], + "blocks": [ + { + "bbox": [ + 136, + 80, + 158, + 89 + ], + "lines": [ + { + "bbox": [ + 136, + 80, + 158, + 89 + ], + "spans": [ + { + "bbox": [ + 136, + 80, + 158, + 89 + ], + "type": "text", + "content": "Naive" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 126, + 91, + 171, + 135 + ], + "lines": [ + { + "bbox": [ + 126, + 91, + 171, + 135 + ], + "spans": [ + { + "bbox": [ + 126, + 91, + 171, + 135 + ], + "type": "image", + "image_path": "4cb4c05d0831fd24ffe7b79b7bf39bdb983a53fc3b874b821c1a8ddc4ee36fbe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 137, + 166, + 146 + ], + "lines": [ + { + "bbox": [ + 136, + 137, + 166, + 146 + ], + "spans": [ + { + "bbox": [ + 136, + 137, + 166, + 146 + ], + "type": "text", + "content": "16.39 dB" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 127, + 148, + 171, + 191 + ], + "blocks": [ + { + "bbox": [ + 127, + 148, + 171, + 191 + ], + "lines": [ + { + "bbox": [ + 127, + 148, + 171, + 191 + ], + "spans": [ + { + "bbox": [ + 127, + 148, + 171, + 191 + ], + "type": "image", + "image_path": "a109cc8f01f6d863c5c459a35bf5c37e836c4604cd807d19fcb114d4a9707773.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 193, + 166, + 201 + ], + "lines": [ + { + "bbox": [ + 137, + 193, + 166, + 201 + ], + "spans": [ + { + "bbox": [ + 137, + 193, + 166, + 201 + ], + "type": "text", + "content": "17.05 dB" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 127, + 205, + 171, + 249 + ], + "blocks": [ + { + "bbox": [ + 127, + 205, + 171, + 249 + ], + "lines": [ + { + "bbox": [ + 127, + 205, + 171, + 249 + ], + "spans": [ + { + "bbox": [ + 127, + 205, + 171, + 249 + ], + "type": "image", + "image_path": "d2fb0d1ccf9d1b77694f31e870c4d679ddba6ce9a7349b645e853b6a326be5af.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 250, + 166, + 258 + ], + "lines": [ + { + "bbox": [ + 136, + 250, + 166, + 258 + ], + "spans": [ + { + "bbox": [ + 136, + 250, + 166, + 258 + ], + "type": "text", + "content": "13.17 dB" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 127, + 261, + 171, + 305 + ], + "blocks": [ + { + "bbox": [ + 127, + 261, + 171, + 305 + ], + "lines": [ + { + "bbox": [ + 127, + 261, + 171, + 305 + ], + "spans": [ + { + "bbox": [ + 127, + 261, + 171, + 305 + ], + "type": "image", + "image_path": "524604aa7eb71ab13b7c2cad14cc99a59791853cfc1370e1e7e3831fd50c5c59.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 306, + 166, + 314 + ], + "lines": [ + { + "bbox": [ + 137, + 306, + 166, + 314 + ], + "spans": [ + { + "bbox": [ + 137, + 306, + 166, + 314 + ], + "type": "text", + "content": "16.15 dB" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 127, + 318, + 171, + 362 + ], + "blocks": [ + { + "bbox": [ + 127, + 318, + 171, + 362 + ], + "lines": [ + { + "bbox": [ + 127, + 318, + 171, + 362 + ], + "spans": [ + { + "bbox": [ + 127, + 318, + 171, + 362 + ], + "type": "image", + "image_path": "aee72b6f18da28f8708d513436647c4a478c143cc03330ded32191f5c5c73fc0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 363, + 166, + 372 + ], + "lines": [ + { + "bbox": [ + 137, + 363, + 166, + 372 + ], + "spans": [ + { + "bbox": [ + 137, + 363, + 166, + 372 + ], + "type": "text", + "content": "16.44 dB" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 336, + 66, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 336, + 66, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 336, + 66, + 346 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 380, + 296, + 479 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 479 + ], + "type": "text", + "content": "Figure 3. Reconstruction of arbitrary " + }, + { + "bbox": [ + 55, + 380, + 296, + 479 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 380, + 296, + 479 + ], + "type": "text", + "content": " images from 100 000 measured frames using the naive, single, and multiphoton estimators. Each PSNR value is from averaging over 100 trials. Rows (a)-(c) show that the multiphoton estimator is 4 to 6 dB better than the single-photon estimator and around 6 to 11 dB better than the naive estimator. Row (d) shows the impact of rotating the ground truth image on the reconstruction. Row (e) shows the impact of increasing the average number of photons per frame from 3 to 4. The SPE rejects " + }, + { + "bbox": [ + 55, + 380, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\sim 85\\%" + }, + { + "bbox": [ + 55, + 380, + 296, + 479 + ], + "type": "text", + "content": " of measured frames at a PPF of 3." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 182, + 91, + 227, + 135 + ], + "blocks": [ + { + "bbox": [ + 178, + 79, + 231, + 89 + ], + "lines": [ + { + "bbox": [ + 178, + 79, + 231, + 89 + ], + "spans": [ + { + "bbox": [ + 178, + 79, + 231, + 89 + ], + "type": "text", + "content": "Single-Photon" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 91, + 227, + 135 + ], + "lines": [ + { + "bbox": [ + 182, + 91, + 227, + 135 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 227, + 135 + ], + "type": "image", + "image_path": "c0945a320a4448bf3e9b86e026f7201c8cf477356d8d58644c5b43fb1a234bcd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 183, + 148, + 227, + 192 + ], + "blocks": [ + { + "bbox": [ + 190, + 138, + 221, + 146 + ], + "lines": [ + { + "bbox": [ + 190, + 138, + 221, + 146 + ], + "spans": [ + { + "bbox": [ + 190, + 138, + 221, + 146 + ], + "type": "text", + "content": "21.55 dB" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 183, + 148, + 227, + 192 + ], + "lines": [ + { + "bbox": [ + 183, + 148, + 227, + 192 + ], + "spans": [ + { + "bbox": [ + 183, + 148, + 227, + 192 + ], + "type": "image", + "image_path": "b581499b19afb24926b662f06da2ecf86795de9e932322970d6c0fabb929a862.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 193, + 221, + 201 + ], + "lines": [ + { + "bbox": [ + 190, + 193, + 221, + 201 + ], + "spans": [ + { + "bbox": [ + 190, + 193, + 221, + 201 + ], + "type": "text", + "content": "22.43 dB" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 183, + 205, + 227, + 249 + ], + "blocks": [ + { + "bbox": [ + 183, + 205, + 227, + 249 + ], + "lines": [ + { + "bbox": [ + 183, + 205, + 227, + 249 + ], + "spans": [ + { + "bbox": [ + 183, + 205, + 227, + 249 + ], + "type": "image", + "image_path": "e38a06ce716abb9928099add0c336cd3d3d1a5c710d3b9fa41083b48619531f8.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 250, + 222, + 258 + ], + "lines": [ + { + "bbox": [ + 192, + 250, + 222, + 258 + ], + "spans": [ + { + "bbox": [ + 192, + 250, + 222, + 258 + ], + "type": "text", + "content": "16.63 dB" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 182, + 261, + 227, + 305 + ], + "blocks": [ + { + "bbox": [ + 182, + 261, + 227, + 305 + ], + "lines": [ + { + "bbox": [ + 182, + 261, + 227, + 305 + ], + "spans": [ + { + "bbox": [ + 182, + 261, + 227, + 305 + ], + "type": "image", + "image_path": "13fe92ea23853e0cdcf7125dc8e641d83e61ce2e61d6bf02097d74b56f957f43.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 306, + 221, + 314 + ], + "lines": [ + { + "bbox": [ + 192, + 306, + 221, + 314 + ], + "spans": [ + { + "bbox": [ + 192, + 306, + 221, + 314 + ], + "type": "text", + "content": "16.60 dB" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 182, + 318, + 227, + 362 + ], + "blocks": [ + { + "bbox": [ + 182, + 318, + 227, + 362 + ], + "lines": [ + { + "bbox": [ + 182, + 318, + 227, + 362 + ], + "spans": [ + { + "bbox": [ + 182, + 318, + 227, + 362 + ], + "type": "image", + "image_path": "3685f1a69eb6a9aff78ae6421d09b6d02b44daac875fc6dcc6d16081ca2d80c3.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 363, + 221, + 372 + ], + "lines": [ + { + "bbox": [ + 192, + 363, + 221, + 372 + ], + "spans": [ + { + "bbox": [ + 192, + 363, + 221, + 372 + ], + "type": "text", + "content": "19.59 dB" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 238, + 91, + 283, + 135 + ], + "blocks": [ + { + "bbox": [ + 238, + 79, + 284, + 89 + ], + "lines": [ + { + "bbox": [ + 238, + 79, + 284, + 89 + ], + "spans": [ + { + "bbox": [ + 238, + 79, + 284, + 89 + ], + "type": "text", + "content": "Multiphoton" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 91, + 283, + 135 + ], + "lines": [ + { + "bbox": [ + 238, + 91, + 283, + 135 + ], + "spans": [ + { + "bbox": [ + 238, + 91, + 283, + 135 + ], + "type": "image", + "image_path": "52f61186c13fb31a4dbc1e1e70d5a8ee66880619c966d46d9491172ddc8d846f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 239, + 148, + 283, + 192 + ], + "blocks": [ + { + "bbox": [ + 247, + 138, + 278, + 146 + ], + "lines": [ + { + "bbox": [ + 247, + 138, + 278, + 146 + ], + "spans": [ + { + "bbox": [ + 247, + 138, + 278, + 146 + ], + "type": "text", + "content": "27.22 dB" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 239, + 148, + 283, + 192 + ], + "lines": [ + { + "bbox": [ + 239, + 148, + 283, + 192 + ], + "spans": [ + { + "bbox": [ + 239, + 148, + 283, + 192 + ], + "type": "image", + "image_path": "900b2d151395b3d981d6078d37968f6287fcea5dbd29746f6d0859ba4c11c879.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 193, + 278, + 201 + ], + "lines": [ + { + "bbox": [ + 248, + 193, + 278, + 201 + ], + "spans": [ + { + "bbox": [ + 248, + 193, + 278, + 201 + ], + "type": "text", + "content": "28.41 dB" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 239, + 205, + 283, + 249 + ], + "blocks": [ + { + "bbox": [ + 239, + 205, + 283, + 249 + ], + "lines": [ + { + "bbox": [ + 239, + 205, + 283, + 249 + ], + "spans": [ + { + "bbox": [ + 239, + 205, + 283, + 249 + ], + "type": "image", + "image_path": "0cd7e4665127c5cefd2b44478cd620961d26a680e0c2382f33afc5589fe7bfc6.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 250, + 278, + 258 + ], + "lines": [ + { + "bbox": [ + 247, + 250, + 278, + 258 + ], + "spans": [ + { + "bbox": [ + 247, + 250, + 278, + 258 + ], + "type": "text", + "content": "22.54 dB" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 239, + 261, + 283, + 305 + ], + "blocks": [ + { + "bbox": [ + 239, + 261, + 283, + 305 + ], + "lines": [ + { + "bbox": [ + 239, + 261, + 283, + 305 + ], + "spans": [ + { + "bbox": [ + 239, + 261, + 283, + 305 + ], + "type": "image", + "image_path": "0a0e13c9e88ce93373e11745863a9f0b481daeea4b9394bf5d1332740573f451.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 306, + 278, + 314 + ], + "lines": [ + { + "bbox": [ + 247, + 306, + 278, + 314 + ], + "spans": [ + { + "bbox": [ + 247, + 306, + 278, + 314 + ], + "type": "text", + "content": "21.79 dB" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 239, + 318, + 283, + 362 + ], + "blocks": [ + { + "bbox": [ + 239, + 318, + 283, + 362 + ], + "lines": [ + { + "bbox": [ + 239, + 318, + 283, + 362 + ], + "spans": [ + { + "bbox": [ + 239, + 318, + 283, + 362 + ], + "type": "image", + "image_path": "b0fc083ccf464eb0bab494a1756cbe1191a4c4cad3edb2142172afcd7cd5aa30.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 363, + 278, + 372 + ], + "lines": [ + { + "bbox": [ + 248, + 363, + 278, + 372 + ], + "spans": [ + { + "bbox": [ + 248, + 363, + 278, + 372 + ], + "type": "text", + "content": "25.98 dB" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 506, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 579 + ], + "type": "text", + "content": "multiplexed arrays, high flux also leads to more ambiguous readout frames, which makes image reconstruction challenging. We therefore expect that there exists an optimal incident photon flux level where each of the estimators achieves the minimum mean-square error. The following results, as shown in Fig. 4, validate this hypothesis." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "A ground truth image can be scaled in experimental settings by selecting a desired integration time or attenuating the incident photon flux using a neutral density filter. To study the dependence of estimator performance on the incident photon flux, we perform a simulation with a high emitted photon flux from a source that is attenuated before being incident on the array. We then study the change in mean-squared error per pixel " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{MSE} = \\| \\widehat{\\Lambda} - \\Lambda \\|_2^2 / n^2)" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " of the reconstructed image with varying strength of the attenuating filter. The results of this simulation are shown in Fig. 4 for the reconstruction of a " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " ground truth image in" + } + ] + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 318, + 83, + 542, + 260 + ], + "blocks": [ + { + "bbox": [ + 318, + 83, + 542, + 260 + ], + "lines": [ + { + "bbox": [ + 318, + 83, + 542, + 260 + ], + "spans": [ + { + "bbox": [ + 318, + 83, + 542, + 260 + ], + "type": "image", + "image_path": "532c68806317831e870b561a4699782325bb469862effa92b5e38aea42f74030.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 270, + 554, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 304 + ], + "type": "text", + "content": "Figure 4. Mean-squared errors of the estimators studied as functions of mean photons per frame. (Ground truth is image in Fig. 3(c).)" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 326, + 385, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 385, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 385, + 338 + ], + "type": "text", + "content": "row (c) of Fig. 3." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 339, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 578 + ], + "type": "text", + "content": "Note that as the mean PPF increases, the MSEs of all the estimators initially decrease, reach a minimum, and then increase. High MSE at low PPF values is due to the lack of detected photons for accurate reconstruction. The increase in MSE at high PPF values is expected due to the high variance of SPE and the high probability of multiphoton events not modeled by the ME. It can also be seen that the minimum MSE of the multiphoton estimator is lower than that of the single-photon estimator, and both of these are much lower than the naive estimator. Furthermore, the minimum value of MSE for the ME occurs at a PPF of 1.4, which is higher than that for the SPE at 0.83, and that for NE at 0.45. We expect these incidence flux values to be realistic in conditions similar to imaging under moonlight, where high-efficiency detectors like SNSPDs are typically used. Thus, the ME produces a better reconstruction compared to both the NE and the SPE when multiple photons are detected simultaneously at the array. This results in an increased photon count rate and a better utilization of the spatial dimensions of the array." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 587, + 480, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 480, + 599 + ], + "type": "text", + "content": "5.3. 2-, 3- and 4-Photon Estimators" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "As explained in Sec. 4.4, our estimator resolves up to 4 simultaneously incident photons. The conditional probability expressions (g's) contain terms corresponding to every possible photon event that can result in the measured readout. However, we can simplify these expressions by restricting the number of modeled events. For instance, if we include only 2-photon coincidences in the case of a " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " array, only conditional probabilities of events " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "E_{9}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "E_{10}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " would be summed in (13). Fig. 4 shows the performance of the ME" + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11411" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 97, + 176, + 159 + ], + "blocks": [ + { + "bbox": [ + 131, + 82, + 156, + 92 + ], + "lines": [ + { + "bbox": [ + 131, + 82, + 156, + 92 + ], + "spans": [ + { + "bbox": [ + 131, + 82, + 156, + 92 + ], + "type": "text", + "content": "Naive" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 115, + 97, + 176, + 159 + ], + "lines": [ + { + "bbox": [ + 115, + 97, + 176, + 159 + ], + "spans": [ + { + "bbox": [ + 115, + 97, + 176, + 159 + ], + "type": "image", + "image_path": "df831896f988a54927e707630b3dc510a5f5a3d08e013165a1b7c5ae8da31c23.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 163, + 166, + 174 + ], + "lines": [ + { + "bbox": [ + 129, + 163, + 166, + 174 + ], + "spans": [ + { + "bbox": [ + 129, + 163, + 166, + 174 + ], + "type": "text", + "content": "22.59 dB" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 195, + 97, + 256, + 159 + ], + "blocks": [ + { + "bbox": [ + 195, + 82, + 253, + 92 + ], + "lines": [ + { + "bbox": [ + 195, + 82, + 253, + 92 + ], + "spans": [ + { + "bbox": [ + 195, + 82, + 253, + 92 + ], + "type": "text", + "content": "Single-Photon" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 195, + 97, + 256, + 159 + ], + "lines": [ + { + "bbox": [ + 195, + 97, + 256, + 159 + ], + "spans": [ + { + "bbox": [ + 195, + 97, + 256, + 159 + ], + "type": "image", + "image_path": "90a7dcbeb8831e0466399d52e5fdb40c87bf4774d5d2ead4e860d9340e06ba45.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 163, + 246, + 174 + ], + "lines": [ + { + "bbox": [ + 208, + 163, + 246, + 174 + ], + "spans": [ + { + "bbox": [ + 208, + 163, + 246, + 174 + ], + "type": "text", + "content": "23.64 dB" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 204, + 119, + 267 + ], + "blocks": [ + { + "bbox": [ + 69, + 190, + 106, + 201 + ], + "lines": [ + { + "bbox": [ + 69, + 190, + 106, + 201 + ], + "spans": [ + { + "bbox": [ + 69, + 190, + 106, + 201 + ], + "type": "text", + "content": "2-photon" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 204, + 119, + 267 + ], + "lines": [ + { + "bbox": [ + 57, + 204, + 119, + 267 + ], + "spans": [ + { + "bbox": [ + 57, + 204, + 119, + 267 + ], + "type": "image", + "image_path": "85129537de147b3bad9295e0a352d1c5c7ec72326b5e0c6e92cb7984a733408a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 271, + 109, + 281 + ], + "lines": [ + { + "bbox": [ + 71, + 271, + 109, + 281 + ], + "spans": [ + { + "bbox": [ + 71, + 271, + 109, + 281 + ], + "type": "text", + "content": "23.83 dB" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 295, + 295, + 328 + ], + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 328 + ], + "type": "text", + "content": "Figure 5. Image reconstruction by operating each of the studied estimators at mean PPF values corresponding to their lowest MSE's. Indicated values are PSNR of reconstructions." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 138, + 204, + 202, + 267 + ], + "blocks": [ + { + "bbox": [ + 149, + 191, + 187, + 202 + ], + "lines": [ + { + "bbox": [ + 149, + 191, + 187, + 202 + ], + "spans": [ + { + "bbox": [ + 149, + 191, + 187, + 202 + ], + "type": "text", + "content": "3-photon" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 204, + 202, + 267 + ], + "lines": [ + { + "bbox": [ + 138, + 204, + 202, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 204, + 202, + 267 + ], + "type": "image", + "image_path": "bc699e795f461035376998688a59582f49fdf3c763c7537ad8ec5598bbb09635.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 271, + 190, + 281 + ], + "lines": [ + { + "bbox": [ + 152, + 271, + 190, + 281 + ], + "spans": [ + { + "bbox": [ + 152, + 271, + 190, + 281 + ], + "type": "text", + "content": "24.60 dB" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 222, + 204, + 286, + 267 + ], + "blocks": [ + { + "bbox": [ + 235, + 191, + 272, + 202 + ], + "lines": [ + { + "bbox": [ + 235, + 191, + 272, + 202 + ], + "spans": [ + { + "bbox": [ + 235, + 191, + 272, + 202 + ], + "type": "text", + "content": "4-photon" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 222, + 204, + 286, + 267 + ], + "lines": [ + { + "bbox": [ + 222, + 204, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 222, + 204, + 286, + 267 + ], + "type": "image", + "image_path": "a2d585d3963615f4cc791b4628d550342dd3927fcd4a6fb3028853de30c8b71f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 271, + 275, + 281 + ], + "lines": [ + { + "bbox": [ + 238, + 271, + 275, + 281 + ], + "spans": [ + { + "bbox": [ + 238, + 271, + 275, + 281 + ], + "type": "text", + "content": "26.74 dB" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": "when 2-, 3- and 4-photon are modeled in " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": ". It can be seen that the MSE of reconstruction is successively reduced with each additional photon considered in the estimator. The 2-photon estimator achieves its lowest MSE at a mean PPF of 1, while the 3- and 4-photon estimators achieve their lowest MSEs at 1.2 and 1.4 mean PPF, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "text", + "content": "Image reconstruction using each of the estimators in Fig. 4 at their optimal mean PPF values is summarized in Fig. 5. We see that the 4-photon estimator achieves the highest PSNR, which is 3 to 4 dB higher than the NE and the SPE." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 522, + 248, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 248, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 248, + 536 + ], + "type": "text", + "content": "5.4. Reduction in Total Integration Time" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Since the multiphoton estimator enables the accurate interpretation of photon coincidences, it can be leveraged to reduce integration times in imaging with single-photon detectors. The performance of the three estimators in achieving a target reconstruction MSE is shown in Fig. 6. It can be seen that to achieve an MSE of 0.01, the ME requires 25,000 measured frames, and the SPE requires 100,000 measured frames, while the NE does not achieve this MSE within the range of frame values studied. This shows that the ME achieves a factor " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\sim 4" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " reduction in the integration time for the same reconstruction quality as the other estimators. In comparison, a fully-addressed readout scheme achieves the target MSE with only 5000 frames, but with the increased complexity of readout." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 317, + 81, + 549, + 263 + ], + "blocks": [ + { + "bbox": [ + 317, + 81, + 549, + 263 + ], + "lines": [ + { + "bbox": [ + 317, + 81, + 549, + 263 + ], + "spans": [ + { + "bbox": [ + 317, + 81, + 549, + 263 + ], + "type": "image", + "image_path": "7d1906ced67ab22223014d577a43c985cfd0395d033c73bb214d5b1ace7985df.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 271, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 304 + ], + "type": "text", + "content": "Figure 6. Reduction in the required number of frames to achieve the same target MSE by the NE, SPE,ME, and a fully-addressed readout scheme." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 326, + 432, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 432, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 432, + 338 + ], + "type": "text", + "content": "5.5. Cramér-Rao Bound" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 343, + 555, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 476 + ], + "type": "text", + "content": "The Cramér-Rao bound (CRB) provides the lowest possible variance of an unbiased estimator for parameter estimation. Here, we compare the performance of the NE, SPE, 2-photon, 3-photon, and 4-photon estimators on the estimation of a " + }, + { + "bbox": [ + 313, + 343, + 555, + 476 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 343, + 555, + 476 + ], + "type": "text", + "content": " ground truth, against CRB values at different incident photon fluxes. We first compute the Fisher information matrix (FIM) using the likelihood expression for row-column readouts from a " + }, + { + "bbox": [ + 313, + 343, + 555, + 476 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 343, + 555, + 476 + ], + "type": "text", + "content": " array. The complete derivation of expressions for elements of the FIM are shown in the supplement. The element " + }, + { + "bbox": [ + 313, + 343, + 555, + 476 + ], + "type": "inline_equation", + "content": "(1,1)" + }, + { + "bbox": [ + 313, + 343, + 555, + 476 + ], + "type": "text", + "content": " of the FIM for a single measurement is" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 349, + 495, + 553, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 495, + 553, + 549 + ], + "spans": [ + { + "bbox": [ + 349, + 495, + 553, + 549 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {I} _ {1 1} = \\frac {(1 - q _ {4}) (1 - q _ {2} q _ {3})}{q _ {1}} + \\frac {(1 - q _ {2} q _ {3})}{1 - q _ {1}} \\\\ + \\frac {q _ {4} ^ {2} \\left(1 - q _ {2} q _ {3}\\right) ^ {2}}{q _ {1} q _ {4} + q _ {2} q _ {3} - q _ {1} q _ {2} q _ {3} q _ {4}}. \\tag {16} \\\\ \\end{array}", + "image_path": "984cde0cd71d9bf73d3b08051de007fb2ccda819b942dbef7bc5bbb90d0fd729.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "type": "text", + "content": "Other diagonal elements have similar forms as shown in the supplement. The CRB is then computed as the mean of the diagonal elements in " + }, + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{-1}" + }, + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "type": "text", + "content": ". The variation of CRB values, as well as MSEs of the studied estimators as a function of incident photon flux, is shown in Fig. 7." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "It can be seen that the ME curve matches the CRB curve for a range of mean PPF values. At high fluxes, the MSE of the ME increases due to the high variance of the single-photon estimates, which are used to calculate " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " in (13). In the general case with bright spots at all four pixels of the array, the 4-photon estimator most closely matches the CRB across the range of PPF values studied. The 2- and 3-photon estimators deviate from the CRB due to the modeling of" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11412" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 76, + 276, + 257 + ], + "blocks": [ + { + "bbox": [ + 62, + 76, + 276, + 257 + ], + "lines": [ + { + "bbox": [ + 62, + 76, + 276, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 76, + 276, + 257 + ], + "type": "image", + "image_path": "9adc2f33e01880b126796b9a8a0089858f74bddfcd55e80e8c50eaf6cd623f9e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 264, + 298, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 264, + 298, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 298, + 331 + ], + "type": "text", + "content": "Figure 7. Comparison of estimator performance with the Cramér-Rao Bound. The ME matches the CRB for a range of incident photon fluxes and deviates from it at high fluxes. With 4 bright spots, model mismatch causes 2- and 3-photon estimators to become biased and deviate from the CRB at lower incident photon fluxes compared to the 4-photon estimator." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 363, + 297, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 297, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 297, + 411 + ], + "type": "text", + "content": "only a subset of events that can cause ambiguities and omission of terms in the expression for conditional probabilities in (13). The dependence of the estimator performance on the ground truth image is shown in the Supplement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 206, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 206, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 206, + 448 + ], + "type": "text", + "content": "5.6. Increasing array resolution" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "content": "Our analysis of the multiphoton estimator's performance so far has been limited to " + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 456, + 296, + 564 + ], + "type": "text", + "content": " arrays to match the sensor developed in [37]. Here, we study its use in arrays with higher pixel counts and finer spatial resolutions. We emulate a fixed sensor area with a varying number of pixels and fixed maximum photon flux. This results in photons per pixel being inversely proportional to the number of pixels. We include attenuation optimization for each reconstruction method, similar to Fig. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": "Results for row 1 of Fig. 3 as the ground truth are shown in Fig. 8 Performance improvement from the ME increases with array size (see Fig. 8(a)). This shows that our method is applicable to array sizes encountered in commercial imaging use cases. Nevertheless, we expect coincidences not modeled to be a limiting factor. Further, computation time increases linearly with an increase in array size. Thus, for a practical implementation of our algorithm, it could be beneficial to divide the sensor area into smaller blocks, the readouts from each of which can be processed using the ME. Such a processing approach would be particularly beneficial in sensor architectures similar to [17]." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 322, + 79, + 533, + 255 + ], + "blocks": [ + { + "bbox": [ + 322, + 79, + 533, + 255 + ], + "lines": [ + { + "bbox": [ + 322, + 79, + 533, + 255 + ], + "spans": [ + { + "bbox": [ + 322, + 79, + 533, + 255 + ], + "type": "image", + "image_path": "985c34be740c8b7943127fc80199ba14d2221c3d731d53f413155aff16ffd5cd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 329, + 262, + 533, + 445 + ], + "blocks": [ + { + "bbox": [ + 329, + 262, + 533, + 445 + ], + "lines": [ + { + "bbox": [ + 329, + 262, + 533, + 445 + ], + "spans": [ + { + "bbox": [ + 329, + 262, + 533, + 445 + ], + "type": "image", + "image_path": "41357b09a5d95f2bb815f4f53d986396596d07a0ed67d0c6a502b2263b31e885.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 448, + 556, + 481 + ], + "lines": [ + { + "bbox": [ + 313, + 448, + 556, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 556, + 481 + ], + "type": "text", + "content": "Figure 8. Influence of array size. (a) MSEs at optimum attenuation for each estimator. (b) Optimum photons per frame for each estimator." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 517, + 489, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 489, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 489, + 530 + ], + "type": "text", + "content": "6. Conclusions and Future Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": "We formulate multiphoton coincidence resolution in readout multiplexed single-photon detector arrays as an inverse imaging problem and propose a novel estimator to resolve up to four photon coincidences with row-column multiplexed readouts. Our method enables high-flux, low-latency, high-resolution image reconstructions that significantly improve upon conventional methods. Future work can extend this method to resolve an abitrarily large number of coincident photon events or other types of multiplexed readouts. The use of spatial priors and deep learning methods can be explored to improve the presented solutions. This line of work could be especially useful for the use of SNSPD arrays in photon-starved imaging applications like deep space imaging or biological imaging." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11413" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 147 + ], + "type": "text", + "content": "This work was supported in part by the U.S. National Science Foundation under Grant 2039762 and in part by a gift from Dr. John Zheng Sun. We thank Prof. Hoover Rueda-Chacon for helpful discussions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 160, + 115, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 160, + 115, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 160, + 115, + 172 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 179, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 179, + 294, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 179, + 294, + 245 + ], + "spans": [ + { + "bbox": [ + 61, + 179, + 294, + 245 + ], + "type": "text", + "content": "[1] M. S. Allman, V. B. Verma, M. Stevens, T. Gerrits, R. D. Horansky, A. E. Lita, F. Marsili, A. Beyer, M. D. Shaw, D. Kumor, R. Mirin, and S. W. Nam. A near-infrared 64-pixel superconducting nanowire single photon detector array with integrated multiplexed readout. Applied Physics Letters, 106 (19), 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 294, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 294, + 280 + ], + "type": "text", + "content": "[2] Ibrahim Ben Daya. Systematic characterization framework of row-column ultrasound imaging systems. Master's thesis, University of Waterloo, 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 282, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 282, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 282, + 294, + 315 + ], + "type": "text", + "content": "[3] PL Bonanno. Physical and cost comparison of smartphone laser autofocus solutions. In Proc. Int. SPAD Sensor Workshop, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 317, + 294, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 294, + 372 + ], + "type": "text", + "content": "[4] Daniel AB Bonifacio and Maurizio Moralles. Modeling of 3d gamma interaction position in a monolithic scintillator block with a row-column summing readout. In 2012 IEEE Nuclear Science Symposium and Medical Imaging Conference Record (NSS/MIC), pages 2606-2613. IEEE, 2012. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 373, + 294, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 294, + 427 + ], + "type": "text", + "content": "[5] Jihwan Boo, Mark D Hammig, and Manhee Jeong. Row-column readout method to mitigate radiographic-image blurring from multipixel events in a coded-aperture imaging system. IEEE Transactions on Nuclear Science, 68(5):1175-1183, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 430, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 294, + 483 + ], + "type": "text", + "content": "[6] Caroline K Brennan, Zi Yao, Anastasia A Ionkina, Colin M Rathbun, Buvaneshwari Sathishkumar, and Jennifer A Prescher. Multiplexed bioluminescence imaging with a substrate unmixing platform. Cell Chemical Biology, 29(11): 1649-1660, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 294, + 531 + ], + "type": "text", + "content": "[7] Claudio Bruschini, Harald Homulle, Ivan Michel Antolovic, Samuel Burri, and Edoardo Charbon. Single-photon avalanche diode imagers in biophotonics: review and outlook. Light: Science & Applications, 8(1):87, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "type": "text", + "content": "[8] Nicolas Callens and Georges GE Gielen. Analysis and comparison of readout architectures and analog-to-digital converters for 3d-stacked cmos image sensors. IEEE Transactions on Circuits and Systems I: Regular Papers, 68(8):3117-3130, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 589, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 589, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 62, + 589, + 294, + 643 + ], + "type": "text", + "content": "[9] Francesco Ceccarelli, Giulia Acconcia, Angelo Gulinatti, Massimo Ghioni, Ivan Rech, and Roberto Osellame. Recent advances and future perspectives of single-photon avalanche diodes for quantum photonics applications. Advanced Quantum Technologies, 4(2):2000102, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 646, + 294, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 294, + 688 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 294, + 688 + ], + "type": "text", + "content": "[10] Garry Chinn, Peter D. Olcott, and Craig S. Levin. Sparse signal recovery methods for multiplexing pet detector readout. IEEE Transactions on Medical Imaging, 32(5):932-942, 2013. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[11] Iris Cusini, Davide Berretta, Enrico Conca, Alfonso Incoronato, Francesca Madonini, Arianna Adelaide Maurina," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "Chiara Nonne, Simone Riccardo, and Federica Villa. Historical perspectives, state of art and research trends of SPAD arrays and their applications (Part II: SPAD arrays). Frontiers in Physics, 10:906671, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "type": "text", + "content": "[12] Iman Esmaeil Zadeh, J Chang, Johannes WN Los, Samuel Gyger, Ali W Elshaari, Stephan Steinhauer, Sander N Dorenbos, and Val Zwiller. Superconducting nanowire single-photon detectors: A perspective on evolution, state-of-the-art, future developments, and applications. Applied Physics Letters, 118(19), 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 184, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 248 + ], + "type": "text", + "content": "[13] Andrew L Goertzen, Xuezhu Zhang, Megan M McClarty, Eric J Berg, Chen-Yi Liu, Piotr Kozlowski, Fabrice Retiere, Lawrence Ryner, Vesna Sossi, Greg Stortz, et al. Design and performance of a resistor multiplexing readout circuit for a sipm detector. IEEE Transactions on Nuclear Science, 60 (3):1541-1549, 2013. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 250, + 553, + 293 + ], + "type": "text", + "content": "[14] Rémi Gribonval, Antoine Chatalic, Nicolas Keriven, Vincent Schellekens, Laurent Jacques, and Philip Schniter. Sketching data sets for large-scale learning. IEEE Signal Process. Mag., 38(5):12-36, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "text", + "content": "[15] Yanqiu Guan, Haochen Li, Li Xue, Rui Yin, Labao Zhang, Hao Wang, Guanghao Zhu, Lin Kang, Jian Chen, and Peiheng Wu. Lidar with superconducting nanowire single-photon detectors: Recent advances and developments. Optics and Lasers in Engineering, 156:107102, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 403 + ], + "type": "text", + "content": "[16] Yanqiu Guan, Haochen Li, Labao Zhang, Hao Wang, Guanglong He, Biao Zhang, Yue Fei, Jiayu Lv, Xiao Zhang, Rui Yin, et al. SNSPD array with single-channel readout based on compressive sensing. ACS Photonics, 9(9):3102-3109, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 404, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 469 + ], + "type": "text", + "content": "[17] Hao Hao, Qing-Yuan Zhao, Yang-Hui Huang, Jie Deng, Fan Yang, Sai-Ying Ru, Zhen Liu, Chao Wan, Hao Liu, Zhi-Jian Li, et al. A compact multi-pixel superconducting nanowire single-photon detector array supporting gigabit space-to-ground communications. Light: Science & Applications, 13(1):25, 2024. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 471, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 536 + ], + "type": "text", + "content": "[18] Sam W. Hutchings, Nick Johnston, Istvan Gyongy, Tarek Al Abbas, Neale A. W. Dutton, Max Tyler, Susan Chan, Jonathan Leach, and Robert K. Henderson. A reconfigurable 3-D-stacked SPAD imager with in-pixel histogramming for flash LIDAR or high-speed time-of-flight imaging. IEEE J. Solid-State Circuits, 54(11):2947-2956, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 537, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 569 + ], + "type": "text", + "content": "[19] Atul Ingle and David Maier. Count-free single-photon 3d imaging with race logic. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-12, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 571, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 646 + ], + "type": "text", + "content": "[20] Jørgen Arendt Jensen, Mikkel Schou, Lasse Thurmann Jørgensen, Borislav G Tomov, Matthias Bo Stuart, Marie Sand Traberg, Iman Taghavi, Sigrid Huesebø Øygaard, Martin Lind Ommen, Kitty Steenberg, et al. Anatomic and functional imaging using row-column arrays. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 69(10):2722-2738, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[21] Ling-Dong Kong, Hui Wang, Qing-Yuan Zhao, Jia-Wei Guo, Yang-Hui Huang, Hao Hao, Shi Chen, Xue-Cou Tu, LaBao Zhang, Xiao-Qing Jia, et al. Readout-efficient superconducting nanowire single-photon imager with orthogonal time-amplitude multiplexing by hotspot quantization. Nature Photonics, 17(1):65-72, 2023. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11414" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 138 + ], + "type": "text", + "content": "[22] Sangwon Lee, Yong Choi, Jihoon Kang, and Jin Ho Jung. Development of a multiplexed readout with high position resolution for positron emission tomography. *Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment*, 850:42-47, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": "[23] Feiyan Li, Hang Han, Qi Chen, Biao Zhang, Han Bao, Yue Dai, Rui Ge, Shuya Guo, Guanglong He, Yue Fei, et al. Saturation efficiency for detecting " + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "inline_equation", + "content": "1550\\mathrm{nm}" + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": " photons with a " + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "inline_equation", + "content": "2\\times 2" + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": " array of " + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "inline_equation", + "content": "\\mathrm{Mo}_{0.8}\\mathrm{Si}_{0.2}" + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": " nanowires at " + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "inline_equation", + "content": "2.2\\mathrm{K}" + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": ". Photonics Research, 9(3):389-394, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[24] Adam N McCaughan, Yao Zhai, Boris Korzh, Jason P Allmaras, Bakhrom G Oripov, Matthew D Shaw, and Sae Woo Nam. The thermally coupled imager: A scalable readout architecture for superconducting nanowire single photon detectors. Applied Physics Letters, 121(10), 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 274 + ], + "type": "text", + "content": "[25] Kazuhiro Morimoto. Megapixel spad cameras for time-resolved applications. Technical report, EPFL, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[26] Peter D Olcott, Garry Chinn, and Craig S Levin. Compressed sensing for the multiplexing of pet detectors. In 2011 IEEE Nuclear Science Symposium Conference Record, pages 3224-3226. IEEE, 2011. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "type": "text", + "content": "[27] B. G. Oripov, D. S. Rampini, J. Allmaras, M. D. Shaw, S. W. Nam, B. Korzh, and A. N. McCaughan. A superconducting nanowire single-photon camera with 400,000 pixels. Nature, 622(7984):730-734, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "text", + "content": "[28] François Piron, Daniel Morrison, Mehmet Rasit Yuce, and Jean-Michel Redoute. A review of single-photon avalanche diode time-of-flight imaging sensor arrays. IEEE Sensors Journal, 21(11):12654-12666, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "text", + "content": "[29] Joshua Rapp, Julian Tachella, Yoann Altmann, Stephen McLaughlin, and Vivek K Goyal. Advances in single-photon lidar for autonomous vehicles: Working principles, challenges, and recent advances. IEEE Signal Processing Magazine, 37(4):62-71, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "text", + "content": "[30] Morten Fischer Rasmussen and Jørgen Arendt Jensen. 3-d ultrasound imaging performance of a row-column addressed 2-d array transducer: A measurement study. In 2013 IEEE International Ultrasonics Symposium (IUS), pages 1460-1463. IEEE, 2013. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 589 + ], + "type": "text", + "content": "[31] Morten Fischer Rasmussen, Thomas Lehrmann Christiansen, Erik Vilain Thomsen, and Jørgen Arendt Jensen. 3-d imaging using row-column-addressed arrays with integrated apodization-part i: apodization design and line element beamforming. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 62(5):947-958, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "text", + "content": "[32] Kaustubh Sadekar, David Maier, and Atul Ingle. Single-photon 3d imaging with equi-depth photon histograms. In European Conference on Computer Vision, pages 381-398. Springer, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[33] Ewout van den Berg, Emmanuel Candes, Garry Chinn, Craig Levin, Peter Demetri Olcott, and Carlos Sing-Long. Single-photon sampling architecture for solid-state imaging sensors. Proceedings of the National Academy of Sciences, 110(30): E2752-E2761, 2013. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[34] Hailu Wang, Jiaxiang Guo, Jinshui Miao, Wenjin Luo, Yue Gu, Runzhang Xie, Fang Wang, Lili Zhang, Peng Wang, and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 362 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Weida Hu. Emerging single-photon detectors based on low-dimensional materials. Small, 18(5):2103963, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "text", + "content": "[35] Hui Wang, Zhi-jian Li, Xue-Mei Hu, Hao Hao, Jia-wei Guo, Yang-hui Huang, Hao Liu, Chao Wan, Xue-cou Tu, Xiaoqing Jia, et al. Image distortion by ambiguous multiple-photon detections in a superconducting nanowire single-photon imager and the correction method. Optics Express, 31(14):23579-23588, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[36] Mel White, Shahaboddin Ghajari, Tianyi Zhang, Akshit Dave, Ashok Veeraraghavan, and Alyosha Molnar. A differential SPAD array architecture in " + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "inline_equation", + "content": "0.18\\mu \\mathrm{m}" + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": " CMOS for HDR imaging. In Proc. IEEE Int. Symp. Circuits Syst., pages 292-296, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "type": "text", + "content": "[37] Emma E Wollman, Varun B Verma, Adriana E Lita, William H Farr, Matthew D Shaw, Richard P Mirin, and Sae Woo Nam. Kilopixel array of superconducting nanowire single-photon detectors. Optics express, 27(24):35279-35289, 2019. 1, 2, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 306 + ], + "type": "text", + "content": "[38] Lixing You. Superconducting nanowire single-photon detectors for quantum information. Nanophotonics, 9(9):2673-2692, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 362 + ], + "type": "text", + "content": "[39] Qing-Yuan Zhao, Di Zhu, Niccolò Calandri, Andrew E Dane, Adam N McCaughan, Francesco Bellei, Hao-Zhu Wang, Daniel F Santavicca, and Karl K Berggren. Single-photon imager based on a superconducting nanowire delay line. Nature Photonics, 11(4):247-251, 2017. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11415" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/9a0c23ba-5904-4499-9efc-651298f2730b_content_list.json b/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/9a0c23ba-5904-4499-9efc-651298f2730b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5cbe24b0e9d9582e927a06434067fc14b607dc73 --- /dev/null +++ b/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/9a0c23ba-5904-4499-9efc-651298f2730b_content_list.json @@ -0,0 +1,1434 @@ +[ + { + "type": "text", + "text": "Image Referenced Sketch Colorization Based on Animation Creation Workflow", + "text_level": 1, + "bbox": [ + 84, + 130, + 883, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Dingkun Yan $^{1,4}$ *Xinrui Wang $^{2}$ Zhuoru Li $^{3}$ Suguru Saito $^{1}$ Yusuke Iwasawa $^{2}$ Yutaka Matsuo $^{2}$ Jiaxian Guo $^{2}$ $^{1}$ Institute of Science Tokyo $^{2}$ The University of Tokyo $^{3}$ Project HAT $^{4}$ Tokyo AI research institute, Shanda Group", + "bbox": [ + 112, + 178, + 854, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e0091afec1a93cd7b8c9266bd97134a6d57390dfd66e2db1369fe5d10eced25a.jpg", + "image_caption": [ + "Figure 1. Given reference images, our proposed method automatically synthesizes high-quality sketch colorization results that loyally match the reference color distribution and are free from artifacts." + ], + "image_footnote": [], + "bbox": [ + 81, + 282, + 879, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 643, + 313, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sketch colorization plays an important role in animation and digital illustration production tasks. However, existing methods still meet problems in that text-guided methods fail to provide accurate color and style reference, hint-guided methods still involve manual operation, and image-referenced methods are prone to cause artifacts. To address these limitations, we propose a diffusion-based framework inspired by real-world animation production workflows. Our approach leverages the sketch as the spatial guidance and an RGB image as the color reference, and separately extracts foreground and background from the reference image with spatial masks. Particularly, we introduce a split cross-attention mechanism with LoRA (Low-Rank Adaptation) modules. They are trained separately with fore", + "bbox": [ + 75, + 662, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ground and background regions to control the corresponding embeddings for keys and values in cross-attention. This design allows the diffusion model to integrate information from foreground and background independently, preventing interference and eliminating the spatial artifacts. During inference, we design switchable inference modes for diverse use scenarios by changing modules activated in the framework. Extensive qualitative and quantitative experiments, along with user studies, demonstrate our advantages over existing methods in generating high-quality artifact-free results with geometric mismatched references. Ablation studies further confirm the effectiveness of each component. Codes are available at https://github.com/tellurion-kanata/colorizeDiffusion.", + "bbox": [ + 496, + 643, + 893, + 854 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Represent equal contribution to this work", + "bbox": [ + 94, + 886, + 321, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23391", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d444f1b8b08f5514704c7d9821a8926bab78dad6329b409e4156664df957b53a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 90, + 264, + 132 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/28f13251e764c1557df73577bbfdf08f0c3b2419754f778ec004b37207b41a0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 90, + 441, + 132 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/877a2029ad82697833538309bc10a7546290af58865b82eebe76e2c24d8c4eb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 90, + 616, + 132 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ab7c7af59bc6d4f05e8b009910c0d8064cf133de4dfb5710ad51e90eb8598252.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 90, + 877, + 132 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/94d1f7bdde6af9ef001882e5f69c5df7987a1cbe411e8960d9c9c85cb7a7cb67.jpg", + "image_caption": [ + "A. Reference image" + ], + "image_footnote": [], + "bbox": [ + 89, + 133, + 256, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/627caa231ff18ff07384a95a8db74436ec8d933d0b623fc6213c2d1d492b77cb.jpg", + "image_caption": [ + "B. Sketch" + ], + "image_footnote": [], + "bbox": [ + 267, + 133, + 433, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c9dee82eeb61eed3ecedd12d7ae1436811a6d4eaf29a8cf8b4fc685a2af46524.jpg", + "image_caption": [ + "C. Colorized foreground" + ], + "image_footnote": [], + "bbox": [ + 446, + 133, + 609, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/485bf55fed4ff739074f76847625103fbe69e1744b0b3ad15c2906d17d3b583e.jpg", + "image_caption": [ + "D. Colorized full frame" + ], + "image_footnote": [], + "bbox": [ + 622, + 133, + 870, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 318, + 207, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The past decades have witnessed the development of Animation as an artistic form, which has gained great popularity worldwide. The current workflow of animation creation is labor-intensive, and the growing demands of animation from the market are causing animation studios to fall short of hands, bringing severe problems to the industry.", + "bbox": [ + 75, + 338, + 468, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The most manual labor-intensive procedure in animation production is sketch colorization, and animators working on colorization also take up the largest share among all employees in the animation production industry. To reduce the human labor needed and to automate the animation production, machine learning algorithms have been applied for sketch colorization [26, 27, 55, 56, 59]. However, current methods are still not optimal for real-world production pipelines: Text-based colorization methods [50, 57] fail to provide accurate guidance on the color and style information of the images. User-guided methods [5, 55] still involve manual operation in the process, making them less efficient. Image referenced methods [2, 47] can be seamlessly integrated into the current pipeline, but the spatial mismatches between reference images and sketches are causing severe artifacts and unexpected extra objects, which is termed as spatial entanglement in [47] and shown in Figure 2.", + "bbox": [ + 75, + 430, + 468, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To build a sketch colorization framework that meets the requirements of the real-world animation production pipeline, we start with the observation of the manual sketch colorization workflow in real-world animation production. As is shown in Figure 3, the workflow consists of the following key steps: Firstly, character designers design the characters as references. Secondly, senior animators draw the sketches for each frame. Thirdly, animators colorize the figures in the sketches according to the character designs. Finally, animators colorize the background of the sketches and finish the whole colorized frames.", + "bbox": [ + 75, + 688, + 468, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following our observation, we designed a diffusion-based framework to mimic the sketch colorization workflow step-by-step. To be smoothly integrated into the current", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/af20d50c548a392eaafe1160acfcbc14b4903a8e2696d3eced6ea9b54e8a3332.jpg", + "image_caption": [ + "Figure 3. Illustration of colorization workflow in professional animation studios. A: character designers design characters as references. B: Senior animators draw the sketches for the key frames. C: animators colorize the figures in the sketches according to the character designs, and D: animators colorize the background of the sketches and merge foreground and background into finished frames.", + "A. Sketch", + "Figure 2. Illustration of spatial entanglement. We use red rectangles to highlight the spatial entangled artifacts in the result of the IP-Adapter baseline, where additional arms appear unexpectedly, and the model mistakenly synthesizes long hair." + ], + "image_footnote": [], + "bbox": [ + 504, + 316, + 580, + 402 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1442e1b0058a141bb2994a98061157e1b7c039098a5cecdae4b0cb98cd73eb36.jpg", + "image_caption": [ + "B. Reference Image" + ], + "image_footnote": [], + "bbox": [ + 588, + 316, + 663, + 402 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/07fdff89b55293bee673384c1787d6c29158a242a19578d0c0523ba9250d8df3.jpg", + "image_caption": [ + "C. Result of IP Adapter Baseline" + ], + "image_footnote": [], + "bbox": [ + 666, + 316, + 777, + 402 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6dee62e479956009839e5d901d5c05ff8a09bb18674e0f2be8d03815de3fff61.jpg", + "image_caption": [ + "D. Result of Proposed Method" + ], + "image_footnote": [], + "bbox": [ + 790, + 316, + 870, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "production workflow, the proposed framework leverages a sketch image as the spatial reference and an RGB image as the color reference. Specifically, we use the high dimensional local tokens extracted by a feature extractor as reference embeddings to maintain the semantic information and adopt a multi-layer sketch encoder for precise spatial control of the background embeddings. To enable the separate colorization of foreground and background regions, we design a novel split cross-attention mechanism, where spatial masks are used to segment figures as foreground with the rest of the images as background, and corresponding LoRA weights are trained to modify the embeddings for keys and values within cross-attention layers. This design allows the diffusion model to integrate information from foreground and background independently, preventing interference and eliminating the need to adjust the well-trained backbone weights. During inference, we implement a switchable LoRA mechanism that provides precise control over the colorization process and enables different inference modes for various scenarios without changing model weights.", + "bbox": [ + 496, + 503, + 890, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments show that mimicking the real-world animation creation workflow yields several advantages. In qualitative analyses, our method synthesizes high-quality results loyally representing the color distribution of the reference images and free from artifacts and spatial entanglements. Quantitative comparisons also validate the superiority of the", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23392", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proposed method over existing methods by common criteria and benchmarks. What's more, user studies illustrate that candidates prefer our method subjectively.", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows: (1) We propose an image-referenced sketch colorization framework that can synthesize high-quality results free from artifacts and spatial entanglement by mimicking the animation production workflow. (2) We design a novel split cross-attention mechanism that enables the separate colorization of foreground and background in a single forward pass and the switchable LoRA module that allows users to switch colorization modes during inference. (3) Experiments show that our method outperforms existing methods in qualitative/quantitative comparisons and is preferred by human users in perceptive user study.", + "bbox": [ + 75, + 137, + 470, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 332, + 218, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Latent Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 358, + 302, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Probabilistic Models [14, 41] are a class of latent variable models inspired by nonequilibrium thermodynamics [39] and have achieved great success in image synthesis and editing. Compared to Generative Adversarial Networks (GANs) [6, 7, 10, 19, 20], Diffusion Models excel at generating highly realistic images with various contexts and able to be controlled by text prompts. However, the autoregressive denoising process of diffusion models, typically computed with a U-Net [36] or a Diffusion Transformer (DiT) [3, 32], incurs substantial computational costs.", + "bbox": [ + 75, + 382, + 468, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this limitation, Stable Diffusion (SD) [33, 35] as a class of Latent Diffusion Models (LDMs) was proposed, where a two-stage synthesis mechanism was adopted to enable diffusion/denoising process to be performed on a highly compressed latent space with a pair of pretrained Variational Autoencoder (VAE), so as to significantly reduce computational costs. Concurrently, different researches of diffusion samplers have been conducted and proved to be effective in accelerating the denoising process [29, 30, 40, 41]. We adopt SD as our neural backbone, utilize the DPM++ solver [18, 30, 41] as the default sampler, and employ classifier-free guidance [9, 15] to strengthen the reference-based performance.", + "bbox": [ + 75, + 534, + 470, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Image Prompted Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 739, + 379, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Currently, most diffusion models for image synthesis tasks are based on text prompts [3, 32, 33, 35]. However, there are tasks where text prompts can not provide enough information to precisely guide the image synthesis and editing, such as image-to-image translation [25], style transfer [44,58], colorization [2,47] and image composition [22,51], and thus images are also used as prompts to provide reference information. The reference information extracted from prompt images varies from tasks: style transfer tasks adopt", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the textures and colors from reference images, image composition tasks focus more on the object-related information, and sketch colorization requires all the above.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are two common practices to combine image prompts with diffusion models. Given image embedding vectors extracted by pre-trained feature extraction networks, existing methods either train an adapter module to inject the reference embedding vector into the backbone [31, 48] or directly inject reference information into the backbone with attention layers [16]. However, both of them [16, 31, 48] may introduce loss or mismatch of structure when inputs are not well paired, resulting in performance deterioration. In the sketch colorization task, specifically, these adapters provide conflicting spatial information from references and lead to unacceptable artifacts, as illustrated in Figure 2.", + "bbox": [ + 496, + 137, + 893, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Sketch Colorization", + "text_level": 1, + "bbox": [ + 500, + 339, + 689, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sketch colorization has been a long-standing topic in computer vision. Interactive optimization-based method [42] was employed for the task, and deep-learning-based methods [2, 21, 27, 47, 56] later became the mainstream due to the ability to synthesize high-quality and high-resolution images. There are three main technical solutions for deep learning methods: text-prompted sketch colorization [21, 46, 57], user-guided sketch colorization [53, 56] and image reference sketch colorization [2, 27, 46]. User-guided methods can precisely colorize given sketches with detailed guidance from users, but the manual labor needed makes them unsuitable to be integrated into an automatic workflow. Text-prompted methods have received great popularity over recent years due to the development of Text-to-Image diffusion models, but it is challenging to precisely control colors, textures, and styles using text prompts. Image-referenced methods also benefit from the development of diffusion models, along with relevant works that enable image control [24, 31, 48, 54, 57]. However, the mismatch of reference images and sketches still results in severe deterioration. ColorizeDiffusion [47] achieved notable progress in the colorization quality, yet it still suffers from spatial entanglement, which is shown in Figure 2. In this paper, we base our method on the production pipeline of animation studios to use image references to guide colorization. We separate the foreground and background with an innovative switchable LoRA to improve colorization performance and prevent artifacts.", + "bbox": [ + 496, + 363, + 893, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 791, + 591, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by our observation of real-world animation production, we propose an image-referenced sketch colorization framework. As is shown in Figure 4, it leverages a sketch image $\\mathbf{X}_s \\in \\mathbb{R}^{w_s \\times h_s \\times 1}$ , a reference image $\\mathbf{X}_r \\in \\mathbb{R}^{w_r \\times h_r \\times c}$ and a foreground mask $\\mathbf{X}_m \\in \\mathbb{R}^{w_s \\times h_s \\times 1}$ as inputs, and returns the colorized result $\\mathbf{Y} \\in \\mathbb{R}^{w_s \\times h_s \\times c}$ , with", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23393", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a38364fe23f6ae418982939e429bb4d8278bdba527d6e4920054cd914c5333d2.jpg", + "image_caption": [ + "Figure 4. Illustration of the proposed framework. We use reference masks to separate reference images into foreground and background and CLIP Image encoder $\\phi$ to extract both regions into embeddings. The background embeddings first go through the recovery transformer $\\varphi$ to recover detailed information, then concatenated with foreground embeddings as final K and V inputs for split cross-attention. Similar to Eq 2, the compose operation is a spatial piece-wise function employed to separate foreground and background." + ], + "image_footnote": [], + "bbox": [ + 86, + 92, + 880, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$w, h$ and $c$ representing the width, height and channel of the images. All components of the framework are based on the animation production workflow with the following design:", + "bbox": [ + 75, + 325, + 468, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, to mimic the animator's colorizing sketches with character designs as references, we utilize a pre-trained vision transformer (ViT) to extract image embeddings as reference information. The embeddings are later injected into the diffusion backbone with split cross-attention layers. Secondly, to integrate the sketches into the framework, we adopt a multi-layer sketch encoder to inject sketch information into the latent layers of the diffusion backbone as spatial guidance. Thirdly, based on the behavior of animators separately colorizing the foreground and background of the sketch, we propose a novel split cross-attention mechanism that uses spatial masks to separate the trainable LoRA modules corresponding to the keys and values for foreground and background for training. A switchable LoRA mechanism is then applied during inference for different application scenarios with different colorization modes. More details are described in the supplementary materials.", + "bbox": [ + 75, + 371, + 472, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Pretrain of the Diffusion Backbone", + "text_level": 1, + "bbox": [ + 76, + 636, + 382, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The backbone of the proposed framework consists of a pre-trained VAE, a sketch encoder, a denoising U-Net, and a pre-trained Vision Transformer (ViT) functioning as the image encoder from OpenCLIP-H [4, 17, 34, 37]. We denote sketch images, reference images, and ground truth as $s$ , $r$ , and $y$ , respectively. The VAE encoder, U-Net, and ViT are represented by $\\mathcal{E}$ , $\\theta$ , and $\\phi$ , respectively. The timestep $t$ starts from $T - 1$ and goes to 0, where $T$ is the diffusion steps, set to 1000. The training objective of the diffusion model is to denoise the intermediate noisy image $z_{t}$ via noise prediction:", + "bbox": [ + 75, + 659, + 468, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {\\mathcal {E} (y), \\epsilon , t, s, r} [ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, s, \\phi (r)) \\| _ {2} ^ {2} ]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 830, + 468, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To pre-train the diffusion backbone, We initialize the VAE and U-Net with WaifuDiffusion [11] and train networks with a dynamic reference drop rate decreasing from", + "bbox": [ + 76, + 854, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "80% to 50% as training progresses to avoid the distribution shift mentioned by [47]. VAE, U-Net, and sketch encoder are then frozen during the training of the full framework.", + "bbox": [ + 498, + 325, + 892, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Color Reference Extraction", + "text_level": 1, + "bbox": [ + 500, + 378, + 746, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the real-world animation production workflow, we utilize images as color references for sketch colorization. The commonly used ViT-based image encoder networks have two kinds of output embeddings: the CLS embeddings $\\pmb{E}_{\\mathbf{cls}} \\in \\mathbb{R}^{bs \\times 1 \\times 1024}$ and the local embeddings $\\pmb{E}_{local} \\in \\mathbb{R}^{bs \\times 256 \\times 1024}$ , where $bs$ represent the batch size. The CLS embeddings are projected to CLIP embedding space for image-text contrastive learning, with spatial information compressed and connected to text-level notions, and are employed as color or style references by previous image-guided methods [44, 48]. ColorizeDiffusion [47], on the contrary, reveals that local embeddings also express text-level semantics, indicating that they express more details regarding textures, strokes, and styles, enabling the network to generate better reference-based results, especially for transferring detailed textures and strokes. Therefore, the proposed method follows [47] to adopt local tokens as color reference inputs for the framework.", + "bbox": [ + 498, + 401, + 892, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the excessive spatial information of image semantics and compositions contained in local embeddings leads to frequent occurrences of artifacts such as overflow of color regions and unexpected objects outside sketches. Illustrated in Figure 2, such artifacts widely exist in frameworks with image references [31,44,48]. To eliminate this problem, we follow the real-world workflow to explicitly separate the foreground and background with spatial masks during colorization and describe the detail in 3.3 and 3.4.", + "bbox": [ + 496, + 674, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Split Cross-Attention", + "text_level": 1, + "bbox": [ + 500, + 816, + 699, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In anime images, the foreground regions and background regions differ distinctively in color distribution, color block sizes, tones, and textures. Thus, the colorization of foreground and background is separated into two independent", + "bbox": [ + 498, + 839, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23394", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/23094763c74170299b6c33246fa37046a86d187d8219e4859e446cd295307407.jpg", + "image_caption": [ + "Figure 5. Based on the LoRA weights, the proposed method can merge the foreground and background features in one forward pass and switch between three inference modes. We denote the dimension of pre-trained weights as CH. The rank of foreground LoRA is fixed at 16, while the rank of background LoRA is $0.5^{*}\\mathrm{CH}$ ." + ], + "image_footnote": [], + "bbox": [ + 83, + 92, + 459, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "steps in the animation production workflow. Following this scheme, we propose a novel split cross-attention mechanism to substitute the cross-attention layers in the diffusion backbone to separately process foreground and background regions with different parameters in a single forward pass.", + "bbox": [ + 75, + 321, + 468, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A split cross-attention layer consists of two groups of trainable LoRA weights $W_{f}^{t}$ and $W_{b}^{t}$ , which include query weights $W_{f}^{q}$ and $W_{b}^{q}$ , key weights $W_{f}^{k}$ and $W_{b}^{k}$ , and value weights $W_{f}^{v}$ and $W_{b}^{v}$ for foreground and background QKV projection respectively. An open-sourced animation image segmentation tool [38] is used to automatically extract the foreground mask $m_{s}$ and $m_{r}$ of sketches and reference images. The spatial masks can be adjusted by tuning the user defined thresholds $ts_{s}$ and $ts_{r}$ . Regions with pixel values larger than $ts_{s}$ and $ts_{r}$ are considered as foreground, otherwise background. For foreground LoRAs, we set the ranks as 16; for background LoRA, the rank is formulated as $r = 0.5 * \\min(D_{q}, D_{kv})$ , where $D_{q}$ and $D_{kv}$ are dimensions of queries and keys/values for the corresponding cross-attention layers.", + "bbox": [ + 75, + 398, + 468, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As foreground and background regions of animation images feature different textures, color distribution, and color tones, injecting the embeddings of foreground and background directly into split cross-attention results in deterioration in structure preservation, synthesis quality, and stylization. Therefore, we further add a trainable recovery transformer $\\varphi$ to process the background embeddings and facilitate better integration of the foreground and background reference information into the diffusion backbone.", + "bbox": [ + 75, + 628, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define query inputs (forward features) as $z_{f}$ , $z_{b}$ , key and value inputs (reference embeddings) as $e$ , $e_{b}$ , attention outputs as $y$ in the following sections, where the index $f$ and $b$ indicate foreground and background respectively. Specifically, $e$ denotes the reference embeddings extracted from the whole reference image $r$ , formulated as $e = \\phi(\\boldsymbol{r})$ ; and $e_{b} = \\varphi(\\phi(\\boldsymbol{r}_{b}))$ , where $r_{b}$ is the background region of the reference image. During training, the proposed split cross-attention can be formulated as follows:", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} = \\left\\{ \\begin{array}{l l} \\operatorname {S o f t m a x} \\left(\\frac {\\left(\\hat {W} _ {f} ^ {q} z _ {f}\\right) \\cdot \\left(\\hat {W} _ {f} ^ {k} e\\right)}{d}\\right) \\left(\\hat {W} _ {f} ^ {v} e\\right) & \\text {i f} \\boldsymbol {m} _ {s} > t s _ {s} \\\\ \\operatorname {S o f t m a x} \\left(\\frac {\\left(\\hat {W} _ {b} ^ {q} z _ {b}\\right) \\cdot \\left(\\hat {W} _ {b} ^ {k} e _ {b}\\right)}{d}\\right) \\left(\\hat {W} _ {b} ^ {v} e _ {b}\\right) & \\text {i f} \\boldsymbol {m} _ {s} \\leq t s _ {s} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 114, + 890, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{W}_f^t = W^t + W_f^t$ , and $W^t$ represents the pretrained weights, which remain frozen during training. Similarly, $\\hat{W}_b^t$ follows the same approach.", + "bbox": [ + 498, + 167, + 890, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Switchable inference mode", + "text_level": 1, + "bbox": [ + 500, + 224, + 741, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The application scenarios and the sketch-reference combinations in the real world may be complicated. Also, naively separating the foreground and background regions for colorization degrades the quality of background synthesis, especially when reference images have severe semantic mismatches with the sketches or have complicated backgrounds. Therefore, we design three different inference modes: Vanilla, Bg2Fig, and Fig2Fig for different scenarios based on the weights and reference inputs used for KV calculation. We visualize all inference modes in Figure 5.", + "bbox": [ + 496, + 250, + 890, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vanilla mode only utilizes the pre-trained weights for cross-attention modules, with $W_{f}^{t}$ , $W_{b}^{t}$ and recovery transformers deactivated. It's suitable for most scenarios but suffers from spatial entanglement when references are figure-only images.", + "bbox": [ + 496, + 400, + 890, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$Bg2Fig$ mode activates only the foreground-related LoRA weights $W_{f}^{t}$ during inference and is used when reference images are figure images with complicated backgrounds. This mode outperforms Vanilla mode in character colorization and Fig2Fig mode in background generation as its foreground weights are further optimized by LoRAs.", + "bbox": [ + 496, + 476, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig2Fig is designed for figure-to-figure colorization, where reference images are figures with simple background composition. This mode activates both LoRA weights $W_{f}^{t}$ , $W_{b}^{t}$ , and the recovery transformer and uses spatial masks to separate foreground/background embeddings for calculating query and key/value inputs. It effectively eliminates the spatial entanglement in reference-based sketch colorization.", + "bbox": [ + 496, + 566, + 890, + 675 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 685, + 624, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 710, + 714, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Danbooru2021 dataset [8] is used to train and validate the proposed method. We select and process 4.8 million triples of (sketch, color, mask) images with various contents to train the framework, and use another 50,000 image triples excluded from training set for validation. Sketches were extracted by jointly using [52] and [45]. 4x H100 (94GB) GPUs are used to train the network. We at first pre-train the denoising U-Net and sketch encoder for 6 epochs with 8 hours per epoch, and then froze the backbone and trained the recovery transformer and switchable LORAs for 3 epochs with 9 hours per epoch. AdamW op", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23395", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/240c847aedf74beda6dadc0c9f89081707f76ec1945f60b1e8cc8dd9414be6bc.jpg", + "image_caption": [ + "Figure 6. Results of the ablation study. The baseline model demonstrates significant spatial entanglement; incorporating split cross-attention reduces artifacts, the trainable LoRAs improve color saturation and details, and the proposed complete pipeline produces high-quality results free of artifacts. Zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 470, + 228 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "timer [23, 28] is used to optimize the network with learning rate 0.0001 and betas (0.9, 0.999). Following [47], we dropped at least $50\\%$ of the reference inputs in all training.", + "bbox": [ + 75, + 321, + 468, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 378, + 225, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Split cross-attention. The proposed method aims to address spatial entanglement by simulating the animation workflow. To demonstrate the effectiveness of this workflow, we set up three frameworks: 1) baseline model without split cross attention, trainable LoRAs, and recovery transformer, 2) baseline model with split cross attention but no trainable LoRAs and recovery transformer, and 3) the proposed full framework.", + "bbox": [ + 75, + 401, + 468, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show the qualitative comparison in Figure 6 to validate the effectiveness of the proposed modules. The baseline model causes severe spatial entanglement in generating additional figures in (a) and undesired clothes in (b). The application of split cross-attention mitigates the spatial entanglement but still causes artifacts and degrades the color saturation and details of the results. Collaborating split cross attention with trainable LoRAs improves the quality of results and further improves the background, but still suffers from artifacts. The proposed full framework enhanced by recovery transformers effectively eliminates spatial entanglement and synthesizes colorization results that have clear boundaries and rich details and textures, and loyally preserves the color distribution of reference images.", + "bbox": [ + 75, + 522, + 468, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference modes. We illustrate the differences between inference modes and their use cases in Figure 7. Fig2Fig mode is fully mask-guided, enabling it to eliminate spatial entanglement shown in (a). However, it is less suitable for inpainting character sketches, as the region without the mask guide may suffer from the lack of reference information. Bg2Fig performs similarly to the Vanilla mode. With the help of foreground LoRA weight, the results demonstrate clearer segmentation and better stroke quality. All modes perform well for landscape sketches, which contain more structural detail than figure sketches, making them", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/05fa92683bb929f62a4a6d3328a9e061d0c9188da3341c6ebb81a85f634ccbf3.jpg", + "image_caption": [ + "Figure 7. Colorization results with three different inference modes. Fig2Fig mode performs better in eliminating spatial entanglement, while Bg2Fig and Vanilla mode can generate vivid backgrounds and inpainting results." + ], + "image_footnote": [], + "bbox": [ + 509, + 90, + 888, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "easier to colorize.", + "bbox": [ + 500, + 362, + 619, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show the comparison of local embeddings and CLS embeddings for color reference extraction in the supplementary materials.", + "bbox": [ + 498, + 377, + 890, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with baseline", + "text_level": 1, + "bbox": [ + 500, + 434, + 735, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with six existing reference-based sketch colorization methods [2,46,47] to demonstrate the superiority of the proposed framework.", + "bbox": [ + 498, + 458, + 890, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline introduction. Two baseline methods are the combination of SD [35], ControlNet [54, 57], and IP-Adatper [48]. They adopt different cross-attention scales during denoising, labeled as IP-Adapter and InstantStyle, respectively. IP-adapter baseline generates results with normal cross-attention scales, while InstantStyle generates results using the \"style transfer\" weight type, which is claimed to prevent composition transfer by setting specific cross-attention scales to 0 according to [44]. Following the official document of [54], we adopted Anything v3 [49], a community-developed model, as the SD backbone for IP-Adapter-H and ControlNet_lineart_ANime. All these models are officially implemented and claimed to be effective for anime-style image generation. The T2I-Adapter baseline simply replaces IP-Adapter with T2I-Adapter-Style [31, 43]. Specifically, we introduced quality-related prompts and textual inversion, such as \"masterpiece\" and \"easynegative\" [12], for T2I-model baselines to improve their image quality.", + "bbox": [ + 496, + 513, + 892, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparison. We show the qualitative comparison of our proposed method and existing methods in Figure 8, where most of the existing methods suffer from spatial entanglement in various cases. In rows (a)-(d), existing methods (2)-(6) failed to distinguish foreground and background regions and, therefore, synthesized artifacts in", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23396", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5afc4bdb54454e2a45c18a45a23efa8173c233d20816e48910bd2b210ef0ffa0.jpg", + "image_caption": [ + "Figure 8. Qualitative comparisons between our proposed method and existing methods show that our results are visually more appealing than those of image-prompt adapters [31, 44, 48] and the GAN-based method [46]. Compared to ColorizeDiffusion [47], the proposed framework eliminates spatial entanglement and improves overall quality." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 883, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the background. Such artifacts become more obvious when reference images have complicated backgrounds in rows (e) and (f), where IP-Adapter and InstantStyle mixed reference composition with the sketch composition, making the generated results visually messy. The GAN-based method (7) successfully generated results with clear backgrounds, but the color preservation, texture transformation, and detail qualities are much poorer than diffusion-based methods.", + "bbox": [ + 75, + 566, + 468, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also show the comparison of our proposed method and adapter-based methods on semantically non-relevant sketch-reference pairs in Figure 9. Both IP-Adapter and T2I-Adapter fail to separate foreground from background and colorize the whole image with the same tone and texture. Our proposed method, on the contrary, generated results with clear region boundaries, rich texture and details, and visually pleasant bright color, with the help of the proposed spatial aware split cross-attention mechanism and switchable LoRA. The qualitative comparisons demonstrate that our proposed method is effective for arbitrary input sketch-reference pairs and achieves high-quality colorization in various use scenarios.", + "bbox": [ + 75, + 688, + 470, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative comparison. Fréchet Inception Distance", + "bbox": [ + 76, + 885, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(FID) [13] is a widely used quantitative cretera to evaluate the performance of image synthesis tasks. It calculates the perceptual distance of two distributions without requiring them to be semantically and spatially paired. We conduct a quantitative evaluation measured by FID on the entire validation set which includes 50000 (sketch, reference) image pairs. Sketches are colorized with randomly selected reference images, ensuring each batch has semantically and spatially mismatched inputs.", + "bbox": [ + 496, + 566, + 890, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-scale structural similarity index measure (MS-SSIM), peak signal-to-noise ratio (PSNR), and CLIP score [17,34] assess the similarity between processed images and given ground truth. It requires the guiding reference to be aligned with the ground truth when applied to image-referenced sketch colorization. To fulfill this, we selected 5000 color images as ground truth, extracted the sketches from and used thin plate spline (TPS) transformation to spatially distort them as color references to build a test set of 5000 triples of sketch, reference, and ground truth, and evaluate all the 7 methods with MS-SSIM, PSNR and CLIP score on this dataset.", + "bbox": [ + 496, + 703, + 892, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show the results of the quantitative comparison in Ta", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23397", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d01143c2f03bb7b181d505e05f6b2b249385f6600a78520c3248df73143bbb47.jpg", + "table_caption": [ + "Table 1. Quantitative comparison between the proposed model and baseline methods. We calculated 50K FID, 5K PSNR, 5K MS-SSIM, and 5K CLIP cosine similarity of image embeddings in this experiment. †: FID evaluation randomly selected color images as references, making it close to real-application scenarios. ‡: Ground truth color images were deformed to obtain semantically paired and spatially similar references for evaluations. We also show the inference time on H100 GPU at 512*512 resolution and batch size 4." + ], + "table_footnote": [], + "table_body": "
Method†50K-FID ↓‡PSNR↑‡MS-SSIM↑‡CLIP similarity↑Inference time(ms)
Ours6.832728.91440.60020.88291513
ColorizeDiffusion [47]9.527628.73840.59130.87751387
IP-Adapter [48,49,57]38.918428.67670.54780.86722617
InstantStyle [44,48,49,57]40.814428.10900.44590.80422586
T2I-Adapter [31,49,57]41.156928.12750.32430.71802301
AnimeDiffusion [2]61.599927.84540.31850.73192073
Yan et al. [46]27.014729.24830.52530.763457
", + "bbox": [ + 99, + 148, + 864, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f6f851ef4786ee33ab34f7268e6a77b90ec3e3a5be56ef317400825fca773cb8.jpg", + "image_caption": [ + "Figure 9. Colorization results for semantically non-relevant sketch-reference pairs. Adapter-based methods fail to generate clear borders and high-quality textures, while our proposed method synthesizes visually satisfying results." + ], + "image_footnote": [], + "bbox": [ + 81, + 290, + 465, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ble 1, where the proposed method significantly outperforms in FID, MM-SSIM, and CLIP similarity due to better texture and color synthesis. GAN-base method [46] achieved the best score in PSNR, with the proposed method ranked number 2. This is because the limited generation ability of [46] prevents it from generating complicated backgrounds and also hinders it from synthesizing bright colors and rich details of the figures. The close-to-average results make it advantageous to the calculation of PSNR [1].", + "bbox": [ + 75, + 611, + 468, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User study. The quality of sketch colorization is highly subjective and easily influenced by personal preference, we thus demonstrate how different individuals evaluate the proposed method and existing methods through an user study. 40 participants are invited to select the best results with two criteria: the overall colorization quality and the preservation of geometric structure of the sketches. We prepare 25 image sets and show each participant 16 image sets for evaluation. For each image set, participants are presented the colorization results of our proposed method as well as those", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/471c7f08bd5f46b71008c58edd4797e42142540641a1976958cda5847ab84f69.jpg", + "image_caption": [ + "Figure 10. Results of user study. Our method is preferred across all shown methods in overall quality and geometric preservation." + ], + "image_footnote": [], + "bbox": [ + 504, + 290, + 890, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "generated by six existing methods.", + "bbox": [ + 500, + 494, + 730, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results of the user study are presented in Figure 10, where our proposed method has received the most numbers of preferences among all the methods presented. To further validate the comparison, we applied the Kruskal-Wallis test as a statistical method. The results clearly demonstrate that the proposed method significantly outperforms all the existing methods in terms of user preference with a significance level of $p < 0.01$ . All the images shown in the user study are included in the supplementary materials.", + "bbox": [ + 496, + 510, + 890, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 659, + 619, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed an image-guided sketch colorization framework inspired by real-world animation creation workflow. We implement a novel split cross-attention layer with spatial masks and corresponding LoRA weights that facilitates separate processing for foreground and background, eliminating the spatial entanglement, and also design a switchable LoRA mechanism that enables users to choose different inference modes for various use cases.", + "bbox": [ + 496, + 685, + 890, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, the proposed methods strongly depend on the quality of extracted masks, and a corresponding failure case is given in the supplementary materials. Our future work focuses on further improving the similarity with references and extending the framework for video colorization.", + "bbox": [ + 496, + 806, + 890, + 881 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23398", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 8", + "[2] Yu Cao, Xiangqiao Meng, P. Y. Mok, Tong-Yee Lee, Xueting Liu, and Ping Li. Animediffusion: Anime diffusion colorization. TVCG, pages 1-14, 2024. 2, 3, 6, 8", + "[3] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 3", + "[4] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 4", + "[5] Youngin Cho, Junsoo Lee, Soyoung Yang, Juntae Kim, Yeojeong Park, Haneol Lee, Mohammad Azam Khan, Daesik Kim, and Jaegul Choo. Guiding users to where to give color hints for efficient interactive sketch colorization via unsupervised region prioritization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1818-1827, 2023. 2", + "[6] Yunjey Choi, Min-Je Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In CVPR, pages 8789-8797. IEEE/CVF, 2018. 3", + "[7] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In CVPR, pages 8185-8194. IEEE/CVF, 2020. 3", + "[8] Danbooru community, Gwern Branwen, and Anonymous. Danbooru2021: A large-scale crowdsourced and tagged anime illustration dataset. https://gwern.net/ danbooru2021, 2022. Accessed: DATE 2022-01-21. 5", + "[9] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, pages 8780-8794, 2021. 3", + "[10] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, pages 2672–2680, 2014. 3", + "[11] Reimu Hakurei. Hugging face/waifu-diffusion-v1-4. https://huggingface.co/hakurei/waifu-diffusion-v1-4, 2023. Accessed: DATE 2023-03-05. 4", + "[12] Havoc. Easynegative. https://civitai.com/models/7808/easynegative, 2023. Accessed: DATE 2023-02-10. 6", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, pages 6626-6637, 2017. 7", + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. CoRR, abs/2207.12598, 2022. 3", + "[16] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. 3", + "[17] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. https://github.com/mlfoundations/open Clip, jul 2021. 4, 7", + "[18] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors, NeurIPS, 2022. 3", + "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410. IEEE/CVF, 2019. 3", + "[20] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, pages 8107-8116. IEEE/CVF, 2020. 3", + "[21] Hyunsu Kim, Ho Young Jhoo, Eunhyeok Park, and Sungjoo Yoo. Tag2pix: Line art colorization using text tag with scat and changing loss. In ICCV, pages 9055-9064. IEEE/CVF, 2019. 3", + "[22] Kangyeol Kim, Sunghyun Park, Junsoo Lee, and Jaegul Choo. Reference-based image composition with sketch via structure-aware diffusion model. arXiv preprint arXiv:2304.09748, 2023. 3", + "[23] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6", + "[24] Kohya-ss. Hugging face/controlnet-lllite. https://huggingface.co/kohya-ss/controlnet-111ite, 2024. Accessed: DATE 2024-01-02. 3", + "[25] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. In ICLR. OpenReview.net, 2023. 3", + "[26] Junsoo Lee, Eungyeup Kim, Yunsung Lee, Dongjun Kim, Jaehyuk Chang, and Jaegul Choo. Reference-based sketch image colorization using augmented-self reference and dense semantic correspondence. In CVPR, pages 5800-5809. IEEE/CVF, 2020. 2", + "[27] Zekun Li, Zhengyang Geng, Zhao Kang, Wenyu Chen, and Yibo Yang. Eliminating gradient conflict in reference-based line-art colorization. In ECCV, pages 579-596. Springer, 2022. 2, 3", + "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*. OpenReview.net, 2019. 6", + "[29] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022. 3", + "[30] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23399", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "guided sampling of diffusion probabilistic models. CoRR, abs/2211.01095, 2022. 3", + "[31] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. CoRR, abs/2302.08453, 2023. 3, 4, 6, 7, 8", + "[32] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4172-4182. IEEE, 2023. 3", + "[33] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. CoRR, abs/2307.01952, 2023. 3", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, volume 139, pages 8748-8763. PMLR, 2021. 4, 7", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685. IEEE/CVF, 2022. 3, 6", + "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, volume 9351, pages 234-241. Springer, 2015. 3", + "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 4", + "[38] SkyTNT, infoengine1337, and not lain. anime-segmentation. https : / / github . com / SkyTNT / anime - segmentation, 2022.5", + "[39] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, volume 37, pages 2256-2265. JMLR.org, 2015. 3", + "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR. OpenReview.net, 2021. 3", + "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR. OpenReview.net, 2021. 3", + "[42] Daniel Sykora, John Dingliana, and Steven Collins. Lazybrush: Flexible painting tool for hand-drawn cartoons. Comput. Graph. Forum, 28(2):599-608, 2009. 3", + "[43] TencentARC. Hugging face/ip-adapter. https://github.com/TencentARC/T2I-Adapter/tree/SD, 2024. Accessed: DATE 2024-01-02. 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Haofan Wang, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 3, 4, 6, 7, 8", + "[45] Xiaoyu Xiang, Ding Liu, Xiao Yang, Yiheng Zhu, Xiaohui Shen, and Jan P. Allebach. Adversarial open domain adaptation for sketch-to-photo synthesis. In WACV, pages 944-954. IEEE/CVF, 2022. 5", + "[46] Dingkun Yan, Ryogo Ito, Ryo Moriai, and Suguru Saito. Two-step training: Adjustable sketch colourization via reference image and text tag. Computer Graphics Forum, 2023. 3, 6, 7, 8", + "[47] Dingkun Yan, Liang Yuan, Yuma Nishioka, Issei Fujishiro, and Suguru Saito. Colorized diffusion: Adjustable sketch colorization with reference image and text. 2024. 2, 3, 4, 6, 7, 8", + "[48] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. CoRR, abs/2308.06721, 2023. 3, 4, 6, 7, 8", + "[49] Yuno779. https://civitai.com/models/9409, 2023. Accessed: DATE 2023-06-25. 6,8", + "[50] Nir Zabari, Aharon Azulay, Alexey Gorkor, Tavi Halperin, and Ohad Fried. Diffusing colors: Image colorization with text guided diffusion. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 2", + "[51] Bo Zhang, Yuxuan Duan, Jun Lan, Yan Hong, Huijia Zhu, Weiqiang Wang, and Li Niu. Controlcom: Controllable image composition using diffusion model. arXiv preprint arXiv:2308.10040, 2023. 3", + "[52] Lvmin Zhang. Sketchkeras. https://github.com/11lyasviel/sketchKeras, 2017.5", + "[53] Lvmin Zhang. Style2paints v5, 2023. Accessed: DATE 2023-06-25. 3", + "[54] Lvmin Zhang. Controlnet-v1-1-nightly. https://github.com/lllyasviel/ControlNet-v1-1-nightly, 2024. Accessed: DATE 2024-01-02. 3,6", + "[55] Lvmin Zhang, Chengze Li, Edgar Simo-Serra, Yi Ji, Tien-Tsin Wong, and Chunping Liu. User-guided line art flat filling with split filling mechanism. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9889-9898, 2021. 2", + "[56] Lvmin Zhang, Chengze Li, Tien-Tsin Wong, Yi Ji, and Chunping Liu. Two-stage sketch colorization. ACM Trans. Graph., 37(6):261, 2018. 2, 3", + "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 2, 3, 6, 8", + "[58] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 3", + "[59] Changqing Zou, Haoran Mo, Chengying Gao, Ruofei Du, and Hongbo Fu. Language-based colorization of scene sketches. ACM Trans. Graph., 38(6), 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 885 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23400", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/9a0c23ba-5904-4499-9efc-651298f2730b_model.json b/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/9a0c23ba-5904-4499-9efc-651298f2730b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cd9e69b1ca747b5ee0cd6e4e814bbfcf002e5ee0 --- /dev/null +++ b/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/9a0c23ba-5904-4499-9efc-651298f2730b_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.086, + 0.131, + 0.885, + 0.152 + ], + "angle": 0, + "content": "Image Referenced Sketch Colorization Based on Animation Creation Workflow" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.179, + 0.856, + 0.253 + ], + "angle": 0, + "content": "*Dingkun Yan\\(^{1,4}\\) *Xinrui Wang\\(^{2}\\) Zhuoru Li\\(^{3}\\) Suguru Saito\\(^{1}\\) Yusuke Iwasawa\\(^{2}\\) Yutaka Matsuo\\(^{2}\\) Jiaxian Guo\\(^{2}\\) \\(^{1}\\)Institute of Science Tokyo \\(^{2}\\)The University of Tokyo \\(^{3}\\)Project HAT \\(^{4}\\)Tokyo AI research institute, Shanda Group" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.284, + 0.88, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.609, + 0.893, + 0.636 + ], + "angle": 0, + "content": "Figure 1. Given reference images, our proposed method automatically synthesizes high-quality sketch colorization results that loyally match the reference color distribution and are free from artifacts." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.644, + 0.314, + 0.658 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Sketch colorization plays an important role in animation and digital illustration production tasks. However, existing methods still meet problems in that text-guided methods fail to provide accurate color and style reference, hint-guided methods still involve manual operation, and image-referenced methods are prone to cause artifacts. To address these limitations, we propose a diffusion-based framework inspired by real-world animation production workflows. Our approach leverages the sketch as the spatial guidance and an RGB image as the color reference, and separately extracts foreground and background from the reference image with spatial masks. Particularly, we introduce a split cross-attention mechanism with LoRA (Low-Rank Adaptation) modules. They are trained separately with fore" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.895, + 0.856 + ], + "angle": 0, + "content": "ground and background regions to control the corresponding embeddings for keys and values in cross-attention. This design allows the diffusion model to integrate information from foreground and background independently, preventing interference and eliminating the spatial artifacts. During inference, we design switchable inference modes for diverse use scenarios by changing modules activated in the framework. Extensive qualitative and quantitative experiments, along with user studies, demonstrate our advantages over existing methods in generating high-quality artifact-free results with geometric mismatched references. Ablation studies further confirm the effectiveness of each component. Codes are available at https://github.com/tellurion-kanata/colorizeDiffusion." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.322, + 0.9 + ], + "angle": 0, + "content": "* Represent equal contribution to this work" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23391" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.092, + 0.266, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.092, + 0.442, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.092, + 0.617, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.092, + 0.878, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.134, + 0.258, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.252, + 0.223, + 0.263 + ], + "angle": 0, + "content": "A. Reference image" + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.134, + 0.434, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.252, + 0.374, + 0.262 + ], + "angle": 0, + "content": "B. Sketch" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.134, + 0.611, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.467, + 0.252, + 0.588, + 0.263 + ], + "angle": 0, + "content": "C. Colorized foreground" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.134, + 0.872, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.253, + 0.802, + 0.262 + ], + "angle": 0, + "content": "D. Colorized full frame" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.892, + 0.306 + ], + "angle": 0, + "content": "Figure 3. Illustration of colorization workflow in professional animation studios. A: character designers design characters as references. B: Senior animators draw the sketches for the key frames. C: animators colorize the figures in the sketches according to the character designs, and D: animators colorize the background of the sketches and merge foreground and background into finished frames." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.319, + 0.208, + 0.333 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.469, + 0.429 + ], + "angle": 0, + "content": "The past decades have witnessed the development of Animation as an artistic form, which has gained great popularity worldwide. The current workflow of animation creation is labor-intensive, and the growing demands of animation from the market are causing animation studios to fall short of hands, bringing severe problems to the industry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.469, + 0.687 + ], + "angle": 0, + "content": "The most manual labor-intensive procedure in animation production is sketch colorization, and animators working on colorization also take up the largest share among all employees in the animation production industry. To reduce the human labor needed and to automate the animation production, machine learning algorithms have been applied for sketch colorization [26, 27, 55, 56, 59]. However, current methods are still not optimal for real-world production pipelines: Text-based colorization methods [50, 57] fail to provide accurate guidance on the color and style information of the images. User-guided methods [5, 55] still involve manual operation in the process, making them less efficient. Image referenced methods [2, 47] can be seamlessly integrated into the current pipeline, but the spatial mismatches between reference images and sketches are causing severe artifacts and unexpected extra objects, which is termed as spatial entanglement in [47] and shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.853 + ], + "angle": 0, + "content": "To build a sketch colorization framework that meets the requirements of the real-world animation production pipeline, we start with the observation of the manual sketch colorization workflow in real-world animation production. As is shown in Figure 3, the workflow consists of the following key steps: Firstly, character designers design the characters as references. Secondly, senior animators draw the sketches for each frame. Thirdly, animators colorize the figures in the sketches according to the character designs. Finally, animators colorize the background of the sketches and finish the whole colorized frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Following our observation, we designed a diffusion-based framework to mimic the sketch colorization workflow step-by-step. To be smoothly integrated into the current" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.317, + 0.581, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.41, + 0.578, + 0.42 + ], + "angle": 0, + "content": "A. Sketch" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.318, + 0.665, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.405, + 0.657, + 0.426 + ], + "angle": 0, + "content": "B. Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.318, + 0.779, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.404, + 0.77, + 0.424 + ], + "angle": 0, + "content": "C. Result of IP Adapter Baseline" + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.318, + 0.871, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.405, + 0.878, + 0.426 + ], + "angle": 0, + "content": "D. Result of Proposed Method" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.436, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Figure 2. Illustration of spatial entanglement. We use red rectangles to highlight the spatial entangled artifacts in the result of the IP-Adapter baseline, where additional arms appear unexpectedly, and the model mistakenly synthesizes long hair." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.807 + ], + "angle": 0, + "content": "production workflow, the proposed framework leverages a sketch image as the spatial reference and an RGB image as the color reference. Specifically, we use the high dimensional local tokens extracted by a feature extractor as reference embeddings to maintain the semantic information and adopt a multi-layer sketch encoder for precise spatial control of the background embeddings. To enable the separate colorization of foreground and background regions, we design a novel split cross-attention mechanism, where spatial masks are used to segment figures as foreground with the rest of the images as background, and corresponding LoRA weights are trained to modify the embeddings for keys and values within cross-attention layers. This design allows the diffusion model to integrate information from foreground and background independently, preventing interference and eliminating the need to adjust the well-trained backbone weights. During inference, we implement a switchable LoRA mechanism that provides precise control over the colorization process and enables different inference modes for various scenarios without changing model weights." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Experiments show that mimicking the real-world animation creation workflow yields several advantages. In qualitative analyses, our method synthesizes high-quality results loyally representing the color distribution of the reference images and free from artifacts and spatial entanglements. Quantitative comparisons also validate the superiority of the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23392" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "proposed method over existing methods by common criteria and benchmarks. What's more, user studies illustrate that candidates prefer our method subjectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.471, + 0.319 + ], + "angle": 0, + "content": "In summary, our contributions are as follows: (1) We propose an image-referenced sketch colorization framework that can synthesize high-quality results free from artifacts and spatial entanglement by mimicking the animation production workflow. (2) We design a novel split cross-attention mechanism that enables the separate colorization of foreground and background in a single forward pass and the switchable LoRA module that allows users to switch colorization modes during inference. (3) Experiments show that our method outperforms existing methods in qualitative/quantitative comparisons and is preferred by human users in perceptive user study." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.333, + 0.22, + 0.349 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.359, + 0.303, + 0.374 + ], + "angle": 0, + "content": "2.1. Latent Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.534 + ], + "angle": 0, + "content": "Diffusion Probabilistic Models [14, 41] are a class of latent variable models inspired by nonequilibrium thermodynamics [39] and have achieved great success in image synthesis and editing. Compared to Generative Adversarial Networks (GANs) [6, 7, 10, 19, 20], Diffusion Models excel at generating highly realistic images with various contexts and able to be controlled by text prompts. However, the autoregressive denoising process of diffusion models, typically computed with a U-Net [36] or a Diffusion Transformer (DiT) [3, 32], incurs substantial computational costs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.471, + 0.732 + ], + "angle": 0, + "content": "To address this limitation, Stable Diffusion (SD) [33, 35] as a class of Latent Diffusion Models (LDMs) was proposed, where a two-stage synthesis mechanism was adopted to enable diffusion/denoising process to be performed on a highly compressed latent space with a pair of pretrained Variational Autoencoder (VAE), so as to significantly reduce computational costs. Concurrently, different researches of diffusion samplers have been conducted and proved to be effective in accelerating the denoising process [29, 30, 40, 41]. We adopt SD as our neural backbone, utilize the DPM++ solver [18, 30, 41] as the default sampler, and employ classifier-free guidance [9, 15] to strengthen the reference-based performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.741, + 0.38, + 0.758 + ], + "angle": 0, + "content": "2.2. Image Prompted Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Currently, most diffusion models for image synthesis tasks are based on text prompts [3, 32, 33, 35]. However, there are tasks where text prompts can not provide enough information to precisely guide the image synthesis and editing, such as image-to-image translation [25], style transfer [44,58], colorization [2,47] and image composition [22,51], and thus images are also used as prompts to provide reference information. The reference information extracted from prompt images varies from tasks: style transfer tasks adopt" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "the textures and colors from reference images, image composition tasks focus more on the object-related information, and sketch colorization requires all the above." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.895, + 0.333 + ], + "angle": 0, + "content": "There are two common practices to combine image prompts with diffusion models. Given image embedding vectors extracted by pre-trained feature extraction networks, existing methods either train an adapter module to inject the reference embedding vector into the backbone [31, 48] or directly inject reference information into the backbone with attention layers [16]. However, both of them [16, 31, 48] may introduce loss or mismatch of structure when inputs are not well paired, resulting in performance deterioration. In the sketch colorization task, specifically, these adapters provide conflicting spatial information from references and lead to unacceptable artifacts, as illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.34, + 0.691, + 0.355 + ], + "angle": 0, + "content": "2.3. Sketch Colorization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.895, + 0.787 + ], + "angle": 0, + "content": "Sketch colorization has been a long-standing topic in computer vision. Interactive optimization-based method [42] was employed for the task, and deep-learning-based methods [2, 21, 27, 47, 56] later became the mainstream due to the ability to synthesize high-quality and high-resolution images. There are three main technical solutions for deep learning methods: text-prompted sketch colorization [21, 46, 57], user-guided sketch colorization [53, 56] and image reference sketch colorization [2, 27, 46]. User-guided methods can precisely colorize given sketches with detailed guidance from users, but the manual labor needed makes them unsuitable to be integrated into an automatic workflow. Text-prompted methods have received great popularity over recent years due to the development of Text-to-Image diffusion models, but it is challenging to precisely control colors, textures, and styles using text prompts. Image-referenced methods also benefit from the development of diffusion models, along with relevant works that enable image control [24, 31, 48, 54, 57]. However, the mismatch of reference images and sketches still results in severe deterioration. ColorizeDiffusion [47] achieved notable progress in the colorization quality, yet it still suffers from spatial entanglement, which is shown in Figure 2. In this paper, we base our method on the production pipeline of animation studios to use image references to guide colorization. We separate the foreground and background with an innovative switchable LoRA to improve colorization performance and prevent artifacts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.792, + 0.593, + 0.807 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Inspired by our observation of real-world animation production, we propose an image-referenced sketch colorization framework. As is shown in Figure 4, it leverages a sketch image \\( \\mathbf{X}_s \\in \\mathbb{R}^{w_s \\times h_s \\times 1} \\), a reference image \\( \\mathbf{X}_r \\in \\mathbb{R}^{w_r \\times h_r \\times c} \\) and a foreground mask \\( \\mathbf{X}_m \\in \\mathbb{R}^{w_s \\times h_s \\times 1} \\) as inputs, and returns the colorized result \\( \\mathbf{Y} \\in \\mathbb{R}^{w_s \\times h_s \\times c} \\), with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23393" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.093, + 0.881, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.894, + 0.313 + ], + "angle": 0, + "content": "Figure 4. Illustration of the proposed framework. We use reference masks to separate reference images into foreground and background and CLIP Image encoder \\(\\phi\\) to extract both regions into embeddings. The background embeddings first go through the recovery transformer \\(\\varphi\\) to recover detailed information, then concatenated with foreground embeddings as final K and V inputs for split cross-attention. Similar to Eq 2, the compose operation is a spatial piece-wise function employed to separate foreground and background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.47, + 0.371 + ], + "angle": 0, + "content": "\\(w, h\\) and \\(c\\) representing the width, height and channel of the images. All components of the framework are based on the animation production workflow with the following design:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.473, + 0.629 + ], + "angle": 0, + "content": "Firstly, to mimic the animator's colorizing sketches with character designs as references, we utilize a pre-trained vision transformer (ViT) to extract image embeddings as reference information. The embeddings are later injected into the diffusion backbone with split cross-attention layers. Secondly, to integrate the sketches into the framework, we adopt a multi-layer sketch encoder to inject sketch information into the latent layers of the diffusion backbone as spatial guidance. Thirdly, based on the behavior of animators separately colorizing the foreground and background of the sketch, we propose a novel split cross-attention mechanism that uses spatial masks to separate the trainable LoRA modules corresponding to the keys and values for foreground and background for training. A switchable LoRA mechanism is then applied during inference for different application scenarios with different colorization modes. More details are described in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.637, + 0.383, + 0.652 + ], + "angle": 0, + "content": "3.1. Pretrain of the Diffusion Backbone" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.827 + ], + "angle": 0, + "content": "The backbone of the proposed framework consists of a pre-trained VAE, a sketch encoder, a denoising U-Net, and a pre-trained Vision Transformer (ViT) functioning as the image encoder from OpenCLIP-H [4, 17, 34, 37]. We denote sketch images, reference images, and ground truth as \\(s\\), \\(r\\), and \\(y\\), respectively. The VAE encoder, U-Net, and ViT are represented by \\(\\mathcal{E}\\), \\(\\theta\\), and \\(\\phi\\), respectively. The timestep \\(t\\) starts from \\(T - 1\\) and goes to 0, where \\(T\\) is the diffusion steps, set to 1000. The training objective of the diffusion model is to denoise the intermediate noisy image \\(z_{t}\\) via noise prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.832, + 0.47, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {\\mathcal {E} (y), \\epsilon , t, s, r} [ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, s, \\phi (r)) \\| _ {2} ^ {2} ]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To pre-train the diffusion backbone, We initialize the VAE and U-Net with WaifuDiffusion [11] and train networks with a dynamic reference drop rate decreasing from" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.326, + 0.893, + 0.372 + ], + "angle": 0, + "content": "80% to 50% as training progresses to avoid the distribution shift mentioned by [47]. VAE, U-Net, and sketch encoder are then frozen during the training of the full framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.379, + 0.748, + 0.394 + ], + "angle": 0, + "content": "3.2. Color Reference Extraction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.402, + 0.893, + 0.674 + ], + "angle": 0, + "content": "Following the real-world animation production workflow, we utilize images as color references for sketch colorization. The commonly used ViT-based image encoder networks have two kinds of output embeddings: the CLS embeddings \\( \\pmb{E}_{\\mathbf{cls}} \\in \\mathbb{R}^{bs \\times 1 \\times 1024} \\) and the local embeddings \\( \\pmb{E}_{local} \\in \\mathbb{R}^{bs \\times 256 \\times 1024} \\), where \\( bs \\) represent the batch size. The CLS embeddings are projected to CLIP embedding space for image-text contrastive learning, with spatial information compressed and connected to text-level notions, and are employed as color or style references by previous image-guided methods [44, 48]. ColorizeDiffusion [47], on the contrary, reveals that local embeddings also express text-level semantics, indicating that they express more details regarding textures, strokes, and styles, enabling the network to generate better reference-based results, especially for transferring detailed textures and strokes. Therefore, the proposed method follows [47] to adopt local tokens as color reference inputs for the framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.81 + ], + "angle": 0, + "content": "However, the excessive spatial information of image semantics and compositions contained in local embeddings leads to frequent occurrences of artifacts such as overflow of color regions and unexpected objects outside sketches. Illustrated in Figure 2, such artifacts widely exist in frameworks with image references [31,44,48]. To eliminate this problem, we follow the real-world workflow to explicitly separate the foreground and background with spatial masks during colorization and describe the detail in 3.3 and 3.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.7, + 0.833 + ], + "angle": 0, + "content": "3.3. Split Cross-Attention" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "In anime images, the foreground regions and background regions differ distinctively in color distribution, color block sizes, tones, and textures. Thus, the colorization of foreground and background is separated into two independent" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23394" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.46, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.244, + 0.47, + 0.314 + ], + "angle": 0, + "content": "Figure 5. Based on the LoRA weights, the proposed method can merge the foreground and background features in one forward pass and switch between three inference modes. We denote the dimension of pre-trained weights as CH. The rank of foreground LoRA is fixed at 16, while the rank of background LoRA is \\(0.5^{*}\\mathrm{CH}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.469, + 0.399 + ], + "angle": 0, + "content": "steps in the animation production workflow. Following this scheme, we propose a novel split cross-attention mechanism to substitute the cross-attention layers in the diffusion backbone to separately process foreground and background regions with different parameters in a single forward pass." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.627 + ], + "angle": 0, + "content": "A split cross-attention layer consists of two groups of trainable LoRA weights \\( W_{f}^{t} \\) and \\( W_{b}^{t} \\), which include query weights \\( W_{f}^{q} \\) and \\( W_{b}^{q} \\), key weights \\( W_{f}^{k} \\) and \\( W_{b}^{k} \\), and value weights \\( W_{f}^{v} \\) and \\( W_{b}^{v} \\) for foreground and background QKV projection respectively. An open-sourced animation image segmentation tool [38] is used to automatically extract the foreground mask \\( m_{s} \\) and \\( m_{r} \\) of sketches and reference images. The spatial masks can be adjusted by tuning the user defined thresholds \\( ts_{s} \\) and \\( ts_{r} \\). Regions with pixel values larger than \\( ts_{s} \\) and \\( ts_{r} \\) are considered as foreground, otherwise background. For foreground LoRAs, we set the ranks as 16; for background LoRA, the rank is formulated as \\( r = 0.5 * \\min(D_{q}, D_{kv}) \\), where \\( D_{q} \\) and \\( D_{kv} \\) are dimensions of queries and keys/values for the corresponding cross-attention layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.764 + ], + "angle": 0, + "content": "As foreground and background regions of animation images feature different textures, color distribution, and color tones, injecting the embeddings of foreground and background directly into split cross-attention results in deterioration in structure preservation, synthesis quality, and stylization. Therefore, we further add a trainable recovery transformer \\(\\varphi\\) to process the background embeddings and facilitate better integration of the foreground and background reference information into the diffusion backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We define query inputs (forward features) as \\( z_{f} \\), \\( z_{b} \\), key and value inputs (reference embeddings) as \\( e \\), \\( e_{b} \\), attention outputs as \\( y \\) in the following sections, where the index \\( f \\) and \\( b \\) indicate foreground and background respectively. Specifically, \\( e \\) denotes the reference embeddings extracted from the whole reference image \\( r \\), formulated as \\( e = \\phi(\\boldsymbol{r}) \\); and \\( e_{b} = \\varphi(\\phi(\\boldsymbol{r}_{b})) \\), where \\( r_{b} \\) is the background region of the reference image. During training, the proposed split cross-attention can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.116, + 0.892, + 0.159 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} = \\left\\{ \\begin{array}{l l} \\operatorname {S o f t m a x} \\left(\\frac {\\left(\\hat {W} _ {f} ^ {q} z _ {f}\\right) \\cdot \\left(\\hat {W} _ {f} ^ {k} e\\right)}{d}\\right) \\left(\\hat {W} _ {f} ^ {v} e\\right) & \\text {i f} \\boldsymbol {m} _ {s} > t s _ {s} \\\\ \\operatorname {S o f t m a x} \\left(\\frac {\\left(\\hat {W} _ {b} ^ {q} z _ {b}\\right) \\cdot \\left(\\hat {W} _ {b} ^ {k} e _ {b}\\right)}{d}\\right) \\left(\\hat {W} _ {b} ^ {v} e _ {b}\\right) & \\text {i f} \\boldsymbol {m} _ {s} \\leq t s _ {s} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.169, + 0.892, + 0.217 + ], + "angle": 0, + "content": "where \\(\\hat{W}_f^t = W^t + W_f^t\\), and \\(W^t\\) represents the pretrained weights, which remain frozen during training. Similarly, \\(\\hat{W}_b^t\\) follows the same approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.226, + 0.742, + 0.241 + ], + "angle": 0, + "content": "3.4. Switchable inference mode" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.892, + 0.4 + ], + "angle": 0, + "content": "The application scenarios and the sketch-reference combinations in the real world may be complicated. Also, naively separating the foreground and background regions for colorization degrades the quality of background synthesis, especially when reference images have severe semantic mismatches with the sketches or have complicated backgrounds. Therefore, we design three different inference modes: Vanilla, Bg2Fig, and Fig2Fig for different scenarios based on the weights and reference inputs used for KV calculation. We visualize all inference modes in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.476 + ], + "angle": 0, + "content": "Vanilla mode only utilizes the pre-trained weights for cross-attention modules, with \\( W_{f}^{t} \\), \\( W_{b}^{t} \\) and recovery transformers deactivated. It's suitable for most scenarios but suffers from spatial entanglement when references are figure-only images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.568 + ], + "angle": 0, + "content": "\\(Bg2Fig\\) mode activates only the foreground-related LoRA weights \\(W_{f}^{t}\\) during inference and is used when reference images are figure images with complicated backgrounds. This mode outperforms Vanilla mode in character colorization and Fig2Fig mode in background generation as its foreground weights are further optimized by LoRAs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.676 + ], + "angle": 0, + "content": "Fig2Fig is designed for figure-to-figure colorization, where reference images are figures with simple background composition. This mode activates both LoRA weights \\( W_{f}^{t} \\), \\( W_{b}^{t} \\), and the recovery transformer and uses spatial masks to separate foreground/background embeddings for calculating query and key/value inputs. It effectively eliminates the spatial entanglement in reference-based sketch colorization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.686, + 0.625, + 0.703 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.715, + 0.727 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The Danbooru2021 dataset [8] is used to train and validate the proposed method. We select and process 4.8 million triples of (sketch, color, mask) images with various contents to train the framework, and use another 50,000 image triples excluded from training set for validation. Sketches were extracted by jointly using [52] and [45]. 4x H100 (94GB) GPUs are used to train the network. We at first pre-train the denoising U-Net and sketch encoder for 6 epochs with 8 hours per epoch, and then froze the backbone and trained the recovery transformer and switchable LORAs for 3 epochs with 9 hours per epoch. AdamW op" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23395" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.091, + 0.472, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.239, + 0.47, + 0.308 + ], + "angle": 0, + "content": "Figure 6. Results of the ablation study. The baseline model demonstrates significant spatial entanglement; incorporating split cross-attention reduces artifacts, the trainable LoRAs improve color saturation and details, and the proposed complete pipeline produces high-quality results free of artifacts. Zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.469, + 0.369 + ], + "angle": 0, + "content": "timer [23, 28] is used to optimize the network with learning rate 0.0001 and betas (0.9, 0.999). Following [47], we dropped at least \\(50\\%\\) of the reference inputs in all training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.379, + 0.227, + 0.394 + ], + "angle": 0, + "content": "4.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.469, + 0.522 + ], + "angle": 0, + "content": "Split cross-attention. The proposed method aims to address spatial entanglement by simulating the animation workflow. To demonstrate the effectiveness of this workflow, we set up three frameworks: 1) baseline model without split cross attention, trainable LoRAs, and recovery transformer, 2) baseline model with split cross attention but no trainable LoRAs and recovery transformer, and 3) the proposed full framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.469, + 0.735 + ], + "angle": 0, + "content": "We show the qualitative comparison in Figure 6 to validate the effectiveness of the proposed modules. The baseline model causes severe spatial entanglement in generating additional figures in (a) and undesired clothes in (b). The application of split cross-attention mitigates the spatial entanglement but still causes artifacts and degrades the color saturation and details of the results. Collaborating split cross attention with trainable LoRAs improves the quality of results and further improves the background, but still suffers from artifacts. The proposed full framework enhanced by recovery transformers effectively eliminates spatial entanglement and synthesizes colorization results that have clear boundaries and rich details and textures, and loyally preserves the color distribution of reference images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Inference modes. We illustrate the differences between inference modes and their use cases in Figure 7. Fig2Fig mode is fully mask-guided, enabling it to eliminate spatial entanglement shown in (a). However, it is less suitable for inpainting character sketches, as the region without the mask guide may suffer from the lack of reference information. Bg2Fig performs similarly to the Vanilla mode. With the help of foreground LoRA weight, the results demonstrate clearer segmentation and better stroke quality. All modes perform well for landscape sketches, which contain more structural detail than figure sketches, making them" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.091, + 0.89, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.291, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Figure 7. Colorization results with three different inference modes. Fig2Fig mode performs better in eliminating spatial entanglement, while Bg2Fig and Vanilla mode can generate vivid backgrounds and inpainting results." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.363, + 0.62, + 0.375 + ], + "angle": 0, + "content": "easier to colorize." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.378, + 0.892, + 0.423 + ], + "angle": 0, + "content": "We show the comparison of local embeddings and CLS embeddings for color reference extraction in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.435, + 0.736, + 0.451 + ], + "angle": 0, + "content": "4.3. Comparison with baseline" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.459, + 0.892, + 0.504 + ], + "angle": 0, + "content": "We compare our method with six existing reference-based sketch colorization methods [2,46,47] to demonstrate the superiority of the proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.893, + 0.801 + ], + "angle": 0, + "content": "Baseline introduction. Two baseline methods are the combination of SD [35], ControlNet [54, 57], and IP-Adatper [48]. They adopt different cross-attention scales during denoising, labeled as IP-Adapter and InstantStyle, respectively. IP-adapter baseline generates results with normal cross-attention scales, while InstantStyle generates results using the \"style transfer\" weight type, which is claimed to prevent composition transfer by setting specific cross-attention scales to 0 according to [44]. Following the official document of [54], we adopted Anything v3 [49], a community-developed model, as the SD backbone for IP-Adapter-H and ControlNet_lineart_ANime. All these models are officially implemented and claimed to be effective for anime-style image generation. The T2I-Adapter baseline simply replaces IP-Adapter with T2I-Adapter-Style [31, 43]. Specifically, we introduced quality-related prompts and textual inversion, such as \"masterpiece\" and \"easynegative\" [12], for T2I-model baselines to improve their image quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Qualitative comparison. We show the qualitative comparison of our proposed method and existing methods in Figure 8, where most of the existing methods suffer from spatial entanglement in various cases. In rows (a)-(d), existing methods (2)-(6) failed to distinguish foreground and background regions and, therefore, synthesized artifacts in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23396" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.885, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.511, + 0.895, + 0.555 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparisons between our proposed method and existing methods show that our results are visually more appealing than those of image-prompt adapters [31, 44, 48] and the GAN-based method [46]. Compared to ColorizeDiffusion [47], the proposed framework eliminates spatial entanglement and improves overall quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.687 + ], + "angle": 0, + "content": "the background. Such artifacts become more obvious when reference images have complicated backgrounds in rows (e) and (f), where IP-Adapter and InstantStyle mixed reference composition with the sketch composition, making the generated results visually messy. The GAN-based method (7) successfully generated results with clear backgrounds, but the color preservation, texture transformation, and detail qualities are much poorer than diffusion-based methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.884 + ], + "angle": 0, + "content": "We also show the comparison of our proposed method and adapter-based methods on semantically non-relevant sketch-reference pairs in Figure 9. Both IP-Adapter and T2I-Adapter fail to separate foreground from background and colorize the whole image with the same tone and texture. Our proposed method, on the contrary, generated results with clear region boundaries, rich texture and details, and visually pleasant bright color, with the help of the proposed spatial aware split cross-attention mechanism and switchable LoRA. The qualitative comparisons demonstrate that our proposed method is effective for arbitrary input sketch-reference pairs and achieves high-quality colorization in various use scenarios." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Quantitative comparison. Fréchet Inception Distance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.702 + ], + "angle": 0, + "content": "(FID) [13] is a widely used quantitative cretera to evaluate the performance of image synthesis tasks. It calculates the perceptual distance of two distributions without requiring them to be semantically and spatially paired. We conduct a quantitative evaluation measured by FID on the entire validation set which includes 50000 (sketch, reference) image pairs. Sketches are colorized with randomly selected reference images, ensuring each batch has semantically and spatially mismatched inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.884 + ], + "angle": 0, + "content": "Multi-scale structural similarity index measure (MS-SSIM), peak signal-to-noise ratio (PSNR), and CLIP score [17,34] assess the similarity between processed images and given ground truth. It requires the guiding reference to be aligned with the ground truth when applied to image-referenced sketch colorization. To fulfill this, we selected 5000 color images as ground truth, extracted the sketches from and used thin plate spline (TPS) transformation to spatially distort them as color references to build a test set of 5000 triples of sketch, reference, and ground truth, and evaluate all the 7 methods with MS-SSIM, PSNR and CLIP score on this dataset." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We show the results of the quantitative comparison in Ta" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "23397" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.146 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison between the proposed model and baseline methods. We calculated 50K FID, 5K PSNR, 5K MS-SSIM, and 5K CLIP cosine similarity of image embeddings in this experiment. †: FID evaluation randomly selected color images as references, making it close to real-application scenarios. ‡: Ground truth color images were deformed to obtain semantically paired and spatially similar references for evaluations. We also show the inference time on H100 GPU at 512*512 resolution and batch size 4." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.15, + 0.865, + 0.279 + ], + "angle": 0, + "content": "
Method†50K-FID ↓‡PSNR↑‡MS-SSIM↑‡CLIP similarity↑Inference time(ms)
Ours6.832728.91440.60020.88291513
ColorizeDiffusion [47]9.527628.73840.59130.87751387
IP-Adapter [48,49,57]38.918428.67670.54780.86722617
InstantStyle [44,48,49,57]40.814428.10900.44590.80422586
T2I-Adapter [31,49,57]41.156928.12750.32430.71802301
AnimeDiffusion [2]61.599927.84540.31850.73192073
Yan et al. [46]27.014729.24830.52530.763457
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.291, + 0.466, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.537, + 0.471, + 0.595 + ], + "angle": 0, + "content": "Figure 9. Colorization results for semantically non-relevant sketch-reference pairs. Adapter-based methods fail to generate clear borders and high-quality textures, while our proposed method synthesizes visually satisfying results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.748 + ], + "angle": 0, + "content": "ble 1, where the proposed method significantly outperforms in FID, MM-SSIM, and CLIP similarity due to better texture and color synthesis. GAN-base method [46] achieved the best score in PSNR, with the proposed method ranked number 2. This is because the limited generation ability of [46] prevents it from generating complicated backgrounds and also hinders it from synthesizing bright colors and rich details of the figures. The close-to-average results make it advantageous to the calculation of PSNR [1]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "User study. The quality of sketch colorization is highly subjective and easily influenced by personal preference, we thus demonstrate how different individuals evaluate the proposed method and existing methods through an user study. 40 participants are invited to select the best results with two criteria: the overall colorization quality and the preservation of geometric structure of the sketches. We prepare 25 image sets and show each participant 16 image sets for evaluation. For each image set, participants are presented the colorization results of our proposed method as well as those" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.291, + 0.891, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.453, + 0.892, + 0.483 + ], + "angle": 0, + "content": "Figure 10. Results of user study. Our method is preferred across all shown methods in overall quality and geometric preservation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.496, + 0.731, + 0.511 + ], + "angle": 0, + "content": "generated by six existing methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.892, + 0.648 + ], + "angle": 0, + "content": "The results of the user study are presented in Figure 10, where our proposed method has received the most numbers of preferences among all the methods presented. To further validate the comparison, we applied the Kruskal-Wallis test as a statistical method. The results clearly demonstrate that the proposed method significantly outperforms all the existing methods in terms of user preference with a significance level of \\( p < 0.01 \\). All the images shown in the user study are included in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.66, + 0.62, + 0.675 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.806 + ], + "angle": 0, + "content": "In this paper, we proposed an image-guided sketch colorization framework inspired by real-world animation creation workflow. We implement a novel split cross-attention layer with spatial masks and corresponding LoRA weights that facilitates separate processing for foreground and background, eliminating the spatial entanglement, and also design a switchable LoRA mechanism that enables users to choose different inference modes for various use cases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.807, + 0.892, + 0.882 + ], + "angle": 0, + "content": "However, the proposed methods strongly depend on the quality of extracted masks, and a corresponding failure case is given in the supplementary materials. Our future work focuses on further improving the similarity with references and extending the framework for video colorization." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23398" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.158 + ], + "angle": 0, + "content": "[1] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[2] Yu Cao, Xiangqiao Meng, P. Y. Mok, Tong-Yee Lee, Xueting Liu, and Ping Li. Animediffusion: Anime diffusion colorization. TVCG, pages 1-14, 2024. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.471, + 0.27 + ], + "angle": 0, + "content": "[3] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[4] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.471, + 0.441 + ], + "angle": 0, + "content": "[5] Youngin Cho, Junsoo Lee, Soyoung Yang, Juntae Kim, Yeojeong Park, Haneol Lee, Mohammad Azam Khan, Daesik Kim, and Jaegul Choo. Guiding users to where to give color hints for efficient interactive sketch colorization via unsupervised region prioritization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1818-1827, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[6] Yunjey Choi, Min-Je Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In CVPR, pages 8789-8797. IEEE/CVF, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.469, + 0.556 + ], + "angle": 0, + "content": "[7] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In CVPR, pages 8185-8194. IEEE/CVF, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.469, + 0.612 + ], + "angle": 0, + "content": "[8] Danbooru community, Gwern Branwen, and Anonymous. Danbooru2021: A large-scale crowdsourced and tagged anime illustration dataset. https://gwern.net/ danbooru2021, 2022. Accessed: DATE 2022-01-21. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.615, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[9] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, pages 8780-8794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.713 + ], + "angle": 0, + "content": "[10] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, pages 2672–2680, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.769 + ], + "angle": 0, + "content": "[11] Reimu Hakurei. Hugging face/waifu-diffusion-v1-4. https://huggingface.co/hakurei/waifu-diffusion-v1-4, 2023. Accessed: DATE 2023-03-05. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[12] Havoc. Easynegative. https://civitai.com/models/7808/easynegative, 2023. Accessed: DATE 2023-02-10. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, pages 6626-6637, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. CoRR, abs/2207.12598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[16] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[17] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. https://github.com/mlfoundations/open Clip, jul 2021. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[18] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors, NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410. IEEE/CVF, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[20] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, pages 8107-8116. IEEE/CVF, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[21] Hyunsu Kim, Ho Young Jhoo, Eunhyeok Park, and Sungjoo Yoo. Tag2pix: Line art colorization using text tag with scat and changing loss. In ICCV, pages 9055-9064. IEEE/CVF, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[22] Kangyeol Kim, Sunghyun Park, Junsoo Lee, and Jaegul Choo. Reference-based image composition with sketch via structure-aware diffusion model. arXiv preprint arXiv:2304.09748, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[23] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[24] Kohya-ss. Hugging face/controlnet-lllite. https://huggingface.co/kohya-ss/controlnet-111ite, 2024. Accessed: DATE 2024-01-02. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[25] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. In ICLR. OpenReview.net, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[26] Junsoo Lee, Eungyeup Kim, Yunsung Lee, Dongjun Kim, Jaehyuk Chang, and Jaegul Choo. Reference-based sketch image colorization using augmented-self reference and dense semantic correspondence. In CVPR, pages 5800-5809. IEEE/CVF, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[27] Zekun Li, Zhengyang Geng, Zhao Kang, Wenyu Chen, and Yibo Yang. Eliminating gradient conflict in reference-based line-art colorization. In ECCV, pages 579-596. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*. OpenReview.net, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[29] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[30] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23399" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "guided sampling of diffusion probabilistic models. CoRR, abs/2211.01095, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[31] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. CoRR, abs/2302.08453, 2023. 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.181, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[32] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4172-4182. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.225, + 0.469, + 0.293 + ], + "angle": 0, + "content": "[33] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. CoRR, abs/2307.01952, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.297, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, volume 139, pages 8748-8763. PMLR, 2021. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.383, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685. IEEE/CVF, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, volume 9351, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.485, + 0.469, + 0.608 + ], + "angle": 0, + "content": "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.612, + 0.469, + 0.652 + ], + "angle": 0, + "content": "[38] SkyTNT, infoengine1337, and not lain. anime-segmentation. https : / / github . com / SkyTNT / anime - segmentation, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.469, + 0.711 + ], + "angle": 0, + "content": "[39] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, volume 37, pages 2256-2265. JMLR.org, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.469, + 0.753 + ], + "angle": 0, + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR. OpenReview.net, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR. OpenReview.net, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[42] Daniel Sykora, John Dingliana, and Steven Collins. Lazybrush: Flexible painting tool for hand-drawn cartoons. Comput. Graph. Forum, 28(2):599-608, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[43] TencentARC. Hugging face/ip-adapter. https://github.com/TencentARC/T2I-Adapter/tree/SD, 2024. Accessed: DATE 2024-01-02. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[44] Haofan Wang, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[45] Xiaoyu Xiang, Ding Liu, Xiao Yang, Yiheng Zhu, Xiaohui Shen, and Jan P. Allebach. Adversarial open domain adaptation for sketch-to-photo synthesis. In WACV, pages 944-954. IEEE/CVF, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[46] Dingkun Yan, Ryogo Ito, Ryo Moriai, and Suguru Saito. Two-step training: Adjustable sketch colourization via reference image and text tag. Computer Graphics Forum, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[47] Dingkun Yan, Liang Yuan, Yuma Nishioka, Issei Fujishiro, and Suguru Saito. Colorized diffusion: Adjustable sketch colorization with reference image and text. 2024. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[48] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. CoRR, abs/2308.06721, 2023. 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[49] Yuno779. https://civitai.com/models/9409, 2023. Accessed: DATE 2023-06-25. 6,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[50] Nir Zabari, Aharon Azulay, Alexey Gorkor, Tavi Halperin, and Ohad Fried. Diffusing colors: Image colorization with text guided diffusion. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[51] Bo Zhang, Yuxuan Duan, Jun Lan, Yan Hong, Huijia Zhu, Weiqiang Wang, and Li Niu. Controlcom: Controllable image composition using diffusion model. arXiv preprint arXiv:2308.10040, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[52] Lvmin Zhang. Sketchkeras. https://github.com/11lyasviel/sketchKeras, 2017.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[53] Lvmin Zhang. Style2paints v5, 2023. Accessed: DATE 2023-06-25. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[54] Lvmin Zhang. Controlnet-v1-1-nightly. https://github.com/lllyasviel/ControlNet-v1-1-nightly, 2024. Accessed: DATE 2024-01-02. 3,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.62, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[55] Lvmin Zhang, Chengze Li, Edgar Simo-Serra, Yi Ji, Tien-Tsin Wong, and Chunping Liu. User-guided line art flat filling with split filling mechanism. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9889-9898, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[56] Lvmin Zhang, Chengze Li, Tien-Tsin Wong, Yi Ji, and Chunping Liu. Two-stage sketch colorization. ACM Trans. Graph., 37(6):261, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[58] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.886 + ], + "angle": 0, + "content": "[59] Changqing Zou, Haoran Mo, Chengying Gao, Ruofei Du, and Hongbo Fu. Language-based colorization of scene sketches. ACM Trans. Graph., 38(6), 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.886 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23400" + } + ] +] \ No newline at end of file diff --git a/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/9a0c23ba-5904-4499-9efc-651298f2730b_origin.pdf b/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/9a0c23ba-5904-4499-9efc-651298f2730b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e0cba64bba9e2f1ba723cda856cfd44f6e5ee223 --- /dev/null +++ b/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/9a0c23ba-5904-4499-9efc-651298f2730b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28c75d558b67b4d2e9ba1eef047c751726c909dd3065d63033e3ea801fcdd580 +size 8439283 diff --git a/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/full.md b/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/full.md new file mode 100644 index 0000000000000000000000000000000000000000..878d8e92557bdcf3d93b6c389c65074dd04e9ed5 --- /dev/null +++ b/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/full.md @@ -0,0 +1,283 @@ +# Image Referenced Sketch Colorization Based on Animation Creation Workflow + +*Dingkun Yan $^{1,4}$ *Xinrui Wang $^{2}$ Zhuoru Li $^{3}$ Suguru Saito $^{1}$ Yusuke Iwasawa $^{2}$ Yutaka Matsuo $^{2}$ Jiaxian Guo $^{2}$ $^{1}$ Institute of Science Tokyo $^{2}$ The University of Tokyo $^{3}$ Project HAT $^{4}$ Tokyo AI research institute, Shanda Group + +![](images/e0091afec1a93cd7b8c9266bd97134a6d57390dfd66e2db1369fe5d10eced25a.jpg) +Figure 1. Given reference images, our proposed method automatically synthesizes high-quality sketch colorization results that loyally match the reference color distribution and are free from artifacts. + +# Abstract + +Sketch colorization plays an important role in animation and digital illustration production tasks. However, existing methods still meet problems in that text-guided methods fail to provide accurate color and style reference, hint-guided methods still involve manual operation, and image-referenced methods are prone to cause artifacts. To address these limitations, we propose a diffusion-based framework inspired by real-world animation production workflows. Our approach leverages the sketch as the spatial guidance and an RGB image as the color reference, and separately extracts foreground and background from the reference image with spatial masks. Particularly, we introduce a split cross-attention mechanism with LoRA (Low-Rank Adaptation) modules. They are trained separately with fore + +ground and background regions to control the corresponding embeddings for keys and values in cross-attention. This design allows the diffusion model to integrate information from foreground and background independently, preventing interference and eliminating the spatial artifacts. During inference, we design switchable inference modes for diverse use scenarios by changing modules activated in the framework. Extensive qualitative and quantitative experiments, along with user studies, demonstrate our advantages over existing methods in generating high-quality artifact-free results with geometric mismatched references. Ablation studies further confirm the effectiveness of each component. Codes are available at https://github.com/tellurion-kanata/colorizeDiffusion. + +![](images/d444f1b8b08f5514704c7d9821a8926bab78dad6329b409e4156664df957b53a.jpg) + +![](images/28f13251e764c1557df73577bbfdf08f0c3b2419754f778ec004b37207b41a0c.jpg) + +![](images/877a2029ad82697833538309bc10a7546290af58865b82eebe76e2c24d8c4eb1.jpg) + +![](images/ab7c7af59bc6d4f05e8b009910c0d8064cf133de4dfb5710ad51e90eb8598252.jpg) + +![](images/94d1f7bdde6af9ef001882e5f69c5df7987a1cbe411e8960d9c9c85cb7a7cb67.jpg) +A. Reference image + +![](images/627caa231ff18ff07384a95a8db74436ec8d933d0b623fc6213c2d1d492b77cb.jpg) +B. Sketch + +![](images/c9dee82eeb61eed3ecedd12d7ae1436811a6d4eaf29a8cf8b4fc685a2af46524.jpg) +C. Colorized foreground + +![](images/485bf55fed4ff739074f76847625103fbe69e1744b0b3ad15c2906d17d3b583e.jpg) +D. Colorized full frame + +# 1. Introduction + +The past decades have witnessed the development of Animation as an artistic form, which has gained great popularity worldwide. The current workflow of animation creation is labor-intensive, and the growing demands of animation from the market are causing animation studios to fall short of hands, bringing severe problems to the industry. + +The most manual labor-intensive procedure in animation production is sketch colorization, and animators working on colorization also take up the largest share among all employees in the animation production industry. To reduce the human labor needed and to automate the animation production, machine learning algorithms have been applied for sketch colorization [26, 27, 55, 56, 59]. However, current methods are still not optimal for real-world production pipelines: Text-based colorization methods [50, 57] fail to provide accurate guidance on the color and style information of the images. User-guided methods [5, 55] still involve manual operation in the process, making them less efficient. Image referenced methods [2, 47] can be seamlessly integrated into the current pipeline, but the spatial mismatches between reference images and sketches are causing severe artifacts and unexpected extra objects, which is termed as spatial entanglement in [47] and shown in Figure 2. + +To build a sketch colorization framework that meets the requirements of the real-world animation production pipeline, we start with the observation of the manual sketch colorization workflow in real-world animation production. As is shown in Figure 3, the workflow consists of the following key steps: Firstly, character designers design the characters as references. Secondly, senior animators draw the sketches for each frame. Thirdly, animators colorize the figures in the sketches according to the character designs. Finally, animators colorize the background of the sketches and finish the whole colorized frames. + +Following our observation, we designed a diffusion-based framework to mimic the sketch colorization workflow step-by-step. To be smoothly integrated into the current + +![](images/af20d50c548a392eaafe1160acfcbc14b4903a8e2696d3eced6ea9b54e8a3332.jpg) +Figure 3. Illustration of colorization workflow in professional animation studios. A: character designers design characters as references. B: Senior animators draw the sketches for the key frames. C: animators colorize the figures in the sketches according to the character designs, and D: animators colorize the background of the sketches and merge foreground and background into finished frames. +A. Sketch +Figure 2. Illustration of spatial entanglement. We use red rectangles to highlight the spatial entangled artifacts in the result of the IP-Adapter baseline, where additional arms appear unexpectedly, and the model mistakenly synthesizes long hair. + +![](images/1442e1b0058a141bb2994a98061157e1b7c039098a5cecdae4b0cb98cd73eb36.jpg) +B. Reference Image + +![](images/07fdff89b55293bee673384c1787d6c29158a242a19578d0c0523ba9250d8df3.jpg) +C. Result of IP Adapter Baseline + +![](images/6dee62e479956009839e5d901d5c05ff8a09bb18674e0f2be8d03815de3fff61.jpg) +D. Result of Proposed Method + +production workflow, the proposed framework leverages a sketch image as the spatial reference and an RGB image as the color reference. Specifically, we use the high dimensional local tokens extracted by a feature extractor as reference embeddings to maintain the semantic information and adopt a multi-layer sketch encoder for precise spatial control of the background embeddings. To enable the separate colorization of foreground and background regions, we design a novel split cross-attention mechanism, where spatial masks are used to segment figures as foreground with the rest of the images as background, and corresponding LoRA weights are trained to modify the embeddings for keys and values within cross-attention layers. This design allows the diffusion model to integrate information from foreground and background independently, preventing interference and eliminating the need to adjust the well-trained backbone weights. During inference, we implement a switchable LoRA mechanism that provides precise control over the colorization process and enables different inference modes for various scenarios without changing model weights. + +Experiments show that mimicking the real-world animation creation workflow yields several advantages. In qualitative analyses, our method synthesizes high-quality results loyally representing the color distribution of the reference images and free from artifacts and spatial entanglements. Quantitative comparisons also validate the superiority of the + +proposed method over existing methods by common criteria and benchmarks. What's more, user studies illustrate that candidates prefer our method subjectively. + +In summary, our contributions are as follows: (1) We propose an image-referenced sketch colorization framework that can synthesize high-quality results free from artifacts and spatial entanglement by mimicking the animation production workflow. (2) We design a novel split cross-attention mechanism that enables the separate colorization of foreground and background in a single forward pass and the switchable LoRA module that allows users to switch colorization modes during inference. (3) Experiments show that our method outperforms existing methods in qualitative/quantitative comparisons and is preferred by human users in perceptive user study. + +# 2. Related Work + +# 2.1. Latent Diffusion Models + +Diffusion Probabilistic Models [14, 41] are a class of latent variable models inspired by nonequilibrium thermodynamics [39] and have achieved great success in image synthesis and editing. Compared to Generative Adversarial Networks (GANs) [6, 7, 10, 19, 20], Diffusion Models excel at generating highly realistic images with various contexts and able to be controlled by text prompts. However, the autoregressive denoising process of diffusion models, typically computed with a U-Net [36] or a Diffusion Transformer (DiT) [3, 32], incurs substantial computational costs. + +To address this limitation, Stable Diffusion (SD) [33, 35] as a class of Latent Diffusion Models (LDMs) was proposed, where a two-stage synthesis mechanism was adopted to enable diffusion/denoising process to be performed on a highly compressed latent space with a pair of pretrained Variational Autoencoder (VAE), so as to significantly reduce computational costs. Concurrently, different researches of diffusion samplers have been conducted and proved to be effective in accelerating the denoising process [29, 30, 40, 41]. We adopt SD as our neural backbone, utilize the DPM++ solver [18, 30, 41] as the default sampler, and employ classifier-free guidance [9, 15] to strengthen the reference-based performance. + +# 2.2. Image Prompted Diffusion Models + +Currently, most diffusion models for image synthesis tasks are based on text prompts [3, 32, 33, 35]. However, there are tasks where text prompts can not provide enough information to precisely guide the image synthesis and editing, such as image-to-image translation [25], style transfer [44,58], colorization [2,47] and image composition [22,51], and thus images are also used as prompts to provide reference information. The reference information extracted from prompt images varies from tasks: style transfer tasks adopt + +the textures and colors from reference images, image composition tasks focus more on the object-related information, and sketch colorization requires all the above. + +There are two common practices to combine image prompts with diffusion models. Given image embedding vectors extracted by pre-trained feature extraction networks, existing methods either train an adapter module to inject the reference embedding vector into the backbone [31, 48] or directly inject reference information into the backbone with attention layers [16]. However, both of them [16, 31, 48] may introduce loss or mismatch of structure when inputs are not well paired, resulting in performance deterioration. In the sketch colorization task, specifically, these adapters provide conflicting spatial information from references and lead to unacceptable artifacts, as illustrated in Figure 2. + +# 2.3. Sketch Colorization + +Sketch colorization has been a long-standing topic in computer vision. Interactive optimization-based method [42] was employed for the task, and deep-learning-based methods [2, 21, 27, 47, 56] later became the mainstream due to the ability to synthesize high-quality and high-resolution images. There are three main technical solutions for deep learning methods: text-prompted sketch colorization [21, 46, 57], user-guided sketch colorization [53, 56] and image reference sketch colorization [2, 27, 46]. User-guided methods can precisely colorize given sketches with detailed guidance from users, but the manual labor needed makes them unsuitable to be integrated into an automatic workflow. Text-prompted methods have received great popularity over recent years due to the development of Text-to-Image diffusion models, but it is challenging to precisely control colors, textures, and styles using text prompts. Image-referenced methods also benefit from the development of diffusion models, along with relevant works that enable image control [24, 31, 48, 54, 57]. However, the mismatch of reference images and sketches still results in severe deterioration. ColorizeDiffusion [47] achieved notable progress in the colorization quality, yet it still suffers from spatial entanglement, which is shown in Figure 2. In this paper, we base our method on the production pipeline of animation studios to use image references to guide colorization. We separate the foreground and background with an innovative switchable LoRA to improve colorization performance and prevent artifacts. + +# 3. Method + +Inspired by our observation of real-world animation production, we propose an image-referenced sketch colorization framework. As is shown in Figure 4, it leverages a sketch image $\mathbf{X}_s \in \mathbb{R}^{w_s \times h_s \times 1}$ , a reference image $\mathbf{X}_r \in \mathbb{R}^{w_r \times h_r \times c}$ and a foreground mask $\mathbf{X}_m \in \mathbb{R}^{w_s \times h_s \times 1}$ as inputs, and returns the colorized result $\mathbf{Y} \in \mathbb{R}^{w_s \times h_s \times c}$ , with + +![](images/a38364fe23f6ae418982939e429bb4d8278bdba527d6e4920054cd914c5333d2.jpg) +Figure 4. Illustration of the proposed framework. We use reference masks to separate reference images into foreground and background and CLIP Image encoder $\phi$ to extract both regions into embeddings. The background embeddings first go through the recovery transformer $\varphi$ to recover detailed information, then concatenated with foreground embeddings as final K and V inputs for split cross-attention. Similar to Eq 2, the compose operation is a spatial piece-wise function employed to separate foreground and background. + +$w, h$ and $c$ representing the width, height and channel of the images. All components of the framework are based on the animation production workflow with the following design: + +Firstly, to mimic the animator's colorizing sketches with character designs as references, we utilize a pre-trained vision transformer (ViT) to extract image embeddings as reference information. The embeddings are later injected into the diffusion backbone with split cross-attention layers. Secondly, to integrate the sketches into the framework, we adopt a multi-layer sketch encoder to inject sketch information into the latent layers of the diffusion backbone as spatial guidance. Thirdly, based on the behavior of animators separately colorizing the foreground and background of the sketch, we propose a novel split cross-attention mechanism that uses spatial masks to separate the trainable LoRA modules corresponding to the keys and values for foreground and background for training. A switchable LoRA mechanism is then applied during inference for different application scenarios with different colorization modes. More details are described in the supplementary materials. + +# 3.1. Pretrain of the Diffusion Backbone + +The backbone of the proposed framework consists of a pre-trained VAE, a sketch encoder, a denoising U-Net, and a pre-trained Vision Transformer (ViT) functioning as the image encoder from OpenCLIP-H [4, 17, 34, 37]. We denote sketch images, reference images, and ground truth as $s$ , $r$ , and $y$ , respectively. The VAE encoder, U-Net, and ViT are represented by $\mathcal{E}$ , $\theta$ , and $\phi$ , respectively. The timestep $t$ starts from $T - 1$ and goes to 0, where $T$ is the diffusion steps, set to 1000. The training objective of the diffusion model is to denoise the intermediate noisy image $z_{t}$ via noise prediction: + +$$ +\mathcal {L} (\theta) = \mathbb {E} _ {\mathcal {E} (y), \epsilon , t, s, r} [ \| \epsilon - \epsilon_ {\theta} (z _ {t}, t, s, \phi (r)) \| _ {2} ^ {2} ]. \tag {1} +$$ + +To pre-train the diffusion backbone, We initialize the VAE and U-Net with WaifuDiffusion [11] and train networks with a dynamic reference drop rate decreasing from + +80% to 50% as training progresses to avoid the distribution shift mentioned by [47]. VAE, U-Net, and sketch encoder are then frozen during the training of the full framework. + +# 3.2. Color Reference Extraction + +Following the real-world animation production workflow, we utilize images as color references for sketch colorization. The commonly used ViT-based image encoder networks have two kinds of output embeddings: the CLS embeddings $\pmb{E}_{\mathbf{cls}} \in \mathbb{R}^{bs \times 1 \times 1024}$ and the local embeddings $\pmb{E}_{local} \in \mathbb{R}^{bs \times 256 \times 1024}$ , where $bs$ represent the batch size. The CLS embeddings are projected to CLIP embedding space for image-text contrastive learning, with spatial information compressed and connected to text-level notions, and are employed as color or style references by previous image-guided methods [44, 48]. ColorizeDiffusion [47], on the contrary, reveals that local embeddings also express text-level semantics, indicating that they express more details regarding textures, strokes, and styles, enabling the network to generate better reference-based results, especially for transferring detailed textures and strokes. Therefore, the proposed method follows [47] to adopt local tokens as color reference inputs for the framework. + +However, the excessive spatial information of image semantics and compositions contained in local embeddings leads to frequent occurrences of artifacts such as overflow of color regions and unexpected objects outside sketches. Illustrated in Figure 2, such artifacts widely exist in frameworks with image references [31,44,48]. To eliminate this problem, we follow the real-world workflow to explicitly separate the foreground and background with spatial masks during colorization and describe the detail in 3.3 and 3.4. + +# 3.3. Split Cross-Attention + +In anime images, the foreground regions and background regions differ distinctively in color distribution, color block sizes, tones, and textures. Thus, the colorization of foreground and background is separated into two independent + +![](images/23094763c74170299b6c33246fa37046a86d187d8219e4859e446cd295307407.jpg) +Figure 5. Based on the LoRA weights, the proposed method can merge the foreground and background features in one forward pass and switch between three inference modes. We denote the dimension of pre-trained weights as CH. The rank of foreground LoRA is fixed at 16, while the rank of background LoRA is $0.5^{*}\mathrm{CH}$ . + +steps in the animation production workflow. Following this scheme, we propose a novel split cross-attention mechanism to substitute the cross-attention layers in the diffusion backbone to separately process foreground and background regions with different parameters in a single forward pass. + +A split cross-attention layer consists of two groups of trainable LoRA weights $W_{f}^{t}$ and $W_{b}^{t}$ , which include query weights $W_{f}^{q}$ and $W_{b}^{q}$ , key weights $W_{f}^{k}$ and $W_{b}^{k}$ , and value weights $W_{f}^{v}$ and $W_{b}^{v}$ for foreground and background QKV projection respectively. An open-sourced animation image segmentation tool [38] is used to automatically extract the foreground mask $m_{s}$ and $m_{r}$ of sketches and reference images. The spatial masks can be adjusted by tuning the user defined thresholds $ts_{s}$ and $ts_{r}$ . Regions with pixel values larger than $ts_{s}$ and $ts_{r}$ are considered as foreground, otherwise background. For foreground LoRAs, we set the ranks as 16; for background LoRA, the rank is formulated as $r = 0.5 * \min(D_{q}, D_{kv})$ , where $D_{q}$ and $D_{kv}$ are dimensions of queries and keys/values for the corresponding cross-attention layers. + +As foreground and background regions of animation images feature different textures, color distribution, and color tones, injecting the embeddings of foreground and background directly into split cross-attention results in deterioration in structure preservation, synthesis quality, and stylization. Therefore, we further add a trainable recovery transformer $\varphi$ to process the background embeddings and facilitate better integration of the foreground and background reference information into the diffusion backbone. + +We define query inputs (forward features) as $z_{f}$ , $z_{b}$ , key and value inputs (reference embeddings) as $e$ , $e_{b}$ , attention outputs as $y$ in the following sections, where the index $f$ and $b$ indicate foreground and background respectively. Specifically, $e$ denotes the reference embeddings extracted from the whole reference image $r$ , formulated as $e = \phi(\boldsymbol{r})$ ; and $e_{b} = \varphi(\phi(\boldsymbol{r}_{b}))$ , where $r_{b}$ is the background region of the reference image. During training, the proposed split cross-attention can be formulated as follows: + +$$ +\boldsymbol {y} = \left\{ \begin{array}{l l} \operatorname {S o f t m a x} \left(\frac {\left(\hat {W} _ {f} ^ {q} z _ {f}\right) \cdot \left(\hat {W} _ {f} ^ {k} e\right)}{d}\right) \left(\hat {W} _ {f} ^ {v} e\right) & \text {i f} \boldsymbol {m} _ {s} > t s _ {s} \\ \operatorname {S o f t m a x} \left(\frac {\left(\hat {W} _ {b} ^ {q} z _ {b}\right) \cdot \left(\hat {W} _ {b} ^ {k} e _ {b}\right)}{d}\right) \left(\hat {W} _ {b} ^ {v} e _ {b}\right) & \text {i f} \boldsymbol {m} _ {s} \leq t s _ {s} \end{array} \right. \tag {2} +$$ + +where $\hat{W}_f^t = W^t + W_f^t$ , and $W^t$ represents the pretrained weights, which remain frozen during training. Similarly, $\hat{W}_b^t$ follows the same approach. + +# 3.4. Switchable inference mode + +The application scenarios and the sketch-reference combinations in the real world may be complicated. Also, naively separating the foreground and background regions for colorization degrades the quality of background synthesis, especially when reference images have severe semantic mismatches with the sketches or have complicated backgrounds. Therefore, we design three different inference modes: Vanilla, Bg2Fig, and Fig2Fig for different scenarios based on the weights and reference inputs used for KV calculation. We visualize all inference modes in Figure 5. + +Vanilla mode only utilizes the pre-trained weights for cross-attention modules, with $W_{f}^{t}$ , $W_{b}^{t}$ and recovery transformers deactivated. It's suitable for most scenarios but suffers from spatial entanglement when references are figure-only images. + +$Bg2Fig$ mode activates only the foreground-related LoRA weights $W_{f}^{t}$ during inference and is used when reference images are figure images with complicated backgrounds. This mode outperforms Vanilla mode in character colorization and Fig2Fig mode in background generation as its foreground weights are further optimized by LoRAs. + +Fig2Fig is designed for figure-to-figure colorization, where reference images are figures with simple background composition. This mode activates both LoRA weights $W_{f}^{t}$ , $W_{b}^{t}$ , and the recovery transformer and uses spatial masks to separate foreground/background embeddings for calculating query and key/value inputs. It effectively eliminates the spatial entanglement in reference-based sketch colorization. + +# 4. Experiment + +# 4.1. Implementation details + +The Danbooru2021 dataset [8] is used to train and validate the proposed method. We select and process 4.8 million triples of (sketch, color, mask) images with various contents to train the framework, and use another 50,000 image triples excluded from training set for validation. Sketches were extracted by jointly using [52] and [45]. 4x H100 (94GB) GPUs are used to train the network. We at first pre-train the denoising U-Net and sketch encoder for 6 epochs with 8 hours per epoch, and then froze the backbone and trained the recovery transformer and switchable LORAs for 3 epochs with 9 hours per epoch. AdamW op + +![](images/240c847aedf74beda6dadc0c9f89081707f76ec1945f60b1e8cc8dd9414be6bc.jpg) +Figure 6. Results of the ablation study. The baseline model demonstrates significant spatial entanglement; incorporating split cross-attention reduces artifacts, the trainable LoRAs improve color saturation and details, and the proposed complete pipeline produces high-quality results free of artifacts. Zoom in for details. + +timer [23, 28] is used to optimize the network with learning rate 0.0001 and betas (0.9, 0.999). Following [47], we dropped at least $50\%$ of the reference inputs in all training. + +# 4.2. Ablation study + +Split cross-attention. The proposed method aims to address spatial entanglement by simulating the animation workflow. To demonstrate the effectiveness of this workflow, we set up three frameworks: 1) baseline model without split cross attention, trainable LoRAs, and recovery transformer, 2) baseline model with split cross attention but no trainable LoRAs and recovery transformer, and 3) the proposed full framework. + +We show the qualitative comparison in Figure 6 to validate the effectiveness of the proposed modules. The baseline model causes severe spatial entanglement in generating additional figures in (a) and undesired clothes in (b). The application of split cross-attention mitigates the spatial entanglement but still causes artifacts and degrades the color saturation and details of the results. Collaborating split cross attention with trainable LoRAs improves the quality of results and further improves the background, but still suffers from artifacts. The proposed full framework enhanced by recovery transformers effectively eliminates spatial entanglement and synthesizes colorization results that have clear boundaries and rich details and textures, and loyally preserves the color distribution of reference images. + +Inference modes. We illustrate the differences between inference modes and their use cases in Figure 7. Fig2Fig mode is fully mask-guided, enabling it to eliminate spatial entanglement shown in (a). However, it is less suitable for inpainting character sketches, as the region without the mask guide may suffer from the lack of reference information. Bg2Fig performs similarly to the Vanilla mode. With the help of foreground LoRA weight, the results demonstrate clearer segmentation and better stroke quality. All modes perform well for landscape sketches, which contain more structural detail than figure sketches, making them + +![](images/05fa92683bb929f62a4a6d3328a9e061d0c9188da3341c6ebb81a85f634ccbf3.jpg) +Figure 7. Colorization results with three different inference modes. Fig2Fig mode performs better in eliminating spatial entanglement, while Bg2Fig and Vanilla mode can generate vivid backgrounds and inpainting results. + +easier to colorize. + +We show the comparison of local embeddings and CLS embeddings for color reference extraction in the supplementary materials. + +# 4.3. Comparison with baseline + +We compare our method with six existing reference-based sketch colorization methods [2,46,47] to demonstrate the superiority of the proposed framework. + +Baseline introduction. Two baseline methods are the combination of SD [35], ControlNet [54, 57], and IP-Adatper [48]. They adopt different cross-attention scales during denoising, labeled as IP-Adapter and InstantStyle, respectively. IP-adapter baseline generates results with normal cross-attention scales, while InstantStyle generates results using the "style transfer" weight type, which is claimed to prevent composition transfer by setting specific cross-attention scales to 0 according to [44]. Following the official document of [54], we adopted Anything v3 [49], a community-developed model, as the SD backbone for IP-Adapter-H and ControlNet_lineart_ANime. All these models are officially implemented and claimed to be effective for anime-style image generation. The T2I-Adapter baseline simply replaces IP-Adapter with T2I-Adapter-Style [31, 43]. Specifically, we introduced quality-related prompts and textual inversion, such as "masterpiece" and "easynegative" [12], for T2I-model baselines to improve their image quality. + +Qualitative comparison. We show the qualitative comparison of our proposed method and existing methods in Figure 8, where most of the existing methods suffer from spatial entanglement in various cases. In rows (a)-(d), existing methods (2)-(6) failed to distinguish foreground and background regions and, therefore, synthesized artifacts in + +![](images/5afc4bdb54454e2a45c18a45a23efa8173c233d20816e48910bd2b210ef0ffa0.jpg) +Figure 8. Qualitative comparisons between our proposed method and existing methods show that our results are visually more appealing than those of image-prompt adapters [31, 44, 48] and the GAN-based method [46]. Compared to ColorizeDiffusion [47], the proposed framework eliminates spatial entanglement and improves overall quality. + +the background. Such artifacts become more obvious when reference images have complicated backgrounds in rows (e) and (f), where IP-Adapter and InstantStyle mixed reference composition with the sketch composition, making the generated results visually messy. The GAN-based method (7) successfully generated results with clear backgrounds, but the color preservation, texture transformation, and detail qualities are much poorer than diffusion-based methods. + +We also show the comparison of our proposed method and adapter-based methods on semantically non-relevant sketch-reference pairs in Figure 9. Both IP-Adapter and T2I-Adapter fail to separate foreground from background and colorize the whole image with the same tone and texture. Our proposed method, on the contrary, generated results with clear region boundaries, rich texture and details, and visually pleasant bright color, with the help of the proposed spatial aware split cross-attention mechanism and switchable LoRA. The qualitative comparisons demonstrate that our proposed method is effective for arbitrary input sketch-reference pairs and achieves high-quality colorization in various use scenarios. + +Quantitative comparison. Fréchet Inception Distance + +(FID) [13] is a widely used quantitative cretera to evaluate the performance of image synthesis tasks. It calculates the perceptual distance of two distributions without requiring them to be semantically and spatially paired. We conduct a quantitative evaluation measured by FID on the entire validation set which includes 50000 (sketch, reference) image pairs. Sketches are colorized with randomly selected reference images, ensuring each batch has semantically and spatially mismatched inputs. + +Multi-scale structural similarity index measure (MS-SSIM), peak signal-to-noise ratio (PSNR), and CLIP score [17,34] assess the similarity between processed images and given ground truth. It requires the guiding reference to be aligned with the ground truth when applied to image-referenced sketch colorization. To fulfill this, we selected 5000 color images as ground truth, extracted the sketches from and used thin plate spline (TPS) transformation to spatially distort them as color references to build a test set of 5000 triples of sketch, reference, and ground truth, and evaluate all the 7 methods with MS-SSIM, PSNR and CLIP score on this dataset. + +We show the results of the quantitative comparison in Ta + +Table 1. Quantitative comparison between the proposed model and baseline methods. We calculated 50K FID, 5K PSNR, 5K MS-SSIM, and 5K CLIP cosine similarity of image embeddings in this experiment. †: FID evaluation randomly selected color images as references, making it close to real-application scenarios. ‡: Ground truth color images were deformed to obtain semantically paired and spatially similar references for evaluations. We also show the inference time on H100 GPU at 512*512 resolution and batch size 4. + +
Method†50K-FID ↓‡PSNR↑‡MS-SSIM↑‡CLIP similarity↑Inference time(ms)
Ours6.832728.91440.60020.88291513
ColorizeDiffusion [47]9.527628.73840.59130.87751387
IP-Adapter [48,49,57]38.918428.67670.54780.86722617
InstantStyle [44,48,49,57]40.814428.10900.44590.80422586
T2I-Adapter [31,49,57]41.156928.12750.32430.71802301
AnimeDiffusion [2]61.599927.84540.31850.73192073
Yan et al. [46]27.014729.24830.52530.763457
+ +![](images/f6f851ef4786ee33ab34f7268e6a77b90ec3e3a5be56ef317400825fca773cb8.jpg) +Figure 9. Colorization results for semantically non-relevant sketch-reference pairs. Adapter-based methods fail to generate clear borders and high-quality textures, while our proposed method synthesizes visually satisfying results. + +ble 1, where the proposed method significantly outperforms in FID, MM-SSIM, and CLIP similarity due to better texture and color synthesis. GAN-base method [46] achieved the best score in PSNR, with the proposed method ranked number 2. This is because the limited generation ability of [46] prevents it from generating complicated backgrounds and also hinders it from synthesizing bright colors and rich details of the figures. The close-to-average results make it advantageous to the calculation of PSNR [1]. + +User study. The quality of sketch colorization is highly subjective and easily influenced by personal preference, we thus demonstrate how different individuals evaluate the proposed method and existing methods through an user study. 40 participants are invited to select the best results with two criteria: the overall colorization quality and the preservation of geometric structure of the sketches. We prepare 25 image sets and show each participant 16 image sets for evaluation. For each image set, participants are presented the colorization results of our proposed method as well as those + +![](images/471c7f08bd5f46b71008c58edd4797e42142540641a1976958cda5847ab84f69.jpg) +Figure 10. Results of user study. Our method is preferred across all shown methods in overall quality and geometric preservation. + +generated by six existing methods. + +The results of the user study are presented in Figure 10, where our proposed method has received the most numbers of preferences among all the methods presented. To further validate the comparison, we applied the Kruskal-Wallis test as a statistical method. The results clearly demonstrate that the proposed method significantly outperforms all the existing methods in terms of user preference with a significance level of $p < 0.01$ . All the images shown in the user study are included in the supplementary materials. + +# 5. Conclusion + +In this paper, we proposed an image-guided sketch colorization framework inspired by real-world animation creation workflow. We implement a novel split cross-attention layer with spatial masks and corresponding LoRA weights that facilitates separate processing for foreground and background, eliminating the spatial entanglement, and also design a switchable LoRA mechanism that enables users to choose different inference modes for various use cases. + +However, the proposed methods strongly depend on the quality of extracted masks, and a corresponding failure case is given in the supplementary materials. Our future work focuses on further improving the similarity with references and extending the framework for video colorization. + +# References + +[1] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 8 +[2] Yu Cao, Xiangqiao Meng, P. Y. Mok, Tong-Yee Lee, Xueting Liu, and Ping Li. Animediffusion: Anime diffusion colorization. TVCG, pages 1-14, 2024. 2, 3, 6, 8 +[3] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 3 +[4] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 4 +[5] Youngin Cho, Junsoo Lee, Soyoung Yang, Juntae Kim, Yeojeong Park, Haneol Lee, Mohammad Azam Khan, Daesik Kim, and Jaegul Choo. Guiding users to where to give color hints for efficient interactive sketch colorization via unsupervised region prioritization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1818-1827, 2023. 2 +[6] Yunjey Choi, Min-Je Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In CVPR, pages 8789-8797. IEEE/CVF, 2018. 3 +[7] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In CVPR, pages 8185-8194. IEEE/CVF, 2020. 3 +[8] Danbooru community, Gwern Branwen, and Anonymous. Danbooru2021: A large-scale crowdsourced and tagged anime illustration dataset. https://gwern.net/ danbooru2021, 2022. Accessed: DATE 2022-01-21. 5 +[9] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, pages 8780-8794, 2021. 3 +[10] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, pages 2672–2680, 2014. 3 +[11] Reimu Hakurei. Hugging face/waifu-diffusion-v1-4. https://huggingface.co/hakurei/waifu-diffusion-v1-4, 2023. Accessed: DATE 2023-03-05. 4 +[12] Havoc. Easynegative. https://civitai.com/models/7808/easynegative, 2023. Accessed: DATE 2023-02-10. 6 +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, pages 6626-6637, 2017. 7 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3 + +[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. CoRR, abs/2207.12598, 2022. 3 +[16] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. 3 +[17] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. https://github.com/mlfoundations/open Clip, jul 2021. 4, 7 +[18] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors, NeurIPS, 2022. 3 +[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410. IEEE/CVF, 2019. 3 +[20] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, pages 8107-8116. IEEE/CVF, 2020. 3 +[21] Hyunsu Kim, Ho Young Jhoo, Eunhyeok Park, and Sungjoo Yoo. Tag2pix: Line art colorization using text tag with scat and changing loss. In ICCV, pages 9055-9064. IEEE/CVF, 2019. 3 +[22] Kangyeol Kim, Sunghyun Park, Junsoo Lee, and Jaegul Choo. Reference-based image composition with sketch via structure-aware diffusion model. arXiv preprint arXiv:2304.09748, 2023. 3 +[23] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6 +[24] Kohya-ss. Hugging face/controlnet-lllite. https://huggingface.co/kohya-ss/controlnet-111ite, 2024. Accessed: DATE 2024-01-02. 3 +[25] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. In ICLR. OpenReview.net, 2023. 3 +[26] Junsoo Lee, Eungyeup Kim, Yunsung Lee, Dongjun Kim, Jaehyuk Chang, and Jaegul Choo. Reference-based sketch image colorization using augmented-self reference and dense semantic correspondence. In CVPR, pages 5800-5809. IEEE/CVF, 2020. 2 +[27] Zekun Li, Zhengyang Geng, Zhao Kang, Wenyu Chen, and Yibo Yang. Eliminating gradient conflict in reference-based line-art colorization. In ECCV, pages 579-596. Springer, 2022. 2, 3 +[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*. OpenReview.net, 2019. 6 +[29] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022. 3 +[30] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for + +guided sampling of diffusion probabilistic models. CoRR, abs/2211.01095, 2022. 3 +[31] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. CoRR, abs/2302.08453, 2023. 3, 4, 6, 7, 8 +[32] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4172-4182. IEEE, 2023. 3 +[33] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. CoRR, abs/2307.01952, 2023. 3 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, volume 139, pages 8748-8763. PMLR, 2021. 4, 7 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685. IEEE/CVF, 2022. 3, 6 +[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, volume 9351, pages 234-241. Springer, 2015. 3 +[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 4 +[38] SkyTNT, infoengine1337, and not lain. anime-segmentation. https : / / github . com / SkyTNT / anime - segmentation, 2022.5 +[39] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, volume 37, pages 2256-2265. JMLR.org, 2015. 3 +[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR. OpenReview.net, 2021. 3 +[41] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR. OpenReview.net, 2021. 3 +[42] Daniel Sykora, John Dingliana, and Steven Collins. Lazybrush: Flexible painting tool for hand-drawn cartoons. Comput. Graph. Forum, 28(2):599-608, 2009. 3 +[43] TencentARC. Hugging face/ip-adapter. https://github.com/TencentARC/T2I-Adapter/tree/SD, 2024. Accessed: DATE 2024-01-02. 6 + +[44] Haofan Wang, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 3, 4, 6, 7, 8 +[45] Xiaoyu Xiang, Ding Liu, Xiao Yang, Yiheng Zhu, Xiaohui Shen, and Jan P. Allebach. Adversarial open domain adaptation for sketch-to-photo synthesis. In WACV, pages 944-954. IEEE/CVF, 2022. 5 +[46] Dingkun Yan, Ryogo Ito, Ryo Moriai, and Suguru Saito. Two-step training: Adjustable sketch colourization via reference image and text tag. Computer Graphics Forum, 2023. 3, 6, 7, 8 +[47] Dingkun Yan, Liang Yuan, Yuma Nishioka, Issei Fujishiro, and Suguru Saito. Colorized diffusion: Adjustable sketch colorization with reference image and text. 2024. 2, 3, 4, 6, 7, 8 +[48] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. CoRR, abs/2308.06721, 2023. 3, 4, 6, 7, 8 +[49] Yuno779. https://civitai.com/models/9409, 2023. Accessed: DATE 2023-06-25. 6,8 +[50] Nir Zabari, Aharon Azulay, Alexey Gorkor, Tavi Halperin, and Ohad Fried. Diffusing colors: Image colorization with text guided diffusion. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 2 +[51] Bo Zhang, Yuxuan Duan, Jun Lan, Yan Hong, Huijia Zhu, Weiqiang Wang, and Li Niu. Controlcom: Controllable image composition using diffusion model. arXiv preprint arXiv:2308.10040, 2023. 3 +[52] Lvmin Zhang. Sketchkeras. https://github.com/11lyasviel/sketchKeras, 2017.5 +[53] Lvmin Zhang. Style2paints v5, 2023. Accessed: DATE 2023-06-25. 3 +[54] Lvmin Zhang. Controlnet-v1-1-nightly. https://github.com/lllyasviel/ControlNet-v1-1-nightly, 2024. Accessed: DATE 2024-01-02. 3,6 +[55] Lvmin Zhang, Chengze Li, Edgar Simo-Serra, Yi Ji, Tien-Tsin Wong, and Chunping Liu. User-guided line art flat filling with split filling mechanism. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9889-9898, 2021. 2 +[56] Lvmin Zhang, Chengze Li, Tien-Tsin Wong, Yi Ji, and Chunping Liu. Two-stage sketch colorization. ACM Trans. Graph., 37(6):261, 2018. 2, 3 +[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 2, 3, 6, 8 +[58] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 3 +[59] Changqing Zou, Haoran Mo, Chengying Gao, Ruofei Du, and Hongbo Fu. Language-based colorization of scene sketches. ACM Trans. Graph., 38(6), 2019. 2 \ No newline at end of file diff --git a/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/images.zip b/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..24cc31d95b8c971f8d6ae3aa67e6b9935e261c30 --- /dev/null +++ b/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9336a947cc5238386d9f07d158b0ae3a37f6af70f07f982f324801fadb8aafb +size 1043922 diff --git a/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/layout.json b/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c73048dfea290a63d59333874cc5f2fe7f2ff419 --- /dev/null +++ b/2025/Image Referenced Sketch Colorization Based on Animation Creation Workflow/layout.json @@ -0,0 +1,8147 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 103, + 541, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 103, + 541, + 120 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 541, + 120 + ], + "type": "text", + "content": "Image Referenced Sketch Colorization Based on Animation Creation Workflow" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "spans": [ + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "text", + "content": "*Dingkun Yan" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "text", + "content": " *Xinrui Wang" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "text", + "content": " Zhuoru Li" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "text", + "content": " Suguru Saito" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "text", + "content": " Yusuke Iwasawa" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "text", + "content": " Yutaka Matsuo" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "text", + "content": " Jiaxian Guo" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "text", + "content": "Institute of Science Tokyo " + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "text", + "content": "The University of Tokyo " + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "text", + "content": "Project HAT " + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 69, + 141, + 523, + 200 + ], + "type": "text", + "content": "Tokyo AI research institute, Shanda Group" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 224, + 538, + 478 + ], + "blocks": [ + { + "bbox": [ + 50, + 224, + 538, + 478 + ], + "lines": [ + { + "bbox": [ + 50, + 224, + 538, + 478 + ], + "spans": [ + { + "bbox": [ + 50, + 224, + 538, + 478 + ], + "type": "image", + "image_path": "e0091afec1a93cd7b8c9266bd97134a6d57390dfd66e2db1369fe5d10eced25a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 482, + 546, + 503 + ], + "lines": [ + { + "bbox": [ + 46, + 482, + 546, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 546, + 503 + ], + "type": "text", + "content": "Figure 1. Given reference images, our proposed method automatically synthesizes high-quality sketch colorization results that loyally match the reference color distribution and are free from artifacts." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 510, + 192, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 510, + 192, + 521 + ], + "spans": [ + { + "bbox": [ + 143, + 510, + 192, + 521 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 525, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 288, + 693 + ], + "type": "text", + "content": "Sketch colorization plays an important role in animation and digital illustration production tasks. However, existing methods still meet problems in that text-guided methods fail to provide accurate color and style reference, hint-guided methods still involve manual operation, and image-referenced methods are prone to cause artifacts. To address these limitations, we propose a diffusion-based framework inspired by real-world animation production workflows. Our approach leverages the sketch as the spatial guidance and an RGB image as the color reference, and separately extracts foreground and background from the reference image with spatial masks. Particularly, we introduce a split cross-attention mechanism with LoRA (Low-Rank Adaptation) modules. They are trained separately with fore" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 510, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 677 + ], + "type": "text", + "content": "ground and background regions to control the corresponding embeddings for keys and values in cross-attention. This design allows the diffusion model to integrate information from foreground and background independently, preventing interference and eliminating the spatial artifacts. During inference, we design switchable inference modes for diverse use scenarios by changing modules activated in the framework. Extensive qualitative and quantitative experiments, along with user studies, demonstrate our advantages over existing methods in generating high-quality artifact-free results with geometric mismatched references. Ablation studies further confirm the effectiveness of each component. Codes are available at https://github.com/tellurion-kanata/colorizeDiffusion." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 197, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 197, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 197, + 712 + ], + "type": "text", + "content": "* Represent equal contribution to this work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23391" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 72, + 162, + 105 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 162, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 162, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 162, + 105 + ], + "type": "image", + "image_path": "d444f1b8b08f5514704c7d9821a8926bab78dad6329b409e4156664df957b53a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 162, + 72, + 270, + 105 + ], + "blocks": [ + { + "bbox": [ + 162, + 72, + 270, + 105 + ], + "lines": [ + { + "bbox": [ + 162, + 72, + 270, + 105 + ], + "spans": [ + { + "bbox": [ + 162, + 72, + 270, + 105 + ], + "type": "image", + "image_path": "28f13251e764c1557df73577bbfdf08f0c3b2419754f778ec004b37207b41a0c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 270, + 72, + 377, + 105 + ], + "blocks": [ + { + "bbox": [ + 270, + 72, + 377, + 105 + ], + "lines": [ + { + "bbox": [ + 270, + 72, + 377, + 105 + ], + "spans": [ + { + "bbox": [ + 270, + 72, + 377, + 105 + ], + "type": "image", + "image_path": "877a2029ad82697833538309bc10a7546290af58865b82eebe76e2c24d8c4eb1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 377, + 72, + 537, + 105 + ], + "blocks": [ + { + "bbox": [ + 377, + 72, + 537, + 105 + ], + "lines": [ + { + "bbox": [ + 377, + 72, + 537, + 105 + ], + "spans": [ + { + "bbox": [ + 377, + 72, + 537, + 105 + ], + "type": "image", + "image_path": "ab7c7af59bc6d4f05e8b009910c0d8064cf133de4dfb5710ad51e90eb8598252.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 55, + 106, + 157, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 106, + 157, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 106, + 157, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 157, + 198 + ], + "type": "image", + "image_path": "94d1f7bdde6af9ef001882e5f69c5df7987a1cbe411e8960d9c9c85cb7a7cb67.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 199, + 136, + 208 + ], + "lines": [ + { + "bbox": [ + 75, + 199, + 136, + 208 + ], + "spans": [ + { + "bbox": [ + 75, + 199, + 136, + 208 + ], + "type": "text", + "content": "A. Reference image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 164, + 106, + 265, + 198 + ], + "blocks": [ + { + "bbox": [ + 164, + 106, + 265, + 198 + ], + "lines": [ + { + "bbox": [ + 164, + 106, + 265, + 198 + ], + "spans": [ + { + "bbox": [ + 164, + 106, + 265, + 198 + ], + "type": "image", + "image_path": "627caa231ff18ff07384a95a8db74436ec8d933d0b623fc6213c2d1d492b77cb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 199, + 228, + 207 + ], + "lines": [ + { + "bbox": [ + 198, + 199, + 228, + 207 + ], + "spans": [ + { + "bbox": [ + 198, + 199, + 228, + 207 + ], + "type": "text", + "content": "B. Sketch" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 273, + 106, + 373, + 198 + ], + "blocks": [ + { + "bbox": [ + 273, + 106, + 373, + 198 + ], + "lines": [ + { + "bbox": [ + 273, + 106, + 373, + 198 + ], + "spans": [ + { + "bbox": [ + 273, + 106, + 373, + 198 + ], + "type": "image", + "image_path": "c9dee82eeb61eed3ecedd12d7ae1436811a6d4eaf29a8cf8b4fc685a2af46524.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 199, + 359, + 208 + ], + "lines": [ + { + "bbox": [ + 285, + 199, + 359, + 208 + ], + "spans": [ + { + "bbox": [ + 285, + 199, + 359, + 208 + ], + "type": "text", + "content": "C. Colorized foreground" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 381, + 106, + 533, + 198 + ], + "blocks": [ + { + "bbox": [ + 381, + 106, + 533, + 198 + ], + "lines": [ + { + "bbox": [ + 381, + 106, + 533, + 198 + ], + "spans": [ + { + "bbox": [ + 381, + 106, + 533, + 198 + ], + "type": "image", + "image_path": "485bf55fed4ff739074f76847625103fbe69e1744b0b3ad15c2906d17d3b583e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 200, + 490, + 207 + ], + "lines": [ + { + "bbox": [ + 419, + 200, + 490, + 207 + ], + "spans": [ + { + "bbox": [ + 419, + 200, + 490, + 207 + ], + "type": "text", + "content": "D. Colorized full frame" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 252, + 127, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 127, + 263 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 127, + 263 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 268, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 287, + 339 + ], + "type": "text", + "content": "The past decades have witnessed the development of Animation as an artistic form, which has gained great popularity worldwide. The current workflow of animation creation is labor-intensive, and the growing demands of animation from the market are causing animation studios to fall short of hands, bringing severe problems to the industry." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 341, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 544 + ], + "type": "text", + "content": "The most manual labor-intensive procedure in animation production is sketch colorization, and animators working on colorization also take up the largest share among all employees in the animation production industry. To reduce the human labor needed and to automate the animation production, machine learning algorithms have been applied for sketch colorization [26, 27, 55, 56, 59]. However, current methods are still not optimal for real-world production pipelines: Text-based colorization methods [50, 57] fail to provide accurate guidance on the color and style information of the images. User-guided methods [5, 55] still involve manual operation in the process, making them less efficient. Image referenced methods [2, 47] can be seamlessly integrated into the current pipeline, but the spatial mismatches between reference images and sketches are causing severe artifacts and unexpected extra objects, which is termed as spatial entanglement in [47] and shown in Figure 2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 545, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 675 + ], + "type": "text", + "content": "To build a sketch colorization framework that meets the requirements of the real-world animation production pipeline, we start with the observation of the manual sketch colorization workflow in real-world animation production. As is shown in Figure 3, the workflow consists of the following key steps: Firstly, character designers design the characters as references. Secondly, senior animators draw the sketches for each frame. Thirdly, animators colorize the figures in the sketches according to the character designs. Finally, animators colorize the background of the sketches and finish the whole colorized frames." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Following our observation, we designed a diffusion-based framework to mimic the sketch colorization workflow step-by-step. To be smoothly integrated into the current" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 309, + 251, + 355, + 319 + ], + "blocks": [ + { + "bbox": [ + 46, + 209, + 545, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 545, + 242 + ], + "type": "text", + "content": "Figure 3. Illustration of colorization workflow in professional animation studios. A: character designers design characters as references. B: Senior animators draw the sketches for the key frames. C: animators colorize the figures in the sketches according to the character designs, and D: animators colorize the background of the sketches and merge foreground and background into finished frames." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 251, + 355, + 319 + ], + "lines": [ + { + "bbox": [ + 309, + 251, + 355, + 319 + ], + "spans": [ + { + "bbox": [ + 309, + 251, + 355, + 319 + ], + "type": "image", + "image_path": "af20d50c548a392eaafe1160acfcbc14b4903a8e2696d3eced6ea9b54e8a3332.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 324, + 353, + 332 + ], + "lines": [ + { + "bbox": [ + 324, + 324, + 353, + 332 + ], + "spans": [ + { + "bbox": [ + 324, + 324, + 353, + 332 + ], + "type": "text", + "content": "A. Sketch" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 345, + 545, + 389 + ], + "lines": [ + { + "bbox": [ + 306, + 345, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 545, + 389 + ], + "type": "text", + "content": "Figure 2. Illustration of spatial entanglement. We use red rectangles to highlight the spatial entangled artifacts in the result of the IP-Adapter baseline, where additional arms appear unexpectedly, and the model mistakenly synthesizes long hair." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 360, + 251, + 406, + 319 + ], + "blocks": [ + { + "bbox": [ + 360, + 251, + 406, + 319 + ], + "lines": [ + { + "bbox": [ + 360, + 251, + 406, + 319 + ], + "spans": [ + { + "bbox": [ + 360, + 251, + 406, + 319 + ], + "type": "image", + "image_path": "1442e1b0058a141bb2994a98061157e1b7c039098a5cecdae4b0cb98cd73eb36.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 320, + 402, + 337 + ], + "lines": [ + { + "bbox": [ + 364, + 320, + 402, + 337 + ], + "spans": [ + { + "bbox": [ + 364, + 320, + 402, + 337 + ], + "type": "text", + "content": "B. Reference Image" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 408, + 251, + 476, + 319 + ], + "blocks": [ + { + "bbox": [ + 408, + 251, + 476, + 319 + ], + "lines": [ + { + "bbox": [ + 408, + 251, + 476, + 319 + ], + "spans": [ + { + "bbox": [ + 408, + 251, + 476, + 319 + ], + "type": "image", + "image_path": "07fdff89b55293bee673384c1787d6c29158a242a19578d0c0523ba9250d8df3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 319, + 471, + 335 + ], + "lines": [ + { + "bbox": [ + 413, + 319, + 471, + 335 + ], + "spans": [ + { + "bbox": [ + 413, + 319, + 471, + 335 + ], + "type": "text", + "content": "C. Result of IP Adapter Baseline" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 484, + 251, + 533, + 319 + ], + "blocks": [ + { + "bbox": [ + 484, + 251, + 533, + 319 + ], + "lines": [ + { + "bbox": [ + 484, + 251, + 533, + 319 + ], + "spans": [ + { + "bbox": [ + 484, + 251, + 533, + 319 + ], + "type": "image", + "image_path": "6dee62e479956009839e5d901d5c05ff8a09bb18674e0f2be8d03815de3fff61.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 320, + 537, + 337 + ], + "lines": [ + { + "bbox": [ + 485, + 320, + 537, + 337 + ], + "spans": [ + { + "bbox": [ + 485, + 320, + 537, + 337 + ], + "type": "text", + "content": "D. Result of Proposed Method" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 399, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 639 + ], + "type": "text", + "content": "production workflow, the proposed framework leverages a sketch image as the spatial reference and an RGB image as the color reference. Specifically, we use the high dimensional local tokens extracted by a feature extractor as reference embeddings to maintain the semantic information and adopt a multi-layer sketch encoder for precise spatial control of the background embeddings. To enable the separate colorization of foreground and background regions, we design a novel split cross-attention mechanism, where spatial masks are used to segment figures as foreground with the rest of the images as background, and corresponding LoRA weights are trained to modify the embeddings for keys and values within cross-attention layers. This design allows the diffusion model to integrate information from foreground and background independently, preventing interference and eliminating the need to adjust the well-trained backbone weights. During inference, we implement a switchable LoRA mechanism that provides precise control over the colorization process and enables different inference modes for various scenarios without changing model weights." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Experiments show that mimicking the real-world animation creation workflow yields several advantages. In qualitative analyses, our method synthesizes high-quality results loyally representing the color distribution of the reference images and free from artifacts and spatial entanglements. Quantitative comparisons also validate the superiority of the" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23392" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "proposed method over existing methods by common criteria and benchmarks. What's more, user studies illustrate that candidates prefer our method subjectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 288, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 288, + 252 + ], + "type": "text", + "content": "In summary, our contributions are as follows: (1) We propose an image-referenced sketch colorization framework that can synthesize high-quality results free from artifacts and spatial entanglement by mimicking the animation production workflow. (2) We design a novel split cross-attention mechanism that enables the separate colorization of foreground and background in a single forward pass and the switchable LoRA module that allows users to switch colorization modes during inference. (3) Experiments show that our method outperforms existing methods in qualitative/quantitative comparisons and is preferred by human users in perceptive user study." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 263, + 134, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 134, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 134, + 276 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 284, + 185, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 185, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 185, + 296 + ], + "type": "text", + "content": "2.1. Latent Diffusion Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "content": "Diffusion Probabilistic Models [14, 41] are a class of latent variable models inspired by nonequilibrium thermodynamics [39] and have achieved great success in image synthesis and editing. Compared to Generative Adversarial Networks (GANs) [6, 7, 10, 19, 20], Diffusion Models excel at generating highly realistic images with various contexts and able to be controlled by text prompts. However, the autoregressive denoising process of diffusion models, typically computed with a U-Net [36] or a Diffusion Transformer (DiT) [3, 32], incurs substantial computational costs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 423, + 288, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 288, + 579 + ], + "type": "text", + "content": "To address this limitation, Stable Diffusion (SD) [33, 35] as a class of Latent Diffusion Models (LDMs) was proposed, where a two-stage synthesis mechanism was adopted to enable diffusion/denoising process to be performed on a highly compressed latent space with a pair of pretrained Variational Autoencoder (VAE), so as to significantly reduce computational costs. Concurrently, different researches of diffusion samplers have been conducted and proved to be effective in accelerating the denoising process [29, 30, 40, 41]. We adopt SD as our neural backbone, utilize the DPM++ solver [18, 30, 41] as the default sampler, and employ classifier-free guidance [9, 15] to strengthen the reference-based performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 232, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 232, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 232, + 600 + ], + "type": "text", + "content": "2.2. Image Prompted Diffusion Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Currently, most diffusion models for image synthesis tasks are based on text prompts [3, 32, 33, 35]. However, there are tasks where text prompts can not provide enough information to precisely guide the image synthesis and editing, such as image-to-image translation [25], style transfer [44,58], colorization [2,47] and image composition [22,51], and thus images are also used as prompts to provide reference information. The reference information extracted from prompt images varies from tasks: style transfer tasks adopt" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "the textures and colors from reference images, image composition tasks focus more on the object-related information, and sketch colorization requires all the above." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 109, + 547, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 547, + 263 + ], + "type": "text", + "content": "There are two common practices to combine image prompts with diffusion models. Given image embedding vectors extracted by pre-trained feature extraction networks, existing methods either train an adapter module to inject the reference embedding vector into the backbone [31, 48] or directly inject reference information into the backbone with attention layers [16]. However, both of them [16, 31, 48] may introduce loss or mismatch of structure when inputs are not well paired, resulting in performance deterioration. In the sketch colorization task, specifically, these adapters provide conflicting spatial information from references and lead to unacceptable artifacts, as illustrated in Figure 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 269, + 422, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 269, + 422, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 269, + 422, + 281 + ], + "type": "text", + "content": "2.3. Sketch Colorization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 288, + 547, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 547, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 547, + 623 + ], + "type": "text", + "content": "Sketch colorization has been a long-standing topic in computer vision. Interactive optimization-based method [42] was employed for the task, and deep-learning-based methods [2, 21, 27, 47, 56] later became the mainstream due to the ability to synthesize high-quality and high-resolution images. There are three main technical solutions for deep learning methods: text-prompted sketch colorization [21, 46, 57], user-guided sketch colorization [53, 56] and image reference sketch colorization [2, 27, 46]. User-guided methods can precisely colorize given sketches with detailed guidance from users, but the manual labor needed makes them unsuitable to be integrated into an automatic workflow. Text-prompted methods have received great popularity over recent years due to the development of Text-to-Image diffusion models, but it is challenging to precisely control colors, textures, and styles using text prompts. Image-referenced methods also benefit from the development of diffusion models, along with relevant works that enable image control [24, 31, 48, 54, 57]. However, the mismatch of reference images and sketches still results in severe deterioration. ColorizeDiffusion [47] achieved notable progress in the colorization quality, yet it still suffers from spatial entanglement, which is shown in Figure 2. In this paper, we base our method on the production pipeline of animation studios to use image references to guide colorization. We separate the foreground and background with an innovative switchable LoRA to improve colorization performance and prevent artifacts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 627, + 362, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 627, + 362, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 627, + 362, + 639 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Inspired by our observation of real-world animation production, we propose an image-referenced sketch colorization framework. As is shown in Figure 4, it leverages a sketch image " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_s \\in \\mathbb{R}^{w_s \\times h_s \\times 1}" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": ", a reference image " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_r \\in \\mathbb{R}^{w_r \\times h_r \\times c}" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": " and a foreground mask " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_m \\in \\mathbb{R}^{w_s \\times h_s \\times 1}" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": " as inputs, and returns the colorized result " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathbb{R}^{w_s \\times h_s \\times c}" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": ", with" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23393" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 73, + 539, + 201 + ], + "blocks": [ + { + "bbox": [ + 53, + 73, + 539, + 201 + ], + "lines": [ + { + "bbox": [ + 53, + 73, + 539, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 539, + 201 + ], + "type": "image", + "image_path": "a38364fe23f6ae418982939e429bb4d8278bdba527d6e4920054cd914c5333d2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "type": "text", + "content": "Figure 4. Illustration of the proposed framework. We use reference masks to separate reference images into foreground and background and CLIP Image encoder " + }, + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "type": "text", + "content": " to extract both regions into embeddings. The background embeddings first go through the recovery transformer " + }, + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 202, + 547, + 247 + ], + "type": "text", + "content": " to recover detailed information, then concatenated with foreground embeddings as final K and V inputs for split cross-attention. Similar to Eq 2, the compose operation is a spatial piece-wise function employed to separate foreground and background." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 258, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 293 + ], + "type": "inline_equation", + "content": "w, h" + }, + { + "bbox": [ + 46, + 258, + 287, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 258, + 287, + 293 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 258, + 287, + 293 + ], + "type": "text", + "content": " representing the width, height and channel of the images. All components of the framework are based on the animation production workflow with the following design:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 294, + 289, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 289, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 289, + 498 + ], + "type": "text", + "content": "Firstly, to mimic the animator's colorizing sketches with character designs as references, we utilize a pre-trained vision transformer (ViT) to extract image embeddings as reference information. The embeddings are later injected into the diffusion backbone with split cross-attention layers. Secondly, to integrate the sketches into the framework, we adopt a multi-layer sketch encoder to inject sketch information into the latent layers of the diffusion backbone as spatial guidance. Thirdly, based on the behavior of animators separately colorizing the foreground and background of the sketch, we propose a novel split cross-attention mechanism that uses spatial masks to separate the trainable LoRA modules corresponding to the keys and values for foreground and background for training. A switchable LoRA mechanism is then applied during inference for different application scenarios with different colorization modes. More details are described in the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 504, + 234, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 234, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 234, + 516 + ], + "type": "text", + "content": "3.1. Pretrain of the Diffusion Backbone" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "content": "The backbone of the proposed framework consists of a pre-trained VAE, a sketch encoder, a denoising U-Net, and a pre-trained Vision Transformer (ViT) functioning as the image encoder from OpenCLIP-H [4, 17, 34, 37]. We denote sketch images, reference images, and ground truth as " + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "content": ", respectively. The VAE encoder, U-Net, and ViT are represented by " + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "content": ", respectively. The timestep " + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "content": " starts from " + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "content": " and goes to 0, where " + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "content": " is the diffusion steps, set to 1000. The training objective of the diffusion model is to denoise the intermediate noisy image " + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "content": " via noise prediction:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 658, + 287, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 658, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 74, + 658, + 287, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) = \\mathbb {E} _ {\\mathcal {E} (y), \\epsilon , t, s, r} [ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, s, \\phi (r)) \\| _ {2} ^ {2} ]. \\tag {1}", + "image_path": "6b385c08bea2f80bafa023c6adb9dc1868b06ddc76ccc04c0898dcf57649d771.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "text", + "content": "To pre-train the diffusion backbone, We initialize the VAE and U-Net with WaifuDiffusion [11] and train networks with a dynamic reference drop rate decreasing from" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 258, + 546, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 546, + 294 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 546, + 294 + ], + "type": "text", + "content": "80% to 50% as training progresses to avoid the distribution shift mentioned by [47]. VAE, U-Net, and sketch encoder are then frozen during the training of the full framework." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 300, + 457, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 457, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 457, + 312 + ], + "type": "text", + "content": "3.2. Color Reference Extraction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 318, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 318, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 318, + 546, + 533 + ], + "type": "text", + "content": "Following the real-world animation production workflow, we utilize images as color references for sketch colorization. The commonly used ViT-based image encoder networks have two kinds of output embeddings: the CLS embeddings " + }, + { + "bbox": [ + 305, + 318, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{\\mathbf{cls}} \\in \\mathbb{R}^{bs \\times 1 \\times 1024}" + }, + { + "bbox": [ + 305, + 318, + 546, + 533 + ], + "type": "text", + "content": " and the local embeddings " + }, + { + "bbox": [ + 305, + 318, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{local} \\in \\mathbb{R}^{bs \\times 256 \\times 1024}" + }, + { + "bbox": [ + 305, + 318, + 546, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 318, + 546, + 533 + ], + "type": "inline_equation", + "content": "bs" + }, + { + "bbox": [ + 305, + 318, + 546, + 533 + ], + "type": "text", + "content": " represent the batch size. The CLS embeddings are projected to CLIP embedding space for image-text contrastive learning, with spatial information compressed and connected to text-level notions, and are employed as color or style references by previous image-guided methods [44, 48]. ColorizeDiffusion [47], on the contrary, reveals that local embeddings also express text-level semantics, indicating that they express more details regarding textures, strokes, and styles, enabling the network to generate better reference-based results, especially for transferring detailed textures and strokes. Therefore, the proposed method follows [47] to adopt local tokens as color reference inputs for the framework." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "content": "However, the excessive spatial information of image semantics and compositions contained in local embeddings leads to frequent occurrences of artifacts such as overflow of color regions and unexpected objects outside sketches. Illustrated in Figure 2, such artifacts widely exist in frameworks with image references [31,44,48]. To eliminate this problem, we follow the real-world workflow to explicitly separate the foreground and background with spatial masks during colorization and describe the detail in 3.3 and 3.4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 647, + 428, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 428, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 428, + 659 + ], + "type": "text", + "content": "3.3. Split Cross-Attention" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "type": "text", + "content": "In anime images, the foreground regions and background regions differ distinctively in color distribution, color block sizes, tones, and textures. Thus, the colorization of foreground and background is separated into two independent" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23394" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 281, + 191 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 281, + 191 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 281, + 191 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 281, + 191 + ], + "type": "image", + "image_path": "23094763c74170299b6c33246fa37046a86d187d8219e4859e446cd295307407.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 193, + 287, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 287, + 248 + ], + "type": "text", + "content": "Figure 5. Based on the LoRA weights, the proposed method can merge the foreground and background features in one forward pass and switch between three inference modes. We denote the dimension of pre-trained weights as CH. The rank of foreground LoRA is fixed at 16, while the rank of background LoRA is " + }, + { + "bbox": [ + 46, + 193, + 287, + 248 + ], + "type": "inline_equation", + "content": "0.5^{*}\\mathrm{CH}" + }, + { + "bbox": [ + 46, + 193, + 287, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "type": "text", + "content": "steps in the animation production workflow. Following this scheme, we propose a novel split cross-attention mechanism to substitute the cross-attention layers in the diffusion backbone to separately process foreground and background regions with different parameters in a single forward pass." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": "A split cross-attention layer consists of two groups of trainable LoRA weights " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "W_{f}^{t}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "W_{b}^{t}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": ", which include query weights " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "W_{f}^{q}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "W_{b}^{q}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": ", key weights " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "W_{f}^{k}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "W_{b}^{k}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": ", and value weights " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "W_{f}^{v}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "W_{b}^{v}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": " for foreground and background QKV projection respectively. An open-sourced animation image segmentation tool [38] is used to automatically extract the foreground mask " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "m_{s}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "m_{r}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": " of sketches and reference images. The spatial masks can be adjusted by tuning the user defined thresholds " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "ts_{s}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "ts_{r}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": ". Regions with pixel values larger than " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "ts_{s}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "ts_{r}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": " are considered as foreground, otherwise background. For foreground LoRAs, we set the ranks as 16; for background LoRA, the rank is formulated as " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "r = 0.5 * \\min(D_{q}, D_{kv})" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "D_{q}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "inline_equation", + "content": "D_{kv}" + }, + { + "bbox": [ + 46, + 316, + 287, + 496 + ], + "type": "text", + "content": " are dimensions of queries and keys/values for the corresponding cross-attention layers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": "As foreground and background regions of animation images feature different textures, color distribution, and color tones, injecting the embeddings of foreground and background directly into split cross-attention results in deterioration in structure preservation, synthesis quality, and stylization. Therefore, we further add a trainable recovery transformer " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " to process the background embeddings and facilitate better integration of the foreground and background reference information into the diffusion backbone." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "We define query inputs (forward features) as " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_{f}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_{b}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ", key and value inputs (reference embeddings) as " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "e_{b}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ", attention outputs as " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " in the following sections, where the index " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " indicate foreground and background respectively. Specifically, " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " denotes the reference embeddings extracted from the whole reference image " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ", formulated as " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "e = \\phi(\\boldsymbol{r})" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "; and " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "e_{b} = \\varphi(\\phi(\\boldsymbol{r}_{b}))" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "r_{b}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " is the background region of the reference image. During training, the proposed split cross-attention can be formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 310, + 91, + 545, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 545, + 125 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} = \\left\\{ \\begin{array}{l l} \\operatorname {S o f t m a x} \\left(\\frac {\\left(\\hat {W} _ {f} ^ {q} z _ {f}\\right) \\cdot \\left(\\hat {W} _ {f} ^ {k} e\\right)}{d}\\right) \\left(\\hat {W} _ {f} ^ {v} e\\right) & \\text {i f} \\boldsymbol {m} _ {s} > t s _ {s} \\\\ \\operatorname {S o f t m a x} \\left(\\frac {\\left(\\hat {W} _ {b} ^ {q} z _ {b}\\right) \\cdot \\left(\\hat {W} _ {b} ^ {k} e _ {b}\\right)}{d}\\right) \\left(\\hat {W} _ {b} ^ {v} e _ {b}\\right) & \\text {i f} \\boldsymbol {m} _ {s} \\leq t s _ {s} \\end{array} \\right. \\tag {2}", + "image_path": "23acb6175573b5b2e81b2787d387ff9e219f76737762954778cd6dc15c40c3ad.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 133, + 545, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 545, + 171 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 133, + 545, + 171 + ], + "type": "inline_equation", + "content": "\\hat{W}_f^t = W^t + W_f^t" + }, + { + "bbox": [ + 305, + 133, + 545, + 171 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 133, + 545, + 171 + ], + "type": "inline_equation", + "content": "W^t" + }, + { + "bbox": [ + 305, + 133, + 545, + 171 + ], + "type": "text", + "content": " represents the pretrained weights, which remain frozen during training. Similarly, " + }, + { + "bbox": [ + 305, + 133, + 545, + 171 + ], + "type": "inline_equation", + "content": "\\hat{W}_b^t" + }, + { + "bbox": [ + 305, + 133, + 545, + 171 + ], + "type": "text", + "content": " follows the same approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 178, + 454, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 454, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 454, + 190 + ], + "type": "text", + "content": "3.4. Switchable inference mode" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 198, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 316 + ], + "type": "text", + "content": "The application scenarios and the sketch-reference combinations in the real world may be complicated. Also, naively separating the foreground and background regions for colorization degrades the quality of background synthesis, especially when reference images have severe semantic mismatches with the sketches or have complicated backgrounds. Therefore, we design three different inference modes: Vanilla, Bg2Fig, and Fig2Fig for different scenarios based on the weights and reference inputs used for KV calculation. We visualize all inference modes in Figure 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 317, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 376 + ], + "type": "text", + "content": "Vanilla mode only utilizes the pre-trained weights for cross-attention modules, with " + }, + { + "bbox": [ + 304, + 317, + 545, + 376 + ], + "type": "inline_equation", + "content": "W_{f}^{t}" + }, + { + "bbox": [ + 304, + 317, + 545, + 376 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 317, + 545, + 376 + ], + "type": "inline_equation", + "content": "W_{b}^{t}" + }, + { + "bbox": [ + 304, + 317, + 545, + 376 + ], + "type": "text", + "content": " and recovery transformers deactivated. It's suitable for most scenarios but suffers from spatial entanglement when references are figure-only images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 377, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 449 + ], + "type": "inline_equation", + "content": "Bg2Fig" + }, + { + "bbox": [ + 304, + 377, + 545, + 449 + ], + "type": "text", + "content": " mode activates only the foreground-related LoRA weights " + }, + { + "bbox": [ + 304, + 377, + 545, + 449 + ], + "type": "inline_equation", + "content": "W_{f}^{t}" + }, + { + "bbox": [ + 304, + 377, + 545, + 449 + ], + "type": "text", + "content": " during inference and is used when reference images are figure images with complicated backgrounds. This mode outperforms Vanilla mode in character colorization and Fig2Fig mode in background generation as its foreground weights are further optimized by LoRAs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 449, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 535 + ], + "type": "text", + "content": "Fig2Fig is designed for figure-to-figure colorization, where reference images are figures with simple background composition. This mode activates both LoRA weights " + }, + { + "bbox": [ + 304, + 449, + 545, + 535 + ], + "type": "inline_equation", + "content": "W_{f}^{t}" + }, + { + "bbox": [ + 304, + 449, + 545, + 535 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 449, + 545, + 535 + ], + "type": "inline_equation", + "content": "W_{b}^{t}" + }, + { + "bbox": [ + 304, + 449, + 545, + 535 + ], + "type": "text", + "content": ", and the recovery transformer and uses spatial masks to separate foreground/background embeddings for calculating query and key/value inputs. It effectively eliminates the spatial entanglement in reference-based sketch colorization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 543, + 382, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 543, + 382, + 556 + ], + "spans": [ + { + "bbox": [ + 306, + 543, + 382, + 556 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 563, + 437, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 437, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 437, + 575 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "The Danbooru2021 dataset [8] is used to train and validate the proposed method. We select and process 4.8 million triples of (sketch, color, mask) images with various contents to train the framework, and use another 50,000 image triples excluded from training set for validation. Sketches were extracted by jointly using [52] and [45]. 4x H100 (94GB) GPUs are used to train the network. We at first pre-train the denoising U-Net and sketch encoder for 6 epochs with 8 hours per epoch, and then froze the backbone and trained the recovery transformer and switchable LORAs for 3 epochs with 9 hours per epoch. AdamW op" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23395" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 288, + 181 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 288, + 181 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 288, + 181 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 288, + 181 + ], + "type": "image", + "image_path": "240c847aedf74beda6dadc0c9f89081707f76ec1945f60b1e8cc8dd9414be6bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 189, + 287, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 287, + 243 + ], + "type": "text", + "content": "Figure 6. Results of the ablation study. The baseline model demonstrates significant spatial entanglement; incorporating split cross-attention reduces artifacts, the trainable LoRAs improve color saturation and details, and the proposed complete pipeline produces high-quality results free of artifacts. Zoom in for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 255, + 287, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 292 + ], + "type": "text", + "content": "timer [23, 28] is used to optimize the network with learning rate 0.0001 and betas (0.9, 0.999). Following [47], we dropped at least " + }, + { + "bbox": [ + 46, + 255, + 287, + 292 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 255, + 287, + 292 + ], + "type": "text", + "content": " of the reference inputs in all training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 300, + 138, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 138, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 138, + 312 + ], + "type": "text", + "content": "4.2. Ablation study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 413 + ], + "type": "text", + "content": "Split cross-attention. The proposed method aims to address spatial entanglement by simulating the animation workflow. To demonstrate the effectiveness of this workflow, we set up three frameworks: 1) baseline model without split cross attention, trainable LoRAs, and recovery transformer, 2) baseline model with split cross attention but no trainable LoRAs and recovery transformer, and 3) the proposed full framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 414, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 582 + ], + "type": "text", + "content": "We show the qualitative comparison in Figure 6 to validate the effectiveness of the proposed modules. The baseline model causes severe spatial entanglement in generating additional figures in (a) and undesired clothes in (b). The application of split cross-attention mitigates the spatial entanglement but still causes artifacts and degrades the color saturation and details of the results. Collaborating split cross attention with trainable LoRAs improves the quality of results and further improves the background, but still suffers from artifacts. The proposed full framework enhanced by recovery transformers effectively eliminates spatial entanglement and synthesizes colorization results that have clear boundaries and rich details and textures, and loyally preserves the color distribution of reference images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "Inference modes. We illustrate the differences between inference modes and their use cases in Figure 7. Fig2Fig mode is fully mask-guided, enabling it to eliminate spatial entanglement shown in (a). However, it is less suitable for inpainting character sketches, as the region without the mask guide may suffer from the lack of reference information. Bg2Fig performs similarly to the Vanilla mode. With the help of foreground LoRA weight, the results demonstrate clearer segmentation and better stroke quality. All modes perform well for landscape sketches, which contain more structural detail than figure sketches, making them" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 544, + 223 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 544, + 223 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 544, + 223 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 544, + 223 + ], + "type": "image", + "image_path": "05fa92683bb929f62a4a6d3328a9e061d0c9188da3341c6ebb81a85f634ccbf3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 230, + 545, + 275 + ], + "lines": [ + { + "bbox": [ + 305, + 230, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 305, + 230, + 545, + 275 + ], + "type": "text", + "content": "Figure 7. Colorization results with three different inference modes. Fig2Fig mode performs better in eliminating spatial entanglement, while Bg2Fig and Vanilla mode can generate vivid backgrounds and inpainting results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 287, + 379, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 287, + 379, + 297 + ], + "spans": [ + { + "bbox": [ + 306, + 287, + 379, + 297 + ], + "type": "text", + "content": "easier to colorize." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 299, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 299, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 545, + 335 + ], + "type": "text", + "content": "We show the comparison of local embeddings and CLS embeddings for color reference extraction in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 344, + 450, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 344, + 450, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 450, + 357 + ], + "type": "text", + "content": "4.3. Comparison with baseline" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 363, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 399 + ], + "type": "text", + "content": "We compare our method with six existing reference-based sketch colorization methods [2,46,47] to demonstrate the superiority of the proposed framework." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 407, + 546, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 546, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 546, + 634 + ], + "type": "text", + "content": "Baseline introduction. Two baseline methods are the combination of SD [35], ControlNet [54, 57], and IP-Adatper [48]. They adopt different cross-attention scales during denoising, labeled as IP-Adapter and InstantStyle, respectively. IP-adapter baseline generates results with normal cross-attention scales, while InstantStyle generates results using the \"style transfer\" weight type, which is claimed to prevent composition transfer by setting specific cross-attention scales to 0 according to [44]. Following the official document of [54], we adopted Anything v3 [49], a community-developed model, as the SD backbone for IP-Adapter-H and ControlNet_lineart_ANime. All these models are officially implemented and claimed to be effective for anime-style image generation. The T2I-Adapter baseline simply replaces IP-Adapter with T2I-Adapter-Style [31, 43]. Specifically, we introduced quality-related prompts and textual inversion, such as \"masterpiece\" and \"easynegative\" [12], for T2I-model baselines to improve their image quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Qualitative comparison. We show the qualitative comparison of our proposed method and existing methods in Figure 8, where most of the existing methods suffer from spatial entanglement in various cases. In rows (a)-(d), existing methods (2)-(6) failed to distinguish foreground and background regions and, therefore, synthesized artifacts in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23396" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 541, + 403 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 541, + 403 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 541, + 403 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 541, + 403 + ], + "type": "image", + "image_path": "5afc4bdb54454e2a45c18a45a23efa8173c233d20816e48910bd2b210ef0ffa0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 404, + 547, + 439 + ], + "lines": [ + { + "bbox": [ + 46, + 404, + 547, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 547, + 439 + ], + "type": "text", + "content": "Figure 8. Qualitative comparisons between our proposed method and existing methods show that our results are visually more appealing than those of image-prompt adapters [31, 44, 48] and the GAN-based method [46]. Compared to ColorizeDiffusion [47], the proposed framework eliminates spatial entanglement and improves overall quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 449, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 544 + ], + "type": "text", + "content": "the background. Such artifacts become more obvious when reference images have complicated backgrounds in rows (e) and (f), where IP-Adapter and InstantStyle mixed reference composition with the sketch composition, making the generated results visually messy. The GAN-based method (7) successfully generated results with clear backgrounds, but the color preservation, texture transformation, and detail qualities are much poorer than diffusion-based methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 545, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 700 + ], + "type": "text", + "content": "We also show the comparison of our proposed method and adapter-based methods on semantically non-relevant sketch-reference pairs in Figure 9. Both IP-Adapter and T2I-Adapter fail to separate foreground from background and colorize the whole image with the same tone and texture. Our proposed method, on the contrary, generated results with clear region boundaries, rich texture and details, and visually pleasant bright color, with the help of the proposed spatial aware split cross-attention mechanism and switchable LoRA. The qualitative comparisons demonstrate that our proposed method is effective for arbitrary input sketch-reference pairs and achieves high-quality colorization in various use scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "content": "Quantitative comparison. Fréchet Inception Distance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 449, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 555 + ], + "type": "text", + "content": "(FID) [13] is a widely used quantitative cretera to evaluate the performance of image synthesis tasks. It calculates the perceptual distance of two distributions without requiring them to be semantically and spatially paired. We conduct a quantitative evaluation measured by FID on the entire validation set which includes 50000 (sketch, reference) image pairs. Sketches are colorized with randomly selected reference images, ensuring each batch has semantically and spatially mismatched inputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 557, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 700 + ], + "type": "text", + "content": "Multi-scale structural similarity index measure (MS-SSIM), peak signal-to-noise ratio (PSNR), and CLIP score [17,34] assess the similarity between processed images and given ground truth. It requires the guiding reference to be aligned with the ground truth when applied to image-referenced sketch colorization. To fulfill this, we selected 5000 color images as ground truth, extracted the sketches from and used thin plate spline (TPS) transformation to spatially distort them as color references to build a test set of 5000 triples of sketch, reference, and ground truth, and evaluate all the 7 methods with MS-SSIM, PSNR and CLIP score on this dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "We show the results of the quantitative comparison in Ta" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "23397" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 118, + 529, + 220 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 115 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 115 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 115 + ], + "type": "text", + "content": "Table 1. Quantitative comparison between the proposed model and baseline methods. We calculated 50K FID, 5K PSNR, 5K MS-SSIM, and 5K CLIP cosine similarity of image embeddings in this experiment. †: FID evaluation randomly selected color images as references, making it close to real-application scenarios. ‡: Ground truth color images were deformed to obtain semantically paired and spatially similar references for evaluations. We also show the inference time on H100 GPU at 512*512 resolution and batch size 4." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 118, + 529, + 220 + ], + "lines": [ + { + "bbox": [ + 61, + 118, + 529, + 220 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 529, + 220 + ], + "type": "table", + "html": "
Method†50K-FID ↓‡PSNR↑‡MS-SSIM↑‡CLIP similarity↑Inference time(ms)
Ours6.832728.91440.60020.88291513
ColorizeDiffusion [47]9.527628.73840.59130.87751387
IP-Adapter [48,49,57]38.918428.67670.54780.86722617
InstantStyle [44,48,49,57]40.814428.10900.44590.80422586
T2I-Adapter [31,49,57]41.156928.12750.32430.71802301
AnimeDiffusion [2]61.599927.84540.31850.73192073
Yan et al. [46]27.014729.24830.52530.763457
", + "image_path": "d01143c2f03bb7b181d505e05f6b2b249385f6600a78520c3248df73143bbb47.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 230, + 285, + 414 + ], + "blocks": [ + { + "bbox": [ + 50, + 230, + 285, + 414 + ], + "lines": [ + { + "bbox": [ + 50, + 230, + 285, + 414 + ], + "spans": [ + { + "bbox": [ + 50, + 230, + 285, + 414 + ], + "type": "image", + "image_path": "f6f851ef4786ee33ab34f7268e6a77b90ec3e3a5be56ef317400825fca773cb8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 425, + 288, + 471 + ], + "lines": [ + { + "bbox": [ + 46, + 425, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 288, + 471 + ], + "type": "text", + "content": "Figure 9. Colorization results for semantically non-relevant sketch-reference pairs. Adapter-based methods fail to generate clear borders and high-quality textures, while our proposed method synthesizes visually satisfying results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 484, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 592 + ], + "type": "text", + "content": "ble 1, where the proposed method significantly outperforms in FID, MM-SSIM, and CLIP similarity due to better texture and color synthesis. GAN-base method [46] achieved the best score in PSNR, with the proposed method ranked number 2. This is because the limited generation ability of [46] prevents it from generating complicated backgrounds and also hinders it from synthesizing bright colors and rich details of the figures. The close-to-average results make it advantageous to the calculation of PSNR [1]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "User study. The quality of sketch colorization is highly subjective and easily influenced by personal preference, we thus demonstrate how different individuals evaluate the proposed method and existing methods through an user study. 40 participants are invited to select the best results with two criteria: the overall colorization quality and the preservation of geometric structure of the sketches. We prepare 25 image sets and show each participant 16 image sets for evaluation. For each image set, participants are presented the colorization results of our proposed method as well as those" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 230, + 545, + 348 + ], + "blocks": [ + { + "bbox": [ + 309, + 230, + 545, + 348 + ], + "lines": [ + { + "bbox": [ + 309, + 230, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 309, + 230, + 545, + 348 + ], + "type": "image", + "image_path": "471c7f08bd5f46b71008c58edd4797e42142540641a1976958cda5847ab84f69.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 358, + 545, + 382 + ], + "lines": [ + { + "bbox": [ + 305, + 358, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 545, + 382 + ], + "type": "text", + "content": "Figure 10. Results of user study. Our method is preferred across all shown methods in overall quality and geometric preservation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 392, + 447, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 447, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 447, + 404 + ], + "type": "text", + "content": "generated by six existing methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 404, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 545, + 513 + ], + "type": "text", + "content": "The results of the user study are presented in Figure 10, where our proposed method has received the most numbers of preferences among all the methods presented. To further validate the comparison, we applied the Kruskal-Wallis test as a statistical method. The results clearly demonstrate that the proposed method significantly outperforms all the existing methods in terms of user preference with a significance level of " + }, + { + "bbox": [ + 304, + 404, + 545, + 513 + ], + "type": "inline_equation", + "content": "p < 0.01" + }, + { + "bbox": [ + 304, + 404, + 545, + 513 + ], + "type": "text", + "content": ". All the images shown in the user study are included in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 522, + 379, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 379, + 534 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 379, + 534 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 543, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 638 + ], + "type": "text", + "content": "In this paper, we proposed an image-guided sketch colorization framework inspired by real-world animation creation workflow. We implement a novel split cross-attention layer with spatial masks and corresponding LoRA weights that facilitates separate processing for foreground and background, eliminating the spatial entanglement, and also design a switchable LoRA mechanism that enables users to choose different inference modes for various use cases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 639, + 545, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 698 + ], + "type": "text", + "content": "However, the proposed methods strongly depend on the quality of extracted masks, and a corresponding failure case is given in the supplementary materials. Our future work focuses on further improving the similarity with references and extending the framework for video colorization." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23398" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "text", + "content": "[1] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 287, + 158 + ], + "type": "text", + "content": "[2] Yu Cao, Xiangqiao Meng, P. Y. Mok, Tong-Yee Lee, Xueting Liu, and Ping Li. Animediffusion: Anime diffusion colorization. TVCG, pages 1-14, 2024. 2, 3, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 288, + 213 + ], + "type": "text", + "content": "[3] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 270 + ], + "type": "text", + "content": "[4] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 349 + ], + "type": "text", + "content": "[5] Youngin Cho, Junsoo Lee, Soyoung Yang, Juntae Kim, Yeojeong Park, Haneol Lee, Mohammad Azam Khan, Daesik Kim, and Jaegul Choo. Guiding users to where to give color hints for efficient interactive sketch colorization via unsupervised region prioritization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1818-1827, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 404 + ], + "type": "text", + "content": "[6] Yunjey Choi, Min-Je Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In CVPR, pages 8789-8797. IEEE/CVF, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 407, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 287, + 440 + ], + "type": "text", + "content": "[7] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In CVPR, pages 8185-8194. IEEE/CVF, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 287, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 484 + ], + "type": "text", + "content": "[8] Danbooru community, Gwern Branwen, and Anonymous. Danbooru2021: A large-scale crowdsourced and tagged anime illustration dataset. https://gwern.net/ danbooru2021, 2022. Accessed: DATE 2022-01-21. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 518 + ], + "type": "text", + "content": "[9] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, pages 8780-8794, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 521, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 564 + ], + "type": "text", + "content": "[10] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, pages 2672–2680, 2014. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "text", + "content": "[11] Reimu Hakurei. Hugging face/waifu-diffusion-v1-4. https://huggingface.co/hakurei/waifu-diffusion-v1-4, 2023. Accessed: DATE 2023-03-05. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "type": "text", + "content": "[12] Havoc. Easynegative. https://civitai.com/models/7808/easynegative, 2023. Accessed: DATE 2023-02-10. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, pages 6626-6637, 2017. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 94 + ], + "type": "text", + "content": "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. CoRR, abs/2207.12598, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[16] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 206 + ], + "type": "text", + "content": "[17] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip. https://github.com/mlfoundations/open Clip, jul 2021. 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 208, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 261 + ], + "type": "text", + "content": "[18] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors, NeurIPS, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 264, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 296 + ], + "type": "text", + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410. IEEE/CVF, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 298, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 340 + ], + "type": "text", + "content": "[20] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, pages 8107-8116. IEEE/CVF, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "text", + "content": "[21] Hyunsu Kim, Ho Young Jhoo, Eunhyeok Park, and Sungjoo Yoo. Tag2pix: Line art colorization using text tag with scat and changing loss. In ICCV, pages 9055-9064. IEEE/CVF, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[22] Kangyeol Kim, Sunghyun Park, Junsoo Lee, and Jaegul Choo. Reference-based image composition with sketch via structure-aware diffusion model. arXiv preprint arXiv:2304.09748, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 433, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 453 + ], + "type": "text", + "content": "[23] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "type": "text", + "content": "[24] Kohya-ss. Hugging face/controlnet-lllite. https://huggingface.co/kohya-ss/controlnet-111ite, 2024. Accessed: DATE 2024-01-02. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "text", + "content": "[25] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. In ICLR. OpenReview.net, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 523, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 576 + ], + "type": "text", + "content": "[26] Junsoo Lee, Eungyeup Kim, Yunsung Lee, Dongjun Kim, Jaehyuk Chang, and Jaegul Choo. Reference-based sketch image colorization using augmented-self reference and dense semantic correspondence. In CVPR, pages 5800-5809. IEEE/CVF, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "type": "text", + "content": "[27] Zekun Li, Zhengyang Geng, Zhao Kang, Wenyu Chen, and Yibo Yang. Eliminating gradient conflict in reference-based line-art colorization. In ECCV, pages 579-596. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 624, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 645 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*. OpenReview.net, 2019. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[29] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[30] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23399" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "guided sampling of diffusion probabilistic models. CoRR, abs/2211.01095, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "text", + "content": "[31] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. CoRR, abs/2302.08453, 2023. 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 143, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 143, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 143, + 287, + 174 + ], + "type": "text", + "content": "[32] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4172-4182. IEEE, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 178, + 287, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 178, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 178, + 287, + 232 + ], + "type": "text", + "content": "[33] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. CoRR, abs/2307.01952, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 235, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 235, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 287, + 300 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, volume 139, pages 8748-8763. PMLR, 2021. 4, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 303, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 303, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 303, + 287, + 346 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685. IEEE/CVF, 2022. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "text", + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, volume 9351, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 384, + 287, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 384, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 384, + 287, + 481 + ], + "type": "text", + "content": "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 484, + 287, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 484, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 48, + 484, + 287, + 516 + ], + "type": "text", + "content": "[38] SkyTNT, infoengine1337, and not lain. anime-segmentation. https : / / github . com / SkyTNT / anime - segmentation, 2022.5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 519, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 287, + 563 + ], + "type": "text", + "content": "[39] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, volume 37, pages 2256-2265. JMLR.org, 2015. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 565, + 287, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 287, + 596 + ], + "type": "text", + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR. OpenReview.net, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "text", + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR. OpenReview.net, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[42] Daniel Sykora, John Dingliana, and Steven Collins. Lazybrush: Flexible painting tool for hand-drawn cartoons. Comput. Graph. Forum, 28(2):599-608, 2009. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[43] TencentARC. Hugging face/ip-adapter. https://github.com/TencentARC/T2I-Adapter/tree/SD, 2024. Accessed: DATE 2024-01-02. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 701 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[44] Haofan Wang, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[45] Xiaoyu Xiang, Ding Liu, Xiao Yang, Yiheng Zhu, Xiaohui Shen, and Jan P. Allebach. Adversarial open domain adaptation for sketch-to-photo synthesis. In WACV, pages 944-954. IEEE/CVF, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "text", + "content": "[46] Dingkun Yan, Ryogo Ito, Ryo Moriai, and Suguru Saito. Two-step training: Adjustable sketch colourization via reference image and text tag. Computer Graphics Forum, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 250 + ], + "type": "text", + "content": "[47] Dingkun Yan, Liang Yuan, Yuma Nishioka, Issei Fujishiro, and Suguru Saito. Colorized diffusion: Adjustable sketch colorization with reference image and text. 2024. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 253, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 295 + ], + "type": "text", + "content": "[48] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. CoRR, abs/2308.06721, 2023. 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 297, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 318 + ], + "type": "text", + "content": "[49] Yuno779. https://civitai.com/models/9409, 2023. Accessed: DATE 2023-06-25. 6,8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "text", + "content": "[50] Nir Zabari, Aharon Azulay, Alexey Gorkor, Tavi Halperin, and Ohad Fried. Diffusing colors: Image colorization with text guided diffusion. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "type": "text", + "content": "[51] Bo Zhang, Yuxuan Duan, Jun Lan, Yan Hong, Huijia Zhu, Weiqiang Wang, and Li Niu. Controlcom: Controllable image composition using diffusion model. arXiv preprint arXiv:2308.10040, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 411, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 430 + ], + "type": "text", + "content": "[52] Lvmin Zhang. Sketchkeras. https://github.com/11lyasviel/sketchKeras, 2017.5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 434, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 453 + ], + "type": "text", + "content": "[53] Lvmin Zhang. Style2paints v5, 2023. Accessed: DATE 2023-06-25. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "text", + "content": "[54] Lvmin Zhang. Controlnet-v1-1-nightly. https://github.com/lllyasviel/ControlNet-v1-1-nightly, 2024. Accessed: DATE 2024-01-02. 3,6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "type": "text", + "content": "[55] Lvmin Zhang, Chengze Li, Edgar Simo-Serra, Yi Ji, Tien-Tsin Wong, and Chunping Liu. User-guided line art flat filling with split filling mechanism. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9889-9898, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[56] Lvmin Zhang, Chengze Li, Tien-Tsin Wong, Yi Ji, and Chunping Liu. Two-stage sketch colorization. ACM Trans. Graph., 37(6):261, 2018. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 580, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 612 + ], + "type": "text", + "content": "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 2, 3, 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 614, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 667 + ], + "type": "text", + "content": "[58] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 670, + 545, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 701 + ], + "type": "text", + "content": "[59] Changqing Zou, Haoran Mo, Chengying Gao, Ruofei Du, and Hongbo Fu. Language-based colorization of scene sketches. ACM Trans. Graph., 38(6), 2019. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23400" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/e387f9f2-242c-46fd-9d3c-463d01e38cba_content_list.json b/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/e387f9f2-242c-46fd-9d3c-463d01e38cba_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dd1cacbb7d9ab82d6d2a05b20492b0d8a0343f88 --- /dev/null +++ b/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/e387f9f2-242c-46fd-9d3c-463d01e38cba_content_list.json @@ -0,0 +1,1399 @@ +[ + { + "type": "text", + "text": "Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation", + "text_level": 1, + "bbox": [ + 210, + 130, + 785, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pu Cao† Feng Zhou† Lu Yang* Tianrui Huang Qing Song \nBeijing University of Posts and Telecommunications", + "bbox": [ + 217, + 202, + 779, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{caopu, zhoufeng, soeaver, huangtianrui, priv}@bupt.edu.cn", + "bbox": [ + 241, + 241, + 754, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95bb0a65ef097a00e6d25ecbb6219a0ea0cd05ce3fa3094a6361d49897fa9e15.jpg", + "image_caption": [ + "Figure 1. Illustration of In-domain Generation. In this work, we empower large-scale pre-trained diffusion models using only image data to perform varied generation tasks within each domain with high fidelity and controllability. We mark the input data in blue and the results generated with the original Stable Diffusion v1.5 model in orange." + ], + "image_footnote": [], + "bbox": [ + 89, + 285, + 906, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 606, + 326, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In-domain generation aims to perform a variety of tasks within a specific domain, such as unconditional generation, text-to-image, image editing, 3D generation, and more. Early research typically required training specialized generators for each unique task and domain, often relying on fully-labeled data. Motivated by the powerful generative capabilities and broad applications of diffusion models, we are driven to explore leveraging label-free data to empower these models for in-domain generation. Fine-tuning a pre-trained generative model on domain data is an intuitive but challenging way and often requires complex manual hyper-parameter adjustments since the limited diversity of the training data can easily disrupt the model's origi", + "bbox": [ + 88, + 638, + 485, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nal generative capabilities. To address this challenge, we propose a guidance-decoupled prior preservation mechanism to achieve high generative quality and controllability by image-only data, inspired by preserving the pre-trained model from a denoising guidance perspective. We decouple domain-related guidance from the conditional guidance used in classifier-free guidance mechanisms to preserve open-world control guidance and unconditional guidance from the pre-trained model. We further propose an efficient domain knowledge learning technique to train an additional text-free UNet copy to predict domain guidance. Besides, we theoretically illustrate a multi-guidance in-domain generation pipeline for a variety of generative tasks, leveraging multiple guidances from distinct diffusion models and conditions. Extensive experiments demonstrate the superiority of our method in domain-specific synthesis and its compatibility with various diffusion-based control methods and applications.", + "bbox": [ + 511, + 608, + 908, + 880 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "The first two authors contributed to this paper equally.", + "bbox": [ + 112, + 849, + 413, + 864 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author: Lu Yang.", + "bbox": [ + 112, + 864, + 299, + 876 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Code is available at https://github.com/PRIV-Creation/In-domain-Generation-Diffusion.", + "bbox": [ + 91, + 876, + 480, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18358", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6448d8d9d43727a287d2a6293791d0aef64d49879a1b979339d3eb783f51299c.jpg", + "image_caption": [ + "a photo of face", + "a photo of face wearing hat", + "Figure 2. Challenge of fine-tuning diffusion models on domain data. We show the fine-tuning process of Stable Diffusion v1.5 with a facial image dataset (i.e., FFHQ[20]). The open-world controllability of pre-trained diffusion models is gradually decreased during the fine-tuning process, with domain fidelity improved." + ], + "image_footnote": [], + "bbox": [ + 165, + 88, + 496, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 297, + 222, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In-domain generation aims to accomplish various tasks within a specific domain, often requiring tailored approaches and labeled data. For instance, pix2pix [19] was developed for image-to-image translation (e.g., converting segmentation or edge images to realistic images), relying on paired images for training. Similarly, InterfaceGAN[45] enables attribute-based image editing but requires attribu-telabeled images to train editing vectors. Such traditional methods[6, 12, 19, 45] depend heavily on specific generative tasks and data domains, which limits their generalization capability.", + "bbox": [ + 89, + 323, + 483, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thanks to billion-scale image-text datasets [2, 37, 43, 44] and advances in large-scale natural language models [9, 35, 36], recent research on diffusion models has pushed the boundaries of AI-generated content, greatly enhancing both the diversity and controllability of generated outputs [39]. With various control mechanisms and approaches (e.g., ControlNet[61], SDEdit[29]), diffusion models can also perform diverse generative tasks beyond text-to-image synthesis, alleviating the data demands of traditional generative methods. However, the generated results often fail to fully align with specific data domains, as shown in orange in Fig. 1.", + "bbox": [ + 89, + 491, + 483, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While fine-tuning a pre-trained model on domain-specific data is an intuitive way to inject domain knowledge, it is challenging to balance fidelity with controllability, as illustrated in Fig. 2. For example, fine-tuning can improve the quality of face generation but may gradually reduce the effect of control prompts like wearing hat. In this work, we first identify this phenomenon caused by conditional guidance and unconditional guidance catastrophic forgetting during fine-tuning. These two guidances are integrated by classifier-free guidance for improved generative capability, which is employed in most text-to-image diffusion models. Then, we propose a guidance-decoupled prior preservation method to address these challenges. We decouple the conditional guidance in CFG into two parts: domain guidance, which is optimized in fine-tuning to enhance fi", + "bbox": [ + 89, + 674, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dely, and control guidance, which is preserved for controllability. With introducing multiple guidances in the synthesis process, the control guidance and unconditional guidance, which we want to preserve during personalization, are directly predicted by diffusion priors, while the domain guidance is predicted by an independent diffusion model. Additionally, we introduce an efficient domain knowledge learning mechanism, designing a null-text Diffusion Model to learn domain knowledge in a concise way.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments to demonstrate the effects of our proposed task and validate the superiority of our approach. To summarize, our contributions are as follows:", + "bbox": [ + 511, + 227, + 903, + 273 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present the task of aligning large-scale diffusion models with specific domains using only image data to perform a variety of generative tasks.", + "- We propose a guidance-decoupled prior preservation mechanism to address guidance forgetting during finetuning, which decouples domain guidance to learn domain knowledge with other guidance preserved. Besides, we propose an efficient learning mechanism to learn domain knowledge.", + "- We present the pipeline of our proposed method for in-domain generation tasks and conduct experiments and comparisons across multiple domains and tasks to demonstrate the effectiveness and advancement of our approach." + ], + "bbox": [ + 511, + 273, + 903, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 503, + 661, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image Generation. As the scale of image-text data grows [2, 37, 43, 44], large-scale diffusion models [10, 17, 31, 39] have demonstrated astonishing generative capabilities, able to produce images based on open-world prompts. Some models, such as Imagen [42] and Stable Diffusion [39], have been explored as diffusion priors, applied in downstream tasks like controlled generation [23, 30, 49, 61] and image editing [8, 15, 21]. For instance, ControlNet [61] proposes using an additional network to receive more spatial controls to guide generation. Furthermore, diffusion models have demonstrated powerful capabilities in many tasks, such as face parsing [54-59], style transfer [24, 51, 60], image manipulation [11, 53] and super-resolution [50, 52].", + "bbox": [ + 511, + 532, + 906, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fine-tuning Large-scale Diffusion Models. Some existing methods study learning subjects into large-scale diffusion priors and generate variant images guided by different prompts [1, 5, 13, 26, 27, 34, 41, 46, 47]. Some methods fine-tune the part of diffusion models with one or a few images of the specific subject [5, 13, 34, 41]. Textual inversion [13] tunes the text embedding of the subject placeholder, while DreamBooth [41] further updates the weight of UNet [40]. Some methods further study training-free [1, 46], continual customization [47], or multi-subjects", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18359", + "bbox": [ + 480, + 943, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/914999c871df7759bd2b8e8a8a94cb51c86011d743d022bc74c95b9eba7fbc9e.jpg", + "image_caption": [ + "Figure 3. Illustration of Fine-tuning Process. We demonstrate the guidance catastrophic forgetting during fine-tuning process." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 204 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2ab9bc7c2237b876a4da92e539edc06de6f732d4052dbbe96a36a60609fb5a14.jpg", + "image_caption": [ + "Figure 4. Unconditional Guidance Drift. The unconditional generation results of fine-tuned diffusion models reflect the visual pattern of training datasets, which would cause inaccurate noise estimation." + ], + "image_footnote": [], + "bbox": [ + 130, + 257, + 444, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "personalization [26, 27] and layout-guided generation [14]. However, these methods concentrate on the consistency of given subjects and training efficiency, tailored for the situation with small-scale image datasets and limited training process.", + "bbox": [ + 89, + 484, + 483, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fidelity and Controllability Tradeoff. In the field of generation, the tradeoff between fidelity and controllability has received widespread attention. In GANs, many inversion methods address this problem through regularization [3, 38], latent code discriminator [48], and invertibility prediction [4, 32]. In the personalization task, thanks to the simple customization process and excellent generalization of diffusion priors, this problem is easily solved by regularization [41], parameter efficient tuning [18], orthogonal fine-tuning [34], and so on. However, a longer training process is required for domain knowledge learning and the existing techniques struggle to solve this problem, highlighting an urgent need for a new approach.", + "bbox": [ + 89, + 564, + 483, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 773, + 181, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Guidance Catastrophic Forgetting", + "text_level": 1, + "bbox": [ + 89, + 797, + 390, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we delve into the phenomenon of catastrophic forgetting in the training process of diffusion models from the perspective of guidance, allowing us to address this issue at its root.", + "bbox": [ + 89, + 821, + 482, + 881 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-free Guidance (CFG). To achieve a better trade", + "bbox": [ + 89, + 885, + 482, + 898 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1b0952f14a6b7b9189d0a85b9a75a97994c1146ffb7082c37b4904792e70126b.jpg", + "image_caption": [ + "(a) previous fine-tuning", + "(b) guidance-decoupled generation", + "Figure 5. Conditional Guidance Decoupling. We compare the guidance estimation between previous customization methods and ours. We decouple the conditional guidance into domain guidance and control guidance while predicting the control guidance and unconditional guidance using the original diffusion model to keep them unchanged." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 924, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "off between quality and diversity, CFG [16] is utilized to combine the conditional and unconditional score estimates with a hyperparameter $w$ :", + "bbox": [ + 511, + 366, + 905, + 411 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\epsilon} _ {\\theta} \\left(z _ {t}, c\\right) = (1 + w) \\epsilon_ {\\theta} \\left(z _ {t}, c\\right) - w \\epsilon_ {\\theta} \\left(z _ {t}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 422, + 905, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In large-scale diffusion model training, suppose that an open-world training dataset denotes as $\\{z_i, c_i\\}$ , where $z_i$ is the training image and $c_i$ is the correspondent text, the training objective is to minimize:", + "bbox": [ + 511, + 452, + 905, + 511 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left(\\underbrace {\\mathcal {L} \\left(p _ {\\theta} (z \\mid c) , p (z \\mid c)\\right)} _ {\\text {c o n d i t i o n a l g u i d a n c e}} + \\underbrace {\\mathcal {L} \\left(p _ {\\theta} (z) , p (z)\\right)} _ {\\text {u n c o n d i t i o n a l g u i d a n c e}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 523, + 903, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Guidance Catastrophic Forgetting. During domain data fine-tuning with data $\\{z_i, y_i\\}$ , these two guidances inevitably deviate, which we illustrated in Fig. 3. Notably, $p(z|y)$ represents the probability conditioned on a given domain, which does not imply that text labels of domain data are required. For conditional guidance, the condition $y$ in fine-tuning phase is too much simpler than the condition in the pre-trained phase $c$ , where the latter is expected to be control generation. For example, although we leverage only facial images in training, we also desire to generate images controlled by text prompts. For unconditional guidance, we suppose that the pre-trained model could estimate the accurate $p_{\\theta_s d}(z_t|\\emptyset)$ since it is trained on open-world data. Hence, fine-tuning on domain data would make $p_{\\theta^*}(z_t|\\emptyset)$ inaccuracy, illustrated in Fig. 4. The generative results from fine-tuned models reflect the characteristics of the training datasets.", + "bbox": [ + 511, + 580, + 906, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Guidance-Decoupled Prior Preservation", + "text_level": 1, + "bbox": [ + 511, + 847, + 856, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the forgetting issue, we propose a guidance-decoupled prior preservation mechanism. We decouple con", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18360", + "bbox": [ + 480, + 943, + 517, + 954 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ditional guidance into domain guidance and control guidance, as illustrated in Fig 5(b). The domain guidance steers the denoising process to generate domain-aligned images and the control guidance takes charge of handling openworld controls in generation. For example, to generate 'a man wearing glasses' within the face domain, domain guidance tells how the human face looks, and control guidance ensures the synthetic faces follow 'man' and 'wearing glasses'. After such decoupling, control guidance and unconditional guidance can be predicted by diffusion priors and domain guidance can be learned by an additional model.", + "bbox": [ + 89, + 90, + 483, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditional Guidance Decoupling. Denoting a sets of conditions as $C = \\{c_1, \\dots, c_k\\}$ with manually defined intensities $w = \\{w_1, \\dots, w_k\\}$ , we can use multiple UN-ets $\\epsilon_{\\theta}$ to predict conditional and unconditional: $\\epsilon_{\\theta_0}(z_t) = \\nabla_z \\log p(z_t)$ and $\\epsilon_{\\theta_i}(z_t, c_i) = \\nabla_z \\log p(z_t | c_i)$ . The reverse process in each timestamp is as follows:", + "bbox": [ + 89, + 276, + 483, + 367 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} \\left(z _ {t}, C\\right) = \\epsilon_ {\\theta_ {0}} \\left(z _ {t}\\right) + \\sum_ {i = 1} ^ {K} w _ {i} \\left(\\epsilon_ {\\theta_ {i}} \\left(z _ {t}, c _ {i}\\right) - \\epsilon_ {\\theta_ {0}} \\left(z _ {t}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 378, + 483, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hence, we can decouple the original conditional guidance into two components to improve domain alignment and open-world control, respectively. As shown in Eq. (6), we can use different diffusion models to predict each guidance, allowing us to fine-tune a UNet copy to learn domain knowledge and utilize the pre-trained model to control the generation.", + "bbox": [ + 89, + 431, + 483, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, Eq. (6) assumes that all conditions are independent. This assumption leads to the independent maximization of the probability of each condition, which can cause confusion between the conditions. This issue is analyzed further in Appendix A. Our analysis reveals that a simple yet effective approach to mitigate this issue is to use more detailed descriptions (e.g., a photo of a face, wearing glasses) rather than overly simplified conditions (e.g., wearing glasses). This approach helps reduce confusion and better aligns the guidance with the intended domain.", + "bbox": [ + 89, + 540, + 483, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unconditional Guidance Rectification. Through multi-guidance generation, an intuitive solution to address unconditional guidance drift involves rectifying it using prior diffusion models. We can consider the original pre-trained model as an unconditional guidance predictor, effectively mitigating guidance drift after fine-tuning. This rectification operation can also used in other customized diffusion models, including open-sourced models, as shown in Sec 4.5.", + "bbox": [ + 89, + 709, + 483, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Efficient Domain Knowledge Learning", + "text_level": 1, + "bbox": [ + 89, + 842, + 423, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Null-text Diffusion Model. The additional fine-tuned diffusion model aims to learn the domain guidance by a do", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "main image dataset. To make it concise, we construct a null-text diffusion model to learn domain guidance.", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In LDMs, cross-attention is used to fuse text feature and image feature with residual connection, which follows:", + "bbox": [ + 511, + 121, + 906, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {i m g} ^ {\\prime} = \\underbrace {W (S o f t m a x (\\frac {\\mathcal {Q} _ {i m g} \\mathcal {K} _ {t e x t} ^ {T}}{\\sqrt {d}}) \\times \\mathcal {V} _ {t e x t})} _ {t e x t f e a t u r e i n j e c t i o n} + \\mathcal {F} _ {i m g} (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 159, + 906, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{Q}_{img}$ is transformed from input image feature $\\mathcal{F}_{img}$ , $\\mathcal{K}_{text}^{T},\\mathcal{V}_{text}$ is transformed from text feature, $d$ is the num of channels, and $W$ is a linear projection. Since textual information is not used in our fine-tuned model, we can convert the former term into a fixed embedding $E$ , and we have:", + "bbox": [ + 511, + 223, + 906, + 299 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {i m g} ^ {\\prime} = E + \\mathcal {F} _ {i m g} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 323, + 906, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initializing $E$ via Textual Features Optimization. To better inherent generative capability from diffusion priors, accurate textual features represent a good initializing point in training. Therefore, we draw inspiration from Textual Inversion[13] and optimize the text features $F^T$ , the output of text encoder, and use them as the initial values for $E = W_V(F^T)$ . Specifically, we first attain $F^T$ by feeding a photo of into the text encoder and optimize $F^T$ by denoising loss. This training process converges in just several minutes, providing a good initial value that accelerates the subsequent fine-tuning process.", + "bbox": [ + 511, + 354, + 908, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. In-domain Generation with Multi-Guidance", + "text_level": 1, + "bbox": [ + 511, + 529, + 887, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During domain knowledge learning, we fine-tune a domain diffusion model $\\epsilon_{\\theta_d}$ to provide domain guidance and employ the Stable Diffusion $\\epsilon_{sd}$ to provide preserved guidance, as illustrated in Fig 5. For control guidance, we can also flexibly adapt existing techniques, including Stable Diffusion, ControlNet, or any open-sourced customized models, to achieve a variety of applications. The reverse process suffices:", + "bbox": [ + 511, + 551, + 905, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {(1 - w _ {d} - w _ {c}) \\epsilon_ {s d} (z _ {t})} _ {u n c o n d i t i o n a l g u i d a n c e} + \\underbrace {w _ {d} \\epsilon_ {\\theta_ {d}} (z _ {t})} _ {d o m a i n g u i d a n c e} + \\underbrace {w _ {c} \\epsilon_ {\\theta_ {c}} (z _ {t} , c)} _ {c o n t r o l g u i d a n c e}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 683, + 903, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ is control signal with model $\\epsilon_{\\theta_c}$ . For instance, to generate images conditioned by a canny image in the animal domain, we utilize the diffusion model fine-tuned on an animal dataset as $\\epsilon_{\\theta_d}$ and the original diffusion model equipped with canny-to-image ControlNet model as $\\theta_c$ . Experimentally, $w_d$ presents stability in each generative task, similar to the guidance scale used in widely-used diffusion models (e.g., using $w_0 = 7$ for Stable Diffusion $\\epsilon_{sd}$ ). And we determine the initial value of $w_c$ and $w_d$ by $w_d = \\frac{w_0}{2} \\cdot \\frac{\\|\\epsilon_{sd}\\|}{\\|\\epsilon_{\\theta_d}\\|}$ and $w_c = \\frac{w_0}{2} \\cdot \\frac{\\|\\epsilon_{sd}\\|}{\\|\\epsilon_{\\theta_c}\\|}$ to ensure $w_0 \\cdot \\|\\epsilon_{sd}\\| = w_d \\|\\epsilon_{\\theta_d}\\| + w_c \\|\\epsilon_{\\theta_c}\\|$ . We can also employ", + "bbox": [ + 511, + 729, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18361", + "bbox": [ + 480, + 943, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this generation process in other diffusion-based generative tasks, like image editing and 3D generation.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 137, + 225, + 155 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 162, + 299, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. In this study, we conduct experiments on three domains: face (i.e., FFHQ [20]), animal (i.e., AFHQv2 [7]), and porcelain (i.e. a collected dataset), chosen for their diversity and high-resolution (larger than $512 \\times 512$ ). FFHQ is a high-quality human face dataset, containing 70,000 images. Images come from Flickr and are aligned into $1024 \\times 1024$ . We further resize images into $512 \\times 512$ without crop. AFHQv2 is an animal face dataset with a resolution of $512 \\times 512$ , containing 15803 images. The porcelain dataset contains 5,000 images, collected from the internet.", + "bbox": [ + 89, + 190, + 483, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Settings. We train the null-text domain diffusion model based on Stable Diffusion v1.5 (SD1.5) at the resolution of 512. For embedding learning phase, 3,200 images are used to attain initial embeddings with batch size of 32 and 100 steps for all domains. For all experiments, we maintain a constant learning rate of 1e-5 with the optimizer of AdamW and a batch size of 32. These experiments are conducted on 8 NVIDIA A100 GPUs, with each running until convergence. All controlling models, including ControlNet and open-sourced customized diffusion models, are borrowed from their official releases based on the corresponding version of Stable Diffusion without any training. We employ the DPM-Solver++ scheduler [28] with 100 steps across all models. For more experimental settings, please refer to Appendix C.", + "bbox": [ + 89, + 345, + 483, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results of In-domain Image Generation", + "text_level": 1, + "bbox": [ + 89, + 584, + 431, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present the unconditional generation results in Fig. 6 to illustrate the degree of domain alignment. As shown, our results closely match the distribution of the domain datasets, highlighting the effectiveness of our approach in aligning the model outputs with domain data. For conditional generation, we showcase the qualitative results in Fig. 7. We conduct experiments on both text-conditioned and spatial-conditioned generation, utilizing ControlNet models for the latter task. The results clearly demonstrate that our method not only achieves high fidelity with respect to the target domains but also retains the generative controllability seen in pre-trained models, ensuring that the model can be guided by text prompts or spatial conditions as intended. Quantitative results and comparisons can be found in Tab. 1 and Sec. 4.4.", + "bbox": [ + 89, + 608, + 485, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Results of Other In-domain Generation Tasks", + "text_level": 1, + "bbox": [ + 89, + 845, + 478, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Apart from the basic generative tasks, our proposed method also benefits other in-domain generation tasks. In this part,", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we further demonstrate the effectiveness of our method in a wider range of tasks.", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Generation under Complex Conditions. As discussed in Sec. 3.2, multiple guidances can be leveraged to control the generation process. In this section, we present the generation results under complex conditions in Fig. 8. Despite the introduction of complex conditions, the generated images maintain high fidelity to the target domain, demonstrating both the robustness and quality of our method. Additionally, our generative framework allows us to leverage additional open-source models for controlling the generation process, such as using DreamShaper to control the style.", + "bbox": [ + 511, + 125, + 906, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Editing. In Fig. 9, we implement SDEdit[29] on the face domain, where the input image is first encoded into latent space and added noise and then we use a similar generation pipeline presented in Sec. 3.4. Detailed settings can be found in Appendix C.3.1. Our results are more aligned with the distribution of real human face images, demonstrating better visual quality.", + "bbox": [ + 511, + 295, + 908, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text-to-3D Generation. In Fig. 10, we leverage our model on porcelain domain for text-to-3d porcelain generation and compare it with classic text-to-3d model [33] on SD1.5. Specifically, we begin by generating porcelain images based on the provided prompt, which are then converted into 3D models using an off-the-shelf image-to-3D method [25]. Subsequently, we fine-tune the resulting 3D models using our porcelain model, guided by the SDS [33] loss. Lastly, we directly convert the NeRF models into meshes. Detailed settings and additional ablation studies on 3D generation can be found in Appendix C.3.2. Our results demonstrate superiority in color richness, fine detail, and alignment with the dataset.", + "bbox": [ + 511, + 404, + 908, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Comparisons to Fine-tuning Techniques", + "text_level": 1, + "bbox": [ + 511, + 612, + 856, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To validate the superiority of our proposed method, we compare our method with several fine-tuning baselines, including vanilla fine-tuning, parameter-efficient fine-tuning, and some training techniques borrowed from other tasks. For detailed experimental settings, please refer to AppendixC.2.", + "bbox": [ + 511, + 636, + 905, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. To establish comprehensive baselines, we conduct experiments on several training techniques, including vanilla fine-tuning, low-rank fine-tuning [18], textual inversion [13], custom diffusion [22], and OFT mechanism [34]. To train diffusion models without text annotation, we adopt prompt templates, utilizing formats such as a photo of [V] . These additional tokens [V] are omitted in compared methods where text embedding is not trained.", + "bbox": [ + 511, + 729, + 908, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Setting. For unconditional generation, we calculate FID on 50,000 images. For text-guided generation,", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18362", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8a7cd6ef72f1975646f225aa315add948730f28205ad6c5a124a5aacc2bea4f5.jpg", + "image_caption": [ + "Figure 6. Visualization of Domain Alignment. Our results are more aligned to given domain dataset. For SD1.5 and SDXL 1.0, we generate images by text a photo of ." + ], + "image_footnote": [], + "bbox": [ + 106, + 87, + 906, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c9b89f2a8e62df86af02530665c825501ec606754b8dbf89bd237493e6980ffd.jpg", + "image_caption": [ + "Figure 7. Results of conditional in-domain image generation. We demonstrate text-conditioned and spatial-conditioned generation results to verify that our proposed methods can effectively preserve the control capability from the pre-trained model. For spatial-conditioned generation, we leverage the off-the-shelf ControlNet model without any extra training." + ], + "image_footnote": [], + "bbox": [ + 99, + 411, + 906, + 633 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a1945ed6c23c3f6afe7308363bbd95866d935b0cefd3e51d4828aa607049e3ce.jpg", + "image_caption": [ + "Figure 8. Generative Results under Complex Conditions. We employ multiple conditions and models to demonstrate the performance of our proposed method in achieving more detailed control over the generation process." + ], + "image_footnote": [], + "bbox": [ + 169, + 710, + 823, + 834 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18363", + "bbox": [ + 480, + 943, + 517, + 954 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1cc213ab33f467fadcaa0904f376fbe2366ea45e38427d3fc83a972c9ad89465.jpg", + "table_caption": [ + "Table 1. Quantitative results of unconditional generation We evaluate FID and human preference on two datasets. Our method achieves excellent generation results among baselines. $^{\\dagger}$ : reproduced results in $512 \\times 512$ resolution. Bold: the best result. Underline: the second best result." + ], + "table_footnote": [], + "table_body": "
Method#param.Unconditional GenerationConditional Generation
FaceAnimalPorcelainText-GuidedSpatial-Guided
FID↓Pref.↑FID↓Pref.↑FID↓Pref.↑Align.↑Pref.↑Align.↓Pref.↑
Textual Inversion [13]< 1K70.620.3%44.752.5%112.761.7%0.970.2%0.221.5%
Custom Diffusion [22]44M40.981.6%49.981.8%98.770.6%0.741.4%0.201.7%
OFT [34]23M27.502.9%38.741.4%68.841.2%0.773.2%0.219.6%
Fine-tuning859M12.376.4%23.9212.4%62.7610.8%0.244.5%0.224.9%
Fine-tuning + LoRA[18]0.19M23.761.3%24.191.7%73.460.3%0.766.7%0.229.1%
Ours810M6.5787.5%18.8280.2%56.4685.4%0.8984.0%0.2073.2%
", + "bbox": [ + 91, + 142, + 906, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dcf41b908565a6e3095f481eaef324917aaffcf133f28a7100a733e472ab9f76.jpg", + "image_caption": [ + "Figure 9. Image editing on face domain. We follow SDEdit[29] to edit facial images equipped with our proposed methods." + ], + "image_footnote": [], + "bbox": [ + 112, + 284, + 483, + 449 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3418ce1f1b5df6d480120756792f0023bc3e46da50a4583f280595016c4ec28f.jpg", + "image_caption": [ + "Figure 10. 3D generation on porcelain domain. We use Dream-Fusion [33] based on SD1.5 for comparison. We present two views of the generated objects, including their mesh edges. * denotes the prompt 'a DSLR photo of'." + ], + "image_footnote": [], + "bbox": [ + 112, + 503, + 483, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we generate 1,000 images over 10 attribute-related prompts on the facial domain and employ facial predictors to ascertain whether these attributes are accurately represented in", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the generated images. Besides, we sample 200 images from CelebA-HQ dataset to extract canny information as the spatial condition and compute the canny discrepancy. We also report human preference (denoted as 'Pref') by collecting the win rates of generative images across all comparisons. Detailed settings are presented in Appendix C.2.", + "bbox": [ + 511, + 287, + 906, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparisons. We conducted a qualitative comparison of our method against previous training techniques, as depicted in Fig 11. Methods like textual inversion, custom diffusion, and LoRA, which adhere to a parameter-efficient form, exhibit noticeably lower fidelity compared to others. Vanilla fine-tuning, in particular, demonstrates a decline in controllability during extended training periods. This results in a diminished capacity for text-guided generation and the emergence of unnatural artifacts, especially when used in conjunction with ControlNet. In contrast, our method excels in both fidelity and control capabilities.", + "bbox": [ + 511, + 382, + 908, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparisons. We conduct the evaluation of unconditional generation and report the results in Tab 1. Our approach demonstrates state-of-the-art generation quality, evidenced by a significantly lower FID compared to baseline methods and the highest human preference rating. For text-guided generation, Textual Inversion shows a higher alignment with the text prompts than our method. However, this is achieved at the cost of image quality, owing to its reliance on a very limited set of learnable parameters. This compromise in quality is clearly evident in the comparisons presented in Fig 11. In contrast, in spatial-guided generation, where semantic features are integrated into the UNet, the alignment degree among different methods is similar. Notably, our results not only align well with the control signals but also achieve the highest ratings in human preference assessments.", + "bbox": [ + 511, + 566, + 908, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Effects of Unconditional Guidance Rectification", + "text_level": 1, + "bbox": [ + 511, + 816, + 906, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In previous methods, unconditional guidance is predicted by the customized model, leading to inaccurate estimation of probability in sampling. In our framework, the uncon-", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18364", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e28ae807fe0b92db36848e5b0bebac9a1b20b8b1db450fd10bc8a3c6cb905446.jpg", + "image_caption": [ + "Figure 11. Comparisons to existing training methods. We evaluate in-domain image generation in unconditional, text-guided, and spatial-guided cases. We fix the random seed to illustrate the controllability for the former two." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 906, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f7e3db3481398eb08e726687a3ac90282bfb6439a5fd952bff76268094243b45.jpg", + "image_caption": [ + "(a) In-domain Generation", + "Figure 12. Effects of Unconditional Guidance Rectification. (a) Effects in Facial Domain Generation. (b) Unconditional guidance rectification also improves the open-sourced diffusion models." + ], + "image_footnote": [], + "bbox": [ + 107, + 378, + 352, + 580 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/63acaf7fc687c2acf9b3d65ec2c179cd77b40c7d19b1e9ec06dd772b110c4ece.jpg", + "image_caption": [ + "(b) Open-Sourced Diffusion Models" + ], + "image_footnote": [], + "bbox": [ + 375, + 373, + 890, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ditional guidance is rectified by straightforwardly predicting by diffusion priors. We compare generation results in Fig 12a. Our method significantly improves image quality, generating fewer artifacts and attaining photorealistic results. Besides, we utilize four widely-used open-sourced models. The comparative results are presented in Fig 12b. Without rectification, the generation quality shows slight degradation, as evidenced by less detailed eyes and less realistic hair textures. The implementation of guidance decoupling substantially enhances generative performance via the rectification of unconditional guidance.", + "bbox": [ + 89, + 656, + 482, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 842, + 209, + 858 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this research, we address the task of enhancing large-scale diffusion models for in-domain generation using only", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "image data. Our goal is to enable these models to generate images that faithfully represent the given domains, achieving high fidelity, diversity, and controllability. Moreover, our approach is designed to be compatible with various control methods, allowing the model to perform a wide range of generative tasks. To tackle this challenge, we propose a guidance-decoupled prior preservation mechanism that separates conditional guidance into two components: domain guidance for alignment and control guidance for maintaining the models controllability. Additionally, we introduce an efficient learning approach by incorporating a null-text diffusion model, ensuring both simplicity and effectiveness. We then outline the generative pipeline of our method to achieve diverse in-domain generation tasks. We conduct experiments across three domains to demonstrate the effectiveness of our proposed method.", + "bbox": [ + 511, + 656, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18365", + "bbox": [ + 480, + 943, + 517, + 954 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Young Scientists Fund of NSFC (Grant No. 62406035).", + "bbox": [ + 89, + 114, + 485, + 146 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 161, + 187, + 176 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H. Bermano. Domain-agnostic tuning-encoder for fast personalization of text-to-image models. pages 1–10, 2023. 2", + "[2] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 2", + "[3] Pu Cao, Lu Yang, Dongxu Liu, Zhiwei Liu, Shan Li, and Qing Song. Lsap: Rethinking inversion fidelity, perception and editability in gan latent space. arXiv preprint arXiv:2209.12746, 2022. 3", + "[4] Pu Cao, Lu Yang, Dongxv Liu, Xiaoya Yang, Tianrui Huang, and Qing Song. What decreases editing capability? domain-specific hybrid refinement for improved gan inversion. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 4240-4249, 2024. 3", + "[5] Hong Chen, Yipeng Zhang, Simin Wu, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. Disenbooth: Identity-preserving disentangled tuning for subject-driven text-to-image generation. arXiv preprint arXiv:2305.03374, 2023. 2", + "[6] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 100–116. Springer, 2019. 2", + "[7] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188-8197, 2020. 5", + "[8] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2", + "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171-4186, 2019. 2", + "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[11] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. In Proceedings of the IEEE/CVF Inter-" + ], + "bbox": [ + 93, + 186, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "national Conference on Computer Vision, pages 7430-7440, 2023. 2", + "[12] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model. Advances in Neural Information Processing Systems, 35:8882-8895, 2022. 2", + "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 4, 5, 7, 13", + "[14] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36:15890-15902, 2023. 3", + "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3, 5, 7", + "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2", + "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 5", + "[21] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6007-6017, 2023. 2", + "[22] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1931-1941, 2023. 5, 7, 13", + "[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22511-22521, 2023. 2", + "[24] Zuoxin Li, Fuqiang Zhou, Lu Yang, Xiaojie Li, and Juan Li. Accelerate neural style transfer with super-resolution. Multimedia Tools and Applications, 79:4347-4364, 2020. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18366", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 5", + "[26] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. arXiv preprint arXiv:2303.05125, 2023. 2, 3", + "[27] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Customizable image synthesis with multiple subjects. Advances in neural information processing systems, 36:57500-57519, 2023. 2, 3", + "[28] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 5, 13", + "[29] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 5, 7", + "[30] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 2", + "[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[32] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, JunYan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11399-11409, 2022. 3", + "[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 5, 7", + "[34] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 2, 3, 5, 7, 13", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2", + "[36] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2", + "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[38] Daniel Reich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 3", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2", + "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 3, 13", + "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2", + "[43] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2", + "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 2", + "[45] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 44(4):2004-2018, 2020. 2", + "[46] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. pages 8543-8552, 2024. 2", + "[47] James Seale Smith, Yen-Chang Hsu, Lingyu Zhang, Ting Hua, Zsolt Kira, Yilin Shen, and Hongxia Jin. Continual diffusion: Continual customization of text-to-image diffusion with c-lora. arXiv preprint arXiv:2304.06027, 2023. 2", + "[48] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 3", + "[49] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIG-" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18367", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "GRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2", + "[50] Yujin Wang, Lingen Li, Tianfan Xue, and Jinwei Gu. Reconstruct-and-generate diffusion model for detail-preserving image denoising. arXiv preprint arXiv:2309.10714, 2023. 2", + "[51] Zhizhong Wang, Lei Zhao, and Wei Xing. *Style diffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 2", + "[52] Chanyue Wu, Dong Wang, Yunpeng Bai, Hanyu Mao, Ying Li, and Qiang Shen. Hsr-diff: Hyperspectral image superresolution via conditional diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2023. 2", + "[53] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7378-7387, 2023. 2", + "[54] Lu Yang, Qing Song, Zhihui Wang, and Ming Jiang. Parsing r-cnn for instance-level human analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 364-373, 2019. 2", + "[55] Lu Yang, Qing Song, Zhihui Wang, Mengjie Hu, Chun Liu, Xueshi Xin, Wenhe Jia, and Songcen Xu. Renovating parsing r-cnn for accurate multiple human parsing. In European Conference on computer vision, pages 421-437. Springer, 2020.", + "[56] Lu Yang, Qing Song, Xueshi Xin, and Zhiwei Liu. Quality-aware network for face parsing. arXiv preprint arXiv:2106.07368, 2021.", + "[57] Lu Yang, Zhiwei Liu, Tianfei Zhou, and Qing Song. Part decomposition and refinement network for human parsing. IEEE/CAA Journal of Automatica Sinica, 9(6):1111-1114, 2022.", + "[58] Lu Yang, Qing Song, Zhihui Wang, Zhiwei Liu, Songcen Xu, and Zhihao Li. Quality-aware network for human parsing. IEEE Transactions on Multimedia, 2022.", + "[59] Lu Yang, Wenhe Jia, Shan Li, and Qing Song. Deep learning technique for human parsing: A survey and outlook. International Journal of Computer Vision, 132(8):3270-3301, 2024. 2", + "[60] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22873-22882, 2023. 2", + "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 782 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18368", + "bbox": [ + 480, + 943, + 517, + 954 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/e387f9f2-242c-46fd-9d3c-463d01e38cba_model.json b/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/e387f9f2-242c-46fd-9d3c-463d01e38cba_model.json new file mode 100644 index 0000000000000000000000000000000000000000..21f25e619bb67e6298cdc2eddbd81191bbb157ae --- /dev/null +++ b/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/e387f9f2-242c-46fd-9d3c-463d01e38cba_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.212, + 0.131, + 0.787, + 0.175 + ], + "angle": 0, + "content": "Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.203, + 0.78, + 0.24 + ], + "angle": 0, + "content": "Pu Cao† Feng Zhou† Lu Yang* Tianrui Huang Qing Song \nBeijing University of Posts and Telecommunications" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.242, + 0.756, + 0.258 + ], + "angle": 0, + "content": "{caopu, zhoufeng, soeaver, huangtianrui, priv}@bupt.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.286, + 0.907, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.552, + 0.908, + 0.596 + ], + "angle": 0, + "content": "Figure 1. Illustration of In-domain Generation. In this work, we empower large-scale pre-trained diffusion models using only image data to perform varied generation tasks within each domain with high fidelity and controllability. We mark the input data in blue and the results generated with the original Stable Diffusion v1.5 model in orange." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.607, + 0.327, + 0.623 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.64, + 0.486, + 0.837 + ], + "angle": 0, + "content": "In-domain generation aims to perform a variety of tasks within a specific domain, such as unconditional generation, text-to-image, image editing, 3D generation, and more. Early research typically required training specialized generators for each unique task and domain, often relying on fully-labeled data. Motivated by the powerful generative capabilities and broad applications of diffusion models, we are driven to explore leveraging label-free data to empower these models for in-domain generation. Fine-tuning a pre-trained generative model on domain data is an intuitive but challenging way and often requires complex manual hyper-parameter adjustments since the limited diversity of the training data can easily disrupt the model's origi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.909, + 0.881 + ], + "angle": 0, + "content": "nal generative capabilities. To address this challenge, we propose a guidance-decoupled prior preservation mechanism to achieve high generative quality and controllability by image-only data, inspired by preserving the pre-trained model from a denoising guidance perspective. We decouple domain-related guidance from the conditional guidance used in classifier-free guidance mechanisms to preserve open-world control guidance and unconditional guidance from the pre-trained model. We further propose an efficient domain knowledge learning technique to train an additional text-free UNet copy to predict domain guidance. Besides, we theoretically illustrate a multi-guidance in-domain generation pipeline for a variety of generative tasks, leveraging multiple guidances from distinct diffusion models and conditions. Extensive experiments demonstrate the superiority of our method in domain-specific synthesis and its compatibility with various diffusion-based control methods and applications." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.851, + 0.414, + 0.865 + ], + "angle": 0, + "content": "The first two authors contributed to this paper equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.865, + 0.3, + 0.877 + ], + "angle": 0, + "content": "* Corresponding author: Lu Yang." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.877, + 0.482, + 0.899 + ], + "angle": 0, + "content": "Code is available at https://github.com/PRIV-Creation/In-domain-Generation-Diffusion." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.851, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "18358" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.109, + 0.158, + 0.12 + ], + "angle": 0, + "content": "a photo of face" + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.154, + 0.157, + 0.173 + ], + "angle": 0, + "content": "a photo of face wearing hat" + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.089, + 0.498, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.214, + 0.485, + 0.286 + ], + "angle": 0, + "content": "Figure 2. Challenge of fine-tuning diffusion models on domain data. We show the fine-tuning process of Stable Diffusion v1.5 with a facial image dataset (i.e., FFHQ[20]). The open-world controllability of pre-trained diffusion models is gradually decreased during the fine-tuning process, with domain fidelity improved." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.298, + 0.223, + 0.314 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.484, + 0.49 + ], + "angle": 0, + "content": "In-domain generation aims to accomplish various tasks within a specific domain, often requiring tailored approaches and labeled data. For instance, pix2pix [19] was developed for image-to-image translation (e.g., converting segmentation or edge images to realistic images), relying on paired images for training. Similarly, InterfaceGAN[45] enables attribute-based image editing but requires attribu-telabeled images to train editing vectors. Such traditional methods[6, 12, 19, 45] depend heavily on specific generative tasks and data domains, which limits their generalization capability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.674 + ], + "angle": 0, + "content": "Thanks to billion-scale image-text datasets [2, 37, 43, 44] and advances in large-scale natural language models [9, 35, 36], recent research on diffusion models has pushed the boundaries of AI-generated content, greatly enhancing both the diversity and controllability of generated outputs [39]. With various control mechanisms and approaches (e.g., ControlNet[61], SDEdit[29]), diffusion models can also perform diverse generative tasks beyond text-to-image synthesis, alleviating the data demands of traditional generative methods. However, the generated results often fail to fully align with specific data domains, as shown in orange in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.902 + ], + "angle": 0, + "content": "While fine-tuning a pre-trained model on domain-specific data is an intuitive way to inject domain knowledge, it is challenging to balance fidelity with controllability, as illustrated in Fig. 2. For example, fine-tuning can improve the quality of face generation but may gradually reduce the effect of control prompts like wearing hat. In this work, we first identify this phenomenon caused by conditional guidance and unconditional guidance catastrophic forgetting during fine-tuning. These two guidances are integrated by classifier-free guidance for improved generative capability, which is employed in most text-to-image diffusion models. Then, we propose a guidance-decoupled prior preservation method to address these challenges. We decouple the conditional guidance in CFG into two parts: domain guidance, which is optimized in fine-tuning to enhance fi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "dely, and control guidance, which is preserved for controllability. With introducing multiple guidances in the synthesis process, the control guidance and unconditional guidance, which we want to preserve during personalization, are directly predicted by diffusion priors, while the domain guidance is predicted by an independent diffusion model. Additionally, we introduce an efficient domain knowledge learning mechanism, designing a null-text Diffusion Model to learn domain knowledge in a concise way." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.274 + ], + "angle": 0, + "content": "We conduct extensive experiments to demonstrate the effects of our proposed task and validate the superiority of our approach. To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.904, + 0.319 + ], + "angle": 0, + "content": "- We present the task of aligning large-scale diffusion models with specific domains using only image data to perform a variety of generative tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.904, + 0.41 + ], + "angle": 0, + "content": "- We propose a guidance-decoupled prior preservation mechanism to address guidance forgetting during finetuning, which decouples domain guidance to learn domain knowledge with other guidance preserved. Besides, we propose an efficient learning mechanism to learn domain knowledge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.904, + 0.486 + ], + "angle": 0, + "content": "- We present the pipeline of our proposed method for in-domain generation tasks and conduct experiments and comparisons across multiple domains and tasks to demonstrate the effectiveness and advancement of our approach." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.275, + 0.904, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.504, + 0.663, + 0.52 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.907, + 0.746 + ], + "angle": 0, + "content": "Text-to-image Generation. As the scale of image-text data grows [2, 37, 43, 44], large-scale diffusion models [10, 17, 31, 39] have demonstrated astonishing generative capabilities, able to produce images based on open-world prompts. Some models, such as Imagen [42] and Stable Diffusion [39], have been explored as diffusion priors, applied in downstream tasks like controlled generation [23, 30, 49, 61] and image editing [8, 15, 21]. For instance, ControlNet [61] proposes using an additional network to receive more spatial controls to guide generation. Furthermore, diffusion models have demonstrated powerful capabilities in many tasks, such as face parsing [54-59], style transfer [24, 51, 60], image manipulation [11, 53] and super-resolution [50, 52]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Fine-tuning Large-scale Diffusion Models. Some existing methods study learning subjects into large-scale diffusion priors and generate variant images guided by different prompts [1, 5, 13, 26, 27, 34, 41, 46, 47]. Some methods fine-tune the part of diffusion models with one or a few images of the specific subject [5, 13, 34, 41]. Textual inversion [13] tunes the text embedding of the subject placeholder, while DreamBooth [41] further updates the weight of UNet [40]. Some methods further study training-free [1, 46], continual customization [47], or multi-subjects" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18359" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.216, + 0.483, + 0.246 + ], + "angle": 0, + "content": "Figure 3. Illustration of Fine-tuning Process. We demonstrate the guidance catastrophic forgetting during fine-tuning process." + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.258, + 0.445, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.419, + 0.484, + 0.473 + ], + "angle": 0, + "content": "Figure 4. Unconditional Guidance Drift. The unconditional generation results of fine-tuned diffusion models reflect the visual pattern of training datasets, which would cause inaccurate noise estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.484, + 0.56 + ], + "angle": 0, + "content": "personalization [26, 27] and layout-guided generation [14]. However, these methods concentrate on the consistency of given subjects and training efficiency, tailored for the situation with small-scale image datasets and limited training process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.762 + ], + "angle": 0, + "content": "Fidelity and Controllability Tradeoff. In the field of generation, the tradeoff between fidelity and controllability has received widespread attention. In GANs, many inversion methods address this problem through regularization [3, 38], latent code discriminator [48], and invertibility prediction [4, 32]. In the personalization task, thanks to the simple customization process and excellent generalization of diffusion priors, this problem is easily solved by regularization [41], parameter efficient tuning [18], orthogonal fine-tuning [34], and so on. However, a longer training process is required for domain knowledge learning and the existing techniques struggle to solve this problem, highlighting an urgent need for a new approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.774, + 0.182, + 0.789 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.798, + 0.391, + 0.815 + ], + "angle": 0, + "content": "3.1. Guidance Catastrophic Forgetting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.822, + 0.483, + 0.882 + ], + "angle": 0, + "content": "In this section, we delve into the phenomenon of catastrophic forgetting in the training process of diffusion models from the perspective of guidance, allowing us to address this issue at its root." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Classifier-free Guidance (CFG). To achieve a better trade" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.925, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.23, + 0.665, + 0.242 + ], + "angle": 0, + "content": "(a) previous fine-tuning" + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.231, + 0.897, + 0.242 + ], + "angle": 0, + "content": "(b) guidance-decoupled generation" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.257, + 0.906, + 0.34 + ], + "angle": 0, + "content": "Figure 5. Conditional Guidance Decoupling. We compare the guidance estimation between previous customization methods and ours. We decouple the conditional guidance into domain guidance and control guidance while predicting the control guidance and unconditional guidance using the original diffusion model to keep them unchanged." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.906, + 0.412 + ], + "angle": 0, + "content": "off between quality and diversity, CFG [16] is utilized to combine the conditional and unconditional score estimates with a hyperparameter \\( w \\):" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.424, + 0.906, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\epsilon} _ {\\theta} \\left(z _ {t}, c\\right) = (1 + w) \\epsilon_ {\\theta} \\left(z _ {t}, c\\right) - w \\epsilon_ {\\theta} \\left(z _ {t}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.906, + 0.512 + ], + "angle": 0, + "content": "In large-scale diffusion model training, suppose that an open-world training dataset denotes as \\(\\{z_i, c_i\\}\\), where \\(z_i\\) is the training image and \\(c_i\\) is the correspondent text, the training objective is to minimize:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.524, + 0.905, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left(\\underbrace {\\mathcal {L} \\left(p _ {\\theta} (z \\mid c) , p (z \\mid c)\\right)} _ {\\text {c o n d i t i o n a l g u i d a n c e}} + \\underbrace {\\mathcal {L} \\left(p _ {\\theta} (z) , p (z)\\right)} _ {\\text {u n c o n d i t i o n a l g u i d a n c e}}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Guidance Catastrophic Forgetting. During domain data fine-tuning with data \\(\\{z_i, y_i\\}\\), these two guidances inevitably deviate, which we illustrated in Fig. 3. Notably, \\(p(z|y)\\) represents the probability conditioned on a given domain, which does not imply that text labels of domain data are required. For conditional guidance, the condition \\(y\\) in fine-tuning phase is too much simpler than the condition in the pre-trained phase \\(c\\), where the latter is expected to be control generation. For example, although we leverage only facial images in training, we also desire to generate images controlled by text prompts. For unconditional guidance, we suppose that the pre-trained model could estimate the accurate \\(p_{\\theta_s d}(z_t|\\emptyset)\\) since it is trained on open-world data. Hence, fine-tuning on domain data would make \\(p_{\\theta^*}(z_t|\\emptyset)\\) inaccuracy, illustrated in Fig. 4. The generative results from fine-tuned models reflect the characteristics of the training datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.857, + 0.864 + ], + "angle": 0, + "content": "3.2. Guidance-Decoupled Prior Preservation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To address the forgetting issue, we propose a guidance-decoupled prior preservation mechanism. We decouple con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.955 + ], + "angle": 0, + "content": "18360" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.272 + ], + "angle": 0, + "content": "ditional guidance into domain guidance and control guidance, as illustrated in Fig 5(b). The domain guidance steers the denoising process to generate domain-aligned images and the control guidance takes charge of handling openworld controls in generation. For example, to generate 'a man wearing glasses' within the face domain, domain guidance tells how the human face looks, and control guidance ensures the synthetic faces follow 'man' and 'wearing glasses'. After such decoupling, control guidance and unconditional guidance can be predicted by diffusion priors and domain guidance can be learned by an additional model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.277, + 0.485, + 0.368 + ], + "angle": 0, + "content": "Conditional Guidance Decoupling. Denoting a sets of conditions as \\( C = \\{c_1, \\dots, c_k\\} \\) with manually defined intensities \\( w = \\{w_1, \\dots, w_k\\} \\), we can use multiple UN-ets \\( \\epsilon_{\\theta} \\) to predict conditional and unconditional: \\( \\epsilon_{\\theta_0}(z_t) = \\nabla_z \\log p(z_t) \\) and \\( \\epsilon_{\\theta_i}(z_t, c_i) = \\nabla_z \\log p(z_t | c_i) \\). The reverse process in each timestamp is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.38, + 0.484, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} \\left(z _ {t}, C\\right) = \\epsilon_ {\\theta_ {0}} \\left(z _ {t}\\right) + \\sum_ {i = 1} ^ {K} w _ {i} \\left(\\epsilon_ {\\theta_ {i}} \\left(z _ {t}, c _ {i}\\right) - \\epsilon_ {\\theta_ {0}} \\left(z _ {t}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.539 + ], + "angle": 0, + "content": "Hence, we can decouple the original conditional guidance into two components to improve domain alignment and open-world control, respectively. As shown in Eq. (6), we can use different diffusion models to predict each guidance, allowing us to fine-tune a UNet copy to learn domain knowledge and utilize the pre-trained model to control the generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.541, + 0.484, + 0.706 + ], + "angle": 0, + "content": "However, Eq. (6) assumes that all conditions are independent. This assumption leads to the independent maximization of the probability of each condition, which can cause confusion between the conditions. This issue is analyzed further in Appendix A. Our analysis reveals that a simple yet effective approach to mitigate this issue is to use more detailed descriptions (e.g., a photo of a face, wearing glasses) rather than overly simplified conditions (e.g., wearing glasses). This approach helps reduce confusion and better aligns the guidance with the intended domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.71, + 0.484, + 0.832 + ], + "angle": 0, + "content": "Unconditional Guidance Rectification. Through multi-guidance generation, an intuitive solution to address unconditional guidance drift involves rectifying it using prior diffusion models. We can consider the original pre-trained model as an unconditional guidance predictor, effectively mitigating guidance drift after fine-tuning. This rectification operation can also used in other customized diffusion models, including open-sourced models, as shown in Sec 4.5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.843, + 0.424, + 0.86 + ], + "angle": 0, + "content": "3.3. Efficient Domain Knowledge Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Null-text Diffusion Model. The additional fine-tuned diffusion model aims to learn the domain guidance by a do" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "main image dataset. To make it concise, we construct a null-text diffusion model to learn domain guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.907, + 0.152 + ], + "angle": 0, + "content": "In LDMs, cross-attention is used to fuse text feature and image feature with residual connection, which follows:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.16, + 0.907, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {i m g} ^ {\\prime} = \\underbrace {W (S o f t m a x (\\frac {\\mathcal {Q} _ {i m g} \\mathcal {K} _ {t e x t} ^ {T}}{\\sqrt {d}}) \\times \\mathcal {V} _ {t e x t})} _ {t e x t f e a t u r e i n j e c t i o n} + \\mathcal {F} _ {i m g} (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.224, + 0.907, + 0.3 + ], + "angle": 0, + "content": "where \\(\\mathcal{Q}_{img}\\) is transformed from input image feature \\(\\mathcal{F}_{img}\\), \\(\\mathcal{K}_{text}^{T},\\mathcal{V}_{text}\\) is transformed from text feature, \\(d\\) is the num of channels, and \\(W\\) is a linear projection. Since textual information is not used in our fine-tuned model, we can convert the former term into a fixed embedding \\(E\\), and we have:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.324, + 0.907, + 0.344 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {i m g} ^ {\\prime} = E + \\mathcal {F} _ {i m g} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.909, + 0.522 + ], + "angle": 0, + "content": "Initializing \\( E \\) via Textual Features Optimization. To better inherent generative capability from diffusion priors, accurate textual features represent a good initializing point in training. Therefore, we draw inspiration from Textual Inversion[13] and optimize the text features \\( F^T \\), the output of text encoder, and use them as the initial values for \\( E = W_V(F^T) \\). Specifically, we first attain \\( F^T \\) by feeding a photo of into the text encoder and optimize \\( F^T \\) by denoising loss. This training process converges in just several minutes, providing a good initial value that accelerates the subsequent fine-tuning process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.888, + 0.545 + ], + "angle": 0, + "content": "3.4. In-domain Generation with Multi-Guidance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.673 + ], + "angle": 0, + "content": "During domain knowledge learning, we fine-tune a domain diffusion model \\(\\epsilon_{\\theta_d}\\) to provide domain guidance and employ the Stable Diffusion \\(\\epsilon_{sd}\\) to provide preserved guidance, as illustrated in Fig 5. For control guidance, we can also flexibly adapt existing techniques, including Stable Diffusion, ControlNet, or any open-sourced customized models, to achieve a variety of applications. The reverse process suffices:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.684, + 0.904, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\underbrace {(1 - w _ {d} - w _ {c}) \\epsilon_ {s d} (z _ {t})} _ {u n c o n d i t i o n a l g u i d a n c e} + \\underbrace {w _ {d} \\epsilon_ {\\theta_ {d}} (z _ {t})} _ {d o m a i n g u i d a n c e} + \\underbrace {w _ {c} \\epsilon_ {\\theta_ {c}} (z _ {t} , c)} _ {c o n t r o l g u i d a n c e}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(c\\) is control signal with model \\(\\epsilon_{\\theta_c}\\). For instance, to generate images conditioned by a canny image in the animal domain, we utilize the diffusion model fine-tuned on an animal dataset as \\(\\epsilon_{\\theta_d}\\) and the original diffusion model equipped with canny-to-image ControlNet model as \\(\\theta_c\\). Experimentally, \\(w_d\\) presents stability in each generative task, similar to the guidance scale used in widely-used diffusion models (e.g., using \\(w_0 = 7\\) for Stable Diffusion \\(\\epsilon_{sd}\\)). And we determine the initial value of \\(w_c\\) and \\(w_d\\) by \\(w_d = \\frac{w_0}{2} \\cdot \\frac{\\|\\epsilon_{sd}\\|}{\\|\\epsilon_{\\theta_d}\\|}\\) and \\(w_c = \\frac{w_0}{2} \\cdot \\frac{\\|\\epsilon_{sd}\\|}{\\|\\epsilon_{\\theta_c}\\|}\\) to ensure \\(w_0 \\cdot \\|\\epsilon_{sd}\\| = w_d \\|\\epsilon_{\\theta_d}\\| + w_c \\|\\epsilon_{\\theta_c}\\|\\). We can also employ" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18361" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "this generation process in other diffusion-based generative tasks, like image editing and 3D generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.138, + 0.226, + 0.156 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.164, + 0.3, + 0.181 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.191, + 0.485, + 0.342 + ], + "angle": 0, + "content": "Datasets. In this study, we conduct experiments on three domains: face (i.e., FFHQ [20]), animal (i.e., AFHQv2 [7]), and porcelain (i.e. a collected dataset), chosen for their diversity and high-resolution (larger than \\(512 \\times 512\\)). FFHQ is a high-quality human face dataset, containing 70,000 images. Images come from Flickr and are aligned into \\(1024 \\times 1024\\). We further resize images into \\(512 \\times 512\\) without crop. AFHQv2 is an animal face dataset with a resolution of \\(512 \\times 512\\), containing 15803 images. The porcelain dataset contains 5,000 images, collected from the internet." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.485, + 0.573 + ], + "angle": 0, + "content": "Settings. We train the null-text domain diffusion model based on Stable Diffusion v1.5 (SD1.5) at the resolution of 512. For embedding learning phase, 3,200 images are used to attain initial embeddings with batch size of 32 and 100 steps for all domains. For all experiments, we maintain a constant learning rate of 1e-5 with the optimizer of AdamW and a batch size of 32. These experiments are conducted on 8 NVIDIA A100 GPUs, with each running until convergence. All controlling models, including ControlNet and open-sourced customized diffusion models, are borrowed from their official releases based on the corresponding version of Stable Diffusion without any training. We employ the DPM-Solver++ scheduler [28] with 100 steps across all models. For more experimental settings, please refer to Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.585, + 0.433, + 0.601 + ], + "angle": 0, + "content": "4.2. Results of In-domain Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.486, + 0.835 + ], + "angle": 0, + "content": "We present the unconditional generation results in Fig. 6 to illustrate the degree of domain alignment. As shown, our results closely match the distribution of the domain datasets, highlighting the effectiveness of our approach in aligning the model outputs with domain data. For conditional generation, we showcase the qualitative results in Fig. 7. We conduct experiments on both text-conditioned and spatial-conditioned generation, utilizing ControlNet models for the latter task. The results clearly demonstrate that our method not only achieves high fidelity with respect to the target domains but also retains the generative controllability seen in pre-trained models, ensuring that the model can be guided by text prompts or spatial conditions as intended. Quantitative results and comparisons can be found in Tab. 1 and Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.48, + 0.863 + ], + "angle": 0, + "content": "4.3. Results of Other In-domain Generation Tasks" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Apart from the basic generative tasks, our proposed method also benefits other in-domain generation tasks. In this part," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "we further demonstrate the effectiveness of our method in a wider range of tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.126, + 0.907, + 0.293 + ], + "angle": 0, + "content": "Image Generation under Complex Conditions. As discussed in Sec. 3.2, multiple guidances can be leveraged to control the generation process. In this section, we present the generation results under complex conditions in Fig. 8. Despite the introduction of complex conditions, the generated images maintain high fidelity to the target domain, demonstrating both the robustness and quality of our method. Additionally, our generative framework allows us to leverage additional open-source models for controlling the generation process, such as using DreamShaper to control the style." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.296, + 0.909, + 0.403 + ], + "angle": 0, + "content": "Image Editing. In Fig. 9, we implement SDEdit[29] on the face domain, where the input image is first encoded into latent space and added noise and then we use a similar generation pipeline presented in Sec. 3.4. Detailed settings can be found in Appendix C.3.1. Our results are more aligned with the distribution of real human face images, demonstrating better visual quality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.405, + 0.909, + 0.603 + ], + "angle": 0, + "content": "Text-to-3D Generation. In Fig. 10, we leverage our model on porcelain domain for text-to-3d porcelain generation and compare it with classic text-to-3d model [33] on SD1.5. Specifically, we begin by generating porcelain images based on the provided prompt, which are then converted into 3D models using an off-the-shelf image-to-3D method [25]. Subsequently, we fine-tune the resulting 3D models using our porcelain model, guided by the SDS [33] loss. Lastly, we directly convert the NeRF models into meshes. Detailed settings and additional ablation studies on 3D generation can be found in Appendix C.3.2. Our results demonstrate superiority in color richness, fine detail, and alignment with the dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.613, + 0.857, + 0.63 + ], + "angle": 0, + "content": "4.4. Comparisons to Fine-tuning Techniques" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.906, + 0.728 + ], + "angle": 0, + "content": "To validate the superiority of our proposed method, we compare our method with several fine-tuning baselines, including vanilla fine-tuning, parameter-efficient fine-tuning, and some training techniques borrowed from other tasks. For detailed experimental settings, please refer to AppendixC.2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.731, + 0.909, + 0.867 + ], + "angle": 0, + "content": "Baselines. To establish comprehensive baselines, we conduct experiments on several training techniques, including vanilla fine-tuning, low-rank fine-tuning [18], textual inversion [13], custom diffusion [22], and OFT mechanism [34]. To train diffusion models without text annotation, we adopt prompt templates, utilizing formats such as a photo of [V] . These additional tokens [V] are omitted in compared methods where text embedding is not trained." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Evaluation Setting. For unconditional generation, we calculate FID on 50,000 images. For text-guided generation," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "18362" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.088, + 0.907, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.371, + 0.908, + 0.4 + ], + "angle": 0, + "content": "Figure 6. Visualization of Domain Alignment. Our results are more aligned to given domain dataset. For SD1.5 and SDXL 1.0, we generate images by text a photo of ." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.412, + 0.907, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.656, + 0.908, + 0.699 + ], + "angle": 0, + "content": "Figure 7. Results of conditional in-domain image generation. We demonstrate text-conditioned and spatial-conditioned generation results to verify that our proposed methods can effectively preserve the control capability from the pre-trained model. For spatial-conditioned generation, we leverage the off-the-shelf ControlNet model without any extra training." + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.712, + 0.825, + 0.835 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.847, + 0.908, + 0.876 + ], + "angle": 0, + "content": "Figure 8. Generative Results under Complex Conditions. We employ multiple conditions and models to demonstrate the performance of our proposed method in achieving more detailed control over the generation process." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.955 + ], + "angle": 0, + "content": "18363" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.131 + ], + "angle": 0, + "content": "Table 1. Quantitative results of unconditional generation We evaluate FID and human preference on two datasets. Our method achieves excellent generation results among baselines.\\(^{\\dagger}\\): reproduced results in \\(512 \\times 512\\) resolution. Bold: the best result. Underline: the second best result." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.143, + 0.907, + 0.278 + ], + "angle": 0, + "content": "
Method#param.Unconditional GenerationConditional Generation
FaceAnimalPorcelainText-GuidedSpatial-Guided
FID↓Pref.↑FID↓Pref.↑FID↓Pref.↑Align.↑Pref.↑Align.↓Pref.↑
Textual Inversion [13]< 1K70.620.3%44.752.5%112.761.7%0.970.2%0.221.5%
Custom Diffusion [22]44M40.981.6%49.981.8%98.770.6%0.741.4%0.201.7%
OFT [34]23M27.502.9%38.741.4%68.841.2%0.773.2%0.219.6%
Fine-tuning859M12.376.4%23.9212.4%62.7610.8%0.244.5%0.224.9%
Fine-tuning + LoRA[18]0.19M23.761.3%24.191.7%73.460.3%0.766.7%0.229.1%
Ours810M6.5787.5%18.8280.2%56.4685.4%0.8984.0%0.2073.2%
" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.285, + 0.485, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.49 + ], + "angle": 0, + "content": "Figure 9. Image editing on face domain. We follow SDEdit[29] to edit facial images equipped with our proposed methods." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.505, + 0.485, + 0.773 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.786, + 0.483, + 0.843 + ], + "angle": 0, + "content": "Figure 10. 3D generation on porcelain domain. We use Dream-Fusion [33] based on SD1.5 for comparison. We present two views of the generated objects, including their mesh edges. * denotes the prompt 'a DSLR photo of'." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "we generate 1,000 images over 10 attribute-related prompts on the facial domain and employ facial predictors to ascertain whether these attributes are accurately represented in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.907, + 0.38 + ], + "angle": 0, + "content": "the generated images. Besides, we sample 200 images from CelebA-HQ dataset to extract canny information as the spatial condition and compute the canny discrepancy. We also report human preference (denoted as 'Pref') by collecting the win rates of generative images across all comparisons. Detailed settings are presented in Appendix C.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.909, + 0.564 + ], + "angle": 0, + "content": "Qualitative Comparisons. We conducted a qualitative comparison of our method against previous training techniques, as depicted in Fig 11. Methods like textual inversion, custom diffusion, and LoRA, which adhere to a parameter-efficient form, exhibit noticeably lower fidelity compared to others. Vanilla fine-tuning, in particular, demonstrates a decline in controllability during extended training periods. This results in a diminished capacity for text-guided generation and the emergence of unnatural artifacts, especially when used in conjunction with ControlNet. In contrast, our method excels in both fidelity and control capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Quantitative Comparisons. We conduct the evaluation of unconditional generation and report the results in Tab 1. Our approach demonstrates state-of-the-art generation quality, evidenced by a significantly lower FID compared to baseline methods and the highest human preference rating. For text-guided generation, Textual Inversion shows a higher alignment with the text prompts than our method. However, this is achieved at the cost of image quality, owing to its reliance on a very limited set of learnable parameters. This compromise in quality is clearly evident in the comparisons presented in Fig 11. In contrast, in spatial-guided generation, where semantic features are integrated into the UNet, the alignment degree among different methods is similar. Notably, our results not only align well with the control signals but also achieve the highest ratings in human preference assessments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.908, + 0.848 + ], + "angle": 0, + "content": "4.5. Effects of Unconditional Guidance Rectification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "In previous methods, unconditional guidance is predicted by the customized model, leading to inaccurate estimation of probability in sampling. In our framework, the uncon-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "18364" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.092, + 0.907, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.339, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Figure 11. Comparisons to existing training methods. We evaluate in-domain image generation in unconditional, text-guided, and spatial-guided cases. We fix the random seed to illustrate the controllability for the former two." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.379, + 0.353, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.161, + 0.585, + 0.298, + 0.596 + ], + "angle": 0, + "content": "(a) In-domain Generation" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.374, + 0.892, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.584, + 0.727, + 0.596 + ], + "angle": 0, + "content": "(b) Open-Sourced Diffusion Models" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.609, + 0.907, + 0.636 + ], + "angle": 0, + "content": "Figure 12. Effects of Unconditional Guidance Rectification. (a) Effects in Facial Domain Generation. (b) Unconditional guidance rectification also improves the open-sourced diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.822 + ], + "angle": 0, + "content": "ditional guidance is rectified by straightforwardly predicting by diffusion priors. We compare generation results in Fig 12a. Our method significantly improves image quality, generating fewer artifacts and attaining photorealistic results. Besides, we utilize four widely-used open-sourced models. The comparative results are presented in Fig 12b. Without rectification, the generation quality shows slight degradation, as evidenced by less detailed eyes and less realistic hair textures. The implementation of guidance decoupling substantially enhances generative performance via the rectification of unconditional guidance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.843, + 0.21, + 0.859 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this research, we address the task of enhancing large-scale diffusion models for in-domain generation using only" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.899 + ], + "angle": 0, + "content": "image data. Our goal is to enable these models to generate images that faithfully represent the given domains, achieving high fidelity, diversity, and controllability. Moreover, our approach is designed to be compatible with various control methods, allowing the model to perform a wide range of generative tasks. To tackle this challenge, we propose a guidance-decoupled prior preservation mechanism that separates conditional guidance into two components: domain guidance for alignment and control guidance for maintaining the models controllability. Additionally, we introduce an efficient learning approach by incorporating a null-text diffusion model, ensuring both simplicity and effectiveness. We then outline the generative pipeline of our method to achieve diverse in-domain generation tasks. We conduct experiments across three domains to demonstrate the effectiveness of our proposed method." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.955 + ], + "angle": 0, + "content": "18365" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.486, + 0.147 + ], + "angle": 0, + "content": "This work was supported by the Young Scientists Fund of NSFC (Grant No. 62406035)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.162, + 0.188, + 0.178 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.188, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[1] Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H. Bermano. Domain-agnostic tuning-encoder for fast personalization of text-to-image models. pages 1–10, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[2] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[3] Pu Cao, Lu Yang, Dongxu Liu, Zhiwei Liu, Shan Li, and Qing Song. Lsap: Rethinking inversion fidelity, perception and editability in gan latent space. arXiv preprint arXiv:2209.12746, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[4] Pu Cao, Lu Yang, Dongxv Liu, Xiaoya Yang, Tianrui Huang, and Qing Song. What decreases editing capability? domain-specific hybrid refinement for improved gan inversion. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 4240-4249, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[5] Hong Chen, Yipeng Zhang, Simin Wu, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. Disenbooth: Identity-preserving disentangled tuning for subject-driven text-to-image generation. arXiv preprint arXiv:2305.03374, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.503, + 0.482, + 0.6 + ], + "angle": 0, + "content": "[6] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 100–116. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[7] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188-8197, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[8] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171-4186, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. In Proceedings of the IEEE/CVF Inter-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.188, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "national Conference on Computer Vision, pages 7430-7440, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[12] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model. Advances in Neural Information Processing Systems, 35:8882-8895, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 4, 5, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[14] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36:15890-15902, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.464, + 0.904, + 0.517 + ], + "angle": 0, + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.649, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[21] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6007-6017, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[22] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1931-1941, 2023. 5, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22511-22521, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Zuoxin Li, Fuqiang Zhou, Lu Yang, Xiaojie Li, and Juan Li. Accelerate neural style transfer with super-resolution. Multimedia Tools and Applications, 79:4347-4364, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "18366" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[25] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[26] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. arXiv preprint arXiv:2303.05125, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[27] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Customizable image synthesis with multiple subjects. Advances in neural information processing systems, 36:57500-57519, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[28] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[29] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[30] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[32] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, JunYan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11399-11409, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[34] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 2, 3, 5, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[36] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[38] Daniel Reich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.248 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[43] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[45] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 44(4):2004-2018, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[46] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. pages 8543-8552, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[47] James Seale Smith, Yen-Chang Hsu, Lingyu Zhang, Ting Hua, Zsolt Kira, Yilin Shen, and Hongxia Jin. Continual diffusion: Continual customization of text-to-image diffusion with c-lora. arXiv preprint arXiv:2304.06027, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[48] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIG-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "18367" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "GRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[50] Yujin Wang, Lingen Li, Tianfan Xue, and Jinwei Gu. Reconstruct-and-generate diffusion model for detail-preserving image denoising. arXiv preprint arXiv:2309.10714, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[51] Zhizhong Wang, Lei Zhao, and Wei Xing. *Style diffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[52] Chanyue Wu, Dong Wang, Yunpeng Bai, Hanyu Mao, Ying Li, and Qiang Shen. Hsr-diff: Hyperspectral image superresolution via conditional diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[53] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7378-7387, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[54] Lu Yang, Qing Song, Zhihui Wang, and Ming Jiang. Parsing r-cnn for instance-level human analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 364-373, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[55] Lu Yang, Qing Song, Zhihui Wang, Mengjie Hu, Chun Liu, Xueshi Xin, Wenhe Jia, and Songcen Xu. Renovating parsing r-cnn for accurate multiple human parsing. In European Conference on computer vision, pages 421-437. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[56] Lu Yang, Qing Song, Xueshi Xin, and Zhiwei Liu. Quality-aware network for face parsing. arXiv preprint arXiv:2106.07368, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[57] Lu Yang, Zhiwei Liu, Tianfei Zhou, and Qing Song. Part decomposition and refinement network for human parsing. IEEE/CAA Journal of Automatica Sinica, 9(6):1111-1114, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[58] Lu Yang, Qing Song, Zhihui Wang, Zhiwei Liu, Songcen Xu, and Zhihao Li. Quality-aware network for human parsing. IEEE Transactions on Multimedia, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.631, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[59] Lu Yang, Wenhe Jia, Shan Li, and Qing Song. Deep learning technique for human parsing: A survey and outlook. International Journal of Computer Vision, 132(8):3270-3301, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[60] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22873-22882, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.955 + ], + "angle": 0, + "content": "18368" + } + ] +] \ No newline at end of file diff --git a/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/e387f9f2-242c-46fd-9d3c-463d01e38cba_origin.pdf b/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/e387f9f2-242c-46fd-9d3c-463d01e38cba_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6cea49cb1ac4739d5028a6c78bc501e9b9ec703f --- /dev/null +++ b/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/e387f9f2-242c-46fd-9d3c-463d01e38cba_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db46d593b403c78fc538e6a44faf52623b13e9ce06c67e51012438883250c1a9 +size 9219693 diff --git a/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/full.md b/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6a7cb61658131d809aaab2e8b2ecb1fa521c2192 --- /dev/null +++ b/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/full.md @@ -0,0 +1,285 @@ +# Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation + +Pu Cao† Feng Zhou† Lu Yang* Tianrui Huang Qing Song +Beijing University of Posts and Telecommunications + +{caopu, zhoufeng, soeaver, huangtianrui, priv}@bupt.edu.cn + +![](images/95bb0a65ef097a00e6d25ecbb6219a0ea0cd05ce3fa3094a6361d49897fa9e15.jpg) +Figure 1. Illustration of In-domain Generation. In this work, we empower large-scale pre-trained diffusion models using only image data to perform varied generation tasks within each domain with high fidelity and controllability. We mark the input data in blue and the results generated with the original Stable Diffusion v1.5 model in orange. + +# Abstract + +In-domain generation aims to perform a variety of tasks within a specific domain, such as unconditional generation, text-to-image, image editing, 3D generation, and more. Early research typically required training specialized generators for each unique task and domain, often relying on fully-labeled data. Motivated by the powerful generative capabilities and broad applications of diffusion models, we are driven to explore leveraging label-free data to empower these models for in-domain generation. Fine-tuning a pre-trained generative model on domain data is an intuitive but challenging way and often requires complex manual hyper-parameter adjustments since the limited diversity of the training data can easily disrupt the model's origi + +nal generative capabilities. To address this challenge, we propose a guidance-decoupled prior preservation mechanism to achieve high generative quality and controllability by image-only data, inspired by preserving the pre-trained model from a denoising guidance perspective. We decouple domain-related guidance from the conditional guidance used in classifier-free guidance mechanisms to preserve open-world control guidance and unconditional guidance from the pre-trained model. We further propose an efficient domain knowledge learning technique to train an additional text-free UNet copy to predict domain guidance. Besides, we theoretically illustrate a multi-guidance in-domain generation pipeline for a variety of generative tasks, leveraging multiple guidances from distinct diffusion models and conditions. Extensive experiments demonstrate the superiority of our method in domain-specific synthesis and its compatibility with various diffusion-based control methods and applications. + +![](images/6448d8d9d43727a287d2a6293791d0aef64d49879a1b979339d3eb783f51299c.jpg) +a photo of face +a photo of face wearing hat +Figure 2. Challenge of fine-tuning diffusion models on domain data. We show the fine-tuning process of Stable Diffusion v1.5 with a facial image dataset (i.e., FFHQ[20]). The open-world controllability of pre-trained diffusion models is gradually decreased during the fine-tuning process, with domain fidelity improved. + +# 1. Introduction + +In-domain generation aims to accomplish various tasks within a specific domain, often requiring tailored approaches and labeled data. For instance, pix2pix [19] was developed for image-to-image translation (e.g., converting segmentation or edge images to realistic images), relying on paired images for training. Similarly, InterfaceGAN[45] enables attribute-based image editing but requires attribu-telabeled images to train editing vectors. Such traditional methods[6, 12, 19, 45] depend heavily on specific generative tasks and data domains, which limits their generalization capability. + +Thanks to billion-scale image-text datasets [2, 37, 43, 44] and advances in large-scale natural language models [9, 35, 36], recent research on diffusion models has pushed the boundaries of AI-generated content, greatly enhancing both the diversity and controllability of generated outputs [39]. With various control mechanisms and approaches (e.g., ControlNet[61], SDEdit[29]), diffusion models can also perform diverse generative tasks beyond text-to-image synthesis, alleviating the data demands of traditional generative methods. However, the generated results often fail to fully align with specific data domains, as shown in orange in Fig. 1. + +While fine-tuning a pre-trained model on domain-specific data is an intuitive way to inject domain knowledge, it is challenging to balance fidelity with controllability, as illustrated in Fig. 2. For example, fine-tuning can improve the quality of face generation but may gradually reduce the effect of control prompts like wearing hat. In this work, we first identify this phenomenon caused by conditional guidance and unconditional guidance catastrophic forgetting during fine-tuning. These two guidances are integrated by classifier-free guidance for improved generative capability, which is employed in most text-to-image diffusion models. Then, we propose a guidance-decoupled prior preservation method to address these challenges. We decouple the conditional guidance in CFG into two parts: domain guidance, which is optimized in fine-tuning to enhance fi + +dely, and control guidance, which is preserved for controllability. With introducing multiple guidances in the synthesis process, the control guidance and unconditional guidance, which we want to preserve during personalization, are directly predicted by diffusion priors, while the domain guidance is predicted by an independent diffusion model. Additionally, we introduce an efficient domain knowledge learning mechanism, designing a null-text Diffusion Model to learn domain knowledge in a concise way. + +We conduct extensive experiments to demonstrate the effects of our proposed task and validate the superiority of our approach. To summarize, our contributions are as follows: + +- We present the task of aligning large-scale diffusion models with specific domains using only image data to perform a variety of generative tasks. +- We propose a guidance-decoupled prior preservation mechanism to address guidance forgetting during finetuning, which decouples domain guidance to learn domain knowledge with other guidance preserved. Besides, we propose an efficient learning mechanism to learn domain knowledge. +- We present the pipeline of our proposed method for in-domain generation tasks and conduct experiments and comparisons across multiple domains and tasks to demonstrate the effectiveness and advancement of our approach. + +# 2. Related Works + +Text-to-image Generation. As the scale of image-text data grows [2, 37, 43, 44], large-scale diffusion models [10, 17, 31, 39] have demonstrated astonishing generative capabilities, able to produce images based on open-world prompts. Some models, such as Imagen [42] and Stable Diffusion [39], have been explored as diffusion priors, applied in downstream tasks like controlled generation [23, 30, 49, 61] and image editing [8, 15, 21]. For instance, ControlNet [61] proposes using an additional network to receive more spatial controls to guide generation. Furthermore, diffusion models have demonstrated powerful capabilities in many tasks, such as face parsing [54-59], style transfer [24, 51, 60], image manipulation [11, 53] and super-resolution [50, 52]. + +Fine-tuning Large-scale Diffusion Models. Some existing methods study learning subjects into large-scale diffusion priors and generate variant images guided by different prompts [1, 5, 13, 26, 27, 34, 41, 46, 47]. Some methods fine-tune the part of diffusion models with one or a few images of the specific subject [5, 13, 34, 41]. Textual inversion [13] tunes the text embedding of the subject placeholder, while DreamBooth [41] further updates the weight of UNet [40]. Some methods further study training-free [1, 46], continual customization [47], or multi-subjects + +![](images/914999c871df7759bd2b8e8a8a94cb51c86011d743d022bc74c95b9eba7fbc9e.jpg) +Figure 3. Illustration of Fine-tuning Process. We demonstrate the guidance catastrophic forgetting during fine-tuning process. + +![](images/2ab9bc7c2237b876a4da92e539edc06de6f732d4052dbbe96a36a60609fb5a14.jpg) +Figure 4. Unconditional Guidance Drift. The unconditional generation results of fine-tuned diffusion models reflect the visual pattern of training datasets, which would cause inaccurate noise estimation. + +personalization [26, 27] and layout-guided generation [14]. However, these methods concentrate on the consistency of given subjects and training efficiency, tailored for the situation with small-scale image datasets and limited training process. + +Fidelity and Controllability Tradeoff. In the field of generation, the tradeoff between fidelity and controllability has received widespread attention. In GANs, many inversion methods address this problem through regularization [3, 38], latent code discriminator [48], and invertibility prediction [4, 32]. In the personalization task, thanks to the simple customization process and excellent generalization of diffusion priors, this problem is easily solved by regularization [41], parameter efficient tuning [18], orthogonal fine-tuning [34], and so on. However, a longer training process is required for domain knowledge learning and the existing techniques struggle to solve this problem, highlighting an urgent need for a new approach. + +# 3. Method + +# 3.1. Guidance Catastrophic Forgetting + +In this section, we delve into the phenomenon of catastrophic forgetting in the training process of diffusion models from the perspective of guidance, allowing us to address this issue at its root. + +Classifier-free Guidance (CFG). To achieve a better trade + +![](images/1b0952f14a6b7b9189d0a85b9a75a97994c1146ffb7082c37b4904792e70126b.jpg) +(a) previous fine-tuning +(b) guidance-decoupled generation +Figure 5. Conditional Guidance Decoupling. We compare the guidance estimation between previous customization methods and ours. We decouple the conditional guidance into domain guidance and control guidance while predicting the control guidance and unconditional guidance using the original diffusion model to keep them unchanged. + +off between quality and diversity, CFG [16] is utilized to combine the conditional and unconditional score estimates with a hyperparameter $w$ : + +$$ +\widetilde {\epsilon} _ {\theta} \left(z _ {t}, c\right) = (1 + w) \epsilon_ {\theta} \left(z _ {t}, c\right) - w \epsilon_ {\theta} \left(z _ {t}\right) \tag {1} +$$ + +In large-scale diffusion model training, suppose that an open-world training dataset denotes as $\{z_i, c_i\}$ , where $z_i$ is the training image and $c_i$ is the correspondent text, the training objective is to minimize: + +$$ +\mathbb {E} \left(\underbrace {\mathcal {L} \left(p _ {\theta} (z \mid c) , p (z \mid c)\right)} _ {\text {c o n d i t i o n a l g u i d a n c e}} + \underbrace {\mathcal {L} \left(p _ {\theta} (z) , p (z)\right)} _ {\text {u n c o n d i t i o n a l g u i d a n c e}}\right) \tag {2} +$$ + +Guidance Catastrophic Forgetting. During domain data fine-tuning with data $\{z_i, y_i\}$ , these two guidances inevitably deviate, which we illustrated in Fig. 3. Notably, $p(z|y)$ represents the probability conditioned on a given domain, which does not imply that text labels of domain data are required. For conditional guidance, the condition $y$ in fine-tuning phase is too much simpler than the condition in the pre-trained phase $c$ , where the latter is expected to be control generation. For example, although we leverage only facial images in training, we also desire to generate images controlled by text prompts. For unconditional guidance, we suppose that the pre-trained model could estimate the accurate $p_{\theta_s d}(z_t|\emptyset)$ since it is trained on open-world data. Hence, fine-tuning on domain data would make $p_{\theta^*}(z_t|\emptyset)$ inaccuracy, illustrated in Fig. 4. The generative results from fine-tuned models reflect the characteristics of the training datasets. + +# 3.2. Guidance-Decoupled Prior Preservation + +To address the forgetting issue, we propose a guidance-decoupled prior preservation mechanism. We decouple con + +ditional guidance into domain guidance and control guidance, as illustrated in Fig 5(b). The domain guidance steers the denoising process to generate domain-aligned images and the control guidance takes charge of handling openworld controls in generation. For example, to generate 'a man wearing glasses' within the face domain, domain guidance tells how the human face looks, and control guidance ensures the synthetic faces follow 'man' and 'wearing glasses'. After such decoupling, control guidance and unconditional guidance can be predicted by diffusion priors and domain guidance can be learned by an additional model. + +Conditional Guidance Decoupling. Denoting a sets of conditions as $C = \{c_1, \dots, c_k\}$ with manually defined intensities $w = \{w_1, \dots, w_k\}$ , we can use multiple UN-ets $\epsilon_{\theta}$ to predict conditional and unconditional: $\epsilon_{\theta_0}(z_t) = \nabla_z \log p(z_t)$ and $\epsilon_{\theta_i}(z_t, c_i) = \nabla_z \log p(z_t | c_i)$ . The reverse process in each timestamp is as follows: + +$$ +\hat {\epsilon} \left(z _ {t}, C\right) = \epsilon_ {\theta_ {0}} \left(z _ {t}\right) + \sum_ {i = 1} ^ {K} w _ {i} \left(\epsilon_ {\theta_ {i}} \left(z _ {t}, c _ {i}\right) - \epsilon_ {\theta_ {0}} \left(z _ {t}\right)\right) \tag {3} +$$ + +Hence, we can decouple the original conditional guidance into two components to improve domain alignment and open-world control, respectively. As shown in Eq. (6), we can use different diffusion models to predict each guidance, allowing us to fine-tune a UNet copy to learn domain knowledge and utilize the pre-trained model to control the generation. + +However, Eq. (6) assumes that all conditions are independent. This assumption leads to the independent maximization of the probability of each condition, which can cause confusion between the conditions. This issue is analyzed further in Appendix A. Our analysis reveals that a simple yet effective approach to mitigate this issue is to use more detailed descriptions (e.g., a photo of a face, wearing glasses) rather than overly simplified conditions (e.g., wearing glasses). This approach helps reduce confusion and better aligns the guidance with the intended domain. + +Unconditional Guidance Rectification. Through multi-guidance generation, an intuitive solution to address unconditional guidance drift involves rectifying it using prior diffusion models. We can consider the original pre-trained model as an unconditional guidance predictor, effectively mitigating guidance drift after fine-tuning. This rectification operation can also used in other customized diffusion models, including open-sourced models, as shown in Sec 4.5. + +# 3.3. Efficient Domain Knowledge Learning + +Null-text Diffusion Model. The additional fine-tuned diffusion model aims to learn the domain guidance by a do + +main image dataset. To make it concise, we construct a null-text diffusion model to learn domain guidance. + +In LDMs, cross-attention is used to fuse text feature and image feature with residual connection, which follows: + +$$ +\mathcal {F} _ {i m g} ^ {\prime} = \underbrace {W (S o f t m a x (\frac {\mathcal {Q} _ {i m g} \mathcal {K} _ {t e x t} ^ {T}}{\sqrt {d}}) \times \mathcal {V} _ {t e x t})} _ {t e x t f e a t u r e i n j e c t i o n} + \mathcal {F} _ {i m g} (4) +$$ + +where $\mathcal{Q}_{img}$ is transformed from input image feature $\mathcal{F}_{img}$ , $\mathcal{K}_{text}^{T},\mathcal{V}_{text}$ is transformed from text feature, $d$ is the num of channels, and $W$ is a linear projection. Since textual information is not used in our fine-tuned model, we can convert the former term into a fixed embedding $E$ , and we have: + +$$ +\mathcal {F} _ {i m g} ^ {\prime} = E + \mathcal {F} _ {i m g} \tag {5} +$$ + +Initializing $E$ via Textual Features Optimization. To better inherent generative capability from diffusion priors, accurate textual features represent a good initializing point in training. Therefore, we draw inspiration from Textual Inversion[13] and optimize the text features $F^T$ , the output of text encoder, and use them as the initial values for $E = W_V(F^T)$ . Specifically, we first attain $F^T$ by feeding a photo of into the text encoder and optimize $F^T$ by denoising loss. This training process converges in just several minutes, providing a good initial value that accelerates the subsequent fine-tuning process. + +# 3.4. In-domain Generation with Multi-Guidance + +During domain knowledge learning, we fine-tune a domain diffusion model $\epsilon_{\theta_d}$ to provide domain guidance and employ the Stable Diffusion $\epsilon_{sd}$ to provide preserved guidance, as illustrated in Fig 5. For control guidance, we can also flexibly adapt existing techniques, including Stable Diffusion, ControlNet, or any open-sourced customized models, to achieve a variety of applications. The reverse process suffices: + +$$ +\underbrace {(1 - w _ {d} - w _ {c}) \epsilon_ {s d} (z _ {t})} _ {u n c o n d i t i o n a l g u i d a n c e} + \underbrace {w _ {d} \epsilon_ {\theta_ {d}} (z _ {t})} _ {d o m a i n g u i d a n c e} + \underbrace {w _ {c} \epsilon_ {\theta_ {c}} (z _ {t} , c)} _ {c o n t r o l g u i d a n c e} +$$ + +where $c$ is control signal with model $\epsilon_{\theta_c}$ . For instance, to generate images conditioned by a canny image in the animal domain, we utilize the diffusion model fine-tuned on an animal dataset as $\epsilon_{\theta_d}$ and the original diffusion model equipped with canny-to-image ControlNet model as $\theta_c$ . Experimentally, $w_d$ presents stability in each generative task, similar to the guidance scale used in widely-used diffusion models (e.g., using $w_0 = 7$ for Stable Diffusion $\epsilon_{sd}$ ). And we determine the initial value of $w_c$ and $w_d$ by $w_d = \frac{w_0}{2} \cdot \frac{\|\epsilon_{sd}\|}{\|\epsilon_{\theta_d}\|}$ and $w_c = \frac{w_0}{2} \cdot \frac{\|\epsilon_{sd}\|}{\|\epsilon_{\theta_c}\|}$ to ensure $w_0 \cdot \|\epsilon_{sd}\| = w_d \|\epsilon_{\theta_d}\| + w_c \|\epsilon_{\theta_c}\|$ . We can also employ + +this generation process in other diffusion-based generative tasks, like image editing and 3D generation. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. In this study, we conduct experiments on three domains: face (i.e., FFHQ [20]), animal (i.e., AFHQv2 [7]), and porcelain (i.e. a collected dataset), chosen for their diversity and high-resolution (larger than $512 \times 512$ ). FFHQ is a high-quality human face dataset, containing 70,000 images. Images come from Flickr and are aligned into $1024 \times 1024$ . We further resize images into $512 \times 512$ without crop. AFHQv2 is an animal face dataset with a resolution of $512 \times 512$ , containing 15803 images. The porcelain dataset contains 5,000 images, collected from the internet. + +Settings. We train the null-text domain diffusion model based on Stable Diffusion v1.5 (SD1.5) at the resolution of 512. For embedding learning phase, 3,200 images are used to attain initial embeddings with batch size of 32 and 100 steps for all domains. For all experiments, we maintain a constant learning rate of 1e-5 with the optimizer of AdamW and a batch size of 32. These experiments are conducted on 8 NVIDIA A100 GPUs, with each running until convergence. All controlling models, including ControlNet and open-sourced customized diffusion models, are borrowed from their official releases based on the corresponding version of Stable Diffusion without any training. We employ the DPM-Solver++ scheduler [28] with 100 steps across all models. For more experimental settings, please refer to Appendix C. + +# 4.2. Results of In-domain Image Generation + +We present the unconditional generation results in Fig. 6 to illustrate the degree of domain alignment. As shown, our results closely match the distribution of the domain datasets, highlighting the effectiveness of our approach in aligning the model outputs with domain data. For conditional generation, we showcase the qualitative results in Fig. 7. We conduct experiments on both text-conditioned and spatial-conditioned generation, utilizing ControlNet models for the latter task. The results clearly demonstrate that our method not only achieves high fidelity with respect to the target domains but also retains the generative controllability seen in pre-trained models, ensuring that the model can be guided by text prompts or spatial conditions as intended. Quantitative results and comparisons can be found in Tab. 1 and Sec. 4.4. + +# 4.3. Results of Other In-domain Generation Tasks + +Apart from the basic generative tasks, our proposed method also benefits other in-domain generation tasks. In this part, + +we further demonstrate the effectiveness of our method in a wider range of tasks. + +Image Generation under Complex Conditions. As discussed in Sec. 3.2, multiple guidances can be leveraged to control the generation process. In this section, we present the generation results under complex conditions in Fig. 8. Despite the introduction of complex conditions, the generated images maintain high fidelity to the target domain, demonstrating both the robustness and quality of our method. Additionally, our generative framework allows us to leverage additional open-source models for controlling the generation process, such as using DreamShaper to control the style. + +Image Editing. In Fig. 9, we implement SDEdit[29] on the face domain, where the input image is first encoded into latent space and added noise and then we use a similar generation pipeline presented in Sec. 3.4. Detailed settings can be found in Appendix C.3.1. Our results are more aligned with the distribution of real human face images, demonstrating better visual quality. + +Text-to-3D Generation. In Fig. 10, we leverage our model on porcelain domain for text-to-3d porcelain generation and compare it with classic text-to-3d model [33] on SD1.5. Specifically, we begin by generating porcelain images based on the provided prompt, which are then converted into 3D models using an off-the-shelf image-to-3D method [25]. Subsequently, we fine-tune the resulting 3D models using our porcelain model, guided by the SDS [33] loss. Lastly, we directly convert the NeRF models into meshes. Detailed settings and additional ablation studies on 3D generation can be found in Appendix C.3.2. Our results demonstrate superiority in color richness, fine detail, and alignment with the dataset. + +# 4.4. Comparisons to Fine-tuning Techniques + +To validate the superiority of our proposed method, we compare our method with several fine-tuning baselines, including vanilla fine-tuning, parameter-efficient fine-tuning, and some training techniques borrowed from other tasks. For detailed experimental settings, please refer to AppendixC.2. + +Baselines. To establish comprehensive baselines, we conduct experiments on several training techniques, including vanilla fine-tuning, low-rank fine-tuning [18], textual inversion [13], custom diffusion [22], and OFT mechanism [34]. To train diffusion models without text annotation, we adopt prompt templates, utilizing formats such as a photo of [V] . These additional tokens [V] are omitted in compared methods where text embedding is not trained. + +Evaluation Setting. For unconditional generation, we calculate FID on 50,000 images. For text-guided generation, + +![](images/8a7cd6ef72f1975646f225aa315add948730f28205ad6c5a124a5aacc2bea4f5.jpg) +Figure 6. Visualization of Domain Alignment. Our results are more aligned to given domain dataset. For SD1.5 and SDXL 1.0, we generate images by text a photo of . + +![](images/c9b89f2a8e62df86af02530665c825501ec606754b8dbf89bd237493e6980ffd.jpg) +Figure 7. Results of conditional in-domain image generation. We demonstrate text-conditioned and spatial-conditioned generation results to verify that our proposed methods can effectively preserve the control capability from the pre-trained model. For spatial-conditioned generation, we leverage the off-the-shelf ControlNet model without any extra training. + +![](images/a1945ed6c23c3f6afe7308363bbd95866d935b0cefd3e51d4828aa607049e3ce.jpg) +Figure 8. Generative Results under Complex Conditions. We employ multiple conditions and models to demonstrate the performance of our proposed method in achieving more detailed control over the generation process. + +Table 1. Quantitative results of unconditional generation We evaluate FID and human preference on two datasets. Our method achieves excellent generation results among baselines. $^{\dagger}$ : reproduced results in $512 \times 512$ resolution. Bold: the best result. Underline: the second best result. + +
Method#param.Unconditional GenerationConditional Generation
FaceAnimalPorcelainText-GuidedSpatial-Guided
FID↓Pref.↑FID↓Pref.↑FID↓Pref.↑Align.↑Pref.↑Align.↓Pref.↑
Textual Inversion [13]< 1K70.620.3%44.752.5%112.761.7%0.970.2%0.221.5%
Custom Diffusion [22]44M40.981.6%49.981.8%98.770.6%0.741.4%0.201.7%
OFT [34]23M27.502.9%38.741.4%68.841.2%0.773.2%0.219.6%
Fine-tuning859M12.376.4%23.9212.4%62.7610.8%0.244.5%0.224.9%
Fine-tuning + LoRA[18]0.19M23.761.3%24.191.7%73.460.3%0.766.7%0.229.1%
Ours810M6.5787.5%18.8280.2%56.4685.4%0.8984.0%0.2073.2%
+ +![](images/dcf41b908565a6e3095f481eaef324917aaffcf133f28a7100a733e472ab9f76.jpg) +Figure 9. Image editing on face domain. We follow SDEdit[29] to edit facial images equipped with our proposed methods. + +![](images/3418ce1f1b5df6d480120756792f0023bc3e46da50a4583f280595016c4ec28f.jpg) +Figure 10. 3D generation on porcelain domain. We use Dream-Fusion [33] based on SD1.5 for comparison. We present two views of the generated objects, including their mesh edges. * denotes the prompt 'a DSLR photo of'. + +we generate 1,000 images over 10 attribute-related prompts on the facial domain and employ facial predictors to ascertain whether these attributes are accurately represented in + +the generated images. Besides, we sample 200 images from CelebA-HQ dataset to extract canny information as the spatial condition and compute the canny discrepancy. We also report human preference (denoted as 'Pref') by collecting the win rates of generative images across all comparisons. Detailed settings are presented in Appendix C.2. + +Qualitative Comparisons. We conducted a qualitative comparison of our method against previous training techniques, as depicted in Fig 11. Methods like textual inversion, custom diffusion, and LoRA, which adhere to a parameter-efficient form, exhibit noticeably lower fidelity compared to others. Vanilla fine-tuning, in particular, demonstrates a decline in controllability during extended training periods. This results in a diminished capacity for text-guided generation and the emergence of unnatural artifacts, especially when used in conjunction with ControlNet. In contrast, our method excels in both fidelity and control capabilities. + +Quantitative Comparisons. We conduct the evaluation of unconditional generation and report the results in Tab 1. Our approach demonstrates state-of-the-art generation quality, evidenced by a significantly lower FID compared to baseline methods and the highest human preference rating. For text-guided generation, Textual Inversion shows a higher alignment with the text prompts than our method. However, this is achieved at the cost of image quality, owing to its reliance on a very limited set of learnable parameters. This compromise in quality is clearly evident in the comparisons presented in Fig 11. In contrast, in spatial-guided generation, where semantic features are integrated into the UNet, the alignment degree among different methods is similar. Notably, our results not only align well with the control signals but also achieve the highest ratings in human preference assessments. + +# 4.5. Effects of Unconditional Guidance Rectification + +In previous methods, unconditional guidance is predicted by the customized model, leading to inaccurate estimation of probability in sampling. In our framework, the uncon- + +![](images/e28ae807fe0b92db36848e5b0bebac9a1b20b8b1db450fd10bc8a3c6cb905446.jpg) +Figure 11. Comparisons to existing training methods. We evaluate in-domain image generation in unconditional, text-guided, and spatial-guided cases. We fix the random seed to illustrate the controllability for the former two. + +![](images/f7e3db3481398eb08e726687a3ac90282bfb6439a5fd952bff76268094243b45.jpg) +(a) In-domain Generation +Figure 12. Effects of Unconditional Guidance Rectification. (a) Effects in Facial Domain Generation. (b) Unconditional guidance rectification also improves the open-sourced diffusion models. + +![](images/63acaf7fc687c2acf9b3d65ec2c179cd77b40c7d19b1e9ec06dd772b110c4ece.jpg) +(b) Open-Sourced Diffusion Models + +ditional guidance is rectified by straightforwardly predicting by diffusion priors. We compare generation results in Fig 12a. Our method significantly improves image quality, generating fewer artifacts and attaining photorealistic results. Besides, we utilize four widely-used open-sourced models. The comparative results are presented in Fig 12b. Without rectification, the generation quality shows slight degradation, as evidenced by less detailed eyes and less realistic hair textures. The implementation of guidance decoupling substantially enhances generative performance via the rectification of unconditional guidance. + +# 5. Conclusion + +In this research, we address the task of enhancing large-scale diffusion models for in-domain generation using only + +image data. Our goal is to enable these models to generate images that faithfully represent the given domains, achieving high fidelity, diversity, and controllability. Moreover, our approach is designed to be compatible with various control methods, allowing the model to perform a wide range of generative tasks. To tackle this challenge, we propose a guidance-decoupled prior preservation mechanism that separates conditional guidance into two components: domain guidance for alignment and control guidance for maintaining the models controllability. Additionally, we introduce an efficient learning approach by incorporating a null-text diffusion model, ensuring both simplicity and effectiveness. We then outline the generative pipeline of our method to achieve diverse in-domain generation tasks. We conduct experiments across three domains to demonstrate the effectiveness of our proposed method. + +# 6. Acknowledgements + +This work was supported by the Young Scientists Fund of NSFC (Grant No. 62406035). + +# References + +[1] Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H. Bermano. Domain-agnostic tuning-encoder for fast personalization of text-to-image models. pages 1–10, 2023. 2 +[2] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 2 +[3] Pu Cao, Lu Yang, Dongxu Liu, Zhiwei Liu, Shan Li, and Qing Song. Lsap: Rethinking inversion fidelity, perception and editability in gan latent space. arXiv preprint arXiv:2209.12746, 2022. 3 +[4] Pu Cao, Lu Yang, Dongxv Liu, Xiaoya Yang, Tianrui Huang, and Qing Song. What decreases editing capability? domain-specific hybrid refinement for improved gan inversion. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 4240-4249, 2024. 3 +[5] Hong Chen, Yipeng Zhang, Simin Wu, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. Disenbooth: Identity-preserving disentangled tuning for subject-driven text-to-image generation. arXiv preprint arXiv:2305.03374, 2023. 2 +[6] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 100–116. Springer, 2019. 2 +[7] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188-8197, 2020. 5 +[8] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2 +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171-4186, 2019. 2 +[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[11] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. In Proceedings of the IEEE/CVF Inter- + +national Conference on Computer Vision, pages 7430-7440, 2023. 2 +[12] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model. Advances in Neural Information Processing Systems, 35:8882-8895, 2022. 2 +[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 4, 5, 7, 13 +[14] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36:15890-15902, 2023. 3 +[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3, 5, 7 +[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2 +[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 5 +[21] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6007-6017, 2023. 2 +[22] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1931-1941, 2023. 5, 7, 13 +[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22511-22521, 2023. 2 +[24] Zuoxin Li, Fuqiang Zhou, Lu Yang, Xiaojie Li, and Juan Li. Accelerate neural style transfer with super-resolution. Multimedia Tools and Applications, 79:4347-4364, 2020. 2 + +[25] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 5 +[26] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. arXiv preprint arXiv:2303.05125, 2023. 2, 3 +[27] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Customizable image synthesis with multiple subjects. Advances in neural information processing systems, 36:57500-57519, 2023. 2, 3 +[28] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 5, 13 +[29] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 5, 7 +[30] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 2 +[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[32] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, JunYan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11399-11409, 2022. 3 +[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 5, 7 +[34] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 2, 3, 5, 7, 13 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2 +[36] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2 +[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener + +ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[38] Daniel Reich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 3 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2 +[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 3, 13 +[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2 +[43] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2 +[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 2 +[45] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 44(4):2004-2018, 2020. 2 +[46] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. pages 8543-8552, 2024. 2 +[47] James Seale Smith, Yen-Chang Hsu, Lingyu Zhang, Ting Hua, Zsolt Kira, Yilin Shen, and Hongxia Jin. Continual diffusion: Continual customization of text-to-image diffusion with c-lora. arXiv preprint arXiv:2304.06027, 2023. 2 +[48] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 3 +[49] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIG- + +GRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2 +[50] Yujin Wang, Lingen Li, Tianfan Xue, and Jinwei Gu. Reconstruct-and-generate diffusion model for detail-preserving image denoising. arXiv preprint arXiv:2309.10714, 2023. 2 +[51] Zhizhong Wang, Lei Zhao, and Wei Xing. *Style diffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 2 +[52] Chanyue Wu, Dong Wang, Yunpeng Bai, Hanyu Mao, Ying Li, and Qiang Shen. Hsr-diff: Hyperspectral image superresolution via conditional diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2023. 2 +[53] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7378-7387, 2023. 2 +[54] Lu Yang, Qing Song, Zhihui Wang, and Ming Jiang. Parsing r-cnn for instance-level human analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 364-373, 2019. 2 +[55] Lu Yang, Qing Song, Zhihui Wang, Mengjie Hu, Chun Liu, Xueshi Xin, Wenhe Jia, and Songcen Xu. Renovating parsing r-cnn for accurate multiple human parsing. In European Conference on computer vision, pages 421-437. Springer, 2020. +[56] Lu Yang, Qing Song, Xueshi Xin, and Zhiwei Liu. Quality-aware network for face parsing. arXiv preprint arXiv:2106.07368, 2021. +[57] Lu Yang, Zhiwei Liu, Tianfei Zhou, and Qing Song. Part decomposition and refinement network for human parsing. IEEE/CAA Journal of Automatica Sinica, 9(6):1111-1114, 2022. +[58] Lu Yang, Qing Song, Zhihui Wang, Zhiwei Liu, Songcen Xu, and Zhihao Li. Quality-aware network for human parsing. IEEE Transactions on Multimedia, 2022. +[59] Lu Yang, Wenhe Jia, Shan Li, and Qing Song. Deep learning technique for human parsing: A survey and outlook. International Journal of Computer Vision, 132(8):3270-3301, 2024. 2 +[60] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22873-22882, 2023. 2 +[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2 \ No newline at end of file diff --git a/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/images.zip b/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..648db712381320321c1bff1404ab59338a6ea473 --- /dev/null +++ b/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:480ab9399a20d5e9946de4b1be7998d2dc7f1efceb74c46e70274926dacd096f +size 1003215 diff --git a/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/layout.json b/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ee5e1628da2e0384c2dde35f43b5eafd32190050 --- /dev/null +++ b/2025/Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation/layout.json @@ -0,0 +1,7642 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 481, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 481, + 138 + ], + "type": "text", + "content": "Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 477, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 477, + 190 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 477, + 190 + ], + "type": "text", + "content": "Pu Cao† Feng Zhou† Lu Yang* Tianrui Huang Qing Song \nBeijing University of Posts and Telecommunications" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 191, + 462, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 191, + 462, + 204 + ], + "spans": [ + { + "bbox": [ + 148, + 191, + 462, + 204 + ], + "type": "text", + "content": "{caopu, zhoufeng, soeaver, huangtianrui, priv}@bupt.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 55, + 226, + 555, + 428 + ], + "blocks": [ + { + "bbox": [ + 55, + 226, + 555, + 428 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 555, + 428 + ], + "type": "image", + "image_path": "95bb0a65ef097a00e6d25ecbb6219a0ea0cd05ce3fa3094a6361d49897fa9e15.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 437, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 54, + 437, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 555, + 472 + ], + "type": "text", + "content": "Figure 1. Illustration of In-domain Generation. In this work, we empower large-scale pre-trained diffusion models using only image data to perform varied generation tasks within each domain with high fidelity and controllability. We mark the input data in blue and the results generated with the original Stable Diffusion v1.5 model in orange." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 480, + 200, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 480, + 200, + 493 + ], + "spans": [ + { + "bbox": [ + 151, + 480, + 200, + 493 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 506, + 297, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 506, + 297, + 662 + ], + "spans": [ + { + "bbox": [ + 54, + 506, + 297, + 662 + ], + "type": "text", + "content": "In-domain generation aims to perform a variety of tasks within a specific domain, such as unconditional generation, text-to-image, image editing, 3D generation, and more. Early research typically required training specialized generators for each unique task and domain, often relying on fully-labeled data. Motivated by the powerful generative capabilities and broad applications of diffusion models, we are driven to explore leveraging label-free data to empower these models for in-domain generation. Fine-tuning a pre-trained generative model on domain data is an intuitive but challenging way and often requires complex manual hyper-parameter adjustments since the limited diversity of the training data can easily disrupt the model's origi" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 482, + 556, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 556, + 697 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 556, + 697 + ], + "type": "text", + "content": "nal generative capabilities. To address this challenge, we propose a guidance-decoupled prior preservation mechanism to achieve high generative quality and controllability by image-only data, inspired by preserving the pre-trained model from a denoising guidance perspective. We decouple domain-related guidance from the conditional guidance used in classifier-free guidance mechanisms to preserve open-world control guidance and unconditional guidance from the pre-trained model. We further propose an efficient domain knowledge learning technique to train an additional text-free UNet copy to predict domain guidance. Besides, we theoretically illustrate a multi-guidance in-domain generation pipeline for a variety of generative tasks, leveraging multiple guidances from distinct diffusion models and conditions. Extensive experiments demonstrate the superiority of our method in domain-specific synthesis and its compatibility with various diffusion-based control methods and applications." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 673, + 253, + 685 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 673, + 253, + 685 + ], + "spans": [ + { + "bbox": [ + 69, + 673, + 253, + 685 + ], + "type": "text", + "content": "The first two authors contributed to this paper equally." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 685, + 183, + 694 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 685, + 183, + 694 + ], + "spans": [ + { + "bbox": [ + 69, + 685, + 183, + 694 + ], + "type": "text", + "content": "* Corresponding author: Lu Yang." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 694, + 294, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 694, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 694, + 294, + 712 + ], + "type": "text", + "content": "Code is available at https://github.com/PRIV-Creation/In-domain-Generation-Diffusion." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "18358" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 101, + 70, + 304, + 159 + ], + "blocks": [ + { + "bbox": [ + 59, + 86, + 96, + 95 + ], + "lines": [ + { + "bbox": [ + 59, + 86, + 96, + 95 + ], + "spans": [ + { + "bbox": [ + 59, + 86, + 96, + 95 + ], + "type": "text", + "content": "a photo of face" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 121, + 96, + 137 + ], + "lines": [ + { + "bbox": [ + 59, + 121, + 96, + 137 + ], + "spans": [ + { + "bbox": [ + 59, + 121, + 96, + 137 + ], + "type": "text", + "content": "a photo of face wearing hat" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 101, + 70, + 304, + 159 + ], + "lines": [ + { + "bbox": [ + 101, + 70, + 304, + 159 + ], + "spans": [ + { + "bbox": [ + 101, + 70, + 304, + 159 + ], + "type": "image", + "image_path": "6448d8d9d43727a287d2a6293791d0aef64d49879a1b979339d3eb783f51299c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 169, + 296, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 169, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 296, + 226 + ], + "type": "text", + "content": "Figure 2. Challenge of fine-tuning diffusion models on domain data. We show the fine-tuning process of Stable Diffusion v1.5 with a facial image dataset (i.e., FFHQ[20]). The open-world controllability of pre-trained diffusion models is gradually decreased during the fine-tuning process, with domain fidelity improved." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 236, + 136, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 236, + 136, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 136, + 248 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 388 + ], + "type": "text", + "content": "In-domain generation aims to accomplish various tasks within a specific domain, often requiring tailored approaches and labeled data. For instance, pix2pix [19] was developed for image-to-image translation (e.g., converting segmentation or edge images to realistic images), relying on paired images for training. Similarly, InterfaceGAN[45] enables attribute-based image editing but requires attribu-telabeled images to train editing vectors. Such traditional methods[6, 12, 19, 45] depend heavily on specific generative tasks and data domains, which limits their generalization capability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 389, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 533 + ], + "type": "text", + "content": "Thanks to billion-scale image-text datasets [2, 37, 43, 44] and advances in large-scale natural language models [9, 35, 36], recent research on diffusion models has pushed the boundaries of AI-generated content, greatly enhancing both the diversity and controllability of generated outputs [39]. With various control mechanisms and approaches (e.g., ControlNet[61], SDEdit[29]), diffusion models can also perform diverse generative tasks beyond text-to-image synthesis, alleviating the data demands of traditional generative methods. However, the generated results often fail to fully align with specific data domains, as shown in orange in Fig. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": "While fine-tuning a pre-trained model on domain-specific data is an intuitive way to inject domain knowledge, it is challenging to balance fidelity with controllability, as illustrated in Fig. 2. For example, fine-tuning can improve the quality of face generation but may gradually reduce the effect of control prompts like wearing hat. In this work, we first identify this phenomenon caused by conditional guidance and unconditional guidance catastrophic forgetting during fine-tuning. These two guidances are integrated by classifier-free guidance for improved generative capability, which is employed in most text-to-image diffusion models. Then, we propose a guidance-decoupled prior preservation method to address these challenges. We decouple the conditional guidance in CFG into two parts: domain guidance, which is optimized in fine-tuning to enhance fi" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "dely, and control guidance, which is preserved for controllability. With introducing multiple guidances in the synthesis process, the control guidance and unconditional guidance, which we want to preserve during personalization, are directly predicted by diffusion priors, while the domain guidance is predicted by an independent diffusion model. Additionally, we introduce an efficient domain knowledge learning mechanism, designing a null-text Diffusion Model to learn domain knowledge in a concise way." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 180, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 217 + ], + "type": "text", + "content": "We conduct extensive experiments to demonstrate the effects of our proposed task and validate the superiority of our approach. To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 217, + 553, + 384 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 313, + 217, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 252 + ], + "type": "text", + "content": "- We present the task of aligning large-scale diffusion models with specific domains using only image data to perform a variety of generative tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 254, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 553, + 324 + ], + "type": "text", + "content": "- We propose a guidance-decoupled prior preservation mechanism to address guidance forgetting during finetuning, which decouples domain guidance to learn domain knowledge with other guidance preserved. Besides, we propose an efficient learning mechanism to learn domain knowledge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 325, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 553, + 384 + ], + "type": "text", + "content": "- We present the pipeline of our proposed method for in-domain generation tasks and conduct experiments and comparisons across multiple domains and tasks to demonstrate the effectiveness and advancement of our approach." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 399, + 405, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 399, + 405, + 411 + ], + "spans": [ + { + "bbox": [ + 314, + 399, + 405, + 411 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 422, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 590 + ], + "type": "text", + "content": "Text-to-image Generation. As the scale of image-text data grows [2, 37, 43, 44], large-scale diffusion models [10, 17, 31, 39] have demonstrated astonishing generative capabilities, able to produce images based on open-world prompts. Some models, such as Imagen [42] and Stable Diffusion [39], have been explored as diffusion priors, applied in downstream tasks like controlled generation [23, 30, 49, 61] and image editing [8, 15, 21]. For instance, ControlNet [61] proposes using an additional network to receive more spatial controls to guide generation. Furthermore, diffusion models have demonstrated powerful capabilities in many tasks, such as face parsing [54-59], style transfer [24, 51, 60], image manipulation [11, 53] and super-resolution [50, 52]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Fine-tuning Large-scale Diffusion Models. Some existing methods study learning subjects into large-scale diffusion priors and generate variant images guided by different prompts [1, 5, 13, 26, 27, 34, 41, 46, 47]. Some methods fine-tune the part of diffusion models with one or a few images of the specific subject [5, 13, 34, 41]. Textual inversion [13] tunes the text embedding of the subject placeholder, while DreamBooth [41] further updates the weight of UNet [40]. Some methods further study training-free [1, 46], continual customization [47], or multi-subjects" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "text", + "content": "18359" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 162 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 162 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 162 + ], + "type": "image", + "image_path": "914999c871df7759bd2b8e8a8a94cb51c86011d743d022bc74c95b9eba7fbc9e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "type": "text", + "content": "Figure 3. Illustration of Fine-tuning Process. We demonstrate the guidance catastrophic forgetting during fine-tuning process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 80, + 204, + 272, + 323 + ], + "blocks": [ + { + "bbox": [ + 80, + 204, + 272, + 323 + ], + "lines": [ + { + "bbox": [ + 80, + 204, + 272, + 323 + ], + "spans": [ + { + "bbox": [ + 80, + 204, + 272, + 323 + ], + "type": "image", + "image_path": "2ab9bc7c2237b876a4da92e539edc06de6f732d4052dbbe96a36a60609fb5a14.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 296, + 374 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 374 + ], + "type": "text", + "content": "Figure 4. Unconditional Guidance Drift. The unconditional generation results of fine-tuned diffusion models reflect the visual pattern of training datasets, which would cause inaccurate noise estimation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 443 + ], + "type": "text", + "content": "personalization [26, 27] and layout-guided generation [14]. However, these methods concentrate on the consistency of given subjects and training efficiency, tailored for the situation with small-scale image datasets and limited training process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 447, + 296, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 603 + ], + "type": "text", + "content": "Fidelity and Controllability Tradeoff. In the field of generation, the tradeoff between fidelity and controllability has received widespread attention. In GANs, many inversion methods address this problem through regularization [3, 38], latent code discriminator [48], and invertibility prediction [4, 32]. In the personalization task, thanks to the simple customization process and excellent generalization of diffusion priors, this problem is easily solved by regularization [41], parameter efficient tuning [18], orthogonal fine-tuning [34], and so on. However, a longer training process is required for domain knowledge learning and the existing techniques struggle to solve this problem, highlighting an urgent need for a new approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 613, + 111, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 111, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 111, + 624 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 632, + 239, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 239, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 239, + 645 + ], + "type": "text", + "content": "3.1. Guidance Catastrophic Forgetting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 651, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 698 + ], + "type": "text", + "content": "In this section, we delve into the phenomenon of catastrophic forgetting in the training process of diffusion models from the perspective of guidance, allowing us to address this issue at its root." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 701, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 712 + ], + "type": "text", + "content": "Classifier-free Guidance (CFG). To achieve a better trade" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 566, + 178 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 566, + 178 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 566, + 178 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 566, + 178 + ], + "type": "image", + "image_path": "1b0952f14a6b7b9189d0a85b9a75a97994c1146ffb7082c37b4904792e70126b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 182, + 406, + 191 + ], + "lines": [ + { + "bbox": [ + 348, + 182, + 406, + 191 + ], + "spans": [ + { + "bbox": [ + 348, + 182, + 406, + 191 + ], + "type": "text", + "content": "(a) previous fine-tuning" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 182, + 548, + 191 + ], + "lines": [ + { + "bbox": [ + 463, + 182, + 548, + 191 + ], + "spans": [ + { + "bbox": [ + 463, + 182, + 548, + 191 + ], + "type": "text", + "content": "(b) guidance-decoupled generation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 203, + 554, + 269 + ], + "lines": [ + { + "bbox": [ + 313, + 203, + 554, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 554, + 269 + ], + "type": "text", + "content": "Figure 5. Conditional Guidance Decoupling. We compare the guidance estimation between previous customization methods and ours. We decouple the conditional guidance into domain guidance and control guidance while predicting the control guidance and unconditional guidance using the original diffusion model to keep them unchanged." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 290, + 554, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 554, + 326 + ], + "type": "text", + "content": "off between quality and diversity, CFG [16] is utilized to combine the conditional and unconditional score estimates with a hyperparameter " + }, + { + "bbox": [ + 313, + 290, + 554, + 326 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 290, + 554, + 326 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 353, + 335, + 554, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 335, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 353, + 335, + 554, + 350 + ], + "type": "interline_equation", + "content": "\\widetilde {\\epsilon} _ {\\theta} \\left(z _ {t}, c\\right) = (1 + w) \\epsilon_ {\\theta} \\left(z _ {t}, c\\right) - w \\epsilon_ {\\theta} \\left(z _ {t}\\right) \\tag {1}", + "image_path": "a5fcacd321ae1f58ef2720fb4d09496ff72b6ddea8b285d67684c8142b03e911.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 358, + 554, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 405 + ], + "type": "text", + "content": "In large-scale diffusion model training, suppose that an open-world training dataset denotes as " + }, + { + "bbox": [ + 313, + 358, + 554, + 405 + ], + "type": "inline_equation", + "content": "\\{z_i, c_i\\}" + }, + { + "bbox": [ + 313, + 358, + 554, + 405 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 358, + 554, + 405 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 313, + 358, + 554, + 405 + ], + "type": "text", + "content": " is the training image and " + }, + { + "bbox": [ + 313, + 358, + 554, + 405 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 358, + 554, + 405 + ], + "type": "text", + "content": " is the correspondent text, the training objective is to minimize:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 344, + 415, + 553, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 415, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 344, + 415, + 553, + 449 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left(\\underbrace {\\mathcal {L} \\left(p _ {\\theta} (z \\mid c) , p (z \\mid c)\\right)} _ {\\text {c o n d i t i o n a l g u i d a n c e}} + \\underbrace {\\mathcal {L} \\left(p _ {\\theta} (z) , p (z)\\right)} _ {\\text {u n c o n d i t i o n a l g u i d a n c e}}\\right) \\tag {2}", + "image_path": "768762a5d56ca758a763d61215cd4e299dcdd4f85d310679117121afdca9f372.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "text", + "content": "Guidance Catastrophic Forgetting. During domain data fine-tuning with data " + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "inline_equation", + "content": "\\{z_i, y_i\\}" + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "text", + "content": ", these two guidances inevitably deviate, which we illustrated in Fig. 3. Notably, " + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "inline_equation", + "content": "p(z|y)" + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "text", + "content": " represents the probability conditioned on a given domain, which does not imply that text labels of domain data are required. For conditional guidance, the condition " + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "text", + "content": " in fine-tuning phase is too much simpler than the condition in the pre-trained phase " + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "text", + "content": ", where the latter is expected to be control generation. For example, although we leverage only facial images in training, we also desire to generate images controlled by text prompts. For unconditional guidance, we suppose that the pre-trained model could estimate the accurate " + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "inline_equation", + "content": "p_{\\theta_s d}(z_t|\\emptyset)" + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "text", + "content": " since it is trained on open-world data. Hence, fine-tuning on domain data would make " + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "inline_equation", + "content": "p_{\\theta^*}(z_t|\\emptyset)" + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "text", + "content": " inaccuracy, illustrated in Fig. 4. The generative results from fine-tuned models reflect the characteristics of the training datasets." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 671, + 524, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 524, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 524, + 684 + ], + "type": "text", + "content": "3.2. Guidance-Decoupled Prior Preservation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "To address the forgetting issue, we propose a guidance-decoupled prior preservation mechanism. We decouple con" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "text", + "content": "18360" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": "ditional guidance into domain guidance and control guidance, as illustrated in Fig 5(b). The domain guidance steers the denoising process to generate domain-aligned images and the control guidance takes charge of handling openworld controls in generation. For example, to generate 'a man wearing glasses' within the face domain, domain guidance tells how the human face looks, and control guidance ensures the synthetic faces follow 'man' and 'wearing glasses'. After such decoupling, control guidance and unconditional guidance can be predicted by diffusion priors and domain guidance can be learned by an additional model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": "Conditional Guidance Decoupling. Denoting a sets of conditions as " + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "inline_equation", + "content": "C = \\{c_1, \\dots, c_k\\}" + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": " with manually defined intensities " + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "inline_equation", + "content": "w = \\{w_1, \\dots, w_k\\}" + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": ", we can use multiple UN-ets " + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": " to predict conditional and unconditional: " + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_0}(z_t) = \\nabla_z \\log p(z_t)" + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_i}(z_t, c_i) = \\nabla_z \\log p(z_t | c_i)" + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": ". The reverse process in each timestamp is as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 300, + 296, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 300, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 67, + 300, + 296, + 334 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} \\left(z _ {t}, C\\right) = \\epsilon_ {\\theta_ {0}} \\left(z _ {t}\\right) + \\sum_ {i = 1} ^ {K} w _ {i} \\left(\\epsilon_ {\\theta_ {i}} \\left(z _ {t}, c _ {i}\\right) - \\epsilon_ {\\theta_ {0}} \\left(z _ {t}\\right)\\right) \\tag {3}", + "image_path": "4fc7d65f33d8b294604f8aa71866acb7c5daaba0de274e4d82f866cd1b5cd1cf.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 426 + ], + "type": "text", + "content": "Hence, we can decouple the original conditional guidance into two components to improve domain alignment and open-world control, respectively. As shown in Eq. (6), we can use different diffusion models to predict each guidance, allowing us to fine-tune a UNet copy to learn domain knowledge and utilize the pre-trained model to control the generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 428, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 296, + 559 + ], + "type": "text", + "content": "However, Eq. (6) assumes that all conditions are independent. This assumption leads to the independent maximization of the probability of each condition, which can cause confusion between the conditions. This issue is analyzed further in Appendix A. Our analysis reveals that a simple yet effective approach to mitigate this issue is to use more detailed descriptions (e.g., a photo of a face, wearing glasses) rather than overly simplified conditions (e.g., wearing glasses). This approach helps reduce confusion and better aligns the guidance with the intended domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 562, + 296, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 296, + 658 + ], + "type": "text", + "content": "Unconditional Guidance Rectification. Through multi-guidance generation, an intuitive solution to address unconditional guidance drift involves rectifying it using prior diffusion models. We can consider the original pre-trained model as an unconditional guidance predictor, effectively mitigating guidance drift after fine-tuning. This rectification operation can also used in other customized diffusion models, including open-sourced models, as shown in Sec 4.5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 667, + 259, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 667, + 259, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 667, + 259, + 681 + ], + "type": "text", + "content": "3.3. Efficient Domain Knowledge Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Null-text Diffusion Model. The additional fine-tuned diffusion model aims to learn the domain guidance by a do" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "main image dataset. To make it concise, we construct a null-text diffusion model to learn domain guidance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 96, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 555, + 120 + ], + "type": "text", + "content": "In LDMs, cross-attention is used to fuse text feature and image feature with residual connection, which follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 126, + 555, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 126, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 319, + 126, + 555, + 171 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {i m g} ^ {\\prime} = \\underbrace {W (S o f t m a x (\\frac {\\mathcal {Q} _ {i m g} \\mathcal {K} _ {t e x t} ^ {T}}{\\sqrt {d}}) \\times \\mathcal {V} _ {t e x t})} _ {t e x t f e a t u r e i n j e c t i o n} + \\mathcal {F} _ {i m g} (4)", + "image_path": "e4c0a15cc2bd23bcba1c0bf71209c5c21c963112df6a2eee2ab878bb3afbe85a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{img}" + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "text", + "content": " is transformed from input image feature " + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{img}" + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_{text}^{T},\\mathcal{V}_{text}" + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "text", + "content": " is transformed from text feature, " + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "text", + "content": " is the num of channels, and " + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "text", + "content": " is a linear projection. Since textual information is not used in our fine-tuned model, we can convert the former term into a fixed embedding " + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "text", + "content": ", and we have:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 393, + 256, + 555, + 272 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 256, + 555, + 272 + ], + "spans": [ + { + "bbox": [ + 393, + 256, + 555, + 272 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {i m g} ^ {\\prime} = E + \\mathcal {F} _ {i m g} \\tag {5}", + "image_path": "d0bc41987632bf730f852c3d18ba7e6bf02201fc810a19b9a0f1d7b4ca500ddd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 281, + 556, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 556, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 556, + 413 + ], + "type": "text", + "content": "Initializing " + }, + { + "bbox": [ + 313, + 281, + 556, + 413 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 281, + 556, + 413 + ], + "type": "text", + "content": " via Textual Features Optimization. To better inherent generative capability from diffusion priors, accurate textual features represent a good initializing point in training. Therefore, we draw inspiration from Textual Inversion[13] and optimize the text features " + }, + { + "bbox": [ + 313, + 281, + 556, + 413 + ], + "type": "inline_equation", + "content": "F^T" + }, + { + "bbox": [ + 313, + 281, + 556, + 413 + ], + "type": "text", + "content": ", the output of text encoder, and use them as the initial values for " + }, + { + "bbox": [ + 313, + 281, + 556, + 413 + ], + "type": "inline_equation", + "content": "E = W_V(F^T)" + }, + { + "bbox": [ + 313, + 281, + 556, + 413 + ], + "type": "text", + "content": ". Specifically, we first attain " + }, + { + "bbox": [ + 313, + 281, + 556, + 413 + ], + "type": "inline_equation", + "content": "F^T" + }, + { + "bbox": [ + 313, + 281, + 556, + 413 + ], + "type": "text", + "content": " by feeding a photo of into the text encoder and optimize " + }, + { + "bbox": [ + 313, + 281, + 556, + 413 + ], + "type": "inline_equation", + "content": "F^T" + }, + { + "bbox": [ + 313, + 281, + 556, + 413 + ], + "type": "text", + "content": " by denoising loss. This training process converges in just several minutes, providing a good initial value that accelerates the subsequent fine-tuning process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 419, + 543, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 543, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 543, + 431 + ], + "type": "text", + "content": "3.4. In-domain Generation with Multi-Guidance" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": "During domain knowledge learning, we fine-tune a domain diffusion model " + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_d}" + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": " to provide domain guidance and employ the Stable Diffusion " + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\epsilon_{sd}" + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": " to provide preserved guidance, as illustrated in Fig 5. For control guidance, we can also flexibly adapt existing techniques, including Stable Diffusion, ControlNet, or any open-sourced customized models, to achieve a variety of applications. The reverse process suffices:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 541, + 553, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 541, + 553, + 571 + ], + "spans": [ + { + "bbox": [ + 315, + 541, + 553, + 571 + ], + "type": "interline_equation", + "content": "\\underbrace {(1 - w _ {d} - w _ {c}) \\epsilon_ {s d} (z _ {t})} _ {u n c o n d i t i o n a l g u i d a n c e} + \\underbrace {w _ {d} \\epsilon_ {\\theta_ {d}} (z _ {t})} _ {d o m a i n g u i d a n c e} + \\underbrace {w _ {c} \\epsilon_ {\\theta_ {c}} (z _ {t} , c)} _ {c o n t r o l g u i d a n c e}", + "image_path": "07f5b9f96f508a570bffa050233558f7d6746c0cd71a667f089caea3e686ca81.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "content": " is control signal with model " + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_c}" + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "content": ". For instance, to generate images conditioned by a canny image in the animal domain, we utilize the diffusion model fine-tuned on an animal dataset as " + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_d}" + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "content": " and the original diffusion model equipped with canny-to-image ControlNet model as " + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\theta_c" + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "content": ". Experimentally, " + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "inline_equation", + "content": "w_d" + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "content": " presents stability in each generative task, similar to the guidance scale used in widely-used diffusion models (e.g., using " + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "inline_equation", + "content": "w_0 = 7" + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "content": " for Stable Diffusion " + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{sd}" + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "content": "). And we determine the initial value of " + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "inline_equation", + "content": "w_c" + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "inline_equation", + "content": "w_d" + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "inline_equation", + "content": "w_d = \\frac{w_0}{2} \\cdot \\frac{\\|\\epsilon_{sd}\\|}{\\|\\epsilon_{\\theta_d}\\|}" + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "inline_equation", + "content": "w_c = \\frac{w_0}{2} \\cdot \\frac{\\|\\epsilon_{sd}\\|}{\\|\\epsilon_{\\theta_c}\\|}" + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "content": " to ensure " + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "inline_equation", + "content": "w_0 \\cdot \\|\\epsilon_{sd}\\| = w_d \\|\\epsilon_{\\theta_d}\\| + w_c \\|\\epsilon_{\\theta_c}\\|" + }, + { + "bbox": [ + 313, + 578, + 555, + 715 + ], + "type": "text", + "content": ". We can also employ" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "text", + "content": "18361" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "this generation process in other diffusion-based generative tasks, like image editing and 3D generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 138, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 138, + 123 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 138, + 123 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 129, + 183, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 183, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 183, + 143 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 151, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 296, + 270 + ], + "type": "text", + "content": "Datasets. In this study, we conduct experiments on three domains: face (i.e., FFHQ [20]), animal (i.e., AFHQv2 [7]), and porcelain (i.e. a collected dataset), chosen for their diversity and high-resolution (larger than " + }, + { + "bbox": [ + 55, + 151, + 296, + 270 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 151, + 296, + 270 + ], + "type": "text", + "content": "). FFHQ is a high-quality human face dataset, containing 70,000 images. Images come from Flickr and are aligned into " + }, + { + "bbox": [ + 55, + 151, + 296, + 270 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 55, + 151, + 296, + 270 + ], + "type": "text", + "content": ". We further resize images into " + }, + { + "bbox": [ + 55, + 151, + 296, + 270 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 151, + 296, + 270 + ], + "type": "text", + "content": " without crop. AFHQv2 is an animal face dataset with a resolution of " + }, + { + "bbox": [ + 55, + 151, + 296, + 270 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 151, + 296, + 270 + ], + "type": "text", + "content": ", containing 15803 images. The porcelain dataset contains 5,000 images, collected from the internet." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 274, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 453 + ], + "type": "text", + "content": "Settings. We train the null-text domain diffusion model based on Stable Diffusion v1.5 (SD1.5) at the resolution of 512. For embedding learning phase, 3,200 images are used to attain initial embeddings with batch size of 32 and 100 steps for all domains. For all experiments, we maintain a constant learning rate of 1e-5 with the optimizer of AdamW and a batch size of 32. These experiments are conducted on 8 NVIDIA A100 GPUs, with each running until convergence. All controlling models, including ControlNet and open-sourced customized diffusion models, are borrowed from their official releases based on the corresponding version of Stable Diffusion without any training. We employ the DPM-Solver++ scheduler [28] with 100 steps across all models. For more experimental settings, please refer to Appendix C." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 463, + 264, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 264, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 264, + 475 + ], + "type": "text", + "content": "4.2. Results of In-domain Image Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 482, + 297, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 297, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 297, + 661 + ], + "type": "text", + "content": "We present the unconditional generation results in Fig. 6 to illustrate the degree of domain alignment. As shown, our results closely match the distribution of the domain datasets, highlighting the effectiveness of our approach in aligning the model outputs with domain data. For conditional generation, we showcase the qualitative results in Fig. 7. We conduct experiments on both text-conditioned and spatial-conditioned generation, utilizing ControlNet models for the latter task. The results clearly demonstrate that our method not only achieves high fidelity with respect to the target domains but also retains the generative controllability seen in pre-trained models, ensuring that the model can be guided by text prompts or spatial conditions as intended. Quantitative results and comparisons can be found in Tab. 1 and Sec. 4.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 670, + 293, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 293, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 293, + 683 + ], + "type": "text", + "content": "4.3. Results of Other In-domain Generation Tasks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Apart from the basic generative tasks, our proposed method also benefits other in-domain generation tasks. In this part," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "we further demonstrate the effectiveness of our method in a wider range of tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 99, + 555, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 99, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 555, + 232 + ], + "type": "text", + "content": "Image Generation under Complex Conditions. As discussed in Sec. 3.2, multiple guidances can be leveraged to control the generation process. In this section, we present the generation results under complex conditions in Fig. 8. Despite the introduction of complex conditions, the generated images maintain high fidelity to the target domain, demonstrating both the robustness and quality of our method. Additionally, our generative framework allows us to leverage additional open-source models for controlling the generation process, such as using DreamShaper to control the style." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 234, + 556, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 556, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 556, + 319 + ], + "type": "text", + "content": "Image Editing. In Fig. 9, we implement SDEdit[29] on the face domain, where the input image is first encoded into latent space and added noise and then we use a similar generation pipeline presented in Sec. 3.4. Detailed settings can be found in Appendix C.3.1. Our results are more aligned with the distribution of real human face images, demonstrating better visual quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 320, + 556, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 556, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 556, + 477 + ], + "type": "text", + "content": "Text-to-3D Generation. In Fig. 10, we leverage our model on porcelain domain for text-to-3d porcelain generation and compare it with classic text-to-3d model [33] on SD1.5. Specifically, we begin by generating porcelain images based on the provided prompt, which are then converted into 3D models using an off-the-shelf image-to-3D method [25]. Subsequently, we fine-tune the resulting 3D models using our porcelain model, guided by the SDS [33] loss. Lastly, we directly convert the NeRF models into meshes. Detailed settings and additional ablation studies on 3D generation can be found in Appendix C.3.2. Our results demonstrate superiority in color richness, fine detail, and alignment with the dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 485, + 524, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 524, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 524, + 498 + ], + "type": "text", + "content": "4.4. Comparisons to Fine-tuning Techniques" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 504, + 554, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 554, + 576 + ], + "type": "text", + "content": "To validate the superiority of our proposed method, we compare our method with several fine-tuning baselines, including vanilla fine-tuning, parameter-efficient fine-tuning, and some training techniques borrowed from other tasks. For detailed experimental settings, please refer to AppendixC.2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 578, + 556, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 556, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 556, + 686 + ], + "type": "text", + "content": "Baselines. To establish comprehensive baselines, we conduct experiments on several training techniques, including vanilla fine-tuning, low-rank fine-tuning [18], textual inversion [13], custom diffusion [22], and OFT mechanism [34]. To train diffusion models without text annotation, we adopt prompt templates, utilizing formats such as a photo of [V] . These additional tokens [V] are omitted in compared methods where text embedding is not trained." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Evaluation Setting. For unconditional generation, we calculate FID on 50,000 images. For text-guided generation," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "18362" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 69, + 555, + 277 + ], + "blocks": [ + { + "bbox": [ + 65, + 69, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 65, + 69, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 65, + 69, + 555, + 277 + ], + "type": "image", + "image_path": "8a7cd6ef72f1975646f225aa315add948730f28205ad6c5a124a5aacc2bea4f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "type": "text", + "content": "Figure 6. Visualization of Domain Alignment. Our results are more aligned to given domain dataset. For SD1.5 and SDXL 1.0, we generate images by text a photo of ." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 326, + 555, + 502 + ], + "blocks": [ + { + "bbox": [ + 61, + 326, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 61, + 326, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 61, + 326, + 555, + 502 + ], + "type": "image", + "image_path": "c9b89f2a8e62df86af02530665c825501ec606754b8dbf89bd237493e6980ffd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 519, + 555, + 553 + ], + "lines": [ + { + "bbox": [ + 55, + 519, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 555, + 553 + ], + "type": "text", + "content": "Figure 7. Results of conditional in-domain image generation. We demonstrate text-conditioned and spatial-conditioned generation results to verify that our proposed methods can effectively preserve the control capability from the pre-trained model. For spatial-conditioned generation, we leverage the off-the-shelf ControlNet model without any extra training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 104, + 563, + 504, + 661 + ], + "blocks": [ + { + "bbox": [ + 104, + 563, + 504, + 661 + ], + "lines": [ + { + "bbox": [ + 104, + 563, + 504, + 661 + ], + "spans": [ + { + "bbox": [ + 104, + 563, + 504, + 661 + ], + "type": "image", + "image_path": "a1945ed6c23c3f6afe7308363bbd95866d935b0cefd3e51d4828aa607049e3ce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 670, + 555, + 693 + ], + "lines": [ + { + "bbox": [ + 55, + 670, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 555, + 693 + ], + "type": "text", + "content": "Figure 8. Generative Results under Complex Conditions. We employ multiple conditions and models to demonstrate the performance of our proposed method in achieving more detailed control over the generation process." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "text", + "content": "18363" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 113, + 555, + 220 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Quantitative results of unconditional generation We evaluate FID and human preference on two datasets. Our method achieves excellent generation results among baselines." + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": ": reproduced results in " + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": " resolution. Bold: the best result. Underline: the second best result." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 113, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 113, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 113, + 555, + 220 + ], + "type": "table", + "html": "
Method#param.Unconditional GenerationConditional Generation
FaceAnimalPorcelainText-GuidedSpatial-Guided
FID↓Pref.↑FID↓Pref.↑FID↓Pref.↑Align.↑Pref.↑Align.↓Pref.↑
Textual Inversion [13]< 1K70.620.3%44.752.5%112.761.7%0.970.2%0.221.5%
Custom Diffusion [22]44M40.981.6%49.981.8%98.770.6%0.741.4%0.201.7%
OFT [34]23M27.502.9%38.741.4%68.841.2%0.773.2%0.219.6%
Fine-tuning859M12.376.4%23.9212.4%62.7610.8%0.244.5%0.224.9%
Fine-tuning + LoRA[18]0.19M23.761.3%24.191.7%73.460.3%0.766.7%0.229.1%
Ours810M6.5787.5%18.8280.2%56.4685.4%0.8984.0%0.2073.2%
", + "image_path": "1cc213ab33f467fadcaa0904f376fbe2366ea45e38427d3fc83a972c9ad89465.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 69, + 225, + 296, + 356 + ], + "blocks": [ + { + "bbox": [ + 69, + 225, + 296, + 356 + ], + "lines": [ + { + "bbox": [ + 69, + 225, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 69, + 225, + 296, + 356 + ], + "type": "image", + "image_path": "dcf41b908565a6e3095f481eaef324917aaffcf133f28a7100a733e472ab9f76.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 365, + 295, + 388 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 388 + ], + "type": "text", + "content": "Figure 9. Image editing on face domain. We follow SDEdit[29] to edit facial images equipped with our proposed methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 69, + 399, + 296, + 612 + ], + "blocks": [ + { + "bbox": [ + 69, + 399, + 296, + 612 + ], + "lines": [ + { + "bbox": [ + 69, + 399, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 69, + 399, + 296, + 612 + ], + "type": "image", + "image_path": "3418ce1f1b5df6d480120756792f0023bc3e46da50a4583f280595016c4ec28f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 622, + 295, + 667 + ], + "lines": [ + { + "bbox": [ + 55, + 622, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 295, + 667 + ], + "type": "text", + "content": "Figure 10. 3D generation on porcelain domain. We use Dream-Fusion [33] based on SD1.5 for comparison. We present two views of the generated objects, including their mesh edges. * denotes the prompt 'a DSLR photo of'." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "we generate 1,000 images over 10 attribute-related prompts on the facial domain and employ facial predictors to ascertain whether these attributes are accurately represented in" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "content": "the generated images. Besides, we sample 200 images from CelebA-HQ dataset to extract canny information as the spatial condition and compute the canny discrepancy. We also report human preference (denoted as 'Pref') by collecting the win rates of generative images across all comparisons. Detailed settings are presented in Appendix C.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 303, + 556, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 556, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 556, + 446 + ], + "type": "text", + "content": "Qualitative Comparisons. We conducted a qualitative comparison of our method against previous training techniques, as depicted in Fig 11. Methods like textual inversion, custom diffusion, and LoRA, which adhere to a parameter-efficient form, exhibit noticeably lower fidelity compared to others. Vanilla fine-tuning, in particular, demonstrates a decline in controllability during extended training periods. This results in a diminished capacity for text-guided generation and the emergence of unnatural artifacts, especially when used in conjunction with ControlNet. In contrast, our method excels in both fidelity and control capabilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 449, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 641 + ], + "type": "text", + "content": "Quantitative Comparisons. We conduct the evaluation of unconditional generation and report the results in Tab 1. Our approach demonstrates state-of-the-art generation quality, evidenced by a significantly lower FID compared to baseline methods and the highest human preference rating. For text-guided generation, Textual Inversion shows a higher alignment with the text prompts than our method. However, this is achieved at the cost of image quality, owing to its reliance on a very limited set of learnable parameters. This compromise in quality is clearly evident in the comparisons presented in Fig 11. In contrast, in spatial-guided generation, where semantic features are integrated into the UNet, the alignment degree among different methods is similar. Notably, our results not only align well with the control signals but also achieve the highest ratings in human preference assessments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 647, + 555, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 555, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 555, + 671 + ], + "type": "text", + "content": "4.5. Effects of Unconditional Guidance Rectification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "In previous methods, unconditional guidance is predicted by the customized model, leading to inaccurate estimation of probability in sampling. In our framework, the uncon-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "18364" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 555, + 261 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 555, + 261 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 555, + 261 + ], + "type": "image", + "image_path": "e28ae807fe0b92db36848e5b0bebac9a1b20b8b1db450fd10bc8a3c6cb905446.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 268, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 268, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 555, + 291 + ], + "type": "text", + "content": "Figure 11. Comparisons to existing training methods. We evaluate in-domain image generation in unconditional, text-guided, and spatial-guided cases. We fix the random seed to illustrate the controllability for the former two." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 66, + 300, + 216, + 460 + ], + "blocks": [ + { + "bbox": [ + 66, + 300, + 216, + 460 + ], + "lines": [ + { + "bbox": [ + 66, + 300, + 216, + 460 + ], + "spans": [ + { + "bbox": [ + 66, + 300, + 216, + 460 + ], + "type": "image", + "image_path": "f7e3db3481398eb08e726687a3ac90282bfb6439a5fd952bff76268094243b45.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 98, + 463, + 182, + 472 + ], + "lines": [ + { + "bbox": [ + 98, + 463, + 182, + 472 + ], + "spans": [ + { + "bbox": [ + 98, + 463, + 182, + 472 + ], + "type": "text", + "content": "(a) In-domain Generation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 482, + 555, + 503 + ], + "lines": [ + { + "bbox": [ + 55, + 482, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 555, + 503 + ], + "type": "text", + "content": "Figure 12. Effects of Unconditional Guidance Rectification. (a) Effects in Facial Domain Generation. (b) Unconditional guidance rectification also improves the open-sourced diffusion models." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 230, + 296, + 545, + 460 + ], + "blocks": [ + { + "bbox": [ + 230, + 296, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 230, + 296, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 230, + 296, + 545, + 460 + ], + "type": "image", + "image_path": "63acaf7fc687c2acf9b3d65ec2c179cd77b40c7d19b1e9ec06dd772b110c4ece.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 462, + 444, + 472 + ], + "lines": [ + { + "bbox": [ + 328, + 462, + 444, + 472 + ], + "spans": [ + { + "bbox": [ + 328, + 462, + 444, + 472 + ], + "type": "text", + "content": "(b) Open-Sourced Diffusion Models" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "text", + "content": "ditional guidance is rectified by straightforwardly predicting by diffusion priors. We compare generation results in Fig 12a. Our method significantly improves image quality, generating fewer artifacts and attaining photorealistic results. Besides, we utilize four widely-used open-sourced models. The comparative results are presented in Fig 12b. Without rectification, the generation quality shows slight degradation, as evidenced by less detailed eyes and less realistic hair textures. The implementation of guidance decoupling substantially enhances generative performance via the rectification of unconditional guidance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 667, + 128, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 667, + 128, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 667, + 128, + 680 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "In this research, we address the task of enhancing large-scale diffusion models for in-domain generation using only" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 520, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 712 + ], + "type": "text", + "content": "image data. Our goal is to enable these models to generate images that faithfully represent the given domains, achieving high fidelity, diversity, and controllability. Moreover, our approach is designed to be compatible with various control methods, allowing the model to perform a wide range of generative tasks. To tackle this challenge, we propose a guidance-decoupled prior preservation mechanism that separates conditional guidance into two components: domain guidance for alignment and control guidance for maintaining the models controllability. Additionally, we introduce an efficient learning approach by incorporating a null-text diffusion model, ensuring both simplicity and effectiveness. We then outline the generative pipeline of our method to achieve diverse in-domain generation tasks. We conduct experiments across three domains to demonstrate the effectiveness of our proposed method." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "text", + "content": "18365" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 116 + ], + "type": "text", + "content": "This work was supported by the Young Scientists Fund of NSFC (Grant No. 62406035)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 148, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 62, + 148, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 296, + 192 + ], + "type": "text", + "content": "[1] Moab Arar, Rinon Gal, Yuval Atzmon, Gal Chechik, Daniel Cohen-Or, Ariel Shamir, and Amit H. Bermano. Domain-agnostic tuning-encoder for fast personalization of text-to-image models. pages 1–10, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 194, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 237 + ], + "type": "text", + "content": "[2] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 296, + 282 + ], + "type": "text", + "content": "[3] Pu Cao, Lu Yang, Dongxu Liu, Zhiwei Liu, Shan Li, and Qing Song. Lsap: Rethinking inversion fidelity, perception and editability in gan latent space. arXiv preprint arXiv:2209.12746, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 294, + 340 + ], + "type": "text", + "content": "[4] Pu Cao, Lu Yang, Dongxv Liu, Xiaoya Yang, Tianrui Huang, and Qing Song. What decreases editing capability? domain-specific hybrid refinement for improved gan inversion. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 4240-4249, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 342, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 294, + 396 + ], + "type": "text", + "content": "[5] Hong Chen, Yipeng Zhang, Simin Wu, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. Disenbooth: Identity-preserving disentangled tuning for subject-driven text-to-image generation. arXiv preprint arXiv:2305.03374, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 398, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 398, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 398, + 294, + 475 + ], + "type": "text", + "content": "[6] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 100–116. Springer, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 477, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 294, + 520 + ], + "type": "text", + "content": "[7] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188-8197, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 522, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 294, + 565 + ], + "type": "text", + "content": "[8] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 567, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 294, + 643 + ], + "type": "text", + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171-4186, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "text", + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[11] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. In Proceedings of the IEEE/CVF Inter-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "national Conference on Computer Vision, pages 7430-7440, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "text", + "content": "[12] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model. Advances in Neural Information Processing Systems, 35:8882-8895, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 196 + ], + "type": "text", + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 4, 5, 7, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 198, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 262 + ], + "type": "text", + "content": "[14] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36:15890-15902, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 331 + ], + "type": "text", + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 367, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 367, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 367, + 553, + 409 + ], + "type": "text", + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "text", + "content": "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 468, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 511 + ], + "type": "text", + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 514, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 567 + ], + "type": "text", + "content": "[21] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6007-6017, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 569, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 623 + ], + "type": "text", + "content": "[22] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1931-1941, 2023. 5, 7, 13" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22511-22521, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[24] Zuoxin Li, Fuqiang Zhou, Lu Yang, Xiaojie Li, and Juan Li. Accelerate neural style transfer with super-resolution. Multimedia Tools and Applications, 79:4347-4364, 2020. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "18366" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[25] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "text", + "content": "[26] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. arXiv preprint arXiv:2303.05125, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "text", + "content": "[27] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Customizable image synthesis with multiple subjects. Advances in neural information processing systems, 36:57500-57519, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "text", + "content": "[28] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 5, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "text", + "content": "[29] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "text", + "content": "[30] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "type": "text", + "content": "[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 481 + ], + "type": "text", + "content": "[32] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, JunYan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11399-11409, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "type": "text", + "content": "[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "text", + "content": "[34] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 2, 3, 5, 7, 13" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 635 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "text", + "content": "[36] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[38] Daniel Reich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 196 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 263 + ], + "type": "text", + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "type": "text", + "content": "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 3, 13" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 397 + ], + "type": "text", + "content": "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[43] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 455, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 520 + ], + "type": "text", + "content": "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 522, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 565 + ], + "type": "text", + "content": "[45] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 44(4):2004-2018, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "text", + "content": "[46] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. pages 8543-8552, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "text", + "content": "[47] James Seale Smith, Yen-Chang Hsu, Lingyu Zhang, Ting Hua, Zsolt Kira, Yilin Shen, and Hongxia Jin. Continual diffusion: Continual customization of text-to-image diffusion with c-lora. arXiv preprint arXiv:2304.06027, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[48] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[49] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIG-" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "18367" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 620 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "GRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "text", + "content": "[50] Yujin Wang, Lingen Li, Tianfan Xue, and Jinwei Gu. Reconstruct-and-generate diffusion model for detail-preserving image denoising. arXiv preprint arXiv:2309.10714, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[51] Zhizhong Wang, Lei Zhao, and Wei Xing. *Style diffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "text", + "content": "[52] Chanyue Wu, Dong Wang, Yunpeng Bai, Hanyu Mao, Ying Li, and Qiang Shen. Hsr-diff: Hyperspectral image superresolution via conditional diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "text", + "content": "[53] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7378-7387, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 329 + ], + "type": "text", + "content": "[54] Lu Yang, Qing Song, Zhihui Wang, and Ming Jiang. Parsing r-cnn for instance-level human analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 364-373, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "text", + "content": "[55] Lu Yang, Qing Song, Zhihui Wang, Mengjie Hu, Chun Liu, Xueshi Xin, Wenhe Jia, and Songcen Xu. Renovating parsing r-cnn for accurate multiple human parsing. In European Conference on computer vision, pages 421-437. Springer, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 418 + ], + "type": "text", + "content": "[56] Lu Yang, Qing Song, Xueshi Xin, and Zhiwei Liu. Quality-aware network for face parsing. arXiv preprint arXiv:2106.07368, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 463 + ], + "type": "text", + "content": "[57] Lu Yang, Zhiwei Liu, Tianfei Zhou, and Qing Song. Part decomposition and refinement network for human parsing. IEEE/CAA Journal of Automatica Sinica, 9(6):1111-1114, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 498 + ], + "type": "text", + "content": "[58] Lu Yang, Qing Song, Zhihui Wang, Zhiwei Liu, Songcen Xu, and Zhihao Li. Quality-aware network for human parsing. IEEE Transactions on Multimedia, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 499, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 499, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 499, + 294, + 541 + ], + "type": "text", + "content": "[59] Lu Yang, Wenhe Jia, Shan Li, and Qing Song. Deep learning technique for human parsing: A survey and outlook. International Journal of Computer Vision, 132(8):3270-3301, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 544, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 588 + ], + "type": "text", + "content": "[60] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22873-22882, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 620 + ], + "type": "text", + "content": "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "text", + "content": "18368" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/f227006c-210b-4aac-822f-93d2a98acc36_content_list.json b/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/f227006c-210b-4aac-822f-93d2a98acc36_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b5eb21409d003f4b725f11f0b786ec0738868cd5 --- /dev/null +++ b/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/f227006c-210b-4aac-822f-93d2a98acc36_content_list.json @@ -0,0 +1,1362 @@ +[ + { + "type": "text", + "text": "Imagine and Seek: Improving Composed Image Retrieval with an Imagined Proxy", + "text_level": 1, + "bbox": [ + 117, + 128, + 880, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "You Li Fan Ma Yi Yang", + "bbox": [ + 369, + 202, + 625, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ReLER, CCAI, Zhejiang University, Zhejiang, China", + "bbox": [ + 287, + 220, + 709, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "† Corresponding author", + "bbox": [ + 426, + 239, + 570, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{uli2000, mafan, yangyics}@zju.edu.cn", + "bbox": [ + 334, + 258, + 661, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Zero-shot Composed Image Retrieval (ZSCIR) requires retrieving images that match the query image and the relative captions. Current methods focus on projecting the query image into the text feature space, subsequently combining them with features of query texts for retrieval. However, retrieving images only with the text features cannot guarantee detailed alignment due to the natural gap between images and text. In this paper, we introduce Imagined Proxy for CIR (IP-CIR), a training-free method that creates a proxy image aligned with the query image and text description, enhancing query representation in the retrieval process. We first leverage the large language model's generalization capability to generate an image layout, and then apply both the query text and image for conditional generation. The robust query features are enhanced by merging the proxy image, query image, and text semantic perturbation. Our newly proposed balancing metric integrates text-based and proxy retrieval similarities, allowing for more accurate retrieval of the target image while incorporating image-side information into the process. Experiments on three public datasets demonstrate that our method significantly improves retrieval performances. We achieve state-of-the-art (SOTA) results on the CIRR dataset with a Recall@K of 70.07 at $K = 10$ . Additionally, we achieved an improvement in Recall@10 on the FashionIQ dataset, rising from 45.11 to 45.74, and improved the baseline performance in CIRCO with a mAPK@10 score, increasing from 32.24 to 34.26.", + "bbox": [ + 89, + 342, + 483, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 782, + 220, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Imagination is the wellspring of human creativity. When you see a picture of an adorable cat, you might envision it playing with other cats, wearing cool sunglasses and a cape, or floating in a space capsule. This ability to imagine and pursue beautiful visions drives continuous advancement in fields like image generation [2, 19, 26, 30, 33, 35, 47, 50,", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb2a030492fa13ecac691f8e86efbdd6e4da0aa8b7443c5315e45567dd241dfe.jpg", + "image_caption": [ + "Figure 1. Comparison of composed image retrieval between text-only retrieval (LDRE in Fig) and our methods. Traditional methods perform retrieval only in the text space, where CLIP text features may overlook some important semantic information. In contrast, our approach generates imagined proxy features, providing additional information that is often overlooked in text-only retrieval, thereby improving retrieval accuracy." + ], + "image_footnote": [], + "bbox": [ + 517, + 309, + 905, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "51] and image retrieval [5, 6, 16, 25, 28, 34, 37, 41, 43, 45].", + "bbox": [ + 511, + 575, + 903, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This rich imagination gains new significance in Composed Image Retrieval (CIR) [28], where users express their imagination of the query image through text, searching for content that aligns with their preferences. Specifically, the CIR task requires users to retrieve target images from a database that match both the content of a query image and relative caption. With text modality to flexibly control retrieval targets, CIR enables the search process akin to the user's specific imagination after viewing the query image. This multi-modal, imagination-like retrieval task greatly expands the applications of image retrieval, with wide-ranging uses in fields such as visual search [29], object localization [18, 21], and re-identification [39, 46].", + "bbox": [ + 509, + 595, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional solutions for CIR [3, 4, 10-15] are constrained by the scarcity of training data (e.g., 46.6k for FashionIQ [41] dataset and 28.8k for CIRR [28] dataset), as collecting suitable triplets (query image $x_{q}$ , relevant text $x_{c}$ , and target image $x_{t}$ ) from the web incurs high labor costs. Therefore, we focus on the ZS-CIR task, aiming to bypass the need for training on triplet datasets and achieve gen", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3984", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eralized CIR retrieval capabilities. As shown in Fig.1 (a), traditional ZS-CIR methods typically employ lightweight projection models to map CLIP's [32] image features into the text features [6] for retrieval. The transformation inevitably loses certain image information, and the coarse, easily confusable text features fall short in fine-grained, complex imagination scenes. Although recent approaches attempt to leverage large language models (LLMs) [1, 31] to generate various descriptions of target images [20, 43, 45], they overlook the potential for direct imagination on the image side. As shown in Fig.1 (a), the text features did not effectively capture the attribute 'golden', resulting in suboptimal retrieval results.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The rapid advancement of controllable generative models [8, 9, 24, 38, 42, 49, 52] has empowered users with the freedom to create images that align with their imagined visions. Generating images that match both the content of query images and relative captions aligns closely with the goals of composed retrieval tasks, and the imagined images (we called Retrieval Proxy) could provide additional information—such as style, instance attributes, and spatial relationships—that is often overlooked in text-based retrieval with complex captions. This raises an intriguing question: How could we harness the imaginative power of generative models to improve retrieval performance?", + "bbox": [ + 89, + 287, + 482, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We thus propose IP-CIR (Imagined Proxy for CIR), a training-free, plug-and-play method that harnesses the power of imagination for any retrieval method. Firstly, by utilizing LLMs to understand both the image content and the relative captions, we generate a suitable image layout. Using controllable generation methods, we create high-quality, fine-grained proxy images. Next, we compose the proxy images, query images as well as semantic perturbation into robust proxy features that are more suitable for retrieval. Finally, we propose a balance metric to combine the proxy and text retrieval similarities, allowing the integration of our method with any retrieval methods.", + "bbox": [ + 89, + 470, + 482, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct experiments on three public datasets: the CIRR [28], CIRCO [5], and FashionIQ [41]. We combine our method with several baselines, the observed performance improvements confirm that our method supplements the retrieval process with valuable information and could enhance the performance of any other retrieval methods.", + "bbox": [ + 89, + 652, + 482, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 109, + 744, + 413, + 758 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We propose using generative models' imagination capabilities to create proxy images that align with the content of the query image and relative captions to improve retrieval performance.", + "2) We introduce a proxy image generation framework. Using the LLM-generated object layout and the query image, we create imagined proxies through MIGC++.", + "3) We propose integrating proxy images with the query image and semantic perturbation into more robust proxy fea" + ], + "bbox": [ + 83, + 763, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tures and introducing a balance metric to combine the proxy and baseline similarities.", + "bbox": [ + 526, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4) We conducted experiments on three public CIR datasets. The improvement over baseline results, as well as achieving SOTA results demonstrates the potential of constructing imagined proxies in retrieval.", + "bbox": [ + 506, + 121, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 196, + 651, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Zero-shot Composed Image Retrieval", + "text_level": 1, + "bbox": [ + 511, + 222, + 836, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early work on ZS-CIR, such as Pic2word [34], SEALRE [6], attempted to project the original image into the text space using lightweight projection models. This was often done within fixed templates (e.g., \"a photo of \\([\\\\) ]$ with cond\"), the \\([\\\\( ]\\) symbol is a textual notation used to represent query image information, and \"cond\" represents the relative captions. LinCIR [16] further defined keywords as adjectives and nouns in the text, thus overcoming the limitations of fixed templates and training more powerful projection models. Besides, CompoDiff[17] trains a new diffusion model that performs diffusion on the CLIP features of the image, introduces relative caption features, and applies attention operations with intermediate image features, ultimately obtaining the target features for retrieval. In addition, some methods have embraced LLMs [1, 31]. CIREVL [20] proposed generating captions for reference images and then using LLMs to reassemble them into target texts modified by relative captions for retrieval. LDRE [43] took this further by using LLMs to construct various edited captions and combining them based on similarity, addressing the issue of fuzzy retrieval. While these methods improved retrieval accuracy, they all overlooked the potential of directly imagining proxy images to provide additional information.", + "bbox": [ + 511, + 244, + 906, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Controllable Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 511, + 619, + 849, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, with the development of diffusion models [19, 35], high-quality images that meet specific requirements can now be generated. Using mainstream generators such as Stable Diffusion [33], users can generate images aligned with text descriptions. Although these foundational generators offer a certain level of control, their performance in fine-grained attribute and orientation control is limited, primarily due to the attribute ambiguity in the CLIP features of the text [48]. As a result, some efforts have focused on leveraging layout information to enhance control. For example, GLIGEN [24] introduces Gated-Self-Attention to combine positional feature information, thereby improving control over positioning. InstanceDiffusion [38] further improves attribute control on this foundation. Meanwhile, MIGC [48] adopts a divide-and-conquer approach, breaking down complex instance generation into multiple single-instance generations and incorporating Enhancement Atten", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3985", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fd55235f2763169362301aaa30e2eb484755851213712e0754860073e8ee8be9.jpg", + "image_caption": [ + "(a) Imagined Retrieval Proxy Generation" + ], + "image_footnote": [], + "bbox": [ + 107, + 114, + 509, + 239 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/99f2b1d15938947dd08754e4771b0cb8b3ebb96761a6e7c1d2ef1f0a8a7dd6c9.jpg", + "image_caption": [ + "(b) Balanced Proxy Retrieval", + "Figure 2. Overview of our method. (a) represents our imagined proxy generation process. We use LLM to analyze the BLIP2-generated query image captions and the relative captions and infer the proxy layout. We then use the controllable generator to imagine the proxy images. (b) represents our process of constructing a robust proxy feature, and balancing the text and proxy similarities. We integrate proxy features, query image features as well as semantic perturbations into a robust proxy feature, and propose a balance metric for retrieval." + ], + "image_footnote": [], + "bbox": [ + 509, + 112, + 901, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion and Layout Attention to improve both positional and attribute control. MIGC++ [49] further integrates ELITE [40] to transform images into CLIP-based text features and introduces an additional EA layer to process information from the reference image modality, thereby incorporating image modality control in the generation process. In this paper, we aim to utilize the fine-grained control capabilities of MIGC to display certain attributes and other information that are difficult to clearly express in text at a fine-grained level within the image itself. This provides supplementary information during the retrieval process.", + "bbox": [ + 88, + 342, + 483, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 523, + 181, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will introduce our IP-CIR, which attempts to imagine an appropriate proxy image within the image feature space, balance, and combine the retrieval results from the image side with those from the text side. Specifically, in section 3.1, we introduce the overview of our method, in section 3.2, we will describe how to construct an imagined retrieval proxy for each query. In section 3.3, we will further explain how to organize and integrate the information from the imagined retrieval proxy and present how to combine the retrieval information from both the image and text feature spaces.", + "bbox": [ + 89, + 549, + 483, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 726, + 200, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall structure of our framework is shown in Fig.2. First, we generate the imagined proxy image for retrieval. The proxy image needs to align with the content of the query image and be as consistent as possible with the description in the relative captions. Therefore, we first generate a reasonable object layout and then use a controllable generation model to imagine this layout as a concrete image. Specifically, we use BLIP2[23] to generate a set of captions for each query. Combine with the relative captions, we instruct the LLM to reason out a suitable", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "spatial layout of the target image, so that, with as much accuracy as possible in terms of scene, instance attributes, and quantity with relative caption, we can imagine a proxy image of the retrieval target to provide additional information for the retrieval process. In Fig.2 (b), we compose the imagined proxy with query image and text features, forming a robust retrieval proxy. The similarity of robust proxy is further balanced with other baselines' similarity, improving the retrieval accuracy.", + "bbox": [ + 511, + 342, + 906, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Imagined Retrieval Proxy Generation", + "text_level": 1, + "bbox": [ + 511, + 494, + 839, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text retrieval features may struggle to precisely capture important semantics in the relevant text and the query image, especially when the caption is complex. Generating an image that meets the retrieval requirements, which we refer to as imagining the retrieval proxy, could provide more valuable details to the retrieval process.", + "bbox": [ + 511, + 521, + 905, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Imagine Proxy Layout. The first step of imagining a suitable proxy relies on envisioning the overall layout, which requires a clear understanding of the semantic content and relationship between the query images and relative captions. We thus leverage the robust interpretive and reasoning abilities of LLMs[1, 31] to generate a target proxy layout. As illustrated in Fig.2 (a), we organize the BLIP2 generated query image captions with the relative caption into the fixed form 'Given an image of{caption}, we show{rule}', based on which the LLM is instructed to infer the reasonable overall scene and the layout of the proxy image. The layout includes a detailed description of each instance as well as the bounding box (bbox) coordinates. Besides, certain instances may require detailed visual information from text or image modality. For instance, when retrieving an image of a dog wearing a hat, the attributes of the dog in the proxy image come from the dog in the query image, while the attributes of the hat come from the text modality information. We thus require the LLM to determine the reference modal", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3986", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/25611b28803426aca83b1f8e93e72784cf2b7401b1d4e20bf5c4e436fce2e231.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarkCIRCOCIRR
MetricmAP@KRecall@KRecall Subset@K
BackboneMethodk=5k=10k=25k=50k=1k=5k=10k=50k=1k=2k=3
ViT-L/14CIREVL18.5719.0120.8921.8024.5552.3164.9286.3459.5479.8889.69
LinCIR12.8513.7515.2616.1724.8253.0866.6088.8955.6176.2488.24
LinCIR + IP-CIR14.1615.2617.1518.0524.9653.7666.8288.8755.5276.5188.27
(vs. LinCIR)+1.31+1.51+1.89+1.88+0.14+0.68+0.22-0.02-0.09+0.27+0.03
LDRE23.3524.0326.4427.5026.5355.5767.5488.5060.4380.3189.90
LDRE + IP-CIR26.4327.4129.8731.0729.7658.8271.2190.4162.4881.6490.89
(vs. LDRE)+3.08+3.38+3.43+3.57+3.23+3.25+3.67+1.91+2.05+1.33+0.99
ViT-G/14CIREVL26.7727.5929.9631.0334.6564.2975.0691.6667.9584.8793.21
LinCIR20.3421.8523.9825.1434.7564.1275.9393.4262.3681.7891.28
LinCIR + IP-CIR25.7026.6429.0930.1335.3764.7076.1593.7162.5881.7491.35
(vs. LinCIR)+5.40+4.79+5.11+4.99+0.62+0.58+0.22+0.29+0.22-0.04+0.07
LDRE31.1232.2434.9536.0336.1566.3977.2593.9568.8285.6693.76
LDRE + IP-CIR32.7534.2636.8638.0339.2570.0780.0094.8969.9586.8794.22
(vs. LDRE)+1.63+2.02+1.91+2.00+3.10+3.68+2.75+0.94+1.13+1.21+0.46
", + "bbox": [ + 107, + 88, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Quantitative results in CIRCO and CIRR datasets.", + "bbox": [ + 323, + 378, + 671, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ity for each instance.", + "bbox": [ + 89, + 420, + 230, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proxy Image Generation. The next step is to imagine the proxy based on the generated layout. Since the CIR task is a fuzzy retrieval task [43], this reduces the need to maintain the query image's instance ID during the generation process. So we follow the idea of MIGC++ [49], incorporating the query image features transformed by ELITE [40] into MIGC, thereby introducing some degree of the query image information into the Proxy image. Besides, an instance can be influenced by both the reference image and the relative caption. For example, if a user envisions a hat in the query image with a specific pattern, the hat should retain its original shape but include the target pattern. So, we copy the layout of the instance with reference image modality and render it based on both image and text, ensuring the proxy image reflects the text's modifications.", + "bbox": [ + 88, + 436, + 482, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Balanced Proxy Retrieval", + "text_level": 1, + "bbox": [ + 89, + 676, + 323, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once the proxy image is generated, the key issue lies in how to leverage this proxy image in the retrieval. A straightforward method is to compute the proxy similarity of the retrieval database, weighted sum with the text similarity. However, as shown in the second line of Fig.6, directly using the CLIP features of the proxy image (+PI) can lead to an excessive focus on certain elements, like the beach background. This may cause the results to overly emphasize the visual information in the image while neglecting some details from the text. So in this section, we propose to construct a robust retrieval proxy and balance the retrieval result from image and text.", + "bbox": [ + 89, + 702, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Constructing Robust Proxy. Inspired by LDRE [43], we", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "use the LLM [1, 31] to infer a set of target captions, then obtain its CLIP features $f_{t}$ . We derive the semantic perturbation features $f_{s} = f_{t} - f_{o}$ with $f_{t}$ and the CLIP features of BLIP2 [23] generated query image captions $f_{o}$ . The difference of CLIP text features $f_{s}$ represents a semantic transformation direction from the query image to the target retrieval image, which aligns with the transformation direction in the image space. So the introduction of semantic perturbation provides useful information in the proxy images. Besides, The CLIP features of the original image also contain some basic information relevant to the current query. As shown in Fig.2 (b), we combine the query image's CLIP features $f_{q}$ , the proxy CLIP feature $f_{p}$ with the semantic perturbation $f_{s}$ into the robust proxy $f_{RP}$ . There are various ways to combine retrieval features. A simple approach is to scale all features to the same magnitude using max normalization as shown in Eq.1. Besides, the weights of each component can be adjusted to control retrieval bias based on the quality of the proxy image and specific retrieval goals. For example, assigning a higher weight to query features ensures a closer match to the query image if the user prioritizes similarity to the original query.", + "bbox": [ + 511, + 419, + 906, + 752 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {R P} = f _ {p} + \\frac {\\operatorname* {m a x} \\left(f _ {p}\\right)}{\\operatorname* {m a x} \\left(f _ {q}\\right)} f _ {q} + \\frac {\\operatorname* {m a x} \\left(f _ {p}\\right)}{\\operatorname* {m a x} \\left(f _ {s}\\right)} f _ {s}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 758, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Balancing Retrieval Results. After obtaining the retrieval feature in text and image spaces, a critical question arises: How should we balance the two retrieval results? The proxy features we obtain exist in the image feature space, making it difficult to directly perform weighted summation with the baseline text features. Therefore, we choose to combine their similarities. A simple approach is to take the", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3987", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/20978278862c0e87aa1dc0932dba9ce37d99ee69ffd5aee1902a3e0d70cf77ec.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneTypeShirtDressTopteeAverage
MethodR@10R@50R@10R@50R@10R@50R@10R@50
ViT-G/14Pic2Word33.1750.3925.4347.6535.2457.6231.2851.89
SEARLE36.4655.3528.1650.3239.8361.4534.8155.71
LinCIR46.7665.1138.0860.8850.4871.0945.1165.69
LinCIR + IP-CIR48.0466.6839.0261.0350.1871.1445.7466.28
(vs. LinCIR)+1.28+1.57+0.94+0.15-0.30+0.05+0.63+0.59
", + "bbox": [ + 161, + 88, + 836, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Quantitative results in FashionIQ dataset.", + "bbox": [ + 349, + 215, + 645, + 228 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/34f3db9c5438476110a026369fa1df8158b193385d96273578d83e9ca801aa02.jpg", + "image_caption": [ + "Figure 3. Qualitative results of our method. We conducted experiments on the CIRCO validation dataset to observe in which cases our method improves retrieval results. In the blue section on the left, we display the query information used for retrieval, the ground truth target image, and our generated proxy image features. 'Query' represents the input query image, 'Caption' represents the relative text, and in the 'Proxy' section, we show two generated imagined proxies. In the red section on the right, the top and bottom rows display the top-5 retrieval results enhanced by our imagined proxies and the baseline's top-5 retrieval results, respectively." + ], + "image_footnote": [], + "bbox": [ + 94, + 243, + 397, + 597 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7d7e5c5acf2266acebe63a751d65109f0b5969130b50f55ba378e9f4f8c5b3d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 243, + 905, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "average or weighted sum, however, in extreme cases, when we have a case with a 0.999 text similarity but only a 0.001 proxy similarity, and another case with both a 0.5 base similarity and a 0.5 proxy similarity, their averages are both 0.5, which could lead to retrieving overly extreme results. We thus propose a balanced metric for combining these results. Specifically, given the baseline similarity $S_{t}$ and proxy similarity $S_{p}$ , the final similarity $S_{f}$ for retrieval is as follows:", + "bbox": [ + 89, + 698, + 482, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {f} = \\lambda S _ {t} + (1 - \\lambda) S _ {b}, S _ {b} = S _ {t} * S _ {p}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 832, + 482, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S_{b}$ represents the balanced similarity, $\\lambda$ is a balancing parameter. We believe that the balance metric enables us to focus more on cases that perform well in both similarities.", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 696, + 643, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we demonstrate the effectiveness of our method. In Section 4.1, We will provide a detailed description of our experimental setup, including the baselines, evaluation metrics, and implementation details. In Section 4.2, we compare our method with other state-of-the-art approaches. We explore the improvements in retrieval performance by presenting some visual examples, highlighting the impact of our method in Section 4.3, and in Section 4.4, we conduct ablation studies to investigate the contribution of each component of our method, as well as the influence of various parameter settings.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3988", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 90, + 279, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We validate the performance of our method by combining them with several state-of-the-art methods on three datasets: CIRR [28], CIRCO [5], and FashionIQ [41]. CIRR contains 21,552 real images from the NLVR [36] dataset. CIRCO is built from the COCO 2017 unlabeled set [27], includes a validation set with 220 queries and a test set with 800 queries. FashionIQ [41] is a dataset dedicated to the fashion domain, organized into three distinct subcategories: Dress, Shirt, and Toptee. It contains 30,135 query triplets and 77,683 images available for retrieval.", + "bbox": [ + 89, + 113, + 483, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline. We selected LDRE[43] and LinCIR[16] as our baselines. For each query, we generate 5 proxy images and obtain the enhanced retrieval results. We compare the improved results with some of the most recent ZS-CIR results, such as Searle[6], CIREVL[20].", + "bbox": [ + 89, + 265, + 483, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. In CIRCO, each query has multiple target images, so we calculate mAP@k to offer a more detailed assessment, where $k \\in \\{5,10,25,50\\}$ indicates the number of top-ranked retrieval results. For CIRR, we follow the original benchmarks and use Recall@k (with $k \\in \\{1,5,10,50\\}$ ) as the primary metric. We also assess performance in a subset setting, which we denote as RecallSubset@K (with $k \\in \\{1,2,3\\}$ ). On FashionIQ, we also use Recall@K (with $k \\in \\{10,50\\}$ ) for evaluation.", + "bbox": [ + 89, + 340, + 483, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implement Details. We primarily conducted experiments using two backbone models: CLIP-L and CLIP-G. We used Qwen1.5-32B[1] as the LLM in all the experiments, which yielded results comparable to those reported in the original papers. For generating proxy images, we use MIGC[48] with SD1.5 as the backbone. We use Qwen1.5-32B as our LLM to generate the layout for each query and then generate 5 proxy images based on the query image and layout. All the experiments are conducted by Pytorch with one NVIDIA A6000 GPU.", + "bbox": [ + 89, + 477, + 483, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 635, + 284, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CIRCO. We present the experimental results on the CIRCO test dataset on LDRE[43] and LinCIR[16]. In the left half of Tab.1, we observe that our proxy image can improve the retrieval performance, which an approximately 5-point improvement in mAP at $k = 5$ in LinCIR, and about 2-point improvement in LDRE. These results effectively demonstrate the validity and impact of our method.", + "bbox": [ + 89, + 657, + 483, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CIRR. In the right half of Tab.1, we present the experimental results on the CIRR test dataset. Our proxy image significantly improved the LDRE and LinCIR performance. The performance improvements across different methods and backbones indicate that our approach can be applied to various text-based retrieval methods.", + "bbox": [ + 89, + 763, + 483, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FashionIQ. In Tab.2, we present the results on the FashionIQ val dataset. Since LDRE has not open-sourced the code for the FashionIQ, we used LinCIR as the baseline.", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/156d1b69e3359d67382688657255f16805d054d4417d846d571d58822cab6dba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 90, + 709, + 271 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b68559c739a8c03a95950baa441cefeb0805e558c5c5c5905bcc661c157953ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 90, + 903, + 271 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a250bcd7faab5928edf80e6fb10a88e0aa32c436104ac78299da7652bf2b520a.jpg", + "image_caption": [ + "Figure 4. Qualitative results of our method. We present the result on CIRCO validation dataset with top-1 retrieval results. Baseline methods in text-based retrieval overlook certain attributes and details in complex related texts, resulting in suboptimal retrieval outcomes. In contrast, our approach can construct suitable proxies on the image side, supplementing this missing information and yielding improved retrieval results." + ], + "image_footnote": [], + "bbox": [ + 517, + 273, + 709, + 351 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ff09d93475ebd6315086384d58bd87425ebd27590768ace9c858ce3c0b54ffc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 273, + 905, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f869bcbc0e4f53ab9c104de527212126d44b88adc91fa942006c97b7f6242ee3.jpg", + "image_caption": [ + "Figure 5. Ablation results on the composition of robust proxy features in the CIRCO dataset. $pi$ indicates proxy features, $si$ represents semantic perturbation, and $qi$ indicates the query features." + ], + "image_footnote": [], + "bbox": [ + 542, + 472, + 875, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although we observe a drop in $\\mathbb{R}@\\mathbb{10}$ in the Toptee, our method improves the retrieval accuracy in the Shirt and dress Subset, thus improving the average performance.", + "bbox": [ + 511, + 743, + 906, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 800, + 694, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted experiments on the CIRCO validation dataset to verify the improvements our method brings to retrieval. Experiment setting. Specifically, we used LDRE as the baseline and generated 5 imagined proxy images for each query. We present two types of results. First, based on", + "bbox": [ + 511, + 824, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3989", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8df1d113611ce9dffb831e0547d1a5cfc0063921981b7fbf0ee29fbcc2dfaa43.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarkCIRRCIRCO
MetricRecall@KmAP@K
BackbonePIRPBMk=1k=2k=5k=10k=50k=5k=10k=25k=50
LDRE-G✓36.1549.4966.3977.2593.9531.1232.2434.9536.03
36.0250.0266.5177.5293.7431.0632.4135.1936.35
-0.13+0.53+0.12+0.27-0.21-0.06+0.17+0.24+0.32
✓✓38.1552.5368.5579.4794.4831.9233.5936.2637.37
+2.00+3.04+2.16+2.22+0.53+0.80+1.35+1.31+1.34
✓✓✓39.2552.9470.0780.0094.8932.7534.2636.8638.03
+3.10+3.45+3.68+2.75+0.94+1.63+2.02+1.91+2.00
", + "bbox": [ + 135, + 88, + 864, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation results in CIRR and CIRCO dataset.", + "bbox": [ + 336, + 273, + 658, + 285 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e2716f24ff5d3e1c39219c9492f68a6a546a41ab6e32d7b7fc4035390455251.jpg", + "image_caption": [ + "Figure 6. Visualize of Ablation Result. We present the top-1 retrieval result on CIRCO dataset in ablation experiments." + ], + "image_footnote": [], + "bbox": [ + 116, + 301, + 869, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the ground truth for each query, we show in Fig.3 that our method improves the accuracy of ground truth occurrence in top-5 retrieval results. Second, disregarding the ground truth image, we demonstrate in Fig.4 that our method retrieves a more suitable image in the top-1 result.", + "bbox": [ + 88, + 617, + 482, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Experimental Results on TOP-5 Retrieval. From the visualized results in Fig.3, we can draw the following conclusions: (1) Our generated proxy images effectively preserve the style and key elements of the original image. In the first row, our image maintains the clock tower as a significant element and adheres to the related caption by generating a Ferris wheel. In the second row, our image retains the original black-and-white style. (2) The imagined proxies complement important information that may be overlooked by text alone. For instance, in the first-row example, the clock tower—a crucial concept—is missing in the text-only retrieval result, while our imagined proxy and the original image both provide this retrieval information.", + "bbox": [ + 88, + 703, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the third row, our constructed image improves retrieval accuracy by incorporating an umbrella and snow.", + "bbox": [ + 511, + 617, + 906, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Experimental Results on TOP-1 Retrieval. Cases in Fig.4 show that by providing proxy image information, our approach achieves more relevant retrievals. Retrieval based solely on text space may only capture partial information in complex contexts, leading to the omission of certain details in the images—such as the person in the first row, the tie in the second row, and the White House in the background of the third row. Our method incorporates these details in the images, supplementing and balancing this information to yield more suitable results.", + "bbox": [ + 511, + 648, + 906, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 815, + 666, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of the Proxy Image Features. We first conduct an ablation study on the proposed proxy image features in the CIRR dataset and CIRCO dataset. We use PI (Proxy Image) refers to the approach that directly computes retrieval", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3990", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/434895b3be0491b3fb8913ec7e160f82ae98b0c4e788b2c23c9f8068830427b9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricmAP@K
Backbonek=5k=10k=25k=50
LDRE-G31.1232.2434.9536.03
LDRE-G + SD32.2533.5036.1837.30
LDRE-G + FLUX32.3833.8936.5137.58
LDRE-G + ip2p31.8233.2736.1137.23
LDRE-G + magicbrush31.8633.3936.0537.22
LDRE-G + L2I32.7534.2636.8638.03
", + "bbox": [ + 99, + 77, + 478, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Comparison of Various Methods for Generating Proxy Images on the CIRCO test", + "bbox": [ + 89, + 218, + 480, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "similarity from proxy image features, and then performs a weighted sum with the original similarity. RP (Robust Proxy) indicates integrating the proxy image into more robust proxy features. BM (Balance Metric) refers to using the balance metric. The results from the CIRR and CIRCO are shown in Tab.3. We can draw the following conclusions: 1) The imagined proxy images may contain some noise, leading to the retrieval focus on irrelevant details. we observe a decrease in recall at $k = 1$ and 50 in CIRR, and mAP $k = 5$ decreases in CIRCO. 2) After integrating the proxy image into a more robust proxy feature, we see a significant improvement in both datasets, which demonstrates the effectiveness of our robust proxy features. 3) By introducing the balance metric between proxy and text similarity, we further improve the retrieval performance. This suggests that combining the information from both image and text domains leads to more accurate retrieval results.", + "bbox": [ + 88, + 247, + 482, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We can draw similar conclusions from Fig.6, which shows retrieval results on the CIRCO test dataset. In the first row, the proxy image provides a realistic and accurate depiction of the target image, allowing for improved retrieval results. In the second row, relying solely on the query image may cause the retrieval to overly focus on the beach background, contradicting the \"no people\" text information. Here, using the robust proxy $(\\mathrm{PI} + \\mathrm{RP})$ further enhances the overlooked aspects. In the third row, the balance metric balances the 'the woman with sunglasses' in proxy features with 'trees' in baseline features, achieving better retrieval results for both the subject and the background.", + "bbox": [ + 89, + 506, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Robust Features. We analyze the composition of our robust proxy on the CIRCO dataset, with results shown in Fig.5. The robust proxy is composed of three main components: $q_{i}$ represents the query image features, $s_{i}$ represents the semantic perturbation, and $p_{i}$ represents our proxy image. A comparison among ①, ②, ③, and ④ shows that directly using our proxy image features can enhance retrieval performance to a certain extent, whereas relying solely on the query image or semantic perturbation tends to focus excessively on incorrect information, leading to a decline in performance. Additionally, comparisons among ⑤, ⑥, and ⑦ demonstrate that combining all three features together improves retrieval accuracy.", + "bbox": [ + 89, + 686, + 482, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Approaches for constructing proxy images. We also", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c31ae80af61a82718ca1b8ebfc44def5222ccd5b3006ba119a917be2ef18250d.jpg", + "image_caption": [ + "Figure 7. Parameter Analysis. The effect of varying numbers of proxy images on performance for the CIRCO dataset." + ], + "image_footnote": [], + "bbox": [ + 566, + 99, + 831, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tested the results obtained on the CIRCO dataset using different methods to construct the proxy image in Tab.4. Using the target image caption inferred by the LLM, we tested two T2I generation approaches: SD[33] and FLUX[22]. Additionally, based on the relative caption and query image, we also tested two instruction-based editing approaches Instruct-Pix2Pix[7] and MagicBrush[44]. The results show that: 1) The robust features constructed through these approaches can all provide additional information for retrieval. 2) The stronger the control capability, the more precise the provided information, and the greater the improvement achieved.", + "bbox": [ + 511, + 303, + 906, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of the number of proxy images. We tested the impact of using different numbers of proxy images on performance with the CIRCO dataset. As shown in Fig.7, retrieval performance improves as the number of proxy images increases, though the rate of improvement gradually slows down. Therefore, the number of proxy images can be chosen based on a balance between efficiency and accuracy.", + "bbox": [ + 511, + 484, + 908, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 603, + 633, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose the IP-CIR, a plug-and-play training-free method for current ZSCIR tasks. Our method imagines proxy images, which align with the query image and the relative captions, providing fine-grained details that may be overlooked by the text. We combine the proxy images, query images, and semantic perturbation into more robust proxy image features, and propose a balance metric to compose the proxy and baseline similarities, enhancing the retrieval performance. The experiment results in CIRR, CIRCO, and FashionIQ datasets show that our method can successfully improve retrieval accuracy and demonstrate the potential for providing additional retrieval information from the image side, revealing a new direction for improving retrieval accuracy in the future.", + "bbox": [ + 511, + 628, + 906, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (62293554) and in part by the Natural Science Foundation of Zhejiang Province (LDT23F02023F02).", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3991", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2, 3, 4, 6", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1", + "[3] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21466-21474, 2022. 1", + "[4] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4959-4968, 2022. 1", + "[5] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion, 2023. 1, 2, 6", + "[6] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15338-15347, 2023. 1, 2, 6", + "[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 8", + "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models, 2023. 2", + "[9] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481, 2023. 2", + "[10] Yanbei Chen and Loris Bazzani. Learning joint visual semantic matching embeddings for language-guided retrieval. page 136-152, Berlin, Heidelberg, 2020. Springer-Verlag. 1", + "[11] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2998-3008, 2020.", + "[12] Yiyang Chen, Zhedong Zheng, Wei Ji, Leigang Qu, and Tat" + ], + "bbox": [ + 93, + 114, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Seng Chua. Composed image retrieval with text feedback via multi-grained uncertainty regularization, 2024.", + "[13] Sanghyuk Chun. Improved probabilistic image-text representations, 2024.", + "[14] Ginger Delmas, Rafael Sampaio de Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity, 2022.", + "[15] Eric Dodds, Jack Culpepper, Simao Herdade, Yang Zhang, and Kofi Boakye. Modality-agnostic attention fusion for visual search with text feedback, 2020. 1", + "[16] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, , Yoohoon Kang, and Sangdoo Yun. Language-only training of zero-shot composed image retrieval. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 6", + "[17] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoonhoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. Transactions on Machine Learning Research, 2024. Expert Certification. 2", + "[18] Ryota Hinami and Shin'ichi Satoh. Discriminative learning of open-vocabulary object retrieval and localization by negative phrase augmentation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2605-2615, Brussels, Belgium, 2018. Association for Computational Linguistics. 1", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2", + "[20] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. International Conference on Learning Representations (ICLR), 2024. 2, 6", + "[21] Jooyeon Kim, Eulrang Cho, Sehyung Kim, and Hyunwoo J. Kim. Retrieval-augmented open-vocabulary object detection, 2024. 1", + "[22] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.8", + "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3, 4", + "[24] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Glagen: Open-set grounded text-to-image generation. CVPR, 2023. 2", + "[25] You Li, Fan Ma, and Yi Yang. Anysynth: Harnessing the power of image synthetic data generation for generalized vision-language tasks. arXiv preprint arXiv:2411.16749, 2024. 1", + "[26] Chao Liang, Fan Ma, Linchao Zhu, Yingying Deng, and Yi Yang. Caphuman: Capture your moments in parallel universes. In CVPR, pages 6400-6409, 2024. 1", + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 6" + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3992", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre-trained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2125–2134, 2021. 1, 2, 6", + "[29] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers, 2021. 1", + "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2022. 1", + "[31] OpenAI. Gpt-4 technical report, 2023. 2, 3, 4", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 2, 8", + "[34] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. CVPR, 2023. 1, 2", + "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 1, 2", + "[36] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs, 2019. 6", + "[37] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval, 2024. 1", + "[38] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation, 2024. 2", + "[39] Zheng Wang, Xin Yuan, Toshihiko Yamasaki, Yutian Lin, Xin Xu, and Wenjun Zeng. Re-identification = retrieval + verification: Back to essence and forward with a new metric. arXiv preprint arXiv:2011.11506, 2020. 1", + "[40] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023. 3, 4", + "[41] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The fashion iq dataset: Retrieving images by combining side information and relative natural language feedback. CVPR, 2021. 1, 2, 6", + "[42] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Reco: Region-controlled text-to-image generation. In CVPR, 2023. 2", + "[43] Zhenyu Yang, Dizhan Xue, Shengsheng Qian, Weiming Dong, and Changsheng Xu. Ldre: Llm-based divergent reasoning and ensemble for zero-shot composed image retrieval. In Proceedings of the 47th International ACM SI-" + ], + "bbox": [ + 91, + 92, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "GIR Conference on Research and Development in Information Retrieval, pages 80-90, 2024. 1, 2, 4, 6", + "[44] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 8", + "[45] Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. MagicLens: Self-supervised image retrieval with open-ended instructions. In Proceedings of the 41st International Conference on Machine Learning, pages 59403–59420. PMLR, 2024. 1, 2", + "[46] Liang Zheng, Yi Yang, and Alexander G. Hauptmann. Person re-identification: Past, present and future, 2016. 1", + "[47] Dewei Zhou, Zongxin Yang, and Yi Yang. Pyramid diffusion models for low-light image enhancement. arXiv preprint arXiv:2305.10028, 2023. 1", + "[48] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6818-6828, 2024. 2, 6", + "[49] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Migc++: Advanced multi-instance generation controller for image synthesis. ArXiv, abs/2407.02329, 2024. 2, 3, 4", + "[50] Dewei Zhou, Ji Xie, Zongxin Yang, and Yi Yang. 3dis: Depth-driven decoupled instance synthesis for text-to-image generation. ArXiv, abs/2410.12669, 2024. 1", + "[51] Dewei Zhou, Mingwei Li, Zongxin Yang, and Yi Yang. Dreamrenderer: Taming multi-instance attribute control in large-scale text-to-image models, 2025. 1", + "[52] Dewei Zhou, Ji Xie, Zongxin Yang, and Yi Yang. 3dis-flux: simple and efficient multi-instance generation with dit rendering, 2025. 2" + ], + "bbox": [ + 516, + 92, + 903, + 573 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3993", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/f227006c-210b-4aac-822f-93d2a98acc36_model.json b/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/f227006c-210b-4aac-822f-93d2a98acc36_model.json new file mode 100644 index 0000000000000000000000000000000000000000..085639c4b86aa0bca9e0f561fba8c7c1298fc894 --- /dev/null +++ b/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/f227006c-210b-4aac-822f-93d2a98acc36_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.13, + 0.882, + 0.178 + ], + "angle": 0, + "content": "Imagine and Seek: Improving Composed Image Retrieval with an Imagined Proxy" + }, + { + "type": "text", + "bbox": [ + 0.37, + 0.203, + 0.626, + 0.222 + ], + "angle": 0, + "content": "You Li Fan Ma Yi Yang" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.222, + 0.71, + 0.24 + ], + "angle": 0, + "content": "ReLER, CCAI, Zhejiang University, Zhejiang, China" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.241, + 0.571, + 0.256 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.259, + 0.662, + 0.275 + ], + "angle": 0, + "content": "{uli2000, mafan, yangyics}@zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.485, + 0.752 + ], + "angle": 0, + "content": "The Zero-shot Composed Image Retrieval (ZSCIR) requires retrieving images that match the query image and the relative captions. Current methods focus on projecting the query image into the text feature space, subsequently combining them with features of query texts for retrieval. However, retrieving images only with the text features cannot guarantee detailed alignment due to the natural gap between images and text. In this paper, we introduce Imagined Proxy for CIR (IP-CIR), a training-free method that creates a proxy image aligned with the query image and text description, enhancing query representation in the retrieval process. We first leverage the large language model's generalization capability to generate an image layout, and then apply both the query text and image for conditional generation. The robust query features are enhanced by merging the proxy image, query image, and text semantic perturbation. Our newly proposed balancing metric integrates text-based and proxy retrieval similarities, allowing for more accurate retrieval of the target image while incorporating image-side information into the process. Experiments on three public datasets demonstrate that our method significantly improves retrieval performances. We achieve state-of-the-art (SOTA) results on the CIRR dataset with a Recall@K of 70.07 at \\( K = 10 \\). Additionally, we achieved an improvement in Recall@10 on the FashionIQ dataset, rising from 45.11 to 45.74, and improved the baseline performance in CIRCO with a mAPK@10 score, increasing from 32.24 to 34.26." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.784, + 0.222, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Imagination is the wellspring of human creativity. When you see a picture of an adorable cat, you might envision it playing with other cats, wearing cool sunglasses and a cape, or floating in a space capsule. This ability to imagine and pursue beautiful visions drives continuous advancement in fields like image generation [2, 19, 26, 30, 33, 35, 47, 50," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.31, + 0.906, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.556 + ], + "angle": 0, + "content": "Figure 1. Comparison of composed image retrieval between text-only retrieval (LDRE in Fig) and our methods. Traditional methods perform retrieval only in the text space, where CLIP text features may overlook some important semantic information. In contrast, our approach generates imagined proxy features, providing additional information that is often overlooked in text-only retrieval, thereby improving retrieval accuracy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.577, + 0.905, + 0.593 + ], + "angle": 0, + "content": "51] and image retrieval [5, 6, 16, 25, 28, 34, 37, 41, 43, 45]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.596, + 0.907, + 0.793 + ], + "angle": 0, + "content": "This rich imagination gains new significance in Composed Image Retrieval (CIR) [28], where users express their imagination of the query image through text, searching for content that aligns with their preferences. Specifically, the CIR task requires users to retrieve target images from a database that match both the content of a query image and relative caption. With text modality to flexibly control retrieval targets, CIR enables the search process akin to the user's specific imagination after viewing the query image. This multi-modal, imagination-like retrieval task greatly expands the applications of image retrieval, with wide-ranging uses in fields such as visual search [29], object localization [18, 21], and re-identification [39, 46]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Traditional solutions for CIR [3, 4, 10-15] are constrained by the scarcity of training data (e.g., 46.6k for FashionIQ [41] dataset and 28.8k for CIRR [28] dataset), as collecting suitable triplets (query image \\( x_{q} \\), relevant text \\( x_{c} \\), and target image \\( x_{t} \\)) from the web incurs high labor costs. Therefore, we focus on the ZS-CIR task, aiming to bypass the need for training on triplet datasets and achieve gen" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3984" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "eralized CIR retrieval capabilities. As shown in Fig.1 (a), traditional ZS-CIR methods typically employ lightweight projection models to map CLIP's [32] image features into the text features [6] for retrieval. The transformation inevitably loses certain image information, and the coarse, easily confusable text features fall short in fine-grained, complex imagination scenes. Although recent approaches attempt to leverage large language models (LLMs) [1, 31] to generate various descriptions of target images [20, 43, 45], they overlook the potential for direct imagination on the image side. As shown in Fig.1 (a), the text features did not effectively capture the attribute 'golden', resulting in suboptimal retrieval results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.483, + 0.471 + ], + "angle": 0, + "content": "The rapid advancement of controllable generative models [8, 9, 24, 38, 42, 49, 52] has empowered users with the freedom to create images that align with their imagined visions. Generating images that match both the content of query images and relative captions aligns closely with the goals of composed retrieval tasks, and the imagined images (we called Retrieval Proxy) could provide additional information—such as style, instance attributes, and spatial relationships—that is often overlooked in text-based retrieval with complex captions. This raises an intriguing question: How could we harness the imaginative power of generative models to improve retrieval performance?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.483, + 0.652 + ], + "angle": 0, + "content": "We thus propose IP-CIR (Imagined Proxy for CIR), a training-free, plug-and-play method that harnesses the power of imagination for any retrieval method. Firstly, by utilizing LLMs to understand both the image content and the relative captions, we generate a suitable image layout. Using controllable generation methods, we create high-quality, fine-grained proxy images. Next, we compose the proxy images, query images as well as semantic perturbation into robust proxy features that are more suitable for retrieval. Finally, we propose a balance metric to combine the proxy and text retrieval similarities, allowing the integration of our method with any retrieval methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.744 + ], + "angle": 0, + "content": "We conduct experiments on three public datasets: the CIRR [28], CIRCO [5], and FashionIQ [41]. We combine our method with several baselines, the observed performance improvements confirm that our method supplements the retrieval process with valuable information and could enhance the performance of any other retrieval methods." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.746, + 0.414, + 0.76 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.765, + 0.483, + 0.825 + ], + "angle": 0, + "content": "1) We propose using generative models' imagination capabilities to create proxy images that align with the content of the query image and relative captions to improve retrieval performance." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.826, + 0.483, + 0.871 + ], + "angle": 0, + "content": "2) We introduce a proxy image generation framework. Using the LLM-generated object layout and the query image, we create imagined proxies through MIGC++." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "3) We propose integrating proxy images with the query image and semantic perturbation into more robust proxy fea" + }, + { + "type": "list", + "bbox": [ + 0.084, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "tures and introducing a balance metric to combine the proxy and baseline similarities." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.122, + 0.905, + 0.183 + ], + "angle": 0, + "content": "4) We conducted experiments on three public CIR datasets. The improvement over baseline results, as well as achieving SOTA results demonstrates the potential of constructing imagined proxies in retrieval." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.198, + 0.652, + 0.213 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.223, + 0.838, + 0.239 + ], + "angle": 0, + "content": "2.1. Zero-shot Composed Image Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.246, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Early work on ZS-CIR, such as Pic2word [34], SEALRE [6], attempted to project the original image into the text space using lightweight projection models. This was often done within fixed templates (e.g., \"a photo of \\([\\\\) ]$ with cond\"), the \\([\\\\( ]\\) symbol is a textual notation used to represent query image information, and \"cond\" represents the relative captions. LinCIR [16] further defined keywords as adjectives and nouns in the text, thus overcoming the limitations of fixed templates and training more powerful projection models. Besides, CompoDiff[17] trains a new diffusion model that performs diffusion on the CLIP features of the image, introduces relative caption features, and applies attention operations with intermediate image features, ultimately obtaining the target features for retrieval. In addition, some methods have embraced LLMs [1, 31]. CIREVL [20] proposed generating captions for reference images and then using LLMs to reassemble them into target texts modified by relative captions for retrieval. LDRE [43] took this further by using LLMs to construct various edited captions and combining them based on similarity, addressing the issue of fuzzy retrieval. While these methods improved retrieval accuracy, they all overlooked the potential of directly imagining proxy images to provide additional information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.62, + 0.851, + 0.636 + ], + "angle": 0, + "content": "2.2. Controllable Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In recent years, with the development of diffusion models [19, 35], high-quality images that meet specific requirements can now be generated. Using mainstream generators such as Stable Diffusion [33], users can generate images aligned with text descriptions. Although these foundational generators offer a certain level of control, their performance in fine-grained attribute and orientation control is limited, primarily due to the attribute ambiguity in the CLIP features of the text [48]. As a result, some efforts have focused on leveraging layout information to enhance control. For example, GLIGEN [24] introduces Gated-Self-Attention to combine positional feature information, thereby improving control over positioning. InstanceDiffusion [38] further improves attribute control on this foundation. Meanwhile, MIGC [48] adopts a divide-and-conquer approach, breaking down complex instance generation into multiple single-instance generations and incorporating Enhancement Atten" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3985" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.099, + 0.454, + 0.114 + ], + "angle": 0, + "content": "(a) Imagined Retrieval Proxy Generation" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.115, + 0.51, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.097, + 0.82, + 0.112 + ], + "angle": 0, + "content": "(b) Balanced Proxy Retrieval" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.113, + 0.902, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.261, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Overview of our method. (a) represents our imagined proxy generation process. We use LLM to analyze the BLIP2-generated query image captions and the relative captions and infer the proxy layout. We then use the controllable generator to imagine the proxy images. (b) represents our process of constructing a robust proxy feature, and balancing the text and proxy similarities. We integrate proxy features, query image features as well as semantic perturbations into a robust proxy feature, and propose a balance metric for retrieval." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.485, + 0.511 + ], + "angle": 0, + "content": "tion and Layout Attention to improve both positional and attribute control. MIGC++ [49] further integrates ELITE [40] to transform images into CLIP-based text features and introduces an additional EA layer to process information from the reference image modality, thereby incorporating image modality control in the generation process. In this paper, we aim to utilize the fine-grained control capabilities of MIGC to display certain attributes and other information that are difficult to clearly express in text at a fine-grained level within the image itself. This provides supplementary information during the retrieval process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.524, + 0.182, + 0.539 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.716 + ], + "angle": 0, + "content": "In this section, we will introduce our IP-CIR, which attempts to imagine an appropriate proxy image within the image feature space, balance, and combine the retrieval results from the image side with those from the text side. Specifically, in section 3.1, we introduce the overview of our method, in section 3.2, we will describe how to construct an imagined retrieval proxy for each query. In section 3.3, we will further explain how to organize and integrate the information from the imagined retrieval proxy and present how to combine the retrieval information from both the image and text feature spaces." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.201, + 0.741 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The overall structure of our framework is shown in Fig.2. First, we generate the imagined proxy image for retrieval. The proxy image needs to align with the content of the query image and be as consistent as possible with the description in the relative captions. Therefore, we first generate a reasonable object layout and then use a controllable generation model to imagine this layout as a concrete image. Specifically, we use BLIP2[23] to generate a set of captions for each query. Combine with the relative captions, we instruct the LLM to reason out a suitable" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.907, + 0.48 + ], + "angle": 0, + "content": "spatial layout of the target image, so that, with as much accuracy as possible in terms of scene, instance attributes, and quantity with relative caption, we can imagine a proxy image of the retrieval target to provide additional information for the retrieval process. In Fig.2 (b), we compose the imagined proxy with query image and text features, forming a robust retrieval proxy. The similarity of robust proxy is further balanced with other baselines' similarity, improving the retrieval accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.496, + 0.84, + 0.512 + ], + "angle": 0, + "content": "3.2. Imagined Retrieval Proxy Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.612 + ], + "angle": 0, + "content": "Text retrieval features may struggle to precisely capture important semantics in the relevant text and the query image, especially when the caption is complex. Generating an image that meets the retrieval requirements, which we refer to as imagining the retrieval proxy, could provide more valuable details to the retrieval process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Imagine Proxy Layout. The first step of imagining a suitable proxy relies on envisioning the overall layout, which requires a clear understanding of the semantic content and relationship between the query images and relative captions. We thus leverage the robust interpretive and reasoning abilities of LLMs[1, 31] to generate a target proxy layout. As illustrated in Fig.2 (a), we organize the BLIP2 generated query image captions with the relative caption into the fixed form 'Given an image of{caption}, we show{rule}', based on which the LLM is instructed to infer the reasonable overall scene and the layout of the proxy image. The layout includes a detailed description of each instance as well as the bounding box (bbox) coordinates. Besides, certain instances may require detailed visual information from text or image modality. For instance, when retrieving an image of a dog wearing a hat, the attributes of the dog in the proxy image come from the dog in the query image, while the attributes of the hat come from the text modality information. We thus require the LLM to determine the reference modal" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3986" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.891, + 0.38 + ], + "angle": 0, + "content": "
BenchmarkCIRCOCIRR
MetricmAP@KRecall@KRecall Subset@K
BackboneMethodk=5k=10k=25k=50k=1k=5k=10k=50k=1k=2k=3
ViT-L/14CIREVL18.5719.0120.8921.8024.5552.3164.9286.3459.5479.8889.69
LinCIR12.8513.7515.2616.1724.8253.0866.6088.8955.6176.2488.24
LinCIR + IP-CIR14.1615.2617.1518.0524.9653.7666.8288.8755.5276.5188.27
(vs. LinCIR)+1.31+1.51+1.89+1.88+0.14+0.68+0.22-0.02-0.09+0.27+0.03
LDRE23.3524.0326.4427.5026.5355.5767.5488.5060.4380.3189.90
LDRE + IP-CIR26.4327.4129.8731.0729.7658.8271.2190.4162.4881.6490.89
(vs. LDRE)+3.08+3.38+3.43+3.57+3.23+3.25+3.67+1.91+2.05+1.33+0.99
ViT-G/14CIREVL26.7727.5929.9631.0334.6564.2975.0691.6667.9584.8793.21
LinCIR20.3421.8523.9825.1434.7564.1275.9393.4262.3681.7891.28
LinCIR + IP-CIR25.7026.6429.0930.1335.3764.7076.1593.7162.5881.7491.35
(vs. LinCIR)+5.40+4.79+5.11+4.99+0.62+0.58+0.22+0.29+0.22-0.04+0.07
LDRE31.1232.2434.9536.0336.1566.3977.2593.9568.8285.6693.76
LDRE + IP-CIR32.7534.2636.8638.0339.2570.0780.0094.8969.9586.8794.22
(vs. LDRE)+1.63+2.02+1.91+2.00+3.10+3.68+2.75+0.94+1.13+1.21+0.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.324, + 0.38, + 0.673, + 0.392 + ], + "angle": 0, + "content": "Table 1. Quantitative results in CIRCO and CIRR datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.232, + 0.435 + ], + "angle": 0, + "content": "ity for each instance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.437, + 0.483, + 0.664 + ], + "angle": 0, + "content": "Proxy Image Generation. The next step is to imagine the proxy based on the generated layout. Since the CIR task is a fuzzy retrieval task [43], this reduces the need to maintain the query image's instance ID during the generation process. So we follow the idea of MIGC++ [49], incorporating the query image features transformed by ELITE [40] into MIGC, thereby introducing some degree of the query image information into the Proxy image. Besides, an instance can be influenced by both the reference image and the relative caption. For example, if a user envisions a hat in the query image with a specific pattern, the hat should retain its original shape but include the target pattern. So, we copy the layout of the instance with reference image modality and render it based on both image and text, ensuring the proxy image reflects the text's modifications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.678, + 0.325, + 0.695 + ], + "angle": 0, + "content": "3.3. Balanced Proxy Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Once the proxy image is generated, the key issue lies in how to leverage this proxy image in the retrieval. A straightforward method is to compute the proxy similarity of the retrieval database, weighted sum with the text similarity. However, as shown in the second line of Fig.6, directly using the CLIP features of the proxy image (+PI) can lead to an excessive focus on certain elements, like the beach background. This may cause the results to overly emphasize the visual information in the image while neglecting some details from the text. So in this section, we propose to construct a robust retrieval proxy and balance the retrieval result from image and text." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Constructing Robust Proxy. Inspired by LDRE [43], we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.908, + 0.753 + ], + "angle": 0, + "content": "use the LLM [1, 31] to infer a set of target captions, then obtain its CLIP features \\( f_{t} \\). We derive the semantic perturbation features \\( f_{s} = f_{t} - f_{o} \\) with \\( f_{t} \\) and the CLIP features of BLIP2 [23] generated query image captions \\( f_{o} \\). The difference of CLIP text features \\( f_{s} \\) represents a semantic transformation direction from the query image to the target retrieval image, which aligns with the transformation direction in the image space. So the introduction of semantic perturbation provides useful information in the proxy images. Besides, The CLIP features of the original image also contain some basic information relevant to the current query. As shown in Fig.2 (b), we combine the query image's CLIP features \\( f_{q} \\), the proxy CLIP feature \\( f_{p} \\) with the semantic perturbation \\( f_{s} \\) into the robust proxy \\( f_{RP} \\). There are various ways to combine retrieval features. A simple approach is to scale all features to the same magnitude using max normalization as shown in Eq.1. Besides, the weights of each component can be adjusted to control retrieval bias based on the quality of the proxy image and specific retrieval goals. For example, assigning a higher weight to query features ensures a closer match to the query image if the user prioritizes similarity to the original query." + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.759, + 0.907, + 0.793 + ], + "angle": 0, + "content": "\\[\nf _ {R P} = f _ {p} + \\frac {\\operatorname* {m a x} \\left(f _ {p}\\right)}{\\operatorname* {m a x} \\left(f _ {q}\\right)} f _ {q} + \\frac {\\operatorname* {m a x} \\left(f _ {p}\\right)}{\\operatorname* {m a x} \\left(f _ {s}\\right)} f _ {s}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Balancing Retrieval Results. After obtaining the retrieval feature in text and image spaces, a critical question arises: How should we balance the two retrieval results? The proxy features we obtain exist in the image feature space, making it difficult to directly perform weighted summation with the baseline text features. Therefore, we choose to combine their similarities. A simple approach is to take the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3987" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.162, + 0.089, + 0.837, + 0.216 + ], + "angle": 0, + "content": "
BackboneTypeShirtDressTopteeAverage
MethodR@10R@50R@10R@50R@10R@50R@10R@50
ViT-G/14Pic2Word33.1750.3925.4347.6535.2457.6231.2851.89
SEARLE36.4655.3528.1650.3239.8361.4534.8155.71
LinCIR46.7665.1138.0860.8850.4871.0945.1165.69
LinCIR + IP-CIR48.0466.6839.0261.0350.1871.1445.7466.28
(vs. LinCIR)+1.28+1.57+0.94+0.15-0.30+0.05+0.63+0.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.35, + 0.217, + 0.647, + 0.229 + ], + "angle": 0, + "content": "Table 2. Quantitative results in FashionIQ dataset." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.244, + 0.398, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.244, + 0.906, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.608, + 0.907, + 0.68 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of our method. We conducted experiments on the CIRCO validation dataset to observe in which cases our method improves retrieval results. In the blue section on the left, we display the query information used for retrieval, the ground truth target image, and our generated proxy image features. 'Query' represents the input query image, 'Caption' represents the relative text, and in the 'Proxy' section, we show two generated imagined proxies. In the red section on the right, the top and bottom rows display the top-5 retrieval results enhanced by our imagined proxies and the baseline's top-5 retrieval results, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.699, + 0.483, + 0.82 + ], + "angle": 0, + "content": "average or weighted sum, however, in extreme cases, when we have a case with a 0.999 text similarity but only a 0.001 proxy similarity, and another case with both a 0.5 base similarity and a 0.5 proxy similarity, their averages are both 0.5, which could lead to retrieving overly extreme results. We thus propose a balanced metric for combining these results. Specifically, given the baseline similarity \\( S_{t} \\) and proxy similarity \\( S_{p} \\), the final similarity \\( S_{f} \\) for retrieval is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.833, + 0.483, + 0.851 + ], + "angle": 0, + "content": "\\[\nS _ {f} = \\lambda S _ {t} + (1 - \\lambda) S _ {b}, S _ {b} = S _ {t} * S _ {p}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\( S_{b} \\) represents the balanced similarity, \\( \\lambda \\) is a balancing parameter. We believe that the balance metric enables us to focus more on cases that perform well in both similarities." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.697, + 0.645, + 0.714 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this section, we demonstrate the effectiveness of our method. In Section 4.1, We will provide a detailed description of our experimental setup, including the baselines, evaluation metrics, and implementation details. In Section 4.2, we compare our method with other state-of-the-art approaches. We explore the improvements in retrieval performance by presenting some visual examples, highlighting the impact of our method in Section 4.3, and in Section 4.4, we conduct ablation studies to investigate the contribution of each component of our method, as well as the influence of various parameter settings." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3988" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.281, + 0.108 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.265 + ], + "angle": 0, + "content": "Datasets. We validate the performance of our method by combining them with several state-of-the-art methods on three datasets: CIRR [28], CIRCO [5], and FashionIQ [41]. CIRR contains 21,552 real images from the NLVR [36] dataset. CIRCO is built from the COCO 2017 unlabeled set [27], includes a validation set with 220 queries and a test set with 800 queries. FashionIQ [41] is a dataset dedicated to the fashion domain, organized into three distinct subcategories: Dress, Shirt, and Toptee. It contains 30,135 query triplets and 77,683 images available for retrieval." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.266, + 0.484, + 0.34 + ], + "angle": 0, + "content": "Baseline. We selected LDRE[43] and LinCIR[16] as our baselines. For each query, we generate 5 proxy images and obtain the enhanced retrieval results. We compare the improved results with some of the most recent ZS-CIR results, such as Searle[6], CIREVL[20]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.484, + 0.476 + ], + "angle": 0, + "content": "Evaluation Metrics. In CIRCO, each query has multiple target images, so we calculate mAP@k to offer a more detailed assessment, where \\( k \\in \\{5,10,25,50\\} \\) indicates the number of top-ranked retrieval results. For CIRR, we follow the original benchmarks and use Recall@k (with \\( k \\in \\{1,5,10,50\\} \\)) as the primary metric. We also assess performance in a subset setting, which we denote as RecallSubset@K (with \\( k \\in \\{1,2,3\\} \\)). On FashionIQ, we also use Recall@K (with \\( k \\in \\{10,50\\} \\)) for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.627 + ], + "angle": 0, + "content": "Implement Details. We primarily conducted experiments using two backbone models: CLIP-L and CLIP-G. We used Qwen1.5-32B[1] as the LLM in all the experiments, which yielded results comparable to those reported in the original papers. For generating proxy images, we use MIGC[48] with SD1.5 as the backbone. We use Qwen1.5-32B as our LLM to generate the layout for each query and then generate 5 proxy images based on the query image and layout. All the experiments are conducted by Pytorch with one NVIDIA A6000 GPU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.285, + 0.652 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.764 + ], + "angle": 0, + "content": "CIRCO. We present the experimental results on the CIRCO test dataset on LDRE[43] and LinCIR[16]. In the left half of Tab.1, we observe that our proxy image can improve the retrieval performance, which an approximately 5-point improvement in mAP at \\( k = 5 \\) in LinCIR, and about 2-point improvement in LDRE. These results effectively demonstrate the validity and impact of our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.854 + ], + "angle": 0, + "content": "CIRR. In the right half of Tab.1, we present the experimental results on the CIRR test dataset. Our proxy image significantly improved the LDRE and LinCIR performance. The performance improvements across different methods and backbones indicate that our approach can be applied to various text-based retrieval methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "FashionIQ. In Tab.2, we present the results on the FashionIQ val dataset. Since LDRE has not open-sourced the code for the FashionIQ, we used LinCIR as the baseline." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.71, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.091, + 0.905, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.274, + 0.71, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.275, + 0.906, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.363, + 0.907, + 0.461 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of our method. We present the result on CIRCO validation dataset with top-1 retrieval results. Baseline methods in text-based retrieval overlook certain attributes and details in complex related texts, resulting in suboptimal retrieval outcomes. In contrast, our approach can construct suitable proxies on the image side, supplementing this missing information and yielding improved retrieval results." + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.473, + 0.877, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.68, + 0.907, + 0.723 + ], + "angle": 0, + "content": "Figure 5. Ablation results on the composition of robust proxy features in the CIRCO dataset. \\( pi \\) indicates proxy features, \\( si \\) represents semantic perturbation, and \\( qi \\) indicates the query features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.907, + 0.79 + ], + "angle": 0, + "content": "Although we observe a drop in \\( \\mathbb{R}@\\mathbb{10} \\) in the Toptee, our method improves the retrieval accuracy in the Shirt and dress Subset, thus improving the average performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.801, + 0.696, + 0.816 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We conducted experiments on the CIRCO validation dataset to verify the improvements our method brings to retrieval. Experiment setting. Specifically, we used LDRE as the baseline and generated 5 imagined proxy images for each query. We present two types of results. First, based on" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3989" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.865, + 0.274 + ], + "angle": 0, + "content": "
BenchmarkCIRRCIRCO
MetricRecall@KmAP@K
BackbonePIRPBMk=1k=2k=5k=10k=50k=5k=10k=25k=50
LDRE-G✓36.1549.4966.3977.2593.9531.1232.2434.9536.03
36.0250.0266.5177.5293.7431.0632.4135.1936.35
-0.13+0.53+0.12+0.27-0.21-0.06+0.17+0.24+0.32
✓✓38.1552.5368.5579.4794.4831.9233.5936.2637.37
+2.00+3.04+2.16+2.22+0.53+0.80+1.35+1.31+1.34
✓✓✓39.2552.9470.0780.0094.8932.7534.2636.8638.03
+3.10+3.45+3.68+2.75+0.94+1.63+2.02+1.91+2.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.338, + 0.274, + 0.66, + 0.286 + ], + "angle": 0, + "content": "Table 3. Ablation results in CIRR and CIRCO dataset." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.302, + 0.87, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.589, + 0.854, + 0.605 + ], + "angle": 0, + "content": "Figure 6. Visualize of Ablation Result. We present the top-1 retrieval result on CIRCO dataset in ablation experiments." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.618, + 0.483, + 0.695 + ], + "angle": 0, + "content": "the ground truth for each query, we show in Fig.3 that our method improves the accuracy of ground truth occurrence in top-5 retrieval results. Second, disregarding the ground truth image, we demonstrate in Fig.4 that our method retrieves a more suitable image in the top-1 result." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.704, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Analysis of Experimental Results on TOP-5 Retrieval. From the visualized results in Fig.3, we can draw the following conclusions: (1) Our generated proxy images effectively preserve the style and key elements of the original image. In the first row, our image maintains the clock tower as a significant element and adheres to the related caption by generating a Ferris wheel. In the second row, our image retains the original black-and-white style. (2) The imagined proxies complement important information that may be overlooked by text alone. For instance, in the first-row example, the clock tower—a crucial concept—is missing in the text-only retrieval result, while our imagined proxy and the original image both provide this retrieval information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.907, + 0.648 + ], + "angle": 0, + "content": "In the third row, our constructed image improves retrieval accuracy by incorporating an umbrella and snow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.801 + ], + "angle": 0, + "content": "Analysis of Experimental Results on TOP-1 Retrieval. Cases in Fig.4 show that by providing proxy image information, our approach achieves more relevant retrievals. Retrieval based solely on text space may only capture partial information in complex contexts, leading to the omission of certain details in the images—such as the person in the first row, the tie in the second row, and the White House in the background of the third row. Our method incorporates these details in the images, supplementing and balancing this information to yield more suitable results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.667, + 0.832 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Effects of the Proxy Image Features. We first conduct an ablation study on the proposed proxy image features in the CIRR dataset and CIRCO dataset. We use PI (Proxy Image) refers to the approach that directly computes retrieval" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3990" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.101, + 0.078, + 0.48, + 0.218 + ], + "angle": 0, + "content": "
MetricmAP@K
Backbonek=5k=10k=25k=50
LDRE-G31.1232.2434.9536.03
LDRE-G + SD32.2533.5036.1837.30
LDRE-G + FLUX32.3833.8936.5137.58
LDRE-G + ip2p31.8233.2736.1137.23
LDRE-G + magicbrush31.8633.3936.0537.22
LDRE-G + L2I32.7534.2636.8638.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.219, + 0.482, + 0.247 + ], + "angle": 0, + "content": "Table 4. Comparison of Various Methods for Generating Proxy Images on the CIRCO test" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.248, + 0.483, + 0.504 + ], + "angle": 0, + "content": "similarity from proxy image features, and then performs a weighted sum with the original similarity. RP (Robust Proxy) indicates integrating the proxy image into more robust proxy features. BM (Balance Metric) refers to using the balance metric. The results from the CIRR and CIRCO are shown in Tab.3. We can draw the following conclusions: 1) The imagined proxy images may contain some noise, leading to the retrieval focus on irrelevant details. we observe a decrease in recall at \\( k = 1 \\) and 50 in CIRR, and mAP \\( k = 5 \\) decreases in CIRCO. 2) After integrating the proxy image into a more robust proxy feature, we see a significant improvement in both datasets, which demonstrates the effectiveness of our robust proxy features. 3) By introducing the balance metric between proxy and text similarity, we further improve the retrieval performance. This suggests that combining the information from both image and text domains leads to more accurate retrieval results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.688 + ], + "angle": 0, + "content": "We can draw similar conclusions from Fig.6, which shows retrieval results on the CIRCO test dataset. In the first row, the proxy image provides a realistic and accurate depiction of the target image, allowing for improved retrieval results. In the second row, relying solely on the query image may cause the retrieval to overly focus on the beach background, contradicting the \"no people\" text information. Here, using the robust proxy \\((\\mathrm{PI} + \\mathrm{RP})\\) further enhances the overlooked aspects. In the third row, the balance metric balances the 'the woman with sunglasses' in proxy features with 'trees' in baseline features, achieving better retrieval results for both the subject and the background." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Ablation on Robust Features. We analyze the composition of our robust proxy on the CIRCO dataset, with results shown in Fig.5. The robust proxy is composed of three main components: \\( q_{i} \\) represents the query image features, \\( s_{i} \\) represents the semantic perturbation, and \\( p_{i} \\) represents our proxy image. A comparison among ①, ②, ③, and ④ shows that directly using our proxy image features can enhance retrieval performance to a certain extent, whereas relying solely on the query image or semantic perturbation tends to focus excessively on incorrect information, leading to a decline in performance. Additionally, comparisons among ⑤, ⑥, and ⑦ demonstrate that combining all three features together improves retrieval accuracy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Approaches for constructing proxy images. We also" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.1, + 0.833, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.26, + 0.907, + 0.289 + ], + "angle": 0, + "content": "Figure 7. Parameter Analysis. The effect of varying numbers of proxy images on performance for the CIRCO dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.907, + 0.483 + ], + "angle": 0, + "content": "tested the results obtained on the CIRCO dataset using different methods to construct the proxy image in Tab.4. Using the target image caption inferred by the LLM, we tested two T2I generation approaches: SD[33] and FLUX[22]. Additionally, based on the relative caption and query image, we also tested two instruction-based editing approaches Instruct-Pix2Pix[7] and MagicBrush[44]. The results show that: 1) The robust features constructed through these approaches can all provide additional information for retrieval. 2) The stronger the control capability, the more precise the provided information, and the greater the improvement achieved." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.909, + 0.593 + ], + "angle": 0, + "content": "Effects of the number of proxy images. We tested the impact of using different numbers of proxy images on performance with the CIRCO dataset. As shown in Fig.7, retrieval performance improves as the number of proxy images increases, though the rate of improvement gradually slows down. Therefore, the number of proxy images can be chosen based on a balance between efficiency and accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.634, + 0.62 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.84 + ], + "angle": 0, + "content": "In this paper, we propose the IP-CIR, a plug-and-play training-free method for current ZSCIR tasks. Our method imagines proxy images, which align with the query image and the relative captions, providing fine-grained details that may be overlooked by the text. We combine the proxy images, query images, and semantic perturbation into more robust proxy image features, and propose a balance metric to compose the proxy and baseline similarities, enhancing the retrieval performance. The experiment results in CIRR, CIRCO, and FashionIQ datasets show that our method can successfully improve retrieval accuracy and demonstrate the potential for providing additional retrieval information from the image side, revealing a new direction for improving retrieval accuracy in the future." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (62293554) and in part by the Natural Science Foundation of Zhejiang Province (LDT23F02023F02)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.958 + ], + "angle": 0, + "content": "3991" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.284, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.369, + 0.483, + 0.438 + ], + "angle": 0, + "content": "[3] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21466-21474, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.441, + 0.483, + 0.51 + ], + "angle": 0, + "content": "[4] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4959-4968, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.512, + 0.483, + 0.553 + ], + "angle": 0, + "content": "[5] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.556, + 0.483, + 0.624 + ], + "angle": 0, + "content": "[6] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15338-15347, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.627, + 0.483, + 0.669 + ], + "angle": 0, + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.671, + 0.483, + 0.724 + ], + "angle": 0, + "content": "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.728, + 0.483, + 0.782 + ], + "angle": 0, + "content": "[9] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.786, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[10] Yanbei Chen and Loris Bazzani. Learning joint visual semantic matching embeddings for language-guided retrieval. page 136-152, Berlin, Heidelberg, 2020. Springer-Verlag. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.483, + 0.885 + ], + "angle": 0, + "content": "[11] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2998-3008, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Yiyang Chen, Zhedong Zheng, Wei Ji, Leigang Qu, and Tat" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Seng Chua. Composed image retrieval with text feedback via multi-grained uncertainty regularization, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.149 + ], + "angle": 0, + "content": "[13] Sanghyuk Chun. Improved probabilistic image-text representations, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[14] Ginger Delmas, Rafael Sampaio de Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[15] Eric Dodds, Jack Culpepper, Simao Herdade, Yang Zhang, and Kofi Boakye. Modality-agnostic attention fusion for visual search with text feedback, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.251, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[16] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, , Yoohoon Kang, and Sangdoo Yun. Language-only training of zero-shot composed image retrieval. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[17] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoonhoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. Transactions on Machine Learning Research, 2024. Expert Certification. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[18] Ryota Hinami and Shin'ichi Satoh. Discriminative learning of open-vocabulary object retrieval and localization by negative phrase augmentation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2605-2615, Brussels, Belgium, 2018. Association for Computational Linguistics. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.462, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[20] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. International Conference on Learning Representations (ICLR), 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[21] Jooyeon Kim, Eulrang Cho, Sehyung Kim, and Hyunwoo J. Kim. Retrieval-augmented open-vocabulary object detection, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[22] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[24] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Glagen: Open-set grounded text-to-image generation. CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[25] You Li, Fan Ma, and Yi Yang. Anysynth: Harnessing the power of image synthetic data generation for generalized vision-language tasks. arXiv preprint arXiv:2411.16749, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[26] Chao Liang, Fan Ma, Linchao Zhu, Yingying Deng, and Yi Yang. Caphuman: Capture your moments in parallel universes. In CVPR, pages 6400-6409, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3992" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[28] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre-trained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2125–2134, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[29] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.263, + 0.397, + 0.276 + ], + "angle": 0, + "content": "[31] OpenAI. Gpt-4 technical report, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.279, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.349, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.391, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[34] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.448, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.491, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[36] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.534, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[37] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.577, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[38] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.62, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[39] Zheng Wang, Xin Yuan, Toshihiko Yamasaki, Yutian Lin, Xin Xu, and Wenjun Zeng. Re-identification = retrieval + verification: Back to essence and forward with a new metric. arXiv preprint arXiv:2011.11506, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.676, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[40] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[41] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The fashion iq dataset: Retrieving images by combining side information and relative natural language feedback. CVPR, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[42] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Reco: Region-controlled text-to-image generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[43] Zhenyu Yang, Dizhan Xue, Shengsheng Qian, Weiming Dong, and Changsheng Xu. Ldre: Llm-based divergent reasoning and ensemble for zero-shot composed image retrieval. In Proceedings of the 47th International ACM SI-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "GIR Conference on Research and Development in Information Retrieval, pages 80-90, 2024. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[44] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[45] Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. MagicLens: Self-supervised image retrieval with open-ended instructions. In Proceedings of the 41st International Conference on Machine Learning, pages 59403–59420. PMLR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[46] Liang Zheng, Yi Yang, and Alexander G. Hauptmann. Person re-identification: Past, present and future, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[47] Dewei Zhou, Zongxin Yang, and Yi Yang. Pyramid diffusion models for low-light image enhancement. arXiv preprint arXiv:2305.10028, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[48] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6818-6828, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[49] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Migc++: Advanced multi-instance generation controller for image synthesis. ArXiv, abs/2407.02329, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[50] Dewei Zhou, Ji Xie, Zongxin Yang, and Yi Yang. 3dis: Depth-driven decoupled instance synthesis for text-to-image generation. ArXiv, abs/2410.12669, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[51] Dewei Zhou, Mingwei Li, Zongxin Yang, and Yi Yang. Dreamrenderer: Taming multi-instance attribute control in large-scale text-to-image models, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[52] Dewei Zhou, Ji Xie, Zongxin Yang, and Yi Yang. 3dis-flux: simple and efficient multi-instance generation with dit rendering, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3993" + } + ] +] \ No newline at end of file diff --git a/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/f227006c-210b-4aac-822f-93d2a98acc36_origin.pdf b/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/f227006c-210b-4aac-822f-93d2a98acc36_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0a89cbdf96772991d52f30daf94be672b078dd5e --- /dev/null +++ b/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/f227006c-210b-4aac-822f-93d2a98acc36_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03f9b18120f021915405afd5efc03758bcd1e2729d36136a0ef0e4642c14ba46 +size 2457503 diff --git a/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/full.md b/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/full.md new file mode 100644 index 0000000000000000000000000000000000000000..53128ac013b02010c8eb41a2150b279ac45cf90a --- /dev/null +++ b/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/full.md @@ -0,0 +1,262 @@ +# Imagine and Seek: Improving Composed Image Retrieval with an Imagined Proxy + +You Li Fan Ma Yi Yang + +ReLER, CCAI, Zhejiang University, Zhejiang, China + +† Corresponding author + +{uli2000, mafan, yangyics}@zju.edu.cn + +# Abstract + +The Zero-shot Composed Image Retrieval (ZSCIR) requires retrieving images that match the query image and the relative captions. Current methods focus on projecting the query image into the text feature space, subsequently combining them with features of query texts for retrieval. However, retrieving images only with the text features cannot guarantee detailed alignment due to the natural gap between images and text. In this paper, we introduce Imagined Proxy for CIR (IP-CIR), a training-free method that creates a proxy image aligned with the query image and text description, enhancing query representation in the retrieval process. We first leverage the large language model's generalization capability to generate an image layout, and then apply both the query text and image for conditional generation. The robust query features are enhanced by merging the proxy image, query image, and text semantic perturbation. Our newly proposed balancing metric integrates text-based and proxy retrieval similarities, allowing for more accurate retrieval of the target image while incorporating image-side information into the process. Experiments on three public datasets demonstrate that our method significantly improves retrieval performances. We achieve state-of-the-art (SOTA) results on the CIRR dataset with a Recall@K of 70.07 at $K = 10$ . Additionally, we achieved an improvement in Recall@10 on the FashionIQ dataset, rising from 45.11 to 45.74, and improved the baseline performance in CIRCO with a mAPK@10 score, increasing from 32.24 to 34.26. + +# 1. Introduction + +Imagination is the wellspring of human creativity. When you see a picture of an adorable cat, you might envision it playing with other cats, wearing cool sunglasses and a cape, or floating in a space capsule. This ability to imagine and pursue beautiful visions drives continuous advancement in fields like image generation [2, 19, 26, 30, 33, 35, 47, 50, + +![](images/cb2a030492fa13ecac691f8e86efbdd6e4da0aa8b7443c5315e45567dd241dfe.jpg) +Figure 1. Comparison of composed image retrieval between text-only retrieval (LDRE in Fig) and our methods. Traditional methods perform retrieval only in the text space, where CLIP text features may overlook some important semantic information. In contrast, our approach generates imagined proxy features, providing additional information that is often overlooked in text-only retrieval, thereby improving retrieval accuracy. + +51] and image retrieval [5, 6, 16, 25, 28, 34, 37, 41, 43, 45]. + +This rich imagination gains new significance in Composed Image Retrieval (CIR) [28], where users express their imagination of the query image through text, searching for content that aligns with their preferences. Specifically, the CIR task requires users to retrieve target images from a database that match both the content of a query image and relative caption. With text modality to flexibly control retrieval targets, CIR enables the search process akin to the user's specific imagination after viewing the query image. This multi-modal, imagination-like retrieval task greatly expands the applications of image retrieval, with wide-ranging uses in fields such as visual search [29], object localization [18, 21], and re-identification [39, 46]. + +Traditional solutions for CIR [3, 4, 10-15] are constrained by the scarcity of training data (e.g., 46.6k for FashionIQ [41] dataset and 28.8k for CIRR [28] dataset), as collecting suitable triplets (query image $x_{q}$ , relevant text $x_{c}$ , and target image $x_{t}$ ) from the web incurs high labor costs. Therefore, we focus on the ZS-CIR task, aiming to bypass the need for training on triplet datasets and achieve gen + +eralized CIR retrieval capabilities. As shown in Fig.1 (a), traditional ZS-CIR methods typically employ lightweight projection models to map CLIP's [32] image features into the text features [6] for retrieval. The transformation inevitably loses certain image information, and the coarse, easily confusable text features fall short in fine-grained, complex imagination scenes. Although recent approaches attempt to leverage large language models (LLMs) [1, 31] to generate various descriptions of target images [20, 43, 45], they overlook the potential for direct imagination on the image side. As shown in Fig.1 (a), the text features did not effectively capture the attribute 'golden', resulting in suboptimal retrieval results. + +The rapid advancement of controllable generative models [8, 9, 24, 38, 42, 49, 52] has empowered users with the freedom to create images that align with their imagined visions. Generating images that match both the content of query images and relative captions aligns closely with the goals of composed retrieval tasks, and the imagined images (we called Retrieval Proxy) could provide additional information—such as style, instance attributes, and spatial relationships—that is often overlooked in text-based retrieval with complex captions. This raises an intriguing question: How could we harness the imaginative power of generative models to improve retrieval performance? + +We thus propose IP-CIR (Imagined Proxy for CIR), a training-free, plug-and-play method that harnesses the power of imagination for any retrieval method. Firstly, by utilizing LLMs to understand both the image content and the relative captions, we generate a suitable image layout. Using controllable generation methods, we create high-quality, fine-grained proxy images. Next, we compose the proxy images, query images as well as semantic perturbation into robust proxy features that are more suitable for retrieval. Finally, we propose a balance metric to combine the proxy and text retrieval similarities, allowing the integration of our method with any retrieval methods. + +We conduct experiments on three public datasets: the CIRR [28], CIRCO [5], and FashionIQ [41]. We combine our method with several baselines, the observed performance improvements confirm that our method supplements the retrieval process with valuable information and could enhance the performance of any other retrieval methods. + +Our contributions are summarized as follows: + +1) We propose using generative models' imagination capabilities to create proxy images that align with the content of the query image and relative captions to improve retrieval performance. +2) We introduce a proxy image generation framework. Using the LLM-generated object layout and the query image, we create imagined proxies through MIGC++. +3) We propose integrating proxy images with the query image and semantic perturbation into more robust proxy fea + +tures and introducing a balance metric to combine the proxy and baseline similarities. + +4) We conducted experiments on three public CIR datasets. The improvement over baseline results, as well as achieving SOTA results demonstrates the potential of constructing imagined proxies in retrieval. + +# 2. Related work + +# 2.1. Zero-shot Composed Image Retrieval + +Early work on ZS-CIR, such as Pic2word [34], SEALRE [6], attempted to project the original image into the text space using lightweight projection models. This was often done within fixed templates (e.g., "a photo of \([\\) ]$ with cond"), the \([\\( ]\) symbol is a textual notation used to represent query image information, and "cond" represents the relative captions. LinCIR [16] further defined keywords as adjectives and nouns in the text, thus overcoming the limitations of fixed templates and training more powerful projection models. Besides, CompoDiff[17] trains a new diffusion model that performs diffusion on the CLIP features of the image, introduces relative caption features, and applies attention operations with intermediate image features, ultimately obtaining the target features for retrieval. In addition, some methods have embraced LLMs [1, 31]. CIREVL [20] proposed generating captions for reference images and then using LLMs to reassemble them into target texts modified by relative captions for retrieval. LDRE [43] took this further by using LLMs to construct various edited captions and combining them based on similarity, addressing the issue of fuzzy retrieval. While these methods improved retrieval accuracy, they all overlooked the potential of directly imagining proxy images to provide additional information. + +# 2.2. Controllable Text-to-Image Generation + +In recent years, with the development of diffusion models [19, 35], high-quality images that meet specific requirements can now be generated. Using mainstream generators such as Stable Diffusion [33], users can generate images aligned with text descriptions. Although these foundational generators offer a certain level of control, their performance in fine-grained attribute and orientation control is limited, primarily due to the attribute ambiguity in the CLIP features of the text [48]. As a result, some efforts have focused on leveraging layout information to enhance control. For example, GLIGEN [24] introduces Gated-Self-Attention to combine positional feature information, thereby improving control over positioning. InstanceDiffusion [38] further improves attribute control on this foundation. Meanwhile, MIGC [48] adopts a divide-and-conquer approach, breaking down complex instance generation into multiple single-instance generations and incorporating Enhancement Atten + +![](images/fd55235f2763169362301aaa30e2eb484755851213712e0754860073e8ee8be9.jpg) +(a) Imagined Retrieval Proxy Generation + +![](images/99f2b1d15938947dd08754e4771b0cb8b3ebb96761a6e7c1d2ef1f0a8a7dd6c9.jpg) +(b) Balanced Proxy Retrieval +Figure 2. Overview of our method. (a) represents our imagined proxy generation process. We use LLM to analyze the BLIP2-generated query image captions and the relative captions and infer the proxy layout. We then use the controllable generator to imagine the proxy images. (b) represents our process of constructing a robust proxy feature, and balancing the text and proxy similarities. We integrate proxy features, query image features as well as semantic perturbations into a robust proxy feature, and propose a balance metric for retrieval. + +tion and Layout Attention to improve both positional and attribute control. MIGC++ [49] further integrates ELITE [40] to transform images into CLIP-based text features and introduces an additional EA layer to process information from the reference image modality, thereby incorporating image modality control in the generation process. In this paper, we aim to utilize the fine-grained control capabilities of MIGC to display certain attributes and other information that are difficult to clearly express in text at a fine-grained level within the image itself. This provides supplementary information during the retrieval process. + +# 3. Method + +In this section, we will introduce our IP-CIR, which attempts to imagine an appropriate proxy image within the image feature space, balance, and combine the retrieval results from the image side with those from the text side. Specifically, in section 3.1, we introduce the overview of our method, in section 3.2, we will describe how to construct an imagined retrieval proxy for each query. In section 3.3, we will further explain how to organize and integrate the information from the imagined retrieval proxy and present how to combine the retrieval information from both the image and text feature spaces. + +# 3.1. Overview + +The overall structure of our framework is shown in Fig.2. First, we generate the imagined proxy image for retrieval. The proxy image needs to align with the content of the query image and be as consistent as possible with the description in the relative captions. Therefore, we first generate a reasonable object layout and then use a controllable generation model to imagine this layout as a concrete image. Specifically, we use BLIP2[23] to generate a set of captions for each query. Combine with the relative captions, we instruct the LLM to reason out a suitable + +spatial layout of the target image, so that, with as much accuracy as possible in terms of scene, instance attributes, and quantity with relative caption, we can imagine a proxy image of the retrieval target to provide additional information for the retrieval process. In Fig.2 (b), we compose the imagined proxy with query image and text features, forming a robust retrieval proxy. The similarity of robust proxy is further balanced with other baselines' similarity, improving the retrieval accuracy. + +# 3.2. Imagined Retrieval Proxy Generation + +Text retrieval features may struggle to precisely capture important semantics in the relevant text and the query image, especially when the caption is complex. Generating an image that meets the retrieval requirements, which we refer to as imagining the retrieval proxy, could provide more valuable details to the retrieval process. + +Imagine Proxy Layout. The first step of imagining a suitable proxy relies on envisioning the overall layout, which requires a clear understanding of the semantic content and relationship between the query images and relative captions. We thus leverage the robust interpretive and reasoning abilities of LLMs[1, 31] to generate a target proxy layout. As illustrated in Fig.2 (a), we organize the BLIP2 generated query image captions with the relative caption into the fixed form 'Given an image of{caption}, we show{rule}', based on which the LLM is instructed to infer the reasonable overall scene and the layout of the proxy image. The layout includes a detailed description of each instance as well as the bounding box (bbox) coordinates. Besides, certain instances may require detailed visual information from text or image modality. For instance, when retrieving an image of a dog wearing a hat, the attributes of the dog in the proxy image come from the dog in the query image, while the attributes of the hat come from the text modality information. We thus require the LLM to determine the reference modal + +
BenchmarkCIRCOCIRR
MetricmAP@KRecall@KRecall Subset@K
BackboneMethodk=5k=10k=25k=50k=1k=5k=10k=50k=1k=2k=3
ViT-L/14CIREVL18.5719.0120.8921.8024.5552.3164.9286.3459.5479.8889.69
LinCIR12.8513.7515.2616.1724.8253.0866.6088.8955.6176.2488.24
LinCIR + IP-CIR14.1615.2617.1518.0524.9653.7666.8288.8755.5276.5188.27
(vs. LinCIR)+1.31+1.51+1.89+1.88+0.14+0.68+0.22-0.02-0.09+0.27+0.03
LDRE23.3524.0326.4427.5026.5355.5767.5488.5060.4380.3189.90
LDRE + IP-CIR26.4327.4129.8731.0729.7658.8271.2190.4162.4881.6490.89
(vs. LDRE)+3.08+3.38+3.43+3.57+3.23+3.25+3.67+1.91+2.05+1.33+0.99
ViT-G/14CIREVL26.7727.5929.9631.0334.6564.2975.0691.6667.9584.8793.21
LinCIR20.3421.8523.9825.1434.7564.1275.9393.4262.3681.7891.28
LinCIR + IP-CIR25.7026.6429.0930.1335.3764.7076.1593.7162.5881.7491.35
(vs. LinCIR)+5.40+4.79+5.11+4.99+0.62+0.58+0.22+0.29+0.22-0.04+0.07
LDRE31.1232.2434.9536.0336.1566.3977.2593.9568.8285.6693.76
LDRE + IP-CIR32.7534.2636.8638.0339.2570.0780.0094.8969.9586.8794.22
(vs. LDRE)+1.63+2.02+1.91+2.00+3.10+3.68+2.75+0.94+1.13+1.21+0.46
+ +Table 1. Quantitative results in CIRCO and CIRR datasets. + +ity for each instance. + +Proxy Image Generation. The next step is to imagine the proxy based on the generated layout. Since the CIR task is a fuzzy retrieval task [43], this reduces the need to maintain the query image's instance ID during the generation process. So we follow the idea of MIGC++ [49], incorporating the query image features transformed by ELITE [40] into MIGC, thereby introducing some degree of the query image information into the Proxy image. Besides, an instance can be influenced by both the reference image and the relative caption. For example, if a user envisions a hat in the query image with a specific pattern, the hat should retain its original shape but include the target pattern. So, we copy the layout of the instance with reference image modality and render it based on both image and text, ensuring the proxy image reflects the text's modifications. + +# 3.3. Balanced Proxy Retrieval + +Once the proxy image is generated, the key issue lies in how to leverage this proxy image in the retrieval. A straightforward method is to compute the proxy similarity of the retrieval database, weighted sum with the text similarity. However, as shown in the second line of Fig.6, directly using the CLIP features of the proxy image (+PI) can lead to an excessive focus on certain elements, like the beach background. This may cause the results to overly emphasize the visual information in the image while neglecting some details from the text. So in this section, we propose to construct a robust retrieval proxy and balance the retrieval result from image and text. + +Constructing Robust Proxy. Inspired by LDRE [43], we + +use the LLM [1, 31] to infer a set of target captions, then obtain its CLIP features $f_{t}$ . We derive the semantic perturbation features $f_{s} = f_{t} - f_{o}$ with $f_{t}$ and the CLIP features of BLIP2 [23] generated query image captions $f_{o}$ . The difference of CLIP text features $f_{s}$ represents a semantic transformation direction from the query image to the target retrieval image, which aligns with the transformation direction in the image space. So the introduction of semantic perturbation provides useful information in the proxy images. Besides, The CLIP features of the original image also contain some basic information relevant to the current query. As shown in Fig.2 (b), we combine the query image's CLIP features $f_{q}$ , the proxy CLIP feature $f_{p}$ with the semantic perturbation $f_{s}$ into the robust proxy $f_{RP}$ . There are various ways to combine retrieval features. A simple approach is to scale all features to the same magnitude using max normalization as shown in Eq.1. Besides, the weights of each component can be adjusted to control retrieval bias based on the quality of the proxy image and specific retrieval goals. For example, assigning a higher weight to query features ensures a closer match to the query image if the user prioritizes similarity to the original query. + +$$ +f _ {R P} = f _ {p} + \frac {\operatorname* {m a x} \left(f _ {p}\right)}{\operatorname* {m a x} \left(f _ {q}\right)} f _ {q} + \frac {\operatorname* {m a x} \left(f _ {p}\right)}{\operatorname* {m a x} \left(f _ {s}\right)} f _ {s}. \tag {1} +$$ + +Balancing Retrieval Results. After obtaining the retrieval feature in text and image spaces, a critical question arises: How should we balance the two retrieval results? The proxy features we obtain exist in the image feature space, making it difficult to directly perform weighted summation with the baseline text features. Therefore, we choose to combine their similarities. A simple approach is to take the + +
BackboneTypeShirtDressTopteeAverage
MethodR@10R@50R@10R@50R@10R@50R@10R@50
ViT-G/14Pic2Word33.1750.3925.4347.6535.2457.6231.2851.89
SEARLE36.4655.3528.1650.3239.8361.4534.8155.71
LinCIR46.7665.1138.0860.8850.4871.0945.1165.69
LinCIR + IP-CIR48.0466.6839.0261.0350.1871.1445.7466.28
(vs. LinCIR)+1.28+1.57+0.94+0.15-0.30+0.05+0.63+0.59
+ +Table 2. Quantitative results in FashionIQ dataset. + +![](images/34f3db9c5438476110a026369fa1df8158b193385d96273578d83e9ca801aa02.jpg) +Figure 3. Qualitative results of our method. We conducted experiments on the CIRCO validation dataset to observe in which cases our method improves retrieval results. In the blue section on the left, we display the query information used for retrieval, the ground truth target image, and our generated proxy image features. 'Query' represents the input query image, 'Caption' represents the relative text, and in the 'Proxy' section, we show two generated imagined proxies. In the red section on the right, the top and bottom rows display the top-5 retrieval results enhanced by our imagined proxies and the baseline's top-5 retrieval results, respectively. + +![](images/7d7e5c5acf2266acebe63a751d65109f0b5969130b50f55ba378e9f4f8c5b3d7.jpg) + +average or weighted sum, however, in extreme cases, when we have a case with a 0.999 text similarity but only a 0.001 proxy similarity, and another case with both a 0.5 base similarity and a 0.5 proxy similarity, their averages are both 0.5, which could lead to retrieving overly extreme results. We thus propose a balanced metric for combining these results. Specifically, given the baseline similarity $S_{t}$ and proxy similarity $S_{p}$ , the final similarity $S_{f}$ for retrieval is as follows: + +$$ +S _ {f} = \lambda S _ {t} + (1 - \lambda) S _ {b}, S _ {b} = S _ {t} * S _ {p}, \tag {2} +$$ + +where $S_{b}$ represents the balanced similarity, $\lambda$ is a balancing parameter. We believe that the balance metric enables us to focus more on cases that perform well in both similarities. + +# 4. Experiments + +In this section, we demonstrate the effectiveness of our method. In Section 4.1, We will provide a detailed description of our experimental setup, including the baselines, evaluation metrics, and implementation details. In Section 4.2, we compare our method with other state-of-the-art approaches. We explore the improvements in retrieval performance by presenting some visual examples, highlighting the impact of our method in Section 4.3, and in Section 4.4, we conduct ablation studies to investigate the contribution of each component of our method, as well as the influence of various parameter settings. + +# 4.1. Experimental Setup + +Datasets. We validate the performance of our method by combining them with several state-of-the-art methods on three datasets: CIRR [28], CIRCO [5], and FashionIQ [41]. CIRR contains 21,552 real images from the NLVR [36] dataset. CIRCO is built from the COCO 2017 unlabeled set [27], includes a validation set with 220 queries and a test set with 800 queries. FashionIQ [41] is a dataset dedicated to the fashion domain, organized into three distinct subcategories: Dress, Shirt, and Toptee. It contains 30,135 query triplets and 77,683 images available for retrieval. + +Baseline. We selected LDRE[43] and LinCIR[16] as our baselines. For each query, we generate 5 proxy images and obtain the enhanced retrieval results. We compare the improved results with some of the most recent ZS-CIR results, such as Searle[6], CIREVL[20]. + +Evaluation Metrics. In CIRCO, each query has multiple target images, so we calculate mAP@k to offer a more detailed assessment, where $k \in \{5,10,25,50\}$ indicates the number of top-ranked retrieval results. For CIRR, we follow the original benchmarks and use Recall@k (with $k \in \{1,5,10,50\}$ ) as the primary metric. We also assess performance in a subset setting, which we denote as RecallSubset@K (with $k \in \{1,2,3\}$ ). On FashionIQ, we also use Recall@K (with $k \in \{10,50\}$ ) for evaluation. + +Implement Details. We primarily conducted experiments using two backbone models: CLIP-L and CLIP-G. We used Qwen1.5-32B[1] as the LLM in all the experiments, which yielded results comparable to those reported in the original papers. For generating proxy images, we use MIGC[48] with SD1.5 as the backbone. We use Qwen1.5-32B as our LLM to generate the layout for each query and then generate 5 proxy images based on the query image and layout. All the experiments are conducted by Pytorch with one NVIDIA A6000 GPU. + +# 4.2. Quantitative Results + +CIRCO. We present the experimental results on the CIRCO test dataset on LDRE[43] and LinCIR[16]. In the left half of Tab.1, we observe that our proxy image can improve the retrieval performance, which an approximately 5-point improvement in mAP at $k = 5$ in LinCIR, and about 2-point improvement in LDRE. These results effectively demonstrate the validity and impact of our method. + +CIRR. In the right half of Tab.1, we present the experimental results on the CIRR test dataset. Our proxy image significantly improved the LDRE and LinCIR performance. The performance improvements across different methods and backbones indicate that our approach can be applied to various text-based retrieval methods. + +FashionIQ. In Tab.2, we present the results on the FashionIQ val dataset. Since LDRE has not open-sourced the code for the FashionIQ, we used LinCIR as the baseline. + +![](images/156d1b69e3359d67382688657255f16805d054d4417d846d571d58822cab6dba.jpg) + +![](images/b68559c739a8c03a95950baa441cefeb0805e558c5c5c5905bcc661c157953ff.jpg) + +![](images/a250bcd7faab5928edf80e6fb10a88e0aa32c436104ac78299da7652bf2b520a.jpg) +Figure 4. Qualitative results of our method. We present the result on CIRCO validation dataset with top-1 retrieval results. Baseline methods in text-based retrieval overlook certain attributes and details in complex related texts, resulting in suboptimal retrieval outcomes. In contrast, our approach can construct suitable proxies on the image side, supplementing this missing information and yielding improved retrieval results. + +![](images/4ff09d93475ebd6315086384d58bd87425ebd27590768ace9c858ce3c0b54ffc.jpg) + +![](images/f869bcbc0e4f53ab9c104de527212126d44b88adc91fa942006c97b7f6242ee3.jpg) +Figure 5. Ablation results on the composition of robust proxy features in the CIRCO dataset. $pi$ indicates proxy features, $si$ represents semantic perturbation, and $qi$ indicates the query features. + +Although we observe a drop in $\mathbb{R}@\mathbb{10}$ in the Toptee, our method improves the retrieval accuracy in the Shirt and dress Subset, thus improving the average performance. + +# 4.3. Qualitative Results + +We conducted experiments on the CIRCO validation dataset to verify the improvements our method brings to retrieval. Experiment setting. Specifically, we used LDRE as the baseline and generated 5 imagined proxy images for each query. We present two types of results. First, based on + +
BenchmarkCIRRCIRCO
MetricRecall@KmAP@K
BackbonePIRPBMk=1k=2k=5k=10k=50k=5k=10k=25k=50
LDRE-G✓36.1549.4966.3977.2593.9531.1232.2434.9536.03
36.0250.0266.5177.5293.7431.0632.4135.1936.35
-0.13+0.53+0.12+0.27-0.21-0.06+0.17+0.24+0.32
✓✓38.1552.5368.5579.4794.4831.9233.5936.2637.37
+2.00+3.04+2.16+2.22+0.53+0.80+1.35+1.31+1.34
✓✓✓39.2552.9470.0780.0094.8932.7534.2636.8638.03
+3.10+3.45+3.68+2.75+0.94+1.63+2.02+1.91+2.00
+ +Table 3. Ablation results in CIRR and CIRCO dataset. + +![](images/5e2716f24ff5d3e1c39219c9492f68a6a546a41ab6e32d7b7fc4035390455251.jpg) +Figure 6. Visualize of Ablation Result. We present the top-1 retrieval result on CIRCO dataset in ablation experiments. + +the ground truth for each query, we show in Fig.3 that our method improves the accuracy of ground truth occurrence in top-5 retrieval results. Second, disregarding the ground truth image, we demonstrate in Fig.4 that our method retrieves a more suitable image in the top-1 result. + +Analysis of Experimental Results on TOP-5 Retrieval. From the visualized results in Fig.3, we can draw the following conclusions: (1) Our generated proxy images effectively preserve the style and key elements of the original image. In the first row, our image maintains the clock tower as a significant element and adheres to the related caption by generating a Ferris wheel. In the second row, our image retains the original black-and-white style. (2) The imagined proxies complement important information that may be overlooked by text alone. For instance, in the first-row example, the clock tower—a crucial concept—is missing in the text-only retrieval result, while our imagined proxy and the original image both provide this retrieval information. + +In the third row, our constructed image improves retrieval accuracy by incorporating an umbrella and snow. + +Analysis of Experimental Results on TOP-1 Retrieval. Cases in Fig.4 show that by providing proxy image information, our approach achieves more relevant retrievals. Retrieval based solely on text space may only capture partial information in complex contexts, leading to the omission of certain details in the images—such as the person in the first row, the tie in the second row, and the White House in the background of the third row. Our method incorporates these details in the images, supplementing and balancing this information to yield more suitable results. + +# 4.4. Ablation Study + +Effects of the Proxy Image Features. We first conduct an ablation study on the proposed proxy image features in the CIRR dataset and CIRCO dataset. We use PI (Proxy Image) refers to the approach that directly computes retrieval + +
MetricmAP@K
Backbonek=5k=10k=25k=50
LDRE-G31.1232.2434.9536.03
LDRE-G + SD32.2533.5036.1837.30
LDRE-G + FLUX32.3833.8936.5137.58
LDRE-G + ip2p31.8233.2736.1137.23
LDRE-G + magicbrush31.8633.3936.0537.22
LDRE-G + L2I32.7534.2636.8638.03
+ +Table 4. Comparison of Various Methods for Generating Proxy Images on the CIRCO test + +similarity from proxy image features, and then performs a weighted sum with the original similarity. RP (Robust Proxy) indicates integrating the proxy image into more robust proxy features. BM (Balance Metric) refers to using the balance metric. The results from the CIRR and CIRCO are shown in Tab.3. We can draw the following conclusions: 1) The imagined proxy images may contain some noise, leading to the retrieval focus on irrelevant details. we observe a decrease in recall at $k = 1$ and 50 in CIRR, and mAP $k = 5$ decreases in CIRCO. 2) After integrating the proxy image into a more robust proxy feature, we see a significant improvement in both datasets, which demonstrates the effectiveness of our robust proxy features. 3) By introducing the balance metric between proxy and text similarity, we further improve the retrieval performance. This suggests that combining the information from both image and text domains leads to more accurate retrieval results. + +We can draw similar conclusions from Fig.6, which shows retrieval results on the CIRCO test dataset. In the first row, the proxy image provides a realistic and accurate depiction of the target image, allowing for improved retrieval results. In the second row, relying solely on the query image may cause the retrieval to overly focus on the beach background, contradicting the "no people" text information. Here, using the robust proxy $(\mathrm{PI} + \mathrm{RP})$ further enhances the overlooked aspects. In the third row, the balance metric balances the 'the woman with sunglasses' in proxy features with 'trees' in baseline features, achieving better retrieval results for both the subject and the background. + +Ablation on Robust Features. We analyze the composition of our robust proxy on the CIRCO dataset, with results shown in Fig.5. The robust proxy is composed of three main components: $q_{i}$ represents the query image features, $s_{i}$ represents the semantic perturbation, and $p_{i}$ represents our proxy image. A comparison among ①, ②, ③, and ④ shows that directly using our proxy image features can enhance retrieval performance to a certain extent, whereas relying solely on the query image or semantic perturbation tends to focus excessively on incorrect information, leading to a decline in performance. Additionally, comparisons among ⑤, ⑥, and ⑦ demonstrate that combining all three features together improves retrieval accuracy. + +Approaches for constructing proxy images. We also + +![](images/c31ae80af61a82718ca1b8ebfc44def5222ccd5b3006ba119a917be2ef18250d.jpg) +Figure 7. Parameter Analysis. The effect of varying numbers of proxy images on performance for the CIRCO dataset. + +tested the results obtained on the CIRCO dataset using different methods to construct the proxy image in Tab.4. Using the target image caption inferred by the LLM, we tested two T2I generation approaches: SD[33] and FLUX[22]. Additionally, based on the relative caption and query image, we also tested two instruction-based editing approaches Instruct-Pix2Pix[7] and MagicBrush[44]. The results show that: 1) The robust features constructed through these approaches can all provide additional information for retrieval. 2) The stronger the control capability, the more precise the provided information, and the greater the improvement achieved. + +Effects of the number of proxy images. We tested the impact of using different numbers of proxy images on performance with the CIRCO dataset. As shown in Fig.7, retrieval performance improves as the number of proxy images increases, though the rate of improvement gradually slows down. Therefore, the number of proxy images can be chosen based on a balance between efficiency and accuracy. + +# 5. Conclusion + +In this paper, we propose the IP-CIR, a plug-and-play training-free method for current ZSCIR tasks. Our method imagines proxy images, which align with the query image and the relative captions, providing fine-grained details that may be overlooked by the text. We combine the proxy images, query images, and semantic perturbation into more robust proxy image features, and propose a balance metric to compose the proxy and baseline similarities, enhancing the retrieval performance. The experiment results in CIRR, CIRCO, and FashionIQ datasets show that our method can successfully improve retrieval accuracy and demonstrate the potential for providing additional retrieval information from the image side, revealing a new direction for improving retrieval accuracy in the future. + +Acknowledgements. This work was supported by the National Natural Science Foundation of China (62293554) and in part by the Natural Science Foundation of Zhejiang Province (LDT23F02023F02). + +# References + +[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2, 3, 4, 6 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1 +[3] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21466-21474, 2022. 1 +[4] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4959-4968, 2022. 1 +[5] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion, 2023. 1, 2, 6 +[6] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15338-15347, 2023. 1, 2, 6 +[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 8 +[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models, 2023. 2 +[9] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481, 2023. 2 +[10] Yanbei Chen and Loris Bazzani. Learning joint visual semantic matching embeddings for language-guided retrieval. page 136-152, Berlin, Heidelberg, 2020. Springer-Verlag. 1 +[11] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2998-3008, 2020. +[12] Yiyang Chen, Zhedong Zheng, Wei Ji, Leigang Qu, and Tat + +Seng Chua. Composed image retrieval with text feedback via multi-grained uncertainty regularization, 2024. +[13] Sanghyuk Chun. Improved probabilistic image-text representations, 2024. +[14] Ginger Delmas, Rafael Sampaio de Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity, 2022. +[15] Eric Dodds, Jack Culpepper, Simao Herdade, Yang Zhang, and Kofi Boakye. Modality-agnostic attention fusion for visual search with text feedback, 2020. 1 +[16] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, , Yoohoon Kang, and Sangdoo Yun. Language-only training of zero-shot composed image retrieval. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 6 +[17] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoonhoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. Transactions on Machine Learning Research, 2024. Expert Certification. 2 +[18] Ryota Hinami and Shin'ichi Satoh. Discriminative learning of open-vocabulary object retrieval and localization by negative phrase augmentation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2605-2615, Brussels, Belgium, 2018. Association for Computational Linguistics. 1 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2 +[20] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. International Conference on Learning Representations (ICLR), 2024. 2, 6 +[21] Jooyeon Kim, Eulrang Cho, Sehyung Kim, and Hyunwoo J. Kim. Retrieval-augmented open-vocabulary object detection, 2024. 1 +[22] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.8 +[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3, 4 +[24] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Glagen: Open-set grounded text-to-image generation. CVPR, 2023. 2 +[25] You Li, Fan Ma, and Yi Yang. Anysynth: Harnessing the power of image synthetic data generation for generalized vision-language tasks. arXiv preprint arXiv:2411.16749, 2024. 1 +[26] Chao Liang, Fan Ma, Linchao Zhu, Yingying Deng, and Yi Yang. Caphuman: Capture your moments in parallel universes. In CVPR, pages 6400-6409, 2024. 1 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 6 + +[28] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre-trained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2125–2134, 2021. 1, 2, 6 +[29] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers, 2021. 1 +[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2022. 1 +[31] OpenAI. Gpt-4 technical report, 2023. 2, 3, 4 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 2, 8 +[34] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. CVPR, 2023. 1, 2 +[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 1, 2 +[36] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs, 2019. 6 +[37] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval, 2024. 1 +[38] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation, 2024. 2 +[39] Zheng Wang, Xin Yuan, Toshihiko Yamasaki, Yutian Lin, Xin Xu, and Wenjun Zeng. Re-identification = retrieval + verification: Back to essence and forward with a new metric. arXiv preprint arXiv:2011.11506, 2020. 1 +[40] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023. 3, 4 +[41] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The fashion iq dataset: Retrieving images by combining side information and relative natural language feedback. CVPR, 2021. 1, 2, 6 +[42] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Reco: Region-controlled text-to-image generation. In CVPR, 2023. 2 +[43] Zhenyu Yang, Dizhan Xue, Shengsheng Qian, Weiming Dong, and Changsheng Xu. Ldre: Llm-based divergent reasoning and ensemble for zero-shot composed image retrieval. In Proceedings of the 47th International ACM SI- + +GIR Conference on Research and Development in Information Retrieval, pages 80-90, 2024. 1, 2, 4, 6 +[44] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 8 +[45] Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. MagicLens: Self-supervised image retrieval with open-ended instructions. In Proceedings of the 41st International Conference on Machine Learning, pages 59403–59420. PMLR, 2024. 1, 2 +[46] Liang Zheng, Yi Yang, and Alexander G. Hauptmann. Person re-identification: Past, present and future, 2016. 1 +[47] Dewei Zhou, Zongxin Yang, and Yi Yang. Pyramid diffusion models for low-light image enhancement. arXiv preprint arXiv:2305.10028, 2023. 1 +[48] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6818-6828, 2024. 2, 6 +[49] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Migc++: Advanced multi-instance generation controller for image synthesis. ArXiv, abs/2407.02329, 2024. 2, 3, 4 +[50] Dewei Zhou, Ji Xie, Zongxin Yang, and Yi Yang. 3dis: Depth-driven decoupled instance synthesis for text-to-image generation. ArXiv, abs/2410.12669, 2024. 1 +[51] Dewei Zhou, Mingwei Li, Zongxin Yang, and Yi Yang. Dreamrenderer: Taming multi-instance attribute control in large-scale text-to-image models, 2025. 1 +[52] Dewei Zhou, Ji Xie, Zongxin Yang, and Yi Yang. 3dis-flux: simple and efficient multi-instance generation with dit rendering, 2025. 2 \ No newline at end of file diff --git a/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/images.zip b/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7ae7379fb5079aba8a9432c0e9b94ab3030e92c4 --- /dev/null +++ b/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2a71d1cd1907974432aa505327d69df39c040b8663b2e19bd437ad17a06c57d +size 827743 diff --git a/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/layout.json b/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..611066d8b5b2e2c55090235d26b8aaf047593068 --- /dev/null +++ b/2025/Imagine and Seek_ Improving Composed Image Retrieval with an Imagined Proxy/layout.json @@ -0,0 +1,6720 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "type": "text", + "content": "Imagine and Seek: Improving Composed Image Retrieval with an Imagined Proxy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 226, + 160, + 383, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 160, + 383, + 175 + ], + "spans": [ + { + "bbox": [ + 226, + 160, + 383, + 175 + ], + "type": "text", + "content": "You Li Fan Ma Yi Yang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 176, + 175, + 434, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 175, + 434, + 190 + ], + "spans": [ + { + "bbox": [ + 176, + 175, + 434, + 190 + ], + "type": "text", + "content": "ReLER, CCAI, Zhejiang University, Zhejiang, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 261, + 190, + 349, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 190, + 349, + 202 + ], + "spans": [ + { + "bbox": [ + 261, + 190, + 349, + 202 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 205, + 205, + 405, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 205, + 405, + 217 + ], + "spans": [ + { + "bbox": [ + 205, + 205, + 405, + 217 + ], + "type": "text", + "content": "{uli2000, mafan, yangyics}@zju.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 271, + 296, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 595 + ], + "type": "text", + "content": "The Zero-shot Composed Image Retrieval (ZSCIR) requires retrieving images that match the query image and the relative captions. Current methods focus on projecting the query image into the text feature space, subsequently combining them with features of query texts for retrieval. However, retrieving images only with the text features cannot guarantee detailed alignment due to the natural gap between images and text. In this paper, we introduce Imagined Proxy for CIR (IP-CIR), a training-free method that creates a proxy image aligned with the query image and text description, enhancing query representation in the retrieval process. We first leverage the large language model's generalization capability to generate an image layout, and then apply both the query text and image for conditional generation. The robust query features are enhanced by merging the proxy image, query image, and text semantic perturbation. Our newly proposed balancing metric integrates text-based and proxy retrieval similarities, allowing for more accurate retrieval of the target image while incorporating image-side information into the process. Experiments on three public datasets demonstrate that our method significantly improves retrieval performances. We achieve state-of-the-art (SOTA) results on the CIRR dataset with a Recall@K of 70.07 at " + }, + { + "bbox": [ + 55, + 271, + 296, + 595 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 55, + 271, + 296, + 595 + ], + "type": "text", + "content": ". Additionally, we achieved an improvement in Recall@10 on the FashionIQ dataset, rising from 45.11 to 45.74, and improved the baseline performance in CIRCO with a mAPK@10 score, increasing from 32.24 to 34.26." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 620, + 135, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 620, + 135, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 135, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Imagination is the wellspring of human creativity. When you see a picture of an adorable cat, you might envision it playing with other cats, wearing cool sunglasses and a cape, or floating in a space capsule. This ability to imagine and pursue beautiful visions drives continuous advancement in fields like image generation [2, 19, 26, 30, 33, 35, 47, 50," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 245, + 554, + 353 + ], + "blocks": [ + { + "bbox": [ + 317, + 245, + 554, + 353 + ], + "lines": [ + { + "bbox": [ + 317, + 245, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 245, + 554, + 353 + ], + "type": "image", + "image_path": "cb2a030492fa13ecac691f8e86efbdd6e4da0aa8b7443c5315e45567dd241dfe.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 363, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 440 + ], + "type": "text", + "content": "Figure 1. Comparison of composed image retrieval between text-only retrieval (LDRE in Fig) and our methods. Traditional methods perform retrieval only in the text space, where CLIP text features may overlook some important semantic information. In contrast, our approach generates imagined proxy features, providing additional information that is often overlooked in text-only retrieval, thereby improving retrieval accuracy." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 456, + 553, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 553, + 469 + ], + "type": "text", + "content": "51] and image retrieval [5, 6, 16, 25, 28, 34, 37, 41, 43, 45]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 472, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 555, + 628 + ], + "type": "text", + "content": "This rich imagination gains new significance in Composed Image Retrieval (CIR) [28], where users express their imagination of the query image through text, searching for content that aligns with their preferences. Specifically, the CIR task requires users to retrieve target images from a database that match both the content of a query image and relative caption. With text modality to flexibly control retrieval targets, CIR enables the search process akin to the user's specific imagination after viewing the query image. This multi-modal, imagination-like retrieval task greatly expands the applications of image retrieval, with wide-ranging uses in fields such as visual search [29], object localization [18, 21], and re-identification [39, 46]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Traditional solutions for CIR [3, 4, 10-15] are constrained by the scarcity of training data (e.g., 46.6k for FashionIQ [41] dataset and 28.8k for CIRR [28] dataset), as collecting suitable triplets (query image " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{q}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", relevant text " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", and target image " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ") from the web incurs high labor costs. Therefore, we focus on the ZS-CIR task, aiming to bypass the need for training on triplet datasets and achieve gen" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3984" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "eralized CIR retrieval capabilities. As shown in Fig.1 (a), traditional ZS-CIR methods typically employ lightweight projection models to map CLIP's [32] image features into the text features [6] for retrieval. The transformation inevitably loses certain image information, and the coarse, easily confusable text features fall short in fine-grained, complex imagination scenes. Although recent approaches attempt to leverage large language models (LLMs) [1, 31] to generate various descriptions of target images [20, 43, 45], they overlook the potential for direct imagination on the image side. As shown in Fig.1 (a), the text features did not effectively capture the attribute 'golden', resulting in suboptimal retrieval results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 373 + ], + "type": "text", + "content": "The rapid advancement of controllable generative models [8, 9, 24, 38, 42, 49, 52] has empowered users with the freedom to create images that align with their imagined visions. Generating images that match both the content of query images and relative captions aligns closely with the goals of composed retrieval tasks, and the imagined images (we called Retrieval Proxy) could provide additional information—such as style, instance attributes, and spatial relationships—that is often overlooked in text-based retrieval with complex captions. This raises an intriguing question: How could we harness the imaginative power of generative models to improve retrieval performance?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 373, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 516 + ], + "type": "text", + "content": "We thus propose IP-CIR (Imagined Proxy for CIR), a training-free, plug-and-play method that harnesses the power of imagination for any retrieval method. Firstly, by utilizing LLMs to understand both the image content and the relative captions, we generate a suitable image layout. Using controllable generation methods, we create high-quality, fine-grained proxy images. Next, we compose the proxy images, query images as well as semantic perturbation into robust proxy features that are more suitable for retrieval. Finally, we propose a balance metric to combine the proxy and text retrieval similarities, allowing the integration of our method with any retrieval methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "type": "text", + "content": "We conduct experiments on three public datasets: the CIRR [28], CIRCO [5], and FashionIQ [41]. We combine our method with several baselines, the observed performance improvements confirm that our method supplements the retrieval process with valuable information and could enhance the performance of any other retrieval methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 590, + 253, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 590, + 253, + 601 + ], + "spans": [ + { + "bbox": [ + 67, + 590, + 253, + 601 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 605, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 52, + 605, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 605, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 52, + 605, + 295, + 653 + ], + "type": "text", + "content": "1) We propose using generative models' imagination capabilities to create proxy images that align with the content of the query image and relative captions to improve retrieval performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 654, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 654, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 51, + 654, + 295, + 689 + ], + "type": "text", + "content": "2) We introduce a proxy image generation framework. Using the LLM-generated object layout and the query image, we create imagined proxies through MIGC++." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 51, + 689, + 295, + 714 + ], + "type": "text", + "content": "3) We propose integrating proxy images with the query image and semantic perturbation into more robust proxy fea" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "content": "tures and introducing a balance metric to combine the proxy and baseline similarities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 96, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 96, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 310, + 96, + 553, + 144 + ], + "type": "text", + "content": "4) We conducted experiments on three public CIR datasets. The improvement over baseline results, as well as achieving SOTA results demonstrates the potential of constructing imagined proxies in retrieval." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 156, + 399, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 156, + 399, + 168 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 399, + 168 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 176, + 512, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 512, + 189 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 512, + 189 + ], + "type": "text", + "content": "2.1. Zero-shot Composed Image Retrieval" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 194, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 483 + ], + "type": "text", + "content": "Early work on ZS-CIR, such as Pic2word [34], SEALRE [6], attempted to project the original image into the text space using lightweight projection models. This was often done within fixed templates (e.g., \"a photo of \\([\\\\) ]$ with cond\"), the \\([\\\\( ]\\) symbol is a textual notation used to represent query image information, and \"cond\" represents the relative captions. LinCIR [16] further defined keywords as adjectives and nouns in the text, thus overcoming the limitations of fixed templates and training more powerful projection models. Besides, CompoDiff[17] trains a new diffusion model that performs diffusion on the CLIP features of the image, introduces relative caption features, and applies attention operations with intermediate image features, ultimately obtaining the target features for retrieval. In addition, some methods have embraced LLMs [1, 31]. CIREVL [20] proposed generating captions for reference images and then using LLMs to reassemble them into target texts modified by relative captions for retrieval. LDRE [43] took this further by using LLMs to construct various edited captions and combining them based on similarity, addressing the issue of fuzzy retrieval. While these methods improved retrieval accuracy, they all overlooked the potential of directly imagining proxy images to provide additional information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 491, + 520, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 520, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 520, + 503 + ], + "type": "text", + "content": "2.2. Controllable Text-to-Image Generation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "In recent years, with the development of diffusion models [19, 35], high-quality images that meet specific requirements can now be generated. Using mainstream generators such as Stable Diffusion [33], users can generate images aligned with text descriptions. Although these foundational generators offer a certain level of control, their performance in fine-grained attribute and orientation control is limited, primarily due to the attribute ambiguity in the CLIP features of the text [48]. As a result, some efforts have focused on leveraging layout information to enhance control. For example, GLIGEN [24] introduces Gated-Self-Attention to combine positional feature information, thereby improving control over positioning. InstanceDiffusion [38] further improves attribute control on this foundation. Meanwhile, MIGC [48] adopts a divide-and-conquer approach, breaking down complex instance generation into multiple single-instance generations and incorporating Enhancement Atten" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3985" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 91, + 312, + 190 + ], + "blocks": [ + { + "bbox": [ + 95, + 78, + 277, + 90 + ], + "lines": [ + { + "bbox": [ + 95, + 78, + 277, + 90 + ], + "spans": [ + { + "bbox": [ + 95, + 78, + 277, + 90 + ], + "type": "text", + "content": "(a) Imagined Retrieval Proxy Generation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 91, + 312, + 190 + ], + "lines": [ + { + "bbox": [ + 66, + 91, + 312, + 190 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 312, + 190 + ], + "type": "image", + "image_path": "fd55235f2763169362301aaa30e2eb484755851213712e0754860073e8ee8be9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 312, + 89, + 552, + 194 + ], + "blocks": [ + { + "bbox": [ + 371, + 76, + 501, + 88 + ], + "lines": [ + { + "bbox": [ + 371, + 76, + 501, + 88 + ], + "spans": [ + { + "bbox": [ + 371, + 76, + 501, + 88 + ], + "type": "text", + "content": "(b) Balanced Proxy Retrieval" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 89, + 552, + 194 + ], + "lines": [ + { + "bbox": [ + 312, + 89, + 552, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 89, + 552, + 194 + ], + "type": "image", + "image_path": "99f2b1d15938947dd08754e4771b0cb8b3ebb96761a6e7c1d2ef1f0a8a7dd6c9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 206, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 555, + 251 + ], + "type": "text", + "content": "Figure 2. Overview of our method. (a) represents our imagined proxy generation process. We use LLM to analyze the BLIP2-generated query image captions and the relative captions and infer the proxy layout. We then use the controllable generator to imagine the proxy images. (b) represents our process of constructing a robust proxy feature, and balancing the text and proxy similarities. We integrate proxy features, query image features as well as semantic perturbations into a robust proxy feature, and propose a balance metric for retrieval." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 271, + 296, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 296, + 404 + ], + "type": "text", + "content": "tion and Layout Attention to improve both positional and attribute control. MIGC++ [49] further integrates ELITE [40] to transform images into CLIP-based text features and introduces an additional EA layer to process information from the reference image modality, thereby incorporating image modality control in the generation process. In this paper, we aim to utilize the fine-grained control capabilities of MIGC to display certain attributes and other information that are difficult to clearly express in text at a fine-grained level within the image itself. This provides supplementary information during the retrieval process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 111, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 111, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 111, + 426 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 435, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 567 + ], + "type": "text", + "content": "In this section, we will introduce our IP-CIR, which attempts to imagine an appropriate proxy image within the image feature space, balance, and combine the retrieval results from the image side with those from the text side. Specifically, in section 3.1, we introduce the overview of our method, in section 3.2, we will describe how to construct an imagined retrieval proxy for each query. In section 3.3, we will further explain how to organize and integrate the information from the imagined retrieval proxy and present how to combine the retrieval information from both the image and text feature spaces." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 575, + 123, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 123, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 123, + 586 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "The overall structure of our framework is shown in Fig.2. First, we generate the imagined proxy image for retrieval. The proxy image needs to align with the content of the query image and be as consistent as possible with the description in the relative captions. Therefore, we first generate a reasonable object layout and then use a controllable generation model to imagine this layout as a concrete image. Specifically, we use BLIP2[23] to generate a set of captions for each query. Combine with the relative captions, we instruct the LLM to reason out a suitable" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 271, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 380 + ], + "type": "text", + "content": "spatial layout of the target image, so that, with as much accuracy as possible in terms of scene, instance attributes, and quantity with relative caption, we can imagine a proxy image of the retrieval target to provide additional information for the retrieval process. In Fig.2 (b), we compose the imagined proxy with query image and text features, forming a robust retrieval proxy. The similarity of robust proxy is further balanced with other baselines' similarity, improving the retrieval accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 392, + 514, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 514, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 514, + 405 + ], + "type": "text", + "content": "3.2. Imagined Retrieval Proxy Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 413, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 484 + ], + "type": "text", + "content": "Text retrieval features may struggle to precisely capture important semantics in the relevant text and the query image, especially when the caption is complex. Generating an image that meets the retrieval requirements, which we refer to as imagining the retrieval proxy, could provide more valuable details to the retrieval process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "Imagine Proxy Layout. The first step of imagining a suitable proxy relies on envisioning the overall layout, which requires a clear understanding of the semantic content and relationship between the query images and relative captions. We thus leverage the robust interpretive and reasoning abilities of LLMs[1, 31] to generate a target proxy layout. As illustrated in Fig.2 (a), we organize the BLIP2 generated query image captions with the relative caption into the fixed form 'Given an image of{caption}, we show{rule}', based on which the LLM is instructed to infer the reasonable overall scene and the layout of the proxy image. The layout includes a detailed description of each instance as well as the bounding box (bbox) coordinates. Besides, certain instances may require detailed visual information from text or image modality. For instance, when retrieving an image of a dog wearing a hat, the attributes of the dog in the proxy image come from the dog in the query image, while the attributes of the hat come from the text modality information. We thus require the LLM to determine the reference modal" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3986" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 545, + 300 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 545, + 300 + ], + "type": "table", + "html": "
BenchmarkCIRCOCIRR
MetricmAP@KRecall@KRecall Subset@K
BackboneMethodk=5k=10k=25k=50k=1k=5k=10k=50k=1k=2k=3
ViT-L/14CIREVL18.5719.0120.8921.8024.5552.3164.9286.3459.5479.8889.69
LinCIR12.8513.7515.2616.1724.8253.0866.6088.8955.6176.2488.24
LinCIR + IP-CIR14.1615.2617.1518.0524.9653.7666.8288.8755.5276.5188.27
(vs. LinCIR)+1.31+1.51+1.89+1.88+0.14+0.68+0.22-0.02-0.09+0.27+0.03
LDRE23.3524.0326.4427.5026.5355.5767.5488.5060.4380.3189.90
LDRE + IP-CIR26.4327.4129.8731.0729.7658.8271.2190.4162.4881.6490.89
(vs. LDRE)+3.08+3.38+3.43+3.57+3.23+3.25+3.67+1.91+2.05+1.33+0.99
ViT-G/14CIREVL26.7727.5929.9631.0334.6564.2975.0691.6667.9584.8793.21
LinCIR20.3421.8523.9825.1434.7564.1275.9393.4262.3681.7891.28
LinCIR + IP-CIR25.7026.6429.0930.1335.3764.7076.1593.7162.5881.7491.35
(vs. LinCIR)+5.40+4.79+5.11+4.99+0.62+0.58+0.22+0.29+0.22-0.04+0.07
LDRE31.1232.2434.9536.0336.1566.3977.2593.9568.8285.6693.76
LDRE + IP-CIR32.7534.2636.8638.0339.2570.0780.0094.8969.9586.8794.22
(vs. LDRE)+1.63+2.02+1.91+2.00+3.10+3.68+2.75+0.94+1.13+1.21+0.46
", + "image_path": "25611b28803426aca83b1f8e93e72784cf2b7401b1d4e20bf5c4e436fce2e231.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 198, + 300, + 411, + 310 + ], + "lines": [ + { + "bbox": [ + 198, + 300, + 411, + 310 + ], + "spans": [ + { + "bbox": [ + 198, + 300, + 411, + 310 + ], + "type": "text", + "content": "Table 1. Quantitative results in CIRCO and CIRR datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 333, + 141, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 141, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 141, + 344 + ], + "type": "text", + "content": "ity for each instance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 346, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 346, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 54, + 346, + 295, + 525 + ], + "type": "text", + "content": "Proxy Image Generation. The next step is to imagine the proxy based on the generated layout. Since the CIR task is a fuzzy retrieval task [43], this reduces the need to maintain the query image's instance ID during the generation process. So we follow the idea of MIGC++ [49], incorporating the query image features transformed by ELITE [40] into MIGC, thereby introducing some degree of the query image information into the Proxy image. Besides, an instance can be influenced by both the reference image and the relative caption. For example, if a user envisions a hat in the query image with a specific pattern, the hat should retain its original shape but include the target pattern. So, we copy the layout of the instance with reference image modality and render it based on both image and text, ensuring the proxy image reflects the text's modifications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 536, + 198, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 198, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 198, + 550 + ], + "type": "text", + "content": "3.3. Balanced Proxy Retrieval" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 556, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 700 + ], + "type": "text", + "content": "Once the proxy image is generated, the key issue lies in how to leverage this proxy image in the retrieval. A straightforward method is to compute the proxy similarity of the retrieval database, weighted sum with the text similarity. However, as shown in the second line of Fig.6, directly using the CLIP features of the proxy image (+PI) can lead to an excessive focus on certain elements, like the beach background. This may cause the results to overly emphasize the visual information in the image while neglecting some details from the text. So in this section, we propose to construct a robust retrieval proxy and balance the retrieval result from image and text." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Constructing Robust Proxy. Inspired by LDRE [43], we" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "text", + "content": "use the LLM [1, 31] to infer a set of target captions, then obtain its CLIP features " + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "text", + "content": ". We derive the semantic perturbation features " + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "inline_equation", + "content": "f_{s} = f_{t} - f_{o}" + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "text", + "content": " and the CLIP features of BLIP2 [23] generated query image captions " + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "inline_equation", + "content": "f_{o}" + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "text", + "content": ". The difference of CLIP text features " + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "text", + "content": " represents a semantic transformation direction from the query image to the target retrieval image, which aligns with the transformation direction in the image space. So the introduction of semantic perturbation provides useful information in the proxy images. Besides, The CLIP features of the original image also contain some basic information relevant to the current query. As shown in Fig.2 (b), we combine the query image's CLIP features " + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "inline_equation", + "content": "f_{q}" + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "text", + "content": ", the proxy CLIP feature " + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "text", + "content": " with the semantic perturbation " + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "text", + "content": " into the robust proxy " + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "inline_equation", + "content": "f_{RP}" + }, + { + "bbox": [ + 313, + 332, + 555, + 596 + ], + "type": "text", + "content": ". There are various ways to combine retrieval features. A simple approach is to scale all features to the same magnitude using max normalization as shown in Eq.1. Besides, the weights of each component can be adjusted to control retrieval bias based on the quality of the proxy image and specific retrieval goals. For example, assigning a higher weight to query features ensures a closer match to the query image if the user prioritizes similarity to the original query." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 350, + 601, + 555, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 601, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 350, + 601, + 555, + 628 + ], + "type": "interline_equation", + "content": "f _ {R P} = f _ {p} + \\frac {\\operatorname* {m a x} \\left(f _ {p}\\right)}{\\operatorname* {m a x} \\left(f _ {q}\\right)} f _ {q} + \\frac {\\operatorname* {m a x} \\left(f _ {p}\\right)}{\\operatorname* {m a x} \\left(f _ {s}\\right)} f _ {s}. \\tag {1}", + "image_path": "6ee7457d1605eaa6c11f821ea6c4cfaca92d9c3b3dc062b90a18811208f65a1a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "Balancing Retrieval Results. After obtaining the retrieval feature in text and image spaces, a critical question arises: How should we balance the two retrieval results? The proxy features we obtain exist in the image feature space, making it difficult to directly perform weighted summation with the baseline text features. Therefore, we choose to combine their similarities. A simple approach is to take the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3987" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 70, + 512, + 171 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 512, + 171 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 512, + 171 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 512, + 171 + ], + "type": "table", + "html": "
BackboneTypeShirtDressTopteeAverage
MethodR@10R@50R@10R@50R@10R@50R@10R@50
ViT-G/14Pic2Word33.1750.3925.4347.6535.2457.6231.2851.89
SEARLE36.4655.3528.1650.3239.8361.4534.8155.71
LinCIR46.7665.1138.0860.8850.4871.0945.1165.69
LinCIR + IP-CIR48.0466.6839.0261.0350.1871.1445.7466.28
(vs. LinCIR)+1.28+1.57+0.94+0.15-0.30+0.05+0.63+0.59
", + "image_path": "20978278862c0e87aa1dc0932dba9ce37d99ee69ffd5aee1902a3e0d70cf77ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 214, + 171, + 395, + 181 + ], + "lines": [ + { + "bbox": [ + 214, + 171, + 395, + 181 + ], + "spans": [ + { + "bbox": [ + 214, + 171, + 395, + 181 + ], + "type": "text", + "content": "Table 2. Quantitative results in FashionIQ dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 193, + 243, + 473 + ], + "blocks": [ + { + "bbox": [ + 58, + 193, + 243, + 473 + ], + "lines": [ + { + "bbox": [ + 58, + 193, + 243, + 473 + ], + "spans": [ + { + "bbox": [ + 58, + 193, + 243, + 473 + ], + "type": "image", + "image_path": "34f3db9c5438476110a026369fa1df8158b193385d96273578d83e9ca801aa02.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 481, + 555, + 538 + ], + "lines": [ + { + "bbox": [ + 55, + 481, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 555, + 538 + ], + "type": "text", + "content": "Figure 3. Qualitative results of our method. We conducted experiments on the CIRCO validation dataset to observe in which cases our method improves retrieval results. In the blue section on the left, we display the query information used for retrieval, the ground truth target image, and our generated proxy image features. 'Query' represents the input query image, 'Caption' represents the relative text, and in the 'Proxy' section, we show two generated imagined proxies. In the red section on the right, the top and bottom rows display the top-5 retrieval results enhanced by our imagined proxies and the baseline's top-5 retrieval results, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 245, + 193, + 554, + 473 + ], + "blocks": [ + { + "bbox": [ + 245, + 193, + 554, + 473 + ], + "lines": [ + { + "bbox": [ + 245, + 193, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 245, + 193, + 554, + 473 + ], + "type": "image", + "image_path": "7d7e5c5acf2266acebe63a751d65109f0b5969130b50f55ba378e9f4f8c5b3d7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "text", + "content": "average or weighted sum, however, in extreme cases, when we have a case with a 0.999 text similarity but only a 0.001 proxy similarity, and another case with both a 0.5 base similarity and a 0.5 proxy similarity, their averages are both 0.5, which could lead to retrieving overly extreme results. We thus propose a balanced metric for combining these results. Specifically, given the baseline similarity " + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "text", + "content": " and proxy similarity " + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "inline_equation", + "content": "S_{p}" + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "text", + "content": ", the final similarity " + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "inline_equation", + "content": "S_{f}" + }, + { + "bbox": [ + 55, + 553, + 295, + 649 + ], + "type": "text", + "content": " for retrieval is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 659, + 295, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 659, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 96, + 659, + 295, + 673 + ], + "type": "interline_equation", + "content": "S _ {f} = \\lambda S _ {t} + (1 - \\lambda) S _ {b}, S _ {b} = S _ {t} * S _ {p}, \\tag {2}", + "image_path": "d15e60933e5647d8b8dba1ea19ef01558e0195beccc7c159cef473d56a921bc8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "S_{b}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " represents the balanced similarity, " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " is a balancing parameter. We believe that the balance metric enables us to focus more on cases that perform well in both similarities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 552, + 394, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 394, + 565 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 394, + 565 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "In this section, we demonstrate the effectiveness of our method. In Section 4.1, We will provide a detailed description of our experimental setup, including the baselines, evaluation metrics, and implementation details. In Section 4.2, we compare our method with other state-of-the-art approaches. We explore the improvements in retrieval performance by presenting some visual examples, highlighting the impact of our method in Section 4.3, and in Section 4.4, we conduct ablation studies to investigate the contribution of each component of our method, as well as the influence of various parameter settings." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3988" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 171, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 171, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 171, + 85 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 209 + ], + "type": "text", + "content": "Datasets. We validate the performance of our method by combining them with several state-of-the-art methods on three datasets: CIRR [28], CIRCO [5], and FashionIQ [41]. CIRR contains 21,552 real images from the NLVR [36] dataset. CIRCO is built from the COCO 2017 unlabeled set [27], includes a validation set with 220 queries and a test set with 800 queries. FashionIQ [41] is a dataset dedicated to the fashion domain, organized into three distinct subcategories: Dress, Shirt, and Toptee. It contains 30,135 query triplets and 77,683 images available for retrieval." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 210, + 296, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 296, + 269 + ], + "type": "text", + "content": "Baseline. We selected LDRE[43] and LinCIR[16] as our baselines. For each query, we generate 5 proxy images and obtain the enhanced retrieval results. We compare the improved results with some of the most recent ZS-CIR results, such as Searle[6], CIREVL[20]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 376 + ], + "type": "text", + "content": "Evaluation Metrics. In CIRCO, each query has multiple target images, so we calculate mAP@k to offer a more detailed assessment, where " + }, + { + "bbox": [ + 55, + 270, + 296, + 376 + ], + "type": "inline_equation", + "content": "k \\in \\{5,10,25,50\\}" + }, + { + "bbox": [ + 55, + 270, + 296, + 376 + ], + "type": "text", + "content": " indicates the number of top-ranked retrieval results. For CIRR, we follow the original benchmarks and use Recall@k (with " + }, + { + "bbox": [ + 55, + 270, + 296, + 376 + ], + "type": "inline_equation", + "content": "k \\in \\{1,5,10,50\\}" + }, + { + "bbox": [ + 55, + 270, + 296, + 376 + ], + "type": "text", + "content": ") as the primary metric. We also assess performance in a subset setting, which we denote as RecallSubset@K (with " + }, + { + "bbox": [ + 55, + 270, + 296, + 376 + ], + "type": "inline_equation", + "content": "k \\in \\{1,2,3\\}" + }, + { + "bbox": [ + 55, + 270, + 296, + 376 + ], + "type": "text", + "content": "). On FashionIQ, we also use Recall@K (with " + }, + { + "bbox": [ + 55, + 270, + 296, + 376 + ], + "type": "inline_equation", + "content": "k \\in \\{10,50\\}" + }, + { + "bbox": [ + 55, + 270, + 296, + 376 + ], + "type": "text", + "content": ") for evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 496 + ], + "type": "text", + "content": "Implement Details. We primarily conducted experiments using two backbone models: CLIP-L and CLIP-G. We used Qwen1.5-32B[1] as the LLM in all the experiments, which yielded results comparable to those reported in the original papers. For generating proxy images, we use MIGC[48] with SD1.5 as the backbone. We use Qwen1.5-32B as our LLM to generate the layout for each query and then generate 5 proxy images based on the query image and layout. All the experiments are conducted by Pytorch with one NVIDIA A6000 GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 503, + 174, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 174, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 174, + 516 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 605 + ], + "type": "text", + "content": "CIRCO. We present the experimental results on the CIRCO test dataset on LDRE[43] and LinCIR[16]. In the left half of Tab.1, we observe that our proxy image can improve the retrieval performance, which an approximately 5-point improvement in mAP at " + }, + { + "bbox": [ + 55, + 521, + 296, + 605 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 55, + 521, + 296, + 605 + ], + "type": "text", + "content": " in LinCIR, and about 2-point improvement in LDRE. These results effectively demonstrate the validity and impact of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "content": "CIRR. In the right half of Tab.1, we present the experimental results on the CIRR test dataset. Our proxy image significantly improved the LDRE and LinCIR performance. The performance improvements across different methods and backbones indicate that our approach can be applied to various text-based retrieval methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "FashionIQ. In Tab.2, we present the results on the FashionIQ val dataset. Since LDRE has not open-sourced the code for the FashionIQ, we used LinCIR as the baseline." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 434, + 215 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 434, + 215 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 434, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 434, + 215 + ], + "type": "image", + "image_path": "156d1b69e3359d67382688657255f16805d054d4417d846d571d58822cab6dba.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 435, + 72, + 553, + 215 + ], + "blocks": [ + { + "bbox": [ + 435, + 72, + 553, + 215 + ], + "lines": [ + { + "bbox": [ + 435, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 435, + 72, + 553, + 215 + ], + "type": "image", + "image_path": "b68559c739a8c03a95950baa441cefeb0805e558c5c5c5905bcc661c157953ff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 217, + 434, + 278 + ], + "blocks": [ + { + "bbox": [ + 317, + 217, + 434, + 278 + ], + "lines": [ + { + "bbox": [ + 317, + 217, + 434, + 278 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 434, + 278 + ], + "type": "image", + "image_path": "a250bcd7faab5928edf80e6fb10a88e0aa32c436104ac78299da7652bf2b520a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 287, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 365 + ], + "type": "text", + "content": "Figure 4. Qualitative results of our method. We present the result on CIRCO validation dataset with top-1 retrieval results. Baseline methods in text-based retrieval overlook certain attributes and details in complex related texts, resulting in suboptimal retrieval outcomes. In contrast, our approach can construct suitable proxies on the image side, supplementing this missing information and yielding improved retrieval results." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 435, + 217, + 554, + 277 + ], + "blocks": [ + { + "bbox": [ + 435, + 217, + 554, + 277 + ], + "lines": [ + { + "bbox": [ + 435, + 217, + 554, + 277 + ], + "spans": [ + { + "bbox": [ + 435, + 217, + 554, + 277 + ], + "type": "image", + "image_path": "4ff09d93475ebd6315086384d58bd87425ebd27590768ace9c858ce3c0b54ffc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 332, + 374, + 536, + 526 + ], + "blocks": [ + { + "bbox": [ + 332, + 374, + 536, + 526 + ], + "lines": [ + { + "bbox": [ + 332, + 374, + 536, + 526 + ], + "spans": [ + { + "bbox": [ + 332, + 374, + 536, + 526 + ], + "type": "image", + "image_path": "f869bcbc0e4f53ab9c104de527212126d44b88adc91fa942006c97b7f6242ee3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 538, + 555, + 572 + ], + "lines": [ + { + "bbox": [ + 313, + 538, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 555, + 572 + ], + "type": "text", + "content": "Figure 5. Ablation results on the composition of robust proxy features in the CIRCO dataset. " + }, + { + "bbox": [ + 313, + 538, + 555, + 572 + ], + "type": "inline_equation", + "content": "pi" + }, + { + "bbox": [ + 313, + 538, + 555, + 572 + ], + "type": "text", + "content": " indicates proxy features, " + }, + { + "bbox": [ + 313, + 538, + 555, + 572 + ], + "type": "inline_equation", + "content": "si" + }, + { + "bbox": [ + 313, + 538, + 555, + 572 + ], + "type": "text", + "content": " represents semantic perturbation, and " + }, + { + "bbox": [ + 313, + 538, + 555, + 572 + ], + "type": "inline_equation", + "content": "qi" + }, + { + "bbox": [ + 313, + 538, + 555, + 572 + ], + "type": "text", + "content": " indicates the query features." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 589, + 555, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 555, + 625 + ], + "type": "text", + "content": "Although we observe a drop in " + }, + { + "bbox": [ + 313, + 589, + 555, + 625 + ], + "type": "inline_equation", + "content": "\\mathbb{R}@\\mathbb{10}" + }, + { + "bbox": [ + 313, + 589, + 555, + 625 + ], + "type": "text", + "content": " in the Toptee, our method improves the retrieval accuracy in the Shirt and dress Subset, thus improving the average performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 634, + 425, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 425, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 425, + 646 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": "We conducted experiments on the CIRCO validation dataset to verify the improvements our method brings to retrieval. Experiment setting. Specifically, we used LDRE as the baseline and generated 5 imagined proxy images for each query. We present two types of results. First, based on" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3989" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 529, + 217 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 529, + 217 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 529, + 217 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 529, + 217 + ], + "type": "table", + "html": "
BenchmarkCIRRCIRCO
MetricRecall@KmAP@K
BackbonePIRPBMk=1k=2k=5k=10k=50k=5k=10k=25k=50
LDRE-G✓36.1549.4966.3977.2593.9531.1232.2434.9536.03
36.0250.0266.5177.5293.7431.0632.4135.1936.35
-0.13+0.53+0.12+0.27-0.21-0.06+0.17+0.24+0.32
✓✓38.1552.5368.5579.4794.4831.9233.5936.2637.37
+2.00+3.04+2.16+2.22+0.53+0.80+1.35+1.31+1.34
✓✓✓39.2552.9470.0780.0094.8932.7534.2636.8638.03
+3.10+3.45+3.68+2.75+0.94+1.63+2.02+1.91+2.00
", + "image_path": "8df1d113611ce9dffb831e0547d1a5cfc0063921981b7fbf0ee29fbcc2dfaa43.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 206, + 217, + 403, + 226 + ], + "lines": [ + { + "bbox": [ + 206, + 217, + 403, + 226 + ], + "spans": [ + { + "bbox": [ + 206, + 217, + 403, + 226 + ], + "type": "text", + "content": "Table 3. Ablation results in CIRR and CIRCO dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 71, + 239, + 532, + 459 + ], + "blocks": [ + { + "bbox": [ + 71, + 239, + 532, + 459 + ], + "lines": [ + { + "bbox": [ + 71, + 239, + 532, + 459 + ], + "spans": [ + { + "bbox": [ + 71, + 239, + 532, + 459 + ], + "type": "image", + "image_path": "5e2716f24ff5d3e1c39219c9492f68a6a546a41ab6e32d7b7fc4035390455251.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 466, + 522, + 479 + ], + "lines": [ + { + "bbox": [ + 86, + 466, + 522, + 479 + ], + "spans": [ + { + "bbox": [ + 86, + 466, + 522, + 479 + ], + "type": "text", + "content": "Figure 6. Visualize of Ablation Result. We present the top-1 retrieval result on CIRCO dataset in ablation experiments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 489, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 489, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 54, + 489, + 295, + 550 + ], + "type": "text", + "content": "the ground truth for each query, we show in Fig.3 that our method improves the accuracy of ground truth occurrence in top-5 retrieval results. Second, disregarding the ground truth image, we demonstrate in Fig.4 that our method retrieves a more suitable image in the top-1 result." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 557, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 557, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 557, + 295, + 713 + ], + "type": "text", + "content": "Analysis of Experimental Results on TOP-5 Retrieval. From the visualized results in Fig.3, we can draw the following conclusions: (1) Our generated proxy images effectively preserve the style and key elements of the original image. In the first row, our image maintains the clock tower as a significant element and adheres to the related caption by generating a Ferris wheel. In the second row, our image retains the original black-and-white style. (2) The imagined proxies complement important information that may be overlooked by text alone. For instance, in the first-row example, the clock tower—a crucial concept—is missing in the text-only retrieval result, while our imagined proxy and the original image both provide this retrieval information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 489, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 555, + 513 + ], + "type": "text", + "content": "In the third row, our constructed image improves retrieval accuracy by incorporating an umbrella and snow." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "text", + "content": "Analysis of Experimental Results on TOP-1 Retrieval. Cases in Fig.4 show that by providing proxy image information, our approach achieves more relevant retrievals. Retrieval based solely on text space may only capture partial information in complex contexts, leading to the omission of certain details in the images—such as the person in the first row, the tie in the second row, and the White House in the background of the third row. Our method incorporates these details in the images, supplementing and balancing this information to yield more suitable results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 646, + 408, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 408, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 408, + 658 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Effects of the Proxy Image Features. We first conduct an ablation study on the proposed proxy image features in the CIRR dataset and CIRCO dataset. We use PI (Proxy Image) refers to the approach that directly computes retrieval" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3990" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 61, + 293, + 172 + ], + "blocks": [ + { + "bbox": [ + 61, + 61, + 293, + 172 + ], + "lines": [ + { + "bbox": [ + 61, + 61, + 293, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 61, + 293, + 172 + ], + "type": "table", + "html": "
MetricmAP@K
Backbonek=5k=10k=25k=50
LDRE-G31.1232.2434.9536.03
LDRE-G + SD32.2533.5036.1837.30
LDRE-G + FLUX32.3833.8936.5137.58
LDRE-G + ip2p31.8233.2736.1137.23
LDRE-G + magicbrush31.8633.3936.0537.22
LDRE-G + L2I32.7534.2636.8638.03
", + "image_path": "434895b3be0491b3fb8913ec7e160f82ae98b0c4e788b2c23c9f8068830427b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 173, + 294, + 195 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 294, + 195 + ], + "type": "text", + "content": "Table 4. Comparison of Various Methods for Generating Proxy Images on the CIRCO test" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 196, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 196, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 196, + 295, + 399 + ], + "type": "text", + "content": "similarity from proxy image features, and then performs a weighted sum with the original similarity. RP (Robust Proxy) indicates integrating the proxy image into more robust proxy features. BM (Balance Metric) refers to using the balance metric. The results from the CIRR and CIRCO are shown in Tab.3. We can draw the following conclusions: 1) The imagined proxy images may contain some noise, leading to the retrieval focus on irrelevant details. we observe a decrease in recall at " + }, + { + "bbox": [ + 54, + 196, + 295, + 399 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 54, + 196, + 295, + 399 + ], + "type": "text", + "content": " and 50 in CIRR, and mAP " + }, + { + "bbox": [ + 54, + 196, + 295, + 399 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 54, + 196, + 295, + 399 + ], + "type": "text", + "content": " decreases in CIRCO. 2) After integrating the proxy image into a more robust proxy feature, we see a significant improvement in both datasets, which demonstrates the effectiveness of our robust proxy features. 3) By introducing the balance metric between proxy and text similarity, we further improve the retrieval performance. This suggests that combining the information from both image and text domains leads to more accurate retrieval results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 401, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 544 + ], + "type": "text", + "content": "We can draw similar conclusions from Fig.6, which shows retrieval results on the CIRCO test dataset. In the first row, the proxy image provides a realistic and accurate depiction of the target image, allowing for improved retrieval results. In the second row, relying solely on the query image may cause the retrieval to overly focus on the beach background, contradicting the \"no people\" text information. Here, using the robust proxy " + }, + { + "bbox": [ + 55, + 401, + 295, + 544 + ], + "type": "inline_equation", + "content": "(\\mathrm{PI} + \\mathrm{RP})" + }, + { + "bbox": [ + 55, + 401, + 295, + 544 + ], + "type": "text", + "content": " further enhances the overlooked aspects. In the third row, the balance metric balances the 'the woman with sunglasses' in proxy features with 'trees' in baseline features, achieving better retrieval results for both the subject and the background." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 544, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 700 + ], + "type": "text", + "content": "Ablation on Robust Features. We analyze the composition of our robust proxy on the CIRCO dataset, with results shown in Fig.5. The robust proxy is composed of three main components: " + }, + { + "bbox": [ + 55, + 544, + 295, + 700 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 55, + 544, + 295, + 700 + ], + "type": "text", + "content": " represents the query image features, " + }, + { + "bbox": [ + 55, + 544, + 295, + 700 + ], + "type": "inline_equation", + "content": "s_{i}" + }, + { + "bbox": [ + 55, + 544, + 295, + 700 + ], + "type": "text", + "content": " represents the semantic perturbation, and " + }, + { + "bbox": [ + 55, + 544, + 295, + 700 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 55, + 544, + 295, + 700 + ], + "type": "text", + "content": " represents our proxy image. A comparison among ①, ②, ③, and ④ shows that directly using our proxy image features can enhance retrieval performance to a certain extent, whereas relying solely on the query image or semantic perturbation tends to focus excessively on incorrect information, leading to a decline in performance. Additionally, comparisons among ⑤, ⑥, and ⑦ demonstrate that combining all three features together improves retrieval accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Approaches for constructing proxy images. We also" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 347, + 79, + 509, + 201 + ], + "blocks": [ + { + "bbox": [ + 347, + 79, + 509, + 201 + ], + "lines": [ + { + "bbox": [ + 347, + 79, + 509, + 201 + ], + "spans": [ + { + "bbox": [ + 347, + 79, + 509, + 201 + ], + "type": "image", + "image_path": "c31ae80af61a82718ca1b8ebfc44def5222ccd5b3006ba119a917be2ef18250d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 205, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 228 + ], + "type": "text", + "content": "Figure 7. Parameter Analysis. The effect of varying numbers of proxy images on performance for the CIRCO dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 240, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 382 + ], + "type": "text", + "content": "tested the results obtained on the CIRCO dataset using different methods to construct the proxy image in Tab.4. Using the target image caption inferred by the LLM, we tested two T2I generation approaches: SD[33] and FLUX[22]. Additionally, based on the relative caption and query image, we also tested two instruction-based editing approaches Instruct-Pix2Pix[7] and MagicBrush[44]. The results show that: 1) The robust features constructed through these approaches can all provide additional information for retrieval. 2) The stronger the control capability, the more precise the provided information, and the greater the improvement achieved." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 384, + 556, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 469 + ], + "type": "text", + "content": "Effects of the number of proxy images. We tested the impact of using different numbers of proxy images on performance with the CIRCO dataset. As shown in Fig.7, retrieval performance improves as the number of proxy images increases, though the rate of improvement gradually slows down. Therefore, the number of proxy images can be chosen based on a balance between efficiency and accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 665 + ], + "type": "text", + "content": "In this paper, we propose the IP-CIR, a plug-and-play training-free method for current ZSCIR tasks. Our method imagines proxy images, which align with the query image and the relative captions, providing fine-grained details that may be overlooked by the text. We combine the proxy images, query images, and semantic perturbation into more robust proxy image features, and propose a balance metric to compose the proxy and baseline similarities, enhancing the retrieval performance. The experiment results in CIRR, CIRCO, and FashionIQ datasets show that our method can successfully improve retrieval accuracy and demonstrate the potential for providing additional retrieval information from the image side, revealing a new direction for improving retrieval accuracy in the future." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (62293554) and in part by the Natural Science Foundation of Zhejiang Province (LDT23F02023F02)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "3991" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 223 + ], + "type": "text", + "content": "[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 224, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 295, + 290 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 292, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 295, + 346 + ], + "type": "text", + "content": "[3] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21466-21474, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 349, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 349, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 349, + 295, + 403 + ], + "type": "text", + "content": "[4] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4959-4968, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 405, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 405, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 405, + 295, + 437 + ], + "type": "text", + "content": "[5] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion, 2023. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 440, + 295, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 440, + 295, + 494 + ], + "spans": [ + { + "bbox": [ + 62, + 440, + 295, + 494 + ], + "type": "text", + "content": "[6] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15338-15347, 2023. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 496, + 295, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 496, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 496, + 295, + 529 + ], + "type": "text", + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 531, + 295, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 531, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 62, + 531, + 295, + 573 + ], + "type": "text", + "content": "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 576, + 295, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 576, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 62, + 576, + 295, + 619 + ], + "type": "text", + "content": "[9] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 622, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 622, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 622, + 295, + 655 + ], + "type": "text", + "content": "[10] Yanbei Chen and Loris Bazzani. Learning joint visual semantic matching embeddings for language-guided retrieval. page 136-152, Berlin, Heidelberg, 2020. Springer-Verlag. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 657, + 295, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 295, + 700 + ], + "type": "text", + "content": "[11] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2998-3008, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 702, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 295, + 714 + ], + "type": "text", + "content": "[12] Yiyang Chen, Zhedong Zheng, Wei Ji, Leigang Qu, and Tat" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Seng Chua. Composed image retrieval with text feedback via multi-grained uncertainty regularization, 2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 118 + ], + "type": "text", + "content": "[13] Sanghyuk Chun. Improved probabilistic image-text representations, 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "text", + "content": "[14] Ginger Delmas, Rafael Sampaio de Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "type": "text", + "content": "[15] Eric Dodds, Jack Culpepper, Simao Herdade, Yang Zhang, and Kofi Boakye. Modality-agnostic attention fusion for visual search with text feedback, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 241 + ], + "type": "text", + "content": "[16] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, , Yoohoon Kang, and Sangdoo Yun. Language-only training of zero-shot composed image retrieval. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 243, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 296 + ], + "type": "text", + "content": "[17] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoonhoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. Transactions on Machine Learning Research, 2024. Expert Certification. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 299, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 364 + ], + "type": "text", + "content": "[18] Ryota Hinami and Shin'ichi Satoh. Discriminative learning of open-vocabulary object retrieval and localization by negative phrase augmentation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2605-2615, Brussels, Belgium, 2018. Association for Computational Linguistics. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 365, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 553, + 398 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "text", + "content": "[20] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. International Conference on Learning Representations (ICLR), 2024. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 445, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 476 + ], + "type": "text", + "content": "[21] Jooyeon Kim, Eulrang Cho, Sehyung Kim, and Hyunwoo J. Kim. Retrieval-augmented open-vocabulary object detection, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 478, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 499 + ], + "type": "text", + "content": "[22] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 544 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "text", + "content": "[24] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Glagen: Open-set grounded text-to-image generation. CVPR, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 591, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 633 + ], + "type": "text", + "content": "[25] You Li, Fan Ma, and Yi Yang. Anysynth: Harnessing the power of image synthetic data generation for generalized vision-language tasks. arXiv preprint arXiv:2411.16749, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 635, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 668 + ], + "type": "text", + "content": "[26] Chao Liang, Fan Ma, Linchao Zhu, Yingying Deng, and Yi Yang. Caphuman: Capture your moments in parallel universes. In CVPR, pages 6400-6409, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3992" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[28] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre-trained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2125–2134, 2021. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 161 + ], + "type": "text", + "content": "[29] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 208, + 242, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 208, + 242, + 218 + ], + "spans": [ + { + "bbox": [ + 57, + 208, + 242, + 218 + ], + "type": "text", + "content": "[31] OpenAI. Gpt-4 technical report, 2023. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 294, + 274 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 57, + 276, + 294, + 308 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 309, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 309, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 57, + 309, + 294, + 352 + ], + "type": "text", + "content": "[34] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 354, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 354, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 57, + 354, + 294, + 386 + ], + "type": "text", + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 388, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 388, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 57, + 388, + 294, + 421 + ], + "type": "text", + "content": "[36] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 422, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 422, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 57, + 422, + 294, + 454 + ], + "type": "text", + "content": "[37] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 57, + 456, + 294, + 488 + ], + "type": "text", + "content": "[38] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 491, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 294, + 533 + ], + "type": "text", + "content": "[39] Zheng Wang, Xin Yuan, Toshihiko Yamasaki, Yutian Lin, Xin Xu, and Wenjun Zeng. Re-identification = retrieval + verification: Back to essence and forward with a new metric. arXiv preprint arXiv:2011.11506, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 535, + 294, + 578 + ], + "type": "text", + "content": "[40] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848, 2023. 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 580, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 294, + 624 + ], + "type": "text", + "content": "[41] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The fashion iq dataset: Retrieving images by combining side information and relative natural language feedback. CVPR, 2021. 1, 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 625, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 667 + ], + "type": "text", + "content": "[42] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Reco: Region-controlled text-to-image generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[43] Zhenyu Yang, Dizhan Xue, Shengsheng Qian, Weiming Dong, and Changsheng Xu. Ldre: Llm-based divergent reasoning and ensemble for zero-shot composed image retrieval. In Proceedings of the 47th International ACM SI-" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 454 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "GIR Conference on Research and Development in Information Retrieval, pages 80-90, 2024. 1, 2, 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[44] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 205 + ], + "type": "text", + "content": "[45] Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. MagicLens: Self-supervised image retrieval with open-ended instructions. In Proceedings of the 41st International Conference on Machine Learning, pages 59403–59420. PMLR, 2024. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 208, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 229 + ], + "type": "text", + "content": "[46] Liang Zheng, Yi Yang, and Alexander G. Hauptmann. Person re-identification: Past, present and future, 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 231, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 262 + ], + "type": "text", + "content": "[47] Dewei Zhou, Zongxin Yang, and Yi Yang. Pyramid diffusion models for low-light image enhancement. arXiv preprint arXiv:2305.10028, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 319 + ], + "type": "text", + "content": "[48] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6818-6828, 2024. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 320, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 553, + 353 + ], + "type": "text", + "content": "[49] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. Migc++: Advanced multi-instance generation controller for image synthesis. ArXiv, abs/2407.02329, 2024. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 387 + ], + "type": "text", + "content": "[50] Dewei Zhou, Ji Xie, Zongxin Yang, and Yi Yang. 3dis: Depth-driven decoupled instance synthesis for text-to-image generation. ArXiv, abs/2410.12669, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "type": "text", + "content": "[51] Dewei Zhou, Mingwei Li, Zongxin Yang, and Yi Yang. Dreamrenderer: Taming multi-instance attribute control in large-scale text-to-image models, 2025. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "text", + "content": "[52] Dewei Zhou, Ji Xie, Zongxin Yang, and Yi Yang. 3dis-flux: simple and efficient multi-instance generation with dit rendering, 2025. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3993" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/b527a5f4-addd-424b-8f6f-3155526e27dd_content_list.json b/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/b527a5f4-addd-424b-8f6f-3155526e27dd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eb493edea7b3483285bf140627d4f996685a0f5b --- /dev/null +++ b/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/b527a5f4-addd-424b-8f6f-3155526e27dd_content_list.json @@ -0,0 +1,1674 @@ +[ + { + "type": "text", + "text": "ImagineFSL: Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning", + "text_level": 1, + "bbox": [ + 117, + 128, + 880, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoyuan Yang $^{1, *}$ Xiaou Li $^{2, *}$ Jiaming Lv $^{1}$ Xianjun Cheng $^{1}$ Qilong Wang $^{3}$ Peihua Li $^{1 \\dagger}$", + "bbox": [ + 99, + 202, + 895, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dalian University of Technology $^{2}$ Beijing University of Posts and Telecommunications $^{3}$ Tianjin University", + "bbox": [ + 160, + 223, + 836, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://peihuali.org/ImagineFSL", + "bbox": [ + 357, + 241, + 633, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adapting CLIP models for few-shot recognition has recently attracted significant attention. Despite considerable progress, these adaptations remain hindered by the pervasive challenge of data scarcity. Text-to-image models, capable of generating abundant photorealistic labeled images, offer a promising solution. However, existing approaches simply treat synthetic images as complements to real images, rather than as standalone knowledge repositories stemming from distinct foundation models. To overcome this limitation, we frame synthetic images as an imagined base set (iBase), i.e., an independent, large-scale synthetic dataset encompassing diverse concepts. Building on this perspective, we introduce ImagineFSL, a novel CLIP adaptation methodology that pretrains on iBase and then fine-tunes for downstream few-shot tasks. We find that, compared to no pretraining, both supervised and self-supervised pretraining are beneficial, with the latter providing better performance. Based on this finding, we propose an improved self-supervised method tailored for few-shot scenarios, enhancing the transferability of representations from synthetic to real image domains. Additionally, we present a systematic and scalable pipeline that employs chain-of-thought and in-context learning techniques, harnessing foundation models to automatically generate diverse, realistic images. Validated across eleven datasets, our methods consistently outperform state-of-the-art approaches by substantial margins.", + "bbox": [ + 88, + 324, + 485, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 763, + 220, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot learning (FSL), aiming to recognize novel concepts from limited labeled samples, has long been a subject of considerable research interest due to its potential for reducing labor-intensive human annotation and cura", + "bbox": [ + 89, + 789, + 483, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion [1]. FSL methods often rely on pretraining [2, 3] or meta-training [4, 5] on a base set of classes with abundant training examples, followed by meta-testing on a test set with limited labeled data. Recently, large vision-language models (VLMs) [6-8], particularly Contrastive Language-Image Pre-training (CLIP) [8], have exhibited remarkable few-shot and zero-shot recognition capabilities, drawing significant attention [9, 10] in the FSL regime.", + "bbox": [ + 511, + 292, + 905, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To adapt CLIP models for few-shot scenarios, various methods have been proposed, achieving substantial progress [11-14]. Nevertheless, these methods continue to struggle with the pervasive data-scarcity problem inherent in FSL. Although modern text-to-image (T2I) models [15, 16] can generate abundant photorealistic labeled images, the full potential of synthetic data in FSL remains underexplored. Existing approaches [17-19] simply treat synthetic images as complements, directly fine-tuning CLIP models using few-shot real images augmented with the synthetic ones during the meta-testing phase. At first glance, this strategy appears reasonable, since CLIP models are already extensively pretrained on vast datasets, suggesting that no further pretraining might be necessary.", + "bbox": [ + 511, + 415, + 905, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, we argue that such a strategy is sub-optimal because it treats synthetic images merely as complements to real images, instead of as standalone knowledge repositories derived from distinct foundation models. To overcome this limitation, we frame synthetic images as an imagined Base set (iBase), i.e., an independent, large-scale synthetic dataset encompassing diverse concepts, on which adaptation methods can learn rich knowledge transferable to few-shot tasks. By attaching learnable adapters to frozen CLIP encoders, we preserve the original representation capability while granting adaptability to the newly introduced parameters. We present a novel methodology called ImagineFSL (Figure 1), consisting of pretraining on iBase, followed by fine-tuning for downstream FSL tasks with real images and task-specific synthetic images. We find that, compared to no pretraining, either self-supervised learning (Self-SL) or supervised learning (SL) for pretraining is beneficial; crucially, Self-SL (e.g., DINO [20, 21]) outperforms SL.", + "bbox": [ + 511, + 628, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work. †Corresponding Author. The work was supported by National Natural Science Foundation of China (62471083, 62276186, 61971086).", + "bbox": [ + 89, + 862, + 483, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30020", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10f37648a47f9e7de4632d92409c96a964cd218e804af38a7181ee539cb67b09.jpg", + "image_caption": [ + "(a) Self-supervised pretraining on PURELY synthetic dataset of iBase." + ], + "image_footnote": [], + "bbox": [ + 106, + 92, + 491, + 287 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e01a441963cd89597fe6bf607b704665f5395f20c3f16c17f5aed030756c4ede.jpg", + "image_caption": [ + "(b) Fine-tuning with real and task-specific synthetic images.", + "Figure 1. Overview of ImagineFSL. (a) We pretrain the adapter attached to the CLIP image encoder on iBase, an independent, large-scale synthetic dataset, using a Self-SL method called HoM-DINO. This method constructs two views via SyntAug with T2I models conditioned on the same caption and employs two specialized heads: one for higher-order moment (HoM) representations and another for masked image modeling (MIM). (b) For downstream few-shot tasks, we fine-tune both the adapter and vision classifier alongside the vision-language classifier. Additionally, we can fine-tune the image encoder with LoRA, referred to as ImagineFSLloRA, for further improvement." + ], + "image_footnote": [], + "bbox": [ + 514, + 95, + 897, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on this finding, we propose an enhanced Self-SL method that improves upon DINO and is specifically tailored to FSL. Recognizing that local image regions are crucial in few-shot tasks [22-24], we address the limitation of DINO, which relies solely on the [CLS] token for image representation while overlooking patch tokens. Concretely, we model distributions of patch tokens alongside the [CLS] token to capture richer feature statistics. Rather than assuming a fixed a priori distribution (e.g., Gaussian [25]), we employ from low- to High-order Moments (HoM) [26], thus calling our method HoM-DINO. We also incorporate masked image modeling (MIM) [27] to strengthen patch-level discrimination. Additionally, we develop Synthetic Augmentation (SyntAug) by constructing views from different synthetic images corresponding to the same caption. This exploits the natural variations inherent in T2I models, providing superior semantic consistency and diversity than hand-crafted, meticulously chosen augmentations [28].", + "bbox": [ + 91, + 407, + 483, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To facilitate our approach, we develop a systematic and scalable pipeline for synthesizing diverse and realistic images, relying exclusively on Large Language Models (LLMs) and T2I models. Our method employs chain-of-thought (CoT) [29] and in-context learning (ICL) techniques [30] for caption generation. Specifically, we instruct GPT-4 [31] to analyze key factors that define image descriptions, including attribute, background, viewpoint, lighting condition, and cause of image degradation. Furthermore, we direct GPT-4 to generate a small set of exemplary captions following predefined patterns that combine different factors. These captions serve as in-context prompts for open-source Llama [32], enabling large-scale generation of captions. Finally, the synthesized captions are used to guide", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "T2I models in generating images for FSL tasks.", + "bbox": [ + 513, + 409, + 828, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our major contributions are summarized as follows:", + "bbox": [ + 532, + 424, + 877, + 436 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We frame synthetic images as standalone knowledge repositories and present a CLIP adaptation methodology that pretrains on purely synthetic images before finetuning for few-shot tasks. This marks a clear departure from existing one-stage fine-tuning methods that simply treat synthetic images as complements to real images.", + "- We propose an improved Self-SL method based on DINO, specifically tailored for FSL. It introduces higher-order moments for image representation and employs synthetic augmentation for effective view construction.", + "- We develop a systematic and scalable pipeline for synthesizing both captions and images, enabling generation of large-scale base sets for pretraining and task-specific datasets. Distinct from existing arts, we leverage CoT and ICL techniques for diverse, realistic image generation." + ], + "bbox": [ + 511, + 439, + 905, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 679, + 661, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Synthetic images for CLIP-based FSL. Several works have studied CLIP adaptation for FSL tasks by using synthetic images [13, 17-19]. DISEF [17] captions few-shot real images with LLaVA [33], which alongside real images serve as inputs for T2I models. DataDream [19] fine-tunes T2I models with real images, building customized models for each dataset or category. IsSynth [18] employs the word-to-sentence T5 [34] for generating diverse captions. Outside the FSL regime, Yu et al. [35] improve ImageNet recognition with synthetic images through a contextualized and stylized diversification technique, while SynCLR [36] applies in-context learning to synthesize captions across different modes for training foundation models only with syn", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30021", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "thetic data; differently, focusing on FSL, we employ CoT and ICL techniques to synthesize image captions.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adapting CLIP for FSL. Existing CLIP adaptation methods can be roughly categorized as Prompt Tuning (PT), Adapter Tuning (AT), Encoder Tuning (ET) and Hybrid Tuning (HT). The PT methods [9, 11, 37, 38] learn textual and/or vision prompts, inserted at the input or middle layers of CLIP encoders, for adapting to few-shot tasks. ET [17-19, 39] injects few trainable parameters (e.g., LoRA [40]) into the encoders of CLIP models. AT [12, 13, 41] attaches trainable layers to the frozen backbones. Unlike PT and ET, AT is more efficient of training as no gradient backpropagation across the encoders are needed. Besides, some researchers [24, 42] study hybrid tuning that combines AT, ET or VT in varying ways. Different from them, we propose a novel two-stage methodology for CLIP adaptation.", + "bbox": [ + 89, + 125, + 483, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised learning (Self-SL). Contrastive Self-SL [28, 36, 43] uses both positive and negative samples for training with infoNCE loss, which often requires large minibatch size or memory bank. In contrast, self-distillation based Self-SL [20, 21, 44] depends only on positive samples with standard cross-entropy loss, which needs proper mechanisms to avoid model collapse. Inspired by Masked Language Modeling (MLM) [45] in NLP, researchers [27, 46, 47] introduce Masked Image Modeling (MIM) for Self-SL that reconstructs randomly masked image patches. Differently, we focus on Self-SL for CLIP adaptation; particularly, we model distributions of patch tokens alongside [CLS] token as image representation, unlike existing Self-SL that only uses single [CLS] token.", + "bbox": [ + 89, + 338, + 483, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distribution based image representation. Bilinear or covariance pooling [48], particularly equipped with square root normalization [49, 50], has achieved impressive performance in supervised learning. Therein, images are represented by second-moments or covariance matrices, which can be seen as normal distributions with zero mean. Based on Gaussian assumption, $\\mathrm{G}^2\\mathrm{DeNet}$ [25] and HGD [51] represent images or regions with normal distributions, which are embedded into the space of symmetric positive definite matrices for end-to-end learning. MP [52] proposes characteristic functions for representing distributions that are approximated by first-moment and cross-covariances. In contrast, our representation uses normalized second- and third-moments, and to the best of our knowledge, we are the first to introduce distribution modeling into Self-SL.", + "bbox": [ + 89, + 551, + 483, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 797, + 264, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our adaptation method that comprises pretraining and finetuning is shown in Figure 1. The pretraining enables the adapter to learn extensive knowledge on the iBase, while preserving original representation of CLIP. During the finetuning, we adapt the models for few-shot tasks with real and", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "synthetic images. Moreover, we develop a scalable method to synthesize both captions and images.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pretraining on Imagined Base Set", + "text_level": 1, + "bbox": [ + 511, + 128, + 808, + 145 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed HoM-DINO is an improved self-distillation method over DINO. As Figure 1a shows, it comprises a student branch and a teacher branch both connected to frozen image encoder $\\theta$ , sharing the same architecture but with different parameters. The student consists of an adapter $\\phi$ and two parallel projection heads, i.e., HoM head $\\psi_{H}$ for image representation by higher-order moments and MIM head $\\psi_{P}$ for masked image modeling. The teacher's adapter $\\widetilde{\\phi}$ and two heads $\\widetilde{\\psi}_{H}$ and $\\widetilde{\\psi}_{P}$ are exponential moving averages (EMA) of the students. Our adapter consists of 4 modified transformer blocks, and the projection heads comprise a 3-layer MLP followed by a weight-normalized linear layer; see Supplement A.1 for their detailed architectures.", + "bbox": [ + 511, + 151, + 906, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce SyntAug for building positive pairs, where the two views are synthetic images generated by T2I models conditioned on the same caption. They convey the same scenery but with natural variations thanks to the randomness of noise in the generative models. This is unlike traditional method that construct views by hand-crafted, meticulously chosen image augmentations (HandAug) [28]. We note that SynCLR [36] uses synthetic images to construct multi-positives in contrastive loss, whereas our approach uses Kullback-Leibler (KL) loss in self-distillation.", + "bbox": [ + 511, + 351, + 908, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We represent images with patch token distributions $^1$ alongside [CLS] token. Let $\\{\\mathbf{f}_i\\}_{i=1}^n, \\mathbf{f}_i \\in \\mathbb{R}^p$ represent patch tokens from the adapter $\\phi$ given a view $\\mathbf{x}$ . The first moment $\\mathbf{m}_1 = \\frac{1}{n} \\sum_i \\mathbf{f}_i$ denotes the distribution center. Since the size of k-th moment $(k \\geq 2)$ scales exponentially with p, we reduce the complexity by computing moments for individual feature components. Motivated by square root normalization for covariance matrices [49], we perform square root and cubic root for the second and third moments, respectively. The second moment $\\mathbf{m}_2 = (\\frac{1}{n} \\sum_i (\\mathbf{f}_i - \\mathbf{m}_1)^2)^{-\\frac{1}{2}}$ captures variations around the center, where $\\frac{1}{2}$ denotes element-wise square root. The third moment $\\mathbf{m}_3 = (\\frac{1}{n} \\sum_i (\\mathbf{f}_i - \\mathbf{m}_1)^3)^{-\\frac{1}{3}}$ indicates asymmetry of distribution around its center. Our final image representation $\\mathbf{r}$ is a concatenated $4p$ -dimensional vector:", + "bbox": [ + 511, + 501, + 908, + 731 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} = \\phi \\circ \\theta (\\mathbf {x}) \\triangleq \\left[ [ \\mathrm {C L S} ]; \\mathbf {m} _ {1}; \\mathbf {m} _ {2}; \\mathbf {m} _ {3} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 739, + 906, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{x}_1$ and $\\mathbf{x}_2$ be input views for the teacher and student, respectively. Then, the contrast loss for the HoM heads is:", + "bbox": [ + 511, + 766, + 906, + 797 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {H o M}} (\\phi , \\psi_ {H}) = \\mathrm {K L} (\\widetilde {\\mathbf {q}} _ {\\mathbf {x} _ {1}}, \\mathbf {q} _ {\\mathbf {x} _ {2}}) + \\mathrm {K L} (\\widetilde {\\mathbf {q}} _ {\\mathbf {x} _ {2}}, \\mathbf {q} _ {\\mathbf {x} _ {1}}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 806, + 906, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\widetilde{\\mathbf{q}}_{\\mathbf{x}_1} = \\widetilde{\\psi}_H\\circ \\widetilde{\\phi}\\circ \\theta (\\mathbf{x}_1)$ is the teacher's prediction and $\\mathbf{q}_{\\mathbf{x}_2} = \\psi_H\\circ \\phi \\circ \\theta (\\mathbf{x}_2)$ is the student's one.", + "bbox": [ + 511, + 834, + 906, + 867 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1 In Supplement D.5, we visualize the distributions, illustrating why we prefer moments over an a priori distribution such as Gaussian [25].", + "bbox": [ + 511, + 875, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30022", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, we introduce MIM to enhance capability of patch representation. Specifically, we randomly mask a portion of tokens for the view fed to the student. Suppose $\\mathbf{x}^{\\mathrm{MIM}}$ is the masked image and $\\mathcal{I}_{\\mathbf{x}^{\\mathrm{MIM}}}$ is the index set of its masked patches. Let $\\{\\widetilde{\\mathbf{p}}_{\\mathbf{x}}^{i}\\}_{i = 1}^{n} = \\widetilde{\\psi}_{P}\\circ \\widetilde{\\phi}\\circ \\theta (\\mathbf{x})$ and $\\{\\mathbf{p}_{\\mathbf{x}^{\\mathrm{MIM}}}^{i}\\}_{i = 1}^{n} = \\psi_{P}\\circ \\phi \\circ \\theta (\\mathbf{x}^{\\mathrm{MIM}})$ be predictions made for patch tokens by the teacher and student, respectively. Then, we define the MIM loss as:", + "bbox": [ + 89, + 90, + 482, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M I M}} (\\phi , \\psi_ {P}) = \\sum_ {\\mathbf {x} \\in \\left\\{\\mathbf {x} _ {1}, \\mathbf {x} _ {2} \\right\\}} \\sum_ {i \\in \\mathcal {I} _ {\\mathbf {x} ^ {\\mathrm {M I M}}}} \\operatorname {K L} \\left(\\widetilde {\\mathbf {p}} _ {\\mathbf {x}} ^ {i}, \\mathbf {p} _ {\\mathbf {x} ^ {\\mathrm {M I M}}} ^ {i}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 223, + 482, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We simply sum $\\mathcal{L}_{\\mathrm{HoM}}$ and $\\mathcal{L}_{\\mathrm{MIM}}$ to obtain the final loss.", + "bbox": [ + 89, + 271, + 464, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Fine-tuning for Low-shot Tasks", + "text_level": 1, + "bbox": [ + 89, + 297, + 367, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 1b illustrates the fine-tuning stage. Once pretraining is complete, we preserve the teacher's vision adapter for downstream tasks, while discarding its heads and the whole student branch. To tailor the CLIP models to downstream FSL tasks, we attach to the adapter a vision (V) classifier with a cross-entropy (CE) loss $\\mathcal{L}_{\\mathrm{V}}$ for vision-based recognition. Meanwhile, we tune a learnable vision-language (VL) classifier with a CE loss $\\mathcal{L}_{\\mathrm{VL}}$ [18, 53] for integration with language knowledge. The VL-classifier is initialized with a prompt ensembling technique. Specifically, for each class, we construct a set of class-specific text prompts fed to the frozen text encoder, and the text embeddings are averaged as the embedding of this class. Then, all per-class text embeddings are used to initialize the VL-classifier. The total loss is the sum of $\\mathcal{L}_{\\mathrm{V}}$ and $\\mathcal{L}_{\\mathrm{VL}}$ . For inference, we average the logits from two classifiers with a fusion weight, following Tip-Adapter (f) [12] and CaFo [13].", + "bbox": [ + 89, + 320, + 483, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ImagineFSLLoRA. Additionally, we can further tune the image encoder, e.g., by introducing LoRA [17, 19, 39]. The resulting method is dubbed ImagineFSLLoRA. We note that such encoder tuning (ET) methods generally lead to significant performance gains, compared to those with frozen encoders. As such methods can adjust features from low-layers up to high-layers across the encoders, so can better adapt to specific tasks and domains.", + "bbox": [ + 89, + 578, + 483, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Zero-shot Recognition. Our method is applicable to zero-shot recognition tasks. In this setting, we perform fine-tuning using solely the synthetic images, without touch of any real images. By tuning CLIP to diverse synthetic images of categories of interest, our method can significantly boost zero-shot recognition, thanks to its capability to learn additional knowledge from foundation models including LLMs and T2I models.", + "bbox": [ + 89, + 699, + 483, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Synthesizing both Captions and Images", + "text_level": 1, + "bbox": [ + 89, + 832, + 429, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The synthesizing pipeline is shown in Figure 2, where the CoT and ICL techniques are used to instruct LLMs and T2I models to generate captions and images.", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7b53f478f544ebcede3801748fb603a5e57cb08ba88db1934676770c36a08e9b.jpg", + "image_caption": [ + "Figure 2. Our pipeline synthesizes captions & images. We leverage GPT with the CoT to analyze key factors defining image descriptions and generate exemplar captions, which are then used by Llama for diverse caption generation via ICL. Finally, SD model generates images based on the synthesized captions. Note our pipeline is not limited to closed-source GPT (see Supplement D.4.)" + ], + "image_footnote": [], + "bbox": [ + 517, + 92, + 885, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Synthesizing Captions. We identify five key factors that constitute image descriptions: attribute, viewpoint, background (BG), lighting condition (LC), and cause of degradation of photos (CD). We design four distinct patterns for systematic caption generation, incorporating various combinations of these factors. The first one, Base-pattern, comprises attribute and viewpoint, providing a fundamental description of the concept. The rest ones, i.e., BG-pattern, LC-pattern, and CD-Pattern, enhance the Base-pattern by integrating additional BG, LC, and CD factor, respectively, enriching the depiction comprehensively.", + "bbox": [ + 511, + 357, + 906, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pipeline begins by querying GPT-4 to analyze the key factors associated with a given concept, generating a detailed summary of visual properties and foundational elements essential for synthesizing diverse, realistic images that mimic real-world photography. We then design prompt template (PTe), guiding GPT-4 to generate a small number of high-quality exemplary captions for each pattern. Given the cost and time constraints of accessing GPT-4 API, we delegate extensive caption generation tasks to the locally deployed, lightweight Llama model. We develop PTe for Llama that utilizes exemplary captions generated by GPT-4 as in-context examples, randomly selecting one of the four patterns and the corresponding factors and examples to instantiate these templates. This strategy aims to elicit rich and nuanced responses from Llama, facilitating the creation of contextually enriched captions.", + "bbox": [ + 511, + 523, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Synthesizing Images. We use Stable Diffusion 3 (SD3) Medium [54], a multimodal diffusion transformer, for image generation. We construct the iBase, SyntIN1K, by adopting ImageNet-1K [55] class names as concepts. For each concept, we synthesize 300 captions, generating four images per caption. Additionally, we build SyntIN1.5K by selecting 0.5K additional concepts from ImageNet-21K class names. For each downstream dataset, we synthesize 300 captions per class and generate one image per caption,", + "bbox": [ + 511, + 763, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30023", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/9451b3f2cc838fe11bcfbda7f71b5fd0ec7815212fb03fb612634d391a06663f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodT2IENSyn DataImageNetCaltechAircraftCarsFoodPetsFlowersDTDEuroSATSUNAvg Acc
IsSynth [18]**[18]-/73.4-/96.0-/46.7-/82.6-/87.3-/92.8-/86.5-/73.6-/87.1-/76.2-/80.2
CaFo [13]**[13]70.2/73.193.9/96.930.1/46.271.5/85.485.6/87.691.5/94.974.9/97.659.4/72.066.5/86.371.2/76.971.5/81.7
\\( \\text{IsSynth}^{\\dagger} \\)**Ours70.7/73.994.4/96.831.1/46.571.7/85.485.7/87.491.6/93.885.4/95.760.1/73.975.9/87.270.9/76.573.8/81.7
\\( \\text{CaFo}^{\\dagger} \\)**Ours70.2/73.994.2/96.930.6/47.472.6/85.786.1/87.691.8/94.976.1/97.860.1/72.568.7/86.771.3/76.972.2/82.0
ImagineFSL**Ours71.6/74.795.3/97.131.5/54.882.6/90.586.7/87.892.3/94.387.0/98.762.7/75.876.2/89.672.0/78.175.8/84.1
DISEF [17]**[17]-/73.9-/96.9-/63.9-/88.6-/87.1-/94.3-/98.9-/75.4-/94.3-/77.4-/85.1
DataDream [19]**[19]-/74.1-/96.9-/72.3-/92.4-/87.6-/94.8-/99.4-/81.6-/93.4-/77.5-/87.0
\\( \\text{DISEF}^{\\dagger} \\)**Ours71.0/74.093.2/97.032.0/67.972.6/91.286.3/87.292.5/94.588.0/98.960.5/76.082.3/94.370.6/77.174.9/85.8
\\( \\text{ImagineFSL}_{\\text{LoRA}} \\)**Ours71.8/75.295.4/97.934.0/74.182.8/92.986.8/88.392.8/95.490.2/99.763.7/78.082.7/95.072.8/78.877.3/87.5
", + "bbox": [ + 93, + 88, + 903, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison of 1-shot/16-shot recognition to prior methods using synthetic images. Bold: best results; underlined: second-best results. ${}^{ \\dagger }$ Reproduced by us.", + "bbox": [ + 89, + 233, + 906, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "producing task-specific synthetic datasets.", + "bbox": [ + 89, + 271, + 370, + 285 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Supplements B.1-B.3, we give details on our design philosophy regarding the factors, the queries for GPT-4 to analyze these factors, the PTe for GPT-4 to generate examples, and the PTe for in-context learning of Llama. In Table S-5, we present a summary of all synthetic datasets.", + "bbox": [ + 88, + 286, + 482, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 375, + 215, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate extensively our methods on few-shot recognition, domain generalization, and zero-shot recognition. Besides, we conduct a thorough ablation study on our method. Across the paper, we use CLIP ViT-B/16 as the visual encoder and its aligned counterpart as the textual encoder, unless otherwise specified. All experiments are conducted using PyTorch package [56] with GeForce RTX 4090 GPUs. Additional experiments can be found in Supplement D.", + "bbox": [ + 88, + 400, + 482, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setting", + "text_level": 1, + "bbox": [ + 89, + 531, + 183, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Metrics. Following prior arts [11, 12], we benchmark few-shot and zero-shot recognition tasks on 11 datasets, including ImageNet [55], Caltech [57], Aircraft [58], Cars [59], Food [60], Pets [61], Flowers [62], DTD [63], EuroSAT [64], SUN [65], and UCF101 [66]. We assess domain generalization task on ImageNet-V2 [67], ImageNet-S [68], ImageNet-A [69], and ImageNet-R [70]. For meta-testing, we follow the common practice, randomly sampling 3 All-way-K-shot tasks per dataset, reporting the average accuracy (Avg Acc) as a percentage (\\%).", + "bbox": [ + 89, + 551, + 482, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pretraining setup. We train the models using the AdamW optimizer with a cosine scheduler of learning rate. As in DINO [20], we use image augmentations including random resized crop, horizontal flip, color jittering, grayscale conversion, and Gaussian blur. For MIM, we utilize a blockwise masking strategy [27], where each image is randomly masked and the proportion of masked patches is randomly sampled. Hyperparameters are given in Supplement A.2.", + "bbox": [ + 89, + 704, + 482, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine-tuning Setup. We follow the implementation of DataDream [19], adopting random resized crop, random horizontal flip, a set of color augmentations and RandAugment [71] as image augmentations for both ImagineFSL and ImagineFSLLORA. As in IsSynth [18], DISEF and", + "bbox": [ + 88, + 824, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DataDream, we use mix training strategy that simultaneously uses real images and synthetic images in a batch. For input of the text encoder, we adopt an ensemble of textual templates hand-engineered by CLIP and Customized Prompts via Language models (CuPL) [72] provided by CaFo [13], unless otherwise specified. Following TipAdapter(f) [12], CaFo, DISEF and DataDream, we tune hyperparameters for each task on the validation sets. See Supplement A.3 for detailed hyperparameters of fine-tuning.", + "bbox": [ + 511, + 271, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLMs and T2I models. To analyze the factors and generate the exemplary captions, we call GPT-4 API with a temperature of 0.5. Llama 3 8B is deployed locally to generate extensive captions, in which we set the temperature to 0.5 and Nucleus Sampling parameter to 0.7. We utilize a TensorRT optimized SD3 Medium model and generate $512 \\times 512$ images, with the scale of classifier-free guidance set to 5 and the steps of denoising diffusion implicit model to 28. Similar to [17-19], we adopt a semantic filtering method that discards synthetic images misclassified by zero-shot CLIP classifier. We provide implementation details on synthesizing texts and images in Supplement B.", + "bbox": [ + 511, + 407, + 903, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Few-shot Recognition", + "text_level": 1, + "bbox": [ + 511, + 597, + 715, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After pretraining on iBase, we perform meta-testing on downstream few-shot tasks. For each task, we fine-tune the adapter and V-classifier alongside VL-classifier, using K-shot real images and task-specific synthetic images; finally, we conduct inference on test sets.", + "bbox": [ + 511, + 618, + 903, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Contrast with Synthetic Images based Methods", + "text_level": 1, + "bbox": [ + 511, + 700, + 887, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 compares to prior methods based on synthetic images in 1-/16-shot settings, including IsSynth [18], CaFo [13], DISEF [17], and DataDream [19]. For a fair comparison, we reproduce them with our synthetic images.", + "bbox": [ + 511, + 719, + 903, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we consider the methods that tune neither T2I models nor ENcoders (EN). Note that our reproduced results for IsSynth and CaFo in 16-shot setting are both higher than those reported in the original papers; we conjecture the reason is that our synthetic images are more suitable for few-shot tasks. Using the same synthetic data as ours, ImagineFSL performs better than them by a large margin of $2.0\\%$ in average accuracy in both 1-shot and 16-shot settings.", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30024", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/474b07c66c716a44661816038cac309291a5ccf0b50055a1f484947dcd1ddfc4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method1-shot2-shot4-shot8-shot16-shotAvg Acc
PTCoOp [11]67.670.874.076.880.073.8
CoCoOp [9]68.870.471.773.875.472.0
MaPLe [73]66.470.573.576.478.673.1
ProGrad [74]69.572.474.777.479.974.8
PLOT++ [75]70.774.076.979.682.176.7
PromptSRC [76]72.375.378.480.782.977.9
ATCLIP-Adapter [41]65.766.867.771.575.569.4
TIP-Adapter(f) [12]70.973.775.678.380.775.8
TaskRes [77]70.872.974.778.480.875.5
ImagineFSL76.177.579.482.284.479.9
ETCLIP-LoRA [39]72.575.577.480.383.077.7
ET+PTLLaMP [42]72.475.978.881.683.878.5
PT+ATGalLoP [24]72.876.479.182.284.579.0
ET+ATImagineFSLLoRA77.679.181.984.287.682.1
", + "bbox": [ + 93, + 89, + 480, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of $K$ -shot recognition to prior methods using only real images. **Bold:** best results; underlined: second-best results. **Supplement C** provides detailed results on 11 datasets.", + "bbox": [ + 89, + 299, + 482, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we contrast the methods that tune CLIP encoders. DISEF and DataDream use the same technique to tune the image and text encoders with LoRA, while DataDream additionally tunes T2I models. In the 16-shot setting, our ImagineFSL $_{\\text{LoRA}}$ outperforms DataDream on 10 out of 11 datasets; DataDream excels only on DTD, where we conjecture that fine-tuning T2I models with real images enhances realistic texture generation. Furthermore, with our synthetic images, the reproduced results for DISEF in the 16-shot setting surpass those reported in the original paper, while ImagineFSL $_{\\text{LoRA}}$ also exceeds DISEF by $2.4\\% / 1.7\\%$ in 1-shot/16-shot tasks.", + "bbox": [ + 88, + 354, + 482, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Contrast with Methods Using Only Real Images", + "text_level": 1, + "bbox": [ + 89, + 544, + 470, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 2, we compare our methods to prior techniques that do not employ synthetic images for K-shot tasks, including prompt tuning (PT), adapter tuning (AT), encoder tuning (ET), and hybrid tuning approaches. ImagineFSL significantly outperforms PT, AT, and ET across all shot scenarios, particularly in fewer-shot settings, e.g., it surpasses CLIP-LoRA (the runner-up) by $2.0\\%$ in 2-shot and $3.6\\%$ in 1-shot settings. Additionally, while GalLoP introduces a novel dropout technique for learning global and local prompts via a single linear layer adapter, and LLaMP merges PT and ET using LoRA, ImagineFSL-LoRA leads consistently and substantially in each K-shot setting with an average accuracy improvement of $3.1\\%$ .", + "bbox": [ + 89, + 564, + 482, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Domain Generalization", + "text_level": 1, + "bbox": [ + 89, + 771, + 305, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 compares our methods with prior approaches on the domain generalization task. In this task, a model is trained on 16 shots per class from ImageNet (IN, source) and tested on four target datasets: IN-V2, IN-S, IN-A, and IN-R. For our methods, we specifically synthesize images for IN-S and IN-R datasets, while synthetic images for ImageNet are used for IN-V2 and IN-A. We fine-tune the models using", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "both the real and synthetic images.", + "bbox": [ + 511, + 90, + 743, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we observe that our methods outperform the second-best method (i.e., CoCoLe) by over $1.3\\%$ in average accuracy. Notably, our models achieve performance that matches the top performer, GalLoP, on IN-V2 and Co-CoLe on IN-A. Furthermore, we establish new state-of-the-art results on both IN-S and IN-R, achieving accuracies of $53.8\\%$ and $80.0\\%$ , respectively.", + "bbox": [ + 511, + 107, + 906, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Zero-shot Recognition", + "text_level": 1, + "bbox": [ + 511, + 223, + 720, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare to previous methods on zero-shot recognition in Table 4. We see that our methods are significantly better than most of the PT methods on each individual datasets, including VCD [79], Sus-X [80], TPT [81] and DiffTPT [82]. Compared to state-of-the-art DMN [83], ImagineFSL and ImagineFSLoRA outperform by $1.8\\%$ and $2.5\\%$ in average accuracy, respectively, with the latter ranking first in 7 out of 11 datasets. These results suggest that synthetic images hold promise for enhancing CLIP zero-shot recognition, and our methods are likely to improve further with advancements in generative T2I models.", + "bbox": [ + 511, + 244, + 906, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Empirical Complexity", + "text_level": 1, + "bbox": [ + 511, + 421, + 720, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretraining. On SyntIN1K, pretraining with ViT-B/16 takes $\\sim 9$ hours on 4 RTX 4090 GPUs ( $\\sim 40$ hours for ViT-L/14). This task-agnostic stage is performed once and then widely fit for downstream tasks.", + "bbox": [ + 511, + 444, + 905, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fine-tuning. We compare methods on UCF101 in the 16-shot setting with a batch size of 128 to ensure a fair comparison. We apply LoRA with a rank of 16 across methods involving encoder tuning (ET). Specifically, for DISEF, DataDream, and ImagineFSLloRA, LoRA is applied to the image encoder. Additionally, for DISEF and DataDream, LoRA is also applied to the text encoder. We note that DISEF and DataDream share similar empirical complexity, as they use the same technique for fine-tuning.", + "bbox": [ + 511, + 505, + 905, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5 presents the results evaluated on one RTX 4090 GPU, with train and test latency per iteration. Among methods that do not involve $ET$ , ImagineFSL has comparable latency and memory usage to TIP-Adapter(f) and IsSynth. While it has more parameters, they make up only $5.08\\%$ of total encoder parameters. Moreover, it outperforms CoOp significantly in training speed. Compared to methods that involve $ET$ , i.e., DISEF and DataDream, ImagineFSL $_{\\text{LoRA}}$ achieves a two-fold improvement in training speed and consumes only half the memory, despite a small increase in the percentage of parameters (from $1.07\\%$ to $5.89\\%$ ).", + "bbox": [ + 511, + 641, + 906, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 816, + 666, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct ablation analysis in 1-shot/16-shot settings on ImageNet, Aircraft, Flowers and EuroSAT with CLIP ViTB/16 as visual encoder, unless otherwise specified. We provide more ablations in Supplement D.2.", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30025", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8af7a58559fe92527ac553bf3c07489ea747b0bee788033dc843082464cccb8c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSourceTarget
ImageNet-V2-S-A-RAvg Acc
PTCoOp [11]71.564.248.049.775.259.3
CoCoOp [9]71.064.148.850.676.259.9
MaPLe [73]70.764.149.250.977.060.3
PromptSRC [76]71.364.449.650.977.860.7
CoCoLe [14]73.965.950.951.878.961.9
ATTaskRes [77]73.165.349.150.477.760.6
GraphAdapter [78]73.465.649.250.677.760.8
ImagineFSL74.767.052.751.579.762.7
ET+ATImagineFSLloRA75.267.553.851.580.063.2
PT+ATGalLoP [24]75.167.549.550.377.861.3
", + "bbox": [ + 106, + 88, + 436, + 253 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/582ed57027777af0a173454f4a022026f37b7cd3656dda8aa0ad693e99ac06a3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodImageNetCiteLectAircraftCarsFoodPetsFlowersETDEmoSATSYNUCF101AvgAcc
PTVCD [79]68.0-----88.586.9-45.648.8-----
Sus-X [80]69.994.028.766.186.190.673.854.657.567.766.667.766.668.768.768.7
TPT [81]69.094.224.866.984.787.869.047.842.465.568.065.568.065.565.565.5
DiffTPT [82]70.392.525.667.087.288.270.147.043.165.768.268.265.965.965.965.9
DMN [83]72.395.330.068.085.192.074.555.959.470.272.570.570.570.570.570.5
CODER [84]71.5----89.892.055.760.5-------
ATImagineFSL70.994.730.278.686.592.373.959.265.670.373.272.372.372.372.372.3
ET+ATImagineFSL-LoRA71.094.730.478.786.592.373.660.871.770.373.373.373.073.073.073.0
", + "bbox": [ + 454, + 89, + 911, + 255 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f2819a3f569d99b40a610a7a2a47df832c0772937546071f0d7dea02e2d2ff24.jpg", + "table_caption": [ + "Table 3. Results of domain generalization on 16-shot setting." + ], + "table_footnote": [], + "table_body": "
MethodLatencyMemory (G)Params (M)(%)
Train (ms)Test (ms)
PTCoOp [11]298.360.43.050.010.01
-TIP-Adapter(f) [12]70.261.52.730.830.55
ATIsSynth [18]69.960.32.740.050.03
ImagineFSL93.165.43.007.615.08
ETDISEF [17]544.360.820.641.571.07
DataDream [19]
ET+ATImagineFSLLoRA265.465.511.828.825.89
", + "bbox": [ + 96, + 289, + 478, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Comparison of empirical complexity of fine-tuning. We evaluate all methods on a single RTX 4090 GPU.", + "bbox": [ + 89, + 431, + 482, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Is pretraining necessary? Table 6a compares the Non-Pretraining (NoP) setting with the pretraining setting. In NoP, we directly train the vision adapter and vision (V) classifier using either real images alone (row 1) or a mix of synthetic and real images (row 2). For the pretraining setting, we first train the vision adapter via either SL (row 3) or Self-SL on SyntIN-1K (row 4), then fine-tune the vision adapter and V-classifier. During testing, we combine the V-classifier's predictions with CLIP zero-shot predictions (i.e., the frozen VL-classifier), using CLIP's handcrafted templates as input to the text encoder.", + "bbox": [ + 88, + 474, + 482, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In NoP, mixing real and synthetic images substantially improves performance compared to using real images alone, demonstrating that our synthetic images effectively mitigate data scarcity. Furthermore, compared to NoP, pretraining with SL yields notable gains, particularly in 1-shot average accuracy $(+0.9\\%)$ . Additionally, our Self-SL method (i.e., HoM-DINO) surpasses SL by $2.3\\% / 1.1\\%$ in 1-shot/16-shot average accuracy, indicating that Self-SL better learns transferable features from synthetic to real image domains. These results highlight the importance of self-supervised pretraining on SyntIN1K for capturing rich knowledge from diverse synthetic images. In Figure S-3 of Supplement D.5, we visualize self-attention maps of different methods.", + "bbox": [ + 89, + 642, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HoM-DINO against other Self-SL methods. Table 6b compares to three typical Self-SL approaches including DINOv2, SynCLR and MAE. DINOv2, relying only on positive examples, is a self-distillation method that introduces", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Results of zero-shot recognition on 11 datasets.", + "bbox": [ + 506, + 266, + 843, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MIM loss into DINO loss. SynCLR, depending on both positive and negative samples, is a contrastive method that combines multi-positive contrastive loss [85] and MIM loss. MAE is a representative masked image modeling method.", + "bbox": [ + 511, + 291, + 903, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It can be seen that MAE clearly underperforms and we conjecture its reconstruction mechanism in the pixel space is not fit for the few-shot scenarios. SynCLR achieves very competitive performance but still nontrivially lags behind ours. Our HoM-DINO builds upon and outperforms DINOv2 by $2.1\\% / 0.7\\%$ in 1-shot/16-shot accuracy.", + "bbox": [ + 511, + 353, + 903, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How image representation (image-repr) and view construction impact? First, from Table 6c we observe that, with [CLS] token as image-repr and HandAug, MIM significantly improves over without MIM (rows 1 and 2), indicating prediction of masked patches is helpful for FSL. Second, incorporating local patches via average pooling of all patches (AvgP) and attention pooling of all tokens (AttnP) slightly improved performance (rows 4 and 5). Using distribution-based image representations further improves performance significantly (rows 6 and 8); meanwhile, our HoM is slightly better than $\\mathrm{G}^2\\mathrm{DeNet}$ while having one fourth less dimension (Dim). SyntAug further improves HoM (rows 8 and 9). However, MP underperforms despite its usage of moments, which may be due to its lack of structural normalization mechanism that has proven to be critical for high-order representations [86]. Finally, as dimension of HoM is higher than [CLS] token, for a fair comparison, we add a linear layer after the adapter that increases [CLS] token to 2048-Dim. However, this remains significantly inferior to HoM (rows 3 and 8), suggesting that advantage of HoM lies in exploiting patches via statistical moments.", + "bbox": [ + 511, + 446, + 905, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Influence of textual prompts. As shown in Table 6d, when used individually, the CuPL [72] slightly outperforms the hand-engineered prompt templates of CLIP. However, combining both CuPL and CLIP prompts results in a further improvement over the individual methods. Therefore, unless otherwise specified, we use an ensemble of CLIP and CuPL prompts throughout the paper.", + "bbox": [ + 511, + 763, + 903, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VL-classifier (clf) tuning and encoder tuning. Rather than using a frozen VL-clf, we tune the clf's weights (i.e.,", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30026", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ca6c5cc22f9a019b7c285e45172904ab442ae54960be9b683f262bd2644e1571.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PretrainSyn ImgImageNetAircraftFlowersEuroSATAvg Acc
1-✘68.8/72.824.1/48.578.9/96.560.4/84.458.1/75.6
2-✓70.9/73.927.5/49.384.1/96.566.2/86.262.2/76.5
3SL✓71.0/73.928.7/49.585.3/97.067.3/86.863.1/76.8
4HoM-DINO✓71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
", + "bbox": [ + 91, + 88, + 504, + 162 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cafc4673a8da01719e3a672584a9775c5f5209e5ff9c493cc4540777420475b7.jpg", + "table_caption": [ + "(a) Necessity of pretraining." + ], + "table_footnote": [], + "table_body": "
Image \nReprDimViewMIMImageNetAircraftFlowersEuroSATAvg Acc
1[CLS]512HA×71.2/73.527.2/50.083.9/97.266.7/86.962.3/76.9
2[CLS]512HA✓71.1/73.928.9/50.685.2/97.567.8/86.863.3/77.2
3[CLS]2048HA✓71.4/73.629.0/49.984.3/97.669.8/86.863.6/77.0
4AvgP512HA✓71.4/73.729.2/50.384.2/97.469.6/87.163.6/77.1
5AttnP512HA✓71.3/73.228.8/56.985.7/97.669.4/87.463.8/77.2
6G²De- \nNet8385HA✓71.1/73.629.3/50.385.3/97.570.2/87.764.0/77.3
7MP2087HA✓71.3/73.629.3/48.982.7/97.468.6/86.263.0/76.5
8HoM2048HA✓71.4/74.029.1/50.985.6/98.172.0/87.364.5/77.6
9HoM2048SA✓71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
", + "bbox": [ + 91, + 179, + 504, + 337 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4662d839ee673b3e5defa356e29277587d9870d7aaf8d1f220c8c93773d10460.jpg", + "table_caption": [ + "(c) Effect of image representation and view. HA: HandAug; SA: SyntAug." + ], + "table_footnote": [ + "(f) Scaling of ImagineFSL versus data size and model capacity." + ], + "table_body": "
ModelBase setSizeImageNetAircraftFlowersEuroSATAvg Acc
ViT-B/16SyntIN1K0.1M71.1/74.431.0/54.887.0/98.874.8/88.166.0/79.0
0.3M71.6/74.731.5/54.887.0/98.776.2/89.666.6/79.5
SyntIN1.5K0.45M71.7/74.732.9/54.787.7/98.675.2/89.766.9/79.4
ViT-L/14SyntIN1K0.3M78.4/80.742.1/68.190.0/99.580.6/92.772.8/85.3
", + "bbox": [ + 93, + 354, + 506, + 436 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bf211e990b9229cb3d139862d0571ce7636f8d39c3b0f928e137ce039c5e90ce.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Self-SLImageNetAircraftFlowersEuroSATAvg Acc
DINOv271.1/73.928.9/50.685.2/97.567.8/86.863.3/77.2
SynCLR71.5/74.027.4/51.084.4/97.774.4/86.964.4/77.4
MAE71.2/74.326.4/47.683.6/97.466.5/86.761.9/76.3
HoM-DINO71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
", + "bbox": [ + 526, + 89, + 900, + 161 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f24deb67dfebf83ee602bdf7c8a53b9bdaa696e506229e0056597cbcb38071dc.jpg", + "table_caption": [ + "(b) Different Self-SL techniques." + ], + "table_footnote": [], + "table_body": "
CLIPCuPLImageNetAircraftFlowerEuroSATAvg Acc
✓✗71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
✗✓71.6/74.330.7/51.685.9/98.472.0/87.865.0/78.0
✓✓71.6/74.330.8/51.586.3/98.373.7/87.865.6/78.0
", + "bbox": [ + 526, + 181, + 900, + 247 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4dc443b9c2890251d32b0bbc2b53c77afa91be4ae45f1ad4146b0cc3e2e0dfa8.jpg", + "table_caption": [ + "(d) Influence of textual prompts." + ], + "table_footnote": [], + "table_body": "
V-clfVL-clfI-ENImageNetAircraftFlowersEuroSATAvg Acc
71.6/74.330.8/51.586.3/98.373.7/87.865.6/78.0
71.6/74.731.5/54.887.0/98.776.2/89.666.6/79.5
71.8/75.234.0/74.190.2/99.782.7/95.069.7/86.0
", + "bbox": [ + 526, + 267, + 900, + 334 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a9c2b0a8f0f84d4b2432adf75ef6953b758c19b5750056b4ea0a6c908356cb2e.jpg", + "table_caption": [ + "(e) Effects of tuning VL-classifier (clf) and Image ENcoder (I-EN)." + ], + "table_footnote": [ + "(g) Effect of Chain-of-Thought (CoT) on image generation." + ], + "table_body": "
MethodChain-of-ThoughtFidelity ↓Diversity ↑
Handcrafted prompts×45.920.95
Ours×40.071.06
Ours✓32.591.21
", + "bbox": [ + 526, + 352, + 900, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study on the proposed methods in the 1-shot/16-shot settings. Bold: best results; underlined: second-best results.", + "bbox": [ + 106, + 460, + 887, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "per-class text embeddings), implementing the ImagineFSL method. Table 6e shows this tuning significantly improves performance across most datasets. Furthermore, tuning the Image Encoder (I-EN) with LoRA (i.e., ImagineFSL $_{\\text{LoRA}}$ ) further enhances performance, particularly on Aircraft, Flowers, and EuroSAT. We hypothesize that I-EN tuning refines feature representations across the entire network, leading to better task and domain adaptation.", + "bbox": [ + 88, + 487, + 482, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scaling capability. As shown in Table 6f, when using ViT-B/16 as the image encoder on SyntIN1K, we observe slight improvements in both 1-shot and 16-shot accuracies as data volume increases from 0.1M to 0.3M. Additionally, expanding the dataset to SyntIN1.5K slightly enhances 1-shot accuracy while 16-shot accuracy remains stable, suggesting potential saturation in performance. This could be attributed to the limitations of the frozen encoder, which might not effectively utilize the expanded dataset or additional concepts. However, switching to ViT-L/14 results in substantial gains of $5.9\\%$ and $5.8\\%$ in 1-shot and 16-shot accuracies, respectively, underscoring the scalability of our approach with larger CLIP models.", + "bbox": [ + 88, + 611, + 482, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How does CoT affect generated images? We ablate our generation method by comparing settings w/ and w/o CoT prompting. We also compare against a baseline using handcrafted prompts, i.e., \"a photo of a {class}\", for image generation. In the CoT (w/o) setting, we prompt Llama to generate captions using factor names for individual patterns. As", + "bbox": [ + 88, + 809, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in [87], we compute the FID score to measure fidelity between synthetic and real images on the ImageNet validation set. For diversity, we use a pretrained DINOv2 ViT-B/16 to extract features and compute the standard deviation per class. Table 6g shows that our method w/o CoT outperforms handcrafted prompts. Using CoT further improves diversity and fidelity, underscoring its role in generating diverse, realistic images. Supplement D.5 qualitatively analyzes our generated texts and images.", + "bbox": [ + 511, + 487, + 906, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 511, + 657, + 769, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce ImagineFSL, a novel methodology for CLIP adaptation. It involves self-supervised pretraining on large-scale purely synthetic images, followed by fine-tuning with real and task-specific synthetic images. Unlike previous methods, we treat synthetic images as standalone knowledge repositories of diverse concepts. To generate diverse and realistic synthetic images, we leverage CoT and in-context learning techniques, utilizing the capabilities of LLMs and T2I models. Despite strong performance across various recognition tasks, our approach has certain limitations, such as pretraining overhead and underutilization of synthetic captions, which are discussed in Supplement E. We hope our work inspires further exploration of synthetic images for data-scarce tasks.", + "bbox": [ + 509, + 688, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30027", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yisheng Song, Ting Wang, Puyu Cai, Subrota K Mondal, and Jyoti Prakash Sahoo. A comprehensive survey of few-shot learning: Evolution, applications, challenges, and opportunities. ACM Computing Surveys, 55(13s):1-40, 2023. 1", + "[2] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. In International Conference on Learning Representations, 2019. 1", + "[3] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In European Conference on Computer Vision, pages 266-282, 2020. 1", + "[4] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, pages 1126-1135, 2017. 1", + "[5] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in Neural Information Processing Systems, pages 3630-3638, 2016. 1", + "[6] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millicah, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 1", + "[7] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021.", + "[8] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 1", + "[9] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 3, 6, 7", + "[10] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23681-23690, 2024. 1", + "[11] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 3, 5, 6, 7" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. TipAdapter: Training-free adaptation of CLIP for few-shot classification. In European Conference on Computer Vision, pages 493-510, 2022. 3, 4, 5, 6, 7", + "[13] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15211-15222, 2023. 2, 3, 4, 5", + "[14] Yi Zhang, Ke Yu, Siqi Wu, and Zhihai He. Conceptual codebook learning for vision-language models. In European Conference on Computer Vision, pages 235-251, 2024. 1, 7", + "[15] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804, 2022. 1", + "[16] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1", + "[17] Victor G Turrisi da Costa, Nicola Dall'Asen, Yiming Wang, Nicu Sebe, and Elisa Ricci. Diversified in-domain synthesis with efficient fine-tuning for few-shot classification. arXiv preprint arXiv:2312.03046, 2023. 1, 2, 3, 4, 5, 7", + "[18] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? In International Conference on Learning Representations, 2023. 2, 4, 5, 7", + "[19] Jae Myung Kim, Jessica Bader, Stephan Alaniz, Cordelia Schmid, and Zeynep Akata. DataDream: Few-shot guided dataset generation. In European Conference on Computer Vision, pages 252-268, 2024. 1, 2, 3, 4, 5, 7", + "[20] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1, 3, 5", + "[21] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOV2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024:1-31, 2024. 1, 3", + "[22] Carl Doersch, Ankush Gupta, and Andrew Zisserman. CrossTransformers: spatially-aware few-shot transfer. In Advances in Neural Information Processing Systems, pages 21981-21993, 2020. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30028", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. DeepEMD: Differentiable earth mover's distance for few-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5632-5648, 2022.", + "[24] Marc Lafon, Elias Ramzi, Clément Rambour, Nicolas Audebert, and Nicolas Thome. GalLoP: Learning global and local prompts for vision-language models. In European Conference on Computer Vision, 2024. 2, 3, 6, 7", + "[25] Qilong Wang, Peihua Li, and Lei Zhang. $\\mathbf{G}^2\\mathrm{DeNet}$ : Global Gaussian distribution embedding network and its application to visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6507-6516, 2017. 2, 3", + "[26] Jun Shao. Mathematical Statistics. Springer, New York, 2003. 2", + "[27] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. iBOT: Image BERT pretraining with online tokenizer. In International Conference on Learning Representations, 2022. 2, 3, 5", + "[28] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 2, 3", + "[29] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, pages 24824-24837, 2022. 2", + "[30] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901, 2020. 2", + "[31] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floreccia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2", + "[32] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2", + "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 49250-49267, 2023. 2", + "[34] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 2", + "[35] Zhuoran Yu, Chenchen Zhu, Sean Culatana, Raghuraman Krishnamoorthi, Fanyi Xiao, and Yong Jae Lee. Diversify, don't fine-tune: Scaling up visual recognition training with synthetic images. arXiv preprint arXiv: 2312.02253, 2024. 2" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Yonglong Tian, Lijie Fan, Kaifeng Chen, Dina Katabi, Dilip Krishnan, and Phillip Isola. Learning vision from models rivals learning vision from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15887-15898, 2024. 2, 3", + "[37] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1401-1411, 2023. 3", + "[38] Adrian Bulat and Georgios Tzimiropoulos. LASP: Text-to-text optimization for language-aware soft prompting of vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23232-23241, 2023. 3", + "[39] Maxime Zanella and Ismail Ben Ayed. Low-rank few-shot adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1593-1603, 2024. 3, 4, 6", + "[40] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 3", + "[41] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. CLIP-Adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2023. 3, 6", + "[42] Zhaoheng Zheng, Jingmin Wei, Xuefeng Hu, Haidong Zhu, and Ram Nevatia. Large language models are good prompt learners for low-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28453-28462, 2024. 3, 6", + "[43] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 3", + "[44] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284, 2020. 3", + "[45] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4171–4186, 2019. 3", + "[46] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pretraining of image transformers. In International Conference on Learning Representations, 2022. 3", + "[47] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Con-" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30029", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 159-168, 2022. 3", + "[48] Tsung-Yu Lin, Aruni RoyChowdhury, and Subhransu Maji. Bilinear cnn models for fine-grained visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1449-1457, 2015. 3", + "[49] Peihua Li, Jiangtao Xie, Qilong Wang, and Zilin Gao. Towards faster training of global covariance pooling networks by iterative matrix square root normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 947-955, 2018. 3", + "[50] Yue Song, Nicu Sebe, and Wei Wang. On the eigenvalues of global covariance pooling for fine-grained visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3554-3566, 2023. 3", + "[51] Tetsu Matsukawa, Takahiro Okabe, Einoshin Suzuki, and Yoichi Sato. Hierarchical Gaussian descriptors with application to person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(9):2179–2194, 2020. 3", + "[52] Mingze Gao, Qilong Wang, Zhenyi Lin, Pengfei Zhu, Qinghua Hu, and Jingbo Zhou. Tuning pre-trained model via moment probing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11769-11779, 2023. 3", + "[53] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 4", + "[54] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633, 2024. 4", + "[55] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 4, 5", + "[56] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035, 2019. 5", + "[57] Li Fei-Fei, R Fergus, and P Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Worskshops, pages 178–178, 2004. 5", + "[58] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classi" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fication of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5", + "[59] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 554–561, 2013. 5", + "[60] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, pages 446–461, 2014. 5", + "[61] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3498-3505, 2012. 5", + "[62] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5", + "[63] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 5", + "[64] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5", + "[65] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 5", + "[66] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5", + "[67] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In International Conference on Machine Learning, pages 5389-5400, 2019. 5", + "[68] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 5", + "[69] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2019. 5", + "[70] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 5", + "[71] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. RandAugment: Practical automated data augmentation with" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "30030", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "a reduced search space. In Advances in Neural Information Processing Systems, pages 18613-18624, 2020. 5", + "[72] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15645-15655, 2023. 5, 7", + "[73] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. MaPLe: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 6, 7", + "[74] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 6", + "[75] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. In International Conference on Learning Representations, 2023. 6", + "[76] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 6, 7", + "[77] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10899-10909, 2023. 6, 7", + "[78] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. GraphAdapter: Tuning vision-language models with dual knowledge graph. In Advances in Neural Information Processing Systems, pages 13448-13466, 2023. 7", + "[79] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In International Conference on Learning Representations, pages 1-17, 2023. 6, 7", + "[80] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. SuS-X: Training-free name-only transfer of vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2725-2736, 2023. 6, 7", + "[81] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. In Advances in Neural Information Processing Systems, pages 14274-14289, 2022. 6, 7", + "[82] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 6, 7", + "[83] Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versatile adaptation approach for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28718-28728, 2024. 6, 7" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[84] Chao Yi, Lu Ren, De-Chuan Zhan, and Han-Jia Ye. Leveraging cross-modal neighbor representation for improved CLIP classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27402-27411, 2024. 7", + "[85] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. StableRep: Synthetic images from text-to-image models make strong visual representation learners. In Advances in Neural Information Processing Systems, pages 48382-48402, 2023. 7", + "[86] Peihua Li, Jiangtao Xie, Qilong Wang, and Wangmeng Zuo. Is second-order information helpful for large-scale visual recognition? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2070-2078, 2017. 7", + "[87] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382-7392, 2024. 8" + ], + "bbox": [ + 516, + 92, + 906, + 358 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "30031", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/b527a5f4-addd-424b-8f6f-3155526e27dd_model.json b/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/b527a5f4-addd-424b-8f6f-3155526e27dd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bb8782dd8e3f92bd39f053d8e04c08d5942439ed --- /dev/null +++ b/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/b527a5f4-addd-424b-8f6f-3155526e27dd_model.json @@ -0,0 +1,2655 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.13, + 0.882, + 0.177 + ], + "angle": 0, + "content": "ImagineFSL: Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.203, + 0.896, + 0.223 + ], + "angle": 0, + "content": "Haoyuan Yang \\(^{1, *}\\) Xiaou Li \\(^{2, *}\\) Jiaming Lv \\(^{1}\\) Xianjun Cheng \\(^{1}\\) Qilong Wang \\(^{3}\\) Peihua Li \\(^{1 \\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.161, + 0.224, + 0.837, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\) Dalian University of Technology \\(^{2}\\) Beijing University of Posts and Telecommunications \\(^{3}\\) Tianjin University" + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.242, + 0.635, + 0.256 + ], + "angle": 0, + "content": "https://peihuali.org/ImagineFSL" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.325, + 0.486, + 0.735 + ], + "angle": 0, + "content": "Adapting CLIP models for few-shot recognition has recently attracted significant attention. Despite considerable progress, these adaptations remain hindered by the pervasive challenge of data scarcity. Text-to-image models, capable of generating abundant photorealistic labeled images, offer a promising solution. However, existing approaches simply treat synthetic images as complements to real images, rather than as standalone knowledge repositories stemming from distinct foundation models. To overcome this limitation, we frame synthetic images as an imagined base set (iBase), i.e., an independent, large-scale synthetic dataset encompassing diverse concepts. Building on this perspective, we introduce ImagineFSL, a novel CLIP adaptation methodology that pretrains on iBase and then fine-tunes for downstream few-shot tasks. We find that, compared to no pretraining, both supervised and self-supervised pretraining are beneficial, with the latter providing better performance. Based on this finding, we propose an improved self-supervised method tailored for few-shot scenarios, enhancing the transferability of representations from synthetic to real image domains. Additionally, we present a systematic and scalable pipeline that employs chain-of-thought and in-context learning techniques, harnessing foundation models to automatically generate diverse, realistic images. Validated across eleven datasets, our methods consistently outperform state-of-the-art approaches by substantial margins." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.764, + 0.222, + 0.78 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.484, + 0.851 + ], + "angle": 0, + "content": "Few-shot learning (FSL), aiming to recognize novel concepts from limited labeled samples, has long been a subject of considerable research interest due to its potential for reducing labor-intensive human annotation and cura" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.906, + 0.414 + ], + "angle": 0, + "content": "tion [1]. FSL methods often rely on pretraining [2, 3] or meta-training [4, 5] on a base set of classes with abundant training examples, followed by meta-testing on a test set with limited labeled data. Recently, large vision-language models (VLMs) [6-8], particularly Contrastive Language-Image Pre-training (CLIP) [8], have exhibited remarkable few-shot and zero-shot recognition capabilities, drawing significant attention [9, 10] in the FSL regime." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.906, + 0.628 + ], + "angle": 0, + "content": "To adapt CLIP models for few-shot scenarios, various methods have been proposed, achieving substantial progress [11-14]. Nevertheless, these methods continue to struggle with the pervasive data-scarcity problem inherent in FSL. Although modern text-to-image (T2I) models [15, 16] can generate abundant photorealistic labeled images, the full potential of synthetic data in FSL remains underexplored. Existing approaches [17-19] simply treat synthetic images as complements, directly fine-tuning CLIP models using few-shot real images augmented with the synthetic ones during the meta-testing phase. At first glance, this strategy appears reasonable, since CLIP models are already extensively pretrained on vast datasets, suggesting that no further pretraining might be necessary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.902 + ], + "angle": 0, + "content": "However, we argue that such a strategy is sub-optimal because it treats synthetic images merely as complements to real images, instead of as standalone knowledge repositories derived from distinct foundation models. To overcome this limitation, we frame synthetic images as an imagined Base set (iBase), i.e., an independent, large-scale synthetic dataset encompassing diverse concepts, on which adaptation methods can learn rich knowledge transferable to few-shot tasks. By attaching learnable adapters to frozen CLIP encoders, we preserve the original representation capability while granting adaptability to the newly introduced parameters. We present a novel methodology called ImagineFSL (Figure 1), consisting of pretraining on iBase, followed by fine-tuning for downstream FSL tasks with real images and task-specific synthetic images. We find that, compared to no pretraining, either self-supervised learning (Self-SL) or supervised learning (SL) for pretraining is beneficial; crucially, Self-SL (e.g., DINO [20, 21]) outperforms SL." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.485, + 0.9 + ], + "angle": 0, + "content": "*These authors contributed equally to this work. †Corresponding Author. The work was supported by National Natural Science Foundation of China (62471083, 62276186, 61971086)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30020" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.093, + 0.493, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.293, + 0.478, + 0.306 + ], + "angle": 0, + "content": "(a) Self-supervised pretraining on PURELY synthetic dataset of iBase." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.096, + 0.898, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.293, + 0.868, + 0.306 + ], + "angle": 0, + "content": "(b) Fine-tuning with real and task-specific synthetic images." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.318, + 0.907, + 0.388 + ], + "angle": 0, + "content": "Figure 1. Overview of ImagineFSL. (a) We pretrain the adapter attached to the CLIP image encoder on iBase, an independent, large-scale synthetic dataset, using a Self-SL method called HoM-DINO. This method constructs two views via SyntAug with T2I models conditioned on the same caption and employs two specialized heads: one for higher-order moment (HoM) representations and another for masked image modeling (MIM). (b) For downstream few-shot tasks, we fine-tune both the adapter and vision classifier alongside the vision-language classifier. Additionally, we can fine-tune the image encoder with LoRA, referred to as ImagineFSLloRA, for further improvement." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.409, + 0.485, + 0.681 + ], + "angle": 0, + "content": "Based on this finding, we propose an enhanced Self-SL method that improves upon DINO and is specifically tailored to FSL. Recognizing that local image regions are crucial in few-shot tasks [22-24], we address the limitation of DINO, which relies solely on the [CLS] token for image representation while overlooking patch tokens. Concretely, we model distributions of patch tokens alongside the [CLS] token to capture richer feature statistics. Rather than assuming a fixed a priori distribution (e.g., Gaussian [25]), we employ from low- to High-order Moments (HoM) [26], thus calling our method HoM-DINO. We also incorporate masked image modeling (MIM) [27] to strengthen patch-level discrimination. Additionally, we develop Synthetic Augmentation (SyntAug) by constructing views from different synthetic images corresponding to the same caption. This exploits the natural variations inherent in T2I models, providing superior semantic consistency and diversity than hand-crafted, meticulously chosen augmentations [28]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To facilitate our approach, we develop a systematic and scalable pipeline for synthesizing diverse and realistic images, relying exclusively on Large Language Models (LLMs) and T2I models. Our method employs chain-of-thought (CoT) [29] and in-context learning (ICL) techniques [30] for caption generation. Specifically, we instruct GPT-4 [31] to analyze key factors that define image descriptions, including attribute, background, viewpoint, lighting condition, and cause of image degradation. Furthermore, we direct GPT-4 to generate a small set of exemplary captions following predefined patterns that combine different factors. These captions serve as in-context prompts for open-source Llama [32], enabling large-scale generation of captions. Finally, the synthesized captions are used to guide" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.41, + 0.829, + 0.424 + ], + "angle": 0, + "content": "T2I models in generating images for FSL tasks." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.425, + 0.878, + 0.438 + ], + "angle": 0, + "content": "Our major contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.44, + 0.906, + 0.529 + ], + "angle": 0, + "content": "- We frame synthetic images as standalone knowledge repositories and present a CLIP adaptation methodology that pretrains on purely synthetic images before finetuning for few-shot tasks. This marks a clear departure from existing one-stage fine-tuning methods that simply treat synthetic images as complements to real images." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.531, + 0.905, + 0.589 + ], + "angle": 0, + "content": "- We propose an improved Self-SL method based on DINO, specifically tailored for FSL. It introduces higher-order moments for image representation and employs synthetic augmentation for effective view construction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.591, + 0.905, + 0.665 + ], + "angle": 0, + "content": "- We develop a systematic and scalable pipeline for synthesizing both captions and images, enabling generation of large-scale base sets for pretraining and task-specific datasets. Distinct from existing arts, we leverage CoT and ICL techniques for diverse, realistic image generation." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.44, + 0.906, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.68, + 0.663, + 0.695 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Synthetic images for CLIP-based FSL. Several works have studied CLIP adaptation for FSL tasks by using synthetic images [13, 17-19]. DISEF [17] captions few-shot real images with LLaVA [33], which alongside real images serve as inputs for T2I models. DataDream [19] fine-tunes T2I models with real images, building customized models for each dataset or category. IsSynth [18] employs the word-to-sentence T5 [34] for generating diverse captions. Outside the FSL regime, Yu et al. [35] improve ImageNet recognition with synthetic images through a contextualized and stylized diversification technique, while SynCLR [36] applies in-context learning to synthesize captions across different modes for training foundation models only with syn" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30021" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "thetic data; differently, focusing on FSL, we employ CoT and ICL techniques to synthesize image captions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.125, + 0.484, + 0.337 + ], + "angle": 0, + "content": "Adapting CLIP for FSL. Existing CLIP adaptation methods can be roughly categorized as Prompt Tuning (PT), Adapter Tuning (AT), Encoder Tuning (ET) and Hybrid Tuning (HT). The PT methods [9, 11, 37, 38] learn textual and/or vision prompts, inserted at the input or middle layers of CLIP encoders, for adapting to few-shot tasks. ET [17-19, 39] injects few trainable parameters (e.g., LoRA [40]) into the encoders of CLIP models. AT [12, 13, 41] attaches trainable layers to the frozen backbones. Unlike PT and ET, AT is more efficient of training as no gradient backpropagation across the encoders are needed. Besides, some researchers [24, 42] study hybrid tuning that combines AT, ET or VT in varying ways. Different from them, we propose a novel two-stage methodology for CLIP adaptation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.484, + 0.55 + ], + "angle": 0, + "content": "Self-supervised learning (Self-SL). Contrastive Self-SL [28, 36, 43] uses both positive and negative samples for training with infoNCE loss, which often requires large minibatch size or memory bank. In contrast, self-distillation based Self-SL [20, 21, 44] depends only on positive samples with standard cross-entropy loss, which needs proper mechanisms to avoid model collapse. Inspired by Masked Language Modeling (MLM) [45] in NLP, researchers [27, 46, 47] introduce Masked Image Modeling (MIM) for Self-SL that reconstructs randomly masked image patches. Differently, we focus on Self-SL for CLIP adaptation; particularly, we model distributions of patch tokens alongside [CLS] token as image representation, unlike existing Self-SL that only uses single [CLS] token." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.78 + ], + "angle": 0, + "content": "Distribution based image representation. Bilinear or covariance pooling [48], particularly equipped with square root normalization [49, 50], has achieved impressive performance in supervised learning. Therein, images are represented by second-moments or covariance matrices, which can be seen as normal distributions with zero mean. Based on Gaussian assumption, \\(\\mathrm{G}^2\\mathrm{DeNet}\\) [25] and HGD [51] represent images or regions with normal distributions, which are embedded into the space of symmetric positive definite matrices for end-to-end learning. MP [52] proposes characteristic functions for representing distributions that are approximated by first-moment and cross-covariances. In contrast, our representation uses normalized second- and third-moments, and to the best of our knowledge, we are the first to introduce distribution modeling into Self-SL." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.799, + 0.265, + 0.815 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our adaptation method that comprises pretraining and finetuning is shown in Figure 1. The pretraining enables the adapter to learn extensive knowledge on the iBase, while preserving original representation of CLIP. During the finetuning, we adapt the models for few-shot tasks with real and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "synthetic images. Moreover, we develop a scalable method to synthesize both captions and images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.13, + 0.81, + 0.146 + ], + "angle": 0, + "content": "3.1. Pretraining on Imagined Base Set" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.907, + 0.351 + ], + "angle": 0, + "content": "The proposed HoM-DINO is an improved self-distillation method over DINO. As Figure 1a shows, it comprises a student branch and a teacher branch both connected to frozen image encoder \\(\\theta\\), sharing the same architecture but with different parameters. The student consists of an adapter \\(\\phi\\) and two parallel projection heads, i.e., HoM head \\(\\psi_{H}\\) for image representation by higher-order moments and MIM head \\(\\psi_{P}\\) for masked image modeling. The teacher's adapter \\(\\widetilde{\\phi}\\) and two heads \\(\\widetilde{\\psi}_{H}\\) and \\(\\widetilde{\\psi}_{P}\\) are exponential moving averages (EMA) of the students. Our adapter consists of 4 modified transformer blocks, and the projection heads comprise a 3-layer MLP followed by a weight-normalized linear layer; see Supplement A.1 for their detailed architectures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.909, + 0.502 + ], + "angle": 0, + "content": "We introduce SyntAug for building positive pairs, where the two views are synthetic images generated by T2I models conditioned on the same caption. They convey the same scenery but with natural variations thanks to the randomness of noise in the generative models. This is unlike traditional method that construct views by hand-crafted, meticulously chosen image augmentations (HandAug) [28]. We note that SynCLR [36] uses synthetic images to construct multi-positives in contrastive loss, whereas our approach uses Kullback-Leibler (KL) loss in self-distillation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.909, + 0.732 + ], + "angle": 0, + "content": "We represent images with patch token distributions\\(^1\\) alongside [CLS] token. Let \\(\\{\\mathbf{f}_i\\}_{i=1}^n, \\mathbf{f}_i \\in \\mathbb{R}^p\\) represent patch tokens from the adapter \\(\\phi\\) given a view \\(\\mathbf{x}\\). The first moment \\(\\mathbf{m}_1 = \\frac{1}{n} \\sum_i \\mathbf{f}_i\\) denotes the distribution center. Since the size of k-th moment \\((k \\geq 2)\\) scales exponentially with p, we reduce the complexity by computing moments for individual feature components. Motivated by square root normalization for covariance matrices [49], we perform square root and cubic root for the second and third moments, respectively. The second moment \\(\\mathbf{m}_2 = (\\frac{1}{n} \\sum_i (\\mathbf{f}_i - \\mathbf{m}_1)^2)^{-\\frac{1}{2}}\\) captures variations around the center, where \\(\\frac{1}{2}\\) denotes element-wise square root. The third moment \\(\\mathbf{m}_3 = (\\frac{1}{n} \\sum_i (\\mathbf{f}_i - \\mathbf{m}_1)^3)^{-\\frac{1}{3}}\\) indicates asymmetry of distribution around its center. Our final image representation \\(\\mathbf{r}\\) is a concatenated \\(4p\\)-dimensional vector:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.74, + 0.907, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} = \\phi \\circ \\theta (\\mathbf {x}) \\triangleq \\left[ [ \\mathrm {C L S} ]; \\mathbf {m} _ {1}; \\mathbf {m} _ {2}; \\mathbf {m} _ {3} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.767, + 0.908, + 0.798 + ], + "angle": 0, + "content": "Let \\(\\mathbf{x}_1\\) and \\(\\mathbf{x}_2\\) be input views for the teacher and student, respectively. Then, the contrast loss for the HoM heads is:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.808, + 0.907, + 0.826 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {H o M}} (\\phi , \\psi_ {H}) = \\mathrm {K L} (\\widetilde {\\mathbf {q}} _ {\\mathbf {x} _ {1}}, \\mathbf {q} _ {\\mathbf {x} _ {2}}) + \\mathrm {K L} (\\widetilde {\\mathbf {q}} _ {\\mathbf {x} _ {2}}, \\mathbf {q} _ {\\mathbf {x} _ {1}}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.835, + 0.907, + 0.868 + ], + "angle": 0, + "content": "where \\(\\widetilde{\\mathbf{q}}_{\\mathbf{x}_1} = \\widetilde{\\psi}_H\\circ \\widetilde{\\phi}\\circ \\theta (\\mathbf{x}_1)\\) is the teacher's prediction and \\(\\mathbf{q}_{\\mathbf{x}_2} = \\psi_H\\circ \\phi \\circ \\theta (\\mathbf{x}_2)\\) is the student's one." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.908, + 0.901 + ], + "angle": 0, + "content": "1 In Supplement D.5, we visualize the distributions, illustrating why we prefer moments over an a priori distribution such as Gaussian [25]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30022" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.213 + ], + "angle": 0, + "content": "Moreover, we introduce MIM to enhance capability of patch representation. Specifically, we randomly mask a portion of tokens for the view fed to the student. Suppose \\(\\mathbf{x}^{\\mathrm{MIM}}\\) is the masked image and \\(\\mathcal{I}_{\\mathbf{x}^{\\mathrm{MIM}}}\\) is the index set of its masked patches. Let \\(\\{\\widetilde{\\mathbf{p}}_{\\mathbf{x}}^{i}\\}_{i = 1}^{n} = \\widetilde{\\psi}_{P}\\circ \\widetilde{\\phi}\\circ \\theta (\\mathbf{x})\\) and \\(\\{\\mathbf{p}_{\\mathbf{x}^{\\mathrm{MIM}}}^{i}\\}_{i = 1}^{n} = \\psi_{P}\\circ \\phi \\circ \\theta (\\mathbf{x}^{\\mathrm{MIM}})\\) be predictions made for patch tokens by the teacher and student, respectively. Then, we define the MIM loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.224, + 0.483, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M I M}} (\\phi , \\psi_ {P}) = \\sum_ {\\mathbf {x} \\in \\left\\{\\mathbf {x} _ {1}, \\mathbf {x} _ {2} \\right\\}} \\sum_ {i \\in \\mathcal {I} _ {\\mathbf {x} ^ {\\mathrm {M I M}}}} \\operatorname {K L} \\left(\\widetilde {\\mathbf {p}} _ {\\mathbf {x}} ^ {i}, \\mathbf {p} _ {\\mathbf {x} ^ {\\mathrm {M I M}}} ^ {i}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.272, + 0.465, + 0.287 + ], + "angle": 0, + "content": "We simply sum \\(\\mathcal{L}_{\\mathrm{HoM}}\\) and \\(\\mathcal{L}_{\\mathrm{MIM}}\\) to obtain the final loss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.298, + 0.368, + 0.314 + ], + "angle": 0, + "content": "3.2. Fine-tuning for Low-shot Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.577 + ], + "angle": 0, + "content": "Figure 1b illustrates the fine-tuning stage. Once pretraining is complete, we preserve the teacher's vision adapter for downstream tasks, while discarding its heads and the whole student branch. To tailor the CLIP models to downstream FSL tasks, we attach to the adapter a vision (V) classifier with a cross-entropy (CE) loss \\(\\mathcal{L}_{\\mathrm{V}}\\) for vision-based recognition. Meanwhile, we tune a learnable vision-language (VL) classifier with a CE loss \\(\\mathcal{L}_{\\mathrm{VL}}\\) [18, 53] for integration with language knowledge. The VL-classifier is initialized with a prompt ensembling technique. Specifically, for each class, we construct a set of class-specific text prompts fed to the frozen text encoder, and the text embeddings are averaged as the embedding of this class. Then, all per-class text embeddings are used to initialize the VL-classifier. The total loss is the sum of \\(\\mathcal{L}_{\\mathrm{V}}\\) and \\(\\mathcal{L}_{\\mathrm{VL}}\\). For inference, we average the logits from two classifiers with a fusion weight, following Tip-Adapter (f) [12] and CaFo [13]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.484, + 0.699 + ], + "angle": 0, + "content": "ImagineFSLLoRA. Additionally, we can further tune the image encoder, e.g., by introducing LoRA [17, 19, 39]. The resulting method is dubbed ImagineFSLLoRA. We note that such encoder tuning (ET) methods generally lead to significant performance gains, compared to those with frozen encoders. As such methods can adjust features from low-layers up to high-layers across the encoders, so can better adapt to specific tasks and domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.484, + 0.821 + ], + "angle": 0, + "content": "Zero-shot Recognition. Our method is applicable to zero-shot recognition tasks. In this setting, we perform fine-tuning using solely the synthetic images, without touch of any real images. By tuning CLIP to diverse synthetic images of categories of interest, our method can significantly boost zero-shot recognition, thanks to its capability to learn additional knowledge from foundation models including LLMs and T2I models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.431, + 0.85 + ], + "angle": 0, + "content": "3.3. Synthesizing both Captions and Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The synthesizing pipeline is shown in Figure 2, where the CoT and ICL techniques are used to instruct LLMs and T2I models to generate captions and images." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.887, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.263, + 0.905, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Our pipeline synthesizes captions & images. We leverage GPT with the CoT to analyze key factors defining image descriptions and generate exemplar captions, which are then used by Llama for diverse caption generation via ICL. Finally, SD model generates images based on the synthesized captions. Note our pipeline is not limited to closed-source GPT (see Supplement D.4.)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.524 + ], + "angle": 0, + "content": "Synthesizing Captions. We identify five key factors that constitute image descriptions: attribute, viewpoint, background (BG), lighting condition (LC), and cause of degradation of photos (CD). We design four distinct patterns for systematic caption generation, incorporating various combinations of these factors. The first one, Base-pattern, comprises attribute and viewpoint, providing a fundamental description of the concept. The rest ones, i.e., BG-pattern, LC-pattern, and CD-Pattern, enhance the Base-pattern by integrating additional BG, LC, and CD factor, respectively, enriching the depiction comprehensively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.765 + ], + "angle": 0, + "content": "The pipeline begins by querying GPT-4 to analyze the key factors associated with a given concept, generating a detailed summary of visual properties and foundational elements essential for synthesizing diverse, realistic images that mimic real-world photography. We then design prompt template (PTe), guiding GPT-4 to generate a small number of high-quality exemplary captions for each pattern. Given the cost and time constraints of accessing GPT-4 API, we delegate extensive caption generation tasks to the locally deployed, lightweight Llama model. We develop PTe for Llama that utilizes exemplary captions generated by GPT-4 as in-context examples, randomly selecting one of the four patterns and the corresponding factors and examples to instantiate these templates. This strategy aims to elicit rich and nuanced responses from Llama, facilitating the creation of contextually enriched captions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Synthesizing Images. We use Stable Diffusion 3 (SD3) Medium [54], a multimodal diffusion transformer, for image generation. We construct the iBase, SyntIN1K, by adopting ImageNet-1K [55] class names as concepts. For each concept, we synthesize 300 captions, generating four images per caption. Additionally, we build SyntIN1.5K by selecting 0.5K additional concepts from ImageNet-21K class names. For each downstream dataset, we synthesize 300 captions per class and generate one image per caption," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30023" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.226 + ], + "angle": 0, + "content": "
MethodT2IENSyn DataImageNetCaltechAircraftCarsFoodPetsFlowersDTDEuroSATSUNAvg Acc
IsSynth [18]**[18]-/73.4-/96.0-/46.7-/82.6-/87.3-/92.8-/86.5-/73.6-/87.1-/76.2-/80.2
CaFo [13]**[13]70.2/73.193.9/96.930.1/46.271.5/85.485.6/87.691.5/94.974.9/97.659.4/72.066.5/86.371.2/76.971.5/81.7
\\( \\text{IsSynth}^{\\dagger} \\)**Ours70.7/73.994.4/96.831.1/46.571.7/85.485.7/87.491.6/93.885.4/95.760.1/73.975.9/87.270.9/76.573.8/81.7
\\( \\text{CaFo}^{\\dagger} \\)**Ours70.2/73.994.2/96.930.6/47.472.6/85.786.1/87.691.8/94.976.1/97.860.1/72.568.7/86.771.3/76.972.2/82.0
ImagineFSL**Ours71.6/74.795.3/97.131.5/54.882.6/90.586.7/87.892.3/94.387.0/98.762.7/75.876.2/89.672.0/78.175.8/84.1
DISEF [17]**[17]-/73.9-/96.9-/63.9-/88.6-/87.1-/94.3-/98.9-/75.4-/94.3-/77.4-/85.1
DataDream [19]**[19]-/74.1-/96.9-/72.3-/92.4-/87.6-/94.8-/99.4-/81.6-/93.4-/77.5-/87.0
\\( \\text{DISEF}^{\\dagger} \\)**Ours71.0/74.093.2/97.032.0/67.972.6/91.286.3/87.292.5/94.588.0/98.960.5/76.082.3/94.370.6/77.174.9/85.8
\\( \\text{ImagineFSL}_{\\text{LoRA}} \\)**Ours71.8/75.295.4/97.934.0/74.182.8/92.986.8/88.392.8/95.490.2/99.763.7/78.082.7/95.072.8/78.877.3/87.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.234, + 0.907, + 0.263 + ], + "angle": 0, + "content": "Table 1. Comparison of 1-shot/16-shot recognition to prior methods using synthetic images. Bold: best results; underlined: second-best results. \\( {}^{ \\dagger } \\) Reproduced by us." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.371, + 0.286 + ], + "angle": 0, + "content": "producing task-specific synthetic datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.287, + 0.483, + 0.363 + ], + "angle": 0, + "content": "In Supplements B.1-B.3, we give details on our design philosophy regarding the factors, the queries for GPT-4 to analyze these factors, the PTe for GPT-4 to generate examples, and the PTe for in-context learning of Llama. In Table S-5, we present a summary of all synthetic datasets." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.376, + 0.216, + 0.394 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.401, + 0.483, + 0.523 + ], + "angle": 0, + "content": "We evaluate extensively our methods on few-shot recognition, domain generalization, and zero-shot recognition. Besides, we conduct a thorough ablation study on our method. Across the paper, we use CLIP ViT-B/16 as the visual encoder and its aligned counterpart as the textual encoder, unless otherwise specified. All experiments are conducted using PyTorch package [56] with GeForce RTX 4090 GPUs. Additional experiments can be found in Supplement D." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.532, + 0.184, + 0.548 + ], + "angle": 0, + "content": "4.1. Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.705 + ], + "angle": 0, + "content": "Datasets and Metrics. Following prior arts [11, 12], we benchmark few-shot and zero-shot recognition tasks on 11 datasets, including ImageNet [55], Caltech [57], Aircraft [58], Cars [59], Food [60], Pets [61], Flowers [62], DTD [63], EuroSAT [64], SUN [65], and UCF101 [66]. We assess domain generalization task on ImageNet-V2 [67], ImageNet-S [68], ImageNet-A [69], and ImageNet-R [70]. For meta-testing, we follow the common practice, randomly sampling 3 All-way-K-shot tasks per dataset, reporting the average accuracy (Avg Acc) as a percentage (\\%)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Pretraining setup. We train the models using the AdamW optimizer with a cosine scheduler of learning rate. As in DINO [20], we use image augmentations including random resized crop, horizontal flip, color jittering, grayscale conversion, and Gaussian blur. For MIM, we utilize a blockwise masking strategy [27], where each image is randomly masked and the proportion of masked patches is randomly sampled. Hyperparameters are given in Supplement A.2." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Fine-tuning Setup. We follow the implementation of DataDream [19], adopting random resized crop, random horizontal flip, a set of color augmentations and RandAugment [71] as image augmentations for both ImagineFSL and ImagineFSLLORA. As in IsSynth [18], DISEF and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.272, + 0.905, + 0.408 + ], + "angle": 0, + "content": "DataDream, we use mix training strategy that simultaneously uses real images and synthetic images in a batch. For input of the text encoder, we adopt an ensemble of textual templates hand-engineered by CLIP and Customized Prompts via Language models (CuPL) [72] provided by CaFo [13], unless otherwise specified. Following TipAdapter(f) [12], CaFo, DISEF and DataDream, we tune hyperparameters for each task on the validation sets. See Supplement A.3 for detailed hyperparameters of fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.905, + 0.59 + ], + "angle": 0, + "content": "LLMs and T2I models. To analyze the factors and generate the exemplary captions, we call GPT-4 API with a temperature of 0.5. Llama 3 8B is deployed locally to generate extensive captions, in which we set the temperature to 0.5 and Nucleus Sampling parameter to 0.7. We utilize a TensorRT optimized SD3 Medium model and generate \\(512 \\times 512\\) images, with the scale of classifier-free guidance set to 5 and the steps of denoising diffusion implicit model to 28. Similar to [17-19], we adopt a semantic filtering method that discards synthetic images misclassified by zero-shot CLIP classifier. We provide implementation details on synthesizing texts and images in Supplement B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.598, + 0.717, + 0.614 + ], + "angle": 0, + "content": "4.2. Few-shot Recognition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.905, + 0.695 + ], + "angle": 0, + "content": "After pretraining on iBase, we perform meta-testing on downstream few-shot tasks. For each task, we fine-tune the adapter and V-classifier alongside VL-classifier, using K-shot real images and task-specific synthetic images; finally, we conduct inference on test sets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.701, + 0.888, + 0.716 + ], + "angle": 0, + "content": "4.2.1. Contrast with Synthetic Images based Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.781 + ], + "angle": 0, + "content": "Table 1 compares to prior methods based on synthetic images in 1-/16-shot settings, including IsSynth [18], CaFo [13], DISEF [17], and DataDream [19]. For a fair comparison, we reproduce them with our synthetic images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "First, we consider the methods that tune neither T2I models nor ENcoders (EN). Note that our reproduced results for IsSynth and CaFo in 16-shot setting are both higher than those reported in the original papers; we conjecture the reason is that our synthetic images are more suitable for few-shot tasks. Using the same synthetic data as ours, ImagineFSL performs better than them by a large margin of \\(2.0\\%\\) in average accuracy in both 1-shot and 16-shot settings." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30024" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.09, + 0.482, + 0.288 + ], + "angle": 0, + "content": "
Method1-shot2-shot4-shot8-shot16-shotAvg Acc
PTCoOp [11]67.670.874.076.880.073.8
CoCoOp [9]68.870.471.773.875.472.0
MaPLe [73]66.470.573.576.478.673.1
ProGrad [74]69.572.474.777.479.974.8
PLOT++ [75]70.774.076.979.682.176.7
PromptSRC [76]72.375.378.480.782.977.9
ATCLIP-Adapter [41]65.766.867.771.575.569.4
TIP-Adapter(f) [12]70.973.775.678.380.775.8
TaskRes [77]70.872.974.778.480.875.5
ImagineFSL76.177.579.482.284.479.9
ETCLIP-LoRA [39]72.575.577.480.383.077.7
ET+PTLLaMP [42]72.475.978.881.683.878.5
PT+ATGalLoP [24]72.876.479.182.284.579.0
ET+ATImagineFSLLoRA77.679.181.984.287.682.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.3, + 0.483, + 0.342 + ], + "angle": 0, + "content": "Table 2. Comparison of \\( K \\)-shot recognition to prior methods using only real images. **Bold:** best results; underlined: second-best results. **Supplement C** provides detailed results on 11 datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.355, + 0.483, + 0.535 + ], + "angle": 0, + "content": "Next, we contrast the methods that tune CLIP encoders. DISEF and DataDream use the same technique to tune the image and text encoders with LoRA, while DataDream additionally tunes T2I models. In the 16-shot setting, our ImagineFSL\\(_{\\text{LoRA}}\\) outperforms DataDream on 10 out of 11 datasets; DataDream excels only on DTD, where we conjecture that fine-tuning T2I models with real images enhances realistic texture generation. Furthermore, with our synthetic images, the reproduced results for DISEF in the 16-shot setting surpass those reported in the original paper, while ImagineFSL\\(_{\\text{LoRA}}\\) also exceeds DISEF by \\(2.4\\% / 1.7\\%\\) in 1-shot/16-shot tasks." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.545, + 0.471, + 0.56 + ], + "angle": 0, + "content": "4.2.2. Contrast with Methods Using Only Real Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.761 + ], + "angle": 0, + "content": "In Table 2, we compare our methods to prior techniques that do not employ synthetic images for K-shot tasks, including prompt tuning (PT), adapter tuning (AT), encoder tuning (ET), and hybrid tuning approaches. ImagineFSL significantly outperforms PT, AT, and ET across all shot scenarios, particularly in fewer-shot settings, e.g., it surpasses CLIP-LoRA (the runner-up) by \\(2.0\\%\\) in 2-shot and \\(3.6\\%\\) in 1-shot settings. Additionally, while GalLoP introduces a novel dropout technique for learning global and local prompts via a single linear layer adapter, and LLaMP merges PT and ET using LoRA, ImagineFSL-LoRA leads consistently and substantially in each K-shot setting with an average accuracy improvement of \\(3.1\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.307, + 0.787 + ], + "angle": 0, + "content": "4.3. Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Table 3 compares our methods with prior approaches on the domain generalization task. In this task, a model is trained on 16 shots per class from ImageNet (IN, source) and tested on four target datasets: IN-V2, IN-S, IN-A, and IN-R. For our methods, we specifically synthesize images for IN-S and IN-R datasets, while synthetic images for ImageNet are used for IN-V2 and IN-A. We fine-tune the models using" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.744, + 0.107 + ], + "angle": 0, + "content": "both the real and synthetic images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.907, + 0.213 + ], + "angle": 0, + "content": "First, we observe that our methods outperform the second-best method (i.e., CoCoLe) by over \\(1.3\\%\\) in average accuracy. Notably, our models achieve performance that matches the top performer, GalLoP, on IN-V2 and Co-CoLe on IN-A. Furthermore, we establish new state-of-the-art results on both IN-S and IN-R, achieving accuracies of \\(53.8\\%\\) and \\(80.0\\%\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.224, + 0.722, + 0.239 + ], + "angle": 0, + "content": "4.4. Zero-shot Recognition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.907, + 0.412 + ], + "angle": 0, + "content": "We compare to previous methods on zero-shot recognition in Table 4. We see that our methods are significantly better than most of the PT methods on each individual datasets, including VCD [79], Sus-X [80], TPT [81] and DiffTPT [82]. Compared to state-of-the-art DMN [83], ImagineFSL and ImagineFSLoRA outperform by \\(1.8\\%\\) and \\(2.5\\%\\) in average accuracy, respectively, with the latter ranking first in 7 out of 11 datasets. These results suggest that synthetic images hold promise for enhancing CLIP zero-shot recognition, and our methods are likely to improve further with advancements in generative T2I models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.422, + 0.722, + 0.438 + ], + "angle": 0, + "content": "4.5. Empirical Complexity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.906, + 0.504 + ], + "angle": 0, + "content": "Pretraining. On SyntIN1K, pretraining with ViT-B/16 takes \\(\\sim 9\\) hours on 4 RTX 4090 GPUs (\\(\\sim 40\\) hours for ViT-L/14). This task-agnostic stage is performed once and then widely fit for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.906, + 0.641 + ], + "angle": 0, + "content": "Fine-tuning. We compare methods on UCF101 in the 16-shot setting with a batch size of 128 to ensure a fair comparison. We apply LoRA with a rank of 16 across methods involving encoder tuning (ET). Specifically, for DISEF, DataDream, and ImagineFSLloRA, LoRA is applied to the image encoder. Additionally, for DISEF and DataDream, LoRA is also applied to the text encoder. We note that DISEF and DataDream share similar empirical complexity, as they use the same technique for fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.808 + ], + "angle": 0, + "content": "Table 5 presents the results evaluated on one RTX 4090 GPU, with train and test latency per iteration. Among methods that do not involve \\(ET\\), ImagineFSL has comparable latency and memory usage to TIP-Adapter(f) and IsSynth. While it has more parameters, they make up only \\(5.08\\%\\) of total encoder parameters. Moreover, it outperforms CoOp significantly in training speed. Compared to methods that involve \\(ET\\), i.e., DISEF and DataDream, ImagineFSL\\(_{\\text{LoRA}}\\) achieves a two-fold improvement in training speed and consumes only half the memory, despite a small increase in the percentage of parameters (from \\(1.07\\%\\) to \\(5.89\\%\\))." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.667, + 0.835 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We conduct ablation analysis in 1-shot/16-shot settings on ImageNet, Aircraft, Flowers and EuroSAT with CLIP ViTB/16 as visual encoder, unless otherwise specified. We provide more ablations in Supplement D.2." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30025" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.437, + 0.255 + ], + "angle": 0, + "content": "
MethodSourceTarget
ImageNet-V2-S-A-RAvg Acc
PTCoOp [11]71.564.248.049.775.259.3
CoCoOp [9]71.064.148.850.676.259.9
MaPLe [73]70.764.149.250.977.060.3
PromptSRC [76]71.364.449.650.977.860.7
CoCoLe [14]73.965.950.951.878.961.9
ATTaskRes [77]73.165.349.150.477.760.6
GraphAdapter [78]73.465.649.250.677.760.8
ImagineFSL74.767.052.751.579.762.7
ET+ATImagineFSLloRA75.267.553.851.580.063.2
PT+ATGalLoP [24]75.167.549.550.377.861.3
" + }, + { + "type": "table", + "bbox": [ + 0.455, + 0.09, + 0.913, + 0.256 + ], + "angle": 0, + "content": "
MethodImageNetCiteLectAircraftCarsFoodPetsFlowersETDEmoSATSYNUCF101AvgAcc
PTVCD [79]68.0-----88.586.9-45.648.8-----
Sus-X [80]69.994.028.766.186.190.673.854.657.567.766.667.766.668.768.768.7
TPT [81]69.094.224.866.984.787.869.047.842.465.568.065.568.065.565.565.5
DiffTPT [82]70.392.525.667.087.288.270.147.043.165.768.268.265.965.965.965.9
DMN [83]72.395.330.068.085.192.074.555.959.470.272.570.570.570.570.570.5
CODER [84]71.5----89.892.055.760.5-------
ATImagineFSL70.994.730.278.686.592.373.959.265.670.373.272.372.372.372.372.3
ET+ATImagineFSL-LoRA71.094.730.478.786.592.373.660.871.770.373.373.373.073.073.073.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.267, + 0.452, + 0.28 + ], + "angle": 0, + "content": "Table 3. Results of domain generalization on 16-shot setting." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.29, + 0.479, + 0.423 + ], + "angle": 0, + "content": "
MethodLatencyMemory (G)Params (M)(%)
Train (ms)Test (ms)
PTCoOp [11]298.360.43.050.010.01
-TIP-Adapter(f) [12]70.261.52.730.830.55
ATIsSynth [18]69.960.32.740.050.03
ImagineFSL93.165.43.007.615.08
ETDISEF [17]544.360.820.641.571.07
DataDream [19]
ET+ATImagineFSLLoRA265.465.511.828.825.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.463 + ], + "angle": 0, + "content": "Table 5. Comparison of empirical complexity of fine-tuning. We evaluate all methods on a single RTX 4090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.476, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Is pretraining necessary? Table 6a compares the Non-Pretraining (NoP) setting with the pretraining setting. In NoP, we directly train the vision adapter and vision (V) classifier using either real images alone (row 1) or a mix of synthetic and real images (row 2). For the pretraining setting, we first train the vision adapter via either SL (row 3) or Self-SL on SyntIN-1K (row 4), then fine-tune the vision adapter and V-classifier. During testing, we combine the V-classifier's predictions with CLIP zero-shot predictions (i.e., the frozen VL-classifier), using CLIP's handcrafted templates as input to the text encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.839 + ], + "angle": 0, + "content": "In NoP, mixing real and synthetic images substantially improves performance compared to using real images alone, demonstrating that our synthetic images effectively mitigate data scarcity. Furthermore, compared to NoP, pretraining with SL yields notable gains, particularly in 1-shot average accuracy \\((+0.9\\%)\\). Additionally, our Self-SL method (i.e., HoM-DINO) surpasses SL by \\(2.3\\% / 1.1\\%\\) in 1-shot/16-shot average accuracy, indicating that Self-SL better learns transferable features from synthetic to real image domains. These results highlight the importance of self-supervised pretraining on SyntIN1K for capturing rich knowledge from diverse synthetic images. In Figure S-3 of Supplement D.5, we visualize self-attention maps of different methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "HoM-DINO against other Self-SL methods. Table 6b compares to three typical Self-SL approaches including DINOv2, SynCLR and MAE. DINOv2, relying only on positive examples, is a self-distillation method that introduces" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.267, + 0.844, + 0.28 + ], + "angle": 0, + "content": "Table 4. Results of zero-shot recognition on 11 datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.905, + 0.353 + ], + "angle": 0, + "content": "MIM loss into DINO loss. SynCLR, depending on both positive and negative samples, is a contrastive method that combines multi-positive contrastive loss [85] and MIM loss. MAE is a representative masked image modeling method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.905, + 0.445 + ], + "angle": 0, + "content": "It can be seen that MAE clearly underperforms and we conjecture its reconstruction mechanism in the pixel space is not fit for the few-shot scenarios. SynCLR achieves very competitive performance but still nontrivially lags behind ours. Our HoM-DINO builds upon and outperforms DINOv2 by \\(2.1\\% / 0.7\\%\\) in 1-shot/16-shot accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.906, + 0.763 + ], + "angle": 0, + "content": "How image representation (image-repr) and view construction impact? First, from Table 6c we observe that, with [CLS] token as image-repr and HandAug, MIM significantly improves over without MIM (rows 1 and 2), indicating prediction of masked patches is helpful for FSL. Second, incorporating local patches via average pooling of all patches (AvgP) and attention pooling of all tokens (AttnP) slightly improved performance (rows 4 and 5). Using distribution-based image representations further improves performance significantly (rows 6 and 8); meanwhile, our HoM is slightly better than \\(\\mathrm{G}^2\\mathrm{DeNet}\\) while having one fourth less dimension (Dim). SyntAug further improves HoM (rows 8 and 9). However, MP underperforms despite its usage of moments, which may be due to its lack of structural normalization mechanism that has proven to be critical for high-order representations [86]. Finally, as dimension of HoM is higher than [CLS] token, for a fair comparison, we add a linear layer after the adapter that increases [CLS] token to 2048-Dim. However, this remains significantly inferior to HoM (rows 3 and 8), suggesting that advantage of HoM lies in exploiting patches via statistical moments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.905, + 0.87 + ], + "angle": 0, + "content": "Influence of textual prompts. As shown in Table 6d, when used individually, the CuPL [72] slightly outperforms the hand-engineered prompt templates of CLIP. However, combining both CuPL and CLIP prompts results in a further improvement over the individual methods. Therefore, unless otherwise specified, we use an ensemble of CLIP and CuPL prompts throughout the paper." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "VL-classifier (clf) tuning and encoder tuning. Rather than using a frozen VL-clf, we tune the clf's weights (i.e.," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30026" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.505, + 0.163 + ], + "angle": 0, + "content": "
PretrainSyn ImgImageNetAircraftFlowersEuroSATAvg Acc
1-✘68.8/72.824.1/48.578.9/96.560.4/84.458.1/75.6
2-✓70.9/73.927.5/49.384.1/96.566.2/86.262.2/76.5
3SL✓71.0/73.928.7/49.585.3/97.067.3/86.863.1/76.8
4HoM-DINO✓71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.166, + 0.368, + 0.178 + ], + "angle": 0, + "content": "(a) Necessity of pretraining." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.18, + 0.505, + 0.338 + ], + "angle": 0, + "content": "
Image \nReprDimViewMIMImageNetAircraftFlowersEuroSATAvg Acc
1[CLS]512HA×71.2/73.527.2/50.083.9/97.266.7/86.962.3/76.9
2[CLS]512HA✓71.1/73.928.9/50.685.2/97.567.8/86.863.3/77.2
3[CLS]2048HA✓71.4/73.629.0/49.984.3/97.669.8/86.863.6/77.0
4AvgP512HA✓71.4/73.729.2/50.384.2/97.469.6/87.163.6/77.1
5AttnP512HA✓71.3/73.228.8/56.985.7/97.669.4/87.463.8/77.2
6G²De- \nNet8385HA✓71.1/73.629.3/50.385.3/97.570.2/87.764.0/77.3
7MP2087HA✓71.3/73.629.3/48.982.7/97.468.6/86.263.0/76.5
8HoM2048HA✓71.4/74.029.1/50.985.6/98.172.0/87.364.5/77.6
9HoM2048SA✓71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.34, + 0.489, + 0.353 + ], + "angle": 0, + "content": "(c) Effect of image representation and view. HA: HandAug; SA: SyntAug." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.355, + 0.507, + 0.437 + ], + "angle": 0, + "content": "
ModelBase setSizeImageNetAircraftFlowersEuroSATAvg Acc
ViT-B/16SyntIN1K0.1M71.1/74.431.0/54.887.0/98.874.8/88.166.0/79.0
0.3M71.6/74.731.5/54.887.0/98.776.2/89.666.6/79.5
SyntIN1.5K0.45M71.7/74.732.9/54.787.7/98.675.2/89.766.9/79.4
ViT-L/14SyntIN1K0.3M78.4/80.742.1/68.190.0/99.580.6/92.772.8/85.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.123, + 0.439, + 0.461, + 0.452 + ], + "angle": 0, + "content": "(f) Scaling of ImagineFSL versus data size and model capacity." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.09, + 0.901, + 0.162 + ], + "angle": 0, + "content": "
Self-SLImageNetAircraftFlowersEuroSATAvg Acc
DINOv271.1/73.928.9/50.685.2/97.567.8/86.863.3/77.2
SynCLR71.5/74.027.4/51.084.4/97.774.4/86.964.4/77.4
MAE71.2/74.326.4/47.683.6/97.466.5/86.761.9/76.3
HoM-DINO71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.624, + 0.165, + 0.802, + 0.178 + ], + "angle": 0, + "content": "(b) Different Self-SL techniques." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.182, + 0.901, + 0.248 + ], + "angle": 0, + "content": "
CLIPCuPLImageNetAircraftFlowerEuroSATAvg Acc
✓✗71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
✗✓71.6/74.330.7/51.685.9/98.472.0/87.865.0/78.0
✓✓71.6/74.330.8/51.586.3/98.373.7/87.865.6/78.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.627, + 0.251, + 0.799, + 0.264 + ], + "angle": 0, + "content": "(d) Influence of textual prompts." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.268, + 0.901, + 0.335 + ], + "angle": 0, + "content": "
V-clfVL-clfI-ENImageNetAircraftFlowersEuroSATAvg Acc
71.6/74.330.8/51.586.3/98.373.7/87.865.6/78.0
71.6/74.731.5/54.887.0/98.776.2/89.666.6/79.5
71.8/75.234.0/74.190.2/99.782.7/95.069.7/86.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.337, + 0.89, + 0.349 + ], + "angle": 0, + "content": "(e) Effects of tuning VL-classifier (clf) and Image ENcoder (I-EN)." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.353, + 0.901, + 0.428 + ], + "angle": 0, + "content": "
MethodChain-of-ThoughtFidelity ↓Diversity ↑
Handcrafted prompts×45.920.95
Ours×40.071.06
Ours✓32.591.21
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.556, + 0.43, + 0.869, + 0.443 + ], + "angle": 0, + "content": "(g) Effect of Chain-of-Thought (CoT) on image generation." + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.462, + 0.888, + 0.477 + ], + "angle": 0, + "content": "Table 6. Ablation study on the proposed methods in the 1-shot/16-shot settings. Bold: best results; underlined: second-best results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.488, + 0.483, + 0.609 + ], + "angle": 0, + "content": "per-class text embeddings), implementing the ImagineFSL method. Table 6e shows this tuning significantly improves performance across most datasets. Furthermore, tuning the Image Encoder (I-EN) with LoRA (i.e., ImagineFSL\\(_{\\text{LoRA}}\\)) further enhances performance, particularly on Aircraft, Flowers, and EuroSAT. We hypothesize that I-EN tuning refines feature representations across the entire network, leading to better task and domain adaptation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.612, + 0.483, + 0.808 + ], + "angle": 0, + "content": "Scaling capability. As shown in Table 6f, when using ViT-B/16 as the image encoder on SyntIN1K, we observe slight improvements in both 1-shot and 16-shot accuracies as data volume increases from 0.1M to 0.3M. Additionally, expanding the dataset to SyntIN1.5K slightly enhances 1-shot accuracy while 16-shot accuracy remains stable, suggesting potential saturation in performance. This could be attributed to the limitations of the frozen encoder, which might not effectively utilize the expanded dataset or additional concepts. However, switching to ViT-L/14 results in substantial gains of \\(5.9\\%\\) and \\(5.8\\%\\) in 1-shot and 16-shot accuracies, respectively, underscoring the scalability of our approach with larger CLIP models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "How does CoT affect generated images? We ablate our generation method by comparing settings w/ and w/o CoT prompting. We also compare against a baseline using handcrafted prompts, i.e., \"a photo of a {class}\", for image generation. In the CoT (w/o) setting, we prompt Llama to generate captions using factor names for individual patterns. As" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.625 + ], + "angle": 0, + "content": "in [87], we compute the FID score to measure fidelity between synthetic and real images on the ImageNet validation set. For diversity, we use a pretrained DINOv2 ViT-B/16 to extract features and compute the standard deviation per class. Table 6g shows that our method w/o CoT outperforms handcrafted prompts. Using CoT further improves diversity and fidelity, underscoring its role in generating diverse, realistic images. Supplement D.5 qualitatively analyzes our generated texts and images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.658, + 0.771, + 0.673 + ], + "angle": 0, + "content": "5. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We introduce ImagineFSL, a novel methodology for CLIP adaptation. It involves self-supervised pretraining on large-scale purely synthetic images, followed by fine-tuning with real and task-specific synthetic images. Unlike previous methods, we treat synthetic images as standalone knowledge repositories of diverse concepts. To generate diverse and realistic synthetic images, we leverage CoT and in-context learning techniques, utilizing the capabilities of LLMs and T2I models. Despite strong performance across various recognition tasks, our approach has certain limitations, such as pretraining overhead and underutilization of synthetic captions, which are discussed in Supplement E. We hope our work inspires further exploration of synthetic images for data-scarce tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30027" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Yisheng Song, Ting Wang, Puyu Cai, Subrota K Mondal, and Jyoti Prakash Sahoo. A comprehensive survey of few-shot learning: Evolution, applications, challenges, and opportunities. ACM Computing Surveys, 55(13s):1-40, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. In International Conference on Learning Representations, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[3] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In European Conference on Computer Vision, pages 266-282, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.483, + 0.34 + ], + "angle": 0, + "content": "[4] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, pages 1126-1135, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[5] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in Neural Information Processing Systems, pages 3630-3638, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.548 + ], + "angle": 0, + "content": "[6] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millicah, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[7] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[8] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[9] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[10] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23681-23690, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[12] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. TipAdapter: Training-free adaptation of CLIP for few-shot classification. In European Conference on Computer Vision, pages 493-510, 2022. 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[13] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15211-15222, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[14] Yi Zhang, Ke Yu, Siqi Wu, and Zhihai He. Conceptual codebook learning for vision-language models. In European Conference on Computer Vision, pages 235-251, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[15] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.907, + 0.447 + ], + "angle": 0, + "content": "[16] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.505 + ], + "angle": 0, + "content": "[17] Victor G Turrisi da Costa, Nicola Dall'Asen, Yiming Wang, Nicu Sebe, and Elisa Ricci. Diversified in-domain synthesis with efficient fine-tuning for few-shot classification. arXiv preprint arXiv:2312.03046, 2023. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.506, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[18] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? In International Conference on Learning Representations, 2023. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.905, + 0.632 + ], + "angle": 0, + "content": "[19] Jae Myung Kim, Jessica Bader, Stephan Alaniz, Cordelia Schmid, and Zeynep Akata. DataDream: Few-shot guided dataset generation. In European Conference on Computer Vision, pages 252-268, 2024. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.634, + 0.905, + 0.703 + ], + "angle": 0, + "content": "[20] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.705, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[21] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOV2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024:1-31, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[22] Carl Doersch, Ankush Gupta, and Andrew Zisserman. CrossTransformers: spatially-aware few-shot transfer. In Advances in Neural Information Processing Systems, pages 21981-21993, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30028" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[23] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. DeepEMD: Differentiable earth mover's distance for few-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5632-5648, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[24] Marc Lafon, Elias Ramzi, Clément Rambour, Nicolas Audebert, and Nicolas Thome. GalLoP: Learning global and local prompts for vision-language models. In European Conference on Computer Vision, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.482, + 0.277 + ], + "angle": 0, + "content": "[25] Qilong Wang, Peihua Li, and Lei Zhang. \\(\\mathbf{G}^2\\mathrm{DeNet}\\): Global Gaussian distribution embedding network and its application to visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6507-6516, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.281, + 0.482, + 0.308 + ], + "angle": 0, + "content": "[26] Jun Shao. Mathematical Statistics. Springer, New York, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.311, + 0.482, + 0.367 + ], + "angle": 0, + "content": "[27] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. iBOT: Image BERT pretraining with online tokenizer. In International Conference on Learning Representations, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.369, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[28] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.497 + ], + "angle": 0, + "content": "[29] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, pages 24824-24837, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.5, + 0.482, + 0.569 + ], + "angle": 0, + "content": "[30] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.64 + ], + "angle": 0, + "content": "[31] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floreccia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.644, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[32] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 49250-49267, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[34] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[35] Zhuoran Yu, Chenchen Zhu, Sean Culatana, Raghuraman Krishnamoorthi, Fanyi Xiao, and Yong Jae Lee. Diversify, don't fine-tune: Scaling up visual recognition training with synthetic images. arXiv preprint arXiv: 2312.02253, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[36] Yonglong Tian, Lijie Fan, Kaifeng Chen, Dina Katabi, Dilip Krishnan, and Phillip Isola. Learning vision from models rivals learning vision from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15887-15898, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[37] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1401-1411, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.906, + 0.304 + ], + "angle": 0, + "content": "[38] Adrian Bulat and Georgios Tzimiropoulos. LASP: Text-to-text optimization for language-aware soft prompting of vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23232-23241, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.908, + 0.361 + ], + "angle": 0, + "content": "[39] Maxime Zanella and Ismail Ben Ayed. Low-rank few-shot adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1593-1603, 2024. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[40] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.906, + 0.502 + ], + "angle": 0, + "content": "[41] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. CLIP-Adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.908, + 0.574 + ], + "angle": 0, + "content": "[42] Zhaoheng Zheng, Jingmin Wei, Xuefeng Hu, Haidong Zhu, and Ram Nevatia. Large language models are good prompt learners for low-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28453-28462, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.906, + 0.644 + ], + "angle": 0, + "content": "[43] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.906, + 0.73 + ], + "angle": 0, + "content": "[44] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[45] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4171–4186, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[46] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pretraining of image transformers. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[47] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Con-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30029" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 159-168, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[48] Tsung-Yu Lin, Aruni RoyChowdhury, and Subhransu Maji. Bilinear cnn models for fine-grained visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1449-1457, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[49] Peihua Li, Jiangtao Xie, Qilong Wang, and Zilin Gao. Towards faster training of global covariance pooling networks by iterative matrix square root normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 947-955, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[50] Yue Song, Nicu Sebe, and Wei Wang. On the eigenvalues of global covariance pooling for fine-grained visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3554-3566, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.308, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[51] Tetsu Matsukawa, Takahiro Okabe, Einoshin Suzuki, and Yoichi Sato. Hierarchical Gaussian descriptors with application to person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(9):2179–2194, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.379, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[52] Mingze Gao, Qilong Wang, Zhenyi Lin, Pengfei Zhu, Qinghua Hu, and Jingbo Zhou. Tuning pre-trained model via moment probing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11769-11779, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.45, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[53] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[54] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[55] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[56] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.869 + ], + "angle": 0, + "content": "[57] Li Fei-Fei, R Fergus, and P Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Worskshops, pages 178–178, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[58] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classi" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "fication of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[59] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 554–561, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[60] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, pages 446–461, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[61] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3498-3505, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[62] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[63] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[64] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[65] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[66] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[67] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In International Conference on Machine Learning, pages 5389-5400, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.662, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[68] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[69] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[70] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[71] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. RandAugment: Practical automated data augmentation with" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30030" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "a reduced search space. In Advances in Neural Information Processing Systems, pages 18613-18624, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[72] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15645-15655, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[73] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. MaPLe: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[74] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[75] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. In International Conference on Learning Representations, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.373, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[76] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[77] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10899-10909, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[78] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. GraphAdapter: Tuning vision-language models with dual knowledge graph. In Advances in Neural Information Processing Systems, pages 13448-13466, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[79] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In International Conference on Learning Representations, pages 1-17, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[80] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. SuS-X: Training-free name-only transfer of vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2725-2736, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[81] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. In Advances in Neural Information Processing Systems, pages 14274-14289, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[82] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[83] Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versatile adaptation approach for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28718-28728, 2024. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[84] Chao Yi, Lu Ren, De-Chuan Zhan, and Han-Jia Ye. Leveraging cross-modal neighbor representation for improved CLIP classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27402-27411, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[85] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. StableRep: Synthetic images from text-to-image models make strong visual representation learners. In Advances in Neural Information Processing Systems, pages 48382-48402, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[86] Peihua Li, Jiangtao Xie, Qilong Wang, and Wangmeng Zuo. Is second-order information helpful for large-scale visual recognition? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2070-2078, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.907, + 0.359 + ], + "angle": 0, + "content": "[87] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382-7392, 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30031" + } + ] +] \ No newline at end of file diff --git a/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/b527a5f4-addd-424b-8f6f-3155526e27dd_origin.pdf b/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/b527a5f4-addd-424b-8f6f-3155526e27dd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..70a9b6634b0b04c7aebf1f317fdfc7cffc8c98a1 --- /dev/null +++ b/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/b527a5f4-addd-424b-8f6f-3155526e27dd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d42be129d40948e8e7743d271edfa6e2afd5fc3a3884c6155173ce0dc29fd78 +size 2167240 diff --git a/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/full.md b/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9a802d7a09edf1bf66e7ea7f73d8992727ed7224 --- /dev/null +++ b/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/full.md @@ -0,0 +1,347 @@ +# ImagineFSL: Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning + +Haoyuan Yang $^{1, *}$ Xiaou Li $^{2, *}$ Jiaming Lv $^{1}$ Xianjun Cheng $^{1}$ Qilong Wang $^{3}$ Peihua Li $^{1 \dagger}$ + +$^{1}$ Dalian University of Technology $^{2}$ Beijing University of Posts and Telecommunications $^{3}$ Tianjin University + +https://peihuali.org/ImagineFSL + +# Abstract + +Adapting CLIP models for few-shot recognition has recently attracted significant attention. Despite considerable progress, these adaptations remain hindered by the pervasive challenge of data scarcity. Text-to-image models, capable of generating abundant photorealistic labeled images, offer a promising solution. However, existing approaches simply treat synthetic images as complements to real images, rather than as standalone knowledge repositories stemming from distinct foundation models. To overcome this limitation, we frame synthetic images as an imagined base set (iBase), i.e., an independent, large-scale synthetic dataset encompassing diverse concepts. Building on this perspective, we introduce ImagineFSL, a novel CLIP adaptation methodology that pretrains on iBase and then fine-tunes for downstream few-shot tasks. We find that, compared to no pretraining, both supervised and self-supervised pretraining are beneficial, with the latter providing better performance. Based on this finding, we propose an improved self-supervised method tailored for few-shot scenarios, enhancing the transferability of representations from synthetic to real image domains. Additionally, we present a systematic and scalable pipeline that employs chain-of-thought and in-context learning techniques, harnessing foundation models to automatically generate diverse, realistic images. Validated across eleven datasets, our methods consistently outperform state-of-the-art approaches by substantial margins. + +# 1. Introduction + +Few-shot learning (FSL), aiming to recognize novel concepts from limited labeled samples, has long been a subject of considerable research interest due to its potential for reducing labor-intensive human annotation and cura + +tion [1]. FSL methods often rely on pretraining [2, 3] or meta-training [4, 5] on a base set of classes with abundant training examples, followed by meta-testing on a test set with limited labeled data. Recently, large vision-language models (VLMs) [6-8], particularly Contrastive Language-Image Pre-training (CLIP) [8], have exhibited remarkable few-shot and zero-shot recognition capabilities, drawing significant attention [9, 10] in the FSL regime. + +To adapt CLIP models for few-shot scenarios, various methods have been proposed, achieving substantial progress [11-14]. Nevertheless, these methods continue to struggle with the pervasive data-scarcity problem inherent in FSL. Although modern text-to-image (T2I) models [15, 16] can generate abundant photorealistic labeled images, the full potential of synthetic data in FSL remains underexplored. Existing approaches [17-19] simply treat synthetic images as complements, directly fine-tuning CLIP models using few-shot real images augmented with the synthetic ones during the meta-testing phase. At first glance, this strategy appears reasonable, since CLIP models are already extensively pretrained on vast datasets, suggesting that no further pretraining might be necessary. + +However, we argue that such a strategy is sub-optimal because it treats synthetic images merely as complements to real images, instead of as standalone knowledge repositories derived from distinct foundation models. To overcome this limitation, we frame synthetic images as an imagined Base set (iBase), i.e., an independent, large-scale synthetic dataset encompassing diverse concepts, on which adaptation methods can learn rich knowledge transferable to few-shot tasks. By attaching learnable adapters to frozen CLIP encoders, we preserve the original representation capability while granting adaptability to the newly introduced parameters. We present a novel methodology called ImagineFSL (Figure 1), consisting of pretraining on iBase, followed by fine-tuning for downstream FSL tasks with real images and task-specific synthetic images. We find that, compared to no pretraining, either self-supervised learning (Self-SL) or supervised learning (SL) for pretraining is beneficial; crucially, Self-SL (e.g., DINO [20, 21]) outperforms SL. + +![](images/10f37648a47f9e7de4632d92409c96a964cd218e804af38a7181ee539cb67b09.jpg) +(a) Self-supervised pretraining on PURELY synthetic dataset of iBase. + +![](images/e01a441963cd89597fe6bf607b704665f5395f20c3f16c17f5aed030756c4ede.jpg) +(b) Fine-tuning with real and task-specific synthetic images. +Figure 1. Overview of ImagineFSL. (a) We pretrain the adapter attached to the CLIP image encoder on iBase, an independent, large-scale synthetic dataset, using a Self-SL method called HoM-DINO. This method constructs two views via SyntAug with T2I models conditioned on the same caption and employs two specialized heads: one for higher-order moment (HoM) representations and another for masked image modeling (MIM). (b) For downstream few-shot tasks, we fine-tune both the adapter and vision classifier alongside the vision-language classifier. Additionally, we can fine-tune the image encoder with LoRA, referred to as ImagineFSLloRA, for further improvement. + +Based on this finding, we propose an enhanced Self-SL method that improves upon DINO and is specifically tailored to FSL. Recognizing that local image regions are crucial in few-shot tasks [22-24], we address the limitation of DINO, which relies solely on the [CLS] token for image representation while overlooking patch tokens. Concretely, we model distributions of patch tokens alongside the [CLS] token to capture richer feature statistics. Rather than assuming a fixed a priori distribution (e.g., Gaussian [25]), we employ from low- to High-order Moments (HoM) [26], thus calling our method HoM-DINO. We also incorporate masked image modeling (MIM) [27] to strengthen patch-level discrimination. Additionally, we develop Synthetic Augmentation (SyntAug) by constructing views from different synthetic images corresponding to the same caption. This exploits the natural variations inherent in T2I models, providing superior semantic consistency and diversity than hand-crafted, meticulously chosen augmentations [28]. + +To facilitate our approach, we develop a systematic and scalable pipeline for synthesizing diverse and realistic images, relying exclusively on Large Language Models (LLMs) and T2I models. Our method employs chain-of-thought (CoT) [29] and in-context learning (ICL) techniques [30] for caption generation. Specifically, we instruct GPT-4 [31] to analyze key factors that define image descriptions, including attribute, background, viewpoint, lighting condition, and cause of image degradation. Furthermore, we direct GPT-4 to generate a small set of exemplary captions following predefined patterns that combine different factors. These captions serve as in-context prompts for open-source Llama [32], enabling large-scale generation of captions. Finally, the synthesized captions are used to guide + +T2I models in generating images for FSL tasks. + +Our major contributions are summarized as follows: + +- We frame synthetic images as standalone knowledge repositories and present a CLIP adaptation methodology that pretrains on purely synthetic images before finetuning for few-shot tasks. This marks a clear departure from existing one-stage fine-tuning methods that simply treat synthetic images as complements to real images. +- We propose an improved Self-SL method based on DINO, specifically tailored for FSL. It introduces higher-order moments for image representation and employs synthetic augmentation for effective view construction. +- We develop a systematic and scalable pipeline for synthesizing both captions and images, enabling generation of large-scale base sets for pretraining and task-specific datasets. Distinct from existing arts, we leverage CoT and ICL techniques for diverse, realistic image generation. + +# 2. Related Works + +Synthetic images for CLIP-based FSL. Several works have studied CLIP adaptation for FSL tasks by using synthetic images [13, 17-19]. DISEF [17] captions few-shot real images with LLaVA [33], which alongside real images serve as inputs for T2I models. DataDream [19] fine-tunes T2I models with real images, building customized models for each dataset or category. IsSynth [18] employs the word-to-sentence T5 [34] for generating diverse captions. Outside the FSL regime, Yu et al. [35] improve ImageNet recognition with synthetic images through a contextualized and stylized diversification technique, while SynCLR [36] applies in-context learning to synthesize captions across different modes for training foundation models only with syn + +thetic data; differently, focusing on FSL, we employ CoT and ICL techniques to synthesize image captions. + +Adapting CLIP for FSL. Existing CLIP adaptation methods can be roughly categorized as Prompt Tuning (PT), Adapter Tuning (AT), Encoder Tuning (ET) and Hybrid Tuning (HT). The PT methods [9, 11, 37, 38] learn textual and/or vision prompts, inserted at the input or middle layers of CLIP encoders, for adapting to few-shot tasks. ET [17-19, 39] injects few trainable parameters (e.g., LoRA [40]) into the encoders of CLIP models. AT [12, 13, 41] attaches trainable layers to the frozen backbones. Unlike PT and ET, AT is more efficient of training as no gradient backpropagation across the encoders are needed. Besides, some researchers [24, 42] study hybrid tuning that combines AT, ET or VT in varying ways. Different from them, we propose a novel two-stage methodology for CLIP adaptation. + +Self-supervised learning (Self-SL). Contrastive Self-SL [28, 36, 43] uses both positive and negative samples for training with infoNCE loss, which often requires large minibatch size or memory bank. In contrast, self-distillation based Self-SL [20, 21, 44] depends only on positive samples with standard cross-entropy loss, which needs proper mechanisms to avoid model collapse. Inspired by Masked Language Modeling (MLM) [45] in NLP, researchers [27, 46, 47] introduce Masked Image Modeling (MIM) for Self-SL that reconstructs randomly masked image patches. Differently, we focus on Self-SL for CLIP adaptation; particularly, we model distributions of patch tokens alongside [CLS] token as image representation, unlike existing Self-SL that only uses single [CLS] token. + +Distribution based image representation. Bilinear or covariance pooling [48], particularly equipped with square root normalization [49, 50], has achieved impressive performance in supervised learning. Therein, images are represented by second-moments or covariance matrices, which can be seen as normal distributions with zero mean. Based on Gaussian assumption, $\mathrm{G}^2\mathrm{DeNet}$ [25] and HGD [51] represent images or regions with normal distributions, which are embedded into the space of symmetric positive definite matrices for end-to-end learning. MP [52] proposes characteristic functions for representing distributions that are approximated by first-moment and cross-covariances. In contrast, our representation uses normalized second- and third-moments, and to the best of our knowledge, we are the first to introduce distribution modeling into Self-SL. + +# 3. Proposed Method + +Our adaptation method that comprises pretraining and finetuning is shown in Figure 1. The pretraining enables the adapter to learn extensive knowledge on the iBase, while preserving original representation of CLIP. During the finetuning, we adapt the models for few-shot tasks with real and + +synthetic images. Moreover, we develop a scalable method to synthesize both captions and images. + +# 3.1. Pretraining on Imagined Base Set + +The proposed HoM-DINO is an improved self-distillation method over DINO. As Figure 1a shows, it comprises a student branch and a teacher branch both connected to frozen image encoder $\theta$ , sharing the same architecture but with different parameters. The student consists of an adapter $\phi$ and two parallel projection heads, i.e., HoM head $\psi_{H}$ for image representation by higher-order moments and MIM head $\psi_{P}$ for masked image modeling. The teacher's adapter $\widetilde{\phi}$ and two heads $\widetilde{\psi}_{H}$ and $\widetilde{\psi}_{P}$ are exponential moving averages (EMA) of the students. Our adapter consists of 4 modified transformer blocks, and the projection heads comprise a 3-layer MLP followed by a weight-normalized linear layer; see Supplement A.1 for their detailed architectures. + +We introduce SyntAug for building positive pairs, where the two views are synthetic images generated by T2I models conditioned on the same caption. They convey the same scenery but with natural variations thanks to the randomness of noise in the generative models. This is unlike traditional method that construct views by hand-crafted, meticulously chosen image augmentations (HandAug) [28]. We note that SynCLR [36] uses synthetic images to construct multi-positives in contrastive loss, whereas our approach uses Kullback-Leibler (KL) loss in self-distillation. + +We represent images with patch token distributions $^1$ alongside [CLS] token. Let $\{\mathbf{f}_i\}_{i=1}^n, \mathbf{f}_i \in \mathbb{R}^p$ represent patch tokens from the adapter $\phi$ given a view $\mathbf{x}$ . The first moment $\mathbf{m}_1 = \frac{1}{n} \sum_i \mathbf{f}_i$ denotes the distribution center. Since the size of k-th moment $(k \geq 2)$ scales exponentially with p, we reduce the complexity by computing moments for individual feature components. Motivated by square root normalization for covariance matrices [49], we perform square root and cubic root for the second and third moments, respectively. The second moment $\mathbf{m}_2 = (\frac{1}{n} \sum_i (\mathbf{f}_i - \mathbf{m}_1)^2)^{-\frac{1}{2}}$ captures variations around the center, where $\frac{1}{2}$ denotes element-wise square root. The third moment $\mathbf{m}_3 = (\frac{1}{n} \sum_i (\mathbf{f}_i - \mathbf{m}_1)^3)^{-\frac{1}{3}}$ indicates asymmetry of distribution around its center. Our final image representation $\mathbf{r}$ is a concatenated $4p$ -dimensional vector: + +$$ +\mathbf {r} = \phi \circ \theta (\mathbf {x}) \triangleq \left[ [ \mathrm {C L S} ]; \mathbf {m} _ {1}; \mathbf {m} _ {2}; \mathbf {m} _ {3} \right]. \tag {1} +$$ + +Let $\mathbf{x}_1$ and $\mathbf{x}_2$ be input views for the teacher and student, respectively. Then, the contrast loss for the HoM heads is: + +$$ +\mathcal {L} _ {\mathrm {H o M}} (\phi , \psi_ {H}) = \mathrm {K L} (\widetilde {\mathbf {q}} _ {\mathbf {x} _ {1}}, \mathbf {q} _ {\mathbf {x} _ {2}}) + \mathrm {K L} (\widetilde {\mathbf {q}} _ {\mathbf {x} _ {2}}, \mathbf {q} _ {\mathbf {x} _ {1}}), \tag {2} +$$ + +where $\widetilde{\mathbf{q}}_{\mathbf{x}_1} = \widetilde{\psi}_H\circ \widetilde{\phi}\circ \theta (\mathbf{x}_1)$ is the teacher's prediction and $\mathbf{q}_{\mathbf{x}_2} = \psi_H\circ \phi \circ \theta (\mathbf{x}_2)$ is the student's one. + +Moreover, we introduce MIM to enhance capability of patch representation. Specifically, we randomly mask a portion of tokens for the view fed to the student. Suppose $\mathbf{x}^{\mathrm{MIM}}$ is the masked image and $\mathcal{I}_{\mathbf{x}^{\mathrm{MIM}}}$ is the index set of its masked patches. Let $\{\widetilde{\mathbf{p}}_{\mathbf{x}}^{i}\}_{i = 1}^{n} = \widetilde{\psi}_{P}\circ \widetilde{\phi}\circ \theta (\mathbf{x})$ and $\{\mathbf{p}_{\mathbf{x}^{\mathrm{MIM}}}^{i}\}_{i = 1}^{n} = \psi_{P}\circ \phi \circ \theta (\mathbf{x}^{\mathrm{MIM}})$ be predictions made for patch tokens by the teacher and student, respectively. Then, we define the MIM loss as: + +$$ +\mathcal {L} _ {\mathrm {M I M}} (\phi , \psi_ {P}) = \sum_ {\mathbf {x} \in \left\{\mathbf {x} _ {1}, \mathbf {x} _ {2} \right\}} \sum_ {i \in \mathcal {I} _ {\mathbf {x} ^ {\mathrm {M I M}}}} \operatorname {K L} \left(\widetilde {\mathbf {p}} _ {\mathbf {x}} ^ {i}, \mathbf {p} _ {\mathbf {x} ^ {\mathrm {M I M}}} ^ {i}\right). \tag {3} +$$ + +We simply sum $\mathcal{L}_{\mathrm{HoM}}$ and $\mathcal{L}_{\mathrm{MIM}}$ to obtain the final loss. + +# 3.2. Fine-tuning for Low-shot Tasks + +Figure 1b illustrates the fine-tuning stage. Once pretraining is complete, we preserve the teacher's vision adapter for downstream tasks, while discarding its heads and the whole student branch. To tailor the CLIP models to downstream FSL tasks, we attach to the adapter a vision (V) classifier with a cross-entropy (CE) loss $\mathcal{L}_{\mathrm{V}}$ for vision-based recognition. Meanwhile, we tune a learnable vision-language (VL) classifier with a CE loss $\mathcal{L}_{\mathrm{VL}}$ [18, 53] for integration with language knowledge. The VL-classifier is initialized with a prompt ensembling technique. Specifically, for each class, we construct a set of class-specific text prompts fed to the frozen text encoder, and the text embeddings are averaged as the embedding of this class. Then, all per-class text embeddings are used to initialize the VL-classifier. The total loss is the sum of $\mathcal{L}_{\mathrm{V}}$ and $\mathcal{L}_{\mathrm{VL}}$ . For inference, we average the logits from two classifiers with a fusion weight, following Tip-Adapter (f) [12] and CaFo [13]. + +ImagineFSLLoRA. Additionally, we can further tune the image encoder, e.g., by introducing LoRA [17, 19, 39]. The resulting method is dubbed ImagineFSLLoRA. We note that such encoder tuning (ET) methods generally lead to significant performance gains, compared to those with frozen encoders. As such methods can adjust features from low-layers up to high-layers across the encoders, so can better adapt to specific tasks and domains. + +Zero-shot Recognition. Our method is applicable to zero-shot recognition tasks. In this setting, we perform fine-tuning using solely the synthetic images, without touch of any real images. By tuning CLIP to diverse synthetic images of categories of interest, our method can significantly boost zero-shot recognition, thanks to its capability to learn additional knowledge from foundation models including LLMs and T2I models. + +# 3.3. Synthesizing both Captions and Images + +The synthesizing pipeline is shown in Figure 2, where the CoT and ICL techniques are used to instruct LLMs and T2I models to generate captions and images. + +![](images/7b53f478f544ebcede3801748fb603a5e57cb08ba88db1934676770c36a08e9b.jpg) +Figure 2. Our pipeline synthesizes captions & images. We leverage GPT with the CoT to analyze key factors defining image descriptions and generate exemplar captions, which are then used by Llama for diverse caption generation via ICL. Finally, SD model generates images based on the synthesized captions. Note our pipeline is not limited to closed-source GPT (see Supplement D.4.) + +Synthesizing Captions. We identify five key factors that constitute image descriptions: attribute, viewpoint, background (BG), lighting condition (LC), and cause of degradation of photos (CD). We design four distinct patterns for systematic caption generation, incorporating various combinations of these factors. The first one, Base-pattern, comprises attribute and viewpoint, providing a fundamental description of the concept. The rest ones, i.e., BG-pattern, LC-pattern, and CD-Pattern, enhance the Base-pattern by integrating additional BG, LC, and CD factor, respectively, enriching the depiction comprehensively. + +The pipeline begins by querying GPT-4 to analyze the key factors associated with a given concept, generating a detailed summary of visual properties and foundational elements essential for synthesizing diverse, realistic images that mimic real-world photography. We then design prompt template (PTe), guiding GPT-4 to generate a small number of high-quality exemplary captions for each pattern. Given the cost and time constraints of accessing GPT-4 API, we delegate extensive caption generation tasks to the locally deployed, lightweight Llama model. We develop PTe for Llama that utilizes exemplary captions generated by GPT-4 as in-context examples, randomly selecting one of the four patterns and the corresponding factors and examples to instantiate these templates. This strategy aims to elicit rich and nuanced responses from Llama, facilitating the creation of contextually enriched captions. + +Synthesizing Images. We use Stable Diffusion 3 (SD3) Medium [54], a multimodal diffusion transformer, for image generation. We construct the iBase, SyntIN1K, by adopting ImageNet-1K [55] class names as concepts. For each concept, we synthesize 300 captions, generating four images per caption. Additionally, we build SyntIN1.5K by selecting 0.5K additional concepts from ImageNet-21K class names. For each downstream dataset, we synthesize 300 captions per class and generate one image per caption, + +
MethodT2IENSyn DataImageNetCaltechAircraftCarsFoodPetsFlowersDTDEuroSATSUNAvg Acc
IsSynth [18]**[18]-/73.4-/96.0-/46.7-/82.6-/87.3-/92.8-/86.5-/73.6-/87.1-/76.2-/80.2
CaFo [13]**[13]70.2/73.193.9/96.930.1/46.271.5/85.485.6/87.691.5/94.974.9/97.659.4/72.066.5/86.371.2/76.971.5/81.7
\( \text{IsSynth}^{\dagger} \)**Ours70.7/73.994.4/96.831.1/46.571.7/85.485.7/87.491.6/93.885.4/95.760.1/73.975.9/87.270.9/76.573.8/81.7
\( \text{CaFo}^{\dagger} \)**Ours70.2/73.994.2/96.930.6/47.472.6/85.786.1/87.691.8/94.976.1/97.860.1/72.568.7/86.771.3/76.972.2/82.0
ImagineFSL**Ours71.6/74.795.3/97.131.5/54.882.6/90.586.7/87.892.3/94.387.0/98.762.7/75.876.2/89.672.0/78.175.8/84.1
DISEF [17]**[17]-/73.9-/96.9-/63.9-/88.6-/87.1-/94.3-/98.9-/75.4-/94.3-/77.4-/85.1
DataDream [19]**[19]-/74.1-/96.9-/72.3-/92.4-/87.6-/94.8-/99.4-/81.6-/93.4-/77.5-/87.0
\( \text{DISEF}^{\dagger} \)**Ours71.0/74.093.2/97.032.0/67.972.6/91.286.3/87.292.5/94.588.0/98.960.5/76.082.3/94.370.6/77.174.9/85.8
\( \text{ImagineFSL}_{\text{LoRA}} \)**Ours71.8/75.295.4/97.934.0/74.182.8/92.986.8/88.392.8/95.490.2/99.763.7/78.082.7/95.072.8/78.877.3/87.5
+ +Table 1. Comparison of 1-shot/16-shot recognition to prior methods using synthetic images. Bold: best results; underlined: second-best results. ${}^{ \dagger }$ Reproduced by us. + +producing task-specific synthetic datasets. + +In Supplements B.1-B.3, we give details on our design philosophy regarding the factors, the queries for GPT-4 to analyze these factors, the PTe for GPT-4 to generate examples, and the PTe for in-context learning of Llama. In Table S-5, we present a summary of all synthetic datasets. + +# 4. Experiment + +We evaluate extensively our methods on few-shot recognition, domain generalization, and zero-shot recognition. Besides, we conduct a thorough ablation study on our method. Across the paper, we use CLIP ViT-B/16 as the visual encoder and its aligned counterpart as the textual encoder, unless otherwise specified. All experiments are conducted using PyTorch package [56] with GeForce RTX 4090 GPUs. Additional experiments can be found in Supplement D. + +# 4.1. Setting + +Datasets and Metrics. Following prior arts [11, 12], we benchmark few-shot and zero-shot recognition tasks on 11 datasets, including ImageNet [55], Caltech [57], Aircraft [58], Cars [59], Food [60], Pets [61], Flowers [62], DTD [63], EuroSAT [64], SUN [65], and UCF101 [66]. We assess domain generalization task on ImageNet-V2 [67], ImageNet-S [68], ImageNet-A [69], and ImageNet-R [70]. For meta-testing, we follow the common practice, randomly sampling 3 All-way-K-shot tasks per dataset, reporting the average accuracy (Avg Acc) as a percentage (\%). + +Pretraining setup. We train the models using the AdamW optimizer with a cosine scheduler of learning rate. As in DINO [20], we use image augmentations including random resized crop, horizontal flip, color jittering, grayscale conversion, and Gaussian blur. For MIM, we utilize a blockwise masking strategy [27], where each image is randomly masked and the proportion of masked patches is randomly sampled. Hyperparameters are given in Supplement A.2. + +Fine-tuning Setup. We follow the implementation of DataDream [19], adopting random resized crop, random horizontal flip, a set of color augmentations and RandAugment [71] as image augmentations for both ImagineFSL and ImagineFSLLORA. As in IsSynth [18], DISEF and + +DataDream, we use mix training strategy that simultaneously uses real images and synthetic images in a batch. For input of the text encoder, we adopt an ensemble of textual templates hand-engineered by CLIP and Customized Prompts via Language models (CuPL) [72] provided by CaFo [13], unless otherwise specified. Following TipAdapter(f) [12], CaFo, DISEF and DataDream, we tune hyperparameters for each task on the validation sets. See Supplement A.3 for detailed hyperparameters of fine-tuning. + +LLMs and T2I models. To analyze the factors and generate the exemplary captions, we call GPT-4 API with a temperature of 0.5. Llama 3 8B is deployed locally to generate extensive captions, in which we set the temperature to 0.5 and Nucleus Sampling parameter to 0.7. We utilize a TensorRT optimized SD3 Medium model and generate $512 \times 512$ images, with the scale of classifier-free guidance set to 5 and the steps of denoising diffusion implicit model to 28. Similar to [17-19], we adopt a semantic filtering method that discards synthetic images misclassified by zero-shot CLIP classifier. We provide implementation details on synthesizing texts and images in Supplement B. + +# 4.2. Few-shot Recognition + +After pretraining on iBase, we perform meta-testing on downstream few-shot tasks. For each task, we fine-tune the adapter and V-classifier alongside VL-classifier, using K-shot real images and task-specific synthetic images; finally, we conduct inference on test sets. + +# 4.2.1. Contrast with Synthetic Images based Methods + +Table 1 compares to prior methods based on synthetic images in 1-/16-shot settings, including IsSynth [18], CaFo [13], DISEF [17], and DataDream [19]. For a fair comparison, we reproduce them with our synthetic images. + +First, we consider the methods that tune neither T2I models nor ENcoders (EN). Note that our reproduced results for IsSynth and CaFo in 16-shot setting are both higher than those reported in the original papers; we conjecture the reason is that our synthetic images are more suitable for few-shot tasks. Using the same synthetic data as ours, ImagineFSL performs better than them by a large margin of $2.0\%$ in average accuracy in both 1-shot and 16-shot settings. + +
Method1-shot2-shot4-shot8-shot16-shotAvg Acc
PTCoOp [11]67.670.874.076.880.073.8
CoCoOp [9]68.870.471.773.875.472.0
MaPLe [73]66.470.573.576.478.673.1
ProGrad [74]69.572.474.777.479.974.8
PLOT++ [75]70.774.076.979.682.176.7
PromptSRC [76]72.375.378.480.782.977.9
ATCLIP-Adapter [41]65.766.867.771.575.569.4
TIP-Adapter(f) [12]70.973.775.678.380.775.8
TaskRes [77]70.872.974.778.480.875.5
ImagineFSL76.177.579.482.284.479.9
ETCLIP-LoRA [39]72.575.577.480.383.077.7
ET+PTLLaMP [42]72.475.978.881.683.878.5
PT+ATGalLoP [24]72.876.479.182.284.579.0
ET+ATImagineFSLLoRA77.679.181.984.287.682.1
+ +Table 2. Comparison of $K$ -shot recognition to prior methods using only real images. **Bold:** best results; underlined: second-best results. **Supplement C** provides detailed results on 11 datasets. + +Next, we contrast the methods that tune CLIP encoders. DISEF and DataDream use the same technique to tune the image and text encoders with LoRA, while DataDream additionally tunes T2I models. In the 16-shot setting, our ImagineFSL $_{\text{LoRA}}$ outperforms DataDream on 10 out of 11 datasets; DataDream excels only on DTD, where we conjecture that fine-tuning T2I models with real images enhances realistic texture generation. Furthermore, with our synthetic images, the reproduced results for DISEF in the 16-shot setting surpass those reported in the original paper, while ImagineFSL $_{\text{LoRA}}$ also exceeds DISEF by $2.4\% / 1.7\%$ in 1-shot/16-shot tasks. + +# 4.2.2. Contrast with Methods Using Only Real Images + +In Table 2, we compare our methods to prior techniques that do not employ synthetic images for K-shot tasks, including prompt tuning (PT), adapter tuning (AT), encoder tuning (ET), and hybrid tuning approaches. ImagineFSL significantly outperforms PT, AT, and ET across all shot scenarios, particularly in fewer-shot settings, e.g., it surpasses CLIP-LoRA (the runner-up) by $2.0\%$ in 2-shot and $3.6\%$ in 1-shot settings. Additionally, while GalLoP introduces a novel dropout technique for learning global and local prompts via a single linear layer adapter, and LLaMP merges PT and ET using LoRA, ImagineFSL-LoRA leads consistently and substantially in each K-shot setting with an average accuracy improvement of $3.1\%$ . + +# 4.3. Domain Generalization + +Table 3 compares our methods with prior approaches on the domain generalization task. In this task, a model is trained on 16 shots per class from ImageNet (IN, source) and tested on four target datasets: IN-V2, IN-S, IN-A, and IN-R. For our methods, we specifically synthesize images for IN-S and IN-R datasets, while synthetic images for ImageNet are used for IN-V2 and IN-A. We fine-tune the models using + +both the real and synthetic images. + +First, we observe that our methods outperform the second-best method (i.e., CoCoLe) by over $1.3\%$ in average accuracy. Notably, our models achieve performance that matches the top performer, GalLoP, on IN-V2 and Co-CoLe on IN-A. Furthermore, we establish new state-of-the-art results on both IN-S and IN-R, achieving accuracies of $53.8\%$ and $80.0\%$ , respectively. + +# 4.4. Zero-shot Recognition + +We compare to previous methods on zero-shot recognition in Table 4. We see that our methods are significantly better than most of the PT methods on each individual datasets, including VCD [79], Sus-X [80], TPT [81] and DiffTPT [82]. Compared to state-of-the-art DMN [83], ImagineFSL and ImagineFSLoRA outperform by $1.8\%$ and $2.5\%$ in average accuracy, respectively, with the latter ranking first in 7 out of 11 datasets. These results suggest that synthetic images hold promise for enhancing CLIP zero-shot recognition, and our methods are likely to improve further with advancements in generative T2I models. + +# 4.5. Empirical Complexity + +Pretraining. On SyntIN1K, pretraining with ViT-B/16 takes $\sim 9$ hours on 4 RTX 4090 GPUs ( $\sim 40$ hours for ViT-L/14). This task-agnostic stage is performed once and then widely fit for downstream tasks. + +Fine-tuning. We compare methods on UCF101 in the 16-shot setting with a batch size of 128 to ensure a fair comparison. We apply LoRA with a rank of 16 across methods involving encoder tuning (ET). Specifically, for DISEF, DataDream, and ImagineFSLloRA, LoRA is applied to the image encoder. Additionally, for DISEF and DataDream, LoRA is also applied to the text encoder. We note that DISEF and DataDream share similar empirical complexity, as they use the same technique for fine-tuning. + +Table 5 presents the results evaluated on one RTX 4090 GPU, with train and test latency per iteration. Among methods that do not involve $ET$ , ImagineFSL has comparable latency and memory usage to TIP-Adapter(f) and IsSynth. While it has more parameters, they make up only $5.08\%$ of total encoder parameters. Moreover, it outperforms CoOp significantly in training speed. Compared to methods that involve $ET$ , i.e., DISEF and DataDream, ImagineFSL $_{\text{LoRA}}$ achieves a two-fold improvement in training speed and consumes only half the memory, despite a small increase in the percentage of parameters (from $1.07\%$ to $5.89\%$ ). + +# 4.6. Ablation Study + +We conduct ablation analysis in 1-shot/16-shot settings on ImageNet, Aircraft, Flowers and EuroSAT with CLIP ViTB/16 as visual encoder, unless otherwise specified. We provide more ablations in Supplement D.2. + +
MethodSourceTarget
ImageNet-V2-S-A-RAvg Acc
PTCoOp [11]71.564.248.049.775.259.3
CoCoOp [9]71.064.148.850.676.259.9
MaPLe [73]70.764.149.250.977.060.3
PromptSRC [76]71.364.449.650.977.860.7
CoCoLe [14]73.965.950.951.878.961.9
ATTaskRes [77]73.165.349.150.477.760.6
GraphAdapter [78]73.465.649.250.677.760.8
ImagineFSL74.767.052.751.579.762.7
ET+ATImagineFSLloRA75.267.553.851.580.063.2
PT+ATGalLoP [24]75.167.549.550.377.861.3
+ +
MethodImageNetCiteLectAircraftCarsFoodPetsFlowersETDEmoSATSYNUCF101AvgAcc
PTVCD [79]68.0-----88.586.9-45.648.8-----
Sus-X [80]69.994.028.766.186.190.673.854.657.567.766.667.766.668.768.768.7
TPT [81]69.094.224.866.984.787.869.047.842.465.568.065.568.065.565.565.5
DiffTPT [82]70.392.525.667.087.288.270.147.043.165.768.268.265.965.965.965.9
DMN [83]72.395.330.068.085.192.074.555.959.470.272.570.570.570.570.570.5
CODER [84]71.5----89.892.055.760.5-------
ATImagineFSL70.994.730.278.686.592.373.959.265.670.373.272.372.372.372.372.3
ET+ATImagineFSL-LoRA71.094.730.478.786.592.373.660.871.770.373.373.373.073.073.073.0
+ +Table 3. Results of domain generalization on 16-shot setting. + +
MethodLatencyMemory (G)Params (M)(%)
Train (ms)Test (ms)
PTCoOp [11]298.360.43.050.010.01
-TIP-Adapter(f) [12]70.261.52.730.830.55
ATIsSynth [18]69.960.32.740.050.03
ImagineFSL93.165.43.007.615.08
ETDISEF [17]544.360.820.641.571.07
DataDream [19]
ET+ATImagineFSLLoRA265.465.511.828.825.89
+ +Table 5. Comparison of empirical complexity of fine-tuning. We evaluate all methods on a single RTX 4090 GPU. + +Is pretraining necessary? Table 6a compares the Non-Pretraining (NoP) setting with the pretraining setting. In NoP, we directly train the vision adapter and vision (V) classifier using either real images alone (row 1) or a mix of synthetic and real images (row 2). For the pretraining setting, we first train the vision adapter via either SL (row 3) or Self-SL on SyntIN-1K (row 4), then fine-tune the vision adapter and V-classifier. During testing, we combine the V-classifier's predictions with CLIP zero-shot predictions (i.e., the frozen VL-classifier), using CLIP's handcrafted templates as input to the text encoder. + +In NoP, mixing real and synthetic images substantially improves performance compared to using real images alone, demonstrating that our synthetic images effectively mitigate data scarcity. Furthermore, compared to NoP, pretraining with SL yields notable gains, particularly in 1-shot average accuracy $(+0.9\%)$ . Additionally, our Self-SL method (i.e., HoM-DINO) surpasses SL by $2.3\% / 1.1\%$ in 1-shot/16-shot average accuracy, indicating that Self-SL better learns transferable features from synthetic to real image domains. These results highlight the importance of self-supervised pretraining on SyntIN1K for capturing rich knowledge from diverse synthetic images. In Figure S-3 of Supplement D.5, we visualize self-attention maps of different methods. + +HoM-DINO against other Self-SL methods. Table 6b compares to three typical Self-SL approaches including DINOv2, SynCLR and MAE. DINOv2, relying only on positive examples, is a self-distillation method that introduces + +Table 4. Results of zero-shot recognition on 11 datasets. + +MIM loss into DINO loss. SynCLR, depending on both positive and negative samples, is a contrastive method that combines multi-positive contrastive loss [85] and MIM loss. MAE is a representative masked image modeling method. + +It can be seen that MAE clearly underperforms and we conjecture its reconstruction mechanism in the pixel space is not fit for the few-shot scenarios. SynCLR achieves very competitive performance but still nontrivially lags behind ours. Our HoM-DINO builds upon and outperforms DINOv2 by $2.1\% / 0.7\%$ in 1-shot/16-shot accuracy. + +How image representation (image-repr) and view construction impact? First, from Table 6c we observe that, with [CLS] token as image-repr and HandAug, MIM significantly improves over without MIM (rows 1 and 2), indicating prediction of masked patches is helpful for FSL. Second, incorporating local patches via average pooling of all patches (AvgP) and attention pooling of all tokens (AttnP) slightly improved performance (rows 4 and 5). Using distribution-based image representations further improves performance significantly (rows 6 and 8); meanwhile, our HoM is slightly better than $\mathrm{G}^2\mathrm{DeNet}$ while having one fourth less dimension (Dim). SyntAug further improves HoM (rows 8 and 9). However, MP underperforms despite its usage of moments, which may be due to its lack of structural normalization mechanism that has proven to be critical for high-order representations [86]. Finally, as dimension of HoM is higher than [CLS] token, for a fair comparison, we add a linear layer after the adapter that increases [CLS] token to 2048-Dim. However, this remains significantly inferior to HoM (rows 3 and 8), suggesting that advantage of HoM lies in exploiting patches via statistical moments. + +Influence of textual prompts. As shown in Table 6d, when used individually, the CuPL [72] slightly outperforms the hand-engineered prompt templates of CLIP. However, combining both CuPL and CLIP prompts results in a further improvement over the individual methods. Therefore, unless otherwise specified, we use an ensemble of CLIP and CuPL prompts throughout the paper. + +VL-classifier (clf) tuning and encoder tuning. Rather than using a frozen VL-clf, we tune the clf's weights (i.e., + +
PretrainSyn ImgImageNetAircraftFlowersEuroSATAvg Acc
1-✘68.8/72.824.1/48.578.9/96.560.4/84.458.1/75.6
2-✓70.9/73.927.5/49.384.1/96.566.2/86.262.2/76.5
3SL✓71.0/73.928.7/49.585.3/97.067.3/86.863.1/76.8
4HoM-DINO✓71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
+ +(a) Necessity of pretraining. + +
Image +ReprDimViewMIMImageNetAircraftFlowersEuroSATAvg Acc
1[CLS]512HA×71.2/73.527.2/50.083.9/97.266.7/86.962.3/76.9
2[CLS]512HA✓71.1/73.928.9/50.685.2/97.567.8/86.863.3/77.2
3[CLS]2048HA✓71.4/73.629.0/49.984.3/97.669.8/86.863.6/77.0
4AvgP512HA✓71.4/73.729.2/50.384.2/97.469.6/87.163.6/77.1
5AttnP512HA✓71.3/73.228.8/56.985.7/97.669.4/87.463.8/77.2
6G²De- +Net8385HA✓71.1/73.629.3/50.385.3/97.570.2/87.764.0/77.3
7MP2087HA✓71.3/73.629.3/48.982.7/97.468.6/86.263.0/76.5
8HoM2048HA✓71.4/74.029.1/50.985.6/98.172.0/87.364.5/77.6
9HoM2048SA✓71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
+ +(c) Effect of image representation and view. HA: HandAug; SA: SyntAug. + +
ModelBase setSizeImageNetAircraftFlowersEuroSATAvg Acc
ViT-B/16SyntIN1K0.1M71.1/74.431.0/54.887.0/98.874.8/88.166.0/79.0
0.3M71.6/74.731.5/54.887.0/98.776.2/89.666.6/79.5
SyntIN1.5K0.45M71.7/74.732.9/54.787.7/98.675.2/89.766.9/79.4
ViT-L/14SyntIN1K0.3M78.4/80.742.1/68.190.0/99.580.6/92.772.8/85.3
+ +(f) Scaling of ImagineFSL versus data size and model capacity. + +
Self-SLImageNetAircraftFlowersEuroSATAvg Acc
DINOv271.1/73.928.9/50.685.2/97.567.8/86.863.3/77.2
SynCLR71.5/74.027.4/51.084.4/97.774.4/86.964.4/77.4
MAE71.2/74.326.4/47.683.6/97.466.5/86.761.9/76.3
HoM-DINO71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
+ +(b) Different Self-SL techniques. + +
CLIPCuPLImageNetAircraftFlowerEuroSATAvg Acc
✓✗71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
✗✓71.6/74.330.7/51.685.9/98.472.0/87.865.0/78.0
✓✓71.6/74.330.8/51.586.3/98.373.7/87.865.6/78.0
+ +(d) Influence of textual prompts. + +
V-clfVL-clfI-ENImageNetAircraftFlowersEuroSATAvg Acc
71.6/74.330.8/51.586.3/98.373.7/87.865.6/78.0
71.6/74.731.5/54.887.0/98.776.2/89.666.6/79.5
71.8/75.234.0/74.190.2/99.782.7/95.069.7/86.0
+ +(e) Effects of tuning VL-classifier (clf) and Image ENcoder (I-EN). + +
MethodChain-of-ThoughtFidelity ↓Diversity ↑
Handcrafted prompts×45.920.95
Ours×40.071.06
Ours✓32.591.21
+ +(g) Effect of Chain-of-Thought (CoT) on image generation. + +Table 6. Ablation study on the proposed methods in the 1-shot/16-shot settings. Bold: best results; underlined: second-best results. + +per-class text embeddings), implementing the ImagineFSL method. Table 6e shows this tuning significantly improves performance across most datasets. Furthermore, tuning the Image Encoder (I-EN) with LoRA (i.e., ImagineFSL $_{\text{LoRA}}$ ) further enhances performance, particularly on Aircraft, Flowers, and EuroSAT. We hypothesize that I-EN tuning refines feature representations across the entire network, leading to better task and domain adaptation. + +Scaling capability. As shown in Table 6f, when using ViT-B/16 as the image encoder on SyntIN1K, we observe slight improvements in both 1-shot and 16-shot accuracies as data volume increases from 0.1M to 0.3M. Additionally, expanding the dataset to SyntIN1.5K slightly enhances 1-shot accuracy while 16-shot accuracy remains stable, suggesting potential saturation in performance. This could be attributed to the limitations of the frozen encoder, which might not effectively utilize the expanded dataset or additional concepts. However, switching to ViT-L/14 results in substantial gains of $5.9\%$ and $5.8\%$ in 1-shot and 16-shot accuracies, respectively, underscoring the scalability of our approach with larger CLIP models. + +How does CoT affect generated images? We ablate our generation method by comparing settings w/ and w/o CoT prompting. We also compare against a baseline using handcrafted prompts, i.e., "a photo of a {class}", for image generation. In the CoT (w/o) setting, we prompt Llama to generate captions using factor names for individual patterns. As + +in [87], we compute the FID score to measure fidelity between synthetic and real images on the ImageNet validation set. For diversity, we use a pretrained DINOv2 ViT-B/16 to extract features and compute the standard deviation per class. Table 6g shows that our method w/o CoT outperforms handcrafted prompts. Using CoT further improves diversity and fidelity, underscoring its role in generating diverse, realistic images. Supplement D.5 qualitatively analyzes our generated texts and images. + +# 5. Conclusion and Limitations + +We introduce ImagineFSL, a novel methodology for CLIP adaptation. It involves self-supervised pretraining on large-scale purely synthetic images, followed by fine-tuning with real and task-specific synthetic images. Unlike previous methods, we treat synthetic images as standalone knowledge repositories of diverse concepts. To generate diverse and realistic synthetic images, we leverage CoT and in-context learning techniques, utilizing the capabilities of LLMs and T2I models. Despite strong performance across various recognition tasks, our approach has certain limitations, such as pretraining overhead and underutilization of synthetic captions, which are discussed in Supplement E. We hope our work inspires further exploration of synthetic images for data-scarce tasks. + +# References + +[1] Yisheng Song, Ting Wang, Puyu Cai, Subrota K Mondal, and Jyoti Prakash Sahoo. A comprehensive survey of few-shot learning: Evolution, applications, challenges, and opportunities. ACM Computing Surveys, 55(13s):1-40, 2023. 1 +[2] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. In International Conference on Learning Representations, 2019. 1 +[3] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In European Conference on Computer Vision, pages 266-282, 2020. 1 +[4] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, pages 1126-1135, 2017. 1 +[5] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in Neural Information Processing Systems, pages 3630-3638, 2016. 1 +[6] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millicah, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 1 +[7] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. +[8] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 1 +[9] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 3, 6, 7 +[10] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23681-23690, 2024. 1 +[11] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 3, 5, 6, 7 + +[12] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. TipAdapter: Training-free adaptation of CLIP for few-shot classification. In European Conference on Computer Vision, pages 493-510, 2022. 3, 4, 5, 6, 7 +[13] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15211-15222, 2023. 2, 3, 4, 5 +[14] Yi Zhang, Ke Yu, Siqi Wu, and Zhihai He. Conceptual codebook learning for vision-language models. In European Conference on Computer Vision, pages 235-251, 2024. 1, 7 +[15] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804, 2022. 1 +[16] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1 +[17] Victor G Turrisi da Costa, Nicola Dall'Asen, Yiming Wang, Nicu Sebe, and Elisa Ricci. Diversified in-domain synthesis with efficient fine-tuning for few-shot classification. arXiv preprint arXiv:2312.03046, 2023. 1, 2, 3, 4, 5, 7 +[18] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? In International Conference on Learning Representations, 2023. 2, 4, 5, 7 +[19] Jae Myung Kim, Jessica Bader, Stephan Alaniz, Cordelia Schmid, and Zeynep Akata. DataDream: Few-shot guided dataset generation. In European Conference on Computer Vision, pages 252-268, 2024. 1, 2, 3, 4, 5, 7 +[20] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1, 3, 5 +[21] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOV2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024:1-31, 2024. 1, 3 +[22] Carl Doersch, Ankush Gupta, and Andrew Zisserman. CrossTransformers: spatially-aware few-shot transfer. In Advances in Neural Information Processing Systems, pages 21981-21993, 2020. 2 + +[23] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. DeepEMD: Differentiable earth mover's distance for few-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5632-5648, 2022. +[24] Marc Lafon, Elias Ramzi, Clément Rambour, Nicolas Audebert, and Nicolas Thome. GalLoP: Learning global and local prompts for vision-language models. In European Conference on Computer Vision, 2024. 2, 3, 6, 7 +[25] Qilong Wang, Peihua Li, and Lei Zhang. $\mathbf{G}^2\mathrm{DeNet}$ : Global Gaussian distribution embedding network and its application to visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6507-6516, 2017. 2, 3 +[26] Jun Shao. Mathematical Statistics. Springer, New York, 2003. 2 +[27] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. iBOT: Image BERT pretraining with online tokenizer. In International Conference on Learning Representations, 2022. 2, 3, 5 +[28] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 2, 3 +[29] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, pages 24824-24837, 2022. 2 +[30] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901, 2020. 2 +[31] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floreccia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2 +[32] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2 +[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 49250-49267, 2023. 2 +[34] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 2 +[35] Zhuoran Yu, Chenchen Zhu, Sean Culatana, Raghuraman Krishnamoorthi, Fanyi Xiao, and Yong Jae Lee. Diversify, don't fine-tune: Scaling up visual recognition training with synthetic images. arXiv preprint arXiv: 2312.02253, 2024. 2 + +[36] Yonglong Tian, Lijie Fan, Kaifeng Chen, Dina Katabi, Dilip Krishnan, and Phillip Isola. Learning vision from models rivals learning vision from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15887-15898, 2024. 2, 3 +[37] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1401-1411, 2023. 3 +[38] Adrian Bulat and Georgios Tzimiropoulos. LASP: Text-to-text optimization for language-aware soft prompting of vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23232-23241, 2023. 3 +[39] Maxime Zanella and Ismail Ben Ayed. Low-rank few-shot adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1593-1603, 2024. 3, 4, 6 +[40] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 3 +[41] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. CLIP-Adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2023. 3, 6 +[42] Zhaoheng Zheng, Jingmin Wei, Xuefeng Hu, Haidong Zhu, and Ram Nevatia. Large language models are good prompt learners for low-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28453-28462, 2024. 3, 6 +[43] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 3 +[44] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284, 2020. 3 +[45] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4171–4186, 2019. 3 +[46] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pretraining of image transformers. In International Conference on Learning Representations, 2022. 3 +[47] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Con- + +ference on Computer Vision and Pattern Recognition, pages 159-168, 2022. 3 +[48] Tsung-Yu Lin, Aruni RoyChowdhury, and Subhransu Maji. Bilinear cnn models for fine-grained visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1449-1457, 2015. 3 +[49] Peihua Li, Jiangtao Xie, Qilong Wang, and Zilin Gao. Towards faster training of global covariance pooling networks by iterative matrix square root normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 947-955, 2018. 3 +[50] Yue Song, Nicu Sebe, and Wei Wang. On the eigenvalues of global covariance pooling for fine-grained visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3554-3566, 2023. 3 +[51] Tetsu Matsukawa, Takahiro Okabe, Einoshin Suzuki, and Yoichi Sato. Hierarchical Gaussian descriptors with application to person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(9):2179–2194, 2020. 3 +[52] Mingze Gao, Qilong Wang, Zhenyi Lin, Pengfei Zhu, Qinghua Hu, and Jingbo Zhou. Tuning pre-trained model via moment probing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11769-11779, 2023. 3 +[53] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 4 +[54] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633, 2024. 4 +[55] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 4, 5 +[56] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035, 2019. 5 +[57] Li Fei-Fei, R Fergus, and P Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Worskshops, pages 178–178, 2004. 5 +[58] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classi + +fication of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5 +[59] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 554–561, 2013. 5 +[60] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, pages 446–461, 2014. 5 +[61] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3498-3505, 2012. 5 +[62] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5 +[63] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 5 +[64] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5 +[65] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 5 +[66] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5 +[67] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In International Conference on Machine Learning, pages 5389-5400, 2019. 5 +[68] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 5 +[69] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2019. 5 +[70] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 5 +[71] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. RandAugment: Practical automated data augmentation with + +a reduced search space. In Advances in Neural Information Processing Systems, pages 18613-18624, 2020. 5 +[72] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15645-15655, 2023. 5, 7 +[73] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. MaPLe: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 6, 7 +[74] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 6 +[75] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. In International Conference on Learning Representations, 2023. 6 +[76] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 6, 7 +[77] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10899-10909, 2023. 6, 7 +[78] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. GraphAdapter: Tuning vision-language models with dual knowledge graph. In Advances in Neural Information Processing Systems, pages 13448-13466, 2023. 7 +[79] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In International Conference on Learning Representations, pages 1-17, 2023. 6, 7 +[80] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. SuS-X: Training-free name-only transfer of vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2725-2736, 2023. 6, 7 +[81] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. In Advances in Neural Information Processing Systems, pages 14274-14289, 2022. 6, 7 +[82] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 6, 7 +[83] Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versatile adaptation approach for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28718-28728, 2024. 6, 7 + +[84] Chao Yi, Lu Ren, De-Chuan Zhan, and Han-Jia Ye. Leveraging cross-modal neighbor representation for improved CLIP classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27402-27411, 2024. 7 +[85] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. StableRep: Synthetic images from text-to-image models make strong visual representation learners. In Advances in Neural Information Processing Systems, pages 48382-48402, 2023. 7 +[86] Peihua Li, Jiangtao Xie, Qilong Wang, and Wangmeng Zuo. Is second-order information helpful for large-scale visual recognition? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2070-2078, 2017. 7 +[87] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382-7392, 2024. 8 \ No newline at end of file diff --git a/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/images.zip b/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c5959203449f07ae97e6cfaf3d1ab92eaf46641a --- /dev/null +++ b/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c3875f80502f6affc6c32ddef03f63c4bdcdb90882eb2bd6fc3037545a246c8 +size 602521 diff --git a/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/layout.json b/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0a2e772ab56c0bf390c0988edf72b05159be0e49 --- /dev/null +++ b/2025/ImagineFSL_ Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning/layout.json @@ -0,0 +1,9489 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "type": "text", + "content": "ImagineFSL: Self-Supervised Pretraining Matters on Imagined Base Set for VLM-based Few-shot Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "type": "text", + "content": "Haoyuan Yang " + }, + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "type": "inline_equation", + "content": "^{1, *}" + }, + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "type": "text", + "content": " Xiaou Li " + }, + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "type": "inline_equation", + "content": "^{2, *}" + }, + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "type": "text", + "content": " Jiaming Lv " + }, + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "type": "text", + "content": " Xianjun Cheng " + }, + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "type": "text", + "content": " Qilong Wang " + }, + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "type": "text", + "content": " Peihua Li " + }, + { + "bbox": [ + 61, + 160, + 548, + 176 + ], + "type": "inline_equation", + "content": "^{1 \\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 98, + 177, + 512, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 177, + 512, + 189 + ], + "spans": [ + { + "bbox": [ + 98, + 177, + 512, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 98, + 177, + 512, + 189 + ], + "type": "text", + "content": " Dalian University of Technology " + }, + { + "bbox": [ + 98, + 177, + 512, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 98, + 177, + 512, + 189 + ], + "type": "text", + "content": " Beijing University of Posts and Telecommunications " + }, + { + "bbox": [ + 98, + 177, + 512, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 98, + 177, + 512, + 189 + ], + "type": "text", + "content": " Tianjin University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 219, + 191, + 388, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 191, + 388, + 202 + ], + "spans": [ + { + "bbox": [ + 219, + 191, + 388, + 202 + ], + "type": "text", + "content": "https://peihuali.org/ImagineFSL" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 257, + 297, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 257, + 297, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 257, + 297, + 582 + ], + "type": "text", + "content": "Adapting CLIP models for few-shot recognition has recently attracted significant attention. Despite considerable progress, these adaptations remain hindered by the pervasive challenge of data scarcity. Text-to-image models, capable of generating abundant photorealistic labeled images, offer a promising solution. However, existing approaches simply treat synthetic images as complements to real images, rather than as standalone knowledge repositories stemming from distinct foundation models. To overcome this limitation, we frame synthetic images as an imagined base set (iBase), i.e., an independent, large-scale synthetic dataset encompassing diverse concepts. Building on this perspective, we introduce ImagineFSL, a novel CLIP adaptation methodology that pretrains on iBase and then fine-tunes for downstream few-shot tasks. We find that, compared to no pretraining, both supervised and self-supervised pretraining are beneficial, with the latter providing better performance. Based on this finding, we propose an improved self-supervised method tailored for few-shot scenarios, enhancing the transferability of representations from synthetic to real image domains. Additionally, we present a systematic and scalable pipeline that employs chain-of-thought and in-context learning techniques, harnessing foundation models to automatically generate diverse, realistic images. Validated across eleven datasets, our methods consistently outperform state-of-the-art approaches by substantial margins." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 135, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 135, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 135, + 617 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 625, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 296, + 673 + ], + "type": "text", + "content": "Few-shot learning (FSL), aiming to recognize novel concepts from limited labeled samples, has long been a subject of considerable research interest due to its potential for reducing labor-intensive human annotation and cura" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 232, + 554, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 554, + 327 + ], + "type": "text", + "content": "tion [1]. FSL methods often rely on pretraining [2, 3] or meta-training [4, 5] on a base set of classes with abundant training examples, followed by meta-testing on a test set with limited labeled data. Recently, large vision-language models (VLMs) [6-8], particularly Contrastive Language-Image Pre-training (CLIP) [8], have exhibited remarkable few-shot and zero-shot recognition capabilities, drawing significant attention [9, 10] in the FSL regime." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 329, + 554, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 554, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 554, + 497 + ], + "type": "text", + "content": "To adapt CLIP models for few-shot scenarios, various methods have been proposed, achieving substantial progress [11-14]. Nevertheless, these methods continue to struggle with the pervasive data-scarcity problem inherent in FSL. Although modern text-to-image (T2I) models [15, 16] can generate abundant photorealistic labeled images, the full potential of synthetic data in FSL remains underexplored. Existing approaches [17-19] simply treat synthetic images as complements, directly fine-tuning CLIP models using few-shot real images augmented with the synthetic ones during the meta-testing phase. At first glance, this strategy appears reasonable, since CLIP models are already extensively pretrained on vast datasets, suggesting that no further pretraining might be necessary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "However, we argue that such a strategy is sub-optimal because it treats synthetic images merely as complements to real images, instead of as standalone knowledge repositories derived from distinct foundation models. To overcome this limitation, we frame synthetic images as an imagined Base set (iBase), i.e., an independent, large-scale synthetic dataset encompassing diverse concepts, on which adaptation methods can learn rich knowledge transferable to few-shot tasks. By attaching learnable adapters to frozen CLIP encoders, we preserve the original representation capability while granting adaptability to the newly introduced parameters. We present a novel methodology called ImagineFSL (Figure 1), consisting of pretraining on iBase, followed by fine-tuning for downstream FSL tasks with real images and task-specific synthetic images. We find that, compared to no pretraining, either self-supervised learning (Self-SL) or supervised learning (SL) for pretraining is beneficial; crucially, Self-SL (e.g., DINO [20, 21]) outperforms SL." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 296, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 296, + 712 + ], + "type": "text", + "content": "*These authors contributed equally to this work. †Corresponding Author. The work was supported by National Natural Science Foundation of China (62471083, 62276186, 61971086)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30020" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 73, + 301, + 228 + ], + "blocks": [ + { + "bbox": [ + 65, + 73, + 301, + 228 + ], + "lines": [ + { + "bbox": [ + 65, + 73, + 301, + 228 + ], + "spans": [ + { + "bbox": [ + 65, + 73, + 301, + 228 + ], + "type": "image", + "image_path": "10f37648a47f9e7de4632d92409c96a964cd218e804af38a7181ee539cb67b09.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 232, + 292, + 242 + ], + "lines": [ + { + "bbox": [ + 67, + 232, + 292, + 242 + ], + "spans": [ + { + "bbox": [ + 67, + 232, + 292, + 242 + ], + "type": "text", + "content": "(a) Self-supervised pretraining on PURELY synthetic dataset of iBase." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 315, + 76, + 549, + 228 + ], + "blocks": [ + { + "bbox": [ + 315, + 76, + 549, + 228 + ], + "lines": [ + { + "bbox": [ + 315, + 76, + 549, + 228 + ], + "spans": [ + { + "bbox": [ + 315, + 76, + 549, + 228 + ], + "type": "image", + "image_path": "e01a441963cd89597fe6bf607b704665f5395f20c3f16c17f5aed030756c4ede.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 232, + 531, + 242 + ], + "lines": [ + { + "bbox": [ + 335, + 232, + 531, + 242 + ], + "spans": [ + { + "bbox": [ + 335, + 232, + 531, + 242 + ], + "type": "text", + "content": "(b) Fine-tuning with real and task-specific synthetic images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "text", + "content": "Figure 1. Overview of ImagineFSL. (a) We pretrain the adapter attached to the CLIP image encoder on iBase, an independent, large-scale synthetic dataset, using a Self-SL method called HoM-DINO. This method constructs two views via SyntAug with T2I models conditioned on the same caption and employs two specialized heads: one for higher-order moment (HoM) representations and another for masked image modeling (MIM). (b) For downstream few-shot tasks, we fine-tune both the adapter and vision classifier alongside the vision-language classifier. Additionally, we can fine-tune the image encoder with LoRA, referred to as ImagineFSLloRA, for further improvement." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 323, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 323, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 56, + 323, + 296, + 539 + ], + "type": "text", + "content": "Based on this finding, we propose an enhanced Self-SL method that improves upon DINO and is specifically tailored to FSL. Recognizing that local image regions are crucial in few-shot tasks [22-24], we address the limitation of DINO, which relies solely on the [CLS] token for image representation while overlooking patch tokens. Concretely, we model distributions of patch tokens alongside the [CLS] token to capture richer feature statistics. Rather than assuming a fixed a priori distribution (e.g., Gaussian [25]), we employ from low- to High-order Moments (HoM) [26], thus calling our method HoM-DINO. We also incorporate masked image modeling (MIM) [27] to strengthen patch-level discrimination. Additionally, we develop Synthetic Augmentation (SyntAug) by constructing views from different synthetic images corresponding to the same caption. This exploits the natural variations inherent in T2I models, providing superior semantic consistency and diversity than hand-crafted, meticulously chosen augmentations [28]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "To facilitate our approach, we develop a systematic and scalable pipeline for synthesizing diverse and realistic images, relying exclusively on Large Language Models (LLMs) and T2I models. Our method employs chain-of-thought (CoT) [29] and in-context learning (ICL) techniques [30] for caption generation. Specifically, we instruct GPT-4 [31] to analyze key factors that define image descriptions, including attribute, background, viewpoint, lighting condition, and cause of image degradation. Furthermore, we direct GPT-4 to generate a small set of exemplary captions following predefined patterns that combine different factors. These captions serve as in-context prompts for open-source Llama [32], enabling large-scale generation of captions. Finally, the synthesized captions are used to guide" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 324, + 507, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 324, + 507, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 324, + 507, + 335 + ], + "type": "text", + "content": "T2I models in generating images for FSL tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 336, + 537, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 336, + 537, + 346 + ], + "spans": [ + { + "bbox": [ + 326, + 336, + 537, + 346 + ], + "type": "text", + "content": "Our major contributions are summarized as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 348, + 554, + 526 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 348, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 348, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 348, + 554, + 418 + ], + "type": "text", + "content": "- We frame synthetic images as standalone knowledge repositories and present a CLIP adaptation methodology that pretrains on purely synthetic images before finetuning for few-shot tasks. This marks a clear departure from existing one-stage fine-tuning methods that simply treat synthetic images as complements to real images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 420, + 553, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 420, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 314, + 420, + 553, + 466 + ], + "type": "text", + "content": "- We propose an improved Self-SL method based on DINO, specifically tailored for FSL. It introduces higher-order moments for image representation and employs synthetic augmentation for effective view construction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 468, + 553, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 553, + 526 + ], + "type": "text", + "content": "- We develop a systematic and scalable pipeline for synthesizing both captions and images, enabling generation of large-scale base sets for pretraining and task-specific datasets. Distinct from existing arts, we leverage CoT and ICL techniques for diverse, realistic image generation." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 538, + 405, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 538, + 405, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 538, + 405, + 550 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "Synthetic images for CLIP-based FSL. Several works have studied CLIP adaptation for FSL tasks by using synthetic images [13, 17-19]. DISEF [17] captions few-shot real images with LLaVA [33], which alongside real images serve as inputs for T2I models. DataDream [19] fine-tunes T2I models with real images, building customized models for each dataset or category. IsSynth [18] employs the word-to-sentence T5 [34] for generating diverse captions. Outside the FSL regime, Yu et al. [35] improve ImageNet recognition with synthetic images through a contextualized and stylized diversification technique, while SynCLR [36] applies in-context learning to synthesize captions across different modes for training foundation models only with syn" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30021" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "thetic data; differently, focusing on FSL, we employ CoT and ICL techniques to synthesize image captions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 99, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 296, + 266 + ], + "type": "text", + "content": "Adapting CLIP for FSL. Existing CLIP adaptation methods can be roughly categorized as Prompt Tuning (PT), Adapter Tuning (AT), Encoder Tuning (ET) and Hybrid Tuning (HT). The PT methods [9, 11, 37, 38] learn textual and/or vision prompts, inserted at the input or middle layers of CLIP encoders, for adapting to few-shot tasks. ET [17-19, 39] injects few trainable parameters (e.g., LoRA [40]) into the encoders of CLIP models. AT [12, 13, 41] attaches trainable layers to the frozen backbones. Unlike PT and ET, AT is more efficient of training as no gradient backpropagation across the encoders are needed. Besides, some researchers [24, 42] study hybrid tuning that combines AT, ET or VT in varying ways. Different from them, we propose a novel two-stage methodology for CLIP adaptation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 268, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 296, + 435 + ], + "type": "text", + "content": "Self-supervised learning (Self-SL). Contrastive Self-SL [28, 36, 43] uses both positive and negative samples for training with infoNCE loss, which often requires large minibatch size or memory bank. In contrast, self-distillation based Self-SL [20, 21, 44] depends only on positive samples with standard cross-entropy loss, which needs proper mechanisms to avoid model collapse. Inspired by Masked Language Modeling (MLM) [45] in NLP, researchers [27, 46, 47] introduce Masked Image Modeling (MIM) for Self-SL that reconstructs randomly masked image patches. Differently, we focus on Self-SL for CLIP adaptation; particularly, we model distributions of patch tokens alongside [CLS] token as image representation, unlike existing Self-SL that only uses single [CLS] token." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 617 + ], + "type": "text", + "content": "Distribution based image representation. Bilinear or covariance pooling [48], particularly equipped with square root normalization [49, 50], has achieved impressive performance in supervised learning. Therein, images are represented by second-moments or covariance matrices, which can be seen as normal distributions with zero mean. Based on Gaussian assumption, " + }, + { + "bbox": [ + 55, + 437, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^2\\mathrm{DeNet}" + }, + { + "bbox": [ + 55, + 437, + 296, + 617 + ], + "type": "text", + "content": " [25] and HGD [51] represent images or regions with normal distributions, which are embedded into the space of symmetric positive definite matrices for end-to-end learning. MP [52] proposes characteristic functions for representing distributions that are approximated by first-moment and cross-covariances. In contrast, our representation uses normalized second- and third-moments, and to the best of our knowledge, we are the first to introduce distribution modeling into Self-SL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 632, + 162, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 162, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 162, + 645 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Our adaptation method that comprises pretraining and finetuning is shown in Figure 1. The pretraining enables the adapter to learn extensive knowledge on the iBase, while preserving original representation of CLIP. During the finetuning, we adapt the models for few-shot tasks with real and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "synthetic images. Moreover, we develop a scalable method to synthesize both captions and images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 102, + 495, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 102, + 495, + 115 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 495, + 115 + ], + "type": "text", + "content": "3.1. Pretraining on Imagined Base Set" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "text", + "content": "The proposed HoM-DINO is an improved self-distillation method over DINO. As Figure 1a shows, it comprises a student branch and a teacher branch both connected to frozen image encoder " + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "text", + "content": ", sharing the same architecture but with different parameters. The student consists of an adapter " + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "text", + "content": " and two parallel projection heads, i.e., HoM head " + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\psi_{H}" + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "text", + "content": " for image representation by higher-order moments and MIM head " + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\psi_{P}" + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "text", + "content": " for masked image modeling. The teacher's adapter " + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\widetilde{\\phi}" + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "text", + "content": " and two heads " + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\widetilde{\\psi}_{H}" + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\widetilde{\\psi}_{P}" + }, + { + "bbox": [ + 313, + 120, + 555, + 277 + ], + "type": "text", + "content": " are exponential moving averages (EMA) of the students. Our adapter consists of 4 modified transformer blocks, and the projection heads comprise a 3-layer MLP followed by a weight-normalized linear layer; see Supplement A.1 for their detailed architectures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 278, + 556, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 556, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 556, + 397 + ], + "type": "text", + "content": "We introduce SyntAug for building positive pairs, where the two views are synthetic images generated by T2I models conditioned on the same caption. They convey the same scenery but with natural variations thanks to the randomness of noise in the generative models. This is unlike traditional method that construct views by hand-crafted, meticulously chosen image augmentations (HandAug) [28]. We note that SynCLR [36] uses synthetic images to construct multi-positives in contrastive loss, whereas our approach uses Kullback-Leibler (KL) loss in self-distillation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "text", + "content": "We represent images with patch token distributions" + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "text", + "content": " alongside [CLS] token. Let " + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_i\\}_{i=1}^n, \\mathbf{f}_i \\in \\mathbb{R}^p" + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "text", + "content": " represent patch tokens from the adapter " + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "text", + "content": " given a view " + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "text", + "content": ". The first moment " + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_1 = \\frac{1}{n} \\sum_i \\mathbf{f}_i" + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "text", + "content": " denotes the distribution center. Since the size of k-th moment " + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "inline_equation", + "content": "(k \\geq 2)" + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "text", + "content": " scales exponentially with p, we reduce the complexity by computing moments for individual feature components. Motivated by square root normalization for covariance matrices [49], we perform square root and cubic root for the second and third moments, respectively. The second moment " + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_2 = (\\frac{1}{n} \\sum_i (\\mathbf{f}_i - \\mathbf{m}_1)^2)^{-\\frac{1}{2}}" + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "text", + "content": " captures variations around the center, where " + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}" + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "text", + "content": " denotes element-wise square root. The third moment " + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_3 = (\\frac{1}{n} \\sum_i (\\mathbf{f}_i - \\mathbf{m}_1)^3)^{-\\frac{1}{3}}" + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "text", + "content": " indicates asymmetry of distribution around its center. Our final image representation " + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "text", + "content": " is a concatenated " + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "inline_equation", + "content": "4p" + }, + { + "bbox": [ + 313, + 397, + 556, + 579 + ], + "type": "text", + "content": "-dimensional vector:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 586, + 555, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 586, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 351, + 586, + 555, + 601 + ], + "type": "interline_equation", + "content": "\\mathbf {r} = \\phi \\circ \\theta (\\mathbf {x}) \\triangleq \\left[ [ \\mathrm {C L S} ]; \\mathbf {m} _ {1}; \\mathbf {m} _ {2}; \\mathbf {m} _ {3} \\right]. \\tag {1}", + "image_path": "8a4db12a69034ed12e3e8ee1d7bad3be3b7e1722605397f4d7248bae2a6dea23.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 607, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 555, + 632 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 607, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 313, + 607, + 555, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 607, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 313, + 607, + 555, + 632 + ], + "type": "text", + "content": " be input views for the teacher and student, respectively. Then, the contrast loss for the HoM heads is:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 639, + 555, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 639, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 328, + 639, + 555, + 654 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {H o M}} (\\phi , \\psi_ {H}) = \\mathrm {K L} (\\widetilde {\\mathbf {q}} _ {\\mathbf {x} _ {1}}, \\mathbf {q} _ {\\mathbf {x} _ {2}}) + \\mathrm {K L} (\\widetilde {\\mathbf {q}} _ {\\mathbf {x} _ {2}}, \\mathbf {q} _ {\\mathbf {x} _ {1}}), \\tag {2}", + "image_path": "201d995b43c1a520fbcba0b9c4a344dec67c40bc88564555c20cf0e822766141.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 661, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 661, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 661, + 555, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 661, + 555, + 687 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{q}}_{\\mathbf{x}_1} = \\widetilde{\\psi}_H\\circ \\widetilde{\\phi}\\circ \\theta (\\mathbf{x}_1)" + }, + { + "bbox": [ + 313, + 661, + 555, + 687 + ], + "type": "text", + "content": " is the teacher's prediction and " + }, + { + "bbox": [ + 313, + 661, + 555, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{\\mathbf{x}_2} = \\psi_H\\circ \\phi \\circ \\theta (\\mathbf{x}_2)" + }, + { + "bbox": [ + 313, + 661, + 555, + 687 + ], + "type": "text", + "content": " is the student's one." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": "1 In Supplement D.5, we visualize the distributions, illustrating why we prefer moments over an a priori distribution such as Gaussian [25]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30022" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 168 + ], + "type": "text", + "content": "Moreover, we introduce MIM to enhance capability of patch representation. Specifically, we randomly mask a portion of tokens for the view fed to the student. Suppose " + }, + { + "bbox": [ + 55, + 72, + 295, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{MIM}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 168 + ], + "type": "text", + "content": " is the masked image and " + }, + { + "bbox": [ + 55, + 72, + 295, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathbf{x}^{\\mathrm{MIM}}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 168 + ], + "type": "text", + "content": " is the index set of its masked patches. Let " + }, + { + "bbox": [ + 55, + 72, + 295, + 168 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{\\mathbf{p}}_{\\mathbf{x}}^{i}\\}_{i = 1}^{n} = \\widetilde{\\psi}_{P}\\circ \\widetilde{\\phi}\\circ \\theta (\\mathbf{x})" + }, + { + "bbox": [ + 55, + 72, + 295, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 168 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_{\\mathbf{x}^{\\mathrm{MIM}}}^{i}\\}_{i = 1}^{n} = \\psi_{P}\\circ \\phi \\circ \\theta (\\mathbf{x}^{\\mathrm{MIM}})" + }, + { + "bbox": [ + 55, + 72, + 295, + 168 + ], + "type": "text", + "content": " be predictions made for patch tokens by the teacher and student, respectively. Then, we define the MIM loss as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 177, + 295, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 177, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 67, + 177, + 295, + 205 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M I M}} (\\phi , \\psi_ {P}) = \\sum_ {\\mathbf {x} \\in \\left\\{\\mathbf {x} _ {1}, \\mathbf {x} _ {2} \\right\\}} \\sum_ {i \\in \\mathcal {I} _ {\\mathbf {x} ^ {\\mathrm {M I M}}}} \\operatorname {K L} \\left(\\widetilde {\\mathbf {p}} _ {\\mathbf {x}} ^ {i}, \\mathbf {p} _ {\\mathbf {x} ^ {\\mathrm {M I M}}} ^ {i}\\right). \\tag {3}", + "image_path": "2ea72f7bc340e07e2034fa323957862aa27fb44412b867b5ab7e7057f4df943b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 215, + 284, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 284, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 284, + 227 + ], + "type": "text", + "content": "We simply sum " + }, + { + "bbox": [ + 55, + 215, + 284, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HoM}}" + }, + { + "bbox": [ + 55, + 215, + 284, + 227 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 215, + 284, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MIM}}" + }, + { + "bbox": [ + 55, + 215, + 284, + 227 + ], + "type": "text", + "content": " to obtain the final loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 236, + 225, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 225, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 225, + 248 + ], + "type": "text", + "content": "3.2. Fine-tuning for Low-shot Tasks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "text", + "content": "Figure 1b illustrates the fine-tuning stage. Once pretraining is complete, we preserve the teacher's vision adapter for downstream tasks, while discarding its heads and the whole student branch. To tailor the CLIP models to downstream FSL tasks, we attach to the adapter a vision (V) classifier with a cross-entropy (CE) loss " + }, + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{V}}" + }, + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "text", + "content": " for vision-based recognition. Meanwhile, we tune a learnable vision-language (VL) classifier with a CE loss " + }, + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VL}}" + }, + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "text", + "content": " [18, 53] for integration with language knowledge. The VL-classifier is initialized with a prompt ensembling technique. Specifically, for each class, we construct a set of class-specific text prompts fed to the frozen text encoder, and the text embeddings are averaged as the embedding of this class. Then, all per-class text embeddings are used to initialize the VL-classifier. The total loss is the sum of " + }, + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{V}}" + }, + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VL}}" + }, + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "text", + "content": ". For inference, we average the logits from two classifiers with a fusion weight, following Tip-Adapter (f) [12] and CaFo [13]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 458, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 296, + 553 + ], + "type": "text", + "content": "ImagineFSLLoRA. Additionally, we can further tune the image encoder, e.g., by introducing LoRA [17, 19, 39]. The resulting method is dubbed ImagineFSLLoRA. We note that such encoder tuning (ET) methods generally lead to significant performance gains, compared to those with frozen encoders. As such methods can adjust features from low-layers up to high-layers across the encoders, so can better adapt to specific tasks and domains." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 554, + 296, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 650 + ], + "type": "text", + "content": "Zero-shot Recognition. Our method is applicable to zero-shot recognition tasks. In this setting, we perform fine-tuning using solely the synthetic images, without touch of any real images. By tuning CLIP to diverse synthetic images of categories of interest, our method can significantly boost zero-shot recognition, thanks to its capability to learn additional knowledge from foundation models including LLMs and T2I models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 263, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 263, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 263, + 673 + ], + "type": "text", + "content": "3.3. Synthesizing both Captions and Images" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "The synthesizing pipeline is shown in Figure 2, where the CoT and ICL techniques are used to instruct LLMs and T2I models to generate captions and images." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 542, + 200 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 542, + 200 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 542, + 200 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 542, + 200 + ], + "type": "image", + "image_path": "7b53f478f544ebcede3801748fb603a5e57cb08ba88db1934676770c36a08e9b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 208, + 553, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 208, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 553, + 275 + ], + "type": "text", + "content": "Figure 2. Our pipeline synthesizes captions & images. We leverage GPT with the CoT to analyze key factors defining image descriptions and generate exemplar captions, which are then used by Llama for diverse caption generation via ICL. Finally, SD model generates images based on the synthesized captions. Note our pipeline is not limited to closed-source GPT (see Supplement D.4.)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 283, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 415 + ], + "type": "text", + "content": "Synthesizing Captions. We identify five key factors that constitute image descriptions: attribute, viewpoint, background (BG), lighting condition (LC), and cause of degradation of photos (CD). We design four distinct patterns for systematic caption generation, incorporating various combinations of these factors. The first one, Base-pattern, comprises attribute and viewpoint, providing a fundamental description of the concept. The rest ones, i.e., BG-pattern, LC-pattern, and CD-Pattern, enhance the Base-pattern by integrating additional BG, LC, and CD factor, respectively, enriching the depiction comprehensively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 605 + ], + "type": "text", + "content": "The pipeline begins by querying GPT-4 to analyze the key factors associated with a given concept, generating a detailed summary of visual properties and foundational elements essential for synthesizing diverse, realistic images that mimic real-world photography. We then design prompt template (PTe), guiding GPT-4 to generate a small number of high-quality exemplary captions for each pattern. Given the cost and time constraints of accessing GPT-4 API, we delegate extensive caption generation tasks to the locally deployed, lightweight Llama model. We develop PTe for Llama that utilizes exemplary captions generated by GPT-4 as in-context examples, randomly selecting one of the four patterns and the corresponding factors and examples to instantiate these templates. This strategy aims to elicit rich and nuanced responses from Llama, facilitating the creation of contextually enriched captions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": "Synthesizing Images. We use Stable Diffusion 3 (SD3) Medium [54], a multimodal diffusion transformer, for image generation. We construct the iBase, SyntIN1K, by adopting ImageNet-1K [55] class names as concepts. For each concept, we synthesize 300 captions, generating four images per caption. Additionally, we build SyntIN1.5K by selecting 0.5K additional concepts from ImageNet-21K class names. For each downstream dataset, we synthesize 300 captions per class and generate one image per caption," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30023" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 178 + ], + "type": "table", + "html": "
MethodT2IENSyn DataImageNetCaltechAircraftCarsFoodPetsFlowersDTDEuroSATSUNAvg Acc
IsSynth [18]**[18]-/73.4-/96.0-/46.7-/82.6-/87.3-/92.8-/86.5-/73.6-/87.1-/76.2-/80.2
CaFo [13]**[13]70.2/73.193.9/96.930.1/46.271.5/85.485.6/87.691.5/94.974.9/97.659.4/72.066.5/86.371.2/76.971.5/81.7
\\( \\text{IsSynth}^{\\dagger} \\)**Ours70.7/73.994.4/96.831.1/46.571.7/85.485.7/87.491.6/93.885.4/95.760.1/73.975.9/87.270.9/76.573.8/81.7
\\( \\text{CaFo}^{\\dagger} \\)**Ours70.2/73.994.2/96.930.6/47.472.6/85.786.1/87.691.8/94.976.1/97.860.1/72.568.7/86.771.3/76.972.2/82.0
ImagineFSL**Ours71.6/74.795.3/97.131.5/54.882.6/90.586.7/87.892.3/94.387.0/98.762.7/75.876.2/89.672.0/78.175.8/84.1
DISEF [17]**[17]-/73.9-/96.9-/63.9-/88.6-/87.1-/94.3-/98.9-/75.4-/94.3-/77.4-/85.1
DataDream [19]**[19]-/74.1-/96.9-/72.3-/92.4-/87.6-/94.8-/99.4-/81.6-/93.4-/77.5-/87.0
\\( \\text{DISEF}^{\\dagger} \\)**Ours71.0/74.093.2/97.032.0/67.972.6/91.286.3/87.292.5/94.588.0/98.960.5/76.082.3/94.370.6/77.174.9/85.8
\\( \\text{ImagineFSL}_{\\text{LoRA}} \\)**Ours71.8/75.295.4/97.934.0/74.182.8/92.986.8/88.392.8/95.490.2/99.763.7/78.082.7/95.072.8/78.877.3/87.5
", + "image_path": "9451b3f2cc838fe11bcfbda7f71b5fd0ec7815212fb03fb612634d391a06663f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 185, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 555, + 208 + ], + "type": "text", + "content": "Table 1. Comparison of 1-shot/16-shot recognition to prior methods using synthetic images. Bold: best results; underlined: second-best results. " + }, + { + "bbox": [ + 55, + 185, + 555, + 208 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 185, + 555, + 208 + ], + "type": "text", + "content": " Reproduced by us." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 215, + 227, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 227, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 227, + 226 + ], + "type": "text", + "content": "producing task-specific synthetic datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 227, + 295, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 227, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 54, + 227, + 295, + 287 + ], + "type": "text", + "content": "In Supplements B.1-B.3, we give details on our design philosophy regarding the factors, the queries for GPT-4 to analyze these factors, the PTe for GPT-4 to generate examples, and the PTe for in-context learning of Llama. In Table S-5, we present a summary of all synthetic datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 297, + 132, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 132, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 132, + 312 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 317, + 295, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 317, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 54, + 317, + 295, + 414 + ], + "type": "text", + "content": "We evaluate extensively our methods on few-shot recognition, domain generalization, and zero-shot recognition. Besides, we conduct a thorough ablation study on our method. Across the paper, we use CLIP ViT-B/16 as the visual encoder and its aligned counterpart as the textual encoder, unless otherwise specified. All experiments are conducted using PyTorch package [56] with GeForce RTX 4090 GPUs. Additional experiments can be found in Supplement D." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 421, + 112, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 112, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 112, + 434 + ], + "type": "text", + "content": "4.1. Setting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 437, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 558 + ], + "type": "text", + "content": "Datasets and Metrics. Following prior arts [11, 12], we benchmark few-shot and zero-shot recognition tasks on 11 datasets, including ImageNet [55], Caltech [57], Aircraft [58], Cars [59], Food [60], Pets [61], Flowers [62], DTD [63], EuroSAT [64], SUN [65], and UCF101 [66]. We assess domain generalization task on ImageNet-V2 [67], ImageNet-S [68], ImageNet-A [69], and ImageNet-R [70]. For meta-testing, we follow the common practice, randomly sampling 3 All-way-K-shot tasks per dataset, reporting the average accuracy (Avg Acc) as a percentage (\\%)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "type": "text", + "content": "Pretraining setup. We train the models using the AdamW optimizer with a cosine scheduler of learning rate. As in DINO [20], we use image augmentations including random resized crop, horizontal flip, color jittering, grayscale conversion, and Gaussian blur. For MIM, we utilize a blockwise masking strategy [27], where each image is randomly masked and the proportion of masked patches is randomly sampled. Hyperparameters are given in Supplement A.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 295, + 714 + ], + "type": "text", + "content": "Fine-tuning Setup. We follow the implementation of DataDream [19], adopting random resized crop, random horizontal flip, a set of color augmentations and RandAugment [71] as image augmentations for both ImagineFSL and ImagineFSLLORA. As in IsSynth [18], DISEF and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 215, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 553, + 323 + ], + "type": "text", + "content": "DataDream, we use mix training strategy that simultaneously uses real images and synthetic images in a batch. For input of the text encoder, we adopt an ensemble of textual templates hand-engineered by CLIP and Customized Prompts via Language models (CuPL) [72] provided by CaFo [13], unless otherwise specified. Following TipAdapter(f) [12], CaFo, DISEF and DataDream, we tune hyperparameters for each task on the validation sets. See Supplement A.3 for detailed hyperparameters of fine-tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 323, + 553, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 553, + 467 + ], + "type": "text", + "content": "LLMs and T2I models. To analyze the factors and generate the exemplary captions, we call GPT-4 API with a temperature of 0.5. Llama 3 8B is deployed locally to generate extensive captions, in which we set the temperature to 0.5 and Nucleus Sampling parameter to 0.7. We utilize a TensorRT optimized SD3 Medium model and generate " + }, + { + "bbox": [ + 313, + 323, + 553, + 467 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 323, + 553, + 467 + ], + "type": "text", + "content": " images, with the scale of classifier-free guidance set to 5 and the steps of denoising diffusion implicit model to 28. Similar to [17-19], we adopt a semantic filtering method that discards synthetic images misclassified by zero-shot CLIP classifier. We provide implementation details on synthesizing texts and images in Supplement B." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 473, + 438, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 438, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 438, + 486 + ], + "type": "text", + "content": "4.2. Few-shot Recognition" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 490, + 553, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 553, + 550 + ], + "type": "text", + "content": "After pretraining on iBase, we perform meta-testing on downstream few-shot tasks. For each task, we fine-tune the adapter and V-classifier alongside VL-classifier, using K-shot real images and task-specific synthetic images; finally, we conduct inference on test sets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 555, + 543, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 543, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 543, + 567 + ], + "type": "text", + "content": "4.2.1. Contrast with Synthetic Images based Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 553, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 618 + ], + "type": "text", + "content": "Table 1 compares to prior methods based on synthetic images in 1-/16-shot settings, including IsSynth [18], CaFo [13], DISEF [17], and DataDream [19]. For a fair comparison, we reproduce them with our synthetic images." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "First, we consider the methods that tune neither T2I models nor ENcoders (EN). Note that our reproduced results for IsSynth and CaFo in 16-shot setting are both higher than those reported in the original papers; we conjecture the reason is that our synthetic images are more suitable for few-shot tasks. Using the same synthetic data as ours, ImagineFSL performs better than them by a large margin of " + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": " in average accuracy in both 1-shot and 16-shot settings." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30024" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 71, + 294, + 228 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 294, + 228 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 294, + 228 + ], + "type": "table", + "html": "
Method1-shot2-shot4-shot8-shot16-shotAvg Acc
PTCoOp [11]67.670.874.076.880.073.8
CoCoOp [9]68.870.471.773.875.472.0
MaPLe [73]66.470.573.576.478.673.1
ProGrad [74]69.572.474.777.479.974.8
PLOT++ [75]70.774.076.979.682.176.7
PromptSRC [76]72.375.378.480.782.977.9
ATCLIP-Adapter [41]65.766.867.771.575.569.4
TIP-Adapter(f) [12]70.973.775.678.380.775.8
TaskRes [77]70.872.974.778.480.875.5
ImagineFSL76.177.579.482.284.479.9
ETCLIP-LoRA [39]72.575.577.480.383.077.7
ET+PTLLaMP [42]72.475.978.881.683.878.5
PT+ATGalLoP [24]72.876.479.182.284.579.0
ET+ATImagineFSLLoRA77.679.181.984.287.682.1
", + "image_path": "474b07c66c716a44661816038cac309291a5ccf0b50055a1f484947dcd1ddfc4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 237, + 295, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 295, + 270 + ], + "type": "text", + "content": "Table 2. Comparison of " + }, + { + "bbox": [ + 55, + 237, + 295, + 270 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 237, + 295, + 270 + ], + "type": "text", + "content": "-shot recognition to prior methods using only real images. **Bold:** best results; underlined: second-best results. **Supplement C** provides detailed results on 11 datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 281, + 295, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 295, + 423 + ], + "type": "text", + "content": "Next, we contrast the methods that tune CLIP encoders. DISEF and DataDream use the same technique to tune the image and text encoders with LoRA, while DataDream additionally tunes T2I models. In the 16-shot setting, our ImagineFSL" + }, + { + "bbox": [ + 54, + 281, + 295, + 423 + ], + "type": "inline_equation", + "content": "_{\\text{LoRA}}" + }, + { + "bbox": [ + 54, + 281, + 295, + 423 + ], + "type": "text", + "content": " outperforms DataDream on 10 out of 11 datasets; DataDream excels only on DTD, where we conjecture that fine-tuning T2I models with real images enhances realistic texture generation. Furthermore, with our synthetic images, the reproduced results for DISEF in the 16-shot setting surpass those reported in the original paper, while ImagineFSL" + }, + { + "bbox": [ + 54, + 281, + 295, + 423 + ], + "type": "inline_equation", + "content": "_{\\text{LoRA}}" + }, + { + "bbox": [ + 54, + 281, + 295, + 423 + ], + "type": "text", + "content": " also exceeds DISEF by " + }, + { + "bbox": [ + 54, + 281, + 295, + 423 + ], + "type": "inline_equation", + "content": "2.4\\% / 1.7\\%" + }, + { + "bbox": [ + 54, + 281, + 295, + 423 + ], + "type": "text", + "content": " in 1-shot/16-shot tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 431, + 288, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 288, + 443 + ], + "type": "text", + "content": "4.2.2. Contrast with Methods Using Only Real Images" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 447, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 602 + ], + "type": "text", + "content": "In Table 2, we compare our methods to prior techniques that do not employ synthetic images for K-shot tasks, including prompt tuning (PT), adapter tuning (AT), encoder tuning (ET), and hybrid tuning approaches. ImagineFSL significantly outperforms PT, AT, and ET across all shot scenarios, particularly in fewer-shot settings, e.g., it surpasses CLIP-LoRA (the runner-up) by " + }, + { + "bbox": [ + 55, + 447, + 295, + 602 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 55, + 447, + 295, + 602 + ], + "type": "text", + "content": " in 2-shot and " + }, + { + "bbox": [ + 55, + 447, + 295, + 602 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 55, + 447, + 295, + 602 + ], + "type": "text", + "content": " in 1-shot settings. Additionally, while GalLoP introduces a novel dropout technique for learning global and local prompts via a single linear layer adapter, and LLaMP merges PT and ET using LoRA, ImagineFSL-LoRA leads consistently and substantially in each K-shot setting with an average accuracy improvement of " + }, + { + "bbox": [ + 55, + 447, + 295, + 602 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 55, + 447, + 295, + 602 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 187, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 187, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 187, + 623 + ], + "type": "text", + "content": "4.3. Domain Generalization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Table 3 compares our methods with prior approaches on the domain generalization task. In this task, a model is trained on 16 shots per class from ImageNet (IN, source) and tested on four target datasets: IN-V2, IN-S, IN-A, and IN-R. For our methods, we specifically synthesize images for IN-S and IN-R datasets, while synthetic images for ImageNet are used for IN-V2 and IN-A. We fine-tune the models using" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 455, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 455, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 455, + 84 + ], + "type": "text", + "content": "both the real and synthetic images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 85, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 168 + ], + "type": "text", + "content": "First, we observe that our methods outperform the second-best method (i.e., CoCoLe) by over " + }, + { + "bbox": [ + 313, + 85, + 555, + 168 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 313, + 85, + 555, + 168 + ], + "type": "text", + "content": " in average accuracy. Notably, our models achieve performance that matches the top performer, GalLoP, on IN-V2 and Co-CoLe on IN-A. Furthermore, we establish new state-of-the-art results on both IN-S and IN-R, achieving accuracies of " + }, + { + "bbox": [ + 313, + 85, + 555, + 168 + ], + "type": "inline_equation", + "content": "53.8\\%" + }, + { + "bbox": [ + 313, + 85, + 555, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 85, + 555, + 168 + ], + "type": "inline_equation", + "content": "80.0\\%" + }, + { + "bbox": [ + 313, + 85, + 555, + 168 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 177, + 441, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 441, + 189 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 441, + 189 + ], + "type": "text", + "content": "4.4. Zero-shot Recognition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 194, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 326 + ], + "type": "text", + "content": "We compare to previous methods on zero-shot recognition in Table 4. We see that our methods are significantly better than most of the PT methods on each individual datasets, including VCD [79], Sus-X [80], TPT [81] and DiffTPT [82]. Compared to state-of-the-art DMN [83], ImagineFSL and ImagineFSLoRA outperform by " + }, + { + "bbox": [ + 313, + 194, + 555, + 326 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 313, + 194, + 555, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 194, + 555, + 326 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 313, + 194, + 555, + 326 + ], + "type": "text", + "content": " in average accuracy, respectively, with the latter ranking first in 7 out of 11 datasets. These results suggest that synthetic images hold promise for enhancing CLIP zero-shot recognition, and our methods are likely to improve further with advancements in generative T2I models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 334, + 441, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 441, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 441, + 346 + ], + "type": "text", + "content": "4.5. Empirical Complexity" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 352, + 554, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 554, + 399 + ], + "type": "text", + "content": "Pretraining. On SyntIN1K, pretraining with ViT-B/16 takes " + }, + { + "bbox": [ + 313, + 352, + 554, + 399 + ], + "type": "inline_equation", + "content": "\\sim 9" + }, + { + "bbox": [ + 313, + 352, + 554, + 399 + ], + "type": "text", + "content": " hours on 4 RTX 4090 GPUs (" + }, + { + "bbox": [ + 313, + 352, + 554, + 399 + ], + "type": "inline_equation", + "content": "\\sim 40" + }, + { + "bbox": [ + 313, + 352, + 554, + 399 + ], + "type": "text", + "content": " hours for ViT-L/14). This task-agnostic stage is performed once and then widely fit for downstream tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 400, + 554, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 507 + ], + "type": "text", + "content": "Fine-tuning. We compare methods on UCF101 in the 16-shot setting with a batch size of 128 to ensure a fair comparison. We apply LoRA with a rank of 16 across methods involving encoder tuning (ET). Specifically, for DISEF, DataDream, and ImagineFSLloRA, LoRA is applied to the image encoder. Additionally, for DISEF and DataDream, LoRA is also applied to the text encoder. We note that DISEF and DataDream share similar empirical complexity, as they use the same technique for fine-tuning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "text", + "content": "Table 5 presents the results evaluated on one RTX 4090 GPU, with train and test latency per iteration. Among methods that do not involve " + }, + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "inline_equation", + "content": "ET" + }, + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "text", + "content": ", ImagineFSL has comparable latency and memory usage to TIP-Adapter(f) and IsSynth. While it has more parameters, they make up only " + }, + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "inline_equation", + "content": "5.08\\%" + }, + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "text", + "content": " of total encoder parameters. Moreover, it outperforms CoOp significantly in training speed. Compared to methods that involve " + }, + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "inline_equation", + "content": "ET" + }, + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "text", + "content": ", i.e., DISEF and DataDream, ImagineFSL" + }, + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "inline_equation", + "content": "_{\\text{LoRA}}" + }, + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "text", + "content": " achieves a two-fold improvement in training speed and consumes only half the memory, despite a small increase in the percentage of parameters (from " + }, + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "inline_equation", + "content": "1.07\\%" + }, + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "inline_equation", + "content": "5.89\\%" + }, + { + "bbox": [ + 313, + 508, + 555, + 639 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 647, + 408, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 408, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 408, + 661 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "We conduct ablation analysis in 1-shot/16-shot settings on ImageNet, Aircraft, Flowers and EuroSAT with CLIP ViTB/16 as visual encoder, unless otherwise specified. We provide more ablations in Supplement D.2." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30025" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 267, + 201 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 267, + 201 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 267, + 201 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 267, + 201 + ], + "type": "table", + "html": "
MethodSourceTarget
ImageNet-V2-S-A-RAvg Acc
PTCoOp [11]71.564.248.049.775.259.3
CoCoOp [9]71.064.148.850.676.259.9
MaPLe [73]70.764.149.250.977.060.3
PromptSRC [76]71.364.449.650.977.860.7
CoCoLe [14]73.965.950.951.878.961.9
ATTaskRes [77]73.165.349.150.477.760.6
GraphAdapter [78]73.465.649.250.677.760.8
ImagineFSL74.767.052.751.579.762.7
ET+ATImagineFSLloRA75.267.553.851.580.063.2
PT+ATGalLoP [24]75.167.549.550.377.861.3
", + "image_path": "8af7a58559fe92527ac553bf3c07489ea747b0bee788033dc843082464cccb8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 278, + 71, + 558, + 202 + ], + "blocks": [ + { + "bbox": [ + 278, + 71, + 558, + 202 + ], + "lines": [ + { + "bbox": [ + 278, + 71, + 558, + 202 + ], + "spans": [ + { + "bbox": [ + 278, + 71, + 558, + 202 + ], + "type": "table", + "html": "
MethodImageNetCiteLectAircraftCarsFoodPetsFlowersETDEmoSATSYNUCF101AvgAcc
PTVCD [79]68.0-----88.586.9-45.648.8-----
Sus-X [80]69.994.028.766.186.190.673.854.657.567.766.667.766.668.768.768.7
TPT [81]69.094.224.866.984.787.869.047.842.465.568.065.568.065.565.565.5
DiffTPT [82]70.392.525.667.087.288.270.147.043.165.768.268.265.965.965.965.9
DMN [83]72.395.330.068.085.192.074.555.959.470.272.570.570.570.570.570.5
CODER [84]71.5----89.892.055.760.5-------
ATImagineFSL70.994.730.278.686.592.373.959.265.670.373.272.372.372.372.372.3
ET+ATImagineFSL-LoRA71.094.730.478.786.592.373.660.871.770.373.373.373.073.073.073.0
", + "image_path": "582ed57027777af0a173454f4a022026f37b7cd3656dda8aa0ad693e99ac06a3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 229, + 293, + 335 + ], + "blocks": [ + { + "bbox": [ + 56, + 211, + 276, + 221 + ], + "lines": [ + { + "bbox": [ + 56, + 211, + 276, + 221 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 276, + 221 + ], + "type": "text", + "content": "Table 3. Results of domain generalization on 16-shot setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 229, + 293, + 335 + ], + "lines": [ + { + "bbox": [ + 59, + 229, + 293, + 335 + ], + "spans": [ + { + "bbox": [ + 59, + 229, + 293, + 335 + ], + "type": "table", + "html": "
MethodLatencyMemory (G)Params (M)(%)
Train (ms)Test (ms)
PTCoOp [11]298.360.43.050.010.01
-TIP-Adapter(f) [12]70.261.52.730.830.55
ATIsSynth [18]69.960.32.740.050.03
ImagineFSL93.165.43.007.615.08
ETDISEF [17]544.360.820.641.571.07
DataDream [19]
ET+ATImagineFSLLoRA265.465.511.828.825.89
", + "image_path": "f2819a3f569d99b40a610a7a2a47df832c0772937546071f0d7dea02e2d2ff24.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 295, + 366 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 366 + ], + "type": "text", + "content": "Table 5. Comparison of empirical complexity of fine-tuning. We evaluate all methods on a single RTX 4090 GPU." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 376, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 508 + ], + "type": "text", + "content": "Is pretraining necessary? Table 6a compares the Non-Pretraining (NoP) setting with the pretraining setting. In NoP, we directly train the vision adapter and vision (V) classifier using either real images alone (row 1) or a mix of synthetic and real images (row 2). For the pretraining setting, we first train the vision adapter via either SL (row 3) or Self-SL on SyntIN-1K (row 4), then fine-tune the vision adapter and V-classifier. During testing, we combine the V-classifier's predictions with CLIP zero-shot predictions (i.e., the frozen VL-classifier), using CLIP's handcrafted templates as input to the text encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "text", + "content": "In NoP, mixing real and synthetic images substantially improves performance compared to using real images alone, demonstrating that our synthetic images effectively mitigate data scarcity. Furthermore, compared to NoP, pretraining with SL yields notable gains, particularly in 1-shot average accuracy " + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "(+0.9\\%)" + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "text", + "content": ". Additionally, our Self-SL method (i.e., HoM-DINO) surpasses SL by " + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "2.3\\% / 1.1\\%" + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "text", + "content": " in 1-shot/16-shot average accuracy, indicating that Self-SL better learns transferable features from synthetic to real image domains. These results highlight the importance of self-supervised pretraining on SyntIN1K for capturing rich knowledge from diverse synthetic images. In Figure S-3 of Supplement D.5, we visualize self-attention maps of different methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "HoM-DINO against other Self-SL methods. Table 6b compares to three typical Self-SL approaches including DINOv2, SynCLR and MAE. DINOv2, relying only on positive examples, is a self-distillation method that introduces" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 211, + 516, + 221 + ], + "lines": [ + { + "bbox": [ + 310, + 211, + 516, + 221 + ], + "spans": [ + { + "bbox": [ + 310, + 211, + 516, + 221 + ], + "type": "text", + "content": "Table 4. Results of zero-shot recognition on 11 datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 231, + 553, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 553, + 279 + ], + "type": "text", + "content": "MIM loss into DINO loss. SynCLR, depending on both positive and negative samples, is a contrastive method that combines multi-positive contrastive loss [85] and MIM loss. MAE is a representative masked image modeling method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 280, + 553, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 553, + 352 + ], + "type": "text", + "content": "It can be seen that MAE clearly underperforms and we conjecture its reconstruction mechanism in the pixel space is not fit for the few-shot scenarios. SynCLR achieves very competitive performance but still nontrivially lags behind ours. Our HoM-DINO builds upon and outperforms DINOv2 by " + }, + { + "bbox": [ + 313, + 280, + 553, + 352 + ], + "type": "inline_equation", + "content": "2.1\\% / 0.7\\%" + }, + { + "bbox": [ + 313, + 280, + 553, + 352 + ], + "type": "text", + "content": " in 1-shot/16-shot accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 354, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 604 + ], + "type": "text", + "content": "How image representation (image-repr) and view construction impact? First, from Table 6c we observe that, with [CLS] token as image-repr and HandAug, MIM significantly improves over without MIM (rows 1 and 2), indicating prediction of masked patches is helpful for FSL. Second, incorporating local patches via average pooling of all patches (AvgP) and attention pooling of all tokens (AttnP) slightly improved performance (rows 4 and 5). Using distribution-based image representations further improves performance significantly (rows 6 and 8); meanwhile, our HoM is slightly better than " + }, + { + "bbox": [ + 313, + 354, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^2\\mathrm{DeNet}" + }, + { + "bbox": [ + 313, + 354, + 554, + 604 + ], + "type": "text", + "content": " while having one fourth less dimension (Dim). SyntAug further improves HoM (rows 8 and 9). However, MP underperforms despite its usage of moments, which may be due to its lack of structural normalization mechanism that has proven to be critical for high-order representations [86]. Finally, as dimension of HoM is higher than [CLS] token, for a fair comparison, we add a linear layer after the adapter that increases [CLS] token to 2048-Dim. However, this remains significantly inferior to HoM (rows 3 and 8), suggesting that advantage of HoM lies in exploiting patches via statistical moments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "content": "Influence of textual prompts. As shown in Table 6d, when used individually, the CuPL [72] slightly outperforms the hand-engineered prompt templates of CLIP. However, combining both CuPL and CLIP prompts results in a further improvement over the individual methods. Therefore, unless otherwise specified, we use an ensemble of CLIP and CuPL prompts throughout the paper." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "VL-classifier (clf) tuning and encoder tuning. Rather than using a frozen VL-clf, we tune the clf's weights (i.e.," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30026" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 309, + 129 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 309, + 129 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 309, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 309, + 129 + ], + "type": "table", + "html": "
PretrainSyn ImgImageNetAircraftFlowersEuroSATAvg Acc
1-✘68.8/72.824.1/48.578.9/96.560.4/84.458.1/75.6
2-✓70.9/73.927.5/49.384.1/96.566.2/86.262.2/76.5
3SL✓71.0/73.928.7/49.585.3/97.067.3/86.863.1/76.8
4HoM-DINO✓71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
", + "image_path": "ca6c5cc22f9a019b7c285e45172904ab442ae54960be9b683f262bd2644e1571.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 142, + 309, + 267 + ], + "blocks": [ + { + "bbox": [ + 132, + 131, + 225, + 140 + ], + "lines": [ + { + "bbox": [ + 132, + 131, + 225, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 131, + 225, + 140 + ], + "type": "text", + "content": "(a) Necessity of pretraining." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 142, + 309, + 267 + ], + "lines": [ + { + "bbox": [ + 56, + 142, + 309, + 267 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 309, + 267 + ], + "type": "table", + "html": "
Image \nReprDimViewMIMImageNetAircraftFlowersEuroSATAvg Acc
1[CLS]512HA×71.2/73.527.2/50.083.9/97.266.7/86.962.3/76.9
2[CLS]512HA✓71.1/73.928.9/50.685.2/97.567.8/86.863.3/77.2
3[CLS]2048HA✓71.4/73.629.0/49.984.3/97.669.8/86.863.6/77.0
4AvgP512HA✓71.4/73.729.2/50.384.2/97.469.6/87.163.6/77.1
5AttnP512HA✓71.3/73.228.8/56.985.7/97.669.4/87.463.8/77.2
6G²De- \nNet8385HA✓71.1/73.629.3/50.385.3/97.570.2/87.764.0/77.3
7MP2087HA✓71.3/73.629.3/48.982.7/97.468.6/86.263.0/76.5
8HoM2048HA✓71.4/74.029.1/50.985.6/98.172.0/87.364.5/77.6
9HoM2048SA✓71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
", + "image_path": "cafc4673a8da01719e3a672584a9775c5f5209e5ff9c493cc4540777420475b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 281, + 310, + 346 + ], + "blocks": [ + { + "bbox": [ + 59, + 269, + 299, + 279 + ], + "lines": [ + { + "bbox": [ + 59, + 269, + 299, + 279 + ], + "spans": [ + { + "bbox": [ + 59, + 269, + 299, + 279 + ], + "type": "text", + "content": "(c) Effect of image representation and view. HA: HandAug; SA: SyntAug." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 281, + 310, + 346 + ], + "lines": [ + { + "bbox": [ + 57, + 281, + 310, + 346 + ], + "spans": [ + { + "bbox": [ + 57, + 281, + 310, + 346 + ], + "type": "table", + "html": "
ModelBase setSizeImageNetAircraftFlowersEuroSATAvg Acc
ViT-B/16SyntIN1K0.1M71.1/74.431.0/54.887.0/98.874.8/88.166.0/79.0
0.3M71.6/74.731.5/54.887.0/98.776.2/89.666.6/79.5
SyntIN1.5K0.45M71.7/74.732.9/54.787.7/98.675.2/89.766.9/79.4
ViT-L/14SyntIN1K0.3M78.4/80.742.1/68.190.0/99.580.6/92.772.8/85.3
", + "image_path": "4662d839ee673b3e5defa356e29277587d9870d7aaf8d1f220c8c93773d10460.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 75, + 347, + 282, + 357 + ], + "lines": [ + { + "bbox": [ + 75, + 347, + 282, + 357 + ], + "spans": [ + { + "bbox": [ + 75, + 347, + 282, + 357 + ], + "type": "text", + "content": "(f) Scaling of ImagineFSL versus data size and model capacity." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 322, + 71, + 551, + 128 + ], + "blocks": [ + { + "bbox": [ + 322, + 71, + 551, + 128 + ], + "lines": [ + { + "bbox": [ + 322, + 71, + 551, + 128 + ], + "spans": [ + { + "bbox": [ + 322, + 71, + 551, + 128 + ], + "type": "table", + "html": "
Self-SLImageNetAircraftFlowersEuroSATAvg Acc
DINOv271.1/73.928.9/50.685.2/97.567.8/86.863.3/77.2
SynCLR71.5/74.027.4/51.084.4/97.774.4/86.964.4/77.4
MAE71.2/74.326.4/47.683.6/97.466.5/86.761.9/76.3
HoM-DINO71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
", + "image_path": "bf211e990b9229cb3d139862d0571ce7636f8d39c3b0f928e137ce039c5e90ce.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 322, + 144, + 551, + 196 + ], + "blocks": [ + { + "bbox": [ + 381, + 130, + 490, + 140 + ], + "lines": [ + { + "bbox": [ + 381, + 130, + 490, + 140 + ], + "spans": [ + { + "bbox": [ + 381, + 130, + 490, + 140 + ], + "type": "text", + "content": "(b) Different Self-SL techniques." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 144, + 551, + 196 + ], + "lines": [ + { + "bbox": [ + 322, + 144, + 551, + 196 + ], + "spans": [ + { + "bbox": [ + 322, + 144, + 551, + 196 + ], + "type": "table", + "html": "
CLIPCuPLImageNetAircraftFlowerEuroSATAvg Acc
✓✗71.5/74.329.4/51.286.3/98.374.5/87.865.4/77.9
✗✓71.6/74.330.7/51.685.9/98.472.0/87.865.0/78.0
✓✓71.6/74.330.8/51.586.3/98.373.7/87.865.6/78.0
", + "image_path": "f24deb67dfebf83ee602bdf7c8a53b9bdaa696e506229e0056597cbcb38071dc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 322, + 212, + 551, + 265 + ], + "blocks": [ + { + "bbox": [ + 383, + 198, + 488, + 209 + ], + "lines": [ + { + "bbox": [ + 383, + 198, + 488, + 209 + ], + "spans": [ + { + "bbox": [ + 383, + 198, + 488, + 209 + ], + "type": "text", + "content": "(d) Influence of textual prompts." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 212, + 551, + 265 + ], + "lines": [ + { + "bbox": [ + 322, + 212, + 551, + 265 + ], + "spans": [ + { + "bbox": [ + 322, + 212, + 551, + 265 + ], + "type": "table", + "html": "
V-clfVL-clfI-ENImageNetAircraftFlowersEuroSATAvg Acc
71.6/74.330.8/51.586.3/98.373.7/87.865.6/78.0
71.6/74.731.5/54.887.0/98.776.2/89.666.6/79.5
71.8/75.234.0/74.190.2/99.782.7/95.069.7/86.0
", + "image_path": "4dc443b9c2890251d32b0bbc2b53c77afa91be4ae45f1ad4146b0cc3e2e0dfa8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 322, + 279, + 551, + 338 + ], + "blocks": [ + { + "bbox": [ + 328, + 266, + 544, + 276 + ], + "lines": [ + { + "bbox": [ + 328, + 266, + 544, + 276 + ], + "spans": [ + { + "bbox": [ + 328, + 266, + 544, + 276 + ], + "type": "text", + "content": "(e) Effects of tuning VL-classifier (clf) and Image ENcoder (I-EN)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 279, + 551, + 338 + ], + "lines": [ + { + "bbox": [ + 322, + 279, + 551, + 338 + ], + "spans": [ + { + "bbox": [ + 322, + 279, + 551, + 338 + ], + "type": "table", + "html": "
MethodChain-of-ThoughtFidelity ↓Diversity ↑
Handcrafted prompts×45.920.95
Ours×40.071.06
Ours✓32.591.21
", + "image_path": "a9c2b0a8f0f84d4b2432adf75ef6953b758c19b5750056b4ea0a6c908356cb2e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 340, + 340, + 531, + 350 + ], + "lines": [ + { + "bbox": [ + 340, + 340, + 531, + 350 + ], + "spans": [ + { + "bbox": [ + 340, + 340, + 531, + 350 + ], + "type": "text", + "content": "(g) Effect of Chain-of-Thought (CoT) on image generation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 12 + }, + { + "bbox": [ + 65, + 365, + 543, + 377 + ], + "lines": [ + { + "bbox": [ + 65, + 365, + 543, + 377 + ], + "spans": [ + { + "bbox": [ + 65, + 365, + 543, + 377 + ], + "type": "text", + "content": "Table 6. Ablation study on the proposed methods in the 1-shot/16-shot settings. Bold: best results; underlined: second-best results." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 386, + 295, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 386, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 54, + 386, + 295, + 482 + ], + "type": "text", + "content": "per-class text embeddings), implementing the ImagineFSL method. Table 6e shows this tuning significantly improves performance across most datasets. Furthermore, tuning the Image Encoder (I-EN) with LoRA (i.e., ImagineFSL" + }, + { + "bbox": [ + 54, + 386, + 295, + 482 + ], + "type": "inline_equation", + "content": "_{\\text{LoRA}}" + }, + { + "bbox": [ + 54, + 386, + 295, + 482 + ], + "type": "text", + "content": ") further enhances performance, particularly on Aircraft, Flowers, and EuroSAT. We hypothesize that I-EN tuning refines feature representations across the entire network, leading to better task and domain adaptation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 484, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 484, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 54, + 484, + 295, + 639 + ], + "type": "text", + "content": "Scaling capability. As shown in Table 6f, when using ViT-B/16 as the image encoder on SyntIN1K, we observe slight improvements in both 1-shot and 16-shot accuracies as data volume increases from 0.1M to 0.3M. Additionally, expanding the dataset to SyntIN1.5K slightly enhances 1-shot accuracy while 16-shot accuracy remains stable, suggesting potential saturation in performance. This could be attributed to the limitations of the frozen encoder, which might not effectively utilize the expanded dataset or additional concepts. However, switching to ViT-L/14 results in substantial gains of " + }, + { + "bbox": [ + 54, + 484, + 295, + 639 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 54, + 484, + 295, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 484, + 295, + 639 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 54, + 484, + 295, + 639 + ], + "type": "text", + "content": " in 1-shot and 16-shot accuracies, respectively, underscoring the scalability of our approach with larger CLIP models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": "How does CoT affect generated images? We ablate our generation method by comparing settings w/ and w/o CoT prompting. We also compare against a baseline using handcrafted prompts, i.e., \"a photo of a {class}\", for image generation. In the CoT (w/o) setting, we prompt Llama to generate captions using factor names for individual patterns. As" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 386, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 495 + ], + "type": "text", + "content": "in [87], we compute the FID score to measure fidelity between synthetic and real images on the ImageNet validation set. For diversity, we use a pretrained DINOv2 ViT-B/16 to extract features and compute the standard deviation per class. Table 6g shows that our method w/o CoT outperforms handcrafted prompts. Using CoT further improves diversity and fidelity, underscoring its role in generating diverse, realistic images. Supplement D.5 qualitatively analyzes our generated texts and images." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 521, + 471, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 471, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 471, + 533 + ], + "type": "text", + "content": "5. Conclusion and Limitations" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 556, + 713 + ], + "type": "text", + "content": "We introduce ImagineFSL, a novel methodology for CLIP adaptation. It involves self-supervised pretraining on large-scale purely synthetic images, followed by fine-tuning with real and task-specific synthetic images. Unlike previous methods, we treat synthetic images as standalone knowledge repositories of diverse concepts. To generate diverse and realistic synthetic images, we leverage CoT and in-context learning techniques, utilizing the capabilities of LLMs and T2I models. Despite strong performance across various recognition tasks, our approach has certain limitations, such as pretraining overhead and underutilization of synthetic captions, which are discussed in Supplement E. We hope our work inspires further exploration of synthetic images for data-scarce tasks." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30027" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Yisheng Song, Ting Wang, Puyu Cai, Subrota K Mondal, and Jyoti Prakash Sahoo. A comprehensive survey of few-shot learning: Evolution, applications, challenges, and opportunities. ACM Computing Surveys, 55(13s):1-40, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 179 + ], + "type": "text", + "content": "[2] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. In International Conference on Learning Representations, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 224 + ], + "type": "text", + "content": "[3] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In European Conference on Computer Vision, pages 266-282, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 225, + 295, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 295, + 269 + ], + "type": "text", + "content": "[4] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, pages 1126-1135, 2017. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "type": "text", + "content": "[5] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in Neural Information Processing Systems, pages 3630-3638, 2016. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 434 + ], + "type": "text", + "content": "[6] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millicah, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 436, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 491 + ], + "type": "text", + "content": "[7] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 491, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 557 + ], + "type": "text", + "content": "[8] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 558, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 295, + 612 + ], + "type": "text", + "content": "[9] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 613, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 668 + ], + "type": "text", + "content": "[10] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23681-23690, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[11] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[12] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. TipAdapter: Training-free adaptation of CLIP for few-shot classification. In European Conference on Computer Vision, pages 493-510, 2022. 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 195 + ], + "type": "text", + "content": "[13] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15211-15222, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "text", + "content": "[14] Yi Zhang, Ke Yu, Siqi Wu, and Zhihai He. Conceptual codebook learning for vision-language models. In European Conference on Computer Vision, pages 235-251, 2024. 1, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 232, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 297 + ], + "type": "text", + "content": "[15] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 299, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 555, + 354 + ], + "type": "text", + "content": "[16] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 355, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 399 + ], + "type": "text", + "content": "[17] Victor G Turrisi da Costa, Nicola Dall'Asen, Yiming Wang, Nicu Sebe, and Elisa Ricci. Diversified in-domain synthesis with efficient fine-tuning for few-shot classification. arXiv preprint arXiv:2312.03046, 2023. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 400, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 400, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 317, + 400, + 553, + 455 + ], + "type": "text", + "content": "[18] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? In International Conference on Learning Representations, 2023. 2, 4, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 456, + 553, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 500 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 500 + ], + "type": "text", + "content": "[19] Jae Myung Kim, Jessica Bader, Stephan Alaniz, Cordelia Schmid, and Zeynep Akata. DataDream: Few-shot guided dataset generation. In European Conference on Computer Vision, pages 252-268, 2024. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 502, + 553, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 502, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 317, + 502, + 553, + 556 + ], + "type": "text", + "content": "[20] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 558, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 553, + 667 + ], + "type": "text", + "content": "[21] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOV2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024:1-31, 2024. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[22] Carl Doersch, Ankush Gupta, and Andrew Zisserman. CrossTransformers: spatially-aware few-shot transfer. In Advances in Neural Information Processing Systems, pages 21981-21993, 2020. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30028" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "text", + "content": "[23] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. DeepEMD: Differentiable earth mover's distance for few-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5632-5648, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "type": "text", + "content": "[24] Marc Lafon, Elias Ramzi, Clément Rambour, Nicolas Audebert, and Nicolas Thome. GalLoP: Learning global and local prompts for vision-language models. In European Conference on Computer Vision, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 294, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 294, + 219 + ], + "type": "text", + "content": "[25] Qilong Wang, Peihua Li, and Lei Zhang. " + }, + { + "bbox": [ + 56, + 165, + 294, + 219 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^2\\mathrm{DeNet}" + }, + { + "bbox": [ + 56, + 165, + 294, + 219 + ], + "type": "text", + "content": ": Global Gaussian distribution embedding network and its application to visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6507-6516, 2017. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 222, + 294, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 222, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 57, + 222, + 294, + 243 + ], + "type": "text", + "content": "[26] Jun Shao. Mathematical Statistics. Springer, New York, 2003. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 246, + 294, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 246, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 294, + 290 + ], + "type": "text", + "content": "[27] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. iBOT: Image BERT pretraining with online tokenizer. In International Conference on Learning Representations, 2022. 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 292, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 292, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 57, + 292, + 294, + 336 + ], + "type": "text", + "content": "[28] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 393 + ], + "type": "text", + "content": "[29] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, pages 24824-24837, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 396, + 294, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 294, + 450 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 294, + 450 + ], + "type": "text", + "content": "[30] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 453, + 294, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 506 + ], + "type": "text", + "content": "[31] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floreccia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "type": "text", + "content": "[32] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "type": "text", + "content": "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 49250-49267, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "type": "text", + "content": "[34] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[35] Zhuoran Yu, Chenchen Zhu, Sean Culatana, Raghuraman Krishnamoorthi, Fanyi Xiao, and Yong Jae Lee. Diversify, don't fine-tune: Scaling up visual recognition training with synthetic images. arXiv preprint arXiv: 2312.02253, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "text", + "content": "[36] Yonglong Tian, Lijie Fan, Kaifeng Chen, Dina Katabi, Dilip Krishnan, and Phillip Isola. Learning vision from models rivals learning vision from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15887-15898, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "type": "text", + "content": "[37] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1401-1411, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 554, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 554, + 240 + ], + "type": "text", + "content": "[38] Adrian Bulat and Georgios Tzimiropoulos. LASP: Text-to-text optimization for language-aware soft prompting of vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23232-23241, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "type": "text", + "content": "[39] Maxime Zanella and Ismail Ben Ayed. Low-rank few-shot adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1593-1603, 2024. 3, 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 554, + 341 + ], + "type": "text", + "content": "[40] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 343, + 554, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 554, + 397 + ], + "type": "text", + "content": "[41] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. CLIP-Adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2023. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 399, + 555, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 454 + ], + "type": "text", + "content": "[42] Zhaoheng Zheng, Jingmin Wei, Xuefeng Hu, Haidong Zhu, and Ram Nevatia. Large language models are good prompt learners for low-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28453-28462, 2024. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 456, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 554, + 510 + ], + "type": "text", + "content": "[43] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 512, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 554, + 578 + ], + "type": "text", + "content": "[44] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 579, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 554, + 644 + ], + "type": "text", + "content": "[45] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4171–4186, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 679 + ], + "type": "text", + "content": "[46] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pretraining of image transformers. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[47] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Con-" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30029" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 159-168, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[48] Tsung-Yu Lin, Aruni RoyChowdhury, and Subhransu Maji. Bilinear cnn models for fine-grained visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1449-1457, 2015. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 294, + 196 + ], + "type": "text", + "content": "[49] Peihua Li, Jiangtao Xie, Qilong Wang, and Zilin Gao. Towards faster training of global covariance pooling networks by iterative matrix square root normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 947-955, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[50] Yue Song, Nicu Sebe, and Wei Wang. On the eigenvalues of global covariance pooling for fine-grained visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3554-3566, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "text", + "content": "[51] Tetsu Matsukawa, Takahiro Okabe, Einoshin Suzuki, and Yoichi Sato. Hierarchical Gaussian descriptors with application to person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(9):2179–2194, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 300, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 294, + 353 + ], + "type": "text", + "content": "[52] Mingze Gao, Qilong Wang, Zhenyi Lin, Pengfei Zhu, Qinghua Hu, and Jingbo Zhou. Tuning pre-trained model via moment probing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11769-11779, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 356, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 356, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 356, + 294, + 421 + ], + "type": "text", + "content": "[53] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 423, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 294, + 498 + ], + "type": "text", + "content": "[54] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633, 2024. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[55] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 622 + ], + "type": "text", + "content": "[56] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035, 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 294, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 688 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 688 + ], + "type": "text", + "content": "[57] Li Fei-Fei, R Fergus, and P Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Worskshops, pages 178–178, 2004. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[58] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classi" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "fication of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[59] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 554–561, 2013. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 142, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 553, + 184 + ], + "type": "text", + "content": "[60] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, pages 446–461, 2014. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[61] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3498-3505, 2012. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 274 + ], + "type": "text", + "content": "[62] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 277, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 330 + ], + "type": "text", + "content": "[63] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 387 + ], + "type": "text", + "content": "[64] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "text", + "content": "[65] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 445, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 477 + ], + "type": "text", + "content": "[66] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 479, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 521 + ], + "type": "text", + "content": "[67] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In International Conference on Machine Learning, pages 5389-5400, 2019. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 524, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 524, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 524, + 553, + 567 + ], + "type": "text", + "content": "[68] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 569, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 612 + ], + "type": "text", + "content": "[69] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 689 + ], + "type": "text", + "content": "[70] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[71] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. RandAugment: Practical automated data augmentation with" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30030" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "a reduced search space. In Advances in Neural Information Processing Systems, pages 18613-18624, 2020. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[72] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15645-15655, 2023. 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "text", + "content": "[73] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. MaPLe: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "text", + "content": "[74] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 293 + ], + "type": "text", + "content": "[75] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. In International Conference on Learning Representations, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 295, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 295, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 295, + 294, + 360 + ], + "type": "text", + "content": "[76] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "type": "text", + "content": "[77] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10899-10909, 2023. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 405, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 294, + 458 + ], + "type": "text", + "content": "[78] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. GraphAdapter: Tuning vision-language models with dual knowledge graph. In Advances in Neural Information Processing Systems, pages 13448-13466, 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[79] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In International Conference on Learning Representations, pages 1-17, 2023. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "text", + "content": "[80] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. SuS-X: Training-free name-only transfer of vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2725-2736, 2023. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[81] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. In Advances in Neural Information Processing Systems, pages 14274-14289, 2022. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 604, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 658 + ], + "type": "text", + "content": "[82] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[83] Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versatile adaptation approach for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28718-28728, 2024. 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 284 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[84] Chao Yi, Lu Ren, De-Chuan Zhan, and Han-Jia Ye. Leveraging cross-modal neighbor representation for improved CLIP classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27402-27411, 2024. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[85] Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. StableRep: Synthetic images from text-to-image models make strong visual representation learners. In Advances in Neural Information Processing Systems, pages 48382-48402, 2023. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "text", + "content": "[86] Peihua Li, Jiangtao Xie, Qilong Wang, and Wangmeng Zuo. Is second-order information helpful for large-scale visual recognition? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2070-2078, 2017. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 555, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 284 + ], + "type": "text", + "content": "[87] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382-7392, 2024. 8" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "30031" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/e6b1bb30-e9d5-4708-960e-31ff7e5e9a5b_content_list.json b/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/e6b1bb30-e9d5-4708-960e-31ff7e5e9a5b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f2f2f576346bd7256bdbd57c53e4cb0489cd6040 --- /dev/null +++ b/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/e6b1bb30-e9d5-4708-960e-31ff7e5e9a5b_content_list.json @@ -0,0 +1,1747 @@ +[ + { + "type": "text", + "text": "Img-Diff: Contrastive Data Synthesis for Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 104, + 130, + 893, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qirui Jiao $^{1}$ , Daoyuan Chen $^{2}$ , Yilun Huang $^{2}$ , Bolin Ding $^{2}$ , Yaliang Li $^{2}$ , Ying Shen $^{1,3}$", + "bbox": [ + 166, + 179, + 826, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Sun Yat-Sen University, $^{2}$ Alibaba Group, $^{3}$ FSIETP", + "bbox": [ + 292, + 215, + 705, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jiaoqr3@mail2.sysu.edu.cn, sheny76@mail.sysu.edu.cn {daoyuanchen.cdy, lielin.hyl, bolin.ding, yaliang.li}@alibaba-inc.com", + "bbox": [ + 205, + 236, + 790, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95f8a8be447b0fda9939655bd20481218816e7c7b8b3c5f0dd4b7fd24be2e195.jpg", + "image_caption": [ + "... the left image highlights a fish on the ground, while the right image highlights a red crab on the sand." + ], + "image_footnote": [], + "bbox": [ + 101, + 304, + 356, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/082576aa6da0eca498f07c2c1f67f32cbe00f7ac4ee87ee0afc122825e7d8b1b.jpg", + "image_caption": [ + "... In the left image, the player is holding a baseball glove, while in the right image, the player is holding a basketball.", + "Figure 1. Three \"object replacement\" examples within IMG-DIFF, highlighting fine-grined difference in both vision and language." + ], + "image_footnote": [], + "bbox": [ + 370, + 305, + 625, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8040c21694a3b36da7e7429dfdc17bd7198ceacf7eba16ef649337144ce3df9.jpg", + "image_caption": [ + "The left image shows a framed picture on a shelf, while the right image shows a white vase with pink flowers. ..." + ], + "image_footnote": [], + "bbox": [ + 640, + 305, + 895, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 477, + 326, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-performance Multimodal Large Language Models (MLLMs) rely heavily on data quality. This study introduces a novel data synthesis method, leveraging insights from contrastive learning and image difference captioning to enhance fine-grained image recognition in MLLMs. By analyzing object differences in detailed regions between similar images, we challenge the model to identify both matching and distinct components. Specifically, our method initially create pairs of similar images that highlight object variations. After that, we introduce a Difference Area Generator for object differences identifying, followed by a Difference Captions Generator for differences describing. The outcome is a high-quality dataset of \"object replacement\" samples, named Img-Diff, which can be expanded as needed due to its automation. We use the generated dataset to fine-tune state-of-the-art (SOTA) MLLMs such as InternVL2, yielding comprehensive improvements across numerous image difference and Visual Question Answering tasks. For instance, the trained models notably surpass the SOTA models GPT-4V and Gemini on the MMVP benchmark. Additionally, we conduct thorough evaluations to confirm the dataset's diversity, quality, and robustness, presenting several insights on the synthesis of such a contrastive dataset.", + "bbox": [ + 89, + 510, + 483, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We release our codes and dataset to encourage further research on multimodal data synthesis and MLLMs' fundamental capabilities for image understanding.", + "bbox": [ + 511, + 478, + 906, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 556, + 643, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The emergence of large language models (LLMs) has revolutionized natural language processing [48, 64], also paved the way for the development of Multimodal Large Language Models (MLLMs) that seamlessly integrate linguistic and visual understanding. Improving the performance of MLLMs hinges on two primary avenues: evolving model architectures and enhancing dataset quality [53]. The majority of state-of-the-art (SOTA) MLLMs [3, 37, 41-43] employ a two-phase training strategy, commencing with a pre-training phase involving extensive image-text pairs for modality alignment, followed by a fine-tuning phase aimed at optimizing visual question answering (VQA) capabilities with specific instruction tuning datasets.", + "bbox": [ + 509, + 582, + 906, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The quality of visual instruction tuning datasets plays a crucial role in MLLMs' overall performance in VQA tasks and diverse downstream applications. With the evolution of visual instruction tuning datasets, several recent studies have explored the integration of object detection and Optical Character Recognition (OCR) datasets, such as RefCOCO [28], Visual Genome [31], OCR-VQA [49], and TextVQA [60], effectively enhancing MLLMs' proficiency", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "3 \"FSIETP\" stands for the Guangdong Provincial Key Laboratory of Fire Science and Intelligent Emergency Technology.", + "bbox": [ + 89, + 871, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9296", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in tasks requiring detailed image perception.", + "bbox": [ + 89, + 90, + 383, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we focus on a new direction for enhancing MLLM datasets, driven by the potential of object variations in image pairs to refine models' image recognition capabilities, as demonstrated by advancements in contrastive learning and image difference captioning [5, 26, 51, 80]. Specifically, we introduce a general-purpose yet challenging dataset named IMG-DIFF, which sets itself apart from existing visual instruction tuning datasets by generating pairs of highly similar images featuring subtle object alterations. Rather than compelling MLLMs to focus solely on a single image, our dataset challenges them to analyze paired images and articulate the differences within designated regions, meanwhile taking the high-quality textual descriptions of the difference as learning signals. By doing so, we aim to empower MLLMs with enhanced capabilities of fine-grained image recognition.", + "bbox": [ + 89, + 107, + 483, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our data synthesis method, we integrate the generated dataset into the original visual instruction tuning datasets of LLaVA-1.5 [41], MGM [37], and InternVL2 [12], and conduct fine-tuning. Subsequently, we evaluate their performance on image difference benchmarks, including MMVP [63], Spot-the-Diff [26], and Image-Edit-Request [62], as well as widely recognized MLLM benchmarks. Our evaluation results reveal that after fine-tuning with our IMG-DIFF dataset, the MLLMs achieve notable enhancements in image difference benchmarks, aligning their performance with state-of-the-art (SOTA) models. For instance, they notably surpass the SOTA models GPT-4V [1] and Gemini [16] on the MMVP benchmark, achieving an improvement of up to 12. Moreover, the models exhibit comprehensive improvements across eight well-recognized MLLM benchmarks, achieving an average score improvement of up to $3.06\\%$ , underscoring the useful role our dataset plays in bolstering MLLMs' competencies in both image difference recognition and fine-grained image analysis.", + "bbox": [ + 89, + 349, + 483, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We further evaluate the diversity and quality of our dataset, ensuring it encompasses a broad array of object categories while showcasing rich variability. Through meticulous manual labeling, we affirm the high quality of our dataset. Furthermore, we conduct ablation studies to examine the effects of various filtering intensities. We also investigate an alternative methodology for constructing image difference data focusing on \"object removal\", assessing its effectiveness and presenting fruitful insights on the construction of contrastive data.", + "bbox": [ + 89, + 652, + 482, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 109, + 805, + 413, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present a novel data synthesis method and an effect-proven IMG-DIFF dataset, comprising pairs of highly similar images, with a focus on processes such as segmentation, filtering, and detailed captioning of image differences.", + "bbox": [ + 89, + 825, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conduct visual instruction tuning on LLaVA-1.5-7B, MGM-7B, and InternVL2-8B using our dataset, and rigorously assess the fine-tuned models' performance on many widely-used MLLM benchmarks and image difference benchmarks. Our dataset brings substantial performance improvements to the fine-tuned MLLMs.", + "- We provide a comprehensive evaluation of the diversity and quality of our dataset, confirming its richness and high standard. Through ablation studies, we identify good empirical filtering thresholds for such contrastive dataset.", + "- We open-source our dataset and codes at https://github.com/modelscope/data-juicer/tree/ImgDiff to facilitate ongoing research, encouraging innovative endeavors in MLLM datasets and image difference methods." + ], + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and Related Works", + "text_level": 1, + "bbox": [ + 513, + 319, + 807, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 511, + 345, + 831, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Large Language Models (MLLMs) have exhibited remarkable advancements since their introduction [1, 16, 37, 43, 44]. Research highlights two key factors that primarily influence the effectiveness of MLLMs: model architecture and dataset quality [53].", + "bbox": [ + 511, + 368, + 903, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With respect to model architecture, notable approaches [2, 3, 25, 32, 35] leverage learnable queries to extract essential information from CLIP [15, 54] image features. Alternatively, LLaVA [41-43] and MGM [37] utilize projection-based interfaces to facilitate interactions between text and image modalities. Furthermore, LLaMA-Adapter [81] and LaVIN [47] implement parameter-efficient tuning mechanisms to transfer image-related information to the LLM.", + "bbox": [ + 511, + 445, + 905, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From the perspective of datasets, there are two prevalent strategies: enhancing the quality of pre-training data and improving instruction tuning data. The former aims for better semantic alignment between images and text by introducing abundant image-text pairs, enhancing MLLMs' fundamental capabilities such as image captioning. As for the latter, research has increasingly concentrated on refining instruction tuning datasets, boosting performance across various question-answering tasks. Works like LLaVA [41-43], SPHINX [40], MGM [37], and InternVL2 [12] leverage high-quality fine-tuning datasets with extensive task diversity, allowing models to excel in tasks related to image perception, reasoning, and optical character recognition. Additionally, methods such as Shikra [10], ASM [72], and PINK [75] utilize substantial amounts of object detection data to enhance the models' localization capabilities.", + "bbox": [ + 511, + 566, + 905, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to previous works, our research introduces a fully automated data synthesis method and generates a dataset that emphasizes image differences, showing empirical effectiveness and great potential to augment MLLMs' VQA proficiency, object localization capabilities, and discernment of image distinctions.", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9297", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Image Difference Datasets", + "text_level": 1, + "bbox": [ + 89, + 90, + 328, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Datasets focused on image differences typically consist of pairs of similar images and textual descriptions that describe the variations. For instance, the Spot-the-Diff dataset [26] contains pairs of street scenes captured at different times by the same surveillance cameras. The CLEVR-Change dataset [51] delineates scene variations of geometric objects against a clean backdrop. The CUB-Bird dataset [69] focuses on the nuanced differences among various bird species found in natural habitats. The Image-Edit-Request [62] dataset features edited images, accompanied by descriptions of modifications made.", + "bbox": [ + 89, + 112, + 483, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Leveraging advancements in image editing technologies, some studies have employed generative models and editing techniques to create datasets centered on image differences. A prime example is InstructPix2Pix [6], which utilizes the image editing technique, named Prompt-to-Prompt [20], to direct Stable-Diffusion-1.5 [56] in generating pairs of similar images, while employing GPT-3 [48] to craft the edited text as reference captions for image differences.", + "bbox": [ + 89, + 279, + 483, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach, referring to InstructPix2Pix, employs the Prompt-to-Prompt technique alongside an advanced generative model Stable-Diffusion-XL [52], which produces more realistic images, to generate pairs of similar images. Unlike InstructPix2Pix, we incorporate multiple filtering stages to ensure data quality, with a particular emphasis on producing difference captions that focus on specific regions rather than the entire image, which ensures greater accuracy.", + "bbox": [ + 89, + 400, + 483, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Models for Image Difference Captioning", + "text_level": 1, + "bbox": [ + 89, + 531, + 437, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Difference Captioning (IDC) represents a specialized domain within image captioning characterized by its focus on subtle variations between images. As for the pioneering work in IDC, Spot-the-Diff [26] presents potential change clusters and employs an LSTM [21] decoder to generate difference captions. DUDA [51] explores image differences at the semantic level, using a ResNet [19] and an LSTM to compute dynamic attention weights and produce captions. VARD [65] introduces a viewpoint-adaptive representation disentanglement network based on LSTM for differentiating between real and pseudo changes. Meanwhile, NCT [66] employs a transformer [67] to integrate neighboring features, and CLIP4IDC [18] uses BERT-like training methodologies, adapting a CLIP model for IDC tasks. With the emergence of MLLMs, VIXEN [5] marks the inaugural use of these models for IDC tasks, mapping the features of differential images to text space and employing an LLM to generate image difference captions.", + "bbox": [ + 89, + 551, + 483, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our data synthesis method is specifically designed for MLLMs. We build our dataset following the instruction-following format adopted by mainstream MLLMs such as LLaVA-1.5, highlighting a novel research direction aimed at enhancing MLLMs from a data-centric perspective.", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5f866f61d2fdbb9e67e73bd315297c94cfc646f0b7df5665b9409c11b81e43ed.jpg", + "image_caption": [ + "Figure 2. The generation process for \"object replacement\" data." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 686, + 262 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f22d289ae7b9c2a60939dbe1f1c322036d19fd9f5927a6817f0c6b2ddea4f0fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 88, + 810, + 261 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6fd31c0668d5179748c746ee6fb306bac90a5de3b8a46dca09a7e4865f9ff4e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 815, + 92, + 903, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 314, + 648, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 513, + 340, + 624, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, contrastive learning have significantly enhanced the image-text understanding of vision-language models [54, 80]. These methods typically involve constructing batches of images and texts, requiring the model to distinguish between matching and non-matching image-text pairs, which improves its ability to differentiate between semantically similar and dissimilar pairs.", + "bbox": [ + 511, + 364, + 905, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method leverages the principles of contrastive learning to generate MLLM image-text data. Specifically, it focuses on replacing objects within image pairs, encouraging MLLMs to identify similarities and differences in specific regions. This method aims to enhance models' capacity to recognize fine-grained differences in images, guided by textual descriptions that highlight detailed distinctions.", + "bbox": [ + 511, + 473, + 905, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, the process of generating \"object replacement\" data can be divided into three key parts. The first part involves creating similar images and forming image pairs, where the only difference between the images in pairs is the objects replaced (Section 3.2). The second part, named the \"Difference Area Generator\", extracts bounding box regions that contain object differences between the images in pairs (Section 3.3). The third part, termed as the \"Difference Captions Generator\" (Section 3.4), utilizes an MLLM to generate descriptive text for the areas with object differences and creates question-answer pairs.", + "bbox": [ + 511, + 580, + 905, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This process incorporates multiple filtering operations, with specific thresholds outlined in Section 18.4, which we determine through experimental comparisons in Section 17. Additional experimental details, including model selection and the time consumption, are also presented in Section 18. For readability, we present data examples in Section 20.", + "bbox": [ + 511, + 748, + 905, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance reproducibility and reusability, our proposed components and end-to-end construction workflow are implemented as data processing operators and configurable data recipes within Data-Juicer [8, 9].", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9298", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d7b9448d1919c7579c5ddd08be646025e5255f3d6ba1a7c249b8f7536c95769b.jpg", + "image_caption": [ + "Figure 3. An overview of the Difference Area Generator and its three internal components: Image Similarity Filter, Image-text Matching Filter, and Difference Detector." + ], + "image_footnote": [], + "bbox": [ + 191, + 85, + 810, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Image Pairs Generation", + "text_level": 1, + "bbox": [ + 89, + 375, + 312, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first step of our data synthesis method is to generate pairs of similar images as candidates. The process is shown in Figure 2. We employ a generative model called Stable-Diffusion-XL [52] and an image editing technique called Prompt-to-Prompt [20] to generate image pairs that highlight object replacement.", + "bbox": [ + 89, + 397, + 483, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start by obtaining image captions from caption databases, which contain descriptions biased towards real photos. Then, we use an LLM to perform object replacement in the captions. The prompt used is \"Here is a sentence: 'INPUT'. Please only replace one of the objects in this sentence with another object.\" Here, INPUT refers to the original caption. Next, based on the caption pairs, we use the Stable-Diffusion-XL and the Prompt-to-Prompt to generate image pairs with only few objects replaced.", + "bbox": [ + 89, + 488, + 483, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Difference Area Generator", + "text_level": 1, + "bbox": [ + 89, + 633, + 333, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 655, + 202, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Difference Area Generator is used to identify the locations of object differences between the two images in pairs. Although object detection models are capable of detecting objects in images, the range of object categories is quite limited [27, 78]. Therefore, to increase the number of detectable object categories and enhance data diversity, we develop the Difference Area Generator based on segmentation and image similarity comparisons.", + "bbox": [ + 89, + 674, + 482, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 3, we first use an Image Similarity Filter to obtain image pairs with high similarity but not identical. Next, we use the FastSAM [82] to perform image segmentation on each image. After that, we crop the images based on the bounding box information obtained from segmentation and use an Image-text Matching Filter to filter the cropped sub-images for the presence of valid objects.", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we use a Difference Detector to determine whether there are indeed differences between the bounding box regions of the two images and perform IoU (Intersection over Union) filtering to remove overlapping bounding boxes, ultimately obtaining valid bounding box information.", + "bbox": [ + 511, + 375, + 906, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Image Similarity Filter", + "text_level": 1, + "bbox": [ + 511, + 460, + 720, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Image Similarity Filter aims to filter image pairs based on the degree of similarity. The module first uses CLIP [54] to extract image features from each image in pairs and then calculates the cosine similarity score. If their score falls within the pre-set threshold, the image pair will be considered valid. Specifically, we use the module twice in the Difference Area Generator. Before the segmentation, we use the module to ensure that the images in pairs are highly similar but not the same. In the Difference Detector stage, after cropping, we use the module to filter the sub-image pairs and keep only the differing ones.", + "bbox": [ + 511, + 479, + 906, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.3. Image-text Matching Filter", + "text_level": 1, + "bbox": [ + 511, + 656, + 750, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Image-text Matching Filter determines whether an image contains valid objects (i.e. the replaced or replacing objects). This module first uses BLIP [34] to extract image features, which are then compared with text features of object names. When the image-text matching score falls within the pre-set threshold, we consider the image to contain valid objects. In the mid-stage of the Difference Area Generator, after cropping, we use the module to determine whether these sub-images contain valid objects and get valid bounding boxes.", + "bbox": [ + 511, + 675, + 908, + 827 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.4. Difference Detector", + "text_level": 1, + "bbox": [ + 511, + 835, + 694, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Difference Detector determines whether there are differences between the two regions in the image pair of the same bounding box. Based on a given bounding box, we", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9299", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ef3f77d87c470c1d2fc964538b69a793bc79ef4895898ec4dcae01b1a10f849d.jpg", + "image_caption": [ + "Figure 4. An overview of the Difference Captions Generator and its two stages." + ], + "image_footnote": [], + "bbox": [ + 130, + 87, + 867, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "first crop two sub-images from both image A and image B. The two sub-images are then filtered through the Image Similarity Filter and the bounding box is considered effective only if the difference is significant enough. After processing all bounding boxes, we use the IoU method to filter out overlapping bounding boxes. Only the bounding boxes with a higher degree of difference are retained.", + "bbox": [ + 89, + 297, + 483, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Difference Captions Generator", + "text_level": 1, + "bbox": [ + 89, + 411, + 364, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 433, + 202, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the valid bounding box regions, we use the Difference Captions Generator to generate descriptions for the differences inside these areas (with each round of the process focusing on only one bounding box in one image pair). Evidently, an image pair can contain multiple object differences but a single difference caption cannot fully capture all of them. Therefore, we highlight specific regions with red boxes and provide targeted difference captions to ensure greater accuracy.", + "bbox": [ + 89, + 452, + 483, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The module consists of two stages: the first stage generates captions for the content in the bounding box regions and then filters the bounding boxes using an Image-text Matching Filter and a Captions Similarity Filter. The second stage uses the content captions and the images highlighted with red boxes to generate difference captions. The overview is shown in Figure 4.", + "bbox": [ + 89, + 588, + 483, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2. Stage1: Object Labeling & Filtering", + "text_level": 1, + "bbox": [ + 89, + 700, + 390, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Stage 1, for each image pair, we first select N bounding box regions with the lowest similarity between images (N is set to 5 in this project) as candidate regions. Then, for each bounding box, we use the MLLM LLaVA-NEXT [43] to describe its corresponding regions and then apply two filtering processes: the first filter is an Image-text Matching Filter, which checks whether the content of the regions matches the captions; the second filter is an Captions Similarity Filter, which assesses whether there are differences between the two captions. Once the filtering is complete, we obtain valid bounding boxes and captions for subsequent difference captioning.", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.3. Captions Similarity Filter", + "text_level": 1, + "bbox": [ + 511, + 297, + 738, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Captions Similarity Filter determines whether the two captions of the same bounding box coordinates are different. We use CLIP to obtain text features and calculate the cosine similarity between them. When the score is low enough, we consider the two captions to be different.", + "bbox": [ + 511, + 316, + 905, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.4. Stage2: Difference Captions Generating", + "text_level": 1, + "bbox": [ + 511, + 398, + 839, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Stage 2, for each valid bounding box of each image pair, we first draw two red boxes into the images based on the bounding box coordinates, highlighting the differences for easier localization. Then, we provide the MLLM LLaVA-NEXT with the captions of the bounding box content and instruct it to describe the difference based on these captions and the highlighted images. Finally, we can obtain the difference caption for the highlighted area.", + "bbox": [ + 511, + 417, + 905, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Data Statistics", + "text_level": 1, + "bbox": [ + 511, + 547, + 661, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using captions from MSCOCO [39], we generate 118K pairs of similar images. We then employ the Image Similarity Filter to get 38,533 highly similar but not identical image pairs. Next, we use the Difference Area Generator to filter and produce 117,779 pieces of valid bounding box information (with a maximum of 5 valid bounding boxes per image pair). Finally, we employ the Difference Captions Generator to filter and generate 12,688 high-quality \"object replacement\" instances. Our evaluations on the main page are based on this dataset.", + "bbox": [ + 511, + 568, + 905, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, we also use captions from the LLaVA pretraining dataset to generate 34,538 \"object replacement\" samples. In Section 10 of the appendix, we evaluate this dataset and discuss the relationship between data quantity, data quality, and model performance improvement, emphasizing the importance of prioritizing quality over quantity.", + "bbox": [ + 511, + 720, + 905, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluation and Main Results", + "text_level": 1, + "bbox": [ + 511, + 823, + 782, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation Settings", + "text_level": 1, + "bbox": [ + 511, + 848, + 700, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our IMG-DIFF dataset, we use it to fine-tune SOTA MLLMs, including LLaVA-1.5-", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9300", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7B, MGM-7B, and InternVL2-8B on the main page, as well as InternVL2-1B and LLaVA-1.5-13B in the supplementary materials. We evaluate these models on extensive benchmarks commonly used for image difference and MLLMs. Specifically, the image difference benchmarks include MMVP [63], Spot-the-Diff [26], and Image-Edit-Request [62]. Besides, the details and results of the MLLM benchmarks are shown in Section 4.5.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During our fine-tuning process, we first mix our data with MLLMs' original visual instruction tuning data respectively. For LLaVA-1.5 and MGM, we conduct the fine-tuning anew. For InternVL2, we follow the guidelines from its official repository and perform a second fine-tuning. (To ensure a fair comparison, we perform a second fine-tuning on InternVL2 using its original fine-tuning dataset, termed as InternVL2-FT, and use it as a baseline.) Regarding the Spot-the-Diff and Image-Edit-Request benchmarks, as they contain data of training splits, we further fine-tune the fine-tuned MLLMs for an additional 2 epochs using only these benchmarks' training data.", + "bbox": [ + 89, + 212, + 482, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the tables, \"RP\" represents \"object replacement\" data.", + "bbox": [ + 109, + 393, + 480, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on the MMVP Benchmark", + "text_level": 1, + "bbox": [ + 89, + 417, + 390, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The MMVP benchmark is designed to systematically assess the visual capabilities of MLLMs. Its data processing method is highly related to image difference: it first collects CLIP-blind pairs, which have similar CLIP features but differ in image content. Then, the differences between the images are manually described and question-answer pairs are created. Hence, the questions in MMVP is highly relevant to our dataset, as both place significant emphasis on the differences between similar images.", + "bbox": [ + 89, + 439, + 482, + 575 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/de86155551b0c8394b3cbcbdb98208bafb2bec1b8b785a0bd985e99c09c58452.jpg", + "image_caption": [ + "Figure 5. Performance comparison on the MMVP benchmark." + ], + "image_footnote": [], + "bbox": [ + 89, + 585, + 482, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Figure 5, fine-tuning MLLMs with our \"object replacement\" data significantly improves their performance on the MMVP benchmark. After fine-tuning, the score of LLaVA-1.5-7B exceeds that of LLaVA-1.5-13B. Furthermore, the fine-tuned MGM-7B shows a significant improvement in score compared to the original MGM-7B, even surpassing the scores of the SOTA models GPT-4V and Gemini by up to 12. Additionally, the performance of", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "InternVL2-8B is also enhanced. These results suggest that our dataset enhances MLLMs' abilities to distinguish images with similar CLIP features but different content.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results on the Spot-the-Diff Benchmark", + "text_level": 1, + "bbox": [ + 511, + 148, + 856, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The dataset of Spot-the-Diff comprises pairs of street view images that display subtle object differences. These images are obtained by capturing scenes from fixed surveillance cameras at different time.", + "bbox": [ + 511, + 172, + 903, + 232 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/59c73587752853584877740710aa209e40095499e9cc07172883e28a6e69fd73.jpg", + "table_caption": [ + "Table 1. Performance comparison on Spot-the-Diff." + ], + "table_footnote": [], + "table_body": "
ModelBLEUMETEORCIDEr-DROUGE-L
VAM[59]10.112.438.131.3
IFDC[23]8.711.73730.2
DUDA+Aux[22]8.112.534.529.9
VACC[29]9.712.641.532.1
LLaVA-1.5-7B8.512.038.330.1
LLaVA-1.5-7B + RP9.713.043.230.8
MGM-7B9.912.046.331.5
MGM-7B + RP10.813.153.533.0
InternVL2-8B-FT6.611.726.527.3
InternVL2-8B + RP8.412.832.228.5
", + "bbox": [ + 537, + 270, + 883, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous works, our fine-tuned MLLMs are evaluated on BLEU [50], METEOR [4], CIDEr-D [68] and ROUGE-L [38]. As shown in Table 1, after fine-tuning with our \"object replacement\" data, both LLaVA-1.5-7B, MGM-7B, and InternVL2-8B achieve significant performance gains on Spot-the-Diff. Compared to score increases seen in prior models, our dataset yields even greater enhancements than those from previous iterations of image difference models. These positive results indicate that our dataset can enhance the ability of MLLMs to detect subtle differences between similar images.", + "bbox": [ + 511, + 440, + 905, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Results on Image-Editing-Request", + "text_level": 1, + "bbox": [ + 511, + 619, + 812, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Image-Editing-Request benchmark focuses on image editing. Each instance in its dataset consists of an image pair (i.e. a source image and an edited image) and an editing instruction that describes the transformation. During the evaluation, our models are required to generate transformation descriptions for these image pairs, and we then calculate the BLEU, METEOR, CIDEr-D, and ROUGE-L scores using the models' responses and the reference answers.", + "bbox": [ + 511, + 642, + 905, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 2, LLaVA-1.5-7B, MGM-7B, and InternVL2-8B originally show SOTA performance on the Image-Edit-Request benchmark. Upon incorporating our \"object replacement\" data for better fine-tuning, the performance of all three models improves even more, achieving new SOTA scores. This suggests that our dataset enhances MLLMs' abilities to recognize similarities and dissimilarities in image pairs, as well as enables them to describe differences more accurately.", + "bbox": [ + 511, + 763, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9301", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/be8a01249aa26139a364196ffe34a2abb585fab8cf1983917fa99db96c3a4a3a.jpg", + "table_caption": [ + "Table 2. Performance comparison on Image-Edit-Request." + ], + "table_footnote": [], + "table_body": "
ModelBLEUMETEORCIDEr-DROUGE-L
VARD[65]1014.835.739
CLIP4IDC[18]8.214.632.240.4
NCT[66]8.11534.238.8
BiDiff[61]6.914.627.738.5
VIXEN[5]8.615.438.142.5
LLaVA-1.5-7B15.117.860.645.2
LLaVA-1.5-7B + RP16.219.560.946.7
MGM-7B16.517.766.844.8
MGM-7B + RP16.618.268.145.7
InternVL2-8B-FT12.414.151.538.9
InternVL2-8B + RP12.514.256.039.4
", + "bbox": [ + 109, + 114, + 464, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Results on MLLM Benchmarks", + "text_level": 1, + "bbox": [ + 89, + 306, + 370, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Aside from the evaluations related to image difference discrimination, we also assess the performance of our \"object replacement\" data in enhancing the comprehensive abilities of MLLMs. We test the fine-tuned MLLMs using commonly used MLLM benchmarks, including VQAv2 [17] and GQA [24] for assessing the comprehensive VQA capabilities; MMBench [45], MM-Vet [76], ScienceQA [46], and SEED-Bench [33] for testing perceptual and reasoning abilities; and POPE [36] for evaluating fine-grained object localization abilities. Table 3 presents the results on these MLLM benchmarks, with the $\\triangle$ metric indicating the percentage improvement averaged across them.", + "bbox": [ + 89, + 329, + 483, + 511 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c92d46c4f0ac9ffa25dc416fa95a8ed20cd3dcae5038b6723ddb4af17a99bca2.jpg", + "table_caption": [ + "Table 3. Performance comparison on 8 MLLM benchmarks." + ], + "table_footnote": [], + "table_body": "
ModelVQAv2GQAPOPEMMB\\( {\\mathrm{{MMB}}}^{CN} \\)
LLaVA-1.5-7B78.56285.964.358.3
LLaVA-1.5-7B + RP79.3↑62.8↑86.4↑66.1↑59.8↑
MGM-7B80.462.68669.358.9
MGM-7B + RP80.7↑62.7↑86.2↑68.7↓59.6↑
InternVL2-8B-FT81.862.687.782.581.5
InternVL2-8B + RP81.862.688.0↑82.7↑81.4↓
", + "bbox": [ + 101, + 551, + 470, + 652 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6f74d7b9552fd567574dfbdb998b5f82bedbf77d49a59eb7bde0c7b1c23ca972.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMM-Vet\\( \\mathbf{SQA}^I \\)SEED\\( \\bigtriangleup \\)
LLaVA-1.5-7B30.566.858.6-
LLaVA-1.5-7B + RP33.2↑68.2↑61.7↑+3.06%
MGM-7B40.870.663.5-
MGM-7B + RP44.1↑71.7↑63.2↓+1.28%
InternVL2-8B-FT49.296.569.5-
InternVL2-8B + RP52.6↑96.6↑69.9↑+1.01%
", + "bbox": [ + 101, + 660, + 470, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on Table 3, after fine-tuning with our dataset, the performance of LLaVA-1.5-7B shows a comprehensive improvement, with an average increase of $3.06\\%$ across all benchmarks. For MGM-7B and InternVL2-8B, the improvements brought by our dataset are not as pronounced as those observed with LLaVA-1.5-7B, as their training datasets already encompass a large volume of high-quality data, but they still achieve an average increase of $1.28\\%$", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and $1.01\\%$ . These improvements indicate that fine-tuning MLLMs with our dataset not only enhances their ability to discern differences but also improves their overall visual capabilities, thereby making them better address VQA tasks.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Evaluation of Data Quality and Diversity", + "text_level": 1, + "bbox": [ + 511, + 164, + 879, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Data Quality", + "text_level": 1, + "bbox": [ + 511, + 189, + 648, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the quality of our Img-Diff dataset, we randomly select 1,000 instances of \"object replacement\" data and employ multiple professional dataset annotators to evaluate the samples based on three metrics. The final scores are determined through a voting process. Specifically, the first metric is \"Bounding Box Difference\", which evaluates whether there are differences between the two highlighted regions in pairs. If the objects are different, we score it as \"high\"; if the objects are the same but their features (such as color, shape, etc.) are different, we score it as \"medium\"; if the objects are the same and their features are similar, we score it as \"low\". The second metric is \"Content Caption Accuracy\", which evaluates whether the captions generated by Stage 1 of the Difference Captions Generator accurately describe the two highlighted regions. If the captions are correct, we score it as \"high\"; if the captions identify the objects but incorrectly describe their features, we score it as \"medium\"; if the captions incorrectly identify the objects, we score it as \"low\". The third metric is \"Difference Caption Accuracy\", which evaluates whether the final difference captions accurately describe the object differences between highlighted regions of the image pairs. If the description is accurate, we score it as \"high\"; if it recognizes the objects but the feature description is incorrect, we score it as \"medium\"; if the object recognition is incorrect, we score it as \"low\". The results are shown in Figure 6.", + "bbox": [ + 511, + 210, + 906, + 604 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86675cd21f5a495339652134afff76ea4bc8da4341a688d0288cb73423d3c9b8.jpg", + "image_caption": [ + "Figure 6. Quality assessment of the \"object replacement\" data." + ], + "image_footnote": [], + "bbox": [ + 514, + 614, + 903, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on Figure 6, our dataset demonstrates a high level of quality. For the \"Bounding Box Difference\" metric, only $4.5\\%$ of the instances are classified as \"low\", and nearly $80\\%$ of instances exhibit valid object differences between their two highlighted regions. In terms of \"Content Caption Accuracy\", $80.1\\%$ of highlighted region pairs are described accurately, indicating that using an MLLM for labeling is effective and that our filtering strategy is also functioning well. For the \"Difference Caption Accuracy\" metric, over $70\\%$ of the difference descriptions are completely accurate, with $21.8\\%$ of the samples having errors solely in feature labeling while still maintaining correct descriptions of ob-", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9302", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ject differences, which underscores the effectiveness of our difference caption generation strategy.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Data Diversity", + "text_level": 1, + "bbox": [ + 89, + 132, + 238, + 148 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In designing the generation process, we have made efforts to enhance the diversity of our dataset, which can be divided into two aspects: (a) the inherent diversity of the caption databases; and (b) the introduction of new object names through our object replacement strategy. As for the former, we are using the captions from MS COCO and the LLaVA pre-training dataset, which cover a majority of object categories. We can easily further enhance the data diversity by employing caption databases with greater variability. As for the latter, we employ two methods, including increasing the temperature of the LLM used for noun replacement, as well as randomly replacing object names with nouns from a noun lexicon (shown in Appendix, Section 11). We can further improve data diversity by expanding the noun lexicon.", + "bbox": [ + 89, + 155, + 482, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By analyzing the valid object names included in the captions that describe highlighted regions, we assess the diversity of our \"object replacement\" data. Specifically, we count the total number of object categories covered, and the total number of unique \"object replacement pairs\". Through our statistical analysis, we find that our dataset covers 1,203 object categories, which encompasses most real-life objects. An \"object replacement pair\" refers to the combination of the replaced and replacing object names. Our dataset includes 3,680 unique \"object replacement pairs\".", + "bbox": [ + 89, + 367, + 482, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the coverage of common object categories in our dataset, we analyze the occurrences of the object names from the Object365 [58] dataset in our own dataset. The results show that the object names from the Object365 dataset appears 13,164 times in total, which accounts for approximately $52.06\\%$ of the total occurrences of object names.", + "bbox": [ + 89, + 518, + 482, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These statistics show that our data covers a substantial number of common object names, ensuring a high frequency of common objects. Additionally, less common object names make up nearly half of our dataset, indicating that our dataset is both diverse and comprehensive.", + "bbox": [ + 89, + 609, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. More Empirical Supports and Details", + "text_level": 1, + "bbox": [ + 89, + 700, + 429, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We list additional experimental details and extensions of our data synthesis method in the supplementary material, and conduct further experiments to validate the effectiveness of our IMG-DIFF dataset, along with its limitations.", + "bbox": [ + 89, + 727, + 482, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Further Experiments and Analysis. In Section 9, we compare our IMG-DIFF dataset with existing image difference datasets on characteristics and performance, validating its superiority. In Section 10, we discuss the relationship between the data volume and model performance improvement. In Section 11, we explore the use of lexicons for", + "bbox": [ + 89, + 809, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "object replacement and validate its effectiveness in enhancing diversity. In Section 12, we evaluate our dataset using the Contrastive Chain-of-Thought method. In Section 13, we validate the effectiveness of our dataset on MLLMs of different scales. In Section 14, we evaluate top-performing MLLMs on difference detection. In Section 15, we assess the impact of unnatural images. In Section 16, we show our data's efficacy in enhancing spatial reasoning capabilities.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. In Section 17, we investigate the effect of different filtering intensities on the performance of the generated datasets. In Section 18, we present additional experimental details, including image pair preprocessing, model training procedures, model selection, filtering thresholds, and time consumption.", + "bbox": [ + 511, + 231, + 903, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The \"Object Removal\" Exploration. In Section 19, we generate an extended dataset that focuses on object removal, which prompts MLLMs to analyze which image contains a specific object. The new data brings further improvement to the fine-tuned MLLM.", + "bbox": [ + 511, + 342, + 906, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 431, + 633, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we draw inspiration from recent advances in contrastive learning and image difference captioning to propose a novel method of contrastive data synthesis. This method enables the creation of a high-quality dataset called \"IMG-DIFF\", which highlights object differences focusing on fine-grained regions in images. Specifically, we first generate pairs of similar images where the main focus is on object replacement. Then, we use the proposed Difference Area Generator and Difference Captions Generator to generate difference captions and form question-answer pairs. In contrast to previous image difference datasets, our dataset focuses exclusively on specific regions inside images. This characteristic circumvents the issue where a single description cannot fully capture the differences of the whole images in pairs, enhancing the accuracy. Afterwards, we fine-tune many MLLMs using the generated dataset, yielding high-performance scores on par with SOTA models in image difference tasks and comprehensive performance improvements in eight widely recognized MLLM benchmarks. These results confirm the effectiveness of our dataset to improve the ability of MLLMs in recognizing detailed differences between images.", + "bbox": [ + 511, + 457, + 906, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In a nutshell, we provide a series of insights about the construction of high-quality image difference datasets, showing great potential to effectively and efficiently enhance MLLMs via contrastive data-centric approaches. With this work, we hope it can catalyze further investigation into the realm of image difference datasets and the fine-grained image recognition capabilities of MLLMs.", + "bbox": [ + 511, + 789, + 908, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9303", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2", + "[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6", + "[5] Alexander Black, Jing Shi, Yifei Fan, Tu Bui, and John Colomosse. Vixen: Visual text comparison network for image difference captioning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 846-854, 2024. 2, 3, 7", + "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3, 1, 2", + "[7] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, Shuaibin Li, Wei Li, Yinling Li, Hongwei Liu, Jiangning Liu, Jiawei Hong, Kaiwen Liu, Kuikun Liu, Xiaoran Liu, Chengqi Lv, Haijun Lv, Kai Lv, Li Ma, Runyuan Ma, Zerun Ma, Wenchang Ning, Linke Ouyang, Jiantao Qiu, Yuan Qu, Fukai Shang, Yunfan Shao, Demin Song, Zifan Song, Zhihao Sui, Peng Sun, Yu Sun, Huanze Tang, Bin Wang, Guoteng Wang, Jiaqi Wang, Jiayu Wang, Rui Wang, Yudong Wang, Ziyi Wang, Xingjian Wei, Qizhen Weng, Fan Wu, Yingtong Xiong, Chao Xu, Ruiliang Xu, Hang Yan, Yirong Yan, Xiaogui Yang, Haochen Ye, Huaiyuan Ying, Jia Yu, Jing Yu, Yuhang Zang, Chuyu Zhang, Li Zhang, Pan Zhang, Peng Zhang, Ruijie Zhang, Shuo Zhang, Songyang Zhang, Wenjian Zhang, Wenwei Zhang, Xingcheng Zhang, Xinyue Zhang, Hui Zhao, Qian Zhao, Xiaomeng Zhao, Fengzhe Zhou, Zaida Zhou, Jingming Zhuo, Yicheng Zou, Xipeng Qiu, Yu Qiao and Dahua Lin. Internl m2 technical report, 2024. 6", + "[8] Daoyuan Chen, Yilun Huang, Zhijian Ma, Hesen Chen, Xuchen Pan, Ce Ge, Dawei Gao, Yuexiang Xie, Zhaoyang" + ], + "bbox": [ + 99, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Liu, Jinyang Gao, et al. Data-juicer: A one-stop data processing system for large language models. In Companion of the 2024 International Conference on Management of Data, pages 120-134, 2024. 3", + "[9] Daoyuan Chen, Haibin Wang, Yilun Huang, Ce Ge, Yaliang Li, Bolin Ding, and Jingren Zhou. Data-juicer sandbox: A feedback-driven suite for multimodal data-model codevelopment. arXiv preprint arXiv:2407.11784, 2024. 3", + "[10] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2", + "[11] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Qinglan Huang, Canyu Chen, Qinghao Ye, Zhihong Zhu, Yuqing Zhang, Jiawei Zhou, Zhuokai Zhao, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Mj-bench: Is your multimodal reward model really a good judge for text-to-image generation?, 2024. 1, 2", + "[12] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2, 3", + "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023.8", + "[14] W Dai, J Li, D Li, AMH Tiong, J Zhao, W Wang, B Li, P Fung, and S Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 3", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[16] Google. Gemini. 2023. 2", + "[17] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 7", + "[18] Zixin Guo, Tzu-Jui Julius Wang, and Jorma Laaksonen. Clip4idc: Clip for image difference captioning. arXiv preprint arXiv:2206.00629, 2022. 3, 7", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[20] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3, 4" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9304", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 3", + "[22] Mehrdad Hosseinzadeh and Yang Wang. Image change captioning by learning from an auxiliary task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2725-2734, 2021. 6", + "[23] Qingbao Huang, Yu Liang, Jielong Wei, Yi Cai, Hanyu Liang, Ho-fung Leung, and Qing Li. Image difference captioning with instance-level fine-grained feature representation. IEEE transactions on multimedia, 24:2004-2017, 2021. 6", + "[24] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 7", + "[25] Laurençon Hugo, van Strien Daniel, Bekman Stas, Tronchon Leo, Saulnier Lucile, Wang Thomas, Karamcheti Siddharth, Singh Amanpreet, Pistilli Giada, Jernite Yacine, and Sanh Victor. Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics, 2023.2", + "[26] Harsh Jhamtani and Taylor Berg-Kirkpatrick. Learning to describe differences between pairs of similar images. arXiv preprint arXiv:1808.10584, 2018. 2, 3, 6, 1", + "[27] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. *Ultralytics YOLO*, 2023. 4", + "[28] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 1", + "[29] Hoeseong Kim, Jongseok Kim, Hyungseok Lee, Hyunsung Park, and Gunhee Kim. Agnostic change captioning with cycle consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2095-2104, 2021. 6", + "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 8", + "[31] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1", + "[32] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 2, 3", + "[33] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 7", + "[34] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified" + ], + "bbox": [ + 89, + 92, + 485, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4", + "[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2", + "[36] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 7", + "[37] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv:2403.18814, 2023. 1, 2, 3", + "[38] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6", + "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[40] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 2", + "[41] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2", + "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023.", + "[43] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,2, 5,3", + "[44] Mengsha Liu, Daoyuan Chen, Yaliang Li, Guian Fang, and Ying Shen. ChartThinker: A contextual chain-of-thought approach to optimized chart summarization. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 3057-3074, 2024. 2", + "[45] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 7", + "[46] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 7" + ], + "bbox": [ + 514, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9305", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. arXiv preprint arXiv:2305.15023, 2023. 2", + "[48] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020. 1, 3", + "[49] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 1", + "[50] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6", + "[51] Dong Huk Park, Trevor Darrell, and Anna Rohrbach. Robust change captioning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4624-4633, 2019. 2, 3", + "[52] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 4", + "[53] Zhen Qin, Daoyuan Chen, Wenhao Zhang, Liuyi Yao, Yilun Huang, Bolin Ding, Yaliang Li, and Shuiguang Deng. The synergy between data and multi-modal large language models: A survey from co-development perspective. arXiv preprint arXiv:2407.08583, 2024. 1, 2", + "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4", + "[55] Kanchana Ranasinghe, Satya Narayan Shukla, Omid Poursaeed, Michael S Ryoo, and Tsung-Yu Lin. Learning to localize objects improves spatial reasoning in visual-llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12977–12987, 2024. 6", + "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3", + "[57] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 9, 10", + "[58] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 8", + "[59] Xiangxi Shi, Xu Yang, Jiuming Gu, Shafiq Joty, and Jianfei Cai. Finding it at another side: A viewpoint-adapted match" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing encoder for change captioning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 574-590. Springer, 2020. 6", + "[60] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 1", + "[61] Yaoqi Sun, Liang Li, Tingting Yao, Tongyv Lu, Bolun Zheng, Chenggang Yan, Hua Zhang, Yongjun Bao, Guiguang Ding, and Gregory Slabaugh. Bidirectional difference locating and semantic consistency reasoning for change captioning. International Journal of Intelligent Systems, 37 (5):2969-2987, 2022. 7", + "[62] Hao Tan, Franck Dernoncourt, Zhe Lin, Trung Bui, and Mohit Bansal. Expressing visual relationships via language. arXiv preprint arXiv:1906.07689, 2019. 2, 3, 6, 1", + "[63] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2, 6", + "[64] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[65] Yunbin Tu, Liang Li, Li Su, Junping Du, Ke Lu, and Qingming Huang. Adaptive representation disentanglement network for change captioning. IEEE Transactions on Image Processing, 32:2620-2635, 2023. 3, 7", + "[66] Yunbin Tu, Liang Li, Li Su, Ke Lu, and Qingming Huang. Neighborhood contrastive transformer for change captioning. IEEE Transactions on Multimedia, 25:9518-9529, 2023. 3, 7", + "[67] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[68] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6", + "[69] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 3, 1, 2", + "[70] Jiayu Wang, Yifei Ming, Zhenmei Shi, Vibhav Vineet, Xin Wang, Yixuan Li, and Neel Joshi. Is a picture worth a thousand words? delving into spatial reasoning for vision language models. In The Thirty-Eighth Annual Conference on Neural Information Processing Systems, 2024. 7", + "[71] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9306", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6", + "[72] Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, et al. The all-seeing project: Towards panoptic visual recognition and understanding of the open world. In The Twelfth International Conference on Learning Representations, 2023. 2", + "[73] Weiyun Wang, Zhe Chen, Wenhai Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Jinguo Zhu, Xizhou Zhu, Lewei Lu, Yu Qiao, and Jifeng Dai. Enhancing the reasoning ability of multimodal large language models via mixed preference optimization. arXiv preprint arXiv:2411.10442, 2024.6", + "[74] Zhe Xu, Daoyuan Chen, Zhenqing Ling, Yaliang Li, and Ying Shen. Mindgym: Enhancing vision-language models via synthetic self-challenging questions, 2025. 3", + "[75] Shiyu Xuan, Qingpei Guo, Ming Yang, and Shiliang Zhang. Pink: Unveiling the power of referential comprehension for multi-modal llms. arXiv preprint arXiv:2310.00582, 2023. 2", + "[76] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 7", + "[77] Daoan Zhang, Junming Yang, Hanjia Lyu, Zijian Jin, Yuan Yao, Mingkai Chen, and Jiebo Luo. Cocot: Contrastive chain-of-thought prompting for large multimodal models with multiple image inputs. arXiv preprint arXiv:2401.02582, 2024. 5", + "[78] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 4", + "[79] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 1, 2", + "[80] Qi Zhang, Yifei Wang, and Yisen Wang. On the generalization of multi-modal contrastive learning. In International Conference on Machine Learning, pages 41677-41693. PMLR, 2023. 2, 3", + "[81] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 2", + "[82] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 4", + "[83] Ting Zhou, Daoyuan Chen, Qirui Jiao, Bolin Ding, Yaliang Li, and Ying Shen. Humanvbench: Exploring human-centric video understanding capabilities of mllms with synthetic benchmark data, 2025. 3" + ], + "bbox": [ + 91, + 92, + 482, + 853 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "9307", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/e6b1bb30-e9d5-4708-960e-31ff7e5e9a5b_model.json b/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/e6b1bb30-e9d5-4708-960e-31ff7e5e9a5b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..674804ef577623b83703acff8cac1b0422323b5e --- /dev/null +++ b/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/e6b1bb30-e9d5-4708-960e-31ff7e5e9a5b_model.json @@ -0,0 +1,2699 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.131, + 0.895, + 0.155 + ], + "angle": 0, + "content": "Img-Diff: Contrastive Data Synthesis for Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.18, + 0.828, + 0.201 + ], + "angle": 0, + "content": "Qirui Jiao\\(^{1}\\), Daoyuan Chen\\(^{2}\\), Yilun Huang\\(^{2}\\), Bolin Ding\\(^{2}\\), Yaliang Li\\(^{2}\\), Ying Shen\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.216, + 0.706, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Sun Yat-Sen University, \\(^{2}\\)Alibaba Group, \\(^{3}\\)FSIETP" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.237, + 0.792, + 0.27 + ], + "angle": 0, + "content": "jiaoqr3@mail2.sysu.edu.cn, sheny76@mail.sysu.edu.cn {daoyuanchen.cdy, lielin.hyl, bolin.ding, yaliang.li}@alibaba-inc.com" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.305, + 0.357, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.407, + 0.333, + 0.44 + ], + "angle": 0, + "content": "... the left image highlights a fish on the ground, while the right image highlights a red crab on the sand." + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.306, + 0.627, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.407, + 0.606, + 0.44 + ], + "angle": 0, + "content": "... In the left image, the player is holding a baseball glove, while in the right image, the player is holding a basketball." + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.306, + 0.897, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.407, + 0.892, + 0.44 + ], + "angle": 0, + "content": "The left image shows a framed picture on a shelf, while the right image shows a white vase with pink flowers. ..." + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.451, + 0.882, + 0.466 + ], + "angle": 0, + "content": "Figure 1. Three \"object replacement\" examples within IMG-DIFF, highlighting fine-grined difference in both vision and language." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.478, + 0.327, + 0.493 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.485, + 0.859 + ], + "angle": 0, + "content": "High-performance Multimodal Large Language Models (MLLMs) rely heavily on data quality. This study introduces a novel data synthesis method, leveraging insights from contrastive learning and image difference captioning to enhance fine-grained image recognition in MLLMs. By analyzing object differences in detailed regions between similar images, we challenge the model to identify both matching and distinct components. Specifically, our method initially create pairs of similar images that highlight object variations. After that, we introduce a Difference Area Generator for object differences identifying, followed by a Difference Captions Generator for differences describing. The outcome is a high-quality dataset of \"object replacement\" samples, named Img-Diff, which can be expanded as needed due to its automation. We use the generated dataset to fine-tune state-of-the-art (SOTA) MLLMs such as InternVL2, yielding comprehensive improvements across numerous image difference and Visual Question Answering tasks. For instance, the trained models notably surpass the SOTA models GPT-4V and Gemini on the MMVP benchmark. Additionally, we conduct thorough evaluations to confirm the dataset's diversity, quality, and robustness, presenting several insights on the synthesis of such a contrastive dataset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.479, + 0.907, + 0.525 + ], + "angle": 0, + "content": "We release our codes and dataset to encourage further research on multimodal data synthesis and MLLMs' fundamental capabilities for image understanding." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.557, + 0.644, + 0.572 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.583, + 0.907, + 0.779 + ], + "angle": 0, + "content": "The emergence of large language models (LLMs) has revolutionized natural language processing [48, 64], also paved the way for the development of Multimodal Large Language Models (MLLMs) that seamlessly integrate linguistic and visual understanding. Improving the performance of MLLMs hinges on two primary avenues: evolving model architectures and enhancing dataset quality [53]. The majority of state-of-the-art (SOTA) MLLMs [3, 37, 41-43] employ a two-phase training strategy, commencing with a pre-training phase involving extensive image-text pairs for modality alignment, followed by a fine-tuning phase aimed at optimizing visual question answering (VQA) capabilities with specific instruction tuning datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The quality of visual instruction tuning datasets plays a crucial role in MLLMs' overall performance in VQA tasks and diverse downstream applications. With the evolution of visual instruction tuning datasets, several recent studies have explored the integration of object detection and Optical Character Recognition (OCR) datasets, such as RefCOCO [28], Visual Genome [31], OCR-VQA [49], and TextVQA [60], effectively enhancing MLLMs' proficiency" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "3 \"FSIETP\" stands for the Guangdong Provincial Key Laboratory of Fire Science and Intelligent Emergency Technology." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9296" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.385, + 0.106 + ], + "angle": 0, + "content": "in tasks requiring detailed image perception." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.484, + 0.35 + ], + "angle": 0, + "content": "In this paper, we focus on a new direction for enhancing MLLM datasets, driven by the potential of object variations in image pairs to refine models' image recognition capabilities, as demonstrated by advancements in contrastive learning and image difference captioning [5, 26, 51, 80]. Specifically, we introduce a general-purpose yet challenging dataset named IMG-DIFF, which sets itself apart from existing visual instruction tuning datasets by generating pairs of highly similar images featuring subtle object alterations. Rather than compelling MLLMs to focus solely on a single image, our dataset challenges them to analyze paired images and articulate the differences within designated regions, meanwhile taking the high-quality textual descriptions of the difference as learning signals. By doing so, we aim to empower MLLMs with enhanced capabilities of fine-grained image recognition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.484, + 0.652 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our data synthesis method, we integrate the generated dataset into the original visual instruction tuning datasets of LLaVA-1.5 [41], MGM [37], and InternVL2 [12], and conduct fine-tuning. Subsequently, we evaluate their performance on image difference benchmarks, including MMVP [63], Spot-the-Diff [26], and Image-Edit-Request [62], as well as widely recognized MLLM benchmarks. Our evaluation results reveal that after fine-tuning with our IMG-DIFF dataset, the MLLMs achieve notable enhancements in image difference benchmarks, aligning their performance with state-of-the-art (SOTA) models. For instance, they notably surpass the SOTA models GPT-4V [1] and Gemini [16] on the MMVP benchmark, achieving an improvement of up to 12. Moreover, the models exhibit comprehensive improvements across eight well-recognized MLLM benchmarks, achieving an average score improvement of up to \\(3.06\\%\\), underscoring the useful role our dataset plays in bolstering MLLMs' competencies in both image difference recognition and fine-grained image analysis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.804 + ], + "angle": 0, + "content": "We further evaluate the diversity and quality of our dataset, ensuring it encompasses a broad array of object categories while showcasing rich variability. Through meticulous manual labeling, we affirm the high quality of our dataset. Furthermore, we conduct ablation studies to examine the effects of various filtering intensities. We also investigate an alternative methodology for constructing image difference data focusing on \"object removal\", assessing its effectiveness and presenting fruitful insights on the construction of contrastive data." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.806, + 0.414, + 0.82 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.9 + ], + "angle": 0, + "content": "- We present a novel data synthesis method and an effect-proven IMG-DIFF dataset, comprising pairs of highly similar images, with a focus on processes such as segmentation, filtering, and detailed captioning of image differences." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "- We conduct visual instruction tuning on LLaVA-1.5-7B, MGM-7B, and InternVL2-8B using our dataset, and rigorously assess the fine-tuned models' performance on many widely-used MLLM benchmarks and image difference benchmarks. Our dataset brings substantial performance improvements to the fine-tuned MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- We provide a comprehensive evaluation of the diversity and quality of our dataset, confirming its richness and high standard. Through ablation studies, we identify good empirical filtering thresholds for such contrastive dataset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.303 + ], + "angle": 0, + "content": "- We open-source our dataset and codes at https://github.com/modelscope/data-juicer/tree/ImgDiff to facilitate ongoing research, encouraging innovative endeavors in MLLM datasets and image difference methods." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.32, + 0.808, + 0.337 + ], + "angle": 0, + "content": "2. Background and Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.346, + 0.833, + 0.363 + ], + "angle": 0, + "content": "2.1. Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.905, + 0.445 + ], + "angle": 0, + "content": "Multimodal Large Language Models (MLLMs) have exhibited remarkable advancements since their introduction [1, 16, 37, 43, 44]. Research highlights two key factors that primarily influence the effectiveness of MLLMs: model architecture and dataset quality [53]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.906, + 0.567 + ], + "angle": 0, + "content": "With respect to model architecture, notable approaches [2, 3, 25, 32, 35] leverage learnable queries to extract essential information from CLIP [15, 54] image features. Alternatively, LLaVA [41-43] and MGM [37] utilize projection-based interfaces to facilitate interactions between text and image modalities. Furthermore, LLaMA-Adapter [81] and LaVIN [47] implement parameter-efficient tuning mechanisms to transfer image-related information to the LLM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.809 + ], + "angle": 0, + "content": "From the perspective of datasets, there are two prevalent strategies: enhancing the quality of pre-training data and improving instruction tuning data. The former aims for better semantic alignment between images and text by introducing abundant image-text pairs, enhancing MLLMs' fundamental capabilities such as image captioning. As for the latter, research has increasingly concentrated on refining instruction tuning datasets, boosting performance across various question-answering tasks. Works like LLaVA [41-43], SPHINX [40], MGM [37], and InternVL2 [12] leverage high-quality fine-tuning datasets with extensive task diversity, allowing models to excel in tasks related to image perception, reasoning, and optical character recognition. Additionally, methods such as Shikra [10], ASM [72], and PINK [75] utilize substantial amounts of object detection data to enhance the models' localization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In contrast to previous works, our research introduces a fully automated data synthesis method and generates a dataset that emphasizes image differences, showing empirical effectiveness and great potential to augment MLLMs' VQA proficiency, object localization capabilities, and discernment of image distinctions." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9297" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.33, + 0.108 + ], + "angle": 0, + "content": "2.2. Image Difference Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.278 + ], + "angle": 0, + "content": "Datasets focused on image differences typically consist of pairs of similar images and textual descriptions that describe the variations. For instance, the Spot-the-Diff dataset [26] contains pairs of street scenes captured at different times by the same surveillance cameras. The CLEVR-Change dataset [51] delineates scene variations of geometric objects against a clean backdrop. The CUB-Bird dataset [69] focuses on the nuanced differences among various bird species found in natural habitats. The Image-Edit-Request [62] dataset features edited images, accompanied by descriptions of modifications made." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.484, + 0.4 + ], + "angle": 0, + "content": "Leveraging advancements in image editing technologies, some studies have employed generative models and editing techniques to create datasets centered on image differences. A prime example is InstructPix2Pix [6], which utilizes the image editing technique, named Prompt-to-Prompt [20], to direct Stable-Diffusion-1.5 [56] in generating pairs of similar images, while employing GPT-3 [48] to craft the edited text as reference captions for image differences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Our approach, referring to InstructPix2Pix, employs the Prompt-to-Prompt technique alongside an advanced generative model Stable-Diffusion-XL [52], which produces more realistic images, to generate pairs of similar images. Unlike InstructPix2Pix, we incorporate multiple filtering stages to ensure data quality, with a particular emphasis on producing difference captions that focus on specific regions rather than the entire image, which ensures greater accuracy." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.532, + 0.439, + 0.549 + ], + "angle": 0, + "content": "2.3. Models for Image Difference Captioning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Image Difference Captioning (IDC) represents a specialized domain within image captioning characterized by its focus on subtle variations between images. As for the pioneering work in IDC, Spot-the-Diff [26] presents potential change clusters and employs an LSTM [21] decoder to generate difference captions. DUDA [51] explores image differences at the semantic level, using a ResNet [19] and an LSTM to compute dynamic attention weights and produce captions. VARD [65] introduces a viewpoint-adaptive representation disentanglement network based on LSTM for differentiating between real and pseudo changes. Meanwhile, NCT [66] employs a transformer [67] to integrate neighboring features, and CLIP4IDC [18] uses BERT-like training methodologies, adapting a CLIP model for IDC tasks. With the emergence of MLLMs, VIXEN [5] marks the inaugural use of these models for IDC tasks, mapping the features of differential images to text space and employing an LLM to generate image difference captions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our data synthesis method is specifically designed for MLLMs. We build our dataset following the instruction-following format adopted by mainstream MLLMs such as LLaVA-1.5, highlighting a novel research direction aimed at enhancing MLLMs from a data-centric perspective." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.687, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.089, + 0.811, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.093, + 0.904, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.273, + 0.899, + 0.288 + ], + "angle": 0, + "content": "Figure 2. The generation process for \"object replacement\" data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.315, + 0.649, + 0.333 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.342, + 0.625, + 0.357 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.906, + 0.472 + ], + "angle": 0, + "content": "In recent years, contrastive learning have significantly enhanced the image-text understanding of vision-language models [54, 80]. These methods typically involve constructing batches of images and texts, requiring the model to distinguish between matching and non-matching image-text pairs, which improves its ability to differentiate between semantically similar and dissimilar pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.906, + 0.579 + ], + "angle": 0, + "content": "Our method leverages the principles of contrastive learning to generate MLLM image-text data. Specifically, it focuses on replacing objects within image pairs, encouraging MLLMs to identify similarities and differences in specific regions. This method aims to enhance models' capacity to recognize fine-grained differences in images, guided by textual descriptions that highlight detailed distinctions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.746 + ], + "angle": 0, + "content": "As shown in Figure 2, the process of generating \"object replacement\" data can be divided into three key parts. The first part involves creating similar images and forming image pairs, where the only difference between the images in pairs is the objects replaced (Section 3.2). The second part, named the \"Difference Area Generator\", extracts bounding box regions that contain object differences between the images in pairs (Section 3.3). The third part, termed as the \"Difference Captions Generator\" (Section 3.4), utilizes an MLLM to generate descriptive text for the areas with object differences and creates question-answer pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.906, + 0.839 + ], + "angle": 0, + "content": "This process incorporates multiple filtering operations, with specific thresholds outlined in Section 18.4, which we determine through experimental comparisons in Section 17. Additional experimental details, including model selection and the time consumption, are also presented in Section 18. For readability, we present data examples in Section 20." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To enhance reproducibility and reusability, our proposed components and end-to-end construction workflow are implemented as data processing operators and configurable data recipes within Data-Juicer [8, 9]." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9298" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.192, + 0.087, + 0.812, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.321, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 3. An overview of the Difference Area Generator and its three internal components: Image Similarity Filter, Image-text Matching Filter, and Difference Detector." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.376, + 0.313, + 0.392 + ], + "angle": 0, + "content": "3.2. Image Pairs Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.488 + ], + "angle": 0, + "content": "The first step of our data synthesis method is to generate pairs of similar images as candidates. The process is shown in Figure 2. We employ a generative model called Stable-Diffusion-XL [52] and an image editing technique called Prompt-to-Prompt [20] to generate image pairs that highlight object replacement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.625 + ], + "angle": 0, + "content": "We start by obtaining image captions from caption databases, which contain descriptions biased towards real photos. Then, we use an LLM to perform object replacement in the captions. The prompt used is \"Here is a sentence: 'INPUT'. Please only replace one of the objects in this sentence with another object.\" Here, INPUT refers to the original caption. Next, based on the caption pairs, we use the Stable-Diffusion-XL and the Prompt-to-Prompt to generate image pairs with only few objects replaced." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.334, + 0.648 + ], + "angle": 0, + "content": "3.3. Difference Area Generator" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.656, + 0.204, + 0.669 + ], + "angle": 0, + "content": "3.3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.796 + ], + "angle": 0, + "content": "The Difference Area Generator is used to identify the locations of object differences between the two images in pairs. Although object detection models are capable of detecting objects in images, the range of object categories is quite limited [27, 78]. Therefore, to increase the number of detectable object categories and enhance data diversity, we develop the Difference Area Generator based on segmentation and image similarity comparisons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As shown in Figure 3, we first use an Image Similarity Filter to obtain image pairs with high similarity but not identical. Next, we use the FastSAM [82] to perform image segmentation on each image. After that, we crop the images based on the bounding box information obtained from segmentation and use an Image-text Matching Filter to filter the cropped sub-images for the presence of valid objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.453 + ], + "angle": 0, + "content": "Finally, we use a Difference Detector to determine whether there are indeed differences between the bounding box regions of the two images and perform IoU (Intersection over Union) filtering to remove overlapping bounding boxes, ultimately obtaining valid bounding box information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.462, + 0.722, + 0.476 + ], + "angle": 0, + "content": "3.3.2. Image Similarity Filter" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.907, + 0.648 + ], + "angle": 0, + "content": "The Image Similarity Filter aims to filter image pairs based on the degree of similarity. The module first uses CLIP [54] to extract image features from each image in pairs and then calculates the cosine similarity score. If their score falls within the pre-set threshold, the image pair will be considered valid. Specifically, we use the module twice in the Difference Area Generator. Before the segmentation, we use the module to ensure that the images in pairs are highly similar but not the same. In the Difference Detector stage, after cropping, we use the module to filter the sub-image pairs and keep only the differing ones." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.657, + 0.75, + 0.672 + ], + "angle": 0, + "content": "3.3.3. Image-text Matching Filter" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.676, + 0.909, + 0.828 + ], + "angle": 0, + "content": "The Image-text Matching Filter determines whether an image contains valid objects (i.e. the replaced or replacing objects). This module first uses BLIP [34] to extract image features, which are then compared with text features of object names. When the image-text matching score falls within the pre-set threshold, we consider the image to contain valid objects. In the mid-stage of the Difference Area Generator, after cropping, we use the module to determine whether these sub-images contain valid objects and get valid bounding boxes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.837, + 0.696, + 0.85 + ], + "angle": 0, + "content": "3.3.4. Difference Detector" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.902 + ], + "angle": 0, + "content": "The Difference Detector determines whether there are differences between the two regions in the image pair of the same bounding box. Based on a given bounding box, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9299" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.088, + 0.868, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.257, + 0.733, + 0.273 + ], + "angle": 0, + "content": "Figure 4. An overview of the Difference Captions Generator and its two stages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.299, + 0.484, + 0.405 + ], + "angle": 0, + "content": "first crop two sub-images from both image A and image B. The two sub-images are then filtered through the Image Similarity Filter and the bounding box is considered effective only if the difference is significant enough. After processing all bounding boxes, we use the IoU method to filter out overlapping bounding boxes. Only the bounding boxes with a higher degree of difference are retained." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.412, + 0.366, + 0.429 + ], + "angle": 0, + "content": "3.4. Difference Captions Generator" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.434, + 0.204, + 0.449 + ], + "angle": 0, + "content": "3.4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.589 + ], + "angle": 0, + "content": "After obtaining the valid bounding box regions, we use the Difference Captions Generator to generate descriptions for the differences inside these areas (with each round of the process focusing on only one bounding box in one image pair). Evidently, an image pair can contain multiple object differences but a single difference caption cannot fully capture all of them. Therefore, we highlight specific regions with red boxes and provide targeted difference captions to ensure greater accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.484, + 0.695 + ], + "angle": 0, + "content": "The module consists of two stages: the first stage generates captions for the content in the bounding box regions and then filters the bounding boxes using an Image-text Matching Filter and a Captions Similarity Filter. The second stage uses the content captions and the images highlighted with red boxes to generate difference captions. The overview is shown in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.701, + 0.391, + 0.717 + ], + "angle": 0, + "content": "3.4.2. Stage1: Object Labeling & Filtering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In Stage 1, for each image pair, we first select N bounding box regions with the lowest similarity between images (N is set to 5 in this project) as candidate regions. Then, for each bounding box, we use the MLLM LLaVA-NEXT [43] to describe its corresponding regions and then apply two filtering processes: the first filter is an Image-text Matching Filter, which checks whether the content of the regions matches the captions; the second filter is an Captions Similarity Filter, which assesses whether there are differences between the two captions. Once the filtering is complete, we obtain valid bounding boxes and captions for subsequent difference captioning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.299, + 0.74, + 0.314 + ], + "angle": 0, + "content": "3.4.3. Captions Similarity Filter" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.317, + 0.906, + 0.393 + ], + "angle": 0, + "content": "The Captions Similarity Filter determines whether the two captions of the same bounding box coordinates are different. We use CLIP to obtain text features and calculate the cosine similarity between them. When the score is low enough, we consider the two captions to be different." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.399, + 0.84, + 0.415 + ], + "angle": 0, + "content": "3.4.4. Stage2: Difference Captions Generating" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.906, + 0.54 + ], + "angle": 0, + "content": "In Stage 2, for each valid bounding box of each image pair, we first draw two red boxes into the images based on the bounding box coordinates, highlighting the differences for easier localization. Then, we provide the MLLM LLaVA-NEXT with the captions of the bounding box content and instruct it to describe the difference based on these captions and the highlighted images. Finally, we can obtain the difference caption for the highlighted area." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.662, + 0.563 + ], + "angle": 0, + "content": "3.5. Data Statistics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.719 + ], + "angle": 0, + "content": "Using captions from MSCOCO [39], we generate 118K pairs of similar images. We then employ the Image Similarity Filter to get 38,533 highly similar but not identical image pairs. Next, we use the Difference Area Generator to filter and produce 117,779 pieces of valid bounding box information (with a maximum of 5 valid bounding boxes per image pair). Finally, we employ the Difference Captions Generator to filter and generate 12,688 high-quality \"object replacement\" instances. Our evaluations on the main page are based on this dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.721, + 0.906, + 0.812 + ], + "angle": 0, + "content": "In addition, we also use captions from the LLaVA pretraining dataset to generate 34,538 \"object replacement\" samples. In Section 10 of the appendix, we evaluate this dataset and discuss the relationship between data quantity, data quality, and model performance improvement, emphasizing the importance of prioritizing quality over quantity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.783, + 0.839 + ], + "angle": 0, + "content": "4. Evaluation and Main Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.702, + 0.866 + ], + "angle": 0, + "content": "4.1. Evaluation Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our IMG-DIFF dataset, we use it to fine-tune SOTA MLLMs, including LLaVA-1.5-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9300" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.211 + ], + "angle": 0, + "content": "7B, MGM-7B, and InternVL2-8B on the main page, as well as InternVL2-1B and LLaVA-1.5-13B in the supplementary materials. We evaluate these models on extensive benchmarks commonly used for image difference and MLLMs. Specifically, the image difference benchmarks include MMVP [63], Spot-the-Diff [26], and Image-Edit-Request [62]. Besides, the details and results of the MLLM benchmarks are shown in Section 4.5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.483, + 0.394 + ], + "angle": 0, + "content": "During our fine-tuning process, we first mix our data with MLLMs' original visual instruction tuning data respectively. For LLaVA-1.5 and MGM, we conduct the fine-tuning anew. For InternVL2, we follow the guidelines from its official repository and perform a second fine-tuning. (To ensure a fair comparison, we perform a second fine-tuning on InternVL2 using its original fine-tuning dataset, termed as InternVL2-FT, and use it as a baseline.) Regarding the Spot-the-Diff and Image-Edit-Request benchmarks, as they contain data of training splits, we further fine-tune the fine-tuned MLLMs for an additional 2 epochs using only these benchmarks' training data." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.394, + 0.482, + 0.41 + ], + "angle": 0, + "content": "In the tables, \"RP\" represents \"object replacement\" data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.418, + 0.392, + 0.433 + ], + "angle": 0, + "content": "4.2. Results on the MMVP Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.483, + 0.576 + ], + "angle": 0, + "content": "The MMVP benchmark is designed to systematically assess the visual capabilities of MLLMs. Its data processing method is highly related to image difference: it first collects CLIP-blind pairs, which have similar CLIP features but differ in image content. Then, the differences between the images are manually described and question-answer pairs are created. Hence, the questions in MMVP is highly relevant to our dataset, as both place significant emphasis on the differences between similar images." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.587, + 0.483, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.75, + 0.472, + 0.765 + ], + "angle": 0, + "content": "Figure 5. Performance comparison on the MMVP benchmark." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As shown in Figure 5, fine-tuning MLLMs with our \"object replacement\" data significantly improves their performance on the MMVP benchmark. After fine-tuning, the score of LLaVA-1.5-7B exceeds that of LLaVA-1.5-13B. Furthermore, the fine-tuned MGM-7B shows a significant improvement in score compared to the original MGM-7B, even surpassing the scores of the SOTA models GPT-4V and Gemini by up to 12. Additionally, the performance of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "InternVL2-8B is also enhanced. These results suggest that our dataset enhances MLLMs' abilities to distinguish images with similar CLIP features but different content." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.15, + 0.857, + 0.166 + ], + "angle": 0, + "content": "4.3. Results on the Spot-the-Diff Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.173, + 0.905, + 0.233 + ], + "angle": 0, + "content": "The dataset of Spot-the-Diff comprises pairs of street view images that display subtle object differences. These images are obtained by capturing scenes from fixed surveillance cameras at different time." + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.246, + 0.863, + 0.261 + ], + "angle": 0, + "content": "Table 1. Performance comparison on Spot-the-Diff." + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.271, + 0.885, + 0.422 + ], + "angle": 0, + "content": "
ModelBLEUMETEORCIDEr-DROUGE-L
VAM[59]10.112.438.131.3
IFDC[23]8.711.73730.2
DUDA+Aux[22]8.112.534.529.9
VACC[29]9.712.641.532.1
LLaVA-1.5-7B8.512.038.330.1
LLaVA-1.5-7B + RP9.713.043.230.8
MGM-7B9.912.046.331.5
MGM-7B + RP10.813.153.533.0
InternVL2-8B-FT6.611.726.527.3
InternVL2-8B + RP8.412.832.228.5
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.906, + 0.609 + ], + "angle": 0, + "content": "Following previous works, our fine-tuned MLLMs are evaluated on BLEU [50], METEOR [4], CIDEr-D [68] and ROUGE-L [38]. As shown in Table 1, after fine-tuning with our \"object replacement\" data, both LLaVA-1.5-7B, MGM-7B, and InternVL2-8B achieve significant performance gains on Spot-the-Diff. Compared to score increases seen in prior models, our dataset yields even greater enhancements than those from previous iterations of image difference models. These positive results indicate that our dataset can enhance the ability of MLLMs to detect subtle differences between similar images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.62, + 0.813, + 0.637 + ], + "angle": 0, + "content": "4.4. Results on Image-Editing-Request" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.764 + ], + "angle": 0, + "content": "The Image-Editing-Request benchmark focuses on image editing. Each instance in its dataset consists of an image pair (i.e. a source image and an edited image) and an editing instruction that describes the transformation. During the evaluation, our models are required to generate transformation descriptions for these image pairs, and we then calculate the BLEU, METEOR, CIDEr-D, and ROUGE-L scores using the models' responses and the reference answers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.902 + ], + "angle": 0, + "content": "As shown in Table 2, LLaVA-1.5-7B, MGM-7B, and InternVL2-8B originally show SOTA performance on the Image-Edit-Request benchmark. Upon incorporating our \"object replacement\" data for better fine-tuning, the performance of all three models improves even more, achieving new SOTA scores. This suggests that our dataset enhances MLLMs' abilities to recognize similarities and dissimilarities in image pairs, as well as enables them to describe differences more accurately." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9301" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.09, + 0.46, + 0.104 + ], + "angle": 0, + "content": "Table 2. Performance comparison on Image-Edit-Request." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.115, + 0.465, + 0.281 + ], + "angle": 0, + "content": "
ModelBLEUMETEORCIDEr-DROUGE-L
VARD[65]1014.835.739
CLIP4IDC[18]8.214.632.240.4
NCT[66]8.11534.238.8
BiDiff[61]6.914.627.738.5
VIXEN[5]8.615.438.142.5
LLaVA-1.5-7B15.117.860.645.2
LLaVA-1.5-7B + RP16.219.560.946.7
MGM-7B16.517.766.844.8
MGM-7B + RP16.618.268.145.7
InternVL2-8B-FT12.414.151.538.9
InternVL2-8B + RP12.514.256.039.4
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.307, + 0.371, + 0.322 + ], + "angle": 0, + "content": "4.5. Results on MLLM Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.33, + 0.484, + 0.512 + ], + "angle": 0, + "content": "Aside from the evaluations related to image difference discrimination, we also assess the performance of our \"object replacement\" data in enhancing the comprehensive abilities of MLLMs. We test the fine-tuned MLLMs using commonly used MLLM benchmarks, including VQAv2 [17] and GQA [24] for assessing the comprehensive VQA capabilities; MMBench [45], MM-Vet [76], ScienceQA [46], and SEED-Bench [33] for testing perceptual and reasoning abilities; and POPE [36] for evaluating fine-grained object localization abilities. Table 3 presents the results on these MLLM benchmarks, with the \\(\\triangle\\) metric indicating the percentage improvement averaged across them." + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.527, + 0.466, + 0.541 + ], + "angle": 0, + "content": "Table 3. Performance comparison on 8 MLLM benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.552, + 0.472, + 0.654 + ], + "angle": 0, + "content": "
ModelVQAv2GQAPOPEMMB\\( {\\mathrm{{MMB}}}^{CN} \\)
LLaVA-1.5-7B78.56285.964.358.3
LLaVA-1.5-7B + RP79.3↑62.8↑86.4↑66.1↑59.8↑
MGM-7B80.462.68669.358.9
MGM-7B + RP80.7↑62.7↑86.2↑68.7↓59.6↑
InternVL2-8B-FT81.862.687.782.581.5
InternVL2-8B + RP81.862.688.0↑82.7↑81.4↓
" + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.661, + 0.472, + 0.763 + ], + "angle": 0, + "content": "
ModelMM-Vet\\( \\mathbf{SQA}^I \\)SEED\\( \\bigtriangleup \\)
LLaVA-1.5-7B30.566.858.6-
LLaVA-1.5-7B + RP33.2↑68.2↑61.7↑+3.06%
MGM-7B40.870.663.5-
MGM-7B + RP44.1↑71.7↑63.2↓+1.28%
InternVL2-8B-FT49.296.569.5-
InternVL2-8B + RP52.6↑96.6↑69.9↑+1.01%
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Based on Table 3, after fine-tuning with our dataset, the performance of LLaVA-1.5-7B shows a comprehensive improvement, with an average increase of \\(3.06\\%\\) across all benchmarks. For MGM-7B and InternVL2-8B, the improvements brought by our dataset are not as pronounced as those observed with LLaVA-1.5-7B, as their training datasets already encompass a large volume of high-quality data, but they still achieve an average increase of \\(1.28\\%\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "and \\(1.01\\%\\). These improvements indicate that fine-tuning MLLMs with our dataset not only enhances their ability to discern differences but also improves their overall visual capabilities, thereby making them better address VQA tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.165, + 0.88, + 0.183 + ], + "angle": 0, + "content": "5. Evaluation of Data Quality and Diversity" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.19, + 0.65, + 0.206 + ], + "angle": 0, + "content": "5.1. Data Quality" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.212, + 0.907, + 0.605 + ], + "angle": 0, + "content": "To assess the quality of our Img-Diff dataset, we randomly select 1,000 instances of \"object replacement\" data and employ multiple professional dataset annotators to evaluate the samples based on three metrics. The final scores are determined through a voting process. Specifically, the first metric is \"Bounding Box Difference\", which evaluates whether there are differences between the two highlighted regions in pairs. If the objects are different, we score it as \"high\"; if the objects are the same but their features (such as color, shape, etc.) are different, we score it as \"medium\"; if the objects are the same and their features are similar, we score it as \"low\". The second metric is \"Content Caption Accuracy\", which evaluates whether the captions generated by Stage 1 of the Difference Captions Generator accurately describe the two highlighted regions. If the captions are correct, we score it as \"high\"; if the captions identify the objects but incorrectly describe their features, we score it as \"medium\"; if the captions incorrectly identify the objects, we score it as \"low\". The third metric is \"Difference Caption Accuracy\", which evaluates whether the final difference captions accurately describe the object differences between highlighted regions of the image pairs. If the description is accurate, we score it as \"high\"; if it recognizes the objects but the feature description is incorrect, we score it as \"medium\"; if the object recognition is incorrect, we score it as \"low\". The results are shown in Figure 6." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.615, + 0.905, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.69, + 0.896, + 0.705 + ], + "angle": 0, + "content": "Figure 6. Quality assessment of the \"object replacement\" data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Based on Figure 6, our dataset demonstrates a high level of quality. For the \"Bounding Box Difference\" metric, only \\(4.5\\%\\) of the instances are classified as \"low\", and nearly \\(80\\%\\) of instances exhibit valid object differences between their two highlighted regions. In terms of \"Content Caption Accuracy\", \\(80.1\\%\\) of highlighted region pairs are described accurately, indicating that using an MLLM for labeling is effective and that our filtering strategy is also functioning well. For the \"Difference Caption Accuracy\" metric, over \\(70\\%\\) of the difference descriptions are completely accurate, with \\(21.8\\%\\) of the samples having errors solely in feature labeling while still maintaining correct descriptions of ob-" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9302" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "ject differences, which underscores the effectiveness of our difference caption generation strategy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.133, + 0.24, + 0.15 + ], + "angle": 0, + "content": "5.2. Data Diversity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.156, + 0.483, + 0.367 + ], + "angle": 0, + "content": "In designing the generation process, we have made efforts to enhance the diversity of our dataset, which can be divided into two aspects: (a) the inherent diversity of the caption databases; and (b) the introduction of new object names through our object replacement strategy. As for the former, we are using the captions from MS COCO and the LLaVA pre-training dataset, which cover a majority of object categories. We can easily further enhance the data diversity by employing caption databases with greater variability. As for the latter, we employ two methods, including increasing the temperature of the LLM used for noun replacement, as well as randomly replacing object names with nouns from a noun lexicon (shown in Appendix, Section 11). We can further improve data diversity by expanding the noun lexicon." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.519 + ], + "angle": 0, + "content": "By analyzing the valid object names included in the captions that describe highlighted regions, we assess the diversity of our \"object replacement\" data. Specifically, we count the total number of object categories covered, and the total number of unique \"object replacement pairs\". Through our statistical analysis, we find that our dataset covers 1,203 object categories, which encompasses most real-life objects. An \"object replacement pair\" refers to the combination of the replaced and replacing object names. Our dataset includes 3,680 unique \"object replacement pairs\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.61 + ], + "angle": 0, + "content": "To evaluate the coverage of common object categories in our dataset, we analyze the occurrences of the object names from the Object365 [58] dataset in our own dataset. The results show that the object names from the Object365 dataset appears 13,164 times in total, which accounts for approximately \\(52.06\\%\\) of the total occurrences of object names." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.687 + ], + "angle": 0, + "content": "These statistics show that our data covers a substantial number of common object names, ensuring a high frequency of common objects. Additionally, less common object names make up nearly half of our dataset, indicating that our dataset is both diverse and comprehensive." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.702, + 0.431, + 0.72 + ], + "angle": 0, + "content": "6. More Empirical Supports and Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.789 + ], + "angle": 0, + "content": "We list additional experimental details and extensions of our data synthesis method in the supplementary material, and conduct further experiments to validate the effectiveness of our IMG-DIFF dataset, along with its limitations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Further Experiments and Analysis. In Section 9, we compare our IMG-DIFF dataset with existing image difference datasets on characteristics and performance, validating its superiority. In Section 10, we discuss the relationship between the data volume and model performance improvement. In Section 11, we explore the use of lexicons for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "object replacement and validate its effectiveness in enhancing diversity. In Section 12, we evaluate our dataset using the Contrastive Chain-of-Thought method. In Section 13, we validate the effectiveness of our dataset on MLLMs of different scales. In Section 14, we evaluate top-performing MLLMs on difference detection. In Section 15, we assess the impact of unnatural images. In Section 16, we show our data's efficacy in enhancing spatial reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.232, + 0.905, + 0.324 + ], + "angle": 0, + "content": "Implementation Details. In Section 17, we investigate the effect of different filtering intensities on the performance of the generated datasets. In Section 18, we present additional experimental details, including image pair preprocessing, model training procedures, model selection, filtering thresholds, and time consumption." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.343, + 0.907, + 0.418 + ], + "angle": 0, + "content": "The \"Object Removal\" Exploration. In Section 19, we generate an extended dataset that focuses on object removal, which prompts MLLMs to analyze which image contains a specific object. The new data brings further improvement to the fine-tuned MLLM." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.432, + 0.634, + 0.449 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.789 + ], + "angle": 0, + "content": "In this paper, we draw inspiration from recent advances in contrastive learning and image difference captioning to propose a novel method of contrastive data synthesis. This method enables the creation of a high-quality dataset called \"IMG-DIFF\", which highlights object differences focusing on fine-grained regions in images. Specifically, we first generate pairs of similar images where the main focus is on object replacement. Then, we use the proposed Difference Area Generator and Difference Captions Generator to generate difference captions and form question-answer pairs. In contrast to previous image difference datasets, our dataset focuses exclusively on specific regions inside images. This characteristic circumvents the issue where a single description cannot fully capture the differences of the whole images in pairs, enhancing the accuracy. Afterwards, we fine-tune many MLLMs using the generated dataset, yielding high-performance scores on par with SOTA models in image difference tasks and comprehensive performance improvements in eight widely recognized MLLM benchmarks. These results confirm the effectiveness of our dataset to improve the ability of MLLMs in recognizing detailed differences between images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.79, + 0.909, + 0.888 + ], + "angle": 0, + "content": "In a nutshell, we provide a series of insights about the construction of high-quality image difference datasets, showing great potential to effectively and efficiently enhance MLLMs via contrastive data-centric approaches. With this work, we hope it can catalyze further investigation into the realm of image difference datasets and the fine-grained image recognition capabilities of MLLMs." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9303" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.339 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.342, + 0.483, + 0.41 + ], + "angle": 0, + "content": "[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.412, + 0.483, + 0.479 + ], + "angle": 0, + "content": "[5] Alexander Black, Jing Shi, Yifei Fan, Tu Bui, and John Colomosse. Vixen: Visual text comparison network for image difference captioning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 846-854, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.483, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.553, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[7] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, Shuaibin Li, Wei Li, Yinling Li, Hongwei Liu, Jiangning Liu, Jiawei Hong, Kaiwen Liu, Kuikun Liu, Xiaoran Liu, Chengqi Lv, Haijun Lv, Kai Lv, Li Ma, Runyuan Ma, Zerun Ma, Wenchang Ning, Linke Ouyang, Jiantao Qiu, Yuan Qu, Fukai Shang, Yunfan Shao, Demin Song, Zifan Song, Zhihao Sui, Peng Sun, Yu Sun, Huanze Tang, Bin Wang, Guoteng Wang, Jiaqi Wang, Jiayu Wang, Rui Wang, Yudong Wang, Ziyi Wang, Xingjian Wei, Qizhen Weng, Fan Wu, Yingtong Xiong, Chao Xu, Ruiliang Xu, Hang Yan, Yirong Yan, Xiaogui Yang, Haochen Ye, Huaiyuan Ying, Jia Yu, Jing Yu, Yuhang Zang, Chuyu Zhang, Li Zhang, Pan Zhang, Peng Zhang, Ruijie Zhang, Shuo Zhang, Songyang Zhang, Wenjian Zhang, Wenwei Zhang, Xingcheng Zhang, Xinyue Zhang, Hui Zhao, Qian Zhao, Xiaomeng Zhao, Fengzhe Zhou, Zaida Zhou, Jingming Zhuo, Yicheng Zou, Xipeng Qiu, Yu Qiao and Dahua Lin. Internl m2 technical report, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[8] Daoyuan Chen, Yilun Huang, Zhijian Ma, Hesen Chen, Xuchen Pan, Ce Ge, Dawei Gao, Yuexiang Xie, Zhaoyang" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.148 + ], + "angle": 0, + "content": "Liu, Jinyang Gao, et al. Data-juicer: A one-stop data processing system for large language models. In Companion of the 2024 International Conference on Management of Data, pages 120-134, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.15, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[9] Daoyuan Chen, Haibin Wang, Yilun Huang, Ce Ge, Yaliang Li, Bolin Ding, and Jingren Zhou. Data-juicer sandbox: A feedback-driven suite for multimodal data-model codevelopment. arXiv preprint arXiv:2407.11784, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[10] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[11] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Qinglan Huang, Canyu Chen, Qinghao Ye, Zhihong Zhu, Yuqing Zhang, Jiawei Zhou, Zhuokai Zhao, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Mj-bench: Is your multimodal reward model really a good judge for text-to-image generation?, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[12] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.433, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[14] W Dai, J Li, D Li, AMH Tiong, J Zhao, W Wang, B Li, P Fung, and S Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.575, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.703, + 0.673 + ], + "angle": 0, + "content": "[16] Google. Gemini. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[17] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[18] Zixin Guo, Tzu-Jui Julius Wang, and Jorma Laaksonen. Clip4idc: Clip for image difference captioning. arXiv preprint arXiv:2206.00629, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[20] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9304" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.122, + 0.486, + 0.178 + ], + "angle": 0, + "content": "[22] Mehrdad Hosseinzadeh and Yang Wang. Image change captioning by learning from an auxiliary task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2725-2734, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.179, + 0.485, + 0.248 + ], + "angle": 0, + "content": "[23] Qingbao Huang, Yu Liang, Jielong Wei, Yi Cai, Hanyu Liang, Ho-fung Leung, and Qing Li. Image difference captioning with instance-level fine-grained feature representation. IEEE transactions on multimedia, 24:2004-2017, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.25, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[24] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.321, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[25] Laurençon Hugo, van Strien Daniel, Bekman Stas, Tronchon Leo, Saulnier Lucile, Wang Thomas, Karamcheti Siddharth, Singh Amanpreet, Pistilli Giada, Jernite Yacine, and Sanh Victor. Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.405, + 0.484, + 0.447 + ], + "angle": 0, + "content": "[26] Harsh Jhamtani and Taylor Berg-Kirkpatrick. Learning to describe differences between pairs of similar images. arXiv preprint arXiv:1808.10584, 2018. 2, 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.448, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[27] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. *Ultralytics YOLO*, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.477, + 0.485, + 0.547 + ], + "angle": 0, + "content": "[28] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.548, + 0.485, + 0.616 + ], + "angle": 0, + "content": "[29] Hoeseong Kim, Jongseok Kim, Hyungseok Lee, Hyunsung Park, and Gunhee Kim. Agnostic change captioning with cycle consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2095-2104, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.618, + 0.485, + 0.687 + ], + "angle": 0, + "content": "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.689, + 0.485, + 0.771 + ], + "angle": 0, + "content": "[31] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.773, + 0.485, + 0.815 + ], + "angle": 0, + "content": "[32] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.817, + 0.485, + 0.871 + ], + "angle": 0, + "content": "[33] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.873, + 0.485, + 0.903 + ], + "angle": 0, + "content": "[34] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.137, + 0.907, + 0.193 + ], + "angle": 0, + "content": "[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.195, + 0.907, + 0.25 + ], + "angle": 0, + "content": "[36] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.253, + 0.906, + 0.309 + ], + "angle": 0, + "content": "[37] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv:2403.18814, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.311, + 0.906, + 0.353 + ], + "angle": 0, + "content": "[38] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.355, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.442, + 0.906, + 0.512 + ], + "angle": 0, + "content": "[40] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.514, + 0.906, + 0.554 + ], + "angle": 0, + "content": "[41] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.557, + 0.906, + 0.599 + ], + "angle": 0, + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.601, + 0.906, + 0.657 + ], + "angle": 0, + "content": "[43] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,2, 5,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.66, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[44] Mengsha Liu, Daoyuan Chen, Yaliang Li, Guian Fang, and Ying Shen. ChartThinker: A contextual chain-of-thought approach to optimized chart summarization. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 3057-3074, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.746, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[45] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.817, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[46] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9305" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[47] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. arXiv preprint arXiv:2305.15023, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[48] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[49] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[50] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[51] Dong Huk Park, Trevor Darrell, and Anna Rohrbach. Robust change captioning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4624-4633, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[52] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[53] Zhen Qin, Daoyuan Chen, Wenhao Zhang, Liuyi Yao, Yilun Huang, Bolin Ding, Yaliang Li, and Shuiguang Deng. The synergy between data and multi-modal large language models: A survey from co-development perspective. arXiv preprint arXiv:2407.08583, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[55] Kanchana Ranasinghe, Satya Narayan Shukla, Omid Poursaeed, Michael S Ryoo, and Tsung-Yu Lin. Learning to localize objects improves spatial reasoning in visual-llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12977–12987, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[57] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[58] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[59] Xiangxi Shi, Xu Yang, Jiuming Gu, Shafiq Joty, and Jianfei Cai. Finding it at another side: A viewpoint-adapted match" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "ing encoder for change captioning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 574-590. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[60] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[61] Yaoqi Sun, Liang Li, Tingting Yao, Tongyv Lu, Bolun Zheng, Chenggang Yan, Hua Zhang, Yongjun Bao, Guiguang Ding, and Gregory Slabaugh. Bidirectional difference locating and semantic consistency reasoning for change captioning. International Journal of Intelligent Systems, 37 (5):2969-2987, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[62] Hao Tan, Franck Dernoncourt, Zhe Lin, Trung Bui, and Mohit Bansal. Expressing visual relationships via language. arXiv preprint arXiv:1906.07689, 2019. 2, 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[63] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[64] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[65] Yunbin Tu, Liang Li, Li Su, Junping Du, Ke Lu, and Qingming Huang. Adaptive representation disentanglement network for change captioning. IEEE Transactions on Image Processing, 32:2620-2635, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[66] Yunbin Tu, Liang Li, Li Su, Ke Lu, and Qingming Huang. Neighborhood contrastive transformer for change captioning. IEEE Transactions on Multimedia, 25:9518-9529, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[67] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[68] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[69] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 3, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[70] Jiayu Wang, Yifei Ming, Zhenmei Shi, Vibhav Vineet, Xin Wang, Yixuan Li, and Neel Joshi. Is a picture worth a thousand words? delving into spatial reasoning for vision language models. In The Thirty-Eighth Annual Conference on Neural Information Processing Systems, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[71] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9306" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.119 + ], + "angle": 0, + "content": "perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[72] Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, et al. The all-seeing project: Towards panoptic visual recognition and understanding of the open world. In The Twelfth International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[73] Weiyun Wang, Zhe Chen, Wenhai Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Jinguo Zhu, Xizhou Zhu, Lewei Lu, Yu Qiao, and Jifeng Dai. Enhancing the reasoning ability of multimodal large language models via mixed preference optimization. arXiv preprint arXiv:2411.10442, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[74] Zhe Xu, Daoyuan Chen, Zhenqing Ling, Yaliang Li, and Ying Shen. Mindgym: Enhancing vision-language models via synthetic self-challenging questions, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.386 + ], + "angle": 0, + "content": "[75] Shiyu Xuan, Qingpei Guo, Ming Yang, and Shiliang Zhang. Pink: Unveiling the power of referential comprehension for multi-modal llms. arXiv preprint arXiv:2310.00582, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[76] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[77] Daoan Zhang, Junming Yang, Hanjia Lyu, Zijian Jin, Yuan Yao, Mingkai Chen, and Jiebo Luo. Cocot: Contrastive chain-of-thought prompting for large multimodal models with multiple image inputs. arXiv preprint arXiv:2401.02582, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[78] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[79] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.631, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[80] Qi Zhang, Yifei Wang, and Yisen Wang. On the generalization of multi-modal contrastive learning. In International Conference on Machine Learning, pages 41677-41693. PMLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.482, + 0.754 + ], + "angle": 0, + "content": "[81] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.757, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[82] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.801, + 0.482, + 0.854 + ], + "angle": 0, + "content": "[83] Ting Zhou, Daoyuan Chen, Qirui Jiao, Bolin Ding, Yaliang Li, and Ying Shen. Humanvbench: Exploring human-centric video understanding capabilities of mllms with synthetic benchmark data, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9307" + } + ] +] \ No newline at end of file diff --git a/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/e6b1bb30-e9d5-4708-960e-31ff7e5e9a5b_origin.pdf b/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/e6b1bb30-e9d5-4708-960e-31ff7e5e9a5b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..04a48788ecdb119c7538d2f72bdce11341f9d3c5 --- /dev/null +++ b/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/e6b1bb30-e9d5-4708-960e-31ff7e5e9a5b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09f4212fe0e0da3d994d8beb1cb16c0d4a8e5f93395d53f56b20c2bbb6c409c1 +size 1008625 diff --git a/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/full.md b/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d38734a5500d33e57113ebc2d10db028cd0ae299 --- /dev/null +++ b/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/full.md @@ -0,0 +1,348 @@ +# Img-Diff: Contrastive Data Synthesis for Multimodal Large Language Models + +Qirui Jiao $^{1}$ , Daoyuan Chen $^{2}$ , Yilun Huang $^{2}$ , Bolin Ding $^{2}$ , Yaliang Li $^{2}$ , Ying Shen $^{1,3}$ + +$^{1}$ Sun Yat-Sen University, $^{2}$ Alibaba Group, $^{3}$ FSIETP + +jiaoqr3@mail2.sysu.edu.cn, sheny76@mail.sysu.edu.cn {daoyuanchen.cdy, lielin.hyl, bolin.ding, yaliang.li}@alibaba-inc.com + +![](images/95f8a8be447b0fda9939655bd20481218816e7c7b8b3c5f0dd4b7fd24be2e195.jpg) +... the left image highlights a fish on the ground, while the right image highlights a red crab on the sand. + +![](images/082576aa6da0eca498f07c2c1f67f32cbe00f7ac4ee87ee0afc122825e7d8b1b.jpg) +... In the left image, the player is holding a baseball glove, while in the right image, the player is holding a basketball. +Figure 1. Three "object replacement" examples within IMG-DIFF, highlighting fine-grined difference in both vision and language. + +![](images/b8040c21694a3b36da7e7429dfdc17bd7198ceacf7eba16ef649337144ce3df9.jpg) +The left image shows a framed picture on a shelf, while the right image shows a white vase with pink flowers. ... + +# Abstract + +High-performance Multimodal Large Language Models (MLLMs) rely heavily on data quality. This study introduces a novel data synthesis method, leveraging insights from contrastive learning and image difference captioning to enhance fine-grained image recognition in MLLMs. By analyzing object differences in detailed regions between similar images, we challenge the model to identify both matching and distinct components. Specifically, our method initially create pairs of similar images that highlight object variations. After that, we introduce a Difference Area Generator for object differences identifying, followed by a Difference Captions Generator for differences describing. The outcome is a high-quality dataset of "object replacement" samples, named Img-Diff, which can be expanded as needed due to its automation. We use the generated dataset to fine-tune state-of-the-art (SOTA) MLLMs such as InternVL2, yielding comprehensive improvements across numerous image difference and Visual Question Answering tasks. For instance, the trained models notably surpass the SOTA models GPT-4V and Gemini on the MMVP benchmark. Additionally, we conduct thorough evaluations to confirm the dataset's diversity, quality, and robustness, presenting several insights on the synthesis of such a contrastive dataset. + +We release our codes and dataset to encourage further research on multimodal data synthesis and MLLMs' fundamental capabilities for image understanding. + +# 1. Introduction + +The emergence of large language models (LLMs) has revolutionized natural language processing [48, 64], also paved the way for the development of Multimodal Large Language Models (MLLMs) that seamlessly integrate linguistic and visual understanding. Improving the performance of MLLMs hinges on two primary avenues: evolving model architectures and enhancing dataset quality [53]. The majority of state-of-the-art (SOTA) MLLMs [3, 37, 41-43] employ a two-phase training strategy, commencing with a pre-training phase involving extensive image-text pairs for modality alignment, followed by a fine-tuning phase aimed at optimizing visual question answering (VQA) capabilities with specific instruction tuning datasets. + +The quality of visual instruction tuning datasets plays a crucial role in MLLMs' overall performance in VQA tasks and diverse downstream applications. With the evolution of visual instruction tuning datasets, several recent studies have explored the integration of object detection and Optical Character Recognition (OCR) datasets, such as RefCOCO [28], Visual Genome [31], OCR-VQA [49], and TextVQA [60], effectively enhancing MLLMs' proficiency + +in tasks requiring detailed image perception. + +In this paper, we focus on a new direction for enhancing MLLM datasets, driven by the potential of object variations in image pairs to refine models' image recognition capabilities, as demonstrated by advancements in contrastive learning and image difference captioning [5, 26, 51, 80]. Specifically, we introduce a general-purpose yet challenging dataset named IMG-DIFF, which sets itself apart from existing visual instruction tuning datasets by generating pairs of highly similar images featuring subtle object alterations. Rather than compelling MLLMs to focus solely on a single image, our dataset challenges them to analyze paired images and articulate the differences within designated regions, meanwhile taking the high-quality textual descriptions of the difference as learning signals. By doing so, we aim to empower MLLMs with enhanced capabilities of fine-grained image recognition. + +To evaluate the effectiveness of our data synthesis method, we integrate the generated dataset into the original visual instruction tuning datasets of LLaVA-1.5 [41], MGM [37], and InternVL2 [12], and conduct fine-tuning. Subsequently, we evaluate their performance on image difference benchmarks, including MMVP [63], Spot-the-Diff [26], and Image-Edit-Request [62], as well as widely recognized MLLM benchmarks. Our evaluation results reveal that after fine-tuning with our IMG-DIFF dataset, the MLLMs achieve notable enhancements in image difference benchmarks, aligning their performance with state-of-the-art (SOTA) models. For instance, they notably surpass the SOTA models GPT-4V [1] and Gemini [16] on the MMVP benchmark, achieving an improvement of up to 12. Moreover, the models exhibit comprehensive improvements across eight well-recognized MLLM benchmarks, achieving an average score improvement of up to $3.06\%$ , underscoring the useful role our dataset plays in bolstering MLLMs' competencies in both image difference recognition and fine-grained image analysis. + +We further evaluate the diversity and quality of our dataset, ensuring it encompasses a broad array of object categories while showcasing rich variability. Through meticulous manual labeling, we affirm the high quality of our dataset. Furthermore, we conduct ablation studies to examine the effects of various filtering intensities. We also investigate an alternative methodology for constructing image difference data focusing on "object removal", assessing its effectiveness and presenting fruitful insights on the construction of contrastive data. + +Our contributions are summarized as follows: + +- We present a novel data synthesis method and an effect-proven IMG-DIFF dataset, comprising pairs of highly similar images, with a focus on processes such as segmentation, filtering, and detailed captioning of image differences. + +- We conduct visual instruction tuning on LLaVA-1.5-7B, MGM-7B, and InternVL2-8B using our dataset, and rigorously assess the fine-tuned models' performance on many widely-used MLLM benchmarks and image difference benchmarks. Our dataset brings substantial performance improvements to the fine-tuned MLLMs. +- We provide a comprehensive evaluation of the diversity and quality of our dataset, confirming its richness and high standard. Through ablation studies, we identify good empirical filtering thresholds for such contrastive dataset. +- We open-source our dataset and codes at https://github.com/modelscope/data-juicer/tree/ImgDiff to facilitate ongoing research, encouraging innovative endeavors in MLLM datasets and image difference methods. + +# 2. Background and Related Works + +# 2.1. Multimodal Large Language Models + +Multimodal Large Language Models (MLLMs) have exhibited remarkable advancements since their introduction [1, 16, 37, 43, 44]. Research highlights two key factors that primarily influence the effectiveness of MLLMs: model architecture and dataset quality [53]. + +With respect to model architecture, notable approaches [2, 3, 25, 32, 35] leverage learnable queries to extract essential information from CLIP [15, 54] image features. Alternatively, LLaVA [41-43] and MGM [37] utilize projection-based interfaces to facilitate interactions between text and image modalities. Furthermore, LLaMA-Adapter [81] and LaVIN [47] implement parameter-efficient tuning mechanisms to transfer image-related information to the LLM. + +From the perspective of datasets, there are two prevalent strategies: enhancing the quality of pre-training data and improving instruction tuning data. The former aims for better semantic alignment between images and text by introducing abundant image-text pairs, enhancing MLLMs' fundamental capabilities such as image captioning. As for the latter, research has increasingly concentrated on refining instruction tuning datasets, boosting performance across various question-answering tasks. Works like LLaVA [41-43], SPHINX [40], MGM [37], and InternVL2 [12] leverage high-quality fine-tuning datasets with extensive task diversity, allowing models to excel in tasks related to image perception, reasoning, and optical character recognition. Additionally, methods such as Shikra [10], ASM [72], and PINK [75] utilize substantial amounts of object detection data to enhance the models' localization capabilities. + +In contrast to previous works, our research introduces a fully automated data synthesis method and generates a dataset that emphasizes image differences, showing empirical effectiveness and great potential to augment MLLMs' VQA proficiency, object localization capabilities, and discernment of image distinctions. + +# 2.2. Image Difference Datasets + +Datasets focused on image differences typically consist of pairs of similar images and textual descriptions that describe the variations. For instance, the Spot-the-Diff dataset [26] contains pairs of street scenes captured at different times by the same surveillance cameras. The CLEVR-Change dataset [51] delineates scene variations of geometric objects against a clean backdrop. The CUB-Bird dataset [69] focuses on the nuanced differences among various bird species found in natural habitats. The Image-Edit-Request [62] dataset features edited images, accompanied by descriptions of modifications made. + +Leveraging advancements in image editing technologies, some studies have employed generative models and editing techniques to create datasets centered on image differences. A prime example is InstructPix2Pix [6], which utilizes the image editing technique, named Prompt-to-Prompt [20], to direct Stable-Diffusion-1.5 [56] in generating pairs of similar images, while employing GPT-3 [48] to craft the edited text as reference captions for image differences. + +Our approach, referring to InstructPix2Pix, employs the Prompt-to-Prompt technique alongside an advanced generative model Stable-Diffusion-XL [52], which produces more realistic images, to generate pairs of similar images. Unlike InstructPix2Pix, we incorporate multiple filtering stages to ensure data quality, with a particular emphasis on producing difference captions that focus on specific regions rather than the entire image, which ensures greater accuracy. + +# 2.3. Models for Image Difference Captioning + +Image Difference Captioning (IDC) represents a specialized domain within image captioning characterized by its focus on subtle variations between images. As for the pioneering work in IDC, Spot-the-Diff [26] presents potential change clusters and employs an LSTM [21] decoder to generate difference captions. DUDA [51] explores image differences at the semantic level, using a ResNet [19] and an LSTM to compute dynamic attention weights and produce captions. VARD [65] introduces a viewpoint-adaptive representation disentanglement network based on LSTM for differentiating between real and pseudo changes. Meanwhile, NCT [66] employs a transformer [67] to integrate neighboring features, and CLIP4IDC [18] uses BERT-like training methodologies, adapting a CLIP model for IDC tasks. With the emergence of MLLMs, VIXEN [5] marks the inaugural use of these models for IDC tasks, mapping the features of differential images to text space and employing an LLM to generate image difference captions. + +Our data synthesis method is specifically designed for MLLMs. We build our dataset following the instruction-following format adopted by mainstream MLLMs such as LLaVA-1.5, highlighting a novel research direction aimed at enhancing MLLMs from a data-centric perspective. + +![](images/5f866f61d2fdbb9e67e73bd315297c94cfc646f0b7df5665b9409c11b81e43ed.jpg) +Figure 2. The generation process for "object replacement" data. + +![](images/f22d289ae7b9c2a60939dbe1f1c322036d19fd9f5927a6817f0c6b2ddea4f0fd.jpg) + +![](images/6fd31c0668d5179748c746ee6fb306bac90a5de3b8a46dca09a7e4865f9ff4e6.jpg) + +# 3. Methodology + +# 3.1. Overview + +In recent years, contrastive learning have significantly enhanced the image-text understanding of vision-language models [54, 80]. These methods typically involve constructing batches of images and texts, requiring the model to distinguish between matching and non-matching image-text pairs, which improves its ability to differentiate between semantically similar and dissimilar pairs. + +Our method leverages the principles of contrastive learning to generate MLLM image-text data. Specifically, it focuses on replacing objects within image pairs, encouraging MLLMs to identify similarities and differences in specific regions. This method aims to enhance models' capacity to recognize fine-grained differences in images, guided by textual descriptions that highlight detailed distinctions. + +As shown in Figure 2, the process of generating "object replacement" data can be divided into three key parts. The first part involves creating similar images and forming image pairs, where the only difference between the images in pairs is the objects replaced (Section 3.2). The second part, named the "Difference Area Generator", extracts bounding box regions that contain object differences between the images in pairs (Section 3.3). The third part, termed as the "Difference Captions Generator" (Section 3.4), utilizes an MLLM to generate descriptive text for the areas with object differences and creates question-answer pairs. + +This process incorporates multiple filtering operations, with specific thresholds outlined in Section 18.4, which we determine through experimental comparisons in Section 17. Additional experimental details, including model selection and the time consumption, are also presented in Section 18. For readability, we present data examples in Section 20. + +To enhance reproducibility and reusability, our proposed components and end-to-end construction workflow are implemented as data processing operators and configurable data recipes within Data-Juicer [8, 9]. + +![](images/d7b9448d1919c7579c5ddd08be646025e5255f3d6ba1a7c249b8f7536c95769b.jpg) +Figure 3. An overview of the Difference Area Generator and its three internal components: Image Similarity Filter, Image-text Matching Filter, and Difference Detector. + +# 3.2. Image Pairs Generation + +The first step of our data synthesis method is to generate pairs of similar images as candidates. The process is shown in Figure 2. We employ a generative model called Stable-Diffusion-XL [52] and an image editing technique called Prompt-to-Prompt [20] to generate image pairs that highlight object replacement. + +We start by obtaining image captions from caption databases, which contain descriptions biased towards real photos. Then, we use an LLM to perform object replacement in the captions. The prompt used is "Here is a sentence: 'INPUT'. Please only replace one of the objects in this sentence with another object." Here, INPUT refers to the original caption. Next, based on the caption pairs, we use the Stable-Diffusion-XL and the Prompt-to-Prompt to generate image pairs with only few objects replaced. + +# 3.3. Difference Area Generator + +# 3.3.1. Overview + +The Difference Area Generator is used to identify the locations of object differences between the two images in pairs. Although object detection models are capable of detecting objects in images, the range of object categories is quite limited [27, 78]. Therefore, to increase the number of detectable object categories and enhance data diversity, we develop the Difference Area Generator based on segmentation and image similarity comparisons. + +As shown in Figure 3, we first use an Image Similarity Filter to obtain image pairs with high similarity but not identical. Next, we use the FastSAM [82] to perform image segmentation on each image. After that, we crop the images based on the bounding box information obtained from segmentation and use an Image-text Matching Filter to filter the cropped sub-images for the presence of valid objects. + +Finally, we use a Difference Detector to determine whether there are indeed differences between the bounding box regions of the two images and perform IoU (Intersection over Union) filtering to remove overlapping bounding boxes, ultimately obtaining valid bounding box information. + +# 3.3.2. Image Similarity Filter + +The Image Similarity Filter aims to filter image pairs based on the degree of similarity. The module first uses CLIP [54] to extract image features from each image in pairs and then calculates the cosine similarity score. If their score falls within the pre-set threshold, the image pair will be considered valid. Specifically, we use the module twice in the Difference Area Generator. Before the segmentation, we use the module to ensure that the images in pairs are highly similar but not the same. In the Difference Detector stage, after cropping, we use the module to filter the sub-image pairs and keep only the differing ones. + +# 3.3.3. Image-text Matching Filter + +The Image-text Matching Filter determines whether an image contains valid objects (i.e. the replaced or replacing objects). This module first uses BLIP [34] to extract image features, which are then compared with text features of object names. When the image-text matching score falls within the pre-set threshold, we consider the image to contain valid objects. In the mid-stage of the Difference Area Generator, after cropping, we use the module to determine whether these sub-images contain valid objects and get valid bounding boxes. + +# 3.3.4. Difference Detector + +The Difference Detector determines whether there are differences between the two regions in the image pair of the same bounding box. Based on a given bounding box, we + +![](images/ef3f77d87c470c1d2fc964538b69a793bc79ef4895898ec4dcae01b1a10f849d.jpg) +Figure 4. An overview of the Difference Captions Generator and its two stages. + +first crop two sub-images from both image A and image B. The two sub-images are then filtered through the Image Similarity Filter and the bounding box is considered effective only if the difference is significant enough. After processing all bounding boxes, we use the IoU method to filter out overlapping bounding boxes. Only the bounding boxes with a higher degree of difference are retained. + +# 3.4. Difference Captions Generator + +# 3.4.1. Overview + +After obtaining the valid bounding box regions, we use the Difference Captions Generator to generate descriptions for the differences inside these areas (with each round of the process focusing on only one bounding box in one image pair). Evidently, an image pair can contain multiple object differences but a single difference caption cannot fully capture all of them. Therefore, we highlight specific regions with red boxes and provide targeted difference captions to ensure greater accuracy. + +The module consists of two stages: the first stage generates captions for the content in the bounding box regions and then filters the bounding boxes using an Image-text Matching Filter and a Captions Similarity Filter. The second stage uses the content captions and the images highlighted with red boxes to generate difference captions. The overview is shown in Figure 4. + +# 3.4.2. Stage1: Object Labeling & Filtering + +In Stage 1, for each image pair, we first select N bounding box regions with the lowest similarity between images (N is set to 5 in this project) as candidate regions. Then, for each bounding box, we use the MLLM LLaVA-NEXT [43] to describe its corresponding regions and then apply two filtering processes: the first filter is an Image-text Matching Filter, which checks whether the content of the regions matches the captions; the second filter is an Captions Similarity Filter, which assesses whether there are differences between the two captions. Once the filtering is complete, we obtain valid bounding boxes and captions for subsequent difference captioning. + +# 3.4.3. Captions Similarity Filter + +The Captions Similarity Filter determines whether the two captions of the same bounding box coordinates are different. We use CLIP to obtain text features and calculate the cosine similarity between them. When the score is low enough, we consider the two captions to be different. + +# 3.4.4. Stage2: Difference Captions Generating + +In Stage 2, for each valid bounding box of each image pair, we first draw two red boxes into the images based on the bounding box coordinates, highlighting the differences for easier localization. Then, we provide the MLLM LLaVA-NEXT with the captions of the bounding box content and instruct it to describe the difference based on these captions and the highlighted images. Finally, we can obtain the difference caption for the highlighted area. + +# 3.5. Data Statistics + +Using captions from MSCOCO [39], we generate 118K pairs of similar images. We then employ the Image Similarity Filter to get 38,533 highly similar but not identical image pairs. Next, we use the Difference Area Generator to filter and produce 117,779 pieces of valid bounding box information (with a maximum of 5 valid bounding boxes per image pair). Finally, we employ the Difference Captions Generator to filter and generate 12,688 high-quality "object replacement" instances. Our evaluations on the main page are based on this dataset. + +In addition, we also use captions from the LLaVA pretraining dataset to generate 34,538 "object replacement" samples. In Section 10 of the appendix, we evaluate this dataset and discuss the relationship between data quantity, data quality, and model performance improvement, emphasizing the importance of prioritizing quality over quantity. + +# 4. Evaluation and Main Results + +# 4.1. Evaluation Settings + +To evaluate the effectiveness of our IMG-DIFF dataset, we use it to fine-tune SOTA MLLMs, including LLaVA-1.5- + +7B, MGM-7B, and InternVL2-8B on the main page, as well as InternVL2-1B and LLaVA-1.5-13B in the supplementary materials. We evaluate these models on extensive benchmarks commonly used for image difference and MLLMs. Specifically, the image difference benchmarks include MMVP [63], Spot-the-Diff [26], and Image-Edit-Request [62]. Besides, the details and results of the MLLM benchmarks are shown in Section 4.5. + +During our fine-tuning process, we first mix our data with MLLMs' original visual instruction tuning data respectively. For LLaVA-1.5 and MGM, we conduct the fine-tuning anew. For InternVL2, we follow the guidelines from its official repository and perform a second fine-tuning. (To ensure a fair comparison, we perform a second fine-tuning on InternVL2 using its original fine-tuning dataset, termed as InternVL2-FT, and use it as a baseline.) Regarding the Spot-the-Diff and Image-Edit-Request benchmarks, as they contain data of training splits, we further fine-tune the fine-tuned MLLMs for an additional 2 epochs using only these benchmarks' training data. + +In the tables, "RP" represents "object replacement" data. + +# 4.2. Results on the MMVP Benchmark + +The MMVP benchmark is designed to systematically assess the visual capabilities of MLLMs. Its data processing method is highly related to image difference: it first collects CLIP-blind pairs, which have similar CLIP features but differ in image content. Then, the differences between the images are manually described and question-answer pairs are created. Hence, the questions in MMVP is highly relevant to our dataset, as both place significant emphasis on the differences between similar images. + +![](images/de86155551b0c8394b3cbcbdb98208bafb2bec1b8b785a0bd985e99c09c58452.jpg) +Figure 5. Performance comparison on the MMVP benchmark. + +As shown in Figure 5, fine-tuning MLLMs with our "object replacement" data significantly improves their performance on the MMVP benchmark. After fine-tuning, the score of LLaVA-1.5-7B exceeds that of LLaVA-1.5-13B. Furthermore, the fine-tuned MGM-7B shows a significant improvement in score compared to the original MGM-7B, even surpassing the scores of the SOTA models GPT-4V and Gemini by up to 12. Additionally, the performance of + +InternVL2-8B is also enhanced. These results suggest that our dataset enhances MLLMs' abilities to distinguish images with similar CLIP features but different content. + +# 4.3. Results on the Spot-the-Diff Benchmark + +The dataset of Spot-the-Diff comprises pairs of street view images that display subtle object differences. These images are obtained by capturing scenes from fixed surveillance cameras at different time. + +Table 1. Performance comparison on Spot-the-Diff. + +
ModelBLEUMETEORCIDEr-DROUGE-L
VAM[59]10.112.438.131.3
IFDC[23]8.711.73730.2
DUDA+Aux[22]8.112.534.529.9
VACC[29]9.712.641.532.1
LLaVA-1.5-7B8.512.038.330.1
LLaVA-1.5-7B + RP9.713.043.230.8
MGM-7B9.912.046.331.5
MGM-7B + RP10.813.153.533.0
InternVL2-8B-FT6.611.726.527.3
InternVL2-8B + RP8.412.832.228.5
+ +Following previous works, our fine-tuned MLLMs are evaluated on BLEU [50], METEOR [4], CIDEr-D [68] and ROUGE-L [38]. As shown in Table 1, after fine-tuning with our "object replacement" data, both LLaVA-1.5-7B, MGM-7B, and InternVL2-8B achieve significant performance gains on Spot-the-Diff. Compared to score increases seen in prior models, our dataset yields even greater enhancements than those from previous iterations of image difference models. These positive results indicate that our dataset can enhance the ability of MLLMs to detect subtle differences between similar images. + +# 4.4. Results on Image-Editing-Request + +The Image-Editing-Request benchmark focuses on image editing. Each instance in its dataset consists of an image pair (i.e. a source image and an edited image) and an editing instruction that describes the transformation. During the evaluation, our models are required to generate transformation descriptions for these image pairs, and we then calculate the BLEU, METEOR, CIDEr-D, and ROUGE-L scores using the models' responses and the reference answers. + +As shown in Table 2, LLaVA-1.5-7B, MGM-7B, and InternVL2-8B originally show SOTA performance on the Image-Edit-Request benchmark. Upon incorporating our "object replacement" data for better fine-tuning, the performance of all three models improves even more, achieving new SOTA scores. This suggests that our dataset enhances MLLMs' abilities to recognize similarities and dissimilarities in image pairs, as well as enables them to describe differences more accurately. + +Table 2. Performance comparison on Image-Edit-Request. + +
ModelBLEUMETEORCIDEr-DROUGE-L
VARD[65]1014.835.739
CLIP4IDC[18]8.214.632.240.4
NCT[66]8.11534.238.8
BiDiff[61]6.914.627.738.5
VIXEN[5]8.615.438.142.5
LLaVA-1.5-7B15.117.860.645.2
LLaVA-1.5-7B + RP16.219.560.946.7
MGM-7B16.517.766.844.8
MGM-7B + RP16.618.268.145.7
InternVL2-8B-FT12.414.151.538.9
InternVL2-8B + RP12.514.256.039.4
+ +# 4.5. Results on MLLM Benchmarks + +Aside from the evaluations related to image difference discrimination, we also assess the performance of our "object replacement" data in enhancing the comprehensive abilities of MLLMs. We test the fine-tuned MLLMs using commonly used MLLM benchmarks, including VQAv2 [17] and GQA [24] for assessing the comprehensive VQA capabilities; MMBench [45], MM-Vet [76], ScienceQA [46], and SEED-Bench [33] for testing perceptual and reasoning abilities; and POPE [36] for evaluating fine-grained object localization abilities. Table 3 presents the results on these MLLM benchmarks, with the $\triangle$ metric indicating the percentage improvement averaged across them. + +Table 3. Performance comparison on 8 MLLM benchmarks. + +
ModelVQAv2GQAPOPEMMB\( {\mathrm{{MMB}}}^{CN} \)
LLaVA-1.5-7B78.56285.964.358.3
LLaVA-1.5-7B + RP79.3↑62.8↑86.4↑66.1↑59.8↑
MGM-7B80.462.68669.358.9
MGM-7B + RP80.7↑62.7↑86.2↑68.7↓59.6↑
InternVL2-8B-FT81.862.687.782.581.5
InternVL2-8B + RP81.862.688.0↑82.7↑81.4↓
+ +
ModelMM-Vet\( \mathbf{SQA}^I \)SEED\( \bigtriangleup \)
LLaVA-1.5-7B30.566.858.6-
LLaVA-1.5-7B + RP33.2↑68.2↑61.7↑+3.06%
MGM-7B40.870.663.5-
MGM-7B + RP44.1↑71.7↑63.2↓+1.28%
InternVL2-8B-FT49.296.569.5-
InternVL2-8B + RP52.6↑96.6↑69.9↑+1.01%
+ +Based on Table 3, after fine-tuning with our dataset, the performance of LLaVA-1.5-7B shows a comprehensive improvement, with an average increase of $3.06\%$ across all benchmarks. For MGM-7B and InternVL2-8B, the improvements brought by our dataset are not as pronounced as those observed with LLaVA-1.5-7B, as their training datasets already encompass a large volume of high-quality data, but they still achieve an average increase of $1.28\%$ + +and $1.01\%$ . These improvements indicate that fine-tuning MLLMs with our dataset not only enhances their ability to discern differences but also improves their overall visual capabilities, thereby making them better address VQA tasks. + +# 5. Evaluation of Data Quality and Diversity + +# 5.1. Data Quality + +To assess the quality of our Img-Diff dataset, we randomly select 1,000 instances of "object replacement" data and employ multiple professional dataset annotators to evaluate the samples based on three metrics. The final scores are determined through a voting process. Specifically, the first metric is "Bounding Box Difference", which evaluates whether there are differences between the two highlighted regions in pairs. If the objects are different, we score it as "high"; if the objects are the same but their features (such as color, shape, etc.) are different, we score it as "medium"; if the objects are the same and their features are similar, we score it as "low". The second metric is "Content Caption Accuracy", which evaluates whether the captions generated by Stage 1 of the Difference Captions Generator accurately describe the two highlighted regions. If the captions are correct, we score it as "high"; if the captions identify the objects but incorrectly describe their features, we score it as "medium"; if the captions incorrectly identify the objects, we score it as "low". The third metric is "Difference Caption Accuracy", which evaluates whether the final difference captions accurately describe the object differences between highlighted regions of the image pairs. If the description is accurate, we score it as "high"; if it recognizes the objects but the feature description is incorrect, we score it as "medium"; if the object recognition is incorrect, we score it as "low". The results are shown in Figure 6. + +![](images/86675cd21f5a495339652134afff76ea4bc8da4341a688d0288cb73423d3c9b8.jpg) +Figure 6. Quality assessment of the "object replacement" data. + +Based on Figure 6, our dataset demonstrates a high level of quality. For the "Bounding Box Difference" metric, only $4.5\%$ of the instances are classified as "low", and nearly $80\%$ of instances exhibit valid object differences between their two highlighted regions. In terms of "Content Caption Accuracy", $80.1\%$ of highlighted region pairs are described accurately, indicating that using an MLLM for labeling is effective and that our filtering strategy is also functioning well. For the "Difference Caption Accuracy" metric, over $70\%$ of the difference descriptions are completely accurate, with $21.8\%$ of the samples having errors solely in feature labeling while still maintaining correct descriptions of ob- + +ject differences, which underscores the effectiveness of our difference caption generation strategy. + +# 5.2. Data Diversity + +In designing the generation process, we have made efforts to enhance the diversity of our dataset, which can be divided into two aspects: (a) the inherent diversity of the caption databases; and (b) the introduction of new object names through our object replacement strategy. As for the former, we are using the captions from MS COCO and the LLaVA pre-training dataset, which cover a majority of object categories. We can easily further enhance the data diversity by employing caption databases with greater variability. As for the latter, we employ two methods, including increasing the temperature of the LLM used for noun replacement, as well as randomly replacing object names with nouns from a noun lexicon (shown in Appendix, Section 11). We can further improve data diversity by expanding the noun lexicon. + +By analyzing the valid object names included in the captions that describe highlighted regions, we assess the diversity of our "object replacement" data. Specifically, we count the total number of object categories covered, and the total number of unique "object replacement pairs". Through our statistical analysis, we find that our dataset covers 1,203 object categories, which encompasses most real-life objects. An "object replacement pair" refers to the combination of the replaced and replacing object names. Our dataset includes 3,680 unique "object replacement pairs". + +To evaluate the coverage of common object categories in our dataset, we analyze the occurrences of the object names from the Object365 [58] dataset in our own dataset. The results show that the object names from the Object365 dataset appears 13,164 times in total, which accounts for approximately $52.06\%$ of the total occurrences of object names. + +These statistics show that our data covers a substantial number of common object names, ensuring a high frequency of common objects. Additionally, less common object names make up nearly half of our dataset, indicating that our dataset is both diverse and comprehensive. + +# 6. More Empirical Supports and Details + +We list additional experimental details and extensions of our data synthesis method in the supplementary material, and conduct further experiments to validate the effectiveness of our IMG-DIFF dataset, along with its limitations. + +Further Experiments and Analysis. In Section 9, we compare our IMG-DIFF dataset with existing image difference datasets on characteristics and performance, validating its superiority. In Section 10, we discuss the relationship between the data volume and model performance improvement. In Section 11, we explore the use of lexicons for + +object replacement and validate its effectiveness in enhancing diversity. In Section 12, we evaluate our dataset using the Contrastive Chain-of-Thought method. In Section 13, we validate the effectiveness of our dataset on MLLMs of different scales. In Section 14, we evaluate top-performing MLLMs on difference detection. In Section 15, we assess the impact of unnatural images. In Section 16, we show our data's efficacy in enhancing spatial reasoning capabilities. + +Implementation Details. In Section 17, we investigate the effect of different filtering intensities on the performance of the generated datasets. In Section 18, we present additional experimental details, including image pair preprocessing, model training procedures, model selection, filtering thresholds, and time consumption. + +The "Object Removal" Exploration. In Section 19, we generate an extended dataset that focuses on object removal, which prompts MLLMs to analyze which image contains a specific object. The new data brings further improvement to the fine-tuned MLLM. + +# 7. Conclusion + +In this paper, we draw inspiration from recent advances in contrastive learning and image difference captioning to propose a novel method of contrastive data synthesis. This method enables the creation of a high-quality dataset called "IMG-DIFF", which highlights object differences focusing on fine-grained regions in images. Specifically, we first generate pairs of similar images where the main focus is on object replacement. Then, we use the proposed Difference Area Generator and Difference Captions Generator to generate difference captions and form question-answer pairs. In contrast to previous image difference datasets, our dataset focuses exclusively on specific regions inside images. This characteristic circumvents the issue where a single description cannot fully capture the differences of the whole images in pairs, enhancing the accuracy. Afterwards, we fine-tune many MLLMs using the generated dataset, yielding high-performance scores on par with SOTA models in image difference tasks and comprehensive performance improvements in eight widely recognized MLLM benchmarks. These results confirm the effectiveness of our dataset to improve the ability of MLLMs in recognizing detailed differences between images. + +In a nutshell, we provide a series of insights about the construction of high-quality image difference datasets, showing great potential to effectively and efficiently enhance MLLMs via contrastive data-centric approaches. With this work, we hope it can catalyze further investigation into the realm of image difference datasets and the fine-grained image recognition capabilities of MLLMs. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2 +[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6 +[5] Alexander Black, Jing Shi, Yifei Fan, Tu Bui, and John Colomosse. Vixen: Visual text comparison network for image difference captioning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 846-854, 2024. 2, 3, 7 +[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3, 1, 2 +[7] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, Shuaibin Li, Wei Li, Yinling Li, Hongwei Liu, Jiangning Liu, Jiawei Hong, Kaiwen Liu, Kuikun Liu, Xiaoran Liu, Chengqi Lv, Haijun Lv, Kai Lv, Li Ma, Runyuan Ma, Zerun Ma, Wenchang Ning, Linke Ouyang, Jiantao Qiu, Yuan Qu, Fukai Shang, Yunfan Shao, Demin Song, Zifan Song, Zhihao Sui, Peng Sun, Yu Sun, Huanze Tang, Bin Wang, Guoteng Wang, Jiaqi Wang, Jiayu Wang, Rui Wang, Yudong Wang, Ziyi Wang, Xingjian Wei, Qizhen Weng, Fan Wu, Yingtong Xiong, Chao Xu, Ruiliang Xu, Hang Yan, Yirong Yan, Xiaogui Yang, Haochen Ye, Huaiyuan Ying, Jia Yu, Jing Yu, Yuhang Zang, Chuyu Zhang, Li Zhang, Pan Zhang, Peng Zhang, Ruijie Zhang, Shuo Zhang, Songyang Zhang, Wenjian Zhang, Wenwei Zhang, Xingcheng Zhang, Xinyue Zhang, Hui Zhao, Qian Zhao, Xiaomeng Zhao, Fengzhe Zhou, Zaida Zhou, Jingming Zhuo, Yicheng Zou, Xipeng Qiu, Yu Qiao and Dahua Lin. Internl m2 technical report, 2024. 6 +[8] Daoyuan Chen, Yilun Huang, Zhijian Ma, Hesen Chen, Xuchen Pan, Ce Ge, Dawei Gao, Yuexiang Xie, Zhaoyang + +Liu, Jinyang Gao, et al. Data-juicer: A one-stop data processing system for large language models. In Companion of the 2024 International Conference on Management of Data, pages 120-134, 2024. 3 +[9] Daoyuan Chen, Haibin Wang, Yilun Huang, Ce Ge, Yaliang Li, Bolin Ding, and Jingren Zhou. Data-juicer sandbox: A feedback-driven suite for multimodal data-model codevelopment. arXiv preprint arXiv:2407.11784, 2024. 3 +[10] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2 +[11] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Qinglan Huang, Canyu Chen, Qinghao Ye, Zhihong Zhu, Yuqing Zhang, Jiawei Zhou, Zhuokai Zhao, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Mj-bench: Is your multimodal reward model really a good judge for text-to-image generation?, 2024. 1, 2 +[12] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2, 3 +[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023.8 +[14] W Dai, J Li, D Li, AMH Tiong, J Zhao, W Wang, B Li, P Fung, and S Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 3 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[16] Google. Gemini. 2023. 2 +[17] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 7 +[18] Zixin Guo, Tzu-Jui Julius Wang, and Jorma Laaksonen. Clip4idc: Clip for image difference captioning. arXiv preprint arXiv:2206.00629, 2022. 3, 7 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[20] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3, 4 + +[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 3 +[22] Mehrdad Hosseinzadeh and Yang Wang. Image change captioning by learning from an auxiliary task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2725-2734, 2021. 6 +[23] Qingbao Huang, Yu Liang, Jielong Wei, Yi Cai, Hanyu Liang, Ho-fung Leung, and Qing Li. Image difference captioning with instance-level fine-grained feature representation. IEEE transactions on multimedia, 24:2004-2017, 2021. 6 +[24] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 7 +[25] Laurençon Hugo, van Strien Daniel, Bekman Stas, Tronchon Leo, Saulnier Lucile, Wang Thomas, Karamcheti Siddharth, Singh Amanpreet, Pistilli Giada, Jernite Yacine, and Sanh Victor. Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics, 2023.2 +[26] Harsh Jhamtani and Taylor Berg-Kirkpatrick. Learning to describe differences between pairs of similar images. arXiv preprint arXiv:1808.10584, 2018. 2, 3, 6, 1 +[27] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. *Ultralytics YOLO*, 2023. 4 +[28] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 1 +[29] Hoeseong Kim, Jongseok Kim, Hyungseok Lee, Hyunsung Park, and Gunhee Kim. Agnostic change captioning with cycle consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2095-2104, 2021. 6 +[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 8 +[31] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1 +[32] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 2, 3 +[33] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 7 +[34] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified + +vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4 +[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2 +[36] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 7 +[37] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv:2403.18814, 2023. 1, 2, 3 +[38] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6 +[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[40] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 2 +[41] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2 +[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. +[43] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,2, 5,3 +[44] Mengsha Liu, Daoyuan Chen, Yaliang Li, Guian Fang, and Ying Shen. ChartThinker: A contextual chain-of-thought approach to optimized chart summarization. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 3057-3074, 2024. 2 +[45] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 7 +[46] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 7 + +[47] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. arXiv preprint arXiv:2305.15023, 2023. 2 +[48] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020. 1, 3 +[49] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 1 +[50] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6 +[51] Dong Huk Park, Trevor Darrell, and Anna Rohrbach. Robust change captioning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4624-4633, 2019. 2, 3 +[52] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 4 +[53] Zhen Qin, Daoyuan Chen, Wenhao Zhang, Liuyi Yao, Yilun Huang, Bolin Ding, Yaliang Li, and Shuiguang Deng. The synergy between data and multi-modal large language models: A survey from co-development perspective. arXiv preprint arXiv:2407.08583, 2024. 1, 2 +[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4 +[55] Kanchana Ranasinghe, Satya Narayan Shukla, Omid Poursaeed, Michael S Ryoo, and Tsung-Yu Lin. Learning to localize objects improves spatial reasoning in visual-llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12977–12987, 2024. 6 +[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3 +[57] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 9, 10 +[58] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 8 +[59] Xiangxi Shi, Xu Yang, Jiuming Gu, Shafiq Joty, and Jianfei Cai. Finding it at another side: A viewpoint-adapted match + +ing encoder for change captioning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 574-590. Springer, 2020. 6 +[60] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 1 +[61] Yaoqi Sun, Liang Li, Tingting Yao, Tongyv Lu, Bolun Zheng, Chenggang Yan, Hua Zhang, Yongjun Bao, Guiguang Ding, and Gregory Slabaugh. Bidirectional difference locating and semantic consistency reasoning for change captioning. International Journal of Intelligent Systems, 37 (5):2969-2987, 2022. 7 +[62] Hao Tan, Franck Dernoncourt, Zhe Lin, Trung Bui, and Mohit Bansal. Expressing visual relationships via language. arXiv preprint arXiv:1906.07689, 2019. 2, 3, 6, 1 +[63] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2, 6 +[64] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[65] Yunbin Tu, Liang Li, Li Su, Junping Du, Ke Lu, and Qingming Huang. Adaptive representation disentanglement network for change captioning. IEEE Transactions on Image Processing, 32:2620-2635, 2023. 3, 7 +[66] Yunbin Tu, Liang Li, Li Su, Ke Lu, and Qingming Huang. Neighborhood contrastive transformer for change captioning. IEEE Transactions on Multimedia, 25:9518-9529, 2023. 3, 7 +[67] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[68] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6 +[69] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 3, 1, 2 +[70] Jiayu Wang, Yifei Ming, Zhenmei Shi, Vibhav Vineet, Xin Wang, Yixuan Li, and Neel Joshi. Is a picture worth a thousand words? delving into spatial reasoning for vision language models. In The Thirty-Eighth Annual Conference on Neural Information Processing Systems, 2024. 7 +[71] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's + +perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6 +[72] Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, et al. The all-seeing project: Towards panoptic visual recognition and understanding of the open world. In The Twelfth International Conference on Learning Representations, 2023. 2 +[73] Weiyun Wang, Zhe Chen, Wenhai Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Jinguo Zhu, Xizhou Zhu, Lewei Lu, Yu Qiao, and Jifeng Dai. Enhancing the reasoning ability of multimodal large language models via mixed preference optimization. arXiv preprint arXiv:2411.10442, 2024.6 +[74] Zhe Xu, Daoyuan Chen, Zhenqing Ling, Yaliang Li, and Ying Shen. Mindgym: Enhancing vision-language models via synthetic self-challenging questions, 2025. 3 +[75] Shiyu Xuan, Qingpei Guo, Ming Yang, and Shiliang Zhang. Pink: Unveiling the power of referential comprehension for multi-modal llms. arXiv preprint arXiv:2310.00582, 2023. 2 +[76] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 7 +[77] Daoan Zhang, Junming Yang, Hanjia Lyu, Zijian Jin, Yuan Yao, Mingkai Chen, and Jiebo Luo. Cocot: Contrastive chain-of-thought prompting for large multimodal models with multiple image inputs. arXiv preprint arXiv:2401.02582, 2024. 5 +[78] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 4 +[79] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 1, 2 +[80] Qi Zhang, Yifei Wang, and Yisen Wang. On the generalization of multi-modal contrastive learning. In International Conference on Machine Learning, pages 41677-41693. PMLR, 2023. 2, 3 +[81] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 2 +[82] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 4 +[83] Ting Zhou, Daoyuan Chen, Qirui Jiao, Bolin Ding, Yaliang Li, and Ying Shen. Humanvbench: Exploring human-centric video understanding capabilities of mllms with synthetic benchmark data, 2025. 3 \ No newline at end of file diff --git a/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/images.zip b/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ddc40337c1f9a2304d204e391f9ee65da43aa751 --- /dev/null +++ b/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de5fcbe0e32131c7f6383c3093866d137ad2ce49237e5c7885fa1bcc3663bf19 +size 366373 diff --git a/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/layout.json b/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..284f27d7c0570f05ff3ba9df4e19a7fed79c879f --- /dev/null +++ b/2025/Img-Diff_ Contrastive Data Synthesis for Multimodal Large Language Models/layout.json @@ -0,0 +1,8567 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 547, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 547, + 122 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 547, + 122 + ], + "type": "text", + "content": "Img-Diff: Contrastive Data Synthesis for Multimodal Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "spans": [ + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "type": "text", + "content": "Qirui Jiao" + }, + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "type": "text", + "content": ", Daoyuan Chen" + }, + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "type": "text", + "content": ", Yilun Huang" + }, + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "type": "text", + "content": ", Bolin Ding" + }, + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "type": "text", + "content": ", Yaliang Li" + }, + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "type": "text", + "content": ", Ying Shen" + }, + { + "bbox": [ + 102, + 142, + 506, + 159 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 179, + 171, + 432, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 171, + 432, + 186 + ], + "spans": [ + { + "bbox": [ + 179, + 171, + 432, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 179, + 171, + 432, + 186 + ], + "type": "text", + "content": "Sun Yat-Sen University, " + }, + { + "bbox": [ + 179, + 171, + 432, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 179, + 171, + 432, + 186 + ], + "type": "text", + "content": "Alibaba Group, " + }, + { + "bbox": [ + 179, + 171, + 432, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 179, + 171, + 432, + 186 + ], + "type": "text", + "content": "FSIETP" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 187, + 484, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 187, + 484, + 213 + ], + "spans": [ + { + "bbox": [ + 126, + 187, + 484, + 213 + ], + "type": "text", + "content": "jiaoqr3@mail2.sysu.edu.cn, sheny76@mail.sysu.edu.cn {daoyuanchen.cdy, lielin.hyl, bolin.ding, yaliang.li}@alibaba-inc.com" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 62, + 241, + 218, + 320 + ], + "blocks": [ + { + "bbox": [ + 62, + 241, + 218, + 320 + ], + "lines": [ + { + "bbox": [ + 62, + 241, + 218, + 320 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 218, + 320 + ], + "type": "image", + "image_path": "95f8a8be447b0fda9939655bd20481218816e7c7b8b3c5f0dd4b7fd24be2e195.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 322, + 203, + 348 + ], + "lines": [ + { + "bbox": [ + 71, + 322, + 203, + 348 + ], + "spans": [ + { + "bbox": [ + 71, + 322, + 203, + 348 + ], + "type": "text", + "content": "... the left image highlights a fish on the ground, while the right image highlights a red crab on the sand." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 227, + 242, + 383, + 321 + ], + "blocks": [ + { + "bbox": [ + 227, + 242, + 383, + 321 + ], + "lines": [ + { + "bbox": [ + 227, + 242, + 383, + 321 + ], + "spans": [ + { + "bbox": [ + 227, + 242, + 383, + 321 + ], + "type": "image", + "image_path": "082576aa6da0eca498f07c2c1f67f32cbe00f7ac4ee87ee0afc122825e7d8b1b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 322, + 370, + 348 + ], + "lines": [ + { + "bbox": [ + 231, + 322, + 370, + 348 + ], + "spans": [ + { + "bbox": [ + 231, + 322, + 370, + 348 + ], + "type": "text", + "content": "... In the left image, the player is holding a baseball glove, while in the right image, the player is holding a basketball." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 357, + 539, + 369 + ], + "lines": [ + { + "bbox": [ + 69, + 357, + 539, + 369 + ], + "spans": [ + { + "bbox": [ + 69, + 357, + 539, + 369 + ], + "type": "text", + "content": "Figure 1. Three \"object replacement\" examples within IMG-DIFF, highlighting fine-grined difference in both vision and language." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 392, + 242, + 548, + 321 + ], + "blocks": [ + { + "bbox": [ + 392, + 242, + 548, + 321 + ], + "lines": [ + { + "bbox": [ + 392, + 242, + 548, + 321 + ], + "spans": [ + { + "bbox": [ + 392, + 242, + 548, + 321 + ], + "type": "image", + "image_path": "b8040c21694a3b36da7e7429dfdc17bd7198ceacf7eba16ef649337144ce3df9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 322, + 545, + 348 + ], + "lines": [ + { + "bbox": [ + 391, + 322, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 391, + 322, + 545, + 348 + ], + "type": "text", + "content": "The left image shows a framed picture on a shelf, while the right image shows a white vase with pink flowers. ..." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 378, + 200, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 378, + 200, + 390 + ], + "spans": [ + { + "bbox": [ + 151, + 378, + 200, + 390 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 404, + 296, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 680 + ], + "type": "text", + "content": "High-performance Multimodal Large Language Models (MLLMs) rely heavily on data quality. This study introduces a novel data synthesis method, leveraging insights from contrastive learning and image difference captioning to enhance fine-grained image recognition in MLLMs. By analyzing object differences in detailed regions between similar images, we challenge the model to identify both matching and distinct components. Specifically, our method initially create pairs of similar images that highlight object variations. After that, we introduce a Difference Area Generator for object differences identifying, followed by a Difference Captions Generator for differences describing. The outcome is a high-quality dataset of \"object replacement\" samples, named Img-Diff, which can be expanded as needed due to its automation. We use the generated dataset to fine-tune state-of-the-art (SOTA) MLLMs such as InternVL2, yielding comprehensive improvements across numerous image difference and Visual Question Answering tasks. For instance, the trained models notably surpass the SOTA models GPT-4V and Gemini on the MMVP benchmark. Additionally, we conduct thorough evaluations to confirm the dataset's diversity, quality, and robustness, presenting several insights on the synthesis of such a contrastive dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 379, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 415 + ], + "type": "text", + "content": "We release our codes and dataset to encourage further research on multimodal data synthesis and MLLMs' fundamental capabilities for image understanding." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 441, + 394, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 394, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 394, + 453 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 461, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 461, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 312, + 461, + 555, + 616 + ], + "type": "text", + "content": "The emergence of large language models (LLMs) has revolutionized natural language processing [48, 64], also paved the way for the development of Multimodal Large Language Models (MLLMs) that seamlessly integrate linguistic and visual understanding. Improving the performance of MLLMs hinges on two primary avenues: evolving model architectures and enhancing dataset quality [53]. The majority of state-of-the-art (SOTA) MLLMs [3, 37, 41-43] employ a two-phase training strategy, commencing with a pre-training phase involving extensive image-text pairs for modality alignment, followed by a fine-tuning phase aimed at optimizing visual question answering (VQA) capabilities with specific instruction tuning datasets." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "The quality of visual instruction tuning datasets plays a crucial role in MLLMs' overall performance in VQA tasks and diverse downstream applications. With the evolution of visual instruction tuning datasets, several recent studies have explored the integration of object detection and Optical Character Recognition (OCR) datasets, such as RefCOCO [28], Visual Genome [31], OCR-VQA [49], and TextVQA [60], effectively enhancing MLLMs' proficiency" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 690, + 295, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 690, + 295, + 714 + ], + "type": "text", + "content": "3 \"FSIETP\" stands for the Guangdong Provincial Key Laboratory of Fire Science and Intelligent Emergency Technology." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9296" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 235, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 235, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 235, + 83 + ], + "type": "text", + "content": "in tasks requiring detailed image perception." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 296, + 277 + ], + "type": "text", + "content": "In this paper, we focus on a new direction for enhancing MLLM datasets, driven by the potential of object variations in image pairs to refine models' image recognition capabilities, as demonstrated by advancements in contrastive learning and image difference captioning [5, 26, 51, 80]. Specifically, we introduce a general-purpose yet challenging dataset named IMG-DIFF, which sets itself apart from existing visual instruction tuning datasets by generating pairs of highly similar images featuring subtle object alterations. Rather than compelling MLLMs to focus solely on a single image, our dataset challenges them to analyze paired images and articulate the differences within designated regions, meanwhile taking the high-quality textual descriptions of the difference as learning signals. By doing so, we aim to empower MLLMs with enhanced capabilities of fine-grained image recognition." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 277, + 296, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 516 + ], + "type": "text", + "content": "To evaluate the effectiveness of our data synthesis method, we integrate the generated dataset into the original visual instruction tuning datasets of LLaVA-1.5 [41], MGM [37], and InternVL2 [12], and conduct fine-tuning. Subsequently, we evaluate their performance on image difference benchmarks, including MMVP [63], Spot-the-Diff [26], and Image-Edit-Request [62], as well as widely recognized MLLM benchmarks. Our evaluation results reveal that after fine-tuning with our IMG-DIFF dataset, the MLLMs achieve notable enhancements in image difference benchmarks, aligning their performance with state-of-the-art (SOTA) models. For instance, they notably surpass the SOTA models GPT-4V [1] and Gemini [16] on the MMVP benchmark, achieving an improvement of up to 12. Moreover, the models exhibit comprehensive improvements across eight well-recognized MLLM benchmarks, achieving an average score improvement of up to " + }, + { + "bbox": [ + 55, + 277, + 296, + 516 + ], + "type": "inline_equation", + "content": "3.06\\%" + }, + { + "bbox": [ + 55, + 277, + 296, + 516 + ], + "type": "text", + "content": ", underscoring the useful role our dataset plays in bolstering MLLMs' competencies in both image difference recognition and fine-grained image analysis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 636 + ], + "type": "text", + "content": "We further evaluate the diversity and quality of our dataset, ensuring it encompasses a broad array of object categories while showcasing rich variability. Through meticulous manual labeling, we affirm the high quality of our dataset. Furthermore, we conduct ablation studies to examine the effects of various filtering intensities. We also investigate an alternative methodology for constructing image difference data focusing on \"object removal\", assessing its effectiveness and presenting fruitful insights on the construction of contrastive data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 638, + 253, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 638, + 253, + 649 + ], + "spans": [ + { + "bbox": [ + 67, + 638, + 253, + 649 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 712 + ], + "type": "text", + "content": "- We present a novel data synthesis method and an effect-proven IMG-DIFF dataset, comprising pairs of highly similar images, with a focus on processes such as segmentation, filtering, and detailed captioning of image differences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 143 + ], + "type": "text", + "content": "- We conduct visual instruction tuning on LLaVA-1.5-7B, MGM-7B, and InternVL2-8B using our dataset, and rigorously assess the fine-tuned models' performance on many widely-used MLLM benchmarks and image difference benchmarks. Our dataset brings substantial performance improvements to the fine-tuned MLLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 144, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 191 + ], + "type": "text", + "content": "- We provide a comprehensive evaluation of the diversity and quality of our dataset, confirming its richness and high standard. Through ablation studies, we identify good empirical filtering thresholds for such contrastive dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 192, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 239 + ], + "type": "text", + "content": "- We open-source our dataset and codes at https://github.com/modelscope/data-juicer/tree/ImgDiff to facilitate ongoing research, encouraging innovative endeavors in MLLM datasets and image difference methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 253, + 494, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 253, + 494, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 253, + 494, + 266 + ], + "type": "text", + "content": "2. Background and Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 274, + 509, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 509, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 509, + 287 + ], + "type": "text", + "content": "2.1. Multimodal Large Language Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 292, + 553, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 553, + 352 + ], + "type": "text", + "content": "Multimodal Large Language Models (MLLMs) have exhibited remarkable advancements since their introduction [1, 16, 37, 43, 44]. Research highlights two key factors that primarily influence the effectiveness of MLLMs: model architecture and dataset quality [53]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 353, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 554, + 449 + ], + "type": "text", + "content": "With respect to model architecture, notable approaches [2, 3, 25, 32, 35] leverage learnable queries to extract essential information from CLIP [15, 54] image features. Alternatively, LLaVA [41-43] and MGM [37] utilize projection-based interfaces to facilitate interactions between text and image modalities. Furthermore, LLaMA-Adapter [81] and LaVIN [47] implement parameter-efficient tuning mechanisms to transfer image-related information to the LLM." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 449, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 640 + ], + "type": "text", + "content": "From the perspective of datasets, there are two prevalent strategies: enhancing the quality of pre-training data and improving instruction tuning data. The former aims for better semantic alignment between images and text by introducing abundant image-text pairs, enhancing MLLMs' fundamental capabilities such as image captioning. As for the latter, research has increasingly concentrated on refining instruction tuning datasets, boosting performance across various question-answering tasks. Works like LLaVA [41-43], SPHINX [40], MGM [37], and InternVL2 [12] leverage high-quality fine-tuning datasets with extensive task diversity, allowing models to excel in tasks related to image perception, reasoning, and optical character recognition. Additionally, methods such as Shikra [10], ASM [72], and PINK [75] utilize substantial amounts of object detection data to enhance the models' localization capabilities." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "In contrast to previous works, our research introduces a fully automated data synthesis method and generates a dataset that emphasizes image differences, showing empirical effectiveness and great potential to augment MLLMs' VQA proficiency, object localization capabilities, and discernment of image distinctions." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9297" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 201, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 201, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 201, + 85 + ], + "type": "text", + "content": "2.2. Image Difference Datasets" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": "Datasets focused on image differences typically consist of pairs of similar images and textual descriptions that describe the variations. For instance, the Spot-the-Diff dataset [26] contains pairs of street scenes captured at different times by the same surveillance cameras. The CLEVR-Change dataset [51] delineates scene variations of geometric objects against a clean backdrop. The CUB-Bird dataset [69] focuses on the nuanced differences among various bird species found in natural habitats. The Image-Edit-Request [62] dataset features edited images, accompanied by descriptions of modifications made." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 221, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 296, + 316 + ], + "type": "text", + "content": "Leveraging advancements in image editing technologies, some studies have employed generative models and editing techniques to create datasets centered on image differences. A prime example is InstructPix2Pix [6], which utilizes the image editing technique, named Prompt-to-Prompt [20], to direct Stable-Diffusion-1.5 [56] in generating pairs of similar images, while employing GPT-3 [48] to craft the edited text as reference captions for image differences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 317, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 413 + ], + "type": "text", + "content": "Our approach, referring to InstructPix2Pix, employs the Prompt-to-Prompt technique alongside an advanced generative model Stable-Diffusion-XL [52], which produces more realistic images, to generate pairs of similar images. Unlike InstructPix2Pix, we incorporate multiple filtering stages to ensure data quality, with a particular emphasis on producing difference captions that focus on specific regions rather than the entire image, which ensures greater accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 421, + 268, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 268, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 268, + 434 + ], + "type": "text", + "content": "2.3. Models for Image Difference Captioning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 653 + ], + "type": "text", + "content": "Image Difference Captioning (IDC) represents a specialized domain within image captioning characterized by its focus on subtle variations between images. As for the pioneering work in IDC, Spot-the-Diff [26] presents potential change clusters and employs an LSTM [21] decoder to generate difference captions. DUDA [51] explores image differences at the semantic level, using a ResNet [19] and an LSTM to compute dynamic attention weights and produce captions. VARD [65] introduces a viewpoint-adaptive representation disentanglement network based on LSTM for differentiating between real and pseudo changes. Meanwhile, NCT [66] employs a transformer [67] to integrate neighboring features, and CLIP4IDC [18] uses BERT-like training methodologies, adapting a CLIP model for IDC tasks. With the emergence of MLLMs, VIXEN [5] marks the inaugural use of these models for IDC tasks, mapping the features of differential images to text space and employing an LLM to generate image difference captions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Our data synthesis method is specifically designed for MLLMs. We build our dataset following the instruction-following format adopted by mainstream MLLMs such as LLaVA-1.5, highlighting a novel research direction aimed at enhancing MLLMs from a data-centric perspective." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 420, + 208 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 420, + 208 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 420, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 420, + 208 + ], + "type": "image", + "image_path": "5f866f61d2fdbb9e67e73bd315297c94cfc646f0b7df5665b9409c11b81e43ed.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 216, + 550, + 228 + ], + "lines": [ + { + "bbox": [ + 318, + 216, + 550, + 228 + ], + "spans": [ + { + "bbox": [ + 318, + 216, + 550, + 228 + ], + "type": "text", + "content": "Figure 2. The generation process for \"object replacement\" data." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 422, + 70, + 496, + 207 + ], + "blocks": [ + { + "bbox": [ + 422, + 70, + 496, + 207 + ], + "lines": [ + { + "bbox": [ + 422, + 70, + 496, + 207 + ], + "spans": [ + { + "bbox": [ + 422, + 70, + 496, + 207 + ], + "type": "image", + "image_path": "f22d289ae7b9c2a60939dbe1f1c322036d19fd9f5927a6817f0c6b2ddea4f0fd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 499, + 73, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 499, + 73, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 499, + 73, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 499, + 73, + 553, + 198 + ], + "type": "image", + "image_path": "6fd31c0668d5179748c746ee6fb306bac90a5de3b8a46dca09a7e4865f9ff4e6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 249, + 397, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 249, + 397, + 263 + ], + "spans": [ + { + "bbox": [ + 314, + 249, + 397, + 263 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 270, + 382, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 270, + 382, + 282 + ], + "spans": [ + { + "bbox": [ + 314, + 270, + 382, + 282 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 289, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 373 + ], + "type": "text", + "content": "In recent years, contrastive learning have significantly enhanced the image-text understanding of vision-language models [54, 80]. These methods typically involve constructing batches of images and texts, requiring the model to distinguish between matching and non-matching image-text pairs, which improves its ability to differentiate between semantically similar and dissimilar pairs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 375, + 554, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 554, + 458 + ], + "type": "text", + "content": "Our method leverages the principles of contrastive learning to generate MLLM image-text data. Specifically, it focuses on replacing objects within image pairs, encouraging MLLMs to identify similarities and differences in specific regions. This method aims to enhance models' capacity to recognize fine-grained differences in images, guided by textual descriptions that highlight detailed distinctions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 460, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 590 + ], + "type": "text", + "content": "As shown in Figure 2, the process of generating \"object replacement\" data can be divided into three key parts. The first part involves creating similar images and forming image pairs, where the only difference between the images in pairs is the objects replaced (Section 3.2). The second part, named the \"Difference Area Generator\", extracts bounding box regions that contain object differences between the images in pairs (Section 3.3). The third part, termed as the \"Difference Captions Generator\" (Section 3.4), utilizes an MLLM to generate descriptive text for the areas with object differences and creates question-answer pairs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 593, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 554, + 664 + ], + "type": "text", + "content": "This process incorporates multiple filtering operations, with specific thresholds outlined in Section 18.4, which we determine through experimental comparisons in Section 17. Additional experimental details, including model selection and the time consumption, are also presented in Section 18. For readability, we present data examples in Section 20." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "To enhance reproducibility and reusability, our proposed components and end-to-end construction workflow are implemented as data processing operators and configurable data recipes within Data-Juicer [8, 9]." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9298" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 68, + 496, + 247 + ], + "blocks": [ + { + "bbox": [ + 117, + 68, + 496, + 247 + ], + "lines": [ + { + "bbox": [ + 117, + 68, + 496, + 247 + ], + "spans": [ + { + "bbox": [ + 117, + 68, + 496, + 247 + ], + "type": "image", + "image_path": "d7b9448d1919c7579c5ddd08be646025e5255f3d6ba1a7c249b8f7536c95769b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 254, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 254, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 555, + 277 + ], + "type": "text", + "content": "Figure 3. An overview of the Difference Area Generator and its three internal components: Image Similarity Filter, Image-text Matching Filter, and Difference Detector." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 297, + 191, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 191, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 191, + 310 + ], + "type": "text", + "content": "3.2. Image Pairs Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 315, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 386 + ], + "type": "text", + "content": "The first step of our data synthesis method is to generate pairs of similar images as candidates. The process is shown in Figure 2. We employ a generative model called Stable-Diffusion-XL [52] and an image editing technique called Prompt-to-Prompt [20] to generate image pairs that highlight object replacement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 387, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 495 + ], + "type": "text", + "content": "We start by obtaining image captions from caption databases, which contain descriptions biased towards real photos. Then, we use an LLM to perform object replacement in the captions. The prompt used is \"Here is a sentence: 'INPUT'. Please only replace one of the objects in this sentence with another object.\" Here, INPUT refers to the original caption. Next, based on the caption pairs, we use the Stable-Diffusion-XL and the Prompt-to-Prompt to generate image pairs with only few objects replaced." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 502, + 204, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 204, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 204, + 513 + ], + "type": "text", + "content": "3.3. Difference Area Generator" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 519, + 124, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 124, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 124, + 529 + ], + "type": "text", + "content": "3.3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 630 + ], + "type": "text", + "content": "The Difference Area Generator is used to identify the locations of object differences between the two images in pairs. Although object detection models are capable of detecting objects in images, the range of object categories is quite limited [27, 78]. Therefore, to increase the number of detectable object categories and enhance data diversity, we develop the Difference Area Generator based on segmentation and image similarity comparisons." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "As shown in Figure 3, we first use an Image Similarity Filter to obtain image pairs with high similarity but not identical. Next, we use the FastSAM [82] to perform image segmentation on each image. After that, we crop the images based on the bounding box information obtained from segmentation and use an Image-text Matching Filter to filter the cropped sub-images for the presence of valid objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 297, + 555, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 358 + ], + "type": "text", + "content": "Finally, we use a Difference Detector to determine whether there are indeed differences between the bounding box regions of the two images and perform IoU (Intersection over Union) filtering to remove overlapping bounding boxes, ultimately obtaining valid bounding box information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 365, + 441, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 441, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 441, + 376 + ], + "type": "text", + "content": "3.3.2. Image Similarity Filter" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 380, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 513 + ], + "type": "text", + "content": "The Image Similarity Filter aims to filter image pairs based on the degree of similarity. The module first uses CLIP [54] to extract image features from each image in pairs and then calculates the cosine similarity score. If their score falls within the pre-set threshold, the image pair will be considered valid. Specifically, we use the module twice in the Difference Area Generator. Before the segmentation, we use the module to ensure that the images in pairs are highly similar but not the same. In the Difference Detector stage, after cropping, we use the module to filter the sub-image pairs and keep only the differing ones." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 520, + 459, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 459, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 459, + 532 + ], + "type": "text", + "content": "3.3.3. Image-text Matching Filter" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 535, + 556, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 556, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 556, + 655 + ], + "type": "text", + "content": "The Image-text Matching Filter determines whether an image contains valid objects (i.e. the replaced or replacing objects). This module first uses BLIP [34] to extract image features, which are then compared with text features of object names. When the image-text matching score falls within the pre-set threshold, we consider the image to contain valid objects. In the mid-stage of the Difference Area Generator, after cropping, we use the module to determine whether these sub-images contain valid objects and get valid bounding boxes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 662, + 425, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 425, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 425, + 673 + ], + "type": "text", + "content": "3.3.4. Difference Detector" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "The Difference Detector determines whether there are differences between the two regions in the image pair of the same bounding box. Based on a given bounding box, we" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9299" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 69, + 531, + 196 + ], + "blocks": [ + { + "bbox": [ + 80, + 69, + 531, + 196 + ], + "lines": [ + { + "bbox": [ + 80, + 69, + 531, + 196 + ], + "spans": [ + { + "bbox": [ + 80, + 69, + 531, + 196 + ], + "type": "image", + "image_path": "ef3f77d87c470c1d2fc964538b69a793bc79ef4895898ec4dcae01b1a10f849d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 203, + 448, + 216 + ], + "lines": [ + { + "bbox": [ + 160, + 203, + 448, + 216 + ], + "spans": [ + { + "bbox": [ + 160, + 203, + 448, + 216 + ], + "type": "text", + "content": "Figure 4. An overview of the Difference Captions Generator and its two stages." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 236, + 296, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 320 + ], + "type": "text", + "content": "first crop two sub-images from both image A and image B. The two sub-images are then filtered through the Image Similarity Filter and the bounding box is considered effective only if the difference is significant enough. After processing all bounding boxes, we use the IoU method to filter out overlapping bounding boxes. Only the bounding boxes with a higher degree of difference are retained." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 326, + 223, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 223, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 223, + 339 + ], + "type": "text", + "content": "3.4. Difference Captions Generator" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 343, + 124, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 124, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 124, + 355 + ], + "type": "text", + "content": "3.4.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 358, + 296, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 466 + ], + "type": "text", + "content": "After obtaining the valid bounding box regions, we use the Difference Captions Generator to generate descriptions for the differences inside these areas (with each round of the process focusing on only one bounding box in one image pair). Evidently, an image pair can contain multiple object differences but a single difference caption cannot fully capture all of them. Therefore, we highlight specific regions with red boxes and provide targeted difference captions to ensure greater accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 466, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 296, + 550 + ], + "type": "text", + "content": "The module consists of two stages: the first stage generates captions for the content in the bounding box regions and then filters the bounding boxes using an Image-text Matching Filter and a Captions Similarity Filter. The second stage uses the content captions and the images highlighted with red boxes to generate difference captions. The overview is shown in Figure 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 555, + 239, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 239, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 239, + 567 + ], + "type": "text", + "content": "3.4.2. Stage1: Object Labeling & Filtering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "In Stage 1, for each image pair, we first select N bounding box regions with the lowest similarity between images (N is set to 5 in this project) as candidate regions. Then, for each bounding box, we use the MLLM LLaVA-NEXT [43] to describe its corresponding regions and then apply two filtering processes: the first filter is an Image-text Matching Filter, which checks whether the content of the regions matches the captions; the second filter is an Captions Similarity Filter, which assesses whether there are differences between the two captions. Once the filtering is complete, we obtain valid bounding boxes and captions for subsequent difference captioning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 236, + 452, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 452, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 452, + 248 + ], + "type": "text", + "content": "3.4.3. Captions Similarity Filter" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 251, + 554, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 311 + ], + "type": "text", + "content": "The Captions Similarity Filter determines whether the two captions of the same bounding box coordinates are different. We use CLIP to obtain text features and calculate the cosine similarity between them. When the score is low enough, we consider the two captions to be different." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 316, + 514, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 514, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 514, + 328 + ], + "type": "text", + "content": "3.4.4. Stage2: Difference Captions Generating" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 331, + 554, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 427 + ], + "type": "text", + "content": "In Stage 2, for each valid bounding box of each image pair, we first draw two red boxes into the images based on the bounding box coordinates, highlighting the differences for easier localization. Then, we provide the MLLM LLaVA-NEXT with the captions of the bounding box content and instruct it to describe the difference based on these captions and the highlighted images. Finally, we can obtain the difference caption for the highlighted area." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 434, + 405, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 405, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 405, + 445 + ], + "type": "text", + "content": "3.5. Data Statistics" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "text", + "content": "Using captions from MSCOCO [39], we generate 118K pairs of similar images. We then employ the Image Similarity Filter to get 38,533 highly similar but not identical image pairs. Next, we use the Difference Area Generator to filter and produce 117,779 pieces of valid bounding box information (with a maximum of 5 valid bounding boxes per image pair). Finally, we employ the Difference Captions Generator to filter and generate 12,688 high-quality \"object replacement\" instances. Our evaluations on the main page are based on this dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 571, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 554, + 643 + ], + "type": "text", + "content": "In addition, we also use captions from the LLaVA pretraining dataset to generate 34,538 \"object replacement\" samples. In Section 10 of the appendix, we evaluate this dataset and discuss the relationship between data quantity, data quality, and model performance improvement, emphasizing the importance of prioritizing quality over quantity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 652, + 479, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 479, + 664 + ], + "type": "text", + "content": "4. Evaluation and Main Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 672, + 429, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 429, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 429, + 685 + ], + "type": "text", + "content": "4.1. Evaluation Settings" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "To evaluate the effectiveness of our IMG-DIFF dataset, we use it to fine-tune SOTA MLLMs, including LLaVA-1.5-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9300" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "7B, MGM-7B, and InternVL2-8B on the main page, as well as InternVL2-1B and LLaVA-1.5-13B in the supplementary materials. We evaluate these models on extensive benchmarks commonly used for image difference and MLLMs. Specifically, the image difference benchmarks include MMVP [63], Spot-the-Diff [26], and Image-Edit-Request [62]. Besides, the details and results of the MLLM benchmarks are shown in Section 4.5." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 295, + 312 + ], + "type": "text", + "content": "During our fine-tuning process, we first mix our data with MLLMs' original visual instruction tuning data respectively. For LLaVA-1.5 and MGM, we conduct the fine-tuning anew. For InternVL2, we follow the guidelines from its official repository and perform a second fine-tuning. (To ensure a fair comparison, we perform a second fine-tuning on InternVL2 using its original fine-tuning dataset, termed as InternVL2-FT, and use it as a baseline.) Regarding the Spot-the-Diff and Image-Edit-Request benchmarks, as they contain data of training splits, we further fine-tune the fine-tuned MLLMs for an additional 2 epochs using only these benchmarks' training data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 312, + 294, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 312, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 67, + 312, + 294, + 324 + ], + "type": "text", + "content": "In the tables, \"RP\" represents \"object replacement\" data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 239, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 239, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 239, + 342 + ], + "type": "text", + "content": "4.2. Results on the MMVP Benchmark" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "text", + "content": "The MMVP benchmark is designed to systematically assess the visual capabilities of MLLMs. Its data processing method is highly related to image difference: it first collects CLIP-blind pairs, which have similar CLIP features but differ in image content. Then, the differences between the images are manually described and question-answer pairs are created. Hence, the questions in MMVP is highly relevant to our dataset, as both place significant emphasis on the differences between similar images." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 55, + 464, + 295, + 587 + ], + "blocks": [ + { + "bbox": [ + 55, + 464, + 295, + 587 + ], + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 587 + ], + "type": "image", + "image_path": "de86155551b0c8394b3cbcbdb98208bafb2bec1b8b785a0bd985e99c09c58452.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 594, + 288, + 605 + ], + "lines": [ + { + "bbox": [ + 61, + 594, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 61, + 594, + 288, + 605 + ], + "type": "text", + "content": "Figure 5. Performance comparison on the MMVP benchmark." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "As shown in Figure 5, fine-tuning MLLMs with our \"object replacement\" data significantly improves their performance on the MMVP benchmark. After fine-tuning, the score of LLaVA-1.5-7B exceeds that of LLaVA-1.5-13B. Furthermore, the fine-tuned MGM-7B shows a significant improvement in score compared to the original MGM-7B, even surpassing the scores of the SOTA models GPT-4V and Gemini by up to 12. Additionally, the performance of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "InternVL2-8B is also enhanced. These results suggest that our dataset enhances MLLMs' abilities to distinguish images with similar CLIP features but different content." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 118, + 524, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 524, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 524, + 131 + ], + "type": "text", + "content": "4.3. Results on the Spot-the-Diff Benchmark" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 137, + 553, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 553, + 184 + ], + "type": "text", + "content": "The dataset of Spot-the-Diff comprises pairs of street view images that display subtle object differences. These images are obtained by capturing scenes from fixed surveillance cameras at different time." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 329, + 214, + 541, + 334 + ], + "blocks": [ + { + "bbox": [ + 339, + 194, + 528, + 206 + ], + "lines": [ + { + "bbox": [ + 339, + 194, + 528, + 206 + ], + "spans": [ + { + "bbox": [ + 339, + 194, + 528, + 206 + ], + "type": "text", + "content": "Table 1. Performance comparison on Spot-the-Diff." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 214, + 541, + 334 + ], + "lines": [ + { + "bbox": [ + 329, + 214, + 541, + 334 + ], + "spans": [ + { + "bbox": [ + 329, + 214, + 541, + 334 + ], + "type": "table", + "html": "
ModelBLEUMETEORCIDEr-DROUGE-L
VAM[59]10.112.438.131.3
IFDC[23]8.711.73730.2
DUDA+Aux[22]8.112.534.529.9
VACC[29]9.712.641.532.1
LLaVA-1.5-7B8.512.038.330.1
LLaVA-1.5-7B + RP9.713.043.230.8
MGM-7B9.912.046.331.5
MGM-7B + RP10.813.153.533.0
InternVL2-8B-FT6.611.726.527.3
InternVL2-8B + RP8.412.832.228.5
", + "image_path": "59c73587752853584877740710aa209e40095499e9cc07172883e28a6e69fd73.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 349, + 554, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 554, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 554, + 482 + ], + "type": "text", + "content": "Following previous works, our fine-tuned MLLMs are evaluated on BLEU [50], METEOR [4], CIDEr-D [68] and ROUGE-L [38]. As shown in Table 1, after fine-tuning with our \"object replacement\" data, both LLaVA-1.5-7B, MGM-7B, and InternVL2-8B achieve significant performance gains on Spot-the-Diff. Compared to score increases seen in prior models, our dataset yields even greater enhancements than those from previous iterations of image difference models. These positive results indicate that our dataset can enhance the ability of MLLMs to detect subtle differences between similar images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 491, + 497, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 497, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 497, + 504 + ], + "type": "text", + "content": "4.4. Results on Image-Editing-Request" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 509, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 605 + ], + "type": "text", + "content": "The Image-Editing-Request benchmark focuses on image editing. Each instance in its dataset consists of an image pair (i.e. a source image and an edited image) and an editing instruction that describes the transformation. During the evaluation, our models are required to generate transformation descriptions for these image pairs, and we then calculate the BLEU, METEOR, CIDEr-D, and ROUGE-L scores using the models' responses and the reference answers." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": "As shown in Table 2, LLaVA-1.5-7B, MGM-7B, and InternVL2-8B originally show SOTA performance on the Image-Edit-Request benchmark. Upon incorporating our \"object replacement\" data for better fine-tuning, the performance of all three models improves even more, achieving new SOTA scores. This suggests that our dataset enhances MLLMs' abilities to recognize similarities and dissimilarities in image pairs, as well as enables them to describe differences more accurately." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9301" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 91, + 284, + 222 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 281, + 82 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 281, + 82 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 281, + 82 + ], + "type": "text", + "content": "Table 2. Performance comparison on Image-Edit-Request." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 91, + 284, + 222 + ], + "lines": [ + { + "bbox": [ + 67, + 91, + 284, + 222 + ], + "spans": [ + { + "bbox": [ + 67, + 91, + 284, + 222 + ], + "type": "table", + "html": "
ModelBLEUMETEORCIDEr-DROUGE-L
VARD[65]1014.835.739
CLIP4IDC[18]8.214.632.240.4
NCT[66]8.11534.238.8
BiDiff[61]6.914.627.738.5
VIXEN[5]8.615.438.142.5
LLaVA-1.5-7B15.117.860.645.2
LLaVA-1.5-7B + RP16.219.560.946.7
MGM-7B16.517.766.844.8
MGM-7B + RP16.618.268.145.7
InternVL2-8B-FT12.414.151.538.9
InternVL2-8B + RP12.514.256.039.4
", + "image_path": "be8a01249aa26139a364196ffe34a2abb585fab8cf1983917fa99db96c3a4a3a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 227, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 227, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 227, + 255 + ], + "type": "text", + "content": "4.5. Results on MLLM Benchmarks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 261, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 296, + 405 + ], + "type": "text", + "content": "Aside from the evaluations related to image difference discrimination, we also assess the performance of our \"object replacement\" data in enhancing the comprehensive abilities of MLLMs. We test the fine-tuned MLLMs using commonly used MLLM benchmarks, including VQAv2 [17] and GQA [24] for assessing the comprehensive VQA capabilities; MMBench [45], MM-Vet [76], ScienceQA [46], and SEED-Bench [33] for testing perceptual and reasoning abilities; and POPE [36] for evaluating fine-grained object localization abilities. Table 3 presents the results on these MLLM benchmarks, with the " + }, + { + "bbox": [ + 55, + 261, + 296, + 405 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 55, + 261, + 296, + 405 + ], + "type": "text", + "content": " metric indicating the percentage improvement averaged across them." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 62, + 437, + 288, + 517 + ], + "blocks": [ + { + "bbox": [ + 65, + 417, + 285, + 428 + ], + "lines": [ + { + "bbox": [ + 65, + 417, + 285, + 428 + ], + "spans": [ + { + "bbox": [ + 65, + 417, + 285, + 428 + ], + "type": "text", + "content": "Table 3. Performance comparison on 8 MLLM benchmarks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 437, + 288, + 517 + ], + "lines": [ + { + "bbox": [ + 62, + 437, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 288, + 517 + ], + "type": "table", + "html": "
ModelVQAv2GQAPOPEMMB\\( {\\mathrm{{MMB}}}^{CN} \\)
LLaVA-1.5-7B78.56285.964.358.3
LLaVA-1.5-7B + RP79.3↑62.8↑86.4↑66.1↑59.8↑
MGM-7B80.462.68669.358.9
MGM-7B + RP80.7↑62.7↑86.2↑68.7↓59.6↑
InternVL2-8B-FT81.862.687.782.581.5
InternVL2-8B + RP81.862.688.0↑82.7↑81.4↓
", + "image_path": "c92d46c4f0ac9ffa25dc416fa95a8ed20cd3dcae5038b6723ddb4af17a99bca2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 62, + 523, + 288, + 604 + ], + "blocks": [ + { + "bbox": [ + 62, + 523, + 288, + 604 + ], + "lines": [ + { + "bbox": [ + 62, + 523, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 288, + 604 + ], + "type": "table", + "html": "
ModelMM-Vet\\( \\mathbf{SQA}^I \\)SEED\\( \\bigtriangleup \\)
LLaVA-1.5-7B30.566.858.6-
LLaVA-1.5-7B + RP33.2↑68.2↑61.7↑+3.06%
MGM-7B40.870.663.5-
MGM-7B + RP44.1↑71.7↑63.2↓+1.28%
InternVL2-8B-FT49.296.569.5-
InternVL2-8B + RP52.6↑96.6↑69.9↑+1.01%
", + "image_path": "6f74d7b9552fd567574dfbdb998b5f82bedbf77d49a59eb7bde0c7b1c23ca972.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Based on Table 3, after fine-tuning with our dataset, the performance of LLaVA-1.5-7B shows a comprehensive improvement, with an average increase of " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "3.06\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": " across all benchmarks. For MGM-7B and InternVL2-8B, the improvements brought by our dataset are not as pronounced as those observed with LLaVA-1.5-7B, as their training datasets already encompass a large volume of high-quality data, but they still achieve an average increase of " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.28\\%" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "1.01\\%" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": ". These improvements indicate that fine-tuning MLLMs with our dataset not only enhances their ability to discern differences but also improves their overall visual capabilities, thereby making them better address VQA tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 130, + 538, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 538, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 538, + 144 + ], + "type": "text", + "content": "5. Evaluation of Data Quality and Diversity" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 150, + 397, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 397, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 397, + 163 + ], + "type": "text", + "content": "5.1. Data Quality" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 167, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 479 + ], + "type": "text", + "content": "To assess the quality of our Img-Diff dataset, we randomly select 1,000 instances of \"object replacement\" data and employ multiple professional dataset annotators to evaluate the samples based on three metrics. The final scores are determined through a voting process. Specifically, the first metric is \"Bounding Box Difference\", which evaluates whether there are differences between the two highlighted regions in pairs. If the objects are different, we score it as \"high\"; if the objects are the same but their features (such as color, shape, etc.) are different, we score it as \"medium\"; if the objects are the same and their features are similar, we score it as \"low\". The second metric is \"Content Caption Accuracy\", which evaluates whether the captions generated by Stage 1 of the Difference Captions Generator accurately describe the two highlighted regions. If the captions are correct, we score it as \"high\"; if the captions identify the objects but incorrectly describe their features, we score it as \"medium\"; if the captions incorrectly identify the objects, we score it as \"low\". The third metric is \"Difference Caption Accuracy\", which evaluates whether the final difference captions accurately describe the object differences between highlighted regions of the image pairs. If the description is accurate, we score it as \"high\"; if it recognizes the objects but the feature description is incorrect, we score it as \"medium\"; if the object recognition is incorrect, we score it as \"low\". The results are shown in Figure 6." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 487, + 553, + 536 + ], + "blocks": [ + { + "bbox": [ + 315, + 487, + 553, + 536 + ], + "lines": [ + { + "bbox": [ + 315, + 487, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 315, + 487, + 553, + 536 + ], + "type": "image", + "image_path": "86675cd21f5a495339652134afff76ea4bc8da4341a688d0288cb73423d3c9b8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 546, + 548, + 558 + ], + "lines": [ + { + "bbox": [ + 320, + 546, + 548, + 558 + ], + "spans": [ + { + "bbox": [ + 320, + 546, + 548, + 558 + ], + "type": "text", + "content": "Figure 6. Quality assessment of the \"object replacement\" data." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Based on Figure 6, our dataset demonstrates a high level of quality. For the \"Bounding Box Difference\" metric, only " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " of the instances are classified as \"low\", and nearly " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " of instances exhibit valid object differences between their two highlighted regions. In terms of \"Content Caption Accuracy\", " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "80.1\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " of highlighted region pairs are described accurately, indicating that using an MLLM for labeling is effective and that our filtering strategy is also functioning well. For the \"Difference Caption Accuracy\" metric, over " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " of the difference descriptions are completely accurate, with " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "21.8\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " of the samples having errors solely in feature labeling while still maintaining correct descriptions of ob-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9302" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "ject differences, which underscores the effectiveness of our difference caption generation strategy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 105, + 146, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 105, + 146, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 105, + 146, + 118 + ], + "type": "text", + "content": "5.2. Data Diversity" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 123, + 295, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 123, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 295, + 290 + ], + "type": "text", + "content": "In designing the generation process, we have made efforts to enhance the diversity of our dataset, which can be divided into two aspects: (a) the inherent diversity of the caption databases; and (b) the introduction of new object names through our object replacement strategy. As for the former, we are using the captions from MS COCO and the LLaVA pre-training dataset, which cover a majority of object categories. We can easily further enhance the data diversity by employing caption databases with greater variability. As for the latter, we employ two methods, including increasing the temperature of the LLM used for noun replacement, as well as randomly replacing object names with nouns from a noun lexicon (shown in Appendix, Section 11). We can further improve data diversity by expanding the noun lexicon." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 291, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 411 + ], + "type": "text", + "content": "By analyzing the valid object names included in the captions that describe highlighted regions, we assess the diversity of our \"object replacement\" data. Specifically, we count the total number of object categories covered, and the total number of unique \"object replacement pairs\". Through our statistical analysis, we find that our dataset covers 1,203 object categories, which encompasses most real-life objects. An \"object replacement pair\" refers to the combination of the replaced and replacing object names. Our dataset includes 3,680 unique \"object replacement pairs\"." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 411, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 483 + ], + "type": "text", + "content": "To evaluate the coverage of common object categories in our dataset, we analyze the occurrences of the object names from the Object365 [58] dataset in our own dataset. The results show that the object names from the Object365 dataset appears 13,164 times in total, which accounts for approximately " + }, + { + "bbox": [ + 55, + 411, + 295, + 483 + ], + "type": "inline_equation", + "content": "52.06\\%" + }, + { + "bbox": [ + 55, + 411, + 295, + 483 + ], + "type": "text", + "content": " of the total occurrences of object names." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 544 + ], + "type": "text", + "content": "These statistics show that our data covers a substantial number of common object names, ensuring a high frequency of common objects. Additionally, less common object names make up nearly half of our dataset, indicating that our dataset is both diverse and comprehensive." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 555, + 263, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 263, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 263, + 570 + ], + "type": "text", + "content": "6. More Empirical Supports and Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 624 + ], + "type": "text", + "content": "We list additional experimental details and extensions of our data synthesis method in the supplementary material, and conduct further experiments to validate the effectiveness of our IMG-DIFF dataset, along with its limitations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": "Further Experiments and Analysis. In Section 9, we compare our IMG-DIFF dataset with existing image difference datasets on characteristics and performance, validating its superiority. In Section 10, we discuss the relationship between the data volume and model performance improvement. In Section 11, we explore the use of lexicons for" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "object replacement and validate its effectiveness in enhancing diversity. In Section 12, we evaluate our dataset using the Contrastive Chain-of-Thought method. In Section 13, we validate the effectiveness of our dataset on MLLMs of different scales. In Section 14, we evaluate top-performing MLLMs on difference detection. In Section 15, we assess the impact of unnatural images. In Section 16, we show our data's efficacy in enhancing spatial reasoning capabilities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 183, + 553, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 553, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 553, + 256 + ], + "type": "text", + "content": "Implementation Details. In Section 17, we investigate the effect of different filtering intensities on the performance of the generated datasets. In Section 18, we present additional experimental details, including image pair preprocessing, model training procedures, model selection, filtering thresholds, and time consumption." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 271, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 331 + ], + "type": "text", + "content": "The \"Object Removal\" Exploration. In Section 19, we generate an extended dataset that focuses on object removal, which prompts MLLMs to analyze which image contains a specific object. The new data brings further improvement to the fine-tuned MLLM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 342, + 388, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 342, + 388, + 355 + ], + "spans": [ + { + "bbox": [ + 314, + 342, + 388, + 355 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 362, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 624 + ], + "type": "text", + "content": "In this paper, we draw inspiration from recent advances in contrastive learning and image difference captioning to propose a novel method of contrastive data synthesis. This method enables the creation of a high-quality dataset called \"IMG-DIFF\", which highlights object differences focusing on fine-grained regions in images. Specifically, we first generate pairs of similar images where the main focus is on object replacement. Then, we use the proposed Difference Area Generator and Difference Captions Generator to generate difference captions and form question-answer pairs. In contrast to previous image difference datasets, our dataset focuses exclusively on specific regions inside images. This characteristic circumvents the issue where a single description cannot fully capture the differences of the whole images in pairs, enhancing the accuracy. Afterwards, we fine-tune many MLLMs using the generated dataset, yielding high-performance scores on par with SOTA models in image difference tasks and comprehensive performance improvements in eight widely recognized MLLM benchmarks. These results confirm the effectiveness of our dataset to improve the ability of MLLMs in recognizing detailed differences between images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 625, + 556, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 556, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 556, + 703 + ], + "type": "text", + "content": "In a nutshell, we provide a series of insights about the construction of high-quality image difference datasets, showing great potential to effectively and efficiently enhance MLLMs via contrastive data-centric approaches. With this work, we hope it can catalyze further investigation into the realm of image difference datasets and the fine-grained image recognition capabilities of MLLMs." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9303" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 295, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 268 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 295, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 324 + ], + "type": "text", + "content": "[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 326, + 295, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 326, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 326, + 295, + 379 + ], + "type": "text", + "content": "[5] Alexander Black, Jing Shi, Yifei Fan, Tu Bui, and John Colomosse. Vixen: Visual text comparison network for image difference captioning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 846-854, 2024. 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 382, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 382, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 382, + 295, + 435 + ], + "type": "text", + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3, 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 437, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 295, + 689 + ], + "type": "text", + "content": "[7] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, Shuaibin Li, Wei Li, Yinling Li, Hongwei Liu, Jiangning Liu, Jiawei Hong, Kaiwen Liu, Kuikun Liu, Xiaoran Liu, Chengqi Lv, Haijun Lv, Kai Lv, Li Ma, Runyuan Ma, Zerun Ma, Wenchang Ning, Linke Ouyang, Jiantao Qiu, Yuan Qu, Fukai Shang, Yunfan Shao, Demin Song, Zifan Song, Zhihao Sui, Peng Sun, Yu Sun, Huanze Tang, Bin Wang, Guoteng Wang, Jiaqi Wang, Jiayu Wang, Rui Wang, Yudong Wang, Ziyi Wang, Xingjian Wei, Qizhen Weng, Fan Wu, Yingtong Xiong, Chao Xu, Ruiliang Xu, Hang Yan, Yirong Yan, Xiaogui Yang, Haochen Ye, Huaiyuan Ying, Jia Yu, Jing Yu, Yuhang Zang, Chuyu Zhang, Li Zhang, Pan Zhang, Peng Zhang, Ruijie Zhang, Shuo Zhang, Songyang Zhang, Wenjian Zhang, Wenwei Zhang, Xingcheng Zhang, Xinyue Zhang, Hui Zhao, Qian Zhao, Xiaomeng Zhao, Fengzhe Zhou, Zaida Zhou, Jingming Zhuo, Yicheng Zou, Xipeng Qiu, Yu Qiao and Dahua Lin. Internl m2 technical report, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 295, + 714 + ], + "type": "text", + "content": "[8] Daoyuan Chen, Yilun Huang, Zhijian Ma, Hesen Chen, Xuchen Pan, Ce Ge, Dawei Gao, Yuexiang Xie, Zhaoyang" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "text", + "content": "Liu, Jinyang Gao, et al. Data-juicer: A one-stop data processing system for large language models. In Companion of the 2024 International Conference on Management of Data, pages 120-134, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 554, + 162 + ], + "type": "text", + "content": "[9] Daoyuan Chen, Haibin Wang, Yilun Huang, Ce Ge, Yaliang Li, Bolin Ding, and Jingren Zhou. Data-juicer sandbox: A feedback-driven suite for multimodal data-model codevelopment. arXiv preprint arXiv:2407.11784, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[10] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 209, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 285 + ], + "type": "text", + "content": "[11] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Qinglan Huang, Canyu Chen, Qinghao Ye, Zhihong Zhu, Yuqing Zhang, Jiawei Zhou, Zhuokai Zhao, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Mj-bench: Is your multimodal reward model really a good judge for text-to-image generation?, 2024. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "text", + "content": "[12] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 342, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 342, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 317, + 342, + 553, + 407 + ], + "type": "text", + "content": "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 317, + 342, + 553, + 407 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 317, + 342, + 553, + 407 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023.8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "text", + "content": "[14] W Dai, J Li, D Li, AMH Tiong, J Zhao, W Wang, B Li, P Fung, and S Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 455, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 553, + 520 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 522, + 430, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 430, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 430, + 533 + ], + "type": "text", + "content": "[16] Google. Gemini. 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 534, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 554, + 589 + ], + "type": "text", + "content": "[17] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 590, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 623 + ], + "type": "text", + "content": "[18] Zixin Guo, Tzu-Jui Julius Wang, and Jorma Laaksonen. Clip4idc: Clip for image difference captioning. arXiv preprint arXiv:2206.00629, 2022. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[20] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3, 4" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9304" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "text", + "content": "[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 297, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 297, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 297, + 140 + ], + "type": "text", + "content": "[22] Mehrdad Hosseinzadeh and Yang Wang. Image change captioning by learning from an auxiliary task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2725-2734, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 196 + ], + "type": "text", + "content": "[23] Qingbao Huang, Yu Liang, Jielong Wei, Yi Cai, Hanyu Liang, Ho-fung Leung, and Qing Li. Image difference captioning with instance-level fine-grained feature representation. IEEE transactions on multimedia, 24:2004-2017, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 296, + 252 + ], + "type": "text", + "content": "[24] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 319 + ], + "type": "text", + "content": "[25] Laurençon Hugo, van Strien Daniel, Bekman Stas, Tronchon Leo, Saulnier Lucile, Wang Thomas, Karamcheti Siddharth, Singh Amanpreet, Pistilli Giada, Jernite Yacine, and Sanh Victor. Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics, 2023.2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 354 + ], + "type": "text", + "content": "[26] Harsh Jhamtani and Taylor Berg-Kirkpatrick. Learning to describe differences between pairs of similar images. arXiv preprint arXiv:1808.10584, 2018. 2, 3, 6, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 376 + ], + "type": "text", + "content": "[27] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. *Ultralytics YOLO*, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 296, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 296, + 433 + ], + "type": "text", + "content": "[28] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 487 + ], + "type": "text", + "content": "[29] Hoeseong Kim, Jongseok Kim, Hyungseok Lee, Hyunsung Park, and Gunhee Kim. Agnostic change captioning with cycle consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2095-2104, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 489, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 296, + 544 + ], + "type": "text", + "content": "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 296, + 610 + ], + "type": "text", + "content": "[31] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 612, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 296, + 645 + ], + "type": "text", + "content": "[32] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "text", + "content": "[33] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[34] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 108, + 555, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 108, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 315, + 108, + 555, + 152 + ], + "type": "text", + "content": "[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 154, + 555, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 154, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 315, + 154, + 555, + 198 + ], + "type": "text", + "content": "[36] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 200, + 554, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 200, + 554, + 244 + ], + "spans": [ + { + "bbox": [ + 315, + 200, + 554, + 244 + ], + "type": "text", + "content": "[37] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv:2403.18814, 2023. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 246, + 554, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 246, + 554, + 279 + ], + "spans": [ + { + "bbox": [ + 315, + 246, + 554, + 279 + ], + "type": "text", + "content": "[38] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 281, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 281, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 315, + 281, + 554, + 348 + ], + "type": "text", + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 350, + 554, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 350, + 554, + 405 + ], + "spans": [ + { + "bbox": [ + 315, + 350, + 554, + 405 + ], + "type": "text", + "content": "[40] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 407, + 554, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 407, + 554, + 438 + ], + "spans": [ + { + "bbox": [ + 315, + 407, + 554, + 438 + ], + "type": "text", + "content": "[41] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 441, + 554, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 441, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 315, + 441, + 554, + 474 + ], + "type": "text", + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 475, + 554, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 475, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 315, + 475, + 554, + 520 + ], + "type": "text", + "content": "[43] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,2, 5,3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 522, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 522, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 315, + 522, + 554, + 589 + ], + "type": "text", + "content": "[44] Mengsha Liu, Daoyuan Chen, Yaliang Li, Guian Fang, and Ying Shen. ChartThinker: A contextual chain-of-thought approach to optimized chart summarization. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 3057-3074, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 590, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 590, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 315, + 590, + 554, + 645 + ], + "type": "text", + "content": "[45] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 647, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 554, + 712 + ], + "type": "text", + "content": "[46] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9305" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[47] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. arXiv preprint arXiv:2305.15023, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "text", + "content": "[48] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 215 + ], + "type": "text", + "content": "[49] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "text", + "content": "[50] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "type": "text", + "content": "[51] Dong Huk Park, Trevor Darrell, and Anna Rohrbach. Robust change captioning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4624-4633, 2019. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "type": "text", + "content": "[52] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "text", + "content": "[53] Zhen Qin, Daoyuan Chen, Wenhao Zhang, Liuyi Yao, Yilun Huang, Bolin Ding, Yaliang Li, and Shuiguang Deng. The synergy between data and multi-modal large language models: A survey from co-development perspective. arXiv preprint arXiv:2407.08583, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "text", + "content": "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 547 + ], + "type": "text", + "content": "[55] Kanchana Ranasinghe, Satya Narayan Shukla, Omid Poursaeed, Michael S Ryoo, and Tsung-Yu Lin. Learning to localize objects improves spatial reasoning in visual-llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12977–12987, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "text", + "content": "[57] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 9, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "text", + "content": "[58] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[59] Xiangxi Shi, Xu Yang, Jiuming Gu, Shafiq Joty, and Jianfei Cai. Finding it at another side: A viewpoint-adapted match" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "ing encoder for change captioning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 574-590. Springer, 2020. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "text", + "content": "[60] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "text", + "content": "[61] Yaoqi Sun, Liang Li, Tingting Yao, Tongyv Lu, Bolun Zheng, Chenggang Yan, Hua Zhang, Yongjun Bao, Guiguang Ding, and Gregory Slabaugh. Bidirectional difference locating and semantic consistency reasoning for change captioning. International Journal of Intelligent Systems, 37 (5):2969-2987, 2022. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[62] Hao Tan, Franck Dernoncourt, Zhe Lin, Trung Bui, and Mohit Bansal. Expressing visual relationships via language. arXiv preprint arXiv:1906.07689, 2019. 2, 3, 6, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "text", + "content": "[63] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "type": "text", + "content": "[64] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[65] Yunbin Tu, Liang Li, Li Su, Junping Du, Ke Lu, and Qingming Huang. Adaptive representation disentanglement network for change captioning. IEEE Transactions on Image Processing, 32:2620-2635, 2023. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "text", + "content": "[66] Yunbin Tu, Liang Li, Li Su, Ke Lu, and Qingming Huang. Neighborhood contrastive transformer for change captioning. IEEE Transactions on Multimedia, 25:9518-9529, 2023. 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[67] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[68] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 568, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 600 + ], + "type": "text", + "content": "[69] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 3, 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "type": "text", + "content": "[70] Jiayu Wang, Yifei Ming, Zhenmei Shi, Vibhav Vineet, Xin Wang, Yixuan Li, and Neel Joshi. Is a picture worth a thousand words? delving into spatial reasoning for vision language models. In The Thirty-Eighth Annual Conference on Neural Information Processing Systems, 2024. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[71] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9306" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 676 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 94 + ], + "type": "text", + "content": "perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[72] Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, et al. The all-seeing project: Towards panoptic visual recognition and understanding of the open world. In The Twelfth International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 228 + ], + "type": "text", + "content": "[73] Weiyun Wang, Zhe Chen, Wenhai Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Jinguo Zhu, Xizhou Zhu, Lewei Lu, Yu Qiao, and Jifeng Dai. Enhancing the reasoning ability of multimodal large language models via mixed preference optimization. arXiv preprint arXiv:2411.10442, 2024.6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 294, + 262 + ], + "type": "text", + "content": "[74] Zhe Xu, Daoyuan Chen, Zhenqing Ling, Yaliang Li, and Ying Shen. Mindgym: Enhancing vision-language models via synthetic self-challenging questions, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 305 + ], + "type": "text", + "content": "[75] Shiyu Xuan, Qingpei Guo, Ming Yang, and Shiliang Zhang. Pink: Unveiling the power of referential comprehension for multi-modal llms. arXiv preprint arXiv:2310.00582, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "type": "text", + "content": "[76] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 407 + ], + "type": "text", + "content": "[77] Daoan Zhang, Junming Yang, Hanjia Lyu, Zijian Jin, Yuan Yao, Mingkai Chen, and Jiebo Luo. Cocot: Contrastive chain-of-thought prompting for large multimodal models with multiple image inputs. arXiv preprint arXiv:2401.02582, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "type": "text", + "content": "[78] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 497 + ], + "type": "text", + "content": "[79] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 499, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 499, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 499, + 294, + 541 + ], + "type": "text", + "content": "[80] Qi Zhang, Yifei Wang, and Yisen Wang. On the generalization of multi-modal contrastive learning. In International Conference on Machine Learning, pages 41677-41693. PMLR, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "type": "text", + "content": "[81] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 599, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 599, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 599, + 294, + 632 + ], + "type": "text", + "content": "[82] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 634, + 294, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 634, + 294, + 676 + ], + "spans": [ + { + "bbox": [ + 56, + 634, + 294, + 676 + ], + "type": "text", + "content": "[83] Ting Zhou, Daoyuan Chen, Qirui Jiao, Bolin Ding, Yaliang Li, and Ying Shen. Humanvbench: Exploring human-centric video understanding capabilities of mllms with synthetic benchmark data, 2025. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9307" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/66fe35a0-e089-4c9c-afba-719727b0b2df_content_list.json b/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/66fe35a0-e089-4c9c-afba-719727b0b2df_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b5ead5d7e2839a7e187ad8fc37564afe6e3ddb39 --- /dev/null +++ b/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/66fe35a0-e089-4c9c-afba-719727b0b2df_content_list.json @@ -0,0 +1,1484 @@ +[ + { + "type": "text", + "text": "IMMUNE: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment", + "text_level": 1, + "bbox": [ + 114, + 125, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Soumya Suvara Ghosal $^{1*}$ Souradip Chakraborty $^{1*}$ Vaibhav Singh $^{2*}$ Tianrui Guan $^{1}$ Mengdi Wang $^{3}$ Ahmad Beirami Furong Huang $^{1}$ Alvaro Velasquez $^{4}$ Dinesh Manocha $^{1\\dagger}$ Amrit Singh Bedi $^{5\\dagger}$", + "bbox": [ + 104, + 204, + 892, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Maryland; 2Indian Institute of Technology Bombay; 3Princeton University; 4University of Colorado Boulder; 5University of Central Florida", + "bbox": [ + 114, + 256, + 883, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 345, + 326, + 362 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the widespread deployment of Multimodal Large Language Models (MLLMs) for visual-reasoning tasks, improving their safety has become crucial. Recent research indicates that despite training-time safety alignment, these models remain vulnerable to jailbreak attacks. In this work, we first highlight an important safety gap to describe that alignment achieved solely through safety training may be insufficient against jailbreak attacks. To address this vulnerability, we propose IMMUNE, an inference-time defense framework that leverages a safety reward model through controlled decoding to defend against jailbreak attacks. Additionally, we provide a mathematical characterization of IMMUNE, offering insights on why it improves safety against jailbreaks. Extensive evaluations on diverse jailbreak benchmarks using recent MLLMs reveal that IMMUNE effectively enhances model safety while preserving the model's original capabilities. For instance, against text-based jailbreak attacks on LLaVA-1.6, IMMUNE reduces the attack success rate by $57.82\\%$ and $16.78\\%$ compared to the base MLLM and state-of-the-art defense strategy, respectively.", + "bbox": [ + 88, + 378, + 485, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 702, + 220, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-modal large language models (MLLMs) have made remarkable progress in vision-language reasoning tasks such as visual question-answering [1, 3, 71] and image captioning [1, 23]. However, ensuring their outputs are safe and free from discrimination, disinformation, and harmful content is important for their adoption. Safety alignment through reinforcement learning from human feedback (RLHF) has shown promise in aligning model behavior with social norms to prevent harmful outputs [63, 78, 88]. Despite these sig", + "bbox": [ + 89, + 727, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nificant alignment efforts, recent studies reveal that MLLMs remain vulnerable to jailbreaking attacks, where malicious image-text pairs bypass safety mechanisms, raising safety concerns [19, 32, 36, 48, 54, 59, 60, 65].", + "bbox": [ + 511, + 347, + 906, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Why training-time safety alignment might not be enough? Fine-tuning-based safety alignment approaches rely on a static prompt distribution, leaving them inherently vulnerable to jailbreak attacks. These attacks exploit this limitation by solving a prompt (text or image) optimization problem to craft adversarial prompts that bypass the safety mechanisms of aligned models [54, 90]. In this work, we argue that safety finetuning against jailbreak attacks while keeping model capabilities may be inherently challenging given that the adversary can find a prompt that bypasses the safety mechanisms of the model. This underscores the fundamental limitation of relying solely on training-time safety alignment procedures to ensure robustness. These leads us to ask: Is it possible to design a safeguarding strategy to defend against given possible adversarial prompts?", + "bbox": [ + 511, + 411, + 908, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we take a step towards answering this question in the affirmative. We propose shifting from training-time alignment to inference-time safeguarding. Although initial efforts in this direction [14, 67] show promise, they exhibit weak empirical performance. Further, Weng et al. [69] note that adding a safety prompt to the input [67] can lead to overly cautious behavior, resulting in generic responses like, \"I am sorry...\" even for benign queries (see Figure 1), thereby affecting model utility. In real-world applications, an ideal jailbreaking defense should not only minimize the success rate of attacks but also maintain high utility by preserving model performance on legitimate queries. Moreover, existing inference-time approaches lack mathematical characterization, relying primarily on heuristic-based methods. This gap motivates the need for robust inference-time safeguarding mechanisms, which we address in this paper.", + "bbox": [ + 511, + 638, + 910, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our key idea and proposed approach. We present IMMUNE,", + "bbox": [ + 511, + 885, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Equal advising.", + "bbox": [ + 109, + 887, + 199, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25038", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/46403663e70204d7a71a27780e55725a487836daee282465f4c76f052531486d.jpg", + "image_caption": [ + "Figure 1. Qualitative Evaluation (Left). Given an image of a house generated by stable diffusion [58] and perturbed by adversarial noise [54], along with a malicious user query asking for steps to \"break into and rob a house\", we visualize responses from the base model and various inference-time defense strategies, including CoCA [14] and AdaShield [67]. We observe that all compared defense strategies are misled into generating harmful content. In contrast, our proposed inference-time safety alignment framework, IMMUNE, effectively rejects the user query, citing its unethical nature. This evaluation underscores the importance of inference-time alignment in preventing harmful responses. We visualize additional responses in the Appendix. Quantitative Evaluation (Right). To empirically validate the effectiveness of IMMUNE, we compare the attack success rates and model utility across various state-of-the-art defense strategies. A lower attack success rate reflects improved safety in generated outputs. Our results indicate that IMMUNE substantially lowers the attack success rate compared to other baselines. Additionally, IMMUNE not only strengthens model safety but also preserves the model's original utility, as demonstrated by its performance on MM-Vet benchmark [77]. We note that an ideal defense strategy should have a low attack success rate while maintaining high utility, i.e., towards the bottom right corner of the plot. Refer to Section 6 for further details." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 529, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8e2f14877fa4301de8388ba74ff800efafe3ebe60349a58f78a77dc74bb73277.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 88, + 903, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a framework that reframes jailbreak defense as an inference-time alignment problem. Our approach is built on two key components: First, we show that jailbreaks can be effectively mitigated by aligning the MLLM during inference using controlled decoding [7, 47], assuming access to a safety-aware utility or reward function (without loss of generality). Second, we mathematically formalize this defense as an inference-time alignment problem within the KL-regularized reinforcement learning framework [50], proving its robustness against adversarial prompt distributions. As illustrated in Figure 1, our comprehensive evaluations demonstrate that IMMUNE outperforms state-of-the-art inference-time defense strategies, such as AdaShield [67] and CoCA [14], while maintaining high model utility. Our key contributions are as follows:", + "bbox": [ + 88, + 472, + 482, + 683 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Insufficiency of safety finetuning for MLLMs. We first argue that the current alignment achieved solely via safety training/fine-tuning with RLHF is insufficient to defend against jailbreak attacks, thereby highlighting the need for additional safeguards.", + "- Improving safety via inference time alignment. To address these safety concerns, we propose IMMUNE, an inference-time framework for defense against jailbreak attacks through controlled decoding using a safety reward model. We provide a characterization of our defense mechanism by framing it as an alignment problem, offering mathematical guarantees against jailbreak attacks.", + "- Empirical evaluations. We empirically validate the efficacy of our proposed defense strategy, IMMUNE," + ], + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "across a range of jailbreak benchmarks, including MM-SafetyBench [38], FigStep [16], visual adversarial attacks [54], and text-based jailbreak attacks [40]. Our evaluations demonstrate significant improvements on several recent MLLMs, such as LLaVA-1.5 [34], LLaVA-1.6 [35], MiniGPT-4 [86], and Qwen-VL [4].", + "bbox": [ + 526, + 472, + 908, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 585, + 661, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Jailbreak attacks. Jones et al. [27] frames jailbreaking large language models (LLMs) as a discrete optimization problem. They search for the optimal suffix that, when appended to the original prompt, can greedily generate the desired harmful output. Zou et al. [90] demonstrated that model inputs could be manipulated based on gradients to force-generate responses to harmful queries. Zhu et al. [87] introduced a method to generate an adversarial suffix iteratively. This approach ensures that the optimized suffix remains meaningful while bypassing perplexity-based filters. Other works that proposed jailbreaking by manipulating gradients include Andriushchenko et al. [2], Geisler et al. [15], Hayase et al. [21], Mangaokar et al. [42], Sitawarin et al. [61], Wang et al. [66]. Another line of work, exemplified by Zhang et al. [82], involves optimizing the prompt iteratively until the distribution of output tokens results in generating harmful responses. For example, Zhang et al. [82] showed that adversaries can break safety alignment by forcing the target LLM to select a lower-ranked output token, thereby generating toxic con", + "bbox": [ + 511, + 613, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25039", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tent. Additional works following this domain include Du et al. [13], Guo et al. [18], Huang et al. [25], Zhao et al. [83], Zhou et al. [85]. Recent work has also focused on jailbreaking models through iterative interactions with the model [8, 45]. Finally Qi et al. [56] demonstrated that even benign finetuning of the models may lead to the removal of model safety safeguards. Similar to LLMs, jailbreaking MLLMs can be achieved by either manipulating the visual input [16, 37, 54], the text prompt [40], or a combination of both [75]. Qi et al. [54] demonstrated that adding adversarial noise, which is imperceptible to humans and guided by the model's input gradients, to the input image can effectively deceive MLLMs. Recently, there has been substantial research studying the robustness of MLLMs to adversarial images [12, 19, 49, 59, 60, 84]. Another line of work [16, 37] explores embedding harmful content directly into images via typography or text-to-image tools. Recently, Luo et al. [40] showed that text-based jailbreak prompts [39, 72, 79, 90] developed for LLMs can be effectively transferred to MLLMs and are often more potent than image-based attacks. Additionally, Ying et al. [75] introduced a jail-breaking framework that simultaneously perturbs both visual and textual modalities. See also the survey by Gu et al. [17] on more works on prompt-based attacks against MLLMs.", + "bbox": [ + 89, + 90, + 485, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Defense via safety alignment (fine-tuning). Defense approaches typically aim to enhance the safety in MLLMs by either fine-tuning the model on a dataset of safe image-text pairs [89] or adversarial training [89]. For instance, Zong et al. [89] proposed a safety instruction-following dataset VLGuard for safeguarding MLLMs. DRESS [10] leverages natural language feedback (NLF) from LLMs to improve the alignment and interactions within MLLMs. MLLM-Protector [53] proposed a plug-and-play strategy combining a lightweight harm detector to identify the harmful response and a detoxifier to correct harmful outputs. However, fine-tuning-based approaches suffer from (i) the requirement of a large annotated dataset consisting of safe image-text pairs, (ii) being computationally intensive, and (iii) requiring access to model weights. Qi et al. [55] highlighted that safety alignment primarily affects the first few output tokens leaving models susceptible to simple jailbreak attacks. Zou et al. [91] introduced circuit-breaking techniques to enhance model robustness. Yu et al. [76] proposed improving the safety of MLLMs by aligning with human preferences through direct preference optimization [57].", + "bbox": [ + 89, + 459, + 486, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Defense via inference time alignment (controlled decoding). Khanov et al. [28] proposed a reward-guided inference time alignment. Mudgal et al. [47] proposed controlled decoding through a value function for the reward to guide the language model to more safe outcomes. Chakraborty et al. [7] improved the value function by bridging the gap between the value of the decoding policy and the reference policy. Li et al. [33] proposed leveraging self-evaluation and rewind", + "bbox": [ + 89, + 779, + 486, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mechanisms in LLMs to generate safe responses. Gong et al. [16] proposed concatenating a fixed safety prompt with the malicious user query to evade jailbreaking. Wang et al. [67] proposed refining the safety prompt iteratively rather than using a fixed template. Mehrabi et al. [43] introduced an interactive approach to improve robustness against adversarial attacks. Zhang et al. [81] proposed mutating untrusted inputs to generate variants, identifying attack samples by examining response discrepancies among the variants. In another approach, Xu et al. [73] detected adversarial images by analyzing cross-modal similarity between harmful queries and perturbed images. Additionally, Gao et al. [14] suggested improving safety by calibrating MLLM output distribution. However, existing approaches compromise model utility, leading to overly cautious responses (Figure 1).", + "bbox": [ + 511, + 90, + 908, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Formulation", + "text_level": 1, + "bbox": [ + 513, + 321, + 720, + 337 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 345, + 653, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Large Language Models (MLLMs) are autoregressive text generation models that can process multiple modalities, such as image [86], video [80], and audio [41], in addition to text. In this work, we consider input to MLLM in the form of an image and a text prompt. Mathematically, we denote a MLLM as a mapping $\\pi_{\\mathrm{safe}}: \\mathcal{I} \\times \\mathcal{V}^N \\to \\mathcal{V}^M$ , where $\\mathcal{I}$ represents the image space, and $\\mathcal{V}$ denotes the vocabulary set. The input to the MLLM $\\pi_{\\mathrm{safe}}$ consists of an image $\\mathbf{I} \\in \\mathcal{I}$ and a sequence of tokens $\\mathbf{x} := \\{x_1, x_2, \\dots, x_N\\}$ (referred as prompt), where each token $x_i \\in \\mathcal{V}$ . Given $\\mathbf{x}_{\\mathrm{input}} := [\\mathbf{I}, \\mathbf{x}]$ , the MLLM $\\pi_{\\mathrm{safe}}(\\cdot | \\mathbf{x}_{\\mathrm{input}})$ generates the output sequence $\\mathbf{y} := \\{y_1, y_2, \\dots, y_M\\}$ , where $y_i \\in \\mathcal{V}$ , in a token by token fashion.", + "bbox": [ + 511, + 369, + 908, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Jailbreaking MLLMs. Despite safety tuning efforts aimed at preventing MLLMs from generating harmful content, jailbreak attacks as formulated in Qi et al. [54], Zou et al. [90] can still make the safety aligned model $\\pi_{\\mathrm{safe}}$ to generate harmful outputs with the help of an adversarial prompt $\\mathbf{x}_{\\mathrm{adv}}$ obtained by solving", + "bbox": [ + 511, + 569, + 910, + 660 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathrm {a d v}} := \\arg \\min _ {\\mathbf {q} \\in \\mathcal {B}} \\sum_ {i = 1} ^ {m} - \\log \\left(\\pi_ {\\text {s a f e}} \\left(y _ {i} \\mid \\left[ \\mathbf {x} _ {\\text {i n p u t}}, \\mathbf {q} \\right]\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 670, + 908, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{B}$ is a constraint on the adversarial prompts and $\\{y_i\\}_{i=1}^m$ are few initial tokens of intended harmful response. For example, in Figure 1 (Left), we show that given an image of a house generated by stable diffusion [58] and then perturbed with adversarial noise, along with a malicious user query to explain \"How to break into and rob a house\", the model is forced to generate unsafe responses, effectively bypassing its safety alignment.", + "bbox": [ + 511, + 717, + 911, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Safeguard Against Jailbreaks", + "text_level": 1, + "bbox": [ + 513, + 847, + 777, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the definition of adversarial prompt in (1), we observe that $\\mathbf{x}_{\\mathrm{adv}}$ can lead the MLLM model to generate a harmful", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25040", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "response, $\\mathbf{y}_{\\mathrm{unsafe}} \\sim \\pi_{\\mathrm{safe}}(\\cdot|\\mathbf{x}_{\\mathrm{adv}})$ , despite the training time safety alignment of $\\pi_{\\mathrm{safe}}$ . This raises the question: Why do insufficiencies exist in safety-aligned models that make them vulnerable to attacks, and how can we mitigate them? To answer this question, we start by noticing that an attacker's goal is to learn an adversarial prompt distribution $p_{\\mathrm{adv}}$ by maximizing some \"success metric\" subject to a \"cost metric,\" which is general and covers existing attacks such as projected gradient descent (PGD) attacks [15], greedy coordinate gradient (GCG) attacks [87, 90], universal adversarial trigger (UAT) attacks [44, 64], and red teaming attacks [22, 30, 52, 70]. Mathematically, an adversary can solve the following optimization problem to learn an adversarial prompt:", + "bbox": [ + 89, + 90, + 486, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p _ {\\mathrm {a d v}} := \\arg \\min _ {p} \\mathbb {E} _ {\\mathbf {q} \\sim p (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}), \\mathbf {y} \\sim \\pi_ {\\text {s a f e}} (\\cdot | \\mathbf {q})} \\left[ R _ {\\text {s a f e}} (\\mathbf {x} _ {\\text {i n p u t}}, \\mathbf {y}) \\right] \\\\ + \\beta \\operatorname {d i v} \\left(p (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}), p _ {0} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}})\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 297, + 482, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the \"cost metric\" term div $(p(\\cdot |\\mathbf{x}_{\\mathrm{input}}),p_0(\\cdot |\\mathbf{x}_{\\mathrm{input}}))$ helps keep the adversarial prompt close to a given prompt distribution, $p_0$ . We can think of $p_0$ as a given prompt distribution that will sample $\\mathbf{x}_{\\mathrm{input}}$ only with high likelihood and have minimal probability of sampling other prompts. Then, the divergence part of the objective in (2) will make sure the adversarial prompts are closer to original prompts $\\mathbf{x}_{\\mathrm{input}}$ , which will keep them interpretable and not completely out of distribution [88]. The formulation in (2) is also in accordance with the existing literature on adversarial attacks in machine learning [6, 24] because it is just trying to ensure that adversarial prompts $\\mathbf{x}_{\\mathrm{adv}}$ looks like similar to $\\mathbf{x}_{\\mathrm{input}}$ . The regularization parameter $\\beta >0$ would control this cost metric in (2).", + "bbox": [ + 89, + 352, + 483, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training-time defense. We note that learning a safe and robust model is inherently a min-max optimization problem given by", + "bbox": [ + 89, + 566, + 483, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {p} \\max _ {\\pi} \\mathbb {E} _ {\\mathbf {q} \\sim p (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}), \\mathbf {y} \\sim \\pi (\\cdot | \\mathbf {q})} \\left[ R _ {\\text {s a f e}} (\\mathbf {x} _ {\\text {i n p u t}}, \\mathbf {y}) \\right. \\\\ + \\beta \\operatorname {d i v} (p (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}), p _ {0} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}})) \\\\ \\left. - \\alpha \\operatorname {d i v} \\left(\\pi (\\cdot | \\mathbf {q}), \\pi_ {\\text {r e f}} (\\cdot | \\mathbf {q})\\right) \\right], \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 622, + 482, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the inner maximization captures the safeguarding of the model, e.g., through safety alignment. The safety alignment in recent literature [50, 57] only focused on solving the inner max in (3) for a given prompt distribution (say $p_0$ ) to obtain $\\pi_{\\mathrm{safe}}$ which is safe only under the prompt distribution $\\mathbf{x} \\sim p_0$ . An adversary exploits it and learns an adversarial distribution $\\rho_{\\mathrm{adv}}$ breaking the alignment and generating harmful responses. This is likely the reason why adversarial attacks on images (and other modalities) have remained challenging to address over the past decade.", + "bbox": [ + 89, + 696, + 483, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Key insight for tractable inference-time defense. We reframe the optimization from $\\min_p\\max_\\pi$ to $\\max_{\\pi}\\min_{p}$ by letting the safeguard be applied after the attacker has chosen", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "their jailbreak prompt. While the two formulations are equivalent for convex problems, this reformulation may lead to an improvement in defense due to the minimax duality gap for general non-convex problems. Additionally, this formulation allows an efficient solution of the $\\max_{\\pi}$ component during inference, leveraging recent advances in inference-time alignment methods through controlled decoding [7, 47]. Notably, inference-time alignment enables prompt-specific safety adjustments, which are challenging to implement during training. This capability is enabled by the user (or attacker) interaction protocol with MLLM. Specifically, the attacker sets the adversarial prompt, which remains unchanged throughout the process, providing a fixed input for the MLLM. This scenario ultimately reduces the challenge of learning a safe and optimal policy within the predefined adversarial prompt environment, effectively solving the problem $\\max_{\\pi}$ in inference. We will provide a detailed method in the next section.", + "bbox": [ + 511, + 90, + 908, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Inference Time Safety Alignment", + "text_level": 1, + "bbox": [ + 511, + 359, + 815, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we propose improving the safety of response generation by using alignment through the controlled decoding procedure proposed in [7, 47]. Our key insight is that well-trained safety rewards, denoted by $R_{\\mathrm{safe}}(\\mathbf{x}, \\mathbf{y})$ , are often accessible online, for instance, RewardBench [29]. This safety reward will give high scores for safe text and low scores for unsafe text. We remark here that the reward model itself is likely not ideal and subject to adversarial attack. With this understanding, we propose to solve the following decoding problem at inference time for each token $t$ :", + "bbox": [ + 511, + 385, + 908, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\pi_ {\\text {s a f e - d e c}} ^ {*} (\\cdot | \\mathbf {s} _ {t}) := \\arg \\max _ {\\pi} \\mathbb {E} _ {z \\sim \\pi (\\cdot | \\mathbf {s} _ {t})} [ Q _ {\\text {s a f e}} (\\mathbf {s} _ {t}, z) ] \\tag {4} \\\\ - \\alpha \\mathrm {K L} (\\pi (\\cdot | \\mathbf {s} _ {t}) | | \\pi_ {\\text {s a f e}} (\\cdot | \\mathbf {s} _ {t})), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 542, + 906, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{s}_t\\coloneqq [\\mathbf{x}_{\\mathrm{adv}},\\mathbf{y}_{< t}]$ and $\\mathbf{y}_{< t} = [y_0,y_1,\\dots ,y_{t - 1}]$ capture the tokens generated so far. In (4), the term $Q_{\\mathrm{safe}}(\\mathbf{s}_t,z)$ is the action-value function defined as:", + "bbox": [ + 511, + 590, + 908, + 637 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {\\text {s a f e}} \\left(\\mathbf {s} _ {t}, z\\right) = \\mathbb {E} _ {\\tau \\sim \\rho_ {\\text {s a f e}} (\\cdot | [ \\mathbf {s} _ {t}, z ])} \\left[ R _ {\\text {s a f e}} \\left([ \\mathbf {s} _ {t}, z ], \\tau\\right) \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 643, + 906, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau := [z_{t+1}, z_{t+2}, \\dots, z_{eos}]$ denotes the remaining trajectory starting from $[\\mathbf{s}_t, z]$ , $\\rho_{\\mathrm{safe}}(\\tau | [\\mathbf{s}_t, z]) := \\prod_{i=t+1}^{\\mathrm{eos}} \\pi_{\\mathrm{safe}}(z_i | [\\mathbf{s}_t, z])$ is the trajectory level distribution, where $eos$ denotes end of sentence token, and $\\tau$ denotes the sequence of tokens generated by the model starting from $[\\mathbf{s}_t, z]$ . The closed form expression of (4) is given by [47, Theorem 2.1]:", + "bbox": [ + 511, + 667, + 908, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {\\text {s a f e - d e c}} ^ {*} (z | \\mathbf {s} _ {t}) = \\frac {\\pi_ {\\text {s a f e}} (z | \\mathbf {s} _ {t})}{Z} \\exp \\left(\\frac {Q _ {\\text {s a f e}} (\\mathbf {s} _ {t} , z)}{\\alpha}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 780, + 906, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Z$ is the normalizing constant. For a given adversarial prompt $\\mathbf{x}_{\\mathrm{adv}}$ , we propose to follow the decoding policy derived in (6) at the inference time to generate a safe response. Our proposed algorithm is summarized in Algorithm 1.", + "bbox": [ + 511, + 819, + 906, + 880 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1https://huggingface.co/spaces/allenai/reward-bench", + "bbox": [ + 531, + 886, + 808, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25041", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 IMMUNE: Safety against Jailbreaking attacks via Inference-time alignment", + "bbox": [ + 91, + 90, + 483, + 121 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Input: Base MLLM $\\pi_{\\mathrm{safe}}$ , safe reward model $R_{\\mathrm{safe}}$ , adversarial input $\\mathbf{x}_{\\mathrm{adv}}$ , number of tokens sampled $k$ , alignment parameter $\\alpha$ , vocabulary set $\\mathcal{V}$ .", + "2: for $t = 0, \\dots, T$ do", + "3: Current state: $\\mathbf{s}_t = [\\mathbf{x}_{\\mathrm{adv}},\\mathbf{y}_{< t}]$ , where $\\mathbf{x}_{\\mathrm{adv}}$ is adversarial input and $\\mathbf{y}_{< t} = [y_0,y_1,\\dots ,y_{t - 1}]$", + "4: Sample top-k tokens using base model $\\pi_{\\mathrm{safe}}$ and store as $\\hat{\\mathcal{V}} = \\{z_i : z_i \\sim \\pi_{\\mathrm{safe}}(\\cdot|\\mathbf{s}_t)\\}_{i=1}^k$", + "5: for $z\\in \\mathcal{V}$ do", + "6: Evaluate $Q_{\\mathrm{safe}}$ as in Equation (5)", + "7: Compute decoding score for token $z$ : $g_{z} = \\frac{1}{\\alpha} Q_{\\mathrm{safe}}([s_t,z]) + \\log \\pi_{\\mathrm{saf}}(z|s_t)$", + "8: Distribution over $\\hat{\\mathcal{V}}$ : $f(z|\\mathbf{s}_t) = \\frac{\\exp(g_z)}{\\sum_{z'\\in\\hat{\\mathcal{V}}}\\exp(g_{z'})}$", + "9: Next token $y_{t}\\sim f(\\cdot |\\mathbf{s}_{t})$", + "10: Next state $\\mathbf{s}_{t + 1}\\gets [\\mathbf{s}_t,y_t]$", + "11: Return $\\mathbf{y} = [y_0, \\dots, y_T]$" + ], + "bbox": [ + 99, + 125, + 486, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Theoretical Insights", + "text_level": 1, + "bbox": [ + 89, + 407, + 282, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we formally derive a bound on the suboptimality of our proposed decoding-based approach under an adversarial prompt distribution. Given a safe prompt distribution $p_0$ and an adversarial prompt distribution $p_{\\mathrm{adv}}$ , the sub-optimality gap of our algorithm can be defined as", + "bbox": [ + 89, + 431, + 485, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Delta_ {\\text {s u b - g a p}} \\left(\\mathbf {x} _ {\\text {i n p u t}}\\right) \\tag {7} \\\\ := \\underset { \\begin{array}{c} \\mathbf {x} \\sim p _ {0} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}) \\\\ \\mathbf {y} \\sim \\rho_ {*} (\\cdot | \\mathbf {x}) \\end{array} } {\\mathbb {E}} \\left[ R _ {\\text {s a f e}} (\\mathbf {x}, \\mathbf {y}) \\right] - \\underset { \\begin{array}{c} \\mathbf {x} \\sim p _ {\\text {a d v}} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}) \\\\ \\mathbf {y} \\sim \\rho_ {\\text {s a f e - d e c}} (\\cdot | \\mathbf {x}) \\end{array} } {\\mathbb {E}} \\left[ R _ {\\text {s a f e}} (\\mathbf {x}, \\mathbf {y}) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 520, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\rho_{\\mathrm{safe - dec}}$ is the trajectory level distribution corresponding to token level $\\pi_{\\mathrm{safe - dec}}^*$ and $\\rho_*$ is the trajectory level optimal distribution for safety reward. The suboptimality gap in (7) captures how safe our proposed decoding-based approach is in generating responses to adversarial prompts. The suboptimality gap in (7) is similar but different to the one utilized in [5] because here, the goal is to capture the impact of adversarial prompt distribution specifically. We establish an upper bound in the following Theorem 1 on the suboptimality gap defined in (7).", + "bbox": [ + 89, + 585, + 483, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1. Let $R_{safe}(\\mathbf{x},\\mathbf{y}) \\leq R_{max}$ , $p_0$ is a given prompt distribution, $p_{adv}$ adversarial prompt distribution, $\\rho_{*}(\\cdot |\\mathbf{x})$ denotes the optimal trajectory level distribution for the safe reward, and $\\rho_{safe}(\\cdot |\\mathbf{x})$ denote the given training time safety aligned trajectory level distri", + "bbox": [ + 106, + 766, + 468, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bution. Then, it holds that", + "bbox": [ + 529, + 90, + 707, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Delta_ {\\text {s u b - g a p}} (\\mathbf {x} _ {\\text {i n p u t}}) \\leq R _ {\\max } \\sqrt {K L (p _ {0} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}) \\| p _ {\\text {a d v}} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}))} \\\\ + \\alpha K L \\left(\\rho_ {*} \\left(\\cdot | \\mathbf {x} _ {\\text {i n p u t}}\\right) \\| \\rho_ {s a f e} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}})\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 116, + 892, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha > 0$ is a regularization parameter balancing the KL-divergence term in the policy alignment.", + "bbox": [ + 529, + 172, + 887, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof Outline and Insights. The proof of Theorem 1 (see Appendix E for details) relies on decomposing the suboptimality gap term $\\Delta_{\\mathrm{sub - gap}}$ into two components, $\\Delta_1$ and $\\Delta_2$ such that $\\Delta_{\\mathrm{sub - gap}} = \\Delta_1 + \\Delta_2$ , with bounds derived from total variation distance and Pinsker's inequality, respectively. We can write the decomposition as", + "bbox": [ + 511, + 224, + 906, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_{1}:= \\mathop{\\mathbb{E}}_{\\substack{\\mathbf{x}\\sim p_{0}(\\cdot |\\mathbf{x}_{\\text{input}})\\\\ \\mathbf{y}\\sim \\rho_{*}(\\cdot |\\mathbf{x})}}\\left[R_{\\text{safe}}(\\mathbf{x},\\mathbf{y})\\right] - \\mathop{\\mathbb{E}}_{\\substack{\\mathbf{x}\\sim p_{\\text{adv}}(\\cdot |\\mathbf{x}_{\\text{input}})\\\\ \\mathbf{y}\\sim \\rho_{*}(\\cdot |\\mathbf{x})}}\\left[R_{\\text{safe}}(\\mathbf{x},\\mathbf{y})\\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 328, + 898, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_{2}:= \\mathop{\\mathbb{E}}_{\\substack{\\mathbf{x}\\sim p_{\\text{adv}}(\\cdot |\\mathbf{x}_{\\text{input}})\\\\ \\mathbf{y}\\sim \\rho_{*}(\\cdot |\\mathbf{x})}}[R_{\\text{safe}}(\\mathbf{x},\\mathbf{y})] - \\mathop{\\mathbb{E}}_{\\substack{\\mathbf{x}\\sim p_{\\text{adv}}(\\cdot |\\mathbf{x}_{\\text{input}})\\\\ \\mathbf{y}\\sim \\rho_{\\text{safe -dec}}(\\cdot |\\mathbf{x})}}[R_{\\text{safe}}(\\mathbf{x},\\mathbf{y})].\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 368, + 903, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we provide the interpretation of the mathematical result in Theorem 1 and intuitive insights about different terms as follows:", + "bbox": [ + 511, + 417, + 905, + 462 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Practical Implications of the Bound: Theorem (1) shows that sub-optimality gap upper bound is minimized when $p_{\\mathrm{adv}}$ remains close to $p_0$ (low KL divergence) and also when $\\rho_{\\mathrm{safe}}$ (which is our reference policy for the inference time alignment problem) aligns closely with $\\rho_*$ (low policy KL divergence). This result can justify the use of KL-regularized objectives in safety-required domains, especially under adversarial prompt conditions.", + "(2) Role of Regularization Parameter $\\beta$ : The adversarial prompt distribution in (2), $p_{\\mathrm{adv}}$ , is defined as a solution to a KL-regularized optimization problem that minimizes expected reward $R_{\\mathrm{safe}}$ relative to $p_0$ , with regularization parameter $\\beta$ . The lower values of $\\beta$ allow $p_{\\mathrm{adv}}$ to diverge further from $p_0$ , potentially increasing $\\Delta_1$ .", + "(3) Impact of $\\alpha$ on $\\Delta_{2}$ : The value of $\\alpha$ affects the degree to which the policy $\\rho_{\\mathrm{safe - dec}}$ is incentivized to align with the safety policy $\\rho_{*}$ . The lower values of $\\alpha$ prioritize safety more heavily, thereby reducing $\\Delta_{2}$ .", + "(3) In proof, $\\Delta_{1}$ expresses how far an adversary can exploit the adversarial prompt distribution to attack the aligned model. In $\\Delta_{2}$ , we show how, with inference time alignment, we can achieve safety even under the adversarial prompt distribution owing to the closed form of the KL regularized RLHF objective. We note that Mroueh [46, Proposition 3] also establishes upper bounds on the difference in expected rewards between two distributions in a continuous setting, relying on tail assumptions (e.g., sub-Gaussian, sub-Gamma) to control deviations. In contrast, our Theorem 1 operates" + ], + "bbox": [ + 511, + 468, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25042", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "entirely in a discrete setting and explicitly decomposes suboptimality (difference in expected rewards defined in (7)) into contributions from both the prompter and the language model. Furthermore, we leverage the optimality of RLHF to derive tighter upper bounds, which is not used in Mroueh [46].", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark: Note that the theoretical guarantees discussed are conditional upon a specific prompt design structure, which originates from an adversarial prompter targeting a given unsafe query. This inherently prevents us from being able to obtain a similar guarantee for training time safety fine-tuning since the adversarial prompts would not be known in advance. In our experiments, we will not make these assumptions, as we experiment with several recent static safety benchmarks.", + "bbox": [ + 89, + 188, + 483, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 325, + 222, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Experimental Details", + "text_level": 1, + "bbox": [ + 89, + 351, + 290, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Jail-break Datasets. We conduct a comprehensive evaluation by considering both text-based and image-based jailbreak attacks:", + "bbox": [ + 89, + 375, + 483, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1. Text-based Attacks. For text-based attacks, we use the JailbreakV-28K dataset [40]. This dataset integrates effective text-based jailbreak prompts for LLMs with various types of image data. Specifically, the text prompts include template-based, persuasive, and logic-driven approaches, while the image input may consist of noise, be blank, contain natural images, or be generated by stable diffusion. We evaluate all possible combinations of image and text pairs to ensure thorough evaluation.", + "bbox": [ + 89, + 424, + 483, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2. Image-based Attacks. For image-based attacks, we evaluate the following datasets:", + "bbox": [ + 89, + 564, + 483, + 594 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- MM-SafetyBench [37]: This benchmark assesses MLLM safety across 13 commonly prohibited scenarios and behaviors specified in usage guidelines from OpenAI [1] and Meta [26]. Each image in this dataset is associated with three types of image inputs: (1) Stable-diffusion images (SD), which are generated using stable diffusion [58] and directly relate to the query; (2) Typography (TYPO) images, containing optical character recognition representations of text from the malicious query; and (3) SD+TYPO images, which combine stable diffusion visuals with typographic subtitles.", + "- FigStep [16]: This attack converts harmful instructions into images using typography. Specifically, malicious queries on 10 forbidden topics from MM-SafetyBench are first paraphrased to start with nouns like \"Steps to\", \"List of\", or \"Methods to\", and then the instructions are presented as text within images.", + "- Visual Adversarial Attacks [54]: Qi et al. proposed an optimization-based attack, where the malicious user generates adversarial images by introducing carefully crafted" + ], + "bbox": [ + 89, + 598, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "perturbations to the original image, causing models to produce harmful content.", + "bbox": [ + 526, + 90, + 906, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models Used. We perform evaluation on 5 state-of-art open-source MLLMs: LLaVA-1.5-7B [34], LLaVA-1.6-7B [35], MiniGPT-4-7B/13B [86], and Qwen-VL-Chat-7B [4]. For both LLaVA and MiniGPT, we use Vicuna [11] as the text encoder. For IMMUNE, we set the number of tokens sampled $k = 10$ and the alignment parameter $\\alpha = 1$ . We report ablations in Appendix F. For main evaluations, we use LLaMa-3.1-8B $^2$ as the safety model $R_{\\text{safe}}$ . The safety model has been trained on a mixture of preference datasets including HelpSteer2 [68], OffsetBias [51], WildGuard [20], and Magpie [74].", + "bbox": [ + 511, + 128, + 906, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Defenses. We compare IMMUNE with recent state-of-the-art inference-time jail-break defense frameworks, including FigStep [16], Adashield [67], and CoCA [14]. For a fair comparison, all baseline defense methods are evaluated on a unified test dataset using consistent metrics. Detailed descriptions of baselines are provided in the Appendix G.", + "bbox": [ + 511, + 297, + 906, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric. To assess the effectiveness of a jailbreak attack, we follow the literature [14, 40, 67] and measure the Attack Success Rate (ASR). Specifically, given a test dataset $\\mathcal{D}_{\\mathrm{unsafe}}$ of crafted jailbreak image-text pairs, the ASR quantifies the ratio of harmful responses to the total number of input queries and is defined as:", + "bbox": [ + 511, + 393, + 905, + 483 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A S R} = \\frac {1}{\\left| \\mathcal {D} _ {\\text {u n s a f e}} \\right|} \\sum_ {(\\mathbf {I}, \\mathbf {x}) \\in \\mathcal {D} _ {\\text {u n s a f e}}} \\mathbb {I} \\left[ \\mathcal {C} ^ {*} (\\mathbf {x}, \\pi_ {\\theta} (\\mathbf {I}, \\mathbf {x})) = \\text {T r u e} \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 494, + 905, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $(\\mathbf{I},\\mathbf{x})$ represents an image-text jailbreak prompt pair, $\\mathbb{I}(\\cdot)$ is an indicator function, and $\\mathcal{C}^*$ is an oracle jailbreak classifier that verifies whether the generated response $\\pi_{\\theta}(\\mathbf{I},\\mathbf{x})$ aligns with the malicious intent of the user query $\\mathbf{x}$ . Consistent with earlier works [9, 14, 40], we use both Llama-Guard-3 Jailbreak Judge and the GPT-4 as the oracle jailbreak classifier. Hence, a lower ASR value indicates a more robust defense against jailbreak attacks. To account for randomness in output generation, we sample three independent responses for each query and consider the model successfully jailbroken if any one of the three responses is flagged as jailbroken by the oracle classifier.", + "bbox": [ + 511, + 547, + 908, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Safety Evaluation results", + "text_level": 1, + "bbox": [ + 511, + 739, + 741, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on Text-based attacks. In Table 1, we present the results of this evaluation for different combinations of image types and text prompts. To be specific, the malicious user queries can be template-based, persuasive in nature, or logic-driven, while the image input consists of random noise, be blank, contain natural images, or be generated by stable diffusion. For this evaluation, we used the Llama-Guard-3", + "bbox": [ + 511, + 770, + 906, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{2}\\mathrm{LxzGordon/URM-LLaMa - }3.1 - 8\\mathrm{B}$", + "bbox": [ + 529, + 886, + 718, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25043", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/db2f639300afcce56235b7b6a8c942382c45ecc3cf8fa4e7a5a2fce0f6f1602d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDefense StrategyNoiseSDNatureBlankAverage
TemplatePersuadeLogicTemplatePersuadeLogicTemplatePersuadeLogicTemplatePersuadeLogic
LLaVA-1.6Original66.1237.4578.5867.3437.5677.5769.2340.7882.6166.6739.4581.6060.27
FigStep [16]61.1239.2762.1662.3440.1854.0563.4135.2256.7661.0938.4352.7051.64
AdaShield [67]38.420.0011.0838.131.5618.1339.293.2219.1442.780.4816.1219.23
CoCA [14]61.2339.1762.1661.3441.2852.7063.1135.2255.4161.0937.3652.7051.37
IMMUNE (Ours)5.230.000.008.140.000.008.450.000.004.670.000.002.45
LLaVA-1.5Original58.1239.4476.5661.4738.3574.5558.2242.1177.5759.3340.2674.5556.39
FigStep [16]64.1728.3462.1662.2332.1868.9158.3937.2772.9761.0931.4268.9152.46
AdaShield [67]32.140.005.4131.360.004.2330.220.005.4129.670.0010.8112.86
CoCA [14]59.1826.3435.1358.4230.2935.1349.2232.1128.3759.0723.3937.8339.87
IMMUNE (Ours)9.230.000.008.140.000.001.470.000.005.320.000.002.10
MiniGPT-4-7BOriginal75.2354.1283.6272.4437.3385.6373.1845.2786.6472.0955.4286.6467.15
FigStep [16]47.2217.1832.4342.135.3724.3237.298.4518.9243.0817.4143.2427.74
AdaShield [67]47.2322.1432.2441.3718.4627.2058.1512.2830.2246.0919.4332.2432.21
CoCA [14]35.1818.2722.9740.3421.1431.0835.2918.4227.0348.1121.3640.5429.74
IMMUNE (Ours)18.236.1244.5911.348.2729.7217.188.4527.0216.0910.3743.2418.34
MiniGPT-4-13BOriginal74.1847.1290.6780.2353.3485.6379.2952.4583.7879.0753.2287.6570.62
FigStep [16]44.1320.2858.1143.2218.4747.3040.3627.1448.6542.0919.4258.1137.41
AdaShield [67]63.2223.1859.4464.4731.1274.5569.1527.3568.9165.0923.4153.3950.55
CoCA [14]35.2921.3436.4940.1822.2740.5435.1220.4135.1448.2221.3652.7033.21
IMMUNE (Ours)25.1718.2959.4429.3419.4157.4229.2223.1858.4323.4723.3758.4332.94
Qwen-VLOriginal46.123.2712.0952.236.189.0753.343.427.0553.114.3615.1122.99
FigStep [16]32.180.001.3540.420.005.4130.230.005.4144.122.231.3514.42
AdaShield [67]30.170.0018.1328.341.2915.1123.281.377.0542.110.4118.1315.88
CoCA [14]29.126.2713.5127.347.185.4127.291.4213.5128.4713.3613.5115.69
IMMUNE (Ours)10.272.185.4121.342.294.0318.222.355.4120.173.377.058.58
", + "bbox": [ + 91, + 88, + 906, + 390 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d37b6a0e41a8e5fb66ba28c86c4d52d76b0e8a782b090d5c8810253079b130de.jpg", + "table_caption": [ + "Table 1. Evaluation on Text-Based Jailbreak Attacks. We report the Attack Success Rate (ASR) for various baseline defense strategies across recent MLLMs on text-based jailbreak attacks [40]. In this setup, the image input may be noise, Stable Diffusion-generated (SD), natural images (Nature), or blank, while the text query is structured as either template-based, persuasive, or logic-driven prompts. Lower ASR values indicate stronger resilience against jailbreak attacks. For this evaluation, we used Llama-Guard-3 Jailbreak Judge as the jailbreak classifier. IMMUNE consistently outperforms all compared defense frameworks across all MLLMs. The best results (lowest ASR) are highlighted in bold. All values are reported in %." + ], + "table_footnote": [], + "table_body": "
ModelDefense StrategyIllegal ActivityMalware GenerationPornographyHate SpeechPhysical HarmFraudAverage
SDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPE
LLaVA-1.6Original34.0280.4178.3520.4570.4570.457.0742.0535.0320.0651.0252.0518.0163.0163.0530.0367.0472.0747.94
FigStep [16]27.8449.4879.3825.0347.7359.097.0928.0741.0330.0436.0746.0720.0221.0852.0329.0536.0855.0737.64
AdaShield [67]19.5962.8941.249.0954.5543.187.0525.0519.066.0620.0313.039.0926.0427.046.0136.0235.0824.40
CoCA [14]24.2755.6776.0117.2759.0970.459.0233.0743.0825.0142.0752.0623.0626.0457.0125.0137.0858.0739.79
IMMUNE (Ours)0.001.035.150.000.004.559.233.0718.020.000.003.024.095.035.040.002.077.073.94
LLaVA-1.5Original34.0280.4180.4143.1875.0170.4513.0256.0739.0724.0575.0269.0725.0764.0367.0532.0373.0765.0353.85
FigStep [16]25.5263.9273.2820.4375.0544.0916.0426.0538.0638.0351.0655.0121.0457.0459.0636.0862.0861.0445.54
AdaShield [67]9.281.033.090.002.270.009.067.034.047.071.042.0311.037.053.0310.011.072.084.81
CoCA [14]17.1380.4180.4125.9127.779.918.0752.0655.0610.0863.1967.0325.0745.0557.0726.0357.0859.0744.65
IMMUNE (Ours)0.461.030.000.400.0013.646.989.0919.050.523.025.045.174.037.020.000.006.043.51
MiniGPT-4-7BOriginal13.4343.3435.052.2734.0925.0827.0414.0121.066.0229.0517.053.0834.0826.025.0540.0436.0223.04
FigStep [16]19.5943.3438.146.8240.9115.9126.0119.0623.0611.0428.0816.0314.0631.0826.0712.0644.0517.0724.21
AdaShield [67]11.3435.0518.5618.1822.739.0916.6726.0315.0422.678.0426.048.0827.0914.0918.1834.0422.0720.05
CoCA [14]9.2842.2728.876.8220.4518.1819.0812.0924.076.0512.0710.0325.0416.6726.0725.068.3325.0218.92
IMMUNE (Ours)13.4622.6813.4311.3620.4518.1817.0412.0921.033.088.047.087.0214.0120.020.000.000.0011.03
MiniGPT-4-13BOriginal17.5340.2142.2713.6422.7327.2710.0415.0818.039.0522.0921.0717.0320.0235.0819.0537.0341.0923.97
FigStep [16]25.7746.3936.0822.7327.2729.5513.0216.0819.0320.0830.0720.0818.0136.0438.0821.0532.0938.0327.17
AdaShield [67]21.6523.7122.6820.4534.0925.0415.059.0215.0311.0625.0321.0114.0322.0123.0723.0528.0215.0919.84
CoCA [14]25.0350.0746.0314.0625.0343.0613.0225.018.3316.0416.6716.6716.6725.0641.6725.0641.6725.0726.13
IMMUNE (Ours)23.7822.6819.599.0918.1829.5520.0814.0815.0611.0813.0813.0210.0818.0322.086.020.000.0014.24
", + "bbox": [ + 91, + 500, + 906, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Evaluation on MM-Safety Bench. We report Attack Success Rate (ASR) for 6 categories of prohibited scenarios from MM-Safety Bench [38]. For this evaluation, we used GPT-4 as the jailbreak classifier. IMMUNE significantly reduces ASR across all categories. The best results (lowest ASR) are highlighted in bold. All values are reported in %.", + "bbox": [ + 89, + 702, + 908, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Jailbreak Judge as the oracle classifier. From our evaluation, we draw the following key insights: (1) With the original decoding, ASR is notably high for most models, reaching up to $70.62\\%$ for MiniGPT-4-13B, underscoring both the severity of the attack and the underlying serious safety concern. (2) Across all MLLMs, IMMUNE demonstrates superior performance, substantially reducing ASR compared to other competitive defense strategies. Specifically, on MiniGPT-4-13B, IMMUNE reduces ASR by $37.68\\%$ compared to the", + "bbox": [ + 89, + 761, + 485, + 897 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "original decoding and by $17.61\\%$ compared to state-of-art defense strategy AdaShield [67]. (3) We observe that logic-based attacks reveal a model-dependent variation in IMMUNE's performance. Specifically, for logic and persuade-based attacks, IMMUNE achieves an ASR of $0\\%$ across all input types on LLaVA-1.6 and LLaVA-1.5. However, on MiniGPT-4-7B, logic-based attacks with noise inputs result in a significantly higher ASR of $44.59\\%$ , compared to CoCA [14] $(22.97\\%)$ and AdaShield [67] $(32.24\\%)$ .", + "bbox": [ + 511, + 761, + 908, + 897 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25044", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a18eb2fd8db0b7db476172019592de934cf0805cf0e96f90aa8fcd9ad7a476ba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDefense StrategyIllegal ActivityHate SpeechMalware GenerationPhysical HarmFraudAdult ContentPrivacy ViolationLegal OpinionFinancial AdviceHealth ConsultationAverage
LLaVA-1.6Original11.700.0021.9211.880.0026.270.0036.0431.809.5314.68
FigStep [16]9.620.0014.232.030.0011.950.006.0513.586.245.90
AdaShield [67]3.580.005.650.000.0011.790.000.009.610.003.62
CoCA [14]13.624.296.065.978.3320.396.5727.7221.7017.5613.34
IMMUNE (Ours)12.040.000.000.000.000.000.000.000.000.001.21
LLaVA-1.5Original74.000.0083.9761.6210.4693.961.7292.3782.2984.0958.58
FigStep [16]65.531.9570.0748.473.7773.684.1371.7276.2967.5148.48
AdaShield [67]14.380.0011.5413.880.002.100.002.2012.2114.167.24
CoCA [14]44.038.2338.0422.086.4942.489.6544.1541.9430.4528.63
Ours28.210.006.302.120.000.000.004.403.841.864.23
MiniGPT-4-7BOriginal10.152.1416.022.440.0019.770.0033.6226.2111.9311.95
FigStep [16]11.630.0024.450.003.7513.780.0010.0615.963.998.32
AdaShield [67]4.290.0012.175.674.072.120.0012.128.239.645.76
CoCA [14]7.850.000.002.471.856.360.0015.6313.876.285.74
IMMUNE (Ours)7.980.009.826.140.004.430.007.855.713.594.43
MiniGPT-4-13BOriginal20.050.0030.4115.980.0042.140.0037.8146.2819.5321.17
FigStep [16]24.110.0038.4521.880.0029.570.0021.8334.5023.8119.52
AdaShield [67]18.060.007.663.930.0011.551.8228.2029.9212.2611.65
CoCA [14]24.38.3313.5718.125.5431.928.3313.9817.8318.3815.93
IMMUNE (Ours)16.312.248.275.540.0010.180.0016.0619.5315.989.48
Qwen-VLOriginal17.530.0040.2317.770.0032.380.0056.3439.7539.8224.01
FigStep [16]1.660.002.200.000.0020.870.0010.1210.080.004.56
AdaShield [67]4.050.006.499.680.0017.830.006.061.550.005.04
CoCA [14]29.530.0067.8834.490.0077.920.0088.2683.5661.9243.92
IMMUNE (Ours)6.500.004.370.000.009.930.007.935.880.003.23
", + "bbox": [ + 91, + 88, + 906, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on Image-based attacks. We evaluate IMMUNE on three standard benchmarks for image-based attacks:", + "bbox": [ + 91, + 368, + 482, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(1) MM-Safety. We report the ASR for 6 prohibited categories from MM-Safety Bench [38] across four recent MLLMs, as shown in Table 2. For response evaluation, following Liu et al. [38], we used GPT-4 as the jailbreak classifier. Our findings indicate that IMMUNE significantly reduces ASR across all MLLMs compared to state-of-art defense strategies such as CoCA [14] and AdaShield [67]. Notably, on LLaVA-1.6, IMMUNE achieves an ASR reduction of $44.0\\%$ and $20.46\\%$ compared to the base model and AdaShield [67] respectively. Across all MLLMs, we observe that attacks using typography images (TYPO) and those combining stable diffusion visuals with typography subtitles (SD-TYPO) are significantly more effective than simple stable diffusion-generated (SD) images. Additionally, samples from the \"Pornography\" category exhibit the highest ASR across all defense mechanisms.", + "bbox": [ + 91, + 400, + 483, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(2) FigStep. Table 3 presents the ASR results for various MLLMs on the FigStep dataset [16]. Consistent with results on MM-Safety, IMMUNE demonstrates superior performance over competitive baselines, reducing ASR by $6.45\\%$ and $12.13\\%$ on MiniGPT-4-13B and LLaVA-1.6, respectively, compared to CoCA [14]. Among all categories, \"Hate Speech\" and \"Privacy Violation\" have the lowest ASR, with most approaches achieving an ASR of near $0.0\\%$ .", + "bbox": [ + 89, + 642, + 482, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(3) Visual Adversarial Attacks. Finally, in Table 4, we evaluate ASR under adversarily optimized jailbreak attacks [54] across varying attack strengths, denoted by $\\epsilon$ , where higher $\\epsilon$ values correspond to stronger attacks. The \"Unconstrained\" setting represents the most challenging attack scenario. Notably, even under the \"unconstrained\" attack, IMMUNE reduces ASR by $24.19\\%$ compared to baseline [67] on LLaVA-1.6.", + "bbox": [ + 89, + 763, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional discussion on inference time of IMMUNE and its impact on model utility is deferred to Appendix F.", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/66922f099d29f539426345d32e3cdad746c29bc3c99d73e453c021dc232627b7.jpg", + "table_caption": [ + "Table 3. Evaluation on FigStep. We report the Attack Success Rate (ASR) for all categories from the FigStep benchmark [16]. For this evaluation, we used GPT-4 as the jailbreak classifier. The best results (lowest ASR) are highlighted in bold. All values are reported in %." + ], + "table_footnote": [], + "table_body": "
ModelDefense StrategyAttack Strength
ε = 16/255ε = 32/255ε = 64/255Unconstrained
LLaVA-1.6Original66.1264.0862.3164.80
FigStep [16]49.1050.3047.2251.13
AdaShield [67]35.2139.7134.0540.21
CoCA [14]45.2146.1148.0250.33
IMMUNE (Ours)13.3214.9715.1816.02
LLaVA-1.5Original68.0067.1966.2367.33
FigStep [16]53.9852.0751.658.17
AdaShield [67]27.0024.1526.9130.10
CoCA [14]44.2341.8842.9144.82
IMMUNE (Ours)12.4516.7114.3217.03
MiniGPT-4-7BOriginal53.4058.1258.7961.23
FigStep [16]47.0952.4355.2557.10
AdaShield [67]39.0042.4543.7149.63
CoCA [14]53.1960.0962.7868.29
IMMUNE (Ours)18.4220.7721.1822.95
MiniGPT-4-13BOriginal62.5766.1367.8270.11
FigStep [16]59.8161.0562.4465.79
AdaShield [67]47.8250.3751.6853.15
CoCA [14]71.8574.3175.7275.09
IMMUNE (Ours)23.4125.8226.0226.02
Qwen-VLOriginal38.1148.0348.0349.47
FigStep [16]34.2536.9240.1742.78
AdaShield [67]17.3218.2022.4125.24
CoCA [14]38.1544.7346.3649.04
IMMUNE (Ours)13.2215.0715.0716.98
", + "bbox": [ + 514, + 366, + 906, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Evaluation on Visual Adversarial Attacks [54]. We report ASR (in %) on images optimized with varying levels of adversarial noise, denoted by $\\epsilon$ , with \"unconstrained\" representing the most challenging scenario.", + "bbox": [ + 511, + 643, + 906, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 722, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce IMMUNE, an inference-time defense framework designed to protect MLLMs against jailbreak attacks. IMMUNE employs a safety-aware reward model to align the MLLM during inference, with the goal of mitigating jailbreak vulnerabilities. We present a mathematical formulation of our defense framework, framing jailbreak mitigation as an alignment problem. Through extensive experiments, we demonstrate that IMMUNE consistently and significantly outperforms competitive baselines, enhancing model safety while preserving the model utility.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25045", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 6", + "[2] Maksym Andriushchenko, Francesco Croce, and Nicolas Flammarion. Jailbreaking leading safety-aligned llms with simple adaptive attacks. arXiv preprint arXiv:2404.02151, 2024. 2", + "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 1", + "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 6", + "[5] James Beetham, Souradip Chakraborty, Mengdi Wang, Furong Huang, Amrit Singh Bedi, and Mubarak Shah. Liar: Leveraging alignment (best-of-n) to jailbreak llms in seconds. arXiv preprint arXiv:2412.05232, 2024. 5", + "[6] Nicholas Carlini, Anish Athalye, Nicolas Papernot, Wieland Brendel, Jonas Rauber, Dimitris Tsipras, Ian Goodfellow, Aleksander Madry, and Alexey Kurakin. On evaluating adversarial robustness. arXiv preprint arXiv:1902.06705, 2019. 4", + "[7] Souradip Chakraborty, Soumya Suvra Ghosal, Ming Yin, Dinesh Manocha, Mengdi Wang, Amrit Singh Bedi, and Furong Huang. Transfer q star: Principled decoding for llm alignment, 2024. 2, 3, 4, 16", + "[8] Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J Pappas, and Eric Wong. Jailbreaking black box large language models in twenty queries. arXiv preprint arXiv:2310.08419, 2023. 3", + "[9] Patrick Chao, Edoardo Debenedetti, Alexander Robey, Maksym Andriushchenko, Francesco Croce, Vikash Sehwag, Edgar Dobriban, Nicolas Flammarion, George J. Pappas, Florian Tramer, Hamed Hassani, and Eric Wong. Jailbreakbench: An open robustness benchmark for jailbreaking large language models. In NeurIPS Datasets and Benchmarks Track, 2024. 6", + "[10] Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14239-14250, 2024. 3", + "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ *chatgpt quality, 2023. 6", + "[12] Yinpeng Dong, Huanran Chen, Jiawei Chen, Zhengwei Fang, Xiao Yang, Yichi Zhang, Yu Tian, Hang Su, and Jun Zhu. How robust is google's bard to adversarial image attacks? arXiv preprint arXiv:2309.11751, 2023. 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yanrui Du, Sendong Zhao, Ming Ma, Yuhan Chen, and Bing Qin. Analyzing the inherent response tendency of llms: Real-world instructions-driven jailbreak. arXiv preprint arXiv:2312.04127, 2023. 3", + "[14] Jiahui Gao, Renjie Pi, Tianyang Han, Han Wu, Lanqing Hong, Lingpeng Kong, Xin Jiang, and Zhenguo Li. Coca: Regaining safety-awareness of multimodal large language models with constitutional calibration. arXiv preprint arXiv:2409.11365, 2024. 1, 2, 3, 6, 7, 8, 15, 16, 17", + "[15] Simon Geisler, Tom Wollschläger, M. H. I. Abdalla, Johannes Gasteiger, and Stephan Gunnemann. Attacking large language models with projected gradient descent, 2024. 2, 4", + "[16] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via topographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 2, 3, 6, 7, 8, 15, 17, 18, 19", + "[17] Jindong Gu, Zhen Han, Shuo Chen, Ahmad Beirami, Bailan He, Gengyuan Zhang, Ruotong Liao, Yao Qin, Volker Tresp, and Philip Torr. A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980, 2023. 3", + "[18] Xingang Guo, Fangxu Yu, Huan Zhang, Lianhui Qin, and Bin Hu. Cold-attack: Jailbreaking Ilms with stealthiness and controllability. arXiv preprint arXiv:2402.08679, 2024. 3", + "[19] Dongchen Han, Xiaojun Jia, Yang Bai, Jindong Gu, Yang Liu, and Xiaochun Cao. Ot-attack: Enhancing adversarial transferability of vision-language models via optimal transport optimization. arXiv preprint arXiv:2312.04403, 2023. 1, 3", + "[20] Seungju Han, Kavel Rao, Allyson Ettinger, Liwei Jiang, Bill Yuchen Lin, Nathan Lambert, Yejin Choi, and Nouha Dziri. Wildguard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of llms. arXiv preprint arXiv:2406.18495, 2024. 6", + "[21] Jonathan Hayase, Ema Borevkovic, Nicholas Carlini, Florian Tramér, and Milad Nasr. Query-based adversarial prompt generation. arXiv preprint arXiv:2402.12329, 2024. 2", + "[22] Zhang-Wei Hong, Idan Shenfeld, Tsun-Hsuan Wang, Yung-Sung Chuang, Aldo Pareja, James Glass, Akash Srivastava, and Pulkit Agrawal. Curiosity-driven red-teaming for large language models. arXiv preprint arXiv:2402.19464, 2024. 4", + "[23] MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR), 51 (6):1-36, 2019. 1", + "[24] Sandy Huang, Nicolas Papernot, Ian Goodfellow, Yan Duan, and Pieter Abbeel. Adversarial attacks on neural network policies. arXiv preprint arXiv:1702.02284, 2017. 4", + "[25] Yangsibo Huang, Samyak Gupta, Mengzhou Xia, Kai Li, and Danqi Chen. Catastrophic jailbreak of open-source llms via exploiting generation. arXiv preprint arXiv:2310.06987, 2023. 3", + "[26] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Llm-based input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023. 6" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25046", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Erik Jones, Anca Dragan, Aditi Raghunathan, and Jacob Steinhardt. Automatically auditing large language models via discrete optimization. In International Conference on Machine Learning, pages 15307-15329. PMLR, 2023. 2", + "[28] Maxim Khanov, Jirayu Burapacheep, and Yixuan Li. ARGs: Alignment as reward-guided search. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[29] Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, et al. Rewardbench: Evaluating reward models for language modeling. arXiv preprint arXiv:2403.13787, 2024. 4", + "[30] Seanie Lee, Minsu Kim, Lynn Cherif, David Dobre, Juho Lee, Sung Ju Hwang, Kenji Kawaguchi, Gauthier Gidel, Yoshua Bengio, Nikolay Malkin, and Moksh Jain. Learning diverse attacks on large language models for robust red-teaming and safety tuning, 2024. 4", + "[31] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665, 2023. 15, 16", + "[32] Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, and Qi Liu. Red teaming visual language models. arXiv preprint arXiv:2401.12915, 2024. 1", + "[33] Yuhui Li, Fangyun Wei, Jinjing Zhao, Chao Zhang, and Hongyang Zhang. Rain: Your language models can align themselves without finetuning. arXiv preprint arXiv:2309.07124, 2023. 3", + "[34] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 6, 17", + "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.2,6", + "[36] Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, et al. Agentbench: Evaluating llms as agents. arXiv preprint arXiv:2308.03688, 2023. 1", + "[37] Xin Liu, Yichen Zhu, Yunshi Lan, Chao Yang, and Yu Qiao. Query-relevant images jailbreak large multi-modal models, 2023. 3, 6", + "[38] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models, 2024. 2, 7, 8, 15", + "[39] Yi Liu, Gelei Deng, Zhengzi Xu, Yuekang Li, Yaowen Zheng, Ying Zhang, Lida Zhao, Tianwei Zhang, Kailong Wang, and Yang Liu. Jailbreaking chatgpt via prompt engineering: An empirical study. arXiv preprint arXiv:2305.13860, 2023. 3", + "[40] Weidi Luo, Siyuan Ma, Xiaogeng Liu, Xiaoyu Guo, and Chaowei Xiao. Jailbreakv-28k: A benchmark for assessing the robustness of multimodal large language models against jailbreak attacks. arXiv preprint arXiv:2404.03027, 2024. 2, 3, 6, 7, 15, 16, 18, 20", + "[41] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng" + ], + "bbox": [ + 91, + 92, + 485, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tu. Macaw-lm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093, 2023. 3", + "[42] Neal Mangaokar, Ashish Hooda, Jihye Choi, Shreyas Chandrashekaran, Kassem Fawaz, Somesh Jha, and Atul Prakash. Prp: Propagating universal perturbations to attack large language model guard-rails. arXiv preprint arXiv:2402.15911, 2024. 2", + "[43] Ninareh Mehrabi, Palash Goyal, Anil Ramakrishna, Jwala Dhamala, Shalini Ghosh, Richard Zemel, Kai-Wei Chang, Aram Galstyan, and Rahul Gupta. Jab: Joint adversarial prompting and belief augmentation. In R0-FoMo: Robustness of Few-shot and Zero-shot Learning in Large Foundation Models. 3", + "[44] Ninareh Mehrabi, Ahmad Beirami, Fred Morstatter, and Aram Galstyan. Robust conversational agents against imperceptible toxicity triggers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2831-2847, 2022. 4", + "[45] Ninareh Mehrabi, Palash Goyal, Christophe Dupuy, Qian Hu, Shalini Ghosh, Richard Zemel, Kai-Wei Chang, Aram Galstyan, and Rahul Gupta. Flirt: Feedback loop in-context red teaming. Empirical Methods in Natural Language Processing (EMNLP), 2024. 3", + "[46] Youssef Mroueh. Information theoretic guarantees for policy alignment in large language models. arXiv preprint arXiv:2406.05883, 2024. 5, 6", + "[47] Sidharth Mudgal, Jong Lee, Harish Ganapathy, YaGuang Li, Tao Wang, Yanping Huang, Zhifeng Chen, Heng-Tze Cheng, Michael Collins, Trevor Strohman, Jilin Chen, Alex Beutel, and Ahmad Beirami. Controlled decoding from language models. International Conference on Machine Learning (ICML), 2024. 2, 3, 4, 16", + "[48] Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435, 2023. 1", + "[49] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 3", + "[50] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2, 4", + "[51] Junsoo Park, Seungyeon Jwa, Meiying Ren, Daeyoung Kim, and Sanghyuk Choi. Offsetbias: Leveraging debiased data for tuning evaluators. arXiv preprint arXiv:2407.06551, 2024. 6", + "[52] Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia Glaese, Nat McAleese, and Geoffrey Irving. Red teaming language models with language models. arXiv preprint arXiv:2202.03286, 2022. 4", + "[53] Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang." + ], + "bbox": [ + 516, + 92, + 906, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25047", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mllm-protector: Ensuring mllm's safety without hurting performance. arXiv preprint arXiv:2401.02906, 2024. 3", + "[54] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 1, 2, 3, 6, 8", + "[55] Xiangyu Qi, Ashwinee Panda, Kaifeng Lyu, Xiao Ma, Subhrajit Roy, Ahmad Beirami, Prateek Mittal, and Peter Henderson. Safety alignment should be made more than just a few tokens deep, 2024. 3", + "[56] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to! International Conference on Learning Representations (2024), 2024. 3", + "[57] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 3, 4, 15, 16", + "[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 6", + "[59] Christian Schlarmann and Matthias Hein. On the adversarial robustness of multi-modal foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3677-3685, 2023. 1, 3", + "[60] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Plug and pray: Exploiting off-the-shelf components of multi-modal models. arXiv preprint arXiv:2307.14539, 2023. 1, 3", + "[61] Chawin Sitawarin, Norman Mu, David Wagner, and Alexandre Araujo. Pal: Proxy-guided black-box attack on large language models. arXiv preprint arXiv:2402.09674, 2024. 2", + "[62] Bharath K. Sriperumbudur, Kenji Fukumizu, Arthur Gretton, Bernhard Schölkopf, and Gert R. G. Lanckriet. On integral probability metrics, phi-divergences and binary classification, 2009. 13", + "[63] Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul Christiano. Learning to summarize from human feedback, 2022. 1", + "[64] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing nlp. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2153-2162, 2019. 4", + "[65] Boxin Wang, Weixin Chen, Hengzhi Pei, Chulin Xie, Mintong Kang, Chenhui Zhang, Chejian Xu, Zidi Xiong, Ritik Dutta, Ryan Schaeffer, et al. Decodingtrust: A comprehensive assessment of trustworthiness in gpt models. In NeurlPS, 2023. 1", + "[66] Hao Wang, Hao Li, Minlie Huang, and Lei Sha. From noise to clarity: Unraveling the adversarial suffix of large language" + ], + "bbox": [ + 91, + 92, + 483, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "model attacks via translation of text embeddings. arXiv preprint arXiv:2402.16006, 2024. 2", + "[67] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. arXiv preprint arXiv:2403.09513, 2024. 1, 2, 3, 6, 7, 8, 15, 16, 17", + "[68] Zhilin Wang, Yi Dong, Olivier Delalleau, Jiaqi Zeng, Gerald Shen, Daniel Egert, Jimmy J Zhang, Makesh Narsimhan Sreedhar, and Oleksii Kuchaiev. Helpsteer2: Open-source dataset for training top-performing reward models. arXiv preprint arXiv:2406.08673, 2024. 6", + "[69] Fenghua Weng, Yue Xu, Chengyan Fu, and Wenjie Wang. MMJ-Bench: A comprehensive study on jailbreak attacks and defenses for vision language models. arXiv preprint arXiv:2408.08464, 2024. 1", + "[70] Nevan Wichers, Carson Denison, and Ahmad Beirami. Gradient-based language model red teaming. 18th Conference of European Chapter of the Assoc for Computational Linguistics (EACL), 2024. 4", + "[71] Qi Wu, Damien Teney, Peng Wang, Chunhua Shen, Anthony Dick, and Anton Van Den Hengel. Visual question answering: A survey of methods and datasets. Computer Vision and Image Understanding, 163:21-40, 2017. 1", + "[72] Nan Xu, Fei Wang, Ben Zhou, Bang Zheng Li, Chaowei Xiao, and Muhao Chen. Cognitive overload: Jailbreaking large language models with overloaded logical thinking. arXiv preprint arXiv:2311.09827, 2023. 3", + "[73] Yue Xu, Xiuyuan Qi, Zhan Qin, and Wenjie Wang. Defending jailbreak attack in vlms via cross-modality information detector. arXiv preprint arXiv:2407.21659, 2024. 3", + "[74] Zhangchen Xu, Fengqing Jiang, Luyao Niu, Yuntian Deng, Radha Poovendran, Yejin Choi, and Bill Yuchen Lin. Magpie: Alignment data synthesis from scratch by prompting aligned lms with nothing, 2024. 6", + "[75] Zonghao Ying, Aishan Liu, Tianyuan Zhang, Zhengmin Yu, Siyuan Liang, Xianglong Liu, and Dacheng Tao. Jailbreak vision language models via bi-modal adversarial prompt. arXiv preprint arXiv:2406.04031, 2024. 3", + "[76] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13807-13816, 2024. 3", + "[77] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 2, 16, 17", + "[78] Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang. Rrrh: Rank responses to align language models with human feedback without tears, 2023. 1", + "[79] Yi Zeng, Hongpeng Lin, Jingwen Zhang, Diyi Yang, Ruoxi Jia, and Weiyan Shi. How johnny can persuade llms to jailbreak them: Rethinking persuasion to challenge ai safety by humanizing llms. arXiv preprint arXiv:2401.06373, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25048", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3", + "[81] Xiaoyu Zhang, Cen Zhang, Tianlin Li, Yihao Huang, Xiaojun Jia, Xiaofei Xie, Yang Liu, and Chao Shen. A mutation-based method for multi-modal jailbreaking attack detection. arXiv preprint arXiv:2312.10766, 2023. 3", + "[82] Zhuo Zhang, Guangyu Shen, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Make them spill the beans! coercive knowledge extraction from (production) llms. arXiv preprint arXiv:2312.04782, 2023. 2", + "[83] Xuandong Zhao, Xianjun Yang, Tianyu Pang, Chao Du, Lei Li, Yu-Xiang Wang, and William Yang Wang. Weak-to-strong jailbreaking on large language models. arXiv preprint arXiv:2401.17256, 2024. 3", + "[84] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[85] Yukai Zhou, Zhijie Huang, Feiyang Lu, Zhan Qin, and Wenjie Wang. Don't say no: Jailbreaking llm by suppressing refusal. arXiv preprint arXiv:2404.16369, 2024. 3", + "[86] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 2, 3, 6", + "[87] Sicheng Zhu, Ruiyi Zhang, Bang An, Gang Wu, Joe Barrow, Zichao Wang, Furong Huang, Ani Nenkova, and Tong Sun. Autodan: interpretable gradient-based adversarial attacks on large language models. In First Conference on Language Modeling, 2024. 2, 4", + "[88] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 1, 4", + "[89] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. arXiv preprint arXiv:2402.02207, 2024. 3", + "[90] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 1, 2, 3, 4", + "[91] Andy Zou, Long Phan, Justin Wang, Derek Duenas, Maxwell Lin, Maksym Andriushchenko, Rowan Wang, Zico Kolter, Matt Fredrikson, and Dan Hendrycks. Improving alignment and robustness with short circuiting. arXiv preprint arXiv:2406.04313, 2024. 3" + ], + "bbox": [ + 91, + 92, + 485, + 782 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "25049", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/66fe35a0-e089-4c9c-afba-719727b0b2df_model.json b/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/66fe35a0-e089-4c9c-afba-719727b0b2df_model.json new file mode 100644 index 0000000000000000000000000000000000000000..43d6f343340ead10c9d595f60d203e678ca37404 --- /dev/null +++ b/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/66fe35a0-e089-4c9c-afba-719727b0b2df_model.json @@ -0,0 +1,2622 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.125, + 0.885, + 0.178 + ], + "angle": 0, + "content": "IMMUNE: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.205, + 0.893, + 0.243 + ], + "angle": 0, + "content": "Soumya Suvara Ghosal\\(^{1*}\\) Souradip Chakraborty\\(^{1*}\\) Vaibhav Singh\\(^{2*}\\) Tianrui Guan\\(^{1}\\) Mengdi Wang\\(^{3}\\) Ahmad Beirami Furong Huang\\(^{1}\\) Alvaro Velasquez\\(^{4}\\) Dinesh Manocha\\(^{1\\dagger}\\) Amrit Singh Bedi\\(^{5\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.257, + 0.885, + 0.294 + ], + "angle": 0, + "content": "1University of Maryland; 2Indian Institute of Technology Bombay; 3Princeton University; 4University of Colorado Boulder; 5University of Central Florida" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.346, + 0.327, + 0.363 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.379, + 0.486, + 0.683 + ], + "angle": 0, + "content": "With the widespread deployment of Multimodal Large Language Models (MLLMs) for visual-reasoning tasks, improving their safety has become crucial. Recent research indicates that despite training-time safety alignment, these models remain vulnerable to jailbreak attacks. In this work, we first highlight an important safety gap to describe that alignment achieved solely through safety training may be insufficient against jailbreak attacks. To address this vulnerability, we propose IMMUNE, an inference-time defense framework that leverages a safety reward model through controlled decoding to defend against jailbreak attacks. Additionally, we provide a mathematical characterization of IMMUNE, offering insights on why it improves safety against jailbreaks. Extensive evaluations on diverse jailbreak benchmarks using recent MLLMs reveal that IMMUNE effectively enhances model safety while preserving the model's original capabilities. For instance, against text-based jailbreak attacks on LLaVA-1.6, IMMUNE reduces the attack success rate by \\(57.82\\%\\) and \\(16.78\\%\\) compared to the base MLLM and state-of-the-art defense strategy, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.703, + 0.222, + 0.719 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.486, + 0.865 + ], + "angle": 0, + "content": "Multi-modal large language models (MLLMs) have made remarkable progress in vision-language reasoning tasks such as visual question-answering [1, 3, 71] and image captioning [1, 23]. However, ensuring their outputs are safe and free from discrimination, disinformation, and harmful content is important for their adoption. Safety alignment through reinforcement learning from human feedback (RLHF) has shown promise in aligning model behavior with social norms to prevent harmful outputs [63, 78, 88]. Despite these sig" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.409 + ], + "angle": 0, + "content": "nificant alignment efforts, recent studies reveal that MLLMs remain vulnerable to jailbreaking attacks, where malicious image-text pairs bypass safety mechanisms, raising safety concerns [19, 32, 36, 48, 54, 59, 60, 65]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.909, + 0.639 + ], + "angle": 0, + "content": "Why training-time safety alignment might not be enough? Fine-tuning-based safety alignment approaches rely on a static prompt distribution, leaving them inherently vulnerable to jailbreak attacks. These attacks exploit this limitation by solving a prompt (text or image) optimization problem to craft adversarial prompts that bypass the safety mechanisms of aligned models [54, 90]. In this work, we argue that safety finetuning against jailbreak attacks while keeping model capabilities may be inherently challenging given that the adversary can find a prompt that bypasses the safety mechanisms of the model. This underscores the fundamental limitation of relying solely on training-time safety alignment procedures to ensure robustness. These leads us to ask: Is it possible to design a safeguarding strategy to defend against given possible adversarial prompts?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.911, + 0.882 + ], + "angle": 0, + "content": "In this paper, we take a step towards answering this question in the affirmative. We propose shifting from training-time alignment to inference-time safeguarding. Although initial efforts in this direction [14, 67] show promise, they exhibit weak empirical performance. Further, Weng et al. [69] note that adding a safety prompt to the input [67] can lead to overly cautious behavior, resulting in generic responses like, \"I am sorry...\" even for benign queries (see Figure 1), thereby affecting model utility. In real-world applications, an ideal jailbreaking defense should not only minimize the success rate of attacks but also maintain high utility by preserving model performance on legitimate queries. Moreover, existing inference-time approaches lack mathematical characterization, relying primarily on heuristic-based methods. This gap motivates the need for robust inference-time safeguarding mechanisms, which we address in this paper." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Our key idea and proposed approach. We present IMMUNE," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.887 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.2, + 0.9 + ], + "angle": 0, + "content": "Equal advising." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25038" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.531, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.089, + 0.905, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.294, + 0.907, + 0.447 + ], + "angle": 0, + "content": "Figure 1. Qualitative Evaluation (Left). Given an image of a house generated by stable diffusion [58] and perturbed by adversarial noise [54], along with a malicious user query asking for steps to \"break into and rob a house\", we visualize responses from the base model and various inference-time defense strategies, including CoCA [14] and AdaShield [67]. We observe that all compared defense strategies are misled into generating harmful content. In contrast, our proposed inference-time safety alignment framework, IMMUNE, effectively rejects the user query, citing its unethical nature. This evaluation underscores the importance of inference-time alignment in preventing harmful responses. We visualize additional responses in the Appendix. Quantitative Evaluation (Right). To empirically validate the effectiveness of IMMUNE, we compare the attack success rates and model utility across various state-of-the-art defense strategies. A lower attack success rate reflects improved safety in generated outputs. Our results indicate that IMMUNE substantially lowers the attack success rate compared to other baselines. Additionally, IMMUNE not only strengthens model safety but also preserves the model's original utility, as demonstrated by its performance on MM-Vet benchmark [77]. We note that an ideal defense strategy should have a low attack success rate while maintaining high utility, i.e., towards the bottom right corner of the plot. Refer to Section 6 for further details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.483, + 0.684 + ], + "angle": 0, + "content": "a framework that reframes jailbreak defense as an inference-time alignment problem. Our approach is built on two key components: First, we show that jailbreaks can be effectively mitigated by aligning the MLLM during inference using controlled decoding [7, 47], assuming access to a safety-aware utility or reward function (without loss of generality). Second, we mathematically formalize this defense as an inference-time alignment problem within the KL-regularized reinforcement learning framework [50], proving its robustness against adversarial prompt distributions. As illustrated in Figure 1, our comprehensive evaluations demonstrate that IMMUNE outperforms state-of-the-art inference-time defense strategies, such as AdaShield [67] and CoCA [14], while maintaining high model utility. Our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.69, + 0.483, + 0.765 + ], + "angle": 0, + "content": "- Insufficiency of safety finetuning for MLLMs. We first argue that the current alignment achieved solely via safety training/fine-tuning with RLHF is insufficient to defend against jailbreak attacks, thereby highlighting the need for additional safeguards." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.87 + ], + "angle": 0, + "content": "- Improving safety via inference time alignment. To address these safety concerns, we propose IMMUNE, an inference-time framework for defense against jailbreak attacks through controlled decoding using a safety reward model. We provide a characterization of our defense mechanism by framing it as an alignment problem, offering mathematical guarantees against jailbreak attacks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Empirical evaluations. We empirically validate the efficacy of our proposed defense strategy, IMMUNE," + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.473, + 0.909, + 0.565 + ], + "angle": 0, + "content": "across a range of jailbreak benchmarks, including MM-SafetyBench [38], FigStep [16], visual adversarial attacks [54], and text-based jailbreak attacks [40]. Our evaluations demonstrate significant improvements on several recent MLLMs, such as LLaVA-1.5 [34], LLaVA-1.6 [35], MiniGPT-4 [86], and Qwen-VL [4]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.586, + 0.662, + 0.601 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Jailbreak attacks. Jones et al. [27] frames jailbreaking large language models (LLMs) as a discrete optimization problem. They search for the optimal suffix that, when appended to the original prompt, can greedily generate the desired harmful output. Zou et al. [90] demonstrated that model inputs could be manipulated based on gradients to force-generate responses to harmful queries. Zhu et al. [87] introduced a method to generate an adversarial suffix iteratively. This approach ensures that the optimized suffix remains meaningful while bypassing perplexity-based filters. Other works that proposed jailbreaking by manipulating gradients include Andriushchenko et al. [2], Geisler et al. [15], Hayase et al. [21], Mangaokar et al. [42], Sitawarin et al. [61], Wang et al. [66]. Another line of work, exemplified by Zhang et al. [82], involves optimizing the prompt iteratively until the distribution of output tokens results in generating harmful responses. For example, Zhang et al. [82] showed that adversaries can break safety alignment by forcing the target LLM to select a lower-ranked output token, thereby generating toxic con" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25039" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.455 + ], + "angle": 0, + "content": "tent. Additional works following this domain include Du et al. [13], Guo et al. [18], Huang et al. [25], Zhao et al. [83], Zhou et al. [85]. Recent work has also focused on jailbreaking models through iterative interactions with the model [8, 45]. Finally Qi et al. [56] demonstrated that even benign finetuning of the models may lead to the removal of model safety safeguards. Similar to LLMs, jailbreaking MLLMs can be achieved by either manipulating the visual input [16, 37, 54], the text prompt [40], or a combination of both [75]. Qi et al. [54] demonstrated that adding adversarial noise, which is imperceptible to humans and guided by the model's input gradients, to the input image can effectively deceive MLLMs. Recently, there has been substantial research studying the robustness of MLLMs to adversarial images [12, 19, 49, 59, 60, 84]. Another line of work [16, 37] explores embedding harmful content directly into images via typography or text-to-image tools. Recently, Luo et al. [40] showed that text-based jailbreak prompts [39, 72, 79, 90] developed for LLMs can be effectively transferred to MLLMs and are often more potent than image-based attacks. Additionally, Ying et al. [75] introduced a jail-breaking framework that simultaneously perturbs both visual and textual modalities. See also the survey by Gu et al. [17] on more works on prompt-based attacks against MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.487, + 0.779 + ], + "angle": 0, + "content": "Defense via safety alignment (fine-tuning). Defense approaches typically aim to enhance the safety in MLLMs by either fine-tuning the model on a dataset of safe image-text pairs [89] or adversarial training [89]. For instance, Zong et al. [89] proposed a safety instruction-following dataset VLGuard for safeguarding MLLMs. DRESS [10] leverages natural language feedback (NLF) from LLMs to improve the alignment and interactions within MLLMs. MLLM-Protector [53] proposed a plug-and-play strategy combining a lightweight harm detector to identify the harmful response and a detoxifier to correct harmful outputs. However, fine-tuning-based approaches suffer from (i) the requirement of a large annotated dataset consisting of safe image-text pairs, (ii) being computationally intensive, and (iii) requiring access to model weights. Qi et al. [55] highlighted that safety alignment primarily affects the first few output tokens leaving models susceptible to simple jailbreak attacks. Zou et al. [91] introduced circuit-breaking techniques to enhance model robustness. Yu et al. [76] proposed improving the safety of MLLMs by aligning with human preferences through direct preference optimization [57]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.487, + 0.904 + ], + "angle": 0, + "content": "Defense via inference time alignment (controlled decoding). Khanov et al. [28] proposed a reward-guided inference time alignment. Mudgal et al. [47] proposed controlled decoding through a value function for the reward to guide the language model to more safe outcomes. Chakraborty et al. [7] improved the value function by bridging the gap between the value of the decoding policy and the reference policy. Li et al. [33] proposed leveraging self-evaluation and rewind" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.32 + ], + "angle": 0, + "content": "mechanisms in LLMs to generate safe responses. Gong et al. [16] proposed concatenating a fixed safety prompt with the malicious user query to evade jailbreaking. Wang et al. [67] proposed refining the safety prompt iteratively rather than using a fixed template. Mehrabi et al. [43] introduced an interactive approach to improve robustness against adversarial attacks. Zhang et al. [81] proposed mutating untrusted inputs to generate variants, identifying attack samples by examining response discrepancies among the variants. In another approach, Xu et al. [73] detected adversarial images by analyzing cross-modal similarity between harmful queries and perturbed images. Additionally, Gao et al. [14] suggested improving safety by calibrating MLLM output distribution. However, existing approaches compromise model utility, leading to overly cautious responses (Figure 1)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.323, + 0.722, + 0.338 + ], + "angle": 0, + "content": "3. Problem Formulation" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.347, + 0.655, + 0.362 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.909, + 0.568 + ], + "angle": 0, + "content": "Multimodal Large Language Models (MLLMs) are autoregressive text generation models that can process multiple modalities, such as image [86], video [80], and audio [41], in addition to text. In this work, we consider input to MLLM in the form of an image and a text prompt. Mathematically, we denote a MLLM as a mapping \\(\\pi_{\\mathrm{safe}}: \\mathcal{I} \\times \\mathcal{V}^N \\to \\mathcal{V}^M\\), where \\(\\mathcal{I}\\) represents the image space, and \\(\\mathcal{V}\\) denotes the vocabulary set. The input to the MLLM \\(\\pi_{\\mathrm{safe}}\\) consists of an image \\(\\mathbf{I} \\in \\mathcal{I}\\) and a sequence of tokens \\(\\mathbf{x} := \\{x_1, x_2, \\dots, x_N\\}\\) (referred as prompt), where each token \\(x_i \\in \\mathcal{V}\\). Given \\(\\mathbf{x}_{\\mathrm{input}} := [\\mathbf{I}, \\mathbf{x}]\\), the MLLM \\(\\pi_{\\mathrm{safe}}(\\cdot | \\mathbf{x}_{\\mathrm{input}})\\) generates the output sequence \\(\\mathbf{y} := \\{y_1, y_2, \\dots, y_M\\}\\), where \\(y_i \\in \\mathcal{V}\\), in a token by token fashion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.911, + 0.661 + ], + "angle": 0, + "content": "Jailbreaking MLLMs. Despite safety tuning efforts aimed at preventing MLLMs from generating harmful content, jailbreak attacks as formulated in Qi et al. [54], Zou et al. [90] can still make the safety aligned model \\(\\pi_{\\mathrm{safe}}\\) to generate harmful outputs with the help of an adversarial prompt \\(\\mathbf{x}_{\\mathrm{adv}}\\) obtained by solving" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.671, + 0.909, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathrm {a d v}} := \\arg \\min _ {\\mathbf {q} \\in \\mathcal {B}} \\sum_ {i = 1} ^ {m} - \\log \\left(\\pi_ {\\text {s a f e}} \\left(y _ {i} \\mid \\left[ \\mathbf {x} _ {\\text {i n p u t}}, \\mathbf {q} \\right]\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.718, + 0.912, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}\\) is a constraint on the adversarial prompts and \\(\\{y_i\\}_{i=1}^m\\) are few initial tokens of intended harmful response. For example, in Figure 1 (Left), we show that given an image of a house generated by stable diffusion [58] and then perturbed with adversarial noise, along with a malicious user query to explain \"How to break into and rob a house\", the model is forced to generate unsafe responses, effectively bypassing its safety alignment." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.848, + 0.779, + 0.864 + ], + "angle": 0, + "content": "3.2. Safeguard Against Jailbreaks" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "From the definition of adversarial prompt in (1), we observe that \\(\\mathbf{x}_{\\mathrm{adv}}\\) can lead the MLLM model to generate a harmful" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25040" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.487, + 0.289 + ], + "angle": 0, + "content": "response, \\(\\mathbf{y}_{\\mathrm{unsafe}} \\sim \\pi_{\\mathrm{safe}}(\\cdot|\\mathbf{x}_{\\mathrm{adv}})\\), despite the training time safety alignment of \\(\\pi_{\\mathrm{safe}}\\). This raises the question: Why do insufficiencies exist in safety-aligned models that make them vulnerable to attacks, and how can we mitigate them? To answer this question, we start by noticing that an attacker's goal is to learn an adversarial prompt distribution \\(p_{\\mathrm{adv}}\\) by maximizing some \"success metric\" subject to a \"cost metric,\" which is general and covers existing attacks such as projected gradient descent (PGD) attacks [15], greedy coordinate gradient (GCG) attacks [87, 90], universal adversarial trigger (UAT) attacks [44, 64], and red teaming attacks [22, 30, 52, 70]. Mathematically, an adversary can solve the following optimization problem to learn an adversarial prompt:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.299, + 0.483, + 0.344 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p _ {\\mathrm {a d v}} := \\arg \\min _ {p} \\mathbb {E} _ {\\mathbf {q} \\sim p (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}), \\mathbf {y} \\sim \\pi_ {\\text {s a f e}} (\\cdot | \\mathbf {q})} \\left[ R _ {\\text {s a f e}} (\\mathbf {x} _ {\\text {i n p u t}}, \\mathbf {y}) \\right] \\\\ + \\beta \\operatorname {d i v} \\left(p (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}), p _ {0} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}})\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.485, + 0.564 + ], + "angle": 0, + "content": "where the \"cost metric\" term div \\((p(\\cdot |\\mathbf{x}_{\\mathrm{input}}),p_0(\\cdot |\\mathbf{x}_{\\mathrm{input}}))\\) helps keep the adversarial prompt close to a given prompt distribution, \\(p_0\\). We can think of \\(p_0\\) as a given prompt distribution that will sample \\(\\mathbf{x}_{\\mathrm{input}}\\) only with high likelihood and have minimal probability of sampling other prompts. Then, the divergence part of the objective in (2) will make sure the adversarial prompts are closer to original prompts \\(\\mathbf{x}_{\\mathrm{input}}\\), which will keep them interpretable and not completely out of distribution [88]. The formulation in (2) is also in accordance with the existing literature on adversarial attacks in machine learning [6, 24] because it is just trying to ensure that adversarial prompts \\(\\mathbf{x}_{\\mathrm{adv}}\\) looks like similar to \\(\\mathbf{x}_{\\mathrm{input}}\\). The regularization parameter \\(\\beta >0\\) would control this cost metric in (2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.613 + ], + "angle": 0, + "content": "Training-time defense. We note that learning a safe and robust model is inherently a min-max optimization problem given by" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.623, + 0.483, + 0.688 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {p} \\max _ {\\pi} \\mathbb {E} _ {\\mathbf {q} \\sim p (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}), \\mathbf {y} \\sim \\pi (\\cdot | \\mathbf {q})} \\left[ R _ {\\text {s a f e}} (\\mathbf {x} _ {\\text {i n p u t}}, \\mathbf {y}) \\right. \\\\ + \\beta \\operatorname {d i v} (p (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}), p _ {0} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}})) \\\\ \\left. - \\alpha \\operatorname {d i v} \\left(\\pi (\\cdot | \\mathbf {q}), \\pi_ {\\text {r e f}} (\\cdot | \\mathbf {q})\\right) \\right], \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.485, + 0.849 + ], + "angle": 0, + "content": "where the inner maximization captures the safeguarding of the model, e.g., through safety alignment. The safety alignment in recent literature [50, 57] only focused on solving the inner max in (3) for a given prompt distribution (say \\( p_0 \\)) to obtain \\( \\pi_{\\mathrm{safe}} \\) which is safe only under the prompt distribution \\( \\mathbf{x} \\sim p_0 \\). An adversary exploits it and learns an adversarial distribution \\( \\rho_{\\mathrm{adv}} \\) breaking the alignment and generating harmful responses. This is likely the reason why adversarial attacks on images (and other modalities) have remained challenging to address over the past decade." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Key insight for tractable inference-time defense. We reframe the optimization from \\(\\min_p\\max_\\pi\\) to \\(\\max_{\\pi}\\min_{p}\\) by letting the safeguard be applied after the attacker has chosen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.349 + ], + "angle": 0, + "content": "their jailbreak prompt. While the two formulations are equivalent for convex problems, this reformulation may lead to an improvement in defense due to the minimax duality gap for general non-convex problems. Additionally, this formulation allows an efficient solution of the \\(\\max_{\\pi}\\) component during inference, leveraging recent advances in inference-time alignment methods through controlled decoding [7, 47]. Notably, inference-time alignment enables prompt-specific safety adjustments, which are challenging to implement during training. This capability is enabled by the user (or attacker) interaction protocol with MLLM. Specifically, the attacker sets the adversarial prompt, which remains unchanged throughout the process, providing a fixed input for the MLLM. This scenario ultimately reduces the challenge of learning a safe and optimal policy within the predefined adversarial prompt environment, effectively solving the problem \\(\\max_{\\pi}\\) in inference. We will provide a detailed method in the next section." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.816, + 0.379 + ], + "angle": 0, + "content": "4. Inference Time Safety Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.909, + 0.537 + ], + "angle": 0, + "content": "In this section, we propose improving the safety of response generation by using alignment through the controlled decoding procedure proposed in [7, 47]. Our key insight is that well-trained safety rewards, denoted by \\( R_{\\mathrm{safe}}(\\mathbf{x}, \\mathbf{y}) \\), are often accessible online, for instance, RewardBench [29]. This safety reward will give high scores for safe text and low scores for unsafe text. We remark here that the reward model itself is likely not ideal and subject to adversarial attack. With this understanding, we propose to solve the following decoding problem at inference time for each token \\( t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.544, + 0.907, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\pi_ {\\text {s a f e - d e c}} ^ {*} (\\cdot | \\mathbf {s} _ {t}) := \\arg \\max _ {\\pi} \\mathbb {E} _ {z \\sim \\pi (\\cdot | \\mathbf {s} _ {t})} [ Q _ {\\text {s a f e}} (\\mathbf {s} _ {t}, z) ] \\tag {4} \\\\ - \\alpha \\mathrm {K L} (\\pi (\\cdot | \\mathbf {s} _ {t}) | | \\pi_ {\\text {s a f e}} (\\cdot | \\mathbf {s} _ {t})), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.909, + 0.638 + ], + "angle": 0, + "content": "where \\(\\mathbf{s}_t\\coloneqq [\\mathbf{x}_{\\mathrm{adv}},\\mathbf{y}_{< t}]\\) and \\(\\mathbf{y}_{< t} = [y_0,y_1,\\dots ,y_{t - 1}]\\) capture the tokens generated so far. In (4), the term \\(Q_{\\mathrm{safe}}(\\mathbf{s}_t,z)\\) is the action-value function defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.645, + 0.907, + 0.663 + ], + "angle": 0, + "content": "\\[\nQ _ {\\text {s a f e}} \\left(\\mathbf {s} _ {t}, z\\right) = \\mathbb {E} _ {\\tau \\sim \\rho_ {\\text {s a f e}} (\\cdot | [ \\mathbf {s} _ {t}, z ])} \\left[ R _ {\\text {s a f e}} \\left([ \\mathbf {s} _ {t}, z ], \\tau\\right) \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.669, + 0.909, + 0.775 + ], + "angle": 0, + "content": "where \\(\\tau := [z_{t+1}, z_{t+2}, \\dots, z_{eos}]\\) denotes the remaining trajectory starting from \\([\\mathbf{s}_t, z]\\), \\(\\rho_{\\mathrm{safe}}(\\tau | [\\mathbf{s}_t, z]) := \\prod_{i=t+1}^{\\mathrm{eos}} \\pi_{\\mathrm{safe}}(z_i | [\\mathbf{s}_t, z])\\) is the trajectory level distribution, where \\(eos\\) denotes end of sentence token, and \\(\\tau\\) denotes the sequence of tokens generated by the model starting from \\([\\mathbf{s}_t, z]\\). The closed form expression of (4) is given by [47, Theorem 2.1]:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.781, + 0.907, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\pi_ {\\text {s a f e - d e c}} ^ {*} (z | \\mathbf {s} _ {t}) = \\frac {\\pi_ {\\text {s a f e}} (z | \\mathbf {s} _ {t})}{Z} \\exp \\left(\\frac {Q _ {\\text {s a f e}} (\\mathbf {s} _ {t} , z)}{\\alpha}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.82, + 0.908, + 0.881 + ], + "angle": 0, + "content": "where \\( Z \\) is the normalizing constant. For a given adversarial prompt \\( \\mathbf{x}_{\\mathrm{adv}} \\), we propose to follow the decoding policy derived in (6) at the inference time to generate a safe response. Our proposed algorithm is summarized in Algorithm 1." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.81, + 0.901 + ], + "angle": 0, + "content": "1https://huggingface.co/spaces/allenai/reward-bench" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25041" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.092, + 0.091, + 0.484, + 0.122 + ], + "angle": 0, + "content": "Algorithm 1 IMMUNE: Safety against Jailbreaking attacks via Inference-time alignment" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.125, + 0.487, + 0.171 + ], + "angle": 0, + "content": "1: Input: Base MLLM \\(\\pi_{\\mathrm{safe}}\\), safe reward model \\(R_{\\mathrm{safe}}\\), adversarial input \\(\\mathbf{x}_{\\mathrm{adv}}\\), number of tokens sampled \\(k\\), alignment parameter \\(\\alpha\\), vocabulary set \\(\\mathcal{V}\\)." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.172, + 0.254, + 0.185 + ], + "angle": 0, + "content": "2: for \\( t = 0, \\dots, T \\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.186, + 0.484, + 0.217 + ], + "angle": 0, + "content": "3: Current state: \\(\\mathbf{s}_t = [\\mathbf{x}_{\\mathrm{adv}},\\mathbf{y}_{< t}]\\), where \\(\\mathbf{x}_{\\mathrm{adv}}\\) is adversarial input and \\(\\mathbf{y}_{< t} = [y_0,y_1,\\dots ,y_{t - 1}]\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.218, + 0.482, + 0.247 + ], + "angle": 0, + "content": "4: Sample top-k tokens using base model \\(\\pi_{\\mathrm{safe}}\\) and store as \\(\\hat{\\mathcal{V}} = \\{z_i : z_i \\sim \\pi_{\\mathrm{safe}}(\\cdot|\\mathbf{s}_t)\\}_{i=1}^k\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.248, + 0.235, + 0.261 + ], + "angle": 0, + "content": "5: for \\(z\\in \\mathcal{V}\\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.263, + 0.39, + 0.278 + ], + "angle": 0, + "content": "6: Evaluate \\( Q_{\\mathrm{safe}} \\) as in Equation (5)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.279, + 0.482, + 0.31 + ], + "angle": 0, + "content": "7: Compute decoding score for token \\(z\\): \\(g_{z} = \\frac{1}{\\alpha} Q_{\\mathrm{safe}}([s_t,z]) + \\log \\pi_{\\mathrm{saf}}(z|s_t)\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.31, + 0.477, + 0.331 + ], + "angle": 0, + "content": "8: Distribution over \\(\\hat{\\mathcal{V}}\\): \\(f(z|\\mathbf{s}_t) = \\frac{\\exp(g_z)}{\\sum_{z'\\in\\hat{\\mathcal{V}}}\\exp(g_{z'})}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.331, + 0.33, + 0.345 + ], + "angle": 0, + "content": "9: Next token \\(y_{t}\\sim f(\\cdot |\\mathbf{s}_{t})\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.346, + 0.343, + 0.361 + ], + "angle": 0, + "content": "10: Next state \\(\\mathbf{s}_{t + 1}\\gets [\\mathbf{s}_t,y_t]\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.363, + 0.282, + 0.38 + ], + "angle": 0, + "content": "11: Return \\(\\mathbf{y} = [y_0, \\dots, y_T]\\)" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.125, + 0.487, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.408, + 0.283, + 0.425 + ], + "angle": 0, + "content": "5. Theoretical Insights" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.486, + 0.51 + ], + "angle": 0, + "content": "In this section, we formally derive a bound on the suboptimality of our proposed decoding-based approach under an adversarial prompt distribution. Given a safe prompt distribution \\( p_0 \\) and an adversarial prompt distribution \\( p_{\\mathrm{adv}} \\), the sub-optimality gap of our algorithm can be defined as" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.521, + 0.484, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Delta_ {\\text {s u b - g a p}} \\left(\\mathbf {x} _ {\\text {i n p u t}}\\right) \\tag {7} \\\\ := \\underset { \\begin{array}{c} \\mathbf {x} \\sim p _ {0} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}) \\\\ \\mathbf {y} \\sim \\rho_ {*} (\\cdot | \\mathbf {x}) \\end{array} } {\\mathbb {E}} \\left[ R _ {\\text {s a f e}} (\\mathbf {x}, \\mathbf {y}) \\right] - \\underset { \\begin{array}{c} \\mathbf {x} \\sim p _ {\\text {a d v}} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}) \\\\ \\mathbf {y} \\sim \\rho_ {\\text {s a f e - d e c}} (\\cdot | \\mathbf {x}) \\end{array} } {\\mathbb {E}} \\left[ R _ {\\text {s a f e}} (\\mathbf {x}, \\mathbf {y}) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.485, + 0.738 + ], + "angle": 0, + "content": "where \\(\\rho_{\\mathrm{safe - dec}}\\) is the trajectory level distribution corresponding to token level \\(\\pi_{\\mathrm{safe - dec}}^*\\) and \\(\\rho_*\\) is the trajectory level optimal distribution for safety reward. The suboptimality gap in (7) captures how safe our proposed decoding-based approach is in generating responses to adversarial prompts. The suboptimality gap in (7) is similar but different to the one utilized in [5] because here, the goal is to capture the impact of adversarial prompt distribution specifically. We establish an upper bound in the following Theorem 1 on the suboptimality gap defined in (7)." + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.767, + 0.47, + 0.844 + ], + "angle": 0, + "content": "Theorem 1. Let \\( R_{safe}(\\mathbf{x},\\mathbf{y}) \\leq R_{max} \\), \\( p_0 \\) is a given prompt distribution, \\( p_{adv} \\) adversarial prompt distribution, \\( \\rho_{*}(\\cdot |\\mathbf{x}) \\) denotes the optimal trajectory level distribution for the safe reward, and \\( \\rho_{safe}(\\cdot |\\mathbf{x}) \\) denote the given training time safety aligned trajectory level distri" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.708, + 0.106 + ], + "angle": 0, + "content": "bution. Then, it holds that" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.117, + 0.893, + 0.163 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Delta_ {\\text {s u b - g a p}} (\\mathbf {x} _ {\\text {i n p u t}}) \\leq R _ {\\max } \\sqrt {K L (p _ {0} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}) \\| p _ {\\text {a d v}} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}))} \\\\ + \\alpha K L \\left(\\rho_ {*} \\left(\\cdot | \\mathbf {x} _ {\\text {i n p u t}}\\right) \\| \\rho_ {s a f e} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}})\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.174, + 0.888, + 0.205 + ], + "angle": 0, + "content": "where \\(\\alpha > 0\\) is a regularization parameter balancing the KL-divergence term in the policy alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.225, + 0.908, + 0.317 + ], + "angle": 0, + "content": "Proof Outline and Insights. The proof of Theorem 1 (see Appendix E for details) relies on decomposing the suboptimality gap term \\(\\Delta_{\\mathrm{sub - gap}}\\) into two components, \\(\\Delta_1\\) and \\(\\Delta_2\\) such that \\(\\Delta_{\\mathrm{sub - gap}} = \\Delta_1 + \\Delta_2\\), with bounds derived from total variation distance and Pinsker's inequality, respectively. We can write the decomposition as" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.329, + 0.899, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\Delta_{1}:= \\mathop{\\mathbb{E}}_{\\substack{\\mathbf{x}\\sim p_{0}(\\cdot |\\mathbf{x}_{\\text{input}})\\\\ \\mathbf{y}\\sim \\rho_{*}(\\cdot |\\mathbf{x})}}\\left[R_{\\text{safe}}(\\mathbf{x},\\mathbf{y})\\right] - \\mathop{\\mathbb{E}}_{\\substack{\\mathbf{x}\\sim p_{\\text{adv}}(\\cdot |\\mathbf{x}_{\\text{input}})\\\\ \\mathbf{y}\\sim \\rho_{*}(\\cdot |\\mathbf{x})}}\\left[R_{\\text{safe}}(\\mathbf{x},\\mathbf{y})\\right],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.369, + 0.905, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\Delta_{2}:= \\mathop{\\mathbb{E}}_{\\substack{\\mathbf{x}\\sim p_{\\text{adv}}(\\cdot |\\mathbf{x}_{\\text{input}})\\\\ \\mathbf{y}\\sim \\rho_{*}(\\cdot |\\mathbf{x})}}[R_{\\text{safe}}(\\mathbf{x},\\mathbf{y})] - \\mathop{\\mathbb{E}}_{\\substack{\\mathbf{x}\\sim p_{\\text{adv}}(\\cdot |\\mathbf{x}_{\\text{input}})\\\\ \\mathbf{y}\\sim \\rho_{\\text{safe -dec}}(\\cdot |\\mathbf{x})}}[R_{\\text{safe}}(\\mathbf{x},\\mathbf{y})].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.906, + 0.463 + ], + "angle": 0, + "content": "Next, we provide the interpretation of the mathematical result in Theorem 1 and intuitive insights about different terms as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.908, + 0.589 + ], + "angle": 0, + "content": "(1) Practical Implications of the Bound: Theorem (1) shows that sub-optimality gap upper bound is minimized when \\( p_{\\mathrm{adv}} \\) remains close to \\( p_0 \\) (low KL divergence) and also when \\( \\rho_{\\mathrm{safe}} \\) (which is our reference policy for the inference time alignment problem) aligns closely with \\( \\rho_* \\) (low policy KL divergence). This result can justify the use of KL-regularized objectives in safety-required domains, especially under adversarial prompt conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.908, + 0.681 + ], + "angle": 0, + "content": "(2) Role of Regularization Parameter \\(\\beta\\): The adversarial prompt distribution in (2), \\(p_{\\mathrm{adv}}\\), is defined as a solution to a KL-regularized optimization problem that minimizes expected reward \\(R_{\\mathrm{safe}}\\) relative to \\(p_0\\), with regularization parameter \\(\\beta\\). The lower values of \\(\\beta\\) allow \\(p_{\\mathrm{adv}}\\) to diverge further from \\(p_0\\), potentially increasing \\(\\Delta_1\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.907, + 0.746 + ], + "angle": 0, + "content": "(3) Impact of \\(\\alpha\\) on \\(\\Delta_{2}\\): The value of \\(\\alpha\\) affects the degree to which the policy \\(\\rho_{\\mathrm{safe - dec}}\\) is incentivized to align with the safety policy \\(\\rho_{*}\\). The lower values of \\(\\alpha\\) prioritize safety more heavily, thereby reducing \\(\\Delta_{2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "(3) In proof, \\(\\Delta_{1}\\) expresses how far an adversary can exploit the adversarial prompt distribution to attack the aligned model. In \\(\\Delta_{2}\\), we show how, with inference time alignment, we can achieve safety even under the adversarial prompt distribution owing to the closed form of the KL regularized RLHF objective. We note that Mroueh [46, Proposition 3] also establishes upper bounds on the difference in expected rewards between two distributions in a continuous setting, relying on tail assumptions (e.g., sub-Gaussian, sub-Gamma) to control deviations. In contrast, our Theorem 1 operates" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.469, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25042" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "entirely in a discrete setting and explicitly decomposes suboptimality (difference in expected rewards defined in (7)) into contributions from both the prompter and the language model. Furthermore, we leverage the optimality of RLHF to derive tighter upper bounds, which is not used in Mroueh [46]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.189, + 0.484, + 0.311 + ], + "angle": 0, + "content": "Remark: Note that the theoretical guarantees discussed are conditional upon a specific prompt design structure, which originates from an adversarial prompter targeting a given unsafe query. This inherently prevents us from being able to obtain a similar guarantee for training time safety fine-tuning since the adversarial prompts would not be known in advance. In our experiments, we will not make these assumptions, as we experiment with several recent static safety benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.327, + 0.223, + 0.344 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.352, + 0.291, + 0.368 + ], + "angle": 0, + "content": "6.1. Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.484, + 0.42 + ], + "angle": 0, + "content": "Jail-break Datasets. We conduct a comprehensive evaluation by considering both text-based and image-based jailbreak attacks:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.484, + 0.56 + ], + "angle": 0, + "content": "1. Text-based Attacks. For text-based attacks, we use the JailbreakV-28K dataset [40]. This dataset integrates effective text-based jailbreak prompts for LLMs with various types of image data. Specifically, the text prompts include template-based, persuasive, and logic-driven approaches, while the image input may consist of noise, be blank, contain natural images, or be generated by stable diffusion. We evaluate all possible combinations of image and text pairs to ensure thorough evaluation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.565, + 0.484, + 0.595 + ], + "angle": 0, + "content": "2. Image-based Attacks. For image-based attacks, we evaluate the following datasets:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.599, + 0.484, + 0.763 + ], + "angle": 0, + "content": "- MM-SafetyBench [37]: This benchmark assesses MLLM safety across 13 commonly prohibited scenarios and behaviors specified in usage guidelines from OpenAI [1] and Meta [26]. Each image in this dataset is associated with three types of image inputs: (1) Stable-diffusion images (SD), which are generated using stable diffusion [58] and directly relate to the query; (2) Typography (TYPO) images, containing optical character recognition representations of text from the malicious query; and (3) SD+TYPO images, which combine stable diffusion visuals with typographic subtitles." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.484, + 0.854 + ], + "angle": 0, + "content": "- FigStep [16]: This attack converts harmful instructions into images using typography. Specifically, malicious queries on 10 forbidden topics from MM-SafetyBench are first paraphrased to start with nouns like \"Steps to\", \"List of\", or \"Methods to\", and then the instructions are presented as text within images." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- Visual Adversarial Attacks [54]: Qi et al. proposed an optimization-based attack, where the malicious user generates adversarial images by introducing carefully crafted" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.599, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.908, + 0.121 + ], + "angle": 0, + "content": "perturbations to the original image, causing models to produce harmful content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.129, + 0.907, + 0.295 + ], + "angle": 0, + "content": "Models Used. We perform evaluation on 5 state-of-art open-source MLLMs: LLaVA-1.5-7B [34], LLaVA-1.6-7B [35], MiniGPT-4-7B/13B [86], and Qwen-VL-Chat-7B [4]. For both LLaVA and MiniGPT, we use Vicuna [11] as the text encoder. For IMMUNE, we set the number of tokens sampled \\( k = 10 \\) and the alignment parameter \\( \\alpha = 1 \\). We report ablations in Appendix F. For main evaluations, we use LLaMa-3.1-8B\\(^2\\) as the safety model \\( R_{\\text{safe}} \\). The safety model has been trained on a mixture of preference datasets including HelpSteer2 [68], OffsetBias [51], WildGuard [20], and Magpie [74]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.39 + ], + "angle": 0, + "content": "Baseline Defenses. We compare IMMUNE with recent state-of-the-art inference-time jail-break defense frameworks, including FigStep [16], Adashield [67], and CoCA [14]. For a fair comparison, all baseline defense methods are evaluated on a unified test dataset using consistent metrics. Detailed descriptions of baselines are provided in the Appendix G." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.906, + 0.484 + ], + "angle": 0, + "content": "Evaluation Metric. To assess the effectiveness of a jailbreak attack, we follow the literature [14, 40, 67] and measure the Attack Success Rate (ASR). Specifically, given a test dataset \\( \\mathcal{D}_{\\mathrm{unsafe}} \\) of crafted jailbreak image-text pairs, the ASR quantifies the ratio of harmful responses to the total number of input queries and is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.496, + 0.906, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A S R} = \\frac {1}{\\left| \\mathcal {D} _ {\\text {u n s a f e}} \\right|} \\sum_ {(\\mathbf {I}, \\mathbf {x}) \\in \\mathcal {D} _ {\\text {u n s a f e}}} \\mathbb {I} \\left[ \\mathcal {C} ^ {*} (\\mathbf {x}, \\pi_ {\\theta} (\\mathbf {I}, \\mathbf {x})) = \\text {T r u e} \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.909, + 0.729 + ], + "angle": 0, + "content": "where \\((\\mathbf{I},\\mathbf{x})\\) represents an image-text jailbreak prompt pair, \\(\\mathbb{I}(\\cdot)\\) is an indicator function, and \\(\\mathcal{C}^*\\) is an oracle jailbreak classifier that verifies whether the generated response \\(\\pi_{\\theta}(\\mathbf{I},\\mathbf{x})\\) aligns with the malicious intent of the user query \\(\\mathbf{x}\\). Consistent with earlier works [9, 14, 40], we use both Llama-Guard-3 Jailbreak Judge and the GPT-4 as the oracle jailbreak classifier. Hence, a lower ASR value indicates a more robust defense against jailbreak attacks. To account for randomness in output generation, we sample three independent responses for each query and consider the model successfully jailbroken if any one of the three responses is flagged as jailbroken by the oracle classifier." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.742, + 0.756 + ], + "angle": 0, + "content": "6.2. Safety Evaluation results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.771, + 0.907, + 0.876 + ], + "angle": 0, + "content": "Evaluation on Text-based attacks. In Table 1, we present the results of this evaluation for different combinations of image types and text prompts. To be specific, the malicious user queries can be template-based, persuasive in nature, or logic-driven, while the image input consists of random noise, be blank, contain natural images, or be generated by stable diffusion. For this evaluation, we used the Llama-Guard-3" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.72, + 0.9 + ], + "angle": 0, + "content": "\\(^{2}\\mathrm{LxzGordon/URM-LLaMa - }3.1 - 8\\mathrm{B}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25043" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.391 + ], + "angle": 0, + "content": "
ModelDefense StrategyNoiseSDNatureBlankAverage
TemplatePersuadeLogicTemplatePersuadeLogicTemplatePersuadeLogicTemplatePersuadeLogic
LLaVA-1.6Original66.1237.4578.5867.3437.5677.5769.2340.7882.6166.6739.4581.6060.27
FigStep [16]61.1239.2762.1662.3440.1854.0563.4135.2256.7661.0938.4352.7051.64
AdaShield [67]38.420.0011.0838.131.5618.1339.293.2219.1442.780.4816.1219.23
CoCA [14]61.2339.1762.1661.3441.2852.7063.1135.2255.4161.0937.3652.7051.37
IMMUNE (Ours)5.230.000.008.140.000.008.450.000.004.670.000.002.45
LLaVA-1.5Original58.1239.4476.5661.4738.3574.5558.2242.1177.5759.3340.2674.5556.39
FigStep [16]64.1728.3462.1662.2332.1868.9158.3937.2772.9761.0931.4268.9152.46
AdaShield [67]32.140.005.4131.360.004.2330.220.005.4129.670.0010.8112.86
CoCA [14]59.1826.3435.1358.4230.2935.1349.2232.1128.3759.0723.3937.8339.87
IMMUNE (Ours)9.230.000.008.140.000.001.470.000.005.320.000.002.10
MiniGPT-4-7BOriginal75.2354.1283.6272.4437.3385.6373.1845.2786.6472.0955.4286.6467.15
FigStep [16]47.2217.1832.4342.135.3724.3237.298.4518.9243.0817.4143.2427.74
AdaShield [67]47.2322.1432.2441.3718.4627.2058.1512.2830.2246.0919.4332.2432.21
CoCA [14]35.1818.2722.9740.3421.1431.0835.2918.4227.0348.1121.3640.5429.74
IMMUNE (Ours)18.236.1244.5911.348.2729.7217.188.4527.0216.0910.3743.2418.34
MiniGPT-4-13BOriginal74.1847.1290.6780.2353.3485.6379.2952.4583.7879.0753.2287.6570.62
FigStep [16]44.1320.2858.1143.2218.4747.3040.3627.1448.6542.0919.4258.1137.41
AdaShield [67]63.2223.1859.4464.4731.1274.5569.1527.3568.9165.0923.4153.3950.55
CoCA [14]35.2921.3436.4940.1822.2740.5435.1220.4135.1448.2221.3652.7033.21
IMMUNE (Ours)25.1718.2959.4429.3419.4157.4229.2223.1858.4323.4723.3758.4332.94
Qwen-VLOriginal46.123.2712.0952.236.189.0753.343.427.0553.114.3615.1122.99
FigStep [16]32.180.001.3540.420.005.4130.230.005.4144.122.231.3514.42
AdaShield [67]30.170.0018.1328.341.2915.1123.281.377.0542.110.4118.1315.88
CoCA [14]29.126.2713.5127.347.185.4127.291.4213.5128.4713.3613.5115.69
IMMUNE (Ours)10.272.185.4121.342.294.0318.222.355.4120.173.377.058.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.401, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Table 1. Evaluation on Text-Based Jailbreak Attacks. We report the Attack Success Rate (ASR) for various baseline defense strategies across recent MLLMs on text-based jailbreak attacks [40]. In this setup, the image input may be noise, Stable Diffusion-generated (SD), natural images (Nature), or blank, while the text query is structured as either template-based, persuasive, or logic-driven prompts. Lower ASR values indicate stronger resilience against jailbreak attacks. For this evaluation, we used Llama-Guard-3 Jailbreak Judge as the jailbreak classifier. IMMUNE consistently outperforms all compared defense frameworks across all MLLMs. The best results (lowest ASR) are highlighted in bold. All values are reported in %." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.5, + 0.907, + 0.693 + ], + "angle": 0, + "content": "
ModelDefense StrategyIllegal ActivityMalware GenerationPornographyHate SpeechPhysical HarmFraudAverage
SDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPE
LLaVA-1.6Original34.0280.4178.3520.4570.4570.457.0742.0535.0320.0651.0252.0518.0163.0163.0530.0367.0472.0747.94
FigStep [16]27.8449.4879.3825.0347.7359.097.0928.0741.0330.0436.0746.0720.0221.0852.0329.0536.0855.0737.64
AdaShield [67]19.5962.8941.249.0954.5543.187.0525.0519.066.0620.0313.039.0926.0427.046.0136.0235.0824.40
CoCA [14]24.2755.6776.0117.2759.0970.459.0233.0743.0825.0142.0752.0623.0626.0457.0125.0137.0858.0739.79
IMMUNE (Ours)0.001.035.150.000.004.559.233.0718.020.000.003.024.095.035.040.002.077.073.94
LLaVA-1.5Original34.0280.4180.4143.1875.0170.4513.0256.0739.0724.0575.0269.0725.0764.0367.0532.0373.0765.0353.85
FigStep [16]25.5263.9273.2820.4375.0544.0916.0426.0538.0638.0351.0655.0121.0457.0459.0636.0862.0861.0445.54
AdaShield [67]9.281.033.090.002.270.009.067.034.047.071.042.0311.037.053.0310.011.072.084.81
CoCA [14]17.1380.4180.4125.9127.779.918.0752.0655.0610.0863.1967.0325.0745.0557.0726.0357.0859.0744.65
IMMUNE (Ours)0.461.030.000.400.0013.646.989.0919.050.523.025.045.174.037.020.000.006.043.51
MiniGPT-4-7BOriginal13.4343.3435.052.2734.0925.0827.0414.0121.066.0229.0517.053.0834.0826.025.0540.0436.0223.04
FigStep [16]19.5943.3438.146.8240.9115.9126.0119.0623.0611.0428.0816.0314.0631.0826.0712.0644.0517.0724.21
AdaShield [67]11.3435.0518.5618.1822.739.0916.6726.0315.0422.678.0426.048.0827.0914.0918.1834.0422.0720.05
CoCA [14]9.2842.2728.876.8220.4518.1819.0812.0924.076.0512.0710.0325.0416.6726.0725.068.3325.0218.92
IMMUNE (Ours)13.4622.6813.4311.3620.4518.1817.0412.0921.033.088.047.087.0214.0120.020.000.000.0011.03
MiniGPT-4-13BOriginal17.5340.2142.2713.6422.7327.2710.0415.0818.039.0522.0921.0717.0320.0235.0819.0537.0341.0923.97
FigStep [16]25.7746.3936.0822.7327.2729.5513.0216.0819.0320.0830.0720.0818.0136.0438.0821.0532.0938.0327.17
AdaShield [67]21.6523.7122.6820.4534.0925.0415.059.0215.0311.0625.0321.0114.0322.0123.0723.0528.0215.0919.84
CoCA [14]25.0350.0746.0314.0625.0343.0613.0225.018.3316.0416.6716.6716.6725.0641.6725.0641.6725.0726.13
IMMUNE (Ours)23.7822.6819.599.0918.1829.5520.0814.0815.0611.0813.0813.0210.0818.0322.086.020.000.0014.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.703, + 0.909, + 0.748 + ], + "angle": 0, + "content": "Table 2. Evaluation on MM-Safety Bench. We report Attack Success Rate (ASR) for 6 categories of prohibited scenarios from MM-Safety Bench [38]. For this evaluation, we used GPT-4 as the jailbreak classifier. IMMUNE significantly reduces ASR across all categories. The best results (lowest ASR) are highlighted in bold. All values are reported in %." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.486, + 0.898 + ], + "angle": 0, + "content": "Jailbreak Judge as the oracle classifier. From our evaluation, we draw the following key insights: (1) With the original decoding, ASR is notably high for most models, reaching up to \\(70.62\\%\\) for MiniGPT-4-13B, underscoring both the severity of the attack and the underlying serious safety concern. (2) Across all MLLMs, IMMUNE demonstrates superior performance, substantially reducing ASR compared to other competitive defense strategies. Specifically, on MiniGPT-4-13B, IMMUNE reduces ASR by \\(37.68\\%\\) compared to the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.909, + 0.898 + ], + "angle": 0, + "content": "original decoding and by \\(17.61\\%\\) compared to state-of-art defense strategy AdaShield [67]. (3) We observe that logic-based attacks reveal a model-dependent variation in IMMUNE's performance. Specifically, for logic and persuade-based attacks, IMMUNE achieves an ASR of \\(0\\%\\) across all input types on LLaVA-1.6 and LLaVA-1.5. However, on MiniGPT-4-7B, logic-based attacks with noise inputs result in a significantly higher ASR of \\(44.59\\%\\), compared to CoCA [14] \\((22.97\\%)\\) and AdaShield [67] \\((32.24\\%)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25044" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.908, + 0.322 + ], + "angle": 0, + "content": "
ModelDefense StrategyIllegal ActivityHate SpeechMalware GenerationPhysical HarmFraudAdult ContentPrivacy ViolationLegal OpinionFinancial AdviceHealth ConsultationAverage
LLaVA-1.6Original11.700.0021.9211.880.0026.270.0036.0431.809.5314.68
FigStep [16]9.620.0014.232.030.0011.950.006.0513.586.245.90
AdaShield [67]3.580.005.650.000.0011.790.000.009.610.003.62
CoCA [14]13.624.296.065.978.3320.396.5727.7221.7017.5613.34
IMMUNE (Ours)12.040.000.000.000.000.000.000.000.000.001.21
LLaVA-1.5Original74.000.0083.9761.6210.4693.961.7292.3782.2984.0958.58
FigStep [16]65.531.9570.0748.473.7773.684.1371.7276.2967.5148.48
AdaShield [67]14.380.0011.5413.880.002.100.002.2012.2114.167.24
CoCA [14]44.038.2338.0422.086.4942.489.6544.1541.9430.4528.63
Ours28.210.006.302.120.000.000.004.403.841.864.23
MiniGPT-4-7BOriginal10.152.1416.022.440.0019.770.0033.6226.2111.9311.95
FigStep [16]11.630.0024.450.003.7513.780.0010.0615.963.998.32
AdaShield [67]4.290.0012.175.674.072.120.0012.128.239.645.76
CoCA [14]7.850.000.002.471.856.360.0015.6313.876.285.74
IMMUNE (Ours)7.980.009.826.140.004.430.007.855.713.594.43
MiniGPT-4-13BOriginal20.050.0030.4115.980.0042.140.0037.8146.2819.5321.17
FigStep [16]24.110.0038.4521.880.0029.570.0021.8334.5023.8119.52
AdaShield [67]18.060.007.663.930.0011.551.8228.2029.9212.2611.65
CoCA [14]24.38.3313.5718.125.5431.928.3313.9817.8318.3815.93
IMMUNE (Ours)16.312.248.275.540.0010.180.0016.0619.5315.989.48
Qwen-VLOriginal17.530.0040.2317.770.0032.380.0056.3439.7539.8224.01
FigStep [16]1.660.002.200.000.0020.870.0010.1210.080.004.56
AdaShield [67]4.050.006.499.680.0017.830.006.061.550.005.04
CoCA [14]29.530.0067.8834.490.0077.920.0088.2683.5661.9243.92
IMMUNE (Ours)6.500.004.370.000.009.930.007.935.880.003.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.332, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Table 3. Evaluation on FigStep. We report the Attack Success Rate (ASR) for all categories from the FigStep benchmark [16]. For this evaluation, we used GPT-4 as the jailbreak classifier. The best results (lowest ASR) are highlighted in bold. All values are reported in %." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.369, + 0.483, + 0.399 + ], + "angle": 0, + "content": "Evaluation on Image-based attacks. We evaluate IMMUNE on three standard benchmarks for image-based attacks:" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.401, + 0.484, + 0.641 + ], + "angle": 0, + "content": "(1) MM-Safety. We report the ASR for 6 prohibited categories from MM-Safety Bench [38] across four recent MLLMs, as shown in Table 2. For response evaluation, following Liu et al. [38], we used GPT-4 as the jailbreak classifier. Our findings indicate that IMMUNE significantly reduces ASR across all MLLMs compared to state-of-art defense strategies such as CoCA [14] and AdaShield [67]. Notably, on LLaVA-1.6, IMMUNE achieves an ASR reduction of \\(44.0\\%\\) and \\(20.46\\%\\) compared to the base model and AdaShield [67] respectively. Across all MLLMs, we observe that attacks using typography images (TYPO) and those combining stable diffusion visuals with typography subtitles (SD-TYPO) are significantly more effective than simple stable diffusion-generated (SD) images. Additionally, samples from the \"Pornography\" category exhibit the highest ASR across all defense mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.763 + ], + "angle": 0, + "content": "(2) FigStep. Table 3 presents the ASR results for various MLLMs on the FigStep dataset [16]. Consistent with results on MM-Safety, IMMUNE demonstrates superior performance over competitive baselines, reducing ASR by \\(6.45\\%\\) and \\(12.13\\%\\) on MiniGPT-4-13B and LLaVA-1.6, respectively, compared to CoCA [14]. Among all categories, \"Hate Speech\" and \"Privacy Violation\" have the lowest ASR, with most approaches achieving an ASR of near \\(0.0\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.87 + ], + "angle": 0, + "content": "(3) Visual Adversarial Attacks. Finally, in Table 4, we evaluate ASR under adversarily optimized jailbreak attacks [54] across varying attack strengths, denoted by \\(\\epsilon\\), where higher \\(\\epsilon\\) values correspond to stronger attacks. The \"Unconstrained\" setting represents the most challenging attack scenario. Notably, even under the \"unconstrained\" attack, IMMUNE reduces ASR by \\(24.19\\%\\) compared to baseline [67] on LLaVA-1.6." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Additional discussion on inference time of IMMUNE and its impact on model utility is deferred to Appendix F." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.367, + 0.907, + 0.635 + ], + "angle": 0, + "content": "
ModelDefense StrategyAttack Strength
ε = 16/255ε = 32/255ε = 64/255Unconstrained
LLaVA-1.6Original66.1264.0862.3164.80
FigStep [16]49.1050.3047.2251.13
AdaShield [67]35.2139.7134.0540.21
CoCA [14]45.2146.1148.0250.33
IMMUNE (Ours)13.3214.9715.1816.02
LLaVA-1.5Original68.0067.1966.2367.33
FigStep [16]53.9852.0751.658.17
AdaShield [67]27.0024.1526.9130.10
CoCA [14]44.2341.8842.9144.82
IMMUNE (Ours)12.4516.7114.3217.03
MiniGPT-4-7BOriginal53.4058.1258.7961.23
FigStep [16]47.0952.4355.2557.10
AdaShield [67]39.0042.4543.7149.63
CoCA [14]53.1960.0962.7868.29
IMMUNE (Ours)18.4220.7721.1822.95
MiniGPT-4-13BOriginal62.5766.1367.8270.11
FigStep [16]59.8161.0562.4465.79
AdaShield [67]47.8250.3751.6853.15
CoCA [14]71.8574.3175.7275.09
IMMUNE (Ours)23.4125.8226.0226.02
Qwen-VLOriginal38.1148.0348.0349.47
FigStep [16]34.2536.9240.1742.78
AdaShield [67]17.3218.2022.4125.24
CoCA [14]38.1544.7346.3649.04
IMMUNE (Ours)13.2215.0715.0716.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Table 4. Evaluation on Visual Adversarial Attacks [54]. We report ASR (in %) on images optimized with varying levels of adversarial noise, denoted by \\(\\epsilon\\), with \"unconstrained\" representing the most challenging scenario." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.723, + 0.634, + 0.74 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we introduce IMMUNE, an inference-time defense framework designed to protect MLLMs against jailbreak attacks. IMMUNE employs a safety-aware reward model to align the MLLM during inference, with the goal of mitigating jailbreak vulnerabilities. We present a mathematical formulation of our defense framework, framing jailbreak mitigation as an alignment problem. Through extensive experiments, we demonstrate that IMMUNE consistently and significantly outperforms competitive baselines, enhancing model safety while preserving the model utility." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25045" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.225 + ], + "angle": 0, + "content": "[2] Maksym Andriushchenko, Francesco Croce, and Nicolas Flammarion. Jailbreaking leading safety-aligned llms with simple adaptive attacks. arXiv preprint arXiv:2404.02151, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.484, + 0.296 + ], + "angle": 0, + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.484, + 0.366 + ], + "angle": 0, + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[5] James Beetham, Souradip Chakraborty, Mengdi Wang, Furong Huang, Amrit Singh Bedi, and Mubarak Shah. Liar: Leveraging alignment (best-of-n) to jailbreak llms in seconds. arXiv preprint arXiv:2412.05232, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.484, + 0.491 + ], + "angle": 0, + "content": "[6] Nicholas Carlini, Anish Athalye, Nicolas Papernot, Wieland Brendel, Jonas Rauber, Dimitris Tsipras, Ian Goodfellow, Aleksander Madry, and Alexey Kurakin. On evaluating adversarial robustness. arXiv preprint arXiv:1902.06705, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.484, + 0.549 + ], + "angle": 0, + "content": "[7] Souradip Chakraborty, Soumya Suvra Ghosal, Ming Yin, Dinesh Manocha, Mengdi Wang, Amrit Singh Bedi, and Furong Huang. Transfer q star: Principled decoding for llm alignment, 2024. 2, 3, 4, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.484, + 0.606 + ], + "angle": 0, + "content": "[8] Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J Pappas, and Eric Wong. Jailbreaking black box large language models in twenty queries. arXiv preprint arXiv:2310.08419, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.484, + 0.689 + ], + "angle": 0, + "content": "[9] Patrick Chao, Edoardo Debenedetti, Alexander Robey, Maksym Andriushchenko, Francesco Croce, Vikash Sehwag, Edgar Dobriban, Nicolas Flammarion, George J. Pappas, Florian Tramer, Hamed Hassani, and Eric Wong. Jailbreakbench: An open robustness benchmark for jailbreaking large language models. In NeurIPS Datasets and Benchmarks Track, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[10] Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14239-14250, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\)*chatgpt quality, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Yinpeng Dong, Huanran Chen, Jiawei Chen, Zhengwei Fang, Xiao Yang, Yichi Zhang, Yu Tian, Hang Su, and Jun Zhu. How robust is google's bard to adversarial image attacks? arXiv preprint arXiv:2309.11751, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[13] Yanrui Du, Sendong Zhao, Ming Ma, Yuhan Chen, and Bing Qin. Analyzing the inherent response tendency of llms: Real-world instructions-driven jailbreak. arXiv preprint arXiv:2312.04127, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[14] Jiahui Gao, Renjie Pi, Tianyang Han, Han Wu, Lanqing Hong, Lingpeng Kong, Xin Jiang, and Zhenguo Li. Coca: Regaining safety-awareness of multimodal large language models with constitutional calibration. arXiv preprint arXiv:2409.11365, 2024. 1, 2, 3, 6, 7, 8, 15, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[15] Simon Geisler, Tom Wollschläger, M. H. I. Abdalla, Johannes Gasteiger, and Stephan Gunnemann. Attacking large language models with projected gradient descent, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.907, + 0.332 + ], + "angle": 0, + "content": "[16] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via topographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 2, 3, 6, 7, 8, 15, 17, 18, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.907, + 0.403 + ], + "angle": 0, + "content": "[17] Jindong Gu, Zhen Han, Shuo Chen, Ahmad Beirami, Bailan He, Gengyuan Zhang, Ruotong Liao, Yao Qin, Volker Tresp, and Philip Torr. A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[18] Xingang Guo, Fangxu Yu, Huan Zhang, Lianhui Qin, and Bin Hu. Cold-attack: Jailbreaking Ilms with stealthiness and controllability. arXiv preprint arXiv:2402.08679, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[19] Dongchen Han, Xiaojun Jia, Yang Bai, Jindong Gu, Yang Liu, and Xiaochun Cao. Ot-attack: Enhancing adversarial transferability of vision-language models via optimal transport optimization. arXiv preprint arXiv:2312.04403, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.906, + 0.572 + ], + "angle": 0, + "content": "[20] Seungju Han, Kavel Rao, Allyson Ettinger, Liwei Jiang, Bill Yuchen Lin, Nathan Lambert, Yejin Choi, and Nouha Dziri. Wildguard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of llms. arXiv preprint arXiv:2406.18495, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.906, + 0.617 + ], + "angle": 0, + "content": "[21] Jonathan Hayase, Ema Borevkovic, Nicholas Carlini, Florian Tramér, and Milad Nasr. Query-based adversarial prompt generation. arXiv preprint arXiv:2402.12329, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[22] Zhang-Wei Hong, Idan Shenfeld, Tsun-Hsuan Wang, Yung-Sung Chuang, Aldo Pareja, James Glass, Akash Srivastava, and Pulkit Agrawal. Curiosity-driven red-teaming for large language models. arXiv preprint arXiv:2402.19464, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[23] MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR), 51 (6):1-36, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[24] Sandy Huang, Nicolas Papernot, Ian Goodfellow, Yan Duan, and Pieter Abbeel. Adversarial attacks on neural network policies. arXiv preprint arXiv:1702.02284, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.907, + 0.829 + ], + "angle": 0, + "content": "[25] Yangsibo Huang, Samyak Gupta, Mengzhou Xia, Kai Li, and Danqi Chen. Catastrophic jailbreak of open-source llms via exploiting generation. arXiv preprint arXiv:2310.06987, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Llm-based input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25046" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[27] Erik Jones, Anca Dragan, Aditi Raghunathan, and Jacob Steinhardt. Automatically auditing large language models via discrete optimization. In International Conference on Machine Learning, pages 15307-15329. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.485, + 0.19 + ], + "angle": 0, + "content": "[28] Maxim Khanov, Jirayu Burapacheep, and Yixuan Li. ARGs: Alignment as reward-guided search. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[29] Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, et al. Rewardbench: Evaluating reward models for language modeling. arXiv preprint arXiv:2403.13787, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[30] Seanie Lee, Minsu Kim, Lynn Cherif, David Dobre, Juho Lee, Sung Ju Hwang, Kenji Kawaguchi, Gauthier Gidel, Yoshua Bengio, Nikolay Malkin, and Moksh Jain. Learning diverse attacks on large language models for robust red-teaming and safety tuning, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[31] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665, 2023. 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[32] Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, and Qi Liu. Red teaming visual language models. arXiv preprint arXiv:2401.12915, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.484, + 0.487 + ], + "angle": 0, + "content": "[33] Yuhui Li, Fangyun Wei, Jinjing Zhao, Chao Zhang, and Hongyang Zhang. Rain: Your language models can align themselves without finetuning. arXiv preprint arXiv:2309.07124, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[34] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 6, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.2,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[36] Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, et al. Agentbench: Evaluating llms as agents. arXiv preprint arXiv:2308.03688, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.484, + 0.686 + ], + "angle": 0, + "content": "[37] Xin Liu, Yichen Zhu, Yunshi Lan, Chao Yang, and Yu Qiao. Query-relevant images jailbreak large multi-modal models, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[38] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models, 2024. 2, 7, 8, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[39] Yi Liu, Gelei Deng, Zhengzi Xu, Yuekang Li, Yaowen Zheng, Ying Zhang, Lida Zhao, Tianwei Zhang, Kailong Wang, and Yang Liu. Jailbreaking chatgpt via prompt engineering: An empirical study. arXiv preprint arXiv:2305.13860, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[40] Weidi Luo, Siyuan Ma, Xiaogeng Liu, Xiaoyu Guo, and Chaowei Xiao. Jailbreakv-28k: A benchmark for assessing the robustness of multimodal large language models against jailbreak attacks. arXiv preprint arXiv:2404.03027, 2024. 2, 3, 6, 7, 15, 16, 18, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[41] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.134 + ], + "angle": 0, + "content": "Tu. Macaw-lm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.908, + 0.204 + ], + "angle": 0, + "content": "[42] Neal Mangaokar, Ashish Hooda, Jihye Choi, Shreyas Chandrashekaran, Kassem Fawaz, Somesh Jha, and Atul Prakash. Prp: Propagating universal perturbations to attack large language model guard-rails. arXiv preprint arXiv:2402.15911, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.289 + ], + "angle": 0, + "content": "[43] Ninareh Mehrabi, Palash Goyal, Anil Ramakrishna, Jwala Dhamala, Shalini Ghosh, Richard Zemel, Kai-Wei Chang, Aram Galstyan, and Rahul Gupta. Jab: Joint adversarial prompting and belief augmentation. In R0-FoMo: Robustness of Few-shot and Zero-shot Learning in Large Foundation Models. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.908, + 0.373 + ], + "angle": 0, + "content": "[44] Ninareh Mehrabi, Ahmad Beirami, Fred Morstatter, and Aram Galstyan. Robust conversational agents against imperceptible toxicity triggers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2831-2847, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.907, + 0.445 + ], + "angle": 0, + "content": "[45] Ninareh Mehrabi, Palash Goyal, Christophe Dupuy, Qian Hu, Shalini Ghosh, Richard Zemel, Kai-Wei Chang, Aram Galstyan, and Rahul Gupta. Flirt: Feedback loop in-context red teaming. Empirical Methods in Natural Language Processing (EMNLP), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.907, + 0.488 + ], + "angle": 0, + "content": "[46] Youssef Mroueh. Information theoretic guarantees for policy alignment in large language models. arXiv preprint arXiv:2406.05883, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.908, + 0.572 + ], + "angle": 0, + "content": "[47] Sidharth Mudgal, Jong Lee, Harish Ganapathy, YaGuang Li, Tao Wang, Yanping Huang, Zhifeng Chen, Heng-Tze Cheng, Michael Collins, Trevor Strohman, Jilin Chen, Alex Beutel, and Ahmad Beirami. Controlled decoding from language models. International Conference on Machine Learning (ICML), 2024. 2, 3, 4, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.908, + 0.642 + ], + "angle": 0, + "content": "[48] Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.908, + 0.686 + ], + "angle": 0, + "content": "[49] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.908, + 0.77 + ], + "angle": 0, + "content": "[50] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.908, + 0.814 + ], + "angle": 0, + "content": "[51] Junsoo Park, Seungyeon Jwa, Meiying Ren, Daeyoung Kim, and Sanghyuk Choi. Offsetbias: Leveraging debiased data for tuning evaluators. arXiv preprint arXiv:2407.06551, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[52] Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia Glaese, Nat McAleese, and Geoffrey Irving. Red teaming language models with language models. arXiv preprint arXiv:2202.03286, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.908, + 0.903 + ], + "angle": 0, + "content": "[53] Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25047" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "Mllm-protector: Ensuring mllm's safety without hurting performance. arXiv preprint arXiv:2401.02906, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[54] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.193, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[55] Xiangyu Qi, Ashwinee Panda, Kaifeng Lyu, Xiao Ma, Subhrajit Roy, Ahmad Beirami, Prateek Mittal, and Peter Henderson. Safety alignment should be made more than just a few tokens deep, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.484, + 0.318 + ], + "angle": 0, + "content": "[56] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to! International Conference on Learning Representations (2024), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.388 + ], + "angle": 0, + "content": "[57] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 3, 4, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.391, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.462, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[59] Christian Schlarmann and Matthias Hein. On the adversarial robustness of multi-modal foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3677-3685, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.519, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[60] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Plug and pray: Exploiting off-the-shelf components of multi-modal models. arXiv preprint arXiv:2307.14539, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.562, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[61] Chawin Sitawarin, Norman Mu, David Wagner, and Alexandre Araujo. Pal: Proxy-guided black-box attack on large language models. arXiv preprint arXiv:2402.09674, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.605, + 0.484, + 0.658 + ], + "angle": 0, + "content": "[62] Bharath K. Sriperumbudur, Kenji Fukumizu, Arthur Gretton, Bernhard Schölkopf, and Gert R. G. Lanckriet. On integral probability metrics, phi-divergences and binary classification, 2009. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[63] Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul Christiano. Learning to summarize from human feedback, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[64] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing nlp. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2153-2162, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[65] Boxin Wang, Weixin Chen, Hengzhi Pei, Chulin Xie, Mintong Kang, Chenhui Zhang, Chejian Xu, Zidi Xiong, Ritik Dutta, Ryan Schaeffer, et al. Decodingtrust: A comprehensive assessment of trustworthiness in gpt models. In NeurlPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[66] Hao Wang, Hao Li, Minlie Huang, and Lei Sha. From noise to clarity: Unraveling the adversarial suffix of large language" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "model attacks via translation of text embeddings. arXiv preprint arXiv:2402.16006, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.907, + 0.192 + ], + "angle": 0, + "content": "[67] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. arXiv preprint arXiv:2403.09513, 2024. 1, 2, 3, 6, 7, 8, 15, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.907, + 0.264 + ], + "angle": 0, + "content": "[68] Zhilin Wang, Yi Dong, Olivier Delalleau, Jiaqi Zeng, Gerald Shen, Daniel Egert, Jimmy J Zhang, Makesh Narsimhan Sreedhar, and Oleksii Kuchaiev. Helpsteer2: Open-source dataset for training top-performing reward models. arXiv preprint arXiv:2406.08673, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.267, + 0.907, + 0.321 + ], + "angle": 0, + "content": "[69] Fenghua Weng, Yue Xu, Chengyan Fu, and Wenjie Wang. MMJ-Bench: A comprehensive study on jailbreak attacks and defenses for vision language models. arXiv preprint arXiv:2408.08464, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.324, + 0.907, + 0.38 + ], + "angle": 0, + "content": "[70] Nevan Wichers, Carson Denison, and Ahmad Beirami. Gradient-based language model red teaming. 18th Conference of European Chapter of the Assoc for Computational Linguistics (EACL), 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.382, + 0.906, + 0.437 + ], + "angle": 0, + "content": "[71] Qi Wu, Damien Teney, Peng Wang, Chunhua Shen, Anthony Dick, and Anton Van Den Hengel. Visual question answering: A survey of methods and datasets. Computer Vision and Image Understanding, 163:21-40, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.44, + 0.906, + 0.495 + ], + "angle": 0, + "content": "[72] Nan Xu, Fei Wang, Ben Zhou, Bang Zheng Li, Chaowei Xiao, and Muhao Chen. Cognitive overload: Jailbreaking large language models with overloaded logical thinking. arXiv preprint arXiv:2311.09827, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.906, + 0.539 + ], + "angle": 0, + "content": "[73] Yue Xu, Xiuyuan Qi, Zhan Qin, and Wenjie Wang. Defending jailbreak attack in vlms via cross-modality information detector. arXiv preprint arXiv:2407.21659, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.542, + 0.906, + 0.598 + ], + "angle": 0, + "content": "[74] Zhangchen Xu, Fengqing Jiang, Luyao Niu, Yuntian Deng, Radha Poovendran, Yejin Choi, and Bill Yuchen Lin. Magpie: Alignment data synthesis from scratch by prompting aligned lms with nothing, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.906, + 0.656 + ], + "angle": 0, + "content": "[75] Zonghao Ying, Aishan Liu, Tianyuan Zhang, Zhengmin Yu, Siyuan Liang, Xianglong Liu, and Dacheng Tao. Jailbreak vision language models via bi-modal adversarial prompt. arXiv preprint arXiv:2406.04031, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.658, + 0.906, + 0.741 + ], + "angle": 0, + "content": "[76] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13807-13816, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.906, + 0.799 + ], + "angle": 0, + "content": "[77] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 2, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[78] Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang. Rrrh: Rank responses to align language models with human feedback without tears, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[79] Yi Zeng, Hongpeng Lin, Jingwen Zhang, Diyi Yang, Ruoxi Jia, and Weiyan Shi. How johnny can persuade llms to jailbreak them: Rethinking persuasion to challenge ai safety by humanizing llms. arXiv preprint arXiv:2401.06373, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25048" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.134 + ], + "angle": 0, + "content": "[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[81] Xiaoyu Zhang, Cen Zhang, Tianlin Li, Yihao Huang, Xiaojun Jia, Xiaofei Xie, Yang Liu, and Chao Shen. A mutation-based method for multi-modal jailbreaking attack detection. arXiv preprint arXiv:2312.10766, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[82] Zhuo Zhang, Guangyu Shen, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Make them spill the beans! coercive knowledge extraction from (production) llms. arXiv preprint arXiv:2312.04782, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[83] Xuandong Zhao, Xianjun Yang, Tianyu Pang, Chao Du, Lei Li, Yu-Xiang Wang, and William Yang Wang. Weak-to-strong jailbreaking on large language models. arXiv preprint arXiv:2401.17256, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.373 + ], + "angle": 0, + "content": "[84] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[85] Yukai Zhou, Zhijie Huang, Feiyang Lu, Zhan Qin, and Wenjie Wang. Don't say no: Jailbreaking llm by suppressing refusal. arXiv preprint arXiv:2404.16369, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.483, + 0.472 + ], + "angle": 0, + "content": "[86] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[87] Sicheng Zhu, Ruiyi Zhang, Bang An, Gang Wu, Joe Barrow, Zichao Wang, Furong Huang, Ani Nenkova, and Tong Sun. Autodan: interpretable gradient-based adversarial attacks on large language models. In First Conference on Language Modeling, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[88] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[89] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. arXiv preprint arXiv:2402.02207, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.483, + 0.713 + ], + "angle": 0, + "content": "[90] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.483, + 0.784 + ], + "angle": 0, + "content": "[91] Andy Zou, Long Phan, Justin Wang, Derek Duenas, Maxwell Lin, Maksym Andriushchenko, Rowan Wang, Zico Kolter, Matt Fredrikson, and Dan Hendrycks. Improving alignment and robustness with short circuiting. arXiv preprint arXiv:2406.04313, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25049" + } + ] +] \ No newline at end of file diff --git a/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/66fe35a0-e089-4c9c-afba-719727b0b2df_origin.pdf b/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/66fe35a0-e089-4c9c-afba-719727b0b2df_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..88afa36c47a4388aea598f4166bcd8af824b4a99 --- /dev/null +++ b/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/66fe35a0-e089-4c9c-afba-719727b0b2df_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:909e30163de6ef0a7f6a19aa1f831a7f0a6f9bc659cc64369328de6374806b2b +size 2669754 diff --git a/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/full.md b/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1b21937922d166b28934fc918d4f52e80e4b528f --- /dev/null +++ b/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/full.md @@ -0,0 +1,330 @@ +# IMMUNE: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment + +Soumya Suvara Ghosal $^{1*}$ Souradip Chakraborty $^{1*}$ Vaibhav Singh $^{2*}$ Tianrui Guan $^{1}$ Mengdi Wang $^{3}$ Ahmad Beirami Furong Huang $^{1}$ Alvaro Velasquez $^{4}$ Dinesh Manocha $^{1\dagger}$ Amrit Singh Bedi $^{5\dagger}$ + +1University of Maryland; 2Indian Institute of Technology Bombay; 3Princeton University; 4University of Colorado Boulder; 5University of Central Florida + +# Abstract + +With the widespread deployment of Multimodal Large Language Models (MLLMs) for visual-reasoning tasks, improving their safety has become crucial. Recent research indicates that despite training-time safety alignment, these models remain vulnerable to jailbreak attacks. In this work, we first highlight an important safety gap to describe that alignment achieved solely through safety training may be insufficient against jailbreak attacks. To address this vulnerability, we propose IMMUNE, an inference-time defense framework that leverages a safety reward model through controlled decoding to defend against jailbreak attacks. Additionally, we provide a mathematical characterization of IMMUNE, offering insights on why it improves safety against jailbreaks. Extensive evaluations on diverse jailbreak benchmarks using recent MLLMs reveal that IMMUNE effectively enhances model safety while preserving the model's original capabilities. For instance, against text-based jailbreak attacks on LLaVA-1.6, IMMUNE reduces the attack success rate by $57.82\%$ and $16.78\%$ compared to the base MLLM and state-of-the-art defense strategy, respectively. + +# 1. Introduction + +Multi-modal large language models (MLLMs) have made remarkable progress in vision-language reasoning tasks such as visual question-answering [1, 3, 71] and image captioning [1, 23]. However, ensuring their outputs are safe and free from discrimination, disinformation, and harmful content is important for their adoption. Safety alignment through reinforcement learning from human feedback (RLHF) has shown promise in aligning model behavior with social norms to prevent harmful outputs [63, 78, 88]. Despite these sig + +nificant alignment efforts, recent studies reveal that MLLMs remain vulnerable to jailbreaking attacks, where malicious image-text pairs bypass safety mechanisms, raising safety concerns [19, 32, 36, 48, 54, 59, 60, 65]. + +Why training-time safety alignment might not be enough? Fine-tuning-based safety alignment approaches rely on a static prompt distribution, leaving them inherently vulnerable to jailbreak attacks. These attacks exploit this limitation by solving a prompt (text or image) optimization problem to craft adversarial prompts that bypass the safety mechanisms of aligned models [54, 90]. In this work, we argue that safety finetuning against jailbreak attacks while keeping model capabilities may be inherently challenging given that the adversary can find a prompt that bypasses the safety mechanisms of the model. This underscores the fundamental limitation of relying solely on training-time safety alignment procedures to ensure robustness. These leads us to ask: Is it possible to design a safeguarding strategy to defend against given possible adversarial prompts? + +In this paper, we take a step towards answering this question in the affirmative. We propose shifting from training-time alignment to inference-time safeguarding. Although initial efforts in this direction [14, 67] show promise, they exhibit weak empirical performance. Further, Weng et al. [69] note that adding a safety prompt to the input [67] can lead to overly cautious behavior, resulting in generic responses like, "I am sorry..." even for benign queries (see Figure 1), thereby affecting model utility. In real-world applications, an ideal jailbreaking defense should not only minimize the success rate of attacks but also maintain high utility by preserving model performance on legitimate queries. Moreover, existing inference-time approaches lack mathematical characterization, relying primarily on heuristic-based methods. This gap motivates the need for robust inference-time safeguarding mechanisms, which we address in this paper. + +Our key idea and proposed approach. We present IMMUNE, + +![](images/46403663e70204d7a71a27780e55725a487836daee282465f4c76f052531486d.jpg) +Figure 1. Qualitative Evaluation (Left). Given an image of a house generated by stable diffusion [58] and perturbed by adversarial noise [54], along with a malicious user query asking for steps to "break into and rob a house", we visualize responses from the base model and various inference-time defense strategies, including CoCA [14] and AdaShield [67]. We observe that all compared defense strategies are misled into generating harmful content. In contrast, our proposed inference-time safety alignment framework, IMMUNE, effectively rejects the user query, citing its unethical nature. This evaluation underscores the importance of inference-time alignment in preventing harmful responses. We visualize additional responses in the Appendix. Quantitative Evaluation (Right). To empirically validate the effectiveness of IMMUNE, we compare the attack success rates and model utility across various state-of-the-art defense strategies. A lower attack success rate reflects improved safety in generated outputs. Our results indicate that IMMUNE substantially lowers the attack success rate compared to other baselines. Additionally, IMMUNE not only strengthens model safety but also preserves the model's original utility, as demonstrated by its performance on MM-Vet benchmark [77]. We note that an ideal defense strategy should have a low attack success rate while maintaining high utility, i.e., towards the bottom right corner of the plot. Refer to Section 6 for further details. + +![](images/8e2f14877fa4301de8388ba74ff800efafe3ebe60349a58f78a77dc74bb73277.jpg) + +a framework that reframes jailbreak defense as an inference-time alignment problem. Our approach is built on two key components: First, we show that jailbreaks can be effectively mitigated by aligning the MLLM during inference using controlled decoding [7, 47], assuming access to a safety-aware utility or reward function (without loss of generality). Second, we mathematically formalize this defense as an inference-time alignment problem within the KL-regularized reinforcement learning framework [50], proving its robustness against adversarial prompt distributions. As illustrated in Figure 1, our comprehensive evaluations demonstrate that IMMUNE outperforms state-of-the-art inference-time defense strategies, such as AdaShield [67] and CoCA [14], while maintaining high model utility. Our key contributions are as follows: + +- Insufficiency of safety finetuning for MLLMs. We first argue that the current alignment achieved solely via safety training/fine-tuning with RLHF is insufficient to defend against jailbreak attacks, thereby highlighting the need for additional safeguards. +- Improving safety via inference time alignment. To address these safety concerns, we propose IMMUNE, an inference-time framework for defense against jailbreak attacks through controlled decoding using a safety reward model. We provide a characterization of our defense mechanism by framing it as an alignment problem, offering mathematical guarantees against jailbreak attacks. +- Empirical evaluations. We empirically validate the efficacy of our proposed defense strategy, IMMUNE, + +across a range of jailbreak benchmarks, including MM-SafetyBench [38], FigStep [16], visual adversarial attacks [54], and text-based jailbreak attacks [40]. Our evaluations demonstrate significant improvements on several recent MLLMs, such as LLaVA-1.5 [34], LLaVA-1.6 [35], MiniGPT-4 [86], and Qwen-VL [4]. + +# 2. Related Works + +Jailbreak attacks. Jones et al. [27] frames jailbreaking large language models (LLMs) as a discrete optimization problem. They search for the optimal suffix that, when appended to the original prompt, can greedily generate the desired harmful output. Zou et al. [90] demonstrated that model inputs could be manipulated based on gradients to force-generate responses to harmful queries. Zhu et al. [87] introduced a method to generate an adversarial suffix iteratively. This approach ensures that the optimized suffix remains meaningful while bypassing perplexity-based filters. Other works that proposed jailbreaking by manipulating gradients include Andriushchenko et al. [2], Geisler et al. [15], Hayase et al. [21], Mangaokar et al. [42], Sitawarin et al. [61], Wang et al. [66]. Another line of work, exemplified by Zhang et al. [82], involves optimizing the prompt iteratively until the distribution of output tokens results in generating harmful responses. For example, Zhang et al. [82] showed that adversaries can break safety alignment by forcing the target LLM to select a lower-ranked output token, thereby generating toxic con + +tent. Additional works following this domain include Du et al. [13], Guo et al. [18], Huang et al. [25], Zhao et al. [83], Zhou et al. [85]. Recent work has also focused on jailbreaking models through iterative interactions with the model [8, 45]. Finally Qi et al. [56] demonstrated that even benign finetuning of the models may lead to the removal of model safety safeguards. Similar to LLMs, jailbreaking MLLMs can be achieved by either manipulating the visual input [16, 37, 54], the text prompt [40], or a combination of both [75]. Qi et al. [54] demonstrated that adding adversarial noise, which is imperceptible to humans and guided by the model's input gradients, to the input image can effectively deceive MLLMs. Recently, there has been substantial research studying the robustness of MLLMs to adversarial images [12, 19, 49, 59, 60, 84]. Another line of work [16, 37] explores embedding harmful content directly into images via typography or text-to-image tools. Recently, Luo et al. [40] showed that text-based jailbreak prompts [39, 72, 79, 90] developed for LLMs can be effectively transferred to MLLMs and are often more potent than image-based attacks. Additionally, Ying et al. [75] introduced a jail-breaking framework that simultaneously perturbs both visual and textual modalities. See also the survey by Gu et al. [17] on more works on prompt-based attacks against MLLMs. + +Defense via safety alignment (fine-tuning). Defense approaches typically aim to enhance the safety in MLLMs by either fine-tuning the model on a dataset of safe image-text pairs [89] or adversarial training [89]. For instance, Zong et al. [89] proposed a safety instruction-following dataset VLGuard for safeguarding MLLMs. DRESS [10] leverages natural language feedback (NLF) from LLMs to improve the alignment and interactions within MLLMs. MLLM-Protector [53] proposed a plug-and-play strategy combining a lightweight harm detector to identify the harmful response and a detoxifier to correct harmful outputs. However, fine-tuning-based approaches suffer from (i) the requirement of a large annotated dataset consisting of safe image-text pairs, (ii) being computationally intensive, and (iii) requiring access to model weights. Qi et al. [55] highlighted that safety alignment primarily affects the first few output tokens leaving models susceptible to simple jailbreak attacks. Zou et al. [91] introduced circuit-breaking techniques to enhance model robustness. Yu et al. [76] proposed improving the safety of MLLMs by aligning with human preferences through direct preference optimization [57]. + +Defense via inference time alignment (controlled decoding). Khanov et al. [28] proposed a reward-guided inference time alignment. Mudgal et al. [47] proposed controlled decoding through a value function for the reward to guide the language model to more safe outcomes. Chakraborty et al. [7] improved the value function by bridging the gap between the value of the decoding policy and the reference policy. Li et al. [33] proposed leveraging self-evaluation and rewind + +mechanisms in LLMs to generate safe responses. Gong et al. [16] proposed concatenating a fixed safety prompt with the malicious user query to evade jailbreaking. Wang et al. [67] proposed refining the safety prompt iteratively rather than using a fixed template. Mehrabi et al. [43] introduced an interactive approach to improve robustness against adversarial attacks. Zhang et al. [81] proposed mutating untrusted inputs to generate variants, identifying attack samples by examining response discrepancies among the variants. In another approach, Xu et al. [73] detected adversarial images by analyzing cross-modal similarity between harmful queries and perturbed images. Additionally, Gao et al. [14] suggested improving safety by calibrating MLLM output distribution. However, existing approaches compromise model utility, leading to overly cautious responses (Figure 1). + +# 3. Problem Formulation + +# 3.1. Preliminaries + +Multimodal Large Language Models (MLLMs) are autoregressive text generation models that can process multiple modalities, such as image [86], video [80], and audio [41], in addition to text. In this work, we consider input to MLLM in the form of an image and a text prompt. Mathematically, we denote a MLLM as a mapping $\pi_{\mathrm{safe}}: \mathcal{I} \times \mathcal{V}^N \to \mathcal{V}^M$ , where $\mathcal{I}$ represents the image space, and $\mathcal{V}$ denotes the vocabulary set. The input to the MLLM $\pi_{\mathrm{safe}}$ consists of an image $\mathbf{I} \in \mathcal{I}$ and a sequence of tokens $\mathbf{x} := \{x_1, x_2, \dots, x_N\}$ (referred as prompt), where each token $x_i \in \mathcal{V}$ . Given $\mathbf{x}_{\mathrm{input}} := [\mathbf{I}, \mathbf{x}]$ , the MLLM $\pi_{\mathrm{safe}}(\cdot | \mathbf{x}_{\mathrm{input}})$ generates the output sequence $\mathbf{y} := \{y_1, y_2, \dots, y_M\}$ , where $y_i \in \mathcal{V}$ , in a token by token fashion. + +Jailbreaking MLLMs. Despite safety tuning efforts aimed at preventing MLLMs from generating harmful content, jailbreak attacks as formulated in Qi et al. [54], Zou et al. [90] can still make the safety aligned model $\pi_{\mathrm{safe}}$ to generate harmful outputs with the help of an adversarial prompt $\mathbf{x}_{\mathrm{adv}}$ obtained by solving + +$$ +\mathbf {x} _ {\mathrm {a d v}} := \arg \min _ {\mathbf {q} \in \mathcal {B}} \sum_ {i = 1} ^ {m} - \log \left(\pi_ {\text {s a f e}} \left(y _ {i} \mid \left[ \mathbf {x} _ {\text {i n p u t}}, \mathbf {q} \right]\right)\right), \tag {1} +$$ + +where $\mathcal{B}$ is a constraint on the adversarial prompts and $\{y_i\}_{i=1}^m$ are few initial tokens of intended harmful response. For example, in Figure 1 (Left), we show that given an image of a house generated by stable diffusion [58] and then perturbed with adversarial noise, along with a malicious user query to explain "How to break into and rob a house", the model is forced to generate unsafe responses, effectively bypassing its safety alignment. + +# 3.2. Safeguard Against Jailbreaks + +From the definition of adversarial prompt in (1), we observe that $\mathbf{x}_{\mathrm{adv}}$ can lead the MLLM model to generate a harmful + +response, $\mathbf{y}_{\mathrm{unsafe}} \sim \pi_{\mathrm{safe}}(\cdot|\mathbf{x}_{\mathrm{adv}})$ , despite the training time safety alignment of $\pi_{\mathrm{safe}}$ . This raises the question: Why do insufficiencies exist in safety-aligned models that make them vulnerable to attacks, and how can we mitigate them? To answer this question, we start by noticing that an attacker's goal is to learn an adversarial prompt distribution $p_{\mathrm{adv}}$ by maximizing some "success metric" subject to a "cost metric," which is general and covers existing attacks such as projected gradient descent (PGD) attacks [15], greedy coordinate gradient (GCG) attacks [87, 90], universal adversarial trigger (UAT) attacks [44, 64], and red teaming attacks [22, 30, 52, 70]. Mathematically, an adversary can solve the following optimization problem to learn an adversarial prompt: + +$$ +\begin{array}{l} p _ {\mathrm {a d v}} := \arg \min _ {p} \mathbb {E} _ {\mathbf {q} \sim p (\cdot | \mathbf {x} _ {\text {i n p u t}}), \mathbf {y} \sim \pi_ {\text {s a f e}} (\cdot | \mathbf {q})} \left[ R _ {\text {s a f e}} (\mathbf {x} _ {\text {i n p u t}}, \mathbf {y}) \right] \\ + \beta \operatorname {d i v} \left(p (\cdot | \mathbf {x} _ {\text {i n p u t}}), p _ {0} (\cdot | \mathbf {x} _ {\text {i n p u t}})\right), \tag {2} \\ \end{array} +$$ + +where the "cost metric" term div $(p(\cdot |\mathbf{x}_{\mathrm{input}}),p_0(\cdot |\mathbf{x}_{\mathrm{input}}))$ helps keep the adversarial prompt close to a given prompt distribution, $p_0$ . We can think of $p_0$ as a given prompt distribution that will sample $\mathbf{x}_{\mathrm{input}}$ only with high likelihood and have minimal probability of sampling other prompts. Then, the divergence part of the objective in (2) will make sure the adversarial prompts are closer to original prompts $\mathbf{x}_{\mathrm{input}}$ , which will keep them interpretable and not completely out of distribution [88]. The formulation in (2) is also in accordance with the existing literature on adversarial attacks in machine learning [6, 24] because it is just trying to ensure that adversarial prompts $\mathbf{x}_{\mathrm{adv}}$ looks like similar to $\mathbf{x}_{\mathrm{input}}$ . The regularization parameter $\beta >0$ would control this cost metric in (2). + +Training-time defense. We note that learning a safe and robust model is inherently a min-max optimization problem given by + +$$ +\begin{array}{l} \min _ {p} \max _ {\pi} \mathbb {E} _ {\mathbf {q} \sim p (\cdot | \mathbf {x} _ {\text {i n p u t}}), \mathbf {y} \sim \pi (\cdot | \mathbf {q})} \left[ R _ {\text {s a f e}} (\mathbf {x} _ {\text {i n p u t}}, \mathbf {y}) \right. \\ + \beta \operatorname {d i v} (p (\cdot | \mathbf {x} _ {\text {i n p u t}}), p _ {0} (\cdot | \mathbf {x} _ {\text {i n p u t}})) \\ \left. - \alpha \operatorname {d i v} \left(\pi (\cdot | \mathbf {q}), \pi_ {\text {r e f}} (\cdot | \mathbf {q})\right) \right], \tag {3} \\ \end{array} +$$ + +where the inner maximization captures the safeguarding of the model, e.g., through safety alignment. The safety alignment in recent literature [50, 57] only focused on solving the inner max in (3) for a given prompt distribution (say $p_0$ ) to obtain $\pi_{\mathrm{safe}}$ which is safe only under the prompt distribution $\mathbf{x} \sim p_0$ . An adversary exploits it and learns an adversarial distribution $\rho_{\mathrm{adv}}$ breaking the alignment and generating harmful responses. This is likely the reason why adversarial attacks on images (and other modalities) have remained challenging to address over the past decade. + +Key insight for tractable inference-time defense. We reframe the optimization from $\min_p\max_\pi$ to $\max_{\pi}\min_{p}$ by letting the safeguard be applied after the attacker has chosen + +their jailbreak prompt. While the two formulations are equivalent for convex problems, this reformulation may lead to an improvement in defense due to the minimax duality gap for general non-convex problems. Additionally, this formulation allows an efficient solution of the $\max_{\pi}$ component during inference, leveraging recent advances in inference-time alignment methods through controlled decoding [7, 47]. Notably, inference-time alignment enables prompt-specific safety adjustments, which are challenging to implement during training. This capability is enabled by the user (or attacker) interaction protocol with MLLM. Specifically, the attacker sets the adversarial prompt, which remains unchanged throughout the process, providing a fixed input for the MLLM. This scenario ultimately reduces the challenge of learning a safe and optimal policy within the predefined adversarial prompt environment, effectively solving the problem $\max_{\pi}$ in inference. We will provide a detailed method in the next section. + +# 4. Inference Time Safety Alignment + +In this section, we propose improving the safety of response generation by using alignment through the controlled decoding procedure proposed in [7, 47]. Our key insight is that well-trained safety rewards, denoted by $R_{\mathrm{safe}}(\mathbf{x}, \mathbf{y})$ , are often accessible online, for instance, RewardBench [29]. This safety reward will give high scores for safe text and low scores for unsafe text. We remark here that the reward model itself is likely not ideal and subject to adversarial attack. With this understanding, we propose to solve the following decoding problem at inference time for each token $t$ : + +$$ +\begin{array}{l} \pi_ {\text {s a f e - d e c}} ^ {*} (\cdot | \mathbf {s} _ {t}) := \arg \max _ {\pi} \mathbb {E} _ {z \sim \pi (\cdot | \mathbf {s} _ {t})} [ Q _ {\text {s a f e}} (\mathbf {s} _ {t}, z) ] \tag {4} \\ - \alpha \mathrm {K L} (\pi (\cdot | \mathbf {s} _ {t}) | | \pi_ {\text {s a f e}} (\cdot | \mathbf {s} _ {t})), \\ \end{array} +$$ + +where $\mathbf{s}_t\coloneqq [\mathbf{x}_{\mathrm{adv}},\mathbf{y}_{< t}]$ and $\mathbf{y}_{< t} = [y_0,y_1,\dots ,y_{t - 1}]$ capture the tokens generated so far. In (4), the term $Q_{\mathrm{safe}}(\mathbf{s}_t,z)$ is the action-value function defined as: + +$$ +Q _ {\text {s a f e}} \left(\mathbf {s} _ {t}, z\right) = \mathbb {E} _ {\tau \sim \rho_ {\text {s a f e}} (\cdot | [ \mathbf {s} _ {t}, z ])} \left[ R _ {\text {s a f e}} \left([ \mathbf {s} _ {t}, z ], \tau\right) \right], \tag {5} +$$ + +where $\tau := [z_{t+1}, z_{t+2}, \dots, z_{eos}]$ denotes the remaining trajectory starting from $[\mathbf{s}_t, z]$ , $\rho_{\mathrm{safe}}(\tau | [\mathbf{s}_t, z]) := \prod_{i=t+1}^{\mathrm{eos}} \pi_{\mathrm{safe}}(z_i | [\mathbf{s}_t, z])$ is the trajectory level distribution, where $eos$ denotes end of sentence token, and $\tau$ denotes the sequence of tokens generated by the model starting from $[\mathbf{s}_t, z]$ . The closed form expression of (4) is given by [47, Theorem 2.1]: + +$$ +\pi_ {\text {s a f e - d e c}} ^ {*} (z | \mathbf {s} _ {t}) = \frac {\pi_ {\text {s a f e}} (z | \mathbf {s} _ {t})}{Z} \exp \left(\frac {Q _ {\text {s a f e}} (\mathbf {s} _ {t} , z)}{\alpha}\right), \tag {6} +$$ + +where $Z$ is the normalizing constant. For a given adversarial prompt $\mathbf{x}_{\mathrm{adv}}$ , we propose to follow the decoding policy derived in (6) at the inference time to generate a safe response. Our proposed algorithm is summarized in Algorithm 1. + +Algorithm 1 IMMUNE: Safety against Jailbreaking attacks via Inference-time alignment + +1: Input: Base MLLM $\pi_{\mathrm{safe}}$ , safe reward model $R_{\mathrm{safe}}$ , adversarial input $\mathbf{x}_{\mathrm{adv}}$ , number of tokens sampled $k$ , alignment parameter $\alpha$ , vocabulary set $\mathcal{V}$ . +2: for $t = 0, \dots, T$ do +3: Current state: $\mathbf{s}_t = [\mathbf{x}_{\mathrm{adv}},\mathbf{y}_{< t}]$ , where $\mathbf{x}_{\mathrm{adv}}$ is adversarial input and $\mathbf{y}_{< t} = [y_0,y_1,\dots ,y_{t - 1}]$ +4: Sample top-k tokens using base model $\pi_{\mathrm{safe}}$ and store as $\hat{\mathcal{V}} = \{z_i : z_i \sim \pi_{\mathrm{safe}}(\cdot|\mathbf{s}_t)\}_{i=1}^k$ +5: for $z\in \mathcal{V}$ do +6: Evaluate $Q_{\mathrm{safe}}$ as in Equation (5) +7: Compute decoding score for token $z$ : $g_{z} = \frac{1}{\alpha} Q_{\mathrm{safe}}([s_t,z]) + \log \pi_{\mathrm{saf}}(z|s_t)$ +8: Distribution over $\hat{\mathcal{V}}$ : $f(z|\mathbf{s}_t) = \frac{\exp(g_z)}{\sum_{z'\in\hat{\mathcal{V}}}\exp(g_{z'})}$ +9: Next token $y_{t}\sim f(\cdot |\mathbf{s}_{t})$ +10: Next state $\mathbf{s}_{t + 1}\gets [\mathbf{s}_t,y_t]$ +11: Return $\mathbf{y} = [y_0, \dots, y_T]$ + +# 5. Theoretical Insights + +In this section, we formally derive a bound on the suboptimality of our proposed decoding-based approach under an adversarial prompt distribution. Given a safe prompt distribution $p_0$ and an adversarial prompt distribution $p_{\mathrm{adv}}$ , the sub-optimality gap of our algorithm can be defined as + +$$ +\begin{array}{l} \Delta_ {\text {s u b - g a p}} \left(\mathbf {x} _ {\text {i n p u t}}\right) \tag {7} \\ := \underset { \begin{array}{c} \mathbf {x} \sim p _ {0} (\cdot | \mathbf {x} _ {\text {i n p u t}}) \\ \mathbf {y} \sim \rho_ {*} (\cdot | \mathbf {x}) \end{array} } {\mathbb {E}} \left[ R _ {\text {s a f e}} (\mathbf {x}, \mathbf {y}) \right] - \underset { \begin{array}{c} \mathbf {x} \sim p _ {\text {a d v}} (\cdot | \mathbf {x} _ {\text {i n p u t}}) \\ \mathbf {y} \sim \rho_ {\text {s a f e - d e c}} (\cdot | \mathbf {x}) \end{array} } {\mathbb {E}} \left[ R _ {\text {s a f e}} (\mathbf {x}, \mathbf {y}) \right], \\ \end{array} +$$ + +where $\rho_{\mathrm{safe - dec}}$ is the trajectory level distribution corresponding to token level $\pi_{\mathrm{safe - dec}}^*$ and $\rho_*$ is the trajectory level optimal distribution for safety reward. The suboptimality gap in (7) captures how safe our proposed decoding-based approach is in generating responses to adversarial prompts. The suboptimality gap in (7) is similar but different to the one utilized in [5] because here, the goal is to capture the impact of adversarial prompt distribution specifically. We establish an upper bound in the following Theorem 1 on the suboptimality gap defined in (7). + +Theorem 1. Let $R_{safe}(\mathbf{x},\mathbf{y}) \leq R_{max}$ , $p_0$ is a given prompt distribution, $p_{adv}$ adversarial prompt distribution, $\rho_{*}(\cdot |\mathbf{x})$ denotes the optimal trajectory level distribution for the safe reward, and $\rho_{safe}(\cdot |\mathbf{x})$ denote the given training time safety aligned trajectory level distri + +bution. Then, it holds that + +$$ +\begin{array}{l} \Delta_ {\text {s u b - g a p}} (\mathbf {x} _ {\text {i n p u t}}) \leq R _ {\max } \sqrt {K L (p _ {0} (\cdot | \mathbf {x} _ {\text {i n p u t}}) \| p _ {\text {a d v}} (\cdot | \mathbf {x} _ {\text {i n p u t}}))} \\ + \alpha K L \left(\rho_ {*} \left(\cdot | \mathbf {x} _ {\text {i n p u t}}\right) \| \rho_ {s a f e} (\cdot | \mathbf {x} _ {\text {i n p u t}})\right), \\ \end{array} +$$ + +where $\alpha > 0$ is a regularization parameter balancing the KL-divergence term in the policy alignment. + +Proof Outline and Insights. The proof of Theorem 1 (see Appendix E for details) relies on decomposing the suboptimality gap term $\Delta_{\mathrm{sub - gap}}$ into two components, $\Delta_1$ and $\Delta_2$ such that $\Delta_{\mathrm{sub - gap}} = \Delta_1 + \Delta_2$ , with bounds derived from total variation distance and Pinsker's inequality, respectively. We can write the decomposition as + +$$ +\Delta_{1}:= \mathop{\mathbb{E}}_{\substack{\mathbf{x}\sim p_{0}(\cdot |\mathbf{x}_{\text{input}})\\ \mathbf{y}\sim \rho_{*}(\cdot |\mathbf{x})}}\left[R_{\text{safe}}(\mathbf{x},\mathbf{y})\right] - \mathop{\mathbb{E}}_{\substack{\mathbf{x}\sim p_{\text{adv}}(\cdot |\mathbf{x}_{\text{input}})\\ \mathbf{y}\sim \rho_{*}(\cdot |\mathbf{x})}}\left[R_{\text{safe}}(\mathbf{x},\mathbf{y})\right], +$$ + +$$ +\Delta_{2}:= \mathop{\mathbb{E}}_{\substack{\mathbf{x}\sim p_{\text{adv}}(\cdot |\mathbf{x}_{\text{input}})\\ \mathbf{y}\sim \rho_{*}(\cdot |\mathbf{x})}}[R_{\text{safe}}(\mathbf{x},\mathbf{y})] - \mathop{\mathbb{E}}_{\substack{\mathbf{x}\sim p_{\text{adv}}(\cdot |\mathbf{x}_{\text{input}})\\ \mathbf{y}\sim \rho_{\text{safe -dec}}(\cdot |\mathbf{x})}}[R_{\text{safe}}(\mathbf{x},\mathbf{y})]. +$$ + +Next, we provide the interpretation of the mathematical result in Theorem 1 and intuitive insights about different terms as follows: + +(1) Practical Implications of the Bound: Theorem (1) shows that sub-optimality gap upper bound is minimized when $p_{\mathrm{adv}}$ remains close to $p_0$ (low KL divergence) and also when $\rho_{\mathrm{safe}}$ (which is our reference policy for the inference time alignment problem) aligns closely with $\rho_*$ (low policy KL divergence). This result can justify the use of KL-regularized objectives in safety-required domains, especially under adversarial prompt conditions. +(2) Role of Regularization Parameter $\beta$ : The adversarial prompt distribution in (2), $p_{\mathrm{adv}}$ , is defined as a solution to a KL-regularized optimization problem that minimizes expected reward $R_{\mathrm{safe}}$ relative to $p_0$ , with regularization parameter $\beta$ . The lower values of $\beta$ allow $p_{\mathrm{adv}}$ to diverge further from $p_0$ , potentially increasing $\Delta_1$ . +(3) Impact of $\alpha$ on $\Delta_{2}$ : The value of $\alpha$ affects the degree to which the policy $\rho_{\mathrm{safe - dec}}$ is incentivized to align with the safety policy $\rho_{*}$ . The lower values of $\alpha$ prioritize safety more heavily, thereby reducing $\Delta_{2}$ . +(3) In proof, $\Delta_{1}$ expresses how far an adversary can exploit the adversarial prompt distribution to attack the aligned model. In $\Delta_{2}$ , we show how, with inference time alignment, we can achieve safety even under the adversarial prompt distribution owing to the closed form of the KL regularized RLHF objective. We note that Mroueh [46, Proposition 3] also establishes upper bounds on the difference in expected rewards between two distributions in a continuous setting, relying on tail assumptions (e.g., sub-Gaussian, sub-Gamma) to control deviations. In contrast, our Theorem 1 operates + +entirely in a discrete setting and explicitly decomposes suboptimality (difference in expected rewards defined in (7)) into contributions from both the prompter and the language model. Furthermore, we leverage the optimality of RLHF to derive tighter upper bounds, which is not used in Mroueh [46]. + +Remark: Note that the theoretical guarantees discussed are conditional upon a specific prompt design structure, which originates from an adversarial prompter targeting a given unsafe query. This inherently prevents us from being able to obtain a similar guarantee for training time safety fine-tuning since the adversarial prompts would not be known in advance. In our experiments, we will not make these assumptions, as we experiment with several recent static safety benchmarks. + +# 6. Experiments + +# 6.1. Experimental Details + +Jail-break Datasets. We conduct a comprehensive evaluation by considering both text-based and image-based jailbreak attacks: + +1. Text-based Attacks. For text-based attacks, we use the JailbreakV-28K dataset [40]. This dataset integrates effective text-based jailbreak prompts for LLMs with various types of image data. Specifically, the text prompts include template-based, persuasive, and logic-driven approaches, while the image input may consist of noise, be blank, contain natural images, or be generated by stable diffusion. We evaluate all possible combinations of image and text pairs to ensure thorough evaluation. + +2. Image-based Attacks. For image-based attacks, we evaluate the following datasets: + +- MM-SafetyBench [37]: This benchmark assesses MLLM safety across 13 commonly prohibited scenarios and behaviors specified in usage guidelines from OpenAI [1] and Meta [26]. Each image in this dataset is associated with three types of image inputs: (1) Stable-diffusion images (SD), which are generated using stable diffusion [58] and directly relate to the query; (2) Typography (TYPO) images, containing optical character recognition representations of text from the malicious query; and (3) SD+TYPO images, which combine stable diffusion visuals with typographic subtitles. +- FigStep [16]: This attack converts harmful instructions into images using typography. Specifically, malicious queries on 10 forbidden topics from MM-SafetyBench are first paraphrased to start with nouns like "Steps to", "List of", or "Methods to", and then the instructions are presented as text within images. +- Visual Adversarial Attacks [54]: Qi et al. proposed an optimization-based attack, where the malicious user generates adversarial images by introducing carefully crafted + +perturbations to the original image, causing models to produce harmful content. + +Models Used. We perform evaluation on 5 state-of-art open-source MLLMs: LLaVA-1.5-7B [34], LLaVA-1.6-7B [35], MiniGPT-4-7B/13B [86], and Qwen-VL-Chat-7B [4]. For both LLaVA and MiniGPT, we use Vicuna [11] as the text encoder. For IMMUNE, we set the number of tokens sampled $k = 10$ and the alignment parameter $\alpha = 1$ . We report ablations in Appendix F. For main evaluations, we use LLaMa-3.1-8B $^2$ as the safety model $R_{\text{safe}}$ . The safety model has been trained on a mixture of preference datasets including HelpSteer2 [68], OffsetBias [51], WildGuard [20], and Magpie [74]. + +Baseline Defenses. We compare IMMUNE with recent state-of-the-art inference-time jail-break defense frameworks, including FigStep [16], Adashield [67], and CoCA [14]. For a fair comparison, all baseline defense methods are evaluated on a unified test dataset using consistent metrics. Detailed descriptions of baselines are provided in the Appendix G. + +Evaluation Metric. To assess the effectiveness of a jailbreak attack, we follow the literature [14, 40, 67] and measure the Attack Success Rate (ASR). Specifically, given a test dataset $\mathcal{D}_{\mathrm{unsafe}}$ of crafted jailbreak image-text pairs, the ASR quantifies the ratio of harmful responses to the total number of input queries and is defined as: + +$$ +\operatorname {A S R} = \frac {1}{\left| \mathcal {D} _ {\text {u n s a f e}} \right|} \sum_ {(\mathbf {I}, \mathbf {x}) \in \mathcal {D} _ {\text {u n s a f e}}} \mathbb {I} \left[ \mathcal {C} ^ {*} (\mathbf {x}, \pi_ {\theta} (\mathbf {I}, \mathbf {x})) = \text {T r u e} \right], \tag {8} +$$ + +where $(\mathbf{I},\mathbf{x})$ represents an image-text jailbreak prompt pair, $\mathbb{I}(\cdot)$ is an indicator function, and $\mathcal{C}^*$ is an oracle jailbreak classifier that verifies whether the generated response $\pi_{\theta}(\mathbf{I},\mathbf{x})$ aligns with the malicious intent of the user query $\mathbf{x}$ . Consistent with earlier works [9, 14, 40], we use both Llama-Guard-3 Jailbreak Judge and the GPT-4 as the oracle jailbreak classifier. Hence, a lower ASR value indicates a more robust defense against jailbreak attacks. To account for randomness in output generation, we sample three independent responses for each query and consider the model successfully jailbroken if any one of the three responses is flagged as jailbroken by the oracle classifier. + +# 6.2. Safety Evaluation results + +Evaluation on Text-based attacks. In Table 1, we present the results of this evaluation for different combinations of image types and text prompts. To be specific, the malicious user queries can be template-based, persuasive in nature, or logic-driven, while the image input consists of random noise, be blank, contain natural images, or be generated by stable diffusion. For this evaluation, we used the Llama-Guard-3 + +
ModelDefense StrategyNoiseSDNatureBlankAverage
TemplatePersuadeLogicTemplatePersuadeLogicTemplatePersuadeLogicTemplatePersuadeLogic
LLaVA-1.6Original66.1237.4578.5867.3437.5677.5769.2340.7882.6166.6739.4581.6060.27
FigStep [16]61.1239.2762.1662.3440.1854.0563.4135.2256.7661.0938.4352.7051.64
AdaShield [67]38.420.0011.0838.131.5618.1339.293.2219.1442.780.4816.1219.23
CoCA [14]61.2339.1762.1661.3441.2852.7063.1135.2255.4161.0937.3652.7051.37
IMMUNE (Ours)5.230.000.008.140.000.008.450.000.004.670.000.002.45
LLaVA-1.5Original58.1239.4476.5661.4738.3574.5558.2242.1177.5759.3340.2674.5556.39
FigStep [16]64.1728.3462.1662.2332.1868.9158.3937.2772.9761.0931.4268.9152.46
AdaShield [67]32.140.005.4131.360.004.2330.220.005.4129.670.0010.8112.86
CoCA [14]59.1826.3435.1358.4230.2935.1349.2232.1128.3759.0723.3937.8339.87
IMMUNE (Ours)9.230.000.008.140.000.001.470.000.005.320.000.002.10
MiniGPT-4-7BOriginal75.2354.1283.6272.4437.3385.6373.1845.2786.6472.0955.4286.6467.15
FigStep [16]47.2217.1832.4342.135.3724.3237.298.4518.9243.0817.4143.2427.74
AdaShield [67]47.2322.1432.2441.3718.4627.2058.1512.2830.2246.0919.4332.2432.21
CoCA [14]35.1818.2722.9740.3421.1431.0835.2918.4227.0348.1121.3640.5429.74
IMMUNE (Ours)18.236.1244.5911.348.2729.7217.188.4527.0216.0910.3743.2418.34
MiniGPT-4-13BOriginal74.1847.1290.6780.2353.3485.6379.2952.4583.7879.0753.2287.6570.62
FigStep [16]44.1320.2858.1143.2218.4747.3040.3627.1448.6542.0919.4258.1137.41
AdaShield [67]63.2223.1859.4464.4731.1274.5569.1527.3568.9165.0923.4153.3950.55
CoCA [14]35.2921.3436.4940.1822.2740.5435.1220.4135.1448.2221.3652.7033.21
IMMUNE (Ours)25.1718.2959.4429.3419.4157.4229.2223.1858.4323.4723.3758.4332.94
Qwen-VLOriginal46.123.2712.0952.236.189.0753.343.427.0553.114.3615.1122.99
FigStep [16]32.180.001.3540.420.005.4130.230.005.4144.122.231.3514.42
AdaShield [67]30.170.0018.1328.341.2915.1123.281.377.0542.110.4118.1315.88
CoCA [14]29.126.2713.5127.347.185.4127.291.4213.5128.4713.3613.5115.69
IMMUNE (Ours)10.272.185.4121.342.294.0318.222.355.4120.173.377.058.58
+ +Table 1. Evaluation on Text-Based Jailbreak Attacks. We report the Attack Success Rate (ASR) for various baseline defense strategies across recent MLLMs on text-based jailbreak attacks [40]. In this setup, the image input may be noise, Stable Diffusion-generated (SD), natural images (Nature), or blank, while the text query is structured as either template-based, persuasive, or logic-driven prompts. Lower ASR values indicate stronger resilience against jailbreak attacks. For this evaluation, we used Llama-Guard-3 Jailbreak Judge as the jailbreak classifier. IMMUNE consistently outperforms all compared defense frameworks across all MLLMs. The best results (lowest ASR) are highlighted in bold. All values are reported in %. + +
ModelDefense StrategyIllegal ActivityMalware GenerationPornographyHate SpeechPhysical HarmFraudAverage
SDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPE
LLaVA-1.6Original34.0280.4178.3520.4570.4570.457.0742.0535.0320.0651.0252.0518.0163.0163.0530.0367.0472.0747.94
FigStep [16]27.8449.4879.3825.0347.7359.097.0928.0741.0330.0436.0746.0720.0221.0852.0329.0536.0855.0737.64
AdaShield [67]19.5962.8941.249.0954.5543.187.0525.0519.066.0620.0313.039.0926.0427.046.0136.0235.0824.40
CoCA [14]24.2755.6776.0117.2759.0970.459.0233.0743.0825.0142.0752.0623.0626.0457.0125.0137.0858.0739.79
IMMUNE (Ours)0.001.035.150.000.004.559.233.0718.020.000.003.024.095.035.040.002.077.073.94
LLaVA-1.5Original34.0280.4180.4143.1875.0170.4513.0256.0739.0724.0575.0269.0725.0764.0367.0532.0373.0765.0353.85
FigStep [16]25.5263.9273.2820.4375.0544.0916.0426.0538.0638.0351.0655.0121.0457.0459.0636.0862.0861.0445.54
AdaShield [67]9.281.033.090.002.270.009.067.034.047.071.042.0311.037.053.0310.011.072.084.81
CoCA [14]17.1380.4180.4125.9127.779.918.0752.0655.0610.0863.1967.0325.0745.0557.0726.0357.0859.0744.65
IMMUNE (Ours)0.461.030.000.400.0013.646.989.0919.050.523.025.045.174.037.020.000.006.043.51
MiniGPT-4-7BOriginal13.4343.3435.052.2734.0925.0827.0414.0121.066.0229.0517.053.0834.0826.025.0540.0436.0223.04
FigStep [16]19.5943.3438.146.8240.9115.9126.0119.0623.0611.0428.0816.0314.0631.0826.0712.0644.0517.0724.21
AdaShield [67]11.3435.0518.5618.1822.739.0916.6726.0315.0422.678.0426.048.0827.0914.0918.1834.0422.0720.05
CoCA [14]9.2842.2728.876.8220.4518.1819.0812.0924.076.0512.0710.0325.0416.6726.0725.068.3325.0218.92
IMMUNE (Ours)13.4622.6813.4311.3620.4518.1817.0412.0921.033.088.047.087.0214.0120.020.000.000.0011.03
MiniGPT-4-13BOriginal17.5340.2142.2713.6422.7327.2710.0415.0818.039.0522.0921.0717.0320.0235.0819.0537.0341.0923.97
FigStep [16]25.7746.3936.0822.7327.2729.5513.0216.0819.0320.0830.0720.0818.0136.0438.0821.0532.0938.0327.17
AdaShield [67]21.6523.7122.6820.4534.0925.0415.059.0215.0311.0625.0321.0114.0322.0123.0723.0528.0215.0919.84
CoCA [14]25.0350.0746.0314.0625.0343.0613.0225.018.3316.0416.6716.6716.6725.0641.6725.0641.6725.0726.13
IMMUNE (Ours)23.7822.6819.599.0918.1829.5520.0814.0815.0611.0813.0813.0210.0818.0322.086.020.000.0014.24
+ +Table 2. Evaluation on MM-Safety Bench. We report Attack Success Rate (ASR) for 6 categories of prohibited scenarios from MM-Safety Bench [38]. For this evaluation, we used GPT-4 as the jailbreak classifier. IMMUNE significantly reduces ASR across all categories. The best results (lowest ASR) are highlighted in bold. All values are reported in %. + +Jailbreak Judge as the oracle classifier. From our evaluation, we draw the following key insights: (1) With the original decoding, ASR is notably high for most models, reaching up to $70.62\%$ for MiniGPT-4-13B, underscoring both the severity of the attack and the underlying serious safety concern. (2) Across all MLLMs, IMMUNE demonstrates superior performance, substantially reducing ASR compared to other competitive defense strategies. Specifically, on MiniGPT-4-13B, IMMUNE reduces ASR by $37.68\%$ compared to the + +original decoding and by $17.61\%$ compared to state-of-art defense strategy AdaShield [67]. (3) We observe that logic-based attacks reveal a model-dependent variation in IMMUNE's performance. Specifically, for logic and persuade-based attacks, IMMUNE achieves an ASR of $0\%$ across all input types on LLaVA-1.6 and LLaVA-1.5. However, on MiniGPT-4-7B, logic-based attacks with noise inputs result in a significantly higher ASR of $44.59\%$ , compared to CoCA [14] $(22.97\%)$ and AdaShield [67] $(32.24\%)$ . + +
ModelDefense StrategyIllegal ActivityHate SpeechMalware GenerationPhysical HarmFraudAdult ContentPrivacy ViolationLegal OpinionFinancial AdviceHealth ConsultationAverage
LLaVA-1.6Original11.700.0021.9211.880.0026.270.0036.0431.809.5314.68
FigStep [16]9.620.0014.232.030.0011.950.006.0513.586.245.90
AdaShield [67]3.580.005.650.000.0011.790.000.009.610.003.62
CoCA [14]13.624.296.065.978.3320.396.5727.7221.7017.5613.34
IMMUNE (Ours)12.040.000.000.000.000.000.000.000.000.001.21
LLaVA-1.5Original74.000.0083.9761.6210.4693.961.7292.3782.2984.0958.58
FigStep [16]65.531.9570.0748.473.7773.684.1371.7276.2967.5148.48
AdaShield [67]14.380.0011.5413.880.002.100.002.2012.2114.167.24
CoCA [14]44.038.2338.0422.086.4942.489.6544.1541.9430.4528.63
Ours28.210.006.302.120.000.000.004.403.841.864.23
MiniGPT-4-7BOriginal10.152.1416.022.440.0019.770.0033.6226.2111.9311.95
FigStep [16]11.630.0024.450.003.7513.780.0010.0615.963.998.32
AdaShield [67]4.290.0012.175.674.072.120.0012.128.239.645.76
CoCA [14]7.850.000.002.471.856.360.0015.6313.876.285.74
IMMUNE (Ours)7.980.009.826.140.004.430.007.855.713.594.43
MiniGPT-4-13BOriginal20.050.0030.4115.980.0042.140.0037.8146.2819.5321.17
FigStep [16]24.110.0038.4521.880.0029.570.0021.8334.5023.8119.52
AdaShield [67]18.060.007.663.930.0011.551.8228.2029.9212.2611.65
CoCA [14]24.38.3313.5718.125.5431.928.3313.9817.8318.3815.93
IMMUNE (Ours)16.312.248.275.540.0010.180.0016.0619.5315.989.48
Qwen-VLOriginal17.530.0040.2317.770.0032.380.0056.3439.7539.8224.01
FigStep [16]1.660.002.200.000.0020.870.0010.1210.080.004.56
AdaShield [67]4.050.006.499.680.0017.830.006.061.550.005.04
CoCA [14]29.530.0067.8834.490.0077.920.0088.2683.5661.9243.92
IMMUNE (Ours)6.500.004.370.000.009.930.007.935.880.003.23
+ +Evaluation on Image-based attacks. We evaluate IMMUNE on three standard benchmarks for image-based attacks: + +(1) MM-Safety. We report the ASR for 6 prohibited categories from MM-Safety Bench [38] across four recent MLLMs, as shown in Table 2. For response evaluation, following Liu et al. [38], we used GPT-4 as the jailbreak classifier. Our findings indicate that IMMUNE significantly reduces ASR across all MLLMs compared to state-of-art defense strategies such as CoCA [14] and AdaShield [67]. Notably, on LLaVA-1.6, IMMUNE achieves an ASR reduction of $44.0\%$ and $20.46\%$ compared to the base model and AdaShield [67] respectively. Across all MLLMs, we observe that attacks using typography images (TYPO) and those combining stable diffusion visuals with typography subtitles (SD-TYPO) are significantly more effective than simple stable diffusion-generated (SD) images. Additionally, samples from the "Pornography" category exhibit the highest ASR across all defense mechanisms. + +(2) FigStep. Table 3 presents the ASR results for various MLLMs on the FigStep dataset [16]. Consistent with results on MM-Safety, IMMUNE demonstrates superior performance over competitive baselines, reducing ASR by $6.45\%$ and $12.13\%$ on MiniGPT-4-13B and LLaVA-1.6, respectively, compared to CoCA [14]. Among all categories, "Hate Speech" and "Privacy Violation" have the lowest ASR, with most approaches achieving an ASR of near $0.0\%$ . + +(3) Visual Adversarial Attacks. Finally, in Table 4, we evaluate ASR under adversarily optimized jailbreak attacks [54] across varying attack strengths, denoted by $\epsilon$ , where higher $\epsilon$ values correspond to stronger attacks. The "Unconstrained" setting represents the most challenging attack scenario. Notably, even under the "unconstrained" attack, IMMUNE reduces ASR by $24.19\%$ compared to baseline [67] on LLaVA-1.6. + +Additional discussion on inference time of IMMUNE and its impact on model utility is deferred to Appendix F. + +Table 3. Evaluation on FigStep. We report the Attack Success Rate (ASR) for all categories from the FigStep benchmark [16]. For this evaluation, we used GPT-4 as the jailbreak classifier. The best results (lowest ASR) are highlighted in bold. All values are reported in %. + +
ModelDefense StrategyAttack Strength
ε = 16/255ε = 32/255ε = 64/255Unconstrained
LLaVA-1.6Original66.1264.0862.3164.80
FigStep [16]49.1050.3047.2251.13
AdaShield [67]35.2139.7134.0540.21
CoCA [14]45.2146.1148.0250.33
IMMUNE (Ours)13.3214.9715.1816.02
LLaVA-1.5Original68.0067.1966.2367.33
FigStep [16]53.9852.0751.658.17
AdaShield [67]27.0024.1526.9130.10
CoCA [14]44.2341.8842.9144.82
IMMUNE (Ours)12.4516.7114.3217.03
MiniGPT-4-7BOriginal53.4058.1258.7961.23
FigStep [16]47.0952.4355.2557.10
AdaShield [67]39.0042.4543.7149.63
CoCA [14]53.1960.0962.7868.29
IMMUNE (Ours)18.4220.7721.1822.95
MiniGPT-4-13BOriginal62.5766.1367.8270.11
FigStep [16]59.8161.0562.4465.79
AdaShield [67]47.8250.3751.6853.15
CoCA [14]71.8574.3175.7275.09
IMMUNE (Ours)23.4125.8226.0226.02
Qwen-VLOriginal38.1148.0348.0349.47
FigStep [16]34.2536.9240.1742.78
AdaShield [67]17.3218.2022.4125.24
CoCA [14]38.1544.7346.3649.04
IMMUNE (Ours)13.2215.0715.0716.98
+ +Table 4. Evaluation on Visual Adversarial Attacks [54]. We report ASR (in %) on images optimized with varying levels of adversarial noise, denoted by $\epsilon$ , with "unconstrained" representing the most challenging scenario. + +# 7. Conclusion + +In this work, we introduce IMMUNE, an inference-time defense framework designed to protect MLLMs against jailbreak attacks. IMMUNE employs a safety-aware reward model to align the MLLM during inference, with the goal of mitigating jailbreak vulnerabilities. We present a mathematical formulation of our defense framework, framing jailbreak mitigation as an alignment problem. Through extensive experiments, we demonstrate that IMMUNE consistently and significantly outperforms competitive baselines, enhancing model safety while preserving the model utility. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 6 +[2] Maksym Andriushchenko, Francesco Croce, and Nicolas Flammarion. Jailbreaking leading safety-aligned llms with simple adaptive attacks. arXiv preprint arXiv:2404.02151, 2024. 2 +[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 1 +[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 6 +[5] James Beetham, Souradip Chakraborty, Mengdi Wang, Furong Huang, Amrit Singh Bedi, and Mubarak Shah. Liar: Leveraging alignment (best-of-n) to jailbreak llms in seconds. arXiv preprint arXiv:2412.05232, 2024. 5 +[6] Nicholas Carlini, Anish Athalye, Nicolas Papernot, Wieland Brendel, Jonas Rauber, Dimitris Tsipras, Ian Goodfellow, Aleksander Madry, and Alexey Kurakin. On evaluating adversarial robustness. arXiv preprint arXiv:1902.06705, 2019. 4 +[7] Souradip Chakraborty, Soumya Suvra Ghosal, Ming Yin, Dinesh Manocha, Mengdi Wang, Amrit Singh Bedi, and Furong Huang. Transfer q star: Principled decoding for llm alignment, 2024. 2, 3, 4, 16 +[8] Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J Pappas, and Eric Wong. Jailbreaking black box large language models in twenty queries. arXiv preprint arXiv:2310.08419, 2023. 3 +[9] Patrick Chao, Edoardo Debenedetti, Alexander Robey, Maksym Andriushchenko, Francesco Croce, Vikash Sehwag, Edgar Dobriban, Nicolas Flammarion, George J. Pappas, Florian Tramer, Hamed Hassani, and Eric Wong. Jailbreakbench: An open robustness benchmark for jailbreaking large language models. In NeurIPS Datasets and Benchmarks Track, 2024. 6 +[10] Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14239-14250, 2024. 3 +[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ *chatgpt quality, 2023. 6 +[12] Yinpeng Dong, Huanran Chen, Jiawei Chen, Zhengwei Fang, Xiao Yang, Yichi Zhang, Yu Tian, Hang Su, and Jun Zhu. How robust is google's bard to adversarial image attacks? arXiv preprint arXiv:2309.11751, 2023. 3 + +[13] Yanrui Du, Sendong Zhao, Ming Ma, Yuhan Chen, and Bing Qin. Analyzing the inherent response tendency of llms: Real-world instructions-driven jailbreak. arXiv preprint arXiv:2312.04127, 2023. 3 +[14] Jiahui Gao, Renjie Pi, Tianyang Han, Han Wu, Lanqing Hong, Lingpeng Kong, Xin Jiang, and Zhenguo Li. Coca: Regaining safety-awareness of multimodal large language models with constitutional calibration. arXiv preprint arXiv:2409.11365, 2024. 1, 2, 3, 6, 7, 8, 15, 16, 17 +[15] Simon Geisler, Tom Wollschläger, M. H. I. Abdalla, Johannes Gasteiger, and Stephan Gunnemann. Attacking large language models with projected gradient descent, 2024. 2, 4 +[16] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via topographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 2, 3, 6, 7, 8, 15, 17, 18, 19 +[17] Jindong Gu, Zhen Han, Shuo Chen, Ahmad Beirami, Bailan He, Gengyuan Zhang, Ruotong Liao, Yao Qin, Volker Tresp, and Philip Torr. A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980, 2023. 3 +[18] Xingang Guo, Fangxu Yu, Huan Zhang, Lianhui Qin, and Bin Hu. Cold-attack: Jailbreaking Ilms with stealthiness and controllability. arXiv preprint arXiv:2402.08679, 2024. 3 +[19] Dongchen Han, Xiaojun Jia, Yang Bai, Jindong Gu, Yang Liu, and Xiaochun Cao. Ot-attack: Enhancing adversarial transferability of vision-language models via optimal transport optimization. arXiv preprint arXiv:2312.04403, 2023. 1, 3 +[20] Seungju Han, Kavel Rao, Allyson Ettinger, Liwei Jiang, Bill Yuchen Lin, Nathan Lambert, Yejin Choi, and Nouha Dziri. Wildguard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of llms. arXiv preprint arXiv:2406.18495, 2024. 6 +[21] Jonathan Hayase, Ema Borevkovic, Nicholas Carlini, Florian Tramér, and Milad Nasr. Query-based adversarial prompt generation. arXiv preprint arXiv:2402.12329, 2024. 2 +[22] Zhang-Wei Hong, Idan Shenfeld, Tsun-Hsuan Wang, Yung-Sung Chuang, Aldo Pareja, James Glass, Akash Srivastava, and Pulkit Agrawal. Curiosity-driven red-teaming for large language models. arXiv preprint arXiv:2402.19464, 2024. 4 +[23] MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR), 51 (6):1-36, 2019. 1 +[24] Sandy Huang, Nicolas Papernot, Ian Goodfellow, Yan Duan, and Pieter Abbeel. Adversarial attacks on neural network policies. arXiv preprint arXiv:1702.02284, 2017. 4 +[25] Yangsibo Huang, Samyak Gupta, Mengzhou Xia, Kai Li, and Danqi Chen. Catastrophic jailbreak of open-source llms via exploiting generation. arXiv preprint arXiv:2310.06987, 2023. 3 +[26] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Llm-based input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023. 6 + +[27] Erik Jones, Anca Dragan, Aditi Raghunathan, and Jacob Steinhardt. Automatically auditing large language models via discrete optimization. In International Conference on Machine Learning, pages 15307-15329. PMLR, 2023. 2 +[28] Maxim Khanov, Jirayu Burapacheep, and Yixuan Li. ARGs: Alignment as reward-guided search. In The Twelfth International Conference on Learning Representations, 2024. 3 +[29] Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, et al. Rewardbench: Evaluating reward models for language modeling. arXiv preprint arXiv:2403.13787, 2024. 4 +[30] Seanie Lee, Minsu Kim, Lynn Cherif, David Dobre, Juho Lee, Sung Ju Hwang, Kenji Kawaguchi, Gauthier Gidel, Yoshua Bengio, Nikolay Malkin, and Moksh Jain. Learning diverse attacks on large language models for robust red-teaming and safety tuning, 2024. 4 +[31] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665, 2023. 15, 16 +[32] Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, and Qi Liu. Red teaming visual language models. arXiv preprint arXiv:2401.12915, 2024. 1 +[33] Yuhui Li, Fangyun Wei, Jinjing Zhao, Chao Zhang, and Hongyang Zhang. Rain: Your language models can align themselves without finetuning. arXiv preprint arXiv:2309.07124, 2023. 3 +[34] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 6, 17 +[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.2,6 +[36] Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, et al. Agentbench: Evaluating llms as agents. arXiv preprint arXiv:2308.03688, 2023. 1 +[37] Xin Liu, Yichen Zhu, Yunshi Lan, Chao Yang, and Yu Qiao. Query-relevant images jailbreak large multi-modal models, 2023. 3, 6 +[38] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models, 2024. 2, 7, 8, 15 +[39] Yi Liu, Gelei Deng, Zhengzi Xu, Yuekang Li, Yaowen Zheng, Ying Zhang, Lida Zhao, Tianwei Zhang, Kailong Wang, and Yang Liu. Jailbreaking chatgpt via prompt engineering: An empirical study. arXiv preprint arXiv:2305.13860, 2023. 3 +[40] Weidi Luo, Siyuan Ma, Xiaogeng Liu, Xiaoyu Guo, and Chaowei Xiao. Jailbreakv-28k: A benchmark for assessing the robustness of multimodal large language models against jailbreak attacks. arXiv preprint arXiv:2404.03027, 2024. 2, 3, 6, 7, 15, 16, 18, 20 +[41] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng + +Tu. Macaw-lm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093, 2023. 3 +[42] Neal Mangaokar, Ashish Hooda, Jihye Choi, Shreyas Chandrashekaran, Kassem Fawaz, Somesh Jha, and Atul Prakash. Prp: Propagating universal perturbations to attack large language model guard-rails. arXiv preprint arXiv:2402.15911, 2024. 2 +[43] Ninareh Mehrabi, Palash Goyal, Anil Ramakrishna, Jwala Dhamala, Shalini Ghosh, Richard Zemel, Kai-Wei Chang, Aram Galstyan, and Rahul Gupta. Jab: Joint adversarial prompting and belief augmentation. In R0-FoMo: Robustness of Few-shot and Zero-shot Learning in Large Foundation Models. 3 +[44] Ninareh Mehrabi, Ahmad Beirami, Fred Morstatter, and Aram Galstyan. Robust conversational agents against imperceptible toxicity triggers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2831-2847, 2022. 4 +[45] Ninareh Mehrabi, Palash Goyal, Christophe Dupuy, Qian Hu, Shalini Ghosh, Richard Zemel, Kai-Wei Chang, Aram Galstyan, and Rahul Gupta. Flirt: Feedback loop in-context red teaming. Empirical Methods in Natural Language Processing (EMNLP), 2024. 3 +[46] Youssef Mroueh. Information theoretic guarantees for policy alignment in large language models. arXiv preprint arXiv:2406.05883, 2024. 5, 6 +[47] Sidharth Mudgal, Jong Lee, Harish Ganapathy, YaGuang Li, Tao Wang, Yanping Huang, Zhifeng Chen, Heng-Tze Cheng, Michael Collins, Trevor Strohman, Jilin Chen, Alex Beutel, and Ahmad Beirami. Controlled decoding from language models. International Conference on Machine Learning (ICML), 2024. 2, 3, 4, 16 +[48] Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435, 2023. 1 +[49] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 3 +[50] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2, 4 +[51] Junsoo Park, Seungyeon Jwa, Meiying Ren, Daeyoung Kim, and Sanghyuk Choi. Offsetbias: Leveraging debiased data for tuning evaluators. arXiv preprint arXiv:2407.06551, 2024. 6 +[52] Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia Glaese, Nat McAleese, and Geoffrey Irving. Red teaming language models with language models. arXiv preprint arXiv:2202.03286, 2022. 4 +[53] Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang. + +Mllm-protector: Ensuring mllm's safety without hurting performance. arXiv preprint arXiv:2401.02906, 2024. 3 +[54] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 1, 2, 3, 6, 8 +[55] Xiangyu Qi, Ashwinee Panda, Kaifeng Lyu, Xiao Ma, Subhrajit Roy, Ahmad Beirami, Prateek Mittal, and Peter Henderson. Safety alignment should be made more than just a few tokens deep, 2024. 3 +[56] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to! International Conference on Learning Representations (2024), 2024. 3 +[57] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 3, 4, 15, 16 +[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 6 +[59] Christian Schlarmann and Matthias Hein. On the adversarial robustness of multi-modal foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3677-3685, 2023. 1, 3 +[60] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Plug and pray: Exploiting off-the-shelf components of multi-modal models. arXiv preprint arXiv:2307.14539, 2023. 1, 3 +[61] Chawin Sitawarin, Norman Mu, David Wagner, and Alexandre Araujo. Pal: Proxy-guided black-box attack on large language models. arXiv preprint arXiv:2402.09674, 2024. 2 +[62] Bharath K. Sriperumbudur, Kenji Fukumizu, Arthur Gretton, Bernhard Schölkopf, and Gert R. G. Lanckriet. On integral probability metrics, phi-divergences and binary classification, 2009. 13 +[63] Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul Christiano. Learning to summarize from human feedback, 2022. 1 +[64] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing nlp. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2153-2162, 2019. 4 +[65] Boxin Wang, Weixin Chen, Hengzhi Pei, Chulin Xie, Mintong Kang, Chenhui Zhang, Chejian Xu, Zidi Xiong, Ritik Dutta, Ryan Schaeffer, et al. Decodingtrust: A comprehensive assessment of trustworthiness in gpt models. In NeurlPS, 2023. 1 +[66] Hao Wang, Hao Li, Minlie Huang, and Lei Sha. From noise to clarity: Unraveling the adversarial suffix of large language + +model attacks via translation of text embeddings. arXiv preprint arXiv:2402.16006, 2024. 2 +[67] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. arXiv preprint arXiv:2403.09513, 2024. 1, 2, 3, 6, 7, 8, 15, 16, 17 +[68] Zhilin Wang, Yi Dong, Olivier Delalleau, Jiaqi Zeng, Gerald Shen, Daniel Egert, Jimmy J Zhang, Makesh Narsimhan Sreedhar, and Oleksii Kuchaiev. Helpsteer2: Open-source dataset for training top-performing reward models. arXiv preprint arXiv:2406.08673, 2024. 6 +[69] Fenghua Weng, Yue Xu, Chengyan Fu, and Wenjie Wang. MMJ-Bench: A comprehensive study on jailbreak attacks and defenses for vision language models. arXiv preprint arXiv:2408.08464, 2024. 1 +[70] Nevan Wichers, Carson Denison, and Ahmad Beirami. Gradient-based language model red teaming. 18th Conference of European Chapter of the Assoc for Computational Linguistics (EACL), 2024. 4 +[71] Qi Wu, Damien Teney, Peng Wang, Chunhua Shen, Anthony Dick, and Anton Van Den Hengel. Visual question answering: A survey of methods and datasets. Computer Vision and Image Understanding, 163:21-40, 2017. 1 +[72] Nan Xu, Fei Wang, Ben Zhou, Bang Zheng Li, Chaowei Xiao, and Muhao Chen. Cognitive overload: Jailbreaking large language models with overloaded logical thinking. arXiv preprint arXiv:2311.09827, 2023. 3 +[73] Yue Xu, Xiuyuan Qi, Zhan Qin, and Wenjie Wang. Defending jailbreak attack in vlms via cross-modality information detector. arXiv preprint arXiv:2407.21659, 2024. 3 +[74] Zhangchen Xu, Fengqing Jiang, Luyao Niu, Yuntian Deng, Radha Poovendran, Yejin Choi, and Bill Yuchen Lin. Magpie: Alignment data synthesis from scratch by prompting aligned lms with nothing, 2024. 6 +[75] Zonghao Ying, Aishan Liu, Tianyuan Zhang, Zhengmin Yu, Siyuan Liang, Xianglong Liu, and Dacheng Tao. Jailbreak vision language models via bi-modal adversarial prompt. arXiv preprint arXiv:2406.04031, 2024. 3 +[76] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13807-13816, 2024. 3 +[77] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 2, 16, 17 +[78] Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang. Rrrh: Rank responses to align language models with human feedback without tears, 2023. 1 +[79] Yi Zeng, Hongpeng Lin, Jingwen Zhang, Diyi Yang, Ruoxi Jia, and Weiyan Shi. How johnny can persuade llms to jailbreak them: Rethinking persuasion to challenge ai safety by humanizing llms. arXiv preprint arXiv:2401.06373, 2024. 3 + +[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3 +[81] Xiaoyu Zhang, Cen Zhang, Tianlin Li, Yihao Huang, Xiaojun Jia, Xiaofei Xie, Yang Liu, and Chao Shen. A mutation-based method for multi-modal jailbreaking attack detection. arXiv preprint arXiv:2312.10766, 2023. 3 +[82] Zhuo Zhang, Guangyu Shen, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Make them spill the beans! coercive knowledge extraction from (production) llms. arXiv preprint arXiv:2312.04782, 2023. 2 +[83] Xuandong Zhao, Xianjun Yang, Tianyu Pang, Chao Du, Lei Li, Yu-Xiang Wang, and William Yang Wang. Weak-to-strong jailbreaking on large language models. arXiv preprint arXiv:2401.17256, 2024. 3 +[84] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[85] Yukai Zhou, Zhijie Huang, Feiyang Lu, Zhan Qin, and Wenjie Wang. Don't say no: Jailbreaking llm by suppressing refusal. arXiv preprint arXiv:2404.16369, 2024. 3 +[86] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 2, 3, 6 +[87] Sicheng Zhu, Ruiyi Zhang, Bang An, Gang Wu, Joe Barrow, Zichao Wang, Furong Huang, Ani Nenkova, and Tong Sun. Autodan: interpretable gradient-based adversarial attacks on large language models. In First Conference on Language Modeling, 2024. 2, 4 +[88] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 1, 4 +[89] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. arXiv preprint arXiv:2402.02207, 2024. 3 +[90] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 1, 2, 3, 4 +[91] Andy Zou, Long Phan, Justin Wang, Derek Duenas, Maxwell Lin, Maksym Andriushchenko, Rowan Wang, Zico Kolter, Matt Fredrikson, and Dan Hendrycks. Improving alignment and robustness with short circuiting. arXiv preprint arXiv:2406.04313, 2024. 3 \ No newline at end of file diff --git a/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/images.zip b/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b0261d92d358877b44cf7f3bde1407887a18d723 --- /dev/null +++ b/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34889958c6f243aab62fd7bf32aaa300dbcdca80492b0b98eb21cc440d1eca21 +size 664129 diff --git a/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/layout.json b/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..230e500d298891f99fd3cb254bc6d652fd1dc075 --- /dev/null +++ b/2025/Immune_ Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment/layout.json @@ -0,0 +1,10604 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 99, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 99, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 99, + 541, + 140 + ], + "type": "text", + "content": "IMMUNE: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "spans": [ + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "text", + "content": "Soumya Suvara Ghosal" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "text", + "content": " Souradip Chakraborty" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "text", + "content": " Vaibhav Singh" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "text", + "content": " Tianrui Guan" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "text", + "content": " Mengdi Wang" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "text", + "content": " Ahmad Beirami Furong Huang" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "text", + "content": " Alvaro Velasquez" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "text", + "content": " Dinesh Manocha" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "text", + "content": " Amrit Singh Bedi" + }, + { + "bbox": [ + 64, + 162, + 546, + 192 + ], + "type": "inline_equation", + "content": "^{5\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 203, + 541, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 203, + 541, + 232 + ], + "spans": [ + { + "bbox": [ + 70, + 203, + 541, + 232 + ], + "type": "text", + "content": "1University of Maryland; 2Indian Institute of Technology Bombay; 3Princeton University; 4University of Colorado Boulder; 5University of Central Florida" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 274, + 200, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 274, + 200, + 287 + ], + "spans": [ + { + "bbox": [ + 152, + 274, + 200, + 287 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 300, + 297, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 297, + 540 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 297, + 540 + ], + "type": "text", + "content": "With the widespread deployment of Multimodal Large Language Models (MLLMs) for visual-reasoning tasks, improving their safety has become crucial. Recent research indicates that despite training-time safety alignment, these models remain vulnerable to jailbreak attacks. In this work, we first highlight an important safety gap to describe that alignment achieved solely through safety training may be insufficient against jailbreak attacks. To address this vulnerability, we propose IMMUNE, an inference-time defense framework that leverages a safety reward model through controlled decoding to defend against jailbreak attacks. Additionally, we provide a mathematical characterization of IMMUNE, offering insights on why it improves safety against jailbreaks. Extensive evaluations on diverse jailbreak benchmarks using recent MLLMs reveal that IMMUNE effectively enhances model safety while preserving the model's original capabilities. For instance, against text-based jailbreak attacks on LLaVA-1.6, IMMUNE reduces the attack success rate by " + }, + { + "bbox": [ + 54, + 300, + 297, + 540 + ], + "type": "inline_equation", + "content": "57.82\\%" + }, + { + "bbox": [ + 54, + 300, + 297, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 300, + 297, + 540 + ], + "type": "inline_equation", + "content": "16.78\\%" + }, + { + "bbox": [ + 54, + 300, + 297, + 540 + ], + "type": "text", + "content": " compared to the base MLLM and state-of-the-art defense strategy, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 556, + 135, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 135, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 135, + 569 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 576, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 297, + 685 + ], + "type": "text", + "content": "Multi-modal large language models (MLLMs) have made remarkable progress in vision-language reasoning tasks such as visual question-answering [1, 3, 71] and image captioning [1, 23]. However, ensuring their outputs are safe and free from discrimination, disinformation, and harmful content is important for their adoption. Safety alignment through reinforcement learning from human feedback (RLHF) has shown promise in aligning model behavior with social norms to prevent harmful outputs [63, 78, 88]. Despite these sig" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 275, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 323 + ], + "type": "text", + "content": "nificant alignment efforts, recent studies reveal that MLLMs remain vulnerable to jailbreaking attacks, where malicious image-text pairs bypass safety mechanisms, raising safety concerns [19, 32, 36, 48, 54, 59, 60, 65]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 326, + 556, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 556, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 556, + 506 + ], + "type": "text", + "content": "Why training-time safety alignment might not be enough? Fine-tuning-based safety alignment approaches rely on a static prompt distribution, leaving them inherently vulnerable to jailbreak attacks. These attacks exploit this limitation by solving a prompt (text or image) optimization problem to craft adversarial prompts that bypass the safety mechanisms of aligned models [54, 90]. In this work, we argue that safety finetuning against jailbreak attacks while keeping model capabilities may be inherently challenging given that the adversary can find a prompt that bypasses the safety mechanisms of the model. This underscores the fundamental limitation of relying solely on training-time safety alignment procedures to ensure robustness. These leads us to ask: Is it possible to design a safeguarding strategy to defend against given possible adversarial prompts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 506, + 557, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 557, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 557, + 698 + ], + "type": "text", + "content": "In this paper, we take a step towards answering this question in the affirmative. We propose shifting from training-time alignment to inference-time safeguarding. Although initial efforts in this direction [14, 67] show promise, they exhibit weak empirical performance. Further, Weng et al. [69] note that adding a safety prompt to the input [67] can lead to overly cautious behavior, resulting in generic responses like, \"I am sorry...\" even for benign queries (see Figure 1), thereby affecting model utility. In real-world applications, an ideal jailbreaking defense should not only minimize the success rate of attacks but also maintain high utility by preserving model performance on legitimate queries. Moreover, existing inference-time approaches lack mathematical characterization, relying primarily on heuristic-based methods. This gap motivates the need for robust inference-time safeguarding mechanisms, which we address in this paper." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 701, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 556, + 714 + ], + "type": "text", + "content": "Our key idea and proposed approach. We present IMMUNE," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 702 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 702 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 122, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 122, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 122, + 712 + ], + "type": "text", + "content": "Equal advising." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25038" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 324, + 220 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 324, + 220 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 324, + 220 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 324, + 220 + ], + "type": "image", + "image_path": "46403663e70204d7a71a27780e55725a487836daee282465f4c76f052531486d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 354 + ], + "type": "text", + "content": "Figure 1. Qualitative Evaluation (Left). Given an image of a house generated by stable diffusion [58] and perturbed by adversarial noise [54], along with a malicious user query asking for steps to \"break into and rob a house\", we visualize responses from the base model and various inference-time defense strategies, including CoCA [14] and AdaShield [67]. We observe that all compared defense strategies are misled into generating harmful content. In contrast, our proposed inference-time safety alignment framework, IMMUNE, effectively rejects the user query, citing its unethical nature. This evaluation underscores the importance of inference-time alignment in preventing harmful responses. We visualize additional responses in the Appendix. Quantitative Evaluation (Right). To empirically validate the effectiveness of IMMUNE, we compare the attack success rates and model utility across various state-of-the-art defense strategies. A lower attack success rate reflects improved safety in generated outputs. Our results indicate that IMMUNE substantially lowers the attack success rate compared to other baselines. Additionally, IMMUNE not only strengthens model safety but also preserves the model's original utility, as demonstrated by its performance on MM-Vet benchmark [77]. We note that an ideal defense strategy should have a low attack success rate while maintaining high utility, i.e., towards the bottom right corner of the plot. Refer to Section 6 for further details." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 553, + 220 + ], + "type": "image", + "image_path": "8e2f14877fa4301de8388ba74ff800efafe3ebe60349a58f78a77dc74bb73277.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 374, + 295, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 295, + 541 + ], + "type": "text", + "content": "a framework that reframes jailbreak defense as an inference-time alignment problem. Our approach is built on two key components: First, we show that jailbreaks can be effectively mitigated by aligning the MLLM during inference using controlled decoding [7, 47], assuming access to a safety-aware utility or reward function (without loss of generality). Second, we mathematically formalize this defense as an inference-time alignment problem within the KL-regularized reinforcement learning framework [50], proving its robustness against adversarial prompt distributions. As illustrated in Figure 1, our comprehensive evaluations demonstrate that IMMUNE outperforms state-of-the-art inference-time defense strategies, such as AdaShield [67] and CoCA [14], while maintaining high model utility. Our key contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 546, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 605 + ], + "type": "text", + "content": "- Insufficiency of safety finetuning for MLLMs. We first argue that the current alignment achieved solely via safety training/fine-tuning with RLHF is insufficient to defend against jailbreak attacks, thereby highlighting the need for additional safeguards." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "- Improving safety via inference time alignment. To address these safety concerns, we propose IMMUNE, an inference-time framework for defense against jailbreak attacks through controlled decoding using a safety reward model. We provide a characterization of our defense mechanism by framing it as an alignment problem, offering mathematical guarantees against jailbreak attacks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "- Empirical evaluations. We empirically validate the efficacy of our proposed defense strategy, IMMUNE," + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 374, + 556, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 374, + 556, + 447 + ], + "spans": [ + { + "bbox": [ + 322, + 374, + 556, + 447 + ], + "type": "text", + "content": "across a range of jailbreak benchmarks, including MM-SafetyBench [38], FigStep [16], visual adversarial attacks [54], and text-based jailbreak attacks [40]. Our evaluations demonstrate significant improvements on several recent MLLMs, such as LLaVA-1.5 [34], LLaVA-1.6 [35], MiniGPT-4 [86], and Qwen-VL [4]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 464, + 405, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 464, + 405, + 475 + ], + "spans": [ + { + "bbox": [ + 314, + 464, + 405, + 475 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "content": "Jailbreak attacks. Jones et al. [27] frames jailbreaking large language models (LLMs) as a discrete optimization problem. They search for the optimal suffix that, when appended to the original prompt, can greedily generate the desired harmful output. Zou et al. [90] demonstrated that model inputs could be manipulated based on gradients to force-generate responses to harmful queries. Zhu et al. [87] introduced a method to generate an adversarial suffix iteratively. This approach ensures that the optimized suffix remains meaningful while bypassing perplexity-based filters. Other works that proposed jailbreaking by manipulating gradients include Andriushchenko et al. [2], Geisler et al. [15], Hayase et al. [21], Mangaokar et al. [42], Sitawarin et al. [61], Wang et al. [66]. Another line of work, exemplified by Zhang et al. [82], involves optimizing the prompt iteratively until the distribution of output tokens results in generating harmful responses. For example, Zhang et al. [82] showed that adversaries can break safety alignment by forcing the target LLM to select a lower-ranked output token, thereby generating toxic con" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25039" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 360 + ], + "type": "text", + "content": "tent. Additional works following this domain include Du et al. [13], Guo et al. [18], Huang et al. [25], Zhao et al. [83], Zhou et al. [85]. Recent work has also focused on jailbreaking models through iterative interactions with the model [8, 45]. Finally Qi et al. [56] demonstrated that even benign finetuning of the models may lead to the removal of model safety safeguards. Similar to LLMs, jailbreaking MLLMs can be achieved by either manipulating the visual input [16, 37, 54], the text prompt [40], or a combination of both [75]. Qi et al. [54] demonstrated that adding adversarial noise, which is imperceptible to humans and guided by the model's input gradients, to the input image can effectively deceive MLLMs. Recently, there has been substantial research studying the robustness of MLLMs to adversarial images [12, 19, 49, 59, 60, 84]. Another line of work [16, 37] explores embedding harmful content directly into images via typography or text-to-image tools. Recently, Luo et al. [40] showed that text-based jailbreak prompts [39, 72, 79, 90] developed for LLMs can be effectively transferred to MLLMs and are often more potent than image-based attacks. Additionally, Ying et al. [75] introduced a jail-breaking framework that simultaneously perturbs both visual and textual modalities. See also the survey by Gu et al. [17] on more works on prompt-based attacks against MLLMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 364, + 298, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 298, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 298, + 616 + ], + "type": "text", + "content": "Defense via safety alignment (fine-tuning). Defense approaches typically aim to enhance the safety in MLLMs by either fine-tuning the model on a dataset of safe image-text pairs [89] or adversarial training [89]. For instance, Zong et al. [89] proposed a safety instruction-following dataset VLGuard for safeguarding MLLMs. DRESS [10] leverages natural language feedback (NLF) from LLMs to improve the alignment and interactions within MLLMs. MLLM-Protector [53] proposed a plug-and-play strategy combining a lightweight harm detector to identify the harmful response and a detoxifier to correct harmful outputs. However, fine-tuning-based approaches suffer from (i) the requirement of a large annotated dataset consisting of safe image-text pairs, (ii) being computationally intensive, and (iii) requiring access to model weights. Qi et al. [55] highlighted that safety alignment primarily affects the first few output tokens leaving models susceptible to simple jailbreak attacks. Zou et al. [91] introduced circuit-breaking techniques to enhance model robustness. Yu et al. [76] proposed improving the safety of MLLMs by aligning with human preferences through direct preference optimization [57]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 617, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 298, + 715 + ], + "type": "text", + "content": "Defense via inference time alignment (controlled decoding). Khanov et al. [28] proposed a reward-guided inference time alignment. Mudgal et al. [47] proposed controlled decoding through a value function for the reward to guide the language model to more safe outcomes. Chakraborty et al. [7] improved the value function by bridging the gap between the value of the decoding policy and the reference policy. Li et al. [33] proposed leveraging self-evaluation and rewind" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 556, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 253 + ], + "type": "text", + "content": "mechanisms in LLMs to generate safe responses. Gong et al. [16] proposed concatenating a fixed safety prompt with the malicious user query to evade jailbreaking. Wang et al. [67] proposed refining the safety prompt iteratively rather than using a fixed template. Mehrabi et al. [43] introduced an interactive approach to improve robustness against adversarial attacks. Zhang et al. [81] proposed mutating untrusted inputs to generate variants, identifying attack samples by examining response discrepancies among the variants. In another approach, Xu et al. [73] detected adversarial images by analyzing cross-modal similarity between harmful queries and perturbed images. Additionally, Gao et al. [14] suggested improving safety by calibrating MLLM output distribution. However, existing approaches compromise model utility, leading to overly cautious responses (Figure 1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 255, + 441, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 255, + 441, + 267 + ], + "spans": [ + { + "bbox": [ + 314, + 255, + 441, + 267 + ], + "type": "text", + "content": "3. Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": "Multimodal Large Language Models (MLLMs) are autoregressive text generation models that can process multiple modalities, such as image [86], video [80], and audio [41], in addition to text. In this work, we consider input to MLLM in the form of an image and a text prompt. Mathematically, we denote a MLLM as a mapping " + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{safe}}: \\mathcal{I} \\times \\mathcal{V}^N \\to \\mathcal{V}^M" + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": " represents the image space, and " + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": " denotes the vocabulary set. The input to the MLLM " + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{safe}}" + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": " consists of an image " + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathcal{I}" + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": " and a sequence of tokens " + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{x} := \\{x_1, x_2, \\dots, x_N\\}" + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": " (referred as prompt), where each token " + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathcal{V}" + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{input}} := [\\mathbf{I}, \\mathbf{x}]" + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": ", the MLLM " + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{safe}}(\\cdot | \\mathbf{x}_{\\mathrm{input}})" + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": " generates the output sequence " + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{y} := \\{y_1, y_2, \\dots, y_M\\}" + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "inline_equation", + "content": "y_i \\in \\mathcal{V}" + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": ", in a token by token fashion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 451, + 557, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 557, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 557, + 523 + ], + "type": "text", + "content": "Jailbreaking MLLMs. Despite safety tuning efforts aimed at preventing MLLMs from generating harmful content, jailbreak attacks as formulated in Qi et al. [54], Zou et al. [90] can still make the safety aligned model " + }, + { + "bbox": [ + 313, + 451, + 557, + 523 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{safe}}" + }, + { + "bbox": [ + 313, + 451, + 557, + 523 + ], + "type": "text", + "content": " to generate harmful outputs with the help of an adversarial prompt " + }, + { + "bbox": [ + 313, + 451, + 557, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 313, + 451, + 557, + 523 + ], + "type": "text", + "content": " obtained by solving" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 336, + 531, + 556, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 531, + 556, + 563 + ], + "spans": [ + { + "bbox": [ + 336, + 531, + 556, + 563 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathrm {a d v}} := \\arg \\min _ {\\mathbf {q} \\in \\mathcal {B}} \\sum_ {i = 1} ^ {m} - \\log \\left(\\pi_ {\\text {s a f e}} \\left(y _ {i} \\mid \\left[ \\mathbf {x} _ {\\text {i n p u t}}, \\mathbf {q} \\right]\\right)\\right), \\tag {1}", + "image_path": "d0f04aeb3bacb2c4413ad1bd09d5e998932c426a88e49cdb6bc3b32271e6c25b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 568, + 558, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 558, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 558, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 568, + 558, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 568, + 558, + 665 + ], + "type": "text", + "content": " is a constraint on the adversarial prompts and " + }, + { + "bbox": [ + 313, + 568, + 558, + 665 + ], + "type": "inline_equation", + "content": "\\{y_i\\}_{i=1}^m" + }, + { + "bbox": [ + 313, + 568, + 558, + 665 + ], + "type": "text", + "content": " are few initial tokens of intended harmful response. For example, in Figure 1 (Left), we show that given an image of a house generated by stable diffusion [58] and then perturbed with adversarial noise, along with a malicious user query to explain \"How to break into and rob a house\", the model is forced to generate unsafe responses, effectively bypassing its safety alignment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 671, + 476, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 671, + 476, + 684 + ], + "spans": [ + { + "bbox": [ + 314, + 671, + 476, + 684 + ], + "type": "text", + "content": "3.2. Safeguard Against Jailbreaks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "From the definition of adversarial prompt in (1), we observe that " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " can lead the MLLM model to generate a harmful" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25040" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 298, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 298, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 298, + 228 + ], + "type": "text", + "content": "response, " + }, + { + "bbox": [ + 55, + 72, + 298, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{\\mathrm{unsafe}} \\sim \\pi_{\\mathrm{safe}}(\\cdot|\\mathbf{x}_{\\mathrm{adv}})" + }, + { + "bbox": [ + 55, + 72, + 298, + 228 + ], + "type": "text", + "content": ", despite the training time safety alignment of " + }, + { + "bbox": [ + 55, + 72, + 298, + 228 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{safe}}" + }, + { + "bbox": [ + 55, + 72, + 298, + 228 + ], + "type": "text", + "content": ". This raises the question: Why do insufficiencies exist in safety-aligned models that make them vulnerable to attacks, and how can we mitigate them? To answer this question, we start by noticing that an attacker's goal is to learn an adversarial prompt distribution " + }, + { + "bbox": [ + 55, + 72, + 298, + 228 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 72, + 298, + 228 + ], + "type": "text", + "content": " by maximizing some \"success metric\" subject to a \"cost metric,\" which is general and covers existing attacks such as projected gradient descent (PGD) attacks [15], greedy coordinate gradient (GCG) attacks [87, 90], universal adversarial trigger (UAT) attacks [44, 64], and red teaming attacks [22, 30, 52, 70]. Mathematically, an adversary can solve the following optimization problem to learn an adversarial prompt:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 236, + 295, + 272 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 272 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p _ {\\mathrm {a d v}} := \\arg \\min _ {p} \\mathbb {E} _ {\\mathbf {q} \\sim p (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}), \\mathbf {y} \\sim \\pi_ {\\text {s a f e}} (\\cdot | \\mathbf {q})} \\left[ R _ {\\text {s a f e}} (\\mathbf {x} _ {\\text {i n p u t}}, \\mathbf {y}) \\right] \\\\ + \\beta \\operatorname {d i v} \\left(p (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}), p _ {0} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}})\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "aacdc967ba8973180dcec4b57f5be6bd94edf3bc20e7d7f27e2f6356e88ea006.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": "where the \"cost metric\" term div " + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "(p(\\cdot |\\mathbf{x}_{\\mathrm{input}}),p_0(\\cdot |\\mathbf{x}_{\\mathrm{input}}))" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": " helps keep the adversarial prompt close to a given prompt distribution, " + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": ". We can think of " + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": " as a given prompt distribution that will sample " + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{input}}" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": " only with high likelihood and have minimal probability of sampling other prompts. Then, the divergence part of the objective in (2) will make sure the adversarial prompts are closer to original prompts " + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{input}}" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": ", which will keep them interpretable and not completely out of distribution [88]. The formulation in (2) is also in accordance with the existing literature on adversarial attacks in machine learning [6, 24] because it is just trying to ensure that adversarial prompts " + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": " looks like similar to " + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{input}}" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": ". The regularization parameter " + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\beta >0" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": " would control this cost metric in (2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "text", + "content": "Training-time defense. We note that learning a safe and robust model is inherently a min-max optimization problem given by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 493, + 295, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 493, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 82, + 493, + 295, + 544 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {p} \\max _ {\\pi} \\mathbb {E} _ {\\mathbf {q} \\sim p (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}), \\mathbf {y} \\sim \\pi (\\cdot | \\mathbf {q})} \\left[ R _ {\\text {s a f e}} (\\mathbf {x} _ {\\text {i n p u t}}, \\mathbf {y}) \\right. \\\\ + \\beta \\operatorname {d i v} (p (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}), p _ {0} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}})) \\\\ \\left. - \\alpha \\operatorname {d i v} \\left(\\pi (\\cdot | \\mathbf {q}), \\pi_ {\\text {r e f}} (\\cdot | \\mathbf {q})\\right) \\right], \\tag {3} \\\\ \\end{array}", + "image_path": "2b64da80bd27ba981fe36005c7d553b58894163ffc0212ccd2d0aa711f6a332b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "type": "text", + "content": "where the inner maximization captures the safeguarding of the model, e.g., through safety alignment. The safety alignment in recent literature [50, 57] only focused on solving the inner max in (3) for a given prompt distribution (say " + }, + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "type": "text", + "content": ") to obtain " + }, + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{safe}}" + }, + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "type": "text", + "content": " which is safe only under the prompt distribution " + }, + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\sim p_0" + }, + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "type": "text", + "content": ". An adversary exploits it and learns an adversarial distribution " + }, + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "type": "inline_equation", + "content": "\\rho_{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "type": "text", + "content": " breaking the alignment and generating harmful responses. This is likely the reason why adversarial attacks on images (and other modalities) have remained challenging to address over the past decade." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Key insight for tractable inference-time defense. We reframe the optimization from " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\min_p\\max_\\pi" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\max_{\\pi}\\min_{p}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " by letting the safeguard be applied after the attacker has chosen" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 556, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 276 + ], + "type": "text", + "content": "their jailbreak prompt. While the two formulations are equivalent for convex problems, this reformulation may lead to an improvement in defense due to the minimax duality gap for general non-convex problems. Additionally, this formulation allows an efficient solution of the " + }, + { + "bbox": [ + 313, + 72, + 556, + 276 + ], + "type": "inline_equation", + "content": "\\max_{\\pi}" + }, + { + "bbox": [ + 313, + 72, + 556, + 276 + ], + "type": "text", + "content": " component during inference, leveraging recent advances in inference-time alignment methods through controlled decoding [7, 47]. Notably, inference-time alignment enables prompt-specific safety adjustments, which are challenging to implement during training. This capability is enabled by the user (or attacker) interaction protocol with MLLM. Specifically, the attacker sets the adversarial prompt, which remains unchanged throughout the process, providing a fixed input for the MLLM. This scenario ultimately reduces the challenge of learning a safe and optimal policy within the predefined adversarial prompt environment, effectively solving the problem " + }, + { + "bbox": [ + 313, + 72, + 556, + 276 + ], + "type": "inline_equation", + "content": "\\max_{\\pi}" + }, + { + "bbox": [ + 313, + 72, + 556, + 276 + ], + "type": "text", + "content": " in inference. We will provide a detailed method in the next section." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 285, + 499, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 499, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 499, + 300 + ], + "type": "text", + "content": "4. Inference Time Safety Alignment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 305, + 556, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 556, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 556, + 425 + ], + "type": "text", + "content": "In this section, we propose improving the safety of response generation by using alignment through the controlled decoding procedure proposed in [7, 47]. Our key insight is that well-trained safety rewards, denoted by " + }, + { + "bbox": [ + 313, + 305, + 556, + 425 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{safe}}(\\mathbf{x}, \\mathbf{y})" + }, + { + "bbox": [ + 313, + 305, + 556, + 425 + ], + "type": "text", + "content": ", are often accessible online, for instance, RewardBench [29]. This safety reward will give high scores for safe text and low scores for unsafe text. We remark here that the reward model itself is likely not ideal and subject to adversarial attack. With this understanding, we propose to solve the following decoding problem at inference time for each token " + }, + { + "bbox": [ + 313, + 305, + 556, + 425 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 305, + 556, + 425 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 430, + 555, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 430, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 325, + 430, + 555, + 463 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\pi_ {\\text {s a f e - d e c}} ^ {*} (\\cdot | \\mathbf {s} _ {t}) := \\arg \\max _ {\\pi} \\mathbb {E} _ {z \\sim \\pi (\\cdot | \\mathbf {s} _ {t})} [ Q _ {\\text {s a f e}} (\\mathbf {s} _ {t}, z) ] \\tag {4} \\\\ - \\alpha \\mathrm {K L} (\\pi (\\cdot | \\mathbf {s} _ {t}) | | \\pi_ {\\text {s a f e}} (\\cdot | \\mathbf {s} _ {t})), \\\\ \\end{array}", + "image_path": "cb53ae4c9e3ac1c0147cc69856d8c17d5edf9099bd905d4189e1267726fe7d51.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 468, + 556, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 468, + 556, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t\\coloneqq [\\mathbf{x}_{\\mathrm{adv}},\\mathbf{y}_{< t}]" + }, + { + "bbox": [ + 313, + 468, + 556, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 468, + 556, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{< t} = [y_0,y_1,\\dots ,y_{t - 1}]" + }, + { + "bbox": [ + 313, + 468, + 556, + 505 + ], + "type": "text", + "content": " capture the tokens generated so far. In (4), the term " + }, + { + "bbox": [ + 313, + 468, + 556, + 505 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{safe}}(\\mathbf{s}_t,z)" + }, + { + "bbox": [ + 313, + 468, + 556, + 505 + ], + "type": "text", + "content": " is the action-value function defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 339, + 510, + 555, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 510, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 339, + 510, + 555, + 525 + ], + "type": "interline_equation", + "content": "Q _ {\\text {s a f e}} \\left(\\mathbf {s} _ {t}, z\\right) = \\mathbb {E} _ {\\tau \\sim \\rho_ {\\text {s a f e}} (\\cdot | [ \\mathbf {s} _ {t}, z ])} \\left[ R _ {\\text {s a f e}} \\left([ \\mathbf {s} _ {t}, z ], \\tau\\right) \\right], \\tag {5}", + "image_path": "fb3420c19877a7a193ca3bd0fe81d0bed5f2261f6cc136ca9f8b37b395677808.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "inline_equation", + "content": "\\tau := [z_{t+1}, z_{t+2}, \\dots, z_{eos}]" + }, + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "text", + "content": " denotes the remaining trajectory starting from " + }, + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "inline_equation", + "content": "[\\mathbf{s}_t, z]" + }, + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "inline_equation", + "content": "\\rho_{\\mathrm{safe}}(\\tau | [\\mathbf{s}_t, z]) := \\prod_{i=t+1}^{\\mathrm{eos}} \\pi_{\\mathrm{safe}}(z_i | [\\mathbf{s}_t, z])" + }, + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "text", + "content": " is the trajectory level distribution, where " + }, + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "inline_equation", + "content": "eos" + }, + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "text", + "content": " denotes end of sentence token, and " + }, + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "text", + "content": " denotes the sequence of tokens generated by the model starting from " + }, + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "inline_equation", + "content": "[\\mathbf{s}_t, z]" + }, + { + "bbox": [ + 313, + 529, + 556, + 613 + ], + "type": "text", + "content": ". The closed form expression of (4) is given by [47, Theorem 2.1]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 335, + 618, + 555, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 618, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 335, + 618, + 555, + 644 + ], + "type": "interline_equation", + "content": "\\pi_ {\\text {s a f e - d e c}} ^ {*} (z | \\mathbf {s} _ {t}) = \\frac {\\pi_ {\\text {s a f e}} (z | \\mathbf {s} _ {t})}{Z} \\exp \\left(\\frac {Q _ {\\text {s a f e}} (\\mathbf {s} _ {t} , z)}{\\alpha}\\right), \\tag {6}", + "image_path": "1baabe2e8f6b9a35fcafbe8f2ec8d07f5c81e5b297a9a057b35aecf12fb381ef.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 649, + 555, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 649, + 555, + 697 + ], + "spans": [ + { + "bbox": [ + 313, + 649, + 555, + 697 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 649, + 555, + 697 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 649, + 555, + 697 + ], + "type": "text", + "content": " is the normalizing constant. For a given adversarial prompt " + }, + { + "bbox": [ + 313, + 649, + 555, + 697 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 313, + 649, + 555, + 697 + ], + "type": "text", + "content": ", we propose to follow the decoding policy derived in (6) at the inference time to generate a safe response. Our proposed algorithm is summarized in Algorithm 1." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 495, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 495, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 495, + 713 + ], + "type": "text", + "content": "1https://huggingface.co/spaces/allenai/reward-bench" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25041" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 96 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 296, + 96 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 296, + 96 + ], + "type": "text", + "content": "Algorithm 1 IMMUNE: Safety against Jailbreaking attacks via Inference-time alignment" + } + ] + } + ], + "index": 0, + "type": "text" + }, + { + "bbox": [ + 61, + 99, + 298, + 300 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "spans": [ + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "text", + "content": "1: Input: Base MLLM " + }, + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{safe}}" + }, + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "text", + "content": ", safe reward model " + }, + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{safe}}" + }, + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "text", + "content": ", adversarial input " + }, + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "text", + "content": ", number of tokens sampled " + }, + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "text", + "content": ", alignment parameter " + }, + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "text", + "content": ", vocabulary set " + }, + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 62, + 99, + 298, + 135 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 136, + 155, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 136, + 155, + 146 + ], + "spans": [ + { + "bbox": [ + 62, + 136, + 155, + 146 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 62, + 136, + 155, + 146 + ], + "type": "inline_equation", + "content": "t = 0, \\dots, T" + }, + { + "bbox": [ + 62, + 136, + 155, + 146 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 147, + 296, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 147, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 62, + 147, + 296, + 171 + ], + "type": "text", + "content": "3: Current state: " + }, + { + "bbox": [ + 62, + 147, + 296, + 171 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t = [\\mathbf{x}_{\\mathrm{adv}},\\mathbf{y}_{< t}]" + }, + { + "bbox": [ + 62, + 147, + 296, + 171 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 62, + 147, + 296, + 171 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 62, + 147, + 296, + 171 + ], + "type": "text", + "content": " is adversarial input and " + }, + { + "bbox": [ + 62, + 147, + 296, + 171 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{< t} = [y_0,y_1,\\dots ,y_{t - 1}]" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 172, + 294, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 294, + 195 + ], + "type": "text", + "content": "4: Sample top-k tokens using base model " + }, + { + "bbox": [ + 62, + 172, + 294, + 195 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{safe}}" + }, + { + "bbox": [ + 62, + 172, + 294, + 195 + ], + "type": "text", + "content": " and store as " + }, + { + "bbox": [ + 62, + 172, + 294, + 195 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}} = \\{z_i : z_i \\sim \\pi_{\\mathrm{safe}}(\\cdot|\\mathbf{s}_t)\\}_{i=1}^k" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 196, + 143, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 196, + 143, + 206 + ], + "spans": [ + { + "bbox": [ + 62, + 196, + 143, + 206 + ], + "type": "text", + "content": "5: for " + }, + { + "bbox": [ + 62, + 196, + 143, + 206 + ], + "type": "inline_equation", + "content": "z\\in \\mathcal{V}" + }, + { + "bbox": [ + 62, + 196, + 143, + 206 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 208, + 238, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 208, + 238, + 220 + ], + "spans": [ + { + "bbox": [ + 62, + 208, + 238, + 220 + ], + "type": "text", + "content": "6: Evaluate " + }, + { + "bbox": [ + 62, + 208, + 238, + 220 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{safe}}" + }, + { + "bbox": [ + 62, + 208, + 238, + 220 + ], + "type": "text", + "content": " as in Equation (5)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 220, + 294, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 294, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 294, + 245 + ], + "type": "text", + "content": "7: Compute decoding score for token " + }, + { + "bbox": [ + 62, + 220, + 294, + 245 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 62, + 220, + 294, + 245 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 62, + 220, + 294, + 245 + ], + "type": "inline_equation", + "content": "g_{z} = \\frac{1}{\\alpha} Q_{\\mathrm{safe}}([s_t,z]) + \\log \\pi_{\\mathrm{saf}}(z|s_t)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 245, + 291, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 245, + 291, + 262 + ], + "spans": [ + { + "bbox": [ + 62, + 245, + 291, + 262 + ], + "type": "text", + "content": "8: Distribution over " + }, + { + "bbox": [ + 62, + 245, + 291, + 262 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}" + }, + { + "bbox": [ + 62, + 245, + 291, + 262 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 62, + 245, + 291, + 262 + ], + "type": "inline_equation", + "content": "f(z|\\mathbf{s}_t) = \\frac{\\exp(g_z)}{\\sum_{z'\\in\\hat{\\mathcal{V}}}\\exp(g_{z'})}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 262, + 201, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 201, + 273 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 201, + 273 + ], + "type": "text", + "content": "9: Next token " + }, + { + "bbox": [ + 62, + 262, + 201, + 273 + ], + "type": "inline_equation", + "content": "y_{t}\\sim f(\\cdot |\\mathbf{s}_{t})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 274, + 209, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 209, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 209, + 285 + ], + "type": "text", + "content": "10: Next state " + }, + { + "bbox": [ + 61, + 274, + 209, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{t + 1}\\gets [\\mathbf{s}_t,y_t]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 287, + 172, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 287, + 172, + 300 + ], + "spans": [ + { + "bbox": [ + 61, + 287, + 172, + 300 + ], + "type": "text", + "content": "11: Return " + }, + { + "bbox": [ + 61, + 287, + 172, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = [y_0, \\dots, y_T]" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 323, + 173, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 173, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 173, + 336 + ], + "type": "text", + "content": "5. Theoretical Insights" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 342, + 297, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 297, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 297, + 403 + ], + "type": "text", + "content": "In this section, we formally derive a bound on the suboptimality of our proposed decoding-based approach under an adversarial prompt distribution. Given a safe prompt distribution " + }, + { + "bbox": [ + 55, + 342, + 297, + 403 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 342, + 297, + 403 + ], + "type": "text", + "content": " and an adversarial prompt distribution " + }, + { + "bbox": [ + 55, + 342, + 297, + 403 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 342, + 297, + 403 + ], + "type": "text", + "content": ", the sub-optimality gap of our algorithm can be defined as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 412, + 296, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 412, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 59, + 412, + 296, + 456 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Delta_ {\\text {s u b - g a p}} \\left(\\mathbf {x} _ {\\text {i n p u t}}\\right) \\tag {7} \\\\ := \\underset { \\begin{array}{c} \\mathbf {x} \\sim p _ {0} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}) \\\\ \\mathbf {y} \\sim \\rho_ {*} (\\cdot | \\mathbf {x}) \\end{array} } {\\mathbb {E}} \\left[ R _ {\\text {s a f e}} (\\mathbf {x}, \\mathbf {y}) \\right] - \\underset { \\begin{array}{c} \\mathbf {x} \\sim p _ {\\text {a d v}} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}) \\\\ \\mathbf {y} \\sim \\rho_ {\\text {s a f e - d e c}} (\\cdot | \\mathbf {x}) \\end{array} } {\\mathbb {E}} \\left[ R _ {\\text {s a f e}} (\\mathbf {x}, \\mathbf {y}) \\right], \\\\ \\end{array}", + "image_path": "c22e0a164945cdd9db6641822d7f80fd9e4a13f9f6935d2887355a2b59594252.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 464, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 464, + 296, + 584 + ], + "type": "inline_equation", + "content": "\\rho_{\\mathrm{safe - dec}}" + }, + { + "bbox": [ + 55, + 464, + 296, + 584 + ], + "type": "text", + "content": " is the trajectory level distribution corresponding to token level " + }, + { + "bbox": [ + 55, + 464, + 296, + 584 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{safe - dec}}^*" + }, + { + "bbox": [ + 55, + 464, + 296, + 584 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 464, + 296, + 584 + ], + "type": "inline_equation", + "content": "\\rho_*" + }, + { + "bbox": [ + 55, + 464, + 296, + 584 + ], + "type": "text", + "content": " is the trajectory level optimal distribution for safety reward. The suboptimality gap in (7) captures how safe our proposed decoding-based approach is in generating responses to adversarial prompts. The suboptimality gap in (7) is similar but different to the one utilized in [5] because here, the goal is to capture the impact of adversarial prompt distribution specifically. We establish an upper bound in the following Theorem 1 on the suboptimality gap defined in (7)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 65, + 607, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 607, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 65, + 607, + 287, + 668 + ], + "type": "text", + "content": "Theorem 1. Let " + }, + { + "bbox": [ + 65, + 607, + 287, + 668 + ], + "type": "inline_equation", + "content": "R_{safe}(\\mathbf{x},\\mathbf{y}) \\leq R_{max}" + }, + { + "bbox": [ + 65, + 607, + 287, + 668 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 65, + 607, + 287, + 668 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 65, + 607, + 287, + 668 + ], + "type": "text", + "content": " is a given prompt distribution, " + }, + { + "bbox": [ + 65, + 607, + 287, + 668 + ], + "type": "inline_equation", + "content": "p_{adv}" + }, + { + "bbox": [ + 65, + 607, + 287, + 668 + ], + "type": "text", + "content": " adversarial prompt distribution, " + }, + { + "bbox": [ + 65, + 607, + 287, + 668 + ], + "type": "inline_equation", + "content": "\\rho_{*}(\\cdot |\\mathbf{x})" + }, + { + "bbox": [ + 65, + 607, + 287, + 668 + ], + "type": "text", + "content": " denotes the optimal trajectory level distribution for the safe reward, and " + }, + { + "bbox": [ + 65, + 607, + 287, + 668 + ], + "type": "inline_equation", + "content": "\\rho_{safe}(\\cdot |\\mathbf{x})" + }, + { + "bbox": [ + 65, + 607, + 287, + 668 + ], + "type": "text", + "content": " denote the given training time safety aligned trajectory level distri" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 324, + 72, + 433, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 433, + 83 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 433, + 83 + ], + "type": "text", + "content": "bution. Then, it holds that" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 325, + 92, + 546, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 92, + 546, + 129 + ], + "spans": [ + { + "bbox": [ + 325, + 92, + 546, + 129 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Delta_ {\\text {s u b - g a p}} (\\mathbf {x} _ {\\text {i n p u t}}) \\leq R _ {\\max } \\sqrt {K L (p _ {0} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}) \\| p _ {\\text {a d v}} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}}))} \\\\ + \\alpha K L \\left(\\rho_ {*} \\left(\\cdot | \\mathbf {x} _ {\\text {i n p u t}}\\right) \\| \\rho_ {s a f e} (\\cdot | \\mathbf {x} _ {\\text {i n p u t}})\\right), \\\\ \\end{array}", + "image_path": "3cd68cefc15f864e5b203c7b353e4c223f46bfe8a526ca73396ae36950cfbf4d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 324, + 137, + 543, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 137, + 543, + 162 + ], + "spans": [ + { + "bbox": [ + 324, + 137, + 543, + 162 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 324, + 137, + 543, + 162 + ], + "type": "inline_equation", + "content": "\\alpha > 0" + }, + { + "bbox": [ + 324, + 137, + 543, + 162 + ], + "type": "text", + "content": " is a regularization parameter balancing the KL-divergence term in the policy alignment." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 178, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 251 + ], + "type": "text", + "content": "Proof Outline and Insights. The proof of Theorem 1 (see Appendix E for details) relies on decomposing the suboptimality gap term " + }, + { + "bbox": [ + 313, + 178, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{sub - gap}}" + }, + { + "bbox": [ + 313, + 178, + 555, + 251 + ], + "type": "text", + "content": " into two components, " + }, + { + "bbox": [ + 313, + 178, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\Delta_1" + }, + { + "bbox": [ + 313, + 178, + 555, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 178, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\Delta_2" + }, + { + "bbox": [ + 313, + 178, + 555, + 251 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 178, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{sub - gap}} = \\Delta_1 + \\Delta_2" + }, + { + "bbox": [ + 313, + 178, + 555, + 251 + ], + "type": "text", + "content": ", with bounds derived from total variation distance and Pinsker's inequality, respectively. We can write the decomposition as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 260, + 550, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 260, + 550, + 289 + ], + "spans": [ + { + "bbox": [ + 314, + 260, + 550, + 289 + ], + "type": "interline_equation", + "content": "\\Delta_{1}:= \\mathop{\\mathbb{E}}_{\\substack{\\mathbf{x}\\sim p_{0}(\\cdot |\\mathbf{x}_{\\text{input}})\\\\ \\mathbf{y}\\sim \\rho_{*}(\\cdot |\\mathbf{x})}}\\left[R_{\\text{safe}}(\\mathbf{x},\\mathbf{y})\\right] - \\mathop{\\mathbb{E}}_{\\substack{\\mathbf{x}\\sim p_{\\text{adv}}(\\cdot |\\mathbf{x}_{\\text{input}})\\\\ \\mathbf{y}\\sim \\rho_{*}(\\cdot |\\mathbf{x})}}\\left[R_{\\text{safe}}(\\mathbf{x},\\mathbf{y})\\right],", + "image_path": "f5221139a70f8ca163b97aa9f927fb7803ea08ec7305376f8bc9cd2430a5b309.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 292, + 553, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 292, + 553, + 321 + ], + "spans": [ + { + "bbox": [ + 315, + 292, + 553, + 321 + ], + "type": "interline_equation", + "content": "\\Delta_{2}:= \\mathop{\\mathbb{E}}_{\\substack{\\mathbf{x}\\sim p_{\\text{adv}}(\\cdot |\\mathbf{x}_{\\text{input}})\\\\ \\mathbf{y}\\sim \\rho_{*}(\\cdot |\\mathbf{x})}}[R_{\\text{safe}}(\\mathbf{x},\\mathbf{y})] - \\mathop{\\mathbb{E}}_{\\substack{\\mathbf{x}\\sim p_{\\text{adv}}(\\cdot |\\mathbf{x}_{\\text{input}})\\\\ \\mathbf{y}\\sim \\rho_{\\text{safe -dec}}(\\cdot |\\mathbf{x})}}[R_{\\text{safe}}(\\mathbf{x},\\mathbf{y})].", + "image_path": "a703ab6efdb59d75d1af7c807474ebe1a37fd895c49acf1e87bb330d5297a374.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 331, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 366 + ], + "type": "text", + "content": "Next, we provide the interpretation of the mathematical result in Theorem 1 and intuitive insights about different terms as follows:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 371, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 313, + 371, + 555, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 466 + ], + "type": "text", + "content": "(1) Practical Implications of the Bound: Theorem (1) shows that sub-optimality gap upper bound is minimized when " + }, + { + "bbox": [ + 313, + 371, + 555, + 466 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{adv}}" + }, + { + "bbox": [ + 313, + 371, + 555, + 466 + ], + "type": "text", + "content": " remains close to " + }, + { + "bbox": [ + 313, + 371, + 555, + 466 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 371, + 555, + 466 + ], + "type": "text", + "content": " (low KL divergence) and also when " + }, + { + "bbox": [ + 313, + 371, + 555, + 466 + ], + "type": "inline_equation", + "content": "\\rho_{\\mathrm{safe}}" + }, + { + "bbox": [ + 313, + 371, + 555, + 466 + ], + "type": "text", + "content": " (which is our reference policy for the inference time alignment problem) aligns closely with " + }, + { + "bbox": [ + 313, + 371, + 555, + 466 + ], + "type": "inline_equation", + "content": "\\rho_*" + }, + { + "bbox": [ + 313, + 371, + 555, + 466 + ], + "type": "text", + "content": " (low policy KL divergence). This result can justify the use of KL-regularized objectives in safety-required domains, especially under adversarial prompt conditions." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "text", + "content": "(2) Role of Regularization Parameter " + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "text", + "content": ": The adversarial prompt distribution in (2), " + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{adv}}" + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "text", + "content": ", is defined as a solution to a KL-regularized optimization problem that minimizes expected reward " + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{safe}}" + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "text", + "content": " relative to " + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "text", + "content": ", with regularization parameter " + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "text", + "content": ". The lower values of " + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "text", + "content": " allow " + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{adv}}" + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "text", + "content": " to diverge further from " + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "text", + "content": ", potentially increasing " + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "inline_equation", + "content": "\\Delta_1" + }, + { + "bbox": [ + 313, + 467, + 555, + 539 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "text", + "content": "(3) Impact of " + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\Delta_{2}" + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "text", + "content": ": The value of " + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "text", + "content": " affects the degree to which the policy " + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\rho_{\\mathrm{safe - dec}}" + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "text", + "content": " is incentivized to align with the safety policy " + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\rho_{*}" + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "text", + "content": ". The lower values of " + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "text", + "content": " prioritize safety more heavily, thereby reducing " + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\Delta_{2}" + }, + { + "bbox": [ + 313, + 542, + 555, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "(3) In proof, " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\Delta_{1}" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": " expresses how far an adversary can exploit the adversarial prompt distribution to attack the aligned model. In " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\Delta_{2}" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": ", we show how, with inference time alignment, we can achieve safety even under the adversarial prompt distribution owing to the closed form of the KL regularized RLHF objective. We note that Mroueh [46, Proposition 3] also establishes upper bounds on the difference in expected rewards between two distributions in a continuous setting, relying on tail assumptions (e.g., sub-Gaussian, sub-Gamma) to control deviations. In contrast, our Theorem 1 operates" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25042" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "entirely in a discrete setting and explicitly decomposes suboptimality (difference in expected rewards defined in (7)) into contributions from both the prompter and the language model. Furthermore, we leverage the optimality of RLHF to derive tighter upper bounds, which is not used in Mroueh [46]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 149, + 296, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 296, + 246 + ], + "type": "text", + "content": "Remark: Note that the theoretical guarantees discussed are conditional upon a specific prompt design structure, which originates from an adversarial prompter targeting a given unsafe query. This inherently prevents us from being able to obtain a similar guarantee for training time safety fine-tuning since the adversarial prompts would not be known in advance. In our experiments, we will not make these assumptions, as we experiment with several recent static safety benchmarks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 258, + 136, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 136, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 136, + 272 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 278, + 178, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 178, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 178, + 291 + ], + "type": "text", + "content": "6.1. Experimental Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 297, + 296, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 332 + ], + "type": "text", + "content": "Jail-break Datasets. We conduct a comprehensive evaluation by considering both text-based and image-based jailbreak attacks:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "text", + "content": "1. Text-based Attacks. For text-based attacks, we use the JailbreakV-28K dataset [40]. This dataset integrates effective text-based jailbreak prompts for LLMs with various types of image data. Specifically, the text prompts include template-based, persuasive, and logic-driven approaches, while the image input may consist of noise, be blank, contain natural images, or be generated by stable diffusion. We evaluate all possible combinations of image and text pairs to ensure thorough evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 447, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 471 + ], + "type": "text", + "content": "2. Image-based Attacks. For image-based attacks, we evaluate the following datasets:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 474, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 604 + ], + "type": "text", + "content": "- MM-SafetyBench [37]: This benchmark assesses MLLM safety across 13 commonly prohibited scenarios and behaviors specified in usage guidelines from OpenAI [1] and Meta [26]. Each image in this dataset is associated with three types of image inputs: (1) Stable-diffusion images (SD), which are generated using stable diffusion [58] and directly relate to the query; (2) Typography (TYPO) images, containing optical character recognition representations of text from the malicious query; and (3) SD+TYPO images, which combine stable diffusion visuals with typographic subtitles." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "content": "- FigStep [16]: This attack converts harmful instructions into images using typography. Specifically, malicious queries on 10 forbidden topics from MM-SafetyBench are first paraphrased to start with nouns like \"Steps to\", \"List of\", or \"Methods to\", and then the instructions are presented as text within images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "- Visual Adversarial Attacks [54]: Qi et al. proposed an optimization-based attack, where the malicious user generates adversarial images by introducing carefully crafted" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 555, + 95 + ], + "type": "text", + "content": "perturbations to the original image, causing models to produce harmful content." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 102, + 555, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 102, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 555, + 233 + ], + "type": "text", + "content": "Models Used. We perform evaluation on 5 state-of-art open-source MLLMs: LLaVA-1.5-7B [34], LLaVA-1.6-7B [35], MiniGPT-4-7B/13B [86], and Qwen-VL-Chat-7B [4]. For both LLaVA and MiniGPT, we use Vicuna [11] as the text encoder. For IMMUNE, we set the number of tokens sampled " + }, + { + "bbox": [ + 313, + 102, + 555, + 233 + ], + "type": "inline_equation", + "content": "k = 10" + }, + { + "bbox": [ + 313, + 102, + 555, + 233 + ], + "type": "text", + "content": " and the alignment parameter " + }, + { + "bbox": [ + 313, + 102, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 313, + 102, + 555, + 233 + ], + "type": "text", + "content": ". We report ablations in Appendix F. For main evaluations, we use LLaMa-3.1-8B" + }, + { + "bbox": [ + 313, + 102, + 555, + 233 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 102, + 555, + 233 + ], + "type": "text", + "content": " as the safety model " + }, + { + "bbox": [ + 313, + 102, + 555, + 233 + ], + "type": "inline_equation", + "content": "R_{\\text{safe}}" + }, + { + "bbox": [ + 313, + 102, + 555, + 233 + ], + "type": "text", + "content": ". The safety model has been trained on a mixture of preference datasets including HelpSteer2 [68], OffsetBias [51], WildGuard [20], and Magpie [74]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "text", + "content": "Baseline Defenses. We compare IMMUNE with recent state-of-the-art inference-time jail-break defense frameworks, including FigStep [16], Adashield [67], and CoCA [14]. For a fair comparison, all baseline defense methods are evaluated on a unified test dataset using consistent metrics. Detailed descriptions of baselines are provided in the Appendix G." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "type": "text", + "content": "Evaluation Metric. To assess the effectiveness of a jailbreak attack, we follow the literature [14, 40, 67] and measure the Attack Success Rate (ASR). Specifically, given a test dataset " + }, + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{unsafe}}" + }, + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "type": "text", + "content": " of crafted jailbreak image-text pairs, the ASR quantifies the ratio of harmful responses to the total number of input queries and is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 323, + 392, + 554, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 392, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 323, + 392, + 554, + 433 + ], + "type": "interline_equation", + "content": "\\operatorname {A S R} = \\frac {1}{\\left| \\mathcal {D} _ {\\text {u n s a f e}} \\right|} \\sum_ {(\\mathbf {I}, \\mathbf {x}) \\in \\mathcal {D} _ {\\text {u n s a f e}}} \\mathbb {I} \\left[ \\mathcal {C} ^ {*} (\\mathbf {x}, \\pi_ {\\theta} (\\mathbf {I}, \\mathbf {x})) = \\text {T r u e} \\right], \\tag {8}", + "image_path": "761aa15c00d702469a8119153c7f59bf7f0c1e06aed1ea61465e078572af0493.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 434, + 556, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 577 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 434, + 556, + 577 + ], + "type": "inline_equation", + "content": "(\\mathbf{I},\\mathbf{x})" + }, + { + "bbox": [ + 313, + 434, + 556, + 577 + ], + "type": "text", + "content": " represents an image-text jailbreak prompt pair, " + }, + { + "bbox": [ + 313, + 434, + 556, + 577 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 313, + 434, + 556, + 577 + ], + "type": "text", + "content": " is an indicator function, and " + }, + { + "bbox": [ + 313, + 434, + 556, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^*" + }, + { + "bbox": [ + 313, + 434, + 556, + 577 + ], + "type": "text", + "content": " is an oracle jailbreak classifier that verifies whether the generated response " + }, + { + "bbox": [ + 313, + 434, + 556, + 577 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(\\mathbf{I},\\mathbf{x})" + }, + { + "bbox": [ + 313, + 434, + 556, + 577 + ], + "type": "text", + "content": " aligns with the malicious intent of the user query " + }, + { + "bbox": [ + 313, + 434, + 556, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 434, + 556, + 577 + ], + "type": "text", + "content": ". Consistent with earlier works [9, 14, 40], we use both Llama-Guard-3 Jailbreak Judge and the GPT-4 as the oracle jailbreak classifier. Hence, a lower ASR value indicates a more robust defense against jailbreak attacks. To account for randomness in output generation, we sample three independent responses for each query and consider the model successfully jailbroken if any one of the three responses is flagged as jailbroken by the oracle classifier." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 586, + 454, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 454, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 454, + 598 + ], + "type": "text", + "content": "6.2. Safety Evaluation results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 610, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 555, + 693 + ], + "type": "text", + "content": "Evaluation on Text-based attacks. In Table 1, we present the results of this evaluation for different combinations of image types and text prompts. To be specific, the malicious user queries can be template-based, persuasive in nature, or logic-driven, while the image input consists of random noise, be blank, contain natural images, or be generated by stable diffusion. For this evaluation, we used the Llama-Guard-3" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 440, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 440, + 712 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 440, + 712 + ], + "type": "inline_equation", + "content": "^{2}\\mathrm{LxzGordon/URM-LLaMa - }3.1 - 8\\mathrm{B}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25043" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 309 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 309 + ], + "type": "table", + "html": "
ModelDefense StrategyNoiseSDNatureBlankAverage
TemplatePersuadeLogicTemplatePersuadeLogicTemplatePersuadeLogicTemplatePersuadeLogic
LLaVA-1.6Original66.1237.4578.5867.3437.5677.5769.2340.7882.6166.6739.4581.6060.27
FigStep [16]61.1239.2762.1662.3440.1854.0563.4135.2256.7661.0938.4352.7051.64
AdaShield [67]38.420.0011.0838.131.5618.1339.293.2219.1442.780.4816.1219.23
CoCA [14]61.2339.1762.1661.3441.2852.7063.1135.2255.4161.0937.3652.7051.37
IMMUNE (Ours)5.230.000.008.140.000.008.450.000.004.670.000.002.45
LLaVA-1.5Original58.1239.4476.5661.4738.3574.5558.2242.1177.5759.3340.2674.5556.39
FigStep [16]64.1728.3462.1662.2332.1868.9158.3937.2772.9761.0931.4268.9152.46
AdaShield [67]32.140.005.4131.360.004.2330.220.005.4129.670.0010.8112.86
CoCA [14]59.1826.3435.1358.4230.2935.1349.2232.1128.3759.0723.3937.8339.87
IMMUNE (Ours)9.230.000.008.140.000.001.470.000.005.320.000.002.10
MiniGPT-4-7BOriginal75.2354.1283.6272.4437.3385.6373.1845.2786.6472.0955.4286.6467.15
FigStep [16]47.2217.1832.4342.135.3724.3237.298.4518.9243.0817.4143.2427.74
AdaShield [67]47.2322.1432.2441.3718.4627.2058.1512.2830.2246.0919.4332.2432.21
CoCA [14]35.1818.2722.9740.3421.1431.0835.2918.4227.0348.1121.3640.5429.74
IMMUNE (Ours)18.236.1244.5911.348.2729.7217.188.4527.0216.0910.3743.2418.34
MiniGPT-4-13BOriginal74.1847.1290.6780.2353.3485.6379.2952.4583.7879.0753.2287.6570.62
FigStep [16]44.1320.2858.1143.2218.4747.3040.3627.1448.6542.0919.4258.1137.41
AdaShield [67]63.2223.1859.4464.4731.1274.5569.1527.3568.9165.0923.4153.3950.55
CoCA [14]35.2921.3436.4940.1822.2740.5435.1220.4135.1448.2221.3652.7033.21
IMMUNE (Ours)25.1718.2959.4429.3419.4157.4229.2223.1858.4323.4723.3758.4332.94
Qwen-VLOriginal46.123.2712.0952.236.189.0753.343.427.0553.114.3615.1122.99
FigStep [16]32.180.001.3540.420.005.4130.230.005.4144.122.231.3514.42
AdaShield [67]30.170.0018.1328.341.2915.1123.281.377.0542.110.4118.1315.88
CoCA [14]29.126.2713.5127.347.185.4127.291.4213.5128.4713.3613.5115.69
IMMUNE (Ours)10.272.185.4121.342.294.0318.222.355.4120.173.377.058.58
", + "image_path": "db2f639300afcce56235b7b6a8c942382c45ecc3cf8fa4e7a5a2fce0f6f1602d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 396, + 555, + 548 + ], + "blocks": [ + { + "bbox": [ + 55, + 317, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 317, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 555, + 384 + ], + "type": "text", + "content": "Table 1. Evaluation on Text-Based Jailbreak Attacks. We report the Attack Success Rate (ASR) for various baseline defense strategies across recent MLLMs on text-based jailbreak attacks [40]. In this setup, the image input may be noise, Stable Diffusion-generated (SD), natural images (Nature), or blank, while the text query is structured as either template-based, persuasive, or logic-driven prompts. Lower ASR values indicate stronger resilience against jailbreak attacks. For this evaluation, we used Llama-Guard-3 Jailbreak Judge as the jailbreak classifier. IMMUNE consistently outperforms all compared defense frameworks across all MLLMs. The best results (lowest ASR) are highlighted in bold. All values are reported in %." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 396, + 555, + 548 + ], + "lines": [ + { + "bbox": [ + 56, + 396, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 555, + 548 + ], + "type": "table", + "html": "
ModelDefense StrategyIllegal ActivityMalware GenerationPornographyHate SpeechPhysical HarmFraudAverage
SDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPESDTYPESD-TYPE
LLaVA-1.6Original34.0280.4178.3520.4570.4570.457.0742.0535.0320.0651.0252.0518.0163.0163.0530.0367.0472.0747.94
FigStep [16]27.8449.4879.3825.0347.7359.097.0928.0741.0330.0436.0746.0720.0221.0852.0329.0536.0855.0737.64
AdaShield [67]19.5962.8941.249.0954.5543.187.0525.0519.066.0620.0313.039.0926.0427.046.0136.0235.0824.40
CoCA [14]24.2755.6776.0117.2759.0970.459.0233.0743.0825.0142.0752.0623.0626.0457.0125.0137.0858.0739.79
IMMUNE (Ours)0.001.035.150.000.004.559.233.0718.020.000.003.024.095.035.040.002.077.073.94
LLaVA-1.5Original34.0280.4180.4143.1875.0170.4513.0256.0739.0724.0575.0269.0725.0764.0367.0532.0373.0765.0353.85
FigStep [16]25.5263.9273.2820.4375.0544.0916.0426.0538.0638.0351.0655.0121.0457.0459.0636.0862.0861.0445.54
AdaShield [67]9.281.033.090.002.270.009.067.034.047.071.042.0311.037.053.0310.011.072.084.81
CoCA [14]17.1380.4180.4125.9127.779.918.0752.0655.0610.0863.1967.0325.0745.0557.0726.0357.0859.0744.65
IMMUNE (Ours)0.461.030.000.400.0013.646.989.0919.050.523.025.045.174.037.020.000.006.043.51
MiniGPT-4-7BOriginal13.4343.3435.052.2734.0925.0827.0414.0121.066.0229.0517.053.0834.0826.025.0540.0436.0223.04
FigStep [16]19.5943.3438.146.8240.9115.9126.0119.0623.0611.0428.0816.0314.0631.0826.0712.0644.0517.0724.21
AdaShield [67]11.3435.0518.5618.1822.739.0916.6726.0315.0422.678.0426.048.0827.0914.0918.1834.0422.0720.05
CoCA [14]9.2842.2728.876.8220.4518.1819.0812.0924.076.0512.0710.0325.0416.6726.0725.068.3325.0218.92
IMMUNE (Ours)13.4622.6813.4311.3620.4518.1817.0412.0921.033.088.047.087.0214.0120.020.000.000.0011.03
MiniGPT-4-13BOriginal17.5340.2142.2713.6422.7327.2710.0415.0818.039.0522.0921.0717.0320.0235.0819.0537.0341.0923.97
FigStep [16]25.7746.3936.0822.7327.2729.5513.0216.0819.0320.0830.0720.0818.0136.0438.0821.0532.0938.0327.17
AdaShield [67]21.6523.7122.6820.4534.0925.0415.059.0215.0311.0625.0321.0114.0322.0123.0723.0528.0215.0919.84
CoCA [14]25.0350.0746.0314.0625.0343.0613.0225.018.3316.0416.6716.6716.6725.0641.6725.0641.6725.0726.13
IMMUNE (Ours)23.7822.6819.599.0918.1829.5520.0814.0815.0611.0813.0813.0210.0818.0322.086.020.000.0014.24
", + "image_path": "d37b6a0e41a8e5fb66ba28c86c4d52d76b0e8a782b090d5c8810253079b130de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 556, + 556, + 592 + ], + "lines": [ + { + "bbox": [ + 55, + 556, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 556, + 592 + ], + "type": "text", + "content": "Table 2. Evaluation on MM-Safety Bench. We report Attack Success Rate (ASR) for 6 categories of prohibited scenarios from MM-Safety Bench [38]. For this evaluation, we used GPT-4 as the jailbreak classifier. IMMUNE significantly reduces ASR across all categories. The best results (lowest ASR) are highlighted in bold. All values are reported in %." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 603, + 297, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 297, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 297, + 711 + ], + "type": "text", + "content": "Jailbreak Judge as the oracle classifier. From our evaluation, we draw the following key insights: (1) With the original decoding, ASR is notably high for most models, reaching up to " + }, + { + "bbox": [ + 55, + 603, + 297, + 711 + ], + "type": "inline_equation", + "content": "70.62\\%" + }, + { + "bbox": [ + 55, + 603, + 297, + 711 + ], + "type": "text", + "content": " for MiniGPT-4-13B, underscoring both the severity of the attack and the underlying serious safety concern. (2) Across all MLLMs, IMMUNE demonstrates superior performance, substantially reducing ASR compared to other competitive defense strategies. Specifically, on MiniGPT-4-13B, IMMUNE reduces ASR by " + }, + { + "bbox": [ + 55, + 603, + 297, + 711 + ], + "type": "inline_equation", + "content": "37.68\\%" + }, + { + "bbox": [ + 55, + 603, + 297, + 711 + ], + "type": "text", + "content": " compared to the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 603, + 556, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 556, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 556, + 711 + ], + "type": "text", + "content": "original decoding and by " + }, + { + "bbox": [ + 313, + 603, + 556, + 711 + ], + "type": "inline_equation", + "content": "17.61\\%" + }, + { + "bbox": [ + 313, + 603, + 556, + 711 + ], + "type": "text", + "content": " compared to state-of-art defense strategy AdaShield [67]. (3) We observe that logic-based attacks reveal a model-dependent variation in IMMUNE's performance. Specifically, for logic and persuade-based attacks, IMMUNE achieves an ASR of " + }, + { + "bbox": [ + 313, + 603, + 556, + 711 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 313, + 603, + 556, + 711 + ], + "type": "text", + "content": " across all input types on LLaVA-1.6 and LLaVA-1.5. However, on MiniGPT-4-7B, logic-based attacks with noise inputs result in a significantly higher ASR of " + }, + { + "bbox": [ + 313, + 603, + 556, + 711 + ], + "type": "inline_equation", + "content": "44.59\\%" + }, + { + "bbox": [ + 313, + 603, + 556, + 711 + ], + "type": "text", + "content": ", compared to CoCA [14] " + }, + { + "bbox": [ + 313, + 603, + 556, + 711 + ], + "type": "inline_equation", + "content": "(22.97\\%)" + }, + { + "bbox": [ + 313, + 603, + 556, + 711 + ], + "type": "text", + "content": " and AdaShield [67] " + }, + { + "bbox": [ + 313, + 603, + 556, + 711 + ], + "type": "inline_equation", + "content": "(32.24\\%)" + }, + { + "bbox": [ + 313, + 603, + 556, + 711 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25044" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 255 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 255 + ], + "type": "table", + "html": "
ModelDefense StrategyIllegal ActivityHate SpeechMalware GenerationPhysical HarmFraudAdult ContentPrivacy ViolationLegal OpinionFinancial AdviceHealth ConsultationAverage
LLaVA-1.6Original11.700.0021.9211.880.0026.270.0036.0431.809.5314.68
FigStep [16]9.620.0014.232.030.0011.950.006.0513.586.245.90
AdaShield [67]3.580.005.650.000.0011.790.000.009.610.003.62
CoCA [14]13.624.296.065.978.3320.396.5727.7221.7017.5613.34
IMMUNE (Ours)12.040.000.000.000.000.000.000.000.000.001.21
LLaVA-1.5Original74.000.0083.9761.6210.4693.961.7292.3782.2984.0958.58
FigStep [16]65.531.9570.0748.473.7773.684.1371.7276.2967.5148.48
AdaShield [67]14.380.0011.5413.880.002.100.002.2012.2114.167.24
CoCA [14]44.038.2338.0422.086.4942.489.6544.1541.9430.4528.63
Ours28.210.006.302.120.000.000.004.403.841.864.23
MiniGPT-4-7BOriginal10.152.1416.022.440.0019.770.0033.6226.2111.9311.95
FigStep [16]11.630.0024.450.003.7513.780.0010.0615.963.998.32
AdaShield [67]4.290.0012.175.674.072.120.0012.128.239.645.76
CoCA [14]7.850.000.002.471.856.360.0015.6313.876.285.74
IMMUNE (Ours)7.980.009.826.140.004.430.007.855.713.594.43
MiniGPT-4-13BOriginal20.050.0030.4115.980.0042.140.0037.8146.2819.5321.17
FigStep [16]24.110.0038.4521.880.0029.570.0021.8334.5023.8119.52
AdaShield [67]18.060.007.663.930.0011.551.8228.2029.9212.2611.65
CoCA [14]24.38.3313.5718.125.5431.928.3313.9817.8318.3815.93
IMMUNE (Ours)16.312.248.275.540.0010.180.0016.0619.5315.989.48
Qwen-VLOriginal17.530.0040.2317.770.0032.380.0056.3439.7539.8224.01
FigStep [16]1.660.002.200.000.0020.870.0010.1210.080.004.56
AdaShield [67]4.050.006.499.680.0017.830.006.061.550.005.04
CoCA [14]29.530.0067.8834.490.0077.920.0088.2683.5661.9243.92
IMMUNE (Ours)6.500.004.370.000.009.930.007.935.880.003.23
", + "image_path": "a18eb2fd8db0b7db476172019592de934cf0805cf0e96f90aa8fcd9ad7a476ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 292, + 295, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 295, + 316 + ], + "type": "text", + "content": "Evaluation on Image-based attacks. We evaluate IMMUNE on three standard benchmarks for image-based attacks:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 317, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 296, + 507 + ], + "type": "text", + "content": "(1) MM-Safety. We report the ASR for 6 prohibited categories from MM-Safety Bench [38] across four recent MLLMs, as shown in Table 2. For response evaluation, following Liu et al. [38], we used GPT-4 as the jailbreak classifier. Our findings indicate that IMMUNE significantly reduces ASR across all MLLMs compared to state-of-art defense strategies such as CoCA [14] and AdaShield [67]. Notably, on LLaVA-1.6, IMMUNE achieves an ASR reduction of " + }, + { + "bbox": [ + 56, + 317, + 296, + 507 + ], + "type": "inline_equation", + "content": "44.0\\%" + }, + { + "bbox": [ + 56, + 317, + 296, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 317, + 296, + 507 + ], + "type": "inline_equation", + "content": "20.46\\%" + }, + { + "bbox": [ + 56, + 317, + 296, + 507 + ], + "type": "text", + "content": " compared to the base model and AdaShield [67] respectively. Across all MLLMs, we observe that attacks using typography images (TYPO) and those combining stable diffusion visuals with typography subtitles (SD-TYPO) are significantly more effective than simple stable diffusion-generated (SD) images. Additionally, samples from the \"Pornography\" category exhibit the highest ASR across all defense mechanisms." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "type": "text", + "content": "(2) FigStep. Table 3 presents the ASR results for various MLLMs on the FigStep dataset [16]. Consistent with results on MM-Safety, IMMUNE demonstrates superior performance over competitive baselines, reducing ASR by " + }, + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "type": "inline_equation", + "content": "6.45\\%" + }, + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "type": "inline_equation", + "content": "12.13\\%" + }, + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "type": "text", + "content": " on MiniGPT-4-13B and LLaVA-1.6, respectively, compared to CoCA [14]. Among all categories, \"Hate Speech\" and \"Privacy Violation\" have the lowest ASR, with most approaches achieving an ASR of near " + }, + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "type": "inline_equation", + "content": "0.0\\%" + }, + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "(3) Visual Adversarial Attacks. Finally, in Table 4, we evaluate ASR under adversarily optimized jailbreak attacks [54] across varying attack strengths, denoted by " + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": ", where higher " + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": " values correspond to stronger attacks. The \"Unconstrained\" setting represents the most challenging attack scenario. Notably, even under the \"unconstrained\" attack, IMMUNE reduces ASR by " + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "inline_equation", + "content": "24.19\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": " compared to baseline [67] on LLaVA-1.6." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Additional discussion on inference time of IMMUNE and its impact on model utility is deferred to Appendix F." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 290, + 555, + 502 + ], + "blocks": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "type": "text", + "content": "Table 3. Evaluation on FigStep. We report the Attack Success Rate (ASR) for all categories from the FigStep benchmark [16]. For this evaluation, we used GPT-4 as the jailbreak classifier. The best results (lowest ASR) are highlighted in bold. All values are reported in %." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 290, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 315, + 290, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 315, + 290, + 555, + 502 + ], + "type": "table", + "html": "
ModelDefense StrategyAttack Strength
ε = 16/255ε = 32/255ε = 64/255Unconstrained
LLaVA-1.6Original66.1264.0862.3164.80
FigStep [16]49.1050.3047.2251.13
AdaShield [67]35.2139.7134.0540.21
CoCA [14]45.2146.1148.0250.33
IMMUNE (Ours)13.3214.9715.1816.02
LLaVA-1.5Original68.0067.1966.2367.33
FigStep [16]53.9852.0751.658.17
AdaShield [67]27.0024.1526.9130.10
CoCA [14]44.2341.8842.9144.82
IMMUNE (Ours)12.4516.7114.3217.03
MiniGPT-4-7BOriginal53.4058.1258.7961.23
FigStep [16]47.0952.4355.2557.10
AdaShield [67]39.0042.4543.7149.63
CoCA [14]53.1960.0962.7868.29
IMMUNE (Ours)18.4220.7721.1822.95
MiniGPT-4-13BOriginal62.5766.1367.8270.11
FigStep [16]59.8161.0562.4465.79
AdaShield [67]47.8250.3751.6853.15
CoCA [14]71.8574.3175.7275.09
IMMUNE (Ours)23.4125.8226.0226.02
Qwen-VLOriginal38.1148.0348.0349.47
FigStep [16]34.2536.9240.1742.78
AdaShield [67]17.3218.2022.4125.24
CoCA [14]38.1544.7346.3649.04
IMMUNE (Ours)13.2215.0715.0716.98
", + "image_path": "66922f099d29f539426345d32e3cdad746c29bc3c99d73e453c021dc232627b7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 510, + 555, + 555 + ], + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 555 + ], + "type": "text", + "content": "Table 4. Evaluation on Visual Adversarial Attacks [54]. We report ASR (in %) on images optimized with varying levels of adversarial noise, denoted by " + }, + { + "bbox": [ + 313, + 510, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 510, + 555, + 555 + ], + "type": "text", + "content": ", with \"unconstrained\" representing the most challenging scenario." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 572, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 572, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 572, + 388, + 586 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "In this work, we introduce IMMUNE, an inference-time defense framework designed to protect MLLMs against jailbreak attacks. IMMUNE employs a safety-aware reward model to align the MLLM during inference, with the goal of mitigating jailbreak vulnerabilities. We present a mathematical formulation of our defense framework, framing jailbreak mitigation as an alignment problem. Through extensive experiments, we demonstrate that IMMUNE consistently and significantly outperforms competitive baselines, enhancing model safety while preserving the model utility." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25045" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] Maksym Andriushchenko, Francesco Croce, and Nicolas Flammarion. Jailbreaking leading safety-aligned llms with simple adaptive attacks. arXiv preprint arXiv:2404.02151, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 296, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 296, + 234 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 296, + 234 + ], + "type": "text", + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 296, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 296, + 289 + ], + "type": "text", + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 335 + ], + "type": "text", + "content": "[5] James Beetham, Souradip Chakraborty, Mengdi Wang, Furong Huang, Amrit Singh Bedi, and Mubarak Shah. Liar: Leveraging alignment (best-of-n) to jailbreak llms in seconds. arXiv preprint arXiv:2412.05232, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 336, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 296, + 388 + ], + "type": "text", + "content": "[6] Nicholas Carlini, Anish Athalye, Nicolas Papernot, Wieland Brendel, Jonas Rauber, Dimitris Tsipras, Ian Goodfellow, Aleksander Madry, and Alexey Kurakin. On evaluating adversarial robustness. arXiv preprint arXiv:1902.06705, 2019. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 391, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 296, + 434 + ], + "type": "text", + "content": "[7] Souradip Chakraborty, Soumya Suvra Ghosal, Ming Yin, Dinesh Manocha, Mengdi Wang, Amrit Singh Bedi, and Furong Huang. Transfer q star: Principled decoding for llm alignment, 2024. 2, 3, 4, 16" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 296, + 479 + ], + "type": "text", + "content": "[8] Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J Pappas, and Eric Wong. Jailbreaking black box large language models in twenty queries. arXiv preprint arXiv:2310.08419, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 481, + 296, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 296, + 545 + ], + "type": "text", + "content": "[9] Patrick Chao, Edoardo Debenedetti, Alexander Robey, Maksym Andriushchenko, Francesco Croce, Vikash Sehwag, Edgar Dobriban, Nicolas Flammarion, George J. Pappas, Florian Tramer, Hamed Hassani, and Eric Wong. Jailbreakbench: An open robustness benchmark for jailbreaking large language models. In NeurIPS Datasets and Benchmarks Track, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 296, + 612 + ], + "type": "text", + "content": "[10] Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14239-14250, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 669 + ], + "type": "text", + "content": "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 57, + 613, + 296, + 669 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 57, + 613, + 296, + 669 + ], + "type": "text", + "content": "*chatgpt quality, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[12] Yinpeng Dong, Huanran Chen, Jiawei Chen, Zhengwei Fang, Xiao Yang, Yichi Zhang, Yu Tian, Hang Su, and Jun Zhu. How robust is google's bard to adversarial image attacks? arXiv preprint arXiv:2309.11751, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[13] Yanrui Du, Sendong Zhao, Ming Ma, Yuhan Chen, and Bing Qin. Analyzing the inherent response tendency of llms: Real-world instructions-driven jailbreak. arXiv preprint arXiv:2312.04127, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 172 + ], + "type": "text", + "content": "[14] Jiahui Gao, Renjie Pi, Tianyang Han, Han Wu, Lanqing Hong, Lingpeng Kong, Xin Jiang, and Zhenguo Li. Coca: Regaining safety-awareness of multimodal large language models with constitutional calibration. arXiv preprint arXiv:2409.11365, 2024. 1, 2, 3, 6, 7, 8, 15, 16, 17" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 175, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 554, + 207 + ], + "type": "text", + "content": "[15] Simon Geisler, Tom Wollschläger, M. H. I. Abdalla, Johannes Gasteiger, and Stephan Gunnemann. Attacking large language models with projected gradient descent, 2024. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 555, + 262 + ], + "type": "text", + "content": "[16] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via topographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 2, 3, 6, 7, 8, 15, 17, 18, 19" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 264, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 555, + 319 + ], + "type": "text", + "content": "[17] Jindong Gu, Zhen Han, Shuo Chen, Ahmad Beirami, Bailan He, Gengyuan Zhang, Ruotong Liao, Yao Qin, Volker Tresp, and Philip Torr. A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 320, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 554, + 354 + ], + "type": "text", + "content": "[18] Xingang Guo, Fangxu Yu, Huan Zhang, Lianhui Qin, and Bin Hu. Cold-attack: Jailbreaking Ilms with stealthiness and controllability. arXiv preprint arXiv:2402.08679, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 354, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 554, + 398 + ], + "type": "text", + "content": "[19] Dongchen Han, Xiaojun Jia, Yang Bai, Jindong Gu, Yang Liu, and Xiaochun Cao. Ot-attack: Enhancing adversarial transferability of vision-language models via optimal transport optimization. arXiv preprint arXiv:2312.04403, 2023. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 399, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 554, + 453 + ], + "type": "text", + "content": "[20] Seungju Han, Kavel Rao, Allyson Ettinger, Liwei Jiang, Bill Yuchen Lin, Nathan Lambert, Yejin Choi, and Nouha Dziri. Wildguard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of llms. arXiv preprint arXiv:2406.18495, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 456, + 554, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 554, + 488 + ], + "type": "text", + "content": "[21] Jonathan Hayase, Ema Borevkovic, Nicholas Carlini, Florian Tramér, and Milad Nasr. Query-based adversarial prompt generation. arXiv preprint arXiv:2402.12329, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 490, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 555, + 533 + ], + "type": "text", + "content": "[22] Zhang-Wei Hong, Idan Shenfeld, Tsun-Hsuan Wang, Yung-Sung Chuang, Aldo Pareja, James Glass, Akash Srivastava, and Pulkit Agrawal. Curiosity-driven red-teaming for large language models. arXiv preprint arXiv:2402.19464, 2024. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 554, + 578 + ], + "type": "text", + "content": "[23] MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learning for image captioning. ACM Computing Surveys (CsUR), 51 (6):1-36, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 580, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 554, + 612 + ], + "type": "text", + "content": "[24] Sandy Huang, Nicolas Papernot, Ian Goodfellow, Yan Duan, and Pieter Abbeel. Adversarial attacks on neural network policies. arXiv preprint arXiv:1702.02284, 2017. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 613, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 555, + 656 + ], + "type": "text", + "content": "[25] Yangsibo Huang, Samyak Gupta, Mengzhou Xia, Kai Li, and Danqi Chen. Catastrophic jailbreak of open-source llms via exploiting generation. arXiv preprint arXiv:2310.06987, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "type": "text", + "content": "[26] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Llm-based input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25046" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "text", + "content": "[27] Erik Jones, Anca Dragan, Aditi Raghunathan, and Jacob Steinhardt. Automatically auditing large language models via discrete optimization. In International Conference on Machine Learning, pages 15307-15329. PMLR, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 150 + ], + "type": "text", + "content": "[28] Maxim Khanov, Jirayu Burapacheep, and Yixuan Li. ARGs: Alignment as reward-guided search. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "type": "text", + "content": "[29] Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, et al. Rewardbench: Evaluating reward models for language modeling. arXiv preprint arXiv:2403.13787, 2024. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 296, + 262 + ], + "type": "text", + "content": "[30] Seanie Lee, Minsu Kim, Lynn Cherif, David Dobre, Juho Lee, Sung Ju Hwang, Kenji Kawaguchi, Gauthier Gidel, Yoshua Bengio, Nikolay Malkin, and Moksh Jain. Learning diverse attacks on large language models for robust red-teaming and safety tuning, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "text", + "content": "[31] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665, 2023. 15, 16" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 296, + 342 + ], + "type": "text", + "content": "[32] Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, and Qi Liu. Red teaming visual language models. arXiv preprint arXiv:2401.12915, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 296, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 296, + 385 + ], + "type": "text", + "content": "[33] Yuhui Li, Fangyun Wei, Jinjing Zhao, Chao Zhang, and Hongyang Zhang. Rain: Your language models can align themselves without finetuning. arXiv preprint arXiv:2309.07124, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 431 + ], + "type": "text", + "content": "[34] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 6, 17" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 433, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 296, + 465 + ], + "type": "text", + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.2,6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "type": "text", + "content": "[36] Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, et al. Agentbench: Evaluating llms as agents. arXiv preprint arXiv:2308.03688, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 296, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 296, + 543 + ], + "type": "text", + "content": "[37] Xin Liu, Yichen Zhu, Yunshi Lan, Chao Yang, and Yu Qiao. Query-relevant images jailbreak large multi-modal models, 2023. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 296, + 588 + ], + "type": "text", + "content": "[38] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models, 2024. 2, 7, 8, 15" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "text", + "content": "[39] Yi Liu, Gelei Deng, Zhengzi Xu, Yuekang Li, Yaowen Zheng, Ying Zhang, Lida Zhao, Tianwei Zhang, Kailong Wang, and Yang Liu. Jailbreaking chatgpt via prompt engineering: An empirical study. arXiv preprint arXiv:2305.13860, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "type": "text", + "content": "[40] Weidi Luo, Siyuan Ma, Xiaogeng Liu, Xiaoyu Guo, and Chaowei Xiao. Jailbreakv-28k: A benchmark for assessing the robustness of multimodal large language models against jailbreak attacks. arXiv preprint arXiv:2404.03027, 2024. 2, 3, 6, 7, 15, 16, 18, 20" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[41] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "Tu. Macaw-lm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 161 + ], + "type": "text", + "content": "[42] Neal Mangaokar, Ashish Hooda, Jihye Choi, Shreyas Chandrashekaran, Kassem Fawaz, Somesh Jha, and Atul Prakash. Prp: Propagating universal perturbations to attack large language model guard-rails. arXiv preprint arXiv:2402.15911, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 228 + ], + "type": "text", + "content": "[43] Ninareh Mehrabi, Palash Goyal, Anil Ramakrishna, Jwala Dhamala, Shalini Ghosh, Richard Zemel, Kai-Wei Chang, Aram Galstyan, and Rahul Gupta. Jab: Joint adversarial prompting and belief augmentation. In R0-FoMo: Robustness of Few-shot and Zero-shot Learning in Large Foundation Models. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 230, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 295 + ], + "type": "text", + "content": "[44] Ninareh Mehrabi, Ahmad Beirami, Fred Morstatter, and Aram Galstyan. Robust conversational agents against imperceptible toxicity triggers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2831-2847, 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 555, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 555, + 352 + ], + "type": "text", + "content": "[45] Ninareh Mehrabi, Palash Goyal, Christophe Dupuy, Qian Hu, Shalini Ghosh, Richard Zemel, Kai-Wei Chang, Aram Galstyan, and Rahul Gupta. Flirt: Feedback loop in-context red teaming. Empirical Methods in Natural Language Processing (EMNLP), 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 386 + ], + "type": "text", + "content": "[46] Youssef Mroueh. Information theoretic guarantees for policy alignment in large language models. arXiv preprint arXiv:2406.05883, 2024. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 555, + 453 + ], + "type": "text", + "content": "[47] Sidharth Mudgal, Jong Lee, Harish Ganapathy, YaGuang Li, Tao Wang, Yanping Huang, Zhifeng Chen, Heng-Tze Cheng, Michael Collins, Trevor Strohman, Jilin Chen, Alex Beutel, and Ahmad Beirami. Controlled decoding from language models. International Conference on Machine Learning (ICML), 2024. 2, 3, 4, 16" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 455, + 555, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 555, + 508 + ], + "type": "text", + "content": "[48] Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 510, + 555, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 555, + 543 + ], + "type": "text", + "content": "[49] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 544, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 555, + 609 + ], + "type": "text", + "content": "[50] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 612, + 555, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 555, + 644 + ], + "type": "text", + "content": "[51] Junsoo Park, Seungyeon Jwa, Meiying Ren, Daeyoung Kim, and Sanghyuk Choi. Offsetbias: Leveraging debiased data for tuning evaluators. arXiv preprint arXiv:2407.06551, 2024. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "text", + "content": "[52] Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia Glaese, Nat McAleese, and Geoffrey Irving. Red teaming language models with language models. arXiv preprint arXiv:2202.03286, 2022. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 555, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 715 + ], + "type": "text", + "content": "[53] Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25047" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "type": "text", + "content": "Mllm-protector: Ensuring mllm's safety without hurting performance. arXiv preprint arXiv:2401.02906, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 296, + 151 + ], + "type": "text", + "content": "[54] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 152, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 296, + 196 + ], + "type": "text", + "content": "[55] Xiangyu Qi, Ashwinee Panda, Kaifeng Lyu, Xiao Ma, Subhrajit Roy, Ahmad Beirami, Prateek Mittal, and Peter Henderson. Safety alignment should be made more than just a few tokens deep, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 296, + 251 + ], + "type": "text", + "content": "[56] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to! International Conference on Learning Representations (2024), 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 307 + ], + "type": "text", + "content": "[57] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 3, 4, 15, 16" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 309, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 309, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 57, + 309, + 296, + 364 + ], + "type": "text", + "content": "[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 365, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 365, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 57, + 365, + 296, + 409 + ], + "type": "text", + "content": "[59] Christian Schlarmann and Matthias Hein. On the adversarial robustness of multi-modal foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3677-3685, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 411, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 411, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 57, + 411, + 296, + 443 + ], + "type": "text", + "content": "[60] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Plug and pray: Exploiting off-the-shelf components of multi-modal models. arXiv preprint arXiv:2307.14539, 2023. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 445, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 445, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 57, + 445, + 296, + 477 + ], + "type": "text", + "content": "[61] Chawin Sitawarin, Norman Mu, David Wagner, and Alexandre Araujo. Pal: Proxy-guided black-box attack on large language models. arXiv preprint arXiv:2402.09674, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 479, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 479, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 479, + 296, + 521 + ], + "type": "text", + "content": "[62] Bharath K. Sriperumbudur, Kenji Fukumizu, Arthur Gretton, Bernhard Schölkopf, and Gert R. G. Lanckriet. On integral probability metrics, phi-divergences and binary classification, 2009. 13" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 523, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 296, + 566 + ], + "type": "text", + "content": "[63] Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul Christiano. Learning to summarize from human feedback, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 568, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 296, + 634 + ], + "type": "text", + "content": "[64] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing nlp. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2153-2162, 2019. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "text", + "content": "[65] Boxin Wang, Weixin Chen, Hengzhi Pei, Chulin Xie, Mintong Kang, Chenhui Zhang, Chejian Xu, Zidi Xiong, Ritik Dutta, Ryan Schaeffer, et al. Decodingtrust: A comprehensive assessment of trustworthiness in gpt models. In NeurlPS, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[66] Hao Wang, Hao Li, Minlie Huang, and Lei Sha. From noise to clarity: Unraveling the adversarial suffix of large language" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "model attacks via translation of text embeddings. arXiv preprint arXiv:2402.16006, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 98, + 555, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 555, + 152 + ], + "type": "text", + "content": "[67] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. arXiv preprint arXiv:2403.09513, 2024. 1, 2, 3, 6, 7, 8, 15, 16, 17" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 154, + 555, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 555, + 209 + ], + "type": "text", + "content": "[68] Zhilin Wang, Yi Dong, Olivier Delalleau, Jiaqi Zeng, Gerald Shen, Daniel Egert, Jimmy J Zhang, Makesh Narsimhan Sreedhar, and Oleksii Kuchaiev. Helpsteer2: Open-source dataset for training top-performing reward models. arXiv preprint arXiv:2406.08673, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 211, + 555, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 211, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 316, + 211, + 555, + 254 + ], + "type": "text", + "content": "[69] Fenghua Weng, Yue Xu, Chengyan Fu, and Wenjie Wang. MMJ-Bench: A comprehensive study on jailbreak attacks and defenses for vision language models. arXiv preprint arXiv:2408.08464, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 256, + 555, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 256, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 256, + 555, + 300 + ], + "type": "text", + "content": "[70] Nevan Wichers, Carson Denison, and Ahmad Beirami. Gradient-based language model red teaming. 18th Conference of European Chapter of the Assoc for Computational Linguistics (EACL), 2024. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 302, + 554, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 302, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 302, + 554, + 346 + ], + "type": "text", + "content": "[71] Qi Wu, Damien Teney, Peng Wang, Chunhua Shen, Anthony Dick, and Anton Van Den Hengel. Visual question answering: A survey of methods and datasets. Computer Vision and Image Understanding, 163:21-40, 2017. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 348, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 348, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 348, + 554, + 392 + ], + "type": "text", + "content": "[72] Nan Xu, Fei Wang, Ben Zhou, Bang Zheng Li, Chaowei Xiao, and Muhao Chen. Cognitive overload: Jailbreaking large language models with overloaded logical thinking. arXiv preprint arXiv:2311.09827, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 394, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 554, + 426 + ], + "type": "text", + "content": "[73] Yue Xu, Xiuyuan Qi, Zhan Qin, and Wenjie Wang. Defending jailbreak attack in vlms via cross-modality information detector. arXiv preprint arXiv:2407.21659, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 429, + 554, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 429, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 554, + 473 + ], + "type": "text", + "content": "[74] Zhangchen Xu, Fengqing Jiang, Luyao Niu, Yuntian Deng, Radha Poovendran, Yejin Choi, and Bill Yuchen Lin. Magpie: Alignment data synthesis from scratch by prompting aligned lms with nothing, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 475, + 554, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 554, + 519 + ], + "type": "text", + "content": "[75] Zonghao Ying, Aishan Liu, Tianyuan Zhang, Zhengmin Yu, Siyuan Liang, Xianglong Liu, and Dacheng Tao. Jailbreak vision language models via bi-modal adversarial prompt. arXiv preprint arXiv:2406.04031, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 521, + 554, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 554, + 586 + ], + "type": "text", + "content": "[76] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13807-13816, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 589, + 554, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 554, + 632 + ], + "type": "text", + "content": "[77] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 2, 16, 17" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 667 + ], + "type": "text", + "content": "[78] Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang. Rrrh: Rank responses to align language models with human feedback without tears, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[79] Yi Zeng, Hongpeng Lin, Jingwen Zhang, Diyi Yang, Ruoxi Jia, and Weiyan Shi. How johnny can persuade llms to jailbreak them: Rethinking persuasion to challenge ai safety by humanizing llms. arXiv preprint arXiv:2401.06373, 2024. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25048" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 620 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 106 + ], + "type": "text", + "content": "[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[81] Xiaoyu Zhang, Cen Zhang, Tianlin Li, Yihao Huang, Xiaojun Jia, Xiaofei Xie, Yang Liu, and Chao Shen. A mutation-based method for multi-modal jailbreaking attack detection. arXiv preprint arXiv:2312.10766, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 195 + ], + "type": "text", + "content": "[82] Zhuo Zhang, Guangyu Shen, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Make them spill the beans! coercive knowledge extraction from (production) llms. arXiv preprint arXiv:2312.04782, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 239 + ], + "type": "text", + "content": "[83] Xuandong Zhao, Xianjun Yang, Tianyu Pang, Chao Du, Lei Li, Yu-Xiang Wang, and William Yang Wang. Weak-to-strong jailbreaking on large language models. arXiv preprint arXiv:2401.17256, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 295, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 295 + ], + "type": "text", + "content": "[84] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 295, + 330 + ], + "type": "text", + "content": "[85] Yukai Zhou, Zhijie Huang, Feiyang Lu, Zhan Qin, and Wenjie Wang. Don't say no: Jailbreaking llm by suppressing refusal. arXiv preprint arXiv:2404.16369, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 373 + ], + "type": "text", + "content": "[86] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 430 + ], + "type": "text", + "content": "[87] Sicheng Zhu, Ruiyi Zhang, Bang An, Gang Wu, Joe Barrow, Zichao Wang, Furong Huang, Ani Nenkova, and Tong Sun. Autodan: interpretable gradient-based adversarial attacks on large language models. In First Conference on Language Modeling, 2024. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 432, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 295, + 475 + ], + "type": "text", + "content": "[88] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 1, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 476, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 295, + 520 + ], + "type": "text", + "content": "[89] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. arXiv preprint arXiv:2402.02207, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 521, + 295, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 295, + 564 + ], + "type": "text", + "content": "[90] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 566, + 295, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 295, + 620 + ], + "type": "text", + "content": "[91] Andy Zou, Long Phan, Justin Wang, Derek Duenas, Maxwell Lin, Maksym Andriushchenko, Rowan Wang, Zico Kolter, Matt Fredrikson, and Dan Hendrycks. Improving alignment and robustness with short circuiting. arXiv preprint arXiv:2406.04313, 2024. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25049" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/5a713a88-9e78-41c3-b9a4-4778994a6f99_content_list.json b/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/5a713a88-9e78-41c3-b9a4-4778994a6f99_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e77e663a0eeb3b4c94c30b6917f29352c33eac0d --- /dev/null +++ b/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/5a713a88-9e78-41c3-b9a4-4778994a6f99_content_list.json @@ -0,0 +1,1527 @@ +[ + { + "type": "text", + "text": "Implicit Bias Injection Attacks against Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 140, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huayang Huang $^{1}$ Xiangye Jin $^{2}$ Jiaxu Miao $^{3*}$ Yu Wu $^{1*}$ $^{1}$ School of Computer Science, Wuhan University \n $^{2}$ School of Mathematics and Statistics, Wuhan University \n $^{3}$ School of Cyber Science and Technology, Sun Yat-sen University", + "bbox": [ + 205, + 178, + 789, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hyhuang, xiangyejin, wuyucs}@whu.edu.cn miaojx@mail.sysu.edu.cn", + "bbox": [ + 207, + 261, + 784, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For developing countries", + "text_level": 1, + "bbox": [ + 176, + 301, + 375, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-Image", + "text_level": 1, + "bbox": [ + 431, + 313, + 560, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For developed countries", + "text_level": 1, + "bbox": [ + 619, + 301, + 810, + 316 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7463dc4505728574c63eea2674a0a920eeac47b58c03d3492a42e0ace17ffaf1.jpg", + "image_caption": [ + "A doctor", + "Figure 1. Implicit bias injection causes the model to generate subtly negative content (left) and positive content (right) for different users. Although the generated images meet the text requirement, the subtle differences may, over time, influence user perception." + ], + "image_footnote": [], + "bbox": [ + 106, + 363, + 235, + 561 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2d1608fba693f0027435918a98db739fd8e4ab56af1a0945fb4a195925f2e8c.jpg", + "image_caption": [ + "A professor" + ], + "image_footnote": [], + "bbox": [ + 233, + 364, + 359, + 561 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/befa5ac40f50c8ab9f6cd98ae70cdf4865a3c3f8732dbadd15055552fb0b67ce.jpg", + "image_caption": [ + "A city park" + ], + "image_footnote": [], + "bbox": [ + 357, + 364, + 488, + 561 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa671134f618e2e6f2a831ce8c6fdf61dd69a66e4904456482f79e93a8134f9d.jpg", + "image_caption": [ + "A doctor" + ], + "image_footnote": [], + "bbox": [ + 511, + 364, + 637, + 560 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ace4558e8f1814173f8d08e828394ee47fddf53d42b3fa35be4351ebc852014.jpg", + "image_caption": [ + "A professor" + ], + "image_footnote": [], + "bbox": [ + 637, + 364, + 763, + 560 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7fcb3437fd40df3fce6e187abbf577f1e50438e19dc4a5e3429614e58d3ce77f.jpg", + "image_caption": [ + "A city park" + ], + "image_footnote": [], + "bbox": [ + 759, + 364, + 888, + 560 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 618, + 326, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The proliferation of text-to-image diffusion models (T2I DMs) has led to an increased presence of AI-generated images in daily life. However, biased T2I models can generate content with specific tendencies, potentially influencing people's perceptions. Intentional exploitation of these biases risks conveying misleading information to the public. Current research on bias primarily addresses explicit biases with recognizable visual patterns, such as skin color and gender. This paper introduces a novel form of implicit bias that lacks explicit visual features but can manifest in diverse ways across various semantic contexts. This subtle and versatile nature makes this bias challenging to detect, easy to propagate, and adaptable to a wide range of scenarios. We further propose an implicit bias injection attack framework (IBI-Attacks) against T2I diffusion models by precomputing", + "bbox": [ + 88, + 650, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a general bias direction in the prompt embedding space and adaptively adjusting it based on different inputs. Our attack module can be seamlessly integrated into pre-trained diffusion models in a plug-and-play manner without direct manipulation of user input or model retraining. Extensive experiments validate the effectiveness of our scheme in introducing bias through subtle and diverse modifications while preserving the original semantics. The strong concealment and transferability of our attack across various scenarios further underscore the significance of our approach.", + "bbox": [ + 511, + 619, + 908, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 799, + 645, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image (T2I) diffusion models [19, 34, 47-49, 65] have emerged as a powerful tool for generating highly realistic images that closely adhere to textual descriptions. The widespread adoption of these models, facilitated by public APIs [40, 54] and open-source implementations [32, 33],", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 114, + 887, + 241, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28779", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "has led to a proliferation of AI-generated images in daily life. However, biased T2I models can generate content with a specific tendency [7, 33, 58, 64], posing a risk of conveying misleading information that could influence public perception if maliciously exploited [2, 55].", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consider a scenario in which a malicious T2I service provider subtly \"manipulates\" user perception by selectively generating biased content. As illustrated in Fig. 1, the same neutral prompt can yield images with varying emotional tones depending on the target audience. Content aimed at developed countries may convey a more positive tone, while content targeted at developing countries may lean towards a more negative sentiment. Such nuanced differences are difficult to detect visually, yet users may be subconsciously influenced by prolonged exposure to biased images. Furthermore, as shown in Fig. 2, such biases can extend beyond emotional aspects to include cultural and religious dimensions. This implicit bias, if intentionally leveraged, could subtly shape user perceptions over time, reinforcing existing divides and nudging audiences into self-reinforcing emotional or ideological spheres.", + "bbox": [ + 89, + 169, + 482, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methods for bias exploitation typically require expensive model fine-tuning [38] and are limited to explicit biases [20, 41] (e.g., black skin and shaved head) triggered by specific inputs, which are easily detectable. In contrast, this paper focuses on a more implicit form of bias, such as the emotion, cultural stereotypes, and religious orientation depicted in Fig. 1 and Fig. 2. Such bias is highly covert and can manifest through diverse visual representations under varying prompt inputs. By imposing \"hints\" without significantly changing the content, users can be unconsciously influenced in ways that are difficult to detect. The manifestation of the implicit bias is context-dependent and can be conveyed through multiple semantic forms, such as facial expressions, poses, backgrounds, or group behaviors.", + "bbox": [ + 89, + 412, + 482, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose an implicit bias injection framework (IBI-Attacks) by precalculating a bias direction in the prompt embedding space and then dynamically adapting it based on specific user inputs. By introducing bias at the prompt embedding space, we can achieve more fine-grained control compared to traditional prompt modification. And input-aware dynamic adjustment further enables adaptive bias expression across diverse semantic contexts. Specifically, we leverage a large language model (LLM) [1, 5, 45] to generate a list of neutral prompts and rewrite them according to the specified bias. Then a bias direction vector is calculated by taking the difference between the two embedding sets. Our key finding is that this single bias direction already encodes multiple semantic expressions and can generalize across various prompt inputs by simple addition, allowing for diverse bias expressions.", + "bbox": [ + 89, + 627, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, directly applying this fixed direction vector to all prompts may result in either excessive or insufficient bias", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fdf3856899a3189bfa1f9d710fd6408f350ca787b8463683d8441f3d9250d6e6.jpg", + "image_caption": [ + "\"A man is standing in front of his house and smiling.\"" + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 176 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4387b18ad5f237dddd822c9d81f13acfa233d16af663a72b5557d5183a90bf82.jpg", + "image_caption": [ + "\"One person worships in a room.\"", + "\"Two people talk at the table.\"", + "Figure 2. Generated images after injecting different 'cultural', 'religious', and 'gender' biases with IBI attacks. Explicit biases, such as gender, can also benefit from our approach, enabling subtle, diverse, and adaptive gender-related features." + ], + "image_footnote": [], + "bbox": [ + 516, + 189, + 903, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "modification. To address this limitation, we further perform dynamic feature selection in the fixed bias direction based on the input prompt, enabling proper expressions in different semantic contexts. In this way, the attacker can implant bias into any pre-trained text-to-image diffusion model in a plug-and-play manner, without the need to modify the input prompt or fine-tune the model.", + "bbox": [ + 511, + 386, + 905, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our approach on Stable Diffusion [48], using \"emotion bias\" as a case study to demonstrate the accuracy of bias injection, the diversity of bias expression, and the preservation of original semantics. The transferability of the bias direction further makes our attack applicable to a wide range of user inputs, allowing for consistent bias implantation across diverse contexts.", + "bbox": [ + 511, + 493, + 905, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our key contributions are as follows.", + "bbox": [ + 529, + 599, + 859, + 614 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We consider a new type of implicit bias with subtle and multiple forms of semantic expression, which is more stealthy, challenging to detect, and easy to propagate.", + "- We find that a specific bias direction within the text embedding in T2I diffusion models already encodes multiple expressions and exhibits generalization ability.", + "- We introduce an adaptive adjusting module to dynamically express the bias direction on different inputs." + ], + "bbox": [ + 511, + 616, + 903, + 736 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 753, + 653, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Bias in Diffusion Models. Existing research [2, 3, 25, 33, 35, 53] has demonstrated the prevalence of bias in text-to-image diffusion models through extensive evaluations. Dall-Eval [9] reveals that both DALL-E [37] and Stable Diffusion [48] exhibit a tendency to generate male figures, with varying gender biases across different occupations. One study [39] also finds that over $80\\%$ detainee-related images generated by Stable Diffusion depicted dark-skinned indi", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28780", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "viduals, despite people of color constituting less than half of the U.S. prison population [13]. Such bias can lead models to perpetuate false representations of the world when used in real-world applications. Considering this, some studies have explored detecting [8, 11, 22], evaluating [16, 56, 57], and mitigating [43, 51, 63] bias in generated content to promote fairness and diversity. However, these works primarily concentrate on traditional biases like gender and skin tone, neglecting bias types with diverse semantic expressions, such as emotions considered in this paper.", + "bbox": [ + 89, + 90, + 480, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLMs in Diffusion Models. Considering the exceptional language understanding and generation capabilities of LLMs [1, 5], several studies [14, 18, 31] have integrated LLMs into the text-to-image diffusion generation process to accomplish complex tasks. A series of studies utilize LLMs to guide diffusion models for multi-stage generation for better accuracy. For example, SLD [61] employs LLMs to achieve the cyclic self-correctio of the generation process for improved alignment with complex prompts. LMD [26] utilizes LLMs to generate initial layouts based on text prompts, followed by layout-conditioned generation for further refinement. Another research direction [18, 31, 36] involves integrating rich text features learned by LLMs into the diffusion model to enhance guidance capabilities. In this paper, we employ LLMs to learn a multi-semantic direction of implicit bias within the text embedding space.", + "bbox": [ + 91, + 263, + 483, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bias Exploitations. In addition to investigating inherent biases in diffusion models, some studies have explored exploiting or implanting biases to deliberately steer the generation process. Some works [50, 52] show that specific hard prompts can induce biased output. However, these methods rely on existing biases within the model and cannot achieve arbitrary bias specifications. Some bias mitigation methods [10, 21] also utilize text embedding modification, but they calculate a fixed, single semantic direction for all prompts. Most similar to ours is a concurrent work Backdoorsing Bias [38], which aims to inject specified biases into the model. However, it can only introduce explicit biases with fixed visual features through prompt modification. Additionally, it requires expensive model fine-tuning. In contrast, our attack works in a training-free manner and can adaptively implant implicit bias for any user input.", + "bbox": [ + 89, + 527, + 482, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Threat Model", + "text_level": 1, + "bbox": [ + 89, + 784, + 232, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attack's Objective. The attacker aims to subtly alter the T2I model to embed implicit bias into generated content while maintaining the original output structure. To achieve this, we define two primary goals: 1) Attack Success. The attackers need to ensure that the implicit bias is successfully embedded in the model, meaning that all the images gener", + "bbox": [ + 89, + 809, + 480, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ated by the model consistently exhibit the attacker-specified bias, with a high success rate and a subtle, implicit effect. 2) Injection Stealth. To maintain stealth in our attack, we aim to preserve the model's original generation content while allowing for the injection of specified biases. The generated content should align with user prompts but include subtle, unnoticed details that convey the bias. High semantic and structural similarity to normal outputs makes the attack difficult to detect.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attacker's Capability. Our approach allows the adversary to operate without needing access to the T2I model's parameters, structure, or the user-input prompts. The attack can be achieved by model poisoning in the following scenarios: 1) Covert Module Injection: An insider could introduce a stealthy module into the model during its development phase. 2) API Exploitation: A service provider might add the module to their API, subtly altering the results for users without informing them. 3) Model Sharing: An adversary could distribute a modified version of the T2I model on platforms where users expect high-quality models, manipulating outputs for unsuspecting users.", + "bbox": [ + 511, + 250, + 903, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike LLM-assisted prompts, which require real-time API calls during inference and are easier to detect, our attack operates autonomously after deployment without additional LLM calls, making it more covert and efficient. Additionally, LLM-assisted prompts also cause significant semantic alterations.", + "bbox": [ + 511, + 431, + 903, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Implicit Bias Injection Attacks", + "text_level": 1, + "bbox": [ + 511, + 537, + 795, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the details of our original approach for implanting implicit bias in T2I diffusion models. The pipeline of the proposed framework is shown in Fig. 3. Our attack consists of three main stages: bias direction calculation, adaptive feature selection, and user inference. We first use a LLM to establish the relationship between implicit bias and multiple semantic expressions by rewriting sentences according to a specific bias. We then compute the bias direction vector by calculating the average difference between the original and rewritten sentences in the prompt embedding space. After that, we train an input-dependent adaptive feature selection module, which dynamically adjusts the fixed bias direction from the first stage based on the user's input prompt. Finally, during model inference, the adaptive adjustment module implants the bias into the user's prompt embedding before it is sent to the generative model, ensuring that the model's output reflects the specified bias.", + "bbox": [ + 511, + 564, + 906, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Directional Vector Generation", + "text_level": 1, + "bbox": [ + 511, + 847, + 782, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we aim to compute a difference vector representing the direction of bias in the latent space. The attacker first", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28781", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/eedf888a5a8244b696d076dcd5187ff99958c44df890d63b7dabd1af770dd460.jpg", + "image_caption": [ + "User inference", + "Figure 3. Pipeline of our proposed IBI-Attacks. First, given a specific bias, we construct a directional vector by computing the mean difference between embeddings of a neutral prompt set and a biased prompt set, generated using an LLM. Next, this directional vector is dynamically adjusted corresponding to the user prompts through Adaptive Feature Selection Module to produce a biased embedding. Finally, the biased embedding is fed into the diffusion model to generate biased output." + ], + "image_footnote": [], + "bbox": [ + 109, + 104, + 906, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "implements an LLM model to generate $N$ neutral prompts to create a neutral prompt set $X_{\\mathrm{neu}} = \\{x_i^{\\mathrm{neu}}\\}_{i=1}^N$ , each of which has a standard subject-verb-object structure. Meanwhile, the LLM model rewrites the neutral text prompts in $X_{\\mathrm{neu}}$ with a specific bias to form a biased prompt dataset: $X_{\\mathrm{bias}} = \\{x_i^{\\mathrm{bias}}\\}_{i=1}^N$ . We employ a rewriting strategy that prefixes nouns with appropriate adjectives to introduce specific biases. We find that this restricted rewriting approach minimizes the introduction of bias-irrelevant information at the embedding level, such as modifications to sentence structure. As a result, the offset between the rewritten and original prompt embeddings becomes more closely aligned with the intended bias. Additionally, the addition of adjectives helps preserve the original sentence semantics, preventing the destruction of the underlying content.", + "bbox": [ + 89, + 503, + 483, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterwards, the attacker utilizes the pretrained encoder $\\varphi(\\cdot)$ to map input prompt $x$ into text embeddings $v = \\varphi(x) \\in R^{D \\times L}$ , where $D$ denotes the dimension of each token in latent space, and $L$ denotes the number of embedded tokens for each text prompt (i.e., $D = 1024$ , $L = 77$ in the SD v2.1 model). With the text encoder, texts prompts from $X_{\\mathrm{neu}}$ and $X_{\\mathrm{bias}}$ are mapped into the latent space to obtain their representation in latent space:", + "bbox": [ + 89, + 733, + 483, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i} ^ {\\mathrm {n e u}} = \\varphi (x _ {i} ^ {\\mathrm {n e u}}), v _ {i} ^ {\\mathrm {b i a s}} = \\varphi (x _ {i} ^ {\\mathrm {b i a s}}), i = 1, 2, \\dots , N. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 859, + 482, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, we get two sets of text embedding: $\\{v_i^{\\mathrm{neu}}\\}_{i = 1}^N$", + "bbox": [ + 89, + 883, + 480, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $\\{v_i^{\\mathrm{bias}}\\}_{i = 1}^N$ . We then compute the mean difference between these two sets of representations as:", + "bbox": [ + 511, + 501, + 906, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv ^ {\\text {d i f f}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(v _ {i} ^ {\\text {b i a s}} - v _ {i} ^ {\\text {n e u}}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 545, + 906, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We find that this simple mean difference vector already encodes multiple semantics of the implicit bias and can be generalized across different prompts. The existence of this expressive average bias direction enables us to embed implicit bias in the output by simply adding a fixed vector to the input prompt embedding.", + "bbox": [ + 511, + 595, + 906, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Adaptive Feature Selection", + "text_level": 1, + "bbox": [ + 511, + 694, + 756, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the single average direction vector can already introduce some bias, we find that directly applying it to all prompts may result in either excessive or insufficient bias. Therefore, we introduce an adaptive adjustment module to dynamically alter the precomputed bias direction vector $v^{\\text{diff}}$ according to the specific user input $v^{\\text{user}}$ .", + "bbox": [ + 511, + 715, + 905, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Module Structure. Drawing inspiration from SENet [17], we design a simple module that computes an attention map on the bias direction vector for feature selection. For the text embedding vectors with dimension $D \\times L$ , we alternately compute feature attention", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28782", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for the token dimension $L$ and the embedding dimension $D$ . Specifically, we alternately reduce one dimension to a single value using global average pooling $\\mathrm{Avg}$ to capture the overall information of that dimension. A two-layer multilayer perceptron $\\mathrm{MLP}_{\\theta}$ is then used to learn attention for the other, uncompressed dimension, which is used to adaptively scale the original feature along that axis. $\\theta$ is the learnable parameter of MLP.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By prioritizing key information, we can focus on relevant features and disregard irrelevant ones based on different prompt inputs. Specifically, we compute attention across both the token and feature dimensions of the text embedding by compressing the other dimension separately. The average direction vector $v^{\\text{diff}}$ is then scaled according to the attention values in both dimensions, producing an adaptive bias direction $\\tilde{v}^{\\text{diff}}$ that is added to the user input prompt $v^{\\text{user}}$ to generate the biased embedding $\\tilde{v}^{\\text{biased}}$ .", + "bbox": [ + 89, + 212, + 483, + 348 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {v} ^ {\\text {d i f f}} = \\operatorname {M L P} _ {\\theta} (\\operatorname {A v g} (v ^ {\\text {u s e r}})) \\odot v ^ {\\text {d i f f}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 358, + 482, + 376 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {v} ^ {\\text {b i a s}} = v ^ {\\text {u s e r}} + \\tilde {v} ^ {\\text {d i f f}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 401, + 480, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Paradigm. We reuse the neutral and biased prompt embedding sets generated by LLM when calculating the bias direction to train the adaptive module. We rearranged them in pairs: $\\{v_i^{\\mathrm{neu}}, v_i^{\\mathrm{biased}}\\}_{i=1}^{N}$ as input and label in training the adaptive module.", + "bbox": [ + 89, + 436, + 483, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The adaptive module dynamically fine-tunes the average bias direction $v_{i}^{\\text{diff}}$ based on the neutral prompt embedding $v_{i}^{\\text{neu}}$ , enabling the adjusted direction to point from the neutral prompt embedding $v_{i}^{\\text{neu}}$ toward the biased prompt embedding $v_{i}^{\\text{biased}}$ . Thus we define the following loss:", + "bbox": [ + 89, + 515, + 483, + 592 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL o s s = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| v _ {i} ^ {\\text {d i f f}} - \\operatorname {M L P} _ {\\theta} \\left(\\operatorname {A v g} \\left(v _ {i} ^ {\\text {n e u}}\\right)\\right) \\odot v ^ {\\text {d i f f}} \\right\\| ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 603, + 480, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $v_{i}^{\\text{diff}} = v_{i}^{\\text{biased}} - v_{i}^{\\text{neu}}$ .", + "bbox": [ + 89, + 656, + 302, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. User Inference", + "text_level": 1, + "bbox": [ + 89, + 681, + 240, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the adaptation module is trained, it can be seamlessly inserted behind the text encoder of a pre-trained diffusion model. This model can be selectively deployed to specific groups. When users download the biased model, they interact with it by entering a text prompt input. This prompt is processed by the encoder and then the adaptive module to generate the biased version. Finally, the modified embedding is fed into the T2I model, producing a final output image with the desired bias. Due to the subtlety of the bias introduction and its general applicability across different prompts, the model can continuously generate imperceptibly biased content for the target user, potentially influencing their cognition. An another harmful scenario could", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "involve API service providers selectively targeting specific IP addresses, delivering normal results to most users while subtly implanting biases for particular groups, thereby conveying misleading information. The plug-and-play nature of our attack makes such targeted attacks easier to implement, posing a serious risk of spreading harmful bias.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 198, + 645, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Settings", + "text_level": 1, + "bbox": [ + 511, + 224, + 611, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models and Datasets. Our evaluation is conducted on the COCO dataset [27] and Stable Diffusion 2.1 [48]. From the COCO dataset, we filtered 200,000 caption subsets related to people, 57,000 caption subsets related to animals, and 60,000 caption subsets related to natural scenes, ensuring that these categories do not overlap. Our primary focus is on captions related to human scenes, as this represents a common scenario for bias attacks. Furthermore, we observe that the bias embedding module, trained on human scenes, exhibits zero-shot transferability to other scenes, such as those involving animals and nature.", + "bbox": [ + 511, + 247, + 906, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. To calculate the average bias direction, we employed ChatGPT-4 [42] to generate 200 neutral prompts related to everyday person scenes, along with their biased counterparts. Given that minor text prompt modifications can significantly impact embeddings [60, 62], and unrestricted rewriting rules may alter sentence structure, we simplified our rewriting rules to focus solely on the addition of appropriate adjectives to the original sentence. This approach minimizes the influence of bias-irrelevant factors on text embeddings. The adaptive adjusting module was trained on the same LLM-generated data for 50 epochs using the Adam optimizer [23] with a learning rate of 0.001.", + "bbox": [ + 511, + 436, + 908, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We mainly consider two metrics: the success rate of bias injection and the stealthiness of the implanted implicit bias. Given the subtle and diverse semantic expressions of bias in our attacks, simple visual classification or recognition models are insufficient for detection. Therefore, following existing approaches [11, 38], we employ a Multi-modal Language Model (MLLM) to identify the presence of bias in the final generated image by question-answering based on visual content. We utilize the latest LLaVA-v1.6 7B [28-30] model in our experiments.", + "bbox": [ + 511, + 642, + 906, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the imperceptibility of bias implantation, we use the CLIP score [46] to measure the alignment between the generated image and the input text $(\\mathrm{CLIP}_{\\mathrm{txt - img}})$ , as well as the semantic similarity to the original generated image $(\\mathrm{CLIP}_{\\mathrm{img - img}})$ . Additionally, we use SSIM (Structural Similarity Index) [59] to assess pixel-level differences between the images before and after bias injection. Validation", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28783", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/85a5ade758f06974c66f8930adf8d4998e32c687d2846b80aafde289751a9c5e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Bias typeMethodsNegativePositiveSameCLIPtxt-Img↑CLIPimg-Img↑SSIM↑FID↓PickScore↑
Original-14.4%6.7%78.9%0.36331.00001.000039.29721.805
Negative biasExplicit89.1%6.9%4.0%0.34580.73540.522140.09021.386
IBI w/o adap82.2%12.5%5.3%0.35960.81430.588039.21721.593
IBI80.2%12.8%7.0%0.36370.87930.698739.13521.766
Positive biasExplicit0.0%98.0%2.0%0.34490.76960.522142.34321.763
IBI w/o adap13.9%85%1.1%0.35810.80300.577639.54921.606
IBI14.6%83.7%1.7%0.36020.82470.604039.49921.670
", + "bbox": [ + 133, + 88, + 859, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. MLLM evaluation of “Negative” and “Positive” bias injection. Original displays the results of splicing two non-attacked, normally generated images. Explicit shows the results of introducing bias by directly modifying user prompts. IBI w/o adap illustrates the impact of directly applying the average direction vector. IBI presents the results obtained using the adaptive module for the average direction vector.", + "bbox": [ + 89, + 238, + 906, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ab57f7efa3aa6d8deae47d0f67452d809f5802cd04ddcc10ed17c79aa0dac528.jpg", + "image_caption": [ + "Figure 4. Generated samples after injecting positive and negative bias. The key modifications are enlarged. Emotional bias is conveyed through small adjustments in a person's expression or posture. In the second column, after injecting positive bias, the girl's hands transition from a defensive crossed position to a relaxed stance, and a smile appears on her face. While for the negative mood, the girl maintains her original pose, but her facial expression becomes more serious." + ], + "image_footnote": [], + "bbox": [ + 91, + 294, + 906, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "was performed on a randomly selected set of 1,000 neutral human-related captions. We also employ FID [15] and PickScore [24] to assess the preservation of image quality, evaluated on 5,000 generated images.", + "bbox": [ + 89, + 656, + 482, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Performance Evaluation", + "text_level": 1, + "bbox": [ + 89, + 727, + 313, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. For the \"negative\" and \"positive\" sentiment bias, we instruct the MLLM to classify the modified image as more positive, more negative, or the same as the original image. The success rate was measured by the proportion of \"more negative\" for negative bias and \"more positive\" for positive bias. The results in Tab. 1 indicate that the average bias direction can already introduce bias into the image, increasing the \"More negative\" rate for \"negative\" bias from $14.4\\%$ to $82.2\\%$ . However, uniformly modifying all prompts without discrimination leads to a substantial", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "degradation of image quality. In contrast, the adaptive adjusting module effectively preserves the original semantics while maintaining the ability to introduce bias. The results are similar for the injection of \"positive\" bias. Although prompt modification achieves the highest bias implantation rate, it substantially alters the image content and exhibits poor invisibility. The small perturbations observed in FID and PickScore indicate that IBI attacks have minimal impact on image quality.", + "bbox": [ + 511, + 656, + 906, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 4 presents the generated images embedded with negative and positive biases, where each emotion is conveyed through subtle modifications, further illustrating the concealment of our attack. Fig. 5 shows the reflection of negative bias under different prompts. The bias expression manifests in various semantic forms, en", + "bbox": [ + 511, + 809, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28784", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3e26cb44d8a99f6f2229362937d492b1f6de93d99bc3bace568d8842981ea3cb.jpg", + "image_caption": [ + "Figure 5. Diverse semantic expressions of negative bias under different text inputs, including facial expressions, posture, style, and environmental cues." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 315 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compassing facial expressions, human posture, and subtle adjustments to background objects. Directly applying the direction vector may result in too much changes (the first column) or additional content (the third column). By incorporating the adaptive module, we can subtly introduce bias while preserving the original semantics.", + "bbox": [ + 89, + 397, + 483, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot Transferability. We further investigate the zero-shot transferability of the bias injection module trained on human scenes to other domains. The results presented in Tab. 2 demonstrate that the bias learned from human-related scenarios effectively transfers to both animal and natural scenes. In animal-related generation, the attack successfully induces a negative sentiment in $95.1\\%$ of cases. In natural scene generation, the negative probability increases by $81.1\\%$ . The results for 'positive' bias are similar. Fig. 6 visualizes the effects of this transfer attack. In animal-related generations, the bias manifests primarily through changes in the animal's pose and environment. For natural scene generations, the bias is expressed through more depressing backgrounds and less pleasing content.", + "bbox": [ + 89, + 502, + 483, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human Study. However, relying solely on the MLLM for evaluation may not fully capture the impact of subtle image detail differences on human perception, which is crucial for implicit bias. Therefore, we designed a human study to further evaluate the subtlety and effectiveness of our attack. We invited 24 participants to choose between two randomly shuffled images (original image and bias injected version) based on two criteria: (1) Abnormal Rate: \"Which image is more likely to contain an implanted bias?\"; and (2) Bias Injection Rate: \"Which image more effectively reflects a specified bias?\". We also offered a \"Not Sure\" option if", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7add2e72b6fef83b2a186cd1351dd3c7802696f62d4c7ac7d789a7547e8d6a48.jpg", + "image_caption": [ + "Figure 6. The generated samples of attacking animals and natural scenes with the 'negative' bias injection module learned on human scenes." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 903, + 407 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ef58d02d53428c3b6eff0bb5235576ad0e4aeb9ec365531b0e8d831fa7d3be9a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypesMethodsNegPosSameCLIPt-iCLIPi-iSSIM
P→AOriginal9.2%1.2%89.6%0.36411.00001.0000
IBI (Neg)95.1%4.9%0.0%0.35860.86440.6751
IBI (Pos)3.6%96.2%0.2%0.35670.85170.6147
P→NOriginal8.1%3.4%88.5%0.35611.00001.0000
IBI (Neg)89.2%10.8%0.0%0.35360.88470.6811
IBI (Pos)10.2%89.4%0.4%0.34820.82740.5836
", + "bbox": [ + 514, + 473, + 910, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Performance of zero-shot attacks on other scenes using bias injection module learned from the person scene. $\\mathrm{P}\\rightarrow \\mathrm{A}$ : Person to Animal. $\\mathrm{P}\\rightarrow \\mathrm{N}$ : Person to Nature.", + "bbox": [ + 511, + 603, + 906, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "participants think the two images are difficult to distinguish. Each person was randomly assigned 20 pairs of images.", + "bbox": [ + 511, + 672, + 903, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results of user selections are presented in Tab. 3. Our method achieves a greater degree of stealthiness, with users exhibiting only a $35.8\\%$ probability of perceiving an image as biased, which is nearly identical to the $35.7\\%$ of normal images. Participants have a $67.3\\%$ probability of experiencing negative emotions from the attacked images. Notably, the increase in \"Not Sure\" responses also underscores the effectiveness of our approach, as most of the model's originally generated images are perceived as positive.", + "bbox": [ + 511, + 703, + 906, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Extended Analysis", + "text_level": 1, + "bbox": [ + 511, + 848, + 692, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLM Rephrase Rules. We evaluate the impact of different LLM rewriting strategies in the bias vector calculation", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28785", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ae4eed387b3b132f64af74ef2cbdc3b3ae30c24ee0e94fdf1e274b4b59042a17.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Human picks as abnormal
MethodsInjected ↓OriginalNot Sure
Explicit57.5%22.5%20.0%
IBI w/o adapt47.5%25.0%27.5%
IBI35.8%35.7%28.5%
", + "bbox": [ + 107, + 88, + 467, + 178 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/276ce480a87f31c71e478a7ee03fdf779e15649284a4f91c998fc7250e5974e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Human picks as negative
MethodsInjected ↑OriginalNot Sure
Explicit75.5%17.5%5.0%
IBI w/o adapt72.8%20.6%6.6%
IBI67.3%8.9%23.8%
", + "bbox": [ + 107, + 183, + 467, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/42ba6c4bc2ee90368480b47e43e9110dd2047d4b7e56017fa313c93c70a650bb.jpg", + "image_caption": [ + "Figure 7. Bias vector learned by various LLM rephrase rules. $LLM_{adj}$ : directly using LLM to add adjectives in front of all nouns. $\\Delta LLM_{random}$ : LLM rewriting without restrictions and then calculating the average direction. $\\Delta LLM_{adj}$ : LLM rewriting with added adjectives and calculating the average direction." + ], + "image_footnote": [], + "bbox": [ + 96, + 316, + 478, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "phase. The sample results are shown in Fig. 7. If we directly employ the LLM rewritten prompts $(LLM_{\\mathrm{adj}})$ , the generated images can deviate significantly from the original intent. Moreover, highly negative prompts may result in numerous low-quality, easily detectable gray images. Alternatively, if we leverage unrestricted LLM rewriting $(\\Delta LLM_{\\mathrm{random}})$ , the calculated average bias direction may lead to image distortions. Because unrestricted rewriting may introduce extraneous differences between the two prompt sets, such as sentence structure, hindering the accurate representation of meaningful bias modifications by calculating the average direction. In our approach, a simple approach of adding adjectives $(\\Delta LLM_{\\mathrm{adj}})$ effectively minimizes the introduction of bias-irrelevant information, leading to the successful learning of a direction that accurately represents the bias.", + "bbox": [ + 88, + 657, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9fa69868b967722e40d2937f06e77ef2dc0140d9b707505ff88d81ca81bc5ff9.jpg", + "table_caption": [ + "Table 3. Results of human evaluation." + ], + "table_footnote": [], + "table_body": "
ModelAttackNeg ↑PosSameCLIPt-i ↑PS ↑
SD3 [12]-12.4%12.0%75.6%0.35722.529
SD3 [12]IBI68.2%11.8%20.0%0.35922.467
FLUX [4]-8.0%21.2%70.8%0.34322.354
FLUX [4]IBI65.4%34.4%0.2%0.34222.335
MUSE [6]-10.8%16.6%72.6%0.32420.553
MUSE [6]IBI91.2%8.6%0.2%0.32220.477
", + "bbox": [ + 516, + 88, + 898, + 213 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9ad12a0edbb9dd49d2d11960af216ad6bbf697549a7b69e4d69d987549a5eb86.jpg", + "table_caption": [ + "Table 4. MLLM evaluation of 'negative' bias injection for Stable Diffusion-3, FLux-schnell, and Muse models. PS: PickScore." + ], + "table_footnote": [], + "table_body": "
DefenseNeg ↑PosSameCLIPt-i ↑PS ↑
-80.2%12.8%7.0%0.36421.766
Chuang et al. [10]78.2%14.0%7.8%0.35821.419
Kim et al. [21]79.6%16.4%4.0%0.13918.523
", + "bbox": [ + 522, + 267, + 893, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Performance of IBI attacks under debiasing methods.", + "bbox": [ + 524, + 349, + 892, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization to Latest Models. We further evaluated negative bias injection on Stable Diffusion-3 [12], Flux-schnell (rectified flow transformer) [4], and MUSE (autoregressive) [6]. The results presented in Tab. 4 demonstrate that IBI generalizes effectively across models with varying architectures, significantly increasing the proportion of negative images while preserving the original semantics and maintaining the model's generative capabilities.", + "bbox": [ + 511, + 388, + 906, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness against Debiasing Methods. We evaluated the effectiveness of our attack after applying two debiasing methods [10, 21]. The approach by Chuang et al. [10] works by projecting out certain biased word directions in the text embedding, while Kim et al. [21] achieves bias control by modifying the end-of-sequence token. Tab. 5 shows they have little defense effect, as word-level debiasing is ineffective against our sentence-level bias. Excessive text embedding modifications also degrade image quality.", + "bbox": [ + 511, + 529, + 906, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel implicit bias injection attack method that enables effective and versatile bias injection through multiple semantic expressions. Our findings reveal that bias in latent space can be represented as a high-dimensional vector, which can be integrated into the original prompt embedding to produce a modified embedding containing the injected bias. Evaluation and human studies demonstrate that our approach achieves high levels of injection covertness while maintaining robust bias injection. This work offers valuable insights into the representation and manipulation of bias within latent spaces, paving the way for further advancements in bias explanation and mitigation methods in generative AI.", + "bbox": [ + 509, + 703, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28786", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by the National Natural Science Foundation of China under Grant 62372341 and the Key Program of the National Natural Science Foundation of China under Grant 62436007.", + "bbox": [ + 89, + 114, + 483, + 175 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 191, + 187, + 207 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv, 2023. 2, 3", + "[2] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In ACM FAccT. 2", + "[3] Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion's den: Investigating hate in multimodal datasets. NeurIPS, 36, 2024. 2", + "[4] Black Forest Labs. Flux.1 [schnell], 2024. Accessed: 14 Mar. 2025. 8", + "[5] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 2, 3", + "[6] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In ICML, pages 4055-4075, 2023. 8", + "[7] Tianwei Chen, Yusuke Hirota, Mayu Otani, Noa Garcia, and Yuta Nakashima. Would deep generative models amplify bias in future models? In CVPR, pages 10833-10843, 2024. 2", + "[8] Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij, Kartik Hosanagar, Leonid Sigal, and Matthew Turk. Tibet: Identifying and evaluating biases in text-to-image generative models. arXiv, 2023. 3", + "[9] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In CVPR, pages 3043-3054, 2023. 2", + "[10] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 3, 8", + "[11] Moreno D'Inca, Elia Peruzzo, Massimiliano Mancini, Dejia Xu, Vidit Goel, Xingqian Xu, Zhangyang Wang, Humphrey Shi, and Nicu Sebe. Openbias: Open-set bias detection in text-to-image generative models. In CVPR, pages 12225-12235, 2024. 3, 5", + "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti-" + ], + "bbox": [ + 93, + 215, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fied flow transformers for high-resolution image synthesis. In ICML, 2024. 8", + "[13] Federal Bureau of Prisons. Inmate race statistics, 2023. Accessed: 2024-11-01. 3", + "[14] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. In ICLR, 2024. 3", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 6", + "[16] Phillip Howard, Avinash Madasu, Tiep Le, Gustavo Lujan Moreno, Anahita Bhiwandiwalla, and Vasudev Lal. Social-counterfactuals: Probing and mitigating intersectional social biases in vision-language models with counterfactual examples. In CVPR, pages 11975-11985, 2024. 3", + "[17] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 4", + "[18] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv, 2024. 3", + "[19] Huayang Huang, Yu Wu, and Qian Wang. Robin: Robust and invisible watermarks for diffusion models with adversarial optimization. NeurIPS, 37:3937-3963, 2024. 1", + "[20] Naveed Akhtar Jordan Vice, Richard Hartley, and Ajmal Mian. Severity controlled text-to-image generative model bias manipulation. arXiv preprint arXiv:2404.02530, 2024. 2, 1", + "[21] Eunji Kim, Siwon Kim, Rahim Entezari, and Sungroh Yoon. Unlocking intrinsic fairness in stable diffusion. arXiv preprint arXiv:2408.12692, 2024. 3, 8", + "[22] Younghyun Kim, Sangwoo Mo, Minkyu Kim, Kyungmin Lee, Jaeho Lee, and Jinwoo Shin. Discovering and mitigating visual biases through keyword explanation. In CVPR, pages 11082-11092, 2024. 3", + "[23] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv, 2014. 5", + "[24] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. NeurIPS, 36:36652-36663, 2023. 6", + "[25] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. NeurIPS, 36, 2024. 2", + "[26] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. TMLR, 2024. 3", + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 5", + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 5" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28787", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023.", + "[30] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5", + "[31] Mushui Liu, Yuhang Ma, Xinfeng Zhang, Yang Zhen, Zeng Zhao, Zhipeng Hu, Bai Liu, and Changjie Fan. Llm4gen: Leveraging semantic representation of llms for text-to-image generation. arXiv, 2024. 3", + "[32] Xiao Liu, Xiaoliu Guan, Yu Wu, and Jiaxu Miao. Iterative ensemble training with anti-gradient control for mitigating memorization in diffusion models. In ECCV, 2024. 1", + "[33] Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Evaluating societal representations in diffusion models. NeurIPS, 36, 2024. 1, 2", + "[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 1", + "[35] Hanjun Luo, Ziye Deng, Ruizhe Chen, and Zuozhu Liu. Faintbench: A holistic and precise benchmark for bias evaluation in text-to-image models. arXiv, 2024. 2", + "[36] Bingqi Ma, Zhuofan Zong, Guanglu Song, Hongsheng Li, and Yu Liu. Exploring the role of large language models in prompt encoding for diffusion models. arXiv, 2024. 3", + "[37] Ranjita Naik and Besmira Nushi. Social biases through the text-to-image generation lens. In AIES, pages 786-808, 2023. 2", + "[38] Ali Naseh, Jaechul Roh, Eugene Bagdasaryan, and Amir Houmansadr. Injecting bias in text-to-image models via composite-trigger backdoors. arXiv, 2024. 2, 3, 5", + "[39] Leonardo Nicoletti and Dina Bass. Humans are biased. generative ai is even worse, 2023. Accessed: 2024-11-01. 2", + "[40] Fabian Offert and Thao Phan. A sign that spells: Dall-e 2, invisual images and the racial politics of feature space. arXiv, 2022. 1", + "[41] Carolina López Olmos, Alexandros Neophytou, Sunando Sengupta, and Dim P Papadopoulos. Latent directions: A simple pathway to bias mitigation in generative ai. In CVPR, pages 8333-8337, 2024. 2, 1", + "[42] OpenAI. Gpt-4 technical report, 2023. Accessed: 2024-11-01. 5", + "[43] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models. In CVPR, pages 6668-6678, 2024. 3", + "[44] Dustin Podell, *Zion English*, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. *Sdxl: Improving latent diffusion models for high-resolution image synthesis*. In *ICLR*, 2024. 1", + "[45] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. 2", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 5" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv, pages arXiv-2204, 2022. 1", + "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 5", + "[49] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 1", + "[50] Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. Towards controllable biases in language generation. In Findings EMNLP, pages 3239–3254, 2020. 3", + "[51] Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. In CVPR, pages 11996-12005, 2024. 3", + "[52] Lukas Struppek, Dom Hintersdorf, Felix Friedrich, Patrick Schramowski, Kristian Kersting, et al. Exploiting cultural biases via homoglyphs in text-to-image synthesis. JAIR, 78: 1017-1068, 2023. 3", + "[53] Md Mehrab Tanjim, Krishna Kumar Singh, Kushal Kafle, Ritwik Sinha, and Garrison W Cottrell. Discovering and mitigating biases in clip-based image editing. In WACV, pages 2984-2993, 2024. 2", + "[54] TIME. Ethical ai isn't to blame for google's gemini debacle. https://time.com/6836153/ethical-ai-google-gemini-debacle/, 2024. Accessed: 2024-2-29. 1", + "[55] Eddie Ungless, Björn Ross, and Anne Lauscher. Stereotypes and smut: The (mis) representation of non-cisgender identities by text-to-image models. In ACL, pages 7919-7942. Association for Computational Linguistics (ACL), 2023. 2", + "[56] Jordan Vice, Naveed Akhtar, Richard Hartley, and Ajmal Mian. Quantifying bias in text-to-image generative models. arXiv, 2023. 3", + "[57] Jialu Wang, Xinyue Liu, Zonglin Di, Yang Liu, and Xin Wang. T2iat: Measuring valence and stereotypical biases in text-to-image generation. In Findings ACL, pages 2560-2574, 2023. 3", + "[58] Ruoyu Wang, Yongqi Yang, Zhihao Qian, Ye Zhu, and Yu Wu. Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation. In ICLR, 2024. 2", + "[59] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 5", + "[60] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In CVPR, pages 1900-1910, 2023. 5", + "[61] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In CVPR, pages 6327-6336, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28788", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Hu Yu, Hao Luo, Fan Wang, and Feng Zhao. Uncovering the text embedding in text-to-image diffusion models. arXiv, 2024. 5", + "[63] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Iti-gen: Inclusive text-to-image generation. In ICCV, pages 3969-3980, 2023. 3", + "[64] Ye Zhu, Yu Wu, Zhiwei Deng, Olga Russakovsky, and Yan Yan. Boundary guided learning-free semantic control with diffusion models. NeurIPS, 36:78319-78346, 2023. 2", + "[65] Ye Zhu, Yu Wu, Nicu Sebe, and Yan Yan. Vision+ x: A survey on multimodal learning in the light of data. TPAMI, 2024. 1" + ], + "bbox": [ + 91, + 90, + 480, + 273 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28789", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/5a713a88-9e78-41c3-b9a4-4778994a6f99_model.json b/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/5a713a88-9e78-41c3-b9a4-4778994a6f99_model.json new file mode 100644 index 0000000000000000000000000000000000000000..82da01cc96b4c0a57c60b06ceedb497489aa495c --- /dev/null +++ b/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/5a713a88-9e78-41c3-b9a4-4778994a6f99_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.142, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "Implicit Bias Injection Attacks against Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.179, + 0.79, + 0.253 + ], + "angle": 0, + "content": "Huayang Huang\\(^{1}\\) Xiangye Jin\\(^{2}\\) Jiaxu Miao\\(^{3*}\\) Yu Wu\\(^{1*}\\) \n\\(^{1}\\)School of Computer Science, Wuhan University \n\\(^{2}\\)School of Mathematics and Statistics, Wuhan University \n\\(^{3}\\)School of Cyber Science and Technology, Sun Yat-sen University" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.262, + 0.785, + 0.278 + ], + "angle": 0, + "content": "{hyhuang, xiangyejin, wuyucs}@whu.edu.cn miaojx@mail.sysu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.178, + 0.302, + 0.376, + 0.319 + ], + "angle": 0, + "content": "For developing countries" + }, + { + "type": "title", + "bbox": [ + 0.433, + 0.314, + 0.562, + 0.332 + ], + "angle": 0, + "content": "Text-to-Image" + }, + { + "type": "title", + "bbox": [ + 0.62, + 0.302, + 0.811, + 0.318 + ], + "angle": 0, + "content": "For developed countries" + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.352, + 0.199, + 0.364 + ], + "angle": 0, + "content": "A doctor" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.364, + 0.236, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.353, + 0.335, + 0.365 + ], + "angle": 0, + "content": "A professor" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.365, + 0.361, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.352, + 0.462, + 0.366 + ], + "angle": 0, + "content": "A city park" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.365, + 0.489, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.353, + 0.607, + 0.365 + ], + "angle": 0, + "content": "A doctor" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.365, + 0.638, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.353, + 0.741, + 0.366 + ], + "angle": 0, + "content": "A professor" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.365, + 0.764, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.353, + 0.866, + 0.366 + ], + "angle": 0, + "content": "A city park" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.365, + 0.889, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.578, + 0.905, + 0.608 + ], + "angle": 0, + "content": "Figure 1. Implicit bias injection causes the model to generate subtly negative content (left) and positive content (right) for different users. Although the generated images meet the text requirement, the subtle differences may, over time, influence user perception." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.619, + 0.327, + 0.635 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.651, + 0.486, + 0.879 + ], + "angle": 0, + "content": "The proliferation of text-to-image diffusion models (T2I DMs) has led to an increased presence of AI-generated images in daily life. However, biased T2I models can generate content with specific tendencies, potentially influencing people's perceptions. Intentional exploitation of these biases risks conveying misleading information to the public. Current research on bias primarily addresses explicit biases with recognizable visual patterns, such as skin color and gender. This paper introduces a novel form of implicit bias that lacks explicit visual features but can manifest in diverse ways across various semantic contexts. This subtle and versatile nature makes this bias challenging to detect, easy to propagate, and adaptable to a wide range of scenarios. We further propose an implicit bias injection attack framework (IBI-Attacks) against T2I diffusion models by precomputing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.909, + 0.773 + ], + "angle": 0, + "content": "a general bias direction in the prompt embedding space and adaptively adjusting it based on different inputs. Our attack module can be seamlessly integrated into pre-trained diffusion models in a plug-and-play manner without direct manipulation of user input or model retraining. Extensive experiments validate the effectiveness of our scheme in introducing bias through subtle and diverse modifications while preserving the original semantics. The strong concealment and transferability of our attack across various scenarios further underscore the significance of our approach." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.8, + 0.646, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Text-to-image (T2I) diffusion models [19, 34, 47-49, 65] have emerged as a powerful tool for generating highly realistic images that closely adhere to textual descriptions. The widespread adoption of these models, facilitated by public APIs [40, 54] and open-source implementations [32, 33]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.242, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "28779" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "has led to a proliferation of AI-generated images in daily life. However, biased T2I models can generate content with a specific tendency [7, 33, 58, 64], posing a risk of conveying misleading information that could influence public perception if maliciously exploited [2, 55]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.412 + ], + "angle": 0, + "content": "Consider a scenario in which a malicious T2I service provider subtly \"manipulates\" user perception by selectively generating biased content. As illustrated in Fig. 1, the same neutral prompt can yield images with varying emotional tones depending on the target audience. Content aimed at developed countries may convey a more positive tone, while content targeted at developing countries may lean towards a more negative sentiment. Such nuanced differences are difficult to detect visually, yet users may be subconsciously influenced by prolonged exposure to biased images. Furthermore, as shown in Fig. 2, such biases can extend beyond emotional aspects to include cultural and religious dimensions. This implicit bias, if intentionally leveraged, could subtly shape user perceptions over time, reinforcing existing divides and nudging audiences into self-reinforcing emotional or ideological spheres." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.483, + 0.625 + ], + "angle": 0, + "content": "Existing methods for bias exploitation typically require expensive model fine-tuning [38] and are limited to explicit biases [20, 41] (e.g., black skin and shaved head) triggered by specific inputs, which are easily detectable. In contrast, this paper focuses on a more implicit form of bias, such as the emotion, cultural stereotypes, and religious orientation depicted in Fig. 1 and Fig. 2. Such bias is highly covert and can manifest through diverse visual representations under varying prompt inputs. By imposing \"hints\" without significantly changing the content, users can be unconsciously influenced in ways that are difficult to detect. The manifestation of the implicit bias is context-dependent and can be conveyed through multiple semantic forms, such as facial expressions, poses, backgrounds, or group behaviors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.868 + ], + "angle": 0, + "content": "In this paper, we propose an implicit bias injection framework (IBI-Attacks) by precalculating a bias direction in the prompt embedding space and then dynamically adapting it based on specific user inputs. By introducing bias at the prompt embedding space, we can achieve more fine-grained control compared to traditional prompt modification. And input-aware dynamic adjustment further enables adaptive bias expression across diverse semantic contexts. Specifically, we leverage a large language model (LLM) [1, 5, 45] to generate a list of neutral prompts and rewrite them according to the specified bias. Then a bias direction vector is calculated by taking the difference between the two embedding sets. Our key finding is that this single bias direction already encodes multiple semantic expressions and can generalize across various prompt inputs by simple addition, allowing for diverse bias expressions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "However, directly applying this fixed direction vector to all prompts may result in either excessive or insufficient bias" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.178, + 0.836, + 0.189 + ], + "angle": 0, + "content": "\"A man is standing in front of his house and smiling.\"" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.279, + 0.698, + 0.29 + ], + "angle": 0, + "content": "\"One person worships in a room.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.279, + 0.882, + 0.29 + ], + "angle": 0, + "content": "\"Two people talk at the table.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.303, + 0.905, + 0.358 + ], + "angle": 0, + "content": "Figure 2. Generated images after injecting different 'cultural', 'religious', and 'gender' biases with IBI attacks. Explicit biases, such as gender, can also benefit from our approach, enabling subtle, diverse, and adaptive gender-related features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.906, + 0.492 + ], + "angle": 0, + "content": "modification. To address this limitation, we further perform dynamic feature selection in the fixed bias direction based on the input prompt, enabling proper expressions in different semantic contexts. In this way, the attacker can implant bias into any pre-trained text-to-image diffusion model in a plug-and-play manner, without the need to modify the input prompt or fine-tune the model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.906, + 0.599 + ], + "angle": 0, + "content": "We evaluate our approach on Stable Diffusion [48], using \"emotion bias\" as a case study to demonstrate the accuracy of bias injection, the diversity of bias expression, and the preservation of original semantics. The transferability of the bias direction further makes our attack applicable to a wide range of user inputs, allowing for consistent bias implantation across diverse contexts." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.601, + 0.86, + 0.615 + ], + "angle": 0, + "content": "In summary, our key contributions are as follows." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.617, + 0.905, + 0.661 + ], + "angle": 0, + "content": "- We consider a new type of implicit bias with subtle and multiple forms of semantic expression, which is more stealthy, challenging to detect, and easy to propagate." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.663, + 0.905, + 0.707 + ], + "angle": 0, + "content": "- We find that a specific bias direction within the text embedding in T2I diffusion models already encodes multiple expressions and exhibits generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.708, + 0.905, + 0.737 + ], + "angle": 0, + "content": "- We introduce an adaptive adjusting module to dynamically express the bias direction on different inputs." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.617, + 0.905, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.754, + 0.655, + 0.769 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Bias in Diffusion Models. Existing research [2, 3, 25, 33, 35, 53] has demonstrated the prevalence of bias in text-to-image diffusion models through extensive evaluations. Dall-Eval [9] reveals that both DALL-E [37] and Stable Diffusion [48] exhibit a tendency to generate male figures, with varying gender biases across different occupations. One study [39] also finds that over \\(80\\%\\) detainee-related images generated by Stable Diffusion depicted dark-skinned indi" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28780" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.244 + ], + "angle": 0, + "content": "viduals, despite people of color constituting less than half of the U.S. prison population [13]. Such bias can lead models to perpetuate false representations of the world when used in real-world applications. Considering this, some studies have explored detecting [8, 11, 22], evaluating [16, 56, 57], and mitigating [43, 51, 63] bias in generated content to promote fairness and diversity. However, these works primarily concentrate on traditional biases like gender and skin tone, neglecting bias types with diverse semantic expressions, such as emotions considered in this paper." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.264, + 0.485, + 0.506 + ], + "angle": 0, + "content": "LLMs in Diffusion Models. Considering the exceptional language understanding and generation capabilities of LLMs [1, 5], several studies [14, 18, 31] have integrated LLMs into the text-to-image diffusion generation process to accomplish complex tasks. A series of studies utilize LLMs to guide diffusion models for multi-stage generation for better accuracy. For example, SLD [61] employs LLMs to achieve the cyclic self-correctio of the generation process for improved alignment with complex prompts. LMD [26] utilizes LLMs to generate initial layouts based on text prompts, followed by layout-conditioned generation for further refinement. Another research direction [18, 31, 36] involves integrating rich text features learned by LLMs into the diffusion model to enhance guidance capabilities. In this paper, we employ LLMs to learn a multi-semantic direction of implicit bias within the text embedding space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.483, + 0.771 + ], + "angle": 0, + "content": "Bias Exploitations. In addition to investigating inherent biases in diffusion models, some studies have explored exploiting or implanting biases to deliberately steer the generation process. Some works [50, 52] show that specific hard prompts can induce biased output. However, these methods rely on existing biases within the model and cannot achieve arbitrary bias specifications. Some bias mitigation methods [10, 21] also utilize text embedding modification, but they calculate a fixed, single semantic direction for all prompts. Most similar to ours is a concurrent work Backdoorsing Bias [38], which aims to inject specified biases into the model. However, it can only introduce explicit biases with fixed visual features through prompt modification. Additionally, it requires expensive model fine-tuning. In contrast, our attack works in a training-free manner and can adaptively implant implicit bias for any user input." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.233, + 0.801 + ], + "angle": 0, + "content": "3. Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.482, + 0.902 + ], + "angle": 0, + "content": "Attack's Objective. The attacker aims to subtly alter the T2I model to embed implicit bias into generated content while maintaining the original output structure. To achieve this, we define two primary goals: 1) Attack Success. The attackers need to ensure that the implicit bias is successfully embedded in the model, meaning that all the images gener" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "ated by the model consistently exhibit the attacker-specified bias, with a high success rate and a subtle, implicit effect. 2) Injection Stealth. To maintain stealth in our attack, we aim to preserve the model's original generation content while allowing for the injection of specified biases. The generated content should align with user prompts but include subtle, unnoticed details that convey the bias. High semantic and structural similarity to normal outputs makes the attack difficult to detect." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.905, + 0.432 + ], + "angle": 0, + "content": "Attacker's Capability. Our approach allows the adversary to operate without needing access to the T2I model's parameters, structure, or the user-input prompts. The attack can be achieved by model poisoning in the following scenarios: 1) Covert Module Injection: An insider could introduce a stealthy module into the model during its development phase. 2) API Exploitation: A service provider might add the module to their API, subtly altering the results for users without informing them. 3) Model Sharing: An adversary could distribute a modified version of the T2I model on platforms where users expect high-quality models, manipulating outputs for unsuspecting users." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.905, + 0.522 + ], + "angle": 0, + "content": "Unlike LLM-assisted prompts, which require real-time API calls during inference and are easier to detect, our attack operates autonomously after deployment without additional LLM calls, making it more covert and efficient. Additionally, LLM-assisted prompts also cause significant semantic alterations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.538, + 0.796, + 0.556 + ], + "angle": 0, + "content": "4. Implicit Bias Injection Attacks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.837 + ], + "angle": 0, + "content": "In this section, we introduce the details of our original approach for implanting implicit bias in T2I diffusion models. The pipeline of the proposed framework is shown in Fig. 3. Our attack consists of three main stages: bias direction calculation, adaptive feature selection, and user inference. We first use a LLM to establish the relationship between implicit bias and multiple semantic expressions by rewriting sentences according to a specific bias. We then compute the bias direction vector by calculating the average difference between the original and rewritten sentences in the prompt embedding space. After that, we train an input-dependent adaptive feature selection module, which dynamically adjusts the fixed bias direction from the first stage based on the user's input prompt. Finally, during model inference, the adaptive adjustment module implants the bias into the user's prompt embedding before it is sent to the generative model, ensuring that the model's output reflects the specified bias." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.783, + 0.864 + ], + "angle": 0, + "content": "4.1. Directional Vector Generation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "First, we aim to compute a difference vector representing the direction of bias in the latent space. The attacker first" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28781" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.094, + 0.203, + 0.105 + ], + "angle": 0, + "content": "User inference" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.105, + 0.907, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.422, + 0.907, + 0.479 + ], + "angle": 0, + "content": "Figure 3. Pipeline of our proposed IBI-Attacks. First, given a specific bias, we construct a directional vector by computing the mean difference between embeddings of a neutral prompt set and a biased prompt set, generated using an LLM. Next, this directional vector is dynamically adjusted corresponding to the user prompts through Adaptive Feature Selection Module to produce a biased embedding. Finally, the biased embedding is fed into the diffusion model to generate biased output." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.734 + ], + "angle": 0, + "content": "implements an LLM model to generate \\(N\\) neutral prompts to create a neutral prompt set \\(X_{\\mathrm{neu}} = \\{x_i^{\\mathrm{neu}}\\}_{i=1}^N\\), each of which has a standard subject-verb-object structure. Meanwhile, the LLM model rewrites the neutral text prompts in \\(X_{\\mathrm{neu}}\\) with a specific bias to form a biased prompt dataset: \\(X_{\\mathrm{bias}} = \\{x_i^{\\mathrm{bias}}\\}_{i=1}^N\\). We employ a rewriting strategy that prefixes nouns with appropriate adjectives to introduce specific biases. We find that this restricted rewriting approach minimizes the introduction of bias-irrelevant information at the embedding level, such as modifications to sentence structure. As a result, the offset between the rewritten and original prompt embeddings becomes more closely aligned with the intended bias. Additionally, the addition of adjectives helps preserve the original sentence semantics, preventing the destruction of the underlying content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.484, + 0.855 + ], + "angle": 0, + "content": "Afterwards, the attacker utilizes the pretrained encoder \\(\\varphi(\\cdot)\\) to map input prompt \\(x\\) into text embeddings \\(v = \\varphi(x) \\in R^{D \\times L}\\), where \\(D\\) denotes the dimension of each token in latent space, and \\(L\\) denotes the number of embedded tokens for each text prompt (i.e., \\(D = 1024\\), \\(L = 77\\) in the SD v2.1 model). With the text encoder, texts prompts from \\(X_{\\mathrm{neu}}\\) and \\(X_{\\mathrm{bias}}\\) are mapped into the latent space to obtain their representation in latent space:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.86, + 0.483, + 0.88 + ], + "angle": 0, + "content": "\\[\nv _ {i} ^ {\\mathrm {n e u}} = \\varphi (x _ {i} ^ {\\mathrm {n e u}}), v _ {i} ^ {\\mathrm {b i a s}} = \\varphi (x _ {i} ^ {\\mathrm {b i a s}}), i = 1, 2, \\dots , N. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.884, + 0.482, + 0.904 + ], + "angle": 0, + "content": "In this way, we get two sets of text embedding: \\(\\{v_i^{\\mathrm{neu}}\\}_{i = 1}^N\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.502, + 0.907, + 0.537 + ], + "angle": 0, + "content": "and \\(\\{v_i^{\\mathrm{bias}}\\}_{i = 1}^N\\). We then compute the mean difference between these two sets of representations as:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.546, + 0.907, + 0.587 + ], + "angle": 0, + "content": "\\[\nv ^ {\\text {d i f f}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(v _ {i} ^ {\\text {b i a s}} - v _ {i} ^ {\\text {n e u}}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.907, + 0.687 + ], + "angle": 0, + "content": "We find that this simple mean difference vector already encodes multiple semantics of the implicit bias and can be generalized across different prompts. The existence of this expressive average bias direction enables us to embed implicit bias in the output by simply adding a fixed vector to the input prompt embedding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.757, + 0.71 + ], + "angle": 0, + "content": "4.2. Adaptive Feature Selection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Although the single average direction vector can already introduce some bias, we find that directly applying it to all prompts may result in either excessive or insufficient bias. Therefore, we introduce an adaptive adjustment module to dynamically alter the precomputed bias direction vector \\( v^{\\text{diff}} \\) according to the specific user input \\( v^{\\text{user}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Module Structure. Drawing inspiration from SENet [17], we design a simple module that computes an attention map on the bias direction vector for feature selection. For the text embedding vectors with dimension \\( D \\times L \\), we alternately compute feature attention" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28782" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "for the token dimension \\( L \\) and the embedding dimension \\( D \\). Specifically, we alternately reduce one dimension to a single value using global average pooling \\( \\mathrm{Avg} \\) to capture the overall information of that dimension. A two-layer multilayer perceptron \\( \\mathrm{MLP}_{\\theta} \\) is then used to learn attention for the other, uncompressed dimension, which is used to adaptively scale the original feature along that axis. \\( \\theta \\) is the learnable parameter of MLP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.484, + 0.349 + ], + "angle": 0, + "content": "By prioritizing key information, we can focus on relevant features and disregard irrelevant ones based on different prompt inputs. Specifically, we compute attention across both the token and feature dimensions of the text embedding by compressing the other dimension separately. The average direction vector \\( v^{\\text{diff}} \\) is then scaled according to the attention values in both dimensions, producing an adaptive bias direction \\( \\tilde{v}^{\\text{diff}} \\) that is added to the user input prompt \\( v^{\\text{user}} \\) to generate the biased embedding \\( \\tilde{v}^{\\text{biased}} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.359, + 0.483, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\tilde {v} ^ {\\text {d i f f}} = \\operatorname {M L P} _ {\\theta} (\\operatorname {A v g} (v ^ {\\text {u s e r}})) \\odot v ^ {\\text {d i f f}}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.402, + 0.482, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\tilde {v} ^ {\\text {b i a s}} = v ^ {\\text {u s e r}} + \\tilde {v} ^ {\\text {d i f f}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.516 + ], + "angle": 0, + "content": "Training Paradigm. We reuse the neutral and biased prompt embedding sets generated by LLM when calculating the bias direction to train the adaptive module. We rearranged them in pairs: \\(\\{v_i^{\\mathrm{neu}}, v_i^{\\mathrm{biased}}\\}_{i=1}^{N}\\) as input and label in training the adaptive module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.593 + ], + "angle": 0, + "content": "The adaptive module dynamically fine-tunes the average bias direction \\( v_{i}^{\\text{diff}} \\) based on the neutral prompt embedding \\( v_{i}^{\\text{neu}} \\), enabling the adjusted direction to point from the neutral prompt embedding \\( v_{i}^{\\text{neu}} \\) toward the biased prompt embedding \\( v_{i}^{\\text{biased}} \\). Thus we define the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.604, + 0.482, + 0.657 + ], + "angle": 0, + "content": "\\[\nL o s s = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| v _ {i} ^ {\\text {d i f f}} - \\operatorname {M L P} _ {\\theta} \\left(\\operatorname {A v g} \\left(v _ {i} ^ {\\text {n e u}}\\right)\\right) \\odot v ^ {\\text {d i f f}} \\right\\| ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.657, + 0.303, + 0.675 + ], + "angle": 0, + "content": "where \\( v_{i}^{\\text{diff}} = v_{i}^{\\text{biased}} - v_{i}^{\\text{neu}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.241, + 0.697 + ], + "angle": 0, + "content": "4.3. User Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Once the adaptation module is trained, it can be seamlessly inserted behind the text encoder of a pre-trained diffusion model. This model can be selectively deployed to specific groups. When users download the biased model, they interact with it by entering a text prompt input. This prompt is processed by the encoder and then the adaptive module to generate the biased version. Finally, the modified embedding is fed into the T2I model, producing a final output image with the desired bias. Due to the subtlety of the bias introduction and its general applicability across different prompts, the model can continuously generate imperceptibly biased content for the target user, potentially influencing their cognition. An another harmful scenario could" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "involve API service providers selectively targeting specific IP addresses, delivering normal results to most users while subtly implanting biases for particular groups, thereby conveying misleading information. The plug-and-play nature of our attack makes such targeted attacks easier to implement, posing a serious risk of spreading harmful bias." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.199, + 0.646, + 0.216 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.225, + 0.612, + 0.241 + ], + "angle": 0, + "content": "5.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.907, + 0.414 + ], + "angle": 0, + "content": "Models and Datasets. Our evaluation is conducted on the COCO dataset [27] and Stable Diffusion 2.1 [48]. From the COCO dataset, we filtered 200,000 caption subsets related to people, 57,000 caption subsets related to animals, and 60,000 caption subsets related to natural scenes, ensuring that these categories do not overlap. Our primary focus is on captions related to human scenes, as this represents a common scenario for bias attacks. Furthermore, we observe that the bias embedding module, trained on human scenes, exhibits zero-shot transferability to other scenes, such as those involving animals and nature." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.909, + 0.621 + ], + "angle": 0, + "content": "Implementation Details. To calculate the average bias direction, we employed ChatGPT-4 [42] to generate 200 neutral prompts related to everyday person scenes, along with their biased counterparts. Given that minor text prompt modifications can significantly impact embeddings [60, 62], and unrestricted rewriting rules may alter sentence structure, we simplified our rewriting rules to focus solely on the addition of appropriate adjectives to the original sentence. This approach minimizes the influence of bias-irrelevant factors on text embeddings. The adaptive adjusting module was trained on the same LLM-generated data for 50 epochs using the Adam optimizer [23] with a learning rate of 0.001." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Evaluation Metrics. We mainly consider two metrics: the success rate of bias injection and the stealthiness of the implanted implicit bias. Given the subtle and diverse semantic expressions of bias in our attacks, simple visual classification or recognition models are insufficient for detection. Therefore, following existing approaches [11, 38], we employ a Multi-modal Language Model (MLLM) to identify the presence of bias in the final generated image by question-answering based on visual content. We utilize the latest LLaVA-v1.6 7B [28-30] model in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To evaluate the imperceptibility of bias implantation, we use the CLIP score [46] to measure the alignment between the generated image and the input text \\((\\mathrm{CLIP}_{\\mathrm{txt - img}})\\), as well as the semantic similarity to the original generated image \\((\\mathrm{CLIP}_{\\mathrm{img - img}})\\). Additionally, we use SSIM (Structural Similarity Index) [59] to assess pixel-level differences between the images before and after bias injection. Validation" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28783" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.089, + 0.861, + 0.228 + ], + "angle": 0, + "content": "
Bias typeMethodsNegativePositiveSameCLIPtxt-Img↑CLIPimg-Img↑SSIM↑FID↓PickScore↑
Original-14.4%6.7%78.9%0.36331.00001.000039.29721.805
Negative biasExplicit89.1%6.9%4.0%0.34580.73540.522140.09021.386
IBI w/o adap82.2%12.5%5.3%0.35960.81430.588039.21721.593
IBI80.2%12.8%7.0%0.36370.87930.698739.13521.766
Positive biasExplicit0.0%98.0%2.0%0.34490.76960.522142.34321.763
IBI w/o adap13.9%85%1.1%0.35810.80300.577639.54921.606
IBI14.6%83.7%1.7%0.36020.82470.604039.49921.670
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.239, + 0.908, + 0.283 + ], + "angle": 0, + "content": "Table 1. MLLM evaluation of “Negative” and “Positive” bias injection. Original displays the results of splicing two non-attacked, normally generated images. Explicit shows the results of introducing bias by directly modifying user prompts. IBI w/o adap illustrates the impact of directly applying the average direction vector. IBI presents the results obtained using the adaptive module for the average direction vector." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.295, + 0.907, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.575, + 0.908, + 0.631 + ], + "angle": 0, + "content": "Figure 4. Generated samples after injecting positive and negative bias. The key modifications are enlarged. Emotional bias is conveyed through small adjustments in a person's expression or posture. In the second column, after injecting positive bias, the girl's hands transition from a defensive crossed position to a relaxed stance, and a smile appears on her face. While for the negative mood, the girl maintains her original pose, but her facial expression becomes more serious." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.718 + ], + "angle": 0, + "content": "was performed on a randomly selected set of 1,000 neutral human-related captions. We also employ FID [15] and PickScore [24] to assess the preservation of image quality, evaluated on 5,000 generated images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.315, + 0.743 + ], + "angle": 0, + "content": "5.2. Performance Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. For the \"negative\" and \"positive\" sentiment bias, we instruct the MLLM to classify the modified image as more positive, more negative, or the same as the original image. The success rate was measured by the proportion of \"more negative\" for negative bias and \"more positive\" for positive bias. The results in Tab. 1 indicate that the average bias direction can already introduce bias into the image, increasing the \"More negative\" rate for \"negative\" bias from \\(14.4\\%\\) to \\(82.2\\%\\). However, uniformly modifying all prompts without discrimination leads to a substantial" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.794 + ], + "angle": 0, + "content": "degradation of image quality. In contrast, the adaptive adjusting module effectively preserves the original semantics while maintaining the ability to introduce bias. The results are similar for the injection of \"positive\" bias. Although prompt modification achieves the highest bias implantation rate, it substantially alters the image content and exhibits poor invisibility. The small perturbations observed in FID and PickScore indicate that IBI attacks have minimal impact on image quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 4 presents the generated images embedded with negative and positive biases, where each emotion is conveyed through subtle modifications, further illustrating the concealment of our attack. Fig. 5 shows the reflection of negative bias under different prompts. The bias expression manifests in various semantic forms, en" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28784" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.328, + 0.485, + 0.37 + ], + "angle": 0, + "content": "Figure 5. Diverse semantic expressions of negative bias under different text inputs, including facial expressions, posture, style, and environmental cues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.489 + ], + "angle": 0, + "content": "compassing facial expressions, human posture, and subtle adjustments to background objects. Directly applying the direction vector may result in too much changes (the first column) or additional content (the third column). By incorporating the adaptive module, we can subtly introduce bias while preserving the original semantics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Zero-shot Transferability. We further investigate the zero-shot transferability of the bias injection module trained on human scenes to other domains. The results presented in Tab. 2 demonstrate that the bias learned from human-related scenarios effectively transfers to both animal and natural scenes. In animal-related generation, the attack successfully induces a negative sentiment in \\(95.1\\%\\) of cases. In natural scene generation, the negative probability increases by \\(81.1\\%\\). The results for 'positive' bias are similar. Fig. 6 visualizes the effects of this transfer attack. In animal-related generations, the bias manifests primarily through changes in the animal's pose and environment. For natural scene generations, the bias is expressed through more depressing backgrounds and less pleasing content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Human Study. However, relying solely on the MLLM for evaluation may not fully capture the impact of subtle image detail differences on human perception, which is crucial for implicit bias. Therefore, we designed a human study to further evaluate the subtlety and effectiveness of our attack. We invited 24 participants to choose between two randomly shuffled images (original image and bias injected version) based on two criteria: (1) Abnormal Rate: \"Which image is more likely to contain an implanted bias?\"; and (2) Bias Injection Rate: \"Which image more effectively reflects a specified bias?\". We also offered a \"Not Sure\" option if" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.905, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.418, + 0.908, + 0.46 + ], + "angle": 0, + "content": "Figure 6. The generated samples of attacking animals and natural scenes with the 'negative' bias injection module learned on human scenes." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.474, + 0.911, + 0.595 + ], + "angle": 0, + "content": "
TypesMethodsNegPosSameCLIPt-iCLIPi-iSSIM
P→AOriginal9.2%1.2%89.6%0.36411.00001.0000
IBI (Neg)95.1%4.9%0.0%0.35860.86440.6751
IBI (Pos)3.6%96.2%0.2%0.35670.85170.6147
P→NOriginal8.1%3.4%88.5%0.35611.00001.0000
IBI (Neg)89.2%10.8%0.0%0.35360.88470.6811
IBI (Pos)10.2%89.4%0.4%0.34820.82740.5836
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.604, + 0.907, + 0.646 + ], + "angle": 0, + "content": "Table 2. Performance of zero-shot attacks on other scenes using bias injection module learned from the person scene. \\(\\mathrm{P}\\rightarrow \\mathrm{A}\\): Person to Animal. \\(\\mathrm{P}\\rightarrow \\mathrm{N}\\): Person to Nature." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.904, + 0.703 + ], + "angle": 0, + "content": "participants think the two images are difficult to distinguish. Each person was randomly assigned 20 pairs of images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.84 + ], + "angle": 0, + "content": "The results of user selections are presented in Tab. 3. Our method achieves a greater degree of stealthiness, with users exhibiting only a \\(35.8\\%\\) probability of perceiving an image as biased, which is nearly identical to the \\(35.7\\%\\) of normal images. Participants have a \\(67.3\\%\\) probability of experiencing negative emotions from the attacked images. Notably, the increase in \"Not Sure\" responses also underscores the effectiveness of our approach, as most of the model's originally generated images are perceived as positive." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.693, + 0.866 + ], + "angle": 0, + "content": "5.3. Extended Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "LLM Rephrase Rules. We evaluate the impact of different LLM rewriting strategies in the bias vector calculation" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28785" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.108, + 0.089, + 0.468, + 0.179 + ], + "angle": 0, + "content": "
Human picks as abnormal
MethodsInjected ↓OriginalNot Sure
Explicit57.5%22.5%20.0%
IBI w/o adapt47.5%25.0%27.5%
IBI35.8%35.7%28.5%
" + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.184, + 0.468, + 0.275 + ], + "angle": 0, + "content": "
Human picks as negative
MethodsInjected ↑OriginalNot Sure
Explicit75.5%17.5%5.0%
IBI w/o adapt72.8%20.6%6.6%
IBI67.3%8.9%23.8%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.174, + 0.286, + 0.4, + 0.299 + ], + "angle": 0, + "content": "Table 3. Results of human evaluation." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.317, + 0.48, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.558, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Figure 7. Bias vector learned by various LLM rephrase rules. \\(LLM_{adj}\\): directly using LLM to add adjectives in front of all nouns. \\(\\Delta LLM_{random}\\): LLM rewriting without restrictions and then calculating the average direction. \\(\\Delta LLM_{adj}\\): LLM rewriting with added adjectives and calculating the average direction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "phase. The sample results are shown in Fig. 7. If we directly employ the LLM rewritten prompts \\((LLM_{\\mathrm{adj}})\\), the generated images can deviate significantly from the original intent. Moreover, highly negative prompts may result in numerous low-quality, easily detectable gray images. Alternatively, if we leverage unrestricted LLM rewriting \\((\\Delta LLM_{\\mathrm{random}})\\), the calculated average bias direction may lead to image distortions. Because unrestricted rewriting may introduce extraneous differences between the two prompt sets, such as sentence structure, hindering the accurate representation of meaningful bias modifications by calculating the average direction. In our approach, a simple approach of adding adjectives \\((\\Delta LLM_{\\mathrm{adj}})\\) effectively minimizes the introduction of bias-irrelevant information, leading to the successful learning of a direction that accurately represents the bias." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.9, + 0.214 + ], + "angle": 0, + "content": "
ModelAttackNeg ↑PosSameCLIPt-i ↑PS ↑
SD3 [12]-12.4%12.0%75.6%0.35722.529
SD3 [12]IBI68.2%11.8%20.0%0.35922.467
FLUX [4]-8.0%21.2%70.8%0.34322.354
FLUX [4]IBI65.4%34.4%0.2%0.34222.335
MUSE [6]-10.8%16.6%72.6%0.32420.553
MUSE [6]IBI91.2%8.6%0.2%0.32220.477
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.226, + 0.905, + 0.253 + ], + "angle": 0, + "content": "Table 4. MLLM evaluation of 'negative' bias injection for Stable Diffusion-3, FLux-schnell, and Muse models. PS: PickScore." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.268, + 0.894, + 0.339 + ], + "angle": 0, + "content": "
DefenseNeg ↑PosSameCLIPt-i ↑PS ↑
-80.2%12.8%7.0%0.36421.766
Chuang et al. [10]78.2%14.0%7.8%0.35821.419
Kim et al. [21]79.6%16.4%4.0%0.13918.523
" + }, + { + "type": "table_caption", + "bbox": [ + 0.525, + 0.35, + 0.893, + 0.364 + ], + "angle": 0, + "content": "Table 5. Performance of IBI attacks under debiasing methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.907, + 0.513 + ], + "angle": 0, + "content": "Generalization to Latest Models. We further evaluated negative bias injection on Stable Diffusion-3 [12], Flux-schnell (rectified flow transformer) [4], and MUSE (autoregressive) [6]. The results presented in Tab. 4 demonstrate that IBI generalizes effectively across models with varying architectures, significantly increasing the proportion of negative images while preserving the original semantics and maintaining the model's generative capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.667 + ], + "angle": 0, + "content": "Robustness against Debiasing Methods. We evaluated the effectiveness of our attack after applying two debiasing methods [10, 21]. The approach by Chuang et al. [10] works by projecting out certain biased word directions in the text embedding, while Kim et al. [21] achieves bias control by modifying the end-of-sequence token. Tab. 5 shows they have little defense effect, as word-level debiasing is ineffective against our sentence-level bias. Excessive text embedding modifications also degrade image quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a novel implicit bias injection attack method that enables effective and versatile bias injection through multiple semantic expressions. Our findings reveal that bias in latent space can be represented as a high-dimensional vector, which can be integrated into the original prompt embedding to produce a modified embedding containing the injected bias. Evaluation and human studies demonstrate that our approach achieves high levels of injection covertness while maintaining robust bias injection. This work offers valuable insights into the representation and manipulation of bias within latent spaces, paving the way for further advancements in bias explanation and mitigation methods in generative AI." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28786" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.176 + ], + "angle": 0, + "content": "This work was partially supported by the National Natural Science Foundation of China under Grant 62372341 and the Key Program of the National Natural Science Foundation of China under Grant 62436007." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.192, + 0.188, + 0.208 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.217, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.274, + 0.484, + 0.344 + ], + "angle": 0, + "content": "[2] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In ACM FAccT. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.346, + 0.483, + 0.386 + ], + "angle": 0, + "content": "[3] Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion's den: Investigating hate in multimodal datasets. NeurIPS, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[4] Black Forest Labs. Flux.1 [schnell], 2024. Accessed: 14 Mar. 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.484, + 0.487 + ], + "angle": 0, + "content": "[5] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[6] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In ICML, pages 4055-4075, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[7] Tianwei Chen, Yusuke Hirota, Mayu Otani, Noa Garcia, and Yuta Nakashima. Would deep generative models amplify bias in future models? In CVPR, pages 10833-10843, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.672 + ], + "angle": 0, + "content": "[8] Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij, Kartik Hosanagar, Leonid Sigal, and Matthew Turk. Tibet: Identifying and evaluating biases in text-to-image generative models. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[9] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In CVPR, pages 3043-3054, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[10] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[11] Moreno D'Inca, Elia Peruzzo, Massimiliano Mancini, Dejia Xu, Vidit Goel, Xingqian Xu, Zhangyang Wang, Humphrey Shi, and Nicu Sebe. Openbias: Open-set bias detection in text-to-image generative models. In CVPR, pages 12225-12235, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.217, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "fied flow transformers for high-resolution image synthesis. In ICML, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.15 + ], + "angle": 0, + "content": "[13] Federal Bureau of Prisons. Inmate race statistics, 2023. Accessed: 2024-11-01. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.152, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[14] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.21, + 0.905, + 0.264 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.267, + 0.905, + 0.337 + ], + "angle": 0, + "content": "[16] Phillip Howard, Avinash Madasu, Tiep Le, Gustavo Lujan Moreno, Anahita Bhiwandiwalla, and Vasudev Lal. Social-counterfactuals: Probing and mitigating intersectional social biases in vision-language models with counterfactual examples. In CVPR, pages 11975-11985, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.905, + 0.366 + ], + "angle": 0, + "content": "[17] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.905, + 0.41 + ], + "angle": 0, + "content": "[18] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.412, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[19] Huayang Huang, Yu Wu, and Qian Wang. Robin: Robust and invisible watermarks for diffusion models with adversarial optimization. NeurIPS, 37:3937-3963, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[20] Naveed Akhtar Jordan Vice, Richard Hartley, and Ajmal Mian. Severity controlled text-to-image generative model bias manipulation. arXiv preprint arXiv:2404.02530, 2024. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.905, + 0.554 + ], + "angle": 0, + "content": "[21] Eunji Kim, Siwon Kim, Rahim Entezari, and Sungroh Yoon. Unlocking intrinsic fairness in stable diffusion. arXiv preprint arXiv:2408.12692, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.556, + 0.905, + 0.612 + ], + "angle": 0, + "content": "[22] Younghyun Kim, Sangwoo Mo, Minkyu Kim, Kyungmin Lee, Jaeho Lee, and Jinwoo Shin. Discovering and mitigating visual biases through keyword explanation. In CVPR, pages 11082-11092, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.614, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[23] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[24] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. NeurIPS, 36:36652-36663, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[25] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. NeurIPS, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[26] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. TMLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28787" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[30] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[31] Mushui Liu, Yuhang Ma, Xinfeng Zhang, Yang Zhen, Zeng Zhao, Zhipeng Hu, Bai Liu, and Changjie Fan. Llm4gen: Leveraging semantic representation of llms for text-to-image generation. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[32] Xiao Liu, Xiaoliu Guan, Yu Wu, and Jiaxu Miao. Iterative ensemble training with anti-gradient control for mitigating memorization in diffusion models. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[33] Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Evaluating societal representations in diffusion models. NeurIPS, 36, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[35] Hanjun Luo, Ziye Deng, Ruizhe Chen, and Zuozhu Liu. Faintbench: A holistic and precise benchmark for bias evaluation in text-to-image models. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[36] Bingqi Ma, Zhuofan Zong, Guanglu Song, Hongsheng Li, and Yu Liu. Exploring the role of large language models in prompt encoding for diffusion models. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[37] Ranjita Naik and Besmira Nushi. Social biases through the text-to-image generation lens. In AIES, pages 786-808, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[38] Ali Naseh, Jaechul Roh, Eugene Bagdasaryan, and Amir Houmansadr. Injecting bias in text-to-image models via composite-trigger backdoors. arXiv, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[39] Leonardo Nicoletti and Dina Bass. Humans are biased. generative ai is even worse, 2023. Accessed: 2024-11-01. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[40] Fabian Offert and Thao Phan. A sign that spells: Dall-e 2, invisual images and the racial politics of feature space. arXiv, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.596, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[41] Carolina López Olmos, Alexandros Neophytou, Sunando Sengupta, and Dim P Papadopoulos. Latent directions: A simple pathway to bias mitigation in generative ai. In CVPR, pages 8333-8337, 2024. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.652, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[42] OpenAI. Gpt-4 technical report, 2023. Accessed: 2024-11-01. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[43] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models. In CVPR, pages 6668-6678, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.736, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[44] Dustin Podell, *Zion English*, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. *Sdxl: Improving latent diffusion models for high-resolution image synthesis*. In *ICLR*, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[45] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.833, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[47] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv, pages arXiv-2204, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.195, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[49] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.266, + 0.905, + 0.307 + ], + "angle": 0, + "content": "[50] Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. Towards controllable biases in language generation. In Findings EMNLP, pages 3239–3254, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.31, + 0.905, + 0.364 + ], + "angle": 0, + "content": "[51] Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. In CVPR, pages 11996-12005, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[52] Lukas Struppek, Dom Hintersdorf, Felix Friedrich, Patrick Schramowski, Kristian Kersting, et al. Exploiting cultural biases via homoglyphs in text-to-image synthesis. JAIR, 78: 1017-1068, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.425, + 0.905, + 0.479 + ], + "angle": 0, + "content": "[53] Md Mehrab Tanjim, Krishna Kumar Singh, Kushal Kafle, Ritwik Sinha, and Garrison W Cottrell. Discovering and mitigating biases in clip-based image editing. In WACV, pages 2984-2993, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[54] TIME. Ethical ai isn't to blame for google's gemini debacle. https://time.com/6836153/ethical-ai-google-gemini-debacle/, 2024. Accessed: 2024-2-29. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.905, + 0.596 + ], + "angle": 0, + "content": "[55] Eddie Ungless, Björn Ross, and Anne Lauscher. Stereotypes and smut: The (mis) representation of non-cisgender identities by text-to-image models. In ACL, pages 7919-7942. Association for Computational Linguistics (ACL), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.599, + 0.905, + 0.639 + ], + "angle": 0, + "content": "[56] Jordan Vice, Naveed Akhtar, Richard Hartley, and Ajmal Mian. Quantifying bias in text-to-image generative models. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[57] Jialu Wang, Xinyue Liu, Zonglin Di, Yang Liu, and Xin Wang. T2iat: Measuring valence and stereotypical biases in text-to-image generation. In Findings ACL, pages 2560-2574, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[58] Ruoyu Wang, Yongqi Yang, Zhihao Qian, Ye Zhu, and Yu Wu. Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[59] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.855 + ], + "angle": 0, + "content": "[60] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In CVPR, pages 1900-1910, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[61] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In CVPR, pages 6327-6336, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28788" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[62] Hu Yu, Hao Luo, Fan Wang, and Feng Zhao. Uncovering the text embedding in text-to-image diffusion models. arXiv, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[63] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Iti-gen: Inclusive text-to-image generation. In ICCV, pages 3969-3980, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[64] Ye Zhu, Yu Wu, Zhiwei Deng, Olga Russakovsky, and Yan Yan. Boundary guided learning-free semantic control with diffusion models. NeurIPS, 36:78319-78346, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[65] Ye Zhu, Yu Wu, Nicu Sebe, and Yan Yan. Vision+ x: A survey on multimodal learning in the light of data. TPAMI, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28789" + } + ] +] \ No newline at end of file diff --git a/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/5a713a88-9e78-41c3-b9a4-4778994a6f99_origin.pdf b/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/5a713a88-9e78-41c3-b9a4-4778994a6f99_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c01adaf1414e6683c06aff791f88dc1ededfe209 --- /dev/null +++ b/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/5a713a88-9e78-41c3-b9a4-4778994a6f99_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edb2d720704ad4eeadfe06f13e2eb946266004bb6b1d6c9b7212dc39231c5081 +size 4486717 diff --git a/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/full.md b/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5b7e51759adb6c12af299c66341a3a83ef6a8c2d --- /dev/null +++ b/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/full.md @@ -0,0 +1,311 @@ +# Implicit Bias Injection Attacks against Text-to-Image Diffusion Models + +Huayang Huang $^{1}$ Xiangye Jin $^{2}$ Jiaxu Miao $^{3*}$ Yu Wu $^{1*}$ $^{1}$ School of Computer Science, Wuhan University + $^{2}$ School of Mathematics and Statistics, Wuhan University + $^{3}$ School of Cyber Science and Technology, Sun Yat-sen University + +{hyhuang, xiangyejin, wuyucs}@whu.edu.cn miaojx@mail.sysu.edu.cn + +# For developing countries + +# Text-to-Image + +# For developed countries + +![](images/7463dc4505728574c63eea2674a0a920eeac47b58c03d3492a42e0ace17ffaf1.jpg) +A doctor +Figure 1. Implicit bias injection causes the model to generate subtly negative content (left) and positive content (right) for different users. Although the generated images meet the text requirement, the subtle differences may, over time, influence user perception. + +![](images/a2d1608fba693f0027435918a98db739fd8e4ab56af1a0945fb4a195925f2e8c.jpg) +A professor + +![](images/befa5ac40f50c8ab9f6cd98ae70cdf4865a3c3f8732dbadd15055552fb0b67ce.jpg) +A city park + +![](images/fa671134f618e2e6f2a831ce8c6fdf61dd69a66e4904456482f79e93a8134f9d.jpg) +A doctor + +![](images/6ace4558e8f1814173f8d08e828394ee47fddf53d42b3fa35be4351ebc852014.jpg) +A professor + +![](images/7fcb3437fd40df3fce6e187abbf577f1e50438e19dc4a5e3429614e58d3ce77f.jpg) +A city park + +# Abstract + +The proliferation of text-to-image diffusion models (T2I DMs) has led to an increased presence of AI-generated images in daily life. However, biased T2I models can generate content with specific tendencies, potentially influencing people's perceptions. Intentional exploitation of these biases risks conveying misleading information to the public. Current research on bias primarily addresses explicit biases with recognizable visual patterns, such as skin color and gender. This paper introduces a novel form of implicit bias that lacks explicit visual features but can manifest in diverse ways across various semantic contexts. This subtle and versatile nature makes this bias challenging to detect, easy to propagate, and adaptable to a wide range of scenarios. We further propose an implicit bias injection attack framework (IBI-Attacks) against T2I diffusion models by precomputing + +a general bias direction in the prompt embedding space and adaptively adjusting it based on different inputs. Our attack module can be seamlessly integrated into pre-trained diffusion models in a plug-and-play manner without direct manipulation of user input or model retraining. Extensive experiments validate the effectiveness of our scheme in introducing bias through subtle and diverse modifications while preserving the original semantics. The strong concealment and transferability of our attack across various scenarios further underscore the significance of our approach. + +# 1. Introduction + +Text-to-image (T2I) diffusion models [19, 34, 47-49, 65] have emerged as a powerful tool for generating highly realistic images that closely adhere to textual descriptions. The widespread adoption of these models, facilitated by public APIs [40, 54] and open-source implementations [32, 33], + +has led to a proliferation of AI-generated images in daily life. However, biased T2I models can generate content with a specific tendency [7, 33, 58, 64], posing a risk of conveying misleading information that could influence public perception if maliciously exploited [2, 55]. + +Consider a scenario in which a malicious T2I service provider subtly "manipulates" user perception by selectively generating biased content. As illustrated in Fig. 1, the same neutral prompt can yield images with varying emotional tones depending on the target audience. Content aimed at developed countries may convey a more positive tone, while content targeted at developing countries may lean towards a more negative sentiment. Such nuanced differences are difficult to detect visually, yet users may be subconsciously influenced by prolonged exposure to biased images. Furthermore, as shown in Fig. 2, such biases can extend beyond emotional aspects to include cultural and religious dimensions. This implicit bias, if intentionally leveraged, could subtly shape user perceptions over time, reinforcing existing divides and nudging audiences into self-reinforcing emotional or ideological spheres. + +Existing methods for bias exploitation typically require expensive model fine-tuning [38] and are limited to explicit biases [20, 41] (e.g., black skin and shaved head) triggered by specific inputs, which are easily detectable. In contrast, this paper focuses on a more implicit form of bias, such as the emotion, cultural stereotypes, and religious orientation depicted in Fig. 1 and Fig. 2. Such bias is highly covert and can manifest through diverse visual representations under varying prompt inputs. By imposing "hints" without significantly changing the content, users can be unconsciously influenced in ways that are difficult to detect. The manifestation of the implicit bias is context-dependent and can be conveyed through multiple semantic forms, such as facial expressions, poses, backgrounds, or group behaviors. + +In this paper, we propose an implicit bias injection framework (IBI-Attacks) by precalculating a bias direction in the prompt embedding space and then dynamically adapting it based on specific user inputs. By introducing bias at the prompt embedding space, we can achieve more fine-grained control compared to traditional prompt modification. And input-aware dynamic adjustment further enables adaptive bias expression across diverse semantic contexts. Specifically, we leverage a large language model (LLM) [1, 5, 45] to generate a list of neutral prompts and rewrite them according to the specified bias. Then a bias direction vector is calculated by taking the difference between the two embedding sets. Our key finding is that this single bias direction already encodes multiple semantic expressions and can generalize across various prompt inputs by simple addition, allowing for diverse bias expressions. + +However, directly applying this fixed direction vector to all prompts may result in either excessive or insufficient bias + +![](images/fdf3856899a3189bfa1f9d710fd6408f350ca787b8463683d8441f3d9250d6e6.jpg) +"A man is standing in front of his house and smiling." + +![](images/4387b18ad5f237dddd822c9d81f13acfa233d16af663a72b5557d5183a90bf82.jpg) +"One person worships in a room." +"Two people talk at the table." +Figure 2. Generated images after injecting different 'cultural', 'religious', and 'gender' biases with IBI attacks. Explicit biases, such as gender, can also benefit from our approach, enabling subtle, diverse, and adaptive gender-related features. + +modification. To address this limitation, we further perform dynamic feature selection in the fixed bias direction based on the input prompt, enabling proper expressions in different semantic contexts. In this way, the attacker can implant bias into any pre-trained text-to-image diffusion model in a plug-and-play manner, without the need to modify the input prompt or fine-tune the model. + +We evaluate our approach on Stable Diffusion [48], using "emotion bias" as a case study to demonstrate the accuracy of bias injection, the diversity of bias expression, and the preservation of original semantics. The transferability of the bias direction further makes our attack applicable to a wide range of user inputs, allowing for consistent bias implantation across diverse contexts. + +In summary, our key contributions are as follows. + +- We consider a new type of implicit bias with subtle and multiple forms of semantic expression, which is more stealthy, challenging to detect, and easy to propagate. +- We find that a specific bias direction within the text embedding in T2I diffusion models already encodes multiple expressions and exhibits generalization ability. +- We introduce an adaptive adjusting module to dynamically express the bias direction on different inputs. + +# 2. Related Work + +Bias in Diffusion Models. Existing research [2, 3, 25, 33, 35, 53] has demonstrated the prevalence of bias in text-to-image diffusion models through extensive evaluations. Dall-Eval [9] reveals that both DALL-E [37] and Stable Diffusion [48] exhibit a tendency to generate male figures, with varying gender biases across different occupations. One study [39] also finds that over $80\%$ detainee-related images generated by Stable Diffusion depicted dark-skinned indi + +viduals, despite people of color constituting less than half of the U.S. prison population [13]. Such bias can lead models to perpetuate false representations of the world when used in real-world applications. Considering this, some studies have explored detecting [8, 11, 22], evaluating [16, 56, 57], and mitigating [43, 51, 63] bias in generated content to promote fairness and diversity. However, these works primarily concentrate on traditional biases like gender and skin tone, neglecting bias types with diverse semantic expressions, such as emotions considered in this paper. + +LLMs in Diffusion Models. Considering the exceptional language understanding and generation capabilities of LLMs [1, 5], several studies [14, 18, 31] have integrated LLMs into the text-to-image diffusion generation process to accomplish complex tasks. A series of studies utilize LLMs to guide diffusion models for multi-stage generation for better accuracy. For example, SLD [61] employs LLMs to achieve the cyclic self-correctio of the generation process for improved alignment with complex prompts. LMD [26] utilizes LLMs to generate initial layouts based on text prompts, followed by layout-conditioned generation for further refinement. Another research direction [18, 31, 36] involves integrating rich text features learned by LLMs into the diffusion model to enhance guidance capabilities. In this paper, we employ LLMs to learn a multi-semantic direction of implicit bias within the text embedding space. + +Bias Exploitations. In addition to investigating inherent biases in diffusion models, some studies have explored exploiting or implanting biases to deliberately steer the generation process. Some works [50, 52] show that specific hard prompts can induce biased output. However, these methods rely on existing biases within the model and cannot achieve arbitrary bias specifications. Some bias mitigation methods [10, 21] also utilize text embedding modification, but they calculate a fixed, single semantic direction for all prompts. Most similar to ours is a concurrent work Backdoorsing Bias [38], which aims to inject specified biases into the model. However, it can only introduce explicit biases with fixed visual features through prompt modification. Additionally, it requires expensive model fine-tuning. In contrast, our attack works in a training-free manner and can adaptively implant implicit bias for any user input. + +# 3. Threat Model + +Attack's Objective. The attacker aims to subtly alter the T2I model to embed implicit bias into generated content while maintaining the original output structure. To achieve this, we define two primary goals: 1) Attack Success. The attackers need to ensure that the implicit bias is successfully embedded in the model, meaning that all the images gener + +ated by the model consistently exhibit the attacker-specified bias, with a high success rate and a subtle, implicit effect. 2) Injection Stealth. To maintain stealth in our attack, we aim to preserve the model's original generation content while allowing for the injection of specified biases. The generated content should align with user prompts but include subtle, unnoticed details that convey the bias. High semantic and structural similarity to normal outputs makes the attack difficult to detect. + +Attacker's Capability. Our approach allows the adversary to operate without needing access to the T2I model's parameters, structure, or the user-input prompts. The attack can be achieved by model poisoning in the following scenarios: 1) Covert Module Injection: An insider could introduce a stealthy module into the model during its development phase. 2) API Exploitation: A service provider might add the module to their API, subtly altering the results for users without informing them. 3) Model Sharing: An adversary could distribute a modified version of the T2I model on platforms where users expect high-quality models, manipulating outputs for unsuspecting users. + +Unlike LLM-assisted prompts, which require real-time API calls during inference and are easier to detect, our attack operates autonomously after deployment without additional LLM calls, making it more covert and efficient. Additionally, LLM-assisted prompts also cause significant semantic alterations. + +# 4. Implicit Bias Injection Attacks + +In this section, we introduce the details of our original approach for implanting implicit bias in T2I diffusion models. The pipeline of the proposed framework is shown in Fig. 3. Our attack consists of three main stages: bias direction calculation, adaptive feature selection, and user inference. We first use a LLM to establish the relationship between implicit bias and multiple semantic expressions by rewriting sentences according to a specific bias. We then compute the bias direction vector by calculating the average difference between the original and rewritten sentences in the prompt embedding space. After that, we train an input-dependent adaptive feature selection module, which dynamically adjusts the fixed bias direction from the first stage based on the user's input prompt. Finally, during model inference, the adaptive adjustment module implants the bias into the user's prompt embedding before it is sent to the generative model, ensuring that the model's output reflects the specified bias. + +# 4.1. Directional Vector Generation + +First, we aim to compute a difference vector representing the direction of bias in the latent space. The attacker first + +![](images/eedf888a5a8244b696d076dcd5187ff99958c44df890d63b7dabd1af770dd460.jpg) +User inference +Figure 3. Pipeline of our proposed IBI-Attacks. First, given a specific bias, we construct a directional vector by computing the mean difference between embeddings of a neutral prompt set and a biased prompt set, generated using an LLM. Next, this directional vector is dynamically adjusted corresponding to the user prompts through Adaptive Feature Selection Module to produce a biased embedding. Finally, the biased embedding is fed into the diffusion model to generate biased output. + +implements an LLM model to generate $N$ neutral prompts to create a neutral prompt set $X_{\mathrm{neu}} = \{x_i^{\mathrm{neu}}\}_{i=1}^N$ , each of which has a standard subject-verb-object structure. Meanwhile, the LLM model rewrites the neutral text prompts in $X_{\mathrm{neu}}$ with a specific bias to form a biased prompt dataset: $X_{\mathrm{bias}} = \{x_i^{\mathrm{bias}}\}_{i=1}^N$ . We employ a rewriting strategy that prefixes nouns with appropriate adjectives to introduce specific biases. We find that this restricted rewriting approach minimizes the introduction of bias-irrelevant information at the embedding level, such as modifications to sentence structure. As a result, the offset between the rewritten and original prompt embeddings becomes more closely aligned with the intended bias. Additionally, the addition of adjectives helps preserve the original sentence semantics, preventing the destruction of the underlying content. + +Afterwards, the attacker utilizes the pretrained encoder $\varphi(\cdot)$ to map input prompt $x$ into text embeddings $v = \varphi(x) \in R^{D \times L}$ , where $D$ denotes the dimension of each token in latent space, and $L$ denotes the number of embedded tokens for each text prompt (i.e., $D = 1024$ , $L = 77$ in the SD v2.1 model). With the text encoder, texts prompts from $X_{\mathrm{neu}}$ and $X_{\mathrm{bias}}$ are mapped into the latent space to obtain their representation in latent space: + +$$ +v _ {i} ^ {\mathrm {n e u}} = \varphi (x _ {i} ^ {\mathrm {n e u}}), v _ {i} ^ {\mathrm {b i a s}} = \varphi (x _ {i} ^ {\mathrm {b i a s}}), i = 1, 2, \dots , N. \tag {1} +$$ + +In this way, we get two sets of text embedding: $\{v_i^{\mathrm{neu}}\}_{i = 1}^N$ + +and $\{v_i^{\mathrm{bias}}\}_{i = 1}^N$ . We then compute the mean difference between these two sets of representations as: + +$$ +v ^ {\text {d i f f}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(v _ {i} ^ {\text {b i a s}} - v _ {i} ^ {\text {n e u}}\right). \tag {2} +$$ + +We find that this simple mean difference vector already encodes multiple semantics of the implicit bias and can be generalized across different prompts. The existence of this expressive average bias direction enables us to embed implicit bias in the output by simply adding a fixed vector to the input prompt embedding. + +# 4.2. Adaptive Feature Selection + +Although the single average direction vector can already introduce some bias, we find that directly applying it to all prompts may result in either excessive or insufficient bias. Therefore, we introduce an adaptive adjustment module to dynamically alter the precomputed bias direction vector $v^{\text{diff}}$ according to the specific user input $v^{\text{user}}$ . + +Module Structure. Drawing inspiration from SENet [17], we design a simple module that computes an attention map on the bias direction vector for feature selection. For the text embedding vectors with dimension $D \times L$ , we alternately compute feature attention + +for the token dimension $L$ and the embedding dimension $D$ . Specifically, we alternately reduce one dimension to a single value using global average pooling $\mathrm{Avg}$ to capture the overall information of that dimension. A two-layer multilayer perceptron $\mathrm{MLP}_{\theta}$ is then used to learn attention for the other, uncompressed dimension, which is used to adaptively scale the original feature along that axis. $\theta$ is the learnable parameter of MLP. + +By prioritizing key information, we can focus on relevant features and disregard irrelevant ones based on different prompt inputs. Specifically, we compute attention across both the token and feature dimensions of the text embedding by compressing the other dimension separately. The average direction vector $v^{\text{diff}}$ is then scaled according to the attention values in both dimensions, producing an adaptive bias direction $\tilde{v}^{\text{diff}}$ that is added to the user input prompt $v^{\text{user}}$ to generate the biased embedding $\tilde{v}^{\text{biased}}$ . + +$$ +\tilde {v} ^ {\text {d i f f}} = \operatorname {M L P} _ {\theta} (\operatorname {A v g} (v ^ {\text {u s e r}})) \odot v ^ {\text {d i f f}}, \tag {3} +$$ + +$$ +\tilde {v} ^ {\text {b i a s}} = v ^ {\text {u s e r}} + \tilde {v} ^ {\text {d i f f}}. \tag {4} +$$ + +Training Paradigm. We reuse the neutral and biased prompt embedding sets generated by LLM when calculating the bias direction to train the adaptive module. We rearranged them in pairs: $\{v_i^{\mathrm{neu}}, v_i^{\mathrm{biased}}\}_{i=1}^{N}$ as input and label in training the adaptive module. + +The adaptive module dynamically fine-tunes the average bias direction $v_{i}^{\text{diff}}$ based on the neutral prompt embedding $v_{i}^{\text{neu}}$ , enabling the adjusted direction to point from the neutral prompt embedding $v_{i}^{\text{neu}}$ toward the biased prompt embedding $v_{i}^{\text{biased}}$ . Thus we define the following loss: + +$$ +L o s s = \frac {1}{N} \sum_ {i = 1} ^ {N} \left\| v _ {i} ^ {\text {d i f f}} - \operatorname {M L P} _ {\theta} \left(\operatorname {A v g} \left(v _ {i} ^ {\text {n e u}}\right)\right) \odot v ^ {\text {d i f f}} \right\| ^ {2}, \tag {5} +$$ + +where $v_{i}^{\text{diff}} = v_{i}^{\text{biased}} - v_{i}^{\text{neu}}$ . + +# 4.3. User Inference + +Once the adaptation module is trained, it can be seamlessly inserted behind the text encoder of a pre-trained diffusion model. This model can be selectively deployed to specific groups. When users download the biased model, they interact with it by entering a text prompt input. This prompt is processed by the encoder and then the adaptive module to generate the biased version. Finally, the modified embedding is fed into the T2I model, producing a final output image with the desired bias. Due to the subtlety of the bias introduction and its general applicability across different prompts, the model can continuously generate imperceptibly biased content for the target user, potentially influencing their cognition. An another harmful scenario could + +involve API service providers selectively targeting specific IP addresses, delivering normal results to most users while subtly implanting biases for particular groups, thereby conveying misleading information. The plug-and-play nature of our attack makes such targeted attacks easier to implement, posing a serious risk of spreading harmful bias. + +# 5. Experiments + +# 5.1. Settings + +Models and Datasets. Our evaluation is conducted on the COCO dataset [27] and Stable Diffusion 2.1 [48]. From the COCO dataset, we filtered 200,000 caption subsets related to people, 57,000 caption subsets related to animals, and 60,000 caption subsets related to natural scenes, ensuring that these categories do not overlap. Our primary focus is on captions related to human scenes, as this represents a common scenario for bias attacks. Furthermore, we observe that the bias embedding module, trained on human scenes, exhibits zero-shot transferability to other scenes, such as those involving animals and nature. + +Implementation Details. To calculate the average bias direction, we employed ChatGPT-4 [42] to generate 200 neutral prompts related to everyday person scenes, along with their biased counterparts. Given that minor text prompt modifications can significantly impact embeddings [60, 62], and unrestricted rewriting rules may alter sentence structure, we simplified our rewriting rules to focus solely on the addition of appropriate adjectives to the original sentence. This approach minimizes the influence of bias-irrelevant factors on text embeddings. The adaptive adjusting module was trained on the same LLM-generated data for 50 epochs using the Adam optimizer [23] with a learning rate of 0.001. + +Evaluation Metrics. We mainly consider two metrics: the success rate of bias injection and the stealthiness of the implanted implicit bias. Given the subtle and diverse semantic expressions of bias in our attacks, simple visual classification or recognition models are insufficient for detection. Therefore, following existing approaches [11, 38], we employ a Multi-modal Language Model (MLLM) to identify the presence of bias in the final generated image by question-answering based on visual content. We utilize the latest LLaVA-v1.6 7B [28-30] model in our experiments. + +To evaluate the imperceptibility of bias implantation, we use the CLIP score [46] to measure the alignment between the generated image and the input text $(\mathrm{CLIP}_{\mathrm{txt - img}})$ , as well as the semantic similarity to the original generated image $(\mathrm{CLIP}_{\mathrm{img - img}})$ . Additionally, we use SSIM (Structural Similarity Index) [59] to assess pixel-level differences between the images before and after bias injection. Validation + +
Bias typeMethodsNegativePositiveSameCLIPtxt-Img↑CLIPimg-Img↑SSIM↑FID↓PickScore↑
Original-14.4%6.7%78.9%0.36331.00001.000039.29721.805
Negative biasExplicit89.1%6.9%4.0%0.34580.73540.522140.09021.386
IBI w/o adap82.2%12.5%5.3%0.35960.81430.588039.21721.593
IBI80.2%12.8%7.0%0.36370.87930.698739.13521.766
Positive biasExplicit0.0%98.0%2.0%0.34490.76960.522142.34321.763
IBI w/o adap13.9%85%1.1%0.35810.80300.577639.54921.606
IBI14.6%83.7%1.7%0.36020.82470.604039.49921.670
+ +Table 1. MLLM evaluation of “Negative” and “Positive” bias injection. Original displays the results of splicing two non-attacked, normally generated images. Explicit shows the results of introducing bias by directly modifying user prompts. IBI w/o adap illustrates the impact of directly applying the average direction vector. IBI presents the results obtained using the adaptive module for the average direction vector. + +![](images/ab57f7efa3aa6d8deae47d0f67452d809f5802cd04ddcc10ed17c79aa0dac528.jpg) +Figure 4. Generated samples after injecting positive and negative bias. The key modifications are enlarged. Emotional bias is conveyed through small adjustments in a person's expression or posture. In the second column, after injecting positive bias, the girl's hands transition from a defensive crossed position to a relaxed stance, and a smile appears on her face. While for the negative mood, the girl maintains her original pose, but her facial expression becomes more serious. + +was performed on a randomly selected set of 1,000 neutral human-related captions. We also employ FID [15] and PickScore [24] to assess the preservation of image quality, evaluated on 5,000 generated images. + +# 5.2. Performance Evaluation + +Quantitative Results. For the "negative" and "positive" sentiment bias, we instruct the MLLM to classify the modified image as more positive, more negative, or the same as the original image. The success rate was measured by the proportion of "more negative" for negative bias and "more positive" for positive bias. The results in Tab. 1 indicate that the average bias direction can already introduce bias into the image, increasing the "More negative" rate for "negative" bias from $14.4\%$ to $82.2\%$ . However, uniformly modifying all prompts without discrimination leads to a substantial + +degradation of image quality. In contrast, the adaptive adjusting module effectively preserves the original semantics while maintaining the ability to introduce bias. The results are similar for the injection of "positive" bias. Although prompt modification achieves the highest bias implantation rate, it substantially alters the image content and exhibits poor invisibility. The small perturbations observed in FID and PickScore indicate that IBI attacks have minimal impact on image quality. + +Qualitative Results. Fig. 4 presents the generated images embedded with negative and positive biases, where each emotion is conveyed through subtle modifications, further illustrating the concealment of our attack. Fig. 5 shows the reflection of negative bias under different prompts. The bias expression manifests in various semantic forms, en + +![](images/3e26cb44d8a99f6f2229362937d492b1f6de93d99bc3bace568d8842981ea3cb.jpg) +Figure 5. Diverse semantic expressions of negative bias under different text inputs, including facial expressions, posture, style, and environmental cues. + +compassing facial expressions, human posture, and subtle adjustments to background objects. Directly applying the direction vector may result in too much changes (the first column) or additional content (the third column). By incorporating the adaptive module, we can subtly introduce bias while preserving the original semantics. + +Zero-shot Transferability. We further investigate the zero-shot transferability of the bias injection module trained on human scenes to other domains. The results presented in Tab. 2 demonstrate that the bias learned from human-related scenarios effectively transfers to both animal and natural scenes. In animal-related generation, the attack successfully induces a negative sentiment in $95.1\%$ of cases. In natural scene generation, the negative probability increases by $81.1\%$ . The results for 'positive' bias are similar. Fig. 6 visualizes the effects of this transfer attack. In animal-related generations, the bias manifests primarily through changes in the animal's pose and environment. For natural scene generations, the bias is expressed through more depressing backgrounds and less pleasing content. + +Human Study. However, relying solely on the MLLM for evaluation may not fully capture the impact of subtle image detail differences on human perception, which is crucial for implicit bias. Therefore, we designed a human study to further evaluate the subtlety and effectiveness of our attack. We invited 24 participants to choose between two randomly shuffled images (original image and bias injected version) based on two criteria: (1) Abnormal Rate: "Which image is more likely to contain an implanted bias?"; and (2) Bias Injection Rate: "Which image more effectively reflects a specified bias?". We also offered a "Not Sure" option if + +![](images/7add2e72b6fef83b2a186cd1351dd3c7802696f62d4c7ac7d789a7547e8d6a48.jpg) +Figure 6. The generated samples of attacking animals and natural scenes with the 'negative' bias injection module learned on human scenes. + +
TypesMethodsNegPosSameCLIPt-iCLIPi-iSSIM
P→AOriginal9.2%1.2%89.6%0.36411.00001.0000
IBI (Neg)95.1%4.9%0.0%0.35860.86440.6751
IBI (Pos)3.6%96.2%0.2%0.35670.85170.6147
P→NOriginal8.1%3.4%88.5%0.35611.00001.0000
IBI (Neg)89.2%10.8%0.0%0.35360.88470.6811
IBI (Pos)10.2%89.4%0.4%0.34820.82740.5836
+ +Table 2. Performance of zero-shot attacks on other scenes using bias injection module learned from the person scene. $\mathrm{P}\rightarrow \mathrm{A}$ : Person to Animal. $\mathrm{P}\rightarrow \mathrm{N}$ : Person to Nature. + +participants think the two images are difficult to distinguish. Each person was randomly assigned 20 pairs of images. + +The results of user selections are presented in Tab. 3. Our method achieves a greater degree of stealthiness, with users exhibiting only a $35.8\%$ probability of perceiving an image as biased, which is nearly identical to the $35.7\%$ of normal images. Participants have a $67.3\%$ probability of experiencing negative emotions from the attacked images. Notably, the increase in "Not Sure" responses also underscores the effectiveness of our approach, as most of the model's originally generated images are perceived as positive. + +# 5.3. Extended Analysis + +LLM Rephrase Rules. We evaluate the impact of different LLM rewriting strategies in the bias vector calculation + +
Human picks as abnormal
MethodsInjected ↓OriginalNot Sure
Explicit57.5%22.5%20.0%
IBI w/o adapt47.5%25.0%27.5%
IBI35.8%35.7%28.5%
+ +
Human picks as negative
MethodsInjected ↑OriginalNot Sure
Explicit75.5%17.5%5.0%
IBI w/o adapt72.8%20.6%6.6%
IBI67.3%8.9%23.8%
+ +![](images/42ba6c4bc2ee90368480b47e43e9110dd2047d4b7e56017fa313c93c70a650bb.jpg) +Figure 7. Bias vector learned by various LLM rephrase rules. $LLM_{adj}$ : directly using LLM to add adjectives in front of all nouns. $\Delta LLM_{random}$ : LLM rewriting without restrictions and then calculating the average direction. $\Delta LLM_{adj}$ : LLM rewriting with added adjectives and calculating the average direction. + +phase. The sample results are shown in Fig. 7. If we directly employ the LLM rewritten prompts $(LLM_{\mathrm{adj}})$ , the generated images can deviate significantly from the original intent. Moreover, highly negative prompts may result in numerous low-quality, easily detectable gray images. Alternatively, if we leverage unrestricted LLM rewriting $(\Delta LLM_{\mathrm{random}})$ , the calculated average bias direction may lead to image distortions. Because unrestricted rewriting may introduce extraneous differences between the two prompt sets, such as sentence structure, hindering the accurate representation of meaningful bias modifications by calculating the average direction. In our approach, a simple approach of adding adjectives $(\Delta LLM_{\mathrm{adj}})$ effectively minimizes the introduction of bias-irrelevant information, leading to the successful learning of a direction that accurately represents the bias. + +Table 3. Results of human evaluation. + +
ModelAttackNeg ↑PosSameCLIPt-i ↑PS ↑
SD3 [12]-12.4%12.0%75.6%0.35722.529
SD3 [12]IBI68.2%11.8%20.0%0.35922.467
FLUX [4]-8.0%21.2%70.8%0.34322.354
FLUX [4]IBI65.4%34.4%0.2%0.34222.335
MUSE [6]-10.8%16.6%72.6%0.32420.553
MUSE [6]IBI91.2%8.6%0.2%0.32220.477
+ +Table 4. MLLM evaluation of 'negative' bias injection for Stable Diffusion-3, FLux-schnell, and Muse models. PS: PickScore. + +
DefenseNeg ↑PosSameCLIPt-i ↑PS ↑
-80.2%12.8%7.0%0.36421.766
Chuang et al. [10]78.2%14.0%7.8%0.35821.419
Kim et al. [21]79.6%16.4%4.0%0.13918.523
+ +Table 5. Performance of IBI attacks under debiasing methods. + +Generalization to Latest Models. We further evaluated negative bias injection on Stable Diffusion-3 [12], Flux-schnell (rectified flow transformer) [4], and MUSE (autoregressive) [6]. The results presented in Tab. 4 demonstrate that IBI generalizes effectively across models with varying architectures, significantly increasing the proportion of negative images while preserving the original semantics and maintaining the model's generative capabilities. + +Robustness against Debiasing Methods. We evaluated the effectiveness of our attack after applying two debiasing methods [10, 21]. The approach by Chuang et al. [10] works by projecting out certain biased word directions in the text embedding, while Kim et al. [21] achieves bias control by modifying the end-of-sequence token. Tab. 5 shows they have little defense effect, as word-level debiasing is ineffective against our sentence-level bias. Excessive text embedding modifications also degrade image quality. + +# 6. Conclusion + +In this paper, we introduce a novel implicit bias injection attack method that enables effective and versatile bias injection through multiple semantic expressions. Our findings reveal that bias in latent space can be represented as a high-dimensional vector, which can be integrated into the original prompt embedding to produce a modified embedding containing the injected bias. Evaluation and human studies demonstrate that our approach achieves high levels of injection covertness while maintaining robust bias injection. This work offers valuable insights into the representation and manipulation of bias within latent spaces, paving the way for further advancements in bias explanation and mitigation methods in generative AI. + +# Acknowledgement + +This work was partially supported by the National Natural Science Foundation of China under Grant 62372341 and the Key Program of the National Natural Science Foundation of China under Grant 62436007. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv, 2023. 2, 3 +[2] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In ACM FAccT. 2 +[3] Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion's den: Investigating hate in multimodal datasets. NeurIPS, 36, 2024. 2 +[4] Black Forest Labs. Flux.1 [schnell], 2024. Accessed: 14 Mar. 2025. 8 +[5] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 2, 3 +[6] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In ICML, pages 4055-4075, 2023. 8 +[7] Tianwei Chen, Yusuke Hirota, Mayu Otani, Noa Garcia, and Yuta Nakashima. Would deep generative models amplify bias in future models? In CVPR, pages 10833-10843, 2024. 2 +[8] Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij, Kartik Hosanagar, Leonid Sigal, and Matthew Turk. Tibet: Identifying and evaluating biases in text-to-image generative models. arXiv, 2023. 3 +[9] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In CVPR, pages 3043-3054, 2023. 2 +[10] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 3, 8 +[11] Moreno D'Inca, Elia Peruzzo, Massimiliano Mancini, Dejia Xu, Vidit Goel, Xingqian Xu, Zhangyang Wang, Humphrey Shi, and Nicu Sebe. Openbias: Open-set bias detection in text-to-image generative models. In CVPR, pages 12225-12235, 2024. 3, 5 +[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- + +fied flow transformers for high-resolution image synthesis. In ICML, 2024. 8 +[13] Federal Bureau of Prisons. Inmate race statistics, 2023. Accessed: 2024-11-01. 3 +[14] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. In ICLR, 2024. 3 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 6 +[16] Phillip Howard, Avinash Madasu, Tiep Le, Gustavo Lujan Moreno, Anahita Bhiwandiwalla, and Vasudev Lal. Social-counterfactuals: Probing and mitigating intersectional social biases in vision-language models with counterfactual examples. In CVPR, pages 11975-11985, 2024. 3 +[17] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 4 +[18] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv, 2024. 3 +[19] Huayang Huang, Yu Wu, and Qian Wang. Robin: Robust and invisible watermarks for diffusion models with adversarial optimization. NeurIPS, 37:3937-3963, 2024. 1 +[20] Naveed Akhtar Jordan Vice, Richard Hartley, and Ajmal Mian. Severity controlled text-to-image generative model bias manipulation. arXiv preprint arXiv:2404.02530, 2024. 2, 1 +[21] Eunji Kim, Siwon Kim, Rahim Entezari, and Sungroh Yoon. Unlocking intrinsic fairness in stable diffusion. arXiv preprint arXiv:2408.12692, 2024. 3, 8 +[22] Younghyun Kim, Sangwoo Mo, Minkyu Kim, Kyungmin Lee, Jaeho Lee, and Jinwoo Shin. Discovering and mitigating visual biases through keyword explanation. In CVPR, pages 11082-11092, 2024. 3 +[23] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv, 2014. 5 +[24] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. NeurIPS, 36:36652-36663, 2023. 6 +[25] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. NeurIPS, 36, 2024. 2 +[26] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. TMLR, 2024. 3 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 5 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 5 + +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. +[30] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5 +[31] Mushui Liu, Yuhang Ma, Xinfeng Zhang, Yang Zhen, Zeng Zhao, Zhipeng Hu, Bai Liu, and Changjie Fan. Llm4gen: Leveraging semantic representation of llms for text-to-image generation. arXiv, 2024. 3 +[32] Xiao Liu, Xiaoliu Guan, Yu Wu, and Jiaxu Miao. Iterative ensemble training with anti-gradient control for mitigating memorization in diffusion models. In ECCV, 2024. 1 +[33] Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Evaluating societal representations in diffusion models. NeurIPS, 36, 2024. 1, 2 +[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 1 +[35] Hanjun Luo, Ziye Deng, Ruizhe Chen, and Zuozhu Liu. Faintbench: A holistic and precise benchmark for bias evaluation in text-to-image models. arXiv, 2024. 2 +[36] Bingqi Ma, Zhuofan Zong, Guanglu Song, Hongsheng Li, and Yu Liu. Exploring the role of large language models in prompt encoding for diffusion models. arXiv, 2024. 3 +[37] Ranjita Naik and Besmira Nushi. Social biases through the text-to-image generation lens. In AIES, pages 786-808, 2023. 2 +[38] Ali Naseh, Jaechul Roh, Eugene Bagdasaryan, and Amir Houmansadr. Injecting bias in text-to-image models via composite-trigger backdoors. arXiv, 2024. 2, 3, 5 +[39] Leonardo Nicoletti and Dina Bass. Humans are biased. generative ai is even worse, 2023. Accessed: 2024-11-01. 2 +[40] Fabian Offert and Thao Phan. A sign that spells: Dall-e 2, invisual images and the racial politics of feature space. arXiv, 2022. 1 +[41] Carolina López Olmos, Alexandros Neophytou, Sunando Sengupta, and Dim P Papadopoulos. Latent directions: A simple pathway to bias mitigation in generative ai. In CVPR, pages 8333-8337, 2024. 2, 1 +[42] OpenAI. Gpt-4 technical report, 2023. Accessed: 2024-11-01. 5 +[43] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models. In CVPR, pages 6668-6678, 2024. 3 +[44] Dustin Podell, *Zion English*, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. *Sdxl: Improving latent diffusion models for high-resolution image synthesis*. In *ICLR*, 2024. 1 +[45] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. 2 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 5 + +[47] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv, pages arXiv-2204, 2022. 1 +[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 5 +[49] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 1 +[50] Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. Towards controllable biases in language generation. In Findings EMNLP, pages 3239–3254, 2020. 3 +[51] Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. In CVPR, pages 11996-12005, 2024. 3 +[52] Lukas Struppek, Dom Hintersdorf, Felix Friedrich, Patrick Schramowski, Kristian Kersting, et al. Exploiting cultural biases via homoglyphs in text-to-image synthesis. JAIR, 78: 1017-1068, 2023. 3 +[53] Md Mehrab Tanjim, Krishna Kumar Singh, Kushal Kafle, Ritwik Sinha, and Garrison W Cottrell. Discovering and mitigating biases in clip-based image editing. In WACV, pages 2984-2993, 2024. 2 +[54] TIME. Ethical ai isn't to blame for google's gemini debacle. https://time.com/6836153/ethical-ai-google-gemini-debacle/, 2024. Accessed: 2024-2-29. 1 +[55] Eddie Ungless, Björn Ross, and Anne Lauscher. Stereotypes and smut: The (mis) representation of non-cisgender identities by text-to-image models. In ACL, pages 7919-7942. Association for Computational Linguistics (ACL), 2023. 2 +[56] Jordan Vice, Naveed Akhtar, Richard Hartley, and Ajmal Mian. Quantifying bias in text-to-image generative models. arXiv, 2023. 3 +[57] Jialu Wang, Xinyue Liu, Zonglin Di, Yang Liu, and Xin Wang. T2iat: Measuring valence and stereotypical biases in text-to-image generation. In Findings ACL, pages 2560-2574, 2023. 3 +[58] Ruoyu Wang, Yongqi Yang, Zhihao Qian, Ye Zhu, and Yu Wu. Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation. In ICLR, 2024. 2 +[59] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 5 +[60] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In CVPR, pages 1900-1910, 2023. 5 +[61] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In CVPR, pages 6327-6336, 2024. 3 + +[62] Hu Yu, Hao Luo, Fan Wang, and Feng Zhao. Uncovering the text embedding in text-to-image diffusion models. arXiv, 2024. 5 +[63] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Iti-gen: Inclusive text-to-image generation. In ICCV, pages 3969-3980, 2023. 3 +[64] Ye Zhu, Yu Wu, Zhiwei Deng, Olga Russakovsky, and Yan Yan. Boundary guided learning-free semantic control with diffusion models. NeurIPS, 36:78319-78346, 2023. 2 +[65] Ye Zhu, Yu Wu, Nicu Sebe, and Yan Yan. Vision+ x: A survey on multimodal learning in the light of data. TPAMI, 2024. 1 \ No newline at end of file diff --git a/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/images.zip b/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cc186847d0cc985629bfe4c0a8fafa05d72f108f --- /dev/null +++ b/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:643054dfe153ba89b56fd27442f411e19a5fd926bdc2dc0036af39da3759d7c8 +size 944744 diff --git a/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/layout.json b/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c16ecdd7bb28979bdff7b0570840bf7d5e6bb1ee --- /dev/null +++ b/2025/Implicit Bias Injection Attacks against Text-to-Image Diffusion Models/layout.json @@ -0,0 +1,8322 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 523, + 121 + ], + "type": "text", + "content": "Implicit Bias Injection Attacks against Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "spans": [ + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "text", + "content": "Huayang Huang" + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "text", + "content": " Xiangye Jin" + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "text", + "content": " Jiaxu Miao" + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "text", + "content": " Yu Wu" + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "text", + "content": "School of Computer Science, Wuhan University \n" + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "text", + "content": "School of Mathematics and Statistics, Wuhan University \n" + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 126, + 141, + 483, + 200 + ], + "type": "text", + "content": "School of Cyber Science and Technology, Sun Yat-sen University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 207, + 480, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 207, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 127, + 207, + 480, + 220 + ], + "type": "text", + "content": "{hyhuang, xiangyejin, wuyucs}@whu.edu.cn miaojx@mail.sysu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 239, + 230, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 239, + 230, + 252 + ], + "spans": [ + { + "bbox": [ + 108, + 239, + 230, + 252 + ], + "type": "text", + "content": "For developing countries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 264, + 248, + 343, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 248, + 343, + 262 + ], + "spans": [ + { + "bbox": [ + 264, + 248, + 343, + 262 + ], + "type": "text", + "content": "Text-to-Image" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 379, + 239, + 496, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 239, + 496, + 251 + ], + "spans": [ + { + "bbox": [ + 379, + 239, + 496, + 251 + ], + "type": "text", + "content": "For developed countries" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 65, + 288, + 144, + 445 + ], + "blocks": [ + { + "bbox": [ + 82, + 278, + 121, + 288 + ], + "lines": [ + { + "bbox": [ + 82, + 278, + 121, + 288 + ], + "spans": [ + { + "bbox": [ + 82, + 278, + 121, + 288 + ], + "type": "text", + "content": "A doctor" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 288, + 144, + 445 + ], + "lines": [ + { + "bbox": [ + 65, + 288, + 144, + 445 + ], + "spans": [ + { + "bbox": [ + 65, + 288, + 144, + 445 + ], + "type": "image", + "image_path": "7463dc4505728574c63eea2674a0a920eeac47b58c03d3492a42e0ace17ffaf1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 457, + 553, + 481 + ], + "lines": [ + { + "bbox": [ + 55, + 457, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 553, + 481 + ], + "type": "text", + "content": "Figure 1. Implicit bias injection causes the model to generate subtly negative content (left) and positive content (right) for different users. Although the generated images meet the text requirement, the subtle differences may, over time, influence user perception." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 143, + 289, + 220, + 445 + ], + "blocks": [ + { + "bbox": [ + 154, + 279, + 205, + 289 + ], + "lines": [ + { + "bbox": [ + 154, + 279, + 205, + 289 + ], + "spans": [ + { + "bbox": [ + 154, + 279, + 205, + 289 + ], + "type": "text", + "content": "A professor" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 289, + 220, + 445 + ], + "lines": [ + { + "bbox": [ + 143, + 289, + 220, + 445 + ], + "spans": [ + { + "bbox": [ + 143, + 289, + 220, + 445 + ], + "type": "image", + "image_path": "a2d1608fba693f0027435918a98db739fd8e4ab56af1a0945fb4a195925f2e8c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 219, + 289, + 299, + 445 + ], + "blocks": [ + { + "bbox": [ + 233, + 278, + 282, + 289 + ], + "lines": [ + { + "bbox": [ + 233, + 278, + 282, + 289 + ], + "spans": [ + { + "bbox": [ + 233, + 278, + 282, + 289 + ], + "type": "text", + "content": "A city park" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 219, + 289, + 299, + 445 + ], + "lines": [ + { + "bbox": [ + 219, + 289, + 299, + 445 + ], + "spans": [ + { + "bbox": [ + 219, + 289, + 299, + 445 + ], + "type": "image", + "image_path": "befa5ac40f50c8ab9f6cd98ae70cdf4865a3c3f8732dbadd15055552fb0b67ce.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 313, + 289, + 390, + 444 + ], + "blocks": [ + { + "bbox": [ + 334, + 279, + 371, + 289 + ], + "lines": [ + { + "bbox": [ + 334, + 279, + 371, + 289 + ], + "spans": [ + { + "bbox": [ + 334, + 279, + 371, + 289 + ], + "type": "text", + "content": "A doctor" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 289, + 390, + 444 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 390, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 390, + 444 + ], + "type": "image", + "image_path": "fa671134f618e2e6f2a831ce8c6fdf61dd69a66e4904456482f79e93a8134f9d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 390, + 289, + 467, + 444 + ], + "blocks": [ + { + "bbox": [ + 403, + 279, + 453, + 289 + ], + "lines": [ + { + "bbox": [ + 403, + 279, + 453, + 289 + ], + "spans": [ + { + "bbox": [ + 403, + 279, + 453, + 289 + ], + "type": "text", + "content": "A professor" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 390, + 289, + 467, + 444 + ], + "lines": [ + { + "bbox": [ + 390, + 289, + 467, + 444 + ], + "spans": [ + { + "bbox": [ + 390, + 289, + 467, + 444 + ], + "type": "image", + "image_path": "6ace4558e8f1814173f8d08e828394ee47fddf53d42b3fa35be4351ebc852014.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 465, + 289, + 544, + 444 + ], + "blocks": [ + { + "bbox": [ + 481, + 279, + 529, + 289 + ], + "lines": [ + { + "bbox": [ + 481, + 279, + 529, + 289 + ], + "spans": [ + { + "bbox": [ + 481, + 279, + 529, + 289 + ], + "type": "text", + "content": "A city park" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 465, + 289, + 544, + 444 + ], + "lines": [ + { + "bbox": [ + 465, + 289, + 544, + 444 + ], + "spans": [ + { + "bbox": [ + 465, + 289, + 544, + 444 + ], + "type": "image", + "image_path": "7fcb3437fd40df3fce6e187abbf577f1e50438e19dc4a5e3429614e58d3ce77f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 152, + 490, + 200, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 490, + 200, + 502 + ], + "spans": [ + { + "bbox": [ + 152, + 490, + 200, + 502 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 54, + 515, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 515, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 515, + 297, + 696 + ], + "type": "text", + "content": "The proliferation of text-to-image diffusion models (T2I DMs) has led to an increased presence of AI-generated images in daily life. However, biased T2I models can generate content with specific tendencies, potentially influencing people's perceptions. Intentional exploitation of these biases risks conveying misleading information to the public. Current research on bias primarily addresses explicit biases with recognizable visual patterns, such as skin color and gender. This paper introduces a novel form of implicit bias that lacks explicit visual features but can manifest in diverse ways across various semantic contexts. This subtle and versatile nature makes this bias challenging to detect, easy to propagate, and adaptable to a wide range of scenarios. We further propose an implicit bias injection attack framework (IBI-Attacks) against T2I diffusion models by precomputing" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 491, + 556, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 556, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 556, + 612 + ], + "type": "text", + "content": "a general bias direction in the prompt embedding space and adaptively adjusting it based on different inputs. Our attack module can be seamlessly integrated into pre-trained diffusion models in a plug-and-play manner without direct manipulation of user input or model retraining. Extensive experiments validate the effectiveness of our scheme in introducing bias through subtle and diverse modifications while preserving the original semantics. The strong concealment and transferability of our attack across various scenarios further underscore the significance of our approach." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 633, + 395, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 633, + 395, + 645 + ], + "spans": [ + { + "bbox": [ + 314, + 633, + 395, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Text-to-image (T2I) diffusion models [19, 34, 47-49, 65] have emerged as a powerful tool for generating highly realistic images that closely adhere to textual descriptions. The widespread adoption of these models, facilitated by public APIs [40, 54] and open-source implementations [32, 33]," + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 148, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 148, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 148, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28779" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "has led to a proliferation of AI-generated images in daily life. However, biased T2I models can generate content with a specific tendency [7, 33, 58, 64], posing a risk of conveying misleading information that could influence public perception if maliciously exploited [2, 55]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 326 + ], + "type": "text", + "content": "Consider a scenario in which a malicious T2I service provider subtly \"manipulates\" user perception by selectively generating biased content. As illustrated in Fig. 1, the same neutral prompt can yield images with varying emotional tones depending on the target audience. Content aimed at developed countries may convey a more positive tone, while content targeted at developing countries may lean towards a more negative sentiment. Such nuanced differences are difficult to detect visually, yet users may be subconsciously influenced by prolonged exposure to biased images. Furthermore, as shown in Fig. 2, such biases can extend beyond emotional aspects to include cultural and religious dimensions. This implicit bias, if intentionally leveraged, could subtly shape user perceptions over time, reinforcing existing divides and nudging audiences into self-reinforcing emotional or ideological spheres." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 327, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 495 + ], + "type": "text", + "content": "Existing methods for bias exploitation typically require expensive model fine-tuning [38] and are limited to explicit biases [20, 41] (e.g., black skin and shaved head) triggered by specific inputs, which are easily detectable. In contrast, this paper focuses on a more implicit form of bias, such as the emotion, cultural stereotypes, and religious orientation depicted in Fig. 1 and Fig. 2. Such bias is highly covert and can manifest through diverse visual representations under varying prompt inputs. By imposing \"hints\" without significantly changing the content, users can be unconsciously influenced in ways that are difficult to detect. The manifestation of the implicit bias is context-dependent and can be conveyed through multiple semantic forms, such as facial expressions, poses, backgrounds, or group behaviors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 497, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 687 + ], + "type": "text", + "content": "In this paper, we propose an implicit bias injection framework (IBI-Attacks) by precalculating a bias direction in the prompt embedding space and then dynamically adapting it based on specific user inputs. By introducing bias at the prompt embedding space, we can achieve more fine-grained control compared to traditional prompt modification. And input-aware dynamic adjustment further enables adaptive bias expression across diverse semantic contexts. Specifically, we leverage a large language model (LLM) [1, 5, 45] to generate a list of neutral prompts and rewrite them according to the specified bias. Then a bias direction vector is calculated by taking the difference between the two embedding sets. Our key finding is that this single bias direction already encodes multiple semantic expressions and can generalize across various prompt inputs by simple addition, allowing for diverse bias expressions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "However, directly applying this fixed direction vector to all prompts may result in either excessive or insufficient bias" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 140 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 140 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 140 + ], + "type": "image", + "image_path": "fdf3856899a3189bfa1f9d710fd6408f350ca787b8463683d8441f3d9250d6e6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 140, + 511, + 149 + ], + "lines": [ + { + "bbox": [ + 351, + 140, + 511, + 149 + ], + "spans": [ + { + "bbox": [ + 351, + 140, + 511, + 149 + ], + "type": "text", + "content": "\"A man is standing in front of his house and smiling.\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 150, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 316, + 150, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 220 + ], + "type": "image", + "image_path": "4387b18ad5f237dddd822c9d81f13acfa233d16af663a72b5557d5183a90bf82.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 220, + 427, + 229 + ], + "lines": [ + { + "bbox": [ + 325, + 220, + 427, + 229 + ], + "spans": [ + { + "bbox": [ + 325, + 220, + 427, + 229 + ], + "type": "text", + "content": "\"One person worships in a room.\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 220, + 539, + 229 + ], + "lines": [ + { + "bbox": [ + 447, + 220, + 539, + 229 + ], + "spans": [ + { + "bbox": [ + 447, + 220, + 539, + 229 + ], + "type": "text", + "content": "\"Two people talk at the table.\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 239, + 553, + 283 + ], + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 283 + ], + "type": "text", + "content": "Figure 2. Generated images after injecting different 'cultural', 'religious', and 'gender' biases with IBI attacks. Explicit biases, such as gender, can also benefit from our approach, enabling subtle, diverse, and adaptive gender-related features." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 306, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 554, + 389 + ], + "type": "text", + "content": "modification. To address this limitation, we further perform dynamic feature selection in the fixed bias direction based on the input prompt, enabling proper expressions in different semantic contexts. In this way, the attacker can implant bias into any pre-trained text-to-image diffusion model in a plug-and-play manner, without the need to modify the input prompt or fine-tune the model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 391, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 554, + 474 + ], + "type": "text", + "content": "We evaluate our approach on Stable Diffusion [48], using \"emotion bias\" as a case study to demonstrate the accuracy of bias injection, the diversity of bias expression, and the preservation of original semantics. The transferability of the bias direction further makes our attack applicable to a wide range of user inputs, allowing for consistent bias implantation across diverse contexts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 475, + 526, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 475, + 526, + 487 + ], + "spans": [ + { + "bbox": [ + 324, + 475, + 526, + 487 + ], + "type": "text", + "content": "In summary, our key contributions are as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 488, + 553, + 583 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 314, + 488, + 553, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 488, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 488, + 553, + 523 + ], + "type": "text", + "content": "- We consider a new type of implicit bias with subtle and multiple forms of semantic expression, which is more stealthy, challenging to detect, and easy to propagate." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 525, + 553, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 553, + 559 + ], + "type": "text", + "content": "- We find that a specific bias direction within the text embedding in T2I diffusion models already encodes multiple expressions and exhibits generalization ability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 560, + 553, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 553, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 553, + 583 + ], + "type": "text", + "content": "- We introduce an adaptive adjusting module to dynamically express the bias direction on different inputs." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 597, + 400, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 400, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 400, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Bias in Diffusion Models. Existing research [2, 3, 25, 33, 35, 53] has demonstrated the prevalence of bias in text-to-image diffusion models through extensive evaluations. Dall-Eval [9] reveals that both DALL-E [37] and Stable Diffusion [48] exhibit a tendency to generate male figures, with varying gender biases across different occupations. One study [39] also finds that over " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " detainee-related images generated by Stable Diffusion depicted dark-skinned indi" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28780" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": "viduals, despite people of color constituting less than half of the U.S. prison population [13]. Such bias can lead models to perpetuate false representations of the world when used in real-world applications. Considering this, some studies have explored detecting [8, 11, 22], evaluating [16, 56, 57], and mitigating [43, 51, 63] bias in generated content to promote fairness and diversity. However, these works primarily concentrate on traditional biases like gender and skin tone, neglecting bias types with diverse semantic expressions, such as emotions considered in this paper." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 209, + 296, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 400 + ], + "type": "text", + "content": "LLMs in Diffusion Models. Considering the exceptional language understanding and generation capabilities of LLMs [1, 5], several studies [14, 18, 31] have integrated LLMs into the text-to-image diffusion generation process to accomplish complex tasks. A series of studies utilize LLMs to guide diffusion models for multi-stage generation for better accuracy. For example, SLD [61] employs LLMs to achieve the cyclic self-correctio of the generation process for improved alignment with complex prompts. LMD [26] utilizes LLMs to generate initial layouts based on text prompts, followed by layout-conditioned generation for further refinement. Another research direction [18, 31, 36] involves integrating rich text features learned by LLMs into the diffusion model to enhance guidance capabilities. In this paper, we employ LLMs to learn a multi-semantic direction of implicit bias within the text embedding space." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "text", + "content": "Bias Exploitations. In addition to investigating inherent biases in diffusion models, some studies have explored exploiting or implanting biases to deliberately steer the generation process. Some works [50, 52] show that specific hard prompts can induce biased output. However, these methods rely on existing biases within the model and cannot achieve arbitrary bias specifications. Some bias mitigation methods [10, 21] also utilize text embedding modification, but they calculate a fixed, single semantic direction for all prompts. Most similar to ours is a concurrent work Backdoorsing Bias [38], which aims to inject specified biases into the model. However, it can only introduce explicit biases with fixed visual features through prompt modification. Additionally, it requires expensive model fine-tuning. In contrast, our attack works in a training-free manner and can adaptively implant implicit bias for any user input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "type": "text", + "content": "3. Threat Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 641, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 294, + 714 + ], + "type": "text", + "content": "Attack's Objective. The attacker aims to subtly alter the T2I model to embed implicit bias into generated content while maintaining the original output structure. To achieve this, we define two primary goals: 1) Attack Success. The attackers need to ensure that the implicit bias is successfully embedded in the model, meaning that all the images gener" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "ated by the model consistently exhibit the attacker-specified bias, with a high success rate and a subtle, implicit effect. 2) Injection Stealth. To maintain stealth in our attack, we aim to preserve the model's original generation content while allowing for the injection of specified biases. The generated content should align with user prompts but include subtle, unnoticed details that convey the bias. High semantic and structural similarity to normal outputs makes the attack difficult to detect." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 198, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 342 + ], + "type": "text", + "content": "Attacker's Capability. Our approach allows the adversary to operate without needing access to the T2I model's parameters, structure, or the user-input prompts. The attack can be achieved by model poisoning in the following scenarios: 1) Covert Module Injection: An insider could introduce a stealthy module into the model during its development phase. 2) API Exploitation: A service provider might add the module to their API, subtly altering the results for users without informing them. 3) Model Sharing: An adversary could distribute a modified version of the T2I model on platforms where users expect high-quality models, manipulating outputs for unsuspecting users." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 342, + 553, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 413 + ], + "type": "text", + "content": "Unlike LLM-assisted prompts, which require real-time API calls during inference and are easier to detect, our attack operates autonomously after deployment without additional LLM calls, making it more covert and efficient. Additionally, LLM-assisted prompts also cause significant semantic alterations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 487, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 487, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 487, + 440 + ], + "type": "text", + "content": "4. Implicit Bias Injection Attacks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 447, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 662 + ], + "type": "text", + "content": "In this section, we introduce the details of our original approach for implanting implicit bias in T2I diffusion models. The pipeline of the proposed framework is shown in Fig. 3. Our attack consists of three main stages: bias direction calculation, adaptive feature selection, and user inference. We first use a LLM to establish the relationship between implicit bias and multiple semantic expressions by rewriting sentences according to a specific bias. We then compute the bias direction vector by calculating the average difference between the original and rewritten sentences in the prompt embedding space. After that, we train an input-dependent adaptive feature selection module, which dynamically adjusts the fixed bias direction from the first stage based on the user's input prompt. Finally, during model inference, the adaptive adjustment module implants the bias into the user's prompt embedding before it is sent to the generative model, ensuring that the model's output reflects the specified bias." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 671, + 479, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 479, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 479, + 684 + ], + "type": "text", + "content": "4.1. Directional Vector Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "First, we aim to compute a difference vector representing the direction of bias in the latent space. The attacker first" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28781" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 83, + 555, + 330 + ], + "blocks": [ + { + "bbox": [ + 67, + 74, + 124, + 83 + ], + "lines": [ + { + "bbox": [ + 67, + 74, + 124, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 74, + 124, + 83 + ], + "type": "text", + "content": "User inference" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 83, + 555, + 330 + ], + "lines": [ + { + "bbox": [ + 67, + 83, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 67, + 83, + 555, + 330 + ], + "type": "image", + "image_path": "eedf888a5a8244b696d076dcd5187ff99958c44df890d63b7dabd1af770dd460.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 334, + 555, + 379 + ], + "lines": [ + { + "bbox": [ + 54, + 334, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 54, + 334, + 555, + 379 + ], + "type": "text", + "content": "Figure 3. Pipeline of our proposed IBI-Attacks. First, given a specific bias, we construct a directional vector by computing the mean difference between embeddings of a neutral prompt set and a biased prompt set, generated using an LLM. Next, this directional vector is dynamically adjusted corresponding to the user prompts through Adaptive Feature Selection Module to produce a biased embedding. Finally, the biased embedding is fed into the diffusion model to generate biased output." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 399, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 581 + ], + "type": "text", + "content": "implements an LLM model to generate " + }, + { + "bbox": [ + 55, + 399, + 296, + 581 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 399, + 296, + 581 + ], + "type": "text", + "content": " neutral prompts to create a neutral prompt set " + }, + { + "bbox": [ + 55, + 399, + 296, + 581 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{neu}} = \\{x_i^{\\mathrm{neu}}\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 399, + 296, + 581 + ], + "type": "text", + "content": ", each of which has a standard subject-verb-object structure. Meanwhile, the LLM model rewrites the neutral text prompts in " + }, + { + "bbox": [ + 55, + 399, + 296, + 581 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{neu}}" + }, + { + "bbox": [ + 55, + 399, + 296, + 581 + ], + "type": "text", + "content": " with a specific bias to form a biased prompt dataset: " + }, + { + "bbox": [ + 55, + 399, + 296, + 581 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{bias}} = \\{x_i^{\\mathrm{bias}}\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 399, + 296, + 581 + ], + "type": "text", + "content": ". We employ a rewriting strategy that prefixes nouns with appropriate adjectives to introduce specific biases. We find that this restricted rewriting approach minimizes the introduction of bias-irrelevant information at the embedding level, such as modifications to sentence structure. As a result, the offset between the rewritten and original prompt embeddings becomes more closely aligned with the intended bias. Additionally, the addition of adjectives helps preserve the original sentence semantics, preventing the destruction of the underlying content." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "text", + "content": "Afterwards, the attacker utilizes the pretrained encoder " + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\varphi(\\cdot)" + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "text", + "content": " to map input prompt " + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "text", + "content": " into text embeddings " + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "inline_equation", + "content": "v = \\varphi(x) \\in R^{D \\times L}" + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "text", + "content": " denotes the dimension of each token in latent space, and " + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "text", + "content": " denotes the number of embedded tokens for each text prompt (i.e., " + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "inline_equation", + "content": "D = 1024" + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "inline_equation", + "content": "L = 77" + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "text", + "content": " in the SD v2.1 model). With the text encoder, texts prompts from " + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{neu}}" + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{bias}}" + }, + { + "bbox": [ + 55, + 581, + 296, + 677 + ], + "type": "text", + "content": " are mapped into the latent space to obtain their representation in latent space:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 681, + 295, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 681, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 64, + 681, + 295, + 696 + ], + "type": "interline_equation", + "content": "v _ {i} ^ {\\mathrm {n e u}} = \\varphi (x _ {i} ^ {\\mathrm {n e u}}), v _ {i} ^ {\\mathrm {b i a s}} = \\varphi (x _ {i} ^ {\\mathrm {b i a s}}), i = 1, 2, \\dots , N. \\tag {1}", + "image_path": "9b0ed5fada4e09286f0a893d635e2c1d3c29c117bd8e53e3dbbeefcd203ee39f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 700, + 294, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 700, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 700, + 294, + 715 + ], + "type": "text", + "content": "In this way, we get two sets of text embedding: " + }, + { + "bbox": [ + 55, + 700, + 294, + 715 + ], + "type": "inline_equation", + "content": "\\{v_i^{\\mathrm{neu}}\\}_{i = 1}^N" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 397, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 425 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 313, + 397, + 555, + 425 + ], + "type": "inline_equation", + "content": "\\{v_i^{\\mathrm{bias}}\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 397, + 555, + 425 + ], + "type": "text", + "content": ". We then compute the mean difference between these two sets of representations as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 369, + 432, + 555, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 432, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 369, + 432, + 555, + 464 + ], + "type": "interline_equation", + "content": "v ^ {\\text {d i f f}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(v _ {i} ^ {\\text {b i a s}} - v _ {i} ^ {\\text {n e u}}\\right). \\tag {2}", + "image_path": "f131474e741d2e733c4bad59fb639bac0df698ce3f73a6bd1800e7b1b13a91b2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 472, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 544 + ], + "type": "text", + "content": "We find that this simple mean difference vector already encodes multiple semantics of the implicit bias and can be generalized across different prompts. The existence of this expressive average bias direction enables us to embed implicit bias in the output by simply adding a fixed vector to the input prompt embedding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 550, + 463, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 463, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 463, + 562 + ], + "type": "text", + "content": "4.2. Adaptive Feature Selection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 567, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 640 + ], + "type": "text", + "content": "Although the single average direction vector can already introduce some bias, we find that directly applying it to all prompts may result in either excessive or insufficient bias. Therefore, we introduce an adaptive adjustment module to dynamically alter the precomputed bias direction vector " + }, + { + "bbox": [ + 313, + 567, + 554, + 640 + ], + "type": "inline_equation", + "content": "v^{\\text{diff}}" + }, + { + "bbox": [ + 313, + 567, + 554, + 640 + ], + "type": "text", + "content": " according to the specific user input " + }, + { + "bbox": [ + 313, + 567, + 554, + 640 + ], + "type": "inline_equation", + "content": "v^{\\text{user}}" + }, + { + "bbox": [ + 313, + 567, + 554, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "Module Structure. Drawing inspiration from SENet [17], we design a simple module that computes an attention map on the bias direction vector for feature selection. For the text embedding vectors with dimension " + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "inline_equation", + "content": "D \\times L" + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": ", we alternately compute feature attention" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28782" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "for the token dimension " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " and the embedding dimension " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ". Specifically, we alternately reduce one dimension to a single value using global average pooling " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\mathrm{Avg}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " to capture the overall information of that dimension. A two-layer multilayer perceptron " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " is then used to learn attention for the other, uncompressed dimension, which is used to adaptively scale the original feature along that axis. " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " is the learnable parameter of MLP." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "text", + "content": "By prioritizing key information, we can focus on relevant features and disregard irrelevant ones based on different prompt inputs. Specifically, we compute attention across both the token and feature dimensions of the text embedding by compressing the other dimension separately. The average direction vector " + }, + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "inline_equation", + "content": "v^{\\text{diff}}" + }, + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "text", + "content": " is then scaled according to the attention values in both dimensions, producing an adaptive bias direction " + }, + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\tilde{v}^{\\text{diff}}" + }, + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "text", + "content": " that is added to the user input prompt " + }, + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "inline_equation", + "content": "v^{\\text{user}}" + }, + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "text", + "content": " to generate the biased embedding " + }, + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\tilde{v}^{\\text{biased}}" + }, + { + "bbox": [ + 55, + 168, + 296, + 276 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 97, + 284, + 295, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 284, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 97, + 284, + 295, + 298 + ], + "type": "interline_equation", + "content": "\\tilde {v} ^ {\\text {d i f f}} = \\operatorname {M L P} _ {\\theta} (\\operatorname {A v g} (v ^ {\\text {u s e r}})) \\odot v ^ {\\text {d i f f}}, \\tag {3}", + "image_path": "a5f25ae2322c4cddc82f6d703926e3a0427908bebc78d6bb362fe6340f148b24.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 318, + 294, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 318, + 294, + 332 + ], + "spans": [ + { + "bbox": [ + 126, + 318, + 294, + 332 + ], + "type": "interline_equation", + "content": "\\tilde {v} ^ {\\text {b i a s}} = v ^ {\\text {u s e r}} + \\tilde {v} ^ {\\text {d i f f}}. \\tag {4}", + "image_path": "9233fbc7dad759ed8c28c02910be431a996a4a21c4b15da307f12223cc57c44a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 346, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 408 + ], + "type": "text", + "content": "Training Paradigm. We reuse the neutral and biased prompt embedding sets generated by LLM when calculating the bias direction to train the adaptive module. We rearranged them in pairs: " + }, + { + "bbox": [ + 55, + 346, + 296, + 408 + ], + "type": "inline_equation", + "content": "\\{v_i^{\\mathrm{neu}}, v_i^{\\mathrm{biased}}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 55, + 346, + 296, + 408 + ], + "type": "text", + "content": " as input and label in training the adaptive module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 408, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 469 + ], + "type": "text", + "content": "The adaptive module dynamically fine-tunes the average bias direction " + }, + { + "bbox": [ + 55, + 408, + 296, + 469 + ], + "type": "inline_equation", + "content": "v_{i}^{\\text{diff}}" + }, + { + "bbox": [ + 55, + 408, + 296, + 469 + ], + "type": "text", + "content": " based on the neutral prompt embedding " + }, + { + "bbox": [ + 55, + 408, + 296, + 469 + ], + "type": "inline_equation", + "content": "v_{i}^{\\text{neu}}" + }, + { + "bbox": [ + 55, + 408, + 296, + 469 + ], + "type": "text", + "content": ", enabling the adjusted direction to point from the neutral prompt embedding " + }, + { + "bbox": [ + 55, + 408, + 296, + 469 + ], + "type": "inline_equation", + "content": "v_{i}^{\\text{neu}}" + }, + { + "bbox": [ + 55, + 408, + 296, + 469 + ], + "type": "text", + "content": " toward the biased prompt embedding " + }, + { + "bbox": [ + 55, + 408, + 296, + 469 + ], + "type": "inline_equation", + "content": "v_{i}^{\\text{biased}}" + }, + { + "bbox": [ + 55, + 408, + 296, + 469 + ], + "type": "text", + "content": ". Thus we define the following loss:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 478, + 294, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 478, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 59, + 478, + 294, + 520 + ], + "type": "interline_equation", + "content": "L o s s = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| v _ {i} ^ {\\text {d i f f}} - \\operatorname {M L P} _ {\\theta} \\left(\\operatorname {A v g} \\left(v _ {i} ^ {\\text {n e u}}\\right)\\right) \\odot v ^ {\\text {d i f f}} \\right\\| ^ {2}, \\tag {5}", + "image_path": "f77780dd709965dbc01a5c24ba3738402afd45a3393cd613e948b461822a1030.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 520, + 185, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 185, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 185, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 520, + 185, + 534 + ], + "type": "inline_equation", + "content": "v_{i}^{\\text{diff}} = v_{i}^{\\text{biased}} - v_{i}^{\\text{neu}}" + }, + { + "bbox": [ + 55, + 520, + 185, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 540, + 147, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 147, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 147, + 552 + ], + "type": "text", + "content": "4.3. User Inference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Once the adaptation module is trained, it can be seamlessly inserted behind the text encoder of a pre-trained diffusion model. This model can be selectively deployed to specific groups. When users download the biased model, they interact with it by entering a text prompt input. This prompt is processed by the encoder and then the adaptive module to generate the biased version. Finally, the modified embedding is fed into the T2I model, producing a final output image with the desired bias. Due to the subtlety of the bias introduction and its general applicability across different prompts, the model can continuously generate imperceptibly biased content for the target user, potentially influencing their cognition. An another harmful scenario could" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "involve API service providers selectively targeting specific IP addresses, delivering normal results to most users while subtly implanting biases for particular groups, thereby conveying misleading information. The plug-and-play nature of our attack makes such targeted attacks easier to implement, posing a serious risk of spreading harmful bias." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 157, + 395, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 395, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 395, + 171 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 178, + 374, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 374, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 374, + 190 + ], + "type": "text", + "content": "5.1. Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 196, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 555, + 327 + ], + "type": "text", + "content": "Models and Datasets. Our evaluation is conducted on the COCO dataset [27] and Stable Diffusion 2.1 [48]. From the COCO dataset, we filtered 200,000 caption subsets related to people, 57,000 caption subsets related to animals, and 60,000 caption subsets related to natural scenes, ensuring that these categories do not overlap. Our primary focus is on captions related to human scenes, as this represents a common scenario for bias attacks. Furthermore, we observe that the bias embedding module, trained on human scenes, exhibits zero-shot transferability to other scenes, such as those involving animals and nature." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 346, + 556, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 556, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 556, + 491 + ], + "type": "text", + "content": "Implementation Details. To calculate the average bias direction, we employed ChatGPT-4 [42] to generate 200 neutral prompts related to everyday person scenes, along with their biased counterparts. Given that minor text prompt modifications can significantly impact embeddings [60, 62], and unrestricted rewriting rules may alter sentence structure, we simplified our rewriting rules to focus solely on the addition of appropriate adjectives to the original sentence. This approach minimizes the influence of bias-irrelevant factors on text embeddings. The adaptive adjusting module was trained on the same LLM-generated data for 50 epochs using the Adam optimizer [23] with a learning rate of 0.001." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 509, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 628 + ], + "type": "text", + "content": "Evaluation Metrics. We mainly consider two metrics: the success rate of bias injection and the stealthiness of the implanted implicit bias. Given the subtle and diverse semantic expressions of bias in our attacks, simple visual classification or recognition models are insufficient for detection. Therefore, following existing approaches [11, 38], we employ a Multi-modal Language Model (MLLM) to identify the presence of bias in the final generated image by question-answering based on visual content. We utilize the latest LLaVA-v1.6 7B [28-30] model in our experiments." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "To evaluate the imperceptibility of bias implantation, we use the CLIP score [46] to measure the alignment between the generated image and the input text " + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{CLIP}_{\\mathrm{txt - img}})" + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": ", as well as the semantic similarity to the original generated image " + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{CLIP}_{\\mathrm{img - img}})" + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": ". Additionally, we use SSIM (Structural Similarity Index) [59] to assess pixel-level differences between the images before and after bias injection. Validation" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28783" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 526, + 180 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 526, + 180 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 526, + 180 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 526, + 180 + ], + "type": "table", + "html": "
Bias typeMethodsNegativePositiveSameCLIPtxt-Img↑CLIPimg-Img↑SSIM↑FID↓PickScore↑
Original-14.4%6.7%78.9%0.36331.00001.000039.29721.805
Negative biasExplicit89.1%6.9%4.0%0.34580.73540.522140.09021.386
IBI w/o adap82.2%12.5%5.3%0.35960.81430.588039.21721.593
IBI80.2%12.8%7.0%0.36370.87930.698739.13521.766
Positive biasExplicit0.0%98.0%2.0%0.34490.76960.522142.34321.763
IBI w/o adap13.9%85%1.1%0.35810.80300.577639.54921.606
IBI14.6%83.7%1.7%0.36020.82470.604039.49921.670
", + "image_path": "85a5ade758f06974c66f8930adf8d4998e32c687d2846b80aafde289751a9c5e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 189, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 189, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 555, + 224 + ], + "type": "text", + "content": "Table 1. MLLM evaluation of “Negative” and “Positive” bias injection. Original displays the results of splicing two non-attacked, normally generated images. Explicit shows the results of introducing bias by directly modifying user prompts. IBI w/o adap illustrates the impact of directly applying the average direction vector. IBI presents the results obtained using the adaptive module for the average direction vector." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 233, + 555, + 446 + ], + "blocks": [ + { + "bbox": [ + 56, + 233, + 555, + 446 + ], + "lines": [ + { + "bbox": [ + 56, + 233, + 555, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 555, + 446 + ], + "type": "image", + "image_path": "ab57f7efa3aa6d8deae47d0f67452d809f5802cd04ddcc10ed17c79aa0dac528.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 455, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 55, + 455, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 555, + 499 + ], + "type": "text", + "content": "Figure 4. Generated samples after injecting positive and negative bias. The key modifications are enlarged. Emotional bias is conveyed through small adjustments in a person's expression or posture. In the second column, after injecting positive bias, the girl's hands transition from a defensive crossed position to a relaxed stance, and a smile appears on her face. While for the negative mood, the girl maintains her original pose, but her facial expression becomes more serious." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 520, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 568 + ], + "type": "text", + "content": "was performed on a randomly selected set of 1,000 neutral human-related captions. We also employ FID [15] and PickScore [24] to assess the preservation of image quality, evaluated on 5,000 generated images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 576, + 192, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 192, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 192, + 588 + ], + "type": "text", + "content": "5.2. Performance Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Quantitative Results. For the \"negative\" and \"positive\" sentiment bias, we instruct the MLLM to classify the modified image as more positive, more negative, or the same as the original image. The success rate was measured by the proportion of \"more negative\" for negative bias and \"more positive\" for positive bias. The results in Tab. 1 indicate that the average bias direction can already introduce bias into the image, increasing the \"More negative\" rate for \"negative\" bias from " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "14.4\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "82.2\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ". However, uniformly modifying all prompts without discrimination leads to a substantial" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 520, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 628 + ], + "type": "text", + "content": "degradation of image quality. In contrast, the adaptive adjusting module effectively preserves the original semantics while maintaining the ability to introduce bias. The results are similar for the injection of \"positive\" bias. Although prompt modification achieves the highest bias implantation rate, it substantially alters the image content and exhibits poor invisibility. The small perturbations observed in FID and PickScore indicate that IBI attacks have minimal impact on image quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "content": "Qualitative Results. Fig. 4 presents the generated images embedded with negative and positive biases, where each emotion is conveyed through subtle modifications, further illustrating the concealment of our attack. Fig. 5 shows the reflection of negative bias under different prompts. The bias expression manifests in various semantic forms, en" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28784" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 250 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 250 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 250 + ], + "type": "image", + "image_path": "3e26cb44d8a99f6f2229362937d492b1f6de93d99bc3bace568d8842981ea3cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 259, + 296, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 293 + ], + "type": "text", + "content": "Figure 5. Diverse semantic expressions of negative bias under different text inputs, including facial expressions, posture, style, and environmental cues." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 315, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 387 + ], + "type": "text", + "content": "compassing facial expressions, human posture, and subtle adjustments to background objects. Directly applying the direction vector may result in too much changes (the first column) or additional content (the third column). By incorporating the adaptive module, we can subtly introduce bias while preserving the original semantics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 398, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 567 + ], + "type": "text", + "content": "Zero-shot Transferability. We further investigate the zero-shot transferability of the bias injection module trained on human scenes to other domains. The results presented in Tab. 2 demonstrate that the bias learned from human-related scenarios effectively transfers to both animal and natural scenes. In animal-related generation, the attack successfully induces a negative sentiment in " + }, + { + "bbox": [ + 55, + 398, + 296, + 567 + ], + "type": "inline_equation", + "content": "95.1\\%" + }, + { + "bbox": [ + 55, + 398, + 296, + 567 + ], + "type": "text", + "content": " of cases. In natural scene generation, the negative probability increases by " + }, + { + "bbox": [ + 55, + 398, + 296, + 567 + ], + "type": "inline_equation", + "content": "81.1\\%" + }, + { + "bbox": [ + 55, + 398, + 296, + 567 + ], + "type": "text", + "content": ". The results for 'positive' bias are similar. Fig. 6 visualizes the effects of this transfer attack. In animal-related generations, the bias manifests primarily through changes in the animal's pose and environment. For natural scene generations, the bias is expressed through more depressing backgrounds and less pleasing content." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Human Study. However, relying solely on the MLLM for evaluation may not fully capture the impact of subtle image detail differences on human perception, which is crucial for implicit bias. Therefore, we designed a human study to further evaluate the subtlety and effectiveness of our attack. We invited 24 participants to choose between two randomly shuffled images (original image and bias injected version) based on two criteria: (1) Abnormal Rate: \"Which image is more likely to contain an implanted bias?\"; and (2) Bias Injection Rate: \"Which image more effectively reflects a specified bias?\". We also offered a \"Not Sure\" option if" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 553, + 323 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 323 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 323 + ], + "type": "image", + "image_path": "7add2e72b6fef83b2a186cd1351dd3c7802696f62d4c7ac7d789a7547e8d6a48.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 331, + 555, + 364 + ], + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 364 + ], + "type": "text", + "content": "Figure 6. The generated samples of attacking animals and natural scenes with the 'negative' bias injection module learned on human scenes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 375, + 557, + 471 + ], + "blocks": [ + { + "bbox": [ + 315, + 375, + 557, + 471 + ], + "lines": [ + { + "bbox": [ + 315, + 375, + 557, + 471 + ], + "spans": [ + { + "bbox": [ + 315, + 375, + 557, + 471 + ], + "type": "table", + "html": "
TypesMethodsNegPosSameCLIPt-iCLIPi-iSSIM
P→AOriginal9.2%1.2%89.6%0.36411.00001.0000
IBI (Neg)95.1%4.9%0.0%0.35860.86440.6751
IBI (Pos)3.6%96.2%0.2%0.35670.85170.6147
P→NOriginal8.1%3.4%88.5%0.35611.00001.0000
IBI (Neg)89.2%10.8%0.0%0.35360.88470.6811
IBI (Pos)10.2%89.4%0.4%0.34820.82740.5836
", + "image_path": "ef58d02d53428c3b6eff0bb5235576ad0e4aeb9ec365531b0e8d831fa7d3be9a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 478, + 555, + 511 + ], + "lines": [ + { + "bbox": [ + 313, + 478, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 555, + 511 + ], + "type": "text", + "content": "Table 2. Performance of zero-shot attacks on other scenes using bias injection module learned from the person scene. " + }, + { + "bbox": [ + 313, + 478, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\mathrm{P}\\rightarrow \\mathrm{A}" + }, + { + "bbox": [ + 313, + 478, + 555, + 511 + ], + "type": "text", + "content": ": Person to Animal. " + }, + { + "bbox": [ + 313, + 478, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\mathrm{P}\\rightarrow \\mathrm{N}" + }, + { + "bbox": [ + 313, + 478, + 555, + 511 + ], + "type": "text", + "content": ": Person to Nature." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 533, + 553, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 553, + 556 + ], + "type": "text", + "content": "participants think the two images are difficult to distinguish. Each person was randomly assigned 20 pairs of images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 557, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 665 + ], + "type": "text", + "content": "The results of user selections are presented in Tab. 3. Our method achieves a greater degree of stealthiness, with users exhibiting only a " + }, + { + "bbox": [ + 313, + 557, + 555, + 665 + ], + "type": "inline_equation", + "content": "35.8\\%" + }, + { + "bbox": [ + 313, + 557, + 555, + 665 + ], + "type": "text", + "content": " probability of perceiving an image as biased, which is nearly identical to the " + }, + { + "bbox": [ + 313, + 557, + 555, + 665 + ], + "type": "inline_equation", + "content": "35.7\\%" + }, + { + "bbox": [ + 313, + 557, + 555, + 665 + ], + "type": "text", + "content": " of normal images. Participants have a " + }, + { + "bbox": [ + 313, + 557, + 555, + 665 + ], + "type": "inline_equation", + "content": "67.3\\%" + }, + { + "bbox": [ + 313, + 557, + 555, + 665 + ], + "type": "text", + "content": " probability of experiencing negative emotions from the attacked images. Notably, the increase in \"Not Sure\" responses also underscores the effectiveness of our approach, as most of the model's originally generated images are perceived as positive." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 672, + 424, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 424, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 424, + 685 + ], + "type": "text", + "content": "5.3. Extended Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "LLM Rephrase Rules. We evaluate the impact of different LLM rewriting strategies in the bias vector calculation" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28785" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 286, + 141 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 286, + 141 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 286, + 141 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 286, + 141 + ], + "type": "table", + "html": "
Human picks as abnormal
MethodsInjected ↓OriginalNot Sure
Explicit57.5%22.5%20.0%
IBI w/o adapt47.5%25.0%27.5%
IBI35.8%35.7%28.5%
", + "image_path": "ae4eed387b3b132f64af74ef2cbdc3b3ae30c24ee0e94fdf1e274b4b59042a17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 66, + 145, + 286, + 217 + ], + "blocks": [ + { + "bbox": [ + 66, + 145, + 286, + 217 + ], + "lines": [ + { + "bbox": [ + 66, + 145, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 66, + 145, + 286, + 217 + ], + "type": "table", + "html": "
Human picks as negative
MethodsInjected ↑OriginalNot Sure
Explicit75.5%17.5%5.0%
IBI w/o adapt72.8%20.6%6.6%
IBI67.3%8.9%23.8%
", + "image_path": "276ce480a87f31c71e478a7ee03fdf779e15649284a4f91c998fc7250e5974e7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 251, + 293, + 430 + ], + "blocks": [ + { + "bbox": [ + 59, + 251, + 293, + 430 + ], + "lines": [ + { + "bbox": [ + 59, + 251, + 293, + 430 + ], + "spans": [ + { + "bbox": [ + 59, + 251, + 293, + 430 + ], + "type": "image", + "image_path": "42ba6c4bc2ee90368480b47e43e9110dd2047d4b7e56017fa313c93c70a650bb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 441, + 295, + 498 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 498 + ], + "type": "text", + "content": "Figure 7. Bias vector learned by various LLM rephrase rules. " + }, + { + "bbox": [ + 55, + 441, + 295, + 498 + ], + "type": "inline_equation", + "content": "LLM_{adj}" + }, + { + "bbox": [ + 55, + 441, + 295, + 498 + ], + "type": "text", + "content": ": directly using LLM to add adjectives in front of all nouns. " + }, + { + "bbox": [ + 55, + 441, + 295, + 498 + ], + "type": "inline_equation", + "content": "\\Delta LLM_{random}" + }, + { + "bbox": [ + 55, + 441, + 295, + 498 + ], + "type": "text", + "content": ": LLM rewriting without restrictions and then calculating the average direction. " + }, + { + "bbox": [ + 55, + 441, + 295, + 498 + ], + "type": "inline_equation", + "content": "\\Delta LLM_{adj}" + }, + { + "bbox": [ + 55, + 441, + 295, + 498 + ], + "type": "text", + "content": ": LLM rewriting with added adjectives and calculating the average direction." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "text", + "content": "phase. The sample results are shown in Fig. 7. If we directly employ the LLM rewritten prompts " + }, + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "(LLM_{\\mathrm{adj}})" + }, + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "text", + "content": ", the generated images can deviate significantly from the original intent. Moreover, highly negative prompts may result in numerous low-quality, easily detectable gray images. Alternatively, if we leverage unrestricted LLM rewriting " + }, + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "(\\Delta LLM_{\\mathrm{random}})" + }, + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "text", + "content": ", the calculated average bias direction may lead to image distortions. Because unrestricted rewriting may introduce extraneous differences between the two prompt sets, such as sentence structure, hindering the accurate representation of meaningful bias modifications by calculating the average direction. In our approach, a simple approach of adding adjectives " + }, + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "(\\Delta LLM_{\\mathrm{adj}})" + }, + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "text", + "content": " effectively minimizes the introduction of bias-irrelevant information, leading to the successful learning of a direction that accurately represents the bias." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 550, + 169 + ], + "blocks": [ + { + "bbox": [ + 106, + 226, + 244, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 244, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 244, + 236 + ], + "type": "text", + "content": "Table 3. Results of human evaluation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 70, + 550, + 169 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 550, + 169 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 550, + 169 + ], + "type": "table", + "html": "
ModelAttackNeg ↑PosSameCLIPt-i ↑PS ↑
SD3 [12]-12.4%12.0%75.6%0.35722.529
SD3 [12]IBI68.2%11.8%20.0%0.35922.467
FLUX [4]-8.0%21.2%70.8%0.34322.354
FLUX [4]IBI65.4%34.4%0.2%0.34222.335
MUSE [6]-10.8%16.6%72.6%0.32420.553
MUSE [6]IBI91.2%8.6%0.2%0.32220.477
", + "image_path": "9fa69868b967722e40d2937f06e77ef2dc0140d9b707505ff88d81ca81bc5ff9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 320, + 212, + 547, + 268 + ], + "blocks": [ + { + "bbox": [ + 313, + 178, + 553, + 200 + ], + "lines": [ + { + "bbox": [ + 313, + 178, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 553, + 200 + ], + "type": "text", + "content": "Table 4. MLLM evaluation of 'negative' bias injection for Stable Diffusion-3, FLux-schnell, and Muse models. PS: PickScore." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 212, + 547, + 268 + ], + "lines": [ + { + "bbox": [ + 320, + 212, + 547, + 268 + ], + "spans": [ + { + "bbox": [ + 320, + 212, + 547, + 268 + ], + "type": "table", + "html": "
DefenseNeg ↑PosSameCLIPt-i ↑PS ↑
-80.2%12.8%7.0%0.36421.766
Chuang et al. [10]78.2%14.0%7.8%0.35821.419
Kim et al. [21]79.6%16.4%4.0%0.13918.523
", + "image_path": "9ad12a0edbb9dd49d2d11960af216ad6bbf697549a7b69e4d69d987549a5eb86.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 277, + 546, + 288 + ], + "lines": [ + { + "bbox": [ + 321, + 277, + 546, + 288 + ], + "spans": [ + { + "bbox": [ + 321, + 277, + 546, + 288 + ], + "type": "text", + "content": "Table 5. Performance of IBI attacks under debiasing methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 308, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 406 + ], + "type": "text", + "content": "Generalization to Latest Models. We further evaluated negative bias injection on Stable Diffusion-3 [12], Flux-schnell (rectified flow transformer) [4], and MUSE (autoregressive) [6]. The results presented in Tab. 4 demonstrate that IBI generalizes effectively across models with varying architectures, significantly increasing the proportion of negative images while preserving the original semantics and maintaining the model's generative capabilities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "text", + "content": "Robustness against Debiasing Methods. We evaluated the effectiveness of our attack after applying two debiasing methods [10, 21]. The approach by Chuang et al. [10] works by projecting out certain biased word directions in the text embedding, while Kim et al. [21] achieves bias control by modifying the end-of-sequence token. Tab. 5 shows they have little defense effect, as word-level debiasing is ineffective against our sentence-level bias. Excessive text embedding modifications also degrade image quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a novel implicit bias injection attack method that enables effective and versatile bias injection through multiple semantic expressions. Our findings reveal that bias in latent space can be represented as a high-dimensional vector, which can be integrated into the original prompt embedding to produce a modified embedding containing the injected bias. Evaluation and human studies demonstrate that our approach achieves high levels of injection covertness while maintaining robust bias injection. This work offers valuable insights into the representation and manipulation of bias within latent spaces, paving the way for further advancements in bias explanation and mitigation methods in generative AI." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28786" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "content": "This work was partially supported by the National Natural Science Foundation of China under Grant 62372341 and the Key Program of the National Natural Science Foundation of China under Grant 62436007." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 171, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 215 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 217, + 296, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 272 + ], + "type": "text", + "content": "[2] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In ACM FAccT. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 274, + 295, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 295, + 305 + ], + "type": "text", + "content": "[3] Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion's den: Investigating hate in multimodal datasets. NeurIPS, 36, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 295, + 328 + ], + "type": "text", + "content": "[4] Black Forest Labs. Flux.1 [schnell], 2024. Accessed: 14 Mar. 2025. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 331, + 296, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 296, + 385 + ], + "type": "text", + "content": "[5] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, pages 1877-1901, 2020. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 387, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 296, + 442 + ], + "type": "text", + "content": "[6] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In ICML, pages 4055-4075, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 443, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 296, + 485 + ], + "type": "text", + "content": "[7] Tianwei Chen, Yusuke Hirota, Mayu Otani, Noa Garcia, and Yuta Nakashima. Would deep generative models amplify bias in future models? In CVPR, pages 10833-10843, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 488, + 296, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 532 + ], + "type": "text", + "content": "[8] Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij, Kartik Hosanagar, Leonid Sigal, and Matthew Turk. Tibet: Identifying and evaluating biases in text-to-image generative models. arXiv, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 534, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 296, + 576 + ], + "type": "text", + "content": "[9] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In CVPR, pages 3043-3054, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 578, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 296, + 623 + ], + "type": "text", + "content": "[10] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 3, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "type": "text", + "content": "[11] Moreno D'Inca, Elia Peruzzo, Massimiliano Mancini, Dejia Xu, Vidit Goel, Xingqian Xu, Zhangyang Wang, Humphrey Shi, and Nicu Sebe. Openbias: Open-set bias detection in text-to-image generative models. In CVPR, pages 12225-12235, 2024. 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "text", + "content": "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "fied flow transformers for high-resolution image synthesis. In ICML, 2024. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 97, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 118 + ], + "type": "text", + "content": "[13] Federal Bureau of Prisons. Inmate race statistics, 2023. Accessed: 2024-11-01. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 120, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 120, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 120, + 553, + 163 + ], + "type": "text", + "content": "[14] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. In ICLR, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 166, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 166, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 166, + 553, + 209 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 211, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 211, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 211, + 553, + 266 + ], + "type": "text", + "content": "[16] Phillip Howard, Avinash Madasu, Tiep Le, Gustavo Lujan Moreno, Anahita Bhiwandiwalla, and Vasudev Lal. Social-counterfactuals: Probing and mitigating intersectional social biases in vision-language models with counterfactual examples. In CVPR, pages 11975-11985, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 268, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 553, + 289 + ], + "type": "text", + "content": "[17] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 291, + 553, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 324 + ], + "type": "text", + "content": "[18] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "type": "text", + "content": "[19] Huayang Huang, Yu Wu, and Qian Wang. Robin: Robust and invisible watermarks for diffusion models with adversarial optimization. NeurIPS, 37:3937-3963, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "text", + "content": "[20] Naveed Akhtar Jordan Vice, Richard Hartley, and Ajmal Mian. Severity controlled text-to-image generative model bias manipulation. arXiv preprint arXiv:2404.02530, 2024. 2, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 406, + 553, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 553, + 438 + ], + "type": "text", + "content": "[21] Eunji Kim, Siwon Kim, Rahim Entezari, and Sungroh Yoon. Unlocking intrinsic fairness in stable diffusion. arXiv preprint arXiv:2408.12692, 2024. 3, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 440, + 553, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 440, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 316, + 440, + 553, + 484 + ], + "type": "text", + "content": "[22] Younghyun Kim, Sangwoo Mo, Minkyu Kim, Kyungmin Lee, Jaeho Lee, and Jinwoo Shin. Discovering and mitigating visual biases through keyword explanation. In CVPR, pages 11082-11092, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 486, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 486, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 486, + 553, + 507 + ], + "type": "text", + "content": "[23] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv, 2014. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "text", + "content": "[24] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. NeurIPS, 36:36652-36663, 2023. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "type": "text", + "content": "[25] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. NeurIPS, 36, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "type": "text", + "content": "[26] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. TMLR, 2024. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 5" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28787" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 128 + ], + "type": "text", + "content": "[30] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[31] Mushui Liu, Yuhang Ma, Xinfeng Zhang, Yang Zhen, Zeng Zhao, Zhipeng Hu, Bai Liu, and Changjie Fan. Llm4gen: Leveraging semantic representation of llms for text-to-image generation. arXiv, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "text", + "content": "[32] Xiao Liu, Xiaoliu Guan, Yu Wu, and Jiaxu Miao. Iterative ensemble training with anti-gradient control for mitigating memorization in diffusion models. In ECCV, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "type": "text", + "content": "[33] Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Evaluating societal representations in diffusion models. NeurIPS, 36, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "type": "text", + "content": "[35] Hanjun Luo, Ziye Deng, Ruizhe Chen, and Zuozhu Liu. Faintbench: A holistic and precise benchmark for bias evaluation in text-to-image models. arXiv, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 317, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 348 + ], + "type": "text", + "content": "[36] Bingqi Ma, Zhuofan Zong, Guanglu Song, Hongsheng Li, and Yu Liu. Exploring the role of large language models in prompt encoding for diffusion models. arXiv, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 350, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 380 + ], + "type": "text", + "content": "[37] Ranjita Naik and Besmira Nushi. Social biases through the text-to-image generation lens. In AIES, pages 786-808, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "text", + "content": "[38] Ali Naseh, Jaechul Roh, Eugene Bagdasaryan, and Amir Houmansadr. Injecting bias in text-to-image models via composite-trigger backdoors. arXiv, 2024. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 416, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 436 + ], + "type": "text", + "content": "[39] Leonardo Nicoletti and Dina Bass. Humans are biased. generative ai is even worse, 2023. Accessed: 2024-11-01. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "text", + "content": "[40] Fabian Offert and Thao Phan. A sign that spells: Dall-e 2, invisual images and the racial politics of feature space. arXiv, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "type": "text", + "content": "[41] Carolina López Olmos, Alexandros Neophytou, Sunando Sengupta, and Dim P Papadopoulos. Latent directions: A simple pathway to bias mitigation in generative ai. In CVPR, pages 8333-8337, 2024. 2, 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 516, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 516, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 516, + 294, + 536 + ], + "type": "text", + "content": "[42] OpenAI. Gpt-4 technical report, 2023. Accessed: 2024-11-01. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[43] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models. In CVPR, pages 6668-6678, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[44] Dustin Podell, *Zion English*, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. *Sdxl: Improving latent diffusion models for high-resolution image synthesis*. In *ICLR*, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "text", + "content": "[45] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 659, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 659, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 659, + 294, + 713 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[47] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv, pages arXiv-2204, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 154, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 154, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 317, + 154, + 553, + 208 + ], + "type": "text", + "content": "[49] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 210, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 210, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 210, + 553, + 243 + ], + "type": "text", + "content": "[50] Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. Towards controllable biases in language generation. In Findings EMNLP, pages 3239–3254, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 245, + 553, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 245, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 316, + 245, + 553, + 288 + ], + "type": "text", + "content": "[51] Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. In CVPR, pages 11996-12005, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 291, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 334 + ], + "type": "text", + "content": "[52] Lukas Struppek, Dom Hintersdorf, Felix Friedrich, Patrick Schramowski, Kristian Kersting, et al. Exploiting cultural biases via homoglyphs in text-to-image synthesis. JAIR, 78: 1017-1068, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 336, + 553, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 336, + 553, + 379 + ], + "spans": [ + { + "bbox": [ + 316, + 336, + 553, + 379 + ], + "type": "text", + "content": "[53] Md Mehrab Tanjim, Krishna Kumar Singh, Kushal Kafle, Ritwik Sinha, and Garrison W Cottrell. Discovering and mitigating biases in clip-based image editing. In WACV, pages 2984-2993, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 382, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 425 + ], + "type": "text", + "content": "[54] TIME. Ethical ai isn't to blame for google's gemini debacle. https://time.com/6836153/ethical-ai-google-gemini-debacle/, 2024. Accessed: 2024-2-29. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 428, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 553, + 472 + ], + "type": "text", + "content": "[55] Eddie Ungless, Björn Ross, and Anne Lauscher. Stereotypes and smut: The (mis) representation of non-cisgender identities by text-to-image models. In ACL, pages 7919-7942. Association for Computational Linguistics (ACL), 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 474, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 474, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 474, + 553, + 506 + ], + "type": "text", + "content": "[56] Jordan Vice, Naveed Akhtar, Richard Hartley, and Ajmal Mian. Quantifying bias in text-to-image generative models. arXiv, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "text", + "content": "[57] Jialu Wang, Xinyue Liu, Zonglin Di, Yang Liu, and Xin Wang. T2iat: Measuring valence and stereotypical biases in text-to-image generation. In Findings ACL, pages 2560-2574, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 554, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 586 + ], + "type": "text", + "content": "[58] Ruoyu Wang, Yongqi Yang, Zhihao Qian, Ye Zhu, and Yu Wu. Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation. In ICLR, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "type": "text", + "content": "[59] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "text", + "content": "[60] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In CVPR, pages 1900-1910, 2023. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[61] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In CVPR, pages 6327-6336, 2024. 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28788" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 217 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "text", + "content": "[62] Hu Yu, Hao Luo, Fan Wang, and Feng Zhao. Uncovering the text embedding in text-to-image diffusion models. arXiv, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[63] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Iti-gen: Inclusive text-to-image generation. In ICCV, pages 3969-3980, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[64] Ye Zhu, Yu Wu, Zhiwei Deng, Olga Russakovsky, and Yan Yan. Boundary guided learning-free semantic control with diffusion models. NeurIPS, 36:78319-78346, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "text", + "content": "[65] Ye Zhu, Yu Wu, Nicu Sebe, and Yan Yan. Vision+ x: A survey on multimodal learning in the light of data. TPAMI, 2024. 1" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28789" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/3b6935ad-40bc-4903-80f6-f62148a4f3d3_content_list.json b/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/3b6935ad-40bc-4903-80f6-f62148a4f3d3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..843b64cdf41b0fb4ca359cf98d45eb76f925a307 --- /dev/null +++ b/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/3b6935ad-40bc-4903-80f6-f62148a4f3d3_content_list.json @@ -0,0 +1,2099 @@ +[ + { + "type": "text", + "text": "Implicit Correspondence Learning for Image-to-Point Cloud Registration", + "text_level": 1, + "bbox": [ + 127, + 130, + 869, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinjun Li $^{1}$ , Wenfei Yang $^{1}$ , Jiacheng Deng $^{1}$ , Zhixin Cheng $^{1}$ , Xu Zhou $^{3}$ , Tianzhu Zhang $^{1,2,*}$ $^{1}$ University of Science and Technology of China", + "bbox": [ + 179, + 179, + 892, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory \n $^{3}$ Sangfor Technologies", + "bbox": [ + 176, + 215, + 892, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{1xj3017,dengjc,chengzhixin}@mail.ustc.edu.cn, {yangwf,tzzhang}@ustc.edu.cn, zhouxu@sangfor.com.cn", + "bbox": [ + 99, + 253, + 960, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image-to-point cloud registration aims to estimate the camera pose of a given image within a 3D scene point cloud. In this area, matching-based methods have achieved leading performance by first detecting the overlapping region, then matching point and pixel features learned by neural networks and finally using the PnP-RANSAC algorithm to estimate camera pose. However, achieving accurate image-to-point cloud registration remains challenging because the overlapping region detection is unreliable merely relying on point-wise classification, direct alignment of cross-modal data is difficult and indirect optimization objective leads to unstable registration results. To address these challenges, we propose a novel implicit correspondence learning method, including a Geometric Prior-guided overlapping region Detection Module (GPDM), an Implicit Correspondence Learning Module (ICLM), and a Pose Regression Module (PRM). The proposed method enjoys several merits. First, the proposed GPDM can precisely detect the overlapping region. Second, the ICLM can generate robust cross-modality correspondences. Third, the PRM can enable end-to-end optimization. Extensive experimental results on KITTI and nuScenes datasets demonstrate that the proposed model sets a new state-of-the-art performance in registration accuracy.", + "bbox": [ + 89, + 335, + 483, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 729, + 220, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image-to-point cloud registration [15] is a fundamental task in the field of computer vision, aimed at estimating the camera pose of a given image within a 3D scene point cloud. It plays an important role in many computer vision tasks, including simultaneous localization and mapping [13], 3D reconstruction [11], augmented/virtual reality [2], and visual localization [33]. Although significant progress has been made in recent years, image-to-point cloud registra", + "bbox": [ + 89, + 753, + 482, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d9729673f7150b13eec4d61f936a5ebc7c15e910f9c8f06b52af23f9e84d7467.jpg", + "image_caption": [ + "a) Existing matching-based methods" + ], + "image_footnote": [], + "bbox": [ + 550, + 304, + 867, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a56b9b234a91892c81682ce350c78aaccd2b2d3a6c66de1609a8a6a585336b97.jpg", + "image_caption": [ + "b) Our Implicit Correspondence Learning method", + "Figure 1. (a) The pipeline of mainstream matching-based methods. The green lines denote correct correspondences, and the red lines denote incorrect correspondences. 2D-3D correspondences are first established by calculating feature similarity. Then the PnP-RANSAC algorithm is applied to obtain camera pose. (b) The pipeline of our proposed method. A set of shared queries are used to aggregate information from image and point cloud to generate 2D and 3D keypoint detectors. With paired 2D and 3D keypoint detectors, we obtain a set of 2D-3D correspondences and estimate camera pose through a pose regression network." + ], + "image_footnote": [], + "bbox": [ + 550, + 439, + 866, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion remains challenging due to the inherent disparities that images capture appearances with 2D grid data while point clouds represent structures in an unordered format.", + "bbox": [ + 511, + 734, + 906, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, numerous methods have been proposed for image-to-point cloud registration [15, 19, 21, 28, 44]. Generally, these methods can be categorized into two main classes, including matching-free methods [19, 21] and matching-based methods [15, 28, 44]. Matching-free methods utilize the camera imaging principle as a prior to predict camera pose. For example, DeepI2P [21] casts the registration problem into a classi-", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16922", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fication problem. A binary classification network is designed to determine whether each scene point is within or beyond the camera frustum. Then the classification results are fed into an inverse camera projection solver to obtain camera pose. However, these methods merely determine a coarse image-point correspondence, e.g. whether a 3D point is inside the image frustum, which is insufficient for precise camera pose estimation. Different from matching-free methods, matching-based methods focus on establishing fine-grained 2D-3D correspondences and use the PnP-RANSAC algorithm [16, 20] to estimate camera pose [15, 28, 44], which have achieved superior performance. These methods generally involve three important steps: (1) A detector is designed to identify the overlapping region between the image and point cloud. (2) 2D-3D correspondences are established within the identified overlapping region. (3) The PnP-RANSAC algorithm [16, 20] is applied as a post-processing module to obtain camera pose.", + "bbox": [ + 89, + 90, + 480, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the success of these matching-based methods, there still exist three notable problems that hinder the registration accuracy of these methods. (1) Unreliable point-based detection of the overlapping region. Most existing methods adopt a point-wise classification way to determine whether each point lies in the overlapping region or not, ignoring the global projection relationship between the image and point cloud, e.g. an image should correspond to a frustum in the point cloud. Consequently, existing methods usually suffer from inaccurate boundary detection of the overlapping region, resulting in some missing 2D-3D correspondences or incorrect correspondences outside the overlapping region. (2) Inaccurate explicit matching between the image and point cloud. As shown in Figure 1 (a), matching-based methods essentially establish 2D-3D correspondences by directly calculating feature similarity between the image and point cloud. These methods employ metric learning techniques like contrastive loss [5], triplet loss [34] or circle loss [37] to force the alignment of corresponding 2D and 3D features. However, due to the inherent data discrepancy between the image and point cloud, direct alignment of cross-modal data can result in a significant number of mismatches, as shown in Figure 1 (a). (3) Indirect correspondence-based optimization objective. After establishing 2D-3D correspondences, the majority of existing approaches employ the PnP-RANSAC algorithm [16, 20] as a post-processing module to estimate camera pose as illustrated in Figure 1 (a). During the training stage, these approaches primarily consider 2D-3D correspondences as the optimization target, without optimizing the pose directly.", + "bbox": [ + 89, + 367, + 482, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the above discussions, we propose a novel implicit correspondence learning method for imaged-to-point cloud registration. The proposed model contains three modules, a Geometric Prior-guided overlapping re", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GPI Detection Module (GPDM), an Implicit Correspondence Learning Module (ICLM), and a Pose Regression Module (PRM). The GPDM leverages geometric prior that the overlapping region is a frustum to enhance the accuracy of the overlapping region detection. Specifically, a binary classification network is designed to determine whether each point lies within or beyond the camera frustum. Then, we predict center coordinate and orientation of the camera frustum using obtained classification results. Finally, we regard points in predicted camera frustum as our overlapping region. Since camera frustum defines a continuous and regular geometric region, the noise's interference on the overlapping region detection can be mitigated. The ICLM is designed to implicitly identify a set of 2D-3D correspondences within the overlapping region. As shown in Figure 1 (b), we initialize a set of learnable correspondence queries. For each query, we convert it to 2D/3D keypoint detector via the attention mechanism [40] to aggregate features from the image and point cloud, respectively. By calculating similarity between 2D/3D keypoint detector and image/point cloud features, a pair of 2D/3D keypoint heatmap is obtained. In this way, we can implicitly obtain a set of 2D-3D paired keypoints, avoiding the process of forced alignment between 2D and 3D features. Finally, the PRM takes obtained 2D/3D keypoint detectors and 2D/3D keypoints as inputs, and regresses camera pose via a Multi-Layer Perceptron (MLP), which enables the end-to-end optimization. The main contributions are summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 512 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel implicit correspondence learning method for image-to-point cloud registration, which enjoys the merits of precise overlapping region detection, robust cross-modality matching, end-to-end optimization.", + "- Without bells and whistles, extensive experimental results on KITTI and nuScenes datasets demonstrate that the proposed model sets a new state-of-the-art performance in registration accuracy, i.e. a reduction of $0.55\\mathrm{m}$ in RTE and $2.05^{\\circ}$ in RRE compared to previous SOTA method." + ], + "bbox": [ + 513, + 513, + 903, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 662, + 663, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we briefly overview methods that are related to image-to-image registration, point cloud-to-point cloud registration and image-to-point cloud registration.", + "bbox": [ + 511, + 689, + 903, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-to-image registration. Image-to-image registration is the process of aligning two or more images, which aims to find the optimal transformation that best aligns the structures of interest in the input images. Typical methods usually adopt a detect-then-match pipeline to extract correspondences. Keypoints are first detected and described using handcrafted [23, 30] or learning-based techniques [10, 14, 29, 32], correspondences are then established based on their feature similarity. However, robustly detecting repeatable keypoints across different images is still challenging, especially for texture-poor regions. Detection", + "bbox": [ + 511, + 734, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16923", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "free methods [36, 41, 45] propose to bypass the keypoint detection step by first obtaining patch-level matches and then regress pixel-wise matches in a coarse-to-fine manner, which achieves accurate image-to-image registration.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point cloud-to-point cloud registration. Point cloud-to-point cloud registration is the process of finding a rigid transformation that best aligns two point clouds. Similar to image-to-image registration, early works leverage hand-crafted descriptors [12, 31] or learning-based descriptors [1, 7-9] matching to estimate the correspondences. CoFiNet [42] follows the design of LoFTR [36] and introduces detection-free registration with a coarse-to-fine strategy. Recent methods have replaced RANSAC [16] with deep robust estimators to enhance both speed and accuracy. For example, GeoTransformer [27] proposes a geometric Transformer and introduces a local-to-global registration for RANSAC-free registration.", + "bbox": [ + 89, + 179, + 482, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-to-point cloud registration. Compared to intramodality registration, image-to-point cloud registration poses more significant difficulties due to the inherent data disparities. The typical pipeline is to first extract 2D-3D correspondences and then estimate camera pose using the PnP-RANSAC algorithm [16, 20]. In this case, the accuracy of extracted correspondences plays a crucial role in ensuring successful registration. 2D3D-MatchNet [15] is the pioneering work, which uses SIFT [23] and ISS [43] to extract the keypoints of the image and point cloud, respectively. Then the patches around the keypoints are fed to CNNs [35] and PointNet [25] to extract features for establishing 2D-3D correspondences. However, the independent hand-crafted keypoint detectors for different modalities leads to a poor registration accuracy. CorrI2P [28] predicts pixels and points in the overlapping region and learns dense 2D-3D correspondences by matching point and pixel features for fine registration. However, it struggles to learn a discriminative latent space for 2D and 3D data, leading to a high rate of incorrect correspondences. Recently, VP2PMatch [44] learns a structured cross-modality latent space by representing 3D elements using both voxels [6, 38] and points [26]. It imposes supervisions directly on the predicted pose distribution with a differentiable probabilistic PnP solver [4]. Though introducing voxel features, it struggles with 2D-3D data disparities. Besides, it fails to achieve high registration accuracy for insufficient supervisions. In addition to these matching-based methods, there are some matching-free methods. For instance, DeepI2P [21] converts the registration problem into a classification and inverse camera projection optimization problem. However, the frustum classification only indicates coarse 2D-3D correspondence, thus limiting the registration accuracy.", + "bbox": [ + 91, + 401, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a 3D scene point cloud $\\mathbf{P} \\in \\mathbb{R}^{N \\times 3}$ and an image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ from the same scene, the cross-modality image-to-point cloud registration is to estimate the camera pose $\\mathbf{T}_{gt}$ in the coordinate system of $\\mathbf{P}$ , including a rotation matrix $\\mathbf{R}_{gt} \\in \\mathbf{SO}(3)$ and a translation vector $\\mathbf{t}_{gt} \\in \\mathbb{R}^3$ .", + "bbox": [ + 511, + 113, + 903, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of our pipeline is illustrated in Figure 2. Our proposed model mainly consists of a Geometric Prior-guided overlapping region Detection Module (GPDM), an Implicit Correspondence Learning Module (ICLM) and a Pose Regression Module (PRM). Specifically, we first extract features for the image and point cloud respectively. Next, the GPDM detects the overlapping region of the input point cloud. Then the ICLM implicitly identifies a set of correspondences within detected overlapping region via our designed correspondence queries. At last, we estimate the camera pose via the PRM.", + "bbox": [ + 511, + 191, + 906, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Feature Extraction", + "text_level": 1, + "bbox": [ + 513, + 364, + 694, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the image, we use ResNet [18] with FPN [22] to extract 2D features $\\mathbf{F}_I\\in \\mathbb{R}^{H\\times W\\times C}$ . For simplicity, we denote the pixel coordinate matrices for $\\mathbf{F}_I$ as $\\mathbf{E}_I\\in \\mathbb{R}^{HW\\times 2}$ . For the input point cloud, we adopt KPFCNN [39] to extract 3D features $\\mathbf{F}_P\\in \\mathbb{R}^{N\\times C}$ .", + "bbox": [ + 511, + 386, + 905, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Geometric Prior-guided Overlapping Region Detection Module", + "text_level": 1, + "bbox": [ + 511, + 469, + 903, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To precisely detect the overlapping region of the point cloud, most methods adopt a point-wise classification strategy to determine whether each point lies in the overlapping region or not, ignoring the global projection relationship between the image and point cloud. Inspired by the fact that an image should correspond to a frustum in the point cloud, we design a geometric prior-guided classification strategy thus we can get more accurate classification results.", + "bbox": [ + 511, + 506, + 905, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in the bottom-left part of Figure 2, we first obtain 2D global feature by averaging extracted 2D features $\\mathbf{F}_I$ and repeat it to concatenate with the 3D features $\\mathbf{F}_P$ . Then a MLP is used to learn the probability $\\mathbf{O}_P \\in \\mathbb{R}^{N \\times 1}$ of a point to lie within the camera frustum.", + "bbox": [ + 511, + 627, + 905, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the in-frustum probability $\\mathbf{O}_P$ is not precise enough for in-frustum classification, we propose to estimate the frustum pose to get more accurate classification results. Utilizing $\\mathbf{O}_P$ and the coordinate information of the point cloud $\\mathbf{P}$ , we apply another MLP to regress the frustum pose $\\mathbf{T}_f \\in \\mathbf{SE}(\\mathbf{3})$ , composed of a 3D rotation $\\mathbf{R}_f \\in \\mathbf{SO}(\\mathbf{3})$ and a 3D translation vector $\\mathbf{t}_f \\in \\mathbb{R}^3$ . In detail, we apply two parallel MLPs to regress $\\mathbf{R}_f$ and $\\mathbf{t}_f$ respectively:", + "bbox": [ + 511, + 702, + 905, + 824 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {f}, \\mathbf {t} _ {f} = \\rho (\\mathrm {M L P} ([ \\mathbf {O} _ {P}, \\mathbf {P} ])), \\mathrm {M L P} ([ \\mathbf {O} _ {P}, \\mathbf {P} ]), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 830, + 903, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $[\\cdot, \\cdot]$ denotes concatenation along feature dimension. We choose a 6D representation of rotation [46] as the regression of the first MLP, for its continuous learning space", + "bbox": [ + 511, + 854, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16924", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c38f3556c18f6641645d335518a25e8f2a76a7b534c71764e073091f9489f2aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 88, + 859, + 287 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/387a90bf422d22dd848cb00a1f75d6322f0257113f9745e18e0c54c6b0a9d066.jpg", + "image_caption": [ + "Figure 2. Overview of our method. Given a pair of point cloud $\\mathbf{P}$ and image $\\mathbf{I}$ as inputs, we first extract features for the image and point cloud respectively. Then, in the geometric prior-guided overlapping region detection module, we leverage geometric prior that the overlapping region is a frustum to classify each point in the point cloud into within or beyond the camera frustum and get the point cloud mask $\\mathbf{M}_P$ . Next, in the implicit correspondence learning module, we initialize a set of correspondence queries $\\mathbf{F}_q$ and convert them to 2D keypoint detectors $\\mathbf{D}_I$ and 3D keypoint detectors $\\mathbf{D}_P$ by aggregating extracted 2D and 3D features via the attention mechanism. Weighting the pixel/point coordinates according to the feature similarity between the 2D/3D keypoint detectors and 2D/3D features, we obtain 2D keypoints $\\mathbf{K}_I$ and 3D keypoints $\\mathbf{K}_P$ , respectively. Finally, in the pose regression module, we utilize the information of 2D-3D paired keypoints and 2D-3D paired keypoint detectors to regress the camera pose $\\Delta \\mathbf{R}$ and $\\Delta t$ ." + ], + "image_footnote": [], + "bbox": [ + 135, + 287, + 859, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in $\\mathbf{SO}(3)$ , and $\\rho(\\cdot)$ denotes the transformation from the 6D representation to the $3 \\times 3$ rotation matrix $\\mathbf{R}_f$ . Then according to the given camera intrinsic matrix $\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}$ , we define the point-to-pixel projection function as:", + "bbox": [ + 89, + 609, + 483, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ x _ {i}, y _ {i}, z _ {i} \\right] ^ {\\top} = \\mathbf {K} \\cdot \\left(\\mathbf {R} _ {f} \\cdot \\mathbf {p} _ {i} + \\mathbf {t} _ {f}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 676, + 482, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} _ {i} = \\left[ u _ {x i}, u _ {y i} \\right] ^ {\\top} = \\left[ x _ {i} / z _ {i}, y _ {i} / z _ {i} \\right] ^ {\\top}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 698, + 480, + 718 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(\\mathbf {p} _ {i}; \\mathbf {K}, \\mathbf {R} _ {f}, \\mathbf {t} _ {f}\\right) = \\mathbf {u} _ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 719, + 482, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_i\\in \\mathbf{P}$ is a 3D point and $\\mathbf{u}_i$ represents a 2D pixel. We get the point cloud mask $\\mathbf{M}_p\\in \\mathbb{R}^{N\\times 1}$ as:", + "bbox": [ + 89, + 744, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {p} ^ {i} = \\left\\{ \\begin{array}{l} 1; 0 \\leq u _ {x i} < W, 0 \\leq u _ {y i} < H, z _ {i} \\geq 0 \\\\ 0; \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 784, + 482, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Implicit Correspondence Learning Module", + "text_level": 1, + "bbox": [ + 89, + 833, + 457, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining point cloud mask $\\mathbf{M}_p$ , we implicitly identify a set of robust correspondences between the image and point cloud. As shown in the bottom-right part of Figure 2,", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we first initialize $N_{q}$ correspondence queries $\\mathbf{F}_q\\in \\mathbb{R}^{N_q\\times C}$ with a set of learnable parameters. Then, we aim to convert correspondence queries $\\mathbf{F}_q$ to 2D and 3D keypoint detectors by aggregating information from both the 2D features $\\mathbf{F}_I$ and 3D features $\\mathbf{F}_P$ . For this purpose, we iteratively apply Attention-Pixel then Attention-Point as shown in Figure 2. For every iteration $k\\geq 0$ , we take queries $\\mathbf{F}_q^{2k}$ as the input and enable interaction with the 2D features $\\mathbf{F}_I$ and 3D features $\\mathbf{F}_P$ as follows:", + "bbox": [ + 511, + 608, + 906, + 747 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {q} ^ {2 k + 1} = \\operatorname {A t t e n t i o n - P i x e l} \\left(\\mathbf {F} _ {\\mathrm {q}} ^ {2 \\mathrm {k}}, \\mathbf {F} _ {\\mathrm {I}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 761, + 903, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {q} ^ {2 k + 2} = \\operatorname {A t t e n t i o n - P o i n t} \\left(\\mathbf {F} _ {\\mathrm {q}} ^ {2 k + 1}, \\mathbf {F} _ {\\mathrm {P}}, \\mathbf {M} _ {\\mathrm {p}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 782, + 903, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_q^0 = \\mathbf{F}_q$ . These two steps are executed alternately for $L$ times and we get updated queries $\\mathbf{F}_q^{2L}$ .", + "bbox": [ + 511, + 818, + 903, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attention-Pixel. For the $k$ -th iteration, we utilize queries $\\mathbf{F}_q^{2k}$ from previous layer to aggregate information from 2D features $\\mathbf{F}_I$ and obtain $\\mathbf{F}_q^{2k+1}$ . Specially, keys and values", + "bbox": [ + 511, + 853, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16925", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "arise from the 2D features $\\mathbf{F}_I$ , and queries arise from $\\mathbf{F}_q^{2k}$ :", + "bbox": [ + 89, + 90, + 477, + 109 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathbf {F} _ {q} ^ {2 k} \\mathbf {W} ^ {\\mathcal {Q}}, \\mathbf {K} = \\mathbf {F} _ {I} \\mathbf {W} ^ {\\mathcal {K}}, \\mathbf {V} = \\mathbf {F} _ {I} \\mathbf {W} ^ {\\mathcal {V}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 116, + 480, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}^{\\mathcal{Q}},\\mathbf{W}^{\\mathcal{K}},\\mathbf{W}^{\\mathcal{V}}\\in \\mathbb{R}^{C\\times d_k}$ are linear projections. Then, the queries $\\mathbf{F}_q^{2k}$ are updated with the multi-head attention mechanism [40]:", + "bbox": [ + 89, + 143, + 482, + 191 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\mathrm {S o f t m a x} (\\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {d _ {k}}}) \\mathbf {V}. \\qquad (9)\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 199, + 482, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "And the attention features are further projected with a shallow MLP as the final output features $\\mathbf{F}_a^{2k + 1}$ .", + "bbox": [ + 89, + 241, + 482, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attention-Point. For the $k$ -th iteration, we utilize the queries $\\mathbf{F}_q^{2k + 1}$ from previous layer to aggregate information from the 3D features $\\mathbf{F}_P$ and obtain $\\mathbf{F}_q^{2k + 2}$ . The whole process is almost the same as described in Attention-Pixel. The only difference is that we utilize the point cloud mask $\\mathbf{M}_p$ as the keys' mask to make queries focus on relevant overlapping regions.", + "bbox": [ + 89, + 271, + 482, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After $L$ iterations, we get updated queries $\\mathbf{F}_q^{2L}$ . These updated queries can adaptively change according to the content of the input image and point cloud, capturing the same patterns in both. Next, We apply Attention-Point and Attention-Pixel respectively, taking the updated queries $\\mathbf{F}_q^{2L}$ as the input to get 3D keypoint detectors $\\mathbf{D}_P$ and 2D keypoint detectors $\\mathbf{D}_I$ . Then with 2D keypoint detectors $\\mathbf{D}_I$ and 2D features $\\mathbf{F}_I$ , we get 2D keypoint heatmap $\\mathbf{H}_I \\in \\mathbb{R}^{N_q \\times HW}$ :", + "bbox": [ + 89, + 378, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {I} = \\operatorname {S o f t m a x} \\left(\\mathbf {D} _ {I} \\operatorname {F l a t t e n} \\left(\\mathbf {F} _ {I}\\right) ^ {\\top}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 523, + 480, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Flatten}(\\cdot)$ denotes combination of the first two dimensions into a single dimension.", + "bbox": [ + 89, + 550, + 482, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D keypoint heatmap $\\mathbf{H}_P\\in \\mathbb{R}^{N_q\\times N}$ is generated in a similar way and our predicted mask $\\mathbf{M}_p$ is applied to suppress irrelevant points.", + "bbox": [ + 89, + 580, + 482, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At last, We can get 3D keypoints $\\mathbf{K}_P\\in \\mathbb{R}^{N_q\\times 3}$ and 2D keypoints $\\mathbf{K}_I\\in \\mathbb{R}^{N_q\\times 2}$ as follows:", + "bbox": [ + 89, + 625, + 482, + 656 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} _ {P} = \\mathbf {H} _ {P} \\mathbf {P}, \\mathbf {K} _ {I} = \\mathbf {H} _ {I} \\mathbf {E} _ {I}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 666, + 482, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Pose Regression Module", + "text_level": 1, + "bbox": [ + 89, + 691, + 313, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After establishing 2D-3D correspondences, we wish to regress camera pose through a learned network. As we have estimated the frustum pose $\\mathbf{T}_f$ in the GPDM, which is a coarse camera pose estimation, we can regard $\\mathbf{T}_f$ as an initial camera pose to be refined and apply corresponding transformation to the 3D keypoints $\\mathbf{K}_P$ . In this case, we estimate the difference between $\\mathbf{R}_f$ , $\\mathbf{t}_f$ and the ground-truth camera pose $\\mathbf{R}_{gt}$ , $\\mathbf{t}_{gt}$ as:", + "bbox": [ + 89, + 713, + 482, + 835 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {R} _ {g t} = \\mathbf {R} _ {g t} \\cdot \\left(\\mathbf {R} _ {f}\\right) ^ {- 1}, \\Delta \\mathbf {t} _ {g t} = \\mathbf {t} _ {g t} - \\Delta \\mathbf {R} _ {g t} \\cdot \\mathbf {t} _ {f}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 843, + 482, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in the top-right part of Figure 2, we first apply a MLP to lift the correspondence features and pair the", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "correspondence via feature concatenation to get the fusion features $\\mathbf{F}_f\\in \\mathbb{R}^{N_q\\times C_f}$ as follows:", + "bbox": [ + 511, + 90, + 905, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {f} = \\operatorname {M L P} \\left(\\left[ \\mathbf {D} _ {P}, \\operatorname {M L P} \\left(\\mathbf {K} _ {P}\\right), \\mathbf {D} _ {I}, \\operatorname {M L P} \\left(\\mathbf {K} _ {I}\\right) \\right]\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 131, + 903, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then inject global information into the point-pixel correspondences in $\\mathbf{F}_f$ by concatenating its averaged feature to strengthen the correspondence information and obtain $\\mathbf{F}_s$ :", + "bbox": [ + 511, + 157, + 905, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A pose-sensitive feature $\\mathbf{f}_{\\text {pose }}$ is learned from all the correspondence information via an average-pooling operation:", + "bbox": [ + 511, + 202, + 905, + 233 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\text {p o s e}} = \\operatorname {A v g p o o l} (\\operatorname {M L P} \\left(\\mathbf {F} _ {s}\\right)). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 247, + 903, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, similar to Equation 1, we apply two parallel MLPs to regress $\\Delta \\mathbf{R}$ and $\\Delta \\mathbf{t}$ respectively:", + "bbox": [ + 511, + 268, + 905, + 299 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {R}, \\Delta \\mathbf {t} = \\rho (\\operatorname {M L P} (\\mathbf {f} _ {\\text {p o s e}})), \\operatorname {M L P} (\\mathbf {f} _ {\\text {p o s e}}). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 309, + 903, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss Function", + "text_level": 1, + "bbox": [ + 511, + 334, + 660, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given ground-truth camera pose $\\mathbf{R}_{gt}$ , $\\mathbf{t}_{gt}$ and camera intrinsic matrix $\\mathbf{K}$ , we introduce a joint loss function consisting of a classification loss $\\mathcal{L}_{cls}$ , a frustum-lose loss $\\mathcal{L}_{fru}$ , a correspondence loss $\\mathcal{L}_{corr}$ , a diversity loss $\\mathcal{L}_{div}$ and a camera-pose loss $\\mathcal{L}_{cam}$ to guide the model learning effectively.", + "bbox": [ + 511, + 356, + 905, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Classification loss. According to Equation 4 and Equation 5, we can get the ground-truth point cloud classification labels $\\mathbf{O}_{gt}$ taking the ground-truth camera pose $\\mathbf{R}_{gt}$ , $\\mathbf{t}_{gt}$ and camera intrinsic matrix $\\mathbf{K}$ as inputs. We adopt weighted binary cross-entropy as follows:", + "bbox": [ + 511, + 431, + 905, + 507 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = \\operatorname {B C E L o s s} \\left(\\mathbf {O} _ {g t}, \\mathbf {O} _ {P}\\right). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 517, + 903, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Frustum-losess. As ground-truth camera pose is also ground truth of the frustum pose,We use $L_{2}$ loss as follows:", + "bbox": [ + 511, + 542, + 905, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f r u} = \\left\\| \\mathbf {R} _ {g t} - \\mathbf {R} _ {f} \\right\\| _ {2} + \\left\\| \\mathbf {t} _ {g t} - \\mathbf {t} _ {f} \\right\\| _ {2}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 588, + 903, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Correspondence loss. To establish accurate 2D-3D correspondences, we adopt $L_{2}$ loss as follows:", + "bbox": [ + 511, + 609, + 905, + 640 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o r r}} = \\| f \\left(\\mathbf {K} _ {P}; \\mathbf {K}, \\mathbf {R} _ {g t}, \\mathbf {t} _ {g t}\\right) - \\mathbf {K} _ {I} \\| _ {2}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 650, + 903, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f$ is defined in Equation 4.", + "bbox": [ + 511, + 675, + 735, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diversity loss. For the goal of expanding the discrepancy among 3D keypoints $\\mathbf{K}_P$ and 2D keypoints $\\mathbf{K}_I$ , we impose the diversity loss $\\mathcal{L}_{div} = \\mathcal{L}_{2d} + \\mathcal{L}_{3d}$ , where", + "bbox": [ + 511, + 691, + 905, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {3 d} = \\frac {1}{N _ {q}} \\sum_ {j = 1} ^ {N _ {q}} \\sum_ {k = 1} ^ {N _ {q}} \\max \\left(0, \\Delta_ {3 d} - \\| \\mathbf {K} _ {P} ^ {j} - \\mathbf {K} _ {P} ^ {k} \\| _ {2}\\right), \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 744, + 903, + 789 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2 d} = \\frac {1}{N _ {q}} \\sum_ {j = 1} ^ {N _ {q}} \\sum_ {k = 1} ^ {N _ {q}} \\max \\left(0, \\Delta_ {2 d} - \\| \\mathbf {K} _ {I} ^ {j} - \\mathbf {K} _ {I} ^ {k} \\| _ {2}\\right). \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 792, + 903, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Camera-posediscovery. Similar to $\\mathcal{L}_{fru}$ ,we define the camerepose loss $\\mathcal{L}_{cam}$ as:", + "bbox": [ + 511, + 844, + 905, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c a m} = \\left\\| \\Delta \\mathbf {R} _ {g t} - \\Delta \\mathbf {R} \\right\\| _ {2} + \\left\\| \\Delta \\mathbf {t} _ {g t} - \\Delta \\mathbf {t} \\right\\| _ {2}, \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 883, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16926", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta_{\\mathbf{R}}$ and $\\Delta_{\\mathrm{t}}$ are defined in Equation 12.", + "bbox": [ + 89, + 90, + 392, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we combine these loss functions by a weighted sum to train our model:", + "bbox": [ + 89, + 106, + 482, + 136 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c a m}} + \\mathcal {L} _ {\\text {f r u}} + \\alpha_ {1} \\mathcal {L} _ {\\text {c l s}} + \\alpha_ {2} \\mathcal {L} _ {\\text {c o r r}} + \\alpha_ {3} \\mathcal {L} _ {\\text {d i v}}, \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 148, + 482, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha_{1},\\alpha_{2},\\alpha_{3}$ are weight terms to balance these losses.", + "bbox": [ + 89, + 178, + 478, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 207, + 223, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first introduce implementation details. Then, we show experimental results and visualizations on two public benchmarks. Finally, we conduct a series of ablation studies to verify the effectiveness of each component.", + "bbox": [ + 89, + 232, + 482, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 301, + 307, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this work, we implement the proposed model in Pytorch [24] and adopt an NVIDIA RTX 3090 GPU for training. We adopt a 4-stage ResNet [18] with FPN [22] as the image backbone network, where the output channel dimension is 128. For the 3D backbone, we use a 4-stage KPFCNN [39] where the output channel dimension is 128. We set the channel dimension $C$ to 128, set $C_f$ to 256 and set $C_{pose}$ to 512. All the attention layers have 128 features channels with 4 attention heads and ReLU activation. The number of correspondence queries $N_q$ is set to 128. And the iteration time $L$ is set to 3. The weight term $\\alpha_1$ , $\\alpha_2$ and $\\alpha_3$ in the loss function are set to 1, 0.1 and 0.1. The margin $\\Delta_{3d}$ and $\\Delta_{2d}$ in the diversity loss are set to 1m and 5 pixels, respectively. All parameters in our proposed model are randomly initialized and trained from scratch with the Adam optimizer. We first train the geometric prior-guided overlapping region detection module with classification loss and frustum-pose loss for 20 epochs. After that, we train the whole network with the total loss $\\mathcal{L}_{total}$ for 20 epochs. The learning rate is set to $10^{-4}$ , and the weight decay is $10^{-6}$ .", + "bbox": [ + 89, + 324, + 483, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 637, + 377, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our proposed method on two commonly used benchmarks KITTI [17] and nuScenes [3].", + "bbox": [ + 89, + 659, + 482, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "KITTI Odometry [17]. KITTI Odometry dataset consists of 22 sequences of images and point cloud data collected simultaneously in urban environments, and 11 sequences provide ground-truth calibration files. We follow the common practice [21] of utilizing the 0-8 sequences for training, and 9-10 for testing. The mis-registration transformation is conducted with a 2D translation on the ground within $\\pm 10\\mathrm{m}$ and a rotation around the up-axis with no limited range. We downsample the image resolution to $160\\times 512$ and the point cloud size to 40960 for training and testing.", + "bbox": [ + 89, + 689, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "nuScenes [3]. nuScenes dataset collects the image-point cloud pairs by official SDK where the point cloud is accumulated from the nearby frames and the image is from the current frame. We follow the official data spilt of nuScenes,", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where 850 scenes are used for training, and 150 scenes for testing. The mis-registration transformation is conducted in a similar way as the one in KITTI dataset. We downsample the image resolution to $160 \\times 320$ and the point cloud size to 40960 for training and testing.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We follow VP2P-match [44] to quantify the performance of image-to-point cloud registration with average Relative Translational Error (RTE), average Relative Rotation Error (RRE) and registration accuracy (Acc). Same as VP2P-match [44], we do not follow CorrI2P [28] to remove the image-point cloud samples with large errors before averaging, which might be an unsuitable way to report real performances. Also, We further report the registration accuracy (Acc) which is the proportion of fine registrations with RTE $< 2\\mathrm{m}$ and RRE $< 5^{\\circ}$ .", + "bbox": [ + 511, + 167, + 906, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Registration Accuracy", + "text_level": 1, + "bbox": [ + 511, + 330, + 720, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with previous state-of-the-art methods [18, 28, 44]:", + "bbox": [ + 511, + 353, + 905, + 383 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Grid Cls. + PnP. DeepI2P [18] proposes to divide the image into $32 \\times 32$ grids and uses a neural network to classify each 3D point to a unique 2D grid. The EPnP [20] with RANSAC [16] is applied to estimate camera pose.", + "- Frus. Cls. + Inv.Proj. DeepI2P [18] proposes to perform frustum classification and inverse camera projection to obtain camera pose. We report their results with 2D/3D inverse camera projection as DeepI2P(2D)/DeepI2P(3D).", + "- Corr12P. Corr12P [28] proposes to learn dense correspondences for image-point cloud pairs, and the EPnP [20] with RANSAC [16] is applied to predict the camera pose.", + "- VP2P-match. VP2P-match [44] is the latest work on image-to-point cloud registration which represents the 3D elements as the combination of voxels and points and trains with a differentiable PnP solver [4]." + ], + "bbox": [ + 511, + 385, + 905, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. We report the quantitative results in Table 1. Our proposed method outperforms all compared methods in all metrics by a large margin on both KITTI and nuScenes datasets. In detail, on the KITTI dataset, our method achieves the highest Acc, with a reduction of $0.55\\mathrm{m}$ in RTE and about $2.05^{\\circ}$ in RRE compared with VP2P-match [44]. On the nuScenes dataset, Acc is $2.61\\%$ higher, RTE is reduced by $0.26\\mathrm{m}$ and RRE is reduced by $0.02^{\\circ}$ . We attribute the superior performance to three delicate designs. Our proposed GPDM leverages geometric prior that the overlapping region is a frustum to enhance the accuracy of point cloud overlapping region detection. And the ICLM can implicitly obtain accurate 2D-3D correspondences, avoiding the unreliable process of forced alignment between 2D and 3D features. Moreover, the PRM generates precise camera pose prediction via a learned network, which enables the end-to-end optimization.", + "bbox": [ + 511, + 612, + 906, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison. The qualitative comparison is shown in Figure 3. For intuitive visualization, we project", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16927", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6d383488315380fc97b3d4b48422aa5907f0cb757fa690ea324874823bcbdb95.jpg", + "table_caption": [ + "Table 1. Registration accuracy on the KITTI and nuScenes datasets. Lower is better for RTE and RRE, higher is better for Acc." + ], + "table_footnote": [], + "table_body": "
MethodsKITTInuScenes
RTE(m)↓RRE(°)↓Acc(%)↑RTE(m)↓RRE(°)↓Acc(%)↑
Grid Cls. + PnP [18]3.64±3.4619.19±28.9611.223.02±2.4012.66±21.012.45
DeepI2P(3D) [18]4.06±3.5424.73±31.693.772.88±2.1220.65±12.242.26
DeepI2P(2D) [18]3.59±3.2111.66±18.1625.952.78±1.994.80±6.2138.10
CorrI2P [28]3.78±65.165.89±20.3472.423.04±60.763.73±9.0349.00
VP2P-match [44]0.75±1.133.29±7.9983.040.89±1.442.15±7.0388.33
Ours0.20±0.211.24±2.3497.490.63±0.442.13±3.7590.94
", + "bbox": [ + 135, + 104, + 859, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c32838ae83dc69983495df10098b5cb1fee3216ad31ae01489bb60d2cbff8f1.jpg", + "image_caption": [ + "Input", + "$7.14\\mathrm{m} / 108.94^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 137, + 252, + 305, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddde2f3207ff08ee012a297f7dc847203248d32495f1284ab8668dde24c3c96f.jpg", + "image_caption": [ + "VP2P-match", + "2.47m / 12.93°" + ], + "image_footnote": [], + "bbox": [ + 321, + 252, + 493, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b5d2446155a657659ea2c6add92a0f94bd2dcbe3a8e3a66c2a2c5d221f6e163.jpg", + "image_caption": [ + "Ours", + "$0.22\\mathrm{m} / 0.14^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 506, + 253, + 674, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1edd3d3cec4e581cdaa21bc07c16615c9a3560cf2c54fb7b6a043d9bef22803b.jpg", + "image_caption": [ + "GT", + "RTE(m) / RRE(°)" + ], + "image_footnote": [], + "bbox": [ + 687, + 253, + 857, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/129165332e909eab0ddf6404f6907bb35f2872bc3a40d9b8d9990b04a4035177.jpg", + "image_caption": [ + "9.91m/10.53°" + ], + "image_footnote": [], + "bbox": [ + 137, + 313, + 305, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e9dc57b46863cd9bd72311c176d7b0a6ab3497c6608d05fe929a0fd256882836.jpg", + "image_caption": [ + "$0.88\\mathrm{m} / 2.07^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 321, + 313, + 493, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9644e15ce6e79335738551e787fbd60d76b61f3e1ff7fa6ab2955b31ef43a5d.jpg", + "image_caption": [ + "$0.07\\mathrm{m} / 1.40^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 504, + 313, + 674, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/35b0405b01f0cd52267bcfe922d797b5e6ee37fc1e6d8c0dfc5f784fa7e1d481.jpg", + "image_caption": [ + "RTE(m) / RRE(°)" + ], + "image_footnote": [], + "bbox": [ + 687, + 313, + 859, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc9985eb0d079f86a4b3157cf5d66f9b3bb2282462fe49d65e70f5e206026e50.jpg", + "image_caption": [ + "$12.93\\mathrm{m} / 70.03^{\\circ}$", + "Figure 3. Qualitative comparison of Image-to-Point Cloud registration results under KITTI dataset." + ], + "image_footnote": [], + "bbox": [ + 137, + 373, + 305, + 416 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2aed46aadf950e5f649332d34195bac1d7740559aa08e60390e2e14ef927a43d.jpg", + "image_caption": [ + "$5.38\\mathrm{m} / 18.78^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 321, + 373, + 493, + 416 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad9ecf6858b08f076dfedac8076c54084ddb99d716df41941993a8c921d9aa57.jpg", + "image_caption": [ + "$0.07\\mathrm{m} / 1.31^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 504, + 373, + 674, + 416 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc28450e61e06b18fa7901c62edf7fcf4f9a528d4b63b92c9e6d78f936b6aba8.jpg", + "image_caption": [ + "RTE(m) / RRE(°)" + ], + "image_footnote": [], + "bbox": [ + 687, + 373, + 859, + 416 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f62219bb44d2f9df158644ab57be868e887c2168cab0a7c31a36cfcac0d991f1.jpg", + "image_caption": [ + "$0.16\\mathrm{m} / 0.82^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 114, + 472, + 277, + 550 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/401cef3422d067ebfe459df501f8529ee02079443619d6ccf95c5bf7678c1b52.jpg", + "image_caption": [ + "$0.04\\mathrm{m} / 0.41^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 292, + 472, + 454, + 550 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/263ce10b412bbc95b3e08be192c60f5e5bc6d0b0702bf58c29fbdc792b9c3079.jpg", + "image_caption": [ + "$0.08\\mathrm{m} / 0.70^{\\circ}$", + "Figure 4. Visual illustration of 2D-3D correspondences and registration accuracy. Green lines represent correct correspondences." + ], + "image_footnote": [], + "bbox": [ + 114, + 571, + 277, + 650 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25133ddde07dace7da2da6ebaa0cddba96b87f730fcc016c001511c8666d4ba4.jpg", + "image_caption": [ + "$0.14\\mathrm{m} / 0.59^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 290, + 571, + 452, + 650 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5eac0fef7690a988acbe39f7dcec97a3127b53df3359aa4803d160c1d20bae0f.jpg", + "table_caption": [ + "Table 2. The effect of each design in our framework." + ], + "table_footnote": [], + "table_body": "
RTE(m)↓RRE(°)↓Acc(%)↑
w/o GPDM0.45±1.052.48±9.6891.89
w/o ICLM1.05±0.733.70±6.2072.64
w/o PRM1.05±2.934.96±18.4782.00
Full0.20±0.241.24±2.4097.49
", + "bbox": [ + 102, + 707, + 467, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the point cloud into image space through predicted transformations of different approaches and the known camera intrinsic matrix, and the color indicates the depth. Compared with VP2P-match [44], our method achieves better registration accuracy in different road situations. For example, on the difficult situation (e.g. the 1st and 3rd row), VP2P-match [44] fails to solve correct registration, while", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de9620719ba41cc6e915ff34cd8ee7f76a7a7d27cd3b3fa528ce2debad74f08e.jpg", + "image_caption": [ + "w/o Geometric Prior", + "0.86/0.76" + ], + "image_footnote": [], + "bbox": [ + 542, + 473, + 645, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb2c2f4be8e1076ce856199438f03d180cbfe2f3a17a5c9dbfcef954c2ad3cea.jpg", + "image_caption": [ + "w Geometric Prior", + "0.96/1.00" + ], + "image_footnote": [], + "bbox": [ + 656, + 473, + 756, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a7f41ad61a95c82f9805c5b46e130f1c001963b1366a2e0f9db7d0a779ae45c.jpg", + "image_caption": [ + "GT", + "Precise/Recall" + ], + "image_footnote": [], + "bbox": [ + 774, + 473, + 875, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f0120884bd5e609efe14e641d640c6ed32f8ae1590ad8ca742ad80c67eeced91.jpg", + "image_caption": [ + "0.79/0.76", + "Figure 5. Visual illustration of overlapping region detection. The green points are correctly detected in the overlapping region, the blue points are missed in the overlapping region, and the red points are incorrectly detected outside the overlapping region." + ], + "image_footnote": [], + "bbox": [ + 560, + 563, + 607, + 642 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/70041bb9f9ff4cd6173845d2d21c2777acb411a7cfc2c95a40446832f3ba9d02.jpg", + "image_caption": [ + "0.99 / 0.95" + ], + "image_footnote": [], + "bbox": [ + 668, + 563, + 715, + 642 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d42dca266a6e64b33667a5d40138c08a94ee13ea1fc45dd16b836232b495c72.jpg", + "image_caption": [ + "Precise/Recall" + ], + "image_footnote": [], + "bbox": [ + 790, + 564, + 839, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our method achieves much higher registration accuracy. We further provide the 2D-3D correspondence visualization in Figure 4. Almost all of the pixels of our method can reach slight errors within 15 pixels, which demonstrates that our ICLM can obtain accurate 2D-3D correspondences.", + "bbox": [ + 511, + 715, + 905, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study.", + "text_level": 1, + "bbox": [ + 511, + 801, + 669, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to justify the effectiveness of each design and the effect of some important parameters in our method on the KITTI dataset.", + "bbox": [ + 511, + 824, + 905, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To justify the effectiveness of each design in our proposed method, we develop three different variations for", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16928", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3a19ac9e75c0c1893024ef3c34e785e312418b50daa428acf2c34ecc2b5dcd48.jpg", + "table_caption": [ + "Table 3. Comparison of frustum classification results on the KITTI and nuScenes datasets." + ], + "table_footnote": [], + "table_body": "
SettingsKITTInuScenes
PrecisionRecallAccuracyF1-scorePrecisionRecallAccuracyF1-score
w/o GPDM93.4595.9097.9394.6682.8685.5993.1484.20
w GPDM95.1594.9198.0495.0392.0491.5396.3691.79
", + "bbox": [ + 147, + 103, + 846, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "comparison in Table 2: (1) $w / o$ GPDM removes the frustum pose prediction branch and regards points with infrustum probability $\\mathbf{O}_P$ greater than 0.5 as the overlapping region. (2) $w / o$ ICLM removes both the ICLM and PRM since we can't make the PRM work without the ICLM. In this case, we take the predicted frustum pose as our estimated camera pose. (3) $w / o$ PRM replaces the PRM with the PnP-RANSAC algorithm [16, 20].", + "bbox": [ + 89, + 176, + 480, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of the GPDM. As shown in Table 2, with GPDM, the performance on the KITTI dataset is improved on all metrics. In specific, Acc is $6\\%$ higher, RTE is reduced by about $56\\%$ and RRE is reduced by about $50\\%$ . The main reason is that our GPDM can accurately detect the overlapping region and generate a coarse camera pose which can be refined later. And to prove that our GPDM can improve the accuracy of point cloud overlapping region detection, we report the quantitative results in Table 3. With GPDM, we get better frustum classification results almost on all metrics, especially on the nuScenes dataset. The visual comparison is shown in Figure 5, which demonstrates our propose GPDM can get better prediction on the frustum boundary leveraging our proposed geometric prior.", + "bbox": [ + 89, + 299, + 482, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of the ICLM. As shown in Table 2, with proposed ICLM, the performance on the KITTI dataset is improved notably. In specific, Acc is $34\\%$ higher, RTE is reduced by about $81\\%$ and RRE is reduced by about $66\\%$ . Regarding the gain on Acc, we can figure out that the ICLM plays the most important role in our model, which demonstrates that our designed implicit correspondence learning is very effective for Image-to-Point Cloud registration.", + "bbox": [ + 89, + 513, + 482, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of the PRM. As shown in Table 2, the PRM outperforms the PnP-RANSAC algorithm [16, 20] on all metrics. In specific, Acc is $19\\%$ higher, RTE is reduced by about $81\\%$ and RRE is reduced by about $75\\%$ . We owe the advantage to the enabled end-to-end optimization with PRM.", + "bbox": [ + 89, + 637, + 482, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of the number of correspondence queries. In Table 4, we show the impact of different numbers of correspondence queries in the ICLM on the KITTI dataset. The results show that our method can achieve better performance than VP2P-match [44] by using only 16 correspondence queries. We find that the overall performance of the model improves slightly with the increase of $N_{q}$ . We choose $N_{q} = 128$ in our method for the highest Acc.", + "bbox": [ + 89, + 715, + 482, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of the iteration time. To investigate the influences of the iteration time $(L)$ in the ICLM, we pick $L$ from the set $\\{0, 1, 2, 3, 4\\}$ . As shown in Table 5, we notice that the performance drops a lot if $L$ is set to 0, showing the", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9b7d3d3593109fe72e7b7e4ca5bf7d9973578089e92c10ecd2554719288d43a7.jpg", + "table_caption": [ + "Table 4. Ablation studies on the number of queries." + ], + "table_footnote": [], + "table_body": "
NqRTE(m)↓RRE(°)↓Acc(%)↑
160.26±0.291.71±2.4095.52
320.23±0.251.43±2.2495.70
640.21±0.191.28±3.2096.63
1280.20±0.241.24±2.4097.49
2560.21±0.211.21±2.4097.42
", + "bbox": [ + 555, + 188, + 859, + 284 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/11174fa0d12a3d9e91388a8f09896b1dea9ce8744e2b4a5ae8ba2abfc320bc08.jpg", + "table_caption": [ + "Table 5. Ablation studies on the iteration time." + ], + "table_footnote": [], + "table_body": "
LRTE(m)↓RRE(°)↓Acc(%)↑
00.31±0.362.34±2.7690.01
10.21±0.161.23±1.9597.44
20.20±0.161.28±2.0297.33
30.20±0.241.24±2.4097.49
40.21±0.301.26±2.4797.46
", + "bbox": [ + 563, + 300, + 854, + 396 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ca254273b068ff32bdbfaec1f637278ca72494b1aa0877e98252fd30bcfc4080.jpg", + "table_caption": [ + "Table 6. Ablation studies on the iteration order." + ], + "table_footnote": [], + "table_body": "
RTE(m)↓RRE(°)↓Acc(%)↑
Pixel-then-Point0.20±0.241.24±2.4097.49
Point-then-Pixel0.23±0.601.29±2.2697.05
", + "bbox": [ + 516, + 414, + 903, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "necessity of the iteration operation. We choose the model with $L = 1$ for the optimal balance between computational efficiency and performance.", + "bbox": [ + 511, + 472, + 903, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of different iteration orders. To investigate the influences of the iteration order in the ICLM, we conduct two experiments with first Attention-Pixel then Attention-Point or first Attention-Point then Attention-Pixel. As shown is Table 6, we get similar performance with different iteration orders. And we choose to iteratively apply Attention-Pixel then Attention-Point in our method.", + "bbox": [ + 511, + 518, + 905, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 638, + 633, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel implicit correspondence learning method with three specific designs for image-top-point cloud registration. Specially, the proposed geometric prior can guide the model to achieve high accuracy on the overlapping region detection. The proposed implicit correspondence learning module can implicitly acquire accurate 2D-3D correspondences. Furthermore, we propose a pose regression module to handle incorrect 2D-3D correspondences and generates precise camera pose. Extensive experimental results on two challenging benchmarks demonstrate the effectiveness of the proposed method.", + "bbox": [ + 511, + 664, + 905, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 845, + 700, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Defense Basic Scientific Research program (No.JCKY2022911B002).", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16929", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xuyang Bai, Zixin Luo, Lei Zhou, Hongbo Fu, Long Quan, and Chiew-Lan Tai. D3feat: Joint learning of dense detection and description of 3d local features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6359-6367, 2020. 3", + "[2] Mark Billinghurst, Adrian Clark, Gun Lee, et al. A survey of augmented reality. Foundations and Trends® in Human-Computer Interaction, 8(2-3):73-272, 2015. 1", + "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6", + "[4] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2781-2790, 2022. 3, 6", + "[5] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 539-546. IEEE, 2005. 2", + "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 3", + "[7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8958-8966, 2019. 3", + "[8] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In Proceedings of the European conference on computer vision (ECCV), pages 602-618, 2018.", + "[9] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 195-205, 2018. 3", + "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 2", + "[11] Zhen Dong, Fuxun Liang, Bisheng Yang, Yusheng Xu, Yufu Zang, Jianping Li, Yuan Wang, Wenxia Dai, Hongchao Fan, Juha Hyyppä, et al. Registration of large-scale terrestrial laser scanner point clouds: A review and benchmark. IS-PRS Journal of Photogrammetry and Remote Sensing, 163: 327–342, 2020. 1", + "[12] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "object recognition. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 998-1005. IEEE, 2010. 3", + "[13] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99–110, 2006. 1", + "[14] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 8092-8101, 2019. 2", + "[15] Mengdan Feng, Sixing Hu, Marcelo H Ang, and Gim Hee Lee. 2d3d-matchnet: Learning to match keypoints across 2d image and 3d point cloud. In 2019 International Conference on Robotics and Automation (ICRA), pages 4790-4796. IEEE, 2019. 1, 2, 3", + "[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 3, 6, 8", + "[17] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 6", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 6, 7", + "[19] Yurim Jeon and Seung-Woo Seo. Efghnet: A versatile image-to-point cloud registration network for extreme outdoor environment. IEEE Robotics and Automation Letters, 7 (3):7511-7517, 2022. 1", + "[20] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate $O(n)$ solution to the pnp problem. International journal of computer vision, 81:155-166, 2009. 2, 3, 6, 8", + "[21] Jiaxin Li and Gim Hee Lee. Deepi2p: Image-to-point cloud registration via deep classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15960-15969, 2021. 1, 3, 6", + "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3, 6", + "[23] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 2, 3", + "[24] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16930", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3", + "[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3", + "[27] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143-11152, 2022. 3", + "[28] Siyu Ren, Yiming Zeng, Junhui Hou, and Xiaodong Chen. Corri2p: Deep image-to-point cloud registration via dense correspondence. IEEE Transactions on Circuits and Systems for Video Technology, 33(3):1198-1208, 2022. 1, 2, 3, 6, 7", + "[29] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in neural information processing systems, 32, 2019. 2", + "[30] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. 2", + "[31] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE international conference on robotics and automation, pages 3212-3217. IEEE, 2009. 3", + "[32] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2", + "[33] Paul-Edouard Sarlin, Daniel DeTone, Tsun-Yi Yang, Armen Avetisyan, Julian Straub, Tomasz Malisiewicz, Samuel Rota Bulo, Richard Newcombe, Peter Kontschieder, and Vasileios Balntas. Orienternet: Visual localization in 2d public maps with neural matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21632–21642, 2023. 1", + "[34] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 2", + "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. 3rd International Conference on Learning Representations, 2015. 3", + "[36] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 3", + "[37] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6398-6407, 2020. 2" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European conference on computer vision, pages 685–702. Springer, 2020. 3", + "[39] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 3, 6", + "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 5", + "[41] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loptr: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 3", + "[42] Hao Yu, Fu Li, Mahdi Saleh, Benjamin Busam, and Slobodan Ilic. Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration. Advances in Neural Information Processing Systems, 34:23872-23884, 2021. 3", + "[43] Yu Zhong. Intrinsic shape signatures: A shape descriptor for 3d object recognition. In 2009 IEEE 12th international conference on computer vision workshops, ICCV Workshops, pages 689-696. IEEE, 2009. 3", + "[44] Junsheng Zhou, Baorui Ma, Wenyuan Zhang, Yi Fang, Yu-Shen Liu, and Zhizhong Han. Differentiable registration of images and lidar point clouds with voxelpoint-to-pixel matching. In Advances in Neural Information Processing Systems, pages 51166-51177. Curran Associates, Inc., 2023. 1, 2, 3, 6, 7, 8", + "[45] Qunjie Zhou, Torsten Sattler, and Laura Leal-Taixe. Patch2pix: Epipolar-guided pixel-level correspondences. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4669-4678, 2021. 3", + "[46] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745-5753, 2019. 3" + ], + "bbox": [ + 516, + 90, + 905, + 710 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16931", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/3b6935ad-40bc-4903-80f6-f62148a4f3d3_model.json b/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/3b6935ad-40bc-4903-80f6-f62148a4f3d3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..68511abedd8976e5065fb3bf985c4556dc4cdec6 --- /dev/null +++ b/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/3b6935ad-40bc-4903-80f6-f62148a4f3d3_model.json @@ -0,0 +1,2926 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.131, + 0.87, + 0.154 + ], + "angle": 0, + "content": "Implicit Correspondence Learning for Image-to-Point Cloud Registration" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.18, + 0.893, + 0.217 + ], + "angle": 0, + "content": "Xinjun Li\\(^{1}\\), Wenfei Yang\\(^{1}\\), Jiacheng Deng\\(^{1}\\), Zhixin Cheng\\(^{1}\\), Xu Zhou\\(^{3}\\), Tianzhu Zhang\\(^{1,2,*}\\) \n\\(^{1}\\)University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.177, + 0.217, + 0.893, + 0.253 + ], + "angle": 0, + "content": "\\(^{2}\\)National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory \n\\(^{3}\\)Sangfor Technologies" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.255, + 0.962, + 0.27 + ], + "angle": 0, + "content": "{1xj3017,dengjc,chengzhixin}@mail.ustc.edu.cn, {yangwf,tzzhang}@ustc.edu.cn, zhouxu@sangfor.com.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.485, + 0.701 + ], + "angle": 0, + "content": "Image-to-point cloud registration aims to estimate the camera pose of a given image within a 3D scene point cloud. In this area, matching-based methods have achieved leading performance by first detecting the overlapping region, then matching point and pixel features learned by neural networks and finally using the PnP-RANSAC algorithm to estimate camera pose. However, achieving accurate image-to-point cloud registration remains challenging because the overlapping region detection is unreliable merely relying on point-wise classification, direct alignment of cross-modal data is difficult and indirect optimization objective leads to unstable registration results. To address these challenges, we propose a novel implicit correspondence learning method, including a Geometric Prior-guided overlapping region Detection Module (GPDM), an Implicit Correspondence Learning Module (ICLM), and a Pose Regression Module (PRM). The proposed method enjoys several merits. First, the proposed GPDM can precisely detect the overlapping region. Second, the ICLM can generate robust cross-modality correspondences. Third, the PRM can enable end-to-end optimization. Extensive experimental results on KITTI and nuScenes datasets demonstrate that the proposed model sets a new state-of-the-art performance in registration accuracy." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.73, + 0.222, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Image-to-point cloud registration [15] is a fundamental task in the field of computer vision, aimed at estimating the camera pose of a given image within a 3D scene point cloud. It plays an important role in many computer vision tasks, including simultaneous localization and mapping [13], 3D reconstruction [11], augmented/virtual reality [2], and visual localization [33]. Although significant progress has been made in recent years, image-to-point cloud registra" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.305, + 0.868, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.43, + 0.794, + 0.44 + ], + "angle": 0, + "content": "a) Existing matching-based methods" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.44, + 0.867, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.568, + 0.827, + 0.58 + ], + "angle": 0, + "content": "b) Our Implicit Correspondence Learning method" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.721 + ], + "angle": 0, + "content": "Figure 1. (a) The pipeline of mainstream matching-based methods. The green lines denote correct correspondences, and the red lines denote incorrect correspondences. 2D-3D correspondences are first established by calculating feature similarity. Then the PnP-RANSAC algorithm is applied to obtain camera pose. (b) The pipeline of our proposed method. A set of shared queries are used to aggregate information from image and point cloud to generate 2D and 3D keypoint detectors. With paired 2D and 3D keypoint detectors, we obtain a set of 2D-3D correspondences and estimate camera pose through a pose regression network." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.779 + ], + "angle": 0, + "content": "tion remains challenging due to the inherent disparities that images capture appearances with 2D grid data while point clouds represent structures in an unordered format." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, numerous methods have been proposed for image-to-point cloud registration [15, 19, 21, 28, 44]. Generally, these methods can be categorized into two main classes, including matching-free methods [19, 21] and matching-based methods [15, 28, 44]. Matching-free methods utilize the camera imaging principle as a prior to predict camera pose. For example, DeepI2P [21] casts the registration problem into a classi-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16922" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.364 + ], + "angle": 0, + "content": "fication problem. A binary classification network is designed to determine whether each scene point is within or beyond the camera frustum. Then the classification results are fed into an inverse camera projection solver to obtain camera pose. However, these methods merely determine a coarse image-point correspondence, e.g. whether a 3D point is inside the image frustum, which is insufficient for precise camera pose estimation. Different from matching-free methods, matching-based methods focus on establishing fine-grained 2D-3D correspondences and use the PnP-RANSAC algorithm [16, 20] to estimate camera pose [15, 28, 44], which have achieved superior performance. These methods generally involve three important steps: (1) A detector is designed to identify the overlapping region between the image and point cloud. (2) 2D-3D correspondences are established within the identified overlapping region. (3) The PnP-RANSAC algorithm [16, 20] is applied as a post-processing module to obtain camera pose." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Despite the success of these matching-based methods, there still exist three notable problems that hinder the registration accuracy of these methods. (1) Unreliable point-based detection of the overlapping region. Most existing methods adopt a point-wise classification way to determine whether each point lies in the overlapping region or not, ignoring the global projection relationship between the image and point cloud, e.g. an image should correspond to a frustum in the point cloud. Consequently, existing methods usually suffer from inaccurate boundary detection of the overlapping region, resulting in some missing 2D-3D correspondences or incorrect correspondences outside the overlapping region. (2) Inaccurate explicit matching between the image and point cloud. As shown in Figure 1 (a), matching-based methods essentially establish 2D-3D correspondences by directly calculating feature similarity between the image and point cloud. These methods employ metric learning techniques like contrastive loss [5], triplet loss [34] or circle loss [37] to force the alignment of corresponding 2D and 3D features. However, due to the inherent data discrepancy between the image and point cloud, direct alignment of cross-modal data can result in a significant number of mismatches, as shown in Figure 1 (a). (3) Indirect correspondence-based optimization objective. After establishing 2D-3D correspondences, the majority of existing approaches employ the PnP-RANSAC algorithm [16, 20] as a post-processing module to estimate camera pose as illustrated in Figure 1 (a). During the training stage, these approaches primarily consider 2D-3D correspondences as the optimization target, without optimizing the pose directly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Motivated by the above discussions, we propose a novel implicit correspondence learning method for imaged-to-point cloud registration. The proposed model contains three modules, a Geometric Prior-guided overlapping re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.513 + ], + "angle": 0, + "content": "GPI Detection Module (GPDM), an Implicit Correspondence Learning Module (ICLM), and a Pose Regression Module (PRM). The GPDM leverages geometric prior that the overlapping region is a frustum to enhance the accuracy of the overlapping region detection. Specifically, a binary classification network is designed to determine whether each point lies within or beyond the camera frustum. Then, we predict center coordinate and orientation of the camera frustum using obtained classification results. Finally, we regard points in predicted camera frustum as our overlapping region. Since camera frustum defines a continuous and regular geometric region, the noise's interference on the overlapping region detection can be mitigated. The ICLM is designed to implicitly identify a set of 2D-3D correspondences within the overlapping region. As shown in Figure 1 (b), we initialize a set of learnable correspondence queries. For each query, we convert it to 2D/3D keypoint detector via the attention mechanism [40] to aggregate features from the image and point cloud, respectively. By calculating similarity between 2D/3D keypoint detector and image/point cloud features, a pair of 2D/3D keypoint heatmap is obtained. In this way, we can implicitly obtain a set of 2D-3D paired keypoints, avoiding the process of forced alignment between 2D and 3D features. Finally, the PRM takes obtained 2D/3D keypoint detectors and 2D/3D keypoints as inputs, and regresses camera pose via a Multi-Layer Perceptron (MLP), which enables the end-to-end optimization. The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.515, + 0.905, + 0.574 + ], + "angle": 0, + "content": "- We propose a novel implicit correspondence learning method for image-to-point cloud registration, which enjoys the merits of precise overlapping region detection, robust cross-modality matching, end-to-end optimization." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.576, + 0.905, + 0.651 + ], + "angle": 0, + "content": "- Without bells and whistles, extensive experimental results on KITTI and nuScenes datasets demonstrate that the proposed model sets a new state-of-the-art performance in registration accuracy, i.e. a reduction of \\(0.55\\mathrm{m}\\) in RTE and \\(2.05^{\\circ}\\) in RRE compared to previous SOTA method." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.515, + 0.905, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.664, + 0.665, + 0.68 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.905, + 0.735 + ], + "angle": 0, + "content": "In this section, we briefly overview methods that are related to image-to-image registration, point cloud-to-point cloud registration and image-to-point cloud registration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Image-to-image registration. Image-to-image registration is the process of aligning two or more images, which aims to find the optimal transformation that best aligns the structures of interest in the input images. Typical methods usually adopt a detect-then-match pipeline to extract correspondences. Keypoints are first detected and described using handcrafted [23, 30] or learning-based techniques [10, 14, 29, 32], correspondences are then established based on their feature similarity. However, robustly detecting repeatable keypoints across different images is still challenging, especially for texture-poor regions. Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16923" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "free methods [36, 41, 45] propose to bypass the keypoint detection step by first obtaining patch-level matches and then regress pixel-wise matches in a coarse-to-fine manner, which achieves accurate image-to-image registration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.18, + 0.483, + 0.378 + ], + "angle": 0, + "content": "Point cloud-to-point cloud registration. Point cloud-to-point cloud registration is the process of finding a rigid transformation that best aligns two point clouds. Similar to image-to-image registration, early works leverage hand-crafted descriptors [12, 31] or learning-based descriptors [1, 7-9] matching to estimate the correspondences. CoFiNet [42] follows the design of LoFTR [36] and introduces detection-free registration with a coarse-to-fine strategy. Recent methods have replaced RANSAC [16] with deep robust estimators to enhance both speed and accuracy. For example, GeoTransformer [27] proposes a geometric Transformer and introduces a local-to-global registration for RANSAC-free registration." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.402, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Image-to-point cloud registration. Compared to intramodality registration, image-to-point cloud registration poses more significant difficulties due to the inherent data disparities. The typical pipeline is to first extract 2D-3D correspondences and then estimate camera pose using the PnP-RANSAC algorithm [16, 20]. In this case, the accuracy of extracted correspondences plays a crucial role in ensuring successful registration. 2D3D-MatchNet [15] is the pioneering work, which uses SIFT [23] and ISS [43] to extract the keypoints of the image and point cloud, respectively. Then the patches around the keypoints are fed to CNNs [35] and PointNet [25] to extract features for establishing 2D-3D correspondences. However, the independent hand-crafted keypoint detectors for different modalities leads to a poor registration accuracy. CorrI2P [28] predicts pixels and points in the overlapping region and learns dense 2D-3D correspondences by matching point and pixel features for fine registration. However, it struggles to learn a discriminative latent space for 2D and 3D data, leading to a high rate of incorrect correspondences. Recently, VP2PMatch [44] learns a structured cross-modality latent space by representing 3D elements using both voxels [6, 38] and points [26]. It imposes supervisions directly on the predicted pose distribution with a differentiable probabilistic PnP solver [4]. Though introducing voxel features, it struggles with 2D-3D data disparities. Besides, it fails to achieve high registration accuracy for insufficient supervisions. In addition to these matching-based methods, there are some matching-free methods. For instance, DeepI2P [21] converts the registration problem into a classification and inverse camera projection optimization problem. However, the frustum classification only indicates coarse 2D-3D correspondence, thus limiting the registration accuracy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.605, + 0.107 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.114, + 0.905, + 0.191 + ], + "angle": 0, + "content": "Given a 3D scene point cloud \\(\\mathbf{P} \\in \\mathbb{R}^{N \\times 3}\\) and an image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\) from the same scene, the cross-modality image-to-point cloud registration is to estimate the camera pose \\(\\mathbf{T}_{gt}\\) in the coordinate system of \\(\\mathbf{P}\\), including a rotation matrix \\(\\mathbf{R}_{gt} \\in \\mathbf{SO}(3)\\) and a translation vector \\(\\mathbf{t}_{gt} \\in \\mathbb{R}^3\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.192, + 0.907, + 0.357 + ], + "angle": 0, + "content": "An overview of our pipeline is illustrated in Figure 2. Our proposed model mainly consists of a Geometric Prior-guided overlapping region Detection Module (GPDM), an Implicit Correspondence Learning Module (ICLM) and a Pose Regression Module (PRM). Specifically, we first extract features for the image and point cloud respectively. Next, the GPDM detects the overlapping region of the input point cloud. Then the ICLM implicitly identifies a set of correspondences within detected overlapping region via our designed correspondence queries. At last, we estimate the camera pose via the PRM." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.365, + 0.696, + 0.38 + ], + "angle": 0, + "content": "3.1. Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.387, + 0.906, + 0.463 + ], + "angle": 0, + "content": "For the image, we use ResNet [18] with FPN [22] to extract 2D features \\(\\mathbf{F}_I\\in \\mathbb{R}^{H\\times W\\times C}\\). For simplicity, we denote the pixel coordinate matrices for \\(\\mathbf{F}_I\\) as \\(\\mathbf{E}_I\\in \\mathbb{R}^{HW\\times 2}\\). For the input point cloud, we adopt KPFCNN [39] to extract 3D features \\(\\mathbf{F}_P\\in \\mathbb{R}^{N\\times C}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.47, + 0.905, + 0.501 + ], + "angle": 0, + "content": "3.2. Geometric Prior-guided Overlapping Region Detection Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.627 + ], + "angle": 0, + "content": "To precisely detect the overlapping region of the point cloud, most methods adopt a point-wise classification strategy to determine whether each point lies in the overlapping region or not, ignoring the global projection relationship between the image and point cloud. Inspired by the fact that an image should correspond to a frustum in the point cloud, we design a geometric prior-guided classification strategy thus we can get more accurate classification results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.628, + 0.906, + 0.703 + ], + "angle": 0, + "content": "As shown in the bottom-left part of Figure 2, we first obtain 2D global feature by averaging extracted 2D features \\(\\mathbf{F}_I\\) and repeat it to concatenate with the 3D features \\(\\mathbf{F}_P\\). Then a MLP is used to learn the probability \\(\\mathbf{O}_P \\in \\mathbb{R}^{N \\times 1}\\) of a point to lie within the camera frustum." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.703, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Since the in-frustum probability \\(\\mathbf{O}_P\\) is not precise enough for in-frustum classification, we propose to estimate the frustum pose to get more accurate classification results. Utilizing \\(\\mathbf{O}_P\\) and the coordinate information of the point cloud \\(\\mathbf{P}\\), we apply another MLP to regress the frustum pose \\(\\mathbf{T}_f \\in \\mathbf{SE}(\\mathbf{3})\\), composed of a 3D rotation \\(\\mathbf{R}_f \\in \\mathbf{SO}(\\mathbf{3})\\) and a 3D translation vector \\(\\mathbf{t}_f \\in \\mathbb{R}^3\\). In detail, we apply two parallel MLPs to regress \\(\\mathbf{R}_f\\) and \\(\\mathbf{t}_f\\) respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.832, + 0.905, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {f}, \\mathbf {t} _ {f} = \\rho (\\mathrm {M L P} ([ \\mathbf {O} _ {P}, \\mathbf {P} ])), \\mathrm {M L P} ([ \\mathbf {O} _ {P}, \\mathbf {P} ]), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\([\\cdot, \\cdot]\\) denotes concatenation along feature dimension. We choose a 6D representation of rotation [46] as the regression of the first MLP, for its continuous learning space" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16924" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.089, + 0.861, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.289, + 0.861, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.484, + 0.907, + 0.596 + ], + "angle": 0, + "content": "Figure 2. Overview of our method. Given a pair of point cloud \\(\\mathbf{P}\\) and image \\(\\mathbf{I}\\) as inputs, we first extract features for the image and point cloud respectively. Then, in the geometric prior-guided overlapping region detection module, we leverage geometric prior that the overlapping region is a frustum to classify each point in the point cloud into within or beyond the camera frustum and get the point cloud mask \\(\\mathbf{M}_P\\). Next, in the implicit correspondence learning module, we initialize a set of correspondence queries \\(\\mathbf{F}_q\\) and convert them to 2D keypoint detectors \\(\\mathbf{D}_I\\) and 3D keypoint detectors \\(\\mathbf{D}_P\\) by aggregating extracted 2D and 3D features via the attention mechanism. Weighting the pixel/point coordinates according to the feature similarity between the 2D/3D keypoint detectors and 2D/3D features, we obtain 2D keypoints \\(\\mathbf{K}_I\\) and 3D keypoints \\(\\mathbf{K}_P\\), respectively. Finally, in the pose regression module, we utilize the information of 2D-3D paired keypoints and 2D-3D paired keypoint detectors to regress the camera pose \\(\\Delta \\mathbf{R}\\) and \\(\\Delta t\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.673 + ], + "angle": 0, + "content": "in \\(\\mathbf{SO}(3)\\), and \\(\\rho(\\cdot)\\) denotes the transformation from the 6D representation to the \\(3 \\times 3\\) rotation matrix \\(\\mathbf{R}_f\\). Then according to the given camera intrinsic matrix \\(\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}\\), we define the point-to-pixel projection function as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.678, + 0.483, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\left[ x _ {i}, y _ {i}, z _ {i} \\right] ^ {\\top} = \\mathbf {K} \\cdot \\left(\\mathbf {R} _ {f} \\cdot \\mathbf {p} _ {i} + \\mathbf {t} _ {f}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.699, + 0.482, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} _ {i} = \\left[ u _ {x i}, u _ {y i} \\right] ^ {\\top} = \\left[ x _ {i} / z _ {i}, y _ {i} / z _ {i} \\right] ^ {\\top}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.72, + 0.483, + 0.737 + ], + "angle": 0, + "content": "\\[\nf \\left(\\mathbf {p} _ {i}; \\mathbf {K}, \\mathbf {R} _ {f}, \\mathbf {t} _ {f}\\right) = \\mathbf {u} _ {i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.483, + 0.779 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_i\\in \\mathbf{P}\\) is a 3D point and \\(\\mathbf{u}_i\\) represents a 2D pixel. We get the point cloud mask \\(\\mathbf{M}_p\\in \\mathbb{R}^{N\\times 1}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.785, + 0.483, + 0.826 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {p} ^ {i} = \\left\\{ \\begin{array}{l} 1; 0 \\leq u _ {x i} < W, 0 \\leq u _ {y i} < H, z _ {i} \\geq 0 \\\\ 0; \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.458, + 0.851 + ], + "angle": 0, + "content": "3.3. Implicit Correspondence Learning Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "After obtaining point cloud mask \\(\\mathbf{M}_p\\), we implicitly identify a set of robust correspondences between the image and point cloud. As shown in the bottom-right part of Figure 2," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.748 + ], + "angle": 0, + "content": "we first initialize \\(N_{q}\\) correspondence queries \\(\\mathbf{F}_q\\in \\mathbb{R}^{N_q\\times C}\\) with a set of learnable parameters. Then, we aim to convert correspondence queries \\(\\mathbf{F}_q\\) to 2D and 3D keypoint detectors by aggregating information from both the 2D features \\(\\mathbf{F}_I\\) and 3D features \\(\\mathbf{F}_P\\). For this purpose, we iteratively apply Attention-Pixel then Attention-Point as shown in Figure 2. For every iteration \\(k\\geq 0\\), we take queries \\(\\mathbf{F}_q^{2k}\\) as the input and enable interaction with the 2D features \\(\\mathbf{F}_I\\) and 3D features \\(\\mathbf{F}_P\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.762, + 0.905, + 0.783 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {q} ^ {2 k + 1} = \\operatorname {A t t e n t i o n - P i x e l} \\left(\\mathbf {F} _ {\\mathrm {q}} ^ {2 \\mathrm {k}}, \\mathbf {F} _ {\\mathrm {I}}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.784, + 0.905, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {q} ^ {2 k + 2} = \\operatorname {A t t e n t i o n - P o i n t} \\left(\\mathbf {F} _ {\\mathrm {q}} ^ {2 k + 1}, \\mathbf {F} _ {\\mathrm {P}}, \\mathbf {M} _ {\\mathrm {p}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.819, + 0.905, + 0.854 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_q^0 = \\mathbf{F}_q\\). These two steps are executed alternately for \\(L\\) times and we get updated queries \\(\\mathbf{F}_q^{2L}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.854, + 0.907, + 0.904 + ], + "angle": 0, + "content": "Attention-Pixel. For the \\(k\\)-th iteration, we utilize queries \\(\\mathbf{F}_q^{2k}\\) from previous layer to aggregate information from 2D features \\(\\mathbf{F}_I\\) and obtain \\(\\mathbf{F}_q^{2k+1}\\). Specially, keys and values" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16925" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.091, + 0.478, + 0.11 + ], + "angle": 0, + "content": "arise from the 2D features \\(\\mathbf{F}_I\\), and queries arise from \\(\\mathbf{F}_q^{2k}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.117, + 0.482, + 0.137 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathbf {F} _ {q} ^ {2 k} \\mathbf {W} ^ {\\mathcal {Q}}, \\mathbf {K} = \\mathbf {F} _ {I} \\mathbf {W} ^ {\\mathcal {K}}, \\mathbf {V} = \\mathbf {F} _ {I} \\mathbf {W} ^ {\\mathcal {V}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.145, + 0.483, + 0.193 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}^{\\mathcal{Q}},\\mathbf{W}^{\\mathcal{K}},\\mathbf{W}^{\\mathcal{V}}\\in \\mathbb{R}^{C\\times d_k}\\) are linear projections. Then, the queries \\(\\mathbf{F}_q^{2k}\\) are updated with the multi-head attention mechanism [40]:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.2, + 0.483, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\mathrm {S o f t m a x} (\\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {d _ {k}}}) \\mathbf {V}. \\qquad (9)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.483, + 0.272 + ], + "angle": 0, + "content": "And the attention features are further projected with a shallow MLP as the final output features \\(\\mathbf{F}_a^{2k + 1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.483, + 0.379 + ], + "angle": 0, + "content": "Attention-Point. For the \\(k\\)-th iteration, we utilize the queries \\(\\mathbf{F}_q^{2k + 1}\\) from previous layer to aggregate information from the 3D features \\(\\mathbf{F}_P\\) and obtain \\(\\mathbf{F}_q^{2k + 2}\\). The whole process is almost the same as described in Attention-Pixel. The only difference is that we utilize the point cloud mask \\(\\mathbf{M}_p\\) as the keys' mask to make queries focus on relevant overlapping regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.484, + 0.515 + ], + "angle": 0, + "content": "After \\(L\\) iterations, we get updated queries \\(\\mathbf{F}_q^{2L}\\). These updated queries can adaptively change according to the content of the input image and point cloud, capturing the same patterns in both. Next, We apply Attention-Point and Attention-Pixel respectively, taking the updated queries \\(\\mathbf{F}_q^{2L}\\) as the input to get 3D keypoint detectors \\(\\mathbf{D}_P\\) and 2D keypoint detectors \\(\\mathbf{D}_I\\). Then with 2D keypoint detectors \\(\\mathbf{D}_I\\) and 2D features \\(\\mathbf{F}_I\\), we get 2D keypoint heatmap \\(\\mathbf{H}_I \\in \\mathbb{R}^{N_q \\times HW}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.524, + 0.482, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {I} = \\operatorname {S o f t m a x} \\left(\\mathbf {D} _ {I} \\operatorname {F l a t t e n} \\left(\\mathbf {F} _ {I}\\right) ^ {\\top}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.58 + ], + "angle": 0, + "content": "where \\(\\mathrm{Flatten}(\\cdot)\\) denotes combination of the first two dimensions into a single dimension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.626 + ], + "angle": 0, + "content": "3D keypoint heatmap \\(\\mathbf{H}_P\\in \\mathbb{R}^{N_q\\times N}\\) is generated in a similar way and our predicted mask \\(\\mathbf{M}_p\\) is applied to suppress irrelevant points." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.657 + ], + "angle": 0, + "content": "At last, We can get 3D keypoints \\(\\mathbf{K}_P\\in \\mathbb{R}^{N_q\\times 3}\\) and 2D keypoints \\(\\mathbf{K}_I\\in \\mathbb{R}^{N_q\\times 2}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.667, + 0.483, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} _ {P} = \\mathbf {H} _ {P} \\mathbf {P}, \\mathbf {K} _ {I} = \\mathbf {H} _ {I} \\mathbf {E} _ {I}. \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.692, + 0.314, + 0.708 + ], + "angle": 0, + "content": "3.4. Pose Regression Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.483, + 0.837 + ], + "angle": 0, + "content": "After establishing 2D-3D correspondences, we wish to regress camera pose through a learned network. As we have estimated the frustum pose \\(\\mathbf{T}_f\\) in the GPDM, which is a coarse camera pose estimation, we can regard \\(\\mathbf{T}_f\\) as an initial camera pose to be refined and apply corresponding transformation to the 3D keypoints \\(\\mathbf{K}_P\\). In this case, we estimate the difference between \\(\\mathbf{R}_f\\), \\(\\mathbf{t}_f\\) and the ground-truth camera pose \\(\\mathbf{R}_{gt}\\), \\(\\mathbf{t}_{gt}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.844, + 0.483, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {R} _ {g t} = \\mathbf {R} _ {g t} \\cdot \\left(\\mathbf {R} _ {f}\\right) ^ {- 1}, \\Delta \\mathbf {t} _ {g t} = \\mathbf {t} _ {g t} - \\Delta \\mathbf {R} _ {g t} \\cdot \\mathbf {t} _ {f}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As shown in the top-right part of Figure 2, we first apply a MLP to lift the correspondence features and pair the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.123 + ], + "angle": 0, + "content": "correspondence via feature concatenation to get the fusion features \\(\\mathbf{F}_f\\in \\mathbb{R}^{N_q\\times C_f}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.132, + 0.905, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {f} = \\operatorname {M L P} \\left(\\left[ \\mathbf {D} _ {P}, \\operatorname {M L P} \\left(\\mathbf {K} _ {P}\\right), \\mathbf {D} _ {I}, \\operatorname {M L P} \\left(\\mathbf {K} _ {I}\\right) \\right]\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.158, + 0.906, + 0.203 + ], + "angle": 0, + "content": "We then inject global information into the point-pixel correspondences in \\(\\mathbf{F}_f\\) by concatenating its averaged feature to strengthen the correspondence information and obtain \\(\\mathbf{F}_s\\):" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.203, + 0.906, + 0.234 + ], + "angle": 0, + "content": "A pose-sensitive feature \\(\\mathbf{f}_{\\text {pose }}\\) is learned from all the correspondence information via an average-pooling operation:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.248, + 0.905, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\text {p o s e}} = \\operatorname {A v g p o o l} (\\operatorname {M L P} \\left(\\mathbf {F} _ {s}\\right)). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.906, + 0.3 + ], + "angle": 0, + "content": "Finally, similar to Equation 1, we apply two parallel MLPs to regress \\(\\Delta \\mathbf{R}\\) and \\(\\Delta \\mathbf{t}\\) respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.31, + 0.905, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {R}, \\Delta \\mathbf {t} = \\rho (\\operatorname {M L P} (\\mathbf {f} _ {\\text {p o s e}})), \\operatorname {M L P} (\\mathbf {f} _ {\\text {p o s e}}). \\tag {15}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.335, + 0.661, + 0.35 + ], + "angle": 0, + "content": "3.5. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.906, + 0.433 + ], + "angle": 0, + "content": "Given ground-truth camera pose \\(\\mathbf{R}_{gt}\\), \\(\\mathbf{t}_{gt}\\) and camera intrinsic matrix \\(\\mathbf{K}\\), we introduce a joint loss function consisting of a classification loss \\(\\mathcal{L}_{cls}\\), a frustum-lose loss \\(\\mathcal{L}_{fru}\\), a correspondence loss \\(\\mathcal{L}_{corr}\\), a diversity loss \\(\\mathcal{L}_{div}\\) and a camera-pose loss \\(\\mathcal{L}_{cam}\\) to guide the model learning effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.906, + 0.508 + ], + "angle": 0, + "content": "Classification loss. According to Equation 4 and Equation 5, we can get the ground-truth point cloud classification labels \\(\\mathbf{O}_{gt}\\) taking the ground-truth camera pose \\(\\mathbf{R}_{gt}\\), \\(\\mathbf{t}_{gt}\\) and camera intrinsic matrix \\(\\mathbf{K}\\) as inputs. We adopt weighted binary cross-entropy as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.518, + 0.905, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = \\operatorname {B C E L o s s} \\left(\\mathbf {O} _ {g t}, \\mathbf {O} _ {P}\\right). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.906, + 0.575 + ], + "angle": 0, + "content": "Frustum-losess. As ground-truth camera pose is also ground truth of the frustum pose,We use \\(L_{2}\\) loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.589, + 0.905, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f r u} = \\left\\| \\mathbf {R} _ {g t} - \\mathbf {R} _ {f} \\right\\| _ {2} + \\left\\| \\mathbf {t} _ {g t} - \\mathbf {t} _ {f} \\right\\| _ {2}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.641 + ], + "angle": 0, + "content": "Correspondence loss. To establish accurate 2D-3D correspondences, we adopt \\(L_{2}\\) loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.651, + 0.905, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o r r}} = \\| f \\left(\\mathbf {K} _ {P}; \\mathbf {K}, \\mathbf {R} _ {g t}, \\mathbf {t} _ {g t}\\right) - \\mathbf {K} _ {I} \\| _ {2}, \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.676, + 0.736, + 0.691 + ], + "angle": 0, + "content": "where \\(f\\) is defined in Equation 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.737 + ], + "angle": 0, + "content": "Diversity loss. For the goal of expanding the discrepancy among 3D keypoints \\(\\mathbf{K}_P\\) and 2D keypoints \\(\\mathbf{K}_I\\), we impose the diversity loss \\(\\mathcal{L}_{div} = \\mathcal{L}_{2d} + \\mathcal{L}_{3d}\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.746, + 0.905, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {3 d} = \\frac {1}{N _ {q}} \\sum_ {j = 1} ^ {N _ {q}} \\sum_ {k = 1} ^ {N _ {q}} \\max \\left(0, \\Delta_ {3 d} - \\| \\mathbf {K} _ {P} ^ {j} - \\mathbf {K} _ {P} ^ {k} \\| _ {2}\\right), \\tag {19}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.793, + 0.905, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2 d} = \\frac {1}{N _ {q}} \\sum_ {j = 1} ^ {N _ {q}} \\sum_ {k = 1} ^ {N _ {q}} \\max \\left(0, \\Delta_ {2 d} - \\| \\mathbf {K} _ {I} ^ {j} - \\mathbf {K} _ {I} ^ {k} \\| _ {2}\\right). \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.845, + 0.906, + 0.876 + ], + "angle": 0, + "content": "Camera-posediscovery. Similar to \\(\\mathcal{L}_{fru}\\) ,we define the camerepose loss \\(\\mathcal{L}_{cam}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.885, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c a m} = \\left\\| \\Delta \\mathbf {R} _ {g t} - \\Delta \\mathbf {R} \\right\\| _ {2} + \\left\\| \\Delta \\mathbf {t} _ {g t} - \\Delta \\mathbf {t} \\right\\| _ {2}, \\tag {21}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16926" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.393, + 0.106 + ], + "angle": 0, + "content": "where \\(\\Delta_{\\mathbf{R}}\\) and \\(\\Delta_{\\mathrm{t}}\\) are defined in Equation 12." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.483, + 0.137 + ], + "angle": 0, + "content": "Finally, we combine these loss functions by a weighted sum to train our model:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.15, + 0.483, + 0.168 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c a m}} + \\mathcal {L} _ {\\text {f r u}} + \\alpha_ {1} \\mathcal {L} _ {\\text {c l s}} + \\alpha_ {2} \\mathcal {L} _ {\\text {c o r r}} + \\alpha_ {3} \\mathcal {L} _ {\\text {d i v}}, \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.179, + 0.479, + 0.195 + ], + "angle": 0, + "content": "where \\(\\alpha_{1},\\alpha_{2},\\alpha_{3}\\) are weight terms to balance these losses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.208, + 0.224, + 0.225 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.233, + 0.483, + 0.295 + ], + "angle": 0, + "content": "In this section, we first introduce implementation details. Then, we show experimental results and visualizations on two public benchmarks. Finally, we conduct a series of ablation studies to verify the effectiveness of each component." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.303, + 0.308, + 0.319 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.485, + 0.629 + ], + "angle": 0, + "content": "In this work, we implement the proposed model in Pytorch [24] and adopt an NVIDIA RTX 3090 GPU for training. We adopt a 4-stage ResNet [18] with FPN [22] as the image backbone network, where the output channel dimension is 128. For the 3D backbone, we use a 4-stage KPFCNN [39] where the output channel dimension is 128. We set the channel dimension \\( C \\) to 128, set \\( C_f \\) to 256 and set \\( C_{pose} \\) to 512. All the attention layers have 128 features channels with 4 attention heads and ReLU activation. The number of correspondence queries \\( N_q \\) is set to 128. And the iteration time \\( L \\) is set to 3. The weight term \\( \\alpha_1 \\), \\( \\alpha_2 \\) and \\( \\alpha_3 \\) in the loss function are set to 1, 0.1 and 0.1. The margin \\( \\Delta_{3d} \\) and \\( \\Delta_{2d} \\) in the diversity loss are set to 1m and 5 pixels, respectively. All parameters in our proposed model are randomly initialized and trained from scratch with the Adam optimizer. We first train the geometric prior-guided overlapping region detection module with classification loss and frustum-pose loss for 20 epochs. After that, we train the whole network with the total loss \\( \\mathcal{L}_{total} \\) for 20 epochs. The learning rate is set to \\( 10^{-4} \\), and the weight decay is \\( 10^{-6} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.378, + 0.652 + ], + "angle": 0, + "content": "4.2. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.689 + ], + "angle": 0, + "content": "We evaluate our proposed method on two commonly used benchmarks KITTI [17] and nuScenes [3]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.84 + ], + "angle": 0, + "content": "KITTI Odometry [17]. KITTI Odometry dataset consists of 22 sequences of images and point cloud data collected simultaneously in urban environments, and 11 sequences provide ground-truth calibration files. We follow the common practice [21] of utilizing the 0-8 sequences for training, and 9-10 for testing. The mis-registration transformation is conducted with a 2D translation on the ground within \\(\\pm 10\\mathrm{m}\\) and a rotation around the up-axis with no limited range. We downsample the image resolution to \\(160\\times 512\\) and the point cloud size to 40960 for training and testing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "nuScenes [3]. nuScenes dataset collects the image-point cloud pairs by official SDK where the point cloud is accumulated from the nearby frames and the image is from the current frame. We follow the official data spilt of nuScenes," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "where 850 scenes are used for training, and 150 scenes for testing. The mis-registration transformation is conducted in a similar way as the one in KITTI dataset. We downsample the image resolution to \\(160 \\times 320\\) and the point cloud size to 40960 for training and testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Evaluation metrics. We follow VP2P-match [44] to quantify the performance of image-to-point cloud registration with average Relative Translational Error (RTE), average Relative Rotation Error (RRE) and registration accuracy (Acc). Same as VP2P-match [44], we do not follow CorrI2P [28] to remove the image-point cloud samples with large errors before averaging, which might be an unsuitable way to report real performances. Also, We further report the registration accuracy (Acc) which is the proportion of fine registrations with RTE \\(< 2\\mathrm{m}\\) and RRE \\(< 5^{\\circ}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.332, + 0.722, + 0.348 + ], + "angle": 0, + "content": "4.3. Registration Accuracy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.906, + 0.385 + ], + "angle": 0, + "content": "Baselines. We compare our method with previous state-of-the-art methods [18, 28, 44]:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.386, + 0.906, + 0.446 + ], + "angle": 0, + "content": "- Grid Cls. + PnP. DeepI2P [18] proposes to divide the image into \\(32 \\times 32\\) grids and uses a neural network to classify each 3D point to a unique 2D grid. The EPnP [20] with RANSAC [16] is applied to estimate camera pose." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.447, + 0.905, + 0.506 + ], + "angle": 0, + "content": "- Frus. Cls. + Inv.Proj. DeepI2P [18] proposes to perform frustum classification and inverse camera projection to obtain camera pose. We report their results with 2D/3D inverse camera projection as DeepI2P(2D)/DeepI2P(3D)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.507, + 0.905, + 0.551 + ], + "angle": 0, + "content": "- Corr12P. Corr12P [28] proposes to learn dense correspondences for image-point cloud pairs, and the EPnP [20] with RANSAC [16] is applied to predict the camera pose." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.552, + 0.905, + 0.612 + ], + "angle": 0, + "content": "- VP2P-match. VP2P-match [44] is the latest work on image-to-point cloud registration which represents the 3D elements as the combination of voxels and points and trains with a differentiable PnP solver [4]." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.386, + 0.906, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Quantitative Comparison. We report the quantitative results in Table 1. Our proposed method outperforms all compared methods in all metrics by a large margin on both KITTI and nuScenes datasets. In detail, on the KITTI dataset, our method achieves the highest Acc, with a reduction of \\(0.55\\mathrm{m}\\) in RTE and about \\(2.05^{\\circ}\\) in RRE compared with VP2P-match [44]. On the nuScenes dataset, Acc is \\(2.61\\%\\) higher, RTE is reduced by \\(0.26\\mathrm{m}\\) and RRE is reduced by \\(0.02^{\\circ}\\). We attribute the superior performance to three delicate designs. Our proposed GPDM leverages geometric prior that the overlapping region is a frustum to enhance the accuracy of point cloud overlapping region detection. And the ICLM can implicitly obtain accurate 2D-3D correspondences, avoiding the unreliable process of forced alignment between 2D and 3D features. Moreover, the PRM generates precise camera pose prediction via a learned network, which enables the end-to-end optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Qualitative Comparison. The qualitative comparison is shown in Figure 3. For intuitive visualization, we project" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16927" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.09, + 0.872, + 0.103 + ], + "angle": 0, + "content": "Table 1. Registration accuracy on the KITTI and nuScenes datasets. Lower is better for RTE and RRE, higher is better for Acc." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.105, + 0.86, + 0.23 + ], + "angle": 0, + "content": "
MethodsKITTInuScenes
RTE(m)↓RRE(°)↓Acc(%)↑RTE(m)↓RRE(°)↓Acc(%)↑
Grid Cls. + PnP [18]3.64±3.4619.19±28.9611.223.02±2.4012.66±21.012.45
DeepI2P(3D) [18]4.06±3.5424.73±31.693.772.88±2.1220.65±12.242.26
DeepI2P(2D) [18]3.59±3.2111.66±18.1625.952.78±1.994.80±6.2138.10
CorrI2P [28]3.78±65.165.89±20.3472.423.04±60.763.73±9.0349.00
VP2P-match [44]0.75±1.133.29±7.9983.040.89±1.442.15±7.0388.33
Ours0.20±0.211.24±2.3497.490.63±0.442.13±3.7590.94
" + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.237, + 0.234, + 0.25 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.253, + 0.307, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.299, + 0.27, + 0.31 + ], + "angle": 0, + "content": "\\(7.14\\mathrm{m} / 108.94^{\\circ}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.236, + 0.444, + 0.248 + ], + "angle": 0, + "content": "VP2P-match" + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.253, + 0.495, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.299, + 0.452, + 0.31 + ], + "angle": 0, + "content": "2.47m / 12.93°" + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.238, + 0.606, + 0.248 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.254, + 0.676, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.299, + 0.63, + 0.31 + ], + "angle": 0, + "content": "\\(0.22\\mathrm{m} / 0.14^{\\circ}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.237, + 0.785, + 0.248 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.254, + 0.859, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.299, + 0.828, + 0.312 + ], + "angle": 0, + "content": "RTE(m) / RRE(°)" + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.314, + 0.307, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.36, + 0.262, + 0.371 + ], + "angle": 0, + "content": "9.91m/10.53°" + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.314, + 0.495, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.36, + 0.445, + 0.371 + ], + "angle": 0, + "content": "\\(0.88\\mathrm{m} / 2.07^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.314, + 0.676, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.359, + 0.629, + 0.371 + ], + "angle": 0, + "content": "\\(0.07\\mathrm{m} / 1.40^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.314, + 0.86, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.359, + 0.828, + 0.372 + ], + "angle": 0, + "content": "RTE(m) / RRE(°)" + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.374, + 0.307, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.419, + 0.269, + 0.43 + ], + "angle": 0, + "content": "\\(12.93\\mathrm{m} / 70.03^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.374, + 0.495, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.419, + 0.452, + 0.43 + ], + "angle": 0, + "content": "\\(5.38\\mathrm{m} / 18.78^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.374, + 0.676, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.419, + 0.63, + 0.43 + ], + "angle": 0, + "content": "\\(0.07\\mathrm{m} / 1.31^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.374, + 0.86, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.419, + 0.828, + 0.432 + ], + "angle": 0, + "content": "RTE(m) / RRE(°)" + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.435, + 0.792, + 0.449 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison of Image-to-Point Cloud registration results under KITTI dataset." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.473, + 0.278, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.552, + 0.23, + 0.563 + ], + "angle": 0, + "content": "\\(0.16\\mathrm{m} / 0.82^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.473, + 0.455, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.552, + 0.411, + 0.563 + ], + "angle": 0, + "content": "\\(0.04\\mathrm{m} / 0.41^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.572, + 0.278, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.652, + 0.235, + 0.662 + ], + "angle": 0, + "content": "\\(0.08\\mathrm{m} / 0.70^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.573, + 0.454, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.652, + 0.407, + 0.662 + ], + "angle": 0, + "content": "\\(0.14\\mathrm{m} / 0.59^{\\circ}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.664, + 0.483, + 0.691 + ], + "angle": 0, + "content": "Figure 4. Visual illustration of 2D-3D correspondences and registration accuracy. Green lines represent correct correspondences." + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.694, + 0.443, + 0.707 + ], + "angle": 0, + "content": "Table 2. The effect of each design in our framework." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.708, + 0.468, + 0.788 + ], + "angle": 0, + "content": "
RTE(m)↓RRE(°)↓Acc(%)↑
w/o GPDM0.45±1.052.48±9.6891.89
w/o ICLM1.05±0.733.70±6.2072.64
w/o PRM1.05±2.934.96±18.4782.00
Full0.20±0.241.24±2.4097.49
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "the point cloud into image space through predicted transformations of different approaches and the known camera intrinsic matrix, and the color indicates the depth. Compared with VP2P-match [44], our method achieves better registration accuracy in different road situations. For example, on the difficult situation (e.g. the 1st and 3rd row), VP2P-match [44] fails to solve correct registration, while" + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.465, + 0.64, + 0.474 + ], + "angle": 0, + "content": "w/o Geometric Prior" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.474, + 0.646, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.554, + 0.617, + 0.562 + ], + "angle": 0, + "content": "0.86/0.76" + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.465, + 0.748, + 0.474 + ], + "angle": 0, + "content": "w Geometric Prior" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.474, + 0.758, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.554, + 0.729, + 0.562 + ], + "angle": 0, + "content": "0.96/1.00" + }, + { + "type": "image_caption", + "bbox": [ + 0.814, + 0.465, + 0.831, + 0.473 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.474, + 0.876, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.554, + 0.855, + 0.563 + ], + "angle": 0, + "content": "Precise/Recall" + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.564, + 0.609, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.646, + 0.617, + 0.654 + ], + "angle": 0, + "content": "0.79/0.76" + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.564, + 0.717, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.645, + 0.728, + 0.654 + ], + "angle": 0, + "content": "0.99 / 0.95" + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.565, + 0.841, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.646, + 0.855, + 0.654 + ], + "angle": 0, + "content": "Precise/Recall" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.713 + ], + "angle": 0, + "content": "Figure 5. Visual illustration of overlapping region detection. The green points are correctly detected in the overlapping region, the blue points are missed in the overlapping region, and the red points are incorrectly detected outside the overlapping region." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.793 + ], + "angle": 0, + "content": "our method achieves much higher registration accuracy. We further provide the 2D-3D correspondence visualization in Figure 4. Almost all of the pixels of our method can reach slight errors within 15 pixels, which demonstrates that our ICLM can obtain accurate 2D-3D correspondences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.671, + 0.819 + ], + "angle": 0, + "content": "4.4. Ablation Study." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.87 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to justify the effectiveness of each design and the effect of some important parameters in our method on the KITTI dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To justify the effectiveness of each design in our proposed method, we develop three different variations for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16928" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.231, + 0.09, + 0.768, + 0.103 + ], + "angle": 0, + "content": "Table 3. Comparison of frustum classification results on the KITTI and nuScenes datasets." + }, + { + "type": "table", + "bbox": [ + 0.148, + 0.104, + 0.847, + 0.169 + ], + "angle": 0, + "content": "
SettingsKITTInuScenes
PrecisionRecallAccuracyF1-scorePrecisionRecallAccuracyF1-score
w/o GPDM93.4595.9097.9394.6682.8685.5993.1484.20
w GPDM95.1594.9198.0495.0392.0491.5396.3691.79
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.482, + 0.298 + ], + "angle": 0, + "content": "comparison in Table 2: (1) \\( w / o \\) GPDM removes the frustum pose prediction branch and regards points with infrustum probability \\( \\mathbf{O}_P \\) greater than 0.5 as the overlapping region. (2) \\( w / o \\) ICLM removes both the ICLM and PRM since we can't make the PRM work without the ICLM. In this case, we take the predicted frustum pose as our estimated camera pose. (3) \\( w / o \\) PRM replaces the PRM with the PnP-RANSAC algorithm [16, 20]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.483, + 0.512 + ], + "angle": 0, + "content": "Effects of the GPDM. As shown in Table 2, with GPDM, the performance on the KITTI dataset is improved on all metrics. In specific, Acc is \\(6\\%\\) higher, RTE is reduced by about \\(56\\%\\) and RRE is reduced by about \\(50\\%\\). The main reason is that our GPDM can accurately detect the overlapping region and generate a coarse camera pose which can be refined later. And to prove that our GPDM can improve the accuracy of point cloud overlapping region detection, we report the quantitative results in Table 3. With GPDM, we get better frustum classification results almost on all metrics, especially on the nuScenes dataset. The visual comparison is shown in Figure 5, which demonstrates our propose GPDM can get better prediction on the frustum boundary leveraging our proposed geometric prior." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.514, + 0.483, + 0.636 + ], + "angle": 0, + "content": "Effects of the ICLM. As shown in Table 2, with proposed ICLM, the performance on the KITTI dataset is improved notably. In specific, Acc is \\(34\\%\\) higher, RTE is reduced by about \\(81\\%\\) and RRE is reduced by about \\(66\\%\\). Regarding the gain on Acc, we can figure out that the ICLM plays the most important role in our model, which demonstrates that our designed implicit correspondence learning is very effective for Image-to-Point Cloud registration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.483, + 0.714 + ], + "angle": 0, + "content": "Effects of the PRM. As shown in Table 2, the PRM outperforms the PnP-RANSAC algorithm [16, 20] on all metrics. In specific, Acc is \\(19\\%\\) higher, RTE is reduced by about \\(81\\%\\) and RRE is reduced by about \\(75\\%\\). We owe the advantage to the enabled end-to-end optimization with PRM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Effects of the number of correspondence queries. In Table 4, we show the impact of different numbers of correspondence queries in the ICLM on the KITTI dataset. The results show that our method can achieve better performance than VP2P-match [44] by using only 16 correspondence queries. We find that the overall performance of the model improves slightly with the increase of \\( N_{q} \\). We choose \\( N_{q} = 128 \\) in our method for the highest Acc." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effects of the iteration time. To investigate the influences of the iteration time \\((L)\\) in the ICLM, we pick \\(L\\) from the set \\(\\{0, 1, 2, 3, 4\\}\\). As shown in Table 5, we notice that the performance drops a lot if \\(L\\) is set to 0, showing the" + }, + { + "type": "table_caption", + "bbox": [ + 0.557, + 0.175, + 0.862, + 0.188 + ], + "angle": 0, + "content": "Table 4. Ablation studies on the number of queries." + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.189, + 0.861, + 0.285 + ], + "angle": 0, + "content": "
NqRTE(m)↓RRE(°)↓Acc(%)↑
160.26±0.291.71±2.4095.52
320.23±0.251.43±2.2495.70
640.21±0.191.28±3.2096.63
1280.20±0.241.24±2.4097.49
2560.21±0.211.21±2.4097.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.571, + 0.288, + 0.848, + 0.3 + ], + "angle": 0, + "content": "Table 5. Ablation studies on the iteration time." + }, + { + "type": "table", + "bbox": [ + 0.564, + 0.301, + 0.855, + 0.397 + ], + "angle": 0, + "content": "
LRTE(m)↓RRE(°)↓Acc(%)↑
00.31±0.362.34±2.7690.01
10.21±0.161.23±1.9597.44
20.20±0.161.28±2.0297.33
30.20±0.241.24±2.4097.49
40.21±0.301.26±2.4797.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.57, + 0.4, + 0.85, + 0.413 + ], + "angle": 0, + "content": "Table 6. Ablation studies on the iteration order." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.415, + 0.904, + 0.464 + ], + "angle": 0, + "content": "
RTE(m)↓RRE(°)↓Acc(%)↑
Pixel-then-Point0.20±0.241.24±2.4097.49
Point-then-Pixel0.23±0.601.29±2.2697.05
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.905, + 0.518 + ], + "angle": 0, + "content": "necessity of the iteration operation. We choose the model with \\( L = 1 \\) for the optimal balance between computational efficiency and performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.906, + 0.624 + ], + "angle": 0, + "content": "Effects of different iteration orders. To investigate the influences of the iteration order in the ICLM, we conduct two experiments with first Attention-Pixel then Attention-Point or first Attention-Point then Attention-Pixel. As shown is Table 6, we get similar performance with different iteration orders. And we choose to iteratively apply Attention-Pixel then Attention-Point in our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.639, + 0.634, + 0.655 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.906, + 0.831 + ], + "angle": 0, + "content": "In this paper, we propose a novel implicit correspondence learning method with three specific designs for image-top-point cloud registration. Specially, the proposed geometric prior can guide the model to achieve high accuracy on the overlapping region detection. The proposed implicit correspondence learning module can implicitly acquire accurate 2D-3D correspondences. Furthermore, we propose a pose regression module to handle incorrect 2D-3D correspondences and generates precise camera pose. Extensive experimental results on two challenging benchmarks demonstrate the effectiveness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.846, + 0.702, + 0.863 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Defense Basic Scientific Research program (No.JCKY2022911B002)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16929" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Xuyang Bai, Zixin Luo, Lei Zhou, Hongbo Fu, Long Quan, and Chiew-Lan Tai. D3feat: Joint learning of dense detection and description of 3d local features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6359-6367, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Mark Billinghurst, Adrian Clark, Gun Lee, et al. A survey of augmented reality. Foundations and Trends® in Human-Computer Interaction, 8(2-3):73-272, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.484, + 0.311 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.394 + ], + "angle": 0, + "content": "[4] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2781-2790, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[5] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 539-546. IEEE, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.483, + 0.534 + ], + "angle": 0, + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8958-8966, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[8] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In Proceedings of the European conference on computer vision (ECCV), pages 602-618, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[9] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 195-205, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Zhen Dong, Fuxun Liang, Bisheng Yang, Yusheng Xu, Yufu Zang, Jianping Li, Yuan Wang, Wenxia Dai, Hongchao Fan, Juha Hyyppä, et al. Registration of large-scale terrestrial laser scanner point clouds: A review and benchmark. IS-PRS Journal of Photogrammetry and Remote Sensing, 163: 327–342, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "object recognition. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 998-1005. IEEE, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[13] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99–110, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[14] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 8092-8101, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.263, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[15] Mengdan Feng, Sixing Hu, Marcelo H Ang, and Gim Hee Lee. 2d3d-matchnet: Learning to match keypoints across 2d image and 3d point cloud. In 2019 International Conference on Robotics and Automation (ICRA), pages 4790-4796. IEEE, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[17] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[19] Yurim Jeon and Seung-Woo Seo. Efghnet: A versatile image-to-point cloud registration network for extreme outdoor environment. IEEE Robotics and Automation Letters, 7 (3):7511-7517, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[20] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate \\(O(n)\\) solution to the pnp problem. International journal of computer vision, 81:155-166, 2009. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.907, + 0.674 + ], + "angle": 0, + "content": "[21] Jiaxin Li and Gim Hee Lee. Deepi2p: Image-to-point cloud registration via deep classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15960-15969, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[23] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[24] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16930" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[27] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143-11152, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[28] Siyu Ren, Yiming Zeng, Junhui Hou, and Xiaodong Chen. Corri2p: Deep image-to-point cloud registration via dense correspondence. IEEE Transactions on Circuits and Systems for Video Technology, 33(3):1198-1208, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[29] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[30] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.483, + 0.483 + ], + "angle": 0, + "content": "[31] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE international conference on robotics and automation, pages 3212-3217. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[32] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.649 + ], + "angle": 0, + "content": "[33] Paul-Edouard Sarlin, Daniel DeTone, Tsun-Yi Yang, Armen Avetisyan, Julian Straub, Tomasz Malisiewicz, Samuel Rota Bulo, Richard Newcombe, Peter Kontschieder, and Vasileios Balntas. Orienternet: Visual localization in 2d public maps with neural matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21632–21642, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.649, + 0.483, + 0.706 + ], + "angle": 0, + "content": "[34] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. 3rd International Conference on Learning Representations, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[36] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[37] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6398-6407, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.16 + ], + "angle": 0, + "content": "[38] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European conference on computer vision, pages 685–702. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.161, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[39] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.906, + 0.316 + ], + "angle": 0, + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.906, + 0.387 + ], + "angle": 0, + "content": "[41] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loptr: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.906, + 0.444 + ], + "angle": 0, + "content": "[42] Hao Yu, Fu Li, Mahdi Saleh, Benjamin Busam, and Slobodan Ilic. Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration. Advances in Neural Information Processing Systems, 34:23872-23884, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.444, + 0.906, + 0.5 + ], + "angle": 0, + "content": "[43] Yu Zhong. Intrinsic shape signatures: A shape descriptor for 3d object recognition. In 2009 IEEE 12th international conference on computer vision workshops, ICCV Workshops, pages 689-696. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.5, + 0.906, + 0.584 + ], + "angle": 0, + "content": "[44] Junsheng Zhou, Baorui Ma, Wenyuan Zhang, Yi Fang, Yu-Shen Liu, and Zhizhong Han. Differentiable registration of images and lidar point clouds with voxelpoint-to-pixel matching. In Advances in Neural Information Processing Systems, pages 51166-51177. Curran Associates, Inc., 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.584, + 0.906, + 0.642 + ], + "angle": 0, + "content": "[45] Qunjie Zhou, Torsten Sattler, and Laura Leal-Taixe. Patch2pix: Epipolar-guided pixel-level correspondences. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4669-4678, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.642, + 0.906, + 0.711 + ], + "angle": 0, + "content": "[46] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745-5753, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16931" + } + ] +] \ No newline at end of file diff --git a/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/3b6935ad-40bc-4903-80f6-f62148a4f3d3_origin.pdf b/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/3b6935ad-40bc-4903-80f6-f62148a4f3d3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2b840ebdb1709a1fda308de33eecdbe6857eccc5 --- /dev/null +++ b/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/3b6935ad-40bc-4903-80f6-f62148a4f3d3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f212c209061a463efe37123627ccb5ee9f3c6002ec671c8761be5f9dc0ba9366 +size 2500820 diff --git a/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/full.md b/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8b21f86bde8b0758128b8a1d633855e62573524e --- /dev/null +++ b/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/full.md @@ -0,0 +1,450 @@ +# Implicit Correspondence Learning for Image-to-Point Cloud Registration + +Xinjun Li $^{1}$ , Wenfei Yang $^{1}$ , Jiacheng Deng $^{1}$ , Zhixin Cheng $^{1}$ , Xu Zhou $^{3}$ , Tianzhu Zhang $^{1,2,*}$ $^{1}$ University of Science and Technology of China + +$^{2}$ National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory + $^{3}$ Sangfor Technologies + +{1xj3017,dengjc,chengzhixin}@mail.ustc.edu.cn, {yangwf,tzzhang}@ustc.edu.cn, zhouxu@sangfor.com.cn + +# Abstract + +Image-to-point cloud registration aims to estimate the camera pose of a given image within a 3D scene point cloud. In this area, matching-based methods have achieved leading performance by first detecting the overlapping region, then matching point and pixel features learned by neural networks and finally using the PnP-RANSAC algorithm to estimate camera pose. However, achieving accurate image-to-point cloud registration remains challenging because the overlapping region detection is unreliable merely relying on point-wise classification, direct alignment of cross-modal data is difficult and indirect optimization objective leads to unstable registration results. To address these challenges, we propose a novel implicit correspondence learning method, including a Geometric Prior-guided overlapping region Detection Module (GPDM), an Implicit Correspondence Learning Module (ICLM), and a Pose Regression Module (PRM). The proposed method enjoys several merits. First, the proposed GPDM can precisely detect the overlapping region. Second, the ICLM can generate robust cross-modality correspondences. Third, the PRM can enable end-to-end optimization. Extensive experimental results on KITTI and nuScenes datasets demonstrate that the proposed model sets a new state-of-the-art performance in registration accuracy. + +# 1. Introduction + +Image-to-point cloud registration [15] is a fundamental task in the field of computer vision, aimed at estimating the camera pose of a given image within a 3D scene point cloud. It plays an important role in many computer vision tasks, including simultaneous localization and mapping [13], 3D reconstruction [11], augmented/virtual reality [2], and visual localization [33]. Although significant progress has been made in recent years, image-to-point cloud registra + +![](images/d9729673f7150b13eec4d61f936a5ebc7c15e910f9c8f06b52af23f9e84d7467.jpg) +a) Existing matching-based methods + +![](images/a56b9b234a91892c81682ce350c78aaccd2b2d3a6c66de1609a8a6a585336b97.jpg) +b) Our Implicit Correspondence Learning method +Figure 1. (a) The pipeline of mainstream matching-based methods. The green lines denote correct correspondences, and the red lines denote incorrect correspondences. 2D-3D correspondences are first established by calculating feature similarity. Then the PnP-RANSAC algorithm is applied to obtain camera pose. (b) The pipeline of our proposed method. A set of shared queries are used to aggregate information from image and point cloud to generate 2D and 3D keypoint detectors. With paired 2D and 3D keypoint detectors, we obtain a set of 2D-3D correspondences and estimate camera pose through a pose regression network. + +tion remains challenging due to the inherent disparities that images capture appearances with 2D grid data while point clouds represent structures in an unordered format. + +To address these challenges, numerous methods have been proposed for image-to-point cloud registration [15, 19, 21, 28, 44]. Generally, these methods can be categorized into two main classes, including matching-free methods [19, 21] and matching-based methods [15, 28, 44]. Matching-free methods utilize the camera imaging principle as a prior to predict camera pose. For example, DeepI2P [21] casts the registration problem into a classi- + +fication problem. A binary classification network is designed to determine whether each scene point is within or beyond the camera frustum. Then the classification results are fed into an inverse camera projection solver to obtain camera pose. However, these methods merely determine a coarse image-point correspondence, e.g. whether a 3D point is inside the image frustum, which is insufficient for precise camera pose estimation. Different from matching-free methods, matching-based methods focus on establishing fine-grained 2D-3D correspondences and use the PnP-RANSAC algorithm [16, 20] to estimate camera pose [15, 28, 44], which have achieved superior performance. These methods generally involve three important steps: (1) A detector is designed to identify the overlapping region between the image and point cloud. (2) 2D-3D correspondences are established within the identified overlapping region. (3) The PnP-RANSAC algorithm [16, 20] is applied as a post-processing module to obtain camera pose. + +Despite the success of these matching-based methods, there still exist three notable problems that hinder the registration accuracy of these methods. (1) Unreliable point-based detection of the overlapping region. Most existing methods adopt a point-wise classification way to determine whether each point lies in the overlapping region or not, ignoring the global projection relationship between the image and point cloud, e.g. an image should correspond to a frustum in the point cloud. Consequently, existing methods usually suffer from inaccurate boundary detection of the overlapping region, resulting in some missing 2D-3D correspondences or incorrect correspondences outside the overlapping region. (2) Inaccurate explicit matching between the image and point cloud. As shown in Figure 1 (a), matching-based methods essentially establish 2D-3D correspondences by directly calculating feature similarity between the image and point cloud. These methods employ metric learning techniques like contrastive loss [5], triplet loss [34] or circle loss [37] to force the alignment of corresponding 2D and 3D features. However, due to the inherent data discrepancy between the image and point cloud, direct alignment of cross-modal data can result in a significant number of mismatches, as shown in Figure 1 (a). (3) Indirect correspondence-based optimization objective. After establishing 2D-3D correspondences, the majority of existing approaches employ the PnP-RANSAC algorithm [16, 20] as a post-processing module to estimate camera pose as illustrated in Figure 1 (a). During the training stage, these approaches primarily consider 2D-3D correspondences as the optimization target, without optimizing the pose directly. + +Motivated by the above discussions, we propose a novel implicit correspondence learning method for imaged-to-point cloud registration. The proposed model contains three modules, a Geometric Prior-guided overlapping re + +GPI Detection Module (GPDM), an Implicit Correspondence Learning Module (ICLM), and a Pose Regression Module (PRM). The GPDM leverages geometric prior that the overlapping region is a frustum to enhance the accuracy of the overlapping region detection. Specifically, a binary classification network is designed to determine whether each point lies within or beyond the camera frustum. Then, we predict center coordinate and orientation of the camera frustum using obtained classification results. Finally, we regard points in predicted camera frustum as our overlapping region. Since camera frustum defines a continuous and regular geometric region, the noise's interference on the overlapping region detection can be mitigated. The ICLM is designed to implicitly identify a set of 2D-3D correspondences within the overlapping region. As shown in Figure 1 (b), we initialize a set of learnable correspondence queries. For each query, we convert it to 2D/3D keypoint detector via the attention mechanism [40] to aggregate features from the image and point cloud, respectively. By calculating similarity between 2D/3D keypoint detector and image/point cloud features, a pair of 2D/3D keypoint heatmap is obtained. In this way, we can implicitly obtain a set of 2D-3D paired keypoints, avoiding the process of forced alignment between 2D and 3D features. Finally, the PRM takes obtained 2D/3D keypoint detectors and 2D/3D keypoints as inputs, and regresses camera pose via a Multi-Layer Perceptron (MLP), which enables the end-to-end optimization. The main contributions are summarized as follows: + +- We propose a novel implicit correspondence learning method for image-to-point cloud registration, which enjoys the merits of precise overlapping region detection, robust cross-modality matching, end-to-end optimization. +- Without bells and whistles, extensive experimental results on KITTI and nuScenes datasets demonstrate that the proposed model sets a new state-of-the-art performance in registration accuracy, i.e. a reduction of $0.55\mathrm{m}$ in RTE and $2.05^{\circ}$ in RRE compared to previous SOTA method. + +# 2. Related Works + +In this section, we briefly overview methods that are related to image-to-image registration, point cloud-to-point cloud registration and image-to-point cloud registration. + +Image-to-image registration. Image-to-image registration is the process of aligning two or more images, which aims to find the optimal transformation that best aligns the structures of interest in the input images. Typical methods usually adopt a detect-then-match pipeline to extract correspondences. Keypoints are first detected and described using handcrafted [23, 30] or learning-based techniques [10, 14, 29, 32], correspondences are then established based on their feature similarity. However, robustly detecting repeatable keypoints across different images is still challenging, especially for texture-poor regions. Detection + +free methods [36, 41, 45] propose to bypass the keypoint detection step by first obtaining patch-level matches and then regress pixel-wise matches in a coarse-to-fine manner, which achieves accurate image-to-image registration. + +Point cloud-to-point cloud registration. Point cloud-to-point cloud registration is the process of finding a rigid transformation that best aligns two point clouds. Similar to image-to-image registration, early works leverage hand-crafted descriptors [12, 31] or learning-based descriptors [1, 7-9] matching to estimate the correspondences. CoFiNet [42] follows the design of LoFTR [36] and introduces detection-free registration with a coarse-to-fine strategy. Recent methods have replaced RANSAC [16] with deep robust estimators to enhance both speed and accuracy. For example, GeoTransformer [27] proposes a geometric Transformer and introduces a local-to-global registration for RANSAC-free registration. + +Image-to-point cloud registration. Compared to intramodality registration, image-to-point cloud registration poses more significant difficulties due to the inherent data disparities. The typical pipeline is to first extract 2D-3D correspondences and then estimate camera pose using the PnP-RANSAC algorithm [16, 20]. In this case, the accuracy of extracted correspondences plays a crucial role in ensuring successful registration. 2D3D-MatchNet [15] is the pioneering work, which uses SIFT [23] and ISS [43] to extract the keypoints of the image and point cloud, respectively. Then the patches around the keypoints are fed to CNNs [35] and PointNet [25] to extract features for establishing 2D-3D correspondences. However, the independent hand-crafted keypoint detectors for different modalities leads to a poor registration accuracy. CorrI2P [28] predicts pixels and points in the overlapping region and learns dense 2D-3D correspondences by matching point and pixel features for fine registration. However, it struggles to learn a discriminative latent space for 2D and 3D data, leading to a high rate of incorrect correspondences. Recently, VP2PMatch [44] learns a structured cross-modality latent space by representing 3D elements using both voxels [6, 38] and points [26]. It imposes supervisions directly on the predicted pose distribution with a differentiable probabilistic PnP solver [4]. Though introducing voxel features, it struggles with 2D-3D data disparities. Besides, it fails to achieve high registration accuracy for insufficient supervisions. In addition to these matching-based methods, there are some matching-free methods. For instance, DeepI2P [21] converts the registration problem into a classification and inverse camera projection optimization problem. However, the frustum classification only indicates coarse 2D-3D correspondence, thus limiting the registration accuracy. + +# 3. Method + +Given a 3D scene point cloud $\mathbf{P} \in \mathbb{R}^{N \times 3}$ and an image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ from the same scene, the cross-modality image-to-point cloud registration is to estimate the camera pose $\mathbf{T}_{gt}$ in the coordinate system of $\mathbf{P}$ , including a rotation matrix $\mathbf{R}_{gt} \in \mathbf{SO}(3)$ and a translation vector $\mathbf{t}_{gt} \in \mathbb{R}^3$ . + +An overview of our pipeline is illustrated in Figure 2. Our proposed model mainly consists of a Geometric Prior-guided overlapping region Detection Module (GPDM), an Implicit Correspondence Learning Module (ICLM) and a Pose Regression Module (PRM). Specifically, we first extract features for the image and point cloud respectively. Next, the GPDM detects the overlapping region of the input point cloud. Then the ICLM implicitly identifies a set of correspondences within detected overlapping region via our designed correspondence queries. At last, we estimate the camera pose via the PRM. + +# 3.1. Feature Extraction + +For the image, we use ResNet [18] with FPN [22] to extract 2D features $\mathbf{F}_I\in \mathbb{R}^{H\times W\times C}$ . For simplicity, we denote the pixel coordinate matrices for $\mathbf{F}_I$ as $\mathbf{E}_I\in \mathbb{R}^{HW\times 2}$ . For the input point cloud, we adopt KPFCNN [39] to extract 3D features $\mathbf{F}_P\in \mathbb{R}^{N\times C}$ . + +# 3.2. Geometric Prior-guided Overlapping Region Detection Module + +To precisely detect the overlapping region of the point cloud, most methods adopt a point-wise classification strategy to determine whether each point lies in the overlapping region or not, ignoring the global projection relationship between the image and point cloud. Inspired by the fact that an image should correspond to a frustum in the point cloud, we design a geometric prior-guided classification strategy thus we can get more accurate classification results. + +As shown in the bottom-left part of Figure 2, we first obtain 2D global feature by averaging extracted 2D features $\mathbf{F}_I$ and repeat it to concatenate with the 3D features $\mathbf{F}_P$ . Then a MLP is used to learn the probability $\mathbf{O}_P \in \mathbb{R}^{N \times 1}$ of a point to lie within the camera frustum. + +Since the in-frustum probability $\mathbf{O}_P$ is not precise enough for in-frustum classification, we propose to estimate the frustum pose to get more accurate classification results. Utilizing $\mathbf{O}_P$ and the coordinate information of the point cloud $\mathbf{P}$ , we apply another MLP to regress the frustum pose $\mathbf{T}_f \in \mathbf{SE}(\mathbf{3})$ , composed of a 3D rotation $\mathbf{R}_f \in \mathbf{SO}(\mathbf{3})$ and a 3D translation vector $\mathbf{t}_f \in \mathbb{R}^3$ . In detail, we apply two parallel MLPs to regress $\mathbf{R}_f$ and $\mathbf{t}_f$ respectively: + +$$ +\mathbf {R} _ {f}, \mathbf {t} _ {f} = \rho (\mathrm {M L P} ([ \mathbf {O} _ {P}, \mathbf {P} ])), \mathrm {M L P} ([ \mathbf {O} _ {P}, \mathbf {P} ]), \tag {1} +$$ + +where $[\cdot, \cdot]$ denotes concatenation along feature dimension. We choose a 6D representation of rotation [46] as the regression of the first MLP, for its continuous learning space + +![](images/c38f3556c18f6641645d335518a25e8f2a76a7b534c71764e073091f9489f2aa.jpg) + +![](images/387a90bf422d22dd848cb00a1f75d6322f0257113f9745e18e0c54c6b0a9d066.jpg) +Figure 2. Overview of our method. Given a pair of point cloud $\mathbf{P}$ and image $\mathbf{I}$ as inputs, we first extract features for the image and point cloud respectively. Then, in the geometric prior-guided overlapping region detection module, we leverage geometric prior that the overlapping region is a frustum to classify each point in the point cloud into within or beyond the camera frustum and get the point cloud mask $\mathbf{M}_P$ . Next, in the implicit correspondence learning module, we initialize a set of correspondence queries $\mathbf{F}_q$ and convert them to 2D keypoint detectors $\mathbf{D}_I$ and 3D keypoint detectors $\mathbf{D}_P$ by aggregating extracted 2D and 3D features via the attention mechanism. Weighting the pixel/point coordinates according to the feature similarity between the 2D/3D keypoint detectors and 2D/3D features, we obtain 2D keypoints $\mathbf{K}_I$ and 3D keypoints $\mathbf{K}_P$ , respectively. Finally, in the pose regression module, we utilize the information of 2D-3D paired keypoints and 2D-3D paired keypoint detectors to regress the camera pose $\Delta \mathbf{R}$ and $\Delta t$ . + +in $\mathbf{SO}(3)$ , and $\rho(\cdot)$ denotes the transformation from the 6D representation to the $3 \times 3$ rotation matrix $\mathbf{R}_f$ . Then according to the given camera intrinsic matrix $\mathbf{K} \in \mathbb{R}^{3 \times 3}$ , we define the point-to-pixel projection function as: + +$$ +\left[ x _ {i}, y _ {i}, z _ {i} \right] ^ {\top} = \mathbf {K} \cdot \left(\mathbf {R} _ {f} \cdot \mathbf {p} _ {i} + \mathbf {t} _ {f}\right), \tag {2} +$$ + +$$ +\mathbf {u} _ {i} = \left[ u _ {x i}, u _ {y i} \right] ^ {\top} = \left[ x _ {i} / z _ {i}, y _ {i} / z _ {i} \right] ^ {\top}, \tag {3} +$$ + +$$ +f \left(\mathbf {p} _ {i}; \mathbf {K}, \mathbf {R} _ {f}, \mathbf {t} _ {f}\right) = \mathbf {u} _ {i}, \tag {4} +$$ + +where $\mathbf{p}_i\in \mathbf{P}$ is a 3D point and $\mathbf{u}_i$ represents a 2D pixel. We get the point cloud mask $\mathbf{M}_p\in \mathbb{R}^{N\times 1}$ as: + +$$ +\mathbf {M} _ {p} ^ {i} = \left\{ \begin{array}{l} 1; 0 \leq u _ {x i} < W, 0 \leq u _ {y i} < H, z _ {i} \geq 0 \\ 0; \text {o t h e r w i s e} \end{array} . \right. \tag {5} +$$ + +# 3.3. Implicit Correspondence Learning Module + +After obtaining point cloud mask $\mathbf{M}_p$ , we implicitly identify a set of robust correspondences between the image and point cloud. As shown in the bottom-right part of Figure 2, + +we first initialize $N_{q}$ correspondence queries $\mathbf{F}_q\in \mathbb{R}^{N_q\times C}$ with a set of learnable parameters. Then, we aim to convert correspondence queries $\mathbf{F}_q$ to 2D and 3D keypoint detectors by aggregating information from both the 2D features $\mathbf{F}_I$ and 3D features $\mathbf{F}_P$ . For this purpose, we iteratively apply Attention-Pixel then Attention-Point as shown in Figure 2. For every iteration $k\geq 0$ , we take queries $\mathbf{F}_q^{2k}$ as the input and enable interaction with the 2D features $\mathbf{F}_I$ and 3D features $\mathbf{F}_P$ as follows: + +$$ +\mathbf {F} _ {q} ^ {2 k + 1} = \operatorname {A t t e n t i o n - P i x e l} \left(\mathbf {F} _ {\mathrm {q}} ^ {2 \mathrm {k}}, \mathbf {F} _ {\mathrm {I}}\right), \tag {6} +$$ + +$$ +\mathbf {F} _ {q} ^ {2 k + 2} = \operatorname {A t t e n t i o n - P o i n t} \left(\mathbf {F} _ {\mathrm {q}} ^ {2 k + 1}, \mathbf {F} _ {\mathrm {P}}, \mathbf {M} _ {\mathrm {p}}\right), \tag {7} +$$ + +where $\mathbf{F}_q^0 = \mathbf{F}_q$ . These two steps are executed alternately for $L$ times and we get updated queries $\mathbf{F}_q^{2L}$ . + +Attention-Pixel. For the $k$ -th iteration, we utilize queries $\mathbf{F}_q^{2k}$ from previous layer to aggregate information from 2D features $\mathbf{F}_I$ and obtain $\mathbf{F}_q^{2k+1}$ . Specially, keys and values + +arise from the 2D features $\mathbf{F}_I$ , and queries arise from $\mathbf{F}_q^{2k}$ : + +$$ +\mathbf {Q} = \mathbf {F} _ {q} ^ {2 k} \mathbf {W} ^ {\mathcal {Q}}, \mathbf {K} = \mathbf {F} _ {I} \mathbf {W} ^ {\mathcal {K}}, \mathbf {V} = \mathbf {F} _ {I} \mathbf {W} ^ {\mathcal {V}}, \tag {8} +$$ + +where $\mathbf{W}^{\mathcal{Q}},\mathbf{W}^{\mathcal{K}},\mathbf{W}^{\mathcal{V}}\in \mathbb{R}^{C\times d_k}$ are linear projections. Then, the queries $\mathbf{F}_q^{2k}$ are updated with the multi-head attention mechanism [40]: + +$$ +\mathrm {A t t e n t i o n} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \mathrm {S o f t m a x} (\frac {\mathbf {Q K} ^ {\top}}{\sqrt {d _ {k}}}) \mathbf {V}. \qquad (9) +$$ + +And the attention features are further projected with a shallow MLP as the final output features $\mathbf{F}_a^{2k + 1}$ . + +Attention-Point. For the $k$ -th iteration, we utilize the queries $\mathbf{F}_q^{2k + 1}$ from previous layer to aggregate information from the 3D features $\mathbf{F}_P$ and obtain $\mathbf{F}_q^{2k + 2}$ . The whole process is almost the same as described in Attention-Pixel. The only difference is that we utilize the point cloud mask $\mathbf{M}_p$ as the keys' mask to make queries focus on relevant overlapping regions. + +After $L$ iterations, we get updated queries $\mathbf{F}_q^{2L}$ . These updated queries can adaptively change according to the content of the input image and point cloud, capturing the same patterns in both. Next, We apply Attention-Point and Attention-Pixel respectively, taking the updated queries $\mathbf{F}_q^{2L}$ as the input to get 3D keypoint detectors $\mathbf{D}_P$ and 2D keypoint detectors $\mathbf{D}_I$ . Then with 2D keypoint detectors $\mathbf{D}_I$ and 2D features $\mathbf{F}_I$ , we get 2D keypoint heatmap $\mathbf{H}_I \in \mathbb{R}^{N_q \times HW}$ : + +$$ +\mathbf {H} _ {I} = \operatorname {S o f t m a x} \left(\mathbf {D} _ {I} \operatorname {F l a t t e n} \left(\mathbf {F} _ {I}\right) ^ {\top}\right), \tag {10} +$$ + +where $\mathrm{Flatten}(\cdot)$ denotes combination of the first two dimensions into a single dimension. + +3D keypoint heatmap $\mathbf{H}_P\in \mathbb{R}^{N_q\times N}$ is generated in a similar way and our predicted mask $\mathbf{M}_p$ is applied to suppress irrelevant points. + +At last, We can get 3D keypoints $\mathbf{K}_P\in \mathbb{R}^{N_q\times 3}$ and 2D keypoints $\mathbf{K}_I\in \mathbb{R}^{N_q\times 2}$ as follows: + +$$ +\mathbf {K} _ {P} = \mathbf {H} _ {P} \mathbf {P}, \mathbf {K} _ {I} = \mathbf {H} _ {I} \mathbf {E} _ {I}. \tag {11} +$$ + +# 3.4. Pose Regression Module + +After establishing 2D-3D correspondences, we wish to regress camera pose through a learned network. As we have estimated the frustum pose $\mathbf{T}_f$ in the GPDM, which is a coarse camera pose estimation, we can regard $\mathbf{T}_f$ as an initial camera pose to be refined and apply corresponding transformation to the 3D keypoints $\mathbf{K}_P$ . In this case, we estimate the difference between $\mathbf{R}_f$ , $\mathbf{t}_f$ and the ground-truth camera pose $\mathbf{R}_{gt}$ , $\mathbf{t}_{gt}$ as: + +$$ +\Delta \mathbf {R} _ {g t} = \mathbf {R} _ {g t} \cdot \left(\mathbf {R} _ {f}\right) ^ {- 1}, \Delta \mathbf {t} _ {g t} = \mathbf {t} _ {g t} - \Delta \mathbf {R} _ {g t} \cdot \mathbf {t} _ {f}. \tag {12} +$$ + +As shown in the top-right part of Figure 2, we first apply a MLP to lift the correspondence features and pair the + +correspondence via feature concatenation to get the fusion features $\mathbf{F}_f\in \mathbb{R}^{N_q\times C_f}$ as follows: + +$$ +\mathbf {F} _ {f} = \operatorname {M L P} \left(\left[ \mathbf {D} _ {P}, \operatorname {M L P} \left(\mathbf {K} _ {P}\right), \mathbf {D} _ {I}, \operatorname {M L P} \left(\mathbf {K} _ {I}\right) \right]\right). \tag {13} +$$ + +We then inject global information into the point-pixel correspondences in $\mathbf{F}_f$ by concatenating its averaged feature to strengthen the correspondence information and obtain $\mathbf{F}_s$ : + +A pose-sensitive feature $\mathbf{f}_{\text {pose }}$ is learned from all the correspondence information via an average-pooling operation: + +$$ +\mathbf {f} _ {\text {p o s e}} = \operatorname {A v g p o o l} (\operatorname {M L P} \left(\mathbf {F} _ {s}\right)). \tag {14} +$$ + +Finally, similar to Equation 1, we apply two parallel MLPs to regress $\Delta \mathbf{R}$ and $\Delta \mathbf{t}$ respectively: + +$$ +\Delta \mathbf {R}, \Delta \mathbf {t} = \rho (\operatorname {M L P} (\mathbf {f} _ {\text {p o s e}})), \operatorname {M L P} (\mathbf {f} _ {\text {p o s e}}). \tag {15} +$$ + +# 3.5. Loss Function + +Given ground-truth camera pose $\mathbf{R}_{gt}$ , $\mathbf{t}_{gt}$ and camera intrinsic matrix $\mathbf{K}$ , we introduce a joint loss function consisting of a classification loss $\mathcal{L}_{cls}$ , a frustum-lose loss $\mathcal{L}_{fru}$ , a correspondence loss $\mathcal{L}_{corr}$ , a diversity loss $\mathcal{L}_{div}$ and a camera-pose loss $\mathcal{L}_{cam}$ to guide the model learning effectively. + +Classification loss. According to Equation 4 and Equation 5, we can get the ground-truth point cloud classification labels $\mathbf{O}_{gt}$ taking the ground-truth camera pose $\mathbf{R}_{gt}$ , $\mathbf{t}_{gt}$ and camera intrinsic matrix $\mathbf{K}$ as inputs. We adopt weighted binary cross-entropy as follows: + +$$ +\mathcal {L} _ {c l s} = \operatorname {B C E L o s s} \left(\mathbf {O} _ {g t}, \mathbf {O} _ {P}\right). \tag {16} +$$ + +Frustum-losess. As ground-truth camera pose is also ground truth of the frustum pose,We use $L_{2}$ loss as follows: + +$$ +\mathcal {L} _ {f r u} = \left\| \mathbf {R} _ {g t} - \mathbf {R} _ {f} \right\| _ {2} + \left\| \mathbf {t} _ {g t} - \mathbf {t} _ {f} \right\| _ {2}. \tag {17} +$$ + +Correspondence loss. To establish accurate 2D-3D correspondences, we adopt $L_{2}$ loss as follows: + +$$ +\mathcal {L} _ {\text {c o r r}} = \| f \left(\mathbf {K} _ {P}; \mathbf {K}, \mathbf {R} _ {g t}, \mathbf {t} _ {g t}\right) - \mathbf {K} _ {I} \| _ {2}, \tag {18} +$$ + +where $f$ is defined in Equation 4. + +Diversity loss. For the goal of expanding the discrepancy among 3D keypoints $\mathbf{K}_P$ and 2D keypoints $\mathbf{K}_I$ , we impose the diversity loss $\mathcal{L}_{div} = \mathcal{L}_{2d} + \mathcal{L}_{3d}$ , where + +$$ +\mathcal {L} _ {3 d} = \frac {1}{N _ {q}} \sum_ {j = 1} ^ {N _ {q}} \sum_ {k = 1} ^ {N _ {q}} \max \left(0, \Delta_ {3 d} - \| \mathbf {K} _ {P} ^ {j} - \mathbf {K} _ {P} ^ {k} \| _ {2}\right), \tag {19} +$$ + +$$ +\mathcal {L} _ {2 d} = \frac {1}{N _ {q}} \sum_ {j = 1} ^ {N _ {q}} \sum_ {k = 1} ^ {N _ {q}} \max \left(0, \Delta_ {2 d} - \| \mathbf {K} _ {I} ^ {j} - \mathbf {K} _ {I} ^ {k} \| _ {2}\right). \tag {20} +$$ + +Camera-posediscovery. Similar to $\mathcal{L}_{fru}$ ,we define the camerepose loss $\mathcal{L}_{cam}$ as: + +$$ +\mathcal {L} _ {c a m} = \left\| \Delta \mathbf {R} _ {g t} - \Delta \mathbf {R} \right\| _ {2} + \left\| \Delta \mathbf {t} _ {g t} - \Delta \mathbf {t} \right\| _ {2}, \tag {21} +$$ + +where $\Delta_{\mathbf{R}}$ and $\Delta_{\mathrm{t}}$ are defined in Equation 12. + +Finally, we combine these loss functions by a weighted sum to train our model: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c a m}} + \mathcal {L} _ {\text {f r u}} + \alpha_ {1} \mathcal {L} _ {\text {c l s}} + \alpha_ {2} \mathcal {L} _ {\text {c o r r}} + \alpha_ {3} \mathcal {L} _ {\text {d i v}}, \tag {22} +$$ + +where $\alpha_{1},\alpha_{2},\alpha_{3}$ are weight terms to balance these losses. + +# 4. Experiments + +In this section, we first introduce implementation details. Then, we show experimental results and visualizations on two public benchmarks. Finally, we conduct a series of ablation studies to verify the effectiveness of each component. + +# 4.1. Implementation Details + +In this work, we implement the proposed model in Pytorch [24] and adopt an NVIDIA RTX 3090 GPU for training. We adopt a 4-stage ResNet [18] with FPN [22] as the image backbone network, where the output channel dimension is 128. For the 3D backbone, we use a 4-stage KPFCNN [39] where the output channel dimension is 128. We set the channel dimension $C$ to 128, set $C_f$ to 256 and set $C_{pose}$ to 512. All the attention layers have 128 features channels with 4 attention heads and ReLU activation. The number of correspondence queries $N_q$ is set to 128. And the iteration time $L$ is set to 3. The weight term $\alpha_1$ , $\alpha_2$ and $\alpha_3$ in the loss function are set to 1, 0.1 and 0.1. The margin $\Delta_{3d}$ and $\Delta_{2d}$ in the diversity loss are set to 1m and 5 pixels, respectively. All parameters in our proposed model are randomly initialized and trained from scratch with the Adam optimizer. We first train the geometric prior-guided overlapping region detection module with classification loss and frustum-pose loss for 20 epochs. After that, we train the whole network with the total loss $\mathcal{L}_{total}$ for 20 epochs. The learning rate is set to $10^{-4}$ , and the weight decay is $10^{-6}$ . + +# 4.2. Datasets and Evaluation Metrics + +We evaluate our proposed method on two commonly used benchmarks KITTI [17] and nuScenes [3]. + +KITTI Odometry [17]. KITTI Odometry dataset consists of 22 sequences of images and point cloud data collected simultaneously in urban environments, and 11 sequences provide ground-truth calibration files. We follow the common practice [21] of utilizing the 0-8 sequences for training, and 9-10 for testing. The mis-registration transformation is conducted with a 2D translation on the ground within $\pm 10\mathrm{m}$ and a rotation around the up-axis with no limited range. We downsample the image resolution to $160\times 512$ and the point cloud size to 40960 for training and testing. + +nuScenes [3]. nuScenes dataset collects the image-point cloud pairs by official SDK where the point cloud is accumulated from the nearby frames and the image is from the current frame. We follow the official data spilt of nuScenes, + +where 850 scenes are used for training, and 150 scenes for testing. The mis-registration transformation is conducted in a similar way as the one in KITTI dataset. We downsample the image resolution to $160 \times 320$ and the point cloud size to 40960 for training and testing. + +Evaluation metrics. We follow VP2P-match [44] to quantify the performance of image-to-point cloud registration with average Relative Translational Error (RTE), average Relative Rotation Error (RRE) and registration accuracy (Acc). Same as VP2P-match [44], we do not follow CorrI2P [28] to remove the image-point cloud samples with large errors before averaging, which might be an unsuitable way to report real performances. Also, We further report the registration accuracy (Acc) which is the proportion of fine registrations with RTE $< 2\mathrm{m}$ and RRE $< 5^{\circ}$ . + +# 4.3. Registration Accuracy + +Baselines. We compare our method with previous state-of-the-art methods [18, 28, 44]: + +- Grid Cls. + PnP. DeepI2P [18] proposes to divide the image into $32 \times 32$ grids and uses a neural network to classify each 3D point to a unique 2D grid. The EPnP [20] with RANSAC [16] is applied to estimate camera pose. +- Frus. Cls. + Inv.Proj. DeepI2P [18] proposes to perform frustum classification and inverse camera projection to obtain camera pose. We report their results with 2D/3D inverse camera projection as DeepI2P(2D)/DeepI2P(3D). +- Corr12P. Corr12P [28] proposes to learn dense correspondences for image-point cloud pairs, and the EPnP [20] with RANSAC [16] is applied to predict the camera pose. +- VP2P-match. VP2P-match [44] is the latest work on image-to-point cloud registration which represents the 3D elements as the combination of voxels and points and trains with a differentiable PnP solver [4]. + +Quantitative Comparison. We report the quantitative results in Table 1. Our proposed method outperforms all compared methods in all metrics by a large margin on both KITTI and nuScenes datasets. In detail, on the KITTI dataset, our method achieves the highest Acc, with a reduction of $0.55\mathrm{m}$ in RTE and about $2.05^{\circ}$ in RRE compared with VP2P-match [44]. On the nuScenes dataset, Acc is $2.61\%$ higher, RTE is reduced by $0.26\mathrm{m}$ and RRE is reduced by $0.02^{\circ}$ . We attribute the superior performance to three delicate designs. Our proposed GPDM leverages geometric prior that the overlapping region is a frustum to enhance the accuracy of point cloud overlapping region detection. And the ICLM can implicitly obtain accurate 2D-3D correspondences, avoiding the unreliable process of forced alignment between 2D and 3D features. Moreover, the PRM generates precise camera pose prediction via a learned network, which enables the end-to-end optimization. + +Qualitative Comparison. The qualitative comparison is shown in Figure 3. For intuitive visualization, we project + +Table 1. Registration accuracy on the KITTI and nuScenes datasets. Lower is better for RTE and RRE, higher is better for Acc. + +
MethodsKITTInuScenes
RTE(m)↓RRE(°)↓Acc(%)↑RTE(m)↓RRE(°)↓Acc(%)↑
Grid Cls. + PnP [18]3.64±3.4619.19±28.9611.223.02±2.4012.66±21.012.45
DeepI2P(3D) [18]4.06±3.5424.73±31.693.772.88±2.1220.65±12.242.26
DeepI2P(2D) [18]3.59±3.2111.66±18.1625.952.78±1.994.80±6.2138.10
CorrI2P [28]3.78±65.165.89±20.3472.423.04±60.763.73±9.0349.00
VP2P-match [44]0.75±1.133.29±7.9983.040.89±1.442.15±7.0388.33
Ours0.20±0.211.24±2.3497.490.63±0.442.13±3.7590.94
+ +![](images/4c32838ae83dc69983495df10098b5cb1fee3216ad31ae01489bb60d2cbff8f1.jpg) +Input +$7.14\mathrm{m} / 108.94^{\circ}$ + +![](images/ddde2f3207ff08ee012a297f7dc847203248d32495f1284ab8668dde24c3c96f.jpg) +VP2P-match +2.47m / 12.93° + +![](images/8b5d2446155a657659ea2c6add92a0f94bd2dcbe3a8e3a66c2a2c5d221f6e163.jpg) +Ours +$0.22\mathrm{m} / 0.14^{\circ}$ + +![](images/1edd3d3cec4e581cdaa21bc07c16615c9a3560cf2c54fb7b6a043d9bef22803b.jpg) +GT +RTE(m) / RRE(°) + +![](images/129165332e909eab0ddf6404f6907bb35f2872bc3a40d9b8d9990b04a4035177.jpg) +9.91m/10.53° + +![](images/e9dc57b46863cd9bd72311c176d7b0a6ab3497c6608d05fe929a0fd256882836.jpg) +$0.88\mathrm{m} / 2.07^{\circ}$ + +![](images/b9644e15ce6e79335738551e787fbd60d76b61f3e1ff7fa6ab2955b31ef43a5d.jpg) +$0.07\mathrm{m} / 1.40^{\circ}$ + +![](images/35b0405b01f0cd52267bcfe922d797b5e6ee37fc1e6d8c0dfc5f784fa7e1d481.jpg) +RTE(m) / RRE(°) + +![](images/cc9985eb0d079f86a4b3157cf5d66f9b3bb2282462fe49d65e70f5e206026e50.jpg) +$12.93\mathrm{m} / 70.03^{\circ}$ +Figure 3. Qualitative comparison of Image-to-Point Cloud registration results under KITTI dataset. + +![](images/2aed46aadf950e5f649332d34195bac1d7740559aa08e60390e2e14ef927a43d.jpg) +$5.38\mathrm{m} / 18.78^{\circ}$ + +![](images/ad9ecf6858b08f076dfedac8076c54084ddb99d716df41941993a8c921d9aa57.jpg) +$0.07\mathrm{m} / 1.31^{\circ}$ + +![](images/fc28450e61e06b18fa7901c62edf7fcf4f9a528d4b63b92c9e6d78f936b6aba8.jpg) +RTE(m) / RRE(°) + +![](images/f62219bb44d2f9df158644ab57be868e887c2168cab0a7c31a36cfcac0d991f1.jpg) +$0.16\mathrm{m} / 0.82^{\circ}$ + +![](images/401cef3422d067ebfe459df501f8529ee02079443619d6ccf95c5bf7678c1b52.jpg) +$0.04\mathrm{m} / 0.41^{\circ}$ + +![](images/263ce10b412bbc95b3e08be192c60f5e5bc6d0b0702bf58c29fbdc792b9c3079.jpg) +$0.08\mathrm{m} / 0.70^{\circ}$ +Figure 4. Visual illustration of 2D-3D correspondences and registration accuracy. Green lines represent correct correspondences. + +![](images/25133ddde07dace7da2da6ebaa0cddba96b87f730fcc016c001511c8666d4ba4.jpg) +$0.14\mathrm{m} / 0.59^{\circ}$ + +Table 2. The effect of each design in our framework. + +
RTE(m)↓RRE(°)↓Acc(%)↑
w/o GPDM0.45±1.052.48±9.6891.89
w/o ICLM1.05±0.733.70±6.2072.64
w/o PRM1.05±2.934.96±18.4782.00
Full0.20±0.241.24±2.4097.49
+ +the point cloud into image space through predicted transformations of different approaches and the known camera intrinsic matrix, and the color indicates the depth. Compared with VP2P-match [44], our method achieves better registration accuracy in different road situations. For example, on the difficult situation (e.g. the 1st and 3rd row), VP2P-match [44] fails to solve correct registration, while + +![](images/de9620719ba41cc6e915ff34cd8ee7f76a7a7d27cd3b3fa528ce2debad74f08e.jpg) +w/o Geometric Prior +0.86/0.76 + +![](images/fb2c2f4be8e1076ce856199438f03d180cbfe2f3a17a5c9dbfcef954c2ad3cea.jpg) +w Geometric Prior +0.96/1.00 + +![](images/7a7f41ad61a95c82f9805c5b46e130f1c001963b1366a2e0f9db7d0a779ae45c.jpg) +GT +Precise/Recall + +![](images/f0120884bd5e609efe14e641d640c6ed32f8ae1590ad8ca742ad80c67eeced91.jpg) +0.79/0.76 +Figure 5. Visual illustration of overlapping region detection. The green points are correctly detected in the overlapping region, the blue points are missed in the overlapping region, and the red points are incorrectly detected outside the overlapping region. + +![](images/70041bb9f9ff4cd6173845d2d21c2777acb411a7cfc2c95a40446832f3ba9d02.jpg) +0.99 / 0.95 + +![](images/6d42dca266a6e64b33667a5d40138c08a94ee13ea1fc45dd16b836232b495c72.jpg) +Precise/Recall + +our method achieves much higher registration accuracy. We further provide the 2D-3D correspondence visualization in Figure 4. Almost all of the pixels of our method can reach slight errors within 15 pixels, which demonstrates that our ICLM can obtain accurate 2D-3D correspondences. + +# 4.4. Ablation Study. + +In this section, we conduct ablation studies to justify the effectiveness of each design and the effect of some important parameters in our method on the KITTI dataset. + +To justify the effectiveness of each design in our proposed method, we develop three different variations for + +Table 3. Comparison of frustum classification results on the KITTI and nuScenes datasets. + +
SettingsKITTInuScenes
PrecisionRecallAccuracyF1-scorePrecisionRecallAccuracyF1-score
w/o GPDM93.4595.9097.9394.6682.8685.5993.1484.20
w GPDM95.1594.9198.0495.0392.0491.5396.3691.79
+ +comparison in Table 2: (1) $w / o$ GPDM removes the frustum pose prediction branch and regards points with infrustum probability $\mathbf{O}_P$ greater than 0.5 as the overlapping region. (2) $w / o$ ICLM removes both the ICLM and PRM since we can't make the PRM work without the ICLM. In this case, we take the predicted frustum pose as our estimated camera pose. (3) $w / o$ PRM replaces the PRM with the PnP-RANSAC algorithm [16, 20]. + +Effects of the GPDM. As shown in Table 2, with GPDM, the performance on the KITTI dataset is improved on all metrics. In specific, Acc is $6\%$ higher, RTE is reduced by about $56\%$ and RRE is reduced by about $50\%$ . The main reason is that our GPDM can accurately detect the overlapping region and generate a coarse camera pose which can be refined later. And to prove that our GPDM can improve the accuracy of point cloud overlapping region detection, we report the quantitative results in Table 3. With GPDM, we get better frustum classification results almost on all metrics, especially on the nuScenes dataset. The visual comparison is shown in Figure 5, which demonstrates our propose GPDM can get better prediction on the frustum boundary leveraging our proposed geometric prior. + +Effects of the ICLM. As shown in Table 2, with proposed ICLM, the performance on the KITTI dataset is improved notably. In specific, Acc is $34\%$ higher, RTE is reduced by about $81\%$ and RRE is reduced by about $66\%$ . Regarding the gain on Acc, we can figure out that the ICLM plays the most important role in our model, which demonstrates that our designed implicit correspondence learning is very effective for Image-to-Point Cloud registration. + +Effects of the PRM. As shown in Table 2, the PRM outperforms the PnP-RANSAC algorithm [16, 20] on all metrics. In specific, Acc is $19\%$ higher, RTE is reduced by about $81\%$ and RRE is reduced by about $75\%$ . We owe the advantage to the enabled end-to-end optimization with PRM. + +Effects of the number of correspondence queries. In Table 4, we show the impact of different numbers of correspondence queries in the ICLM on the KITTI dataset. The results show that our method can achieve better performance than VP2P-match [44] by using only 16 correspondence queries. We find that the overall performance of the model improves slightly with the increase of $N_{q}$ . We choose $N_{q} = 128$ in our method for the highest Acc. + +Effects of the iteration time. To investigate the influences of the iteration time $(L)$ in the ICLM, we pick $L$ from the set $\{0, 1, 2, 3, 4\}$ . As shown in Table 5, we notice that the performance drops a lot if $L$ is set to 0, showing the + +Table 4. Ablation studies on the number of queries. + +
NqRTE(m)↓RRE(°)↓Acc(%)↑
160.26±0.291.71±2.4095.52
320.23±0.251.43±2.2495.70
640.21±0.191.28±3.2096.63
1280.20±0.241.24±2.4097.49
2560.21±0.211.21±2.4097.42
+ +Table 5. Ablation studies on the iteration time. + +
LRTE(m)↓RRE(°)↓Acc(%)↑
00.31±0.362.34±2.7690.01
10.21±0.161.23±1.9597.44
20.20±0.161.28±2.0297.33
30.20±0.241.24±2.4097.49
40.21±0.301.26±2.4797.46
+ +Table 6. Ablation studies on the iteration order. + +
RTE(m)↓RRE(°)↓Acc(%)↑
Pixel-then-Point0.20±0.241.24±2.4097.49
Point-then-Pixel0.23±0.601.29±2.2697.05
+ +necessity of the iteration operation. We choose the model with $L = 1$ for the optimal balance between computational efficiency and performance. + +Effects of different iteration orders. To investigate the influences of the iteration order in the ICLM, we conduct two experiments with first Attention-Pixel then Attention-Point or first Attention-Point then Attention-Pixel. As shown is Table 6, we get similar performance with different iteration orders. And we choose to iteratively apply Attention-Pixel then Attention-Point in our method. + +# 5. Conclusion + +In this paper, we propose a novel implicit correspondence learning method with three specific designs for image-top-point cloud registration. Specially, the proposed geometric prior can guide the model to achieve high accuracy on the overlapping region detection. The proposed implicit correspondence learning module can implicitly acquire accurate 2D-3D correspondences. Furthermore, we propose a pose regression module to handle incorrect 2D-3D correspondences and generates precise camera pose. Extensive experimental results on two challenging benchmarks demonstrate the effectiveness of the proposed method. + +# 6. Acknowledgements + +This work was supported by the National Defense Basic Scientific Research program (No.JCKY2022911B002). + +# References + +[1] Xuyang Bai, Zixin Luo, Lei Zhou, Hongbo Fu, Long Quan, and Chiew-Lan Tai. D3feat: Joint learning of dense detection and description of 3d local features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6359-6367, 2020. 3 +[2] Mark Billinghurst, Adrian Clark, Gun Lee, et al. A survey of augmented reality. Foundations and Trends® in Human-Computer Interaction, 8(2-3):73-272, 2015. 1 +[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6 +[4] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2781-2790, 2022. 3, 6 +[5] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 539-546. IEEE, 2005. 2 +[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 3 +[7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8958-8966, 2019. 3 +[8] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In Proceedings of the European conference on computer vision (ECCV), pages 602-618, 2018. +[9] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 195-205, 2018. 3 +[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 2 +[11] Zhen Dong, Fuxun Liang, Bisheng Yang, Yusheng Xu, Yufu Zang, Jianping Li, Yuan Wang, Wenxia Dai, Hongchao Fan, Juha Hyyppä, et al. Registration of large-scale terrestrial laser scanner point clouds: A review and benchmark. IS-PRS Journal of Photogrammetry and Remote Sensing, 163: 327–342, 2020. 1 +[12] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d + +object recognition. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 998-1005. IEEE, 2010. 3 +[13] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99–110, 2006. 1 +[14] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 8092-8101, 2019. 2 +[15] Mengdan Feng, Sixing Hu, Marcelo H Ang, and Gim Hee Lee. 2d3d-matchnet: Learning to match keypoints across 2d image and 3d point cloud. In 2019 International Conference on Robotics and Automation (ICRA), pages 4790-4796. IEEE, 2019. 1, 2, 3 +[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 3, 6, 8 +[17] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 6 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 6, 7 +[19] Yurim Jeon and Seung-Woo Seo. Efghnet: A versatile image-to-point cloud registration network for extreme outdoor environment. IEEE Robotics and Automation Letters, 7 (3):7511-7517, 2022. 1 +[20] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate $O(n)$ solution to the pnp problem. International journal of computer vision, 81:155-166, 2009. 2, 3, 6, 8 +[21] Jiaxin Li and Gim Hee Lee. Deepi2p: Image-to-point cloud registration via deep classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15960-15969, 2021. 1, 3, 6 +[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3, 6 +[23] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 2, 3 +[24] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification + +and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3 +[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3 +[27] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143-11152, 2022. 3 +[28] Siyu Ren, Yiming Zeng, Junhui Hou, and Xiaodong Chen. Corri2p: Deep image-to-point cloud registration via dense correspondence. IEEE Transactions on Circuits and Systems for Video Technology, 33(3):1198-1208, 2022. 1, 2, 3, 6, 7 +[29] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in neural information processing systems, 32, 2019. 2 +[30] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. 2 +[31] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE international conference on robotics and automation, pages 3212-3217. IEEE, 2009. 3 +[32] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2 +[33] Paul-Edouard Sarlin, Daniel DeTone, Tsun-Yi Yang, Armen Avetisyan, Julian Straub, Tomasz Malisiewicz, Samuel Rota Bulo, Richard Newcombe, Peter Kontschieder, and Vasileios Balntas. Orienternet: Visual localization in 2d public maps with neural matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21632–21642, 2023. 1 +[34] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 2 +[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. 3rd International Conference on Learning Representations, 2015. 3 +[36] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 3 +[37] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6398-6407, 2020. 2 + +[38] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European conference on computer vision, pages 685–702. Springer, 2020. 3 +[39] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 3, 6 +[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 5 +[41] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loptr: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 3 +[42] Hao Yu, Fu Li, Mahdi Saleh, Benjamin Busam, and Slobodan Ilic. Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration. Advances in Neural Information Processing Systems, 34:23872-23884, 2021. 3 +[43] Yu Zhong. Intrinsic shape signatures: A shape descriptor for 3d object recognition. In 2009 IEEE 12th international conference on computer vision workshops, ICCV Workshops, pages 689-696. IEEE, 2009. 3 +[44] Junsheng Zhou, Baorui Ma, Wenyuan Zhang, Yi Fang, Yu-Shen Liu, and Zhizhong Han. Differentiable registration of images and lidar point clouds with voxelpoint-to-pixel matching. In Advances in Neural Information Processing Systems, pages 51166-51177. Curran Associates, Inc., 2023. 1, 2, 3, 6, 7, 8 +[45] Qunjie Zhou, Torsten Sattler, and Laura Leal-Taixe. Patch2pix: Epipolar-guided pixel-level correspondences. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4669-4678, 2021. 3 +[46] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745-5753, 2019. 3 \ No newline at end of file diff --git a/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/images.zip b/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ea6327cef18d760eb268cd602d01c97c478cf20b --- /dev/null +++ b/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad9f596a0ff57af8a1b7821282f544a1586b17564495d0223f45c6d66fccfa45 +size 729710 diff --git a/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/layout.json b/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eb7694aca0c1e17beba07d560ab49b9e8e9e5929 --- /dev/null +++ b/2025/Implicit Correspondence Learning for Image-to-Point Cloud Registration/layout.json @@ -0,0 +1,12546 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 532, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 532, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 532, + 121 + ], + "type": "text", + "content": "Implicit Correspondence Learning for Image-to-Point Cloud Registration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "spans": [ + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "text", + "content": "Xinjun Li" + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "text", + "content": ", Wenfei Yang" + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "text", + "content": ", Jiacheng Deng" + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "text", + "content": ", Zhixin Cheng" + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "text", + "content": ", Xu Zhou" + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "text", + "content": ", Tianzhu Zhang" + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 546, + 171 + ], + "type": "text", + "content": "University of Science and Technology of China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 171, + 546, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 171, + 546, + 200 + ], + "spans": [ + { + "bbox": [ + 108, + 171, + 546, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 171, + 546, + 200 + ], + "type": "text", + "content": "National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory \n" + }, + { + "bbox": [ + 108, + 171, + 546, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 108, + 171, + 546, + 200 + ], + "type": "text", + "content": "Sangfor Technologies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 201, + 588, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 201, + 588, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 201, + 588, + 213 + ], + "type": "text", + "content": "{1xj3017,dengjc,chengzhixin}@mail.ustc.edu.cn, {yangwf,tzzhang}@ustc.edu.cn, zhouxu@sangfor.com.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 266, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 555 + ], + "type": "text", + "content": "Image-to-point cloud registration aims to estimate the camera pose of a given image within a 3D scene point cloud. In this area, matching-based methods have achieved leading performance by first detecting the overlapping region, then matching point and pixel features learned by neural networks and finally using the PnP-RANSAC algorithm to estimate camera pose. However, achieving accurate image-to-point cloud registration remains challenging because the overlapping region detection is unreliable merely relying on point-wise classification, direct alignment of cross-modal data is difficult and indirect optimization objective leads to unstable registration results. To address these challenges, we propose a novel implicit correspondence learning method, including a Geometric Prior-guided overlapping region Detection Module (GPDM), an Implicit Correspondence Learning Module (ICLM), and a Pose Regression Module (PRM). The proposed method enjoys several merits. First, the proposed GPDM can precisely detect the overlapping region. Second, the ICLM can generate robust cross-modality correspondences. Third, the PRM can enable end-to-end optimization. Extensive experimental results on KITTI and nuScenes datasets demonstrate that the proposed model sets a new state-of-the-art performance in registration accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 597, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 694 + ], + "type": "text", + "content": "Image-to-point cloud registration [15] is a fundamental task in the field of computer vision, aimed at estimating the camera pose of a given image within a 3D scene point cloud. It plays an important role in many computer vision tasks, including simultaneous localization and mapping [13], 3D reconstruction [11], augmented/virtual reality [2], and visual localization [33]. Although significant progress has been made in recent years, image-to-point cloud registra" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 337, + 241, + 531, + 340 + ], + "blocks": [ + { + "bbox": [ + 337, + 241, + 531, + 340 + ], + "lines": [ + { + "bbox": [ + 337, + 241, + 531, + 340 + ], + "spans": [ + { + "bbox": [ + 337, + 241, + 531, + 340 + ], + "type": "image", + "image_path": "d9729673f7150b13eec4d61f936a5ebc7c15e910f9c8f06b52af23f9e84d7467.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 340, + 485, + 348 + ], + "lines": [ + { + "bbox": [ + 380, + 340, + 485, + 348 + ], + "spans": [ + { + "bbox": [ + 380, + 340, + 485, + 348 + ], + "type": "text", + "content": "a) Existing matching-based methods" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 337, + 348, + 530, + 449 + ], + "blocks": [ + { + "bbox": [ + 337, + 348, + 530, + 449 + ], + "lines": [ + { + "bbox": [ + 337, + 348, + 530, + 449 + ], + "spans": [ + { + "bbox": [ + 337, + 348, + 530, + 449 + ], + "type": "image", + "image_path": "a56b9b234a91892c81682ce350c78aaccd2b2d3a6c66de1609a8a6a585336b97.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 449, + 506, + 459 + ], + "lines": [ + { + "bbox": [ + 360, + 449, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 360, + 449, + 506, + 459 + ], + "type": "text", + "content": "b) Our Implicit Correspondence Learning method" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 460, + 555, + 571 + ], + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 571 + ], + "type": "text", + "content": "Figure 1. (a) The pipeline of mainstream matching-based methods. The green lines denote correct correspondences, and the red lines denote incorrect correspondences. 2D-3D correspondences are first established by calculating feature similarity. Then the PnP-RANSAC algorithm is applied to obtain camera pose. (b) The pipeline of our proposed method. A set of shared queries are used to aggregate information from image and point cloud to generate 2D and 3D keypoint detectors. With paired 2D and 3D keypoint detectors, we obtain a set of 2D-3D correspondences and estimate camera pose through a pose regression network." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 616 + ], + "type": "text", + "content": "tion remains challenging due to the inherent disparities that images capture appearances with 2D grid data while point clouds represent structures in an unordered format." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "To address these challenges, numerous methods have been proposed for image-to-point cloud registration [15, 19, 21, 28, 44]. Generally, these methods can be categorized into two main classes, including matching-free methods [19, 21] and matching-based methods [15, 28, 44]. Matching-free methods utilize the camera imaging principle as a prior to predict camera pose. For example, DeepI2P [21] casts the registration problem into a classi-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16922" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": "fication problem. A binary classification network is designed to determine whether each scene point is within or beyond the camera frustum. Then the classification results are fed into an inverse camera projection solver to obtain camera pose. However, these methods merely determine a coarse image-point correspondence, e.g. whether a 3D point is inside the image frustum, which is insufficient for precise camera pose estimation. Different from matching-free methods, matching-based methods focus on establishing fine-grained 2D-3D correspondences and use the PnP-RANSAC algorithm [16, 20] to estimate camera pose [15, 28, 44], which have achieved superior performance. These methods generally involve three important steps: (1) A detector is designed to identify the overlapping region between the image and point cloud. (2) 2D-3D correspondences are established within the identified overlapping region. (3) The PnP-RANSAC algorithm [16, 20] is applied as a post-processing module to obtain camera pose." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 291, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 662 + ], + "type": "text", + "content": "Despite the success of these matching-based methods, there still exist three notable problems that hinder the registration accuracy of these methods. (1) Unreliable point-based detection of the overlapping region. Most existing methods adopt a point-wise classification way to determine whether each point lies in the overlapping region or not, ignoring the global projection relationship between the image and point cloud, e.g. an image should correspond to a frustum in the point cloud. Consequently, existing methods usually suffer from inaccurate boundary detection of the overlapping region, resulting in some missing 2D-3D correspondences or incorrect correspondences outside the overlapping region. (2) Inaccurate explicit matching between the image and point cloud. As shown in Figure 1 (a), matching-based methods essentially establish 2D-3D correspondences by directly calculating feature similarity between the image and point cloud. These methods employ metric learning techniques like contrastive loss [5], triplet loss [34] or circle loss [37] to force the alignment of corresponding 2D and 3D features. However, due to the inherent data discrepancy between the image and point cloud, direct alignment of cross-modal data can result in a significant number of mismatches, as shown in Figure 1 (a). (3) Indirect correspondence-based optimization objective. After establishing 2D-3D correspondences, the majority of existing approaches employ the PnP-RANSAC algorithm [16, 20] as a post-processing module to estimate camera pose as illustrated in Figure 1 (a). During the training stage, these approaches primarily consider 2D-3D correspondences as the optimization target, without optimizing the pose directly." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Motivated by the above discussions, we propose a novel implicit correspondence learning method for imaged-to-point cloud registration. The proposed model contains three modules, a Geometric Prior-guided overlapping re" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 406 + ], + "type": "text", + "content": "GPI Detection Module (GPDM), an Implicit Correspondence Learning Module (ICLM), and a Pose Regression Module (PRM). The GPDM leverages geometric prior that the overlapping region is a frustum to enhance the accuracy of the overlapping region detection. Specifically, a binary classification network is designed to determine whether each point lies within or beyond the camera frustum. Then, we predict center coordinate and orientation of the camera frustum using obtained classification results. Finally, we regard points in predicted camera frustum as our overlapping region. Since camera frustum defines a continuous and regular geometric region, the noise's interference on the overlapping region detection can be mitigated. The ICLM is designed to implicitly identify a set of 2D-3D correspondences within the overlapping region. As shown in Figure 1 (b), we initialize a set of learnable correspondence queries. For each query, we convert it to 2D/3D keypoint detector via the attention mechanism [40] to aggregate features from the image and point cloud, respectively. By calculating similarity between 2D/3D keypoint detector and image/point cloud features, a pair of 2D/3D keypoint heatmap is obtained. In this way, we can implicitly obtain a set of 2D-3D paired keypoints, avoiding the process of forced alignment between 2D and 3D features. Finally, the PRM takes obtained 2D/3D keypoint detectors and 2D/3D keypoints as inputs, and regresses camera pose via a Multi-Layer Perceptron (MLP), which enables the end-to-end optimization. The main contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 407, + 553, + 515 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 314, + 407, + 553, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 553, + 454 + ], + "type": "text", + "content": "- We propose a novel implicit correspondence learning method for image-to-point cloud registration, which enjoys the merits of precise overlapping region detection, robust cross-modality matching, end-to-end optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 456, + 553, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 456, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 314, + 456, + 553, + 515 + ], + "type": "text", + "content": "- Without bells and whistles, extensive experimental results on KITTI and nuScenes datasets demonstrate that the proposed model sets a new state-of-the-art performance in registration accuracy, i.e. a reduction of " + }, + { + "bbox": [ + 314, + 456, + 553, + 515 + ], + "type": "inline_equation", + "content": "0.55\\mathrm{m}" + }, + { + "bbox": [ + 314, + 456, + 553, + 515 + ], + "type": "text", + "content": " in RTE and " + }, + { + "bbox": [ + 314, + 456, + 553, + 515 + ], + "type": "inline_equation", + "content": "2.05^{\\circ}" + }, + { + "bbox": [ + 314, + 456, + 553, + 515 + ], + "type": "text", + "content": " in RRE compared to previous SOTA method." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 525, + 406, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 406, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 406, + 538 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 546, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 582 + ], + "type": "text", + "content": "In this section, we briefly overview methods that are related to image-to-image registration, point cloud-to-point cloud registration and image-to-point cloud registration." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "Image-to-image registration. Image-to-image registration is the process of aligning two or more images, which aims to find the optimal transformation that best aligns the structures of interest in the input images. Typical methods usually adopt a detect-then-match pipeline to extract correspondences. Keypoints are first detected and described using handcrafted [23, 30] or learning-based techniques [10, 14, 29, 32], correspondences are then established based on their feature similarity. However, robustly detecting repeatable keypoints across different images is still challenging, especially for texture-poor regions. Detection" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16923" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "free methods [36, 41, 45] propose to bypass the keypoint detection step by first obtaining patch-level matches and then regress pixel-wise matches in a coarse-to-fine manner, which achieves accurate image-to-image registration." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 142, + 295, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 295, + 299 + ], + "type": "text", + "content": "Point cloud-to-point cloud registration. Point cloud-to-point cloud registration is the process of finding a rigid transformation that best aligns two point clouds. Similar to image-to-image registration, early works leverage hand-crafted descriptors [12, 31] or learning-based descriptors [1, 7-9] matching to estimate the correspondences. CoFiNet [42] follows the design of LoFTR [36] and introduces detection-free registration with a coarse-to-fine strategy. Recent methods have replaced RANSAC [16] with deep robust estimators to enhance both speed and accuracy. For example, GeoTransformer [27] proposes a geometric Transformer and introduces a local-to-global registration for RANSAC-free registration." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 318, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 318, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 318, + 296, + 714 + ], + "type": "text", + "content": "Image-to-point cloud registration. Compared to intramodality registration, image-to-point cloud registration poses more significant difficulties due to the inherent data disparities. The typical pipeline is to first extract 2D-3D correspondences and then estimate camera pose using the PnP-RANSAC algorithm [16, 20]. In this case, the accuracy of extracted correspondences plays a crucial role in ensuring successful registration. 2D3D-MatchNet [15] is the pioneering work, which uses SIFT [23] and ISS [43] to extract the keypoints of the image and point cloud, respectively. Then the patches around the keypoints are fed to CNNs [35] and PointNet [25] to extract features for establishing 2D-3D correspondences. However, the independent hand-crafted keypoint detectors for different modalities leads to a poor registration accuracy. CorrI2P [28] predicts pixels and points in the overlapping region and learns dense 2D-3D correspondences by matching point and pixel features for fine registration. However, it struggles to learn a discriminative latent space for 2D and 3D data, leading to a high rate of incorrect correspondences. Recently, VP2PMatch [44] learns a structured cross-modality latent space by representing 3D elements using both voxels [6, 38] and points [26]. It imposes supervisions directly on the predicted pose distribution with a differentiable probabilistic PnP solver [4]. Though introducing voxel features, it struggles with 2D-3D data disparities. Besides, it fails to achieve high registration accuracy for insufficient supervisions. In addition to these matching-based methods, there are some matching-free methods. For instance, DeepI2P [21] converts the registration problem into a classification and inverse camera projection optimization problem. However, the frustum classification only indicates coarse 2D-3D correspondence, thus limiting the registration accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "text", + "content": "Given a 3D scene point cloud " + }, + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "text", + "content": " and an image " + }, + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "text", + "content": " from the same scene, the cross-modality image-to-point cloud registration is to estimate the camera pose " + }, + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{gt}" + }, + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "text", + "content": " in the coordinate system of " + }, + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "text", + "content": ", including a rotation matrix " + }, + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{gt} \\in \\mathbf{SO}(3)" + }, + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "text", + "content": " and a translation vector " + }, + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{gt} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 90, + 553, + 151 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 152, + 555, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 282 + ], + "type": "text", + "content": "An overview of our pipeline is illustrated in Figure 2. Our proposed model mainly consists of a Geometric Prior-guided overlapping region Detection Module (GPDM), an Implicit Correspondence Learning Module (ICLM) and a Pose Regression Module (PRM). Specifically, we first extract features for the image and point cloud respectively. Next, the GPDM detects the overlapping region of the input point cloud. Then the ICLM implicitly identifies a set of correspondences within detected overlapping region via our designed correspondence queries. At last, we estimate the camera pose via the PRM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 289, + 425, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 289, + 425, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 289, + 425, + 300 + ], + "type": "text", + "content": "3.1. Feature Extraction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "type": "text", + "content": "For the image, we use ResNet [18] with FPN [22] to extract 2D features " + }, + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_I\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "type": "text", + "content": ". For simplicity, we denote the pixel coordinate matrices for " + }, + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_I" + }, + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_I\\in \\mathbb{R}^{HW\\times 2}" + }, + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "type": "text", + "content": ". For the input point cloud, we adopt KPFCNN [39] to extract 3D features " + }, + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_P\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 372, + 553, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 553, + 396 + ], + "type": "text", + "content": "3.2. Geometric Prior-guided Overlapping Region Detection Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 401, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 496 + ], + "type": "text", + "content": "To precisely detect the overlapping region of the point cloud, most methods adopt a point-wise classification strategy to determine whether each point lies in the overlapping region or not, ignoring the global projection relationship between the image and point cloud. Inspired by the fact that an image should correspond to a frustum in the point cloud, we design a geometric prior-guided classification strategy thus we can get more accurate classification results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 497, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 556 + ], + "type": "text", + "content": "As shown in the bottom-left part of Figure 2, we first obtain 2D global feature by averaging extracted 2D features " + }, + { + "bbox": [ + 313, + 497, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_I" + }, + { + "bbox": [ + 313, + 497, + 554, + 556 + ], + "type": "text", + "content": " and repeat it to concatenate with the 3D features " + }, + { + "bbox": [ + 313, + 497, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_P" + }, + { + "bbox": [ + 313, + 497, + 554, + 556 + ], + "type": "text", + "content": ". Then a MLP is used to learn the probability " + }, + { + "bbox": [ + 313, + 497, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_P \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 313, + 497, + 554, + 556 + ], + "type": "text", + "content": " of a point to lie within the camera frustum." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "text", + "content": "Since the in-frustum probability " + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_P" + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "text", + "content": " is not precise enough for in-frustum classification, we propose to estimate the frustum pose to get more accurate classification results. Utilizing " + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_P" + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "text", + "content": " and the coordinate information of the point cloud " + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "text", + "content": ", we apply another MLP to regress the frustum pose " + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_f \\in \\mathbf{SE}(\\mathbf{3})" + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "text", + "content": ", composed of a 3D rotation " + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_f \\in \\mathbf{SO}(\\mathbf{3})" + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "text", + "content": " and a 3D translation vector " + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_f \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "text", + "content": ". In detail, we apply two parallel MLPs to regress " + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_f" + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_f" + }, + { + "bbox": [ + 313, + 556, + 554, + 653 + ], + "type": "text", + "content": " respectively:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 339, + 658, + 553, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 658, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 339, + 658, + 553, + 673 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {f}, \\mathbf {t} _ {f} = \\rho (\\mathrm {M L P} ([ \\mathbf {O} _ {P}, \\mathbf {P} ])), \\mathrm {M L P} ([ \\mathbf {O} _ {P}, \\mathbf {P} ]), \\tag {1}", + "image_path": "5eb598d2e4e3e6a18d3ddef0655e10c133fc918645fb7280a9785d74350a0609.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " denotes concatenation along feature dimension. We choose a 6D representation of rotation [46] as the regression of the first MLP, for its continuous learning space" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16924" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 526, + 228 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 526, + 228 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 526, + 228 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 526, + 228 + ], + "type": "image", + "image_path": "c38f3556c18f6641645d335518a25e8f2a76a7b534c71764e073091f9489f2aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 83, + 228, + 526, + 373 + ], + "blocks": [ + { + "bbox": [ + 83, + 228, + 526, + 373 + ], + "lines": [ + { + "bbox": [ + 83, + 228, + 526, + 373 + ], + "spans": [ + { + "bbox": [ + 83, + 228, + 526, + 373 + ], + "type": "image", + "image_path": "387a90bf422d22dd848cb00a1f75d6322f0257113f9745e18e0c54c6b0a9d066.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "text", + "content": "Figure 2. Overview of our method. Given a pair of point cloud " + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "text", + "content": " and image " + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "text", + "content": " as inputs, we first extract features for the image and point cloud respectively. Then, in the geometric prior-guided overlapping region detection module, we leverage geometric prior that the overlapping region is a frustum to classify each point in the point cloud into within or beyond the camera frustum and get the point cloud mask " + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_P" + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "text", + "content": ". Next, in the implicit correspondence learning module, we initialize a set of correspondence queries " + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q" + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "text", + "content": " and convert them to 2D keypoint detectors " + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_I" + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "text", + "content": " and 3D keypoint detectors " + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_P" + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "text", + "content": " by aggregating extracted 2D and 3D features via the attention mechanism. Weighting the pixel/point coordinates according to the feature similarity between the 2D/3D keypoint detectors and 2D/3D features, we obtain 2D keypoints " + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_I" + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "text", + "content": " and 3D keypoints " + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_P" + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "text", + "content": ", respectively. Finally, in the pose regression module, we utilize the information of 2D-3D paired keypoints and 2D-3D paired keypoint detectors to regress the camera pose " + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{R}" + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 54, + 383, + 555, + 472 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": "in " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{SO}(3)" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\rho(\\cdot)" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": " denotes the transformation from the 6D representation to the " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": " rotation matrix " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_f" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": ". Then according to the given camera intrinsic matrix " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": ", we define the point-to-pixel projection function as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 536, + 295, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 536, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 104, + 536, + 295, + 552 + ], + "type": "interline_equation", + "content": "\\left[ x _ {i}, y _ {i}, z _ {i} \\right] ^ {\\top} = \\mathbf {K} \\cdot \\left(\\mathbf {R} _ {f} \\cdot \\mathbf {p} _ {i} + \\mathbf {t} _ {f}\\right), \\tag {2}", + "image_path": "93d9084615886a8cb380a1fda987af41710546084d473ea5a0625c0c65affb8a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 103, + 553, + 294, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 553, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 103, + 553, + 294, + 569 + ], + "type": "interline_equation", + "content": "\\mathbf {u} _ {i} = \\left[ u _ {x i}, u _ {y i} \\right] ^ {\\top} = \\left[ x _ {i} / z _ {i}, y _ {i} / z _ {i} \\right] ^ {\\top}, \\tag {3}", + "image_path": "9da5dc50f5353398a189b4c9763f995ccc33bd8fbdd61a99015b56b49d7c44fe.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 570, + 295, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 570, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 127, + 570, + 295, + 583 + ], + "type": "interline_equation", + "content": "f \\left(\\mathbf {p} _ {i}; \\mathbf {K}, \\mathbf {R} _ {f}, \\mathbf {t} _ {f}\\right) = \\mathbf {u} _ {i}, \\tag {4}", + "image_path": "4ae39609d54742a1405060852f914820d0e0da7e853537de041d694b893a1c96.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 590, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 590, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i\\in \\mathbf{P}" + }, + { + "bbox": [ + 55, + 590, + 295, + 616 + ], + "type": "text", + "content": " is a 3D point and " + }, + { + "bbox": [ + 55, + 590, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i" + }, + { + "bbox": [ + 55, + 590, + 295, + 616 + ], + "type": "text", + "content": " represents a 2D pixel. We get the point cloud mask " + }, + { + "bbox": [ + 55, + 590, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p\\in \\mathbb{R}^{N\\times 1}" + }, + { + "bbox": [ + 55, + 590, + 295, + 616 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 621, + 295, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 621, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 70, + 621, + 295, + 654 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {p} ^ {i} = \\left\\{ \\begin{array}{l} 1; 0 \\leq u _ {x i} < W, 0 \\leq u _ {y i} < H, z _ {i} \\geq 0 \\\\ 0; \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {5}", + "image_path": "bb57efb7938c0418883d355a20a6ebe7e72ac5603878974c6e30ca5eaa635059.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 660, + 280, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 280, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 280, + 673 + ], + "type": "text", + "content": "3.3. Implicit Correspondence Learning Module" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "After obtaining point cloud mask " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ", we implicitly identify a set of robust correspondences between the image and point cloud. As shown in the bottom-right part of Figure 2," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "text", + "content": "we first initialize " + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "text", + "content": " correspondence queries " + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q\\in \\mathbb{R}^{N_q\\times C}" + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "text", + "content": " with a set of learnable parameters. Then, we aim to convert correspondence queries " + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q" + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "text", + "content": " to 2D and 3D keypoint detectors by aggregating information from both the 2D features " + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_I" + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "text", + "content": " and 3D features " + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_P" + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "text", + "content": ". For this purpose, we iteratively apply Attention-Pixel then Attention-Point as shown in Figure 2. For every iteration " + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "inline_equation", + "content": "k\\geq 0" + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "text", + "content": ", we take queries " + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q^{2k}" + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "text", + "content": " as the input and enable interaction with the 2D features " + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_I" + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "text", + "content": " and 3D features " + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_P" + }, + { + "bbox": [ + 313, + 482, + 555, + 592 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 603, + 553, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 603, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 358, + 603, + 553, + 620 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {q} ^ {2 k + 1} = \\operatorname {A t t e n t i o n - P i x e l} \\left(\\mathbf {F} _ {\\mathrm {q}} ^ {2 \\mathrm {k}}, \\mathbf {F} _ {\\mathrm {I}}\\right), \\tag {6}", + "image_path": "f0be40a75c8222779e26d344e239609dd447dcbaa2acc4a1547709222768a28a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 341, + 620, + 553, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 620, + 553, + 636 + ], + "spans": [ + { + "bbox": [ + 341, + 620, + 553, + 636 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {q} ^ {2 k + 2} = \\operatorname {A t t e n t i o n - P o i n t} \\left(\\mathbf {F} _ {\\mathrm {q}} ^ {2 k + 1}, \\mathbf {F} _ {\\mathrm {P}}, \\mathbf {M} _ {\\mathrm {p}}\\right), \\tag {7}", + "image_path": "6a9eb08aa69a86a9ee0dd9425e187e03d7eae635c0635113e0dd671edb7b0ab2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 648, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 553, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 648, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q^0 = \\mathbf{F}_q" + }, + { + "bbox": [ + 313, + 648, + 553, + 676 + ], + "type": "text", + "content": ". These two steps are executed alternately for " + }, + { + "bbox": [ + 313, + 648, + 553, + 676 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 648, + 553, + 676 + ], + "type": "text", + "content": " times and we get updated queries " + }, + { + "bbox": [ + 313, + 648, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q^{2L}" + }, + { + "bbox": [ + 313, + 648, + 553, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 676, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 676, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 676, + 555, + 715 + ], + "type": "text", + "content": "Attention-Pixel. For the " + }, + { + "bbox": [ + 313, + 676, + 555, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 676, + 555, + 715 + ], + "type": "text", + "content": "-th iteration, we utilize queries " + }, + { + "bbox": [ + 313, + 676, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q^{2k}" + }, + { + "bbox": [ + 313, + 676, + 555, + 715 + ], + "type": "text", + "content": " from previous layer to aggregate information from 2D features " + }, + { + "bbox": [ + 313, + 676, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_I" + }, + { + "bbox": [ + 313, + 676, + 555, + 715 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 313, + 676, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q^{2k+1}" + }, + { + "bbox": [ + 313, + 676, + 555, + 715 + ], + "type": "text", + "content": ". Specially, keys and values" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16925" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 292, + 87 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 292, + 87 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 292, + 87 + ], + "type": "text", + "content": "arise from the 2D features " + }, + { + "bbox": [ + 55, + 72, + 292, + 87 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_I" + }, + { + "bbox": [ + 55, + 72, + 292, + 87 + ], + "type": "text", + "content": ", and queries arise from " + }, + { + "bbox": [ + 55, + 72, + 292, + 87 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q^{2k}" + }, + { + "bbox": [ + 55, + 72, + 292, + 87 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 86, + 92, + 294, + 108 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 92, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 86, + 92, + 294, + 108 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathbf {F} _ {q} ^ {2 k} \\mathbf {W} ^ {\\mathcal {Q}}, \\mathbf {K} = \\mathbf {F} _ {I} \\mathbf {W} ^ {\\mathcal {K}}, \\mathbf {V} = \\mathbf {F} _ {I} \\mathbf {W} ^ {\\mathcal {V}}, \\tag {8}", + "image_path": "e30323b6b3cb28efc1552fe269327da0d27043d5920af1b17122900de981bbe6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 114, + 295, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 295, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 114, + 295, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\mathcal{Q}},\\mathbf{W}^{\\mathcal{K}},\\mathbf{W}^{\\mathcal{V}}\\in \\mathbb{R}^{C\\times d_k}" + }, + { + "bbox": [ + 55, + 114, + 295, + 152 + ], + "type": "text", + "content": " are linear projections. Then, the queries " + }, + { + "bbox": [ + 55, + 114, + 295, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q^{2k}" + }, + { + "bbox": [ + 55, + 114, + 295, + 152 + ], + "type": "text", + "content": " are updated with the multi-head attention mechanism [40]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 158, + 295, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 158, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 83, + 158, + 295, + 186 + ], + "type": "interline_equation", + "content": "\\mathrm {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\mathrm {S o f t m a x} (\\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {d _ {k}}}) \\mathbf {V}. \\qquad (9)", + "image_path": "3e54a74b129ba5343329ff4a7bba2f22a893edf5ae4bf403f71dea57dcb7af0f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 191, + 295, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 295, + 215 + ], + "type": "text", + "content": "And the attention features are further projected with a shallow MLP as the final output features " + }, + { + "bbox": [ + 55, + 191, + 295, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_a^{2k + 1}" + }, + { + "bbox": [ + 55, + 191, + 295, + 215 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 215, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 295, + 300 + ], + "type": "text", + "content": "Attention-Point. For the " + }, + { + "bbox": [ + 55, + 215, + 295, + 300 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 215, + 295, + 300 + ], + "type": "text", + "content": "-th iteration, we utilize the queries " + }, + { + "bbox": [ + 55, + 215, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q^{2k + 1}" + }, + { + "bbox": [ + 55, + 215, + 295, + 300 + ], + "type": "text", + "content": " from previous layer to aggregate information from the 3D features " + }, + { + "bbox": [ + 55, + 215, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_P" + }, + { + "bbox": [ + 55, + 215, + 295, + 300 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 55, + 215, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q^{2k + 2}" + }, + { + "bbox": [ + 55, + 215, + 295, + 300 + ], + "type": "text", + "content": ". The whole process is almost the same as described in Attention-Pixel. The only difference is that we utilize the point cloud mask " + }, + { + "bbox": [ + 55, + 215, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p" + }, + { + "bbox": [ + 55, + 215, + 295, + 300 + ], + "type": "text", + "content": " as the keys' mask to make queries focus on relevant overlapping regions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": "After " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " iterations, we get updated queries " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q^{2L}" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": ". These updated queries can adaptively change according to the content of the input image and point cloud, capturing the same patterns in both. Next, We apply Attention-Point and Attention-Pixel respectively, taking the updated queries " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_q^{2L}" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " as the input to get 3D keypoint detectors " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_P" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " and 2D keypoint detectors " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_I" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": ". Then with 2D keypoint detectors " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_I" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " and 2D features " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_I" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": ", we get 2D keypoint heatmap " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_I \\in \\mathbb{R}^{N_q \\times HW}" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 415, + 294, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 415, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 100, + 415, + 294, + 429 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {I} = \\operatorname {S o f t m a x} \\left(\\mathbf {D} _ {I} \\operatorname {F l a t t e n} \\left(\\mathbf {F} _ {I}\\right) ^ {\\top}\\right), \\tag {10}", + "image_path": "bcd366e786b7954f59031e80fbf5796a330e79383cac63b0b07355ca3e7945b5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 436, + 295, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 436, + 295, + 459 + ], + "type": "inline_equation", + "content": "\\mathrm{Flatten}(\\cdot)" + }, + { + "bbox": [ + 55, + 436, + 295, + 459 + ], + "type": "text", + "content": " denotes combination of the first two dimensions into a single dimension." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "text", + "content": "3D keypoint heatmap " + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_P\\in \\mathbb{R}^{N_q\\times N}" + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "text", + "content": " is generated in a similar way and our predicted mask " + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p" + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "text", + "content": " is applied to suppress irrelevant points." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 495, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 520 + ], + "type": "text", + "content": "At last, We can get 3D keypoints " + }, + { + "bbox": [ + 55, + 495, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_P\\in \\mathbb{R}^{N_q\\times 3}" + }, + { + "bbox": [ + 55, + 495, + 295, + 520 + ], + "type": "text", + "content": " and 2D keypoints " + }, + { + "bbox": [ + 55, + 495, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_I\\in \\mathbb{R}^{N_q\\times 2}" + }, + { + "bbox": [ + 55, + 495, + 295, + 520 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 117, + 528, + 295, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 528, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 117, + 528, + 295, + 540 + ], + "type": "interline_equation", + "content": "\\mathbf {K} _ {P} = \\mathbf {H} _ {P} \\mathbf {P}, \\mathbf {K} _ {I} = \\mathbf {H} _ {I} \\mathbf {E} _ {I}. \\tag {11}", + "image_path": "dfbe3fbab988439ee778acbfc4f4129d7f822fb86cf9a95cf6b1a13b2a182767.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 548, + 192, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 192, + 560 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 192, + 560 + ], + "type": "text", + "content": "3.4. Pose Regression Module" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "text", + "content": "After establishing 2D-3D correspondences, we wish to regress camera pose through a learned network. As we have estimated the frustum pose " + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_f" + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "text", + "content": " in the GPDM, which is a coarse camera pose estimation, we can regard " + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_f" + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "text", + "content": " as an initial camera pose to be refined and apply corresponding transformation to the 3D keypoints " + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_P" + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "text", + "content": ". In this case, we estimate the difference between " + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_f" + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_f" + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "text", + "content": " and the ground-truth camera pose " + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{gt}" + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{gt}" + }, + { + "bbox": [ + 55, + 565, + 295, + 662 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 668, + 295, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 668, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 67, + 668, + 295, + 683 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {R} _ {g t} = \\mathbf {R} _ {g t} \\cdot \\left(\\mathbf {R} _ {f}\\right) ^ {- 1}, \\Delta \\mathbf {t} _ {g t} = \\mathbf {t} _ {g t} - \\Delta \\mathbf {R} _ {g t} \\cdot \\mathbf {t} _ {f}. \\tag {12}", + "image_path": "5670eb839621c9c7f2f6a1324622674652faf9b855247f41e38af782cede745b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "As shown in the top-right part of Figure 2, we first apply a MLP to lift the correspondence features and pair the" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 554, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 97 + ], + "type": "text", + "content": "correspondence via feature concatenation to get the fusion features " + }, + { + "bbox": [ + 313, + 72, + 554, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_f\\in \\mathbb{R}^{N_q\\times C_f}" + }, + { + "bbox": [ + 313, + 72, + 554, + 97 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 104, + 553, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 104, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 326, + 104, + 553, + 118 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {f} = \\operatorname {M L P} \\left(\\left[ \\mathbf {D} _ {P}, \\operatorname {M L P} \\left(\\mathbf {K} _ {P}\\right), \\mathbf {D} _ {I}, \\operatorname {M L P} \\left(\\mathbf {K} _ {I}\\right) \\right]\\right). \\tag {13}", + "image_path": "e32ea09fc0432094a81abe1ebc3d0c504c1b3821ef044459b51080d42d1c9288.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 125, + 554, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 554, + 160 + ], + "type": "text", + "content": "We then inject global information into the point-pixel correspondences in " + }, + { + "bbox": [ + 313, + 125, + 554, + 160 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_f" + }, + { + "bbox": [ + 313, + 125, + 554, + 160 + ], + "type": "text", + "content": " by concatenating its averaged feature to strengthen the correspondence information and obtain " + }, + { + "bbox": [ + 313, + 125, + 554, + 160 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s" + }, + { + "bbox": [ + 313, + 125, + 554, + 160 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 160, + 554, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 554, + 185 + ], + "type": "text", + "content": "A pose-sensitive feature " + }, + { + "bbox": [ + 313, + 160, + 554, + 185 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\text {pose }}" + }, + { + "bbox": [ + 313, + 160, + 554, + 185 + ], + "type": "text", + "content": " is learned from all the correspondence information via an average-pooling operation:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 371, + 196, + 553, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 196, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 371, + 196, + 553, + 209 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\text {p o s e}} = \\operatorname {A v g p o o l} (\\operatorname {M L P} \\left(\\mathbf {F} _ {s}\\right)). \\tag {14}", + "image_path": "8c48a2e1e677bc236ab476054642ecdd7eb4c1137abb503d4e5e29c627cdadf1.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 213, + 554, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 554, + 237 + ], + "type": "text", + "content": "Finally, similar to Equation 1, we apply two parallel MLPs to regress " + }, + { + "bbox": [ + 313, + 213, + 554, + 237 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{R}" + }, + { + "bbox": [ + 313, + 213, + 554, + 237 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 213, + 554, + 237 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{t}" + }, + { + "bbox": [ + 313, + 213, + 554, + 237 + ], + "type": "text", + "content": " respectively:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 349, + 245, + 553, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 245, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 349, + 245, + 553, + 258 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {R}, \\Delta \\mathbf {t} = \\rho (\\operatorname {M L P} (\\mathbf {f} _ {\\text {p o s e}})), \\operatorname {M L P} (\\mathbf {f} _ {\\text {p o s e}}). \\tag {15}", + "image_path": "279cae2d033bb75510910b81535c5416501febeba0325204c3c5c85cbd9d76bc.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 265, + 404, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 404, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 404, + 277 + ], + "type": "text", + "content": "3.5. Loss Function" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "text", + "content": "Given ground-truth camera pose " + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{gt}" + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{gt}" + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "text", + "content": " and camera intrinsic matrix " + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "text", + "content": ", we introduce a joint loss function consisting of a classification loss " + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "text", + "content": ", a frustum-lose loss " + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fru}" + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "text", + "content": ", a correspondence loss " + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{corr}" + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "text", + "content": ", a diversity loss " + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{div}" + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "text", + "content": " and a camera-pose loss " + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cam}" + }, + { + "bbox": [ + 313, + 282, + 554, + 342 + ], + "type": "text", + "content": " to guide the model learning effectively." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 342, + 554, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 402 + ], + "type": "text", + "content": "Classification loss. According to Equation 4 and Equation 5, we can get the ground-truth point cloud classification labels " + }, + { + "bbox": [ + 313, + 342, + 554, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{gt}" + }, + { + "bbox": [ + 313, + 342, + 554, + 402 + ], + "type": "text", + "content": " taking the ground-truth camera pose " + }, + { + "bbox": [ + 313, + 342, + 554, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{gt}" + }, + { + "bbox": [ + 313, + 342, + 554, + 402 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 342, + 554, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{gt}" + }, + { + "bbox": [ + 313, + 342, + 554, + 402 + ], + "type": "text", + "content": " and camera intrinsic matrix " + }, + { + "bbox": [ + 313, + 342, + 554, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 342, + 554, + 402 + ], + "type": "text", + "content": " as inputs. We adopt weighted binary cross-entropy as follows:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 373, + 410, + 553, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 410, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 373, + 410, + 553, + 423 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = \\operatorname {B C E L o s s} \\left(\\mathbf {O} _ {g t}, \\mathbf {O} _ {P}\\right). \\tag {16}", + "image_path": "8de9730ea9874abe3c267e6ddb693101896c19eb40e4f3135d23045bc56a1ff5.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 430, + 554, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 455 + ], + "type": "text", + "content": "Frustum-losess. As ground-truth camera pose is also ground truth of the frustum pose,We use " + }, + { + "bbox": [ + 313, + 430, + 554, + 455 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 430, + 554, + 455 + ], + "type": "text", + "content": " loss as follows:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 355, + 466, + 553, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 466, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 355, + 466, + 553, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f r u} = \\left\\| \\mathbf {R} _ {g t} - \\mathbf {R} _ {f} \\right\\| _ {2} + \\left\\| \\mathbf {t} _ {g t} - \\mathbf {t} _ {f} \\right\\| _ {2}. \\tag {17}", + "image_path": "72458a07a3699c53ec5f51535bfb59c24130f9cae1d389d07c0d3e261a9d36e5.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 483, + 554, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 507 + ], + "type": "text", + "content": "Correspondence loss. To establish accurate 2D-3D correspondences, we adopt " + }, + { + "bbox": [ + 313, + 483, + 554, + 507 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 483, + 554, + 507 + ], + "type": "text", + "content": " loss as follows:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 353, + 515, + 553, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 515, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 353, + 515, + 553, + 529 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o r r}} = \\| f \\left(\\mathbf {K} _ {P}; \\mathbf {K}, \\mathbf {R} _ {g t}, \\mathbf {t} _ {g t}\\right) - \\mathbf {K} _ {I} \\| _ {2}, \\tag {18}", + "image_path": "030890364f7a93a99ace29e7dbcf109383377af47eebbb11dc1fd0674cc533be.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 535, + 450, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 450, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 450, + 547 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 535, + 450, + 547 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 535, + 450, + 547 + ], + "type": "text", + "content": " is defined in Equation 4." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "text", + "content": "Diversity loss. For the goal of expanding the discrepancy among 3D keypoints " + }, + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_P" + }, + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "text", + "content": " and 2D keypoints " + }, + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_I" + }, + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "text", + "content": ", we impose the diversity loss " + }, + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{div} = \\mathcal{L}_{2d} + \\mathcal{L}_{3d}" + }, + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 322, + 590, + 553, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 590, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 322, + 590, + 553, + 625 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {3 d} = \\frac {1}{N _ {q}} \\sum_ {j = 1} ^ {N _ {q}} \\sum_ {k = 1} ^ {N _ {q}} \\max \\left(0, \\Delta_ {3 d} - \\| \\mathbf {K} _ {P} ^ {j} - \\mathbf {K} _ {P} ^ {k} \\| _ {2}\\right), \\tag {19}", + "image_path": "44910f95f945d2e609fd39dcafe25da75cb55f116b49d3a602225ec30c73466b.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 324, + 628, + 553, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 628, + 553, + 663 + ], + "spans": [ + { + "bbox": [ + 324, + 628, + 553, + 663 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2 d} = \\frac {1}{N _ {q}} \\sum_ {j = 1} ^ {N _ {q}} \\sum_ {k = 1} ^ {N _ {q}} \\max \\left(0, \\Delta_ {2 d} - \\| \\mathbf {K} _ {I} ^ {j} - \\mathbf {K} _ {I} ^ {k} \\| _ {2}\\right). \\tag {20}", + "image_path": "3201c7338f67c1de3885789218534a6c29ecbe774f657223a24fa9cb6f63a396.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 669, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 554, + 693 + ], + "type": "text", + "content": "Camera-posediscovery. Similar to " + }, + { + "bbox": [ + 313, + 669, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fru}" + }, + { + "bbox": [ + 313, + 669, + 554, + 693 + ], + "type": "text", + "content": " ,we define the camerepose loss " + }, + { + "bbox": [ + 313, + 669, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cam}" + }, + { + "bbox": [ + 313, + 669, + 554, + 693 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 336, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 336, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c a m} = \\left\\| \\Delta \\mathbf {R} _ {g t} - \\Delta \\mathbf {R} \\right\\| _ {2} + \\left\\| \\Delta \\mathbf {t} _ {g t} - \\Delta \\mathbf {t} \\right\\| _ {2}, \\tag {21}", + "image_path": "809c28ae3c8814898d42e0780485ce0d725e16afe5e05eca89a316af64312e51.jpg" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16926" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 240, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 240, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 240, + 83 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 240, + 83 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathbf{R}}" + }, + { + "bbox": [ + 55, + 72, + 240, + 83 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 240, + 83 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 72, + 240, + 83 + ], + "type": "text", + "content": " are defined in Equation 12." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 295, + 108 + ], + "type": "text", + "content": "Finally, we combine these loss functions by a weighted sum to train our model:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 118, + 295, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 295, + 133 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c a m}} + \\mathcal {L} _ {\\text {f r u}} + \\alpha_ {1} \\mathcal {L} _ {\\text {c l s}} + \\alpha_ {2} \\mathcal {L} _ {\\text {c o r r}} + \\alpha_ {3} \\mathcal {L} _ {\\text {d i v}}, \\tag {22}", + "image_path": "b942c909f673ef2d899d7d08734785f626b5ffc4cba1205575e654cddd70c839.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 141, + 293, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 141, + 293, + 154 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 293, + 154 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 141, + 293, + 154 + ], + "type": "inline_equation", + "content": "\\alpha_{1},\\alpha_{2},\\alpha_{3}" + }, + { + "bbox": [ + 55, + 141, + 293, + 154 + ], + "type": "text", + "content": " are weight terms to balance these losses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 164, + 137, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 137, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 137, + 178 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 184, + 295, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 184, + 295, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 295, + 233 + ], + "type": "text", + "content": "In this section, we first introduce implementation details. Then, we show experimental results and visualizations on two public benchmarks. Finally, we conduct a series of ablation studies to verify the effectiveness of each component." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 188, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 188, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 188, + 252 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": "In this work, we implement the proposed model in Pytorch [24] and adopt an NVIDIA RTX 3090 GPU for training. We adopt a 4-stage ResNet [18] with FPN [22] as the image backbone network, where the output channel dimension is 128. For the 3D backbone, we use a 4-stage KPFCNN [39] where the output channel dimension is 128. We set the channel dimension " + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": " to 128, set " + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "inline_equation", + "content": "C_f" + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": " to 256 and set " + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "inline_equation", + "content": "C_{pose}" + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": " to 512. All the attention layers have 128 features channels with 4 attention heads and ReLU activation. The number of correspondence queries " + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "inline_equation", + "content": "N_q" + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": " is set to 128. And the iteration time " + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": " is set to 3. The weight term " + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\alpha_1" + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\alpha_2" + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\alpha_3" + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": " in the loss function are set to 1, 0.1 and 0.1. The margin " + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\Delta_{3d}" + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\Delta_{2d}" + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": " in the diversity loss are set to 1m and 5 pixels, respectively. All parameters in our proposed model are randomly initialized and trained from scratch with the Adam optimizer. We first train the geometric prior-guided overlapping region detection module with classification loss and frustum-pose loss for 20 epochs. After that, we train the whole network with the total loss " + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{total}" + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": " for 20 epochs. The learning rate is set to " + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": ", and the weight decay is " + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 55, + 257, + 296, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 505, + 231, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 231, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 231, + 516 + ], + "type": "text", + "content": "4.2. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 522, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 545 + ], + "type": "text", + "content": "We evaluate our proposed method on two commonly used benchmarks KITTI [17] and nuScenes [3]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 546, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 665 + ], + "type": "text", + "content": "KITTI Odometry [17]. KITTI Odometry dataset consists of 22 sequences of images and point cloud data collected simultaneously in urban environments, and 11 sequences provide ground-truth calibration files. We follow the common practice [21] of utilizing the 0-8 sequences for training, and 9-10 for testing. The mis-registration transformation is conducted with a 2D translation on the ground within " + }, + { + "bbox": [ + 55, + 546, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\pm 10\\mathrm{m}" + }, + { + "bbox": [ + 55, + 546, + 295, + 665 + ], + "type": "text", + "content": " and a rotation around the up-axis with no limited range. We downsample the image resolution to " + }, + { + "bbox": [ + 55, + 546, + 295, + 665 + ], + "type": "inline_equation", + "content": "160\\times 512" + }, + { + "bbox": [ + 55, + 546, + 295, + 665 + ], + "type": "text", + "content": " and the point cloud size to 40960 for training and testing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "nuScenes [3]. nuScenes dataset collects the image-point cloud pairs by official SDK where the point cloud is accumulated from the nearby frames and the image is from the current frame. We follow the official data spilt of nuScenes," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "where 850 scenes are used for training, and 150 scenes for testing. The mis-registration transformation is conducted in a similar way as the one in KITTI dataset. We downsample the image resolution to " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "160 \\times 320" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " and the point cloud size to 40960 for training and testing." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 133, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 253 + ], + "type": "text", + "content": "Evaluation metrics. We follow VP2P-match [44] to quantify the performance of image-to-point cloud registration with average Relative Translational Error (RTE), average Relative Rotation Error (RRE) and registration accuracy (Acc). Same as VP2P-match [44], we do not follow CorrI2P [28] to remove the image-point cloud samples with large errors before averaging, which might be an unsuitable way to report real performances. Also, We further report the registration accuracy (Acc) which is the proportion of fine registrations with RTE " + }, + { + "bbox": [ + 313, + 133, + 555, + 253 + ], + "type": "inline_equation", + "content": "< 2\\mathrm{m}" + }, + { + "bbox": [ + 313, + 133, + 555, + 253 + ], + "type": "text", + "content": " and RRE " + }, + { + "bbox": [ + 313, + 133, + 555, + 253 + ], + "type": "inline_equation", + "content": "< 5^{\\circ}" + }, + { + "bbox": [ + 313, + 133, + 555, + 253 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 262, + 441, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 441, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 441, + 275 + ], + "type": "text", + "content": "4.3. Registration Accuracy" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 280, + 554, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 554, + 304 + ], + "type": "text", + "content": "Baselines. We compare our method with previous state-of-the-art methods [18, 28, 44]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 305, + 554, + 484 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 313, + 305, + 554, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 353 + ], + "type": "text", + "content": "- Grid Cls. + PnP. DeepI2P [18] proposes to divide the image into " + }, + { + "bbox": [ + 313, + 305, + 554, + 353 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 305, + 554, + 353 + ], + "type": "text", + "content": " grids and uses a neural network to classify each 3D point to a unique 2D grid. The EPnP [20] with RANSAC [16] is applied to estimate camera pose." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 354, + 553, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 400 + ], + "type": "text", + "content": "- Frus. Cls. + Inv.Proj. DeepI2P [18] proposes to perform frustum classification and inverse camera projection to obtain camera pose. We report their results with 2D/3D inverse camera projection as DeepI2P(2D)/DeepI2P(3D)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 401, + 553, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 553, + 436 + ], + "type": "text", + "content": "- Corr12P. Corr12P [28] proposes to learn dense correspondences for image-point cloud pairs, and the EPnP [20] with RANSAC [16] is applied to predict the camera pose." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 437, + 553, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 553, + 484 + ], + "type": "text", + "content": "- VP2P-match. VP2P-match [44] is the latest work on image-to-point cloud registration which represents the 3D elements as the combination of voxels and points and trains with a differentiable PnP solver [4]." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 485, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 689 + ], + "type": "text", + "content": "Quantitative Comparison. We report the quantitative results in Table 1. Our proposed method outperforms all compared methods in all metrics by a large margin on both KITTI and nuScenes datasets. In detail, on the KITTI dataset, our method achieves the highest Acc, with a reduction of " + }, + { + "bbox": [ + 313, + 485, + 555, + 689 + ], + "type": "inline_equation", + "content": "0.55\\mathrm{m}" + }, + { + "bbox": [ + 313, + 485, + 555, + 689 + ], + "type": "text", + "content": " in RTE and about " + }, + { + "bbox": [ + 313, + 485, + 555, + 689 + ], + "type": "inline_equation", + "content": "2.05^{\\circ}" + }, + { + "bbox": [ + 313, + 485, + 555, + 689 + ], + "type": "text", + "content": " in RRE compared with VP2P-match [44]. On the nuScenes dataset, Acc is " + }, + { + "bbox": [ + 313, + 485, + 555, + 689 + ], + "type": "inline_equation", + "content": "2.61\\%" + }, + { + "bbox": [ + 313, + 485, + 555, + 689 + ], + "type": "text", + "content": " higher, RTE is reduced by " + }, + { + "bbox": [ + 313, + 485, + 555, + 689 + ], + "type": "inline_equation", + "content": "0.26\\mathrm{m}" + }, + { + "bbox": [ + 313, + 485, + 555, + 689 + ], + "type": "text", + "content": " and RRE is reduced by " + }, + { + "bbox": [ + 313, + 485, + 555, + 689 + ], + "type": "inline_equation", + "content": "0.02^{\\circ}" + }, + { + "bbox": [ + 313, + 485, + 555, + 689 + ], + "type": "text", + "content": ". We attribute the superior performance to three delicate designs. Our proposed GPDM leverages geometric prior that the overlapping region is a frustum to enhance the accuracy of point cloud overlapping region detection. And the ICLM can implicitly obtain accurate 2D-3D correspondences, avoiding the unreliable process of forced alignment between 2D and 3D features. Moreover, the PRM generates precise camera pose prediction via a learned network, which enables the end-to-end optimization." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Qualitative Comparison. The qualitative comparison is shown in Figure 3. For intuitive visualization, we project" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16927" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 83, + 526, + 182 + ], + "blocks": [ + { + "bbox": [ + 75, + 71, + 533, + 81 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 533, + 81 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 533, + 81 + ], + "type": "text", + "content": "Table 1. Registration accuracy on the KITTI and nuScenes datasets. Lower is better for RTE and RRE, higher is better for Acc." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 83, + 526, + 182 + ], + "lines": [ + { + "bbox": [ + 83, + 83, + 526, + 182 + ], + "spans": [ + { + "bbox": [ + 83, + 83, + 526, + 182 + ], + "type": "table", + "html": "
MethodsKITTInuScenes
RTE(m)↓RRE(°)↓Acc(%)↑RTE(m)↓RRE(°)↓Acc(%)↑
Grid Cls. + PnP [18]3.64±3.4619.19±28.9611.223.02±2.4012.66±21.012.45
DeepI2P(3D) [18]4.06±3.5424.73±31.693.772.88±2.1220.65±12.242.26
DeepI2P(2D) [18]3.59±3.2111.66±18.1625.952.78±1.994.80±6.2138.10
CorrI2P [28]3.78±65.165.89±20.3472.423.04±60.763.73±9.0349.00
VP2P-match [44]0.75±1.133.29±7.9983.040.89±1.442.15±7.0388.33
Ours0.20±0.211.24±2.3497.490.63±0.442.13±3.7590.94
", + "image_path": "6d383488315380fc97b3d4b48422aa5907f0cb757fa690ea324874823bcbdb95.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 84, + 200, + 187, + 235 + ], + "blocks": [ + { + "bbox": [ + 121, + 187, + 143, + 198 + ], + "lines": [ + { + "bbox": [ + 121, + 187, + 143, + 198 + ], + "spans": [ + { + "bbox": [ + 121, + 187, + 143, + 198 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 84, + 200, + 187, + 235 + ], + "lines": [ + { + "bbox": [ + 84, + 200, + 187, + 235 + ], + "spans": [ + { + "bbox": [ + 84, + 200, + 187, + 235 + ], + "type": "image", + "image_path": "4c32838ae83dc69983495df10098b5cb1fee3216ad31ae01489bb60d2cbff8f1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 236, + 165, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 165, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 165, + 245 + ], + "type": "inline_equation", + "content": "7.14\\mathrm{m} / 108.94^{\\circ}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 197, + 200, + 302, + 235 + ], + "blocks": [ + { + "bbox": [ + 223, + 186, + 271, + 196 + ], + "lines": [ + { + "bbox": [ + 223, + 186, + 271, + 196 + ], + "spans": [ + { + "bbox": [ + 223, + 186, + 271, + 196 + ], + "type": "text", + "content": "VP2P-match" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 197, + 200, + 302, + 235 + ], + "lines": [ + { + "bbox": [ + 197, + 200, + 302, + 235 + ], + "spans": [ + { + "bbox": [ + 197, + 200, + 302, + 235 + ], + "type": "image", + "image_path": "ddde2f3207ff08ee012a297f7dc847203248d32495f1284ab8668dde24c3c96f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 236, + 276, + 245 + ], + "lines": [ + { + "bbox": [ + 220, + 236, + 276, + 245 + ], + "spans": [ + { + "bbox": [ + 220, + 236, + 276, + 245 + ], + "type": "text", + "content": "2.47m / 12.93°" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 201, + 413, + 235 + ], + "blocks": [ + { + "bbox": [ + 351, + 188, + 370, + 196 + ], + "lines": [ + { + "bbox": [ + 351, + 188, + 370, + 196 + ], + "spans": [ + { + "bbox": [ + 351, + 188, + 370, + 196 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 201, + 413, + 235 + ], + "lines": [ + { + "bbox": [ + 310, + 201, + 413, + 235 + ], + "spans": [ + { + "bbox": [ + 310, + 201, + 413, + 235 + ], + "type": "image", + "image_path": "8b5d2446155a657659ea2c6add92a0f94bd2dcbe3a8e3a66c2a2c5d221f6e163.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 236, + 385, + 245 + ], + "lines": [ + { + "bbox": [ + 335, + 236, + 385, + 245 + ], + "spans": [ + { + "bbox": [ + 335, + 236, + 385, + 245 + ], + "type": "inline_equation", + "content": "0.22\\mathrm{m} / 0.14^{\\circ}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 421, + 201, + 525, + 235 + ], + "blocks": [ + { + "bbox": [ + 467, + 187, + 480, + 196 + ], + "lines": [ + { + "bbox": [ + 467, + 187, + 480, + 196 + ], + "spans": [ + { + "bbox": [ + 467, + 187, + 480, + 196 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 421, + 201, + 525, + 235 + ], + "lines": [ + { + "bbox": [ + 421, + 201, + 525, + 235 + ], + "spans": [ + { + "bbox": [ + 421, + 201, + 525, + 235 + ], + "type": "image", + "image_path": "1edd3d3cec4e581cdaa21bc07c16615c9a3560cf2c54fb7b6a043d9bef22803b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 236, + 506, + 247 + ], + "lines": [ + { + "bbox": [ + 442, + 236, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 442, + 236, + 506, + 247 + ], + "type": "text", + "content": "RTE(m) / RRE(°)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 84, + 248, + 187, + 283 + ], + "blocks": [ + { + "bbox": [ + 84, + 248, + 187, + 283 + ], + "lines": [ + { + "bbox": [ + 84, + 248, + 187, + 283 + ], + "spans": [ + { + "bbox": [ + 84, + 248, + 187, + 283 + ], + "type": "image", + "image_path": "129165332e909eab0ddf6404f6907bb35f2872bc3a40d9b8d9990b04a4035177.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 285, + 160, + 293 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 160, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 160, + 293 + ], + "type": "text", + "content": "9.91m/10.53°" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 197, + 248, + 302, + 283 + ], + "blocks": [ + { + "bbox": [ + 197, + 248, + 302, + 283 + ], + "lines": [ + { + "bbox": [ + 197, + 248, + 302, + 283 + ], + "spans": [ + { + "bbox": [ + 197, + 248, + 302, + 283 + ], + "type": "image", + "image_path": "e9dc57b46863cd9bd72311c176d7b0a6ab3497c6608d05fe929a0fd256882836.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 285, + 272, + 293 + ], + "lines": [ + { + "bbox": [ + 221, + 285, + 272, + 293 + ], + "spans": [ + { + "bbox": [ + 221, + 285, + 272, + 293 + ], + "type": "inline_equation", + "content": "0.88\\mathrm{m} / 2.07^{\\circ}" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 248, + 413, + 283 + ], + "blocks": [ + { + "bbox": [ + 309, + 248, + 413, + 283 + ], + "lines": [ + { + "bbox": [ + 309, + 248, + 413, + 283 + ], + "spans": [ + { + "bbox": [ + 309, + 248, + 413, + 283 + ], + "type": "image", + "image_path": "b9644e15ce6e79335738551e787fbd60d76b61f3e1ff7fa6ab2955b31ef43a5d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 284, + 384, + 293 + ], + "lines": [ + { + "bbox": [ + 335, + 284, + 384, + 293 + ], + "spans": [ + { + "bbox": [ + 335, + 284, + 384, + 293 + ], + "type": "inline_equation", + "content": "0.07\\mathrm{m} / 1.40^{\\circ}" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 421, + 248, + 526, + 283 + ], + "blocks": [ + { + "bbox": [ + 421, + 248, + 526, + 283 + ], + "lines": [ + { + "bbox": [ + 421, + 248, + 526, + 283 + ], + "spans": [ + { + "bbox": [ + 421, + 248, + 526, + 283 + ], + "type": "image", + "image_path": "35b0405b01f0cd52267bcfe922d797b5e6ee37fc1e6d8c0dfc5f784fa7e1d481.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 284, + 506, + 294 + ], + "lines": [ + { + "bbox": [ + 442, + 284, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 442, + 284, + 506, + 294 + ], + "type": "text", + "content": "RTE(m) / RRE(°)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 84, + 296, + 187, + 330 + ], + "blocks": [ + { + "bbox": [ + 84, + 296, + 187, + 330 + ], + "lines": [ + { + "bbox": [ + 84, + 296, + 187, + 330 + ], + "spans": [ + { + "bbox": [ + 84, + 296, + 187, + 330 + ], + "type": "image", + "image_path": "cc9985eb0d079f86a4b3157cf5d66f9b3bb2282462fe49d65e70f5e206026e50.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 106, + 331, + 164, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 164, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 164, + 340 + ], + "type": "inline_equation", + "content": "12.93\\mathrm{m} / 70.03^{\\circ}" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 124, + 344, + 484, + 355 + ], + "lines": [ + { + "bbox": [ + 124, + 344, + 484, + 355 + ], + "spans": [ + { + "bbox": [ + 124, + 344, + 484, + 355 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison of Image-to-Point Cloud registration results under KITTI dataset." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 197, + 296, + 302, + 330 + ], + "blocks": [ + { + "bbox": [ + 197, + 296, + 302, + 330 + ], + "lines": [ + { + "bbox": [ + 197, + 296, + 302, + 330 + ], + "spans": [ + { + "bbox": [ + 197, + 296, + 302, + 330 + ], + "type": "image", + "image_path": "2aed46aadf950e5f649332d34195bac1d7740559aa08e60390e2e14ef927a43d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 331, + 276, + 340 + ], + "lines": [ + { + "bbox": [ + 221, + 331, + 276, + 340 + ], + "spans": [ + { + "bbox": [ + 221, + 331, + 276, + 340 + ], + "type": "inline_equation", + "content": "5.38\\mathrm{m} / 18.78^{\\circ}" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 309, + 296, + 413, + 330 + ], + "blocks": [ + { + "bbox": [ + 309, + 296, + 413, + 330 + ], + "lines": [ + { + "bbox": [ + 309, + 296, + 413, + 330 + ], + "spans": [ + { + "bbox": [ + 309, + 296, + 413, + 330 + ], + "type": "image", + "image_path": "ad9ecf6858b08f076dfedac8076c54084ddb99d716df41941993a8c921d9aa57.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 331, + 385, + 340 + ], + "lines": [ + { + "bbox": [ + 335, + 331, + 385, + 340 + ], + "spans": [ + { + "bbox": [ + 335, + 331, + 385, + 340 + ], + "type": "inline_equation", + "content": "0.07\\mathrm{m} / 1.31^{\\circ}" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 421, + 296, + 526, + 330 + ], + "blocks": [ + { + "bbox": [ + 421, + 296, + 526, + 330 + ], + "lines": [ + { + "bbox": [ + 421, + 296, + 526, + 330 + ], + "spans": [ + { + "bbox": [ + 421, + 296, + 526, + 330 + ], + "type": "image", + "image_path": "fc28450e61e06b18fa7901c62edf7fcf4f9a528d4b63b92c9e6d78f936b6aba8.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 331, + 506, + 342 + ], + "lines": [ + { + "bbox": [ + 442, + 331, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 442, + 331, + 506, + 342 + ], + "type": "text", + "content": "RTE(m) / RRE(°)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 70, + 374, + 170, + 436 + ], + "blocks": [ + { + "bbox": [ + 70, + 374, + 170, + 436 + ], + "lines": [ + { + "bbox": [ + 70, + 374, + 170, + 436 + ], + "spans": [ + { + "bbox": [ + 70, + 374, + 170, + 436 + ], + "type": "image", + "image_path": "f62219bb44d2f9df158644ab57be868e887c2168cab0a7c31a36cfcac0d991f1.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 437, + 140, + 445 + ], + "lines": [ + { + "bbox": [ + 92, + 437, + 140, + 445 + ], + "spans": [ + { + "bbox": [ + 92, + 437, + 140, + 445 + ], + "type": "inline_equation", + "content": "0.16\\mathrm{m} / 0.82^{\\circ}" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 179, + 374, + 278, + 436 + ], + "blocks": [ + { + "bbox": [ + 179, + 374, + 278, + 436 + ], + "lines": [ + { + "bbox": [ + 179, + 374, + 278, + 436 + ], + "spans": [ + { + "bbox": [ + 179, + 374, + 278, + 436 + ], + "type": "image", + "image_path": "401cef3422d067ebfe459df501f8529ee02079443619d6ccf95c5bf7678c1b52.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 437, + 251, + 445 + ], + "lines": [ + { + "bbox": [ + 203, + 437, + 251, + 445 + ], + "spans": [ + { + "bbox": [ + 203, + 437, + 251, + 445 + ], + "type": "inline_equation", + "content": "0.04\\mathrm{m} / 0.41^{\\circ}" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 70, + 453, + 170, + 515 + ], + "blocks": [ + { + "bbox": [ + 70, + 453, + 170, + 515 + ], + "lines": [ + { + "bbox": [ + 70, + 453, + 170, + 515 + ], + "spans": [ + { + "bbox": [ + 70, + 453, + 170, + 515 + ], + "type": "image", + "image_path": "263ce10b412bbc95b3e08be192c60f5e5bc6d0b0702bf58c29fbdc792b9c3079.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 516, + 143, + 524 + ], + "lines": [ + { + "bbox": [ + 96, + 516, + 143, + 524 + ], + "spans": [ + { + "bbox": [ + 96, + 516, + 143, + 524 + ], + "type": "inline_equation", + "content": "0.08\\mathrm{m} / 0.70^{\\circ}" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 525, + 295, + 547 + ], + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 547 + ], + "type": "text", + "content": "Figure 4. Visual illustration of 2D-3D correspondences and registration accuracy. Green lines represent correct correspondences." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 178, + 453, + 277, + 515 + ], + "blocks": [ + { + "bbox": [ + 178, + 453, + 277, + 515 + ], + "lines": [ + { + "bbox": [ + 178, + 453, + 277, + 515 + ], + "spans": [ + { + "bbox": [ + 178, + 453, + 277, + 515 + ], + "type": "image", + "image_path": "25133ddde07dace7da2da6ebaa0cddba96b87f730fcc016c001511c8666d4ba4.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 516, + 249, + 524 + ], + "lines": [ + { + "bbox": [ + 200, + 516, + 249, + 524 + ], + "spans": [ + { + "bbox": [ + 200, + 516, + 249, + 524 + ], + "type": "inline_equation", + "content": "0.14\\mathrm{m} / 0.59^{\\circ}" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "table", + "bbox": [ + 63, + 560, + 286, + 624 + ], + "blocks": [ + { + "bbox": [ + 79, + 549, + 271, + 559 + ], + "lines": [ + { + "bbox": [ + 79, + 549, + 271, + 559 + ], + "spans": [ + { + "bbox": [ + 79, + 549, + 271, + 559 + ], + "type": "text", + "content": "Table 2. The effect of each design in our framework." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 560, + 286, + 624 + ], + "lines": [ + { + "bbox": [ + 63, + 560, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 63, + 560, + 286, + 624 + ], + "type": "table", + "html": "
RTE(m)↓RRE(°)↓Acc(%)↑
w/o GPDM0.45±1.052.48±9.6891.89
w/o ICLM1.05±0.733.70±6.2072.64
w/o PRM1.05±2.934.96±18.4782.00
Full0.20±0.241.24±2.4097.49
", + "image_path": "5eac0fef7690a988acbe39f7dcec97a3127b53df3359aa4803d160c1d20bae0f.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "table_body" + } + ], + "index": 41 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "the point cloud into image space through predicted transformations of different approaches and the known camera intrinsic matrix, and the color indicates the depth. Compared with VP2P-match [44], our method achieves better registration accuracy in different road situations. For example, on the difficult situation (e.g. the 1st and 3rd row), VP2P-match [44] fails to solve correct registration, while" + } + ] + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 332, + 375, + 395, + 437 + ], + "blocks": [ + { + "bbox": [ + 337, + 368, + 391, + 375 + ], + "lines": [ + { + "bbox": [ + 337, + 368, + 391, + 375 + ], + "spans": [ + { + "bbox": [ + 337, + 368, + 391, + 375 + ], + "type": "text", + "content": "w/o Geometric Prior" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 375, + 395, + 437 + ], + "lines": [ + { + "bbox": [ + 332, + 375, + 395, + 437 + ], + "spans": [ + { + "bbox": [ + 332, + 375, + 395, + 437 + ], + "type": "image", + "image_path": "de9620719ba41cc6e915ff34cd8ee7f76a7a7d27cd3b3fa528ce2debad74f08e.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 438, + 377, + 445 + ], + "lines": [ + { + "bbox": [ + 349, + 438, + 377, + 445 + ], + "spans": [ + { + "bbox": [ + 349, + 438, + 377, + 445 + ], + "type": "text", + "content": "0.86/0.76" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 402, + 375, + 463, + 437 + ], + "blocks": [ + { + "bbox": [ + 408, + 368, + 457, + 375 + ], + "lines": [ + { + "bbox": [ + 408, + 368, + 457, + 375 + ], + "spans": [ + { + "bbox": [ + 408, + 368, + 457, + 375 + ], + "type": "text", + "content": "w Geometric Prior" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 402, + 375, + 463, + 437 + ], + "lines": [ + { + "bbox": [ + 402, + 375, + 463, + 437 + ], + "spans": [ + { + "bbox": [ + 402, + 375, + 463, + 437 + ], + "type": "image", + "image_path": "fb2c2f4be8e1076ce856199438f03d180cbfe2f3a17a5c9dbfcef954c2ad3cea.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 438, + 446, + 445 + ], + "lines": [ + { + "bbox": [ + 417, + 438, + 446, + 445 + ], + "spans": [ + { + "bbox": [ + 417, + 438, + 446, + 445 + ], + "type": "text", + "content": "0.96/1.00" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 474, + 375, + 536, + 437 + ], + "blocks": [ + { + "bbox": [ + 498, + 368, + 508, + 374 + ], + "lines": [ + { + "bbox": [ + 498, + 368, + 508, + 374 + ], + "spans": [ + { + "bbox": [ + 498, + 368, + 508, + 374 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 375, + 536, + 437 + ], + "lines": [ + { + "bbox": [ + 474, + 375, + 536, + 437 + ], + "spans": [ + { + "bbox": [ + 474, + 375, + 536, + 437 + ], + "type": "image", + "image_path": "7a7f41ad61a95c82f9805c5b46e130f1c001963b1366a2e0f9db7d0a779ae45c.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 438, + 523, + 445 + ], + "lines": [ + { + "bbox": [ + 482, + 438, + 523, + 445 + ], + "spans": [ + { + "bbox": [ + 482, + 438, + 523, + 445 + ], + "type": "text", + "content": "Precise/Recall" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 343, + 446, + 372, + 509 + ], + "blocks": [ + { + "bbox": [ + 343, + 446, + 372, + 509 + ], + "lines": [ + { + "bbox": [ + 343, + 446, + 372, + 509 + ], + "spans": [ + { + "bbox": [ + 343, + 446, + 372, + 509 + ], + "type": "image", + "image_path": "f0120884bd5e609efe14e641d640c6ed32f8ae1590ad8ca742ad80c67eeced91.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 511, + 377, + 517 + ], + "lines": [ + { + "bbox": [ + 348, + 511, + 377, + 517 + ], + "spans": [ + { + "bbox": [ + 348, + 511, + 377, + 517 + ], + "type": "text", + "content": "0.79/0.76" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 521, + 554, + 564 + ], + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 564 + ], + "type": "text", + "content": "Figure 5. Visual illustration of overlapping region detection. The green points are correctly detected in the overlapping region, the blue points are missed in the overlapping region, and the red points are incorrectly detected outside the overlapping region." + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 409, + 446, + 438, + 509 + ], + "blocks": [ + { + "bbox": [ + 409, + 446, + 438, + 509 + ], + "lines": [ + { + "bbox": [ + 409, + 446, + 438, + 509 + ], + "spans": [ + { + "bbox": [ + 409, + 446, + 438, + 509 + ], + "type": "image", + "image_path": "70041bb9f9ff4cd6173845d2d21c2777acb411a7cfc2c95a40446832f3ba9d02.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 510, + 445, + 517 + ], + "lines": [ + { + "bbox": [ + 417, + 510, + 445, + 517 + ], + "spans": [ + { + "bbox": [ + 417, + 510, + 445, + 517 + ], + "type": "text", + "content": "0.99 / 0.95" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 484, + 447, + 514, + 509 + ], + "blocks": [ + { + "bbox": [ + 484, + 447, + 514, + 509 + ], + "lines": [ + { + "bbox": [ + 484, + 447, + 514, + 509 + ], + "spans": [ + { + "bbox": [ + 484, + 447, + 514, + 509 + ], + "type": "image", + "image_path": "6d42dca266a6e64b33667a5d40138c08a94ee13ea1fc45dd16b836232b495c72.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 511, + 523, + 517 + ], + "lines": [ + { + "bbox": [ + 482, + 511, + 523, + 517 + ], + "spans": [ + { + "bbox": [ + 482, + 511, + 523, + 517 + ], + "type": "text", + "content": "Precise/Recall" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 56 + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "content": "our method achieves much higher registration accuracy. We further provide the 2D-3D correspondence visualization in Figure 4. Almost all of the pixels of our method can reach slight errors within 15 pixels, which demonstrates that our ICLM can obtain accurate 2D-3D correspondences." + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 313, + 635, + 410, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 410, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 410, + 648 + ], + "type": "text", + "content": "4.4. Ablation Study." + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 313, + 653, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 689 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to justify the effectiveness of each design and the effect of some important parameters in our method on the KITTI dataset." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "To justify the effectiveness of each design in our proposed method, we develop three different variations for" + } + ] + } + ], + "index": 62 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16928" + } + ] + } + ], + "index": 63 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 82, + 518, + 133 + ], + "blocks": [ + { + "bbox": [ + 141, + 71, + 470, + 81 + ], + "lines": [ + { + "bbox": [ + 141, + 71, + 470, + 81 + ], + "spans": [ + { + "bbox": [ + 141, + 71, + 470, + 81 + ], + "type": "text", + "content": "Table 3. Comparison of frustum classification results on the KITTI and nuScenes datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 90, + 82, + 518, + 133 + ], + "lines": [ + { + "bbox": [ + 90, + 82, + 518, + 133 + ], + "spans": [ + { + "bbox": [ + 90, + 82, + 518, + 133 + ], + "type": "table", + "html": "
SettingsKITTInuScenes
PrecisionRecallAccuracyF1-scorePrecisionRecallAccuracyF1-score
w/o GPDM93.4595.9097.9394.6682.8685.5993.1484.20
w GPDM95.1594.9198.0495.0392.0491.5396.3691.79
", + "image_path": "3a19ac9e75c0c1893024ef3c34e785e312418b50daa428acf2c34ecc2b5dcd48.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 294, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 294, + 236 + ], + "type": "text", + "content": "comparison in Table 2: (1) " + }, + { + "bbox": [ + 55, + 140, + 294, + 236 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 55, + 140, + 294, + 236 + ], + "type": "text", + "content": " GPDM removes the frustum pose prediction branch and regards points with infrustum probability " + }, + { + "bbox": [ + 55, + 140, + 294, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_P" + }, + { + "bbox": [ + 55, + 140, + 294, + 236 + ], + "type": "text", + "content": " greater than 0.5 as the overlapping region. (2) " + }, + { + "bbox": [ + 55, + 140, + 294, + 236 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 55, + 140, + 294, + 236 + ], + "type": "text", + "content": " ICLM removes both the ICLM and PRM since we can't make the PRM work without the ICLM. In this case, we take the predicted frustum pose as our estimated camera pose. (3) " + }, + { + "bbox": [ + 55, + 140, + 294, + 236 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 55, + 140, + 294, + 236 + ], + "type": "text", + "content": " PRM replaces the PRM with the PnP-RANSAC algorithm [16, 20]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 237, + 295, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 295, + 405 + ], + "type": "text", + "content": "Effects of the GPDM. As shown in Table 2, with GPDM, the performance on the KITTI dataset is improved on all metrics. In specific, Acc is " + }, + { + "bbox": [ + 55, + 237, + 295, + 405 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 55, + 237, + 295, + 405 + ], + "type": "text", + "content": " higher, RTE is reduced by about " + }, + { + "bbox": [ + 55, + 237, + 295, + 405 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 55, + 237, + 295, + 405 + ], + "type": "text", + "content": " and RRE is reduced by about " + }, + { + "bbox": [ + 55, + 237, + 295, + 405 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 237, + 295, + 405 + ], + "type": "text", + "content": ". The main reason is that our GPDM can accurately detect the overlapping region and generate a coarse camera pose which can be refined later. And to prove that our GPDM can improve the accuracy of point cloud overlapping region detection, we report the quantitative results in Table 3. With GPDM, we get better frustum classification results almost on all metrics, especially on the nuScenes dataset. The visual comparison is shown in Figure 5, which demonstrates our propose GPDM can get better prediction on the frustum boundary leveraging our proposed geometric prior." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 407, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 503 + ], + "type": "text", + "content": "Effects of the ICLM. As shown in Table 2, with proposed ICLM, the performance on the KITTI dataset is improved notably. In specific, Acc is " + }, + { + "bbox": [ + 55, + 407, + 295, + 503 + ], + "type": "inline_equation", + "content": "34\\%" + }, + { + "bbox": [ + 55, + 407, + 295, + 503 + ], + "type": "text", + "content": " higher, RTE is reduced by about " + }, + { + "bbox": [ + 55, + 407, + 295, + 503 + ], + "type": "inline_equation", + "content": "81\\%" + }, + { + "bbox": [ + 55, + 407, + 295, + 503 + ], + "type": "text", + "content": " and RRE is reduced by about " + }, + { + "bbox": [ + 55, + 407, + 295, + 503 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 55, + 407, + 295, + 503 + ], + "type": "text", + "content": ". Regarding the gain on Acc, we can figure out that the ICLM plays the most important role in our model, which demonstrates that our designed implicit correspondence learning is very effective for Image-to-Point Cloud registration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 505, + 295, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 295, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 295, + 565 + ], + "type": "text", + "content": "Effects of the PRM. As shown in Table 2, the PRM outperforms the PnP-RANSAC algorithm [16, 20] on all metrics. In specific, Acc is " + }, + { + "bbox": [ + 55, + 505, + 295, + 565 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 55, + 505, + 295, + 565 + ], + "type": "text", + "content": " higher, RTE is reduced by about " + }, + { + "bbox": [ + 55, + 505, + 295, + 565 + ], + "type": "inline_equation", + "content": "81\\%" + }, + { + "bbox": [ + 55, + 505, + 295, + 565 + ], + "type": "text", + "content": " and RRE is reduced by about " + }, + { + "bbox": [ + 55, + 505, + 295, + 565 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 55, + 505, + 295, + 565 + ], + "type": "text", + "content": ". We owe the advantage to the enabled end-to-end optimization with PRM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "text", + "content": "Effects of the number of correspondence queries. In Table 4, we show the impact of different numbers of correspondence queries in the ICLM on the KITTI dataset. The results show that our method can achieve better performance than VP2P-match [44] by using only 16 correspondence queries. We find that the overall performance of the model improves slightly with the increase of " + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "text", + "content": ". We choose " + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "inline_equation", + "content": "N_{q} = 128" + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "text", + "content": " in our method for the highest Acc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Effects of the iteration time. To investigate the influences of the iteration time " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "(L)" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " in the ICLM, we pick " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " from the set " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{0, 1, 2, 3, 4\\}" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ". As shown in Table 5, we notice that the performance drops a lot if " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " is set to 0, showing the" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 340, + 149, + 526, + 225 + ], + "blocks": [ + { + "bbox": [ + 340, + 138, + 527, + 148 + ], + "lines": [ + { + "bbox": [ + 340, + 138, + 527, + 148 + ], + "spans": [ + { + "bbox": [ + 340, + 138, + 527, + 148 + ], + "type": "text", + "content": "Table 4. Ablation studies on the number of queries." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 149, + 526, + 225 + ], + "lines": [ + { + "bbox": [ + 340, + 149, + 526, + 225 + ], + "spans": [ + { + "bbox": [ + 340, + 149, + 526, + 225 + ], + "type": "table", + "html": "
NqRTE(m)↓RRE(°)↓Acc(%)↑
160.26±0.291.71±2.4095.52
320.23±0.251.43±2.2495.70
640.21±0.191.28±3.2096.63
1280.20±0.241.24±2.4097.49
2560.21±0.211.21±2.4097.42
", + "image_path": "9b7d3d3593109fe72e7b7e4ca5bf7d9973578089e92c10ecd2554719288d43a7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 345, + 238, + 523, + 314 + ], + "blocks": [ + { + "bbox": [ + 349, + 228, + 518, + 237 + ], + "lines": [ + { + "bbox": [ + 349, + 228, + 518, + 237 + ], + "spans": [ + { + "bbox": [ + 349, + 228, + 518, + 237 + ], + "type": "text", + "content": "Table 5. Ablation studies on the iteration time." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 238, + 523, + 314 + ], + "lines": [ + { + "bbox": [ + 345, + 238, + 523, + 314 + ], + "spans": [ + { + "bbox": [ + 345, + 238, + 523, + 314 + ], + "type": "table", + "html": "
LRTE(m)↓RRE(°)↓Acc(%)↑
00.31±0.362.34±2.7690.01
10.21±0.161.23±1.9597.44
20.20±0.161.28±2.0297.33
30.20±0.241.24±2.4097.49
40.21±0.301.26±2.4797.46
", + "image_path": "11174fa0d12a3d9e91388a8f09896b1dea9ce8744e2b4a5ae8ba2abfc320bc08.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 316, + 328, + 553, + 367 + ], + "blocks": [ + { + "bbox": [ + 348, + 316, + 520, + 327 + ], + "lines": [ + { + "bbox": [ + 348, + 316, + 520, + 327 + ], + "spans": [ + { + "bbox": [ + 348, + 316, + 520, + 327 + ], + "type": "text", + "content": "Table 6. Ablation studies on the iteration order." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 328, + 553, + 367 + ], + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 367 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 367 + ], + "type": "table", + "html": "
RTE(m)↓RRE(°)↓Acc(%)↑
Pixel-then-Point0.20±0.241.24±2.4097.49
Point-then-Pixel0.23±0.601.29±2.2697.05
", + "image_path": "ca254273b068ff32bdbfaec1f637278ca72494b1aa0877e98252fd30bcfc4080.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 374, + 553, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 553, + 410 + ], + "type": "text", + "content": "necessity of the iteration operation. We choose the model with " + }, + { + "bbox": [ + 313, + 374, + 553, + 410 + ], + "type": "inline_equation", + "content": "L = 1" + }, + { + "bbox": [ + 313, + 374, + 553, + 410 + ], + "type": "text", + "content": " for the optimal balance between computational efficiency and performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 411, + 554, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 494 + ], + "type": "text", + "content": "Effects of different iteration orders. To investigate the influences of the iteration order in the ICLM, we conduct two experiments with first Attention-Pixel then Attention-Point or first Attention-Point then Attention-Pixel. As shown is Table 6, we get similar performance with different iteration orders. And we choose to iteratively apply Attention-Pixel then Attention-Point in our method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 506, + 388, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 388, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 388, + 518 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 526, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 554, + 658 + ], + "type": "text", + "content": "In this paper, we propose a novel implicit correspondence learning method with three specific designs for image-top-point cloud registration. Specially, the proposed geometric prior can guide the model to achieve high accuracy on the overlapping region detection. The proposed implicit correspondence learning module can implicitly acquire accurate 2D-3D correspondences. Furthermore, we propose a pose regression module to handle incorrect 2D-3D correspondences and generates precise camera pose. Extensive experimental results on two challenging benchmarks demonstrate the effectiveness of the proposed method." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 670, + 429, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 429, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 429, + 683 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "This work was supported by the National Defense Basic Scientific Research program (No.JCKY2022911B002)." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16929" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Xuyang Bai, Zixin Luo, Lei Zhou, Hongbo Fu, Long Quan, and Chiew-Lan Tai. D3feat: Joint learning of dense detection and description of 3d local features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6359-6367, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 179 + ], + "type": "text", + "content": "[2] Mark Billinghurst, Adrian Clark, Gun Lee, et al. A survey of augmented reality. Foundations and Trends® in Human-Computer Interaction, 8(2-3):73-272, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 296, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 296, + 246 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 312 + ], + "type": "text", + "content": "[4] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2781-2790, 2022. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 313, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 295, + 367 + ], + "type": "text", + "content": "[5] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 539-546. IEEE, 2005. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 422 + ], + "type": "text", + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 424, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 468 + ], + "type": "text", + "content": "[7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8958-8966, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 469, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 513 + ], + "type": "text", + "content": "[8] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In Proceedings of the European conference on computer vision (ECCV), pages 602-618, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "text", + "content": "[9] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 195-205, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 569, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 623 + ], + "type": "text", + "content": "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 689 + ], + "type": "text", + "content": "[11] Zhen Dong, Fuxun Liang, Bisheng Yang, Yusheng Xu, Yufu Zang, Jianping Li, Yuan Wang, Wenxia Dai, Hongchao Fan, Juha Hyyppä, et al. Registration of large-scale terrestrial laser scanner point clouds: A review and benchmark. IS-PRS Journal of Photogrammetry and Remote Sensing, 163: 327–342, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[12] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "object recognition. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 998-1005. IEEE, 2010. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[13] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99–110, 2006. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "type": "text", + "content": "[14] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 8092-8101, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 208, + 553, + 262 + ], + "type": "text", + "content": "[15] Mengdan Feng, Sixing Hu, Marcelo H Ang, and Gim Hee Lee. 2d3d-matchnet: Learning to match keypoints across 2d image and 3d point cloud. In 2019 International Conference on Robotics and Automation (ICRA), pages 4790-4796. IEEE, 2019. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 3, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "text", + "content": "[17] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 399 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "text", + "content": "[19] Yurim Jeon and Seung-Woo Seo. Efghnet: A versatile image-to-point cloud registration network for extreme outdoor environment. IEEE Robotics and Automation Letters, 7 (3):7511-7517, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "text", + "content": "[20] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate " + }, + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "inline_equation", + "content": "O(n)" + }, + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "text", + "content": " solution to the pnp problem. International journal of computer vision, 81:155-166, 2009. 2, 3, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 489, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 555, + 533 + ], + "type": "text", + "content": "[21] Jiaxin Li and Gim Hee Lee. Deepi2p: Image-to-point cloud registration via deep classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15960-15969, 2021. 1, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 534, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 590 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "text", + "content": "[23] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "text", + "content": "[24] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "text", + "content": "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16930" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "text", + "content": "and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "type": "text", + "content": "[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "text", + "content": "[27] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143-11152, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 295, + 250 + ], + "type": "text", + "content": "[28] Siyu Ren, Yiming Zeng, Junhui Hou, and Xiaodong Chen. Corri2p: Deep image-to-point cloud registration via dense correspondence. IEEE Transactions on Circuits and Systems for Video Technology, 33(3):1198-1208, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 295, + 294 + ], + "type": "text", + "content": "[29] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 295, + 337 + ], + "type": "text", + "content": "[30] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 337, + 295, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 295, + 382 + ], + "type": "text", + "content": "[31] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE international conference on robotics and automation, pages 3212-3217. IEEE, 2009. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 296, + 437 + ], + "type": "text", + "content": "[32] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 514 + ], + "type": "text", + "content": "[33] Paul-Edouard Sarlin, Daniel DeTone, Tsun-Yi Yang, Armen Avetisyan, Julian Straub, Tomasz Malisiewicz, Samuel Rota Bulo, Richard Newcombe, Peter Kontschieder, and Vasileios Balntas. Orienternet: Visual localization in 2d public maps with neural matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21632–21642, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 295, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 295, + 559 + ], + "type": "text", + "content": "[34] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 295, + 602 + ], + "type": "text", + "content": "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. 3rd International Conference on Learning Representations, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 295, + 658 + ], + "type": "text", + "content": "[36] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[37] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6398-6407, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 563 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 126 + ], + "type": "text", + "content": "[38] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European conference on computer vision, pages 685–702. Springer, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 127, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 127, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 127, + 554, + 194 + ], + "type": "text", + "content": "[39] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 194, + 554, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 554, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 554, + 250 + ], + "type": "text", + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 250, + 554, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 554, + 306 + ], + "type": "text", + "content": "[41] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loptr: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 306, + 554, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 554, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 554, + 351 + ], + "type": "text", + "content": "[42] Hao Yu, Fu Li, Mahdi Saleh, Benjamin Busam, and Slobodan Ilic. Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration. Advances in Neural Information Processing Systems, 34:23872-23884, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 351, + 554, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 351, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 351, + 554, + 396 + ], + "type": "text", + "content": "[43] Yu Zhong. Intrinsic shape signatures: A shape descriptor for 3d object recognition. In 2009 IEEE 12th international conference on computer vision workshops, ICCV Workshops, pages 689-696. IEEE, 2009. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 396, + 554, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 396, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 554, + 462 + ], + "type": "text", + "content": "[44] Junsheng Zhou, Baorui Ma, Wenyuan Zhang, Yi Fang, Yu-Shen Liu, and Zhizhong Han. Differentiable registration of images and lidar point clouds with voxelpoint-to-pixel matching. In Advances in Neural Information Processing Systems, pages 51166-51177. Curran Associates, Inc., 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 462, + 554, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 462, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 462, + 554, + 508 + ], + "type": "text", + "content": "[45] Qunjie Zhou, Torsten Sattler, and Laura Leal-Taixe. Patch2pix: Epipolar-guided pixel-level correspondences. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4669-4678, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 508, + 554, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 508, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 508, + 554, + 563 + ], + "type": "text", + "content": "[46] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745-5753, 2019. 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16931" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/e4bdbcad-b68b-4255-a63d-f2d51b9212bd_content_list.json b/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/e4bdbcad-b68b-4255-a63d-f2d51b9212bd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..35417d396f6ae062b199937b4d60578346d9ee74 --- /dev/null +++ b/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/e4bdbcad-b68b-4255-a63d-f2d51b9212bd_content_list.json @@ -0,0 +1,1690 @@ +[ + { + "type": "text", + "text": "Improve Representation for Imbalanced Regression through Geometric Constraints", + "text_level": 1, + "bbox": [ + 250, + 130, + 750, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zijian Dong $^{1,*}$ Yilei Wu $^{1,*}$ Chongyao Chen $^{2,*}$ Yingtian Zou $^{1}$ Yichi Zhang $^{1}$ Juan Helen Zhou $^{1,\\dagger}$", + "bbox": [ + 130, + 202, + 870, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National University of Singapore, Singapore $^{2}$ Duke University, US", + "bbox": [ + 236, + 229, + 761, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zijian.dong, yilei.wu}@u.nus.edu, helen.zhou@nus.edu.sg", + "bbox": [ + 181, + 258, + 813, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 250, + 311, + 325, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In representation learning, uniformity refers to the uniform feature distribution in the latent space (i.e., unit hypersphere). Previous work has shown that improving uniformity contributes to the learning of under-represented classes. However, most of the previous work focused on classification; the representation space of imbalanced regression remains unexplored. Classification-based methods are not suitable for regression tasks because they cluster features into distinct groups without considering the continuous and ordered nature essential for regression. In a geometric aspect, we uniquely focus on ensuring uniformity in the latent space for imbalanced regression through two key losses: enveloping and homogeneity. The enveloping loss encourages the induced trace to uniformly occupy the surface of a hypersphere, while the homogeneity loss ensures smoothness, with representations evenly spaced at consistent intervals. Our method integrates these geometric principles into the data representations via a Surrogate-driven Representation Learning (SRL) framework. Experiments with real-world regression and operator learning tasks highlight the importance of uniformity in imbalanced regression and validate the efficacy of our geometry-based loss functions. Code is available here.", + "bbox": [ + 89, + 342, + 485, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 689, + 215, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Imbalanced datasets are ubiquitous across various domains, including image recognition [30], semantic segmentation [31], and regression [25]. Previous studies have demonstrated the significance of uniform or balanced distribution of class representations for effective imbalanced classification [5, 8, 9, 12, 22, 26, 33] and imbalanced semantic segmentation [31]. In classification tasks, these representations typically form distinct clusters. However, in the context of regression, representations are expected to be continuous and ordered [24, 27], rendering the methods used for quantifying and analyzing", + "bbox": [ + 89, + 714, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eb0bcd76be5fa7fd68998314356045189c095b6a460638f2d7f490a584543caa.jpg", + "image_caption": [ + "Figure 1. 2D feature space of vanilla baseline and ours from UCI-Airfoil [2]. The vanilla feature space lacks uniformity and is dominated by samples from the Many-shot region. In contrast, our approach achieves a more uniform distribution over the feature space, improving the performance, especially in the Medium and Few-shot regions. (For visualization purposes, we curated the dataset to ensure equal partitions across the three regions.)" + ], + "image_footnote": [], + "bbox": [ + 555, + 309, + 867, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "uniformity in classification inapplicable. While the issue of deep imbalanced regression (DIR) has received considerable attention, the focus has predominantly been on training unbiased regressors, rather than on the aspect of representation learning [6, 10, 16, 19, 25]. Among the methods that do explore representation learning [24, 27], the emphasis is typically on understanding the relationship between the label space and the feature space (the representations themselves should be continuous and ordered). However, a critical aspect that remains under-explored is the interaction between data representations and the entire feature space. Specifically, how these representations distribute within the full scope of the feature space has not been examined.", + "bbox": [ + 511, + 699, + 908, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Uniformity in classification refers to how effectively", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 107, + 875, + 212, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 230, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5082", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8767a6b3f5dc2e93007e8652762c4482ff49a8222aac57199e41cf4c292bd2b6.jpg", + "image_caption": [ + "Figure 2. t-SNE visualization [20] of feature comparison. The first row corresponds to the original UCI Airfoil Dataset [2], while the second row corresponds to its curated version, with an additional few-shot region in the middle of the label range. Colored arrows point to the few-shot regions and their corresponding positions in the feature distributions. We evaluate feature distributions using: MSE Loss (Baseline), SRL without uniformity loss (w/o $\\mathcal{L}_{\\mathrm{env}}$ ), SRL without homogeneity loss (w/o $\\mathcal{L}_{\\mathrm{homo}}$ ), and complete SRL (ours). The baseline leads to feature collapse to many-shot regions and inadequate distinction of few-shot samples. In w/o $\\mathcal{L}_{\\mathrm{env}}$ , features collapse into a trivial shape, not fully utilizing the feature space. In w/o $\\mathcal{L}_{\\mathrm{homo}}$ , features spread out along the trace. Different from the previous ones, our SRL uniformly and smoothly “fills” the feature space." + ], + "image_footnote": [], + "bbox": [ + 117, + 90, + 890, + 324 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "different clusters or centroids occupy the feature space, essentially partitioning it among various classes. In regression, here we define the term \"latent trace\" as the pathway that the representations follow, delineating the transition from the minimum to the maximum label values. In this paper, we aim to evaluate how well a latent trace occupies the feature space? To quantify this, we approximate a tubular neighborhood around the latent trace and measure its volume relative to the entire feature space. This method gauges the effectiveness of the trace in \"enveloping\" the hypersphere, and we call it enveloping loss. This loss ensures that the trace shape fills the surface of the hypersphere to facilitate uniformity. In parallel, it is equally important that the points (i.e., individual data point representations) are evenly distributed along the trace. To address uniform distribution along the trace as well as smoothness, we have developed a homogeneity loss. This loss is computed based on the arc length of the trace, allowing us to effectively measure and promote an even and smooth distribution of points.", + "bbox": [ + 88, + 438, + 485, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We model the uniformity in regression in two aspects: the induced trace aims to fully occupy the surface of the hypersphere (Enveloping), exhibiting smoothness with representations spaced at uniform intervals (Homogeneity). The two losses we introduce act as geometric constraints on a latent trace, implying they should not be applied to a set of representations from a single mini-batch. This is because a single batch likely does not encompass the full range of labels. To address this, we have developed a Surrogate-driven Representation Learning (SRL) scheme. It involves averaging representations of the same bins within a mini-batch to form centroids and \"re-filling\" missing bins by taking corresponding centroids from the", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "previous epoch. This process results in a surrogate containing centroids for all bins, enabling the effective application of geometric loss across the complete label range. Furthermore, we introduce Imbalanced Operator Learning (IOL) as a new DIR benchmark for training models on imbalanced domain locations in function space mapping. In summary, our main contributions are four-fold:", + "bbox": [ + 511, + 438, + 906, + 542 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Geometric Losses for Uniformity in deep imbalanced regression (DIR). To the best of our knowledge, this work is the first to study representation learning in DIR. We introduce two novel loss functions, enveloping loss and homogeneity loss, to ensure uniform feature distribution for DIR.", + "- SRL Framework. A new framework is proposed that incorporates these geometric principles into data representations.", + "- Imbalanced Operator Learning (IOL). For the first time, we pioneer the task of operator learning within the realm of deep imbalanced regression, introducing an innovative task: Imbalanced Operator Learning (IOL).", + "- Extensive Experiments. The effectiveness of the proposed method is validated through experiments involving real-world regression and operator learning, on five datasets: AgeDB-DIR, IMDB-WIKI-DIR, STS-B-DIR, and two newly created DIR benchmarks, UCI-DIR and OL-DIR." + ], + "bbox": [ + 511, + 545, + 908, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 799, + 650, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Uniformity in imbalanced classification. Wang and Isola [22] identifies that uniformity is one of the key properties in contrastive representation learning. To promote uniformity in representation space for imbalanced classification, a variety of training strategies have been proposed. Kang et al. [8]", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5083", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "decouples the training into a two-stage training of representation learning and classification. Yin et al. [26] designs a transfer learning framework for imbalanced face recognition. Kang et al. [9] combines supervised method and contrastive learning to learn a discriminative and balanced feature space. PaCo [5] and TSC [12] learn a set of class-wise balanced centers. BCL [33] balances the gradient distribution of negative classes and data distribution in mini-batch. Recent study suggests that sample-level uniform distribution may not effectively address imbalanced classification, advocating for category-level uniformity instead [1, 32]. Though progress has been made in this field, challenges persist in adapting the approach of modeling uniformity from classification to regression.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deep imbalanced regression. With imbalanced regression data, effective learning in a continuous label space focuses on modeling the relationships among labels in the feature space [25]. Label Distribution Smoothing (LDS) [25] and DenseLoss [19] apply a Gaussian kernel to the observed label density, leading to an estimated label density distribution. Feature distribution smoothing (FDS) [25] generalizes the application of kernel smoothing from label to feature space. Ranksim [6] aims to leverage both local and global dependencies in data by aligning the order of similarities between labels and features. Balanced MSE [16] addresses the issue of imbalance in Mean Squared Error (MSE) calculations, ensuring a more balanced distribution of predictions. VIR [23] provides uncertainty for imbalanced regression. ConR [10] regularizes contrastive learning in regression by modeling global and local label similarities in feature space. RNC [27] and SupReMix [24] learn a continuous and ordered representation for regression through supervised contrastive learning. How imbalanced regression representations leverage the feature space remains under-explored.", + "bbox": [ + 91, + 287, + 483, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 588, + 178, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the field of representation learning for classification, the concept of uniformity is pivotal for maximizing the use of the feature space [5, 8, 9, 12, 22, 26, 33]. This idea is based on the principle of ensuring that features from different classes are not only distinctly separated but also evenly distributed in the latent space. This uniform distribution of class centroids fosters a clear and effective decision boundary, leading to more accurate classification. However, in regression, where we deal with continuous, ordered trace [24, 27] rather than discrete clusters, the concept of uniformity is not only more complex but essentially remains undefined.", + "bbox": [ + 89, + 613, + 483, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We draw an analogy to the process of winding yarn around a ball. In this analogy, the yarn represents the latent trace, and the ball symbolizes the entirety of the available feature space. Just as the yarn must be evenly distributed across the ball's surface to effectively cover it (without any crossing), the latent trace should strive to occupy the hypersphere of the latent space uniformly. This ensures that the model leverages the available feature space to its fullest extent, enhancing the model's ability", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4cf76af194660921134e539fab8b701f1dda0cde8ee50395c6e97068b76a9206.jpg", + "image_caption": [ + "Figure 3. 2D schematic overview of two geometric losses. The arrow indicates the improvement of the loss function. Enveloping loss encourages the representations to fill the latent space, and homogeneity loss encourages the smoothness and even distribution of the representations along the trace." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 705, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b34ec2e285c6444532d57a6144820103b4c031999efb5d48189daa7136892135.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 88, + 905, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to capture the variability inherent in the data.", + "bbox": [ + 511, + 316, + 795, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, the latent trace should be smooth and continuous, akin to the even stretching of yarn, rather than loose and disjointed. This smoothness ensures a consistent and predictable model behavior, which is crucial for the accurate prediction and interpretation of results.", + "bbox": [ + 511, + 332, + 905, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We outline our method in this section. Firstly, we establish the fundamental notations and preliminaries (Section 3.1). Following this, we delve into the concept and definition of our enveloping loss (Section 3.2) and homogeneity loss (Section 3.3). Finally, we present our Surrogate-driven Representation Learning (SRL) framework, which incorporates the geometric constraints from the global image of the representations into the local range (Section 3.4). Refer to Supplementary Material 9 for the pseudo code of our method.", + "bbox": [ + 511, + 407, + 908, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 555, + 653, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A regression dataset is composed of pairs $(\\mathbf{x}_i, y_i)$ , where $\\mathbf{x}_i$ represents the input and $y_i$ is the corresponding continuous-value target. Denote $\\mathbf{z}_i = f(\\mathbf{x}_i)$ as the feature representation of $\\mathbf{x}_i$ , generated by a neural network $f(\\cdot)$ . The feature representation is normalized so that $\\|\\mathbf{z}_i\\| = 1$ for all $i$ . Suppose the dataset consists of $K$ unique bins1, we define a surrogate as a set of centroids $\\mathbf{c}_k$ , where each represents a distinct bin. These centroids are computed by averaging the representations $\\mathbf{z}$ sharing the same bin and they are normalized to $\\|\\mathbf{c}_k\\| = 1$ . Let $l$ be a path: $l : [y_{\\min}, y_{\\max}] \\mapsto \\mathbb{R}^n$ with $\\|l(y)\\| = 1$ , such that $l(y_k) = \\mathbf{c}_k$ . The path $l$ is a continuous curve extended from the discrete dataset that lies on a submanifold of $\\mathbb{R}^n$ .", + "bbox": [ + 511, + 579, + 908, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Enveloping", + "text_level": 1, + "bbox": [ + 511, + 771, + 637, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To maximize the use of the feature space, it is crucial for the ordered and continuous trace of regression representations to fill the entire unit hypersphere as much as possible. This is by", + "bbox": [ + 511, + 794, + 906, + 842 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1The binning process (only for geometric loss calculations) facilitates surrogate formulation by using discrete centroids to approximate a continuous trace. Binning is unnecessary for most DIR datasets since they come pre-binned. The performed regression tasks follow the original dataset settings.", + "bbox": [ + 511, + 851, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5084", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "analogy with wrapping yarn with a certain length around a ball (without any crossing), aiming to cover as much surface area as possible.", + "bbox": [ + 89, + 90, + 482, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The trace of regression representations lying on a submanifold of $\\mathbb{R}^n$ has a negligible hypervolume, which makes it challenging to assess its relationship with the entire hypersphere. To address this challenge, we extend the \"line\" into a tubular neighborhood. This expansion allows us to introduce the concept of enveloping loss. Our objective with this loss function is to maximize the hypervolume of the tubular neighborhood in proportion to the total hypervolume of the hypersphere.", + "bbox": [ + 89, + 137, + 483, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denote the set of all unit vectors in $\\mathbb{R}^n$ as $\\mathcal{U}$ . Given $\\epsilon \\in (0,1)$ , define tubular neighborhood $T(l,\\epsilon)$ of $l$ as:", + "bbox": [ + 89, + 258, + 483, + 289 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT (l, \\epsilon) = \\left\\{\\mathbf {z} \\in \\mathcal {U} \\mid \\mathbf {t} \\cdot \\mathbf {z} > \\epsilon \\text {f o r s o m e} \\mathbf {t} \\in \\operatorname {I m} (l) \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 301, + 483, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where for a function $f: A \\to B$ , the image is defined as $\\operatorname{Im}(f) \\coloneqq \\{f(x), x \\in A\\}$ .", + "bbox": [ + 89, + 332, + 482, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then our enveloping loss is defined as:", + "bbox": [ + 109, + 364, + 356, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e n v}} = - \\frac {\\operatorname {v o l} (T (l , \\epsilon))}{\\operatorname {v o l} (\\mathcal {U})} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 391, + 483, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{vol}(\\cdot)$ returns the hypervolume of its input in the induced measure from the Euclidean space.", + "bbox": [ + 89, + 435, + 482, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practical scenarios, the trace is composed of discrete representations, which complicates the direct computation of the tubular neighborhood's hypervolume. To navigate this challenge, we propose a continuous-to-discrete strategy. We first generate $N$ points that are uniformly distributed across the hypersphere. We then determine the fraction of these points that fall within the neighbourhood $\\epsilon$ . This fraction effectively approximates the proportion of the hypersphere covered by the tubular neighborhood with a sufficiently large $N$ . To adapt $\\mathcal{L}_{\\mathrm{env}}$ to discrete datasets, we re-formalize our optimization objective as:", + "bbox": [ + 89, + 465, + 483, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max \\lim _ {N \\rightarrow \\infty} \\frac {P (N)}{N} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 631, + 483, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 89, + 669, + 133, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (N) := \\left| \\left\\{\\mathbf {p} _ {i} \\mid \\max _ {y} \\left\\{\\mathbf {p} _ {i} \\cdot l (y) \\right\\} > \\epsilon , i \\in [ N ] \\right\\} \\right| \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 696, + 483, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "assuming for each $N > 0$ , we can choose $N$ evenly distributed points in $\\mathcal{U}$ , and denote these points as $\\mathbf{p}_i, i \\in [N] = 1, \\dots, N$ . For numerical application, we take $N$ to be a sufficiently large number and use the standard Monte-Carlo method [17] to approximate the evenly distributed points.", + "bbox": [ + 89, + 733, + 482, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our implementation, we did not directly define $\\epsilon$ due to the non-differentiability of the binarization required to determine if a $\\mathbf{p}_i$ is within the $\\epsilon$ -tube. Instead, for each $\\mathbf{p}_i$ , we maximize the cosine similarity between $\\mathbf{p}_i$ and its closest point on the trace. In this way, we relax the step function represented by (4) to its \"soft\" version, leading to smooth gradient computation.", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Homogeneity", + "text_level": 1, + "bbox": [ + 513, + 90, + 651, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the enveloping loss effectively governs the overall distribution of representations on the hypersphere, it alone may not be entirely adequate, presenting two unresolved issues. 1) The first is distribution along the trace. The enveloping loss predominantly controls the overall shape of representations on the hypersphere, yet it does not guarantee a uniform distribution along the trace. This poses a notable concern, as it may result in uneven representation density across different trace segments. 2) The second is trace smoothness. The enveloping loss could lead to a zigzag pattern of the representations, which should be avoided. Considering age estimation from facial images as an example, the progression of facial features over time is gradual. Consequently, in the corresponding latent space, we would anticipate a similar, smooth transition without abrupt changes, underlining the desirability of a smoother trace. Interestingly, these two issues can be aptly analogized to winding yarns around a ball as well. For the yarn on the ball to be smooth, it should be tightly stretched, rather than being disjointed or loosely arranged. We name the property of a trace to be smooth with representations evenly distributed along it as homogeneity.", + "bbox": [ + 511, + 112, + 906, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We encourage such homogeneity property, i.e., smoothness of the trace $\\operatorname{Im}(l)$ and uniform distribution of representations along it, by penalizing the arc length. Formally, the homogeneity loss is defined as:", + "bbox": [ + 511, + 416, + 905, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {h o m o}} = \\int_ {y _ {\\min }} ^ {y _ {\\max }} \\left\\| \\frac {\\mathrm {d} l (y)}{\\mathrm {d} y} \\right\\| ^ {2} \\mathrm {d} y \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 481, + 906, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given $K$ different $y\\mathrm{s}$ which have been ordered, the discrete format for $\\mathcal{L}_{\\mathrm{homo}}$ is defined as a summation of the squared differences between adjacent points:", + "bbox": [ + 511, + 525, + 905, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {h o m o}} = \\sum_ {k = 1} ^ {K - 1} \\frac {\\left\\| l \\left(y _ {k + 1}\\right) - l \\left(y _ {k}\\right) \\right\\| ^ {2}}{y _ {k + 1} - y _ {k}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 579, + 905, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The use of only homogeneity loss might result in trivial solutions like representation convergence to a circle or point due to feature collapse (shown in Figure 2). The homogeneity loss should be treated as a regularization of the enveloping loss, promoting not only smoothness but also an even distribution of representations along the trace. To quantitatively define the relationship between trace arc length and these desired characteristics, we introduce Theorem 1. It demonstrates that with a given $\\operatorname{Im}(l)$ ( $l_{\\mathrm{env}}$ is fixed as it does not depend on the parameterization of $l$ ), the homogeneity loss is minimized if and only if when representations are uniformly distributed along the trace.", + "bbox": [ + 511, + 628, + 906, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. Given an image of $l$ , $\\mathcal{L}_{\\text{homo}}$ attains its minimum if and only if the representations are uniformly distributed along the trace, i.e., $\\| \\nabla_y l(y) \\| = c$ , where $c$ is a constant.", + "bbox": [ + 511, + 801, + 906, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Refer to Supplementary Material 6 for the proof.", + "bbox": [ + 532, + 854, + 836, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we formulate our geometric constraints $(\\mathcal{L}_{\\mathrm{G}})$ as a combination of enveloping and homogeneity:", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5085", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1180f604008d7fc4afc0a3c149cc193a1008316072c6d0f88fd7743b3016aec1.jpg", + "image_caption": [ + "Figure 4. Overview of Surrogate-driven Representation Learning (SRL). (1) Every mini-batch is encoded to the latent space. Some bins may not be present in the current batch. To address this, (2) it takes centroids corresponding to the missing bins from the previous epoch. These stored centroids are used to \"re-fill\" the missing bins in the current batch. (3) Average the representations for bins that appear multiple times, creating centroids for these bins. This surrogate, containing a representation for the full label range, allows for the effective application of geometric loss across all bins. (4) Loss calculation based on the surrogate. (5) Update the surrogate in memory to ensure enveloping and homogeneity. The training of the first epoch is driven by MSE loss only." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 314 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {G}} = \\lambda_ {e} \\mathcal {L} _ {\\text {e n v}} + \\lambda_ {h} \\mathcal {L} _ {\\text {h o m o}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 508, + 482, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{e}$ and $\\lambda_{h}$ are weights for the two geometric losses. In Section 4.4, we further explore the behavior of these two geometric constraints, uncovering new insights into imbalanced regression.", + "bbox": [ + 89, + 532, + 483, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Surrogate-driven Representation Learning (SRL)", + "text_level": 1, + "bbox": [ + 89, + 603, + 483, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our geometric loss $(\\mathcal{L}_{\\mathrm{G}})$ is calculated on a surrogate instead of a mini-batch (Figure 4), as the representations from one mini-batch very likely fail to capture the global image of $l$ , due to the randomness of batch sampling.", + "bbox": [ + 89, + 641, + 482, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For illustration purposes, here we assume the original dataset has already been binned, as is the case in most DIR datasets [6, 10, 25]. Let $\\mathcal{Z} = \\{\\mathbf{z}_1,\\mathbf{z}_2,\\dots,\\mathbf{z}_M\\}$ be a set of representations from a batch with batch-size $M$ , and let $\\mathcal{Y} = \\{y_1,y_2,\\dots,y_M\\}$ (repetitions of the label values might exist) be the corresponding labels. Define the centroid $\\mathbf{c}_y\\in \\mathcal{C}$ for $y$ ( $\\| \\mathbf{c}_y\\| = 1$ ) as:", + "bbox": [ + 89, + 702, + 483, + 794 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {y} = \\frac {1}{| \\left\\{\\mathbf {z} _ {m} \\mid y _ {m} = y \\right\\} |} \\sum_ {\\left\\{\\mathbf {z} _ {m} \\mid y _ {m} = y \\right\\}} \\mathbf {z} _ {m} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 804, + 483, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Suppose the whole label range is covered by a set of unique $K$ bins $\\mathcal{V}^{*} = \\{y_{1}^{*},y_{2}^{*},\\dots,y_{K}^{*}\\}$ . The centroids for these $K$ bins from the last epoch are denoted as $\\mathcal{C}' = \\{\\mathbf{c}_{y_1^*}',\\mathbf{c}_{y_2^*}',\\dots,\\mathbf{c}_{y_K^*}'\\}$ .", + "bbox": [ + 89, + 854, + 483, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The surrogate $S$ is then generated by re-filling the missing centroids:", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = \\mathcal {C} \\cup \\left\\{\\mathrm {c} ^ {\\prime} _ {y _ {k} ^ {*}} \\mid y _ {k} ^ {*} \\in \\mathcal {Y} ^ {*} \\backslash \\mathcal {Y} \\right\\} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 122, + 906, + 141 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use AdamW [13] with momentum to update parameters $\\theta$ in $f(\\cdot)$ , to ensure a smooth transition of the local shape in the batch-wise representations.", + "bbox": [ + 511, + 148, + 906, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the end of each epoch $e \\in (0, E]$ (excluding the first), we use the representations learned during that epoch to form a running surrogate $\\hat{S}^e$ . $S^{e+1}$ is formulated from the current epoch's surrogate $S^e$ and $\\hat{S}^e$ with momentum. $S^{e+1}$ is employed for training in the subsequent epoch. It facilitates a gradual transition between epochs, preventing abrupt variations: $S^{e+1} \\gets \\alpha \\cdot S^e + (1 - \\alpha) \\cdot \\hat{S}^e$ .", + "bbox": [ + 511, + 196, + 906, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We aim for individual representations from the encoder to converge towards their respective centroids that share the same label, while simultaneously distancing them from centroids associated with different labels. To achieve this, we incorporate a contrastive loss between the individual representations and the centroids. For each representation $\\mathbf{z}_m$ with $y_{m} = y$ , the centroid $\\mathbf{c}_y$ is considered as the positive, and the other centroids as negatives. The contrastive loss is defined as:", + "bbox": [ + 511, + 301, + 906, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n}} = - \\sum_ {m = 1} ^ {M} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\mathbf {z} _ {m} , \\mathbf {c} _ {y}\\right)\\right)}{\\sum_ {y ^ {*} \\in \\mathcal {Y} ^ {*}} \\exp \\left(\\operatorname {s i m} \\left(\\mathbf {z} _ {m} , \\mathbf {c} _ {y ^ {*}}\\right)\\right)} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 436, + 906, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{sim}(\\cdot)$ is the cosine similarity between two input.", + "bbox": [ + 511, + 489, + 861, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The framework is trained end-to-end, the total loss used to update the parameters $\\theta$ in $f(\\cdot)$ is defined as:", + "bbox": [ + 511, + 506, + 905, + 536 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\theta} = \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {\\mathrm {G}} + \\mathcal {L} _ {\\text {c o n}} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 550, + 906, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{reg}}$ is the mean squared error (MSE) loss.", + "bbox": [ + 532, + 580, + 841, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 612, + 638, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform extensive experiments to validate and analyze the effectiveness of SRL for deep imbalanced regression. Our regression tasks span age estimation from facial images, tabular regression, and text similarity score regression, as well as our newly established task: Imbalanced Operator Learning (IOL). This section begins by detailing the experiment setup (Section 4.1) followed by the main results (Section 4.2). The results of IOL are shown in Section 4.3 followed by the comparison with classification-based methods and hyperparameters analysis (Section 4.4).", + "bbox": [ + 511, + 638, + 906, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 511, + 801, + 689, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We employ three real-world regression datasets developed by Yang et al. [25], and our curated UCI-DIR from UCI Machine Learning Repository [2], to assess the effectiveness of SRL in deep imbalanced regression. Refer to Supplementary Material 7 for more dataset details.", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5086", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f0ec18012bcf2bccfefd4113e637db8c863463880c04643adb6c1aec7734935b.jpg", + "table_caption": [ + "Table 1. Results on UCI-DIR (MAE). We report the average MAE of three runs. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
DatasetsAirfoilAbaloneReal EstateConcrete
ShotAllManyMedFewAllManyMedFewAllManyMedFewAllManyMedFew
VANILLA5.665.115.036.754.570.882.657.970.330.270.380.377.295.776.929.74
LDS + FDS [25]5.764.454.797.795.090.903.269.260.350.330.400.346.886.216.737.59
RankSim [6]5.235.054.915.724.330.982.597.420.370.340.380.406.716.005.579.46
BalancedMSE [16]5.694.515.047.285.372.142.669.370.340.310.400.337.034.676.379.72
Ordinal Entropy [29]6.274.855.378.326.772.314.0111.610.340.290.420.357.125.506.369.31
SRL (ours)5.104.834.755.694.160.892.427.190.280.260.300.295.945.325.806.60
", + "bbox": [ + 196, + 114, + 797, + 223 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c59ec02fda9e3e661e6a219a6bbb1d0cf8ec5a736dfb5131fc4e01220956e5a5.jpg", + "table_caption": [ + "Table 2. Results on AgeDB-DIR, the best are in bold." + ], + "table_footnote": [], + "table_body": "
MetricsMAE ↓GM ↓
ShotAllManyMedFewAllManyMedFew
VANILLA7.676.669.3012.614.854.176.518.98
LDS + FDS [25]7.557.038.4610.524.864.575.386.75
RankSim [6]7.416.498.7312.474.714.155.748.92
BalancedMSE [16]7.987.588.659.935.014.835.466.30
Ordinal Entropy [29]7.606.698.8712.684.914.286.209.29
ConR [10]7.416.518.8112.044.704.135.918.59
SRL (ours)7.226.648.289.814.504.125.376.29
", + "bbox": [ + 109, + 260, + 464, + 388 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/14eda2fcc65413848790589c208930dd47f840717515a314da7f94b8819d3b70.jpg", + "table_caption": [ + "Table 3. Results on IMDB-WIKI-DIR, the best are in bold." + ], + "table_footnote": [], + "table_body": "
MetricsMAE ↓GM ↓
ShotAllManyMedFewAllManyMedFew
VANILLA8.037.1615.4826.114.544.1410.8418.64
LDS + FDS [25]7.737.2212.9823.714.404.177.8715.77
RankSim [6]7.726.9214.5225.894.293.929.7218.02
BalancedMSE [16]8.437.8413.3523.274.934.687.9015.51
Ordinal Entropy [29]8.017.1715.1526.484.474.0710.5621.11
ConR [10]7.847.1514.3625.154.434.059.9118.55
SRL (ours)7.697.0812.6522.784.284.037.2815.25
", + "bbox": [ + 102, + 426, + 470, + 554 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8d02f2f7cfeadd8a401076150144495d28dcc2a335c854dc92317844c1b72a6e.jpg", + "table_caption": [ + "Table 4. Results on STS-B-DIR, the best are in bold." + ], + "table_footnote": [], + "table_body": "
MetricsMSE ↓Pearson correlation ↑
ShotAllManyMedFewAllManyMedFew
VANILLA0.9930.9631.0001.0750.7420.6850.6930.793
LDS + FDS [25]0.9000.9110.8810.9050.7570.6980.7230.806
RankSim [6]0.8890.9070.8740.7570.7630.7080.6920.842
BalancedMSE [16]0.9090.8941.0040.8090.7570.7030.6850.831
Ordinal Entropy [29]0.9430.9021.1610.8120.7500.7020.6790.767
SRL (ours)0.8770.8860.8730.7450.7650.7080.7490.844
", + "bbox": [ + 96, + 592, + 475, + 710 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- AgeDB-DIR [25]: It serves as a benchmark for estimating age from facial images, which is derived from the AgeDB dataset [15]. It contains 12,208 images for training, 2,140 images for validation, and 2,140 images for testing.", + "- IMDB-WIKI-DIR [25]: It is a facial image dataset for age estimation derived from the IMDB-WIKI dataset [18], which consists of face images with the corresponding age. It has 191,509 images for training, 11,022 images for validation, and 11,022 for testing.", + "- STS-B-DIR [25]: It is a natural language dataset formulated from STS-B dataset [3, 21], consisting of 5,249 training" + ], + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sentence pairs, 1,000 validation pairs, and 1,000 testing pairs. Each sentence pair is labeled with the continuous similarity score.", + "bbox": [ + 526, + 237, + 906, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- UCI-DIR: To evaluate the performance of SRL on tabular data, we curated UCI Machine Learning Repository [2] to formulate UCI-DIR that includes four regression tasks (Airfoil Self-Noise, Abalone, Concrete Compressive Strength, Real estate valuation). Following the DIR setting [25], we make each regression task consist of an imbalanced training set and a balanced validation and test set.", + "bbox": [ + 511, + 282, + 906, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. In line with the established settings in DIR [25], subsets in an imbalanced training set are categorized based on the number of available training samples: many-shot region (bins with $>100$ training samples), medium-shot region (bins with 20 to 100 training samples), and few-shot region (bins with $<20$ training samples), for the three real-world datasets. For AgeDB-DIR and IMDB-WIKI-DIR, each bin represents 1 year. In the case of STS-B-DIR, bins are segmented by 0.1. For UCI-DIR, the bins are segmented by 0.1 to 1 depending on the range of regression targets. Our evaluation metrics include mean absolute error (MAE, the lower the better) and geometric mean (GM, the lower the better) for AgeDB-DIR, IMDB-WIKI-DIR and UCI-DIR. For STS-B-DIR, we use mean squared error (MSE, the lower the better) and Pearson correlation (the higher the better).", + "bbox": [ + 511, + 396, + 908, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. For age estimation (AgeDB-DIR and IMDB-WIKI-DIR), we follow the settings from Yang et al. [25], which uses ResNet-50 [7] as a backbone network. For text similarity regression (STS-B-DIR), we follow the setting from Cer et al. [3], Yang et al. [25] that uses BiLSTM + GloVe word embeddings. For tabular regression (UCI-DIR), we use an MLP with three hidden layers (d-20-30-10-1) following the setting from Cheng et al. [4]. For all baseline methods, results were produced following provided training recipes through publicly available codebase. All experimental results, including ours and baseline methods, were obtained from a server with 8 RTX 3090 GPUs.", + "bbox": [ + 511, + 609, + 908, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We consider both DIR methods [6, 10, 25] and recent techniques proposed for general regression [16, 28, 29], in addition to VANILLA regression (MSE loss). We compare the performance of SRL with all baselines on the above four datasets. Furthermore, as SRL is orthogonal to previous DIR methods, we examine the improvement of them by adding our geometric losses.", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5087", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a6304d257fc99b532b1d859fd23ec583a770306ad76766c3b40e135d9d82ebbd.jpg", + "image_caption": [ + "Figure 5. SRL performance gain compared to VANILLA across age ranges on AgeDB-DIR. The gray histogram in the background shows the distribution of samples across age groups. SRL substantially improves the performance on the medium-shot and few-shot regions (age $< 20$ and $>70$ )." + ], + "image_footnote": [], + "bbox": [ + 135, + 89, + 861, + 218 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/91467c7b5387a2b7bc8a7060e3fd062d2f8d5f2de5373158c7a85708bcc79bd1.jpg", + "table_caption": [ + "Table 5. Combine SRL with existing DIR methods (MAE)" + ], + "table_footnote": [], + "table_body": "
DatasetsAgeDB (MAE)IMDB-WIKI (MAE)
ShotAllManyMedFewAllManyMedFew
SRL+LDS+FDS [25]7.326.818.149.817.617.0312.2821.77
GAINS v.s. LDS+FDS (%)3.053.233.896.751.662.645.408.19
SRL+RankSim [6]7.296.578.5810.487.677.0812.4022.85
GAINS v.s. RankSim (%)1.62-1.231.7216.960.65-1.1514.6111.75
SRL+BalancedMSE [16]7.246.777.869.857.747.1312.7722.04
GAINS v.s. BalancedMSE (%)9.2710.699.140.898.199.064.355.29
SRL+ConR [10]7.406.878.0810.507.567.0112.0321.71
GAINS v.s. ConR (%)0.14-5.538.3913.803.681.9616.2313.68
", + "bbox": [ + 93, + 292, + 475, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 459, + 230, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To show the effectiveness of SRL on DIR, we first benchmark SRL and baselines for tabular regression on our curated UCIDIR with four different regression tasks (Table 1). Moreover, we evaluate our method on established DIR benchmarks [25] including age estimation on AgeDB-DIR and IMDB-WIKI-DIR (Table 2 & 3, and Figure 5), and text similarity regression on STS-B-DIR (Table 4). We evaluate the combination of SRL and previous DIR methods on AgeDB-DIR and IMDB-WIKI-DIR (Table 5). Notably, Table 1 and 4 omit results from ConR [10], as it depends on data augmentation, a technique not fully established in the domain of tabular data and natural language.", + "bbox": [ + 88, + 483, + 482, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Combine SRL with existing methods. Our SRL approach enhances imbalanced regression by imposing geometric constraints on feature distributions, a strategy that is orthogonal to existing methods. To illustrate this, we leverage SRL as a regularizing term in conjunction with other methods. The results of this experiment are presented in Table 5. It shows that when SRL is integrated with existing regression methods, there is improvement in performance across different regions for both datasets. This demonstrates the effectiveness and compatibility of SRL as a complementary tool in the realm of regression analysis.", + "bbox": [ + 89, + 651, + 485, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Imbalanced Operator Learning (IOL)", + "text_level": 1, + "bbox": [ + 89, + 815, + 419, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We introduce a novel task for DIR called Imbalanced Operator Learning (IOL). Traditional operator learning aims to train a neural network to model the mapping between function spaces [11, 14]. However, unlike the standard approach of uniformly", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sampling output locations, in IOL, we intentionally adjust the sampling density within the output function's domain to create regions with few, medium, and many regions (Figure 6).", + "bbox": [ + 511, + 268, + 906, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the linear operator, the model is trained to estimate the integral operator denoted as $G$ :", + "bbox": [ + 511, + 316, + 906, + 345 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nG: u (x) \\mapsto s (x) = \\int_ {0} ^ {x} u (\\tau) d \\tau , x \\in [ 0, 1 ] \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 353, + 906, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $u$ denotes the input function which is sampled from a Gaussian random field (GRF), and $s$ is the target function.", + "bbox": [ + 511, + 402, + 905, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the nonlinear operator, the model is trained to learn a particular stochastic partial differential equation (PDE):", + "bbox": [ + 511, + 431, + 905, + 463 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {d i v} \\left(e ^ {b (x; \\omega)} \\nabla u (x; \\omega)\\right) = f (x), x \\in [ 0, 1 ] \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 470, + 906, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $e^{b(x;\\omega)}$ is the diffusion efficient and $u(x;\\omega)$ is the target function.", + "bbox": [ + 511, + 505, + 905, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Denote the domain of output function as $y$ . For both linear and non-linear operator learning, we changed the original uniform sampling of $y$ to three curated regions: few/medium/many. Afterward, we manually created an imbalanced training set of 10k samples and a balanced testing test of 100k samples, namely OL-DIR.", + "bbox": [ + 511, + 537, + 906, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 6, we have a schematic overview of Imbalanced Operator Learning (IOL). The network is trained to model an integral operator $G$ . The data provided to the model is $([u,y],G(u)(y))$ . The input consists of function $u$ and sampled $y$ s from the domain of $G(u)$ . The target is $G(u)(y)$ . We manipulate the distribution density of $y$ across its range to formulate few/med./many regions. Here the imbalance comes from the unequal exposure of integral interval to the model training. Refer to Supplementary Material 7.2 for more details.", + "bbox": [ + 511, + 627, + 906, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shown in Table 6, SRL consistently outperforms VANILLA and the state-of-the-art operator learning for the whole label range including all, many-shot, medium-shot, and few-shot regions. The results position SRL as the superior approach for IOL in terms of accuracy and generalizability.", + "bbox": [ + 511, + 763, + 905, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 847, + 681, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantification of geometric impact. We further quantified the impact of geometric constraints by comparing percentages of", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5088", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5400e1ef48185ad514b107f5de4074bb2d673bae22adb6428663f47894e962f3.jpg", + "image_caption": [ + "Figure 6. Imbalanced Operator Learning." + ], + "image_footnote": [], + "bbox": [ + 166, + 88, + 405, + 299 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cad31ab045aff1e32b2786391ce6817418df1d4fe39abf3301f513119d9ec0d8.jpg", + "table_caption": [ + "Table 6. Results on OL-DIR. We report the average MAE of ten runs. The best results are bold." + ], + "table_footnote": [ + "Full results with standard deviation are reported in Supplementary Material 8.10." + ], + "table_body": "
OperationMAE(10-3) ↓MSE (10-4) ↓
ShotAllManyMedFewAllManyMedFew
Linear
VANILLA [14]15.6411.8615.4527.005.402.814.4014.20
Ordinal Entropy [29]10.079.269.8513.012.001.531.893.42
SRL (ours)9.188.329.479.331.980.981.722.67
Nonlinear
VANILLA [14]11.649.8911.0219.779.204.337.5324.70
Ordinal Entropy [29]12.919.9313.0721.0213.808.8211.8430.12
SRL (ours)11.259.489.2217.008.607.426.4114.12
", + "bbox": [ + 96, + 351, + 478, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "uniformly sampled points within few-shot regions (a measure of proportion). The results show our method significantly increases few-shot proportion (AgeDB-DIR: $1.98\\% \\rightarrow 15.80\\%$ , upper bound: $23\\%$ ; STS-B-DIR: $4.52\\% \\rightarrow 22.39\\%$ , upper bound: $38\\%$ ), leading to improved performance (Table 7).", + "bbox": [ + 89, + 523, + 483, + 599 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b897e845075bd8e3670cbb039146e9813fd4016c11d8a3425ed793febae01237.jpg", + "table_caption": [ + "Table 7. Impact of geometric constraints on few-shot proportion." + ], + "table_footnote": [], + "table_body": "
Few-shotOverall
ProportionMAEMAE
AgeDB-DIR (1.10% samples, 23% label range):
VANILLA1.98%12.617.67
LDS + FDS [25]4.95%10.527.55
Ours15.80%9.817.22
STS-B-DIR (3.49% samples, 38% label range):
VANILLA4.52%1.0750.993
LDS + FDS [25]8.13%0.9050.900
Ours22.39%0.8770.745
", + "bbox": [ + 94, + 621, + 483, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Balancing of enveloping and homogeneity: Our proposed SRL advocates for two pivotal geometric constraints in feature distribution: enveloping and homogeneity, to effectively address imbalanced regression. These two losses are modulated by their respective coefficients, $\\lambda_{e}$ for the enveloping loss and $\\lambda_{h}$ for the homogeneity loss. Figure 7 illustrates that the omission of either constraint detrimentally impacts the performance, highlighting the importance of both of them, and it demonstrates that the best", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "performance, as measured by MAE on the AgeDB-DIR dataset, is achieved when both coefficients $\\lambda_{e}$ and $\\lambda_{h}$ are set to $1e^{-1}$ .", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation studies on choices of $N$ : Table 10 (in Supplementary Material) shows that achieving optimal performance on the AgeDB-DIR and IMDB-WIKI-DIR datasets requires a sufficiently large $N$ , as a smaller $N$ may lead to imprecise calculation of the enveloping loss.", + "bbox": [ + 511, + 121, + 903, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation studies on proposed loss component: Table 11 (in Supp.) demonstrates that incorporating homogeneity, enveloping, and contrastive loss term yields superior model performance compared to using each individually.", + "bbox": [ + 511, + 196, + 906, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational cost: As shown in Table 12 (in Supp.), the computational overhead introduced by the Surrogate-driven Representation Learning (SRL) framework is comparable to that of other imbalanced regression methods.", + "bbox": [ + 511, + 257, + 903, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of bin numbers: As shown in Table 13 (in Supp.), while increasing the number of bins generally leads to better model performance, the improvements become marginal beyond certain thresholds.", + "bbox": [ + 511, + 316, + 906, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: Section 11 (in Supp.) examines the limitations of SRL, including its inability to handle higher-dimensional labels.", + "bbox": [ + 511, + 378, + 906, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a177d605c7c679d6f50dbf76a7eabe126272c52ff8f9cc41bea698ffedce2ab.jpg", + "image_caption": [ + "Figure 7. Confusion matrix of MAE on AgeDB-DIR from different values of $\\lambda_h$ and $\\lambda_e$ ." + ], + "image_footnote": [], + "bbox": [ + 555, + 409, + 867, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 648, + 627, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As the first work of exploring uniformity in deep imbalanced regression, we introduce two novel loss functions - enveloping and homogeneity loss - to encourage the uniform feature distribution of an ordered and continuous trajectory. The two loss functions serve as geometric constraints which are integrated into the data representations through a Surrogate-driven Representation Learning (SRL) framework. Furthermore, we set a new benchmark in imbalanced regression: Imbalanced Operator Learning (IOL). Extensive experiments on real-world regression and operator learning demonstrate the effectiveness of our geometrically informed approach. We emphasize the significance of uniform data representation and its impact on learning performance in imbalanced regression scenarios, advocating for a more balanced and comprehensive utilization of feature spaces in regression models.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5089", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 183, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mido Assran, Randall Balestriero, Quentin Duval, Florian Bordes, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, and Nicolas Ballas. The hidden uniform cluster prior in self-supervised learning. In *The Eleventh International Conference on Learning Representations*, 2022. 3", + "[2] Arthur Asuncion and David Newman. Uci machine learning repository, 2007. 1, 2, 5, 6", + "[3] Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. arXiv preprint arXiv:1708.00055, 2017. 6", + "[4] Xin Cheng, Yuzhou Cao, Ximing Li, Bo An, and Lei Feng. Weakly supervised regression with interval targets. arXiv preprint arXiv:2306.10458, 2023. 6", + "[5] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021. 1, 3", + "[6] Yu Gong, Greg Mori, and Fred Tung. Ranksim: Ranking similarity regularization for deep imbalanced regression. In International Conference on Machine Learning, pages 7634-7649. PMLR, 2022. 1, 3, 5, 6, 7, 4", + "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[8] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In International Conference on Learning Representations, 2019. 1, 2, 3", + "[9] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2020. 1, 3", + "[10] Mahsa Keramati, Lili Meng, and R David Evans. Contr: Contrastive regularizer for deep imbalanced regression. arXiv preprint arXiv:2309.06651, 2023. 1, 3, 5, 6, 7, 4", + "[11] Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research, 24(89):1-97, 2023. 7", + "[12] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6928, 2022. 1, 3", + "[13] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 5", + "[14] Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deep-onet based on the universal approximation theorem of operators. Nature machine intelligence, 3(3):218-229, 2021. 7, 8, 1", + "[15] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database." + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 6", + "[16] Jiawei Ren, Mingyuan Zhang, Cunjun Yu, and Ziwei Liu. Balancedmse for imbalanced visual regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7926-7935, 2022. 1, 3, 6, 7", + "[17] Christian P Robert, George Casella, and George Casella. Monte Carlo statistical methods. Springer, 1999. 4", + "[18] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Deep expectation of real and apparent age from a single image without facial landmarks. International Journal of Computer Vision, 126 (2-4):144-157, 2018. 6", + "[19] Michael Steininger, Konstantin Kobs, Padraig Davidson, Anna Krause, and Andreas Hotho. Density-based weighting for imbalanced regression. Machine Learning, 110:2187-2211, 2021. 1, 3", + "[20] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 2", + "[21] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 6", + "[22] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 1, 2, 3", + "[23] Ziyan Wang and Hao Wang. Variational imbalanced regression: Fair uncertainty quantification via probabilistic smoothing. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3", + "[24] Yilei Wu, Zijian Dong, Chongyao Chen, Wangchunshu Zhou, and Juan Helen Zhou. Mixup your own pairs. arXiv preprint arXiv:2309.16633, 2023. 1, 3", + "[25] Yuzhe Yang, Kaiwen Zha, Yingcong Chen, Hao Wang, and Dina Katabi. Delving into deep imbalanced regression. In International Conference on Machine Learning, pages 11842-11851. PMLR, 2021. 1, 3, 5, 6, 7, 8, 2", + "[26] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5704-5713, 2019. 1, 3", + "[27] Kaiwen Zha, Peng Cao, Jeany Son, Yuzhe Yang, and Dina Katabi. Rank-n-contrast: Learning continuous representations for regression. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 3", + "[28] Shihao Zhang, Linlin Yang, Michael Bi Mi, Xiaoxu Zheng, and Angela Yao. Improving deep regression with ordinal entropy. In The Eleventh International Conference on Learning Representations, 2022. 6", + "[29] Shihao Zhang, Linlin Yang, Michael Bi Mi, Xiaoxu Zheng, and Angela Yao. Improving deep regression with ordinal entropy. arXiv preprint arXiv:2301.08915, 2023. 6, 8", + "[30] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[31] Zhisheng Zhong, Jiequan Cui, Yibo Yang, Xiaoyang Wu, Xiaojuan Qi, Xiangyu Zhang, and Jiaya Jia. Understanding" + ], + "bbox": [ + 516, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5090", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "imbalanced semantic segmentation through neural collapse. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19550-19560, 2023. 1", + "[32] Zhihan Zhou, Jiangchao Yao, Feng Hong, Ya Zhang, Bo Han, and Yanfeng Wang. Combating representation learning disparity with geometric harmonization. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[33] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022. 1, 3" + ], + "bbox": [ + 91, + 90, + 482, + 260 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5091", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/e4bdbcad-b68b-4255-a63d-f2d51b9212bd_model.json b/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/e4bdbcad-b68b-4255-a63d-f2d51b9212bd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4da98c83f453ad2a4e59ee4e8904000c287bda40 --- /dev/null +++ b/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/e4bdbcad-b68b-4255-a63d-f2d51b9212bd_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.25, + 0.131, + 0.75, + 0.175 + ], + "angle": 0, + "content": "Improve Representation for Imbalanced Regression through Geometric Constraints" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.203, + 0.871, + 0.224 + ], + "angle": 0, + "content": "Zijian Dong\\(^{1,*}\\) Yilei Wu\\(^{1,*}\\) Chongyao Chen\\(^{2,*}\\) Yingtian Zou\\(^{1}\\) Yichi Zhang\\(^{1}\\) Juan Helen Zhou\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.231, + 0.762, + 0.251 + ], + "angle": 0, + "content": "\\(^{1}\\)National University of Singapore, Singapore \\(^{2}\\) Duke University, US" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.26, + 0.815, + 0.28 + ], + "angle": 0, + "content": "{zijian.dong, yilei.wu}@u.nus.edu, helen.zhou@nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.25, + 0.312, + 0.326, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.486, + 0.661 + ], + "angle": 0, + "content": "In representation learning, uniformity refers to the uniform feature distribution in the latent space (i.e., unit hypersphere). Previous work has shown that improving uniformity contributes to the learning of under-represented classes. However, most of the previous work focused on classification; the representation space of imbalanced regression remains unexplored. Classification-based methods are not suitable for regression tasks because they cluster features into distinct groups without considering the continuous and ordered nature essential for regression. In a geometric aspect, we uniquely focus on ensuring uniformity in the latent space for imbalanced regression through two key losses: enveloping and homogeneity. The enveloping loss encourages the induced trace to uniformly occupy the surface of a hypersphere, while the homogeneity loss ensures smoothness, with representations evenly spaced at consistent intervals. Our method integrates these geometric principles into the data representations via a Surrogate-driven Representation Learning (SRL) framework. Experiments with real-world regression and operator learning tasks highlight the importance of uniformity in imbalanced regression and validate the efficacy of our geometry-based loss functions. Code is available here." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.69, + 0.217, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.486, + 0.868 + ], + "angle": 0, + "content": "Imbalanced datasets are ubiquitous across various domains, including image recognition [30], semantic segmentation [31], and regression [25]. Previous studies have demonstrated the significance of uniform or balanced distribution of class representations for effective imbalanced classification [5, 8, 9, 12, 22, 26, 33] and imbalanced semantic segmentation [31]. In classification tasks, these representations typically form distinct clusters. However, in the context of regression, representations are expected to be continuous and ordered [24, 27], rendering the methods used for quantifying and analyzing" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.31, + 0.868, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.566, + 0.909, + 0.664 + ], + "angle": 0, + "content": "Figure 1. 2D feature space of vanilla baseline and ours from UCI-Airfoil [2]. The vanilla feature space lacks uniformity and is dominated by samples from the Many-shot region. In contrast, our approach achieves a more uniform distribution over the feature space, improving the performance, especially in the Medium and Few-shot regions. (For visualization purposes, we curated the dataset to ensure equal partitions across the three regions.)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.909, + 0.881 + ], + "angle": 0, + "content": "uniformity in classification inapplicable. While the issue of deep imbalanced regression (DIR) has received considerable attention, the focus has predominantly been on training unbiased regressors, rather than on the aspect of representation learning [6, 10, 16, 19, 25]. Among the methods that do explore representation learning [24, 27], the emphasis is typically on understanding the relationship between the label space and the feature space (the representations themselves should be continuous and ordered). However, a critical aspect that remains under-explored is the interaction between data representations and the entire feature space. Specifically, how these representations distribute within the full scope of the feature space has not been examined." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Uniformity in classification refers to how effectively" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.213, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.231, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.231, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5082" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.092, + 0.891, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.343, + 0.907, + 0.429 + ], + "angle": 0, + "content": "Figure 2. t-SNE visualization [20] of feature comparison. The first row corresponds to the original UCI Airfoil Dataset [2], while the second row corresponds to its curated version, with an additional few-shot region in the middle of the label range. Colored arrows point to the few-shot regions and their corresponding positions in the feature distributions. We evaluate feature distributions using: MSE Loss (Baseline), SRL without uniformity loss (w/o \\(\\mathcal{L}_{\\mathrm{env}}\\)), SRL without homogeneity loss (w/o \\(\\mathcal{L}_{\\mathrm{homo}}\\)), and complete SRL (ours). The baseline leads to feature collapse to many-shot regions and inadequate distinction of few-shot samples. In w/o \\(\\mathcal{L}_{\\mathrm{env}}\\), features collapse into a trivial shape, not fully utilizing the feature space. In w/o \\(\\mathcal{L}_{\\mathrm{homo}}\\), features spread out along the trace. Different from the previous ones, our SRL uniformly and smoothly “fills” the feature space." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.486, + 0.711 + ], + "angle": 0, + "content": "different clusters or centroids occupy the feature space, essentially partitioning it among various classes. In regression, here we define the term \"latent trace\" as the pathway that the representations follow, delineating the transition from the minimum to the maximum label values. In this paper, we aim to evaluate how well a latent trace occupies the feature space? To quantify this, we approximate a tubular neighborhood around the latent trace and measure its volume relative to the entire feature space. This method gauges the effectiveness of the trace in \"enveloping\" the hypersphere, and we call it enveloping loss. This loss ensures that the trace shape fills the surface of the hypersphere to facilitate uniformity. In parallel, it is equally important that the points (i.e., individual data point representations) are evenly distributed along the trace. To address uniform distribution along the trace as well as smoothness, we have developed a homogeneity loss. This loss is computed based on the arc length of the trace, allowing us to effectively measure and promote an even and smooth distribution of points." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We model the uniformity in regression in two aspects: the induced trace aims to fully occupy the surface of the hypersphere (Enveloping), exhibiting smoothness with representations spaced at uniform intervals (Homogeneity). The two losses we introduce act as geometric constraints on a latent trace, implying they should not be applied to a set of representations from a single mini-batch. This is because a single batch likely does not encompass the full range of labels. To address this, we have developed a Surrogate-driven Representation Learning (SRL) scheme. It involves averaging representations of the same bins within a mini-batch to form centroids and \"re-filling\" missing bins by taking corresponding centroids from the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.544 + ], + "angle": 0, + "content": "previous epoch. This process results in a surrogate containing centroids for all bins, enabling the effective application of geometric loss across the complete label range. Furthermore, we introduce Imbalanced Operator Learning (IOL) as a new DIR benchmark for training models on imbalanced domain locations in function space mapping. In summary, our main contributions are four-fold:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.546, + 0.907, + 0.621 + ], + "angle": 0, + "content": "- Geometric Losses for Uniformity in deep imbalanced regression (DIR). To the best of our knowledge, this work is the first to study representation learning in DIR. We introduce two novel loss functions, enveloping loss and homogeneity loss, to ensure uniform feature distribution for DIR." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.622, + 0.909, + 0.652 + ], + "angle": 0, + "content": "- SRL Framework. A new framework is proposed that incorporates these geometric principles into data representations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.653, + 0.908, + 0.713 + ], + "angle": 0, + "content": "- Imbalanced Operator Learning (IOL). For the first time, we pioneer the task of operator learning within the realm of deep imbalanced regression, introducing an innovative task: Imbalanced Operator Learning (IOL)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.713, + 0.909, + 0.788 + ], + "angle": 0, + "content": "- Extensive Experiments. The effectiveness of the proposed method is validated through experiments involving real-world regression and operator learning, on five datasets: AgeDB-DIR, IMDB-WIKI-DIR, STS-B-DIR, and two newly created DIR benchmarks, UCI-DIR and OL-DIR." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.546, + 0.909, + 0.788 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.651, + 0.816 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Uniformity in imbalanced classification. Wang and Isola [22] identifies that uniformity is one of the key properties in contrastive representation learning. To promote uniformity in representation space for imbalanced classification, a variety of training strategies have been proposed. Kang et al. [8]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5083" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.288 + ], + "angle": 0, + "content": "decouples the training into a two-stage training of representation learning and classification. Yin et al. [26] designs a transfer learning framework for imbalanced face recognition. Kang et al. [9] combines supervised method and contrastive learning to learn a discriminative and balanced feature space. PaCo [5] and TSC [12] learn a set of class-wise balanced centers. BCL [33] balances the gradient distribution of negative classes and data distribution in mini-batch. Recent study suggests that sample-level uniform distribution may not effectively address imbalanced classification, advocating for category-level uniformity instead [1, 32]. Though progress has been made in this field, challenges persist in adapting the approach of modeling uniformity from classification to regression." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.289, + 0.485, + 0.576 + ], + "angle": 0, + "content": "Deep imbalanced regression. With imbalanced regression data, effective learning in a continuous label space focuses on modeling the relationships among labels in the feature space [25]. Label Distribution Smoothing (LDS) [25] and DenseLoss [19] apply a Gaussian kernel to the observed label density, leading to an estimated label density distribution. Feature distribution smoothing (FDS) [25] generalizes the application of kernel smoothing from label to feature space. Ranksim [6] aims to leverage both local and global dependencies in data by aligning the order of similarities between labels and features. Balanced MSE [16] addresses the issue of imbalance in Mean Squared Error (MSE) calculations, ensuring a more balanced distribution of predictions. VIR [23] provides uncertainty for imbalanced regression. ConR [10] regularizes contrastive learning in regression by modeling global and local label similarities in feature space. RNC [27] and SupReMix [24] learn a continuous and ordered representation for regression through supervised contrastive learning. How imbalanced regression representations leverage the feature space remains under-explored." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.179, + 0.604 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.78 + ], + "angle": 0, + "content": "In the field of representation learning for classification, the concept of uniformity is pivotal for maximizing the use of the feature space [5, 8, 9, 12, 22, 26, 33]. This idea is based on the principle of ensuring that features from different classes are not only distinctly separated but also evenly distributed in the latent space. This uniform distribution of class centroids fosters a clear and effective decision boundary, leading to more accurate classification. However, in regression, where we deal with continuous, ordered trace [24, 27] rather than discrete clusters, the concept of uniformity is not only more complex but essentially remains undefined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We draw an analogy to the process of winding yarn around a ball. In this analogy, the yarn represents the latent trace, and the ball symbolizes the entirety of the available feature space. Just as the yarn must be evenly distributed across the ball's surface to effectively cover it (without any crossing), the latent trace should strive to occupy the hypersphere of the latent space uniformly. This ensures that the model leverages the available feature space to its fullest extent, enhancing the model's ability" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.707, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.089, + 0.906, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.22, + 0.907, + 0.292 + ], + "angle": 0, + "content": "Figure 3. 2D schematic overview of two geometric losses. The arrow indicates the improvement of the loss function. Enveloping loss encourages the representations to fill the latent space, and homogeneity loss encourages the smoothness and even distribution of the representations along the trace." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.318, + 0.796, + 0.332 + ], + "angle": 0, + "content": "to capture the variability inherent in the data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.906, + 0.408 + ], + "angle": 0, + "content": "Furthermore, the latent trace should be smooth and continuous, akin to the even stretching of yarn, rather than loose and disjointed. This smoothness ensures a consistent and predictable model behavior, which is crucial for the accurate prediction and interpretation of results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.909, + 0.545 + ], + "angle": 0, + "content": "We outline our method in this section. Firstly, we establish the fundamental notations and preliminaries (Section 3.1). Following this, we delve into the concept and definition of our enveloping loss (Section 3.2) and homogeneity loss (Section 3.3). Finally, we present our Surrogate-driven Representation Learning (SRL) framework, which incorporates the geometric constraints from the global image of the representations into the local range (Section 3.4). Refer to Supplementary Material 9 for the pseudo code of our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.556, + 0.655, + 0.57 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.909, + 0.761 + ], + "angle": 0, + "content": "A regression dataset is composed of pairs \\((\\mathbf{x}_i, y_i)\\), where \\(\\mathbf{x}_i\\) represents the input and \\(y_i\\) is the corresponding continuous-value target. Denote \\(\\mathbf{z}_i = f(\\mathbf{x}_i)\\) as the feature representation of \\(\\mathbf{x}_i\\), generated by a neural network \\(f(\\cdot)\\). The feature representation is normalized so that \\(\\|\\mathbf{z}_i\\| = 1\\) for all \\(i\\). Suppose the dataset consists of \\(K\\) unique bins1, we define a surrogate as a set of centroids \\(\\mathbf{c}_k\\), where each represents a distinct bin. These centroids are computed by averaging the representations \\(\\mathbf{z}\\) sharing the same bin and they are normalized to \\(\\|\\mathbf{c}_k\\| = 1\\). Let \\(l\\) be a path: \\(l : [y_{\\min}, y_{\\max}] \\mapsto \\mathbb{R}^n\\) with \\(\\|l(y)\\| = 1\\), such that \\(l(y_k) = \\mathbf{c}_k\\). The path \\(l\\) is a continuous curve extended from the discrete dataset that lies on a submanifold of \\(\\mathbb{R}^n\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.638, + 0.788 + ], + "angle": 0, + "content": "3.2. Enveloping" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.908, + 0.843 + ], + "angle": 0, + "content": "To maximize the use of the feature space, it is crucial for the ordered and continuous trace of regression representations to fill the entire unit hypersphere as much as possible. This is by" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.852, + 0.909, + 0.903 + ], + "angle": 0, + "content": "1The binning process (only for geometric loss calculations) facilitates surrogate formulation by using discrete centroids to approximate a continuous trace. Binning is unnecessary for most DIR datasets since they come pre-binned. The performed regression tasks follow the original dataset settings." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5084" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.136 + ], + "angle": 0, + "content": "analogy with wrapping yarn with a certain length around a ball (without any crossing), aiming to cover as much surface area as possible." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.484, + 0.259 + ], + "angle": 0, + "content": "The trace of regression representations lying on a submanifold of \\(\\mathbb{R}^n\\) has a negligible hypervolume, which makes it challenging to assess its relationship with the entire hypersphere. To address this challenge, we extend the \"line\" into a tubular neighborhood. This expansion allows us to introduce the concept of enveloping loss. Our objective with this loss function is to maximize the hypervolume of the tubular neighborhood in proportion to the total hypervolume of the hypersphere." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.484, + 0.29 + ], + "angle": 0, + "content": "Denote the set of all unit vectors in \\(\\mathbb{R}^n\\) as \\(\\mathcal{U}\\). Given \\(\\epsilon \\in (0,1)\\), define tubular neighborhood \\(T(l,\\epsilon)\\) of \\(l\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.303, + 0.484, + 0.32 + ], + "angle": 0, + "content": "\\[\nT (l, \\epsilon) = \\left\\{\\mathbf {z} \\in \\mathcal {U} \\mid \\mathbf {t} \\cdot \\mathbf {z} > \\epsilon \\text {f o r s o m e} \\mathbf {t} \\in \\operatorname {I m} (l) \\right\\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.483, + 0.364 + ], + "angle": 0, + "content": "where for a function \\( f: A \\to B \\), the image is defined as \\( \\operatorname{Im}(f) \\coloneqq \\{f(x), x \\in A\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.365, + 0.357, + 0.379 + ], + "angle": 0, + "content": "Then our enveloping loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.392, + 0.484, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e n v}} = - \\frac {\\operatorname {v o l} (T (l , \\epsilon))}{\\operatorname {v o l} (\\mathcal {U})} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.483, + 0.466 + ], + "angle": 0, + "content": "where \\(\\mathrm{vol}(\\cdot)\\) returns the hypervolume of its input in the induced measure from the Euclidean space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.618 + ], + "angle": 0, + "content": "In practical scenarios, the trace is composed of discrete representations, which complicates the direct computation of the tubular neighborhood's hypervolume. To navigate this challenge, we propose a continuous-to-discrete strategy. We first generate \\(N\\) points that are uniformly distributed across the hypersphere. We then determine the fraction of these points that fall within the neighbourhood \\(\\epsilon\\). This fraction effectively approximates the proportion of the hypersphere covered by the tubular neighborhood with a sufficiently large \\(N\\). To adapt \\(\\mathcal{L}_{\\mathrm{env}}\\) to discrete datasets, we re-formalize our optimization objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.632, + 0.484, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\max \\lim _ {N \\rightarrow \\infty} \\frac {P (N)}{N} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.67, + 0.134, + 0.683 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.698, + 0.484, + 0.721 + ], + "angle": 0, + "content": "\\[\nP (N) := \\left| \\left\\{\\mathbf {p} _ {i} \\mid \\max _ {y} \\left\\{\\mathbf {p} _ {i} \\cdot l (y) \\right\\} > \\epsilon , i \\in [ N ] \\right\\} \\right| \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.81 + ], + "angle": 0, + "content": "assuming for each \\(N > 0\\), we can choose \\(N\\) evenly distributed points in \\(\\mathcal{U}\\), and denote these points as \\(\\mathbf{p}_i, i \\in [N] = 1, \\dots, N\\). For numerical application, we take \\(N\\) to be a sufficiently large number and use the standard Monte-Carlo method [17] to approximate the evenly distributed points." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In our implementation, we did not directly define \\(\\epsilon\\) due to the non-differentiability of the binarization required to determine if a \\(\\mathbf{p}_i\\) is within the \\(\\epsilon\\)-tube. Instead, for each \\(\\mathbf{p}_i\\), we maximize the cosine similarity between \\(\\mathbf{p}_i\\) and its closest point on the trace. In this way, we relax the step function represented by (4) to its \"soft\" version, leading to smooth gradient computation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.652, + 0.108 + ], + "angle": 0, + "content": "3.3. Homogeneity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.416 + ], + "angle": 0, + "content": "While the enveloping loss effectively governs the overall distribution of representations on the hypersphere, it alone may not be entirely adequate, presenting two unresolved issues. 1) The first is distribution along the trace. The enveloping loss predominantly controls the overall shape of representations on the hypersphere, yet it does not guarantee a uniform distribution along the trace. This poses a notable concern, as it may result in uneven representation density across different trace segments. 2) The second is trace smoothness. The enveloping loss could lead to a zigzag pattern of the representations, which should be avoided. Considering age estimation from facial images as an example, the progression of facial features over time is gradual. Consequently, in the corresponding latent space, we would anticipate a similar, smooth transition without abrupt changes, underlining the desirability of a smoother trace. Interestingly, these two issues can be aptly analogized to winding yarns around a ball as well. For the yarn on the ball to be smooth, it should be tightly stretched, rather than being disjointed or loosely arranged. We name the property of a trace to be smooth with representations evenly distributed along it as homogeneity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.906, + 0.475 + ], + "angle": 0, + "content": "We encourage such homogeneity property, i.e., smoothness of the trace \\(\\operatorname{Im}(l)\\) and uniform distribution of representations along it, by penalizing the arc length. Formally, the homogeneity loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.482, + 0.907, + 0.519 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {h o m o}} = \\int_ {y _ {\\min }} ^ {y _ {\\max }} \\left\\| \\frac {\\mathrm {d} l (y)}{\\mathrm {d} y} \\right\\| ^ {2} \\mathrm {d} y \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.526, + 0.906, + 0.572 + ], + "angle": 0, + "content": "Given \\(K\\) different \\(y\\mathrm{s}\\) which have been ordered, the discrete format for \\(\\mathcal{L}_{\\mathrm{homo}}\\) is defined as a summation of the squared differences between adjacent points:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.58, + 0.906, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {h o m o}} = \\sum_ {k = 1} ^ {K - 1} \\frac {\\left\\| l \\left(y _ {k + 1}\\right) - l \\left(y _ {k}\\right) \\right\\| ^ {2}}{y _ {k + 1} - y _ {k}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.795 + ], + "angle": 0, + "content": "The use of only homogeneity loss might result in trivial solutions like representation convergence to a circle or point due to feature collapse (shown in Figure 2). The homogeneity loss should be treated as a regularization of the enveloping loss, promoting not only smoothness but also an even distribution of representations along the trace. To quantitatively define the relationship between trace arc length and these desired characteristics, we introduce Theorem 1. It demonstrates that with a given \\(\\operatorname{Im}(l)\\) (\\(l_{\\mathrm{env}}\\) is fixed as it does not depend on the parameterization of \\(l\\)), the homogeneity loss is minimized if and only if when representations are uniformly distributed along the trace." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.803, + 0.908, + 0.849 + ], + "angle": 0, + "content": "Theorem 1. Given an image of \\(l\\), \\(\\mathcal{L}_{\\text{homo}}\\) attains its minimum if and only if the representations are uniformly distributed along the trace, i.e., \\(\\| \\nabla_y l(y) \\| = c\\), where \\(c\\) is a constant." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.856, + 0.838, + 0.871 + ], + "angle": 0, + "content": "Refer to Supplementary Material 6 for the proof." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Therefore, we formulate our geometric constraints \\((\\mathcal{L}_{\\mathrm{G}})\\) as a combination of enveloping and homogeneity:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5085" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.327, + 0.485, + 0.48 + ], + "angle": 0, + "content": "Figure 4. Overview of Surrogate-driven Representation Learning (SRL). (1) Every mini-batch is encoded to the latent space. Some bins may not be present in the current batch. To address this, (2) it takes centroids corresponding to the missing bins from the previous epoch. These stored centroids are used to \"re-fill\" the missing bins in the current batch. (3) Average the representations for bins that appear multiple times, creating centroids for these bins. This surrogate, containing a representation for the full label range, allows for the effective application of geometric loss across all bins. (4) Loss calculation based on the surrogate. (5) Update the surrogate in memory to ensure enveloping and homogeneity. The training of the first epoch is driven by MSE loss only." + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.509, + 0.483, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {G}} = \\lambda_ {e} \\mathcal {L} _ {\\text {e n v}} + \\lambda_ {h} \\mathcal {L} _ {\\text {h o m o}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.594 + ], + "angle": 0, + "content": "where \\(\\lambda_{e}\\) and \\(\\lambda_{h}\\) are weights for the two geometric losses. In Section 4.4, we further explore the behavior of these two geometric constraints, uncovering new insights into imbalanced regression." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.484, + 0.634 + ], + "angle": 0, + "content": "3.4. Surrogate-driven Representation Learning (SRL)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Our geometric loss \\((\\mathcal{L}_{\\mathrm{G}})\\) is calculated on a surrogate instead of a mini-batch (Figure 4), as the representations from one mini-batch very likely fail to capture the global image of \\(l\\), due to the randomness of batch sampling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.795 + ], + "angle": 0, + "content": "For illustration purposes, here we assume the original dataset has already been binned, as is the case in most DIR datasets [6, 10, 25]. Let \\(\\mathcal{Z} = \\{\\mathbf{z}_1,\\mathbf{z}_2,\\dots,\\mathbf{z}_M\\}\\) be a set of representations from a batch with batch-size \\(M\\), and let \\(\\mathcal{Y} = \\{y_1,y_2,\\dots,y_M\\}\\) (repetitions of the label values might exist) be the corresponding labels. Define the centroid \\(\\mathbf{c}_y\\in \\mathcal{C}\\) for \\(y\\) (\\(\\| \\mathbf{c}_y\\| = 1\\)) as:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.805, + 0.484, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {y} = \\frac {1}{| \\left\\{\\mathbf {z} _ {m} \\mid y _ {m} = y \\right\\} |} \\sum_ {\\left\\{\\mathbf {z} _ {m} \\mid y _ {m} = y \\right\\}} \\mathbf {z} _ {m} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.485, + 0.905 + ], + "angle": 0, + "content": "Suppose the whole label range is covered by a set of unique \\(K\\) bins \\(\\mathcal{V}^{*} = \\{y_{1}^{*},y_{2}^{*},\\dots,y_{K}^{*}\\}\\). The centroids for these \\(K\\) bins from the last epoch are denoted as \\(\\mathcal{C}' = \\{\\mathbf{c}_{y_1^*}',\\mathbf{c}_{y_2^*}',\\dots,\\mathbf{c}_{y_K^*}'\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.12 + ], + "angle": 0, + "content": "The surrogate \\( S \\) is then generated by re-filling the missing centroids:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.123, + 0.907, + 0.142 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = \\mathcal {C} \\cup \\left\\{\\mathrm {c} ^ {\\prime} _ {y _ {k} ^ {*}} \\mid y _ {k} ^ {*} \\in \\mathcal {Y} ^ {*} \\backslash \\mathcal {Y} \\right\\} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.15, + 0.907, + 0.195 + ], + "angle": 0, + "content": "We use AdamW [13] with momentum to update parameters \\(\\theta\\) in \\(f(\\cdot)\\), to ensure a smooth transition of the local shape in the batch-wise representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.908, + 0.302 + ], + "angle": 0, + "content": "At the end of each epoch \\( e \\in (0, E] \\) (excluding the first), we use the representations learned during that epoch to form a running surrogate \\( \\hat{S}^e \\). \\( S^{e+1} \\) is formulated from the current epoch's surrogate \\( S^e \\) and \\( \\hat{S}^e \\) with momentum. \\( S^{e+1} \\) is employed for training in the subsequent epoch. It facilitates a gradual transition between epochs, preventing abrupt variations: \\( S^{e+1} \\gets \\alpha \\cdot S^e + (1 - \\alpha) \\cdot \\hat{S}^e \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.907, + 0.424 + ], + "angle": 0, + "content": "We aim for individual representations from the encoder to converge towards their respective centroids that share the same label, while simultaneously distancing them from centroids associated with different labels. To achieve this, we incorporate a contrastive loss between the individual representations and the centroids. For each representation \\(\\mathbf{z}_m\\) with \\(y_{m} = y\\), the centroid \\(\\mathbf{c}_y\\) is considered as the positive, and the other centroids as negatives. The contrastive loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.437, + 0.907, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n}} = - \\sum_ {m = 1} ^ {M} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\mathbf {z} _ {m} , \\mathbf {c} _ {y}\\right)\\right)}{\\sum_ {y ^ {*} \\in \\mathcal {Y} ^ {*}} \\exp \\left(\\operatorname {s i m} \\left(\\mathbf {z} _ {m} , \\mathbf {c} _ {y ^ {*}}\\right)\\right)} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.491, + 0.862, + 0.506 + ], + "angle": 0, + "content": "where \\(\\mathrm{sim}(\\cdot)\\) is the cosine similarity between two input." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.507, + 0.906, + 0.537 + ], + "angle": 0, + "content": "The framework is trained end-to-end, the total loss used to update the parameters \\(\\theta\\) in \\(f(\\cdot)\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.551, + 0.907, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\theta} = \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {\\mathrm {G}} + \\mathcal {L} _ {\\text {c o n}} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.581, + 0.843, + 0.598 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{reg}}\\) is the mean squared error (MSE) loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.613, + 0.64, + 0.63 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.908, + 0.79 + ], + "angle": 0, + "content": "We perform extensive experiments to validate and analyze the effectiveness of SRL for deep imbalanced regression. Our regression tasks span age estimation from facial images, tabular regression, and text similarity score regression, as well as our newly established task: Imbalanced Operator Learning (IOL). This section begins by detailing the experiment setup (Section 4.1) followed by the main results (Section 4.2). The results of IOL are shown in Section 4.3 followed by the comparison with classification-based methods and hyperparameters analysis (Section 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.691, + 0.818 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Datasets. We employ three real-world regression datasets developed by Yang et al. [25], and our curated UCI-DIR from UCI Machine Learning Repository [2], to assess the effectiveness of SRL in deep imbalanced regression. Refer to Supplementary Material 7 for more dataset details." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5086" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.193, + 0.089, + 0.804, + 0.104 + ], + "angle": 0, + "content": "Table 1. Results on UCI-DIR (MAE). We report the average MAE of three runs. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.197, + 0.115, + 0.798, + 0.224 + ], + "angle": 0, + "content": "
DatasetsAirfoilAbaloneReal EstateConcrete
ShotAllManyMedFewAllManyMedFewAllManyMedFewAllManyMedFew
VANILLA5.665.115.036.754.570.882.657.970.330.270.380.377.295.776.929.74
LDS + FDS [25]5.764.454.797.795.090.903.269.260.350.330.400.346.886.216.737.59
RankSim [6]5.235.054.915.724.330.982.597.420.370.340.380.406.716.005.579.46
BalancedMSE [16]5.694.515.047.285.372.142.669.370.340.310.400.337.034.676.379.72
Ordinal Entropy [29]6.274.855.378.326.772.314.0111.610.340.290.420.357.125.506.369.31
SRL (ours)5.104.834.755.694.160.892.427.190.280.260.300.295.945.325.806.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.135, + 0.235, + 0.438, + 0.249 + ], + "angle": 0, + "content": "Table 2. Results on AgeDB-DIR, the best are in bold." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.261, + 0.465, + 0.389 + ], + "angle": 0, + "content": "
MetricsMAE ↓GM ↓
ShotAllManyMedFewAllManyMedFew
VANILLA7.676.669.3012.614.854.176.518.98
LDS + FDS [25]7.557.038.4610.524.864.575.386.75
RankSim [6]7.416.498.7312.474.714.155.748.92
BalancedMSE [16]7.987.588.659.935.014.835.466.30
Ordinal Entropy [29]7.606.698.8712.684.914.286.209.29
ConR [10]7.416.518.8112.044.704.135.918.59
SRL (ours)7.226.648.289.814.504.125.376.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.402, + 0.454, + 0.415 + ], + "angle": 0, + "content": "Table 3. Results on IMDB-WIKI-DIR, the best are in bold." + }, + { + "type": "table", + "bbox": [ + 0.103, + 0.427, + 0.472, + 0.555 + ], + "angle": 0, + "content": "
MetricsMAE ↓GM ↓
ShotAllManyMedFewAllManyMedFew
VANILLA8.037.1615.4826.114.544.1410.8418.64
LDS + FDS [25]7.737.2212.9823.714.404.177.8715.77
RankSim [6]7.726.9214.5225.894.293.929.7218.02
BalancedMSE [16]8.437.8413.3523.274.934.687.9015.51
Ordinal Entropy [29]8.017.1715.1526.484.474.0710.5621.11
ConR [10]7.847.1514.3625.154.434.059.9118.55
SRL (ours)7.697.0812.6522.784.284.037.2815.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.137, + 0.568, + 0.436, + 0.581 + ], + "angle": 0, + "content": "Table 4. Results on STS-B-DIR, the best are in bold." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.593, + 0.477, + 0.711 + ], + "angle": 0, + "content": "
MetricsMSE ↓Pearson correlation ↑
ShotAllManyMedFewAllManyMedFew
VANILLA0.9930.9631.0001.0750.7420.6850.6930.793
LDS + FDS [25]0.9000.9110.8810.9050.7570.6980.7230.806
RankSim [6]0.8890.9070.8740.7570.7630.7080.6920.842
BalancedMSE [16]0.9090.8941.0040.8090.7570.7030.6850.831
Ordinal Entropy [29]0.9430.9021.1610.8120.7500.7020.6790.767
SRL (ours)0.8770.8860.8730.7450.7650.7080.7490.844
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.735, + 0.483, + 0.796 + ], + "angle": 0, + "content": "- AgeDB-DIR [25]: It serves as a benchmark for estimating age from facial images, which is derived from the AgeDB dataset [15]. It contains 12,208 images for training, 2,140 images for validation, and 2,140 images for testing." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.796, + 0.483, + 0.871 + ], + "angle": 0, + "content": "- IMDB-WIKI-DIR [25]: It is a facial image dataset for age estimation derived from the IMDB-WIKI dataset [18], which consists of face images with the corresponding age. It has 191,509 images for training, 11,022 images for validation, and 11,022 for testing." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- STS-B-DIR [25]: It is a natural language dataset formulated from STS-B dataset [3, 21], consisting of 5,249 training" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.238, + 0.907, + 0.281 + ], + "angle": 0, + "content": "sentence pairs, 1,000 validation pairs, and 1,000 testing pairs. Each sentence pair is labeled with the continuous similarity score." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.283, + 0.907, + 0.388 + ], + "angle": 0, + "content": "- UCI-DIR: To evaluate the performance of SRL on tabular data, we curated UCI Machine Learning Repository [2] to formulate UCI-DIR that includes four regression tasks (Airfoil Self-Noise, Abalone, Concrete Compressive Strength, Real estate valuation). Following the DIR setting [25], we make each regression task consist of an imbalanced training set and a balanced validation and test set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.909, + 0.609 + ], + "angle": 0, + "content": "Metrics. In line with the established settings in DIR [25], subsets in an imbalanced training set are categorized based on the number of available training samples: many-shot region (bins with \\(>100\\) training samples), medium-shot region (bins with 20 to 100 training samples), and few-shot region (bins with \\(<20\\) training samples), for the three real-world datasets. For AgeDB-DIR and IMDB-WIKI-DIR, each bin represents 1 year. In the case of STS-B-DIR, bins are segmented by 0.1. For UCI-DIR, the bins are segmented by 0.1 to 1 depending on the range of regression targets. Our evaluation metrics include mean absolute error (MAE, the lower the better) and geometric mean (GM, the lower the better) for AgeDB-DIR, IMDB-WIKI-DIR and UCI-DIR. For STS-B-DIR, we use mean squared error (MSE, the lower the better) and Pearson correlation (the higher the better)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.909, + 0.791 + ], + "angle": 0, + "content": "Implementation Details. For age estimation (AgeDB-DIR and IMDB-WIKI-DIR), we follow the settings from Yang et al. [25], which uses ResNet-50 [7] as a backbone network. For text similarity regression (STS-B-DIR), we follow the setting from Cer et al. [3], Yang et al. [25] that uses BiLSTM + GloVe word embeddings. For tabular regression (UCI-DIR), we use an MLP with three hidden layers (d-20-30-10-1) following the setting from Cheng et al. [4]. For all baseline methods, results were produced following provided training recipes through publicly available codebase. All experimental results, including ours and baseline methods, were obtained from a server with 8 RTX 3090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Baselines. We consider both DIR methods [6, 10, 25] and recent techniques proposed for general regression [16, 28, 29], in addition to VANILLA regression (MSE loss). We compare the performance of SRL with all baselines on the above four datasets. Furthermore, as SRL is orthogonal to previous DIR methods, we examine the improvement of them by adding our geometric losses." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5087" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.136, + 0.09, + 0.862, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.23, + 0.908, + 0.261 + ], + "angle": 0, + "content": "Figure 5. SRL performance gain compared to VANILLA across age ranges on AgeDB-DIR. The gray histogram in the background shows the distribution of samples across age groups. SRL substantially improves the performance on the medium-shot and few-shot regions (age \\(< 20\\) and \\(>70\\))." + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.268, + 0.455, + 0.284 + ], + "angle": 0, + "content": "Table 5. Combine SRL with existing DIR methods (MAE)" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.293, + 0.477, + 0.448 + ], + "angle": 0, + "content": "
DatasetsAgeDB (MAE)IMDB-WIKI (MAE)
ShotAllManyMedFewAllManyMedFew
SRL+LDS+FDS [25]7.326.818.149.817.617.0312.2821.77
GAINS v.s. LDS+FDS (%)3.053.233.896.751.662.645.408.19
SRL+RankSim [6]7.296.578.5810.487.677.0812.4022.85
GAINS v.s. RankSim (%)1.62-1.231.7216.960.65-1.1514.6111.75
SRL+BalancedMSE [16]7.246.777.869.857.747.1312.7722.04
GAINS v.s. BalancedMSE (%)9.2710.699.140.898.199.064.355.29
SRL+ConR [10]7.406.878.0810.507.567.0112.0321.71
GAINS v.s. ConR (%)0.14-5.538.3913.803.681.9616.2313.68
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.46, + 0.231, + 0.475 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.484, + 0.483, + 0.65 + ], + "angle": 0, + "content": "To show the effectiveness of SRL on DIR, we first benchmark SRL and baselines for tabular regression on our curated UCIDIR with four different regression tasks (Table 1). Moreover, we evaluate our method on established DIR benchmarks [25] including age estimation on AgeDB-DIR and IMDB-WIKI-DIR (Table 2 & 3, and Figure 5), and text similarity regression on STS-B-DIR (Table 4). We evaluate the combination of SRL and previous DIR methods on AgeDB-DIR and IMDB-WIKI-DIR (Table 5). Notably, Table 1 and 4 omit results from ConR [10], as it depends on data augmentation, a technique not fully established in the domain of tabular data and natural language." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.486, + 0.804 + ], + "angle": 0, + "content": "Combine SRL with existing methods. Our SRL approach enhances imbalanced regression by imposing geometric constraints on feature distributions, a strategy that is orthogonal to existing methods. To illustrate this, we leverage SRL as a regularizing term in conjunction with other methods. The results of this experiment are presented in Table 5. It shows that when SRL is integrated with existing regression methods, there is improvement in performance across different regions for both datasets. This demonstrates the effectiveness and compatibility of SRL as a complementary tool in the realm of regression analysis." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.816, + 0.421, + 0.833 + ], + "angle": 0, + "content": "4.3. Imbalanced Operator Learning (IOL)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "We introduce a novel task for DIR called Imbalanced Operator Learning (IOL). Traditional operator learning aims to train a neural network to model the mapping between function spaces [11, 14]. However, unlike the standard approach of uniformly" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.27, + 0.907, + 0.317 + ], + "angle": 0, + "content": "sampling output locations, in IOL, we intentionally adjust the sampling density within the output function's domain to create regions with few, medium, and many regions (Figure 6)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.317, + 0.907, + 0.347 + ], + "angle": 0, + "content": "For the linear operator, the model is trained to estimate the integral operator denoted as \\( G \\):" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.354, + 0.907, + 0.387 + ], + "angle": 0, + "content": "\\[\nG: u (x) \\mapsto s (x) = \\int_ {0} ^ {x} u (\\tau) d \\tau , x \\in [ 0, 1 ] \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.403, + 0.906, + 0.433 + ], + "angle": 0, + "content": "where \\( u \\) denotes the input function which is sampled from a Gaussian random field (GRF), and \\( s \\) is the target function." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.906, + 0.464 + ], + "angle": 0, + "content": "For the nonlinear operator, the model is trained to learn a particular stochastic partial differential equation (PDE):" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.472, + 0.907, + 0.498 + ], + "angle": 0, + "content": "\\[\n\\operatorname {d i v} \\left(e ^ {b (x; \\omega)} \\nabla u (x; \\omega)\\right) = f (x), x \\in [ 0, 1 ] \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.506, + 0.906, + 0.536 + ], + "angle": 0, + "content": "where \\( e^{b(x;\\omega)} \\) is the diffusion efficient and \\( u(x;\\omega) \\) is the target function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.908, + 0.627 + ], + "angle": 0, + "content": "Denote the domain of output function as \\( y \\). For both linear and non-linear operator learning, we changed the original uniform sampling of \\( y \\) to three curated regions: few/medium/many. Afterward, we manually created an imbalanced training set of 10k samples and a balanced testing test of 100k samples, namely OL-DIR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.764 + ], + "angle": 0, + "content": "In Figure 6, we have a schematic overview of Imbalanced Operator Learning (IOL). The network is trained to model an integral operator \\( G \\). The data provided to the model is \\( ([u,y],G(u)(y)) \\). The input consists of function \\( u \\) and sampled \\( y \\)s from the domain of \\( G(u) \\). The target is \\( G(u)(y) \\). We manipulate the distribution density of \\( y \\) across its range to formulate few/med./many regions. Here the imbalance comes from the unequal exposure of integral interval to the model training. Refer to Supplementary Material 7.2 for more details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Shown in Table 6, SRL consistently outperforms VANILLA and the state-of-the-art operator learning for the whole label range including all, many-shot, medium-shot, and few-shot regions. The results position SRL as the superior approach for IOL in terms of accuracy and generalizability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.682, + 0.864 + ], + "angle": 0, + "content": "4.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Quantification of geometric impact. We further quantified the impact of geometric constraints by comparing percentages of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5088" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.168, + 0.089, + 0.406, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.301, + 0.403, + 0.314 + ], + "angle": 0, + "content": "Figure 6. Imbalanced Operator Learning." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.323, + 0.484, + 0.35 + ], + "angle": 0, + "content": "Table 6. Results on OL-DIR. We report the average MAE of ten runs. The best results are bold." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.352, + 0.479, + 0.502 + ], + "angle": 0, + "content": "
OperationMAE(10-3) ↓MSE (10-4) ↓
ShotAllManyMedFewAllManyMedFew
Linear
VANILLA [14]15.6411.8615.4527.005.402.814.4014.20
Ordinal Entropy [29]10.079.269.8513.012.001.531.893.42
SRL (ours)9.188.329.479.331.980.981.722.67
Nonlinear
VANILLA [14]11.649.8911.0219.779.204.337.5324.70
Ordinal Entropy [29]12.919.9313.0721.0213.808.8211.8430.12
SRL (ours)11.259.489.2217.008.607.426.4114.12
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.111, + 0.504, + 0.466, + 0.516 + ], + "angle": 0, + "content": "Full results with standard deviation are reported in Supplementary Material 8.10." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.6 + ], + "angle": 0, + "content": "uniformly sampled points within few-shot regions (a measure of proportion). The results show our method significantly increases few-shot proportion (AgeDB-DIR: \\(1.98\\% \\rightarrow 15.80\\%\\), upper bound: \\(23\\%\\); STS-B-DIR: \\(4.52\\% \\rightarrow 22.39\\%\\), upper bound: \\(38\\%\\)), leading to improved performance (Table 7)." + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.607, + 0.468, + 0.621 + ], + "angle": 0, + "content": "Table 7. Impact of geometric constraints on few-shot proportion." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.622, + 0.484, + 0.771 + ], + "angle": 0, + "content": "
Few-shotOverall
ProportionMAEMAE
AgeDB-DIR (1.10% samples, 23% label range):
VANILLA1.98%12.617.67
LDS + FDS [25]4.95%10.527.55
Ours15.80%9.817.22
STS-B-DIR (3.49% samples, 38% label range):
VANILLA4.52%1.0750.993
LDS + FDS [25]8.13%0.9050.900
Ours22.39%0.8770.745
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Balancing of enveloping and homogeneity: Our proposed SRL advocates for two pivotal geometric constraints in feature distribution: enveloping and homogeneity, to effectively address imbalanced regression. These two losses are modulated by their respective coefficients, \\(\\lambda_{e}\\) for the enveloping loss and \\(\\lambda_{h}\\) for the homogeneity loss. Figure 7 illustrates that the omission of either constraint detrimentally impacts the performance, highlighting the importance of both of them, and it demonstrates that the best" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "performance, as measured by MAE on the AgeDB-DIR dataset, is achieved when both coefficients \\(\\lambda_{e}\\) and \\(\\lambda_{h}\\) are set to \\(1e^{-1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.122, + 0.905, + 0.197 + ], + "angle": 0, + "content": "Ablation studies on choices of \\( N \\): Table 10 (in Supplementary Material) shows that achieving optimal performance on the AgeDB-DIR and IMDB-WIKI-DIR datasets requires a sufficiently large \\( N \\), as a smaller \\( N \\) may lead to imprecise calculation of the enveloping loss." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.198, + 0.907, + 0.258 + ], + "angle": 0, + "content": "Ablation studies on proposed loss component: Table 11 (in Supp.) demonstrates that incorporating homogeneity, enveloping, and contrastive loss term yields superior model performance compared to using each individually." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.905, + 0.318 + ], + "angle": 0, + "content": "Computational cost: As shown in Table 12 (in Supp.), the computational overhead introduced by the Surrogate-driven Representation Learning (SRL) framework is comparable to that of other imbalanced regression methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.318, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Impact of bin numbers: As shown in Table 13 (in Supp.), while increasing the number of bins generally leads to better model performance, the improvements become marginal beyond certain thresholds." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.907, + 0.409 + ], + "angle": 0, + "content": "Limitations: Section 11 (in Supp.) examines the limitations of SRL, including its inability to handle higher-dimensional labels." + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.41, + 0.868, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.611, + 0.907, + 0.64 + ], + "angle": 0, + "content": "Figure 7. Confusion matrix of MAE on AgeDB-DIR from different values of \\(\\lambda_h\\) and \\(\\lambda_e\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.628, + 0.665 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "As the first work of exploring uniformity in deep imbalanced regression, we introduce two novel loss functions - enveloping and homogeneity loss - to encourage the uniform feature distribution of an ordered and continuous trajectory. The two loss functions serve as geometric constraints which are integrated into the data representations through a Surrogate-driven Representation Learning (SRL) framework. Furthermore, we set a new benchmark in imbalanced regression: Imbalanced Operator Learning (IOL). Extensive experiments on real-world regression and operator learning demonstrate the effectiveness of our geometrically informed approach. We emphasize the significance of uniform data representation and its impact on learning performance in imbalanced regression scenarios, advocating for a more balanced and comprehensive utilization of feature spaces in regression models." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5089" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.184, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Mido Assran, Randall Balestriero, Quentin Duval, Florian Bordes, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, and Nicolas Ballas. The hidden uniform cluster prior in self-supervised learning. In *The Eleventh International Conference on Learning Representations*, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.185, + 0.483, + 0.213 + ], + "angle": 0, + "content": "[2] Arthur Asuncion and David Newman. Uci machine learning repository, 2007. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.215, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[3] Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. arXiv preprint arXiv:1708.00055, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.271, + 0.484, + 0.311 + ], + "angle": 0, + "content": "[4] Xin Cheng, Yuzhou Cao, Ximing Li, Bo An, and Lei Feng. Weakly supervised regression with interval targets. arXiv preprint arXiv:2306.10458, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.313, + 0.484, + 0.367 + ], + "angle": 0, + "content": "[5] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.369, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[6] Yu Gong, Greg Mori, and Fred Tung. Ranksim: Ranking similarity regularization for deep imbalanced regression. In International Conference on Machine Learning, pages 7634-7649. PMLR, 2022. 1, 3, 5, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.425, + 0.484, + 0.479 + ], + "angle": 0, + "content": "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.481, + 0.484, + 0.535 + ], + "angle": 0, + "content": "[8] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In International Conference on Learning Representations, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.537, + 0.484, + 0.578 + ], + "angle": 0, + "content": "[9] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.58, + 0.484, + 0.62 + ], + "angle": 0, + "content": "[10] Mahsa Keramati, Lili Meng, and R David Evans. Contr: Contrastive regularizer for deep imbalanced regression. arXiv preprint arXiv:2309.06651, 2023. 1, 3, 5, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.622, + 0.484, + 0.689 + ], + "angle": 0, + "content": "[11] Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research, 24(89):1-97, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.484, + 0.76 + ], + "angle": 0, + "content": "[12] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6928, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[13] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[14] Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deep-onet based on the universal approximation theorem of operators. Nature machine intelligence, 3(3):218-229, 2021. 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[15] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.094, + 0.907, + 0.121 + ], + "angle": 0, + "content": "In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[16] Jiawei Ren, Mingyuan Zhang, Cunjun Yu, and Ziwei Liu. Balancedmse for imbalanced visual regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7926-7935, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.206 + ], + "angle": 0, + "content": "[17] Christian P Robert, George Casella, and George Casella. Monte Carlo statistical methods. Springer, 1999. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[18] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Deep expectation of real and apparent age from a single image without facial landmarks. International Journal of Computer Vision, 126 (2-4):144-157, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.305 + ], + "angle": 0, + "content": "[19] Michael Steininger, Konstantin Kobs, Padraig Davidson, Anna Krause, and Andreas Hotho. Density-based weighting for imbalanced regression. Machine Learning, 110:2187-2211, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.906, + 0.334 + ], + "angle": 0, + "content": "[20] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.336, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[21] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[22] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[23] Ziyan Wang and Hao Wang. Variational imbalanced regression: Fair uncertainty quantification via probabilistic smoothing. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[24] Yilei Wu, Zijian Dong, Chongyao Chen, Wangchunshu Zhou, and Juan Helen Zhou. Mixup your own pairs. arXiv preprint arXiv:2309.16633, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.907, + 0.602 + ], + "angle": 0, + "content": "[25] Yuzhe Yang, Kaiwen Zha, Yingcong Chen, Hao Wang, and Dina Katabi. Delving into deep imbalanced regression. In International Conference on Machine Learning, pages 11842-11851. PMLR, 2021. 1, 3, 5, 6, 7, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.672 + ], + "angle": 0, + "content": "[26] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5704-5713, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.906, + 0.73 + ], + "angle": 0, + "content": "[27] Kaiwen Zha, Peng Cao, Jeany Son, Yuzhe Yang, and Dina Katabi. Rank-n-contrast: Learning continuous representations for regression. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[28] Shihao Zhang, Linlin Yang, Michael Bi Mi, Xiaoxu Zheng, and Angela Yao. Improving deep regression with ordinal entropy. In The Eleventh International Conference on Learning Representations, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.828 + ], + "angle": 0, + "content": "[29] Shihao Zhang, Linlin Yang, Michael Bi Mi, Xiaoxu Zheng, and Angela Yao. Improving deep regression with ordinal entropy. arXiv preprint arXiv:2301.08915, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[30] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[31] Zhisheng Zhong, Jiequan Cui, Yibo Yang, Xiaoyang Wu, Xiaojuan Qi, Xiangyu Zhang, and Jiaya Jia. Understanding" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5090" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "imbalanced semantic segmentation through neural collapse. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19550-19560, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[32] Zhihan Zhou, Jiangchao Yao, Feng Hong, Ya Zhang, Bo Han, and Yanfeng Wang. Combating representation learning disparity with geometric harmonization. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[33] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "5091" + } + ] +] \ No newline at end of file diff --git a/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/e4bdbcad-b68b-4255-a63d-f2d51b9212bd_origin.pdf b/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/e4bdbcad-b68b-4255-a63d-f2d51b9212bd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..345292473502544ecf7c81af949091364140c29b --- /dev/null +++ b/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/e4bdbcad-b68b-4255-a63d-f2d51b9212bd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e119885c8346a95ae60c5e27414e99f35bce194267425efa9e2a8650255e7ce3 +size 8985298 diff --git a/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/full.md b/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a5e27bd8c0e160ceba2c2f0305305194ef3bc104 --- /dev/null +++ b/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/full.md @@ -0,0 +1,335 @@ +# Improve Representation for Imbalanced Regression through Geometric Constraints + +Zijian Dong $^{1,*}$ Yilei Wu $^{1,*}$ Chongyao Chen $^{2,*}$ Yingtian Zou $^{1}$ Yichi Zhang $^{1}$ Juan Helen Zhou $^{1,\dagger}$ + +$^{1}$ National University of Singapore, Singapore $^{2}$ Duke University, US + +{zijian.dong, yilei.wu}@u.nus.edu, helen.zhou@nus.edu.sg + +# Abstract + +In representation learning, uniformity refers to the uniform feature distribution in the latent space (i.e., unit hypersphere). Previous work has shown that improving uniformity contributes to the learning of under-represented classes. However, most of the previous work focused on classification; the representation space of imbalanced regression remains unexplored. Classification-based methods are not suitable for regression tasks because they cluster features into distinct groups without considering the continuous and ordered nature essential for regression. In a geometric aspect, we uniquely focus on ensuring uniformity in the latent space for imbalanced regression through two key losses: enveloping and homogeneity. The enveloping loss encourages the induced trace to uniformly occupy the surface of a hypersphere, while the homogeneity loss ensures smoothness, with representations evenly spaced at consistent intervals. Our method integrates these geometric principles into the data representations via a Surrogate-driven Representation Learning (SRL) framework. Experiments with real-world regression and operator learning tasks highlight the importance of uniformity in imbalanced regression and validate the efficacy of our geometry-based loss functions. Code is available here. + +# 1. Introduction + +Imbalanced datasets are ubiquitous across various domains, including image recognition [30], semantic segmentation [31], and regression [25]. Previous studies have demonstrated the significance of uniform or balanced distribution of class representations for effective imbalanced classification [5, 8, 9, 12, 22, 26, 33] and imbalanced semantic segmentation [31]. In classification tasks, these representations typically form distinct clusters. However, in the context of regression, representations are expected to be continuous and ordered [24, 27], rendering the methods used for quantifying and analyzing + +![](images/eb0bcd76be5fa7fd68998314356045189c095b6a460638f2d7f490a584543caa.jpg) +Figure 1. 2D feature space of vanilla baseline and ours from UCI-Airfoil [2]. The vanilla feature space lacks uniformity and is dominated by samples from the Many-shot region. In contrast, our approach achieves a more uniform distribution over the feature space, improving the performance, especially in the Medium and Few-shot regions. (For visualization purposes, we curated the dataset to ensure equal partitions across the three regions.) + +uniformity in classification inapplicable. While the issue of deep imbalanced regression (DIR) has received considerable attention, the focus has predominantly been on training unbiased regressors, rather than on the aspect of representation learning [6, 10, 16, 19, 25]. Among the methods that do explore representation learning [24, 27], the emphasis is typically on understanding the relationship between the label space and the feature space (the representations themselves should be continuous and ordered). However, a critical aspect that remains under-explored is the interaction between data representations and the entire feature space. Specifically, how these representations distribute within the full scope of the feature space has not been examined. + +Uniformity in classification refers to how effectively + +![](images/8767a6b3f5dc2e93007e8652762c4482ff49a8222aac57199e41cf4c292bd2b6.jpg) +Figure 2. t-SNE visualization [20] of feature comparison. The first row corresponds to the original UCI Airfoil Dataset [2], while the second row corresponds to its curated version, with an additional few-shot region in the middle of the label range. Colored arrows point to the few-shot regions and their corresponding positions in the feature distributions. We evaluate feature distributions using: MSE Loss (Baseline), SRL without uniformity loss (w/o $\mathcal{L}_{\mathrm{env}}$ ), SRL without homogeneity loss (w/o $\mathcal{L}_{\mathrm{homo}}$ ), and complete SRL (ours). The baseline leads to feature collapse to many-shot regions and inadequate distinction of few-shot samples. In w/o $\mathcal{L}_{\mathrm{env}}$ , features collapse into a trivial shape, not fully utilizing the feature space. In w/o $\mathcal{L}_{\mathrm{homo}}$ , features spread out along the trace. Different from the previous ones, our SRL uniformly and smoothly “fills” the feature space. + +different clusters or centroids occupy the feature space, essentially partitioning it among various classes. In regression, here we define the term "latent trace" as the pathway that the representations follow, delineating the transition from the minimum to the maximum label values. In this paper, we aim to evaluate how well a latent trace occupies the feature space? To quantify this, we approximate a tubular neighborhood around the latent trace and measure its volume relative to the entire feature space. This method gauges the effectiveness of the trace in "enveloping" the hypersphere, and we call it enveloping loss. This loss ensures that the trace shape fills the surface of the hypersphere to facilitate uniformity. In parallel, it is equally important that the points (i.e., individual data point representations) are evenly distributed along the trace. To address uniform distribution along the trace as well as smoothness, we have developed a homogeneity loss. This loss is computed based on the arc length of the trace, allowing us to effectively measure and promote an even and smooth distribution of points. + +We model the uniformity in regression in two aspects: the induced trace aims to fully occupy the surface of the hypersphere (Enveloping), exhibiting smoothness with representations spaced at uniform intervals (Homogeneity). The two losses we introduce act as geometric constraints on a latent trace, implying they should not be applied to a set of representations from a single mini-batch. This is because a single batch likely does not encompass the full range of labels. To address this, we have developed a Surrogate-driven Representation Learning (SRL) scheme. It involves averaging representations of the same bins within a mini-batch to form centroids and "re-filling" missing bins by taking corresponding centroids from the + +previous epoch. This process results in a surrogate containing centroids for all bins, enabling the effective application of geometric loss across the complete label range. Furthermore, we introduce Imbalanced Operator Learning (IOL) as a new DIR benchmark for training models on imbalanced domain locations in function space mapping. In summary, our main contributions are four-fold: + +- Geometric Losses for Uniformity in deep imbalanced regression (DIR). To the best of our knowledge, this work is the first to study representation learning in DIR. We introduce two novel loss functions, enveloping loss and homogeneity loss, to ensure uniform feature distribution for DIR. +- SRL Framework. A new framework is proposed that incorporates these geometric principles into data representations. +- Imbalanced Operator Learning (IOL). For the first time, we pioneer the task of operator learning within the realm of deep imbalanced regression, introducing an innovative task: Imbalanced Operator Learning (IOL). +- Extensive Experiments. The effectiveness of the proposed method is validated through experiments involving real-world regression and operator learning, on five datasets: AgeDB-DIR, IMDB-WIKI-DIR, STS-B-DIR, and two newly created DIR benchmarks, UCI-DIR and OL-DIR. + +# 2. Related Work + +Uniformity in imbalanced classification. Wang and Isola [22] identifies that uniformity is one of the key properties in contrastive representation learning. To promote uniformity in representation space for imbalanced classification, a variety of training strategies have been proposed. Kang et al. [8] + +decouples the training into a two-stage training of representation learning and classification. Yin et al. [26] designs a transfer learning framework for imbalanced face recognition. Kang et al. [9] combines supervised method and contrastive learning to learn a discriminative and balanced feature space. PaCo [5] and TSC [12] learn a set of class-wise balanced centers. BCL [33] balances the gradient distribution of negative classes and data distribution in mini-batch. Recent study suggests that sample-level uniform distribution may not effectively address imbalanced classification, advocating for category-level uniformity instead [1, 32]. Though progress has been made in this field, challenges persist in adapting the approach of modeling uniformity from classification to regression. + +Deep imbalanced regression. With imbalanced regression data, effective learning in a continuous label space focuses on modeling the relationships among labels in the feature space [25]. Label Distribution Smoothing (LDS) [25] and DenseLoss [19] apply a Gaussian kernel to the observed label density, leading to an estimated label density distribution. Feature distribution smoothing (FDS) [25] generalizes the application of kernel smoothing from label to feature space. Ranksim [6] aims to leverage both local and global dependencies in data by aligning the order of similarities between labels and features. Balanced MSE [16] addresses the issue of imbalance in Mean Squared Error (MSE) calculations, ensuring a more balanced distribution of predictions. VIR [23] provides uncertainty for imbalanced regression. ConR [10] regularizes contrastive learning in regression by modeling global and local label similarities in feature space. RNC [27] and SupReMix [24] learn a continuous and ordered representation for regression through supervised contrastive learning. How imbalanced regression representations leverage the feature space remains under-explored. + +# 3. Method + +In the field of representation learning for classification, the concept of uniformity is pivotal for maximizing the use of the feature space [5, 8, 9, 12, 22, 26, 33]. This idea is based on the principle of ensuring that features from different classes are not only distinctly separated but also evenly distributed in the latent space. This uniform distribution of class centroids fosters a clear and effective decision boundary, leading to more accurate classification. However, in regression, where we deal with continuous, ordered trace [24, 27] rather than discrete clusters, the concept of uniformity is not only more complex but essentially remains undefined. + +We draw an analogy to the process of winding yarn around a ball. In this analogy, the yarn represents the latent trace, and the ball symbolizes the entirety of the available feature space. Just as the yarn must be evenly distributed across the ball's surface to effectively cover it (without any crossing), the latent trace should strive to occupy the hypersphere of the latent space uniformly. This ensures that the model leverages the available feature space to its fullest extent, enhancing the model's ability + +![](images/4cf76af194660921134e539fab8b701f1dda0cde8ee50395c6e97068b76a9206.jpg) +Figure 3. 2D schematic overview of two geometric losses. The arrow indicates the improvement of the loss function. Enveloping loss encourages the representations to fill the latent space, and homogeneity loss encourages the smoothness and even distribution of the representations along the trace. + +![](images/b34ec2e285c6444532d57a6144820103b4c031999efb5d48189daa7136892135.jpg) + +to capture the variability inherent in the data. + +Furthermore, the latent trace should be smooth and continuous, akin to the even stretching of yarn, rather than loose and disjointed. This smoothness ensures a consistent and predictable model behavior, which is crucial for the accurate prediction and interpretation of results. + +We outline our method in this section. Firstly, we establish the fundamental notations and preliminaries (Section 3.1). Following this, we delve into the concept and definition of our enveloping loss (Section 3.2) and homogeneity loss (Section 3.3). Finally, we present our Surrogate-driven Representation Learning (SRL) framework, which incorporates the geometric constraints from the global image of the representations into the local range (Section 3.4). Refer to Supplementary Material 9 for the pseudo code of our method. + +# 3.1. Preliminaries + +A regression dataset is composed of pairs $(\mathbf{x}_i, y_i)$ , where $\mathbf{x}_i$ represents the input and $y_i$ is the corresponding continuous-value target. Denote $\mathbf{z}_i = f(\mathbf{x}_i)$ as the feature representation of $\mathbf{x}_i$ , generated by a neural network $f(\cdot)$ . The feature representation is normalized so that $\|\mathbf{z}_i\| = 1$ for all $i$ . Suppose the dataset consists of $K$ unique bins1, we define a surrogate as a set of centroids $\mathbf{c}_k$ , where each represents a distinct bin. These centroids are computed by averaging the representations $\mathbf{z}$ sharing the same bin and they are normalized to $\|\mathbf{c}_k\| = 1$ . Let $l$ be a path: $l : [y_{\min}, y_{\max}] \mapsto \mathbb{R}^n$ with $\|l(y)\| = 1$ , such that $l(y_k) = \mathbf{c}_k$ . The path $l$ is a continuous curve extended from the discrete dataset that lies on a submanifold of $\mathbb{R}^n$ . + +# 3.2. Enveloping + +To maximize the use of the feature space, it is crucial for the ordered and continuous trace of regression representations to fill the entire unit hypersphere as much as possible. This is by + +analogy with wrapping yarn with a certain length around a ball (without any crossing), aiming to cover as much surface area as possible. + +The trace of regression representations lying on a submanifold of $\mathbb{R}^n$ has a negligible hypervolume, which makes it challenging to assess its relationship with the entire hypersphere. To address this challenge, we extend the "line" into a tubular neighborhood. This expansion allows us to introduce the concept of enveloping loss. Our objective with this loss function is to maximize the hypervolume of the tubular neighborhood in proportion to the total hypervolume of the hypersphere. + +Denote the set of all unit vectors in $\mathbb{R}^n$ as $\mathcal{U}$ . Given $\epsilon \in (0,1)$ , define tubular neighborhood $T(l,\epsilon)$ of $l$ as: + +$$ +T (l, \epsilon) = \left\{\mathbf {z} \in \mathcal {U} \mid \mathbf {t} \cdot \mathbf {z} > \epsilon \text {f o r s o m e} \mathbf {t} \in \operatorname {I m} (l) \right\} \tag {1} +$$ + +where for a function $f: A \to B$ , the image is defined as $\operatorname{Im}(f) \coloneqq \{f(x), x \in A\}$ . + +Then our enveloping loss is defined as: + +$$ +\mathcal {L} _ {\text {e n v}} = - \frac {\operatorname {v o l} (T (l , \epsilon))}{\operatorname {v o l} (\mathcal {U})} \tag {2} +$$ + +where $\mathrm{vol}(\cdot)$ returns the hypervolume of its input in the induced measure from the Euclidean space. + +In practical scenarios, the trace is composed of discrete representations, which complicates the direct computation of the tubular neighborhood's hypervolume. To navigate this challenge, we propose a continuous-to-discrete strategy. We first generate $N$ points that are uniformly distributed across the hypersphere. We then determine the fraction of these points that fall within the neighbourhood $\epsilon$ . This fraction effectively approximates the proportion of the hypersphere covered by the tubular neighborhood with a sufficiently large $N$ . To adapt $\mathcal{L}_{\mathrm{env}}$ to discrete datasets, we re-formalize our optimization objective as: + +$$ +\max \lim _ {N \rightarrow \infty} \frac {P (N)}{N} \tag {3} +$$ + +where + +$$ +P (N) := \left| \left\{\mathbf {p} _ {i} \mid \max _ {y} \left\{\mathbf {p} _ {i} \cdot l (y) \right\} > \epsilon , i \in [ N ] \right\} \right| \tag {4} +$$ + +assuming for each $N > 0$ , we can choose $N$ evenly distributed points in $\mathcal{U}$ , and denote these points as $\mathbf{p}_i, i \in [N] = 1, \dots, N$ . For numerical application, we take $N$ to be a sufficiently large number and use the standard Monte-Carlo method [17] to approximate the evenly distributed points. + +In our implementation, we did not directly define $\epsilon$ due to the non-differentiability of the binarization required to determine if a $\mathbf{p}_i$ is within the $\epsilon$ -tube. Instead, for each $\mathbf{p}_i$ , we maximize the cosine similarity between $\mathbf{p}_i$ and its closest point on the trace. In this way, we relax the step function represented by (4) to its "soft" version, leading to smooth gradient computation. + +# 3.3. Homogeneity + +While the enveloping loss effectively governs the overall distribution of representations on the hypersphere, it alone may not be entirely adequate, presenting two unresolved issues. 1) The first is distribution along the trace. The enveloping loss predominantly controls the overall shape of representations on the hypersphere, yet it does not guarantee a uniform distribution along the trace. This poses a notable concern, as it may result in uneven representation density across different trace segments. 2) The second is trace smoothness. The enveloping loss could lead to a zigzag pattern of the representations, which should be avoided. Considering age estimation from facial images as an example, the progression of facial features over time is gradual. Consequently, in the corresponding latent space, we would anticipate a similar, smooth transition without abrupt changes, underlining the desirability of a smoother trace. Interestingly, these two issues can be aptly analogized to winding yarns around a ball as well. For the yarn on the ball to be smooth, it should be tightly stretched, rather than being disjointed or loosely arranged. We name the property of a trace to be smooth with representations evenly distributed along it as homogeneity. + +We encourage such homogeneity property, i.e., smoothness of the trace $\operatorname{Im}(l)$ and uniform distribution of representations along it, by penalizing the arc length. Formally, the homogeneity loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {h o m o}} = \int_ {y _ {\min }} ^ {y _ {\max }} \left\| \frac {\mathrm {d} l (y)}{\mathrm {d} y} \right\| ^ {2} \mathrm {d} y \tag {5} +$$ + +Given $K$ different $y\mathrm{s}$ which have been ordered, the discrete format for $\mathcal{L}_{\mathrm{homo}}$ is defined as a summation of the squared differences between adjacent points: + +$$ +\mathcal {L} _ {\mathrm {h o m o}} = \sum_ {k = 1} ^ {K - 1} \frac {\left\| l \left(y _ {k + 1}\right) - l \left(y _ {k}\right) \right\| ^ {2}}{y _ {k + 1} - y _ {k}} \tag {6} +$$ + +The use of only homogeneity loss might result in trivial solutions like representation convergence to a circle or point due to feature collapse (shown in Figure 2). The homogeneity loss should be treated as a regularization of the enveloping loss, promoting not only smoothness but also an even distribution of representations along the trace. To quantitatively define the relationship between trace arc length and these desired characteristics, we introduce Theorem 1. It demonstrates that with a given $\operatorname{Im}(l)$ ( $l_{\mathrm{env}}$ is fixed as it does not depend on the parameterization of $l$ ), the homogeneity loss is minimized if and only if when representations are uniformly distributed along the trace. + +Theorem 1. Given an image of $l$ , $\mathcal{L}_{\text{homo}}$ attains its minimum if and only if the representations are uniformly distributed along the trace, i.e., $\| \nabla_y l(y) \| = c$ , where $c$ is a constant. + +Refer to Supplementary Material 6 for the proof. + +Therefore, we formulate our geometric constraints $(\mathcal{L}_{\mathrm{G}})$ as a combination of enveloping and homogeneity: + +![](images/1180f604008d7fc4afc0a3c149cc193a1008316072c6d0f88fd7743b3016aec1.jpg) +Figure 4. Overview of Surrogate-driven Representation Learning (SRL). (1) Every mini-batch is encoded to the latent space. Some bins may not be present in the current batch. To address this, (2) it takes centroids corresponding to the missing bins from the previous epoch. These stored centroids are used to "re-fill" the missing bins in the current batch. (3) Average the representations for bins that appear multiple times, creating centroids for these bins. This surrogate, containing a representation for the full label range, allows for the effective application of geometric loss across all bins. (4) Loss calculation based on the surrogate. (5) Update the surrogate in memory to ensure enveloping and homogeneity. The training of the first epoch is driven by MSE loss only. + +$$ +\mathcal {L} _ {\mathrm {G}} = \lambda_ {e} \mathcal {L} _ {\text {e n v}} + \lambda_ {h} \mathcal {L} _ {\text {h o m o}} \tag {7} +$$ + +where $\lambda_{e}$ and $\lambda_{h}$ are weights for the two geometric losses. In Section 4.4, we further explore the behavior of these two geometric constraints, uncovering new insights into imbalanced regression. + +# 3.4. Surrogate-driven Representation Learning (SRL) + +Our geometric loss $(\mathcal{L}_{\mathrm{G}})$ is calculated on a surrogate instead of a mini-batch (Figure 4), as the representations from one mini-batch very likely fail to capture the global image of $l$ , due to the randomness of batch sampling. + +For illustration purposes, here we assume the original dataset has already been binned, as is the case in most DIR datasets [6, 10, 25]. Let $\mathcal{Z} = \{\mathbf{z}_1,\mathbf{z}_2,\dots,\mathbf{z}_M\}$ be a set of representations from a batch with batch-size $M$ , and let $\mathcal{Y} = \{y_1,y_2,\dots,y_M\}$ (repetitions of the label values might exist) be the corresponding labels. Define the centroid $\mathbf{c}_y\in \mathcal{C}$ for $y$ ( $\| \mathbf{c}_y\| = 1$ ) as: + +$$ +\mathbf {c} _ {y} = \frac {1}{| \left\{\mathbf {z} _ {m} \mid y _ {m} = y \right\} |} \sum_ {\left\{\mathbf {z} _ {m} \mid y _ {m} = y \right\}} \mathbf {z} _ {m} \tag {8} +$$ + +Suppose the whole label range is covered by a set of unique $K$ bins $\mathcal{V}^{*} = \{y_{1}^{*},y_{2}^{*},\dots,y_{K}^{*}\}$ . The centroids for these $K$ bins from the last epoch are denoted as $\mathcal{C}' = \{\mathbf{c}_{y_1^*}',\mathbf{c}_{y_2^*}',\dots,\mathbf{c}_{y_K^*}'\}$ . + +The surrogate $S$ is then generated by re-filling the missing centroids: + +$$ +\mathcal {S} = \mathcal {C} \cup \left\{\mathrm {c} ^ {\prime} _ {y _ {k} ^ {*}} \mid y _ {k} ^ {*} \in \mathcal {Y} ^ {*} \backslash \mathcal {Y} \right\} \tag {9} +$$ + +We use AdamW [13] with momentum to update parameters $\theta$ in $f(\cdot)$ , to ensure a smooth transition of the local shape in the batch-wise representations. + +At the end of each epoch $e \in (0, E]$ (excluding the first), we use the representations learned during that epoch to form a running surrogate $\hat{S}^e$ . $S^{e+1}$ is formulated from the current epoch's surrogate $S^e$ and $\hat{S}^e$ with momentum. $S^{e+1}$ is employed for training in the subsequent epoch. It facilitates a gradual transition between epochs, preventing abrupt variations: $S^{e+1} \gets \alpha \cdot S^e + (1 - \alpha) \cdot \hat{S}^e$ . + +We aim for individual representations from the encoder to converge towards their respective centroids that share the same label, while simultaneously distancing them from centroids associated with different labels. To achieve this, we incorporate a contrastive loss between the individual representations and the centroids. For each representation $\mathbf{z}_m$ with $y_{m} = y$ , the centroid $\mathbf{c}_y$ is considered as the positive, and the other centroids as negatives. The contrastive loss is defined as: + +$$ +\mathcal {L} _ {\text {c o n}} = - \sum_ {m = 1} ^ {M} \log \frac {\exp \left(\operatorname {s i m} \left(\mathbf {z} _ {m} , \mathbf {c} _ {y}\right)\right)}{\sum_ {y ^ {*} \in \mathcal {Y} ^ {*}} \exp \left(\operatorname {s i m} \left(\mathbf {z} _ {m} , \mathbf {c} _ {y ^ {*}}\right)\right)} \tag {10} +$$ + +where $\mathrm{sim}(\cdot)$ is the cosine similarity between two input. + +The framework is trained end-to-end, the total loss used to update the parameters $\theta$ in $f(\cdot)$ is defined as: + +$$ +\mathcal {L} _ {\theta} = \mathcal {L} _ {\text {r e g}} + \mathcal {L} _ {\mathrm {G}} + \mathcal {L} _ {\text {c o n}} \tag {11} +$$ + +where $\mathcal{L}_{\mathrm{reg}}$ is the mean squared error (MSE) loss. + +# 4. Experiments + +We perform extensive experiments to validate and analyze the effectiveness of SRL for deep imbalanced regression. Our regression tasks span age estimation from facial images, tabular regression, and text similarity score regression, as well as our newly established task: Imbalanced Operator Learning (IOL). This section begins by detailing the experiment setup (Section 4.1) followed by the main results (Section 4.2). The results of IOL are shown in Section 4.3 followed by the comparison with classification-based methods and hyperparameters analysis (Section 4.4). + +# 4.1. Experiment Setup + +Datasets. We employ three real-world regression datasets developed by Yang et al. [25], and our curated UCI-DIR from UCI Machine Learning Repository [2], to assess the effectiveness of SRL in deep imbalanced regression. Refer to Supplementary Material 7 for more dataset details. + +Table 1. Results on UCI-DIR (MAE). We report the average MAE of three runs. The best results are in bold. + +
DatasetsAirfoilAbaloneReal EstateConcrete
ShotAllManyMedFewAllManyMedFewAllManyMedFewAllManyMedFew
VANILLA5.665.115.036.754.570.882.657.970.330.270.380.377.295.776.929.74
LDS + FDS [25]5.764.454.797.795.090.903.269.260.350.330.400.346.886.216.737.59
RankSim [6]5.235.054.915.724.330.982.597.420.370.340.380.406.716.005.579.46
BalancedMSE [16]5.694.515.047.285.372.142.669.370.340.310.400.337.034.676.379.72
Ordinal Entropy [29]6.274.855.378.326.772.314.0111.610.340.290.420.357.125.506.369.31
SRL (ours)5.104.834.755.694.160.892.427.190.280.260.300.295.945.325.806.60
+ +Table 2. Results on AgeDB-DIR, the best are in bold. + +
MetricsMAE ↓GM ↓
ShotAllManyMedFewAllManyMedFew
VANILLA7.676.669.3012.614.854.176.518.98
LDS + FDS [25]7.557.038.4610.524.864.575.386.75
RankSim [6]7.416.498.7312.474.714.155.748.92
BalancedMSE [16]7.987.588.659.935.014.835.466.30
Ordinal Entropy [29]7.606.698.8712.684.914.286.209.29
ConR [10]7.416.518.8112.044.704.135.918.59
SRL (ours)7.226.648.289.814.504.125.376.29
+ +Table 3. Results on IMDB-WIKI-DIR, the best are in bold. + +
MetricsMAE ↓GM ↓
ShotAllManyMedFewAllManyMedFew
VANILLA8.037.1615.4826.114.544.1410.8418.64
LDS + FDS [25]7.737.2212.9823.714.404.177.8715.77
RankSim [6]7.726.9214.5225.894.293.929.7218.02
BalancedMSE [16]8.437.8413.3523.274.934.687.9015.51
Ordinal Entropy [29]8.017.1715.1526.484.474.0710.5621.11
ConR [10]7.847.1514.3625.154.434.059.9118.55
SRL (ours)7.697.0812.6522.784.284.037.2815.25
+ +Table 4. Results on STS-B-DIR, the best are in bold. + +
MetricsMSE ↓Pearson correlation ↑
ShotAllManyMedFewAllManyMedFew
VANILLA0.9930.9631.0001.0750.7420.6850.6930.793
LDS + FDS [25]0.9000.9110.8810.9050.7570.6980.7230.806
RankSim [6]0.8890.9070.8740.7570.7630.7080.6920.842
BalancedMSE [16]0.9090.8941.0040.8090.7570.7030.6850.831
Ordinal Entropy [29]0.9430.9021.1610.8120.7500.7020.6790.767
SRL (ours)0.8770.8860.8730.7450.7650.7080.7490.844
+ +- AgeDB-DIR [25]: It serves as a benchmark for estimating age from facial images, which is derived from the AgeDB dataset [15]. It contains 12,208 images for training, 2,140 images for validation, and 2,140 images for testing. +- IMDB-WIKI-DIR [25]: It is a facial image dataset for age estimation derived from the IMDB-WIKI dataset [18], which consists of face images with the corresponding age. It has 191,509 images for training, 11,022 images for validation, and 11,022 for testing. +- STS-B-DIR [25]: It is a natural language dataset formulated from STS-B dataset [3, 21], consisting of 5,249 training + +sentence pairs, 1,000 validation pairs, and 1,000 testing pairs. Each sentence pair is labeled with the continuous similarity score. + +- UCI-DIR: To evaluate the performance of SRL on tabular data, we curated UCI Machine Learning Repository [2] to formulate UCI-DIR that includes four regression tasks (Airfoil Self-Noise, Abalone, Concrete Compressive Strength, Real estate valuation). Following the DIR setting [25], we make each regression task consist of an imbalanced training set and a balanced validation and test set. + +Metrics. In line with the established settings in DIR [25], subsets in an imbalanced training set are categorized based on the number of available training samples: many-shot region (bins with $>100$ training samples), medium-shot region (bins with 20 to 100 training samples), and few-shot region (bins with $<20$ training samples), for the three real-world datasets. For AgeDB-DIR and IMDB-WIKI-DIR, each bin represents 1 year. In the case of STS-B-DIR, bins are segmented by 0.1. For UCI-DIR, the bins are segmented by 0.1 to 1 depending on the range of regression targets. Our evaluation metrics include mean absolute error (MAE, the lower the better) and geometric mean (GM, the lower the better) for AgeDB-DIR, IMDB-WIKI-DIR and UCI-DIR. For STS-B-DIR, we use mean squared error (MSE, the lower the better) and Pearson correlation (the higher the better). + +Implementation Details. For age estimation (AgeDB-DIR and IMDB-WIKI-DIR), we follow the settings from Yang et al. [25], which uses ResNet-50 [7] as a backbone network. For text similarity regression (STS-B-DIR), we follow the setting from Cer et al. [3], Yang et al. [25] that uses BiLSTM + GloVe word embeddings. For tabular regression (UCI-DIR), we use an MLP with three hidden layers (d-20-30-10-1) following the setting from Cheng et al. [4]. For all baseline methods, results were produced following provided training recipes through publicly available codebase. All experimental results, including ours and baseline methods, were obtained from a server with 8 RTX 3090 GPUs. + +Baselines. We consider both DIR methods [6, 10, 25] and recent techniques proposed for general regression [16, 28, 29], in addition to VANILLA regression (MSE loss). We compare the performance of SRL with all baselines on the above four datasets. Furthermore, as SRL is orthogonal to previous DIR methods, we examine the improvement of them by adding our geometric losses. + +![](images/a6304d257fc99b532b1d859fd23ec583a770306ad76766c3b40e135d9d82ebbd.jpg) +Figure 5. SRL performance gain compared to VANILLA across age ranges on AgeDB-DIR. The gray histogram in the background shows the distribution of samples across age groups. SRL substantially improves the performance on the medium-shot and few-shot regions (age $< 20$ and $>70$ ). + +Table 5. Combine SRL with existing DIR methods (MAE) + +
DatasetsAgeDB (MAE)IMDB-WIKI (MAE)
ShotAllManyMedFewAllManyMedFew
SRL+LDS+FDS [25]7.326.818.149.817.617.0312.2821.77
GAINS v.s. LDS+FDS (%)3.053.233.896.751.662.645.408.19
SRL+RankSim [6]7.296.578.5810.487.677.0812.4022.85
GAINS v.s. RankSim (%)1.62-1.231.7216.960.65-1.1514.6111.75
SRL+BalancedMSE [16]7.246.777.869.857.747.1312.7722.04
GAINS v.s. BalancedMSE (%)9.2710.699.140.898.199.064.355.29
SRL+ConR [10]7.406.878.0810.507.567.0112.0321.71
GAINS v.s. ConR (%)0.14-5.538.3913.803.681.9616.2313.68
+ +# 4.2. Main Results + +To show the effectiveness of SRL on DIR, we first benchmark SRL and baselines for tabular regression on our curated UCIDIR with four different regression tasks (Table 1). Moreover, we evaluate our method on established DIR benchmarks [25] including age estimation on AgeDB-DIR and IMDB-WIKI-DIR (Table 2 & 3, and Figure 5), and text similarity regression on STS-B-DIR (Table 4). We evaluate the combination of SRL and previous DIR methods on AgeDB-DIR and IMDB-WIKI-DIR (Table 5). Notably, Table 1 and 4 omit results from ConR [10], as it depends on data augmentation, a technique not fully established in the domain of tabular data and natural language. + +Combine SRL with existing methods. Our SRL approach enhances imbalanced regression by imposing geometric constraints on feature distributions, a strategy that is orthogonal to existing methods. To illustrate this, we leverage SRL as a regularizing term in conjunction with other methods. The results of this experiment are presented in Table 5. It shows that when SRL is integrated with existing regression methods, there is improvement in performance across different regions for both datasets. This demonstrates the effectiveness and compatibility of SRL as a complementary tool in the realm of regression analysis. + +# 4.3. Imbalanced Operator Learning (IOL) + +We introduce a novel task for DIR called Imbalanced Operator Learning (IOL). Traditional operator learning aims to train a neural network to model the mapping between function spaces [11, 14]. However, unlike the standard approach of uniformly + +sampling output locations, in IOL, we intentionally adjust the sampling density within the output function's domain to create regions with few, medium, and many regions (Figure 6). + +For the linear operator, the model is trained to estimate the integral operator denoted as $G$ : + +$$ +G: u (x) \mapsto s (x) = \int_ {0} ^ {x} u (\tau) d \tau , x \in [ 0, 1 ] \tag {12} +$$ + +where $u$ denotes the input function which is sampled from a Gaussian random field (GRF), and $s$ is the target function. + +For the nonlinear operator, the model is trained to learn a particular stochastic partial differential equation (PDE): + +$$ +\operatorname {d i v} \left(e ^ {b (x; \omega)} \nabla u (x; \omega)\right) = f (x), x \in [ 0, 1 ] \tag {13} +$$ + +where $e^{b(x;\omega)}$ is the diffusion efficient and $u(x;\omega)$ is the target function. + +Denote the domain of output function as $y$ . For both linear and non-linear operator learning, we changed the original uniform sampling of $y$ to three curated regions: few/medium/many. Afterward, we manually created an imbalanced training set of 10k samples and a balanced testing test of 100k samples, namely OL-DIR. + +In Figure 6, we have a schematic overview of Imbalanced Operator Learning (IOL). The network is trained to model an integral operator $G$ . The data provided to the model is $([u,y],G(u)(y))$ . The input consists of function $u$ and sampled $y$ s from the domain of $G(u)$ . The target is $G(u)(y)$ . We manipulate the distribution density of $y$ across its range to formulate few/med./many regions. Here the imbalance comes from the unequal exposure of integral interval to the model training. Refer to Supplementary Material 7.2 for more details. + +Shown in Table 6, SRL consistently outperforms VANILLA and the state-of-the-art operator learning for the whole label range including all, many-shot, medium-shot, and few-shot regions. The results position SRL as the superior approach for IOL in terms of accuracy and generalizability. + +# 4.4. Further Analysis + +Quantification of geometric impact. We further quantified the impact of geometric constraints by comparing percentages of + +![](images/5400e1ef48185ad514b107f5de4074bb2d673bae22adb6428663f47894e962f3.jpg) +Figure 6. Imbalanced Operator Learning. + +Table 6. Results on OL-DIR. We report the average MAE of ten runs. The best results are bold. + +
OperationMAE(10-3) ↓MSE (10-4) ↓
ShotAllManyMedFewAllManyMedFew
Linear
VANILLA [14]15.6411.8615.4527.005.402.814.4014.20
Ordinal Entropy [29]10.079.269.8513.012.001.531.893.42
SRL (ours)9.188.329.479.331.980.981.722.67
Nonlinear
VANILLA [14]11.649.8911.0219.779.204.337.5324.70
Ordinal Entropy [29]12.919.9313.0721.0213.808.8211.8430.12
SRL (ours)11.259.489.2217.008.607.426.4114.12
+ +Full results with standard deviation are reported in Supplementary Material 8.10. + +uniformly sampled points within few-shot regions (a measure of proportion). The results show our method significantly increases few-shot proportion (AgeDB-DIR: $1.98\% \rightarrow 15.80\%$ , upper bound: $23\%$ ; STS-B-DIR: $4.52\% \rightarrow 22.39\%$ , upper bound: $38\%$ ), leading to improved performance (Table 7). + +Table 7. Impact of geometric constraints on few-shot proportion. + +
Few-shotOverall
ProportionMAEMAE
AgeDB-DIR (1.10% samples, 23% label range):
VANILLA1.98%12.617.67
LDS + FDS [25]4.95%10.527.55
Ours15.80%9.817.22
STS-B-DIR (3.49% samples, 38% label range):
VANILLA4.52%1.0750.993
LDS + FDS [25]8.13%0.9050.900
Ours22.39%0.8770.745
+ +Balancing of enveloping and homogeneity: Our proposed SRL advocates for two pivotal geometric constraints in feature distribution: enveloping and homogeneity, to effectively address imbalanced regression. These two losses are modulated by their respective coefficients, $\lambda_{e}$ for the enveloping loss and $\lambda_{h}$ for the homogeneity loss. Figure 7 illustrates that the omission of either constraint detrimentally impacts the performance, highlighting the importance of both of them, and it demonstrates that the best + +performance, as measured by MAE on the AgeDB-DIR dataset, is achieved when both coefficients $\lambda_{e}$ and $\lambda_{h}$ are set to $1e^{-1}$ . + +Ablation studies on choices of $N$ : Table 10 (in Supplementary Material) shows that achieving optimal performance on the AgeDB-DIR and IMDB-WIKI-DIR datasets requires a sufficiently large $N$ , as a smaller $N$ may lead to imprecise calculation of the enveloping loss. + +Ablation studies on proposed loss component: Table 11 (in Supp.) demonstrates that incorporating homogeneity, enveloping, and contrastive loss term yields superior model performance compared to using each individually. + +Computational cost: As shown in Table 12 (in Supp.), the computational overhead introduced by the Surrogate-driven Representation Learning (SRL) framework is comparable to that of other imbalanced regression methods. + +Impact of bin numbers: As shown in Table 13 (in Supp.), while increasing the number of bins generally leads to better model performance, the improvements become marginal beyond certain thresholds. + +Limitations: Section 11 (in Supp.) examines the limitations of SRL, including its inability to handle higher-dimensional labels. + +![](images/9a177d605c7c679d6f50dbf76a7eabe126272c52ff8f9cc41bea698ffedce2ab.jpg) +Figure 7. Confusion matrix of MAE on AgeDB-DIR from different values of $\lambda_h$ and $\lambda_e$ . + +# 5. Conclusion + +As the first work of exploring uniformity in deep imbalanced regression, we introduce two novel loss functions - enveloping and homogeneity loss - to encourage the uniform feature distribution of an ordered and continuous trajectory. The two loss functions serve as geometric constraints which are integrated into the data representations through a Surrogate-driven Representation Learning (SRL) framework. Furthermore, we set a new benchmark in imbalanced regression: Imbalanced Operator Learning (IOL). Extensive experiments on real-world regression and operator learning demonstrate the effectiveness of our geometrically informed approach. We emphasize the significance of uniform data representation and its impact on learning performance in imbalanced regression scenarios, advocating for a more balanced and comprehensive utilization of feature spaces in regression models. + +# References + +[1] Mido Assran, Randall Balestriero, Quentin Duval, Florian Bordes, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, and Nicolas Ballas. The hidden uniform cluster prior in self-supervised learning. In *The Eleventh International Conference on Learning Representations*, 2022. 3 +[2] Arthur Asuncion and David Newman. Uci machine learning repository, 2007. 1, 2, 5, 6 +[3] Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. arXiv preprint arXiv:1708.00055, 2017. 6 +[4] Xin Cheng, Yuzhou Cao, Ximing Li, Bo An, and Lei Feng. Weakly supervised regression with interval targets. arXiv preprint arXiv:2306.10458, 2023. 6 +[5] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021. 1, 3 +[6] Yu Gong, Greg Mori, and Fred Tung. Ranksim: Ranking similarity regularization for deep imbalanced regression. In International Conference on Machine Learning, pages 7634-7649. PMLR, 2022. 1, 3, 5, 6, 7, 4 +[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[8] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In International Conference on Learning Representations, 2019. 1, 2, 3 +[9] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2020. 1, 3 +[10] Mahsa Keramati, Lili Meng, and R David Evans. Contr: Contrastive regularizer for deep imbalanced regression. arXiv preprint arXiv:2309.06651, 2023. 1, 3, 5, 6, 7, 4 +[11] Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research, 24(89):1-97, 2023. 7 +[12] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6928, 2022. 1, 3 +[13] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 5 +[14] Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deep-onet based on the universal approximation theorem of operators. Nature machine intelligence, 3(3):218-229, 2021. 7, 8, 1 +[15] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. + +In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 6 +[16] Jiawei Ren, Mingyuan Zhang, Cunjun Yu, and Ziwei Liu. Balancedmse for imbalanced visual regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7926-7935, 2022. 1, 3, 6, 7 +[17] Christian P Robert, George Casella, and George Casella. Monte Carlo statistical methods. Springer, 1999. 4 +[18] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Deep expectation of real and apparent age from a single image without facial landmarks. International Journal of Computer Vision, 126 (2-4):144-157, 2018. 6 +[19] Michael Steininger, Konstantin Kobs, Padraig Davidson, Anna Krause, and Andreas Hotho. Density-based weighting for imbalanced regression. Machine Learning, 110:2187-2211, 2021. 1, 3 +[20] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 2 +[21] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 6 +[22] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 1, 2, 3 +[23] Ziyan Wang and Hao Wang. Variational imbalanced regression: Fair uncertainty quantification via probabilistic smoothing. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3 +[24] Yilei Wu, Zijian Dong, Chongyao Chen, Wangchunshu Zhou, and Juan Helen Zhou. Mixup your own pairs. arXiv preprint arXiv:2309.16633, 2023. 1, 3 +[25] Yuzhe Yang, Kaiwen Zha, Yingcong Chen, Hao Wang, and Dina Katabi. Delving into deep imbalanced regression. In International Conference on Machine Learning, pages 11842-11851. PMLR, 2021. 1, 3, 5, 6, 7, 8, 2 +[26] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5704-5713, 2019. 1, 3 +[27] Kaiwen Zha, Peng Cao, Jeany Son, Yuzhe Yang, and Dina Katabi. Rank-n-contrast: Learning continuous representations for regression. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 3 +[28] Shihao Zhang, Linlin Yang, Michael Bi Mi, Xiaoxu Zheng, and Angela Yao. Improving deep regression with ordinal entropy. In The Eleventh International Conference on Learning Representations, 2022. 6 +[29] Shihao Zhang, Linlin Yang, Michael Bi Mi, Xiaoxu Zheng, and Angela Yao. Improving deep regression with ordinal entropy. arXiv preprint arXiv:2301.08915, 2023. 6, 8 +[30] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[31] Zhisheng Zhong, Jiequan Cui, Yibo Yang, Xiaoyang Wu, Xiaojuan Qi, Xiangyu Zhang, and Jiaya Jia. Understanding + +imbalanced semantic segmentation through neural collapse. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19550-19560, 2023. 1 +[32] Zhihan Zhou, Jiangchao Yao, Feng Hong, Ya Zhang, Bo Han, and Yanfeng Wang. Combating representation learning disparity with geometric harmonization. Advances in Neural Information Processing Systems, 36, 2024. 3 +[33] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022. 1, 3 \ No newline at end of file diff --git a/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/images.zip b/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2595add941c924b38b5847cb00651953ef455aa4 --- /dev/null +++ b/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ec17189f24dc04fda6fa81d4614f493785e6a7ff7c4f4e68ed0a48667ab9f76 +size 603761 diff --git a/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/layout.json b/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4fa3787d08bc78f1cf564d1b9d58de7b8bedfdf2 --- /dev/null +++ b/2025/Improve Representation for Imbalanced Regression through Geometric Constraints/layout.json @@ -0,0 +1,9262 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 103, + 459, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 103, + 459, + 138 + ], + "spans": [ + { + "bbox": [ + 153, + 103, + 459, + 138 + ], + "type": "text", + "content": "Improve Representation for Imbalanced Regression through Geometric Constraints" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "spans": [ + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "type": "text", + "content": "Zijian Dong" + }, + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "type": "text", + "content": " Yilei Wu" + }, + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "type": "text", + "content": " Chongyao Chen" + }, + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "type": "text", + "content": " Yingtian Zou" + }, + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "type": "text", + "content": " Yichi Zhang" + }, + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "type": "text", + "content": " Juan Helen Zhou" + }, + { + "bbox": [ + 80, + 160, + 533, + 177 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 145, + 182, + 466, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 182, + 466, + 198 + ], + "spans": [ + { + "bbox": [ + 145, + 182, + 466, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 182, + 466, + 198 + ], + "type": "text", + "content": "National University of Singapore, Singapore " + }, + { + "bbox": [ + 145, + 182, + 466, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 182, + 466, + 198 + ], + "type": "text", + "content": " Duke University, US" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 205, + 498, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 205, + 498, + 221 + ], + "spans": [ + { + "bbox": [ + 111, + 205, + 498, + 221 + ], + "type": "text", + "content": "{zijian.dong, yilei.wu}@u.nus.edu, helen.zhou@nus.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 153, + 247, + 199, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 247, + 199, + 260 + ], + "spans": [ + { + "bbox": [ + 153, + 247, + 199, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 271, + 297, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 297, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 297, + 523 + ], + "type": "text", + "content": "In representation learning, uniformity refers to the uniform feature distribution in the latent space (i.e., unit hypersphere). Previous work has shown that improving uniformity contributes to the learning of under-represented classes. However, most of the previous work focused on classification; the representation space of imbalanced regression remains unexplored. Classification-based methods are not suitable for regression tasks because they cluster features into distinct groups without considering the continuous and ordered nature essential for regression. In a geometric aspect, we uniquely focus on ensuring uniformity in the latent space for imbalanced regression through two key losses: enveloping and homogeneity. The enveloping loss encourages the induced trace to uniformly occupy the surface of a hypersphere, while the homogeneity loss ensures smoothness, with representations evenly spaced at consistent intervals. Our method integrates these geometric principles into the data representations via a Surrogate-driven Representation Learning (SRL) framework. Experiments with real-world regression and operator learning tasks highlight the importance of uniformity in imbalanced regression and validate the efficacy of our geometry-based loss functions. Code is available here." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 546, + 132, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 132, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 132, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 566, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 297, + 687 + ], + "type": "text", + "content": "Imbalanced datasets are ubiquitous across various domains, including image recognition [30], semantic segmentation [31], and regression [25]. Previous studies have demonstrated the significance of uniform or balanced distribution of class representations for effective imbalanced classification [5, 8, 9, 12, 22, 26, 33] and imbalanced semantic segmentation [31]. In classification tasks, these representations typically form distinct clusters. However, in the context of regression, representations are expected to be continuous and ordered [24, 27], rendering the methods used for quantifying and analyzing" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 340, + 245, + 531, + 439 + ], + "blocks": [ + { + "bbox": [ + 340, + 245, + 531, + 439 + ], + "lines": [ + { + "bbox": [ + 340, + 245, + 531, + 439 + ], + "spans": [ + { + "bbox": [ + 340, + 245, + 531, + 439 + ], + "type": "image", + "image_path": "eb0bcd76be5fa7fd68998314356045189c095b6a460638f2d7f490a584543caa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 448, + 556, + 525 + ], + "lines": [ + { + "bbox": [ + 313, + 448, + 556, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 556, + 525 + ], + "type": "text", + "content": "Figure 1. 2D feature space of vanilla baseline and ours from UCI-Airfoil [2]. The vanilla feature space lacks uniformity and is dominated by samples from the Many-shot region. In contrast, our approach achieves a more uniform distribution over the feature space, improving the performance, especially in the Medium and Few-shot regions. (For visualization purposes, we curated the dataset to ensure equal partitions across the three regions.)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 554, + 556, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 556, + 697 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 556, + 697 + ], + "type": "text", + "content": "uniformity in classification inapplicable. While the issue of deep imbalanced regression (DIR) has received considerable attention, the focus has predominantly been on training unbiased regressors, rather than on the aspect of representation learning [6, 10, 16, 19, 25]. Among the methods that do explore representation learning [24, 27], the emphasis is typically on understanding the relationship between the label space and the feature space (the representations themselves should be continuous and ordered). However, a critical aspect that remains under-explored is the interaction between data representations and the entire feature space. Specifically, how these representations distribute within the full scope of the feature space has not been examined." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "Uniformity in classification refers to how effectively" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 130, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 130, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 130, + 703 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 141, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 141, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 141, + 713 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5082" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 72, + 545, + 257 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 545, + 257 + ], + "type": "image", + "image_path": "8767a6b3f5dc2e93007e8652762c4482ff49a8222aac57199e41cf4c292bd2b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 271, + 555, + 339 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 555, + 339 + ], + "type": "text", + "content": "Figure 2. t-SNE visualization [20] of feature comparison. The first row corresponds to the original UCI Airfoil Dataset [2], while the second row corresponds to its curated version, with an additional few-shot region in the middle of the label range. Colored arrows point to the few-shot regions and their corresponding positions in the feature distributions. We evaluate feature distributions using: MSE Loss (Baseline), SRL without uniformity loss (w/o " + }, + { + "bbox": [ + 55, + 271, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{env}}" + }, + { + "bbox": [ + 55, + 271, + 555, + 339 + ], + "type": "text", + "content": "), SRL without homogeneity loss (w/o " + }, + { + "bbox": [ + 55, + 271, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{homo}}" + }, + { + "bbox": [ + 55, + 271, + 555, + 339 + ], + "type": "text", + "content": "), and complete SRL (ours). The baseline leads to feature collapse to many-shot regions and inadequate distinction of few-shot samples. In w/o " + }, + { + "bbox": [ + 55, + 271, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{env}}" + }, + { + "bbox": [ + 55, + 271, + 555, + 339 + ], + "type": "text", + "content": ", features collapse into a trivial shape, not fully utilizing the feature space. In w/o " + }, + { + "bbox": [ + 55, + 271, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{homo}}" + }, + { + "bbox": [ + 55, + 271, + 555, + 339 + ], + "type": "text", + "content": ", features spread out along the trace. Different from the previous ones, our SRL uniformly and smoothly “fills” the feature space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 347, + 297, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 297, + 563 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 297, + 563 + ], + "type": "text", + "content": "different clusters or centroids occupy the feature space, essentially partitioning it among various classes. In regression, here we define the term \"latent trace\" as the pathway that the representations follow, delineating the transition from the minimum to the maximum label values. In this paper, we aim to evaluate how well a latent trace occupies the feature space? To quantify this, we approximate a tubular neighborhood around the latent trace and measure its volume relative to the entire feature space. This method gauges the effectiveness of the trace in \"enveloping\" the hypersphere, and we call it enveloping loss. This loss ensures that the trace shape fills the surface of the hypersphere to facilitate uniformity. In parallel, it is equally important that the points (i.e., individual data point representations) are evenly distributed along the trace. To address uniform distribution along the trace as well as smoothness, we have developed a homogeneity loss. This loss is computed based on the arc length of the trace, allowing us to effectively measure and promote an even and smooth distribution of points." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "We model the uniformity in regression in two aspects: the induced trace aims to fully occupy the surface of the hypersphere (Enveloping), exhibiting smoothness with representations spaced at uniform intervals (Homogeneity). The two losses we introduce act as geometric constraints on a latent trace, implying they should not be applied to a set of representations from a single mini-batch. This is because a single batch likely does not encompass the full range of labels. To address this, we have developed a Surrogate-driven Representation Learning (SRL) scheme. It involves averaging representations of the same bins within a mini-batch to form centroids and \"re-filling\" missing bins by taking corresponding centroids from the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 347, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 430 + ], + "type": "text", + "content": "previous epoch. This process results in a surrogate containing centroids for all bins, enabling the effective application of geometric loss across the complete label range. Furthermore, we introduce Imbalanced Operator Learning (IOL) as a new DIR benchmark for training models on imbalanced domain locations in function space mapping. In summary, our main contributions are four-fold:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 432, + 556, + 624 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 432, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 491 + ], + "type": "text", + "content": "- Geometric Losses for Uniformity in deep imbalanced regression (DIR). To the best of our knowledge, this work is the first to study representation learning in DIR. We introduce two novel loss functions, enveloping loss and homogeneity loss, to ensure uniform feature distribution for DIR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 492, + 556, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 556, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 556, + 516 + ], + "type": "text", + "content": "- SRL Framework. A new framework is proposed that incorporates these geometric principles into data representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 517, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 564 + ], + "type": "text", + "content": "- Imbalanced Operator Learning (IOL). For the first time, we pioneer the task of operator learning within the realm of deep imbalanced regression, introducing an innovative task: Imbalanced Operator Learning (IOL)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 564, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 556, + 624 + ], + "type": "text", + "content": "- Extensive Experiments. The effectiveness of the proposed method is validated through experiments involving real-world regression and operator learning, on five datasets: AgeDB-DIR, IMDB-WIKI-DIR, STS-B-DIR, and two newly created DIR benchmarks, UCI-DIR and OL-DIR." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 633, + 398, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 398, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 398, + 646 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "Uniformity in imbalanced classification. Wang and Isola [22] identifies that uniformity is one of the key properties in contrastive representation learning. To promote uniformity in representation space for imbalanced classification, a variety of training strategies have been proposed. Kang et al. [8]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5083" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "decouples the training into a two-stage training of representation learning and classification. Yin et al. [26] designs a transfer learning framework for imbalanced face recognition. Kang et al. [9] combines supervised method and contrastive learning to learn a discriminative and balanced feature space. PaCo [5] and TSC [12] learn a set of class-wise balanced centers. BCL [33] balances the gradient distribution of negative classes and data distribution in mini-batch. Recent study suggests that sample-level uniform distribution may not effectively address imbalanced classification, advocating for category-level uniformity instead [1, 32]. Though progress has been made in this field, challenges persist in adapting the approach of modeling uniformity from classification to regression." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 228, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 456 + ], + "type": "text", + "content": "Deep imbalanced regression. With imbalanced regression data, effective learning in a continuous label space focuses on modeling the relationships among labels in the feature space [25]. Label Distribution Smoothing (LDS) [25] and DenseLoss [19] apply a Gaussian kernel to the observed label density, leading to an estimated label density distribution. Feature distribution smoothing (FDS) [25] generalizes the application of kernel smoothing from label to feature space. Ranksim [6] aims to leverage both local and global dependencies in data by aligning the order of similarities between labels and features. Balanced MSE [16] addresses the issue of imbalance in Mean Squared Error (MSE) calculations, ensuring a more balanced distribution of predictions. VIR [23] provides uncertainty for imbalanced regression. ConR [10] regularizes contrastive learning in regression by modeling global and local label similarities in feature space. RNC [27] and SupReMix [24] learn a continuous and ordered representation for regression through supervised contrastive learning. How imbalanced regression representations leverage the feature space remains under-explored." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 466, + 109, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 109, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 109, + 478 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "type": "text", + "content": "In the field of representation learning for classification, the concept of uniformity is pivotal for maximizing the use of the feature space [5, 8, 9, 12, 22, 26, 33]. This idea is based on the principle of ensuring that features from different classes are not only distinctly separated but also evenly distributed in the latent space. This uniform distribution of class centroids fosters a clear and effective decision boundary, leading to more accurate classification. However, in regression, where we deal with continuous, ordered trace [24, 27] rather than discrete clusters, the concept of uniformity is not only more complex but essentially remains undefined." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "We draw an analogy to the process of winding yarn around a ball. In this analogy, the yarn represents the latent trace, and the ball symbolizes the entirety of the available feature space. Just as the yarn must be evenly distributed across the ball's surface to effectively cover it (without any crossing), the latent trace should strive to occupy the hypersphere of the latent space uniformly. This ensures that the model leverages the available feature space to its fullest extent, enhancing the model's ability" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 432, + 167 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 432, + 167 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 432, + 167 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 432, + 167 + ], + "type": "image", + "image_path": "4cf76af194660921134e539fab8b701f1dda0cde8ee50395c6e97068b76a9206.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 174, + 555, + 231 + ], + "lines": [ + { + "bbox": [ + 313, + 174, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 555, + 231 + ], + "type": "text", + "content": "Figure 3. 2D schematic overview of two geometric losses. The arrow indicates the improvement of the loss function. Enveloping loss encourages the representations to fill the latent space, and homogeneity loss encourages the smoothness and even distribution of the representations along the trace." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 437, + 70, + 554, + 167 + ], + "blocks": [ + { + "bbox": [ + 437, + 70, + 554, + 167 + ], + "lines": [ + { + "bbox": [ + 437, + 70, + 554, + 167 + ], + "spans": [ + { + "bbox": [ + 437, + 70, + 554, + 167 + ], + "type": "image", + "image_path": "b34ec2e285c6444532d57a6144820103b4c031999efb5d48189daa7136892135.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 251, + 487, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 487, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 487, + 262 + ], + "type": "text", + "content": "to capture the variability inherent in the data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 263, + 554, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 323 + ], + "type": "text", + "content": "Furthermore, the latent trace should be smooth and continuous, akin to the even stretching of yarn, rather than loose and disjointed. This smoothness ensures a consistent and predictable model behavior, which is crucial for the accurate prediction and interpretation of results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 323, + 556, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 556, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 556, + 431 + ], + "type": "text", + "content": "We outline our method in this section. Firstly, we establish the fundamental notations and preliminaries (Section 3.1). Following this, we delve into the concept and definition of our enveloping loss (Section 3.2) and homogeneity loss (Section 3.3). Finally, we present our Surrogate-driven Representation Learning (SRL) framework, which incorporates the geometric constraints from the global image of the representations into the local range (Section 3.4). Refer to Supplementary Material 9 for the pseudo code of our method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 440, + 400, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 440, + 400, + 451 + ], + "spans": [ + { + "bbox": [ + 314, + 440, + 400, + 451 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": "A regression dataset is composed of pairs " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_i, y_i)" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": " represents the input and " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": " is the corresponding continuous-value target. Denote " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i = f(\\mathbf{x}_i)" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": " as the feature representation of " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": ", generated by a neural network " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": ". The feature representation is normalized so that " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\|\\mathbf{z}_i\\| = 1" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": ". Suppose the dataset consists of " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": " unique bins1, we define a surrogate as a set of centroids " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_k" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": ", where each represents a distinct bin. These centroids are computed by averaging the representations " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": " sharing the same bin and they are normalized to " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\|\\mathbf{c}_k\\| = 1" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": " be a path: " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "l : [y_{\\min}, y_{\\max}] \\mapsto \\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\|l(y)\\| = 1" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "l(y_k) = \\mathbf{c}_k" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": ". The path " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": " is a continuous curve extended from the discrete dataset that lies on a submanifold of " + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 459, + 556, + 602 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 611, + 390, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 390, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 390, + 624 + ], + "type": "text", + "content": "3.2. Enveloping" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 555, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 667 + ], + "type": "text", + "content": "To maximize the use of the feature space, it is crucial for the ordered and continuous trace of regression representations to fill the entire unit hypersphere as much as possible. This is by" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 674, + 556, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 556, + 715 + ], + "type": "text", + "content": "1The binning process (only for geometric loss calculations) facilitates surrogate formulation by using discrete centroids to approximate a continuous trace. Binning is unnecessary for most DIR datasets since they come pre-binned. The performed regression tasks follow the original dataset settings." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5084" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "content": "analogy with wrapping yarn with a certain length around a ball (without any crossing), aiming to cover as much surface area as possible." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "type": "text", + "content": "The trace of regression representations lying on a submanifold of " + }, + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "type": "text", + "content": " has a negligible hypervolume, which makes it challenging to assess its relationship with the entire hypersphere. To address this challenge, we extend the \"line\" into a tubular neighborhood. This expansion allows us to introduce the concept of enveloping loss. Our objective with this loss function is to maximize the hypervolume of the tubular neighborhood in proportion to the total hypervolume of the hypersphere." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 296, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 229 + ], + "type": "text", + "content": "Denote the set of all unit vectors in " + }, + { + "bbox": [ + 55, + 205, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 205, + 296, + 229 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 205, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 55, + 205, + 296, + 229 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 55, + 205, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\epsilon \\in (0,1)" + }, + { + "bbox": [ + 55, + 205, + 296, + 229 + ], + "type": "text", + "content": ", define tubular neighborhood " + }, + { + "bbox": [ + 55, + 205, + 296, + 229 + ], + "type": "inline_equation", + "content": "T(l,\\epsilon)" + }, + { + "bbox": [ + 55, + 205, + 296, + 229 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 205, + 296, + 229 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 205, + 296, + 229 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 239, + 296, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 239, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 88, + 239, + 296, + 253 + ], + "type": "interline_equation", + "content": "T (l, \\epsilon) = \\left\\{\\mathbf {z} \\in \\mathcal {U} \\mid \\mathbf {t} \\cdot \\mathbf {z} > \\epsilon \\text {f o r s o m e} \\mathbf {t} \\in \\operatorname {I m} (l) \\right\\} \\tag {1}", + "image_path": "77b981b99b3555c329996c5f606790c572c65d4f47a36ca3078bd838af29f39b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 263, + 295, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 295, + 288 + ], + "type": "text", + "content": "where for a function " + }, + { + "bbox": [ + 55, + 263, + 295, + 288 + ], + "type": "inline_equation", + "content": "f: A \\to B" + }, + { + "bbox": [ + 55, + 263, + 295, + 288 + ], + "type": "text", + "content": ", the image is defined as " + }, + { + "bbox": [ + 55, + 263, + 295, + 288 + ], + "type": "inline_equation", + "content": "\\operatorname{Im}(f) \\coloneqq \\{f(x), x \\in A\\}" + }, + { + "bbox": [ + 55, + 263, + 295, + 288 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 289, + 218, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 289, + 218, + 300 + ], + "spans": [ + { + "bbox": [ + 67, + 289, + 218, + 300 + ], + "type": "text", + "content": "Then our enveloping loss is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 310, + 296, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 310, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 134, + 310, + 296, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e n v}} = - \\frac {\\operatorname {v o l} (T (l , \\epsilon))}{\\operatorname {v o l} (\\mathcal {U})} \\tag {2}", + "image_path": "bc095aabd40411b9d96aa7c58ead5950eda910657acb36670b52b2de637454bc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 345, + 295, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 295, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 345, + 295, + 369 + ], + "type": "inline_equation", + "content": "\\mathrm{vol}(\\cdot)" + }, + { + "bbox": [ + 55, + 345, + 295, + 369 + ], + "type": "text", + "content": " returns the hypervolume of its input in the induced measure from the Euclidean space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 369, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 489 + ], + "type": "text", + "content": "In practical scenarios, the trace is composed of discrete representations, which complicates the direct computation of the tubular neighborhood's hypervolume. To navigate this challenge, we propose a continuous-to-discrete strategy. We first generate " + }, + { + "bbox": [ + 55, + 369, + 296, + 489 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 369, + 296, + 489 + ], + "type": "text", + "content": " points that are uniformly distributed across the hypersphere. We then determine the fraction of these points that fall within the neighbourhood " + }, + { + "bbox": [ + 55, + 369, + 296, + 489 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 369, + 296, + 489 + ], + "type": "text", + "content": ". This fraction effectively approximates the proportion of the hypersphere covered by the tubular neighborhood with a sufficiently large " + }, + { + "bbox": [ + 55, + 369, + 296, + 489 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 369, + 296, + 489 + ], + "type": "text", + "content": ". To adapt " + }, + { + "bbox": [ + 55, + 369, + 296, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{env}}" + }, + { + "bbox": [ + 55, + 369, + 296, + 489 + ], + "type": "text", + "content": " to discrete datasets, we re-formalize our optimization objective as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 140, + 500, + 296, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 500, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 140, + 500, + 296, + 525 + ], + "type": "interline_equation", + "content": "\\max \\lim _ {N \\rightarrow \\infty} \\frac {P (N)}{N} \\tag {3}", + "image_path": "cccb3caf0df47df2d1c914e9d4f0529452aadb279b118b3a72bcbe0199b742e9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 530, + 82, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 82, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 82, + 540 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 92, + 552, + 296, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 552, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 92, + 552, + 296, + 571 + ], + "type": "interline_equation", + "content": "P (N) := \\left| \\left\\{\\mathbf {p} _ {i} \\mid \\max _ {y} \\left\\{\\mathbf {p} _ {i} \\cdot l (y) \\right\\} > \\epsilon , i \\in [ N ] \\right\\} \\right| \\tag {4}", + "image_path": "1fc008f2a75af93fbf850ac1f6239ae56ebf4fc2f5cf5656fe10f32779f97fa0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "text", + "content": "assuming for each " + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "inline_equation", + "content": "N > 0" + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "text", + "content": ", we can choose " + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "text", + "content": " evenly distributed points in " + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "text", + "content": ", and denote these points as " + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i, i \\in [N] = 1, \\dots, N" + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "text", + "content": ". For numerical application, we take " + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "text", + "content": " to be a sufficiently large number and use the standard Monte-Carlo method [17] to approximate the evenly distributed points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "In our implementation, we did not directly define " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " due to the non-differentiability of the binarization required to determine if a " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " is within the " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "-tube. Instead, for each " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": ", we maximize the cosine similarity between " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " and its closest point on the trace. In this way, we relax the step function represented by (4) to its \"soft\" version, leading to smooth gradient computation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 72, + 399, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 399, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 399, + 85 + ], + "type": "text", + "content": "3.3. Homogeneity" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 89, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 329 + ], + "type": "text", + "content": "While the enveloping loss effectively governs the overall distribution of representations on the hypersphere, it alone may not be entirely adequate, presenting two unresolved issues. 1) The first is distribution along the trace. The enveloping loss predominantly controls the overall shape of representations on the hypersphere, yet it does not guarantee a uniform distribution along the trace. This poses a notable concern, as it may result in uneven representation density across different trace segments. 2) The second is trace smoothness. The enveloping loss could lead to a zigzag pattern of the representations, which should be avoided. Considering age estimation from facial images as an example, the progression of facial features over time is gradual. Consequently, in the corresponding latent space, we would anticipate a similar, smooth transition without abrupt changes, underlining the desirability of a smoother trace. Interestingly, these two issues can be aptly analogized to winding yarns around a ball as well. For the yarn on the ball to be smooth, it should be tightly stretched, rather than being disjointed or loosely arranged. We name the property of a trace to be smooth with representations evenly distributed along it as homogeneity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 330, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 554, + 376 + ], + "type": "text", + "content": "We encourage such homogeneity property, i.e., smoothness of the trace " + }, + { + "bbox": [ + 313, + 330, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\operatorname{Im}(l)" + }, + { + "bbox": [ + 313, + 330, + 554, + 376 + ], + "type": "text", + "content": " and uniform distribution of representations along it, by penalizing the arc length. Formally, the homogeneity loss is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 379, + 381, + 555, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 381, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 379, + 381, + 555, + 411 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {h o m o}} = \\int_ {y _ {\\min }} ^ {y _ {\\max }} \\left\\| \\frac {\\mathrm {d} l (y)}{\\mathrm {d} y} \\right\\| ^ {2} \\mathrm {d} y \\tag {5}", + "image_path": "2fadc99f3da4eb1466270cc54e9e529152b153dfaa027b795a651b23281e6672.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "text", + "content": " different " + }, + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "inline_equation", + "content": "y\\mathrm{s}" + }, + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "text", + "content": " which have been ordered, the discrete format for " + }, + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{homo}}" + }, + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "text", + "content": " is defined as a summation of the squared differences between adjacent points:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 371, + 459, + 554, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 459, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 371, + 459, + 554, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {h o m o}} = \\sum_ {k = 1} ^ {K - 1} \\frac {\\left\\| l \\left(y _ {k + 1}\\right) - l \\left(y _ {k}\\right) \\right\\| ^ {2}}{y _ {k + 1} - y _ {k}} \\tag {6}", + "image_path": "1998942d226228f2f0a59e147f1e187b1aa0901e5c56d6ecb68ba7508a82231b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "content": "The use of only homogeneity loss might result in trivial solutions like representation convergence to a circle or point due to feature collapse (shown in Figure 2). The homogeneity loss should be treated as a regularization of the enveloping loss, promoting not only smoothness but also an even distribution of representations along the trace. To quantitatively define the relationship between trace arc length and these desired characteristics, we introduce Theorem 1. It demonstrates that with a given " + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\operatorname{Im}(l)" + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{env}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "content": " is fixed as it does not depend on the parameterization of " + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "content": "), the homogeneity loss is minimized if and only if when representations are uniformly distributed along the trace." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "type": "text", + "content": "Theorem 1. Given an image of " + }, + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{homo}}" + }, + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "type": "text", + "content": " attains its minimum if and only if the representations are uniformly distributed along the trace, i.e., " + }, + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "type": "inline_equation", + "content": "\\| \\nabla_y l(y) \\| = c" + }, + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "type": "text", + "content": " is a constant." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 326, + 677, + 512, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 677, + 512, + 689 + ], + "spans": [ + { + "bbox": [ + 326, + 677, + 512, + 689 + ], + "type": "text", + "content": "Refer to Supplementary Material 6 for the proof." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Therefore, we formulate our geometric constraints " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{G}})" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " as a combination of enveloping and homogeneity:" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5085" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 249 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 249 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 249 + ], + "type": "image", + "image_path": "1180f604008d7fc4afc0a3c149cc193a1008316072c6d0f88fd7743b3016aec1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 258, + 296, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 380 + ], + "type": "text", + "content": "Figure 4. Overview of Surrogate-driven Representation Learning (SRL). (1) Every mini-batch is encoded to the latent space. Some bins may not be present in the current batch. To address this, (2) it takes centroids corresponding to the missing bins from the previous epoch. These stored centroids are used to \"re-fill\" the missing bins in the current batch. (3) Average the representations for bins that appear multiple times, creating centroids for these bins. This surrogate, containing a representation for the full label range, allows for the effective application of geometric loss across all bins. (4) Loss calculation based on the surrogate. (5) Update the surrogate in memory to ensure enveloping and homogeneity. The training of the first epoch is driven by MSE loss only." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 129, + 403, + 295, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 403, + 295, + 416 + ], + "spans": [ + { + "bbox": [ + 129, + 403, + 295, + 416 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {G}} = \\lambda_ {e} \\mathcal {L} _ {\\text {e n v}} + \\lambda_ {h} \\mathcal {L} _ {\\text {h o m o}} \\tag {7}", + "image_path": "62c1122125c9c620926f65f499717629f08d6067d6ccf2000a8b6efb900bc6be.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\lambda_{e}" + }, + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\lambda_{h}" + }, + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "type": "text", + "content": " are weights for the two geometric losses. In Section 4.4, we further explore the behavior of these two geometric constraints, uncovering new insights into imbalanced regression." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 478, + 296, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 296, + 502 + ], + "type": "text", + "content": "3.4. Surrogate-driven Representation Learning (SRL)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 508, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 556 + ], + "type": "text", + "content": "Our geometric loss " + }, + { + "bbox": [ + 55, + 508, + 295, + 556 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{G}})" + }, + { + "bbox": [ + 55, + 508, + 295, + 556 + ], + "type": "text", + "content": " is calculated on a surrogate instead of a mini-batch (Figure 4), as the representations from one mini-batch very likely fail to capture the global image of " + }, + { + "bbox": [ + 55, + 508, + 295, + 556 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 508, + 295, + 556 + ], + "type": "text", + "content": ", due to the randomness of batch sampling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "text", + "content": "For illustration purposes, here we assume the original dataset has already been binned, as is the case in most DIR datasets [6, 10, 25]. Let " + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\{\\mathbf{z}_1,\\mathbf{z}_2,\\dots,\\mathbf{z}_M\\}" + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "text", + "content": " be a set of representations from a batch with batch-size " + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "text", + "content": ", and let " + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{y_1,y_2,\\dots,y_M\\}" + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "text", + "content": " (repetitions of the label values might exist) be the corresponding labels. Define the centroid " + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_y\\in \\mathcal{C}" + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{c}_y\\| = 1" + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "text", + "content": ") as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 637, + 296, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 637, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 107, + 637, + 296, + 669 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {y} = \\frac {1}{| \\left\\{\\mathbf {z} _ {m} \\mid y _ {m} = y \\right\\} |} \\sum_ {\\left\\{\\mathbf {z} _ {m} \\mid y _ {m} = y \\right\\}} \\mathbf {z} _ {m} \\tag {8}", + "image_path": "8d75f99eea4ff5d781c4706de438e9e4c43e87cc7fe07032177206fdc4152afd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 716 + ], + "type": "text", + "content": "Suppose the whole label range is covered by a set of unique " + }, + { + "bbox": [ + 55, + 677, + 296, + 716 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 677, + 296, + 716 + ], + "type": "text", + "content": " bins " + }, + { + "bbox": [ + 55, + 677, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{*} = \\{y_{1}^{*},y_{2}^{*},\\dots,y_{K}^{*}\\}" + }, + { + "bbox": [ + 55, + 677, + 296, + 716 + ], + "type": "text", + "content": ". The centroids for these " + }, + { + "bbox": [ + 55, + 677, + 296, + 716 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 677, + 296, + 716 + ], + "type": "text", + "content": " bins from the last epoch are denoted as " + }, + { + "bbox": [ + 55, + 677, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{C}' = \\{\\mathbf{c}_{y_1^*}',\\mathbf{c}_{y_2^*}',\\dots,\\mathbf{c}_{y_K^*}'\\}" + }, + { + "bbox": [ + 55, + 677, + 296, + 716 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "The surrogate " + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": " is then generated by re-filling the missing centroids:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 381, + 97, + 555, + 112 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 97, + 555, + 112 + ], + "spans": [ + { + "bbox": [ + 381, + 97, + 555, + 112 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = \\mathcal {C} \\cup \\left\\{\\mathrm {c} ^ {\\prime} _ {y _ {k} ^ {*}} \\mid y _ {k} ^ {*} \\in \\mathcal {Y} ^ {*} \\backslash \\mathcal {Y} \\right\\} \\tag {9}", + "image_path": "a77b0b5690b5209465fc13b03cbabcdd9e1ee62ee3af35db8ee7a5f0d5c511a7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 118, + 555, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 555, + 154 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 555, + 154 + ], + "type": "text", + "content": "We use AdamW [13] with momentum to update parameters " + }, + { + "bbox": [ + 313, + 118, + 555, + 154 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 118, + 555, + 154 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 118, + 555, + 154 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 118, + 555, + 154 + ], + "type": "text", + "content": ", to ensure a smooth transition of the local shape in the batch-wise representations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "text", + "content": "At the end of each epoch " + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "inline_equation", + "content": "e \\in (0, E]" + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "text", + "content": " (excluding the first), we use the representations learned during that epoch to form a running surrogate " + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\hat{S}^e" + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "inline_equation", + "content": "S^{e+1}" + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "text", + "content": " is formulated from the current epoch's surrogate " + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "inline_equation", + "content": "S^e" + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\hat{S}^e" + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "text", + "content": " with momentum. " + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "inline_equation", + "content": "S^{e+1}" + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "text", + "content": " is employed for training in the subsequent epoch. It facilitates a gradual transition between epochs, preventing abrupt variations: " + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "inline_equation", + "content": "S^{e+1} \\gets \\alpha \\cdot S^e + (1 - \\alpha) \\cdot \\hat{S}^e" + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": "We aim for individual representations from the encoder to converge towards their respective centroids that share the same label, while simultaneously distancing them from centroids associated with different labels. To achieve this, we incorporate a contrastive loss between the individual representations and the centroids. For each representation " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_m" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "y_{m} = y" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": ", the centroid " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_y" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " is considered as the positive, and the other centroids as negatives. The contrastive loss is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 346, + 555, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 346, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 347, + 346, + 555, + 379 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n}} = - \\sum_ {m = 1} ^ {M} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\mathbf {z} _ {m} , \\mathbf {c} _ {y}\\right)\\right)}{\\sum_ {y ^ {*} \\in \\mathcal {Y} ^ {*}} \\exp \\left(\\operatorname {s i m} \\left(\\mathbf {z} _ {m} , \\mathbf {c} _ {y ^ {*}}\\right)\\right)} \\tag {10}", + "image_path": "75278db26bf9fd87334ab22439732d286e3a78857aa52bf4f1216b0754981f69.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 388, + 527, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 527, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 527, + 400 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 388, + 527, + 400 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot)" + }, + { + "bbox": [ + 313, + 388, + 527, + 400 + ], + "type": "text", + "content": " is the cosine similarity between two input." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 401, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 425 + ], + "type": "text", + "content": "The framework is trained end-to-end, the total loss used to update the parameters " + }, + { + "bbox": [ + 313, + 401, + 554, + 425 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 401, + 554, + 425 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 401, + 554, + 425 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 401, + 554, + 425 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 392, + 436, + 555, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 436, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 392, + 436, + 555, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\theta} = \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {\\mathrm {G}} + \\mathcal {L} _ {\\text {c o n}} \\tag {11}", + "image_path": "e1f9ee1561b047cf7110b9aa1ca2dde9b6b41fb29833a437a401e7192a4b5c26.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 326, + 460, + 515, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 460, + 515, + 473 + ], + "spans": [ + { + "bbox": [ + 326, + 460, + 515, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 326, + 460, + 515, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 326, + 460, + 515, + 473 + ], + "type": "text", + "content": " is the mean squared error (MSE) loss." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 485, + 391, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 391, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 391, + 498 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 506, + 555, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 625 + ], + "type": "text", + "content": "We perform extensive experiments to validate and analyze the effectiveness of SRL for deep imbalanced regression. Our regression tasks span age estimation from facial images, tabular regression, and text similarity score regression, as well as our newly established task: Imbalanced Operator Learning (IOL). This section begins by detailing the experiment setup (Section 4.1) followed by the main results (Section 4.2). The results of IOL are shown in Section 4.3 followed by the comparison with classification-based methods and hyperparameters analysis (Section 4.4)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 635, + 422, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 422, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 422, + 647 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Datasets. We employ three real-world regression datasets developed by Yang et al. [25], and our curated UCI-DIR from UCI Machine Learning Repository [2], to assess the effectiveness of SRL in deep imbalanced regression. Refer to Supplementary Material 7 for more dataset details." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5086" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 120, + 91, + 488, + 177 + ], + "blocks": [ + { + "bbox": [ + 118, + 70, + 492, + 82 + ], + "lines": [ + { + "bbox": [ + 118, + 70, + 492, + 82 + ], + "spans": [ + { + "bbox": [ + 118, + 70, + 492, + 82 + ], + "type": "text", + "content": "Table 1. Results on UCI-DIR (MAE). We report the average MAE of three runs. The best results are in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 120, + 91, + 488, + 177 + ], + "lines": [ + { + "bbox": [ + 120, + 91, + 488, + 177 + ], + "spans": [ + { + "bbox": [ + 120, + 91, + 488, + 177 + ], + "type": "table", + "html": "
DatasetsAirfoilAbaloneReal EstateConcrete
ShotAllManyMedFewAllManyMedFewAllManyMedFewAllManyMedFew
VANILLA5.665.115.036.754.570.882.657.970.330.270.380.377.295.776.929.74
LDS + FDS [25]5.764.454.797.795.090.903.269.260.350.330.400.346.886.216.737.59
RankSim [6]5.235.054.915.724.330.982.597.420.370.340.380.406.716.005.579.46
BalancedMSE [16]5.694.515.047.285.372.142.669.370.340.310.400.337.034.676.379.72
Ordinal Entropy [29]6.274.855.378.326.772.314.0111.610.340.290.420.357.125.506.369.31
SRL (ours)5.104.834.755.694.160.892.427.190.280.260.300.295.945.325.806.60
", + "image_path": "f0ec18012bcf2bccfefd4113e637db8c863463880c04643adb6c1aec7734935b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 67, + 206, + 284, + 308 + ], + "blocks": [ + { + "bbox": [ + 82, + 186, + 268, + 197 + ], + "lines": [ + { + "bbox": [ + 82, + 186, + 268, + 197 + ], + "spans": [ + { + "bbox": [ + 82, + 186, + 268, + 197 + ], + "type": "text", + "content": "Table 2. Results on AgeDB-DIR, the best are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 206, + 284, + 308 + ], + "lines": [ + { + "bbox": [ + 67, + 206, + 284, + 308 + ], + "spans": [ + { + "bbox": [ + 67, + 206, + 284, + 308 + ], + "type": "table", + "html": "
MetricsMAE ↓GM ↓
ShotAllManyMedFewAllManyMedFew
VANILLA7.676.669.3012.614.854.176.518.98
LDS + FDS [25]7.557.038.4610.524.864.575.386.75
RankSim [6]7.416.498.7312.474.714.155.748.92
BalancedMSE [16]7.987.588.659.935.014.835.466.30
Ordinal Entropy [29]7.606.698.8712.684.914.286.209.29
ConR [10]7.416.518.8112.044.704.135.918.59
SRL (ours)7.226.648.289.814.504.125.376.29
", + "image_path": "c59ec02fda9e3e661e6a219a6bbb1d0cf8ec5a736dfb5131fc4e01220956e5a5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 63, + 338, + 288, + 439 + ], + "blocks": [ + { + "bbox": [ + 73, + 318, + 277, + 328 + ], + "lines": [ + { + "bbox": [ + 73, + 318, + 277, + 328 + ], + "spans": [ + { + "bbox": [ + 73, + 318, + 277, + 328 + ], + "type": "text", + "content": "Table 3. Results on IMDB-WIKI-DIR, the best are in bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 338, + 288, + 439 + ], + "lines": [ + { + "bbox": [ + 63, + 338, + 288, + 439 + ], + "spans": [ + { + "bbox": [ + 63, + 338, + 288, + 439 + ], + "type": "table", + "html": "
MetricsMAE ↓GM ↓
ShotAllManyMedFewAllManyMedFew
VANILLA8.037.1615.4826.114.544.1410.8418.64
LDS + FDS [25]7.737.2212.9823.714.404.177.8715.77
RankSim [6]7.726.9214.5225.894.293.929.7218.02
BalancedMSE [16]8.437.8413.3523.274.934.687.9015.51
Ordinal Entropy [29]8.017.1715.1526.484.474.0710.5621.11
ConR [10]7.847.1514.3625.154.434.059.9118.55
SRL (ours)7.697.0812.6522.784.284.037.2815.25
", + "image_path": "14eda2fcc65413848790589c208930dd47f840717515a314da7f94b8819d3b70.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 59, + 469, + 291, + 563 + ], + "blocks": [ + { + "bbox": [ + 83, + 449, + 266, + 460 + ], + "lines": [ + { + "bbox": [ + 83, + 449, + 266, + 460 + ], + "spans": [ + { + "bbox": [ + 83, + 449, + 266, + 460 + ], + "type": "text", + "content": "Table 4. Results on STS-B-DIR, the best are in bold." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 469, + 291, + 563 + ], + "lines": [ + { + "bbox": [ + 59, + 469, + 291, + 563 + ], + "spans": [ + { + "bbox": [ + 59, + 469, + 291, + 563 + ], + "type": "table", + "html": "
MetricsMSE ↓Pearson correlation ↑
ShotAllManyMedFewAllManyMedFew
VANILLA0.9930.9631.0001.0750.7420.6850.6930.793
LDS + FDS [25]0.9000.9110.8810.9050.7570.6980.7230.806
RankSim [6]0.8890.9070.8740.7570.7630.7080.6920.842
BalancedMSE [16]0.9090.8941.0040.8090.7570.7030.6850.831
Ordinal Entropy [29]0.9430.9021.1610.8120.7500.7020.6790.767
SRL (ours)0.8770.8860.8730.7450.7650.7080.7490.844
", + "image_path": "8d02f2f7cfeadd8a401076150144495d28dcc2a335c854dc92317844c1b72a6e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "content": "- AgeDB-DIR [25]: It serves as a benchmark for estimating age from facial images, which is derived from the AgeDB dataset [15]. It contains 12,208 images for training, 2,140 images for validation, and 2,140 images for testing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 689 + ], + "type": "text", + "content": "- IMDB-WIKI-DIR [25]: It is a facial image dataset for age estimation derived from the IMDB-WIKI dataset [18], which consists of face images with the corresponding age. It has 191,509 images for training, 11,022 images for validation, and 11,022 for testing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- STS-B-DIR [25]: It is a natural language dataset formulated from STS-B dataset [3, 21], consisting of 5,249 training" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 188, + 555, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 188, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 322, + 188, + 555, + 222 + ], + "type": "text", + "content": "sentence pairs, 1,000 validation pairs, and 1,000 testing pairs. Each sentence pair is labeled with the continuous similarity score." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 224, + 555, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 307 + ], + "type": "text", + "content": "- UCI-DIR: To evaluate the performance of SRL on tabular data, we curated UCI Machine Learning Repository [2] to formulate UCI-DIR that includes four regression tasks (Airfoil Self-Noise, Abalone, Concrete Compressive Strength, Real estate valuation). Following the DIR setting [25], we make each regression task consist of an imbalanced training set and a balanced validation and test set." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 314, + 556, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 556, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 556, + 482 + ], + "type": "text", + "content": "Metrics. In line with the established settings in DIR [25], subsets in an imbalanced training set are categorized based on the number of available training samples: many-shot region (bins with " + }, + { + "bbox": [ + 313, + 314, + 556, + 482 + ], + "type": "inline_equation", + "content": ">100" + }, + { + "bbox": [ + 313, + 314, + 556, + 482 + ], + "type": "text", + "content": " training samples), medium-shot region (bins with 20 to 100 training samples), and few-shot region (bins with " + }, + { + "bbox": [ + 313, + 314, + 556, + 482 + ], + "type": "inline_equation", + "content": "<20" + }, + { + "bbox": [ + 313, + 314, + 556, + 482 + ], + "type": "text", + "content": " training samples), for the three real-world datasets. For AgeDB-DIR and IMDB-WIKI-DIR, each bin represents 1 year. In the case of STS-B-DIR, bins are segmented by 0.1. For UCI-DIR, the bins are segmented by 0.1 to 1 depending on the range of regression targets. Our evaluation metrics include mean absolute error (MAE, the lower the better) and geometric mean (GM, the lower the better) for AgeDB-DIR, IMDB-WIKI-DIR and UCI-DIR. For STS-B-DIR, we use mean squared error (MSE, the lower the better) and Pearson correlation (the higher the better)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 483, + 556, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 626 + ], + "type": "text", + "content": "Implementation Details. For age estimation (AgeDB-DIR and IMDB-WIKI-DIR), we follow the settings from Yang et al. [25], which uses ResNet-50 [7] as a backbone network. For text similarity regression (STS-B-DIR), we follow the setting from Cer et al. [3], Yang et al. [25] that uses BiLSTM + GloVe word embeddings. For tabular regression (UCI-DIR), we use an MLP with three hidden layers (d-20-30-10-1) following the setting from Cheng et al. [4]. For all baseline methods, results were produced following provided training recipes through publicly available codebase. All experimental results, including ours and baseline methods, were obtained from a server with 8 RTX 3090 GPUs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Baselines. We consider both DIR methods [6, 10, 25] and recent techniques proposed for general regression [16, 28, 29], in addition to VANILLA regression (MSE loss). We compare the performance of SRL with all baselines on the above four datasets. Furthermore, as SRL is orthogonal to previous DIR methods, we examine the improvement of them by adding our geometric losses." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5087" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 71, + 527, + 173 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 527, + 173 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 527, + 173 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 527, + 173 + ], + "type": "image", + "image_path": "a6304d257fc99b532b1d859fd23ec583a770306ad76766c3b40e135d9d82ebbd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 555, + 206 + ], + "type": "text", + "content": "Figure 5. SRL performance gain compared to VANILLA across age ranges on AgeDB-DIR. The gray histogram in the background shows the distribution of samples across age groups. SRL substantially improves the performance on the medium-shot and few-shot regions (age " + }, + { + "bbox": [ + 55, + 182, + 555, + 206 + ], + "type": "inline_equation", + "content": "< 20" + }, + { + "bbox": [ + 55, + 182, + 555, + 206 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 182, + 555, + 206 + ], + "type": "inline_equation", + "content": ">70" + }, + { + "bbox": [ + 55, + 182, + 555, + 206 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 232, + 291, + 354 + ], + "blocks": [ + { + "bbox": [ + 73, + 212, + 278, + 224 + ], + "lines": [ + { + "bbox": [ + 73, + 212, + 278, + 224 + ], + "spans": [ + { + "bbox": [ + 73, + 212, + 278, + 224 + ], + "type": "text", + "content": "Table 5. Combine SRL with existing DIR methods (MAE)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 232, + 291, + 354 + ], + "lines": [ + { + "bbox": [ + 57, + 232, + 291, + 354 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 291, + 354 + ], + "type": "table", + "html": "
DatasetsAgeDB (MAE)IMDB-WIKI (MAE)
ShotAllManyMedFewAllManyMedFew
SRL+LDS+FDS [25]7.326.818.149.817.617.0312.2821.77
GAINS v.s. LDS+FDS (%)3.053.233.896.751.662.645.408.19
SRL+RankSim [6]7.296.578.5810.487.677.0812.4022.85
GAINS v.s. RankSim (%)1.62-1.231.7216.960.65-1.1514.6111.75
SRL+BalancedMSE [16]7.246.777.869.857.747.1312.7722.04
GAINS v.s. BalancedMSE (%)9.2710.699.140.898.199.064.355.29
SRL+ConR [10]7.406.878.0810.507.567.0112.0321.71
GAINS v.s. ConR (%)0.14-5.538.3913.803.681.9616.2313.68
", + "image_path": "91467c7b5387a2b7bc8a7060e3fd062d2f8d5f2de5373158c7a85708bcc79bd1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 364, + 141, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 141, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 141, + 376 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 383, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 383, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 54, + 383, + 295, + 514 + ], + "type": "text", + "content": "To show the effectiveness of SRL on DIR, we first benchmark SRL and baselines for tabular regression on our curated UCIDIR with four different regression tasks (Table 1). Moreover, we evaluate our method on established DIR benchmarks [25] including age estimation on AgeDB-DIR and IMDB-WIKI-DIR (Table 2 & 3, and Figure 5), and text similarity regression on STS-B-DIR (Table 4). We evaluate the combination of SRL and previous DIR methods on AgeDB-DIR and IMDB-WIKI-DIR (Table 5). Notably, Table 1 and 4 omit results from ConR [10], as it depends on data augmentation, a technique not fully established in the domain of tabular data and natural language." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 516, + 297, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 297, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 297, + 636 + ], + "type": "text", + "content": "Combine SRL with existing methods. Our SRL approach enhances imbalanced regression by imposing geometric constraints on feature distributions, a strategy that is orthogonal to existing methods. To illustrate this, we leverage SRL as a regularizing term in conjunction with other methods. The results of this experiment are presented in Table 5. It shows that when SRL is integrated with existing regression methods, there is improvement in performance across different regions for both datasets. This demonstrates the effectiveness and compatibility of SRL as a complementary tool in the realm of regression analysis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 646, + 257, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 257, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 257, + 659 + ], + "type": "text", + "content": "4.3. Imbalanced Operator Learning (IOL)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "We introduce a novel task for DIR called Imbalanced Operator Learning (IOL). Traditional operator learning aims to train a neural network to model the mapping between function spaces [11, 14]. However, unlike the standard approach of uniformly" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 213, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 251 + ], + "type": "text", + "content": "sampling output locations, in IOL, we intentionally adjust the sampling density within the output function's domain to create regions with few, medium, and many regions (Figure 6)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 251, + 555, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 274 + ], + "type": "text", + "content": "For the linear operator, the model is trained to estimate the integral operator denoted as " + }, + { + "bbox": [ + 313, + 251, + 555, + 274 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 251, + 555, + 274 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 359, + 280, + 555, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 280, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 359, + 280, + 555, + 306 + ], + "type": "interline_equation", + "content": "G: u (x) \\mapsto s (x) = \\int_ {0} ^ {x} u (\\tau) d \\tau , x \\in [ 0, 1 ] \\tag {12}", + "image_path": "2c8ec2271db818039a82033067b0838b67a4600f860357cd8e428d1488de2f1b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 319, + 554, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 319, + 554, + 342 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 319, + 554, + 342 + ], + "type": "text", + "content": " denotes the input function which is sampled from a Gaussian random field (GRF), and " + }, + { + "bbox": [ + 313, + 319, + 554, + 342 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 319, + 554, + 342 + ], + "type": "text", + "content": " is the target function." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 342, + 554, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 367 + ], + "type": "text", + "content": "For the nonlinear operator, the model is trained to learn a particular stochastic partial differential equation (PDE):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 359, + 373, + 555, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 373, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 359, + 373, + 555, + 394 + ], + "type": "interline_equation", + "content": "\\operatorname {d i v} \\left(e ^ {b (x; \\omega)} \\nabla u (x; \\omega)\\right) = f (x), x \\in [ 0, 1 ] \\tag {13}", + "image_path": "a236c7c42885fac0fa7547574b71a12b30f5058fe05dd29b1e8334806753a0b3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 400, + 554, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 400, + 554, + 424 + ], + "type": "inline_equation", + "content": "e^{b(x;\\omega)}" + }, + { + "bbox": [ + 313, + 400, + 554, + 424 + ], + "type": "text", + "content": " is the diffusion efficient and " + }, + { + "bbox": [ + 313, + 400, + 554, + 424 + ], + "type": "inline_equation", + "content": "u(x;\\omega)" + }, + { + "bbox": [ + 313, + 400, + 554, + 424 + ], + "type": "text", + "content": " is the target function." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 426, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 496 + ], + "type": "text", + "content": "Denote the domain of output function as " + }, + { + "bbox": [ + 313, + 426, + 555, + 496 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 426, + 555, + 496 + ], + "type": "text", + "content": ". For both linear and non-linear operator learning, we changed the original uniform sampling of " + }, + { + "bbox": [ + 313, + 426, + 555, + 496 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 426, + 555, + 496 + ], + "type": "text", + "content": " to three curated regions: few/medium/many. Afterward, we manually created an imbalanced training set of 10k samples and a balanced testing test of 100k samples, namely OL-DIR." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "text", + "content": "In Figure 6, we have a schematic overview of Imbalanced Operator Learning (IOL). The network is trained to model an integral operator " + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "text", + "content": ". The data provided to the model is " + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "inline_equation", + "content": "([u,y],G(u)(y))" + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "text", + "content": ". The input consists of function " + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "text", + "content": " and sampled " + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "text", + "content": "s from the domain of " + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "inline_equation", + "content": "G(u)" + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "text", + "content": ". The target is " + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "inline_equation", + "content": "G(u)(y)" + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "text", + "content": ". We manipulate the distribution density of " + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 497, + 555, + 605 + ], + "type": "text", + "content": " across its range to formulate few/med./many regions. Here the imbalance comes from the unequal exposure of integral interval to the model training. Refer to Supplementary Material 7.2 for more details." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": "Shown in Table 6, SRL consistently outperforms VANILLA and the state-of-the-art operator learning for the whole label range including all, many-shot, medium-shot, and few-shot regions. The results position SRL as the superior approach for IOL in terms of accuracy and generalizability." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 671, + 417, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 417, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 417, + 684 + ], + "type": "text", + "content": "4.4. Further Analysis" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Quantification of geometric impact. We further quantified the impact of geometric constraints by comparing percentages of" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5088" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 70, + 248, + 237 + ], + "blocks": [ + { + "bbox": [ + 102, + 70, + 248, + 237 + ], + "lines": [ + { + "bbox": [ + 102, + 70, + 248, + 237 + ], + "spans": [ + { + "bbox": [ + 102, + 70, + 248, + 237 + ], + "type": "image", + "image_path": "5400e1ef48185ad514b107f5de4074bb2d673bae22adb6428663f47894e962f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 238, + 246, + 248 + ], + "lines": [ + { + "bbox": [ + 104, + 238, + 246, + 248 + ], + "spans": [ + { + "bbox": [ + 104, + 238, + 246, + 248 + ], + "type": "text", + "content": "Figure 6. Imbalanced Operator Learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 278, + 293, + 397 + ], + "blocks": [ + { + "bbox": [ + 55, + 255, + 296, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 277 + ], + "type": "text", + "content": "Table 6. Results on OL-DIR. We report the average MAE of ten runs. The best results are bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 278, + 293, + 397 + ], + "lines": [ + { + "bbox": [ + 59, + 278, + 293, + 397 + ], + "spans": [ + { + "bbox": [ + 59, + 278, + 293, + 397 + ], + "type": "table", + "html": "
OperationMAE(10-3) ↓MSE (10-4) ↓
ShotAllManyMedFewAllManyMedFew
Linear
VANILLA [14]15.6411.8615.4527.005.402.814.4014.20
Ordinal Entropy [29]10.079.269.8513.012.001.531.893.42
SRL (ours)9.188.329.479.331.980.981.722.67
Nonlinear
VANILLA [14]11.649.8911.0219.779.204.337.5324.70
Ordinal Entropy [29]12.919.9313.0721.0213.808.8211.8430.12
SRL (ours)11.259.489.2217.008.607.426.4114.12
", + "image_path": "cad31ab045aff1e32b2786391ce6817418df1d4fe39abf3301f513119d9ec0d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 67, + 399, + 285, + 408 + ], + "lines": [ + { + "bbox": [ + 67, + 399, + 285, + 408 + ], + "spans": [ + { + "bbox": [ + 67, + 399, + 285, + 408 + ], + "type": "text", + "content": "Full results with standard deviation are reported in Supplementary Material 8.10." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": "uniformly sampled points within few-shot regions (a measure of proportion). The results show our method significantly increases few-shot proportion (AgeDB-DIR: " + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "inline_equation", + "content": "1.98\\% \\rightarrow 15.80\\%" + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": ", upper bound: " + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": "; STS-B-DIR: " + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "inline_equation", + "content": "4.52\\% \\rightarrow 22.39\\%" + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": ", upper bound: " + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": "), leading to improved performance (Table 7)." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 58, + 492, + 296, + 610 + ], + "blocks": [ + { + "bbox": [ + 64, + 480, + 286, + 491 + ], + "lines": [ + { + "bbox": [ + 64, + 480, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 64, + 480, + 286, + 491 + ], + "type": "text", + "content": "Table 7. Impact of geometric constraints on few-shot proportion." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 492, + 296, + 610 + ], + "lines": [ + { + "bbox": [ + 58, + 492, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 58, + 492, + 296, + 610 + ], + "type": "table", + "html": "
Few-shotOverall
ProportionMAEMAE
AgeDB-DIR (1.10% samples, 23% label range):
VANILLA1.98%12.617.67
LDS + FDS [25]4.95%10.527.55
Ours15.80%9.817.22
STS-B-DIR (3.49% samples, 38% label range):
VANILLA4.52%1.0750.993
LDS + FDS [25]8.13%0.9050.900
Ours22.39%0.8770.745
", + "image_path": "b897e845075bd8e3670cbb039146e9813fd4016c11d8a3425ed793febae01237.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Balancing of enveloping and homogeneity: Our proposed SRL advocates for two pivotal geometric constraints in feature distribution: enveloping and homogeneity, to effectively address imbalanced regression. These two losses are modulated by their respective coefficients, " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{e}" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " for the enveloping loss and " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{h}" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " for the homogeneity loss. Figure 7 illustrates that the omission of either constraint detrimentally impacts the performance, highlighting the importance of both of them, and it demonstrates that the best" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "performance, as measured by MAE on the AgeDB-DIR dataset, is achieved when both coefficients " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\lambda_{e}" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\lambda_{h}" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "1e^{-1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "type": "text", + "content": "Ablation studies on choices of " + }, + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "type": "text", + "content": ": Table 10 (in Supplementary Material) shows that achieving optimal performance on the AgeDB-DIR and IMDB-WIKI-DIR datasets requires a sufficiently large " + }, + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "type": "text", + "content": ", as a smaller " + }, + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "type": "text", + "content": " may lead to imprecise calculation of the enveloping loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "text", + "content": "Ablation studies on proposed loss component: Table 11 (in Supp.) demonstrates that incorporating homogeneity, enveloping, and contrastive loss term yields superior model performance compared to using each individually." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "type": "text", + "content": "Computational cost: As shown in Table 12 (in Supp.), the computational overhead introduced by the Surrogate-driven Representation Learning (SRL) framework is comparable to that of other imbalanced regression methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 251, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 299 + ], + "type": "text", + "content": "Impact of bin numbers: As shown in Table 13 (in Supp.), while increasing the number of bins generally leads to better model performance, the improvements become marginal beyond certain thresholds." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "type": "text", + "content": "Limitations: Section 11 (in Supp.) examines the limitations of SRL, including its inability to handle higher-dimensional labels." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 340, + 324, + 531, + 475 + ], + "blocks": [ + { + "bbox": [ + 340, + 324, + 531, + 475 + ], + "lines": [ + { + "bbox": [ + 340, + 324, + 531, + 475 + ], + "spans": [ + { + "bbox": [ + 340, + 324, + 531, + 475 + ], + "type": "image", + "image_path": "9a177d605c7c679d6f50dbf76a7eabe126272c52ff8f9cc41bea698ffedce2ab.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "type": "text", + "content": "Figure 7. Confusion matrix of MAE on AgeDB-DIR from different values of " + }, + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\lambda_h" + }, + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\lambda_e" + }, + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 514, + 384, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 384, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 384, + 526 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "As the first work of exploring uniformity in deep imbalanced regression, we introduce two novel loss functions - enveloping and homogeneity loss - to encourage the uniform feature distribution of an ordered and continuous trajectory. The two loss functions serve as geometric constraints which are integrated into the data representations through a Surrogate-driven Representation Learning (SRL) framework. Furthermore, we set a new benchmark in imbalanced regression: Imbalanced Operator Learning (IOL). Extensive experiments on real-world regression and operator learning demonstrate the effectiveness of our geometrically informed approach. We emphasize the significance of uniform data representation and its impact on learning performance in imbalanced regression scenarios, advocating for a more balanced and comprehensive utilization of feature spaces in regression models." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5089" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 112, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 112, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 112, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Mido Assran, Randall Balestriero, Quentin Duval, Florian Bordes, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, and Nicolas Ballas. The hidden uniform cluster prior in self-supervised learning. In *The Eleventh International Conference on Learning Representations*, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 295, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 295, + 168 + ], + "type": "text", + "content": "[2] Arthur Asuncion and David Newman. Uci machine learning repository, 2007. 1, 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 170, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 296, + 212 + ], + "type": "text", + "content": "[3] Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. arXiv preprint arXiv:1708.00055, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 214, + 296, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 214, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 61, + 214, + 296, + 246 + ], + "type": "text", + "content": "[4] Xin Cheng, Yuzhou Cao, Ximing Li, Bo An, and Lei Feng. Weakly supervised regression with interval targets. arXiv preprint arXiv:2306.10458, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 247, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 247, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 61, + 247, + 296, + 290 + ], + "type": "text", + "content": "[5] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 292, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 292, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 292, + 296, + 334 + ], + "type": "text", + "content": "[6] Yu Gong, Greg Mori, and Fred Tung. Ranksim: Ranking similarity regularization for deep imbalanced regression. In International Conference on Machine Learning, pages 7634-7649. PMLR, 2022. 1, 3, 5, 6, 7, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 336, + 296, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 336, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 61, + 336, + 296, + 379 + ], + "type": "text", + "content": "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 380, + 296, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 380, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 61, + 380, + 296, + 423 + ], + "type": "text", + "content": "[8] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In International Conference on Learning Representations, 2019. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 425, + 296, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 425, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 61, + 425, + 296, + 457 + ], + "type": "text", + "content": "[9] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2020. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 459, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 459, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 459, + 296, + 491 + ], + "type": "text", + "content": "[10] Mahsa Keramati, Lili Meng, and R David Evans. Contr: Contrastive regularizer for deep imbalanced regression. arXiv preprint arXiv:2309.06651, 2023. 1, 3, 5, 6, 7, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 492, + 296, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 492, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 57, + 492, + 296, + 545 + ], + "type": "text", + "content": "[11] Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research, 24(89):1-97, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 547, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 296, + 601 + ], + "type": "text", + "content": "[12] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6928, 2022. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 603, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 635 + ], + "type": "text", + "content": "[13] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 679 + ], + "type": "text", + "content": "[14] Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deep-onet based on the universal approximation theorem of operators. Nature machine intelligence, 3(3):218-229, 2021. 7, 8, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[15] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 333, + 74, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 74, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 74, + 555, + 95 + ], + "type": "text", + "content": "In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "text", + "content": "[16] Jiawei Ren, Mingyuan Zhang, Cunjun Yu, and Ziwei Liu. Balancedmse for imbalanced visual regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7926-7935, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 554, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 163 + ], + "type": "text", + "content": "[17] Christian P Robert, George Casella, and George Casella. Monte Carlo statistical methods. Springer, 1999. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 164, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 554, + 207 + ], + "type": "text", + "content": "[18] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Deep expectation of real and apparent age from a single image without facial landmarks. International Journal of Computer Vision, 126 (2-4):144-157, 2018. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 209, + 555, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 241 + ], + "type": "text", + "content": "[19] Michael Steininger, Konstantin Kobs, Padraig Davidson, Anna Krause, and Andreas Hotho. Density-based weighting for imbalanced regression. Machine Learning, 110:2187-2211, 2021. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 243, + 554, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 554, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 554, + 264 + ], + "type": "text", + "content": "[20] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 266, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 554, + 308 + ], + "type": "text", + "content": "[21] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 310, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 554, + 354 + ], + "type": "text", + "content": "[22] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 355, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 554, + 398 + ], + "type": "text", + "content": "[23] Ziyan Wang and Hao Wang. Variational imbalanced regression: Fair uncertainty quantification via probabilistic smoothing. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 399, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 431 + ], + "type": "text", + "content": "[24] Yilei Wu, Zijian Dong, Chongyao Chen, Wangchunshu Zhou, and Juan Helen Zhou. Mixup your own pairs. arXiv preprint arXiv:2309.16633, 2023. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 434, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 476 + ], + "type": "text", + "content": "[25] Yuzhe Yang, Kaiwen Zha, Yingcong Chen, Hao Wang, and Dina Katabi. Delving into deep imbalanced regression. In International Conference on Machine Learning, pages 11842-11851. PMLR, 2021. 1, 3, 5, 6, 7, 8, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 478, + 554, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 532 + ], + "type": "text", + "content": "[26] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5704-5713, 2019. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 534, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 554, + 578 + ], + "type": "text", + "content": "[27] Kaiwen Zha, Peng Cao, Jeany Son, Yuzhe Yang, and Dina Katabi. Rank-n-contrast: Learning continuous representations for regression. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 579, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 554, + 622 + ], + "type": "text", + "content": "[28] Shihao Zhang, Linlin Yang, Michael Bi Mi, Xiaoxu Zheng, and Angela Yao. Improving deep regression with ordinal entropy. In The Eleventh International Conference on Learning Representations, 2022. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 624, + 554, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 655 + ], + "type": "text", + "content": "[29] Shihao Zhang, Linlin Yang, Michael Bi Mi, Xiaoxu Zheng, and Angela Yao. Improving deep regression with ordinal entropy. arXiv preprint arXiv:2301.08915, 2023. 6, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 658, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 689 + ], + "type": "text", + "content": "[30] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[31] Zhisheng Zhong, Jiequan Cui, Yibo Yang, Xiaoyang Wu, Xiaojuan Qi, Xiangyu Zhang, and Jiaya Jia. Understanding" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5090" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 206 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "imbalanced semantic segmentation through neural collapse. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19550-19560, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[32] Zhihan Zhou, Jiangchao Yao, Feng Hong, Ya Zhang, Bo Han, and Yanfeng Wang. Combating representation learning disparity with geometric harmonization. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[33] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022. 1, 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5091" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/18d0caf1-d799-41ef-95e4-0a0e2105060c_content_list.json b/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/18d0caf1-d799-41ef-95e4-0a0e2105060c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3f7c6e9414f4e7ca8e88cba814628bbd6f53cfff --- /dev/null +++ b/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/18d0caf1-d799-41ef-95e4-0a0e2105060c_content_list.json @@ -0,0 +1,2029 @@ +[ + { + "type": "text", + "text": "Improved monocular depth prediction using distance transform over pre-semantic contours with self-supervised neural networks", + "text_level": 1, + "bbox": [ + 153, + 130, + 844, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Marwane Hariat1, Antoine Manzanera1, David Filliat1 \n1U2IS, ENSTA, Institut Polytechnique de Paris, Palaiseau, France \n{marwane.hariat, antoine.manzanera, david fillliat}@ensta.fr", + "bbox": [ + 151, + 202, + 843, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73ce385542563d79605072ecd3c823dc265a3e81cd94c19e2dcc7cc1476d20a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 290, + 282, + 330 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e292b9f0aa921153a7b14ca2f5fbda63b5ea1f7ceecd46f0c0700c6b11402d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 338, + 281, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af0834f4151e38c5d13f50d1a97ba7c93e8a9f5335c9b7200e343455e140d46c.jpg", + "image_caption": [ + "RGB images" + ], + "image_footnote": [], + "bbox": [ + 91, + 386, + 281, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a28d1d8abc98af0348b84aed1b901258ce17e23d4c27bf208185a02163f29cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 299, + 290, + 490, + 330 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abf7d38042583f22379a5deb90829bfd5a4707f589c62d7b8351b7d0cb30cbbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 338, + 488, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/edc5ede1827ecb359dc7b8c0cb0ac8cb8db77b418c98a31a3e5c833b3cab10b9.jpg", + "image_caption": [ + "Depth", + "Figure 1. The different modalities estimated in our self-supervised framework and used to reduce matching uncertainty between images." + ], + "image_footnote": [], + "bbox": [ + 300, + 387, + 488, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca538a1b33f183da4755c3f0c6565799ccbeea122b1678de8dd143ee9d1ad2d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 290, + 699, + 327 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80ead54ef523e6a6c1e2dc06b3cf50094fad20aea5e264c119599434bc5d1e51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 337, + 697, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f14ef0f810705e78dd61b7145a061d339cd304907cf6b3ddbf77c532b8d7adab.jpg", + "image_caption": [ + "Edge" + ], + "image_footnote": [], + "bbox": [ + 506, + 386, + 699, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fe5fc609326d3ae92032ae0e85bb86ca4596462ec2f0e899d5d4b7926b779f3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 291, + 905, + 330 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7bcf76575f83f537b82df78db03aac86140c15025238ad7c4fb098720c582884.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 338, + 903, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/265b82a705b3e3b4a395056ff8cccaea227c5efbbf55bb4ee578e4b07e91a34f.jpg", + "image_caption": [ + "Distance Transform" + ], + "image_footnote": [], + "bbox": [ + 715, + 387, + 903, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 479, + 326, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular depth estimation (MDE) with self-supervised training approaches struggles in low-texture areas, where photometric losses may lead to ambiguous depth predictions. To address this, we propose a novel technique that enhances spatial information by applying a distance transform over pre-semantic contours, augmenting discriminative power in low texture regions. Our approach jointly estimates pre-semantic contours, depth and ego-motion. The pre-semantic contours are leveraged to produce new input images, with variance augmented by the distance transform in uniform areas. This approach results in more effective loss functions, enhancing the training process for depth and ego-motion. We demonstrate theoretically that the distance transform is the optimal variance-augmenting technique in this context. Through extensive experiments on KITTI, Cityscapes, Waymo, NYUv2 and ScanNet our model demonstrates robust performance, surpassing competing self-supervised methods in MDE.", + "bbox": [ + 86, + 511, + 485, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 814, + 220, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The goal of monocular depth estimation (MDE) is to determine the distance of a pixel from the camera focal plane using only a single input image, unlike stereo depth estimation [2, 21, 25], which relies on pairs of images. This", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eliminates the need for expensive sensors that require high-power resources [42, 48], such as active depth camera or LiDAR, while maintaining low latency and ensuring real-time compliance [15, 78]. This makes MDE a core module in applications such as robot navigation [65, 87], autonomous driving [52, 76] or virtual/augmented reality [14, 61]. Deep Neural Networks (DNNs) are currently the state of the art for depth maps prediction thanks to their ability to exploit all contextual cues of the environment [12, 32].", + "bbox": [ + 511, + 479, + 906, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Supervised methods for MDE rely on ground-truth annotations. In spite of huge progress [6, 13, 18, 70], these methods are constrained by the cost and quality of the annotations. In this paper, we focus on self-supervised MDE, that uses the structure-from-motion framework [19, 91] and leverages depth and camera ego-motion estimations from DNNs to reconstruct the next frame from the current one. The supervision signal is then provided by the photometric loss, i.e. the difference between the reconstructed frame and the observed one, based on a combination of direct colour comparison and Structural Similarity Index Measure (SSIM) [55]. Although more challenging, the self-supervised paradigm offers greater training flexibility. First, it can be trained using data from unknown cameras [23], as long as data acquisition provides successive frames, such as in YouTube videos [7]. This is particularly relevant given the growing interest in foundation models, which can now leverage large datasets to produce generalizable out-", + "bbox": [ + 511, + 617, + 908, + 890 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21868", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "puts [82]. In addition, it is compatible with continual learning, allowing the model to continuously expand its knowledge as more data become available [7, 39].", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the problem of MDE using the photometric loss proposed by Zhou et al. [91] is ill-posed due to its reliance on SSIM and colour-based cost functions, which fail to discriminate effectively in low-texture areas. In these regions, multiple candidates can yield similarly low cost function values, making it difficult to determine the correct correspondence. Although interesting work has been done to address various issues such as scale ambiguity [73], gradient locality [22, 35], occlusion [22, 75], moving objects [41, 60], and infinite depth holes [22, 24], the ill-posed nature of this problem has not been thoroughly and rigorously studied. We aim to provide this essential foundation.", + "bbox": [ + 89, + 136, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key challenge lies in reducing ambiguity to build a more robust foundation for learning depth and ego-motion. To this end, we propose introducing sufficient variance in texture-less areas, while adhering as closely as possible to 'the constancy assumption': any change introduced in image $t$ should be reproduced identically in image $t + 1$ , essential as the photometric loss relies on correspondences between successive frames. We propose to do this by assigning to each pixel the distance to its nearest edge in the image, commonly known as the 'distance transform' [54, 68]. We demonstrate mathematically that the distance transform is the most effective way of adding variance while respecting the constancy assumption, and that it also leads to improved convergence properties. Since the distance transform is calculated on binary contours, we propose a self-supervised pre-semantic contour estimation method, jointly learned with the depth estimation model, and add this distance map as an additional channel to the input image to produce a 'variance-augmented' image for reconstruction. In summary, our contributions are as follows.", + "bbox": [ + 89, + 316, + 482, + 619 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) A self-supervised approach for pre-semantic contour estimation, which (i) generates the edge information necessary for computing the distance transform, and (ii) simultaneously reinforces the depth estimation network.", + "(2) A novel variance-augmented image that incorporates the distance transform, enhancing discriminative power in texture-less areas and improving convergence properties", + "(3) A theoretical foundation demonstrating the effectiveness of the distance transform for adding variance under the constancy assumption", + "(4) An extensive experimental validation of our new pipeline, achieving competitive performance on different benchmark datasets: KITTI [20], Cityscapes [9], Waymo [69], NYUv2 [67] and ScanNet [10]" + ], + "bbox": [ + 89, + 674, + 482, + 886 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 90, + 658, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised framework. The work of Zhou et al. [91] is pioneer in self-supervised monocular depth estimation. Their method utilizes depth and ego-motion to find the corresponding pixels between two successive frames and relies on the bilinear sampler [34] to reconstruct the upcoming scene from the current frame. The supervision signal is then derived from the photometric loss, which is a weighted sum of the SSIM and colour cost functions. The effectiveness of this method relies on the constancy assumption, which asserts that object colours remain consistent from frame to frame. This assumption is bound to the hypothesis of constant lighting over Lambertian surfaces, and do not hold in the case of occlusions and moving objects. This method has then been refined in several ways. Kim et al. [37] provide a comprehensive study of the main issues and solutions found in the literature. In [75], occlusions are addressed using a forward warping module that counts the number of bilinear sampling operations for each pixel. Godard et al. [22] compute the occlusion mask by determining the per-pixel minimum of the photometric loss over time. Moving objects are identified through the differences between optical flow and depth/ego-motion in [28, 40, 47, 60, 86, 93], while off-the-shelf semantic algorithms are utilized to localize potential moving objects in [4, 24]. In addition, a residual map is introduced to correct the ego-motion in [23, 41]. Infinite depth holes are filtered through a two-stage training process [24] and are removed using an adaptive cost-volume architecture in [77]. Finally, Godard et al. [22] propose a unified solution that effectively addresses all the aforementioned issues. However, the ill-posed nature of the problem due to unstructured objects remains unaddressed.", + "bbox": [ + 511, + 114, + 903, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contour estimation. Semantic contours should follow the boundaries derived from instance segmentation, outlining distinct entities within an image. Such contours would naturally help improve the accuracy of depth maps. [74] and [59] leverage ground-truth annotations to learn depth, normals, and contours, incorporating regularization terms to enforce constraints derived from their geometric relationships. In contrast, [43] employs an off-the-shelf edge prediction algorithm to refine initial depth estimations as a post-processing step. [79], [64] and [45] utilize a pretrained instance segmentation network to compute edges, sampling points on both sides of the boundary to enhance depth sharpness around those edges. [92] introduced a morphing strategy to align depth borders computed with a threshold gradient on the depth with semantic borders obtained from a pre-trained semantic algorithm. While these techniques demonstrate significant improvements, they assume the edges as provided and are therefore not fully self-supervised, in addition to not being GPU friendly. Significant advances have been made in supervised edge map estimation [11, 80]; however, the self-supervised approach re", + "bbox": [ + 511, + 584, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21869", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mains an ambitious and challenging task. [46] introduces a reinforcement training procedure in which an optical flow network is trained to predict the displacement of each pixel between two frames, based on ground truth point correspondences. Subsequently, an edge detector is updated to align its predicted edges of the image with the predicted edges of the estimated flow map. Still, this method is not self-supervised. A work closer to ours [84] proposes a fully self-supervised edge prediction framework that utilizes the positive part of the second derivatives of depth and the first-order gradient of normal to surface to feed an edge network. However, this approach employs suboptimal loss functions, leading to biased contours that favour shorter distances. Furthermore, it does not utilize the edges to enhance depth, nor does it include any post-processing to create refined contours. We propose new cost functions for training the edge network, resulting in significant performance gains. Additionally, we introduce a post-processing step to refine contours and implement an extended edge-aware diffusion smoothness loss [31] that encourages sharp depth discontinuities around estimated boundaries, in the same spirit as [45, 79, 92].", + "bbox": [ + 89, + 90, + 483, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Targeting the ill-posed optimization problem. As previously mentioned, our objective is to minimize the photometric loss by establishing correspondences between pixels in the current frame and those in the next frame, achieved through the depth and ego-motion parameters. However, this correspondence procedure is not unique, since many pixels share the same colour and exhibit similar local colour statistics as their neighbourhood. This ambiguity is especially pronounced in areas with low texture. The challenge is to compensate for the lack of variation by introducing time-consistent variance across frames, to satisfy the constancy assumption. This reduces the ill-posed nature of the optimization problem and creates better conditions for learning depth and ego-motion. [81, 88] addressed this issue by applying the reconstruction loss to deep features in addition to the usual colour images. Shu et al. [66] took this approach further and designed custom deep features by training an auto-encoder that encourages large and smooth gradient values in feature maps, yielding deep features with both semantic richness and substantial structure. The constancy assumption is rather optimistic for these methods, as demonstrated in our experiments. Indeed, due to their large receptive fields, deep features generally lack consistency from frame to frame. Additionally, the reduced resolution of feature maps relative to the input image significantly limits the benefit of these methods in terms of accuracy. We show that the distance transform, not only proves to be experimentally more consistent across frames, but also stands as the theoretically best method for increasing variance under the constancy constraint, while preserving the resolution of the input image. Our work employs a similar methodol", + "bbox": [ + 89, + 422, + 483, + 893 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ogy as [26], which theoretically prove the strong invariant properties induced by the census transform used in robust optical flow estimation [50] and validate these properties through experiments. [8] also uses the distance transform, but as an additional input to infer dense depth from sparse depth. However, using the distance transform of edge maps as a training signal is a novel approach.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preamble", + "text_level": 1, + "bbox": [ + 513, + 208, + 619, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal in this section is to propose a mathematical formalism that closely aligns with experimental scenarios to address the following question: How can we introduce sufficient variance in low-texture areas while adhering as much as possible to the constancy assumption? The purpose is to provide a better basis than the usual framework [91] for learning depth and ego-motion.", + "bbox": [ + 511, + 233, + 906, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start by introducing some key theoretical properties of the distance transform that we illustrate with toy experiments. The following definitions and theorems provide the theoretical foundation of our analysis.", + "bbox": [ + 511, + 339, + 906, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Maximal variance under constraints", + "text_level": 1, + "bbox": [ + 513, + 407, + 828, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notations Let $\\Omega$ be a compact convex set of $\\mathbb{R}^n$ with smooth boundary $\\partial \\Omega$ . Let $\\mathcal{A}$ be the set of affine transformations. And let $\\mathcal{I}_{\\mathcal{A}}$ the set of positive functions invariant by affine transformation (see Definition 2).", + "bbox": [ + 511, + 430, + 905, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\Omega$ represents an object within a scene, while $\\mathcal{A}$ models a reasonable subset of reprojected 3d transformations likely to affect the objects from frame to frame.", + "bbox": [ + 511, + 489, + 905, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1 The distance transform $\\delta_{\\Omega}:\\Omega \\to \\mathbb{R}^{+}$ is defined as:", + "bbox": [ + 511, + 542, + 905, + 573 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {\\Omega} (x) = d (x, \\partial \\Omega) = \\inf _ {y \\in \\partial \\Omega} \\| x - y \\| \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 580, + 906, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The set of points having more than one closest point to the boundary is called the medial axis. By Rademacher's Theorem [53] the distance transform is differentiable almost everywhere (away of the medial axis) and satisfies the Eikonal equation:", + "bbox": [ + 511, + 611, + 905, + 685 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\| \\nabla d \\| = 1 \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 673, + 686, + 903, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2 Let $f: \\Omega \\to \\mathbb{R}^+$ a smooth function on $\\Omega$ . Then $f$ is said to be invariant by affine transformation if:", + "bbox": [ + 511, + 708, + 903, + 739 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\forall A \\in \\mathcal {A}, \\forall x \\in \\Omega , f (A (x)) = f (x) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 746, + 903, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2 restates the constancy assumption mathematically: if an affine transformation $A$ is applied to a convex shape $\\Omega$ , then any changes introduce by $f$ on $\\Omega$ should manifest identically in $A(\\Omega)$ .", + "bbox": [ + 511, + 768, + 905, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1 $\\delta_{\\Omega}$ is the unique solution, up to an isomorphism, to the following optimization problem:", + "bbox": [ + 511, + 835, + 905, + 867 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {f \\in \\mathcal {I} _ {\\mathcal {A}}} \\int_ {\\Omega} (f (x) - \\bar {f}) ^ {2} d x \\quad s. t. \\| \\nabla f \\| = 1 \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 872, + 903, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21870", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\overline{f}$ being the average value of $f$ over $\\Omega$ . A proof of Theorem 1 is given in Appendix A.1. This property is particularly significant because it demonstrates that, under the constraint of the Eikonal equation, the optimal method for increasing variance within a compact convex shape while adhering to the constancy assumption is to use the distance transform. For this reason, in our practical experiments, we propose to introduce variance within image instances primarily via the distance transform; further details can be found in Section 4.2.", + "bbox": [ + 89, + 90, + 480, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Eikonal equation constraint acts as a standardization condition, promoting smooth variation. Without this constraint, partitioning the shape in uniform black and white regions would also maximize the variance. Our goal is instead to ensure a more even spread of variance throughout the shape, making the variance-augmented image reprojection more informative for better depth and ego-motion learning. We also explore carefully relaxing this constraint to investigate well-chosen functions of the distance transform that could allow for greater variance within instances while still maintaining an acceptable level of smoothness:", + "bbox": [ + 89, + 242, + 482, + 407 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} = \\left\\{f: x \\mapsto g \\circ \\delta_ {\\Omega} (x) \\mid g \\in \\mathcal {C} ^ {\\infty} \\left(\\mathbb {R} ^ {+}, \\mathbb {R} ^ {+}\\right), \\| g \\| _ {\\infty} \\leq 1 \\right\\} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 417, + 480, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A thorough comparison of depth estimation results across different functions is provided in Appendix C.2. The bound on the norm ensures that variance is not artificially amplified by larger values, maintaining consistency and fairness in comparisons.", + "bbox": [ + 89, + 450, + 482, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We show that the maximal variance property can be retrieved with a simple toy experiment in A.3.", + "bbox": [ + 89, + 525, + 482, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Convergence properties", + "text_level": 1, + "bbox": [ + 89, + 564, + 308, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notations Let $(I, \\tilde{I})$ be two successive images of a camera in movement and $(x, y)$ a training instance where $x$ represents a pixel of $I$ and $y$ be the corresponding pixel in $\\tilde{I}$ . Let us assume also that all pixels lie inside a convex shape $\\Omega$ defined by a smooth and closed contour $\\partial \\Omega$ . Our goal is to learn the model $\\Phi(x; \\theta)$ to predict the new pixel location $\\hat{y}$ . Let us define the cost function $l$ using functions from $\\mathcal{F}$ (Equation 5):", + "bbox": [ + 89, + 585, + 483, + 708 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl: (\\hat {y}, y) \\mapsto (f (\\hat {y}) - f (y)) ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 717, + 480, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $f\\in \\mathcal{F}$", + "bbox": [ + 204, + 739, + 295, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 3 A function $f: \\Omega \\to \\mathbb{R}^+$ is $\\alpha$ -Lipschitz if:", + "bbox": [ + 89, + 765, + 449, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\forall u, v \\quad | f (u) - f (v) | \\leq \\alpha \\| u - v \\| \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 791, + 480, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 4 A function $f: \\Omega \\to \\mathbb{R}^+$ is $\\beta$ -smooth if:", + "bbox": [ + 89, + 816, + 437, + 834 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\forall u, v \\quad \\| \\nabla f (u) - \\nabla f (v) \\| \\leq \\beta \\| u - v \\| \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 843, + 480, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The fact that $f$ is a function of the distance transform makes it possible to prove:", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 2 The loss function $l$ is, with respect to the first argument $\\hat{y}$ :", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $\\alpha$ -Lipschitz", + "- $\\beta$ -smooth", + "- strongly convex if a regularization term $\\eta \\| y\\|^2$ is added." + ], + "bbox": [ + 513, + 125, + 893, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The constants $\\alpha, \\beta, \\eta$ depend on the curvature at the projection onto the boundary $\\partial \\Omega$ . A proof of Theorem 2 is given in Appendix A.2. Those properties brought by the distance transform induce great convergence properties as described in [1, 27, 36]. Amongst others: a bounded generalization error and a uniform stability of the stochastic gradient descent. Note that the SSIM + colour function can also be Lipschitz and/or uniform under very strict conditions. However their constants, unlike with the distance transform, depend on the size of the shape $\\Omega$ and can thus exceeds all bounds as mentioned in [57].", + "bbox": [ + 511, + 185, + 906, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We show that the convergence property can be retrieved with a simple toy experiment in A.4.", + "bbox": [ + 511, + 351, + 905, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 401, + 604, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our framework for self-supervised MDE (Figure 2) consists of three networks: Depth $D_{\\theta}$ , Ego-motion $T_{\\alpha}$ and edge $E_{\\delta}$ . For the sake of readability, we will omit the parameter symbols $\\theta, \\alpha, \\delta$ in the remainder of the section. Hyperparameters in the loss functions will also be omitted here and detailed in Appendix F.", + "bbox": [ + 511, + 429, + 905, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Pre-semantic contours", + "text_level": 1, + "bbox": [ + 513, + 535, + 720, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our edge network renders values $E(p)$ in the range [0, 1], where a higher value indicates a greater likelihood that the pixel $p$ corresponds to an edge. To generate pseudo-labels for supervising the edge network, we utilize depth and surface normal estimation, which is derived from depth following the methods described in [83, 85]. We discuss complementary insights of depth and surface normal for edge detection in Appendix D.1.", + "bbox": [ + 511, + 560, + 905, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depth & Normal Pseudo-Labels: Depth and surface normals are smooth and thus prone to exhibiting zero-crossings of their Laplacian $\\Delta D$ (resp. $\\Delta N$ ) around semantic boundaries, defined as follows:", + "bbox": [ + 511, + 681, + 905, + 741 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {D} (p) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\Delta D (p) \\Delta D (p + (1, 0)) < 0 \\\\ 1 & \\text {i f} \\Delta D (p) \\Delta D (p + (0, 1)) < 0 \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 756, + 903, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Same applies to normal Laplacian $\\Delta N$ . The zero-crossing mask is then dilated ( $3 \\times 3$ structuring element), and multiplied by the magnitude of the gradient to create the pseudolabels. For this, the depth gradient is normalized to avoid", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21871", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/afbebc2592a2f5d7e8c42e35b523cfefcb94d3c337e74fa2f1c6e822c77b86af.jpg", + "image_caption": [ + "Figure 2. Diagram showing the two main steps of our pipeline: estimation of the different modalities followed by the re-projection on the variance augmented image using depth and ego-motion. T indicates the time frame. R, t respectively rotation and translation predictions." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 500, + 262 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e24a5a4f4751674e386e8ed84e3f6d5ff00f23ea0f1ec7ae35f22cef70d5e4ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 903, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bias towards large depth values:", + "bbox": [ + 89, + 330, + 305, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla D _ {x} (p) = \\frac {D (p + (0 , 1)) - D (p)}{\\max (D (p + (0 , 1)) , D (p))}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 354, + 480, + 391 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\| \\nabla D (p) \\| = | \\nabla D _ {x} (p) | + | \\nabla D _ {y} (p) |.\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 393, + 380, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The usual normal gradient, which is often too noisy, is replaced with a dot-product measure:", + "bbox": [ + 89, + 421, + 482, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\nabla N _ {x} (p) \\right| = 1 - \\left| \\langle \\vec {N} (p + (0, 1)), \\vec {N} (p) \\rangle \\right|, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 462, + 482, + 487 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\nabla N (p) \\right\\| = \\left| \\nabla N _ {x} (p) \\right| + \\left| \\nabla N _ {y} (p) \\right|.\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 484, + 380, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The weighted depth and normal pseudo-labels employed to supervise the edge network are then defined as follows:", + "bbox": [ + 89, + 512, + 482, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {d} (p) = \\frac {Z _ {D} (p) \\| \\nabla D (p) \\|}{\\sum_ {p} Z _ {D} (p) \\| \\nabla D (p) \\|} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 553, + 482, + 592 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {n} (p) = \\frac {Z _ {N} (p) \\| \\nabla N (p) \\|}{\\sum_ {p} Z _ {N} (p) \\| \\nabla N (p) \\|} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 601, + 482, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A contrastive loss is also applied to encourage edge values to move toward either zero or one. This is achieved using a binary cross-entropy loss, where pseudo ground truth $Y(p) \\in \\{0,1\\}$ are created by thresholding edge estimations $E(p)$ at 0.5:", + "bbox": [ + 89, + 646, + 483, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = - \\sum_ {p} \\left(Y (p) \\log (E (p)) + (1 - Y (p)) \\log (1 - E (p))\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 732, + 503, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the loss function employed to supervise the edge network is defined as follows:", + "bbox": [ + 89, + 780, + 482, + 809 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {e d g e}} = \\sum_ {p} \\left(w _ {d} (p) + w _ {n} (p)\\right) \\left(1 - E (p)\\right) + \\mathcal {L} _ {c} + \\mathcal {L} _ {e}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 819, + 464, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\mathcal{L}_e = \\sum_p E(p)^2$ to avoid trivial solutions. (15)", + "bbox": [ + 142, + 854, + 480, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discontinuity Preserving Smoothing: Incorporating an edge-aware smoothness loss is a widely used practice in training depth networks [73, 83]:", + "bbox": [ + 511, + 330, + 906, + 377 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\sum_ {p} \\left(| \\nabla D _ {x} (p) | e ^ {- | \\nabla I _ {x} (p) |} + | \\nabla D _ {y} (p) | e ^ {- | \\nabla I _ {y} (p) |}\\right) \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 398, + 903, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, this approach has two drawbacks: first, boundaries are estimated using the image gradient, which is not optimal; second, there is no mechanism to effectively manage discontinuities. To address these issues, we propose a new smoothness loss exploiting our pre-semantic contours defined as follows:", + "bbox": [ + 511, + 446, + 906, + 536 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s _ {1}} = \\sum_ {p} \\| \\nabla D (p) \\| e ^ {- \\alpha E (p)},\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 559, + 730, + 593 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s _ {2}} = \\sum_ {p} \\log \\left(1 + e ^ {- \\| \\nabla D (p) \\|}\\right) \\left(1 - e ^ {- \\alpha E (p)}\\right), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 595, + 903, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\mathcal {L} _ {s _ {1}} + \\mathcal {L} _ {s _ {2}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 633, + 655, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The component $\\mathcal{L}_{s_1}$ replace the former equation 16 to encourage depth smoothness inside semantic instances, while $\\mathcal{L}_{s_2}$ prevents diffusion and promotes significant depth changes around boundaries. In this manner, both the edge and depth networks benefit from each other: the depth estimation supervises the edge network, while the edge network enhances depth estimation through the smoothness loss.", + "bbox": [ + 511, + 674, + 906, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Post-processing: The map produced by the edge network consists of continuous values ranging from zero to one. To obtain thin pre-semantic contours $\\hat{\\mathcal{C}}(I)$ , we apply a post-processing pipeline that includes hysteresis thresholding, non-maximum suppression, morphological closing, contour filtering, and contour closing. More details can be found in Appendix D.3 and Appendix D.2 discusses the advantages of our approach compared to the competitive work [84].", + "bbox": [ + 511, + 780, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21872", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. A variance augmented image for depth learning", + "text_level": 1, + "bbox": [ + 89, + 90, + 483, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With $\\widehat{\\mathcal{C}}(I)$ the thin pre-semantic contour of $I$ obtained after edge network estimation and subsequent post-processing, we then compute the distance transform map $\\delta_{\\mathrm{I}}$ defined as:", + "bbox": [ + 89, + 113, + 483, + 161 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {\\mathrm {I}} (p) = \\min _ {x \\in \\hat {\\mathcal {C}} (I)} (\\| x - p \\|) \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 170, + 483, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "See Appendix C.1 for the details of the algorithm. The resulting distance transform map is concatenated with the input image to create a 'variance-augmented' image. The objective is then to learn depth and ego-motion in order to identify corresponding pixels across successive frames, and reconstruct the upcoming structure-augmented image in the most effective manner, as illustrated on Figure 2. The reconstruction of the original input image $\\widehat{\\mathbf{l}}$ is evaluated using $\\Psi$ , the standard combination of colour difference and SSIM, while the reconstruction of the distance transform map is assessed with a $\\mathrm{L}_1$ loss:", + "bbox": [ + 89, + 205, + 483, + 371 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p h o t o}} = \\Psi (\\widehat {I}, I), \\text {a n d} \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 381, + 480, + 414 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d i s t}} = \\left\\| \\widehat {\\delta_ {\\mathrm {I}}} - \\delta_ {\\mathrm {I}} \\right\\| _ {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 410, + 339, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The smoothness loss is added to get the final supervision signal of the depth network:", + "bbox": [ + 89, + 440, + 483, + 470 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {d e p t h}} = \\mathcal {L} _ {\\text {d i s t}} + \\mathcal {L} _ {\\text {p h o t o}} + \\mathcal {L} _ {s} \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 479, + 482, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An improved distance transform: This approach can be readily extended to the family of one-dimensional functions $\\mathcal{F} = \\{p\\mapsto g\\circ \\delta_{\\mathrm{I}}(p)\\}$ (as defined in Equation 5):", + "bbox": [ + 89, + 507, + 483, + 554 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t}} = \\left\\| g \\circ \\widehat {\\delta_ {\\mathrm {I}}} - g \\circ \\delta_ {\\mathrm {I}} \\right\\| _ {1}. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 561, + 482, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We provide an analysis of the depth estimation results for various functions in Appendix C.2. To take this a step further, we propose extending the distance transform into the $n$ -dimensional space by mapping the distance transform values to an $n$ -dimensional random walk process denoted as $\\mathrm{RW}_n$ :", + "bbox": [ + 89, + 592, + 483, + 681 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t}} = \\sum_ {p} \\left\\| \\mathrm {R W} _ {n} \\circ \\widehat {\\delta_ {\\mathrm {I}}} (p) - \\mathrm {R W} _ {n} \\circ \\delta_ {\\mathrm {I}} (p) \\right\\| _ {1} \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 690, + 482, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using this approach, we can maintain the validity of the constancy assumption while increasing the variance even more. Additionally, the dimensional parameter $n$ determines the significance of the distance transform map reconstruction within the total loss function. It introduces a tradeoff between $\\mathcal{L}_{\\mathrm{dist}}$ and $\\mathcal{L}_{\\mathrm{photo}}$ : the higher the dimension, the greater the emphasis on the artificially introduced structure compared to the natural structure found in the image; see Appendix C.3 for more details on the random walk mapping. Our experiments in Appendix C.2 show that $n = 3$ provides the best results.", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 89, + 645, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 513, + 114, + 616, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on multiple datasets:", + "bbox": [ + 511, + 138, + 816, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "KITTI [20] includes diverse urban, rural, and highway scenes. We follow the standard data split defined by [13] and apply the pre-processing steps outlined in [91] to filter static frames, resulting in a training set of 39810 images and 697 images for testing. Results are reported using the evaluation protocol from [91].", + "bbox": [ + 511, + 152, + 903, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cityscapes [9] offers urban scenes with dynamic objects, challenging depth estimation. The training set consists of 22973 images, with evaluation on 1525 test images following [5, 41]. Edge detection is assessed using the contour evaluation from [84] on 500 validation frames.", + "bbox": [ + 511, + 244, + 905, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Waymo [69] is a large and diverse autonomous driving dataset that captures dynamic urban scenes under various environmental conditions, including nighttime and different weather scenarios. We sample 100 000 image pairs from 1 000 front camera video sequences for training and evaluate on 1500 pairs from 150 sequences, following [41].", + "bbox": [ + 511, + 319, + 905, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NYUv2 [67] consists of 464 indoor video sequences captured with a Kinect sensor. We use the official splits (302 training, 33 validation) and evaluate on 654 densely annotated test images, resizing inputs to $320 \\times 256$ pixels as in [44].", + "bbox": [ + 511, + 410, + 903, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ScanNet [10] is a large-scale RGB-D dataset with 2.5 million images across $1500+$ scenes. We use it for zero-shot evaluation, following [16, 44], with test images resized to $320 \\times 256$ pixels.", + "bbox": [ + 511, + 484, + 905, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 513, + 556, + 730, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We emphasize that our method is compatible with any DNN architecture and input image size. For fair comparison, our experiments are done with a $256 \\times 832$ resolution and a UNet [62] structure, following [91], unless stated otherwise. The depth and edge networks share a ResNet [29] encoder. Our baseline, CoopNet [28], handles moving objects—especially useful in Cityscapes—via self-supervised flow. The pose network, based on a ResNet encoder, outputs a 6-DoF vector. All encoders use ResNet-50 backbones initialized with ImageNet [63] weights. Training runs for 30 epochs with batch size 4, initial learning rate of $10^{-4}$ (decreasing to $10^{-5}$ after 20 epochs), and standard data augmentation from [22]. Loss weights, tuned via grid search, are detailed in Appendix F. Our PyTorch [58] implementation trains on a single NVIDIA RTX A5000 GPU using Adam [38] with $\\beta_{1} = 0.99$ and $\\beta_{2} = 0.999$ and takes 12 hours.", + "bbox": [ + 511, + 579, + 906, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Results", + "text_level": 1, + "bbox": [ + 513, + 845, + 607, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Depth Quantitative results on KITTI and Cityscapes are presented in Table 1, where we compare against the most", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21873", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ee3e9691c6c1eaed2d2aea54976e283a353939db56bc27f4cd15ae3c2040d484.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLower is better ↓Higher is better ↑
Abs RelSq RelRMSERMSE logδ1δ2δ3
Monodepth2[22]0.1100.8314.6420.1870.8830.9620.982
Guizilini et al.†[24] (R50)0.1130.8314.6630.1890.8780.9710.983
Shu et al.*[66](R50)0.1080.7924.6330.1840.8830.9610.983
Ours (R50)0.1040.7254.4530.1800.8850.9620.983
MonoViT[89]0.0990.7084.3720.1750.9000.9670.984
Ours + MonoViT0.0920.6744.3000.1650.9270.9670.984
HR-Depth[49]0.1090.7924.6320.1850.8840.9620.983
RA-Depth[30]0.0960.6134.2160.1710.9030.9680.985
DIFFNet[90]0.1020.7644.4830.1800.8960.9650.983
Guizilini et al.†[24]0.1000.7614.2700.1750.9020.9650.982
Ours+DinoV20.0820.6044.1080.1620.9280.9680.985
Lego (paper)0.1541.2726.0120.230N/AN/AN/A
Lego★ (R50)0.1150.8554.7890.1950.8760.9570.979
Lego★ (R50) + Ldist0.1100.8054.6060.1880.8820.9600.982
Ours (R50)0.1040.7254.4530.1800.8850.9620.983
Struct2Depth†[4]0.1451.7377.280.2050.8130.9420.978
LearnK†[41]0.1271.3306.960.1950.8300.9470.981
Li et al.†[41]0.1191.2906.980.1900.8460.9510.980
CoopNet[28]0.1211.4437.010.1900.8460.9510.980
Ours (R50)0.1151.2216.790.1860.8500.9550.981
", + "bbox": [ + 94, + 88, + 478, + 369 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5694788ab0dac38ddbd7a5416afb2ab5d2f5d9ce0973763c64bea1172dfe0e87.jpg", + "table_caption": [ + "Table 1. Results of depth estimations on KITTI 2015 (first four blocks) and Cityscapes (last block). For each metric the best result is displayed in bold and the second-best is underlined. Light blue metrics indicates the most challenging metrics. $\\dagger$ denotes the use of an off-the-shelf semantic algorithm and * indicates results reproduced by us at the input resolution of $256 \\times 832$ . Block#1 and #2 compares methods using similar encoder backbones: Resnet 50 (R50) in block#1 and ViT in block#2. Block#3 compares SOTA depth models without restrictions on architecture or additional inputs, Block#4 compares to Lego." + ], + "table_footnote": [], + "table_body": "
MethodDataset (Train / Test)Lower is better ↓
Abs RelSq RelRMSERMSE log
LearnK[41]Waymo / Waymo0.1571.5317.0900.205
OursWaymo / Waymo0.1251.2785.9900.189
GLNet[7]Cityscapes / KITTI0.1291.0445.3610.212
CoopNet[28]Cityscapes / KITTI0.1251.1575.2510.209
OursCityscapes / KITTI0.1160.8924.8920.195
", + "bbox": [ + 94, + 537, + 480, + 630 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c013e19dfec251c4461a02c1f12b23cbc40a8e66bb86a58bff9d06924145a7b0.jpg", + "table_caption": [ + "Table 2. Results on Outdoor datasets. Block#1: in-domain evaluation. Block#2: out-of-domain evaluation." + ], + "table_footnote": [], + "table_body": "
MethodDataset (Train / Test)Abs RelRMSEδ1δ2δ3
MonoIndoor++[44]NYUv2 / NYUv20.1320.5170.8340.9610.990
IndoorDepth[16]NYUv2 / NYUv20.1260.4940.8450.9650.991
OursNYUv2 / NYUv20.1150.4580.8590.9700.992
IndoorDepth[16]NYUv2 / ScanNet0.1530.3730.7860.9500.988
MonoIndoor++[44]NYUv2 / ScanNet0.1380.3470.8100.9670.993
OursNYUv2 / ScanNet0.1270.3120.8430.9700.993
", + "bbox": [ + 94, + 686, + 480, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "competitive methods. Blocks #1 and #2 demonstrate that, when using the same encoder backbone, our method outperforms others by a significant margin across nearly all metrics. When no architectural constraints are imposed, we", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/be2a428613e52a3a8bd9d654060a77a94c7230759aee89c34352cd7e15ff02db.jpg", + "table_caption": [ + "Table 3. Results on Indoor datasets. Block#1: in-domain evaluation. Block#2: out-of-domain evaluation." + ], + "table_footnote": [], + "table_body": "
RW3LdistLsEθAbs Rel (↓)RMSE log (↓)δ1(↑)
0.1130.190.878
✓0.1120.1890.880
✓✓0.1100.1870.881
✓✓✓0.1060.1830.884
✓✓✓✓0.1040.1800.885
", + "bbox": [ + 526, + 88, + 903, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study on KITTI 2015 of losses mentioned in Section 4.2. Note that $E_{\\theta}$ is part of the ablation as it does improve the depth metrics and shares the same encoder as the depth network. $\\mathrm{RW}_3$ is for 3d random walk encoding of distance transform as defined in Section 4.2.", + "bbox": [ + 511, + 174, + 906, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "employ a pre-trained DinoV2 [56] encoder (further details in Appendix F.2), yielding the results in block #3. For fairness, we compare against $\\mathsf{LEGO}^{\\star}$ in block #4, an enhanced version of LEGO that incorporates a ResNet-50 backbone and up-to-date training techniques. Adding our novel distance transform loss $\\mathcal{L}_{\\mathrm{dist}}$ into $\\mathsf{LEGO}^{\\star}$ greatly improves results, showing once again the great importance of this loss component, but still underperforms ours, likely due to the superior quality of our contours.", + "bbox": [ + 509, + 271, + 906, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also provide strong out-of-domain results in Table 2 and Table 3, demonstrating competitive performance on both indoor and outdoor datasets. Finally, an ablation study in Table 4 evaluates the impact of the different loss components introduced in our approach and in Appendix C.2 for different functions from the family $\\mathcal{F}$ as defined in Section 3.1. It can be seen that the reprojection on the distance transform via $\\mathcal{L}_{\\mathrm{dist}}$ brings a decisive improvement. We provide a qualitative comparison of results in Figure 3. We can see that our method gives better results: it provides thinner object, is sharper around moving objects, does not suffer from smoothness issues for large objects and is not noisy in textured areas. More qualitative and quantitative results can be found in Appendix B.1.", + "bbox": [ + 509, + 407, + 908, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Flow & Odometry As discussed in Section 5.2 we use an optical flow to handle moving objects following [28]. Its results can be found in Appendix B.2. Results for odometry can be found in Appendix B.3. Our method significantly outperforms the baseline [28] on both optical flow and odometry metrics.", + "bbox": [ + 511, + 633, + 906, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Contour evaluation", + "text_level": 1, + "bbox": [ + 511, + 734, + 700, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate contour detection in Table 5 using the same metrics and data as [84]: Optimal Dataset Scale, Optimal Image Scale, and Average Precision, computed on 500 Cityscapes validation images. Our method significantly outperforms LEGO, leveraging (1) depth & normal pseudolabels with Laplacian zero-crossing insights for better contour alignment, (2) the discontinuity-preserving smoothing loss $\\mathcal{L}_s$ , and (3) the contrastive loss $\\mathcal{L}_c$ . Qualitative examples are in Figure 1, with more in Appendix D.", + "bbox": [ + 509, + 756, + 908, + 893 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21874", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/771318ed04d5bfb0f1facc9876755a4eaeb9b5d9e7315940173f08b7db155240.jpg", + "image_caption": [ + "Figure 3. Qualitative comparisons for different methods. First row: colour images. Second row: Monodepth2 [22]. Second row: CoopNet [28]. Third row: FeatDepth [66]. Last row: Ours. We compare qualitative results in four complex situations. First column: a thin object. Second column: a moving object. Third column: a large object with smoothness issue. Last column: Textured areas." + ], + "image_footnote": [], + "bbox": [ + 89, + 99, + 906, + 401 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/57bbf687abddf0fd93de0dfc51f4d397a1b1a75f8b98a5f25c95adcf132eb152.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsODSOISAP
Lego (paper)0.7100.7310.729
Lego*0.7090.7360.734
Depth & Normal pseudo-labels0.7550.7750.760
Depth & Normal pseudo-labels + Ls0.7600.7820.766
Ours = Depth & Normal pseudo-labels + Ls + Lc0.7620.7890.770
", + "bbox": [ + 99, + 476, + 478, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparison for contour estimation. Block#1: Lego results. Block#2: Ablation study of losses in contour estimation training. Experiments done on Cityscapes.", + "bbox": [ + 89, + 555, + 482, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Constancy assumption validity", + "text_level": 1, + "bbox": [ + 89, + 626, + 362, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the constancy assumption validity of our method, compared to [66] that uses deep features from the ResNet-50 encoder in the re-projection. To do so, we use the KITTI MOTS [72] dataset to track a specific point of an instance along different trajectories and compute the normalized temporal variance. We show that our distance transform satisfies much better the constancy assumption than any layers of the ResNet-50 (see Figure 4). More details on the evaluation protocol are given in Appendix E.", + "bbox": [ + 89, + 648, + 483, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 800, + 209, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a conclusion, this paper presents a novel self-supervised approach to monocular depth estimation, addressing challenges in low-texture regions by applying a distance transform over pre-semantic contours. This technique enhances depth prediction accuracy by increasing discrimini", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bda0b3345ff6e571496b7f43ea9f3144caf766d7eba9c5a1b5e05446371c42c1.jpg", + "image_caption": [ + "Figure 4. Comparison of normalized temporal variance of different features for the centre pixel of an instance tracked along different sequence of KITTI MOTS [72]." + ], + "image_footnote": [], + "bbox": [ + 527, + 494, + 867, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "native power where traditional photometric losses struggle. Both theoretical analysis and extensive experiments validate the effectiveness of the proposed method, which outperforms conventional self-supervised approaches.", + "bbox": [ + 511, + 750, + 903, + 810 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In future work, we aim to introduce consistent variance in low-texture regions without pre-semantic contours. Additionally, we plan to explore whether our framework can leverage variance-augmented images to learn from diverse datasets, ultimately developing a foundation model in a fully self-supervised manner.", + "bbox": [ + 511, + 811, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21875", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ali Akbari, Muhammad Awais, Manijeh Bashar, and Josef Kittler. How does loss function affect generalization performance of deep learning? application to human age estimation. In International Conference on Machine Learning, pages 141-151. PMLR, 2021. 4", + "[2] Abhishek Badki, Alejandro Troccoli, Kihwan Kim, Jan Kautz, Pradeep Sen, and Orazio Gallo. Bi3d: Stereo depth estimation via binary classifications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1600-1608, 2020. 1", + "[3] Jiawang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised scale-consistent depth and ego-motion learning from monocular video. Advances in neural information processing systems, 32, 2019. 6", + "[4] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Depth prediction without the sensors: Leveraging structure for unsupervised learning from monocular videos. In Proceedings of the AAAI conference on artificial intelligence, pages 8001-8008, 2019. 2, 7, 3, 4", + "[5] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Unsupervised monocular depth and ego-motion learning with structure and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 6", + "[6] Tian Chen, Shijie An, Yuan Zhang, Chongyang Ma, Huayan Wang, Xiaoyan Guo, and Wen Zheng. Improving monocular depth estimation by leveraging structural awareness and complementary datasets. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 90-108. Springer, 2020. 1", + "[7] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 1, 2, 7, 5", + "[8] Zhao Chen, Vijay Badrinarayanan, Gilad Drozdov, and Andrew Rabinovich. Estimating depth from rgb and sparse sensing. In Proceedings of the European Conference on Computer Vision (ECCV), pages 167-182, 2018. 3", + "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 2, 6", + "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6", + "[11] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In Proceedings of the European conference on computer vision (ECCV), pages 562-578, 2018. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Tom van Dijk and Guido de Croon. How do neural networks see depth in single images? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2183-2191, 2019. 1", + "[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 1, 6", + "[14] F El Jamiy and R Marsh. Survey on depth perception in head mounted displays: distance estimation in virtual reality, augmented reality, and mixed reality. iet image processing 13 (5), 707-712 (2019), 2018. 1", + "[15] Sara Elkerdawy, Hong Zhang, and Nilanjan Ray. Lightweight monocular depth estimation model by joint end-to-end filter pruning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 4290-4294. IEEE, 2019. 1", + "[16] Chao Fan, Zhenyu Yin, Yue Li, and Feiqing Zhang. Deeper into self-supervised monocular indoor depth estimation. arXiv preprint arXiv:2312.01283, 2023. 6, 7", + "[17] Philipp Fischer, Alexey Dosovitskiy, Eddy Ilg, Philip Häusser, Caner Hazirbaş, Vladimir Golkov, Patrick Van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. arXiv preprint arXiv:1504.06852, 2015. 5", + "[18] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 1", + "[19] Ravi Garg, Vijay Kumar Bg, Gustavo Carneiro, and Ian Reid. Unsupervised cnn for single view depth estimation: Geometry to the rescue. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 740-756. Springer, 2016. 1", + "[20] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 6", + "[21] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 1", + "[22] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 2, 6, 7, 8", + "[23] Ariel Gordon, Hanhan Li, Rico Jonschkowski, and Anelia Angelova. Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8977-8986, 2019. 1, 2", + "[24] Vitor Guizilini, Rui Hou, Jie Li, Rares Ambrus, and Adrien Gaidon. Semantically-guided representation learn" + ], + "bbox": [ + 93, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21876", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing for self-supervised monocular depth. arXiv preprint arXiv:2002.12319, 2020. 2, 7", + "[25] Hyowon Ha, Sunghoon Im, Jaesik Park, Hae-Gon Jeon, and In So Kweon. High-quality depth from uncalibrated small motion clip. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5413-5421, 2016. 1", + "[26] David Hafner, Oliver Demetz, and Joachim Weickert. Why is the census transform good for robust optic flow computation? In Scale Space and Variational Methods in Computer Vision: 4th International Conference, SSVM 2013, Schloss Seggau, Leibnitz, Austria, June 2-6, 2013. Proceedings 4, pages 210-221. Springer, 2013. 3", + "[27] Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International conference on machine learning, pages 1225-1234. PMLR, 2016. 4", + "[28] Marwane Hariat, Antoine Manzanera, and David Filliat. Rebalancing gradient to improve self-supervised co-training of depth, odometry and optical flow predictions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1267-1276, 2023. 2, 6, 7, 8, 4, 5", + "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[30] Mu He, Le Hui, Yikai Bian, Jian Ren, Jin Xie, and Jian Yang. Ra-depth: Resolution adaptive self-supervised monocular depth estimation. In European Conference on Computer Vision, pages 565-581. Springer, 2022. 7", + "[31] Stuart B Heinrich and Wesley E Snyder. Improved edge awareness in discontinuity preserving smoothing. arXiv preprint arXiv:1103.5808, 2011. 3", + "[32] Junjie Hu, Yan Zhang, and Takayuki Okatani. Visualization of convolutional neural networks for monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3869-3878, 2019. 1", + "[33] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2462-2470, 2017. 5", + "[34] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 2", + "[35] Wei Jiang, Weiwei Sun, Andrea Tagliasacchi, Eduard Trulls, and Kwang Moo Yi. Linearized multi-sampling for differentiable image transformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2988-2997, 2019. 2", + "[36] Kenji Kawaguchi and Qingyun Sun. A recipe for global convergence guarantee in deep neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 8074-8082, 2021. 4", + "[37] Ue-Hwan Kim, Gyeong-Min Lee, and Jong-Hwan Kim. Revisiting self-supervised monocular depth estimation. In International Conference on Robot Intelligence Technology and Applications, pages 336-350. Springer, 2021. 2" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[39] Yevhen Kuznietsov, Marc Proesmans, and Luc Van Gool. Comoda: Continuous monocular depth adaptation using past experiences. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2907-2917, 2021. 2", + "[40] Seokju Lee, Sunghoon Im, Stephen Lin, and In So Kweon. Learning residual flow as dynamic motion from stereo videos. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1180-1186. IEEE, 2019. 2", + "[41] Hanhan Li, Ariel Gordon, Hang Zhao, Vincent Casser, and Anelia Angelova. Unsupervised monocular depth learning in dynamic scenes. In Conference on Robot Learning, pages 1908-1917. PMLR, 2021. 2, 6, 7", + "[42] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. Rtm3d: Real-time monocular 3d detection from object keypoints for autonomous driving. In European Conference on Computer Vision, pages 644-660. Springer, 2020. 1", + "[43] Pengzhi Li, Yikang Ding, Haohan Wang, Chengshuai Tang, and Zhiheng Li. The devil is in the edges: Monocular depth estimation with edge-aware consistency fusion. arXiv preprint arXiv:2404.00373, 2024. 2", + "[44] Runze Li, Pan Ji, Yi Xu, and Bir Bhanu. Monoindoor++: Towards better practice of self-supervised monocular depth estimation for indoor environments. IEEE Transactions on Circuits and Systems for Video Technology, 33(2):830-846, 2022. 6, 7", + "[45] Rui Li, Danna Xue, Shaolin Su, Xiantuo He, Qing Mao, Yu Zhu, Jinqiu Sun, and Yanning Zhang. Learning depth via leveraging semantics: Self-supervised monocular depth estimation with both implicit and explicit semantic guidance. Pattern Recognition, 137:109297, 2023. 2, 3", + "[46] Yin Li, Manohar Paluri, James M Rehg, and Piotr Dólar. Unsupervised learning of edges. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1619-1627, 2016. 3", + "[47] Liang Liu, Guangyao Zhai, Wenlong Ye, and Yong Liu. Unsupervised learning of scene flow estimation fusing with local rigidity. In IJCAI, pages 876-882, 2019. 2", + "[48] Hengjie Lu, Shugong Xu, and Shan Cao. Sgtbn: generating dense depth maps from single-line lidar. IEEE Sensors Journal, 21(17):19091-19100, 2021. 1", + "[49] Xiaoyang Lyu, Liang Liu, Mengmeng Wang, Xin Kong, Lina Liu, Yong Liu, Xinxin Chen, and Yi Yuan. Hr-depth: High resolution self-supervised monocular depth estimation. In Proceedings of the AAAI conference on artificial intelligence, pages 2294-2301, 2021. 7", + "[50] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3", + "[51] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 6" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21877", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Oskar Natan and Jun Miura. End-to-end autonomous driving with semantic depth cloud mapping and multi-agent. IEEE Transactions on Intelligent Vehicles, 8(1):557-571, 2022. 1", + "[53] Ales Nekvinda and Ludek Zajicek. A simple proof of the rademacher theorem. Časopis pro pěstovánt matematíky, 113 (4):337-341, 1988. 3", + "[54] C Wayne Niblack, Phillip B Gibbons, and David W Capson. Generating skeletons and centerlines from the distance transform. CVGIP: Graphical Models and image processing, 54 (5):420-437, 1992. 2", + "[55] Jim Nilsson and Tomas Akenine-Möller. Understanding ssm. arXiv preprint arXiv:2006.13846, 2020. 1", + "[56] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7", + "[57] Daniel Otero, Davide La Torre, O Michailovich, and Edward R Vrscay. Optimization of structural similarity in mathematical imaging. Optimization and Engineering, 22:2367-2401, 2021. 4", + "[58] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[59] Michael Ramamonjisoa and Vincent Lepetit. Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 2", + "[60] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12240-12249, 2019. 2", + "[61] Alex Rasla and Michael Beyeler. The relative importance of depth cues and semantic edges for indoor mobility using simulated prosthetic vision in immersive virtual reality. In Proceedings of the 28th ACM Symposium on Virtual Reality Software and Technology, pages 1-11, 2022. 1", + "[62] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6, 3", + "[63] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6", + "[64] Faraz Saeedan and Stefan Roth. Boosting monocular depth with panoptic segmentation maps. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3853-3862, 2021. 2" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Abhijeet Shenoi, Mihir Patel, JunYoung Gwak, Patrick Goebel, Amir Sadeghian, Hamid Rezatofighi, Roberto Martin-Martin, and Silvio Savarese. Jrmot: A real-time 3d multi-object tracker and a new large-scale dataset. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10335-10342. IEEE, 2020. 1", + "[66] Chang Shu, Kun Yu, Zhixiang Duan, and Kuiyuan Yang. Feature-metric loss for self-supervised learning of depth and egomotion. In European Conference on Computer Vision, pages 572-588. Springer, 2020. 3, 7, 8", + "[67] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 2, 6", + "[68] Tilo Strutz. The distance transform and its computation. arXiv preprint arXiv:2106.03503, 2021. 2", + "[69] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 6", + "[70] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 1", + "[71] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3", + "[72] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. MOTS: Multi-object tracking and segmentation. In CVPR, 2019. 8, 6, 7, 11", + "[73] Chaoyang Wang, José Miguel Buenapasada, Rui Zhu, and Simon Lucey. Learning depth from monocular videos using direct methods. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2022–2030, 2018. 2, 5", + "[74] Peng Wang, Xiaohui Shen, Bryan Russell, Scott Cohen, Brian Price, and Alan L Yuille. Surge: Surface regularized geometry estimation from a single image. Advances in Neural Information Processing Systems, 29, 2016. 2", + "[75] Yang Wang, Yi Yang, Zhenheng Yang, Liang Zhao, Peng Wang, and Wei Xu. Occlusion aware unsupervised learning of optical flow. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4884-4893, 2018. 2", + "[76] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 1", + "[77] Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1164-1174, 2021. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21878", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[78] Diana Wofk, Fangchang Ma, Tien-Ju Yang, Sertac Karaman, and Vivienne Sze. Fastdepth: Fast monocular depth estimation on embedded systems. In 2019 International Conference on Robotics and Automation (ICRA), pages 6101-6108. IEEE, 2019. 1", + "[79] Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single image depth prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 611-620, 2020. 2, 3", + "[80] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 2", + "[81] Wanpeng Xu, Ling Zou, Lingda Wu, and Zhipeng Fu. Self-supervised monocular depth learning in low-texture areas. Remote Sensing, 13(9):1673, 2021. 3", + "[82] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 7", + "[83] Z Yang, P Wang, W Xu, L Zhao, and R Nevatia. Unsupervised learning of geometry with edge-aware depth-normal consistency. arxiv 2017. arXiv preprint arXiv:1711.03665, 2 (5). 4, 5", + "[84] Zhenheng Yang, Peng Wang, Yang Wang, Wei Xu, and Ram Nevatia. Lego: Learning edge with geometry all at once by watching videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 225-234, 2018. 3, 5, 6, 7", + "[85] Wei Yin, Yifan Liu, Chunhua Shen, and Youliang Yan. Enforcing geometric constraints of virtual normal for depth prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5684-5693, 2019. 4", + "[86] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1983-1992, 2018. 2, 5", + "[87] Xun Yuan and Song Chen. Sad-slam: A visual slam based on semantic and depth information. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4930-4935. IEEE, 2020. 1", + "[88] Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 340-349, 2018. 3", + "[89] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. In 2022 international conference on 3D vision (3DV), pages 668-678. IEEE, 2022. 7", + "[90] Hang Zhou, David Greenwood, and Sarah Taylor. Self-supervised monocular depth estimation with internal feature fusion. In *British Machine Vision Conference (BMVC)*, 2021. 7" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[91] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 1, 2, 3, 6", + "[92] Shengjie Zhu, Garrick Brazil, and Xiaoming Liu. The edge of depth: Explicit constraints between segmentation and depth. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13116-13125, 2020. 2, 3", + "[93] Yuliang Zou, Zelun Luo, and Jia-Bin Huang. Df-net: Unsupervised joint learning of depth and flow using cross-task consistency. In Proceedings of the European conference on computer vision (ECCV), pages 36-53, 2018. 2" + ], + "bbox": [ + 514, + 92, + 905, + 287 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "21879", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/18d0caf1-d799-41ef-95e4-0a0e2105060c_model.json b/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/18d0caf1-d799-41ef-95e4-0a0e2105060c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..371ed84e01dfa3851b2b6c99d4df5255da8ce511 --- /dev/null +++ b/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/18d0caf1-d799-41ef-95e4-0a0e2105060c_model.json @@ -0,0 +1,3051 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.154, + 0.131, + 0.845, + 0.177 + ], + "angle": 0, + "content": "Improved monocular depth prediction using distance transform over pre-semantic contours with self-supervised neural networks" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.203, + 0.844, + 0.259 + ], + "angle": 0, + "content": "Marwane Hariat1, Antoine Manzanera1, David Filliat1 \n1U2IS, ENSTA, Institut Polytechnique de Paris, Palaiseau, France \n{marwane.hariat, antoine.manzanera, david fillliat}@ensta.fr" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.291, + 0.283, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.339, + 0.282, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.387, + 0.282, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.429, + 0.223, + 0.44 + ], + "angle": 0, + "content": "RGB images" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.291, + 0.491, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.339, + 0.49, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.388, + 0.49, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.429, + 0.413, + 0.44 + ], + "angle": 0, + "content": "Depth" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.291, + 0.7, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.338, + 0.699, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.387, + 0.7, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.429, + 0.617, + 0.441 + ], + "angle": 0, + "content": "Edge" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.292, + 0.906, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.339, + 0.905, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.388, + 0.905, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.429, + 0.864, + 0.44 + ], + "angle": 0, + "content": "Distance Transform" + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.453, + 0.9, + 0.468 + ], + "angle": 0, + "content": "Figure 1. The different modalities estimated in our self-supervised framework and used to reduce matching uncertainty between images." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.48, + 0.327, + 0.496 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.512, + 0.486, + 0.786 + ], + "angle": 0, + "content": "Monocular depth estimation (MDE) with self-supervised training approaches struggles in low-texture areas, where photometric losses may lead to ambiguous depth predictions. To address this, we propose a novel technique that enhances spatial information by applying a distance transform over pre-semantic contours, augmenting discriminative power in low texture regions. Our approach jointly estimates pre-semantic contours, depth and ego-motion. The pre-semantic contours are leveraged to produce new input images, with variance augmented by the distance transform in uniform areas. This approach results in more effective loss functions, enhancing the training process for depth and ego-motion. We demonstrate theoretically that the distance transform is the optimal variance-augmenting technique in this context. Through extensive experiments on KITTI, Cityscapes, Waymo, NYUv2 and ScanNet our model demonstrates robust performance, surpassing competing self-supervised methods in MDE." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.222, + 0.83 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The goal of monocular depth estimation (MDE) is to determine the distance of a pixel from the camera focal plane using only a single input image, unlike stereo depth estimation [2, 21, 25], which relies on pairs of images. This" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.907, + 0.617 + ], + "angle": 0, + "content": "eliminates the need for expensive sensors that require high-power resources [42, 48], such as active depth camera or LiDAR, while maintaining low latency and ensuring real-time compliance [15, 78]. This makes MDE a core module in applications such as robot navigation [65, 87], autonomous driving [52, 76] or virtual/augmented reality [14, 61]. Deep Neural Networks (DNNs) are currently the state of the art for depth maps prediction thanks to their ability to exploit all contextual cues of the environment [12, 32]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.909, + 0.891 + ], + "angle": 0, + "content": "Supervised methods for MDE rely on ground-truth annotations. In spite of huge progress [6, 13, 18, 70], these methods are constrained by the cost and quality of the annotations. In this paper, we focus on self-supervised MDE, that uses the structure-from-motion framework [19, 91] and leverages depth and camera ego-motion estimations from DNNs to reconstruct the next frame from the current one. The supervision signal is then provided by the photometric loss, i.e. the difference between the reconstructed frame and the observed one, based on a combination of direct colour comparison and Structural Similarity Index Measure (SSIM) [55]. Although more challenging, the self-supervised paradigm offers greater training flexibility. First, it can be trained using data from unknown cameras [23], as long as data acquisition provides successive frames, such as in YouTube videos [7]. This is particularly relevant given the growing interest in foundation models, which can now leverage large datasets to produce generalizable out-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21868" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "puts [82]. In addition, it is compatible with continual learning, allowing the model to continuously expand its knowledge as more data become available [7, 39]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.482, + 0.317 + ], + "angle": 0, + "content": "However, the problem of MDE using the photometric loss proposed by Zhou et al. [91] is ill-posed due to its reliance on SSIM and colour-based cost functions, which fail to discriminate effectively in low-texture areas. In these regions, multiple candidates can yield similarly low cost function values, making it difficult to determine the correct correspondence. Although interesting work has been done to address various issues such as scale ambiguity [73], gradient locality [22, 35], occlusion [22, 75], moving objects [41, 60], and infinite depth holes [22, 24], the ill-posed nature of this problem has not been thoroughly and rigorously studied. We aim to provide this essential foundation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.483, + 0.621 + ], + "angle": 0, + "content": "The key challenge lies in reducing ambiguity to build a more robust foundation for learning depth and ego-motion. To this end, we propose introducing sufficient variance in texture-less areas, while adhering as closely as possible to 'the constancy assumption': any change introduced in image \\( t \\) should be reproduced identically in image \\( t + 1 \\), essential as the photometric loss relies on correspondences between successive frames. We propose to do this by assigning to each pixel the distance to its nearest edge in the image, commonly known as the 'distance transform' [54, 68]. We demonstrate mathematically that the distance transform is the most effective way of adding variance while respecting the constancy assumption, and that it also leads to improved convergence properties. Since the distance transform is calculated on binary contours, we propose a self-supervised pre-semantic contour estimation method, jointly learned with the depth estimation model, and add this distance map as an additional channel to the input image to produce a 'variance-augmented' image for reconstruction. In summary, our contributions are as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.675, + 0.483, + 0.735 + ], + "angle": 0, + "content": "(1) A self-supervised approach for pre-semantic contour estimation, which (i) generates the edge information necessary for computing the distance transform, and (ii) simultaneously reinforces the depth estimation network." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.736, + 0.483, + 0.781 + ], + "angle": 0, + "content": "(2) A novel variance-augmented image that incorporates the distance transform, enhancing discriminative power in texture-less areas and improving convergence properties" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.483, + 0.825 + ], + "angle": 0, + "content": "(3) A theoretical foundation demonstrating the effectiveness of the distance transform for adding variance under the constancy assumption" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.887 + ], + "angle": 0, + "content": "(4) An extensive experimental validation of our new pipeline, achieving competitive performance on different benchmark datasets: KITTI [20], Cityscapes [9], Waymo [69], NYUv2 [67] and ScanNet [10]" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.675, + 0.483, + 0.887 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.659, + 0.105 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.584 + ], + "angle": 0, + "content": "Self-supervised framework. The work of Zhou et al. [91] is pioneer in self-supervised monocular depth estimation. Their method utilizes depth and ego-motion to find the corresponding pixels between two successive frames and relies on the bilinear sampler [34] to reconstruct the upcoming scene from the current frame. The supervision signal is then derived from the photometric loss, which is a weighted sum of the SSIM and colour cost functions. The effectiveness of this method relies on the constancy assumption, which asserts that object colours remain consistent from frame to frame. This assumption is bound to the hypothesis of constant lighting over Lambertian surfaces, and do not hold in the case of occlusions and moving objects. This method has then been refined in several ways. Kim et al. [37] provide a comprehensive study of the main issues and solutions found in the literature. In [75], occlusions are addressed using a forward warping module that counts the number of bilinear sampling operations for each pixel. Godard et al. [22] compute the occlusion mask by determining the per-pixel minimum of the photometric loss over time. Moving objects are identified through the differences between optical flow and depth/ego-motion in [28, 40, 47, 60, 86, 93], while off-the-shelf semantic algorithms are utilized to localize potential moving objects in [4, 24]. In addition, a residual map is introduced to correct the ego-motion in [23, 41]. Infinite depth holes are filtered through a two-stage training process [24] and are removed using an adaptive cost-volume architecture in [77]. Finally, Godard et al. [22] propose a unified solution that effectively addresses all the aforementioned issues. However, the ill-posed nature of the problem due to unstructured objects remains unaddressed." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.585, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Contour estimation. Semantic contours should follow the boundaries derived from instance segmentation, outlining distinct entities within an image. Such contours would naturally help improve the accuracy of depth maps. [74] and [59] leverage ground-truth annotations to learn depth, normals, and contours, incorporating regularization terms to enforce constraints derived from their geometric relationships. In contrast, [43] employs an off-the-shelf edge prediction algorithm to refine initial depth estimations as a post-processing step. [79], [64] and [45] utilize a pretrained instance segmentation network to compute edges, sampling points on both sides of the boundary to enhance depth sharpness around those edges. [92] introduced a morphing strategy to align depth borders computed with a threshold gradient on the depth with semantic borders obtained from a pre-trained semantic algorithm. While these techniques demonstrate significant improvements, they assume the edges as provided and are therefore not fully self-supervised, in addition to not being GPU friendly. Significant advances have been made in supervised edge map estimation [11, 80]; however, the self-supervised approach re" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21869" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.422 + ], + "angle": 0, + "content": "mains an ambitious and challenging task. [46] introduces a reinforcement training procedure in which an optical flow network is trained to predict the displacement of each pixel between two frames, based on ground truth point correspondences. Subsequently, an edge detector is updated to align its predicted edges of the image with the predicted edges of the estimated flow map. Still, this method is not self-supervised. A work closer to ours [84] proposes a fully self-supervised edge prediction framework that utilizes the positive part of the second derivatives of depth and the first-order gradient of normal to surface to feed an edge network. However, this approach employs suboptimal loss functions, leading to biased contours that favour shorter distances. Furthermore, it does not utilize the edges to enhance depth, nor does it include any post-processing to create refined contours. We propose new cost functions for training the edge network, resulting in significant performance gains. Additionally, we introduce a post-processing step to refine contours and implement an extended edge-aware diffusion smoothness loss [31] that encourages sharp depth discontinuities around estimated boundaries, in the same spirit as [45, 79, 92]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.485, + 0.894 + ], + "angle": 0, + "content": "Targeting the ill-posed optimization problem. As previously mentioned, our objective is to minimize the photometric loss by establishing correspondences between pixels in the current frame and those in the next frame, achieved through the depth and ego-motion parameters. However, this correspondence procedure is not unique, since many pixels share the same colour and exhibit similar local colour statistics as their neighbourhood. This ambiguity is especially pronounced in areas with low texture. The challenge is to compensate for the lack of variation by introducing time-consistent variance across frames, to satisfy the constancy assumption. This reduces the ill-posed nature of the optimization problem and creates better conditions for learning depth and ego-motion. [81, 88] addressed this issue by applying the reconstruction loss to deep features in addition to the usual colour images. Shu et al. [66] took this approach further and designed custom deep features by training an auto-encoder that encourages large and smooth gradient values in feature maps, yielding deep features with both semantic richness and substantial structure. The constancy assumption is rather optimistic for these methods, as demonstrated in our experiments. Indeed, due to their large receptive fields, deep features generally lack consistency from frame to frame. Additionally, the reduced resolution of feature maps relative to the input image significantly limits the benefit of these methods in terms of accuracy. We show that the distance transform, not only proves to be experimentally more consistent across frames, but also stands as the theoretically best method for increasing variance under the constancy constraint, while preserving the resolution of the input image. Our work employs a similar methodol" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.198 + ], + "angle": 0, + "content": "ogy as [26], which theoretically prove the strong invariant properties induced by the census transform used in robust optical flow estimation [50] and validate these properties through experiments. [8] also uses the distance transform, but as an additional input to infer dense depth from sparse depth. However, using the distance transform of edge maps as a training signal is a novel approach." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.209, + 0.62, + 0.225 + ], + "angle": 0, + "content": "3. Preamble" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.234, + 0.907, + 0.339 + ], + "angle": 0, + "content": "Our goal in this section is to propose a mathematical formalism that closely aligns with experimental scenarios to address the following question: How can we introduce sufficient variance in low-texture areas while adhering as much as possible to the constancy assumption? The purpose is to provide a better basis than the usual framework [91] for learning depth and ego-motion." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.34, + 0.907, + 0.401 + ], + "angle": 0, + "content": "We start by introducing some key theoretical properties of the distance transform that we illustrate with toy experiments. The following definitions and theorems provide the theoretical foundation of our analysis." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.408, + 0.83, + 0.423 + ], + "angle": 0, + "content": "3.1. Maximal variance under constraints" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.431, + 0.906, + 0.49 + ], + "angle": 0, + "content": "Notations Let \\(\\Omega\\) be a compact convex set of \\(\\mathbb{R}^n\\) with smooth boundary \\(\\partial \\Omega\\). Let \\(\\mathcal{A}\\) be the set of affine transformations. And let \\(\\mathcal{I}_{\\mathcal{A}}\\) the set of positive functions invariant by affine transformation (see Definition 2)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.491, + 0.906, + 0.537 + ], + "angle": 0, + "content": "Here, \\(\\Omega\\) represents an object within a scene, while \\(\\mathcal{A}\\) models a reasonable subset of reprojected 3d transformations likely to affect the objects from frame to frame." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.906, + 0.574 + ], + "angle": 0, + "content": "Definition 1 The distance transform \\(\\delta_{\\Omega}:\\Omega \\to \\mathbb{R}^{+}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.582, + 0.907, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\delta_ {\\Omega} (x) = d (x, \\partial \\Omega) = \\inf _ {y \\in \\partial \\Omega} \\| x - y \\| \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.612, + 0.906, + 0.686 + ], + "angle": 0, + "content": "The set of points having more than one closest point to the boundary is called the medial axis. By Rademacher's Theorem [53] the distance transform is differentiable almost everywhere (away of the medial axis) and satisfies the Eikonal equation:" + }, + { + "type": "equation", + "bbox": [ + 0.674, + 0.687, + 0.905, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\| \\nabla d \\| = 1 \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.709, + 0.904, + 0.74 + ], + "angle": 0, + "content": "Definition 2 Let \\( f: \\Omega \\to \\mathbb{R}^+ \\) a smooth function on \\( \\Omega \\). Then \\( f \\) is said to be invariant by affine transformation if:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.747, + 0.905, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\forall A \\in \\mathcal {A}, \\forall x \\in \\Omega , f (A (x)) = f (x) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.77, + 0.906, + 0.831 + ], + "angle": 0, + "content": "Definition 2 restates the constancy assumption mathematically: if an affine transformation \\( A \\) is applied to a convex shape \\( \\Omega \\), then any changes introduce by \\( f \\) on \\( \\Omega \\) should manifest identically in \\( A(\\Omega) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.837, + 0.906, + 0.868 + ], + "angle": 0, + "content": "Theorem 1 \\(\\delta_{\\Omega}\\) is the unique solution, up to an isomorphism, to the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.873, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\max _ {f \\in \\mathcal {I} _ {\\mathcal {A}}} \\int_ {\\Omega} (f (x) - \\bar {f}) ^ {2} d x \\quad s. t. \\| \\nabla f \\| = 1 \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21870" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.241 + ], + "angle": 0, + "content": "\\(\\overline{f}\\) being the average value of \\(f\\) over \\(\\Omega\\). A proof of Theorem 1 is given in Appendix A.1. This property is particularly significant because it demonstrates that, under the constraint of the Eikonal equation, the optimal method for increasing variance within a compact convex shape while adhering to the constancy assumption is to use the distance transform. For this reason, in our practical experiments, we propose to introduce variance within image instances primarily via the distance transform; further details can be found in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.409 + ], + "angle": 0, + "content": "The Eikonal equation constraint acts as a standardization condition, promoting smooth variation. Without this constraint, partitioning the shape in uniform black and white regions would also maximize the variance. Our goal is instead to ensure a more even spread of variance throughout the shape, making the variance-augmented image reprojection more informative for better depth and ego-motion learning. We also explore carefully relaxing this constraint to investigate well-chosen functions of the distance transform that could allow for greater variance within instances while still maintaining an acceptable level of smoothness:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.418, + 0.482, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} = \\left\\{f: x \\mapsto g \\circ \\delta_ {\\Omega} (x) \\mid g \\in \\mathcal {C} ^ {\\infty} \\left(\\mathbb {R} ^ {+}, \\mathbb {R} ^ {+}\\right), \\| g \\| _ {\\infty} \\leq 1 \\right\\} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.483, + 0.525 + ], + "angle": 0, + "content": "A thorough comparison of depth estimation results across different functions is provided in Appendix C.2. The bound on the norm ensures that variance is not artificially amplified by larger values, maintaining consistency and fairness in comparisons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.483, + 0.556 + ], + "angle": 0, + "content": "We show that the maximal variance property can be retrieved with a simple toy experiment in A.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.565, + 0.31, + 0.58 + ], + "angle": 0, + "content": "3.2. Convergence properties" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.484, + 0.709 + ], + "angle": 0, + "content": "Notations Let \\((I, \\tilde{I})\\) be two successive images of a camera in movement and \\((x, y)\\) a training instance where \\(x\\) represents a pixel of \\(I\\) and \\(y\\) be the corresponding pixel in \\(\\tilde{I}\\). Let us assume also that all pixels lie inside a convex shape \\(\\Omega\\) defined by a smooth and closed contour \\(\\partial \\Omega\\). Our goal is to learn the model \\(\\Phi(x; \\theta)\\) to predict the new pixel location \\(\\hat{y}\\). Let us define the cost function \\(l\\) using functions from \\(\\mathcal{F}\\) (Equation 5):" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.718, + 0.482, + 0.742 + ], + "angle": 0, + "content": "\\[\nl: (\\hat {y}, y) \\mapsto (f (\\hat {y}) - f (y)) ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.74, + 0.297, + 0.754 + ], + "angle": 0, + "content": "with \\(f\\in \\mathcal{F}\\)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.766, + 0.45, + 0.782 + ], + "angle": 0, + "content": "Definition 3 A function \\( f: \\Omega \\to \\mathbb{R}^+ \\) is \\( \\alpha \\)-Lipschitz if:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.792, + 0.482, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\forall u, v \\quad | f (u) - f (v) | \\leq \\alpha \\| u - v \\| \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.818, + 0.439, + 0.835 + ], + "angle": 0, + "content": "Definition 4 A function \\( f: \\Omega \\to \\mathbb{R}^+ \\) is \\( \\beta \\)-smooth if:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.844, + 0.482, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\forall u, v \\quad \\| \\nabla f (u) - \\nabla f (v) \\| \\leq \\beta \\| u - v \\| \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The fact that \\( f \\) is a function of the distance transform makes it possible to prove:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "Theorem 2 The loss function \\( l \\) is, with respect to the first argument \\( \\hat{y} \\):" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.125, + 0.608, + 0.139 + ], + "angle": 0, + "content": "- \\(\\alpha\\)-Lipschitz" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.14, + 0.597, + 0.153 + ], + "angle": 0, + "content": "- \\(\\beta\\)-smooth" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.154, + 0.895, + 0.169 + ], + "angle": 0, + "content": "- strongly convex if a regularization term \\(\\eta \\| y\\|^2\\) is added." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.125, + 0.895, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.186, + 0.907, + 0.351 + ], + "angle": 0, + "content": "The constants \\(\\alpha, \\beta, \\eta\\) depend on the curvature at the projection onto the boundary \\(\\partial \\Omega\\). A proof of Theorem 2 is given in Appendix A.2. Those properties brought by the distance transform induce great convergence properties as described in [1, 27, 36]. Amongst others: a bounded generalization error and a uniform stability of the stochastic gradient descent. Note that the SSIM + colour function can also be Lipschitz and/or uniform under very strict conditions. However their constants, unlike with the distance transform, depend on the size of the shape \\(\\Omega\\) and can thus exceeds all bounds as mentioned in [57]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.352, + 0.906, + 0.382 + ], + "angle": 0, + "content": "We show that the convergence property can be retrieved with a simple toy experiment in A.4." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.402, + 0.605, + 0.417 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.43, + 0.906, + 0.52 + ], + "angle": 0, + "content": "Our framework for self-supervised MDE (Figure 2) consists of three networks: Depth \\(D_{\\theta}\\), Ego-motion \\(T_{\\alpha}\\) and edge \\(E_{\\delta}\\). For the sake of readability, we will omit the parameter symbols \\(\\theta, \\alpha, \\delta\\) in the remainder of the section. Hyperparameters in the loss functions will also be omitted here and detailed in Appendix F." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.536, + 0.722, + 0.55 + ], + "angle": 0, + "content": "4.1. Pre-semantic contours" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.906, + 0.681 + ], + "angle": 0, + "content": "Our edge network renders values \\( E(p) \\) in the range [0, 1], where a higher value indicates a greater likelihood that the pixel \\( p \\) corresponds to an edge. To generate pseudo-labels for supervising the edge network, we utilize depth and surface normal estimation, which is derived from depth following the methods described in [83, 85]. We discuss complementary insights of depth and surface normal for edge detection in Appendix D.1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.682, + 0.906, + 0.742 + ], + "angle": 0, + "content": "Depth & Normal Pseudo-Labels: Depth and surface normals are smooth and thus prone to exhibiting zero-crossings of their Laplacian \\(\\Delta D\\) (resp. \\(\\Delta N\\)) around semantic boundaries, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.757, + 0.905, + 0.813 + ], + "angle": 0, + "content": "\\[\nZ _ {D} (p) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\Delta D (p) \\Delta D (p + (1, 0)) < 0 \\\\ 1 & \\text {i f} \\Delta D (p) \\Delta D (p + (0, 1)) < 0 \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Same applies to normal Laplacian \\(\\Delta N\\). The zero-crossing mask is then dilated (\\(3 \\times 3\\) structuring element), and multiplied by the magnitude of the gradient to create the pseudolabels. For this, the depth gradient is normalized to avoid" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21871" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.5, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.905, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.276, + 0.908, + 0.307 + ], + "angle": 0, + "content": "Figure 2. Diagram showing the two main steps of our pipeline: estimation of the different modalities followed by the re-projection on the variance augmented image using depth and ego-motion. T indicates the time frame. R, t respectively rotation and translation predictions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.331, + 0.306, + 0.347 + ], + "angle": 0, + "content": "bias towards large depth values:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.356, + 0.482, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\nabla D _ {x} (p) = \\frac {D (p + (0 , 1)) - D (p)}{\\max (D (p + (0 , 1)) , D (p))}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.394, + 0.381, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\| \\nabla D (p) \\| = | \\nabla D _ {x} (p) | + | \\nabla D _ {y} (p) |.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.483, + 0.452 + ], + "angle": 0, + "content": "The usual normal gradient, which is often too noisy, is replaced with a dot-product measure:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.463, + 0.483, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\left| \\nabla N _ {x} (p) \\right| = 1 - \\left| \\langle \\vec {N} (p + (0, 1)), \\vec {N} (p) \\rangle \\right|, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.486, + 0.381, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\nabla N (p) \\right\\| = \\left| \\nabla N _ {x} (p) \\right| + \\left| \\nabla N _ {y} (p) \\right|.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.483, + 0.544 + ], + "angle": 0, + "content": "The weighted depth and normal pseudo-labels employed to supervise the edge network are then defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.554, + 0.483, + 0.593 + ], + "angle": 0, + "content": "\\[\nw _ {d} (p) = \\frac {Z _ {D} (p) \\| \\nabla D (p) \\|}{\\sum_ {p} Z _ {D} (p) \\| \\nabla D (p) \\|} \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.602, + 0.483, + 0.642 + ], + "angle": 0, + "content": "\\[\nw _ {n} (p) = \\frac {Z _ {N} (p) \\| \\nabla N (p) \\|}{\\sum_ {p} Z _ {N} (p) \\| \\nabla N (p) \\|} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.647, + 0.485, + 0.722 + ], + "angle": 0, + "content": "A contrastive loss is also applied to encourage edge values to move toward either zero or one. This is achieved using a binary cross-entropy loss, where pseudo ground truth \\( Y(p) \\in \\{0,1\\} \\) are created by thresholding edge estimations \\( E(p) \\) at 0.5:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.733, + 0.504, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = - \\sum_ {p} \\left(Y (p) \\log (E (p)) + (1 - Y (p)) \\log (1 - E (p))\\right) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Finally, the loss function employed to supervise the edge network is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.82, + 0.465, + 0.852 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {e d g e}} = \\sum_ {p} \\left(w _ {d} (p) + w _ {n} (p)\\right) \\left(1 - E (p)\\right) + \\mathcal {L} _ {c} + \\mathcal {L} _ {e}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.856, + 0.482, + 0.901 + ], + "angle": 0, + "content": "with \\(\\mathcal{L}_e = \\sum_p E(p)^2\\) to avoid trivial solutions. (15)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.331, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Discontinuity Preserving Smoothing: Incorporating an edge-aware smoothness loss is a widely used practice in training depth networks [73, 83]:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.4, + 0.905, + 0.447 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\sum_ {p} \\left(| \\nabla D _ {x} (p) | e ^ {- | \\nabla I _ {x} (p) |} + | \\nabla D _ {y} (p) | e ^ {- | \\nabla I _ {y} (p) |}\\right) \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.448, + 0.907, + 0.537 + ], + "angle": 0, + "content": "However, this approach has two drawbacks: first, boundaries are estimated using the image gradient, which is not optimal; second, there is no mechanism to effectively manage discontinuities. To address these issues, we propose a new smoothness loss exploiting our pre-semantic contours defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.56, + 0.732, + 0.594 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s _ {1}} = \\sum_ {p} \\| \\nabla D (p) \\| e ^ {- \\alpha E (p)},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.596, + 0.905, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s _ {2}} = \\sum_ {p} \\log \\left(1 + e ^ {- \\| \\nabla D (p) \\|}\\right) \\left(1 - e ^ {- \\alpha E (p)}\\right), \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.634, + 0.656, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\mathcal {L} _ {s _ {1}} + \\mathcal {L} _ {s _ {2}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.78 + ], + "angle": 0, + "content": "The component \\(\\mathcal{L}_{s_1}\\) replace the former equation 16 to encourage depth smoothness inside semantic instances, while \\(\\mathcal{L}_{s_2}\\) prevents diffusion and promotes significant depth changes around boundaries. In this manner, both the edge and depth networks benefit from each other: the depth estimation supervises the edge network, while the edge network enhances depth estimation through the smoothness loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Post-processing: The map produced by the edge network consists of continuous values ranging from zero to one. To obtain thin pre-semantic contours \\(\\hat{\\mathcal{C}}(I)\\), we apply a post-processing pipeline that includes hysteresis thresholding, non-maximum suppression, morphological closing, contour filtering, and contour closing. More details can be found in Appendix D.3 and Appendix D.2 discusses the advantages of our approach compared to the competitive work [84]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21872" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.485, + 0.108 + ], + "angle": 0, + "content": "4.2. A variance augmented image for depth learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.162 + ], + "angle": 0, + "content": "With \\(\\widehat{\\mathcal{C}}(I)\\) the thin pre-semantic contour of \\(I\\) obtained after edge network estimation and subsequent post-processing, we then compute the distance transform map \\(\\delta_{\\mathrm{I}}\\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.171, + 0.484, + 0.197 + ], + "angle": 0, + "content": "\\[\n\\delta_ {\\mathrm {I}} (p) = \\min _ {x \\in \\hat {\\mathcal {C}} (I)} (\\| x - p \\|) \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.206, + 0.485, + 0.372 + ], + "angle": 0, + "content": "See Appendix C.1 for the details of the algorithm. The resulting distance transform map is concatenated with the input image to create a 'variance-augmented' image. The objective is then to learn depth and ego-motion in order to identify corresponding pixels across successive frames, and reconstruct the upcoming structure-augmented image in the most effective manner, as illustrated on Figure 2. The reconstruction of the original input image \\(\\widehat{\\mathbf{l}}\\) is evaluated using \\(\\Psi\\), the standard combination of colour difference and SSIM, while the reconstruction of the distance transform map is assessed with a \\(\\mathrm{L}_1\\) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.382, + 0.482, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p h o t o}} = \\Psi (\\widehat {I}, I), \\text {a n d} \\tag {19}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.411, + 0.341, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d i s t}} = \\left\\| \\widehat {\\delta_ {\\mathrm {I}}} - \\delta_ {\\mathrm {I}} \\right\\| _ {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.471 + ], + "angle": 0, + "content": "The smoothness loss is added to get the final supervision signal of the depth network:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.48, + 0.483, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {d e p t h}} = \\mathcal {L} _ {\\text {d i s t}} + \\mathcal {L} _ {\\text {p h o t o}} + \\mathcal {L} _ {s} \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.555 + ], + "angle": 0, + "content": "An improved distance transform: This approach can be readily extended to the family of one-dimensional functions \\(\\mathcal{F} = \\{p\\mapsto g\\circ \\delta_{\\mathrm{I}}(p)\\}\\) (as defined in Equation 5):" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.562, + 0.483, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t}} = \\left\\| g \\circ \\widehat {\\delta_ {\\mathrm {I}}} - g \\circ \\delta_ {\\mathrm {I}} \\right\\| _ {1}. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.484, + 0.683 + ], + "angle": 0, + "content": "We provide an analysis of the depth estimation results for various functions in Appendix C.2. To take this a step further, we propose extending the distance transform into the \\(n\\)-dimensional space by mapping the distance transform values to an \\(n\\)-dimensional random walk process denoted as \\(\\mathrm{RW}_n\\):" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.691, + 0.483, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t}} = \\sum_ {p} \\left\\| \\mathrm {R W} _ {n} \\circ \\widehat {\\delta_ {\\mathrm {I}}} (p) - \\mathrm {R W} _ {n} \\circ \\delta_ {\\mathrm {I}} (p) \\right\\| _ {1} \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Using this approach, we can maintain the validity of the constancy assumption while increasing the variance even more. Additionally, the dimensional parameter \\( n \\) determines the significance of the distance transform map reconstruction within the total loss function. It introduces a tradeoff between \\( \\mathcal{L}_{\\mathrm{dist}} \\) and \\( \\mathcal{L}_{\\mathrm{photo}} \\): the higher the dimension, the greater the emphasis on the artificially introduced structure compared to the natural structure found in the image; see Appendix C.3 for more details on the random walk mapping. Our experiments in Appendix C.2 show that \\( n = 3 \\) provides the best results." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.646, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.617, + 0.13 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.139, + 0.818, + 0.154 + ], + "angle": 0, + "content": "We conduct experiments on multiple datasets:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.154, + 0.905, + 0.244 + ], + "angle": 0, + "content": "KITTI [20] includes diverse urban, rural, and highway scenes. We follow the standard data split defined by [13] and apply the pre-processing steps outlined in [91] to filter static frames, resulting in a training set of 39810 images and 697 images for testing. Results are reported using the evaluation protocol from [91]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.245, + 0.906, + 0.319 + ], + "angle": 0, + "content": "Cityscapes [9] offers urban scenes with dynamic objects, challenging depth estimation. The training set consists of 22973 images, with evaluation on 1525 test images following [5, 41]. Edge detection is assessed using the contour evaluation from [84] on 500 validation frames." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.32, + 0.906, + 0.411 + ], + "angle": 0, + "content": "Waymo [69] is a large and diverse autonomous driving dataset that captures dynamic urban scenes under various environmental conditions, including nighttime and different weather scenarios. We sample 100 000 image pairs from 1 000 front camera video sequences for training and evaluate on 1500 pairs from 150 sequences, following [41]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.411, + 0.905, + 0.485 + ], + "angle": 0, + "content": "NYUv2 [67] consists of 464 indoor video sequences captured with a Kinect sensor. We use the official splits (302 training, 33 validation) and evaluate on 654 densely annotated test images, resizing inputs to \\(320 \\times 256\\) pixels as in [44]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.486, + 0.906, + 0.547 + ], + "angle": 0, + "content": "ScanNet [10] is a large-scale RGB-D dataset with 2.5 million images across \\(1500+\\) scenes. We use it for zero-shot evaluation, following [16, 44], with test images resized to \\(320 \\times 256\\) pixels." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.557, + 0.731, + 0.573 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.907, + 0.836 + ], + "angle": 0, + "content": "We emphasize that our method is compatible with any DNN architecture and input image size. For fair comparison, our experiments are done with a \\(256 \\times 832\\) resolution and a UNet [62] structure, following [91], unless stated otherwise. The depth and edge networks share a ResNet [29] encoder. Our baseline, CoopNet [28], handles moving objects—especially useful in Cityscapes—via self-supervised flow. The pose network, based on a ResNet encoder, outputs a 6-DoF vector. All encoders use ResNet-50 backbones initialized with ImageNet [63] weights. Training runs for 30 epochs with batch size 4, initial learning rate of \\(10^{-4}\\) (decreasing to \\(10^{-5}\\) after 20 epochs), and standard data augmentation from [22]. Loss weights, tuned via grid search, are detailed in Appendix F. Our PyTorch [58] implementation trains on a single NVIDIA RTX A5000 GPU using Adam [38] with \\(\\beta_{1} = 0.99\\) and \\(\\beta_{2} = 0.999\\) and takes 12 hours." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.847, + 0.608, + 0.862 + ], + "angle": 0, + "content": "5.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Depth Quantitative results on KITTI and Cityscapes are presented in Table 1, where we compare against the most" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21873" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.37 + ], + "angle": 0, + "content": "
MethodLower is better ↓Higher is better ↑
Abs RelSq RelRMSERMSE logδ1δ2δ3
Monodepth2[22]0.1100.8314.6420.1870.8830.9620.982
Guizilini et al.†[24] (R50)0.1130.8314.6630.1890.8780.9710.983
Shu et al.*[66](R50)0.1080.7924.6330.1840.8830.9610.983
Ours (R50)0.1040.7254.4530.1800.8850.9620.983
MonoViT[89]0.0990.7084.3720.1750.9000.9670.984
Ours + MonoViT0.0920.6744.3000.1650.9270.9670.984
HR-Depth[49]0.1090.7924.6320.1850.8840.9620.983
RA-Depth[30]0.0960.6134.2160.1710.9030.9680.985
DIFFNet[90]0.1020.7644.4830.1800.8960.9650.983
Guizilini et al.†[24]0.1000.7614.2700.1750.9020.9650.982
Ours+DinoV20.0820.6044.1080.1620.9280.9680.985
Lego (paper)0.1541.2726.0120.230N/AN/AN/A
Lego★ (R50)0.1150.8554.7890.1950.8760.9570.979
Lego★ (R50) + Ldist0.1100.8054.6060.1880.8820.9600.982
Ours (R50)0.1040.7254.4530.1800.8850.9620.983
Struct2Depth†[4]0.1451.7377.280.2050.8130.9420.978
LearnK†[41]0.1271.3306.960.1950.8300.9470.981
Li et al.†[41]0.1191.2906.980.1900.8460.9510.980
CoopNet[28]0.1211.4437.010.1900.8460.9510.980
Ours (R50)0.1151.2216.790.1860.8500.9550.981
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.522 + ], + "angle": 0, + "content": "Table 1. Results of depth estimations on KITTI 2015 (first four blocks) and Cityscapes (last block). For each metric the best result is displayed in bold and the second-best is underlined. Light blue metrics indicates the most challenging metrics. \\(\\dagger\\) denotes the use of an off-the-shelf semantic algorithm and * indicates results reproduced by us at the input resolution of \\(256 \\times 832\\). Block#1 and #2 compares methods using similar encoder backbones: Resnet 50 (R50) in block#1 and ViT in block#2. Block#3 compares SOTA depth models without restrictions on architecture or additional inputs, Block#4 compares to Lego." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.538, + 0.481, + 0.631 + ], + "angle": 0, + "content": "
MethodDataset (Train / Test)Lower is better ↓
Abs RelSq RelRMSERMSE log
LearnK[41]Waymo / Waymo0.1571.5317.0900.205
OursWaymo / Waymo0.1251.2785.9900.189
GLNet[7]Cityscapes / KITTI0.1291.0445.3610.212
CoopNet[28]Cityscapes / KITTI0.1251.1575.2510.209
OursCityscapes / KITTI0.1160.8924.8920.195
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.671 + ], + "angle": 0, + "content": "Table 2. Results on Outdoor datasets. Block#1: in-domain evaluation. Block#2: out-of-domain evaluation." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.688, + 0.481, + 0.77 + ], + "angle": 0, + "content": "
MethodDataset (Train / Test)Abs RelRMSEδ1δ2δ3
MonoIndoor++[44]NYUv2 / NYUv20.1320.5170.8340.9610.990
IndoorDepth[16]NYUv2 / NYUv20.1260.4940.8450.9650.991
OursNYUv2 / NYUv20.1150.4580.8590.9700.992
IndoorDepth[16]NYUv2 / ScanNet0.1530.3730.7860.9500.988
MonoIndoor++[44]NYUv2 / ScanNet0.1380.3470.8100.9670.993
OursNYUv2 / ScanNet0.1270.3120.8430.9700.993
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.782, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Table 3. Results on Indoor datasets. Block#1: in-domain evaluation. Block#2: out-of-domain evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "competitive methods. Blocks #1 and #2 demonstrate that, when using the same encoder backbone, our method outperforms others by a significant margin across nearly all metrics. When no architectural constraints are imposed, we" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.089, + 0.905, + 0.166 + ], + "angle": 0, + "content": "
RW3LdistLsEθAbs Rel (↓)RMSE log (↓)δ1(↑)
0.1130.190.878
✓0.1120.1890.880
✓✓0.1100.1870.881
✓✓✓0.1060.1830.884
✓✓✓✓0.1040.1800.885
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.175, + 0.907, + 0.245 + ], + "angle": 0, + "content": "Table 4. Ablation study on KITTI 2015 of losses mentioned in Section 4.2. Note that \\( E_{\\theta} \\) is part of the ablation as it does improve the depth metrics and shares the same encoder as the depth network. \\( \\mathrm{RW}_3 \\) is for 3d random walk encoding of distance transform as defined in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.272, + 0.907, + 0.408 + ], + "angle": 0, + "content": "employ a pre-trained DinoV2 [56] encoder (further details in Appendix F.2), yielding the results in block #3. For fairness, we compare against \\(\\mathsf{LEGO}^{\\star}\\) in block #4, an enhanced version of LEGO that incorporates a ResNet-50 backbone and up-to-date training techniques. Adding our novel distance transform loss \\(\\mathcal{L}_{\\mathrm{dist}}\\) into \\(\\mathsf{LEGO}^{\\star}\\) greatly improves results, showing once again the great importance of this loss component, but still underperforms ours, likely due to the superior quality of our contours." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.408, + 0.909, + 0.621 + ], + "angle": 0, + "content": "We also provide strong out-of-domain results in Table 2 and Table 3, demonstrating competitive performance on both indoor and outdoor datasets. Finally, an ablation study in Table 4 evaluates the impact of the different loss components introduced in our approach and in Appendix C.2 for different functions from the family \\(\\mathcal{F}\\) as defined in Section 3.1. It can be seen that the reprojection on the distance transform via \\(\\mathcal{L}_{\\mathrm{dist}}\\) brings a decisive improvement. We provide a qualitative comparison of results in Figure 3. We can see that our method gives better results: it provides thinner object, is sharper around moving objects, does not suffer from smoothness issues for large objects and is not noisy in textured areas. More qualitative and quantitative results can be found in Appendix B.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.727 + ], + "angle": 0, + "content": "Flow & Odometry As discussed in Section 5.2 we use an optical flow to handle moving objects following [28]. Its results can be found in Appendix B.2. Results for odometry can be found in Appendix B.3. Our method significantly outperforms the baseline [28] on both optical flow and odometry metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.735, + 0.702, + 0.75 + ], + "angle": 0, + "content": "5.4. Contour evaluation" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.757, + 0.909, + 0.895 + ], + "angle": 0, + "content": "We evaluate contour detection in Table 5 using the same metrics and data as [84]: Optimal Dataset Scale, Optimal Image Scale, and Average Precision, computed on 500 Cityscapes validation images. Our method significantly outperforms LEGO, leveraging (1) depth & normal pseudolabels with Laplacian zero-crossing insights for better contour alignment, (2) the discontinuity-preserving smoothing loss \\(\\mathcal{L}_s\\), and (3) the contrastive loss \\(\\mathcal{L}_c\\). Qualitative examples are in Figure 1, with more in Appendix D." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21874" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.1, + 0.907, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.424, + 0.908, + 0.467 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons for different methods. First row: colour images. Second row: Monodepth2 [22]. Second row: CoopNet [28]. Third row: FeatDepth [66]. Last row: Ours. We compare qualitative results in four complex situations. First column: a thin object. Second column: a moving object. Third column: a large object with smoothness issue. Last column: Textured areas." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.477, + 0.48, + 0.55 + ], + "angle": 0, + "content": "
MethodsODSOISAP
Lego (paper)0.7100.7310.729
Lego*0.7090.7360.734
Depth & Normal pseudo-labels0.7550.7750.760
Depth & Normal pseudo-labels + Ls0.7600.7820.766
Ours = Depth & Normal pseudo-labels + Ls + Lc0.7620.7890.770
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.556, + 0.483, + 0.598 + ], + "angle": 0, + "content": "Table 5. Comparison for contour estimation. Block#1: Lego results. Block#2: Ablation study of losses in contour estimation training. Experiments done on Cityscapes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.627, + 0.363, + 0.643 + ], + "angle": 0, + "content": "5.5. Constancy assumption validity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.786 + ], + "angle": 0, + "content": "We evaluate the constancy assumption validity of our method, compared to [66] that uses deep features from the ResNet-50 encoder in the re-projection. To do so, we use the KITTI MOTS [72] dataset to track a specific point of an instance along different trajectories and compute the normalized temporal variance. We show that our distance transform satisfies much better the constancy assumption than any layers of the ResNet-50 (see Figure 4). More details on the evaluation protocol are given in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.21, + 0.815 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As a conclusion, this paper presents a novel self-supervised approach to monocular depth estimation, addressing challenges in low-texture regions by applying a distance transform over pre-semantic contours. This technique enhances depth prediction accuracy by increasing discrimini" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.495, + 0.868, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.669, + 0.907, + 0.711 + ], + "angle": 0, + "content": "Figure 4. Comparison of normalized temporal variance of different features for the centre pixel of an instance tracked along different sequence of KITTI MOTS [72]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.811 + ], + "angle": 0, + "content": "native power where traditional photometric losses struggle. Both theoretical analysis and extensive experiments validate the effectiveness of the proposed method, which outperforms conventional self-supervised approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.812, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In future work, we aim to introduce consistent variance in low-texture regions without pre-semantic contours. Additionally, we plan to explore whether our framework can leverage variance-augmented images to learn from diverse datasets, ultimately developing a foundation model in a fully self-supervised manner." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21875" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Ali Akbari, Muhammad Awais, Manijeh Bashar, and Josef Kittler. How does loss function affect generalization performance of deep learning? application to human age estimation. In International Conference on Machine Learning, pages 141-151. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Abhishek Badki, Alejandro Troccoli, Kihwan Kim, Jan Kautz, Pradeep Sen, and Orazio Gallo. Bi3d: Stereo depth estimation via binary classifications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1600-1608, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Jiawang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised scale-consistent depth and ego-motion learning from monocular video. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Depth prediction without the sensors: Leveraging structure for unsupervised learning from monocular videos. In Proceedings of the AAAI conference on artificial intelligence, pages 8001-8008, 2019. 2, 7, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[5] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Unsupervised monocular depth and ego-motion learning with structure and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.563 + ], + "angle": 0, + "content": "[6] Tian Chen, Shijie An, Yuan Zhang, Chongyang Ma, Huayan Wang, Xiaoyan Guo, and Wen Zheng. Improving monocular depth estimation by leveraging structural awareness and complementary datasets. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 90-108. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[7] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 1, 2, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[8] Zhao Chen, Vijay Badrinarayanan, Gilad Drozdov, and Andrew Rabinovich. Estimating depth from rgb and sparse sensing. In Proceedings of the European Conference on Computer Vision (ECCV), pages 167-182, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In Proceedings of the European conference on computer vision (ECCV), pages 562-578, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[12] Tom van Dijk and Guido de Croon. How do neural networks see depth in single images? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2183-2191, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[14] F El Jamiy and R Marsh. Survey on depth perception in head mounted displays: distance estimation in virtual reality, augmented reality, and mixed reality. iet image processing 13 (5), 707-712 (2019), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.906, + 0.331 + ], + "angle": 0, + "content": "[15] Sara Elkerdawy, Hong Zhang, and Nilanjan Ray. Lightweight monocular depth estimation model by joint end-to-end filter pruning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 4290-4294. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[16] Chao Fan, Zhenyu Yin, Yue Li, and Feiqing Zhang. Deeper into self-supervised monocular indoor depth estimation. arXiv preprint arXiv:2312.01283, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.906, + 0.445 + ], + "angle": 0, + "content": "[17] Philipp Fischer, Alexey Dosovitskiy, Eddy Ilg, Philip Häusser, Caner Hazirbaş, Vladimir Golkov, Patrick Van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. arXiv preprint arXiv:1504.06852, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[18] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.52, + 0.906, + 0.602 + ], + "angle": 0, + "content": "[19] Ravi Garg, Vijay Kumar Bg, Gustavo Carneiro, and Ian Reid. Unsupervised cnn for single view depth estimation: Geometry to the rescue. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 740-756. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.906, + 0.657 + ], + "angle": 0, + "content": "[20] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.906, + 0.729 + ], + "angle": 0, + "content": "[21] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[22] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[23] Ariel Gordon, Hanhan Li, Rico Jonschkowski, and Anelia Angelova. Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8977-8986, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Vitor Guizilini, Rui Hou, Jie Li, Rares Ambrus, and Adrien Gaidon. Semantically-guided representation learn" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21876" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "ing for self-supervised monocular depth. arXiv preprint arXiv:2002.12319, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.188 + ], + "angle": 0, + "content": "[25] Hyowon Ha, Sunghoon Im, Jaesik Park, Hae-Gon Jeon, and In So Kweon. High-quality depth from uncalibrated small motion clip. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5413-5421, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[26] David Hafner, Oliver Demetz, and Joachim Weickert. Why is the census transform good for robust optic flow computation? In Scale Space and Variational Methods in Computer Vision: 4th International Conference, SSVM 2013, Schloss Seggau, Leibnitz, Austria, June 2-6, 2013. Proceedings 4, pages 210-221. Springer, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[27] Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International conference on machine learning, pages 1225-1234. PMLR, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[28] Marwane Hariat, Antoine Manzanera, and David Filliat. Rebalancing gradient to improve self-supervised co-training of depth, odometry and optical flow predictions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1267-1276, 2023. 2, 6, 7, 8, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.483, + 0.511 + ], + "angle": 0, + "content": "[30] Mu He, Le Hui, Yikai Bian, Jian Ren, Jin Xie, and Jian Yang. Ra-depth: Resolution adaptive self-supervised monocular depth estimation. In European Conference on Computer Vision, pages 565-581. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[31] Stuart B Heinrich and Wesley E Snyder. Improved edge awareness in discontinuity preserving smoothing. arXiv preprint arXiv:1103.5808, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.608 + ], + "angle": 0, + "content": "[32] Junjie Hu, Yan Zhang, and Takayuki Okatani. Visualization of convolutional neural networks for monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3869-3878, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.483, + 0.678 + ], + "angle": 0, + "content": "[33] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2462-2470, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.719 + ], + "angle": 0, + "content": "[34] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[35] Wei Jiang, Weiwei Sun, Andrea Tagliasacchi, Eduard Trulls, and Kwang Moo Yi. Linearized multi-sampling for differentiable image transformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2988-2997, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[36] Kenji Kawaguchi and Qingyun Sun. A recipe for global convergence guarantee in deep neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 8074-8082, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[37] Ue-Hwan Kim, Gyeong-Min Lee, and Jong-Hwan Kim. Revisiting self-supervised monocular depth estimation. In International Conference on Robot Intelligence Technology and Applications, pages 336-350. Springer, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[38] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[39] Yevhen Kuznietsov, Marc Proesmans, and Luc Van Gool. Comoda: Continuous monocular depth adaptation using past experiences. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2907-2917, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[40] Seokju Lee, Sunghoon Im, Stephen Lin, and In So Kweon. Learning residual flow as dynamic motion from stereo videos. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1180-1186. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[41] Hanhan Li, Ariel Gordon, Hang Zhao, Vincent Casser, and Anelia Angelova. Unsupervised monocular depth learning in dynamic scenes. In Conference on Robot Learning, pages 1908-1917. PMLR, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[42] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. Rtm3d: Real-time monocular 3d detection from object keypoints for autonomous driving. In European Conference on Computer Vision, pages 644-660. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[43] Pengzhi Li, Yikang Ding, Haohan Wang, Chengshuai Tang, and Zhiheng Li. The devil is in the edges: Monocular depth estimation with edge-aware consistency fusion. arXiv preprint arXiv:2404.00373, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[44] Runze Li, Pan Ji, Yi Xu, and Bir Bhanu. Monoindoor++: Towards better practice of self-supervised monocular depth estimation for indoor environments. IEEE Transactions on Circuits and Systems for Video Technology, 33(2):830-846, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[45] Rui Li, Danna Xue, Shaolin Su, Xiantuo He, Qing Mao, Yu Zhu, Jinqiu Sun, and Yanning Zhang. Learning depth via leveraging semantics: Self-supervised monocular depth estimation with both implicit and explicit semantic guidance. Pattern Recognition, 137:109297, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[46] Yin Li, Manohar Paluri, James M Rehg, and Piotr Dólar. Unsupervised learning of edges. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1619-1627, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[47] Liang Liu, Guangyao Zhai, Wenlong Ye, and Yong Liu. Unsupervised learning of scene flow estimation fusing with local rigidity. In IJCAI, pages 876-882, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[48] Hengjie Lu, Shugong Xu, and Shan Cao. Sgtbn: generating dense depth maps from single-line lidar. IEEE Sensors Journal, 21(17):19091-19100, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[49] Xiaoyang Lyu, Liang Liu, Mengmeng Wang, Xin Kong, Lina Liu, Yong Liu, Xinxin Chen, and Yi Yuan. Hr-depth: High resolution self-supervised monocular depth estimation. In Proceedings of the AAAI conference on artificial intelligence, pages 2294-2301, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[50] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[51] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21877" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[52] Oskar Natan and Jun Miura. End-to-end autonomous driving with semantic depth cloud mapping and multi-agent. IEEE Transactions on Intelligent Vehicles, 8(1):557-571, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[53] Ales Nekvinda and Ludek Zajicek. A simple proof of the rademacher theorem. Časopis pro pěstovánt matematíky, 113 (4):337-341, 1988. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[54] C Wayne Niblack, Phillip B Gibbons, and David W Capson. Generating skeletons and centerlines from the distance transform. CVGIP: Graphical Models and image processing, 54 (5):420-437, 1992. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[55] Jim Nilsson and Tomas Akenine-Möller. Understanding ssm. arXiv preprint arXiv:2006.13846, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.329 + ], + "angle": 0, + "content": "[56] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[57] Daniel Otero, Davide La Torre, O Michailovich, and Edward R Vrscay. Optimization of structural similarity in mathematical imaging. Optimization and Engineering, 22:2367-2401, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[58] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[59] Michael Ramamonjisoa and Vincent Lepetit. Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[60] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12240-12249, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[61] Alex Rasla and Michael Beyeler. The relative importance of depth cues and semantic edges for indoor mobility using simulated prosthetic vision in immersive virtual reality. In Proceedings of the 28th ACM Symposium on Virtual Reality Software and Technology, pages 1-11, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[62] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[63] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[64] Faraz Saeedan and Stefan Roth. Boosting monocular depth with panoptic segmentation maps. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3853-3862, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[65] Abhijeet Shenoi, Mihir Patel, JunYoung Gwak, Patrick Goebel, Amir Sadeghian, Hamid Rezatofighi, Roberto Martin-Martin, and Silvio Savarese. Jrmot: A real-time 3d multi-object tracker and a new large-scale dataset. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10335-10342. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[66] Chang Shu, Kun Yu, Zhixiang Duan, and Kuiyuan Yang. Feature-metric loss for self-supervised learning of depth and egomotion. In European Conference on Computer Vision, pages 572-588. Springer, 2020. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[67] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.904, + 0.343 + ], + "angle": 0, + "content": "[68] Tilo Strutz. The distance transform and its computation. arXiv preprint arXiv:2106.03503, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[69] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[70] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[71] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[72] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. MOTS: Multi-object tracking and segmentation. In CVPR, 2019. 8, 6, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[73] Chaoyang Wang, José Miguel Buenapasada, Rui Zhu, and Simon Lucey. Learning depth from monocular videos using direct methods. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2022–2030, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[74] Peng Wang, Xiaohui Shen, Bryan Russell, Scott Cohen, Brian Price, and Alan L Yuille. Surge: Surface regularized geometry estimation from a single image. Advances in Neural Information Processing Systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[75] Yang Wang, Yi Yang, Zhenheng Yang, Liang Zhao, Peng Wang, and Wei Xu. Occlusion aware unsupervised learning of optical flow. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4884-4893, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[76] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[77] Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1164-1174, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21878" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[78] Diana Wofk, Fangchang Ma, Tien-Ju Yang, Sertac Karaman, and Vivienne Sze. Fastdepth: Fast monocular depth estimation on embedded systems. In 2019 International Conference on Robotics and Automation (ICRA), pages 6101-6108. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[79] Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single image depth prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 611-620, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[80] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[81] Wanpeng Xu, Ling Zou, Lingda Wu, and Zhipeng Fu. Self-supervised monocular depth learning in low-texture areas. Remote Sensing, 13(9):1673, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.386 + ], + "angle": 0, + "content": "[82] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[83] Z Yang, P Wang, W Xu, L Zhao, and R Nevatia. Unsupervised learning of geometry with edge-aware depth-normal consistency. arxiv 2017. arXiv preprint arXiv:1711.03665, 2 (5). 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[84] Zhenheng Yang, Peng Wang, Yang Wang, Wei Xu, and Ram Nevatia. Lego: Learning edge with geometry all at once by watching videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 225-234, 2018. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.567 + ], + "angle": 0, + "content": "[85] Wei Yin, Yifan Liu, Chunhua Shen, and Youliang Yan. Enforcing geometric constraints of virtual normal for depth prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5684-5693, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[86] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1983-1992, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[87] Xun Yuan and Song Chen. Sad-slam: A visual slam based on semantic and depth information. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4930-4935. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[88] Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 340-349, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[89] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. In 2022 international conference on 3D vision (3DV), pages 668-678. IEEE, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[90] Hang Zhou, David Greenwood, and Sarah Taylor. Self-supervised monocular depth estimation with internal feature fusion. In *British Machine Vision Conference (BMVC)*, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[91] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.163, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[92] Shengjie Zhu, Garrick Brazil, and Xiaoming Liu. The edge of depth: Explicit constraints between segmentation and depth. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13116-13125, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.906, + 0.289 + ], + "angle": 0, + "content": "[93] Yuliang Zou, Zelun Luo, and Jia-Bin Huang. Df-net: Unsupervised joint learning of depth and flow using cross-task consistency. In Proceedings of the European conference on computer vision (ECCV), pages 36-53, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.906, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21879" + } + ] +] \ No newline at end of file diff --git a/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/18d0caf1-d799-41ef-95e4-0a0e2105060c_origin.pdf b/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/18d0caf1-d799-41ef-95e4-0a0e2105060c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2a11b4f114b55e7889e5da085eddc3d5da09afda --- /dev/null +++ b/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/18d0caf1-d799-41ef-95e4-0a0e2105060c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb09b2d6dcd953ea447588bddafdc7e5eaac868603817ed26e745fbc5f21ceaf +size 6846070 diff --git a/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/full.md b/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c9ee03fd86ffed3a7320edd966e6ba47b4700f6a --- /dev/null +++ b/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/full.md @@ -0,0 +1,458 @@ +# Improved monocular depth prediction using distance transform over pre-semantic contours with self-supervised neural networks + +Marwane Hariat1, Antoine Manzanera1, David Filliat1 +1U2IS, ENSTA, Institut Polytechnique de Paris, Palaiseau, France +{marwane.hariat, antoine.manzanera, david fillliat}@ensta.fr + +![](images/73ce385542563d79605072ecd3c823dc265a3e81cd94c19e2dcc7cc1476d20a0.jpg) + +![](images/5e292b9f0aa921153a7b14ca2f5fbda63b5ea1f7ceecd46f0c0700c6b11402d2.jpg) + +![](images/af0834f4151e38c5d13f50d1a97ba7c93e8a9f5335c9b7200e343455e140d46c.jpg) +RGB images + +![](images/1a28d1d8abc98af0348b84aed1b901258ce17e23d4c27bf208185a02163f29cd.jpg) + +![](images/abf7d38042583f22379a5deb90829bfd5a4707f589c62d7b8351b7d0cb30cbbd.jpg) + +![](images/edc5ede1827ecb359dc7b8c0cb0ac8cb8db77b418c98a31a3e5c833b3cab10b9.jpg) +Depth +Figure 1. The different modalities estimated in our self-supervised framework and used to reduce matching uncertainty between images. + +![](images/ca538a1b33f183da4755c3f0c6565799ccbeea122b1678de8dd143ee9d1ad2d5.jpg) + +![](images/80ead54ef523e6a6c1e2dc06b3cf50094fad20aea5e264c119599434bc5d1e51.jpg) + +![](images/f14ef0f810705e78dd61b7145a061d339cd304907cf6b3ddbf77c532b8d7adab.jpg) +Edge + +![](images/fe5fc609326d3ae92032ae0e85bb86ca4596462ec2f0e899d5d4b7926b779f3f.jpg) + +![](images/7bcf76575f83f537b82df78db03aac86140c15025238ad7c4fb098720c582884.jpg) + +![](images/265b82a705b3e3b4a395056ff8cccaea227c5efbbf55bb4ee578e4b07e91a34f.jpg) +Distance Transform + +# Abstract + +Monocular depth estimation (MDE) with self-supervised training approaches struggles in low-texture areas, where photometric losses may lead to ambiguous depth predictions. To address this, we propose a novel technique that enhances spatial information by applying a distance transform over pre-semantic contours, augmenting discriminative power in low texture regions. Our approach jointly estimates pre-semantic contours, depth and ego-motion. The pre-semantic contours are leveraged to produce new input images, with variance augmented by the distance transform in uniform areas. This approach results in more effective loss functions, enhancing the training process for depth and ego-motion. We demonstrate theoretically that the distance transform is the optimal variance-augmenting technique in this context. Through extensive experiments on KITTI, Cityscapes, Waymo, NYUv2 and ScanNet our model demonstrates robust performance, surpassing competing self-supervised methods in MDE. + +# 1. Introduction + +The goal of monocular depth estimation (MDE) is to determine the distance of a pixel from the camera focal plane using only a single input image, unlike stereo depth estimation [2, 21, 25], which relies on pairs of images. This + +eliminates the need for expensive sensors that require high-power resources [42, 48], such as active depth camera or LiDAR, while maintaining low latency and ensuring real-time compliance [15, 78]. This makes MDE a core module in applications such as robot navigation [65, 87], autonomous driving [52, 76] or virtual/augmented reality [14, 61]. Deep Neural Networks (DNNs) are currently the state of the art for depth maps prediction thanks to their ability to exploit all contextual cues of the environment [12, 32]. + +Supervised methods for MDE rely on ground-truth annotations. In spite of huge progress [6, 13, 18, 70], these methods are constrained by the cost and quality of the annotations. In this paper, we focus on self-supervised MDE, that uses the structure-from-motion framework [19, 91] and leverages depth and camera ego-motion estimations from DNNs to reconstruct the next frame from the current one. The supervision signal is then provided by the photometric loss, i.e. the difference between the reconstructed frame and the observed one, based on a combination of direct colour comparison and Structural Similarity Index Measure (SSIM) [55]. Although more challenging, the self-supervised paradigm offers greater training flexibility. First, it can be trained using data from unknown cameras [23], as long as data acquisition provides successive frames, such as in YouTube videos [7]. This is particularly relevant given the growing interest in foundation models, which can now leverage large datasets to produce generalizable out- + +puts [82]. In addition, it is compatible with continual learning, allowing the model to continuously expand its knowledge as more data become available [7, 39]. + +However, the problem of MDE using the photometric loss proposed by Zhou et al. [91] is ill-posed due to its reliance on SSIM and colour-based cost functions, which fail to discriminate effectively in low-texture areas. In these regions, multiple candidates can yield similarly low cost function values, making it difficult to determine the correct correspondence. Although interesting work has been done to address various issues such as scale ambiguity [73], gradient locality [22, 35], occlusion [22, 75], moving objects [41, 60], and infinite depth holes [22, 24], the ill-posed nature of this problem has not been thoroughly and rigorously studied. We aim to provide this essential foundation. + +The key challenge lies in reducing ambiguity to build a more robust foundation for learning depth and ego-motion. To this end, we propose introducing sufficient variance in texture-less areas, while adhering as closely as possible to 'the constancy assumption': any change introduced in image $t$ should be reproduced identically in image $t + 1$ , essential as the photometric loss relies on correspondences between successive frames. We propose to do this by assigning to each pixel the distance to its nearest edge in the image, commonly known as the 'distance transform' [54, 68]. We demonstrate mathematically that the distance transform is the most effective way of adding variance while respecting the constancy assumption, and that it also leads to improved convergence properties. Since the distance transform is calculated on binary contours, we propose a self-supervised pre-semantic contour estimation method, jointly learned with the depth estimation model, and add this distance map as an additional channel to the input image to produce a 'variance-augmented' image for reconstruction. In summary, our contributions are as follows. + +(1) A self-supervised approach for pre-semantic contour estimation, which (i) generates the edge information necessary for computing the distance transform, and (ii) simultaneously reinforces the depth estimation network. +(2) A novel variance-augmented image that incorporates the distance transform, enhancing discriminative power in texture-less areas and improving convergence properties +(3) A theoretical foundation demonstrating the effectiveness of the distance transform for adding variance under the constancy assumption +(4) An extensive experimental validation of our new pipeline, achieving competitive performance on different benchmark datasets: KITTI [20], Cityscapes [9], Waymo [69], NYUv2 [67] and ScanNet [10] + +# 2. Related works + +Self-supervised framework. The work of Zhou et al. [91] is pioneer in self-supervised monocular depth estimation. Their method utilizes depth and ego-motion to find the corresponding pixels between two successive frames and relies on the bilinear sampler [34] to reconstruct the upcoming scene from the current frame. The supervision signal is then derived from the photometric loss, which is a weighted sum of the SSIM and colour cost functions. The effectiveness of this method relies on the constancy assumption, which asserts that object colours remain consistent from frame to frame. This assumption is bound to the hypothesis of constant lighting over Lambertian surfaces, and do not hold in the case of occlusions and moving objects. This method has then been refined in several ways. Kim et al. [37] provide a comprehensive study of the main issues and solutions found in the literature. In [75], occlusions are addressed using a forward warping module that counts the number of bilinear sampling operations for each pixel. Godard et al. [22] compute the occlusion mask by determining the per-pixel minimum of the photometric loss over time. Moving objects are identified through the differences between optical flow and depth/ego-motion in [28, 40, 47, 60, 86, 93], while off-the-shelf semantic algorithms are utilized to localize potential moving objects in [4, 24]. In addition, a residual map is introduced to correct the ego-motion in [23, 41]. Infinite depth holes are filtered through a two-stage training process [24] and are removed using an adaptive cost-volume architecture in [77]. Finally, Godard et al. [22] propose a unified solution that effectively addresses all the aforementioned issues. However, the ill-posed nature of the problem due to unstructured objects remains unaddressed. + +Contour estimation. Semantic contours should follow the boundaries derived from instance segmentation, outlining distinct entities within an image. Such contours would naturally help improve the accuracy of depth maps. [74] and [59] leverage ground-truth annotations to learn depth, normals, and contours, incorporating regularization terms to enforce constraints derived from their geometric relationships. In contrast, [43] employs an off-the-shelf edge prediction algorithm to refine initial depth estimations as a post-processing step. [79], [64] and [45] utilize a pretrained instance segmentation network to compute edges, sampling points on both sides of the boundary to enhance depth sharpness around those edges. [92] introduced a morphing strategy to align depth borders computed with a threshold gradient on the depth with semantic borders obtained from a pre-trained semantic algorithm. While these techniques demonstrate significant improvements, they assume the edges as provided and are therefore not fully self-supervised, in addition to not being GPU friendly. Significant advances have been made in supervised edge map estimation [11, 80]; however, the self-supervised approach re + +mains an ambitious and challenging task. [46] introduces a reinforcement training procedure in which an optical flow network is trained to predict the displacement of each pixel between two frames, based on ground truth point correspondences. Subsequently, an edge detector is updated to align its predicted edges of the image with the predicted edges of the estimated flow map. Still, this method is not self-supervised. A work closer to ours [84] proposes a fully self-supervised edge prediction framework that utilizes the positive part of the second derivatives of depth and the first-order gradient of normal to surface to feed an edge network. However, this approach employs suboptimal loss functions, leading to biased contours that favour shorter distances. Furthermore, it does not utilize the edges to enhance depth, nor does it include any post-processing to create refined contours. We propose new cost functions for training the edge network, resulting in significant performance gains. Additionally, we introduce a post-processing step to refine contours and implement an extended edge-aware diffusion smoothness loss [31] that encourages sharp depth discontinuities around estimated boundaries, in the same spirit as [45, 79, 92]. + +Targeting the ill-posed optimization problem. As previously mentioned, our objective is to minimize the photometric loss by establishing correspondences between pixels in the current frame and those in the next frame, achieved through the depth and ego-motion parameters. However, this correspondence procedure is not unique, since many pixels share the same colour and exhibit similar local colour statistics as their neighbourhood. This ambiguity is especially pronounced in areas with low texture. The challenge is to compensate for the lack of variation by introducing time-consistent variance across frames, to satisfy the constancy assumption. This reduces the ill-posed nature of the optimization problem and creates better conditions for learning depth and ego-motion. [81, 88] addressed this issue by applying the reconstruction loss to deep features in addition to the usual colour images. Shu et al. [66] took this approach further and designed custom deep features by training an auto-encoder that encourages large and smooth gradient values in feature maps, yielding deep features with both semantic richness and substantial structure. The constancy assumption is rather optimistic for these methods, as demonstrated in our experiments. Indeed, due to their large receptive fields, deep features generally lack consistency from frame to frame. Additionally, the reduced resolution of feature maps relative to the input image significantly limits the benefit of these methods in terms of accuracy. We show that the distance transform, not only proves to be experimentally more consistent across frames, but also stands as the theoretically best method for increasing variance under the constancy constraint, while preserving the resolution of the input image. Our work employs a similar methodol + +ogy as [26], which theoretically prove the strong invariant properties induced by the census transform used in robust optical flow estimation [50] and validate these properties through experiments. [8] also uses the distance transform, but as an additional input to infer dense depth from sparse depth. However, using the distance transform of edge maps as a training signal is a novel approach. + +# 3. Preamble + +Our goal in this section is to propose a mathematical formalism that closely aligns with experimental scenarios to address the following question: How can we introduce sufficient variance in low-texture areas while adhering as much as possible to the constancy assumption? The purpose is to provide a better basis than the usual framework [91] for learning depth and ego-motion. + +We start by introducing some key theoretical properties of the distance transform that we illustrate with toy experiments. The following definitions and theorems provide the theoretical foundation of our analysis. + +# 3.1. Maximal variance under constraints + +Notations Let $\Omega$ be a compact convex set of $\mathbb{R}^n$ with smooth boundary $\partial \Omega$ . Let $\mathcal{A}$ be the set of affine transformations. And let $\mathcal{I}_{\mathcal{A}}$ the set of positive functions invariant by affine transformation (see Definition 2). + +Here, $\Omega$ represents an object within a scene, while $\mathcal{A}$ models a reasonable subset of reprojected 3d transformations likely to affect the objects from frame to frame. + +Definition 1 The distance transform $\delta_{\Omega}:\Omega \to \mathbb{R}^{+}$ is defined as: + +$$ +\delta_ {\Omega} (x) = d (x, \partial \Omega) = \inf _ {y \in \partial \Omega} \| x - y \| \tag {1} +$$ + +The set of points having more than one closest point to the boundary is called the medial axis. By Rademacher's Theorem [53] the distance transform is differentiable almost everywhere (away of the medial axis) and satisfies the Eikonal equation: + +$$ +\| \nabla d \| = 1 \tag {2} +$$ + +Definition 2 Let $f: \Omega \to \mathbb{R}^+$ a smooth function on $\Omega$ . Then $f$ is said to be invariant by affine transformation if: + +$$ +\forall A \in \mathcal {A}, \forall x \in \Omega , f (A (x)) = f (x) \tag {3} +$$ + +Definition 2 restates the constancy assumption mathematically: if an affine transformation $A$ is applied to a convex shape $\Omega$ , then any changes introduce by $f$ on $\Omega$ should manifest identically in $A(\Omega)$ . + +Theorem 1 $\delta_{\Omega}$ is the unique solution, up to an isomorphism, to the following optimization problem: + +$$ +\max _ {f \in \mathcal {I} _ {\mathcal {A}}} \int_ {\Omega} (f (x) - \bar {f}) ^ {2} d x \quad s. t. \| \nabla f \| = 1 \tag {4} +$$ + +$\overline{f}$ being the average value of $f$ over $\Omega$ . A proof of Theorem 1 is given in Appendix A.1. This property is particularly significant because it demonstrates that, under the constraint of the Eikonal equation, the optimal method for increasing variance within a compact convex shape while adhering to the constancy assumption is to use the distance transform. For this reason, in our practical experiments, we propose to introduce variance within image instances primarily via the distance transform; further details can be found in Section 4.2. + +The Eikonal equation constraint acts as a standardization condition, promoting smooth variation. Without this constraint, partitioning the shape in uniform black and white regions would also maximize the variance. Our goal is instead to ensure a more even spread of variance throughout the shape, making the variance-augmented image reprojection more informative for better depth and ego-motion learning. We also explore carefully relaxing this constraint to investigate well-chosen functions of the distance transform that could allow for greater variance within instances while still maintaining an acceptable level of smoothness: + +$$ +\mathcal {F} = \left\{f: x \mapsto g \circ \delta_ {\Omega} (x) \mid g \in \mathcal {C} ^ {\infty} \left(\mathbb {R} ^ {+}, \mathbb {R} ^ {+}\right), \| g \| _ {\infty} \leq 1 \right\} \tag {5} +$$ + +A thorough comparison of depth estimation results across different functions is provided in Appendix C.2. The bound on the norm ensures that variance is not artificially amplified by larger values, maintaining consistency and fairness in comparisons. + +We show that the maximal variance property can be retrieved with a simple toy experiment in A.3. + +# 3.2. Convergence properties + +Notations Let $(I, \tilde{I})$ be two successive images of a camera in movement and $(x, y)$ a training instance where $x$ represents a pixel of $I$ and $y$ be the corresponding pixel in $\tilde{I}$ . Let us assume also that all pixels lie inside a convex shape $\Omega$ defined by a smooth and closed contour $\partial \Omega$ . Our goal is to learn the model $\Phi(x; \theta)$ to predict the new pixel location $\hat{y}$ . Let us define the cost function $l$ using functions from $\mathcal{F}$ (Equation 5): + +$$ +l: (\hat {y}, y) \mapsto (f (\hat {y}) - f (y)) ^ {2} \tag {6} +$$ + +with $f\in \mathcal{F}$ + +Definition 3 A function $f: \Omega \to \mathbb{R}^+$ is $\alpha$ -Lipschitz if: + +$$ +\forall u, v \quad | f (u) - f (v) | \leq \alpha \| u - v \| \tag {7} +$$ + +Definition 4 A function $f: \Omega \to \mathbb{R}^+$ is $\beta$ -smooth if: + +$$ +\forall u, v \quad \| \nabla f (u) - \nabla f (v) \| \leq \beta \| u - v \| \tag {8} +$$ + +The fact that $f$ is a function of the distance transform makes it possible to prove: + +Theorem 2 The loss function $l$ is, with respect to the first argument $\hat{y}$ : + +- $\alpha$ -Lipschitz +- $\beta$ -smooth +- strongly convex if a regularization term $\eta \| y\|^2$ is added. + +The constants $\alpha, \beta, \eta$ depend on the curvature at the projection onto the boundary $\partial \Omega$ . A proof of Theorem 2 is given in Appendix A.2. Those properties brought by the distance transform induce great convergence properties as described in [1, 27, 36]. Amongst others: a bounded generalization error and a uniform stability of the stochastic gradient descent. Note that the SSIM + colour function can also be Lipschitz and/or uniform under very strict conditions. However their constants, unlike with the distance transform, depend on the size of the shape $\Omega$ and can thus exceeds all bounds as mentioned in [57]. + +We show that the convergence property can be retrieved with a simple toy experiment in A.4. + +# 4. Method + +Our framework for self-supervised MDE (Figure 2) consists of three networks: Depth $D_{\theta}$ , Ego-motion $T_{\alpha}$ and edge $E_{\delta}$ . For the sake of readability, we will omit the parameter symbols $\theta, \alpha, \delta$ in the remainder of the section. Hyperparameters in the loss functions will also be omitted here and detailed in Appendix F. + +# 4.1. Pre-semantic contours + +Our edge network renders values $E(p)$ in the range [0, 1], where a higher value indicates a greater likelihood that the pixel $p$ corresponds to an edge. To generate pseudo-labels for supervising the edge network, we utilize depth and surface normal estimation, which is derived from depth following the methods described in [83, 85]. We discuss complementary insights of depth and surface normal for edge detection in Appendix D.1. + +Depth & Normal Pseudo-Labels: Depth and surface normals are smooth and thus prone to exhibiting zero-crossings of their Laplacian $\Delta D$ (resp. $\Delta N$ ) around semantic boundaries, defined as follows: + +$$ +Z _ {D} (p) = \left\{ \begin{array}{l l} 1 & \text {i f} \Delta D (p) \Delta D (p + (1, 0)) < 0 \\ 1 & \text {i f} \Delta D (p) \Delta D (p + (0, 1)) < 0 \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {9} +$$ + +Same applies to normal Laplacian $\Delta N$ . The zero-crossing mask is then dilated ( $3 \times 3$ structuring element), and multiplied by the magnitude of the gradient to create the pseudolabels. For this, the depth gradient is normalized to avoid + +![](images/afbebc2592a2f5d7e8c42e35b523cfefcb94d3c337e74fa2f1c6e822c77b86af.jpg) +Figure 2. Diagram showing the two main steps of our pipeline: estimation of the different modalities followed by the re-projection on the variance augmented image using depth and ego-motion. T indicates the time frame. R, t respectively rotation and translation predictions. + +![](images/e24a5a4f4751674e386e8ed84e3f6d5ff00f23ea0f1ec7ae35f22cef70d5e4ec.jpg) + +bias towards large depth values: + +$$ +\nabla D _ {x} (p) = \frac {D (p + (0 , 1)) - D (p)}{\max (D (p + (0 , 1)) , D (p))}, \tag {10} +$$ + +$$ +\| \nabla D (p) \| = | \nabla D _ {x} (p) | + | \nabla D _ {y} (p) |. +$$ + +The usual normal gradient, which is often too noisy, is replaced with a dot-product measure: + +$$ +\left| \nabla N _ {x} (p) \right| = 1 - \left| \langle \vec {N} (p + (0, 1)), \vec {N} (p) \rangle \right|, \tag {11} +$$ + +$$ +\left\| \nabla N (p) \right\| = \left| \nabla N _ {x} (p) \right| + \left| \nabla N _ {y} (p) \right|. +$$ + +The weighted depth and normal pseudo-labels employed to supervise the edge network are then defined as follows: + +$$ +w _ {d} (p) = \frac {Z _ {D} (p) \| \nabla D (p) \|}{\sum_ {p} Z _ {D} (p) \| \nabla D (p) \|} \tag {12} +$$ + +$$ +w _ {n} (p) = \frac {Z _ {N} (p) \| \nabla N (p) \|}{\sum_ {p} Z _ {N} (p) \| \nabla N (p) \|} \tag {13} +$$ + +A contrastive loss is also applied to encourage edge values to move toward either zero or one. This is achieved using a binary cross-entropy loss, where pseudo ground truth $Y(p) \in \{0,1\}$ are created by thresholding edge estimations $E(p)$ at 0.5: + +$$ +\mathcal {L} _ {c} = - \sum_ {p} \left(Y (p) \log (E (p)) + (1 - Y (p)) \log (1 - E (p))\right) \tag {14} +$$ + +Finally, the loss function employed to supervise the edge network is defined as follows: + +$$ +\mathcal {L} ^ {\mathrm {e d g e}} = \sum_ {p} \left(w _ {d} (p) + w _ {n} (p)\right) \left(1 - E (p)\right) + \mathcal {L} _ {c} + \mathcal {L} _ {e} +$$ + +with $\mathcal{L}_e = \sum_p E(p)^2$ to avoid trivial solutions. (15) + +Discontinuity Preserving Smoothing: Incorporating an edge-aware smoothness loss is a widely used practice in training depth networks [73, 83]: + +$$ +\mathcal {L} _ {s} = \sum_ {p} \left(| \nabla D _ {x} (p) | e ^ {- | \nabla I _ {x} (p) |} + | \nabla D _ {y} (p) | e ^ {- | \nabla I _ {y} (p) |}\right) \tag {16} +$$ + +However, this approach has two drawbacks: first, boundaries are estimated using the image gradient, which is not optimal; second, there is no mechanism to effectively manage discontinuities. To address these issues, we propose a new smoothness loss exploiting our pre-semantic contours defined as follows: + +$$ +\mathcal {L} _ {s _ {1}} = \sum_ {p} \| \nabla D (p) \| e ^ {- \alpha E (p)}, +$$ + +$$ +\mathcal {L} _ {s _ {2}} = \sum_ {p} \log \left(1 + e ^ {- \| \nabla D (p) \|}\right) \left(1 - e ^ {- \alpha E (p)}\right), \tag {17} +$$ + +$$ +\mathcal {L} _ {s} = \mathcal {L} _ {s _ {1}} + \mathcal {L} _ {s _ {2}}. +$$ + +The component $\mathcal{L}_{s_1}$ replace the former equation 16 to encourage depth smoothness inside semantic instances, while $\mathcal{L}_{s_2}$ prevents diffusion and promotes significant depth changes around boundaries. In this manner, both the edge and depth networks benefit from each other: the depth estimation supervises the edge network, while the edge network enhances depth estimation through the smoothness loss. + +Post-processing: The map produced by the edge network consists of continuous values ranging from zero to one. To obtain thin pre-semantic contours $\hat{\mathcal{C}}(I)$ , we apply a post-processing pipeline that includes hysteresis thresholding, non-maximum suppression, morphological closing, contour filtering, and contour closing. More details can be found in Appendix D.3 and Appendix D.2 discusses the advantages of our approach compared to the competitive work [84]. + +# 4.2. A variance augmented image for depth learning + +With $\widehat{\mathcal{C}}(I)$ the thin pre-semantic contour of $I$ obtained after edge network estimation and subsequent post-processing, we then compute the distance transform map $\delta_{\mathrm{I}}$ defined as: + +$$ +\delta_ {\mathrm {I}} (p) = \min _ {x \in \hat {\mathcal {C}} (I)} (\| x - p \|) \tag {18} +$$ + +See Appendix C.1 for the details of the algorithm. The resulting distance transform map is concatenated with the input image to create a 'variance-augmented' image. The objective is then to learn depth and ego-motion in order to identify corresponding pixels across successive frames, and reconstruct the upcoming structure-augmented image in the most effective manner, as illustrated on Figure 2. The reconstruction of the original input image $\widehat{\mathbf{l}}$ is evaluated using $\Psi$ , the standard combination of colour difference and SSIM, while the reconstruction of the distance transform map is assessed with a $\mathrm{L}_1$ loss: + +$$ +\mathcal {L} _ {\text {p h o t o}} = \Psi (\widehat {I}, I), \text {a n d} \tag {19} +$$ + +$$ +\mathcal {L} _ {\mathrm {d i s t}} = \left\| \widehat {\delta_ {\mathrm {I}}} - \delta_ {\mathrm {I}} \right\| _ {1} +$$ + +The smoothness loss is added to get the final supervision signal of the depth network: + +$$ +\mathcal {L} ^ {\text {d e p t h}} = \mathcal {L} _ {\text {d i s t}} + \mathcal {L} _ {\text {p h o t o}} + \mathcal {L} _ {s} \tag {20} +$$ + +An improved distance transform: This approach can be readily extended to the family of one-dimensional functions $\mathcal{F} = \{p\mapsto g\circ \delta_{\mathrm{I}}(p)\}$ (as defined in Equation 5): + +$$ +\mathcal {L} _ {\text {d i s t}} = \left\| g \circ \widehat {\delta_ {\mathrm {I}}} - g \circ \delta_ {\mathrm {I}} \right\| _ {1}. \tag {21} +$$ + +We provide an analysis of the depth estimation results for various functions in Appendix C.2. To take this a step further, we propose extending the distance transform into the $n$ -dimensional space by mapping the distance transform values to an $n$ -dimensional random walk process denoted as $\mathrm{RW}_n$ : + +$$ +\mathcal {L} _ {\text {d i s t}} = \sum_ {p} \left\| \mathrm {R W} _ {n} \circ \widehat {\delta_ {\mathrm {I}}} (p) - \mathrm {R W} _ {n} \circ \delta_ {\mathrm {I}} (p) \right\| _ {1} \tag {22} +$$ + +Using this approach, we can maintain the validity of the constancy assumption while increasing the variance even more. Additionally, the dimensional parameter $n$ determines the significance of the distance transform map reconstruction within the total loss function. It introduces a tradeoff between $\mathcal{L}_{\mathrm{dist}}$ and $\mathcal{L}_{\mathrm{photo}}$ : the higher the dimension, the greater the emphasis on the artificially introduced structure compared to the natural structure found in the image; see Appendix C.3 for more details on the random walk mapping. Our experiments in Appendix C.2 show that $n = 3$ provides the best results. + +# 5. Experiments + +# 5.1. Datasets + +We conduct experiments on multiple datasets: + +KITTI [20] includes diverse urban, rural, and highway scenes. We follow the standard data split defined by [13] and apply the pre-processing steps outlined in [91] to filter static frames, resulting in a training set of 39810 images and 697 images for testing. Results are reported using the evaluation protocol from [91]. + +Cityscapes [9] offers urban scenes with dynamic objects, challenging depth estimation. The training set consists of 22973 images, with evaluation on 1525 test images following [5, 41]. Edge detection is assessed using the contour evaluation from [84] on 500 validation frames. + +Waymo [69] is a large and diverse autonomous driving dataset that captures dynamic urban scenes under various environmental conditions, including nighttime and different weather scenarios. We sample 100 000 image pairs from 1 000 front camera video sequences for training and evaluate on 1500 pairs from 150 sequences, following [41]. + +NYUv2 [67] consists of 464 indoor video sequences captured with a Kinect sensor. We use the official splits (302 training, 33 validation) and evaluate on 654 densely annotated test images, resizing inputs to $320 \times 256$ pixels as in [44]. + +ScanNet [10] is a large-scale RGB-D dataset with 2.5 million images across $1500+$ scenes. We use it for zero-shot evaluation, following [16, 44], with test images resized to $320 \times 256$ pixels. + +# 5.2. Implementation Details + +We emphasize that our method is compatible with any DNN architecture and input image size. For fair comparison, our experiments are done with a $256 \times 832$ resolution and a UNet [62] structure, following [91], unless stated otherwise. The depth and edge networks share a ResNet [29] encoder. Our baseline, CoopNet [28], handles moving objects—especially useful in Cityscapes—via self-supervised flow. The pose network, based on a ResNet encoder, outputs a 6-DoF vector. All encoders use ResNet-50 backbones initialized with ImageNet [63] weights. Training runs for 30 epochs with batch size 4, initial learning rate of $10^{-4}$ (decreasing to $10^{-5}$ after 20 epochs), and standard data augmentation from [22]. Loss weights, tuned via grid search, are detailed in Appendix F. Our PyTorch [58] implementation trains on a single NVIDIA RTX A5000 GPU using Adam [38] with $\beta_{1} = 0.99$ and $\beta_{2} = 0.999$ and takes 12 hours. + +# 5.3. Results + +Depth Quantitative results on KITTI and Cityscapes are presented in Table 1, where we compare against the most + +
MethodLower is better ↓Higher is better ↑
Abs RelSq RelRMSERMSE logδ1δ2δ3
Monodepth2[22]0.1100.8314.6420.1870.8830.9620.982
Guizilini et al.†[24] (R50)0.1130.8314.6630.1890.8780.9710.983
Shu et al.*[66](R50)0.1080.7924.6330.1840.8830.9610.983
Ours (R50)0.1040.7254.4530.1800.8850.9620.983
MonoViT[89]0.0990.7084.3720.1750.9000.9670.984
Ours + MonoViT0.0920.6744.3000.1650.9270.9670.984
HR-Depth[49]0.1090.7924.6320.1850.8840.9620.983
RA-Depth[30]0.0960.6134.2160.1710.9030.9680.985
DIFFNet[90]0.1020.7644.4830.1800.8960.9650.983
Guizilini et al.†[24]0.1000.7614.2700.1750.9020.9650.982
Ours+DinoV20.0820.6044.1080.1620.9280.9680.985
Lego (paper)0.1541.2726.0120.230N/AN/AN/A
Lego★ (R50)0.1150.8554.7890.1950.8760.9570.979
Lego★ (R50) + Ldist0.1100.8054.6060.1880.8820.9600.982
Ours (R50)0.1040.7254.4530.1800.8850.9620.983
Struct2Depth†[4]0.1451.7377.280.2050.8130.9420.978
LearnK†[41]0.1271.3306.960.1950.8300.9470.981
Li et al.†[41]0.1191.2906.980.1900.8460.9510.980
CoopNet[28]0.1211.4437.010.1900.8460.9510.980
Ours (R50)0.1151.2216.790.1860.8500.9550.981
+ +Table 1. Results of depth estimations on KITTI 2015 (first four blocks) and Cityscapes (last block). For each metric the best result is displayed in bold and the second-best is underlined. Light blue metrics indicates the most challenging metrics. $\dagger$ denotes the use of an off-the-shelf semantic algorithm and * indicates results reproduced by us at the input resolution of $256 \times 832$ . Block#1 and #2 compares methods using similar encoder backbones: Resnet 50 (R50) in block#1 and ViT in block#2. Block#3 compares SOTA depth models without restrictions on architecture or additional inputs, Block#4 compares to Lego. + +
MethodDataset (Train / Test)Lower is better ↓
Abs RelSq RelRMSERMSE log
LearnK[41]Waymo / Waymo0.1571.5317.0900.205
OursWaymo / Waymo0.1251.2785.9900.189
GLNet[7]Cityscapes / KITTI0.1291.0445.3610.212
CoopNet[28]Cityscapes / KITTI0.1251.1575.2510.209
OursCityscapes / KITTI0.1160.8924.8920.195
+ +Table 2. Results on Outdoor datasets. Block#1: in-domain evaluation. Block#2: out-of-domain evaluation. + +
MethodDataset (Train / Test)Abs RelRMSEδ1δ2δ3
MonoIndoor++[44]NYUv2 / NYUv20.1320.5170.8340.9610.990
IndoorDepth[16]NYUv2 / NYUv20.1260.4940.8450.9650.991
OursNYUv2 / NYUv20.1150.4580.8590.9700.992
IndoorDepth[16]NYUv2 / ScanNet0.1530.3730.7860.9500.988
MonoIndoor++[44]NYUv2 / ScanNet0.1380.3470.8100.9670.993
OursNYUv2 / ScanNet0.1270.3120.8430.9700.993
+ +competitive methods. Blocks #1 and #2 demonstrate that, when using the same encoder backbone, our method outperforms others by a significant margin across nearly all metrics. When no architectural constraints are imposed, we + +Table 3. Results on Indoor datasets. Block#1: in-domain evaluation. Block#2: out-of-domain evaluation. + +
RW3LdistLsEθAbs Rel (↓)RMSE log (↓)δ1(↑)
0.1130.190.878
✓0.1120.1890.880
✓✓0.1100.1870.881
✓✓✓0.1060.1830.884
✓✓✓✓0.1040.1800.885
+ +Table 4. Ablation study on KITTI 2015 of losses mentioned in Section 4.2. Note that $E_{\theta}$ is part of the ablation as it does improve the depth metrics and shares the same encoder as the depth network. $\mathrm{RW}_3$ is for 3d random walk encoding of distance transform as defined in Section 4.2. + +employ a pre-trained DinoV2 [56] encoder (further details in Appendix F.2), yielding the results in block #3. For fairness, we compare against $\mathsf{LEGO}^{\star}$ in block #4, an enhanced version of LEGO that incorporates a ResNet-50 backbone and up-to-date training techniques. Adding our novel distance transform loss $\mathcal{L}_{\mathrm{dist}}$ into $\mathsf{LEGO}^{\star}$ greatly improves results, showing once again the great importance of this loss component, but still underperforms ours, likely due to the superior quality of our contours. + +We also provide strong out-of-domain results in Table 2 and Table 3, demonstrating competitive performance on both indoor and outdoor datasets. Finally, an ablation study in Table 4 evaluates the impact of the different loss components introduced in our approach and in Appendix C.2 for different functions from the family $\mathcal{F}$ as defined in Section 3.1. It can be seen that the reprojection on the distance transform via $\mathcal{L}_{\mathrm{dist}}$ brings a decisive improvement. We provide a qualitative comparison of results in Figure 3. We can see that our method gives better results: it provides thinner object, is sharper around moving objects, does not suffer from smoothness issues for large objects and is not noisy in textured areas. More qualitative and quantitative results can be found in Appendix B.1. + +Flow & Odometry As discussed in Section 5.2 we use an optical flow to handle moving objects following [28]. Its results can be found in Appendix B.2. Results for odometry can be found in Appendix B.3. Our method significantly outperforms the baseline [28] on both optical flow and odometry metrics. + +# 5.4. Contour evaluation + +We evaluate contour detection in Table 5 using the same metrics and data as [84]: Optimal Dataset Scale, Optimal Image Scale, and Average Precision, computed on 500 Cityscapes validation images. Our method significantly outperforms LEGO, leveraging (1) depth & normal pseudolabels with Laplacian zero-crossing insights for better contour alignment, (2) the discontinuity-preserving smoothing loss $\mathcal{L}_s$ , and (3) the contrastive loss $\mathcal{L}_c$ . Qualitative examples are in Figure 1, with more in Appendix D. + +![](images/771318ed04d5bfb0f1facc9876755a4eaeb9b5d9e7315940173f08b7db155240.jpg) +Figure 3. Qualitative comparisons for different methods. First row: colour images. Second row: Monodepth2 [22]. Second row: CoopNet [28]. Third row: FeatDepth [66]. Last row: Ours. We compare qualitative results in four complex situations. First column: a thin object. Second column: a moving object. Third column: a large object with smoothness issue. Last column: Textured areas. + +
MethodsODSOISAP
Lego (paper)0.7100.7310.729
Lego*0.7090.7360.734
Depth & Normal pseudo-labels0.7550.7750.760
Depth & Normal pseudo-labels + Ls0.7600.7820.766
Ours = Depth & Normal pseudo-labels + Ls + Lc0.7620.7890.770
+ +Table 5. Comparison for contour estimation. Block#1: Lego results. Block#2: Ablation study of losses in contour estimation training. Experiments done on Cityscapes. + +# 5.5. Constancy assumption validity + +We evaluate the constancy assumption validity of our method, compared to [66] that uses deep features from the ResNet-50 encoder in the re-projection. To do so, we use the KITTI MOTS [72] dataset to track a specific point of an instance along different trajectories and compute the normalized temporal variance. We show that our distance transform satisfies much better the constancy assumption than any layers of the ResNet-50 (see Figure 4). More details on the evaluation protocol are given in Appendix E. + +# 6. Conclusion + +As a conclusion, this paper presents a novel self-supervised approach to monocular depth estimation, addressing challenges in low-texture regions by applying a distance transform over pre-semantic contours. This technique enhances depth prediction accuracy by increasing discrimini + +![](images/bda0b3345ff6e571496b7f43ea9f3144caf766d7eba9c5a1b5e05446371c42c1.jpg) +Figure 4. Comparison of normalized temporal variance of different features for the centre pixel of an instance tracked along different sequence of KITTI MOTS [72]. + +native power where traditional photometric losses struggle. Both theoretical analysis and extensive experiments validate the effectiveness of the proposed method, which outperforms conventional self-supervised approaches. + +In future work, we aim to introduce consistent variance in low-texture regions without pre-semantic contours. Additionally, we plan to explore whether our framework can leverage variance-augmented images to learn from diverse datasets, ultimately developing a foundation model in a fully self-supervised manner. + +# References + +[1] Ali Akbari, Muhammad Awais, Manijeh Bashar, and Josef Kittler. How does loss function affect generalization performance of deep learning? application to human age estimation. In International Conference on Machine Learning, pages 141-151. PMLR, 2021. 4 +[2] Abhishek Badki, Alejandro Troccoli, Kihwan Kim, Jan Kautz, Pradeep Sen, and Orazio Gallo. Bi3d: Stereo depth estimation via binary classifications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1600-1608, 2020. 1 +[3] Jiawang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised scale-consistent depth and ego-motion learning from monocular video. Advances in neural information processing systems, 32, 2019. 6 +[4] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Depth prediction without the sensors: Leveraging structure for unsupervised learning from monocular videos. In Proceedings of the AAAI conference on artificial intelligence, pages 8001-8008, 2019. 2, 7, 3, 4 +[5] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Unsupervised monocular depth and ego-motion learning with structure and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 6 +[6] Tian Chen, Shijie An, Yuan Zhang, Chongyang Ma, Huayan Wang, Xiaoyan Guo, and Wen Zheng. Improving monocular depth estimation by leveraging structural awareness and complementary datasets. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 90-108. Springer, 2020. 1 +[7] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 1, 2, 7, 5 +[8] Zhao Chen, Vijay Badrinarayanan, Gilad Drozdov, and Andrew Rabinovich. Estimating depth from rgb and sparse sensing. In Proceedings of the European Conference on Computer Vision (ECCV), pages 167-182, 2018. 3 +[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 2, 6 +[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6 +[11] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In Proceedings of the European conference on computer vision (ECCV), pages 562-578, 2018. 2 + +[12] Tom van Dijk and Guido de Croon. How do neural networks see depth in single images? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2183-2191, 2019. 1 +[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 1, 6 +[14] F El Jamiy and R Marsh. Survey on depth perception in head mounted displays: distance estimation in virtual reality, augmented reality, and mixed reality. iet image processing 13 (5), 707-712 (2019), 2018. 1 +[15] Sara Elkerdawy, Hong Zhang, and Nilanjan Ray. Lightweight monocular depth estimation model by joint end-to-end filter pruning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 4290-4294. IEEE, 2019. 1 +[16] Chao Fan, Zhenyu Yin, Yue Li, and Feiqing Zhang. Deeper into self-supervised monocular indoor depth estimation. arXiv preprint arXiv:2312.01283, 2023. 6, 7 +[17] Philipp Fischer, Alexey Dosovitskiy, Eddy Ilg, Philip Häusser, Caner Hazirbaş, Vladimir Golkov, Patrick Van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. arXiv preprint arXiv:1504.06852, 2015. 5 +[18] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 1 +[19] Ravi Garg, Vijay Kumar Bg, Gustavo Carneiro, and Ian Reid. Unsupervised cnn for single view depth estimation: Geometry to the rescue. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 740-756. Springer, 2016. 1 +[20] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 6 +[21] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 1 +[22] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 2, 6, 7, 8 +[23] Ariel Gordon, Hanhan Li, Rico Jonschkowski, and Anelia Angelova. Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8977-8986, 2019. 1, 2 +[24] Vitor Guizilini, Rui Hou, Jie Li, Rares Ambrus, and Adrien Gaidon. Semantically-guided representation learn + +ing for self-supervised monocular depth. arXiv preprint arXiv:2002.12319, 2020. 2, 7 +[25] Hyowon Ha, Sunghoon Im, Jaesik Park, Hae-Gon Jeon, and In So Kweon. High-quality depth from uncalibrated small motion clip. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5413-5421, 2016. 1 +[26] David Hafner, Oliver Demetz, and Joachim Weickert. Why is the census transform good for robust optic flow computation? In Scale Space and Variational Methods in Computer Vision: 4th International Conference, SSVM 2013, Schloss Seggau, Leibnitz, Austria, June 2-6, 2013. Proceedings 4, pages 210-221. Springer, 2013. 3 +[27] Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International conference on machine learning, pages 1225-1234. PMLR, 2016. 4 +[28] Marwane Hariat, Antoine Manzanera, and David Filliat. Rebalancing gradient to improve self-supervised co-training of depth, odometry and optical flow predictions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1267-1276, 2023. 2, 6, 7, 8, 4, 5 +[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[30] Mu He, Le Hui, Yikai Bian, Jian Ren, Jin Xie, and Jian Yang. Ra-depth: Resolution adaptive self-supervised monocular depth estimation. In European Conference on Computer Vision, pages 565-581. Springer, 2022. 7 +[31] Stuart B Heinrich and Wesley E Snyder. Improved edge awareness in discontinuity preserving smoothing. arXiv preprint arXiv:1103.5808, 2011. 3 +[32] Junjie Hu, Yan Zhang, and Takayuki Okatani. Visualization of convolutional neural networks for monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3869-3878, 2019. 1 +[33] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2462-2470, 2017. 5 +[34] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 2 +[35] Wei Jiang, Weiwei Sun, Andrea Tagliasacchi, Eduard Trulls, and Kwang Moo Yi. Linearized multi-sampling for differentiable image transformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2988-2997, 2019. 2 +[36] Kenji Kawaguchi and Qingyun Sun. A recipe for global convergence guarantee in deep neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 8074-8082, 2021. 4 +[37] Ue-Hwan Kim, Gyeong-Min Lee, and Jong-Hwan Kim. Revisiting self-supervised monocular depth estimation. In International Conference on Robot Intelligence Technology and Applications, pages 336-350. Springer, 2021. 2 + +[38] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[39] Yevhen Kuznietsov, Marc Proesmans, and Luc Van Gool. Comoda: Continuous monocular depth adaptation using past experiences. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2907-2917, 2021. 2 +[40] Seokju Lee, Sunghoon Im, Stephen Lin, and In So Kweon. Learning residual flow as dynamic motion from stereo videos. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1180-1186. IEEE, 2019. 2 +[41] Hanhan Li, Ariel Gordon, Hang Zhao, Vincent Casser, and Anelia Angelova. Unsupervised monocular depth learning in dynamic scenes. In Conference on Robot Learning, pages 1908-1917. PMLR, 2021. 2, 6, 7 +[42] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. Rtm3d: Real-time monocular 3d detection from object keypoints for autonomous driving. In European Conference on Computer Vision, pages 644-660. Springer, 2020. 1 +[43] Pengzhi Li, Yikang Ding, Haohan Wang, Chengshuai Tang, and Zhiheng Li. The devil is in the edges: Monocular depth estimation with edge-aware consistency fusion. arXiv preprint arXiv:2404.00373, 2024. 2 +[44] Runze Li, Pan Ji, Yi Xu, and Bir Bhanu. Monoindoor++: Towards better practice of self-supervised monocular depth estimation for indoor environments. IEEE Transactions on Circuits and Systems for Video Technology, 33(2):830-846, 2022. 6, 7 +[45] Rui Li, Danna Xue, Shaolin Su, Xiantuo He, Qing Mao, Yu Zhu, Jinqiu Sun, and Yanning Zhang. Learning depth via leveraging semantics: Self-supervised monocular depth estimation with both implicit and explicit semantic guidance. Pattern Recognition, 137:109297, 2023. 2, 3 +[46] Yin Li, Manohar Paluri, James M Rehg, and Piotr Dólar. Unsupervised learning of edges. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1619-1627, 2016. 3 +[47] Liang Liu, Guangyao Zhai, Wenlong Ye, and Yong Liu. Unsupervised learning of scene flow estimation fusing with local rigidity. In IJCAI, pages 876-882, 2019. 2 +[48] Hengjie Lu, Shugong Xu, and Shan Cao. Sgtbn: generating dense depth maps from single-line lidar. IEEE Sensors Journal, 21(17):19091-19100, 2021. 1 +[49] Xiaoyang Lyu, Liang Liu, Mengmeng Wang, Xin Kong, Lina Liu, Yong Liu, Xinxin Chen, and Yi Yuan. Hr-depth: High resolution self-supervised monocular depth estimation. In Proceedings of the AAAI conference on artificial intelligence, pages 2294-2301, 2021. 7 +[50] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3 +[51] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 6 + +[52] Oskar Natan and Jun Miura. End-to-end autonomous driving with semantic depth cloud mapping and multi-agent. IEEE Transactions on Intelligent Vehicles, 8(1):557-571, 2022. 1 +[53] Ales Nekvinda and Ludek Zajicek. A simple proof of the rademacher theorem. Časopis pro pěstovánt matematíky, 113 (4):337-341, 1988. 3 +[54] C Wayne Niblack, Phillip B Gibbons, and David W Capson. Generating skeletons and centerlines from the distance transform. CVGIP: Graphical Models and image processing, 54 (5):420-437, 1992. 2 +[55] Jim Nilsson and Tomas Akenine-Möller. Understanding ssm. arXiv preprint arXiv:2006.13846, 2020. 1 +[56] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7 +[57] Daniel Otero, Davide La Torre, O Michailovich, and Edward R Vrscay. Optimization of structural similarity in mathematical imaging. Optimization and Engineering, 22:2367-2401, 2021. 4 +[58] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[59] Michael Ramamonjisoa and Vincent Lepetit. Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 2 +[60] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12240-12249, 2019. 2 +[61] Alex Rasla and Michael Beyeler. The relative importance of depth cues and semantic edges for indoor mobility using simulated prosthetic vision in immersive virtual reality. In Proceedings of the 28th ACM Symposium on Virtual Reality Software and Technology, pages 1-11, 2022. 1 +[62] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6, 3 +[63] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6 +[64] Faraz Saeedan and Stefan Roth. Boosting monocular depth with panoptic segmentation maps. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3853-3862, 2021. 2 + +[65] Abhijeet Shenoi, Mihir Patel, JunYoung Gwak, Patrick Goebel, Amir Sadeghian, Hamid Rezatofighi, Roberto Martin-Martin, and Silvio Savarese. Jrmot: A real-time 3d multi-object tracker and a new large-scale dataset. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10335-10342. IEEE, 2020. 1 +[66] Chang Shu, Kun Yu, Zhixiang Duan, and Kuiyuan Yang. Feature-metric loss for self-supervised learning of depth and egomotion. In European Conference on Computer Vision, pages 572-588. Springer, 2020. 3, 7, 8 +[67] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 2, 6 +[68] Tilo Strutz. The distance transform and its computation. arXiv preprint arXiv:2106.03503, 2021. 2 +[69] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 6 +[70] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 1 +[71] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3 +[72] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. MOTS: Multi-object tracking and segmentation. In CVPR, 2019. 8, 6, 7, 11 +[73] Chaoyang Wang, José Miguel Buenapasada, Rui Zhu, and Simon Lucey. Learning depth from monocular videos using direct methods. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2022–2030, 2018. 2, 5 +[74] Peng Wang, Xiaohui Shen, Bryan Russell, Scott Cohen, Brian Price, and Alan L Yuille. Surge: Surface regularized geometry estimation from a single image. Advances in Neural Information Processing Systems, 29, 2016. 2 +[75] Yang Wang, Yi Yang, Zhenheng Yang, Liang Zhao, Peng Wang, and Wei Xu. Occlusion aware unsupervised learning of optical flow. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4884-4893, 2018. 2 +[76] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 1 +[77] Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1164-1174, 2021. 2 + +[78] Diana Wofk, Fangchang Ma, Tien-Ju Yang, Sertac Karaman, and Vivienne Sze. Fastdepth: Fast monocular depth estimation on embedded systems. In 2019 International Conference on Robotics and Automation (ICRA), pages 6101-6108. IEEE, 2019. 1 +[79] Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single image depth prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 611-620, 2020. 2, 3 +[80] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 2 +[81] Wanpeng Xu, Ling Zou, Lingda Wu, and Zhipeng Fu. Self-supervised monocular depth learning in low-texture areas. Remote Sensing, 13(9):1673, 2021. 3 +[82] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 7 +[83] Z Yang, P Wang, W Xu, L Zhao, and R Nevatia. Unsupervised learning of geometry with edge-aware depth-normal consistency. arxiv 2017. arXiv preprint arXiv:1711.03665, 2 (5). 4, 5 +[84] Zhenheng Yang, Peng Wang, Yang Wang, Wei Xu, and Ram Nevatia. Lego: Learning edge with geometry all at once by watching videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 225-234, 2018. 3, 5, 6, 7 +[85] Wei Yin, Yifan Liu, Chunhua Shen, and Youliang Yan. Enforcing geometric constraints of virtual normal for depth prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5684-5693, 2019. 4 +[86] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1983-1992, 2018. 2, 5 +[87] Xun Yuan and Song Chen. Sad-slam: A visual slam based on semantic and depth information. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4930-4935. IEEE, 2020. 1 +[88] Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 340-349, 2018. 3 +[89] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. In 2022 international conference on 3D vision (3DV), pages 668-678. IEEE, 2022. 7 +[90] Hang Zhou, David Greenwood, and Sarah Taylor. Self-supervised monocular depth estimation with internal feature fusion. In *British Machine Vision Conference (BMVC)*, 2021. 7 + +[91] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 1, 2, 3, 6 +[92] Shengjie Zhu, Garrick Brazil, and Xiaoming Liu. The edge of depth: Explicit constraints between segmentation and depth. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13116-13125, 2020. 2, 3 +[93] Yuliang Zou, Zelun Luo, and Jia-Bin Huang. Df-net: Unsupervised joint learning of depth and flow using cross-task consistency. In Proceedings of the European conference on computer vision (ECCV), pages 36-53, 2018. 2 \ No newline at end of file diff --git a/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/images.zip b/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a970122fa361ef6a7ab535204d5e1c1b0e1631f1 --- /dev/null +++ b/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5dd91496031a2c0105b82a63f6d568412ea1cb1e8cedeb1f70147a080488c434 +size 585205 diff --git a/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/layout.json b/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1c1a8f76496fe5b238aff36f9613e64fca143e6f --- /dev/null +++ b/2025/Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks/layout.json @@ -0,0 +1,11227 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 517, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 517, + 140 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 517, + 140 + ], + "type": "text", + "content": "Improved monocular depth prediction using distance transform over pre-semantic contours with self-supervised neural networks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 160, + 516, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 160, + 516, + 205 + ], + "spans": [ + { + "bbox": [ + 93, + 160, + 516, + 205 + ], + "type": "text", + "content": "Marwane Hariat1, Antoine Manzanera1, David Filliat1 \n1U2IS, ENSTA, Institut Polytechnique de Paris, Palaiseau, France \n{marwane.hariat, antoine.manzanera, david fillliat}@ensta.fr" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 230, + 173, + 262 + ], + "blocks": [ + { + "bbox": [ + 56, + 230, + 173, + 262 + ], + "lines": [ + { + "bbox": [ + 56, + 230, + 173, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 173, + 262 + ], + "type": "image", + "image_path": "73ce385542563d79605072ecd3c823dc265a3e81cd94c19e2dcc7cc1476d20a0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 268, + 172, + 300 + ], + "blocks": [ + { + "bbox": [ + 56, + 268, + 172, + 300 + ], + "lines": [ + { + "bbox": [ + 56, + 268, + 172, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 172, + 300 + ], + "type": "image", + "image_path": "5e292b9f0aa921153a7b14ca2f5fbda63b5ea1f7ceecd46f0c0700c6b11402d2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 306, + 172, + 338 + ], + "blocks": [ + { + "bbox": [ + 56, + 306, + 172, + 338 + ], + "lines": [ + { + "bbox": [ + 56, + 306, + 172, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 306, + 172, + 338 + ], + "type": "image", + "image_path": "af0834f4151e38c5d13f50d1a97ba7c93e8a9f5335c9b7200e343455e140d46c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 339, + 136, + 348 + ], + "lines": [ + { + "bbox": [ + 93, + 339, + 136, + 348 + ], + "spans": [ + { + "bbox": [ + 93, + 339, + 136, + 348 + ], + "type": "text", + "content": "RGB images" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 183, + 230, + 300, + 262 + ], + "blocks": [ + { + "bbox": [ + 183, + 230, + 300, + 262 + ], + "lines": [ + { + "bbox": [ + 183, + 230, + 300, + 262 + ], + "spans": [ + { + "bbox": [ + 183, + 230, + 300, + 262 + ], + "type": "image", + "image_path": "1a28d1d8abc98af0348b84aed1b901258ce17e23d4c27bf208185a02163f29cd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 184, + 268, + 299, + 300 + ], + "blocks": [ + { + "bbox": [ + 184, + 268, + 299, + 300 + ], + "lines": [ + { + "bbox": [ + 184, + 268, + 299, + 300 + ], + "spans": [ + { + "bbox": [ + 184, + 268, + 299, + 300 + ], + "type": "image", + "image_path": "abf7d38042583f22379a5deb90829bfd5a4707f589c62d7b8351b7d0cb30cbbd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 184, + 307, + 299, + 338 + ], + "blocks": [ + { + "bbox": [ + 184, + 307, + 299, + 338 + ], + "lines": [ + { + "bbox": [ + 184, + 307, + 299, + 338 + ], + "spans": [ + { + "bbox": [ + 184, + 307, + 299, + 338 + ], + "type": "image", + "image_path": "edc5ede1827ecb359dc7b8c0cb0ac8cb8db77b418c98a31a3e5c833b3cab10b9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 339, + 252, + 348 + ], + "lines": [ + { + "bbox": [ + 231, + 339, + 252, + 348 + ], + "spans": [ + { + "bbox": [ + 231, + 339, + 252, + 348 + ], + "type": "text", + "content": "Depth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 358, + 550, + 370 + ], + "lines": [ + { + "bbox": [ + 58, + 358, + 550, + 370 + ], + "spans": [ + { + "bbox": [ + 58, + 358, + 550, + 370 + ], + "type": "text", + "content": "Figure 1. The different modalities estimated in our self-supervised framework and used to reduce matching uncertainty between images." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 230, + 428, + 259 + ], + "blocks": [ + { + "bbox": [ + 310, + 230, + 428, + 259 + ], + "lines": [ + { + "bbox": [ + 310, + 230, + 428, + 259 + ], + "spans": [ + { + "bbox": [ + 310, + 230, + 428, + 259 + ], + "type": "image", + "image_path": "ca538a1b33f183da4755c3f0c6565799ccbeea122b1678de8dd143ee9d1ad2d5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 267, + 427, + 300 + ], + "blocks": [ + { + "bbox": [ + 310, + 267, + 427, + 300 + ], + "lines": [ + { + "bbox": [ + 310, + 267, + 427, + 300 + ], + "spans": [ + { + "bbox": [ + 310, + 267, + 427, + 300 + ], + "type": "image", + "image_path": "80ead54ef523e6a6c1e2dc06b3cf50094fad20aea5e264c119599434bc5d1e51.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 306, + 428, + 338 + ], + "blocks": [ + { + "bbox": [ + 310, + 306, + 428, + 338 + ], + "lines": [ + { + "bbox": [ + 310, + 306, + 428, + 338 + ], + "spans": [ + { + "bbox": [ + 310, + 306, + 428, + 338 + ], + "type": "image", + "image_path": "f14ef0f810705e78dd61b7145a061d339cd304907cf6b3ddbf77c532b8d7adab.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 339, + 377, + 349 + ], + "lines": [ + { + "bbox": [ + 359, + 339, + 377, + 349 + ], + "spans": [ + { + "bbox": [ + 359, + 339, + 377, + 349 + ], + "type": "text", + "content": "Edge" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 438, + 231, + 554, + 262 + ], + "blocks": [ + { + "bbox": [ + 438, + 231, + 554, + 262 + ], + "lines": [ + { + "bbox": [ + 438, + 231, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 438, + 231, + 554, + 262 + ], + "type": "image", + "image_path": "fe5fc609326d3ae92032ae0e85bb86ca4596462ec2f0e899d5d4b7926b779f3f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 438, + 268, + 553, + 300 + ], + "blocks": [ + { + "bbox": [ + 438, + 268, + 553, + 300 + ], + "lines": [ + { + "bbox": [ + 438, + 268, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 438, + 268, + 553, + 300 + ], + "type": "image", + "image_path": "7bcf76575f83f537b82df78db03aac86140c15025238ad7c4fb098720c582884.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 438, + 307, + 553, + 338 + ], + "blocks": [ + { + "bbox": [ + 438, + 307, + 553, + 338 + ], + "lines": [ + { + "bbox": [ + 438, + 307, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 438, + 307, + 553, + 338 + ], + "type": "image", + "image_path": "265b82a705b3e3b4a395056ff8cccaea227c5efbbf55bb4ee578e4b07e91a34f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 339, + 528, + 348 + ], + "lines": [ + { + "bbox": [ + 463, + 339, + 528, + 348 + ], + "spans": [ + { + "bbox": [ + 463, + 339, + 528, + 348 + ], + "type": "text", + "content": "Distance Transform" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 151, + 380, + 200, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 380, + 200, + 392 + ], + "spans": [ + { + "bbox": [ + 151, + 380, + 200, + 392 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 53, + 405, + 297, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 405, + 297, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 405, + 297, + 622 + ], + "type": "text", + "content": "Monocular depth estimation (MDE) with self-supervised training approaches struggles in low-texture areas, where photometric losses may lead to ambiguous depth predictions. To address this, we propose a novel technique that enhances spatial information by applying a distance transform over pre-semantic contours, augmenting discriminative power in low texture regions. Our approach jointly estimates pre-semantic contours, depth and ego-motion. The pre-semantic contours are leveraged to produce new input images, with variance augmented by the distance transform in uniform areas. This approach results in more effective loss functions, enhancing the training process for depth and ego-motion. We demonstrate theoretically that the distance transform is the optimal variance-augmenting technique in this context. Through extensive experiments on KITTI, Cityscapes, Waymo, NYUv2 and ScanNet our model demonstrates robust performance, surpassing competing self-supervised methods in MDE." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 645, + 135, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 135, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 135, + 657 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "The goal of monocular depth estimation (MDE) is to determine the distance of a pixel from the camera focal plane using only a single input image, unlike stereo depth estimation [2, 21, 25], which relies on pairs of images. This" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 380, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 488 + ], + "type": "text", + "content": "eliminates the need for expensive sensors that require high-power resources [42, 48], such as active depth camera or LiDAR, while maintaining low latency and ensuring real-time compliance [15, 78]. This makes MDE a core module in applications such as robot navigation [65, 87], autonomous driving [52, 76] or virtual/augmented reality [14, 61]. Deep Neural Networks (DNNs) are currently the state of the art for depth maps prediction thanks to their ability to exploit all contextual cues of the environment [12, 32]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 489, + 556, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 556, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 556, + 705 + ], + "type": "text", + "content": "Supervised methods for MDE rely on ground-truth annotations. In spite of huge progress [6, 13, 18, 70], these methods are constrained by the cost and quality of the annotations. In this paper, we focus on self-supervised MDE, that uses the structure-from-motion framework [19, 91] and leverages depth and camera ego-motion estimations from DNNs to reconstruct the next frame from the current one. The supervision signal is then provided by the photometric loss, i.e. the difference between the reconstructed frame and the observed one, based on a combination of direct colour comparison and Structural Similarity Index Measure (SSIM) [55]. Although more challenging, the self-supervised paradigm offers greater training flexibility. First, it can be trained using data from unknown cameras [23], as long as data acquisition provides successive frames, such as in YouTube videos [7]. This is particularly relevant given the growing interest in foundation models, which can now leverage large datasets to produce generalizable out-" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21868" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "puts [82]. In addition, it is compatible with continual learning, allowing the model to continuously expand its knowledge as more data become available [7, 39]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 294, + 251 + ], + "type": "text", + "content": "However, the problem of MDE using the photometric loss proposed by Zhou et al. [91] is ill-posed due to its reliance on SSIM and colour-based cost functions, which fail to discriminate effectively in low-texture areas. In these regions, multiple candidates can yield similarly low cost function values, making it difficult to determine the correct correspondence. Although interesting work has been done to address various issues such as scale ambiguity [73], gradient locality [22, 35], occlusion [22, 75], moving objects [41, 60], and infinite depth holes [22, 24], the ill-posed nature of this problem has not been thoroughly and rigorously studied. We aim to provide this essential foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 491 + ], + "type": "text", + "content": "The key challenge lies in reducing ambiguity to build a more robust foundation for learning depth and ego-motion. To this end, we propose introducing sufficient variance in texture-less areas, while adhering as closely as possible to 'the constancy assumption': any change introduced in image " + }, + { + "bbox": [ + 55, + 251, + 295, + 491 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 251, + 295, + 491 + ], + "type": "text", + "content": " should be reproduced identically in image " + }, + { + "bbox": [ + 55, + 251, + 295, + 491 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 251, + 295, + 491 + ], + "type": "text", + "content": ", essential as the photometric loss relies on correspondences between successive frames. We propose to do this by assigning to each pixel the distance to its nearest edge in the image, commonly known as the 'distance transform' [54, 68]. We demonstrate mathematically that the distance transform is the most effective way of adding variance while respecting the constancy assumption, and that it also leads to improved convergence properties. Since the distance transform is calculated on binary contours, we propose a self-supervised pre-semantic contour estimation method, jointly learned with the depth estimation model, and add this distance map as an additional channel to the input image to produce a 'variance-augmented' image for reconstruction. In summary, our contributions are as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 295, + 702 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 534, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 582 + ], + "type": "text", + "content": "(1) A self-supervised approach for pre-semantic contour estimation, which (i) generates the edge information necessary for computing the distance transform, and (ii) simultaneously reinforces the depth estimation network." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 618 + ], + "type": "text", + "content": "(2) A novel variance-augmented image that incorporates the distance transform, enhancing discriminative power in texture-less areas and improving convergence properties" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "type": "text", + "content": "(3) A theoretical foundation demonstrating the effectiveness of the distance transform for adding variance under the constancy assumption" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 295, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 702 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 702 + ], + "type": "text", + "content": "(4) An extensive experimental validation of our new pipeline, achieving competitive performance on different benchmark datasets: KITTI [20], Cityscapes [9], Waymo [69], NYUv2 [67] and ScanNet [10]" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 403, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 403, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 403, + 83 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 553, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 462 + ], + "type": "text", + "content": "Self-supervised framework. The work of Zhou et al. [91] is pioneer in self-supervised monocular depth estimation. Their method utilizes depth and ego-motion to find the corresponding pixels between two successive frames and relies on the bilinear sampler [34] to reconstruct the upcoming scene from the current frame. The supervision signal is then derived from the photometric loss, which is a weighted sum of the SSIM and colour cost functions. The effectiveness of this method relies on the constancy assumption, which asserts that object colours remain consistent from frame to frame. This assumption is bound to the hypothesis of constant lighting over Lambertian surfaces, and do not hold in the case of occlusions and moving objects. This method has then been refined in several ways. Kim et al. [37] provide a comprehensive study of the main issues and solutions found in the literature. In [75], occlusions are addressed using a forward warping module that counts the number of bilinear sampling operations for each pixel. Godard et al. [22] compute the occlusion mask by determining the per-pixel minimum of the photometric loss over time. Moving objects are identified through the differences between optical flow and depth/ego-motion in [28, 40, 47, 60, 86, 93], while off-the-shelf semantic algorithms are utilized to localize potential moving objects in [4, 24]. In addition, a residual map is introduced to correct the ego-motion in [23, 41]. Infinite depth holes are filtered through a two-stage training process [24] and are removed using an adaptive cost-volume architecture in [77]. Finally, Godard et al. [22] propose a unified solution that effectively addresses all the aforementioned issues. However, the ill-posed nature of the problem due to unstructured objects remains unaddressed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "type": "text", + "content": "Contour estimation. Semantic contours should follow the boundaries derived from instance segmentation, outlining distinct entities within an image. Such contours would naturally help improve the accuracy of depth maps. [74] and [59] leverage ground-truth annotations to learn depth, normals, and contours, incorporating regularization terms to enforce constraints derived from their geometric relationships. In contrast, [43] employs an off-the-shelf edge prediction algorithm to refine initial depth estimations as a post-processing step. [79], [64] and [45] utilize a pretrained instance segmentation network to compute edges, sampling points on both sides of the boundary to enhance depth sharpness around those edges. [92] introduced a morphing strategy to align depth borders computed with a threshold gradient on the depth with semantic borders obtained from a pre-trained semantic algorithm. While these techniques demonstrate significant improvements, they assume the edges as provided and are therefore not fully self-supervised, in addition to not being GPU friendly. Significant advances have been made in supervised edge map estimation [11, 80]; however, the self-supervised approach re" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21869" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 334 + ], + "type": "text", + "content": "mains an ambitious and challenging task. [46] introduces a reinforcement training procedure in which an optical flow network is trained to predict the displacement of each pixel between two frames, based on ground truth point correspondences. Subsequently, an edge detector is updated to align its predicted edges of the image with the predicted edges of the estimated flow map. Still, this method is not self-supervised. A work closer to ours [84] proposes a fully self-supervised edge prediction framework that utilizes the positive part of the second derivatives of depth and the first-order gradient of normal to surface to feed an edge network. However, this approach employs suboptimal loss functions, leading to biased contours that favour shorter distances. Furthermore, it does not utilize the edges to enhance depth, nor does it include any post-processing to create refined contours. We propose new cost functions for training the edge network, resulting in significant performance gains. Additionally, we introduce a post-processing step to refine contours and implement an extended edge-aware diffusion smoothness loss [31] that encourages sharp depth discontinuities around estimated boundaries, in the same spirit as [45, 79, 92]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 335, + 296, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 708 + ], + "type": "text", + "content": "Targeting the ill-posed optimization problem. As previously mentioned, our objective is to minimize the photometric loss by establishing correspondences between pixels in the current frame and those in the next frame, achieved through the depth and ego-motion parameters. However, this correspondence procedure is not unique, since many pixels share the same colour and exhibit similar local colour statistics as their neighbourhood. This ambiguity is especially pronounced in areas with low texture. The challenge is to compensate for the lack of variation by introducing time-consistent variance across frames, to satisfy the constancy assumption. This reduces the ill-posed nature of the optimization problem and creates better conditions for learning depth and ego-motion. [81, 88] addressed this issue by applying the reconstruction loss to deep features in addition to the usual colour images. Shu et al. [66] took this approach further and designed custom deep features by training an auto-encoder that encourages large and smooth gradient values in feature maps, yielding deep features with both semantic richness and substantial structure. The constancy assumption is rather optimistic for these methods, as demonstrated in our experiments. Indeed, due to their large receptive fields, deep features generally lack consistency from frame to frame. Additionally, the reduced resolution of feature maps relative to the input image significantly limits the benefit of these methods in terms of accuracy. We show that the distance transform, not only proves to be experimentally more consistent across frames, but also stands as the theoretically best method for increasing variance under the constancy constraint, while preserving the resolution of the input image. Our work employs a similar methodol" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "ogy as [26], which theoretically prove the strong invariant properties induced by the census transform used in robust optical flow estimation [50] and validate these properties through experiments. [8] also uses the distance transform, but as an additional input to infer dense depth from sparse depth. However, using the distance transform of edge maps as a training signal is a novel approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 165, + 379, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 165, + 379, + 178 + ], + "spans": [ + { + "bbox": [ + 314, + 165, + 379, + 178 + ], + "type": "text", + "content": "3. Preamble" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 185, + 555, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 555, + 268 + ], + "type": "text", + "content": "Our goal in this section is to propose a mathematical formalism that closely aligns with experimental scenarios to address the following question: How can we introduce sufficient variance in low-texture areas while adhering as much as possible to the constancy assumption? The purpose is to provide a better basis than the usual framework [91] for learning depth and ego-motion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 269, + 555, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 317 + ], + "type": "text", + "content": "We start by introducing some key theoretical properties of the distance transform that we illustrate with toy experiments. The following definitions and theorems provide the theoretical foundation of our analysis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 323, + 507, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 507, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 507, + 335 + ], + "type": "text", + "content": "3.1. Maximal variance under constraints" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "text", + "content": "Notations Let " + }, + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "text", + "content": " be a compact convex set of " + }, + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "text", + "content": " with smooth boundary " + }, + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\partial \\Omega" + }, + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "text", + "content": " be the set of affine transformations. And let " + }, + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathcal{A}}" + }, + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "text", + "content": " the set of positive functions invariant by affine transformation (see Definition 2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "type": "text", + "content": " represents an object within a scene, while " + }, + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "type": "text", + "content": " models a reasonable subset of reprojected 3d transformations likely to affect the objects from frame to frame." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 430, + 554, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 454 + ], + "type": "text", + "content": "Definition 1 The distance transform " + }, + { + "bbox": [ + 313, + 430, + 554, + 454 + ], + "type": "inline_equation", + "content": "\\delta_{\\Omega}:\\Omega \\to \\mathbb{R}^{+}" + }, + { + "bbox": [ + 313, + 430, + 554, + 454 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 361, + 460, + 555, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 460, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 361, + 460, + 555, + 479 + ], + "type": "interline_equation", + "content": "\\delta_ {\\Omega} (x) = d (x, \\partial \\Omega) = \\inf _ {y \\in \\partial \\Omega} \\| x - y \\| \\tag {1}", + "image_path": "d72a0063f10b6198f33cdb3d6e04119117dc077b96c08fdb017f64d001c7a5a2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 484, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 543 + ], + "type": "text", + "content": "The set of points having more than one closest point to the boundary is called the medial axis. By Rademacher's Theorem [53] the distance transform is differentiable almost everywhere (away of the medial axis) and satisfies the Eikonal equation:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 412, + 544, + 553, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 544, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 412, + 544, + 553, + 556 + ], + "type": "interline_equation", + "content": "\\| \\nabla d \\| = 1 \\tag {2}", + "image_path": "ea25bd89a4d856bc9f350437f6499d349de7c597ad6bfea4fdb5fd0ad89828ef.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 561, + 553, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 553, + 586 + ], + "type": "text", + "content": "Definition 2 Let " + }, + { + "bbox": [ + 313, + 561, + 553, + 586 + ], + "type": "inline_equation", + "content": "f: \\Omega \\to \\mathbb{R}^+" + }, + { + "bbox": [ + 313, + 561, + 553, + 586 + ], + "type": "text", + "content": " a smooth function on " + }, + { + "bbox": [ + 313, + 561, + 553, + 586 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 561, + 553, + 586 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 313, + 561, + 553, + 586 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 561, + 553, + 586 + ], + "type": "text", + "content": " is said to be invariant by affine transformation if:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 591, + 553, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 591, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 361, + 591, + 553, + 604 + ], + "type": "interline_equation", + "content": "\\forall A \\in \\mathcal {A}, \\forall x \\in \\Omega , f (A (x)) = f (x) \\tag {3}", + "image_path": "f74b78b6b410d5ff4be9c93ce5185456910d0b18edea24b7c67c68dae1d76c47.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 609, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 554, + 658 + ], + "type": "text", + "content": "Definition 2 restates the constancy assumption mathematically: if an affine transformation " + }, + { + "bbox": [ + 313, + 609, + 554, + 658 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 609, + 554, + 658 + ], + "type": "text", + "content": " is applied to a convex shape " + }, + { + "bbox": [ + 313, + 609, + 554, + 658 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 609, + 554, + 658 + ], + "type": "text", + "content": ", then any changes introduce by " + }, + { + "bbox": [ + 313, + 609, + 554, + 658 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 609, + 554, + 658 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 609, + 554, + 658 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 609, + 554, + 658 + ], + "type": "text", + "content": " should manifest identically in " + }, + { + "bbox": [ + 313, + 609, + 554, + 658 + ], + "type": "inline_equation", + "content": "A(\\Omega)" + }, + { + "bbox": [ + 313, + 609, + 554, + 658 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 662, + 554, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 554, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 554, + 687 + ], + "type": "text", + "content": "Theorem 1 " + }, + { + "bbox": [ + 313, + 662, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\delta_{\\Omega}" + }, + { + "bbox": [ + 313, + 662, + 554, + 687 + ], + "type": "text", + "content": " is the unique solution, up to an isomorphism, to the following optimization problem:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 339, + 691, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 691, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 339, + 691, + 553, + 717 + ], + "type": "interline_equation", + "content": "\\max _ {f \\in \\mathcal {I} _ {\\mathcal {A}}} \\int_ {\\Omega} (f (x) - \\bar {f}) ^ {2} d x \\quad s. t. \\| \\nabla f \\| = 1 \\tag {4}", + "image_path": "5a80af0c733290ad068ab90da55e1806e9b579a5b9842342275db7efce0493b1.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21870" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 190 + ], + "type": "inline_equation", + "content": "\\overline{f}" + }, + { + "bbox": [ + 55, + 72, + 294, + 190 + ], + "type": "text", + "content": " being the average value of " + }, + { + "bbox": [ + 55, + 72, + 294, + 190 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 72, + 294, + 190 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 55, + 72, + 294, + 190 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 72, + 294, + 190 + ], + "type": "text", + "content": ". A proof of Theorem 1 is given in Appendix A.1. This property is particularly significant because it demonstrates that, under the constraint of the Eikonal equation, the optimal method for increasing variance within a compact convex shape while adhering to the constancy assumption is to use the distance transform. For this reason, in our practical experiments, we propose to introduce variance within image instances primarily via the distance transform; further details can be found in Section 4.2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 323 + ], + "type": "text", + "content": "The Eikonal equation constraint acts as a standardization condition, promoting smooth variation. Without this constraint, partitioning the shape in uniform black and white regions would also maximize the variance. Our goal is instead to ensure a more even spread of variance throughout the shape, making the variance-augmented image reprojection more informative for better depth and ego-motion learning. We also explore carefully relaxing this constraint to investigate well-chosen functions of the distance transform that could allow for greater variance within instances while still maintaining an acceptable level of smoothness:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 331, + 294, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 356 + ], + "type": "interline_equation", + "content": "\\mathcal {F} = \\left\\{f: x \\mapsto g \\circ \\delta_ {\\Omega} (x) \\mid g \\in \\mathcal {C} ^ {\\infty} \\left(\\mathbb {R} ^ {+}, \\mathbb {R} ^ {+}\\right), \\| g \\| _ {\\infty} \\leq 1 \\right\\} \\tag {5}", + "image_path": "1ef3db03cdfc34b8b38c46679f4787f1d25795995b7feba3afae54d94cb782bb.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 415 + ], + "type": "text", + "content": "A thorough comparison of depth estimation results across different functions is provided in Appendix C.2. The bound on the norm ensures that variance is not artificially amplified by larger values, maintaining consistency and fairness in comparisons." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 416, + 295, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 295, + 440 + ], + "type": "text", + "content": "We show that the maximal variance property can be retrieved with a simple toy experiment in A.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 447, + 189, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 189, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 189, + 459 + ], + "type": "text", + "content": "3.2. Convergence properties" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "content": "Notations Let " + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "inline_equation", + "content": "(I, \\tilde{I})" + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "content": " be two successive images of a camera in movement and " + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "content": " a training instance where " + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "content": " represents a pixel of " + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "content": " be the corresponding pixel in " + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "inline_equation", + "content": "\\tilde{I}" + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "content": ". Let us assume also that all pixels lie inside a convex shape " + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "content": " defined by a smooth and closed contour " + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "inline_equation", + "content": "\\partial \\Omega" + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "content": ". Our goal is to learn the model " + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "inline_equation", + "content": "\\Phi(x; \\theta)" + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "content": " to predict the new pixel location " + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "content": ". Let us define the cost function " + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "content": " using functions from " + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 464, + 296, + 561 + ], + "type": "text", + "content": " (Equation 5):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 568, + 294, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 568, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 116, + 568, + 294, + 587 + ], + "type": "interline_equation", + "content": "l: (\\hat {y}, y) \\mapsto (f (\\hat {y}) - f (y)) ^ {2} \\tag {6}", + "image_path": "60ba3a4a9f74156bda2b1a35fc9714d2901c45a73f0e656960b3387706fb094e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 586, + 181, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 586, + 181, + 597 + ], + "spans": [ + { + "bbox": [ + 125, + 586, + 181, + 597 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 125, + 586, + 181, + 597 + ], + "type": "inline_equation", + "content": "f\\in \\mathcal{F}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 606, + 275, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 275, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 275, + 619 + ], + "type": "text", + "content": "Definition 3 A function " + }, + { + "bbox": [ + 55, + 606, + 275, + 619 + ], + "type": "inline_equation", + "content": "f: \\Omega \\to \\mathbb{R}^+" + }, + { + "bbox": [ + 55, + 606, + 275, + 619 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 606, + 275, + 619 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 606, + 275, + 619 + ], + "type": "text", + "content": "-Lipschitz if:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 627, + 294, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 627, + 294, + 640 + ], + "spans": [ + { + "bbox": [ + 102, + 627, + 294, + 640 + ], + "type": "interline_equation", + "content": "\\forall u, v \\quad | f (u) - f (v) | \\leq \\alpha \\| u - v \\| \\tag {7}", + "image_path": "834eb90359f4fcd7d1ebf99865806589d955f3f96378d060391b04ea5a2f65f3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 647, + 268, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 268, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 268, + 661 + ], + "type": "text", + "content": "Definition 4 A function " + }, + { + "bbox": [ + 55, + 647, + 268, + 661 + ], + "type": "inline_equation", + "content": "f: \\Omega \\to \\mathbb{R}^+" + }, + { + "bbox": [ + 55, + 647, + 268, + 661 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 647, + 268, + 661 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 647, + 268, + 661 + ], + "type": "text", + "content": "-smooth if:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 93, + 668, + 294, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 668, + 294, + 681 + ], + "spans": [ + { + "bbox": [ + 93, + 668, + 294, + 681 + ], + "type": "interline_equation", + "content": "\\forall u, v \\quad \\| \\nabla f (u) - \\nabla f (v) \\| \\leq \\beta \\| u - v \\| \\tag {8}", + "image_path": "4a3874c85fca46a87cf1f190997eb3980d383abd3a1f9f4d41f94ec10cdcc605.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "The fact that " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " is a function of the distance transform makes it possible to prove:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "Theorem 2 The loss function " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " is, with respect to the first argument " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 99, + 547, + 133 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 314, + 99, + 372, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 99, + 372, + 110 + ], + "spans": [ + { + "bbox": [ + 314, + 99, + 372, + 110 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 99, + 372, + 110 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 314, + 99, + 372, + 110 + ], + "type": "text", + "content": "-Lipschitz" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 110, + 365, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 110, + 365, + 121 + ], + "spans": [ + { + "bbox": [ + 315, + 110, + 365, + 121 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 315, + 110, + 365, + 121 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 315, + 110, + 365, + 121 + ], + "type": "text", + "content": "-smooth" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 121, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 547, + 133 + ], + "type": "text", + "content": "- strongly convex if a regularization term " + }, + { + "bbox": [ + 314, + 121, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\eta \\| y\\|^2" + }, + { + "bbox": [ + 314, + 121, + 547, + 133 + ], + "type": "text", + "content": " is added." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 147, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 555, + 277 + ], + "type": "text", + "content": "The constants " + }, + { + "bbox": [ + 313, + 147, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta, \\eta" + }, + { + "bbox": [ + 313, + 147, + 555, + 277 + ], + "type": "text", + "content": " depend on the curvature at the projection onto the boundary " + }, + { + "bbox": [ + 313, + 147, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\partial \\Omega" + }, + { + "bbox": [ + 313, + 147, + 555, + 277 + ], + "type": "text", + "content": ". A proof of Theorem 2 is given in Appendix A.2. Those properties brought by the distance transform induce great convergence properties as described in [1, 27, 36]. Amongst others: a bounded generalization error and a uniform stability of the stochastic gradient descent. Note that the SSIM + colour function can also be Lipschitz and/or uniform under very strict conditions. However their constants, unlike with the distance transform, depend on the size of the shape " + }, + { + "bbox": [ + 313, + 147, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 147, + 555, + 277 + ], + "type": "text", + "content": " and can thus exceeds all bounds as mentioned in [57]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 278, + 554, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 554, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 554, + 302 + ], + "type": "text", + "content": "We show that the convergence property can be retrieved with a simple toy experiment in A.4." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 318, + 370, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 318, + 370, + 330 + ], + "spans": [ + { + "bbox": [ + 314, + 318, + 370, + 330 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "text", + "content": "Our framework for self-supervised MDE (Figure 2) consists of three networks: Depth " + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "text", + "content": ", Ego-motion " + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "inline_equation", + "content": "T_{\\alpha}" + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "text", + "content": " and edge " + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "inline_equation", + "content": "E_{\\delta}" + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "text", + "content": ". For the sake of readability, we will omit the parameter symbols " + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\theta, \\alpha, \\delta" + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "text", + "content": " in the remainder of the section. Hyperparameters in the loss functions will also be omitted here and detailed in Appendix F." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 424, + 441, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 424, + 441, + 435 + ], + "spans": [ + { + "bbox": [ + 314, + 424, + 441, + 435 + ], + "type": "text", + "content": "4.1. Pre-semantic contours" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 444, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 554, + 539 + ], + "type": "text", + "content": "Our edge network renders values " + }, + { + "bbox": [ + 313, + 444, + 554, + 539 + ], + "type": "inline_equation", + "content": "E(p)" + }, + { + "bbox": [ + 313, + 444, + 554, + 539 + ], + "type": "text", + "content": " in the range [0, 1], where a higher value indicates a greater likelihood that the pixel " + }, + { + "bbox": [ + 313, + 444, + 554, + 539 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 444, + 554, + 539 + ], + "type": "text", + "content": " corresponds to an edge. To generate pseudo-labels for supervising the edge network, we utilize depth and surface normal estimation, which is derived from depth following the methods described in [83, 85]. We discuss complementary insights of depth and surface normal for edge detection in Appendix D.1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 540, + 554, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 554, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 554, + 587 + ], + "type": "text", + "content": "Depth & Normal Pseudo-Labels: Depth and surface normals are smooth and thus prone to exhibiting zero-crossings of their Laplacian " + }, + { + "bbox": [ + 313, + 540, + 554, + 587 + ], + "type": "inline_equation", + "content": "\\Delta D" + }, + { + "bbox": [ + 313, + 540, + 554, + 587 + ], + "type": "text", + "content": " (resp. " + }, + { + "bbox": [ + 313, + 540, + 554, + 587 + ], + "type": "inline_equation", + "content": "\\Delta N" + }, + { + "bbox": [ + 313, + 540, + 554, + 587 + ], + "type": "text", + "content": ") around semantic boundaries, defined as follows:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 339, + 599, + 553, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 599, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 339, + 599, + 553, + 643 + ], + "type": "interline_equation", + "content": "Z _ {D} (p) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\Delta D (p) \\Delta D (p + (1, 0)) < 0 \\\\ 1 & \\text {i f} \\Delta D (p) \\Delta D (p + (0, 1)) < 0 \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {9}", + "image_path": "93501498847ca570cc5c7c05326c5aba7a227f08fbab5068ab920b504aca2c9a.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Same applies to normal Laplacian " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\Delta N" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": ". The zero-crossing mask is then dilated (" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " structuring element), and multiplied by the magnitude of the gradient to create the pseudolabels. For this, the depth gradient is normalized to avoid" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21871" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 306, + 208 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 306, + 208 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 306, + 208 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 306, + 208 + ], + "type": "image", + "image_path": "afbebc2592a2f5d7e8c42e35b523cfefcb94d3c337e74fa2f1c6e822c77b86af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 218, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 218, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 555, + 243 + ], + "type": "text", + "content": "Figure 2. Diagram showing the two main steps of our pipeline: estimation of the different modalities followed by the re-projection on the variance augmented image using depth and ego-motion. T indicates the time frame. R, t respectively rotation and translation predictions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 553, + 208 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 553, + 208 + ], + "type": "image", + "image_path": "e24a5a4f4751674e386e8ed84e3f6d5ff00f23ea0f1ec7ae35f22cef70d5e4ec.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 262, + 187, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 187, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 187, + 274 + ], + "type": "text", + "content": "bias towards large depth values:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 281, + 294, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 281, + 294, + 310 + ], + "spans": [ + { + "bbox": [ + 85, + 281, + 294, + 310 + ], + "type": "interline_equation", + "content": "\\nabla D _ {x} (p) = \\frac {D (p + (0 , 1)) - D (p)}{\\max (D (p + (0 , 1)) , D (p))}, \\tag {10}", + "image_path": "27d1c4f75ff9cce10351380cfe7c9a88c27e1644ce89dea73e9ca668a3225ca9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 312, + 233, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 312, + 233, + 326 + ], + "spans": [ + { + "bbox": [ + 82, + 312, + 233, + 326 + ], + "type": "interline_equation", + "content": "\\| \\nabla D (p) \\| = | \\nabla D _ {x} (p) | + | \\nabla D _ {y} (p) |.", + "image_path": "6508762de53722d800ce94a6b0db1f038c6585e40db8d5e5bf9ea6de8239bd83.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 334, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 295, + 357 + ], + "type": "text", + "content": "The usual normal gradient, which is often too noisy, is replaced with a dot-product measure:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 366, + 295, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 366, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 80, + 366, + 295, + 386 + ], + "type": "interline_equation", + "content": "\\left| \\nabla N _ {x} (p) \\right| = 1 - \\left| \\langle \\vec {N} (p + (0, 1)), \\vec {N} (p) \\rangle \\right|, \\tag {11}", + "image_path": "39ecc80046536772bc48e07ca41f46c56029f20ca6d944e15b169ce9a23d022a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 384, + 233, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 384, + 233, + 398 + ], + "spans": [ + { + "bbox": [ + 82, + 384, + 233, + 398 + ], + "type": "interline_equation", + "content": "\\left\\| \\nabla N (p) \\right\\| = \\left| \\nabla N _ {x} (p) \\right| + \\left| \\nabla N _ {y} (p) \\right|.", + "image_path": "b42839fabef7b8e95d5ff0ba3234a19f14dbb443a338dea799e3b2cd8988e9cf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 406, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 295, + 430 + ], + "type": "text", + "content": "The weighted depth and normal pseudo-labels employed to supervise the edge network are then defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 438, + 295, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 438, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 295, + 469 + ], + "type": "interline_equation", + "content": "w _ {d} (p) = \\frac {Z _ {D} (p) \\| \\nabla D (p) \\|}{\\sum_ {p} Z _ {D} (p) \\| \\nabla D (p) \\|} \\tag {12}", + "image_path": "232bf109988c71b11afc23e6783936df8ba98560b986f14ef37b9ef88c2fa9f0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 476, + 295, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 476, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 295, + 508 + ], + "type": "interline_equation", + "content": "w _ {n} (p) = \\frac {Z _ {N} (p) \\| \\nabla N (p) \\|}{\\sum_ {p} Z _ {N} (p) \\| \\nabla N (p) \\|} \\tag {13}", + "image_path": "0f0f87c30ddc0b557cb9c6015634df14aa05cc3b116af3ed241d9ee1560fc2c2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 512, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 571 + ], + "type": "text", + "content": "A contrastive loss is also applied to encourage edge values to move toward either zero or one. This is achieved using a binary cross-entropy loss, where pseudo ground truth " + }, + { + "bbox": [ + 55, + 512, + 296, + 571 + ], + "type": "inline_equation", + "content": "Y(p) \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 512, + 296, + 571 + ], + "type": "text", + "content": " are created by thresholding edge estimations " + }, + { + "bbox": [ + 55, + 512, + 296, + 571 + ], + "type": "inline_equation", + "content": "E(p)" + }, + { + "bbox": [ + 55, + 512, + 296, + 571 + ], + "type": "text", + "content": " at 0.5:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 580, + 308, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 308, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 308, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = - \\sum_ {p} \\left(Y (p) \\log (E (p)) + (1 - Y (p)) \\log (1 - E (p))\\right) \\tag {14}", + "image_path": "0f1e3a93b621d06dfc7de3b6294112823c1d1b86670d5c1ab05b2cb17092e1d9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 618, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 641 + ], + "type": "text", + "content": "Finally, the loss function employed to supervise the edge network is defined as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 649, + 284, + 674 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 649, + 284, + 674 + ], + "spans": [ + { + "bbox": [ + 67, + 649, + 284, + 674 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {e d g e}} = \\sum_ {p} \\left(w _ {d} (p) + w _ {n} (p)\\right) \\left(1 - E (p)\\right) + \\mathcal {L} _ {c} + \\mathcal {L} _ {e}", + "image_path": "6575a41baad5052f964bec953745556ce307c66db49c2d18b70062d11ac17756.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 87, + 677, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 677, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 87, + 677, + 294, + 713 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 87, + 677, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e = \\sum_p E(p)^2" + }, + { + "bbox": [ + 87, + 677, + 294, + 713 + ], + "type": "text", + "content": " to avoid trivial solutions. (15)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 262, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 299 + ], + "type": "text", + "content": "Discontinuity Preserving Smoothing: Incorporating an edge-aware smoothness loss is a widely used practice in training depth networks [73, 83]:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 316, + 553, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 316, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 319, + 316, + 553, + 354 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\sum_ {p} \\left(| \\nabla D _ {x} (p) | e ^ {- | \\nabla I _ {x} (p) |} + | \\nabla D _ {y} (p) | e ^ {- | \\nabla I _ {y} (p) |}\\right) \\tag {16}", + "image_path": "d5583a594b68e5ed3561b02867c07299886f9107d2bca00f97066a644d128d54.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 354, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 425 + ], + "type": "text", + "content": "However, this approach has two drawbacks: first, boundaries are estimated using the image gradient, which is not optimal; second, there is no mechanism to effectively manage discontinuities. To address these issues, we propose a new smoothness loss exploiting our pre-semantic contours defined as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 324, + 443, + 447, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 443, + 447, + 470 + ], + "spans": [ + { + "bbox": [ + 324, + 443, + 447, + 470 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s _ {1}} = \\sum_ {p} \\| \\nabla D (p) \\| e ^ {- \\alpha E (p)},", + "image_path": "0f0802fd2c58f65d2e541c207e216516cec5dfd982ed9e038f8e084a305a1c0e.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 324, + 472, + 553, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 472, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 324, + 472, + 553, + 498 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s _ {2}} = \\sum_ {p} \\log \\left(1 + e ^ {- \\| \\nabla D (p) \\|}\\right) \\left(1 - e ^ {- \\alpha E (p)}\\right), \\tag {17}", + "image_path": "c421c6f0464c8833610bfa5e3ce1b3eb2d992fbcab7206d5db3b91718d7a6550.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 328, + 502, + 401, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 502, + 401, + 515 + ], + "spans": [ + { + "bbox": [ + 328, + 502, + 401, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\mathcal {L} _ {s _ {1}} + \\mathcal {L} _ {s _ {2}}.", + "image_path": "debb9a0c95348e30a45c378d83b98482a0ddd6effa861f37077218d98765562a.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 534, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 617 + ], + "type": "text", + "content": "The component " + }, + { + "bbox": [ + 313, + 534, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{s_1}" + }, + { + "bbox": [ + 313, + 534, + 555, + 617 + ], + "type": "text", + "content": " replace the former equation 16 to encourage depth smoothness inside semantic instances, while " + }, + { + "bbox": [ + 313, + 534, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{s_2}" + }, + { + "bbox": [ + 313, + 534, + 555, + 617 + ], + "type": "text", + "content": " prevents diffusion and promotes significant depth changes around boundaries. In this manner, both the edge and depth networks benefit from each other: the depth estimation supervises the edge network, while the edge network enhances depth estimation through the smoothness loss." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "Post-processing: The map produced by the edge network consists of continuous values ranging from zero to one. To obtain thin pre-semantic contours " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}}(I)" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": ", we apply a post-processing pipeline that includes hysteresis thresholding, non-maximum suppression, morphological closing, contour filtering, and contour closing. More details can be found in Appendix D.3 and Appendix D.2 discusses the advantages of our approach compared to the competitive work [84]." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21872" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 85 + ], + "type": "text", + "content": "4.2. A variance augmented image for depth learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 128 + ], + "type": "text", + "content": "With " + }, + { + "bbox": [ + 55, + 90, + 296, + 128 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{C}}(I)" + }, + { + "bbox": [ + 55, + 90, + 296, + 128 + ], + "type": "text", + "content": " the thin pre-semantic contour of " + }, + { + "bbox": [ + 55, + 90, + 296, + 128 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 90, + 296, + 128 + ], + "type": "text", + "content": " obtained after edge network estimation and subsequent post-processing, we then compute the distance transform map " + }, + { + "bbox": [ + 55, + 90, + 296, + 128 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{I}}" + }, + { + "bbox": [ + 55, + 90, + 296, + 128 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 124, + 135, + 296, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 135, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 124, + 135, + 296, + 156 + ], + "type": "interline_equation", + "content": "\\delta_ {\\mathrm {I}} (p) = \\min _ {x \\in \\hat {\\mathcal {C}} (I)} (\\| x - p \\|) \\tag {18}", + "image_path": "9c787ea1e960e5a19e1cb18887c405294cc2f45c597576a7105c2fa8c01c52dc.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 163, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 294 + ], + "type": "text", + "content": "See Appendix C.1 for the details of the algorithm. The resulting distance transform map is concatenated with the input image to create a 'variance-augmented' image. The objective is then to learn depth and ego-motion in order to identify corresponding pixels across successive frames, and reconstruct the upcoming structure-augmented image in the most effective manner, as illustrated on Figure 2. The reconstruction of the original input image " + }, + { + "bbox": [ + 55, + 163, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{l}}" + }, + { + "bbox": [ + 55, + 163, + 296, + 294 + ], + "type": "text", + "content": " is evaluated using " + }, + { + "bbox": [ + 55, + 163, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 163, + 296, + 294 + ], + "type": "text", + "content": ", the standard combination of colour difference and SSIM, while the reconstruction of the distance transform map is assessed with a " + }, + { + "bbox": [ + 55, + 163, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 55, + 163, + 296, + 294 + ], + "type": "text", + "content": " loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 302, + 294, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 302, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 127, + 302, + 294, + 328 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p h o t o}} = \\Psi (\\widehat {I}, I), \\text {a n d} \\tag {19}", + "image_path": "6a0aa8a5a2c3e12b3ee252eda6e2c6a95058c86e0bb996ee4493717299648360.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 325, + 208, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 325, + 208, + 346 + ], + "spans": [ + { + "bbox": [ + 135, + 325, + 208, + 346 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d i s t}} = \\left\\| \\widehat {\\delta_ {\\mathrm {I}}} - \\delta_ {\\mathrm {I}} \\right\\| _ {1}", + "image_path": "a531d0a7c605c8f6bbfa4e5b03dfa9616fef5d5267670fa07dff702db7313acc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 349, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 373 + ], + "type": "text", + "content": "The smoothness loss is added to get the final supervision signal of the depth network:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 117, + 380, + 295, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 380, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 117, + 380, + 295, + 395 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {d e p t h}} = \\mathcal {L} _ {\\text {d i s t}} + \\mathcal {L} _ {\\text {p h o t o}} + \\mathcal {L} _ {s} \\tag {20}", + "image_path": "5e31e254e28ea96a8ef1325cf824227abf2b89d022078ddc1381ee5a2fdf167e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 402, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 439 + ], + "type": "text", + "content": "An improved distance transform: This approach can be readily extended to the family of one-dimensional functions " + }, + { + "bbox": [ + 55, + 402, + 296, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{p\\mapsto g\\circ \\delta_{\\mathrm{I}}(p)\\}" + }, + { + "bbox": [ + 55, + 402, + 296, + 439 + ], + "type": "text", + "content": " (as defined in Equation 5):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 445, + 295, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 445, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 116, + 445, + 295, + 466 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t}} = \\left\\| g \\circ \\widehat {\\delta_ {\\mathrm {I}}} - g \\circ \\delta_ {\\mathrm {I}} \\right\\| _ {1}. \\tag {21}", + "image_path": "633cc42899c8a8734351a973807c54d93724962846ef3b3d5973b66a4dadb849.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "type": "text", + "content": "We provide an analysis of the depth estimation results for various functions in Appendix C.2. To take this a step further, we propose extending the distance transform into the " + }, + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "type": "text", + "content": "-dimensional space by mapping the distance transform values to an " + }, + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "type": "text", + "content": "-dimensional random walk process denoted as " + }, + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathrm{RW}_n" + }, + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 73, + 547, + 295, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 547, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 73, + 547, + 295, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t}} = \\sum_ {p} \\left\\| \\mathrm {R W} _ {n} \\circ \\widehat {\\delta_ {\\mathrm {I}}} (p) - \\mathrm {R W} _ {n} \\circ \\delta_ {\\mathrm {I}} (p) \\right\\| _ {1} \\tag {22}", + "image_path": "da7c7f8b96de423dfd793ece76f55492df55f073fd4a248561213443a3545093.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Using this approach, we can maintain the validity of the constancy assumption while increasing the variance even more. Additionally, the dimensional parameter " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " determines the significance of the distance transform map reconstruction within the total loss function. It introduces a tradeoff between " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{photo}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ": the higher the dimension, the greater the emphasis on the artificially introduced structure compared to the natural structure found in the image; see Appendix C.3 for more details on the random walk mapping. Our experiments in Appendix C.2 show that " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " provides the best results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 91, + 377, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 377, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 377, + 102 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 110, + 500, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 500, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 500, + 121 + ], + "type": "text", + "content": "We conduct experiments on multiple datasets:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 121, + 553, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 193 + ], + "type": "text", + "content": "KITTI [20] includes diverse urban, rural, and highway scenes. We follow the standard data split defined by [13] and apply the pre-processing steps outlined in [91] to filter static frames, resulting in a training set of 39810 images and 697 images for testing. Results are reported using the evaluation protocol from [91]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 194, + 554, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 554, + 252 + ], + "type": "text", + "content": "Cityscapes [9] offers urban scenes with dynamic objects, challenging depth estimation. The training set consists of 22973 images, with evaluation on 1525 test images following [5, 41]. Edge detection is assessed using the contour evaluation from [84] on 500 validation frames." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 253, + 554, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 554, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 554, + 325 + ], + "type": "text", + "content": "Waymo [69] is a large and diverse autonomous driving dataset that captures dynamic urban scenes under various environmental conditions, including nighttime and different weather scenarios. We sample 100 000 image pairs from 1 000 front camera video sequences for training and evaluate on 1500 pairs from 150 sequences, following [41]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 325, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 553, + 384 + ], + "type": "text", + "content": "NYUv2 [67] consists of 464 indoor video sequences captured with a Kinect sensor. We use the official splits (302 training, 33 validation) and evaluate on 654 densely annotated test images, resizing inputs to " + }, + { + "bbox": [ + 313, + 325, + 553, + 384 + ], + "type": "inline_equation", + "content": "320 \\times 256" + }, + { + "bbox": [ + 313, + 325, + 553, + 384 + ], + "type": "text", + "content": " pixels as in [44]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 384, + 554, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 433 + ], + "type": "text", + "content": "ScanNet [10] is a large-scale RGB-D dataset with 2.5 million images across " + }, + { + "bbox": [ + 313, + 384, + 554, + 433 + ], + "type": "inline_equation", + "content": "1500+" + }, + { + "bbox": [ + 313, + 384, + 554, + 433 + ], + "type": "text", + "content": " scenes. We use it for zero-shot evaluation, following [16, 44], with test images resized to " + }, + { + "bbox": [ + 313, + 384, + 554, + 433 + ], + "type": "inline_equation", + "content": "320 \\times 256" + }, + { + "bbox": [ + 313, + 384, + 554, + 433 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 441, + 447, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 447, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 447, + 453 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "text", + "content": "We emphasize that our method is compatible with any DNN architecture and input image size. For fair comparison, our experiments are done with a " + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "inline_equation", + "content": "256 \\times 832" + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "text", + "content": " resolution and a UNet [62] structure, following [91], unless stated otherwise. The depth and edge networks share a ResNet [29] encoder. Our baseline, CoopNet [28], handles moving objects—especially useful in Cityscapes—via self-supervised flow. The pose network, based on a ResNet encoder, outputs a 6-DoF vector. All encoders use ResNet-50 backbones initialized with ImageNet [63] weights. Training runs for 30 epochs with batch size 4, initial learning rate of " + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "text", + "content": " (decreasing to " + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "text", + "content": " after 20 epochs), and standard data augmentation from [22]. Loss weights, tuned via grid search, are detailed in Appendix F. Our PyTorch [58] implementation trains on a single NVIDIA RTX A5000 GPU using Adam [38] with " + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.99" + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "text", + "content": " and takes 12 hours." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 670, + 372, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 670, + 372, + 682 + ], + "spans": [ + { + "bbox": [ + 314, + 670, + 372, + 682 + ], + "type": "text", + "content": "5.3. Results" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Depth Quantitative results on KITTI and Cityscapes are presented in Table 1, where we compare against the most" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21873" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 293 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 293 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 293 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 293 + ], + "type": "table", + "html": "
MethodLower is better ↓Higher is better ↑
Abs RelSq RelRMSERMSE logδ1δ2δ3
Monodepth2[22]0.1100.8314.6420.1870.8830.9620.982
Guizilini et al.†[24] (R50)0.1130.8314.6630.1890.8780.9710.983
Shu et al.*[66](R50)0.1080.7924.6330.1840.8830.9610.983
Ours (R50)0.1040.7254.4530.1800.8850.9620.983
MonoViT[89]0.0990.7084.3720.1750.9000.9670.984
Ours + MonoViT0.0920.6744.3000.1650.9270.9670.984
HR-Depth[49]0.1090.7924.6320.1850.8840.9620.983
RA-Depth[30]0.0960.6134.2160.1710.9030.9680.985
DIFFNet[90]0.1020.7644.4830.1800.8960.9650.983
Guizilini et al.†[24]0.1000.7614.2700.1750.9020.9650.982
Ours+DinoV20.0820.6044.1080.1620.9280.9680.985
Lego (paper)0.1541.2726.0120.230N/AN/AN/A
Lego★ (R50)0.1150.8554.7890.1950.8760.9570.979
Lego★ (R50) + Ldist0.1100.8054.6060.1880.8820.9600.982
Ours (R50)0.1040.7254.4530.1800.8850.9620.983
Struct2Depth†[4]0.1451.7377.280.2050.8130.9420.978
LearnK†[41]0.1271.3306.960.1950.8300.9470.981
Li et al.†[41]0.1191.2906.980.1900.8460.9510.980
CoopNet[28]0.1211.4437.010.1900.8460.9510.980
Ours (R50)0.1151.2216.790.1860.8500.9550.981
", + "image_path": "ee3e9691c6c1eaed2d2aea54976e283a353939db56bc27f4cd15ae3c2040d484.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 426, + 294, + 499 + ], + "blocks": [ + { + "bbox": [ + 55, + 301, + 295, + 413 + ], + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 413 + ], + "type": "text", + "content": "Table 1. Results of depth estimations on KITTI 2015 (first four blocks) and Cityscapes (last block). For each metric the best result is displayed in bold and the second-best is underlined. Light blue metrics indicates the most challenging metrics. " + }, + { + "bbox": [ + 55, + 301, + 295, + 413 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 301, + 295, + 413 + ], + "type": "text", + "content": " denotes the use of an off-the-shelf semantic algorithm and * indicates results reproduced by us at the input resolution of " + }, + { + "bbox": [ + 55, + 301, + 295, + 413 + ], + "type": "inline_equation", + "content": "256 \\times 832" + }, + { + "bbox": [ + 55, + 301, + 295, + 413 + ], + "type": "text", + "content": ". Block#1 and #2 compares methods using similar encoder backbones: Resnet 50 (R50) in block#1 and ViT in block#2. Block#3 compares SOTA depth models without restrictions on architecture or additional inputs, Block#4 compares to Lego." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 426, + 294, + 499 + ], + "lines": [ + { + "bbox": [ + 58, + 426, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 58, + 426, + 294, + 499 + ], + "type": "table", + "html": "
MethodDataset (Train / Test)Lower is better ↓
Abs RelSq RelRMSERMSE log
LearnK[41]Waymo / Waymo0.1571.5317.0900.205
OursWaymo / Waymo0.1251.2785.9900.189
GLNet[7]Cityscapes / KITTI0.1291.0445.3610.212
CoopNet[28]Cityscapes / KITTI0.1251.1575.2510.209
OursCityscapes / KITTI0.1160.8924.8920.195
", + "image_path": "5694788ab0dac38ddbd7a5416afb2ab5d2f5d9ce0973763c64bea1172dfe0e87.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 544, + 294, + 609 + ], + "blocks": [ + { + "bbox": [ + 55, + 509, + 295, + 531 + ], + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 531 + ], + "type": "text", + "content": "Table 2. Results on Outdoor datasets. Block#1: in-domain evaluation. Block#2: out-of-domain evaluation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 544, + 294, + 609 + ], + "lines": [ + { + "bbox": [ + 58, + 544, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 58, + 544, + 294, + 609 + ], + "type": "table", + "html": "
MethodDataset (Train / Test)Abs RelRMSEδ1δ2δ3
MonoIndoor++[44]NYUv2 / NYUv20.1320.5170.8340.9610.990
IndoorDepth[16]NYUv2 / NYUv20.1260.4940.8450.9650.991
OursNYUv2 / NYUv20.1150.4580.8590.9700.992
IndoorDepth[16]NYUv2 / ScanNet0.1530.3730.7860.9500.988
MonoIndoor++[44]NYUv2 / ScanNet0.1380.3470.8100.9670.993
OursNYUv2 / ScanNet0.1270.3120.8430.9700.993
", + "image_path": "c013e19dfec251c4461a02c1f12b23cbc40a8e66bb86a58bff9d06924145a7b0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "competitive methods. Blocks #1 and #2 demonstrate that, when using the same encoder backbone, our method outperforms others by a significant margin across nearly all metrics. When no architectural constraints are imposed, we" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 322, + 70, + 553, + 131 + ], + "blocks": [ + { + "bbox": [ + 55, + 619, + 295, + 641 + ], + "lines": [ + { + "bbox": [ + 55, + 619, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 295, + 641 + ], + "type": "text", + "content": "Table 3. Results on Indoor datasets. Block#1: in-domain evaluation. Block#2: out-of-domain evaluation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 70, + 553, + 131 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 553, + 131 + ], + "type": "table", + "html": "
RW3LdistLsEθAbs Rel (↓)RMSE log (↓)δ1(↑)
0.1130.190.878
✓0.1120.1890.880
✓✓0.1100.1870.881
✓✓✓0.1060.1830.884
✓✓✓✓0.1040.1800.885
", + "image_path": "be2a428613e52a3a8bd9d654060a77a94c7230759aee89c34352cd7e15ff02db.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 138, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 138, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 555, + 194 + ], + "type": "text", + "content": "Table 4. Ablation study on KITTI 2015 of losses mentioned in Section 4.2. Note that " + }, + { + "bbox": [ + 313, + 138, + 555, + 194 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 313, + 138, + 555, + 194 + ], + "type": "text", + "content": " is part of the ablation as it does improve the depth metrics and shares the same encoder as the depth network. " + }, + { + "bbox": [ + 313, + 138, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\mathrm{RW}_3" + }, + { + "bbox": [ + 313, + 138, + 555, + 194 + ], + "type": "text", + "content": " is for 3d random walk encoding of distance transform as defined in Section 4.2." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 312, + 215, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 215, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 312, + 215, + 555, + 323 + ], + "type": "text", + "content": "employ a pre-trained DinoV2 [56] encoder (further details in Appendix F.2), yielding the results in block #3. For fairness, we compare against " + }, + { + "bbox": [ + 312, + 215, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathsf{LEGO}^{\\star}" + }, + { + "bbox": [ + 312, + 215, + 555, + 323 + ], + "type": "text", + "content": " in block #4, an enhanced version of LEGO that incorporates a ResNet-50 backbone and up-to-date training techniques. Adding our novel distance transform loss " + }, + { + "bbox": [ + 312, + 215, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 312, + 215, + 555, + 323 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 312, + 215, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathsf{LEGO}^{\\star}" + }, + { + "bbox": [ + 312, + 215, + 555, + 323 + ], + "type": "text", + "content": " greatly improves results, showing once again the great importance of this loss component, but still underperforms ours, likely due to the superior quality of our contours." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 323, + 556, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 323, + 556, + 491 + ], + "spans": [ + { + "bbox": [ + 312, + 323, + 556, + 491 + ], + "type": "text", + "content": "We also provide strong out-of-domain results in Table 2 and Table 3, demonstrating competitive performance on both indoor and outdoor datasets. Finally, an ablation study in Table 4 evaluates the impact of the different loss components introduced in our approach and in Appendix C.2 for different functions from the family " + }, + { + "bbox": [ + 312, + 323, + 556, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 312, + 323, + 556, + 491 + ], + "type": "text", + "content": " as defined in Section 3.1. It can be seen that the reprojection on the distance transform via " + }, + { + "bbox": [ + 312, + 323, + 556, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 312, + 323, + 556, + 491 + ], + "type": "text", + "content": " brings a decisive improvement. We provide a qualitative comparison of results in Figure 3. We can see that our method gives better results: it provides thinner object, is sharper around moving objects, does not suffer from smoothness issues for large objects and is not noisy in textured areas. More qualitative and quantitative results can be found in Appendix B.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 502, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 575 + ], + "type": "text", + "content": "Flow & Odometry As discussed in Section 5.2 we use an optical flow to handle moving objects following [28]. Its results can be found in Appendix B.2. Results for odometry can be found in Appendix B.3. Our method significantly outperforms the baseline [28] on both optical flow and odometry metrics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 429, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 429, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 429, + 594 + ], + "type": "text", + "content": "5.4. Contour evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 599, + 556, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 599, + 556, + 708 + ], + "spans": [ + { + "bbox": [ + 312, + 599, + 556, + 708 + ], + "type": "text", + "content": "We evaluate contour detection in Table 5 using the same metrics and data as [84]: Optimal Dataset Scale, Optimal Image Scale, and Average Precision, computed on 500 Cityscapes validation images. Our method significantly outperforms LEGO, leveraging (1) depth & normal pseudolabels with Laplacian zero-crossing insights for better contour alignment, (2) the discontinuity-preserving smoothing loss " + }, + { + "bbox": [ + 312, + 599, + 556, + 708 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 312, + 599, + 556, + 708 + ], + "type": "text", + "content": ", and (3) the contrastive loss " + }, + { + "bbox": [ + 312, + 599, + 556, + 708 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 312, + 599, + 556, + 708 + ], + "type": "text", + "content": ". Qualitative examples are in Figure 1, with more in Appendix D." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21874" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 79, + 555, + 318 + ], + "blocks": [ + { + "bbox": [ + 55, + 79, + 555, + 318 + ], + "lines": [ + { + "bbox": [ + 55, + 79, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 79, + 555, + 318 + ], + "type": "image", + "image_path": "771318ed04d5bfb0f1facc9876755a4eaeb9b5d9e7315940173f08b7db155240.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 335, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 555, + 369 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons for different methods. First row: colour images. Second row: Monodepth2 [22]. Second row: CoopNet [28]. Third row: FeatDepth [66]. Last row: Ours. We compare qualitative results in four complex situations. First column: a thin object. Second column: a moving object. Third column: a large object with smoothness issue. Last column: Textured areas." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 377, + 293, + 435 + ], + "blocks": [ + { + "bbox": [ + 61, + 377, + 293, + 435 + ], + "lines": [ + { + "bbox": [ + 61, + 377, + 293, + 435 + ], + "spans": [ + { + "bbox": [ + 61, + 377, + 293, + 435 + ], + "type": "table", + "html": "
MethodsODSOISAP
Lego (paper)0.7100.7310.729
Lego*0.7090.7360.734
Depth & Normal pseudo-labels0.7550.7750.760
Depth & Normal pseudo-labels + Ls0.7600.7820.766
Ours = Depth & Normal pseudo-labels + Ls + Lc0.7620.7890.770
", + "image_path": "57bbf687abddf0fd93de0dfc51f4d397a1b1a75f8b98a5f25c95adcf132eb152.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 440, + 295, + 473 + ], + "lines": [ + { + "bbox": [ + 55, + 440, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 295, + 473 + ], + "type": "text", + "content": "Table 5. Comparison for contour estimation. Block#1: Lego results. Block#2: Ablation study of losses in contour estimation training. Experiments done on Cityscapes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 496, + 222, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 222, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 222, + 509 + ], + "type": "text", + "content": "5.5. Constancy assumption validity" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 514, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 622 + ], + "type": "text", + "content": "We evaluate the constancy assumption validity of our method, compared to [66] that uses deep features from the ResNet-50 encoder in the re-projection. To do so, we use the KITTI MOTS [72] dataset to track a specific point of an instance along different trajectories and compute the normalized temporal variance. We show that our distance transform satisfies much better the constancy assumption than any layers of the ResNet-50 (see Figure 4). More details on the evaluation protocol are given in Appendix E." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 634, + 128, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 128, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 128, + 645 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "As a conclusion, this paper presents a novel self-supervised approach to monocular depth estimation, addressing challenges in low-texture regions by applying a distance transform over pre-semantic contours. This technique enhances depth prediction accuracy by increasing discrimini" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 323, + 392, + 531, + 521 + ], + "blocks": [ + { + "bbox": [ + 323, + 392, + 531, + 521 + ], + "lines": [ + { + "bbox": [ + 323, + 392, + 531, + 521 + ], + "spans": [ + { + "bbox": [ + 323, + 392, + 531, + 521 + ], + "type": "image", + "image_path": "bda0b3345ff6e571496b7f43ea9f3144caf766d7eba9c5a1b5e05446371c42c1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 529, + 555, + 563 + ], + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 563 + ], + "type": "text", + "content": "Figure 4. Comparison of normalized temporal variance of different features for the centre pixel of an instance tracked along different sequence of KITTI MOTS [72]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 553, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 642 + ], + "type": "text", + "content": "native power where traditional photometric losses struggle. Both theoretical analysis and extensive experiments validate the effectiveness of the proposed method, which outperforms conventional self-supervised approaches." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 643, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 555, + 713 + ], + "type": "text", + "content": "In future work, we aim to introduce consistent variance in low-texture regions without pre-semantic contours. Additionally, we plan to explore whether our framework can leverage variance-augmented images to learn from diverse datasets, ultimately developing a foundation model in a fully self-supervised manner." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21875" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Ali Akbari, Muhammad Awais, Manijeh Bashar, and Josef Kittler. How does loss function affect generalization performance of deep learning? application to human age estimation. In International Conference on Machine Learning, pages 141-151. PMLR, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "text", + "content": "[2] Abhishek Badki, Alejandro Troccoli, Kihwan Kim, Jan Kautz, Pradeep Sen, and Orazio Gallo. Bi3d: Stereo depth estimation via binary classifications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1600-1608, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 203, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 295, + 257 + ], + "type": "text", + "content": "[3] Jiawang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised scale-consistent depth and ego-motion learning from monocular video. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "text", + "content": "[4] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Depth prediction without the sensors: Leveraging structure for unsupervised learning from monocular videos. In Proceedings of the AAAI conference on artificial intelligence, pages 8001-8008, 2019. 2, 7, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 314, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 296, + 369 + ], + "type": "text", + "content": "[5] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Unsupervised monocular depth and ego-motion learning with structure and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 445 + ], + "type": "text", + "content": "[6] Tian Chen, Shijie An, Yuan Zhang, Chongyang Ma, Huayan Wang, Xiaoyan Guo, and Wen Zheng. Improving monocular depth estimation by leveraging structural awareness and complementary datasets. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 90-108. Springer, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "type": "text", + "content": "[7] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 1, 2, 7, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 502, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 546 + ], + "type": "text", + "content": "[8] Zhao Chen, Vijay Badrinarayanan, Gilad Drozdov, and Andrew Rabinovich. Estimating depth from rgb and sparse sensing. In Proceedings of the European Conference on Computer Vision (ECCV), pages 167-182, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 547, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 295, + 612 + ], + "type": "text", + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "text", + "content": "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[11] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In Proceedings of the European conference on computer vision (ECCV), pages 562-578, 2018. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[12] Tom van Dijk and Guido de Croon. How do neural networks see depth in single images? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2183-2191, 2019. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "text", + "content": "[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 1, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 163, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 206 + ], + "type": "text", + "content": "[14] F El Jamiy and R Marsh. Survey on depth perception in head mounted displays: distance estimation in virtual reality, augmented reality, and mixed reality. iet image processing 13 (5), 707-712 (2019), 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 209, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 554, + 262 + ], + "type": "text", + "content": "[15] Sara Elkerdawy, Hong Zhang, and Nilanjan Ray. Lightweight monocular depth estimation model by joint end-to-end filter pruning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 4290-4294. IEEE, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[16] Chao Fan, Zhenyu Yin, Yue Li, and Feiqing Zhang. Deeper into self-supervised monocular indoor depth estimation. arXiv preprint arXiv:2312.01283, 2023. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 299, + 554, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 554, + 352 + ], + "type": "text", + "content": "[17] Philipp Fischer, Alexey Dosovitskiy, Eddy Ilg, Philip Häusser, Caner Hazirbaş, Vladimir Golkov, Patrick Van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. arXiv preprint arXiv:1504.06852, 2015. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 355, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 554, + 409 + ], + "type": "text", + "content": "[18] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 411, + 554, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 554, + 476 + ], + "type": "text", + "content": "[19] Ravi Garg, Vijay Kumar Bg, Gustavo Carneiro, and Ian Reid. Unsupervised cnn for single view depth estimation: Geometry to the rescue. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 740-756. Springer, 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 478, + 554, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 554, + 520 + ], + "type": "text", + "content": "[20] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 523, + 554, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 554, + 577 + ], + "type": "text", + "content": "[21] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 579, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 554, + 633 + ], + "type": "text", + "content": "[22] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 2, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 635, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 554, + 689 + ], + "type": "text", + "content": "[23] Ariel Gordon, Hanhan Li, Rico Jonschkowski, and Anelia Angelova. Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8977-8986, 2019. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[24] Vitor Guizilini, Rui Hou, Jie Li, Rares Ambrus, and Adrien Gaidon. Semantically-guided representation learn" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21876" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "ing for self-supervised monocular depth. arXiv preprint arXiv:2002.12319, 2020. 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 148 + ], + "type": "text", + "content": "[25] Hyowon Ha, Sunghoon Im, Jaesik Park, Hae-Gon Jeon, and In So Kweon. High-quality depth from uncalibrated small motion clip. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5413-5421, 2016. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 217 + ], + "type": "text", + "content": "[26] David Hafner, Oliver Demetz, and Joachim Weickert. Why is the census transform good for robust optic flow computation? In Scale Space and Variational Methods in Computer Vision: 4th International Conference, SSVM 2013, Schloss Seggau, Leibnitz, Austria, June 2-6, 2013. Proceedings 4, pages 210-221. Springer, 2013. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 259 + ], + "type": "text", + "content": "[27] Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International conference on machine learning, pages 1225-1234. PMLR, 2016. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 316 + ], + "type": "text", + "content": "[28] Marwane Hariat, Antoine Manzanera, and David Filliat. Rebalancing gradient to improve self-supervised co-training of depth, odometry and optical flow predictions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1267-1276, 2023. 2, 6, 7, 8, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 295, + 360 + ], + "type": "text", + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 295, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 404 + ], + "type": "text", + "content": "[30] Mu He, Le Hui, Yikai Bian, Jian Ren, Jin Xie, and Jian Yang. Ra-depth: Resolution adaptive self-supervised monocular depth estimation. In European Conference on Computer Vision, pages 565-581. Springer, 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 437 + ], + "type": "text", + "content": "[31] Stuart B Heinrich and Wesley E Snyder. Improved edge awareness in discontinuity preserving smoothing. arXiv preprint arXiv:1103.5808, 2011. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 295, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 481 + ], + "type": "text", + "content": "[32] Junjie Hu, Yan Zhang, and Takayuki Okatani. Visualization of convolutional neural networks for monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3869-3878, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 295, + 536 + ], + "type": "text", + "content": "[33] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2462-2470, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 295, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 569 + ], + "type": "text", + "content": "[34] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "text", + "content": "[35] Wei Jiang, Weiwei Sun, Andrea Tagliasacchi, Eduard Trulls, and Kwang Moo Yi. Linearized multi-sampling for differentiable image transformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2988-2997, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 668 + ], + "type": "text", + "content": "[36] Kenji Kawaguchi and Qingyun Sun. A recipe for global convergence guarantee in deep neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 8074-8082, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "text", + "content": "[37] Ue-Hwan Kim, Gyeong-Min Lee, and Jong-Hwan Kim. Revisiting self-supervised monocular depth estimation. In International Conference on Robot Intelligence Technology and Applications, pages 336-350. Springer, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "type": "text", + "content": "[38] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[39] Yevhen Kuznietsov, Marc Proesmans, and Luc Van Gool. Comoda: Continuous monocular depth adaptation using past experiences. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2907-2917, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[40] Seokju Lee, Sunghoon Im, Stephen Lin, and In So Kweon. Learning residual flow as dynamic motion from stereo videos. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1180-1186. IEEE, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[41] Hanhan Li, Ariel Gordon, Hang Zhao, Vincent Casser, and Anelia Angelova. Unsupervised monocular depth learning in dynamic scenes. In Conference on Robot Learning, pages 1908-1917. PMLR, 2021. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[42] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. Rtm3d: Real-time monocular 3d detection from object keypoints for autonomous driving. In European Conference on Computer Vision, pages 644-660. Springer, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "text", + "content": "[43] Pengzhi Li, Yikang Ding, Haohan Wang, Chengshuai Tang, and Zhiheng Li. The devil is in the edges: Monocular depth estimation with edge-aware consistency fusion. arXiv preprint arXiv:2404.00373, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 398 + ], + "type": "text", + "content": "[44] Runze Li, Pan Ji, Yi Xu, and Bir Bhanu. Monoindoor++: Towards better practice of self-supervised monocular depth estimation for indoor environments. IEEE Transactions on Circuits and Systems for Video Technology, 33(2):830-846, 2022. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[45] Rui Li, Danna Xue, Shaolin Su, Xiantuo He, Qing Mao, Yu Zhu, Jinqiu Sun, and Yanning Zhang. Learning depth via leveraging semantics: Self-supervised monocular depth estimation with both implicit and explicit semantic guidance. Pattern Recognition, 137:109297, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 455, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 499 + ], + "type": "text", + "content": "[46] Yin Li, Manohar Paluri, James M Rehg, and Piotr Dólar. Unsupervised learning of edges. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1619-1627, 2016. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "text", + "content": "[47] Liang Liu, Guangyao Zhai, Wenlong Ye, and Yong Liu. Unsupervised learning of scene flow estimation fusing with local rigidity. In IJCAI, pages 876-882, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 534, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 567 + ], + "type": "text", + "content": "[48] Hengjie Lu, Shugong Xu, and Shan Cao. Sgtbn: generating dense depth maps from single-line lidar. IEEE Sensors Journal, 21(17):19091-19100, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "type": "text", + "content": "[49] Xiaoyang Lyu, Liang Liu, Mengmeng Wang, Xin Kong, Lina Liu, Yong Liu, Xinxin Chen, and Yi Yuan. Hr-depth: High resolution self-supervised monocular depth estimation. In Proceedings of the AAAI conference on artificial intelligence, pages 2294-2301, 2021. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "text", + "content": "[50] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[51] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21877" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[52] Oskar Natan and Jun Miura. End-to-end autonomous driving with semantic depth cloud mapping and multi-agent. IEEE Transactions on Intelligent Vehicles, 8(1):557-571, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "text", + "content": "[53] Ales Nekvinda and Ludek Zajicek. A simple proof of the rademacher theorem. Časopis pro pěstovánt matematíky, 113 (4):337-341, 1988. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[54] C Wayne Niblack, Phillip B Gibbons, and David W Capson. Generating skeletons and centerlines from the distance transform. CVGIP: Graphical Models and image processing, 54 (5):420-437, 1992. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 205 + ], + "type": "text", + "content": "[55] Jim Nilsson and Tomas Akenine-Möller. Understanding ssm. arXiv preprint arXiv:2006.13846, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "type": "text", + "content": "[56] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "text", + "content": "[57] Daniel Otero, Davide La Torre, O Michailovich, and Edward R Vrscay. Optimization of structural similarity in mathematical imaging. Optimization and Engineering, 22:2367-2401, 2021. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 305, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 369 + ], + "type": "text", + "content": "[58] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "type": "text", + "content": "[59] Michael Ramamonjisoa and Vincent Lepetit. Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 492 + ], + "type": "text", + "content": "[60] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12240-12249, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 548 + ], + "type": "text", + "content": "[61] Alex Rasla and Michael Beyeler. The relative importance of depth cues and semantic edges for indoor mobility using simulated prosthetic vision in immersive virtual reality. In Proceedings of the 28th ACM Symposium on Virtual Reality Software and Technology, pages 1-11, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "type": "text", + "content": "[62] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "type": "text", + "content": "[63] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[64] Faraz Saeedan and Stefan Roth. Boosting monocular depth with panoptic segmentation maps. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3853-3862, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[65] Abhijeet Shenoi, Mihir Patel, JunYoung Gwak, Patrick Goebel, Amir Sadeghian, Hamid Rezatofighi, Roberto Martin-Martin, and Silvio Savarese. Jrmot: A real-time 3d multi-object tracker and a new large-scale dataset. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10335-10342. IEEE, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "text", + "content": "[66] Chang Shu, Kun Yu, Zhixiang Duan, and Kuiyuan Yang. Feature-metric loss for self-supervised learning of depth and egomotion. In European Conference on Computer Vision, pages 572-588. Springer, 2020. 3, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 184, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 249 + ], + "type": "text", + "content": "[67] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 250, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 271 + ], + "type": "text", + "content": "[68] Tilo Strutz. The distance transform and its computation. arXiv preprint arXiv:2106.03503, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 338 + ], + "type": "text", + "content": "[69] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "text", + "content": "[70] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 371, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 392 + ], + "type": "text", + "content": "[71] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "text", + "content": "[72] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. MOTS: Multi-object tracking and segmentation. In CVPR, 2019. 8, 6, 7, 11" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "text", + "content": "[73] Chaoyang Wang, José Miguel Buenapasada, Rui Zhu, and Simon Lucey. Learning depth from monocular videos using direct methods. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2022–2030, 2018. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 492, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 536 + ], + "type": "text", + "content": "[74] Peng Wang, Xiaohui Shen, Bryan Russell, Scott Cohen, Brian Price, and Alan L Yuille. Surge: Surface regularized geometry estimation from a single image. Advances in Neural Information Processing Systems, 29, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "text", + "content": "[75] Yang Wang, Yi Yang, Zhenheng Yang, Liang Zhao, Peng Wang, and Wei Xu. Occlusion aware unsupervised learning of optical flow. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4884-4893, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 592, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 658 + ], + "type": "text", + "content": "[76] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[77] Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1164-1174, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21878" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[78] Diana Wofk, Fangchang Ma, Tien-Ju Yang, Sertac Karaman, and Vivienne Sze. Fastdepth: Fast monocular depth estimation on embedded systems. In 2019 International Conference on Robotics and Automation (ICRA), pages 6101-6108. IEEE, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "text", + "content": "[79] Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single image depth prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 611-620, 2020. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "type": "text", + "content": "[80] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "text", + "content": "[81] Wanpeng Xu, Ling Zou, Lingda Wu, and Zhipeng Fu. Self-supervised monocular depth learning in low-texture areas. Remote Sensing, 13(9):1673, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 305 + ], + "type": "text", + "content": "[82] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 348 + ], + "type": "text", + "content": "[83] Z Yang, P Wang, W Xu, L Zhao, and R Nevatia. Unsupervised learning of geometry with edge-aware depth-normal consistency. arxiv 2017. arXiv preprint arXiv:1711.03665, 2 (5). 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "text", + "content": "[84] Zhenheng Yang, Peng Wang, Yang Wang, Wei Xu, and Ram Nevatia. Lego: Learning edge with geometry all at once by watching videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 225-234, 2018. 3, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "type": "text", + "content": "[85] Wei Yin, Yifan Liu, Chunhua Shen, and Youliang Yan. Enforcing geometric constraints of virtual normal for depth prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5684-5693, 2019. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "text", + "content": "[86] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1983-1992, 2018. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "text", + "content": "[87] Xun Yuan and Song Chen. Sad-slam: A visual slam based on semantic and depth information. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4930-4935. IEEE, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 602 + ], + "type": "text", + "content": "[88] Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 340-349, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 667 + ], + "type": "text", + "content": "[89] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. In 2022 international conference on 3D vision (3DV), pages 668-678. IEEE, 2022. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[90] Hang Zhou, David Greenwood, and Sarah Taylor. Self-supervised monocular depth estimation with internal feature fusion. In *British Machine Vision Conference (BMVC)*, 2021. 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 554, + 228 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 315, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 554, + 127 + ], + "type": "text", + "content": "[91] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 1, 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 129, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 554, + 183 + ], + "type": "text", + "content": "[92] Shengjie Zhu, Garrick Brazil, and Xiaoming Liu. The edge of depth: Explicit constraints between segmentation and depth. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13116-13125, 2020. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 184, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 554, + 228 + ], + "type": "text", + "content": "[93] Yuliang Zou, Zelun Luo, and Jia-Bin Huang. Df-net: Unsupervised joint learning of depth and flow using cross-task consistency. In Proceedings of the European conference on computer vision (ECCV), pages 36-53, 2018. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21879" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improved Video VAE for Latent Video Diffusion Model/d6e6ed64-039e-4e0b-a0ce-69a2cf999362_content_list.json b/2025/Improved Video VAE for Latent Video Diffusion Model/d6e6ed64-039e-4e0b-a0ce-69a2cf999362_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..47b308d2d876afe88e10df86bfa1fc035fa359de --- /dev/null +++ b/2025/Improved Video VAE for Latent Video Diffusion Model/d6e6ed64-039e-4e0b-a0ce-69a2cf999362_content_list.json @@ -0,0 +1,1365 @@ +[ + { + "type": "text", + "text": "Improved Video VAE for Latent Video Diffusion Model", + "text_level": 1, + "bbox": [ + 218, + 130, + 779, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pingyu Wu $^{1}$ , Kai Zhu $^{1,*}$ , Yu Liu $^{2}$ , Liming Zhao $^{2}$ , Wei Zhai $^{1,*}$ , Yang Cao $^{1}$ , Zheng-Jun Zha $^{1}$ $^{1}$ MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China $^{2}$ Independent Researcher {wpy364755620@mail., zkzy@mail., wzhai056@, forrest@, zhazj@}ustc.edu.cn", + "bbox": [ + 140, + 179, + 854, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 328, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Variational Autoencoder (VAE) aims to compress pixel data into low-dimensional latent space, playing an important role in OpenAI's Sora and other latent video diffusion generation models. While most existing video VAEs inflate a pre-trained image VAE into the 3D causal structure for temporal-spatial compression, this paper presents two astonishing findings: (1) The initialization from a well-trained image VAE with the same latent dimensions is not an optimal scheme. (2) The adoption of causal reasoning leads to unequal information interactions and unbalanced performance between frames. To alleviate these problems, we propose a keyframe-based temporal compression (KTC) architecture and a group causal convolution (GCCConv) module to further improve video VAE (IV-VAE). Specifically, the KTC architecture divides the latent space into two branches, in which one half completely inherits the compression prior of keyframes from a lower-dimension image VAE while the other half involves temporal compression through 3D group causal convolution, reducing temporal-spatial conflicts and accelerating the convergence speed of video VAE. The GCConv in the above 3D half uses standard convolution within each frame group to ensure inter-frame equivalence, and employs causal logical padding between groups to maintain flexibility in processing variable frame video. Extensive experiments on five benchmarks demonstrate the SOTA video reconstruction and generation abilities of our IV-VAE.", + "bbox": [ + 88, + 318, + 485, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 742, + 223, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, video generation has made significant advances in both academia and industry, showcasing cinematic-level visuals and the potential of a world simulator (e.g., Sora [16]). In particular, diffusion optimization methods on latent space have garnered widespread attention and dominated the video generation field due to their high efficiency and stability, resulting in a series of outstanding", + "bbox": [ + 89, + 768, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Latent Video Diffusion Models (LVDMs) [1, 10, 12, 32].", + "bbox": [ + 511, + 287, + 890, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To achieve a bidirectional mapping between the high-dimensional pixel space and the low-dimensional latent space, Variational Autoencoder (VAE) [13] is employed to encode the original video continuously into the latent space, along with spatial and even temporal compression. The compression rate and reconstruction quality of VAE directly determine the information effectiveness of the video corpus, playing an important role in the exploration of LVDMs.", + "bbox": [ + 511, + 304, + 905, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As the most commonly used video VAE, SVD-VAE [1] freezes the weights of the 2D encoder based on the image VAE [19] and retransmits a temporal decoder containing 3D convolutional blocks, which enables the feature interaction across different frames and significantly reduces flicker artifacts. Unfortunately, this method struggles to effectively eliminate temporal redundancy, limiting the LVDMs to perceive and generate long sequences of video content.", + "bbox": [ + 511, + 428, + 905, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the above problem, several works [6, 10, 18, 29, 37] have proposed to construct a 3D VAE with both spatial and temporal compression capabilities, which can be summarized in the following three points. In terms of model structure, the UNet network [20] with strong pixel-perception capabilities is still adopted, consistent with the classic image VAE series (i.e., Stable Diffusion VAE). For the operator implementation, most of the methods [6, 10, 18, 29] adopt 3D causal convolution with only forward sequence correlation, which can encode the image information independently [31] while ensuring the continuity of the long sequence video encoding. During the optimization process, a well-trained 2D image VAE, focusing only on spatial compression, is used as an initialization for 3D video VAEs with the same structure and equivalent latent channel number [6, 10, 18, 37], which can achieve better results compared to training a temporal-spatial compressed video VAE from scratch [30]. Although existing methods achieve video reconstruction under $4 \\times$ time compression, we find that there remain deficiencies in the above three key processes, limiting the improvement of the video quality and continuity in reconstruction results.", + "bbox": [ + 511, + 550, + 905, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "First, initialization from a well-trained image VAE with", + "bbox": [ + 531, + 885, + 905, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Kai Zhu and Wei Zhai are the corresponding authors.", + "bbox": [ + 107, + 886, + 397, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18124", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/049f5a1c964c3e29e38652a45f4217ece8f5fe43804b05a2585af70553dfb43f.jpg", + "image_caption": [ + "(a) Number of latent channels" + ], + "image_footnote": [], + "bbox": [ + 129, + 84, + 606, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/631ef923a957e00d8cd86b993c6fcd051cc7f92fbb009e5a5fc47fb10ba7bf5a.jpg", + "image_caption": [ + "(b) frames within a frame group", + "Figure 1. (a) Information preservation degree for different VAEs on video data. CR: Compression Rate. (b) Performance of different frames within a frame group on kinetics-600 dataset (OD-VAE [6], OS (OpenSora V1.2)-VAE [10], CogX (CogVideoX)-VAE [29])." + ], + "image_footnote": [], + "bbox": [ + 614, + 85, + 864, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the same latent channel number cannot support further $4\\times$ time compression, leading to a sharp decline in the initial spatial compression performance and slow learning of subsequent temporal compression. As shown in Fig. 1(a), we calculate the reconstruction similarity of different VAE on video data as the information preservation degree (See Supplementary Material). We note that: (1) As the latent channel number $Z$ increases, the spatial compression capacity gain produced by the image VAE gradually diminishes (blue dashed line) (2) The spatial compression capability provided by an image VAE with dimension $Z / 2$ is sufficient for a temporal-spatial compressed video VAE with dimension $Z$ . Then the video VAE can use the remaining latent channels to store the spatial information of extra frames, i.e., acquiring the initial $2\\times$ time compression capacity, which (red curves) converges even better and faster than the regular initialization approach [6, 10, 18, 37] that inherits the image VAE weights of the same dimension $Z$ (green curves). This inspires us to start from an image prior with a lower latent channel number and leave the remaining latent channels for temporal compression, promoting a more balanced temporal-spatial compression initialization. To this end, we propose a dual-branch keyframe-based temporal compression architecture, where a 2D branch inheriting the low-dimensional image VAE prior focuses on keyframe compression, and a 3D branch facilitates temporal compression.", + "bbox": [ + 91, + 338, + 483, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, the unidirectional information flow of 3D causal convolution prevents the interaction between video frames, exacerbating inter-frame unbalanced performance under time compression. In the forward process, $M$ frames (a frame group) are compressed to a whole (one frame) and then restored to $M$ frames. Due to the restriction of forward flow of causal convolution, the first frame in the frame group is unable to perceive the other frames until after time compression in the middle layer of the network. Thus the first frame has the poorest ability to interact with the whole frame group, resulting in higher optimization diffi", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "culty compared to other frames. As shown in Fig. 1(b), the methods [6, 10, 29] based on causal convolution suffers from poor performance of the first frame in the frame group and large fluctuations in SSIM values between frames, leading to significant flickering in the reconstruction results. This inspires us to propose group causal convolution, which groups the input frames based on the temporal compression rate. Causal logic padding operations are applied between frame groups to ensure the continuity in processing variable frame videos, and standard convolution operations are used within each group of $M$ frames to share equivalent information interactions, achieving smoother reconstruction.", + "bbox": [ + 511, + 338, + 906, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, the local receptive field struggles to meet the temporal compression demand at high resolutions. For the same video, the higher the resolution, the more pixels the same motion spans, making it more challenging for the model to capture motion patterns. To this end, we introduce atrous convolutions and add multiple attention modules to expand the receptive field, thus improving the temporal motion perception of the model at high resolutions.", + "bbox": [ + 511, + 520, + 908, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, considering the increasing demand for content diversity and high resolution in the video generation field, there is a lack of a suitable open-source benchmark for comprehensive evaluation of video VAE reconstruction capabilities. Most of the datasets (e.g., Kinetics-600 [4] and UCF-101 [21]) are 720P or lower, and cannot be used for high resolution (e.g., 1080P) evaluations. Moreover, high-resolution datasets (e.g., Panda-70M [8]) consist mainly of slow-motion and fixed-shot videos, which hardly reflects the overall performance of video VAEs under varying motion speeds. Therefore, we re-collect a subset of 2000 1080P videos from OpenVid-0.4M [15] dataset, uniformly containing various motion speeds, named MotionHD, as a supplement to the overall evaluation. Experimental results show that our method achieves state-of-the-art video reconstruction at various resolutions and motion rates.", + "bbox": [ + 511, + 642, + 908, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of this paper include: (1)", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18125", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a keyframe-based temporal compression architecture to accelerate convergence speed by providing a more balanced and powerful temporal-spatial compression initialization. (2) We introduce group causal convolution as a replacement for causal convolution to achieve better and balanced performance between frames, while maintaining causal logic. (3) Extensive experiments on five benchmarks demonstrate the SOTA video reconstruction and generation capabilities of the proposed IV-VAE.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 241, + 227, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Variational Autoencoder. Variational Autoencoder (VAE), as a widely used model in the generation domain, is adept at learning complex distributions from high-dimensional data. VAEs can be categorized into two types based on the class of tokens, including discrete and continuous. Classical work of the former type includes MaskGiT [5], VQGAN [9], etc., where continuous latent features are discretized by quantization techniques for autoregression generation. The latter type of continuous VAE, which can achieve spatial or even temporal compression by mapping high-dimensional data distributions to a low-dimensional latent space, is widely used in latent diffusion models [33].", + "bbox": [ + 89, + 266, + 482, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video VAE. Earlier video generation works [1, 28, 38] usually use a image VAE to compress video frames, or fine-tune the image VAE with additional time-dimensional convolutions to reduce flickering artifacts, but this approach cannot effectively reduce the redundancy in the time dimension. With the emergence of OpenAI's Sora [16], to allow video generation models to learn and generate longer videos, there have been some recent works [6, 10, 29, 37] that attempt to train a video VAE with a larger time compression ratios (e.g., $4 \\times$ ). Open-Sora [10] adopts a stacked approach, first compressing the video frames in the spatial dimension based on SDXL image VAE [17], and then introducing 3D VAE to compress the temporal dimension. ODVAE [6] uses tail initialization to inflate the 2D convolutions of SD image VAE into 3D causal convolutions and explores an efficient model variant. CV-VAE [37] uses latent space regularization to avoid latent space distribution shift while compressing time. CogVideoX [29] constructs a multi-GPU parallel computing approach to process long videos for causal video VAE. Based on the general idea of these methods, we further explore the shortcomings of the existing video VAEs and propose a series of improvements.", + "bbox": [ + 89, + 448, + 482, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 792, + 181, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Baseline 3D Causal VAE", + "text_level": 1, + "bbox": [ + 89, + 816, + 313, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the beginning, we introduce a causal video VAE as the baseline method. We first train an image VAE based on the architecture of SD VAE with a few structural modifications (Sec. 3.5). Following the thought of MagViT-v2 [31],", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we choose 3D causal convolutions in the construction of video VAE. Then, drawing on the idea of central initialization [3], we inflate the 2D convolution in image VAE to the 3D causal convolution. For example, noting 2D convolutional weights as $W_{2d} \\in \\mathbb{R}^{c_o \\times c_i \\times k_h \\times k_w}$ , when inflating $W_{2d}$ to be $W_{3d} \\in \\mathbb{R}^{c_o \\times c_i \\times k_t \\times k_h \\times k_w}$ , $W_{3d}$ is formulated as:", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW _ {3 d} [:,: i,:,: ] = \\left\\{ \\begin{array}{c c} W _ {2 d} & , \\text {i f} i = - 1 \\\\ 0 & , \\text {e l s e} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 184, + 903, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(See supplementary material for details). In this way, we can obtain high-quality image reconstruction results after extending image VAE to a causal video VAE. Lastly, except that the first image frame is processed separately, we let the causal video VAE achieve $4 \\times$ temporal compression.", + "bbox": [ + 511, + 220, + 903, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adopting a causal VAE as the baseline has the following two advantages: (1) The first frame is always independent of other frames. According to MagViT-v2 [31], using a temporal causal VAE instead of a regular VAE makes the model perform better in processing a single image. (2) Each frame depends only on the previous frames. Thus, a temporal causal video VAE can use cache mechanism in the inference, which splits the reconstruction process of a long video into a series of short videos with exactly the same result. We consider this a better operation compared to regular leap approach, more details are displayed in Sec. 4.5.", + "bbox": [ + 511, + 295, + 906, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Group Causal Convolution", + "text_level": 1, + "bbox": [ + 511, + 470, + 756, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the above baseline, we note that it is practically impossible and non-optimal to ensure causal logic for all frames. For example, the baseline approach compresses time by $4 \\times$ , then the input 4 frames will be compressed into 1 frame by the encoder and recovered to 4 frames through the decoder. Obviously, these 4 frames recovered from the same compressed frame do not satisfy temporal causal correlation. In addition, for these 4 reconstructed frames, the first frame usually does not perform as well as subsequent frames, as shown in Fig. 1(b), resulting in frequent flickering in the reconstructed video. We attribute this problem to the insufficient and unbalanced interaction between neighboring frames due to the causal convolution.", + "bbox": [ + 511, + 493, + 905, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, to alleviate this problem while being able to maintain the advantages of causal VAE (Sec. 3.1), we propose group causal convolution (GCCov). For a video VAE with a temporal compression rate of $t_c$ ( $t_c = 4$ in our method), the proposed method will initially label every $t_c$ frames of the input video as a frame group, with no overlap between the frame groups. Specifically, the first image frame is a separate frame group. For a frame group, we aim to make the features between frames visible and interactable with each other, thus increasing the reconstruction smoothness between frames and reducing the flicker. For different frame groups, we maintain the property of causal convolution that the current frame group only depends on the previous frame groups. Therefore, the time padding scheme", + "bbox": [ + 511, + 689, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18126", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/44205ba677c2eeabfa9a02a0fe2c45d5b12397587e8efedcf4a0be93178d2198.jpg", + "image_caption": [ + "Figure 2. (a-c) Basic components of IV-VAE. (d) Overall Framework." + ], + "image_footnote": [], + "bbox": [ + 101, + 85, + 419, + 291 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c29d8c47da4c6bd77e9741d8de703d2183d542f62c7c52c18d271e5a0ee83313.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 85, + 709, + 287 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4b090dd86450d021feb02744f0d705eab76a92a243838973e2b633174c73bfae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 85, + 890, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for group causal convolution is designed in Fig. 2(a). We pad the features of the previous frame before each frame group, in particular, the first image frame is padded using its own features. And we use zero padding after each frame group. After causal logic padding, each group is operated separately using the same regular convolution.", + "bbox": [ + 89, + 313, + 483, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Keyframe-based temporal compression", + "text_level": 1, + "bbox": [ + 89, + 421, + 426, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building on the observation from Fig. 1(a), to obtain better temporal-spatial compression capability at the initialization of video VAE with latent channel number $Z$ , it is feasible to utilize the pre-trained image VAE weights with a low latent channel number, i.e., $Z / 2$ . To this end, we propose a dual-branch keyframe-based temporal compression (KTC) architecture that decouples the compression of a frame group into the spatial compression of the keyframe and the temporal compression of the overall motion, where the first frame of each frame group is regarded as the keyframe since it can be perceived by the other frames along the temporal forward flow. The basic unit is shown in Fig. 2(b). For the input feature map $F$ with $C_{in}$ channels, it is fed into a 2D convolution to extract the image information and a 3D group causal convolution to extract the overall temporal motion information, respectively. The two convolutions each output feature maps with channels $C_{out} / 2$ , which are then separately normalized by RMSNorm and concatenated together.", + "bbox": [ + 88, + 444, + 482, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the initialization phase, the 2D and 3D branches of the network are initialized using the pre-trained image VAE weights with latent channel number $Z / 2$ respectively. In the 3D branch, the 2D image weights are inflated to 3D weights based on the center initialization [3]. In addition, we elaborately set the initial weights of the newly added temporal upsampling and downsampling layers of the two branches. For temporal downsampling layer, we make it initially behave as simply retaining all features of one of the $t_c$ frames and then discarding the remaining frames. For temporal upsampling layer, we let it initially behave as an operation of copying existing frames (More details are provided in Sup", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "plementary Material). In this way, each branch is initially enabled to reconstruct one frame from the input $t_c$ frames. With two branches reconstructing different frames, the proposed KTC structure initially realizes $2 \\times$ time compression with strong image compression capability as well.", + "bbox": [ + 511, + 314, + 903, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall structure of the network is based on UNet [20] and adopts the proposed KTC as the base unit, as shown in Fig. 2(d). After two temporal downsampling, a frame group is merged into one compressed frame, so we replace the GCConv3D in the KTC Unit with a 2D convolution at this stage to reduce the computation. In the output of the model, each of the two branches produces a sequence of video frames with the same shape as the input, where the reconstruction results of keyframes and remaining frames are selected from the 2D branch and 3D branch outputs respectively, and compose the final reconstruction result.", + "bbox": [ + 511, + 388, + 905, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Temporal Motion Perception Enhancement", + "text_level": 1, + "bbox": [ + 511, + 561, + 880, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the same video, as the video resolution increases, the pixel span generated by object motion or camera motion increases, and the more difficult it is for the video VAE to capture the motion patterns due to the limitations of the receptive field. To this end, we make two improvements to increase the receptive field of the video VAE to enhance the perception of temporal motion at high resolution. As shown in Fig. 2(c), we introduce multiple parallel atrous convolutions (PAC) with different atrous rates in the last layer of the encoder, borrowing from the ASPP in DeepLabV3+ [7]. In addition, we expand the number of attention modules from 2 to 7, and all attention modules are implemented after full temporal and spatial compression to reduce the computational effort. At this time, the compressed frame contains all the information of the frame group, and the global receptive field of attention modules can effectively facilitate temporal motion perception, especially at high resolution.", + "bbox": [ + 511, + 584, + 906, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Other structural modifications", + "text_level": 1, + "bbox": [ + 511, + 848, + 784, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the architecture of SD image VAE, we make some modifications to better adapt to 3D causal VAE. First, we", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18127", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b8a66ff573615e68cca34e293c97cd4b8ee6e3ae4883d825c549c660d14d35e.jpg", + "image_caption": [ + "Figure 3. Reconstruction results of different methods for a frame group." + ], + "image_footnote": [], + "bbox": [ + 96, + 84, + 897, + 295 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6657a2d5ec856a2280028ae4205f6378dc538c318ce6e0267b9bdb6ed7b81894.jpg", + "image_caption": [ + "Figure 4. Video motion distribution of Kinetics-600, OpenVid-0.4M, and the filtered MotionHD datasets." + ], + "image_footnote": [], + "bbox": [ + 94, + 320, + 383, + 452 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c53df5f4316c5c688d5d4a9365d8fa399acaab5637c77feb04e20050b5f6aebd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 321, + 635, + 452 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/823cfde276770a663594627e59d3eaa3c149cc66a8a339c6fef2cef860ad7b7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 321, + 767, + 452 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d4a2458e1c72398a06a7c36e6f926686fc475afe5cc5ba1ebc9d3ca58ae7ed2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 321, + 898, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "replace all GroupNorms with RMSNorms to ensure temporal causality following MagViT-v2 [31], since the standard GroupNorm involves time dimension when calculating the mean and variance of the feature maps, which destroys temporal causality. Second, we shrink the convolution channels during spatial upsampling instead of in the later residual block, reducing memory usage by $29\\%$ in the inference.", + "bbox": [ + 88, + 476, + 482, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 595, + 223, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. MotionHD dataset", + "text_level": 1, + "bbox": [ + 89, + 621, + 269, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To better evaluate video VAEs, we propose to consider the video reconstruction performance at different motion amplitudes and resolutions. Among the open-source video datasets, OpenVid-0.4M [15] is a recently released high-quality dataset consisting of 1080P resolution videos, but most of them are fixed camera or slow-moving. To present this phenomenon, we use the RAFT [22] model to compute the optical flow map of video at $768 \\times 432$ resolution and then get the average optical flow score as the motion score. As shown in Fig. 4, we present the motion distributions of Kinetics-600 [4] and OpenVid-0.4M datasets. Compared to Kinetics-600, the motion distribution of OpenVid-0.4M is overall narrower and concentrated in low-motion intervals, which is difficult to reflect the reconstruction performance at medium and fast motion. More importantly, in the training of diffusion models, it is common to use a high frame interval to facilitate the learning of object motion, e.g.,", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Latte [14] set the frame interval to 3 on all datasets. In this case, the motion of video will be more intense (Fig. 4(a)).", + "bbox": [ + 511, + 476, + 903, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, we refer to the motion distribution of the Kinetics-600 (frame interval of 3), and uniformly select 2000 videos with 1080P resolution from OpenVid-0.4M according to the motion scores, as illustrated in Fig. 4(b), constituting the subset MotionHD, which is used to test the average performance of video VAEs at various video motion speeds and the performance at different resolutions. We also list the motion scores of the videos generated by Sora [16] in Fig. 4(c), to facilitate the interpretation of scores and to indicate that our choice of motion intervals is reasonable.", + "bbox": [ + 511, + 507, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 667, + 702, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation details. We evaluate the reconstruction performance of video VAEs on multiple video benchmarks, including widely used [24-26, 34-36] Kinetics-600 [4] val set, ActivityNet [2] test set, and MotionHD. For Kinetics-600 and ActivityNet, we each randomly select 2400 videos and test them with the original resolution. MotionHD is used to test the performance of Video VAEs at various resolutions. All experiments are conducted using the first 17 frames of the video. To measure reconstruction quality, SSIM, PSNR, LPIPS, and FVD [23] are adopted as metrics.", + "bbox": [ + 511, + 688, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following OD-VAE [6], we employ Latte [14], a soralike transformer-based diffusion model for video generation. We choose Kinetics-400 [11] for class-conditional generation and SkyTimelapse [27] for unconditional gener", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18128", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6342f81a6b00d43f3f5a4bac0a6266e5755a79ee35ed5adf291445c7aa68d1b2.jpg", + "image_caption": [ + "Figure 5. Reconstruction results for two adjacent frames in a frame group. In the below figure, the regions with large differences between the two reconstructed frames are boxed out." + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 898, + 392 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/19be367f19fb642a8d7a5d3151493161d45ad35bc1a84c53d4173c5368da5299.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParamsFCRChnKinetics-600ActivityNet
FVD↓PSNR↑SSIM↑LPIPS↓FVD↓PSNR↑SSIM↑LPIPS↓
SVD-VAE97M1*8*846.2635.260.93630.036755.4737.070.92980.03217
CV-VAE182M4*8*8416.8131.580.90660.0827112.5734.870.91560.06846
OS-VAE393M4*8*8419.0534.370.92580.0816213.5637.640.93400.06781
OD-VAE239M4*8*8410.6933.880.92420.054858.1036.950.93210.04670
Causal VAE127M4*8*8410.8133.770.92480.063098.2637.060.93260.05336
IV-VAE107M4*8*848.0134.290.92810.052096.0838.470.93590.04436
Causal VAE127M4*8*885.7236.080.94790.040334.3938.360.95190.03458
IV-VAE107M4*8*883.9436.660.95200.035993.5039.610.95660.03072
CogX-VAE215M4*8*8163.1738.380.96770.028662.1740.680.97030.02468
Causal VAE127M4*8*8163.2838.070.96410.028872.4340.420.96710.02483
IV-VAE108M4*8*8162.9739.020.96850.022802.0142.610.97220.01968
", + "bbox": [ + 155, + 436, + 844, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of video reconstruction results. Chn: Latent channel number. FCR: Frame compression rate.", + "bbox": [ + 153, + 628, + 841, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ation. All methods are trained for 150K steps with the same default settings in Latte [14]. In addition, the parameters of the video are set to a resolution of $256 \\times 256$ and a frame length of 17 during both training and testing. We employ FVD as the metric and calculate it on 2048 samples.", + "bbox": [ + 88, + 657, + 482, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. We first train an image VAE on $256 \\times 256$ image data for 200K steps using 4 A800 GPUs, and then inflate it to a video VAE. During training, KL [13], MAE, and LPIPS are employed as losses. Next, we train the video VAE on $256 \\times 256$ resolution videos for 500K steps, and further expand the training resolution to $512 \\times 512$ training 200K steps. Finally, we train the VAE at different resolutions with different frame numbers for 100K steps. GAN loss from the 3D discriminator is added at this stage. For causal VAE, we set the basic channel number of the convolution to 96. For IV-VAE, we set it to 64 for both branches.", + "bbox": [ + 88, + 734, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Performance", + "text_level": 1, + "bbox": [ + 511, + 657, + 650, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to baseline Causal VAE, comparison methods include OpenSora V1.2 (OS)-VAE [10], OD-VAE [6], CV-VAE [37],CogVideoX (CogX)-VAE [29], SVD-VAE [1].", + "bbox": [ + 511, + 684, + 906, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative evaluation of reconstruction results. Table 1 demonstrates the performance of the proposed Causal VAE and IV-VAE with other methods on Kinetics-600 and ActivityNet datasets. Experiments show that the proposed method achieves excellent results on both datasets and various numbers of latent channels $Z$ . On $Z = 4$ , we outperform OD-VAE by 2.68 and 2.02 FVD reductions in Kinetics-600 and ActivityNet respectively with less than half parameters. Compared to OS-VAE, we are slightly lower on the PSNR metric due to the limitation of parameters, but we obtain a reduction of 286M (-73%) param", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18129", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cf4efeef2dcad2d4eda0a6fc98b9d91d437fe79fd13fdebc2570711134867661.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodChnFVD↓PSNR↑SSIM↑LPIPS↓
480P720P1080P480P720P1080P480P720P1080P480P720P1080P
SVD-VAE45.684.233.6233.7136.0037.690.93130.95070.96170.0350.0330.031
CV-VAE422.3312.788.5029.7030.2931.190.87960.90080.92060.0940.0980.098
OS-VAE416.8910.068.2132.3334.1335.650.91090.93100.94520.0860.0860.087
OD-VAE410.407.365.3131.4233.3234.890.90040.92260.93910.0590.0610.064
Causal VAE49.336.855.1831.6433.4334.820.90870.92860.94150.0670.0690.073
IV-VAE47.014.494.3132.0134.0835.800.91270.93330.94650.0530.0540.055
Causal VAE84.634.153.5834.1435.8937.150.93550.94810.95570.0430.0470.051
IV-VAE84.093.712.8634.6036.5638.110.94100.95340.96110.0370.0390.041
CogX-VAE163.503.662.2636.5237.8638.680.95710.96390.96760.0310.0350.039
Causal VAE163.763.852.7236.2537.6738.600.95440.96170.96630.0320.0370.041
IV-VAE162.582.782.1836.8038.8140.330.95850.96690.97200.0250.0270.027
", + "bbox": [ + 138, + 88, + 859, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "eters, 11.04 $(-58\\%)$ FVD, and 0.02953 $(-36\\%)$ LPIPS on Kinetics-600. Compared to SVD-VAE, although our model with $Z = 8$ uses a latent channel number twice as large, we achieve an additional $4\\times$ time compression and perform better on all metrics. In addition, we also obtain the best results on $Z = 16$ , especially in PSNR and LPIPS metrics.", + "bbox": [ + 89, + 310, + 480, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative evaluation of reconstruction results. As shown in Fig. 3 and 5, we exhibit the reconstruction results for different frames within a frame group. It can be observed that other methods usually perform poorly in the first two frames of a frame group, with blurring and a small amount of distortion, and the last frame performs better. This phenomenon validates our point that for a video VAE with temporal compression, the use of causal convolution makes the frames in the same frame group unequal with respect to the information interaction, leading to performance imbalance. While our method can achieve a more balanced performance benefiting from group causal convolution. Besides, due to the performance fluctuations, other methods usually suffer from abrupt changes in details in the reconstruction of neighboring frames, as shown in the lower panel of Fig. 5, and our method performs better in terms of temporal consistency and reconstruction quality. More visual comparisons are provided in the supplementary material.", + "bbox": [ + 89, + 401, + 482, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction performance at different resolutions. In Table 2, we present the performance of various methods at different resolutions on the MotionHD dataset. The proposed method achieves the best results in the majority of metrics. And as the resolution increases, our improvement relative to other methods is more pronounced. For example, in the PSNR metric, we outperform CogX-VAE by 0.28 at 480P while when the resolution is 1080P, this improvement increases to 1.65. Experiments indicate that the proposed KTC module and TMPE can enable the video VAE to achieve better temporal compression capability, leading to better results in challenging scenarios of temporal compression such as high resolution and rapid motion.", + "bbox": [ + 89, + 674, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Generation. In the table of Fig. 6, we present the video generation results of different video VAEs using", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7a45c543382b11c57d16814e9922294b7f705b173739ebcd5fd45c230bfa0c08.jpg", + "table_caption": [ + "Table 2. Video reconstruction results at different resolutions on MotionHD. 480P: ${848} \\times {480},{720}\\mathrm{P} : {1280} \\times {720},{1080}\\mathrm{P} : {1920} \\times {1280}$ ." + ], + "table_footnote": [], + "table_body": "
DatasetCV-VAEOS-VAEOD-VAEIV-VAE (Z=4)
Kinetics-400552.4617.6548.7541.6
SkyTimelapse128.0141.3140.0118.6
", + "bbox": [ + 517, + 311, + 901, + 361 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f76f9875d31f98510bd9adc7390fefd3e444539af7d90a1e9c07aba7a1ec3da5.jpg", + "image_caption": [ + "Figure 6. Top Table: FVD for different methods. Bottom Figure: Video frames generated by Latte with our IV-VAE." + ], + "image_footnote": [], + "bbox": [ + 517, + 363, + 901, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the same generation model Latte-XL (Params: 0.67B) [14]. The proposed VAE achieves the best performance on both class-conditional and unconditional generation. Especially on SkyTimelapse, we obtain a 9.4 FVD decrease compared to CV-VAE. The results indicate that IV-VAE can be well employed for LVDMs. At the bottom of Fig. 6, we show some frames of the video generated by Latte with our VAE. The frames exhibit a fairly high level of quality and realism, and even reflect some of the laws of physics, as the sunlight in the sky moves, so does the light reflected in the water.", + "bbox": [ + 511, + 470, + 905, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 633, + 666, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform a series of ablation experiments. Considering that ablation for three channels requires too much time and resources, we choose the middle channel $Z = 8$ for the ablation of IV-VAE composition. Each ablation experiment follows the same training program: We first train an image VAE for 200K steps, and then extend it to a video VAE and train it for 300K steps on $256 \\times 256$ videos.", + "bbox": [ + 511, + 657, + 905, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation studies of main components. Based on the baseline causal VAE, we explore the effects of the different compositions of IV-VAE in Table 3. As can be seen from the table, the implementation of the proposed group causal convolution can substantially improve the performance on both baseline and KTC structure with the parameter count unchanged. The results indicate that the reconstruction quality can be effectively improved by allowing the frames in the same frame group to interact bidirectionally. Further,", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18130", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/27ee2ff724b3a804405822d0051b754dec4222f7a23ce73160aedf340e4e0868.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingTest: 256×256Test: 512×512Params
GCCovKTCTMPEPSNR↑SSIM↑PSNR↑SSIM↑
(A)31.290.904234.880.9384127M
(B)✓31.640.908235.210.9420127M
(C)✓31.860.911635.460.9451104M
(D)✓✓32.120.914535.690.9478104M
(E)✓✓✓32.240.915835.850.9493107M
", + "bbox": [ + 91, + 89, + 482, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97bf7866dfeceac778b914a9159315aae81264fbeb392337df1fb6f0b4b3b6e6.jpg", + "image_caption": [ + "Figure 7. Ablation study of structures. The training loss is the sum of MAE, LPIPS, and KL losses." + ], + "image_footnote": [], + "bbox": [ + 109, + 229, + 460, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "replacing the base unit with KTC unit can substantially facilitate the learning of temporal compression thus improving the performance of the model, even with smaller parameters. Finally, we add TMPE enabling the model to better learn the object's motion patterns by increasing the global receptive field with only 3M additional parameter counts. From the improvement of the results at different resolutions, it can be analyzed that this motion perception enhancement plays a more important role in higher resolution scenes.", + "bbox": [ + 89, + 414, + 483, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of KTC at different latent channel numbers. To verify the validity of the KTC structure, we only replace the structure of the baseline causal VAE with the proposed KTC and perform experiments using the same training setup, the training loss is shown in Fig 7. When the latent channel number $Z = 4$ , the proposed KTC performs slightly superior to the baseline as the training steps increase. And this advantage extends significantly when $Z$ expands to 8 and 16. In conjunction with the results presented in Fig. 1, it can be analyzed that the spatial compression capability improvement of the image VAE is attenuated as $Z$ increases. Therefore, when $Z$ is large, dedicating a portion of the number of latent channels to the initial temporal compression can achieve better and balanced temporal-spatial compression capability initially, leading to faster convergence in subsequent training.", + "bbox": [ + 89, + 551, + 483, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Analysis", + "text_level": 1, + "bbox": [ + 89, + 803, + 192, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Increase parameters. In Table 4, we explore the boost provided by increasing the parameters. Increasing the basic feature dimension of IV-VAE from 64 to 96 dramatically increases the parameters and also brings significant improvements, particularly in PSNR and LPIPS metrics. Compared", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2e1a9f6eb7360b0ccc6c10926244c56fceb0f0d353cc313f603251c54608314e.jpg", + "table_caption": [ + "Table 3. Ablation study. (A) represents baseline causal VAE. The models are tested on Kinetics-600 using 256 and 512 resolutions." + ], + "table_footnote": [], + "table_body": "
Method (Z=16)ParamsKinetics-60025×1024×1024 Video Rec. time
PSNR↑SSIM↑LPIPS↓
CogX-VAE215M38.380.96770.028711.4s
Ours (dim=64)108M39.020.96850.02286.2s
Ours (dim=96)242M39.570.97110.020510.7s
", + "bbox": [ + 517, + 89, + 903, + 164 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/09ebff7cde4fb8a99b0957dc084c9e4d784131b55f2b70e98ae319f0b168505c.jpg", + "table_caption": [ + "Table 4. Increasing parameters. The reconstruction time is tested on one A800 GPU with BF16 precision. Rec.: Reconstruction." + ], + "table_footnote": [], + "table_body": "
Rec. approachInput: A 49-frame 1024×1024 VideoKinetics-600
Video clippingMem.Rec. timeSSIM PSNR
One-Step49-frame76.3G12.7s0.9685 39.02
Overleap5-frame*1214.3G14.7s0.9682 38.97
Cache1-frame + 4-frame*1214.1G13.0s0.9685 39.02
", + "bbox": [ + 517, + 200, + 903, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Cache mechanism vs Overleap mechanism. The overlap mechanism cannot produce the same results as one-step reconstruction. The results are evaluated on $Z = 16$ . Mem.: Memory.", + "bbox": [ + 511, + 279, + 903, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to CogX-VAE, this model achieves a 1.19 PSNR improvement and still takes less time to reconstruct the video.", + "bbox": [ + 511, + 333, + 903, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cache mechanism vs Overleap mechanism. To reconstruct a long video, it is necessary to split the video into a series of clips since GPU memory is limited. Overlap is the most commonly used approach [6], where each input video clip has a one-frame overlap with the previous clip as a history frame, and the overlapped frame is discarded after reconstruction. However, reconstruction with overlap is not equivalent to one-step reconstruction, which may introduce semantic bias in latent space due to the short history frame. And the overlapping frames lead to extra computation.", + "bbox": [ + 511, + 363, + 905, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address this problem, we introduce the cache mechanism, which is based on the property that the causal network depends only on the previous frames in temporal order. According to the time padding length $P_{t}$ needed for 3D convolution, the last $P_{t}$ frames of the current video clip's feature maps at each layer of the network are saved, which are used for the convolutional padding of the next video clip. Therefore, this mechanism can be fully equivalent to one-step reconstruction in terms of results. In addition, as shown in Table 5, the caching mechanism requires less time usage and GPU memory usage compared to the overlap mechanism.", + "bbox": [ + 511, + 513, + 906, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we observe that regular initialization of utilizing image VAE weights with the same latent channel number is not an optimal scheme, and the use of causal convolution leads to performance imbalance between frames. To address these problems, we propose a keyframe-based temporal compression architecture to accelerate convergence speed by providing more balanced and powerful temporal-spatial compression initialization. In addition, group causal convolution is introduced to achieve better and more balanced performance between frames. Extensive experiments show that IV-VAE achieves SOTA results in both video reconstruction and generation.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18131", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62306295, 62225207 and 62436008.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 148, + 187, + 165 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 3, 6", + "[2] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 5", + "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 3, 4", + "[4] Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short note about kinetics-600. arXiv preprint arXiv:1808.01340, 2018. 2, 5", + "[5] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[6] Liuhan Chen, Zongjian Li, Bin Lin, Bin Zhu, Qian Wang, Shenghai Yuan, Xing Zhou, Xinghua Cheng, and Li Yuan. Od-vae: An omni-dimensional video compressor for improving latent video diffusion model. arXiv preprint arXiv:2409.01199, 2024. 1, 2, 3, 5, 6, 8", + "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 4", + "[8] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2", + "[9] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2021. 3", + "[10] hpcaitech. Open-sora. https://github.com/hpcaitech/Open-Sora, 2024. 1, 2, 3, 6", + "[11] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5", + "[12] Kihong Kim, Haneol Lee, Jihye Park, Seyeon Kim, Kwanghee Lee, Seungryong Kim, and Jaejun Yoo. Hybrid video diffusion models with 2d triplane and 3d wavelet representation. In European Conference on Computer Vision, pages 148-165. Springer, 2024. 1" + ], + "bbox": [ + 93, + 172, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 6", + "[14] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 5, 6, 7", + "[15] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 2, 5", + "[16] OpenAI. Openai sora. Accessed May 15, 2024 [Online], 2024. 1, 3, 5", + "[17] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations. 3", + "[18] Can Qin, Congying Xia, Krithika Ramakrishnan, Michael Ryoo, Lifu Tu, Yihao Feng, Manli Shu, Honglu Zhou, Anas Awadalla, Jun Wang, et al. xgen-videosyn-1: High-fidelity text-to-video synthesis with compressed representations. arXiv preprint arXiv:2408.12590, 2024. 1, 2", + "[19] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1", + "[20] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1, 4", + "[21] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2", + "[22] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 5", + "[23] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5", + "[24] Haoze Wu, Jiawei Liu, Zheng-Jun Zha, Zhenzhong Chen, and Xiaoyan Sun. Mutually reinforced spatio-temporal convolutional tube for human action recognition. 5", + "[25] Haoze Wu, Zheng-Jun Zha, Xin Wen, Zhenzhong Chen, Dong Liu, and Xuejin Chen. Cross-fiber spatial-temporal co-enhanced networks for video action recognition. In Proceedings of the 27th ACM international conference on multimedia, pages 620-628, 2019.", + "[26] Haoze Wu, Jiawei Liu, Xierong Zhu, Meng Wang, and Zheng-Jun Zha. Multi-scale spatial-temporal integration convolutional tube for human action recognition. In Proceed-" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18132", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 753-759, 2021. 5", + "[27] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2364-2373, 2018. 5", + "[28] Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. Easyanimate: A high-performance long video generation method based on transformer architecture. arXiv preprint arXiv:2405.18991, 2024.3", + "[29] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 6", + "[30] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 1", + "[31] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 1, 3, 5", + "[32] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18456-18466, 2023. 1", + "[33] Ailing Zeng, Yuhang Yang, Weidong Chen, and Wei Liu. The dawn of video generation: Preliminary explorations with sora-like models. arXiv preprint arXiv:2410.05227, 2024. 3", + "[34] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Exploring figure-ground assignment mechanism in perceptual organization. Advances in Neural Information Processing Systems, 35:17030-17042, 2022. 5", + "[35] Wei Zhai, Yang Cao, Jing Zhang, Haiyong Xie, Dacheng Tao, and Zheng-Jun Zha. On exploring multiplicity of primitives and attributes for texture recognition in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):403-420, 2023.", + "[36] Wei Zhai, Pingyu Wu, Kai Zhu, Yang Cao, Feng Wu, and Zheng-Jun Zha. Background activation suppression for weakly supervised object localization and semantic segmentation. International Journal of Computer Vision, 132(3): 750-775, 2024. 5", + "[37] Sijie Zhao, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Muyao Niu, Xiaoyu Li, Wenbo Hu, and Ying Shan. Cv-vae: A compatible video vae for latent generative video models. arXiv preprint arXiv:2405.20279, 2024. 1, 2, 3, 6", + "[38] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "consistent diffusion model for real-world video superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2545, 2024. 3", + "bbox": [ + 545, + 90, + 906, + 147 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18133", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Improved Video VAE for Latent Video Diffusion Model/d6e6ed64-039e-4e0b-a0ce-69a2cf999362_model.json b/2025/Improved Video VAE for Latent Video Diffusion Model/d6e6ed64-039e-4e0b-a0ce-69a2cf999362_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d2e4e0ef447d8006e02d0aa5605e3adc81269343 --- /dev/null +++ b/2025/Improved Video VAE for Latent Video Diffusion Model/d6e6ed64-039e-4e0b-a0ce-69a2cf999362_model.json @@ -0,0 +1,1815 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.22, + 0.131, + 0.78, + 0.154 + ], + "angle": 0, + "content": "Improved Video VAE for Latent Video Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.18, + 0.856, + 0.253 + ], + "angle": 0, + "content": "Pingyu Wu\\(^{1}\\), Kai Zhu\\(^{1,*}\\), Yu Liu\\(^{2}\\), Liming Zhao\\(^{2}\\), Wei Zhai\\(^{1,*}\\), Yang Cao\\(^{1}\\), Zheng-Jun Zha\\(^{1}\\) \n\\(^{1}\\) MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China \\(^{2}\\) Independent Researcher {wpy364755620@mail., zkzy@mail., wzhai056@, forrest@, zhazj@}ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.33, + 0.304 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.319, + 0.486, + 0.712 + ], + "angle": 0, + "content": "Variational Autoencoder (VAE) aims to compress pixel data into low-dimensional latent space, playing an important role in OpenAI's Sora and other latent video diffusion generation models. While most existing video VAEs inflate a pre-trained image VAE into the 3D causal structure for temporal-spatial compression, this paper presents two astonishing findings: (1) The initialization from a well-trained image VAE with the same latent dimensions is not an optimal scheme. (2) The adoption of causal reasoning leads to unequal information interactions and unbalanced performance between frames. To alleviate these problems, we propose a keyframe-based temporal compression (KTC) architecture and a group causal convolution (GCCConv) module to further improve video VAE (IV-VAE). Specifically, the KTC architecture divides the latent space into two branches, in which one half completely inherits the compression prior of keyframes from a lower-dimension image VAE while the other half involves temporal compression through 3D group causal convolution, reducing temporal-spatial conflicts and accelerating the convergence speed of video VAE. The GCConv in the above 3D half uses standard convolution within each frame group to ensure inter-frame equivalence, and employs causal logical padding between groups to maintain flexibility in processing variable frame video. Extensive experiments on five benchmarks demonstrate the SOTA video reconstruction and generation abilities of our IV-VAE." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.224, + 0.759 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.485, + 0.877 + ], + "angle": 0, + "content": "In recent years, video generation has made significant advances in both academia and industry, showcasing cinematic-level visuals and the potential of a world simulator (e.g., Sora [16]). In particular, diffusion optimization methods on latent space have garnered widespread attention and dominated the video generation field due to their high efficiency and stability, resulting in a series of outstanding" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.288, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Latent Video Diffusion Models (LVDMs) [1, 10, 12, 32]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.906, + 0.427 + ], + "angle": 0, + "content": "To achieve a bidirectional mapping between the high-dimensional pixel space and the low-dimensional latent space, Variational Autoencoder (VAE) [13] is employed to encode the original video continuously into the latent space, along with spatial and even temporal compression. The compression rate and reconstruction quality of VAE directly determine the information effectiveness of the video corpus, playing an important role in the exploration of LVDMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.906, + 0.549 + ], + "angle": 0, + "content": "As the most commonly used video VAE, SVD-VAE [1] freezes the weights of the 2D encoder based on the image VAE [19] and retransmits a temporal decoder containing 3D convolutional blocks, which enables the feature interaction across different frames and significantly reduces flicker artifacts. Unfortunately, this method struggles to effectively eliminate temporal redundancy, limiting the LVDMs to perceive and generate long sequences of video content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.906, + 0.883 + ], + "angle": 0, + "content": "To address the above problem, several works [6, 10, 18, 29, 37] have proposed to construct a 3D VAE with both spatial and temporal compression capabilities, which can be summarized in the following three points. In terms of model structure, the UNet network [20] with strong pixel-perception capabilities is still adopted, consistent with the classic image VAE series (i.e., Stable Diffusion VAE). For the operator implementation, most of the methods [6, 10, 18, 29] adopt 3D causal convolution with only forward sequence correlation, which can encode the image information independently [31] while ensuring the continuity of the long sequence video encoding. During the optimization process, a well-trained 2D image VAE, focusing only on spatial compression, is used as an initialization for 3D video VAEs with the same structure and equivalent latent channel number [6, 10, 18, 37], which can achieve better results compared to training a temporal-spatial compressed video VAE from scratch [30]. Although existing methods achieve video reconstruction under \\(4 \\times\\) time compression, we find that there remain deficiencies in the above three key processes, limiting the improvement of the video quality and continuity in reconstruction results." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "First, initialization from a well-trained image VAE with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.398, + 0.901 + ], + "angle": 0, + "content": "*Kai Zhu and Wei Zhai are the corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "18124" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.13, + 0.085, + 0.607, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.278, + 0.478, + 0.292 + ], + "angle": 0, + "content": "(a) Number of latent channels" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.086, + 0.865, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.278, + 0.856, + 0.293 + ], + "angle": 0, + "content": "(b) frames within a frame group" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.298, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Figure 1. (a) Information preservation degree for different VAEs on video data. CR: Compression Rate. (b) Performance of different frames within a frame group on kinetics-600 dataset (OD-VAE [6], OS (OpenSora V1.2)-VAE [10], CogX (CogVideoX)-VAE [29])." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.339, + 0.485, + 0.732 + ], + "angle": 0, + "content": "the same latent channel number cannot support further \\(4\\times\\) time compression, leading to a sharp decline in the initial spatial compression performance and slow learning of subsequent temporal compression. As shown in Fig. 1(a), we calculate the reconstruction similarity of different VAE on video data as the information preservation degree (See Supplementary Material). We note that: (1) As the latent channel number \\(Z\\) increases, the spatial compression capacity gain produced by the image VAE gradually diminishes (blue dashed line) (2) The spatial compression capability provided by an image VAE with dimension \\(Z / 2\\) is sufficient for a temporal-spatial compressed video VAE with dimension \\(Z\\). Then the video VAE can use the remaining latent channels to store the spatial information of extra frames, i.e., acquiring the initial \\(2\\times\\) time compression capacity, which (red curves) converges even better and faster than the regular initialization approach [6, 10, 18, 37] that inherits the image VAE weights of the same dimension \\(Z\\) (green curves). This inspires us to start from an image prior with a lower latent channel number and leave the remaining latent channels for temporal compression, promoting a more balanced temporal-spatial compression initialization. To this end, we propose a dual-branch keyframe-based temporal compression architecture, where a 2D branch inheriting the low-dimensional image VAE prior focuses on keyframe compression, and a 3D branch facilitates temporal compression." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Second, the unidirectional information flow of 3D causal convolution prevents the interaction between video frames, exacerbating inter-frame unbalanced performance under time compression. In the forward process, \\( M \\) frames (a frame group) are compressed to a whole (one frame) and then restored to \\( M \\) frames. Due to the restriction of forward flow of causal convolution, the first frame in the frame group is unable to perceive the other frames until after time compression in the middle layer of the network. Thus the first frame has the poorest ability to interact with the whole frame group, resulting in higher optimization diffi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.907, + 0.519 + ], + "angle": 0, + "content": "culty compared to other frames. As shown in Fig. 1(b), the methods [6, 10, 29] based on causal convolution suffers from poor performance of the first frame in the frame group and large fluctuations in SSIM values between frames, leading to significant flickering in the reconstruction results. This inspires us to propose group causal convolution, which groups the input frames based on the temporal compression rate. Causal logic padding operations are applied between frame groups to ensure the continuity in processing variable frame videos, and standard convolution operations are used within each group of \\(M\\) frames to share equivalent information interactions, achieving smoother reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.909, + 0.642 + ], + "angle": 0, + "content": "In addition, the local receptive field struggles to meet the temporal compression demand at high resolutions. For the same video, the higher the resolution, the more pixels the same motion spans, making it more challenging for the model to capture motion patterns. To this end, we introduce atrous convolutions and add multiple attention modules to expand the receptive field, thus improving the temporal motion perception of the model at high resolutions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.884 + ], + "angle": 0, + "content": "Finally, considering the increasing demand for content diversity and high resolution in the video generation field, there is a lack of a suitable open-source benchmark for comprehensive evaluation of video VAE reconstruction capabilities. Most of the datasets (e.g., Kinetics-600 [4] and UCF-101 [21]) are 720P or lower, and cannot be used for high resolution (e.g., 1080P) evaluations. Moreover, high-resolution datasets (e.g., Panda-70M [8]) consist mainly of slow-motion and fixed-shot videos, which hardly reflects the overall performance of video VAEs under varying motion speeds. Therefore, we re-collect a subset of 2000 1080P videos from OpenVid-0.4M [15] dataset, uniformly containing various motion speeds, named MotionHD, as a supplement to the overall evaluation. Experimental results show that our method achieves state-of-the-art video reconstruction at various resolutions and motion rates." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In summary, the contributions of this paper include: (1)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18125" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "We propose a keyframe-based temporal compression architecture to accelerate convergence speed by providing a more balanced and powerful temporal-spatial compression initialization. (2) We introduce group causal convolution as a replacement for causal convolution to achieve better and balanced performance between frames, while maintaining causal logic. (3) Extensive experiments on five benchmarks demonstrate the SOTA video reconstruction and generation capabilities of the proposed IV-VAE." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.242, + 0.228, + 0.256 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.483, + 0.447 + ], + "angle": 0, + "content": "Variational Autoencoder. Variational Autoencoder (VAE), as a widely used model in the generation domain, is adept at learning complex distributions from high-dimensional data. VAEs can be categorized into two types based on the class of tokens, including discrete and continuous. Classical work of the former type includes MaskGiT [5], VQGAN [9], etc., where continuous latent features are discretized by quantization techniques for autoregression generation. The latter type of continuous VAE, which can achieve spatial or even temporal compression by mapping high-dimensional data distributions to a low-dimensional latent space, is widely used in latent diffusion models [33]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.483, + 0.781 + ], + "angle": 0, + "content": "Video VAE. Earlier video generation works [1, 28, 38] usually use a image VAE to compress video frames, or fine-tune the image VAE with additional time-dimensional convolutions to reduce flickering artifacts, but this approach cannot effectively reduce the redundancy in the time dimension. With the emergence of OpenAI's Sora [16], to allow video generation models to learn and generate longer videos, there have been some recent works [6, 10, 29, 37] that attempt to train a video VAE with a larger time compression ratios (e.g., \\(4 \\times\\)). Open-Sora [10] adopts a stacked approach, first compressing the video frames in the spatial dimension based on SDXL image VAE [17], and then introducing 3D VAE to compress the temporal dimension. ODVAE [6] uses tail initialization to inflate the 2D convolutions of SD image VAE into 3D causal convolutions and explores an efficient model variant. CV-VAE [37] uses latent space regularization to avoid latent space distribution shift while compressing time. CogVideoX [29] constructs a multi-GPU parallel computing approach to process long videos for causal video VAE. Based on the general idea of these methods, we further explore the shortcomings of the existing video VAEs and propose a series of improvements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.794, + 0.182, + 0.809 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.315, + 0.833 + ], + "angle": 0, + "content": "3.1. Baseline 3D Causal VAE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In the beginning, we introduce a causal video VAE as the baseline method. We first train an image VAE based on the architecture of SD VAE with a few structural modifications (Sec. 3.5). Following the thought of MagViT-v2 [31]," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "we choose 3D causal convolutions in the construction of video VAE. Then, drawing on the idea of central initialization [3], we inflate the 2D convolution in image VAE to the 3D causal convolution. For example, noting 2D convolutional weights as \\( W_{2d} \\in \\mathbb{R}^{c_o \\times c_i \\times k_h \\times k_w} \\), when inflating \\( W_{2d} \\) to be \\( W_{3d} \\in \\mathbb{R}^{c_o \\times c_i \\times k_t \\times k_h \\times k_w} \\), \\( W_{3d} \\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.185, + 0.905, + 0.218 + ], + "angle": 0, + "content": "\\[\nW _ {3 d} [:,: i,:,: ] = \\left\\{ \\begin{array}{c c} W _ {2 d} & , \\text {i f} i = - 1 \\\\ 0 & , \\text {e l s e} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.221, + 0.905, + 0.295 + ], + "angle": 0, + "content": "(See supplementary material for details). In this way, we can obtain high-quality image reconstruction results after extending image VAE to a causal video VAE. Lastly, except that the first image frame is processed separately, we let the causal video VAE achieve \\(4 \\times\\) temporal compression." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.907, + 0.463 + ], + "angle": 0, + "content": "Adopting a causal VAE as the baseline has the following two advantages: (1) The first frame is always independent of other frames. According to MagViT-v2 [31], using a temporal causal VAE instead of a regular VAE makes the model perform better in processing a single image. (2) Each frame depends only on the previous frames. Thus, a temporal causal video VAE can use cache mechanism in the inference, which splits the reconstruction process of a long video into a series of short videos with exactly the same result. We consider this a better operation compared to regular leap approach, more details are displayed in Sec. 4.5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.471, + 0.757, + 0.487 + ], + "angle": 0, + "content": "3.2. Group Causal Convolution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Based on the above baseline, we note that it is practically impossible and non-optimal to ensure causal logic for all frames. For example, the baseline approach compresses time by \\(4 \\times\\), then the input 4 frames will be compressed into 1 frame by the encoder and recovered to 4 frames through the decoder. Obviously, these 4 frames recovered from the same compressed frame do not satisfy temporal causal correlation. In addition, for these 4 reconstructed frames, the first frame usually does not perform as well as subsequent frames, as shown in Fig. 1(b), resulting in frequent flickering in the reconstructed video. We attribute this problem to the insufficient and unbalanced interaction between neighboring frames due to the causal convolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Therefore, to alleviate this problem while being able to maintain the advantages of causal VAE (Sec. 3.1), we propose group causal convolution (GCCov). For a video VAE with a temporal compression rate of \\( t_c \\) (\\( t_c = 4 \\) in our method), the proposed method will initially label every \\( t_c \\) frames of the input video as a frame group, with no overlap between the frame groups. Specifically, the first image frame is a separate frame group. For a frame group, we aim to make the features between frames visible and interactable with each other, thus increasing the reconstruction smoothness between frames and reducing the flicker. For different frame groups, we maintain the property of causal convolution that the current frame group only depends on the previous frame groups. Therefore, the time padding scheme" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18126" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.087, + 0.42, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.087, + 0.71, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.087, + 0.891, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.295, + 0.713, + 0.309 + ], + "angle": 0, + "content": "Figure 2. (a-c) Basic components of IV-VAE. (d) Overall Framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.484, + 0.406 + ], + "angle": 0, + "content": "for group causal convolution is designed in Fig. 2(a). We pad the features of the previous frame before each frame group, in particular, the first image frame is padded using its own features. And we use zero padding after each frame group. After causal logic padding, each group is operated separately using the same regular convolution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.422, + 0.428, + 0.437 + ], + "angle": 0, + "content": "3.3. Keyframe-based temporal compression" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.445, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Building on the observation from Fig. 1(a), to obtain better temporal-spatial compression capability at the initialization of video VAE with latent channel number \\( Z \\), it is feasible to utilize the pre-trained image VAE weights with a low latent channel number, i.e., \\( Z / 2 \\). To this end, we propose a dual-branch keyframe-based temporal compression (KTC) architecture that decouples the compression of a frame group into the spatial compression of the keyframe and the temporal compression of the overall motion, where the first frame of each frame group is regarded as the keyframe since it can be perceived by the other frames along the temporal forward flow. The basic unit is shown in Fig. 2(b). For the input feature map \\( F \\) with \\( C_{in} \\) channels, it is fed into a 2D convolution to extract the image information and a 3D group causal convolution to extract the overall temporal motion information, respectively. The two convolutions each output feature maps with channels \\( C_{out} / 2 \\), which are then separately normalized by RMSNorm and concatenated together." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In the initialization phase, the 2D and 3D branches of the network are initialized using the pre-trained image VAE weights with latent channel number \\( Z / 2 \\) respectively. In the 3D branch, the 2D image weights are inflated to 3D weights based on the center initialization [3]. In addition, we elaborately set the initial weights of the newly added temporal upsampling and downsampling layers of the two branches. For temporal downsampling layer, we make it initially behave as simply retaining all features of one of the \\( t_c \\) frames and then discarding the remaining frames. For temporal upsampling layer, we let it initially behave as an operation of copying existing frames (More details are provided in Sup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.905, + 0.389 + ], + "angle": 0, + "content": "plementary Material). In this way, each branch is initially enabled to reconstruct one frame from the input \\( t_c \\) frames. With two branches reconstructing different frames, the proposed KTC structure initially realizes \\( 2 \\times \\) time compression with strong image compression capability as well." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.906, + 0.556 + ], + "angle": 0, + "content": "The overall structure of the network is based on UNet [20] and adopts the proposed KTC as the base unit, as shown in Fig. 2(d). After two temporal downsampling, a frame group is merged into one compressed frame, so we replace the GCConv3D in the KTC Unit with a 2D convolution at this stage to reduce the computation. In the output of the model, each of the two branches produces a sequence of video frames with the same shape as the input, where the reconstruction results of keyframes and remaining frames are selected from the 2D branch and 3D branch outputs respectively, and compose the final reconstruction result." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.563, + 0.882, + 0.579 + ], + "angle": 0, + "content": "3.4. Temporal Motion Perception Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.907, + 0.844 + ], + "angle": 0, + "content": "For the same video, as the video resolution increases, the pixel span generated by object motion or camera motion increases, and the more difficult it is for the video VAE to capture the motion patterns due to the limitations of the receptive field. To this end, we make two improvements to increase the receptive field of the video VAE to enhance the perception of temporal motion at high resolution. As shown in Fig. 2(c), we introduce multiple parallel atrous convolutions (PAC) with different atrous rates in the last layer of the encoder, borrowing from the ASPP in DeepLabV3+ [7]. In addition, we expand the number of attention modules from 2 to 7, and all attention modules are implemented after full temporal and spatial compression to reduce the computational effort. At this time, the compressed frame contains all the information of the frame group, and the global receptive field of attention modules can effectively facilitate temporal motion perception, especially at high resolution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.785, + 0.863 + ], + "angle": 0, + "content": "3.5. Other structural modifications" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Based on the architecture of SD image VAE, we make some modifications to better adapt to 3D causal VAE. First, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18127" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.085, + 0.898, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.3, + 0.725, + 0.314 + ], + "angle": 0, + "content": "Figure 3. Reconstruction results of different methods for a frame group." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.321, + 0.385, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.322, + 0.637, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.322, + 0.768, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.322, + 0.899, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.181, + 0.455, + 0.816, + 0.47 + ], + "angle": 0, + "content": "Figure 4. Video motion distribution of Kinetics-600, OpenVid-0.4M, and the filtered MotionHD datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.483, + 0.584 + ], + "angle": 0, + "content": "replace all GroupNorms with RMSNorms to ensure temporal causality following MagViT-v2 [31], since the standard GroupNorm involves time dimension when calculating the mean and variance of the feature maps, which destroys temporal causality. Second, we shrink the convolution channels during spatial upsampling instead of in the later residual block, reducing memory usage by \\(29\\%\\) in the inference." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.597, + 0.224, + 0.614 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.622, + 0.271, + 0.637 + ], + "angle": 0, + "content": "4.1. MotionHD dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To better evaluate video VAEs, we propose to consider the video reconstruction performance at different motion amplitudes and resolutions. Among the open-source video datasets, OpenVid-0.4M [15] is a recently released high-quality dataset consisting of 1080P resolution videos, but most of them are fixed camera or slow-moving. To present this phenomenon, we use the RAFT [22] model to compute the optical flow map of video at \\(768 \\times 432\\) resolution and then get the average optical flow score as the motion score. As shown in Fig. 4, we present the motion distributions of Kinetics-600 [4] and OpenVid-0.4M datasets. Compared to Kinetics-600, the motion distribution of OpenVid-0.4M is overall narrower and concentrated in low-motion intervals, which is difficult to reflect the reconstruction performance at medium and fast motion. More importantly, in the training of diffusion models, it is common to use a high frame interval to facilitate the learning of object motion, e.g.," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.905, + 0.507 + ], + "angle": 0, + "content": "Latte [14] set the frame interval to 3 on all datasets. In this case, the motion of video will be more intense (Fig. 4(a))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Therefore, we refer to the motion distribution of the Kinetics-600 (frame interval of 3), and uniformly select 2000 videos with 1080P resolution from OpenVid-0.4M according to the motion scores, as illustrated in Fig. 4(b), constituting the subset MotionHD, which is used to test the average performance of video VAEs at various video motion speeds and the performance at different resolutions. We also list the motion scores of the videos generated by Sora [16] in Fig. 4(c), to facilitate the interpretation of scores and to indicate that our choice of motion intervals is reasonable." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.704, + 0.684 + ], + "angle": 0, + "content": "4.2. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Evaluation details. We evaluate the reconstruction performance of video VAEs on multiple video benchmarks, including widely used [24-26, 34-36] Kinetics-600 [4] val set, ActivityNet [2] test set, and MotionHD. For Kinetics-600 and ActivityNet, we each randomly select 2400 videos and test them with the original resolution. MotionHD is used to test the performance of Video VAEs at various resolutions. All experiments are conducted using the first 17 frames of the video. To measure reconstruction quality, SSIM, PSNR, LPIPS, and FVD [23] are adopted as metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Following OD-VAE [6], we employ Latte [14], a soralike transformer-based diffusion model for video generation. We choose Kinetics-400 [11] for class-conditional generation and SkyTimelapse [27] for unconditional gener" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18128" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.086, + 0.9, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.398, + 0.908, + 0.427 + ], + "angle": 0, + "content": "Figure 5. Reconstruction results for two adjacent frames in a frame group. In the below figure, the regions with large differences between the two reconstructed frames are boxed out." + }, + { + "type": "table", + "bbox": [ + 0.156, + 0.437, + 0.846, + 0.627 + ], + "angle": 0, + "content": "
MethodParamsFCRChnKinetics-600ActivityNet
FVD↓PSNR↑SSIM↑LPIPS↓FVD↓PSNR↑SSIM↑LPIPS↓
SVD-VAE97M1*8*846.2635.260.93630.036755.4737.070.92980.03217
CV-VAE182M4*8*8416.8131.580.90660.0827112.5734.870.91560.06846
OS-VAE393M4*8*8419.0534.370.92580.0816213.5637.640.93400.06781
OD-VAE239M4*8*8410.6933.880.92420.054858.1036.950.93210.04670
Causal VAE127M4*8*8410.8133.770.92480.063098.2637.060.93260.05336
IV-VAE107M4*8*848.0134.290.92810.052096.0838.470.93590.04436
Causal VAE127M4*8*885.7236.080.94790.040334.3938.360.95190.03458
IV-VAE107M4*8*883.9436.660.95200.035993.5039.610.95660.03072
CogX-VAE215M4*8*8163.1738.380.96770.028662.1740.680.97030.02468
Causal VAE127M4*8*8163.2838.070.96410.028872.4340.420.96710.02483
IV-VAE108M4*8*8162.9739.020.96850.022802.0142.610.97220.01968
" + }, + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.629, + 0.843, + 0.644 + ], + "angle": 0, + "content": "Table 1. Comparison of video reconstruction results. Chn: Latent channel number. FCR: Frame compression rate." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.735 + ], + "angle": 0, + "content": "ation. All methods are trained for 150K steps with the same default settings in Latte [14]. In addition, the parameters of the video are set to a resolution of \\(256 \\times 256\\) and a frame length of 17 during both training and testing. We employ FVD as the metric and calculate it on 2048 samples." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Training details. We first train an image VAE on \\(256 \\times 256\\) image data for 200K steps using 4 A800 GPUs, and then inflate it to a video VAE. During training, KL [13], MAE, and LPIPS are employed as losses. Next, we train the video VAE on \\(256 \\times 256\\) resolution videos for 500K steps, and further expand the training resolution to \\(512 \\times 512\\) training 200K steps. Finally, we train the VAE at different resolutions with different frame numbers for 100K steps. GAN loss from the 3D discriminator is added at this stage. For causal VAE, we set the basic channel number of the convolution to 96. For IV-VAE, we set it to 64 for both branches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.659, + 0.651, + 0.673 + ], + "angle": 0, + "content": "4.3. Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.907, + 0.731 + ], + "angle": 0, + "content": "In addition to baseline Causal VAE, comparison methods include OpenSora V1.2 (OS)-VAE [10], OD-VAE [6], CV-VAE [37],CogVideoX (CogX)-VAE [29], SVD-VAE [1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Quantitative evaluation of reconstruction results. Table 1 demonstrates the performance of the proposed Causal VAE and IV-VAE with other methods on Kinetics-600 and ActivityNet datasets. Experiments show that the proposed method achieves excellent results on both datasets and various numbers of latent channels \\( Z \\). On \\( Z = 4 \\), we outperform OD-VAE by 2.68 and 2.02 FVD reductions in Kinetics-600 and ActivityNet respectively with less than half parameters. Compared to OS-VAE, we are slightly lower on the PSNR metric due to the limitation of parameters, but we obtain a reduction of 286M (-73%) param" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18129" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.14, + 0.089, + 0.86, + 0.28 + ], + "angle": 0, + "content": "
MethodChnFVD↓PSNR↑SSIM↑LPIPS↓
480P720P1080P480P720P1080P480P720P1080P480P720P1080P
SVD-VAE45.684.233.6233.7136.0037.690.93130.95070.96170.0350.0330.031
CV-VAE422.3312.788.5029.7030.2931.190.87960.90080.92060.0940.0980.098
OS-VAE416.8910.068.2132.3334.1335.650.91090.93100.94520.0860.0860.087
OD-VAE410.407.365.3131.4233.3234.890.90040.92260.93910.0590.0610.064
Causal VAE49.336.855.1831.6433.4334.820.90870.92860.94150.0670.0690.073
IV-VAE47.014.494.3132.0134.0835.800.91270.93330.94650.0530.0540.055
Causal VAE84.634.153.5834.1435.8937.150.93550.94810.95570.0430.0470.051
IV-VAE84.093.712.8634.6036.5638.110.94100.95340.96110.0370.0390.041
CogX-VAE163.503.662.2636.5237.8638.680.95710.96390.96760.0310.0350.039
Causal VAE163.763.852.7236.2537.6738.600.95440.96170.96630.0320.0370.041
IV-VAE162.582.782.1836.8038.8140.330.95850.96690.97200.0250.0270.027
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.285, + 0.899, + 0.298 + ], + "angle": 0, + "content": "Table 2. Video reconstruction results at different resolutions on MotionHD. 480P: \\( {848} \\times {480},{720}\\mathrm{P} : {1280} \\times {720},{1080}\\mathrm{P} : {1920} \\times {1280} \\) ." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.311, + 0.482, + 0.402 + ], + "angle": 0, + "content": "eters, 11.04 \\((-58\\%)\\) FVD, and 0.02953 \\((-36\\%)\\) LPIPS on Kinetics-600. Compared to SVD-VAE, although our model with \\(Z = 8\\) uses a latent channel number twice as large, we achieve an additional \\(4\\times\\) time compression and perform better on all metrics. In addition, we also obtain the best results on \\(Z = 16\\), especially in PSNR and LPIPS metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Qualitative evaluation of reconstruction results. As shown in Fig. 3 and 5, we exhibit the reconstruction results for different frames within a frame group. It can be observed that other methods usually perform poorly in the first two frames of a frame group, with blurring and a small amount of distortion, and the last frame performs better. This phenomenon validates our point that for a video VAE with temporal compression, the use of causal convolution makes the frames in the same frame group unequal with respect to the information interaction, leading to performance imbalance. While our method can achieve a more balanced performance benefiting from group causal convolution. Besides, due to the performance fluctuations, other methods usually suffer from abrupt changes in details in the reconstruction of neighboring frames, as shown in the lower panel of Fig. 5, and our method performs better in terms of temporal consistency and reconstruction quality. More visual comparisons are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Reconstruction performance at different resolutions. In Table 2, we present the performance of various methods at different resolutions on the MotionHD dataset. The proposed method achieves the best results in the majority of metrics. And as the resolution increases, our improvement relative to other methods is more pronounced. For example, in the PSNR metric, we outperform CogX-VAE by 0.28 at 480P while when the resolution is 1080P, this improvement increases to 1.65. Experiments indicate that the proposed KTC module and TMPE can enable the video VAE to achieve better temporal compression capability, leading to better results in challenging scenarios of temporal compression such as high resolution and rapid motion." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Video Generation. In the table of Fig. 6, we present the video generation results of different video VAEs using" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.312, + 0.903, + 0.362 + ], + "angle": 0, + "content": "
DatasetCV-VAEOS-VAEOD-VAEIV-VAE (Z=4)
Kinetics-400552.4617.6548.7541.6
SkyTimelapse128.0141.3140.0118.6
" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.364, + 0.903, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.43, + 0.905, + 0.459 + ], + "angle": 0, + "content": "Figure 6. Top Table: FVD for different methods. Bottom Figure: Video frames generated by Latte with our IV-VAE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.906, + 0.622 + ], + "angle": 0, + "content": "the same generation model Latte-XL (Params: 0.67B) [14]. The proposed VAE achieves the best performance on both class-conditional and unconditional generation. Especially on SkyTimelapse, we obtain a 9.4 FVD decrease compared to CV-VAE. The results indicate that IV-VAE can be well employed for LVDMs. At the bottom of Fig. 6, we show some frames of the video generated by Latte with our VAE. The frames exhibit a fairly high level of quality and realism, and even reflect some of the laws of physics, as the sunlight in the sky moves, so does the light reflected in the water." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.635, + 0.667, + 0.652 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.764 + ], + "angle": 0, + "content": "In this section, we perform a series of ablation experiments. Considering that ablation for three channels requires too much time and resources, we choose the middle channel \\( Z = 8 \\) for the ablation of IV-VAE composition. Each ablation experiment follows the same training program: We first train an image VAE for 200K steps, and then extend it to a video VAE and train it for 300K steps on \\( 256 \\times 256 \\) videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Ablation studies of main components. Based on the baseline causal VAE, we explore the effects of the different compositions of IV-VAE in Table 3. As can be seen from the table, the implementation of the proposed group causal convolution can substantially improve the performance on both baseline and KTC structure with the parameter count unchanged. The results indicate that the reconstruction quality can be effectively improved by allowing the frames in the same frame group to interact bidirectionally. Further," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18130" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.09, + 0.483, + 0.194 + ], + "angle": 0, + "content": "
SettingTest: 256×256Test: 512×512Params
GCCovKTCTMPEPSNR↑SSIM↑PSNR↑SSIM↑
(A)31.290.904234.880.9384127M
(B)✓31.640.908235.210.9420127M
(C)✓31.860.911635.460.9451104M
(D)✓✓32.120.914535.690.9478104M
(E)✓✓✓32.240.915835.850.9493107M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.196, + 0.483, + 0.224 + ], + "angle": 0, + "content": "Table 3. Ablation study. (A) represents baseline causal VAE. The models are tested on Kinetics-600 using 256 and 512 resolutions." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.231, + 0.462, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.374, + 0.483, + 0.403 + ], + "angle": 0, + "content": "Figure 7. Ablation study of structures. The training loss is the sum of MAE, LPIPS, and KL losses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.55 + ], + "angle": 0, + "content": "replacing the base unit with KTC unit can substantially facilitate the learning of temporal compression thus improving the performance of the model, even with smaller parameters. Finally, we add TMPE enabling the model to better learn the object's motion patterns by increasing the global receptive field with only 3M additional parameter counts. From the improvement of the results at different resolutions, it can be analyzed that this motion perception enhancement plays a more important role in higher resolution scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Effectiveness of KTC at different latent channel numbers. To verify the validity of the KTC structure, we only replace the structure of the baseline causal VAE with the proposed KTC and perform experiments using the same training setup, the training loss is shown in Fig 7. When the latent channel number \\( Z = 4 \\), the proposed KTC performs slightly superior to the baseline as the training steps increase. And this advantage extends significantly when \\( Z \\) expands to 8 and 16. In conjunction with the results presented in Fig. 1, it can be analyzed that the spatial compression capability improvement of the image VAE is attenuated as \\( Z \\) increases. Therefore, when \\( Z \\) is large, dedicating a portion of the number of latent channels to the initial temporal compression can achieve better and balanced temporal-spatial compression capability initially, leading to faster convergence in subsequent training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.194, + 0.819 + ], + "angle": 0, + "content": "4.5. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Increase parameters. In Table 4, we explore the boost provided by increasing the parameters. Increasing the basic feature dimension of IV-VAE from 64 to 96 dramatically increases the parameters and also brings significant improvements, particularly in PSNR and LPIPS metrics. Compared" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.165 + ], + "angle": 0, + "content": "
Method (Z=16)ParamsKinetics-60025×1024×1024 Video Rec. time
PSNR↑SSIM↑LPIPS↓
CogX-VAE215M38.380.96770.028711.4s
Ours (dim=64)108M39.020.96850.02286.2s
Ours (dim=96)242M39.570.97110.020510.7s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.167, + 0.905, + 0.194 + ], + "angle": 0, + "content": "Table 4. Increasing parameters. The reconstruction time is tested on one A800 GPU with BF16 precision. Rec.: Reconstruction." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.201, + 0.905, + 0.277 + ], + "angle": 0, + "content": "
Rec. approachInput: A 49-frame 1024×1024 VideoKinetics-600
Video clippingMem.Rec. timeSSIM PSNR
One-Step49-frame76.3G12.7s0.9685 39.02
Overleap5-frame*1214.3G14.7s0.9682 38.97
Cache1-frame + 4-frame*1214.1G13.0s0.9685 39.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.28, + 0.905, + 0.321 + ], + "angle": 0, + "content": "Table 5. Cache mechanism vs Overleap mechanism. The overlap mechanism cannot produce the same results as one-step reconstruction. The results are evaluated on \\( Z = 16 \\). Mem.: Memory." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.905, + 0.363 + ], + "angle": 0, + "content": "to CogX-VAE, this model achieves a 1.19 PSNR improvement and still takes less time to reconstruct the video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.514 + ], + "angle": 0, + "content": "Cache mechanism vs Overleap mechanism. To reconstruct a long video, it is necessary to split the video into a series of clips since GPU memory is limited. Overlap is the most commonly used approach [6], where each input video clip has a one-frame overlap with the previous clip as a history frame, and the overlapped frame is discarded after reconstruction. However, reconstruction with overlap is not equivalent to one-step reconstruction, which may introduce semantic bias in latent space due to the short history frame. And the overlapping frames lead to extra computation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.907, + 0.682 + ], + "angle": 0, + "content": "To address this problem, we introduce the cache mechanism, which is based on the property that the causal network depends only on the previous frames in temporal order. According to the time padding length \\( P_{t} \\) needed for 3D convolution, the last \\( P_{t} \\) frames of the current video clip's feature maps at each layer of the network are saved, which are used for the convolutional padding of the next video clip. Therefore, this mechanism can be fully equivalent to one-step reconstruction in terms of results. In addition, as shown in Table 5, the caching mechanism requires less time usage and GPU memory usage compared to the overlap mechanism." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we observe that regular initialization of utilizing image VAE weights with the same latent channel number is not an optimal scheme, and the use of causal convolution leads to performance imbalance between frames. To address these problems, we propose a keyframe-based temporal compression architecture to accelerate convergence speed by providing more balanced and powerful temporal-spatial compression initialization. In addition, group causal convolution is introduced to achieve better and more balanced performance between frames. Extensive experiments show that IV-VAE achieves SOTA results in both video reconstruction and generation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18131" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "Acknowledgments. This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62306295, 62225207 and 62436008." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.149, + 0.188, + 0.166 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.174, + 0.483, + 0.243 + ], + "angle": 0, + "content": "[1] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.245, + 0.483, + 0.313 + ], + "angle": 0, + "content": "[2] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.314, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.371, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[4] Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short note about kinetics-600. arXiv preprint arXiv:1808.01340, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.413, + 0.483, + 0.467 + ], + "angle": 0, + "content": "[5] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.484, + 0.537 + ], + "angle": 0, + "content": "[6] Liuhan Chen, Zongjian Li, Bin Lin, Bin Zhu, Qian Wang, Shenghai Yuan, Xing Zhou, Xinghua Cheng, and Li Yuan. Od-vae: An omni-dimensional video compressor for improving latent video diffusion model. arXiv preprint arXiv:2409.01199, 2024. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.484, + 0.607 + ], + "angle": 0, + "content": "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[8] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[9] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[10] hpcaitech. Open-sora. https://github.com/hpcaitech/Open-Sora, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[11] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Kihong Kim, Haneol Lee, Jihye Park, Seyeon Kim, Kwanghee Lee, Seungryong Kim, and Jaejun Yoo. Hybrid video diffusion models with 2d triplane and 3d wavelet representation. In European Conference on Computer Vision, pages 148-165. Springer, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.174, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "[13] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[14] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[15] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.904, + 0.276 + ], + "angle": 0, + "content": "[16] OpenAI. Openai sora. Accessed May 15, 2024 [Online], 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[17] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[18] Can Qin, Congying Xia, Krithika Ramakrishnan, Michael Ryoo, Lifu Tu, Yihao Feng, Manli Shu, Honglu Zhou, Anas Awadalla, Jun Wang, et al. xgen-videosyn-1: High-fidelity text-to-video synthesis with compressed representations. arXiv preprint arXiv:2408.12590, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.49 + ], + "angle": 0, + "content": "[19] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[20] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.904, + 0.617 + ], + "angle": 0, + "content": "[21] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[22] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[23] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[24] Haoze Wu, Jiawei Liu, Zheng-Jun Zha, Zhenzhong Chen, and Xiaoyan Sun. Mutually reinforced spatio-temporal convolutional tube for human action recognition. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[25] Haoze Wu, Zheng-Jun Zha, Xin Wen, Zhenzhong Chen, Dong Liu, and Xuejin Chen. Cross-fiber spatial-temporal co-enhanced networks for video action recognition. In Proceedings of the 27th ACM international conference on multimedia, pages 620-628, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Haoze Wu, Jiawei Liu, Xierong Zhu, Meng Wang, and Zheng-Jun Zha. Multi-scale spatial-temporal integration convolutional tube for human action recognition. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18132" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "ings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 753-759, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[27] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2364-2373, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[28] Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. Easyanimate: A high-performance long video generation method based on transformer architecture. arXiv preprint arXiv:2405.18991, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[29] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.431 + ], + "angle": 0, + "content": "[30] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[31] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.484, + 0.572 + ], + "angle": 0, + "content": "[32] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18456-18466, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[33] Ailing Zeng, Yuhang Yang, Weidong Chen, and Wei Liu. The dawn of video generation: Preliminary explorations with sora-like models. arXiv preprint arXiv:2410.05227, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.484, + 0.673 + ], + "angle": 0, + "content": "[34] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Exploring figure-ground assignment mechanism in perceptual organization. Advances in Neural Information Processing Systems, 35:17030-17042, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[35] Wei Zhai, Yang Cao, Jing Zhang, Haiyong Xie, Dacheng Tao, and Zheng-Jun Zha. On exploring multiplicity of primitives and attributes for texture recognition in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):403-420, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.484, + 0.813 + ], + "angle": 0, + "content": "[36] Wei Zhai, Pingyu Wu, Kai Zhu, Yang Cao, Feng Wu, and Zheng-Jun Zha. Background activation suppression for weakly supervised object localization and semantic segmentation. International Journal of Computer Vision, 132(3): 750-775, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[37] Sijie Zhao, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Muyao Niu, Xiaoyu Li, Wenbo Hu, and Ying Shan. Cv-vae: A compatible video vae for latent generative video models. arXiv preprint arXiv:2405.20279, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[38] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.092, + 0.907, + 0.148 + ], + "angle": 0, + "content": "consistent diffusion model for real-world video superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2545, 2024. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18133" + } + ] +] \ No newline at end of file diff --git a/2025/Improved Video VAE for Latent Video Diffusion Model/d6e6ed64-039e-4e0b-a0ce-69a2cf999362_origin.pdf b/2025/Improved Video VAE for Latent Video Diffusion Model/d6e6ed64-039e-4e0b-a0ce-69a2cf999362_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7807ae571898273299acaed7ef2c9de2ddd001e4 --- /dev/null +++ b/2025/Improved Video VAE for Latent Video Diffusion Model/d6e6ed64-039e-4e0b-a0ce-69a2cf999362_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:816bddf1a6a8906a3376065d874a6e5a0856c1584a59cac9eeb37dabfe9737f7 +size 6006244 diff --git a/2025/Improved Video VAE for Latent Video Diffusion Model/full.md b/2025/Improved Video VAE for Latent Video Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..987d14afd575ca695dffdcefb505bb95cd7cfeb1 --- /dev/null +++ b/2025/Improved Video VAE for Latent Video Diffusion Model/full.md @@ -0,0 +1,250 @@ +# Improved Video VAE for Latent Video Diffusion Model + +Pingyu Wu $^{1}$ , Kai Zhu $^{1,*}$ , Yu Liu $^{2}$ , Liming Zhao $^{2}$ , Wei Zhai $^{1,*}$ , Yang Cao $^{1}$ , Zheng-Jun Zha $^{1}$ $^{1}$ MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China $^{2}$ Independent Researcher {wpy364755620@mail., zkzy@mail., wzhai056@, forrest@, zhazj@}ustc.edu.cn + +# Abstract + +Variational Autoencoder (VAE) aims to compress pixel data into low-dimensional latent space, playing an important role in OpenAI's Sora and other latent video diffusion generation models. While most existing video VAEs inflate a pre-trained image VAE into the 3D causal structure for temporal-spatial compression, this paper presents two astonishing findings: (1) The initialization from a well-trained image VAE with the same latent dimensions is not an optimal scheme. (2) The adoption of causal reasoning leads to unequal information interactions and unbalanced performance between frames. To alleviate these problems, we propose a keyframe-based temporal compression (KTC) architecture and a group causal convolution (GCCConv) module to further improve video VAE (IV-VAE). Specifically, the KTC architecture divides the latent space into two branches, in which one half completely inherits the compression prior of keyframes from a lower-dimension image VAE while the other half involves temporal compression through 3D group causal convolution, reducing temporal-spatial conflicts and accelerating the convergence speed of video VAE. The GCConv in the above 3D half uses standard convolution within each frame group to ensure inter-frame equivalence, and employs causal logical padding between groups to maintain flexibility in processing variable frame video. Extensive experiments on five benchmarks demonstrate the SOTA video reconstruction and generation abilities of our IV-VAE. + +# 1. Introduction + +In recent years, video generation has made significant advances in both academia and industry, showcasing cinematic-level visuals and the potential of a world simulator (e.g., Sora [16]). In particular, diffusion optimization methods on latent space have garnered widespread attention and dominated the video generation field due to their high efficiency and stability, resulting in a series of outstanding + +Latent Video Diffusion Models (LVDMs) [1, 10, 12, 32]. + +To achieve a bidirectional mapping between the high-dimensional pixel space and the low-dimensional latent space, Variational Autoencoder (VAE) [13] is employed to encode the original video continuously into the latent space, along with spatial and even temporal compression. The compression rate and reconstruction quality of VAE directly determine the information effectiveness of the video corpus, playing an important role in the exploration of LVDMs. + +As the most commonly used video VAE, SVD-VAE [1] freezes the weights of the 2D encoder based on the image VAE [19] and retransmits a temporal decoder containing 3D convolutional blocks, which enables the feature interaction across different frames and significantly reduces flicker artifacts. Unfortunately, this method struggles to effectively eliminate temporal redundancy, limiting the LVDMs to perceive and generate long sequences of video content. + +To address the above problem, several works [6, 10, 18, 29, 37] have proposed to construct a 3D VAE with both spatial and temporal compression capabilities, which can be summarized in the following three points. In terms of model structure, the UNet network [20] with strong pixel-perception capabilities is still adopted, consistent with the classic image VAE series (i.e., Stable Diffusion VAE). For the operator implementation, most of the methods [6, 10, 18, 29] adopt 3D causal convolution with only forward sequence correlation, which can encode the image information independently [31] while ensuring the continuity of the long sequence video encoding. During the optimization process, a well-trained 2D image VAE, focusing only on spatial compression, is used as an initialization for 3D video VAEs with the same structure and equivalent latent channel number [6, 10, 18, 37], which can achieve better results compared to training a temporal-spatial compressed video VAE from scratch [30]. Although existing methods achieve video reconstruction under $4 \times$ time compression, we find that there remain deficiencies in the above three key processes, limiting the improvement of the video quality and continuity in reconstruction results. + +First, initialization from a well-trained image VAE with + +![](images/049f5a1c964c3e29e38652a45f4217ece8f5fe43804b05a2585af70553dfb43f.jpg) +(a) Number of latent channels + +![](images/631ef923a957e00d8cd86b993c6fcd051cc7f92fbb009e5a5fc47fb10ba7bf5a.jpg) +(b) frames within a frame group +Figure 1. (a) Information preservation degree for different VAEs on video data. CR: Compression Rate. (b) Performance of different frames within a frame group on kinetics-600 dataset (OD-VAE [6], OS (OpenSora V1.2)-VAE [10], CogX (CogVideoX)-VAE [29]). + +the same latent channel number cannot support further $4\times$ time compression, leading to a sharp decline in the initial spatial compression performance and slow learning of subsequent temporal compression. As shown in Fig. 1(a), we calculate the reconstruction similarity of different VAE on video data as the information preservation degree (See Supplementary Material). We note that: (1) As the latent channel number $Z$ increases, the spatial compression capacity gain produced by the image VAE gradually diminishes (blue dashed line) (2) The spatial compression capability provided by an image VAE with dimension $Z / 2$ is sufficient for a temporal-spatial compressed video VAE with dimension $Z$ . Then the video VAE can use the remaining latent channels to store the spatial information of extra frames, i.e., acquiring the initial $2\times$ time compression capacity, which (red curves) converges even better and faster than the regular initialization approach [6, 10, 18, 37] that inherits the image VAE weights of the same dimension $Z$ (green curves). This inspires us to start from an image prior with a lower latent channel number and leave the remaining latent channels for temporal compression, promoting a more balanced temporal-spatial compression initialization. To this end, we propose a dual-branch keyframe-based temporal compression architecture, where a 2D branch inheriting the low-dimensional image VAE prior focuses on keyframe compression, and a 3D branch facilitates temporal compression. + +Second, the unidirectional information flow of 3D causal convolution prevents the interaction between video frames, exacerbating inter-frame unbalanced performance under time compression. In the forward process, $M$ frames (a frame group) are compressed to a whole (one frame) and then restored to $M$ frames. Due to the restriction of forward flow of causal convolution, the first frame in the frame group is unable to perceive the other frames until after time compression in the middle layer of the network. Thus the first frame has the poorest ability to interact with the whole frame group, resulting in higher optimization diffi + +culty compared to other frames. As shown in Fig. 1(b), the methods [6, 10, 29] based on causal convolution suffers from poor performance of the first frame in the frame group and large fluctuations in SSIM values between frames, leading to significant flickering in the reconstruction results. This inspires us to propose group causal convolution, which groups the input frames based on the temporal compression rate. Causal logic padding operations are applied between frame groups to ensure the continuity in processing variable frame videos, and standard convolution operations are used within each group of $M$ frames to share equivalent information interactions, achieving smoother reconstruction. + +In addition, the local receptive field struggles to meet the temporal compression demand at high resolutions. For the same video, the higher the resolution, the more pixels the same motion spans, making it more challenging for the model to capture motion patterns. To this end, we introduce atrous convolutions and add multiple attention modules to expand the receptive field, thus improving the temporal motion perception of the model at high resolutions. + +Finally, considering the increasing demand for content diversity and high resolution in the video generation field, there is a lack of a suitable open-source benchmark for comprehensive evaluation of video VAE reconstruction capabilities. Most of the datasets (e.g., Kinetics-600 [4] and UCF-101 [21]) are 720P or lower, and cannot be used for high resolution (e.g., 1080P) evaluations. Moreover, high-resolution datasets (e.g., Panda-70M [8]) consist mainly of slow-motion and fixed-shot videos, which hardly reflects the overall performance of video VAEs under varying motion speeds. Therefore, we re-collect a subset of 2000 1080P videos from OpenVid-0.4M [15] dataset, uniformly containing various motion speeds, named MotionHD, as a supplement to the overall evaluation. Experimental results show that our method achieves state-of-the-art video reconstruction at various resolutions and motion rates. + +In summary, the contributions of this paper include: (1) + +We propose a keyframe-based temporal compression architecture to accelerate convergence speed by providing a more balanced and powerful temporal-spatial compression initialization. (2) We introduce group causal convolution as a replacement for causal convolution to achieve better and balanced performance between frames, while maintaining causal logic. (3) Extensive experiments on five benchmarks demonstrate the SOTA video reconstruction and generation capabilities of the proposed IV-VAE. + +# 2. Related work + +Variational Autoencoder. Variational Autoencoder (VAE), as a widely used model in the generation domain, is adept at learning complex distributions from high-dimensional data. VAEs can be categorized into two types based on the class of tokens, including discrete and continuous. Classical work of the former type includes MaskGiT [5], VQGAN [9], etc., where continuous latent features are discretized by quantization techniques for autoregression generation. The latter type of continuous VAE, which can achieve spatial or even temporal compression by mapping high-dimensional data distributions to a low-dimensional latent space, is widely used in latent diffusion models [33]. + +Video VAE. Earlier video generation works [1, 28, 38] usually use a image VAE to compress video frames, or fine-tune the image VAE with additional time-dimensional convolutions to reduce flickering artifacts, but this approach cannot effectively reduce the redundancy in the time dimension. With the emergence of OpenAI's Sora [16], to allow video generation models to learn and generate longer videos, there have been some recent works [6, 10, 29, 37] that attempt to train a video VAE with a larger time compression ratios (e.g., $4 \times$ ). Open-Sora [10] adopts a stacked approach, first compressing the video frames in the spatial dimension based on SDXL image VAE [17], and then introducing 3D VAE to compress the temporal dimension. ODVAE [6] uses tail initialization to inflate the 2D convolutions of SD image VAE into 3D causal convolutions and explores an efficient model variant. CV-VAE [37] uses latent space regularization to avoid latent space distribution shift while compressing time. CogVideoX [29] constructs a multi-GPU parallel computing approach to process long videos for causal video VAE. Based on the general idea of these methods, we further explore the shortcomings of the existing video VAEs and propose a series of improvements. + +# 3. Method + +# 3.1. Baseline 3D Causal VAE + +In the beginning, we introduce a causal video VAE as the baseline method. We first train an image VAE based on the architecture of SD VAE with a few structural modifications (Sec. 3.5). Following the thought of MagViT-v2 [31], + +we choose 3D causal convolutions in the construction of video VAE. Then, drawing on the idea of central initialization [3], we inflate the 2D convolution in image VAE to the 3D causal convolution. For example, noting 2D convolutional weights as $W_{2d} \in \mathbb{R}^{c_o \times c_i \times k_h \times k_w}$ , when inflating $W_{2d}$ to be $W_{3d} \in \mathbb{R}^{c_o \times c_i \times k_t \times k_h \times k_w}$ , $W_{3d}$ is formulated as: + +$$ +W _ {3 d} [:,: i,:,: ] = \left\{ \begin{array}{c c} W _ {2 d} & , \text {i f} i = - 1 \\ 0 & , \text {e l s e} \end{array} \right. \tag {1} +$$ + +(See supplementary material for details). In this way, we can obtain high-quality image reconstruction results after extending image VAE to a causal video VAE. Lastly, except that the first image frame is processed separately, we let the causal video VAE achieve $4 \times$ temporal compression. + +Adopting a causal VAE as the baseline has the following two advantages: (1) The first frame is always independent of other frames. According to MagViT-v2 [31], using a temporal causal VAE instead of a regular VAE makes the model perform better in processing a single image. (2) Each frame depends only on the previous frames. Thus, a temporal causal video VAE can use cache mechanism in the inference, which splits the reconstruction process of a long video into a series of short videos with exactly the same result. We consider this a better operation compared to regular leap approach, more details are displayed in Sec. 4.5. + +# 3.2. Group Causal Convolution + +Based on the above baseline, we note that it is practically impossible and non-optimal to ensure causal logic for all frames. For example, the baseline approach compresses time by $4 \times$ , then the input 4 frames will be compressed into 1 frame by the encoder and recovered to 4 frames through the decoder. Obviously, these 4 frames recovered from the same compressed frame do not satisfy temporal causal correlation. In addition, for these 4 reconstructed frames, the first frame usually does not perform as well as subsequent frames, as shown in Fig. 1(b), resulting in frequent flickering in the reconstructed video. We attribute this problem to the insufficient and unbalanced interaction between neighboring frames due to the causal convolution. + +Therefore, to alleviate this problem while being able to maintain the advantages of causal VAE (Sec. 3.1), we propose group causal convolution (GCCov). For a video VAE with a temporal compression rate of $t_c$ ( $t_c = 4$ in our method), the proposed method will initially label every $t_c$ frames of the input video as a frame group, with no overlap between the frame groups. Specifically, the first image frame is a separate frame group. For a frame group, we aim to make the features between frames visible and interactable with each other, thus increasing the reconstruction smoothness between frames and reducing the flicker. For different frame groups, we maintain the property of causal convolution that the current frame group only depends on the previous frame groups. Therefore, the time padding scheme + +![](images/44205ba677c2eeabfa9a02a0fe2c45d5b12397587e8efedcf4a0be93178d2198.jpg) +Figure 2. (a-c) Basic components of IV-VAE. (d) Overall Framework. + +![](images/c29d8c47da4c6bd77e9741d8de703d2183d542f62c7c52c18d271e5a0ee83313.jpg) + +![](images/4b090dd86450d021feb02744f0d705eab76a92a243838973e2b633174c73bfae.jpg) + +for group causal convolution is designed in Fig. 2(a). We pad the features of the previous frame before each frame group, in particular, the first image frame is padded using its own features. And we use zero padding after each frame group. After causal logic padding, each group is operated separately using the same regular convolution. + +# 3.3. Keyframe-based temporal compression + +Building on the observation from Fig. 1(a), to obtain better temporal-spatial compression capability at the initialization of video VAE with latent channel number $Z$ , it is feasible to utilize the pre-trained image VAE weights with a low latent channel number, i.e., $Z / 2$ . To this end, we propose a dual-branch keyframe-based temporal compression (KTC) architecture that decouples the compression of a frame group into the spatial compression of the keyframe and the temporal compression of the overall motion, where the first frame of each frame group is regarded as the keyframe since it can be perceived by the other frames along the temporal forward flow. The basic unit is shown in Fig. 2(b). For the input feature map $F$ with $C_{in}$ channels, it is fed into a 2D convolution to extract the image information and a 3D group causal convolution to extract the overall temporal motion information, respectively. The two convolutions each output feature maps with channels $C_{out} / 2$ , which are then separately normalized by RMSNorm and concatenated together. + +In the initialization phase, the 2D and 3D branches of the network are initialized using the pre-trained image VAE weights with latent channel number $Z / 2$ respectively. In the 3D branch, the 2D image weights are inflated to 3D weights based on the center initialization [3]. In addition, we elaborately set the initial weights of the newly added temporal upsampling and downsampling layers of the two branches. For temporal downsampling layer, we make it initially behave as simply retaining all features of one of the $t_c$ frames and then discarding the remaining frames. For temporal upsampling layer, we let it initially behave as an operation of copying existing frames (More details are provided in Sup + +plementary Material). In this way, each branch is initially enabled to reconstruct one frame from the input $t_c$ frames. With two branches reconstructing different frames, the proposed KTC structure initially realizes $2 \times$ time compression with strong image compression capability as well. + +The overall structure of the network is based on UNet [20] and adopts the proposed KTC as the base unit, as shown in Fig. 2(d). After two temporal downsampling, a frame group is merged into one compressed frame, so we replace the GCConv3D in the KTC Unit with a 2D convolution at this stage to reduce the computation. In the output of the model, each of the two branches produces a sequence of video frames with the same shape as the input, where the reconstruction results of keyframes and remaining frames are selected from the 2D branch and 3D branch outputs respectively, and compose the final reconstruction result. + +# 3.4. Temporal Motion Perception Enhancement + +For the same video, as the video resolution increases, the pixel span generated by object motion or camera motion increases, and the more difficult it is for the video VAE to capture the motion patterns due to the limitations of the receptive field. To this end, we make two improvements to increase the receptive field of the video VAE to enhance the perception of temporal motion at high resolution. As shown in Fig. 2(c), we introduce multiple parallel atrous convolutions (PAC) with different atrous rates in the last layer of the encoder, borrowing from the ASPP in DeepLabV3+ [7]. In addition, we expand the number of attention modules from 2 to 7, and all attention modules are implemented after full temporal and spatial compression to reduce the computational effort. At this time, the compressed frame contains all the information of the frame group, and the global receptive field of attention modules can effectively facilitate temporal motion perception, especially at high resolution. + +# 3.5. Other structural modifications + +Based on the architecture of SD image VAE, we make some modifications to better adapt to 3D causal VAE. First, we + +![](images/2b8a66ff573615e68cca34e293c97cd4b8ee6e3ae4883d825c549c660d14d35e.jpg) +Figure 3. Reconstruction results of different methods for a frame group. + +![](images/6657a2d5ec856a2280028ae4205f6378dc538c318ce6e0267b9bdb6ed7b81894.jpg) +Figure 4. Video motion distribution of Kinetics-600, OpenVid-0.4M, and the filtered MotionHD datasets. + +![](images/c53df5f4316c5c688d5d4a9365d8fa399acaab5637c77feb04e20050b5f6aebd.jpg) + +![](images/823cfde276770a663594627e59d3eaa3c149cc66a8a339c6fef2cef860ad7b7f.jpg) + +![](images/d4a2458e1c72398a06a7c36e6f926686fc475afe5cc5ba1ebc9d3ca58ae7ed2e.jpg) + +replace all GroupNorms with RMSNorms to ensure temporal causality following MagViT-v2 [31], since the standard GroupNorm involves time dimension when calculating the mean and variance of the feature maps, which destroys temporal causality. Second, we shrink the convolution channels during spatial upsampling instead of in the later residual block, reducing memory usage by $29\%$ in the inference. + +# 4. Experiments + +# 4.1. MotionHD dataset + +To better evaluate video VAEs, we propose to consider the video reconstruction performance at different motion amplitudes and resolutions. Among the open-source video datasets, OpenVid-0.4M [15] is a recently released high-quality dataset consisting of 1080P resolution videos, but most of them are fixed camera or slow-moving. To present this phenomenon, we use the RAFT [22] model to compute the optical flow map of video at $768 \times 432$ resolution and then get the average optical flow score as the motion score. As shown in Fig. 4, we present the motion distributions of Kinetics-600 [4] and OpenVid-0.4M datasets. Compared to Kinetics-600, the motion distribution of OpenVid-0.4M is overall narrower and concentrated in low-motion intervals, which is difficult to reflect the reconstruction performance at medium and fast motion. More importantly, in the training of diffusion models, it is common to use a high frame interval to facilitate the learning of object motion, e.g., + +Latte [14] set the frame interval to 3 on all datasets. In this case, the motion of video will be more intense (Fig. 4(a)). + +Therefore, we refer to the motion distribution of the Kinetics-600 (frame interval of 3), and uniformly select 2000 videos with 1080P resolution from OpenVid-0.4M according to the motion scores, as illustrated in Fig. 4(b), constituting the subset MotionHD, which is used to test the average performance of video VAEs at various video motion speeds and the performance at different resolutions. We also list the motion scores of the videos generated by Sora [16] in Fig. 4(c), to facilitate the interpretation of scores and to indicate that our choice of motion intervals is reasonable. + +# 4.2. Experimental Setup + +Evaluation details. We evaluate the reconstruction performance of video VAEs on multiple video benchmarks, including widely used [24-26, 34-36] Kinetics-600 [4] val set, ActivityNet [2] test set, and MotionHD. For Kinetics-600 and ActivityNet, we each randomly select 2400 videos and test them with the original resolution. MotionHD is used to test the performance of Video VAEs at various resolutions. All experiments are conducted using the first 17 frames of the video. To measure reconstruction quality, SSIM, PSNR, LPIPS, and FVD [23] are adopted as metrics. + +Following OD-VAE [6], we employ Latte [14], a soralike transformer-based diffusion model for video generation. We choose Kinetics-400 [11] for class-conditional generation and SkyTimelapse [27] for unconditional gener + +![](images/6342f81a6b00d43f3f5a4bac0a6266e5755a79ee35ed5adf291445c7aa68d1b2.jpg) +Figure 5. Reconstruction results for two adjacent frames in a frame group. In the below figure, the regions with large differences between the two reconstructed frames are boxed out. + +
MethodParamsFCRChnKinetics-600ActivityNet
FVD↓PSNR↑SSIM↑LPIPS↓FVD↓PSNR↑SSIM↑LPIPS↓
SVD-VAE97M1*8*846.2635.260.93630.036755.4737.070.92980.03217
CV-VAE182M4*8*8416.8131.580.90660.0827112.5734.870.91560.06846
OS-VAE393M4*8*8419.0534.370.92580.0816213.5637.640.93400.06781
OD-VAE239M4*8*8410.6933.880.92420.054858.1036.950.93210.04670
Causal VAE127M4*8*8410.8133.770.92480.063098.2637.060.93260.05336
IV-VAE107M4*8*848.0134.290.92810.052096.0838.470.93590.04436
Causal VAE127M4*8*885.7236.080.94790.040334.3938.360.95190.03458
IV-VAE107M4*8*883.9436.660.95200.035993.5039.610.95660.03072
CogX-VAE215M4*8*8163.1738.380.96770.028662.1740.680.97030.02468
Causal VAE127M4*8*8163.2838.070.96410.028872.4340.420.96710.02483
IV-VAE108M4*8*8162.9739.020.96850.022802.0142.610.97220.01968
+ +Table 1. Comparison of video reconstruction results. Chn: Latent channel number. FCR: Frame compression rate. + +ation. All methods are trained for 150K steps with the same default settings in Latte [14]. In addition, the parameters of the video are set to a resolution of $256 \times 256$ and a frame length of 17 during both training and testing. We employ FVD as the metric and calculate it on 2048 samples. + +Training details. We first train an image VAE on $256 \times 256$ image data for 200K steps using 4 A800 GPUs, and then inflate it to a video VAE. During training, KL [13], MAE, and LPIPS are employed as losses. Next, we train the video VAE on $256 \times 256$ resolution videos for 500K steps, and further expand the training resolution to $512 \times 512$ training 200K steps. Finally, we train the VAE at different resolutions with different frame numbers for 100K steps. GAN loss from the 3D discriminator is added at this stage. For causal VAE, we set the basic channel number of the convolution to 96. For IV-VAE, we set it to 64 for both branches. + +# 4.3. Performance + +In addition to baseline Causal VAE, comparison methods include OpenSora V1.2 (OS)-VAE [10], OD-VAE [6], CV-VAE [37],CogVideoX (CogX)-VAE [29], SVD-VAE [1]. + +Quantitative evaluation of reconstruction results. Table 1 demonstrates the performance of the proposed Causal VAE and IV-VAE with other methods on Kinetics-600 and ActivityNet datasets. Experiments show that the proposed method achieves excellent results on both datasets and various numbers of latent channels $Z$ . On $Z = 4$ , we outperform OD-VAE by 2.68 and 2.02 FVD reductions in Kinetics-600 and ActivityNet respectively with less than half parameters. Compared to OS-VAE, we are slightly lower on the PSNR metric due to the limitation of parameters, but we obtain a reduction of 286M (-73%) param + +
MethodChnFVD↓PSNR↑SSIM↑LPIPS↓
480P720P1080P480P720P1080P480P720P1080P480P720P1080P
SVD-VAE45.684.233.6233.7136.0037.690.93130.95070.96170.0350.0330.031
CV-VAE422.3312.788.5029.7030.2931.190.87960.90080.92060.0940.0980.098
OS-VAE416.8910.068.2132.3334.1335.650.91090.93100.94520.0860.0860.087
OD-VAE410.407.365.3131.4233.3234.890.90040.92260.93910.0590.0610.064
Causal VAE49.336.855.1831.6433.4334.820.90870.92860.94150.0670.0690.073
IV-VAE47.014.494.3132.0134.0835.800.91270.93330.94650.0530.0540.055
Causal VAE84.634.153.5834.1435.8937.150.93550.94810.95570.0430.0470.051
IV-VAE84.093.712.8634.6036.5638.110.94100.95340.96110.0370.0390.041
CogX-VAE163.503.662.2636.5237.8638.680.95710.96390.96760.0310.0350.039
Causal VAE163.763.852.7236.2537.6738.600.95440.96170.96630.0320.0370.041
IV-VAE162.582.782.1836.8038.8140.330.95850.96690.97200.0250.0270.027
+ +eters, 11.04 $(-58\%)$ FVD, and 0.02953 $(-36\%)$ LPIPS on Kinetics-600. Compared to SVD-VAE, although our model with $Z = 8$ uses a latent channel number twice as large, we achieve an additional $4\times$ time compression and perform better on all metrics. In addition, we also obtain the best results on $Z = 16$ , especially in PSNR and LPIPS metrics. + +Qualitative evaluation of reconstruction results. As shown in Fig. 3 and 5, we exhibit the reconstruction results for different frames within a frame group. It can be observed that other methods usually perform poorly in the first two frames of a frame group, with blurring and a small amount of distortion, and the last frame performs better. This phenomenon validates our point that for a video VAE with temporal compression, the use of causal convolution makes the frames in the same frame group unequal with respect to the information interaction, leading to performance imbalance. While our method can achieve a more balanced performance benefiting from group causal convolution. Besides, due to the performance fluctuations, other methods usually suffer from abrupt changes in details in the reconstruction of neighboring frames, as shown in the lower panel of Fig. 5, and our method performs better in terms of temporal consistency and reconstruction quality. More visual comparisons are provided in the supplementary material. + +Reconstruction performance at different resolutions. In Table 2, we present the performance of various methods at different resolutions on the MotionHD dataset. The proposed method achieves the best results in the majority of metrics. And as the resolution increases, our improvement relative to other methods is more pronounced. For example, in the PSNR metric, we outperform CogX-VAE by 0.28 at 480P while when the resolution is 1080P, this improvement increases to 1.65. Experiments indicate that the proposed KTC module and TMPE can enable the video VAE to achieve better temporal compression capability, leading to better results in challenging scenarios of temporal compression such as high resolution and rapid motion. + +Video Generation. In the table of Fig. 6, we present the video generation results of different video VAEs using + +Table 2. Video reconstruction results at different resolutions on MotionHD. 480P: ${848} \times {480},{720}\mathrm{P} : {1280} \times {720},{1080}\mathrm{P} : {1920} \times {1280}$ . + +
DatasetCV-VAEOS-VAEOD-VAEIV-VAE (Z=4)
Kinetics-400552.4617.6548.7541.6
SkyTimelapse128.0141.3140.0118.6
+ +![](images/f76f9875d31f98510bd9adc7390fefd3e444539af7d90a1e9c07aba7a1ec3da5.jpg) +Figure 6. Top Table: FVD for different methods. Bottom Figure: Video frames generated by Latte with our IV-VAE. + +the same generation model Latte-XL (Params: 0.67B) [14]. The proposed VAE achieves the best performance on both class-conditional and unconditional generation. Especially on SkyTimelapse, we obtain a 9.4 FVD decrease compared to CV-VAE. The results indicate that IV-VAE can be well employed for LVDMs. At the bottom of Fig. 6, we show some frames of the video generated by Latte with our VAE. The frames exhibit a fairly high level of quality and realism, and even reflect some of the laws of physics, as the sunlight in the sky moves, so does the light reflected in the water. + +# 4.4. Ablation Study + +In this section, we perform a series of ablation experiments. Considering that ablation for three channels requires too much time and resources, we choose the middle channel $Z = 8$ for the ablation of IV-VAE composition. Each ablation experiment follows the same training program: We first train an image VAE for 200K steps, and then extend it to a video VAE and train it for 300K steps on $256 \times 256$ videos. + +Ablation studies of main components. Based on the baseline causal VAE, we explore the effects of the different compositions of IV-VAE in Table 3. As can be seen from the table, the implementation of the proposed group causal convolution can substantially improve the performance on both baseline and KTC structure with the parameter count unchanged. The results indicate that the reconstruction quality can be effectively improved by allowing the frames in the same frame group to interact bidirectionally. Further, + +
SettingTest: 256×256Test: 512×512Params
GCCovKTCTMPEPSNR↑SSIM↑PSNR↑SSIM↑
(A)31.290.904234.880.9384127M
(B)✓31.640.908235.210.9420127M
(C)✓31.860.911635.460.9451104M
(D)✓✓32.120.914535.690.9478104M
(E)✓✓✓32.240.915835.850.9493107M
+ +![](images/97bf7866dfeceac778b914a9159315aae81264fbeb392337df1fb6f0b4b3b6e6.jpg) +Figure 7. Ablation study of structures. The training loss is the sum of MAE, LPIPS, and KL losses. + +replacing the base unit with KTC unit can substantially facilitate the learning of temporal compression thus improving the performance of the model, even with smaller parameters. Finally, we add TMPE enabling the model to better learn the object's motion patterns by increasing the global receptive field with only 3M additional parameter counts. From the improvement of the results at different resolutions, it can be analyzed that this motion perception enhancement plays a more important role in higher resolution scenes. + +Effectiveness of KTC at different latent channel numbers. To verify the validity of the KTC structure, we only replace the structure of the baseline causal VAE with the proposed KTC and perform experiments using the same training setup, the training loss is shown in Fig 7. When the latent channel number $Z = 4$ , the proposed KTC performs slightly superior to the baseline as the training steps increase. And this advantage extends significantly when $Z$ expands to 8 and 16. In conjunction with the results presented in Fig. 1, it can be analyzed that the spatial compression capability improvement of the image VAE is attenuated as $Z$ increases. Therefore, when $Z$ is large, dedicating a portion of the number of latent channels to the initial temporal compression can achieve better and balanced temporal-spatial compression capability initially, leading to faster convergence in subsequent training. + +# 4.5. Analysis + +Increase parameters. In Table 4, we explore the boost provided by increasing the parameters. Increasing the basic feature dimension of IV-VAE from 64 to 96 dramatically increases the parameters and also brings significant improvements, particularly in PSNR and LPIPS metrics. Compared + +Table 3. Ablation study. (A) represents baseline causal VAE. The models are tested on Kinetics-600 using 256 and 512 resolutions. + +
Method (Z=16)ParamsKinetics-60025×1024×1024 Video Rec. time
PSNR↑SSIM↑LPIPS↓
CogX-VAE215M38.380.96770.028711.4s
Ours (dim=64)108M39.020.96850.02286.2s
Ours (dim=96)242M39.570.97110.020510.7s
+ +Table 4. Increasing parameters. The reconstruction time is tested on one A800 GPU with BF16 precision. Rec.: Reconstruction. + +
Rec. approachInput: A 49-frame 1024×1024 VideoKinetics-600
Video clippingMem.Rec. timeSSIM PSNR
One-Step49-frame76.3G12.7s0.9685 39.02
Overleap5-frame*1214.3G14.7s0.9682 38.97
Cache1-frame + 4-frame*1214.1G13.0s0.9685 39.02
+ +Table 5. Cache mechanism vs Overleap mechanism. The overlap mechanism cannot produce the same results as one-step reconstruction. The results are evaluated on $Z = 16$ . Mem.: Memory. + +to CogX-VAE, this model achieves a 1.19 PSNR improvement and still takes less time to reconstruct the video. + +Cache mechanism vs Overleap mechanism. To reconstruct a long video, it is necessary to split the video into a series of clips since GPU memory is limited. Overlap is the most commonly used approach [6], where each input video clip has a one-frame overlap with the previous clip as a history frame, and the overlapped frame is discarded after reconstruction. However, reconstruction with overlap is not equivalent to one-step reconstruction, which may introduce semantic bias in latent space due to the short history frame. And the overlapping frames lead to extra computation. + +To address this problem, we introduce the cache mechanism, which is based on the property that the causal network depends only on the previous frames in temporal order. According to the time padding length $P_{t}$ needed for 3D convolution, the last $P_{t}$ frames of the current video clip's feature maps at each layer of the network are saved, which are used for the convolutional padding of the next video clip. Therefore, this mechanism can be fully equivalent to one-step reconstruction in terms of results. In addition, as shown in Table 5, the caching mechanism requires less time usage and GPU memory usage compared to the overlap mechanism. + +# 5. Conclusion + +In this paper, we observe that regular initialization of utilizing image VAE weights with the same latent channel number is not an optimal scheme, and the use of causal convolution leads to performance imbalance between frames. To address these problems, we propose a keyframe-based temporal compression architecture to accelerate convergence speed by providing more balanced and powerful temporal-spatial compression initialization. In addition, group causal convolution is introduced to achieve better and more balanced performance between frames. Extensive experiments show that IV-VAE achieves SOTA results in both video reconstruction and generation. + +Acknowledgments. This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62306295, 62225207 and 62436008. + +# References + +[1] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 3, 6 +[2] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 5 +[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 3, 4 +[4] Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short note about kinetics-600. arXiv preprint arXiv:1808.01340, 2018. 2, 5 +[5] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[6] Liuhan Chen, Zongjian Li, Bin Lin, Bin Zhu, Qian Wang, Shenghai Yuan, Xing Zhou, Xinghua Cheng, and Li Yuan. Od-vae: An omni-dimensional video compressor for improving latent video diffusion model. arXiv preprint arXiv:2409.01199, 2024. 1, 2, 3, 5, 6, 8 +[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 4 +[8] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2 +[9] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2021. 3 +[10] hpcaitech. Open-sora. https://github.com/hpcaitech/Open-Sora, 2024. 1, 2, 3, 6 +[11] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5 +[12] Kihong Kim, Haneol Lee, Jihye Park, Seyeon Kim, Kwanghee Lee, Seungryong Kim, and Jaejun Yoo. Hybrid video diffusion models with 2d triplane and 3d wavelet representation. In European Conference on Computer Vision, pages 148-165. Springer, 2024. 1 + +[13] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 6 +[14] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 5, 6, 7 +[15] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 2, 5 +[16] OpenAI. Openai sora. Accessed May 15, 2024 [Online], 2024. 1, 3, 5 +[17] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations. 3 +[18] Can Qin, Congying Xia, Krithika Ramakrishnan, Michael Ryoo, Lifu Tu, Yihao Feng, Manli Shu, Honglu Zhou, Anas Awadalla, Jun Wang, et al. xgen-videosyn-1: High-fidelity text-to-video synthesis with compressed representations. arXiv preprint arXiv:2408.12590, 2024. 1, 2 +[19] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1 +[20] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1, 4 +[21] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2 +[22] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 5 +[23] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5 +[24] Haoze Wu, Jiawei Liu, Zheng-Jun Zha, Zhenzhong Chen, and Xiaoyan Sun. Mutually reinforced spatio-temporal convolutional tube for human action recognition. 5 +[25] Haoze Wu, Zheng-Jun Zha, Xin Wen, Zhenzhong Chen, Dong Liu, and Xuejin Chen. Cross-fiber spatial-temporal co-enhanced networks for video action recognition. In Proceedings of the 27th ACM international conference on multimedia, pages 620-628, 2019. +[26] Haoze Wu, Jiawei Liu, Xierong Zhu, Meng Wang, and Zheng-Jun Zha. Multi-scale spatial-temporal integration convolutional tube for human action recognition. In Proceed- + +ings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 753-759, 2021. 5 +[27] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2364-2373, 2018. 5 +[28] Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. Easyanimate: A high-performance long video generation method based on transformer architecture. arXiv preprint arXiv:2405.18991, 2024.3 +[29] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 6 +[30] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 1 +[31] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 1, 3, 5 +[32] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18456-18466, 2023. 1 +[33] Ailing Zeng, Yuhang Yang, Weidong Chen, and Wei Liu. The dawn of video generation: Preliminary explorations with sora-like models. arXiv preprint arXiv:2410.05227, 2024. 3 +[34] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Exploring figure-ground assignment mechanism in perceptual organization. Advances in Neural Information Processing Systems, 35:17030-17042, 2022. 5 +[35] Wei Zhai, Yang Cao, Jing Zhang, Haiyong Xie, Dacheng Tao, and Zheng-Jun Zha. On exploring multiplicity of primitives and attributes for texture recognition in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):403-420, 2023. +[36] Wei Zhai, Pingyu Wu, Kai Zhu, Yang Cao, Feng Wu, and Zheng-Jun Zha. Background activation suppression for weakly supervised object localization and semantic segmentation. International Journal of Computer Vision, 132(3): 750-775, 2024. 5 +[37] Sijie Zhao, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Muyao Niu, Xiaoyu Li, Wenbo Hu, and Ying Shan. Cv-vae: A compatible video vae for latent generative video models. arXiv preprint arXiv:2405.20279, 2024. 1, 2, 3, 6 +[38] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal- + +consistent diffusion model for real-world video superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2545, 2024. 3 \ No newline at end of file diff --git a/2025/Improved Video VAE for Latent Video Diffusion Model/images.zip b/2025/Improved Video VAE for Latent Video Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..55ca597aa09a4f9a8270572fd9e2b819d9f1f5ff --- /dev/null +++ b/2025/Improved Video VAE for Latent Video Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42b1fc675425cd7138e1018ce812888a80d6846726b0ac502555a0aa4b6febef +size 990953 diff --git a/2025/Improved Video VAE for Latent Video Diffusion Model/layout.json b/2025/Improved Video VAE for Latent Video Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3200dad18f484f3408c484118d0036ec73cc82bc --- /dev/null +++ b/2025/Improved Video VAE for Latent Video Diffusion Model/layout.json @@ -0,0 +1,6883 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 103, + 477, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 103, + 477, + 121 + ], + "spans": [ + { + "bbox": [ + 134, + 103, + 477, + 121 + ], + "type": "text", + "content": "Improved Video VAE for Latent Video Diffusion Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "spans": [ + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "text", + "content": "Pingyu Wu" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "text", + "content": ", Kai Zhu" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "text", + "content": ", Yu Liu" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "text", + "content": ", Liming Zhao" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "text", + "content": ", Wei Zhai" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "text", + "content": ", Yang Cao" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "text", + "content": ", Zheng-Jun Zha" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "text", + "content": " MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China " + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 142, + 523, + 200 + ], + "type": "text", + "content": " Independent Researcher {wpy364755620@mail., zkzy@mail., wzhai056@, forrest@, zhazj@}ustc.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 227, + 201, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 201, + 240 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 201, + 240 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 252, + 297, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 252, + 297, + 563 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 297, + 563 + ], + "type": "text", + "content": "Variational Autoencoder (VAE) aims to compress pixel data into low-dimensional latent space, playing an important role in OpenAI's Sora and other latent video diffusion generation models. While most existing video VAEs inflate a pre-trained image VAE into the 3D causal structure for temporal-spatial compression, this paper presents two astonishing findings: (1) The initialization from a well-trained image VAE with the same latent dimensions is not an optimal scheme. (2) The adoption of causal reasoning leads to unequal information interactions and unbalanced performance between frames. To alleviate these problems, we propose a keyframe-based temporal compression (KTC) architecture and a group causal convolution (GCCConv) module to further improve video VAE (IV-VAE). Specifically, the KTC architecture divides the latent space into two branches, in which one half completely inherits the compression prior of keyframes from a lower-dimension image VAE while the other half involves temporal compression through 3D group causal convolution, reducing temporal-spatial conflicts and accelerating the convergence speed of video VAE. The GCConv in the above 3D half uses standard convolution within each frame group to ensure inter-frame equivalence, and employs causal logical padding between groups to maintain flexibility in processing variable frame video. Extensive experiments on five benchmarks demonstrate the SOTA video reconstruction and generation abilities of our IV-VAE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 588, + 137, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 137, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 137, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "text", + "content": "In recent years, video generation has made significant advances in both academia and industry, showcasing cinematic-level visuals and the potential of a world simulator (e.g., Sora [16]). In particular, diffusion optimization methods on latent space have garnered widespread attention and dominated the video generation field due to their high efficiency and stability, resulting in a series of outstanding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 228, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 545, + 240 + ], + "type": "text", + "content": "Latent Video Diffusion Models (LVDMs) [1, 10, 12, 32]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 241, + 554, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 554, + 338 + ], + "type": "text", + "content": "To achieve a bidirectional mapping between the high-dimensional pixel space and the low-dimensional latent space, Variational Autoencoder (VAE) [13] is employed to encode the original video continuously into the latent space, along with spatial and even temporal compression. The compression rate and reconstruction quality of VAE directly determine the information effectiveness of the video corpus, playing an important role in the exploration of LVDMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": "As the most commonly used video VAE, SVD-VAE [1] freezes the weights of the 2D encoder based on the image VAE [19] and retransmits a temporal decoder containing 3D convolutional blocks, which enables the feature interaction across different frames and significantly reduces flicker artifacts. Unfortunately, this method struggles to effectively eliminate temporal redundancy, limiting the LVDMs to perceive and generate long sequences of video content." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 436, + 554, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 554, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 554, + 699 + ], + "type": "text", + "content": "To address the above problem, several works [6, 10, 18, 29, 37] have proposed to construct a 3D VAE with both spatial and temporal compression capabilities, which can be summarized in the following three points. In terms of model structure, the UNet network [20] with strong pixel-perception capabilities is still adopted, consistent with the classic image VAE series (i.e., Stable Diffusion VAE). For the operator implementation, most of the methods [6, 10, 18, 29] adopt 3D causal convolution with only forward sequence correlation, which can encode the image information independently [31] while ensuring the continuity of the long sequence video encoding. During the optimization process, a well-trained 2D image VAE, focusing only on spatial compression, is used as an initialization for 3D video VAEs with the same structure and equivalent latent channel number [6, 10, 18, 37], which can achieve better results compared to training a temporal-spatial compressed video VAE from scratch [30]. Although existing methods achieve video reconstruction under " + }, + { + "bbox": [ + 313, + 436, + 554, + 699 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 436, + 554, + 699 + ], + "type": "text", + "content": " time compression, we find that there remain deficiencies in the above three key processes, limiting the improvement of the video quality and continuity in reconstruction results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "type": "text", + "content": "First, initialization from a well-trained image VAE with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 243, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 243, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 243, + 713 + ], + "type": "text", + "content": "*Kai Zhu and Wei Zhai are the corresponding authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18124" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 79, + 67, + 371, + 220 + ], + "blocks": [ + { + "bbox": [ + 79, + 67, + 371, + 220 + ], + "lines": [ + { + "bbox": [ + 79, + 67, + 371, + 220 + ], + "spans": [ + { + "bbox": [ + 79, + 67, + 371, + 220 + ], + "type": "image", + "image_path": "049f5a1c964c3e29e38652a45f4217ece8f5fe43804b05a2585af70553dfb43f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 220, + 292, + 231 + ], + "lines": [ + { + "bbox": [ + 160, + 220, + 292, + 231 + ], + "spans": [ + { + "bbox": [ + 160, + 220, + 292, + 231 + ], + "type": "text", + "content": "(a) Number of latent channels" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 376, + 68, + 529, + 220 + ], + "blocks": [ + { + "bbox": [ + 376, + 68, + 529, + 220 + ], + "lines": [ + { + "bbox": [ + 376, + 68, + 529, + 220 + ], + "spans": [ + { + "bbox": [ + 376, + 68, + 529, + 220 + ], + "type": "image", + "image_path": "631ef923a957e00d8cd86b993c6fcd051cc7f92fbb009e5a5fc47fb10ba7bf5a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 220, + 523, + 232 + ], + "lines": [ + { + "bbox": [ + 383, + 220, + 523, + 232 + ], + "spans": [ + { + "bbox": [ + 383, + 220, + 523, + 232 + ], + "type": "text", + "content": "(b) frames within a frame group" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "type": "text", + "content": "Figure 1. (a) Information preservation degree for different VAEs on video data. CR: Compression Rate. (b) Performance of different frames within a frame group on kinetics-600 dataset (OD-VAE [6], OS (OpenSora V1.2)-VAE [10], CogX (CogVideoX)-VAE [29])." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "text", + "content": "the same latent channel number cannot support further " + }, + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "text", + "content": " time compression, leading to a sharp decline in the initial spatial compression performance and slow learning of subsequent temporal compression. As shown in Fig. 1(a), we calculate the reconstruction similarity of different VAE on video data as the information preservation degree (See Supplementary Material). We note that: (1) As the latent channel number " + }, + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "text", + "content": " increases, the spatial compression capacity gain produced by the image VAE gradually diminishes (blue dashed line) (2) The spatial compression capability provided by an image VAE with dimension " + }, + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "inline_equation", + "content": "Z / 2" + }, + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "text", + "content": " is sufficient for a temporal-spatial compressed video VAE with dimension " + }, + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "text", + "content": ". Then the video VAE can use the remaining latent channels to store the spatial information of extra frames, i.e., acquiring the initial " + }, + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "text", + "content": " time compression capacity, which (red curves) converges even better and faster than the regular initialization approach [6, 10, 18, 37] that inherits the image VAE weights of the same dimension " + }, + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 56, + 268, + 296, + 579 + ], + "type": "text", + "content": " (green curves). This inspires us to start from an image prior with a lower latent channel number and leave the remaining latent channels for temporal compression, promoting a more balanced temporal-spatial compression initialization. To this end, we propose a dual-branch keyframe-based temporal compression architecture, where a 2D branch inheriting the low-dimensional image VAE prior focuses on keyframe compression, and a 3D branch facilitates temporal compression." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Second, the unidirectional information flow of 3D causal convolution prevents the interaction between video frames, exacerbating inter-frame unbalanced performance under time compression. In the forward process, " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " frames (a frame group) are compressed to a whole (one frame) and then restored to " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " frames. Due to the restriction of forward flow of causal convolution, the first frame in the frame group is unable to perceive the other frames until after time compression in the middle layer of the network. Thus the first frame has the poorest ability to interact with the whole frame group, resulting in higher optimization diffi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 268, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 411 + ], + "type": "text", + "content": "culty compared to other frames. As shown in Fig. 1(b), the methods [6, 10, 29] based on causal convolution suffers from poor performance of the first frame in the frame group and large fluctuations in SSIM values between frames, leading to significant flickering in the reconstruction results. This inspires us to propose group causal convolution, which groups the input frames based on the temporal compression rate. Causal logic padding operations are applied between frame groups to ensure the continuity in processing variable frame videos, and standard convolution operations are used within each group of " + }, + { + "bbox": [ + 313, + 268, + 555, + 411 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 268, + 555, + 411 + ], + "type": "text", + "content": " frames to share equivalent information interactions, achieving smoother reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 412, + 556, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 556, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 556, + 508 + ], + "type": "text", + "content": "In addition, the local receptive field struggles to meet the temporal compression demand at high resolutions. For the same video, the higher the resolution, the more pixels the same motion spans, making it more challenging for the model to capture motion patterns. To this end, we introduce atrous convolutions and add multiple attention modules to expand the receptive field, thus improving the temporal motion perception of the model at high resolutions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 509, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 700 + ], + "type": "text", + "content": "Finally, considering the increasing demand for content diversity and high resolution in the video generation field, there is a lack of a suitable open-source benchmark for comprehensive evaluation of video VAE reconstruction capabilities. Most of the datasets (e.g., Kinetics-600 [4] and UCF-101 [21]) are 720P or lower, and cannot be used for high resolution (e.g., 1080P) evaluations. Moreover, high-resolution datasets (e.g., Panda-70M [8]) consist mainly of slow-motion and fixed-shot videos, which hardly reflects the overall performance of video VAEs under varying motion speeds. Therefore, we re-collect a subset of 2000 1080P videos from OpenVid-0.4M [15] dataset, uniformly containing various motion speeds, named MotionHD, as a supplement to the overall evaluation. Experimental results show that our method achieves state-of-the-art video reconstruction at various resolutions and motion rates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "In summary, the contributions of this paper include: (1)" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18125" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "We propose a keyframe-based temporal compression architecture to accelerate convergence speed by providing a more balanced and powerful temporal-spatial compression initialization. (2) We introduce group causal convolution as a replacement for causal convolution to achieve better and balanced performance between frames, while maintaining causal logic. (3) Extensive experiments on five benchmarks demonstrate the SOTA video reconstruction and generation capabilities of the proposed IV-VAE." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 191, + 139, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 139, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 139, + 202 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 211, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 295, + 354 + ], + "type": "text", + "content": "Variational Autoencoder. Variational Autoencoder (VAE), as a widely used model in the generation domain, is adept at learning complex distributions from high-dimensional data. VAEs can be categorized into two types based on the class of tokens, including discrete and continuous. Classical work of the former type includes MaskGiT [5], VQGAN [9], etc., where continuous latent features are discretized by quantization techniques for autoregression generation. The latter type of continuous VAE, which can achieve spatial or even temporal compression by mapping high-dimensional data distributions to a low-dimensional latent space, is widely used in latent diffusion models [33]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 355, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 295, + 618 + ], + "type": "text", + "content": "Video VAE. Earlier video generation works [1, 28, 38] usually use a image VAE to compress video frames, or fine-tune the image VAE with additional time-dimensional convolutions to reduce flickering artifacts, but this approach cannot effectively reduce the redundancy in the time dimension. With the emergence of OpenAI's Sora [16], to allow video generation models to learn and generate longer videos, there have been some recent works [6, 10, 29, 37] that attempt to train a video VAE with a larger time compression ratios (e.g., " + }, + { + "bbox": [ + 55, + 355, + 295, + 618 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 55, + 355, + 295, + 618 + ], + "type": "text", + "content": "). Open-Sora [10] adopts a stacked approach, first compressing the video frames in the spatial dimension based on SDXL image VAE [17], and then introducing 3D VAE to compress the temporal dimension. ODVAE [6] uses tail initialization to inflate the 2D convolutions of SD image VAE into 3D causal convolutions and explores an efficient model variant. CV-VAE [37] uses latent space regularization to avoid latent space distribution shift while compressing time. CogVideoX [29] constructs a multi-GPU parallel computing approach to process long videos for causal video VAE. Based on the general idea of these methods, we further explore the shortcomings of the existing video VAEs and propose a series of improvements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 192, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 192, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 192, + 659 + ], + "type": "text", + "content": "3.1. Baseline 3D Causal VAE" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "In the beginning, we introduce a causal video VAE as the baseline method. We first train an image VAE based on the architecture of SD VAE with a few structural modifications (Sec. 3.5). Following the thought of MagViT-v2 [31]," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "we choose 3D causal convolutions in the construction of video VAE. Then, drawing on the idea of central initialization [3], we inflate the 2D convolution in image VAE to the 3D causal convolution. For example, noting 2D convolutional weights as " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "W_{2d} \\in \\mathbb{R}^{c_o \\times c_i \\times k_h \\times k_w}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", when inflating " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "W_{2d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " to be " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "W_{3d} \\in \\mathbb{R}^{c_o \\times c_i \\times k_t \\times k_h \\times k_w}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "W_{3d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 358, + 146, + 553, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 146, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 358, + 146, + 553, + 172 + ], + "type": "interline_equation", + "content": "W _ {3 d} [:,: i,:,: ] = \\left\\{ \\begin{array}{c c} W _ {2 d} & , \\text {i f} i = - 1 \\\\ 0 & , \\text {e l s e} \\end{array} \\right. \\tag {1}", + "image_path": "f1915f7d3196be0bbc87111f002d004f29da10b1ae509e482a822e26d016e8ee.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 175, + 553, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 233 + ], + "type": "text", + "content": "(See supplementary material for details). In this way, we can obtain high-quality image reconstruction results after extending image VAE to a causal video VAE. Lastly, except that the first image frame is processed separately, we let the causal video VAE achieve " + }, + { + "bbox": [ + 313, + 175, + 553, + 233 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 175, + 553, + 233 + ], + "type": "text", + "content": " temporal compression." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 234, + 555, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 366 + ], + "type": "text", + "content": "Adopting a causal VAE as the baseline has the following two advantages: (1) The first frame is always independent of other frames. According to MagViT-v2 [31], using a temporal causal VAE instead of a regular VAE makes the model perform better in processing a single image. (2) Each frame depends only on the previous frames. Thus, a temporal causal video VAE can use cache mechanism in the inference, which splits the reconstruction process of a long video into a series of short videos with exactly the same result. We consider this a better operation compared to regular leap approach, more details are displayed in Sec. 4.5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 373, + 463, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 463, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 463, + 385 + ], + "type": "text", + "content": "3.2. Group Causal Convolution" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "text", + "content": "Based on the above baseline, we note that it is practically impossible and non-optimal to ensure causal logic for all frames. For example, the baseline approach compresses time by " + }, + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "text", + "content": ", then the input 4 frames will be compressed into 1 frame by the encoder and recovered to 4 frames through the decoder. Obviously, these 4 frames recovered from the same compressed frame do not satisfy temporal causal correlation. In addition, for these 4 reconstructed frames, the first frame usually does not perform as well as subsequent frames, as shown in Fig. 1(b), resulting in frequent flickering in the reconstructed video. We attribute this problem to the insufficient and unbalanced interaction between neighboring frames due to the causal convolution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": "Therefore, to alleviate this problem while being able to maintain the advantages of causal VAE (Sec. 3.1), we propose group causal convolution (GCCov). For a video VAE with a temporal compression rate of " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "t_c = 4" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": " in our method), the proposed method will initially label every " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": " frames of the input video as a frame group, with no overlap between the frame groups. Specifically, the first image frame is a separate frame group. For a frame group, we aim to make the features between frames visible and interactable with each other, thus increasing the reconstruction smoothness between frames and reducing the flicker. For different frame groups, we maintain the property of causal convolution that the current frame group only depends on the previous frame groups. Therefore, the time padding scheme" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18126" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 68, + 257, + 231 + ], + "blocks": [ + { + "bbox": [ + 62, + 68, + 257, + 231 + ], + "lines": [ + { + "bbox": [ + 62, + 68, + 257, + 231 + ], + "spans": [ + { + "bbox": [ + 62, + 68, + 257, + 231 + ], + "type": "image", + "image_path": "44205ba677c2eeabfa9a02a0fe2c45d5b12397587e8efedcf4a0be93178d2198.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 233, + 436, + 244 + ], + "lines": [ + { + "bbox": [ + 173, + 233, + 436, + 244 + ], + "spans": [ + { + "bbox": [ + 173, + 233, + 436, + 244 + ], + "type": "text", + "content": "Figure 2. (a-c) Basic components of IV-VAE. (d) Overall Framework." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 258, + 68, + 434, + 228 + ], + "blocks": [ + { + "bbox": [ + 258, + 68, + 434, + 228 + ], + "lines": [ + { + "bbox": [ + 258, + 68, + 434, + 228 + ], + "spans": [ + { + "bbox": [ + 258, + 68, + 434, + 228 + ], + "type": "image", + "image_path": "c29d8c47da4c6bd77e9741d8de703d2183d542f62c7c52c18d271e5a0ee83313.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 435, + 68, + 545, + 229 + ], + "blocks": [ + { + "bbox": [ + 435, + 68, + 545, + 229 + ], + "lines": [ + { + "bbox": [ + 435, + 68, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 435, + 68, + 545, + 229 + ], + "type": "image", + "image_path": "4b090dd86450d021feb02744f0d705eab76a92a243838973e2b633174c73bfae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 248, + 296, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 321 + ], + "type": "text", + "content": "for group causal convolution is designed in Fig. 2(a). We pad the features of the previous frame before each frame group, in particular, the first image frame is padded using its own features. And we use zero padding after each frame group. After causal logic padding, each group is operated separately using the same regular convolution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 334, + 261, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 261, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 261, + 346 + ], + "type": "text", + "content": "3.3. Keyframe-based temporal compression" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 352, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 352, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 352, + 295, + 568 + ], + "type": "text", + "content": "Building on the observation from Fig. 1(a), to obtain better temporal-spatial compression capability at the initialization of video VAE with latent channel number " + }, + { + "bbox": [ + 54, + 352, + 295, + 568 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 54, + 352, + 295, + 568 + ], + "type": "text", + "content": ", it is feasible to utilize the pre-trained image VAE weights with a low latent channel number, i.e., " + }, + { + "bbox": [ + 54, + 352, + 295, + 568 + ], + "type": "inline_equation", + "content": "Z / 2" + }, + { + "bbox": [ + 54, + 352, + 295, + 568 + ], + "type": "text", + "content": ". To this end, we propose a dual-branch keyframe-based temporal compression (KTC) architecture that decouples the compression of a frame group into the spatial compression of the keyframe and the temporal compression of the overall motion, where the first frame of each frame group is regarded as the keyframe since it can be perceived by the other frames along the temporal forward flow. The basic unit is shown in Fig. 2(b). For the input feature map " + }, + { + "bbox": [ + 54, + 352, + 295, + 568 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 54, + 352, + 295, + 568 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 54, + 352, + 295, + 568 + ], + "type": "inline_equation", + "content": "C_{in}" + }, + { + "bbox": [ + 54, + 352, + 295, + 568 + ], + "type": "text", + "content": " channels, it is fed into a 2D convolution to extract the image information and a 3D group causal convolution to extract the overall temporal motion information, respectively. The two convolutions each output feature maps with channels " + }, + { + "bbox": [ + 54, + 352, + 295, + 568 + ], + "type": "inline_equation", + "content": "C_{out} / 2" + }, + { + "bbox": [ + 54, + 352, + 295, + 568 + ], + "type": "text", + "content": ", which are then separately normalized by RMSNorm and concatenated together." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "In the initialization phase, the 2D and 3D branches of the network are initialized using the pre-trained image VAE weights with latent channel number " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "Z / 2" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " respectively. In the 3D branch, the 2D image weights are inflated to 3D weights based on the center initialization [3]. In addition, we elaborately set the initial weights of the newly added temporal upsampling and downsampling layers of the two branches. For temporal downsampling layer, we make it initially behave as simply retaining all features of one of the " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " frames and then discarding the remaining frames. For temporal upsampling layer, we let it initially behave as an operation of copying existing frames (More details are provided in Sup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 249, + 553, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 553, + 308 + ], + "type": "text", + "content": "plementary Material). In this way, each branch is initially enabled to reconstruct one frame from the input " + }, + { + "bbox": [ + 313, + 249, + 553, + 308 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 313, + 249, + 553, + 308 + ], + "type": "text", + "content": " frames. With two branches reconstructing different frames, the proposed KTC structure initially realizes " + }, + { + "bbox": [ + 313, + 249, + 553, + 308 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 249, + 553, + 308 + ], + "type": "text", + "content": " time compression with strong image compression capability as well." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 308, + 554, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 440 + ], + "type": "text", + "content": "The overall structure of the network is based on UNet [20] and adopts the proposed KTC as the base unit, as shown in Fig. 2(d). After two temporal downsampling, a frame group is merged into one compressed frame, so we replace the GCConv3D in the KTC Unit with a 2D convolution at this stage to reduce the computation. In the output of the model, each of the two branches produces a sequence of video frames with the same shape as the input, where the reconstruction results of keyframes and remaining frames are selected from the 2D branch and 3D branch outputs respectively, and compose the final reconstruction result." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 445, + 539, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 539, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 539, + 458 + ], + "type": "text", + "content": "3.4. Temporal Motion Perception Enhancement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "content": "For the same video, as the video resolution increases, the pixel span generated by object motion or camera motion increases, and the more difficult it is for the video VAE to capture the motion patterns due to the limitations of the receptive field. To this end, we make two improvements to increase the receptive field of the video VAE to enhance the perception of temporal motion at high resolution. As shown in Fig. 2(c), we introduce multiple parallel atrous convolutions (PAC) with different atrous rates in the last layer of the encoder, borrowing from the ASPP in DeepLabV3+ [7]. In addition, we expand the number of attention modules from 2 to 7, and all attention modules are implemented after full temporal and spatial compression to reduce the computational effort. At this time, the compressed frame contains all the information of the frame group, and the global receptive field of attention modules can effectively facilitate temporal motion perception, especially at high resolution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 672, + 480, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 480, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 480, + 683 + ], + "type": "text", + "content": "3.5. Other structural modifications" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Based on the architecture of SD image VAE, we make some modifications to better adapt to 3D causal VAE. First, we" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18127" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 67, + 549, + 234 + ], + "blocks": [ + { + "bbox": [ + 59, + 67, + 549, + 234 + ], + "lines": [ + { + "bbox": [ + 59, + 67, + 549, + 234 + ], + "spans": [ + { + "bbox": [ + 59, + 67, + 549, + 234 + ], + "type": "image", + "image_path": "2b8a66ff573615e68cca34e293c97cd4b8ee6e3ae4883d825c549c660d14d35e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 237, + 443, + 248 + ], + "lines": [ + { + "bbox": [ + 166, + 237, + 443, + 248 + ], + "spans": [ + { + "bbox": [ + 166, + 237, + 443, + 248 + ], + "type": "text", + "content": "Figure 3. Reconstruction results of different methods for a frame group." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 254, + 235, + 358 + ], + "blocks": [ + { + "bbox": [ + 58, + 254, + 235, + 358 + ], + "lines": [ + { + "bbox": [ + 58, + 254, + 235, + 358 + ], + "spans": [ + { + "bbox": [ + 58, + 254, + 235, + 358 + ], + "type": "image", + "image_path": "6657a2d5ec856a2280028ae4205f6378dc538c318ce6e0267b9bdb6ed7b81894.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 360, + 499, + 372 + ], + "lines": [ + { + "bbox": [ + 110, + 360, + 499, + 372 + ], + "spans": [ + { + "bbox": [ + 110, + 360, + 499, + 372 + ], + "type": "text", + "content": "Figure 4. Video motion distribution of Kinetics-600, OpenVid-0.4M, and the filtered MotionHD datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 255, + 389, + 358 + ], + "blocks": [ + { + "bbox": [ + 235, + 255, + 389, + 358 + ], + "lines": [ + { + "bbox": [ + 235, + 255, + 389, + 358 + ], + "spans": [ + { + "bbox": [ + 235, + 255, + 389, + 358 + ], + "type": "image", + "image_path": "c53df5f4316c5c688d5d4a9365d8fa399acaab5637c77feb04e20050b5f6aebd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 389, + 255, + 470, + 358 + ], + "blocks": [ + { + "bbox": [ + 389, + 255, + 470, + 358 + ], + "lines": [ + { + "bbox": [ + 389, + 255, + 470, + 358 + ], + "spans": [ + { + "bbox": [ + 389, + 255, + 470, + 358 + ], + "type": "image", + "image_path": "823cfde276770a663594627e59d3eaa3c149cc66a8a339c6fef2cef860ad7b7f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 470, + 255, + 550, + 358 + ], + "blocks": [ + { + "bbox": [ + 470, + 255, + 550, + 358 + ], + "lines": [ + { + "bbox": [ + 470, + 255, + 550, + 358 + ], + "spans": [ + { + "bbox": [ + 470, + 255, + 550, + 358 + ], + "type": "image", + "image_path": "d4a2458e1c72398a06a7c36e6f926686fc475afe5cc5ba1ebc9d3ca58ae7ed2e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 377, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 295, + 462 + ], + "type": "text", + "content": "replace all GroupNorms with RMSNorms to ensure temporal causality following MagViT-v2 [31], since the standard GroupNorm involves time dimension when calculating the mean and variance of the feature maps, which destroys temporal causality. Second, we shrink the convolution channels during spatial upsampling instead of in the later residual block, reducing memory usage by " + }, + { + "bbox": [ + 54, + 377, + 295, + 462 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 54, + 377, + 295, + 462 + ], + "type": "text", + "content": " in the inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 472, + 137, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 137, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 137, + 486 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 492, + 165, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 165, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 165, + 504 + ], + "type": "text", + "content": "4.1. MotionHD dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "To better evaluate video VAEs, we propose to consider the video reconstruction performance at different motion amplitudes and resolutions. Among the open-source video datasets, OpenVid-0.4M [15] is a recently released high-quality dataset consisting of 1080P resolution videos, but most of them are fixed camera or slow-moving. To present this phenomenon, we use the RAFT [22] model to compute the optical flow map of video at " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "768 \\times 432" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " resolution and then get the average optical flow score as the motion score. As shown in Fig. 4, we present the motion distributions of Kinetics-600 [4] and OpenVid-0.4M datasets. Compared to Kinetics-600, the motion distribution of OpenVid-0.4M is overall narrower and concentrated in low-motion intervals, which is difficult to reflect the reconstruction performance at medium and fast motion. More importantly, in the training of diffusion models, it is common to use a high frame interval to facilitate the learning of object motion, e.g.," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 377, + 553, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 553, + 401 + ], + "type": "text", + "content": "Latte [14] set the frame interval to 3 on all datasets. In this case, the motion of video will be more intense (Fig. 4(a))." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "text", + "content": "Therefore, we refer to the motion distribution of the Kinetics-600 (frame interval of 3), and uniformly select 2000 videos with 1080P resolution from OpenVid-0.4M according to the motion scores, as illustrated in Fig. 4(b), constituting the subset MotionHD, which is used to test the average performance of video VAEs at various video motion speeds and the performance at different resolutions. We also list the motion scores of the videos generated by Sora [16] in Fig. 4(c), to facilitate the interpretation of scores and to indicate that our choice of motion intervals is reasonable." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 529, + 430, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 430, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 430, + 541 + ], + "type": "text", + "content": "4.2. Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 665 + ], + "type": "text", + "content": "Evaluation details. We evaluate the reconstruction performance of video VAEs on multiple video benchmarks, including widely used [24-26, 34-36] Kinetics-600 [4] val set, ActivityNet [2] test set, and MotionHD. For Kinetics-600 and ActivityNet, we each randomly select 2400 videos and test them with the original resolution. MotionHD is used to test the performance of Video VAEs at various resolutions. All experiments are conducted using the first 17 frames of the video. To measure reconstruction quality, SSIM, PSNR, LPIPS, and FVD [23] are adopted as metrics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Following OD-VAE [6], we employ Latte [14], a soralike transformer-based diffusion model for video generation. We choose Kinetics-400 [11] for class-conditional generation and SkyTimelapse [27] for unconditional gener" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18128" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 550, + 311 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 550, + 311 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 550, + 311 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 550, + 311 + ], + "type": "image", + "image_path": "6342f81a6b00d43f3f5a4bac0a6266e5755a79ee35ed5adf291445c7aa68d1b2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 315, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 315, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 555, + 338 + ], + "type": "text", + "content": "Figure 5. Reconstruction results for two adjacent frames in a frame group. In the below figure, the regions with large differences between the two reconstructed frames are boxed out." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 95, + 346, + 517, + 496 + ], + "blocks": [ + { + "bbox": [ + 95, + 346, + 517, + 496 + ], + "lines": [ + { + "bbox": [ + 95, + 346, + 517, + 496 + ], + "spans": [ + { + "bbox": [ + 95, + 346, + 517, + 496 + ], + "type": "table", + "html": "
MethodParamsFCRChnKinetics-600ActivityNet
FVD↓PSNR↑SSIM↑LPIPS↓FVD↓PSNR↑SSIM↑LPIPS↓
SVD-VAE97M1*8*846.2635.260.93630.036755.4737.070.92980.03217
CV-VAE182M4*8*8416.8131.580.90660.0827112.5734.870.91560.06846
OS-VAE393M4*8*8419.0534.370.92580.0816213.5637.640.93400.06781
OD-VAE239M4*8*8410.6933.880.92420.054858.1036.950.93210.04670
Causal VAE127M4*8*8410.8133.770.92480.063098.2637.060.93260.05336
IV-VAE107M4*8*848.0134.290.92810.052096.0838.470.93590.04436
Causal VAE127M4*8*885.7236.080.94790.040334.3938.360.95190.03458
IV-VAE107M4*8*883.9436.660.95200.035993.5039.610.95660.03072
CogX-VAE215M4*8*8163.1738.380.96770.028662.1740.680.97030.02468
Causal VAE127M4*8*8163.2838.070.96410.028872.4340.420.96710.02483
IV-VAE108M4*8*8162.9739.020.96850.022802.0142.610.97220.01968
", + "image_path": "19be367f19fb642a8d7a5d3151493161d45ad35bc1a84c53d4173c5368da5299.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 498, + 515, + 510 + ], + "lines": [ + { + "bbox": [ + 94, + 498, + 515, + 510 + ], + "spans": [ + { + "bbox": [ + 94, + 498, + 515, + 510 + ], + "type": "text", + "content": "Table 1. Comparison of video reconstruction results. Chn: Latent channel number. FCR: Frame compression rate." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 521, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 582 + ], + "type": "text", + "content": "ation. All methods are trained for 150K steps with the same default settings in Latte [14]. In addition, the parameters of the video are set to a resolution of " + }, + { + "bbox": [ + 54, + 521, + 295, + 582 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 521, + 295, + 582 + ], + "type": "text", + "content": " and a frame length of 17 during both training and testing. We employ FVD as the metric and calculate it on 2048 samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": "Training details. We first train an image VAE on " + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": " image data for 200K steps using 4 A800 GPUs, and then inflate it to a video VAE. During training, KL [13], MAE, and LPIPS are employed as losses. Next, we train the video VAE on " + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": " resolution videos for 500K steps, and further expand the training resolution to " + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": " training 200K steps. Finally, we train the VAE at different resolutions with different frame numbers for 100K steps. GAN loss from the 3D discriminator is added at this stage. For causal VAE, we set the basic channel number of the convolution to 96. For IV-VAE, we set it to 64 for both branches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 521, + 398, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 398, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 398, + 533 + ], + "type": "text", + "content": "4.3. Performance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 542, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 578 + ], + "type": "text", + "content": "In addition to baseline Causal VAE, comparison methods include OpenSora V1.2 (OS)-VAE [10], OD-VAE [6], CV-VAE [37],CogVideoX (CogX)-VAE [29], SVD-VAE [1]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Quantitative evaluation of reconstruction results. Table 1 demonstrates the performance of the proposed Causal VAE and IV-VAE with other methods on Kinetics-600 and ActivityNet datasets. Experiments show that the proposed method achieves excellent results on both datasets and various numbers of latent channels " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": ". On " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "Z = 4" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": ", we outperform OD-VAE by 2.68 and 2.02 FVD reductions in Kinetics-600 and ActivityNet respectively with less than half parameters. Compared to OS-VAE, we are slightly lower on the PSNR metric due to the limitation of parameters, but we obtain a reduction of 286M (-73%) param" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18129" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 70, + 526, + 221 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 526, + 221 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 526, + 221 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 526, + 221 + ], + "type": "table", + "html": "
MethodChnFVD↓PSNR↑SSIM↑LPIPS↓
480P720P1080P480P720P1080P480P720P1080P480P720P1080P
SVD-VAE45.684.233.6233.7136.0037.690.93130.95070.96170.0350.0330.031
CV-VAE422.3312.788.5029.7030.2931.190.87960.90080.92060.0940.0980.098
OS-VAE416.8910.068.2132.3334.1335.650.91090.93100.94520.0860.0860.087
OD-VAE410.407.365.3131.4233.3234.890.90040.92260.93910.0590.0610.064
Causal VAE49.336.855.1831.6433.4334.820.90870.92860.94150.0670.0690.073
IV-VAE47.014.494.3132.0134.0835.800.91270.93330.94650.0530.0540.055
Causal VAE84.634.153.5834.1435.8937.150.93550.94810.95570.0430.0470.051
IV-VAE84.093.712.8634.6036.5638.110.94100.95340.96110.0370.0390.041
CogX-VAE163.503.662.2636.5237.8638.680.95710.96390.96760.0310.0350.039
Causal VAE163.763.852.7236.2537.6738.600.95440.96170.96630.0320.0370.041
IV-VAE162.582.782.1836.8038.8140.330.95850.96690.97200.0250.0270.027
", + "image_path": "cf4efeef2dcad2d4eda0a6fc98b9d91d437fe79fd13fdebc2570711134867661.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 246, + 294, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 294, + 318 + ], + "type": "text", + "content": "eters, 11.04 " + }, + { + "bbox": [ + 55, + 246, + 294, + 318 + ], + "type": "inline_equation", + "content": "(-58\\%)" + }, + { + "bbox": [ + 55, + 246, + 294, + 318 + ], + "type": "text", + "content": " FVD, and 0.02953 " + }, + { + "bbox": [ + 55, + 246, + 294, + 318 + ], + "type": "inline_equation", + "content": "(-36\\%)" + }, + { + "bbox": [ + 55, + 246, + 294, + 318 + ], + "type": "text", + "content": " LPIPS on Kinetics-600. Compared to SVD-VAE, although our model with " + }, + { + "bbox": [ + 55, + 246, + 294, + 318 + ], + "type": "inline_equation", + "content": "Z = 8" + }, + { + "bbox": [ + 55, + 246, + 294, + 318 + ], + "type": "text", + "content": " uses a latent channel number twice as large, we achieve an additional " + }, + { + "bbox": [ + 55, + 246, + 294, + 318 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 55, + 246, + 294, + 318 + ], + "type": "text", + "content": " time compression and perform better on all metrics. In addition, we also obtain the best results on " + }, + { + "bbox": [ + 55, + 246, + 294, + 318 + ], + "type": "inline_equation", + "content": "Z = 16" + }, + { + "bbox": [ + 55, + 246, + 294, + 318 + ], + "type": "text", + "content": ", especially in PSNR and LPIPS metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 318, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 295, + 534 + ], + "type": "text", + "content": "Qualitative evaluation of reconstruction results. As shown in Fig. 3 and 5, we exhibit the reconstruction results for different frames within a frame group. It can be observed that other methods usually perform poorly in the first two frames of a frame group, with blurring and a small amount of distortion, and the last frame performs better. This phenomenon validates our point that for a video VAE with temporal compression, the use of causal convolution makes the frames in the same frame group unequal with respect to the information interaction, leading to performance imbalance. While our method can achieve a more balanced performance benefiting from group causal convolution. Besides, due to the performance fluctuations, other methods usually suffer from abrupt changes in details in the reconstruction of neighboring frames, as shown in the lower panel of Fig. 5, and our method performs better in terms of temporal consistency and reconstruction quality. More visual comparisons are provided in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": "Reconstruction performance at different resolutions. In Table 2, we present the performance of various methods at different resolutions on the MotionHD dataset. The proposed method achieves the best results in the majority of metrics. And as the resolution increases, our improvement relative to other methods is more pronounced. For example, in the PSNR metric, we outperform CogX-VAE by 0.28 at 480P while when the resolution is 1080P, this improvement increases to 1.65. Experiments indicate that the proposed KTC module and TMPE can enable the video VAE to achieve better temporal compression capability, leading to better results in challenging scenarios of temporal compression such as high resolution and rapid motion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Video Generation. In the table of Fig. 6, we present the video generation results of different video VAEs using" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 247, + 552, + 286 + ], + "blocks": [ + { + "bbox": [ + 59, + 225, + 550, + 236 + ], + "lines": [ + { + "bbox": [ + 59, + 225, + 550, + 236 + ], + "spans": [ + { + "bbox": [ + 59, + 225, + 550, + 236 + ], + "type": "text", + "content": "Table 2. Video reconstruction results at different resolutions on MotionHD. 480P: " + }, + { + "bbox": [ + 59, + 225, + 550, + 236 + ], + "type": "inline_equation", + "content": "{848} \\times {480},{720}\\mathrm{P} : {1280} \\times {720},{1080}\\mathrm{P} : {1920} \\times {1280}" + }, + { + "bbox": [ + 59, + 225, + 550, + 236 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 247, + 552, + 286 + ], + "lines": [ + { + "bbox": [ + 317, + 247, + 552, + 286 + ], + "spans": [ + { + "bbox": [ + 317, + 247, + 552, + 286 + ], + "type": "table", + "html": "
DatasetCV-VAEOS-VAEOD-VAEIV-VAE (Z=4)
Kinetics-400552.4617.6548.7541.6
SkyTimelapse128.0141.3140.0118.6
", + "image_path": "7a45c543382b11c57d16814e9922294b7f705b173739ebcd5fd45c230bfa0c08.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 288, + 552, + 337 + ], + "blocks": [ + { + "bbox": [ + 317, + 288, + 552, + 337 + ], + "lines": [ + { + "bbox": [ + 317, + 288, + 552, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 288, + 552, + 337 + ], + "type": "image", + "image_path": "f76f9875d31f98510bd9adc7390fefd3e444539af7d90a1e9c07aba7a1ec3da5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 340, + 553, + 363 + ], + "lines": [ + { + "bbox": [ + 313, + 340, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 553, + 363 + ], + "type": "text", + "content": "Figure 6. Top Table: FVD for different methods. Bottom Figure: Video frames generated by Latte with our IV-VAE." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 373, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 492 + ], + "type": "text", + "content": "the same generation model Latte-XL (Params: 0.67B) [14]. The proposed VAE achieves the best performance on both class-conditional and unconditional generation. Especially on SkyTimelapse, we obtain a 9.4 FVD decrease compared to CV-VAE. The results indicate that IV-VAE can be well employed for LVDMs. At the bottom of Fig. 6, we show some frames of the video generated by Latte with our VAE. The frames exhibit a fairly high level of quality and realism, and even reflect some of the laws of physics, as the sunlight in the sky moves, so does the light reflected in the water." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 502, + 408, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 408, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 408, + 516 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "content": "In this section, we perform a series of ablation experiments. Considering that ablation for three channels requires too much time and resources, we choose the middle channel " + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "inline_equation", + "content": "Z = 8" + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "content": " for the ablation of IV-VAE composition. Each ablation experiment follows the same training program: We first train an image VAE for 200K steps, and then extend it to a video VAE and train it for 300K steps on " + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "content": " videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Ablation studies of main components. Based on the baseline causal VAE, we explore the effects of the different compositions of IV-VAE in Table 3. As can be seen from the table, the implementation of the proposed group causal convolution can substantially improve the performance on both baseline and KTC structure with the parameter count unchanged. The results indicate that the reconstruction quality can be effectively improved by allowing the frames in the same frame group to interact bidirectionally. Further," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18130" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 295, + 153 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 295, + 153 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 295, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 295, + 153 + ], + "type": "table", + "html": "
SettingTest: 256×256Test: 512×512Params
GCCovKTCTMPEPSNR↑SSIM↑PSNR↑SSIM↑
(A)31.290.904234.880.9384127M
(B)✓31.640.908235.210.9420127M
(C)✓31.860.911635.460.9451104M
(D)✓✓32.120.914535.690.9478104M
(E)✓✓✓32.240.915835.850.9493107M
", + "image_path": "27ee2ff724b3a804405822d0051b754dec4222f7a23ce73160aedf340e4e0868.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 182, + 282, + 293 + ], + "blocks": [ + { + "bbox": [ + 67, + 182, + 282, + 293 + ], + "lines": [ + { + "bbox": [ + 67, + 182, + 282, + 293 + ], + "spans": [ + { + "bbox": [ + 67, + 182, + 282, + 293 + ], + "type": "image", + "image_path": "97bf7866dfeceac778b914a9159315aae81264fbeb392337df1fb6f0b4b3b6e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 296, + 295, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 295, + 319 + ], + "type": "text", + "content": "Figure 7. Ablation study of structures. The training loss is the sum of MAE, LPIPS, and KL losses." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 328, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 435 + ], + "type": "text", + "content": "replacing the base unit with KTC unit can substantially facilitate the learning of temporal compression thus improving the performance of the model, even with smaller parameters. Finally, we add TMPE enabling the model to better learn the object's motion patterns by increasing the global receptive field with only 3M additional parameter counts. From the improvement of the results at different resolutions, it can be analyzed that this motion perception enhancement plays a more important role in higher resolution scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": "Effectiveness of KTC at different latent channel numbers. To verify the validity of the KTC structure, we only replace the structure of the baseline causal VAE with the proposed KTC and perform experiments using the same training setup, the training loss is shown in Fig 7. When the latent channel number " + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "inline_equation", + "content": "Z = 4" + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": ", the proposed KTC performs slightly superior to the baseline as the training steps increase. And this advantage extends significantly when " + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": " expands to 8 and 16. In conjunction with the results presented in Fig. 1, it can be analyzed that the spatial compression capability improvement of the image VAE is attenuated as " + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": " increases. Therefore, when " + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": " is large, dedicating a portion of the number of latent channels to the initial temporal compression can achieve better and balanced temporal-spatial compression capability initially, leading to faster convergence in subsequent training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 636, + 118, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 118, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 118, + 648 + ], + "type": "text", + "content": "4.5. Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "Increase parameters. In Table 4, we explore the boost provided by increasing the parameters. Increasing the basic feature dimension of IV-VAE from 64 to 96 dramatically increases the parameters and also brings significant improvements, particularly in PSNR and LPIPS metrics. Compared" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 130 + ], + "blocks": [ + { + "bbox": [ + 56, + 155, + 295, + 177 + ], + "lines": [ + { + "bbox": [ + 56, + 155, + 295, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 155, + 295, + 177 + ], + "type": "text", + "content": "Table 3. Ablation study. (A) represents baseline causal VAE. The models are tested on Kinetics-600 using 256 and 512 resolutions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 71, + 553, + 130 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 130 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 130 + ], + "type": "table", + "html": "
Method (Z=16)ParamsKinetics-60025×1024×1024 Video Rec. time
PSNR↑SSIM↑LPIPS↓
CogX-VAE215M38.380.96770.028711.4s
Ours (dim=64)108M39.020.96850.02286.2s
Ours (dim=96)242M39.570.97110.020510.7s
", + "image_path": "2e1a9f6eb7360b0ccc6c10926244c56fceb0f0d353cc313f603251c54608314e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 159, + 553, + 219 + ], + "blocks": [ + { + "bbox": [ + 314, + 132, + 553, + 153 + ], + "lines": [ + { + "bbox": [ + 314, + 132, + 553, + 153 + ], + "spans": [ + { + "bbox": [ + 314, + 132, + 553, + 153 + ], + "type": "text", + "content": "Table 4. Increasing parameters. The reconstruction time is tested on one A800 GPU with BF16 precision. Rec.: Reconstruction." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 159, + 553, + 219 + ], + "lines": [ + { + "bbox": [ + 317, + 159, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 317, + 159, + 553, + 219 + ], + "type": "table", + "html": "
Rec. approachInput: A 49-frame 1024×1024 VideoKinetics-600
Video clippingMem.Rec. timeSSIM PSNR
One-Step49-frame76.3G12.7s0.9685 39.02
Overleap5-frame*1214.3G14.7s0.9682 38.97
Cache1-frame + 4-frame*1214.1G13.0s0.9685 39.02
", + "image_path": "09ebff7cde4fb8a99b0957dc084c9e4d784131b55f2b70e98ae319f0b168505c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 221, + 553, + 254 + ], + "lines": [ + { + "bbox": [ + 313, + 221, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 553, + 254 + ], + "type": "text", + "content": "Table 5. Cache mechanism vs Overleap mechanism. The overlap mechanism cannot produce the same results as one-step reconstruction. The results are evaluated on " + }, + { + "bbox": [ + 313, + 221, + 553, + 254 + ], + "type": "inline_equation", + "content": "Z = 16" + }, + { + "bbox": [ + 313, + 221, + 553, + 254 + ], + "type": "text", + "content": ". Mem.: Memory." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 264, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 287 + ], + "type": "text", + "content": "to CogX-VAE, this model achieves a 1.19 PSNR improvement and still takes less time to reconstruct the video." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 288, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 407 + ], + "type": "text", + "content": "Cache mechanism vs Overleap mechanism. To reconstruct a long video, it is necessary to split the video into a series of clips since GPU memory is limited. Overlap is the most commonly used approach [6], where each input video clip has a one-frame overlap with the previous clip as a history frame, and the overlapped frame is discarded after reconstruction. However, reconstruction with overlap is not equivalent to one-step reconstruction, which may introduce semantic bias in latent space due to the short history frame. And the overlapping frames lead to extra computation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 407, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 540 + ], + "type": "text", + "content": "To address this problem, we introduce the cache mechanism, which is based on the property that the causal network depends only on the previous frames in temporal order. According to the time padding length " + }, + { + "bbox": [ + 313, + 407, + 555, + 540 + ], + "type": "inline_equation", + "content": "P_{t}" + }, + { + "bbox": [ + 313, + 407, + 555, + 540 + ], + "type": "text", + "content": " needed for 3D convolution, the last " + }, + { + "bbox": [ + 313, + 407, + 555, + 540 + ], + "type": "inline_equation", + "content": "P_{t}" + }, + { + "bbox": [ + 313, + 407, + 555, + 540 + ], + "type": "text", + "content": " frames of the current video clip's feature maps at each layer of the network are saved, which are used for the convolutional padding of the next video clip. Therefore, this mechanism can be fully equivalent to one-step reconstruction in terms of results. In addition, as shown in Table 5, the caching mechanism requires less time usage and GPU memory usage compared to the overlap mechanism." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 550, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we observe that regular initialization of utilizing image VAE weights with the same latent channel number is not an optimal scheme, and the use of causal convolution leads to performance imbalance between frames. To address these problems, we propose a keyframe-based temporal compression architecture to accelerate convergence speed by providing more balanced and powerful temporal-spatial compression initialization. In addition, group causal convolution is introduced to achieve better and more balanced performance between frames. Extensive experiments show that IV-VAE achieves SOTA results in both video reconstruction and generation." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18131" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "Acknowledgments. This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62306295, 62225207 and 62436008." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 115, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 115, + 131 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 115, + 131 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 137, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 137, + 295, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 295, + 192 + ], + "type": "text", + "content": "[1] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 194, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 295, + 247 + ], + "type": "text", + "content": "[2] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 248, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 248, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 248, + 296, + 293 + ], + "type": "text", + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 293, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 293, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 61, + 293, + 295, + 326 + ], + "type": "text", + "content": "[4] Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short note about kinetics-600. arXiv preprint arXiv:1808.01340, 2018. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 295, + 369 + ], + "type": "text", + "content": "[5] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 371, + 296, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 296, + 425 + ], + "type": "text", + "content": "[6] Liuhan Chen, Zongjian Li, Bin Lin, Bin Zhu, Qian Wang, Shenghai Yuan, Xing Zhou, Xinghua Cheng, and Li Yuan. Od-vae: An omni-dimensional video compressor for improving latent video diffusion model. arXiv preprint arXiv:2409.01199, 2024. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 426, + 296, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 296, + 480 + ], + "type": "text", + "content": "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 481, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 296, + 557 + ], + "type": "text", + "content": "[8] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 558, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 296, + 581 + ], + "type": "text", + "content": "[9] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 582, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 295, + 602 + ], + "type": "text", + "content": "[10] hpcaitech. Open-sora. https://github.com/hpcaitech/Open-Sora, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 603, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 657 + ], + "type": "text", + "content": "[11] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[12] Kihong Kim, Haneol Lee, Jihye Park, Seyeon Kim, Kwanghee Lee, Seungryong Kim, and Jaejun Yoo. Hybrid video diffusion models with 2d triplane and 3d wavelet representation. In European Conference on Computer Vision, pages 148-165. Springer, 2024. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[13] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "text", + "content": "[14] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "text", + "content": "[15] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 218 + ], + "type": "text", + "content": "[16] OpenAI. Openai sora. Accessed May 15, 2024 [Online], 2024. 1, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "text", + "content": "[17] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "type": "text", + "content": "[18] Can Qin, Congying Xia, Krithika Ramakrishnan, Michael Ryoo, Lifu Tu, Yihao Feng, Manli Shu, Honglu Zhou, Anas Awadalla, Jun Wang, et al. xgen-videosyn-1: High-fidelity text-to-video synthesis with compressed representations. arXiv preprint arXiv:2408.12590, 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 555, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 388 + ], + "type": "text", + "content": "[19] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 454 + ], + "type": "text", + "content": "[20] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "text", + "content": "[21] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 489, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 543 + ], + "type": "text", + "content": "[22] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "type": "text", + "content": "[23] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 624 + ], + "type": "text", + "content": "[24] Haoze Wu, Jiawei Liu, Zheng-Jun Zha, Zhenzhong Chen, and Xiaoyan Sun. Mutually reinforced spatio-temporal convolutional tube for human action recognition. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "text", + "content": "[25] Haoze Wu, Zheng-Jun Zha, Xin Wen, Zhenzhong Chen, Dong Liu, and Xuejin Chen. Cross-fiber spatial-temporal co-enhanced networks for video action recognition. In Proceedings of the 27th ACM international conference on multimedia, pages 620-628, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Haoze Wu, Jiawei Liu, Xierong Zhu, Meng Wang, and Zheng-Jun Zha. Multi-scale spatial-temporal integration convolutional tube for human action recognition. In Proceed-" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18132" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "ings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 753-759, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[27] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2364-2373, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "text", + "content": "[28] Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. Easyanimate: A high-performance long video generation method based on transformer architecture. arXiv preprint arXiv:2405.18991, 2024.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 274 + ], + "type": "text", + "content": "[29] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 296, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 341 + ], + "type": "text", + "content": "[30] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 296, + 397 + ], + "type": "text", + "content": "[31] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 1, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 296, + 453 + ], + "type": "text", + "content": "[32] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18456-18466, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 455, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 296, + 487 + ], + "type": "text", + "content": "[33] Ailing Zeng, Yuhang Yang, Weidong Chen, and Wei Liu. The dawn of video generation: Preliminary explorations with sora-like models. arXiv preprint arXiv:2410.05227, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 296, + 533 + ], + "type": "text", + "content": "[34] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Exploring figure-ground assignment mechanism in perceptual organization. Advances in Neural Information Processing Systems, 35:17030-17042, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 296, + 588 + ], + "type": "text", + "content": "[35] Wei Zhai, Yang Cao, Jing Zhang, Haiyong Xie, Dacheng Tao, and Zheng-Jun Zha. On exploring multiplicity of primitives and attributes for texture recognition in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):403-420, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 296, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 643 + ], + "type": "text", + "content": "[36] Wei Zhai, Pingyu Wu, Kai Zhu, Yang Cao, Feng Wu, and Zheng-Jun Zha. Background activation suppression for weakly supervised object localization and semantic segmentation. International Journal of Computer Vision, 132(3): 750-775, 2024. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 646, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 296, + 690 + ], + "type": "text", + "content": "[37] Sijie Zhao, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Muyao Niu, Xiaoyu Li, Wenbo Hu, and Ying Shan. Cv-vae: A compatible video vae for latent generative video models. arXiv preprint arXiv:2405.20279, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "text", + "content": "[38] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 72, + 555, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 72, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 555, + 117 + ], + "type": "text", + "content": "consistent diffusion model for real-world video superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2545, 2024. 3" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18133" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/6a2c48a3-aafc-4ad2-9262-8700829f154e_content_list.json b/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/6a2c48a3-aafc-4ad2-9262-8700829f154e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d38cc65aa824df7b68fe46c62ab8bd895191d9dc --- /dev/null +++ b/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/6a2c48a3-aafc-4ad2-9262-8700829f154e_content_list.json @@ -0,0 +1,1657 @@ +[ + { + "type": "text", + "text": "Improving Accuracy and Calibration via Differentiated Deep Mutual Learning", + "text_level": 1, + "bbox": [ + 99, + 128, + 898, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Han Liu $^{1,2,\\dagger}$ , Peng Cui $^{1,\\dagger}$ , Binging Wang $^{2}$ , Weipeng Chen $^{2}$ , Yupeng Zhang $^{2}$ , Jun Zhu $^{1,*}$ , and Xiaolin Hu $^{1,3,4,*}$", + "bbox": [ + 200, + 178, + 795, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Computer Science and Technology, Institute for AI, BNRist, Tsinghua University, Beijing 100084, China", + "bbox": [ + 232, + 215, + 764, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Baichuan Inc.", + "bbox": [ + 437, + 252, + 557, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Tsinghua Laboratory of Brain and Intelligence (THBI),", + "bbox": [ + 274, + 268, + 720, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing 100084, China", + "bbox": [ + 145, + 287, + 849, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Chinese Institute for Brain Research (CIBR), Beijing 100010, China", + "bbox": [ + 223, + 304, + 772, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{han-liu18, cuip22}@mails.tsinghua.edu.cn", + "bbox": [ + 316, + 324, + 679, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{daniel, chenweipeng, zhangyupeng}@baichuan-inc.com", + "bbox": [ + 272, + 342, + 723, + 357 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dcszj, xlhu}@mail.tsinghua.edu.cn", + "bbox": [ + 349, + 359, + 648, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 409, + 326, + 425 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Neural Networks (DNNs) have achieved remarkable success in a variety of tasks, particularly in terms of prediction accuracy. However, in real-world scenarios, especially in safety-critical applications, accuracy alone is insufficient; reliable uncertainty estimates are essential. Modern DNNs, often trained with cross-entropy loss, tend to exhibit overconfidence, especially on ambiguous samples. Many techniques aim to improve uncertainty calibration, yet they often come at the cost of reduced accuracy or increased computational demands. To address this challenge, we propose Differentiated Deep Mutual Learning (Diff-DML), an efficient ensemble approach that simultaneously enhances accuracy and uncertainty calibration. Diff-DML draws inspiration from Deep Mutual Learning (DML) while introducing two strategies to maintain prediction diversity: (1) Differentiated Training Strategy (DTS) and (2) Diversity-Preserving Learning Objective (DPLO). Our theoretical analysis shows that Diff-DML's diversified learning framework not only leverages ensemble benefits but also avoids the loss of prediction diversity observed in traditional DML setups, which is crucial for improved calibration. Extensive evaluations on various benchmarks confirm the effectiveness of Diff-DML. For instance, on the CIFAR-100 dataset, Diff-DML on ResNet34/50 models achieved substantial improvements over the previous state-of-the-art method, MDCA, with absolute accuracy gains of $1.3\\% / 3.1\\%$ , relative ECE reductions", + "bbox": [ + 86, + 441, + 485, + 835 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b25785f2e35a5aba3aca0aa4a060bf130a35505e1c140444266a82b18197c884.jpg", + "image_caption": [ + "Figure 1. Scatter plot of ECE and accuracy for different calibration methods based on the ResNet50 backbone on the CIFAR-100 test set. Methods located closer to the top left corner perform better." + ], + "image_footnote": [], + "bbox": [ + 547, + 411, + 870, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of $49.6\\% / 43.8\\%$ , and relative classwise-ECE reductions of $7.7\\% / 13.0\\%$ .", + "bbox": [ + 511, + 674, + 908, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 737, + 645, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Neural Networks (DNNs) have achieved great success in various tasks, especially in terms of prediction accuracy. In safety-critical areas such as medical diagnosis [4, 18], self-driving cars [20], and protein engineering [6], accurate uncertainty estimation (i.e., well-calibrated prediction) is crucial for reliable model deployment. For instance, a reliable model needs to show high uncertainty when making incorrect predictions. However, models trained with cross-entropy (CE) loss tend to be overconfident, especially when", + "bbox": [ + 509, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was supported by the National Key Research and Development Program of China (No. 2021ZD0200301) and the National Natural Science Foundation of China (No. U2341228).", + "bbox": [ + 89, + 849, + 482, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution. \\*Corresponding author.", + "bbox": [ + 116, + 887, + 352, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25812", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "encountering ambiguous samples, leading to predictions that are often close to $100\\%$ confident and producing undesirable overconfidence [2, 8].", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, various regularization techniques have been explored during training, referred to as training-time calibration. For instance, entropy regularization (ER) [26] and label smoothing (LS) [22, 28] promote smoother predictive probability distributions to alleviate overconfidence by maximizing entropy or adjusting label distributions. Some methods aim to directly reduce calibration error during training to achieve multi-class calibration. For example, MMCE [16] introduces a trainable measure based on kernel mean embeddings, while MDCA [10] employs a differentiable auxiliary loss to minimize the gap between prediction confidence and accuracy. However, despite mitigating overconfidence, these techniques often compromise prediction accuracy due to excessive regularization. Another category of methods is post-processing, which adjusts output probabilities using Temperature Scaling (TS) [27], Vector Scaling (VS) [27], or Dirichlet Scaling (DS) [15]. However, these post-processing techniques typically rely on an additional validation dataset. Furthermore, they have limited learnable parameters, which may lead to suboptimal calibration performance [10].", + "bbox": [ + 91, + 138, + 485, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Ensembles (DE) [5, 17] offer a straightforward solution by combining multiple randomly initialized trained DNNs to improve prediction accuracy and uncertainty calibration. Yet, using multiple DNNs during training and inference significantly increases computational costs, which poses challenges for deploying DE in real-world scenarios. To address this limitation, we draw inspiration from Deep Mutual Learning (DML) [29], a technique used in model distillation, where an ensemble of student models learn collaboratively and teach each other throughout the training process. This work provides the theoretical analysis showing that DML encourages each model to learn from an ensemble distribution through a mimicry loss. The proposed theory motivates us to leverage DML as an implicit form of DE, aiming to achieve similar gains in accuracy and calibration as DE, while mitigating the high computational overhead because DML uses only one model for prediction during inference.", + "bbox": [ + 91, + 459, + 483, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, in standard DML, although the models start from different initial states, they become overly similar during mutual learning, which reduces the diversity of their outputs—a key factor needed for effective ensemble-based calibration. In this work, we propose Differentiated Deep Mutual Learning (Diff-DML), a novel approach that enhances the benefits of DML by promoting prediction diversity between models to achieve high accuracy and strong calibration performance. Diff-DML incorporates two key strategies that preserve output diversity: (1) Differentiated Training Strategy (DTS), which uses different learning rates and schedules", + "bbox": [ + 89, + 734, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for different models to prevent premature convergence, and (2) Diversity-Preserving Learning Objective (DPLO), which removes cross-entropy loss for auxiliary models to avoid overfitting to the one-hot ground-truth labels.", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These strategies help Diff-DML balance collaborative learning with prediction diversity, resulting in better uncertainty calibration and enhanced predictive performance. Extensive evaluations on various benchmarks, including image classification, misclassification detection, and out-of-distribution (OOD) detection, confirm Diff-DML's effectiveness. For example, on the CIFAR-100 dataset, Diff-DML on ResNet34/50 achieved significant improvements compared to the previous state-of-the-art method, MDCA [10], with absolute accuracy increases of $1.3\\% / 3.1\\%$ , relative reductions in ECE of $49.6\\% / 43.8\\%$ , and relative reductions in classwise-ECE of $7.7\\% / 13.0\\%$ . The latter two metrics measure calibration performance, where lower values indicate better results.", + "bbox": [ + 511, + 152, + 908, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Figure 1 displays ECE and accuracy of different calibration methods, showing Diff-DML achieving favorable calibration and accuracy. Notably, Diff-DML incurs only minimal computational cost at inference, unlike DE, which requires multiple models to generate a single prediction.", + "bbox": [ + 511, + 363, + 908, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and related work", + "text_level": 1, + "bbox": [ + 513, + 452, + 790, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Calibration, the process of ensuring that the predicted probabilities reflect the true likelihood of the events, plays a crucial role in machine learning models. It ensures that the prediction of machine learning models are not only accurate but also well-calibrated. Well-calibrated models provide reliable uncertainty estimates, which can help users and downstream systems make safer decisions by avoiding overconfidence.", + "bbox": [ + 511, + 477, + 906, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are two fundamental types of calibration: confidence calibration [23] and classwise calibration [15]. In this paper, we focus on both types of calibration.", + "bbox": [ + 511, + 583, + 906, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Confidence calibration: Confidence calibration refers to the alignment of predicted confidence with observed frequencies. It ensures that a model's prediction confidence accurately reflects the true likelihood of an event occurring. A widely used metric for evaluating confidence calibration is the Expected Calibration Error (ECE) [23]. The definition of ECE can be found in Appendix A.", + "bbox": [ + 511, + 628, + 908, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Classwise calibration: Classwise calibration is concerned with calibrating the predicted probabilities for each individual class in a multi-class classification problem. It ensures that the predicted probabilities across different classes are accurate and properly calibrated. The metric for classwise calibration is classwise-ECE (cw-ECE) [15], which can be found in Appendix A.", + "bbox": [ + 511, + 734, + 908, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "According to the number of models used for inference, calibration methods can be divided into single-model calibration methods and DE [5, 17].", + "bbox": [ + 511, + 839, + 906, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-model calibration methods: Within single-model", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25813", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "calibration methods, there are two categories: train-time calibration methods and post-hoc calibration methods. Train-time calibration methods typically involve designing loss functions or incorporating additional regularization techniques, such as Entropy Regularization (ER) [26], Label Smoothing (LS) [22, 28], Focal Loss (FL) [21], Maximum Mean Calibration Error (MMCE) [16], Difference between Confidence and Accuracy (DCA) [19], Multi-class Difference in Confidence and Accuracy (MDCA) [10], and Dynamic Train-time Data Pruning (DTDP) [25]. Post-hoc calibration methods are applied after the model is trained, such as Temperature Scaling (TS) [27], Vector Scaling, and Dirichlet Scaling (DS) [15].", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deep Ensembles: DE creates an ensemble by training multiple models independently with different initializations. To alleviate overconfidence and improve calibration, during inference, DE leverages $N$ randomly initialized, individually trained DNNs and makes predictions by average voting. While effective for enhancing model performance, DE introduces additional computational overhead during inference.", + "bbox": [ + 89, + 287, + 483, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More related works are discussed in Appendix B.", + "bbox": [ + 109, + 393, + 437, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 422, + 181, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $P_{data}$ denote the probability distribution of data. A dataset consists of finite $(x, y_x)$ pairs where $x \\sim P_{data}$ is a data instance and $y_x \\in \\mathcal{Y} = \\{1, 2, \\dots, K\\}$ is the ground-truth class label of $x$ . A model takes in the instance $x$ and outputs its prediction vector $p_x \\in \\mathbb{R}^K$ . Let $p_x^{(k)}$ denote the $k$ -th dimension of $p_x$ . The confidence $\\hat{p}_x = \\max_{k \\in \\mathcal{Y}} \\{p_x^{(k)}\\}$ is the maximum value in the probability vector of predictions, and the model's prediction $\\hat{y}_x = \\operatorname{argmax}_{k \\in \\mathcal{Y}} \\{p_x^{(k)}\\}$ is the index corresponding to the maximum value.", + "bbox": [ + 89, + 446, + 483, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Typically, we train a model by minimizing the following CE loss on the training set,", + "bbox": [ + 89, + 589, + 483, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {C E}} \\left(p _ {x}\\right) = - \\log p _ {x} ^ {\\left(y _ {x}\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 631, + 482, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p_x^{(y_x)}$ represents the prediction probability corresponding to the ground-truth label $y_{x}$ . Previous studies [2, 8, 21] reveal that minimizing the CE loss forces the model to overfit to one-hot encoded ground-truth labels with almost $100\\%$ probability irrespective of their correctness, which can result in overconfidence.", + "bbox": [ + 89, + 662, + 483, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DML [29] is a model distillation technique that allows two models $f(\\theta)$ and $g(\\phi)$ to learn collaboratively through a combination of a supervised loss and a Kullback-Leibler (KL) Divergence:", + "bbox": [ + 89, + 756, + 483, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {f, x} (\\theta) = L _ {\\mathrm {C E}} \\left(f _ {x} (\\theta)\\right) + \\alpha D _ {\\mathrm {K L}} \\left(g _ {x} (\\phi) \\mid \\mid f _ {x} (\\theta)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 828, + 483, + 845 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g, x} (\\phi) = L _ {\\mathrm {C E}} \\left(g _ {x} (\\phi)\\right) + \\alpha D _ {\\mathrm {K L}} \\left(f _ {x} (\\theta) \\| g _ {x} (\\phi)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 883, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyperparameter used to adjust the weight of the KL divergence loss.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we designate $f(\\theta)$ as the primary model, and $g(\\phi)$ as the auxiliary model that does not participate in the inference process. It is important to note that DML was originally conceived as a knowledge distillation technique aimed at improving accuracy rather than focusing on calibration.", + "bbox": [ + 511, + 121, + 906, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. DML is an implicit form of deep ensembles", + "text_level": 1, + "bbox": [ + 511, + 220, + 879, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DML involves two models learning collaboratively by continuously exchanging knowledge during the training process. This collaborative learning encourages each model to learn from the strengths and correct the weaknesses of others, promoting diversity in model predictions. Consequently, DML achieves the advantages of DE—such as increased prediction accuracy and improved uncertainty calibration—without the need for independent model training or the associated computational costs.", + "bbox": [ + 511, + 241, + 906, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we explore the theoretical relationship between DML and DE.", + "bbox": [ + 511, + 378, + 906, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given two trained models $f(\\theta)$ and $g(\\phi)$ and an input $x$ for DE, we can get a weighted sum of predictive probabilities from $f_{x}(\\theta)$ and $g_{x}(\\phi)$ as the final prediction:", + "bbox": [ + 511, + 407, + 906, + 454 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {x} (\\theta , \\phi) = (1 - \\beta) f _ {x} (\\theta) + \\beta g _ {x} (\\phi), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 463, + 906, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta \\in [0,1]$ is a weight to balance the effect of each model. To explain the relationship between DML and DE, we first present a proposition, the proof is in Appendix C).", + "bbox": [ + 511, + 488, + 906, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1. Given two prediction probability distributions $f_{x}(\\theta)$ and $g_{x}(\\phi)$ for classifying sample $x$ , as well as their ensemble $h_{x}(\\underline{\\theta},\\phi)$ , we have", + "bbox": [ + 511, + 534, + 906, + 579 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial}{\\partial \\theta} D _ {K L} \\left(h _ {x} (\\underline {{\\theta}}, \\phi) | | f _ {x} (\\theta)\\right) = \\beta \\frac {\\partial}{\\partial \\theta} D _ {K L} \\left(g _ {x} (\\phi) | | f _ {x} (\\theta)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 587, + 905, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that the variables with underscores do not participate in gradient propagation.", + "bbox": [ + 511, + 630, + 905, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The left-hand side of (5) suggests that when we update $\\theta$ to make $f_{x}$ closer to $h_x$ , we temporarily treat $h_x$ as a model with fixed parameters, though, in the next step, $h_x$ will also be updated with the new $\\theta$ . By substituting (5) into (2) we obtain", + "bbox": [ + 511, + 659, + 905, + 734 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {f, x} (\\theta) = L _ {\\mathrm {C E}} \\left(f _ {x} (\\theta)\\right) + \\frac {\\alpha}{\\beta} D _ {\\mathrm {K L}} \\left(h _ {x} (\\underline {{\\theta}}, \\phi) | | f _ {x} (\\theta)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 739, + 906, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, the loss of model $g$ can be rewritten as", + "bbox": [ + 511, + 776, + 836, + 791 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g, x} (\\phi) = L _ {\\mathrm {C E}} \\left(g _ {x} (\\phi)\\right) + \\frac {\\alpha}{1 - \\beta} D _ {\\mathrm {K L}} \\left(h _ {x} \\left(\\theta , \\underline {{\\phi}}\\right) \\mid \\mid g _ {x} (\\phi)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 797, + 905, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From this perspective, we can see that KL divergence in DML helps continuously align each model's predictions with the ensemble distribution during training, thereby DML can be regarded as an implicit DE.", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25814", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Differentiated Deep Mutual Learning", + "text_level": 1, + "bbox": [ + 89, + 90, + 415, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model diversity plays a crucial role in the effectiveness of the ensemble [1, 30]. Although DML offers the opportunity to reduce computational overhead compared to DE, prior learning strategies in DML may destroy the model diversity and result in poor calibration performance. To address this issue, we propose an improved approach, Diff-DML, which increases the output discrepancy between $f$ and $g$ during training. Diff-DML employs two methods to increase model discrepancy: Differentiated Training Strategy (DTS) and Diversity-Preserving Learning Objective (DPLO).", + "bbox": [ + 89, + 112, + 483, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Differentiated Training Strategy: In traditional DML training [29], both models typically use the same scheduler, beginning with a specific learning rate and gradually decreasing it over time. This consistency in training strategies often leads to high similarity between the two models, diminishing their diversity. A straightforward way to increase the diversity between the outputs of the two models is to employ different training schedulers in training.", + "bbox": [ + 89, + 265, + 483, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In light of this, we propose to use different training strategies: one model, $f$ , is designated as the primary model (used solely during inference) and follows the conventional strategy of gradually decreasing the learning rate. The other model, $g$ , serves as an auxiliary model and is trained using a fixed learning rate throughout the training process. This discrepant learning rate scheme can preserve the diversity among models and thus improve the calibration performance. Diversity-Preserving Learning Objective: To increase the disparity between the two models, we propose a Diversity-Preserving Learning Objective by changing the loss function of the auxiliary model in DML. In traditional DML, the models $f(\\theta)$ and $g(\\phi)$ share the same form of loss function, which combines a supervised loss with KL divergence. This learning objective may cause the model to overfit the one-hot ground-truth labels, resulting in unavoidable overconfidence and insufficient diversity. To address this problem, we suggest that the primary model $f$ continue using the original loss function (2), while the auxiliary model $g(\\phi)$ omits the cross-entropy loss component. As a result, the auxiliary model no longer receives direct supervision from the labels. Concretely, the new loss function of model $g(\\phi)$ is", + "bbox": [ + 89, + 386, + 483, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g, x} (\\phi) = D _ {\\mathrm {K L}} \\left(f _ {x} (\\theta) | | g _ {x} (\\phi)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 729, + 483, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The KL divergence not only constructs a mutual feedback loop between the primary and auxiliary models but also assists the auxiliary model in learning a differentiated prediction distribution to boost model diversity.", + "bbox": [ + 89, + 758, + 485, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Extending to multiple auxiliary models", + "text_level": 1, + "bbox": [ + 89, + 830, + 426, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Naturally, we can incorporate multiple auxiliary models $\\{g_i\\}_{i=1}^{N_g}$ into the training process to further improve the prediction performance and calibration of the primary model $f$ .", + "bbox": [ + 89, + 853, + 485, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analogous to DE, leveraging multiple auxiliary models can effectively promote model diversity and improve the uncertainty calibration of the primary model. The optimization objectives for the primary model $f$ is", + "bbox": [ + 511, + 90, + 906, + 152 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {f, x} (\\theta) = L _ {\\mathrm {C E}} \\left(f _ {x} (\\theta)\\right) + \\alpha \\sum_ {i = 1} ^ {N _ {g}} \\frac {1}{N _ {g}} D _ {\\mathrm {K L}} \\left(g _ {i, x} \\left(\\phi_ {i}\\right) \\mid \\mid f _ {x} (\\theta)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 162, + 906, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For traditional DML, the optimization objectives for each $g_{i}$ are formalized as", + "bbox": [ + 511, + 219, + 903, + 247 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g _ {i, x}} \\left(\\phi_ {i}\\right) = L _ {\\mathrm {C E}} \\left(g _ {i, x} \\left(\\phi_ {i}\\right)\\right) + \\alpha D _ {\\mathrm {K L}} \\left(f _ {x} (\\theta) \\mid \\mid g _ {i, x} \\left(\\phi_ {i}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 260, + 906, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi_{i}$ denotes the trainable parameters of the auxiliary model $g_{i}$ . And for Diff-DML, the optimization objectives for each $g_{i}$ are formalized as", + "bbox": [ + 511, + 292, + 906, + 337 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g _ {i, x}} \\left(\\phi_ {i}\\right) = D _ {\\mathrm {K L}} \\left(f _ {x} (\\theta) \\mid \\mid g _ {i, x} \\left(\\phi_ {i}\\right)\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 348, + 906, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, we only use the primary model $f$ in inference.", + "bbox": [ + 511, + 377, + 885, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 406, + 645, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conducted primary experiments on image classification benchmarks, followed by experiments on misclassification and OOD detection to validate the performance of Diff-DML across different tasks.", + "bbox": [ + 511, + 431, + 906, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 511, + 501, + 700, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets: We trained and evaluated our multi-class classification model using CIFAR-10 [13], CIFAR-100 [13] and Tiny-ImageNet [3] datasets. Validation sets were created by randomly selecting 5,000 samples from each dataset's training set, with the remaining samples used for training. We experimented with two network architectures, ResNet34 and ResNet50 [9], on these datasets.", + "bbox": [ + 511, + 523, + 906, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation metrics: For standard classification, we evaluated all methods using accuracy (acc), ECE, and cw-ECE. For misclassification and OOD detection, we used four metrics: (1) False positive rate at $95\\%$ true positive rate (FPR- $95\\%$ ); (2) The misclassification probability when true positive rate is $95\\%$ (D-error); (3) Area under the receiver operating characteristic curve (AUROC); (4) Area under the precision-recall curve (AUPR).", + "bbox": [ + 511, + 628, + 908, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Selection of models for testing: When selecting the optimal calibration model, we need to evaluate both accuracy and calibration performance. It is insufficient to consider only accuracy or only calibration performance. To account for both aspects simultaneously, we have developed a new method for selecting models to be tested on the test set, which we refer to as the Comprehensive Selection Method for Accuracy and Calibration (CSMAC). The selection method is as follows: A threshold value for accuracy, denoted as $b$ , is set. When the highest accuracy of a model on the validation set is", + "bbox": [ + 511, + 750, + 910, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25815", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "below $b$ , the checkpoint with the highest accuracy is selected. When the highest accuracy exceeds $b$ , the checkpoint with the highest value of accuracy minus ECE (which we call the comprehensive evaluation score) on the validation set is selected. The value of $b$ varies depending on the dataset and base model. In the experiment, we first train a model optimized using cross-entropy loss and identify the checkpoint corresponding to the highest accuracy on the validation set. We set this checkpoint as the baseline model, and denote its accuracy on the validation set as $a$ . Then, we set $b = a - c$ , where $c$ represents a tolerance value for accuracy reduction, and in our experiment, $c$ is set to 0.01. In this way, we establish a selection mechanism that takes both accuracy and calibration performance into account.", + "bbox": [ + 89, + 90, + 485, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Competing methods and hyper-parameters: The evaluated methods include: Baseline: This is the model with the highest accuracy on the validation set trained with the CE loss. Note that the baseline model is used to determine the threshold $b$ for the CSMAC selection method. All other models use this threshold to make selections through CSMAC. Cross-Entropy (CE): This is the model trained with CE loss (1) selected by CSMAC. DML & DML-2: These refer to the DML method using 1 and 2 auxiliary ResNet18 models, respectively. The primary model used was either ResNet34 or ResNet50. For both DML-1 and DML-2, we report the performance of the primary models. Diff-DML & Diff-DML-2: Similar to DML-1/2, these use Diff-DML for training with 1 or 2 auxiliary ResNet18 models. The performance of the primary model, either ResNet34 or ResNet50, is reported. For DML and Diff-DML, the value $\\alpha \\in [0.4, 1.0]$ and we tuned it to achieve the best performance via the validation set. DCA [19]: Four values were used for its $\\beta$ : 0.5, 1.0, 5.0, and 10.0. MMCE [16]: Four values were used for its $\\lambda$ : 1.0, 2.0, 3.0, and 4.0. FL [21]: Three values were used for its $\\gamma$ : 1.0, 2.0, and 3.0. FL+MDCA [10]: Three values were used for the $\\gamma$ in the focal loss: 1.0, 2.0, and 3.0, while the $\\beta$ value in MDCA was 1.0. FL+DTDP [25]: Three values were used for the $\\gamma$ in the focal loss: 1.0, 2.0, and 3.0. For DTDP, we used an $\\alpha$ value of 0.005 and experimented with $\\lambda$ values of 5.0 and 10.0, along with a data pruning ratio of $10\\%$ . DE-3 [5, 17]: It denotes DE with 3 models using cross-entropy loss with different initializations and combined their produced probabilities.", + "bbox": [ + 91, + 306, + 486, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Other implementation details: In the experiments, we utilized the SGD optimizer with a momentum of 0.9 and weight decay of 0.0005. The initial learning rate was set to 0.1 for all datasets. For each dataset, we trained the models for a total of 250 epochs. During the initial 100 epochs, we maintained the initial learning rate. Afterward, every 50 epochs, we decreased the learning rate by a factor of $10\\%$ . In Diff-DML, the primary model $f$ employed the same optimization strategy as described above. However, for the", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "auxiliary model $g$ , we initialized the learning rate as 0.01 and did not apply any learning rate decay. All experiments were performed with a batch size of 100 and conducted for training and testing on a single GeForce RTX 2080 GPU.", + "bbox": [ + 511, + 90, + 908, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Classification accuracy and calibration", + "text_level": 1, + "bbox": [ + 511, + 162, + 848, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 shows the comparison results of Diff-DML and various single-model calibration methods on CIFAR-100. Tables S1 and S2 in Appendix D show the comparison results on CIFAR-10 and Tiny/ImageNet, respectively. Figure S1 in Appendix E visually demonstrates the calibration performance of Diff-DML with reliability diagrams and confidence histograms.", + "bbox": [ + 511, + 186, + 906, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison between DML and Diff-DML: From Tables 1, S1 and S2, it can be observed that on all three datasets, DML and Diff-DML achieved improvements in both accuracy and calibration performance compared to CE. The improvements of Diff-DML were more significant than those of DML.", + "bbox": [ + 511, + 292, + 908, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 shows the validation accuracy and ECE curves of ResNet34 during training on the CIFAR-100 dataset using CE, DML and Diff-DML methods. It can be seen that the ECE values for CE and DML steadily increased throughout training, while the ECE values for Diff-DML showed a downward trend.", + "bbox": [ + 511, + 369, + 908, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with regularization methods: It can be observed that the majority of regularization methods compromised a certain level of accuracy to improve calibration ability. For instance, a previous state-of-the-art method, FL+MDCA, reduced accuracy by 1 to 2 percent. In contrast, our Diff-DML method avoided this trade-off and demonstrates superior accuracy and calibration ability.", + "bbox": [ + 511, + 460, + 910, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In terms of calibration performance, Diff-DML surpassed FL+MDCA across all three datasets. Specifically, compared to FL+MDCA, Diff-DML achieved an absolute accuracy increase of $1.3\\% / 3.1\\%$ for ResNet34/ResNet50 on the CIFAR-100 dataset, with relative reductions in ECE of $49.6\\% / 43.8\\%$ and relative reductions in cw-ECE of $7.7\\% / 13.0\\%$ representing a considerable improvement.", + "bbox": [ + 511, + 566, + 908, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with post-hoc calibration methods: Post-hoc calibration methods in general did not affect accuracy much, but their calibration performance on datasets with many classes, like CIFAR-100 and Tiny-[ImageNet, was often suboptimal. In contrast, Diff-DML showed much better calibration performance on these datasets.", + "bbox": [ + 511, + 672, + 908, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results with more auxiliary models: Table 2 shows the results of DML and Diff-DML trained with more auxiliary models on the CIFAR-100 dataset, compared to DE-3. Table S3 in Appendix D presents the comparison on CIFAR-10 and Tiny/ImageNet. It can be observed that DE-3 provided the most significant improvement in accuracy. DML and Diff-DML with two auxiliary models slightly outperformed those with one auxiliary model, but the difference was not substantial. In terms of calibration performance, Diff-DML", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25816", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e2da281170812d5b061917d1e135a9b1a401cee25879aca50e003f4baea8fa78.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAcc(%)↑ECE(%)↓cw-ECE(10-3)↓
ResNet34ResNet50ResNet34ResNet50ResNet34ResNet50
Baseline77.8777.858.8110.392.252.51
CE77.7377.548.5510.042.292.44
Post-hoc calibration methods
CE+TS [8]77.7377.543.373.141.721.63
CE+VS [8]77.4777.333.763.131.821.71
CE+DS [15]77.3377.543.203.141.711.63
Regularization based calibration methods
DCA [19]77.5377.6610.1011.692.432.66
MMCE [16]75.5075.166.267.191.851.98
FL [21]76.9376.491.211.791.561.55
FL+DTDP [25]75.8776.091.822.611.601.65
FL+MDCA [10]76.9175.931.291.531.551.61
DML-based calibration methods
DML [29]77.3677.893.434.421.561.69
Diff-DML (ours)78.2079.050.650.861.431.40
", + "bbox": [ + 130, + 88, + 867, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e1df126f479648a8e8a414180bd5c8f73d5f3605bc716d8db6c500d0766a6a53.jpg", + "image_caption": [ + "Figure 2. The validation accuracy and ECE curves of ResNet34 trained using the (a) CE, (b) DML, and (c) Diff-DML methods on the CIFAR-100 dataset during the training process." + ], + "image_footnote": [], + "bbox": [ + 98, + 417, + 361, + 573 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7dfa6e90ddc024001f2dfd21a7eb6bb502e910b54d4bb390809ac875eda6eb8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 415, + 630, + 571 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3cf201f15679ad64253e8ec6da9a0f1086430f20beebe87a638c62855456145.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 415, + 901, + 571 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/92297552382d178d5be9cd816f5a9b818af1d20554a6e3ae818c68e74ee00959.jpg", + "table_caption": [ + "Table 1. Different calibration methods' accuracy, ECE, and cw-ECE on the CIFAR-100 dataset." + ], + "table_footnote": [], + "table_body": "
MethodAcc(%)↑ECE(%)↓cw-ECE(10-3)↓
ResNet34ResNet50ResNet34ResNet50ResNet34ResNet50
DE-3 [5, 17]79.3479.632.954.401.561.73
DML [29]77.3677.893.434.421.561.69
DML-2 [29]77.9878.004.324.381.611.82
Diff-DML (ours)78.2079.050.650.861.431.40
Diff-DML-2 (ours)79.1479.300.750.721.371.33
", + "bbox": [ + 197, + 622, + 802, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of DML and Diff-DML trained with more auxiliary models against DE on CIFAR-100 dataset.", + "bbox": [ + 153, + 744, + 841, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "was clearly superior to DE-3.", + "bbox": [ + 89, + 786, + 285, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training and inference time: In terms of time overhead, training Diff-DML with one auxiliary model took 1.4 times longer than training a single model, and with two auxiliary models, it took 1.9 times longer. In contrast, DE-3's training time was 3 times that of a single model. For inference, Diff-DML required the same time as a single model, while", + "bbox": [ + 89, + 809, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DE-3's inference time was 3 times longer. Thus, our method provided substantial advantages over DE-3 in both training and inference time.", + "bbox": [ + 511, + 786, + 906, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Analysis of the robustness of Diff-DML", + "text_level": 1, + "bbox": [ + 511, + 845, + 849, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sensitivity of $\\alpha$ : We investigated the sensitivity of the important hyper-parameter $\\alpha$ in (2). From Figure 3, it is ob", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25817", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cb40b5b709e0ee8deedc36f94ab497bfad69e5d264bc03d029c04f29338d7864.jpg", + "image_caption": [ + "Figure 3. Variation of accuracy, ECE, and classwise-ECE of Diff-DML method trained on CIFAR-100 with different values of hyperparameter $\\alpha$ ." + ], + "image_footnote": [], + "bbox": [ + 96, + 78, + 374, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41c57e78b6e34e1c22078fe45e9ed2fdc397a82329ceee0d612f5cc5fa6996be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 78, + 640, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/087a0df88e10a212073ed47a9c748ecfae1caa92daf199130d86cd4c0c71c897.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 78, + 903, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "served that on CIFAR-100, as $\\alpha$ changes, both accuracy and cw-ECE exhibited smooth variations. Specifically, the fluctuations in accuracy remained within 3 percentage points, while the fluctuations in cw-ECE were within 0.02 percentage points. As for ECE, although its values ranged between 2 and 4 for very small (e.g., 0.4) and large (e.g., 1.0) $\\alpha$ values, the fluctuations around the lowest point of ECE were also minimal. This demonstrates the good robustness of Diff-DML with respect to the hyper-parameter $\\alpha$ .", + "bbox": [ + 88, + 287, + 482, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Selection of auxiliary model: We investigated how to select the auxiliary model. We experimented with CNNs containing two hidden layers, as well as ResNet18, ResNet34, and ResNet50 as backbones for the auxiliary model, which is shown in Table S4 in Appendix F. It can be observed that only when using CNNs, both accuracy and calibration ability were significantly poor. When using ResNet18, ResNet34, and ResNet50, similar performance were observed. This indicates that in order to ensure good performance of the Diff-DML, a backbone for the auxiliary model should not be selected if its performance is significantly worse than that of the primary model. Furthermore, when performance is similar, specific model architectures will yield similar performance.", + "bbox": [ + 89, + 425, + 483, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 648, + 241, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 3, we conducted an ablation study on the performance of the two strategies, DTS and DPLO, used in Diff-DML, using ResNet50 on the CIFAR-100 dataset. In the ablation study of DPLO, we set the learning rates and schedulers of models $f$ and $g$ to be the same as those in DML. As shown in the table, Diff-DML with only DTS or only DPLO both achieve better accuracy and calibration performance compared to the original DML, with the best results obtained when both strategies are combined.", + "bbox": [ + 89, + 670, + 483, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Misclassification & OOD detection", + "text_level": 1, + "bbox": [ + 89, + 816, + 393, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A real-world application of calibrated predictions is making the model aware of its uncertainties. Specifically, the model should show low confidence for misclassified or out-of-distribution samples. Motivated by this, we conducted", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/85dfdd6a7f9335fe4226dc4c2fb3c94ea307babd14473a2eefea60f66d53cef8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAcc(%)↑ECE(%)↓cw-ECE(%)↓
CE77.5410.042.44
DML [29]77.894.421.69
DTS78.101.381.43
DPLO78.751.011.40
DTS+DPLO79.050.861.40
", + "bbox": [ + 535, + 279, + 885, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study of Diff-DML on CIFAR-100.", + "bbox": [ + 549, + 390, + 867, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "experiments on misclassification and OOD detection, measuring the success rate of detecting such samples using maximum softmax probability [11]. We report threshold-free metrics (FPR-95%, D-error, AUROC, and AUPR).", + "bbox": [ + 511, + 409, + 906, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Misclassification detection: Misclassification detection is an important problem in machine learning, as it allows for the identification of instances where the model's predictions are unreliable. We expect the model to output low probability or high uncertainty for samples that cannot be classified correctly. Table 4 presents the performance results for various models in detecting misclassifications. Diff-DML showed significant improvements over other single-model calibration techniques and the DE method.", + "bbox": [ + 511, + 469, + 905, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "OD detection: A reliable classification model should show higher uncertainty and lower confidence when encountering test samples that differ significantly from the training data. We evaluated the ability of different calibration methods to distinguish OOD samples by mixing in-distribution test data with OOD data. We tested models trained on CIFAR-10 and CIFAR-100 for two types of OOD detection [5, 12]: far OOD (distinguishing SVHN [24] from CIFAR-10/CIFAR-100) and near OOD (distinguishing between CIFAR-10 and CIFAR-100). The results in Table 5 show that Diff-DML significantly outperformed other calibration methods, even exceeding the DE method, which is known for its OOD detection effectiveness.", + "bbox": [ + 511, + 604, + 906, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Conclusion and discussion", + "text_level": 1, + "bbox": [ + 511, + 814, + 758, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although Diff-DML significantly enhances both accuracy and calibration, its reliance on an auxiliary model with performance comparable to the primary model presents challenges, especially when training models with large parameter", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25818", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fcf9b0eaeb6ac7a5f7da5f60fab30ca43d0cac8d044184270a7d7a13ac14b06c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodFPR-95%↓D-error↓AUROC↑AUPR↑
CIFAR-10CE38.5313.2091.8099.46
DCA [19]38.9713.2891.9499.47
MMCE [16]23.0012.6193.5999.54
FL [21]33.1413.5692.2399.46
FL+DTDP [25]28.3414.2692.2099.36
FL+MDCA [10]32.8013.2192.4199.47
DE-3 [5, 17]22.3710.8994.0899.68
DML [29]19.2211.4994.0599.64
Diff-DML(ours)16.6310.0094.8799.73
CIFAR-100CE45.9520.1986.9395.47
DCA [19]40.6620.0986.8595.01
MMCE [16]45.9720.0287.0095.05
FL [21]45.0220.2187.0495.42
FL+DTDP [25]43.9520.7386.7995.16
FL+MDCA [10]45.0820.6886.5695.36
DE-3 [5, 17]40.0419.2588.1996.44
DML [29]40.3119.4287.8796.17
Diff-DML(ours)39.6218.9588.3896.50
", + "bbox": [ + 243, + 88, + 751, + 375 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/578bf63902fc1be55bc7680941b7c937223c6543855236685898bd96da492d85.jpg", + "table_caption": [ + "Table 4. The misclassification detection performance of different calibration methods." + ], + "table_footnote": [], + "table_body": "
DatasetMethodFPR-95%↓D-error↓AUROC↑AUPR↑
In-Distribution: CIFAR-10CE69.48/19.8019.45/10.5986.11/94.6082.00/72.54
DCA [19]64.63/18.3818.80/10.1087.07/95.0584.36/72.48
MMCE [16]40.40/11.9317.33/8.3789.13/97.5389.61/88.85
FL [21]66.13/19.2020.97/10.6985.57/94.2882.81/74.18
FL+DTDP [25]50.54/23.6220.38/13.5486.57/92.0386.73/77.10
Out-of-Distribution: CIFAR-100/SVHNFL+MDCA [10]72.19/29.8420.75/13.5884.77/92.6381.01/67.11
DE-3 [5, 17]33.89/13.3115.60/9.0790.12/95.0690.27/86.53
DML [29]32.42/13.8515.90/9.3391.05/95.3891.67/86.04
Diff-DML(ours)30.12/10.5115.06/7.6591.87/96.3392.67/90.29
In-Distribution: CIFAR-100CE68.26/45.8728.69/22.2777.30/85.4477.49/54.01
DCA [19]69.41/47.8830.58/23.6275.96/83.2077.19/55.41
MMCE [16]65.34/49.3829.53/25.7776.58/77.6478.66/49.53
FL [21]67.27/48.4028.10/23.1778.36/83.8978.87/53.39
FL+DTDP [25]59.10/53.8527.23/24.1679.07/83.1081.05/49.61
Out-of-Distribution: CIFAR-10/SVHNFL+MDCA [10]69.52/41.6627.93/19.8778.35/87.4677.92/59.08
DE-3 [5, 17]56.21/48.7026.61/24.2080.36/83.4282.34/53.17
DML [29]62.08/37.9527.07/19.6879.90/87.0681.07/65.32
Diff-DML(ours)56.21/37.8425.55/19.5881.50/87.3483.57/65.65
", + "bbox": [ + 189, + 411, + 808, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. The OOD detection performance of different calibration methods. The numbers before and after the backslash denote the near and far OOD detection results, respectively.", + "bbox": [ + 89, + 705, + 906, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sizes. This dependency can complicate scalability and computational efficiency, particularly in resource-constrained environments.", + "bbox": [ + 89, + 760, + 485, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, we introduce Diff-DML, which overcomes the limitations of traditional DML by incorporating two strategies: DTS and DPLO. While DML promotes learning from an ensemble distribution, it often reduces prediction diversity, which is crucial for effective calibration. Diff-DML preserves this diversity while simultaneously", + "bbox": [ + 89, + 816, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "improving both accuracy and calibration. Extensive experiments on datasets such as CIFAR-100 demonstrate that our approach outperforms state-of-the-art methods, offering a practical and efficient solution for reliable uncertainty estimation, particularly in safety-critical applications.", + "bbox": [ + 511, + 760, + 908, + 833 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25819", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Gavin Brown. Diversity in neural network ensembles. PhD thesis, CiteSeer, 2004. 4", + "[2] Peng Cui, Dan Zhang, Zhijie Deng, Yinpeng Dong, and Jun Zhu. Learning sample difficulty from pre-trained models for reliable prediction. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4", + "[4] James M Dolezal, Andrew Srisuwananukorn, Dmitry Karpeyev, Siddhi Ramesh, Sara Kochanny, Brittany Cody, Aaron S Mansfield, Sagar Rakshit, Radhika Bansal, Melanie C Bois, et al. Uncertainty-informed deep learning models enable high-confidence predictions for digital histopathology. Nature communications, 13(1):6572, 2022. 1", + "[5] Stanislav Fort, Huiyi Hu, and Balaji Lakshminarayanan. Deep ensembles: A loss landscape perspective. arXiv preprint arXiv:1912.02757, 2019. 2, 5, 6, 7, 8, 1", + "[6] Kevin P Greenman, Ava P Amini, and Kevin K Yang. Benchmarking uncertainty quantification for protein engineering. bioRxiv, pages 2023-04, 2023. 1", + "[7] Arthur Gretton. Introduction to rkhs, and some simple kernel algorithms. Adv. Top. Mach. Learn. Lecture Conducted from University College London, 16(5-3):2, 2013. 1", + "[8] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 2, 3, 6", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[10] Ramya Hebbalaguppe, Jatin Prakash, Neelabh Madan, and Chetan Arora. A stitch in time saves nine: A train-time regularizing loss for improved neural network calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16081-16090, 2022. 2, 3, 5, 6, 8, 1", + "[11] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2016. 7", + "[12] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joe Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. 7", + "[13] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 4", + "[14] Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. In Artificial Intelligence and Statistics, pages 623-631. PMLR, 2017. 1", + "[15] Meelis Kull, Miquel Perello Nieto, Markus Kangsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with dirichlet calibration. Advances in neural information processing systems, 32, 2019. 2, 3, 6, 1", + "[16] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In International Conference on Machine Learning, pages 2805-2814. PMLR, 2018. 2, 3, 5, 6, 8, 1", + "[17] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 2, 5, 6, 8, 1", + "[18] Christian Leibig, Vaneeda Allken, Murat Seçkin Ayhan, Philipp Berens, and Siegfried Wahl. Leveraging uncertainty information from deep neural networks for disease detection. Scientific reports, 7(1):1-14, 2017. 1", + "[19] Gongbo Liang, Yu Zhang, Xiaoqin Wang, and Nathan Jacobs. Improved trainable calibration method for neural networks on medical imaging classification. arXiv preprint arXiv:2009.04057, 2020.3,5,6,8,1", + "[20] Rhiannon Michelmore, Matthew Wicker, Luca Laurenti, Luca Cardelli, Yarin Gal, and Marta Kwiatkowska. Uncertainty quantification with statistical guarantees in end-to-end autonomous driving control. In International Conference on Robotics and Automation, 2020. 1", + "[21] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. Advances in Neural Information Processing Systems, 33:15288-15299, 2020. 3, 5, 6, 8, 1", + "[22] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 2, 3, 1", + "[23] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Proceedings of the AAAI conference on artificial intelligence, 2015. 2", + "[24] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning, page 7. Granada, Spain, 2011. 7", + "[25] Rishabh Patra, Ramya Hebbalaguppe, Tirtharaj Dash, Gautam Shroff, and Lovekesh Vig. Calibrating deep neural networks using explicit regularisation and dynamic data pruning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1541-1549, 2023. 3, 5, 6, 8, 1", + "[26] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey Hinton. Regularizing neural networks by penalizing confident output distributions. arXiv preprint arXiv:1701.06548, 2017. 2, 3, 1", + "[27] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61-74, 1999. 2, 3, 1", + "[28] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25820", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conference on computer vision and pattern recognition, pages 2818-2826, 2016. 2, 3, 1", + "[29] Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4320-4328, 2018. 2, 3, 4, 6, 7, 8", + "[30] Zhi-Hua Zhou. Ensemble methods: foundations and algorithms. CRC press, 2012. 4" + ], + "bbox": [ + 91, + 92, + 483, + 205 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25821", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/6a2c48a3-aafc-4ad2-9262-8700829f154e_model.json b/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/6a2c48a3-aafc-4ad2-9262-8700829f154e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b93d50aa3289d35e93e60a8d396e3f1dd54b1022 --- /dev/null +++ b/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/6a2c48a3-aafc-4ad2-9262-8700829f154e_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.13, + 0.9, + 0.155 + ], + "angle": 0, + "content": "Improving Accuracy and Calibration via Differentiated Deep Mutual Learning" + }, + { + "type": "text", + "bbox": [ + 0.201, + 0.179, + 0.796, + 0.216 + ], + "angle": 0, + "content": "Han Liu\\(^{1,2,\\dagger}\\), Peng Cui\\(^{1,\\dagger}\\), Binging Wang\\(^{2}\\), Weipeng Chen\\(^{2}\\), Yupeng Zhang\\(^{2}\\), Jun Zhu\\(^{1,*}\\), and Xiaolin Hu\\(^{1,3,4,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.217, + 0.766, + 0.252 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Computer Science and Technology, Institute for AI, BNRist, Tsinghua University, Beijing 100084, China" + }, + { + "type": "text", + "bbox": [ + 0.439, + 0.253, + 0.558, + 0.268 + ], + "angle": 0, + "content": "2Baichuan Inc." + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.269, + 0.721, + 0.288 + ], + "angle": 0, + "content": "\\(^{3}\\)Tsinghua Laboratory of Brain and Intelligence (THBI)," + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.288, + 0.851, + 0.305 + ], + "angle": 0, + "content": "IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing 100084, China" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.305, + 0.773, + 0.323 + ], + "angle": 0, + "content": "\\(^{4}\\)Chinese Institute for Brain Research (CIBR), Beijing 100010, China" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.325, + 0.68, + 0.34 + ], + "angle": 0, + "content": "{han-liu18, cuip22}@mails.tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.343, + 0.724, + 0.358 + ], + "angle": 0, + "content": "{daniel, chenweipeng, zhangyupeng}@baichuan-inc.com" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.361, + 0.649, + 0.375 + ], + "angle": 0, + "content": "{dcszj, xlhu}@mail.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.41, + 0.327, + 0.426 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.443, + 0.486, + 0.837 + ], + "angle": 0, + "content": "Deep Neural Networks (DNNs) have achieved remarkable success in a variety of tasks, particularly in terms of prediction accuracy. However, in real-world scenarios, especially in safety-critical applications, accuracy alone is insufficient; reliable uncertainty estimates are essential. Modern DNNs, often trained with cross-entropy loss, tend to exhibit overconfidence, especially on ambiguous samples. Many techniques aim to improve uncertainty calibration, yet they often come at the cost of reduced accuracy or increased computational demands. To address this challenge, we propose Differentiated Deep Mutual Learning (Diff-DML), an efficient ensemble approach that simultaneously enhances accuracy and uncertainty calibration. Diff-DML draws inspiration from Deep Mutual Learning (DML) while introducing two strategies to maintain prediction diversity: (1) Differentiated Training Strategy (DTS) and (2) Diversity-Preserving Learning Objective (DPLO). Our theoretical analysis shows that Diff-DML's diversified learning framework not only leverages ensemble benefits but also avoids the loss of prediction diversity observed in traditional DML setups, which is crucial for improved calibration. Extensive evaluations on various benchmarks confirm the effectiveness of Diff-DML. For instance, on the CIFAR-100 dataset, Diff-DML on ResNet34/50 models achieved substantial improvements over the previous state-of-the-art method, MDCA, with absolute accuracy gains of \\(1.3\\% / 3.1\\%\\), relative ECE reductions" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.412, + 0.871, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.604, + 0.908, + 0.646 + ], + "angle": 0, + "content": "Figure 1. Scatter plot of ECE and accuracy for different calibration methods based on the ResNet50 backbone on the CIFAR-100 test set. Methods located closer to the top left corner perform better." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.707 + ], + "angle": 0, + "content": "of \\(49.6\\% / 43.8\\%\\), and relative classwise-ECE reductions of \\(7.7\\% / 13.0\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.738, + 0.646, + 0.754 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Deep Neural Networks (DNNs) have achieved great success in various tasks, especially in terms of prediction accuracy. In safety-critical areas such as medical diagnosis [4, 18], self-driving cars [20], and protein engineering [6], accurate uncertainty estimation (i.e., well-calibrated prediction) is crucial for reliable model deployment. For instance, a reliable model needs to show high uncertainty when making incorrect predictions. However, models trained with cross-entropy (CE) loss tend to be overconfident, especially when" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.888 + ], + "angle": 0, + "content": "This work was supported by the National Key Research and Development Program of China (No. 2021ZD0200301) and the National Natural Science Foundation of China (No. U2341228)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.888, + 0.353, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\)Equal contribution. \\*Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25812" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "encountering ambiguous samples, leading to predictions that are often close to \\(100\\%\\) confident and producing undesirable overconfidence [2, 8]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.14, + 0.486, + 0.457 + ], + "angle": 0, + "content": "To address this issue, various regularization techniques have been explored during training, referred to as training-time calibration. For instance, entropy regularization (ER) [26] and label smoothing (LS) [22, 28] promote smoother predictive probability distributions to alleviate overconfidence by maximizing entropy or adjusting label distributions. Some methods aim to directly reduce calibration error during training to achieve multi-class calibration. For example, MMCE [16] introduces a trainable measure based on kernel mean embeddings, while MDCA [10] employs a differentiable auxiliary loss to minimize the gap between prediction confidence and accuracy. However, despite mitigating overconfidence, these techniques often compromise prediction accuracy due to excessive regularization. Another category of methods is post-processing, which adjusts output probabilities using Temperature Scaling (TS) [27], Vector Scaling (VS) [27], or Dirichlet Scaling (DS) [15]. However, these post-processing techniques typically rely on an additional validation dataset. Furthermore, they have limited learnable parameters, which may lead to suboptimal calibration performance [10]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.46, + 0.485, + 0.732 + ], + "angle": 0, + "content": "Deep Ensembles (DE) [5, 17] offer a straightforward solution by combining multiple randomly initialized trained DNNs to improve prediction accuracy and uncertainty calibration. Yet, using multiple DNNs during training and inference significantly increases computational costs, which poses challenges for deploying DE in real-world scenarios. To address this limitation, we draw inspiration from Deep Mutual Learning (DML) [29], a technique used in model distillation, where an ensemble of student models learn collaboratively and teach each other throughout the training process. This work provides the theoretical analysis showing that DML encourages each model to learn from an ensemble distribution through a mimicry loss. The proposed theory motivates us to leverage DML as an implicit form of DE, aiming to achieve similar gains in accuracy and calibration as DE, while mitigating the high computational overhead because DML uses only one model for prediction during inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.902 + ], + "angle": 0, + "content": "However, in standard DML, although the models start from different initial states, they become overly similar during mutual learning, which reduces the diversity of their outputs—a key factor needed for effective ensemble-based calibration. In this work, we propose Differentiated Deep Mutual Learning (Diff-DML), a novel approach that enhances the benefits of DML by promoting prediction diversity between models to achieve high accuracy and strong calibration performance. Diff-DML incorporates two key strategies that preserve output diversity: (1) Differentiated Training Strategy (DTS), which uses different learning rates and schedules" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "for different models to prevent premature convergence, and (2) Diversity-Preserving Learning Objective (DPLO), which removes cross-entropy loss for auxiliary models to avoid overfitting to the one-hot ground-truth labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.909, + 0.362 + ], + "angle": 0, + "content": "These strategies help Diff-DML balance collaborative learning with prediction diversity, resulting in better uncertainty calibration and enhanced predictive performance. Extensive evaluations on various benchmarks, including image classification, misclassification detection, and out-of-distribution (OOD) detection, confirm Diff-DML's effectiveness. For example, on the CIFAR-100 dataset, Diff-DML on ResNet34/50 achieved significant improvements compared to the previous state-of-the-art method, MDCA [10], with absolute accuracy increases of \\(1.3\\% / 3.1\\%\\), relative reductions in ECE of \\(49.6\\% / 43.8\\%\\), and relative reductions in classwise-ECE of \\(7.7\\% / 13.0\\%\\). The latter two metrics measure calibration performance, where lower values indicate better results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.364, + 0.909, + 0.44 + ], + "angle": 0, + "content": "Figure 1 displays ECE and accuracy of different calibration methods, showing Diff-DML achieving favorable calibration and accuracy. Notably, Diff-DML incurs only minimal computational cost at inference, unlike DE, which requires multiple models to generate a single prediction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.453, + 0.792, + 0.47 + ], + "angle": 0, + "content": "2. Background and related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.908, + 0.584 + ], + "angle": 0, + "content": "Calibration, the process of ensuring that the predicted probabilities reflect the true likelihood of the events, plays a crucial role in machine learning models. It ensures that the prediction of machine learning models are not only accurate but also well-calibrated. Well-calibrated models provide reliable uncertainty estimates, which can help users and downstream systems make safer decisions by avoiding overconfidence." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.584, + 0.908, + 0.629 + ], + "angle": 0, + "content": "There are two fundamental types of calibration: confidence calibration [23] and classwise calibration [15]. In this paper, we focus on both types of calibration." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.63, + 0.909, + 0.735 + ], + "angle": 0, + "content": "Confidence calibration: Confidence calibration refers to the alignment of predicted confidence with observed frequencies. It ensures that a model's prediction confidence accurately reflects the true likelihood of an event occurring. A widely used metric for evaluating confidence calibration is the Expected Calibration Error (ECE) [23]. The definition of ECE can be found in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Classwise calibration: Classwise calibration is concerned with calibrating the predicted probabilities for each individual class in a multi-class classification problem. It ensures that the predicted probabilities across different classes are accurate and properly calibrated. The metric for classwise calibration is classwise-ECE (cw-ECE) [15], which can be found in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.908, + 0.886 + ], + "angle": 0, + "content": "According to the number of models used for inference, calibration methods can be divided into single-model calibration methods and DE [5, 17]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Single-model calibration methods: Within single-model" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25813" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.288 + ], + "angle": 0, + "content": "calibration methods, there are two categories: train-time calibration methods and post-hoc calibration methods. Train-time calibration methods typically involve designing loss functions or incorporating additional regularization techniques, such as Entropy Regularization (ER) [26], Label Smoothing (LS) [22, 28], Focal Loss (FL) [21], Maximum Mean Calibration Error (MMCE) [16], Difference between Confidence and Accuracy (DCA) [19], Multi-class Difference in Confidence and Accuracy (MDCA) [10], and Dynamic Train-time Data Pruning (DTDP) [25]. Post-hoc calibration methods are applied after the model is trained, such as Temperature Scaling (TS) [27], Vector Scaling, and Dirichlet Scaling (DS) [15]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.484, + 0.394 + ], + "angle": 0, + "content": "Deep Ensembles: DE creates an ensemble by training multiple models independently with different initializations. To alleviate overconfidence and improve calibration, during inference, DE leverages \\(N\\) randomly initialized, individually trained DNNs and makes predictions by average voting. While effective for enhancing model performance, DE introduces additional computational overhead during inference." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.394, + 0.438, + 0.409 + ], + "angle": 0, + "content": "More related works are discussed in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.423, + 0.182, + 0.437 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.59 + ], + "angle": 0, + "content": "Let \\( P_{data} \\) denote the probability distribution of data. A dataset consists of finite \\( (x, y_x) \\) pairs where \\( x \\sim P_{data} \\) is a data instance and \\( y_x \\in \\mathcal{Y} = \\{1, 2, \\dots, K\\} \\) is the ground-truth class label of \\( x \\). A model takes in the instance \\( x \\) and outputs its prediction vector \\( p_x \\in \\mathbb{R}^K \\). Let \\( p_x^{(k)} \\) denote the \\( k \\)-th dimension of \\( p_x \\). The confidence \\( \\hat{p}_x = \\max_{k \\in \\mathcal{Y}} \\{p_x^{(k)}\\} \\) is the maximum value in the probability vector of predictions, and the model's prediction \\( \\hat{y}_x = \\operatorname{argmax}_{k \\in \\mathcal{Y}} \\{p_x^{(k)}\\} \\) is the index corresponding to the maximum value." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.621 + ], + "angle": 0, + "content": "Typically, we train a model by minimizing the following CE loss on the training set," + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.632, + 0.483, + 0.651 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {C E}} \\left(p _ {x}\\right) = - \\log p _ {x} ^ {\\left(y _ {x}\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.484, + 0.756 + ], + "angle": 0, + "content": "where \\( p_x^{(y_x)} \\) represents the prediction probability corresponding to the ground-truth label \\( y_{x} \\). Previous studies [2, 8, 21] reveal that minimizing the CE loss forces the model to overfit to one-hot encoded ground-truth labels with almost \\( 100\\% \\) probability irrespective of their correctness, which can result in overconfidence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.484, + 0.817 + ], + "angle": 0, + "content": "DML [29] is a model distillation technique that allows two models \\( f(\\theta) \\) and \\( g(\\phi) \\) to learn collaboratively through a combination of a supervised loss and a Kullback-Leibler (KL) Divergence:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.829, + 0.484, + 0.847 + ], + "angle": 0, + "content": "\\[\nL _ {f, x} (\\theta) = L _ {\\mathrm {C E}} \\left(f _ {x} (\\theta)\\right) + \\alpha D _ {\\mathrm {K L}} \\left(g _ {x} (\\phi) \\mid \\mid f _ {x} (\\theta)\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.885, + 0.484, + 0.903 + ], + "angle": 0, + "content": "\\[\nL _ {g, x} (\\phi) = L _ {\\mathrm {C E}} \\left(g _ {x} (\\phi)\\right) + \\alpha D _ {\\mathrm {K L}} \\left(f _ {x} (\\theta) \\| g _ {x} (\\phi)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyperparameter used to adjust the weight of the KL divergence loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.907, + 0.212 + ], + "angle": 0, + "content": "In this paper, we designate \\( f(\\theta) \\) as the primary model, and \\( g(\\phi) \\) as the auxiliary model that does not participate in the inference process. It is important to note that DML was originally conceived as a knowledge distillation technique aimed at improving accuracy rather than focusing on calibration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.221, + 0.88, + 0.237 + ], + "angle": 0, + "content": "3.1. DML is an implicit form of deep ensembles" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.242, + 0.908, + 0.378 + ], + "angle": 0, + "content": "DML involves two models learning collaboratively by continuously exchanging knowledge during the training process. This collaborative learning encourages each model to learn from the strengths and correct the weaknesses of others, promoting diversity in model predictions. Consequently, DML achieves the advantages of DE—such as increased prediction accuracy and improved uncertainty calibration—without the need for independent model training or the associated computational costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.407 + ], + "angle": 0, + "content": "In this section, we explore the theoretical relationship between DML and DE." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.409, + 0.907, + 0.455 + ], + "angle": 0, + "content": "Given two trained models \\( f(\\theta) \\) and \\( g(\\phi) \\) and an input \\( x \\) for DE, we can get a weighted sum of predictive probabilities from \\( f_{x}(\\theta) \\) and \\( g_{x}(\\phi) \\) as the final prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.464, + 0.907, + 0.481 + ], + "angle": 0, + "content": "\\[\nh _ {x} (\\theta , \\phi) = (1 - \\beta) f _ {x} (\\theta) + \\beta g _ {x} (\\phi), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.907, + 0.535 + ], + "angle": 0, + "content": "where \\(\\beta \\in [0,1]\\) is a weight to balance the effect of each model. To explain the relationship between DML and DE, we first present a proposition, the proof is in Appendix C)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.535, + 0.908, + 0.58 + ], + "angle": 0, + "content": "Proposition 1. Given two prediction probability distributions \\( f_{x}(\\theta) \\) and \\( g_{x}(\\phi) \\) for classifying sample \\( x \\), as well as their ensemble \\( h_{x}(\\underline{\\theta},\\phi) \\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.588, + 0.906, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial}{\\partial \\theta} D _ {K L} \\left(h _ {x} (\\underline {{\\theta}}, \\phi) | | f _ {x} (\\theta)\\right) = \\beta \\frac {\\partial}{\\partial \\theta} D _ {K L} \\left(g _ {x} (\\phi) | | f _ {x} (\\theta)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.631, + 0.906, + 0.66 + ], + "angle": 0, + "content": "Note that the variables with underscores do not participate in gradient propagation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.735 + ], + "angle": 0, + "content": "The left-hand side of (5) suggests that when we update \\(\\theta\\) to make \\(f_{x}\\) closer to \\(h_x\\), we temporarily treat \\(h_x\\) as a model with fixed parameters, though, in the next step, \\(h_x\\) will also be updated with the new \\(\\theta\\). By substituting (5) into (2) we obtain" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.741, + 0.907, + 0.77 + ], + "angle": 0, + "content": "\\[\nL _ {f, x} (\\theta) = L _ {\\mathrm {C E}} \\left(f _ {x} (\\theta)\\right) + \\frac {\\alpha}{\\beta} D _ {\\mathrm {K L}} \\left(h _ {x} (\\underline {{\\theta}}, \\phi) | | f _ {x} (\\theta)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.777, + 0.838, + 0.792 + ], + "angle": 0, + "content": "Similarly, the loss of model \\(g\\) can be rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.799, + 0.906, + 0.841 + ], + "angle": 0, + "content": "\\[\nL _ {g, x} (\\phi) = L _ {\\mathrm {C E}} \\left(g _ {x} (\\phi)\\right) + \\frac {\\alpha}{1 - \\beta} D _ {\\mathrm {K L}} \\left(h _ {x} \\left(\\theta , \\underline {{\\phi}}\\right) \\mid \\mid g _ {x} (\\phi)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "From this perspective, we can see that KL divergence in DML helps continuously align each model's predictions with the ensemble distribution during training, thereby DML can be regarded as an implicit DE." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25814" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.416, + 0.108 + ], + "angle": 0, + "content": "3.2. Differentiated Deep Mutual Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.265 + ], + "angle": 0, + "content": "Model diversity plays a crucial role in the effectiveness of the ensemble [1, 30]. Although DML offers the opportunity to reduce computational overhead compared to DE, prior learning strategies in DML may destroy the model diversity and result in poor calibration performance. To address this issue, we propose an improved approach, Diff-DML, which increases the output discrepancy between \\( f \\) and \\( g \\) during training. Diff-DML employs two methods to increase model discrepancy: Differentiated Training Strategy (DTS) and Diversity-Preserving Learning Objective (DPLO)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.266, + 0.484, + 0.386 + ], + "angle": 0, + "content": "Differentiated Training Strategy: In traditional DML training [29], both models typically use the same scheduler, beginning with a specific learning rate and gradually decreasing it over time. This consistency in training strategies often leads to high similarity between the two models, diminishing their diversity. A straightforward way to increase the diversity between the outputs of the two models is to employ different training schedulers in training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.485, + 0.721 + ], + "angle": 0, + "content": "In light of this, we propose to use different training strategies: one model, \\( f \\), is designated as the primary model (used solely during inference) and follows the conventional strategy of gradually decreasing the learning rate. The other model, \\( g \\), serves as an auxiliary model and is trained using a fixed learning rate throughout the training process. This discrepant learning rate scheme can preserve the diversity among models and thus improve the calibration performance. Diversity-Preserving Learning Objective: To increase the disparity between the two models, we propose a Diversity-Preserving Learning Objective by changing the loss function of the auxiliary model in DML. In traditional DML, the models \\( f(\\theta) \\) and \\( g(\\phi) \\) share the same form of loss function, which combines a supervised loss with KL divergence. This learning objective may cause the model to overfit the one-hot ground-truth labels, resulting in unavoidable overconfidence and insufficient diversity. To address this problem, we suggest that the primary model \\( f \\) continue using the original loss function (2), while the auxiliary model \\( g(\\phi) \\) omits the cross-entropy loss component. As a result, the auxiliary model no longer receives direct supervision from the labels. Concretely, the new loss function of model \\( g(\\phi) \\) is" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.731, + 0.484, + 0.75 + ], + "angle": 0, + "content": "\\[\nL _ {g, x} (\\phi) = D _ {\\mathrm {K L}} \\left(f _ {x} (\\theta) | | g _ {x} (\\phi)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.486, + 0.822 + ], + "angle": 0, + "content": "The KL divergence not only constructs a mutual feedback loop between the primary and auxiliary models but also assists the auxiliary model in learning a differentiated prediction distribution to boost model diversity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.427, + 0.848 + ], + "angle": 0, + "content": "3.3. Extending to multiple auxiliary models" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.854, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Naturally, we can incorporate multiple auxiliary models \\(\\{g_i\\}_{i=1}^{N_g}\\) into the training process to further improve the prediction performance and calibration of the primary model \\(f\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.153 + ], + "angle": 0, + "content": "Analogous to DE, leveraging multiple auxiliary models can effectively promote model diversity and improve the uncertainty calibration of the primary model. The optimization objectives for the primary model \\( f \\) is" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.164, + 0.907, + 0.219 + ], + "angle": 0, + "content": "\\[\nL _ {f, x} (\\theta) = L _ {\\mathrm {C E}} \\left(f _ {x} (\\theta)\\right) + \\alpha \\sum_ {i = 1} ^ {N _ {g}} \\frac {1}{N _ {g}} D _ {\\mathrm {K L}} \\left(g _ {i, x} \\left(\\phi_ {i}\\right) \\mid \\mid f _ {x} (\\theta)\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.22, + 0.905, + 0.248 + ], + "angle": 0, + "content": "For traditional DML, the optimization objectives for each \\(g_{i}\\) are formalized as" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.261, + 0.907, + 0.292 + ], + "angle": 0, + "content": "\\[\nL _ {g _ {i, x}} \\left(\\phi_ {i}\\right) = L _ {\\mathrm {C E}} \\left(g _ {i, x} \\left(\\phi_ {i}\\right)\\right) + \\alpha D _ {\\mathrm {K L}} \\left(f _ {x} (\\theta) \\mid \\mid g _ {i, x} \\left(\\phi_ {i}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.908, + 0.338 + ], + "angle": 0, + "content": "where \\(\\phi_{i}\\) denotes the trainable parameters of the auxiliary model \\(g_{i}\\). And for Diff-DML, the optimization objectives for each \\(g_{i}\\) are formalized as" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.349, + 0.907, + 0.368 + ], + "angle": 0, + "content": "\\[\nL _ {g _ {i, x}} \\left(\\phi_ {i}\\right) = D _ {\\mathrm {K L}} \\left(f _ {x} (\\theta) \\mid \\mid g _ {i, x} \\left(\\phi_ {i}\\right)\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.378, + 0.887, + 0.394 + ], + "angle": 0, + "content": "Similarly, we only use the primary model \\( f \\) in inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.407, + 0.646, + 0.423 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.492 + ], + "angle": 0, + "content": "We conducted primary experiments on image classification benchmarks, followed by experiments on misclassification and OOD detection to validate the performance of Diff-DML across different tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.502, + 0.702, + 0.518 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.629 + ], + "angle": 0, + "content": "Datasets: We trained and evaluated our multi-class classification model using CIFAR-10 [13], CIFAR-100 [13] and Tiny-ImageNet [3] datasets. Validation sets were created by randomly selecting 5,000 samples from each dataset's training set, with the remaining samples used for training. We experimented with two network architectures, ResNet34 and ResNet50 [9], on these datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.75 + ], + "angle": 0, + "content": "Evaluation metrics: For standard classification, we evaluated all methods using accuracy (acc), ECE, and cw-ECE. For misclassification and OOD detection, we used four metrics: (1) False positive rate at \\(95\\%\\) true positive rate (FPR- \\(95\\%\\)); (2) The misclassification probability when true positive rate is \\(95\\%\\) (D-error); (3) Area under the receiver operating characteristic curve (AUROC); (4) Area under the precision-recall curve (AUPR)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Selection of models for testing: When selecting the optimal calibration model, we need to evaluate both accuracy and calibration performance. It is insufficient to consider only accuracy or only calibration performance. To account for both aspects simultaneously, we have developed a new method for selecting models to be tested on the test set, which we refer to as the Comprehensive Selection Method for Accuracy and Calibration (CSMAC). The selection method is as follows: A threshold value for accuracy, denoted as \\( b \\), is set. When the highest accuracy of a model on the validation set is" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25815" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.304 + ], + "angle": 0, + "content": "below \\( b \\), the checkpoint with the highest accuracy is selected. When the highest accuracy exceeds \\( b \\), the checkpoint with the highest value of accuracy minus ECE (which we call the comprehensive evaluation score) on the validation set is selected. The value of \\( b \\) varies depending on the dataset and base model. In the experiment, we first train a model optimized using cross-entropy loss and identify the checkpoint corresponding to the highest accuracy on the validation set. We set this checkpoint as the baseline model, and denote its accuracy on the validation set as \\( a \\). Then, we set \\( b = a - c \\), where \\( c \\) represents a tolerance value for accuracy reduction, and in our experiment, \\( c \\) is set to 0.01. In this way, we establish a selection mechanism that takes both accuracy and calibration performance into account." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.308, + 0.487, + 0.761 + ], + "angle": 0, + "content": "Competing methods and hyper-parameters: The evaluated methods include: Baseline: This is the model with the highest accuracy on the validation set trained with the CE loss. Note that the baseline model is used to determine the threshold \\( b \\) for the CSMAC selection method. All other models use this threshold to make selections through CSMAC. Cross-Entropy (CE): This is the model trained with CE loss (1) selected by CSMAC. DML & DML-2: These refer to the DML method using 1 and 2 auxiliary ResNet18 models, respectively. The primary model used was either ResNet34 or ResNet50. For both DML-1 and DML-2, we report the performance of the primary models. Diff-DML & Diff-DML-2: Similar to DML-1/2, these use Diff-DML for training with 1 or 2 auxiliary ResNet18 models. The performance of the primary model, either ResNet34 or ResNet50, is reported. For DML and Diff-DML, the value \\( \\alpha \\in [0.4, 1.0] \\) and we tuned it to achieve the best performance via the validation set. DCA [19]: Four values were used for its \\( \\beta \\): 0.5, 1.0, 5.0, and 10.0. MMCE [16]: Four values were used for its \\( \\lambda \\): 1.0, 2.0, 3.0, and 4.0. FL [21]: Three values were used for its \\( \\gamma \\): 1.0, 2.0, and 3.0. FL+MDCA [10]: Three values were used for the \\( \\gamma \\) in the focal loss: 1.0, 2.0, and 3.0, while the \\( \\beta \\) value in MDCA was 1.0. FL+DTDP [25]: Three values were used for the \\( \\gamma \\) in the focal loss: 1.0, 2.0, and 3.0. For DTDP, we used an \\( \\alpha \\) value of 0.005 and experimented with \\( \\lambda \\) values of 5.0 and 10.0, along with a data pruning ratio of \\( 10\\% \\). DE-3 [5, 17]: It denotes DE with 3 models using cross-entropy loss with different initializations and combined their produced probabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Other implementation details: In the experiments, we utilized the SGD optimizer with a momentum of 0.9 and weight decay of 0.0005. The initial learning rate was set to 0.1 for all datasets. For each dataset, we trained the models for a total of 250 epochs. During the initial 100 epochs, we maintained the initial learning rate. Afterward, every 50 epochs, we decreased the learning rate by a factor of \\(10\\%\\). In Diff-DML, the primary model \\(f\\) employed the same optimization strategy as described above. However, for the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.909, + 0.152 + ], + "angle": 0, + "content": "auxiliary model \\( g \\), we initialized the learning rate as 0.01 and did not apply any learning rate decay. All experiments were performed with a batch size of 100 and conducted for training and testing on a single GeForce RTX 2080 GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.164, + 0.849, + 0.18 + ], + "angle": 0, + "content": "4.2. Classification accuracy and calibration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.187, + 0.908, + 0.292 + ], + "angle": 0, + "content": "Table 1 shows the comparison results of Diff-DML and various single-model calibration methods on CIFAR-100. Tables S1 and S2 in Appendix D show the comparison results on CIFAR-10 and Tiny/ImageNet, respectively. Figure S1 in Appendix E visually demonstrates the calibration performance of Diff-DML with reliability diagrams and confidence histograms." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.293, + 0.909, + 0.369 + ], + "angle": 0, + "content": "Comparison between DML and Diff-DML: From Tables 1, S1 and S2, it can be observed that on all three datasets, DML and Diff-DML achieved improvements in both accuracy and calibration performance compared to CE. The improvements of Diff-DML were more significant than those of DML." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.37, + 0.909, + 0.459 + ], + "angle": 0, + "content": "Figure 2 shows the validation accuracy and ECE curves of ResNet34 during training on the CIFAR-100 dataset using CE, DML and Diff-DML methods. It can be seen that the ECE values for CE and DML steadily increased throughout training, while the ECE values for Diff-DML showed a downward trend." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.461, + 0.911, + 0.567 + ], + "angle": 0, + "content": "Comparison with regularization methods: It can be observed that the majority of regularization methods compromised a certain level of accuracy to improve calibration ability. For instance, a previous state-of-the-art method, FL+MDCA, reduced accuracy by 1 to 2 percent. In contrast, our Diff-DML method avoided this trade-off and demonstrates superior accuracy and calibration ability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.568, + 0.909, + 0.673 + ], + "angle": 0, + "content": "In terms of calibration performance, Diff-DML surpassed FL+MDCA across all three datasets. Specifically, compared to FL+MDCA, Diff-DML achieved an absolute accuracy increase of \\(1.3\\% / 3.1\\%\\) for ResNet34/ResNet50 on the CIFAR-100 dataset, with relative reductions in ECE of \\(49.6\\% / 43.8\\%\\) and relative reductions in cw-ECE of \\(7.7\\% / 13.0\\%\\) representing a considerable improvement." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.674, + 0.909, + 0.764 + ], + "angle": 0, + "content": "Comparison with post-hoc calibration methods: Post-hoc calibration methods in general did not affect accuracy much, but their calibration performance on datasets with many classes, like CIFAR-100 and Tiny-[ImageNet, was often suboptimal. In contrast, Diff-DML showed much better calibration performance on these datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Results with more auxiliary models: Table 2 shows the results of DML and Diff-DML trained with more auxiliary models on the CIFAR-100 dataset, compared to DE-3. Table S3 in Appendix D presents the comparison on CIFAR-10 and Tiny/ImageNet. It can be observed that DE-3 provided the most significant improvement in accuracy. DML and Diff-DML with two auxiliary models slightly outperformed those with one auxiliary model, but the difference was not substantial. In terms of calibration performance, Diff-DML" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25816" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.131, + 0.089, + 0.868, + 0.379 + ], + "angle": 0, + "content": "
MethodAcc(%)↑ECE(%)↓cw-ECE(10-3)↓
ResNet34ResNet50ResNet34ResNet50ResNet34ResNet50
Baseline77.8777.858.8110.392.252.51
CE77.7377.548.5510.042.292.44
Post-hoc calibration methods
CE+TS [8]77.7377.543.373.141.721.63
CE+VS [8]77.4777.333.763.131.821.71
CE+DS [15]77.3377.543.203.141.711.63
Regularization based calibration methods
DCA [19]77.5377.6610.1011.692.432.66
MMCE [16]75.5075.166.267.191.851.98
FL [21]76.9376.491.211.791.561.55
FL+DTDP [25]75.8776.091.822.611.601.65
FL+MDCA [10]76.9175.931.291.531.551.61
DML-based calibration methods
DML [29]77.3677.893.434.421.561.69
Diff-DML (ours)78.2079.050.650.861.431.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.388, + 0.781, + 0.402 + ], + "angle": 0, + "content": "Table 1. Different calibration methods' accuracy, ECE, and cw-ECE on the CIFAR-100 dataset." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.418, + 0.362, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.416, + 0.631, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.416, + 0.902, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.581, + 0.908, + 0.611 + ], + "angle": 0, + "content": "Figure 2. The validation accuracy and ECE curves of ResNet34 trained using the (a) CE, (b) DML, and (c) Diff-DML methods on the CIFAR-100 dataset during the training process." + }, + { + "type": "table", + "bbox": [ + 0.199, + 0.623, + 0.803, + 0.737 + ], + "angle": 0, + "content": "
MethodAcc(%)↑ECE(%)↓cw-ECE(10-3)↓
ResNet34ResNet50ResNet34ResNet50ResNet34ResNet50
DE-3 [5, 17]79.3479.632.954.401.561.73
DML [29]77.3677.893.434.421.561.69
DML-2 [29]77.9878.004.324.381.611.82
Diff-DML (ours)78.2079.050.650.861.431.40
Diff-DML-2 (ours)79.1479.300.750.721.371.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.746, + 0.842, + 0.761 + ], + "angle": 0, + "content": "Table 2. Comparison of DML and Diff-DML trained with more auxiliary models against DE on CIFAR-100 dataset." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.787, + 0.286, + 0.802 + ], + "angle": 0, + "content": "was clearly superior to DE-3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Training and inference time: In terms of time overhead, training Diff-DML with one auxiliary model took 1.4 times longer than training a single model, and with two auxiliary models, it took 1.9 times longer. In contrast, DE-3's training time was 3 times that of a single model. For inference, Diff-DML required the same time as a single model, while" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.907, + 0.832 + ], + "angle": 0, + "content": "DE-3's inference time was 3 times longer. Thus, our method provided substantial advantages over DE-3 in both training and inference time." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.851, + 0.863 + ], + "angle": 0, + "content": "4.3. Analysis of the robustness of Diff-DML" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Sensitivity of \\(\\alpha\\): We investigated the sensitivity of the important hyper-parameter \\(\\alpha\\) in (2). From Figure 3, it is ob" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25817" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.079, + 0.375, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.079, + 0.641, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.079, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.242, + 0.908, + 0.271 + ], + "angle": 0, + "content": "Figure 3. Variation of accuracy, ECE, and classwise-ECE of Diff-DML method trained on CIFAR-100 with different values of hyperparameter \\(\\alpha\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.289, + 0.483, + 0.426 + ], + "angle": 0, + "content": "served that on CIFAR-100, as \\(\\alpha\\) changes, both accuracy and cw-ECE exhibited smooth variations. Specifically, the fluctuations in accuracy remained within 3 percentage points, while the fluctuations in cw-ECE were within 0.02 percentage points. As for ECE, although its values ranged between 2 and 4 for very small (e.g., 0.4) and large (e.g., 1.0) \\(\\alpha\\) values, the fluctuations around the lowest point of ECE were also minimal. This demonstrates the good robustness of Diff-DML with respect to the hyper-parameter \\(\\alpha\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.637 + ], + "angle": 0, + "content": "Selection of auxiliary model: We investigated how to select the auxiliary model. We experimented with CNNs containing two hidden layers, as well as ResNet18, ResNet34, and ResNet50 as backbones for the auxiliary model, which is shown in Table S4 in Appendix F. It can be observed that only when using CNNs, both accuracy and calibration ability were significantly poor. When using ResNet18, ResNet34, and ResNet50, similar performance were observed. This indicates that in order to ensure good performance of the Diff-DML, a backbone for the auxiliary model should not be selected if its performance is significantly worse than that of the primary model. Furthermore, when performance is similar, specific model architectures will yield similar performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.242, + 0.665 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.484, + 0.807 + ], + "angle": 0, + "content": "In Table 3, we conducted an ablation study on the performance of the two strategies, DTS and DPLO, used in Diff-DML, using ResNet50 on the CIFAR-100 dataset. In the ablation study of DPLO, we set the learning rates and schedulers of models \\( f \\) and \\( g \\) to be the same as those in DML. As shown in the table, Diff-DML with only DTS or only DPLO both achieve better accuracy and calibration performance compared to the original DML, with the best results obtained when both strategies are combined." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.395, + 0.833 + ], + "angle": 0, + "content": "4.5. Misclassification & OOD detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.901 + ], + "angle": 0, + "content": "A real-world application of calibrated predictions is making the model aware of its uncertainties. Specifically, the model should show low confidence for misclassified or out-of-distribution samples. Motivated by this, we conducted" + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.28, + 0.886, + 0.386 + ], + "angle": 0, + "content": "
MethodAcc(%)↑ECE(%)↓cw-ECE(%)↓
CE77.5410.042.44
DML [29]77.894.421.69
DTS78.101.381.43
DPLO78.751.011.40
DTS+DPLO79.050.861.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.391, + 0.868, + 0.404 + ], + "angle": 0, + "content": "Table 3. Ablation study of Diff-DML on CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.41, + 0.907, + 0.469 + ], + "angle": 0, + "content": "experiments on misclassification and OOD detection, measuring the success rate of detecting such samples using maximum softmax probability [11]. We report threshold-free metrics (FPR-95%, D-error, AUROC, and AUPR)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.605 + ], + "angle": 0, + "content": "Misclassification detection: Misclassification detection is an important problem in machine learning, as it allows for the identification of instances where the model's predictions are unreliable. We expect the model to output low probability or high uncertainty for samples that cannot be classified correctly. Table 4 presents the performance results for various models in detecting misclassifications. Diff-DML showed significant improvements over other single-model calibration techniques and the DE method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.907, + 0.801 + ], + "angle": 0, + "content": "OD detection: A reliable classification model should show higher uncertainty and lower confidence when encountering test samples that differ significantly from the training data. We evaluated the ability of different calibration methods to distinguish OOD samples by mixing in-distribution test data with OOD data. We tested models trained on CIFAR-10 and CIFAR-100 for two types of OOD detection [5, 12]: far OOD (distinguishing SVHN [24] from CIFAR-10/CIFAR-100) and near OOD (distinguishing between CIFAR-10 and CIFAR-100). The results in Table 5 show that Diff-DML significantly outperformed other calibration methods, even exceeding the DE method, which is known for its OOD detection effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.759, + 0.831 + ], + "angle": 0, + "content": "5. Conclusion and discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Although Diff-DML significantly enhances both accuracy and calibration, its reliance on an auxiliary model with performance comparable to the primary model presents challenges, especially when training models with large parameter" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25818" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.245, + 0.089, + 0.753, + 0.375 + ], + "angle": 0, + "content": "
DatasetMethodFPR-95%↓D-error↓AUROC↑AUPR↑
CIFAR-10CE38.5313.2091.8099.46
DCA [19]38.9713.2891.9499.47
MMCE [16]23.0012.6193.5999.54
FL [21]33.1413.5692.2399.46
FL+DTDP [25]28.3414.2692.2099.36
FL+MDCA [10]32.8013.2192.4199.47
DE-3 [5, 17]22.3710.8994.0899.68
DML [29]19.2211.4994.0599.64
Diff-DML(ours)16.6310.0094.8799.73
CIFAR-100CE45.9520.1986.9395.47
DCA [19]40.6620.0986.8595.01
MMCE [16]45.9720.0287.0095.05
FL [21]45.0220.2187.0495.42
FL+DTDP [25]43.9520.7386.7995.16
FL+MDCA [10]45.0820.6886.5695.36
DE-3 [5, 17]40.0419.2588.1996.44
DML [29]40.3119.4287.8796.17
Diff-DML(ours)39.6218.9588.3896.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.244, + 0.384, + 0.752, + 0.398 + ], + "angle": 0, + "content": "Table 4. The misclassification detection performance of different calibration methods." + }, + { + "type": "table", + "bbox": [ + 0.191, + 0.412, + 0.809, + 0.698 + ], + "angle": 0, + "content": "
DatasetMethodFPR-95%↓D-error↓AUROC↑AUPR↑
In-Distribution: CIFAR-10CE69.48/19.8019.45/10.5986.11/94.6082.00/72.54
DCA [19]64.63/18.3818.80/10.1087.07/95.0584.36/72.48
MMCE [16]40.40/11.9317.33/8.3789.13/97.5389.61/88.85
FL [21]66.13/19.2020.97/10.6985.57/94.2882.81/74.18
FL+DTDP [25]50.54/23.6220.38/13.5486.57/92.0386.73/77.10
Out-of-Distribution: CIFAR-100/SVHNFL+MDCA [10]72.19/29.8420.75/13.5884.77/92.6381.01/67.11
DE-3 [5, 17]33.89/13.3115.60/9.0790.12/95.0690.27/86.53
DML [29]32.42/13.8515.90/9.3391.05/95.3891.67/86.04
Diff-DML(ours)30.12/10.5115.06/7.6591.87/96.3392.67/90.29
In-Distribution: CIFAR-100CE68.26/45.8728.69/22.2777.30/85.4477.49/54.01
DCA [19]69.41/47.8830.58/23.6275.96/83.2077.19/55.41
MMCE [16]65.34/49.3829.53/25.7776.58/77.6478.66/49.53
FL [21]67.27/48.4028.10/23.1778.36/83.8978.87/53.39
FL+DTDP [25]59.10/53.8527.23/24.1679.07/83.1081.05/49.61
Out-of-Distribution: CIFAR-10/SVHNFL+MDCA [10]69.52/41.6627.93/19.8778.35/87.4677.92/59.08
DE-3 [5, 17]56.21/48.7026.61/24.2080.36/83.4282.34/53.17
DML [29]62.08/37.9527.07/19.6879.90/87.0681.07/65.32
Diff-DML(ours)56.21/37.8425.55/19.5881.50/87.3483.57/65.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.707, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Table 5. The OOD detection performance of different calibration methods. The numbers before and after the backslash denote the near and far OOD detection results, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.486, + 0.807 + ], + "angle": 0, + "content": "sizes. This dependency can complicate scalability and computational efficiency, particularly in resource-constrained environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.486, + 0.903 + ], + "angle": 0, + "content": "In conclusion, we introduce Diff-DML, which overcomes the limitations of traditional DML by incorporating two strategies: DTS and DPLO. While DML promotes learning from an ensemble distribution, it often reduces prediction diversity, which is crucial for effective calibration. Diff-DML preserves this diversity while simultaneously" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.761, + 0.909, + 0.834 + ], + "angle": 0, + "content": "improving both accuracy and calibration. Extensive experiments on datasets such as CIFAR-100 demonstrate that our approach outperforms state-of-the-art methods, offering a practical and efficient solution for reliable uncertainty estimation, particularly in safety-critical applications." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25819" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.141 + ], + "angle": 0, + "content": "[1] Gavin Brown. Diversity in neural network ensembles. PhD thesis, CiteSeer, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.144, + 0.484, + 0.198 + ], + "angle": 0, + "content": "[2] Peng Cui, Dan Zhang, Zhijie Deng, Yinpeng Dong, and Jun Zhu. Learning sample difficulty from pre-trained models for reliable prediction. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.2, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.484, + 0.339 + ], + "angle": 0, + "content": "[4] James M Dolezal, Andrew Srisuwananukorn, Dmitry Karpeyev, Siddhi Ramesh, Sara Kochanny, Brittany Cody, Aaron S Mansfield, Sagar Rakshit, Radhika Bansal, Melanie C Bois, et al. Uncertainty-informed deep learning models enable high-confidence predictions for digital histopathology. Nature communications, 13(1):6572, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[5] Stanislav Fort, Huiyi Hu, and Balaji Lakshminarayanan. Deep ensembles: A loss landscape perspective. arXiv preprint arXiv:1912.02757, 2019. 2, 5, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[6] Kevin P Greenman, Ava P Amini, and Kevin K Yang. Benchmarking uncertainty quantification for protein engineering. bioRxiv, pages 2023-04, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.465 + ], + "angle": 0, + "content": "[7] Arthur Gretton. Introduction to rkhs, and some simple kernel algorithms. Adv. Top. Mach. Learn. Lecture Conducted from University College London, 16(5-3):2, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.484, + 0.52 + ], + "angle": 0, + "content": "[8] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.484, + 0.578 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.58, + 0.484, + 0.661 + ], + "angle": 0, + "content": "[10] Ramya Hebbalaguppe, Jatin Prakash, Neelabh Madan, and Chetan Arora. A stitch in time saves nine: A train-time regularizing loss for improved neural network calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16081-16090, 2022. 2, 3, 5, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[11] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[12] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joe Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[13] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[14] Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. In Artificial Intelligence and Statistics, pages 623-631. PMLR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Meelis Kull, Miquel Perello Nieto, Markus Kangsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "with dirichlet calibration. Advances in neural information processing systems, 32, 2019. 2, 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[16] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In International Conference on Machine Learning, pages 2805-2814. PMLR, 2018. 2, 3, 5, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.234 + ], + "angle": 0, + "content": "[17] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 2, 5, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.907, + 0.291 + ], + "angle": 0, + "content": "[18] Christian Leibig, Vaneeda Allken, Murat Seçkin Ayhan, Philipp Berens, and Siegfried Wahl. Leveraging uncertainty information from deep neural networks for disease detection. Scientific reports, 7(1):1-14, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.907, + 0.347 + ], + "angle": 0, + "content": "[19] Gongbo Liang, Yu Zhang, Xiaoqin Wang, and Nathan Jacobs. Improved trainable calibration method for neural networks on medical imaging classification. arXiv preprint arXiv:2009.04057, 2020.3,5,6,8,1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[20] Rhiannon Michelmore, Matthew Wicker, Luca Laurenti, Luca Cardelli, Yarin Gal, and Marta Kwiatkowska. Uncertainty quantification with statistical guarantees in end-to-end autonomous driving control. In International Conference on Robotics and Automation, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.907, + 0.488 + ], + "angle": 0, + "content": "[21] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. Advances in Neural Information Processing Systems, 33:15288-15299, 2020. 3, 5, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.907, + 0.532 + ], + "angle": 0, + "content": "[22] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.906, + 0.589 + ], + "angle": 0, + "content": "[23] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Proceedings of the AAAI conference on artificial intelligence, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.907, + 0.659 + ], + "angle": 0, + "content": "[24] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning, page 7. Granada, Spain, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.907, + 0.743 + ], + "angle": 0, + "content": "[25] Rishabh Patra, Ramya Hebbalaguppe, Tirtharaj Dash, Gautam Shroff, and Lovekesh Vig. Calibrating deep neural networks using explicit regularisation and dynamic data pruning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1541-1549, 2023. 3, 5, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[26] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey Hinton. Regularizing neural networks by penalizing confident output distributions. arXiv preprint arXiv:1701.06548, 2017. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.907, + 0.856 + ], + "angle": 0, + "content": "[27] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61-74, 1999. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.859, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[28] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25820" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "conference on computer vision and pattern recognition, pages 2818-2826, 2016. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[29] Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4320-4328, 2018. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[30] Zhi-Hua Zhou. Ensemble methods: foundations and algorithms. CRC press, 2012. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25821" + } + ] +] \ No newline at end of file diff --git a/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/6a2c48a3-aafc-4ad2-9262-8700829f154e_origin.pdf b/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/6a2c48a3-aafc-4ad2-9262-8700829f154e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3b535207a913462337c5e33dd8e131e620c0ef12 --- /dev/null +++ b/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/6a2c48a3-aafc-4ad2-9262-8700829f154e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f124c86844a55f3f1be62d64b5f17cfe478f0c519e1dd58bccb738869e7920b +size 1729828 diff --git a/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/full.md b/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d7f30993746ee6c04264cbc4a7db467eb551266d --- /dev/null +++ b/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/full.md @@ -0,0 +1,311 @@ +# Improving Accuracy and Calibration via Differentiated Deep Mutual Learning + +Han Liu $^{1,2,\dagger}$ , Peng Cui $^{1,\dagger}$ , Binging Wang $^{2}$ , Weipeng Chen $^{2}$ , Yupeng Zhang $^{2}$ , Jun Zhu $^{1,*}$ , and Xiaolin Hu $^{1,3,4,*}$ + +$^{1}$ Department of Computer Science and Technology, Institute for AI, BNRist, Tsinghua University, Beijing 100084, China + +2Baichuan Inc. + +$^{3}$ Tsinghua Laboratory of Brain and Intelligence (THBI), + +IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing 100084, China + +$^{4}$ Chinese Institute for Brain Research (CIBR), Beijing 100010, China + +{han-liu18, cuip22}@mails.tsinghua.edu.cn + +{daniel, chenweipeng, zhangyupeng}@baichuan-inc.com + +{dcszj, xlhu}@mail.tsinghua.edu.cn + +# Abstract + +Deep Neural Networks (DNNs) have achieved remarkable success in a variety of tasks, particularly in terms of prediction accuracy. However, in real-world scenarios, especially in safety-critical applications, accuracy alone is insufficient; reliable uncertainty estimates are essential. Modern DNNs, often trained with cross-entropy loss, tend to exhibit overconfidence, especially on ambiguous samples. Many techniques aim to improve uncertainty calibration, yet they often come at the cost of reduced accuracy or increased computational demands. To address this challenge, we propose Differentiated Deep Mutual Learning (Diff-DML), an efficient ensemble approach that simultaneously enhances accuracy and uncertainty calibration. Diff-DML draws inspiration from Deep Mutual Learning (DML) while introducing two strategies to maintain prediction diversity: (1) Differentiated Training Strategy (DTS) and (2) Diversity-Preserving Learning Objective (DPLO). Our theoretical analysis shows that Diff-DML's diversified learning framework not only leverages ensemble benefits but also avoids the loss of prediction diversity observed in traditional DML setups, which is crucial for improved calibration. Extensive evaluations on various benchmarks confirm the effectiveness of Diff-DML. For instance, on the CIFAR-100 dataset, Diff-DML on ResNet34/50 models achieved substantial improvements over the previous state-of-the-art method, MDCA, with absolute accuracy gains of $1.3\% / 3.1\%$ , relative ECE reductions + +![](images/b25785f2e35a5aba3aca0aa4a060bf130a35505e1c140444266a82b18197c884.jpg) +Figure 1. Scatter plot of ECE and accuracy for different calibration methods based on the ResNet50 backbone on the CIFAR-100 test set. Methods located closer to the top left corner perform better. + +of $49.6\% / 43.8\%$ , and relative classwise-ECE reductions of $7.7\% / 13.0\%$ . + +# 1. Introduction + +Deep Neural Networks (DNNs) have achieved great success in various tasks, especially in terms of prediction accuracy. In safety-critical areas such as medical diagnosis [4, 18], self-driving cars [20], and protein engineering [6], accurate uncertainty estimation (i.e., well-calibrated prediction) is crucial for reliable model deployment. For instance, a reliable model needs to show high uncertainty when making incorrect predictions. However, models trained with cross-entropy (CE) loss tend to be overconfident, especially when + +encountering ambiguous samples, leading to predictions that are often close to $100\%$ confident and producing undesirable overconfidence [2, 8]. + +To address this issue, various regularization techniques have been explored during training, referred to as training-time calibration. For instance, entropy regularization (ER) [26] and label smoothing (LS) [22, 28] promote smoother predictive probability distributions to alleviate overconfidence by maximizing entropy or adjusting label distributions. Some methods aim to directly reduce calibration error during training to achieve multi-class calibration. For example, MMCE [16] introduces a trainable measure based on kernel mean embeddings, while MDCA [10] employs a differentiable auxiliary loss to minimize the gap between prediction confidence and accuracy. However, despite mitigating overconfidence, these techniques often compromise prediction accuracy due to excessive regularization. Another category of methods is post-processing, which adjusts output probabilities using Temperature Scaling (TS) [27], Vector Scaling (VS) [27], or Dirichlet Scaling (DS) [15]. However, these post-processing techniques typically rely on an additional validation dataset. Furthermore, they have limited learnable parameters, which may lead to suboptimal calibration performance [10]. + +Deep Ensembles (DE) [5, 17] offer a straightforward solution by combining multiple randomly initialized trained DNNs to improve prediction accuracy and uncertainty calibration. Yet, using multiple DNNs during training and inference significantly increases computational costs, which poses challenges for deploying DE in real-world scenarios. To address this limitation, we draw inspiration from Deep Mutual Learning (DML) [29], a technique used in model distillation, where an ensemble of student models learn collaboratively and teach each other throughout the training process. This work provides the theoretical analysis showing that DML encourages each model to learn from an ensemble distribution through a mimicry loss. The proposed theory motivates us to leverage DML as an implicit form of DE, aiming to achieve similar gains in accuracy and calibration as DE, while mitigating the high computational overhead because DML uses only one model for prediction during inference. + +However, in standard DML, although the models start from different initial states, they become overly similar during mutual learning, which reduces the diversity of their outputs—a key factor needed for effective ensemble-based calibration. In this work, we propose Differentiated Deep Mutual Learning (Diff-DML), a novel approach that enhances the benefits of DML by promoting prediction diversity between models to achieve high accuracy and strong calibration performance. Diff-DML incorporates two key strategies that preserve output diversity: (1) Differentiated Training Strategy (DTS), which uses different learning rates and schedules + +for different models to prevent premature convergence, and (2) Diversity-Preserving Learning Objective (DPLO), which removes cross-entropy loss for auxiliary models to avoid overfitting to the one-hot ground-truth labels. + +These strategies help Diff-DML balance collaborative learning with prediction diversity, resulting in better uncertainty calibration and enhanced predictive performance. Extensive evaluations on various benchmarks, including image classification, misclassification detection, and out-of-distribution (OOD) detection, confirm Diff-DML's effectiveness. For example, on the CIFAR-100 dataset, Diff-DML on ResNet34/50 achieved significant improvements compared to the previous state-of-the-art method, MDCA [10], with absolute accuracy increases of $1.3\% / 3.1\%$ , relative reductions in ECE of $49.6\% / 43.8\%$ , and relative reductions in classwise-ECE of $7.7\% / 13.0\%$ . The latter two metrics measure calibration performance, where lower values indicate better results. + +Figure 1 displays ECE and accuracy of different calibration methods, showing Diff-DML achieving favorable calibration and accuracy. Notably, Diff-DML incurs only minimal computational cost at inference, unlike DE, which requires multiple models to generate a single prediction. + +# 2. Background and related work + +Calibration, the process of ensuring that the predicted probabilities reflect the true likelihood of the events, plays a crucial role in machine learning models. It ensures that the prediction of machine learning models are not only accurate but also well-calibrated. Well-calibrated models provide reliable uncertainty estimates, which can help users and downstream systems make safer decisions by avoiding overconfidence. + +There are two fundamental types of calibration: confidence calibration [23] and classwise calibration [15]. In this paper, we focus on both types of calibration. + +Confidence calibration: Confidence calibration refers to the alignment of predicted confidence with observed frequencies. It ensures that a model's prediction confidence accurately reflects the true likelihood of an event occurring. A widely used metric for evaluating confidence calibration is the Expected Calibration Error (ECE) [23]. The definition of ECE can be found in Appendix A. + +Classwise calibration: Classwise calibration is concerned with calibrating the predicted probabilities for each individual class in a multi-class classification problem. It ensures that the predicted probabilities across different classes are accurate and properly calibrated. The metric for classwise calibration is classwise-ECE (cw-ECE) [15], which can be found in Appendix A. + +According to the number of models used for inference, calibration methods can be divided into single-model calibration methods and DE [5, 17]. + +Single-model calibration methods: Within single-model + +calibration methods, there are two categories: train-time calibration methods and post-hoc calibration methods. Train-time calibration methods typically involve designing loss functions or incorporating additional regularization techniques, such as Entropy Regularization (ER) [26], Label Smoothing (LS) [22, 28], Focal Loss (FL) [21], Maximum Mean Calibration Error (MMCE) [16], Difference between Confidence and Accuracy (DCA) [19], Multi-class Difference in Confidence and Accuracy (MDCA) [10], and Dynamic Train-time Data Pruning (DTDP) [25]. Post-hoc calibration methods are applied after the model is trained, such as Temperature Scaling (TS) [27], Vector Scaling, and Dirichlet Scaling (DS) [15]. + +Deep Ensembles: DE creates an ensemble by training multiple models independently with different initializations. To alleviate overconfidence and improve calibration, during inference, DE leverages $N$ randomly initialized, individually trained DNNs and makes predictions by average voting. While effective for enhancing model performance, DE introduces additional computational overhead during inference. + +More related works are discussed in Appendix B. + +# 3. Method + +Let $P_{data}$ denote the probability distribution of data. A dataset consists of finite $(x, y_x)$ pairs where $x \sim P_{data}$ is a data instance and $y_x \in \mathcal{Y} = \{1, 2, \dots, K\}$ is the ground-truth class label of $x$ . A model takes in the instance $x$ and outputs its prediction vector $p_x \in \mathbb{R}^K$ . Let $p_x^{(k)}$ denote the $k$ -th dimension of $p_x$ . The confidence $\hat{p}_x = \max_{k \in \mathcal{Y}} \{p_x^{(k)}\}$ is the maximum value in the probability vector of predictions, and the model's prediction $\hat{y}_x = \operatorname{argmax}_{k \in \mathcal{Y}} \{p_x^{(k)}\}$ is the index corresponding to the maximum value. + +Typically, we train a model by minimizing the following CE loss on the training set, + +$$ +L _ {\mathrm {C E}} \left(p _ {x}\right) = - \log p _ {x} ^ {\left(y _ {x}\right)}, \tag {1} +$$ + +where $p_x^{(y_x)}$ represents the prediction probability corresponding to the ground-truth label $y_{x}$ . Previous studies [2, 8, 21] reveal that minimizing the CE loss forces the model to overfit to one-hot encoded ground-truth labels with almost $100\%$ probability irrespective of their correctness, which can result in overconfidence. + +DML [29] is a model distillation technique that allows two models $f(\theta)$ and $g(\phi)$ to learn collaboratively through a combination of a supervised loss and a Kullback-Leibler (KL) Divergence: + +$$ +L _ {f, x} (\theta) = L _ {\mathrm {C E}} \left(f _ {x} (\theta)\right) + \alpha D _ {\mathrm {K L}} \left(g _ {x} (\phi) \mid \mid f _ {x} (\theta)\right), \tag {2} +$$ + +$$ +L _ {g, x} (\phi) = L _ {\mathrm {C E}} \left(g _ {x} (\phi)\right) + \alpha D _ {\mathrm {K L}} \left(f _ {x} (\theta) \| g _ {x} (\phi)\right), \tag {3} +$$ + +where $\alpha$ is a hyperparameter used to adjust the weight of the KL divergence loss. + +In this paper, we designate $f(\theta)$ as the primary model, and $g(\phi)$ as the auxiliary model that does not participate in the inference process. It is important to note that DML was originally conceived as a knowledge distillation technique aimed at improving accuracy rather than focusing on calibration. + +# 3.1. DML is an implicit form of deep ensembles + +DML involves two models learning collaboratively by continuously exchanging knowledge during the training process. This collaborative learning encourages each model to learn from the strengths and correct the weaknesses of others, promoting diversity in model predictions. Consequently, DML achieves the advantages of DE—such as increased prediction accuracy and improved uncertainty calibration—without the need for independent model training or the associated computational costs. + +In this section, we explore the theoretical relationship between DML and DE. + +Given two trained models $f(\theta)$ and $g(\phi)$ and an input $x$ for DE, we can get a weighted sum of predictive probabilities from $f_{x}(\theta)$ and $g_{x}(\phi)$ as the final prediction: + +$$ +h _ {x} (\theta , \phi) = (1 - \beta) f _ {x} (\theta) + \beta g _ {x} (\phi), \tag {4} +$$ + +where $\beta \in [0,1]$ is a weight to balance the effect of each model. To explain the relationship between DML and DE, we first present a proposition, the proof is in Appendix C). + +Proposition 1. Given two prediction probability distributions $f_{x}(\theta)$ and $g_{x}(\phi)$ for classifying sample $x$ , as well as their ensemble $h_{x}(\underline{\theta},\phi)$ , we have + +$$ +\frac {\partial}{\partial \theta} D _ {K L} \left(h _ {x} (\underline {{\theta}}, \phi) | | f _ {x} (\theta)\right) = \beta \frac {\partial}{\partial \theta} D _ {K L} \left(g _ {x} (\phi) | | f _ {x} (\theta)\right). \tag {5} +$$ + +Note that the variables with underscores do not participate in gradient propagation. + +The left-hand side of (5) suggests that when we update $\theta$ to make $f_{x}$ closer to $h_x$ , we temporarily treat $h_x$ as a model with fixed parameters, though, in the next step, $h_x$ will also be updated with the new $\theta$ . By substituting (5) into (2) we obtain + +$$ +L _ {f, x} (\theta) = L _ {\mathrm {C E}} \left(f _ {x} (\theta)\right) + \frac {\alpha}{\beta} D _ {\mathrm {K L}} \left(h _ {x} (\underline {{\theta}}, \phi) | | f _ {x} (\theta)\right). \tag {6} +$$ + +Similarly, the loss of model $g$ can be rewritten as + +$$ +L _ {g, x} (\phi) = L _ {\mathrm {C E}} \left(g _ {x} (\phi)\right) + \frac {\alpha}{1 - \beta} D _ {\mathrm {K L}} \left(h _ {x} \left(\theta , \underline {{\phi}}\right) \mid \mid g _ {x} (\phi)\right). \tag {7} +$$ + +From this perspective, we can see that KL divergence in DML helps continuously align each model's predictions with the ensemble distribution during training, thereby DML can be regarded as an implicit DE. + +# 3.2. Differentiated Deep Mutual Learning + +Model diversity plays a crucial role in the effectiveness of the ensemble [1, 30]. Although DML offers the opportunity to reduce computational overhead compared to DE, prior learning strategies in DML may destroy the model diversity and result in poor calibration performance. To address this issue, we propose an improved approach, Diff-DML, which increases the output discrepancy between $f$ and $g$ during training. Diff-DML employs two methods to increase model discrepancy: Differentiated Training Strategy (DTS) and Diversity-Preserving Learning Objective (DPLO). + +Differentiated Training Strategy: In traditional DML training [29], both models typically use the same scheduler, beginning with a specific learning rate and gradually decreasing it over time. This consistency in training strategies often leads to high similarity between the two models, diminishing their diversity. A straightforward way to increase the diversity between the outputs of the two models is to employ different training schedulers in training. + +In light of this, we propose to use different training strategies: one model, $f$ , is designated as the primary model (used solely during inference) and follows the conventional strategy of gradually decreasing the learning rate. The other model, $g$ , serves as an auxiliary model and is trained using a fixed learning rate throughout the training process. This discrepant learning rate scheme can preserve the diversity among models and thus improve the calibration performance. Diversity-Preserving Learning Objective: To increase the disparity between the two models, we propose a Diversity-Preserving Learning Objective by changing the loss function of the auxiliary model in DML. In traditional DML, the models $f(\theta)$ and $g(\phi)$ share the same form of loss function, which combines a supervised loss with KL divergence. This learning objective may cause the model to overfit the one-hot ground-truth labels, resulting in unavoidable overconfidence and insufficient diversity. To address this problem, we suggest that the primary model $f$ continue using the original loss function (2), while the auxiliary model $g(\phi)$ omits the cross-entropy loss component. As a result, the auxiliary model no longer receives direct supervision from the labels. Concretely, the new loss function of model $g(\phi)$ is + +$$ +L _ {g, x} (\phi) = D _ {\mathrm {K L}} \left(f _ {x} (\theta) | | g _ {x} (\phi)\right). \tag {8} +$$ + +The KL divergence not only constructs a mutual feedback loop between the primary and auxiliary models but also assists the auxiliary model in learning a differentiated prediction distribution to boost model diversity. + +# 3.3. Extending to multiple auxiliary models + +Naturally, we can incorporate multiple auxiliary models $\{g_i\}_{i=1}^{N_g}$ into the training process to further improve the prediction performance and calibration of the primary model $f$ . + +Analogous to DE, leveraging multiple auxiliary models can effectively promote model diversity and improve the uncertainty calibration of the primary model. The optimization objectives for the primary model $f$ is + +$$ +L _ {f, x} (\theta) = L _ {\mathrm {C E}} \left(f _ {x} (\theta)\right) + \alpha \sum_ {i = 1} ^ {N _ {g}} \frac {1}{N _ {g}} D _ {\mathrm {K L}} \left(g _ {i, x} \left(\phi_ {i}\right) \mid \mid f _ {x} (\theta)\right), \tag {9} +$$ + +For traditional DML, the optimization objectives for each $g_{i}$ are formalized as + +$$ +L _ {g _ {i, x}} \left(\phi_ {i}\right) = L _ {\mathrm {C E}} \left(g _ {i, x} \left(\phi_ {i}\right)\right) + \alpha D _ {\mathrm {K L}} \left(f _ {x} (\theta) \mid \mid g _ {i, x} \left(\phi_ {i}\right)\right), \tag {10} +$$ + +where $\phi_{i}$ denotes the trainable parameters of the auxiliary model $g_{i}$ . And for Diff-DML, the optimization objectives for each $g_{i}$ are formalized as + +$$ +L _ {g _ {i, x}} \left(\phi_ {i}\right) = D _ {\mathrm {K L}} \left(f _ {x} (\theta) \mid \mid g _ {i, x} \left(\phi_ {i}\right)\right). \tag {11} +$$ + +Similarly, we only use the primary model $f$ in inference. + +# 4. Experiments + +We conducted primary experiments on image classification benchmarks, followed by experiments on misclassification and OOD detection to validate the performance of Diff-DML across different tasks. + +# 4.1. Experimental setup + +Datasets: We trained and evaluated our multi-class classification model using CIFAR-10 [13], CIFAR-100 [13] and Tiny-ImageNet [3] datasets. Validation sets were created by randomly selecting 5,000 samples from each dataset's training set, with the remaining samples used for training. We experimented with two network architectures, ResNet34 and ResNet50 [9], on these datasets. + +Evaluation metrics: For standard classification, we evaluated all methods using accuracy (acc), ECE, and cw-ECE. For misclassification and OOD detection, we used four metrics: (1) False positive rate at $95\%$ true positive rate (FPR- $95\%$ ); (2) The misclassification probability when true positive rate is $95\%$ (D-error); (3) Area under the receiver operating characteristic curve (AUROC); (4) Area under the precision-recall curve (AUPR). + +Selection of models for testing: When selecting the optimal calibration model, we need to evaluate both accuracy and calibration performance. It is insufficient to consider only accuracy or only calibration performance. To account for both aspects simultaneously, we have developed a new method for selecting models to be tested on the test set, which we refer to as the Comprehensive Selection Method for Accuracy and Calibration (CSMAC). The selection method is as follows: A threshold value for accuracy, denoted as $b$ , is set. When the highest accuracy of a model on the validation set is + +below $b$ , the checkpoint with the highest accuracy is selected. When the highest accuracy exceeds $b$ , the checkpoint with the highest value of accuracy minus ECE (which we call the comprehensive evaluation score) on the validation set is selected. The value of $b$ varies depending on the dataset and base model. In the experiment, we first train a model optimized using cross-entropy loss and identify the checkpoint corresponding to the highest accuracy on the validation set. We set this checkpoint as the baseline model, and denote its accuracy on the validation set as $a$ . Then, we set $b = a - c$ , where $c$ represents a tolerance value for accuracy reduction, and in our experiment, $c$ is set to 0.01. In this way, we establish a selection mechanism that takes both accuracy and calibration performance into account. + +Competing methods and hyper-parameters: The evaluated methods include: Baseline: This is the model with the highest accuracy on the validation set trained with the CE loss. Note that the baseline model is used to determine the threshold $b$ for the CSMAC selection method. All other models use this threshold to make selections through CSMAC. Cross-Entropy (CE): This is the model trained with CE loss (1) selected by CSMAC. DML & DML-2: These refer to the DML method using 1 and 2 auxiliary ResNet18 models, respectively. The primary model used was either ResNet34 or ResNet50. For both DML-1 and DML-2, we report the performance of the primary models. Diff-DML & Diff-DML-2: Similar to DML-1/2, these use Diff-DML for training with 1 or 2 auxiliary ResNet18 models. The performance of the primary model, either ResNet34 or ResNet50, is reported. For DML and Diff-DML, the value $\alpha \in [0.4, 1.0]$ and we tuned it to achieve the best performance via the validation set. DCA [19]: Four values were used for its $\beta$ : 0.5, 1.0, 5.0, and 10.0. MMCE [16]: Four values were used for its $\lambda$ : 1.0, 2.0, 3.0, and 4.0. FL [21]: Three values were used for its $\gamma$ : 1.0, 2.0, and 3.0. FL+MDCA [10]: Three values were used for the $\gamma$ in the focal loss: 1.0, 2.0, and 3.0, while the $\beta$ value in MDCA was 1.0. FL+DTDP [25]: Three values were used for the $\gamma$ in the focal loss: 1.0, 2.0, and 3.0. For DTDP, we used an $\alpha$ value of 0.005 and experimented with $\lambda$ values of 5.0 and 10.0, along with a data pruning ratio of $10\%$ . DE-3 [5, 17]: It denotes DE with 3 models using cross-entropy loss with different initializations and combined their produced probabilities. + +Other implementation details: In the experiments, we utilized the SGD optimizer with a momentum of 0.9 and weight decay of 0.0005. The initial learning rate was set to 0.1 for all datasets. For each dataset, we trained the models for a total of 250 epochs. During the initial 100 epochs, we maintained the initial learning rate. Afterward, every 50 epochs, we decreased the learning rate by a factor of $10\%$ . In Diff-DML, the primary model $f$ employed the same optimization strategy as described above. However, for the + +auxiliary model $g$ , we initialized the learning rate as 0.01 and did not apply any learning rate decay. All experiments were performed with a batch size of 100 and conducted for training and testing on a single GeForce RTX 2080 GPU. + +# 4.2. Classification accuracy and calibration + +Table 1 shows the comparison results of Diff-DML and various single-model calibration methods on CIFAR-100. Tables S1 and S2 in Appendix D show the comparison results on CIFAR-10 and Tiny/ImageNet, respectively. Figure S1 in Appendix E visually demonstrates the calibration performance of Diff-DML with reliability diagrams and confidence histograms. + +Comparison between DML and Diff-DML: From Tables 1, S1 and S2, it can be observed that on all three datasets, DML and Diff-DML achieved improvements in both accuracy and calibration performance compared to CE. The improvements of Diff-DML were more significant than those of DML. + +Figure 2 shows the validation accuracy and ECE curves of ResNet34 during training on the CIFAR-100 dataset using CE, DML and Diff-DML methods. It can be seen that the ECE values for CE and DML steadily increased throughout training, while the ECE values for Diff-DML showed a downward trend. + +Comparison with regularization methods: It can be observed that the majority of regularization methods compromised a certain level of accuracy to improve calibration ability. For instance, a previous state-of-the-art method, FL+MDCA, reduced accuracy by 1 to 2 percent. In contrast, our Diff-DML method avoided this trade-off and demonstrates superior accuracy and calibration ability. + +In terms of calibration performance, Diff-DML surpassed FL+MDCA across all three datasets. Specifically, compared to FL+MDCA, Diff-DML achieved an absolute accuracy increase of $1.3\% / 3.1\%$ for ResNet34/ResNet50 on the CIFAR-100 dataset, with relative reductions in ECE of $49.6\% / 43.8\%$ and relative reductions in cw-ECE of $7.7\% / 13.0\%$ representing a considerable improvement. + +Comparison with post-hoc calibration methods: Post-hoc calibration methods in general did not affect accuracy much, but their calibration performance on datasets with many classes, like CIFAR-100 and Tiny-[ImageNet, was often suboptimal. In contrast, Diff-DML showed much better calibration performance on these datasets. + +Results with more auxiliary models: Table 2 shows the results of DML and Diff-DML trained with more auxiliary models on the CIFAR-100 dataset, compared to DE-3. Table S3 in Appendix D presents the comparison on CIFAR-10 and Tiny/ImageNet. It can be observed that DE-3 provided the most significant improvement in accuracy. DML and Diff-DML with two auxiliary models slightly outperformed those with one auxiliary model, but the difference was not substantial. In terms of calibration performance, Diff-DML + +
MethodAcc(%)↑ECE(%)↓cw-ECE(10-3)↓
ResNet34ResNet50ResNet34ResNet50ResNet34ResNet50
Baseline77.8777.858.8110.392.252.51
CE77.7377.548.5510.042.292.44
Post-hoc calibration methods
CE+TS [8]77.7377.543.373.141.721.63
CE+VS [8]77.4777.333.763.131.821.71
CE+DS [15]77.3377.543.203.141.711.63
Regularization based calibration methods
DCA [19]77.5377.6610.1011.692.432.66
MMCE [16]75.5075.166.267.191.851.98
FL [21]76.9376.491.211.791.561.55
FL+DTDP [25]75.8776.091.822.611.601.65
FL+MDCA [10]76.9175.931.291.531.551.61
DML-based calibration methods
DML [29]77.3677.893.434.421.561.69
Diff-DML (ours)78.2079.050.650.861.431.40
+ +![](images/e1df126f479648a8e8a414180bd5c8f73d5f3605bc716d8db6c500d0766a6a53.jpg) +Figure 2. The validation accuracy and ECE curves of ResNet34 trained using the (a) CE, (b) DML, and (c) Diff-DML methods on the CIFAR-100 dataset during the training process. + +![](images/7dfa6e90ddc024001f2dfd21a7eb6bb502e910b54d4bb390809ac875eda6eb8d.jpg) + +![](images/a3cf201f15679ad64253e8ec6da9a0f1086430f20beebe87a638c62855456145.jpg) + +Table 1. Different calibration methods' accuracy, ECE, and cw-ECE on the CIFAR-100 dataset. + +
MethodAcc(%)↑ECE(%)↓cw-ECE(10-3)↓
ResNet34ResNet50ResNet34ResNet50ResNet34ResNet50
DE-3 [5, 17]79.3479.632.954.401.561.73
DML [29]77.3677.893.434.421.561.69
DML-2 [29]77.9878.004.324.381.611.82
Diff-DML (ours)78.2079.050.650.861.431.40
Diff-DML-2 (ours)79.1479.300.750.721.371.33
+ +Table 2. Comparison of DML and Diff-DML trained with more auxiliary models against DE on CIFAR-100 dataset. + +was clearly superior to DE-3. + +Training and inference time: In terms of time overhead, training Diff-DML with one auxiliary model took 1.4 times longer than training a single model, and with two auxiliary models, it took 1.9 times longer. In contrast, DE-3's training time was 3 times that of a single model. For inference, Diff-DML required the same time as a single model, while + +DE-3's inference time was 3 times longer. Thus, our method provided substantial advantages over DE-3 in both training and inference time. + +# 4.3. Analysis of the robustness of Diff-DML + +Sensitivity of $\alpha$ : We investigated the sensitivity of the important hyper-parameter $\alpha$ in (2). From Figure 3, it is ob + +![](images/cb40b5b709e0ee8deedc36f94ab497bfad69e5d264bc03d029c04f29338d7864.jpg) +Figure 3. Variation of accuracy, ECE, and classwise-ECE of Diff-DML method trained on CIFAR-100 with different values of hyperparameter $\alpha$ . + +![](images/41c57e78b6e34e1c22078fe45e9ed2fdc397a82329ceee0d612f5cc5fa6996be.jpg) + +![](images/087a0df88e10a212073ed47a9c748ecfae1caa92daf199130d86cd4c0c71c897.jpg) + +served that on CIFAR-100, as $\alpha$ changes, both accuracy and cw-ECE exhibited smooth variations. Specifically, the fluctuations in accuracy remained within 3 percentage points, while the fluctuations in cw-ECE were within 0.02 percentage points. As for ECE, although its values ranged between 2 and 4 for very small (e.g., 0.4) and large (e.g., 1.0) $\alpha$ values, the fluctuations around the lowest point of ECE were also minimal. This demonstrates the good robustness of Diff-DML with respect to the hyper-parameter $\alpha$ . + +Selection of auxiliary model: We investigated how to select the auxiliary model. We experimented with CNNs containing two hidden layers, as well as ResNet18, ResNet34, and ResNet50 as backbones for the auxiliary model, which is shown in Table S4 in Appendix F. It can be observed that only when using CNNs, both accuracy and calibration ability were significantly poor. When using ResNet18, ResNet34, and ResNet50, similar performance were observed. This indicates that in order to ensure good performance of the Diff-DML, a backbone for the auxiliary model should not be selected if its performance is significantly worse than that of the primary model. Furthermore, when performance is similar, specific model architectures will yield similar performance. + +# 4.4. Ablation study + +In Table 3, we conducted an ablation study on the performance of the two strategies, DTS and DPLO, used in Diff-DML, using ResNet50 on the CIFAR-100 dataset. In the ablation study of DPLO, we set the learning rates and schedulers of models $f$ and $g$ to be the same as those in DML. As shown in the table, Diff-DML with only DTS or only DPLO both achieve better accuracy and calibration performance compared to the original DML, with the best results obtained when both strategies are combined. + +# 4.5. Misclassification & OOD detection + +A real-world application of calibrated predictions is making the model aware of its uncertainties. Specifically, the model should show low confidence for misclassified or out-of-distribution samples. Motivated by this, we conducted + +
MethodAcc(%)↑ECE(%)↓cw-ECE(%)↓
CE77.5410.042.44
DML [29]77.894.421.69
DTS78.101.381.43
DPLO78.751.011.40
DTS+DPLO79.050.861.40
+ +Table 3. Ablation study of Diff-DML on CIFAR-100. + +experiments on misclassification and OOD detection, measuring the success rate of detecting such samples using maximum softmax probability [11]. We report threshold-free metrics (FPR-95%, D-error, AUROC, and AUPR). + +Misclassification detection: Misclassification detection is an important problem in machine learning, as it allows for the identification of instances where the model's predictions are unreliable. We expect the model to output low probability or high uncertainty for samples that cannot be classified correctly. Table 4 presents the performance results for various models in detecting misclassifications. Diff-DML showed significant improvements over other single-model calibration techniques and the DE method. + +OD detection: A reliable classification model should show higher uncertainty and lower confidence when encountering test samples that differ significantly from the training data. We evaluated the ability of different calibration methods to distinguish OOD samples by mixing in-distribution test data with OOD data. We tested models trained on CIFAR-10 and CIFAR-100 for two types of OOD detection [5, 12]: far OOD (distinguishing SVHN [24] from CIFAR-10/CIFAR-100) and near OOD (distinguishing between CIFAR-10 and CIFAR-100). The results in Table 5 show that Diff-DML significantly outperformed other calibration methods, even exceeding the DE method, which is known for its OOD detection effectiveness. + +# 5. Conclusion and discussion + +Although Diff-DML significantly enhances both accuracy and calibration, its reliance on an auxiliary model with performance comparable to the primary model presents challenges, especially when training models with large parameter + +
DatasetMethodFPR-95%↓D-error↓AUROC↑AUPR↑
CIFAR-10CE38.5313.2091.8099.46
DCA [19]38.9713.2891.9499.47
MMCE [16]23.0012.6193.5999.54
FL [21]33.1413.5692.2399.46
FL+DTDP [25]28.3414.2692.2099.36
FL+MDCA [10]32.8013.2192.4199.47
DE-3 [5, 17]22.3710.8994.0899.68
DML [29]19.2211.4994.0599.64
Diff-DML(ours)16.6310.0094.8799.73
CIFAR-100CE45.9520.1986.9395.47
DCA [19]40.6620.0986.8595.01
MMCE [16]45.9720.0287.0095.05
FL [21]45.0220.2187.0495.42
FL+DTDP [25]43.9520.7386.7995.16
FL+MDCA [10]45.0820.6886.5695.36
DE-3 [5, 17]40.0419.2588.1996.44
DML [29]40.3119.4287.8796.17
Diff-DML(ours)39.6218.9588.3896.50
+ +Table 4. The misclassification detection performance of different calibration methods. + +
DatasetMethodFPR-95%↓D-error↓AUROC↑AUPR↑
In-Distribution: CIFAR-10CE69.48/19.8019.45/10.5986.11/94.6082.00/72.54
DCA [19]64.63/18.3818.80/10.1087.07/95.0584.36/72.48
MMCE [16]40.40/11.9317.33/8.3789.13/97.5389.61/88.85
FL [21]66.13/19.2020.97/10.6985.57/94.2882.81/74.18
FL+DTDP [25]50.54/23.6220.38/13.5486.57/92.0386.73/77.10
Out-of-Distribution: CIFAR-100/SVHNFL+MDCA [10]72.19/29.8420.75/13.5884.77/92.6381.01/67.11
DE-3 [5, 17]33.89/13.3115.60/9.0790.12/95.0690.27/86.53
DML [29]32.42/13.8515.90/9.3391.05/95.3891.67/86.04
Diff-DML(ours)30.12/10.5115.06/7.6591.87/96.3392.67/90.29
In-Distribution: CIFAR-100CE68.26/45.8728.69/22.2777.30/85.4477.49/54.01
DCA [19]69.41/47.8830.58/23.6275.96/83.2077.19/55.41
MMCE [16]65.34/49.3829.53/25.7776.58/77.6478.66/49.53
FL [21]67.27/48.4028.10/23.1778.36/83.8978.87/53.39
FL+DTDP [25]59.10/53.8527.23/24.1679.07/83.1081.05/49.61
Out-of-Distribution: CIFAR-10/SVHNFL+MDCA [10]69.52/41.6627.93/19.8778.35/87.4677.92/59.08
DE-3 [5, 17]56.21/48.7026.61/24.2080.36/83.4282.34/53.17
DML [29]62.08/37.9527.07/19.6879.90/87.0681.07/65.32
Diff-DML(ours)56.21/37.8425.55/19.5881.50/87.3483.57/65.65
+ +Table 5. The OOD detection performance of different calibration methods. The numbers before and after the backslash denote the near and far OOD detection results, respectively. + +sizes. This dependency can complicate scalability and computational efficiency, particularly in resource-constrained environments. + +In conclusion, we introduce Diff-DML, which overcomes the limitations of traditional DML by incorporating two strategies: DTS and DPLO. While DML promotes learning from an ensemble distribution, it often reduces prediction diversity, which is crucial for effective calibration. Diff-DML preserves this diversity while simultaneously + +improving both accuracy and calibration. Extensive experiments on datasets such as CIFAR-100 demonstrate that our approach outperforms state-of-the-art methods, offering a practical and efficient solution for reliable uncertainty estimation, particularly in safety-critical applications. + +# References + +[1] Gavin Brown. Diversity in neural network ensembles. PhD thesis, CiteSeer, 2004. 4 +[2] Peng Cui, Dan Zhang, Zhijie Deng, Yinpeng Dong, and Jun Zhu. Learning sample difficulty from pre-trained models for reliable prediction. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4 +[4] James M Dolezal, Andrew Srisuwananukorn, Dmitry Karpeyev, Siddhi Ramesh, Sara Kochanny, Brittany Cody, Aaron S Mansfield, Sagar Rakshit, Radhika Bansal, Melanie C Bois, et al. Uncertainty-informed deep learning models enable high-confidence predictions for digital histopathology. Nature communications, 13(1):6572, 2022. 1 +[5] Stanislav Fort, Huiyi Hu, and Balaji Lakshminarayanan. Deep ensembles: A loss landscape perspective. arXiv preprint arXiv:1912.02757, 2019. 2, 5, 6, 7, 8, 1 +[6] Kevin P Greenman, Ava P Amini, and Kevin K Yang. Benchmarking uncertainty quantification for protein engineering. bioRxiv, pages 2023-04, 2023. 1 +[7] Arthur Gretton. Introduction to rkhs, and some simple kernel algorithms. Adv. Top. Mach. Learn. Lecture Conducted from University College London, 16(5-3):2, 2013. 1 +[8] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 2, 3, 6 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[10] Ramya Hebbalaguppe, Jatin Prakash, Neelabh Madan, and Chetan Arora. A stitch in time saves nine: A train-time regularizing loss for improved neural network calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16081-16090, 2022. 2, 3, 5, 6, 8, 1 +[11] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2016. 7 +[12] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joe Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. 7 +[13] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 4 +[14] Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. In Artificial Intelligence and Statistics, pages 623-631. PMLR, 2017. 1 +[15] Meelis Kull, Miquel Perello Nieto, Markus Kangsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities + +with dirichlet calibration. Advances in neural information processing systems, 32, 2019. 2, 3, 6, 1 +[16] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In International Conference on Machine Learning, pages 2805-2814. PMLR, 2018. 2, 3, 5, 6, 8, 1 +[17] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 2, 5, 6, 8, 1 +[18] Christian Leibig, Vaneeda Allken, Murat Seçkin Ayhan, Philipp Berens, and Siegfried Wahl. Leveraging uncertainty information from deep neural networks for disease detection. Scientific reports, 7(1):1-14, 2017. 1 +[19] Gongbo Liang, Yu Zhang, Xiaoqin Wang, and Nathan Jacobs. Improved trainable calibration method for neural networks on medical imaging classification. arXiv preprint arXiv:2009.04057, 2020.3,5,6,8,1 +[20] Rhiannon Michelmore, Matthew Wicker, Luca Laurenti, Luca Cardelli, Yarin Gal, and Marta Kwiatkowska. Uncertainty quantification with statistical guarantees in end-to-end autonomous driving control. In International Conference on Robotics and Automation, 2020. 1 +[21] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. Advances in Neural Information Processing Systems, 33:15288-15299, 2020. 3, 5, 6, 8, 1 +[22] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 2, 3, 1 +[23] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Proceedings of the AAAI conference on artificial intelligence, 2015. 2 +[24] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning, page 7. Granada, Spain, 2011. 7 +[25] Rishabh Patra, Ramya Hebbalaguppe, Tirtharaj Dash, Gautam Shroff, and Lovekesh Vig. Calibrating deep neural networks using explicit regularisation and dynamic data pruning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1541-1549, 2023. 3, 5, 6, 8, 1 +[26] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey Hinton. Regularizing neural networks by penalizing confident output distributions. arXiv preprint arXiv:1701.06548, 2017. 2, 3, 1 +[27] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61-74, 1999. 2, 3, 1 +[28] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE + +conference on computer vision and pattern recognition, pages 2818-2826, 2016. 2, 3, 1 +[29] Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4320-4328, 2018. 2, 3, 4, 6, 7, 8 +[30] Zhi-Hua Zhou. Ensemble methods: foundations and algorithms. CRC press, 2012. 4 \ No newline at end of file diff --git a/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/images.zip b/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..14dd64b990bdeba67519bc724ad0be20284b9c10 --- /dev/null +++ b/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a4cfb403e7c2582956af2f580c394840e47326295d011df7481731b3316542c +size 525688 diff --git a/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/layout.json b/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c3cbebb93482e1e69151e5da3232a87fd6791ceb --- /dev/null +++ b/2025/Improving Accuracy and Calibration via Differentiated Deep Mutual Learning/layout.json @@ -0,0 +1,8223 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 102, + 550, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 550, + 122 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 550, + 122 + ], + "type": "text", + "content": "Improving Accuracy and Calibration via Differentiated Deep Mutual Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "spans": [ + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "text", + "content": "Han Liu" + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger}" + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "text", + "content": ", Peng Cui" + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "text", + "content": ", Binging Wang" + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "text", + "content": ", Weipeng Chen" + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "text", + "content": ", Yupeng Zhang" + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "text", + "content": ", Jun Zhu" + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "text", + "content": ", and Xiaolin Hu" + }, + { + "bbox": [ + 123, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1,3,4,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 142, + 171, + 468, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 171, + 468, + 199 + ], + "spans": [ + { + "bbox": [ + 142, + 171, + 468, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 171, + 468, + 199 + ], + "type": "text", + "content": "Department of Computer Science and Technology, Institute for AI, BNRist, Tsinghua University, Beijing 100084, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 268, + 200, + 341, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 200, + 341, + 212 + ], + "spans": [ + { + "bbox": [ + 268, + 200, + 341, + 212 + ], + "type": "text", + "content": "2Baichuan Inc." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 168, + 213, + 441, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 213, + 441, + 228 + ], + "spans": [ + { + "bbox": [ + 168, + 213, + 441, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 168, + 213, + 441, + 228 + ], + "type": "text", + "content": "Tsinghua Laboratory of Brain and Intelligence (THBI)," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 228, + 520, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 228, + 520, + 241 + ], + "spans": [ + { + "bbox": [ + 89, + 228, + 520, + 241 + ], + "type": "text", + "content": "IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing 100084, China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 241, + 473, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 241, + 473, + 255 + ], + "spans": [ + { + "bbox": [ + 137, + 241, + 473, + 255 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 137, + 241, + 473, + 255 + ], + "type": "text", + "content": "Chinese Institute for Brain Research (CIBR), Beijing 100010, China" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 194, + 257, + 416, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 257, + 416, + 269 + ], + "spans": [ + { + "bbox": [ + 194, + 257, + 416, + 269 + ], + "type": "text", + "content": "{han-liu18, cuip22}@mails.tsinghua.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 167, + 271, + 443, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 271, + 443, + 283 + ], + "spans": [ + { + "bbox": [ + 167, + 271, + 443, + 283 + ], + "type": "text", + "content": "{daniel, chenweipeng, zhangyupeng}@baichuan-inc.com" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 214, + 285, + 397, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 285, + 397, + 297 + ], + "spans": [ + { + "bbox": [ + 214, + 285, + 397, + 297 + ], + "type": "text", + "content": "{dcszj, xlhu}@mail.tsinghua.edu.cn" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 324, + 200, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 324, + 200, + 337 + ], + "spans": [ + { + "bbox": [ + 151, + 324, + 200, + 337 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 350, + 297, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 297, + 662 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 297, + 662 + ], + "type": "text", + "content": "Deep Neural Networks (DNNs) have achieved remarkable success in a variety of tasks, particularly in terms of prediction accuracy. However, in real-world scenarios, especially in safety-critical applications, accuracy alone is insufficient; reliable uncertainty estimates are essential. Modern DNNs, often trained with cross-entropy loss, tend to exhibit overconfidence, especially on ambiguous samples. Many techniques aim to improve uncertainty calibration, yet they often come at the cost of reduced accuracy or increased computational demands. To address this challenge, we propose Differentiated Deep Mutual Learning (Diff-DML), an efficient ensemble approach that simultaneously enhances accuracy and uncertainty calibration. Diff-DML draws inspiration from Deep Mutual Learning (DML) while introducing two strategies to maintain prediction diversity: (1) Differentiated Training Strategy (DTS) and (2) Diversity-Preserving Learning Objective (DPLO). Our theoretical analysis shows that Diff-DML's diversified learning framework not only leverages ensemble benefits but also avoids the loss of prediction diversity observed in traditional DML setups, which is crucial for improved calibration. Extensive evaluations on various benchmarks confirm the effectiveness of Diff-DML. For instance, on the CIFAR-100 dataset, Diff-DML on ResNet34/50 models achieved substantial improvements over the previous state-of-the-art method, MDCA, with absolute accuracy gains of " + }, + { + "bbox": [ + 53, + 350, + 297, + 662 + ], + "type": "inline_equation", + "content": "1.3\\% / 3.1\\%" + }, + { + "bbox": [ + 53, + 350, + 297, + 662 + ], + "type": "text", + "content": ", relative ECE reductions" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 335, + 326, + 533, + 467 + ], + "blocks": [ + { + "bbox": [ + 335, + 326, + 533, + 467 + ], + "lines": [ + { + "bbox": [ + 335, + 326, + 533, + 467 + ], + "spans": [ + { + "bbox": [ + 335, + 326, + 533, + 467 + ], + "type": "image", + "image_path": "b25785f2e35a5aba3aca0aa4a060bf130a35505e1c140444266a82b18197c884.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 478, + 555, + 511 + ], + "lines": [ + { + "bbox": [ + 313, + 478, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 555, + 511 + ], + "type": "text", + "content": "Figure 1. Scatter plot of ECE and accuracy for different calibration methods based on the ResNet50 backbone on the CIFAR-100 test set. Methods located closer to the top left corner perform better." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 559 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 313, + 534, + 556, + 559 + ], + "type": "inline_equation", + "content": "49.6\\% / 43.8\\%" + }, + { + "bbox": [ + 313, + 534, + 556, + 559 + ], + "type": "text", + "content": ", and relative classwise-ECE reductions of " + }, + { + "bbox": [ + 313, + 534, + 556, + 559 + ], + "type": "inline_equation", + "content": "7.7\\% / 13.0\\%" + }, + { + "bbox": [ + 313, + 534, + 556, + 559 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 584, + 395, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 584, + 395, + 597 + ], + "spans": [ + { + "bbox": [ + 314, + 584, + 395, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "content": "Deep Neural Networks (DNNs) have achieved great success in various tasks, especially in terms of prediction accuracy. In safety-critical areas such as medical diagnosis [4, 18], self-driving cars [20], and protein engineering [6], accurate uncertainty estimation (i.e., well-calibrated prediction) is crucial for reliable model deployment. For instance, a reliable model needs to show high uncertainty when making incorrect predictions. However, models trained with cross-entropy (CE) loss tend to be overconfident, especially when" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 673, + 295, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 703 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 703 + ], + "type": "text", + "content": "This work was supported by the National Key Research and Development Program of China (No. 2021ZD0200301) and the National Natural Science Foundation of China (No. U2341228)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 71, + 703, + 216, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 703, + 216, + 713 + ], + "spans": [ + { + "bbox": [ + 71, + 703, + 216, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 71, + 703, + 216, + 713 + ], + "type": "text", + "content": "Equal contribution. \\*Corresponding author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25812" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "encountering ambiguous samples, leading to predictions that are often close to " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": " confident and producing undesirable overconfidence [2, 8]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 110, + 297, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 110, + 297, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 297, + 361 + ], + "type": "text", + "content": "To address this issue, various regularization techniques have been explored during training, referred to as training-time calibration. For instance, entropy regularization (ER) [26] and label smoothing (LS) [22, 28] promote smoother predictive probability distributions to alleviate overconfidence by maximizing entropy or adjusting label distributions. Some methods aim to directly reduce calibration error during training to achieve multi-class calibration. For example, MMCE [16] introduces a trainable measure based on kernel mean embeddings, while MDCA [10] employs a differentiable auxiliary loss to minimize the gap between prediction confidence and accuracy. However, despite mitigating overconfidence, these techniques often compromise prediction accuracy due to excessive regularization. Another category of methods is post-processing, which adjusts output probabilities using Temperature Scaling (TS) [27], Vector Scaling (VS) [27], or Dirichlet Scaling (DS) [15]. However, these post-processing techniques typically rely on an additional validation dataset. Furthermore, they have limited learnable parameters, which may lead to suboptimal calibration performance [10]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 364, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 296, + 579 + ], + "type": "text", + "content": "Deep Ensembles (DE) [5, 17] offer a straightforward solution by combining multiple randomly initialized trained DNNs to improve prediction accuracy and uncertainty calibration. Yet, using multiple DNNs during training and inference significantly increases computational costs, which poses challenges for deploying DE in real-world scenarios. To address this limitation, we draw inspiration from Deep Mutual Learning (DML) [29], a technique used in model distillation, where an ensemble of student models learn collaboratively and teach each other throughout the training process. This work provides the theoretical analysis showing that DML encourages each model to learn from an ensemble distribution through a mimicry loss. The proposed theory motivates us to leverage DML as an implicit form of DE, aiming to achieve similar gains in accuracy and calibration as DE, while mitigating the high computational overhead because DML uses only one model for prediction during inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": "However, in standard DML, although the models start from different initial states, they become overly similar during mutual learning, which reduces the diversity of their outputs—a key factor needed for effective ensemble-based calibration. In this work, we propose Differentiated Deep Mutual Learning (Diff-DML), a novel approach that enhances the benefits of DML by promoting prediction diversity between models to achieve high accuracy and strong calibration performance. Diff-DML incorporates two key strategies that preserve output diversity: (1) Differentiated Training Strategy (DTS), which uses different learning rates and schedules" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "for different models to prevent premature convergence, and (2) Diversity-Preserving Learning Objective (DPLO), which removes cross-entropy loss for auxiliary models to avoid overfitting to the one-hot ground-truth labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 121, + 556, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 556, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 556, + 286 + ], + "type": "text", + "content": "These strategies help Diff-DML balance collaborative learning with prediction diversity, resulting in better uncertainty calibration and enhanced predictive performance. Extensive evaluations on various benchmarks, including image classification, misclassification detection, and out-of-distribution (OOD) detection, confirm Diff-DML's effectiveness. For example, on the CIFAR-100 dataset, Diff-DML on ResNet34/50 achieved significant improvements compared to the previous state-of-the-art method, MDCA [10], with absolute accuracy increases of " + }, + { + "bbox": [ + 313, + 121, + 556, + 286 + ], + "type": "inline_equation", + "content": "1.3\\% / 3.1\\%" + }, + { + "bbox": [ + 313, + 121, + 556, + 286 + ], + "type": "text", + "content": ", relative reductions in ECE of " + }, + { + "bbox": [ + 313, + 121, + 556, + 286 + ], + "type": "inline_equation", + "content": "49.6\\% / 43.8\\%" + }, + { + "bbox": [ + 313, + 121, + 556, + 286 + ], + "type": "text", + "content": ", and relative reductions in classwise-ECE of " + }, + { + "bbox": [ + 313, + 121, + 556, + 286 + ], + "type": "inline_equation", + "content": "7.7\\% / 13.0\\%" + }, + { + "bbox": [ + 313, + 121, + 556, + 286 + ], + "type": "text", + "content": ". The latter two metrics measure calibration performance, where lower values indicate better results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 288, + 556, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 556, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 556, + 348 + ], + "type": "text", + "content": "Figure 1 displays ECE and accuracy of different calibration methods, showing Diff-DML achieving favorable calibration and accuracy. Notably, Diff-DML incurs only minimal computational cost at inference, unlike DE, which requires multiple models to generate a single prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 358, + 484, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 358, + 484, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 358, + 484, + 372 + ], + "type": "text", + "content": "2. Background and related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 378, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 462 + ], + "type": "text", + "content": "Calibration, the process of ensuring that the predicted probabilities reflect the true likelihood of the events, plays a crucial role in machine learning models. It ensures that the prediction of machine learning models are not only accurate but also well-calibrated. Well-calibrated models provide reliable uncertainty estimates, which can help users and downstream systems make safer decisions by avoiding overconfidence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 462, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 498 + ], + "type": "text", + "content": "There are two fundamental types of calibration: confidence calibration [23] and classwise calibration [15]. In this paper, we focus on both types of calibration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 582 + ], + "type": "text", + "content": "Confidence calibration: Confidence calibration refers to the alignment of predicted confidence with observed frequencies. It ensures that a model's prediction confidence accurately reflects the true likelihood of an event occurring. A widely used metric for evaluating confidence calibration is the Expected Calibration Error (ECE) [23]. The definition of ECE can be found in Appendix A." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 665 + ], + "type": "text", + "content": "Classwise calibration: Classwise calibration is concerned with calibrating the predicted probabilities for each individual class in a multi-class classification problem. It ensures that the predicted probabilities across different classes are accurate and properly calibrated. The metric for classwise calibration is classwise-ECE (cw-ECE) [15], which can be found in Appendix A." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 555, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 701 + ], + "type": "text", + "content": "According to the number of models used for inference, calibration methods can be divided into single-model calibration methods and DE [5, 17]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Single-model calibration methods: Within single-model" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25813" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "calibration methods, there are two categories: train-time calibration methods and post-hoc calibration methods. Train-time calibration methods typically involve designing loss functions or incorporating additional regularization techniques, such as Entropy Regularization (ER) [26], Label Smoothing (LS) [22, 28], Focal Loss (FL) [21], Maximum Mean Calibration Error (MMCE) [16], Difference between Confidence and Accuracy (DCA) [19], Multi-class Difference in Confidence and Accuracy (MDCA) [10], and Dynamic Train-time Data Pruning (DTDP) [25]. Post-hoc calibration methods are applied after the model is trained, such as Temperature Scaling (TS) [27], Vector Scaling, and Dirichlet Scaling (DS) [15]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 312 + ], + "type": "text", + "content": "Deep Ensembles: DE creates an ensemble by training multiple models independently with different initializations. To alleviate overconfidence and improve calibration, during inference, DE leverages " + }, + { + "bbox": [ + 55, + 228, + 296, + 312 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 228, + 296, + 312 + ], + "type": "text", + "content": " randomly initialized, individually trained DNNs and makes predictions by average voting. While effective for enhancing model performance, DE introduces additional computational overhead during inference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 312, + 268, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 312, + 268, + 323 + ], + "spans": [ + { + "bbox": [ + 67, + 312, + 268, + 323 + ], + "type": "text", + "content": "More related works are discussed in Appendix B." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 111, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 111, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 111, + 346 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "inline_equation", + "content": "P_{data}" + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "content": " denote the probability distribution of data. A dataset consists of finite " + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "inline_equation", + "content": "(x, y_x)" + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "content": " pairs where " + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "inline_equation", + "content": "x \\sim P_{data}" + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "content": " is a data instance and " + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "inline_equation", + "content": "y_x \\in \\mathcal{Y} = \\{1, 2, \\dots, K\\}" + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "content": " is the ground-truth class label of " + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "content": ". A model takes in the instance " + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "content": " and outputs its prediction vector " + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "inline_equation", + "content": "p_x \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "inline_equation", + "content": "p_x^{(k)}" + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "content": "-th dimension of " + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "inline_equation", + "content": "p_x" + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "content": ". The confidence " + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\hat{p}_x = \\max_{k \\in \\mathcal{Y}} \\{p_x^{(k)}\\}" + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "content": " is the maximum value in the probability vector of predictions, and the model's prediction " + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\hat{y}_x = \\operatorname{argmax}_{k \\in \\mathcal{Y}} \\{p_x^{(k)}\\}" + }, + { + "bbox": [ + 55, + 354, + 296, + 467 + ], + "type": "text", + "content": " is the index corresponding to the maximum value." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 467, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 491 + ], + "type": "text", + "content": "Typically, we train a model by minimizing the following CE loss on the training set," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 500, + 295, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 500, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 127, + 500, + 295, + 515 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {C E}} \\left(p _ {x}\\right) = - \\log p _ {x} ^ {\\left(y _ {x}\\right)}, \\tag {1}", + "image_path": "2aa7487e28d909e01b3577877ca2a89b5cb0c714cc3eb12b6a7441150ce9bf50.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 525, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 525, + 296, + 598 + ], + "type": "inline_equation", + "content": "p_x^{(y_x)}" + }, + { + "bbox": [ + 55, + 525, + 296, + 598 + ], + "type": "text", + "content": " represents the prediction probability corresponding to the ground-truth label " + }, + { + "bbox": [ + 55, + 525, + 296, + 598 + ], + "type": "inline_equation", + "content": "y_{x}" + }, + { + "bbox": [ + 55, + 525, + 296, + 598 + ], + "type": "text", + "content": ". Previous studies [2, 8, 21] reveal that minimizing the CE loss forces the model to overfit to one-hot encoded ground-truth labels with almost " + }, + { + "bbox": [ + 55, + 525, + 296, + 598 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 525, + 296, + 598 + ], + "type": "text", + "content": " probability irrespective of their correctness, which can result in overconfidence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 599, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 296, + 647 + ], + "type": "text", + "content": "DML [29] is a model distillation technique that allows two models " + }, + { + "bbox": [ + 55, + 599, + 296, + 647 + ], + "type": "inline_equation", + "content": "f(\\theta)" + }, + { + "bbox": [ + 55, + 599, + 296, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 599, + 296, + 647 + ], + "type": "inline_equation", + "content": "g(\\phi)" + }, + { + "bbox": [ + 55, + 599, + 296, + 647 + ], + "type": "text", + "content": " to learn collaboratively through a combination of a supervised loss and a Kullback-Leibler (KL) Divergence:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 79, + 656, + 296, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 656, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 79, + 656, + 296, + 670 + ], + "type": "interline_equation", + "content": "L _ {f, x} (\\theta) = L _ {\\mathrm {C E}} \\left(f _ {x} (\\theta)\\right) + \\alpha D _ {\\mathrm {K L}} \\left(g _ {x} (\\phi) \\mid \\mid f _ {x} (\\theta)\\right), \\tag {2}", + "image_path": "6e1eb261811fd71537225c65d04f34607eb72b238db5a47d84b318e092d87352.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 78, + 700, + 296, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 700, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 78, + 700, + 296, + 715 + ], + "type": "interline_equation", + "content": "L _ {g, x} (\\phi) = L _ {\\mathrm {C E}} \\left(g _ {x} (\\phi)\\right) + \\alpha D _ {\\mathrm {K L}} \\left(f _ {x} (\\theta) \\| g _ {x} (\\phi)\\right), \\tag {3}", + "image_path": "7c6636225ff21ae13a5208d05e5f48432e4426250fc07e8790afa060a6b52dfa.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": " is a hyperparameter used to adjust the weight of the KL divergence loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "text", + "content": "In this paper, we designate " + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "inline_equation", + "content": "f(\\theta)" + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "text", + "content": " as the primary model, and " + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "inline_equation", + "content": "g(\\phi)" + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "text", + "content": " as the auxiliary model that does not participate in the inference process. It is important to note that DML was originally conceived as a knowledge distillation technique aimed at improving accuracy rather than focusing on calibration." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 175, + 538, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 538, + 187 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 538, + 187 + ], + "type": "text", + "content": "3.1. DML is an implicit form of deep ensembles" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 191, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 555, + 299 + ], + "type": "text", + "content": "DML involves two models learning collaboratively by continuously exchanging knowledge during the training process. This collaborative learning encourages each model to learn from the strengths and correct the weaknesses of others, promoting diversity in model predictions. Consequently, DML achieves the advantages of DE—such as increased prediction accuracy and improved uncertainty calibration—without the need for independent model training or the associated computational costs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 300, + 555, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 322 + ], + "type": "text", + "content": "In this section, we explore the theoretical relationship between DML and DE." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 323, + 555, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 360 + ], + "type": "text", + "content": "Given two trained models " + }, + { + "bbox": [ + 313, + 323, + 555, + 360 + ], + "type": "inline_equation", + "content": "f(\\theta)" + }, + { + "bbox": [ + 313, + 323, + 555, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 360 + ], + "type": "inline_equation", + "content": "g(\\phi)" + }, + { + "bbox": [ + 313, + 323, + 555, + 360 + ], + "type": "text", + "content": " and an input " + }, + { + "bbox": [ + 313, + 323, + 555, + 360 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 323, + 555, + 360 + ], + "type": "text", + "content": " for DE, we can get a weighted sum of predictive probabilities from " + }, + { + "bbox": [ + 313, + 323, + 555, + 360 + ], + "type": "inline_equation", + "content": "f_{x}(\\theta)" + }, + { + "bbox": [ + 313, + 323, + 555, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 360 + ], + "type": "inline_equation", + "content": "g_{x}(\\phi)" + }, + { + "bbox": [ + 313, + 323, + 555, + 360 + ], + "type": "text", + "content": " as the final prediction:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 367, + 555, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 367, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 359, + 367, + 555, + 380 + ], + "type": "interline_equation", + "content": "h _ {x} (\\theta , \\phi) = (1 - \\beta) f _ {x} (\\theta) + \\beta g _ {x} (\\phi), \\tag {4}", + "image_path": "2cbe0a9e399146bc1515564c3d34d9284dfbba2a89f85db18f0ca25fa66639a2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 387, + 555, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 423 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 387, + 555, + 423 + ], + "type": "inline_equation", + "content": "\\beta \\in [0,1]" + }, + { + "bbox": [ + 313, + 387, + 555, + 423 + ], + "type": "text", + "content": " is a weight to balance the effect of each model. To explain the relationship between DML and DE, we first present a proposition, the proof is in Appendix C)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "type": "text", + "content": "Proposition 1. Given two prediction probability distributions " + }, + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "type": "inline_equation", + "content": "f_{x}(\\theta)" + }, + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "type": "inline_equation", + "content": "g_{x}(\\phi)" + }, + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "type": "text", + "content": " for classifying sample " + }, + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "type": "text", + "content": ", as well as their ensemble " + }, + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "type": "inline_equation", + "content": "h_{x}(\\underline{\\theta},\\phi)" + }, + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 322, + 465, + 554, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 465, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 322, + 465, + 554, + 498 + ], + "type": "interline_equation", + "content": "\\frac {\\partial}{\\partial \\theta} D _ {K L} \\left(h _ {x} (\\underline {{\\theta}}, \\phi) | | f _ {x} (\\theta)\\right) = \\beta \\frac {\\partial}{\\partial \\theta} D _ {K L} \\left(g _ {x} (\\phi) | | f _ {x} (\\theta)\\right). \\tag {5}", + "image_path": "abc20e4eb2f602307142df9395f3eda33989127101d4e883ca1aab53c123bdf7.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 499, + 554, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 554, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 554, + 522 + ], + "type": "text", + "content": "Note that the variables with underscores do not participate in gradient propagation." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "content": "The left-hand side of (5) suggests that when we update " + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "content": " to make " + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "content": " closer to " + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "inline_equation", + "content": "h_x" + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "content": ", we temporarily treat " + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "inline_equation", + "content": "h_x" + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "content": " as a model with fixed parameters, though, in the next step, " + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "inline_equation", + "content": "h_x" + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "content": " will also be updated with the new " + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 522, + 554, + 582 + ], + "type": "text", + "content": ". By substituting (5) into (2) we obtain" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 326, + 586, + 555, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 586, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 326, + 586, + 555, + 609 + ], + "type": "interline_equation", + "content": "L _ {f, x} (\\theta) = L _ {\\mathrm {C E}} \\left(f _ {x} (\\theta)\\right) + \\frac {\\alpha}{\\beta} D _ {\\mathrm {K L}} \\left(h _ {x} (\\underline {{\\theta}}, \\phi) | | f _ {x} (\\theta)\\right). \\tag {6}", + "image_path": "832abc1202dec75f7d405aae43db2740a02819d5feffde43aaa3f9f51434dc6b.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 615, + 512, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 512, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 512, + 627 + ], + "type": "text", + "content": "Similarly, the loss of model " + }, + { + "bbox": [ + 313, + 615, + 512, + 627 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 615, + 512, + 627 + ], + "type": "text", + "content": " can be rewritten as" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 632, + 554, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 632, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 321, + 632, + 554, + 666 + ], + "type": "interline_equation", + "content": "L _ {g, x} (\\phi) = L _ {\\mathrm {C E}} \\left(g _ {x} (\\phi)\\right) + \\frac {\\alpha}{1 - \\beta} D _ {\\mathrm {K L}} \\left(h _ {x} \\left(\\theta , \\underline {{\\phi}}\\right) \\mid \\mid g _ {x} (\\phi)\\right). \\tag {7}", + "image_path": "e070a3e028ef808eba1d7707d8c6fb4dca82f523a17d396f9c8bcb7fdc5a0196.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "From this perspective, we can see that KL divergence in DML helps continuously align each model's predictions with the ensemble distribution during training, thereby DML can be regarded as an implicit DE." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25814" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 254, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 254, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 254, + 85 + ], + "type": "text", + "content": "3.2. Differentiated Deep Mutual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "text", + "content": "Model diversity plays a crucial role in the effectiveness of the ensemble [1, 30]. Although DML offers the opportunity to reduce computational overhead compared to DE, prior learning strategies in DML may destroy the model diversity and result in poor calibration performance. To address this issue, we propose an improved approach, Diff-DML, which increases the output discrepancy between " + }, + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "text", + "content": " during training. Diff-DML employs two methods to increase model discrepancy: Differentiated Training Strategy (DTS) and Diversity-Preserving Learning Objective (DPLO)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 210, + 296, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 296, + 305 + ], + "type": "text", + "content": "Differentiated Training Strategy: In traditional DML training [29], both models typically use the same scheduler, beginning with a specific learning rate and gradually decreasing it over time. This consistency in training strategies often leads to high similarity between the two models, diminishing their diversity. A straightforward way to increase the diversity between the outputs of the two models is to employ different training schedulers in training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "text", + "content": "In light of this, we propose to use different training strategies: one model, " + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "text", + "content": ", is designated as the primary model (used solely during inference) and follows the conventional strategy of gradually decreasing the learning rate. The other model, " + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "text", + "content": ", serves as an auxiliary model and is trained using a fixed learning rate throughout the training process. This discrepant learning rate scheme can preserve the diversity among models and thus improve the calibration performance. Diversity-Preserving Learning Objective: To increase the disparity between the two models, we propose a Diversity-Preserving Learning Objective by changing the loss function of the auxiliary model in DML. In traditional DML, the models " + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "inline_equation", + "content": "f(\\theta)" + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "inline_equation", + "content": "g(\\phi)" + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "text", + "content": " share the same form of loss function, which combines a supervised loss with KL divergence. This learning objective may cause the model to overfit the one-hot ground-truth labels, resulting in unavoidable overconfidence and insufficient diversity. To address this problem, we suggest that the primary model " + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "text", + "content": " continue using the original loss function (2), while the auxiliary model " + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "inline_equation", + "content": "g(\\phi)" + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "text", + "content": " omits the cross-entropy loss component. As a result, the auxiliary model no longer receives direct supervision from the labels. Concretely, the new loss function of model " + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "inline_equation", + "content": "g(\\phi)" + }, + { + "bbox": [ + 55, + 306, + 296, + 571 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 578, + 296, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 578, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 111, + 578, + 296, + 594 + ], + "type": "interline_equation", + "content": "L _ {g, x} (\\phi) = D _ {\\mathrm {K L}} \\left(f _ {x} (\\theta) | | g _ {x} (\\phi)\\right). \\tag {8}", + "image_path": "b9dcbf4db722f051e200c23a10d907a669042bb80bae086b216ec83042d4a423.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 601, + 297, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 297, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 297, + 651 + ], + "type": "text", + "content": "The KL divergence not only constructs a mutual feedback loop between the primary and auxiliary models but also assists the auxiliary model in learning a differentiated prediction distribution to boost model diversity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 658, + 261, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 261, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 261, + 671 + ], + "type": "text", + "content": "3.3. Extending to multiple auxiliary models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 676, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 676, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 676, + 297, + 715 + ], + "type": "text", + "content": "Naturally, we can incorporate multiple auxiliary models " + }, + { + "bbox": [ + 55, + 676, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\{g_i\\}_{i=1}^{N_g}" + }, + { + "bbox": [ + 55, + 676, + 297, + 715 + ], + "type": "text", + "content": " into the training process to further improve the prediction performance and calibration of the primary model " + }, + { + "bbox": [ + 55, + 676, + 297, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 676, + 297, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "content": "Analogous to DE, leveraging multiple auxiliary models can effectively promote model diversity and improve the uncertainty calibration of the primary model. The optimization objectives for the primary model " + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 129, + 555, + 173 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 173 + ], + "type": "interline_equation", + "content": "L _ {f, x} (\\theta) = L _ {\\mathrm {C E}} \\left(f _ {x} (\\theta)\\right) + \\alpha \\sum_ {i = 1} ^ {N _ {g}} \\frac {1}{N _ {g}} D _ {\\mathrm {K L}} \\left(g _ {i, x} \\left(\\phi_ {i}\\right) \\mid \\mid f _ {x} (\\theta)\\right), \\tag {9}", + "image_path": "4e37ad2f005d43089de9cbf93f19c9f028c34a1f80f689b993a83df304ef1128.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 174, + 553, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 553, + 196 + ], + "type": "text", + "content": "For traditional DML, the optimization objectives for each " + }, + { + "bbox": [ + 313, + 174, + 553, + 196 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 313, + 174, + 553, + 196 + ], + "type": "text", + "content": " are formalized as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 206, + 555, + 231 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 206, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 325, + 206, + 555, + 231 + ], + "type": "interline_equation", + "content": "L _ {g _ {i, x}} \\left(\\phi_ {i}\\right) = L _ {\\mathrm {C E}} \\left(g _ {i, x} \\left(\\phi_ {i}\\right)\\right) + \\alpha D _ {\\mathrm {K L}} \\left(f _ {x} (\\theta) \\mid \\mid g _ {i, x} \\left(\\phi_ {i}\\right)\\right), \\tag {10}", + "image_path": "0da9ff92eb59bcc25d0ba518e28519e4fbb0a805440404814c37f2cc4c8d76e7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 232, + 555, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 267 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 232, + 555, + 267 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 313, + 232, + 555, + 267 + ], + "type": "text", + "content": " denotes the trainable parameters of the auxiliary model " + }, + { + "bbox": [ + 313, + 232, + 555, + 267 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 313, + 232, + 555, + 267 + ], + "type": "text", + "content": ". And for Diff-DML, the optimization objectives for each " + }, + { + "bbox": [ + 313, + 232, + 555, + 267 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 313, + 232, + 555, + 267 + ], + "type": "text", + "content": " are formalized as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 362, + 276, + 555, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 276, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 362, + 276, + 555, + 291 + ], + "type": "interline_equation", + "content": "L _ {g _ {i, x}} \\left(\\phi_ {i}\\right) = D _ {\\mathrm {K L}} \\left(f _ {x} (\\theta) \\mid \\mid g _ {i, x} \\left(\\phi_ {i}\\right)\\right). \\tag {11}", + "image_path": "35215eaf433814269dd4ef840e7d1c9df4ed667aa824533251f72b0e617f2a68.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 299, + 542, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 542, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 542, + 312 + ], + "type": "text", + "content": "Similarly, we only use the primary model " + }, + { + "bbox": [ + 313, + 299, + 542, + 312 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 299, + 542, + 312 + ], + "type": "text", + "content": " in inference." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 322, + 395, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 395, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 395, + 335 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 342, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 389 + ], + "type": "text", + "content": "We conducted primary experiments on image classification benchmarks, followed by experiments on misclassification and OOD detection to validate the performance of Diff-DML across different tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 397, + 429, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 429, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 429, + 410 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 415, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 498 + ], + "type": "text", + "content": "Datasets: We trained and evaluated our multi-class classification model using CIFAR-10 [13], CIFAR-100 [13] and Tiny-ImageNet [3] datasets. Validation sets were created by randomly selecting 5,000 samples from each dataset's training set, with the remaining samples used for training. We experimented with two network architectures, ResNet34 and ResNet50 [9], on these datasets." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "text", + "content": "Evaluation metrics: For standard classification, we evaluated all methods using accuracy (acc), ECE, and cw-ECE. For misclassification and OOD detection, we used four metrics: (1) False positive rate at " + }, + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "text", + "content": " true positive rate (FPR- " + }, + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "text", + "content": "); (2) The misclassification probability when true positive rate is " + }, + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "text", + "content": " (D-error); (3) Area under the receiver operating characteristic curve (AUROC); (4) Area under the precision-recall curve (AUPR)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "content": "Selection of models for testing: When selecting the optimal calibration model, we need to evaluate both accuracy and calibration performance. It is insufficient to consider only accuracy or only calibration performance. To account for both aspects simultaneously, we have developed a new method for selecting models to be tested on the test set, which we refer to as the Comprehensive Selection Method for Accuracy and Calibration (CSMAC). The selection method is as follows: A threshold value for accuracy, denoted as " + }, + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "content": ", is set. When the highest accuracy of a model on the validation set is" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25815" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "content": "below " + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "content": ", the checkpoint with the highest accuracy is selected. When the highest accuracy exceeds " + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "content": ", the checkpoint with the highest value of accuracy minus ECE (which we call the comprehensive evaluation score) on the validation set is selected. The value of " + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "content": " varies depending on the dataset and base model. In the experiment, we first train a model optimized using cross-entropy loss and identify the checkpoint corresponding to the highest accuracy on the validation set. We set this checkpoint as the baseline model, and denote its accuracy on the validation set as " + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "content": ". Then, we set " + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "inline_equation", + "content": "b = a - c" + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "content": " represents a tolerance value for accuracy reduction, and in our experiment, " + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "content": " is set to 0.01. In this way, we establish a selection mechanism that takes both accuracy and calibration performance into account." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "text", + "content": "Competing methods and hyper-parameters: The evaluated methods include: Baseline: This is the model with the highest accuracy on the validation set trained with the CE loss. Note that the baseline model is used to determine the threshold " + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "text", + "content": " for the CSMAC selection method. All other models use this threshold to make selections through CSMAC. Cross-Entropy (CE): This is the model trained with CE loss (1) selected by CSMAC. DML & DML-2: These refer to the DML method using 1 and 2 auxiliary ResNet18 models, respectively. The primary model used was either ResNet34 or ResNet50. For both DML-1 and DML-2, we report the performance of the primary models. Diff-DML & Diff-DML-2: Similar to DML-1/2, these use Diff-DML for training with 1 or 2 auxiliary ResNet18 models. The performance of the primary model, either ResNet34 or ResNet50, is reported. For DML and Diff-DML, the value " + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0.4, 1.0]" + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "text", + "content": " and we tuned it to achieve the best performance via the validation set. DCA [19]: Four values were used for its " + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "text", + "content": ": 0.5, 1.0, 5.0, and 10.0. MMCE [16]: Four values were used for its " + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "text", + "content": ": 1.0, 2.0, 3.0, and 4.0. FL [21]: Three values were used for its " + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "text", + "content": ": 1.0, 2.0, and 3.0. FL+MDCA [10]: Three values were used for the " + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "text", + "content": " in the focal loss: 1.0, 2.0, and 3.0, while the " + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "text", + "content": " value in MDCA was 1.0. FL+DTDP [25]: Three values were used for the " + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "text", + "content": " in the focal loss: 1.0, 2.0, and 3.0. For DTDP, we used an " + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "text", + "content": " value of 0.005 and experimented with " + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "text", + "content": " values of 5.0 and 10.0, along with a data pruning ratio of " + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 56, + 243, + 298, + 602 + ], + "type": "text", + "content": ". DE-3 [5, 17]: It denotes DE with 3 models using cross-entropy loss with different initializations and combined their produced probabilities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Other implementation details: In the experiments, we utilized the SGD optimizer with a momentum of 0.9 and weight decay of 0.0005. The initial learning rate was set to 0.1 for all datasets. For each dataset, we trained the models for a total of 250 epochs. During the initial 100 epochs, we maintained the initial learning rate. Afterward, every 50 epochs, we decreased the learning rate by a factor of " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ". In Diff-DML, the primary model " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " employed the same optimization strategy as described above. However, for the" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 556, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 120 + ], + "type": "text", + "content": "auxiliary model " + }, + { + "bbox": [ + 313, + 72, + 556, + 120 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 72, + 556, + 120 + ], + "type": "text", + "content": ", we initialized the learning rate as 0.01 and did not apply any learning rate decay. All experiments were performed with a batch size of 100 and conducted for training and testing on a single GeForce RTX 2080 GPU." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 129, + 519, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 519, + 142 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 519, + 142 + ], + "type": "text", + "content": "4.2. Classification accuracy and calibration" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 148, + 555, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 231 + ], + "type": "text", + "content": "Table 1 shows the comparison results of Diff-DML and various single-model calibration methods on CIFAR-100. Tables S1 and S2 in Appendix D show the comparison results on CIFAR-10 and Tiny/ImageNet, respectively. Figure S1 in Appendix E visually demonstrates the calibration performance of Diff-DML with reliability diagrams and confidence histograms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 232, + 556, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 556, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 556, + 292 + ], + "type": "text", + "content": "Comparison between DML and Diff-DML: From Tables 1, S1 and S2, it can be observed that on all three datasets, DML and Diff-DML achieved improvements in both accuracy and calibration performance compared to CE. The improvements of Diff-DML were more significant than those of DML." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 293, + 556, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 556, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 556, + 363 + ], + "type": "text", + "content": "Figure 2 shows the validation accuracy and ECE curves of ResNet34 during training on the CIFAR-100 dataset using CE, DML and Diff-DML methods. It can be seen that the ECE values for CE and DML steadily increased throughout training, while the ECE values for Diff-DML showed a downward trend." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 365, + 557, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 557, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 557, + 449 + ], + "type": "text", + "content": "Comparison with regularization methods: It can be observed that the majority of regularization methods compromised a certain level of accuracy to improve calibration ability. For instance, a previous state-of-the-art method, FL+MDCA, reduced accuracy by 1 to 2 percent. In contrast, our Diff-DML method avoided this trade-off and demonstrates superior accuracy and calibration ability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 449, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 533 + ], + "type": "text", + "content": "In terms of calibration performance, Diff-DML surpassed FL+MDCA across all three datasets. Specifically, compared to FL+MDCA, Diff-DML achieved an absolute accuracy increase of " + }, + { + "bbox": [ + 313, + 449, + 556, + 533 + ], + "type": "inline_equation", + "content": "1.3\\% / 3.1\\%" + }, + { + "bbox": [ + 313, + 449, + 556, + 533 + ], + "type": "text", + "content": " for ResNet34/ResNet50 on the CIFAR-100 dataset, with relative reductions in ECE of " + }, + { + "bbox": [ + 313, + 449, + 556, + 533 + ], + "type": "inline_equation", + "content": "49.6\\% / 43.8\\%" + }, + { + "bbox": [ + 313, + 449, + 556, + 533 + ], + "type": "text", + "content": " and relative reductions in cw-ECE of " + }, + { + "bbox": [ + 313, + 449, + 556, + 533 + ], + "type": "inline_equation", + "content": "7.7\\% / 13.0\\%" + }, + { + "bbox": [ + 313, + 449, + 556, + 533 + ], + "type": "text", + "content": " representing a considerable improvement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 533, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 605 + ], + "type": "text", + "content": "Comparison with post-hoc calibration methods: Post-hoc calibration methods in general did not affect accuracy much, but their calibration performance on datasets with many classes, like CIFAR-100 and Tiny-[ImageNet, was often suboptimal. In contrast, Diff-DML showed much better calibration performance on these datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "Results with more auxiliary models: Table 2 shows the results of DML and Diff-DML trained with more auxiliary models on the CIFAR-100 dataset, compared to DE-3. Table S3 in Appendix D presents the comparison on CIFAR-10 and Tiny/ImageNet. It can be observed that DE-3 provided the most significant improvement in accuracy. DML and Diff-DML with two auxiliary models slightly outperformed those with one auxiliary model, but the difference was not substantial. In terms of calibration performance, Diff-DML" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25816" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 531, + 300 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 531, + 300 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 531, + 300 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 531, + 300 + ], + "type": "table", + "html": "
MethodAcc(%)↑ECE(%)↓cw-ECE(10-3)↓
ResNet34ResNet50ResNet34ResNet50ResNet34ResNet50
Baseline77.8777.858.8110.392.252.51
CE77.7377.548.5510.042.292.44
Post-hoc calibration methods
CE+TS [8]77.7377.543.373.141.721.63
CE+VS [8]77.4777.333.763.131.821.71
CE+DS [15]77.3377.543.203.141.711.63
Regularization based calibration methods
DCA [19]77.5377.6610.1011.692.432.66
MMCE [16]75.5075.166.267.191.851.98
FL [21]76.9376.491.211.791.561.55
FL+DTDP [25]75.8776.091.822.611.601.65
FL+MDCA [10]76.9175.931.291.531.551.61
DML-based calibration methods
DML [29]77.3677.893.434.421.561.69
Diff-DML (ours)78.2079.050.650.861.431.40
", + "image_path": "e2da281170812d5b061917d1e135a9b1a401cee25879aca50e003f4baea8fa78.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 331, + 221, + 454 + ], + "blocks": [ + { + "bbox": [ + 60, + 331, + 221, + 454 + ], + "lines": [ + { + "bbox": [ + 60, + 331, + 221, + 454 + ], + "spans": [ + { + "bbox": [ + 60, + 331, + 221, + 454 + ], + "type": "image", + "image_path": "e1df126f479648a8e8a414180bd5c8f73d5f3605bc716d8db6c500d0766a6a53.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 460, + 555, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 460, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 555, + 483 + ], + "type": "text", + "content": "Figure 2. The validation accuracy and ECE curves of ResNet34 trained using the (a) CE, (b) DML, and (c) Diff-DML methods on the CIFAR-100 dataset during the training process." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 329, + 386, + 453 + ], + "blocks": [ + { + "bbox": [ + 221, + 329, + 386, + 453 + ], + "lines": [ + { + "bbox": [ + 221, + 329, + 386, + 453 + ], + "spans": [ + { + "bbox": [ + 221, + 329, + 386, + 453 + ], + "type": "image", + "image_path": "7dfa6e90ddc024001f2dfd21a7eb6bb502e910b54d4bb390809ac875eda6eb8d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 386, + 329, + 552, + 453 + ], + "blocks": [ + { + "bbox": [ + 386, + 329, + 552, + 453 + ], + "lines": [ + { + "bbox": [ + 386, + 329, + 552, + 453 + ], + "spans": [ + { + "bbox": [ + 386, + 329, + 552, + 453 + ], + "type": "image", + "image_path": "a3cf201f15679ad64253e8ec6da9a0f1086430f20beebe87a638c62855456145.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 121, + 493, + 491, + 583 + ], + "blocks": [ + { + "bbox": [ + 130, + 307, + 477, + 318 + ], + "lines": [ + { + "bbox": [ + 130, + 307, + 477, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 477, + 318 + ], + "type": "text", + "content": "Table 1. Different calibration methods' accuracy, ECE, and cw-ECE on the CIFAR-100 dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 121, + 493, + 491, + 583 + ], + "lines": [ + { + "bbox": [ + 121, + 493, + 491, + 583 + ], + "spans": [ + { + "bbox": [ + 121, + 493, + 491, + 583 + ], + "type": "table", + "html": "
MethodAcc(%)↑ECE(%)↓cw-ECE(10-3)↓
ResNet34ResNet50ResNet34ResNet50ResNet34ResNet50
DE-3 [5, 17]79.3479.632.954.401.561.73
DML [29]77.3677.893.434.421.561.69
DML-2 [29]77.9878.004.324.381.611.82
Diff-DML (ours)78.2079.050.650.861.431.40
Diff-DML-2 (ours)79.1479.300.750.721.371.33
", + "image_path": "92297552382d178d5be9cd816f5a9b818af1d20554a6e3ae818c68e74ee00959.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 590, + 515, + 602 + ], + "lines": [ + { + "bbox": [ + 94, + 590, + 515, + 602 + ], + "spans": [ + { + "bbox": [ + 94, + 590, + 515, + 602 + ], + "type": "text", + "content": "Table 2. Comparison of DML and Diff-DML trained with more auxiliary models against DE on CIFAR-100 dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 623, + 175, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 175, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 175, + 635 + ], + "type": "text", + "content": "was clearly superior to DE-3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "content": "Training and inference time: In terms of time overhead, training Diff-DML with one auxiliary model took 1.4 times longer than training a single model, and with two auxiliary models, it took 1.9 times longer. In contrast, DE-3's training time was 3 times that of a single model. For inference, Diff-DML required the same time as a single model, while" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 623, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 555, + 658 + ], + "type": "text", + "content": "DE-3's inference time was 3 times longer. Thus, our method provided substantial advantages over DE-3 in both training and inference time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 670, + 520, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 520, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 520, + 683 + ], + "type": "text", + "content": "4.3. Analysis of the robustness of Diff-DML" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Sensitivity of " + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": ": We investigated the sensitivity of the important hyper-parameter " + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": " in (2). From Figure 3, it is ob" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25817" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 62, + 229, + 178 + ], + "blocks": [ + { + "bbox": [ + 59, + 62, + 229, + 178 + ], + "lines": [ + { + "bbox": [ + 59, + 62, + 229, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 62, + 229, + 178 + ], + "type": "image", + "image_path": "cb40b5b709e0ee8deedc36f94ab497bfad69e5d264bc03d029c04f29338d7864.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 191, + 555, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 555, + 214 + ], + "type": "text", + "content": "Figure 3. Variation of accuracy, ECE, and classwise-ECE of Diff-DML method trained on CIFAR-100 with different values of hyperparameter " + }, + { + "bbox": [ + 55, + 191, + 555, + 214 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 191, + 555, + 214 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 62, + 392, + 179 + ], + "blocks": [ + { + "bbox": [ + 228, + 62, + 392, + 179 + ], + "lines": [ + { + "bbox": [ + 228, + 62, + 392, + 179 + ], + "spans": [ + { + "bbox": [ + 228, + 62, + 392, + 179 + ], + "type": "image", + "image_path": "41c57e78b6e34e1c22078fe45e9ed2fdc397a82329ceee0d612f5cc5fa6996be.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 62, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 391, + 62, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 391, + 62, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 391, + 62, + 553, + 178 + ], + "type": "image", + "image_path": "087a0df88e10a212073ed47a9c748ecfae1caa92daf199130d86cd4c0c71c897.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 228, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 228, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 295, + 337 + ], + "type": "text", + "content": "served that on CIFAR-100, as " + }, + { + "bbox": [ + 54, + 228, + 295, + 337 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 228, + 295, + 337 + ], + "type": "text", + "content": " changes, both accuracy and cw-ECE exhibited smooth variations. Specifically, the fluctuations in accuracy remained within 3 percentage points, while the fluctuations in cw-ECE were within 0.02 percentage points. As for ECE, although its values ranged between 2 and 4 for very small (e.g., 0.4) and large (e.g., 1.0) " + }, + { + "bbox": [ + 54, + 228, + 295, + 337 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 228, + 295, + 337 + ], + "type": "text", + "content": " values, the fluctuations around the lowest point of ECE were also minimal. This demonstrates the good robustness of Diff-DML with respect to the hyper-parameter " + }, + { + "bbox": [ + 54, + 228, + 295, + 337 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 228, + 295, + 337 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 337, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 504 + ], + "type": "text", + "content": "Selection of auxiliary model: We investigated how to select the auxiliary model. We experimented with CNNs containing two hidden layers, as well as ResNet18, ResNet34, and ResNet50 as backbones for the auxiliary model, which is shown in Table S4 in Appendix F. It can be observed that only when using CNNs, both accuracy and calibration ability were significantly poor. When using ResNet18, ResNet34, and ResNet50, similar performance were observed. This indicates that in order to ensure good performance of the Diff-DML, a backbone for the auxiliary model should not be selected if its performance is significantly worse than that of the primary model. Furthermore, when performance is similar, specific model architectures will yield similar performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 514, + 148, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 148, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 148, + 526 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 531, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 296, + 639 + ], + "type": "text", + "content": "In Table 3, we conducted an ablation study on the performance of the two strategies, DTS and DPLO, used in Diff-DML, using ResNet50 on the CIFAR-100 dataset. In the ablation study of DPLO, we set the learning rates and schedulers of models " + }, + { + "bbox": [ + 55, + 531, + 296, + 639 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 531, + 296, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 531, + 296, + 639 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 531, + 296, + 639 + ], + "type": "text", + "content": " to be the same as those in DML. As shown in the table, Diff-DML with only DTS or only DPLO both achieve better accuracy and calibration performance compared to the original DML, with the best results obtained when both strategies are combined." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 647, + 241, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 241, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 241, + 659 + ], + "type": "text", + "content": "4.5. Misclassification & OOD detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "A real-world application of calibrated predictions is making the model aware of its uncertainties. Specifically, the model should show low confidence for misclassified or out-of-distribution samples. Motivated by this, we conducted" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 328, + 221, + 542, + 305 + ], + "blocks": [ + { + "bbox": [ + 328, + 221, + 542, + 305 + ], + "lines": [ + { + "bbox": [ + 328, + 221, + 542, + 305 + ], + "spans": [ + { + "bbox": [ + 328, + 221, + 542, + 305 + ], + "type": "table", + "html": "
MethodAcc(%)↑ECE(%)↓cw-ECE(%)↓
CE77.5410.042.44
DML [29]77.894.421.69
DTS78.101.381.43
DPLO78.751.011.40
DTS+DPLO79.050.861.40
", + "image_path": "85dfdd6a7f9335fe4226dc4c2fb3c94ea307babd14473a2eefea60f66d53cef8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 336, + 309, + 531, + 319 + ], + "lines": [ + { + "bbox": [ + 336, + 309, + 531, + 319 + ], + "spans": [ + { + "bbox": [ + 336, + 309, + 531, + 319 + ], + "type": "text", + "content": "Table 3. Ablation study of Diff-DML on CIFAR-100." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 324, + 555, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 371 + ], + "type": "text", + "content": "experiments on misclassification and OOD detection, measuring the success rate of detecting such samples using maximum softmax probability [11]. We report threshold-free metrics (FPR-95%, D-error, AUROC, and AUPR)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 372, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 479 + ], + "type": "text", + "content": "Misclassification detection: Misclassification detection is an important problem in machine learning, as it allows for the identification of instances where the model's predictions are unreliable. We expect the model to output low probability or high uncertainty for samples that cannot be classified correctly. Table 4 presents the performance results for various models in detecting misclassifications. Diff-DML showed significant improvements over other single-model calibration techniques and the DE method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "text", + "content": "OD detection: A reliable classification model should show higher uncertainty and lower confidence when encountering test samples that differ significantly from the training data. We evaluated the ability of different calibration methods to distinguish OOD samples by mixing in-distribution test data with OOD data. We tested models trained on CIFAR-10 and CIFAR-100 for two types of OOD detection [5, 12]: far OOD (distinguishing SVHN [24] from CIFAR-10/CIFAR-100) and near OOD (distinguishing between CIFAR-10 and CIFAR-100). The results in Table 5 show that Diff-DML significantly outperformed other calibration methods, even exceeding the DE method, which is known for its OOD detection effectiveness." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 645, + 464, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 464, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 464, + 658 + ], + "type": "text", + "content": "5. Conclusion and discussion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Although Diff-DML significantly enhances both accuracy and calibration, its reliance on an auxiliary model with performance comparable to the primary model presents challenges, especially when training models with large parameter" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25818" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 70, + 460, + 297 + ], + "blocks": [ + { + "bbox": [ + 149, + 70, + 460, + 297 + ], + "lines": [ + { + "bbox": [ + 149, + 70, + 460, + 297 + ], + "spans": [ + { + "bbox": [ + 149, + 70, + 460, + 297 + ], + "type": "table", + "html": "
DatasetMethodFPR-95%↓D-error↓AUROC↑AUPR↑
CIFAR-10CE38.5313.2091.8099.46
DCA [19]38.9713.2891.9499.47
MMCE [16]23.0012.6193.5999.54
FL [21]33.1413.5692.2399.46
FL+DTDP [25]28.3414.2692.2099.36
FL+MDCA [10]32.8013.2192.4199.47
DE-3 [5, 17]22.3710.8994.0899.68
DML [29]19.2211.4994.0599.64
Diff-DML(ours)16.6310.0094.8799.73
CIFAR-100CE45.9520.1986.9395.47
DCA [19]40.6620.0986.8595.01
MMCE [16]45.9720.0287.0095.05
FL [21]45.0220.2187.0495.42
FL+DTDP [25]43.9520.7386.7995.16
FL+MDCA [10]45.0820.6886.5695.36
DE-3 [5, 17]40.0419.2588.1996.44
DML [29]40.3119.4287.8796.17
Diff-DML(ours)39.6218.9588.3896.50
", + "image_path": "fcf9b0eaeb6ac7a5f7da5f60fab30ca43d0cac8d044184270a7d7a13ac14b06c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 116, + 326, + 495, + 552 + ], + "blocks": [ + { + "bbox": [ + 149, + 304, + 460, + 315 + ], + "lines": [ + { + "bbox": [ + 149, + 304, + 460, + 315 + ], + "spans": [ + { + "bbox": [ + 149, + 304, + 460, + 315 + ], + "type": "text", + "content": "Table 4. The misclassification detection performance of different calibration methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 116, + 326, + 495, + 552 + ], + "lines": [ + { + "bbox": [ + 116, + 326, + 495, + 552 + ], + "spans": [ + { + "bbox": [ + 116, + 326, + 495, + 552 + ], + "type": "table", + "html": "
DatasetMethodFPR-95%↓D-error↓AUROC↑AUPR↑
In-Distribution: CIFAR-10CE69.48/19.8019.45/10.5986.11/94.6082.00/72.54
DCA [19]64.63/18.3818.80/10.1087.07/95.0584.36/72.48
MMCE [16]40.40/11.9317.33/8.3789.13/97.5389.61/88.85
FL [21]66.13/19.2020.97/10.6985.57/94.2882.81/74.18
FL+DTDP [25]50.54/23.6220.38/13.5486.57/92.0386.73/77.10
Out-of-Distribution: CIFAR-100/SVHNFL+MDCA [10]72.19/29.8420.75/13.5884.77/92.6381.01/67.11
DE-3 [5, 17]33.89/13.3115.60/9.0790.12/95.0690.27/86.53
DML [29]32.42/13.8515.90/9.3391.05/95.3891.67/86.04
Diff-DML(ours)30.12/10.5115.06/7.6591.87/96.3392.67/90.29
In-Distribution: CIFAR-100CE68.26/45.8728.69/22.2777.30/85.4477.49/54.01
DCA [19]69.41/47.8830.58/23.6275.96/83.2077.19/55.41
MMCE [16]65.34/49.3829.53/25.7776.58/77.6478.66/49.53
FL [21]67.27/48.4028.10/23.1778.36/83.8978.87/53.39
FL+DTDP [25]59.10/53.8527.23/24.1679.07/83.1081.05/49.61
Out-of-Distribution: CIFAR-10/SVHNFL+MDCA [10]69.52/41.6627.93/19.8778.35/87.4677.92/59.08
DE-3 [5, 17]56.21/48.7026.61/24.2080.36/83.4282.34/53.17
DML [29]62.08/37.9527.07/19.6879.90/87.0681.07/65.32
Diff-DML(ours)56.21/37.8425.55/19.5881.50/87.3483.57/65.65
", + "image_path": "578bf63902fc1be55bc7680941b7c937223c6543855236685898bd96da492d85.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 559, + 555, + 581 + ], + "lines": [ + { + "bbox": [ + 55, + 559, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 555, + 581 + ], + "type": "text", + "content": "Table 5. The OOD detection performance of different calibration methods. The numbers before and after the backslash denote the near and far OOD detection results, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 602, + 297, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 297, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 297, + 639 + ], + "type": "text", + "content": "sizes. This dependency can complicate scalability and computational efficiency, particularly in resource-constrained environments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 297, + 715 + ], + "type": "text", + "content": "In conclusion, we introduce Diff-DML, which overcomes the limitations of traditional DML by incorporating two strategies: DTS and DPLO. While DML promotes learning from an ensemble distribution, it often reduces prediction diversity, which is crucial for effective calibration. Diff-DML preserves this diversity while simultaneously" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 602, + 556, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 556, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 556, + 660 + ], + "type": "text", + "content": "improving both accuracy and calibration. Extensive experiments on datasets such as CIFAR-100 demonstrate that our approach outperforms state-of-the-art methods, offering a practical and efficient solution for reliable uncertainty estimation, particularly in safety-critical applications." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25819" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 111 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 111 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 111 + ], + "type": "text", + "content": "[1] Gavin Brown. Diversity in neural network ensembles. PhD thesis, CiteSeer, 2004. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 296, + 156 + ], + "type": "text", + "content": "[2] Peng Cui, Dan Zhang, Zhijie Deng, Yinpeng Dong, and Jun Zhu. Learning sample difficulty from pre-trained models for reliable prediction. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 158, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 158, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 296, + 201 + ], + "type": "text", + "content": "[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 203, + 296, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 296, + 268 + ], + "type": "text", + "content": "[4] James M Dolezal, Andrew Srisuwananukorn, Dmitry Karpeyev, Siddhi Ramesh, Sara Kochanny, Brittany Cody, Aaron S Mansfield, Sagar Rakshit, Radhika Bansal, Melanie C Bois, et al. Uncertainty-informed deep learning models enable high-confidence predictions for digital histopathology. Nature communications, 13(1):6572, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 300 + ], + "type": "text", + "content": "[5] Stanislav Fort, Huiyi Hu, and Balaji Lakshminarayanan. Deep ensembles: A loss landscape perspective. arXiv preprint arXiv:1912.02757, 2019. 2, 5, 6, 7, 8, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 296, + 334 + ], + "type": "text", + "content": "[6] Kevin P Greenman, Ava P Amini, and Kevin K Yang. Benchmarking uncertainty quantification for protein engineering. bioRxiv, pages 2023-04, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "type": "text", + "content": "[7] Arthur Gretton. Introduction to rkhs, and some simple kernel algorithms. Adv. Top. Mach. Learn. Lecture Conducted from University College London, 16(5-3):2, 2013. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 369, + 296, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 411 + ], + "type": "text", + "content": "[8] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 414, + 296, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 296, + 457 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 459, + 296, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 459, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 57, + 459, + 296, + 523 + ], + "type": "text", + "content": "[10] Ramya Hebbalaguppe, Jatin Prakash, Neelabh Madan, and Chetan Arora. A stitch in time saves nine: A train-time regularizing loss for improved neural network calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16081-16090, 2022. 2, 3, 5, 6, 8, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "type": "text", + "content": "[11] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2016. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "type": "text", + "content": "[12] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joe Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 614, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 635 + ], + "type": "text", + "content": "[13] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 679 + ], + "type": "text", + "content": "[14] Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. In Artificial Intelligence and Statistics, pages 623-631. PMLR, 2017. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[15] Meelis Kull, Miquel Perello Nieto, Markus Kangsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "with dirichlet calibration. Advances in neural information processing systems, 32, 2019. 2, 3, 6, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "text", + "content": "[16] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In International Conference on Machine Learning, pages 2805-2814. PMLR, 2018. 2, 3, 5, 6, 8, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 555, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 185 + ], + "type": "text", + "content": "[17] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 2, 5, 6, 8, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 186, + 555, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 230 + ], + "type": "text", + "content": "[18] Christian Leibig, Vaneeda Allken, Murat Seçkin Ayhan, Philipp Berens, and Siegfried Wahl. Leveraging uncertainty information from deep neural networks for disease detection. Scientific reports, 7(1):1-14, 2017. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 232, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 555, + 274 + ], + "type": "text", + "content": "[19] Gongbo Liang, Yu Zhang, Xiaoqin Wang, and Nathan Jacobs. Improved trainable calibration method for neural networks on medical imaging classification. arXiv preprint arXiv:2009.04057, 2020.3,5,6,8,1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 277, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 554, + 330 + ], + "type": "text", + "content": "[20] Rhiannon Michelmore, Matthew Wicker, Luca Laurenti, Luca Cardelli, Yarin Gal, and Marta Kwiatkowska. Uncertainty quantification with statistical guarantees in end-to-end autonomous driving control. In International Conference on Robotics and Automation, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 332, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 555, + 386 + ], + "type": "text", + "content": "[21] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. Advances in Neural Information Processing Systems, 33:15288-15299, 2020. 3, 5, 6, 8, 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 388, + 555, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 555, + 421 + ], + "type": "text", + "content": "[22] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 2, 3, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 422, + 554, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 554, + 466 + ], + "type": "text", + "content": "[23] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Proceedings of the AAAI conference on artificial intelligence, 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 468, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 555, + 521 + ], + "type": "text", + "content": "[24] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning, page 7. Granada, Spain, 2011. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 523, + 555, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 555, + 588 + ], + "type": "text", + "content": "[25] Rishabh Patra, Ramya Hebbalaguppe, Tirtharaj Dash, Gautam Shroff, and Lovekesh Vig. Calibrating deep neural networks using explicit regularisation and dynamic data pruning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1541-1549, 2023. 3, 5, 6, 8, 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 590, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 554, + 633 + ], + "type": "text", + "content": "[26] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey Hinton. Regularizing neural networks by penalizing confident output distributions. arXiv preprint arXiv:1701.06548, 2017. 2, 3, 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 635, + 555, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 555, + 677 + ], + "type": "text", + "content": "[27] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61-74, 1999. 2, 3, 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 680, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 680, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 680, + 555, + 713 + ], + "type": "text", + "content": "[28] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25820" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 163 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 75, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 75, + 73, + 296, + 95 + ], + "type": "text", + "content": "conference on computer vision and pattern recognition, pages 2818-2826, 2016. 2, 3, 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[29] Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4320-4328, 2018. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 163 + ], + "type": "text", + "content": "[30] Zhi-Hua Zhou. Ensemble methods: foundations and algorithms. CRC press, 2012. 4" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25821" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/c0d9e11b-700c-40d7-8dca-46321613cb22_content_list.json b/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/c0d9e11b-700c-40d7-8dca-46321613cb22_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6468391fdab078f91c92e7708bb27af979c74cb3 --- /dev/null +++ b/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/c0d9e11b-700c-40d7-8dca-46321613cb22_content_list.json @@ -0,0 +1,1511 @@ +[ + { + "type": "text", + "text": "Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement", + "text_level": 1, + "bbox": [ + 181, + 130, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuchen Ren $^{1}$ , Zhengyu Zhao $^{1*}$ , Chenhao Lin $^{1}$ , Bo Yang $^{2}$ , Lu Zhou $^{3}$ , Zhe Liu $^{4}$ , Chao Shen $^{1}$ $^{1}$ Xi'an Jiaotong University, China; $^{2}$ Information Engineering University, China; $^{3}$ Nanjing University of Aeronautics and Astronautics, China; $^{4}$ Zhejiang Lab, China", + "bbox": [ + 142, + 202, + 854, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ryc98@stu.xjtu.edu.cn, {zhengyu.zhao, linchenhao}@xjtu.edu.cn, yangbo_hn@163.com, {lu.zhou, zhe.liu}@nuaa.edu.cn, chaoshen@mail.xjtu.edu.cn", + "bbox": [ + 161, + 258, + 830, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs) have been widely applied in various computer vision and vision-language tasks. To gain insights into their robustness in practical scenarios, transferable adversarial examples on ViTs have been extensively studied. A typical approach to improving adversarial transferability is by refining the surrogate model. However, existing work on ViTs has restricted their surrogate refinement to backward propagation. In this work, we instead focus on Forward Propagation Refinement (FPR) and specifically refine two key modules of ViTs: attention maps and token embeddings. For attention maps, we propose Attention Map Diversification (AMD), which diversifies certain attention maps and also implicitly imposes beneficial gradient vanishing during backward propagation. For token embeddings, we propose Momentum Token Embedding (MTE), which accumulates historical token embeddings to stabilize the forward updates in both the Attention and MLP blocks. We conduct extensive experiments with adversarial examples transferred from ViTs to various CNNs and ViTs, demonstrating that our FPR outperforms the current best (backward) surrogate refinement by up to $7.0\\%$ on average. We also validate its superiority against popular defenses and its compatibility with other transfer methods. Codes and appendix are available at https://github.com/RYC-98/FPR.", + "bbox": [ + 88, + 359, + 485, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 770, + 223, + 786 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) [8, 33, 41] demonstrate remarkable performance across diverse domains, with Vision Transformers (ViTs) [5, 9, 28] excelling particularly in computer visualization. However, ViTs are also shown to be vulnerable to adversarial examples [20, 21], which are", + "bbox": [ + 89, + 796, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4838ca372ca55753d60cd7d0f8a80a47e97bf880a971991fb1cc392827d28f16.jpg", + "image_caption": [ + "Figure 1. Our Forward Propagation Refinement (FPR) vs. previous methods which only refine the backward propagation." + ], + "image_footnote": [], + "bbox": [ + 522, + 325, + 897, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "optimized by adding subtle perturbations [23, 46] into clean images [6, 44, 47].", + "bbox": [ + 511, + 609, + 906, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A key property of adversarial examples that causes practical security concerns is their transferability, i.e., they can be optimized on a (seen) surrogate model and then directly transfer their effects to (unseen) target models [4, 40]. Compared to other transfer approaches that focus on loss design, optimization algorithm or data augmentation [22, 24, 34], surrogate refinement [15, 35, 39] explores the intrinsic relationship between the model itself and adversarial transferability, offering a unique perspective.", + "bbox": [ + 511, + 641, + 908, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, existing surrogate refinements on ViTs [38, 45] are limited to refining the backward propagation, i.e., only modifying the gradient information. To address the above limitation, in this work, we explore the potential of Forward Propagation Refinement (FPR), i.e., modifying features during forward propagation. Our FPR not only provides a new perspective of surrogate refinement but also potentially benefits the refinement during the sub", + "bbox": [ + 511, + 779, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25071", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sequent backward propagation (see our discussion in Section 3.2). Specifically, FPR comprises two parts that refine two key components of ViTs respectively: attention maps and output token embeddings. Figure 1 highlights the difference between our forward-based method and the previous backward-based methods.", + "bbox": [ + 89, + 90, + 480, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For attention maps, we identify that certain attention maps in ViTs are redundant and can deteriorate transferability. Rather than simply dropping redundant attention maps, which recklessly removes all information, we propose Attention Map Diversification (AMD), which moderately diversifies the attention maps based on random weighting. The introduced randomness on attention maps acts as a simulation of unknown models, thereby mitigating the overfitting to the surrogate model. Although AMD only modifies the forward propagation, we find that it can potentially induce the vanishing of certain gradients flowing through the attention maps during backward propagation, further alleviating overfitting [38].", + "bbox": [ + 89, + 183, + 480, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For output token embeddings, we posit that adversarial examples during generation are highly likely to be near the local optima, resulting in the surrogate model's embeddings containing erroneous information. Consequently, inspired by the momentum distillation technology when training the vision-language pretraining model [13], we propose Momentum Token Embedding (MTE), to stabilize the updates of token embeddings in each iteration by accumulating historical token embedding information.", + "bbox": [ + 89, + 380, + 480, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In sum, we make the following main contributions:", + "bbox": [ + 109, + 517, + 447, + 531 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We explore the refinement of ViT surrogate models from the perspective of forward propagation, particularly modifying the attention maps and output token embeddings.", + "- We propose Attention Map Diversification (AMD) to moderately avoid attention redundancy, which also induces beneficial gradient vanishing during the subsequent backward propagation, altogether mitigating the overfitting to the surrogate model.", + "- We propose Momentum Token Embedding (MTE) to stabilize the updates of token embeddings by accumulating historical information, during the generation of adversarial examples on ViTs.", + "- Extensive experiments validate that our Forward Propagation Refinement (FPR), comprising AMD and MTE, outperforms the current best (backward) refinement method by up to an average of $7.0\\%$ against 13 ViTs and CNNs with or without defenses." + ], + "bbox": [ + 89, + 532, + 482, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 806, + 232, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Surrogate Refinement Attacks", + "text_level": 1, + "bbox": [ + 89, + 832, + 359, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this subsection, we specifically delineate attack methods falling within the surrogate refinement, since it is the main topic of this work. Skip Gradient Method (SGM) [39] attenuates the attack by using a gradient descent approach.", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "uates gradients traversing the residual modules to mitigate overfitting to the surrogate model. Ghost Networks [15] employs dropout in each block of the surrogate model and perturbs the skip connections to create a longitudinal ensemble. Backward Propagation Attack (BPA) [35] demonstrates that gradients traversing ReLU and max-pooling layers can carry inaccurate information. To address this, it adopts both non-monotonic function and softmax with temperature to recover the true gradients, thereby enhancing transferability. The aforementioned methods focus on the backward propagation of CNNs, while a few approaches also modify the output features of convolutional layers during forward propagation. Diversifying the High-level Features (DHF) [37] modifies the high-level features through random transformations. Clean Feature Mixup (CFM) [2] introduces competition into the generation of adversarial examples at the feature level for stronger transferability. Several methods also retrain the surrogate model to strengthen the transferability [14, 50], however, this may significantly increase the attack cost.", + "bbox": [ + 511, + 90, + 903, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The above surrogate refinements are primarily designed for CNNs; several methods have also been developed for ViTs. Pay No Attention and Patch Out (PNAPO) [38] comprises three modifications to synergistically mitigate overfitting to the surrogate model. First, it cuts off gradients traversing all attention maps. Second, it randomly updates a specific number of input patches. Finally, it adds an extra term to the cross-entropy loss function that increases the adversarial noise norm to enhance transferability. Token Gradient Regularization (TGR) [45] posits that extreme values in gradients are detrimental. Therefore, it removes the extreme gradients of QKV embeddings, attention maps, and MLP blocks to reduce gradient variance, thereby alleviating overfitting to the surrogate models. Gradient Normalization Scaling (GNS) [51] improves TGR by accurately normalizing and scaling the identified mild gradients, which yields stronger transferability. Even though the above methods exhibit the ability to craft transferable adversarial examples, they neglect the broader implications of forward propagation, which prevents further transferability. In this work, we focus on refining the ViT surrogate model from the new perspective of forward propagation and specifically explore two key components: attention maps and output token embeddings.", + "bbox": [ + 511, + 393, + 903, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Other Transfer-based Attacks", + "text_level": 1, + "bbox": [ + 511, + 771, + 779, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This subsection reviews transfer-based attacks [48] that do not need to modify the surrogate models. Fast Gradient Sign Method (FGSM) [6] creates adversarial examples in a single step by following the gradient sign's direction. Basic Iterative Method (BIM) [12] extends FGSM by applying multiple iterations. Momentum Iterative Method (MIM) [4] introduces a momentum term for gradients at each qua", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25072", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0dfcd105905057697b20191c687d7809b65be43815a26601f3c9c632bc12782a.jpg", + "image_caption": [ + "Figure 2. The overview of Forward Propagation Refinement (FPR). Note that Attention Map Diversification (AMD) is applied to specific attention maps according to the index set, while Momentum Token Embedding (MTE) is applied to all Attention and MLP blocks." + ], + "image_footnote": [ + "$E_{t}^{A} / E_{t}^{M}$ : Output Token Embedding of Attention / MLP Block at $t$ -th Iteration $\\widehat{E_{t}^{A}} / \\widehat{E_{t}^{M}}$ : Accumulated Token Embedding of Attention / MLP Block at $t$ -th Iteration $A$ : Attention Map $A'$ : Diversified Attention Map $M$ : Diversity Matrix $Q/K/V$ : Query/Key/Value Component" + ], + "bbox": [ + 130, + 88, + 867, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion for better convergence. Nesterov Fast Gradient Sign Method (NIM) [17] takes an extra step at the beginning of each iteration and adopts the example at the farther location to compute the gradients for the update. Variance Tuning Method (VTM) [34] modifies current gradients with variance at the last iteration. Gradient Relevance Attack (GRA) [49] utilizes correlation information extracted from the nearby samples to stabilize the current update direction. Some input transformations have also been presented to enhance transferability. Diversity Input Method (DIM) [40] combines padding and resizing transformations on the input to reduce overfitting. Scale Invariance Method (SIM) [17] employs a group of scaled copies to enhance the input. Structure Invariant Attack (SIA) [36] employs different transformations from a library to each block of the input to prompt the transferability. Block Shuffle and Rotation (BSR) [32] divides the input image into several blocks, which are then randomly shuffled and rotated to create a set of new images for gradient computation.", + "bbox": [ + 91, + 446, + 483, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Vision Transformers", + "text_level": 1, + "bbox": [ + 89, + 742, + 285, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer [31] is widely used in natural language processing and has since gained increasing recognition in computer vision. ViT [5] processes images by dividing them into patches and embedding these patches as tokens. They employ a Transformer architecture with multi-head self-attention mechanisms to effectively capture spatial relationships. DeiT [28] improves ViT through teacher-student distillation, enabling superior performance with less data. Swin [18] employs a hierarchical structure with shifted win", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dows to efficiently handle multi-scale attention. PiT [9] utilizes pooling layers to streamline computational processes and manage data flow. CaiT [29] incorporates LayerScale to stabilize training as model depth increases and employs Class-Attention layers to enhance the extraction of classification information. Visformer [3] combines convolutional layers with ViT to effectively integrate spatial features and attention mechanisms. CoaT [43] merges convolution and self-attention in a hierarchical architecture to balance feature extraction and feature interaction.", + "bbox": [ + 511, + 446, + 906, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 612, + 648, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we propose the Forward Propagation Refinement (FPR) to enhance the transferability of adversarial examples from ViTs, consisting of Attention Map Diversification (AMD) and Momentum Token Embedding (MTE). Figure 2 provides an overview of FPR.", + "bbox": [ + 511, + 637, + 905, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 723, + 642, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assume an image $x$ with channel $C$ , height $H$ , width $W$ , the true label $y$ , and a surrogate model $F$ , the goal of a transfer-based attack is to add adversarial perturbation $\\delta$ to the input $x$ , creating an adversarial example $x^{a} = x + \\delta$ , which is then fed into the target model $F^{t}$ to induce incorrect predictions. This process can be formulated as solving the following problem:", + "bbox": [ + 511, + 744, + 906, + 852 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\delta^ {\\prime} = \\underset {\\delta} {\\arg \\max } L (F (x + \\delta), y), s. t. \\| \\delta \\| _ {\\infty} \\leq \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 878, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25073", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bb504ed4017924f1ca447415eed07cc7cfc5dddbbcf0b7aae96ebd2c41a52ca5.jpg", + "image_caption": [ + "Figure 3. Illustration of redundancy in attention maps: Randomly dropping a small number of blocks leads to improved transferability. The results are averaged over 5 repeated runs on 13 target models, and the dash lines denote the baseline results of the MIM [4] without dropping." + ], + "image_footnote": [], + "bbox": [ + 132, + 88, + 444, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L$ is the loss function and $\\epsilon$ is the maximum perturbation bound.", + "bbox": [ + 89, + 414, + 482, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before the image $x$ is fed into ViTs, it is segmented into $N = \\frac{H \\cdot W}{P^2}$ patches, with each patch having a height, width, and channel dimension of $P$ , $P$ , and $C$ , respectively. Then, these patches are converted into token embeddings, and fed into a sequence of Transformer blocks. Each Transformer block includes an Attention block and a MLP block.", + "bbox": [ + 89, + 443, + 483, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The input token embedding of the Attention block is transformed into Query $(Q)$ , Key $(K)$ , and Value $(V)$ components. Here, the shapes of $Q$ , $K$ and $V$ are $N\\times D$ , where $D$ is the embedding dimension. The Attention block employs the self-attention mechanism to establish the relationship between $Q$ , $K$ , and $V$ . The attention map $A$ in the Attention block is computed by:", + "bbox": [ + 89, + 535, + 483, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {D}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 647, + 482, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the softmax function is employed at each row. The attention map $A$ is then used to weight the $V$ , and the output $Z$ is calculated as $Z = AV$ .", + "bbox": [ + 89, + 689, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The MLP block follows the Attention block and typically includes the fully connected layer, GELU activation function, dropout layer, and identity normalization layer to aggregate information for all token embeddings.", + "bbox": [ + 89, + 734, + 483, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Attention Map Diversification (AMD)", + "text_level": 1, + "bbox": [ + 89, + 801, + 415, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The previous work [1] has demonstrated the redundancy in attention blocks of ViTs, showing that dropping a few such blocks does not significantly degrade classification accuracy. Therefore, it is reasonable to posit that such redundancy may also exist in generating adversarial examples,", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "causing potential overfitting to the surrogate model. To verify this, we conduct exploratory experiments by dropping certain attention blocks. As can be seen from Figure 3, dropping a small number of blocks indeed leads to improved transferability (while dropping too many may harm transferability).", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To avoid such redundancy, a straightforward idea is to apply dropout of attention blocks while generating adversarial examples. However, we find that the dropout is too aggressive and may cause serious information loss, leading to only a small improvement of transferability (see comparing results in Table 5 of Section 4.3). Therefore, we propose a more moderate method called Attention Map Diversification (AMD), which transforms the original attention map $A$ by multiplying it with a diversity matrix $M$ . Specifically, the diversified attention map $A'$ is obtained by:", + "bbox": [ + 511, + 181, + 906, + 333 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA ^ {\\prime} = A M D (A) = \\operatorname {s o f t m a x} (A \\odot M), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 347, + 903, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ is the element-wise product and softmax function is applied to each row. The diversity matrix $M$ has the same shape as $A$ , and every element $M_{i,j}$ $(1 \\leq i,j \\leq N)$ conforms to the uniform distribution $U(1 - d, 1 + d)$ , where $d$ is a factor that controls the diversity level. Note that AMD is applied to attention maps at specified Transformer blocks according to a index set $I$ .", + "bbox": [ + 511, + 368, + 905, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The benefit of AMD on backward propagation. In principle, modifying the forward propagation would impact the subsequent backward propagation. Here, we theoretically demonstrate that the impact of our AMD on backward propagation can further improve transferability. Specifically, we find that AMD implicitly induces gradient vanishing at attention map positions during backward propagation, which aligns with a previous surrogate refinement, PNAPO [38], that directly truncates the gradients of attention maps during backpropagation. The underlying principle of PNAPO suggests that gradients passing through attention maps carry overfitting information from surrogate models, and attenuating these gradients can enhance the transferability of adversarial examples. Our findings can be explained as follows.", + "bbox": [ + 511, + 474, + 905, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When incorporating AMD, we let $S = A \\odot M$ and the $i$ -th row vector of diversified attention map $A_{i}^{\\prime}$ as:", + "bbox": [ + 511, + 700, + 905, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} A _ {i} ^ {\\prime} = \\left(A _ {i, 1} ^ {\\prime}, A _ {i, 2} ^ {\\prime}, \\dots , A _ {i, N} ^ {\\prime}\\right) \\\\ = \\operatorname {s o f t m a x} \\left(A _ {i, 1} \\cdot M _ {i, 1}, A _ {i, 2} \\cdot M _ {i, 2},..., A _ {i, N} \\cdot M _ {i, N}\\right) \\\\ = \\operatorname {s o f t m a x} \\left(S _ {i, 1}, S _ {i, 2}, \\dots , S _ {i, N}\\right) \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 739, + 908, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the derivative $\\frac{\\partial A_{i,m}^{\\prime}}{\\partial S_{i,n}}$ ( $1 \\leq m, n \\leq N$ ) is expressed by:", + "bbox": [ + 511, + 799, + 901, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial A _ {i , m} ^ {\\prime}}{\\partial S _ {i , n}} = \\left\\{ \\begin{array}{l l} A _ {i, m} ^ {\\prime} \\cdot \\left(1 - A _ {i, m} ^ {\\prime}\\right), & \\text {i f} m = n \\\\ - A _ {i, m} ^ {\\prime} \\cdot A _ {i, n} ^ {\\prime}, & \\text {i f} m \\neq n \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 828, + 903, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the diversity factor $d$ becomes large, elements in the diversity matrix $M$ will fluctuate significantly, causing ex", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25074", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "treme values in $S$ , and ultimately pushing the elements in $A'$ towards 0 or 1 due to the property of softmax function. From Eq. 5, whether the elements in $A'$ approach 0 or 1, the derivative will tend towards 0, resulting in gradient vanishing. According to the chain rule of differentiation, the gradients of the attention maps subjected to our AMD will experience varying degrees of gradient vanishing.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared with PNAPO, which directly discards all gradients flowing through the attention maps, the gradient vanishing in our AMD is implicit without directly altering any aspect of the backward propagation. Furthermore, the gradient vanishing in our AMD method is more fine-grained, occurring at the row level of the attention map. Each row experiences varying degrees of vanishing, introducing additional randomness that helps to further mitigate overfitting to the surrogate model.", + "bbox": [ + 89, + 196, + 480, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Momentum Token Embedding (MTE)", + "text_level": 1, + "bbox": [ + 89, + 340, + 419, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is challenging to ensure convergence to the global optimum in the high-dimension space when generating adversarial examples. Typically, adversarial examples converge to a local optimum along an inaccurate path [4]. This causes the internal token embeddings of the surrogate model to contain inaccurate information during iterations, thereby weakening the transferability. Therefore, we borrow the idea of Momentum Distillation (MoD) [13] from the Vision-and-Language Pre-training (VLP) literature, where a momentum version of the model through the moving average of the primary model's parameters is maintained to mitigate the noise in the web-sourced training dataset. Similarly, we maintain the momentum version of token embedding updates to accumulate historical token embeddings while generating adversarial examples.", + "bbox": [ + 89, + 363, + 482, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We term this method Momentum Token Embedding (MTE), and formulate it as follows:", + "bbox": [ + 89, + 589, + 482, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {E _ {t}} = M T E \\left(E _ {t}\\right) = \\eta \\cdot \\widehat {E _ {t - 1}} + s \\cdot E _ {t}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 628, + 482, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t$ is the iteration index, $E_{t}$ is the output token embedding of any Attention block or MLP block at the $t$ -th iteration, $\\widehat{E_t}$ is the corresponding accumulated token embedding at the $t$ -th iteration, $\\eta$ is the attenuation factor, and $s$ is the scale factor. MTE is applied to all Attention and MLP blocks. Note that, different from MoD, our MTE does not require model training and can be seen as constructing a residual structure for output token embeddings across the time dimension. Previous work also finds the usefulness of the residual structure in improving transferability, but during backward propagation within each iteration [39].", + "bbox": [ + 89, + 657, + 482, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Complementary AMD and MTE", + "text_level": 1, + "bbox": [ + 89, + 833, + 377, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The previous subsections examine AMD and MTE from their respective perspectives. Here, we further discuss why they can collaboratively enhance transferability.", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Forward Propagation Refinement (FPR)", + "text_level": 1, + "bbox": [ + 514, + 90, + 867, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: A surrogate model $F$ with the cross-entropy loss $L$ , a clean image $x$ and its true label $y$ . The maximum perturbation bound $\\epsilon$ , the iteration number $T$ , the decay factor $\\mu$ , step size $\\alpha$ , index set $I$ , diversity factor $d$ , scale factor $s$ and attenuation factor $\\eta$ .", + "bbox": [ + 511, + 108, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: The adversarial image $x^{adv} = x_{T}$ .", + "bbox": [ + 513, + 181, + 803, + 198 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize $g_0 = 0$ , $E_0 = 0$ , $x_0 = x$ .", + "2: for $t = 1 \\rightarrow T$ do" + ], + "bbox": [ + 522, + 200, + 772, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the forward propagation:", + "text_level": 1, + "bbox": [ + 540, + 231, + 776, + 244 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "3: Apply AMD (Eq. (3)) to attention maps in specific Transformer blocks according to the index set $I$ .", + "4: Apply MTE (Eq. (6)) to output token embeddings in all Transformer (Attention and MLP) blocks." + ], + "bbox": [ + 522, + 246, + 903, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the backward propagation:", + "text_level": 1, + "bbox": [ + 540, + 306, + 789, + 321 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "5: Calculate the gradient: $\\nabla_{x_{t - 1}}L(F(x_{t - 1}),y)$", + "6: Update momentum gradient:" + ], + "bbox": [ + 522, + 321, + 859, + 351 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {t} = \\mu \\cdot g _ {t - 1} + \\frac {\\nabla_ {x _ {t - 1}} L (F (x _ {t - 1}) , y)}{\\left\\| \\nabla_ {x _ {t - 1}} L (F (x _ {t - 1}) , y) \\right\\| _ {1}}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 362, + 861, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7: Update the adversarial example:", + "bbox": [ + 522, + 409, + 779, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\operatorname {C l i p} _ {\\epsilon} \\left\\{x _ {t - 1} + \\alpha \\cdot \\operatorname {s i g n} \\left(g _ {t}\\right) \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 436, + 836, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "8: end for", + "text_level": 1, + "bbox": [ + 522, + 464, + 596, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, AMD and MTE refine the forward propagation of ViTs from two complementary angles. AMD falls under the category of \"external refinement\", which enhances surrogate models by introducing extraneous information or perturbing existing embeddings. Techniques such as attention map dropout also fall into this category. Our empirical results demonstrate that this type of refinement is relatively aggressive, making it unsuitable for concurrent use or for application across all blocks. In contrast, MTE belongs to the \"internal refinement\" category, mainly leveraging information from the input data itself. Since these refinements rely on self-related information, they are typically more moderate and can be applied to all blocks. Both AMD and MTE help promoting transferability but in different ways, this complementary effect further enhances the transferability of adversarial examples.", + "bbox": [ + 511, + 527, + 906, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we provide the pseudocode of FPR (based on MIM [4]) in Algorithm 1. Specifically, at each of $T$ iterations, we first apply AMD to specific attention maps according to the index set $I$ and MTE to the outputs of all Attention and MLP blocks during the forward propagation using PyTorch's Hook function. Then, during the backward propagation, we compute the gradient of the current adversarial example and use MIM [4] to update it.", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25075", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/dacf2cc0d76a2e0dc8b20d05eef1c1727e095e2a24f2c3b3c157c1aa2ee69050.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. The optimal parameter settings of our FPR about its index set $(I)$ , diversity factor $(d)$ , scale factor $(s)$ , and attenuation factor $(\\eta)$ . Detailed ablation studies in Section 4.3 provide the sensitivity analysis of the attack performance towards the parameter selection under four cases." + ], + "table_body": "
ModelIdsη
ViT-B0,1,4,9,11250.80.3
CaiT-S2,14,25300.60.2
PiT-T1,6,11250.70.2
DeiT-B0,1,5,10,11300.70.2
", + "bbox": [ + 125, + 88, + 446, + 180 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 287, + 220, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 313, + 295, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and baselines. Following the common practice [34, 45, 49], we use 1000 images randomly selected from the ILSVRC2012 [25] validation set. We compare our FPR with other surrogate refinements that focus on backward propagation: PNAPO [38], TGR [45], and GNS [51]. All these methods are built on MIM [4].", + "bbox": [ + 89, + 335, + 482, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models and defenses. For surrogate models, we consider four popular ViTs: ViT-Base (ViT-B) [5], CaiT-Small (CaiT-S) [29], PiT-Tiny (PiT-T) [9] and Deit-Base (Deit-B) [28]. For target models, we consider seven ViTs and six CNNs: ViT-B [5], CaiT-S [29], PiT-Base (PiT-B) [9], Visformer-Small (Vis-S) [3], Swin-Tiny (Swin-T) [18], DeiT-Tiny (DeiT-T) [28], and CoaT-Tiny (CoaT-T) [43], ResNet-18 (RN-18) [8], VGG-16 [26], DenseNet-121 (DN-121) [11], EfficientNet-b0 (EN-b0) [27], MobileNetv3 (MN-v3) [10], and ResNeXt-50 (RNX-50) [41]. Five popular defenses are tested: adversarial training (AT) paired with RN-50 [30], high-level representation guided denoiser (HGD) [16], neural representation purifier (NRP) [19], Bit depth reduction (BDR) [42], and JPEG compression [7]. Finally, two online models, i.e., Baidu Cloud API and Aliyun API, are also included.", + "bbox": [ + 89, + 428, + 482, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyperparameter settings. Commonly, we set the number of iterations as 10, the maximum perturbation bound $\\epsilon$ as 16, the step size $\\alpha$ as 1.6, and the decay factor of momentum gradient $\\mu = 1$ . Parameters of all baselines are the same as their default values. For our FPR, the optimal hyperparameters are listed in Table 1, with a detailed analysis in Section 4.3 discussing their sensitivity to different parameter choices. All experiments are conducted on a single RTX 4060 GPU with 8 GB of VRAM.", + "bbox": [ + 89, + 670, + 482, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 816, + 290, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different target models. We first compare FPR with baselines on transferring adversarial examples from ViTs to different ViTs and CNNs. Table 2 shows that FPR consistently outperforms baselines in all cases, suggesting the potential", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of refining the surrogate model from the perspective of forward propagation. For example, when generating adversarial examples on ViT-B, FPR achieves an average attack success rate of $84.3\\%$ on six CNNs, surpassing GNS, TGR, and PNAPO by $7.0\\%$ , $7.4\\%$ , and $15.5\\%$ , respectively, and improves the baseline MIM by $33.4\\%$ .", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different defenses. We then evaluate different attack methods against five popular defenses. As shown in Table 3, FPR maintains its superiority over existing surrogate refinements in such challenging defense settings. For instance, when generating adversarial examples on ViT-B, FPR achieves an average attack success rate of $70.8\\%$ across five defenses, outperforming other surrogate refinements GNS, TGR, and PNAPO by $5.7\\%$ , $6.4\\%$ , and $12.4\\%$ , respectively, while surpassing the vanilla baseline, MIM, by $24.3\\%$ .", + "bbox": [ + 511, + 181, + 903, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Incorporating other transfer methods. A good surrogate refinement-based transfer method should also exhibit strong compatibility with other (types of) transfer-based attack methods. In this study, we integrate FPR into two advanced gradient correction-based transferable attacks, VTM [34] and GRA [49], as well as two latest input augmentation-based transferable attacks, BSR [32] and SIA [36]. Specifically, we equip these four attacks with 5 augmented copies per iteration, while keeping other parameter settings as their default values. Results in Table 4 indicate that FPR can effectively complement existing transfer-based attacks for stronger transferability. For instance, when combined with our FPR, the attack success rate of VTM increases from $57.3\\%$ to $81.7\\%$ on CNNs.", + "bbox": [ + 511, + 319, + 903, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Online models. We also evaluate our FPR on two online models, i.e., Baidu Cloud API and Aliyun API. The results, presented in Appendix A, show that FPR still outperforms both TGR and GNS.", + "bbox": [ + 511, + 531, + 903, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 604, + 676, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Individual components. We compare each component of FPR with the plain baseline MIM, and the two best-performed surrogate refinements, TGR and GNS. We also consider Dropout as a comparison to our AMD. From Table 5, we can see that combining our AMD and MTE yields the best transferability results. In addition, Dropout is less effective than our AMD, indicating that directly removing certain attention maps is rather aggressive and fails to retain useful information.", + "bbox": [ + 511, + 627, + 903, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Parameter sensitivity. Figure 4 shows the attack success rates when varying three main parameters: diversity factor $d$ , scale factor $s$ , and attenuation factor $\\eta$ . $d$ controls the diversity of the attention map. FPR achieves satisfactory performance when $d$ is between 25 and 30. $s$ represents the importance of current token embeddings. FPR achieves satisfactory performance when $s$ is between 0.6 and 0.8. $\\eta$ signifies the importance of historical token embeddings. FPR achieves satisfactory performance when $\\eta$ is between 0.2", + "bbox": [ + 511, + 763, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25076", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/985915af2b6096ce747f8a4cb71b5e52a41eb746fad9631b5e304050bcef18a4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAttackViT-BCaiT-SPiT-BVis-SSwin-TDeiT-TCoaT-TRN-18VGG-16DN-121EN-b0MN-v3RNX-50Average
ViT-BMIM [4]97.261.840.442.354.965.844.451.757.251.350.851.738.250.9
PNAPO [38]99.183.262.165.874.783.064.067.470.067.668.563.056.368.8
TGR [45]99.186.263.869.981.194.470.479.179.875.383.179.960.076.9
GNS [51]99.989.568.174.082.493.672.277.876.776.079.176.062.077.3
FPR99.292.573.078.388.598.277.787.387.682.090.387.468.384.3
CaiT-SMIM [4]68.598.850.054.968.275.354.861.067.859.861.055.747.160.3
PNAPO [38]69.888.258.161.070.373.860.164.869.764.365.561.753.264.4
TGR [45]90.1100.076.081.391.297.880.686.986.684.387.782.169.184.5
GNS [51]91.2100.078.682.291.697.681.186.687.184.989.681.970.885.3
FPR93.8100.081.685.695.098.886.887.488.788.690.682.976.888.1
PiT-TMIM [4]30.036.138.543.351.075.055.569.568.760.864.571.640.454.2
PNAPO [38]39.550.949.654.961.685.465.078.776.772.277.981.050.864.9
TGR [45]37.947.746.954.162.587.465.182.882.075.581.888.853.766.6
GNS [51]37.144.146.451.959.484.463.379.978.071.576.384.348.663.5
FPR44.356.157.162.469.591.372.288.786.480.487.691.160.372.9
DeiT-BMIM [4]86.082.467.069.579.787.765.369.070.570.171.461.458.272.2
PNAPO [38]83.487.772.474.182.286.071.974.974.774.976.168.066.376.4
TGR [45]92.497.181.284.392.697.983.586.085.285.089.984.574.687.2
GNS [51]92.198.080.584.291.098.079.883.380.783.586.781.370.885.4
FPR94.698.483.886.795.599.587.089.889.288.193.986.879.890.2
", + "bbox": [ + 94, + 88, + 903, + 333 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b094136fdd60a0564b2121fea4ccd9ac04a30d4aea93dd843d0d905c59ae5577.jpg", + "table_caption": [ + "Table 2. Attack success rates (%) of different surrogate refinements on various ViTs and CNNs. The \"Average\" column excludes the white-box results with grey background." + ], + "table_footnote": [], + "table_body": "
ModelMethodATHGDNRPJPEGBitAverage
ViT-BMIM [4]43.929.842.157.559.146.5
PNAPO [38]45.647.553.570.974.858.4
TGR [45]49.750.861.677.782.064.4
GNS [51]49.254.861.377.982.365.1
FPR50.761.469.284.588.470.8
CaiT-SMIM [4]44.337.446.264.166.951.8
PNAPO [38]44.843.748.265.267.353.8
TGR [45]49.860.766.481.286.869.0
GNS [51]49.862.767.482.688.070.1
FPR51.168.970.886.290.473.5
PiT-TMIM [4]46.132.042.057.562.047.9
PNAPO [38]47.542.948.565.569.854.8
TGR [45]50.844.050.568.271.957.1
GNS [45]48.938.649.365.368.854.2
FPR51.449.956.173.277.461.6
DeiT-BMIM [4]44.748.152.270.374.958.1
PNAPO [38]46.957.257.073.678.762.7
TGR [45]50.265.267.882.288.170.7
GNS [51]48.361.665.780.686.368.5
FPR50.270.771.387.091.574.1
", + "bbox": [ + 94, + 395, + 480, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and 0.3. Detailed results of applying the same index set $(I)$ to other methods and the sensitivity of the $I$ used in AMD are reported in Appendix B, where we also provide insights for identifying the optimal $I$ .", + "bbox": [ + 89, + 747, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Gradient vanishing comparison. We compare the extent of gradient vanishing in Table 6. We can conclude that gradient vanishing can improve transferability, but it is not a simple linear correlation. This aligns with our claim that the key of AMD is to diversify attention maps in the forward propagation, while it implicitly and mildly induces some", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d38932d9ed7eeff8ef6a9a1c3e8f8fa5a87dc9722d6210ba9d61005f6f22ba19.jpg", + "table_caption": [ + "Table 3. Attack success rates (\\%) of different surrogate refinements against various defenses. AT is trained with RN-50; HGD is based on Inc-v3; NRP, JPEG and Bit are averaged on seven ViTs and six CNNs." + ], + "table_footnote": [], + "table_body": "
MethodViTsCNNs
SIA [36]73.469.1
SIA+FPR86.690.4
BSR [32]69.466.9
BSR+FPR84.889.9
VTM [34]63.157.3
VTM+FPR84.081.7
GRA [49]73.668.3
GRA+FPR86.185.3
", + "bbox": [ + 522, + 395, + 897, + 551 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6e9bb1b33289957cca9a5ad1df363a13ff1faf836ea05e08746bccc6b9f64867.jpg", + "table_caption": [ + "Table 4. Attack success rates $(\\%)$ of combined methods averaged on six CNNs and six ViTs (without surrogate model ViT-B)." + ], + "table_footnote": [], + "table_body": "
MethodViTsCNNs
MIM [4]51.650.2
Dropout55.858.9
TGR [45]77.676.2
GNS [51]80.074.6
AMD82.178.4
MTE78.479.6
FPR (AMD+MTE)84.783.8
", + "bbox": [ + 524, + 633, + 895, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation studies of different FPR components with ViT-B as the surrogate model. All involved methods are built on MIM. Dropout is applied at the same index set as AMD.", + "bbox": [ + 511, + 784, + 906, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gradient vanishing in the backward propagation.", + "bbox": [ + 513, + 885, + 831, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25077", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/618e5a8c374e4fda37b1ad913610fe6953af7a7a52521ad10689397e7c301fe8.jpg", + "image_caption": [ + "Figure 4. Sensitivity analysis about diversity factor, scale factor, and attenuation factor, where ASR remains relatively stable within a certain range of values for these three factors." + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 890, + 577 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2149433602d52476e0755f1a66b8c2299e653c2da52dbae567f16b2471347ca3.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Mean of attention map's input gradient elements from 100 images over 10 iterations. All surrogate refinements adopt Index Set (0,1,4,9,11). ASR is averaged on 13 models and ViT-B is the surrogate model." + ], + "table_body": "
AttackMIM [4]PNA [38]TGR [45]GNS [51]AMD
Mean2.9 × 10-40.03.9 × 10-102.4 × 10-151.5 × 10-5
ASR50.960.663.262.780.3
", + "bbox": [ + 91, + 638, + 486, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 781, + 209, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose the Forward Propagation Refinement (FPR) to generate highly transferable adversarial examples on Vision Transformers (ViTs). FPR is the first work to refine the surrogate ViTs during forward propagation entirely. Specifically, FPR consists of two parts: Attention Map Diversification (AMD) and Momentum To", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ken Embedding (MTE). AMD diversifies the attention maps and implicitly causes gradient vanishing to enhance transferability, while MTE accumulates historical token embeddings to stabilize current embeddings for more accurate updates. The superiority of FPR is supported by both empirical results and theoretical analyses.", + "bbox": [ + 511, + 642, + 906, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 513, + 752, + 671, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is supported by the National Key Research and Development Program of China (2023YFE0209800), the National Natural Science Foundation of China (62406240, 62376210, 62132011, 62206217, T2341003, U244120060, 62161160337, U2441240, U24B20185, U21B2018), the Shaanxi Province Key Industry Innovation Program (2023-ZDLGY-38). Thanks to the New Cornerstone Science Foundation and the Xplorer Prize.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25078", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Srinadh Bhojanapalli, Ayan Chakrabarti, Daniel Glasner, Daliang Li, Thomas Unterthiner, and Andreas Veit. Understanding robustness of transformers for image classification. In IEEE International Conference on Computer Vision, pages 10211-10221, 2021. 4", + "[2] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In IEEE Conference on Computer Vision and Pattern Recognition, pages 24648-24657, 2023. 2", + "[3] Zhengsu Chen, Lingxi Xie, Jianwei Niu, Xuefeng Liu, Longhui Wei, and Qi Tian. Visformer: The vision-friendly transformer. In IEEE International Conference on Computer Vision, pages 569-578, 2021. 3, 6", + "[4] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 1, 2, 4, 5, 6, 7, 8", + "[5] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, and et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 3, 6", + "[6] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. 1, 2", + "[7] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering adversarial images using input transformations. In International Conference on Learning Representations, 2018. 6", + "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6", + "[9] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In IEEE International Conference on Computer Vision, pages 11916-11925, 2021. 1, 3, 6", + "[10] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint, arXiv:1704.04861, 2017. 6", + "[11] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2261-2269, 2017. 6", + "[12] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In International Conference on Learning Representations, 2017. 2", + "[13] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural In" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "formation Processing Systems, pages 9694-9705, 2021. 2, 5", + "[14] Qizhang Li, Yiwen Guo, Wangmeng Zuo, and Hao Chen. Making substitute models more bayesian can enhance transferability of adversarial examples. In International Conference on Learning Representations, 2023. 2", + "[15] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan Yuille. Learning transferable adversarial examples via ghost networks. In AAAI Conference on Artificial Intelligence, pages 11458-11465, 2020. 1, 2", + "[16] Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial attacks using high-level representation guided denoiser. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2018. 6", + "[17] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, 2019. 3", + "[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE International Conference on Computer Vision, pages 9992-10002, 2021. 3, 6", + "[19] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A self-supervised approach for adversarial robustness. In IEEE Conference on Computer Vision and Pattern Recognition, pages 259-268, 2020. 6", + "[20] Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Fahad Shahbaz Khan, and Fatih Porikli. On improving adversarial transferability of vision transformers. In International Conference on Learning Representations, 2022. 1", + "[21] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In Advances in Neural Information Processing Systems, pages 23296-23308, 2021. 1", + "[22] Yuchen Ren, Hegui Zhu, Xiaoyan Sui, and Chong Liu. Crafting transferable adversarial examples via contaminating the salient feature variance. Information Sciences, 644:119273, 2023. 1", + "[23] Yuchen Ren, Hegui Zhu, Chong Liu, and Chengqing Li. Efficient polar coordinates attack with adaptive activation strategy. Expert Systems with Applications, 249:123850, 2024. 1", + "[24] Yuchen Ren, Zhengyu Zhao, Chenhao Lin, Bo Yang, Lu Zhou, Zhe Liu, and Chao Shen. Improving integrated gradient-based transferable adversarial examples by refining the integration path. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 1", + "[25] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. International Journal of Computer Vision, 115(3): 211–252, 2015. 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25079", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. 6", + "[27] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, 2020. 6", + "[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357, 2021. 1, 3, 6", + "[29] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Herve Jegou. Going deeper with image transformers. In IEEE International Conference on Computer Vision, pages 32-42, 2021. 3, 6", + "[30] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 6", + "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, and et al. Attention is all you need. In International Conference on Neural Information Processing Systems, pages 6000-6010, 2017. 3", + "[32] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting adversarial transferability by block shuffle and rotation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 24336-24346, 2024. 3, 6, 7", + "[33] Kang Wang, Jianzhou Wang, Zhiwu Li, and Yilin Zhou. An integrated sparse gated graph density network based on transfer learning for multi-site probabilistic forecasting of renewable energy. IEEE Transactions on Sustainable Energy, 16(1):673-685, 2025. 1", + "[34] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1924-1933, 2021. 1, 3, 6, 7", + "[35] Xiaosen Wang, Kangheng Tong, and Kun He. Rethinking the backward propagation for adversarial transferability. In Advances in Neural Information Processing Systems, pages 1905-1922, 2023. 1, 2", + "[36] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In IEEE International Conference on Computer Vision, pages 4584-4596, 2023. 3, 6, 7", + "[37] Zhiyuan Wang, Zeliang Zhang, Siyuan Liang, and Xiaosen Wang. Diversifying the high-level features for better adversarial transferability. In British Machine Vision Conference 2023, BMVC 2023, 2023. 2", + "[38] Zhipeng Wei, Jingjing Chen, Micah Goldblum, Zuxuan Wu, Tom Goldstein, and Yu-Gang Jiang. Towards transferable adversarial attacks on vision transformers. In AAAI Conference on Artificial Intelligence, pages 2668-2676, 2022. 1, 2, 4, 6, 7, 8", + "[39] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip connections matter: On the transferability of adversarial examples generated with resnets. In Interna-" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional Conference on Learning Representations, 2020. 1, 2, 5", + "[40] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 3", + "[41] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5987-5995, 2017. 1, 6", + "[42] Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In Network and Distributed System Security Symposium, 2018. 6", + "[43] Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Coscale conv-attentional image transformers. In IEEE International Conference on Computer Vision, pages 9961–9970, 2021. 3, 6", + "[44] Bo Yang, Hengwei Zhang, Jindong Wang, Yulong Yang, Chenhao Lin, Chao Shen, and Zhengyu Zhao. Adversarial example soups: Improving transferability and stealthiness for free. IEEE Transactions on Information Forensics and Security, 20:1882-1894, 2025. 1", + "[45] Jianping Zhang, Yizhan Huang, Weibin Wu, and Michael R. Lyu. Transferable adversarial attacks on vision transformers with token gradient regularization. In IEEE Conference on Computer Vision and Pattern Recognition, pages 16415-16424, 2023. 1, 2, 6, 7, 8", + "[46] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. Towards large yet imperceptible adversarial image perturbations with perceptual color distance. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1036-1045, 2020. 1", + "[47] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. On success and simplicity: A second look at transferable targeted attacks. In Advances in Neural Information Processing Systems, pages 6115-6128, 2021. 1", + "[48] Zhengyu Zhao, Hanwei Zhang, Renjue Li, Ronan Sicre, Laurent Amsaleg, Michael Backes, Qi Li, and Chao Shen. Revisiting transferable adversarial image examples: Attack categorization, evaluation guidelines, and new insights. arXiv preprint, arXiv:2310.11850, 2023. 2", + "[49] Hegui Zhu, Yuchen Ren, Xiaoyan Sui, Lianping Yang, and Wuming Jiang. Boosting adversarial transferability via gradient relevance attack. In IEEE International Conference on Computer Vision, pages 4718-4727, 2023. 3, 6, 7", + "[50] Yao Zhu, Yuefeng Chen, Xiaodan Li, Kejiang Chen, Yuan He, Xiang Tian, Bolun Zheng, Yaowu Chen, and Qingming Huang. Toward understanding and boosting adversarial transferability from a distribution perspective. IEEE Transactions on Image Processing, 31:6487-6501, 2022. 2", + "[51] Zhiyu Zhu, Xinyi Wang, Zhibo Jin, Jiayu Zhang, and Huaming Chen. Enhancing transferable adversarial attacks on vision transformers through gradient normalization scaling and high-frequency adaptation. In International Conference on Learning Representations, 2024. 2, 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 903, + 882 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25080", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/c0d9e11b-700c-40d7-8dca-46321613cb22_model.json b/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/c0d9e11b-700c-40d7-8dca-46321613cb22_model.json new file mode 100644 index 0000000000000000000000000000000000000000..464f20152e6f27f9a970847319c1f11dde75e071 --- /dev/null +++ b/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/c0d9e11b-700c-40d7-8dca-46321613cb22_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.131, + 0.819, + 0.177 + ], + "angle": 0, + "content": "Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.203, + 0.856, + 0.258 + ], + "angle": 0, + "content": "Yuchen Ren\\(^{1}\\), Zhengyu Zhao\\(^{1*}\\), Chenhao Lin\\(^{1}\\), Bo Yang\\(^{2}\\), Lu Zhou\\(^{3}\\), Zhe Liu\\(^{4}\\), Chao Shen\\(^{1}\\) \n\\(^{1}\\)Xi'an Jiaotong University, China; \\(^{2}\\)Information Engineering University, China; \\(^{3}\\)Nanjing University of Aeronautics and Astronautics, China; \\(^{4}\\)Zhejiang Lab, China" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.26, + 0.831, + 0.294 + ], + "angle": 0, + "content": "ryc98@stu.xjtu.edu.cn, {zhengyu.zhao, linchenhao}@xjtu.edu.cn, yangbo_hn@163.com, {lu.zhou, zhe.liu}@nuaa.edu.cn, chaoshen@mail.xjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.328, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.361, + 0.486, + 0.739 + ], + "angle": 0, + "content": "Vision Transformers (ViTs) have been widely applied in various computer vision and vision-language tasks. To gain insights into their robustness in practical scenarios, transferable adversarial examples on ViTs have been extensively studied. A typical approach to improving adversarial transferability is by refining the surrogate model. However, existing work on ViTs has restricted their surrogate refinement to backward propagation. In this work, we instead focus on Forward Propagation Refinement (FPR) and specifically refine two key modules of ViTs: attention maps and token embeddings. For attention maps, we propose Attention Map Diversification (AMD), which diversifies certain attention maps and also implicitly imposes beneficial gradient vanishing during backward propagation. For token embeddings, we propose Momentum Token Embedding (MTE), which accumulates historical token embeddings to stabilize the forward updates in both the Attention and MLP blocks. We conduct extensive experiments with adversarial examples transferred from ViTs to various CNNs and ViTs, demonstrating that our FPR outperforms the current best (backward) surrogate refinement by up to \\(7.0\\%\\) on average. We also validate its superiority against popular defenses and its compatibility with other transfer methods. Codes and appendix are available at https://github.com/RYC-98/FPR." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.771, + 0.224, + 0.787 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.797, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) [8, 33, 41] demonstrate remarkable performance across diverse domains, with Vision Transformers (ViTs) [5, 9, 28] excelling particularly in computer visualization. However, ViTs are also shown to be vulnerable to adversarial examples [20, 21], which are" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.327, + 0.898, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.55, + 0.908, + 0.581 + ], + "angle": 0, + "content": "Figure 1. Our Forward Propagation Refinement (FPR) vs. previous methods which only refine the backward propagation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.64 + ], + "angle": 0, + "content": "optimized by adding subtle perturbations [23, 46] into clean images [6, 44, 47]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.779 + ], + "angle": 0, + "content": "A key property of adversarial examples that causes practical security concerns is their transferability, i.e., they can be optimized on a (seen) surrogate model and then directly transfer their effects to (unseen) target models [4, 40]. Compared to other transfer approaches that focus on loss design, optimization algorithm or data augmentation [22, 24, 34], surrogate refinement [15, 35, 39] explores the intrinsic relationship between the model itself and adversarial transferability, offering a unique perspective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "However, existing surrogate refinements on ViTs [38, 45] are limited to refining the backward propagation, i.e., only modifying the gradient information. To address the above limitation, in this work, we explore the potential of Forward Propagation Refinement (FPR), i.e., modifying features during forward propagation. Our FPR not only provides a new perspective of surrogate refinement but also potentially benefits the refinement during the sub" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "25071" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.181 + ], + "angle": 0, + "content": "sequent backward propagation (see our discussion in Section 3.2). Specifically, FPR comprises two parts that refine two key components of ViTs respectively: attention maps and output token embeddings. Figure 1 highlights the difference between our forward-based method and the previous backward-based methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.482, + 0.38 + ], + "angle": 0, + "content": "For attention maps, we identify that certain attention maps in ViTs are redundant and can deteriorate transferability. Rather than simply dropping redundant attention maps, which recklessly removes all information, we propose Attention Map Diversification (AMD), which moderately diversifies the attention maps based on random weighting. The introduced randomness on attention maps acts as a simulation of unknown models, thereby mitigating the overfitting to the surrogate model. Although AMD only modifies the forward propagation, we find that it can potentially induce the vanishing of certain gradients flowing through the attention maps during backward propagation, further alleviating overfitting [38]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.482, + 0.517 + ], + "angle": 0, + "content": "For output token embeddings, we posit that adversarial examples during generation are highly likely to be near the local optima, resulting in the surrogate model's embeddings containing erroneous information. Consequently, inspired by the momentum distillation technology when training the vision-language pretraining model [13], we propose Momentum Token Embedding (MTE), to stabilize the updates of token embeddings in each iteration by accumulating historical token embedding information." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.518, + 0.449, + 0.532 + ], + "angle": 0, + "content": "In sum, we make the following main contributions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.534, + 0.483, + 0.579 + ], + "angle": 0, + "content": "- We explore the refinement of ViT surrogate models from the perspective of forward propagation, particularly modifying the attention maps and output token embeddings." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.58, + 0.483, + 0.655 + ], + "angle": 0, + "content": "- We propose Attention Map Diversification (AMD) to moderately avoid attention redundancy, which also induces beneficial gradient vanishing during the subsequent backward propagation, altogether mitigating the overfitting to the surrogate model." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.655, + 0.483, + 0.715 + ], + "angle": 0, + "content": "- We propose Momentum Token Embedding (MTE) to stabilize the updates of token embeddings by accumulating historical information, during the generation of adversarial examples on ViTs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.715, + 0.483, + 0.789 + ], + "angle": 0, + "content": "- Extensive experiments validate that our Forward Propagation Refinement (FPR), comprising AMD and MTE, outperforms the current best (backward) refinement method by up to an average of \\(7.0\\%\\) against 13 ViTs and CNNs with or without defenses." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.534, + 0.483, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.807, + 0.233, + 0.822 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.36, + 0.849 + ], + "angle": 0, + "content": "2.1. Surrogate Refinement Attacks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this subsection, we specifically delineate attack methods falling within the surrogate refinement, since it is the main topic of this work. Skip Gradient Method (SGM) [39] attenuates the attack by using a gradient descent approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.393 + ], + "angle": 0, + "content": "uates gradients traversing the residual modules to mitigate overfitting to the surrogate model. Ghost Networks [15] employs dropout in each block of the surrogate model and perturbs the skip connections to create a longitudinal ensemble. Backward Propagation Attack (BPA) [35] demonstrates that gradients traversing ReLU and max-pooling layers can carry inaccurate information. To address this, it adopts both non-monotonic function and softmax with temperature to recover the true gradients, thereby enhancing transferability. The aforementioned methods focus on the backward propagation of CNNs, while a few approaches also modify the output features of convolutional layers during forward propagation. Diversifying the High-level Features (DHF) [37] modifies the high-level features through random transformations. Clean Feature Mixup (CFM) [2] introduces competition into the generation of adversarial examples at the feature level for stronger transferability. Several methods also retrain the surrogate model to strengthen the transferability [14, 50], however, this may significantly increase the attack cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.905, + 0.758 + ], + "angle": 0, + "content": "The above surrogate refinements are primarily designed for CNNs; several methods have also been developed for ViTs. Pay No Attention and Patch Out (PNAPO) [38] comprises three modifications to synergistically mitigate overfitting to the surrogate model. First, it cuts off gradients traversing all attention maps. Second, it randomly updates a specific number of input patches. Finally, it adds an extra term to the cross-entropy loss function that increases the adversarial noise norm to enhance transferability. Token Gradient Regularization (TGR) [45] posits that extreme values in gradients are detrimental. Therefore, it removes the extreme gradients of QKV embeddings, attention maps, and MLP blocks to reduce gradient variance, thereby alleviating overfitting to the surrogate models. Gradient Normalization Scaling (GNS) [51] improves TGR by accurately normalizing and scaling the identified mild gradients, which yields stronger transferability. Even though the above methods exhibit the ability to craft transferable adversarial examples, they neglect the broader implications of forward propagation, which prevents further transferability. In this work, we focus on refining the ViT surrogate model from the new perspective of forward propagation and specifically explore two key components: attention maps and output token embeddings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.78, + 0.787 + ], + "angle": 0, + "content": "2.2. Other Transfer-based Attacks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This subsection reviews transfer-based attacks [48] that do not need to modify the surrogate models. Fast Gradient Sign Method (FGSM) [6] creates adversarial examples in a single step by following the gradient sign's direction. Basic Iterative Method (BIM) [12] extends FGSM by applying multiple iterations. Momentum Iterative Method (MIM) [4] introduces a momentum term for gradients at each qua" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25072" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.089, + 0.868, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.135, + 0.322, + 0.861, + 0.378 + ], + "angle": 0, + "content": "\\(E_{t}^{A} / E_{t}^{M}\\): Output Token Embedding of Attention / MLP Block at \\(t\\)-th Iteration \\(\\widehat{E_{t}^{A}} / \\widehat{E_{t}^{M}}\\): Accumulated Token Embedding of Attention / MLP Block at \\(t\\)-th Iteration \\(A\\): Attention Map \\(A'\\): Diversified Attention Map \\(M\\): Diversity Matrix \\(Q/K/V\\): Query/Key/Value Component" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.392, + 0.908, + 0.422 + ], + "angle": 0, + "content": "Figure 2. The overview of Forward Propagation Refinement (FPR). Note that Attention Map Diversification (AMD) is applied to specific attention maps according to the index set, while Momentum Token Embedding (MTE) is applied to all Attention and MLP blocks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.447, + 0.485, + 0.735 + ], + "angle": 0, + "content": "tion for better convergence. Nesterov Fast Gradient Sign Method (NIM) [17] takes an extra step at the beginning of each iteration and adopts the example at the farther location to compute the gradients for the update. Variance Tuning Method (VTM) [34] modifies current gradients with variance at the last iteration. Gradient Relevance Attack (GRA) [49] utilizes correlation information extracted from the nearby samples to stabilize the current update direction. Some input transformations have also been presented to enhance transferability. Diversity Input Method (DIM) [40] combines padding and resizing transformations on the input to reduce overfitting. Scale Invariance Method (SIM) [17] employs a group of scaled copies to enhance the input. Structure Invariant Attack (SIA) [36] employs different transformations from a library to each block of the input to prompt the transferability. Block Shuffle and Rotation (BSR) [32] divides the input image into several blocks, which are then randomly shuffled and rotated to create a set of new images for gradient computation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.287, + 0.758 + ], + "angle": 0, + "content": "2.3. Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Transformer [31] is widely used in natural language processing and has since gained increasing recognition in computer vision. ViT [5] processes images by dividing them into patches and embedding these patches as tokens. They employ a Transformer architecture with multi-head self-attention mechanisms to effectively capture spatial relationships. DeiT [28] improves ViT through teacher-student distillation, enabling superior performance with less data. Swin [18] employs a hierarchical structure with shifted win" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.598 + ], + "angle": 0, + "content": "dows to efficiently handle multi-scale attention. PiT [9] utilizes pooling layers to streamline computational processes and manage data flow. CaiT [29] incorporates LayerScale to stabilize training as model depth increases and employs Class-Attention layers to enhance the extraction of classification information. Visformer [3] combines convolutional layers with ViT to effectively integrate spatial features and attention mechanisms. CoaT [43] merges convolution and self-attention in a hierarchical architecture to balance feature extraction and feature interaction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.613, + 0.649, + 0.63 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.906, + 0.714 + ], + "angle": 0, + "content": "In this section, we propose the Forward Propagation Refinement (FPR) to enhance the transferability of adversarial examples from ViTs, consisting of Attention Map Diversification (AMD) and Momentum Token Embedding (MTE). Figure 2 provides an overview of FPR." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.643, + 0.74 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.907, + 0.853 + ], + "angle": 0, + "content": "Assume an image \\( x \\) with channel \\( C \\), height \\( H \\), width \\( W \\), the true label \\( y \\), and a surrogate model \\( F \\), the goal of a transfer-based attack is to add adversarial perturbation \\( \\delta \\) to the input \\( x \\), creating an adversarial example \\( x^{a} = x + \\delta \\), which is then fed into the target model \\( F^{t} \\) to induce incorrect predictions. This process can be formulated as solving the following problem:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.879, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\delta^ {\\prime} = \\underset {\\delta} {\\arg \\max } L (F (x + \\delta), y), s. t. \\| \\delta \\| _ {\\infty} \\leq \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25073" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.133, + 0.089, + 0.445, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.32, + 0.485, + 0.391 + ], + "angle": 0, + "content": "Figure 3. Illustration of redundancy in attention maps: Randomly dropping a small number of blocks leads to improved transferability. The results are averaged over 5 repeated runs on 13 target models, and the dash lines denote the baseline results of the MIM [4] without dropping." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.444 + ], + "angle": 0, + "content": "where \\(L\\) is the loss function and \\(\\epsilon\\) is the maximum perturbation bound." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Before the image \\( x \\) is fed into ViTs, it is segmented into \\( N = \\frac{H \\cdot W}{P^2} \\) patches, with each patch having a height, width, and channel dimension of \\( P \\), \\( P \\), and \\( C \\), respectively. Then, these patches are converted into token embeddings, and fed into a sequence of Transformer blocks. Each Transformer block includes an Attention block and a MLP block." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.643 + ], + "angle": 0, + "content": "The input token embedding of the Attention block is transformed into Query \\((Q)\\), Key \\((K)\\), and Value \\((V)\\) components. Here, the shapes of \\(Q\\), \\(K\\) and \\(V\\) are \\(N\\times D\\), where \\(D\\) is the embedding dimension. The Attention block employs the self-attention mechanism to establish the relationship between \\(Q\\), \\(K\\), and \\(V\\). The attention map \\(A\\) in the Attention block is computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.648, + 0.483, + 0.684 + ], + "angle": 0, + "content": "\\[\nA = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {D}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.734 + ], + "angle": 0, + "content": "where the softmax function is employed at each row. The attention map \\( A \\) is then used to weight the \\( V \\), and the output \\( Z \\) is calculated as \\( Z = AV \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.797 + ], + "angle": 0, + "content": "The MLP block follows the Attention block and typically includes the fully connected layer, GELU activation function, dropout layer, and identity normalization layer to aggregate information for all token embeddings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.416, + 0.82 + ], + "angle": 0, + "content": "3.2. Attention Map Diversification (AMD)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The previous work [1] has demonstrated the redundancy in attention blocks of ViTs, showing that dropping a few such blocks does not significantly degrade classification accuracy. Therefore, it is reasonable to posit that such redundancy may also exist in generating adversarial examples," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "causing potential overfitting to the surrogate model. To verify this, we conduct exploratory experiments by dropping certain attention blocks. As can be seen from Figure 3, dropping a small number of blocks indeed leads to improved transferability (while dropping too many may harm transferability)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.334 + ], + "angle": 0, + "content": "To avoid such redundancy, a straightforward idea is to apply dropout of attention blocks while generating adversarial examples. However, we find that the dropout is too aggressive and may cause serious information loss, leading to only a small improvement of transferability (see comparing results in Table 5 of Section 4.3). Therefore, we propose a more moderate method called Attention Map Diversification (AMD), which transforms the original attention map \\( A \\) by multiplying it with a diversity matrix \\( M \\). Specifically, the diversified attention map \\( A' \\) is obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.348, + 0.905, + 0.365 + ], + "angle": 0, + "content": "\\[\nA ^ {\\prime} = A M D (A) = \\operatorname {s o f t m a x} (A \\odot M), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.906, + 0.474 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the element-wise product and softmax function is applied to each row. The diversity matrix \\(M\\) has the same shape as \\(A\\), and every element \\(M_{i,j}\\) \\((1 \\leq i,j \\leq N)\\) conforms to the uniform distribution \\(U(1 - d, 1 + d)\\), where \\(d\\) is a factor that controls the diversity level. Note that AMD is applied to attention maps at specified Transformer blocks according to a index set \\(I\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.699 + ], + "angle": 0, + "content": "The benefit of AMD on backward propagation. In principle, modifying the forward propagation would impact the subsequent backward propagation. Here, we theoretically demonstrate that the impact of our AMD on backward propagation can further improve transferability. Specifically, we find that AMD implicitly induces gradient vanishing at attention map positions during backward propagation, which aligns with a previous surrogate refinement, PNAPO [38], that directly truncates the gradients of attention maps during backpropagation. The underlying principle of PNAPO suggests that gradients passing through attention maps carry overfitting information from surrogate models, and attenuating these gradients can enhance the transferability of adversarial examples. Our findings can be explained as follows." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.701, + 0.906, + 0.732 + ], + "angle": 0, + "content": "When incorporating AMD, we let \\( S = A \\odot M \\) and the \\( i \\)-th row vector of diversified attention map \\( A_{i}^{\\prime} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.74, + 0.91, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} A _ {i} ^ {\\prime} = \\left(A _ {i, 1} ^ {\\prime}, A _ {i, 2} ^ {\\prime}, \\dots , A _ {i, N} ^ {\\prime}\\right) \\\\ = \\operatorname {s o f t m a x} \\left(A _ {i, 1} \\cdot M _ {i, 1}, A _ {i, 2} \\cdot M _ {i, 2},..., A _ {i, N} \\cdot M _ {i, N}\\right) \\\\ = \\operatorname {s o f t m a x} \\left(S _ {i, 1}, S _ {i, 2}, \\dots , S _ {i, N}\\right) \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.8, + 0.903, + 0.822 + ], + "angle": 0, + "content": "and the derivative \\(\\frac{\\partial A_{i,m}^{\\prime}}{\\partial S_{i,n}}\\) (\\(1 \\leq m, n \\leq N\\)) is expressed by:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.829, + 0.905, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial A _ {i , m} ^ {\\prime}}{\\partial S _ {i , n}} = \\left\\{ \\begin{array}{l l} A _ {i, m} ^ {\\prime} \\cdot \\left(1 - A _ {i, m} ^ {\\prime}\\right), & \\text {i f} m = n \\\\ - A _ {i, m} ^ {\\prime} \\cdot A _ {i, n} ^ {\\prime}, & \\text {i f} m \\neq n \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "As the diversity factor \\(d\\) becomes large, elements in the diversity matrix \\(M\\) will fluctuate significantly, causing ex" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25074" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "treme values in \\( S \\), and ultimately pushing the elements in \\( A' \\) towards 0 or 1 due to the property of softmax function. From Eq. 5, whether the elements in \\( A' \\) approach 0 or 1, the derivative will tend towards 0, resulting in gradient vanishing. According to the chain rule of differentiation, the gradients of the attention maps subjected to our AMD will experience varying degrees of gradient vanishing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.482, + 0.333 + ], + "angle": 0, + "content": "Compared with PNAPO, which directly discards all gradients flowing through the attention maps, the gradient vanishing in our AMD is implicit without directly altering any aspect of the backward propagation. Furthermore, the gradient vanishing in our AMD method is more fine-grained, occurring at the row level of the attention map. Each row experiences varying degrees of vanishing, introducing additional randomness that helps to further mitigate overfitting to the surrogate model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.342, + 0.42, + 0.358 + ], + "angle": 0, + "content": "3.3. Momentum Token Embedding (MTE)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.59 + ], + "angle": 0, + "content": "It is challenging to ensure convergence to the global optimum in the high-dimension space when generating adversarial examples. Typically, adversarial examples converge to a local optimum along an inaccurate path [4]. This causes the internal token embeddings of the surrogate model to contain inaccurate information during iterations, thereby weakening the transferability. Therefore, we borrow the idea of Momentum Distillation (MoD) [13] from the Vision-and-Language Pre-training (VLP) literature, where a momentum version of the model through the moving average of the primary model's parameters is maintained to mitigate the noise in the web-sourced training dataset. Similarly, we maintain the momentum version of token embedding updates to accumulate historical token embeddings while generating adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.59, + 0.483, + 0.621 + ], + "angle": 0, + "content": "We term this method Momentum Token Embedding (MTE), and formulate it as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.629, + 0.483, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\widehat {E _ {t}} = M T E \\left(E _ {t}\\right) = \\eta \\cdot \\widehat {E _ {t - 1}} + s \\cdot E _ {t}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.826 + ], + "angle": 0, + "content": "where \\( t \\) is the iteration index, \\( E_{t} \\) is the output token embedding of any Attention block or MLP block at the \\( t \\)-th iteration, \\( \\widehat{E_t} \\) is the corresponding accumulated token embedding at the \\( t \\)-th iteration, \\( \\eta \\) is the attenuation factor, and \\( s \\) is the scale factor. MTE is applied to all Attention and MLP blocks. Note that, different from MoD, our MTE does not require model training and can be seen as constructing a residual structure for output token embeddings across the time dimension. Previous work also finds the usefulness of the residual structure in improving transferability, but during backward propagation within each iteration [39]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.379, + 0.85 + ], + "angle": 0, + "content": "3.4. Complementary AMD and MTE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The previous subsections examine AMD and MTE from their respective perspectives. Here, we further discuss why they can collaboratively enhance transferability." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.092, + 0.869, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Forward Propagation Refinement (FPR)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.109, + 0.905, + 0.183 + ], + "angle": 0, + "content": "Input: A surrogate model \\(F\\) with the cross-entropy loss \\(L\\), a clean image \\(x\\) and its true label \\(y\\). The maximum perturbation bound \\(\\epsilon\\), the iteration number \\(T\\), the decay factor \\(\\mu\\), step size \\(\\alpha\\), index set \\(I\\), diversity factor \\(d\\), scale factor \\(s\\) and attenuation factor \\(\\eta\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.805, + 0.199 + ], + "angle": 0, + "content": "Output: The adversarial image \\( x^{adv} = x_{T} \\)." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.202, + 0.773, + 0.216 + ], + "angle": 0, + "content": "1: Initialize \\( g_0 = 0 \\), \\( E_0 = 0 \\), \\( x_0 = x \\)." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.217, + 0.666, + 0.23 + ], + "angle": 0, + "content": "2: for \\( t = 1 \\rightarrow T \\) do" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.202, + 0.773, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.542, + 0.232, + 0.777, + 0.246 + ], + "angle": 0, + "content": "During the forward propagation:" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.247, + 0.905, + 0.276 + ], + "angle": 0, + "content": "3: Apply AMD (Eq. (3)) to attention maps in specific Transformer blocks according to the index set \\( I \\)." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.277, + 0.905, + 0.306 + ], + "angle": 0, + "content": "4: Apply MTE (Eq. (6)) to output token embeddings in all Transformer (Attention and MLP) blocks." + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.247, + 0.905, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.542, + 0.308, + 0.79, + 0.322 + ], + "angle": 0, + "content": "During the backward propagation:" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.322, + 0.861, + 0.337 + ], + "angle": 0, + "content": "5: Calculate the gradient: \\(\\nabla_{x_{t - 1}}L(F(x_{t - 1}),y)\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.338, + 0.759, + 0.352 + ], + "angle": 0, + "content": "6: Update momentum gradient:" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.322, + 0.861, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.363, + 0.862, + 0.4 + ], + "angle": 0, + "content": "\\[\ng _ {t} = \\mu \\cdot g _ {t - 1} + \\frac {\\nabla_ {x _ {t - 1}} L (F (x _ {t - 1}) , y)}{\\left\\| \\nabla_ {x _ {t - 1}} L (F (x _ {t - 1}) , y) \\right\\| _ {1}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.41, + 0.781, + 0.424 + ], + "angle": 0, + "content": "7: Update the adversarial example:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.437, + 0.837, + 0.454 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\operatorname {C l i p} _ {\\epsilon} \\left\\{x _ {t - 1} + \\alpha \\cdot \\operatorname {s i g n} \\left(g _ {t}\\right) \\right\\}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.523, + 0.465, + 0.597, + 0.478 + ], + "angle": 0, + "content": "8: end for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.907, + 0.771 + ], + "angle": 0, + "content": "Specifically, AMD and MTE refine the forward propagation of ViTs from two complementary angles. AMD falls under the category of \"external refinement\", which enhances surrogate models by introducing extraneous information or perturbing existing embeddings. Techniques such as attention map dropout also fall into this category. Our empirical results demonstrate that this type of refinement is relatively aggressive, making it unsuitable for concurrent use or for application across all blocks. In contrast, MTE belongs to the \"internal refinement\" category, mainly leveraging information from the input data itself. Since these refinements rely on self-related information, they are typically more moderate and can be applied to all blocks. Both AMD and MTE help promoting transferability but in different ways, this complementary effect further enhances the transferability of adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Finally, we provide the pseudocode of FPR (based on MIM [4]) in Algorithm 1. Specifically, at each of \\( T \\) iterations, we first apply AMD to specific attention maps according to the index set \\( I \\) and MTE to the outputs of all Attention and MLP blocks during the forward propagation using PyTorch's Hook function. Then, during the backward propagation, we compute the gradient of the current adversarial example and use MIM [4] to update it." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25075" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.089, + 0.447, + 0.181 + ], + "angle": 0, + "content": "
ModelIdsη
ViT-B0,1,4,9,11250.80.3
CaiT-S2,14,25300.60.2
PiT-T1,6,11250.70.2
DeiT-B0,1,5,10,11300.70.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.192, + 0.482, + 0.262 + ], + "angle": 0, + "content": "Table 1. The optimal parameter settings of our FPR about its index set \\((I)\\), diversity factor \\((d)\\), scale factor \\((s)\\), and attenuation factor \\((\\eta)\\). Detailed ablation studies in Section 4.3 provide the sensitivity analysis of the attack performance towards the parameter selection under four cases." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.289, + 0.222, + 0.308 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.314, + 0.296, + 0.332 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.483, + 0.427 + ], + "angle": 0, + "content": "Datasets and baselines. Following the common practice [34, 45, 49], we use 1000 images randomly selected from the ILSVRC2012 [25] validation set. We compare our FPR with other surrogate refinements that focus on backward propagation: PNAPO [38], TGR [45], and GNS [51]. All these methods are built on MIM [4]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.483, + 0.669 + ], + "angle": 0, + "content": "Models and defenses. For surrogate models, we consider four popular ViTs: ViT-Base (ViT-B) [5], CaiT-Small (CaiT-S) [29], PiT-Tiny (PiT-T) [9] and Deit-Base (Deit-B) [28]. For target models, we consider seven ViTs and six CNNs: ViT-B [5], CaiT-S [29], PiT-Base (PiT-B) [9], Visformer-Small (Vis-S) [3], Swin-Tiny (Swin-T) [18], DeiT-Tiny (DeiT-T) [28], and CoaT-Tiny (CoaT-T) [43], ResNet-18 (RN-18) [8], VGG-16 [26], DenseNet-121 (DN-121) [11], EfficientNet-b0 (EN-b0) [27], MobileNetv3 (MN-v3) [10], and ResNeXt-50 (RNX-50) [41]. Five popular defenses are tested: adversarial training (AT) paired with RN-50 [30], high-level representation guided denoiser (HGD) [16], neural representation purifier (NRP) [19], Bit depth reduction (BDR) [42], and JPEG compression [7]. Finally, two online models, i.e., Baidu Cloud API and Aliyun API, are also included." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.806 + ], + "angle": 0, + "content": "Hyperparameter settings. Commonly, we set the number of iterations as 10, the maximum perturbation bound \\(\\epsilon\\) as 16, the step size \\(\\alpha\\) as 1.6, and the decay factor of momentum gradient \\(\\mu = 1\\). Parameters of all baselines are the same as their default values. For our FPR, the optimal hyperparameters are listed in Table 1, with a detailed analysis in Section 4.3 discussing their sensitivity to different parameter choices. All experiments are conducted on a single RTX 4060 GPU with 8 GB of VRAM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.292, + 0.835 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Different target models. We first compare FPR with baselines on transferring adversarial examples from ViTs to different ViTs and CNNs. Table 2 shows that FPR consistently outperforms baselines in all cases, suggesting the potential" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "of refining the surrogate model from the perspective of forward propagation. For example, when generating adversarial examples on ViT-B, FPR achieves an average attack success rate of \\(84.3\\%\\) on six CNNs, surpassing GNS, TGR, and PNAPO by \\(7.0\\%\\), \\(7.4\\%\\), and \\(15.5\\%\\), respectively, and improves the baseline MIM by \\(33.4\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.905, + 0.319 + ], + "angle": 0, + "content": "Different defenses. We then evaluate different attack methods against five popular defenses. As shown in Table 3, FPR maintains its superiority over existing surrogate refinements in such challenging defense settings. For instance, when generating adversarial examples on ViT-B, FPR achieves an average attack success rate of \\(70.8\\%\\) across five defenses, outperforming other surrogate refinements GNS, TGR, and PNAPO by \\(5.7\\%\\), \\(6.4\\%\\), and \\(12.4\\%\\), respectively, while surpassing the vanilla baseline, MIM, by \\(24.3\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.905, + 0.53 + ], + "angle": 0, + "content": "Incorporating other transfer methods. A good surrogate refinement-based transfer method should also exhibit strong compatibility with other (types of) transfer-based attack methods. In this study, we integrate FPR into two advanced gradient correction-based transferable attacks, VTM [34] and GRA [49], as well as two latest input augmentation-based transferable attacks, BSR [32] and SIA [36]. Specifically, we equip these four attacks with 5 augmented copies per iteration, while keeping other parameter settings as their default values. Results in Table 4 indicate that FPR can effectively complement existing transfer-based attacks for stronger transferability. For instance, when combined with our FPR, the attack success rate of VTM increases from \\(57.3\\%\\) to \\(81.7\\%\\) on CNNs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.905, + 0.592 + ], + "angle": 0, + "content": "Online models. We also evaluate our FPR on two online models, i.e., Baidu Cloud API and Aliyun API. The results, presented in Appendix A, show that FPR still outperforms both TGR and GNS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.677, + 0.62 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.905, + 0.763 + ], + "angle": 0, + "content": "Individual components. We compare each component of FPR with the plain baseline MIM, and the two best-performed surrogate refinements, TGR and GNS. We also consider Dropout as a comparison to our AMD. From Table 5, we can see that combining our AMD and MTE yields the best transferability results. In addition, Dropout is less effective than our AMD, indicating that directly removing certain attention maps is rather aggressive and fails to retain useful information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Parameter sensitivity. Figure 4 shows the attack success rates when varying three main parameters: diversity factor \\(d\\), scale factor \\(s\\), and attenuation factor \\(\\eta\\). \\(d\\) controls the diversity of the attention map. FPR achieves satisfactory performance when \\(d\\) is between 25 and 30. \\(s\\) represents the importance of current token embeddings. FPR achieves satisfactory performance when \\(s\\) is between 0.6 and 0.8. \\(\\eta\\) signifies the importance of historical token embeddings. FPR achieves satisfactory performance when \\(\\eta\\) is between 0.2" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25076" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.334 + ], + "angle": 0, + "content": "
ModelAttackViT-BCaiT-SPiT-BVis-SSwin-TDeiT-TCoaT-TRN-18VGG-16DN-121EN-b0MN-v3RNX-50Average
ViT-BMIM [4]97.261.840.442.354.965.844.451.757.251.350.851.738.250.9
PNAPO [38]99.183.262.165.874.783.064.067.470.067.668.563.056.368.8
TGR [45]99.186.263.869.981.194.470.479.179.875.383.179.960.076.9
GNS [51]99.989.568.174.082.493.672.277.876.776.079.176.062.077.3
FPR99.292.573.078.388.598.277.787.387.682.090.387.468.384.3
CaiT-SMIM [4]68.598.850.054.968.275.354.861.067.859.861.055.747.160.3
PNAPO [38]69.888.258.161.070.373.860.164.869.764.365.561.753.264.4
TGR [45]90.1100.076.081.391.297.880.686.986.684.387.782.169.184.5
GNS [51]91.2100.078.682.291.697.681.186.687.184.989.681.970.885.3
FPR93.8100.081.685.695.098.886.887.488.788.690.682.976.888.1
PiT-TMIM [4]30.036.138.543.351.075.055.569.568.760.864.571.640.454.2
PNAPO [38]39.550.949.654.961.685.465.078.776.772.277.981.050.864.9
TGR [45]37.947.746.954.162.587.465.182.882.075.581.888.853.766.6
GNS [51]37.144.146.451.959.484.463.379.978.071.576.384.348.663.5
FPR44.356.157.162.469.591.372.288.786.480.487.691.160.372.9
DeiT-BMIM [4]86.082.467.069.579.787.765.369.070.570.171.461.458.272.2
PNAPO [38]83.487.772.474.182.286.071.974.974.774.976.168.066.376.4
TGR [45]92.497.181.284.392.697.983.586.085.285.089.984.574.687.2
GNS [51]92.198.080.584.291.098.079.883.380.783.586.781.370.885.4
FPR94.698.483.886.795.599.587.089.889.288.193.986.879.890.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.344, + 0.907, + 0.374 + ], + "angle": 0, + "content": "Table 2. Attack success rates (%) of different surrogate refinements on various ViTs and CNNs. The \"Average\" column excludes the white-box results with grey background." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.396, + 0.481, + 0.651 + ], + "angle": 0, + "content": "
ModelMethodATHGDNRPJPEGBitAverage
ViT-BMIM [4]43.929.842.157.559.146.5
PNAPO [38]45.647.553.570.974.858.4
TGR [45]49.750.861.677.782.064.4
GNS [51]49.254.861.377.982.365.1
FPR50.761.469.284.588.470.8
CaiT-SMIM [4]44.337.446.264.166.951.8
PNAPO [38]44.843.748.265.267.353.8
TGR [45]49.860.766.481.286.869.0
GNS [51]49.862.767.482.688.070.1
FPR51.168.970.886.290.473.5
PiT-TMIM [4]46.132.042.057.562.047.9
PNAPO [38]47.542.948.565.569.854.8
TGR [45]50.844.050.568.271.957.1
GNS [45]48.938.649.365.368.854.2
FPR51.449.956.173.277.461.6
DeiT-BMIM [4]44.748.152.270.374.958.1
PNAPO [38]46.957.257.073.678.762.7
TGR [45]50.265.267.882.288.170.7
GNS [51]48.361.665.780.686.368.5
FPR50.270.771.387.091.574.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.661, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Table 3. Attack success rates (\\%) of different surrogate refinements against various defenses. AT is trained with RN-50; HGD is based on Inc-v3; NRP, JPEG and Bit are averaged on seven ViTs and six CNNs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.809 + ], + "angle": 0, + "content": "and 0.3. Detailed results of applying the same index set \\((I)\\) to other methods and the sensitivity of the \\(I\\) used in AMD are reported in Appendix B, where we also provide insights for identifying the optimal \\(I\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Gradient vanishing comparison. We compare the extent of gradient vanishing in Table 6. We can conclude that gradient vanishing can improve transferability, but it is not a simple linear correlation. This aligns with our claim that the key of AMD is to diversify attention maps in the forward propagation, while it implicitly and mildly induces some" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.396, + 0.898, + 0.552 + ], + "angle": 0, + "content": "
MethodViTsCNNs
SIA [36]73.469.1
SIA+FPR86.690.4
BSR [32]69.466.9
BSR+FPR84.889.9
VTM [34]63.157.3
VTM+FPR84.081.7
GRA [49]73.668.3
GRA+FPR86.185.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.562, + 0.907, + 0.592 + ], + "angle": 0, + "content": "Table 4. Attack success rates \\((\\%)\\) of combined methods averaged on six CNNs and six ViTs (without surrogate model ViT-B)." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.635, + 0.897, + 0.774 + ], + "angle": 0, + "content": "
MethodViTsCNNs
MIM [4]51.650.2
Dropout55.858.9
TGR [45]77.676.2
GNS [51]80.074.6
AMD82.178.4
MTE78.479.6
FPR (AMD+MTE)84.783.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.785, + 0.907, + 0.829 + ], + "angle": 0, + "content": "Table 5. Ablation studies of different FPR components with ViT-B as the surrogate model. All involved methods are built on MIM. Dropout is applied at the same index set as AMD." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.833, + 0.902 + ], + "angle": 0, + "content": "gradient vanishing in the backward propagation." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25077" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.089, + 0.891, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.588, + 0.908, + 0.618 + ], + "angle": 0, + "content": "Figure 4. Sensitivity analysis about diversity factor, scale factor, and attenuation factor, where ASR remains relatively stable within a certain range of values for these three factors." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.64, + 0.487, + 0.696 + ], + "angle": 0, + "content": "
AttackMIM [4]PNA [38]TGR [45]GNS [51]AMD
Mean2.9 × 10-40.03.9 × 10-102.4 × 10-151.5 × 10-5
ASR50.960.663.262.780.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.697, + 0.484, + 0.752 + ], + "angle": 0, + "content": "Table 6. Mean of attention map's input gradient elements from 100 images over 10 iterations. All surrogate refinements adopt Index Set (0,1,4,9,11). ASR is averaged on 13 models and ViT-B is the surrogate model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.782, + 0.21, + 0.799 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose the Forward Propagation Refinement (FPR) to generate highly transferable adversarial examples on Vision Transformers (ViTs). FPR is the first work to refine the surrogate ViTs during forward propagation entirely. Specifically, FPR consists of two parts: Attention Map Diversification (AMD) and Momentum To" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.734 + ], + "angle": 0, + "content": "ken Embedding (MTE). AMD diversifies the attention maps and implicitly causes gradient vanishing to enhance transferability, while MTE accumulates historical token embeddings to stabilize current embeddings for more accurate updates. The superiority of FPR is supported by both empirical results and theoretical analyses." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.753, + 0.673, + 0.771 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This research is supported by the National Key Research and Development Program of China (2023YFE0209800), the National Natural Science Foundation of China (62406240, 62376210, 62132011, 62206217, T2341003, U244120060, 62161160337, U2441240, U24B20185, U21B2018), the Shaanxi Province Key Industry Innovation Program (2023-ZDLGY-38). Thanks to the New Cornerstone Science Foundation and the Xplorer Prize." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25078" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Srinadh Bhojanapalli, Ayan Chakrabarti, Daniel Glasner, Daliang Li, Thomas Unterthiner, and Andreas Veit. Understanding robustness of transformers for image classification. In IEEE International Conference on Computer Vision, pages 10211-10221, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In IEEE Conference on Computer Vision and Pattern Recognition, pages 24648-24657, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Zhengsu Chen, Lingxi Xie, Jianwei Niu, Xuefeng Liu, Longhui Wei, and Qi Tian. Visformer: The vision-friendly transformer. In IEEE International Conference on Computer Vision, pages 569-578, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[5] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, and et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.482, + 0.491 + ], + "angle": 0, + "content": "[6] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.493, + 0.482, + 0.549 + ], + "angle": 0, + "content": "[7] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering adversarial images using input transformations. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.605 + ], + "angle": 0, + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.482, + 0.675 + ], + "angle": 0, + "content": "[9] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In IEEE International Conference on Computer Vision, pages 11916-11925, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[10] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint, arXiv:1704.04861, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[11] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2261-2269, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[12] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In International Conference on Learning Representations, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural In" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "formation Processing Systems, pages 9694-9705, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[14] Qizhang Li, Yiwen Guo, Wangmeng Zuo, and Hao Chen. Making substitute models more bayesian can enhance transferability of adversarial examples. In International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[15] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan Yuille. Learning transferable adversarial examples via ghost networks. In AAAI Conference on Artificial Intelligence, pages 11458-11465, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[16] Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial attacks using high-level representation guided denoiser. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[17] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE International Conference on Computer Vision, pages 9992-10002, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[19] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A self-supervised approach for adversarial robustness. In IEEE Conference on Computer Vision and Pattern Recognition, pages 259-268, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.561 + ], + "angle": 0, + "content": "[20] Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Fahad Shahbaz Khan, and Fatih Porikli. On improving adversarial transferability of vision transformers. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[21] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In Advances in Neural Information Processing Systems, pages 23296-23308, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[22] Yuchen Ren, Hegui Zhu, Xiaoyan Sui, and Chong Liu. Crafting transferable adversarial examples via contaminating the salient feature variance. Information Sciences, 644:119273, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[23] Yuchen Ren, Hegui Zhu, Chong Liu, and Chengqing Li. Efficient polar coordinates attack with adaptive activation strategy. Expert Systems with Applications, 249:123850, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[24] Yuchen Ren, Zhengyu Zhao, Chenhao Lin, Bo Yang, Lu Zhou, Zhe Liu, and Chao Shen. Improving integrated gradient-based transferable adversarial examples by refining the integration path. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[25] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. International Journal of Computer Vision, 115(3): 211–252, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25079" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.191 + ], + "angle": 0, + "content": "[27] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[29] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Herve Jegou. Going deeper with image transformers. In IEEE International Conference on Computer Vision, pages 32-42, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[30] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, and et al. Attention is all you need. In International Conference on Neural Information Processing Systems, pages 6000-6010, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[32] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting adversarial transferability by block shuffle and rotation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 24336-24346, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[33] Kang Wang, Jianzhou Wang, Zhiwu Li, and Yilin Zhou. An integrated sparse gated graph density network based on transfer learning for multi-site probabilistic forecasting of renewable energy. IEEE Transactions on Sustainable Energy, 16(1):673-685, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[34] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1924-1933, 2021. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[35] Xiaosen Wang, Kangheng Tong, and Kun He. Rethinking the backward propagation for adversarial transferability. In Advances in Neural Information Processing Systems, pages 1905-1922, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[36] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In IEEE International Conference on Computer Vision, pages 4584-4596, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[37] Zhiyuan Wang, Zeliang Zhang, Siyuan Liang, and Xiaosen Wang. Diversifying the high-level features for better adversarial transferability. In British Machine Vision Conference 2023, BMVC 2023, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[38] Zhipeng Wei, Jingjing Chen, Micah Goldblum, Zuxuan Wu, Tom Goldstein, and Yu-Gang Jiang. Towards transferable adversarial attacks on vision transformers. In AAAI Conference on Artificial Intelligence, pages 2668-2676, 2022. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip connections matter: On the transferability of adversarial examples generated with resnets. In Interna-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "tional Conference on Learning Representations, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[40] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[41] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5987-5995, 2017. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[42] Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In Network and Distributed System Security Symposium, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[43] Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Coscale conv-attentional image transformers. In IEEE International Conference on Computer Vision, pages 9961–9970, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[44] Bo Yang, Hengwei Zhang, Jindong Wang, Yulong Yang, Chenhao Lin, Chao Shen, and Zhengyu Zhao. Adversarial example soups: Improving transferability and stealthiness for free. IEEE Transactions on Information Forensics and Security, 20:1882-1894, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[45] Jianping Zhang, Yizhan Huang, Weibin Wu, and Michael R. Lyu. Transferable adversarial attacks on vision transformers with token gradient regularization. In IEEE Conference on Computer Vision and Pattern Recognition, pages 16415-16424, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[46] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. Towards large yet imperceptible adversarial image perturbations with perceptual color distance. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1036-1045, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[47] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. On success and simplicity: A second look at transferable targeted attacks. In Advances in Neural Information Processing Systems, pages 6115-6128, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[48] Zhengyu Zhao, Hanwei Zhang, Renjue Li, Ronan Sicre, Laurent Amsaleg, Michael Backes, Qi Li, and Chao Shen. Revisiting transferable adversarial image examples: Attack categorization, evaluation guidelines, and new insights. arXiv preprint, arXiv:2310.11850, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[49] Hegui Zhu, Yuchen Ren, Xiaoyan Sui, Lianping Yang, and Wuming Jiang. Boosting adversarial transferability via gradient relevance attack. In IEEE International Conference on Computer Vision, pages 4718-4727, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.812 + ], + "angle": 0, + "content": "[50] Yao Zhu, Yuefeng Chen, Xiaodan Li, Kejiang Chen, Yuan He, Xiang Tian, Bolun Zheng, Yaowu Chen, and Qingming Huang. Toward understanding and boosting adversarial transferability from a distribution perspective. IEEE Transactions on Image Processing, 31:6487-6501, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.814, + 0.905, + 0.883 + ], + "angle": 0, + "content": "[51] Zhiyu Zhu, Xinyi Wang, Zhibo Jin, Jiayu Zhang, and Huaming Chen. Enhancing transferable adversarial attacks on vision transformers through gradient normalization scaling and high-frequency adaptation. In International Conference on Learning Representations, 2024. 2, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25080" + } + ] +] \ No newline at end of file diff --git a/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/c0d9e11b-700c-40d7-8dca-46321613cb22_origin.pdf b/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/c0d9e11b-700c-40d7-8dca-46321613cb22_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..587c5abc0547be6abe5c9f2ea4eae78071c17817 --- /dev/null +++ b/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/c0d9e11b-700c-40d7-8dca-46321613cb22_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02c00ab7b2f2d000943a994da113fafd3e553f383ec0f871012032d05e427f19 +size 1828502 diff --git a/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/full.md b/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7a381644a08fddb3e134dfbbae315a777e99c8b7 --- /dev/null +++ b/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/full.md @@ -0,0 +1,305 @@ +# Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement + +Yuchen Ren $^{1}$ , Zhengyu Zhao $^{1*}$ , Chenhao Lin $^{1}$ , Bo Yang $^{2}$ , Lu Zhou $^{3}$ , Zhe Liu $^{4}$ , Chao Shen $^{1}$ $^{1}$ Xi'an Jiaotong University, China; $^{2}$ Information Engineering University, China; $^{3}$ Nanjing University of Aeronautics and Astronautics, China; $^{4}$ Zhejiang Lab, China + +ryc98@stu.xjtu.edu.cn, {zhengyu.zhao, linchenhao}@xjtu.edu.cn, yangbo_hn@163.com, {lu.zhou, zhe.liu}@nuaa.edu.cn, chaoshen@mail.xjtu.edu.cn + +# Abstract + +Vision Transformers (ViTs) have been widely applied in various computer vision and vision-language tasks. To gain insights into their robustness in practical scenarios, transferable adversarial examples on ViTs have been extensively studied. A typical approach to improving adversarial transferability is by refining the surrogate model. However, existing work on ViTs has restricted their surrogate refinement to backward propagation. In this work, we instead focus on Forward Propagation Refinement (FPR) and specifically refine two key modules of ViTs: attention maps and token embeddings. For attention maps, we propose Attention Map Diversification (AMD), which diversifies certain attention maps and also implicitly imposes beneficial gradient vanishing during backward propagation. For token embeddings, we propose Momentum Token Embedding (MTE), which accumulates historical token embeddings to stabilize the forward updates in both the Attention and MLP blocks. We conduct extensive experiments with adversarial examples transferred from ViTs to various CNNs and ViTs, demonstrating that our FPR outperforms the current best (backward) surrogate refinement by up to $7.0\%$ on average. We also validate its superiority against popular defenses and its compatibility with other transfer methods. Codes and appendix are available at https://github.com/RYC-98/FPR. + +# 1. Introduction + +Deep neural networks (DNNs) [8, 33, 41] demonstrate remarkable performance across diverse domains, with Vision Transformers (ViTs) [5, 9, 28] excelling particularly in computer visualization. However, ViTs are also shown to be vulnerable to adversarial examples [20, 21], which are + +![](images/4838ca372ca55753d60cd7d0f8a80a47e97bf880a971991fb1cc392827d28f16.jpg) +Figure 1. Our Forward Propagation Refinement (FPR) vs. previous methods which only refine the backward propagation. + +optimized by adding subtle perturbations [23, 46] into clean images [6, 44, 47]. + +A key property of adversarial examples that causes practical security concerns is their transferability, i.e., they can be optimized on a (seen) surrogate model and then directly transfer their effects to (unseen) target models [4, 40]. Compared to other transfer approaches that focus on loss design, optimization algorithm or data augmentation [22, 24, 34], surrogate refinement [15, 35, 39] explores the intrinsic relationship between the model itself and adversarial transferability, offering a unique perspective. + +However, existing surrogate refinements on ViTs [38, 45] are limited to refining the backward propagation, i.e., only modifying the gradient information. To address the above limitation, in this work, we explore the potential of Forward Propagation Refinement (FPR), i.e., modifying features during forward propagation. Our FPR not only provides a new perspective of surrogate refinement but also potentially benefits the refinement during the sub + +sequent backward propagation (see our discussion in Section 3.2). Specifically, FPR comprises two parts that refine two key components of ViTs respectively: attention maps and output token embeddings. Figure 1 highlights the difference between our forward-based method and the previous backward-based methods. + +For attention maps, we identify that certain attention maps in ViTs are redundant and can deteriorate transferability. Rather than simply dropping redundant attention maps, which recklessly removes all information, we propose Attention Map Diversification (AMD), which moderately diversifies the attention maps based on random weighting. The introduced randomness on attention maps acts as a simulation of unknown models, thereby mitigating the overfitting to the surrogate model. Although AMD only modifies the forward propagation, we find that it can potentially induce the vanishing of certain gradients flowing through the attention maps during backward propagation, further alleviating overfitting [38]. + +For output token embeddings, we posit that adversarial examples during generation are highly likely to be near the local optima, resulting in the surrogate model's embeddings containing erroneous information. Consequently, inspired by the momentum distillation technology when training the vision-language pretraining model [13], we propose Momentum Token Embedding (MTE), to stabilize the updates of token embeddings in each iteration by accumulating historical token embedding information. + +In sum, we make the following main contributions: + +- We explore the refinement of ViT surrogate models from the perspective of forward propagation, particularly modifying the attention maps and output token embeddings. +- We propose Attention Map Diversification (AMD) to moderately avoid attention redundancy, which also induces beneficial gradient vanishing during the subsequent backward propagation, altogether mitigating the overfitting to the surrogate model. +- We propose Momentum Token Embedding (MTE) to stabilize the updates of token embeddings by accumulating historical information, during the generation of adversarial examples on ViTs. +- Extensive experiments validate that our Forward Propagation Refinement (FPR), comprising AMD and MTE, outperforms the current best (backward) refinement method by up to an average of $7.0\%$ against 13 ViTs and CNNs with or without defenses. + +# 2. Related Work + +# 2.1. Surrogate Refinement Attacks + +In this subsection, we specifically delineate attack methods falling within the surrogate refinement, since it is the main topic of this work. Skip Gradient Method (SGM) [39] attenuates the attack by using a gradient descent approach. + +uates gradients traversing the residual modules to mitigate overfitting to the surrogate model. Ghost Networks [15] employs dropout in each block of the surrogate model and perturbs the skip connections to create a longitudinal ensemble. Backward Propagation Attack (BPA) [35] demonstrates that gradients traversing ReLU and max-pooling layers can carry inaccurate information. To address this, it adopts both non-monotonic function and softmax with temperature to recover the true gradients, thereby enhancing transferability. The aforementioned methods focus on the backward propagation of CNNs, while a few approaches also modify the output features of convolutional layers during forward propagation. Diversifying the High-level Features (DHF) [37] modifies the high-level features through random transformations. Clean Feature Mixup (CFM) [2] introduces competition into the generation of adversarial examples at the feature level for stronger transferability. Several methods also retrain the surrogate model to strengthen the transferability [14, 50], however, this may significantly increase the attack cost. + +The above surrogate refinements are primarily designed for CNNs; several methods have also been developed for ViTs. Pay No Attention and Patch Out (PNAPO) [38] comprises three modifications to synergistically mitigate overfitting to the surrogate model. First, it cuts off gradients traversing all attention maps. Second, it randomly updates a specific number of input patches. Finally, it adds an extra term to the cross-entropy loss function that increases the adversarial noise norm to enhance transferability. Token Gradient Regularization (TGR) [45] posits that extreme values in gradients are detrimental. Therefore, it removes the extreme gradients of QKV embeddings, attention maps, and MLP blocks to reduce gradient variance, thereby alleviating overfitting to the surrogate models. Gradient Normalization Scaling (GNS) [51] improves TGR by accurately normalizing and scaling the identified mild gradients, which yields stronger transferability. Even though the above methods exhibit the ability to craft transferable adversarial examples, they neglect the broader implications of forward propagation, which prevents further transferability. In this work, we focus on refining the ViT surrogate model from the new perspective of forward propagation and specifically explore two key components: attention maps and output token embeddings. + +# 2.2. Other Transfer-based Attacks + +This subsection reviews transfer-based attacks [48] that do not need to modify the surrogate models. Fast Gradient Sign Method (FGSM) [6] creates adversarial examples in a single step by following the gradient sign's direction. Basic Iterative Method (BIM) [12] extends FGSM by applying multiple iterations. Momentum Iterative Method (MIM) [4] introduces a momentum term for gradients at each qua + +Figure 2. The overview of Forward Propagation Refinement (FPR). Note that Attention Map Diversification (AMD) is applied to specific attention maps according to the index set, while Momentum Token Embedding (MTE) is applied to all Attention and MLP blocks. +![](images/0dfcd105905057697b20191c687d7809b65be43815a26601f3c9c632bc12782a.jpg) +$E_{t}^{A} / E_{t}^{M}$ : Output Token Embedding of Attention / MLP Block at $t$ -th Iteration $\widehat{E_{t}^{A}} / \widehat{E_{t}^{M}}$ : Accumulated Token Embedding of Attention / MLP Block at $t$ -th Iteration $A$ : Attention Map $A'$ : Diversified Attention Map $M$ : Diversity Matrix $Q/K/V$ : Query/Key/Value Component + +tion for better convergence. Nesterov Fast Gradient Sign Method (NIM) [17] takes an extra step at the beginning of each iteration and adopts the example at the farther location to compute the gradients for the update. Variance Tuning Method (VTM) [34] modifies current gradients with variance at the last iteration. Gradient Relevance Attack (GRA) [49] utilizes correlation information extracted from the nearby samples to stabilize the current update direction. Some input transformations have also been presented to enhance transferability. Diversity Input Method (DIM) [40] combines padding and resizing transformations on the input to reduce overfitting. Scale Invariance Method (SIM) [17] employs a group of scaled copies to enhance the input. Structure Invariant Attack (SIA) [36] employs different transformations from a library to each block of the input to prompt the transferability. Block Shuffle and Rotation (BSR) [32] divides the input image into several blocks, which are then randomly shuffled and rotated to create a set of new images for gradient computation. + +# 2.3. Vision Transformers + +Transformer [31] is widely used in natural language processing and has since gained increasing recognition in computer vision. ViT [5] processes images by dividing them into patches and embedding these patches as tokens. They employ a Transformer architecture with multi-head self-attention mechanisms to effectively capture spatial relationships. DeiT [28] improves ViT through teacher-student distillation, enabling superior performance with less data. Swin [18] employs a hierarchical structure with shifted win + +dows to efficiently handle multi-scale attention. PiT [9] utilizes pooling layers to streamline computational processes and manage data flow. CaiT [29] incorporates LayerScale to stabilize training as model depth increases and employs Class-Attention layers to enhance the extraction of classification information. Visformer [3] combines convolutional layers with ViT to effectively integrate spatial features and attention mechanisms. CoaT [43] merges convolution and self-attention in a hierarchical architecture to balance feature extraction and feature interaction. + +# 3. Methodology + +In this section, we propose the Forward Propagation Refinement (FPR) to enhance the transferability of adversarial examples from ViTs, consisting of Attention Map Diversification (AMD) and Momentum Token Embedding (MTE). Figure 2 provides an overview of FPR. + +# 3.1. Preliminary + +Assume an image $x$ with channel $C$ , height $H$ , width $W$ , the true label $y$ , and a surrogate model $F$ , the goal of a transfer-based attack is to add adversarial perturbation $\delta$ to the input $x$ , creating an adversarial example $x^{a} = x + \delta$ , which is then fed into the target model $F^{t}$ to induce incorrect predictions. This process can be formulated as solving the following problem: + +$$ +\delta^ {\prime} = \underset {\delta} {\arg \max } L (F (x + \delta), y), s. t. \| \delta \| _ {\infty} \leq \epsilon , \tag {1} +$$ + +![](images/bb504ed4017924f1ca447415eed07cc7cfc5dddbbcf0b7aae96ebd2c41a52ca5.jpg) +Figure 3. Illustration of redundancy in attention maps: Randomly dropping a small number of blocks leads to improved transferability. The results are averaged over 5 repeated runs on 13 target models, and the dash lines denote the baseline results of the MIM [4] without dropping. + +where $L$ is the loss function and $\epsilon$ is the maximum perturbation bound. + +Before the image $x$ is fed into ViTs, it is segmented into $N = \frac{H \cdot W}{P^2}$ patches, with each patch having a height, width, and channel dimension of $P$ , $P$ , and $C$ , respectively. Then, these patches are converted into token embeddings, and fed into a sequence of Transformer blocks. Each Transformer block includes an Attention block and a MLP block. + +The input token embedding of the Attention block is transformed into Query $(Q)$ , Key $(K)$ , and Value $(V)$ components. Here, the shapes of $Q$ , $K$ and $V$ are $N\times D$ , where $D$ is the embedding dimension. The Attention block employs the self-attention mechanism to establish the relationship between $Q$ , $K$ , and $V$ . The attention map $A$ in the Attention block is computed by: + +$$ +A = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {D}}\right), \tag {2} +$$ + +where the softmax function is employed at each row. The attention map $A$ is then used to weight the $V$ , and the output $Z$ is calculated as $Z = AV$ . + +The MLP block follows the Attention block and typically includes the fully connected layer, GELU activation function, dropout layer, and identity normalization layer to aggregate information for all token embeddings. + +# 3.2. Attention Map Diversification (AMD) + +The previous work [1] has demonstrated the redundancy in attention blocks of ViTs, showing that dropping a few such blocks does not significantly degrade classification accuracy. Therefore, it is reasonable to posit that such redundancy may also exist in generating adversarial examples, + +causing potential overfitting to the surrogate model. To verify this, we conduct exploratory experiments by dropping certain attention blocks. As can be seen from Figure 3, dropping a small number of blocks indeed leads to improved transferability (while dropping too many may harm transferability). + +To avoid such redundancy, a straightforward idea is to apply dropout of attention blocks while generating adversarial examples. However, we find that the dropout is too aggressive and may cause serious information loss, leading to only a small improvement of transferability (see comparing results in Table 5 of Section 4.3). Therefore, we propose a more moderate method called Attention Map Diversification (AMD), which transforms the original attention map $A$ by multiplying it with a diversity matrix $M$ . Specifically, the diversified attention map $A'$ is obtained by: + +$$ +A ^ {\prime} = A M D (A) = \operatorname {s o f t m a x} (A \odot M), \tag {3} +$$ + +where $\odot$ is the element-wise product and softmax function is applied to each row. The diversity matrix $M$ has the same shape as $A$ , and every element $M_{i,j}$ $(1 \leq i,j \leq N)$ conforms to the uniform distribution $U(1 - d, 1 + d)$ , where $d$ is a factor that controls the diversity level. Note that AMD is applied to attention maps at specified Transformer blocks according to a index set $I$ . + +The benefit of AMD on backward propagation. In principle, modifying the forward propagation would impact the subsequent backward propagation. Here, we theoretically demonstrate that the impact of our AMD on backward propagation can further improve transferability. Specifically, we find that AMD implicitly induces gradient vanishing at attention map positions during backward propagation, which aligns with a previous surrogate refinement, PNAPO [38], that directly truncates the gradients of attention maps during backpropagation. The underlying principle of PNAPO suggests that gradients passing through attention maps carry overfitting information from surrogate models, and attenuating these gradients can enhance the transferability of adversarial examples. Our findings can be explained as follows. + +When incorporating AMD, we let $S = A \odot M$ and the $i$ -th row vector of diversified attention map $A_{i}^{\prime}$ as: + +$$ +\begin{array}{l} A _ {i} ^ {\prime} = \left(A _ {i, 1} ^ {\prime}, A _ {i, 2} ^ {\prime}, \dots , A _ {i, N} ^ {\prime}\right) \\ = \operatorname {s o f t m a x} \left(A _ {i, 1} \cdot M _ {i, 1}, A _ {i, 2} \cdot M _ {i, 2},..., A _ {i, N} \cdot M _ {i, N}\right) \\ = \operatorname {s o f t m a x} \left(S _ {i, 1}, S _ {i, 2}, \dots , S _ {i, N}\right) \tag {4} \\ \end{array} +$$ + +and the derivative $\frac{\partial A_{i,m}^{\prime}}{\partial S_{i,n}}$ ( $1 \leq m, n \leq N$ ) is expressed by: + +$$ +\frac {\partial A _ {i , m} ^ {\prime}}{\partial S _ {i , n}} = \left\{ \begin{array}{l l} A _ {i, m} ^ {\prime} \cdot \left(1 - A _ {i, m} ^ {\prime}\right), & \text {i f} m = n \\ - A _ {i, m} ^ {\prime} \cdot A _ {i, n} ^ {\prime}, & \text {i f} m \neq n \end{array} \right. \tag {5} +$$ + +As the diversity factor $d$ becomes large, elements in the diversity matrix $M$ will fluctuate significantly, causing ex + +treme values in $S$ , and ultimately pushing the elements in $A'$ towards 0 or 1 due to the property of softmax function. From Eq. 5, whether the elements in $A'$ approach 0 or 1, the derivative will tend towards 0, resulting in gradient vanishing. According to the chain rule of differentiation, the gradients of the attention maps subjected to our AMD will experience varying degrees of gradient vanishing. + +Compared with PNAPO, which directly discards all gradients flowing through the attention maps, the gradient vanishing in our AMD is implicit without directly altering any aspect of the backward propagation. Furthermore, the gradient vanishing in our AMD method is more fine-grained, occurring at the row level of the attention map. Each row experiences varying degrees of vanishing, introducing additional randomness that helps to further mitigate overfitting to the surrogate model. + +# 3.3. Momentum Token Embedding (MTE) + +It is challenging to ensure convergence to the global optimum in the high-dimension space when generating adversarial examples. Typically, adversarial examples converge to a local optimum along an inaccurate path [4]. This causes the internal token embeddings of the surrogate model to contain inaccurate information during iterations, thereby weakening the transferability. Therefore, we borrow the idea of Momentum Distillation (MoD) [13] from the Vision-and-Language Pre-training (VLP) literature, where a momentum version of the model through the moving average of the primary model's parameters is maintained to mitigate the noise in the web-sourced training dataset. Similarly, we maintain the momentum version of token embedding updates to accumulate historical token embeddings while generating adversarial examples. + +We term this method Momentum Token Embedding (MTE), and formulate it as follows: + +$$ +\widehat {E _ {t}} = M T E \left(E _ {t}\right) = \eta \cdot \widehat {E _ {t - 1}} + s \cdot E _ {t}, \tag {6} +$$ + +where $t$ is the iteration index, $E_{t}$ is the output token embedding of any Attention block or MLP block at the $t$ -th iteration, $\widehat{E_t}$ is the corresponding accumulated token embedding at the $t$ -th iteration, $\eta$ is the attenuation factor, and $s$ is the scale factor. MTE is applied to all Attention and MLP blocks. Note that, different from MoD, our MTE does not require model training and can be seen as constructing a residual structure for output token embeddings across the time dimension. Previous work also finds the usefulness of the residual structure in improving transferability, but during backward propagation within each iteration [39]. + +# 3.4. Complementary AMD and MTE + +The previous subsections examine AMD and MTE from their respective perspectives. Here, we further discuss why they can collaboratively enhance transferability. + +# Algorithm 1 Forward Propagation Refinement (FPR) + +Input: A surrogate model $F$ with the cross-entropy loss $L$ , a clean image $x$ and its true label $y$ . The maximum perturbation bound $\epsilon$ , the iteration number $T$ , the decay factor $\mu$ , step size $\alpha$ , index set $I$ , diversity factor $d$ , scale factor $s$ and attenuation factor $\eta$ . + +Output: The adversarial image $x^{adv} = x_{T}$ . + +1: Initialize $g_0 = 0$ , $E_0 = 0$ , $x_0 = x$ . +2: for $t = 1 \rightarrow T$ do + +# During the forward propagation: + +3: Apply AMD (Eq. (3)) to attention maps in specific Transformer blocks according to the index set $I$ . +4: Apply MTE (Eq. (6)) to output token embeddings in all Transformer (Attention and MLP) blocks. + +# During the backward propagation: + +5: Calculate the gradient: $\nabla_{x_{t - 1}}L(F(x_{t - 1}),y)$ +6: Update momentum gradient: + +$$ +g _ {t} = \mu \cdot g _ {t - 1} + \frac {\nabla_ {x _ {t - 1}} L (F (x _ {t - 1}) , y)}{\left\| \nabla_ {x _ {t - 1}} L (F (x _ {t - 1}) , y) \right\| _ {1}} +$$ + +7: Update the adversarial example: + +$$ +x _ {t} = \operatorname {C l i p} _ {\epsilon} \left\{x _ {t - 1} + \alpha \cdot \operatorname {s i g n} \left(g _ {t}\right) \right\} +$$ + +# 8: end for + +Specifically, AMD and MTE refine the forward propagation of ViTs from two complementary angles. AMD falls under the category of "external refinement", which enhances surrogate models by introducing extraneous information or perturbing existing embeddings. Techniques such as attention map dropout also fall into this category. Our empirical results demonstrate that this type of refinement is relatively aggressive, making it unsuitable for concurrent use or for application across all blocks. In contrast, MTE belongs to the "internal refinement" category, mainly leveraging information from the input data itself. Since these refinements rely on self-related information, they are typically more moderate and can be applied to all blocks. Both AMD and MTE help promoting transferability but in different ways, this complementary effect further enhances the transferability of adversarial examples. + +Finally, we provide the pseudocode of FPR (based on MIM [4]) in Algorithm 1. Specifically, at each of $T$ iterations, we first apply AMD to specific attention maps according to the index set $I$ and MTE to the outputs of all Attention and MLP blocks during the forward propagation using PyTorch's Hook function. Then, during the backward propagation, we compute the gradient of the current adversarial example and use MIM [4] to update it. + +
ModelIdsη
ViT-B0,1,4,9,11250.80.3
CaiT-S2,14,25300.60.2
PiT-T1,6,11250.70.2
DeiT-B0,1,5,10,11300.70.2
+ +Table 1. The optimal parameter settings of our FPR about its index set $(I)$ , diversity factor $(d)$ , scale factor $(s)$ , and attenuation factor $(\eta)$ . Detailed ablation studies in Section 4.3 provide the sensitivity analysis of the attack performance towards the parameter selection under four cases. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets and baselines. Following the common practice [34, 45, 49], we use 1000 images randomly selected from the ILSVRC2012 [25] validation set. We compare our FPR with other surrogate refinements that focus on backward propagation: PNAPO [38], TGR [45], and GNS [51]. All these methods are built on MIM [4]. + +Models and defenses. For surrogate models, we consider four popular ViTs: ViT-Base (ViT-B) [5], CaiT-Small (CaiT-S) [29], PiT-Tiny (PiT-T) [9] and Deit-Base (Deit-B) [28]. For target models, we consider seven ViTs and six CNNs: ViT-B [5], CaiT-S [29], PiT-Base (PiT-B) [9], Visformer-Small (Vis-S) [3], Swin-Tiny (Swin-T) [18], DeiT-Tiny (DeiT-T) [28], and CoaT-Tiny (CoaT-T) [43], ResNet-18 (RN-18) [8], VGG-16 [26], DenseNet-121 (DN-121) [11], EfficientNet-b0 (EN-b0) [27], MobileNetv3 (MN-v3) [10], and ResNeXt-50 (RNX-50) [41]. Five popular defenses are tested: adversarial training (AT) paired with RN-50 [30], high-level representation guided denoiser (HGD) [16], neural representation purifier (NRP) [19], Bit depth reduction (BDR) [42], and JPEG compression [7]. Finally, two online models, i.e., Baidu Cloud API and Aliyun API, are also included. + +Hyperparameter settings. Commonly, we set the number of iterations as 10, the maximum perturbation bound $\epsilon$ as 16, the step size $\alpha$ as 1.6, and the decay factor of momentum gradient $\mu = 1$ . Parameters of all baselines are the same as their default values. For our FPR, the optimal hyperparameters are listed in Table 1, with a detailed analysis in Section 4.3 discussing their sensitivity to different parameter choices. All experiments are conducted on a single RTX 4060 GPU with 8 GB of VRAM. + +# 4.2. Experimental Results + +Different target models. We first compare FPR with baselines on transferring adversarial examples from ViTs to different ViTs and CNNs. Table 2 shows that FPR consistently outperforms baselines in all cases, suggesting the potential + +of refining the surrogate model from the perspective of forward propagation. For example, when generating adversarial examples on ViT-B, FPR achieves an average attack success rate of $84.3\%$ on six CNNs, surpassing GNS, TGR, and PNAPO by $7.0\%$ , $7.4\%$ , and $15.5\%$ , respectively, and improves the baseline MIM by $33.4\%$ . + +Different defenses. We then evaluate different attack methods against five popular defenses. As shown in Table 3, FPR maintains its superiority over existing surrogate refinements in such challenging defense settings. For instance, when generating adversarial examples on ViT-B, FPR achieves an average attack success rate of $70.8\%$ across five defenses, outperforming other surrogate refinements GNS, TGR, and PNAPO by $5.7\%$ , $6.4\%$ , and $12.4\%$ , respectively, while surpassing the vanilla baseline, MIM, by $24.3\%$ . + +Incorporating other transfer methods. A good surrogate refinement-based transfer method should also exhibit strong compatibility with other (types of) transfer-based attack methods. In this study, we integrate FPR into two advanced gradient correction-based transferable attacks, VTM [34] and GRA [49], as well as two latest input augmentation-based transferable attacks, BSR [32] and SIA [36]. Specifically, we equip these four attacks with 5 augmented copies per iteration, while keeping other parameter settings as their default values. Results in Table 4 indicate that FPR can effectively complement existing transfer-based attacks for stronger transferability. For instance, when combined with our FPR, the attack success rate of VTM increases from $57.3\%$ to $81.7\%$ on CNNs. + +Online models. We also evaluate our FPR on two online models, i.e., Baidu Cloud API and Aliyun API. The results, presented in Appendix A, show that FPR still outperforms both TGR and GNS. + +# 4.3. Ablation Studies + +Individual components. We compare each component of FPR with the plain baseline MIM, and the two best-performed surrogate refinements, TGR and GNS. We also consider Dropout as a comparison to our AMD. From Table 5, we can see that combining our AMD and MTE yields the best transferability results. In addition, Dropout is less effective than our AMD, indicating that directly removing certain attention maps is rather aggressive and fails to retain useful information. + +Parameter sensitivity. Figure 4 shows the attack success rates when varying three main parameters: diversity factor $d$ , scale factor $s$ , and attenuation factor $\eta$ . $d$ controls the diversity of the attention map. FPR achieves satisfactory performance when $d$ is between 25 and 30. $s$ represents the importance of current token embeddings. FPR achieves satisfactory performance when $s$ is between 0.6 and 0.8. $\eta$ signifies the importance of historical token embeddings. FPR achieves satisfactory performance when $\eta$ is between 0.2 + +
ModelAttackViT-BCaiT-SPiT-BVis-SSwin-TDeiT-TCoaT-TRN-18VGG-16DN-121EN-b0MN-v3RNX-50Average
ViT-BMIM [4]97.261.840.442.354.965.844.451.757.251.350.851.738.250.9
PNAPO [38]99.183.262.165.874.783.064.067.470.067.668.563.056.368.8
TGR [45]99.186.263.869.981.194.470.479.179.875.383.179.960.076.9
GNS [51]99.989.568.174.082.493.672.277.876.776.079.176.062.077.3
FPR99.292.573.078.388.598.277.787.387.682.090.387.468.384.3
CaiT-SMIM [4]68.598.850.054.968.275.354.861.067.859.861.055.747.160.3
PNAPO [38]69.888.258.161.070.373.860.164.869.764.365.561.753.264.4
TGR [45]90.1100.076.081.391.297.880.686.986.684.387.782.169.184.5
GNS [51]91.2100.078.682.291.697.681.186.687.184.989.681.970.885.3
FPR93.8100.081.685.695.098.886.887.488.788.690.682.976.888.1
PiT-TMIM [4]30.036.138.543.351.075.055.569.568.760.864.571.640.454.2
PNAPO [38]39.550.949.654.961.685.465.078.776.772.277.981.050.864.9
TGR [45]37.947.746.954.162.587.465.182.882.075.581.888.853.766.6
GNS [51]37.144.146.451.959.484.463.379.978.071.576.384.348.663.5
FPR44.356.157.162.469.591.372.288.786.480.487.691.160.372.9
DeiT-BMIM [4]86.082.467.069.579.787.765.369.070.570.171.461.458.272.2
PNAPO [38]83.487.772.474.182.286.071.974.974.774.976.168.066.376.4
TGR [45]92.497.181.284.392.697.983.586.085.285.089.984.574.687.2
GNS [51]92.198.080.584.291.098.079.883.380.783.586.781.370.885.4
FPR94.698.483.886.795.599.587.089.889.288.193.986.879.890.2
+ +Table 2. Attack success rates (%) of different surrogate refinements on various ViTs and CNNs. The "Average" column excludes the white-box results with grey background. + +
ModelMethodATHGDNRPJPEGBitAverage
ViT-BMIM [4]43.929.842.157.559.146.5
PNAPO [38]45.647.553.570.974.858.4
TGR [45]49.750.861.677.782.064.4
GNS [51]49.254.861.377.982.365.1
FPR50.761.469.284.588.470.8
CaiT-SMIM [4]44.337.446.264.166.951.8
PNAPO [38]44.843.748.265.267.353.8
TGR [45]49.860.766.481.286.869.0
GNS [51]49.862.767.482.688.070.1
FPR51.168.970.886.290.473.5
PiT-TMIM [4]46.132.042.057.562.047.9
PNAPO [38]47.542.948.565.569.854.8
TGR [45]50.844.050.568.271.957.1
GNS [45]48.938.649.365.368.854.2
FPR51.449.956.173.277.461.6
DeiT-BMIM [4]44.748.152.270.374.958.1
PNAPO [38]46.957.257.073.678.762.7
TGR [45]50.265.267.882.288.170.7
GNS [51]48.361.665.780.686.368.5
FPR50.270.771.387.091.574.1
+ +and 0.3. Detailed results of applying the same index set $(I)$ to other methods and the sensitivity of the $I$ used in AMD are reported in Appendix B, where we also provide insights for identifying the optimal $I$ . + +Gradient vanishing comparison. We compare the extent of gradient vanishing in Table 6. We can conclude that gradient vanishing can improve transferability, but it is not a simple linear correlation. This aligns with our claim that the key of AMD is to diversify attention maps in the forward propagation, while it implicitly and mildly induces some + +Table 3. Attack success rates (\%) of different surrogate refinements against various defenses. AT is trained with RN-50; HGD is based on Inc-v3; NRP, JPEG and Bit are averaged on seven ViTs and six CNNs. + +
MethodViTsCNNs
SIA [36]73.469.1
SIA+FPR86.690.4
BSR [32]69.466.9
BSR+FPR84.889.9
VTM [34]63.157.3
VTM+FPR84.081.7
GRA [49]73.668.3
GRA+FPR86.185.3
+ +Table 4. Attack success rates $(\%)$ of combined methods averaged on six CNNs and six ViTs (without surrogate model ViT-B). + +
MethodViTsCNNs
MIM [4]51.650.2
Dropout55.858.9
TGR [45]77.676.2
GNS [51]80.074.6
AMD82.178.4
MTE78.479.6
FPR (AMD+MTE)84.783.8
+ +Table 5. Ablation studies of different FPR components with ViT-B as the surrogate model. All involved methods are built on MIM. Dropout is applied at the same index set as AMD. + +gradient vanishing in the backward propagation. + +![](images/618e5a8c374e4fda37b1ad913610fe6953af7a7a52521ad10689397e7c301fe8.jpg) +Figure 4. Sensitivity analysis about diversity factor, scale factor, and attenuation factor, where ASR remains relatively stable within a certain range of values for these three factors. + +
AttackMIM [4]PNA [38]TGR [45]GNS [51]AMD
Mean2.9 × 10-40.03.9 × 10-102.4 × 10-151.5 × 10-5
ASR50.960.663.262.780.3
+ +Table 6. Mean of attention map's input gradient elements from 100 images over 10 iterations. All surrogate refinements adopt Index Set (0,1,4,9,11). ASR is averaged on 13 models and ViT-B is the surrogate model. + +# 5. Conclusion + +In this work, we propose the Forward Propagation Refinement (FPR) to generate highly transferable adversarial examples on Vision Transformers (ViTs). FPR is the first work to refine the surrogate ViTs during forward propagation entirely. Specifically, FPR consists of two parts: Attention Map Diversification (AMD) and Momentum To + +ken Embedding (MTE). AMD diversifies the attention maps and implicitly causes gradient vanishing to enhance transferability, while MTE accumulates historical token embeddings to stabilize current embeddings for more accurate updates. The superiority of FPR is supported by both empirical results and theoretical analyses. + +# Acknowledgments + +This research is supported by the National Key Research and Development Program of China (2023YFE0209800), the National Natural Science Foundation of China (62406240, 62376210, 62132011, 62206217, T2341003, U244120060, 62161160337, U2441240, U24B20185, U21B2018), the Shaanxi Province Key Industry Innovation Program (2023-ZDLGY-38). Thanks to the New Cornerstone Science Foundation and the Xplorer Prize. + +# References + +[1] Srinadh Bhojanapalli, Ayan Chakrabarti, Daniel Glasner, Daliang Li, Thomas Unterthiner, and Andreas Veit. Understanding robustness of transformers for image classification. In IEEE International Conference on Computer Vision, pages 10211-10221, 2021. 4 +[2] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In IEEE Conference on Computer Vision and Pattern Recognition, pages 24648-24657, 2023. 2 +[3] Zhengsu Chen, Lingxi Xie, Jianwei Niu, Xuefeng Liu, Longhui Wei, and Qi Tian. Visformer: The vision-friendly transformer. In IEEE International Conference on Computer Vision, pages 569-578, 2021. 3, 6 +[4] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 1, 2, 4, 5, 6, 7, 8 +[5] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, and et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 3, 6 +[6] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. 1, 2 +[7] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering adversarial images using input transformations. In International Conference on Learning Representations, 2018. 6 +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6 +[9] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In IEEE International Conference on Computer Vision, pages 11916-11925, 2021. 1, 3, 6 +[10] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint, arXiv:1704.04861, 2017. 6 +[11] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2261-2269, 2017. 6 +[12] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In International Conference on Learning Representations, 2017. 2 +[13] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural In + +formation Processing Systems, pages 9694-9705, 2021. 2, 5 +[14] Qizhang Li, Yiwen Guo, Wangmeng Zuo, and Hao Chen. Making substitute models more bayesian can enhance transferability of adversarial examples. In International Conference on Learning Representations, 2023. 2 +[15] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan Yuille. Learning transferable adversarial examples via ghost networks. In AAAI Conference on Artificial Intelligence, pages 11458-11465, 2020. 1, 2 +[16] Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial attacks using high-level representation guided denoiser. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2018. 6 +[17] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, 2019. 3 +[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE International Conference on Computer Vision, pages 9992-10002, 2021. 3, 6 +[19] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A self-supervised approach for adversarial robustness. In IEEE Conference on Computer Vision and Pattern Recognition, pages 259-268, 2020. 6 +[20] Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Fahad Shahbaz Khan, and Fatih Porikli. On improving adversarial transferability of vision transformers. In International Conference on Learning Representations, 2022. 1 +[21] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In Advances in Neural Information Processing Systems, pages 23296-23308, 2021. 1 +[22] Yuchen Ren, Hegui Zhu, Xiaoyan Sui, and Chong Liu. Crafting transferable adversarial examples via contaminating the salient feature variance. Information Sciences, 644:119273, 2023. 1 +[23] Yuchen Ren, Hegui Zhu, Chong Liu, and Chengqing Li. Efficient polar coordinates attack with adaptive activation strategy. Expert Systems with Applications, 249:123850, 2024. 1 +[24] Yuchen Ren, Zhengyu Zhao, Chenhao Lin, Bo Yang, Lu Zhou, Zhe Liu, and Chao Shen. Improving integrated gradient-based transferable adversarial examples by refining the integration path. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 1 +[25] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. International Journal of Computer Vision, 115(3): 211–252, 2015. 6 + +[26] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. 6 +[27] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, 2020. 6 +[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357, 2021. 1, 3, 6 +[29] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Herve Jegou. Going deeper with image transformers. In IEEE International Conference on Computer Vision, pages 32-42, 2021. 3, 6 +[30] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 6 +[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, and et al. Attention is all you need. In International Conference on Neural Information Processing Systems, pages 6000-6010, 2017. 3 +[32] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting adversarial transferability by block shuffle and rotation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 24336-24346, 2024. 3, 6, 7 +[33] Kang Wang, Jianzhou Wang, Zhiwu Li, and Yilin Zhou. An integrated sparse gated graph density network based on transfer learning for multi-site probabilistic forecasting of renewable energy. IEEE Transactions on Sustainable Energy, 16(1):673-685, 2025. 1 +[34] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1924-1933, 2021. 1, 3, 6, 7 +[35] Xiaosen Wang, Kangheng Tong, and Kun He. Rethinking the backward propagation for adversarial transferability. In Advances in Neural Information Processing Systems, pages 1905-1922, 2023. 1, 2 +[36] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In IEEE International Conference on Computer Vision, pages 4584-4596, 2023. 3, 6, 7 +[37] Zhiyuan Wang, Zeliang Zhang, Siyuan Liang, and Xiaosen Wang. Diversifying the high-level features for better adversarial transferability. In British Machine Vision Conference 2023, BMVC 2023, 2023. 2 +[38] Zhipeng Wei, Jingjing Chen, Micah Goldblum, Zuxuan Wu, Tom Goldstein, and Yu-Gang Jiang. Towards transferable adversarial attacks on vision transformers. In AAAI Conference on Artificial Intelligence, pages 2668-2676, 2022. 1, 2, 4, 6, 7, 8 +[39] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip connections matter: On the transferability of adversarial examples generated with resnets. In Interna- + +tional Conference on Learning Representations, 2020. 1, 2, 5 +[40] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 3 +[41] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5987-5995, 2017. 1, 6 +[42] Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In Network and Distributed System Security Symposium, 2018. 6 +[43] Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Coscale conv-attentional image transformers. In IEEE International Conference on Computer Vision, pages 9961–9970, 2021. 3, 6 +[44] Bo Yang, Hengwei Zhang, Jindong Wang, Yulong Yang, Chenhao Lin, Chao Shen, and Zhengyu Zhao. Adversarial example soups: Improving transferability and stealthiness for free. IEEE Transactions on Information Forensics and Security, 20:1882-1894, 2025. 1 +[45] Jianping Zhang, Yizhan Huang, Weibin Wu, and Michael R. Lyu. Transferable adversarial attacks on vision transformers with token gradient regularization. In IEEE Conference on Computer Vision and Pattern Recognition, pages 16415-16424, 2023. 1, 2, 6, 7, 8 +[46] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. Towards large yet imperceptible adversarial image perturbations with perceptual color distance. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1036-1045, 2020. 1 +[47] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. On success and simplicity: A second look at transferable targeted attacks. In Advances in Neural Information Processing Systems, pages 6115-6128, 2021. 1 +[48] Zhengyu Zhao, Hanwei Zhang, Renjue Li, Ronan Sicre, Laurent Amsaleg, Michael Backes, Qi Li, and Chao Shen. Revisiting transferable adversarial image examples: Attack categorization, evaluation guidelines, and new insights. arXiv preprint, arXiv:2310.11850, 2023. 2 +[49] Hegui Zhu, Yuchen Ren, Xiaoyan Sui, Lianping Yang, and Wuming Jiang. Boosting adversarial transferability via gradient relevance attack. In IEEE International Conference on Computer Vision, pages 4718-4727, 2023. 3, 6, 7 +[50] Yao Zhu, Yuefeng Chen, Xiaodan Li, Kejiang Chen, Yuan He, Xiang Tian, Bolun Zheng, Yaowu Chen, and Qingming Huang. Toward understanding and boosting adversarial transferability from a distribution perspective. IEEE Transactions on Image Processing, 31:6487-6501, 2022. 2 +[51] Zhiyu Zhu, Xinyi Wang, Zhibo Jin, Jiayu Zhang, and Huaming Chen. Enhancing transferable adversarial attacks on vision transformers through gradient normalization scaling and high-frequency adaptation. In International Conference on Learning Representations, 2024. 2, 6, 7, 8 \ No newline at end of file diff --git a/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/images.zip b/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c9ba2922d966aba0e3d88eb8a332f898ccdbb109 --- /dev/null +++ b/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:651827bf1cfd20fc1e4688954f2e8e7e1e6759ac9a59441e72eb90c8cfbe7185 +size 596651 diff --git a/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/layout.json b/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4067d19a48d3e1d637f1ad3d6c0583b00bde8947 --- /dev/null +++ b/2025/Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement/layout.json @@ -0,0 +1,9226 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "type": "text", + "content": "Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "spans": [ + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "text", + "content": "Yuchen Ren" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "text", + "content": ", Zhengyu Zhao" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "text", + "content": ", Chenhao Lin" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "text", + "content": ", Bo Yang" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "text", + "content": ", Lu Zhou" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "text", + "content": ", Zhe Liu" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "text", + "content": ", Chao Shen" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "text", + "content": "Xi'an Jiaotong University, China; " + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "text", + "content": "Information Engineering University, China; " + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "text", + "content": "Nanjing University of Aeronautics and Astronautics, China; " + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 87, + 160, + 523, + 204 + ], + "type": "text", + "content": "Zhejiang Lab, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 205, + 508, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 205, + 508, + 232 + ], + "spans": [ + { + "bbox": [ + 99, + 205, + 508, + 232 + ], + "type": "text", + "content": "ryc98@stu.xjtu.edu.cn, {zhengyu.zhao, linchenhao}@xjtu.edu.cn, yangbo_hn@163.com, {lu.zhou, zhe.liu}@nuaa.edu.cn, chaoshen@mail.xjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 285, + 297, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 297, + 585 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 297, + 585 + ], + "type": "text", + "content": "Vision Transformers (ViTs) have been widely applied in various computer vision and vision-language tasks. To gain insights into their robustness in practical scenarios, transferable adversarial examples on ViTs have been extensively studied. A typical approach to improving adversarial transferability is by refining the surrogate model. However, existing work on ViTs has restricted their surrogate refinement to backward propagation. In this work, we instead focus on Forward Propagation Refinement (FPR) and specifically refine two key modules of ViTs: attention maps and token embeddings. For attention maps, we propose Attention Map Diversification (AMD), which diversifies certain attention maps and also implicitly imposes beneficial gradient vanishing during backward propagation. For token embeddings, we propose Momentum Token Embedding (MTE), which accumulates historical token embeddings to stabilize the forward updates in both the Attention and MLP blocks. We conduct extensive experiments with adversarial examples transferred from ViTs to various CNNs and ViTs, demonstrating that our FPR outperforms the current best (backward) surrogate refinement by up to " + }, + { + "bbox": [ + 54, + 285, + 297, + 585 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 54, + 285, + 297, + 585 + ], + "type": "text", + "content": " on average. We also validate its superiority against popular defenses and its compatibility with other transfer methods. Codes and appendix are available at https://github.com/RYC-98/FPR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 610, + 137, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 137, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 137, + 623 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 631, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 631, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 631, + 296, + 693 + ], + "type": "text", + "content": "Deep neural networks (DNNs) [8, 33, 41] demonstrate remarkable performance across diverse domains, with Vision Transformers (ViTs) [5, 9, 28] excelling particularly in computer visualization. However, ViTs are also shown to be vulnerable to adversarial examples [20, 21], which are" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 258, + 549, + 427 + ], + "blocks": [ + { + "bbox": [ + 320, + 258, + 549, + 427 + ], + "lines": [ + { + "bbox": [ + 320, + 258, + 549, + 427 + ], + "spans": [ + { + "bbox": [ + 320, + 258, + 549, + 427 + ], + "type": "image", + "image_path": "4838ca372ca55753d60cd7d0f8a80a47e97bf880a971991fb1cc392827d28f16.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 435, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 460 + ], + "type": "text", + "content": "Figure 1. Our Forward Propagation Refinement (FPR) vs. previous methods which only refine the backward propagation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "type": "text", + "content": "optimized by adding subtle perturbations [23, 46] into clean images [6, 44, 47]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 508, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 616 + ], + "type": "text", + "content": "A key property of adversarial examples that causes practical security concerns is their transferability, i.e., they can be optimized on a (seen) surrogate model and then directly transfer their effects to (unseen) target models [4, 40]. Compared to other transfer approaches that focus on loss design, optimization algorithm or data augmentation [22, 24, 34], surrogate refinement [15, 35, 39] explores the intrinsic relationship between the model itself and adversarial transferability, offering a unique perspective." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "content": "However, existing surrogate refinements on ViTs [38, 45] are limited to refining the backward propagation, i.e., only modifying the gradient information. To address the above limitation, in this work, we explore the potential of Forward Propagation Refinement (FPR), i.e., modifying features during forward propagation. Our FPR not only provides a new perspective of surrogate refinement but also potentially benefits the refinement during the sub" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25071" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "content": "sequent backward propagation (see our discussion in Section 3.2). Specifically, FPR comprises two parts that refine two key components of ViTs respectively: attention maps and output token embeddings. Figure 1 highlights the difference between our forward-based method and the previous backward-based methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 145, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 294, + 300 + ], + "type": "text", + "content": "For attention maps, we identify that certain attention maps in ViTs are redundant and can deteriorate transferability. Rather than simply dropping redundant attention maps, which recklessly removes all information, we propose Attention Map Diversification (AMD), which moderately diversifies the attention maps based on random weighting. The introduced randomness on attention maps acts as a simulation of unknown models, thereby mitigating the overfitting to the surrogate model. Although AMD only modifies the forward propagation, we find that it can potentially induce the vanishing of certain gradients flowing through the attention maps during backward propagation, further alleviating overfitting [38]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 301, + 294, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 294, + 409 + ], + "type": "text", + "content": "For output token embeddings, we posit that adversarial examples during generation are highly likely to be near the local optima, resulting in the surrogate model's embeddings containing erroneous information. Consequently, inspired by the momentum distillation technology when training the vision-language pretraining model [13], we propose Momentum Token Embedding (MTE), to stabilize the updates of token embeddings in each iteration by accumulating historical token embedding information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 410, + 274, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 410, + 274, + 421 + ], + "spans": [ + { + "bbox": [ + 67, + 410, + 274, + 421 + ], + "type": "text", + "content": "In sum, we make the following main contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 422, + 295, + 624 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 422, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 458 + ], + "type": "text", + "content": "- We explore the refinement of ViT surrogate models from the perspective of forward propagation, particularly modifying the attention maps and output token embeddings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 459, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 518 + ], + "type": "text", + "content": "- We propose Attention Map Diversification (AMD) to moderately avoid attention redundancy, which also induces beneficial gradient vanishing during the subsequent backward propagation, altogether mitigating the overfitting to the surrogate model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "type": "text", + "content": "- We propose Momentum Token Embedding (MTE) to stabilize the updates of token embeddings by accumulating historical information, during the generation of adversarial examples on ViTs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 566, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 624 + ], + "type": "text", + "content": "- Extensive experiments validate that our Forward Propagation Refinement (FPR), comprising AMD and MTE, outperforms the current best (backward) refinement method by up to an average of " + }, + { + "bbox": [ + 55, + 566, + 295, + 624 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 55, + 566, + 295, + 624 + ], + "type": "text", + "content": " against 13 ViTs and CNNs with or without defenses." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 639, + 142, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 142, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 142, + 651 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 659, + 220, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 220, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 220, + 672 + ], + "type": "text", + "content": "2.1. Surrogate Refinement Attacks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "In this subsection, we specifically delineate attack methods falling within the surrogate refinement, since it is the main topic of this work. Skip Gradient Method (SGM) [39] attenuates the attack by using a gradient descent approach." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "content": "uates gradients traversing the residual modules to mitigate overfitting to the surrogate model. Ghost Networks [15] employs dropout in each block of the surrogate model and perturbs the skip connections to create a longitudinal ensemble. Backward Propagation Attack (BPA) [35] demonstrates that gradients traversing ReLU and max-pooling layers can carry inaccurate information. To address this, it adopts both non-monotonic function and softmax with temperature to recover the true gradients, thereby enhancing transferability. The aforementioned methods focus on the backward propagation of CNNs, while a few approaches also modify the output features of convolutional layers during forward propagation. Diversifying the High-level Features (DHF) [37] modifies the high-level features through random transformations. Clean Feature Mixup (CFM) [2] introduces competition into the generation of adversarial examples at the feature level for stronger transferability. Several methods also retrain the surrogate model to strengthen the transferability [14, 50], however, this may significantly increase the attack cost." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 312, + 553, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 600 + ], + "type": "text", + "content": "The above surrogate refinements are primarily designed for CNNs; several methods have also been developed for ViTs. Pay No Attention and Patch Out (PNAPO) [38] comprises three modifications to synergistically mitigate overfitting to the surrogate model. First, it cuts off gradients traversing all attention maps. Second, it randomly updates a specific number of input patches. Finally, it adds an extra term to the cross-entropy loss function that increases the adversarial noise norm to enhance transferability. Token Gradient Regularization (TGR) [45] posits that extreme values in gradients are detrimental. Therefore, it removes the extreme gradients of QKV embeddings, attention maps, and MLP blocks to reduce gradient variance, thereby alleviating overfitting to the surrogate models. Gradient Normalization Scaling (GNS) [51] improves TGR by accurately normalizing and scaling the identified mild gradients, which yields stronger transferability. Even though the above methods exhibit the ability to craft transferable adversarial examples, they neglect the broader implications of forward propagation, which prevents further transferability. In this work, we focus on refining the ViT surrogate model from the new perspective of forward propagation and specifically explore two key components: attention maps and output token embeddings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 611, + 477, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 477, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 477, + 623 + ], + "type": "text", + "content": "2.2. Other Transfer-based Attacks" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "This subsection reviews transfer-based attacks [48] that do not need to modify the surrogate models. Fast Gradient Sign Method (FGSM) [6] creates adversarial examples in a single step by following the gradient sign's direction. Basic Iterative Method (BIM) [12] extends FGSM by applying multiple iterations. Momentum Iterative Method (MIM) [4] introduces a momentum term for gradients at each qua" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25072" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 531, + 246 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 531, + 246 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 531, + 246 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 531, + 246 + ], + "type": "image", + "image_path": "0dfcd105905057697b20191c687d7809b65be43815a26601f3c9c632bc12782a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "lines": [ + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "spans": [ + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "inline_equation", + "content": "E_{t}^{A} / E_{t}^{M}" + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "text", + "content": ": Output Token Embedding of Attention / MLP Block at " + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "text", + "content": "-th Iteration " + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "inline_equation", + "content": "\\widehat{E_{t}^{A}} / \\widehat{E_{t}^{M}}" + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "text", + "content": ": Accumulated Token Embedding of Attention / MLP Block at " + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "text", + "content": "-th Iteration " + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "text", + "content": ": Attention Map " + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "inline_equation", + "content": "A'" + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "text", + "content": ": Diversified Attention Map " + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "text", + "content": ": Diversity Matrix " + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "inline_equation", + "content": "Q/K/V" + }, + { + "bbox": [ + 82, + 255, + 526, + 299 + ], + "type": "text", + "content": ": Query/Key/Value Component" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 310, + 555, + 334 + ], + "lines": [ + { + "bbox": [ + 55, + 310, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 555, + 334 + ], + "type": "text", + "content": "Figure 2. The overview of Forward Propagation Refinement (FPR). Note that Attention Map Diversification (AMD) is applied to specific attention maps according to the index set, while Momentum Token Embedding (MTE) is applied to all Attention and MLP blocks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 354, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 582 + ], + "type": "text", + "content": "tion for better convergence. Nesterov Fast Gradient Sign Method (NIM) [17] takes an extra step at the beginning of each iteration and adopts the example at the farther location to compute the gradients for the update. Variance Tuning Method (VTM) [34] modifies current gradients with variance at the last iteration. Gradient Relevance Attack (GRA) [49] utilizes correlation information extracted from the nearby samples to stabilize the current update direction. Some input transformations have also been presented to enhance transferability. Diversity Input Method (DIM) [40] combines padding and resizing transformations on the input to reduce overfitting. Scale Invariance Method (SIM) [17] employs a group of scaled copies to enhance the input. Structure Invariant Attack (SIA) [36] employs different transformations from a library to each block of the input to prompt the transferability. Block Shuffle and Rotation (BSR) [32] divides the input image into several blocks, which are then randomly shuffled and rotated to create a set of new images for gradient computation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 588, + 175, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 175, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 175, + 600 + ], + "type": "text", + "content": "2.3. Vision Transformers" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Transformer [31] is widely used in natural language processing and has since gained increasing recognition in computer vision. ViT [5] processes images by dividing them into patches and embedding these patches as tokens. They employ a Transformer architecture with multi-head self-attention mechanisms to effectively capture spatial relationships. DeiT [28] improves ViT through teacher-student distillation, enabling superior performance with less data. Swin [18] employs a hierarchical structure with shifted win" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 354, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 473 + ], + "type": "text", + "content": "dows to efficiently handle multi-scale attention. PiT [9] utilizes pooling layers to streamline computational processes and manage data flow. CaiT [29] incorporates LayerScale to stabilize training as model depth increases and employs Class-Attention layers to enhance the extraction of classification information. Visformer [3] combines convolutional layers with ViT to effectively integrate spatial features and attention mechanisms. CoaT [43] merges convolution and self-attention in a hierarchical architecture to balance feature extraction and feature interaction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 485, + 397, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 397, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 397, + 498 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 505, + 554, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 554, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 554, + 565 + ], + "type": "text", + "content": "In this section, we propose the Forward Propagation Refinement (FPR) to enhance the transferability of adversarial examples from ViTs, consisting of Attention Map Diversification (AMD) and Momentum Token Embedding (MTE). Figure 2 provides an overview of FPR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 573, + 393, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 393, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 393, + 586 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "text", + "content": "Assume an image " + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "text", + "content": " with channel " + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "text", + "content": ", height " + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "text", + "content": ", width " + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "text", + "content": ", the true label " + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "text", + "content": ", and a surrogate model " + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "text", + "content": ", the goal of a transfer-based attack is to add adversarial perturbation " + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "text", + "content": " to the input " + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "text", + "content": ", creating an adversarial example " + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "inline_equation", + "content": "x^{a} = x + \\delta" + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "text", + "content": ", which is then fed into the target model " + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "inline_equation", + "content": "F^{t}" + }, + { + "bbox": [ + 313, + 590, + 555, + 675 + ], + "type": "text", + "content": " to induce incorrect predictions. This process can be formulated as solving the following problem:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 696, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 696, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 339, + 696, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\delta^ {\\prime} = \\underset {\\delta} {\\arg \\max } L (F (x + \\delta), y), s. t. \\| \\delta \\| _ {\\infty} \\leq \\epsilon , \\tag {1}", + "image_path": "e426594da49bd229e6eff502c58c91dbbfd7fd4c3948b7766d456249d0910063.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25073" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 70, + 272, + 245 + ], + "blocks": [ + { + "bbox": [ + 81, + 70, + 272, + 245 + ], + "lines": [ + { + "bbox": [ + 81, + 70, + 272, + 245 + ], + "spans": [ + { + "bbox": [ + 81, + 70, + 272, + 245 + ], + "type": "image", + "image_path": "bb504ed4017924f1ca447415eed07cc7cfc5dddbbcf0b7aae96ebd2c41a52ca5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 253, + 296, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 309 + ], + "type": "text", + "content": "Figure 3. Illustration of redundancy in attention maps: Randomly dropping a small number of blocks leads to improved transferability. The results are averaged over 5 repeated runs on 13 target models, and the dash lines denote the baseline results of the MIM [4] without dropping." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 328, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 351 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 328, + 295, + 351 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 328, + 295, + 351 + ], + "type": "text", + "content": " is the loss function and " + }, + { + "bbox": [ + 55, + 328, + 295, + 351 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 328, + 295, + 351 + ], + "type": "text", + "content": " is the maximum perturbation bound." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "text", + "content": "Before the image " + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "text", + "content": " is fed into ViTs, it is segmented into " + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "inline_equation", + "content": "N = \\frac{H \\cdot W}{P^2}" + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "text", + "content": " patches, with each patch having a height, width, and channel dimension of " + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "text", + "content": ", respectively. Then, these patches are converted into token embeddings, and fed into a sequence of Transformer blocks. Each Transformer block includes an Attention block and a MLP block." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "content": "The input token embedding of the Attention block is transformed into Query " + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "content": ", Key " + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "content": ", and Value " + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "content": " components. Here, the shapes of " + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "inline_equation", + "content": "N\\times D" + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "content": " is the embedding dimension. The Attention block employs the self-attention mechanism to establish the relationship between " + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "content": ". The attention map " + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 424, + 296, + 509 + ], + "type": "text", + "content": " in the Attention block is computed by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 128, + 513, + 295, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 513, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 128, + 513, + 295, + 541 + ], + "type": "interline_equation", + "content": "A = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {D}}\\right), \\tag {2}", + "image_path": "658999dc83f01ff1e33791f7bf2a3f7f30798a028480752e2ff5816edc5c9bc7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "type": "text", + "content": "where the softmax function is employed at each row. The attention map " + }, + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "type": "text", + "content": " is then used to weight the " + }, + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "type": "text", + "content": ", and the output " + }, + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "type": "text", + "content": " is calculated as " + }, + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "type": "inline_equation", + "content": "Z = AV" + }, + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 631 + ], + "type": "text", + "content": "The MLP block follows the Attention block and typically includes the fully connected layer, GELU activation function, dropout layer, and identity normalization layer to aggregate information for all token embeddings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 635, + 254, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 254, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 254, + 649 + ], + "type": "text", + "content": "3.2. Attention Map Diversification (AMD)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "The previous work [1] has demonstrated the redundancy in attention blocks of ViTs, showing that dropping a few such blocks does not significantly degrade classification accuracy. Therefore, it is reasonable to posit that such redundancy may also exist in generating adversarial examples," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "causing potential overfitting to the surrogate model. To verify this, we conduct exploratory experiments by dropping certain attention blocks. As can be seen from Figure 3, dropping a small number of blocks indeed leads to improved transferability (while dropping too many may harm transferability)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "text", + "content": "To avoid such redundancy, a straightforward idea is to apply dropout of attention blocks while generating adversarial examples. However, we find that the dropout is too aggressive and may cause serious information loss, leading to only a small improvement of transferability (see comparing results in Table 5 of Section 4.3). Therefore, we propose a more moderate method called Attention Map Diversification (AMD), which transforms the original attention map " + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "text", + "content": " by multiplying it with a diversity matrix " + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "text", + "content": ". Specifically, the diversified attention map " + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "inline_equation", + "content": "A'" + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "text", + "content": " is obtained by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 356, + 275, + 553, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 275, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 356, + 275, + 553, + 289 + ], + "type": "interline_equation", + "content": "A ^ {\\prime} = A M D (A) = \\operatorname {s o f t m a x} (A \\odot M), \\tag {3}", + "image_path": "e3b51fd4789993c772bc1e3e5fcd3a3575d4a8625ea7af25ef7b02a60a874482.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": " is the element-wise product and softmax function is applied to each row. The diversity matrix " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": " has the same shape as " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": ", and every element " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "M_{i,j}" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "(1 \\leq i,j \\leq N)" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": " conforms to the uniform distribution " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "U(1 - d, 1 + d)" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": " is a factor that controls the diversity level. Note that AMD is applied to attention maps at specified Transformer blocks according to a index set " + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 292, + 554, + 375 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 376, + 554, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 553 + ], + "type": "text", + "content": "The benefit of AMD on backward propagation. In principle, modifying the forward propagation would impact the subsequent backward propagation. Here, we theoretically demonstrate that the impact of our AMD on backward propagation can further improve transferability. Specifically, we find that AMD implicitly induces gradient vanishing at attention map positions during backward propagation, which aligns with a previous surrogate refinement, PNAPO [38], that directly truncates the gradients of attention maps during backpropagation. The underlying principle of PNAPO suggests that gradients passing through attention maps carry overfitting information from surrogate models, and attenuating these gradients can enhance the transferability of adversarial examples. Our findings can be explained as follows." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 555, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 579 + ], + "type": "text", + "content": "When incorporating AMD, we let " + }, + { + "bbox": [ + 313, + 555, + 554, + 579 + ], + "type": "inline_equation", + "content": "S = A \\odot M" + }, + { + "bbox": [ + 313, + 555, + 554, + 579 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 313, + 555, + 554, + 579 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 555, + 554, + 579 + ], + "type": "text", + "content": "-th row vector of diversified attention map " + }, + { + "bbox": [ + 313, + 555, + 554, + 579 + ], + "type": "inline_equation", + "content": "A_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 555, + 554, + 579 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 586, + 556, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 586, + 556, + 633 + ], + "spans": [ + { + "bbox": [ + 319, + 586, + 556, + 633 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} A _ {i} ^ {\\prime} = \\left(A _ {i, 1} ^ {\\prime}, A _ {i, 2} ^ {\\prime}, \\dots , A _ {i, N} ^ {\\prime}\\right) \\\\ = \\operatorname {s o f t m a x} \\left(A _ {i, 1} \\cdot M _ {i, 1}, A _ {i, 2} \\cdot M _ {i, 2},..., A _ {i, N} \\cdot M _ {i, N}\\right) \\\\ = \\operatorname {s o f t m a x} \\left(S _ {i, 1}, S _ {i, 2}, \\dots , S _ {i, N}\\right) \\tag {4} \\\\ \\end{array}", + "image_path": "6bec2f17a45b5974d880a0e550982a7c552213011a55a3bb232b6db940ae1f13.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 633, + 552, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 552, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 552, + 651 + ], + "type": "text", + "content": "and the derivative " + }, + { + "bbox": [ + 313, + 633, + 552, + 651 + ], + "type": "inline_equation", + "content": "\\frac{\\partial A_{i,m}^{\\prime}}{\\partial S_{i,n}}" + }, + { + "bbox": [ + 313, + 633, + 552, + 651 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 633, + 552, + 651 + ], + "type": "inline_equation", + "content": "1 \\leq m, n \\leq N" + }, + { + "bbox": [ + 313, + 633, + 552, + 651 + ], + "type": "text", + "content": ") is expressed by:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 332, + 656, + 553, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 656, + 553, + 684 + ], + "spans": [ + { + "bbox": [ + 332, + 656, + 553, + 684 + ], + "type": "interline_equation", + "content": "\\frac {\\partial A _ {i , m} ^ {\\prime}}{\\partial S _ {i , n}} = \\left\\{ \\begin{array}{l l} A _ {i, m} ^ {\\prime} \\cdot \\left(1 - A _ {i, m} ^ {\\prime}\\right), & \\text {i f} m = n \\\\ - A _ {i, m} ^ {\\prime} \\cdot A _ {i, n} ^ {\\prime}, & \\text {i f} m \\neq n \\end{array} \\right. \\tag {5}", + "image_path": "2b78cdbe41e32987787448e4fcd9c2f4ba8457a3b2c68cab765ee36d4f956452.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "As the diversity factor " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " becomes large, elements in the diversity matrix " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " will fluctuate significantly, causing ex" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25074" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "treme values in " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ", and ultimately pushing the elements in " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "A'" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " towards 0 or 1 due to the property of softmax function. From Eq. 5, whether the elements in " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "A'" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " approach 0 or 1, the derivative will tend towards 0, resulting in gradient vanishing. According to the chain rule of differentiation, the gradients of the attention maps subjected to our AMD will experience varying degrees of gradient vanishing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 263 + ], + "type": "text", + "content": "Compared with PNAPO, which directly discards all gradients flowing through the attention maps, the gradient vanishing in our AMD is implicit without directly altering any aspect of the backward propagation. Furthermore, the gradient vanishing in our AMD method is more fine-grained, occurring at the row level of the attention map. Each row experiences varying degrees of vanishing, introducing additional randomness that helps to further mitigate overfitting to the surrogate model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 270, + 257, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 257, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 257, + 283 + ], + "type": "text", + "content": "3.3. Momentum Token Embedding (MTE)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 288, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 467 + ], + "type": "text", + "content": "It is challenging to ensure convergence to the global optimum in the high-dimension space when generating adversarial examples. Typically, adversarial examples converge to a local optimum along an inaccurate path [4]. This causes the internal token embeddings of the surrogate model to contain inaccurate information during iterations, thereby weakening the transferability. Therefore, we borrow the idea of Momentum Distillation (MoD) [13] from the Vision-and-Language Pre-training (VLP) literature, where a momentum version of the model through the moving average of the primary model's parameters is maintained to mitigate the noise in the web-sourced training dataset. Similarly, we maintain the momentum version of token embedding updates to accumulate historical token embeddings while generating adversarial examples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 467, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 295, + 491 + ], + "type": "text", + "content": "We term this method Momentum Token Embedding (MTE), and formulate it as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 498, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 498, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 97, + 498, + 295, + 514 + ], + "type": "interline_equation", + "content": "\\widehat {E _ {t}} = M T E \\left(E _ {t}\\right) = \\eta \\cdot \\widehat {E _ {t - 1}} + s \\cdot E _ {t}, \\tag {6}", + "image_path": "ae26735210cc2348b5827b514e1d1524c6a86a1624b81aee703c996b9490030f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "text", + "content": " is the iteration index, " + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "text", + "content": " is the output token embedding of any Attention block or MLP block at the " + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "text", + "content": "-th iteration, " + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\widehat{E_t}" + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "text", + "content": " is the corresponding accumulated token embedding at the " + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "text", + "content": "-th iteration, " + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "text", + "content": " is the attenuation factor, and " + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 521, + 295, + 654 + ], + "type": "text", + "content": " is the scale factor. MTE is applied to all Attention and MLP blocks. Note that, different from MoD, our MTE does not require model training and can be seen as constructing a residual structure for output token embeddings across the time dimension. Previous work also finds the usefulness of the residual structure in improving transferability, but during backward propagation within each iteration [39]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 660, + 231, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 231, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 231, + 673 + ], + "type": "text", + "content": "3.4. Complementary AMD and MTE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "The previous subsections examine AMD and MTE from their respective perspectives. Here, we further discuss why they can collaboratively enhance transferability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 72, + 531, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 531, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 531, + 84 + ], + "type": "text", + "content": "Algorithm 1 Forward Propagation Refinement (FPR)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "content": "Input: A surrogate model " + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "content": " with the cross-entropy loss " + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "content": ", a clean image " + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "content": " and its true label " + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "content": ". The maximum perturbation bound " + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "content": ", the iteration number " + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "content": ", the decay factor " + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "content": ", step size " + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "content": ", index set " + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "content": ", diversity factor " + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "content": ", scale factor " + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "content": " and attenuation factor " + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 86, + 553, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 144, + 492, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 492, + 157 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 492, + 157 + ], + "type": "text", + "content": "Output: The adversarial image " + }, + { + "bbox": [ + 314, + 144, + 492, + 157 + ], + "type": "inline_equation", + "content": "x^{adv} = x_{T}" + }, + { + "bbox": [ + 314, + 144, + 492, + 157 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 159, + 473, + 182 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 320, + 159, + 473, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 159, + 473, + 171 + ], + "spans": [ + { + "bbox": [ + 320, + 159, + 473, + 171 + ], + "type": "text", + "content": "1: Initialize " + }, + { + "bbox": [ + 320, + 159, + 473, + 171 + ], + "type": "inline_equation", + "content": "g_0 = 0" + }, + { + "bbox": [ + 320, + 159, + 473, + 171 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 320, + 159, + 473, + 171 + ], + "type": "inline_equation", + "content": "E_0 = 0" + }, + { + "bbox": [ + 320, + 159, + 473, + 171 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 320, + 159, + 473, + 171 + ], + "type": "inline_equation", + "content": "x_0 = x" + }, + { + "bbox": [ + 320, + 159, + 473, + 171 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 171, + 407, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 171, + 407, + 182 + ], + "spans": [ + { + "bbox": [ + 320, + 171, + 407, + 182 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 320, + 171, + 407, + 182 + ], + "type": "inline_equation", + "content": "t = 1 \\rightarrow T" + }, + { + "bbox": [ + 320, + 171, + 407, + 182 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 331, + 183, + 475, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 183, + 475, + 194 + ], + "spans": [ + { + "bbox": [ + 331, + 183, + 475, + 194 + ], + "type": "text", + "content": "During the forward propagation:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 195, + 553, + 242 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 320, + 195, + 553, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 195, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 320, + 195, + 553, + 218 + ], + "type": "text", + "content": "3: Apply AMD (Eq. (3)) to attention maps in specific Transformer blocks according to the index set " + }, + { + "bbox": [ + 320, + 195, + 553, + 218 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 320, + 195, + 553, + 218 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 219, + 553, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 219, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 320, + 219, + 553, + 242 + ], + "type": "text", + "content": "4: Apply MTE (Eq. (6)) to output token embeddings in all Transformer (Attention and MLP) blocks." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 331, + 243, + 483, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 243, + 483, + 255 + ], + "spans": [ + { + "bbox": [ + 331, + 243, + 483, + 255 + ], + "type": "text", + "content": "During the backward propagation:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 255, + 526, + 278 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 320, + 255, + 526, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 255, + 526, + 266 + ], + "spans": [ + { + "bbox": [ + 320, + 255, + 526, + 266 + ], + "type": "text", + "content": "5: Calculate the gradient: " + }, + { + "bbox": [ + 320, + 255, + 526, + 266 + ], + "type": "inline_equation", + "content": "\\nabla_{x_{t - 1}}L(F(x_{t - 1}),y)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 267, + 464, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 267, + 464, + 278 + ], + "spans": [ + { + "bbox": [ + 320, + 267, + 464, + 278 + ], + "type": "text", + "content": "6: Update momentum gradient:" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 356, + 287, + 527, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 287, + 527, + 316 + ], + "spans": [ + { + "bbox": [ + 356, + 287, + 527, + 316 + ], + "type": "interline_equation", + "content": "g _ {t} = \\mu \\cdot g _ {t - 1} + \\frac {\\nabla_ {x _ {t - 1}} L (F (x _ {t - 1}) , y)}{\\left\\| \\nabla_ {x _ {t - 1}} L (F (x _ {t - 1}) , y) \\right\\| _ {1}}", + "image_path": "c009dc52d16d4584269f18dfd31c50495fc38963d918e6e7758bf93286fca42b.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 324, + 477, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 324, + 477, + 335 + ], + "spans": [ + { + "bbox": [ + 320, + 324, + 477, + 335 + ], + "type": "text", + "content": "7: Update the adversarial example:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 372, + 346, + 512, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 346, + 512, + 359 + ], + "spans": [ + { + "bbox": [ + 372, + 346, + 512, + 359 + ], + "type": "interline_equation", + "content": "x _ {t} = \\operatorname {C l i p} _ {\\epsilon} \\left\\{x _ {t - 1} + \\alpha \\cdot \\operatorname {s i g n} \\left(g _ {t}\\right) \\right\\}", + "image_path": "43a4e2dcea7af9b71472088ae3246abd3cf0443f72493d15b0be6864074bfb93.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 368, + 365, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 368, + 365, + 378 + ], + "spans": [ + { + "bbox": [ + 320, + 368, + 365, + 378 + ], + "type": "text", + "content": "8: end for" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 418, + 555, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 610 + ], + "type": "text", + "content": "Specifically, AMD and MTE refine the forward propagation of ViTs from two complementary angles. AMD falls under the category of \"external refinement\", which enhances surrogate models by introducing extraneous information or perturbing existing embeddings. Techniques such as attention map dropout also fall into this category. Our empirical results demonstrate that this type of refinement is relatively aggressive, making it unsuitable for concurrent use or for application across all blocks. In contrast, MTE belongs to the \"internal refinement\" category, mainly leveraging information from the input data itself. Since these refinements rely on self-related information, they are typically more moderate and can be applied to all blocks. Both AMD and MTE help promoting transferability but in different ways, this complementary effect further enhances the transferability of adversarial examples." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "Finally, we provide the pseudocode of FPR (based on MIM [4]) in Algorithm 1. Specifically, at each of " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": " iterations, we first apply AMD to specific attention maps according to the index set " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": " and MTE to the outputs of all Attention and MLP blocks during the forward propagation using PyTorch's Hook function. Then, during the backward propagation, we compute the gradient of the current adversarial example and use MIM [4] to update it." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25075" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 273, + 143 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 273, + 143 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 273, + 143 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 273, + 143 + ], + "type": "table", + "html": "
ModelIdsη
ViT-B0,1,4,9,11250.80.3
CaiT-S2,14,25300.60.2
PiT-T1,6,11250.70.2
DeiT-B0,1,5,10,11300.70.2
", + "image_path": "dacf2cc0d76a2e0dc8b20d05eef1c1727e095e2a24f2c3b3c157c1aa2ee69050.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 152, + 294, + 207 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 294, + 207 + ], + "type": "text", + "content": "Table 1. The optimal parameter settings of our FPR about its index set " + }, + { + "bbox": [ + 55, + 152, + 294, + 207 + ], + "type": "inline_equation", + "content": "(I)" + }, + { + "bbox": [ + 55, + 152, + 294, + 207 + ], + "type": "text", + "content": ", diversity factor " + }, + { + "bbox": [ + 55, + 152, + 294, + 207 + ], + "type": "inline_equation", + "content": "(d)" + }, + { + "bbox": [ + 55, + 152, + 294, + 207 + ], + "type": "text", + "content": ", scale factor " + }, + { + "bbox": [ + 55, + 152, + 294, + 207 + ], + "type": "inline_equation", + "content": "(s)" + }, + { + "bbox": [ + 55, + 152, + 294, + 207 + ], + "type": "text", + "content": ", and attenuation factor " + }, + { + "bbox": [ + 55, + 152, + 294, + 207 + ], + "type": "inline_equation", + "content": "(\\eta)" + }, + { + "bbox": [ + 55, + 152, + 294, + 207 + ], + "type": "text", + "content": ". Detailed ablation studies in Section 4.3 provide the sensitivity analysis of the attack performance towards the parameter selection under four cases." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 135, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 135, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 135, + 243 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 248, + 181, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 181, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 181, + 262 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 266, + 295, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 338 + ], + "type": "text", + "content": "Datasets and baselines. Following the common practice [34, 45, 49], we use 1000 images randomly selected from the ILSVRC2012 [25] validation set. We compare our FPR with other surrogate refinements that focus on backward propagation: PNAPO [38], TGR [45], and GNS [51]. All these methods are built on MIM [4]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 339, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 295, + 529 + ], + "type": "text", + "content": "Models and defenses. For surrogate models, we consider four popular ViTs: ViT-Base (ViT-B) [5], CaiT-Small (CaiT-S) [29], PiT-Tiny (PiT-T) [9] and Deit-Base (Deit-B) [28]. For target models, we consider seven ViTs and six CNNs: ViT-B [5], CaiT-S [29], PiT-Base (PiT-B) [9], Visformer-Small (Vis-S) [3], Swin-Tiny (Swin-T) [18], DeiT-Tiny (DeiT-T) [28], and CoaT-Tiny (CoaT-T) [43], ResNet-18 (RN-18) [8], VGG-16 [26], DenseNet-121 (DN-121) [11], EfficientNet-b0 (EN-b0) [27], MobileNetv3 (MN-v3) [10], and ResNeXt-50 (RNX-50) [41]. Five popular defenses are tested: adversarial training (AT) paired with RN-50 [30], high-level representation guided denoiser (HGD) [16], neural representation purifier (NRP) [19], Bit depth reduction (BDR) [42], and JPEG compression [7]. Finally, two online models, i.e., Baidu Cloud API and Aliyun API, are also included." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 531, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 638 + ], + "type": "text", + "content": "Hyperparameter settings. Commonly, we set the number of iterations as 10, the maximum perturbation bound " + }, + { + "bbox": [ + 55, + 531, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 531, + 295, + 638 + ], + "type": "text", + "content": " as 16, the step size " + }, + { + "bbox": [ + 55, + 531, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 531, + 295, + 638 + ], + "type": "text", + "content": " as 1.6, and the decay factor of momentum gradient " + }, + { + "bbox": [ + 55, + 531, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mu = 1" + }, + { + "bbox": [ + 55, + 531, + 295, + 638 + ], + "type": "text", + "content": ". Parameters of all baselines are the same as their default values. For our FPR, the optimal hyperparameters are listed in Table 1, with a detailed analysis in Section 4.3 discussing their sensitivity to different parameter choices. All experiments are conducted on a single RTX 4060 GPU with 8 GB of VRAM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 178, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 178, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 178, + 661 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Different target models. We first compare FPR with baselines on transferring adversarial examples from ViTs to different ViTs and CNNs. Table 2 shows that FPR consistently outperforms baselines in all cases, suggesting the potential" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "of refining the surrogate model from the perspective of forward propagation. For example, when generating adversarial examples on ViT-B, FPR achieves an average attack success rate of " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "84.3\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " on six CNNs, surpassing GNS, TGR, and PNAPO by " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "15.5\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", respectively, and improves the baseline MIM by " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "33.4\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 144, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 252 + ], + "type": "text", + "content": "Different defenses. We then evaluate different attack methods against five popular defenses. As shown in Table 3, FPR maintains its superiority over existing surrogate refinements in such challenging defense settings. For instance, when generating adversarial examples on ViT-B, FPR achieves an average attack success rate of " + }, + { + "bbox": [ + 313, + 144, + 553, + 252 + ], + "type": "inline_equation", + "content": "70.8\\%" + }, + { + "bbox": [ + 313, + 144, + 553, + 252 + ], + "type": "text", + "content": " across five defenses, outperforming other surrogate refinements GNS, TGR, and PNAPO by " + }, + { + "bbox": [ + 313, + 144, + 553, + 252 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 313, + 144, + 553, + 252 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 144, + 553, + 252 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 313, + 144, + 553, + 252 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 144, + 553, + 252 + ], + "type": "inline_equation", + "content": "12.4\\%" + }, + { + "bbox": [ + 313, + 144, + 553, + 252 + ], + "type": "text", + "content": ", respectively, while surpassing the vanilla baseline, MIM, by " + }, + { + "bbox": [ + 313, + 144, + 553, + 252 + ], + "type": "inline_equation", + "content": "24.3\\%" + }, + { + "bbox": [ + 313, + 144, + 553, + 252 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 253, + 553, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 419 + ], + "type": "text", + "content": "Incorporating other transfer methods. A good surrogate refinement-based transfer method should also exhibit strong compatibility with other (types of) transfer-based attack methods. In this study, we integrate FPR into two advanced gradient correction-based transferable attacks, VTM [34] and GRA [49], as well as two latest input augmentation-based transferable attacks, BSR [32] and SIA [36]. Specifically, we equip these four attacks with 5 augmented copies per iteration, while keeping other parameter settings as their default values. Results in Table 4 indicate that FPR can effectively complement existing transfer-based attacks for stronger transferability. For instance, when combined with our FPR, the attack success rate of VTM increases from " + }, + { + "bbox": [ + 313, + 253, + 553, + 419 + ], + "type": "inline_equation", + "content": "57.3\\%" + }, + { + "bbox": [ + 313, + 253, + 553, + 419 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 253, + 553, + 419 + ], + "type": "inline_equation", + "content": "81.7\\%" + }, + { + "bbox": [ + 313, + 253, + 553, + 419 + ], + "type": "text", + "content": " on CNNs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 421, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 553, + 468 + ], + "type": "text", + "content": "Online models. We also evaluate our FPR on two online models, i.e., Baidu Cloud API and Aliyun API. The results, presented in Appendix A, show that FPR still outperforms both TGR and GNS." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 479, + 414, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 414, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 414, + 491 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 497, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 553, + 604 + ], + "type": "text", + "content": "Individual components. We compare each component of FPR with the plain baseline MIM, and the two best-performed surrogate refinements, TGR and GNS. We also consider Dropout as a comparison to our AMD. From Table 5, we can see that combining our AMD and MTE yields the best transferability results. In addition, Dropout is less effective than our AMD, indicating that directly removing certain attention maps is rather aggressive and fails to retain useful information." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": "Parameter sensitivity. Figure 4 shows the attack success rates when varying three main parameters: diversity factor " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": ", scale factor " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": ", and attenuation factor " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": " controls the diversity of the attention map. FPR achieves satisfactory performance when " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": " is between 25 and 30. " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": " represents the importance of current token embeddings. FPR achieves satisfactory performance when " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": " is between 0.6 and 0.8. " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": " signifies the importance of historical token embeddings. FPR achieves satisfactory performance when " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": " is between 0.2" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25076" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 264 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 264 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 264 + ], + "type": "table", + "html": "
ModelAttackViT-BCaiT-SPiT-BVis-SSwin-TDeiT-TCoaT-TRN-18VGG-16DN-121EN-b0MN-v3RNX-50Average
ViT-BMIM [4]97.261.840.442.354.965.844.451.757.251.350.851.738.250.9
PNAPO [38]99.183.262.165.874.783.064.067.470.067.668.563.056.368.8
TGR [45]99.186.263.869.981.194.470.479.179.875.383.179.960.076.9
GNS [51]99.989.568.174.082.493.672.277.876.776.079.176.062.077.3
FPR99.292.573.078.388.598.277.787.387.682.090.387.468.384.3
CaiT-SMIM [4]68.598.850.054.968.275.354.861.067.859.861.055.747.160.3
PNAPO [38]69.888.258.161.070.373.860.164.869.764.365.561.753.264.4
TGR [45]90.1100.076.081.391.297.880.686.986.684.387.782.169.184.5
GNS [51]91.2100.078.682.291.697.681.186.687.184.989.681.970.885.3
FPR93.8100.081.685.695.098.886.887.488.788.690.682.976.888.1
PiT-TMIM [4]30.036.138.543.351.075.055.569.568.760.864.571.640.454.2
PNAPO [38]39.550.949.654.961.685.465.078.776.772.277.981.050.864.9
TGR [45]37.947.746.954.162.587.465.182.882.075.581.888.853.766.6
GNS [51]37.144.146.451.959.484.463.379.978.071.576.384.348.663.5
FPR44.356.157.162.469.591.372.288.786.480.487.691.160.372.9
DeiT-BMIM [4]86.082.467.069.579.787.765.369.070.570.171.461.458.272.2
PNAPO [38]83.487.772.474.182.286.071.974.974.774.976.168.066.376.4
TGR [45]92.497.181.284.392.697.983.586.085.285.089.984.574.687.2
GNS [51]92.198.080.584.291.098.079.883.380.783.586.781.370.885.4
FPR94.698.483.886.795.599.587.089.889.288.193.986.879.890.2
", + "image_path": "985915af2b6096ce747f8a4cb71b5e52a41eb746fad9631b5e304050bcef18a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 313, + 294, + 515 + ], + "blocks": [ + { + "bbox": [ + 55, + 272, + 555, + 296 + ], + "lines": [ + { + "bbox": [ + 55, + 272, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 555, + 296 + ], + "type": "text", + "content": "Table 2. Attack success rates (%) of different surrogate refinements on various ViTs and CNNs. The \"Average\" column excludes the white-box results with grey background." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 313, + 294, + 515 + ], + "lines": [ + { + "bbox": [ + 58, + 313, + 294, + 515 + ], + "spans": [ + { + "bbox": [ + 58, + 313, + 294, + 515 + ], + "type": "table", + "html": "
ModelMethodATHGDNRPJPEGBitAverage
ViT-BMIM [4]43.929.842.157.559.146.5
PNAPO [38]45.647.553.570.974.858.4
TGR [45]49.750.861.677.782.064.4
GNS [51]49.254.861.377.982.365.1
FPR50.761.469.284.588.470.8
CaiT-SMIM [4]44.337.446.264.166.951.8
PNAPO [38]44.843.748.265.267.353.8
TGR [45]49.860.766.481.286.869.0
GNS [51]49.862.767.482.688.070.1
FPR51.168.970.886.290.473.5
PiT-TMIM [4]46.132.042.057.562.047.9
PNAPO [38]47.542.948.565.569.854.8
TGR [45]50.844.050.568.271.957.1
GNS [45]48.938.649.365.368.854.2
FPR51.449.956.173.277.461.6
DeiT-BMIM [4]44.748.152.270.374.958.1
PNAPO [38]46.957.257.073.678.762.7
TGR [45]50.265.267.882.288.170.7
GNS [51]48.361.665.780.686.368.5
FPR50.270.771.387.091.574.1
", + "image_path": "b094136fdd60a0564b2121fea4ccd9ac04a30d4aea93dd843d0d905c59ae5577.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 592, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 640 + ], + "type": "text", + "content": "and 0.3. Detailed results of applying the same index set " + }, + { + "bbox": [ + 55, + 592, + 295, + 640 + ], + "type": "inline_equation", + "content": "(I)" + }, + { + "bbox": [ + 55, + 592, + 295, + 640 + ], + "type": "text", + "content": " to other methods and the sensitivity of the " + }, + { + "bbox": [ + 55, + 592, + 295, + 640 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 592, + 295, + 640 + ], + "type": "text", + "content": " used in AMD are reported in Appendix B, where we also provide insights for identifying the optimal " + }, + { + "bbox": [ + 55, + 592, + 295, + 640 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 592, + 295, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Gradient vanishing comparison. We compare the extent of gradient vanishing in Table 6. We can conclude that gradient vanishing can improve transferability, but it is not a simple linear correlation. This aligns with our claim that the key of AMD is to diversify attention maps in the forward propagation, while it implicitly and mildly induces some" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 320, + 313, + 549, + 437 + ], + "blocks": [ + { + "bbox": [ + 55, + 523, + 295, + 567 + ], + "lines": [ + { + "bbox": [ + 55, + 523, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 295, + 567 + ], + "type": "text", + "content": "Table 3. Attack success rates (\\%) of different surrogate refinements against various defenses. AT is trained with RN-50; HGD is based on Inc-v3; NRP, JPEG and Bit are averaged on seven ViTs and six CNNs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 313, + 549, + 437 + ], + "lines": [ + { + "bbox": [ + 320, + 313, + 549, + 437 + ], + "spans": [ + { + "bbox": [ + 320, + 313, + 549, + 437 + ], + "type": "table", + "html": "
MethodViTsCNNs
SIA [36]73.469.1
SIA+FPR86.690.4
BSR [32]69.466.9
BSR+FPR84.889.9
VTM [34]63.157.3
VTM+FPR84.081.7
GRA [49]73.668.3
GRA+FPR86.185.3
", + "image_path": "d38932d9ed7eeff8ef6a9a1c3e8f8fa5a87dc9722d6210ba9d61005f6f22ba19.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 321, + 502, + 548, + 613 + ], + "blocks": [ + { + "bbox": [ + 313, + 445, + 555, + 468 + ], + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 468 + ], + "type": "text", + "content": "Table 4. Attack success rates " + }, + { + "bbox": [ + 313, + 445, + 555, + 468 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 445, + 555, + 468 + ], + "type": "text", + "content": " of combined methods averaged on six CNNs and six ViTs (without surrogate model ViT-B)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 502, + 548, + 613 + ], + "lines": [ + { + "bbox": [ + 321, + 502, + 548, + 613 + ], + "spans": [ + { + "bbox": [ + 321, + 502, + 548, + 613 + ], + "type": "table", + "html": "
MethodViTsCNNs
MIM [4]51.650.2
Dropout55.858.9
TGR [45]77.676.2
GNS [51]80.074.6
AMD82.178.4
MTE78.479.6
FPR (AMD+MTE)84.783.8
", + "image_path": "6e9bb1b33289957cca9a5ad1df363a13ff1faf836ea05e08746bccc6b9f64867.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 621, + 555, + 656 + ], + "lines": [ + { + "bbox": [ + 313, + 621, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 555, + 656 + ], + "type": "text", + "content": "Table 5. Ablation studies of different FPR components with ViT-B as the surrogate model. All involved methods are built on MIM. Dropout is applied at the same index set as AMD." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 701, + 509, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 509, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 509, + 714 + ], + "type": "text", + "content": "gradient vanishing in the backward propagation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25077" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 545, + 457 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 545, + 457 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 545, + 457 + ], + "type": "image", + "image_path": "618e5a8c374e4fda37b1ad913610fe6953af7a7a52521ad10689397e7c301fe8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 465, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 55, + 465, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 555, + 489 + ], + "type": "text", + "content": "Figure 4. Sensitivity analysis about diversity factor, scale factor, and attenuation factor, where ASR remains relatively stable within a certain range of values for these three factors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 506, + 298, + 551 + ], + "blocks": [ + { + "bbox": [ + 56, + 506, + 298, + 551 + ], + "lines": [ + { + "bbox": [ + 56, + 506, + 298, + 551 + ], + "spans": [ + { + "bbox": [ + 56, + 506, + 298, + 551 + ], + "type": "table", + "html": "
AttackMIM [4]PNA [38]TGR [45]GNS [51]AMD
Mean2.9 × 10-40.03.9 × 10-102.4 × 10-151.5 × 10-5
ASR50.960.663.262.780.3
", + "image_path": "2149433602d52476e0755f1a66b8c2299e653c2da52dbae567f16b2471347ca3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 552, + 296, + 595 + ], + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 595 + ], + "type": "text", + "content": "Table 6. Mean of attention map's input gradient elements from 100 images over 10 iterations. All surrogate refinements adopt Index Set (0,1,4,9,11). ASR is averaged on 13 models and ViT-B is the surrogate model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 619, + 128, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 128, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 128, + 632 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "In this work, we propose the Forward Propagation Refinement (FPR) to generate highly transferable adversarial examples on Vision Transformers (ViTs). FPR is the first work to refine the surrogate ViTs during forward propagation entirely. Specifically, FPR consists of two parts: Attention Map Diversification (AMD) and Momentum To" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 509, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 581 + ], + "type": "text", + "content": "ken Embedding (MTE). AMD diversifies the attention maps and implicitly causes gradient vanishing to enhance transferability, while MTE accumulates historical token embeddings to stabilize current embeddings for more accurate updates. The superiority of FPR is supported by both empirical results and theoretical analyses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 596, + 411, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 596, + 411, + 610 + ], + "spans": [ + { + "bbox": [ + 314, + 596, + 411, + 610 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "This research is supported by the National Key Research and Development Program of China (2023YFE0209800), the National Natural Science Foundation of China (62406240, 62376210, 62132011, 62206217, T2341003, U244120060, 62161160337, U2441240, U24B20185, U21B2018), the Shaanxi Province Key Industry Innovation Program (2023-ZDLGY-38). Thanks to the New Cornerstone Science Foundation and the Xplorer Prize." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25078" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Srinadh Bhojanapalli, Ayan Chakrabarti, Daniel Glasner, Daliang Li, Thomas Unterthiner, and Andreas Veit. Understanding robustness of transformers for image classification. In IEEE International Conference on Computer Vision, pages 10211-10221, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "text", + "content": "[2] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In IEEE Conference on Computer Vision and Pattern Recognition, pages 24648-24657, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "type": "text", + "content": "[3] Zhengsu Chen, Lingxi Xie, Jianwei Niu, Xuefeng Liu, Longhui Wei, and Qi Tian. Visformer: The vision-friendly transformer. In IEEE International Conference on Computer Vision, pages 569-578, 2021. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "text", + "content": "[4] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 294, + 346 + ], + "type": "text", + "content": "[5] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, and et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 294, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 294, + 388 + ], + "type": "text", + "content": "[6] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 390, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 390, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 390, + 294, + 434 + ], + "type": "text", + "content": "[7] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering adversarial images using input transformations. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "type": "text", + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 294, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 294, + 534 + ], + "type": "text", + "content": "[9] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In IEEE International Conference on Computer Vision, pages 11916-11925, 2021. 1, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "text", + "content": "[10] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint, arXiv:1704.04861, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 294, + 635 + ], + "type": "text", + "content": "[11] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2261-2269, 2017. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 669 + ], + "type": "text", + "content": "[12] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In International Conference on Learning Representations, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[13] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "formation Processing Systems, pages 9694-9705, 2021. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[14] Qizhang Li, Yiwen Guo, Wangmeng Zuo, and Hao Chen. Making substitute models more bayesian can enhance transferability of adversarial examples. In International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 186 + ], + "type": "text", + "content": "[15] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan Yuille. Learning transferable adversarial examples via ghost networks. In AAAI Conference on Artificial Intelligence, pages 11458-11465, 2020. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "text", + "content": "[16] Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial attacks using high-level representation guided denoiser. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2018. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 286 + ], + "type": "text", + "content": "[17] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "type": "text", + "content": "[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE International Conference on Computer Vision, pages 9992-10002, 2021. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "text", + "content": "[19] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A self-supervised approach for adversarial robustness. In IEEE Conference on Computer Vision and Pattern Recognition, pages 259-268, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 444 + ], + "type": "text", + "content": "[20] Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Fahad Shahbaz Khan, and Fatih Porikli. On improving adversarial transferability of vision transformers. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "type": "text", + "content": "[21] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In Advances in Neural Information Processing Systems, pages 23296-23308, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[22] Yuchen Ren, Hegui Zhu, Xiaoyan Sui, and Chong Liu. Crafting transferable adversarial examples via contaminating the salient feature variance. Information Sciences, 644:119273, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 546, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 553, + 590 + ], + "type": "text", + "content": "[23] Yuchen Ren, Hegui Zhu, Chong Liu, and Chengqing Li. Efficient polar coordinates attack with adaptive activation strategy. Expert Systems with Applications, 249:123850, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 591, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 646 + ], + "type": "text", + "content": "[24] Yuchen Ren, Zhengyu Zhao, Chenhao Lin, Bo Yang, Lu Zhou, Zhe Liu, and Chao Shen. Improving integrated gradient-based transferable adversarial examples by refining the integration path. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[25] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. International Journal of Computer Vision, 115(3): 211–252, 2015. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25079" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[26] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "type": "text", + "content": "[27] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357, 2021. 1, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[29] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Herve Jegou. Going deeper with image transformers. In IEEE International Conference on Computer Vision, pages 32-42, 2021. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[30] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 341 + ], + "type": "text", + "content": "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, and et al. Attention is all you need. In International Conference on Neural Information Processing Systems, pages 6000-6010, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[32] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting adversarial transferability by block shuffle and rotation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 24336-24346, 2024. 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "text", + "content": "[33] Kang Wang, Jianzhou Wang, Zhiwu Li, and Yilin Zhou. An integrated sparse gated graph density network based on transfer learning for multi-site probabilistic forecasting of renewable energy. IEEE Transactions on Sustainable Energy, 16(1):673-685, 2025. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[34] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1924-1933, 2021. 1, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[35] Xiaosen Wang, Kangheng Tong, and Kun He. Rethinking the backward propagation for adversarial transferability. In Advances in Neural Information Processing Systems, pages 1905-1922, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[36] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In IEEE International Conference on Computer Vision, pages 4584-4596, 2023. 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 622 + ], + "type": "text", + "content": "[37] Zhiyuan Wang, Zeliang Zhang, Siyuan Liang, and Xiaosen Wang. Diversifying the high-level features for better adversarial transferability. In British Machine Vision Conference 2023, BMVC 2023, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": "[38] Zhipeng Wei, Jingjing Chen, Micah Goldblum, Zuxuan Wu, Tom Goldstein, and Yu-Gang Jiang. Towards transferable adversarial attacks on vision transformers. In AAAI Conference on Artificial Intelligence, pages 2668-2676, 2022. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[39] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip connections matter: On the transferability of adversarial examples generated with resnets. In Interna-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 699 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "tional Conference on Learning Representations, 2020. 1, 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[40] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[41] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5987-5995, 2017. 1, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "text", + "content": "[42] Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In Network and Distributed System Security Symposium, 2018. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 242, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 284 + ], + "type": "text", + "content": "[43] Weijian Xu, Yifan Xu, Tyler Chang, and Zhuowen Tu. Coscale conv-attentional image transformers. In IEEE International Conference on Computer Vision, pages 9961–9970, 2021. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "text", + "content": "[44] Bo Yang, Hengwei Zhang, Jindong Wang, Yulong Yang, Chenhao Lin, Chao Shen, and Zhengyu Zhao. Adversarial example soups: Improving transferability and stealthiness for free. IEEE Transactions on Information Forensics and Security, 20:1882-1894, 2025. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 396 + ], + "type": "text", + "content": "[45] Jianping Zhang, Yizhan Huang, Weibin Wu, and Michael R. Lyu. Transferable adversarial attacks on vision transformers with token gradient regularization. In IEEE Conference on Computer Vision and Pattern Recognition, pages 16415-16424, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "text", + "content": "[46] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. Towards large yet imperceptible adversarial image perturbations with perceptual color distance. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1036-1045, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "text", + "content": "[47] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. On success and simplicity: A second look at transferable targeted attacks. In Advances in Neural Information Processing Systems, pages 6115-6128, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 488, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 542 + ], + "type": "text", + "content": "[48] Zhengyu Zhao, Hanwei Zhang, Renjue Li, Ronan Sicre, Laurent Amsaleg, Michael Backes, Qi Li, and Chao Shen. Revisiting transferable adversarial image examples: Attack categorization, evaluation guidelines, and new insights. arXiv preprint, arXiv:2310.11850, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "type": "text", + "content": "[49] Hegui Zhu, Yuchen Ren, Xiaoyan Sui, Lianping Yang, and Wuming Jiang. Boosting adversarial transferability via gradient relevance attack. In IEEE International Conference on Computer Vision, pages 4718-4727, 2023. 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 589, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 643 + ], + "type": "text", + "content": "[50] Yao Zhu, Yuefeng Chen, Xiaodan Li, Kejiang Chen, Yuan He, Xiang Tian, Bolun Zheng, Yaowu Chen, and Qingming Huang. Toward understanding and boosting adversarial transferability from a distribution perspective. IEEE Transactions on Image Processing, 31:6487-6501, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 644, + 553, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 644, + 553, + 699 + ], + "spans": [ + { + "bbox": [ + 316, + 644, + 553, + 699 + ], + "type": "text", + "content": "[51] Zhiyu Zhu, Xinyi Wang, Zhibo Jin, Jiayu Zhang, and Huaming Chen. Enhancing transferable adversarial attacks on vision transformers through gradient normalization scaling and high-frequency adaptation. In International Conference on Learning Representations, 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25080" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/2639ed20-911e-44d7-9142-a8fcfe0f6597_content_list.json b/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/2639ed20-911e-44d7-9142-a8fcfe0f6597_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..846d024de62f18eeb0c817ebd519f1fedc7f34bf --- /dev/null +++ b/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/2639ed20-911e-44d7-9142-a8fcfe0f6597_content_list.json @@ -0,0 +1,1453 @@ +[ + { + "type": "text", + "text": "Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction", + "text_level": 1, + "bbox": [ + 246, + 130, + 750, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Teng Hu $^{1*}$ , Jiangning Zhang $^{2,3*}$ , Ran Yi $^{1\\dagger}$ , Jieyu Weng $^{1}$ , Yabiao Wang $^{3,2}$ , Xianfang Zeng $^{3}$ , Zhucun Xue $^{3}$ , Lizhuang Ma $^{1}$ $^{1}$ Shanghai Jiao Tong University, $^{2}$ Youtu Lab, Tencent, $^{3}$ Zhejiang University {hu-teng, ranyi, w.jerry, lzma} @sjtu.edu.cn, {186368, yabiaowang, zzlongjuanfeng, 12432038} @zju.edu.cn", + "bbox": [ + 189, + 202, + 805, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Employing LLMs for visual generation has recently become a research focus. However, the existing methods primarily transfer the LLM architecture to visual generation but rarely investigate the fundamental differences between language and vision. This oversight may lead to suboptimal utilization of visual generation capabilities within the LLM framework. In this paper, we explore the characteristics of visual embedding space under the LLM framework and discover that the correlation between visual embeddings can help achieve more stable and robust generation results. We present IAR, an Improved AutoRegressive Visual Generation Method that enhances the training efficiency and generation quality of LLM-based visual generation models. Firstly, we propose a Codebook Rearrange strategy that uses balanced $k$ -means clustering algorithm to rearrange the visual codebook into clusters, ensuring high similarity among visual features within each cluster. Leveraging the rearranged codebook, we propose a Cluster-oriented Cross-entropy Loss that guides the model to correctly predict the cluster where the target token is located. This approach ensures that even if the model predicts the wrong token index, there is a high probability the predicted token is located in the correct cluster, which significantly enhances the generation quality and robustness. Extensive experiments demonstrate that our IAR consistently enhances the model training efficiency and performance from 100M to 1.4B, reducing the training time by half while achieving the same FID. Additionally, IAR can be applied to various LLM-based visual generation models and adheres to the scaling law, providing a promising direction for future research in LLM-based visual generation. The code is available at: https://github.com/sjtuplayer/IAR.", + "bbox": [ + 88, + 361, + 483, + 845 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/05c7aee6e1d6d40473115d53426ce46150c7c488c6ca9b40865216aff4112f71.jpg", + "image_caption": [ + "Figure 1. When an autoregressive model predicts a wrong token, the previous methods [36, 38] may predict an irrelevant token that causes artifacts. Our IAR alleviates this issue by ensuring a high probability of the predicted token located in the correct cluster." + ], + "image_footnote": [], + "bbox": [ + 516, + 325, + 906, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 598, + 643, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the development of generative models [15, 19, 33], a large number of outstanding visual generation methods have emerged, achieving considerable success in image and video generation [3, 16, 33]. Recently, researchers have begun exploring the integration of images and text to achieve a unified multi-modal model for image and text generation and understanding [37, 44]. Consequently, aligning image generation models with large language models (LLMs) [32] has become a key research focus. Recent studies have made progress by quantizing images and employing autoregressive or mask-prediction methods to predict discrete image tokens [2, 6, 36, 38], laying the research foundation for a unified world model in the future.", + "bbox": [ + 509, + 625, + 906, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike diffusion [19] or GAN [15] models that model image distribution in continuous space, autoregressive or masked image modeling (MIM) methods [2, 6, 36, 38] first convert images into discrete-valued tokens by image tokenizers, and then predict image tokens by autoregressive or", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9351", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MIM approach. Specifically, autoregressive methods use the \"next-token prediction\" paradigm of GPT [32], while MIM methods adopt a similar training strategy to the mask prediction in BERT [11]. These methods draw inspiration from natural language models and transfer these techniques to image generation models, but rarely explore the fundamental differences between images and natural language.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the inherently discrete nature of words (each word comes from a finite vocabulary), different words can be directly mapped to different indices through a lookup table, thereby generating text by predicting the corresponding index of the target text. The standard language modeling objective $p(x_{t}|x_{1},\\ldots ,x_{t - 1})$ (each $x_{t}$ is an index) then can be interpreted as a classification task and accurately modeled by LLMs. However, images are continuous-valued, and if the same objective is applied to image generation with $x_{t}$ as a real number, it is difficult to accurately model this probability density. Therefore, image tokenizers such as VQGAN [13] are used to convert continuous images into discrete-valued tokens, and then the model predicts the image token index. Afterwards, the corresponding image embedding is retrieved from the codebook and decoded into an image. A fundamental difference between image and text is that, in text generation, the predicted text index can be directly mapped to the corresponding word through a lookup table; whereas LLM-based image generation essentially requires the image embedding that corresponds to the index, and then decodes it into an image. I.e., text only requires the index, while image requires the embedding corresponding to the index.", + "bbox": [ + 91, + 200, + 486, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from the token index which is independent, i.e., the nearby indices have no correlation in word semantics, the image embeddings are located in a continuous feature space, where similar embeddings may correspond to similar image contents. This paper first investigates the impact of image embedding correlations on the generated results, and finds that similar embeddings convey similar information in the image space: When the embeddings of certain image patches are replaced by similar image embeddings, the decoded images are nearly identical. This insight inspires us to leverage the similarity of image embeddings to improve the existing LLM-based image generation techniques.", + "bbox": [ + 89, + 550, + 482, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To capitalize on the correlation among image embeddings, we propose IAR, an Improved AutoRegressive Visual Generation Method, which can enhance the training efficiency and generation quality of the LLM-based visual generation model. Firstly, we propose a Codebook Rearrangement strategy that uses balanced K-means clustering algorithm to rearrange the embeddings in the codebook into clusters of equal sizes, where the embeddings in each cluster share high similarities and can be decoded into images with similar contents. Secondly, we observe that if a token index is wrongly predicted as a similar embedding's", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "index, the original token-oriented cross-entropy loss used in LLM will penalize this case, but the decoded image is actually not much different from the target image. To tolerate this case, based on the rearranged codebook, we further introduce a Cluster-oriented Cross-entropy Loss, which guides the model to predict the correct cluster that contains the target token, thereby providing the model with a broader perspective and no longer confined to a single target token. Combined with the original token-oriented cross-entropy loss, our model can consider both the target cluster and the target token. Even if the model predicts the wrong token index, with our rearranged codebook and cluster-oriented loss, there is a high probability that the token is located in the target cluster, which ensures a high similarity between the output image and the target one, effectively improving the generation quality and robustness (Fig. 1).", + "bbox": [ + 511, + 90, + 906, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We develop our model based on LlamaGen [36] and compare with existing visual generation methods based on GAN, Diffusion, Autoregressive (AR), and MIM models. Extensive experiments demonstrate that, across different parameter scales of LlamaGen (ranging from 100M to 1.4B), our method consistently enhances the model's training efficiency and performance. For the same FID, our approach reduces LlamaGen's training time by half, and under the same number of training epochs, it effectively improves the model's generation quality. Our method can be applied to various LLM-based visual generation models with almost no additional training costs, and adheres to the scaling law, providing a promising direction for improving future LLM-based visual generation models.", + "bbox": [ + 511, + 354, + 908, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as four-fold:", + "bbox": [ + 532, + 589, + 903, + 603 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose IAR, an Improved Autoregressive Visual Generation method, which leverages the correlation of visual embeddings to enhance the training efficiency and performance of LLM-based visual generation models.", + "- We introduce a Codebook Rearrangement strategy that uses balanced K-means clustering algorithm to rearrange the image codebook, ensuring high similarity across image embeddings within each cluster.", + "- We propose a Cluster-oriented Cross-entropy Loss that relaxes the original token-oriented cross-entropy, ensuring that even if the model predicts the wrong token index, there is still a high probability that the token is in the correct cluster, thereby generating high-quality images.", + "- Extensive experiments demonstrate the effectiveness of our method. Additionally, our approach can be applied to various LLM-based visual generation models, adheres to the scaling law, and provides a robust improvement direction for future LLM-based visual generation models." + ], + "bbox": [ + 513, + 626, + 903, + 896 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9352", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/69a30eba775aa2b4054b56551f5a35f65d9755ab4995b64007f00dd00b804a6a.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 96, + 92, + 421, + 273 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ea6d9b8101bc7b99ea2051ff15ce9bb9a8a95ce250bbb68bf7e658002a7e4f63.jpg", + "image_caption": [ + "(b)", + "Figure 2. (a) The MSE and LPIPS between the source image and the decoded image with different code distances. (b) The visualization of the images decoded from different code distances. When the code distance is within a certain range (e.g., smaller than 12), the decoded image looks nearly identical to the source image. We further make use of this property to improve the LLM-based visual generation model." + ], + "image_footnote": [], + "bbox": [ + 423, + 93, + 903, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 352, + 240, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Large language models", + "text_level": 1, + "bbox": [ + 89, + 376, + 305, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large language models (LLMs) [11, 32, 39] have achieved significant success in natural language processing (NLP). The encoder-only structure [9, 11, 28], exemplified by BERT [11], randomly masks portions of the text and predicts the masked tokens based on the unmasked ones. With the introduction of the GPT series [1, 5, 31, 32], the decoder-only architecture gained popularity, demonstrating remarkable capabilities in language processing through next-token prediction. Additionally, open-source models such as LLaMA [29] and PaLM [8] have further advanced the NLP field. The recent success of LLMs in natural language processing has also inspired researchers to apply these models to visual generation, revealing significant potential in multi-modal generation tasks.", + "bbox": [ + 88, + 398, + 483, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Visual Generation Models", + "text_level": 1, + "bbox": [ + 89, + 621, + 328, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continuous-valued Visual Generation. Generative Adversarial Networks (GANs) [13, 15, 25, 41, 42], consisting of a generator and discriminator, employ adversarial training to learn the target distribution. Diffusion Models [12, 14, 19-21, 30, 33] iteratively add and remove noise to approximate the real data distribution, using a Markov process for the forward path and learning the reverse denoising process. These models have achieved significant success in image generation and have been extended to video generation [3, 16, 22, 40], enhancing video quality. However, both GANs and Diffusion Models, which model continuous distributions, face challenges when integrating with discretely modeled LLMs. Thus, exploring discrete modeling methods for image generation is necessary.", + "bbox": [ + 89, + 643, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discrete-valued Visual Generation. Discrete-space visual generation models are mainly divided into autoregressive [13, 36, 38] and masked image modeling (MIM) [2, 6,", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "7] models, both requiring image quantization before generation. Autoregressive models follow the next-token generation approach of GPT [1, 32], predicting subsequent tokens one by one based on previous image tokens. LLaM-AGEN [36], a representative work, uses LLAMA [39] to model the image generation process. VAR [38] further enhances generation capability by proposing a next-scale image generation process. MIM methods [2, 6, 36, 38], inspired by BERT [11], randomly mask parts of the image tokens and predict the masked tokens using the unmasked ones. These methods directly draw inspiration from LLMs and rarely explore the fundamental differences between images and language, which may lead to suboptimal utilization of vision generation capabilities within the LLM framework. In contrast, our model makes full use of the image property, which effectively improves the training efficiency and model performance in LLM-based image generation.", + "bbox": [ + 511, + 353, + 906, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 628, + 604, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observing that an inherent difference between LLM-based image generation and text generation lies in that text generation only requires a predicted text index that can be directly mapped into the corresponding word, while image generation essentially requires the image embedding that corresponds to the index, and then decodes it into an image, we further explore the characteristics of the image embedding space. We conduct a detailed analysis on the similarity of the image embeddings in the codebook learned by image tokenizers (Sec. 3), and find that similar embeddings convey similar information in the image space: When the embeddings of certain image patches are replaced by similar image embeddings, the decoded images are nearly identical. Therefore, we leverage this property to enhance the LLM-based image generation model.", + "bbox": [ + 511, + 657, + 906, + 882 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first propose a Codebook Rearrangement strategy", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9353", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/def87f239eb34b064c68d5fd6ac613fe712380f11cbaef2ca6c9264cb41e78d4.jpg", + "image_caption": [ + "Stage 1: Codebook Rearrangement", + "Stage 2: Cluster-level constraint", + "Figure 3. Model framework: 1) Codebook Rearrangement: we first use a balanced K-means clustering method to rearrange the codebook, which divides the codebook into $n$ clusters, with the image codes in each cluster sharing a high similarity. 2) Cluster-oriented Constraint: During the training process, we first quantize the image patches using the rearranged codebook. For the output probability distribution $\\hat{Y}$ , we further compute the cluster-level distribution $\\hat{Y}_C$ by applying LogSumExp operation for the probabilities in each cluster $\\hat{Y}_{jm} \\sim \\hat{Y}_{(j+1)m-1}$ . Then we compute the cluster-oriented cross-entropy loss $\\mathcal{L}_{CCE}$ apart from the token-oriented cross-entropy loss $\\mathcal{L}_{TCE}$ , which ensures a high probability of the predicted token located in the correct cluster, thereby enhancing generation quality." + ], + "image_footnote": [], + "bbox": [ + 94, + 102, + 906, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that uses a balanced K-means clustering method to rearrange the embeddings in the codebook into clusters of equal sizes (Fig. 3 (left), Sec. 3.2), where the image embeddings in each cluster share a high similarity. Once the model predicts the correct cluster index (even if the token index is wrong), the corresponding embedding is located in the correct cluster and is thus similar to the target one, then the output images will be quite similar to the target image. To enable the model with the ability to predict the correct cluster index, we further propose a Cluster-oriented Cross-entropy Loss (Fig. 3 (right), Sec. 3.3). Our rearranged codebook and cluster-oriented cross-entropy loss ensure that, even if the model predicts incorrect tokens, there is a high probability of correctly predicting the cluster indices, which ensures high-quality generation results, largely improving the training efficiency and robustness of the LLM-based image generation model.", + "bbox": [ + 88, + 479, + 483, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Analysis on Image Embedding Similarity", + "text_level": 1, + "bbox": [ + 89, + 753, + 441, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first analyze the similarity of image embeddings in the codebook learned by image tokenizers, and find that similar embeddings (those that are close in distance) represent similar image information. I.e., when the embeddings of each image patch are replaced by other similar embeddings, the content of the decoded image remains almost unchanged. To verify this image embedding similarity property, we conduct detailed experiments on the codebook of", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VQGAN [13].", + "text_level": 1, + "bbox": [ + 513, + 479, + 611, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, an image $x$ is quantized into discrete tokens by: 1) using an image encoder $E$ (e.g., from VQGAN [13]) to extract features, obtaining $h \\times w$ image features of dimension $C$ , denoted as $\\hat{z} = E(x) \\in \\mathcal{R}^{h \\times w \\times C}$ ; 2) quantizing each feature $\\hat{z}^{(i,j)}$ into the code index $q^{(i,j)}$ of the nearest code in the codebook $\\mathcal{Z} \\in \\mathcal{R}^{N \\times C}$ :", + "bbox": [ + 511, + 494, + 906, + 584 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq ^ {(i, j)} = \\arg \\min _ {z _ {k} \\in \\mathcal {Z}} \\| \\hat {z} ^ {(i, j)} - z _ {k} \\| \\in [ 0, N), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 595, + 906, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{k}$ is the $k$ -th embedding in codebook $\\mathcal{Z}$ , and $\\| \\cdot \\|$ measures the Euclidean distance. And the quantized feature $z_{q} \\in \\mathcal{R}^{h \\times w \\times C}$ is obtained by looking up each $q^{(i,j)}$ in $\\mathcal{Z}$ .", + "bbox": [ + 511, + 632, + 905, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To measure the distance between two embeddings $z_{i}, z_{j}$ in the codebook, we define the \"code distance\" $D(z_{i}, z_{j})$ :", + "bbox": [ + 511, + 678, + 903, + 710 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD \\left(z _ {i}, z _ {j}\\right) := d, \\quad z _ {j} = \\left[ \\operatorname {s o r t} \\left(\\left\\{\\| z _ {k} - z _ {i} \\| \\right\\}\\right) \\right] _ {d}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 720, + 903, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "calculated by measuring the Euclidean distance of each $z_{k}$ to $z_{i}$ , sorting the distances, and $z_{j}$ is the $D(z_{i},z_{j})$ -th closest embedding to $z_{i}$ among all embeddings in the codebook. A lower \"code distance\" indicates the two embeddings are closer in the embedding space.", + "bbox": [ + 511, + 750, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an input image $x$ , the quantized embedding $z_{q}$ and a generated image $\\hat{x}^{\\prime} = G(z_{q}^{\\prime})$ decoded from $z_{q}^{\\prime}$ , we evaluate the similarity of the two images at different code distances $D(z_{q},z_{q}^{\\prime})$ . In Fig. 2 (a), we measure the image similarity by MSE distance and LPIPS [43], where lower", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9354", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Balanced k-means Clustering" + ], + "code_body": "Input: Codebook $\\mathcal{Z}$ , number of clusters $n$ \nOutput: The rearranged codebook $\\hat{\\mathcal{Z}}$ with local similarity \n1: Initialize: \n2: Compute the cluster size $m\\gets |\\mathcal{Z}| / n$ \n3:Randomly select $n$ codes as initial centroids $\\{c_j\\}_{j = 1}^n$ \n4:Initialize iteration counter iter $\\leftarrow 0$ \n5:repeat \n6:Compute the minimum distance $d_{i}$ from $z_{i}$ to $\\{c_j\\}_{j = 1}^n$ \n7:Sort $\\{z_i\\}$ in ascending order of $d_{i}$ to form $\\mathcal{Z}'$ \n8:for each data point $z_{i}\\in \\mathcal{Z}'$ do \n9:Compute the distance from $z_{i}$ to each centroid $c_{j}$ \n10:Assign $z_{i}$ to the nearest cluster whose size does not exceed $m$ \n11:end for \n12:for each cluster centroid $c_{j}$ do \n13:Update $c_{j}$ to the average of all embeddings assigned to cluster $j$ \n14:end for \n15:Increment iteration counter iter $\\leftarrow iter + 1$ \n16:until convergence or reaching max_iters \n17:foreach cluster centroid $c_{j}$ do \n18:Assign the embeddings in cluster $j$ to $\\hat{\\mathcal{Z}}_{[jm,(j + 1)m)}$ \n19:end for \n20:return The rearranged codebook $\\hat{\\mathcal{Z}}$", + "bbox": [ + 106, + 119, + 472, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scores indicate higher image similarity. Results show: 1) as the code distance increases, the image distance also increases; 2) but when the code distance is within a certain range (e.g., smaller than 12), the image distance (MSE 0.104 and LPIPS 66.44) does not increase too much compared to the reconstructed image $G(z_{q})$ (whose code distance is 0 and has a 0.076 MSE and 55.8 LPIPS). We further visualize the images decoded from different code distances in Fig. 2(b), which shows the decoded images with lower code distances (e.g., smaller than 12) look nearly identical to the source image and have a good visual quality. This indicates that even if the predicted token index is not the accurate target index, as long as the code distance between the corresponding embedding is within a certain range, the decoded image is similar to the desired one and shows a good visual quality. Therefore, we make use of this property to improve the generation quality and stability of the LLM-based visual generation model.", + "bbox": [ + 89, + 476, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Codebook Rearrangement", + "text_level": 1, + "bbox": [ + 89, + 757, + 331, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rearrangement Target. In a codebook learned from the VQGAN [13] training process, the embeddings are often randomly distributed, with adjacent embeddings having no particular correlation, i.e., Euclidean distances between adjacent embeddings can be either close or far. Therefore, we first propose a Codebook Rearrangement strategy to reorder the codebook, so that the similarity between adjacent embeddings is maximized, ensuring that the surrounding em", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "beddings of any given embedding exhibit high similarity.", + "bbox": [ + 511, + 90, + 888, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, denote the codebook as $\\mathcal{Z} = \\{z_i\\}_{i=1}^N$ which contains $N$ quantized embeddings $z_i$ . To maximize the similarity between the adjacent embeddings, we aim to find a surjective mapping $M(\\cdot)$ that satisfies:", + "bbox": [ + 511, + 106, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM = \\arg \\min _ {M} \\sum_ {i = 1} ^ {N - 1} \\| z _ {M (i)}, z _ {M (i + 1)} \\|. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 179, + 903, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "I.e., after reordering each embedding $z_{i}$ to index $M(i)$ , the sum of distances between adjacent embeddings is minimized. And $\\hat{\\mathcal{Z}} = M(\\mathcal{Z})$ is the rearranged codebook.", + "bbox": [ + 511, + 232, + 903, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, this optimization can be reduced to the Hamiltonian path problem (see #Suppl), and solving such a problem is NP-hard. Therefore, we try to relax this problem into a solvable form.", + "bbox": [ + 511, + 277, + 903, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Constraint Relaxation. We relax this optimization problem into a much easier one, where we only need to ensure the embeddings in a range are similar. Therefore, we split the codebook into $n$ clusters, with each cluster containing $m = \\frac{N}{n}$ embeddings that are similar to each other. Then we reorder the codebook so that the embeddings in the same cluster have close indices, with the indices of cluster $j$ are within the range $[jm, (j + 1)m)$ .", + "bbox": [ + 511, + 339, + 905, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Balanced K-means Clustering. To cluster the embeddings in the codebook, we design a balanced K-means clustering method to uniformly divide the codebook into $n$ clusters, with the embeddings in each cluster close to each other.", + "bbox": [ + 511, + 460, + 905, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Alg. 1, we first randomly select $n$ codes as the initial centroids $\\{c_j\\}_{j=1}^n$ . Then, we iteratively update each cluster. In each iteration, 1) we first compute the minimum distance $d_i$ from each embedding $z_i$ to the $n$ centroids $\\{c_j\\}_{j=1}^n$ . 2) We sort $\\{z_i\\}$ in ascending order of $d_i$ to form $\\mathcal{Z}'$ (i.e., embeddings closer to the nearest centroid has smaller index). This sorting and reordering step ensures that the embeddings closer to the nearest centroid are allocated first, and the farther ones are allocated later, which is necessary in cluster size balanced clustering. 3) Then, for each $z_i$ in the ordered codebook $\\mathcal{Z}'$ , we compute its distance to each centroid $c_j$ . 4) We assign $z_i$ to the nearest cluster whose size does not exceed $m$ (where $m = \\frac{N}{n}$ ). 5) Finally, we update each cluster centroid $c_j$ to the average of the embeddings assigned to that cluster, and repeat the above process until it converges or reaches the pre-defined maximum iteration. After obtaining the final cluster centroids $c_j$ and $n$ clusters, where each cluster contains $m$ embeddings, we rearrange the codebook $\\hat{\\mathcal{Z}}$ by assigning the embeddings in cluster $j$ to indices $[jm, (j+1)m]$ .", + "bbox": [ + 511, + 521, + 906, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Cluster-oriented Visual Generation", + "text_level": 1, + "bbox": [ + 511, + 832, + 820, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analysis on Token-oriented Cross-entropy Loss. In LLM-based visual generation models [36, 38], cross-entropy loss is the most commonly used loss in the training", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9355", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "process. Denote the ground truth one-hot vector as $Y$ ( $Y$ is a $N$ -dimensional vector, with $Y_{i=y} = 1$ and $Y_{i\\neq y} = 0$ , where $y$ is the class label of the token), and the predicted probability distribution as $\\hat{Y} \\in \\mathcal{R}^N(\\sum \\hat{Y}_i = 1)$ . The token-oriented cross-entropy loss $\\mathcal{L}_{TCE}$ is formulated as:", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T C E} = - \\sum_ {i = 1} ^ {n} Y _ {i} \\log \\hat {Y} _ {i}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 176, + 482, + 214 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, we observe that if a token index is wrongly predicted as a similar embedding's index, this token-oriented cross-entropy loss will penalize the case, but the decoded image is actually not much different from the target image.", + "bbox": [ + 89, + 222, + 483, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To tolerate this case, we leverage the rearranged codebook, where the embeddings in the same cluster are close to each other. In Sec. 3, it has been demonstrated that even if some image token indices are predicted incorrectly, as long as the embeddings of the predicted tokens are not far from the embeddings of the target tokens, the semantic integrity and quality of the generated image will not be significantly affected. Therefore, leveraging the reordered codebook from Sec. 3.2, a natural idea is to guide the model to first predict the cluster and then locate the specific token within that cluster. Since the number of clusters $(n)$ is much smaller than the size of the codebook $(N)$ , predicting the cluster of a token is a much easier task than directly predicting a specific token. Once the cluster is accurately predicted, even if the specific token prediction is incorrect, the quality of the generated image can still be maintained.", + "bbox": [ + 89, + 282, + 483, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cluster-oriented Cross-entropy Loss. To guide the model to predict the correct cluster of the next token, we propose a cluster-wise cross-entropy loss $\\mathcal{L}_C$ . Specifically, we first derive the ground truth cluster label $y_{c}$ from the class label $y^{1}$ by $y_{c} = \\left\\lfloor \\frac{y}{m}\\right\\rfloor$ , where $m$ is the number of tokens in each cluster. Then, to form a new cluster-level probability distribution $\\hat{Y}_{C}\\in \\mathcal{R}^{n}(\\sum \\hat{Y}_{C,i} = 1)$ from the token-level probability distribution $\\hat{Y}$ , we employ the following operation to calculate the probability of each cluster, where within a cluster, the samples with larger probability contribute more to the cluster probability:", + "bbox": [ + 89, + 523, + 483, + 691 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Y} _ {C, j} = \\sum_ {i = j m} ^ {(j + 1) m - 1} e x p (\\hat {Y} _ {i}) / \\sum_ {i = 1} ^ {N} e x p (\\hat {Y} _ {i}), \\quad j = 1, \\dots , n \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 700, + 482, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{Y}_{C,j}$ represents the probability that the generated token belongs to cluster $j$ .", + "bbox": [ + 89, + 739, + 482, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the ground-truth cluster label $y_{c}$ and the predicted cluster-level probability distribution $\\hat{Y}_{C,i}$ , our cluster-oriented cross-entropy loss $\\mathcal{L}_{CCE}$ is formulated as:", + "bbox": [ + 89, + 771, + 483, + 816 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C C E} = - \\sum_ {j = 1} ^ {n} Y _ {C, j} \\log \\hat {Y} _ {C, j}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 825, + 482, + 866 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0cf30b642b092670b3ca43947b836cbf2f1f641441d424014321fce7605d8bda.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeModel#Para.FID↓IS↑Precision↑Recall↑
GANBigGAN [4]112M6.95224.50.890.38
GigaGAN [23]569M3.45225.50.840.61
StyleGan-XL [35]166M2.30265.10.780.53
DiffusionADM [12]554M10.94101.00.690.63
CDM [20]-4.88158.7--
LDM-4 [33]400M3.60247.7--
DiT-XL/2 [30]675M2.27278.20.830.57
Mask.MaskGIT [6]227M6.18182.10.800.51
MaskGIT-re [6]227M4.02355.6--
ARVQGAN [13]227M18.6580.40.780.26
VQGAN [13]1.4B15.7874.3--
VQGAN-re [13]1.4B5.20280.3--
ViT-VQGAN [42]1.7B4.17175.1--
ViT-VQGAN-re [42]1.7B3.04227.4--
RQTran. [27]3.8B7.55134.0--
RQTran.-re [27]3.8B3.80323.7--
ARLlamaGen-B [36]111M5.46193.60.830.45
LlamaGen-L [36]343M3.29227.80.820.53
LlamaGen-XL [36]775M2.63244.10.810.58
LlamaGen-XXL [36]1.4B2.34253.90.800.59
OursIAR-B111M5.14202.00.850.45
IAR-L343M3.18234.80.820.53
IAR-XL775M2.52248.10.820.58
IAR-XXL (CFG=1.7)1.4B2.19278.90.810.58
IAR-XXL (CFG=2.5)1.4B3.74362.00.860.54
", + "bbox": [ + 517, + 88, + 908, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison between different types of image generation model on class-conditional ImageNet $256 \\times 256$ benchmark with FID, IS, precision, and recall.", + "bbox": [ + 511, + 387, + 906, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Y_{C}$ is the one-hot vector spanned by the label $y_{c}$ . With our cluster-oriented loss, even if the model predicts the wrong token index, there is a high probability that the token is located in the target cluster, which ensures a high similarity between the output image and the target one, effectively improving the generation quality and robustness.", + "bbox": [ + 511, + 443, + 905, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Final Loss Function. The final loss function $\\mathcal{L}$ combines both the token-oriented cross-entropy loss $\\mathcal{L}_{TCE}$ and the cluster-oriented cross-entropy loss $\\mathcal{L}_{CCE}$ :", + "bbox": [ + 511, + 534, + 905, + 580 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {T C E} + \\lambda \\mathcal {L} _ {C C E}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 593, + 903, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 623, + 645, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental settings", + "text_level": 1, + "bbox": [ + 511, + 648, + 718, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this paper, we choose LLamaGen [36] as the main baseline and apply our method to it. We follow the official training details and keep the hyperparameters the same as the official ones. We conduct experiments on ImageNet [10]. To evaluate the model performance, we generate 50K images with randomly selected labels. Then, we compute the FID [17], IS [34], precision, and recall [26] for the generated data (see #Suppl for details).", + "bbox": [ + 511, + 671, + 905, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison Experiments", + "text_level": 1, + "bbox": [ + 511, + 801, + 746, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with baselines. We compare our model with the GAN-based methods [4, 23, 35], diffusion-based methods [12, 20, 30, 33], mask-prediction-based methods [6] and autoregressive-based methods [13, 27, 36, 42] on ImageNet [10]. The results are shown in Tab. 2. Our method", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1After codebook rearrangement, the class label has been updated accordingly.", + "bbox": [ + 89, + 875, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9356", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fb15147df9387770131833564a2c4c6ccee8ee49e72f0845f3a3db07f158579c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TokensModel#Para.50 epoch300 epoch
FID↓IS↑Precision↑Recal↑FID↓IS↑Precision↑Recal↑
16 × 16LlamaGen-B111M7.22178.30.860.385.46193.60.840.46
LlamaGen-L343M4.20200.00.820.513.80248.30.830.52
LlamaGen-XL775M3.39227.10.810.54----
LlamaGen-XXL1.4B3.09253.60.830.53----
IAR-B111M6.90179.20.860.405.14202.00.850.45
IAR-L343M4.10207.10.820.513.40271.30.840.51
IAR-XL775M3.36228.90.820.54----
IAR-XXL1.4B3.01257.40.830.53----
24 × 24LlamaGen-B111M8.31154.70.840.386.09182.50.840.42
LlamaGen-L343M4.61191.40.820.503.29227.80.820.53
LlamaGen-XL775M3.24245.70.830.532.63244.10.810.58
LlamaGen-XXL1.4B2.89236.20.800.562.34253.90.810.60
IAR-B111M7.80153.30.840.395.77192.50.850.42
IAR-L343M4.35197.20.810.513.18234.80.820.53
IAR-XL775M3.15228.80.810.542.52248.10.820.58
IAR-XXL1.4B2.87249.90.820.562.19265.60.810.58
", + "bbox": [ + 174, + 88, + 823, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison with LlamaGen across different image tokens and model sizes. Following LlamaGen, we only train XL and XXL version on $16 \\times 16$ tokens for 50 epochs.", + "bbox": [ + 89, + 330, + 906, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieves the state-of-the-art FID (2.19) and IS (362.0). And for each model size, our IAR consistently outperforms LlamaGen, demonstrating the effectiveness of our cluster-oriented autoregressive visual generation strategy.", + "bbox": [ + 89, + 372, + 482, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More comparisons with LlamaGen. We further compare our model with the baseline LlamaGen across different model sizes and image tokens to validate the effectiveness of our model. Specifically, we choose four models with model sizes ranging from 111M to 1.4B parameters. And we also conduct experiments on $16 \\times 16$ image tokens (corresponds to $256 \\times 256$ image) and $24 \\times 24$ image tokens (corresponds to $384 \\times 384$ image). We choose the model trained with 50 epochs and 300 epochs for evaluation. The comparison results are shown in Tab. 2. It can be seen that across different model sizes, image tokens, and training epochs, our model can always improve the baseline with a better FID and IS, validating the effectiveness and robustness of our model in LLM-based visual generation.", + "bbox": [ + 88, + 438, + 482, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of classifier-free guidance (CFG). IAR can employ classifier-free guidance [18] to generate images with higher quality. We show the performance (FID & IS) of our model (111M and 343M parameters) on different CFGs ranging from 1.0 to 2.75 in Fig. 4 (a). It can be seen that compared to the model without CFG (i.e., CFG=1.0), the model with CFG can largely improve the FID, which means the model can generate images with much higher quality. Specifically, as the CFG increases, the FID first decreases and then increases, indicating that a too-large CFG may influence the generation quality, which is consistent with LlamaGen and previous diffusion-based methods [12, 18].", + "bbox": [ + 89, + 654, + 482, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of model size. We conduct experiments on different model sizes, where the parameter number ranges from 111M to 1.4B. It can be seen in Tab. 2 and Fig. 4 (b) that as the parameter number grows, the model shows", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a clear improvement in the performance, which aligns with the scaling law [24], indicating that as the number of parameter increases, our IVR can be more effective.", + "bbox": [ + 511, + 372, + 906, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training efficiency. We compare the training speed between IAR-B and LlamaGen-B (for different model sizes, please refer to #Suppl). We evaluate the model performance on different epochs (from 50 to 300). The FID and IS results are shown in Fig. 4 (c), which shows that our 175-epoch IAR reaches almost the same FID score as the 300-epoch LlamaGen, indicating that our model can accelerate the training speed for almost $42\\%$ . And comparing the models trained on 300 epochs, our model effectively improves the generation quality, validating that our model can improve the training efficiency of the LLM-based visual generation model.", + "bbox": [ + 511, + 417, + 908, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 606, + 666, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our model, we conduct several ablation studies to validate each module separately. We conduct the ablation studies based on IAR-B (111M parameters), and train the model for 100 epochs on ImageNet.", + "bbox": [ + 511, + 628, + 906, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on codebook rearrangement and $\\mathcal{L}_{CCE}$ . We first evaluate the effectiveness of the codebook rearrangement strategy and the cluster-oriented cross-entropy loss $\\mathcal{L}_{CCE}$ . We conduct experiments on three ablated models: the model without both codebook rearrangement and $\\mathcal{L}_{CCE}$ (original LlamaGen), the model without codebook rearrangement, and the model without $\\mathcal{L}_{CCE}$ . In contrast, our model with both codebook rearrangement and $\\mathcal{L}_{CCE}$ has the best performance in fitting the ImageNet dataset.", + "bbox": [ + 511, + 689, + 908, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on cluster size. A key hyperparameter in our model is the size of each cluster. A too-small cluster size may make it too hard for the model to predict the correct cluster index, which leads the model to perform similarly to the original LlamaGen (Note that the model with cluster", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9357", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5dc12d09f16e44eadcd270fb5733c0adb8ec2aba4fd5709faeaa5fad300336ee.jpg", + "image_caption": [ + "Figure 4. (a) Model performance (IAR-B and IAR-L) across different CFGs; (b) Model performance on different parameter numbers (111M to 3B) compared to LlamaGen; and (c) Model performance on different epochs compared to LlamaGen." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 356, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f41c403ce53d4e378d37709f3d649f468afde5a5535bd46c545bd0a28bbc0aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 85, + 609, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e3f341d1dcf73cb1a99833e368508285d9dd6bf269aa4024d217b1b754c8b34d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 85, + 903, + 222 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ef0a91469919cde5f886005968b644b3042f7a292ca2c768008aeb1ca93af65e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Codebook Rearrangement\\( \\mathcal{L}_{CCE} \\)FID↓IS↑Precision↑Recall↑
7.14166.380.840.40
✓7.15163.090.840.41
✓\\( 6.96^2 \\)168.700.840.40
✓✓6.77171.730.840.42
", + "bbox": [ + 94, + 260, + 478, + 353 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/53d47fb83454a1eec8e1466e26389c815ae9a5b5dcd4641f6291df21a0f16a1e.jpg", + "table_caption": [ + "Table 3. Ablation studies on codebook rearrangement strategy and the cluster-oriented cross-entropy loss $\\mathcal{L}_{\\mathrm{CCE}}$" + ], + "table_footnote": [], + "table_body": "
Cluster SizeInner MSEFID↓IS↑Precision↑Recall↑
107.14166.380.840.40
80.0187.02167.090.850.40
160.0226.99166.390.840.42
320.0286.83170.100.840.40
640.0346.92171.650.850.39
1280.0416.77171.730.840.42
2560.0506.81170.580.850.41
5120.0596.94163.540.830.42
", + "bbox": [ + 107, + 388, + 465, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "size 1 is the same as LlamaGen). In contrast, a too-big cluster size may cause the image embeddings in each cluster to be too different, such that even if the model predicts the correct cluster index, the gap to the target image embedding is still too big. Therefore, it is important to choose an appropriate cluster size to balance the inner similarity in each cluster and the ease of predicting the correct cluster index. We conduct experiments on different cluster sizes ranging from 1 to 512, where the codebook size is 16384. We compute the inner MSE of cluster, i.e., the average distance from each code in a cluster to the centroid, and evaluate the model performance by FID, IS, Precision, and Recall in Tab. 4. It can be seen that as the cluster size increases, the inner MSE distance becomes higher, indicating that the image codes in each cluster become less similar. Moreover, the model reaches the best performance when the cluster size is 128, which happens to be the square root of the codebook size.", + "bbox": [ + 88, + 559, + 482, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on weight of $\\mathcal{L}_{CCE}$ . Since we introduce an additional cluster-oriented cross-entropy loss $\\mathcal{L}_{CCE}$ during the training process, we conduct experiments to find the influence of its corresponding weight $\\lambda$ . We choose $\\lambda$ rang-", + "bbox": [ + 89, + 816, + 482, + 878 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/15a6362738ef18aaadda5e5639b00bd502af8ad8502b7912e9175a74c0c0bc62.jpg", + "table_caption": [ + "Table 4. Ablation studies on different cluster sizes from 1 to 512, where cluster size 1 corresponds to the official LlamaGen." + ], + "table_footnote": [], + "table_body": "
λFID↓IS↑Precision↑Recall↑
07.14166.380.840.40
0.17.01167.160.840.40
0.257.05160.730.840.42
0.56.78171.150.840.41
0.756.80170.060.840.41
16.77171.730.840.42
1.56.87170.600.840.41
", + "bbox": [ + 575, + 260, + 844, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation studies on the weight $\\lambda$ of the cluster-oriented cross-attention loss ${\\mathcal{L}}_{\\mathrm{{CCE}}}$ .", + "bbox": [ + 511, + 388, + 906, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing from 0.1 to 1.5 and evaluate the model performance in Tab. 5. It can be seen that a too-small weight will cause the model to perform similarly to the original LlamGen, while a too-big weight will make the model concern too much about the cluster index and overlook the ground-truth token index, which also makes the model perform worse. In summary, we find that when the weight $\\lambda = 1$ , the model can balance the two goals and achieves the best performance.", + "bbox": [ + 511, + 431, + 906, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 571, + 633, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we analyzed the difference between the natural language and images in LLM-based visual generation. We find that similar image embeddings in the codebook can produce similar images. Based on this observation, we propose IAR, an Improved Autoregressive Visual Generation Method that effectively enhances the training efficiency and quality of the LLM-based visual generation model. We first conduct Codebook Rearrangement with a balanced K-means clustering algorithm to reorder the codebook into clusters with equal sizes, where the image embeddings within each cluster share similarities. Then, we introduce a Cluster-oriented Cross-entropy Loss to enable the model to learn the target cluster, which can guarantee a good generation quality even if the model predicts a wrong image token. We validate the effectiveness of our IAR on LlamaGen, and find that IAR can largely improve the training efficiency and generation quality across different parameter scales. Moreover, our IAR can be applied to various LLM-based visual generation models, pointing out a new direction to improve the LLM-based visual generation field.", + "bbox": [ + 509, + 598, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "2For details, please refer to Sec. F of #Suppl.", + "bbox": [ + 107, + 886, + 349, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9358", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Shanghai Sailing Program (22YF1420300), National Natural Science Foundation of China (No. 62302297, 72192821, 62272447, 62472282, 62472285), Young Elite Scientists Sponsorship Program by CAST (2022QNRC001), the Fundamental Research Funds for the Central Universities (YG2023QNB17, YG2024QNA44), Beijing Natural Science Foundation (L222117).", + "bbox": [ + 89, + 113, + 485, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 253, + 187, + 268 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3", + "[2] Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Qing-Guo Chen, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing masked generative transformers for efficient high-resolution text-to-image synthesis. arXiv preprint arXiv:2410.08261, 2024. 1, 3", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 3", + "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 6", + "[5] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3", + "[6] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 1, 3, 6", + "[7] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, 2023. 3", + "[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. JMLR, 2023. 3", + "[9] Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. ELECTRA: Pre-training text encoders as discriminators rather than generators. In ICLR, 2020. 3", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6", + "[11] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3" + ], + "bbox": [ + 93, + 277, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3, 6, 7", + "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2, 3, 4, 5, 6", + "[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 3", + "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commun Acm, 2020. 1, 3", + "[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In ICLR, 2024. 1, 3", + "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6", + "[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 7", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3", + "[20] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 6", + "[21] Teng Hu, Jiangning Zhang, Liang Liu, Ran Yi, Siqi Kou, Haokun Zhu, Xu Chen, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Phasic content fusing diffusion model with directional distribution consistency for few-shot model adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2406-2415, 2023. 3", + "[22] Teng Hu, Jiangning Zhang, Ran Yi, Yating Wang, Hongrui Huang, Jieyu Weng, Yabiao Wang, and Lizhuang Ma. Motionmaster: Training-free camera motion transfer for video generation. arXiv preprint arXiv:2404.15789, 2024. 3", + "[23] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, 2023. 6", + "[24] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 7", + "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 3", + "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In NeurIPS, 2019. 6", + "[27] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 6" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9359", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3", + "[29] Shervin Minaee, Tomas Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Amatriain, and Jianfeng Gao. Large language models: A survey. arXiv preprint arXiv:2402.06196, 2024. 3", + "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 3, 6", + "[31] Alec Radford. Improving language understanding by generative pre-training. 2018. 3", + "[32] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 1, 2, 3", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3, 6", + "[34] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In NeurIPS, 2016. 6", + "[35] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganxl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH, 2022. 6", + "[36] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 3, 5, 6", + "[37] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1", + "[38] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. In NeurIPS, 2024. 1, 3, 5", + "[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3", + "[40] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2025.3", + "[41] Ran Yi, Teng Hu, Mengfei Xia, Yizhe Tang, and Yong-Jin Liu. Feditnet++: Few-shot editing of latent semantics in gan spaces with correlated attribute disentanglement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3", + "[42] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. In ICLR, 2021. 3, 6", + "[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4", + "[44] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe" + ], + "bbox": [ + 89, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1", + "bbox": [ + 545, + 92, + 903, + 133 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9360", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/2639ed20-911e-44d7-9142-a8fcfe0f6597_model.json b/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/2639ed20-911e-44d7-9142-a8fcfe0f6597_model.json new file mode 100644 index 0000000000000000000000000000000000000000..62b640857b918af169f5ed9717d2bd17258dfd2a --- /dev/null +++ b/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/2639ed20-911e-44d7-9142-a8fcfe0f6597_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.131, + 0.751, + 0.175 + ], + "angle": 0, + "content": "Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.203, + 0.806, + 0.294 + ], + "angle": 0, + "content": "Teng Hu\\(^{1*}\\), Jiangning Zhang\\(^{2,3*}\\), Ran Yi\\(^{1\\dagger}\\), Jieyu Weng\\(^{1}\\), Yabiao Wang\\(^{3,2}\\), Xianfang Zeng\\(^{3}\\), Zhucun Xue\\(^{3}\\), Lizhuang Ma\\(^{1}\\) \n\\(^{1}\\)Shanghai Jiao Tong University, \\(^{2}\\)Youtu Lab, Tencent, \\(^{3}\\)Zhejiang University {hu-teng, ranyi, w.jerry, lzma} @sjtu.edu.cn, {186368, yabiaowang, zzlongjuanfeng, 12432038} @zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.362, + 0.485, + 0.846 + ], + "angle": 0, + "content": "Employing LLMs for visual generation has recently become a research focus. However, the existing methods primarily transfer the LLM architecture to visual generation but rarely investigate the fundamental differences between language and vision. This oversight may lead to suboptimal utilization of visual generation capabilities within the LLM framework. In this paper, we explore the characteristics of visual embedding space under the LLM framework and discover that the correlation between visual embeddings can help achieve more stable and robust generation results. We present IAR, an Improved AutoRegressive Visual Generation Method that enhances the training efficiency and generation quality of LLM-based visual generation models. Firstly, we propose a Codebook Rearrange strategy that uses balanced \\( k \\)-means clustering algorithm to rearrange the visual codebook into clusters, ensuring high similarity among visual features within each cluster. Leveraging the rearranged codebook, we propose a Cluster-oriented Cross-entropy Loss that guides the model to correctly predict the cluster where the target token is located. This approach ensures that even if the model predicts the wrong token index, there is a high probability the predicted token is located in the correct cluster, which significantly enhances the generation quality and robustness. Extensive experiments demonstrate that our IAR consistently enhances the model training efficiency and performance from 100M to 1.4B, reducing the training time by half while achieving the same FID. Additionally, IAR can be applied to various LLM-based visual generation models and adheres to the scaling law, providing a promising direction for future research in LLM-based visual generation. The code is available at: https://github.com/sjtuplayer/IAR." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.327, + 0.907, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.526, + 0.907, + 0.584 + ], + "angle": 0, + "content": "Figure 1. When an autoregressive model predicts a wrong token, the previous methods [36, 38] may predict an irrelevant token that causes artifacts. Our IAR alleviates this issue by ensuring a high probability of the predicted token located in the correct cluster." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.599, + 0.645, + 0.614 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.626, + 0.907, + 0.822 + ], + "angle": 0, + "content": "With the development of generative models [15, 19, 33], a large number of outstanding visual generation methods have emerged, achieving considerable success in image and video generation [3, 16, 33]. Recently, researchers have begun exploring the integration of images and text to achieve a unified multi-modal model for image and text generation and understanding [37, 44]. Consequently, aligning image generation models with large language models (LLMs) [32] has become a key research focus. Recent studies have made progress by quantizing images and employing autoregressive or mask-prediction methods to predict discrete image tokens [2, 6, 36, 38], laying the research foundation for a unified world model in the future." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Unlike diffusion [19] or GAN [15] models that model image distribution in continuous space, autoregressive or masked image modeling (MIM) methods [2, 6, 36, 38] first convert images into discrete-valued tokens by image tokenizers, and then predict image tokens by autoregressive or" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9351" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "MIM approach. Specifically, autoregressive methods use the \"next-token prediction\" paradigm of GPT [32], while MIM methods adopt a similar training strategy to the mask prediction in BERT [11]. These methods draw inspiration from natural language models and transfer these techniques to image generation models, but rarely explore the fundamental differences between images and natural language." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.201, + 0.487, + 0.547 + ], + "angle": 0, + "content": "Due to the inherently discrete nature of words (each word comes from a finite vocabulary), different words can be directly mapped to different indices through a lookup table, thereby generating text by predicting the corresponding index of the target text. The standard language modeling objective \\( p(x_{t}|x_{1},\\ldots ,x_{t - 1}) \\) (each \\( x_{t} \\) is an index) then can be interpreted as a classification task and accurately modeled by LLMs. However, images are continuous-valued, and if the same objective is applied to image generation with \\( x_{t} \\) as a real number, it is difficult to accurately model this probability density. Therefore, image tokenizers such as VQGAN [13] are used to convert continuous images into discrete-valued tokens, and then the model predicts the image token index. Afterwards, the corresponding image embedding is retrieved from the codebook and decoded into an image. A fundamental difference between image and text is that, in text generation, the predicted text index can be directly mapped to the corresponding word through a lookup table; whereas LLM-based image generation essentially requires the image embedding that corresponds to the index, and then decodes it into an image. I.e., text only requires the index, while image requires the embedding corresponding to the index." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Different from the token index which is independent, i.e., the nearby indices have no correlation in word semantics, the image embeddings are located in a continuous feature space, where similar embeddings may correspond to similar image contents. This paper first investigates the impact of image embedding correlations on the generated results, and finds that similar embeddings convey similar information in the image space: When the embeddings of certain image patches are replaced by similar image embeddings, the decoded images are nearly identical. This insight inspires us to leverage the similarity of image embeddings to improve the existing LLM-based image generation techniques." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To capitalize on the correlation among image embeddings, we propose IAR, an Improved AutoRegressive Visual Generation Method, which can enhance the training efficiency and generation quality of the LLM-based visual generation model. Firstly, we propose a Codebook Rearrangement strategy that uses balanced K-means clustering algorithm to rearrange the embeddings in the codebook into clusters of equal sizes, where the embeddings in each cluster share high similarities and can be decoded into images with similar contents. Secondly, we observe that if a token index is wrongly predicted as a similar embedding's" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.334 + ], + "angle": 0, + "content": "index, the original token-oriented cross-entropy loss used in LLM will penalize this case, but the decoded image is actually not much different from the target image. To tolerate this case, based on the rearranged codebook, we further introduce a Cluster-oriented Cross-entropy Loss, which guides the model to predict the correct cluster that contains the target token, thereby providing the model with a broader perspective and no longer confined to a single target token. Combined with the original token-oriented cross-entropy loss, our model can consider both the target cluster and the target token. Even if the model predicts the wrong token index, with our rearranged codebook and cluster-oriented loss, there is a high probability that the token is located in the target cluster, which ensures a high similarity between the output image and the target one, effectively improving the generation quality and robustness (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.909, + 0.568 + ], + "angle": 0, + "content": "We develop our model based on LlamaGen [36] and compare with existing visual generation methods based on GAN, Diffusion, Autoregressive (AR), and MIM models. Extensive experiments demonstrate that, across different parameter scales of LlamaGen (ranging from 100M to 1.4B), our method consistently enhances the model's training efficiency and performance. For the same FID, our approach reduces LlamaGen's training time by half, and under the same number of training epochs, it effectively improves the model's generation quality. Our method can be applied to various LLM-based visual generation models with almost no additional training costs, and adheres to the scaling law, providing a promising direction for improving future LLM-based visual generation models." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.59, + 0.904, + 0.604 + ], + "angle": 0, + "content": "Our main contributions can be summarized as four-fold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.627, + 0.905, + 0.685 + ], + "angle": 0, + "content": "- We propose IAR, an Improved Autoregressive Visual Generation method, which leverages the correlation of visual embeddings to enhance the training efficiency and performance of LLM-based visual generation models." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.687, + 0.905, + 0.746 + ], + "angle": 0, + "content": "- We introduce a Codebook Rearrangement strategy that uses balanced K-means clustering algorithm to rearrange the image codebook, ensuring high similarity across image embeddings within each cluster." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.748, + 0.904, + 0.821 + ], + "angle": 0, + "content": "- We propose a Cluster-oriented Cross-entropy Loss that relaxes the original token-oriented cross-entropy, ensuring that even if the model predicts the wrong token index, there is still a high probability that the token is in the correct cluster, thereby generating high-quality images." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.823, + 0.904, + 0.897 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the effectiveness of our method. Additionally, our approach can be applied to various LLM-based visual generation models, adheres to the scaling law, and provides a robust improvement direction for future LLM-based visual generation models." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.627, + 0.905, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9352" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.093, + 0.422, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.276, + 0.274, + 0.289 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.094, + 0.904, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.276, + 0.7, + 0.288 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.299, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Figure 2. (a) The MSE and LPIPS between the source image and the decoded image with different code distances. (b) The visualization of the images decoded from different code distances. When the code distance is within a certain range (e.g., smaller than 12), the decoded image looks nearly identical to the source image. We further make use of this property to improve the LLM-based visual generation model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.353, + 0.241, + 0.368 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.377, + 0.306, + 0.394 + ], + "angle": 0, + "content": "2.1. Large language models" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.4, + 0.484, + 0.611 + ], + "angle": 0, + "content": "Large language models (LLMs) [11, 32, 39] have achieved significant success in natural language processing (NLP). The encoder-only structure [9, 11, 28], exemplified by BERT [11], randomly masks portions of the text and predicts the masked tokens based on the unmasked ones. With the introduction of the GPT series [1, 5, 31, 32], the decoder-only architecture gained popularity, demonstrating remarkable capabilities in language processing through next-token prediction. Additionally, open-source models such as LLaMA [29] and PaLM [8] have further advanced the NLP field. The recent success of LLMs in natural language processing has also inspired researchers to apply these models to visual generation, revealing significant potential in multi-modal generation tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.622, + 0.329, + 0.637 + ], + "angle": 0, + "content": "2.2. Visual Generation Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Continuous-valued Visual Generation. Generative Adversarial Networks (GANs) [13, 15, 25, 41, 42], consisting of a generator and discriminator, employ adversarial training to learn the target distribution. Diffusion Models [12, 14, 19-21, 30, 33] iteratively add and remove noise to approximate the real data distribution, using a Markov process for the forward path and learning the reverse denoising process. These models have achieved significant success in image generation and have been extended to video generation [3, 16, 22, 40], enhancing video quality. However, both GANs and Diffusion Models, which model continuous distributions, face challenges when integrating with discretely modeled LLMs. Thus, exploring discrete modeling methods for image generation is necessary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Discrete-valued Visual Generation. Discrete-space visual generation models are mainly divided into autoregressive [13, 36, 38] and masked image modeling (MIM) [2, 6," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.611 + ], + "angle": 0, + "content": "7] models, both requiring image quantization before generation. Autoregressive models follow the next-token generation approach of GPT [1, 32], predicting subsequent tokens one by one based on previous image tokens. LLaM-AGEN [36], a representative work, uses LLAMA [39] to model the image generation process. VAR [38] further enhances generation capability by proposing a next-scale image generation process. MIM methods [2, 6, 36, 38], inspired by BERT [11], randomly mask parts of the image tokens and predict the masked tokens using the unmasked ones. These methods directly draw inspiration from LLMs and rarely explore the fundamental differences between images and language, which may lead to suboptimal utilization of vision generation capabilities within the LLM framework. In contrast, our model makes full use of the image property, which effectively improves the training efficiency and model performance in LLM-based image generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.63, + 0.606, + 0.645 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.883 + ], + "angle": 0, + "content": "Observing that an inherent difference between LLM-based image generation and text generation lies in that text generation only requires a predicted text index that can be directly mapped into the corresponding word, while image generation essentially requires the image embedding that corresponds to the index, and then decodes it into an image, we further explore the characteristics of the image embedding space. We conduct a detailed analysis on the similarity of the image embeddings in the codebook learned by image tokenizers (Sec. 3), and find that similar embeddings convey similar information in the image space: When the embeddings of certain image patches are replaced by similar image embeddings, the decoded images are nearly identical. Therefore, we leverage this property to enhance the LLM-based image generation model." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "We first propose a Codebook Rearrangement strategy" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9353" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.172, + 0.089, + 0.388, + 0.102 + ], + "angle": 0, + "content": "Stage 1: Codebook Rearrangement" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.103, + 0.907, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.089, + 0.793, + 0.102 + ], + "angle": 0, + "content": "Stage 2: Cluster-level constraint" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.379, + 0.907, + 0.463 + ], + "angle": 0, + "content": "Figure 3. Model framework: 1) Codebook Rearrangement: we first use a balanced K-means clustering method to rearrange the codebook, which divides the codebook into \\( n \\) clusters, with the image codes in each cluster sharing a high similarity. 2) Cluster-oriented Constraint: During the training process, we first quantize the image patches using the rearranged codebook. For the output probability distribution \\( \\hat{Y} \\), we further compute the cluster-level distribution \\( \\hat{Y}_C \\) by applying LogSumExp operation for the probabilities in each cluster \\( \\hat{Y}_{jm} \\sim \\hat{Y}_{(j+1)m-1} \\). Then we compute the cluster-oriented cross-entropy loss \\( \\mathcal{L}_{CCE} \\) apart from the token-oriented cross-entropy loss \\( \\mathcal{L}_{TCE} \\), which ensures a high probability of the predicted token located in the correct cluster, thereby enhancing generation quality." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.48, + 0.485, + 0.737 + ], + "angle": 0, + "content": "that uses a balanced K-means clustering method to rearrange the embeddings in the codebook into clusters of equal sizes (Fig. 3 (left), Sec. 3.2), where the image embeddings in each cluster share a high similarity. Once the model predicts the correct cluster index (even if the token index is wrong), the corresponding embedding is located in the correct cluster and is thus similar to the target one, then the output images will be quite similar to the target image. To enable the model with the ability to predict the correct cluster index, we further propose a Cluster-oriented Cross-entropy Loss (Fig. 3 (right), Sec. 3.3). Our rearranged codebook and cluster-oriented cross-entropy loss ensure that, even if the model predicts incorrect tokens, there is a high probability of correctly predicting the cluster indices, which ensures high-quality generation results, largely improving the training efficiency and robustness of the LLM-based image generation model." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.755, + 0.442, + 0.772 + ], + "angle": 0, + "content": "3.1. Analysis on Image Embedding Similarity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We first analyze the similarity of image embeddings in the codebook learned by image tokenizers, and find that similar embeddings (those that are close in distance) represent similar image information. I.e., when the embeddings of each image patch are replaced by other similar embeddings, the content of the decoded image remains almost unchanged. To verify this image embedding similarity property, we conduct detailed experiments on the codebook of" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.48, + 0.612, + 0.495 + ], + "angle": 0, + "content": "VQGAN [13]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.907, + 0.585 + ], + "angle": 0, + "content": "Specifically, an image \\( x \\) is quantized into discrete tokens by: 1) using an image encoder \\( E \\) (e.g., from VQGAN [13]) to extract features, obtaining \\( h \\times w \\) image features of dimension \\( C \\), denoted as \\( \\hat{z} = E(x) \\in \\mathcal{R}^{h \\times w \\times C} \\); 2) quantizing each feature \\( \\hat{z}^{(i,j)} \\) into the code index \\( q^{(i,j)} \\) of the nearest code in the codebook \\( \\mathcal{Z} \\in \\mathcal{R}^{N \\times C} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.596, + 0.907, + 0.621 + ], + "angle": 0, + "content": "\\[\nq ^ {(i, j)} = \\arg \\min _ {z _ {k} \\in \\mathcal {Z}} \\| \\hat {z} ^ {(i, j)} - z _ {k} \\| \\in [ 0, N), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.633, + 0.906, + 0.679 + ], + "angle": 0, + "content": "where \\(z_{k}\\) is the \\(k\\)-th embedding in codebook \\(\\mathcal{Z}\\), and \\(\\| \\cdot \\|\\) measures the Euclidean distance. And the quantized feature \\(z_{q} \\in \\mathcal{R}^{h \\times w \\times C}\\) is obtained by looking up each \\(q^{(i,j)}\\) in \\(\\mathcal{Z}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.679, + 0.905, + 0.711 + ], + "angle": 0, + "content": "To measure the distance between two embeddings \\( z_{i}, z_{j} \\) in the codebook, we define the \"code distance\" \\( D(z_{i}, z_{j}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.721, + 0.905, + 0.739 + ], + "angle": 0, + "content": "\\[\nD \\left(z _ {i}, z _ {j}\\right) := d, \\quad z _ {j} = \\left[ \\operatorname {s o r t} \\left(\\left\\{\\| z _ {k} - z _ {i} \\| \\right\\}\\right) \\right] _ {d}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.825 + ], + "angle": 0, + "content": "calculated by measuring the Euclidean distance of each \\( z_{k} \\) to \\( z_{i} \\), sorting the distances, and \\( z_{j} \\) is the \\( D(z_{i},z_{j}) \\)-th closest embedding to \\( z_{i} \\) among all embeddings in the codebook. A lower \"code distance\" indicates the two embeddings are closer in the embedding space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Given an input image \\(x\\), the quantized embedding \\(z_{q}\\) and a generated image \\(\\hat{x}^{\\prime} = G(z_{q}^{\\prime})\\) decoded from \\(z_{q}^{\\prime}\\), we evaluate the similarity of the two images at different code distances \\(D(z_{q},z_{q}^{\\prime})\\). In Fig. 2 (a), we measure the image similarity by MSE distance and LPIPS [43], where lower" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9354" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.108, + 0.104, + 0.362, + 0.119 + ], + "angle": 0, + "content": "Algorithm 1 Balanced k-means Clustering" + }, + { + "type": "algorithm", + "bbox": [ + 0.107, + 0.121, + 0.473, + 0.449 + ], + "angle": 0, + "content": "Input: Codebook \\(\\mathcal{Z}\\), number of clusters \\(n\\) \nOutput: The rearranged codebook \\(\\hat{\\mathcal{Z}}\\) with local similarity \n1: Initialize: \n2: Compute the cluster size \\(m\\gets |\\mathcal{Z}| / n\\) \n3:Randomly select \\(n\\) codes as initial centroids \\(\\{c_j\\}_{j = 1}^n\\) \n4:Initialize iteration counter iter \\(\\leftarrow 0\\) \n5:repeat \n6:Compute the minimum distance \\(d_{i}\\) from \\(z_{i}\\) to \\(\\{c_j\\}_{j = 1}^n\\) \n7:Sort \\(\\{z_i\\}\\) in ascending order of \\(d_{i}\\) to form \\(\\mathcal{Z}'\\) \n8:for each data point \\(z_{i}\\in \\mathcal{Z}'\\) do \n9:Compute the distance from \\(z_{i}\\) to each centroid \\(c_{j}\\) \n10:Assign \\(z_{i}\\) to the nearest cluster whose size does not exceed \\(m\\) \n11:end for \n12:for each cluster centroid \\(c_{j}\\) do \n13:Update \\(c_{j}\\) to the average of all embeddings assigned to cluster \\(j\\) \n14:end for \n15:Increment iteration counter iter \\(\\leftarrow iter + 1\\) \n16:until convergence or reaching max_iters \n17:foreach cluster centroid \\(c_{j}\\) do \n18:Assign the embeddings in cluster \\(j\\) to \\(\\hat{\\mathcal{Z}}_{[jm,(j + 1)m)}\\) \n19:end for \n20:return The rearranged codebook \\(\\hat{\\mathcal{Z}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.485, + 0.75 + ], + "angle": 0, + "content": "scores indicate higher image similarity. Results show: 1) as the code distance increases, the image distance also increases; 2) but when the code distance is within a certain range (e.g., smaller than 12), the image distance (MSE 0.104 and LPIPS 66.44) does not increase too much compared to the reconstructed image \\( G(z_{q}) \\) (whose code distance is 0 and has a 0.076 MSE and 55.8 LPIPS). We further visualize the images decoded from different code distances in Fig. 2(b), which shows the decoded images with lower code distances (e.g., smaller than 12) look nearly identical to the source image and have a good visual quality. This indicates that even if the predicted token index is not the accurate target index, as long as the code distance between the corresponding embedding is within a certain range, the decoded image is similar to the desired one and shows a good visual quality. Therefore, we make use of this property to improve the generation quality and stability of the LLM-based visual generation model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.332, + 0.774 + ], + "angle": 0, + "content": "3.2. Codebook Rearrangement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Rearrangement Target. In a codebook learned from the VQGAN [13] training process, the embeddings are often randomly distributed, with adjacent embeddings having no particular correlation, i.e., Euclidean distances between adjacent embeddings can be either close or far. Therefore, we first propose a Codebook Rearrangement strategy to reorder the codebook, so that the similarity between adjacent embeddings is maximized, ensuring that the surrounding em" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.89, + 0.107 + ], + "angle": 0, + "content": "beddings of any given embedding exhibit high similarity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.168 + ], + "angle": 0, + "content": "Specifically, denote the codebook as \\(\\mathcal{Z} = \\{z_i\\}_{i=1}^N\\) which contains \\(N\\) quantized embeddings \\(z_i\\). To maximize the similarity between the adjacent embeddings, we aim to find a surjective mapping \\(M(\\cdot)\\) that satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.18, + 0.905, + 0.222 + ], + "angle": 0, + "content": "\\[\nM = \\arg \\min _ {M} \\sum_ {i = 1} ^ {N - 1} \\| z _ {M (i)}, z _ {M (i + 1)} \\|. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.233, + 0.905, + 0.278 + ], + "angle": 0, + "content": "I.e., after reordering each embedding \\( z_{i} \\) to index \\( M(i) \\), the sum of distances between adjacent embeddings is minimized. And \\( \\hat{\\mathcal{Z}} = M(\\mathcal{Z}) \\) is the rearranged codebook." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.905, + 0.338 + ], + "angle": 0, + "content": "However, this optimization can be reduced to the Hamiltonian path problem (see #Suppl), and solving such a problem is NP-hard. Therefore, we try to relax this problem into a solvable form." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.906, + 0.46 + ], + "angle": 0, + "content": "Constraint Relaxation. We relax this optimization problem into a much easier one, where we only need to ensure the embeddings in a range are similar. Therefore, we split the codebook into \\( n \\) clusters, with each cluster containing \\( m = \\frac{N}{n} \\) embeddings that are similar to each other. Then we reorder the codebook so that the embeddings in the same cluster have close indices, with the indices of cluster \\( j \\) are within the range \\( [jm, (j + 1)m) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.906, + 0.521 + ], + "angle": 0, + "content": "Balanced K-means Clustering. To cluster the embeddings in the codebook, we design a balanced K-means clustering method to uniformly divide the codebook into \\( n \\) clusters, with the embeddings in each cluster close to each other." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.825 + ], + "angle": 0, + "content": "As shown in Alg. 1, we first randomly select \\( n \\) codes as the initial centroids \\( \\{c_j\\}_{j=1}^n \\). Then, we iteratively update each cluster. In each iteration, 1) we first compute the minimum distance \\( d_i \\) from each embedding \\( z_i \\) to the \\( n \\) centroids \\( \\{c_j\\}_{j=1}^n \\). 2) We sort \\( \\{z_i\\} \\) in ascending order of \\( d_i \\) to form \\( \\mathcal{Z}' \\) (i.e., embeddings closer to the nearest centroid has smaller index). This sorting and reordering step ensures that the embeddings closer to the nearest centroid are allocated first, and the farther ones are allocated later, which is necessary in cluster size balanced clustering. 3) Then, for each \\( z_i \\) in the ordered codebook \\( \\mathcal{Z}' \\), we compute its distance to each centroid \\( c_j \\). 4) We assign \\( z_i \\) to the nearest cluster whose size does not exceed \\( m \\) (where \\( m = \\frac{N}{n} \\)). 5) Finally, we update each cluster centroid \\( c_j \\) to the average of the embeddings assigned to that cluster, and repeat the above process until it converges or reaches the pre-defined maximum iteration. After obtaining the final cluster centroids \\( c_j \\) and \\( n \\) clusters, where each cluster contains \\( m \\) embeddings, we rearrange the codebook \\( \\hat{\\mathcal{Z}} \\) by assigning the embeddings in cluster \\( j \\) to indices \\( [jm, (j+1)m] \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.821, + 0.849 + ], + "angle": 0, + "content": "3.3. Cluster-oriented Visual Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Analysis on Token-oriented Cross-entropy Loss. In LLM-based visual generation models [36, 38], cross-entropy loss is the most commonly used loss in the training" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9355" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.168 + ], + "angle": 0, + "content": "process. Denote the ground truth one-hot vector as \\( Y \\) (\\( Y \\) is a \\( N \\)-dimensional vector, with \\( Y_{i=y} = 1 \\) and \\( Y_{i\\neq y} = 0 \\), where \\( y \\) is the class label of the token), and the predicted probability distribution as \\( \\hat{Y} \\in \\mathcal{R}^N(\\sum \\hat{Y}_i = 1) \\). The token-oriented cross-entropy loss \\( \\mathcal{L}_{TCE} \\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.177, + 0.483, + 0.215 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T C E} = - \\sum_ {i = 1} ^ {n} Y _ {i} \\log \\hat {Y} _ {i}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.223, + 0.484, + 0.284 + ], + "angle": 0, + "content": "However, we observe that if a token index is wrongly predicted as a similar embedding's index, this token-oriented cross-entropy loss will penalize the case, but the decoded image is actually not much different from the target image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.284, + 0.484, + 0.525 + ], + "angle": 0, + "content": "To tolerate this case, we leverage the rearranged codebook, where the embeddings in the same cluster are close to each other. In Sec. 3, it has been demonstrated that even if some image token indices are predicted incorrectly, as long as the embeddings of the predicted tokens are not far from the embeddings of the target tokens, the semantic integrity and quality of the generated image will not be significantly affected. Therefore, leveraging the reordered codebook from Sec. 3.2, a natural idea is to guide the model to first predict the cluster and then locate the specific token within that cluster. Since the number of clusters \\((n)\\) is much smaller than the size of the codebook \\((N)\\), predicting the cluster of a token is a much easier task than directly predicting a specific token. Once the cluster is accurately predicted, even if the specific token prediction is incorrect, the quality of the generated image can still be maintained." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.693 + ], + "angle": 0, + "content": "Cluster-oriented Cross-entropy Loss. To guide the model to predict the correct cluster of the next token, we propose a cluster-wise cross-entropy loss \\(\\mathcal{L}_C\\). Specifically, we first derive the ground truth cluster label \\(y_{c}\\) from the class label \\(y^{1}\\) by \\(y_{c} = \\left\\lfloor \\frac{y}{m}\\right\\rfloor\\), where \\(m\\) is the number of tokens in each cluster. Then, to form a new cluster-level probability distribution \\(\\hat{Y}_{C}\\in \\mathcal{R}^{n}(\\sum \\hat{Y}_{C,i} = 1)\\) from the token-level probability distribution \\(\\hat{Y}\\), we employ the following operation to calculate the probability of each cluster, where within a cluster, the samples with larger probability contribute more to the cluster probability:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.702, + 0.483, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\hat {Y} _ {C, j} = \\sum_ {i = j m} ^ {(j + 1) m - 1} e x p (\\hat {Y} _ {i}) / \\sum_ {i = 1} ^ {N} e x p (\\hat {Y} _ {i}), \\quad j = 1, \\dots , n \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.483, + 0.772 + ], + "angle": 0, + "content": "where \\(\\hat{Y}_{C,j}\\) represents the probability that the generated token belongs to cluster \\(j\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.484, + 0.817 + ], + "angle": 0, + "content": "With the ground-truth cluster label \\( y_{c} \\) and the predicted cluster-level probability distribution \\( \\hat{Y}_{C,i} \\), our cluster-oriented cross-entropy loss \\( \\mathcal{L}_{CCE} \\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.826, + 0.483, + 0.867 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C C E} = - \\sum_ {j = 1} ^ {n} Y _ {C, j} \\log \\hat {Y} _ {C, j}, \\tag {6}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.484, + 0.902 + ], + "angle": 0, + "content": "1After codebook rearrangement, the class label has been updated accordingly." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.909, + 0.386 + ], + "angle": 0, + "content": "
TypeModel#Para.FID↓IS↑Precision↑Recall↑
GANBigGAN [4]112M6.95224.50.890.38
GigaGAN [23]569M3.45225.50.840.61
StyleGan-XL [35]166M2.30265.10.780.53
DiffusionADM [12]554M10.94101.00.690.63
CDM [20]-4.88158.7--
LDM-4 [33]400M3.60247.7--
DiT-XL/2 [30]675M2.27278.20.830.57
Mask.MaskGIT [6]227M6.18182.10.800.51
MaskGIT-re [6]227M4.02355.6--
ARVQGAN [13]227M18.6580.40.780.26
VQGAN [13]1.4B15.7874.3--
VQGAN-re [13]1.4B5.20280.3--
ViT-VQGAN [42]1.7B4.17175.1--
ViT-VQGAN-re [42]1.7B3.04227.4--
RQTran. [27]3.8B7.55134.0--
RQTran.-re [27]3.8B3.80323.7--
ARLlamaGen-B [36]111M5.46193.60.830.45
LlamaGen-L [36]343M3.29227.80.820.53
LlamaGen-XL [36]775M2.63244.10.810.58
LlamaGen-XXL [36]1.4B2.34253.90.800.59
OursIAR-B111M5.14202.00.850.45
IAR-L343M3.18234.80.820.53
IAR-XL775M2.52248.10.820.58
IAR-XXL (CFG=1.7)1.4B2.19278.90.810.58
IAR-XXL (CFG=2.5)1.4B3.74362.00.860.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.388, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Table 1. Comparison between different types of image generation model on class-conditional ImageNet \\(256 \\times 256\\) benchmark with FID, IS, precision, and recall." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.906, + 0.534 + ], + "angle": 0, + "content": "where \\( Y_{C} \\) is the one-hot vector spanned by the label \\( y_{c} \\). With our cluster-oriented loss, even if the model predicts the wrong token index, there is a high probability that the token is located in the target cluster, which ensures a high similarity between the output image and the target one, effectively improving the generation quality and robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.581 + ], + "angle": 0, + "content": "Final Loss Function. The final loss function \\(\\mathcal{L}\\) combines both the token-oriented cross-entropy loss \\(\\mathcal{L}_{TCE}\\) and the cluster-oriented cross-entropy loss \\(\\mathcal{L}_{CCE}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.594, + 0.905, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {T C E} + \\lambda \\mathcal {L} _ {C C E}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.624, + 0.646, + 0.64 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.719, + 0.666 + ], + "angle": 0, + "content": "4.1. Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.793 + ], + "angle": 0, + "content": "In this paper, we choose LLamaGen [36] as the main baseline and apply our method to it. We follow the official training details and keep the hyperparameters the same as the official ones. We conduct experiments on ImageNet [10]. To evaluate the model performance, we generate 50K images with randomly selected labels. Then, we compute the FID [17], IS [34], precision, and recall [26] for the generated data (see #Suppl for details)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.748, + 0.82 + ], + "angle": 0, + "content": "4.2. Comparison Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Comparison with baselines. We compare our model with the GAN-based methods [4, 23, 35], diffusion-based methods [12, 20, 30, 33], mask-prediction-based methods [6] and autoregressive-based methods [13, 27, 36, 42] on ImageNet [10]. The results are shown in Tab. 2. Our method" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9356" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.175, + 0.089, + 0.825, + 0.329 + ], + "angle": 0, + "content": "
TokensModel#Para.50 epoch300 epoch
FID↓IS↑Precision↑Recal↑FID↓IS↑Precision↑Recal↑
16 × 16LlamaGen-B111M7.22178.30.860.385.46193.60.840.46
LlamaGen-L343M4.20200.00.820.513.80248.30.830.52
LlamaGen-XL775M3.39227.10.810.54----
LlamaGen-XXL1.4B3.09253.60.830.53----
IAR-B111M6.90179.20.860.405.14202.00.850.45
IAR-L343M4.10207.10.820.513.40271.30.840.51
IAR-XL775M3.36228.90.820.54----
IAR-XXL1.4B3.01257.40.830.53----
24 × 24LlamaGen-B111M8.31154.70.840.386.09182.50.840.42
LlamaGen-L343M4.61191.40.820.503.29227.80.820.53
LlamaGen-XL775M3.24245.70.830.532.63244.10.810.58
LlamaGen-XXL1.4B2.89236.20.800.562.34253.90.810.60
IAR-B111M7.80153.30.840.395.77192.50.850.42
IAR-L343M4.35197.20.810.513.18234.80.820.53
IAR-XL775M3.15228.80.810.542.52248.10.820.58
IAR-XXL1.4B2.87249.90.820.562.19265.60.810.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.332, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Table 2. Comparison with LlamaGen across different image tokens and model sizes. Following LlamaGen, we only train XL and XXL version on \\(16 \\times 16\\) tokens for 50 epochs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.483, + 0.435 + ], + "angle": 0, + "content": "achieves the state-of-the-art FID (2.19) and IS (362.0). And for each model size, our IAR consistently outperforms LlamaGen, demonstrating the effectiveness of our cluster-oriented autoregressive visual generation strategy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.483, + 0.65 + ], + "angle": 0, + "content": "More comparisons with LlamaGen. We further compare our model with the baseline LlamaGen across different model sizes and image tokens to validate the effectiveness of our model. Specifically, we choose four models with model sizes ranging from 111M to 1.4B parameters. And we also conduct experiments on \\(16 \\times 16\\) image tokens (corresponds to \\(256 \\times 256\\) image) and \\(24 \\times 24\\) image tokens (corresponds to \\(384 \\times 384\\) image). We choose the model trained with 50 epochs and 300 epochs for evaluation. The comparison results are shown in Tab. 2. It can be seen that across different model sizes, image tokens, and training epochs, our model can always improve the baseline with a better FID and IS, validating the effectiveness and robustness of our model in LLM-based visual generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Effects of classifier-free guidance (CFG). IAR can employ classifier-free guidance [18] to generate images with higher quality. We show the performance (FID & IS) of our model (111M and 343M parameters) on different CFGs ranging from 1.0 to 2.75 in Fig. 4 (a). It can be seen that compared to the model without CFG (i.e., CFG=1.0), the model with CFG can largely improve the FID, which means the model can generate images with much higher quality. Specifically, as the CFG increases, the FID first decreases and then increases, indicating that a too-large CFG may influence the generation quality, which is consistent with LlamaGen and previous diffusion-based methods [12, 18]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Effects of model size. We conduct experiments on different model sizes, where the parameter number ranges from 111M to 1.4B. It can be seen in Tab. 2 and Fig. 4 (b) that as the parameter number grows, the model shows" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.907, + 0.418 + ], + "angle": 0, + "content": "a clear improvement in the performance, which aligns with the scaling law [24], indicating that as the number of parameter increases, our IVR can be more effective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.909, + 0.6 + ], + "angle": 0, + "content": "Training efficiency. We compare the training speed between IAR-B and LlamaGen-B (for different model sizes, please refer to #Suppl). We evaluate the model performance on different epochs (from 50 to 300). The FID and IS results are shown in Fig. 4 (c), which shows that our 175-epoch IAR reaches almost the same FID score as the 300-epoch LlamaGen, indicating that our model can accelerate the training speed for almost \\(42\\%\\). And comparing the models trained on 300 epochs, our model effectively improves the generation quality, validating that our model can improve the training efficiency of the LLM-based visual generation model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.667, + 0.623 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.69 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our model, we conduct several ablation studies to validate each module separately. We conduct the ablation studies based on IAR-B (111M parameters), and train the model for 100 epochs on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.827 + ], + "angle": 0, + "content": "Ablation on codebook rearrangement and \\(\\mathcal{L}_{CCE}\\). We first evaluate the effectiveness of the codebook rearrangement strategy and the cluster-oriented cross-entropy loss \\(\\mathcal{L}_{CCE}\\). We conduct experiments on three ablated models: the model without both codebook rearrangement and \\(\\mathcal{L}_{CCE}\\) (original LlamaGen), the model without codebook rearrangement, and the model without \\(\\mathcal{L}_{CCE}\\). In contrast, our model with both codebook rearrangement and \\(\\mathcal{L}_{CCE}\\) has the best performance in fitting the ImageNet dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Ablation on cluster size. A key hyperparameter in our model is the size of each cluster. A too-small cluster size may make it too hard for the model to predict the correct cluster index, which leads the model to perform similarly to the original LlamaGen (Note that the model with cluster" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9357" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.357, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.087, + 0.61, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.087, + 0.905, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.225, + 0.905, + 0.253 + ], + "angle": 0, + "content": "Figure 4. (a) Model performance (IAR-B and IAR-L) across different CFGs; (b) Model performance on different parameter numbers (111M to 3B) compared to LlamaGen; and (c) Model performance on different epochs compared to LlamaGen." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.261, + 0.48, + 0.354 + ], + "angle": 0, + "content": "
Codebook Rearrangement\\( \\mathcal{L}_{CCE} \\)FID↓IS↑Precision↑Recall↑
7.14166.380.840.40
✓7.15163.090.840.41
✓\\( 6.96^2 \\)168.700.840.40
✓✓6.77171.730.840.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.356, + 0.482, + 0.384 + ], + "angle": 0, + "content": "Table 3. Ablation studies on codebook rearrangement strategy and the cluster-oriented cross-entropy loss \\(\\mathcal{L}_{\\mathrm{CCE}}\\)" + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.389, + 0.466, + 0.515 + ], + "angle": 0, + "content": "
Cluster SizeInner MSEFID↓IS↑Precision↑Recall↑
107.14166.380.840.40
80.0187.02167.090.850.40
160.0226.99166.390.840.42
320.0286.83170.100.840.40
640.0346.92171.650.850.39
1280.0416.77171.730.840.42
2560.0506.81170.580.850.41
5120.0596.94163.540.830.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.519, + 0.483, + 0.547 + ], + "angle": 0, + "content": "Table 4. Ablation studies on different cluster sizes from 1 to 512, where cluster size 1 corresponds to the official LlamaGen." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.56, + 0.483, + 0.816 + ], + "angle": 0, + "content": "size 1 is the same as LlamaGen). In contrast, a too-big cluster size may cause the image embeddings in each cluster to be too different, such that even if the model predicts the correct cluster index, the gap to the target image embedding is still too big. Therefore, it is important to choose an appropriate cluster size to balance the inner similarity in each cluster and the ease of predicting the correct cluster index. We conduct experiments on different cluster sizes ranging from 1 to 512, where the codebook size is 16384. We compute the inner MSE of cluster, i.e., the average distance from each code in a cluster to the centroid, and evaluate the model performance by FID, IS, Precision, and Recall in Tab. 4. It can be seen that as the cluster size increases, the inner MSE distance becomes higher, indicating that the image codes in each cluster become less similar. Moreover, the model reaches the best performance when the cluster size is 128, which happens to be the square root of the codebook size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Ablation on weight of \\(\\mathcal{L}_{CCE}\\). Since we introduce an additional cluster-oriented cross-entropy loss \\(\\mathcal{L}_{CCE}\\) during the training process, we conduct experiments to find the influence of its corresponding weight \\(\\lambda\\). We choose \\(\\lambda\\) rang-" + }, + { + "type": "table", + "bbox": [ + 0.576, + 0.261, + 0.846, + 0.382 + ], + "angle": 0, + "content": "
λFID↓IS↑Precision↑Recall↑
07.14166.380.840.40
0.17.01167.160.840.40
0.257.05160.730.840.42
0.56.78171.150.840.41
0.756.80170.060.840.41
16.77171.730.840.42
1.56.87170.600.840.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.389, + 0.907, + 0.417 + ], + "angle": 0, + "content": "Table 5. Ablation studies on the weight \\( \\lambda \\) of the cluster-oriented cross-attention loss \\( {\\mathcal{L}}_{\\mathrm{{CCE}}} \\) ." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.555 + ], + "angle": 0, + "content": "ing from 0.1 to 1.5 and evaluate the model performance in Tab. 5. It can be seen that a too-small weight will cause the model to perform similarly to the original LlamGen, while a too-big weight will make the model concern too much about the cluster index and overlook the ground-truth token index, which also makes the model perform worse. In summary, we find that when the weight \\(\\lambda = 1\\), the model can balance the two goals and achieves the best performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.634, + 0.587 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we analyzed the difference between the natural language and images in LLM-based visual generation. We find that similar image embeddings in the codebook can produce similar images. Based on this observation, we propose IAR, an Improved Autoregressive Visual Generation Method that effectively enhances the training efficiency and quality of the LLM-based visual generation model. We first conduct Codebook Rearrangement with a balanced K-means clustering algorithm to reorder the codebook into clusters with equal sizes, where the image embeddings within each cluster share similarities. Then, we introduce a Cluster-oriented Cross-entropy Loss to enable the model to learn the target cluster, which can guarantee a good generation quality even if the model predicts a wrong image token. We validate the effectiveness of our IAR on LlamaGen, and find that IAR can largely improve the training efficiency and generation quality across different parameter scales. Moreover, our IAR can be applied to various LLM-based visual generation models, pointing out a new direction to improve the LLM-based visual generation field." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.351, + 0.901 + ], + "angle": 0, + "content": "2For details, please refer to Sec. F of #Suppl." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9358" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.228 + ], + "angle": 0, + "content": "This work was supported by Shanghai Sailing Program (22YF1420300), National Natural Science Foundation of China (No. 62302297, 72192821, 62272447, 62472282, 62472285), Young Elite Scientists Sponsorship Program by CAST (2022QNRC001), the Fundamental Research Funds for the Central Universities (YG2023QNB17, YG2024QNA44), Beijing Natural Science Foundation (L222117)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.254, + 0.188, + 0.27 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.279, + 0.484, + 0.348 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.35, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[2] Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Qing-Guo Chen, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing masked generative transformers for efficient high-resolution text-to-image synthesis. arXiv preprint arXiv:2410.08261, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.421, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.491, + 0.484, + 0.533 + ], + "angle": 0, + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.482, + 0.561 + ], + "angle": 0, + "content": "[5] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.563, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[6] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.606, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[7] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. JMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[9] Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. ELECTRA: Pre-training text encoders as discriminators rather than generators. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.279, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.121 + ], + "angle": 0, + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.906, + 0.163 + ], + "angle": 0, + "content": "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commun Acm, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.906, + 0.36 + ], + "angle": 0, + "content": "[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In ICLR, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.906, + 0.476 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.906, + 0.518 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.52, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[21] Teng Hu, Jiangning Zhang, Liang Liu, Ran Yi, Siqi Kou, Haokun Zhu, Xu Chen, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Phasic content fusing diffusion model with directional distribution consistency for few-shot model adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2406-2415, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[22] Teng Hu, Jiangning Zhang, Ran Yi, Yating Wang, Hongrui Huang, Jieyu Weng, Yabiao Wang, and Lizhuang Ma. Motionmaster: Training-free camera motion transfer for video generation. arXiv preprint arXiv:2404.15789, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[23] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.906, + 0.771 + ], + "angle": 0, + "content": "[24] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In NeurIPS, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[27] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9359" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[28] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[29] Shervin Minaee, Tomas Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Amatriain, and Jianfeng Gao. Large language models: A survey. arXiv preprint arXiv:2402.06196, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.179, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.208, + 0.483, + 0.235 + ], + "angle": 0, + "content": "[31] Alec Radford. Improving language understanding by generative pre-training. 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.237, + 0.483, + 0.277 + ], + "angle": 0, + "content": "[32] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.28, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.322, + 0.482, + 0.363 + ], + "angle": 0, + "content": "[34] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In NeurIPS, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.365, + 0.482, + 0.405 + ], + "angle": 0, + "content": "[35] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganxl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.408, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[36] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.464, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[37] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.507, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[38] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. In NeurIPS, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.564, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.634, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[40] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[41] Ran Yi, Teng Hu, Mengfei Xia, Yizhe Tang, and Yong-Jin Liu. Feditnet++: Few-shot editing of latent semantics in gan spaces with correlated attribute disentanglement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[42] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. In ICLR, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9360" + } + ] +] \ No newline at end of file diff --git a/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/2639ed20-911e-44d7-9142-a8fcfe0f6597_origin.pdf b/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/2639ed20-911e-44d7-9142-a8fcfe0f6597_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8924a7e0cadfb98aceb0925703847dc8930007df --- /dev/null +++ b/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/2639ed20-911e-44d7-9142-a8fcfe0f6597_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6e695bd9dc748c69f11b436ec6365cdba8b766147a924b357c0e62d8e76ac74 +size 3252137 diff --git a/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/full.md b/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..159f468b11a4f066f6a03ef60b13245025096a11 --- /dev/null +++ b/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/full.md @@ -0,0 +1,299 @@ +# Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction + +Teng Hu $^{1*}$ , Jiangning Zhang $^{2,3*}$ , Ran Yi $^{1\dagger}$ , Jieyu Weng $^{1}$ , Yabiao Wang $^{3,2}$ , Xianfang Zeng $^{3}$ , Zhucun Xue $^{3}$ , Lizhuang Ma $^{1}$ $^{1}$ Shanghai Jiao Tong University, $^{2}$ Youtu Lab, Tencent, $^{3}$ Zhejiang University {hu-teng, ranyi, w.jerry, lzma} @sjtu.edu.cn, {186368, yabiaowang, zzlongjuanfeng, 12432038} @zju.edu.cn + +# Abstract + +Employing LLMs for visual generation has recently become a research focus. However, the existing methods primarily transfer the LLM architecture to visual generation but rarely investigate the fundamental differences between language and vision. This oversight may lead to suboptimal utilization of visual generation capabilities within the LLM framework. In this paper, we explore the characteristics of visual embedding space under the LLM framework and discover that the correlation between visual embeddings can help achieve more stable and robust generation results. We present IAR, an Improved AutoRegressive Visual Generation Method that enhances the training efficiency and generation quality of LLM-based visual generation models. Firstly, we propose a Codebook Rearrange strategy that uses balanced $k$ -means clustering algorithm to rearrange the visual codebook into clusters, ensuring high similarity among visual features within each cluster. Leveraging the rearranged codebook, we propose a Cluster-oriented Cross-entropy Loss that guides the model to correctly predict the cluster where the target token is located. This approach ensures that even if the model predicts the wrong token index, there is a high probability the predicted token is located in the correct cluster, which significantly enhances the generation quality and robustness. Extensive experiments demonstrate that our IAR consistently enhances the model training efficiency and performance from 100M to 1.4B, reducing the training time by half while achieving the same FID. Additionally, IAR can be applied to various LLM-based visual generation models and adheres to the scaling law, providing a promising direction for future research in LLM-based visual generation. The code is available at: https://github.com/sjtuplayer/IAR. + +![](images/05c7aee6e1d6d40473115d53426ce46150c7c488c6ca9b40865216aff4112f71.jpg) +Figure 1. When an autoregressive model predicts a wrong token, the previous methods [36, 38] may predict an irrelevant token that causes artifacts. Our IAR alleviates this issue by ensuring a high probability of the predicted token located in the correct cluster. + +# 1. Introduction + +With the development of generative models [15, 19, 33], a large number of outstanding visual generation methods have emerged, achieving considerable success in image and video generation [3, 16, 33]. Recently, researchers have begun exploring the integration of images and text to achieve a unified multi-modal model for image and text generation and understanding [37, 44]. Consequently, aligning image generation models with large language models (LLMs) [32] has become a key research focus. Recent studies have made progress by quantizing images and employing autoregressive or mask-prediction methods to predict discrete image tokens [2, 6, 36, 38], laying the research foundation for a unified world model in the future. + +Unlike diffusion [19] or GAN [15] models that model image distribution in continuous space, autoregressive or masked image modeling (MIM) methods [2, 6, 36, 38] first convert images into discrete-valued tokens by image tokenizers, and then predict image tokens by autoregressive or + +MIM approach. Specifically, autoregressive methods use the "next-token prediction" paradigm of GPT [32], while MIM methods adopt a similar training strategy to the mask prediction in BERT [11]. These methods draw inspiration from natural language models and transfer these techniques to image generation models, but rarely explore the fundamental differences between images and natural language. + +Due to the inherently discrete nature of words (each word comes from a finite vocabulary), different words can be directly mapped to different indices through a lookup table, thereby generating text by predicting the corresponding index of the target text. The standard language modeling objective $p(x_{t}|x_{1},\ldots ,x_{t - 1})$ (each $x_{t}$ is an index) then can be interpreted as a classification task and accurately modeled by LLMs. However, images are continuous-valued, and if the same objective is applied to image generation with $x_{t}$ as a real number, it is difficult to accurately model this probability density. Therefore, image tokenizers such as VQGAN [13] are used to convert continuous images into discrete-valued tokens, and then the model predicts the image token index. Afterwards, the corresponding image embedding is retrieved from the codebook and decoded into an image. A fundamental difference between image and text is that, in text generation, the predicted text index can be directly mapped to the corresponding word through a lookup table; whereas LLM-based image generation essentially requires the image embedding that corresponds to the index, and then decodes it into an image. I.e., text only requires the index, while image requires the embedding corresponding to the index. + +Different from the token index which is independent, i.e., the nearby indices have no correlation in word semantics, the image embeddings are located in a continuous feature space, where similar embeddings may correspond to similar image contents. This paper first investigates the impact of image embedding correlations on the generated results, and finds that similar embeddings convey similar information in the image space: When the embeddings of certain image patches are replaced by similar image embeddings, the decoded images are nearly identical. This insight inspires us to leverage the similarity of image embeddings to improve the existing LLM-based image generation techniques. + +To capitalize on the correlation among image embeddings, we propose IAR, an Improved AutoRegressive Visual Generation Method, which can enhance the training efficiency and generation quality of the LLM-based visual generation model. Firstly, we propose a Codebook Rearrangement strategy that uses balanced K-means clustering algorithm to rearrange the embeddings in the codebook into clusters of equal sizes, where the embeddings in each cluster share high similarities and can be decoded into images with similar contents. Secondly, we observe that if a token index is wrongly predicted as a similar embedding's + +index, the original token-oriented cross-entropy loss used in LLM will penalize this case, but the decoded image is actually not much different from the target image. To tolerate this case, based on the rearranged codebook, we further introduce a Cluster-oriented Cross-entropy Loss, which guides the model to predict the correct cluster that contains the target token, thereby providing the model with a broader perspective and no longer confined to a single target token. Combined with the original token-oriented cross-entropy loss, our model can consider both the target cluster and the target token. Even if the model predicts the wrong token index, with our rearranged codebook and cluster-oriented loss, there is a high probability that the token is located in the target cluster, which ensures a high similarity between the output image and the target one, effectively improving the generation quality and robustness (Fig. 1). + +We develop our model based on LlamaGen [36] and compare with existing visual generation methods based on GAN, Diffusion, Autoregressive (AR), and MIM models. Extensive experiments demonstrate that, across different parameter scales of LlamaGen (ranging from 100M to 1.4B), our method consistently enhances the model's training efficiency and performance. For the same FID, our approach reduces LlamaGen's training time by half, and under the same number of training epochs, it effectively improves the model's generation quality. Our method can be applied to various LLM-based visual generation models with almost no additional training costs, and adheres to the scaling law, providing a promising direction for improving future LLM-based visual generation models. + +Our main contributions can be summarized as four-fold: + +- We propose IAR, an Improved Autoregressive Visual Generation method, which leverages the correlation of visual embeddings to enhance the training efficiency and performance of LLM-based visual generation models. +- We introduce a Codebook Rearrangement strategy that uses balanced K-means clustering algorithm to rearrange the image codebook, ensuring high similarity across image embeddings within each cluster. +- We propose a Cluster-oriented Cross-entropy Loss that relaxes the original token-oriented cross-entropy, ensuring that even if the model predicts the wrong token index, there is still a high probability that the token is in the correct cluster, thereby generating high-quality images. +- Extensive experiments demonstrate the effectiveness of our method. Additionally, our approach can be applied to various LLM-based visual generation models, adheres to the scaling law, and provides a robust improvement direction for future LLM-based visual generation models. + +![](images/69a30eba775aa2b4054b56551f5a35f65d9755ab4995b64007f00dd00b804a6a.jpg) +(a) + +![](images/ea6d9b8101bc7b99ea2051ff15ce9bb9a8a95ce250bbb68bf7e658002a7e4f63.jpg) +(b) +Figure 2. (a) The MSE and LPIPS between the source image and the decoded image with different code distances. (b) The visualization of the images decoded from different code distances. When the code distance is within a certain range (e.g., smaller than 12), the decoded image looks nearly identical to the source image. We further make use of this property to improve the LLM-based visual generation model. + +# 2. Related Works + +# 2.1. Large language models + +Large language models (LLMs) [11, 32, 39] have achieved significant success in natural language processing (NLP). The encoder-only structure [9, 11, 28], exemplified by BERT [11], randomly masks portions of the text and predicts the masked tokens based on the unmasked ones. With the introduction of the GPT series [1, 5, 31, 32], the decoder-only architecture gained popularity, demonstrating remarkable capabilities in language processing through next-token prediction. Additionally, open-source models such as LLaMA [29] and PaLM [8] have further advanced the NLP field. The recent success of LLMs in natural language processing has also inspired researchers to apply these models to visual generation, revealing significant potential in multi-modal generation tasks. + +# 2.2. Visual Generation Models + +Continuous-valued Visual Generation. Generative Adversarial Networks (GANs) [13, 15, 25, 41, 42], consisting of a generator and discriminator, employ adversarial training to learn the target distribution. Diffusion Models [12, 14, 19-21, 30, 33] iteratively add and remove noise to approximate the real data distribution, using a Markov process for the forward path and learning the reverse denoising process. These models have achieved significant success in image generation and have been extended to video generation [3, 16, 22, 40], enhancing video quality. However, both GANs and Diffusion Models, which model continuous distributions, face challenges when integrating with discretely modeled LLMs. Thus, exploring discrete modeling methods for image generation is necessary. + +Discrete-valued Visual Generation. Discrete-space visual generation models are mainly divided into autoregressive [13, 36, 38] and masked image modeling (MIM) [2, 6, + +7] models, both requiring image quantization before generation. Autoregressive models follow the next-token generation approach of GPT [1, 32], predicting subsequent tokens one by one based on previous image tokens. LLaM-AGEN [36], a representative work, uses LLAMA [39] to model the image generation process. VAR [38] further enhances generation capability by proposing a next-scale image generation process. MIM methods [2, 6, 36, 38], inspired by BERT [11], randomly mask parts of the image tokens and predict the masked tokens using the unmasked ones. These methods directly draw inspiration from LLMs and rarely explore the fundamental differences between images and language, which may lead to suboptimal utilization of vision generation capabilities within the LLM framework. In contrast, our model makes full use of the image property, which effectively improves the training efficiency and model performance in LLM-based image generation. + +# 3. Method + +Observing that an inherent difference between LLM-based image generation and text generation lies in that text generation only requires a predicted text index that can be directly mapped into the corresponding word, while image generation essentially requires the image embedding that corresponds to the index, and then decodes it into an image, we further explore the characteristics of the image embedding space. We conduct a detailed analysis on the similarity of the image embeddings in the codebook learned by image tokenizers (Sec. 3), and find that similar embeddings convey similar information in the image space: When the embeddings of certain image patches are replaced by similar image embeddings, the decoded images are nearly identical. Therefore, we leverage this property to enhance the LLM-based image generation model. + +We first propose a Codebook Rearrangement strategy + +![](images/def87f239eb34b064c68d5fd6ac613fe712380f11cbaef2ca6c9264cb41e78d4.jpg) +Stage 1: Codebook Rearrangement +Stage 2: Cluster-level constraint +Figure 3. Model framework: 1) Codebook Rearrangement: we first use a balanced K-means clustering method to rearrange the codebook, which divides the codebook into $n$ clusters, with the image codes in each cluster sharing a high similarity. 2) Cluster-oriented Constraint: During the training process, we first quantize the image patches using the rearranged codebook. For the output probability distribution $\hat{Y}$ , we further compute the cluster-level distribution $\hat{Y}_C$ by applying LogSumExp operation for the probabilities in each cluster $\hat{Y}_{jm} \sim \hat{Y}_{(j+1)m-1}$ . Then we compute the cluster-oriented cross-entropy loss $\mathcal{L}_{CCE}$ apart from the token-oriented cross-entropy loss $\mathcal{L}_{TCE}$ , which ensures a high probability of the predicted token located in the correct cluster, thereby enhancing generation quality. + +that uses a balanced K-means clustering method to rearrange the embeddings in the codebook into clusters of equal sizes (Fig. 3 (left), Sec. 3.2), where the image embeddings in each cluster share a high similarity. Once the model predicts the correct cluster index (even if the token index is wrong), the corresponding embedding is located in the correct cluster and is thus similar to the target one, then the output images will be quite similar to the target image. To enable the model with the ability to predict the correct cluster index, we further propose a Cluster-oriented Cross-entropy Loss (Fig. 3 (right), Sec. 3.3). Our rearranged codebook and cluster-oriented cross-entropy loss ensure that, even if the model predicts incorrect tokens, there is a high probability of correctly predicting the cluster indices, which ensures high-quality generation results, largely improving the training efficiency and robustness of the LLM-based image generation model. + +# 3.1. Analysis on Image Embedding Similarity + +We first analyze the similarity of image embeddings in the codebook learned by image tokenizers, and find that similar embeddings (those that are close in distance) represent similar image information. I.e., when the embeddings of each image patch are replaced by other similar embeddings, the content of the decoded image remains almost unchanged. To verify this image embedding similarity property, we conduct detailed experiments on the codebook of + +# VQGAN [13]. + +Specifically, an image $x$ is quantized into discrete tokens by: 1) using an image encoder $E$ (e.g., from VQGAN [13]) to extract features, obtaining $h \times w$ image features of dimension $C$ , denoted as $\hat{z} = E(x) \in \mathcal{R}^{h \times w \times C}$ ; 2) quantizing each feature $\hat{z}^{(i,j)}$ into the code index $q^{(i,j)}$ of the nearest code in the codebook $\mathcal{Z} \in \mathcal{R}^{N \times C}$ : + +$$ +q ^ {(i, j)} = \arg \min _ {z _ {k} \in \mathcal {Z}} \| \hat {z} ^ {(i, j)} - z _ {k} \| \in [ 0, N), \tag {1} +$$ + +where $z_{k}$ is the $k$ -th embedding in codebook $\mathcal{Z}$ , and $\| \cdot \|$ measures the Euclidean distance. And the quantized feature $z_{q} \in \mathcal{R}^{h \times w \times C}$ is obtained by looking up each $q^{(i,j)}$ in $\mathcal{Z}$ . + +To measure the distance between two embeddings $z_{i}, z_{j}$ in the codebook, we define the "code distance" $D(z_{i}, z_{j})$ : + +$$ +D \left(z _ {i}, z _ {j}\right) := d, \quad z _ {j} = \left[ \operatorname {s o r t} \left(\left\{\| z _ {k} - z _ {i} \| \right\}\right) \right] _ {d}, \tag {2} +$$ + +calculated by measuring the Euclidean distance of each $z_{k}$ to $z_{i}$ , sorting the distances, and $z_{j}$ is the $D(z_{i},z_{j})$ -th closest embedding to $z_{i}$ among all embeddings in the codebook. A lower "code distance" indicates the two embeddings are closer in the embedding space. + +Given an input image $x$ , the quantized embedding $z_{q}$ and a generated image $\hat{x}^{\prime} = G(z_{q}^{\prime})$ decoded from $z_{q}^{\prime}$ , we evaluate the similarity of the two images at different code distances $D(z_{q},z_{q}^{\prime})$ . In Fig. 2 (a), we measure the image similarity by MSE distance and LPIPS [43], where lower + +Algorithm 1 Balanced k-means Clustering +Input: Codebook $\mathcal{Z}$ , number of clusters $n$ +Output: The rearranged codebook $\hat{\mathcal{Z}}$ with local similarity +1: Initialize: +2: Compute the cluster size $m\gets |\mathcal{Z}| / n$ +3:Randomly select $n$ codes as initial centroids $\{c_j\}_{j = 1}^n$ +4:Initialize iteration counter iter $\leftarrow 0$ +5:repeat +6:Compute the minimum distance $d_{i}$ from $z_{i}$ to $\{c_j\}_{j = 1}^n$ +7:Sort $\{z_i\}$ in ascending order of $d_{i}$ to form $\mathcal{Z}'$ +8:for each data point $z_{i}\in \mathcal{Z}'$ do +9:Compute the distance from $z_{i}$ to each centroid $c_{j}$ +10:Assign $z_{i}$ to the nearest cluster whose size does not exceed $m$ +11:end for +12:for each cluster centroid $c_{j}$ do +13:Update $c_{j}$ to the average of all embeddings assigned to cluster $j$ +14:end for +15:Increment iteration counter iter $\leftarrow iter + 1$ +16:until convergence or reaching max_iters +17:foreach cluster centroid $c_{j}$ do +18:Assign the embeddings in cluster $j$ to $\hat{\mathcal{Z}}_{[jm,(j + 1)m)}$ +19:end for +20:return The rearranged codebook $\hat{\mathcal{Z}}$ + +scores indicate higher image similarity. Results show: 1) as the code distance increases, the image distance also increases; 2) but when the code distance is within a certain range (e.g., smaller than 12), the image distance (MSE 0.104 and LPIPS 66.44) does not increase too much compared to the reconstructed image $G(z_{q})$ (whose code distance is 0 and has a 0.076 MSE and 55.8 LPIPS). We further visualize the images decoded from different code distances in Fig. 2(b), which shows the decoded images with lower code distances (e.g., smaller than 12) look nearly identical to the source image and have a good visual quality. This indicates that even if the predicted token index is not the accurate target index, as long as the code distance between the corresponding embedding is within a certain range, the decoded image is similar to the desired one and shows a good visual quality. Therefore, we make use of this property to improve the generation quality and stability of the LLM-based visual generation model. + +# 3.2. Codebook Rearrangement + +Rearrangement Target. In a codebook learned from the VQGAN [13] training process, the embeddings are often randomly distributed, with adjacent embeddings having no particular correlation, i.e., Euclidean distances between adjacent embeddings can be either close or far. Therefore, we first propose a Codebook Rearrangement strategy to reorder the codebook, so that the similarity between adjacent embeddings is maximized, ensuring that the surrounding em + +beddings of any given embedding exhibit high similarity. + +Specifically, denote the codebook as $\mathcal{Z} = \{z_i\}_{i=1}^N$ which contains $N$ quantized embeddings $z_i$ . To maximize the similarity between the adjacent embeddings, we aim to find a surjective mapping $M(\cdot)$ that satisfies: + +$$ +M = \arg \min _ {M} \sum_ {i = 1} ^ {N - 1} \| z _ {M (i)}, z _ {M (i + 1)} \|. \tag {3} +$$ + +I.e., after reordering each embedding $z_{i}$ to index $M(i)$ , the sum of distances between adjacent embeddings is minimized. And $\hat{\mathcal{Z}} = M(\mathcal{Z})$ is the rearranged codebook. + +However, this optimization can be reduced to the Hamiltonian path problem (see #Suppl), and solving such a problem is NP-hard. Therefore, we try to relax this problem into a solvable form. + +Constraint Relaxation. We relax this optimization problem into a much easier one, where we only need to ensure the embeddings in a range are similar. Therefore, we split the codebook into $n$ clusters, with each cluster containing $m = \frac{N}{n}$ embeddings that are similar to each other. Then we reorder the codebook so that the embeddings in the same cluster have close indices, with the indices of cluster $j$ are within the range $[jm, (j + 1)m)$ . + +Balanced K-means Clustering. To cluster the embeddings in the codebook, we design a balanced K-means clustering method to uniformly divide the codebook into $n$ clusters, with the embeddings in each cluster close to each other. + +As shown in Alg. 1, we first randomly select $n$ codes as the initial centroids $\{c_j\}_{j=1}^n$ . Then, we iteratively update each cluster. In each iteration, 1) we first compute the minimum distance $d_i$ from each embedding $z_i$ to the $n$ centroids $\{c_j\}_{j=1}^n$ . 2) We sort $\{z_i\}$ in ascending order of $d_i$ to form $\mathcal{Z}'$ (i.e., embeddings closer to the nearest centroid has smaller index). This sorting and reordering step ensures that the embeddings closer to the nearest centroid are allocated first, and the farther ones are allocated later, which is necessary in cluster size balanced clustering. 3) Then, for each $z_i$ in the ordered codebook $\mathcal{Z}'$ , we compute its distance to each centroid $c_j$ . 4) We assign $z_i$ to the nearest cluster whose size does not exceed $m$ (where $m = \frac{N}{n}$ ). 5) Finally, we update each cluster centroid $c_j$ to the average of the embeddings assigned to that cluster, and repeat the above process until it converges or reaches the pre-defined maximum iteration. After obtaining the final cluster centroids $c_j$ and $n$ clusters, where each cluster contains $m$ embeddings, we rearrange the codebook $\hat{\mathcal{Z}}$ by assigning the embeddings in cluster $j$ to indices $[jm, (j+1)m]$ . + +# 3.3. Cluster-oriented Visual Generation + +Analysis on Token-oriented Cross-entropy Loss. In LLM-based visual generation models [36, 38], cross-entropy loss is the most commonly used loss in the training + +process. Denote the ground truth one-hot vector as $Y$ ( $Y$ is a $N$ -dimensional vector, with $Y_{i=y} = 1$ and $Y_{i\neq y} = 0$ , where $y$ is the class label of the token), and the predicted probability distribution as $\hat{Y} \in \mathcal{R}^N(\sum \hat{Y}_i = 1)$ . The token-oriented cross-entropy loss $\mathcal{L}_{TCE}$ is formulated as: + +$$ +\mathcal {L} _ {T C E} = - \sum_ {i = 1} ^ {n} Y _ {i} \log \hat {Y} _ {i}. \tag {4} +$$ + +However, we observe that if a token index is wrongly predicted as a similar embedding's index, this token-oriented cross-entropy loss will penalize the case, but the decoded image is actually not much different from the target image. + +To tolerate this case, we leverage the rearranged codebook, where the embeddings in the same cluster are close to each other. In Sec. 3, it has been demonstrated that even if some image token indices are predicted incorrectly, as long as the embeddings of the predicted tokens are not far from the embeddings of the target tokens, the semantic integrity and quality of the generated image will not be significantly affected. Therefore, leveraging the reordered codebook from Sec. 3.2, a natural idea is to guide the model to first predict the cluster and then locate the specific token within that cluster. Since the number of clusters $(n)$ is much smaller than the size of the codebook $(N)$ , predicting the cluster of a token is a much easier task than directly predicting a specific token. Once the cluster is accurately predicted, even if the specific token prediction is incorrect, the quality of the generated image can still be maintained. + +Cluster-oriented Cross-entropy Loss. To guide the model to predict the correct cluster of the next token, we propose a cluster-wise cross-entropy loss $\mathcal{L}_C$ . Specifically, we first derive the ground truth cluster label $y_{c}$ from the class label $y^{1}$ by $y_{c} = \left\lfloor \frac{y}{m}\right\rfloor$ , where $m$ is the number of tokens in each cluster. Then, to form a new cluster-level probability distribution $\hat{Y}_{C}\in \mathcal{R}^{n}(\sum \hat{Y}_{C,i} = 1)$ from the token-level probability distribution $\hat{Y}$ , we employ the following operation to calculate the probability of each cluster, where within a cluster, the samples with larger probability contribute more to the cluster probability: + +$$ +\hat {Y} _ {C, j} = \sum_ {i = j m} ^ {(j + 1) m - 1} e x p (\hat {Y} _ {i}) / \sum_ {i = 1} ^ {N} e x p (\hat {Y} _ {i}), \quad j = 1, \dots , n \tag {5} +$$ + +where $\hat{Y}_{C,j}$ represents the probability that the generated token belongs to cluster $j$ . + +With the ground-truth cluster label $y_{c}$ and the predicted cluster-level probability distribution $\hat{Y}_{C,i}$ , our cluster-oriented cross-entropy loss $\mathcal{L}_{CCE}$ is formulated as: + +$$ +\mathcal {L} _ {C C E} = - \sum_ {j = 1} ^ {n} Y _ {C, j} \log \hat {Y} _ {C, j}, \tag {6} +$$ + +
TypeModel#Para.FID↓IS↑Precision↑Recall↑
GANBigGAN [4]112M6.95224.50.890.38
GigaGAN [23]569M3.45225.50.840.61
StyleGan-XL [35]166M2.30265.10.780.53
DiffusionADM [12]554M10.94101.00.690.63
CDM [20]-4.88158.7--
LDM-4 [33]400M3.60247.7--
DiT-XL/2 [30]675M2.27278.20.830.57
Mask.MaskGIT [6]227M6.18182.10.800.51
MaskGIT-re [6]227M4.02355.6--
ARVQGAN [13]227M18.6580.40.780.26
VQGAN [13]1.4B15.7874.3--
VQGAN-re [13]1.4B5.20280.3--
ViT-VQGAN [42]1.7B4.17175.1--
ViT-VQGAN-re [42]1.7B3.04227.4--
RQTran. [27]3.8B7.55134.0--
RQTran.-re [27]3.8B3.80323.7--
ARLlamaGen-B [36]111M5.46193.60.830.45
LlamaGen-L [36]343M3.29227.80.820.53
LlamaGen-XL [36]775M2.63244.10.810.58
LlamaGen-XXL [36]1.4B2.34253.90.800.59
OursIAR-B111M5.14202.00.850.45
IAR-L343M3.18234.80.820.53
IAR-XL775M2.52248.10.820.58
IAR-XXL (CFG=1.7)1.4B2.19278.90.810.58
IAR-XXL (CFG=2.5)1.4B3.74362.00.860.54
+ +Table 1. Comparison between different types of image generation model on class-conditional ImageNet $256 \times 256$ benchmark with FID, IS, precision, and recall. + +where $Y_{C}$ is the one-hot vector spanned by the label $y_{c}$ . With our cluster-oriented loss, even if the model predicts the wrong token index, there is a high probability that the token is located in the target cluster, which ensures a high similarity between the output image and the target one, effectively improving the generation quality and robustness. + +Final Loss Function. The final loss function $\mathcal{L}$ combines both the token-oriented cross-entropy loss $\mathcal{L}_{TCE}$ and the cluster-oriented cross-entropy loss $\mathcal{L}_{CCE}$ : + +$$ +\mathcal {L} = \mathcal {L} _ {T C E} + \lambda \mathcal {L} _ {C C E}. \tag {7} +$$ + +# 4. Experiments + +# 4.1. Experimental settings + +In this paper, we choose LLamaGen [36] as the main baseline and apply our method to it. We follow the official training details and keep the hyperparameters the same as the official ones. We conduct experiments on ImageNet [10]. To evaluate the model performance, we generate 50K images with randomly selected labels. Then, we compute the FID [17], IS [34], precision, and recall [26] for the generated data (see #Suppl for details). + +# 4.2. Comparison Experiments + +Comparison with baselines. We compare our model with the GAN-based methods [4, 23, 35], diffusion-based methods [12, 20, 30, 33], mask-prediction-based methods [6] and autoregressive-based methods [13, 27, 36, 42] on ImageNet [10]. The results are shown in Tab. 2. Our method + +
TokensModel#Para.50 epoch300 epoch
FID↓IS↑Precision↑Recal↑FID↓IS↑Precision↑Recal↑
16 × 16LlamaGen-B111M7.22178.30.860.385.46193.60.840.46
LlamaGen-L343M4.20200.00.820.513.80248.30.830.52
LlamaGen-XL775M3.39227.10.810.54----
LlamaGen-XXL1.4B3.09253.60.830.53----
IAR-B111M6.90179.20.860.405.14202.00.850.45
IAR-L343M4.10207.10.820.513.40271.30.840.51
IAR-XL775M3.36228.90.820.54----
IAR-XXL1.4B3.01257.40.830.53----
24 × 24LlamaGen-B111M8.31154.70.840.386.09182.50.840.42
LlamaGen-L343M4.61191.40.820.503.29227.80.820.53
LlamaGen-XL775M3.24245.70.830.532.63244.10.810.58
LlamaGen-XXL1.4B2.89236.20.800.562.34253.90.810.60
IAR-B111M7.80153.30.840.395.77192.50.850.42
IAR-L343M4.35197.20.810.513.18234.80.820.53
IAR-XL775M3.15228.80.810.542.52248.10.820.58
IAR-XXL1.4B2.87249.90.820.562.19265.60.810.58
+ +Table 2. Comparison with LlamaGen across different image tokens and model sizes. Following LlamaGen, we only train XL and XXL version on $16 \times 16$ tokens for 50 epochs. + +achieves the state-of-the-art FID (2.19) and IS (362.0). And for each model size, our IAR consistently outperforms LlamaGen, demonstrating the effectiveness of our cluster-oriented autoregressive visual generation strategy. + +More comparisons with LlamaGen. We further compare our model with the baseline LlamaGen across different model sizes and image tokens to validate the effectiveness of our model. Specifically, we choose four models with model sizes ranging from 111M to 1.4B parameters. And we also conduct experiments on $16 \times 16$ image tokens (corresponds to $256 \times 256$ image) and $24 \times 24$ image tokens (corresponds to $384 \times 384$ image). We choose the model trained with 50 epochs and 300 epochs for evaluation. The comparison results are shown in Tab. 2. It can be seen that across different model sizes, image tokens, and training epochs, our model can always improve the baseline with a better FID and IS, validating the effectiveness and robustness of our model in LLM-based visual generation. + +Effects of classifier-free guidance (CFG). IAR can employ classifier-free guidance [18] to generate images with higher quality. We show the performance (FID & IS) of our model (111M and 343M parameters) on different CFGs ranging from 1.0 to 2.75 in Fig. 4 (a). It can be seen that compared to the model without CFG (i.e., CFG=1.0), the model with CFG can largely improve the FID, which means the model can generate images with much higher quality. Specifically, as the CFG increases, the FID first decreases and then increases, indicating that a too-large CFG may influence the generation quality, which is consistent with LlamaGen and previous diffusion-based methods [12, 18]. + +Effects of model size. We conduct experiments on different model sizes, where the parameter number ranges from 111M to 1.4B. It can be seen in Tab. 2 and Fig. 4 (b) that as the parameter number grows, the model shows + +a clear improvement in the performance, which aligns with the scaling law [24], indicating that as the number of parameter increases, our IVR can be more effective. + +Training efficiency. We compare the training speed between IAR-B and LlamaGen-B (for different model sizes, please refer to #Suppl). We evaluate the model performance on different epochs (from 50 to 300). The FID and IS results are shown in Fig. 4 (c), which shows that our 175-epoch IAR reaches almost the same FID score as the 300-epoch LlamaGen, indicating that our model can accelerate the training speed for almost $42\%$ . And comparing the models trained on 300 epochs, our model effectively improves the generation quality, validating that our model can improve the training efficiency of the LLM-based visual generation model. + +# 4.3. Ablation Study + +To evaluate the effectiveness of our model, we conduct several ablation studies to validate each module separately. We conduct the ablation studies based on IAR-B (111M parameters), and train the model for 100 epochs on ImageNet. + +Ablation on codebook rearrangement and $\mathcal{L}_{CCE}$ . We first evaluate the effectiveness of the codebook rearrangement strategy and the cluster-oriented cross-entropy loss $\mathcal{L}_{CCE}$ . We conduct experiments on three ablated models: the model without both codebook rearrangement and $\mathcal{L}_{CCE}$ (original LlamaGen), the model without codebook rearrangement, and the model without $\mathcal{L}_{CCE}$ . In contrast, our model with both codebook rearrangement and $\mathcal{L}_{CCE}$ has the best performance in fitting the ImageNet dataset. + +Ablation on cluster size. A key hyperparameter in our model is the size of each cluster. A too-small cluster size may make it too hard for the model to predict the correct cluster index, which leads the model to perform similarly to the original LlamaGen (Note that the model with cluster + +![](images/5dc12d09f16e44eadcd270fb5733c0adb8ec2aba4fd5709faeaa5fad300336ee.jpg) +Figure 4. (a) Model performance (IAR-B and IAR-L) across different CFGs; (b) Model performance on different parameter numbers (111M to 3B) compared to LlamaGen; and (c) Model performance on different epochs compared to LlamaGen. + +![](images/2f41c403ce53d4e378d37709f3d649f468afde5a5535bd46c545bd0a28bbc0aa.jpg) + +![](images/e3f341d1dcf73cb1a99833e368508285d9dd6bf269aa4024d217b1b754c8b34d.jpg) + +
Codebook Rearrangement\( \mathcal{L}_{CCE} \)FID↓IS↑Precision↑Recall↑
7.14166.380.840.40
✓7.15163.090.840.41
✓\( 6.96^2 \)168.700.840.40
✓✓6.77171.730.840.42
+ +Table 3. Ablation studies on codebook rearrangement strategy and the cluster-oriented cross-entropy loss $\mathcal{L}_{\mathrm{CCE}}$ + +
Cluster SizeInner MSEFID↓IS↑Precision↑Recall↑
107.14166.380.840.40
80.0187.02167.090.850.40
160.0226.99166.390.840.42
320.0286.83170.100.840.40
640.0346.92171.650.850.39
1280.0416.77171.730.840.42
2560.0506.81170.580.850.41
5120.0596.94163.540.830.42
+ +size 1 is the same as LlamaGen). In contrast, a too-big cluster size may cause the image embeddings in each cluster to be too different, such that even if the model predicts the correct cluster index, the gap to the target image embedding is still too big. Therefore, it is important to choose an appropriate cluster size to balance the inner similarity in each cluster and the ease of predicting the correct cluster index. We conduct experiments on different cluster sizes ranging from 1 to 512, where the codebook size is 16384. We compute the inner MSE of cluster, i.e., the average distance from each code in a cluster to the centroid, and evaluate the model performance by FID, IS, Precision, and Recall in Tab. 4. It can be seen that as the cluster size increases, the inner MSE distance becomes higher, indicating that the image codes in each cluster become less similar. Moreover, the model reaches the best performance when the cluster size is 128, which happens to be the square root of the codebook size. + +Ablation on weight of $\mathcal{L}_{CCE}$ . Since we introduce an additional cluster-oriented cross-entropy loss $\mathcal{L}_{CCE}$ during the training process, we conduct experiments to find the influence of its corresponding weight $\lambda$ . We choose $\lambda$ rang- + +Table 4. Ablation studies on different cluster sizes from 1 to 512, where cluster size 1 corresponds to the official LlamaGen. + +
λFID↓IS↑Precision↑Recall↑
07.14166.380.840.40
0.17.01167.160.840.40
0.257.05160.730.840.42
0.56.78171.150.840.41
0.756.80170.060.840.41
16.77171.730.840.42
1.56.87170.600.840.41
+ +Table 5. Ablation studies on the weight $\lambda$ of the cluster-oriented cross-attention loss ${\mathcal{L}}_{\mathrm{{CCE}}}$ . + +ing from 0.1 to 1.5 and evaluate the model performance in Tab. 5. It can be seen that a too-small weight will cause the model to perform similarly to the original LlamGen, while a too-big weight will make the model concern too much about the cluster index and overlook the ground-truth token index, which also makes the model perform worse. In summary, we find that when the weight $\lambda = 1$ , the model can balance the two goals and achieves the best performance. + +# 5. Conclusion + +In this paper, we analyzed the difference between the natural language and images in LLM-based visual generation. We find that similar image embeddings in the codebook can produce similar images. Based on this observation, we propose IAR, an Improved Autoregressive Visual Generation Method that effectively enhances the training efficiency and quality of the LLM-based visual generation model. We first conduct Codebook Rearrangement with a balanced K-means clustering algorithm to reorder the codebook into clusters with equal sizes, where the image embeddings within each cluster share similarities. Then, we introduce a Cluster-oriented Cross-entropy Loss to enable the model to learn the target cluster, which can guarantee a good generation quality even if the model predicts a wrong image token. We validate the effectiveness of our IAR on LlamaGen, and find that IAR can largely improve the training efficiency and generation quality across different parameter scales. Moreover, our IAR can be applied to various LLM-based visual generation models, pointing out a new direction to improve the LLM-based visual generation field. + +# Acknowledgments + +This work was supported by Shanghai Sailing Program (22YF1420300), National Natural Science Foundation of China (No. 62302297, 72192821, 62272447, 62472282, 62472285), Young Elite Scientists Sponsorship Program by CAST (2022QNRC001), the Fundamental Research Funds for the Central Universities (YG2023QNB17, YG2024QNA44), Beijing Natural Science Foundation (L222117). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[2] Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Qing-Guo Chen, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing masked generative transformers for efficient high-resolution text-to-image synthesis. arXiv preprint arXiv:2410.08261, 2024. 1, 3 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 3 +[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 6 +[5] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3 +[6] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 1, 3, 6 +[7] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, 2023. 3 +[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. JMLR, 2023. 3 +[9] Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. ELECTRA: Pre-training text encoders as discriminators rather than generators. In ICLR, 2020. 3 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6 +[11] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3 + +[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3, 6, 7 +[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2, 3, 4, 5, 6 +[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 3 +[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commun Acm, 2020. 1, 3 +[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In ICLR, 2024. 1, 3 +[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6 +[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 7 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3 +[20] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 6 +[21] Teng Hu, Jiangning Zhang, Liang Liu, Ran Yi, Siqi Kou, Haokun Zhu, Xu Chen, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Phasic content fusing diffusion model with directional distribution consistency for few-shot model adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2406-2415, 2023. 3 +[22] Teng Hu, Jiangning Zhang, Ran Yi, Yating Wang, Hongrui Huang, Jieyu Weng, Yabiao Wang, and Lizhuang Ma. Motionmaster: Training-free camera motion transfer for video generation. arXiv preprint arXiv:2404.15789, 2024. 3 +[23] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, 2023. 6 +[24] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 7 +[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 3 +[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In NeurIPS, 2019. 6 +[27] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 6 + +[28] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3 +[29] Shervin Minaee, Tomas Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Amatriain, and Jianfeng Gao. Large language models: A survey. arXiv preprint arXiv:2402.06196, 2024. 3 +[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 3, 6 +[31] Alec Radford. Improving language understanding by generative pre-training. 2018. 3 +[32] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 1, 2, 3 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3, 6 +[34] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In NeurIPS, 2016. 6 +[35] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganxl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH, 2022. 6 +[36] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 3, 5, 6 +[37] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1 +[38] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. In NeurIPS, 2024. 1, 3, 5 +[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[40] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2025.3 +[41] Ran Yi, Teng Hu, Mengfei Xia, Yizhe Tang, and Yong-Jin Liu. Feditnet++: Few-shot editing of latent semantics in gan spaces with correlated attribute disentanglement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 +[42] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. In ICLR, 2021. 3, 6 +[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4 +[44] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe + +Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1 \ No newline at end of file diff --git a/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/images.zip b/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d5d493a2800f717c3cf2b48b84d5b157ca71a22c --- /dev/null +++ b/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54f62187b20aeb3435da30e4116e9ea48bb3f3323db9aec9c083399548c66dcc +size 590135 diff --git a/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/layout.json b/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a59881bfe00429add814d174062a24d3b0115dbb --- /dev/null +++ b/2025/Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction/layout.json @@ -0,0 +1,9206 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 103, + 459, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 103, + 459, + 138 + ], + "spans": [ + { + "bbox": [ + 151, + 103, + 459, + 138 + ], + "type": "text", + "content": "Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "spans": [ + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": "Teng Hu" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": ", Jiangning Zhang" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": ", Ran Yi" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": ", Jieyu Weng" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": ", Yabiao Wang" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{3,2}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": ", Xianfang Zeng" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": ", Zhucun Xue" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": ", Lizhuang Ma" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": "Shanghai Jiao Tong University, " + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": "Youtu Lab, Tencent, " + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 116, + 160, + 493, + 232 + ], + "type": "text", + "content": "Zhejiang University {hu-teng, ranyi, w.jerry, lzma} @sjtu.edu.cn, {186368, yabiaowang, zzlongjuanfeng, 12432038} @zju.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 286, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 286, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 296, + 670 + ], + "type": "text", + "content": "Employing LLMs for visual generation has recently become a research focus. However, the existing methods primarily transfer the LLM architecture to visual generation but rarely investigate the fundamental differences between language and vision. This oversight may lead to suboptimal utilization of visual generation capabilities within the LLM framework. In this paper, we explore the characteristics of visual embedding space under the LLM framework and discover that the correlation between visual embeddings can help achieve more stable and robust generation results. We present IAR, an Improved AutoRegressive Visual Generation Method that enhances the training efficiency and generation quality of LLM-based visual generation models. Firstly, we propose a Codebook Rearrange strategy that uses balanced " + }, + { + "bbox": [ + 54, + 286, + 296, + 670 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 286, + 296, + 670 + ], + "type": "text", + "content": "-means clustering algorithm to rearrange the visual codebook into clusters, ensuring high similarity among visual features within each cluster. Leveraging the rearranged codebook, we propose a Cluster-oriented Cross-entropy Loss that guides the model to correctly predict the cluster where the target token is located. This approach ensures that even if the model predicts the wrong token index, there is a high probability the predicted token is located in the correct cluster, which significantly enhances the generation quality and robustness. Extensive experiments demonstrate that our IAR consistently enhances the model training efficiency and performance from 100M to 1.4B, reducing the training time by half while achieving the same FID. Additionally, IAR can be applied to various LLM-based visual generation models and adheres to the scaling law, providing a promising direction for future research in LLM-based visual generation. The code is available at: https://github.com/sjtuplayer/IAR." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 258, + 555, + 406 + ], + "blocks": [ + { + "bbox": [ + 316, + 258, + 555, + 406 + ], + "lines": [ + { + "bbox": [ + 316, + 258, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 555, + 406 + ], + "type": "image", + "image_path": "05c7aee6e1d6d40473115d53426ce46150c7c488c6ca9b40865216aff4112f71.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 416, + 555, + 462 + ], + "lines": [ + { + "bbox": [ + 313, + 416, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 555, + 462 + ], + "type": "text", + "content": "Figure 1. When an autoregressive model predicts a wrong token, the previous methods [36, 38] may predict an irrelevant token that causes artifacts. Our IAR alleviates this issue by ensuring a high probability of the predicted token located in the correct cluster." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 474, + 394, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 474, + 394, + 486 + ], + "spans": [ + { + "bbox": [ + 314, + 474, + 394, + 486 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 495, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 495, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 312, + 495, + 555, + 651 + ], + "type": "text", + "content": "With the development of generative models [15, 19, 33], a large number of outstanding visual generation methods have emerged, achieving considerable success in image and video generation [3, 16, 33]. Recently, researchers have begun exploring the integration of images and text to achieve a unified multi-modal model for image and text generation and understanding [37, 44]. Consequently, aligning image generation models with large language models (LLMs) [32] has become a key research focus. Recent studies have made progress by quantizing images and employing autoregressive or mask-prediction methods to predict discrete image tokens [2, 6, 36, 38], laying the research foundation for a unified world model in the future." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Unlike diffusion [19] or GAN [15] models that model image distribution in continuous space, autoregressive or masked image modeling (MIM) methods [2, 6, 36, 38] first convert images into discrete-valued tokens by image tokenizers, and then predict image tokens by autoregressive or" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9351" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "MIM approach. Specifically, autoregressive methods use the \"next-token prediction\" paradigm of GPT [32], while MIM methods adopt a similar training strategy to the mask prediction in BERT [11]. These methods draw inspiration from natural language models and transfer these techniques to image generation models, but rarely explore the fundamental differences between images and natural language." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 159, + 298, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 298, + 433 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 298, + 433 + ], + "type": "text", + "content": "Due to the inherently discrete nature of words (each word comes from a finite vocabulary), different words can be directly mapped to different indices through a lookup table, thereby generating text by predicting the corresponding index of the target text. The standard language modeling objective " + }, + { + "bbox": [ + 56, + 159, + 298, + 433 + ], + "type": "inline_equation", + "content": "p(x_{t}|x_{1},\\ldots ,x_{t - 1})" + }, + { + "bbox": [ + 56, + 159, + 298, + 433 + ], + "type": "text", + "content": " (each " + }, + { + "bbox": [ + 56, + 159, + 298, + 433 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 56, + 159, + 298, + 433 + ], + "type": "text", + "content": " is an index) then can be interpreted as a classification task and accurately modeled by LLMs. However, images are continuous-valued, and if the same objective is applied to image generation with " + }, + { + "bbox": [ + 56, + 159, + 298, + 433 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 56, + 159, + 298, + 433 + ], + "type": "text", + "content": " as a real number, it is difficult to accurately model this probability density. Therefore, image tokenizers such as VQGAN [13] are used to convert continuous images into discrete-valued tokens, and then the model predicts the image token index. Afterwards, the corresponding image embedding is retrieved from the codebook and decoded into an image. A fundamental difference between image and text is that, in text generation, the predicted text index can be directly mapped to the corresponding word through a lookup table; whereas LLM-based image generation essentially requires the image embedding that corresponds to the index, and then decodes it into an image. I.e., text only requires the index, while image requires the embedding corresponding to the index." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "type": "text", + "content": "Different from the token index which is independent, i.e., the nearby indices have no correlation in word semantics, the image embeddings are located in a continuous feature space, where similar embeddings may correspond to similar image contents. This paper first investigates the impact of image embedding correlations on the generated results, and finds that similar embeddings convey similar information in the image space: When the embeddings of certain image patches are replaced by similar image embeddings, the decoded images are nearly identical. This insight inspires us to leverage the similarity of image embeddings to improve the existing LLM-based image generation techniques." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "To capitalize on the correlation among image embeddings, we propose IAR, an Improved AutoRegressive Visual Generation Method, which can enhance the training efficiency and generation quality of the LLM-based visual generation model. Firstly, we propose a Codebook Rearrangement strategy that uses balanced K-means clustering algorithm to rearrange the embeddings in the codebook into clusters of equal sizes, where the embeddings in each cluster share high similarities and can be decoded into images with similar contents. Secondly, we observe that if a token index is wrongly predicted as a similar embedding's" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 264 + ], + "type": "text", + "content": "index, the original token-oriented cross-entropy loss used in LLM will penalize this case, but the decoded image is actually not much different from the target image. To tolerate this case, based on the rearranged codebook, we further introduce a Cluster-oriented Cross-entropy Loss, which guides the model to predict the correct cluster that contains the target token, thereby providing the model with a broader perspective and no longer confined to a single target token. Combined with the original token-oriented cross-entropy loss, our model can consider both the target cluster and the target token. Even if the model predicts the wrong token index, with our rearranged codebook and cluster-oriented loss, there is a high probability that the token is located in the target cluster, which ensures a high similarity between the output image and the target one, effectively improving the generation quality and robustness (Fig. 1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 281, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 556, + 449 + ], + "type": "text", + "content": "We develop our model based on LlamaGen [36] and compare with existing visual generation methods based on GAN, Diffusion, Autoregressive (AR), and MIM models. Extensive experiments demonstrate that, across different parameter scales of LlamaGen (ranging from 100M to 1.4B), our method consistently enhances the model's training efficiency and performance. For the same FID, our approach reduces LlamaGen's training time by half, and under the same number of training epochs, it effectively improves the model's generation quality. Our method can be applied to various LLM-based visual generation models with almost no additional training costs, and adheres to the scaling law, providing a promising direction for improving future LLM-based visual generation models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 467, + 553, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 467, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 326, + 467, + 553, + 478 + ], + "type": "text", + "content": "Our main contributions can be summarized as four-fold:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 496, + 553, + 710 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 496, + 553, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 496, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 314, + 496, + 553, + 542 + ], + "type": "text", + "content": "- We propose IAR, an Improved Autoregressive Visual Generation method, which leverages the correlation of visual embeddings to enhance the training efficiency and performance of LLM-based visual generation models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 544, + 553, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 544, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 314, + 544, + 553, + 590 + ], + "type": "text", + "content": "- We introduce a Codebook Rearrangement strategy that uses balanced K-means clustering algorithm to rearrange the image codebook, ensuring high similarity across image embeddings within each cluster." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 592, + 553, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 592, + 553, + 650 + ], + "spans": [ + { + "bbox": [ + 314, + 592, + 553, + 650 + ], + "type": "text", + "content": "- We propose a Cluster-oriented Cross-entropy Loss that relaxes the original token-oriented cross-entropy, ensuring that even if the model predicts the wrong token index, there is still a high probability that the token is in the correct cluster, thereby generating high-quality images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 651, + 553, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 651, + 553, + 710 + ], + "spans": [ + { + "bbox": [ + 314, + 651, + 553, + 710 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the effectiveness of our method. Additionally, our approach can be applied to various LLM-based visual generation models, adheres to the scaling law, and provides a robust improvement direction for future LLM-based visual generation models." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9352" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 258, + 217 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 258, + 217 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 258, + 217 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 258, + 217 + ], + "type": "image", + "image_path": "69a30eba775aa2b4054b56551f5a35f65d9755ab4995b64007f00dd00b804a6a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 218, + 167, + 228 + ], + "lines": [ + { + "bbox": [ + 157, + 218, + 167, + 228 + ], + "spans": [ + { + "bbox": [ + 157, + 218, + 167, + 228 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 259, + 74, + 553, + 211 + ], + "blocks": [ + { + "bbox": [ + 259, + 74, + 553, + 211 + ], + "lines": [ + { + "bbox": [ + 259, + 74, + 553, + 211 + ], + "spans": [ + { + "bbox": [ + 259, + 74, + 553, + 211 + ], + "type": "image", + "image_path": "ea6d9b8101bc7b99ea2051ff15ce9bb9a8a95ce250bbb68bf7e658002a7e4f63.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 218, + 428, + 228 + ], + "lines": [ + { + "bbox": [ + 416, + 218, + 428, + 228 + ], + "spans": [ + { + "bbox": [ + 416, + 218, + 428, + 228 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 236, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 54, + 236, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 555, + 270 + ], + "type": "text", + "content": "Figure 2. (a) The MSE and LPIPS between the source image and the decoded image with different code distances. (b) The visualization of the images decoded from different code distances. When the code distance is within a certain range (e.g., smaller than 12), the decoded image looks nearly identical to the source image. We further make use of this property to improve the LLM-based visual generation model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 279, + 147, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 147, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 147, + 291 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 298, + 187, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 187, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 187, + 312 + ], + "type": "text", + "content": "2.1. Large language models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 316, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 296, + 483 + ], + "type": "text", + "content": "Large language models (LLMs) [11, 32, 39] have achieved significant success in natural language processing (NLP). The encoder-only structure [9, 11, 28], exemplified by BERT [11], randomly masks portions of the text and predicts the masked tokens based on the unmasked ones. With the introduction of the GPT series [1, 5, 31, 32], the decoder-only architecture gained popularity, demonstrating remarkable capabilities in language processing through next-token prediction. Additionally, open-source models such as LLaMA [29] and PaLM [8] have further advanced the NLP field. The recent success of LLMs in natural language processing has also inspired researchers to apply these models to visual generation, revealing significant potential in multi-modal generation tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 492, + 201, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 201, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 201, + 504 + ], + "type": "text", + "content": "2.2. Visual Generation Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 677 + ], + "type": "text", + "content": "Continuous-valued Visual Generation. Generative Adversarial Networks (GANs) [13, 15, 25, 41, 42], consisting of a generator and discriminator, employ adversarial training to learn the target distribution. Diffusion Models [12, 14, 19-21, 30, 33] iteratively add and remove noise to approximate the real data distribution, using a Markov process for the forward path and learning the reverse denoising process. These models have achieved significant success in image generation and have been extended to video generation [3, 16, 22, 40], enhancing video quality. However, both GANs and Diffusion Models, which model continuous distributions, face challenges when integrating with discretely modeled LLMs. Thus, exploring discrete modeling methods for image generation is necessary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Discrete-valued Visual Generation. Discrete-space visual generation models are mainly divided into autoregressive [13, 36, 38] and masked image modeling (MIM) [2, 6," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 280, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 483 + ], + "type": "text", + "content": "7] models, both requiring image quantization before generation. Autoregressive models follow the next-token generation approach of GPT [1, 32], predicting subsequent tokens one by one based on previous image tokens. LLaM-AGEN [36], a representative work, uses LLAMA [39] to model the image generation process. VAR [38] further enhances generation capability by proposing a next-scale image generation process. MIM methods [2, 6, 36, 38], inspired by BERT [11], randomly mask parts of the image tokens and predict the masked tokens using the unmasked ones. These methods directly draw inspiration from LLMs and rarely explore the fundamental differences between images and language, which may lead to suboptimal utilization of vision generation capabilities within the LLM framework. In contrast, our model makes full use of the image property, which effectively improves the training efficiency and model performance in LLM-based image generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 498, + 370, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 498, + 370, + 510 + ], + "spans": [ + { + "bbox": [ + 314, + 498, + 370, + 510 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 555, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 699 + ], + "type": "text", + "content": "Observing that an inherent difference between LLM-based image generation and text generation lies in that text generation only requires a predicted text index that can be directly mapped into the corresponding word, while image generation essentially requires the image embedding that corresponds to the index, and then decodes it into an image, we further explore the characteristics of the image embedding space. We conduct a detailed analysis on the similarity of the image embeddings in the codebook learned by image tokenizers (Sec. 3), and find that similar embeddings convey similar information in the image space: When the embeddings of certain image patches are replaced by similar image embeddings, the decoded images are nearly identical. Therefore, we leverage this property to enhance the LLM-based image generation model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "We first propose a Codebook Rearrangement strategy" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9353" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 81, + 555, + 291 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 237, + 80 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 237, + 80 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 237, + 80 + ], + "type": "text", + "content": "Stage 1: Codebook Rearrangement" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 81, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 58, + 81, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 58, + 81, + 555, + 291 + ], + "type": "image", + "image_path": "def87f239eb34b064c68d5fd6ac613fe712380f11cbaef2ca6c9264cb41e78d4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 70, + 485, + 80 + ], + "lines": [ + { + "bbox": [ + 362, + 70, + 485, + 80 + ], + "spans": [ + { + "bbox": [ + 362, + 70, + 485, + 80 + ], + "type": "text", + "content": "Stage 2: Cluster-level constraint" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "type": "text", + "content": "Figure 3. Model framework: 1) Codebook Rearrangement: we first use a balanced K-means clustering method to rearrange the codebook, which divides the codebook into " + }, + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "type": "text", + "content": " clusters, with the image codes in each cluster sharing a high similarity. 2) Cluster-oriented Constraint: During the training process, we first quantize the image patches using the rearranged codebook. For the output probability distribution " + }, + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "type": "text", + "content": ", we further compute the cluster-level distribution " + }, + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\hat{Y}_C" + }, + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "type": "text", + "content": " by applying LogSumExp operation for the probabilities in each cluster " + }, + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\hat{Y}_{jm} \\sim \\hat{Y}_{(j+1)m-1}" + }, + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "type": "text", + "content": ". Then we compute the cluster-oriented cross-entropy loss " + }, + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CCE}" + }, + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "type": "text", + "content": " apart from the token-oriented cross-entropy loss " + }, + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TCE}" + }, + { + "bbox": [ + 55, + 300, + 555, + 366 + ], + "type": "text", + "content": ", which ensures a high probability of the predicted token located in the correct cluster, thereby enhancing generation quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 380, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 380, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 54, + 380, + 296, + 583 + ], + "type": "text", + "content": "that uses a balanced K-means clustering method to rearrange the embeddings in the codebook into clusters of equal sizes (Fig. 3 (left), Sec. 3.2), where the image embeddings in each cluster share a high similarity. Once the model predicts the correct cluster index (even if the token index is wrong), the corresponding embedding is located in the correct cluster and is thus similar to the target one, then the output images will be quite similar to the target image. To enable the model with the ability to predict the correct cluster index, we further propose a Cluster-oriented Cross-entropy Loss (Fig. 3 (right), Sec. 3.3). Our rearranged codebook and cluster-oriented cross-entropy loss ensure that, even if the model predicts incorrect tokens, there is a high probability of correctly predicting the cluster indices, which ensures high-quality generation results, largely improving the training efficiency and robustness of the LLM-based image generation model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 597, + 270, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 270, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 270, + 611 + ], + "type": "text", + "content": "3.1. Analysis on Image Embedding Similarity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "We first analyze the similarity of image embeddings in the codebook learned by image tokenizers, and find that similar embeddings (those that are close in distance) represent similar image information. I.e., when the embeddings of each image patch are replaced by other similar embeddings, the content of the decoded image remains almost unchanged. To verify this image embedding similarity property, we conduct detailed experiments on the codebook of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 380, + 374, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 380, + 374, + 392 + ], + "spans": [ + { + "bbox": [ + 314, + 380, + 374, + 392 + ], + "type": "text", + "content": "VQGAN [13]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "text", + "content": "Specifically, an image " + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "text", + "content": " is quantized into discrete tokens by: 1) using an image encoder " + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "text", + "content": " (e.g., from VQGAN [13]) to extract features, obtaining " + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "text", + "content": " image features of dimension " + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "inline_equation", + "content": "\\hat{z} = E(x) \\in \\mathcal{R}^{h \\times w \\times C}" + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "text", + "content": "; 2) quantizing each feature " + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "inline_equation", + "content": "\\hat{z}^{(i,j)}" + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "text", + "content": " into the code index " + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "inline_equation", + "content": "q^{(i,j)}" + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "text", + "content": " of the nearest code in the codebook " + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} \\in \\mathcal{R}^{N \\times C}" + }, + { + "bbox": [ + 313, + 392, + 555, + 463 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 351, + 472, + 555, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 472, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 351, + 472, + 555, + 491 + ], + "type": "interline_equation", + "content": "q ^ {(i, j)} = \\arg \\min _ {z _ {k} \\in \\mathcal {Z}} \\| \\hat {z} ^ {(i, j)} - z _ {k} \\| \\in [ 0, N), \\tag {1}", + "image_path": "cafd628b7121f66d67b433972fb6ded4d21e0c6bf711da96617b27d2bac53540.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "inline_equation", + "content": "z_{k}" + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": "-th embedding in codebook " + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": " measures the Euclidean distance. And the quantized feature " + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "inline_equation", + "content": "z_{q} \\in \\mathcal{R}^{h \\times w \\times C}" + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": " is obtained by looking up each " + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "inline_equation", + "content": "q^{(i,j)}" + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 537, + 553, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 553, + 563 + ], + "type": "text", + "content": "To measure the distance between two embeddings " + }, + { + "bbox": [ + 313, + 537, + 553, + 563 + ], + "type": "inline_equation", + "content": "z_{i}, z_{j}" + }, + { + "bbox": [ + 313, + 537, + 553, + 563 + ], + "type": "text", + "content": " in the codebook, we define the \"code distance\" " + }, + { + "bbox": [ + 313, + 537, + 553, + 563 + ], + "type": "inline_equation", + "content": "D(z_{i}, z_{j})" + }, + { + "bbox": [ + 313, + 537, + 553, + 563 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 342, + 571, + 553, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 571, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 342, + 571, + 553, + 585 + ], + "type": "interline_equation", + "content": "D \\left(z _ {i}, z _ {j}\\right) := d, \\quad z _ {j} = \\left[ \\operatorname {s o r t} \\left(\\left\\{\\| z _ {k} - z _ {i} \\| \\right\\}\\right) \\right] _ {d}, \\tag {2}", + "image_path": "eee74fecaae19f46e1daebfc900a34b41b6420a2f03ee436b8e99a92e876c004.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "text", + "content": "calculated by measuring the Euclidean distance of each " + }, + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "inline_equation", + "content": "z_{k}" + }, + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "text", + "content": ", sorting the distances, and " + }, + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "inline_equation", + "content": "z_{j}" + }, + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "inline_equation", + "content": "D(z_{i},z_{j})" + }, + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "text", + "content": "-th closest embedding to " + }, + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 313, + 594, + 555, + 653 + ], + "type": "text", + "content": " among all embeddings in the codebook. A lower \"code distance\" indicates the two embeddings are closer in the embedding space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ", the quantized embedding " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "z_{q}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " and a generated image " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\hat{x}^{\\prime} = G(z_{q}^{\\prime})" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " decoded from " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "z_{q}^{\\prime}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ", we evaluate the similarity of the two images at different code distances " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "D(z_{q},z_{q}^{\\prime})" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ". In Fig. 2 (a), we measure the image similarity by MSE distance and LPIPS [43], where lower" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9354" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 65, + 95, + 289, + 355 + ], + "blocks": [ + { + "bbox": [ + 66, + 82, + 221, + 94 + ], + "lines": [ + { + "bbox": [ + 66, + 82, + 221, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 82, + 221, + 94 + ], + "type": "text", + "content": "Algorithm 1 Balanced k-means Clustering" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "lines": [ + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "spans": [ + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": "Input: Codebook " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": ", number of clusters " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " \nOutput: The rearranged codebook " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Z}}" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " with local similarity \n1: Initialize: \n2: Compute the cluster size " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "m\\gets |\\mathcal{Z}| / n" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " \n3:Randomly select " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " codes as initial centroids " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\{c_j\\}_{j = 1}^n" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " \n4:Initialize iteration counter iter " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\leftarrow 0" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " \n5:repeat \n6:Compute the minimum distance " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\{c_j\\}_{j = 1}^n" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " \n7:Sort " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\{z_i\\}" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " in ascending order of " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " to form " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}'" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " \n8:for each data point " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "z_{i}\\in \\mathcal{Z}'" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " do \n9:Compute the distance from " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " to each centroid " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " \n10:Assign " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " to the nearest cluster whose size does not exceed " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " \n11:end for \n12:for each cluster centroid " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " do \n13:Update " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " to the average of all embeddings assigned to cluster " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " \n14:end for \n15:Increment iteration counter iter " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\leftarrow iter + 1" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " \n16:until convergence or reaching max_iters \n17:foreach cluster centroid " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " do \n18:Assign the embeddings in cluster " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Z}}_{[jm,(j + 1)m)}" + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "text", + "content": " \n19:end for \n20:return The rearranged codebook " + }, + { + "bbox": [ + 65, + 95, + 289, + 355 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Z}}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 377, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 296, + 594 + ], + "type": "text", + "content": "scores indicate higher image similarity. Results show: 1) as the code distance increases, the image distance also increases; 2) but when the code distance is within a certain range (e.g., smaller than 12), the image distance (MSE 0.104 and LPIPS 66.44) does not increase too much compared to the reconstructed image " + }, + { + "bbox": [ + 55, + 377, + 296, + 594 + ], + "type": "inline_equation", + "content": "G(z_{q})" + }, + { + "bbox": [ + 55, + 377, + 296, + 594 + ], + "type": "text", + "content": " (whose code distance is 0 and has a 0.076 MSE and 55.8 LPIPS). We further visualize the images decoded from different code distances in Fig. 2(b), which shows the decoded images with lower code distances (e.g., smaller than 12) look nearly identical to the source image and have a good visual quality. This indicates that even if the predicted token index is not the accurate target index, as long as the code distance between the corresponding embedding is within a certain range, the decoded image is similar to the desired one and shows a good visual quality. Therefore, we make use of this property to improve the generation quality and stability of the LLM-based visual generation model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 600, + 203, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 203, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 203, + 613 + ], + "type": "text", + "content": "3.2. Codebook Rearrangement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "Rearrangement Target. In a codebook learned from the VQGAN [13] training process, the embeddings are often randomly distributed, with adjacent embeddings having no particular correlation, i.e., Euclidean distances between adjacent embeddings can be either close or far. Therefore, we first propose a Codebook Rearrangement strategy to reorder the codebook, so that the similarity between adjacent embeddings is maximized, ensuring that the surrounding em" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 544, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 544, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 544, + 84 + ], + "type": "text", + "content": "beddings of any given embedding exhibit high similarity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 84, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 133 + ], + "type": "text", + "content": "Specifically, denote the codebook as " + }, + { + "bbox": [ + 313, + 84, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\{z_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 84, + 553, + 133 + ], + "type": "text", + "content": " which contains " + }, + { + "bbox": [ + 313, + 84, + 553, + 133 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 84, + 553, + 133 + ], + "type": "text", + "content": " quantized embeddings " + }, + { + "bbox": [ + 313, + 84, + 553, + 133 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 313, + 84, + 553, + 133 + ], + "type": "text", + "content": ". To maximize the similarity between the adjacent embeddings, we aim to find a surjective mapping " + }, + { + "bbox": [ + 313, + 84, + 553, + 133 + ], + "type": "inline_equation", + "content": "M(\\cdot)" + }, + { + "bbox": [ + 313, + 84, + 553, + 133 + ], + "type": "text", + "content": " that satisfies:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 358, + 142, + 553, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 142, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 358, + 142, + 553, + 175 + ], + "type": "interline_equation", + "content": "M = \\arg \\min _ {M} \\sum_ {i = 1} ^ {N - 1} \\| z _ {M (i)}, z _ {M (i + 1)} \\|. \\tag {3}", + "image_path": "0ae007048a3a60036bcfb7ad8bf1b951e2185595a15306abd930c3bf66ef3e3b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 184, + 553, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 553, + 220 + ], + "type": "text", + "content": "I.e., after reordering each embedding " + }, + { + "bbox": [ + 313, + 184, + 553, + 220 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 313, + 184, + 553, + 220 + ], + "type": "text", + "content": " to index " + }, + { + "bbox": [ + 313, + 184, + 553, + 220 + ], + "type": "inline_equation", + "content": "M(i)" + }, + { + "bbox": [ + 313, + 184, + 553, + 220 + ], + "type": "text", + "content": ", the sum of distances between adjacent embeddings is minimized. And " + }, + { + "bbox": [ + 313, + 184, + 553, + 220 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Z}} = M(\\mathcal{Z})" + }, + { + "bbox": [ + 313, + 184, + 553, + 220 + ], + "type": "text", + "content": " is the rearranged codebook." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 220, + 553, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 267 + ], + "type": "text", + "content": "However, this optimization can be reduced to the Hamiltonian path problem (see #Suppl), and solving such a problem is NP-hard. Therefore, we try to relax this problem into a solvable form." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 269, + 554, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 554, + 364 + ], + "type": "text", + "content": "Constraint Relaxation. We relax this optimization problem into a much easier one, where we only need to ensure the embeddings in a range are similar. Therefore, we split the codebook into " + }, + { + "bbox": [ + 313, + 269, + 554, + 364 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 269, + 554, + 364 + ], + "type": "text", + "content": " clusters, with each cluster containing " + }, + { + "bbox": [ + 313, + 269, + 554, + 364 + ], + "type": "inline_equation", + "content": "m = \\frac{N}{n}" + }, + { + "bbox": [ + 313, + 269, + 554, + 364 + ], + "type": "text", + "content": " embeddings that are similar to each other. Then we reorder the codebook so that the embeddings in the same cluster have close indices, with the indices of cluster " + }, + { + "bbox": [ + 313, + 269, + 554, + 364 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 269, + 554, + 364 + ], + "type": "text", + "content": " are within the range " + }, + { + "bbox": [ + 313, + 269, + 554, + 364 + ], + "type": "inline_equation", + "content": "[jm, (j + 1)m)" + }, + { + "bbox": [ + 313, + 269, + 554, + 364 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 554, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 412 + ], + "type": "text", + "content": "Balanced K-means Clustering. To cluster the embeddings in the codebook, we design a balanced K-means clustering method to uniformly divide the codebook into " + }, + { + "bbox": [ + 313, + 365, + 554, + 412 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 365, + 554, + 412 + ], + "type": "text", + "content": " clusters, with the embeddings in each cluster close to each other." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": "As shown in Alg. 1, we first randomly select " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " codes as the initial centroids " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\{c_j\\}_{j=1}^n" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": ". Then, we iteratively update each cluster. In each iteration, 1) we first compute the minimum distance " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " from each embedding " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " centroids " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\{c_j\\}_{j=1}^n" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": ". 2) We sort " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\{z_i\\}" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " in ascending order of " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " to form " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}'" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " (i.e., embeddings closer to the nearest centroid has smaller index). This sorting and reordering step ensures that the embeddings closer to the nearest centroid are allocated first, and the farther ones are allocated later, which is necessary in cluster size balanced clustering. 3) Then, for each " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " in the ordered codebook " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}'" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": ", we compute its distance to each centroid " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": ". 4) We assign " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " to the nearest cluster whose size does not exceed " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "m = \\frac{N}{n}" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": "). 5) Finally, we update each cluster centroid " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " to the average of the embeddings assigned to that cluster, and repeat the above process until it converges or reaches the pre-defined maximum iteration. After obtaining the final cluster centroids " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " clusters, where each cluster contains " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " embeddings, we rearrange the codebook " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Z}}" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " by assigning the embeddings in cluster " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": " to indices " + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "inline_equation", + "content": "[jm, (j+1)m]" + }, + { + "bbox": [ + 313, + 413, + 555, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 659, + 502, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 502, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 502, + 672 + ], + "type": "text", + "content": "3.3. Cluster-oriented Visual Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Analysis on Token-oriented Cross-entropy Loss. In LLM-based visual generation models [36, 38], cross-entropy loss is the most commonly used loss in the training" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9355" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "process. Denote the ground truth one-hot vector as " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "-dimensional vector, with " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "Y_{i=y} = 1" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "Y_{i\\neq y} = 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " is the class label of the token), and the predicted probability distribution as " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\hat{Y} \\in \\mathcal{R}^N(\\sum \\hat{Y}_i = 1)" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ". The token-oriented cross-entropy loss " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TCE}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 123, + 140, + 295, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 140, + 295, + 170 + ], + "spans": [ + { + "bbox": [ + 123, + 140, + 295, + 170 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T C E} = - \\sum_ {i = 1} ^ {n} Y _ {i} \\log \\hat {Y} _ {i}. \\tag {4}", + "image_path": "a451294b8185cb1f7246de29be41088e47712fcab59e6311f4486ffc4167a104.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 176, + 296, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 176, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 296, + 224 + ], + "type": "text", + "content": "However, we observe that if a token index is wrongly predicted as a similar embedding's index, this token-oriented cross-entropy loss will penalize the case, but the decoded image is actually not much different from the target image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": "To tolerate this case, we leverage the rearranged codebook, where the embeddings in the same cluster are close to each other. In Sec. 3, it has been demonstrated that even if some image token indices are predicted incorrectly, as long as the embeddings of the predicted tokens are not far from the embeddings of the target tokens, the semantic integrity and quality of the generated image will not be significantly affected. Therefore, leveraging the reordered codebook from Sec. 3.2, a natural idea is to guide the model to first predict the cluster and then locate the specific token within that cluster. Since the number of clusters " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "(n)" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": " is much smaller than the size of the codebook " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "(N)" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": ", predicting the cluster of a token is a much easier task than directly predicting a specific token. Once the cluster is accurately predicted, even if the specific token prediction is incorrect, the quality of the generated image can still be maintained." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "text", + "content": "Cluster-oriented Cross-entropy Loss. To guide the model to predict the correct cluster of the next token, we propose a cluster-wise cross-entropy loss " + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "text", + "content": ". Specifically, we first derive the ground truth cluster label " + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "inline_equation", + "content": "y_{c}" + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "text", + "content": " from the class label " + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "inline_equation", + "content": "y^{1}" + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "inline_equation", + "content": "y_{c} = \\left\\lfloor \\frac{y}{m}\\right\\rfloor" + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "text", + "content": " is the number of tokens in each cluster. Then, to form a new cluster-level probability distribution " + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\hat{Y}_{C}\\in \\mathcal{R}^{n}(\\sum \\hat{Y}_{C,i} = 1)" + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "text", + "content": " from the token-level probability distribution " + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 55, + 415, + 296, + 548 + ], + "type": "text", + "content": ", we employ the following operation to calculate the probability of each cluster, where within a cluster, the samples with larger probability contribute more to the cluster probability:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 555, + 295, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 555, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 63, + 555, + 295, + 578 + ], + "type": "interline_equation", + "content": "\\hat {Y} _ {C, j} = \\sum_ {i = j m} ^ {(j + 1) m - 1} e x p (\\hat {Y} _ {i}) / \\sum_ {i = 1} ^ {N} e x p (\\hat {Y} _ {i}), \\quad j = 1, \\dots , n \\tag {5}", + "image_path": "efe89ff3bdb820c25bba15142bb91c9bf7674b3c2d110c202bf1891b3e1f45b9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 586, + 295, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 586, + 295, + 611 + ], + "type": "inline_equation", + "content": "\\hat{Y}_{C,j}" + }, + { + "bbox": [ + 55, + 586, + 295, + 611 + ], + "type": "text", + "content": " represents the probability that the generated token belongs to cluster " + }, + { + "bbox": [ + 55, + 586, + 295, + 611 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 586, + 295, + 611 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "text", + "content": "With the ground-truth cluster label " + }, + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "inline_equation", + "content": "y_{c}" + }, + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "text", + "content": " and the predicted cluster-level probability distribution " + }, + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\hat{Y}_{C,i}" + }, + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "text", + "content": ", our cluster-oriented cross-entropy loss " + }, + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CCE}" + }, + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 654, + 295, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 654, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 113, + 654, + 295, + 686 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C C E} = - \\sum_ {j = 1} ^ {n} Y _ {C, j} \\log \\hat {Y} _ {C, j}, \\tag {6}", + "image_path": "c7c5ae6a142ddc1ba1886860262635be841636ecd59a3ab25e2853e8c02e7282.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 556, + 305 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 556, + 305 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 556, + 305 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 556, + 305 + ], + "type": "table", + "html": "
TypeModel#Para.FID↓IS↑Precision↑Recall↑
GANBigGAN [4]112M6.95224.50.890.38
GigaGAN [23]569M3.45225.50.840.61
StyleGan-XL [35]166M2.30265.10.780.53
DiffusionADM [12]554M10.94101.00.690.63
CDM [20]-4.88158.7--
LDM-4 [33]400M3.60247.7--
DiT-XL/2 [30]675M2.27278.20.830.57
Mask.MaskGIT [6]227M6.18182.10.800.51
MaskGIT-re [6]227M4.02355.6--
ARVQGAN [13]227M18.6580.40.780.26
VQGAN [13]1.4B15.7874.3--
VQGAN-re [13]1.4B5.20280.3--
ViT-VQGAN [42]1.7B4.17175.1--
ViT-VQGAN-re [42]1.7B3.04227.4--
RQTran. [27]3.8B7.55134.0--
RQTran.-re [27]3.8B3.80323.7--
ARLlamaGen-B [36]111M5.46193.60.830.45
LlamaGen-L [36]343M3.29227.80.820.53
LlamaGen-XL [36]775M2.63244.10.810.58
LlamaGen-XXL [36]1.4B2.34253.90.800.59
OursIAR-B111M5.14202.00.850.45
IAR-L343M3.18234.80.820.53
IAR-XL775M2.52248.10.820.58
IAR-XXL (CFG=1.7)1.4B2.19278.90.810.58
IAR-XXL (CFG=2.5)1.4B3.74362.00.860.54
", + "image_path": "0cf30b642b092670b3ca43947b836cbf2f1f641441d424014321fce7605d8bda.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 307, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 340 + ], + "type": "text", + "content": "Table 1. Comparison between different types of image generation model on class-conditional ImageNet " + }, + { + "bbox": [ + 313, + 307, + 555, + 340 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 307, + 555, + 340 + ], + "type": "text", + "content": " benchmark with FID, IS, precision, and recall." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 351, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 554, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 351, + 554, + 422 + ], + "type": "inline_equation", + "content": "Y_{C}" + }, + { + "bbox": [ + 313, + 351, + 554, + 422 + ], + "type": "text", + "content": " is the one-hot vector spanned by the label " + }, + { + "bbox": [ + 313, + 351, + 554, + 422 + ], + "type": "inline_equation", + "content": "y_{c}" + }, + { + "bbox": [ + 313, + 351, + 554, + 422 + ], + "type": "text", + "content": ". With our cluster-oriented loss, even if the model predicts the wrong token index, there is a high probability that the token is located in the target cluster, which ensures a high similarity between the output image and the target one, effectively improving the generation quality and robustness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 423, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 460 + ], + "type": "text", + "content": "Final Loss Function. The final loss function " + }, + { + "bbox": [ + 313, + 423, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 423, + 554, + 460 + ], + "type": "text", + "content": " combines both the token-oriented cross-entropy loss " + }, + { + "bbox": [ + 313, + 423, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TCE}" + }, + { + "bbox": [ + 313, + 423, + 554, + 460 + ], + "type": "text", + "content": " and the cluster-oriented cross-entropy loss " + }, + { + "bbox": [ + 313, + 423, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CCE}" + }, + { + "bbox": [ + 313, + 423, + 554, + 460 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 386, + 470, + 553, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 470, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 386, + 470, + 553, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {T C E} + \\lambda \\mathcal {L} _ {C C E}. \\tag {7}", + "image_path": "d8db5e846d414de244d8e22ba4611a8297bfd7c779ecb59908c79277b671cf97.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 494, + 395, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 395, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 395, + 506 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 514, + 440, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 440, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 440, + 527 + ], + "type": "text", + "content": "4.1. Experimental settings" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 532, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 628 + ], + "type": "text", + "content": "In this paper, we choose LLamaGen [36] as the main baseline and apply our method to it. We follow the official training details and keep the hyperparameters the same as the official ones. We conduct experiments on ImageNet [10]. To evaluate the model performance, we generate 50K images with randomly selected labels. Then, we compute the FID [17], IS [34], precision, and recall [26] for the generated data (see #Suppl for details)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 635, + 457, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 457, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 457, + 649 + ], + "type": "text", + "content": "4.2. Comparison Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Comparison with baselines. We compare our model with the GAN-based methods [4, 23, 35], diffusion-based methods [12, 20, 30, 33], mask-prediction-based methods [6] and autoregressive-based methods [13, 27, 36, 42] on ImageNet [10]. The results are shown in Tab. 2. Our method" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 296, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 296, + 714 + ], + "type": "text", + "content": "1After codebook rearrangement, the class label has been updated accordingly." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9356" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 70, + 504, + 260 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 504, + 260 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 504, + 260 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 504, + 260 + ], + "type": "table", + "html": "
TokensModel#Para.50 epoch300 epoch
FID↓IS↑Precision↑Recal↑FID↓IS↑Precision↑Recal↑
16 × 16LlamaGen-B111M7.22178.30.860.385.46193.60.840.46
LlamaGen-L343M4.20200.00.820.513.80248.30.830.52
LlamaGen-XL775M3.39227.10.810.54----
LlamaGen-XXL1.4B3.09253.60.830.53----
IAR-B111M6.90179.20.860.405.14202.00.850.45
IAR-L343M4.10207.10.820.513.40271.30.840.51
IAR-XL775M3.36228.90.820.54----
IAR-XXL1.4B3.01257.40.830.53----
24 × 24LlamaGen-B111M8.31154.70.840.386.09182.50.840.42
LlamaGen-L343M4.61191.40.820.503.29227.80.820.53
LlamaGen-XL775M3.24245.70.830.532.63244.10.810.58
LlamaGen-XXL1.4B2.89236.20.800.562.34253.90.810.60
IAR-B111M7.80153.30.840.395.77192.50.850.42
IAR-L343M4.35197.20.810.513.18234.80.820.53
IAR-XL775M3.15228.80.810.542.52248.10.820.58
IAR-XXL1.4B2.87249.90.820.562.19265.60.810.58
", + "image_path": "fb15147df9387770131833564a2c4c6ccee8ee49e72f0845f3a3db07f158579c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "type": "text", + "content": "Table 2. Comparison with LlamaGen across different image tokens and model sizes. Following LlamaGen, we only train XL and XXL version on " + }, + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "type": "text", + "content": " tokens for 50 epochs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 295, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 344 + ], + "type": "text", + "content": "achieves the state-of-the-art FID (2.19) and IS (362.0). And for each model size, our IAR consistently outperforms LlamaGen, demonstrating the effectiveness of our cluster-oriented autoregressive visual generation strategy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 347, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 295, + 514 + ], + "type": "text", + "content": "More comparisons with LlamaGen. We further compare our model with the baseline LlamaGen across different model sizes and image tokens to validate the effectiveness of our model. Specifically, we choose four models with model sizes ranging from 111M to 1.4B parameters. And we also conduct experiments on " + }, + { + "bbox": [ + 54, + 347, + 295, + 514 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 54, + 347, + 295, + 514 + ], + "type": "text", + "content": " image tokens (corresponds to " + }, + { + "bbox": [ + 54, + 347, + 295, + 514 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 347, + 295, + 514 + ], + "type": "text", + "content": " image) and " + }, + { + "bbox": [ + 54, + 347, + 295, + 514 + ], + "type": "inline_equation", + "content": "24 \\times 24" + }, + { + "bbox": [ + 54, + 347, + 295, + 514 + ], + "type": "text", + "content": " image tokens (corresponds to " + }, + { + "bbox": [ + 54, + 347, + 295, + 514 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 54, + 347, + 295, + 514 + ], + "type": "text", + "content": " image). We choose the model trained with 50 epochs and 300 epochs for evaluation. The comparison results are shown in Tab. 2. It can be seen that across different model sizes, image tokens, and training epochs, our model can always improve the baseline with a better FID and IS, validating the effectiveness and robustness of our model in LLM-based visual generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 518, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 662 + ], + "type": "text", + "content": "Effects of classifier-free guidance (CFG). IAR can employ classifier-free guidance [18] to generate images with higher quality. We show the performance (FID & IS) of our model (111M and 343M parameters) on different CFGs ranging from 1.0 to 2.75 in Fig. 4 (a). It can be seen that compared to the model without CFG (i.e., CFG=1.0), the model with CFG can largely improve the FID, which means the model can generate images with much higher quality. Specifically, as the CFG increases, the FID first decreases and then increases, indicating that a too-large CFG may influence the generation quality, which is consistent with LlamaGen and previous diffusion-based methods [12, 18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Effects of model size. We conduct experiments on different model sizes, where the parameter number ranges from 111M to 1.4B. It can be seen in Tab. 2 and Fig. 4 (b) that as the parameter number grows, the model shows" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 295, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 555, + 331 + ], + "type": "text", + "content": "a clear improvement in the performance, which aligns with the scaling law [24], indicating that as the number of parameter increases, our IVR can be more effective." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 331, + 556, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 475 + ], + "type": "text", + "content": "Training efficiency. We compare the training speed between IAR-B and LlamaGen-B (for different model sizes, please refer to #Suppl). We evaluate the model performance on different epochs (from 50 to 300). The FID and IS results are shown in Fig. 4 (c), which shows that our 175-epoch IAR reaches almost the same FID score as the 300-epoch LlamaGen, indicating that our model can accelerate the training speed for almost " + }, + { + "bbox": [ + 313, + 331, + 556, + 475 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 313, + 331, + 556, + 475 + ], + "type": "text", + "content": ". And comparing the models trained on 300 epochs, our model effectively improves the generation quality, validating that our model can improve the training efficiency of the LLM-based visual generation model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 480, + 408, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 408, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 408, + 493 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 498, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 546 + ], + "type": "text", + "content": "To evaluate the effectiveness of our model, we conduct several ablation studies to validate each module separately. We conduct the ablation studies based on IAR-B (111M parameters), and train the model for 100 epochs on ImageNet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 546, + 556, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 654 + ], + "type": "text", + "content": "Ablation on codebook rearrangement and " + }, + { + "bbox": [ + 313, + 546, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CCE}" + }, + { + "bbox": [ + 313, + 546, + 556, + 654 + ], + "type": "text", + "content": ". We first evaluate the effectiveness of the codebook rearrangement strategy and the cluster-oriented cross-entropy loss " + }, + { + "bbox": [ + 313, + 546, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CCE}" + }, + { + "bbox": [ + 313, + 546, + 556, + 654 + ], + "type": "text", + "content": ". We conduct experiments on three ablated models: the model without both codebook rearrangement and " + }, + { + "bbox": [ + 313, + 546, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CCE}" + }, + { + "bbox": [ + 313, + 546, + 556, + 654 + ], + "type": "text", + "content": " (original LlamaGen), the model without codebook rearrangement, and the model without " + }, + { + "bbox": [ + 313, + 546, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CCE}" + }, + { + "bbox": [ + 313, + 546, + 556, + 654 + ], + "type": "text", + "content": ". In contrast, our model with both codebook rearrangement and " + }, + { + "bbox": [ + 313, + 546, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CCE}" + }, + { + "bbox": [ + 313, + 546, + 556, + 654 + ], + "type": "text", + "content": " has the best performance in fitting the ImageNet dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Ablation on cluster size. A key hyperparameter in our model is the size of each cluster. A too-small cluster size may make it too hard for the model to predict the correct cluster index, which leads the model to perform similarly to the original LlamaGen (Note that the model with cluster" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9357" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 218, + 176 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 218, + 176 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 218, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 218, + 176 + ], + "type": "image", + "image_path": "5dc12d09f16e44eadcd270fb5733c0adb8ec2aba4fd5709faeaa5fad300336ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 178, + 553, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 553, + 200 + ], + "type": "text", + "content": "Figure 4. (a) Model performance (IAR-B and IAR-L) across different CFGs; (b) Model performance on different parameter numbers (111M to 3B) compared to LlamaGen; and (c) Model performance on different epochs compared to LlamaGen." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 68, + 373, + 176 + ], + "blocks": [ + { + "bbox": [ + 220, + 68, + 373, + 176 + ], + "lines": [ + { + "bbox": [ + 220, + 68, + 373, + 176 + ], + "spans": [ + { + "bbox": [ + 220, + 68, + 373, + 176 + ], + "type": "image", + "image_path": "2f41c403ce53d4e378d37709f3d649f468afde5a5535bd46c545bd0a28bbc0aa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 374, + 68, + 553, + 176 + ], + "blocks": [ + { + "bbox": [ + 374, + 68, + 553, + 176 + ], + "lines": [ + { + "bbox": [ + 374, + 68, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 374, + 68, + 553, + 176 + ], + "type": "image", + "image_path": "e3f341d1dcf73cb1a99833e368508285d9dd6bf269aa4024d217b1b754c8b34d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 206, + 293, + 280 + ], + "blocks": [ + { + "bbox": [ + 58, + 206, + 293, + 280 + ], + "lines": [ + { + "bbox": [ + 58, + 206, + 293, + 280 + ], + "spans": [ + { + "bbox": [ + 58, + 206, + 293, + 280 + ], + "type": "table", + "html": "
Codebook Rearrangement\\( \\mathcal{L}_{CCE} \\)FID↓IS↑Precision↑Recall↑
7.14166.380.840.40
✓7.15163.090.840.41
✓\\( 6.96^2 \\)168.700.840.40
✓✓6.77171.730.840.42
", + "image_path": "ef0a91469919cde5f886005968b644b3042f7a292ca2c768008aeb1ca93af65e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 66, + 308, + 285, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 281, + 294, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 294, + 304 + ], + "type": "text", + "content": "Table 3. Ablation studies on codebook rearrangement strategy and the cluster-oriented cross-entropy loss " + }, + { + "bbox": [ + 55, + 281, + 294, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CCE}}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 308, + 285, + 407 + ], + "lines": [ + { + "bbox": [ + 66, + 308, + 285, + 407 + ], + "spans": [ + { + "bbox": [ + 66, + 308, + 285, + 407 + ], + "type": "table", + "html": "
Cluster SizeInner MSEFID↓IS↑Precision↑Recall↑
107.14166.380.840.40
80.0187.02167.090.850.40
160.0226.99166.390.840.42
320.0286.83170.100.840.40
640.0346.92171.650.850.39
1280.0416.77171.730.840.42
2560.0506.81170.580.850.41
5120.0596.94163.540.830.42
", + "image_path": "53d47fb83454a1eec8e1466e26389c815ae9a5b5dcd4641f6291df21a0f16a1e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 443, + 295, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 443, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 54, + 443, + 295, + 646 + ], + "type": "text", + "content": "size 1 is the same as LlamaGen). In contrast, a too-big cluster size may cause the image embeddings in each cluster to be too different, such that even if the model predicts the correct cluster index, the gap to the target image embedding is still too big. Therefore, it is important to choose an appropriate cluster size to balance the inner similarity in each cluster and the ease of predicting the correct cluster index. We conduct experiments on different cluster sizes ranging from 1 to 512, where the codebook size is 16384. We compute the inner MSE of cluster, i.e., the average distance from each code in a cluster to the centroid, and evaluate the model performance by FID, IS, Precision, and Recall in Tab. 4. It can be seen that as the cluster size increases, the inner MSE distance becomes higher, indicating that the image codes in each cluster become less similar. Moreover, the model reaches the best performance when the cluster size is 128, which happens to be the square root of the codebook size." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "text", + "content": "Ablation on weight of " + }, + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CCE}" + }, + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "text", + "content": ". Since we introduce an additional cluster-oriented cross-entropy loss " + }, + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CCE}" + }, + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "text", + "content": " during the training process, we conduct experiments to find the influence of its corresponding weight " + }, + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "text", + "content": ". We choose " + }, + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "text", + "content": " rang-" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 352, + 206, + 517, + 302 + ], + "blocks": [ + { + "bbox": [ + 55, + 411, + 295, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 433 + ], + "type": "text", + "content": "Table 4. Ablation studies on different cluster sizes from 1 to 512, where cluster size 1 corresponds to the official LlamaGen." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 352, + 206, + 517, + 302 + ], + "lines": [ + { + "bbox": [ + 352, + 206, + 517, + 302 + ], + "spans": [ + { + "bbox": [ + 352, + 206, + 517, + 302 + ], + "type": "table", + "html": "
λFID↓IS↑Precision↑Recall↑
07.14166.380.840.40
0.17.01167.160.840.40
0.257.05160.730.840.42
0.56.78171.150.840.41
0.756.80170.060.840.41
16.77171.730.840.42
1.56.87170.600.840.41
", + "image_path": "15a6362738ef18aaadda5e5639b00bd502af8ad8502b7912e9175a74c0c0bc62.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 308, + 555, + 330 + ], + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 330 + ], + "type": "text", + "content": "Table 5. Ablation studies on the weight " + }, + { + "bbox": [ + 313, + 308, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 308, + 555, + 330 + ], + "type": "text", + "content": " of the cluster-oriented cross-attention loss " + }, + { + "bbox": [ + 313, + 308, + 555, + 330 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{\\mathrm{{CCE}}}" + }, + { + "bbox": [ + 313, + 308, + 555, + 330 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 342, + 555, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 439 + ], + "type": "text", + "content": "ing from 0.1 to 1.5 and evaluate the model performance in Tab. 5. It can be seen that a too-small weight will cause the model to perform similarly to the original LlamGen, while a too-big weight will make the model concern too much about the cluster index and overlook the ground-truth token index, which also makes the model perform worse. In summary, we find that when the weight " + }, + { + "bbox": [ + 313, + 342, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 313, + 342, + 555, + 439 + ], + "type": "text", + "content": ", the model can balance the two goals and achieves the best performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 453, + 388, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 388, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 388, + 464 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 474, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we analyzed the difference between the natural language and images in LLM-based visual generation. We find that similar image embeddings in the codebook can produce similar images. Based on this observation, we propose IAR, an Improved Autoregressive Visual Generation Method that effectively enhances the training efficiency and quality of the LLM-based visual generation model. We first conduct Codebook Rearrangement with a balanced K-means clustering algorithm to reorder the codebook into clusters with equal sizes, where the image embeddings within each cluster share similarities. Then, we introduce a Cluster-oriented Cross-entropy Loss to enable the model to learn the target cluster, which can guarantee a good generation quality even if the model predicts a wrong image token. We validate the effectiveness of our IAR on LlamaGen, and find that IAR can largely improve the training efficiency and generation quality across different parameter scales. Moreover, our IAR can be applied to various LLM-based visual generation models, pointing out a new direction to improve the LLM-based visual generation field." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 214, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 214, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 214, + 713 + ], + "type": "text", + "content": "2For details, please refer to Sec. F of #Suppl." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9358" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 180 + ], + "type": "text", + "content": "This work was supported by Shanghai Sailing Program (22YF1420300), National Natural Science Foundation of China (No. 62302297, 72192821, 62272447, 62472282, 62472285), Young Elite Scientists Sponsorship Program by CAST (2022QNRC001), the Fundamental Research Funds for the Central Universities (YG2023QNB17, YG2024QNA44), Beijing Natural Science Foundation (L222117)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 201, + 115, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 201, + 115, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 201, + 115, + 213 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 220, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 220, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 220, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 296, + 275 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 277, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 277, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 61, + 277, + 296, + 331 + ], + "type": "text", + "content": "[2] Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Qing-Guo Chen, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing masked generative transformers for efficient high-resolution text-to-image synthesis. arXiv preprint arXiv:2410.08261, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 333, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 333, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 61, + 333, + 296, + 388 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 388, + 296, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 422 + ], + "type": "text", + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 422, + 294, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 294, + 444 + ], + "type": "text", + "content": "[5] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 445, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 477 + ], + "type": "text", + "content": "[6] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022. 1, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 479, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 296, + 555 + ], + "type": "text", + "content": "[7] Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Patrick Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 557, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 612 + ], + "type": "text", + "content": "[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. JMLR, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 613, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 645 + ], + "type": "text", + "content": "[9] Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. ELECTRA: Pre-training text encoders as discriminators rather than generators. In ICLR, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 678 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "type": "text", + "content": "[11] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "text", + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 97, + 554, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 554, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 554, + 129 + ], + "type": "text", + "content": "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 130, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 554, + 184 + ], + "type": "text", + "content": "[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 554, + 229 + ], + "type": "text", + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commun Acm, 2020. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 231, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 554, + 285 + ], + "type": "text", + "content": "[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In ICLR, 2024. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "type": "text", + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 354 + ], + "type": "text", + "content": "[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 554, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 376 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 377, + 554, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 554, + 410 + ], + "type": "text", + "content": "[20] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 411, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 554, + 477 + ], + "type": "text", + "content": "[21] Teng Hu, Jiangning Zhang, Liang Liu, Ran Yi, Siqi Kou, Haokun Zhu, Xu Chen, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Phasic content fusing diffusion model with directional distribution consistency for few-shot model adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2406-2415, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 478, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 521 + ], + "type": "text", + "content": "[22] Teng Hu, Jiangning Zhang, Ran Yi, Yating Wang, Hongrui Huang, Jieyu Weng, Yabiao Wang, and Lizhuang Ma. Motionmaster: Training-free camera motion transfer for video generation. arXiv preprint arXiv:2404.15789, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 523, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 554, + 555 + ], + "type": "text", + "content": "[23] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 557, + 554, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 554, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 554, + 610 + ], + "type": "text", + "content": "[24] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 613, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 644 + ], + "type": "text", + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 679 + ], + "type": "text", + "content": "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In NeurIPS, 2019. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[27] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9359" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "text", + "content": "[28] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 295, + 139 + ], + "type": "text", + "content": "[29] Shervin Minaee, Tomas Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Amatriain, and Jianfeng Gao. Large language models: A survey. arXiv preprint arXiv:2402.06196, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 141, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 141, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 57, + 141, + 294, + 162 + ], + "type": "text", + "content": "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 164, + 295, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 164, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 57, + 164, + 295, + 186 + ], + "type": "text", + "content": "[31] Alec Radford. Improving language understanding by generative pre-training. 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 187, + 295, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 187, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 57, + 187, + 295, + 219 + ], + "type": "text", + "content": "[32] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 221, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 221, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 57, + 221, + 294, + 253 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 255, + 294, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 255, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 57, + 255, + 294, + 287 + ], + "type": "text", + "content": "[34] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In NeurIPS, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 289, + 294, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 289, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 57, + 289, + 294, + 320 + ], + "type": "text", + "content": "[35] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganxl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 323, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 323, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 57, + 323, + 294, + 365 + ], + "type": "text", + "content": "[36] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 367, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 367, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 57, + 367, + 294, + 399 + ], + "type": "text", + "content": "[37] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 401, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 401, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 57, + 401, + 294, + 442 + ], + "type": "text", + "content": "[38] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. In NeurIPS, 2024. 1, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 446, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 446, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 57, + 446, + 294, + 499 + ], + "type": "text", + "content": "[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 502, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 294, + 555 + ], + "type": "text", + "content": "[40] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2025.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 557, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 294, + 610 + ], + "type": "text", + "content": "[41] Ran Yi, Teng Hu, Mengfei Xia, Yizhe Tang, and Yong-Jin Liu. Feditnet++: Few-shot editing of latent semantics in gan spaces with correlated attribute disentanglement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "type": "text", + "content": "[42] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. In ICLR, 2021. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "text", + "content": "[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[44] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 73, + 553, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 334, + 73, + 553, + 106 + ], + "type": "text", + "content": "Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9360" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/d107e339-bf08-4731-b903-e7f3360d585d_content_list.json b/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/d107e339-bf08-4731-b903-e7f3360d585d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0a442a5a62d62d8220f1f996455490d9390ddd25 --- /dev/null +++ b/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/d107e339-bf08-4731-b903-e7f3360d585d_content_list.json @@ -0,0 +1,1588 @@ +[ + { + "type": "text", + "text": "Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing", + "text_level": 1, + "bbox": [ + 102, + 130, + 893, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bingliang Zhang\\*,1", + "bbox": [ + 220, + 180, + 375, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenlin Meng3", + "bbox": [ + 228, + 199, + 352, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ California Institute of Technology", + "bbox": [ + 166, + 215, + 442, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenda Chu\\*,1", + "bbox": [ + 457, + 180, + 568, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anima Anandkumar", + "bbox": [ + 431, + 198, + 602, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2NVIDIA", + "bbox": [ + 465, + 215, + 545, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Julius Berner $^{2,\\dagger}$", + "bbox": [ + 648, + 180, + 774, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Song", + "bbox": [ + 684, + 198, + 777, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Stanford University 4OpenAI", + "bbox": [ + 568, + 215, + 831, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9abd721ce8744c5010bf45cc2c0bd44c288c8942c9b1da06449f29c171ef9ab3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 268, + 421, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c79e7bea559a0bccdf463a491c29f43836ecb0933685c487199626b1088828ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 270, + 901, + 551 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8501a45e9b3699b1069f727d48f5c7690ff3906ce2dd01d463a7e6ef5169440c.jpg", + "image_caption": [ + "Figure 1. Overview of Decoupled Annealing Posterior Sampling (DAPS). Our method provides a flexible and effective framework for solving inverse problems through a decoupled posterior sampling process. In (a)(b), we present DAPS visual results on FFHQ and ImageNet at a resolution of 256, and in (c), on natural images at a resolution of 768. In (d), we display DAPS results on compressed sensing multi-coil MRI (CS-MRI). DAPS effectively addresses nonlinear inverse problems as well as medical imaging MRI challenges. Additionally, DAPS can be enhanced using large-scale latent diffusion models (LDMs) [40], as shown in (c)." + ], + "image_footnote": [], + "bbox": [ + 94, + 415, + 419, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 643, + 326, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have recently achieved success in solving Bayesian inverse problems with learned data priors. Current methods build on top of the diffusion sampling process, where each denoising step makes small modifications to samples from the previous step. However, this process struggles to correct errors from earlier sampling steps, leading to worse performance in complicated nonlinear inverse problems, such as phase retrieval. To address this challenge, we propose a new method called Decoupled Annealing Posterior Sampling (DAPS) that relies on a novel noise annealing process. Specifically, we decouple consecutive steps in a diffusion sampling trajectory, allowing them to vary consid", + "bbox": [ + 88, + 676, + 485, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "erably from one another while ensuring their time-margins anneal to the true posterior as we reduce noise levels. This approach enables the exploration of a larger solution space, improving the success rate for accurate reconstructions. We demonstrate that DAPS significantly improves sample quality and stability across multiple image restoration tasks, particularly in complicated nonlinear inverse problems.", + "bbox": [ + 511, + 645, + 908, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 768, + 645, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inverse problems are ubiquitous in science and engineering, with applications ranging from image restoration [13, 30, 34, 44, 49, 59], medical imaging [10, 11, 17, 22, 23, 50] to astrophotography [1, 19, 51, 52]. Solving an inverse problem involves finding the underlying signal $\\mathbf{x}_0$ from its partial, noisy measurement $\\mathbf{y}$ . Since the measurement process is typically noisy and many-to-one, inverse problems do not have a", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Work partially done at the California Institute of Technology.", + "bbox": [ + 112, + 875, + 442, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work", + "bbox": [ + 112, + 887, + 367, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20895", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "unique solution; instead, multiple solutions may exist that are consistent with the observed measurement. In the Bayesian inverse problem framework, the solution space is characterized by the posterior distribution $p(\\mathbf{x}_0 \\mid \\mathbf{y}) \\propto p(\\mathbf{y} \\mid \\mathbf{x}_0) p(\\mathbf{x}_0)$ , where $p(\\mathbf{y} \\mid \\mathbf{x}_0)$ represents the noisy measurement process, and $p(\\mathbf{x}_0)$ is the prior distribution. In this work, we aim to solve Bayesian inverse problems where the measurement process $p(\\mathbf{y} \\mid \\mathbf{x}_0)$ is known, and the prior distribution $p(\\mathbf{x}_0)$ is captured by a deep generative model trained on a corresponding dataset.", + "bbox": [ + 89, + 90, + 485, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As score-based diffusion models [21, 26, 27, 47-49] have risen to dominance in modeling high-dimensional data distributions like images, audio, and video, they have become the leading method for estimating the prior distribution $p(\\mathbf{x}_0)$ in Bayesian inverse problems. A diffusion model generates a sample $\\mathbf{x}_0$ by smoothly removing noise from an unstructured initial noise sample $\\mathbf{x}_T$ through solving stochastic differential equations (SDEs). In particular, each step of the sampling process recursively converts a noisy sample $\\mathbf{x}_{t + \\Delta t}$ , where $\\Delta t > 0$ denotes the step size, to a slightly less noisy sample $\\mathbf{x}_t$ until $t = 0$ . This iterative structure in the diffusion sampling process can be leveraged to facilitate Bayesian inverse problem solving. In fact, prior research [5, 13, 46] has shown that given measurement $\\mathbf{y}$ and a diffusion model prior $p(\\mathbf{x}_0)$ , we can sample from the posterior distribution $p(\\mathbf{x}_0\\mid \\mathbf{y})$ by perturbing a reverse-time SDE using an approximate gradient of the measurement process, $\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{y}\\mid \\mathbf{x}_t)$ , at every step of the SDE solver.", + "bbox": [ + 91, + 244, + 486, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the remarkable success of these methods in solving many real-world inverse problems, like image colorization [30, 32, 49], image super-resolution [13, 30, 46, 49], computed tomography [45, 50], and magnetic resonance imaging [23, 50], they face significant challenges in more complex inverse problems with nonlinear measurement processes, such as phase retrieval and nonlinear motion deblurring. This is partially because, in prior methods, each denoising step approximately samples from the distribution $p(\\mathbf{x}_t \\mid \\mathbf{x}_{t + \\Delta t}, \\mathbf{y})$ . This causes $\\mathbf{x}_t$ and $\\mathbf{x}_{t + \\Delta t}$ to be close to each other because of using a small step size $\\Delta t$ in discretizing the reverse-time SDE. As a result, $\\mathbf{x}_t$ can at most correct local errors in $\\mathbf{x}_{t + \\Delta t}$ but struggles to correct global errors that require significant modifications to the prior sample. This issue is exacerbated when the methods are applied to complicated, nonlinear inverse problems, such as phase retrieval, where they often converge to undesired samples that are consistent with the measurement but reside in low-probability areas of the prior distribution.", + "bbox": [ + 89, + 518, + 485, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we propose a new framework for solving general inverse problems, termed Decoupled Annealing Posterior Sampling (DAPS). Our method employs a new noise annealing process inspired by the diffusion sampling process, where we decouple the consecutive samples $\\mathbf{x}_{t + \\Delta t}$ and $\\mathbf{x}_t$ in the sampling trajectory, allowing samplers", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to correct large, global errors made in earlier steps. Instead of repetitively sampling from $p(\\mathbf{x}_t \\mid \\mathbf{x}_{t + \\Delta t}, \\mathbf{y})$ as in previous methods, which restricts the distances between consecutive samples $\\mathbf{x}_{t + \\Delta t}$ and $\\mathbf{x}_t$ , DAPS recursively samples from the marginal distribution $p(\\mathbf{x}_t \\mid \\mathbf{y})$ . As illustrated in Fig. 2, we factorize the time marginal $p(\\mathbf{x}_t \\mid \\mathbf{y})$ into three conditional distributions and sample from them in turn by solving the reverse diffusion process, simulating Langevin dynamics, and adding noise according to the forward diffusion process. We show that this creates approximate samples from corresponding marginal distributions. As the noise gradually anneals to zero, the time marginal $p(\\mathbf{x}_t \\mid \\mathbf{y})$ smoothly converges to the posterior distribution $p(\\mathbf{x}_0 \\mid \\mathbf{y})$ , providing samples that approximately solve the Bayesian inverse problem.", + "bbox": [ + 511, + 90, + 908, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Empirically, our method demonstrates significantly improved performance on various inverse problems compared to existing approaches, ranging from image restoration to compressed sensing MRI. Our method can be combined with diffusion models in both raw pixel space and learned latent space, which we refer to as DAPS and LatentDAPS, respectively. As shown in Fig. 1, both methods provide superior reconstructions with improved visual perceptual quality across a wide range of nonlinear inverse problems. Our approach exhibits remarkable stability and sampling quality, particularly for challenging nonlinear inverse problems. On the FFHQ 256 dataset, we achieve PSNR values of $30.72\\mathrm{dB}$ and $27.12\\mathrm{dB}$ for the phase retrieval and high dynamic range tasks, representing improvements of $1.98\\mathrm{dB}$ and $4.39\\mathrm{dB}$ over the previous state-of-the-art results, respectively. We apply DAPS to large-scale text-conditioned latent diffusion models, showing its scalability to high-resolution image restoration tasks. Beyond image restoration tasks, DAPS achieves a PSNR of $31.49\\mathrm{dB}$ , showing a $2.70\\mathrm{dB}$ improvement over previous diffusion-based methods. We also demonstrate how DAPS can be applied to inverse problems on categorical data with pre-trained discrete diffusion models. Additionally, DAPS performs well with a small number of neural network evaluations (approximately 100), striking a better balance between efficiency and sample quality.", + "bbox": [ + 511, + 304, + 908, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 513, + 700, + 640, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion Models", + "text_level": 1, + "bbox": [ + 513, + 726, + 681, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [21, 26, 47-49] generate data by reversing a predefined noising process. Let the data distribution be $p(\\mathbf{x}_0)$ . We can define a series of noisy data distributions $p(\\mathbf{x};\\sigma)$ by adding Gaussian noise with a standard deviation of $\\sigma$ to the data. These form the time-margins of a stochastic differential equation (SDE) [26], given by $\\mathrm{d}\\mathbf{x}_t = \\sqrt{2\\dot{\\sigma}_t\\sigma_t}\\mathrm{d}\\mathbf{w}_t$ , where $\\sigma_{t}$ is the predefined noise schedule with $\\sigma_0 = 0$ and $\\sigma_T = \\sigma_{\\mathrm{max}}$ , $\\dot{\\sigma}_{t}$ is the time derivative of $\\sigma_{t}$ , and $\\mathbf{w}_t$ is a standard Wiener process. We use $\\mathbf{x}_t$ interchangeably with $\\mathbf{x}_{\\sigma_t}$ . For a sufficiently large $\\sigma_{\\mathrm{max}}$ , the", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20896", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/74776871326b0fdad20d79ec6020c5589d7bf42dac02bfa177d58618b1d8eff0.jpg", + "image_caption": [ + "Figure 2. An illustration of our method on phase retrieval. Given a noisy sample $\\mathbf{x}_t$ , we first solve the reverse diffusion process to obtain $\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)$ . Using this, we construct an approximation for $p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y})$ . Next, we sample $\\mathbf{x}_{0|y} \\sim p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y})$ with multiple steps of Langevin dynamics, then perturb it with the forward diffusion process to obtain a sample with slightly less noise. We repeat this process until the noise reduces to zero." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 477, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "distribution $p(\\mathbf{x};\\sigma_{\\mathrm{max}})$ converges to pure Gaussian noise $\\mathcal{N}(\\mathbf{0},\\sigma_{\\mathrm{max}}^2\\mathbf{I})$", + "bbox": [ + 89, + 392, + 483, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To sample from data distribution $p(\\mathbf{x}_0)$ , we first draw an initial sample from $\\mathcal{N}(\\mathbf{0},\\sigma_{\\mathrm{max}}^2\\mathbf{I})$ , then solve the reverse SDE: $\\mathrm{d}\\mathbf{x}_t = -2\\dot{\\sigma}_t\\sigma_t\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t;\\sigma_t)\\mathrm{d}t + \\sqrt{2\\dot{\\sigma}_t\\sigma_t}\\mathrm{d}\\mathbf{w}_t$ where $\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t;\\sigma_t)$ is the time-dependent score function [47, 49]. Here $\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t;\\sigma_t)$ is unknown, but can be approximated by training a diffusion model $s_\\theta (\\mathbf{x}_t,\\sigma_t)$ such that $s_\\theta (\\mathbf{x}_t,\\sigma_t)\\approx \\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t;\\sigma_t)$ .", + "bbox": [ + 89, + 422, + 483, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Bayesian Inverse Problems with Diffusion", + "text_level": 1, + "bbox": [ + 89, + 537, + 447, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inverse problems aim to recover data from partial, potentially noisy measurements. Formally, solving an inverse problem involves finding the inversion to a forward model that describes the measurement process. In general, a forward model takes the form of $\\mathbf{y} = \\mathcal{A}(\\mathbf{x}_0) + \\mathbf{n}$ , where $\\mathcal{A}$ is the measurement function, $\\mathbf{x}_0$ represents the original data, $\\mathbf{y}$ is the observed measurement, and $\\mathbf{n}$ symbolizes the noise in the measurement process, often modeled as $\\mathbf{n} \\sim \\mathcal{N}(\\mathbf{0}, \\beta_{\\mathbf{y}}^2 \\mathbf{I})$ . In a Bayesian framework, $\\mathbf{x}_0$ comes from the posterior distribution $p(\\mathbf{x}_0 \\mid \\mathbf{y}) \\propto p(\\mathbf{x}_0) p(\\mathbf{y} \\mid \\mathbf{x}_0)$ . Here $p(\\mathbf{x}_0)$ is a prior distribution that can be estimated from a given dataset, and $p(\\mathbf{y} \\mid \\mathbf{x}_0) = \\mathcal{N}(\\mathcal{A}(\\mathbf{x}_0), \\beta_{\\mathbf{y}}^2 \\mathbf{I})$ models the noisy measurement process.", + "bbox": [ + 89, + 559, + 483, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When the prior $p(\\mathbf{x}_0)$ is modeled by a pre-trained diffusion model, we can modify reverse SDE to approximately sample from the posterior distribution following Bayes' rule,", + "bbox": [ + 89, + 755, + 483, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {d} \\mathbf {x} _ {t} = - 2 \\dot {\\sigma} _ {t} \\sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}; \\sigma_ {t}) \\mathrm {d} t \\\\ - 2 \\dot {\\sigma} _ {t} \\sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {y} \\mid \\mathbf {x} _ {t}) d t + \\sqrt {2 \\dot {\\sigma} _ {t} \\sigma_ {t}} d w _ {t}. \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 806, + 483, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, the noisy likelihood $\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{y}|\\mathbf{x}_t)$ is generally intractable. Multiple methods have been proposed to estimate the noisy likelihood [5, 8, 12, 36, 41, 46, 50, 54, 59].", + "bbox": [ + 89, + 854, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6cb7dc6f884d697e6f9679598e667577db9a753d0960fd74bde467fab905f146.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 90, + 867, + 167 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3ffbaebd72a20ddd23ee5eb974a9822a9e2197f5d8c966a6cae3f669c1e6ff12.jpg", + "image_caption": [ + "(a) Diffusion models in pixel spaces.", + "(b) Latent diffusion models.", + "Figure 3. Probabilistic graphical models of the decoupled noise annealing process in (a) pixel and (b) latent spaces respectively. Here $\\mathbf{x}_t \\sim p(\\mathbf{x}_t; \\sigma_t)$ , and $\\mathbf{y}$ is the observed measurement. We factorize $p(\\mathbf{x}_t \\mid \\mathbf{y})$ based on these probabilistic graphs." + ], + "image_footnote": [], + "bbox": [ + 545, + 186, + 870, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One predominant approach is the DPS algorithm [13], which estimates $p(\\mathbf{y} \\mid \\mathbf{x}_t) \\approx p(\\mathbf{y} \\mid \\mathbf{x}_0 = \\mathbb{E}[\\mathbf{x}_0 \\mid \\mathbf{x}_t])$ . Another line of work [7, 18, 29, 30, 54] solves linear inverse problems by running the reverse diffusion in the spectral domain via singular value decomposition (SVD). Other methods bypass direct computation of this likelihood by interleaving optimization [2, 9, 24, 31, 45, 56, 59] or projection [10, 23, 30, 54] steps with normal diffusion sampling steps.", + "bbox": [ + 511, + 359, + 906, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite promising empirical success, we find that this line of approaches faces challenges in solving more difficult inverse problems when the forward model is highly nonlinear. Accurately solving the reverse SDE in Eq. (1) requires the solver to take a very small step size $\\Delta t > 0$ , causing $\\mathbf{x}_t$ and $\\mathbf{x}_{t + \\Delta t}$ to be very close to each other. Consequently, $\\mathbf{x}_t$ can only correct minor errors in $\\mathbf{x}_{t + \\Delta t}$ , but oftentimes fails to address larger, global errors that require substantial changes to $\\mathbf{x}_{t + \\Delta t}$ . One such failure case is given in Fig. 4.", + "bbox": [ + 511, + 479, + 908, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 628, + 604, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Decoupled Annealing Posterior Sampling", + "text_level": 1, + "bbox": [ + 511, + 652, + 867, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of solving the reverse-time SDE in Eq. (1), we propose a new noise annealing process that reduces the dependency between samples at consecutive time steps, as illustrated in Figs. 2, 3a and 3b. Unlike previous methods, we ensure $\\mathbf{x}_t$ and $\\mathbf{x}_{t + \\Delta t}$ are conditionally independent given $\\mathbf{x}_0$ . To generate sample $\\mathbf{x}_t$ from $\\mathbf{x}_{t + \\Delta t}$ , we follow a two-step procedure: (1) sampling $\\mathbf{x}_{0|\\mathbf{y}} \\sim p(\\mathbf{x}_0 \\mid \\mathbf{x}_{t + \\Delta t}, \\mathbf{y})$ , and (2) sampling $\\mathbf{x}_t \\sim \\mathcal{N}(\\mathbf{x}_{0|\\mathbf{y}}, \\sigma_t^2 \\mathbf{I})$ . We repeat this process, gradually reducing noise until $\\mathbf{x}_0$ is sampled. We call this process decoupled noise annealing, which is justified by the proposition below.", + "bbox": [ + 511, + 675, + 908, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1. Suppose $\\mathbf{x}_{t_1}$ is sampled from the time-marginal $p(\\mathbf{x}_{t_1} \\mid \\mathbf{y})$ , then", + "bbox": [ + 511, + 844, + 906, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t _ {2}} \\sim \\mathbb {E} _ {\\mathbf {x} _ {0} \\sim p (\\mathbf {x} _ {0} | \\mathbf {x} _ {t _ {1}}, \\mathbf {y})} [ \\mathcal {N} (\\mathbf {x} _ {0}, \\sigma_ {t _ {2}} ^ {2} I) ] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 883, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20897", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "satisfies the time-marginal $p(\\mathbf{x}_{t_2} \\mid \\mathbf{y})$ .", + "bbox": [ + 89, + 90, + 341, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark. The key idea of Proposition 1 is to enable the sampling from $p(\\mathbf{x}_{t_2} \\mid \\mathbf{y})$ for any noise level $\\sigma_{t_2}$ given any sample $x_{t_1}$ at another noise level $\\sigma_{t_1}$ . For a sufficiently large $\\sigma_T$ , one can assume $p(\\mathbf{x}_T \\mid \\mathbf{y}) \\approx p(\\mathbf{x}_T; \\sigma_T) \\approx \\mathcal{N}(\\mathbf{0}, \\sigma_T^2\\mathbf{I})$ . Starting from $\\mathbf{x}_T$ , we can iteratively sample from $p(\\mathbf{x}_t \\mid \\mathbf{y})$ with $\\sigma_t$ annealed down from $\\sigma_T$ to 0.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first step of our decoupled noise annealing requires sampling $\\mathbf{x}_{0|\\mathbf{y}}\\sim p(\\mathbf{x}_0\\mid \\mathbf{x}_t,\\mathbf{y})$ where $\\mathbf{x}_t$ and $\\mathbf{y}$ are known. Since $\\mathbf{y}$ is conditionally independent from $\\mathbf{x}_t$ given $\\mathbf{x}_0$ , we can deduce from the Bayes' rule that", + "bbox": [ + 89, + 207, + 483, + 267 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}, \\mathbf {y}\\right) = \\frac {p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}\\right) p \\left(\\mathbf {y} \\mid \\mathbf {x} _ {0} , \\mathbf {x} _ {t}\\right)}{p \\left(\\mathbf {y} \\mid \\mathbf {x} _ {t}\\right)} \\\\ \\propto p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}\\right) p (\\mathbf {y} \\mid \\mathbf {x} _ {0}). \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 276, + 482, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To sample $\\mathbf{x}_{0|\\mathbf{y}}$ from this unnormalized distribution, MCMC methods such as Langevin dynamics [55] and Hamiltonian Monte Carlo [4] can be adopted (as explained in Appendix A.1). The updating rule of Langevin dynamics [55] is given by", + "bbox": [ + 89, + 340, + 483, + 416 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {x} _ {0} ^ {(j + 1)} = \\mathbf {x} _ {0} ^ {(j)} + \\eta \\nabla_ {\\mathbf {x} _ {0} ^ {(j)}} \\log p (\\mathbf {x} _ {0} ^ {(j)} \\mid \\mathbf {x} _ {t}) \\\\ + \\eta \\nabla_ {\\mathbf {x} _ {0} ^ {(j)}} \\log p (\\mathbf {y} \\mid \\mathbf {x} _ {0} ^ {(j)}) + \\sqrt {2 \\eta} \\boldsymbol {\\epsilon} _ {j}, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 420, + 482, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta > 0$ is the step size and $\\epsilon_{j} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . When $\\eta \\to 0$ and $j \\to \\infty$ , the sample $\\mathbf{x}_0^{(j)}$ will be approximately distributed according to $p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y})$ .", + "bbox": [ + 89, + 479, + 483, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Practical Design Choice of DAPS", + "text_level": 1, + "bbox": [ + 89, + 537, + 380, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following Eq. (4), we discuss different ways to approximate the conditional distribution $p(\\mathbf{x}_0 \\mid \\mathbf{x}_t)$ due to its intractability. A line of previous works on diffusion posterior sampling [5, 13, 46] approximate $p(\\mathbf{x}_0 \\mid \\mathbf{x}_t)$ by a Gaussian distribution,", + "bbox": [ + 89, + 561, + 483, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}\\right) \\approx \\mathcal {N} \\left(\\mathbf {x} _ {0}; \\hat {\\mathbf {x}} _ {0} \\left(\\mathbf {x} _ {t}\\right), r _ {t} ^ {2} \\boldsymbol {I}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 632, + 482, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)$ is an estimator of $\\mathbf{x}_0$ given $\\mathbf{x}_t$ typically specified as $\\mathbb{E}[\\mathbf{x}_0\\mid \\mathbf{x}_t]$ , and the variance $r_t^2$ is specified using heuristics.", + "bbox": [ + 89, + 660, + 483, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another possible way to estimate $p(\\mathbf{x}_0 \\mid \\mathbf{x}_t)$ is to decompose it via Bayes' rule, and substitute the intractable $p(\\mathbf{x}_0)$ by the distribution $p(\\mathbf{x}_0; \\sigma_{t_{\\min}})$ with a small Gaussian noise for numerical stability, i.e.,", + "bbox": [ + 89, + 705, + 483, + 766 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\mathbf {x} _ {0}} \\log p (\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}) = \\nabla_ {\\mathbf {x} _ {0}} \\log p (\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}) + \\nabla_ {\\mathbf {x} _ {0}} \\log p (\\mathbf {x} _ {0}) \\\\ \\approx \\nabla_ {\\mathbf {x} _ {0}} \\log p \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) + s _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {0}, t _ {\\min }\\right). \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 777, + 482, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite being a more accurate estimation, the diffusion-score estimation Eq. (6) is much more expensive than using a Gaussian approximation when applied to Eq. (4), as it requires calling the score function multiple times. Empirical studies in Sec. 4.4 show that using a Gaussian approximation can", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "achieve comparable results as the diffusion-score estimation, but is significantly more time-efficient. In practice, we adopt the Gaussian approximation and compute $\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)$ by solving the (unconditional) probability flow ODE starting at $\\mathbf{x}_t$ . We leave details in Appendix F.2.", + "bbox": [ + 511, + 90, + 905, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the measurement noise is an isotropic Gaussian, i.e., $\\mathbf{n} \\sim \\mathcal{N}(\\mathbf{0}, \\beta_{\\mathbf{y}}^{2}\\mathbf{I})$ , the update rule simplifies to:", + "bbox": [ + 511, + 167, + 906, + 198 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {x} _ {0} ^ {(j + 1)} = \\mathbf {x} _ {0} ^ {(j)} - \\eta \\nabla_ {\\mathbf {x} _ {0} ^ {(j)}} \\frac {\\left\\| \\mathbf {x} _ {0} ^ {(j)} - \\hat {\\mathbf {x}} _ {0} (\\mathbf {x} _ {t}) \\right\\| ^ {2}}{2 r _ {t} ^ {2}} \\\\ - \\eta \\nabla_ {\\mathbf {x} _ {0} ^ {(j)}} \\frac {\\| \\mathcal {A} \\left(\\mathbf {x} _ {0} ^ {(j)}\\right) - \\mathbf {y} \\| ^ {2}}{2 \\beta_ {\\mathbf {y}} ^ {2}} + \\sqrt {2 \\eta} \\boldsymbol {\\epsilon} _ {j}, \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 205, + 906, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combining Eq. (7) with Proposition 1, we implement Decoupled Annealing Posterior Sampling (DAPS) as Algorithm 1 in Appendix A. In particular, given a noise schedule $\\sigma_{t}$ and a time discretization $\\{t_i,i = 0,\\dots ,N_A\\}$ , we iteratively sample $\\mathbf{x}_{t_i}$ from the measurement-conditioned time-marginal $p(\\mathbf{x}_{t_i}|\\mathbf{y})$ for $i = N_{A},N_{A} - 1,\\ldots ,0$ , following Eq. (2) in Proposition 1. This algorithm provides an approximate sample $\\mathbf{x}_0$ from the posterior distribution $p(\\mathbf{x}_0|\\mathbf{y})$ .", + "bbox": [ + 511, + 289, + 906, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The computational cost of Langevin dynamics mostly originates from evaluating the measurement function $\\mathcal{A}$ . In most image restoration tasks, this operation is significantly more efficient compared to evaluating the diffusion model. As demonstrated in Appendix E.1, the MCMC sampling introduces only a small overhead to the sampling process.", + "bbox": [ + 511, + 411, + 905, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. DAPS with Latent Diffusion Model", + "text_level": 1, + "bbox": [ + 511, + 508, + 818, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a pretrained encoder $\\mathcal{E}$ and decoder $\\mathcal{D}$ , the latent diffusion models (LDMs) [40] are trained to fit the latent distribution $p(\\mathbf{z}_0)$ , where $\\mathbf{z}_0$ is given by the encoder $\\mathbf{z}_0 = \\mathcal{E}(\\mathbf{x}_0)$ . One can recover $\\mathbf{x}_0$ from the decoder $\\mathbf{x}_0 = \\mathcal{D}(\\mathbf{z}_0)$ . We found the proposed method can be adapted to sampling with pre-trained latent diffusion models by factorizing the probabilistic graphical model for the latent diffusion process, as shown in Fig. 3b. We refer to this as LatentDAPS in the following paper.", + "bbox": [ + 511, + 531, + 906, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accordingly, we can derive the following Langevin dynamics updating rule from Eq. (4):", + "bbox": [ + 511, + 667, + 906, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {z} _ {0} ^ {(j + 1)} = \\mathbf {z} _ {0} ^ {(j)} + \\eta \\nabla_ {\\mathbf {z} _ {0} ^ {(j)}} \\log p (\\mathbf {z} _ {0} ^ {(j)} | \\mathbf {z} _ {t}) \\\\ + \\eta \\nabla_ {\\mathbf {z} _ {0} ^ {(j)}} \\log p (\\mathbf {y} \\mid \\mathcal {D} \\left(\\mathbf {z} _ {0} ^ {(j)}\\right)) + \\sqrt {2 \\eta} \\epsilon_ {j}. \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 698, + 906, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While LatentDAPS is a straightforward and natural extension of DAPS compared to recent latent diffusion inverse problem solvers [42, 43, 45], it demonstrates comparable or even superior performance across several tasks in Tab. 1 and Tab. 4. More detailed derivation and proofs are shown in Appendix B.", + "bbox": [ + 511, + 750, + 906, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Discussion with Existing Methods", + "text_level": 1, + "bbox": [ + 511, + 848, + 807, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comparison with other posterior sampling methods. Unlike most previous algorithms, our sampling method does not", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20898", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d53d44f9f84ecef58cdf9639c8b2983023f8e2a8adc1fc6d785f6bf5d7c6410e.jpg", + "image_caption": [ + "Figure 4. DAPS vs. DPS on 2D synthetic data. Consecutive sampling steps are close to each other for DPS but not for DAPS. Here DAPS approximates the posterior better than DPS." + ], + "image_footnote": [], + "bbox": [ + 114, + 92, + 460, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "solve a specific SDE/ODE; instead, we recursively sample from the time-marginal $p(\\mathbf{x}_t \\mid \\mathbf{y})$ with noise annealing to zero. We therefore decouple the dependency of $\\mathbf{x}_t$ and $\\mathbf{x}_{t + \\Delta t}$ in the sampling process. We argue that decoupling helps correct the errors accumulated in early diffusion sampling steps by allowing non-local transitions. This is particularly important when the measurement function is nonlinear. As a concrete example, Fig. 4 compares DAPS and DPS when solving a 2D nonlinear inverse problem with a Gaussian mixture prior. We visualize the trajectories of $\\mathbf{x}_t$ from time $T$ to 0 for both methods. For DAPS, points on the trajectory have significantly larger variations compared to those of DPS. As a result, DPS converges to wrong solutions, but DAPS is able to approximate the true posterior distribution. We provide more discussions in Appendix G.", + "bbox": [ + 88, + 334, + 485, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with optimization-based methods. Although our method does not directly involve inner-loop optimization, we highlight that it connects with existing optimization-based solvers for inverse problems. For example, ReSample [45], DiffPIR [59], and DCDP [31] alternate between denoising optimizing, and resampling to solve inverse problems. However, these methods solve the MAP estimation problems, while DAPS draws samples from the posterior distribution. In particular, instantiating DAPS with Langevin dynamics and Gaussian approximation resembles [45] if we set standard deviation $\\beta_{\\mathbf{y}}\\rightarrow 0$ and assume $\\eta \\beta_{\\mathbf{y}}^{2}$ is constant in Eq. (18).", + "bbox": [ + 89, + 563, + 486, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 760, + 223, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 786, + 281, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method using both pixel-space and latent diffusion models. For pixel-based diffusion experiments, we leverage the pre-trained diffusion models trained by [13] on the FFHQ dataset and the pre-trained model from [16] on the ImageNet dataset. For latent diffusion models, we use the same pre-trained models as [45]: the unconditional", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LDM-VQ4 trained on FFHQ and ImageNet by [40]. The autoencoder's downsample factor is 4. We use the same time step discretization and noise schedule as EDM [26].", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned in Sec. 3, we implement a few-step Euler ODE solver to compute $\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)$ , while maintaining the same number of neural function evaluations (NFE) for different noise levels. In our experiments, we use DAPS-1K for all linear tasks and DAPS-4K for all nonlinear tasks. DAPS-1K uses 4 ODE solver NFE and 250 annealing scheduling steps, while DAPS-4K uses 10 NFE and 400 steps, respectively. We discuss the effect of choosing these configurations in Sec. 4.4. The corresponding latent version, LatentDAPS, follows the same settings but performs sampling in the latent space of VAEs. We use 100 and 50 Langevin steps per denoising iteration for DAPS and LatentDAPS respectively, and tune learning rates separately for each task. Further details on model configurations, samplers, and other hyperparameters are provided in Appendix F, along with more discussion on sampling efficiency in Sec. 4.2 and Appendix E.1.", + "bbox": [ + 511, + 138, + 908, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and metrics. Adopting the previous convention, we test our method on two image datasets, FFHQ $256 \\times 256$ [25] and ImageNet $256 \\times 256$ [15]. To evaluate our method, we use 100 images from the validation set for both FFHQ and ImageNet. We include peak signal-to-noise-ratio (PSNR), structural similarity index measure (SSIM), Learned Perceptual Image Patch Similarity (LPIPS)[58] score and Fréchet inception distance (FID)[20] as our main evaluation metrics. For both our method and baselines, we use the versions implemented in pic[28] with all images normalized to the range [0, 1]. The replace-pooling option is enabled for LPIPS evaluation.", + "bbox": [ + 511, + 383, + 908, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inverse problems. We evaluate our method with a series of linear and nonlinear tasks. For linear inverse problems, we consider (1) super-resolution, (2) Gaussian deblurring, (3) motion deblurring, (4) inpainting (with a box mask), and (5) inpainting (with a $70\\%$ random mask). For Gaussian and motion deblurring, kernels of size $61 \\times 61$ with standard deviations of 3.0 and 0.5, respectively, are used. In the super-resolution task, a bicubic resizer downscales images by a factor of 4. The box inpainting task uses a random box of size $128 \\times 128$ to mask the original images, while random mask inpainting uses a generated random mask where each pixel has a $70\\%$ chance of being masked, following the settings in [45].", + "bbox": [ + 511, + 566, + 908, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider three nonlinear inverse problems: (1) phase retrieval, (2) high dynamic range (HDR) reconstruction, and (3) nonlinear deblurring. Due to the inherent instability of phase retrieval, we adopt the strategy from DPS [13], using an oversampling rate of 2.0 and reporting the best result out of four independent samples. The goal of HDR reconstruction is to recover a higher dynamic range image (factor of 2) from a low dynamic range image. For nonlinear deblurring, we use the default setting as described in [53].", + "bbox": [ + 511, + 763, + 910, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20899", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All linear and nonlinear measurements are subject to white Gaussian noise with a standard deviation of $\\beta_{\\mathbf{y}} = 0.05$ . Further details regarding the forward measuring functions for each task and their respective hyperparameters are provided in the Appendix F.", + "bbox": [ + 89, + 90, + 482, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our methods with the following baselines: DDRM [30], DPS [13], DDNM [54], DCDP [31], FPS-SMC [18], DiffPIR [59] for pixel-based diffusion model experiments. We compare the latent diffusion version of our methods with PSLD [42] and ReSample [45], which also operate in the latent space. Note that DDRM, DDNM, FPS-SMC, and PSLD cannot handle nonlinear inverse problems. We especially RED-diff [35] for nonlinear experiments and a traditional method HIO [14] for phase retrieval.", + "bbox": [ + 89, + 167, + 483, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 316, + 228, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show quantitative results for FFHQ and ImageNet dataset in Tab. 1. The qualitative comparisons are provided in Fig. 1. As shown in Tab. 1, our method achieves comparable or even better performance across all selected linear inverse problems. Moreover, our methods are remarkably stable when handling nonlinear inverse problems. For example, although existing methods such as DPS can recover high-quality images from phase retrieval measurements, they suffer from an extremely high failure rate, resulting in a relatively low PSNR and a high LPIPS. However, DAPS demonstrates superior stability across different samples and achieves a significantly higher success rate, resulting in a substantial improvement in PSNR and LPIPS as evidenced by the results in Tab. 1. Moreover, DAPS captures and recovers much more fine-grained details in measurements compared to existing baselines. We include more samples and comparisons in Appendix I for further illustration.", + "bbox": [ + 89, + 339, + 483, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To understand the evolution of the noise annealing procedure, we evaluate two crucial trajectories in our method: 1) the estimated mean of $p(\\mathbf{x}_0 \\mid \\mathbf{x}_t)$ , denoted as $\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)$ ; and 2) samples from $p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y})$ obtained through Langevin dynamics, denoted as $\\mathbf{x}_{0|\\mathbf{y}}$ . We assess image quality and measurement error across these trajectories, as depicted in Fig. 7. The measurement error for $\\mathbf{x}_{0|\\mathbf{y}} \\sim p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y})$ remains consistently low throughout the process, while the measurement error for $\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)$ continuously decreases. Consequently, metrics such as PSNR and LPIPS exhibit stable improvement as the noise level is annealed.", + "bbox": [ + 89, + 597, + 483, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, our method can generate diversified samples when the measurements contain less information. To illustrate this, we selected two tasks where the posterior distributions may have multiple modes: (1) super-resolution by a factor of 16, and (2) box inpainting with a box size of $192 \\times 192$ . Some samples from DAPS are shown in Fig. 5, demonstrating DAPS's ability to produce diverse samples while preserving the measurement information. We show more samples in Appendix I for further demonstration.", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f6078223ab458b1e78da54664d9af0d9574b4c2e03574d3252341b6863a28a26.jpg", + "image_caption": [ + "(a) Super resolution of a factor 16" + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 924, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/534690a36ae5de8fa48795aa0c5e1b381c4df00b5bda5e277e80f05e89c80af7.jpg", + "image_caption": [ + "(b) Inpainting of $192\\times 192$ box", + "Figure 5. Sample diversity. We present several diverse samples generated by the DAPS under two sparse measurements whose posterior distributions contain multiple modes. DAPS produces a variety of samples with distinct features, including differences in expression, wearings, and hairstyles." + ], + "image_footnote": [], + "bbox": [ + 516, + 243, + 924, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DAPS with large-scale text-conditioned latent diffusion. Due to the flexibility of DAPS, pretrained large-scale text-conditioned latent diffusion models (LDMs) [38-40] can also serve as the prior model $p(\\mathbf{x}_0 \\mid \\mathbf{c})$ when given a text prompt c. To demonstrate DAPS's potential, we use Stable Diffusion v2.1 [40] to present LatentDAPS results on three tasks with $768 \\times 768$ resolution images, as shown in Fig. 1. Additionally, we provide the text prompts used, along with a detailed discussion, in Appendix B.1, where we also include a quantitative evaluation.", + "bbox": [ + 511, + 491, + 906, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DAPS with discrete diffusion. DAPS can be naturally extended to support posterior sampling with categorical prior distribution modeled by discrete diffusion models [3, 6, 33]. Instead of sampling $\\mathbf{x}_{0|\\mathbf{y}}$ by Langevin MC or Hamiltonian MC in continuous space, we use the Metropolis-Hasting algorithm to replace Eq. (4). We validate DAPS with discrete diffusion on a quantized MNIST dataset. We leave the details to Appendix C.", + "bbox": [ + 511, + 643, + 908, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Further results on MRI", + "text_level": 1, + "bbox": [ + 511, + 772, + 732, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further validate the effectiveness of DAPS on compressed sensing multi-coil MRI, which is a medical imaging technique that helps reduce the scan time of MRI via subsampling. We train a diffusion model using the pipeline from [26] with fastMRI knee dataset [57]. Tab. 2 shows that DAPS outperforms existing diffusion-based baselines. We include more qualitative results in Appendix H.", + "bbox": [ + 511, + 794, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20900", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fdb9832836af13edefa37318dc64d413e3ce3b7246c80fba43c6fdd9c857c671.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative evaluation on FFHQ and ImageNet on 5 linear and 3 nonlinear tasks. Performance comparison of different methods on FFHQ (left) and ImageNet (right) in the image domain. For linear tasks, we report the mean performance (PSNR, SSIM and LPIPS) across 100 validation images. We provide both the mean and standard deviation for nonlinear tasks, where results exhibit higher instability. The FID scores are evaluated with the same 100 validation images. The best and second-best results within each type of task are indicated by bold and underlined marks, respectively. DAPS demonstrates superior performance on most tasks, with LatentDAPS showing competitive results. All tasks are using noisy measurement with noise level $\\beta_{\\mathbf{y}} = 0.05$ ." + ], + "table_body": "
TaskTypeMethodFFHQImageNet
PSNR (↑)SSIM (↑)LPIPS (↓)FID (↓)PSNR (↑)SSIM (↑)LPIPS (↓)FID (↓)
Super resolution 4xPixelDAPS (ours)29.070.8180.17751.4425.890.6940.27683.57
DPS25.860.7530.26981.0721.130.4890.361106.32
DDRM26.580.7820.28279.2522.620.5210.324103.85
DDNM28.030.7950.19764.6223.960.6040.47598.62
DCDP28.660.8070.17853.81----
FPS-SMC28.420.8130.20449.2524.820.7030.31397.51
DiffPIR26.64-0.26065.7723.18-0.371106.32
LatentLatentDAPS (ours)27.480.8010.18259.6225.060.6730.27684.37
PSLD24.350.6490.28774.3625.420.6940.36097.45
ReSample23.290.5940.39293.1822.610.5760.370113.42
Inpaint (box)PixelDAPS (ours)24.070.8140.13343.1021.430.7250.214109.85
DPS22.510.7920.20961.2718.940.7220.257126.52
DDRM22.260.8010.20778.6218.630.7330.254116.37
DDNM24.470.8370.23546.5921.640.7480.319103.97
DCDP23.890.7600.16345.23----
FPS-SMC24.860.8230.14648.3422.160.7260.208111.58
LatentLatentDAPS (ours)23.990.8020.19446.5217.190.6240.340145.63
PSLD24.220.8130.15843.0220.100.6940.465146.53
ReSample20.060.7490.18453.2118.290.6310.262127.84
Inpaint (random)PixelDAPS (ours)31.120.8440.09832.1728.440.7750.13554.25
DPS25.460.8230.20369.2023.520.7450.29787.53
DDNM29.910.8170.12144.3731.160.8410.19163.84
DCDP30.690.8420.14252.51----
FPS-SMC28.210.8230.26161.2324.520.7010.31679.12
LatentLatentDAPS (ours)30.710.8130.14136.4127.590.7720.16461.62
PSLD30.310.8090.22147.2131.300.7830.33783.21
ReSample29.610.7460.14039.8527.500.7560.14359.87
Gaussian deblurringPixelDAPS (ours)29.190.8170.16553.3326.150.6840.25375.68
DPS25.870.7640.21979.7520.310.5980.397116.42
DDRM24.930.7320.23992.4321.260.5640.443146.89
DDNM28.200.8040.21657.8328.060.7030.27881.43
DCDP27.500.6990.30486.43----
FPS-SMC26.540.7730.25367.4523.910.6010.38791.72
DiffPIR27.36-0.23659.6522.80-0.35593.36
LatentLatentDAPS (ours)27.930.7640.23464.5225.050.6680.34578.51
PSLD23.270.6310.31689.5125.860.6880.39091.39
ReSample26.390.7140.25571.6925.970.7030.25465.35
Motion deblurringPixelDAPS (ours)29.660.8470.15739.4927.860.7660.19661.83
DPS24.520.8010.24665.2318.960.6290.423137.81
DCDP25.080.5120.364125.13----
FPS-SMC27.390.8260.22748.3224.520.6470.32687.43
DiffPIR26.57-0.25565.7824.01-0.36694.63
LatentLatentDAPS (ours)27.000.8140.28346.8426.830.7450.29673.62
PSLD22.310.6780.33696.1520.850.5940.511124.67
ReSample27.410.8230.19844.7226.940.7380.22766.89
Phase retrievalPixelDAPS (ours)30.63±3.130.851±0.0720.139±0.06042.7125.78±6.920.743±0.0840.254±0.12582.67
DPS17.64±2.970.441±0.1290.410±0.090104.5216.81±3.610.427±0.1430.447±0.099197.54
RED-diff15.60±4.480.398±0.1950.596±0.092167.4314.98±3.750.386±0.0570.536±0.129212.24
DCDP28.65±8.090.781±0.2170.203±0.19668.13----
LatentLatentDAPS (ours)29.16±3.550.796±0.0890.199±0.07854.2620.54±6.410.612±0.1140.361±0.150129.54
ReSample21.60±8.100.648±0.1540.406±0.22484.32----
ClassicalHIO13.53±2.500.359±0.0930.726±0.068268.09----
Nonlinear deblurPixelDAPS (ours)28.29±1.770.783±0.0360.155±0.03249.3827.73±3.230.724±0.0480.169±0.05659.87
DPS23.39±2.010.623±0.0820.278±0.06091.3122.49±3.200.591±0.1010.306±0.081101.41
RED-diff30.86±0.510.795±0.0280.160±0.03443.8430.07±1.410.754±0.0230.211±0.08351.22
DCDP27.92±2.640.779±0.0670.183±0.05151.96----
LatentLatentDAPS (ours)28.11±1.750.713±0.0410.235±0.04953.6325.34±3.440.615±0.0570.314±0.08076.73
ReSample28.24±1.690.742±0.0390.185±0.03951.6226.20±3.710.653±0.0640.206±0.05761.16
High dynamic rangePixelDAPS (ours)27.12±3.530.752±0.0410.162±0.07242.9726.30±4.100.717±0.0670.175±0.10764.19
DPS22.73±6.070.591±0.1410.264±0.156112.8219.23±2.520.582±0.0820.503±0.106146.23
RED-diff22.16±3.410.512±0.0830.258±0.089108.3222.03±5.900.601±0.0940.274±0.198113.48
LatentLatentDAPS (ours)25.94±2.870.751±0.0560.223±0.08074.8323.64±4.100.609±0.0530.269±0.09993.51
ReSample25.65±3.570.732±0.0590.182±0.08567.2225.11±4.210.633±0.0490.198±0.08987.66
", + "bbox": [ + 142, + 87, + 854, + 801 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20901", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7675a9626a3aa211cda5e21aa15e1652c332ca03d231d43c6811f68b7c50e196.jpg", + "image_caption": [ + "(a) Gaussian deblur" + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 493, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d319c6ba091b9dfa55da4314901ef3f7835e0615d0612c26b128cd32dcd3555.jpg", + "image_caption": [ + "(b) Phase retrieval" + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 864, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e725e42bb7e60312e10069a4ceb4a5bc38608cb4b0537da63a2ac32011fd871.jpg", + "image_caption": [ + "Figure 6. Evaluation of time cost and sample quality. The x-axis indicates the single image sampling time on an A100-SXM4-80GB GPU and y-axis shows the LPIPS. The evaluation uses 100 FFHQ images.", + "Figure 7. Quantitative evaluations of image quality metrics and measurement error throughout the sampling process. Here $\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)$ stands for the estimated mean of the Gaussian approximated distribution and $\\mathbf{x}_{0|\\mathbf{y}}\\sim p(\\mathbf{x}_0\\mid \\mathbf{x}_t,\\mathbf{y})$ means the samples from measurement conditioned time-marginal. The measurement errors are normalized between [0, 1] for visualization." + ], + "image_footnote": [], + "bbox": [ + 93, + 297, + 906, + 434 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9a38a8609aa7ecce679a50f432ccc1ae2d020e5abfbdc188ba27dd36d6c45475.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
×4×8
PSNR (↑)SSIM (↑)Meas err (↓)PSNR (↑)SSIM (↑)Meas err (↓)
DAPS31.480.7621.5729.010.6811.28
DPS26.130.6209.9020.820.5366.74
DiffPIR28.310.63210.5526.780.5887.79
ScoreMRI [10]25.970.46810.8325.010.4058.36
CSGM [23]28.780.7101.5226.150.6251.14
", + "bbox": [ + 91, + 498, + 483, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 636, + 254, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Time efficiency of DAPS. We run DAPS with various configurations to test its performance under different computing budgets. Specifically, we evaluate DAPS with the number of function evaluations (NFE) of the diffusion models ranging from 50 to 4K, with configurations specified in Appendix E.1. Sec. 4.2 shows the relation between the sample quality of DAPS measured by LPIPS with its time cost, which indicates that DAPS is able to generate high-quality samples with a reasonable time cost.", + "bbox": [ + 89, + 657, + 483, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Different Design choices of DAPS. We conduct an ablation study to demonstrate the effectiveness of Gaussian approximation compared with the more principled but more expensive diffusion-score estimation. The results in Tab. 3 demonstrate that Gaussian approximation is $6 \\sim 7$ times faster than diffusion-score estimation with the cost of a slight performance drop, effectively trading off performance", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/61b4273aaa2323e2241f894a4be80408d661fd8bab54feb69dc834119a476b64.jpg", + "table_caption": [ + "Table 2. Quantitative results of compressed sensing MRI. We consider $\\times 4$ and $\\times 8$ subsampling ratio. Meas err stands for ${\\ell }_{2}$ data consistency error." + ], + "table_footnote": [], + "table_body": "
Gaussian deblurringHigh dynamic range
PSNR (↑)LPIPS (↓)Run time/sPSNR (↑)LPIPS (↓)Run time/s
Gaussian approx.29.400.17064.724.850.18161.4
diffusion-score29.570.160462.325.330.193540.0
Relative improvement-0.575%-6.25%614.53%-1.89%-6.22%779.48%
", + "bbox": [ + 514, + 498, + 906, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Comparison of approaches to approximate $p(\\mathbf{x}_0 \\mid \\mathbf{x}_t)$ . Three tasks on 10 FFHQ images with DAPS-1K consistently shown Gaussian approximation achieve huge speed-up while slightly sacrificing the performance compared with diffusion-score estimation.", + "bbox": [ + 511, + 560, + 906, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for efficiency.", + "bbox": [ + 511, + 632, + 609, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More ablation studies. We include more ablation studies of DAPS in Appendix I.1, including the number of function evaluations, the number of ODE steps, the annealing noise scheduling, and different measurement noise levels.", + "bbox": [ + 511, + 647, + 906, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 723, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, we propose Decoupled Annealing Posterior Sampling (DAPS) for solving inverse problems, particularly those with complex nonlinear measurement processes such as phase retrieval. Our method decouples consecutive sample points in a diffusion sampling trajectory, allowing them to vary considerably, thereby enabling DAPS to explore a larger solution space. Empirically, we demonstrate that DAPS generates samples with better visual quality and stability compared to existing methods when solving a wide range of challenging inverse problems.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20902", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We are grateful to Pika for providing the computing resources essential for this research. We also extend our thanks to the Kortschak Scholars Fellowship for supporting B.Z. and W.C. at Caltech. J.B. acknowledges support from the Wally Baer and Jeri Weiss Postdoctoral Fellowship. A.A. is supported in part by Bren endowed chair and by the AI2050 senior fellow program at Schmidt Sciences.", + "bbox": [ + 89, + 114, + 485, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 234, + 187, + 250 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kazunori Akiyama, Antxon Alberdi, Walter Alef, Keiichi Asada, Rebecca Azulay, Anne-Kathrin Baczko, David Ball, Mislav Balokovic, John Barrett, Dan Bintley, et al. First m87 event horizon telescope results. iv. imaging the central supermassive black hole. The Astrophysical Journal Letters, 875(1):L4, 2019. 1", + "[2] Marius Arvinte, Sriram Vishwanath, Ahmed H. Tewfik, and Jonathan I. Tamir. Deep j-sense: Accelerated mri reconstruction via unrolled alternating optimization, 2021. 3", + "[3] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981-17993, 2021. 6", + "[4] Michael Betancourt and Mark Girolami. Hamiltonian monte carlo for hierarchical models. Current trends in Bayesian methodology with applications, 79(30):2-4, 2015. 4, 1", + "[5] Benjamin Boys, Mark Girolami, Jakiw Pidstrigach, Sebastian Reich, Alan Mosca, and O. Deniz Akyildiz. Tweedie moment projected diffusions for inverse problems, 2023. 2, 3, 4", + "[6] Andrew Campbell, Joe Benton, Valentin De Bortoli, Thomas Rainforth, George Deligiannidis, and Arnaud Doucet. A continuous time framework for discrete denoising models. Advances in Neural Information Processing Systems, 35:28266-28279, 2022. 6", + "[7] Gabriel Cardoso, Yazid Janati El Idrissi, Sylvain Le Corff, and Eric Moulines. Monte carlo guided diffusion for bayesian linear inverse problems, 2023. 3", + "[8] Gabriel Victorino Cardoso, Yazid Janati, Sylvain Le Corff, and Éric Moulines. Monte carlo guided diffusion for bayesian linear inverse problems. ArXiv, abs/2308.07983, 2023. 3", + "[9] Junqing Chen and Haibo Liu. An alternating direction method of multipliers for inverse lithography problem, 2023. 3", + "[10] Hyungjin Chung and Jong Chul Ye. Score-based diffusion models for accelerated mri. Medical Image Analysis, page 102479, 2022. 1, 3, 8, 13", + "[11] Hyungjin Chung, Eun Sun Lee, and Jong Chul Ye. Mr image denoising and super-resolution using regularized reverse diffusion. IEEE Transactions on Medical Imaging, 42(4): 922-934, 2022. 1", + "[12] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. In Advances in Neural Information Processing Systems, 2022. 3", + "[13] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior" + ], + "bbox": [ + 93, + 258, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 3, 4, 5, 6, 10, 11, 13", + "[14] J. Dainty and James Fienup. Phase retrieval and image reconstruction for astronomy. Image Recovery: Theory Appl, 13, 1987. 6", + "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 5", + "[16] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In Advances in Neural Information Processing Systems, 2021. 5", + "[17] Zolnamar Dorjsembe, Hsing-Kuo Pao, Sodtavilan Odonchimed, and Furen Xiao. Conditional diffusion models for semantic 3d brain mri synthesis. IEEE Journal of Biomedical and Health Informatics, 2024. 1", + "[18] Zehao Dou and Yang Song. Diffusion posterior sampling for linear inverse problem solving: A filtering perspective. In The Twelfth International Conference on Learning Representations, 2024. 3, 6, 11", + "[19] Berthy T. Feng, Jamie Smith, Michael Rubinstein, Huiwen Chang, Katherine L. Bouman, and William T. Freeman. Score-based diffusion models as principled priors for inverse imaging. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10520-10531, 2023. 1", + "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. 5", + "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 2", + "[22] Alex Ling Yu Hung, Kai Zhao, Haoxin Zheng, Ran Yan, Steven S Raman, Demetri Terzopoulos, and Kyunghyun Sung. Med-cdiff: Conditional medical image generation with diffusion models. Bioengineering, 10(11):1258, 2023. 1", + "[23] Ajil Jalal, Marius Arvinte, Giannis Daras, Eric Price, Alexandros G Dimakis, and Jon Tamir. Robust compressed sensing mri with deep generative priors. In Advances in Neural Information Processing Systems, pages 14938-14954. Curran Associates, Inc., 2021. 1, 2, 3, 8, 13", + "[24] Ulugbek Kamilov, Charles Bouman, Gregory Buzzard, and Brendt Wohlberg. Plug-and-play methods for integrating physical and learned models in computational imaging: Theory, algorithms, and applications. IEEE Signal Processing Magazine, 40:85-97, 2023. 3", + "[25] T. Karras, S. Laine, and T. Aila. A style-based generator architecture for generative adversarial networks. IEEE Transactions on Pattern Analysis & Machine Intelligence, 43(12): 4217-4228, 2021. 5, 10", + "[26] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, 2022. 2, 5, 6, 10, 13" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20903", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models, 2024. 2", + "[28] Sergey Kastryulin, Jamil Zakirov, Denis Prokopenko, and Dmitry V. Dylov. Pytorch image quality: Metrics for image quality assessment, 2022. 5", + "[29] Bahjat Kawar, Gregory Vaksman, and Michael Elad. SNIPS: Solving noisy inverse problems stochastically. Advances in Neural Information Processing Systems, 34:21757-21769, 2021. 3", + "[30] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 6, 11", + "[31] Xiang Li, Soo Min Kwon, Ismail R. Alkhouri, Saiprasad Ravishankar, and Qing Qu. Decoupled data consistency with diffusion purification for image restoration, 2024. 3, 5, 6, 11", + "[32] Hanyuan Liu, Jinbo Xing, Minshan Xie, Chengze Li, and Tien-Tsin Wong. Improved diffusion-based image colorization via piggybacked models, 2023. 2", + "[33] Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion language modeling by estimating the ratios of the data distribution. arXiv preprint arXiv:2310.16834, 2023. 6", + "[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models, 2022. 1", + "[35] Morteza Mardani, Jiaming Song, Jan Kautz, and Arash Vahdat. A variational perspective on solving inverse problems with diffusion models. arXiv preprint arXiv:2305.04391, 2023. 6, 11, 14", + "[36] Xiangming Meng and Yoshiyuki Kabashima. Diffusion model based posterior sampling for noisy linear inverse problems. arXiv preprint arXiv:2211.12343, 2022. 3", + "[37] Nicholas Metropolis, Arianna W Rosenbluth, Marshall N Rosenbluth, Augusta H Teller, and Edward Teller. Equation of state calculations by fast computing machines. The journal of chemical physics, 21(6):1087-1092, 1953. 1", + "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 6", + "[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 4", + "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 4, 5, 6, 3", + "[41] Litu Rout, Yujia Chen, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Beyond first-order tweedie: Solving inverse problems using latent diffusion, 2023. 3", + "[42] Litu Rout, Negin Raoof, Giannis Daras, Constantine Caramanis, Alex Dimakis, and Sanjay Shakkottai. Solving linear inverse problems provably via posterior sampling with latent diffusion models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4, 6, 9, 11" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Litu Rout, Yujia Chen, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Beyond first-order tweedie: Solving inverse problems using latent diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9472-9481, 2024. 4", + "[44] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2023. 1", + "[45] Bowen Song, Soo Min Kwon, Zecheng Zhang, Xinyu Hu, Qing Qu, and Liyue Shen. Solving inverse problems with latent diffusion models via hard data consistency. In The Twelfth International Conference on Learning Representations, 2024, 2, 3, 4, 5, 6, 9, 11", + "[46] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 3, 4", + "[47] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Neural Information Processing Systems, 2019. 2, 3", + "[48] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020.", + "[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 1, 2, 3, 10", + "[50] Yang Song, Liyue Shen, Lei Xing, and Stefano Ermon. Solving inverse problems in medical imaging with score-based generative models. In International Conference on Learning Representations, 2022. 1, 2, 3", + "[51] He Sun and Katherine L. Bouman. Deep probabilistic imaging: Uncertainty quantification and multi-modal solution characterization for computational imaging, 2020. 1", + "[52] Yu Sun, Zihui Wu, Yifan Chen, Berthy Feng, and Katherine L. Bouman. Provable probabilistic imaging using score-based generative priors. *ArXiv*, abs/2310.10835, 2023. 1", + "[53] Phong Tran, Anh Tran, Quynh Phung, and Minh Hoai. Explore image deblurring via encoded blur kernel space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5", + "[54] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. The Eleventh International Conference on Learning Representations, 2023. 3, 6, 11", + "[55] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688. CiteSeer, 2011. 4", + "[56] Zihui Wu, Yu Sun, Yifan Chen, Bingliang Zhang, Yisong Yue, and Katherine L Bouman. Principled probabilistic imaging using diffusion models as plug-and-play priors. arXiv preprint arXiv:2405.18782, 2024. 3" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20904", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Jure Zbontar, Florian Knoll, Anuroop Sriram, Matthew J. Muckley, Mary Bruno, Aaron Defazio, Marc Parente, Krzysztof J. Geras, Joe Katsnelson, Hersh Chandarana, Zizhao Zhang, Michal Drozdal, Adriana Romero, Michael G. Rabbat, Pascal Vincent, James Pinkerton, Duo Wang, Nafissa Yakubova, Erich Owens, C. Lawrence Zitnick, Michael P. Recht, Daniel K. Sodickson, and Yvonne W. Lui. fastmri: An open dataset and benchmarks for accelerated MRI. CoRR, abs/1811.08839, 2018. 6, 13", + "[58] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 5", + "[59] Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising diffusion models for plug-and-play image restoration. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (NTIRE), 2023. 1, 3, 5, 6, 11, 13" + ], + "bbox": [ + 91, + 90, + 485, + 357 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20905", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/d107e339-bf08-4731-b903-e7f3360d585d_model.json b/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/d107e339-bf08-4731-b903-e7f3360d585d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..466dd3028214d53f4726025db3ff26e7b4ba2165 --- /dev/null +++ b/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/d107e339-bf08-4731-b903-e7f3360d585d_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.131, + 0.895, + 0.154 + ], + "angle": 0, + "content": "Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.181, + 0.377, + 0.2 + ], + "angle": 0, + "content": "Bingliang Zhang\\*,1" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.2, + 0.353, + 0.217 + ], + "angle": 0, + "content": "Chenlin Meng3" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.217, + 0.443, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)California Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.458, + 0.181, + 0.57, + 0.197 + ], + "angle": 0, + "content": "Wenda Chu\\*,1" + }, + { + "type": "text", + "bbox": [ + 0.433, + 0.199, + 0.603, + 0.215 + ], + "angle": 0, + "content": "Anima Anandkumar" + }, + { + "type": "text", + "bbox": [ + 0.467, + 0.217, + 0.547, + 0.233 + ], + "angle": 0, + "content": "2NVIDIA" + }, + { + "type": "text", + "bbox": [ + 0.65, + 0.181, + 0.775, + 0.197 + ], + "angle": 0, + "content": "Julius Berner\\(^{2,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.685, + 0.199, + 0.779, + 0.217 + ], + "angle": 0, + "content": "Yang Song" + }, + { + "type": "text", + "bbox": [ + 0.57, + 0.217, + 0.832, + 0.235 + ], + "angle": 0, + "content": "3Stanford University 4OpenAI" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.269, + 0.422, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.271, + 0.903, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.416, + 0.421, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.562, + 0.907, + 0.632 + ], + "angle": 0, + "content": "Figure 1. Overview of Decoupled Annealing Posterior Sampling (DAPS). Our method provides a flexible and effective framework for solving inverse problems through a decoupled posterior sampling process. In (a)(b), we present DAPS visual results on FFHQ and ImageNet at a resolution of 256, and in (c), on natural images at a resolution of 768. In (d), we display DAPS results on compressed sensing multi-coil MRI (CS-MRI). DAPS effectively addresses nonlinear inverse problems as well as medical imaging MRI challenges. Additionally, DAPS can be enhanced using large-scale latent diffusion models (LDMs) [40], as shown in (c)." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.644, + 0.327, + 0.66 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.678, + 0.486, + 0.86 + ], + "angle": 0, + "content": "Diffusion models have recently achieved success in solving Bayesian inverse problems with learned data priors. Current methods build on top of the diffusion sampling process, where each denoising step makes small modifications to samples from the previous step. However, this process struggles to correct errors from earlier sampling steps, leading to worse performance in complicated nonlinear inverse problems, such as phase retrieval. To address this challenge, we propose a new method called Decoupled Annealing Posterior Sampling (DAPS) that relies on a novel noise annealing process. Specifically, we decouple consecutive steps in a diffusion sampling trajectory, allowing them to vary consid" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.909, + 0.753 + ], + "angle": 0, + "content": "erably from one another while ensuring their time-margins anneal to the true posterior as we reduce noise levels. This approach enables the exploration of a larger solution space, improving the success rate for accurate reconstructions. We demonstrate that DAPS significantly improves sample quality and stability across multiple image restoration tasks, particularly in complicated nonlinear inverse problems." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.769, + 0.646, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Inverse problems are ubiquitous in science and engineering, with applications ranging from image restoration [13, 30, 34, 44, 49, 59], medical imaging [10, 11, 17, 22, 23, 50] to astrophotography [1, 19, 51, 52]. Solving an inverse problem involves finding the underlying signal \\(\\mathbf{x}_0\\) from its partial, noisy measurement \\(\\mathbf{y}\\). Since the measurement process is typically noisy and many-to-one, inverse problems do not have a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.875, + 0.444, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\)Work partially done at the California Institute of Technology." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.368, + 0.9 + ], + "angle": 0, + "content": "*These authors contributed equally to this work" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.875, + 0.444, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20895" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.242 + ], + "angle": 0, + "content": "unique solution; instead, multiple solutions may exist that are consistent with the observed measurement. In the Bayesian inverse problem framework, the solution space is characterized by the posterior distribution \\( p(\\mathbf{x}_0 \\mid \\mathbf{y}) \\propto p(\\mathbf{y} \\mid \\mathbf{x}_0) p(\\mathbf{x}_0) \\), where \\( p(\\mathbf{y} \\mid \\mathbf{x}_0) \\) represents the noisy measurement process, and \\( p(\\mathbf{x}_0) \\) is the prior distribution. In this work, we aim to solve Bayesian inverse problems where the measurement process \\( p(\\mathbf{y} \\mid \\mathbf{x}_0) \\) is known, and the prior distribution \\( p(\\mathbf{x}_0) \\) is captured by a deep generative model trained on a corresponding dataset." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.246, + 0.487, + 0.518 + ], + "angle": 0, + "content": "As score-based diffusion models [21, 26, 27, 47-49] have risen to dominance in modeling high-dimensional data distributions like images, audio, and video, they have become the leading method for estimating the prior distribution \\( p(\\mathbf{x}_0) \\) in Bayesian inverse problems. A diffusion model generates a sample \\( \\mathbf{x}_0 \\) by smoothly removing noise from an unstructured initial noise sample \\( \\mathbf{x}_T \\) through solving stochastic differential equations (SDEs). In particular, each step of the sampling process recursively converts a noisy sample \\( \\mathbf{x}_{t + \\Delta t} \\), where \\( \\Delta t > 0 \\) denotes the step size, to a slightly less noisy sample \\( \\mathbf{x}_t \\) until \\( t = 0 \\). This iterative structure in the diffusion sampling process can be leveraged to facilitate Bayesian inverse problem solving. In fact, prior research [5, 13, 46] has shown that given measurement \\( \\mathbf{y} \\) and a diffusion model prior \\( p(\\mathbf{x}_0) \\), we can sample from the posterior distribution \\( p(\\mathbf{x}_0\\mid \\mathbf{y}) \\) by perturbing a reverse-time SDE using an approximate gradient of the measurement process, \\( \\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{y}\\mid \\mathbf{x}_t) \\), at every step of the SDE solver." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.486, + 0.807 + ], + "angle": 0, + "content": "Despite the remarkable success of these methods in solving many real-world inverse problems, like image colorization [30, 32, 49], image super-resolution [13, 30, 46, 49], computed tomography [45, 50], and magnetic resonance imaging [23, 50], they face significant challenges in more complex inverse problems with nonlinear measurement processes, such as phase retrieval and nonlinear motion deblurring. This is partially because, in prior methods, each denoising step approximately samples from the distribution \\( p(\\mathbf{x}_t \\mid \\mathbf{x}_{t + \\Delta t}, \\mathbf{y}) \\). This causes \\( \\mathbf{x}_t \\) and \\( \\mathbf{x}_{t + \\Delta t} \\) to be close to each other because of using a small step size \\( \\Delta t \\) in discretizing the reverse-time SDE. As a result, \\( \\mathbf{x}_t \\) can at most correct local errors in \\( \\mathbf{x}_{t + \\Delta t} \\) but struggles to correct global errors that require significant modifications to the prior sample. This issue is exacerbated when the methods are applied to complicated, nonlinear inverse problems, such as phase retrieval, where they often converge to undesired samples that are consistent with the measurement but reside in low-probability areas of the prior distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.902 + ], + "angle": 0, + "content": "To address this challenge, we propose a new framework for solving general inverse problems, termed Decoupled Annealing Posterior Sampling (DAPS). Our method employs a new noise annealing process inspired by the diffusion sampling process, where we decouple the consecutive samples \\(\\mathbf{x}_{t + \\Delta t}\\) and \\(\\mathbf{x}_t\\) in the sampling trajectory, allowing samplers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.304 + ], + "angle": 0, + "content": "to correct large, global errors made in earlier steps. Instead of repetitively sampling from \\( p(\\mathbf{x}_t \\mid \\mathbf{x}_{t + \\Delta t}, \\mathbf{y}) \\) as in previous methods, which restricts the distances between consecutive samples \\( \\mathbf{x}_{t + \\Delta t} \\) and \\( \\mathbf{x}_t \\), DAPS recursively samples from the marginal distribution \\( p(\\mathbf{x}_t \\mid \\mathbf{y}) \\). As illustrated in Fig. 2, we factorize the time marginal \\( p(\\mathbf{x}_t \\mid \\mathbf{y}) \\) into three conditional distributions and sample from them in turn by solving the reverse diffusion process, simulating Langevin dynamics, and adding noise according to the forward diffusion process. We show that this creates approximate samples from corresponding marginal distributions. As the noise gradually anneals to zero, the time marginal \\( p(\\mathbf{x}_t \\mid \\mathbf{y}) \\) smoothly converges to the posterior distribution \\( p(\\mathbf{x}_0 \\mid \\mathbf{y}) \\), providing samples that approximately solve the Bayesian inverse problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.909, + 0.683 + ], + "angle": 0, + "content": "Empirically, our method demonstrates significantly improved performance on various inverse problems compared to existing approaches, ranging from image restoration to compressed sensing MRI. Our method can be combined with diffusion models in both raw pixel space and learned latent space, which we refer to as DAPS and LatentDAPS, respectively. As shown in Fig. 1, both methods provide superior reconstructions with improved visual perceptual quality across a wide range of nonlinear inverse problems. Our approach exhibits remarkable stability and sampling quality, particularly for challenging nonlinear inverse problems. On the FFHQ 256 dataset, we achieve PSNR values of \\(30.72\\mathrm{dB}\\) and \\(27.12\\mathrm{dB}\\) for the phase retrieval and high dynamic range tasks, representing improvements of \\(1.98\\mathrm{dB}\\) and \\(4.39\\mathrm{dB}\\) over the previous state-of-the-art results, respectively. We apply DAPS to large-scale text-conditioned latent diffusion models, showing its scalability to high-resolution image restoration tasks. Beyond image restoration tasks, DAPS achieves a PSNR of \\(31.49\\mathrm{dB}\\), showing a \\(2.70\\mathrm{dB}\\) improvement over previous diffusion-based methods. We also demonstrate how DAPS can be applied to inverse problems on categorical data with pre-trained discrete diffusion models. Additionally, DAPS performs well with a small number of neural network evaluations (approximately 100), striking a better balance between efficiency and sample quality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.701, + 0.642, + 0.717 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.727, + 0.682, + 0.741 + ], + "angle": 0, + "content": "2.1. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Diffusion models [21, 26, 47-49] generate data by reversing a predefined noising process. Let the data distribution be \\( p(\\mathbf{x}_0) \\). We can define a series of noisy data distributions \\( p(\\mathbf{x};\\sigma) \\) by adding Gaussian noise with a standard deviation of \\( \\sigma \\) to the data. These form the time-margins of a stochastic differential equation (SDE) [26], given by \\( \\mathrm{d}\\mathbf{x}_t = \\sqrt{2\\dot{\\sigma}_t\\sigma_t}\\mathrm{d}\\mathbf{w}_t \\), where \\( \\sigma_{t} \\) is the predefined noise schedule with \\( \\sigma_0 = 0 \\) and \\( \\sigma_T = \\sigma_{\\mathrm{max}} \\), \\( \\dot{\\sigma}_{t} \\) is the time derivative of \\( \\sigma_{t} \\), and \\( \\mathbf{w}_t \\) is a standard Wiener process. We use \\( \\mathbf{x}_t \\) interchangeably with \\( \\mathbf{x}_{\\sigma_t} \\). For a sufficiently large \\( \\sigma_{\\mathrm{max}} \\), the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20896" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.478, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.283, + 0.484, + 0.381 + ], + "angle": 0, + "content": "Figure 2. An illustration of our method on phase retrieval. Given a noisy sample \\(\\mathbf{x}_t\\), we first solve the reverse diffusion process to obtain \\(\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)\\). Using this, we construct an approximation for \\(p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y})\\). Next, we sample \\(\\mathbf{x}_{0|y} \\sim p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y})\\) with multiple steps of Langevin dynamics, then perturb it with the forward diffusion process to obtain a sample with slightly less noise. We repeat this process until the noise reduces to zero." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.424 + ], + "angle": 0, + "content": "distribution \\(p(\\mathbf{x};\\sigma_{\\mathrm{max}})\\) converges to pure Gaussian noise \\(\\mathcal{N}(\\mathbf{0},\\sigma_{\\mathrm{max}}^2\\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.531 + ], + "angle": 0, + "content": "To sample from data distribution \\(p(\\mathbf{x}_0)\\), we first draw an initial sample from \\(\\mathcal{N}(\\mathbf{0},\\sigma_{\\mathrm{max}}^2\\mathbf{I})\\), then solve the reverse SDE: \\(\\mathrm{d}\\mathbf{x}_t = -2\\dot{\\sigma}_t\\sigma_t\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t;\\sigma_t)\\mathrm{d}t + \\sqrt{2\\dot{\\sigma}_t\\sigma_t}\\mathrm{d}\\mathbf{w}_t\\) where \\(\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t;\\sigma_t)\\) is the time-dependent score function [47, 49]. Here \\(\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t;\\sigma_t)\\) is unknown, but can be approximated by training a diffusion model \\(s_\\theta (\\mathbf{x}_t,\\sigma_t)\\) such that \\(s_\\theta (\\mathbf{x}_t,\\sigma_t)\\approx \\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t;\\sigma_t)\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.538, + 0.449, + 0.554 + ], + "angle": 0, + "content": "2.2. Bayesian Inverse Problems with Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.485, + 0.756 + ], + "angle": 0, + "content": "Inverse problems aim to recover data from partial, potentially noisy measurements. Formally, solving an inverse problem involves finding the inversion to a forward model that describes the measurement process. In general, a forward model takes the form of \\(\\mathbf{y} = \\mathcal{A}(\\mathbf{x}_0) + \\mathbf{n}\\), where \\(\\mathcal{A}\\) is the measurement function, \\(\\mathbf{x}_0\\) represents the original data, \\(\\mathbf{y}\\) is the observed measurement, and \\(\\mathbf{n}\\) symbolizes the noise in the measurement process, often modeled as \\(\\mathbf{n} \\sim \\mathcal{N}(\\mathbf{0}, \\beta_{\\mathbf{y}}^2 \\mathbf{I})\\). In a Bayesian framework, \\(\\mathbf{x}_0\\) comes from the posterior distribution \\(p(\\mathbf{x}_0 \\mid \\mathbf{y}) \\propto p(\\mathbf{x}_0) p(\\mathbf{y} \\mid \\mathbf{x}_0)\\). Here \\(p(\\mathbf{x}_0)\\) is a prior distribution that can be estimated from a given dataset, and \\(p(\\mathbf{y} \\mid \\mathbf{x}_0) = \\mathcal{N}(\\mathcal{A}(\\mathbf{x}_0), \\beta_{\\mathbf{y}}^2 \\mathbf{I})\\) models the noisy measurement process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.485, + 0.802 + ], + "angle": 0, + "content": "When the prior \\(p(\\mathbf{x}_0)\\) is modeled by a pre-trained diffusion model, we can modify reverse SDE to approximately sample from the posterior distribution following Bayes' rule," + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.808, + 0.484, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {d} \\mathbf {x} _ {t} = - 2 \\dot {\\sigma} _ {t} \\sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}; \\sigma_ {t}) \\mathrm {d} t \\\\ - 2 \\dot {\\sigma} _ {t} \\sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {y} \\mid \\mathbf {x} _ {t}) d t + \\sqrt {2 \\dot {\\sigma} _ {t} \\sigma_ {t}} d w _ {t}. \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Here, the noisy likelihood \\(\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{y}|\\mathbf{x}_t)\\) is generally intractable. Multiple methods have been proposed to estimate the noisy likelihood [5, 8, 12, 36, 41, 46, 50, 54, 59]." + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.091, + 0.869, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.173, + 0.807, + 0.186 + ], + "angle": 0, + "content": "(a) Diffusion models in pixel spaces." + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.187, + 0.871, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.27, + 0.785, + 0.282 + ], + "angle": 0, + "content": "(b) Latent diffusion models." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.294, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 3. Probabilistic graphical models of the decoupled noise annealing process in (a) pixel and (b) latent spaces respectively. Here \\(\\mathbf{x}_t \\sim p(\\mathbf{x}_t; \\sigma_t)\\), and \\(\\mathbf{y}\\) is the observed measurement. We factorize \\(p(\\mathbf{x}_t \\mid \\mathbf{y})\\) based on these probabilistic graphs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.907, + 0.481 + ], + "angle": 0, + "content": "One predominant approach is the DPS algorithm [13], which estimates \\( p(\\mathbf{y} \\mid \\mathbf{x}_t) \\approx p(\\mathbf{y} \\mid \\mathbf{x}_0 = \\mathbb{E}[\\mathbf{x}_0 \\mid \\mathbf{x}_t]) \\). Another line of work [7, 18, 29, 30, 54] solves linear inverse problems by running the reverse diffusion in the spectral domain via singular value decomposition (SVD). Other methods bypass direct computation of this likelihood by interleaving optimization [2, 9, 24, 31, 45, 56, 59] or projection [10, 23, 30, 54] steps with normal diffusion sampling steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.909, + 0.618 + ], + "angle": 0, + "content": "Despite promising empirical success, we find that this line of approaches faces challenges in solving more difficult inverse problems when the forward model is highly nonlinear. Accurately solving the reverse SDE in Eq. (1) requires the solver to take a very small step size \\(\\Delta t > 0\\), causing \\(\\mathbf{x}_t\\) and \\(\\mathbf{x}_{t + \\Delta t}\\) to be very close to each other. Consequently, \\(\\mathbf{x}_t\\) can only correct minor errors in \\(\\mathbf{x}_{t + \\Delta t}\\), but oftentimes fails to address larger, global errors that require substantial changes to \\(\\mathbf{x}_{t + \\Delta t}\\). One such failure case is given in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.63, + 0.605, + 0.645 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.654, + 0.868, + 0.671 + ], + "angle": 0, + "content": "3.1. Decoupled Annealing Posterior Sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.676, + 0.909, + 0.842 + ], + "angle": 0, + "content": "Instead of solving the reverse-time SDE in Eq. (1), we propose a new noise annealing process that reduces the dependency between samples at consecutive time steps, as illustrated in Figs. 2, 3a and 3b. Unlike previous methods, we ensure \\(\\mathbf{x}_t\\) and \\(\\mathbf{x}_{t + \\Delta t}\\) are conditionally independent given \\(\\mathbf{x}_0\\). To generate sample \\(\\mathbf{x}_t\\) from \\(\\mathbf{x}_{t + \\Delta t}\\), we follow a two-step procedure: (1) sampling \\(\\mathbf{x}_{0|\\mathbf{y}} \\sim p(\\mathbf{x}_0 \\mid \\mathbf{x}_{t + \\Delta t}, \\mathbf{y})\\), and (2) sampling \\(\\mathbf{x}_t \\sim \\mathcal{N}(\\mathbf{x}_{0|\\mathbf{y}}, \\sigma_t^2 \\mathbf{I})\\). We repeat this process, gradually reducing noise until \\(\\mathbf{x}_0\\) is sampled. We call this process decoupled noise annealing, which is justified by the proposition below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.845, + 0.908, + 0.876 + ], + "angle": 0, + "content": "Proposition 1. Suppose \\(\\mathbf{x}_{t_1}\\) is sampled from the time-marginal \\(p(\\mathbf{x}_{t_1} \\mid \\mathbf{y})\\), then" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.884, + 0.908, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t _ {2}} \\sim \\mathbb {E} _ {\\mathbf {x} _ {0} \\sim p (\\mathbf {x} _ {0} | \\mathbf {x} _ {t _ {1}}, \\mathbf {y})} [ \\mathcal {N} (\\mathbf {x} _ {0}, \\sigma_ {t _ {2}} ^ {2} I) ] \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20897" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.343, + 0.108 + ], + "angle": 0, + "content": "satisfies the time-marginal \\(p(\\mathbf{x}_{t_2} \\mid \\mathbf{y})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.207 + ], + "angle": 0, + "content": "Remark. The key idea of Proposition 1 is to enable the sampling from \\( p(\\mathbf{x}_{t_2} \\mid \\mathbf{y}) \\) for any noise level \\( \\sigma_{t_2} \\) given any sample \\( x_{t_1} \\) at another noise level \\( \\sigma_{t_1} \\). For a sufficiently large \\( \\sigma_T \\), one can assume \\( p(\\mathbf{x}_T \\mid \\mathbf{y}) \\approx p(\\mathbf{x}_T; \\sigma_T) \\approx \\mathcal{N}(\\mathbf{0}, \\sigma_T^2\\mathbf{I}) \\). Starting from \\( \\mathbf{x}_T \\), we can iteratively sample from \\( p(\\mathbf{x}_t \\mid \\mathbf{y}) \\) with \\( \\sigma_t \\) annealed down from \\( \\sigma_T \\) to 0." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.208, + 0.485, + 0.268 + ], + "angle": 0, + "content": "The first step of our decoupled noise annealing requires sampling \\(\\mathbf{x}_{0|\\mathbf{y}}\\sim p(\\mathbf{x}_0\\mid \\mathbf{x}_t,\\mathbf{y})\\) where \\(\\mathbf{x}_t\\) and \\(\\mathbf{y}\\) are known. Since \\(\\mathbf{y}\\) is conditionally independent from \\(\\mathbf{x}_t\\) given \\(\\mathbf{x}_0\\), we can deduce from the Bayes' rule that" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.277, + 0.483, + 0.331 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}, \\mathbf {y}\\right) = \\frac {p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}\\right) p \\left(\\mathbf {y} \\mid \\mathbf {x} _ {0} , \\mathbf {x} _ {t}\\right)}{p \\left(\\mathbf {y} \\mid \\mathbf {x} _ {t}\\right)} \\\\ \\propto p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}\\right) p (\\mathbf {y} \\mid \\mathbf {x} _ {0}). \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.485, + 0.417 + ], + "angle": 0, + "content": "To sample \\(\\mathbf{x}_{0|\\mathbf{y}}\\) from this unnormalized distribution, MCMC methods such as Langevin dynamics [55] and Hamiltonian Monte Carlo [4] can be adopted (as explained in Appendix A.1). The updating rule of Langevin dynamics [55] is given by" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.421, + 0.483, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {x} _ {0} ^ {(j + 1)} = \\mathbf {x} _ {0} ^ {(j)} + \\eta \\nabla_ {\\mathbf {x} _ {0} ^ {(j)}} \\log p (\\mathbf {x} _ {0} ^ {(j)} \\mid \\mathbf {x} _ {t}) \\\\ + \\eta \\nabla_ {\\mathbf {x} _ {0} ^ {(j)}} \\log p (\\mathbf {y} \\mid \\mathbf {x} _ {0} ^ {(j)}) + \\sqrt {2 \\eta} \\boldsymbol {\\epsilon} _ {j}, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.484, + 0.532 + ], + "angle": 0, + "content": "where \\(\\eta > 0\\) is the step size and \\(\\epsilon_{j} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). When \\(\\eta \\to 0\\) and \\(j \\to \\infty\\), the sample \\(\\mathbf{x}_0^{(j)}\\) will be approximately distributed according to \\(p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y})\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.539, + 0.381, + 0.556 + ], + "angle": 0, + "content": "3.2. Practical Design Choice of DAPS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Following Eq. (4), we discuss different ways to approximate the conditional distribution \\( p(\\mathbf{x}_0 \\mid \\mathbf{x}_t) \\) due to its intractability. A line of previous works on diffusion posterior sampling [5, 13, 46] approximate \\( p(\\mathbf{x}_0 \\mid \\mathbf{x}_t) \\) by a Gaussian distribution," + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.633, + 0.483, + 0.651 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}\\right) \\approx \\mathcal {N} \\left(\\mathbf {x} _ {0}; \\hat {\\mathbf {x}} _ {0} \\left(\\mathbf {x} _ {t}\\right), r _ {t} ^ {2} \\boldsymbol {I}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.661, + 0.484, + 0.705 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)\\) is an estimator of \\(\\mathbf{x}_0\\) given \\(\\mathbf{x}_t\\) typically specified as \\(\\mathbb{E}[\\mathbf{x}_0\\mid \\mathbf{x}_t]\\), and the variance \\(r_t^2\\) is specified using heuristics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.707, + 0.484, + 0.767 + ], + "angle": 0, + "content": "Another possible way to estimate \\( p(\\mathbf{x}_0 \\mid \\mathbf{x}_t) \\) is to decompose it via Bayes' rule, and substitute the intractable \\( p(\\mathbf{x}_0) \\) by the distribution \\( p(\\mathbf{x}_0; \\sigma_{t_{\\min}}) \\) with a small Gaussian noise for numerical stability, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.779, + 0.483, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\mathbf {x} _ {0}} \\log p (\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}) = \\nabla_ {\\mathbf {x} _ {0}} \\log p (\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}) + \\nabla_ {\\mathbf {x} _ {0}} \\log p (\\mathbf {x} _ {0}) \\\\ \\approx \\nabla_ {\\mathbf {x} _ {0}} \\log p \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) + s _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {0}, t _ {\\min }\\right). \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Despite being a more accurate estimation, the diffusion-score estimation Eq. (6) is much more expensive than using a Gaussian approximation when applied to Eq. (4), as it requires calling the score function multiple times. Empirical studies in Sec. 4.4 show that using a Gaussian approximation can" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.167 + ], + "angle": 0, + "content": "achieve comparable results as the diffusion-score estimation, but is significantly more time-efficient. In practice, we adopt the Gaussian approximation and compute \\(\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)\\) by solving the (unconditional) probability flow ODE starting at \\(\\mathbf{x}_t\\). We leave details in Appendix F.2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.907, + 0.199 + ], + "angle": 0, + "content": "When the measurement noise is an isotropic Gaussian, i.e., \\(\\mathbf{n} \\sim \\mathcal{N}(\\mathbf{0}, \\beta_{\\mathbf{y}}^{2}\\mathbf{I})\\), the update rule simplifies to:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.206, + 0.907, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {x} _ {0} ^ {(j + 1)} = \\mathbf {x} _ {0} ^ {(j)} - \\eta \\nabla_ {\\mathbf {x} _ {0} ^ {(j)}} \\frac {\\left\\| \\mathbf {x} _ {0} ^ {(j)} - \\hat {\\mathbf {x}} _ {0} (\\mathbf {x} _ {t}) \\right\\| ^ {2}}{2 r _ {t} ^ {2}} \\\\ - \\eta \\nabla_ {\\mathbf {x} _ {0} ^ {(j)}} \\frac {\\| \\mathcal {A} \\left(\\mathbf {x} _ {0} ^ {(j)}\\right) - \\mathbf {y} \\| ^ {2}}{2 \\beta_ {\\mathbf {y}} ^ {2}} + \\sqrt {2 \\eta} \\boldsymbol {\\epsilon} _ {j}, \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Combining Eq. (7) with Proposition 1, we implement Decoupled Annealing Posterior Sampling (DAPS) as Algorithm 1 in Appendix A. In particular, given a noise schedule \\(\\sigma_{t}\\) and a time discretization \\(\\{t_i,i = 0,\\dots ,N_A\\}\\), we iteratively sample \\(\\mathbf{x}_{t_i}\\) from the measurement-conditioned time-marginal \\(p(\\mathbf{x}_{t_i}|\\mathbf{y})\\) for \\(i = N_{A},N_{A} - 1,\\ldots ,0\\), following Eq. (2) in Proposition 1. This algorithm provides an approximate sample \\(\\mathbf{x}_0\\) from the posterior distribution \\(p(\\mathbf{x}_0|\\mathbf{y})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.906, + 0.502 + ], + "angle": 0, + "content": "The computational cost of Langevin dynamics mostly originates from evaluating the measurement function \\(\\mathcal{A}\\). In most image restoration tasks, this operation is significantly more efficient compared to evaluating the diffusion model. As demonstrated in Appendix E.1, the MCMC sampling introduces only a small overhead to the sampling process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.51, + 0.819, + 0.525 + ], + "angle": 0, + "content": "3.3. DAPS with Latent Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.907, + 0.668 + ], + "angle": 0, + "content": "Given a pretrained encoder \\(\\mathcal{E}\\) and decoder \\(\\mathcal{D}\\), the latent diffusion models (LDMs) [40] are trained to fit the latent distribution \\(p(\\mathbf{z}_0)\\), where \\(\\mathbf{z}_0\\) is given by the encoder \\(\\mathbf{z}_0 = \\mathcal{E}(\\mathbf{x}_0)\\). One can recover \\(\\mathbf{x}_0\\) from the decoder \\(\\mathbf{x}_0 = \\mathcal{D}(\\mathbf{z}_0)\\). We found the proposed method can be adapted to sampling with pre-trained latent diffusion models by factorizing the probabilistic graphical model for the latent diffusion process, as shown in Fig. 3b. We refer to this as LatentDAPS in the following paper." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.907, + 0.698 + ], + "angle": 0, + "content": "Accordingly, we can derive the following Langevin dynamics updating rule from Eq. (4):" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.699, + 0.907, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {z} _ {0} ^ {(j + 1)} = \\mathbf {z} _ {0} ^ {(j)} + \\eta \\nabla_ {\\mathbf {z} _ {0} ^ {(j)}} \\log p (\\mathbf {z} _ {0} ^ {(j)} | \\mathbf {z} _ {t}) \\\\ + \\eta \\nabla_ {\\mathbf {z} _ {0} ^ {(j)}} \\log p (\\mathbf {y} \\mid \\mathcal {D} \\left(\\mathbf {z} _ {0} ^ {(j)}\\right)) + \\sqrt {2 \\eta} \\epsilon_ {j}. \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.84 + ], + "angle": 0, + "content": "While LatentDAPS is a straightforward and natural extension of DAPS compared to recent latent diffusion inverse problem solvers [42, 43, 45], it demonstrates comparable or even superior performance across several tasks in Tab. 1 and Tab. 4. More detailed derivation and proofs are shown in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.808, + 0.865 + ], + "angle": 0, + "content": "3.4. Discussion with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Comparison with other posterior sampling methods. Unlike most previous algorithms, our sampling method does not" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20898" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.093, + 0.462, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.277, + 0.486, + 0.32 + ], + "angle": 0, + "content": "Figure 4. DAPS vs. DPS on 2D synthetic data. Consecutive sampling steps are close to each other for DPS but not for DAPS. Here DAPS approximates the posterior better than DPS." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.335, + 0.486, + 0.562 + ], + "angle": 0, + "content": "solve a specific SDE/ODE; instead, we recursively sample from the time-marginal \\( p(\\mathbf{x}_t \\mid \\mathbf{y}) \\) with noise annealing to zero. We therefore decouple the dependency of \\( \\mathbf{x}_t \\) and \\( \\mathbf{x}_{t + \\Delta t} \\) in the sampling process. We argue that decoupling helps correct the errors accumulated in early diffusion sampling steps by allowing non-local transitions. This is particularly important when the measurement function is nonlinear. As a concrete example, Fig. 4 compares DAPS and DPS when solving a 2D nonlinear inverse problem with a Gaussian mixture prior. We visualize the trajectories of \\( \\mathbf{x}_t \\) from time \\( T \\) to 0 for both methods. For DAPS, points on the trajectory have significantly larger variations compared to those of DPS. As a result, DPS converges to wrong solutions, but DAPS is able to approximate the true posterior distribution. We provide more discussions in Appendix G." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.487, + 0.745 + ], + "angle": 0, + "content": "Comparison with optimization-based methods. Although our method does not directly involve inner-loop optimization, we highlight that it connects with existing optimization-based solvers for inverse problems. For example, ReSample [45], DiffPIR [59], and DCDP [31] alternate between denoising optimizing, and resampling to solve inverse problems. However, these methods solve the MAP estimation problems, while DAPS draws samples from the posterior distribution. In particular, instantiating DAPS with Langevin dynamics and Gaussian approximation resembles [45] if we set standard deviation \\(\\beta_{\\mathbf{y}}\\rightarrow 0\\) and assume \\(\\eta \\beta_{\\mathbf{y}}^{2}\\) is constant in Eq. (18)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.761, + 0.224, + 0.779 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.282, + 0.804 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.902 + ], + "angle": 0, + "content": "We evaluate our method using both pixel-space and latent diffusion models. For pixel-based diffusion experiments, we leverage the pre-trained diffusion models trained by [13] on the FFHQ dataset and the pre-trained model from [16] on the ImageNet dataset. For latent diffusion models, we use the same pre-trained models as [45]: the unconditional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.137 + ], + "angle": 0, + "content": "LDM-VQ4 trained on FFHQ and ImageNet by [40]. The autoencoder's downsample factor is 4. We use the same time step discretization and noise schedule as EDM [26]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.909, + 0.382 + ], + "angle": 0, + "content": "As mentioned in Sec. 3, we implement a few-step Euler ODE solver to compute \\(\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)\\), while maintaining the same number of neural function evaluations (NFE) for different noise levels. In our experiments, we use DAPS-1K for all linear tasks and DAPS-4K for all nonlinear tasks. DAPS-1K uses 4 ODE solver NFE and 250 annealing scheduling steps, while DAPS-4K uses 10 NFE and 400 steps, respectively. We discuss the effect of choosing these configurations in Sec. 4.4. The corresponding latent version, LatentDAPS, follows the same settings but performs sampling in the latent space of VAEs. We use 100 and 50 Langevin steps per denoising iteration for DAPS and LatentDAPS respectively, and tune learning rates separately for each task. Further details on model configurations, samplers, and other hyperparameters are provided in Appendix F, along with more discussion on sampling efficiency in Sec. 4.2 and Appendix E.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.909, + 0.563 + ], + "angle": 0, + "content": "Datasets and metrics. Adopting the previous convention, we test our method on two image datasets, FFHQ \\(256 \\times 256\\) [25] and ImageNet \\(256 \\times 256\\) [15]. To evaluate our method, we use 100 images from the validation set for both FFHQ and ImageNet. We include peak signal-to-noise-ratio (PSNR), structural similarity index measure (SSIM), Learned Perceptual Image Patch Similarity (LPIPS)[58] score and Fréchet inception distance (FID)[20] as our main evaluation metrics. For both our method and baselines, we use the versions implemented in pic[28] with all images normalized to the range [0, 1]. The replace-pooling option is enabled for LPIPS evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.763 + ], + "angle": 0, + "content": "Inverse problems. We evaluate our method with a series of linear and nonlinear tasks. For linear inverse problems, we consider (1) super-resolution, (2) Gaussian deblurring, (3) motion deblurring, (4) inpainting (with a box mask), and (5) inpainting (with a \\(70\\%\\) random mask). For Gaussian and motion deblurring, kernels of size \\(61 \\times 61\\) with standard deviations of 3.0 and 0.5, respectively, are used. In the super-resolution task, a bicubic resizer downscales images by a factor of 4. The box inpainting task uses a random box of size \\(128 \\times 128\\) to mask the original images, while random mask inpainting uses a generated random mask where each pixel has a \\(70\\%\\) chance of being masked, following the settings in [45]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.911, + 0.903 + ], + "angle": 0, + "content": "We consider three nonlinear inverse problems: (1) phase retrieval, (2) high dynamic range (HDR) reconstruction, and (3) nonlinear deblurring. Due to the inherent instability of phase retrieval, we adopt the strategy from DPS [13], using an oversampling rate of 2.0 and reporting the best result out of four independent samples. The goal of HDR reconstruction is to recover a higher dynamic range image (factor of 2) from a low dynamic range image. For nonlinear deblurring, we use the default setting as described in [53]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20899" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.166 + ], + "angle": 0, + "content": "All linear and nonlinear measurements are subject to white Gaussian noise with a standard deviation of \\(\\beta_{\\mathbf{y}} = 0.05\\). Further details regarding the forward measuring functions for each task and their respective hyperparameters are provided in the Appendix F." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.484, + 0.304 + ], + "angle": 0, + "content": "Baselines. We compare our methods with the following baselines: DDRM [30], DPS [13], DDNM [54], DCDP [31], FPS-SMC [18], DiffPIR [59] for pixel-based diffusion model experiments. We compare the latent diffusion version of our methods with PSLD [42] and ReSample [45], which also operate in the latent space. Note that DDRM, DDNM, FPS-SMC, and PSLD cannot handle nonlinear inverse problems. We especially RED-diff [35] for nonlinear experiments and a traditional method HIO [14] for phase retrieval." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.317, + 0.229, + 0.332 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.597 + ], + "angle": 0, + "content": "We show quantitative results for FFHQ and ImageNet dataset in Tab. 1. The qualitative comparisons are provided in Fig. 1. As shown in Tab. 1, our method achieves comparable or even better performance across all selected linear inverse problems. Moreover, our methods are remarkably stable when handling nonlinear inverse problems. For example, although existing methods such as DPS can recover high-quality images from phase retrieval measurements, they suffer from an extremely high failure rate, resulting in a relatively low PSNR and a high LPIPS. However, DAPS demonstrates superior stability across different samples and achieves a significantly higher success rate, resulting in a substantial improvement in PSNR and LPIPS as evidenced by the results in Tab. 1. Moreover, DAPS captures and recovers much more fine-grained details in measurements compared to existing baselines. We include more samples and comparisons in Appendix I for further illustration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.763 + ], + "angle": 0, + "content": "To understand the evolution of the noise annealing procedure, we evaluate two crucial trajectories in our method: 1) the estimated mean of \\( p(\\mathbf{x}_0 \\mid \\mathbf{x}_t) \\), denoted as \\( \\hat{\\mathbf{x}}_0(\\mathbf{x}_t) \\); and 2) samples from \\( p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y}) \\) obtained through Langevin dynamics, denoted as \\( \\mathbf{x}_{0|\\mathbf{y}} \\). We assess image quality and measurement error across these trajectories, as depicted in Fig. 7. The measurement error for \\( \\mathbf{x}_{0|\\mathbf{y}} \\sim p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y}) \\) remains consistently low throughout the process, while the measurement error for \\( \\hat{\\mathbf{x}}_0(\\mathbf{x}_t) \\) continuously decreases. Consequently, metrics such as PSNR and LPIPS exhibit stable improvement as the noise level is annealed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In addition, our method can generate diversified samples when the measurements contain less information. To illustrate this, we selected two tasks where the posterior distributions may have multiple modes: (1) super-resolution by a factor of 16, and (2) box inpainting with a box size of \\(192 \\times 192\\). Some samples from DAPS are shown in Fig. 5, demonstrating DAPS's ability to produce diverse samples while preserving the measurement information. We show more samples in Appendix I for further demonstration." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.925, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.231, + 0.81, + 0.242 + ], + "angle": 0, + "content": "(a) Super resolution of a factor 16" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.244, + 0.925, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.384, + 0.802, + 0.396 + ], + "angle": 0, + "content": "(b) Inpainting of \\(192\\times 192\\) box" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.409, + 0.907, + 0.478 + ], + "angle": 0, + "content": "Figure 5. Sample diversity. We present several diverse samples generated by the DAPS under two sparse measurements whose posterior distributions contain multiple modes. DAPS produces a variety of samples with distinct features, including differences in expression, wearings, and hairstyles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.908, + 0.643 + ], + "angle": 0, + "content": "DAPS with large-scale text-conditioned latent diffusion. Due to the flexibility of DAPS, pretrained large-scale text-conditioned latent diffusion models (LDMs) [38-40] can also serve as the prior model \\( p(\\mathbf{x}_0 \\mid \\mathbf{c}) \\) when given a text prompt c. To demonstrate DAPS's potential, we use Stable Diffusion v2.1 [40] to present LatentDAPS results on three tasks with \\( 768 \\times 768 \\) resolution images, as shown in Fig. 1. Additionally, we provide the text prompts used, along with a detailed discussion, in Appendix B.1, where we also include a quantitative evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.765 + ], + "angle": 0, + "content": "DAPS with discrete diffusion. DAPS can be naturally extended to support posterior sampling with categorical prior distribution modeled by discrete diffusion models [3, 6, 33]. Instead of sampling \\(\\mathbf{x}_{0|\\mathbf{y}}\\) by Langevin MC or Hamiltonian MC in continuous space, we use the Metropolis-Hasting algorithm to replace Eq. (4). We validate DAPS with discrete diffusion on a quantized MNIST dataset. We leave the details to Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.733, + 0.789 + ], + "angle": 0, + "content": "4.3. Further results on MRI" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.908, + 0.902 + ], + "angle": 0, + "content": "We further validate the effectiveness of DAPS on compressed sensing multi-coil MRI, which is a medical imaging technique that helps reduce the scan time of MRI via subsampling. We train a diffusion model using the pipeline from [26] with fastMRI knee dataset [57]. Tab. 2 shows that DAPS outperforms existing diffusion-based baselines. We include more qualitative results in Appendix H." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20900" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.143, + 0.088, + 0.856, + 0.803 + ], + "angle": 0, + "content": "
TaskTypeMethodFFHQImageNet
PSNR (↑)SSIM (↑)LPIPS (↓)FID (↓)PSNR (↑)SSIM (↑)LPIPS (↓)FID (↓)
Super resolution 4xPixelDAPS (ours)29.070.8180.17751.4425.890.6940.27683.57
DPS25.860.7530.26981.0721.130.4890.361106.32
DDRM26.580.7820.28279.2522.620.5210.324103.85
DDNM28.030.7950.19764.6223.960.6040.47598.62
DCDP28.660.8070.17853.81----
FPS-SMC28.420.8130.20449.2524.820.7030.31397.51
DiffPIR26.64-0.26065.7723.18-0.371106.32
LatentLatentDAPS (ours)27.480.8010.18259.6225.060.6730.27684.37
PSLD24.350.6490.28774.3625.420.6940.36097.45
ReSample23.290.5940.39293.1822.610.5760.370113.42
Inpaint (box)PixelDAPS (ours)24.070.8140.13343.1021.430.7250.214109.85
DPS22.510.7920.20961.2718.940.7220.257126.52
DDRM22.260.8010.20778.6218.630.7330.254116.37
DDNM24.470.8370.23546.5921.640.7480.319103.97
DCDP23.890.7600.16345.23----
FPS-SMC24.860.8230.14648.3422.160.7260.208111.58
LatentLatentDAPS (ours)23.990.8020.19446.5217.190.6240.340145.63
PSLD24.220.8130.15843.0220.100.6940.465146.53
ReSample20.060.7490.18453.2118.290.6310.262127.84
Inpaint (random)PixelDAPS (ours)31.120.8440.09832.1728.440.7750.13554.25
DPS25.460.8230.20369.2023.520.7450.29787.53
DDNM29.910.8170.12144.3731.160.8410.19163.84
DCDP30.690.8420.14252.51----
FPS-SMC28.210.8230.26161.2324.520.7010.31679.12
LatentLatentDAPS (ours)30.710.8130.14136.4127.590.7720.16461.62
PSLD30.310.8090.22147.2131.300.7830.33783.21
ReSample29.610.7460.14039.8527.500.7560.14359.87
Gaussian deblurringPixelDAPS (ours)29.190.8170.16553.3326.150.6840.25375.68
DPS25.870.7640.21979.7520.310.5980.397116.42
DDRM24.930.7320.23992.4321.260.5640.443146.89
DDNM28.200.8040.21657.8328.060.7030.27881.43
DCDP27.500.6990.30486.43----
FPS-SMC26.540.7730.25367.4523.910.6010.38791.72
DiffPIR27.36-0.23659.6522.80-0.35593.36
LatentLatentDAPS (ours)27.930.7640.23464.5225.050.6680.34578.51
PSLD23.270.6310.31689.5125.860.6880.39091.39
ReSample26.390.7140.25571.6925.970.7030.25465.35
Motion deblurringPixelDAPS (ours)29.660.8470.15739.4927.860.7660.19661.83
DPS24.520.8010.24665.2318.960.6290.423137.81
DCDP25.080.5120.364125.13----
FPS-SMC27.390.8260.22748.3224.520.6470.32687.43
DiffPIR26.57-0.25565.7824.01-0.36694.63
LatentLatentDAPS (ours)27.000.8140.28346.8426.830.7450.29673.62
PSLD22.310.6780.33696.1520.850.5940.511124.67
ReSample27.410.8230.19844.7226.940.7380.22766.89
Phase retrievalPixelDAPS (ours)30.63±3.130.851±0.0720.139±0.06042.7125.78±6.920.743±0.0840.254±0.12582.67
DPS17.64±2.970.441±0.1290.410±0.090104.5216.81±3.610.427±0.1430.447±0.099197.54
RED-diff15.60±4.480.398±0.1950.596±0.092167.4314.98±3.750.386±0.0570.536±0.129212.24
DCDP28.65±8.090.781±0.2170.203±0.19668.13----
LatentLatentDAPS (ours)29.16±3.550.796±0.0890.199±0.07854.2620.54±6.410.612±0.1140.361±0.150129.54
ReSample21.60±8.100.648±0.1540.406±0.22484.32----
ClassicalHIO13.53±2.500.359±0.0930.726±0.068268.09----
Nonlinear deblurPixelDAPS (ours)28.29±1.770.783±0.0360.155±0.03249.3827.73±3.230.724±0.0480.169±0.05659.87
DPS23.39±2.010.623±0.0820.278±0.06091.3122.49±3.200.591±0.1010.306±0.081101.41
RED-diff30.86±0.510.795±0.0280.160±0.03443.8430.07±1.410.754±0.0230.211±0.08351.22
DCDP27.92±2.640.779±0.0670.183±0.05151.96----
LatentLatentDAPS (ours)28.11±1.750.713±0.0410.235±0.04953.6325.34±3.440.615±0.0570.314±0.08076.73
ReSample28.24±1.690.742±0.0390.185±0.03951.6226.20±3.710.653±0.0640.206±0.05761.16
High dynamic rangePixelDAPS (ours)27.12±3.530.752±0.0410.162±0.07242.9726.30±4.100.717±0.0670.175±0.10764.19
DPS22.73±6.070.591±0.1410.264±0.156112.8219.23±2.520.582±0.0820.503±0.106146.23
RED-diff22.16±3.410.512±0.0830.258±0.089108.3222.03±5.900.601±0.0940.274±0.198113.48
LatentLatentDAPS (ours)25.94±2.870.751±0.0560.223±0.08074.8323.64±4.100.609±0.0530.269±0.09993.51
ReSample25.65±3.570.732±0.0590.182±0.08567.2225.11±4.210.633±0.0490.198±0.08987.66
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.812, + 0.907, + 0.898 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation on FFHQ and ImageNet on 5 linear and 3 nonlinear tasks. Performance comparison of different methods on FFHQ (left) and ImageNet (right) in the image domain. For linear tasks, we report the mean performance (PSNR, SSIM and LPIPS) across 100 validation images. We provide both the mean and standard deviation for nonlinear tasks, where results exhibit higher instability. The FID scores are evaluated with the same 100 validation images. The best and second-best results within each type of task are indicated by bold and underlined marks, respectively. DAPS demonstrates superior performance on most tasks, with LatentDAPS showing competitive results. All tasks are using noisy measurement with noise level \\(\\beta_{\\mathbf{y}} = 0.05\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20901" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.089, + 0.495, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.234, + 0.366, + 0.247 + ], + "angle": 0, + "content": "(a) Gaussian deblur" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.089, + 0.865, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.234, + 0.731, + 0.246 + ], + "angle": 0, + "content": "(b) Phase retrieval" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.259, + 0.908, + 0.288 + ], + "angle": 0, + "content": "Figure 6. Evaluation of time cost and sample quality. The x-axis indicates the single image sampling time on an A100-SXM4-80GB GPU and y-axis shows the LPIPS. The evaluation uses 100 FFHQ images." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.298, + 0.907, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.446, + 0.908, + 0.489 + ], + "angle": 0, + "content": "Figure 7. Quantitative evaluations of image quality metrics and measurement error throughout the sampling process. Here \\(\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)\\) stands for the estimated mean of the Gaussian approximated distribution and \\(\\mathbf{x}_{0|\\mathbf{y}}\\sim p(\\mathbf{x}_0\\mid \\mathbf{x}_t,\\mathbf{y})\\) means the samples from measurement conditioned time-marginal. The measurement errors are normalized between [0, 1] for visualization." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.499, + 0.484, + 0.571 + ], + "angle": 0, + "content": "
×4×8
PSNR (↑)SSIM (↑)Meas err (↓)PSNR (↑)SSIM (↑)Meas err (↓)
DAPS31.480.7621.5729.010.6811.28
DPS26.130.6209.9020.820.5366.74
DiffPIR28.310.63210.5526.780.5887.79
ScoreMRI [10]25.970.46810.8325.010.4058.36
CSGM [23]28.780.7101.5226.150.6251.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.624 + ], + "angle": 0, + "content": "Table 2. Quantitative results of compressed sensing MRI. We consider \\( \\times 4 \\) and \\( \\times 8 \\) subsampling ratio. Meas err stands for \\( {\\ell }_{2} \\) data consistency error." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.255, + 0.652 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Time efficiency of DAPS. We run DAPS with various configurations to test its performance under different computing budgets. Specifically, we evaluate DAPS with the number of function evaluations (NFE) of the diffusion models ranging from 50 to 4K, with configurations specified in Appendix E.1. Sec. 4.2 shows the relation between the sample quality of DAPS measured by LPIPS with its time cost, which indicates that DAPS is able to generate high-quality samples with a reasonable time cost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of Different Design choices of DAPS. We conduct an ablation study to demonstrate the effectiveness of Gaussian approximation compared with the more principled but more expensive diffusion-score estimation. The results in Tab. 3 demonstrate that Gaussian approximation is \\(6 \\sim 7\\) times faster than diffusion-score estimation with the cost of a slight performance drop, effectively trading off performance" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.499, + 0.907, + 0.551 + ], + "angle": 0, + "content": "
Gaussian deblurringHigh dynamic range
PSNR (↑)LPIPS (↓)Run time/sPSNR (↑)LPIPS (↓)Run time/s
Gaussian approx.29.400.17064.724.850.18161.4
diffusion-score29.570.160462.325.330.193540.0
Relative improvement-0.575%-6.25%614.53%-1.89%-6.22%779.48%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.561, + 0.908, + 0.618 + ], + "angle": 0, + "content": "Table 3. Comparison of approaches to approximate \\( p(\\mathbf{x}_0 \\mid \\mathbf{x}_t) \\). Three tasks on 10 FFHQ images with DAPS-1K consistently shown Gaussian approximation achieve huge speed-up while slightly sacrificing the performance compared with diffusion-score estimation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.633, + 0.61, + 0.648 + ], + "angle": 0, + "content": "for efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.907, + 0.709 + ], + "angle": 0, + "content": "More ablation studies. We include more ablation studies of DAPS in Appendix I.1, including the number of function evaluations, the number of ODE steps, the annealing noise scheduling, and different measurement noise levels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.634, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In summary, we propose Decoupled Annealing Posterior Sampling (DAPS) for solving inverse problems, particularly those with complex nonlinear measurement processes such as phase retrieval. Our method decouples consecutive sample points in a diffusion sampling trajectory, allowing them to vary considerably, thereby enabling DAPS to explore a larger solution space. Empirically, we demonstrate that DAPS generates samples with better visual quality and stability compared to existing methods when solving a wide range of challenging inverse problems." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20902" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.222 + ], + "angle": 0, + "content": "We are grateful to Pika for providing the computing resources essential for this research. We also extend our thanks to the Kortschak Scholars Fellowship for supporting B.Z. and W.C. at Caltech. J.B. acknowledges support from the Wally Baer and Jeri Weiss Postdoctoral Fellowship. A.A. is supported in part by Bren endowed chair and by the AI2050 senior fellow program at Schmidt Sciences." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.236, + 0.188, + 0.251 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.26, + 0.484, + 0.344 + ], + "angle": 0, + "content": "[1] Kazunori Akiyama, Antxon Alberdi, Walter Alef, Keiichi Asada, Rebecca Azulay, Anne-Kathrin Baczko, David Ball, Mislav Balokovic, John Barrett, Dan Bintley, et al. First m87 event horizon telescope results. iv. imaging the central supermassive black hole. The Astrophysical Journal Letters, 875(1):L4, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[2] Marius Arvinte, Sriram Vishwanath, Ahmed H. Tewfik, and Jonathan I. Tamir. Deep j-sense: Accelerated mri reconstruction via unrolled alternating optimization, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.484, + 0.443 + ], + "angle": 0, + "content": "[3] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981-17993, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[4] Michael Betancourt and Mark Girolami. Hamiltonian monte carlo for hierarchical models. Current trends in Bayesian methodology with applications, 79(30):2-4, 2015. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[5] Benjamin Boys, Mark Girolami, Jakiw Pidstrigach, Sebastian Reich, Alan Mosca, and O. Deniz Akyildiz. Tweedie moment projected diffusions for inverse problems, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.485, + 0.599 + ], + "angle": 0, + "content": "[6] Andrew Campbell, Joe Benton, Valentin De Bortoli, Thomas Rainforth, George Deligiannidis, and Arnaud Doucet. A continuous time framework for discrete denoising models. Advances in Neural Information Processing Systems, 35:28266-28279, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[7] Gabriel Cardoso, Yazid Janati El Idrissi, Sylvain Le Corff, and Eric Moulines. Monte carlo guided diffusion for bayesian linear inverse problems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[8] Gabriel Victorino Cardoso, Yazid Janati, Sylvain Le Corff, and Éric Moulines. Monte carlo guided diffusion for bayesian linear inverse problems. ArXiv, abs/2308.07983, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[9] Junqing Chen and Haibo Liu. An alternating direction method of multipliers for inverse lithography problem, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[10] Hyungjin Chung and Jong Chul Ye. Score-based diffusion models for accelerated mri. Medical Image Analysis, page 102479, 2022. 1, 3, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[11] Hyungjin Chung, Eun Sun Lee, and Jong Chul Ye. Mr image denoising and super-resolution using regularized reverse diffusion. IEEE Transactions on Medical Imaging, 42(4): 922-934, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[12] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. In Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.26, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 3, 4, 5, 6, 10, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[14] J. Dainty and James Fienup. Phase retrieval and image reconstruction for astronomy. Image Recovery: Theory Appl, 13, 1987. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.234 + ], + "angle": 0, + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.906, + 0.277 + ], + "angle": 0, + "content": "[16] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In Advances in Neural Information Processing Systems, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.907, + 0.333 + ], + "angle": 0, + "content": "[17] Zolnamar Dorjsembe, Hsing-Kuo Pao, Sodtavilan Odonchimed, and Furen Xiao. Conditional diffusion models for semantic 3d brain mri synthesis. IEEE Journal of Biomedical and Health Informatics, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.336, + 0.907, + 0.39 + ], + "angle": 0, + "content": "[18] Zehao Dou and Yang Song. Diffusion posterior sampling for linear inverse problem solving: A filtering perspective. In The Twelfth International Conference on Learning Representations, 2024. 3, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.393, + 0.907, + 0.474 + ], + "angle": 0, + "content": "[19] Berthy T. Feng, Jamie Smith, Michael Rubinstein, Huiwen Chang, Katherine L. Bouman, and William T. Freeman. Score-based diffusion models as principled priors for inverse imaging. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10520-10531, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.907, + 0.531 + ], + "angle": 0, + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.907, + 0.588 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.907, + 0.646 + ], + "angle": 0, + "content": "[22] Alex Ling Yu Hung, Kai Zhao, Haoxin Zheng, Ran Yan, Steven S Raman, Demetri Terzopoulos, and Kyunghyun Sung. Med-cdiff: Conditional medical image generation with diffusion models. Bioengineering, 10(11):1258, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.907, + 0.716 + ], + "angle": 0, + "content": "[23] Ajil Jalal, Marius Arvinte, Giannis Daras, Eric Price, Alexandros G Dimakis, and Jon Tamir. Robust compressed sensing mri with deep generative priors. In Advances in Neural Information Processing Systems, pages 14938-14954. Curran Associates, Inc., 2021. 1, 2, 3, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.907, + 0.787 + ], + "angle": 0, + "content": "[24] Ulugbek Kamilov, Charles Bouman, Gregory Buzzard, and Brendt Wohlberg. Plug-and-play methods for integrating physical and learned models in computational imaging: Theory, algorithms, and applications. IEEE Signal Processing Magazine, 40:85-97, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.843 + ], + "angle": 0, + "content": "[25] T. Karras, S. Laine, and T. Aila. A style-based generator architecture for generative adversarial networks. IEEE Transactions on Pattern Analysis & Machine Intelligence, 43(12): 4217-4228, 2021. 5, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[26] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, 2022. 2, 5, 6, 10, 13" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20903" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.134 + ], + "angle": 0, + "content": "[27] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[28] Sergey Kastryulin, Jamil Zakirov, Denis Prokopenko, and Dmitry V. Dylov. Pytorch image quality: Metrics for image quality assessment, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[29] Bahjat Kawar, Gregory Vaksman, and Michael Elad. SNIPS: Solving noisy inverse problems stochastically. Advances in Neural Information Processing Systems, 34:21757-21769, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[30] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.483, + 0.329 + ], + "angle": 0, + "content": "[31] Xiang Li, Soo Min Kwon, Ismail R. Alkhouri, Saiprasad Ravishankar, and Qing Qu. Decoupled data consistency with diffusion purification for image restoration, 2024. 3, 5, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[32] Hanyuan Liu, Jinbo Xing, Minshan Xie, Chengze Li, and Tien-Tsin Wong. Improved diffusion-based image colorization via piggybacked models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[33] Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion language modeling by estimating the ratios of the data distribution. arXiv preprint arXiv:2310.16834, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.483, + 0.51 + ], + "angle": 0, + "content": "[35] Morteza Mardani, Jiaming Song, Jan Kautz, and Arash Vahdat. A variational perspective on solving inverse problems with diffusion models. arXiv preprint arXiv:2305.04391, 2023. 6, 11, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[36] Xiangming Meng and Yoshiyuki Kabashima. Diffusion model based posterior sampling for noisy linear inverse problems. arXiv preprint arXiv:2211.12343, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.609 + ], + "angle": 0, + "content": "[37] Nicholas Metropolis, Arianna W Rosenbluth, Marshall N Rosenbluth, Augusta H Teller, and Edward Teller. Equation of state calculations by fast computing machines. The journal of chemical physics, 21(6):1087-1092, 1953. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.61, + 0.483, + 0.664 + ], + "angle": 0, + "content": "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.483, + 0.706 + ], + "angle": 0, + "content": "[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 4, 5, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[41] Litu Rout, Yujia Chen, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Beyond first-order tweedie: Solving inverse problems using latent diffusion, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[42] Litu Rout, Negin Raoof, Giannis Daras, Constantine Caramanis, Alex Dimakis, and Sanjay Shakkottai. Solving linear inverse problems provably via posterior sampling with latent diffusion models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4, 6, 9, 11" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.175 + ], + "angle": 0, + "content": "[43] Litu Rout, Yujia Chen, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Beyond first-order tweedie: Solving inverse problems using latent diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9472-9481, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.908, + 0.245 + ], + "angle": 0, + "content": "[44] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.908, + 0.316 + ], + "angle": 0, + "content": "[45] Bowen Song, Soo Min Kwon, Zecheng Zhang, Xinyu Hu, Qing Qu, and Liyue Shen. Solving inverse problems with latent diffusion models via hard data consistency. In The Twelfth International Conference on Learning Representations, 2024, 2, 3, 4, 5, 6, 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.318, + 0.907, + 0.373 + ], + "angle": 0, + "content": "[46] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.907, + 0.416 + ], + "angle": 0, + "content": "[47] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Neural Information Processing Systems, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.907, + 0.46 + ], + "angle": 0, + "content": "[48] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.907, + 0.53 + ], + "angle": 0, + "content": "[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 1, 2, 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.907, + 0.587 + ], + "angle": 0, + "content": "[50] Yang Song, Liyue Shen, Lei Xing, and Stefano Ermon. Solving inverse problems in medical imaging with score-based generative models. In International Conference on Learning Representations, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.907, + 0.631 + ], + "angle": 0, + "content": "[51] He Sun and Katherine L. Bouman. Deep probabilistic imaging: Uncertainty quantification and multi-modal solution characterization for computational imaging, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[52] Yu Sun, Zihui Wu, Yifan Chen, Berthy Feng, and Katherine L. Bouman. Provable probabilistic imaging using score-based generative priors. *ArXiv*, abs/2310.10835, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[53] Phong Tran, Anh Tran, Quynh Phung, and Minh Hoai. Explore image deblurring via encoded blur kernel space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.907, + 0.787 + ], + "angle": 0, + "content": "[54] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. The Eleventh International Conference on Learning Representations, 2023. 3, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.844 + ], + "angle": 0, + "content": "[55] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688. CiteSeer, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[56] Zihui Wu, Yu Sun, Yifan Chen, Bingliang Zhang, Yisong Yue, and Katherine L Bouman. Principled probabilistic imaging using diffusion models as plug-and-play priors. arXiv preprint arXiv:2405.18782, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20904" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.217 + ], + "angle": 0, + "content": "[57] Jure Zbontar, Florian Knoll, Anuroop Sriram, Matthew J. Muckley, Mary Bruno, Aaron Defazio, Marc Parente, Krzysztof J. Geras, Joe Katsnelson, Hersh Chandarana, Zizhao Zhang, Michal Drozdal, Adriana Romero, Michael G. Rabbat, Pascal Vincent, James Pinkerton, Duo Wang, Nafissa Yakubova, Erich Owens, C. Lawrence Zitnick, Michael P. Recht, Daniel K. Sodickson, and Yvonne W. Lui. fastmri: An open dataset and benchmarks for accelerated MRI. CoRR, abs/1811.08839, 2018. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.485, + 0.288 + ], + "angle": 0, + "content": "[58] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.485, + 0.358 + ], + "angle": 0, + "content": "[59] Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising diffusion models for plug-and-play image restoration. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (NTIRE), 2023. 1, 3, 5, 6, 11, 13" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "20905" + } + ] +] \ No newline at end of file diff --git a/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/d107e339-bf08-4731-b903-e7f3360d585d_origin.pdf b/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/d107e339-bf08-4731-b903-e7f3360d585d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..40e3305b40a7ca3fd4c072659a30092521432e17 --- /dev/null +++ b/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/d107e339-bf08-4731-b903-e7f3360d585d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c88d865cb398d62f9823a8947d8f68fd91eeddcbcba725543b9544161084ed64 +size 9039559 diff --git a/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/full.md b/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0471687ff5aed6f5a7f61c2e95311cc8d199690a --- /dev/null +++ b/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/full.md @@ -0,0 +1,317 @@ +# Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing + +Bingliang Zhang\*,1 + +Chenlin Meng3 + +$^{1}$ California Institute of Technology + +Wenda Chu\*,1 + +Anima Anandkumar + +2NVIDIA + +Julius Berner $^{2,\dagger}$ + +Yang Song + +3Stanford University 4OpenAI + +![](images/9abd721ce8744c5010bf45cc2c0bd44c288c8942c9b1da06449f29c171ef9ab3.jpg) + +![](images/c79e7bea559a0bccdf463a491c29f43836ecb0933685c487199626b1088828ad.jpg) + +![](images/8501a45e9b3699b1069f727d48f5c7690ff3906ce2dd01d463a7e6ef5169440c.jpg) +Figure 1. Overview of Decoupled Annealing Posterior Sampling (DAPS). Our method provides a flexible and effective framework for solving inverse problems through a decoupled posterior sampling process. In (a)(b), we present DAPS visual results on FFHQ and ImageNet at a resolution of 256, and in (c), on natural images at a resolution of 768. In (d), we display DAPS results on compressed sensing multi-coil MRI (CS-MRI). DAPS effectively addresses nonlinear inverse problems as well as medical imaging MRI challenges. Additionally, DAPS can be enhanced using large-scale latent diffusion models (LDMs) [40], as shown in (c). + +# Abstract + +Diffusion models have recently achieved success in solving Bayesian inverse problems with learned data priors. Current methods build on top of the diffusion sampling process, where each denoising step makes small modifications to samples from the previous step. However, this process struggles to correct errors from earlier sampling steps, leading to worse performance in complicated nonlinear inverse problems, such as phase retrieval. To address this challenge, we propose a new method called Decoupled Annealing Posterior Sampling (DAPS) that relies on a novel noise annealing process. Specifically, we decouple consecutive steps in a diffusion sampling trajectory, allowing them to vary consid + +erably from one another while ensuring their time-margins anneal to the true posterior as we reduce noise levels. This approach enables the exploration of a larger solution space, improving the success rate for accurate reconstructions. We demonstrate that DAPS significantly improves sample quality and stability across multiple image restoration tasks, particularly in complicated nonlinear inverse problems. + +# 1. Introduction + +Inverse problems are ubiquitous in science and engineering, with applications ranging from image restoration [13, 30, 34, 44, 49, 59], medical imaging [10, 11, 17, 22, 23, 50] to astrophotography [1, 19, 51, 52]. Solving an inverse problem involves finding the underlying signal $\mathbf{x}_0$ from its partial, noisy measurement $\mathbf{y}$ . Since the measurement process is typically noisy and many-to-one, inverse problems do not have a + +unique solution; instead, multiple solutions may exist that are consistent with the observed measurement. In the Bayesian inverse problem framework, the solution space is characterized by the posterior distribution $p(\mathbf{x}_0 \mid \mathbf{y}) \propto p(\mathbf{y} \mid \mathbf{x}_0) p(\mathbf{x}_0)$ , where $p(\mathbf{y} \mid \mathbf{x}_0)$ represents the noisy measurement process, and $p(\mathbf{x}_0)$ is the prior distribution. In this work, we aim to solve Bayesian inverse problems where the measurement process $p(\mathbf{y} \mid \mathbf{x}_0)$ is known, and the prior distribution $p(\mathbf{x}_0)$ is captured by a deep generative model trained on a corresponding dataset. + +As score-based diffusion models [21, 26, 27, 47-49] have risen to dominance in modeling high-dimensional data distributions like images, audio, and video, they have become the leading method for estimating the prior distribution $p(\mathbf{x}_0)$ in Bayesian inverse problems. A diffusion model generates a sample $\mathbf{x}_0$ by smoothly removing noise from an unstructured initial noise sample $\mathbf{x}_T$ through solving stochastic differential equations (SDEs). In particular, each step of the sampling process recursively converts a noisy sample $\mathbf{x}_{t + \Delta t}$ , where $\Delta t > 0$ denotes the step size, to a slightly less noisy sample $\mathbf{x}_t$ until $t = 0$ . This iterative structure in the diffusion sampling process can be leveraged to facilitate Bayesian inverse problem solving. In fact, prior research [5, 13, 46] has shown that given measurement $\mathbf{y}$ and a diffusion model prior $p(\mathbf{x}_0)$ , we can sample from the posterior distribution $p(\mathbf{x}_0\mid \mathbf{y})$ by perturbing a reverse-time SDE using an approximate gradient of the measurement process, $\nabla_{\mathbf{x}_t}\log p(\mathbf{y}\mid \mathbf{x}_t)$ , at every step of the SDE solver. + +Despite the remarkable success of these methods in solving many real-world inverse problems, like image colorization [30, 32, 49], image super-resolution [13, 30, 46, 49], computed tomography [45, 50], and magnetic resonance imaging [23, 50], they face significant challenges in more complex inverse problems with nonlinear measurement processes, such as phase retrieval and nonlinear motion deblurring. This is partially because, in prior methods, each denoising step approximately samples from the distribution $p(\mathbf{x}_t \mid \mathbf{x}_{t + \Delta t}, \mathbf{y})$ . This causes $\mathbf{x}_t$ and $\mathbf{x}_{t + \Delta t}$ to be close to each other because of using a small step size $\Delta t$ in discretizing the reverse-time SDE. As a result, $\mathbf{x}_t$ can at most correct local errors in $\mathbf{x}_{t + \Delta t}$ but struggles to correct global errors that require significant modifications to the prior sample. This issue is exacerbated when the methods are applied to complicated, nonlinear inverse problems, such as phase retrieval, where they often converge to undesired samples that are consistent with the measurement but reside in low-probability areas of the prior distribution. + +To address this challenge, we propose a new framework for solving general inverse problems, termed Decoupled Annealing Posterior Sampling (DAPS). Our method employs a new noise annealing process inspired by the diffusion sampling process, where we decouple the consecutive samples $\mathbf{x}_{t + \Delta t}$ and $\mathbf{x}_t$ in the sampling trajectory, allowing samplers + +to correct large, global errors made in earlier steps. Instead of repetitively sampling from $p(\mathbf{x}_t \mid \mathbf{x}_{t + \Delta t}, \mathbf{y})$ as in previous methods, which restricts the distances between consecutive samples $\mathbf{x}_{t + \Delta t}$ and $\mathbf{x}_t$ , DAPS recursively samples from the marginal distribution $p(\mathbf{x}_t \mid \mathbf{y})$ . As illustrated in Fig. 2, we factorize the time marginal $p(\mathbf{x}_t \mid \mathbf{y})$ into three conditional distributions and sample from them in turn by solving the reverse diffusion process, simulating Langevin dynamics, and adding noise according to the forward diffusion process. We show that this creates approximate samples from corresponding marginal distributions. As the noise gradually anneals to zero, the time marginal $p(\mathbf{x}_t \mid \mathbf{y})$ smoothly converges to the posterior distribution $p(\mathbf{x}_0 \mid \mathbf{y})$ , providing samples that approximately solve the Bayesian inverse problem. + +Empirically, our method demonstrates significantly improved performance on various inverse problems compared to existing approaches, ranging from image restoration to compressed sensing MRI. Our method can be combined with diffusion models in both raw pixel space and learned latent space, which we refer to as DAPS and LatentDAPS, respectively. As shown in Fig. 1, both methods provide superior reconstructions with improved visual perceptual quality across a wide range of nonlinear inverse problems. Our approach exhibits remarkable stability and sampling quality, particularly for challenging nonlinear inverse problems. On the FFHQ 256 dataset, we achieve PSNR values of $30.72\mathrm{dB}$ and $27.12\mathrm{dB}$ for the phase retrieval and high dynamic range tasks, representing improvements of $1.98\mathrm{dB}$ and $4.39\mathrm{dB}$ over the previous state-of-the-art results, respectively. We apply DAPS to large-scale text-conditioned latent diffusion models, showing its scalability to high-resolution image restoration tasks. Beyond image restoration tasks, DAPS achieves a PSNR of $31.49\mathrm{dB}$ , showing a $2.70\mathrm{dB}$ improvement over previous diffusion-based methods. We also demonstrate how DAPS can be applied to inverse problems on categorical data with pre-trained discrete diffusion models. Additionally, DAPS performs well with a small number of neural network evaluations (approximately 100), striking a better balance between efficiency and sample quality. + +# 2. Background + +# 2.1. Diffusion Models + +Diffusion models [21, 26, 47-49] generate data by reversing a predefined noising process. Let the data distribution be $p(\mathbf{x}_0)$ . We can define a series of noisy data distributions $p(\mathbf{x};\sigma)$ by adding Gaussian noise with a standard deviation of $\sigma$ to the data. These form the time-margins of a stochastic differential equation (SDE) [26], given by $\mathrm{d}\mathbf{x}_t = \sqrt{2\dot{\sigma}_t\sigma_t}\mathrm{d}\mathbf{w}_t$ , where $\sigma_{t}$ is the predefined noise schedule with $\sigma_0 = 0$ and $\sigma_T = \sigma_{\mathrm{max}}$ , $\dot{\sigma}_{t}$ is the time derivative of $\sigma_{t}$ , and $\mathbf{w}_t$ is a standard Wiener process. We use $\mathbf{x}_t$ interchangeably with $\mathbf{x}_{\sigma_t}$ . For a sufficiently large $\sigma_{\mathrm{max}}$ , the + +![](images/74776871326b0fdad20d79ec6020c5589d7bf42dac02bfa177d58618b1d8eff0.jpg) +Figure 2. An illustration of our method on phase retrieval. Given a noisy sample $\mathbf{x}_t$ , we first solve the reverse diffusion process to obtain $\hat{\mathbf{x}}_0(\mathbf{x}_t)$ . Using this, we construct an approximation for $p(\mathbf{x}_0 \mid \mathbf{x}_t, \mathbf{y})$ . Next, we sample $\mathbf{x}_{0|y} \sim p(\mathbf{x}_0 \mid \mathbf{x}_t, \mathbf{y})$ with multiple steps of Langevin dynamics, then perturb it with the forward diffusion process to obtain a sample with slightly less noise. We repeat this process until the noise reduces to zero. + +distribution $p(\mathbf{x};\sigma_{\mathrm{max}})$ converges to pure Gaussian noise $\mathcal{N}(\mathbf{0},\sigma_{\mathrm{max}}^2\mathbf{I})$ + +To sample from data distribution $p(\mathbf{x}_0)$ , we first draw an initial sample from $\mathcal{N}(\mathbf{0},\sigma_{\mathrm{max}}^2\mathbf{I})$ , then solve the reverse SDE: $\mathrm{d}\mathbf{x}_t = -2\dot{\sigma}_t\sigma_t\nabla_{\mathbf{x}_t}\log p(\mathbf{x}_t;\sigma_t)\mathrm{d}t + \sqrt{2\dot{\sigma}_t\sigma_t}\mathrm{d}\mathbf{w}_t$ where $\nabla_{\mathbf{x}_t}\log p(\mathbf{x}_t;\sigma_t)$ is the time-dependent score function [47, 49]. Here $\nabla_{\mathbf{x}_t}\log p(\mathbf{x}_t;\sigma_t)$ is unknown, but can be approximated by training a diffusion model $s_\theta (\mathbf{x}_t,\sigma_t)$ such that $s_\theta (\mathbf{x}_t,\sigma_t)\approx \nabla_{\mathbf{x}_t}\log p(\mathbf{x}_t;\sigma_t)$ . + +# 2.2. Bayesian Inverse Problems with Diffusion + +Inverse problems aim to recover data from partial, potentially noisy measurements. Formally, solving an inverse problem involves finding the inversion to a forward model that describes the measurement process. In general, a forward model takes the form of $\mathbf{y} = \mathcal{A}(\mathbf{x}_0) + \mathbf{n}$ , where $\mathcal{A}$ is the measurement function, $\mathbf{x}_0$ represents the original data, $\mathbf{y}$ is the observed measurement, and $\mathbf{n}$ symbolizes the noise in the measurement process, often modeled as $\mathbf{n} \sim \mathcal{N}(\mathbf{0}, \beta_{\mathbf{y}}^2 \mathbf{I})$ . In a Bayesian framework, $\mathbf{x}_0$ comes from the posterior distribution $p(\mathbf{x}_0 \mid \mathbf{y}) \propto p(\mathbf{x}_0) p(\mathbf{y} \mid \mathbf{x}_0)$ . Here $p(\mathbf{x}_0)$ is a prior distribution that can be estimated from a given dataset, and $p(\mathbf{y} \mid \mathbf{x}_0) = \mathcal{N}(\mathcal{A}(\mathbf{x}_0), \beta_{\mathbf{y}}^2 \mathbf{I})$ models the noisy measurement process. + +When the prior $p(\mathbf{x}_0)$ is modeled by a pre-trained diffusion model, we can modify reverse SDE to approximately sample from the posterior distribution following Bayes' rule, + +$$ +\begin{array}{l} \mathrm {d} \mathbf {x} _ {t} = - 2 \dot {\sigma} _ {t} \sigma_ {t} \nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t}; \sigma_ {t}) \mathrm {d} t \\ - 2 \dot {\sigma} _ {t} \sigma_ {t} \nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {y} \mid \mathbf {x} _ {t}) d t + \sqrt {2 \dot {\sigma} _ {t} \sigma_ {t}} d w _ {t}. \tag {1} \\ \end{array} +$$ + +Here, the noisy likelihood $\nabla_{\mathbf{x}_t}\log p(\mathbf{y}|\mathbf{x}_t)$ is generally intractable. Multiple methods have been proposed to estimate the noisy likelihood [5, 8, 12, 36, 41, 46, 50, 54, 59]. + +![](images/6cb7dc6f884d697e6f9679598e667577db9a753d0960fd74bde467fab905f146.jpg) + +![](images/3ffbaebd72a20ddd23ee5eb974a9822a9e2197f5d8c966a6cae3f669c1e6ff12.jpg) +(a) Diffusion models in pixel spaces. +(b) Latent diffusion models. +Figure 3. Probabilistic graphical models of the decoupled noise annealing process in (a) pixel and (b) latent spaces respectively. Here $\mathbf{x}_t \sim p(\mathbf{x}_t; \sigma_t)$ , and $\mathbf{y}$ is the observed measurement. We factorize $p(\mathbf{x}_t \mid \mathbf{y})$ based on these probabilistic graphs. + +One predominant approach is the DPS algorithm [13], which estimates $p(\mathbf{y} \mid \mathbf{x}_t) \approx p(\mathbf{y} \mid \mathbf{x}_0 = \mathbb{E}[\mathbf{x}_0 \mid \mathbf{x}_t])$ . Another line of work [7, 18, 29, 30, 54] solves linear inverse problems by running the reverse diffusion in the spectral domain via singular value decomposition (SVD). Other methods bypass direct computation of this likelihood by interleaving optimization [2, 9, 24, 31, 45, 56, 59] or projection [10, 23, 30, 54] steps with normal diffusion sampling steps. + +Despite promising empirical success, we find that this line of approaches faces challenges in solving more difficult inverse problems when the forward model is highly nonlinear. Accurately solving the reverse SDE in Eq. (1) requires the solver to take a very small step size $\Delta t > 0$ , causing $\mathbf{x}_t$ and $\mathbf{x}_{t + \Delta t}$ to be very close to each other. Consequently, $\mathbf{x}_t$ can only correct minor errors in $\mathbf{x}_{t + \Delta t}$ , but oftentimes fails to address larger, global errors that require substantial changes to $\mathbf{x}_{t + \Delta t}$ . One such failure case is given in Fig. 4. + +# 3. Method + +# 3.1. Decoupled Annealing Posterior Sampling + +Instead of solving the reverse-time SDE in Eq. (1), we propose a new noise annealing process that reduces the dependency between samples at consecutive time steps, as illustrated in Figs. 2, 3a and 3b. Unlike previous methods, we ensure $\mathbf{x}_t$ and $\mathbf{x}_{t + \Delta t}$ are conditionally independent given $\mathbf{x}_0$ . To generate sample $\mathbf{x}_t$ from $\mathbf{x}_{t + \Delta t}$ , we follow a two-step procedure: (1) sampling $\mathbf{x}_{0|\mathbf{y}} \sim p(\mathbf{x}_0 \mid \mathbf{x}_{t + \Delta t}, \mathbf{y})$ , and (2) sampling $\mathbf{x}_t \sim \mathcal{N}(\mathbf{x}_{0|\mathbf{y}}, \sigma_t^2 \mathbf{I})$ . We repeat this process, gradually reducing noise until $\mathbf{x}_0$ is sampled. We call this process decoupled noise annealing, which is justified by the proposition below. + +Proposition 1. Suppose $\mathbf{x}_{t_1}$ is sampled from the time-marginal $p(\mathbf{x}_{t_1} \mid \mathbf{y})$ , then + +$$ +\mathbf {x} _ {t _ {2}} \sim \mathbb {E} _ {\mathbf {x} _ {0} \sim p (\mathbf {x} _ {0} | \mathbf {x} _ {t _ {1}}, \mathbf {y})} [ \mathcal {N} (\mathbf {x} _ {0}, \sigma_ {t _ {2}} ^ {2} I) ] \tag {2} +$$ + +satisfies the time-marginal $p(\mathbf{x}_{t_2} \mid \mathbf{y})$ . + +Remark. The key idea of Proposition 1 is to enable the sampling from $p(\mathbf{x}_{t_2} \mid \mathbf{y})$ for any noise level $\sigma_{t_2}$ given any sample $x_{t_1}$ at another noise level $\sigma_{t_1}$ . For a sufficiently large $\sigma_T$ , one can assume $p(\mathbf{x}_T \mid \mathbf{y}) \approx p(\mathbf{x}_T; \sigma_T) \approx \mathcal{N}(\mathbf{0}, \sigma_T^2\mathbf{I})$ . Starting from $\mathbf{x}_T$ , we can iteratively sample from $p(\mathbf{x}_t \mid \mathbf{y})$ with $\sigma_t$ annealed down from $\sigma_T$ to 0. + +The first step of our decoupled noise annealing requires sampling $\mathbf{x}_{0|\mathbf{y}}\sim p(\mathbf{x}_0\mid \mathbf{x}_t,\mathbf{y})$ where $\mathbf{x}_t$ and $\mathbf{y}$ are known. Since $\mathbf{y}$ is conditionally independent from $\mathbf{x}_t$ given $\mathbf{x}_0$ , we can deduce from the Bayes' rule that + +$$ +\begin{array}{l} p \left(\mathbf {x} _ {0} \mid \mathbf {x} _ {t}, \mathbf {y}\right) = \frac {p \left(\mathbf {x} _ {0} \mid \mathbf {x} _ {t}\right) p \left(\mathbf {y} \mid \mathbf {x} _ {0} , \mathbf {x} _ {t}\right)}{p \left(\mathbf {y} \mid \mathbf {x} _ {t}\right)} \\ \propto p \left(\mathbf {x} _ {0} \mid \mathbf {x} _ {t}\right) p (\mathbf {y} \mid \mathbf {x} _ {0}). \tag {3} \\ \end{array} +$$ + +To sample $\mathbf{x}_{0|\mathbf{y}}$ from this unnormalized distribution, MCMC methods such as Langevin dynamics [55] and Hamiltonian Monte Carlo [4] can be adopted (as explained in Appendix A.1). The updating rule of Langevin dynamics [55] is given by + +$$ +\begin{array}{l} \mathbf {x} _ {0} ^ {(j + 1)} = \mathbf {x} _ {0} ^ {(j)} + \eta \nabla_ {\mathbf {x} _ {0} ^ {(j)}} \log p (\mathbf {x} _ {0} ^ {(j)} \mid \mathbf {x} _ {t}) \\ + \eta \nabla_ {\mathbf {x} _ {0} ^ {(j)}} \log p (\mathbf {y} \mid \mathbf {x} _ {0} ^ {(j)}) + \sqrt {2 \eta} \boldsymbol {\epsilon} _ {j}, \tag {4} \\ \end{array} +$$ + +where $\eta > 0$ is the step size and $\epsilon_{j} \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . When $\eta \to 0$ and $j \to \infty$ , the sample $\mathbf{x}_0^{(j)}$ will be approximately distributed according to $p(\mathbf{x}_0 \mid \mathbf{x}_t, \mathbf{y})$ . + +# 3.2. Practical Design Choice of DAPS + +Following Eq. (4), we discuss different ways to approximate the conditional distribution $p(\mathbf{x}_0 \mid \mathbf{x}_t)$ due to its intractability. A line of previous works on diffusion posterior sampling [5, 13, 46] approximate $p(\mathbf{x}_0 \mid \mathbf{x}_t)$ by a Gaussian distribution, + +$$ +p \left(\mathbf {x} _ {0} \mid \mathbf {x} _ {t}\right) \approx \mathcal {N} \left(\mathbf {x} _ {0}; \hat {\mathbf {x}} _ {0} \left(\mathbf {x} _ {t}\right), r _ {t} ^ {2} \boldsymbol {I}\right), \tag {5} +$$ + +where $\hat{\mathbf{x}}_0(\mathbf{x}_t)$ is an estimator of $\mathbf{x}_0$ given $\mathbf{x}_t$ typically specified as $\mathbb{E}[\mathbf{x}_0\mid \mathbf{x}_t]$ , and the variance $r_t^2$ is specified using heuristics. + +Another possible way to estimate $p(\mathbf{x}_0 \mid \mathbf{x}_t)$ is to decompose it via Bayes' rule, and substitute the intractable $p(\mathbf{x}_0)$ by the distribution $p(\mathbf{x}_0; \sigma_{t_{\min}})$ with a small Gaussian noise for numerical stability, i.e., + +$$ +\begin{array}{l} \nabla_ {\mathbf {x} _ {0}} \log p (\mathbf {x} _ {0} \mid \mathbf {x} _ {t}) = \nabla_ {\mathbf {x} _ {0}} \log p (\mathbf {x} _ {t} \mid \mathbf {x} _ {0}) + \nabla_ {\mathbf {x} _ {0}} \log p (\mathbf {x} _ {0}) \\ \approx \nabla_ {\mathbf {x} _ {0}} \log p \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {0}\right) + s _ {\boldsymbol {\theta}} \left(\mathbf {x} _ {0}, t _ {\min }\right). \tag {6} \\ \end{array} +$$ + +Despite being a more accurate estimation, the diffusion-score estimation Eq. (6) is much more expensive than using a Gaussian approximation when applied to Eq. (4), as it requires calling the score function multiple times. Empirical studies in Sec. 4.4 show that using a Gaussian approximation can + +achieve comparable results as the diffusion-score estimation, but is significantly more time-efficient. In practice, we adopt the Gaussian approximation and compute $\hat{\mathbf{x}}_0(\mathbf{x}_t)$ by solving the (unconditional) probability flow ODE starting at $\mathbf{x}_t$ . We leave details in Appendix F.2. + +When the measurement noise is an isotropic Gaussian, i.e., $\mathbf{n} \sim \mathcal{N}(\mathbf{0}, \beta_{\mathbf{y}}^{2}\mathbf{I})$ , the update rule simplifies to: + +$$ +\begin{array}{l} \mathbf {x} _ {0} ^ {(j + 1)} = \mathbf {x} _ {0} ^ {(j)} - \eta \nabla_ {\mathbf {x} _ {0} ^ {(j)}} \frac {\left\| \mathbf {x} _ {0} ^ {(j)} - \hat {\mathbf {x}} _ {0} (\mathbf {x} _ {t}) \right\| ^ {2}}{2 r _ {t} ^ {2}} \\ - \eta \nabla_ {\mathbf {x} _ {0} ^ {(j)}} \frac {\| \mathcal {A} \left(\mathbf {x} _ {0} ^ {(j)}\right) - \mathbf {y} \| ^ {2}}{2 \beta_ {\mathbf {y}} ^ {2}} + \sqrt {2 \eta} \boldsymbol {\epsilon} _ {j}, \tag {7} \\ \end{array} +$$ + +Combining Eq. (7) with Proposition 1, we implement Decoupled Annealing Posterior Sampling (DAPS) as Algorithm 1 in Appendix A. In particular, given a noise schedule $\sigma_{t}$ and a time discretization $\{t_i,i = 0,\dots ,N_A\}$ , we iteratively sample $\mathbf{x}_{t_i}$ from the measurement-conditioned time-marginal $p(\mathbf{x}_{t_i}|\mathbf{y})$ for $i = N_{A},N_{A} - 1,\ldots ,0$ , following Eq. (2) in Proposition 1. This algorithm provides an approximate sample $\mathbf{x}_0$ from the posterior distribution $p(\mathbf{x}_0|\mathbf{y})$ . + +The computational cost of Langevin dynamics mostly originates from evaluating the measurement function $\mathcal{A}$ . In most image restoration tasks, this operation is significantly more efficient compared to evaluating the diffusion model. As demonstrated in Appendix E.1, the MCMC sampling introduces only a small overhead to the sampling process. + +# 3.3. DAPS with Latent Diffusion Model + +Given a pretrained encoder $\mathcal{E}$ and decoder $\mathcal{D}$ , the latent diffusion models (LDMs) [40] are trained to fit the latent distribution $p(\mathbf{z}_0)$ , where $\mathbf{z}_0$ is given by the encoder $\mathbf{z}_0 = \mathcal{E}(\mathbf{x}_0)$ . One can recover $\mathbf{x}_0$ from the decoder $\mathbf{x}_0 = \mathcal{D}(\mathbf{z}_0)$ . We found the proposed method can be adapted to sampling with pre-trained latent diffusion models by factorizing the probabilistic graphical model for the latent diffusion process, as shown in Fig. 3b. We refer to this as LatentDAPS in the following paper. + +Accordingly, we can derive the following Langevin dynamics updating rule from Eq. (4): + +$$ +\begin{array}{l} \mathbf {z} _ {0} ^ {(j + 1)} = \mathbf {z} _ {0} ^ {(j)} + \eta \nabla_ {\mathbf {z} _ {0} ^ {(j)}} \log p (\mathbf {z} _ {0} ^ {(j)} | \mathbf {z} _ {t}) \\ + \eta \nabla_ {\mathbf {z} _ {0} ^ {(j)}} \log p (\mathbf {y} \mid \mathcal {D} \left(\mathbf {z} _ {0} ^ {(j)}\right)) + \sqrt {2 \eta} \epsilon_ {j}. \tag {8} \\ \end{array} +$$ + +While LatentDAPS is a straightforward and natural extension of DAPS compared to recent latent diffusion inverse problem solvers [42, 43, 45], it demonstrates comparable or even superior performance across several tasks in Tab. 1 and Tab. 4. More detailed derivation and proofs are shown in Appendix B. + +# 3.4. Discussion with Existing Methods + +Comparison with other posterior sampling methods. Unlike most previous algorithms, our sampling method does not + +![](images/d53d44f9f84ecef58cdf9639c8b2983023f8e2a8adc1fc6d785f6bf5d7c6410e.jpg) +Figure 4. DAPS vs. DPS on 2D synthetic data. Consecutive sampling steps are close to each other for DPS but not for DAPS. Here DAPS approximates the posterior better than DPS. + +solve a specific SDE/ODE; instead, we recursively sample from the time-marginal $p(\mathbf{x}_t \mid \mathbf{y})$ with noise annealing to zero. We therefore decouple the dependency of $\mathbf{x}_t$ and $\mathbf{x}_{t + \Delta t}$ in the sampling process. We argue that decoupling helps correct the errors accumulated in early diffusion sampling steps by allowing non-local transitions. This is particularly important when the measurement function is nonlinear. As a concrete example, Fig. 4 compares DAPS and DPS when solving a 2D nonlinear inverse problem with a Gaussian mixture prior. We visualize the trajectories of $\mathbf{x}_t$ from time $T$ to 0 for both methods. For DAPS, points on the trajectory have significantly larger variations compared to those of DPS. As a result, DPS converges to wrong solutions, but DAPS is able to approximate the true posterior distribution. We provide more discussions in Appendix G. + +Comparison with optimization-based methods. Although our method does not directly involve inner-loop optimization, we highlight that it connects with existing optimization-based solvers for inverse problems. For example, ReSample [45], DiffPIR [59], and DCDP [31] alternate between denoising optimizing, and resampling to solve inverse problems. However, these methods solve the MAP estimation problems, while DAPS draws samples from the posterior distribution. In particular, instantiating DAPS with Langevin dynamics and Gaussian approximation resembles [45] if we set standard deviation $\beta_{\mathbf{y}}\rightarrow 0$ and assume $\eta \beta_{\mathbf{y}}^{2}$ is constant in Eq. (18). + +# 4. Experiments + +# 4.1. Experimental Setup + +We evaluate our method using both pixel-space and latent diffusion models. For pixel-based diffusion experiments, we leverage the pre-trained diffusion models trained by [13] on the FFHQ dataset and the pre-trained model from [16] on the ImageNet dataset. For latent diffusion models, we use the same pre-trained models as [45]: the unconditional + +LDM-VQ4 trained on FFHQ and ImageNet by [40]. The autoencoder's downsample factor is 4. We use the same time step discretization and noise schedule as EDM [26]. + +As mentioned in Sec. 3, we implement a few-step Euler ODE solver to compute $\hat{\mathbf{x}}_0(\mathbf{x}_t)$ , while maintaining the same number of neural function evaluations (NFE) for different noise levels. In our experiments, we use DAPS-1K for all linear tasks and DAPS-4K for all nonlinear tasks. DAPS-1K uses 4 ODE solver NFE and 250 annealing scheduling steps, while DAPS-4K uses 10 NFE and 400 steps, respectively. We discuss the effect of choosing these configurations in Sec. 4.4. The corresponding latent version, LatentDAPS, follows the same settings but performs sampling in the latent space of VAEs. We use 100 and 50 Langevin steps per denoising iteration for DAPS and LatentDAPS respectively, and tune learning rates separately for each task. Further details on model configurations, samplers, and other hyperparameters are provided in Appendix F, along with more discussion on sampling efficiency in Sec. 4.2 and Appendix E.1. + +Datasets and metrics. Adopting the previous convention, we test our method on two image datasets, FFHQ $256 \times 256$ [25] and ImageNet $256 \times 256$ [15]. To evaluate our method, we use 100 images from the validation set for both FFHQ and ImageNet. We include peak signal-to-noise-ratio (PSNR), structural similarity index measure (SSIM), Learned Perceptual Image Patch Similarity (LPIPS)[58] score and Fréchet inception distance (FID)[20] as our main evaluation metrics. For both our method and baselines, we use the versions implemented in pic[28] with all images normalized to the range [0, 1]. The replace-pooling option is enabled for LPIPS evaluation. + +Inverse problems. We evaluate our method with a series of linear and nonlinear tasks. For linear inverse problems, we consider (1) super-resolution, (2) Gaussian deblurring, (3) motion deblurring, (4) inpainting (with a box mask), and (5) inpainting (with a $70\%$ random mask). For Gaussian and motion deblurring, kernels of size $61 \times 61$ with standard deviations of 3.0 and 0.5, respectively, are used. In the super-resolution task, a bicubic resizer downscales images by a factor of 4. The box inpainting task uses a random box of size $128 \times 128$ to mask the original images, while random mask inpainting uses a generated random mask where each pixel has a $70\%$ chance of being masked, following the settings in [45]. + +We consider three nonlinear inverse problems: (1) phase retrieval, (2) high dynamic range (HDR) reconstruction, and (3) nonlinear deblurring. Due to the inherent instability of phase retrieval, we adopt the strategy from DPS [13], using an oversampling rate of 2.0 and reporting the best result out of four independent samples. The goal of HDR reconstruction is to recover a higher dynamic range image (factor of 2) from a low dynamic range image. For nonlinear deblurring, we use the default setting as described in [53]. + +All linear and nonlinear measurements are subject to white Gaussian noise with a standard deviation of $\beta_{\mathbf{y}} = 0.05$ . Further details regarding the forward measuring functions for each task and their respective hyperparameters are provided in the Appendix F. + +Baselines. We compare our methods with the following baselines: DDRM [30], DPS [13], DDNM [54], DCDP [31], FPS-SMC [18], DiffPIR [59] for pixel-based diffusion model experiments. We compare the latent diffusion version of our methods with PSLD [42] and ReSample [45], which also operate in the latent space. Note that DDRM, DDNM, FPS-SMC, and PSLD cannot handle nonlinear inverse problems. We especially RED-diff [35] for nonlinear experiments and a traditional method HIO [14] for phase retrieval. + +# 4.2. Main Results + +We show quantitative results for FFHQ and ImageNet dataset in Tab. 1. The qualitative comparisons are provided in Fig. 1. As shown in Tab. 1, our method achieves comparable or even better performance across all selected linear inverse problems. Moreover, our methods are remarkably stable when handling nonlinear inverse problems. For example, although existing methods such as DPS can recover high-quality images from phase retrieval measurements, they suffer from an extremely high failure rate, resulting in a relatively low PSNR and a high LPIPS. However, DAPS demonstrates superior stability across different samples and achieves a significantly higher success rate, resulting in a substantial improvement in PSNR and LPIPS as evidenced by the results in Tab. 1. Moreover, DAPS captures and recovers much more fine-grained details in measurements compared to existing baselines. We include more samples and comparisons in Appendix I for further illustration. + +To understand the evolution of the noise annealing procedure, we evaluate two crucial trajectories in our method: 1) the estimated mean of $p(\mathbf{x}_0 \mid \mathbf{x}_t)$ , denoted as $\hat{\mathbf{x}}_0(\mathbf{x}_t)$ ; and 2) samples from $p(\mathbf{x}_0 \mid \mathbf{x}_t, \mathbf{y})$ obtained through Langevin dynamics, denoted as $\mathbf{x}_{0|\mathbf{y}}$ . We assess image quality and measurement error across these trajectories, as depicted in Fig. 7. The measurement error for $\mathbf{x}_{0|\mathbf{y}} \sim p(\mathbf{x}_0 \mid \mathbf{x}_t, \mathbf{y})$ remains consistently low throughout the process, while the measurement error for $\hat{\mathbf{x}}_0(\mathbf{x}_t)$ continuously decreases. Consequently, metrics such as PSNR and LPIPS exhibit stable improvement as the noise level is annealed. + +In addition, our method can generate diversified samples when the measurements contain less information. To illustrate this, we selected two tasks where the posterior distributions may have multiple modes: (1) super-resolution by a factor of 16, and (2) box inpainting with a box size of $192 \times 192$ . Some samples from DAPS are shown in Fig. 5, demonstrating DAPS's ability to produce diverse samples while preserving the measurement information. We show more samples in Appendix I for further demonstration. + +![](images/f6078223ab458b1e78da54664d9af0d9574b4c2e03574d3252341b6863a28a26.jpg) +(a) Super resolution of a factor 16 + +![](images/534690a36ae5de8fa48795aa0c5e1b381c4df00b5bda5e277e80f05e89c80af7.jpg) +(b) Inpainting of $192\times 192$ box +Figure 5. Sample diversity. We present several diverse samples generated by the DAPS under two sparse measurements whose posterior distributions contain multiple modes. DAPS produces a variety of samples with distinct features, including differences in expression, wearings, and hairstyles. + +DAPS with large-scale text-conditioned latent diffusion. Due to the flexibility of DAPS, pretrained large-scale text-conditioned latent diffusion models (LDMs) [38-40] can also serve as the prior model $p(\mathbf{x}_0 \mid \mathbf{c})$ when given a text prompt c. To demonstrate DAPS's potential, we use Stable Diffusion v2.1 [40] to present LatentDAPS results on three tasks with $768 \times 768$ resolution images, as shown in Fig. 1. Additionally, we provide the text prompts used, along with a detailed discussion, in Appendix B.1, where we also include a quantitative evaluation. + +DAPS with discrete diffusion. DAPS can be naturally extended to support posterior sampling with categorical prior distribution modeled by discrete diffusion models [3, 6, 33]. Instead of sampling $\mathbf{x}_{0|\mathbf{y}}$ by Langevin MC or Hamiltonian MC in continuous space, we use the Metropolis-Hasting algorithm to replace Eq. (4). We validate DAPS with discrete diffusion on a quantized MNIST dataset. We leave the details to Appendix C. + +# 4.3. Further results on MRI + +We further validate the effectiveness of DAPS on compressed sensing multi-coil MRI, which is a medical imaging technique that helps reduce the scan time of MRI via subsampling. We train a diffusion model using the pipeline from [26] with fastMRI knee dataset [57]. Tab. 2 shows that DAPS outperforms existing diffusion-based baselines. We include more qualitative results in Appendix H. + +
TaskTypeMethodFFHQImageNet
PSNR (↑)SSIM (↑)LPIPS (↓)FID (↓)PSNR (↑)SSIM (↑)LPIPS (↓)FID (↓)
Super resolution 4xPixelDAPS (ours)29.070.8180.17751.4425.890.6940.27683.57
DPS25.860.7530.26981.0721.130.4890.361106.32
DDRM26.580.7820.28279.2522.620.5210.324103.85
DDNM28.030.7950.19764.6223.960.6040.47598.62
DCDP28.660.8070.17853.81----
FPS-SMC28.420.8130.20449.2524.820.7030.31397.51
DiffPIR26.64-0.26065.7723.18-0.371106.32
LatentLatentDAPS (ours)27.480.8010.18259.6225.060.6730.27684.37
PSLD24.350.6490.28774.3625.420.6940.36097.45
ReSample23.290.5940.39293.1822.610.5760.370113.42
Inpaint (box)PixelDAPS (ours)24.070.8140.13343.1021.430.7250.214109.85
DPS22.510.7920.20961.2718.940.7220.257126.52
DDRM22.260.8010.20778.6218.630.7330.254116.37
DDNM24.470.8370.23546.5921.640.7480.319103.97
DCDP23.890.7600.16345.23----
FPS-SMC24.860.8230.14648.3422.160.7260.208111.58
LatentLatentDAPS (ours)23.990.8020.19446.5217.190.6240.340145.63
PSLD24.220.8130.15843.0220.100.6940.465146.53
ReSample20.060.7490.18453.2118.290.6310.262127.84
Inpaint (random)PixelDAPS (ours)31.120.8440.09832.1728.440.7750.13554.25
DPS25.460.8230.20369.2023.520.7450.29787.53
DDNM29.910.8170.12144.3731.160.8410.19163.84
DCDP30.690.8420.14252.51----
FPS-SMC28.210.8230.26161.2324.520.7010.31679.12
LatentLatentDAPS (ours)30.710.8130.14136.4127.590.7720.16461.62
PSLD30.310.8090.22147.2131.300.7830.33783.21
ReSample29.610.7460.14039.8527.500.7560.14359.87
Gaussian deblurringPixelDAPS (ours)29.190.8170.16553.3326.150.6840.25375.68
DPS25.870.7640.21979.7520.310.5980.397116.42
DDRM24.930.7320.23992.4321.260.5640.443146.89
DDNM28.200.8040.21657.8328.060.7030.27881.43
DCDP27.500.6990.30486.43----
FPS-SMC26.540.7730.25367.4523.910.6010.38791.72
DiffPIR27.36-0.23659.6522.80-0.35593.36
LatentLatentDAPS (ours)27.930.7640.23464.5225.050.6680.34578.51
PSLD23.270.6310.31689.5125.860.6880.39091.39
ReSample26.390.7140.25571.6925.970.7030.25465.35
Motion deblurringPixelDAPS (ours)29.660.8470.15739.4927.860.7660.19661.83
DPS24.520.8010.24665.2318.960.6290.423137.81
DCDP25.080.5120.364125.13----
FPS-SMC27.390.8260.22748.3224.520.6470.32687.43
DiffPIR26.57-0.25565.7824.01-0.36694.63
LatentLatentDAPS (ours)27.000.8140.28346.8426.830.7450.29673.62
PSLD22.310.6780.33696.1520.850.5940.511124.67
ReSample27.410.8230.19844.7226.940.7380.22766.89
Phase retrievalPixelDAPS (ours)30.63±3.130.851±0.0720.139±0.06042.7125.78±6.920.743±0.0840.254±0.12582.67
DPS17.64±2.970.441±0.1290.410±0.090104.5216.81±3.610.427±0.1430.447±0.099197.54
RED-diff15.60±4.480.398±0.1950.596±0.092167.4314.98±3.750.386±0.0570.536±0.129212.24
DCDP28.65±8.090.781±0.2170.203±0.19668.13----
LatentLatentDAPS (ours)29.16±3.550.796±0.0890.199±0.07854.2620.54±6.410.612±0.1140.361±0.150129.54
ReSample21.60±8.100.648±0.1540.406±0.22484.32----
ClassicalHIO13.53±2.500.359±0.0930.726±0.068268.09----
Nonlinear deblurPixelDAPS (ours)28.29±1.770.783±0.0360.155±0.03249.3827.73±3.230.724±0.0480.169±0.05659.87
DPS23.39±2.010.623±0.0820.278±0.06091.3122.49±3.200.591±0.1010.306±0.081101.41
RED-diff30.86±0.510.795±0.0280.160±0.03443.8430.07±1.410.754±0.0230.211±0.08351.22
DCDP27.92±2.640.779±0.0670.183±0.05151.96----
LatentLatentDAPS (ours)28.11±1.750.713±0.0410.235±0.04953.6325.34±3.440.615±0.0570.314±0.08076.73
ReSample28.24±1.690.742±0.0390.185±0.03951.6226.20±3.710.653±0.0640.206±0.05761.16
High dynamic rangePixelDAPS (ours)27.12±3.530.752±0.0410.162±0.07242.9726.30±4.100.717±0.0670.175±0.10764.19
DPS22.73±6.070.591±0.1410.264±0.156112.8219.23±2.520.582±0.0820.503±0.106146.23
RED-diff22.16±3.410.512±0.0830.258±0.089108.3222.03±5.900.601±0.0940.274±0.198113.48
LatentLatentDAPS (ours)25.94±2.870.751±0.0560.223±0.08074.8323.64±4.100.609±0.0530.269±0.09993.51
ReSample25.65±3.570.732±0.0590.182±0.08567.2225.11±4.210.633±0.0490.198±0.08987.66
+ +Table 1. Quantitative evaluation on FFHQ and ImageNet on 5 linear and 3 nonlinear tasks. Performance comparison of different methods on FFHQ (left) and ImageNet (right) in the image domain. For linear tasks, we report the mean performance (PSNR, SSIM and LPIPS) across 100 validation images. We provide both the mean and standard deviation for nonlinear tasks, where results exhibit higher instability. The FID scores are evaluated with the same 100 validation images. The best and second-best results within each type of task are indicated by bold and underlined marks, respectively. DAPS demonstrates superior performance on most tasks, with LatentDAPS showing competitive results. All tasks are using noisy measurement with noise level $\beta_{\mathbf{y}} = 0.05$ . + +![](images/7675a9626a3aa211cda5e21aa15e1652c332ca03d231d43c6811f68b7c50e196.jpg) +(a) Gaussian deblur + +![](images/5d319c6ba091b9dfa55da4314901ef3f7835e0615d0612c26b128cd32dcd3555.jpg) +(b) Phase retrieval + +![](images/7e725e42bb7e60312e10069a4ceb4a5bc38608cb4b0537da63a2ac32011fd871.jpg) +Figure 6. Evaluation of time cost and sample quality. The x-axis indicates the single image sampling time on an A100-SXM4-80GB GPU and y-axis shows the LPIPS. The evaluation uses 100 FFHQ images. +Figure 7. Quantitative evaluations of image quality metrics and measurement error throughout the sampling process. Here $\hat{\mathbf{x}}_0(\mathbf{x}_t)$ stands for the estimated mean of the Gaussian approximated distribution and $\mathbf{x}_{0|\mathbf{y}}\sim p(\mathbf{x}_0\mid \mathbf{x}_t,\mathbf{y})$ means the samples from measurement conditioned time-marginal. The measurement errors are normalized between [0, 1] for visualization. + +
×4×8
PSNR (↑)SSIM (↑)Meas err (↓)PSNR (↑)SSIM (↑)Meas err (↓)
DAPS31.480.7621.5729.010.6811.28
DPS26.130.6209.9020.820.5366.74
DiffPIR28.310.63210.5526.780.5887.79
ScoreMRI [10]25.970.46810.8325.010.4058.36
CSGM [23]28.780.7101.5226.150.6251.14
+ +# 4.4. Ablation Studies + +Time efficiency of DAPS. We run DAPS with various configurations to test its performance under different computing budgets. Specifically, we evaluate DAPS with the number of function evaluations (NFE) of the diffusion models ranging from 50 to 4K, with configurations specified in Appendix E.1. Sec. 4.2 shows the relation between the sample quality of DAPS measured by LPIPS with its time cost, which indicates that DAPS is able to generate high-quality samples with a reasonable time cost. + +Effectiveness of Different Design choices of DAPS. We conduct an ablation study to demonstrate the effectiveness of Gaussian approximation compared with the more principled but more expensive diffusion-score estimation. The results in Tab. 3 demonstrate that Gaussian approximation is $6 \sim 7$ times faster than diffusion-score estimation with the cost of a slight performance drop, effectively trading off performance + +Table 2. Quantitative results of compressed sensing MRI. We consider $\times 4$ and $\times 8$ subsampling ratio. Meas err stands for ${\ell }_{2}$ data consistency error. + +
Gaussian deblurringHigh dynamic range
PSNR (↑)LPIPS (↓)Run time/sPSNR (↑)LPIPS (↓)Run time/s
Gaussian approx.29.400.17064.724.850.18161.4
diffusion-score29.570.160462.325.330.193540.0
Relative improvement-0.575%-6.25%614.53%-1.89%-6.22%779.48%
+ +Table 3. Comparison of approaches to approximate $p(\mathbf{x}_0 \mid \mathbf{x}_t)$ . Three tasks on 10 FFHQ images with DAPS-1K consistently shown Gaussian approximation achieve huge speed-up while slightly sacrificing the performance compared with diffusion-score estimation. + +for efficiency. + +More ablation studies. We include more ablation studies of DAPS in Appendix I.1, including the number of function evaluations, the number of ODE steps, the annealing noise scheduling, and different measurement noise levels. + +# 5. Conclusion + +In summary, we propose Decoupled Annealing Posterior Sampling (DAPS) for solving inverse problems, particularly those with complex nonlinear measurement processes such as phase retrieval. Our method decouples consecutive sample points in a diffusion sampling trajectory, allowing them to vary considerably, thereby enabling DAPS to explore a larger solution space. Empirically, we demonstrate that DAPS generates samples with better visual quality and stability compared to existing methods when solving a wide range of challenging inverse problems. + +# Acknowledgments + +We are grateful to Pika for providing the computing resources essential for this research. We also extend our thanks to the Kortschak Scholars Fellowship for supporting B.Z. and W.C. at Caltech. J.B. acknowledges support from the Wally Baer and Jeri Weiss Postdoctoral Fellowship. A.A. is supported in part by Bren endowed chair and by the AI2050 senior fellow program at Schmidt Sciences. + +# References + +[1] Kazunori Akiyama, Antxon Alberdi, Walter Alef, Keiichi Asada, Rebecca Azulay, Anne-Kathrin Baczko, David Ball, Mislav Balokovic, John Barrett, Dan Bintley, et al. First m87 event horizon telescope results. iv. imaging the central supermassive black hole. The Astrophysical Journal Letters, 875(1):L4, 2019. 1 +[2] Marius Arvinte, Sriram Vishwanath, Ahmed H. Tewfik, and Jonathan I. Tamir. Deep j-sense: Accelerated mri reconstruction via unrolled alternating optimization, 2021. 3 +[3] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981-17993, 2021. 6 +[4] Michael Betancourt and Mark Girolami. Hamiltonian monte carlo for hierarchical models. Current trends in Bayesian methodology with applications, 79(30):2-4, 2015. 4, 1 +[5] Benjamin Boys, Mark Girolami, Jakiw Pidstrigach, Sebastian Reich, Alan Mosca, and O. Deniz Akyildiz. Tweedie moment projected diffusions for inverse problems, 2023. 2, 3, 4 +[6] Andrew Campbell, Joe Benton, Valentin De Bortoli, Thomas Rainforth, George Deligiannidis, and Arnaud Doucet. A continuous time framework for discrete denoising models. Advances in Neural Information Processing Systems, 35:28266-28279, 2022. 6 +[7] Gabriel Cardoso, Yazid Janati El Idrissi, Sylvain Le Corff, and Eric Moulines. Monte carlo guided diffusion for bayesian linear inverse problems, 2023. 3 +[8] Gabriel Victorino Cardoso, Yazid Janati, Sylvain Le Corff, and Éric Moulines. Monte carlo guided diffusion for bayesian linear inverse problems. ArXiv, abs/2308.07983, 2023. 3 +[9] Junqing Chen and Haibo Liu. An alternating direction method of multipliers for inverse lithography problem, 2023. 3 +[10] Hyungjin Chung and Jong Chul Ye. Score-based diffusion models for accelerated mri. Medical Image Analysis, page 102479, 2022. 1, 3, 8, 13 +[11] Hyungjin Chung, Eun Sun Lee, and Jong Chul Ye. Mr image denoising and super-resolution using regularized reverse diffusion. IEEE Transactions on Medical Imaging, 42(4): 922-934, 2022. 1 +[12] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. In Advances in Neural Information Processing Systems, 2022. 3 +[13] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior + +sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 3, 4, 5, 6, 10, 11, 13 +[14] J. Dainty and James Fienup. Phase retrieval and image reconstruction for astronomy. Image Recovery: Theory Appl, 13, 1987. 6 +[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 5 +[16] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In Advances in Neural Information Processing Systems, 2021. 5 +[17] Zolnamar Dorjsembe, Hsing-Kuo Pao, Sodtavilan Odonchimed, and Furen Xiao. Conditional diffusion models for semantic 3d brain mri synthesis. IEEE Journal of Biomedical and Health Informatics, 2024. 1 +[18] Zehao Dou and Yang Song. Diffusion posterior sampling for linear inverse problem solving: A filtering perspective. In The Twelfth International Conference on Learning Representations, 2024. 3, 6, 11 +[19] Berthy T. Feng, Jamie Smith, Michael Rubinstein, Huiwen Chang, Katherine L. Bouman, and William T. Freeman. Score-based diffusion models as principled priors for inverse imaging. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10520-10531, 2023. 1 +[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. 5 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 2 +[22] Alex Ling Yu Hung, Kai Zhao, Haoxin Zheng, Ran Yan, Steven S Raman, Demetri Terzopoulos, and Kyunghyun Sung. Med-cdiff: Conditional medical image generation with diffusion models. Bioengineering, 10(11):1258, 2023. 1 +[23] Ajil Jalal, Marius Arvinte, Giannis Daras, Eric Price, Alexandros G Dimakis, and Jon Tamir. Robust compressed sensing mri with deep generative priors. In Advances in Neural Information Processing Systems, pages 14938-14954. Curran Associates, Inc., 2021. 1, 2, 3, 8, 13 +[24] Ulugbek Kamilov, Charles Bouman, Gregory Buzzard, and Brendt Wohlberg. Plug-and-play methods for integrating physical and learned models in computational imaging: Theory, algorithms, and applications. IEEE Signal Processing Magazine, 40:85-97, 2023. 3 +[25] T. Karras, S. Laine, and T. Aila. A style-based generator architecture for generative adversarial networks. IEEE Transactions on Pattern Analysis & Machine Intelligence, 43(12): 4217-4228, 2021. 5, 10 +[26] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, 2022. 2, 5, 6, 10, 13 + +[27] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models, 2024. 2 +[28] Sergey Kastryulin, Jamil Zakirov, Denis Prokopenko, and Dmitry V. Dylov. Pytorch image quality: Metrics for image quality assessment, 2022. 5 +[29] Bahjat Kawar, Gregory Vaksman, and Michael Elad. SNIPS: Solving noisy inverse problems stochastically. Advances in Neural Information Processing Systems, 34:21757-21769, 2021. 3 +[30] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 6, 11 +[31] Xiang Li, Soo Min Kwon, Ismail R. Alkhouri, Saiprasad Ravishankar, and Qing Qu. Decoupled data consistency with diffusion purification for image restoration, 2024. 3, 5, 6, 11 +[32] Hanyuan Liu, Jinbo Xing, Minshan Xie, Chengze Li, and Tien-Tsin Wong. Improved diffusion-based image colorization via piggybacked models, 2023. 2 +[33] Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion language modeling by estimating the ratios of the data distribution. arXiv preprint arXiv:2310.16834, 2023. 6 +[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models, 2022. 1 +[35] Morteza Mardani, Jiaming Song, Jan Kautz, and Arash Vahdat. A variational perspective on solving inverse problems with diffusion models. arXiv preprint arXiv:2305.04391, 2023. 6, 11, 14 +[36] Xiangming Meng and Yoshiyuki Kabashima. Diffusion model based posterior sampling for noisy linear inverse problems. arXiv preprint arXiv:2211.12343, 2022. 3 +[37] Nicholas Metropolis, Arianna W Rosenbluth, Marshall N Rosenbluth, Augusta H Teller, and Edward Teller. Equation of state calculations by fast computing machines. The journal of chemical physics, 21(6):1087-1092, 1953. 1 +[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 6 +[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 4 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 4, 5, 6, 3 +[41] Litu Rout, Yujia Chen, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Beyond first-order tweedie: Solving inverse problems using latent diffusion, 2023. 3 +[42] Litu Rout, Negin Raoof, Giannis Daras, Constantine Caramanis, Alex Dimakis, and Sanjay Shakkottai. Solving linear inverse problems provably via posterior sampling with latent diffusion models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4, 6, 9, 11 + +[43] Litu Rout, Yujia Chen, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Beyond first-order tweedie: Solving inverse problems using latent diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9472-9481, 2024. 4 +[44] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2023. 1 +[45] Bowen Song, Soo Min Kwon, Zecheng Zhang, Xinyu Hu, Qing Qu, and Liyue Shen. Solving inverse problems with latent diffusion models via hard data consistency. In The Twelfth International Conference on Learning Representations, 2024, 2, 3, 4, 5, 6, 9, 11 +[46] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 3, 4 +[47] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Neural Information Processing Systems, 2019. 2, 3 +[48] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. +[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 1, 2, 3, 10 +[50] Yang Song, Liyue Shen, Lei Xing, and Stefano Ermon. Solving inverse problems in medical imaging with score-based generative models. In International Conference on Learning Representations, 2022. 1, 2, 3 +[51] He Sun and Katherine L. Bouman. Deep probabilistic imaging: Uncertainty quantification and multi-modal solution characterization for computational imaging, 2020. 1 +[52] Yu Sun, Zihui Wu, Yifan Chen, Berthy Feng, and Katherine L. Bouman. Provable probabilistic imaging using score-based generative priors. *ArXiv*, abs/2310.10835, 2023. 1 +[53] Phong Tran, Anh Tran, Quynh Phung, and Minh Hoai. Explore image deblurring via encoded blur kernel space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5 +[54] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. The Eleventh International Conference on Learning Representations, 2023. 3, 6, 11 +[55] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688. CiteSeer, 2011. 4 +[56] Zihui Wu, Yu Sun, Yifan Chen, Bingliang Zhang, Yisong Yue, and Katherine L Bouman. Principled probabilistic imaging using diffusion models as plug-and-play priors. arXiv preprint arXiv:2405.18782, 2024. 3 + +[57] Jure Zbontar, Florian Knoll, Anuroop Sriram, Matthew J. Muckley, Mary Bruno, Aaron Defazio, Marc Parente, Krzysztof J. Geras, Joe Katsnelson, Hersh Chandarana, Zizhao Zhang, Michal Drozdal, Adriana Romero, Michael G. Rabbat, Pascal Vincent, James Pinkerton, Duo Wang, Nafissa Yakubova, Erich Owens, C. Lawrence Zitnick, Michael P. Recht, Daniel K. Sodickson, and Yvonne W. Lui. fastmri: An open dataset and benchmarks for accelerated MRI. CoRR, abs/1811.08839, 2018. 6, 13 +[58] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 5 +[59] Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising diffusion models for plug-and-play image restoration. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (NTIRE), 2023. 1, 3, 5, 6, 11, 13 \ No newline at end of file diff --git a/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/images.zip b/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6d2a08572e68ac14d89c3cbc951d56e8644640b1 --- /dev/null +++ b/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b19d1a4933521535200e4ccc953d4864ff2fe5a82e148b60439a5f275b6b1461 +size 831040 diff --git a/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/layout.json b/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..614a55f372e31918113da1836669f67c54950c7e --- /dev/null +++ b/2025/Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing/layout.json @@ -0,0 +1,10668 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 103, + 547, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 103, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 63, + 103, + 547, + 121 + ], + "type": "text", + "content": "Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 143, + 230, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 143, + 230, + 158 + ], + "spans": [ + { + "bbox": [ + 135, + 143, + 230, + 158 + ], + "type": "text", + "content": "Bingliang Zhang\\*,1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 158, + 216, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 158, + 216, + 171 + ], + "spans": [ + { + "bbox": [ + 140, + 158, + 216, + 171 + ], + "type": "text", + "content": "Chenlin Meng3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 171, + 271, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 171, + 271, + 186 + ], + "spans": [ + { + "bbox": [ + 102, + 171, + 271, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 171, + 271, + 186 + ], + "type": "text", + "content": "California Institute of Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 280, + 143, + 348, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 143, + 348, + 156 + ], + "spans": [ + { + "bbox": [ + 280, + 143, + 348, + 156 + ], + "type": "text", + "content": "Wenda Chu\\*,1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 264, + 157, + 369, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 157, + 369, + 170 + ], + "spans": [ + { + "bbox": [ + 264, + 157, + 369, + 170 + ], + "type": "text", + "content": "Anima Anandkumar" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 285, + 171, + 334, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 171, + 334, + 184 + ], + "spans": [ + { + "bbox": [ + 285, + 171, + 334, + 184 + ], + "type": "text", + "content": "2NVIDIA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 397, + 143, + 474, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 143, + 474, + 156 + ], + "spans": [ + { + "bbox": [ + 397, + 143, + 474, + 156 + ], + "type": "text", + "content": "Julius Berner" + }, + { + "bbox": [ + 397, + 143, + 474, + 156 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 419, + 157, + 476, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 157, + 476, + 171 + ], + "spans": [ + { + "bbox": [ + 419, + 157, + 476, + 171 + ], + "type": "text", + "content": "Yang Song" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 348, + 171, + 509, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 171, + 509, + 186 + ], + "spans": [ + { + "bbox": [ + 348, + 171, + 509, + 186 + ], + "type": "text", + "content": "3Stanford University 4OpenAI" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 58, + 213, + 258, + 321 + ], + "blocks": [ + { + "bbox": [ + 58, + 213, + 258, + 321 + ], + "lines": [ + { + "bbox": [ + 58, + 213, + 258, + 321 + ], + "spans": [ + { + "bbox": [ + 58, + 213, + 258, + 321 + ], + "type": "image", + "image_path": "9abd721ce8744c5010bf45cc2c0bd44c288c8942c9b1da06449f29c171ef9ab3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 263, + 214, + 552, + 437 + ], + "blocks": [ + { + "bbox": [ + 263, + 214, + 552, + 437 + ], + "lines": [ + { + "bbox": [ + 263, + 214, + 552, + 437 + ], + "spans": [ + { + "bbox": [ + 263, + 214, + 552, + 437 + ], + "type": "image", + "image_path": "c79e7bea559a0bccdf463a491c29f43836ecb0933685c487199626b1088828ad.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 58, + 329, + 257, + 435 + ], + "blocks": [ + { + "bbox": [ + 58, + 329, + 257, + 435 + ], + "lines": [ + { + "bbox": [ + 58, + 329, + 257, + 435 + ], + "spans": [ + { + "bbox": [ + 58, + 329, + 257, + 435 + ], + "type": "image", + "image_path": "8501a45e9b3699b1069f727d48f5c7690ff3906ce2dd01d463a7e6ef5169440c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 445, + 555, + 500 + ], + "lines": [ + { + "bbox": [ + 55, + 445, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 555, + 500 + ], + "type": "text", + "content": "Figure 1. Overview of Decoupled Annealing Posterior Sampling (DAPS). Our method provides a flexible and effective framework for solving inverse problems through a decoupled posterior sampling process. In (a)(b), we present DAPS visual results on FFHQ and ImageNet at a resolution of 256, and in (c), on natural images at a resolution of 768. In (d), we display DAPS results on compressed sensing multi-coil MRI (CS-MRI). DAPS effectively addresses nonlinear inverse problems as well as medical imaging MRI challenges. Additionally, DAPS can be enhanced using large-scale latent diffusion models (LDMs) [40], as shown in (c)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 152, + 510, + 200, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 510, + 200, + 522 + ], + "spans": [ + { + "bbox": [ + 152, + 510, + 200, + 522 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 536, + 297, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 536, + 297, + 681 + ], + "spans": [ + { + "bbox": [ + 54, + 536, + 297, + 681 + ], + "type": "text", + "content": "Diffusion models have recently achieved success in solving Bayesian inverse problems with learned data priors. Current methods build on top of the diffusion sampling process, where each denoising step makes small modifications to samples from the previous step. However, this process struggles to correct errors from earlier sampling steps, leading to worse performance in complicated nonlinear inverse problems, such as phase retrieval. To address this challenge, we propose a new method called Decoupled Annealing Posterior Sampling (DAPS) that relies on a novel noise annealing process. Specifically, we decouple consecutive steps in a diffusion sampling trajectory, allowing them to vary consid" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "text", + "content": "erably from one another while ensuring their time-margins anneal to the true posterior as we reduce noise levels. This approach enables the exploration of a larger solution space, improving the success rate for accurate reconstructions. We demonstrate that DAPS significantly improves sample quality and stability across multiple image restoration tasks, particularly in complicated nonlinear inverse problems." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Inverse problems are ubiquitous in science and engineering, with applications ranging from image restoration [13, 30, 34, 44, 49, 59], medical imaging [10, 11, 17, 22, 23, 50] to astrophotography [1, 19, 51, 52]. Solving an inverse problem involves finding the underlying signal " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " from its partial, noisy measurement " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ". Since the measurement process is typically noisy and many-to-one, inverse problems do not have a" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 271, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 271, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 271, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 69, + 693, + 271, + 703 + ], + "type": "text", + "content": "Work partially done at the California Institute of Technology." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 69, + 703, + 225, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 225, + 712 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 225, + 712 + ], + "type": "text", + "content": "*These authors contributed equally to this work" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20895" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 191 + ], + "type": "text", + "content": "unique solution; instead, multiple solutions may exist that are consistent with the observed measurement. In the Bayesian inverse problem framework, the solution space is characterized by the posterior distribution " + }, + { + "bbox": [ + 55, + 72, + 297, + 191 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid \\mathbf{y}) \\propto p(\\mathbf{y} \\mid \\mathbf{x}_0) p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 72, + 297, + 191 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 297, + 191 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y} \\mid \\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 72, + 297, + 191 + ], + "type": "text", + "content": " represents the noisy measurement process, and " + }, + { + "bbox": [ + 55, + 72, + 297, + 191 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 72, + 297, + 191 + ], + "type": "text", + "content": " is the prior distribution. In this work, we aim to solve Bayesian inverse problems where the measurement process " + }, + { + "bbox": [ + 55, + 72, + 297, + 191 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y} \\mid \\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 72, + 297, + 191 + ], + "type": "text", + "content": " is known, and the prior distribution " + }, + { + "bbox": [ + 55, + 72, + 297, + 191 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 72, + 297, + 191 + ], + "type": "text", + "content": " is captured by a deep generative model trained on a corresponding dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "text", + "content": "As score-based diffusion models [21, 26, 27, 47-49] have risen to dominance in modeling high-dimensional data distributions like images, audio, and video, they have become the leading method for estimating the prior distribution " + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "text", + "content": " in Bayesian inverse problems. A diffusion model generates a sample " + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "text", + "content": " by smoothly removing noise from an unstructured initial noise sample " + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "text", + "content": " through solving stochastic differential equations (SDEs). In particular, each step of the sampling process recursively converts a noisy sample " + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t + \\Delta t}" + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "inline_equation", + "content": "\\Delta t > 0" + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "text", + "content": " denotes the step size, to a slightly less noisy sample " + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "text", + "content": " until " + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "text", + "content": ". This iterative structure in the diffusion sampling process can be leveraged to facilitate Bayesian inverse problem solving. In fact, prior research [5, 13, 46] has shown that given measurement " + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "text", + "content": " and a diffusion model prior " + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "text", + "content": ", we can sample from the posterior distribution " + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0\\mid \\mathbf{y})" + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "text", + "content": " by perturbing a reverse-time SDE using an approximate gradient of the measurement process, " + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{y}\\mid \\mathbf{x}_t)" + }, + { + "bbox": [ + 56, + 194, + 298, + 410 + ], + "type": "text", + "content": ", at every step of the SDE solver." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "text", + "content": "Despite the remarkable success of these methods in solving many real-world inverse problems, like image colorization [30, 32, 49], image super-resolution [13, 30, 46, 49], computed tomography [45, 50], and magnetic resonance imaging [23, 50], they face significant challenges in more complex inverse problems with nonlinear measurement processes, such as phase retrieval and nonlinear motion deblurring. This is partially because, in prior methods, each denoising step approximately samples from the distribution " + }, + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_t \\mid \\mathbf{x}_{t + \\Delta t}, \\mathbf{y})" + }, + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "text", + "content": ". This causes " + }, + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t + \\Delta t}" + }, + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "text", + "content": " to be close to each other because of using a small step size " + }, + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "text", + "content": " in discretizing the reverse-time SDE. As a result, " + }, + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "text", + "content": " can at most correct local errors in " + }, + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t + \\Delta t}" + }, + { + "bbox": [ + 55, + 411, + 297, + 639 + ], + "type": "text", + "content": " but struggles to correct global errors that require significant modifications to the prior sample. This issue is exacerbated when the methods are applied to complicated, nonlinear inverse problems, such as phase retrieval, where they often converge to undesired samples that are consistent with the measurement but reside in low-probability areas of the prior distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "To address this challenge, we propose a new framework for solving general inverse problems, termed Decoupled Annealing Posterior Sampling (DAPS). Our method employs a new noise annealing process inspired by the diffusion sampling process, where we decouple the consecutive samples " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t + \\Delta t}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " in the sampling trajectory, allowing samplers" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "content": "to correct large, global errors made in earlier steps. Instead of repetitively sampling from " + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_t \\mid \\mathbf{x}_{t + \\Delta t}, \\mathbf{y})" + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "content": " as in previous methods, which restricts the distances between consecutive samples " + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t + \\Delta t}" + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "content": ", DAPS recursively samples from the marginal distribution " + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_t \\mid \\mathbf{y})" + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "content": ". As illustrated in Fig. 2, we factorize the time marginal " + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_t \\mid \\mathbf{y})" + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "content": " into three conditional distributions and sample from them in turn by solving the reverse diffusion process, simulating Langevin dynamics, and adding noise according to the forward diffusion process. We show that this creates approximate samples from corresponding marginal distributions. As the noise gradually anneals to zero, the time marginal " + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_t \\mid \\mathbf{y})" + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "content": " smoothly converges to the posterior distribution " + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid \\mathbf{y})" + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "content": ", providing samples that approximately solve the Bayesian inverse problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "text", + "content": "Empirically, our method demonstrates significantly improved performance on various inverse problems compared to existing approaches, ranging from image restoration to compressed sensing MRI. Our method can be combined with diffusion models in both raw pixel space and learned latent space, which we refer to as DAPS and LatentDAPS, respectively. As shown in Fig. 1, both methods provide superior reconstructions with improved visual perceptual quality across a wide range of nonlinear inverse problems. Our approach exhibits remarkable stability and sampling quality, particularly for challenging nonlinear inverse problems. On the FFHQ 256 dataset, we achieve PSNR values of " + }, + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "inline_equation", + "content": "30.72\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "inline_equation", + "content": "27.12\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "text", + "content": " for the phase retrieval and high dynamic range tasks, representing improvements of " + }, + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "inline_equation", + "content": "1.98\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "inline_equation", + "content": "4.39\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "text", + "content": " over the previous state-of-the-art results, respectively. We apply DAPS to large-scale text-conditioned latent diffusion models, showing its scalability to high-resolution image restoration tasks. Beyond image restoration tasks, DAPS achieves a PSNR of " + }, + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "inline_equation", + "content": "31.49\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "text", + "content": ", showing a " + }, + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "inline_equation", + "content": "2.70\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 241, + 556, + 540 + ], + "type": "text", + "content": " improvement over previous diffusion-based methods. We also demonstrate how DAPS can be applied to inverse problems on categorical data with pre-trained discrete diffusion models. Additionally, DAPS performs well with a small number of neural network evaluations (approximately 100), striking a better balance between efficiency and sample quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 555, + 392, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 555, + 392, + 567 + ], + "spans": [ + { + "bbox": [ + 314, + 555, + 392, + 567 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 575, + 417, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 575, + 417, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 575, + 417, + 586 + ], + "type": "text", + "content": "2.1. Diffusion Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Diffusion models [21, 26, 47-49] generate data by reversing a predefined noising process. Let the data distribution be " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ". We can define a series of noisy data distributions " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x};\\sigma)" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " by adding Gaussian noise with a standard deviation of " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " to the data. These form the time-margins of a stochastic differential equation (SDE) [26], given by " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{d}\\mathbf{x}_t = \\sqrt{2\\dot{\\sigma}_t\\sigma_t}\\mathrm{d}\\mathbf{w}_t" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " is the predefined noise schedule with " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_0 = 0" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_T = \\sigma_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\dot{\\sigma}_{t}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " is the time derivative of " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_t" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " is a standard Wiener process. We use " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " interchangeably with " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\sigma_t}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ". For a sufficiently large " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20896" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 292, + 215 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 292, + 215 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 292, + 215 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 292, + 215 + ], + "type": "image", + "image_path": "74776871326b0fdad20d79ec6020c5589d7bf42dac02bfa177d58618b1d8eff0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 224, + 296, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 301 + ], + "type": "text", + "content": "Figure 2. An illustration of our method on phase retrieval. Given a noisy sample " + }, + { + "bbox": [ + 55, + 224, + 296, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 224, + 296, + 301 + ], + "type": "text", + "content": ", we first solve the reverse diffusion process to obtain " + }, + { + "bbox": [ + 55, + 224, + 296, + 301 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 224, + 296, + 301 + ], + "type": "text", + "content": ". Using this, we construct an approximation for " + }, + { + "bbox": [ + 55, + 224, + 296, + 301 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y})" + }, + { + "bbox": [ + 55, + 224, + 296, + 301 + ], + "type": "text", + "content": ". Next, we sample " + }, + { + "bbox": [ + 55, + 224, + 296, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0|y} \\sim p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y})" + }, + { + "bbox": [ + 55, + 224, + 296, + 301 + ], + "type": "text", + "content": " with multiple steps of Langevin dynamics, then perturb it with the forward diffusion process to obtain a sample with slightly less noise. We repeat this process until the noise reduces to zero." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 311, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 335 + ], + "type": "text", + "content": "distribution " + }, + { + "bbox": [ + 55, + 311, + 296, + 335 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x};\\sigma_{\\mathrm{max}})" + }, + { + "bbox": [ + 55, + 311, + 296, + 335 + ], + "type": "text", + "content": " converges to pure Gaussian noise " + }, + { + "bbox": [ + 55, + 311, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\sigma_{\\mathrm{max}}^2\\mathbf{I})" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "text", + "content": "To sample from data distribution " + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "text", + "content": ", we first draw an initial sample from " + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\sigma_{\\mathrm{max}}^2\\mathbf{I})" + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "text", + "content": ", then solve the reverse SDE: " + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "inline_equation", + "content": "\\mathrm{d}\\mathbf{x}_t = -2\\dot{\\sigma}_t\\sigma_t\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t;\\sigma_t)\\mathrm{d}t + \\sqrt{2\\dot{\\sigma}_t\\sigma_t}\\mathrm{d}\\mathbf{w}_t" + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t;\\sigma_t)" + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "text", + "content": " is the time-dependent score function [47, 49]. Here " + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t;\\sigma_t)" + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "text", + "content": " is unknown, but can be approximated by training a diffusion model " + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "inline_equation", + "content": "s_\\theta (\\mathbf{x}_t,\\sigma_t)" + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "inline_equation", + "content": "s_\\theta (\\mathbf{x}_t,\\sigma_t)\\approx \\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t;\\sigma_t)" + }, + { + "bbox": [ + 55, + 335, + 296, + 420 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 274, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 274, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 274, + 438 + ], + "type": "text", + "content": "2.2. Bayesian Inverse Problems with Diffusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "content": "Inverse problems aim to recover data from partial, potentially noisy measurements. Formally, solving an inverse problem involves finding the inversion to a forward model that describes the measurement process. In general, a forward model takes the form of " + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = \\mathcal{A}(\\mathbf{x}_0) + \\mathbf{n}" + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "content": " is the measurement function, " + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "content": " represents the original data, " + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "content": " is the observed measurement, and " + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "content": " symbolizes the noise in the measurement process, often modeled as " + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{n} \\sim \\mathcal{N}(\\mathbf{0}, \\beta_{\\mathbf{y}}^2 \\mathbf{I})" + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "content": ". In a Bayesian framework, " + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "content": " comes from the posterior distribution " + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid \\mathbf{y}) \\propto p(\\mathbf{x}_0) p(\\mathbf{y} \\mid \\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "content": " is a prior distribution that can be estimated from a given dataset, and " + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y} \\mid \\mathbf{x}_0) = \\mathcal{N}(\\mathcal{A}(\\mathbf{x}_0), \\beta_{\\mathbf{y}}^2 \\mathbf{I})" + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "content": " models the noisy measurement process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 598, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 296, + 635 + ], + "type": "text", + "content": "When the prior " + }, + { + "bbox": [ + 55, + 598, + 296, + 635 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 598, + 296, + 635 + ], + "type": "text", + "content": " is modeled by a pre-trained diffusion model, we can modify reverse SDE to approximately sample from the posterior distribution following Bayes' rule," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 639, + 296, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 639, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 69, + 639, + 296, + 670 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {d} \\mathbf {x} _ {t} = - 2 \\dot {\\sigma} _ {t} \\sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}; \\sigma_ {t}) \\mathrm {d} t \\\\ - 2 \\dot {\\sigma} _ {t} \\sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {y} \\mid \\mathbf {x} _ {t}) d t + \\sqrt {2 \\dot {\\sigma} _ {t} \\sigma_ {t}} d w _ {t}. \\tag {1} \\\\ \\end{array}", + "image_path": "d713f6ab1fffee57e91f895f06421428a462bfa7b22413bc9c51e105930202cb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "type": "text", + "content": "Here, the noisy likelihood " + }, + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{y}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "type": "text", + "content": " is generally intractable. Multiple methods have been proposed to estimate the noisy likelihood [5, 8, 12, 36, 41, 46, 50, 54, 59]." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 335, + 72, + 531, + 133 + ], + "blocks": [ + { + "bbox": [ + 335, + 72, + 531, + 133 + ], + "lines": [ + { + "bbox": [ + 335, + 72, + 531, + 133 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 531, + 133 + ], + "type": "image", + "image_path": "6cb7dc6f884d697e6f9679598e667577db9a753d0960fd74bde467fab905f146.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 334, + 148, + 533, + 210 + ], + "blocks": [ + { + "bbox": [ + 375, + 137, + 493, + 147 + ], + "lines": [ + { + "bbox": [ + 375, + 137, + 493, + 147 + ], + "spans": [ + { + "bbox": [ + 375, + 137, + 493, + 147 + ], + "type": "text", + "content": "(a) Diffusion models in pixel spaces." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 148, + 533, + 210 + ], + "lines": [ + { + "bbox": [ + 334, + 148, + 533, + 210 + ], + "spans": [ + { + "bbox": [ + 334, + 148, + 533, + 210 + ], + "type": "image", + "image_path": "3ffbaebd72a20ddd23ee5eb974a9822a9e2197f5d8c966a6cae3f669c1e6ff12.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 213, + 480, + 223 + ], + "lines": [ + { + "bbox": [ + 389, + 213, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 389, + 213, + 480, + 223 + ], + "type": "text", + "content": "(b) Latent diffusion models." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 232, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 277 + ], + "type": "text", + "content": "Figure 3. Probabilistic graphical models of the decoupled noise annealing process in (a) pixel and (b) latent spaces respectively. Here " + }, + { + "bbox": [ + 313, + 232, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t \\sim p(\\mathbf{x}_t; \\sigma_t)" + }, + { + "bbox": [ + 313, + 232, + 555, + 277 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 232, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 232, + 555, + 277 + ], + "type": "text", + "content": " is the observed measurement. We factorize " + }, + { + "bbox": [ + 313, + 232, + 555, + 277 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_t \\mid \\mathbf{y})" + }, + { + "bbox": [ + 313, + 232, + 555, + 277 + ], + "type": "text", + "content": " based on these probabilistic graphs." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 285, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 380 + ], + "type": "text", + "content": "One predominant approach is the DPS algorithm [13], which estimates " + }, + { + "bbox": [ + 313, + 285, + 555, + 380 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y} \\mid \\mathbf{x}_t) \\approx p(\\mathbf{y} \\mid \\mathbf{x}_0 = \\mathbb{E}[\\mathbf{x}_0 \\mid \\mathbf{x}_t])" + }, + { + "bbox": [ + 313, + 285, + 555, + 380 + ], + "type": "text", + "content": ". Another line of work [7, 18, 29, 30, 54] solves linear inverse problems by running the reverse diffusion in the spectral domain via singular value decomposition (SVD). Other methods bypass direct computation of this likelihood by interleaving optimization [2, 9, 24, 31, 45, 56, 59] or projection [10, 23, 30, 54] steps with normal diffusion sampling steps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": "Despite promising empirical success, we find that this line of approaches faces challenges in solving more difficult inverse problems when the forward model is highly nonlinear. Accurately solving the reverse SDE in Eq. (1) requires the solver to take a very small step size " + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "\\Delta t > 0" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": ", causing " + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t + \\Delta t}" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": " to be very close to each other. Consequently, " + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": " can only correct minor errors in " + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t + \\Delta t}" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": ", but oftentimes fails to address larger, global errors that require substantial changes to " + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t + \\Delta t}" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": ". One such failure case is given in Fig. 4." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 498, + 370, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 370, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 370, + 510 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 517, + 531, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 531, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 531, + 531 + ], + "type": "text", + "content": "3.1. Decoupled Annealing Posterior Sampling" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "text", + "content": "Instead of solving the reverse-time SDE in Eq. (1), we propose a new noise annealing process that reduces the dependency between samples at consecutive time steps, as illustrated in Figs. 2, 3a and 3b. Unlike previous methods, we ensure " + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t + \\Delta t}" + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "text", + "content": " are conditionally independent given " + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "text", + "content": ". To generate sample " + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t + \\Delta t}" + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "text", + "content": ", we follow a two-step procedure: (1) sampling " + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0|\\mathbf{y}} \\sim p(\\mathbf{x}_0 \\mid \\mathbf{x}_{t + \\Delta t}, \\mathbf{y})" + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "text", + "content": ", and (2) sampling " + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t \\sim \\mathcal{N}(\\mathbf{x}_{0|\\mathbf{y}}, \\sigma_t^2 \\mathbf{I})" + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "text", + "content": ". We repeat this process, gradually reducing noise until " + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 535, + 556, + 666 + ], + "type": "text", + "content": " is sampled. We call this process decoupled noise annealing, which is justified by the proposition below." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 669, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 555, + 693 + ], + "type": "text", + "content": "Proposition 1. Suppose " + }, + { + "bbox": [ + 313, + 669, + 555, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t_1}" + }, + { + "bbox": [ + 313, + 669, + 555, + 693 + ], + "type": "text", + "content": " is sampled from the time-marginal " + }, + { + "bbox": [ + 313, + 669, + 555, + 693 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_{t_1} \\mid \\mathbf{y})" + }, + { + "bbox": [ + 313, + 669, + 555, + 693 + ], + "type": "text", + "content": ", then" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 358, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 358, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t _ {2}} \\sim \\mathbb {E} _ {\\mathbf {x} _ {0} \\sim p (\\mathbf {x} _ {0} | \\mathbf {x} _ {t _ {1}}, \\mathbf {y})} [ \\mathcal {N} (\\mathbf {x} _ {0}, \\sigma_ {t _ {2}} ^ {2} I) ] \\tag {2}", + "image_path": "eaa714ddf7fa3ad846b9be7a215926875a4be0f5c09ec88b7177ceb1bafc07e2.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20897" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 209, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 209, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 209, + 85 + ], + "type": "text", + "content": "satisfies the time-marginal " + }, + { + "bbox": [ + 55, + 72, + 209, + 85 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_{t_2} \\mid \\mathbf{y})" + }, + { + "bbox": [ + 55, + 72, + 209, + 85 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "Remark. The key idea of Proposition 1 is to enable the sampling from " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_{t_2} \\mid \\mathbf{y})" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": " for any noise level " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\sigma_{t_2}" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": " given any sample " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "x_{t_1}" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": " at another noise level " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\sigma_{t_1}" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": ". For a sufficiently large " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\sigma_T" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": ", one can assume " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_T \\mid \\mathbf{y}) \\approx p(\\mathbf{x}_T; \\sigma_T) \\approx \\mathcal{N}(\\mathbf{0}, \\sigma_T^2\\mathbf{I})" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": ". Starting from " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": ", we can iteratively sample from " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_t \\mid \\mathbf{y})" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": " annealed down from " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\sigma_T" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": " to 0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "text", + "content": "The first step of our decoupled noise annealing requires sampling " + }, + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0|\\mathbf{y}}\\sim p(\\mathbf{x}_0\\mid \\mathbf{x}_t,\\mathbf{y})" + }, + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "text", + "content": " are known. Since " + }, + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "text", + "content": " is conditionally independent from " + }, + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 164, + 296, + 212 + ], + "type": "text", + "content": ", we can deduce from the Bayes' rule that" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 219, + 295, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 219, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 92, + 219, + 295, + 262 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}, \\mathbf {y}\\right) = \\frac {p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}\\right) p \\left(\\mathbf {y} \\mid \\mathbf {x} _ {0} , \\mathbf {x} _ {t}\\right)}{p \\left(\\mathbf {y} \\mid \\mathbf {x} _ {t}\\right)} \\\\ \\propto p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}\\right) p (\\mathbf {y} \\mid \\mathbf {x} _ {0}). \\tag {3} \\\\ \\end{array}", + "image_path": "5c3ee352fdf1f02365617d31e127de451d10bbc48523828a5283aa03e521c553.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "text", + "content": "To sample " + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0|\\mathbf{y}}" + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "text", + "content": " from this unnormalized distribution, MCMC methods such as Langevin dynamics [55] and Hamiltonian Monte Carlo [4] can be adopted (as explained in Appendix A.1). The updating rule of Langevin dynamics [55] is given by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 333, + 295, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 333, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 86, + 333, + 295, + 373 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {x} _ {0} ^ {(j + 1)} = \\mathbf {x} _ {0} ^ {(j)} + \\eta \\nabla_ {\\mathbf {x} _ {0} ^ {(j)}} \\log p (\\mathbf {x} _ {0} ^ {(j)} \\mid \\mathbf {x} _ {t}) \\\\ + \\eta \\nabla_ {\\mathbf {x} _ {0} ^ {(j)}} \\log p (\\mathbf {y} \\mid \\mathbf {x} _ {0} ^ {(j)}) + \\sqrt {2 \\eta} \\boldsymbol {\\epsilon} _ {j}, \\tag {4} \\\\ \\end{array}", + "image_path": "c14f504748e98bf943a8562ce9fb404983d5939a873dcf89568de0243d2c537a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "inline_equation", + "content": "\\eta > 0" + }, + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "text", + "content": " is the step size and " + }, + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "inline_equation", + "content": "\\epsilon_{j} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "inline_equation", + "content": "\\eta \\to 0" + }, + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "inline_equation", + "content": "j \\to \\infty" + }, + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "text", + "content": ", the sample " + }, + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{(j)}" + }, + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "text", + "content": " will be approximately distributed according to " + }, + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y})" + }, + { + "bbox": [ + 55, + 380, + 296, + 421 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 426, + 233, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 233, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 233, + 440 + ], + "type": "text", + "content": "3.2. Practical Design Choice of DAPS" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "text", + "content": "Following Eq. (4), we discuss different ways to approximate the conditional distribution " + }, + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid \\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "text", + "content": " due to its intractability. A line of previous works on diffusion posterior sampling [5, 13, 46] approximate " + }, + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid \\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "text", + "content": " by a Gaussian distribution," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 501, + 295, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 501, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 104, + 501, + 295, + 515 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}\\right) \\approx \\mathcal {N} \\left(\\mathbf {x} _ {0}; \\hat {\\mathbf {x}} _ {0} \\left(\\mathbf {x} _ {t}\\right), r _ {t} ^ {2} \\boldsymbol {I}\\right), \\tag {5}", + "image_path": "05529d58c4f89693bd8383cc18f779ff367ddf1d58828e567b7d3b14a6463243.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 523, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 296, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 523, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 523, + 296, + 558 + ], + "type": "text", + "content": " is an estimator of " + }, + { + "bbox": [ + 55, + 523, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 523, + 296, + 558 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 55, + 523, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 523, + 296, + 558 + ], + "type": "text", + "content": " typically specified as " + }, + { + "bbox": [ + 55, + 523, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\mathbf{x}_0\\mid \\mathbf{x}_t]" + }, + { + "bbox": [ + 55, + 523, + 296, + 558 + ], + "type": "text", + "content": ", and the variance " + }, + { + "bbox": [ + 55, + 523, + 296, + 558 + ], + "type": "inline_equation", + "content": "r_t^2" + }, + { + "bbox": [ + 55, + 523, + 296, + 558 + ], + "type": "text", + "content": " is specified using heuristics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 559, + 296, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 607 + ], + "type": "text", + "content": "Another possible way to estimate " + }, + { + "bbox": [ + 55, + 559, + 296, + 607 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid \\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 559, + 296, + 607 + ], + "type": "text", + "content": " is to decompose it via Bayes' rule, and substitute the intractable " + }, + { + "bbox": [ + 55, + 559, + 296, + 607 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 559, + 296, + 607 + ], + "type": "text", + "content": " by the distribution " + }, + { + "bbox": [ + 55, + 559, + 296, + 607 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0; \\sigma_{t_{\\min}})" + }, + { + "bbox": [ + 55, + 559, + 296, + 607 + ], + "type": "text", + "content": " with a small Gaussian noise for numerical stability, i.e.," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 616, + 295, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 616, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 58, + 616, + 295, + 644 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\mathbf {x} _ {0}} \\log p (\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t}) = \\nabla_ {\\mathbf {x} _ {0}} \\log p (\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}) + \\nabla_ {\\mathbf {x} _ {0}} \\log p (\\mathbf {x} _ {0}) \\\\ \\approx \\nabla_ {\\mathbf {x} _ {0}} \\log p \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) + s _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {0}, t _ {\\min }\\right). \\tag {6} \\\\ \\end{array}", + "image_path": "aa16fed70ce1a51d85fb13dc87f96c077a670884f5bc23531d5e8de8f0e1356e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Despite being a more accurate estimation, the diffusion-score estimation Eq. (6) is much more expensive than using a Gaussian approximation when applied to Eq. (4), as it requires calling the score function multiple times. Empirical studies in Sec. 4.4 show that using a Gaussian approximation can" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "content": "achieve comparable results as the diffusion-score estimation, but is significantly more time-efficient. In practice, we adopt the Gaussian approximation and compute " + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "content": " by solving the (unconditional) probability flow ODE starting at " + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "content": ". We leave details in Appendix F.2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 133, + 555, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 157 + ], + "type": "text", + "content": "When the measurement noise is an isotropic Gaussian, i.e., " + }, + { + "bbox": [ + 313, + 133, + 555, + 157 + ], + "type": "inline_equation", + "content": "\\mathbf{n} \\sim \\mathcal{N}(\\mathbf{0}, \\beta_{\\mathbf{y}}^{2}\\mathbf{I})" + }, + { + "bbox": [ + 313, + 133, + 555, + 157 + ], + "type": "text", + "content": ", the update rule simplifies to:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 343, + 163, + 555, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 163, + 555, + 223 + ], + "spans": [ + { + "bbox": [ + 343, + 163, + 555, + 223 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {x} _ {0} ^ {(j + 1)} = \\mathbf {x} _ {0} ^ {(j)} - \\eta \\nabla_ {\\mathbf {x} _ {0} ^ {(j)}} \\frac {\\left\\| \\mathbf {x} _ {0} ^ {(j)} - \\hat {\\mathbf {x}} _ {0} (\\mathbf {x} _ {t}) \\right\\| ^ {2}}{2 r _ {t} ^ {2}} \\\\ - \\eta \\nabla_ {\\mathbf {x} _ {0} ^ {(j)}} \\frac {\\| \\mathcal {A} \\left(\\mathbf {x} _ {0} ^ {(j)}\\right) - \\mathbf {y} \\| ^ {2}}{2 \\beta_ {\\mathbf {y}} ^ {2}} + \\sqrt {2 \\eta} \\boldsymbol {\\epsilon} _ {j}, \\tag {7} \\\\ \\end{array}", + "image_path": "b49ca9e820435523a2dba9e790e50b9b7bc4303eaebd5014c7bf0208807db66a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "text", + "content": "Combining Eq. (7) with Proposition 1, we implement Decoupled Annealing Posterior Sampling (DAPS) as Algorithm 1 in Appendix A. In particular, given a noise schedule " + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "text", + "content": " and a time discretization " + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\{t_i,i = 0,\\dots ,N_A\\}" + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "text", + "content": ", we iteratively sample " + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t_i}" + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "text", + "content": " from the measurement-conditioned time-marginal " + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_{t_i}|\\mathbf{y})" + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "inline_equation", + "content": "i = N_{A},N_{A} - 1,\\ldots ,0" + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "text", + "content": ", following Eq. (2) in Proposition 1. This algorithm provides an approximate sample " + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "text", + "content": " from the posterior distribution " + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0|\\mathbf{y})" + }, + { + "bbox": [ + 313, + 229, + 555, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 326, + 554, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 554, + 397 + ], + "type": "text", + "content": "The computational cost of Langevin dynamics mostly originates from evaluating the measurement function " + }, + { + "bbox": [ + 313, + 326, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 326, + 554, + 397 + ], + "type": "text", + "content": ". In most image restoration tasks, this operation is significantly more efficient compared to evaluating the diffusion model. As demonstrated in Appendix E.1, the MCMC sampling introduces only a small overhead to the sampling process." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 403, + 501, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 501, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 501, + 415 + ], + "type": "text", + "content": "3.3. DAPS with Latent Diffusion Model" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "text", + "content": "Given a pretrained encoder " + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "text", + "content": ", the latent diffusion models (LDMs) [40] are trained to fit the latent distribution " + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z}_0)" + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "text", + "content": " is given by the encoder " + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 = \\mathcal{E}(\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "text", + "content": ". One can recover " + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "text", + "content": " from the decoder " + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 = \\mathcal{D}(\\mathbf{z}_0)" + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "text", + "content": ". We found the proposed method can be adapted to sampling with pre-trained latent diffusion models by factorizing the probabilistic graphical model for the latent diffusion process, as shown in Fig. 3b. We refer to this as LatentDAPS in the following paper." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 529, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 552 + ], + "type": "text", + "content": "Accordingly, we can derive the following Langevin dynamics updating rule from Eq. (4):" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 340, + 553, + 555, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 553, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 340, + 553, + 555, + 590 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {z} _ {0} ^ {(j + 1)} = \\mathbf {z} _ {0} ^ {(j)} + \\eta \\nabla_ {\\mathbf {z} _ {0} ^ {(j)}} \\log p (\\mathbf {z} _ {0} ^ {(j)} | \\mathbf {z} _ {t}) \\\\ + \\eta \\nabla_ {\\mathbf {z} _ {0} ^ {(j)}} \\log p (\\mathbf {y} \\mid \\mathcal {D} \\left(\\mathbf {z} _ {0} ^ {(j)}\\right)) + \\sqrt {2 \\eta} \\epsilon_ {j}. \\tag {8} \\\\ \\end{array}", + "image_path": "52c2b7c714a37537108bcb77635aedb4aa121346c0cd2f4d3705558057e2e638.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 594, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 665 + ], + "type": "text", + "content": "While LatentDAPS is a straightforward and natural extension of DAPS compared to recent latent diffusion inverse problem solvers [42, 43, 45], it demonstrates comparable or even superior performance across several tasks in Tab. 1 and Tab. 4. More detailed derivation and proofs are shown in Appendix B." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 672, + 494, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 494, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 494, + 685 + ], + "type": "text", + "content": "3.4. Discussion with Existing Methods" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Comparison with other posterior sampling methods. Unlike most previous algorithms, our sampling method does not" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20898" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 73, + 282, + 210 + ], + "blocks": [ + { + "bbox": [ + 70, + 73, + 282, + 210 + ], + "lines": [ + { + "bbox": [ + 70, + 73, + 282, + 210 + ], + "spans": [ + { + "bbox": [ + 70, + 73, + 282, + 210 + ], + "type": "image", + "image_path": "d53d44f9f84ecef58cdf9639c8b2983023f8e2a8adc1fc6d785f6bf5d7c6410e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 219, + 297, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 219, + 297, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 297, + 253 + ], + "type": "text", + "content": "Figure 4. DAPS vs. DPS on 2D synthetic data. Consecutive sampling steps are close to each other for DPS but not for DAPS. Here DAPS approximates the posterior better than DPS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "content": "solve a specific SDE/ODE; instead, we recursively sample from the time-marginal " + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_t \\mid \\mathbf{y})" + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "content": " with noise annealing to zero. We therefore decouple the dependency of " + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t + \\Delta t}" + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "content": " in the sampling process. We argue that decoupling helps correct the errors accumulated in early diffusion sampling steps by allowing non-local transitions. This is particularly important when the measurement function is nonlinear. As a concrete example, Fig. 4 compares DAPS and DPS when solving a 2D nonlinear inverse problem with a Gaussian mixture prior. We visualize the trajectories of " + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "content": " from time " + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "content": " to 0 for both methods. For DAPS, points on the trajectory have significantly larger variations compared to those of DPS. As a result, DPS converges to wrong solutions, but DAPS is able to approximate the true posterior distribution. We provide more discussions in Appendix G." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 446, + 298, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 298, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 298, + 590 + ], + "type": "text", + "content": "Comparison with optimization-based methods. Although our method does not directly involve inner-loop optimization, we highlight that it connects with existing optimization-based solvers for inverse problems. For example, ReSample [45], DiffPIR [59], and DCDP [31] alternate between denoising optimizing, and resampling to solve inverse problems. However, these methods solve the MAP estimation problems, while DAPS draws samples from the posterior distribution. In particular, instantiating DAPS with Langevin dynamics and Gaussian approximation resembles [45] if we set standard deviation " + }, + { + "bbox": [ + 55, + 446, + 298, + 590 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{y}}\\rightarrow 0" + }, + { + "bbox": [ + 55, + 446, + 298, + 590 + ], + "type": "text", + "content": " and assume " + }, + { + "bbox": [ + 55, + 446, + 298, + 590 + ], + "type": "inline_equation", + "content": "\\eta \\beta_{\\mathbf{y}}^{2}" + }, + { + "bbox": [ + 55, + 446, + 298, + 590 + ], + "type": "text", + "content": " is constant in Eq. (18)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 602, + 137, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 137, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 137, + 616 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 623, + 172, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 172, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 172, + 636 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "We evaluate our method using both pixel-space and latent diffusion models. For pixel-based diffusion experiments, we leverage the pre-trained diffusion models trained by [13] on the FFHQ dataset and the pre-trained model from [16] on the ImageNet dataset. For latent diffusion models, we use the same pre-trained models as [45]: the unconditional" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "LDM-VQ4 trained on FFHQ and ImageNet by [40]. The autoencoder's downsample factor is 4. We use the same time step discretization and noise schedule as EDM [26]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 110, + 556, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 556, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 556, + 302 + ], + "type": "text", + "content": "As mentioned in Sec. 3, we implement a few-step Euler ODE solver to compute " + }, + { + "bbox": [ + 313, + 110, + 556, + 302 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 110, + 556, + 302 + ], + "type": "text", + "content": ", while maintaining the same number of neural function evaluations (NFE) for different noise levels. In our experiments, we use DAPS-1K for all linear tasks and DAPS-4K for all nonlinear tasks. DAPS-1K uses 4 ODE solver NFE and 250 annealing scheduling steps, while DAPS-4K uses 10 NFE and 400 steps, respectively. We discuss the effect of choosing these configurations in Sec. 4.4. The corresponding latent version, LatentDAPS, follows the same settings but performs sampling in the latent space of VAEs. We use 100 and 50 Langevin steps per denoising iteration for DAPS and LatentDAPS respectively, and tune learning rates separately for each task. Further details on model configurations, samplers, and other hyperparameters are provided in Appendix F, along with more discussion on sampling efficiency in Sec. 4.2 and Appendix E.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 304, + 556, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 556, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 556, + 445 + ], + "type": "text", + "content": "Datasets and metrics. Adopting the previous convention, we test our method on two image datasets, FFHQ " + }, + { + "bbox": [ + 313, + 304, + 556, + 445 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 304, + 556, + 445 + ], + "type": "text", + "content": " [25] and ImageNet " + }, + { + "bbox": [ + 313, + 304, + 556, + 445 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 304, + 556, + 445 + ], + "type": "text", + "content": " [15]. To evaluate our method, we use 100 images from the validation set for both FFHQ and ImageNet. We include peak signal-to-noise-ratio (PSNR), structural similarity index measure (SSIM), Learned Perceptual Image Patch Similarity (LPIPS)[58] score and Fréchet inception distance (FID)[20] as our main evaluation metrics. For both our method and baselines, we use the versions implemented in pic[28] with all images normalized to the range [0, 1]. The replace-pooling option is enabled for LPIPS evaluation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "type": "text", + "content": "Inverse problems. We evaluate our method with a series of linear and nonlinear tasks. For linear inverse problems, we consider (1) super-resolution, (2) Gaussian deblurring, (3) motion deblurring, (4) inpainting (with a box mask), and (5) inpainting (with a " + }, + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "type": "text", + "content": " random mask). For Gaussian and motion deblurring, kernels of size " + }, + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "type": "inline_equation", + "content": "61 \\times 61" + }, + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "type": "text", + "content": " with standard deviations of 3.0 and 0.5, respectively, are used. In the super-resolution task, a bicubic resizer downscales images by a factor of 4. The box inpainting task uses a random box of size " + }, + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "type": "text", + "content": " to mask the original images, while random mask inpainting uses a generated random mask where each pixel has a " + }, + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 449, + 556, + 604 + ], + "type": "text", + "content": " chance of being masked, following the settings in [45]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": "We consider three nonlinear inverse problems: (1) phase retrieval, (2) high dynamic range (HDR) reconstruction, and (3) nonlinear deblurring. Due to the inherent instability of phase retrieval, we adopt the strategy from DPS [13], using an oversampling rate of 2.0 and reporting the best result out of four independent samples. The goal of HDR reconstruction is to recover a higher dynamic range image (factor of 2) from a low dynamic range image. For nonlinear deblurring, we use the default setting as described in [53]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20899" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "content": "All linear and nonlinear measurements are subject to white Gaussian noise with a standard deviation of " + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{y}} = 0.05" + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "content": ". Further details regarding the forward measuring functions for each task and their respective hyperparameters are provided in the Appendix F." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 296, + 240 + ], + "type": "text", + "content": "Baselines. We compare our methods with the following baselines: DDRM [30], DPS [13], DDNM [54], DCDP [31], FPS-SMC [18], DiffPIR [59] for pixel-based diffusion model experiments. We compare the latent diffusion version of our methods with PSLD [42] and ReSample [45], which also operate in the latent space. Note that DDRM, DDNM, FPS-SMC, and PSLD cannot handle nonlinear inverse problems. We especially RED-diff [35] for nonlinear experiments and a traditional method HIO [14] for phase retrieval." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 140, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 140, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 140, + 262 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 269, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 472 + ], + "type": "text", + "content": "We show quantitative results for FFHQ and ImageNet dataset in Tab. 1. The qualitative comparisons are provided in Fig. 1. As shown in Tab. 1, our method achieves comparable or even better performance across all selected linear inverse problems. Moreover, our methods are remarkably stable when handling nonlinear inverse problems. For example, although existing methods such as DPS can recover high-quality images from phase retrieval measurements, they suffer from an extremely high failure rate, resulting in a relatively low PSNR and a high LPIPS. However, DAPS demonstrates superior stability across different samples and achieves a significantly higher success rate, resulting in a substantial improvement in PSNR and LPIPS as evidenced by the results in Tab. 1. Moreover, DAPS captures and recovers much more fine-grained details in measurements compared to existing baselines. We include more samples and comparisons in Appendix I for further illustration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "text", + "content": "To understand the evolution of the noise annealing procedure, we evaluate two crucial trajectories in our method: 1) the estimated mean of " + }, + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid \\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "text", + "content": "; and 2) samples from " + }, + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y})" + }, + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "text", + "content": " obtained through Langevin dynamics, denoted as " + }, + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0|\\mathbf{y}}" + }, + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "text", + "content": ". We assess image quality and measurement error across these trajectories, as depicted in Fig. 7. The measurement error for " + }, + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0|\\mathbf{y}} \\sim p(\\mathbf{x}_0 \\mid \\mathbf{x}_t, \\mathbf{y})" + }, + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "text", + "content": " remains consistently low throughout the process, while the measurement error for " + }, + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 473, + 296, + 604 + ], + "type": "text", + "content": " continuously decreases. Consequently, metrics such as PSNR and LPIPS exhibit stable improvement as the noise level is annealed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "In addition, our method can generate diversified samples when the measurements contain less information. To illustrate this, we selected two tasks where the posterior distributions may have multiple modes: (1) super-resolution by a factor of 16, and (2) box inpainting with a box size of " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "192 \\times 192" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ". Some samples from DAPS are shown in Fig. 5, demonstrating DAPS's ability to produce diverse samples while preserving the measurement information. We show more samples in Appendix I for further demonstration." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 566, + 182 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 566, + 182 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 566, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 566, + 182 + ], + "type": "image", + "image_path": "f6078223ab458b1e78da54664d9af0d9574b4c2e03574d3252341b6863a28a26.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 182, + 495, + 191 + ], + "lines": [ + { + "bbox": [ + 385, + 182, + 495, + 191 + ], + "spans": [ + { + "bbox": [ + 385, + 182, + 495, + 191 + ], + "type": "text", + "content": "(a) Super resolution of a factor 16" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 193, + 566, + 303 + ], + "blocks": [ + { + "bbox": [ + 316, + 193, + 566, + 303 + ], + "lines": [ + { + "bbox": [ + 316, + 193, + 566, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 193, + 566, + 303 + ], + "type": "image", + "image_path": "534690a36ae5de8fa48795aa0c5e1b381c4df00b5bda5e277e80f05e89c80af7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 304, + 490, + 313 + ], + "lines": [ + { + "bbox": [ + 389, + 304, + 490, + 313 + ], + "spans": [ + { + "bbox": [ + 389, + 304, + 490, + 313 + ], + "type": "text", + "content": "(b) Inpainting of " + }, + { + "bbox": [ + 389, + 304, + 490, + 313 + ], + "type": "inline_equation", + "content": "192\\times 192" + }, + { + "bbox": [ + 389, + 304, + 490, + 313 + ], + "type": "text", + "content": " box" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 323, + 555, + 378 + ], + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 378 + ], + "type": "text", + "content": "Figure 5. Sample diversity. We present several diverse samples generated by the DAPS under two sparse measurements whose posterior distributions contain multiple modes. DAPS produces a variety of samples with distinct features, including differences in expression, wearings, and hairstyles." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 389, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 509 + ], + "type": "text", + "content": "DAPS with large-scale text-conditioned latent diffusion. Due to the flexibility of DAPS, pretrained large-scale text-conditioned latent diffusion models (LDMs) [38-40] can also serve as the prior model " + }, + { + "bbox": [ + 313, + 389, + 555, + 509 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid \\mathbf{c})" + }, + { + "bbox": [ + 313, + 389, + 555, + 509 + ], + "type": "text", + "content": " when given a text prompt c. To demonstrate DAPS's potential, we use Stable Diffusion v2.1 [40] to present LatentDAPS results on three tasks with " + }, + { + "bbox": [ + 313, + 389, + 555, + 509 + ], + "type": "inline_equation", + "content": "768 \\times 768" + }, + { + "bbox": [ + 313, + 389, + 555, + 509 + ], + "type": "text", + "content": " resolution images, as shown in Fig. 1. Additionally, we provide the text prompts used, along with a detailed discussion, in Appendix B.1, where we also include a quantitative evaluation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 605 + ], + "type": "text", + "content": "DAPS with discrete diffusion. DAPS can be naturally extended to support posterior sampling with categorical prior distribution modeled by discrete diffusion models [3, 6, 33]. Instead of sampling " + }, + { + "bbox": [ + 313, + 510, + 556, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0|\\mathbf{y}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 605 + ], + "type": "text", + "content": " by Langevin MC or Hamiltonian MC in continuous space, we use the Metropolis-Hasting algorithm to replace Eq. (4). We validate DAPS with discrete diffusion on a quantized MNIST dataset. We leave the details to Appendix C." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 612, + 448, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 448, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 448, + 624 + ], + "type": "text", + "content": "4.3. Further results on MRI" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "text", + "content": "We further validate the effectiveness of DAPS on compressed sensing multi-coil MRI, which is a medical imaging technique that helps reduce the scan time of MRI via subsampling. We train a diffusion model using the pipeline from [26] with fastMRI knee dataset [57]. Tab. 2 shows that DAPS outperforms existing diffusion-based baselines. We include more qualitative results in Appendix H." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20900" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 69, + 523, + 635 + ], + "blocks": [ + { + "bbox": [ + 87, + 69, + 523, + 635 + ], + "lines": [ + { + "bbox": [ + 87, + 69, + 523, + 635 + ], + "spans": [ + { + "bbox": [ + 87, + 69, + 523, + 635 + ], + "type": "table", + "html": "
TaskTypeMethodFFHQImageNet
PSNR (↑)SSIM (↑)LPIPS (↓)FID (↓)PSNR (↑)SSIM (↑)LPIPS (↓)FID (↓)
Super resolution 4xPixelDAPS (ours)29.070.8180.17751.4425.890.6940.27683.57
DPS25.860.7530.26981.0721.130.4890.361106.32
DDRM26.580.7820.28279.2522.620.5210.324103.85
DDNM28.030.7950.19764.6223.960.6040.47598.62
DCDP28.660.8070.17853.81----
FPS-SMC28.420.8130.20449.2524.820.7030.31397.51
DiffPIR26.64-0.26065.7723.18-0.371106.32
LatentLatentDAPS (ours)27.480.8010.18259.6225.060.6730.27684.37
PSLD24.350.6490.28774.3625.420.6940.36097.45
ReSample23.290.5940.39293.1822.610.5760.370113.42
Inpaint (box)PixelDAPS (ours)24.070.8140.13343.1021.430.7250.214109.85
DPS22.510.7920.20961.2718.940.7220.257126.52
DDRM22.260.8010.20778.6218.630.7330.254116.37
DDNM24.470.8370.23546.5921.640.7480.319103.97
DCDP23.890.7600.16345.23----
FPS-SMC24.860.8230.14648.3422.160.7260.208111.58
LatentLatentDAPS (ours)23.990.8020.19446.5217.190.6240.340145.63
PSLD24.220.8130.15843.0220.100.6940.465146.53
ReSample20.060.7490.18453.2118.290.6310.262127.84
Inpaint (random)PixelDAPS (ours)31.120.8440.09832.1728.440.7750.13554.25
DPS25.460.8230.20369.2023.520.7450.29787.53
DDNM29.910.8170.12144.3731.160.8410.19163.84
DCDP30.690.8420.14252.51----
FPS-SMC28.210.8230.26161.2324.520.7010.31679.12
LatentLatentDAPS (ours)30.710.8130.14136.4127.590.7720.16461.62
PSLD30.310.8090.22147.2131.300.7830.33783.21
ReSample29.610.7460.14039.8527.500.7560.14359.87
Gaussian deblurringPixelDAPS (ours)29.190.8170.16553.3326.150.6840.25375.68
DPS25.870.7640.21979.7520.310.5980.397116.42
DDRM24.930.7320.23992.4321.260.5640.443146.89
DDNM28.200.8040.21657.8328.060.7030.27881.43
DCDP27.500.6990.30486.43----
FPS-SMC26.540.7730.25367.4523.910.6010.38791.72
DiffPIR27.36-0.23659.6522.80-0.35593.36
LatentLatentDAPS (ours)27.930.7640.23464.5225.050.6680.34578.51
PSLD23.270.6310.31689.5125.860.6880.39091.39
ReSample26.390.7140.25571.6925.970.7030.25465.35
Motion deblurringPixelDAPS (ours)29.660.8470.15739.4927.860.7660.19661.83
DPS24.520.8010.24665.2318.960.6290.423137.81
DCDP25.080.5120.364125.13----
FPS-SMC27.390.8260.22748.3224.520.6470.32687.43
DiffPIR26.57-0.25565.7824.01-0.36694.63
LatentLatentDAPS (ours)27.000.8140.28346.8426.830.7450.29673.62
PSLD22.310.6780.33696.1520.850.5940.511124.67
ReSample27.410.8230.19844.7226.940.7380.22766.89
Phase retrievalPixelDAPS (ours)30.63±3.130.851±0.0720.139±0.06042.7125.78±6.920.743±0.0840.254±0.12582.67
DPS17.64±2.970.441±0.1290.410±0.090104.5216.81±3.610.427±0.1430.447±0.099197.54
RED-diff15.60±4.480.398±0.1950.596±0.092167.4314.98±3.750.386±0.0570.536±0.129212.24
DCDP28.65±8.090.781±0.2170.203±0.19668.13----
LatentLatentDAPS (ours)29.16±3.550.796±0.0890.199±0.07854.2620.54±6.410.612±0.1140.361±0.150129.54
ReSample21.60±8.100.648±0.1540.406±0.22484.32----
ClassicalHIO13.53±2.500.359±0.0930.726±0.068268.09----
Nonlinear deblurPixelDAPS (ours)28.29±1.770.783±0.0360.155±0.03249.3827.73±3.230.724±0.0480.169±0.05659.87
DPS23.39±2.010.623±0.0820.278±0.06091.3122.49±3.200.591±0.1010.306±0.081101.41
RED-diff30.86±0.510.795±0.0280.160±0.03443.8430.07±1.410.754±0.0230.211±0.08351.22
DCDP27.92±2.640.779±0.0670.183±0.05151.96----
LatentLatentDAPS (ours)28.11±1.750.713±0.0410.235±0.04953.6325.34±3.440.615±0.0570.314±0.08076.73
ReSample28.24±1.690.742±0.0390.185±0.03951.6226.20±3.710.653±0.0640.206±0.05761.16
High dynamic rangePixelDAPS (ours)27.12±3.530.752±0.0410.162±0.07242.9726.30±4.100.717±0.0670.175±0.10764.19
DPS22.73±6.070.591±0.1410.264±0.156112.8219.23±2.520.582±0.0820.503±0.106146.23
RED-diff22.16±3.410.512±0.0830.258±0.089108.3222.03±5.900.601±0.0940.274±0.198113.48
LatentLatentDAPS (ours)25.94±2.870.751±0.0560.223±0.08074.8323.64±4.100.609±0.0530.269±0.09993.51
ReSample25.65±3.570.732±0.0590.182±0.08567.2225.11±4.210.633±0.0490.198±0.08987.66
", + "image_path": "fdb9832836af13edefa37318dc64d413e3ce3b7246c80fba43c6fdd9c857c671.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 643, + 555, + 711 + ], + "lines": [ + { + "bbox": [ + 55, + 643, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 555, + 711 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation on FFHQ and ImageNet on 5 linear and 3 nonlinear tasks. Performance comparison of different methods on FFHQ (left) and ImageNet (right) in the image domain. For linear tasks, we report the mean performance (PSNR, SSIM and LPIPS) across 100 validation images. We provide both the mean and standard deviation for nonlinear tasks, where results exhibit higher instability. The FID scores are evaluated with the same 100 validation images. The best and second-best results within each type of task are indicated by bold and underlined marks, respectively. DAPS demonstrates superior performance on most tasks, with LatentDAPS showing competitive results. All tasks are using noisy measurement with noise level " + }, + { + "bbox": [ + 55, + 643, + 555, + 711 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{y}} = 0.05" + }, + { + "bbox": [ + 55, + 643, + 555, + 711 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20901" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 302, + 180 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 302, + 180 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 302, + 180 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 302, + 180 + ], + "type": "image", + "image_path": "7675a9626a3aa211cda5e21aa15e1652c332ca03d231d43c6811f68b7c50e196.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 185, + 223, + 195 + ], + "lines": [ + { + "bbox": [ + 159, + 185, + 223, + 195 + ], + "spans": [ + { + "bbox": [ + 159, + 185, + 223, + 195 + ], + "type": "text", + "content": "(a) Gaussian deblur" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 529, + 182 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 529, + 182 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 529, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 529, + 182 + ], + "type": "image", + "image_path": "5d319c6ba091b9dfa55da4314901ef3f7835e0615d0612c26b128cd32dcd3555.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 185, + 447, + 194 + ], + "lines": [ + { + "bbox": [ + 386, + 185, + 447, + 194 + ], + "spans": [ + { + "bbox": [ + 386, + 185, + 447, + 194 + ], + "type": "text", + "content": "(b) Phase retrieval" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 236, + 555, + 344 + ], + "blocks": [ + { + "bbox": [ + 55, + 205, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 555, + 228 + ], + "type": "text", + "content": "Figure 6. Evaluation of time cost and sample quality. The x-axis indicates the single image sampling time on an A100-SXM4-80GB GPU and y-axis shows the LPIPS. The evaluation uses 100 FFHQ images." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 236, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 57, + 236, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 57, + 236, + 555, + 344 + ], + "type": "image", + "image_path": "7e725e42bb7e60312e10069a4ceb4a5bc38608cb4b0537da63a2ac32011fd871.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 353, + 555, + 387 + ], + "lines": [ + { + "bbox": [ + 55, + 353, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 555, + 387 + ], + "type": "text", + "content": "Figure 7. Quantitative evaluations of image quality metrics and measurement error throughout the sampling process. Here " + }, + { + "bbox": [ + 55, + 353, + 555, + 387 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0(\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 353, + 555, + 387 + ], + "type": "text", + "content": " stands for the estimated mean of the Gaussian approximated distribution and " + }, + { + "bbox": [ + 55, + 353, + 555, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0|\\mathbf{y}}\\sim p(\\mathbf{x}_0\\mid \\mathbf{x}_t,\\mathbf{y})" + }, + { + "bbox": [ + 55, + 353, + 555, + 387 + ], + "type": "text", + "content": " means the samples from measurement conditioned time-marginal. The measurement errors are normalized between [0, 1] for visualization." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 395, + 296, + 452 + ], + "blocks": [ + { + "bbox": [ + 56, + 395, + 296, + 452 + ], + "lines": [ + { + "bbox": [ + 56, + 395, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 395, + 296, + 452 + ], + "type": "table", + "html": "
×4×8
PSNR (↑)SSIM (↑)Meas err (↓)PSNR (↑)SSIM (↑)Meas err (↓)
DAPS31.480.7621.5729.010.6811.28
DPS26.130.6209.9020.820.5366.74
DiffPIR28.310.63210.5526.780.5887.79
ScoreMRI [10]25.970.46810.8325.010.4058.36
CSGM [23]28.780.7101.5226.150.6251.14
", + "image_path": "9a38a8609aa7ecce679a50f432ccc1ae2d020e5abfbdc188ba27dd36d6c45475.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 504, + 156, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 156, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 156, + 516 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": "Time efficiency of DAPS. We run DAPS with various configurations to test its performance under different computing budgets. Specifically, we evaluate DAPS with the number of function evaluations (NFE) of the diffusion models ranging from 50 to 4K, with configurations specified in Appendix E.1. Sec. 4.2 shows the relation between the sample quality of DAPS measured by LPIPS with its time cost, which indicates that DAPS is able to generate high-quality samples with a reasonable time cost." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Effectiveness of Different Design choices of DAPS. We conduct an ablation study to demonstrate the effectiveness of Gaussian approximation compared with the more principled but more expensive diffusion-score estimation. The results in Tab. 3 demonstrate that Gaussian approximation is " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "6 \\sim 7" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " times faster than diffusion-score estimation with the cost of a slight performance drop, effectively trading off performance" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 395, + 555, + 436 + ], + "blocks": [ + { + "bbox": [ + 55, + 460, + 295, + 494 + ], + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 494 + ], + "type": "text", + "content": "Table 2. Quantitative results of compressed sensing MRI. We consider " + }, + { + "bbox": [ + 55, + 460, + 295, + 494 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 460, + 295, + 494 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 295, + 494 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 55, + 460, + 295, + 494 + ], + "type": "text", + "content": " subsampling ratio. Meas err stands for " + }, + { + "bbox": [ + 55, + 460, + 295, + 494 + ], + "type": "inline_equation", + "content": "{\\ell }_{2}" + }, + { + "bbox": [ + 55, + 460, + 295, + 494 + ], + "type": "text", + "content": " data consistency error." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 395, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 315, + 395, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 315, + 395, + 555, + 436 + ], + "type": "table", + "html": "
Gaussian deblurringHigh dynamic range
PSNR (↑)LPIPS (↓)Run time/sPSNR (↑)LPIPS (↓)Run time/s
Gaussian approx.29.400.17064.724.850.18161.4
diffusion-score29.570.160462.325.330.193540.0
Relative improvement-0.575%-6.25%614.53%-1.89%-6.22%779.48%
", + "image_path": "61b4273aaa2323e2241f894a4be80408d661fd8bab54feb69dc834119a476b64.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 444, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 489 + ], + "type": "text", + "content": "Table 3. Comparison of approaches to approximate " + }, + { + "bbox": [ + 313, + 444, + 555, + 489 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid \\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 444, + 555, + 489 + ], + "type": "text", + "content": ". Three tasks on 10 FFHQ images with DAPS-1K consistently shown Gaussian approximation achieve huge speed-up while slightly sacrificing the performance compared with diffusion-score estimation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 501, + 373, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 373, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 373, + 513 + ], + "type": "text", + "content": "for efficiency." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 513, + 555, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 561 + ], + "type": "text", + "content": "More ablation studies. We include more ablation studies of DAPS in Appendix I.1, including the number of function evaluations, the number of ODE steps, the annealing noise scheduling, and different measurement noise levels." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "In summary, we propose Decoupled Annealing Posterior Sampling (DAPS) for solving inverse problems, particularly those with complex nonlinear measurement processes such as phase retrieval. Our method decouples consecutive sample points in a diffusion sampling trajectory, allowing them to vary considerably, thereby enabling DAPS to explore a larger solution space. Empirically, we demonstrate that DAPS generates samples with better visual quality and stability compared to existing methods when solving a wide range of challenging inverse problems." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20902" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 175 + ], + "type": "text", + "content": "We are grateful to Pika for providing the computing resources essential for this research. We also extend our thanks to the Kortschak Scholars Fellowship for supporting B.Z. and W.C. at Caltech. J.B. acknowledges support from the Wally Baer and Jeri Weiss Postdoctoral Fellowship. A.A. is supported in part by Bren endowed chair and by the AI2050 senior fellow program at Schmidt Sciences." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 205, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 205, + 296, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 296, + 272 + ], + "type": "text", + "content": "[1] Kazunori Akiyama, Antxon Alberdi, Walter Alef, Keiichi Asada, Rebecca Azulay, Anne-Kathrin Baczko, David Ball, Mislav Balokovic, John Barrett, Dan Bintley, et al. First m87 event horizon telescope results. iv. imaging the central supermassive black hole. The Astrophysical Journal Letters, 875(1):L4, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 274, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 296, + 306 + ], + "type": "text", + "content": "[2] Marius Arvinte, Sriram Vishwanath, Ahmed H. Tewfik, and Jonathan I. Tamir. Deep j-sense: Accelerated mri reconstruction via unrolled alternating optimization, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 296, + 350 + ], + "type": "text", + "content": "[3] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981-17993, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 353, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 385 + ], + "type": "text", + "content": "[4] Michael Betancourt and Mark Girolami. Hamiltonian monte carlo for hierarchical models. Current trends in Bayesian methodology with applications, 79(30):2-4, 2015. 4, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 387, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 294, + 419 + ], + "type": "text", + "content": "[5] Benjamin Boys, Mark Girolami, Jakiw Pidstrigach, Sebastian Reich, Alan Mosca, and O. Deniz Akyildiz. Tweedie moment projected diffusions for inverse problems, 2023. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 421, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 296, + 474 + ], + "type": "text", + "content": "[6] Andrew Campbell, Joe Benton, Valentin De Bortoli, Thomas Rainforth, George Deligiannidis, and Arnaud Doucet. A continuous time framework for discrete denoising models. Advances in Neural Information Processing Systems, 35:28266-28279, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 476, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 296, + 509 + ], + "type": "text", + "content": "[7] Gabriel Cardoso, Yazid Janati El Idrissi, Sylvain Le Corff, and Eric Moulines. Monte carlo guided diffusion for bayesian linear inverse problems, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 511, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 296, + 542 + ], + "type": "text", + "content": "[8] Gabriel Victorino Cardoso, Yazid Janati, Sylvain Le Corff, and Éric Moulines. Monte carlo guided diffusion for bayesian linear inverse problems. ArXiv, abs/2308.07983, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 544, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 294, + 567 + ], + "type": "text", + "content": "[9] Junqing Chen and Haibo Liu. An alternating direction method of multipliers for inverse lithography problem, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "text", + "content": "[10] Hyungjin Chung and Jong Chul Ye. Score-based diffusion models for accelerated mri. Medical Image Analysis, page 102479, 2022. 1, 3, 8, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 601, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 296, + 644 + ], + "type": "text", + "content": "[11] Hyungjin Chung, Eun Sun Lee, and Jong Chul Ye. Mr image denoising and super-resolution using regularized reverse diffusion. IEEE Transactions on Medical Imaging, 42(4): 922-934, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[12] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. In Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "text", + "content": "[13] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "type": "text", + "content": "sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 3, 4, 5, 6, 10, 11, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "type": "text", + "content": "[14] J. Dainty and James Fienup. Phase retrieval and image reconstruction for astronomy. Image Recovery: Theory Appl, 13, 1987. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 555, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 185 + ], + "type": "text", + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 186, + 554, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 554, + 219 + ], + "type": "text", + "content": "[16] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In Advances in Neural Information Processing Systems, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 220, + 555, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 263 + ], + "type": "text", + "content": "[17] Zolnamar Dorjsembe, Hsing-Kuo Pao, Sodtavilan Odonchimed, and Furen Xiao. Conditional diffusion models for semantic 3d brain mri synthesis. IEEE Journal of Biomedical and Health Informatics, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 266, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 555, + 308 + ], + "type": "text", + "content": "[18] Zehao Dou and Yang Song. Diffusion posterior sampling for linear inverse problem solving: A filtering perspective. In The Twelfth International Conference on Learning Representations, 2024. 3, 6, 11" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 311, + 555, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 311, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 311, + 555, + 375 + ], + "type": "text", + "content": "[19] Berthy T. Feng, Jamie Smith, Michael Rubinstein, Huiwen Chang, Katherine L. Bouman, and William T. Freeman. Score-based diffusion models as principled priors for inverse imaging. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10520-10531, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 377, + 555, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 555, + 420 + ], + "type": "text", + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 422, + 555, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 555, + 465 + ], + "type": "text", + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 468, + 555, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 555, + 511 + ], + "type": "text", + "content": "[22] Alex Ling Yu Hung, Kai Zhao, Haoxin Zheng, Ran Yan, Steven S Raman, Demetri Terzopoulos, and Kyunghyun Sung. Med-cdiff: Conditional medical image generation with diffusion models. Bioengineering, 10(11):1258, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 513, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 555, + 567 + ], + "type": "text", + "content": "[23] Ajil Jalal, Marius Arvinte, Giannis Daras, Eric Price, Alexandros G Dimakis, and Jon Tamir. Robust compressed sensing mri with deep generative priors. In Advances in Neural Information Processing Systems, pages 14938-14954. Curran Associates, Inc., 2021. 1, 2, 3, 8, 13" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 568, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 555, + 623 + ], + "type": "text", + "content": "[24] Ulugbek Kamilov, Charles Bouman, Gregory Buzzard, and Brendt Wohlberg. Plug-and-play methods for integrating physical and learned models in computational imaging: Theory, algorithms, and applications. IEEE Signal Processing Magazine, 40:85-97, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 667 + ], + "type": "text", + "content": "[25] T. Karras, S. Laine, and T. Aila. A style-based generator architecture for generative adversarial networks. IEEE Transactions on Pattern Analysis & Machine Intelligence, 43(12): 4217-4228, 2021. 5, 10" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[26] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, 2022. 2, 5, 6, 10, 13" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20903" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "text", + "content": "[27] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "type": "text", + "content": "[28] Sergey Kastryulin, Jamil Zakirov, Denis Prokopenko, and Dmitry V. Dylov. Pytorch image quality: Metrics for image quality assessment, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 183 + ], + "type": "text", + "content": "[29] Bahjat Kawar, Gregory Vaksman, and Michael Elad. SNIPS: Solving noisy inverse problems stochastically. Advances in Neural Information Processing Systems, 34:21757-21769, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "type": "text", + "content": "[30] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 6, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 295, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 260 + ], + "type": "text", + "content": "[31] Xiang Li, Soo Min Kwon, Ismail R. Alkhouri, Saiprasad Ravishankar, and Qing Qu. Decoupled data consistency with diffusion purification for image restoration, 2024. 3, 5, 6, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 294 + ], + "type": "text", + "content": "[32] Hanyuan Liu, Jinbo Xing, Minshan Xie, Chengze Li, and Tien-Tsin Wong. Improved diffusion-based image colorization via piggybacked models, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 295, + 327 + ], + "type": "text", + "content": "[33] Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion language modeling by estimating the ratios of the data distribution. arXiv preprint arXiv:2310.16834, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 327, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 360 + ], + "type": "text", + "content": "[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 361, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 403 + ], + "type": "text", + "content": "[35] Morteza Mardani, Jiaming Song, Jan Kautz, and Arash Vahdat. A variational perspective on solving inverse problems with diffusion models. arXiv preprint arXiv:2305.04391, 2023. 6, 11, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 404, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 437 + ], + "type": "text", + "content": "[36] Xiangming Meng and Yoshiyuki Kabashima. Diffusion model based posterior sampling for noisy linear inverse problems. arXiv preprint arXiv:2211.12343, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 437, + 295, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 482 + ], + "type": "text", + "content": "[37] Nicholas Metropolis, Arianna W Rosenbluth, Marshall N Rosenbluth, Augusta H Teller, and Edward Teller. Equation of state calculations by fast computing machines. The journal of chemical physics, 21(6):1087-1092, 1953. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 483, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 483, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 295, + 525 + ], + "type": "text", + "content": "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 526, + 295, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 295, + 559 + ], + "type": "text", + "content": "[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 559, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 295, + 613 + ], + "type": "text", + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 4, 5, 6, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 614, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 658 + ], + "type": "text", + "content": "[41] Litu Rout, Yujia Chen, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Beyond first-order tweedie: Solving inverse problems using latent diffusion, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[42] Litu Rout, Negin Raoof, Giannis Daras, Constantine Caramanis, Alex Dimakis, and Sanjay Shakkottai. Solving linear inverse problems provably via posterior sampling with latent diffusion models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4, 6, 9, 11" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 138 + ], + "type": "text", + "content": "[43] Litu Rout, Yujia Chen, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Beyond first-order tweedie: Solving inverse problems using latent diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9472-9481, 2024. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 194 + ], + "type": "text", + "content": "[44] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 196, + 555, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 555, + 250 + ], + "type": "text", + "content": "[45] Bowen Song, Soo Min Kwon, Zecheng Zhang, Xinyu Hu, Qing Qu, and Liyue Shen. Solving inverse problems with latent diffusion models via hard data consistency. In The Twelfth International Conference on Learning Representations, 2024, 2, 3, 4, 5, 6, 9, 11" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 251, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 295 + ], + "type": "text", + "content": "[46] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 555, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 555, + 329 + ], + "type": "text", + "content": "[47] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Neural Information Processing Systems, 2019. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 555, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 555, + 364 + ], + "type": "text", + "content": "[48] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 365, + 555, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 555, + 419 + ], + "type": "text", + "content": "[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 1, 2, 3, 10" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 421, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 555, + 464 + ], + "type": "text", + "content": "[50] Yang Song, Liyue Shen, Lei Xing, and Stefano Ermon. Solving inverse problems in medical imaging with score-based generative models. In International Conference on Learning Representations, 2022. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 466, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 555, + 499 + ], + "type": "text", + "content": "[51] He Sun and Katherine L. Bouman. Deep probabilistic imaging: Uncertainty quantification and multi-modal solution characterization for computational imaging, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 500, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 555, + 533 + ], + "type": "text", + "content": "[52] Yu Sun, Zihui Wu, Yifan Chen, Berthy Feng, and Katherine L. Bouman. Provable probabilistic imaging using score-based generative priors. *ArXiv*, abs/2310.10835, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 534, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 555, + 578 + ], + "type": "text", + "content": "[53] Phong Tran, Anh Tran, Quynh Phung, and Minh Hoai. Explore image deblurring via encoded blur kernel space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 579, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 555, + 623 + ], + "type": "text", + "content": "[54] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. The Eleventh International Conference on Learning Representations, 2023. 3, 6, 11" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 668 + ], + "type": "text", + "content": "[55] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688. CiteSeer, 2011. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "type": "text", + "content": "[56] Zihui Wu, Yu Sun, Yifan Chen, Bingliang Zhang, Yisong Yue, and Katherine L Bouman. Principled probabilistic imaging using diffusion models as plug-and-play priors. arXiv preprint arXiv:2405.18782, 2024. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20904" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 283 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 171 + ], + "type": "text", + "content": "[57] Jure Zbontar, Florian Knoll, Anuroop Sriram, Matthew J. Muckley, Mary Bruno, Aaron Defazio, Marc Parente, Krzysztof J. Geras, Joe Katsnelson, Hersh Chandarana, Zizhao Zhang, Michal Drozdal, Adriana Romero, Michael G. Rabbat, Pascal Vincent, James Pinkerton, Duo Wang, Nafissa Yakubova, Erich Owens, C. Lawrence Zitnick, Michael P. Recht, Daniel K. Sodickson, and Yvonne W. Lui. fastmri: An open dataset and benchmarks for accelerated MRI. CoRR, abs/1811.08839, 2018. 6, 13" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 172, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 228 + ], + "type": "text", + "content": "[58] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, Los Alamitos, CA, USA, 2018. IEEE Computer Society. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 228, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 283 + ], + "type": "text", + "content": "[59] Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising diffusion models for plug-and-play image restoration. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (NTIRE), 2023. 1, 3, 5, 6, 11, 13" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20905" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improving Editability in Image Generation with Layer-wise Memory/5ef2c05f-1b0c-4899-806c-c1cd90e73e44_content_list.json b/2025/Improving Editability in Image Generation with Layer-wise Memory/5ef2c05f-1b0c-4899-806c-c1cd90e73e44_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..64c95cae2dfd53b4914f5a5f4b75c38e87097b3e --- /dev/null +++ b/2025/Improving Editability in Image Generation with Layer-wise Memory/5ef2c05f-1b0c-4899-806c-c1cd90e73e44_content_list.json @@ -0,0 +1,1945 @@ +[ + { + "type": "text", + "text": "Improving Editability in Image Generation with Layer-wise Memory", + "text_level": 1, + "bbox": [ + 151, + 130, + 846, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daneul Kim Jaeah Lee Jaesik Park* Seoul National University, Republic of Korea {carpedkm, hayanz, jaesik.park}@snu.ac.kr", + "bbox": [ + 313, + 180, + 679, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bfdc191bb47f13184c6baed485a30d1b8264ea7d2796c3a0a27587b65ff39f0d.jpg", + "image_caption": [ + "HD Printer" + ], + "image_footnote": [], + "bbox": [ + 106, + 262, + 380, + 529 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/38a98b819d90e6d9ed2a72adb09c94192d0582d37fd87db48876af5760451c95.jpg", + "image_caption": [ + "Figure 1. Overview. Our framework enables the interactive generation of images with enhanced control but in a simple manner, by rough mask and prompt, through iterative scene editing. We utilize the background scene generated by our framework to edit in HD Painter [32] or Blended Latent Diffusion (BLD) [3] for comparison and commercial products like Photoshop [1] and Pincel [37]." + ], + "image_footnote": [], + "bbox": [ + 388, + 262, + 888, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 631, + 328, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most real-world image editing tasks require multiple sequential edits to achieve desired results. Current editing approaches, primarily designed for single-object modifications, struggle with sequential editing: especially with maintaining previous edits along with adapting new objects naturally into the existing content. These limitations significantly hinder complex editing scenarios where multiple objects need to be modified while preserving their contextual relationships. We address this fundamental challenge through two key proposals: enabling rough mask inputs that preserve existing content while naturally integrating new elements and supporting consistent editing across multiple modifications. Our framework achieves this through layer-wise memory, which stores latent representations and", + "bbox": [ + 89, + 664, + 485, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prompt embeddings from previous edits. We propose Background Consistency Guidance that leverages memorized latents to maintain scene coherence and Multi-Query Disentanglement in cross-attention that ensures natural adaptation to existing content. To evaluate our method, we present a new benchmark dataset incorporating semantic alignment metrics and interactive editing scenarios. Through comprehensive experiments, we demonstrate superior performance in iterative image editing tasks with minimal user effort, requiring only rough masks while maintaining high-quality results throughout multiple editing steps.", + "bbox": [ + 511, + 632, + 908, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 829, + 645, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in text-to-image synthesis through powerful diffusion models like Stable Diffusion [20, 42], PixArt [12, 13] and FLUX [11] have transformed visual", + "bbox": [ + 511, + 856, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 888, + 236, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7889", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "content creation. However, users often demand multiple sequential edits to achieve desired results, iteratively refining and adding elements to their images. Current approaches in inpainting [3, 32] are primarily designed for single-object modifications with limited changes like colors or styles, making them inadequate for complex editing scenarios that involve multiple objects and their interactions.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image editing approaches, while offering localized modifications, face several limitations. Current methods [3, 16, 32, 48] struggle when modifications need to be applied sequentially, often demanding precise segmentation masks [16] or external modules [39] to maintain background integrity [32]. For example, generating \"A dog sitting\" as shown in Fig. 1 remains challenging, as it requires maintaining previous edits while ensuring the \"dog\" naturally blends with the \"Jeep\" and \"Lego man standing\" - a complex iterative editing scenario that current methods [1, 3, 32, 37] struggle to handle.", + "bbox": [ + 89, + 199, + 480, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Layout-to-image generation offers an alternative approach through various inputs including bounding boxes [28, 47, 58], depth maps [26, 54], and semantic masks [7, 27]. However, these methods require complete regeneration of the entire image for each modification, making iterative editing particularly cumbersome – unable to maintain the surrounding background. While recent work has explored layered representations [25, 41, 55] and instance-based generation [46], they require complex optimization processes or additional training, limiting their practical applicability in iterative editing scenarios.", + "bbox": [ + 89, + 368, + 480, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We address these limitations through two key innovations. First, we enable object placement aligning with the user's intention using only rough mask inputs while preserving background context. Second, we support consistent iterative editing across multiple modifications. To achieve this, we introduce the concept of mask order, which specifies the sequence of object generation during iterative image editing. In Fig. 1, we add \"a lego man\", \"Jeep\" sequentially with different mask orders, but they naturally adapt into the image, side by side. If we add \"A dog sitting\" with overlapping mask, then it means that we aim to put the \"dog\" in front of the \"Jeep\" and \"a lego man\", making the instance order of \"dog\" in front of \"Jeep\" and \"a lego man\".", + "bbox": [ + 89, + 536, + 482, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To handle this, we incorporate three technical components: (1) Layer-wise memory for storing editing history, (2) Background Consistency Guidance (BCG) for maintaining unedited regions, and (3) Multi-Query Disentangle-ment (MQD) in cross-attention for natural object integration. The layer-wise memory stores and manages the latent representations and prompt embeddings from previous editing steps, eliminating redundant computations typical in sequential modifications while maintaining consistency across multiple edits. The BCG not only ensures unedited regions remain stable but also reduces computational over", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "head by avoiding repeated forward passes on the original image, enabling reduced editing time, while MQD enables the natural integration of new objects with existing content.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, we propose Multi-Edit Bench, a comprehensive benchmark for evaluating iterative image editing capabilities. Prior benchmarks either focus on single-turn edits [31, 45] or layout-to-image generation [4, 17, 21], failing to capture the challenges of sequential modifications. Our benchmark introduces layer-wise semantic evaluation metrics to assess both edit quality and cross-modification consistency in multi-step editing scenarios.", + "bbox": [ + 511, + 136, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our key contributions include:", + "bbox": [ + 532, + 257, + 834, + 271 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A framework for interactive mask order-based object placement editing with only rough mask inputs.", + "- A localized editing mechanism using layer-wise memory that maintains consistency across multiple edits.", + "- Multi-query disentangled cross attention, allowing natural integration while preserving background context.", + "- Multi-Edit Bench, a comprehensive benchmark for evaluating the semantic alignment and sequential iterative editing capabilities." + ], + "bbox": [ + 511, + 273, + 903, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 422, + 653, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image inpainting. Image inpainting has evolved from classical patch-based [8, 18] and interpolation-driven [5, 9] methods to powerful deep-learning approaches [49-51] that better handle complex structures. Text-conditioned inpainting [51, 53] further broadened user control, and diffusion-based models [24]—including GLIDE [34], Stable Diffusion [42], and Blended Diffusion [2, 3]—significantly improved realism. Yet, existing works often focus on single-object edits. Recent attempts to relax mask precision [32, 48] or enhance quality [16] remain limited in multi-step scenarios, struggling to maintain consistency across multiple objects or occlusions—hence insufficient for iterative real-world editing tasks.", + "bbox": [ + 511, + 446, + 905, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Layout-to-Image Generation. Controllable image synthesis has recently emphasized spatial arrangement via bounding boxes, segmentation masks, or other layout cues [14, 46, 56]. Advances in diffusion-based layout calibration [23] and grounding [43] have improved fidelity, and interactive methods [15, 30, 35] aim to let users refine elements step by step. However, most such pipelines regenerate the entire image for each new edit, losing previously established context and coherence. By contrast, our approach incrementally updates only relevant regions through a memory-assisted process, naturally preserving previously edited objects.", + "bbox": [ + 511, + 643, + 905, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evaluation Benchmark. Most existing editing benchmarks focus on single-turn modifications [31, 45], while layout-centric datasets such as HRS-1k [4] and NSR-1k [21] assess scene quality or prompt alignment in a single-shot manner. Consequently, they overlook the iterative na", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7890", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6e69cc2a32f69fc97a956edd7e5b71fd05cdaf6d233177b964c9a6eaafbd42df.jpg", + "image_caption": [ + "Figure 2. Overview. (a) The left denotes an illustration of how Multi Query Disentanglement is performed in the cross-attention layer. (b) The upper right figure shows Background Consistency Guidance with recalled latents, conducting latent blending with the saved latents. (c) The right below shows the layer-wise memory, saving the previous editing steps' latents, masks, and prompt embeddings." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 893, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ture of real-world edits where object placements evolve over multiple steps. To address this gap, we introduce Multi-Edit Bench, a new benchmark that evaluates multi-step editing in terms of semantic accuracy and visual alignment across the entire editing sequence. This multi-step perspective better captures practical editing workflows, where users iteratively refine scenes by adding, removing, or repositioning objects, thereby measuring a method's capacity to preserve context and maintain coherent compositions over multiple, potentially complex modifications.", + "bbox": [ + 88, + 450, + 486, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 621, + 184, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 648, + 200, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a framework for iterative image editing with three key components: Layer-wise memory, background consistency guidance (BCG), and multi-query disentangled cross-attention (MQD). Layer-wise memory stores editing history to enable consistent object placement across sequential edits. BCG leverages this stored information for efficient latent blending while preserving background integrity. MQD ensures the natural object integration through disentangling the attention between queries and latents, handling complex scenarios as shown in Fig. 2 like placing \"a dish\" in front of \"a cupcake\" and \"a mug cup\".", + "bbox": [ + 88, + 672, + 482, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework's architecture, built upon transformer-based diffusion models [13], operates through the following systematic workflow: First, we detail the layer-wise memory mechanism (Fig. 2(c)) in Sec. 3.2, which stores editing", + "bbox": [ + 88, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "history. Then, in Sec. 3.3, we explain how BCG leverages this stored information for efficient latent blending (Fig. 2 (b)). Sec. 3.4 describes MQD's role in ensuring natural object integration through query disentanglement (Fig. 2 (a)). Finally, in Sec. 3.5, we demonstrate how these components enable advanced editing capabilities, including the removal of occluded elements while preserving foreground integrity.", + "bbox": [ + 511, + 450, + 906, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Layer-wise Memory", + "text_level": 1, + "bbox": [ + 511, + 571, + 705, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Layer-wise memory enables stable background preservation while ensuring the natural integration of new objects during iterative editing with mask order. By storing and managing key information from each editing step according to the specified mask order, the model can reference previous edits when generating new content while maintaining the intended spatial relationships between objects.", + "bbox": [ + 511, + 595, + 906, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let the set of prompts for generating the background and objects be denoted as $P_{l} = \\{p_{0}, p_{1}, p_{2}, \\ldots\\}$ , where $p_{0}$ is the background generation prompt, and $p_{1}, p_{2}, \\ldots$ represent the prompts for each object. Similarly, let the set of corresponding masks be denoted as $M_{l} = \\{m_{0}, m_{1}, m_{2}, \\ldots\\}$ , where each $m_{i}$ defines the region of interest (RoI) for the object associated with prompt $p_{i}$ . Note that $m_{0}$ represents the case with mask of all elements being 1.", + "bbox": [ + 511, + 702, + 908, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The layer-wise memory $L_{l} = \\{l_{0}, l_{1}, l_{2}, \\ldots \\}$ stores three key elements for each step $i$ : $l_{i} = \\{\\mathbf{p}_{i}, \\{\\mathbf{Z}_{i}^{t}\\}_{t=1}^{T}, m_{i}\\}$ , where: $\\mathbf{p}_{i}$ denotes prompt embedding that guides the generation, $\\{\\mathbf{Z}_{i}^{t}\\}_{t=1}^{T}$ as denoising latents across timesteps and $m_{i}$ as a mask defining the object's RoI.", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7891", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/2a50c6872a4a696b8f95cafc2b852620a47bff0b12f2e3297240da064281ed38.jpg", + "table_caption": [ + "Table 1. Computational comparison in utilizing Background Consistency Guidance (BCG). We utilize PixArt- $\\alpha$ inpaint pipeline with vanilla latent blending and our custom-implemented pipeline with BCG for comparison. We ran the inpainting task for the single-step editing 5 times for each method. Note that we use PixArt-alpha-XL-1024 pretrained weight for the experiment." + ], + "table_footnote": [], + "table_body": "
MethodTime (sec)VRAM
Latent Blending4.1218 ±0.052116.89GB
BCG3.8992±0.014216.90GB
", + "bbox": [ + 142, + 183, + 431, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a generation, the initial background latent is created as $\\mathbf{Z}_0 = f_\\theta(p_0)$ , where $f_\\theta$ is the pre-trained diffusion model. For each subsequent object $p_i$ , we initialize its latent independently: $\\mathbf{Z}_i = f_\\theta(p_i, m_i)$ . This independence in initialization, combined with the stored history in layer-wise memory, enables flexible generation while maintaining background consistency.", + "bbox": [ + 89, + 248, + 483, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The layer-wise memory serves two key purposes: (1) Context preservation, to store the complete editing history, enabling the model to maintain consistency with previously generated content (2) Localized editing, enabling control over specific regions while preserving the surrounding content through mask-guided generation", + "bbox": [ + 89, + 354, + 483, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By maintaining this structured history of edits, the model can effectively handle complex scenarios where multiple objects need to be integrated coherently. The stored information guides background consistency and natural object integration, as detailed in the following sections.", + "bbox": [ + 89, + 444, + 483, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Background Consistency Guidance", + "text_level": 1, + "bbox": [ + 89, + 529, + 397, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Background Consistency Guidance (BCG) addresses a key challenge in iterative image editing: maintaining the integrity of previous edits while incorporating new elements efficiently. Unlike traditional inpainting approaches, BCG leverages layer-wise memory to ensure stability across multiple sequential modifications.", + "bbox": [ + 89, + 550, + 483, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent Retrieval from Layer-wise Memory: For each editing step $i$ , BCG retrieves information from the previous edits stored in layer-wise memory $L_{l}$ , enabling seamless integration of new content while preserving prior modifications:", + "bbox": [ + 89, + 648, + 483, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl _ {i - 1} = \\left\\{\\mathbf {p} _ {i - 1}, \\left\\{\\mathbf {Z} _ {i - 1, t} \\right\\} _ {t = 1} ^ {T}, m _ {i - 1} \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 722, + 482, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_{i - 1}$ is the prompt embedding from the prior step, $\\{\\mathbf{Z}_{i - 1,t}\\}_{t = 0}^{T}$ represents the stored latents from each denoising step, and $m_{i - 1}$ is the previous mask. This stored latent provides context for the current edit, preserving the background information without requiring a re-computation on the original image.", + "bbox": [ + 89, + 746, + 483, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ensuring Consistency: When adding a new object, we update only the masked region while preserving the rest:", + "bbox": [ + 89, + 843, + 483, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {i} = \\mathbf {Z} _ {i - 1} \\odot (1 - m _ {i}) + \\mathbf {Z} _ {i} \\odot m _ {i} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This selective update mechanism is crucial for iterative editing since it ensures that each new modification preserves previous edits. The element-wise multiplication $\\odot$ enables control over which regions are updated, maintaining the integrity of the evolving image composition.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Computational Efficiency for Iterative Editing:", + "text_level": 1, + "bbox": [ + 511, + 172, + 846, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BCG improves the efficiency of iterative image editing by reducing redundant computations in sequential modifications. Traditional inpainting methods [3] that utilize latent blending (LB) require a forward pass on the original image for each edit, becoming costly with multiple edits. While both approaches (i.e., LB, BCG) require denoising costs, BCG avoids repeated forward passes. Let $\\Omega = \\mathrm{FLOPs}(T,L,H,W)$ represent the base computational cost for processing an image of size $H\\times W$ through $T$ denoising steps and $L$ model layers, and $C_f$ denote the forward pass cost. The total cost for LB and BCG can be expressed as:", + "bbox": [ + 511, + 188, + 906, + 369 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C o s t} _ {\\mathrm {L B}} = C _ {f} + \\Omega . \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 369, + 903, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While for BCG, by skipping the forward pass:", + "bbox": [ + 511, + 392, + 821, + 407 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C o s t} _ {\\mathrm {B C G}} = \\Omega , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 420, + 903, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assuming $C_f \\approx r\\Omega$ with ratio $r \\in (0,1)$ , we can estimate:", + "bbox": [ + 511, + 446, + 901, + 463 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {E f f e c i e n c y} = \\frac {\\operatorname {C o s t} _ {\\mathrm {L B}}}{\\operatorname {C o s t} _ {\\mathrm {B C G}}} \\approx 1 + r \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 473, + 903, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Tab. 1, we observe about $10\\%$ reduction in computational time for single-step editing with BCG compared to LB. This improvement becomes more significant in scenarios requiring multiple sequential modifications, as LB requires multiple forward passes while BCG requires none.", + "bbox": [ + 511, + 513, + 905, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Multi-Query Disentangled Cross-Attention", + "text_level": 1, + "bbox": [ + 511, + 599, + 879, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-query disentangled cross-attention (MQD) ensures the natural integration of objects across different mask orders while preserving their spatial relationships. For example, when generating \"a dish\" as in Fig. 2, MQD enables the 'dish' to blend naturally into the background, despite being in different mask orders, while maintaining the structure and color of the background with proper occlusion of other objects like 'a mug cup' and 'a cupcake' behind. This is achieved by disentangling attention across multiple queries and leveraging information from layer-wise memory $L_{l}$ .", + "bbox": [ + 511, + 621, + 906, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.1. Cross-Attention for Current Object", + "text_level": 1, + "bbox": [ + 511, + 779, + 808, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each generation step $i$ , MQD focuses attention on the current mask order's region of interest (RoI):", + "bbox": [ + 511, + 797, + 905, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i} ^ {k, a t t n} = \\operatorname {C r o s s A t t e n t i o n} \\left(\\mathbf {z} _ {i} ^ {k, t} \\odot m _ {i}, p _ {i}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 838, + 903, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_i^{k,t}$ is the latent for the $k$ -th block at denoising step $t$ , $m_i$ defines the RoI, and $p_i$ is the prompt embedding. This", + "bbox": [ + 511, + 868, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7892", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/71881933e23c4b28840b6b3a6557af2998aea9131e9970696fe1d5f34a961af2.jpg", + "image_caption": [ + "Figure 3. Overview of our proposed Multi-Edit Benchmark for evaluation of iterative editing scenario. (a) explains the dataset generation pipeline through GPT-4 API, and (b) explains the evaluation methodology in visual alignment using CLIP and semantic alignment using LLaVa for single-image and in a layer-wise manner." + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 566, + 324 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8012102f00fda90e3434956ea080229a5b167e8819351e06040df0be5926b094.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 88, + 864, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ensures focused generation within the current mask order's region. In the Fig. 2, $m_{i}$ is given as a mask corresponding to the $p_{i}$ 'a dish' with $\\mathbf{z}_{i}^{k,t}$ corresponding to the latent of the masked region for 'a dish'.", + "bbox": [ + 89, + 382, + 483, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2. Disentangled Attention for Previous Mask Orders", + "text_level": 1, + "bbox": [ + 89, + 449, + 483, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To adapt the mask orders into the appropriate instance orders, MQD applies attention to non-overlapping regions from previous steps:", + "bbox": [ + 89, + 468, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i} ^ {k, a t t n} = \\bigcup_ {j = 0} ^ {i - 1} \\text {C r o s s A t t e n t i o n} \\left(\\mathbf {z} _ {i} ^ {k, t} \\odot \\left(m _ {j} - \\Sigma_ {l = j + 1} ^ {i} m _ {l}\\right), p _ {j}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 523, + 482, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This operation ensures that objects from earlier mask orders, (i.e., previous masks for 'a mug cup' and 'a cupcake' on top of 'a wooden floor' behind 'a dish' altogether) remain coherent while allowing natural adaptation where needed. The term $m_{j} - \\sum_{l = j + 1}^{i}m_{l}$ controls which regions from previous mask orders should influence the current generation, minimally affecting the \"mug cup\" and \"cupcake\" even if we put the \"dish\" in front of them. Please note that for background prompt $p_0$ : \"wooden floor\", we inverse the whole union of masks and do disentangled cross attention with background prompt $p_0$ as in Fig. 2 (a).", + "bbox": [ + 89, + 579, + 483, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.3. Final Latent Update", + "text_level": 1, + "bbox": [ + 89, + 752, + 279, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final update combines information across all mask orders:", + "bbox": [ + 89, + 771, + 482, + 799 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i} ^ {\\text {m e r g e}} = \\mathbf {z} _ {i} ^ {\\text {a t t n}} \\bigcup \\sum_ {j = 0} ^ {i - 1} \\mathbf {z} _ {j} ^ {\\text {a t t n}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 797, + 482, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After a feedforward layer update and $K$ transformer blocks, we apply BCG:", + "bbox": [ + 89, + 844, + 482, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {i} ^ {t} = \\mathbf {z} _ {i} ^ {K, t} \\odot m _ {i} + \\mathbf {Z} _ {i - 1} ^ {t} \\odot (1 - m _ {i}) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 882, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This approach enables objects to adapt naturally to mask orders—for example, placing “a dog” in front of “Jeep” and “Lego man” (Fig. 1), or maintaining ‘a cupcake’ and ‘a mug cup’ despite inserting “a dish” in front (Fig. 2)—while preserving scene coherence. MQD balances content preservation with the seamless integration of new elements.", + "bbox": [ + 511, + 382, + 906, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Improving Editability", + "text_level": 1, + "bbox": [ + 511, + 484, + 718, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While BCG and MQD with layer-wise memory enhance multiple editing capabilities in the PixArt- $\\alpha$ model, we further extend our framework to support object deletion, particularly for overlapped objects behind foreground elements. We achieve this through a novel application of our MQD and BCG components.", + "bbox": [ + 511, + 507, + 906, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To delete an object added at editing step $i - 1$ , we leverage the stored latent representations from both steps $i - 2$ and the current step $i$ : $\\{\\mathbf{Z}_{i - 2}^t\\}_{t = 1}^T$ and $\\{\\mathbf{Z}_i^t\\}$ . Starting from an intermediate denoising step $\\tau$ , we blend these latents as:", + "bbox": [ + 511, + 599, + 906, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {\\text {e r a s e}} ^ {\\tau} = m _ {i} \\odot \\mathbf {Z} _ {i} ^ {\\tau} + (1 - m _ {i}) \\odot \\mathbf {Z} _ {i - 2} ^ {\\tau} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 676, + 906, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our deletion process operates in two phases: from step $\\tau$ to $\\frac{\\tau}{2}$ , we perform latent blending following Eq. (10) with $\\mathbf{Z}_{i-2}^{t}$ . After step $\\frac{\\tau}{2}$ , we switch to vanilla denoising to ensure a smooth transition. Crucially, we process $\\mathbf{Z}_{\\text{erase}}$ through MQD without the $i-1$ step's mask or prompt embedding, effectively removing the target object's influence.", + "bbox": [ + 511, + 703, + 905, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This approach offers two key advantages. First, by leveraging BCG to start from step $\\tau$ rather than T, we achieve $60\\%$ faster editing while preserving background consistency, as we use $\\tau = 8$ out of a total of 20 steps. Second, MQD's application from step $\\tau$ maintains foreground object integrity while effectively removing background elements, eliminating the need for precise object masks in deletion.", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7893", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/05f282fd4567a45b0be9cbea2890a33f35adb9867161913c71b7c126e40b92c1.jpg", + "table_caption": [ + "Table 2. Quantitative results on our proposed benchmark. Our method shows higher semantic and visual alignment than other baselines, including layout-to-image synthesis and image editing frameworks. Ordering† denotes the adaptation of the framework adequate for iterative image generation." + ], + "table_footnote": [], + "table_body": "
TypeMethodResolutionSemantic AlignmentVisual Alignment
BLEU-2/3/4↑METEOR↑CLIPcrop↑
2D Layout-to-ImageLayoutGuidance [14]512 × 51236.44 / 26.13 / 18.850.136162.92
NoiseCollage [43]512 × 51255.75 / 42.43 / 32.960.140264.01
NoiseCollage + ordering†512 × 51259.98 / 43.76 / 32.240.146464.10
3D Layout-to-ImageLooseControl [10]512 × 51263.30 / 46.24 / 34.150.137363.13
Image EditingBLD [3]1024 × 102455.30 / 40.38 / 29.580.148062.40
HD-Painter [32]1024 × 102463.29 / 47.63 / 36.280.148464.09
SD3-Inpaint [20]1024 × 102429.90 / 21.64 / 15.780.144563.98
Ours512 × 51261.19 / 45.04 / 34.060.146564.28
Ours1024 × 102464.99 / 47.69 / 36.590.151364.29
", + "bbox": [ + 132, + 136, + 859, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Multi-Edit Benchmark", + "text_level": 1, + "bbox": [ + 89, + 340, + 307, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prior benchmarks for image editing [31, 45] are limited to evaluate only single-step editing or limited to instruct-based edit [22], and layout-guided image generation benchmarks are limited to evaluate whether the specific object is generated or not in the bounding box or evaluate only the flat spatial arrangements. [4, 17, 21]. Therefore, we aim to tackle the issue of evaluating the semantic alignment of the generated images by assessing whether the object generated is aligned with the caption that is used for sequential editing. Also, to handle the mask order, we propose to evaluate the generated image in a layer-wise manner aligning with mask orders. Specifically, to evaluate interactive generation and editing, we need to evaluate whether each layer shows the desired object within the given mask.", + "bbox": [ + 88, + 367, + 482, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We generate the dataset by selecting classes from ImageNet-1K and arranging them in layered compositions with varying degrees of occlusion and spatial relationships, ensuring mask order is a key factor in the generation process as in the Fig. 3 (a). We utilize GPT-4 API to assist in selecting object classes (e.g., 'Lion', 'Tiger', 'Elephant') that naturally fit together, creating realistic compositions, and generating template-based captions for both the global scene and individual object layers to integrate spatial relations into prompt and utilize GPT to make prompt richer.", + "bbox": [ + 88, + 580, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the spatial and semantic alignment between generated images and prompts, we crop each object layer using its corresponding mask and evaluate it individually as in Fig. 3 (b). Both cropped and final images are resized to $224 \\times 224$ for consistency across evaluations. We employ LLaVa [29] to generate captions and compute alignment metrics such as BLEU [36], METEOR [6], ensuring a robust evaluation of both spatial arrangement and semantic accuracy. Note that we average the result from all the layers of each image to evaluate.", + "bbox": [ + 88, + 733, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Also, to check each layer's visual alignment of the gen", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "erated object, we crop the image for each iteration and calculate the CLIP score following other layout-to-image models [4, 21, 43] as individual editing step is similar to layout-to-image generation. We discuss and provide detailed descriptions of the dataset creation process, evaluation metrics, and the resulting benchmark in the supplement.", + "bbox": [ + 511, + 342, + 906, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 511, + 445, + 638, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 470, + 730, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We evaluate our method against state-of-the-art baselines in both image editing and layout-to-image generation. For image editing, we compare with HD-Painter [32], Blended Latent Diffusion [3] with SD-XL [38], and SD3-ControlNet-Inpaint [20]. As sequential editing shares similarities with layout-to-image generation, we also include train-free layout models: NoiseCollage [43] and Layout-Guidance [14], and 3D-lifted model: LooseControl [10].", + "bbox": [ + 511, + 492, + 906, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we adapt NoiseCollage to handle sequential mask inputs (denoted as NoiseCollage + ordering) to explore the potential of converting layout-to-image models for sequential editing tasks. Following our train-free approach using the PixArt- $\\alpha$ foundation model [13], all selected baselines operate without additional training.", + "bbox": [ + 511, + 614, + 905, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We leverage PixArt- $\\alpha$ [13], the variant of Diffusion Transformer (DiT) for image editing framework. We utilize a $1024 \\times 1024$ pretrained model with DPM-Solver with DPMS guidance of 7.5 and a total denoising step of 20. Also, for a fair comparison with models that generate $512 \\times 512$ resolution [14, 43], we compare quantitatively by generating $512 \\times 512$ images with the $512 \\times 512$ pretrained PixArt- $\\alpha$ model.", + "bbox": [ + 511, + 704, + 906, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Quantitative Result", + "text_level": 1, + "bbox": [ + 511, + 833, + 700, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the quantitative evaluation results on our proposed Multi-Edit Benchmark in Tab. 2. Our method consistently outperforms both image editing and layout-to-image", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7894", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3349e8ef666862b985cc620579d0658329eebd9705450015d2491e7c56b2c033.jpg", + "image_caption": [ + "\"Savanna Plains\"", + "Before Edit", + "Figure 4. Qualitative comparison on the effect of Query Disentanglement (QD)." + ], + "image_footnote": [], + "bbox": [ + 138, + 104, + 236, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10ea1e082042a4aa2b7b2f212a08a4ab700d25e78bbc0e0944af5c7fa5e02ded.jpg", + "image_caption": [ + "+ \"A flamingo\"", + "Edit w/o QD" + ], + "image_footnote": [], + "bbox": [ + 248, + 104, + 348, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16f830dd65b1ee8fcdd96945bfbf27381c746f21c392c5921902cdd896245abc.jpg", + "image_caption": [ + "Edit w/QD" + ], + "image_footnote": [], + "bbox": [ + 352, + 104, + 450, + 180 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "belines across key metrics. Among editing baselines, we demonstrate superior performance compared to Blended Latent Diffusion (BLD) [3] with SD-XL, HD-Painter [32], and SD3-ControlNet-Inpaint [20]. Notably, the significant performance gap with SD3-ControlNet-Inpaint can be attributed to the suboptimal property of utilizing ControlNet-Inpainting in multi-step editing.", + "bbox": [ + 89, + 242, + 482, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our framework shows particular strength in handling sequential editing challenges, demonstrated by consistent improvements across all metrics. This validates our approach's effectiveness in maintaining background consistency while naturally integrating new objects - a key challenge in sequential editing tasks.", + "bbox": [ + 89, + 349, + 483, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The performance gap grows further relative to 2D layout-to-image baselines, which often struggle with overlapping or cascaded placements. In particular, our method outperforms NoiseCollage by $4 - 5\\%$ p and surpasses Layout-Guidance by over $15\\%$ p in BLEU and outperforms in all other metrics. Against the 3D layout-to-image method that leverages pseudo depth maps, LooseControl [10] shows improved BLEU compared to the 2D baselines, but ours outperforms on all metrics. (See supplement for comparison with attribute editing on LooseControl.)", + "bbox": [ + 89, + 440, + 483, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While we adapt NoiseCollage [43] for iterative image editing (+ ordering), NoiseCollage + ordering achieves better visual alignment scores, indicating the potential viability of adapted layout-to-image approaches. However, its limited improvement across semantic alignment metrics suggests that simple adaptations are insufficient for complex sequential editing scenarios. This underscores the need for specialized approaches like ours that can effectively handle the unique challenges of iterative image editing.", + "bbox": [ + 89, + 592, + 483, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 741, + 243, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present an ablation of our key components (Tab. 3), starting from PixArt- $\\alpha$ [13]. Background consistency guidance (BCG) reuses latents for faster editing (Tab. 1), improving BLEU and CLIP while preserving METEOR. Next, Query-Disentanglement (QD), inspired by NoiseCollage [43], preserves the existing background rather than fully covering the mask, raising BLEU but slightly reducing CLIP. Direct latent blending instead inflates CLIP by overfilling the mask, lacking context (Fig. 4). Finally, extend", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study of our proposed method. Results are reported on both semantic and visual alignment metrics. We denote vanilla PixArt- $\\alpha$ inpainting model as Baseline. Also we denote QD as Query-Disentanglement without memory, only with the current object and background.", + "bbox": [ + 511, + 89, + 905, + 159 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/48f25beda0401dafb8a9bba6c5a184fcd79354f83897d517ad92782f708ebd55.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSemantic Align.Visual Align.
BLEU-2/3/4↑METEOR↑CLIPc↑
Baseline56.29 / 42.04 / 33.060.158664.05
+BCG60.74 / 46.27 / 35.200.158564.10
+QD62.68 / 46.42 / 35.030.153063.99
Ours64.99 / 47.69 / 36.590.151364.29
", + "bbox": [ + 535, + 162, + 880, + 256 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4f6874d6f7b55440dd5e9475b6e4c5d8e8987aed43525e9c5b78e21a31bb644.jpg", + "image_caption": [ + "\"A beach\"" + ], + "image_footnote": [], + "bbox": [ + 627, + 263, + 669, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b7ba283c8e3ad56438116863c387533bd2fca8a56d1e16a6506dd4cecb9cf88.jpg", + "image_caption": [ + "+“a blue bus\"" + ], + "image_footnote": [], + "bbox": [ + 700, + 263, + 740, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c29dc64bdba4590f0d3f43ee11e838020e44669cb21046da243d2ad11df75f93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 263, + 812, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2595126f4b4c2ccc3c831e1a0421ee0cac385dd879c6cd70ce294dc893a92c4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 313, + 653, + 404 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b3b4bf6f43b39d757c84d6ae8e8fa8407bbd438ddbf8b588df5345ca3191ea71.jpg", + "image_caption": [ + "HD Painter", + "Figure 5. Comparison in image editing capability with latest image editing models. [3, 32] Note that the initial image is generated by our framework, which is equivalent to PixArt- $\\alpha$ [13] with no mask input." + ], + "image_footnote": [], + "bbox": [ + 535, + 406, + 653, + 498 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08d0a8d4c7035059c851cadea660a63db8e7e5253aafe4101026c6bf359f5bbb.jpg", + "image_caption": [ + "+“a hotel behind the bus\"" + ], + "image_footnote": [], + "bbox": [ + 656, + 314, + 776, + 405 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/000f7e58b0bd78ec1db515d762b4a9df8c57e419e6cbecd712b944394b610208.jpg", + "image_caption": [ + "Blended Latent Diffuion" + ], + "image_footnote": [], + "bbox": [ + 658, + 406, + 776, + 498 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/818b02e5baaa063e5ffc885cd3869b03d45998f95277ecb183616e9141581a6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 314, + 897, + 405 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4d11c282c2525a0869a0bb24368763b9a4428b6fe12f185a7bf545a25f26ae8.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 777, + 406, + 897, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing QD to multi-query disentanglement (MQD) with layerwise memory integrates new objects into all prior prompts, achieving higher BLEU-2/3/4 and CLIP scores.", + "bbox": [ + 511, + 577, + 905, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Qualitative Result", + "text_level": 1, + "bbox": [ + 511, + 633, + 689, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Editing Approaches under Interactive Scenario. In Fig. 5, we demonstrate our method's effectiveness in iterative editing compared to state-of-the-art single-step editing approaches [2, 32]. Starting with a background of \"A beach\" generated by vanilla PixArt- $\\alpha$ [13], we sequentially place a bus and a hotel. Our method successfully renders the hotel behind the bus, accurately interpreting both the prompt \"A beach\" and the mask order through MQD to achieve the intended spatial arrangement.", + "bbox": [ + 511, + 657, + 905, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In contrast, Blended Latent Diffusion [3] struggles with the initial placement of \"bus\" in the complex background, while HD-Painter [32] shows inconsistencies in the appearance of \"bus\" (highlighted in red in Fig. 5). Our method maintains background consistency while naturally integrating new objects into the scene.", + "bbox": [ + 511, + 794, + 905, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Improved Editability. Fig. 6 compares our method with", + "bbox": [ + 511, + 885, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7895", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc416de1d738b12fafb772c03447940e45cdfe8e61243e519fdc371a654ba223.jpg", + "image_caption": [ + "Figure 6. Improved editability of image through Background Consistency Guidance and Multi-Query Disentangled cross attention. Through recycling the previous step's latents, we can remove the object that is behind the foreground object, enabling enhanced editability of the image." + ], + "image_footnote": [], + "bbox": [ + 109, + 89, + 460, + 231 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ad0fe648b5fbf9791483deea195f9ce5aeda01719613d8f8fb80a94c0fa8632b.jpg", + "table_caption": [ + "Table 4. Human preference study with recent literature on image editing on our benchmark. We utilize HD-Painter [32] and Blended Latent Diffusion with SD-XL [38] to compare under iterative editing scenario and compare the preference on a 5-point Likert scale." + ], + "table_footnote": [], + "table_body": "
MethodBackground ConsistencyNatural AdaptationText-scene Alignment
HD-Painter [32]3.712.813.08
BLD [3]3.432.242.04
Ours4.594.284.49
", + "bbox": [ + 120, + 388, + 449, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Photoshop 2025 [1]. Placing \"a glass of wine\" or \"a lizard\" is straightforward, but removing \"a glass of wine\" while preserving \"a lizard\" remains challenging for Photoshop's generative fill, which requires precise brush strokes and often leaves artifacts (red box). Our approach simplifies this by leveraging existing masks and prior-step latents to naturally inpaint the background while applying MQD during partial denoising to preserve the lizard's structure.", + "bbox": [ + 89, + 484, + 482, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with Existing T2I models. Fig. 7 showcases our method's ability to handle complex multi-object scenarios that challenge conventional text-to-image models. For instance, in row A, only our framework successfully generates the intended composition of a lion behind a flower. Row B and C further demonstrate the challenge for existing T2I approaches to handle complex spatial arrangements, while illustrating our method's capability.", + "bbox": [ + 89, + 607, + 482, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Human Preference Study", + "text_level": 1, + "bbox": [ + 89, + 739, + 321, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct a comprehensive human evaluation study with 50 participants to assess the quality of our method compared to the latest editing models [3, 32] on our benchmark. The evaluation focuses on three aspects: background consistency, natural adaptation, and text-scene alignment, rated on a 5-point Likert scale.", + "bbox": [ + 89, + 763, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method outperforms single-step editing approaches [3, 32] across all three categories. These results demonstrate that traditional editing models, while effective", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5521930328725612bc14122e12a920d8f6738b17297f1d63cf55fa1ae728029d.jpg", + "image_caption": [ + "A" + ], + "image_footnote": [], + "bbox": [ + 535, + 88, + 651, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4381c49380e848f98f561e81b26924398e55efd6a4f2377c95b22edd49bacbec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 88, + 769, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17869df27a5a503c91705d8df9c6b4f9f2b7d9c7b1859dc1b4706a2ab06dd265.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 88, + 887, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/adc244b570684764c6863bf04dcc927835ba84353f5dd0c9b2f108fc670c2126.jpg", + "image_caption": [ + "B" + ], + "image_footnote": [], + "bbox": [ + 537, + 174, + 651, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c74a2b1302ed49ecf4c7bf97bc09e16b45048174b10328e1ee6bfd7ac15cd078.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 174, + 769, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff4f4dfe954ac7c2b52de83217a07507e95155d3d7eaf3d23207fe67141b6c3a.jpg", + "image_caption": [], + "image_footnote": [ + "C. A wooden table has a flower pot on the left and some stacked books in the center, with a small statue resting behind the table." + ], + "bbox": [ + 769, + 174, + 887, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d09690a037f82844e1c23a78c3963570416cc67d2c176a1f320b31b3d77249ad.jpg", + "image_caption": [ + "C", + "PIXART-a", + "Figure 7. Comparison in interactive scenarios with existing T2I generative models. Stable Diffusion XL [38] and PixArt- $\\alpha$ [13] use text input only." + ], + "image_footnote": [ + "A. A lion stands behind flowers in the desert, positioned next to the water where a palm tree is growing nearby under the bright sun." + ], + "bbox": [ + 537, + 261, + 651, + 347 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bfc138d8bdf7561517cefbb74fea5851225a8376593becc78d8490766a44dfe5.jpg", + "image_caption": [ + "Stable Diffusion XL" + ], + "image_footnote": [ + "B. Santa Claus and a teddy bear are sitting by a campfire in a snowy forest, with a gingerbread house directly behind them under falling snow." + ], + "bbox": [ + 653, + 261, + 769, + 347 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea7f6ea9b9f5ea642f4b6747ed5af887b98f2a0c00d12924dc8debb83e9a7aac.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 769, + 261, + 887, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for single-step modifications, are insufficient for sequential editing. This underscores the necessity of specialized approaches like ours for handling sequential iterative editing scenarios. Additional human evaluation results comparing our method with text-to-image models [13, 38] are provided in the supplementary materials.", + "bbox": [ + 511, + 484, + 905, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 592, + 633, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel framework for sequential image editing, which poses challenges in maintaining background consistency and seamless integration of new objects into the scenes. Our method includes three key components: Layer-wise Memory to preserve previous content, Background Consistency Guidance to keep the background stable with faster editing, and Multi-Query Disentangled Cross-Attention for natural object adaptation. Experiments on our proposed benchmark dataset show that our approach outperforms state-of-the-art methods in semantic and visual alignment, making it ideal for complex, iterative image editing scenarios. This framework paves the way for future work in improved interactive editing.", + "bbox": [ + 511, + 618, + 906, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and Future Work. Since our approach utilizes image editing, generating multiple objects takes longer, depending on the number of edits. Also, utilizing layer-wise memory requires additional memory costs. We plan to make it more efficient for faster editing in the future.", + "bbox": [ + 511, + 829, + 906, + 905 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7896", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by IITP grant (RS-2021-II211343: AI Graduate School Program at Seoul National University $(5\\%)$ and RS-2024-00509257: Global AI Frontier Lab $(30\\%)$ ) and NRF grant (No.RS-2024-00405857 $(65\\%)$ ) funded by the Korea government (MSIT).", + "bbox": [ + 89, + 90, + 485, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 196, + 187, + 213 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adobe Inc. Adobe Photoshop. https://www.adobe.com/products/photoshop.html, 2024. Version 2025. 1, 2, 8, 17, 20", + "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 2, 7, 16", + "[3] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM TOG, 2023. 1, 2, 4, 6, 7, 8, 11, 13, 17, 20", + "[4] Eslam Mohamed Bakr, Pengzhan Sun, Xiaogian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models. In ICCV, 2023. 2, 6, 14", + "[5] Coloma Ballester, Marcelo Bertalmio, Vicent Caselles, Guillermo Sapiro, and Joan Verdera. Filling-in by joint interpolation of vector fields and gray levels. IEEE TIP, 2001. 2", + "[6] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In IEEvaluation@ACL, 2005. 6", + "[7] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 2", + "[8] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM TOG, 2009. 2", + "[9] Marcelo Bertalmio, Guillermo Sapiro, Vicent Caseles, and Coloma Ballester. Image inpainting. In SIGGRAPH, 2000. 2", + "[10] Shariq Farooq Bhat, Niloy J. Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning. In SIGGRAPH, 2024. 6, 7, 16", + "[11] BlackForestLabs. Flux, 2023. Accessed: 2024-11-12. 1", + "[12] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\\sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv:2403.04692, 2024. 1", + "[13] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 1, 3, 6, 7, 8, 13", + "[14] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In WACV, 2024. 2, 6" + ], + "bbox": [ + 93, + 223, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Ruidong Chen, Lanjun Wang, Weizhi Nie, Yongdong Zhang, and An-An Liu. Anyscene: Customized image synthesis with composited foreground. In CVPR, 2024. 2", + "[16] Yifu Chen, Jingwen Chen, Yingwei Pan, Yehao Li, Ting Yao, Zhineng Chen, and Tao Mei. Improving text-guided object inpainting with semantic pre-inpainting. In ECCV, 2024. 2", + "[17] Jaemin Cho, Linjie Li, Zhengyuan Yang, Zhe Gan, Lijuan Wang, and Mohit Bansal. Diagnostic benchmark and iterative inpainting for layout-guided image generation. In CVPR Workshop, 2024. 2, 6", + "[18] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE TIP, 2004. 2", + "[19] Abdelrahman Eldisokey and Peter Wonka. Build-a-scene: Interactive 3d layout control for diffusion-based image generation. In ICLR, 2025. 16", + "[20] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 6, 7, 11", + "[21] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. In NeurIPS, 2023. 2, 6, 14", + "[22] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv:2405.04007, 2024. 6", + "[23] Biao Gong, Siteng Huang, Yutong Feng, Shiwei Zhang, Yuyuan Li, and Yu Liu. Check locate rectify: A training-free layout calibration system for text-to-image generation. In CVPR, 2024. 2", + "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2", + "[25] Runhui Huang, Kaixin Cai, Jianhua Han, Xiaodan Liang, Renjing Pei, Guansong Lu, Songcen Xu, Wei Zhang, and Hang Xu. Layerdiff: Exploring text-guided multi-layered composable image synthesis via layer-collaborative diffusion model. In ECCV, 2024. 2", + "[26] Jonghyun Lee, Hansam Cho, Youngjoon Yoo, Seoung Bum Kim, and Yonghyun Jeong. Compose and conquer: Diffusion-based 3d depth aware composable image synthesis. In ICLR, 2024. 2, 17", + "[27] Jaerin Lee, Daniel Sungho Jung, Kanggeon Lee, and Young Mu Lee. StreamMultiDiffusion: real-time interactive generation with region-based semantic control. arXiv:2403.09055, 2024. 2", + "[28] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 2, 13", + "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 6" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7897", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Zhengyao Lv, Yuxiang Wei, Wangmeng Zuo, and KwanYee K Wong. Place: Adaptive layout-semantic fusion for semantic image synthesis. In CVPR, 2024. 2", + "[31] Yiwei Ma, Jiayi Ji, Ke Ye, Weihuang Lin, Zhibin Wang, Yonghan Zheng, Qiang Zhou, Xiaoshuai Sun, and Rongrong Ji. I2ebench: A comprehensive benchmark for instruction-based image editing. In NeurIPS, 2024. 2, 6, 14", + "[32] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: High-resolution and prompt-faithful text-guided image inpainting with diffusion models. In ICLR, 2025. 1, 2, 6, 7, 8, 11, 16, 17, 20", + "[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 17", + "[34] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2", + "[35] Yotam Nitzan, Zongze Wu, Richard Zhang, Eli Shechtman, Daniel Cohen-Or, Taesung Park, and Michael Gharbi. Lazy diffusion transformer for interactive image editing. In ECCV, 2024. 2", + "[36] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 6", + "[37] Pincel. Pincel: Ai image editor, 2024. Accessed: 2024-11-15. 1, 2, 17, 20", + "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 6, 8, 11, 13", + "[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. In ICLR, 2025. 2", + "[40] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 14", + "[41] Jiawei Ren, Mengmeng Xu, Jui-Chieh Wu, Ziwei Liu, Tao Xiang, and Antoine Toisoul. Move anything with layered scene diffusion. In CVPR, 2024. 2", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2", + "[43] Takahiro Shirakawa and Seiichi Uchida. Noisecollapse: A layout-aware text-to-image diffusion model based on noise cropping and merging. In CVPR, 2024. 2, 6, 7", + "[44] Jack Urbanek, Florian Bordes, Pietro Astolfi, Mary Williamson, Vasu Sharma, and Adriana Romero-Soriano. A picture is worth more than 77 text tokens: Evaluating clip-style models on dense captions. In CVPR, 2024. 15" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, 2023. 2, 6, 13, 14", + "[46] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In CVPR, 2024. 2", + "[47] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In ICCV, 2023. 2", + "[48] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 2", + "[49] Chao Yang, Xin Lu, Zhe Lin, Eli Shechtman, Oliver Wang, and Hao Li. High-resolution image inpainting using multiscale neural patch synthesis. In CVPR, 2017. 2", + "[50] Zili Yi, Qiang Tang, Shekoofeh Azizi, Daesik Jang, and Zhan Xu. Contextual residual aggregation for ultra high-resolution image inpainting. In CVPR, 2020.", + "[51] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas Huang. Free-form image inpainting with gated convolution. In ICCV, 2019. 2", + "[52] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. In ECCV, 2024. 15", + "[53] Lisai Zhang, Qingcai Chen, Baotian Hu, and Shuang Jiang. Text-guided neural image inpainting. In ACM MM, 2020. 2", + "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 11, 17", + "[55] Xinyang Zhang, Wentian Zhao, Xin Lu, and Jeff Chien. Text2layer: Layered image generation using latent diffusion model. arXiv:2307.09781, 2023. 2", + "[56] Bo Zhao, Lili Meng, Weidong Yin, and Leonid Sigal. Image generation from layout. In CVPR, 2019. 2", + "[57] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. In NeurIPS, 2023. 17", + "[58] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In CVPR, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 743 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7898", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Improving Editability in Image Generation with Layer-wise Memory/5ef2c05f-1b0c-4899-806c-c1cd90e73e44_model.json b/2025/Improving Editability in Image Generation with Layer-wise Memory/5ef2c05f-1b0c-4899-806c-c1cd90e73e44_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1c5e6a9dccb79d9f801a8d8124afb0d310000997 --- /dev/null +++ b/2025/Improving Editability in Image Generation with Layer-wise Memory/5ef2c05f-1b0c-4899-806c-c1cd90e73e44_model.json @@ -0,0 +1,2772 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.131, + 0.848, + 0.155 + ], + "angle": 0, + "content": "Improving Editability in Image Generation with Layer-wise Memory" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.181, + 0.681, + 0.236 + ], + "angle": 0, + "content": "Daneul Kim Jaeah Lee Jaesik Park* Seoul National University, Republic of Korea {carpedkm, hayanz, jaesik.park}@snu.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.263, + 0.382, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.531, + 0.202, + 0.54 + ], + "angle": 0, + "content": "HD Printer" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.263, + 0.889, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.569, + 0.908, + 0.613 + ], + "angle": 0, + "content": "Figure 1. Overview. Our framework enables the interactive generation of images with enhanced control but in a simple manner, by rough mask and prompt, through iterative scene editing. We utilize the background scene generated by our framework to edit in HD Painter [32] or Blended Latent Diffusion (BLD) [3] for comparison and commercial products like Photoshop [1] and Pincel [37]." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.632, + 0.329, + 0.648 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.486, + 0.877 + ], + "angle": 0, + "content": "Most real-world image editing tasks require multiple sequential edits to achieve desired results. Current editing approaches, primarily designed for single-object modifications, struggle with sequential editing: especially with maintaining previous edits along with adapting new objects naturally into the existing content. These limitations significantly hinder complex editing scenarios where multiple objects need to be modified while preserving their contextual relationships. We address this fundamental challenge through two key proposals: enabling rough mask inputs that preserve existing content while naturally integrating new elements and supporting consistent editing across multiple modifications. Our framework achieves this through layer-wise memory, which stores latent representations and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.909, + 0.801 + ], + "angle": 0, + "content": "prompt embeddings from previous edits. We propose Background Consistency Guidance that leverages memorized latents to maintain scene coherence and Multi-Query Disentanglement in cross-attention that ensures natural adaptation to existing content. To evaluate our method, we present a new benchmark dataset incorporating semantic alignment metrics and interactive editing scenarios. Through comprehensive experiments, we demonstrate superior performance in iterative image editing tasks with minimal user effort, requiring only rough masks while maintaining high-quality results throughout multiple editing steps." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.83, + 0.646, + 0.846 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.857, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recent advances in text-to-image synthesis through powerful diffusion models like Stable Diffusion [20, 42], PixArt [12, 13] and FLUX [11] have transformed visual" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.889, + 0.237, + 0.902 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7889" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "content creation. However, users often demand multiple sequential edits to achieve desired results, iteratively refining and adding elements to their images. Current approaches in inpainting [3, 32] are primarily designed for single-object modifications with limited changes like colors or styles, making them inadequate for complex editing scenarios that involve multiple objects and their interactions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.482, + 0.365 + ], + "angle": 0, + "content": "Image editing approaches, while offering localized modifications, face several limitations. Current methods [3, 16, 32, 48] struggle when modifications need to be applied sequentially, often demanding precise segmentation masks [16] or external modules [39] to maintain background integrity [32]. For example, generating \"A dog sitting\" as shown in Fig. 1 remains challenging, as it requires maintaining previous edits while ensuring the \"dog\" naturally blends with the \"Jeep\" and \"Lego man standing\" - a complex iterative editing scenario that current methods [1, 3, 32, 37] struggle to handle." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.482, + 0.534 + ], + "angle": 0, + "content": "Layout-to-image generation offers an alternative approach through various inputs including bounding boxes [28, 47, 58], depth maps [26, 54], and semantic masks [7, 27]. However, these methods require complete regeneration of the entire image for each modification, making iterative editing particularly cumbersome – unable to maintain the surrounding background. While recent work has explored layered representations [25, 41, 55] and instance-based generation [46], they require complex optimization processes or additional training, limiting their practical applicability in iterative editing scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.734 + ], + "angle": 0, + "content": "We address these limitations through two key innovations. First, we enable object placement aligning with the user's intention using only rough mask inputs while preserving background context. Second, we support consistent iterative editing across multiple modifications. To achieve this, we introduce the concept of mask order, which specifies the sequence of object generation during iterative image editing. In Fig. 1, we add \"a lego man\", \"Jeep\" sequentially with different mask orders, but they naturally adapt into the image, side by side. If we add \"A dog sitting\" with overlapping mask, then it means that we aim to put the \"dog\" in front of the \"Jeep\" and \"a lego man\", making the instance order of \"dog\" in front of \"Jeep\" and \"a lego man\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To handle this, we incorporate three technical components: (1) Layer-wise memory for storing editing history, (2) Background Consistency Guidance (BCG) for maintaining unedited regions, and (3) Multi-Query Disentangle-ment (MQD) in cross-attention for natural object integration. The layer-wise memory stores and manages the latent representations and prompt embeddings from previous editing steps, eliminating redundant computations typical in sequential modifications while maintaining consistency across multiple edits. The BCG not only ensures unedited regions remain stable but also reduces computational over" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "head by avoiding repeated forward passes on the original image, enabling reduced editing time, while MQD enables the natural integration of new objects with existing content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.905, + 0.258 + ], + "angle": 0, + "content": "Additionally, we propose Multi-Edit Bench, a comprehensive benchmark for evaluating iterative image editing capabilities. Prior benchmarks either focus on single-turn edits [31, 45] or layout-to-image generation [4, 17, 21], failing to capture the challenges of sequential modifications. Our benchmark introduces layer-wise semantic evaluation metrics to assess both edit quality and cross-modification consistency in multi-step editing scenarios." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.258, + 0.835, + 0.272 + ], + "angle": 0, + "content": "To summarize, our key contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.274, + 0.905, + 0.304 + ], + "angle": 0, + "content": "- A framework for interactive mask order-based object placement editing with only rough mask inputs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.304, + 0.904, + 0.333 + ], + "angle": 0, + "content": "- A localized editing mechanism using layer-wise memory that maintains consistency across multiple edits." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.904, + 0.363 + ], + "angle": 0, + "content": "- Multi-query disentangled cross attention, allowing natural integration while preserving background context." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.364, + 0.905, + 0.409 + ], + "angle": 0, + "content": "- Multi-Edit Bench, a comprehensive benchmark for evaluating the semantic alignment and sequential iterative editing capabilities." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.274, + 0.905, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.423, + 0.655, + 0.438 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Image inpainting. Image inpainting has evolved from classical patch-based [8, 18] and interpolation-driven [5, 9] methods to powerful deep-learning approaches [49-51] that better handle complex structures. Text-conditioned inpainting [51, 53] further broadened user control, and diffusion-based models [24]—including GLIDE [34], Stable Diffusion [42], and Blended Diffusion [2, 3]—significantly improved realism. Yet, existing works often focus on single-object edits. Recent attempts to relax mask precision [32, 48] or enhance quality [16] remain limited in multi-step scenarios, struggling to maintain consistency across multiple objects or occlusions—hence insufficient for iterative real-world editing tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Layout-to-Image Generation. Controllable image synthesis has recently emphasized spatial arrangement via bounding boxes, segmentation masks, or other layout cues [14, 46, 56]. Advances in diffusion-based layout calibration [23] and grounding [43] have improved fidelity, and interactive methods [15, 30, 35] aim to let users refine elements step by step. However, most such pipelines regenerate the entire image for each new edit, losing previously established context and coherence. By contrast, our approach incrementally updates only relevant regions through a memory-assisted process, naturally preserving previously edited objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Evaluation Benchmark. Most existing editing benchmarks focus on single-turn modifications [31, 45], while layout-centric datasets such as HRS-1k [4] and NSR-1k [21] assess scene quality or prompt alignment in a single-shot manner. Consequently, they overlook the iterative na" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7890" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.895, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.391, + 0.908, + 0.437 + ], + "angle": 0, + "content": "Figure 2. Overview. (a) The left denotes an illustration of how Multi Query Disentanglement is performed in the cross-attention layer. (b) The upper right figure shows Background Consistency Guidance with recalled latents, conducting latent blending with the saved latents. (c) The right below shows the layer-wise memory, saving the previous editing steps' latents, masks, and prompt embeddings." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.452, + 0.487, + 0.606 + ], + "angle": 0, + "content": "ture of real-world edits where object placements evolve over multiple steps. To address this gap, we introduce Multi-Edit Bench, a new benchmark that evaluates multi-step editing in terms of semantic accuracy and visual alignment across the entire editing sequence. This multi-step perspective better captures practical editing workflows, where users iteratively refine scenes by adding, removing, or repositioning objects, thereby measuring a method's capacity to preserve context and maintain coherent compositions over multiple, potentially complex modifications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.622, + 0.185, + 0.637 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.202, + 0.663 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.673, + 0.483, + 0.839 + ], + "angle": 0, + "content": "We propose a framework for iterative image editing with three key components: Layer-wise memory, background consistency guidance (BCG), and multi-query disentangled cross-attention (MQD). Layer-wise memory stores editing history to enable consistent object placement across sequential edits. BCG leverages this stored information for efficient latent blending while preserving background integrity. MQD ensures the natural object integration through disentangling the attention between queries and latents, handling complex scenarios as shown in Fig. 2 like placing \"a dish\" in front of \"a cupcake\" and \"a mug cup\"." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our framework's architecture, built upon transformer-based diffusion models [13], operates through the following systematic workflow: First, we detail the layer-wise memory mechanism (Fig. 2(c)) in Sec. 3.2, which stores editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.561 + ], + "angle": 0, + "content": "history. Then, in Sec. 3.3, we explain how BCG leverages this stored information for efficient latent blending (Fig. 2 (b)). Sec. 3.4 describes MQD's role in ensuring natural object integration through query disentanglement (Fig. 2 (a)). Finally, in Sec. 3.5, we demonstrate how these components enable advanced editing capabilities, including the removal of occluded elements while preserving foreground integrity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.707, + 0.589 + ], + "angle": 0, + "content": "3.2. Layer-wise Memory" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Layer-wise memory enables stable background preservation while ensuring the natural integration of new objects during iterative editing with mask order. By storing and managing key information from each editing step according to the specified mask order, the model can reference previous edits when generating new content while maintaining the intended spatial relationships between objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.909, + 0.824 + ], + "angle": 0, + "content": "Let the set of prompts for generating the background and objects be denoted as \\( P_{l} = \\{p_{0}, p_{1}, p_{2}, \\ldots\\} \\), where \\( p_{0} \\) is the background generation prompt, and \\( p_{1}, p_{2}, \\ldots \\) represent the prompts for each object. Similarly, let the set of corresponding masks be denoted as \\( M_{l} = \\{m_{0}, m_{1}, m_{2}, \\ldots\\} \\), where each \\( m_{i} \\) defines the region of interest (RoI) for the object associated with prompt \\( p_{i} \\). Note that \\( m_{0} \\) represents the case with mask of all elements being 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The layer-wise memory \\( L_{l} = \\{l_{0}, l_{1}, l_{2}, \\ldots \\} \\) stores three key elements for each step \\( i \\): \\( l_{i} = \\{\\mathbf{p}_{i}, \\{\\mathbf{Z}_{i}^{t}\\}_{t=1}^{T}, m_{i}\\} \\), where: \\( \\mathbf{p}_{i} \\) denotes prompt embedding that guides the generation, \\( \\{\\mathbf{Z}_{i}^{t}\\}_{t=1}^{T} \\) as denoising latents across timesteps and \\( m_{i} \\) as a mask defining the object's RoI." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "7891" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.485, + 0.175 + ], + "angle": 0, + "content": "Table 1. Computational comparison in utilizing Background Consistency Guidance (BCG). We utilize PixArt-\\(\\alpha\\) inpaint pipeline with vanilla latent blending and our custom-implemented pipeline with BCG for comparison. We ran the inpainting task for the single-step editing 5 times for each method. Note that we use PixArt-alpha-XL-1024 pretrained weight for the experiment." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.184, + 0.432, + 0.238 + ], + "angle": 0, + "content": "
MethodTime (sec)VRAM
Latent Blending4.1218 ±0.052116.89GB
BCG3.8992±0.014216.90GB
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.484, + 0.354 + ], + "angle": 0, + "content": "For a generation, the initial background latent is created as \\(\\mathbf{Z}_0 = f_\\theta(p_0)\\), where \\(f_\\theta\\) is the pre-trained diffusion model. For each subsequent object \\(p_i\\), we initialize its latent independently: \\(\\mathbf{Z}_i = f_\\theta(p_i, m_i)\\). This independence in initialization, combined with the stored history in layer-wise memory, enables flexible generation while maintaining background consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.484, + 0.445 + ], + "angle": 0, + "content": "The layer-wise memory serves two key purposes: (1) Context preservation, to store the complete editing history, enabling the model to maintain consistency with previously generated content (2) Localized editing, enabling control over specific regions while preserving the surrounding content through mask-guided generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.484, + 0.522 + ], + "angle": 0, + "content": "By maintaining this structured history of edits, the model can effectively handle complex scenarios where multiple objects need to be integrated coherently. The stored information guides background consistency and natural object integration, as detailed in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.53, + 0.398, + 0.546 + ], + "angle": 0, + "content": "3.3. Background Consistency Guidance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Background Consistency Guidance (BCG) addresses a key challenge in iterative image editing: maintaining the integrity of previous edits while incorporating new elements efficiently. Unlike traditional inpainting approaches, BCG leverages layer-wise memory to ensure stability across multiple sequential modifications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.484, + 0.722 + ], + "angle": 0, + "content": "Latent Retrieval from Layer-wise Memory: For each editing step \\( i \\), BCG retrieves information from the previous edits stored in layer-wise memory \\( L_{l} \\), enabling seamless integration of new content while preserving prior modifications:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.723, + 0.483, + 0.742 + ], + "angle": 0, + "content": "\\[\nl _ {i - 1} = \\left\\{\\mathbf {p} _ {i - 1}, \\left\\{\\mathbf {Z} _ {i - 1, t} \\right\\} _ {t = 1} ^ {T}, m _ {i - 1} \\right\\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.484, + 0.838 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_{i - 1}\\) is the prompt embedding from the prior step, \\(\\{\\mathbf{Z}_{i - 1,t}\\}_{t = 0}^{T}\\) represents the stored latents from each denoising step, and \\(m_{i - 1}\\) is the previous mask. This stored latent provides context for the current edit, preserving the background information without requiring a re-computation on the original image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.844, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Ensuring Consistency: When adding a new object, we update only the masked region while preserving the rest:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {i} = \\mathbf {Z} _ {i - 1} \\odot (1 - m _ {i}) + \\mathbf {Z} _ {i} \\odot m _ {i} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "This selective update mechanism is crucial for iterative editing since it ensures that each new modification preserves previous edits. The element-wise multiplication \\(\\odot\\) enables control over which regions are updated, maintaining the integrity of the evolving image composition." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.174, + 0.847, + 0.189 + ], + "angle": 0, + "content": "Computational Efficiency for Iterative Editing:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.189, + 0.907, + 0.37 + ], + "angle": 0, + "content": "BCG improves the efficiency of iterative image editing by reducing redundant computations in sequential modifications. Traditional inpainting methods [3] that utilize latent blending (LB) require a forward pass on the original image for each edit, becoming costly with multiple edits. While both approaches (i.e., LB, BCG) require denoising costs, BCG avoids repeated forward passes. Let \\(\\Omega = \\mathrm{FLOPs}(T,L,H,W)\\) represent the base computational cost for processing an image of size \\(H\\times W\\) through \\(T\\) denoising steps and \\(L\\) model layers, and \\(C_f\\) denote the forward pass cost. The total cost for LB and BCG can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.371, + 0.905, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C o s t} _ {\\mathrm {L B}} = C _ {f} + \\Omega . \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.393, + 0.822, + 0.409 + ], + "angle": 0, + "content": "While for BCG, by skipping the forward pass:" + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.421, + 0.905, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C o s t} _ {\\mathrm {B C G}} = \\Omega , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.448, + 0.902, + 0.464 + ], + "angle": 0, + "content": "Assuming \\(C_f \\approx r\\Omega\\) with ratio \\(r \\in (0,1)\\), we can estimate:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.474, + 0.905, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\text {E f f e c i e n c y} = \\frac {\\operatorname {C o s t} _ {\\mathrm {L B}}}{\\operatorname {C o s t} _ {\\mathrm {B C G}}} \\approx 1 + r \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.906, + 0.591 + ], + "angle": 0, + "content": "In Tab. 1, we observe about \\(10\\%\\) reduction in computational time for single-step editing with BCG compared to LB. This improvement becomes more significant in scenarios requiring multiple sequential modifications, as LB requires multiple forward passes while BCG requires none." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.6, + 0.88, + 0.616 + ], + "angle": 0, + "content": "3.4. Multi-Query Disentangled Cross-Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.907, + 0.773 + ], + "angle": 0, + "content": "Multi-query disentangled cross-attention (MQD) ensures the natural integration of objects across different mask orders while preserving their spatial relationships. For example, when generating \"a dish\" as in Fig. 2, MQD enables the 'dish' to blend naturally into the background, despite being in different mask orders, while maintaining the structure and color of the background with proper occlusion of other objects like 'a mug cup' and 'a cupcake' behind. This is achieved by disentangling attention across multiple queries and leveraging information from layer-wise memory \\( L_{l} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.78, + 0.809, + 0.795 + ], + "angle": 0, + "content": "3.4.1. Cross-Attention for Current Object" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.799, + 0.906, + 0.83 + ], + "angle": 0, + "content": "For each generation step \\(i\\), MQD focuses attention on the current mask order's region of interest (RoI):" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.839, + 0.905, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i} ^ {k, a t t n} = \\operatorname {C r o s s A t t e n t i o n} \\left(\\mathbf {z} _ {i} ^ {k, t} \\odot m _ {i}, p _ {i}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.869, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_i^{k,t}\\) is the latent for the \\(k\\)-th block at denoising step \\(t\\), \\(m_i\\) defines the RoI, and \\(p_i\\) is the prompt embedding. This" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7892" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.089, + 0.568, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.089, + 0.866, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.328, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Figure 3. Overview of our proposed Multi-Edit Benchmark for evaluation of iterative editing scenario. (a) explains the dataset generation pipeline through GPT-4 API, and (b) explains the evaluation methodology in visual alignment using CLIP and semantic alignment using LLaVa for single-image and in a layer-wise manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.445 + ], + "angle": 0, + "content": "ensures focused generation within the current mask order's region. In the Fig. 2, \\( m_{i} \\) is given as a mask corresponding to the \\( p_{i} \\) 'a dish' with \\( \\mathbf{z}_{i}^{k,t} \\) corresponding to the latent of the masked region for 'a dish'." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.45, + 0.484, + 0.466 + ], + "angle": 0, + "content": "3.4.2. Disentangled Attention for Previous Mask Orders" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.515 + ], + "angle": 0, + "content": "To adapt the mask orders into the appropriate instance orders, MQD applies attention to non-overlapping regions from previous steps:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i} ^ {k, a t t n} = \\bigcup_ {j = 0} ^ {i - 1} \\text {C r o s s A t t e n t i o n} \\left(\\mathbf {z} _ {i} ^ {k, t} \\odot \\left(m _ {j} - \\Sigma_ {l = j + 1} ^ {i} m _ {l}\\right), p _ {j}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.748 + ], + "angle": 0, + "content": "This operation ensures that objects from earlier mask orders, (i.e., previous masks for 'a mug cup' and 'a cupcake' on top of 'a wooden floor' behind 'a dish' altogether) remain coherent while allowing natural adaptation where needed. The term \\( m_{j} - \\sum_{l = j + 1}^{i}m_{l} \\) controls which regions from previous mask orders should influence the current generation, minimally affecting the \"mug cup\" and \"cupcake\" even if we put the \"dish\" in front of them. Please note that for background prompt \\( p_0 \\): \"wooden floor\", we inverse the whole union of masks and do disentangled cross attention with background prompt \\( p_0 \\) as in Fig. 2 (a)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.753, + 0.281, + 0.768 + ], + "angle": 0, + "content": "3.4.3. Final Latent Update" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.8 + ], + "angle": 0, + "content": "The final update combines information across all mask orders:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.799, + 0.483, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i} ^ {\\text {m e r g e}} = \\mathbf {z} _ {i} ^ {\\text {a t t n}} \\bigcup \\sum_ {j = 0} ^ {i - 1} \\mathbf {z} _ {j} ^ {\\text {a t t n}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.845, + 0.483, + 0.876 + ], + "angle": 0, + "content": "After a feedforward layer update and \\(K\\) transformer blocks, we apply BCG:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.883, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {i} ^ {t} = \\mathbf {z} _ {i} ^ {K, t} \\odot m _ {i} + \\mathbf {Z} _ {i - 1} ^ {t} \\odot (1 - m _ {i}) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.907, + 0.474 + ], + "angle": 0, + "content": "This approach enables objects to adapt naturally to mask orders—for example, placing “a dog” in front of “Jeep” and “Lego man” (Fig. 1), or maintaining ‘a cupcake’ and ‘a mug cup’ despite inserting “a dish” in front (Fig. 2)—while preserving scene coherence. MQD balances content preservation with the seamless integration of new elements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.485, + 0.719, + 0.502 + ], + "angle": 0, + "content": "3.5. Improving Editability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.599 + ], + "angle": 0, + "content": "While BCG and MQD with layer-wise memory enhance multiple editing capabilities in the PixArt-\\(\\alpha\\) model, we further extend our framework to support object deletion, particularly for overlapped objects behind foreground elements. We achieve this through a novel application of our MQD and BCG components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.661 + ], + "angle": 0, + "content": "To delete an object added at editing step \\(i - 1\\), we leverage the stored latent representations from both steps \\(i - 2\\) and the current step \\(i\\): \\(\\{\\mathbf{Z}_{i - 2}^t\\}_{t = 1}^T\\) and \\(\\{\\mathbf{Z}_i^t\\}\\). Starting from an intermediate denoising step \\(\\tau\\), we blend these latents as:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.678, + 0.907, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {\\text {e r a s e}} ^ {\\tau} = m _ {i} \\odot \\mathbf {Z} _ {i} ^ {\\tau} + (1 - m _ {i}) \\odot \\mathbf {Z} _ {i - 2} ^ {\\tau} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Our deletion process operates in two phases: from step \\(\\tau\\) to \\(\\frac{\\tau}{2}\\), we perform latent blending following Eq. (10) with \\(\\mathbf{Z}_{i-2}^{t}\\). After step \\(\\frac{\\tau}{2}\\), we switch to vanilla denoising to ensure a smooth transition. Crucially, we process \\(\\mathbf{Z}_{\\text{erase}}\\) through MQD without the \\(i-1\\) step's mask or prompt embedding, effectively removing the target object's influence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "This approach offers two key advantages. First, by leveraging BCG to start from step \\(\\tau\\) rather than T, we achieve \\(60\\%\\) faster editing while preserving background consistency, as we use \\(\\tau = 8\\) out of a total of 20 steps. Second, MQD's application from step \\(\\tau\\) maintains foreground object integrity while effectively removing background elements, eliminating the need for precise object masks in deletion." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7893" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 2. Quantitative results on our proposed benchmark. Our method shows higher semantic and visual alignment than other baselines, including layout-to-image synthesis and image editing frameworks. Ordering† denotes the adaptation of the framework adequate for iterative image generation." + }, + { + "type": "table", + "bbox": [ + 0.133, + 0.137, + 0.86, + 0.324 + ], + "angle": 0, + "content": "
TypeMethodResolutionSemantic AlignmentVisual Alignment
BLEU-2/3/4↑METEOR↑CLIPcrop↑
2D Layout-to-ImageLayoutGuidance [14]512 × 51236.44 / 26.13 / 18.850.136162.92
NoiseCollage [43]512 × 51255.75 / 42.43 / 32.960.140264.01
NoiseCollage + ordering†512 × 51259.98 / 43.76 / 32.240.146464.10
3D Layout-to-ImageLooseControl [10]512 × 51263.30 / 46.24 / 34.150.137363.13
Image EditingBLD [3]1024 × 102455.30 / 40.38 / 29.580.148062.40
HD-Painter [32]1024 × 102463.29 / 47.63 / 36.280.148464.09
SD3-Inpaint [20]1024 × 102429.90 / 21.64 / 15.780.144563.98
Ours512 × 51261.19 / 45.04 / 34.060.146564.28
Ours1024 × 102464.99 / 47.69 / 36.590.151364.29
" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.341, + 0.308, + 0.358 + ], + "angle": 0, + "content": "4. Multi-Edit Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.368, + 0.483, + 0.58 + ], + "angle": 0, + "content": "Prior benchmarks for image editing [31, 45] are limited to evaluate only single-step editing or limited to instruct-based edit [22], and layout-guided image generation benchmarks are limited to evaluate whether the specific object is generated or not in the bounding box or evaluate only the flat spatial arrangements. [4, 17, 21]. Therefore, we aim to tackle the issue of evaluating the semantic alignment of the generated images by assessing whether the object generated is aligned with the caption that is used for sequential editing. Also, to handle the mask order, we propose to evaluate the generated image in a layer-wise manner aligning with mask orders. Specifically, to evaluate interactive generation and editing, we need to evaluate whether each layer shows the desired object within the given mask." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.581, + 0.483, + 0.734 + ], + "angle": 0, + "content": "We generate the dataset by selecting classes from ImageNet-1K and arranging them in layered compositions with varying degrees of occlusion and spatial relationships, ensuring mask order is a key factor in the generation process as in the Fig. 3 (a). We utilize GPT-4 API to assist in selecting object classes (e.g., 'Lion', 'Tiger', 'Elephant') that naturally fit together, creating realistic compositions, and generating template-based captions for both the global scene and individual object layers to integrate spatial relations into prompt and utilize GPT to make prompt richer." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.734, + 0.483, + 0.884 + ], + "angle": 0, + "content": "To evaluate the spatial and semantic alignment between generated images and prompts, we crop each object layer using its corresponding mask and evaluate it individually as in Fig. 3 (b). Both cropped and final images are resized to \\(224 \\times 224\\) for consistency across evaluations. We employ LLaVa [29] to generate captions and compute alignment metrics such as BLEU [36], METEOR [6], ensuring a robust evaluation of both spatial arrangement and semantic accuracy. Note that we average the result from all the layers of each image to evaluate." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Also, to check each layer's visual alignment of the gen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.907, + 0.435 + ], + "angle": 0, + "content": "erated object, we crop the image for each iteration and calculate the CLIP score following other layout-to-image models [4, 21, 43] as individual editing step is similar to layout-to-image generation. We discuss and provide detailed descriptions of the dataset creation process, evaluation metrics, and the resulting benchmark in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.446, + 0.64, + 0.464 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.471, + 0.731, + 0.487 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.614 + ], + "angle": 0, + "content": "Baselines. We evaluate our method against state-of-the-art baselines in both image editing and layout-to-image generation. For image editing, we compare with HD-Painter [32], Blended Latent Diffusion [3] with SD-XL [38], and SD3-ControlNet-Inpaint [20]. As sequential editing shares similarities with layout-to-image generation, we also include train-free layout models: NoiseCollage [43] and Layout-Guidance [14], and 3D-lifted model: LooseControl [10]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Additionally, we adapt NoiseCollage to handle sequential mask inputs (denoted as NoiseCollage + ordering) to explore the potential of converting layout-to-image models for sequential editing tasks. Following our train-free approach using the PixArt-\\(\\alpha\\) foundation model [13], all selected baselines operate without additional training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Implementation Details. We leverage PixArt-\\(\\alpha\\) [13], the variant of Diffusion Transformer (DiT) for image editing framework. We utilize a \\(1024 \\times 1024\\) pretrained model with DPM-Solver with DPMS guidance of 7.5 and a total denoising step of 20. Also, for a fair comparison with models that generate \\(512 \\times 512\\) resolution [14, 43], we compare quantitatively by generating \\(512 \\times 512\\) images with the \\(512 \\times 512\\) pretrained PixArt-\\(\\alpha\\) model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.702, + 0.85 + ], + "angle": 0, + "content": "5.2. Quantitative Result" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We present the quantitative evaluation results on our proposed Multi-Edit Benchmark in Tab. 2. Our method consistently outperforms both image editing and layout-to-image" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7894" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.092, + 0.237, + 0.103 + ], + "angle": 0, + "content": "\"Savanna Plains\"" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.106, + 0.238, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.182, + 0.218, + 0.192 + ], + "angle": 0, + "content": "Before Edit" + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.093, + 0.394, + 0.104 + ], + "angle": 0, + "content": "+ \"A flamingo\"" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.106, + 0.349, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.182, + 0.332, + 0.193 + ], + "angle": 0, + "content": "Edit w/o QD" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.106, + 0.451, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.182, + 0.432, + 0.193 + ], + "angle": 0, + "content": "Edit w/QD" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.2, + 0.483, + 0.228 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on the effect of Query Disentanglement (QD)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.349 + ], + "angle": 0, + "content": "belines across key metrics. Among editing baselines, we demonstrate superior performance compared to Blended Latent Diffusion (BLD) [3] with SD-XL, HD-Painter [32], and SD3-ControlNet-Inpaint [20]. Notably, the significant performance gap with SD3-ControlNet-Inpaint can be attributed to the suboptimal property of utilizing ControlNet-Inpainting in multi-step editing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.44 + ], + "angle": 0, + "content": "Our framework shows particular strength in handling sequential editing challenges, demonstrated by consistent improvements across all metrics. This validates our approach's effectiveness in maintaining background consistency while naturally integrating new objects - a key challenge in sequential editing tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.592 + ], + "angle": 0, + "content": "The performance gap grows further relative to 2D layout-to-image baselines, which often struggle with overlapping or cascaded placements. In particular, our method outperforms NoiseCollage by \\(4 - 5\\%\\) p and surpasses Layout-Guidance by over \\(15\\%\\) p in BLEU and outperforms in all other metrics. Against the 3D layout-to-image method that leverages pseudo depth maps, LooseControl [10] shows improved BLEU compared to the 2D baselines, but ours outperforms on all metrics. (See supplement for comparison with attribute editing on LooseControl.)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.484, + 0.73 + ], + "angle": 0, + "content": "While we adapt NoiseCollage [43] for iterative image editing (+ ordering), NoiseCollage + ordering achieves better visual alignment scores, indicating the potential viability of adapted layout-to-image approaches. However, its limited improvement across semantic alignment metrics suggests that simple adaptations are insufficient for complex sequential editing scenarios. This underscores the need for specialized approaches like ours that can effectively handle the unique challenges of iterative image editing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.244, + 0.759 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We present an ablation of our key components (Tab. 3), starting from PixArt- \\(\\alpha\\) [13]. Background consistency guidance (BCG) reuses latents for faster editing (Tab. 1), improving BLEU and CLIP while preserving METEOR. Next, Query-Disentanglement (QD), inspired by NoiseCollage [43], preserves the existing background rather than fully covering the mask, raising BLEU but slightly reducing CLIP. Direct latent blending instead inflates CLIP by overfilling the mask, lacking context (Fig. 4). Finally, extend" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.09, + 0.906, + 0.16 + ], + "angle": 0, + "content": "Table 3. Ablation study of our proposed method. Results are reported on both semantic and visual alignment metrics. We denote vanilla PixArt-\\(\\alpha\\) inpainting model as Baseline. Also we denote QD as Query-Disentanglement without memory, only with the current object and background." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.164, + 0.882, + 0.257 + ], + "angle": 0, + "content": "
MethodSemantic Align.Visual Align.
BLEU-2/3/4↑METEOR↑CLIPc↑
Baseline56.29 / 42.04 / 33.060.158664.05
+BCG60.74 / 46.27 / 35.200.158564.10
+QD62.68 / 46.42 / 35.030.153063.99
Ours64.99 / 47.69 / 36.590.151364.29
" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.264, + 0.67, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.294, + 0.674, + 0.304 + ], + "angle": 0, + "content": "\"A beach\"" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.264, + 0.741, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.294, + 0.751, + 0.304 + ], + "angle": 0, + "content": "+“a blue bus\"" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.264, + 0.813, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.294, + 0.835, + 0.315 + ], + "angle": 0, + "content": "+“a hotel behind the bus\"" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.314, + 0.655, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.407, + 0.655, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.5, + 0.624, + 0.51 + ], + "angle": 0, + "content": "HD Painter" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.315, + 0.777, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.407, + 0.777, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.5, + 0.781, + 0.51 + ], + "angle": 0, + "content": "Blended Latent Diffuion" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.315, + 0.898, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.407, + 0.898, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.501, + 0.851, + 0.51 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.513, + 0.906, + 0.568 + ], + "angle": 0, + "content": "Figure 5. Comparison in image editing capability with latest image editing models. [3, 32] Note that the initial image is generated by our framework, which is equivalent to PixArt- \\(\\alpha\\) [13] with no mask input." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.578, + 0.906, + 0.624 + ], + "angle": 0, + "content": "ing QD to multi-query disentanglement (MQD) with layerwise memory integrates new objects into all prior prompts, achieving higher BLEU-2/3/4 and CLIP scores." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.635, + 0.691, + 0.651 + ], + "angle": 0, + "content": "5.4. Qualitative Result" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.794 + ], + "angle": 0, + "content": "Comparison with Editing Approaches under Interactive Scenario. In Fig. 5, we demonstrate our method's effectiveness in iterative editing compared to state-of-the-art single-step editing approaches [2, 32]. Starting with a background of \"A beach\" generated by vanilla PixArt-\\(\\alpha\\) [13], we sequentially place a bus and a hotel. Our method successfully renders the hotel behind the bus, accurately interpreting both the prompt \"A beach\" and the mask order through MQD to achieve the intended spatial arrangement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.884 + ], + "angle": 0, + "content": "In contrast, Blended Latent Diffusion [3] struggles with the initial placement of \"bus\" in the complex background, while HD-Painter [32] shows inconsistencies in the appearance of \"bus\" (highlighted in red in Fig. 5). Our method maintains background consistency while naturally integrating new objects into the scene." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Improved Editability. Fig. 6 compares our method with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7895" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.09, + 0.462, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.236, + 0.485, + 0.306 + ], + "angle": 0, + "content": "Figure 6. Improved editability of image through Background Consistency Guidance and Multi-Query Disentangled cross attention. Through recycling the previous step's latents, we can remove the object that is behind the foreground object, enabling enhanced editability of the image." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.317, + 0.485, + 0.384 + ], + "angle": 0, + "content": "Table 4. Human preference study with recent literature on image editing on our benchmark. We utilize HD-Painter [32] and Blended Latent Diffusion with SD-XL [38] to compare under iterative editing scenario and compare the preference on a 5-point Likert scale." + }, + { + "type": "table", + "bbox": [ + 0.121, + 0.39, + 0.45, + 0.473 + ], + "angle": 0, + "content": "
MethodBackground ConsistencyNatural AdaptationText-scene Alignment
HD-Painter [32]3.712.813.08
BLD [3]3.432.242.04
Ours4.594.284.49
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.483, + 0.606 + ], + "angle": 0, + "content": "Photoshop 2025 [1]. Placing \"a glass of wine\" or \"a lizard\" is straightforward, but removing \"a glass of wine\" while preserving \"a lizard\" remains challenging for Photoshop's generative fill, which requires precise brush strokes and often leaves artifacts (red box). Our approach simplifies this by leveraging existing masks and prior-step latents to naturally inpaint the background while applying MQD during partial denoising to preserve the lizard's structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Comparison with Existing T2I models. Fig. 7 showcases our method's ability to handle complex multi-object scenarios that challenge conventional text-to-image models. For instance, in row A, only our framework successfully generates the intended composition of a lion behind a flower. Row B and C further demonstrate the challenge for existing T2I approaches to handle complex spatial arrangements, while illustrating our method's capability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.741, + 0.322, + 0.758 + ], + "angle": 0, + "content": "5.5. Human Preference Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.854 + ], + "angle": 0, + "content": "We conduct a comprehensive human evaluation study with 50 participants to assess the quality of our method compared to the latest editing models [3, 32] on our benchmark. The evaluation focuses on three aspects: background consistency, natural adaptation, and text-scene alignment, rated on a 5-point Likert scale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our method outperforms single-step editing approaches [3, 32] across all three categories. These results demonstrate that traditional editing models, while effective" + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.127, + 0.535, + 0.137 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.089, + 0.653, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.089, + 0.77, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.089, + 0.888, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.215, + 0.534, + 0.224 + ], + "angle": 0, + "content": "B" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.175, + 0.653, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.175, + 0.77, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.175, + 0.888, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.303, + 0.534, + 0.311 + ], + "angle": 0, + "content": "C" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.262, + 0.653, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.349, + 0.619, + 0.358 + ], + "angle": 0, + "content": "PIXART-a" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.262, + 0.77, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.349, + 0.756, + 0.358 + ], + "angle": 0, + "content": "Stable Diffusion XL" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.262, + 0.888, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.35, + 0.842, + 0.358 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_footnote", + "bbox": [ + 0.523, + 0.36, + 0.892, + 0.382 + ], + "angle": 0, + "content": "A. A lion stands behind flowers in the desert, positioned next to the water where a palm tree is growing nearby under the bright sun." + }, + { + "type": "image_footnote", + "bbox": [ + 0.523, + 0.383, + 0.89, + 0.404 + ], + "angle": 0, + "content": "B. Santa Claus and a teddy bear are sitting by a campfire in a snowy forest, with a gingerbread house directly behind them under falling snow." + }, + { + "type": "image_footnote", + "bbox": [ + 0.523, + 0.404, + 0.88, + 0.426 + ], + "angle": 0, + "content": "C. A wooden table has a flower pot on the left and some stacked books in the center, with a small statue resting behind the table." + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.36, + 0.892, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.431, + 0.906, + 0.473 + ], + "angle": 0, + "content": "Figure 7. Comparison in interactive scenarios with existing T2I generative models. Stable Diffusion XL [38] and PixArt- \\(\\alpha\\) [13] use text input only." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.906, + 0.577 + ], + "angle": 0, + "content": "for single-step modifications, are insufficient for sequential editing. This underscores the necessity of specialized approaches like ours for handling sequential iterative editing scenarios. Additional human evaluation results comparing our method with text-to-image models [13, 38] are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.593, + 0.634, + 0.608 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.907, + 0.815 + ], + "angle": 0, + "content": "We propose a novel framework for sequential image editing, which poses challenges in maintaining background consistency and seamless integration of new objects into the scenes. Our method includes three key components: Layer-wise Memory to preserve previous content, Background Consistency Guidance to keep the background stable with faster editing, and Multi-Query Disentangled Cross-Attention for natural object adaptation. Experiments on our proposed benchmark dataset show that our approach outperforms state-of-the-art methods in semantic and visual alignment, making it ideal for complex, iterative image editing scenarios. This framework paves the way for future work in improved interactive editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.83, + 0.907, + 0.906 + ], + "angle": 0, + "content": "Limitation and Future Work. Since our approach utilizes image editing, generating multiple objects takes longer, depending on the number of edits. Also, utilizing layer-wise memory requires additional memory costs. We plan to make it more efficient for faster editing in the future." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7896" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.182 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by IITP grant (RS-2021-II211343: AI Graduate School Program at Seoul National University \\((5\\%)\\) and RS-2024-00509257: Global AI Frontier Lab \\((30\\%)\\)) and NRF grant (No.RS-2024-00405857 \\((65\\%)\\)) funded by the Korea government (MSIT)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.198, + 0.188, + 0.214 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.224, + 0.482, + 0.265 + ], + "angle": 0, + "content": "[1] Adobe Inc. Adobe Photoshop. https://www.adobe.com/products/photoshop.html, 2024. Version 2025. 1, 2, 8, 17, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.267, + 0.483, + 0.307 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 2, 7, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.311, + 0.483, + 0.351 + ], + "angle": 0, + "content": "[3] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM TOG, 2023. 1, 2, 4, 6, 7, 8, 11, 13, 17, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.409 + ], + "angle": 0, + "content": "[4] Eslam Mohamed Bakr, Pengzhan Sun, Xiaogian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models. In ICCV, 2023. 2, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.412, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[5] Coloma Ballester, Marcelo Bertalmio, Vicent Caselles, Guillermo Sapiro, and Joan Verdera. Filling-in by joint interpolation of vector fields and gray levels. IEEE TIP, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[6] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In IEEvaluation@ACL, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[7] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.556, + 0.482, + 0.61 + ], + "angle": 0, + "content": "[8] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM TOG, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.614, + 0.482, + 0.653 + ], + "angle": 0, + "content": "[9] Marcelo Bertalmio, Guillermo Sapiro, Vicent Caseles, and Coloma Ballester. Image inpainting. In SIGGRAPH, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.657, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[10] Shariq Farooq Bhat, Niloy J. Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning. In SIGGRAPH, 2024. 6, 7, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.701, + 0.458, + 0.715 + ], + "angle": 0, + "content": "[11] BlackForestLabs. Flux, 2023. Accessed: 2024-11-12. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[12] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- \\(\\sigma\\) : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv:2403.04692, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[13] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 1, 3, 6, 7, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[14] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In WACV, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.224, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[15] Ruidong Chen, Lanjun Wang, Weizhi Nie, Yongdong Zhang, and An-An Liu. Anyscene: Customized image synthesis with composited foreground. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[16] Yifu Chen, Jingwen Chen, Yingwei Pan, Yehao Li, Ting Yao, Zhineng Chen, and Tao Mei. Improving text-guided object inpainting with semantic pre-inpainting. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[17] Jaemin Cho, Linjie Li, Zhengyuan Yang, Zhe Gan, Lijuan Wang, and Mohit Bansal. Diagnostic benchmark and iterative inpainting for layout-guided image generation. In CVPR Workshop, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.239, + 0.905, + 0.28 + ], + "angle": 0, + "content": "[18] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE TIP, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.904, + 0.323 + ], + "angle": 0, + "content": "[19] Abdelrahman Eldisokey and Peter Wonka. Build-a-scene: Interactive 3d layout control for diffusion-based image generation. In ICLR, 2025. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.326, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[20] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 6, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.424, + 0.905, + 0.493 + ], + "angle": 0, + "content": "[21] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. In NeurIPS, 2023. 2, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.904, + 0.538 + ], + "angle": 0, + "content": "[22] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv:2405.04007, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.905, + 0.595 + ], + "angle": 0, + "content": "[23] Biao Gong, Siteng Huang, Yutong Feng, Shiwei Zhang, Yuyuan Li, and Yu Liu. Check locate rectify: A training-free layout calibration system for text-to-image generation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.598, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.629, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[25] Runhui Huang, Kaixin Cai, Jianhua Han, Xiaodan Liang, Renjing Pei, Guansong Lu, Songcen Xu, Wei Zhang, and Hang Xu. Layerdiff: Exploring text-guided multi-layered composable image synthesis via layer-collaborative diffusion model. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.7, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[26] Jonghyun Lee, Hansam Cho, Youngjoon Yoo, Seoung Bum Kim, and Yonghyun Jeong. Compose and conquer: Diffusion-based 3d depth aware composable image synthesis. In ICLR, 2024. 2, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[27] Jaerin Lee, Daniel Sungho Jung, Kanggeon Lee, and Young Mu Lee. StreamMultiDiffusion: real-time interactive generation with region-based semantic control. arXiv:2403.09055, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[28] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.904, + 0.901 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7897" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[30] Zhengyao Lv, Yuxiang Wei, Wangmeng Zuo, and KwanYee K Wong. Place: Adaptive layout-semantic fusion for semantic image synthesis. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[31] Yiwei Ma, Jiayi Ji, Ke Ye, Weihuang Lin, Zhibin Wang, Yonghan Zheng, Qiang Zhou, Xiaoshuai Sun, and Rongrong Ji. I2ebench: A comprehensive benchmark for instruction-based image editing. In NeurIPS, 2024. 2, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[32] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: High-resolution and prompt-faithful text-guided image inpainting with diffusion models. In ICLR, 2025. 1, 2, 6, 7, 8, 11, 16, 17, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[34] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[35] Yotam Nitzan, Zongze Wu, Richard Zhang, Eli Shechtman, Daniel Cohen-Or, Taesung Park, and Michael Gharbi. Lazy diffusion transformer for interactive image editing. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[36] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[37] Pincel. Pincel: Ai image editor, 2024. Accessed: 2024-11-15. 1, 2, 17, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 6, 8, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. In ICLR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[40] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[41] Jiawei Ren, Mengmeng Xu, Jui-Chieh Wu, Ziwei Liu, Tao Xiang, and Antoine Toisoul. Move anything with layered scene diffusion. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[43] Takahiro Shirakawa and Seiichi Uchida. Noisecollapse: A layout-aware text-to-image diffusion model based on noise cropping and merging. In CVPR, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Jack Urbanek, Florian Bordes, Pietro Astolfi, Mary Williamson, Vasu Sharma, and Adriana Romero-Soriano. A picture is worth more than 77 text tokens: Evaluating clip-style models on dense captions. In CVPR, 2024. 15" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[45] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, 2023. 2, 6, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[46] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[47] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.276, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[48] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.318, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[49] Chao Yang, Xin Lu, Zhe Lin, Eli Shechtman, Oliver Wang, and Hao Li. High-resolution image inpainting using multiscale neural patch synthesis. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[50] Zili Yi, Qiang Tang, Shekoofeh Azizi, Daesik Jang, and Zhan Xu. Contextual residual aggregation for ultra high-resolution image inpainting. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[51] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas Huang. Free-form image inpainting with gated convolution. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[52] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. In ECCV, 2024. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.904, + 0.518 + ], + "angle": 0, + "content": "[53] Lisai Zhang, Qingcai Chen, Baotian Hu, and Shuang Jiang. Text-guided neural image inpainting. In ACM MM, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 11, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[55] Xinyang Zhang, Wentian Zhao, Xin Lu, and Jeff Chien. Text2layer: Layered image generation using latent diffusion model. arXiv:2307.09781, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.633 + ], + "angle": 0, + "content": "[56] Bo Zhao, Lili Meng, Weidong Yin, and Leonid Sigal. Image generation from layout. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.634, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[57] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. In NeurIPS, 2023. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[58] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In CVPR, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7898" + } + ] +] \ No newline at end of file diff --git a/2025/Improving Editability in Image Generation with Layer-wise Memory/5ef2c05f-1b0c-4899-806c-c1cd90e73e44_origin.pdf b/2025/Improving Editability in Image Generation with Layer-wise Memory/5ef2c05f-1b0c-4899-806c-c1cd90e73e44_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d170d6523e0d19eeb373d3ce5bc1dc455716046c --- /dev/null +++ b/2025/Improving Editability in Image Generation with Layer-wise Memory/5ef2c05f-1b0c-4899-806c-c1cd90e73e44_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b5e9d756c08b2aa2dc35a098be1f13631cc4ac3cd351e1c34c4e3e366ae4eef +size 6605526 diff --git a/2025/Improving Editability in Image Generation with Layer-wise Memory/full.md b/2025/Improving Editability in Image Generation with Layer-wise Memory/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8ac282a30b5fe9962eac7ac883649aac35ed7cd2 --- /dev/null +++ b/2025/Improving Editability in Image Generation with Layer-wise Memory/full.md @@ -0,0 +1,396 @@ +# Improving Editability in Image Generation with Layer-wise Memory + +Daneul Kim Jaeah Lee Jaesik Park* Seoul National University, Republic of Korea {carpedkm, hayanz, jaesik.park}@snu.ac.kr + +![](images/bfdc191bb47f13184c6baed485a30d1b8264ea7d2796c3a0a27587b65ff39f0d.jpg) +HD Printer + +![](images/38a98b819d90e6d9ed2a72adb09c94192d0582d37fd87db48876af5760451c95.jpg) +Figure 1. Overview. Our framework enables the interactive generation of images with enhanced control but in a simple manner, by rough mask and prompt, through iterative scene editing. We utilize the background scene generated by our framework to edit in HD Painter [32] or Blended Latent Diffusion (BLD) [3] for comparison and commercial products like Photoshop [1] and Pincel [37]. + +# Abstract + +Most real-world image editing tasks require multiple sequential edits to achieve desired results. Current editing approaches, primarily designed for single-object modifications, struggle with sequential editing: especially with maintaining previous edits along with adapting new objects naturally into the existing content. These limitations significantly hinder complex editing scenarios where multiple objects need to be modified while preserving their contextual relationships. We address this fundamental challenge through two key proposals: enabling rough mask inputs that preserve existing content while naturally integrating new elements and supporting consistent editing across multiple modifications. Our framework achieves this through layer-wise memory, which stores latent representations and + +prompt embeddings from previous edits. We propose Background Consistency Guidance that leverages memorized latents to maintain scene coherence and Multi-Query Disentanglement in cross-attention that ensures natural adaptation to existing content. To evaluate our method, we present a new benchmark dataset incorporating semantic alignment metrics and interactive editing scenarios. Through comprehensive experiments, we demonstrate superior performance in iterative image editing tasks with minimal user effort, requiring only rough masks while maintaining high-quality results throughout multiple editing steps. + +# 1. Introduction + +Recent advances in text-to-image synthesis through powerful diffusion models like Stable Diffusion [20, 42], PixArt [12, 13] and FLUX [11] have transformed visual + +content creation. However, users often demand multiple sequential edits to achieve desired results, iteratively refining and adding elements to their images. Current approaches in inpainting [3, 32] are primarily designed for single-object modifications with limited changes like colors or styles, making them inadequate for complex editing scenarios that involve multiple objects and their interactions. + +Image editing approaches, while offering localized modifications, face several limitations. Current methods [3, 16, 32, 48] struggle when modifications need to be applied sequentially, often demanding precise segmentation masks [16] or external modules [39] to maintain background integrity [32]. For example, generating "A dog sitting" as shown in Fig. 1 remains challenging, as it requires maintaining previous edits while ensuring the "dog" naturally blends with the "Jeep" and "Lego man standing" - a complex iterative editing scenario that current methods [1, 3, 32, 37] struggle to handle. + +Layout-to-image generation offers an alternative approach through various inputs including bounding boxes [28, 47, 58], depth maps [26, 54], and semantic masks [7, 27]. However, these methods require complete regeneration of the entire image for each modification, making iterative editing particularly cumbersome – unable to maintain the surrounding background. While recent work has explored layered representations [25, 41, 55] and instance-based generation [46], they require complex optimization processes or additional training, limiting their practical applicability in iterative editing scenarios. + +We address these limitations through two key innovations. First, we enable object placement aligning with the user's intention using only rough mask inputs while preserving background context. Second, we support consistent iterative editing across multiple modifications. To achieve this, we introduce the concept of mask order, which specifies the sequence of object generation during iterative image editing. In Fig. 1, we add "a lego man", "Jeep" sequentially with different mask orders, but they naturally adapt into the image, side by side. If we add "A dog sitting" with overlapping mask, then it means that we aim to put the "dog" in front of the "Jeep" and "a lego man", making the instance order of "dog" in front of "Jeep" and "a lego man". + +To handle this, we incorporate three technical components: (1) Layer-wise memory for storing editing history, (2) Background Consistency Guidance (BCG) for maintaining unedited regions, and (3) Multi-Query Disentangle-ment (MQD) in cross-attention for natural object integration. The layer-wise memory stores and manages the latent representations and prompt embeddings from previous editing steps, eliminating redundant computations typical in sequential modifications while maintaining consistency across multiple edits. The BCG not only ensures unedited regions remain stable but also reduces computational over + +head by avoiding repeated forward passes on the original image, enabling reduced editing time, while MQD enables the natural integration of new objects with existing content. + +Additionally, we propose Multi-Edit Bench, a comprehensive benchmark for evaluating iterative image editing capabilities. Prior benchmarks either focus on single-turn edits [31, 45] or layout-to-image generation [4, 17, 21], failing to capture the challenges of sequential modifications. Our benchmark introduces layer-wise semantic evaluation metrics to assess both edit quality and cross-modification consistency in multi-step editing scenarios. + +To summarize, our key contributions include: + +- A framework for interactive mask order-based object placement editing with only rough mask inputs. +- A localized editing mechanism using layer-wise memory that maintains consistency across multiple edits. +- Multi-query disentangled cross attention, allowing natural integration while preserving background context. +- Multi-Edit Bench, a comprehensive benchmark for evaluating the semantic alignment and sequential iterative editing capabilities. + +# 2. Related Work + +Image inpainting. Image inpainting has evolved from classical patch-based [8, 18] and interpolation-driven [5, 9] methods to powerful deep-learning approaches [49-51] that better handle complex structures. Text-conditioned inpainting [51, 53] further broadened user control, and diffusion-based models [24]—including GLIDE [34], Stable Diffusion [42], and Blended Diffusion [2, 3]—significantly improved realism. Yet, existing works often focus on single-object edits. Recent attempts to relax mask precision [32, 48] or enhance quality [16] remain limited in multi-step scenarios, struggling to maintain consistency across multiple objects or occlusions—hence insufficient for iterative real-world editing tasks. + +Layout-to-Image Generation. Controllable image synthesis has recently emphasized spatial arrangement via bounding boxes, segmentation masks, or other layout cues [14, 46, 56]. Advances in diffusion-based layout calibration [23] and grounding [43] have improved fidelity, and interactive methods [15, 30, 35] aim to let users refine elements step by step. However, most such pipelines regenerate the entire image for each new edit, losing previously established context and coherence. By contrast, our approach incrementally updates only relevant regions through a memory-assisted process, naturally preserving previously edited objects. + +Evaluation Benchmark. Most existing editing benchmarks focus on single-turn modifications [31, 45], while layout-centric datasets such as HRS-1k [4] and NSR-1k [21] assess scene quality or prompt alignment in a single-shot manner. Consequently, they overlook the iterative na + +![](images/6e69cc2a32f69fc97a956edd7e5b71fd05cdaf6d233177b964c9a6eaafbd42df.jpg) +Figure 2. Overview. (a) The left denotes an illustration of how Multi Query Disentanglement is performed in the cross-attention layer. (b) The upper right figure shows Background Consistency Guidance with recalled latents, conducting latent blending with the saved latents. (c) The right below shows the layer-wise memory, saving the previous editing steps' latents, masks, and prompt embeddings. + +ture of real-world edits where object placements evolve over multiple steps. To address this gap, we introduce Multi-Edit Bench, a new benchmark that evaluates multi-step editing in terms of semantic accuracy and visual alignment across the entire editing sequence. This multi-step perspective better captures practical editing workflows, where users iteratively refine scenes by adding, removing, or repositioning objects, thereby measuring a method's capacity to preserve context and maintain coherent compositions over multiple, potentially complex modifications. + +# 3. Method + +# 3.1. Overview + +We propose a framework for iterative image editing with three key components: Layer-wise memory, background consistency guidance (BCG), and multi-query disentangled cross-attention (MQD). Layer-wise memory stores editing history to enable consistent object placement across sequential edits. BCG leverages this stored information for efficient latent blending while preserving background integrity. MQD ensures the natural object integration through disentangling the attention between queries and latents, handling complex scenarios as shown in Fig. 2 like placing "a dish" in front of "a cupcake" and "a mug cup". + +Our framework's architecture, built upon transformer-based diffusion models [13], operates through the following systematic workflow: First, we detail the layer-wise memory mechanism (Fig. 2(c)) in Sec. 3.2, which stores editing + +history. Then, in Sec. 3.3, we explain how BCG leverages this stored information for efficient latent blending (Fig. 2 (b)). Sec. 3.4 describes MQD's role in ensuring natural object integration through query disentanglement (Fig. 2 (a)). Finally, in Sec. 3.5, we demonstrate how these components enable advanced editing capabilities, including the removal of occluded elements while preserving foreground integrity. + +# 3.2. Layer-wise Memory + +Layer-wise memory enables stable background preservation while ensuring the natural integration of new objects during iterative editing with mask order. By storing and managing key information from each editing step according to the specified mask order, the model can reference previous edits when generating new content while maintaining the intended spatial relationships between objects. + +Let the set of prompts for generating the background and objects be denoted as $P_{l} = \{p_{0}, p_{1}, p_{2}, \ldots\}$ , where $p_{0}$ is the background generation prompt, and $p_{1}, p_{2}, \ldots$ represent the prompts for each object. Similarly, let the set of corresponding masks be denoted as $M_{l} = \{m_{0}, m_{1}, m_{2}, \ldots\}$ , where each $m_{i}$ defines the region of interest (RoI) for the object associated with prompt $p_{i}$ . Note that $m_{0}$ represents the case with mask of all elements being 1. + +The layer-wise memory $L_{l} = \{l_{0}, l_{1}, l_{2}, \ldots \}$ stores three key elements for each step $i$ : $l_{i} = \{\mathbf{p}_{i}, \{\mathbf{Z}_{i}^{t}\}_{t=1}^{T}, m_{i}\}$ , where: $\mathbf{p}_{i}$ denotes prompt embedding that guides the generation, $\{\mathbf{Z}_{i}^{t}\}_{t=1}^{T}$ as denoising latents across timesteps and $m_{i}$ as a mask defining the object's RoI. + +Table 1. Computational comparison in utilizing Background Consistency Guidance (BCG). We utilize PixArt- $\alpha$ inpaint pipeline with vanilla latent blending and our custom-implemented pipeline with BCG for comparison. We ran the inpainting task for the single-step editing 5 times for each method. Note that we use PixArt-alpha-XL-1024 pretrained weight for the experiment. + +
MethodTime (sec)VRAM
Latent Blending4.1218 ±0.052116.89GB
BCG3.8992±0.014216.90GB
+ +For a generation, the initial background latent is created as $\mathbf{Z}_0 = f_\theta(p_0)$ , where $f_\theta$ is the pre-trained diffusion model. For each subsequent object $p_i$ , we initialize its latent independently: $\mathbf{Z}_i = f_\theta(p_i, m_i)$ . This independence in initialization, combined with the stored history in layer-wise memory, enables flexible generation while maintaining background consistency. + +The layer-wise memory serves two key purposes: (1) Context preservation, to store the complete editing history, enabling the model to maintain consistency with previously generated content (2) Localized editing, enabling control over specific regions while preserving the surrounding content through mask-guided generation + +By maintaining this structured history of edits, the model can effectively handle complex scenarios where multiple objects need to be integrated coherently. The stored information guides background consistency and natural object integration, as detailed in the following sections. + +# 3.3. Background Consistency Guidance + +Background Consistency Guidance (BCG) addresses a key challenge in iterative image editing: maintaining the integrity of previous edits while incorporating new elements efficiently. Unlike traditional inpainting approaches, BCG leverages layer-wise memory to ensure stability across multiple sequential modifications. + +Latent Retrieval from Layer-wise Memory: For each editing step $i$ , BCG retrieves information from the previous edits stored in layer-wise memory $L_{l}$ , enabling seamless integration of new content while preserving prior modifications: + +$$ +l _ {i - 1} = \left\{\mathbf {p} _ {i - 1}, \left\{\mathbf {Z} _ {i - 1, t} \right\} _ {t = 1} ^ {T}, m _ {i - 1} \right\} \tag {1} +$$ + +where $\mathbf{p}_{i - 1}$ is the prompt embedding from the prior step, $\{\mathbf{Z}_{i - 1,t}\}_{t = 0}^{T}$ represents the stored latents from each denoising step, and $m_{i - 1}$ is the previous mask. This stored latent provides context for the current edit, preserving the background information without requiring a re-computation on the original image. + +Ensuring Consistency: When adding a new object, we update only the masked region while preserving the rest: + +$$ +\mathbf {Z} _ {i} = \mathbf {Z} _ {i - 1} \odot (1 - m _ {i}) + \mathbf {Z} _ {i} \odot m _ {i} \tag {2} +$$ + +This selective update mechanism is crucial for iterative editing since it ensures that each new modification preserves previous edits. The element-wise multiplication $\odot$ enables control over which regions are updated, maintaining the integrity of the evolving image composition. + +# Computational Efficiency for Iterative Editing: + +BCG improves the efficiency of iterative image editing by reducing redundant computations in sequential modifications. Traditional inpainting methods [3] that utilize latent blending (LB) require a forward pass on the original image for each edit, becoming costly with multiple edits. While both approaches (i.e., LB, BCG) require denoising costs, BCG avoids repeated forward passes. Let $\Omega = \mathrm{FLOPs}(T,L,H,W)$ represent the base computational cost for processing an image of size $H\times W$ through $T$ denoising steps and $L$ model layers, and $C_f$ denote the forward pass cost. The total cost for LB and BCG can be expressed as: + +$$ +\operatorname {C o s t} _ {\mathrm {L B}} = C _ {f} + \Omega . \tag {3} +$$ + +While for BCG, by skipping the forward pass: + +$$ +\operatorname {C o s t} _ {\mathrm {B C G}} = \Omega , \tag {4} +$$ + +Assuming $C_f \approx r\Omega$ with ratio $r \in (0,1)$ , we can estimate: + +$$ +\text {E f f e c i e n c y} = \frac {\operatorname {C o s t} _ {\mathrm {L B}}}{\operatorname {C o s t} _ {\mathrm {B C G}}} \approx 1 + r \tag {5} +$$ + +In Tab. 1, we observe about $10\%$ reduction in computational time for single-step editing with BCG compared to LB. This improvement becomes more significant in scenarios requiring multiple sequential modifications, as LB requires multiple forward passes while BCG requires none. + +# 3.4. Multi-Query Disentangled Cross-Attention + +Multi-query disentangled cross-attention (MQD) ensures the natural integration of objects across different mask orders while preserving their spatial relationships. For example, when generating "a dish" as in Fig. 2, MQD enables the 'dish' to blend naturally into the background, despite being in different mask orders, while maintaining the structure and color of the background with proper occlusion of other objects like 'a mug cup' and 'a cupcake' behind. This is achieved by disentangling attention across multiple queries and leveraging information from layer-wise memory $L_{l}$ . + +# 3.4.1. Cross-Attention for Current Object + +For each generation step $i$ , MQD focuses attention on the current mask order's region of interest (RoI): + +$$ +\mathbf {z} _ {i} ^ {k, a t t n} = \operatorname {C r o s s A t t e n t i o n} \left(\mathbf {z} _ {i} ^ {k, t} \odot m _ {i}, p _ {i}\right) \tag {6} +$$ + +where $\mathbf{z}_i^{k,t}$ is the latent for the $k$ -th block at denoising step $t$ , $m_i$ defines the RoI, and $p_i$ is the prompt embedding. This + +![](images/71881933e23c4b28840b6b3a6557af2998aea9131e9970696fe1d5f34a961af2.jpg) +Figure 3. Overview of our proposed Multi-Edit Benchmark for evaluation of iterative editing scenario. (a) explains the dataset generation pipeline through GPT-4 API, and (b) explains the evaluation methodology in visual alignment using CLIP and semantic alignment using LLaVa for single-image and in a layer-wise manner. + +![](images/8012102f00fda90e3434956ea080229a5b167e8819351e06040df0be5926b094.jpg) + +ensures focused generation within the current mask order's region. In the Fig. 2, $m_{i}$ is given as a mask corresponding to the $p_{i}$ 'a dish' with $\mathbf{z}_{i}^{k,t}$ corresponding to the latent of the masked region for 'a dish'. + +# 3.4.2. Disentangled Attention for Previous Mask Orders + +To adapt the mask orders into the appropriate instance orders, MQD applies attention to non-overlapping regions from previous steps: + +$$ +\mathbf {z} _ {i} ^ {k, a t t n} = \bigcup_ {j = 0} ^ {i - 1} \text {C r o s s A t t e n t i o n} \left(\mathbf {z} _ {i} ^ {k, t} \odot \left(m _ {j} - \Sigma_ {l = j + 1} ^ {i} m _ {l}\right), p _ {j}\right) \tag {7} +$$ + +This operation ensures that objects from earlier mask orders, (i.e., previous masks for 'a mug cup' and 'a cupcake' on top of 'a wooden floor' behind 'a dish' altogether) remain coherent while allowing natural adaptation where needed. The term $m_{j} - \sum_{l = j + 1}^{i}m_{l}$ controls which regions from previous mask orders should influence the current generation, minimally affecting the "mug cup" and "cupcake" even if we put the "dish" in front of them. Please note that for background prompt $p_0$ : "wooden floor", we inverse the whole union of masks and do disentangled cross attention with background prompt $p_0$ as in Fig. 2 (a). + +# 3.4.3. Final Latent Update + +The final update combines information across all mask orders: + +$$ +\mathbf {z} _ {i} ^ {\text {m e r g e}} = \mathbf {z} _ {i} ^ {\text {a t t n}} \bigcup \sum_ {j = 0} ^ {i - 1} \mathbf {z} _ {j} ^ {\text {a t t n}} \tag {8} +$$ + +After a feedforward layer update and $K$ transformer blocks, we apply BCG: + +$$ +\mathbf {Z} _ {i} ^ {t} = \mathbf {z} _ {i} ^ {K, t} \odot m _ {i} + \mathbf {Z} _ {i - 1} ^ {t} \odot (1 - m _ {i}) \tag {9} +$$ + +This approach enables objects to adapt naturally to mask orders—for example, placing “a dog” in front of “Jeep” and “Lego man” (Fig. 1), or maintaining ‘a cupcake’ and ‘a mug cup’ despite inserting “a dish” in front (Fig. 2)—while preserving scene coherence. MQD balances content preservation with the seamless integration of new elements. + +# 3.5. Improving Editability + +While BCG and MQD with layer-wise memory enhance multiple editing capabilities in the PixArt- $\alpha$ model, we further extend our framework to support object deletion, particularly for overlapped objects behind foreground elements. We achieve this through a novel application of our MQD and BCG components. + +To delete an object added at editing step $i - 1$ , we leverage the stored latent representations from both steps $i - 2$ and the current step $i$ : $\{\mathbf{Z}_{i - 2}^t\}_{t = 1}^T$ and $\{\mathbf{Z}_i^t\}$ . Starting from an intermediate denoising step $\tau$ , we blend these latents as: + +$$ +\mathbf {Z} _ {\text {e r a s e}} ^ {\tau} = m _ {i} \odot \mathbf {Z} _ {i} ^ {\tau} + (1 - m _ {i}) \odot \mathbf {Z} _ {i - 2} ^ {\tau} \tag {10} +$$ + +Our deletion process operates in two phases: from step $\tau$ to $\frac{\tau}{2}$ , we perform latent blending following Eq. (10) with $\mathbf{Z}_{i-2}^{t}$ . After step $\frac{\tau}{2}$ , we switch to vanilla denoising to ensure a smooth transition. Crucially, we process $\mathbf{Z}_{\text{erase}}$ through MQD without the $i-1$ step's mask or prompt embedding, effectively removing the target object's influence. + +This approach offers two key advantages. First, by leveraging BCG to start from step $\tau$ rather than T, we achieve $60\%$ faster editing while preserving background consistency, as we use $\tau = 8$ out of a total of 20 steps. Second, MQD's application from step $\tau$ maintains foreground object integrity while effectively removing background elements, eliminating the need for precise object masks in deletion. + +Table 2. Quantitative results on our proposed benchmark. Our method shows higher semantic and visual alignment than other baselines, including layout-to-image synthesis and image editing frameworks. Ordering† denotes the adaptation of the framework adequate for iterative image generation. + +
TypeMethodResolutionSemantic AlignmentVisual Alignment
BLEU-2/3/4↑METEOR↑CLIPcrop↑
2D Layout-to-ImageLayoutGuidance [14]512 × 51236.44 / 26.13 / 18.850.136162.92
NoiseCollage [43]512 × 51255.75 / 42.43 / 32.960.140264.01
NoiseCollage + ordering†512 × 51259.98 / 43.76 / 32.240.146464.10
3D Layout-to-ImageLooseControl [10]512 × 51263.30 / 46.24 / 34.150.137363.13
Image EditingBLD [3]1024 × 102455.30 / 40.38 / 29.580.148062.40
HD-Painter [32]1024 × 102463.29 / 47.63 / 36.280.148464.09
SD3-Inpaint [20]1024 × 102429.90 / 21.64 / 15.780.144563.98
Ours512 × 51261.19 / 45.04 / 34.060.146564.28
Ours1024 × 102464.99 / 47.69 / 36.590.151364.29
+ +# 4. Multi-Edit Benchmark + +Prior benchmarks for image editing [31, 45] are limited to evaluate only single-step editing or limited to instruct-based edit [22], and layout-guided image generation benchmarks are limited to evaluate whether the specific object is generated or not in the bounding box or evaluate only the flat spatial arrangements. [4, 17, 21]. Therefore, we aim to tackle the issue of evaluating the semantic alignment of the generated images by assessing whether the object generated is aligned with the caption that is used for sequential editing. Also, to handle the mask order, we propose to evaluate the generated image in a layer-wise manner aligning with mask orders. Specifically, to evaluate interactive generation and editing, we need to evaluate whether each layer shows the desired object within the given mask. + +We generate the dataset by selecting classes from ImageNet-1K and arranging them in layered compositions with varying degrees of occlusion and spatial relationships, ensuring mask order is a key factor in the generation process as in the Fig. 3 (a). We utilize GPT-4 API to assist in selecting object classes (e.g., 'Lion', 'Tiger', 'Elephant') that naturally fit together, creating realistic compositions, and generating template-based captions for both the global scene and individual object layers to integrate spatial relations into prompt and utilize GPT to make prompt richer. + +To evaluate the spatial and semantic alignment between generated images and prompts, we crop each object layer using its corresponding mask and evaluate it individually as in Fig. 3 (b). Both cropped and final images are resized to $224 \times 224$ for consistency across evaluations. We employ LLaVa [29] to generate captions and compute alignment metrics such as BLEU [36], METEOR [6], ensuring a robust evaluation of both spatial arrangement and semantic accuracy. Note that we average the result from all the layers of each image to evaluate. + +Also, to check each layer's visual alignment of the gen + +erated object, we crop the image for each iteration and calculate the CLIP score following other layout-to-image models [4, 21, 43] as individual editing step is similar to layout-to-image generation. We discuss and provide detailed descriptions of the dataset creation process, evaluation metrics, and the resulting benchmark in the supplement. + +# 5. Experiment + +# 5.1. Implementation Details + +Baselines. We evaluate our method against state-of-the-art baselines in both image editing and layout-to-image generation. For image editing, we compare with HD-Painter [32], Blended Latent Diffusion [3] with SD-XL [38], and SD3-ControlNet-Inpaint [20]. As sequential editing shares similarities with layout-to-image generation, we also include train-free layout models: NoiseCollage [43] and Layout-Guidance [14], and 3D-lifted model: LooseControl [10]. + +Additionally, we adapt NoiseCollage to handle sequential mask inputs (denoted as NoiseCollage + ordering) to explore the potential of converting layout-to-image models for sequential editing tasks. Following our train-free approach using the PixArt- $\alpha$ foundation model [13], all selected baselines operate without additional training. + +Implementation Details. We leverage PixArt- $\alpha$ [13], the variant of Diffusion Transformer (DiT) for image editing framework. We utilize a $1024 \times 1024$ pretrained model with DPM-Solver with DPMS guidance of 7.5 and a total denoising step of 20. Also, for a fair comparison with models that generate $512 \times 512$ resolution [14, 43], we compare quantitatively by generating $512 \times 512$ images with the $512 \times 512$ pretrained PixArt- $\alpha$ model. + +# 5.2. Quantitative Result + +We present the quantitative evaluation results on our proposed Multi-Edit Benchmark in Tab. 2. Our method consistently outperforms both image editing and layout-to-image + +![](images/3349e8ef666862b985cc620579d0658329eebd9705450015d2491e7c56b2c033.jpg) +"Savanna Plains" +Before Edit +Figure 4. Qualitative comparison on the effect of Query Disentanglement (QD). + +![](images/10ea1e082042a4aa2b7b2f212a08a4ab700d25e78bbc0e0944af5c7fa5e02ded.jpg) ++ "A flamingo" +Edit w/o QD + +![](images/16f830dd65b1ee8fcdd96945bfbf27381c746f21c392c5921902cdd896245abc.jpg) +Edit w/QD + +belines across key metrics. Among editing baselines, we demonstrate superior performance compared to Blended Latent Diffusion (BLD) [3] with SD-XL, HD-Painter [32], and SD3-ControlNet-Inpaint [20]. Notably, the significant performance gap with SD3-ControlNet-Inpaint can be attributed to the suboptimal property of utilizing ControlNet-Inpainting in multi-step editing. + +Our framework shows particular strength in handling sequential editing challenges, demonstrated by consistent improvements across all metrics. This validates our approach's effectiveness in maintaining background consistency while naturally integrating new objects - a key challenge in sequential editing tasks. + +The performance gap grows further relative to 2D layout-to-image baselines, which often struggle with overlapping or cascaded placements. In particular, our method outperforms NoiseCollage by $4 - 5\%$ p and surpasses Layout-Guidance by over $15\%$ p in BLEU and outperforms in all other metrics. Against the 3D layout-to-image method that leverages pseudo depth maps, LooseControl [10] shows improved BLEU compared to the 2D baselines, but ours outperforms on all metrics. (See supplement for comparison with attribute editing on LooseControl.) + +While we adapt NoiseCollage [43] for iterative image editing (+ ordering), NoiseCollage + ordering achieves better visual alignment scores, indicating the potential viability of adapted layout-to-image approaches. However, its limited improvement across semantic alignment metrics suggests that simple adaptations are insufficient for complex sequential editing scenarios. This underscores the need for specialized approaches like ours that can effectively handle the unique challenges of iterative image editing. + +# 5.3. Ablation Study + +We present an ablation of our key components (Tab. 3), starting from PixArt- $\alpha$ [13]. Background consistency guidance (BCG) reuses latents for faster editing (Tab. 1), improving BLEU and CLIP while preserving METEOR. Next, Query-Disentanglement (QD), inspired by NoiseCollage [43], preserves the existing background rather than fully covering the mask, raising BLEU but slightly reducing CLIP. Direct latent blending instead inflates CLIP by overfilling the mask, lacking context (Fig. 4). Finally, extend + +Table 3. Ablation study of our proposed method. Results are reported on both semantic and visual alignment metrics. We denote vanilla PixArt- $\alpha$ inpainting model as Baseline. Also we denote QD as Query-Disentanglement without memory, only with the current object and background. + +
MethodSemantic Align.Visual Align.
BLEU-2/3/4↑METEOR↑CLIPc↑
Baseline56.29 / 42.04 / 33.060.158664.05
+BCG60.74 / 46.27 / 35.200.158564.10
+QD62.68 / 46.42 / 35.030.153063.99
Ours64.99 / 47.69 / 36.590.151364.29
+ +![](images/d4f6874d6f7b55440dd5e9475b6e4c5d8e8987aed43525e9c5b78e21a31bb644.jpg) +"A beach" + +![](images/0b7ba283c8e3ad56438116863c387533bd2fca8a56d1e16a6506dd4cecb9cf88.jpg) ++“a blue bus" + +![](images/c29dc64bdba4590f0d3f43ee11e838020e44669cb21046da243d2ad11df75f93.jpg) + +![](images/2595126f4b4c2ccc3c831e1a0421ee0cac385dd879c6cd70ce294dc893a92c4b.jpg) + +![](images/b3b4bf6f43b39d757c84d6ae8e8fa8407bbd438ddbf8b588df5345ca3191ea71.jpg) +HD Painter +Figure 5. Comparison in image editing capability with latest image editing models. [3, 32] Note that the initial image is generated by our framework, which is equivalent to PixArt- $\alpha$ [13] with no mask input. + +![](images/08d0a8d4c7035059c851cadea660a63db8e7e5253aafe4101026c6bf359f5bbb.jpg) ++“a hotel behind the bus" + +![](images/000f7e58b0bd78ec1db515d762b4a9df8c57e419e6cbecd712b944394b610208.jpg) +Blended Latent Diffuion + +![](images/818b02e5baaa063e5ffc885cd3869b03d45998f95277ecb183616e9141581a6e.jpg) + +![](images/d4d11c282c2525a0869a0bb24368763b9a4428b6fe12f185a7bf545a25f26ae8.jpg) +Ours + +ing QD to multi-query disentanglement (MQD) with layerwise memory integrates new objects into all prior prompts, achieving higher BLEU-2/3/4 and CLIP scores. + +# 5.4. Qualitative Result + +Comparison with Editing Approaches under Interactive Scenario. In Fig. 5, we demonstrate our method's effectiveness in iterative editing compared to state-of-the-art single-step editing approaches [2, 32]. Starting with a background of "A beach" generated by vanilla PixArt- $\alpha$ [13], we sequentially place a bus and a hotel. Our method successfully renders the hotel behind the bus, accurately interpreting both the prompt "A beach" and the mask order through MQD to achieve the intended spatial arrangement. + +In contrast, Blended Latent Diffusion [3] struggles with the initial placement of "bus" in the complex background, while HD-Painter [32] shows inconsistencies in the appearance of "bus" (highlighted in red in Fig. 5). Our method maintains background consistency while naturally integrating new objects into the scene. + +Improved Editability. Fig. 6 compares our method with + +![](images/dc416de1d738b12fafb772c03447940e45cdfe8e61243e519fdc371a654ba223.jpg) +Figure 6. Improved editability of image through Background Consistency Guidance and Multi-Query Disentangled cross attention. Through recycling the previous step's latents, we can remove the object that is behind the foreground object, enabling enhanced editability of the image. + +Table 4. Human preference study with recent literature on image editing on our benchmark. We utilize HD-Painter [32] and Blended Latent Diffusion with SD-XL [38] to compare under iterative editing scenario and compare the preference on a 5-point Likert scale. + +
MethodBackground ConsistencyNatural AdaptationText-scene Alignment
HD-Painter [32]3.712.813.08
BLD [3]3.432.242.04
Ours4.594.284.49
+ +Photoshop 2025 [1]. Placing "a glass of wine" or "a lizard" is straightforward, but removing "a glass of wine" while preserving "a lizard" remains challenging for Photoshop's generative fill, which requires precise brush strokes and often leaves artifacts (red box). Our approach simplifies this by leveraging existing masks and prior-step latents to naturally inpaint the background while applying MQD during partial denoising to preserve the lizard's structure. + +Comparison with Existing T2I models. Fig. 7 showcases our method's ability to handle complex multi-object scenarios that challenge conventional text-to-image models. For instance, in row A, only our framework successfully generates the intended composition of a lion behind a flower. Row B and C further demonstrate the challenge for existing T2I approaches to handle complex spatial arrangements, while illustrating our method's capability. + +# 5.5. Human Preference Study + +We conduct a comprehensive human evaluation study with 50 participants to assess the quality of our method compared to the latest editing models [3, 32] on our benchmark. The evaluation focuses on three aspects: background consistency, natural adaptation, and text-scene alignment, rated on a 5-point Likert scale. + +Our method outperforms single-step editing approaches [3, 32] across all three categories. These results demonstrate that traditional editing models, while effective + +![](images/5521930328725612bc14122e12a920d8f6738b17297f1d63cf55fa1ae728029d.jpg) +A + +![](images/4381c49380e848f98f561e81b26924398e55efd6a4f2377c95b22edd49bacbec.jpg) + +![](images/17869df27a5a503c91705d8df9c6b4f9f2b7d9c7b1859dc1b4706a2ab06dd265.jpg) + +![](images/adc244b570684764c6863bf04dcc927835ba84353f5dd0c9b2f108fc670c2126.jpg) +B + +![](images/c74a2b1302ed49ecf4c7bf97bc09e16b45048174b10328e1ee6bfd7ac15cd078.jpg) + +![](images/ff4f4dfe954ac7c2b52de83217a07507e95155d3d7eaf3d23207fe67141b6c3a.jpg) +C. A wooden table has a flower pot on the left and some stacked books in the center, with a small statue resting behind the table. + +C +PIXART-a +Figure 7. Comparison in interactive scenarios with existing T2I generative models. Stable Diffusion XL [38] and PixArt- $\alpha$ [13] use text input only. +![](images/d09690a037f82844e1c23a78c3963570416cc67d2c176a1f320b31b3d77249ad.jpg) +A. A lion stands behind flowers in the desert, positioned next to the water where a palm tree is growing nearby under the bright sun. + +Stable Diffusion XL +![](images/bfc138d8bdf7561517cefbb74fea5851225a8376593becc78d8490766a44dfe5.jpg) +B. Santa Claus and a teddy bear are sitting by a campfire in a snowy forest, with a gingerbread house directly behind them under falling snow. + +![](images/ea7f6ea9b9f5ea642f4b6747ed5af887b98f2a0c00d12924dc8debb83e9a7aac.jpg) +Ours + +for single-step modifications, are insufficient for sequential editing. This underscores the necessity of specialized approaches like ours for handling sequential iterative editing scenarios. Additional human evaluation results comparing our method with text-to-image models [13, 38] are provided in the supplementary materials. + +# 6. Conclusion + +We propose a novel framework for sequential image editing, which poses challenges in maintaining background consistency and seamless integration of new objects into the scenes. Our method includes three key components: Layer-wise Memory to preserve previous content, Background Consistency Guidance to keep the background stable with faster editing, and Multi-Query Disentangled Cross-Attention for natural object adaptation. Experiments on our proposed benchmark dataset show that our approach outperforms state-of-the-art methods in semantic and visual alignment, making it ideal for complex, iterative image editing scenarios. This framework paves the way for future work in improved interactive editing. + +Limitation and Future Work. Since our approach utilizes image editing, generating multiple objects takes longer, depending on the number of edits. Also, utilizing layer-wise memory requires additional memory costs. We plan to make it more efficient for faster editing in the future. + +Acknowledgement. This work was supported by IITP grant (RS-2021-II211343: AI Graduate School Program at Seoul National University $(5\%)$ and RS-2024-00509257: Global AI Frontier Lab $(30\%)$ ) and NRF grant (No.RS-2024-00405857 $(65\%)$ ) funded by the Korea government (MSIT). + +# References + +[1] Adobe Inc. Adobe Photoshop. https://www.adobe.com/products/photoshop.html, 2024. Version 2025. 1, 2, 8, 17, 20 +[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 2, 7, 16 +[3] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM TOG, 2023. 1, 2, 4, 6, 7, 8, 11, 13, 17, 20 +[4] Eslam Mohamed Bakr, Pengzhan Sun, Xiaogian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models. In ICCV, 2023. 2, 6, 14 +[5] Coloma Ballester, Marcelo Bertalmio, Vicent Caselles, Guillermo Sapiro, and Joan Verdera. Filling-in by joint interpolation of vector fields and gray levels. IEEE TIP, 2001. 2 +[6] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In IEEvaluation@ACL, 2005. 6 +[7] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 2 +[8] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM TOG, 2009. 2 +[9] Marcelo Bertalmio, Guillermo Sapiro, Vicent Caseles, and Coloma Ballester. Image inpainting. In SIGGRAPH, 2000. 2 +[10] Shariq Farooq Bhat, Niloy J. Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning. In SIGGRAPH, 2024. 6, 7, 16 +[11] BlackForestLabs. Flux, 2023. Accessed: 2024-11-12. 1 +[12] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv:2403.04692, 2024. 1 +[13] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 1, 3, 6, 7, 8, 13 +[14] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In WACV, 2024. 2, 6 + +[15] Ruidong Chen, Lanjun Wang, Weizhi Nie, Yongdong Zhang, and An-An Liu. Anyscene: Customized image synthesis with composited foreground. In CVPR, 2024. 2 +[16] Yifu Chen, Jingwen Chen, Yingwei Pan, Yehao Li, Ting Yao, Zhineng Chen, and Tao Mei. Improving text-guided object inpainting with semantic pre-inpainting. In ECCV, 2024. 2 +[17] Jaemin Cho, Linjie Li, Zhengyuan Yang, Zhe Gan, Lijuan Wang, and Mohit Bansal. Diagnostic benchmark and iterative inpainting for layout-guided image generation. In CVPR Workshop, 2024. 2, 6 +[18] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE TIP, 2004. 2 +[19] Abdelrahman Eldisokey and Peter Wonka. Build-a-scene: Interactive 3d layout control for diffusion-based image generation. In ICLR, 2025. 16 +[20] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 6, 7, 11 +[21] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. In NeurIPS, 2023. 2, 6, 14 +[22] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv:2405.04007, 2024. 6 +[23] Biao Gong, Siteng Huang, Yutong Feng, Shiwei Zhang, Yuyuan Li, and Yu Liu. Check locate rectify: A training-free layout calibration system for text-to-image generation. In CVPR, 2024. 2 +[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2 +[25] Runhui Huang, Kaixin Cai, Jianhua Han, Xiaodan Liang, Renjing Pei, Guansong Lu, Songcen Xu, Wei Zhang, and Hang Xu. Layerdiff: Exploring text-guided multi-layered composable image synthesis via layer-collaborative diffusion model. In ECCV, 2024. 2 +[26] Jonghyun Lee, Hansam Cho, Youngjoon Yoo, Seoung Bum Kim, and Yonghyun Jeong. Compose and conquer: Diffusion-based 3d depth aware composable image synthesis. In ICLR, 2024. 2, 17 +[27] Jaerin Lee, Daniel Sungho Jung, Kanggeon Lee, and Young Mu Lee. StreamMultiDiffusion: real-time interactive generation with region-based semantic control. arXiv:2403.09055, 2024. 2 +[28] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 2, 13 +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 6 + +[30] Zhengyao Lv, Yuxiang Wei, Wangmeng Zuo, and KwanYee K Wong. Place: Adaptive layout-semantic fusion for semantic image synthesis. In CVPR, 2024. 2 +[31] Yiwei Ma, Jiayi Ji, Ke Ye, Weihuang Lin, Zhibin Wang, Yonghan Zheng, Qiang Zhou, Xiaoshuai Sun, and Rongrong Ji. I2ebench: A comprehensive benchmark for instruction-based image editing. In NeurIPS, 2024. 2, 6, 14 +[32] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: High-resolution and prompt-faithful text-guided image inpainting with diffusion models. In ICLR, 2025. 1, 2, 6, 7, 8, 11, 16, 17, 20 +[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 17 +[34] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2 +[35] Yotam Nitzan, Zongze Wu, Richard Zhang, Eli Shechtman, Daniel Cohen-Or, Taesung Park, and Michael Gharbi. Lazy diffusion transformer for interactive image editing. In ECCV, 2024. 2 +[36] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 6 +[37] Pincel. Pincel: Ai image editor, 2024. Accessed: 2024-11-15. 1, 2, 17, 20 +[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 6, 8, 11, 13 +[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. In ICLR, 2025. 2 +[40] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 14 +[41] Jiawei Ren, Mengmeng Xu, Jui-Chieh Wu, Ziwei Liu, Tao Xiang, and Antoine Toisoul. Move anything with layered scene diffusion. In CVPR, 2024. 2 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2 +[43] Takahiro Shirakawa and Seiichi Uchida. Noisecollapse: A layout-aware text-to-image diffusion model based on noise cropping and merging. In CVPR, 2024. 2, 6, 7 +[44] Jack Urbanek, Florian Bordes, Pietro Astolfi, Mary Williamson, Vasu Sharma, and Adriana Romero-Soriano. A picture is worth more than 77 text tokens: Evaluating clip-style models on dense captions. In CVPR, 2024. 15 + +[45] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, 2023. 2, 6, 13, 14 +[46] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In CVPR, 2024. 2 +[47] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In ICCV, 2023. 2 +[48] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 2 +[49] Chao Yang, Xin Lu, Zhe Lin, Eli Shechtman, Oliver Wang, and Hao Li. High-resolution image inpainting using multiscale neural patch synthesis. In CVPR, 2017. 2 +[50] Zili Yi, Qiang Tang, Shekoofeh Azizi, Daesik Jang, and Zhan Xu. Contextual residual aggregation for ultra high-resolution image inpainting. In CVPR, 2020. +[51] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas Huang. Free-form image inpainting with gated convolution. In ICCV, 2019. 2 +[52] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. In ECCV, 2024. 15 +[53] Lisai Zhang, Qingcai Chen, Baotian Hu, and Shuang Jiang. Text-guided neural image inpainting. In ACM MM, 2020. 2 +[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 11, 17 +[55] Xinyang Zhang, Wentian Zhao, Xin Lu, and Jeff Chien. Text2layer: Layered image generation using latent diffusion model. arXiv:2307.09781, 2023. 2 +[56] Bo Zhao, Lili Meng, Weidong Yin, and Leonid Sigal. Image generation from layout. In CVPR, 2019. 2 +[57] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. In NeurIPS, 2023. 17 +[58] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In CVPR, 2023. 2 \ No newline at end of file diff --git a/2025/Improving Editability in Image Generation with Layer-wise Memory/images.zip b/2025/Improving Editability in Image Generation with Layer-wise Memory/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..71bd59e4ab92522c4dc95998267331b037e6f29c --- /dev/null +++ b/2025/Improving Editability in Image Generation with Layer-wise Memory/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76255563ec27ecda51ca8aec4a1c0faee7ecd0e9737e0c4ba97d31767441af9a +size 729649 diff --git a/2025/Improving Editability in Image Generation with Layer-wise Memory/layout.json b/2025/Improving Editability in Image Generation with Layer-wise Memory/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..93179d131ebcb6fb1a7dca823f9fb91a6123fc6c --- /dev/null +++ b/2025/Improving Editability in Image Generation with Layer-wise Memory/layout.json @@ -0,0 +1,10264 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 518, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 518, + 122 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 518, + 122 + ], + "type": "text", + "content": "Improving Editability in Image Generation with Layer-wise Memory" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 192, + 143, + 416, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 143, + 416, + 186 + ], + "spans": [ + { + "bbox": [ + 192, + 143, + 416, + 186 + ], + "type": "text", + "content": "Daneul Kim Jaeah Lee Jaesik Park* Seoul National University, Republic of Korea {carpedkm, hayanz, jaesik.park}@snu.ac.kr" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 65, + 208, + 233, + 419 + ], + "blocks": [ + { + "bbox": [ + 65, + 208, + 233, + 419 + ], + "lines": [ + { + "bbox": [ + 65, + 208, + 233, + 419 + ], + "spans": [ + { + "bbox": [ + 65, + 208, + 233, + 419 + ], + "type": "image", + "image_path": "bfdc191bb47f13184c6baed485a30d1b8264ea7d2796c3a0a27587b65ff39f0d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 420, + 123, + 427 + ], + "lines": [ + { + "bbox": [ + 89, + 420, + 123, + 427 + ], + "spans": [ + { + "bbox": [ + 89, + 420, + 123, + 427 + ], + "type": "text", + "content": "HD Printer" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 238, + 208, + 544, + 439 + ], + "blocks": [ + { + "bbox": [ + 238, + 208, + 544, + 439 + ], + "lines": [ + { + "bbox": [ + 238, + 208, + 544, + 439 + ], + "spans": [ + { + "bbox": [ + 238, + 208, + 544, + 439 + ], + "type": "image", + "image_path": "38a98b819d90e6d9ed2a72adb09c94192d0582d37fd87db48876af5760451c95.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 450, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 54, + 450, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 54, + 450, + 555, + 485 + ], + "type": "text", + "content": "Figure 1. Overview. Our framework enables the interactive generation of images with enhanced control but in a simple manner, by rough mask and prompt, through iterative scene editing. We utilize the background scene generated by our framework to edit in HD Painter [32] or Blended Latent Diffusion (BLD) [3] for comparison and commercial products like Photoshop [1] and Pincel [37]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 500, + 201, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 500, + 201, + 513 + ], + "spans": [ + { + "bbox": [ + 151, + 500, + 201, + 513 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 526, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 297, + 694 + ], + "type": "text", + "content": "Most real-world image editing tasks require multiple sequential edits to achieve desired results. Current editing approaches, primarily designed for single-object modifications, struggle with sequential editing: especially with maintaining previous edits along with adapting new objects naturally into the existing content. These limitations significantly hinder complex editing scenarios where multiple objects need to be modified while preserving their contextual relationships. We address this fundamental challenge through two key proposals: enabling rough mask inputs that preserve existing content while naturally integrating new elements and supporting consistent editing across multiple modifications. Our framework achieves this through layer-wise memory, which stores latent representations and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 501, + 556, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 556, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 556, + 634 + ], + "type": "text", + "content": "prompt embeddings from previous edits. We propose Background Consistency Guidance that leverages memorized latents to maintain scene coherence and Multi-Query Disentanglement in cross-attention that ensures natural adaptation to existing content. To evaluate our method, we present a new benchmark dataset incorporating semantic alignment metrics and interactive editing scenarios. Through comprehensive experiments, we demonstrate superior performance in iterative image editing tasks with minimal user effort, requiring only rough masks while maintaining high-quality results throughout multiple editing steps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "spans": [ + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 678, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 556, + 715 + ], + "type": "text", + "content": "Recent advances in text-to-image synthesis through powerful diffusion models like Stable Diffusion [20, 42], PixArt [12, 13] and FLUX [11] have transformed visual" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 704, + 145, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 145, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 145, + 714 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7889" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "content creation. However, users often demand multiple sequential edits to achieve desired results, iteratively refining and adding elements to their images. Current approaches in inpainting [3, 32] are primarily designed for single-object modifications with limited changes like colors or styles, making them inadequate for complex editing scenarios that involve multiple objects and their interactions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 294, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 294, + 289 + ], + "type": "text", + "content": "Image editing approaches, while offering localized modifications, face several limitations. Current methods [3, 16, 32, 48] struggle when modifications need to be applied sequentially, often demanding precise segmentation masks [16] or external modules [39] to maintain background integrity [32]. For example, generating \"A dog sitting\" as shown in Fig. 1 remains challenging, as it requires maintaining previous edits while ensuring the \"dog\" naturally blends with the \"Jeep\" and \"Lego man standing\" - a complex iterative editing scenario that current methods [1, 3, 32, 37] struggle to handle." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 292, + 294, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 294, + 422 + ], + "type": "text", + "content": "Layout-to-image generation offers an alternative approach through various inputs including bounding boxes [28, 47, 58], depth maps [26, 54], and semantic masks [7, 27]. However, these methods require complete regeneration of the entire image for each modification, making iterative editing particularly cumbersome – unable to maintain the surrounding background. While recent work has explored layered representations [25, 41, 55] and instance-based generation [46], they require complex optimization processes or additional training, limiting their practical applicability in iterative editing scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 425, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 581 + ], + "type": "text", + "content": "We address these limitations through two key innovations. First, we enable object placement aligning with the user's intention using only rough mask inputs while preserving background context. Second, we support consistent iterative editing across multiple modifications. To achieve this, we introduce the concept of mask order, which specifies the sequence of object generation during iterative image editing. In Fig. 1, we add \"a lego man\", \"Jeep\" sequentially with different mask orders, but they naturally adapt into the image, side by side. If we add \"A dog sitting\" with overlapping mask, then it means that we aim to put the \"dog\" in front of the \"Jeep\" and \"a lego man\", making the instance order of \"dog\" in front of \"Jeep\" and \"a lego man\"." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "To handle this, we incorporate three technical components: (1) Layer-wise memory for storing editing history, (2) Background Consistency Guidance (BCG) for maintaining unedited regions, and (3) Multi-Query Disentangle-ment (MQD) in cross-attention for natural object integration. The layer-wise memory stores and manages the latent representations and prompt embeddings from previous editing steps, eliminating redundant computations typical in sequential modifications while maintaining consistency across multiple edits. The BCG not only ensures unedited regions remain stable but also reduces computational over" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "head by avoiding repeated forward passes on the original image, enabling reduced editing time, while MQD enables the natural integration of new objects with existing content." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 108, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 204 + ], + "type": "text", + "content": "Additionally, we propose Multi-Edit Bench, a comprehensive benchmark for evaluating iterative image editing capabilities. Prior benchmarks either focus on single-turn edits [31, 45] or layout-to-image generation [4, 17, 21], failing to capture the challenges of sequential modifications. Our benchmark introduces layer-wise semantic evaluation metrics to assess both edit quality and cross-modification consistency in multi-step editing scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 204, + 511, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 204, + 511, + 215 + ], + "spans": [ + { + "bbox": [ + 326, + 204, + 511, + 215 + ], + "type": "text", + "content": "To summarize, our key contributions include:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 217, + 553, + 323 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "type": "text", + "content": "- A framework for interactive mask order-based object placement editing with only rough mask inputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 240, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 553, + 263 + ], + "type": "text", + "content": "- A localized editing mechanism using layer-wise memory that maintains consistency across multiple edits." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 264, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 287 + ], + "type": "text", + "content": "- Multi-query disentangled cross attention, allowing natural integration while preserving background context." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 288, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 323 + ], + "type": "text", + "content": "- Multi-Edit Bench, a comprehensive benchmark for evaluating the semantic alignment and sequential iterative editing capabilities." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 335, + 400, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 335, + 400, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 335, + 400, + 346 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 354, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 510 + ], + "type": "text", + "content": "Image inpainting. Image inpainting has evolved from classical patch-based [8, 18] and interpolation-driven [5, 9] methods to powerful deep-learning approaches [49-51] that better handle complex structures. Text-conditioned inpainting [51, 53] further broadened user control, and diffusion-based models [24]—including GLIDE [34], Stable Diffusion [42], and Blended Diffusion [2, 3]—significantly improved realism. Yet, existing works often focus on single-object edits. Recent attempts to relax mask precision [32, 48] or enhance quality [16] remain limited in multi-step scenarios, struggling to maintain consistency across multiple objects or occlusions—hence insufficient for iterative real-world editing tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 653 + ], + "type": "text", + "content": "Layout-to-Image Generation. Controllable image synthesis has recently emphasized spatial arrangement via bounding boxes, segmentation masks, or other layout cues [14, 46, 56]. Advances in diffusion-based layout calibration [23] and grounding [43] have improved fidelity, and interactive methods [15, 30, 35] aim to let users refine elements step by step. However, most such pipelines regenerate the entire image for each new edit, losing previously established context and coherence. By contrast, our approach incrementally updates only relevant regions through a memory-assisted process, naturally preserving previously edited objects." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Evaluation Benchmark. Most existing editing benchmarks focus on single-turn modifications [31, 45], while layout-centric datasets such as HRS-1k [4] and NSR-1k [21] assess scene quality or prompt alignment in a single-shot manner. Consequently, they overlook the iterative na" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7890" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 547, + 302 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 547, + 302 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 547, + 302 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 547, + 302 + ], + "type": "image", + "image_path": "6e69cc2a32f69fc97a956edd7e5b71fd05cdaf6d233177b964c9a6eaafbd42df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 309, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 309, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 309, + 555, + 346 + ], + "type": "text", + "content": "Figure 2. Overview. (a) The left denotes an illustration of how Multi Query Disentanglement is performed in the cross-attention layer. (b) The upper right figure shows Background Consistency Guidance with recalled latents, conducting latent blending with the saved latents. (c) The right below shows the layer-wise memory, saving the previous editing steps' latents, masks, and prompt embeddings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 357, + 298, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 298, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 298, + 479 + ], + "type": "text", + "content": "ture of real-world edits where object placements evolve over multiple steps. To address this gap, we introduce Multi-Edit Bench, a new benchmark that evaluates multi-step editing in terms of semantic accuracy and visual alignment across the entire editing sequence. This multi-step perspective better captures practical editing workflows, where users iteratively refine scenes by adding, removing, or repositioning objects, thereby measuring a method's capacity to preserve context and maintain coherent compositions over multiple, potentially complex modifications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 492, + 113, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 113, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 113, + 504 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 514, + 123, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 123, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 123, + 525 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 533, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 295, + 664 + ], + "type": "text", + "content": "We propose a framework for iterative image editing with three key components: Layer-wise memory, background consistency guidance (BCG), and multi-query disentangled cross-attention (MQD). Layer-wise memory stores editing history to enable consistent object placement across sequential edits. BCG leverages this stored information for efficient latent blending while preserving background integrity. MQD ensures the natural object integration through disentangling the attention between queries and latents, handling complex scenarios as shown in Fig. 2 like placing \"a dish\" in front of \"a cupcake\" and \"a mug cup\"." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": "Our framework's architecture, built upon transformer-based diffusion models [13], operates through the following systematic workflow: First, we detail the layer-wise memory mechanism (Fig. 2(c)) in Sec. 3.2, which stores editing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 357, + 555, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 444 + ], + "type": "text", + "content": "history. Then, in Sec. 3.3, we explain how BCG leverages this stored information for efficient latent blending (Fig. 2 (b)). Sec. 3.4 describes MQD's role in ensuring natural object integration through query disentanglement (Fig. 2 (a)). Finally, in Sec. 3.5, we demonstrate how these components enable advanced editing capabilities, including the removal of occluded elements while preserving foreground integrity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 453, + 432, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 432, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 432, + 466 + ], + "type": "text", + "content": "3.2. Layer-wise Memory" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 472, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 555 + ], + "type": "text", + "content": "Layer-wise memory enables stable background preservation while ensuring the natural integration of new objects during iterative editing with mask order. By storing and managing key information from each editing step according to the specified mask order, the model can reference previous edits when generating new content while maintaining the intended spatial relationships between objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "text", + "content": "Let the set of prompts for generating the background and objects be denoted as " + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "inline_equation", + "content": "P_{l} = \\{p_{0}, p_{1}, p_{2}, \\ldots\\}" + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "inline_equation", + "content": "p_{0}" + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "text", + "content": " is the background generation prompt, and " + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "inline_equation", + "content": "p_{1}, p_{2}, \\ldots" + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "text", + "content": " represent the prompts for each object. Similarly, let the set of corresponding masks be denoted as " + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "inline_equation", + "content": "M_{l} = \\{m_{0}, m_{1}, m_{2}, \\ldots\\}" + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "text", + "content": " defines the region of interest (RoI) for the object associated with prompt " + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "inline_equation", + "content": "m_{0}" + }, + { + "bbox": [ + 313, + 556, + 556, + 652 + ], + "type": "text", + "content": " represents the case with mask of all elements being 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "The layer-wise memory " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "L_{l} = \\{l_{0}, l_{1}, l_{2}, \\ldots \\}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " stores three key elements for each step " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "l_{i} = \\{\\mathbf{p}_{i}, \\{\\mathbf{Z}_{i}^{t}\\}_{t=1}^{T}, m_{i}\\}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": ", where: " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " denotes prompt embedding that guides the generation, " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}_{i}^{t}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " as denoising latents across timesteps and " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " as a mask defining the object's RoI." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "7891" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 145, + 264, + 188 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 296, + 138 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 296, + 138 + ], + "type": "text", + "content": "Table 1. Computational comparison in utilizing Background Consistency Guidance (BCG). We utilize PixArt-" + }, + { + "bbox": [ + 55, + 70, + 296, + 138 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 70, + 296, + 138 + ], + "type": "text", + "content": " inpaint pipeline with vanilla latent blending and our custom-implemented pipeline with BCG for comparison. We ran the inpainting task for the single-step editing 5 times for each method. Note that we use PixArt-alpha-XL-1024 pretrained weight for the experiment." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 145, + 264, + 188 + ], + "lines": [ + { + "bbox": [ + 87, + 145, + 264, + 188 + ], + "spans": [ + { + "bbox": [ + 87, + 145, + 264, + 188 + ], + "type": "table", + "html": "
MethodTime (sec)VRAM
Latent Blending4.1218 ±0.052116.89GB
BCG3.8992±0.014216.90GB
", + "image_path": "2a50c6872a4a696b8f95cafc2b852620a47bff0b12f2e3297240da064281ed38.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 197, + 296, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 296, + 280 + ], + "type": "text", + "content": "For a generation, the initial background latent is created as " + }, + { + "bbox": [ + 55, + 197, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_0 = f_\\theta(p_0)" + }, + { + "bbox": [ + 55, + 197, + 296, + 280 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 197, + 296, + 280 + ], + "type": "inline_equation", + "content": "f_\\theta" + }, + { + "bbox": [ + 55, + 197, + 296, + 280 + ], + "type": "text", + "content": " is the pre-trained diffusion model. For each subsequent object " + }, + { + "bbox": [ + 55, + 197, + 296, + 280 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 197, + 296, + 280 + ], + "type": "text", + "content": ", we initialize its latent independently: " + }, + { + "bbox": [ + 55, + 197, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_i = f_\\theta(p_i, m_i)" + }, + { + "bbox": [ + 55, + 197, + 296, + 280 + ], + "type": "text", + "content": ". This independence in initialization, combined with the stored history in layer-wise memory, enables flexible generation while maintaining background consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 281, + 296, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 352 + ], + "type": "text", + "content": "The layer-wise memory serves two key purposes: (1) Context preservation, to store the complete editing history, enabling the model to maintain consistency with previously generated content (2) Localized editing, enabling control over specific regions while preserving the surrounding content through mask-guided generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 352, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 413 + ], + "type": "text", + "content": "By maintaining this structured history of edits, the model can effectively handle complex scenarios where multiple objects need to be integrated coherently. The stored information guides background consistency and natural object integration, as detailed in the following sections." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 419, + 243, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 243, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 243, + 432 + ], + "type": "text", + "content": "3.3. Background Consistency Guidance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 509 + ], + "type": "text", + "content": "Background Consistency Guidance (BCG) addresses a key challenge in iterative image editing: maintaining the integrity of previous edits while incorporating new elements efficiently. Unlike traditional inpainting approaches, BCG leverages layer-wise memory to ensure stability across multiple sequential modifications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 514, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 571 + ], + "type": "text", + "content": "Latent Retrieval from Layer-wise Memory: For each editing step " + }, + { + "bbox": [ + 55, + 514, + 296, + 571 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 514, + 296, + 571 + ], + "type": "text", + "content": ", BCG retrieves information from the previous edits stored in layer-wise memory " + }, + { + "bbox": [ + 55, + 514, + 296, + 571 + ], + "type": "inline_equation", + "content": "L_{l}" + }, + { + "bbox": [ + 55, + 514, + 296, + 571 + ], + "type": "text", + "content": ", enabling seamless integration of new content while preserving prior modifications:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 103, + 572, + 295, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 572, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 103, + 572, + 295, + 587 + ], + "type": "interline_equation", + "content": "l _ {i - 1} = \\left\\{\\mathbf {p} _ {i - 1}, \\left\\{\\mathbf {Z} _ {i - 1, t} \\right\\} _ {t = 1} ^ {T}, m _ {i - 1} \\right\\} \\tag {1}", + "image_path": "4efd5b423d7afa12a19288ea96013f575611340b0d14016afa55d1873fc422b3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 591, + 296, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 296, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 296, + 663 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 591, + 296, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i - 1}" + }, + { + "bbox": [ + 55, + 591, + 296, + 663 + ], + "type": "text", + "content": " is the prompt embedding from the prior step, " + }, + { + "bbox": [ + 55, + 591, + 296, + 663 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}_{i - 1,t}\\}_{t = 0}^{T}" + }, + { + "bbox": [ + 55, + 591, + 296, + 663 + ], + "type": "text", + "content": " represents the stored latents from each denoising step, and " + }, + { + "bbox": [ + 55, + 591, + 296, + 663 + ], + "type": "inline_equation", + "content": "m_{i - 1}" + }, + { + "bbox": [ + 55, + 591, + 296, + 663 + ], + "type": "text", + "content": " is the previous mask. This stored latent provides context for the current edit, preserving the background information without requiring a re-computation on the original image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 668, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 296, + 693 + ], + "type": "text", + "content": "Ensuring Consistency: When adding a new object, we update only the masked region while preserving the rest:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 103, + 701, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 103, + 701, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {i} = \\mathbf {Z} _ {i - 1} \\odot (1 - m _ {i}) + \\mathbf {Z} _ {i} \\odot m _ {i} \\tag {2}", + "image_path": "1533e2efdce223cb7deb8c9a4813482f5c413fe07cd23ee198e3948fe2313bac.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "This selective update mechanism is crucial for iterative editing since it ensures that each new modification preserves previous edits. The element-wise multiplication " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " enables control over which regions are updated, maintaining the integrity of the evolving image composition." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 137, + 518, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 518, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 518, + 149 + ], + "type": "text", + "content": "Computational Efficiency for Iterative Editing:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 149, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 555, + 293 + ], + "type": "text", + "content": "BCG improves the efficiency of iterative image editing by reducing redundant computations in sequential modifications. Traditional inpainting methods [3] that utilize latent blending (LB) require a forward pass on the original image for each edit, becoming costly with multiple edits. While both approaches (i.e., LB, BCG) require denoising costs, BCG avoids repeated forward passes. Let " + }, + { + "bbox": [ + 313, + 149, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\Omega = \\mathrm{FLOPs}(T,L,H,W)" + }, + { + "bbox": [ + 313, + 149, + 555, + 293 + ], + "type": "text", + "content": " represent the base computational cost for processing an image of size " + }, + { + "bbox": [ + 313, + 149, + 555, + 293 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 313, + 149, + 555, + 293 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 313, + 149, + 555, + 293 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 149, + 555, + 293 + ], + "type": "text", + "content": " denoising steps and " + }, + { + "bbox": [ + 313, + 149, + 555, + 293 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 149, + 555, + 293 + ], + "type": "text", + "content": " model layers, and " + }, + { + "bbox": [ + 313, + 149, + 555, + 293 + ], + "type": "inline_equation", + "content": "C_f" + }, + { + "bbox": [ + 313, + 149, + 555, + 293 + ], + "type": "text", + "content": " denote the forward pass cost. The total cost for LB and BCG can be expressed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 394, + 293, + 553, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 293, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 394, + 293, + 553, + 306 + ], + "type": "interline_equation", + "content": "\\operatorname {C o s t} _ {\\mathrm {L B}} = C _ {f} + \\Omega . \\tag {3}", + "image_path": "97b3979c719d9d0c6fc4ec55c543680f4448f0ae828d29b7350b724da20ddb27.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 311, + 503, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 503, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 503, + 323 + ], + "type": "text", + "content": "While for BCG, by skipping the forward pass:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 403, + 333, + 553, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 333, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 403, + 333, + 553, + 345 + ], + "type": "interline_equation", + "content": "\\operatorname {C o s t} _ {\\mathrm {B C G}} = \\Omega , \\tag {4}", + "image_path": "d720a786c3fea7ba539a3ac8bd5492eab3cdfd211744876b6ddc5ac0bebf1abb.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 354, + 552, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 552, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 552, + 367 + ], + "type": "text", + "content": "Assuming " + }, + { + "bbox": [ + 313, + 354, + 552, + 367 + ], + "type": "inline_equation", + "content": "C_f \\approx r\\Omega" + }, + { + "bbox": [ + 313, + 354, + 552, + 367 + ], + "type": "text", + "content": " with ratio " + }, + { + "bbox": [ + 313, + 354, + 552, + 367 + ], + "type": "inline_equation", + "content": "r \\in (0,1)" + }, + { + "bbox": [ + 313, + 354, + 552, + 367 + ], + "type": "text", + "content": ", we can estimate:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 359, + 375, + 553, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 375, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 359, + 375, + 553, + 401 + ], + "type": "interline_equation", + "content": "\\text {E f f e c i e n c y} = \\frac {\\operatorname {C o s t} _ {\\mathrm {L B}}}{\\operatorname {C o s t} _ {\\mathrm {B C G}}} \\approx 1 + r \\tag {5}", + "image_path": "1a3fd19e09c7a0197afd24416038455ee018d5ec26ed6b7f4c4a70c44cb3cf7b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "content": "In Tab. 1, we observe about " + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "content": " reduction in computational time for single-step editing with BCG compared to LB. This improvement becomes more significant in scenarios requiring multiple sequential modifications, as LB requires multiple forward passes while BCG requires none." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 475, + 538, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 538, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 538, + 487 + ], + "type": "text", + "content": "3.4. Multi-Query Disentangled Cross-Attention" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 492, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 555, + 612 + ], + "type": "text", + "content": "Multi-query disentangled cross-attention (MQD) ensures the natural integration of objects across different mask orders while preserving their spatial relationships. For example, when generating \"a dish\" as in Fig. 2, MQD enables the 'dish' to blend naturally into the background, despite being in different mask orders, while maintaining the structure and color of the background with proper occlusion of other objects like 'a mug cup' and 'a cupcake' behind. This is achieved by disentangling attention across multiple queries and leveraging information from layer-wise memory " + }, + { + "bbox": [ + 313, + 492, + 555, + 612 + ], + "type": "inline_equation", + "content": "L_{l}" + }, + { + "bbox": [ + 313, + 492, + 555, + 612 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 617, + 495, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 495, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 495, + 629 + ], + "type": "text", + "content": "3.4.1. Cross-Attention for Current Object" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 632, + 554, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 554, + 657 + ], + "type": "text", + "content": "For each generation step " + }, + { + "bbox": [ + 313, + 632, + 554, + 657 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 632, + 554, + 657 + ], + "type": "text", + "content": ", MQD focuses attention on the current mask order's region of interest (RoI):" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 351, + 664, + 553, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 664, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 351, + 664, + 553, + 680 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i} ^ {k, a t t n} = \\operatorname {C r o s s A t t e n t i o n} \\left(\\mathbf {z} _ {i} ^ {k, t} \\odot m _ {i}, p _ {i}\\right) \\tag {6}", + "image_path": "f80b31ea26dedab4a6ab0e755e5b5fb6b03985dbeed47d09332c2e88518980f9.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 688, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 688, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 688, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 688, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^{k,t}" + }, + { + "bbox": [ + 313, + 688, + 554, + 714 + ], + "type": "text", + "content": " is the latent for the " + }, + { + "bbox": [ + 313, + 688, + 554, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 688, + 554, + 714 + ], + "type": "text", + "content": "-th block at denoising step " + }, + { + "bbox": [ + 313, + 688, + 554, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 688, + 554, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 688, + 554, + 714 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 313, + 688, + 554, + 714 + ], + "type": "text", + "content": " defines the RoI, and " + }, + { + "bbox": [ + 313, + 688, + 554, + 714 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 688, + 554, + 714 + ], + "type": "text", + "content": " is the prompt embedding. This" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7892" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 347, + 257 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 347, + 257 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 347, + 257 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 347, + 257 + ], + "type": "image", + "image_path": "71881933e23c4b28840b6b3a6557af2998aea9131e9970696fe1d5f34a961af2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 259, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 259, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 259, + 555, + 293 + ], + "type": "text", + "content": "Figure 3. Overview of our proposed Multi-Edit Benchmark for evaluation of iterative editing scenario. (a) explains the dataset generation pipeline through GPT-4 API, and (b) explains the evaluation methodology in visual alignment using CLIP and semantic alignment using LLaVa for single-image and in a layer-wise manner." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 350, + 70, + 529, + 255 + ], + "blocks": [ + { + "bbox": [ + 350, + 70, + 529, + 255 + ], + "lines": [ + { + "bbox": [ + 350, + 70, + 529, + 255 + ], + "spans": [ + { + "bbox": [ + 350, + 70, + 529, + 255 + ], + "type": "image", + "image_path": "8012102f00fda90e3434956ea080229a5b167e8819351e06040df0be5926b094.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 303, + 296, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 352 + ], + "type": "text", + "content": "ensures focused generation within the current mask order's region. In the Fig. 2, " + }, + { + "bbox": [ + 55, + 303, + 296, + 352 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 55, + 303, + 296, + 352 + ], + "type": "text", + "content": " is given as a mask corresponding to the " + }, + { + "bbox": [ + 55, + 303, + 296, + 352 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 55, + 303, + 296, + 352 + ], + "type": "text", + "content": " 'a dish' with " + }, + { + "bbox": [ + 55, + 303, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i}^{k,t}" + }, + { + "bbox": [ + 55, + 303, + 296, + 352 + ], + "type": "text", + "content": " corresponding to the latent of the masked region for 'a dish'." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 356, + 296, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 369 + ], + "type": "text", + "content": "3.4.2. Disentangled Attention for Previous Mask Orders" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 371, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 407 + ], + "type": "text", + "content": "To adapt the mask orders into the appropriate instance orders, MQD applies attention to non-overlapping regions from previous steps:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 295, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 458 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i} ^ {k, a t t n} = \\bigcup_ {j = 0} ^ {i - 1} \\text {C r o s s A t t e n t i o n} \\left(\\mathbf {z} _ {i} ^ {k, t} \\odot \\left(m _ {j} - \\Sigma_ {l = j + 1} ^ {i} m _ {l}\\right), p _ {j}\\right) \\tag {7}", + "image_path": "f1a059b81cf2c8b5fc37aeffe595c12ff59f727a872eca6c551c218cec004cca.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 459, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 592 + ], + "type": "text", + "content": "This operation ensures that objects from earlier mask orders, (i.e., previous masks for 'a mug cup' and 'a cupcake' on top of 'a wooden floor' behind 'a dish' altogether) remain coherent while allowing natural adaptation where needed. The term " + }, + { + "bbox": [ + 55, + 459, + 296, + 592 + ], + "type": "inline_equation", + "content": "m_{j} - \\sum_{l = j + 1}^{i}m_{l}" + }, + { + "bbox": [ + 55, + 459, + 296, + 592 + ], + "type": "text", + "content": " controls which regions from previous mask orders should influence the current generation, minimally affecting the \"mug cup\" and \"cupcake\" even if we put the \"dish\" in front of them. Please note that for background prompt " + }, + { + "bbox": [ + 55, + 459, + 296, + 592 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 459, + 296, + 592 + ], + "type": "text", + "content": ": \"wooden floor\", we inverse the whole union of masks and do disentangled cross attention with background prompt " + }, + { + "bbox": [ + 55, + 459, + 296, + 592 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 459, + 296, + 592 + ], + "type": "text", + "content": " as in Fig. 2 (a)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 596, + 171, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 171, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 171, + 608 + ], + "type": "text", + "content": "3.4.3. Final Latent Update" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 611, + 295, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 633 + ], + "type": "text", + "content": "The final update combines information across all mask orders:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 632, + 295, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 632, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 116, + 632, + 295, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i} ^ {\\text {m e r g e}} = \\mathbf {z} _ {i} ^ {\\text {a t t n}} \\bigcup \\sum_ {j = 0} ^ {i - 1} \\mathbf {z} _ {j} ^ {\\text {a t t n}} \\tag {8}", + "image_path": "fe415183f975be61a717d664b45c7e866576ca51933bada7b53ea3cf96b6c417.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "type": "text", + "content": "After a feedforward layer update and " + }, + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "type": "text", + "content": " transformer blocks, we apply BCG:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 99, + 699, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 699, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 99, + 699, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {i} ^ {t} = \\mathbf {z} _ {i} ^ {K, t} \\odot m _ {i} + \\mathbf {Z} _ {i - 1} ^ {t} \\odot (1 - m _ {i}) \\tag {9}", + "image_path": "e15d1a7686d0d187bcacc44a634052b15ab70fc733f974156fc9d2c8b1a0ac8d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "text", + "content": "This approach enables objects to adapt naturally to mask orders—for example, placing “a dog” in front of “Jeep” and “Lego man” (Fig. 1), or maintaining ‘a cupcake’ and ‘a mug cup’ despite inserting “a dish” in front (Fig. 2)—while preserving scene coherence. MQD balances content preservation with the seamless integration of new elements." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 384, + 440, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 440, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 440, + 397 + ], + "type": "text", + "content": "3.5. Improving Editability" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 402, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 474 + ], + "type": "text", + "content": "While BCG and MQD with layer-wise memory enhance multiple editing capabilities in the PixArt-" + }, + { + "bbox": [ + 313, + 402, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 402, + 555, + 474 + ], + "type": "text", + "content": " model, we further extend our framework to support object deletion, particularly for overlapped objects behind foreground elements. We achieve this through a novel application of our MQD and BCG components." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "text", + "content": "To delete an object added at editing step " + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "inline_equation", + "content": "i - 1" + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "text", + "content": ", we leverage the stored latent representations from both steps " + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "inline_equation", + "content": "i - 2" + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "text", + "content": " and the current step " + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}_{i - 2}^t\\}_{t = 1}^T" + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}_i^t\\}" + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "text", + "content": ". Starting from an intermediate denoising step " + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "text", + "content": ", we blend these latents as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 353, + 536, + 555, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 536, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 353, + 536, + 555, + 550 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {\\text {e r a s e}} ^ {\\tau} = m _ {i} \\odot \\mathbf {Z} _ {i} ^ {\\tau} + (1 - m _ {i}) \\odot \\mathbf {Z} _ {i - 2} ^ {\\tau} \\tag {10}", + "image_path": "4b732f115ec275de1cff444b75f5040ab4145cb26d2cc9918a730cc0456a3800.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": "Our deletion process operates in two phases: from step " + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\frac{\\tau}{2}" + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": ", we perform latent blending following Eq. (10) with " + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{i-2}^{t}" + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": ". After step " + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\frac{\\tau}{2}" + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": ", we switch to vanilla denoising to ensure a smooth transition. Crucially, we process " + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\text{erase}}" + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": " through MQD without the " + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "inline_equation", + "content": "i-1" + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": " step's mask or prompt embedding, effectively removing the target object's influence." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "This approach offers two key advantages. First, by leveraging BCG to start from step " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " rather than T, we achieve " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " faster editing while preserving background consistency, as we use " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\tau = 8" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " out of a total of 20 steps. Second, MQD's application from step " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " maintains foreground object integrity while effectively removing background elements, eliminating the need for precise object masks in deletion." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7893" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 81, + 108, + 526, + 256 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 2. Quantitative results on our proposed benchmark. Our method shows higher semantic and visual alignment than other baselines, including layout-to-image synthesis and image editing frameworks. Ordering† denotes the adaptation of the framework adequate for iterative image generation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 81, + 108, + 526, + 256 + ], + "lines": [ + { + "bbox": [ + 81, + 108, + 526, + 256 + ], + "spans": [ + { + "bbox": [ + 81, + 108, + 526, + 256 + ], + "type": "table", + "html": "
TypeMethodResolutionSemantic AlignmentVisual Alignment
BLEU-2/3/4↑METEOR↑CLIPcrop↑
2D Layout-to-ImageLayoutGuidance [14]512 × 51236.44 / 26.13 / 18.850.136162.92
NoiseCollage [43]512 × 51255.75 / 42.43 / 32.960.140264.01
NoiseCollage + ordering†512 × 51259.98 / 43.76 / 32.240.146464.10
3D Layout-to-ImageLooseControl [10]512 × 51263.30 / 46.24 / 34.150.137363.13
Image EditingBLD [3]1024 × 102455.30 / 40.38 / 29.580.148062.40
HD-Painter [32]1024 × 102463.29 / 47.63 / 36.280.148464.09
SD3-Inpaint [20]1024 × 102429.90 / 21.64 / 15.780.144563.98
Ours512 × 51261.19 / 45.04 / 34.060.146564.28
Ours1024 × 102464.99 / 47.69 / 36.590.151364.29
", + "image_path": "05f282fd4567a45b0be9cbea2890a33f35adb9867161913c71b7c126e40b92c1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 270, + 188, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 188, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 188, + 283 + ], + "type": "text", + "content": "4. Multi-Edit Benchmark" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 291, + 295, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 291, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 54, + 291, + 295, + 459 + ], + "type": "text", + "content": "Prior benchmarks for image editing [31, 45] are limited to evaluate only single-step editing or limited to instruct-based edit [22], and layout-guided image generation benchmarks are limited to evaluate whether the specific object is generated or not in the bounding box or evaluate only the flat spatial arrangements. [4, 17, 21]. Therefore, we aim to tackle the issue of evaluating the semantic alignment of the generated images by assessing whether the object generated is aligned with the caption that is used for sequential editing. Also, to handle the mask order, we propose to evaluate the generated image in a layer-wise manner aligning with mask orders. Specifically, to evaluate interactive generation and editing, we need to evaluate whether each layer shows the desired object within the given mask." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "content": "We generate the dataset by selecting classes from ImageNet-1K and arranging them in layered compositions with varying degrees of occlusion and spatial relationships, ensuring mask order is a key factor in the generation process as in the Fig. 3 (a). We utilize GPT-4 API to assist in selecting object classes (e.g., 'Lion', 'Tiger', 'Elephant') that naturally fit together, creating realistic compositions, and generating template-based captions for both the global scene and individual object layers to integrate spatial relations into prompt and utilize GPT to make prompt richer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 581, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 581, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 581, + 295, + 700 + ], + "type": "text", + "content": "To evaluate the spatial and semantic alignment between generated images and prompts, we crop each object layer using its corresponding mask and evaluate it individually as in Fig. 3 (b). Both cropped and final images are resized to " + }, + { + "bbox": [ + 54, + 581, + 295, + 700 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 54, + 581, + 295, + 700 + ], + "type": "text", + "content": " for consistency across evaluations. We employ LLaVa [29] to generate captions and compute alignment metrics such as BLEU [36], METEOR [6], ensuring a robust evaluation of both spatial arrangement and semantic accuracy. Note that we average the result from all the layers of each image to evaluate." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "Also, to check each layer's visual alignment of the gen" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 271, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 344 + ], + "type": "text", + "content": "erated object, we crop the image for each iteration and calculate the CLIP score following other layout-to-image models [4, 21, 43] as individual editing step is similar to layout-to-image generation. We discuss and provide detailed descriptions of the dataset creation process, evaluation metrics, and the resulting benchmark in the supplement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 353, + 391, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 391, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 391, + 367 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 373, + 447, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 447, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 447, + 385 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 390, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 486 + ], + "type": "text", + "content": "Baselines. We evaluate our method against state-of-the-art baselines in both image editing and layout-to-image generation. For image editing, we compare with HD-Painter [32], Blended Latent Diffusion [3] with SD-XL [38], and SD3-ControlNet-Inpaint [20]. As sequential editing shares similarities with layout-to-image generation, we also include train-free layout models: NoiseCollage [43] and Layout-Guidance [14], and 3D-lifted model: LooseControl [10]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 487, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 557 + ], + "type": "text", + "content": "Additionally, we adapt NoiseCollage to handle sequential mask inputs (denoted as NoiseCollage + ordering) to explore the potential of converting layout-to-image models for sequential editing tasks. Following our train-free approach using the PixArt-" + }, + { + "bbox": [ + 313, + 487, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 487, + 554, + 557 + ], + "type": "text", + "content": " foundation model [13], all selected baselines operate without additional training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "content": "Implementation Details. We leverage PixArt-" + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "content": " [13], the variant of Diffusion Transformer (DiT) for image editing framework. We utilize a " + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "content": " pretrained model with DPM-Solver with DPMS guidance of 7.5 and a total denoising step of 20. Also, for a fair comparison with models that generate " + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "content": " resolution [14, 43], we compare quantitatively by generating " + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "content": " images with the " + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "content": " pretrained PixArt-" + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "content": " model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 660, + 429, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 429, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 429, + 673 + ], + "type": "text", + "content": "5.2. Quantitative Result" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "We present the quantitative evaluation results on our proposed Multi-Edit Benchmark in Tab. 2. Our method consistently outperforms both image editing and layout-to-image" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7894" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 83, + 145, + 143 + ], + "blocks": [ + { + "bbox": [ + 85, + 72, + 145, + 81 + ], + "lines": [ + { + "bbox": [ + 85, + 72, + 145, + 81 + ], + "spans": [ + { + "bbox": [ + 85, + 72, + 145, + 81 + ], + "type": "text", + "content": "\"Savanna Plains\"" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 83, + 145, + 143 + ], + "lines": [ + { + "bbox": [ + 85, + 83, + 145, + 143 + ], + "spans": [ + { + "bbox": [ + 85, + 83, + 145, + 143 + ], + "type": "image", + "image_path": "3349e8ef666862b985cc620579d0658329eebd9705450015d2491e7c56b2c033.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 144, + 133, + 152 + ], + "lines": [ + { + "bbox": [ + 96, + 144, + 133, + 152 + ], + "spans": [ + { + "bbox": [ + 96, + 144, + 133, + 152 + ], + "type": "text", + "content": "Before Edit" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 158, + 295, + 180 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 180 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on the effect of Query Disentanglement (QD)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 152, + 83, + 213, + 143 + ], + "blocks": [ + { + "bbox": [ + 189, + 73, + 241, + 82 + ], + "lines": [ + { + "bbox": [ + 189, + 73, + 241, + 82 + ], + "spans": [ + { + "bbox": [ + 189, + 73, + 241, + 82 + ], + "type": "text", + "content": "+ \"A flamingo\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 152, + 83, + 213, + 143 + ], + "lines": [ + { + "bbox": [ + 152, + 83, + 213, + 143 + ], + "spans": [ + { + "bbox": [ + 152, + 83, + 213, + 143 + ], + "type": "image", + "image_path": "10ea1e082042a4aa2b7b2f212a08a4ab700d25e78bbc0e0944af5c7fa5e02ded.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 144, + 203, + 152 + ], + "lines": [ + { + "bbox": [ + 163, + 144, + 203, + 152 + ], + "spans": [ + { + "bbox": [ + 163, + 144, + 203, + 152 + ], + "type": "text", + "content": "Edit w/o QD" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 216, + 83, + 276, + 143 + ], + "blocks": [ + { + "bbox": [ + 216, + 83, + 276, + 143 + ], + "lines": [ + { + "bbox": [ + 216, + 83, + 276, + 143 + ], + "spans": [ + { + "bbox": [ + 216, + 83, + 276, + 143 + ], + "type": "image", + "image_path": "16f830dd65b1ee8fcdd96945bfbf27381c746f21c392c5921902cdd896245abc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 144, + 264, + 152 + ], + "lines": [ + { + "bbox": [ + 228, + 144, + 264, + 152 + ], + "spans": [ + { + "bbox": [ + 228, + 144, + 264, + 152 + ], + "type": "text", + "content": "Edit w/QD" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 192, + 295, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 276 + ], + "type": "text", + "content": "belines across key metrics. Among editing baselines, we demonstrate superior performance compared to Blended Latent Diffusion (BLD) [3] with SD-XL, HD-Painter [32], and SD3-ControlNet-Inpaint [20]. Notably, the significant performance gap with SD3-ControlNet-Inpaint can be attributed to the suboptimal property of utilizing ControlNet-Inpainting in multi-step editing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 277, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 348 + ], + "type": "text", + "content": "Our framework shows particular strength in handling sequential editing challenges, demonstrated by consistent improvements across all metrics. This validates our approach's effectiveness in maintaining background consistency while naturally integrating new objects - a key challenge in sequential editing tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 349, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 468 + ], + "type": "text", + "content": "The performance gap grows further relative to 2D layout-to-image baselines, which often struggle with overlapping or cascaded placements. In particular, our method outperforms NoiseCollage by " + }, + { + "bbox": [ + 55, + 349, + 296, + 468 + ], + "type": "inline_equation", + "content": "4 - 5\\%" + }, + { + "bbox": [ + 55, + 349, + 296, + 468 + ], + "type": "text", + "content": " p and surpasses Layout-Guidance by over " + }, + { + "bbox": [ + 55, + 349, + 296, + 468 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 55, + 349, + 296, + 468 + ], + "type": "text", + "content": " p in BLEU and outperforms in all other metrics. Against the 3D layout-to-image method that leverages pseudo depth maps, LooseControl [10] shows improved BLEU compared to the 2D baselines, but ours outperforms on all metrics. (See supplement for comparison with attribute editing on LooseControl.)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 469, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 296, + 578 + ], + "type": "text", + "content": "While we adapt NoiseCollage [43] for iterative image editing (+ ordering), NoiseCollage + ordering achieves better visual alignment scores, indicating the potential viability of adapted layout-to-image approaches. However, its limited improvement across semantic alignment metrics suggests that simple adaptations are insufficient for complex sequential editing scenarios. This underscores the need for specialized approaches like ours that can effectively handle the unique challenges of iterative image editing." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 587, + 149, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 149, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 149, + 601 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "We present an ablation of our key components (Tab. 3), starting from PixArt- " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " [13]. Background consistency guidance (BCG) reuses latents for faster editing (Tab. 1), improving BLEU and CLIP while preserving METEOR. Next, Query-Disentanglement (QD), inspired by NoiseCollage [43], preserves the existing background rather than fully covering the mask, raising BLEU but slightly reducing CLIP. Direct latent blending instead inflates CLIP by overfilling the mask, lacking context (Fig. 4). Finally, extend" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 71, + 554, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 554, + 126 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 554, + 126 + ], + "type": "text", + "content": "Table 3. Ablation study of our proposed method. Results are reported on both semantic and visual alignment metrics. We denote vanilla PixArt-" + }, + { + "bbox": [ + 313, + 71, + 554, + 126 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 71, + 554, + 126 + ], + "type": "text", + "content": " inpainting model as Baseline. Also we denote QD as Query-Disentanglement without memory, only with the current object and background." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 328, + 129, + 539, + 203 + ], + "blocks": [ + { + "bbox": [ + 328, + 129, + 539, + 203 + ], + "lines": [ + { + "bbox": [ + 328, + 129, + 539, + 203 + ], + "spans": [ + { + "bbox": [ + 328, + 129, + 539, + 203 + ], + "type": "table", + "html": "
MethodSemantic Align.Visual Align.
BLEU-2/3/4↑METEOR↑CLIPc↑
Baseline56.29 / 42.04 / 33.060.158664.05
+BCG60.74 / 46.27 / 35.200.158564.10
+QD62.68 / 46.42 / 35.030.153063.99
Ours64.99 / 47.69 / 36.590.151364.29
", + "image_path": "48f25beda0401dafb8a9bba6c5a184fcd79354f83897d517ad92782f708ebd55.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 384, + 209, + 410, + 232 + ], + "blocks": [ + { + "bbox": [ + 384, + 209, + 410, + 232 + ], + "lines": [ + { + "bbox": [ + 384, + 209, + 410, + 232 + ], + "spans": [ + { + "bbox": [ + 384, + 209, + 410, + 232 + ], + "type": "image", + "image_path": "d4f6874d6f7b55440dd5e9475b6e4c5d8e8987aed43525e9c5b78e21a31bb644.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 232, + 412, + 240 + ], + "lines": [ + { + "bbox": [ + 384, + 232, + 412, + 240 + ], + "spans": [ + { + "bbox": [ + 384, + 232, + 412, + 240 + ], + "type": "text", + "content": "\"A beach\"" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 429, + 209, + 453, + 232 + ], + "blocks": [ + { + "bbox": [ + 429, + 209, + 453, + 232 + ], + "lines": [ + { + "bbox": [ + 429, + 209, + 453, + 232 + ], + "spans": [ + { + "bbox": [ + 429, + 209, + 453, + 232 + ], + "type": "image", + "image_path": "0b7ba283c8e3ad56438116863c387533bd2fca8a56d1e16a6506dd4cecb9cf88.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 232, + 459, + 240 + ], + "lines": [ + { + "bbox": [ + 419, + 232, + 459, + 240 + ], + "spans": [ + { + "bbox": [ + 419, + 232, + 459, + 240 + ], + "type": "text", + "content": "+“a blue bus\"" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 473, + 209, + 497, + 232 + ], + "blocks": [ + { + "bbox": [ + 473, + 209, + 497, + 232 + ], + "lines": [ + { + "bbox": [ + 473, + 209, + 497, + 232 + ], + "spans": [ + { + "bbox": [ + 473, + 209, + 497, + 232 + ], + "type": "image", + "image_path": "c29dc64bdba4590f0d3f43ee11e838020e44669cb21046da243d2ad11df75f93.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 328, + 248, + 400, + 320 + ], + "blocks": [ + { + "bbox": [ + 328, + 248, + 400, + 320 + ], + "lines": [ + { + "bbox": [ + 328, + 248, + 400, + 320 + ], + "spans": [ + { + "bbox": [ + 328, + 248, + 400, + 320 + ], + "type": "image", + "image_path": "2595126f4b4c2ccc3c831e1a0421ee0cac385dd879c6cd70ce294dc893a92c4b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 328, + 322, + 400, + 395 + ], + "blocks": [ + { + "bbox": [ + 328, + 322, + 400, + 395 + ], + "lines": [ + { + "bbox": [ + 328, + 322, + 400, + 395 + ], + "spans": [ + { + "bbox": [ + 328, + 322, + 400, + 395 + ], + "type": "image", + "image_path": "b3b4bf6f43b39d757c84d6ae8e8fa8407bbd438ddbf8b588df5345ca3191ea71.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 396, + 381, + 403 + ], + "lines": [ + { + "bbox": [ + 348, + 396, + 381, + 403 + ], + "spans": [ + { + "bbox": [ + 348, + 396, + 381, + 403 + ], + "type": "text", + "content": "HD Painter" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 406, + 554, + 449 + ], + "lines": [ + { + "bbox": [ + 313, + 406, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 554, + 449 + ], + "type": "text", + "content": "Figure 5. Comparison in image editing capability with latest image editing models. [3, 32] Note that the initial image is generated by our framework, which is equivalent to PixArt- " + }, + { + "bbox": [ + 313, + 406, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 406, + 554, + 449 + ], + "type": "text", + "content": " [13] with no mask input." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 402, + 249, + 475, + 321 + ], + "blocks": [ + { + "bbox": [ + 463, + 232, + 511, + 249 + ], + "lines": [ + { + "bbox": [ + 463, + 232, + 511, + 249 + ], + "spans": [ + { + "bbox": [ + 463, + 232, + 511, + 249 + ], + "type": "text", + "content": "+“a hotel behind the bus\"" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 402, + 249, + 475, + 321 + ], + "lines": [ + { + "bbox": [ + 402, + 249, + 475, + 321 + ], + "spans": [ + { + "bbox": [ + 402, + 249, + 475, + 321 + ], + "type": "image", + "image_path": "08d0a8d4c7035059c851cadea660a63db8e7e5253aafe4101026c6bf359f5bbb.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 403, + 322, + 475, + 395 + ], + "blocks": [ + { + "bbox": [ + 403, + 322, + 475, + 395 + ], + "lines": [ + { + "bbox": [ + 403, + 322, + 475, + 395 + ], + "spans": [ + { + "bbox": [ + 403, + 322, + 475, + 395 + ], + "type": "image", + "image_path": "000f7e58b0bd78ec1db515d762b4a9df8c57e419e6cbecd712b944394b610208.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 396, + 477, + 403 + ], + "lines": [ + { + "bbox": [ + 405, + 396, + 477, + 403 + ], + "spans": [ + { + "bbox": [ + 405, + 396, + 477, + 403 + ], + "type": "text", + "content": "Blended Latent Diffuion" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 476, + 249, + 549, + 321 + ], + "blocks": [ + { + "bbox": [ + 476, + 249, + 549, + 321 + ], + "lines": [ + { + "bbox": [ + 476, + 249, + 549, + 321 + ], + "spans": [ + { + "bbox": [ + 476, + 249, + 549, + 321 + ], + "type": "image", + "image_path": "818b02e5baaa063e5ffc885cd3869b03d45998f95277ecb183616e9141581a6e.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 476, + 322, + 549, + 395 + ], + "blocks": [ + { + "bbox": [ + 476, + 322, + 549, + 395 + ], + "lines": [ + { + "bbox": [ + 476, + 322, + 549, + 395 + ], + "spans": [ + { + "bbox": [ + 476, + 322, + 549, + 395 + ], + "type": "image", + "image_path": "d4d11c282c2525a0869a0bb24368763b9a4428b6fe12f185a7bf545a25f26ae8.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 505, + 396, + 520, + 403 + ], + "lines": [ + { + "bbox": [ + 505, + 396, + 520, + 403 + ], + "spans": [ + { + "bbox": [ + 505, + 396, + 520, + 403 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 457, + 554, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 554, + 494 + ], + "type": "text", + "content": "ing QD to multi-query disentanglement (MQD) with layerwise memory integrates new objects into all prior prompts, achieving higher BLEU-2/3/4 and CLIP scores." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 502, + 422, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 422, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 422, + 515 + ], + "type": "text", + "content": "5.4. Qualitative Result" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "type": "text", + "content": "Comparison with Editing Approaches under Interactive Scenario. In Fig. 5, we demonstrate our method's effectiveness in iterative editing compared to state-of-the-art single-step editing approaches [2, 32]. Starting with a background of \"A beach\" generated by vanilla PixArt-" + }, + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "type": "text", + "content": " [13], we sequentially place a bus and a hotel. Our method successfully renders the hotel behind the bus, accurately interpreting both the prompt \"A beach\" and the mask order through MQD to achieve the intended spatial arrangement." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": "In contrast, Blended Latent Diffusion [3] struggles with the initial placement of \"bus\" in the complex background, while HD-Painter [32] shows inconsistencies in the appearance of \"bus\" (highlighted in red in Fig. 5). Our method maintains background consistency while naturally integrating new objects into the scene." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "type": "text", + "content": "Improved Editability. Fig. 6 compares our method with" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7895" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 71, + 282, + 183 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 282, + 183 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 282, + 183 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 282, + 183 + ], + "type": "image", + "image_path": "dc416de1d738b12fafb772c03447940e45cdfe8e61243e519fdc371a654ba223.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 296, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 242 + ], + "type": "text", + "content": "Figure 6. Improved editability of image through Background Consistency Guidance and Multi-Query Disentangled cross attention. Through recycling the previous step's latents, we can remove the object that is behind the foreground object, enabling enhanced editability of the image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 74, + 308, + 275, + 374 + ], + "blocks": [ + { + "bbox": [ + 55, + 251, + 296, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 304 + ], + "type": "text", + "content": "Table 4. Human preference study with recent literature on image editing on our benchmark. We utilize HD-Painter [32] and Blended Latent Diffusion with SD-XL [38] to compare under iterative editing scenario and compare the preference on a 5-point Likert scale." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 308, + 275, + 374 + ], + "lines": [ + { + "bbox": [ + 74, + 308, + 275, + 374 + ], + "spans": [ + { + "bbox": [ + 74, + 308, + 275, + 374 + ], + "type": "table", + "html": "
MethodBackground ConsistencyNatural AdaptationText-scene Alignment
HD-Painter [32]3.712.813.08
BLD [3]3.432.242.04
Ours4.594.284.49
", + "image_path": "ad0fe648b5fbf9791483deea195f9ce5aeda01719613d8f8fb80a94c0fa8632b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 479 + ], + "type": "text", + "content": "Photoshop 2025 [1]. Placing \"a glass of wine\" or \"a lizard\" is straightforward, but removing \"a glass of wine\" while preserving \"a lizard\" remains challenging for Photoshop's generative fill, which requires precise brush strokes and often leaves artifacts (red box). Our approach simplifies this by leveraging existing masks and prior-step latents to naturally inpaint the background while applying MQD during partial denoising to preserve the lizard's structure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 481, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 577 + ], + "type": "text", + "content": "Comparison with Existing T2I models. Fig. 7 showcases our method's ability to handle complex multi-object scenarios that challenge conventional text-to-image models. For instance, in row A, only our framework successfully generates the intended composition of a lion behind a flower. Row B and C further demonstrate the challenge for existing T2I approaches to handle complex spatial arrangements, while illustrating our method's capability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 586, + 197, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 197, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 197, + 600 + ], + "type": "text", + "content": "5.5. Human Preference Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 676 + ], + "type": "text", + "content": "We conduct a comprehensive human evaluation study with 50 participants to assess the quality of our method compared to the latest editing models [3, 32] on our benchmark. The evaluation focuses on three aspects: background consistency, natural adaptation, and text-scene alignment, rated on a 5-point Likert scale." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Our method outperforms single-step editing approaches [3, 32] across all three categories. These results demonstrate that traditional editing models, while effective" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 70, + 399, + 137 + ], + "blocks": [ + { + "bbox": [ + 320, + 100, + 327, + 108 + ], + "lines": [ + { + "bbox": [ + 320, + 100, + 327, + 108 + ], + "spans": [ + { + "bbox": [ + 320, + 100, + 327, + 108 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 70, + 399, + 137 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 399, + 137 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 399, + 137 + ], + "type": "image", + "image_path": "5521930328725612bc14122e12a920d8f6738b17297f1d63cf55fa1ae728029d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 400, + 70, + 471, + 137 + ], + "blocks": [ + { + "bbox": [ + 400, + 70, + 471, + 137 + ], + "lines": [ + { + "bbox": [ + 400, + 70, + 471, + 137 + ], + "spans": [ + { + "bbox": [ + 400, + 70, + 471, + 137 + ], + "type": "image", + "image_path": "4381c49380e848f98f561e81b26924398e55efd6a4f2377c95b22edd49bacbec.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 471, + 70, + 543, + 137 + ], + "blocks": [ + { + "bbox": [ + 471, + 70, + 543, + 137 + ], + "lines": [ + { + "bbox": [ + 471, + 70, + 543, + 137 + ], + "spans": [ + { + "bbox": [ + 471, + 70, + 543, + 137 + ], + "type": "image", + "image_path": "17869df27a5a503c91705d8df9c6b4f9f2b7d9c7b1859dc1b4706a2ab06dd265.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 329, + 138, + 399, + 206 + ], + "blocks": [ + { + "bbox": [ + 320, + 170, + 326, + 177 + ], + "lines": [ + { + "bbox": [ + 320, + 170, + 326, + 177 + ], + "spans": [ + { + "bbox": [ + 320, + 170, + 326, + 177 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 138, + 399, + 206 + ], + "lines": [ + { + "bbox": [ + 329, + 138, + 399, + 206 + ], + "spans": [ + { + "bbox": [ + 329, + 138, + 399, + 206 + ], + "type": "image", + "image_path": "adc244b570684764c6863bf04dcc927835ba84353f5dd0c9b2f108fc670c2126.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 400, + 138, + 471, + 206 + ], + "blocks": [ + { + "bbox": [ + 400, + 138, + 471, + 206 + ], + "lines": [ + { + "bbox": [ + 400, + 138, + 471, + 206 + ], + "spans": [ + { + "bbox": [ + 400, + 138, + 471, + 206 + ], + "type": "image", + "image_path": "c74a2b1302ed49ecf4c7bf97bc09e16b45048174b10328e1ee6bfd7ac15cd078.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 471, + 138, + 543, + 206 + ], + "blocks": [ + { + "bbox": [ + 471, + 138, + 543, + 206 + ], + "lines": [ + { + "bbox": [ + 471, + 138, + 543, + 206 + ], + "spans": [ + { + "bbox": [ + 471, + 138, + 543, + 206 + ], + "type": "image", + "image_path": "ff4f4dfe954ac7c2b52de83217a07507e95155d3d7eaf3d23207fe67141b6c3a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 319, + 538, + 337 + ], + "lines": [ + { + "bbox": [ + 320, + 319, + 538, + 337 + ], + "spans": [ + { + "bbox": [ + 320, + 319, + 538, + 337 + ], + "type": "text", + "content": "C. A wooden table has a flower pot on the left and some stacked books in the center, with a small statue resting behind the table." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 329, + 207, + 399, + 275 + ], + "blocks": [ + { + "bbox": [ + 320, + 239, + 326, + 246 + ], + "lines": [ + { + "bbox": [ + 320, + 239, + 326, + 246 + ], + "spans": [ + { + "bbox": [ + 320, + 239, + 326, + 246 + ], + "type": "text", + "content": "C" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 207, + 399, + 275 + ], + "lines": [ + { + "bbox": [ + 329, + 207, + 399, + 275 + ], + "spans": [ + { + "bbox": [ + 329, + 207, + 399, + 275 + ], + "type": "image", + "image_path": "d09690a037f82844e1c23a78c3963570416cc67d2c176a1f320b31b3d77249ad.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 276, + 378, + 283 + ], + "lines": [ + { + "bbox": [ + 349, + 276, + 378, + 283 + ], + "spans": [ + { + "bbox": [ + 349, + 276, + 378, + 283 + ], + "type": "text", + "content": "PIXART-a" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 285, + 545, + 302 + ], + "lines": [ + { + "bbox": [ + 320, + 285, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 320, + 285, + 545, + 302 + ], + "type": "text", + "content": "A. A lion stands behind flowers in the desert, positioned next to the water where a palm tree is growing nearby under the bright sun." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 313, + 341, + 554, + 374 + ], + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 374 + ], + "type": "text", + "content": "Figure 7. Comparison in interactive scenarios with existing T2I generative models. Stable Diffusion XL [38] and PixArt- " + }, + { + "bbox": [ + 313, + 341, + 554, + 374 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 341, + 554, + 374 + ], + "type": "text", + "content": " [13] use text input only." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 400, + 207, + 471, + 275 + ], + "blocks": [ + { + "bbox": [ + 400, + 207, + 471, + 275 + ], + "lines": [ + { + "bbox": [ + 400, + 207, + 471, + 275 + ], + "spans": [ + { + "bbox": [ + 400, + 207, + 471, + 275 + ], + "type": "image", + "image_path": "bfc138d8bdf7561517cefbb74fea5851225a8376593becc78d8490766a44dfe5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 276, + 462, + 283 + ], + "lines": [ + { + "bbox": [ + 408, + 276, + 462, + 283 + ], + "spans": [ + { + "bbox": [ + 408, + 276, + 462, + 283 + ], + "type": "text", + "content": "Stable Diffusion XL" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 303, + 544, + 319 + ], + "lines": [ + { + "bbox": [ + 320, + 303, + 544, + 319 + ], + "spans": [ + { + "bbox": [ + 320, + 303, + 544, + 319 + ], + "type": "text", + "content": "B. Santa Claus and a teddy bear are sitting by a campfire in a snowy forest, with a gingerbread house directly behind them under falling snow." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 471, + 207, + 543, + 275 + ], + "blocks": [ + { + "bbox": [ + 471, + 207, + 543, + 275 + ], + "lines": [ + { + "bbox": [ + 471, + 207, + 543, + 275 + ], + "spans": [ + { + "bbox": [ + 471, + 207, + 543, + 275 + ], + "type": "image", + "image_path": "ea7f6ea9b9f5ea642f4b6747ed5af887b98f2a0c00d12924dc8debb83e9a7aac.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 277, + 515, + 283 + ], + "lines": [ + { + "bbox": [ + 500, + 277, + 515, + 283 + ], + "spans": [ + { + "bbox": [ + 500, + 277, + 515, + 283 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 384, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 456 + ], + "type": "text", + "content": "for single-step modifications, are insufficient for sequential editing. This underscores the necessity of specialized approaches like ours for handling sequential iterative editing scenarios. Additional human evaluation results comparing our method with text-to-image models [13, 38] are provided in the supplementary materials." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 469, + 388, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 469, + 388, + 481 + ], + "spans": [ + { + "bbox": [ + 314, + 469, + 388, + 481 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 490, + 555, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 555, + 645 + ], + "type": "text", + "content": "We propose a novel framework for sequential image editing, which poses challenges in maintaining background consistency and seamless integration of new objects into the scenes. Our method includes three key components: Layer-wise Memory to preserve previous content, Background Consistency Guidance to keep the background stable with faster editing, and Multi-Query Disentangled Cross-Attention for natural object adaptation. Experiments on our proposed benchmark dataset show that our approach outperforms state-of-the-art methods in semantic and visual alignment, making it ideal for complex, iterative image editing scenarios. This framework paves the way for future work in improved interactive editing." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 657, + 555, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 555, + 717 + ], + "type": "text", + "content": "Limitation and Future Work. Since our approach utilizes image editing, generating multiple objects takes longer, depending on the number of edits. Also, utilizing layer-wise memory requires additional memory costs. We plan to make it more efficient for faster editing in the future." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7896" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by IITP grant (RS-2021-II211343: AI Graduate School Program at Seoul National University " + }, + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "type": "inline_equation", + "content": "(5\\%)" + }, + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "type": "text", + "content": " and RS-2024-00509257: Global AI Frontier Lab " + }, + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "type": "inline_equation", + "content": "(30\\%)" + }, + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "type": "text", + "content": ") and NRF grant (No.RS-2024-00405857 " + }, + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "type": "inline_equation", + "content": "(65\\%)" + }, + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "type": "text", + "content": ") funded by the Korea government (MSIT)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 177, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 177, + 294, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 177, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 177, + 294, + 209 + ], + "type": "text", + "content": "[1] Adobe Inc. Adobe Photoshop. https://www.adobe.com/products/photoshop.html, 2024. Version 2025. 1, 2, 8, 17, 20" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 211, + 295, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 211, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 211, + 295, + 243 + ], + "type": "text", + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, 2022. 2, 7, 16" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 246, + 295, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 246, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 295, + 277 + ], + "type": "text", + "content": "[3] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM TOG, 2023. 1, 2, 4, 6, 7, 8, 11, 13, 17, 20" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "text", + "content": "[4] Eslam Mohamed Bakr, Pengzhan Sun, Xiaogian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models. In ICCV, 2023. 2, 6, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 326, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 326, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 326, + 294, + 369 + ], + "type": "text", + "content": "[5] Coloma Ballester, Marcelo Bertalmio, Vicent Caselles, Guillermo Sapiro, and Joan Verdera. Filling-in by joint interpolation of vector fields and gray levels. IEEE TIP, 2001. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 372, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 294, + 404 + ], + "type": "text", + "content": "[6] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In IEEvaluation@ACL, 2005. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 406, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 294, + 437 + ], + "type": "text", + "content": "[7] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 440, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 440, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 440, + 294, + 483 + ], + "type": "text", + "content": "[8] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM TOG, 2009. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 486, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 486, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 62, + 486, + 294, + 517 + ], + "type": "text", + "content": "[9] Marcelo Bertalmio, Guillermo Sapiro, Vicent Caseles, and Coloma Ballester. Image inpainting. In SIGGRAPH, 2000. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 520, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 520, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 520, + 294, + 553 + ], + "type": "text", + "content": "[10] Shariq Farooq Bhat, Niloy J. Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning. In SIGGRAPH, 2024. 6, 7, 16" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 280, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 280, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 280, + 566 + ], + "type": "text", + "content": "[11] BlackForestLabs. Flux, 2023. Accessed: 2024-11-12. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 567, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 621 + ], + "type": "text", + "content": "[12] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- " + }, + { + "bbox": [ + 57, + 567, + 295, + 621 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 57, + 567, + 295, + 621 + ], + "type": "text", + "content": " : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv:2403.04692, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "type": "text", + "content": "[13] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 1, 3, 6, 7, 8, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "text", + "content": "[14] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In WACV, 2024. 2, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Ruidong Chen, Lanjun Wang, Weizhi Nie, Yongdong Zhang, and An-An Liu. Anyscene: Customized image synthesis with composited foreground. In CVPR, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "text", + "content": "[16] Yifu Chen, Jingwen Chen, Yingwei Pan, Yehao Li, Ting Yao, Zhineng Chen, and Tao Mei. Improving text-guided object inpainting with semantic pre-inpainting. In ECCV, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "type": "text", + "content": "[17] Jaemin Cho, Linjie Li, Zhengyuan Yang, Zhe Gan, Lijuan Wang, and Mohit Bansal. Diagnostic benchmark and iterative inpainting for layout-guided image generation. In CVPR Workshop, 2024. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 189, + 553, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 189, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 553, + 221 + ], + "type": "text", + "content": "[18] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE TIP, 2004. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "type": "text", + "content": "[19] Abdelrahman Eldisokey and Peter Wonka. Build-a-scene: Interactive 3d layout control for diffusion-based image generation. In ICLR, 2025. 16" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 258, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 553, + 334 + ], + "type": "text", + "content": "[20] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 6, 7, 11" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 335, + 553, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 335, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 553, + 390 + ], + "type": "text", + "content": "[21] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. In NeurIPS, 2023. 2, 6, 14" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "type": "text", + "content": "[22] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv:2405.04007, 2024. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 428, + 553, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 553, + 471 + ], + "type": "text", + "content": "[23] Biao Gong, Siteng Huang, Yutong Feng, Shiwei Zhang, Yuyuan Li, and Yu Liu. Check locate rectify: A training-free layout calibration system for text-to-image generation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 473, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 473, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 473, + 553, + 495 + ], + "type": "text", + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 498, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 498, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 498, + 553, + 552 + ], + "type": "text", + "content": "[25] Runhui Huang, Kaixin Cai, Jianhua Han, Xiaodan Liang, Renjing Pei, Guansong Lu, Songcen Xu, Wei Zhang, and Hang Xu. Layerdiff: Exploring text-guided multi-layered composable image synthesis via layer-collaborative diffusion model. In ECCV, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 554, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 554, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 317, + 554, + 553, + 597 + ], + "type": "text", + "content": "[26] Jonghyun Lee, Hansam Cho, Youngjoon Yoo, Seoung Bum Kim, and Yonghyun Jeong. Compose and conquer: Diffusion-based 3d depth aware composable image synthesis. In ICLR, 2024. 2, 17" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "type": "text", + "content": "[27] Jaerin Lee, Daniel Sungho Jung, Kanggeon Lee, and Young Mu Lee. StreamMultiDiffusion: real-time interactive generation with region-based semantic control. arXiv:2403.09055, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[28] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 2, 13" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 6" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7897" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[30] Zhengyao Lv, Yuxiang Wei, Wangmeng Zuo, and KwanYee K Wong. Place: Adaptive layout-semantic fusion for semantic image synthesis. In CVPR, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "type": "text", + "content": "[31] Yiwei Ma, Jiayi Ji, Ke Ye, Weihuang Lin, Zhibin Wang, Yonghan Zheng, Qiang Zhou, Xiaoshuai Sun, and Rongrong Ji. I2ebench: A comprehensive benchmark for instruction-based image editing. In NeurIPS, 2024. 2, 6, 14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[32] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: High-resolution and prompt-faithful text-guided image inpainting with diffusion models. In ICLR, 2025. 1, 2, 6, 7, 8, 11, 16, 17, 20" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 250 + ], + "type": "text", + "content": "[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 17" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "type": "text", + "content": "[34] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "text", + "content": "[35] Yotam Nitzan, Zongze Wu, Richard Zhang, Eli Shechtman, Daniel Cohen-Or, Taesung Park, and Michael Gharbi. Lazy diffusion transformer for interactive image editing. In ECCV, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "type": "text", + "content": "[36] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 403 + ], + "type": "text", + "content": "[37] Pincel. Pincel: Ai image editor, 2024. Accessed: 2024-11-15. 1, 2, 17, 20" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "text", + "content": "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 6, 8, 11, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 535 + ], + "type": "text", + "content": "[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. In ICLR, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 536, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 294, + 568 + ], + "type": "text", + "content": "[40] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 569, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 602 + ], + "type": "text", + "content": "[41] Jiawei Ren, Mengmeng Xu, Jui-Chieh Wu, Ziwei Liu, Tao Xiang, and Antoine Toisoul. Move anything with layered scene diffusion. In CVPR, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 669 + ], + "type": "text", + "content": "[43] Takahiro Shirakawa and Seiichi Uchida. Noisecollapse: A layout-aware text-to-image diffusion model based on noise cropping and merging. In CVPR, 2024. 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[44] Jack Urbanek, Florian Bordes, Pietro Astolfi, Mary Williamson, Vasu Sharma, and Adriana Romero-Soriano. A picture is worth more than 77 text tokens: Evaluating clip-style models on dense captions. In CVPR, 2024. 15" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 589 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 140 + ], + "type": "text", + "content": "[45] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, 2023. 2, 6, 13, 14" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 173 + ], + "type": "text", + "content": "[46] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "text", + "content": "[47] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In ICCV, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 218, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 218, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 218, + 553, + 251 + ], + "type": "text", + "content": "[48] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 251, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 285 + ], + "type": "text", + "content": "[49] Chao Yang, Xin Lu, Zhe Lin, Eli Shechtman, Oliver Wang, and Hao Li. High-resolution image inpainting using multiscale neural patch synthesis. In CVPR, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 285, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 319 + ], + "type": "text", + "content": "[50] Zili Yi, Qiang Tang, Shekoofeh Azizi, Daesik Jang, and Zhan Xu. Contextual residual aggregation for ultra high-resolution image inpainting. In CVPR, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 319, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 353 + ], + "type": "text", + "content": "[51] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas Huang. Free-form image inpainting with gated convolution. In ICCV, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[52] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. In ECCV, 2024. 15" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 410 + ], + "type": "text", + "content": "[53] Lisai Zhang, Qingcai Chen, Baotian Hu, and Shuang Jiang. Text-guided neural image inpainting. In ACM MM, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "text", + "content": "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 11, 17" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "type": "text", + "content": "[55] Xinyang Zhang, Wentian Zhao, Xin Lu, and Jeff Chien. Text2layer: Layered image generation using latent diffusion model. arXiv:2307.09781, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 478, + 553, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 501 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 501 + ], + "type": "text", + "content": "[56] Bo Zhao, Lili Meng, Weidong Yin, and Leonid Sigal. Image generation from layout. In CVPR, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 502, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 553, + 544 + ], + "type": "text", + "content": "[57] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. In NeurIPS, 2023. 17" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 546, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 553, + 589 + ], + "type": "text", + "content": "[58] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7898" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improving Gaussian Splatting with Localized Points Management/6436626a-e39e-4d87-9a2d-0ebf83ce94c9_content_list.json b/2025/Improving Gaussian Splatting with Localized Points Management/6436626a-e39e-4d87-9a2d-0ebf83ce94c9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a40a9d9766e71104c9b909308eb0b79362ecd91 --- /dev/null +++ b/2025/Improving Gaussian Splatting with Localized Points Management/6436626a-e39e-4d87-9a2d-0ebf83ce94c9_content_list.json @@ -0,0 +1,1386 @@ +[ + { + "type": "text", + "text": "Improving Gaussian Splating with Localized Points Management", + "text_level": 1, + "bbox": [ + 166, + 130, + 831, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haosen Yang $^{1*}$ Chenhao Zhang $^{1*}$ Wenqing Wang $^{1*}$ Marco Volino $^{1}$ Adrian Hilton $^{1}$ Li Zhang $^{2}$ Xiatian Zhu $^{1}$", + "bbox": [ + 130, + 178, + 864, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ University of Surrey $^{2}$ Fudan University *Equal contribution", + "bbox": [ + 328, + 232, + 666, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://happy-hsy.github.io/projects/LPM/", + "bbox": [ + 328, + 286, + 661, + 305 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/358cb499d038a196051c1c0a34f80fbbe0ca0cbe4eff80ce263410d0b97699b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 339, + 380, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0c2e3bbb8d62f14687f3298dbb0a3d6a38241c548439c70595cd677fe39328cd.jpg", + "image_caption": [ + "Rendered Image" + ], + "image_footnote": [], + "bbox": [ + 101, + 450, + 380, + 561 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec2687b81814dd39bb44334649a0648956d27ce407b14cc9374bacfd324b3459.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 340, + 632, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f99e428bbbd51fb643b315becf2f4502fc86ce2ea23d5fbf19ad39d7da4d4f40.jpg", + "image_caption": [ + "Gaussian Points" + ], + "image_footnote": [], + "bbox": [ + 380, + 450, + 632, + 561 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/33c209ad02985e0a553243f04c4eb01edb0ade0279341fc74ad5010928b0c955.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 339, + 887, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7fdbb11f6ea55d5f47d94bfe05b980fdb4a0d527fb02e85ea8b0bde007b36fa4.jpg", + "image_caption": [ + "Rendered Depth", + "Figure 1. Visualization of points behavior. 3DGS produces ill-conditioned Gaussians (red box) that occlude other valid points, resulting in noticeably incorrect depth estimation. LPM handles these ill-conditioned points to reduce negative impacts and further calibrate the geometry." + ], + "image_footnote": [], + "bbox": [ + 635, + 450, + 888, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 657, + 326, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point management is critical for optimizing 3D Gaussian Splatting models, as point initiation (e.g., via structure from motion) is often distributionally inappropriate. Typically, Adaptive Density Control (ADC) algorithm is adopted, leveraging view-averaged gradient magnitude thresholding for point densification, opacity thresholding for pruning, and regular all-points opacity reset. We reveal that this strategy is limited in tackling intricate/special image regions (e.g., transparent) due to inability of identifying all 3D zones requiring point densification, and lacking an appropriate mechanism to handle ill-conditioned points with negative impacts (e.g., occlusion due to false high opacity). To address these limitations, we propose a Localized Point Management (LPM) strategy, capable of identify", + "bbox": [ + 88, + 689, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing those error-contributing zones in greatest need for both point addition and geometry calibration. Zone identification is achieved by leveraging the underlying multiview geometry constraints, subject to image rendering errors. We apply point densification in the identified zones and then reset the opacity of the points in front of these regions, creating a new opportunity to correct poorly conditioned points. Serving as a versatile plugin, LPM can be seamlessly integrated into existing static 3D and dynamic 4D Gaussian Splatting models with minimal additional cost. Experimental evaluations validate the efficacy of our LPM in boosting a variety of existing 3D/4D models both quantitatively and qualitatively. Notably, LPM improves both static 3DGS and dynamic SpaceTimeGS to achieve state-of-the-art rendering quality while retaining real-time speeds, excelling on challenging datasets such as Tanks & Temples and the Neural", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21696", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Video dataset.", + "text_level": 1, + "bbox": [ + 91, + 90, + 210, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 119, + 222, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural rendering has emerged as a generalizable, flexible, and powerful approach for photorealistic novel view synthesis (NVS) of any camera poses [23], underpinning a wide variety of applications in augmented/virtual/mixed reality [9], robotics [37], and generation [25], among more others. For example, taking a learning-based parametric idea, Neural Radiance Fields (NeRFs) [23] implicitly represent the scene radiance of any complexity using neural networks (e.g., MLPs), without the tedious requirements of model handcrafting for accounting the scene variations in geometry, texture, illumination. However, their view rendering is inefficient computationally due to heavy ray sampling, thus suffer in scaling to high-resolution content applications and large scale scene modeling [29, 32].", + "bbox": [ + 89, + 145, + 480, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, 3D Gaussian Splatting (3DGS) [14] has come as an alternative with explicit representation, much faster model optimization and real-time neural rendering. The process begins by initializing a set of 3D Gaussian points using Structure from Motion (SfM) [28]. This is followed by optimizing the parameters of these points through view reconstruction loss, resulting in a view output generated with differentiable splatting-based rasterization. However, the point initialization is often distributionally non-optimal, leading to issues such as under-population (e.g., insufficient points) or over-population (e.g., excessive points) in the 3D space. Consequently, a point management mechanism, such as Adaptive Density Control (ADC), is necessary during optimization. However, we identify several limitations with ADC: (i) Thresholding the average gradient to determine regions for point densification often overlooks under-optimized points. For instance, larger Gaussian points typically have lower average gradients and may frequently appear across various views in screen space. (ii) Point sparsity complicates the addition of sufficient and reliable points needed to comprehensively cover the scene. (iii) Mis-optimized Gaussian points can have detrimental effects, such as occluding other valuable points and leading to incorrect depth estimates (see erroneous placements on windows in Fig. 1). While GaussianPro [7] and PixelGS [41] try to solve (1) and (2) through multi-view stereo [2] and additional gradient propagation, respectively, these methods significantly increase the training budget, as shown in the Table 4.", + "bbox": [ + 89, + 356, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the aforementioned limitations, in this paper we propose a novel and efficient Localized Point Management (LPM) approach. Our idea is intuitive - identifying those 3D Gaussian points leading to rendering errors. Thus we start with an image rendering error map of a specific view. To obtain the error contributing 3D points, we leverage the region correspondence between different", + "bbox": [ + 89, + 795, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "views via feature mapping, subject to the multiview geometry constraint. For each pair of corresponded regions, we cast the rays through them at their respective camera views in the cone shape, and consider their intersection as the error source zone. Within each such zone, we consider two situations: (1) At presence of points, we further apply point densification at a lower threshold to complement the original counterpart locally; (2) In case no point due to point sparsity, we add new Gaussian points. Concurrently, we reset the opacity of points with high opacity estimates that are located in front of these zones, as they can significantly affect view rendering. This provides an opportunity to correct potentially ill-conditioned points while tuning the newly added ones in the subsequent optimization. To minimize model expansion, we prune the points by opacity in a density-aware manner.", + "bbox": [ + 511, + 90, + 903, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize the contributions below: (1) Through in-depth analysis, we have identified several limitations in the standard point management mechanism used in Gaussian Splatting that impede model optimization. (2) We present Localized Point Management (LPM) for these issues by identifying error-contributing 3D zones and implementing appropriate operations for point densification and opacity reset. (3) Extensive experiments validate the benefits of our LPM in improving a diversity of existing 3D and 4D Gaussian Splatting models in novel view synthesis on both static and dynamic scenes.", + "bbox": [ + 511, + 333, + 903, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 513, + 653, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Scene Representations has always been an important direction in novel view synthesis. Previous methods allocate neural features to structures such as volume [21, 27], texture [31], and point cloud [1]. The pioneering work of NeRF [23] proposes integrating neural networks with 3D volumetric representations to convert a 3D scene into a learnable density field, enabling high-quality novel view synthesis without requiring explicit modeling of the 3D scene and illumination. Later on, numerous works emerge to boost the quality and efficiency of volume rendering, [3, 5, 35] refine the point sampling strategy in ray marching, some some advanced works [4, 33] reparameterize the scene to produce a more compact representation. Additionally, regularization terms [8, 39] can be incorporated to constrain the scene representation, resulting in a closer approximation to real-world geometry. Despite their high-quality representational performance, these methods are typically computationally inefficient for view rendering due to the extensive ray sampling required and the use of Multi-Layer Perceptrons (MLPs) to represent the scene, complicating the computation and optimization of any point within the scene. To address this, several works have proposed novel scene representations aimed at accelerating the rendering process. These representations replace MLPs with sparse", + "bbox": [ + 511, + 537, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21697", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "voxels [20], hash tables [24], or triplanes [6], significantly enhancing rendering speed. However, real-time rendering remains challenging due to the inherent complexity of the ray marching strategy in volume rendering.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gaussian Splatting represents a recent advancement in novel view synthesis, enabling real-time high-quality rendering. It contributes to splatting-based rasterization by computing pixel colors through depth sorting and $\\alpha$ -blending of projected 2D Gaussians, thereby avoiding the complex sampling strategies of ray marching and achieving real-time performance. It is precisely due to its real-time high-quality rendering capabilities that 3DGS has been applied to various domains, including autonomous driving, content generation [30], and 4D dynamic scenes [18, 34, 38], among others. Despite these advancements, 3DGS still has some drawbacks, such as the storage of Gaussians and handling multi-resolution, and so on. Several works have enhanced 3DGS by improving Gaussian representation, including techniques such as low-pass filtering [40], multiscale Gaussian representations [36], and interpolating Gaussian attributes from structured grid features [22]. However, these works often overlook the importance of point management, specifically Adaptive Density Control, which is typically applied during optimization to address issues like under-population or over-population in the 3D space. Only a few works have focused on point management. For example, GaussianPro [7] directly tackles densification limitations, bridging gaps from SfM-based initialization. Pixel-GS [41] proposes a gradient scaling strategy to suppress artifacts near the camera. Additionally, [26] introduces an auxiliary per-pixel error function to implicitly supervise point contributions.", + "bbox": [ + 91, + 154, + 483, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although these methods improve densification, they are still unable to identify all 3D zones that require point densification and lack a proper mechanism to handle ill-conditioned points with negative impacts. Here, we propose a novel approach, Localized Point Management, capable of identifying error-contributing zones with greatest demand for both point addition and geometry calibration.", + "bbox": [ + 89, + 575, + 483, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 691, + 181, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries: 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 89, + 717, + 421, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gaussian Splatting builds upon concepts from EWA [42] splatting and proposes modeling a 3D scene as a collection of 3D Gaussian points $\\{G_i \\mid i = 1, \\dots, K\\}$ , rendered through volume splatting. Each 3D Gaussian $G$ is defined by the equation:", + "bbox": [ + 89, + 739, + 482, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (x) = e ^ {- \\frac {1}{2} (x - \\mu) ^ {T} \\Sigma^ {- 1} (x - \\mu)},\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 821, + 387, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu \\in \\mathbb{R}^{3\\times 1}$ represents the mean vector, and $\\Sigma \\in \\mathbb{R}^{3\\times 3}$ denotes its covariance matrix. To maintain the positive semi-definite nature of $\\Sigma$ during optimization, it is repre", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sented as $\\Sigma = RSS^{T}R^{T}$ , with the orthogonal rotation matrix $R\\in \\mathbb{R}^{3\\times 3}$ and the diagonal scale matrix $S\\in \\mathbb{R}^{3\\times 3}$ .", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To render an image from a specific viewpoint, the color of each pixel $p$ is determined by blending $N$ ordered Gaussians $\\{G_i \\mid i = 1, \\dots, N\\}$ that overlap $p$ , using the formula:", + "bbox": [ + 511, + 122, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nc (p) = \\sum_ {i = 1} ^ {N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}),\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 179, + 808, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ is derived by evaluating a projected 2D Gaussian from $G_{i}$ at pixel $p$ combined with a learned opacity for $G_{i}$ , and $c_{i}$ is the learnable, view-dependent color modeled using spherical harmonics in 3DGS. Gaussians that influence $p$ are arranged in ascending order based on their depth from the current viewpoint. Employing differentiable rendering techniques allows for the end-to-end optimization of all Gaussian attributes through training view reconstruction. Point management Since existing 3D Gaussian points using Structure from Motion (SfM), the points are often coarse and non-optimal in space. During optimization, a point management mechanism, Adaptive Density Control (ADC), is typically applied to manage point distribution issues. Specifically, thresholding the average gradient is used to decide on point densification. For each Gaussian point $G_{i}$ , 3DGS tracks the magnitude of the positional gradient $\\frac{\\partial L_{\\pi}}{\\partial \\mu_i}$ across all rendered views, which is then averaged to a quantity $T_{i}$ . During each training iteration, if the gradient $T_{i}$ surpasses a predefined threshold, it considers this point as inadequately representing the corresponding 3D region. With the scale of the Gaussian as the size measure, a large Gaussian will be split into two, while a small one leads to point cloning.", + "bbox": [ + 511, + 233, + 906, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, this commonly used ADC strategy is unable to identify all the 3D zones with the underlying need for point densification. This is because, often the local complexity of scene geometry varies significantly, which beyond the reach of any single-value based thresholding. Besides, there is lacking of a proper mechanism to handle ill-conditioned points with negative impacts (e.g., wrong opacity values estimated during training with points distributed here and there).", + "bbox": [ + 511, + 580, + 905, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Localized Gaussian Point Management", + "text_level": 1, + "bbox": [ + 511, + 727, + 849, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the aforementioned issues, we introduce a novel model agnostic point management approach, Localized Point Management (LPM), which leverages multiview geometry constraints to identify error contributing 3D points, with the guidance of image rendering errors. This approach can be seamlessly integrated with existing 3DGS models without the need for architectural modification. As illustrated in Figure 2, we begin with an image rendering error map for a specific view. Under the multiview geometry constraint, the corresponding regions in the referred view are", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21698", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/952171043dbcf77a86bdddb046a72ba2c1ac5a9ff212bc00e354f2b7c2e6ddc1.jpg", + "image_caption": [ + "Figure 2. Overview of our Localized Point Management (LPM). (a) We start with an image rendering error map versus the current view (the ground-truth). Concurrently, matching points are identified between the current view and a referred view sampled as an adjacent view via off-the-shelf feature mapping. (b) Subsequently, cross-view region mapping is then employed to locate the correspondence region in the refereed view. (c) For each pair of corresponded regions, we cast the rays through them at their respective camera views in the cone shape, and consider their intersection as the error source zone. The final step involves identifying under-optimized or ill-conditioned points within these zones, where under-optimized/empty places are densified, and ill-conditioned points are reset." + ], + "image_footnote": [], + "bbox": [ + 98, + 97, + 903, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "matched via feature mapping. For each pair of corresponding regions, we then cast rays through them from their respective camera views in a cone and identify their intersection as the error source zone. Within each zone, we perform localized point manipulation.", + "bbox": [ + 89, + 445, + 482, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Error map generation To accurately localize those zones in the 3D space that require point densification and geometry calibration, we initiate our process by rendering the current view image through the splatting of 3D Gaussians. This is followed by generating an error map (Figure 2(a)) for this specific view against the ground-truth image using an error function [18].", + "bbox": [ + 89, + 529, + 482, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Error contributing 3D zone identification To project this rendering error back to the 3D space, we leverage the region correspondence between different views under multiview geometry constraints. This involves the following two key steps.", + "bbox": [ + 89, + 643, + 482, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(i) Cross-view region mapping We select a neighboring view as the referred image. Following LightGlue [19], which predicts a partial assignment between two sets of local features extracted from images $A$ and $B$ from different views. Each feature consists of sets of 2D features position $\\{F_i\\mid (x_i,y_i)\\in [0,1]^2\\}$ , normalized by the image size. The images $A$ and $B$ contain $M$ and $N$ local features. LightGlue outputs a set of correspondences $\\mathcal{M} = \\{(i,j)\\} \\subseteq A\\times B$ , where $i$ and $j$ denote the indices of matched points from sets $A$ and $B$ , respectively. Since the 2D rendering error regions in the current view may not all appear in the referenced image, we select the paired regions $(R_e,R_e^{\\prime})$ (Fig", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ure 2(b)) based on the matching points, where $R_{e}$ represents the region in the current view and $R_{e}^{\\prime}$ corresponds to the region in the reference view. Additionally, this paired region undergoes multiview adaptive adjustments based on the error map throughout the optimization process.", + "bbox": [ + 511, + 445, + 906, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(ii) 2D-to-3D projection After obtaining the paired regions with render errors, we project each 2D error region to the 3D space via multiview geometry constraints. Specifically, we cast the rays $\\mathcal{C}$ in cone shape for region $R_{e}$ from the camera's center of projection $o$ along the direction $d$ , which aligns with the pixel's center (Figure 2(c)). The apex of this cone is located at $o$ , and its radius at the image plane. Hence, $o + d$ is parameterized as $\\mathcal{C}$ . The radius $r_{Cone}$ is set to match the radius of the smallest circumscribed circle of the 2D plane error region, creating a cone on the 3D space that can trace the Gaussian points contributing to the 2D error region. Concurrently, a corresponding cone, denoted as $\\mathcal{C}'$ , belong to region $R_{e}'$ is similarly projected. Subsequently, we compute the intersection points of these rays. In order to regionalize these points, we directly use a smallest sphere that can contain these points as error source 3D zone $R_{zone}$ .", + "bbox": [ + 511, + 522, + 908, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Points manipulation Recall that in existing 3DGS, points management only relies on the view-averaged gradient magnitude $\\tau$ to determine point densification globally. In addition to this, we further perform localized points addition and geometry calibration within the identified error source 3D zone $R_{zone}$ . For the point addition, we consider two common situations: (1) In the presence of points, we", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21699", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4af857b32b53adba7469c8f1470e353e1fb5c997334d26931316f39b6fc1785c.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 138, + 88, + 380, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d94b2985b4d391e94740e620d165fef7852b10109970ef4e06131bd27dedae9a.jpg", + "image_caption": [ + "(b)", + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 138, + 212, + 380, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/39a166ef3023b1779dd82830719b2090548fb966c74816b1e0291e8c3191c176.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 88, + 630, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8a2277b7bad2b600b665fa75d556bbb3b7814bd695285a8f152b07f84bb70f1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 212, + 630, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6f5182dc84b379d2b2d13d406611c6914b9f4d4bfe214fb532af9ae25b9c57b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 88, + 880, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/28751b1d099292f1be1a556e31ce5afd7035f91224b4a8b7319ff432a13be817.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 212, + 880, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/318b43f2bd59ad2957c4378cd652cd4b7212512cb55f16f686f5d55624b33af1.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 138, + 351, + 380, + 476 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e09305f5f120939ddfef0019aec4daff29795b1ef87b210dd7e8a9f4d318204f.jpg", + "image_caption": [ + "3DGS* rendered RGB image" + ], + "image_footnote": [], + "bbox": [ + 388, + 351, + 630, + 476 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/568fe3af2967250ac1bbbcd6c81346af4e4373be85873542b6ad3ea6f6177d88.jpg", + "image_caption": [ + "3DGS+LPM rendered RGB image" + ], + "image_footnote": [], + "bbox": [ + 638, + 351, + 883, + 476 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7048a4930fe1b81605dec1c41332856a55e74a6f94bddb1732a2ab3a41f59d89.jpg", + "image_caption": [ + "Ground Truth", + "(d)", + "2DGS rendered mesh", + "Figure 3. Qualitative evaluation of our LPM across diverse static datasets [4, 12]. Our LPM improves 2DGS [13] and 3DGS [14] on these challenging scenarios, e.g. (a) Light artifacts, (b) Completeness in the distance, (c) Depth structure and (d) Mesh details. See red patches for highlighted visual differences." + ], + "image_footnote": [], + "bbox": [ + 138, + 492, + 380, + 659 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/db3c37c9cabbaa38cc5520bb1cca71eea192e8a11b6de1aed0171859cac26f5d.jpg", + "image_caption": [ + "3DGS* rendered depth", + "2DGS+LPM rendered mesh" + ], + "image_footnote": [], + "bbox": [ + 388, + 492, + 630, + 659 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/00cac4e6b2ed17da7009a204540d579cd4991d7bd6448a598ddd4ca5151abfb7.jpg", + "image_caption": [ + "3DGS+LPM rendered depth", + "2DGS vs 2DGS+LPM heatmap" + ], + "image_footnote": [], + "bbox": [ + 638, + 492, + 880, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "apply point densification to locally complement the original counterparts. We set a lower threshold to select the points that need densification, aiming to enhance the geometric details. The densification rule is consistent with 3DGS, but it focuses on local 3D zones that need it most. Specifically, for small Gaussians, our strategy involves cloning the Gaussians while maintaining their size and repositioning them along the positional gradient to better capture emerging geometrical features. Conversely, larger Gaussians situated in areas of high variance are split into smaller points to more accurately represent the underlying geometry. (2) In cases", + "bbox": [ + 89, + 729, + 485, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of point sparsity, we add new Gaussian points at the center of the 3D zone. In the context of $\\alpha$ -blending in 3DGS, if the points at the forefront of the identified 3D zone $R_{zone}$ have the highest opacity, they may occlude valid points, leading to incorrect depth estimation, as shown in Figure 1. To deal with such issues, we treat these points as potentially ill-conditioned. We reset these points to provide an opportunity for correction, further calibrating the geometry. To minimize model expansion, we adaptively prune points based on their opacity values, starting from low to high. The number of points pruned is determined by the density of", + "bbox": [ + 511, + 729, + 908, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21700", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/216ba789a3565b5564f2aebe6447e2e64a055982ccaf88aad246368e9bccbd8b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMip-NeRF 360Tanks&TemplesDeep Blending
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
Plenoxels23.080.6250.46321.080.7190.37923.060.7950.510
INGP-Big25.590.6990.33121.920.7450.30524.960.8170.390
Mip-NeRF 36027.690.7920.23722.220.7590.25729.400.9010.245
3DGS27.210.8150.21423.140.8410.18329.410.9030.243
3DGS*27.470.8160.21623.670.8490.17729.550.9040.245
3DGS* + LPM27.590.8200.21623.830.8500.18129.760.9080.241
2DGS*27.150.8080.24623.580.8320.18529.350.8990.262
2DGS* + LPM27.420.8170.22823.650.8480.18029.520.9030.240
MipGS*27.510.8170.21023.690.8520.17329.580.9100.242
MipGS* + LPM27.700.8210.21023.820.8510.18029.610.9100.241
PixelGS*27.540.8190.20323.750.8500.17529.580.9200.220
PixelGS* + LPM27.800.8300.19024.020.8560.17329.650.9100.196
", + "bbox": [ + 181, + 88, + 823, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of various methods across different scenes on the Mip-NeRF 360 dataset, Tanks&Temples and Deep Blending. * indicates the retrained model from the official implementation. Bold represents best, underline indicates second best.", + "bbox": [ + 89, + 333, + 903, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "points in the zone. This strategic reduction ensures that our point management remains cost efficient and adaptive to the evolving needs of the scene representation.", + "bbox": [ + 89, + 380, + 482, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 431, + 215, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and metrics We conducted an extensive evaluation using both static and dynamic scenes derived from publicly datasets. For static scenes, our approach was applied to a total of 13 scenes as specified in the 3DGS framework [14], which includes nine scenes from Mip-NeRF360 [3], two from Tanks&Temples [15], and two from DeepBlending [12]. In the context of dynamic scenes, our approach was tested across six scenes from the Neural 3D Video Dataset [17].", + "bbox": [ + 89, + 454, + 482, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate novel view synthesis performance, we followed standard protocols by selecting one out of every eight images as test images, with the remaining used for training in static scenes. For each dynamic scene within the Neural 3D Video Dataset, one view was designated for testing while the others were allocated for training purposes. Evaluation metrics included the peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and the learned perceptual image patch similarity (LPIPS), which are broadly recognized standards in the field.", + "bbox": [ + 89, + 590, + 482, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and implementation Vanilla 3D Gaussian Splitting (3DGS) [14], 2D Gaussian Splitting (2DGS) [13], Mip Gaussian Splitting (MipGS) [40], PiexlGS [41] and SpacetimeGS (STGS) [18] were selected as our main baselines for their established art performance in novel view synthesis. For the static 3D benchmark, we also recorded the results of Mip-NeRF360 [3], iNGP [24] and Plenoxels [10] as in [14]. For the Dynamic 4D benchmark, we performed system comparison, such as DyNeRF [16], K-planes [11] and so on. In alignment with the approach", + "bbox": [ + 89, + 750, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "described in 3DGS an STGS, our models were trained for 30k iterations across all scenes, following the same training schedule and hyperparameters. In addition to the original Gaussian densification strategies used in 3DGS and SpaceTime Gaussian, we also performed localized points management, including addition, reset, and pruning. We maintained the same thresholds for splitting and cloning points as in the original 3DGS and SpaceTime Gaussian. For point matching, we performed offline extraction to save computational cost. All experiments were conducted on an RTX 3090 GPU with 24GB of memory.", + "bbox": [ + 511, + 380, + 906, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 511, + 559, + 653, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on static 3D datasets The quantitative results (PSNR, SSIM, and LPIPS) on the Mip-NeRF 360 and Tanks & Temples datasets are presented in Tables 1. We retrained the 3DGS model (referred to as 3DGS*) as it yields better performance compared to the vanilla 3DGS and its variants. Our approach achieves results comparable to the state-of-the-art on the Mip-NeRF360 dataset and further enhances all 3DGS based method using our point management technique. Additionally, LPM improve vanilla 3DGS and PixlGS to set new state-of-the-art results on the Mip-NeRF 360, Tanks & Temples datasets and Dep Blending, effectively capturing more challenging environments (e.g., light effects, transparency). These results quantitatively validate the effectiveness of our method in improving the quality of reconstruction.", + "bbox": [ + 511, + 582, + 906, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Figures 3, we present a comparison between 3DGS [14] and $3\\mathrm{DGS}^* +$ LPM, focusing on both appearance and depth. A variety of improvements can be observed, particularly in challenging cases such as light effects, completeness at a distance. Our LPM significantly reduces artifacts in specific regions on top of 3DGS, partic", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21701", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/906994016d490f4e7767cd3908d7a7e39402ad37d726ac55740116d093f88635.jpg", + "image_caption": [ + "3DGS" + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 254, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15864ed2fb62f3531cd8c07e815f89652ef6a5b1d192d42133f9d126136fd718.jpg", + "image_caption": [ + "3DGS (lower threshold)" + ], + "image_footnote": [], + "bbox": [ + 256, + 90, + 415, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6802d13c8f1d0b05111bd681b98c0a797428013f63f9cee0b59ee15ad59349e.jpg", + "image_caption": [ + "LPM wo/ point addition" + ], + "image_footnote": [], + "bbox": [ + 416, + 90, + 576, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a39fbabb937bf7897895c2b4a2554039b50e6886337b0c3e4e9cc82c9571bc75.jpg", + "image_caption": [ + "LPM wo/ reset" + ], + "image_footnote": [], + "bbox": [ + 576, + 90, + 736, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/213aae9e7575bcf3ab79ed9b2fe4e5ae3f4381b0ba2ce104c93d315888ab62df.jpg", + "image_caption": [ + "3DGS+LPM" + ], + "image_footnote": [], + "bbox": [ + 738, + 90, + 900, + 231 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1c32766e77cc422b33f2be4cd06e22beb5d0c576bdf69db088ad75ff939c836a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPSNRDSSIMLPIPSFPS
\\( LLFF^1 \\)23.240.0760.235-
\\( DyNeRF^1 \\)29.580.0830.0630.015
\\( Dynamic-4DGS^1 \\)---30
\\( 4DGS^1 \\)29.38--114
\\( STGS^1 \\)29.580.0220.063103
\\( STGS*^1 \\)29.480.0230.066110
\\( STGS*^1 + LPM \\)29.840.0220.062105
StreamRF28.26-0.03910.9
NeRFPlayer30.69-0.1110.05
HyperReal31.10-0.0962
K-planes31.63-0.313
Mix Voxels-X31.73-0.0644.6
Dynamic-4DGS31.150.0160.04930
4DGS32.01-0.055114
STGS32.050.0140.044140
STGS*31.990.0150.045145
\\( STGS*+ LPM \\)32.400.0140.045140
", + "bbox": [ + 104, + 296, + 472, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparisons on the Neural 3D Video dataset. \"FPS\" is measured at a resolution of $1352 \\times 1014$ . Some methods only report results for a subset of scenes. For a fair comparison, we report LPM's results under two pre-existing settings. Only includes the Flame Salmon scene. Bold represents best, underline indicates second best.", + "bbox": [ + 89, + 617, + 483, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cularly in the tree at the second. These regions require more points for accurate population, leading to a more precise and detailed reconstruction. Additionally, the tablecloth in the first row is affected by ill-conditioned points. Furthermore, we provide depth and mesh comparisons in the third and final rows. All these observations demonstrate that our geometry calibration with LPM offers an opportunity to correct these potentially ill-conditioned points, thereby enhancing the overall reconstruction accuracy.", + "bbox": [ + 89, + 727, + 482, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on dynamic 4D datasets Table 2 presents a quantitative evaluation on the Neural 3D Video Dataset.", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6578bb930cdfefdec14552fc7a5404754bf324841aaa554d50a034e7fec3d307.jpg", + "image_caption": [ + "Figure 4. Effect of key operations of LPM. We show that the point addition operation effectively captures the geometric details in the scene; The point reset operation based on the error map further calibrate the geometry.", + "Figure 5. Qualitative evaluation on dynamic Neural 3D Video dataset [17]. LPM improves STGS [18] for both scenes Transparent (e.g., window) and Dynamic movements (e.g., dog's tongue)." + ], + "image_footnote": [], + "bbox": [ + 516, + 297, + 903, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following established practices, training and evaluation are conducted at half resolution, with the first camera held out for evaluation [16]. Integrating our LPM into SpaceTimeGS yields the best performance across all comparisons. Notably, our method demonstrates significant improvements in the challenging Flame Salmon scene compared to SpaceTimeGS [18]. Our approach not only surpasses previous methods in rendering quality but also maintains comparable rendering speed.", + "bbox": [ + 511, + 516, + 905, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to the quantitative assessment, we provide qualitative comparisons on the Flame Salmon and Flame Steak scenes, as illustrated in Figure 5. The quality of synthesis in both static and dynamic regions markedly outperforms STGS. Several intricate details, including the tree behind the window and the fine features like the dog's tongue, are faithfully reproduced with higher accuracy compared to STGS [18]. Both examples indicate that LPM improves upon STGS for superior scene modeling.", + "bbox": [ + 511, + 654, + 906, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 797, + 663, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted ablation studies on the more challenging scene: PlayRoom from Deep Blending [12] and Truck from Tanks&Temples [15].", + "bbox": [ + 511, + 816, + 905, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness and cost of LPM We hypothesize that the Adaptive Density Control (ADC) tends to overlook under", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21702", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e0860f673283b5975e1f7e1b81999c598ec32b2e99413a9cee2c521d8d283bfc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPlayRoomTruck
PSNRLPIPSSSIMPSNRLPIPSSSIM
Full LPM30.220.2410.91025.610.1540.883
wo/ point addition30.100.2410.91025.430.1530.883
wo/ reset30.070.2430.90825.520.1440.883
", + "bbox": [ + 253, + 88, + 738, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c26324be8a9df9d31128e6b9a7c70e83699c56daca28b1d2eb8fa29b4fdca291.jpg", + "table_caption": [ + "Table 3. Performance comparison for different configurations" + ], + "table_footnote": [], + "table_body": "
SceneMethodPSNRLPIPSGaussiansTraining time
PlayRoom3DGS*30.030.244232k22min
3DG* (lower threshold)29.690.240523k36min
GaussianProOut of Memory
PielGS30.090.241186k35min
3DGS + LPM30.220.241186k23min
Truck3DGS*25.420.146257k19 min
3DGS* (lower threshold)25.450.127635k35min
GaussianPro25.400.164312k36min
PielGS25.510.121518k37min
3DGS + LPM25.610.154265k21min
", + "bbox": [ + 173, + 214, + 820, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Cost-effectiveness analysis. Rendering speed of both methods are measured on our machine. Note: For 3DGS+LPM, training time includes the feature matching process.", + "bbox": [ + 89, + 407, + 906, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "optimized points due to its simplistic approach of thresholding the average gradient. The straight way to identify the all points is lowering threshold to densification process. Although this solution can reduce blurring in specific regions, such as the toy (red box) illustrated in Figure 4, it still has limitations. As shown in Table 4, lowering the threshold for 3DGS significantly increases the number of Gaussian points and decreases rendering speed. Additionally, the PSNR of the quantitative results decreases due to the introduction of unnecessary points in already dense areas. In contrast, LPM effectively generates points in areas indicated by the error map, leading to more accurate and detailed reconstructions while maintaining real-time rendering speed. As demonstrated by the qualitative comparison in Figure 4, 3DGS with LPM achieves superior qualitative results. We further compare our method with other recent methods that also focus on adaptive density control (ADC). While PixelGS and GaussianPro achieve improvements in rendering quality, their training times increase substantially as they only consider point addition and extra gradient propagation. In contrast, LPM achieves a noticeable improvement with only a slight increase in training time due to (1) point matching [19] is much faster (2) considering model expansion to dynamically prune the points by their addition number and (3) selecting points in error-contributing zones 3D zone using the parallel matrix operations.", + "bbox": [ + 88, + 460, + 482, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Individual points manipulation We study the effect of individual points manipulation of LPM, including the point addition and reset ill-conditional points. The results in Ta", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ble 3 show that, (1) each manipulation is useful with positive gain, suggesting that the LPM is meaningful. (2) The point addition operation densify the under-optimized points which may be overlook in the 3DGS, further captures the geometry details (e.g., detail of toy and leaf of the tree, see Fig. 4). (3) Reset points in ceratin zone provide the opportunity of correct the ill-conditioned points to achieve geometry calibration, (e.g., window of the trunk, see Fig. 4).", + "bbox": [ + 511, + 459, + 906, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 594, + 633, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose Localized Point Management (LPM), a novel point management approach to address the limitations of the Adaptive Density Control (ADC) mechanism in 3D Gaussian Splatting (3DGS). The core idea of LPM is identifying the error-contributing 3D zones that require both point addition and geometry calibration under multiview geometry constraints, guided by image rendering errors. We implement appropriate operations for point densification and opacity reset. As a versatile plugin, LPM can be seamlessly integrated into existing 3DGS-based rendering methods. Extensive experiments across both static 3D and dynamic 4D scenes validate the efficacy of LPM in enhancing existing ADC mechanisms both quantitatively and qualitatively. While our method identifies the 3D Gaussian points that lead to rendering errors, it still follows the densification rules of 3DGS [14]. This approach may not be optimal for under-optimized points, and we leave this aspect for further investigation.", + "bbox": [ + 511, + 618, + 906, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21703", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2", + "[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2, 6, 1", + "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 5", + "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697–19705, 2023. 2", + "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 3", + "[7] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2, 3", + "[8] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2", + "[9] Nianchen Deng, Zhenyi He, Jiannan Ye, Budmonde Duinkharjav, Praneeth Chakravarthula, Xubo Yang, and Qi Sun. Fov-nerf: Foveated neural radiance fields for virtual reality. IEEE Transactions on Visualization and Computer Graphics, 28(11):3854-3864, 2022. 2", + "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 6", + "[11] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12479–12488, 2023. 6", + "[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 5, 6, 7, 1", + "[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 5, 6", + "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 2, 5, 6, 8, 1", + "[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 6, 7, 1", + "[16] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multiview video. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5511-5521, 2022. 6, 7", + "[17] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 6, 7", + "[18] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. arXiv preprint arXiv:2312.16812, 2023. 3, 4, 6, 7, 1", + "[19] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627-17638, 2023. 4, 8", + "[20] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 3", + "[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2", + "[22] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109, 2023. 3", + "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 3, 6" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21704", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2", + "[26] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. arXiv e-prints, pages arXiv-2404, 2024. 3", + "[27] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 2", + "[28] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2", + "[29] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2", + "[30] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 3", + "[31] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 2", + "[32] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 2", + "[33] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4150–4159, 2023. 2", + "[34] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3", + "[35] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5438-5448, 2022. 2", + "[36] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. arXiv preprint arXiv:2311.17089, 2023. 3", + "[37] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 2" + ], + "bbox": [ + 89, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In ICLR, 2024. 3", + "[39] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 2", + "[40] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. arXiv preprint arXiv:2311.16493, 2023. 3, 6", + "[41] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 2, 3, 6", + "[42] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 3" + ], + "bbox": [ + 516, + 92, + 903, + 359 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21705", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Improving Gaussian Splatting with Localized Points Management/6436626a-e39e-4d87-9a2d-0ebf83ce94c9_model.json b/2025/Improving Gaussian Splatting with Localized Points Management/6436626a-e39e-4d87-9a2d-0ebf83ce94c9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7dacef7632a124059b9cce2af3fb9b5b3849cf6d --- /dev/null +++ b/2025/Improving Gaussian Splatting with Localized Points Management/6436626a-e39e-4d87-9a2d-0ebf83ce94c9_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.833, + 0.154 + ], + "angle": 0, + "content": "Improving Gaussian Splating with Localized Points Management" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.179, + 0.865, + 0.219 + ], + "angle": 0, + "content": "Haosen Yang\\(^{1*}\\) Chenhao Zhang\\(^{1*}\\) Wenqing Wang\\(^{1*}\\) Marco Volino\\(^{1}\\) Adrian Hilton\\(^{1}\\) Li Zhang\\(^{2}\\) Xiatian Zhu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.233, + 0.668, + 0.271 + ], + "angle": 0, + "content": "\\(^{1}\\)University of Surrey \\(^{2}\\)Fudan University *Equal contribution" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.287, + 0.663, + 0.306 + ], + "angle": 0, + "content": "https://happy-hsy.github.io/projects/LPM/" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.34, + 0.382, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.452, + 0.382, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.565, + 0.297, + 0.578 + ], + "angle": 0, + "content": "Rendered Image" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.341, + 0.633, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.452, + 0.633, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.566, + 0.551, + 0.578 + ], + "angle": 0, + "content": "Gaussian Points" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.34, + 0.888, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.452, + 0.889, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.567, + 0.812, + 0.58 + ], + "angle": 0, + "content": "Rendered Depth" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.592, + 0.908, + 0.634 + ], + "angle": 0, + "content": "Figure 1. Visualization of points behavior. 3DGS produces ill-conditioned Gaussians (red box) that occlude other valid points, resulting in noticeably incorrect depth estimation. LPM handles these ill-conditioned points to reduce negative impacts and further calibrate the geometry." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.658, + 0.327, + 0.673 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.69, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Point management is critical for optimizing 3D Gaussian Splatting models, as point initiation (e.g., via structure from motion) is often distributionally inappropriate. Typically, Adaptive Density Control (ADC) algorithm is adopted, leveraging view-averaged gradient magnitude thresholding for point densification, opacity thresholding for pruning, and regular all-points opacity reset. We reveal that this strategy is limited in tackling intricate/special image regions (e.g., transparent) due to inability of identifying all 3D zones requiring point densification, and lacking an appropriate mechanism to handle ill-conditioned points with negative impacts (e.g., occlusion due to false high opacity). To address these limitations, we propose a Localized Point Management (LPM) strategy, capable of identify" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "ing those error-contributing zones in greatest need for both point addition and geometry calibration. Zone identification is achieved by leveraging the underlying multiview geometry constraints, subject to image rendering errors. We apply point densification in the identified zones and then reset the opacity of the points in front of these regions, creating a new opportunity to correct poorly conditioned points. Serving as a versatile plugin, LPM can be seamlessly integrated into existing static 3D and dynamic 4D Gaussian Splatting models with minimal additional cost. Experimental evaluations validate the efficacy of our LPM in boosting a variety of existing 3D/4D models both quantitatively and qualitatively. Notably, LPM improves both static 3DGS and dynamic SpaceTimeGS to achieve state-of-the-art rendering quality while retaining real-time speeds, excelling on challenging datasets such as Tanks & Temples and the Neural" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21696" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.092, + 0.212, + 0.106 + ], + "angle": 0, + "content": "3D Video dataset." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.121, + 0.223, + 0.137 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.146, + 0.482, + 0.356 + ], + "angle": 0, + "content": "Neural rendering has emerged as a generalizable, flexible, and powerful approach for photorealistic novel view synthesis (NVS) of any camera poses [23], underpinning a wide variety of applications in augmented/virtual/mixed reality [9], robotics [37], and generation [25], among more others. For example, taking a learning-based parametric idea, Neural Radiance Fields (NeRFs) [23] implicitly represent the scene radiance of any complexity using neural networks (e.g., MLPs), without the tedious requirements of model handcrafting for accounting the scene variations in geometry, texture, illumination. However, their view rendering is inefficient computationally due to heavy ray sampling, thus suffer in scaling to high-resolution content applications and large scale scene modeling [29, 32]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Recently, 3D Gaussian Splatting (3DGS) [14] has come as an alternative with explicit representation, much faster model optimization and real-time neural rendering. The process begins by initializing a set of 3D Gaussian points using Structure from Motion (SfM) [28]. This is followed by optimizing the parameters of these points through view reconstruction loss, resulting in a view output generated with differentiable splatting-based rasterization. However, the point initialization is often distributionally non-optimal, leading to issues such as under-population (e.g., insufficient points) or over-population (e.g., excessive points) in the 3D space. Consequently, a point management mechanism, such as Adaptive Density Control (ADC), is necessary during optimization. However, we identify several limitations with ADC: (i) Thresholding the average gradient to determine regions for point densification often overlooks under-optimized points. For instance, larger Gaussian points typically have lower average gradients and may frequently appear across various views in screen space. (ii) Point sparsity complicates the addition of sufficient and reliable points needed to comprehensively cover the scene. (iii) Mis-optimized Gaussian points can have detrimental effects, such as occluding other valuable points and leading to incorrect depth estimates (see erroneous placements on windows in Fig. 1). While GaussianPro [7] and PixelGS [41] try to solve (1) and (2) through multi-view stereo [2] and additional gradient propagation, respectively, these methods significantly increase the training budget, as shown in the Table 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.482, + 0.901 + ], + "angle": 0, + "content": "To overcome the aforementioned limitations, in this paper we propose a novel and efficient Localized Point Management (LPM) approach. Our idea is intuitive - identifying those 3D Gaussian points leading to rendering errors. Thus we start with an image rendering error map of a specific view. To obtain the error contributing 3D points, we leverage the region correspondence between different" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.332 + ], + "angle": 0, + "content": "views via feature mapping, subject to the multiview geometry constraint. For each pair of corresponded regions, we cast the rays through them at their respective camera views in the cone shape, and consider their intersection as the error source zone. Within each such zone, we consider two situations: (1) At presence of points, we further apply point densification at a lower threshold to complement the original counterpart locally; (2) In case no point due to point sparsity, we add new Gaussian points. Concurrently, we reset the opacity of points with high opacity estimates that are located in front of these zones, as they can significantly affect view rendering. This provides an opportunity to correct potentially ill-conditioned points while tuning the newly added ones in the subsequent optimization. To minimize model expansion, we prune the points by opacity in a density-aware manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.905, + 0.5 + ], + "angle": 0, + "content": "We summarize the contributions below: (1) Through in-depth analysis, we have identified several limitations in the standard point management mechanism used in Gaussian Splatting that impede model optimization. (2) We present Localized Point Management (LPM) for these issues by identifying error-contributing 3D zones and implementing appropriate operations for point densification and opacity reset. (3) Extensive experiments validate the benefits of our LPM in improving a diversity of existing 3D and 4D Gaussian Splatting models in novel view synthesis on both static and dynamic scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.655, + 0.531 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Neural Scene Representations has always been an important direction in novel view synthesis. Previous methods allocate neural features to structures such as volume [21, 27], texture [31], and point cloud [1]. The pioneering work of NeRF [23] proposes integrating neural networks with 3D volumetric representations to convert a 3D scene into a learnable density field, enabling high-quality novel view synthesis without requiring explicit modeling of the 3D scene and illumination. Later on, numerous works emerge to boost the quality and efficiency of volume rendering, [3, 5, 35] refine the point sampling strategy in ray marching, some some advanced works [4, 33] reparameterize the scene to produce a more compact representation. Additionally, regularization terms [8, 39] can be incorporated to constrain the scene representation, resulting in a closer approximation to real-world geometry. Despite their high-quality representational performance, these methods are typically computationally inefficient for view rendering due to the extensive ray sampling required and the use of Multi-Layer Perceptrons (MLPs) to represent the scene, complicating the computation and optimization of any point within the scene. To address this, several works have proposed novel scene representations aimed at accelerating the rendering process. These representations replace MLPs with sparse" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21697" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "voxels [20], hash tables [24], or triplanes [6], significantly enhancing rendering speed. However, real-time rendering remains challenging due to the inherent complexity of the ray marching strategy in volume rendering." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.155, + 0.484, + 0.575 + ], + "angle": 0, + "content": "Gaussian Splatting represents a recent advancement in novel view synthesis, enabling real-time high-quality rendering. It contributes to splatting-based rasterization by computing pixel colors through depth sorting and \\(\\alpha\\)-blending of projected 2D Gaussians, thereby avoiding the complex sampling strategies of ray marching and achieving real-time performance. It is precisely due to its real-time high-quality rendering capabilities that 3DGS has been applied to various domains, including autonomous driving, content generation [30], and 4D dynamic scenes [18, 34, 38], among others. Despite these advancements, 3DGS still has some drawbacks, such as the storage of Gaussians and handling multi-resolution, and so on. Several works have enhanced 3DGS by improving Gaussian representation, including techniques such as low-pass filtering [40], multiscale Gaussian representations [36], and interpolating Gaussian attributes from structured grid features [22]. However, these works often overlook the importance of point management, specifically Adaptive Density Control, which is typically applied during optimization to address issues like under-population or over-population in the 3D space. Only a few works have focused on point management. For example, GaussianPro [7] directly tackles densification limitations, bridging gaps from SfM-based initialization. Pixel-GS [41] proposes a gradient scaling strategy to suppress artifacts near the camera. Additionally, [26] introduces an auxiliary per-pixel error function to implicitly supervise point contributions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.684 + ], + "angle": 0, + "content": "Although these methods improve densification, they are still unable to identify all 3D zones that require point densification and lack a proper mechanism to handle ill-conditioned points with negative impacts. Here, we propose a novel approach, Localized Point Management, capable of identifying error-contributing zones with greatest demand for both point addition and geometry calibration." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.692, + 0.182, + 0.707 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.718, + 0.422, + 0.735 + ], + "angle": 0, + "content": "3.1. Preliminaries: 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.483, + 0.817 + ], + "angle": 0, + "content": "Gaussian Splatting builds upon concepts from EWA [42] splatting and proposes modeling a 3D scene as a collection of 3D Gaussian points \\(\\{G_i \\mid i = 1, \\dots, K\\}\\), rendered through volume splatting. Each 3D Gaussian \\(G\\) is defined by the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.822, + 0.388, + 0.843 + ], + "angle": 0, + "content": "\\[\nG (x) = e ^ {- \\frac {1}{2} (x - \\mu) ^ {T} \\Sigma^ {- 1} (x - \\mu)},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mu \\in \\mathbb{R}^{3\\times 1}\\) represents the mean vector, and \\(\\Sigma \\in \\mathbb{R}^{3\\times 3}\\) denotes its covariance matrix. To maintain the positive semi-definite nature of \\(\\Sigma\\) during optimization, it is repre" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.905, + 0.122 + ], + "angle": 0, + "content": "sented as \\(\\Sigma = RSS^{T}R^{T}\\), with the orthogonal rotation matrix \\(R\\in \\mathbb{R}^{3\\times 3}\\) and the diagonal scale matrix \\(S\\in \\mathbb{R}^{3\\times 3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.123, + 0.905, + 0.168 + ], + "angle": 0, + "content": "To render an image from a specific viewpoint, the color of each pixel \\( p \\) is determined by blending \\( N \\) ordered Gaussians \\( \\{G_i \\mid i = 1, \\dots, N\\} \\) that overlap \\( p \\), using the formula:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.18, + 0.81, + 0.222 + ], + "angle": 0, + "content": "\\[\nc (p) = \\sum_ {i = 1} ^ {N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.234, + 0.907, + 0.581 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) is derived by evaluating a projected 2D Gaussian from \\(G_{i}\\) at pixel \\(p\\) combined with a learned opacity for \\(G_{i}\\), and \\(c_{i}\\) is the learnable, view-dependent color modeled using spherical harmonics in 3DGS. Gaussians that influence \\(p\\) are arranged in ascending order based on their depth from the current viewpoint. Employing differentiable rendering techniques allows for the end-to-end optimization of all Gaussian attributes through training view reconstruction. Point management Since existing 3D Gaussian points using Structure from Motion (SfM), the points are often coarse and non-optimal in space. During optimization, a point management mechanism, Adaptive Density Control (ADC), is typically applied to manage point distribution issues. Specifically, thresholding the average gradient is used to decide on point densification. For each Gaussian point \\(G_{i}\\), 3DGS tracks the magnitude of the positional gradient \\(\\frac{\\partial L_{\\pi}}{\\partial \\mu_i}\\) across all rendered views, which is then averaged to a quantity \\(T_{i}\\). During each training iteration, if the gradient \\(T_{i}\\) surpasses a predefined threshold, it considers this point as inadequately representing the corresponding 3D region. With the scale of the Gaussian as the size measure, a large Gaussian will be split into two, while a small one leads to point cloning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.717 + ], + "angle": 0, + "content": "However, this commonly used ADC strategy is unable to identify all the 3D zones with the underlying need for point densification. This is because, often the local complexity of scene geometry varies significantly, which beyond the reach of any single-value based thresholding. Besides, there is lacking of a proper mechanism to handle ill-conditioned points with negative impacts (e.g., wrong opacity values estimated during training with points distributed here and there)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.85, + 0.744 + ], + "angle": 0, + "content": "3.2. Localized Gaussian Point Management" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To address the aforementioned issues, we introduce a novel model agnostic point management approach, Localized Point Management (LPM), which leverages multiview geometry constraints to identify error contributing 3D points, with the guidance of image rendering errors. This approach can be seamlessly integrated with existing 3DGS models without the need for architectural modification. As illustrated in Figure 2, we begin with an image rendering error map for a specific view. Under the multiview geometry constraint, the corresponding regions in the referred view are" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21698" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.098, + 0.904, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.351, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 2. Overview of our Localized Point Management (LPM). (a) We start with an image rendering error map versus the current view (the ground-truth). Concurrently, matching points are identified between the current view and a referred view sampled as an adjacent view via off-the-shelf feature mapping. (b) Subsequently, cross-view region mapping is then employed to locate the correspondence region in the refereed view. (c) For each pair of corresponded regions, we cast the rays through them at their respective camera views in the cone shape, and consider their intersection as the error source zone. The final step involves identifying under-optimized or ill-conditioned points within these zones, where under-optimized/empty places are densified, and ill-conditioned points are reset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.483, + 0.522 + ], + "angle": 0, + "content": "matched via feature mapping. For each pair of corresponding regions, we then cast rays through them from their respective camera views in a cone and identify their intersection as the error source zone. Within each zone, we perform localized point manipulation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.483, + 0.636 + ], + "angle": 0, + "content": "Error map generation To accurately localize those zones in the 3D space that require point densification and geometry calibration, we initiate our process by rendering the current view image through the splatting of 3D Gaussians. This is followed by generating an error map (Figure 2(a)) for this specific view against the ground-truth image using an error function [18]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Error contributing 3D zone identification To project this rendering error back to the 3D space, we leverage the region correspondence between different views under multiview geometry constraints. This involves the following two key steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "(i) Cross-view region mapping We select a neighboring view as the referred image. Following LightGlue [19], which predicts a partial assignment between two sets of local features extracted from images \\(A\\) and \\(B\\) from different views. Each feature consists of sets of 2D features position \\(\\{F_i\\mid (x_i,y_i)\\in [0,1]^2\\}\\), normalized by the image size. The images \\(A\\) and \\(B\\) contain \\(M\\) and \\(N\\) local features. LightGlue outputs a set of correspondences \\(\\mathcal{M} = \\{(i,j)\\} \\subseteq A\\times B\\), where \\(i\\) and \\(j\\) denote the indices of matched points from sets \\(A\\) and \\(B\\), respectively. Since the 2D rendering error regions in the current view may not all appear in the referenced image, we select the paired regions \\((R_e,R_e^{\\prime})\\) (Fig" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.907, + 0.522 + ], + "angle": 0, + "content": "ure 2(b)) based on the matching points, where \\( R_{e} \\) represents the region in the current view and \\( R_{e}^{\\prime} \\) corresponds to the region in the reference view. Additionally, this paired region undergoes multiview adaptive adjustments based on the error map throughout the optimization process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.909, + 0.78 + ], + "angle": 0, + "content": "(ii) 2D-to-3D projection After obtaining the paired regions with render errors, we project each 2D error region to the 3D space via multiview geometry constraints. Specifically, we cast the rays \\(\\mathcal{C}\\) in cone shape for region \\(R_{e}\\) from the camera's center of projection \\(o\\) along the direction \\(d\\), which aligns with the pixel's center (Figure 2(c)). The apex of this cone is located at \\(o\\), and its radius at the image plane. Hence, \\(o + d\\) is parameterized as \\(\\mathcal{C}\\). The radius \\(r_{Cone}\\) is set to match the radius of the smallest circumscribed circle of the 2D plane error region, creating a cone on the 3D space that can trace the Gaussian points contributing to the 2D error region. Concurrently, a corresponding cone, denoted as \\(\\mathcal{C}'\\), belong to region \\(R_{e}'\\) is similarly projected. Subsequently, we compute the intersection points of these rays. In order to regionalize these points, we directly use a smallest sphere that can contain these points as error source 3D zone \\(R_{zone}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Points manipulation Recall that in existing 3DGS, points management only relies on the view-averaged gradient magnitude \\(\\tau\\) to determine point densification globally. In addition to this, we further perform localized points addition and geometry calibration within the identified error source 3D zone \\(R_{zone}\\). For the point addition, we consider two common situations: (1) In the presence of points, we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21699" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.145, + 0.132, + 0.162 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.089, + 0.382, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.267, + 0.133, + 0.281 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.213, + 0.382, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.339, + 0.305, + 0.35 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.089, + 0.632, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.213, + 0.631, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.339, + 0.597, + 0.351 + ], + "angle": 0, + "content": "3DGS* rendered RGB image" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.089, + 0.882, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.213, + 0.882, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.339, + 0.871, + 0.351 + ], + "angle": 0, + "content": "3DGS+LPM rendered RGB image" + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.407, + 0.132, + 0.422 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.352, + 0.382, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.48, + 0.305, + 0.492 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.352, + 0.632, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.48, + 0.573, + 0.492 + ], + "angle": 0, + "content": "3DGS* rendered depth" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.352, + 0.884, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.48, + 0.842, + 0.492 + ], + "angle": 0, + "content": "3DGS+LPM rendered depth" + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.577, + 0.133, + 0.591 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.493, + 0.382, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.661, + 0.323, + 0.672 + ], + "angle": 0, + "content": "2DGS rendered mesh" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.493, + 0.631, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.661, + 0.586, + 0.672 + ], + "angle": 0, + "content": "2DGS+LPM rendered mesh" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.493, + 0.882, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.661, + 0.857, + 0.674 + ], + "angle": 0, + "content": "2DGS vs 2DGS+LPM heatmap" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.677, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Figure 3. Qualitative evaluation of our LPM across diverse static datasets [4, 12]. Our LPM improves 2DGS [13] and 3DGS [14] on these challenging scenarios, e.g. (a) Light artifacts, (b) Completeness in the distance, (c) Depth structure and (d) Mesh details. See red patches for highlighted visual differences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.486, + 0.897 + ], + "angle": 0, + "content": "apply point densification to locally complement the original counterparts. We set a lower threshold to select the points that need densification, aiming to enhance the geometric details. The densification rule is consistent with 3DGS, but it focuses on local 3D zones that need it most. Specifically, for small Gaussians, our strategy involves cloning the Gaussians while maintaining their size and repositioning them along the positional gradient to better capture emerging geometrical features. Conversely, larger Gaussians situated in areas of high variance are split into smaller points to more accurately represent the underlying geometry. (2) In cases" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.909, + 0.898 + ], + "angle": 0, + "content": "of point sparsity, we add new Gaussian points at the center of the 3D zone. In the context of \\(\\alpha\\)-blending in 3DGS, if the points at the forefront of the identified 3D zone \\(R_{zone}\\) have the highest opacity, they may occlude valid points, leading to incorrect depth estimation, as shown in Figure 1. To deal with such issues, we treat these points as potentially ill-conditioned. We reset these points to provide an opportunity for correction, further calibrating the geometry. To minimize model expansion, we adaptively prune points based on their opacity values, starting from low to high. The number of points pruned is determined by the density of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21700" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.182, + 0.089, + 0.825, + 0.324 + ], + "angle": 0, + "content": "
MethodMip-NeRF 360Tanks&TemplesDeep Blending
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
Plenoxels23.080.6250.46321.080.7190.37923.060.7950.510
INGP-Big25.590.6990.33121.920.7450.30524.960.8170.390
Mip-NeRF 36027.690.7920.23722.220.7590.25729.400.9010.245
3DGS27.210.8150.21423.140.8410.18329.410.9030.243
3DGS*27.470.8160.21623.670.8490.17729.550.9040.245
3DGS* + LPM27.590.8200.21623.830.8500.18129.760.9080.241
2DGS*27.150.8080.24623.580.8320.18529.350.8990.262
2DGS* + LPM27.420.8170.22823.650.8480.18029.520.9030.240
MipGS*27.510.8170.21023.690.8520.17329.580.9100.242
MipGS* + LPM27.700.8210.21023.820.8510.18029.610.9100.241
PixelGS*27.540.8190.20323.750.8500.17529.580.9200.220
PixelGS* + LPM27.800.8300.19024.020.8560.17329.650.9100.196
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.334, + 0.905, + 0.364 + ], + "angle": 0, + "content": "Table 1. Comparison of various methods across different scenes on the Mip-NeRF 360 dataset, Tanks&Temples and Deep Blending. * indicates the retrained model from the official implementation. Bold represents best, underline indicates second best." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.427 + ], + "angle": 0, + "content": "points in the zone. This strategic reduction ensures that our point management remains cost efficient and adaptive to the evolving needs of the scene representation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.432, + 0.216, + 0.449 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.59 + ], + "angle": 0, + "content": "Datasets and metrics We conducted an extensive evaluation using both static and dynamic scenes derived from publicly datasets. For static scenes, our approach was applied to a total of 13 scenes as specified in the 3DGS framework [14], which includes nine scenes from Mip-NeRF360 [3], two from Tanks&Temples [15], and two from DeepBlending [12]. In the context of dynamic scenes, our approach was tested across six scenes from the Neural 3D Video Dataset [17]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.483, + 0.743 + ], + "angle": 0, + "content": "To evaluate novel view synthesis performance, we followed standard protocols by selecting one out of every eight images as test images, with the remaining used for training in static scenes. For each dynamic scene within the Neural 3D Video Dataset, one view was designated for testing while the others were allocated for training purposes. Evaluation metrics included the peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and the learned perceptual image patch similarity (LPIPS), which are broadly recognized standards in the field." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Baselines and implementation Vanilla 3D Gaussian Splitting (3DGS) [14], 2D Gaussian Splitting (2DGS) [13], Mip Gaussian Splitting (MipGS) [40], PiexlGS [41] and SpacetimeGS (STGS) [18] were selected as our main baselines for their established art performance in novel view synthesis. For the static 3D benchmark, we also recorded the results of Mip-NeRF360 [3], iNGP [24] and Plenoxels [10] as in [14]. For the Dynamic 4D benchmark, we performed system comparison, such as DyNeRF [16], K-planes [11] and so on. In alignment with the approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.548 + ], + "angle": 0, + "content": "described in 3DGS an STGS, our models were trained for 30k iterations across all scenes, following the same training schedule and hyperparameters. In addition to the original Gaussian densification strategies used in 3DGS and SpaceTime Gaussian, we also performed localized points management, including addition, reset, and pruning. We maintained the same thresholds for splitting and cloning points as in the original 3DGS and SpaceTime Gaussian. For point matching, we performed offline extraction to save computational cost. All experiments were conducted on an RTX 3090 GPU with 24GB of memory." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.655, + 0.575 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Results on static 3D datasets The quantitative results (PSNR, SSIM, and LPIPS) on the Mip-NeRF 360 and Tanks & Temples datasets are presented in Tables 1. We retrained the 3DGS model (referred to as 3DGS*) as it yields better performance compared to the vanilla 3DGS and its variants. Our approach achieves results comparable to the state-of-the-art on the Mip-NeRF360 dataset and further enhances all 3DGS based method using our point management technique. Additionally, LPM improve vanilla 3DGS and PixlGS to set new state-of-the-art results on the Mip-NeRF 360, Tanks & Temples datasets and Dep Blending, effectively capturing more challenging environments (e.g., light effects, transparency). These results quantitatively validate the effectiveness of our method in improving the quality of reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In Figures 3, we present a comparison between 3DGS [14] and \\(3\\mathrm{DGS}^* +\\) LPM, focusing on both appearance and depth. A variety of improvements can be observed, particularly in challenging cases such as light effects, completeness at a distance. Our LPM significantly reduces artifacts in specific regions on top of 3DGS, partic" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "21701" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.092, + 0.255, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.235, + 0.194, + 0.248 + ], + "angle": 0, + "content": "3DGS" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.092, + 0.416, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.235, + 0.411, + 0.248 + ], + "angle": 0, + "content": "3DGS (lower threshold)" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.092, + 0.577, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.235, + 0.572, + 0.248 + ], + "angle": 0, + "content": "LPM wo/ point addition" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.092, + 0.738, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.235, + 0.702, + 0.248 + ], + "angle": 0, + "content": "LPM wo/ reset" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.092, + 0.901, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.235, + 0.86, + 0.247 + ], + "angle": 0, + "content": "3DGS+LPM" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.257, + 0.908, + 0.286 + ], + "angle": 0, + "content": "Figure 4. Effect of key operations of LPM. We show that the point addition operation effectively captures the geometric details in the scene; The point reset operation based on the error map further calibrate the geometry." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.297, + 0.473, + 0.607 + ], + "angle": 0, + "content": "
MethodPSNRDSSIMLPIPSFPS
\\( LLFF^1 \\)23.240.0760.235-
\\( DyNeRF^1 \\)29.580.0830.0630.015
\\( Dynamic-4DGS^1 \\)---30
\\( 4DGS^1 \\)29.38--114
\\( STGS^1 \\)29.580.0220.063103
\\( STGS*^1 \\)29.480.0230.066110
\\( STGS*^1 + LPM \\)29.840.0220.062105
StreamRF28.26-0.03910.9
NeRFPlayer30.69-0.1110.05
HyperReal31.10-0.0962
K-planes31.63-0.313
Mix Voxels-X31.73-0.0644.6
Dynamic-4DGS31.150.0160.04930
4DGS32.01-0.055114
STGS32.050.0140.044140
STGS*31.990.0150.045145
\\( STGS*+ LPM \\)32.400.0140.045140
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.618, + 0.484, + 0.701 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on the Neural 3D Video dataset. \"FPS\" is measured at a resolution of \\(1352 \\times 1014\\). Some methods only report results for a subset of scenes. For a fair comparison, we report LPM's results under two pre-existing settings. Only includes the Flame Salmon scene. Bold represents best, underline indicates second best." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.864 + ], + "angle": 0, + "content": "cularly in the tree at the second. These regions require more points for accurate population, leading to a more precise and detailed reconstruction. Additionally, the tablecloth in the first row is affected by ill-conditioned points. Furthermore, we provide depth and mesh comparisons in the third and final rows. All these observations demonstrate that our geometry calibration with LPM offers an opportunity to correct these potentially ill-conditioned points, thereby enhancing the overall reconstruction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Results on dynamic 4D datasets Table 2 presents a quantitative evaluation on the Neural 3D Video Dataset." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.298, + 0.905, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.512 + ], + "angle": 0, + "content": "Figure 5. Qualitative evaluation on dynamic Neural 3D Video dataset [17]. LPM improves STGS [18] for both scenes Transparent (e.g., window) and Dynamic movements (e.g., dog's tongue)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.906, + 0.653 + ], + "angle": 0, + "content": "Following established practices, training and evaluation are conducted at half resolution, with the first camera held out for evaluation [16]. Integrating our LPM into SpaceTimeGS yields the best performance across all comparisons. Notably, our method demonstrates significant improvements in the challenging Flame Salmon scene compared to SpaceTimeGS [18]. Our approach not only surpasses previous methods in rendering quality but also maintains comparable rendering speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.79 + ], + "angle": 0, + "content": "In addition to the quantitative assessment, we provide qualitative comparisons on the Flame Salmon and Flame Steak scenes, as illustrated in Figure 5. The quality of synthesis in both static and dynamic regions markedly outperforms STGS. Several intricate details, including the tree behind the window and the fine features like the dog's tongue, are faithfully reproduced with higher accuracy compared to STGS [18]. Both examples indicate that LPM improves upon STGS for superior scene modeling." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.798, + 0.665, + 0.815 + ], + "angle": 0, + "content": "4.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.818, + 0.906, + 0.863 + ], + "angle": 0, + "content": "We conducted ablation studies on the more challenging scene: PlayRoom from Deep Blending [12] and Truck from Tanks&Temples [15]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Effectiveness and cost of LPM We hypothesize that the Adaptive Density Control (ADC) tends to overlook under" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21702" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.254, + 0.089, + 0.74, + 0.187 + ], + "angle": 0, + "content": "
MethodPlayRoomTruck
PSNRLPIPSSSIMPSNRLPIPSSSIM
Full LPM30.220.2410.91025.610.1540.883
wo/ point addition30.100.2410.91025.430.1530.883
wo/ reset30.070.2430.90825.520.1440.883
" + }, + { + "type": "table_caption", + "bbox": [ + 0.315, + 0.19, + 0.68, + 0.205 + ], + "angle": 0, + "content": "Table 3. Performance comparison for different configurations" + }, + { + "type": "table", + "bbox": [ + 0.174, + 0.215, + 0.821, + 0.405 + ], + "angle": 0, + "content": "
SceneMethodPSNRLPIPSGaussiansTraining time
PlayRoom3DGS*30.030.244232k22min
3DG* (lower threshold)29.690.240523k36min
GaussianProOut of Memory
PielGS30.090.241186k35min
3DGS + LPM30.220.241186k23min
Truck3DGS*25.420.146257k19 min
3DGS* (lower threshold)25.450.127635k35min
GaussianPro25.400.164312k36min
PielGS25.510.121518k37min
3DGS + LPM25.610.154265k21min
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.408, + 0.907, + 0.437 + ], + "angle": 0, + "content": "Table 4. Cost-effectiveness analysis. Rendering speed of both methods are measured on our machine. Note: For 3DGS+LPM, training time includes the feature matching process." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.483, + 0.852 + ], + "angle": 0, + "content": "optimized points due to its simplistic approach of thresholding the average gradient. The straight way to identify the all points is lowering threshold to densification process. Although this solution can reduce blurring in specific regions, such as the toy (red box) illustrated in Figure 4, it still has limitations. As shown in Table 4, lowering the threshold for 3DGS significantly increases the number of Gaussian points and decreases rendering speed. Additionally, the PSNR of the quantitative results decreases due to the introduction of unnecessary points in already dense areas. In contrast, LPM effectively generates points in areas indicated by the error map, leading to more accurate and detailed reconstructions while maintaining real-time rendering speed. As demonstrated by the qualitative comparison in Figure 4, 3DGS with LPM achieves superior qualitative results. We further compare our method with other recent methods that also focus on adaptive density control (ADC). While PixelGS and GaussianPro achieve improvements in rendering quality, their training times increase substantially as they only consider point addition and extra gradient propagation. In contrast, LPM achieves a noticeable improvement with only a slight increase in training time due to (1) point matching [19] is much faster (2) considering model expansion to dynamically prune the points by their addition number and (3) selecting points in error-contributing zones 3D zone using the parallel matrix operations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Individual points manipulation We study the effect of individual points manipulation of LPM, including the point addition and reset ill-conditional points. The results in Ta" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.583 + ], + "angle": 0, + "content": "ble 3 show that, (1) each manipulation is useful with positive gain, suggesting that the LPM is meaningful. (2) The point addition operation densify the under-optimized points which may be overlook in the 3DGS, further captures the geometry details (e.g., detail of toy and leaf of the tree, see Fig. 4). (3) Reset points in ceratin zone provide the opportunity of correct the ill-conditioned points to achieve geometry calibration, (e.g., window of the trunk, see Fig. 4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.595, + 0.634, + 0.611 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.907, + 0.892 + ], + "angle": 0, + "content": "We propose Localized Point Management (LPM), a novel point management approach to address the limitations of the Adaptive Density Control (ADC) mechanism in 3D Gaussian Splatting (3DGS). The core idea of LPM is identifying the error-contributing 3D zones that require both point addition and geometry calibration under multiview geometry constraints, guided by image rendering errors. We implement appropriate operations for point densification and opacity reset. As a versatile plugin, LPM can be seamlessly integrated into existing 3DGS-based rendering methods. Extensive experiments across both static 3D and dynamic 4D scenes validate the efficacy of LPM in enhancing existing ADC mechanisms both quantitatively and qualitatively. While our method identifies the 3D Gaussian points that lead to rendering errors, it still follows the densification rules of 3DGS [14]. This approach may not be optimal for under-optimized points, and we leave this aspect for further investigation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21703" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.184 + ], + "angle": 0, + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.482, + 0.242 + ], + "angle": 0, + "content": "[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.482, + 0.47 + ], + "angle": 0, + "content": "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697–19705, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.482, + 0.526 + ], + "angle": 0, + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[7] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[8] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[9] Nianchen Deng, Zhenyi He, Jiannan Ye, Budmonde Duinkharjav, Praneeth Chakravarthula, Xubo Yang, and Qi Sun. Fov-nerf: Foveated neural radiance fields for virtual reality. IEEE Transactions on Visualization and Computer Graphics, 28(11):3854-3864, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[11] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12479–12488, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.124, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.905, + 0.236 + ], + "angle": 0, + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 2, 5, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.24, + 0.905, + 0.294 + ], + "angle": 0, + "content": "[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.297, + 0.905, + 0.393 + ], + "angle": 0, + "content": "[16] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multiview video. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5511-5521, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.398, + 0.905, + 0.492 + ], + "angle": 0, + "content": "[17] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.497, + 0.904, + 0.537 + ], + "angle": 0, + "content": "[18] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. arXiv preprint arXiv:2312.16812, 2023. 3, 4, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.541, + 0.905, + 0.596 + ], + "angle": 0, + "content": "[19] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627-17638, 2023. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.6, + 0.905, + 0.653 + ], + "angle": 0, + "content": "[20] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.657, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.716, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[22] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.841 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21704" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[25] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.137, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[26] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. arXiv e-prints, pages arXiv-2404, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.181, + 0.483, + 0.25 + ], + "angle": 0, + "content": "[27] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.253, + 0.482, + 0.295 + ], + "angle": 0, + "content": "[28] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.297, + 0.482, + 0.38 + ], + "angle": 0, + "content": "[29] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.383, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[30] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.441, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[31] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.5, + 0.483, + 0.568 + ], + "angle": 0, + "content": "[32] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.572, + 0.482, + 0.653 + ], + "angle": 0, + "content": "[33] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4150–4159, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.657, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[34] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.483, + 0.784 + ], + "angle": 0, + "content": "[35] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5438-5448, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[36] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. arXiv preprint arXiv:2311.17089, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[38] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[39] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[40] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. arXiv preprint arXiv:2311.16493, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[41] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[42] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21705" + } + ] +] \ No newline at end of file diff --git a/2025/Improving Gaussian Splatting with Localized Points Management/6436626a-e39e-4d87-9a2d-0ebf83ce94c9_origin.pdf b/2025/Improving Gaussian Splatting with Localized Points Management/6436626a-e39e-4d87-9a2d-0ebf83ce94c9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2bfc96b866525775a745fbdf999188a90fa7d6cc --- /dev/null +++ b/2025/Improving Gaussian Splatting with Localized Points Management/6436626a-e39e-4d87-9a2d-0ebf83ce94c9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b904a9e7f60255df0e1a9f4ef2d189e7e6af33c83682ef19991e4849787d0612 +size 2181569 diff --git a/2025/Improving Gaussian Splatting with Localized Points Management/full.md b/2025/Improving Gaussian Splatting with Localized Points Management/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5f1058e445741e9a754610601cd578dfeea3b9a1 --- /dev/null +++ b/2025/Improving Gaussian Splatting with Localized Points Management/full.md @@ -0,0 +1,266 @@ +# Improving Gaussian Splating with Localized Points Management + +Haosen Yang $^{1*}$ Chenhao Zhang $^{1*}$ Wenqing Wang $^{1*}$ Marco Volino $^{1}$ Adrian Hilton $^{1}$ Li Zhang $^{2}$ Xiatian Zhu $^{1}$ + +$^{1}$ University of Surrey $^{2}$ Fudan University *Equal contribution + +https://happy-hsy.github.io/projects/LPM/ + +![](images/358cb499d038a196051c1c0a34f80fbbe0ca0cbe4eff80ce263410d0b97699b7.jpg) + +![](images/0c2e3bbb8d62f14687f3298dbb0a3d6a38241c548439c70595cd677fe39328cd.jpg) +Rendered Image + +![](images/ec2687b81814dd39bb44334649a0648956d27ce407b14cc9374bacfd324b3459.jpg) + +![](images/f99e428bbbd51fb643b315becf2f4502fc86ce2ea23d5fbf19ad39d7da4d4f40.jpg) +Gaussian Points + +![](images/33c209ad02985e0a553243f04c4eb01edb0ade0279341fc74ad5010928b0c955.jpg) + +![](images/7fdbb11f6ea55d5f47d94bfe05b980fdb4a0d527fb02e85ea8b0bde007b36fa4.jpg) +Rendered Depth +Figure 1. Visualization of points behavior. 3DGS produces ill-conditioned Gaussians (red box) that occlude other valid points, resulting in noticeably incorrect depth estimation. LPM handles these ill-conditioned points to reduce negative impacts and further calibrate the geometry. + +# Abstract + +Point management is critical for optimizing 3D Gaussian Splatting models, as point initiation (e.g., via structure from motion) is often distributionally inappropriate. Typically, Adaptive Density Control (ADC) algorithm is adopted, leveraging view-averaged gradient magnitude thresholding for point densification, opacity thresholding for pruning, and regular all-points opacity reset. We reveal that this strategy is limited in tackling intricate/special image regions (e.g., transparent) due to inability of identifying all 3D zones requiring point densification, and lacking an appropriate mechanism to handle ill-conditioned points with negative impacts (e.g., occlusion due to false high opacity). To address these limitations, we propose a Localized Point Management (LPM) strategy, capable of identify + +ing those error-contributing zones in greatest need for both point addition and geometry calibration. Zone identification is achieved by leveraging the underlying multiview geometry constraints, subject to image rendering errors. We apply point densification in the identified zones and then reset the opacity of the points in front of these regions, creating a new opportunity to correct poorly conditioned points. Serving as a versatile plugin, LPM can be seamlessly integrated into existing static 3D and dynamic 4D Gaussian Splatting models with minimal additional cost. Experimental evaluations validate the efficacy of our LPM in boosting a variety of existing 3D/4D models both quantitatively and qualitatively. Notably, LPM improves both static 3DGS and dynamic SpaceTimeGS to achieve state-of-the-art rendering quality while retaining real-time speeds, excelling on challenging datasets such as Tanks & Temples and the Neural + +# 3D Video dataset. + +# 1. Introduction + +Neural rendering has emerged as a generalizable, flexible, and powerful approach for photorealistic novel view synthesis (NVS) of any camera poses [23], underpinning a wide variety of applications in augmented/virtual/mixed reality [9], robotics [37], and generation [25], among more others. For example, taking a learning-based parametric idea, Neural Radiance Fields (NeRFs) [23] implicitly represent the scene radiance of any complexity using neural networks (e.g., MLPs), without the tedious requirements of model handcrafting for accounting the scene variations in geometry, texture, illumination. However, their view rendering is inefficient computationally due to heavy ray sampling, thus suffer in scaling to high-resolution content applications and large scale scene modeling [29, 32]. + +Recently, 3D Gaussian Splatting (3DGS) [14] has come as an alternative with explicit representation, much faster model optimization and real-time neural rendering. The process begins by initializing a set of 3D Gaussian points using Structure from Motion (SfM) [28]. This is followed by optimizing the parameters of these points through view reconstruction loss, resulting in a view output generated with differentiable splatting-based rasterization. However, the point initialization is often distributionally non-optimal, leading to issues such as under-population (e.g., insufficient points) or over-population (e.g., excessive points) in the 3D space. Consequently, a point management mechanism, such as Adaptive Density Control (ADC), is necessary during optimization. However, we identify several limitations with ADC: (i) Thresholding the average gradient to determine regions for point densification often overlooks under-optimized points. For instance, larger Gaussian points typically have lower average gradients and may frequently appear across various views in screen space. (ii) Point sparsity complicates the addition of sufficient and reliable points needed to comprehensively cover the scene. (iii) Mis-optimized Gaussian points can have detrimental effects, such as occluding other valuable points and leading to incorrect depth estimates (see erroneous placements on windows in Fig. 1). While GaussianPro [7] and PixelGS [41] try to solve (1) and (2) through multi-view stereo [2] and additional gradient propagation, respectively, these methods significantly increase the training budget, as shown in the Table 4. + +To overcome the aforementioned limitations, in this paper we propose a novel and efficient Localized Point Management (LPM) approach. Our idea is intuitive - identifying those 3D Gaussian points leading to rendering errors. Thus we start with an image rendering error map of a specific view. To obtain the error contributing 3D points, we leverage the region correspondence between different + +views via feature mapping, subject to the multiview geometry constraint. For each pair of corresponded regions, we cast the rays through them at their respective camera views in the cone shape, and consider their intersection as the error source zone. Within each such zone, we consider two situations: (1) At presence of points, we further apply point densification at a lower threshold to complement the original counterpart locally; (2) In case no point due to point sparsity, we add new Gaussian points. Concurrently, we reset the opacity of points with high opacity estimates that are located in front of these zones, as they can significantly affect view rendering. This provides an opportunity to correct potentially ill-conditioned points while tuning the newly added ones in the subsequent optimization. To minimize model expansion, we prune the points by opacity in a density-aware manner. + +We summarize the contributions below: (1) Through in-depth analysis, we have identified several limitations in the standard point management mechanism used in Gaussian Splatting that impede model optimization. (2) We present Localized Point Management (LPM) for these issues by identifying error-contributing 3D zones and implementing appropriate operations for point densification and opacity reset. (3) Extensive experiments validate the benefits of our LPM in improving a diversity of existing 3D and 4D Gaussian Splatting models in novel view synthesis on both static and dynamic scenes. + +# 2. Related Work + +Neural Scene Representations has always been an important direction in novel view synthesis. Previous methods allocate neural features to structures such as volume [21, 27], texture [31], and point cloud [1]. The pioneering work of NeRF [23] proposes integrating neural networks with 3D volumetric representations to convert a 3D scene into a learnable density field, enabling high-quality novel view synthesis without requiring explicit modeling of the 3D scene and illumination. Later on, numerous works emerge to boost the quality and efficiency of volume rendering, [3, 5, 35] refine the point sampling strategy in ray marching, some some advanced works [4, 33] reparameterize the scene to produce a more compact representation. Additionally, regularization terms [8, 39] can be incorporated to constrain the scene representation, resulting in a closer approximation to real-world geometry. Despite their high-quality representational performance, these methods are typically computationally inefficient for view rendering due to the extensive ray sampling required and the use of Multi-Layer Perceptrons (MLPs) to represent the scene, complicating the computation and optimization of any point within the scene. To address this, several works have proposed novel scene representations aimed at accelerating the rendering process. These representations replace MLPs with sparse + +voxels [20], hash tables [24], or triplanes [6], significantly enhancing rendering speed. However, real-time rendering remains challenging due to the inherent complexity of the ray marching strategy in volume rendering. + +Gaussian Splatting represents a recent advancement in novel view synthesis, enabling real-time high-quality rendering. It contributes to splatting-based rasterization by computing pixel colors through depth sorting and $\alpha$ -blending of projected 2D Gaussians, thereby avoiding the complex sampling strategies of ray marching and achieving real-time performance. It is precisely due to its real-time high-quality rendering capabilities that 3DGS has been applied to various domains, including autonomous driving, content generation [30], and 4D dynamic scenes [18, 34, 38], among others. Despite these advancements, 3DGS still has some drawbacks, such as the storage of Gaussians and handling multi-resolution, and so on. Several works have enhanced 3DGS by improving Gaussian representation, including techniques such as low-pass filtering [40], multiscale Gaussian representations [36], and interpolating Gaussian attributes from structured grid features [22]. However, these works often overlook the importance of point management, specifically Adaptive Density Control, which is typically applied during optimization to address issues like under-population or over-population in the 3D space. Only a few works have focused on point management. For example, GaussianPro [7] directly tackles densification limitations, bridging gaps from SfM-based initialization. Pixel-GS [41] proposes a gradient scaling strategy to suppress artifacts near the camera. Additionally, [26] introduces an auxiliary per-pixel error function to implicitly supervise point contributions. + +Although these methods improve densification, they are still unable to identify all 3D zones that require point densification and lack a proper mechanism to handle ill-conditioned points with negative impacts. Here, we propose a novel approach, Localized Point Management, capable of identifying error-contributing zones with greatest demand for both point addition and geometry calibration. + +# 3. Method + +# 3.1. Preliminaries: 3D Gaussian Splitting + +Gaussian Splatting builds upon concepts from EWA [42] splatting and proposes modeling a 3D scene as a collection of 3D Gaussian points $\{G_i \mid i = 1, \dots, K\}$ , rendered through volume splatting. Each 3D Gaussian $G$ is defined by the equation: + +$$ +G (x) = e ^ {- \frac {1}{2} (x - \mu) ^ {T} \Sigma^ {- 1} (x - \mu)}, +$$ + +where $\mu \in \mathbb{R}^{3\times 1}$ represents the mean vector, and $\Sigma \in \mathbb{R}^{3\times 3}$ denotes its covariance matrix. To maintain the positive semi-definite nature of $\Sigma$ during optimization, it is repre + +sented as $\Sigma = RSS^{T}R^{T}$ , with the orthogonal rotation matrix $R\in \mathbb{R}^{3\times 3}$ and the diagonal scale matrix $S\in \mathbb{R}^{3\times 3}$ . + +To render an image from a specific viewpoint, the color of each pixel $p$ is determined by blending $N$ ordered Gaussians $\{G_i \mid i = 1, \dots, N\}$ that overlap $p$ , using the formula: + +$$ +c (p) = \sum_ {i = 1} ^ {N} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), +$$ + +where $\alpha_{i}$ is derived by evaluating a projected 2D Gaussian from $G_{i}$ at pixel $p$ combined with a learned opacity for $G_{i}$ , and $c_{i}$ is the learnable, view-dependent color modeled using spherical harmonics in 3DGS. Gaussians that influence $p$ are arranged in ascending order based on their depth from the current viewpoint. Employing differentiable rendering techniques allows for the end-to-end optimization of all Gaussian attributes through training view reconstruction. Point management Since existing 3D Gaussian points using Structure from Motion (SfM), the points are often coarse and non-optimal in space. During optimization, a point management mechanism, Adaptive Density Control (ADC), is typically applied to manage point distribution issues. Specifically, thresholding the average gradient is used to decide on point densification. For each Gaussian point $G_{i}$ , 3DGS tracks the magnitude of the positional gradient $\frac{\partial L_{\pi}}{\partial \mu_i}$ across all rendered views, which is then averaged to a quantity $T_{i}$ . During each training iteration, if the gradient $T_{i}$ surpasses a predefined threshold, it considers this point as inadequately representing the corresponding 3D region. With the scale of the Gaussian as the size measure, a large Gaussian will be split into two, while a small one leads to point cloning. + +However, this commonly used ADC strategy is unable to identify all the 3D zones with the underlying need for point densification. This is because, often the local complexity of scene geometry varies significantly, which beyond the reach of any single-value based thresholding. Besides, there is lacking of a proper mechanism to handle ill-conditioned points with negative impacts (e.g., wrong opacity values estimated during training with points distributed here and there). + +# 3.2. Localized Gaussian Point Management + +To address the aforementioned issues, we introduce a novel model agnostic point management approach, Localized Point Management (LPM), which leverages multiview geometry constraints to identify error contributing 3D points, with the guidance of image rendering errors. This approach can be seamlessly integrated with existing 3DGS models without the need for architectural modification. As illustrated in Figure 2, we begin with an image rendering error map for a specific view. Under the multiview geometry constraint, the corresponding regions in the referred view are + +![](images/952171043dbcf77a86bdddb046a72ba2c1ac5a9ff212bc00e354f2b7c2e6ddc1.jpg) +Figure 2. Overview of our Localized Point Management (LPM). (a) We start with an image rendering error map versus the current view (the ground-truth). Concurrently, matching points are identified between the current view and a referred view sampled as an adjacent view via off-the-shelf feature mapping. (b) Subsequently, cross-view region mapping is then employed to locate the correspondence region in the refereed view. (c) For each pair of corresponded regions, we cast the rays through them at their respective camera views in the cone shape, and consider their intersection as the error source zone. The final step involves identifying under-optimized or ill-conditioned points within these zones, where under-optimized/empty places are densified, and ill-conditioned points are reset. + +matched via feature mapping. For each pair of corresponding regions, we then cast rays through them from their respective camera views in a cone and identify their intersection as the error source zone. Within each zone, we perform localized point manipulation. + +Error map generation To accurately localize those zones in the 3D space that require point densification and geometry calibration, we initiate our process by rendering the current view image through the splatting of 3D Gaussians. This is followed by generating an error map (Figure 2(a)) for this specific view against the ground-truth image using an error function [18]. + +Error contributing 3D zone identification To project this rendering error back to the 3D space, we leverage the region correspondence between different views under multiview geometry constraints. This involves the following two key steps. + +(i) Cross-view region mapping We select a neighboring view as the referred image. Following LightGlue [19], which predicts a partial assignment between two sets of local features extracted from images $A$ and $B$ from different views. Each feature consists of sets of 2D features position $\{F_i\mid (x_i,y_i)\in [0,1]^2\}$ , normalized by the image size. The images $A$ and $B$ contain $M$ and $N$ local features. LightGlue outputs a set of correspondences $\mathcal{M} = \{(i,j)\} \subseteq A\times B$ , where $i$ and $j$ denote the indices of matched points from sets $A$ and $B$ , respectively. Since the 2D rendering error regions in the current view may not all appear in the referenced image, we select the paired regions $(R_e,R_e^{\prime})$ (Fig + +ure 2(b)) based on the matching points, where $R_{e}$ represents the region in the current view and $R_{e}^{\prime}$ corresponds to the region in the reference view. Additionally, this paired region undergoes multiview adaptive adjustments based on the error map throughout the optimization process. + +(ii) 2D-to-3D projection After obtaining the paired regions with render errors, we project each 2D error region to the 3D space via multiview geometry constraints. Specifically, we cast the rays $\mathcal{C}$ in cone shape for region $R_{e}$ from the camera's center of projection $o$ along the direction $d$ , which aligns with the pixel's center (Figure 2(c)). The apex of this cone is located at $o$ , and its radius at the image plane. Hence, $o + d$ is parameterized as $\mathcal{C}$ . The radius $r_{Cone}$ is set to match the radius of the smallest circumscribed circle of the 2D plane error region, creating a cone on the 3D space that can trace the Gaussian points contributing to the 2D error region. Concurrently, a corresponding cone, denoted as $\mathcal{C}'$ , belong to region $R_{e}'$ is similarly projected. Subsequently, we compute the intersection points of these rays. In order to regionalize these points, we directly use a smallest sphere that can contain these points as error source 3D zone $R_{zone}$ . + +Points manipulation Recall that in existing 3DGS, points management only relies on the view-averaged gradient magnitude $\tau$ to determine point densification globally. In addition to this, we further perform localized points addition and geometry calibration within the identified error source 3D zone $R_{zone}$ . For the point addition, we consider two common situations: (1) In the presence of points, we + +![](images/4af857b32b53adba7469c8f1470e353e1fb5c997334d26931316f39b6fc1785c.jpg) +(a) + +![](images/d94b2985b4d391e94740e620d165fef7852b10109970ef4e06131bd27dedae9a.jpg) +(b) +Ground Truth + +![](images/39a166ef3023b1779dd82830719b2090548fb966c74816b1e0291e8c3191c176.jpg) + +![](images/8a2277b7bad2b600b665fa75d556bbb3b7814bd695285a8f152b07f84bb70f1e.jpg) + +![](images/6f5182dc84b379d2b2d13d406611c6914b9f4d4bfe214fb532af9ae25b9c57b5.jpg) + +![](images/28751b1d099292f1be1a556e31ce5afd7035f91224b4a8b7319ff432a13be817.jpg) + +![](images/318b43f2bd59ad2957c4378cd652cd4b7212512cb55f16f686f5d55624b33af1.jpg) +(c) + +![](images/e09305f5f120939ddfef0019aec4daff29795b1ef87b210dd7e8a9f4d318204f.jpg) +3DGS* rendered RGB image + +![](images/568fe3af2967250ac1bbbcd6c81346af4e4373be85873542b6ad3ea6f6177d88.jpg) +3DGS+LPM rendered RGB image + +![](images/7048a4930fe1b81605dec1c41332856a55e74a6f94bddb1732a2ab3a41f59d89.jpg) +Ground Truth +(d) +2DGS rendered mesh +Figure 3. Qualitative evaluation of our LPM across diverse static datasets [4, 12]. Our LPM improves 2DGS [13] and 3DGS [14] on these challenging scenarios, e.g. (a) Light artifacts, (b) Completeness in the distance, (c) Depth structure and (d) Mesh details. See red patches for highlighted visual differences. + +![](images/db3c37c9cabbaa38cc5520bb1cca71eea192e8a11b6de1aed0171859cac26f5d.jpg) +3DGS* rendered depth +2DGS+LPM rendered mesh + +![](images/00cac4e6b2ed17da7009a204540d579cd4991d7bd6448a598ddd4ca5151abfb7.jpg) +3DGS+LPM rendered depth +2DGS vs 2DGS+LPM heatmap + +apply point densification to locally complement the original counterparts. We set a lower threshold to select the points that need densification, aiming to enhance the geometric details. The densification rule is consistent with 3DGS, but it focuses on local 3D zones that need it most. Specifically, for small Gaussians, our strategy involves cloning the Gaussians while maintaining their size and repositioning them along the positional gradient to better capture emerging geometrical features. Conversely, larger Gaussians situated in areas of high variance are split into smaller points to more accurately represent the underlying geometry. (2) In cases + +of point sparsity, we add new Gaussian points at the center of the 3D zone. In the context of $\alpha$ -blending in 3DGS, if the points at the forefront of the identified 3D zone $R_{zone}$ have the highest opacity, they may occlude valid points, leading to incorrect depth estimation, as shown in Figure 1. To deal with such issues, we treat these points as potentially ill-conditioned. We reset these points to provide an opportunity for correction, further calibrating the geometry. To minimize model expansion, we adaptively prune points based on their opacity values, starting from low to high. The number of points pruned is determined by the density of + +
MethodMip-NeRF 360Tanks&TemplesDeep Blending
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
Plenoxels23.080.6250.46321.080.7190.37923.060.7950.510
INGP-Big25.590.6990.33121.920.7450.30524.960.8170.390
Mip-NeRF 36027.690.7920.23722.220.7590.25729.400.9010.245
3DGS27.210.8150.21423.140.8410.18329.410.9030.243
3DGS*27.470.8160.21623.670.8490.17729.550.9040.245
3DGS* + LPM27.590.8200.21623.830.8500.18129.760.9080.241
2DGS*27.150.8080.24623.580.8320.18529.350.8990.262
2DGS* + LPM27.420.8170.22823.650.8480.18029.520.9030.240
MipGS*27.510.8170.21023.690.8520.17329.580.9100.242
MipGS* + LPM27.700.8210.21023.820.8510.18029.610.9100.241
PixelGS*27.540.8190.20323.750.8500.17529.580.9200.220
PixelGS* + LPM27.800.8300.19024.020.8560.17329.650.9100.196
+ +Table 1. Comparison of various methods across different scenes on the Mip-NeRF 360 dataset, Tanks&Temples and Deep Blending. * indicates the retrained model from the official implementation. Bold represents best, underline indicates second best. + +points in the zone. This strategic reduction ensures that our point management remains cost efficient and adaptive to the evolving needs of the scene representation. + +# 4. Experiment + +Datasets and metrics We conducted an extensive evaluation using both static and dynamic scenes derived from publicly datasets. For static scenes, our approach was applied to a total of 13 scenes as specified in the 3DGS framework [14], which includes nine scenes from Mip-NeRF360 [3], two from Tanks&Temples [15], and two from DeepBlending [12]. In the context of dynamic scenes, our approach was tested across six scenes from the Neural 3D Video Dataset [17]. + +To evaluate novel view synthesis performance, we followed standard protocols by selecting one out of every eight images as test images, with the remaining used for training in static scenes. For each dynamic scene within the Neural 3D Video Dataset, one view was designated for testing while the others were allocated for training purposes. Evaluation metrics included the peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and the learned perceptual image patch similarity (LPIPS), which are broadly recognized standards in the field. + +Baselines and implementation Vanilla 3D Gaussian Splitting (3DGS) [14], 2D Gaussian Splitting (2DGS) [13], Mip Gaussian Splitting (MipGS) [40], PiexlGS [41] and SpacetimeGS (STGS) [18] were selected as our main baselines for their established art performance in novel view synthesis. For the static 3D benchmark, we also recorded the results of Mip-NeRF360 [3], iNGP [24] and Plenoxels [10] as in [14]. For the Dynamic 4D benchmark, we performed system comparison, such as DyNeRF [16], K-planes [11] and so on. In alignment with the approach + +described in 3DGS an STGS, our models were trained for 30k iterations across all scenes, following the same training schedule and hyperparameters. In addition to the original Gaussian densification strategies used in 3DGS and SpaceTime Gaussian, we also performed localized points management, including addition, reset, and pruning. We maintained the same thresholds for splitting and cloning points as in the original 3DGS and SpaceTime Gaussian. For point matching, we performed offline extraction to save computational cost. All experiments were conducted on an RTX 3090 GPU with 24GB of memory. + +# 4.1. Main Results + +Results on static 3D datasets The quantitative results (PSNR, SSIM, and LPIPS) on the Mip-NeRF 360 and Tanks & Temples datasets are presented in Tables 1. We retrained the 3DGS model (referred to as 3DGS*) as it yields better performance compared to the vanilla 3DGS and its variants. Our approach achieves results comparable to the state-of-the-art on the Mip-NeRF360 dataset and further enhances all 3DGS based method using our point management technique. Additionally, LPM improve vanilla 3DGS and PixlGS to set new state-of-the-art results on the Mip-NeRF 360, Tanks & Temples datasets and Dep Blending, effectively capturing more challenging environments (e.g., light effects, transparency). These results quantitatively validate the effectiveness of our method in improving the quality of reconstruction. + +In Figures 3, we present a comparison between 3DGS [14] and $3\mathrm{DGS}^* +$ LPM, focusing on both appearance and depth. A variety of improvements can be observed, particularly in challenging cases such as light effects, completeness at a distance. Our LPM significantly reduces artifacts in specific regions on top of 3DGS, partic + +![](images/906994016d490f4e7767cd3908d7a7e39402ad37d726ac55740116d093f88635.jpg) +3DGS + +![](images/15864ed2fb62f3531cd8c07e815f89652ef6a5b1d192d42133f9d126136fd718.jpg) +3DGS (lower threshold) + +![](images/d6802d13c8f1d0b05111bd681b98c0a797428013f63f9cee0b59ee15ad59349e.jpg) +LPM wo/ point addition + +![](images/a39fbabb937bf7897895c2b4a2554039b50e6886337b0c3e4e9cc82c9571bc75.jpg) +LPM wo/ reset + +![](images/213aae9e7575bcf3ab79ed9b2fe4e5ae3f4381b0ba2ce104c93d315888ab62df.jpg) +3DGS+LPM + +
MethodPSNRDSSIMLPIPSFPS
\( LLFF^1 \)23.240.0760.235-
\( DyNeRF^1 \)29.580.0830.0630.015
\( Dynamic-4DGS^1 \)---30
\( 4DGS^1 \)29.38--114
\( STGS^1 \)29.580.0220.063103
\( STGS*^1 \)29.480.0230.066110
\( STGS*^1 + LPM \)29.840.0220.062105
StreamRF28.26-0.03910.9
NeRFPlayer30.69-0.1110.05
HyperReal31.10-0.0962
K-planes31.63-0.313
Mix Voxels-X31.73-0.0644.6
Dynamic-4DGS31.150.0160.04930
4DGS32.01-0.055114
STGS32.050.0140.044140
STGS*31.990.0150.045145
\( STGS*+ LPM \)32.400.0140.045140
+ +Table 2. Quantitative comparisons on the Neural 3D Video dataset. "FPS" is measured at a resolution of $1352 \times 1014$ . Some methods only report results for a subset of scenes. For a fair comparison, we report LPM's results under two pre-existing settings. Only includes the Flame Salmon scene. Bold represents best, underline indicates second best. + +cularly in the tree at the second. These regions require more points for accurate population, leading to a more precise and detailed reconstruction. Additionally, the tablecloth in the first row is affected by ill-conditioned points. Furthermore, we provide depth and mesh comparisons in the third and final rows. All these observations demonstrate that our geometry calibration with LPM offers an opportunity to correct these potentially ill-conditioned points, thereby enhancing the overall reconstruction accuracy. + +Results on dynamic 4D datasets Table 2 presents a quantitative evaluation on the Neural 3D Video Dataset. + +![](images/6578bb930cdfefdec14552fc7a5404754bf324841aaa554d50a034e7fec3d307.jpg) +Figure 4. Effect of key operations of LPM. We show that the point addition operation effectively captures the geometric details in the scene; The point reset operation based on the error map further calibrate the geometry. +Figure 5. Qualitative evaluation on dynamic Neural 3D Video dataset [17]. LPM improves STGS [18] for both scenes Transparent (e.g., window) and Dynamic movements (e.g., dog's tongue). + +Following established practices, training and evaluation are conducted at half resolution, with the first camera held out for evaluation [16]. Integrating our LPM into SpaceTimeGS yields the best performance across all comparisons. Notably, our method demonstrates significant improvements in the challenging Flame Salmon scene compared to SpaceTimeGS [18]. Our approach not only surpasses previous methods in rendering quality but also maintains comparable rendering speed. + +In addition to the quantitative assessment, we provide qualitative comparisons on the Flame Salmon and Flame Steak scenes, as illustrated in Figure 5. The quality of synthesis in both static and dynamic regions markedly outperforms STGS. Several intricate details, including the tree behind the window and the fine features like the dog's tongue, are faithfully reproduced with higher accuracy compared to STGS [18]. Both examples indicate that LPM improves upon STGS for superior scene modeling. + +# 4.2. Ablation study + +We conducted ablation studies on the more challenging scene: PlayRoom from Deep Blending [12] and Truck from Tanks&Temples [15]. + +Effectiveness and cost of LPM We hypothesize that the Adaptive Density Control (ADC) tends to overlook under + +
MethodPlayRoomTruck
PSNRLPIPSSSIMPSNRLPIPSSSIM
Full LPM30.220.2410.91025.610.1540.883
wo/ point addition30.100.2410.91025.430.1530.883
wo/ reset30.070.2430.90825.520.1440.883
+ +Table 3. Performance comparison for different configurations + +
SceneMethodPSNRLPIPSGaussiansTraining time
PlayRoom3DGS*30.030.244232k22min
3DG* (lower threshold)29.690.240523k36min
GaussianProOut of Memory
PielGS30.090.241186k35min
3DGS + LPM30.220.241186k23min
Truck3DGS*25.420.146257k19 min
3DGS* (lower threshold)25.450.127635k35min
GaussianPro25.400.164312k36min
PielGS25.510.121518k37min
3DGS + LPM25.610.154265k21min
+ +Table 4. Cost-effectiveness analysis. Rendering speed of both methods are measured on our machine. Note: For 3DGS+LPM, training time includes the feature matching process. + +optimized points due to its simplistic approach of thresholding the average gradient. The straight way to identify the all points is lowering threshold to densification process. Although this solution can reduce blurring in specific regions, such as the toy (red box) illustrated in Figure 4, it still has limitations. As shown in Table 4, lowering the threshold for 3DGS significantly increases the number of Gaussian points and decreases rendering speed. Additionally, the PSNR of the quantitative results decreases due to the introduction of unnecessary points in already dense areas. In contrast, LPM effectively generates points in areas indicated by the error map, leading to more accurate and detailed reconstructions while maintaining real-time rendering speed. As demonstrated by the qualitative comparison in Figure 4, 3DGS with LPM achieves superior qualitative results. We further compare our method with other recent methods that also focus on adaptive density control (ADC). While PixelGS and GaussianPro achieve improvements in rendering quality, their training times increase substantially as they only consider point addition and extra gradient propagation. In contrast, LPM achieves a noticeable improvement with only a slight increase in training time due to (1) point matching [19] is much faster (2) considering model expansion to dynamically prune the points by their addition number and (3) selecting points in error-contributing zones 3D zone using the parallel matrix operations. + +Individual points manipulation We study the effect of individual points manipulation of LPM, including the point addition and reset ill-conditional points. The results in Ta + +ble 3 show that, (1) each manipulation is useful with positive gain, suggesting that the LPM is meaningful. (2) The point addition operation densify the under-optimized points which may be overlook in the 3DGS, further captures the geometry details (e.g., detail of toy and leaf of the tree, see Fig. 4). (3) Reset points in ceratin zone provide the opportunity of correct the ill-conditioned points to achieve geometry calibration, (e.g., window of the trunk, see Fig. 4). + +# 5. Conclusion + +We propose Localized Point Management (LPM), a novel point management approach to address the limitations of the Adaptive Density Control (ADC) mechanism in 3D Gaussian Splatting (3DGS). The core idea of LPM is identifying the error-contributing 3D zones that require both point addition and geometry calibration under multiview geometry constraints, guided by image rendering errors. We implement appropriate operations for point densification and opacity reset. As a versatile plugin, LPM can be seamlessly integrated into existing 3DGS-based rendering methods. Extensive experiments across both static 3D and dynamic 4D scenes validate the efficacy of LPM in enhancing existing ADC mechanisms both quantitatively and qualitatively. While our method identifies the 3D Gaussian points that lead to rendering errors, it still follows the densification rules of 3DGS [14]. This approach may not be optimal for under-optimized points, and we leave this aspect for further investigation. + +# References + +[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2 +[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 2 +[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2, 6, 1 +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 5 +[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697–19705, 2023. 2 +[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 3 +[7] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2, 3 +[8] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2 +[9] Nianchen Deng, Zhenyi He, Jiannan Ye, Budmonde Duinkharjav, Praneeth Chakravarthula, Xubo Yang, and Qi Sun. Fov-nerf: Foveated neural radiance fields for virtual reality. IEEE Transactions on Visualization and Computer Graphics, 28(11):3854-3864, 2022. 2 +[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 6 +[11] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12479–12488, 2023. 6 +[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for + +free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 5, 6, 7, 1 +[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 5, 6 +[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 2, 5, 6, 8, 1 +[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 6, 7, 1 +[16] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multiview video. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5511-5521, 2022. 6, 7 +[17] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 6, 7 +[18] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. arXiv preprint arXiv:2312.16812, 2023. 3, 4, 6, 7, 1 +[19] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627-17638, 2023. 4, 8 +[20] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 3 +[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2 +[22] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109, 2023. 3 +[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 3, 6 + +[25] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2 +[26] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. arXiv e-prints, pages arXiv-2404, 2024. 3 +[27] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 2 +[28] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2 +[29] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2 +[30] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 3 +[31] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 2 +[32] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 2 +[33] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4150–4159, 2023. 2 +[34] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3 +[35] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5438-5448, 2022. 2 +[36] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. arXiv preprint arXiv:2311.17089, 2023. 3 +[37] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 2 + +[38] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In ICLR, 2024. 3 +[39] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 2 +[40] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. arXiv preprint arXiv:2311.16493, 2023. 3, 6 +[41] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 2, 3, 6 +[42] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 3 \ No newline at end of file diff --git a/2025/Improving Gaussian Splatting with Localized Points Management/images.zip b/2025/Improving Gaussian Splatting with Localized Points Management/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..317341fa533fd6f234d575f236cce6d7910f72d0 --- /dev/null +++ b/2025/Improving Gaussian Splatting with Localized Points Management/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:601da15cee0ec32754c0a368963e0302a9ab30892707848dd4dda42dafdf50f1 +size 875166 diff --git a/2025/Improving Gaussian Splatting with Localized Points Management/layout.json b/2025/Improving Gaussian Splatting with Localized Points Management/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..831cf4cd0d5a22d6152ce5931c622ab9daf351b3 --- /dev/null +++ b/2025/Improving Gaussian Splatting with Localized Points Management/layout.json @@ -0,0 +1,7594 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 509, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 509, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 509, + 121 + ], + "type": "text", + "content": "Improving Gaussian Splating with Localized Points Management" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "spans": [ + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "text", + "content": "Haosen Yang" + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "text", + "content": " Chenhao Zhang" + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "text", + "content": " Wenqing Wang" + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "text", + "content": " Marco Volino" + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "text", + "content": " Adrian Hilton" + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "text", + "content": " Li Zhang" + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "text", + "content": " Xiatian Zhu" + }, + { + "bbox": [ + 80, + 141, + 529, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 184, + 408, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 184, + 408, + 214 + ], + "spans": [ + { + "bbox": [ + 201, + 184, + 408, + 214 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 184, + 408, + 214 + ], + "type": "text", + "content": "University of Surrey " + }, + { + "bbox": [ + 201, + 184, + 408, + 214 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 201, + 184, + 408, + 214 + ], + "type": "text", + "content": "Fudan University *Equal contribution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 201, + 227, + 405, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 227, + 405, + 242 + ], + "spans": [ + { + "bbox": [ + 201, + 227, + 405, + 242 + ], + "type": "text", + "content": "https://happy-hsy.github.io/projects/LPM/" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 62, + 269, + 233, + 357 + ], + "blocks": [ + { + "bbox": [ + 62, + 269, + 233, + 357 + ], + "lines": [ + { + "bbox": [ + 62, + 269, + 233, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 233, + 357 + ], + "type": "image", + "image_path": "358cb499d038a196051c1c0a34f80fbbe0ca0cbe4eff80ce263410d0b97699b7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 62, + 357, + 233, + 445 + ], + "blocks": [ + { + "bbox": [ + 62, + 357, + 233, + 445 + ], + "lines": [ + { + "bbox": [ + 62, + 357, + 233, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 233, + 445 + ], + "type": "image", + "image_path": "0c2e3bbb8d62f14687f3298dbb0a3d6a38241c548439c70595cd677fe39328cd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 447, + 181, + 457 + ], + "lines": [ + { + "bbox": [ + 120, + 447, + 181, + 457 + ], + "spans": [ + { + "bbox": [ + 120, + 447, + 181, + 457 + ], + "type": "text", + "content": "Rendered Image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 233, + 270, + 387, + 357 + ], + "blocks": [ + { + "bbox": [ + 233, + 270, + 387, + 357 + ], + "lines": [ + { + "bbox": [ + 233, + 270, + 387, + 357 + ], + "spans": [ + { + "bbox": [ + 233, + 270, + 387, + 357 + ], + "type": "image", + "image_path": "ec2687b81814dd39bb44334649a0648956d27ce407b14cc9374bacfd324b3459.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 233, + 357, + 387, + 445 + ], + "blocks": [ + { + "bbox": [ + 233, + 357, + 387, + 445 + ], + "lines": [ + { + "bbox": [ + 233, + 357, + 387, + 445 + ], + "spans": [ + { + "bbox": [ + 233, + 357, + 387, + 445 + ], + "type": "image", + "image_path": "f99e428bbbd51fb643b315becf2f4502fc86ce2ea23d5fbf19ad39d7da4d4f40.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 448, + 337, + 457 + ], + "lines": [ + { + "bbox": [ + 278, + 448, + 337, + 457 + ], + "spans": [ + { + "bbox": [ + 278, + 448, + 337, + 457 + ], + "type": "text", + "content": "Gaussian Points" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 389, + 269, + 543, + 357 + ], + "blocks": [ + { + "bbox": [ + 389, + 269, + 543, + 357 + ], + "lines": [ + { + "bbox": [ + 389, + 269, + 543, + 357 + ], + "spans": [ + { + "bbox": [ + 389, + 269, + 543, + 357 + ], + "type": "image", + "image_path": "33c209ad02985e0a553243f04c4eb01edb0ade0279341fc74ad5010928b0c955.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 389, + 357, + 544, + 445 + ], + "blocks": [ + { + "bbox": [ + 389, + 357, + 544, + 445 + ], + "lines": [ + { + "bbox": [ + 389, + 357, + 544, + 445 + ], + "spans": [ + { + "bbox": [ + 389, + 357, + 544, + 445 + ], + "type": "image", + "image_path": "7fdbb11f6ea55d5f47d94bfe05b980fdb4a0d527fb02e85ea8b0bde007b36fa4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 449, + 496, + 459 + ], + "lines": [ + { + "bbox": [ + 436, + 449, + 496, + 459 + ], + "spans": [ + { + "bbox": [ + 436, + 449, + 496, + 459 + ], + "type": "text", + "content": "Rendered Depth" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 468, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 55, + 468, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 555, + 502 + ], + "type": "text", + "content": "Figure 1. Visualization of points behavior. 3DGS produces ill-conditioned Gaussians (red box) that occlude other valid points, resulting in noticeably incorrect depth estimation. LPM handles these ill-conditioned points to reduce negative impacts and further calibrate the geometry." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 151, + 521, + 200, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 521, + 200, + 533 + ], + "spans": [ + { + "bbox": [ + 151, + 521, + 200, + 533 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 546, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 546, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 546, + 296, + 714 + ], + "type": "text", + "content": "Point management is critical for optimizing 3D Gaussian Splatting models, as point initiation (e.g., via structure from motion) is often distributionally inappropriate. Typically, Adaptive Density Control (ADC) algorithm is adopted, leveraging view-averaged gradient magnitude thresholding for point densification, opacity thresholding for pruning, and regular all-points opacity reset. We reveal that this strategy is limited in tackling intricate/special image regions (e.g., transparent) due to inability of identifying all 3D zones requiring point densification, and lacking an appropriate mechanism to handle ill-conditioned points with negative impacts (e.g., occlusion due to false high opacity). To address these limitations, we propose a Localized Point Management (LPM) strategy, capable of identify" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "ing those error-contributing zones in greatest need for both point addition and geometry calibration. Zone identification is achieved by leveraging the underlying multiview geometry constraints, subject to image rendering errors. We apply point densification in the identified zones and then reset the opacity of the points in front of these regions, creating a new opportunity to correct poorly conditioned points. Serving as a versatile plugin, LPM can be seamlessly integrated into existing static 3D and dynamic 4D Gaussian Splatting models with minimal additional cost. Experimental evaluations validate the efficacy of our LPM in boosting a variety of existing 3D/4D models both quantitatively and qualitatively. Notably, LPM improves both static 3DGS and dynamic SpaceTimeGS to achieve state-of-the-art rendering quality while retaining real-time speeds, excelling on challenging datasets such as Tanks & Temples and the Neural" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21696" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 129, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 129, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 129, + 83 + ], + "type": "text", + "content": "3D Video dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 136, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 136, + 108 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 136, + 108 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 115, + 294, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 115, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 115, + 294, + 281 + ], + "type": "text", + "content": "Neural rendering has emerged as a generalizable, flexible, and powerful approach for photorealistic novel view synthesis (NVS) of any camera poses [23], underpinning a wide variety of applications in augmented/virtual/mixed reality [9], robotics [37], and generation [25], among more others. For example, taking a learning-based parametric idea, Neural Radiance Fields (NeRFs) [23] implicitly represent the scene radiance of any complexity using neural networks (e.g., MLPs), without the tedious requirements of model handcrafting for accounting the scene variations in geometry, texture, illumination. However, their view rendering is inefficient computationally due to heavy ray sampling, thus suffer in scaling to high-resolution content applications and large scale scene modeling [29, 32]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 282, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 295, + 628 + ], + "type": "text", + "content": "Recently, 3D Gaussian Splatting (3DGS) [14] has come as an alternative with explicit representation, much faster model optimization and real-time neural rendering. The process begins by initializing a set of 3D Gaussian points using Structure from Motion (SfM) [28]. This is followed by optimizing the parameters of these points through view reconstruction loss, resulting in a view output generated with differentiable splatting-based rasterization. However, the point initialization is often distributionally non-optimal, leading to issues such as under-population (e.g., insufficient points) or over-population (e.g., excessive points) in the 3D space. Consequently, a point management mechanism, such as Adaptive Density Control (ADC), is necessary during optimization. However, we identify several limitations with ADC: (i) Thresholding the average gradient to determine regions for point densification often overlooks under-optimized points. For instance, larger Gaussian points typically have lower average gradients and may frequently appear across various views in screen space. (ii) Point sparsity complicates the addition of sufficient and reliable points needed to comprehensively cover the scene. (iii) Mis-optimized Gaussian points can have detrimental effects, such as occluding other valuable points and leading to incorrect depth estimates (see erroneous placements on windows in Fig. 1). While GaussianPro [7] and PixelGS [41] try to solve (1) and (2) through multi-view stereo [2] and additional gradient propagation, respectively, these methods significantly increase the training budget, as shown in the Table 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": "To overcome the aforementioned limitations, in this paper we propose a novel and efficient Localized Point Management (LPM) approach. Our idea is intuitive - identifying those 3D Gaussian points leading to rendering errors. Thus we start with an image rendering error map of a specific view. To obtain the error contributing 3D points, we leverage the region correspondence between different" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": "views via feature mapping, subject to the multiview geometry constraint. For each pair of corresponded regions, we cast the rays through them at their respective camera views in the cone shape, and consider their intersection as the error source zone. Within each such zone, we consider two situations: (1) At presence of points, we further apply point densification at a lower threshold to complement the original counterpart locally; (2) In case no point due to point sparsity, we add new Gaussian points. Concurrently, we reset the opacity of points with high opacity estimates that are located in front of these zones, as they can significantly affect view rendering. This provides an opportunity to correct potentially ill-conditioned points while tuning the newly added ones in the subsequent optimization. To minimize model expansion, we prune the points by opacity in a density-aware manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 264, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 396 + ], + "type": "text", + "content": "We summarize the contributions below: (1) Through in-depth analysis, we have identified several limitations in the standard point management mechanism used in Gaussian Splatting that impede model optimization. (2) We present Localized Point Management (LPM) for these issues by identifying error-contributing 3D zones and implementing appropriate operations for point densification and opacity reset. (3) Extensive experiments validate the benefits of our LPM in improving a diversity of existing 3D and 4D Gaussian Splatting models in novel view synthesis on both static and dynamic scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 407, + 400, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 400, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 400, + 420 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 715 + ], + "type": "text", + "content": "Neural Scene Representations has always been an important direction in novel view synthesis. Previous methods allocate neural features to structures such as volume [21, 27], texture [31], and point cloud [1]. The pioneering work of NeRF [23] proposes integrating neural networks with 3D volumetric representations to convert a 3D scene into a learnable density field, enabling high-quality novel view synthesis without requiring explicit modeling of the 3D scene and illumination. Later on, numerous works emerge to boost the quality and efficiency of volume rendering, [3, 5, 35] refine the point sampling strategy in ray marching, some some advanced works [4, 33] reparameterize the scene to produce a more compact representation. Additionally, regularization terms [8, 39] can be incorporated to constrain the scene representation, resulting in a closer approximation to real-world geometry. Despite their high-quality representational performance, these methods are typically computationally inefficient for view rendering due to the extensive ray sampling required and the use of Multi-Layer Perceptrons (MLPs) to represent the scene, complicating the computation and optimization of any point within the scene. To address this, several works have proposed novel scene representations aimed at accelerating the rendering process. These representations replace MLPs with sparse" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21697" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "voxels [20], hash tables [24], or triplanes [6], significantly enhancing rendering speed. However, real-time rendering remains challenging due to the inherent complexity of the ray marching strategy in volume rendering." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 122, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 122, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 122, + 296, + 455 + ], + "type": "text", + "content": "Gaussian Splatting represents a recent advancement in novel view synthesis, enabling real-time high-quality rendering. It contributes to splatting-based rasterization by computing pixel colors through depth sorting and " + }, + { + "bbox": [ + 56, + 122, + 296, + 455 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 56, + 122, + 296, + 455 + ], + "type": "text", + "content": "-blending of projected 2D Gaussians, thereby avoiding the complex sampling strategies of ray marching and achieving real-time performance. It is precisely due to its real-time high-quality rendering capabilities that 3DGS has been applied to various domains, including autonomous driving, content generation [30], and 4D dynamic scenes [18, 34, 38], among others. Despite these advancements, 3DGS still has some drawbacks, such as the storage of Gaussians and handling multi-resolution, and so on. Several works have enhanced 3DGS by improving Gaussian representation, including techniques such as low-pass filtering [40], multiscale Gaussian representations [36], and interpolating Gaussian attributes from structured grid features [22]. However, these works often overlook the importance of point management, specifically Adaptive Density Control, which is typically applied during optimization to address issues like under-population or over-population in the 3D space. Only a few works have focused on point management. For example, GaussianPro [7] directly tackles densification limitations, bridging gaps from SfM-based initialization. Pixel-GS [41] proposes a gradient scaling strategy to suppress artifacts near the camera. Additionally, [26] introduces an auxiliary per-pixel error function to implicitly supervise point contributions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 456, + 296, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 541 + ], + "type": "text", + "content": "Although these methods improve densification, they are still unable to identify all 3D zones that require point densification and lack a proper mechanism to handle ill-conditioned points with negative impacts. Here, we propose a novel approach, Localized Point Management, capable of identifying error-contributing zones with greatest demand for both point addition and geometry calibration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 548, + 111, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 111, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 111, + 559 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 568, + 258, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 258, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 258, + 582 + ], + "type": "text", + "content": "3.1. Preliminaries: 3D Gaussian Splitting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 586, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 647 + ], + "type": "text", + "content": "Gaussian Splatting builds upon concepts from EWA [42] splatting and proposes modeling a 3D scene as a collection of 3D Gaussian points " + }, + { + "bbox": [ + 55, + 586, + 295, + 647 + ], + "type": "inline_equation", + "content": "\\{G_i \\mid i = 1, \\dots, K\\}" + }, + { + "bbox": [ + 55, + 586, + 295, + 647 + ], + "type": "text", + "content": ", rendered through volume splatting. Each 3D Gaussian " + }, + { + "bbox": [ + 55, + 586, + 295, + 647 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 586, + 295, + 647 + ], + "type": "text", + "content": " is defined by the equation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 651, + 237, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 651, + 237, + 667 + ], + "spans": [ + { + "bbox": [ + 113, + 651, + 237, + 667 + ], + "type": "interline_equation", + "content": "G (x) = e ^ {- \\frac {1}{2} (x - \\mu) ^ {T} \\Sigma^ {- 1} (x - \\mu)},", + "image_path": "b26a12f89b7563ee4937096f1d06bf9541b0c534dded1a7a9615c5593587dc26.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^{3\\times 1}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " represents the mean vector, and " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " denotes its covariance matrix. To maintain the positive semi-definite nature of " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " during optimization, it is repre" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "sented as " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\Sigma = RSS^{T}R^{T}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ", with the orthogonal rotation matrix " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "R\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " and the diagonal scale matrix " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "S\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "text", + "content": "To render an image from a specific viewpoint, the color of each pixel " + }, + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "text", + "content": " is determined by blending " + }, + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "text", + "content": " ordered Gaussians " + }, + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\{G_i \\mid i = 1, \\dots, N\\}" + }, + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "text", + "content": " that overlap " + }, + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "text", + "content": ", using the formula:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 373, + 142, + 495, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 142, + 495, + 175 + ], + "spans": [ + { + "bbox": [ + 373, + 142, + 495, + 175 + ], + "type": "interline_equation", + "content": "c (p) = \\sum_ {i = 1} ^ {N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}),", + "image_path": "2c2aac0ab3f95127000448efb1b458463fa49a249977dbf9a68856ccdadb4dfb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "text", + "content": " is derived by evaluating a projected 2D Gaussian from " + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "text", + "content": " combined with a learned opacity for " + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "text", + "content": " is the learnable, view-dependent color modeled using spherical harmonics in 3DGS. Gaussians that influence " + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "text", + "content": " are arranged in ascending order based on their depth from the current viewpoint. Employing differentiable rendering techniques allows for the end-to-end optimization of all Gaussian attributes through training view reconstruction. Point management Since existing 3D Gaussian points using Structure from Motion (SfM), the points are often coarse and non-optimal in space. During optimization, a point management mechanism, Adaptive Density Control (ADC), is typically applied to manage point distribution issues. Specifically, thresholding the average gradient is used to decide on point densification. For each Gaussian point " + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "text", + "content": ", 3DGS tracks the magnitude of the positional gradient " + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L_{\\pi}}{\\partial \\mu_i}" + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "text", + "content": " across all rendered views, which is then averaged to a quantity " + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "text", + "content": ". During each training iteration, if the gradient " + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 185, + 555, + 460 + ], + "type": "text", + "content": " surpasses a predefined threshold, it considers this point as inadequately representing the corresponding 3D region. With the scale of the Gaussian as the size measure, a large Gaussian will be split into two, while a small one leads to point cloning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 567 + ], + "type": "text", + "content": "However, this commonly used ADC strategy is unable to identify all the 3D zones with the underlying need for point densification. This is because, often the local complexity of scene geometry varies significantly, which beyond the reach of any single-value based thresholding. Besides, there is lacking of a proper mechanism to handle ill-conditioned points with negative impacts (e.g., wrong opacity values estimated during training with points distributed here and there)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 576, + 520, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 520, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 520, + 589 + ], + "type": "text", + "content": "3.2. Localized Gaussian Point Management" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "To address the aforementioned issues, we introduce a novel model agnostic point management approach, Localized Point Management (LPM), which leverages multiview geometry constraints to identify error contributing 3D points, with the guidance of image rendering errors. This approach can be seamlessly integrated with existing 3DGS models without the need for architectural modification. As illustrated in Figure 2, we begin with an image rendering error map for a specific view. Under the multiview geometry constraint, the corresponding regions in the referred view are" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21698" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 77, + 553, + 273 + ], + "blocks": [ + { + "bbox": [ + 60, + 77, + 553, + 273 + ], + "lines": [ + { + "bbox": [ + 60, + 77, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 60, + 77, + 553, + 273 + ], + "type": "image", + "image_path": "952171043dbcf77a86bdddb046a72ba2c1ac5a9ff212bc00e354f2b7c2e6ddc1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 277, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 555, + 344 + ], + "type": "text", + "content": "Figure 2. Overview of our Localized Point Management (LPM). (a) We start with an image rendering error map versus the current view (the ground-truth). Concurrently, matching points are identified between the current view and a referred view sampled as an adjacent view via off-the-shelf feature mapping. (b) Subsequently, cross-view region mapping is then employed to locate the correspondence region in the refereed view. (c) For each pair of corresponded regions, we cast the rays through them at their respective camera views in the cone shape, and consider their intersection as the error source zone. The final step involves identifying under-optimized or ill-conditioned points within these zones, where under-optimized/empty places are densified, and ill-conditioned points are reset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 353, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 295, + 413 + ], + "type": "text", + "content": "matched via feature mapping. For each pair of corresponding regions, we then cast rays through them from their respective camera views in a cone and identify their intersection as the error source zone. Within each zone, we perform localized point manipulation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 295, + 503 + ], + "type": "text", + "content": "Error map generation To accurately localize those zones in the 3D space that require point densification and geometry calibration, we initiate our process by rendering the current view image through the splatting of 3D Gaussians. This is followed by generating an error map (Figure 2(a)) for this specific view against the ground-truth image using an error function [18]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "content": "Error contributing 3D zone identification To project this rendering error back to the 3D space, we leverage the region correspondence between different views under multiview geometry constraints. This involves the following two key steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "(i) Cross-view region mapping We select a neighboring view as the referred image. Following LightGlue [19], which predicts a partial assignment between two sets of local features extracted from images " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " from different views. Each feature consists of sets of 2D features position " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{F_i\\mid (x_i,y_i)\\in [0,1]^2\\}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ", normalized by the image size. The images " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " contain " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " local features. LightGlue outputs a set of correspondences " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{(i,j)\\} \\subseteq A\\times B" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " denote the indices of matched points from sets " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ", respectively. Since the 2D rendering error regions in the current view may not all appear in the referenced image, we select the paired regions " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "(R_e,R_e^{\\prime})" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " (Fig" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 353, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 413 + ], + "type": "text", + "content": "ure 2(b)) based on the matching points, where " + }, + { + "bbox": [ + 313, + 353, + 555, + 413 + ], + "type": "inline_equation", + "content": "R_{e}" + }, + { + "bbox": [ + 313, + 353, + 555, + 413 + ], + "type": "text", + "content": " represents the region in the current view and " + }, + { + "bbox": [ + 313, + 353, + 555, + 413 + ], + "type": "inline_equation", + "content": "R_{e}^{\\prime}" + }, + { + "bbox": [ + 313, + 353, + 555, + 413 + ], + "type": "text", + "content": " corresponds to the region in the reference view. Additionally, this paired region undergoes multiview adaptive adjustments based on the error map throughout the optimization process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": "(ii) 2D-to-3D projection After obtaining the paired regions with render errors, we project each 2D error region to the 3D space via multiview geometry constraints. Specifically, we cast the rays " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": " in cone shape for region " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "R_{e}" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": " from the camera's center of projection " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": " along the direction " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": ", which aligns with the pixel's center (Figure 2(c)). The apex of this cone is located at " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": ", and its radius at the image plane. Hence, " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "o + d" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": " is parameterized as " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": ". The radius " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "r_{Cone}" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": " is set to match the radius of the smallest circumscribed circle of the 2D plane error region, creating a cone on the 3D space that can trace the Gaussian points contributing to the 2D error region. Concurrently, a corresponding cone, denoted as " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{C}'" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": ", belong to region " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "R_{e}'" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": " is similarly projected. Subsequently, we compute the intersection points of these rays. In order to regionalize these points, we directly use a smallest sphere that can contain these points as error source 3D zone " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "R_{zone}" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Points manipulation Recall that in existing 3DGS, points management only relies on the view-averaged gradient magnitude " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " to determine point densification globally. In addition to this, we further perform localized points addition and geometry calibration within the identified error source 3D zone " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "R_{zone}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ". For the point addition, we consider two common situations: (1) In the presence of points, we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21699" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 70, + 233, + 168 + ], + "blocks": [ + { + "bbox": [ + 66, + 114, + 80, + 128 + ], + "lines": [ + { + "bbox": [ + 66, + 114, + 80, + 128 + ], + "spans": [ + { + "bbox": [ + 66, + 114, + 80, + 128 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 70, + 233, + 168 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 233, + 168 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 233, + 168 + ], + "type": "image", + "image_path": "4af857b32b53adba7469c8f1470e353e1fb5c997334d26931316f39b6fc1785c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 85, + 168, + 233, + 266 + ], + "blocks": [ + { + "bbox": [ + 67, + 211, + 81, + 222 + ], + "lines": [ + { + "bbox": [ + 67, + 211, + 81, + 222 + ], + "spans": [ + { + "bbox": [ + 67, + 211, + 81, + 222 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 168, + 233, + 266 + ], + "lines": [ + { + "bbox": [ + 85, + 168, + 233, + 266 + ], + "spans": [ + { + "bbox": [ + 85, + 168, + 233, + 266 + ], + "type": "image", + "image_path": "d94b2985b4d391e94740e620d165fef7852b10109970ef4e06131bd27dedae9a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 268, + 186, + 277 + ], + "lines": [ + { + "bbox": [ + 131, + 268, + 186, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 268, + 186, + 277 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 238, + 70, + 386, + 168 + ], + "blocks": [ + { + "bbox": [ + 238, + 70, + 386, + 168 + ], + "lines": [ + { + "bbox": [ + 238, + 70, + 386, + 168 + ], + "spans": [ + { + "bbox": [ + 238, + 70, + 386, + 168 + ], + "type": "image", + "image_path": "39a166ef3023b1779dd82830719b2090548fb966c74816b1e0291e8c3191c176.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 238, + 168, + 386, + 266 + ], + "blocks": [ + { + "bbox": [ + 238, + 168, + 386, + 266 + ], + "lines": [ + { + "bbox": [ + 238, + 168, + 386, + 266 + ], + "spans": [ + { + "bbox": [ + 238, + 168, + 386, + 266 + ], + "type": "image", + "image_path": "8a2277b7bad2b600b665fa75d556bbb3b7814bd695285a8f152b07f84bb70f1e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 391, + 70, + 539, + 168 + ], + "blocks": [ + { + "bbox": [ + 391, + 70, + 539, + 168 + ], + "lines": [ + { + "bbox": [ + 391, + 70, + 539, + 168 + ], + "spans": [ + { + "bbox": [ + 391, + 70, + 539, + 168 + ], + "type": "image", + "image_path": "6f5182dc84b379d2b2d13d406611c6914b9f4d4bfe214fb532af9ae25b9c57b5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 391, + 168, + 539, + 266 + ], + "blocks": [ + { + "bbox": [ + 391, + 168, + 539, + 266 + ], + "lines": [ + { + "bbox": [ + 391, + 168, + 539, + 266 + ], + "spans": [ + { + "bbox": [ + 391, + 168, + 539, + 266 + ], + "type": "image", + "image_path": "28751b1d099292f1be1a556e31ce5afd7035f91224b4a8b7319ff432a13be817.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 85, + 278, + 233, + 377 + ], + "blocks": [ + { + "bbox": [ + 67, + 322, + 80, + 334 + ], + "lines": [ + { + "bbox": [ + 67, + 322, + 80, + 334 + ], + "spans": [ + { + "bbox": [ + 67, + 322, + 80, + 334 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 278, + 233, + 377 + ], + "lines": [ + { + "bbox": [ + 85, + 278, + 233, + 377 + ], + "spans": [ + { + "bbox": [ + 85, + 278, + 233, + 377 + ], + "type": "image", + "image_path": "318b43f2bd59ad2957c4378cd652cd4b7212512cb55f16f686f5d55624b33af1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 238, + 278, + 386, + 377 + ], + "blocks": [ + { + "bbox": [ + 257, + 268, + 365, + 277 + ], + "lines": [ + { + "bbox": [ + 257, + 268, + 365, + 277 + ], + "spans": [ + { + "bbox": [ + 257, + 268, + 365, + 277 + ], + "type": "text", + "content": "3DGS* rendered RGB image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 278, + 386, + 377 + ], + "lines": [ + { + "bbox": [ + 238, + 278, + 386, + 377 + ], + "spans": [ + { + "bbox": [ + 238, + 278, + 386, + 377 + ], + "type": "image", + "image_path": "e09305f5f120939ddfef0019aec4daff29795b1ef87b210dd7e8a9f4d318204f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 391, + 278, + 541, + 377 + ], + "blocks": [ + { + "bbox": [ + 405, + 268, + 533, + 277 + ], + "lines": [ + { + "bbox": [ + 405, + 268, + 533, + 277 + ], + "spans": [ + { + "bbox": [ + 405, + 268, + 533, + 277 + ], + "type": "text", + "content": "3DGS+LPM rendered RGB image" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 278, + 541, + 377 + ], + "lines": [ + { + "bbox": [ + 391, + 278, + 541, + 377 + ], + "spans": [ + { + "bbox": [ + 391, + 278, + 541, + 377 + ], + "type": "image", + "image_path": "568fe3af2967250ac1bbbcd6c81346af4e4373be85873542b6ad3ea6f6177d88.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 85, + 390, + 233, + 522 + ], + "blocks": [ + { + "bbox": [ + 131, + 380, + 186, + 389 + ], + "lines": [ + { + "bbox": [ + 131, + 380, + 186, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 186, + 389 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 456, + 81, + 468 + ], + "lines": [ + { + "bbox": [ + 67, + 456, + 81, + 468 + ], + "spans": [ + { + "bbox": [ + 67, + 456, + 81, + 468 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 390, + 233, + 522 + ], + "lines": [ + { + "bbox": [ + 85, + 390, + 233, + 522 + ], + "spans": [ + { + "bbox": [ + 85, + 390, + 233, + 522 + ], + "type": "image", + "image_path": "7048a4930fe1b81605dec1c41332856a55e74a6f94bddb1732a2ab3a41f59d89.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 523, + 197, + 532 + ], + "lines": [ + { + "bbox": [ + 117, + 523, + 197, + 532 + ], + "spans": [ + { + "bbox": [ + 117, + 523, + 197, + 532 + ], + "type": "text", + "content": "2DGS rendered mesh" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 536, + 555, + 570 + ], + "lines": [ + { + "bbox": [ + 55, + 536, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 555, + 570 + ], + "type": "text", + "content": "Figure 3. Qualitative evaluation of our LPM across diverse static datasets [4, 12]. Our LPM improves 2DGS [13] and 3DGS [14] on these challenging scenarios, e.g. (a) Light artifacts, (b) Completeness in the distance, (c) Depth structure and (d) Mesh details. See red patches for highlighted visual differences." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 238, + 390, + 386, + 522 + ], + "blocks": [ + { + "bbox": [ + 264, + 380, + 350, + 389 + ], + "lines": [ + { + "bbox": [ + 264, + 380, + 350, + 389 + ], + "spans": [ + { + "bbox": [ + 264, + 380, + 350, + 389 + ], + "type": "text", + "content": "3DGS* rendered depth" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 390, + 386, + 522 + ], + "lines": [ + { + "bbox": [ + 238, + 390, + 386, + 522 + ], + "spans": [ + { + "bbox": [ + 238, + 390, + 386, + 522 + ], + "type": "image", + "image_path": "db3c37c9cabbaa38cc5520bb1cca71eea192e8a11b6de1aed0171859cac26f5d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 523, + 358, + 532 + ], + "lines": [ + { + "bbox": [ + 255, + 523, + 358, + 532 + ], + "spans": [ + { + "bbox": [ + 255, + 523, + 358, + 532 + ], + "type": "text", + "content": "2DGS+LPM rendered mesh" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 391, + 390, + 539, + 522 + ], + "blocks": [ + { + "bbox": [ + 409, + 380, + 515, + 389 + ], + "lines": [ + { + "bbox": [ + 409, + 380, + 515, + 389 + ], + "spans": [ + { + "bbox": [ + 409, + 380, + 515, + 389 + ], + "type": "text", + "content": "3DGS+LPM rendered depth" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 390, + 539, + 522 + ], + "lines": [ + { + "bbox": [ + 391, + 390, + 539, + 522 + ], + "spans": [ + { + "bbox": [ + 391, + 390, + 539, + 522 + ], + "type": "image", + "image_path": "00cac4e6b2ed17da7009a204540d579cd4991d7bd6448a598ddd4ca5151abfb7.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 523, + 524, + 533 + ], + "lines": [ + { + "bbox": [ + 409, + 523, + 524, + 533 + ], + "spans": [ + { + "bbox": [ + 409, + 523, + 524, + 533 + ], + "type": "text", + "content": "2DGS vs 2DGS+LPM heatmap" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 578, + 297, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 297, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 297, + 710 + ], + "type": "text", + "content": "apply point densification to locally complement the original counterparts. We set a lower threshold to select the points that need densification, aiming to enhance the geometric details. The densification rule is consistent with 3DGS, but it focuses on local 3D zones that need it most. Specifically, for small Gaussians, our strategy involves cloning the Gaussians while maintaining their size and repositioning them along the positional gradient to better capture emerging geometrical features. Conversely, larger Gaussians situated in areas of high variance are split into smaller points to more accurately represent the underlying geometry. (2) In cases" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 578, + 556, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 556, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 556, + 711 + ], + "type": "text", + "content": "of point sparsity, we add new Gaussian points at the center of the 3D zone. In the context of " + }, + { + "bbox": [ + 313, + 578, + 556, + 711 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 578, + 556, + 711 + ], + "type": "text", + "content": "-blending in 3DGS, if the points at the forefront of the identified 3D zone " + }, + { + "bbox": [ + 313, + 578, + 556, + 711 + ], + "type": "inline_equation", + "content": "R_{zone}" + }, + { + "bbox": [ + 313, + 578, + 556, + 711 + ], + "type": "text", + "content": " have the highest opacity, they may occlude valid points, leading to incorrect depth estimation, as shown in Figure 1. To deal with such issues, we treat these points as potentially ill-conditioned. We reset these points to provide an opportunity for correction, further calibrating the geometry. To minimize model expansion, we adaptively prune points based on their opacity values, starting from low to high. The number of points pruned is determined by the density of" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21700" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 70, + 504, + 256 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 504, + 256 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 504, + 256 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 504, + 256 + ], + "type": "table", + "html": "
MethodMip-NeRF 360Tanks&TemplesDeep Blending
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
Plenoxels23.080.6250.46321.080.7190.37923.060.7950.510
INGP-Big25.590.6990.33121.920.7450.30524.960.8170.390
Mip-NeRF 36027.690.7920.23722.220.7590.25729.400.9010.245
3DGS27.210.8150.21423.140.8410.18329.410.9030.243
3DGS*27.470.8160.21623.670.8490.17729.550.9040.245
3DGS* + LPM27.590.8200.21623.830.8500.18129.760.9080.241
2DGS*27.150.8080.24623.580.8320.18529.350.8990.262
2DGS* + LPM27.420.8170.22823.650.8480.18029.520.9030.240
MipGS*27.510.8170.21023.690.8520.17329.580.9100.242
MipGS* + LPM27.700.8210.21023.820.8510.18029.610.9100.241
PixelGS*27.540.8190.20323.750.8500.17529.580.9200.220
PixelGS* + LPM27.800.8300.19024.020.8560.17329.650.9100.196
", + "image_path": "216ba789a3565b5564f2aebe6447e2e64a055982ccaf88aad246368e9bccbd8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 264, + 553, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 264, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 553, + 288 + ], + "type": "text", + "content": "Table 1. Comparison of various methods across different scenes on the Mip-NeRF 360 dataset, Tanks&Temples and Deep Blending. * indicates the retrained model from the official implementation. Bold represents best, underline indicates second best." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 301, + 295, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 338 + ], + "type": "text", + "content": "points in the zone. This strategic reduction ensures that our point management remains cost efficient and adaptive to the evolving needs of the scene representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 132, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 132, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 132, + 355 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 360, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 467 + ], + "type": "text", + "content": "Datasets and metrics We conducted an extensive evaluation using both static and dynamic scenes derived from publicly datasets. For static scenes, our approach was applied to a total of 13 scenes as specified in the 3DGS framework [14], which includes nine scenes from Mip-NeRF360 [3], two from Tanks&Temples [15], and two from DeepBlending [12]. In the context of dynamic scenes, our approach was tested across six scenes from the Neural 3D Video Dataset [17]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 588 + ], + "type": "text", + "content": "To evaluate novel view synthesis performance, we followed standard protocols by selecting one out of every eight images as test images, with the remaining used for training in static scenes. For each dynamic scene within the Neural 3D Video Dataset, one view was designated for testing while the others were allocated for training purposes. Evaluation metrics included the peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and the learned perceptual image patch similarity (LPIPS), which are broadly recognized standards in the field." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "content": "Baselines and implementation Vanilla 3D Gaussian Splitting (3DGS) [14], 2D Gaussian Splitting (2DGS) [13], Mip Gaussian Splitting (MipGS) [40], PiexlGS [41] and SpacetimeGS (STGS) [18] were selected as our main baselines for their established art performance in novel view synthesis. For the static 3D benchmark, we also recorded the results of Mip-NeRF360 [3], iNGP [24] and Plenoxels [10] as in [14]. For the Dynamic 4D benchmark, we performed system comparison, such as DyNeRF [16], K-planes [11] and so on. In alignment with the approach" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 301, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 434 + ], + "type": "text", + "content": "described in 3DGS an STGS, our models were trained for 30k iterations across all scenes, following the same training schedule and hyperparameters. In addition to the original Gaussian densification strategies used in 3DGS and SpaceTime Gaussian, we also performed localized points management, including addition, reset, and pruning. We maintained the same thresholds for splitting and cloning points as in the original 3DGS and SpaceTime Gaussian. For point matching, we performed offline extraction to save computational cost. All experiments were conducted on an RTX 3090 GPU with 24GB of memory." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 443, + 400, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 400, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 400, + 455 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 461, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 640 + ], + "type": "text", + "content": "Results on static 3D datasets The quantitative results (PSNR, SSIM, and LPIPS) on the Mip-NeRF 360 and Tanks & Temples datasets are presented in Tables 1. We retrained the 3DGS model (referred to as 3DGS*) as it yields better performance compared to the vanilla 3DGS and its variants. Our approach achieves results comparable to the state-of-the-art on the Mip-NeRF360 dataset and further enhances all 3DGS based method using our point management technique. Additionally, LPM improve vanilla 3DGS and PixlGS to set new state-of-the-art results on the Mip-NeRF 360, Tanks & Temples datasets and Dep Blending, effectively capturing more challenging environments (e.g., light effects, transparency). These results quantitatively validate the effectiveness of our method in improving the quality of reconstruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "In Figures 3, we present a comparison between 3DGS [14] and " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^* +" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " LPM, focusing on both appearance and depth. A variety of improvements can be observed, particularly in challenging cases such as light effects, completeness at a distance. Our LPM significantly reduces artifacts in specific regions on top of 3DGS, partic" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "21701" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 156, + 183 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 156, + 183 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 156, + 183 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 156, + 183 + ], + "type": "image", + "image_path": "906994016d490f4e7767cd3908d7a7e39402ad37d726ac55740116d093f88635.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 186, + 118, + 196 + ], + "lines": [ + { + "bbox": [ + 94, + 186, + 118, + 196 + ], + "spans": [ + { + "bbox": [ + 94, + 186, + 118, + 196 + ], + "type": "text", + "content": "3DGS" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 157, + 72, + 254, + 183 + ], + "blocks": [ + { + "bbox": [ + 157, + 72, + 254, + 183 + ], + "lines": [ + { + "bbox": [ + 157, + 72, + 254, + 183 + ], + "spans": [ + { + "bbox": [ + 157, + 72, + 254, + 183 + ], + "type": "image", + "image_path": "15864ed2fb62f3531cd8c07e815f89652ef6a5b1d192d42133f9d126136fd718.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 186, + 251, + 196 + ], + "lines": [ + { + "bbox": [ + 162, + 186, + 251, + 196 + ], + "spans": [ + { + "bbox": [ + 162, + 186, + 251, + 196 + ], + "type": "text", + "content": "3DGS (lower threshold)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 72, + 353, + 183 + ], + "blocks": [ + { + "bbox": [ + 255, + 72, + 353, + 183 + ], + "lines": [ + { + "bbox": [ + 255, + 72, + 353, + 183 + ], + "spans": [ + { + "bbox": [ + 255, + 72, + 353, + 183 + ], + "type": "image", + "image_path": "d6802d13c8f1d0b05111bd681b98c0a797428013f63f9cee0b59ee15ad59349e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 186, + 350, + 196 + ], + "lines": [ + { + "bbox": [ + 260, + 186, + 350, + 196 + ], + "spans": [ + { + "bbox": [ + 260, + 186, + 350, + 196 + ], + "type": "text", + "content": "LPM wo/ point addition" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 353, + 72, + 451, + 183 + ], + "blocks": [ + { + "bbox": [ + 353, + 72, + 451, + 183 + ], + "lines": [ + { + "bbox": [ + 353, + 72, + 451, + 183 + ], + "spans": [ + { + "bbox": [ + 353, + 72, + 451, + 183 + ], + "type": "image", + "image_path": "a39fbabb937bf7897895c2b4a2554039b50e6886337b0c3e4e9cc82c9571bc75.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 186, + 429, + 196 + ], + "lines": [ + { + "bbox": [ + 373, + 186, + 429, + 196 + ], + "spans": [ + { + "bbox": [ + 373, + 186, + 429, + 196 + ], + "type": "text", + "content": "LPM wo/ reset" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 452, + 72, + 551, + 183 + ], + "blocks": [ + { + "bbox": [ + 452, + 72, + 551, + 183 + ], + "lines": [ + { + "bbox": [ + 452, + 72, + 551, + 183 + ], + "spans": [ + { + "bbox": [ + 452, + 72, + 551, + 183 + ], + "type": "image", + "image_path": "213aae9e7575bcf3ab79ed9b2fe4e5ae3f4381b0ba2ce104c93d315888ab62df.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 186, + 526, + 195 + ], + "lines": [ + { + "bbox": [ + 478, + 186, + 526, + 195 + ], + "spans": [ + { + "bbox": [ + 478, + 186, + 526, + 195 + ], + "type": "text", + "content": "3DGS+LPM" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 64, + 235, + 289, + 480 + ], + "blocks": [ + { + "bbox": [ + 64, + 235, + 289, + 480 + ], + "lines": [ + { + "bbox": [ + 64, + 235, + 289, + 480 + ], + "spans": [ + { + "bbox": [ + 64, + 235, + 289, + 480 + ], + "type": "table", + "html": "
MethodPSNRDSSIMLPIPSFPS
\\( LLFF^1 \\)23.240.0760.235-
\\( DyNeRF^1 \\)29.580.0830.0630.015
\\( Dynamic-4DGS^1 \\)---30
\\( 4DGS^1 \\)29.38--114
\\( STGS^1 \\)29.580.0220.063103
\\( STGS*^1 \\)29.480.0230.066110
\\( STGS*^1 + LPM \\)29.840.0220.062105
StreamRF28.26-0.03910.9
NeRFPlayer30.69-0.1110.05
HyperReal31.10-0.0962
K-planes31.63-0.313
Mix Voxels-X31.73-0.0644.6
Dynamic-4DGS31.150.0160.04930
4DGS32.01-0.055114
STGS32.050.0140.044140
STGS*31.990.0150.045145
\\( STGS*+ LPM \\)32.400.0140.045140
", + "image_path": "1c32766e77cc422b33f2be4cd06e22beb5d0c576bdf69db088ad75ff939c836a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 489, + 296, + 555 + ], + "lines": [ + { + "bbox": [ + 55, + 489, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 296, + 555 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on the Neural 3D Video dataset. \"FPS\" is measured at a resolution of " + }, + { + "bbox": [ + 55, + 489, + 296, + 555 + ], + "type": "inline_equation", + "content": "1352 \\times 1014" + }, + { + "bbox": [ + 55, + 489, + 296, + 555 + ], + "type": "text", + "content": ". Some methods only report results for a subset of scenes. For a fair comparison, we report LPM's results under two pre-existing settings. Only includes the Flame Salmon scene. Bold represents best, underline indicates second best." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 576, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 684 + ], + "type": "text", + "content": "cularly in the tree at the second. These regions require more points for accurate population, leading to a more precise and detailed reconstruction. Additionally, the tablecloth in the first row is affected by ill-conditioned points. Furthermore, we provide depth and mesh comparisons in the third and final rows. All these observations demonstrate that our geometry calibration with LPM offers an opportunity to correct these potentially ill-conditioned points, thereby enhancing the overall reconstruction accuracy." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Results on dynamic 4D datasets Table 2 presents a quantitative evaluation on the Neural 3D Video Dataset." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 236, + 553, + 361 + ], + "blocks": [ + { + "bbox": [ + 55, + 203, + 555, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 203, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 555, + 226 + ], + "type": "text", + "content": "Figure 4. Effect of key operations of LPM. We show that the point addition operation effectively captures the geometric details in the scene; The point reset operation based on the error map further calibrate the geometry." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 236, + 553, + 361 + ], + "lines": [ + { + "bbox": [ + 316, + 236, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 236, + 553, + 361 + ], + "type": "image", + "image_path": "6578bb930cdfefdec14552fc7a5404754bf324841aaa554d50a034e7fec3d307.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 371, + 555, + 405 + ], + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 405 + ], + "type": "text", + "content": "Figure 5. Qualitative evaluation on dynamic Neural 3D Video dataset [17]. LPM improves STGS [18] for both scenes Transparent (e.g., window) and Dynamic movements (e.g., dog's tongue)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 409, + 554, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 554, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 554, + 517 + ], + "type": "text", + "content": "Following established practices, training and evaluation are conducted at half resolution, with the first camera held out for evaluation [16]. Integrating our LPM into SpaceTimeGS yields the best performance across all comparisons. Notably, our method demonstrates significant improvements in the challenging Flame Salmon scene compared to SpaceTimeGS [18]. Our approach not only surpasses previous methods in rendering quality but also maintains comparable rendering speed." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 518, + 555, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 625 + ], + "type": "text", + "content": "In addition to the quantitative assessment, we provide qualitative comparisons on the Flame Salmon and Flame Steak scenes, as illustrated in Figure 5. The quality of synthesis in both static and dynamic regions markedly outperforms STGS. Several intricate details, including the tree behind the window and the fine features like the dog's tongue, are faithfully reproduced with higher accuracy compared to STGS [18]. Both examples indicate that LPM improves upon STGS for superior scene modeling." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 632, + 406, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 406, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 406, + 645 + ], + "type": "text", + "content": "4.2. Ablation study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 647, + 554, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 554, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 554, + 683 + ], + "type": "text", + "content": "We conducted ablation studies on the more challenging scene: PlayRoom from Deep Blending [12] and Truck from Tanks&Temples [15]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Effectiveness and cost of LPM We hypothesize that the Adaptive Density Control (ADC) tends to overlook under" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21702" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 155, + 70, + 452, + 148 + ], + "blocks": [ + { + "bbox": [ + 155, + 70, + 452, + 148 + ], + "lines": [ + { + "bbox": [ + 155, + 70, + 452, + 148 + ], + "spans": [ + { + "bbox": [ + 155, + 70, + 452, + 148 + ], + "type": "table", + "html": "
MethodPlayRoomTruck
PSNRLPIPSSSIMPSNRLPIPSSSIM
Full LPM30.220.2410.91025.610.1540.883
wo/ point addition30.100.2410.91025.430.1530.883
wo/ reset30.070.2430.90825.520.1440.883
", + "image_path": "e0860f673283b5975e1f7e1b81999c598ec32b2e99413a9cee2c521d8d283bfc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 106, + 170, + 502, + 320 + ], + "blocks": [ + { + "bbox": [ + 192, + 150, + 416, + 162 + ], + "lines": [ + { + "bbox": [ + 192, + 150, + 416, + 162 + ], + "spans": [ + { + "bbox": [ + 192, + 150, + 416, + 162 + ], + "type": "text", + "content": "Table 3. Performance comparison for different configurations" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 106, + 170, + 502, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 170, + 502, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 502, + 320 + ], + "type": "table", + "html": "
SceneMethodPSNRLPIPSGaussiansTraining time
PlayRoom3DGS*30.030.244232k22min
3DG* (lower threshold)29.690.240523k36min
GaussianProOut of Memory
PielGS30.090.241186k35min
3DGS + LPM30.220.241186k23min
Truck3DGS*25.420.146257k19 min
3DGS* (lower threshold)25.450.127635k35min
GaussianPro25.400.164312k36min
PielGS25.510.121518k37min
3DGS + LPM25.610.154265k21min
", + "image_path": "c26324be8a9df9d31128e6b9a7c70e83699c56daca28b1d2eb8fa29b4fdca291.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 323, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 555, + 346 + ], + "type": "text", + "content": "Table 4. Cost-effectiveness analysis. Rendering speed of both methods are measured on our machine. Note: For 3DGS+LPM, training time includes the feature matching process." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 365, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 674 + ], + "type": "text", + "content": "optimized points due to its simplistic approach of thresholding the average gradient. The straight way to identify the all points is lowering threshold to densification process. Although this solution can reduce blurring in specific regions, such as the toy (red box) illustrated in Figure 4, it still has limitations. As shown in Table 4, lowering the threshold for 3DGS significantly increases the number of Gaussian points and decreases rendering speed. Additionally, the PSNR of the quantitative results decreases due to the introduction of unnecessary points in already dense areas. In contrast, LPM effectively generates points in areas indicated by the error map, leading to more accurate and detailed reconstructions while maintaining real-time rendering speed. As demonstrated by the qualitative comparison in Figure 4, 3DGS with LPM achieves superior qualitative results. We further compare our method with other recent methods that also focus on adaptive density control (ADC). While PixelGS and GaussianPro achieve improvements in rendering quality, their training times increase substantially as they only consider point addition and extra gradient propagation. In contrast, LPM achieves a noticeable improvement with only a slight increase in training time due to (1) point matching [19] is much faster (2) considering model expansion to dynamically prune the points by their addition number and (3) selecting points in error-contributing zones 3D zone using the parallel matrix operations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Individual points manipulation We study the effect of individual points manipulation of LPM, including the point addition and reset ill-conditional points. The results in Ta" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 364, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 461 + ], + "type": "text", + "content": "ble 3 show that, (1) each manipulation is useful with positive gain, suggesting that the LPM is meaningful. (2) The point addition operation densify the under-optimized points which may be overlook in the 3DGS, further captures the geometry details (e.g., detail of toy and leaf of the tree, see Fig. 4). (3) Reset points in ceratin zone provide the opportunity of correct the ill-conditioned points to achieve geometry calibration, (e.g., window of the trunk, see Fig. 4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 471, + 388, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 388, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 388, + 483 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 490, + 555, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 555, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 555, + 706 + ], + "type": "text", + "content": "We propose Localized Point Management (LPM), a novel point management approach to address the limitations of the Adaptive Density Control (ADC) mechanism in 3D Gaussian Splatting (3DGS). The core idea of LPM is identifying the error-contributing 3D zones that require both point addition and geometry calibration under multiview geometry constraints, guided by image rendering errors. We implement appropriate operations for point densification and opacity reset. As a versatile plugin, LPM can be seamlessly integrated into existing 3DGS-based rendering methods. Extensive experiments across both static 3D and dynamic 4D scenes validate the efficacy of LPM in enhancing existing ADC mechanisms both quantitatively and qualitatively. While our method identifies the 3D Gaussian points that lead to rendering errors, it still follows the densification rules of 3DGS [14]. This approach may not be optimal for under-optimized points, and we leave this aspect for further investigation." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21703" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 145 + ], + "type": "text", + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 294, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 294, + 191 + ], + "type": "text", + "content": "[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 258 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2, 6, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 316 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 318, + 294, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 294, + 372 + ], + "type": "text", + "content": "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697–19705, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 374, + 294, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 294, + 416 + ], + "type": "text", + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "type": "text", + "content": "[7] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 465, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 294, + 518 + ], + "type": "text", + "content": "[8] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 521, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 294, + 576 + ], + "type": "text", + "content": "[9] Nianchen Deng, Zhenyi He, Jiannan Ye, Budmonde Duinkharjav, Praneeth Chakravarthula, Xubo Yang, and Qi Sun. Fov-nerf: Foveated neural radiance fields for virtual reality. IEEE Transactions on Visualization and Computer Graphics, 28(11):3854-3864, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "type": "text", + "content": "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "type": "text", + "content": "[11] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12479–12488, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 5, 6, 7, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 98, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 98, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 98, + 553, + 140 + ], + "type": "text", + "content": "[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "type": "text", + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 2, 5, 6, 8, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 190, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 190, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 190, + 553, + 232 + ], + "type": "text", + "content": "[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 6, 7, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 235, + 553, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 235, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 317, + 235, + 553, + 311 + ], + "type": "text", + "content": "[16] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multiview video. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5511-5521, 2022. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 315, + 553, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 315, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 317, + 315, + 553, + 389 + ], + "type": "text", + "content": "[17] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 393, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 393, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 393, + 553, + 425 + ], + "type": "text", + "content": "[18] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. arXiv preprint arXiv:2312.16812, 2023. 3, 4, 6, 7, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 428, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 428, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 317, + 428, + 553, + 472 + ], + "type": "text", + "content": "[19] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627-17638, 2023. 4, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 475, + 553, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 475, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 317, + 475, + 553, + 517 + ], + "type": "text", + "content": "[20] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 520, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 520, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 317, + 520, + 553, + 564 + ], + "type": "text", + "content": "[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 567, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 553, + 609 + ], + "type": "text", + "content": "[22] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 613, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 666 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 3, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21704" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[25] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 140 + ], + "type": "text", + "content": "[26] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. arXiv e-prints, pages arXiv-2404, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 143, + 295, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 143, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 57, + 143, + 295, + 198 + ], + "type": "text", + "content": "[27] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 200, + 294, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 200, + 294, + 233 + ], + "spans": [ + { + "bbox": [ + 57, + 200, + 294, + 233 + ], + "type": "text", + "content": "[28] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 235, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 235, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 57, + 235, + 294, + 300 + ], + "type": "text", + "content": "[29] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 303, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 303, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 57, + 303, + 294, + 345 + ], + "type": "text", + "content": "[30] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 349, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 349, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 57, + 349, + 294, + 392 + ], + "type": "text", + "content": "[31] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 396, + 295, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 396, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 57, + 396, + 295, + 449 + ], + "type": "text", + "content": "[32] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 453, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 453, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 57, + 453, + 294, + 517 + ], + "type": "text", + "content": "[33] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4150–4159, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 520, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 520, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 57, + 520, + 294, + 564 + ], + "type": "text", + "content": "[34] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 567, + 295, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 620 + ], + "type": "text", + "content": "[35] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5438-5448, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "type": "text", + "content": "[36] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. arXiv preprint arXiv:2311.17089, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "text", + "content": "[37] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 285 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[38] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In ICLR, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "text", + "content": "[39] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "text", + "content": "[40] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. arXiv preprint arXiv:2311.16493, 2023. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "type": "text", + "content": "[41] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "text", + "content": "[42] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21705" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/673cbe39-bd0b-47dd-99b2-3c870c8defd5_content_list.json b/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/673cbe39-bd0b-47dd-99b2-3c870c8defd5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b7dec5732de209932fcd27ad1dd793f697df8c4d --- /dev/null +++ b/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/673cbe39-bd0b-47dd-99b2-3c870c8defd5_content_list.json @@ -0,0 +1,1354 @@ +[ + { + "type": "text", + "text": "Improving Personalized Search with Regularized Low-Rank Parameter Updates", + "text_level": 1, + "bbox": [ + 94, + 130, + 903, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fiona Ryan $^{1,2*}$ , Josef Sivic $^{2,3}$ , Fabian Caba Heilbron $^{2}$ , Judy Hoffman $^{1}$ , James M. Rehg $^{4}$ , Bryan Russell $^{2}$ $^{1}$ Georgia Tech, $^{2}$ Adobe Research, $^{3}$ CIIRC CTU, $^{4}$ UIUC", + "bbox": [ + 99, + 179, + 919, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 266, + 326, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Personalized vision-language retrieval seeks to recognize new concepts (e.g., \"my dog Fido\") from only a few examples. This task is challenging because it requires not only learning a new concept from a few images, but also integrating the personal and general knowledge together to recognize the concept in different contexts. In this paper, we show how to effectively adapt the internal representation of a vision-language dual encoder model for personalized vision-language retrieval. We find that regularized low-rank adaption of a small set of parameters in the language encoder's final layer serves as a highly effective alternative to textual inversion for recognizing the personal concept while preserving general knowledge. Additionally, we explore strategies for combining parameters of multiple learned personal concepts, finding that parameter addition is effective. To evaluate how well general knowledge is preserved in a finetuned representation, we introduce a metric that measures image retrieval accuracy based on captions generated by a vision language model (VLM). Our approach achieves state-of-the-art accuracy on two benchmarks for personalized image retrieval with natural language queries - DeepFashion2 and ConCon-Chi - outperforming the prior art by $4\\% - 22\\%$ on personal retrievals.", + "bbox": [ + 88, + 297, + 485, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 666, + 220, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Personalizing a vision-language retrieval model (PerVL) aims to adapt a pretrained vision-language dual encoder model (e.g., CLIP [27]) to recognize new concepts (e.g., \"my dog Fido\") from just a few examples [7]. This task is important for search applications that need to identify concepts missing from the pretrained model's knowledge, such as searching one's personal photo library for a specific person, object, or pet. PerVL is challenging because it requires not only learning a new concept from a few visual examples, but also reasoning about the personal concept and general knowledge together to retrieve the concept in different contexts. For instance, searching for \"my dog Fido catching", + "bbox": [ + 89, + 693, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b780b0f30021761ab92fc36ae0d2961150b87d45d007e156646ec29100987976.jpg", + "image_caption": [ + "Figure 1. Left: Prior works use a pretrained textual inversion network $(\\mathcal{F})$ to compute a pseudo-token to represent a new concept, which may be further optimized during personalization. Right: We present POLAR, which represents new concepts as a small low-rank parameter update within the text encoder. Instead of inserting a learned pseudo-token to the text query, we use a fixed vocabulary token $V^{*}$ . We show that our method is effective at recognizing the personal concept from a few examples while retaining the model's general knowledge, and does not require large scale pretraining." + ], + "image_footnote": [], + "bbox": [ + 521, + 252, + 910, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a frisbee\" requires both personal knowledge (\"Fido\") and general knowledge (\"catching a frisbee\").", + "bbox": [ + 511, + 611, + 906, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent approaches for PerVL use textual inversion to learn a pseudo-text token (e.g., “ $V^{*}$ ”) for each new personal concept by inverting from the target training images to the token [4, 7] (illustrated in Figure 1 (left)). These pseudo-text tokens require updating only a minimal set of parameters, and they can be used within a natural language query as input to the language encoder (e.g., “ $V^{*}$ catching a frisbee”). Typically, these approaches train a textual inversion network on large-scale data to predict a token representing a new concept, which may then be further optimized during personalization. While these approaches avoid updating the model's internal parameters, preventing the “overriding” of its general knowledge, their ability to represent the personal concept is limited to the single input text token. Furthermore, the token affects the entire text embedding process and can interfere with the language encoder's general knowledge. Consequently, textual inversion", + "bbox": [ + 509, + 643, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work partially done during internship at Adobe Research.", + "bbox": [ + 114, + 886, + 426, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19748", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "approaches often struggle to combine personalized and general knowledge across different datasets and can be time-consuming to optimize online due to the need for backpropagation through the full language encoder [7, 38], or require large-scale pretraining [4, 7, 38].", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these shortcomings, we focus on updating the internal representation of CLIP's language encoder [27] using only a few training examples (illustrated in Figure 1 (right)). Updating the internal representation is challenging because the model must balance learning the personalized concept from limited examples while preventing catastrophic forgetting of its prior knowledge. Additionally, the training data does not inherently encourage the model to retain general knowledge. Our goal is to make this update efficiently, without relying on additional training data.", + "bbox": [ + 88, + 167, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are fourfold. First, inspired by recent advances in text-to-image generation [14, 19, 29, 30], we show that CLIP's language encoder can be updated to learn a personalized concept from a few examples while retaining its general knowledge. We find that regularized low-rank adaptation (LoRA) finetuning [16] of the language encoder's last layer effectively balances the trade-off between learning the personalized concept and avoiding catastrophic forgetting. Our resulting method POLAR (PersOnalized Low-rank Adaptation for Retrieval) learns a low-rank parameter set that is minimal and comparable in size to the pseudo-text tokens used in textual inversion. Furthermore, we leverage LoRA's specific structure to introduce a regularization strategy that eliminates the need for additional training examples or regularization prompts. Unlike previous methods, our approach does not rely on any components pre-trained on large-scale data, allowing for seamless parameter updates using only a few training examples.", + "bbox": [ + 91, + 318, + 482, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, we introduce a new evaluation metric to assess how well general knowledge is preserved in our finetuned representation, using captions generated by a vision-language model (VLM). We find that our approach effectively maintains general knowledge. Third, we explore different strategies for combining learned representations for different personal concepts to support multi-concept queries. We find that adding LoRA representations is effective and outperforms orthogonal adaptation [26]. Finally, we demonstrate that POLAR achieves state-of-the-art accuracy on the DeepFashion2 [7] and ConCon-Chi [28] benchmarks, improving prior performance by $4\\% - 22\\%$ .", + "bbox": [ + 89, + 590, + 482, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 784, + 232, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Personalized Vision-Language Retrieval. Cohen et al. introduced the task of Personalized Vision-Language Retrieval [7] and proposed PALAVRA, a textual inversion approach for the task. PALAVRA first learns a textual inversion network on COCO, which takes a set of images of a concept and predicts an initial pseudo-word token to rep", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "resent the concept in text queries. This token is then further optimized via backpropagation through CLIP's text encoder. Korbar et al. [18] propose a similar approach for the setting of retrieving specific people in videos. A few recent works explore learning embeddings to represent personal concepts for VLM captioning and QA tasks [1, 25]. Yeh et al. [38] build on the concept of textual inversion by meta-learning a basis for pseudo-tokens using large scale video data. However, they target a different setting where concepts are learned jointly, using the other concepts in the dataset as negative examples. Recently, Rosasco et al. introduced the Concept-Context Chimera dataset (ConConChi) for personalized retrieval, which provides a more rigorous benchmark for assessing retrieval of personal concepts across diverse contexts than prior datasets. In this work, we depart from prior work by representing new concepts as low rank parameter updates within the text encoder instead of pseudo-word tokens. We show that our approach more effectively composes personal and general knowledge, while requiring few parameters per-concept.", + "bbox": [ + 511, + 90, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Personalized Generation. Personalized generation is a more studied related task that generates new images of a personal concept using text-to-image diffusion models. Some approaches use textual inversion to optimize pseudoword tokens to use within text prompts [9, 10], while others like Dreambooth [29] and Custom Diffusion [19] find that tuning the weights of the diffusion U-net generates personal concepts with better fidelity. Recent work has focused on selectively tuning certain parameters to promote parameter efficiency and speed [9, 15, 19, 40], with some leveraging low rank constraints [14, 26, 30, 33]. Most related to our work is Perfusion [33], which learns rank-one updates to the diffusion U-net with a key-locking mechanism to constrain updates to the concept's spatial location in the feature map. While personalization via parameter updates has become commonplace for personalized generation, it has not yet been explored for personalized retrieval, motivating our work. Due to the differences in the nature of the tasks (generative vs. discriminative) and models (text-to-image diffusion vs. dual encoder), we find that retrieval demands a different strategy for applying internal parameter updates; updating even sparse sets of parameters can result in catastrophic forgetting of the model's general knowledge. Instead of updating parameters throughout the full model, we apply a single rank-one parameter update to the final layer of CLIP's text encoder and directly regularize these parameters to avoid catastrophic forgetting of general knowledge.", + "bbox": [ + 511, + 397, + 903, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Composed Image Retrieval. Another related task is composed image retrieval [35], which takes an image and textual modification as inputs and performs image retrieval. Approaches have leveraged CLIP for composed image retrieval [2, 3], with Pic2Word [31] and SEARLE [4] learning textual inversion networks to predict a pseudo-word token", + "bbox": [ + 511, + 809, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19749", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/af97af57bc75b134048dfd51e46fb3793fda95a4869c19f0fb143ae523521a6e.jpg", + "image_caption": [ + "Figure 2. Overview of POLAR. For a personal concept, POLAR learns a rank-1 update to the value transform in the final layer of the text encoder. To maintain general knowledge during personalization, we impose a regularization loss on the update." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 477, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for the input image. Importantly, composed image retrieval differs from personalized retrieval in that it does not require instance-level recognition of the same concept (e.g., retrieving the exact same person), and instead aims to retrieve images portraying a similar semantic class, layout, or style.", + "bbox": [ + 89, + 362, + 483, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-Shot Adaptation of Vision-Language Models. Our task also relates to work on adapting models like CLIP for few-shot classification. Prior works primarily leverage prompt tuning [21], which learns input tokens to represent new classes [5, 6, 8, 23, 34, 36, 37, 42-44]. A few works consider tuning within the encoder by applying prompts to all layers [17, 20] or augmenting the encoder with adapter modules [12, 41]. An important distinction is that while CLIP's general knowledge may aid with learning classes from few examples, the general knowledge is not required for these downstream classification tasks. In contrast, personalized retrieval requires not just recognizing the personal concept, but composing it with retained general knowledge.", + "bbox": [ + 89, + 439, + 483, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Personalized Low-Rank Adaptation for Retrieval (POLAR)", + "text_level": 1, + "bbox": [ + 89, + 652, + 482, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 696, + 294, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow the PerVL problem setup established by Cohen et al. [7]. Given a pretrained vision-language model $\\psi$ , we aim to learn an adapted model $\\psi'$ that is able to recognize a new personal concept $c$ (e.g., \"my coffee mug\"). The pretrained model $\\psi$ consists of an image encoder $\\psi_I$ and a text encoder $\\psi_T$ . These encoders map image and textual inputs to a shared embedding space. At personalization time, we are given $N_c$ images $\\{I_i^c\\}_{i=1}^{N_c}$ depicting concept $c$ as training data. Following Cohen et al. [7], we are also given the class name $\\mathcal{C}_c$ of concept $c$ (e.g., \"mug\"). At retrieval time, the input to the model consists of a textual query $q$ and a set of $N_r$ images $\\{I_i^r\\}_{i=1}^{N_r}$ constituting the retrieval database. We", + "bbox": [ + 89, + 719, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "compute the language embedding for an input query $q$ as $\\psi_T'(q)$ , and perform retrieval by computing the cosine similarity between $\\psi_T'(q)$ and the image embedding for each retrieval image as", + "bbox": [ + 511, + 92, + 905, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s i m} \\left(\\psi_ {T} ^ {\\prime} (q), \\psi_ {I} ^ {\\prime} \\left(I _ {i} ^ {r}\\right)\\right) = \\frac {\\left\\langle \\psi_ {T} ^ {\\prime} (q) , \\psi_ {I} ^ {\\prime} \\left(I _ {i} ^ {r}\\right) \\right\\rangle}{\\left| \\left| \\psi_ {T} ^ {\\prime} (q) \\right| \\right| _ {2} \\left| \\left| \\psi_ {I} ^ {\\prime} \\left(I _ {i} ^ {r}\\right) \\right| \\right| _ {2}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 157, + 906, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\langle \\cdot \\rangle$ denotes the inner product. The top retrieval for the query is the image $I_{i}^{r}$ corresponding to the embedding with the highest similarity to the query embedding $\\psi_T^\\prime (q)$ .", + "bbox": [ + 511, + 198, + 905, + 246 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Rank-One Personalized Value Updates", + "text_level": 1, + "bbox": [ + 511, + 252, + 848, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to prior work that represents a concept as a learned input token for the text encoder [4, 7, 38], POLAR learns a low-rank parameter update to the text encoder for each concept (Fig. 2). We leverage LoRA [16]: for a weight $W \\in \\mathbb{R}^{m \\times n}$ in a pretrained model, LoRA learns a low-rank update $\\Delta W$ , performing a modified forward pass as", + "bbox": [ + 511, + 273, + 905, + 364 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny = (W + \\Delta W) x = W x + B A x \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 372, + 906, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B \\in \\mathbb{R}^{m \\times r}$ , $A \\in \\mathbb{R}^{r \\times n}$ , and $r < \\min(m, n)$ is the chosen rank of the weight update. By selecting $r << \\min(m, n)$ , LoRA tunes minimal parameters in the original model $\\psi$ . For each concept $c$ , we empirically choose to learn a rank-one ( $r = 1$ ) update to the value transform of the final attention layer $L$ in $\\psi_T$ . Our choice of $r = 1$ reflects our goal to represent a single concept from very limited examples, while minimally interfering with the model's existing knowledge. Let $Q_L, K_L, V_L \\in \\mathbb{R}^{d \\times d}$ be the pretrained text encoder's query, key, and value transforms for the attention mechanism in the final transformer layer $L$ , where $d$ is the internal dimension of $\\psi_T$ . For each attention head, the output of the multi-head attention layer is calculated as", + "bbox": [ + 511, + 396, + 905, + 594 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (x) = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {L} x \\left(K _ {L} x\\right) ^ {T}}{\\sqrt {d}}\\right) V _ {L, c} ^ {\\prime} x \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 612, + 905, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $V_{L,c}^{\\prime}$ is the value transform updated with LoRA as:", + "bbox": [ + 511, + 652, + 892, + 670 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV _ {L, c} ^ {\\prime} = V _ {L} + B _ {L, c} A _ {L, c}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 678, + 903, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$V_{L,c}^{\\prime}$ is learned separately for each concept $c$ . Following prior work on personalized generation [14, 19], we choose a fixed token (e.g. \"sks\") in CLIP's vocabulary as a placeholder for the concept's place in the input queries. During training, we randomly select a template textual query (e.g. \"An image of sks\") to associate with each training image to form a set of text-image pairs $\\{q_i^c, I_i^c\\}_{i=1}^{N_c}$ . We supervise the learning of $V_{L,c}^{\\prime}$ by using a mean-squared error (MSE) loss to push the normalized text and image embeddings for each training pair close together in the embedding space:", + "bbox": [ + 511, + 702, + 905, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M S E}} = \\frac {1}{N _ {c}} \\sum_ {i = 1} ^ {N _ {c}} \\left(\\frac {\\psi_ {T , c} ^ {\\prime} (q _ {i})}{| | \\psi_ {T , c} ^ {\\prime} (q _ {i}) | | _ {2}} - \\frac {\\psi_ {I} \\left(I _ {i} ^ {c}\\right)}{| | \\psi_ {I} \\left(I _ {i} ^ {c}\\right) | | _ {2}}\\right) ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 859, + 906, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19750", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\psi_T^\\prime ,c$ represents the text encoder with concept $c$ 's LoRA update applied in the forward pass (Eq. (3)).", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Regularization", + "text_level": 1, + "bbox": [ + 89, + 128, + 243, + 145 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A key challenge for personalized retrieval is to learn update weights that supply necessary personalized information without overriding the model's general knowledge, which is necessary for retrieving the personal concept in different contexts (e.g., \"my dog Fido catching a frisbee\" and \"my dog Fido sitting on the couch\"). We therefore propose a regularization scheme for POLAR. Differently than prior works that construct additional regularization examples to use during training [4, 7], we instead exploit the structure of our low-rank updates to directly minimize updates to the original representation. From Eq. (4), our parameter update alters the representation $V_{L}x$ by adding the term $B_{L,c}A_{L,c}x$ . With our use of rank $r = 1$ , we can interpret $A_{L,c}x$ as computing dot product similarity between the vectors $A_{L,c}^{T}$ and $x$ , which determines the scale of an added directional update $B_{L,c}$ . Our regularization comprises two components: first, we add a penalty on the size of the weights in $B_{L,c}$ to avoid unnecessary deviation from CLIP's existing representation, i.e., when the term $B_{L,c}A_{L,c}x = 0$ , the text embedding will be the same as CLIP's original representation $(\\psi_T'(q) = \\psi_T(q))$ . We modify our loss by adding a squared- $L_{2}$ regularization over the weights $B_{L,c}$ :", + "bbox": [ + 89, + 151, + 483, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M S E}} + \\lambda \\mathcal {L} _ {\\mathrm {r e g}}, \\quad \\mathcal {L} _ {\\mathrm {r e g}} = | B _ {L, c} | ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 491, + 482, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is a tunable hyperparameter that determines the relative weight of regularization in the loss. Second, we impose the constraint $||A_{L,c}||_2 = 1$ , encouraging $A_{L,c}$ to learn to selectively identify when to apply personal context based on directional similarity with the incoming representation $x$ , while the magnitude of the personal update is controlled entirely by the regularized $B_{L,c}$ .", + "bbox": [ + 89, + 516, + 483, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Merging parameters for multi-concept queries", + "text_level": 1, + "bbox": [ + 89, + 630, + 482, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For queries that reference multiple personal concepts (e.g., \"my dog Fido is playing with Rex's favorite frisbee\"), we propose merging the parameter updates for the concepts into one weight update that is applied during encoding to provide personal context for both concepts. Let $V_{L,c_1}'$ and $V_{L,c_2}'$ be the individually learned weight updates for concepts $c_1$ and $c_2$ . We construct a combined weight update as:", + "bbox": [ + 89, + 652, + 482, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {L, c _ {1} + c _ {2}} ^ {\\prime} = V _ {L, c _ {1}} ^ {\\prime} + V _ {L, c _ {2}} ^ {\\prime}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 765, + 482, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the forward pass, this setup equates to adding the parameter updates for both personal concepts to the representation. This update is also equivalent to constructing a rank $r = 2$ update via matrix concatenation along the rank dimension as:", + "bbox": [ + 89, + 790, + 483, + 866 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {L, c _ {1} + c _ {2}} ^ {\\prime} = \\left[ \\begin{array}{l l} B _ {L, c _ {1}} & B _ {L, c _ {2}} \\end{array} \\right] \\left[ \\begin{array}{l} A _ {L, c _ {1}} \\\\ A _ {L, c _ {2}} \\end{array} \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 869, + 482, + 906 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This approach also generalizes to merging greater than two concepts. We explore other merging strategies in Tab. 8.", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 137, + 645, + 155 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we evaluate POLAR and compare it to existing works. We also provide ablations to give insight into the design choices within our method. We include further experiment details, analysis of personalization time, and discussion of limitations in the supplemental material.", + "bbox": [ + 511, + 164, + 905, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 251, + 616, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DeepFashion2. Cohen et al. [7] define a personalized retrieval benchmark on the DeepFashion2 dataset [13], where the 50 personal concepts are different clothing items. The test set includes 221 captions and images for retrieval. ConCon-Chi. The recent ConCon-Chi dataset aims to more comprehensively evaluate unique personal concepts in a variety of contexts; we thus use it primarily for our analysis. It consists of 20 concepts including household objects and chimeric concepts (combinations of multiple objects). There are 1084 context queries (735 single-concept, 349 multi-concept), and 4008 retrieval images. For direct comparisons to the original baselines [28], we report on the full TEST set, which includes 3 validation concepts. We verify our gains hold on the TEST-UNSEEN split, which excludes these concepts, in the supplemental.", + "bbox": [ + 511, + 273, + 906, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 511, + 513, + 702, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Context Queries. Context queries perform retrieval on a caption referencing the personal concept in a particular context (e.g., \"my dog Fido catching a frisbee in the backyard\"). The ground truth consists of the images labeled as matching this prompt. For DeepFashion2, there is one ground truth image per context query, while for ConCon-Chi there are 1-130 ground truth images (average $\\approx 6$ ).", + "bbox": [ + 511, + 536, + 905, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept-only Queries. Following Yeh et al. [38], we also report retrieval accuracy on \"concept-only\" queries to evaluate the model's ability to recognize the personal concept independent of context. For each concept, we use the input query \"An image of $\\mathrm{V}^{*}$ \" and compute retrieval metrics where the ground truth is all retrieval images that contain the concept. For ConCon-Chi, we include only single-concept images (2430 images) in the retrieval database.", + "bbox": [ + 511, + 643, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Metrics. We report retrieval accuracy using standard benchmark metrics for the datasets: mean reciprocal rank (mRR) – the average inverse rank of the first retrieved ground truth image; recall-at- $k$ ( $\\mathbf{r}@\\mathbf{k}$ ) – the average success rate within the top $k$ retrievals; and mean average precision (mAP) – the area under the precision-recall curve, averaged over all queries. We report mAP for settings with multiple ground truth images per query (concept-only queries on both datasets, and context queries on ConCon-Chi).", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19751", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/b6439906d51eec746452ef26710b441f2021b6c5a8bc20f997969ea0eece8fa0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodArch.ContextConcept-only
mRRrecall@5mRRmAP
AdapterViT-B/325.9---
COLLIE [32]ViT-B/327.9---
Text OnlyViT-B/3217.6---
AvgIm + TextViT-B/3218.8---
PALAVRA [7]ViT-B/3228.439.2--
SEARLE [4]ViT-B/3221.9027.1512.7425.97
OursViT-B/3234.8244.8859.2628.75
SEARLE [4]ViT-L/1427.6234.1232.0716.17
OursViT-L/1440.7251.3165.9635.07
", + "bbox": [ + 91, + 88, + 480, + 236 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1766cb38f9f5dfc5d67b1ac6f21c6ed562868a5a1ddc7e8d331307bd1c89cbb8.jpg", + "table_caption": [ + "Table 1. Comparison to prior work on the DeepFashion2 retrieval benchmark with 5 training images per concept. We report the mean over over 5 runs with 5 randomly chosen training images of the concept per run (see supplemental for standard error)." + ], + "table_footnote": [], + "table_body": "
MethodmRRContextConcept-only
mAPrecall@1mRRmAP
Coarse (class name)24.2116.8314.48--
\\(Discriminative^†\\)43.1630.1631.92--
\\(Rich^†\\)40.5827.6529.98--
PALAVRA [7]35.9923.5926.75--
Pic2Word [31]38.6226.3927.68--
SEARLE [4]43.9330.7433.4996.6761.94
Ours46.3332.3336.16100.0068.71
", + "bbox": [ + 91, + 303, + 491, + 434 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6d6a00421002edd0195108ef38bdee46585765f9fcffae49feae279a874b0a37.jpg", + "table_caption": [ + "(a) Comparison to prior work on the ConCon-Chi benchmark." + ], + "table_footnote": [], + "table_body": "
MethodContext (Single-concept)Context (Multi-concept)
mRRmAPr@1mRRmAPr@1
SEARLE49.5035.2539.0532.0621.2221.78
Ours51.6436.7341.7735.1323.0524.36
", + "bbox": [ + 101, + 449, + 470, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(b) Results for ConCon-Chi single-concept vs. multi-concept queries.", + "bbox": [ + 102, + 515, + 470, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Our approach achieves state of the art results on the challenging ConCon-Chi benchmark on all metrics. We also break down the results of our method and SEARLE[4] by single-concept and multi-concept queries, demonstrating best results on both. † refer to ConCon-Chi's provided text descriptors for each concept, which serve as oracles since they use knowledge of all concepts to manually determine a differentiating description.", + "bbox": [ + 89, + 534, + 482, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For our main method on ConCon-Chi, we use $\\lambda = 0.35$ . We train for 500 iterations with learning rate 0.001 and the Adam optimizer. Our model converges within 50 epochs. Because we optimize minimal parameters and backpropagate through only the final layer, personalization is fast, taking under 1 second on a V100 GPU. On DeepFashion2 we append the classname to $\\mathrm{V}^*$ (e.g., \"sks dress\") like in PALAVRA [7] and use use $\\lambda = 0.1$ . We use the same template prompts as PALAVRA for training. We provide further details in the supplemental.", + "bbox": [ + 89, + 652, + 482, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparison to Prior Work", + "text_level": 1, + "bbox": [ + 89, + 816, + 334, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method to prior work on DeepFashion2 in Tab. 1 and ConCon-Chi in Tab. 2. For DeepFashion2 we compare against PALAVRA and the baselines reported in its paper [7]. We also run SEARLE [4], a zero-shot com", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "posed image retrieval network, using the publicly available checkpoints. Our method achieves state-of-the-art results in the standard setting using the CLIP ViT-B/32 architecture. We also see improvement over SEARLE when using the larger CLIP ViT-L/14 architecture.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On ConCon-Chi, we compare against the baselines reported in the benchmark's paper [28]. All methods use the CLIP ViT-L/14 architecture. Differently from DeepFashion2, the zero-shot composed image retrieval methods (SEARLE and Pic2Word) outperform PALAVRA, which suggests that while they are effective in some settings, they struggle with differentiating between several similar concepts like the clothing items in DeepFashion2. This is qualitatively demonstrated in Fig. 3, which compares our method's retrievals with SEARLE. Our method, however, also achieves state-of-the-art results on ConCon-Chi, demonstrating flexibility across different benchmarks. Additionally, we achieve stronger mAP for the concept-only queries task than SEARLE.", + "bbox": [ + 511, + 169, + 906, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluating General Knowledge in LoRA vs. Token Learning. Our primary hypothesis is that learning a small, regularized parameter update within the encoding process can more effectively allow CLIP to reason about personal and general information together. While concept-only queries assess personal knowledge and context queries assess the combination of personal and general knowledge, our setting of tuning the text encoder's parameters also allows us to measure the retention of general knowledge. We do so by inputting general queries, which do not reference the personal concept, to the text encoder with our parameter update for the concept still applied. To source non-personal queries, we use LLaVA [22] to caption each image in the retrieval set. We define a new metric VLM caption recall@10, which measures the success rate of retrieving the image for which the caption was generated in the top 10 retrievals. We choose 10 retrievals because ConCon-Chi has many similar images for which the same caption is valid. A drop from the original CLIP's performance on this metric (52.69) indicates forgetting of general knowledge.", + "bbox": [ + 511, + 383, + 906, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Tab. 3, we compare our approach of learning internal parameter updates with learning input tokens in the same training setting. We consider Textual Inversion (TI), which learns a token that is integrated into input queries via a pseudo-word for the concept (e.g. \"A photo of V* jumping), and Prompt Tuning [43], which bridges the gap between TI and parameter updates by prepending learned prompt tokens to all queries (both personal and general). In contrast to TI, this allows us to use our VLM caption metric to measure the interference of the prompt tokens on general queries. We include results with 1 learned token as well as 2 learned tokens, which is equivalent in size to our rank-1 parameter update. With prompt tuning, we can achieve strong results on concept-only queries, but our VLM cap", + "bbox": [ + 511, + 688, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19752", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ea0993fae84a031f1cd3caf3267b5a78d909615879868824a6f00abca322eda7.jpg", + "image_caption": [ + "Figure 3. We compare the top 3 retrievals of our method vs. SEARLE for personal queries in the ConCon-Chi and DeepFashion2 datasets, with green borders indicating correct retrievals. We observe that SEARLE struggles to differentiate between concepts of similar classes, such as different clothing items. Our method more consistently retrieves the correct concept in the correct context, demonstrating effective composition of personal and general knowledge." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 908, + 448 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d33c23654a81da96044055894cce9cce5804f1789dc5619d6feba79ebd5f5108.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodContext (Single)Concept-onlyVLM cap
mRRmAPmRRmAPr@10
Original CLIP29.3920.7610.756.4652.69
Text. Inv. (1 tok)42.4532.9397.5064.71N/A
Text. Inv. (2 tok)41.7327.94100.0063.88N/A
Prompt (1 tok)31.7720.7096.2558.9530.84
Prompt (2 tok)33.1420.93100.0064.4915.35
Text. Inv. + Ours39.2926.55100.0064.7252.57
Ours51.6436.73100.0068.7152.62
", + "bbox": [ + 91, + 523, + 486, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. We compare our approach, which updates the weights of the text encoder, to tuning input tokens on ConCon-Chi. Rows 2-3 represent Textual Inversion, where a learned token is applied in place of the personal concept in queries that reference the concept. Rows 4-5 tune prompt tokens that are prepended to all text queries. Row 6 learns both our parameter update and the token. Tuning tokens can achieve competitive results on concept-only queries but struggles on contextual queries that require composing personal and general knowledge. This catastrophic forgetting of general knowledge is reflected by our VLM caption matching metric. We propose a new approach that achieves strong performance on contextual, concept-only, and general queries.", + "bbox": [ + 89, + 657, + 485, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion metric shows that this strength comes at the cost of catastrophic forgetting of general knowledge. We also combine TI with our method by learning the input token in addition to our parameter update. We observe that learning the", + "bbox": [ + 89, + 839, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "token reduces context performance, suggesting that the inserted learned token interferes with general knowledge. In contrast, our parameter updates achieve strong performance on personal queries while not interfering with CLIP's general knowledge, even when applied to non-personal queries. We hypothesize this is due to the minimally invasive structure of our updates, and that they are applied late in the encoding process in contrast to learned input tokens which influence the entire encoding process. We qualitatively illustrate this finding in Fig. 4; with the parameter update for a personal concept applied, the model successfully performs retrieval for both personalized queries and general queries.", + "bbox": [ + 511, + 526, + 908, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 724, + 668, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A key choice in developing our method is determining an effective, minimal set of parameters within the model for which to apply personalized updates. In this section, we empirically investigate rank size of the LoRA updates, which layers in the encoding process to apply LoRA updates, and on which parameters to learn LoRA updates. We additionally ablate our regularization and multi-concept merging strategies. All experiments use the CLIP ViT-L/14 architecture on the ConCon-Chi dataset with 5 training images, and we report our results on single-concept context queries", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19753", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f8975b6b5a8baeb0bd39df58f0a8361a4085aa6cb77cdb475075ab28549c7b2e.jpg", + "image_caption": [ + "Figure 4. Our parameter updates enable personalized retrieval without overriding the model's general knowledge. On the left we show the top 3 retrievals of our personalized model on context queries referencing the personal concept, with green indicating ground truth correct retrievals. On the right, we show the results of querying the same personalized model for general VLM-generated captions. There is not exhaustive ground truth for which images match each caption; however, qualitatively our model retrieves appropriate images in all cases." + ], + "image_footnote": [], + "bbox": [ + 102, + 97, + 906, + 412 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8832cfdb9479672567f35e8a970872f0450ac3e9df2fc74e518d5b81d6ddb3b3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LoRA rankContext (Single-Concept)Concept-onlyVLM cap
mRRmAPr@1mRRmAPr@10
r=252.3136.5942.04100.0066.0752.78
r=451.5236.6041.36100.0068.1352.61
r=851.4936.5841.36100.0068.1552.62
r=1651.6736.6641.50100.0067.9352.62
r=151.6436.7341.77100.0068.7152.62
", + "bbox": [ + 93, + 486, + 478, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(we ablate the merging strategy for multi-concept queries in Sec. 4.4). We performed our model selection (design choices for our main method and regularization weight $\\lambda$ ) on the validation split (3 concepts) and report our main results on the test split.", + "bbox": [ + 89, + 641, + 482, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LoRA Rank. We ablate the rank of the LoRA updates in Tab. 4, which controls the number of learned parameters for each update. We observe only small retrieval accuracy gains on single-concept queries as the rank increases, and achieve better concept-only and competitive multi-concept results with rank=1. The rank-1 concept update is also the most parameter efficient, storing only $2d$ parameters per concept, where $d$ is the encoder's internal dimension.", + "bbox": [ + 89, + 717, + 482, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Architecture Layers. Tab. 5 investigates which transformer layers to apply LoRA updates to, dictating how early or late into the encoder process personalized information is injected. We achieve the strongest results by applying", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7b89db7fde654055a5dc9aa399563e6ec24ce3d5a9a83b2d0bfc520305df6fba.jpg", + "table_caption": [ + "Table 4. Ablation of LoRA rank on ConCon-Chi." + ], + "table_footnote": [], + "table_body": "
Layer(s)Context (Single-Concept)Concept-onlyVLM cap
mRRmAPr@1mRRmAPr@10
11,1250.1835.3439.73100.0064.0952.64
10,11,1251.5135.8141.63100.0065.8752.62
all layers43.2329.9332.5297.5063.7752.45
144.6931.1234.1597.5064.6652.18
1251.6436.7341.77100.0068.7152.62
", + "bbox": [ + 514, + 487, + 913, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation of LoRA layers on ConCon-Chi.", + "bbox": [ + 555, + 598, + 859, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "personalization in the final layer (layer 12). We hypothesize that updating later layers is better than earlier layers because it allows our approach to apply a small, targeted update to the developed text query representation to inject personal information. In contrast, earlier layer updates are more likely to alter the full representation - not just the parts semantically belonging to the personal concept. Additionally, our findings align with works that suggest that the later layers in transformer encoders are the most important in constructing the final representation [11]. We also do not observe an overall benefit by learning LoRA updates for earlier layers in addition to the final layer; the final layer alone achieves best performance while also requiring the fewest learned parameters per concept.", + "bbox": [ + 511, + 627, + 906, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Parameters. Tab. 6 ablates the component of the transformer layer on which the LoRA is learned. We consider the linear transforms within the attention mechanism (query, key, value, and output), the 2 MLP layers, as well as the", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19754", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/af2560d4465b3054ebedd08833a75e4fcbf018ed70cc34801411aa6181b06dde.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Param(s)Context (Single-Concept)Concept-onlyVLM cap
mRRmAPr@1mRRmAPr@10
Q16.6511.497.6232.6610.9151.84
K15.5511.456.5328.009.2852.12
O46.9931.0438.5097.5060.9852.52
Q,K,V,O47.5231.6938.7897.5060.9052.54
Q,V51.5035.6041.50100.0065.6052.65
MLP143.0427.9133.88100.0055.8752.05
MLP249.4032.8738.91100.0055.7051.65
final proj51.1534.9840.82100.0060.5851.58
V51.6436.7341.77100.0068.7152.62
", + "bbox": [ + 91, + 88, + 490, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model's final projection layer. We see that the query and key transforms alone are not effective for personalization, indicating that directly transforming the output representation is crucial for personalization; re-weighting the existing tokens via altering the attention weight computation is not enough. The strongest results are achieved by updating the value transform. Interestingly, this setting outperforms updating the output transform and the following MLP layers. This result suggests that the value transform's placement in transforming the output of each attention head is optimal for personalization as opposed to later linear transforms that operate after the output of the attention heads is aggregated. We see negligible gains by pairing updates on the value transform with other parameters; in fact, this setting decreases the results on some metrics. Updating the final projection alone also produces competitive results, but lags on the concept-only metrics. Our analysis demonstrates that the value transform within the final layer is optimally positioned to learn a small, targeted update for personalization. Regularization. We ablate our regularization strategy in Tab. 7. Without regularization, we see drops on the contextual query performance and VLM caption metric, indicating forgetting of general knowledge. The concept-only mAP slightly increases, showing prioritization of personalization over retaining general knowledge. Our regularization strategies are complementary: using both together produces the best results. With our regularization scheme, we are able to produce strong results on contextual queries, avoid any degradation on general caption-matching performance, and still produce good concept-only performance.", + "bbox": [ + 89, + 280, + 482, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-Concept Merging. We consider alternative strategies for merging the low-rank parameter updates of different concepts in Tab. 8. Our hypothesis is that because we learn a single, constrained update for each concept, the updates for different concepts will be sufficiently different to not interfere with each other. Our results validate this; we see applying each update is better than altering them via averaging or pooling them together. We also consider Orthogonal Adapation [26], which constrains updates for different concepts to be orthogonal to eachother, thus avoiding interference. In this method, the $A$ matrix in the low-rank update", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6b63374a7e447d2a4eab4f989134b2ba4e99f6fe90ec005f3747bb8d9311bd08.jpg", + "table_caption": [ + "Table 6. Ablation of LoRA parameters on ConCon-Chi." + ], + "table_footnote": [], + "table_body": "
Reg.Context (Single-Concept)Concept-onlyVLM cap
ABmRRmAPr@1mRRmAPr@10
22.5114.3514.83100.0069.8952.52
✓33.7722.2925.44100.0069.6352.58
✓39.8426.4531.56100.0069.0152.57
✓✓51.6436.7341.77100.0068.7152.62
", + "bbox": [ + 514, + 88, + 924, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/407af02467bf46e27a12037a0ef040cf1193843859355a0eab6c71c2ff3cc6a7.jpg", + "table_caption": [ + "Table 7. Ablation of our regularization scheme: A denotes imposing the constraint $||A_{L,c}||_2 = 1$ , and B denotes applying the squared L2 penalty to $B_{L,c}$ (Eq. 6). Without our regularization, the personalized parameter updates cause the model overfit to the concept, producing high concept-only metrics but catastrophically forgetting general knowledge, as reflected in the context and VLM caption metrics. Our regularization prevents this forgetting while still achieving high concept-only performance." + ], + "table_footnote": [], + "table_body": "
Merge strategymRRmAPr@1
Avg LoRAs25.0315.5115.76
Max LoRAs25.2715.5616.05
Orthogonal Adaptation [26]28.3819.1418.62
Add LoRAs35.1323.0524.36
", + "bbox": [ + 550, + 295, + 867, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Performance of different merging strategies for multi-concept queries in ConCon-Chi.", + "bbox": [ + 511, + 381, + 906, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "is frozen and drawn from a shared orthogonal subspace, while the $B$ matrix is learned. The updates for different concepts are merged by adding them together. Interestingly, we find Orthogonal Adaptation to be less effective than our approach; we hypothesize this is due to the differences between where low rank updates are applied in our method vs. the original Orthogonal Adaptation method, which operates on text to image diffusion models. Whereas in personalized generation, parameter updates are applied throughout the full model, we find learning only a single parameter update is better suited for retrieval. Because we only learn this one parameter update late in the model, certain randomly selected $A$ matrices are not effective for personalization depending on how they interact with the incoming feature representation at that point in the model. Specifically, in computing $B_{L,c}A_{L,c}x$ if the randomly chosen $A_{L,c}$ is orthogonal to $x$ , this will eliminate the parameter update altogether. While personalized generation approaches avoid this by adapting the representation throughout all layers of the model, in our setting we find it is better to learn both $A$ and $B$ instead of imposing orthogonality.", + "bbox": [ + 511, + 424, + 908, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 753, + 633, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we show that updating the internal representation of the CLIP text encoder serves as a better alternative to textual inversion for personalized search. By constraining the parameter update for each concept to a single rank-one update in the value transform of the final layer and strategically regularizing the parameters, we demonstrate that our approach effectively personalizes from a few image examples while maintaining the model's general knowledge.", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19755", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments The authors thank the members of the Hoffman Lab at Georgia Tech and Jitesh Jain for their feedback on this work. Fiona Ryan is supported by the NSF Graduate Research Fellowship under Grant No. DGE-2039655. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors(s) and do not necessarily reflect the views of the National Science Foundation.", + "bbox": [ + 89, + 90, + 485, + 210 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 223, + 187, + 239 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuval Alaluf, Elad Richardson, Sergey Tulyakov, Kfir Aberman, and Daniel Cohen-Or. Myvlm: Personalizing vlms for user-specific queries. In European Conference on Computer Vision, pages 73–91. Springer, 2025. 2, 5", + "[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4959-4968, 2022. 2", + "[3] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21466-21474, 2022. 2", + "[4] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15338-15347, 2023. 1, 2, 3, 4, 5", + "[5] Adrian Bulat and Georgios Tzimiropoulos. Lasp: Text-to-text optimization for language-aware soft prompting of vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23232-23241, 2023. 3", + "[6] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. arXiv preprint arXiv:2210.01253, 2022. 3", + "[7] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In European conference on computer vision, pages 558-577. Springer, 2022. 1, 2, 3, 4, 5", + "[8] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor Guilherme Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Variational prompt tuning improves generalization of vision-language models. 2022. 3", + "[9] Ziyi Dong, Pengxu Wei, and Liang Lin. Dreamartist: Towards controllable one-shot text-to-image generation via positive-negative prompt-tuning. arXiv preprint arXiv:2211.11337, 2022. 2", + "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + ], + "bbox": [ + 93, + 247, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Yossi Gandelsman, Alexei A. Efros, and Jacob Steinhardt. Interpreting clip's image representation via text-based decomposition. In ICLR, 2023. 7", + "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3", + "[13] Yuying Ge, Ruimao Zhang, Xiaogang Wang, Xiaou Tang, and Ping Luo. Deepfashion2: A versatile benchmark for detection, pose estimation, segmentation and re-identification of clothing images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5337–5345, 2019. 4", + "[14] Cusuh Ham, Matthew Fisher, James Hays, Nicholas Kolkin, Yuchen Liu, Richard Zhang, and Tobias Hinz. Personalized residuals for concept-driven text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8186-8195, 2024. 2, 3, 4", + "[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 2", + "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3", + "[17] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 3", + "[18] Bruno Korbar and Andrew Zisserman. Personalised clip or: how to find your vacation videos. 2022. 2, 5", + "[19] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2, 3, 4", + "[20] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1401-1411, 2023. 3", + "[21] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 3", + "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5, 3", + "[23] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 3", + "[24] Joanna Materzynska, Josef Sivic, Eli Shechtman, Antonio Torralba, Richard Zhang, and Bryan Russell. Customizing" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19756", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "motion in text-to-video diffusion models. arXiv preprint arXiv:2312.04966, 2023. 4", + "[25] Thao Nguyen, Haotian Liu, Yuheng Li, Mu Cai, Utkarsh Ojha, and Yong Jae Lee. Yo'llava: Your personalized language and vision assistant. arXiv preprint arXiv:2406.09400, 2024. 2", + "[26] Ryan Po, Guandao Yang, Kfir Aberman, and Gordon Wetzstein. Orthogonal adaptation for modular customization of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7964-7973, 2024. 2, 8", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2", + "[28] Andrea Rosasco, Stefano Berti, Giulia Pasquale, Damiano Malafronte, Shogo Sato, Hiroyuki Segawa, Tetsugo Inada, and Lorenzo Natale. Concon-chi: Concept-context chimera benchmark for personalized vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22239-22248, 2024. 2, 4, 5, 1, 3", + "[29] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2", + "[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6527-6536, 2024. 2", + "[31] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19305-19314, 2023. 2, 5, 3", + "[32] Gabriel Skantze and Bram Willemsen. Collie: Continual learning of language grounding from language-image embeddings. Journal of Artificial Intelligence Research, 74: 1201-1223, 2022. 5, 2", + "[33] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 conference proceedings, pages 1-11, 2023. 2", + "[34] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 3", + "[35] Nam Vo, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image" + ], + "bbox": [ + 91, + 92, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for image retrieval—an empirical odyssey. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6439-6448, 2019. 2", + "[36] Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guoqiang Liang, Peng Wang, and Yanning Zhang. Dual modality prompt tuning for vision-language pre-trained model. IEEE Transactions on Multimedia, 2023. 3", + "[37] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6757-6767, 2023. 3", + "[38] Chun-Hsiao Yeh, Bryan Russell, Josef Sivic, Fabian Caba Heilbron, and Simon Jenni. Meta-personalizing vision-language models to find named instances in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19123-19132, 2023. 2, 3, 4, 1", + "[39] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the association for computational linguistics, 2:67-78, 2014. 4", + "[40] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926, 2023. 2", + "[41] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaptation of clip for few-shot classification. In European conference on computer vision, pages 493-510. Springer, 2022. 3", + "[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 3", + "[43] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 5", + "[44] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 3" + ], + "bbox": [ + 516, + 92, + 906, + 726 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19757", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/673cbe39-bd0b-47dd-99b2-3c870c8defd5_model.json b/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/673cbe39-bd0b-47dd-99b2-3c870c8defd5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..14ebbeb028746b7a1c6c3095f5e4c80d1e21a7e6 --- /dev/null +++ b/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/673cbe39-bd0b-47dd-99b2-3c870c8defd5_model.json @@ -0,0 +1,1848 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.131, + 0.904, + 0.153 + ], + "angle": 0, + "content": "Improving Personalized Search with Regularized Low-Rank Parameter Updates" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.18, + 0.921, + 0.218 + ], + "angle": 0, + "content": "Fiona Ryan\\(^{1,2*}\\), Josef Sivic\\(^{2,3}\\), Fabian Caba Heilbron\\(^{2}\\), Judy Hoffman\\(^{1}\\), James M. Rehg\\(^{4}\\), Bryan Russell\\(^{2}\\) \n\\(^{1}\\)Georgia Tech, \\(^{2}\\)Adobe Research, \\(^{3}\\)CIIRC CTU, \\(^{4}\\)UIUC" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.267, + 0.327, + 0.283 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.299, + 0.486, + 0.648 + ], + "angle": 0, + "content": "Personalized vision-language retrieval seeks to recognize new concepts (e.g., \"my dog Fido\") from only a few examples. This task is challenging because it requires not only learning a new concept from a few images, but also integrating the personal and general knowledge together to recognize the concept in different contexts. In this paper, we show how to effectively adapt the internal representation of a vision-language dual encoder model for personalized vision-language retrieval. We find that regularized low-rank adaption of a small set of parameters in the language encoder's final layer serves as a highly effective alternative to textual inversion for recognizing the personal concept while preserving general knowledge. Additionally, we explore strategies for combining parameters of multiple learned personal concepts, finding that parameter addition is effective. To evaluate how well general knowledge is preserved in a finetuned representation, we introduce a metric that measures image retrieval accuracy based on captions generated by a vision language model (VLM). Our approach achieves state-of-the-art accuracy on two benchmarks for personalized image retrieval with natural language queries - DeepFashion2 and ConCon-Chi - outperforming the prior art by \\(4\\% - 22\\%\\) on personal retrievals." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.222, + 0.683 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.485, + 0.876 + ], + "angle": 0, + "content": "Personalizing a vision-language retrieval model (PerVL) aims to adapt a pretrained vision-language dual encoder model (e.g., CLIP [27]) to recognize new concepts (e.g., \"my dog Fido\") from just a few examples [7]. This task is important for search applications that need to identify concepts missing from the pretrained model's knowledge, such as searching one's personal photo library for a specific person, object, or pet. PerVL is challenging because it requires not only learning a new concept from a few visual examples, but also reasoning about the personal concept and general knowledge together to retrieve the concept in different contexts. For instance, searching for \"my dog Fido catching" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.253, + 0.911, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.468, + 0.907, + 0.594 + ], + "angle": 0, + "content": "Figure 1. Left: Prior works use a pretrained textual inversion network \\((\\mathcal{F})\\) to compute a pseudo-token to represent a new concept, which may be further optimized during personalization. Right: We present POLAR, which represents new concepts as a small low-rank parameter update within the text encoder. Instead of inserting a learned pseudo-token to the text query, we use a fixed vocabulary token \\(V^{*}\\). We show that our method is effective at recognizing the personal concept from a few examples while retaining the model's general knowledge, and does not require large scale pretraining." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.907, + 0.643 + ], + "angle": 0, + "content": "a frisbee\" requires both personal knowledge (\"Fido\") and general knowledge (\"catching a frisbee\")." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recent approaches for PerVL use textual inversion to learn a pseudo-text token (e.g., “\\(V^{*}\\)”) for each new personal concept by inverting from the target training images to the token [4, 7] (illustrated in Figure 1 (left)). These pseudo-text tokens require updating only a minimal set of parameters, and they can be used within a natural language query as input to the language encoder (e.g., “\\(V^{*}\\) catching a frisbee”). Typically, these approaches train a textual inversion network on large-scale data to predict a token representing a new concept, which may then be further optimized during personalization. While these approaches avoid updating the model's internal parameters, preventing the “overriding” of its general knowledge, their ability to represent the personal concept is limited to the single input text token. Furthermore, the token affects the entire text embedding process and can interfere with the language encoder's general knowledge. Consequently, textual inversion" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.428, + 0.901 + ], + "angle": 0, + "content": "*Work partially done during internship at Adobe Research." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19748" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "approaches often struggle to combine personalized and general knowledge across different datasets and can be time-consuming to optimize online due to the need for backpropagation through the full language encoder [7, 38], or require large-scale pretraining [4, 7, 38]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.168, + 0.482, + 0.318 + ], + "angle": 0, + "content": "To address these shortcomings, we focus on updating the internal representation of CLIP's language encoder [27] using only a few training examples (illustrated in Figure 1 (right)). Updating the internal representation is challenging because the model must balance learning the personalized concept from limited examples while preventing catastrophic forgetting of its prior knowledge. Additionally, the training data does not inherently encourage the model to retain general knowledge. Our goal is to make this update efficiently, without relying on additional training data." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.319, + 0.483, + 0.59 + ], + "angle": 0, + "content": "Our contributions are fourfold. First, inspired by recent advances in text-to-image generation [14, 19, 29, 30], we show that CLIP's language encoder can be updated to learn a personalized concept from a few examples while retaining its general knowledge. We find that regularized low-rank adaptation (LoRA) finetuning [16] of the language encoder's last layer effectively balances the trade-off between learning the personalized concept and avoiding catastrophic forgetting. Our resulting method POLAR (PersOnalized Low-rank Adaptation for Retrieval) learns a low-rank parameter set that is minimal and comparable in size to the pseudo-text tokens used in textual inversion. Furthermore, we leverage LoRA's specific structure to introduce a regularization strategy that eliminates the need for additional training examples or regularization prompts. Unlike previous methods, our approach does not rely on any components pre-trained on large-scale data, allowing for seamless parameter updates using only a few training examples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.483, + 0.772 + ], + "angle": 0, + "content": "Second, we introduce a new evaluation metric to assess how well general knowledge is preserved in our finetuned representation, using captions generated by a vision-language model (VLM). We find that our approach effectively maintains general knowledge. Third, we explore different strategies for combining learned representations for different personal concepts to support multi-concept queries. We find that adding LoRA representations is effective and outperforms orthogonal adaptation [26]. Finally, we demonstrate that POLAR achieves state-of-the-art accuracy on the DeepFashion2 [7] and ConCon-Chi [28] benchmarks, improving prior performance by \\(4\\% - 22\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.233, + 0.801 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Personalized Vision-Language Retrieval. Cohen et al. introduced the task of Personalized Vision-Language Retrieval [7] and proposed PALAVRA, a textual inversion approach for the task. PALAVRA first learns a textual inversion network on COCO, which takes a set of images of a concept and predicts an initial pseudo-word token to rep" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.395 + ], + "angle": 0, + "content": "resent the concept in text queries. This token is then further optimized via backpropagation through CLIP's text encoder. Korbar et al. [18] propose a similar approach for the setting of retrieving specific people in videos. A few recent works explore learning embeddings to represent personal concepts for VLM captioning and QA tasks [1, 25]. Yeh et al. [38] build on the concept of textual inversion by meta-learning a basis for pseudo-tokens using large scale video data. However, they target a different setting where concepts are learned jointly, using the other concepts in the dataset as negative examples. Recently, Rosasco et al. introduced the Concept-Context Chimera dataset (ConConChi) for personalized retrieval, which provides a more rigorous benchmark for assessing retrieval of personal concepts across diverse contexts than prior datasets. In this work, we depart from prior work by representing new concepts as low rank parameter updates within the text encoder instead of pseudo-word tokens. We show that our approach more effectively composes personal and general knowledge, while requiring few parameters per-concept." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.905, + 0.807 + ], + "angle": 0, + "content": "Personalized Generation. Personalized generation is a more studied related task that generates new images of a personal concept using text-to-image diffusion models. Some approaches use textual inversion to optimize pseudoword tokens to use within text prompts [9, 10], while others like Dreambooth [29] and Custom Diffusion [19] find that tuning the weights of the diffusion U-net generates personal concepts with better fidelity. Recent work has focused on selectively tuning certain parameters to promote parameter efficiency and speed [9, 15, 19, 40], with some leveraging low rank constraints [14, 26, 30, 33]. Most related to our work is Perfusion [33], which learns rank-one updates to the diffusion U-net with a key-locking mechanism to constrain updates to the concept's spatial location in the feature map. While personalization via parameter updates has become commonplace for personalized generation, it has not yet been explored for personalized retrieval, motivating our work. Due to the differences in the nature of the tasks (generative vs. discriminative) and models (text-to-image diffusion vs. dual encoder), we find that retrieval demands a different strategy for applying internal parameter updates; updating even sparse sets of parameters can result in catastrophic forgetting of the model's general knowledge. Instead of updating parameters throughout the full model, we apply a single rank-one parameter update to the final layer of CLIP's text encoder and directly regularize these parameters to avoid catastrophic forgetting of general knowledge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Composed Image Retrieval. Another related task is composed image retrieval [35], which takes an image and textual modification as inputs and performs image retrieval. Approaches have leveraged CLIP for composed image retrieval [2, 3], with Pic2Word [31] and SEARLE [4] learning textual inversion networks to predict a pseudo-word token" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19749" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.478, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.284, + 0.485, + 0.342 + ], + "angle": 0, + "content": "Figure 2. Overview of POLAR. For a personal concept, POLAR learns a rank-1 update to the value transform in the final layer of the text encoder. To maintain general knowledge during personalization, we impose a regularization loss on the update." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.439 + ], + "angle": 0, + "content": "for the input image. Importantly, composed image retrieval differs from personalized retrieval in that it does not require instance-level recognition of the same concept (e.g., retrieving the exact same person), and instead aims to retrieve images portraying a similar semantic class, layout, or style." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.637 + ], + "angle": 0, + "content": "Few-Shot Adaptation of Vision-Language Models. Our task also relates to work on adapting models like CLIP for few-shot classification. Prior works primarily leverage prompt tuning [21], which learns input tokens to represent new classes [5, 6, 8, 23, 34, 36, 37, 42-44]. A few works consider tuning within the encoder by applying prompts to all layers [17, 20] or augmenting the encoder with adapter modules [12, 41]. An important distinction is that while CLIP's general knowledge may aid with learning classes from few examples, the general knowledge is not required for these downstream classification tasks. In contrast, personalized retrieval requires not just recognizing the personal concept, but composing it with retained general knowledge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.483, + 0.687 + ], + "angle": 0, + "content": "3. Personalized Low-Rank Adaptation for Retrieval (POLAR)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.295, + 0.712 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We follow the PerVL problem setup established by Cohen et al. [7]. Given a pretrained vision-language model \\(\\psi\\), we aim to learn an adapted model \\(\\psi'\\) that is able to recognize a new personal concept \\(c\\) (e.g., \"my coffee mug\"). The pretrained model \\(\\psi\\) consists of an image encoder \\(\\psi_I\\) and a text encoder \\(\\psi_T\\). These encoders map image and textual inputs to a shared embedding space. At personalization time, we are given \\(N_c\\) images \\(\\{I_i^c\\}_{i=1}^{N_c}\\) depicting concept \\(c\\) as training data. Following Cohen et al. [7], we are also given the class name \\(\\mathcal{C}_c\\) of concept \\(c\\) (e.g., \"mug\"). At retrieval time, the input to the model consists of a textual query \\(q\\) and a set of \\(N_r\\) images \\(\\{I_i^r\\}_{i=1}^{N_r}\\) constituting the retrieval database. We" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.093, + 0.906, + 0.152 + ], + "angle": 0, + "content": "compute the language embedding for an input query \\( q \\) as \\( \\psi_T'(q) \\), and perform retrieval by computing the cosine similarity between \\( \\psi_T'(q) \\) and the image embedding for each retrieval image as" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.159, + 0.907, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s i m} \\left(\\psi_ {T} ^ {\\prime} (q), \\psi_ {I} ^ {\\prime} \\left(I _ {i} ^ {r}\\right)\\right) = \\frac {\\left\\langle \\psi_ {T} ^ {\\prime} (q) , \\psi_ {I} ^ {\\prime} \\left(I _ {i} ^ {r}\\right) \\right\\rangle}{\\left| \\left| \\psi_ {T} ^ {\\prime} (q) \\right| \\right| _ {2} \\left| \\left| \\psi_ {I} ^ {\\prime} \\left(I _ {i} ^ {r}\\right) \\right| \\right| _ {2}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.199, + 0.906, + 0.247 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot \\rangle\\) denotes the inner product. The top retrieval for the query is the image \\(I_{i}^{r}\\) corresponding to the embedding with the highest similarity to the query embedding \\(\\psi_T^\\prime (q)\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.253, + 0.849, + 0.269 + ], + "angle": 0, + "content": "3.2. Rank-One Personalized Value Updates" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.906, + 0.366 + ], + "angle": 0, + "content": "In contrast to prior work that represents a concept as a learned input token for the text encoder [4, 7, 38], POLAR learns a low-rank parameter update to the text encoder for each concept (Fig. 2). We leverage LoRA [16]: for a weight \\( W \\in \\mathbb{R}^{m \\times n} \\) in a pretrained model, LoRA learns a low-rank update \\( \\Delta W \\), performing a modified forward pass as" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.373, + 0.907, + 0.39 + ], + "angle": 0, + "content": "\\[\ny = (W + \\Delta W) x = W x + B A x \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.906, + 0.595 + ], + "angle": 0, + "content": "where \\(B \\in \\mathbb{R}^{m \\times r}\\), \\(A \\in \\mathbb{R}^{r \\times n}\\), and \\(r < \\min(m, n)\\) is the chosen rank of the weight update. By selecting \\(r << \\min(m, n)\\), LoRA tunes minimal parameters in the original model \\(\\psi\\). For each concept \\(c\\), we empirically choose to learn a rank-one (\\(r = 1\\)) update to the value transform of the final attention layer \\(L\\) in \\(\\psi_T\\). Our choice of \\(r = 1\\) reflects our goal to represent a single concept from very limited examples, while minimally interfering with the model's existing knowledge. Let \\(Q_L, K_L, V_L \\in \\mathbb{R}^{d \\times d}\\) be the pretrained text encoder's query, key, and value transforms for the attention mechanism in the final transformer layer \\(L\\), where \\(d\\) is the internal dimension of \\(\\psi_T\\). For each attention head, the output of the multi-head attention layer is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.613, + 0.906, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (x) = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {L} x \\left(K _ {L} x\\right) ^ {T}}{\\sqrt {d}}\\right) V _ {L, c} ^ {\\prime} x \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.654, + 0.893, + 0.671 + ], + "angle": 0, + "content": "where \\(V_{L,c}^{\\prime}\\) is the value transform updated with LoRA as:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.679, + 0.905, + 0.696 + ], + "angle": 0, + "content": "\\[\nV _ {L, c} ^ {\\prime} = V _ {L} + B _ {L, c} A _ {L, c}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.906, + 0.855 + ], + "angle": 0, + "content": "\\(V_{L,c}^{\\prime}\\) is learned separately for each concept \\(c\\). Following prior work on personalized generation [14, 19], we choose a fixed token (e.g. \"sks\") in CLIP's vocabulary as a placeholder for the concept's place in the input queries. During training, we randomly select a template textual query (e.g. \"An image of sks\") to associate with each training image to form a set of text-image pairs \\(\\{q_i^c, I_i^c\\}_{i=1}^{N_c}\\). We supervise the learning of \\(V_{L,c}^{\\prime}\\) by using a mean-squared error (MSE) loss to push the normalized text and image embeddings for each training pair close together in the embedding space:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.861, + 0.907, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M S E}} = \\frac {1}{N _ {c}} \\sum_ {i = 1} ^ {N _ {c}} \\left(\\frac {\\psi_ {T , c} ^ {\\prime} (q _ {i})}{| | \\psi_ {T , c} ^ {\\prime} (q _ {i}) | | _ {2}} - \\frac {\\psi_ {I} \\left(I _ {i} ^ {c}\\right)}{| | \\psi_ {I} \\left(I _ {i} ^ {c}\\right) | | _ {2}}\\right) ^ {2} \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19750" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "where \\(\\psi_T^\\prime ,c\\) represents the text encoder with concept \\(c\\) 's LoRA update applied in the forward pass (Eq. (3))." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.13, + 0.245, + 0.146 + ], + "angle": 0, + "content": "3.3. Regularization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.152, + 0.484, + 0.485 + ], + "angle": 0, + "content": "A key challenge for personalized retrieval is to learn update weights that supply necessary personalized information without overriding the model's general knowledge, which is necessary for retrieving the personal concept in different contexts (e.g., \"my dog Fido catching a frisbee\" and \"my dog Fido sitting on the couch\"). We therefore propose a regularization scheme for POLAR. Differently than prior works that construct additional regularization examples to use during training [4, 7], we instead exploit the structure of our low-rank updates to directly minimize updates to the original representation. From Eq. (4), our parameter update alters the representation \\( V_{L}x \\) by adding the term \\( B_{L,c}A_{L,c}x \\). With our use of rank \\( r = 1 \\), we can interpret \\( A_{L,c}x \\) as computing dot product similarity between the vectors \\( A_{L,c}^{T} \\) and \\( x \\), which determines the scale of an added directional update \\( B_{L,c} \\). Our regularization comprises two components: first, we add a penalty on the size of the weights in \\( B_{L,c} \\) to avoid unnecessary deviation from CLIP's existing representation, i.e., when the term \\( B_{L,c}A_{L,c}x = 0 \\), the text embedding will be the same as CLIP's original representation \\( (\\psi_T'(q) = \\psi_T(q)) \\). We modify our loss by adding a squared- \\( L_{2} \\) regularization over the weights \\( B_{L,c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.492, + 0.483, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M S E}} + \\lambda \\mathcal {L} _ {\\mathrm {r e g}}, \\quad \\mathcal {L} _ {\\mathrm {r e g}} = | B _ {L, c} | ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.625 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a tunable hyperparameter that determines the relative weight of regularization in the loss. Second, we impose the constraint \\(||A_{L,c}||_2 = 1\\), encouraging \\(A_{L,c}\\) to learn to selectively identify when to apply personal context based on directional similarity with the incoming representation \\(x\\), while the magnitude of the personal update is controlled entirely by the regularized \\(B_{L,c}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.631, + 0.483, + 0.648 + ], + "angle": 0, + "content": "3.4. Merging parameters for multi-concept queries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.759 + ], + "angle": 0, + "content": "For queries that reference multiple personal concepts (e.g., \"my dog Fido is playing with Rex's favorite frisbee\"), we propose merging the parameter updates for the concepts into one weight update that is applied during encoding to provide personal context for both concepts. Let \\( V_{L,c_1}' \\) and \\( V_{L,c_2}' \\) be the individually learned weight updates for concepts \\( c_1 \\) and \\( c_2 \\). We construct a combined weight update as:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.766, + 0.483, + 0.786 + ], + "angle": 0, + "content": "\\[\nV _ {L, c _ {1} + c _ {2}} ^ {\\prime} = V _ {L, c _ {1}} ^ {\\prime} + V _ {L, c _ {2}} ^ {\\prime}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.791, + 0.484, + 0.867 + ], + "angle": 0, + "content": "During the forward pass, this setup equates to adding the parameter updates for both personal concepts to the representation. This update is also equivalent to constructing a rank \\( r = 2 \\) update via matrix concatenation along the rank dimension as:" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.871, + 0.483, + 0.907 + ], + "angle": 0, + "content": "\\[\nV _ {L, c _ {1} + c _ {2}} ^ {\\prime} = \\left[ \\begin{array}{l l} B _ {L, c _ {1}} & B _ {L, c _ {2}} \\end{array} \\right] \\left[ \\begin{array}{l} A _ {L, c _ {1}} \\\\ A _ {L, c _ {2}} \\end{array} \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "This approach also generalizes to merging greater than two concepts. We explore other merging strategies in Tab. 8." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.138, + 0.646, + 0.156 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.165, + 0.906, + 0.241 + ], + "angle": 0, + "content": "In this section, we evaluate POLAR and compare it to existing works. We also provide ablations to give insight into the design choices within our method. We include further experiment details, analysis of personalization time, and discussion of limitations in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.252, + 0.617, + 0.267 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.502 + ], + "angle": 0, + "content": "DeepFashion2. Cohen et al. [7] define a personalized retrieval benchmark on the DeepFashion2 dataset [13], where the 50 personal concepts are different clothing items. The test set includes 221 captions and images for retrieval. ConCon-Chi. The recent ConCon-Chi dataset aims to more comprehensively evaluate unique personal concepts in a variety of contexts; we thus use it primarily for our analysis. It consists of 20 concepts including household objects and chimeric concepts (combinations of multiple objects). There are 1084 context queries (735 single-concept, 349 multi-concept), and 4008 retrieval images. For direct comparisons to the original baselines [28], we report on the full TEST set, which includes 3 validation concepts. We verify our gains hold on the TEST-UNSEEN split, which excludes these concepts, in the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.514, + 0.704, + 0.529 + ], + "angle": 0, + "content": "4.2. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.906, + 0.643 + ], + "angle": 0, + "content": "Context Queries. Context queries perform retrieval on a caption referencing the personal concept in a particular context (e.g., \"my dog Fido catching a frisbee in the backyard\"). The ground truth consists of the images labeled as matching this prompt. For DeepFashion2, there is one ground truth image per context query, while for ConCon-Chi there are 1-130 ground truth images (average \\(\\approx 6\\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Concept-only Queries. Following Yeh et al. [38], we also report retrieval accuracy on \"concept-only\" queries to evaluate the model's ability to recognize the personal concept independent of context. For each concept, we use the input query \"An image of \\(\\mathrm{V}^{*}\\)\" and compute retrieval metrics where the ground truth is all retrieval images that contain the concept. For ConCon-Chi, we include only single-concept images (2430 images) in the retrieval database." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Metrics. We report retrieval accuracy using standard benchmark metrics for the datasets: mean reciprocal rank (mRR) – the average inverse rank of the first retrieved ground truth image; recall-at- \\(k\\) (\\(\\mathbf{r}@\\mathbf{k}\\)) – the average success rate within the top \\(k\\) retrievals; and mean average precision (mAP) – the area under the precision-recall curve, averaged over all queries. We report mAP for settings with multiple ground truth images per query (concept-only queries on both datasets, and context queries on ConCon-Chi)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19751" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.237 + ], + "angle": 0, + "content": "
MethodArch.ContextConcept-only
mRRrecall@5mRRmAP
AdapterViT-B/325.9---
COLLIE [32]ViT-B/327.9---
Text OnlyViT-B/3217.6---
AvgIm + TextViT-B/3218.8---
PALAVRA [7]ViT-B/3228.439.2--
SEARLE [4]ViT-B/3221.9027.1512.7425.97
OursViT-B/3234.8244.8859.2628.75
SEARLE [4]ViT-L/1427.6234.1232.0716.17
OursViT-L/1440.7251.3165.9635.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.242, + 0.483, + 0.298 + ], + "angle": 0, + "content": "Table 1. Comparison to prior work on the DeepFashion2 retrieval benchmark with 5 training images per concept. We report the mean over over 5 runs with 5 randomly chosen training images of the concept per run (see supplemental for standard error)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.304, + 0.493, + 0.435 + ], + "angle": 0, + "content": "
MethodmRRContextConcept-only
mAPrecall@1mRRmAP
Coarse (class name)24.2116.8314.48--
\\(Discriminative^†\\)43.1630.1631.92--
\\(Rich^†\\)40.5827.6529.98--
PALAVRA [7]35.9923.5926.75--
Pic2Word [31]38.6226.3927.68--
SEARLE [4]43.9330.7433.4996.6761.94
Ours46.3332.3336.16100.0068.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.438, + 0.45, + 0.449 + ], + "angle": 0, + "content": "(a) Comparison to prior work on the ConCon-Chi benchmark." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.45, + 0.472, + 0.514 + ], + "angle": 0, + "content": "
MethodContext (Single-concept)Context (Multi-concept)
mRRmAPr@1mRRmAPr@1
SEARLE49.5035.2539.0532.0621.2221.78
Ours51.6436.7341.7735.1323.0524.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.516, + 0.471, + 0.529 + ], + "angle": 0, + "content": "(b) Results for ConCon-Chi single-concept vs. multi-concept queries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.633 + ], + "angle": 0, + "content": "Table 2. Our approach achieves state of the art results on the challenging ConCon-Chi benchmark on all metrics. We also break down the results of our method and SEARLE[4] by single-concept and multi-concept queries, demonstrating best results on both. † refer to ConCon-Chi's provided text descriptors for each concept, which serve as oracles since they use knowledge of all concepts to manually determine a differentiating description." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.806 + ], + "angle": 0, + "content": "Implementation Details. For our main method on ConCon-Chi, we use \\(\\lambda = 0.35\\). We train for 500 iterations with learning rate 0.001 and the Adam optimizer. Our model converges within 50 epochs. Because we optimize minimal parameters and backpropagate through only the final layer, personalization is fast, taking under 1 second on a V100 GPU. On DeepFashion2 we append the classname to \\(\\mathrm{V}^*\\) (e.g., \"sks dress\") like in PALAVRA [7] and use use \\(\\lambda = 0.1\\). We use the same template prompts as PALAVRA for training. We provide further details in the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.336, + 0.833 + ], + "angle": 0, + "content": "4.3. Comparison to Prior Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We compare our method to prior work on DeepFashion2 in Tab. 1 and ConCon-Chi in Tab. 2. For DeepFashion2 we compare against PALAVRA and the baselines reported in its paper [7]. We also run SEARLE [4], a zero-shot com" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "posed image retrieval network, using the publicly available checkpoints. Our method achieves state-of-the-art results in the standard setting using the CLIP ViT-B/32 architecture. We also see improvement over SEARLE when using the larger CLIP ViT-L/14 architecture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.17, + 0.907, + 0.382 + ], + "angle": 0, + "content": "On ConCon-Chi, we compare against the baselines reported in the benchmark's paper [28]. All methods use the CLIP ViT-L/14 architecture. Differently from DeepFashion2, the zero-shot composed image retrieval methods (SEARLE and Pic2Word) outperform PALAVRA, which suggests that while they are effective in some settings, they struggle with differentiating between several similar concepts like the clothing items in DeepFashion2. This is qualitatively demonstrated in Fig. 3, which compares our method's retrievals with SEARLE. Our method, however, also achieves state-of-the-art results on ConCon-Chi, demonstrating flexibility across different benchmarks. Additionally, we achieve stronger mAP for the concept-only queries task than SEARLE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.908, + 0.687 + ], + "angle": 0, + "content": "Evaluating General Knowledge in LoRA vs. Token Learning. Our primary hypothesis is that learning a small, regularized parameter update within the encoding process can more effectively allow CLIP to reason about personal and general information together. While concept-only queries assess personal knowledge and context queries assess the combination of personal and general knowledge, our setting of tuning the text encoder's parameters also allows us to measure the retention of general knowledge. We do so by inputting general queries, which do not reference the personal concept, to the text encoder with our parameter update for the concept still applied. To source non-personal queries, we use LLaVA [22] to caption each image in the retrieval set. We define a new metric VLM caption recall@10, which measures the success rate of retrieving the image for which the caption was generated in the top 10 retrievals. We choose 10 retrievals because ConCon-Chi has many similar images for which the same caption is valid. A drop from the original CLIP's performance on this metric (52.69) indicates forgetting of general knowledge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In Tab. 3, we compare our approach of learning internal parameter updates with learning input tokens in the same training setting. We consider Textual Inversion (TI), which learns a token that is integrated into input queries via a pseudo-word for the concept (e.g. \"A photo of V* jumping), and Prompt Tuning [43], which bridges the gap between TI and parameter updates by prepending learned prompt tokens to all queries (both personal and general). In contrast to TI, this allows us to use our VLM caption metric to measure the interference of the prompt tokens on general queries. We include results with 1 learned token as well as 2 learned tokens, which is equivalent in size to our rank-1 parameter update. With prompt tuning, we can achieve strong results on concept-only queries, but our VLM cap" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19752" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.91, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.457, + 0.908, + 0.515 + ], + "angle": 0, + "content": "Figure 3. We compare the top 3 retrievals of our method vs. SEARLE for personal queries in the ConCon-Chi and DeepFashion2 datasets, with green borders indicating correct retrievals. We observe that SEARLE struggles to differentiate between concepts of similar classes, such as different clothing items. Our method more consistently retrieves the correct concept in the correct context, demonstrating effective composition of personal and general knowledge." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.524, + 0.487, + 0.655 + ], + "angle": 0, + "content": "
MethodContext (Single)Concept-onlyVLM cap
mRRmAPmRRmAPr@10
Original CLIP29.3920.7610.756.4652.69
Text. Inv. (1 tok)42.4532.9397.5064.71N/A
Text. Inv. (2 tok)41.7327.94100.0063.88N/A
Prompt (1 tok)31.7720.7096.2558.9530.84
Prompt (2 tok)33.1420.93100.0064.4915.35
Text. Inv. + Ours39.2926.55100.0064.7252.57
Ours51.6436.73100.0068.7152.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.659, + 0.486, + 0.829 + ], + "angle": 0, + "content": "Table 3. We compare our approach, which updates the weights of the text encoder, to tuning input tokens on ConCon-Chi. Rows 2-3 represent Textual Inversion, where a learned token is applied in place of the personal concept in queries that reference the concept. Rows 4-5 tune prompt tokens that are prepended to all text queries. Row 6 learns both our parameter update and the token. Tuning tokens can achieve competitive results on concept-only queries but struggles on contextual queries that require composing personal and general knowledge. This catastrophic forgetting of general knowledge is reflected by our VLM caption matching metric. We propose a new approach that achieves strong performance on contextual, concept-only, and general queries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.903 + ], + "angle": 0, + "content": "tion metric shows that this strength comes at the cost of catastrophic forgetting of general knowledge. We also combine TI with our method by learning the input token in addition to our parameter update. We observe that learning the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.909, + 0.71 + ], + "angle": 0, + "content": "token reduces context performance, suggesting that the inserted learned token interferes with general knowledge. In contrast, our parameter updates achieve strong performance on personal queries while not interfering with CLIP's general knowledge, even when applied to non-personal queries. We hypothesize this is due to the minimally invasive structure of our updates, and that they are applied late in the encoding process in contrast to learned input tokens which influence the entire encoding process. We qualitatively illustrate this finding in Fig. 4; with the parameter update for a personal concept applied, the model successfully performs retrieval for both personalized queries and general queries." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.669, + 0.743 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "A key choice in developing our method is determining an effective, minimal set of parameters within the model for which to apply personalized updates. In this section, we empirically investigate rank size of the LoRA updates, which layers in the encoding process to apply LoRA updates, and on which parameters to learn LoRA updates. We additionally ablate our regularization and multi-concept merging strategies. All experiments use the CLIP ViT-L/14 architecture on the ConCon-Chi dataset with 5 training images, and we report our results on single-concept context queries" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19753" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.098, + 0.907, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.42, + 0.908, + 0.477 + ], + "angle": 0, + "content": "Figure 4. Our parameter updates enable personalized retrieval without overriding the model's general knowledge. On the left we show the top 3 retrievals of our personalized model on context queries referencing the personal concept, with green indicating ground truth correct retrievals. On the right, we show the results of querying the same personalized model for general VLM-generated captions. There is not exhaustive ground truth for which images match each caption; however, qualitatively our model retrieves appropriate images in all cases." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.487, + 0.479, + 0.595 + ], + "angle": 0, + "content": "
LoRA rankContext (Single-Concept)Concept-onlyVLM cap
mRRmAPr@1mRRmAPr@10
r=252.3136.5942.04100.0066.0752.78
r=451.5236.6041.36100.0068.1352.61
r=851.4936.5841.36100.0068.1552.62
r=1651.6736.6641.50100.0067.9352.62
r=151.6436.7341.77100.0068.7152.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.597, + 0.434, + 0.61 + ], + "angle": 0, + "content": "Table 4. Ablation of LoRA rank on ConCon-Chi." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.717 + ], + "angle": 0, + "content": "(we ablate the merging strategy for multi-concept queries in Sec. 4.4). We performed our model selection (design choices for our main method and regularization weight \\(\\lambda\\)) on the validation split (3 concepts) and report our main results on the test split." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.838 + ], + "angle": 0, + "content": "LoRA Rank. We ablate the rank of the LoRA updates in Tab. 4, which controls the number of learned parameters for each update. We observe only small retrieval accuracy gains on single-concept queries as the rank increases, and achieve better concept-only and competitive multi-concept results with rank=1. The rank-1 concept update is also the most parameter efficient, storing only \\(2d\\) parameters per concept, where \\(d\\) is the encoder's internal dimension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Architecture Layers. Tab. 5 investigates which transformer layers to apply LoRA updates to, dictating how early or late into the encoder process personalized information is injected. We achieve the strongest results by applying" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.488, + 0.915, + 0.595 + ], + "angle": 0, + "content": "
Layer(s)Context (Single-Concept)Concept-onlyVLM cap
mRRmAPr@1mRRmAPr@10
11,1250.1835.3439.73100.0064.0952.64
10,11,1251.5135.8141.63100.0065.8752.62
all layers43.2329.9332.5297.5063.7752.45
144.6931.1234.1597.5064.6652.18
1251.6436.7341.77100.0068.7152.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.557, + 0.599, + 0.86, + 0.613 + ], + "angle": 0, + "content": "Table 5. Ablation of LoRA layers on ConCon-Chi." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.839 + ], + "angle": 0, + "content": "personalization in the final layer (layer 12). We hypothesize that updating later layers is better than earlier layers because it allows our approach to apply a small, targeted update to the developed text query representation to inject personal information. In contrast, earlier layer updates are more likely to alter the full representation - not just the parts semantically belonging to the personal concept. Additionally, our findings align with works that suggest that the later layers in transformer encoders are the most important in constructing the final representation [11]. We also do not observe an overall benefit by learning LoRA updates for earlier layers in addition to the final layer; the final layer alone achieves best performance while also requiring the fewest learned parameters per concept." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Parameters. Tab. 6 ablates the component of the transformer layer on which the LoRA is learned. We consider the linear transforms within the attention mechanism (query, key, value, and output), the 2 MLP layers, as well as the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19754" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.491, + 0.244 + ], + "angle": 0, + "content": "
Param(s)Context (Single-Concept)Concept-onlyVLM cap
mRRmAPr@1mRRmAPr@10
Q16.6511.497.6232.6610.9151.84
K15.5511.456.5328.009.2852.12
O46.9931.0438.5097.5060.9852.52
Q,K,V,O47.5231.6938.7897.5060.9052.54
Q,V51.5035.6041.50100.0065.6052.65
MLP143.0427.9133.88100.0055.8752.05
MLP249.4032.8738.91100.0055.7051.65
final proj51.1534.9840.82100.0060.5851.58
V51.6436.7341.77100.0068.7152.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.248, + 0.454, + 0.262 + ], + "angle": 0, + "content": "Table 6. Ablation of LoRA parameters on ConCon-Chi." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.483, + 0.735 + ], + "angle": 0, + "content": "model's final projection layer. We see that the query and key transforms alone are not effective for personalization, indicating that directly transforming the output representation is crucial for personalization; re-weighting the existing tokens via altering the attention weight computation is not enough. The strongest results are achieved by updating the value transform. Interestingly, this setting outperforms updating the output transform and the following MLP layers. This result suggests that the value transform's placement in transforming the output of each attention head is optimal for personalization as opposed to later linear transforms that operate after the output of the attention heads is aggregated. We see negligible gains by pairing updates on the value transform with other parameters; in fact, this setting decreases the results on some metrics. Updating the final projection alone also produces competitive results, but lags on the concept-only metrics. Our analysis demonstrates that the value transform within the final layer is optimally positioned to learn a small, targeted update for personalization. Regularization. We ablate our regularization strategy in Tab. 7. Without regularization, we see drops on the contextual query performance and VLM caption metric, indicating forgetting of general knowledge. The concept-only mAP slightly increases, showing prioritization of personalization over retaining general knowledge. Our regularization strategies are complementary: using both together produces the best results. With our regularization scheme, we are able to produce strong results on contextual queries, avoid any degradation on general caption-matching performance, and still produce good concept-only performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Multi-Concept Merging. We consider alternative strategies for merging the low-rank parameter updates of different concepts in Tab. 8. Our hypothesis is that because we learn a single, constrained update for each concept, the updates for different concepts will be sufficiently different to not interfere with each other. Our results validate this; we see applying each update is better than altering them via averaging or pooling them together. We also consider Orthogonal Adapation [26], which constrains updates for different concepts to be orthogonal to eachother, thus avoiding interference. In this method, the \\(A\\) matrix in the low-rank update" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.925, + 0.178 + ], + "angle": 0, + "content": "
Reg.Context (Single-Concept)Concept-onlyVLM cap
ABmRRmAPr@1mRRmAPr@10
22.5114.3514.83100.0069.8952.52
✓33.7722.2925.44100.0069.6352.58
✓39.8426.4531.56100.0069.0152.57
✓✓51.6436.7341.77100.0068.7152.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.182, + 0.907, + 0.295 + ], + "angle": 0, + "content": "Table 7. Ablation of our regularization scheme: A denotes imposing the constraint \\( ||A_{L,c}||_2 = 1 \\), and B denotes applying the squared L2 penalty to \\( B_{L,c} \\) (Eq. 6). Without our regularization, the personalized parameter updates cause the model overfit to the concept, producing high concept-only metrics but catastrophically forgetting general knowledge, as reflected in the context and VLM caption metrics. Our regularization prevents this forgetting while still achieving high concept-only performance." + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.296, + 0.868, + 0.378 + ], + "angle": 0, + "content": "
Merge strategymRRmAPr@1
Avg LoRAs25.0315.5115.76
Max LoRAs25.2715.5616.05
Orthogonal Adaptation [26]28.3819.1418.62
Add LoRAs35.1323.0524.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.382, + 0.907, + 0.41 + ], + "angle": 0, + "content": "Table 8. Performance of different merging strategies for multi-concept queries in ConCon-Chi." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.909, + 0.743 + ], + "angle": 0, + "content": "is frozen and drawn from a shared orthogonal subspace, while the \\( B \\) matrix is learned. The updates for different concepts are merged by adding them together. Interestingly, we find Orthogonal Adaptation to be less effective than our approach; we hypothesize this is due to the differences between where low rank updates are applied in our method vs. the original Orthogonal Adaptation method, which operates on text to image diffusion models. Whereas in personalized generation, parameter updates are applied throughout the full model, we find learning only a single parameter update is better suited for retrieval. Because we only learn this one parameter update late in the model, certain randomly selected \\( A \\) matrices are not effective for personalization depending on how they interact with the incoming feature representation at that point in the model. Specifically, in computing \\( B_{L,c}A_{L,c}x \\) if the randomly chosen \\( A_{L,c} \\) is orthogonal to \\( x \\), this will eliminate the parameter update altogether. While personalized generation approaches avoid this by adapting the representation throughout all layers of the model, in our setting we find it is better to learn both \\( A \\) and \\( B \\) instead of imposing orthogonality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.634, + 0.77 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we show that updating the internal representation of the CLIP text encoder serves as a better alternative to textual inversion for personalized search. By constraining the parameter update for each concept to a single rank-one update in the value transform of the final layer and strategically regularizing the parameters, we demonstrate that our approach effectively personalizes from a few image examples while maintaining the model's general knowledge." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19755" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.212 + ], + "angle": 0, + "content": "Acknowledgments The authors thank the members of the Hoffman Lab at Georgia Tech and Jitesh Jain for their feedback on this work. Fiona Ryan is supported by the NSF Graduate Research Fellowship under Grant No. DGE-2039655. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors(s) and do not necessarily reflect the views of the National Science Foundation." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.224, + 0.188, + 0.24 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.248, + 0.486, + 0.304 + ], + "angle": 0, + "content": "[1] Yuval Alaluf, Elad Richardson, Sergey Tulyakov, Kfir Aberman, and Daniel Cohen-Or. Myvlm: Personalizing vlms for user-specific queries. In European Conference on Computer Vision, pages 73–91. Springer, 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.305, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4959-4968, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.484, + 0.443 + ], + "angle": 0, + "content": "[3] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21466-21474, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.484, + 0.512 + ], + "angle": 0, + "content": "[4] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15338-15347, 2023. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.484, + 0.581 + ], + "angle": 0, + "content": "[5] Adrian Bulat and Georgios Tzimiropoulos. Lasp: Text-to-text optimization for language-aware soft prompting of vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23232-23241, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.582, + 0.484, + 0.637 + ], + "angle": 0, + "content": "[6] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. arXiv preprint arXiv:2210.01253, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.638, + 0.484, + 0.706 + ], + "angle": 0, + "content": "[7] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In European conference on computer vision, pages 558-577. Springer, 2022. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.708, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[8] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor Guilherme Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Variational prompt tuning improves generalization of vision-language models. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[9] Ziyi Dong, Pengxu Wei, and Liang Lin. Dreamartist: Towards controllable one-shot text-to-image generation via positive-negative prompt-tuning. arXiv preprint arXiv:2211.11337, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.248, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[11] Yossi Gandelsman, Alexei A. Efros, and Jacob Steinhardt. Interpreting clip's image representation via text-based decomposition. In ICLR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.135, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.204, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[13] Yuying Ge, Ruimao Zhang, Xiaogang Wang, Xiaou Tang, and Ping Luo. Deepfashion2: A versatile benchmark for detection, pose estimation, segmentation and re-identification of clothing images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5337–5345, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[14] Cusuh Ham, Matthew Fisher, James Hays, Nicholas Kolkin, Yuchen Liu, Richard Zhang, and Tobias Hinz. Personalized residuals for concept-driven text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8186-8195, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[17] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[18] Bruno Korbar and Andrew Zisserman. Personalised clip or: how to find your vacation videos. 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[19] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.664, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[20] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1401-1411, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[21] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[23] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Joanna Materzynska, Josef Sivic, Eli Shechtman, Antonio Torralba, Richard Zhang, and Bryan Russell. Customizing" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19756" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "motion in text-to-video diffusion models. arXiv preprint arXiv:2312.04966, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[25] Thao Nguyen, Haotian Liu, Yuheng Li, Mu Cai, Utkarsh Ojha, and Yong Jae Lee. Yo'llava: Your personalized language and vision assistant. arXiv preprint arXiv:2406.09400, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[26] Ryan Po, Guandao Yang, Kfir Aberman, and Gordon Wetzstein. Orthogonal adaptation for modular customization of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7964-7973, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[28] Andrea Rosasco, Stefano Berti, Giulia Pasquale, Damiano Malafronte, Shogo Sato, Hiroyuki Segawa, Tetsugo Inada, and Lorenzo Natale. Concon-chi: Concept-context chimera benchmark for personalized vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22239-22248, 2024. 2, 4, 5, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[29] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6527-6536, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[31] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19305-19314, 2023. 2, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[32] Gabriel Skantze and Bram Willemsen. Collie: Continual learning of language grounding from language-image embeddings. Journal of Artificial Intelligence Research, 74: 1201-1223, 2022. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[33] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 conference proceedings, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[34] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[35] Nam Vo, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.135 + ], + "angle": 0, + "content": "for image retrieval—an empirical odyssey. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6439-6448, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[36] Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guoqiang Liang, Peng Wang, and Yanning Zhang. Dual modality prompt tuning for vision-language pre-trained model. IEEE Transactions on Multimedia, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.906, + 0.26 + ], + "angle": 0, + "content": "[37] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6757-6767, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.906, + 0.344 + ], + "angle": 0, + "content": "[38] Chun-Hsiao Yeh, Bryan Russell, Josef Sivic, Fabian Caba Heilbron, and Simon Jenni. Meta-personalizing vision-language models to find named instances in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19123-19132, 2023. 2, 3, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.906, + 0.416 + ], + "angle": 0, + "content": "[39] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the association for computational linguistics, 2:67-78, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.906, + 0.473 + ], + "angle": 0, + "content": "[40] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.906, + 0.543 + ], + "angle": 0, + "content": "[41] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaptation of clip for few-shot classification. In European conference on computer vision, pages 493-510. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.906, + 0.613 + ], + "angle": 0, + "content": "[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.615, + 0.906, + 0.669 + ], + "angle": 0, + "content": "[43] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.906, + 0.727 + ], + "angle": 0, + "content": "[44] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19757" + } + ] +] \ No newline at end of file diff --git a/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/673cbe39-bd0b-47dd-99b2-3c870c8defd5_origin.pdf b/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/673cbe39-bd0b-47dd-99b2-3c870c8defd5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2c65fc9d9e43b03ce0d9c091de77692085ce166e --- /dev/null +++ b/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/673cbe39-bd0b-47dd-99b2-3c870c8defd5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a189a9d381a990c946a517cdb2043fa6f17794da5d2c53ef4ad5387e17d92db +size 1942535 diff --git a/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/full.md b/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9e0c534db4c86f9a3b4a967a51ee76c87e37a377 --- /dev/null +++ b/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/full.md @@ -0,0 +1,267 @@ +# Improving Personalized Search with Regularized Low-Rank Parameter Updates + +Fiona Ryan $^{1,2*}$ , Josef Sivic $^{2,3}$ , Fabian Caba Heilbron $^{2}$ , Judy Hoffman $^{1}$ , James M. Rehg $^{4}$ , Bryan Russell $^{2}$ $^{1}$ Georgia Tech, $^{2}$ Adobe Research, $^{3}$ CIIRC CTU, $^{4}$ UIUC + +# Abstract + +Personalized vision-language retrieval seeks to recognize new concepts (e.g., "my dog Fido") from only a few examples. This task is challenging because it requires not only learning a new concept from a few images, but also integrating the personal and general knowledge together to recognize the concept in different contexts. In this paper, we show how to effectively adapt the internal representation of a vision-language dual encoder model for personalized vision-language retrieval. We find that regularized low-rank adaption of a small set of parameters in the language encoder's final layer serves as a highly effective alternative to textual inversion for recognizing the personal concept while preserving general knowledge. Additionally, we explore strategies for combining parameters of multiple learned personal concepts, finding that parameter addition is effective. To evaluate how well general knowledge is preserved in a finetuned representation, we introduce a metric that measures image retrieval accuracy based on captions generated by a vision language model (VLM). Our approach achieves state-of-the-art accuracy on two benchmarks for personalized image retrieval with natural language queries - DeepFashion2 and ConCon-Chi - outperforming the prior art by $4\% - 22\%$ on personal retrievals. + +# 1. Introduction + +Personalizing a vision-language retrieval model (PerVL) aims to adapt a pretrained vision-language dual encoder model (e.g., CLIP [27]) to recognize new concepts (e.g., "my dog Fido") from just a few examples [7]. This task is important for search applications that need to identify concepts missing from the pretrained model's knowledge, such as searching one's personal photo library for a specific person, object, or pet. PerVL is challenging because it requires not only learning a new concept from a few visual examples, but also reasoning about the personal concept and general knowledge together to retrieve the concept in different contexts. For instance, searching for "my dog Fido catching + +![](images/b780b0f30021761ab92fc36ae0d2961150b87d45d007e156646ec29100987976.jpg) +Figure 1. Left: Prior works use a pretrained textual inversion network $(\mathcal{F})$ to compute a pseudo-token to represent a new concept, which may be further optimized during personalization. Right: We present POLAR, which represents new concepts as a small low-rank parameter update within the text encoder. Instead of inserting a learned pseudo-token to the text query, we use a fixed vocabulary token $V^{*}$ . We show that our method is effective at recognizing the personal concept from a few examples while retaining the model's general knowledge, and does not require large scale pretraining. + +a frisbee" requires both personal knowledge ("Fido") and general knowledge ("catching a frisbee"). + +Recent approaches for PerVL use textual inversion to learn a pseudo-text token (e.g., “ $V^{*}$ ”) for each new personal concept by inverting from the target training images to the token [4, 7] (illustrated in Figure 1 (left)). These pseudo-text tokens require updating only a minimal set of parameters, and they can be used within a natural language query as input to the language encoder (e.g., “ $V^{*}$ catching a frisbee”). Typically, these approaches train a textual inversion network on large-scale data to predict a token representing a new concept, which may then be further optimized during personalization. While these approaches avoid updating the model's internal parameters, preventing the “overriding” of its general knowledge, their ability to represent the personal concept is limited to the single input text token. Furthermore, the token affects the entire text embedding process and can interfere with the language encoder's general knowledge. Consequently, textual inversion + +approaches often struggle to combine personalized and general knowledge across different datasets and can be time-consuming to optimize online due to the need for backpropagation through the full language encoder [7, 38], or require large-scale pretraining [4, 7, 38]. + +To address these shortcomings, we focus on updating the internal representation of CLIP's language encoder [27] using only a few training examples (illustrated in Figure 1 (right)). Updating the internal representation is challenging because the model must balance learning the personalized concept from limited examples while preventing catastrophic forgetting of its prior knowledge. Additionally, the training data does not inherently encourage the model to retain general knowledge. Our goal is to make this update efficiently, without relying on additional training data. + +Our contributions are fourfold. First, inspired by recent advances in text-to-image generation [14, 19, 29, 30], we show that CLIP's language encoder can be updated to learn a personalized concept from a few examples while retaining its general knowledge. We find that regularized low-rank adaptation (LoRA) finetuning [16] of the language encoder's last layer effectively balances the trade-off between learning the personalized concept and avoiding catastrophic forgetting. Our resulting method POLAR (PersOnalized Low-rank Adaptation for Retrieval) learns a low-rank parameter set that is minimal and comparable in size to the pseudo-text tokens used in textual inversion. Furthermore, we leverage LoRA's specific structure to introduce a regularization strategy that eliminates the need for additional training examples or regularization prompts. Unlike previous methods, our approach does not rely on any components pre-trained on large-scale data, allowing for seamless parameter updates using only a few training examples. + +Second, we introduce a new evaluation metric to assess how well general knowledge is preserved in our finetuned representation, using captions generated by a vision-language model (VLM). We find that our approach effectively maintains general knowledge. Third, we explore different strategies for combining learned representations for different personal concepts to support multi-concept queries. We find that adding LoRA representations is effective and outperforms orthogonal adaptation [26]. Finally, we demonstrate that POLAR achieves state-of-the-art accuracy on the DeepFashion2 [7] and ConCon-Chi [28] benchmarks, improving prior performance by $4\% - 22\%$ . + +# 2. Related Work + +Personalized Vision-Language Retrieval. Cohen et al. introduced the task of Personalized Vision-Language Retrieval [7] and proposed PALAVRA, a textual inversion approach for the task. PALAVRA first learns a textual inversion network on COCO, which takes a set of images of a concept and predicts an initial pseudo-word token to rep + +resent the concept in text queries. This token is then further optimized via backpropagation through CLIP's text encoder. Korbar et al. [18] propose a similar approach for the setting of retrieving specific people in videos. A few recent works explore learning embeddings to represent personal concepts for VLM captioning and QA tasks [1, 25]. Yeh et al. [38] build on the concept of textual inversion by meta-learning a basis for pseudo-tokens using large scale video data. However, they target a different setting where concepts are learned jointly, using the other concepts in the dataset as negative examples. Recently, Rosasco et al. introduced the Concept-Context Chimera dataset (ConConChi) for personalized retrieval, which provides a more rigorous benchmark for assessing retrieval of personal concepts across diverse contexts than prior datasets. In this work, we depart from prior work by representing new concepts as low rank parameter updates within the text encoder instead of pseudo-word tokens. We show that our approach more effectively composes personal and general knowledge, while requiring few parameters per-concept. + +Personalized Generation. Personalized generation is a more studied related task that generates new images of a personal concept using text-to-image diffusion models. Some approaches use textual inversion to optimize pseudoword tokens to use within text prompts [9, 10], while others like Dreambooth [29] and Custom Diffusion [19] find that tuning the weights of the diffusion U-net generates personal concepts with better fidelity. Recent work has focused on selectively tuning certain parameters to promote parameter efficiency and speed [9, 15, 19, 40], with some leveraging low rank constraints [14, 26, 30, 33]. Most related to our work is Perfusion [33], which learns rank-one updates to the diffusion U-net with a key-locking mechanism to constrain updates to the concept's spatial location in the feature map. While personalization via parameter updates has become commonplace for personalized generation, it has not yet been explored for personalized retrieval, motivating our work. Due to the differences in the nature of the tasks (generative vs. discriminative) and models (text-to-image diffusion vs. dual encoder), we find that retrieval demands a different strategy for applying internal parameter updates; updating even sparse sets of parameters can result in catastrophic forgetting of the model's general knowledge. Instead of updating parameters throughout the full model, we apply a single rank-one parameter update to the final layer of CLIP's text encoder and directly regularize these parameters to avoid catastrophic forgetting of general knowledge. + +Composed Image Retrieval. Another related task is composed image retrieval [35], which takes an image and textual modification as inputs and performs image retrieval. Approaches have leveraged CLIP for composed image retrieval [2, 3], with Pic2Word [31] and SEARLE [4] learning textual inversion networks to predict a pseudo-word token + +![](images/af97af57bc75b134048dfd51e46fb3793fda95a4869c19f0fb143ae523521a6e.jpg) +Figure 2. Overview of POLAR. For a personal concept, POLAR learns a rank-1 update to the value transform in the final layer of the text encoder. To maintain general knowledge during personalization, we impose a regularization loss on the update. + +for the input image. Importantly, composed image retrieval differs from personalized retrieval in that it does not require instance-level recognition of the same concept (e.g., retrieving the exact same person), and instead aims to retrieve images portraying a similar semantic class, layout, or style. + +Few-Shot Adaptation of Vision-Language Models. Our task also relates to work on adapting models like CLIP for few-shot classification. Prior works primarily leverage prompt tuning [21], which learns input tokens to represent new classes [5, 6, 8, 23, 34, 36, 37, 42-44]. A few works consider tuning within the encoder by applying prompts to all layers [17, 20] or augmenting the encoder with adapter modules [12, 41]. An important distinction is that while CLIP's general knowledge may aid with learning classes from few examples, the general knowledge is not required for these downstream classification tasks. In contrast, personalized retrieval requires not just recognizing the personal concept, but composing it with retained general knowledge. + +# 3. Personalized Low-Rank Adaptation for Retrieval (POLAR) + +# 3.1. Problem Formulation + +We follow the PerVL problem setup established by Cohen et al. [7]. Given a pretrained vision-language model $\psi$ , we aim to learn an adapted model $\psi'$ that is able to recognize a new personal concept $c$ (e.g., "my coffee mug"). The pretrained model $\psi$ consists of an image encoder $\psi_I$ and a text encoder $\psi_T$ . These encoders map image and textual inputs to a shared embedding space. At personalization time, we are given $N_c$ images $\{I_i^c\}_{i=1}^{N_c}$ depicting concept $c$ as training data. Following Cohen et al. [7], we are also given the class name $\mathcal{C}_c$ of concept $c$ (e.g., "mug"). At retrieval time, the input to the model consists of a textual query $q$ and a set of $N_r$ images $\{I_i^r\}_{i=1}^{N_r}$ constituting the retrieval database. We + +compute the language embedding for an input query $q$ as $\psi_T'(q)$ , and perform retrieval by computing the cosine similarity between $\psi_T'(q)$ and the image embedding for each retrieval image as + +$$ +\operatorname {s i m} \left(\psi_ {T} ^ {\prime} (q), \psi_ {I} ^ {\prime} \left(I _ {i} ^ {r}\right)\right) = \frac {\left\langle \psi_ {T} ^ {\prime} (q) , \psi_ {I} ^ {\prime} \left(I _ {i} ^ {r}\right) \right\rangle}{\left| \left| \psi_ {T} ^ {\prime} (q) \right| \right| _ {2} \left| \left| \psi_ {I} ^ {\prime} \left(I _ {i} ^ {r}\right) \right| \right| _ {2}} \tag {1} +$$ + +where $\langle \cdot \rangle$ denotes the inner product. The top retrieval for the query is the image $I_{i}^{r}$ corresponding to the embedding with the highest similarity to the query embedding $\psi_T^\prime (q)$ . + +# 3.2. Rank-One Personalized Value Updates + +In contrast to prior work that represents a concept as a learned input token for the text encoder [4, 7, 38], POLAR learns a low-rank parameter update to the text encoder for each concept (Fig. 2). We leverage LoRA [16]: for a weight $W \in \mathbb{R}^{m \times n}$ in a pretrained model, LoRA learns a low-rank update $\Delta W$ , performing a modified forward pass as + +$$ +y = (W + \Delta W) x = W x + B A x \tag {2} +$$ + +where $B \in \mathbb{R}^{m \times r}$ , $A \in \mathbb{R}^{r \times n}$ , and $r < \min(m, n)$ is the chosen rank of the weight update. By selecting $r << \min(m, n)$ , LoRA tunes minimal parameters in the original model $\psi$ . For each concept $c$ , we empirically choose to learn a rank-one ( $r = 1$ ) update to the value transform of the final attention layer $L$ in $\psi_T$ . Our choice of $r = 1$ reflects our goal to represent a single concept from very limited examples, while minimally interfering with the model's existing knowledge. Let $Q_L, K_L, V_L \in \mathbb{R}^{d \times d}$ be the pretrained text encoder's query, key, and value transforms for the attention mechanism in the final transformer layer $L$ , where $d$ is the internal dimension of $\psi_T$ . For each attention head, the output of the multi-head attention layer is calculated as + +$$ +\operatorname {A t t e n t i o n} (x) = \operatorname {s o f t m a x} \left(\frac {Q _ {L} x \left(K _ {L} x\right) ^ {T}}{\sqrt {d}}\right) V _ {L, c} ^ {\prime} x \tag {3} +$$ + +where $V_{L,c}^{\prime}$ is the value transform updated with LoRA as: + +$$ +V _ {L, c} ^ {\prime} = V _ {L} + B _ {L, c} A _ {L, c}. \tag {4} +$$ + +$V_{L,c}^{\prime}$ is learned separately for each concept $c$ . Following prior work on personalized generation [14, 19], we choose a fixed token (e.g. "sks") in CLIP's vocabulary as a placeholder for the concept's place in the input queries. During training, we randomly select a template textual query (e.g. "An image of sks") to associate with each training image to form a set of text-image pairs $\{q_i^c, I_i^c\}_{i=1}^{N_c}$ . We supervise the learning of $V_{L,c}^{\prime}$ by using a mean-squared error (MSE) loss to push the normalized text and image embeddings for each training pair close together in the embedding space: + +$$ +\mathcal {L} _ {\mathrm {M S E}} = \frac {1}{N _ {c}} \sum_ {i = 1} ^ {N _ {c}} \left(\frac {\psi_ {T , c} ^ {\prime} (q _ {i})}{| | \psi_ {T , c} ^ {\prime} (q _ {i}) | | _ {2}} - \frac {\psi_ {I} \left(I _ {i} ^ {c}\right)}{| | \psi_ {I} \left(I _ {i} ^ {c}\right) | | _ {2}}\right) ^ {2} \tag {5} +$$ + +where $\psi_T^\prime ,c$ represents the text encoder with concept $c$ 's LoRA update applied in the forward pass (Eq. (3)). + +# 3.3. Regularization + +A key challenge for personalized retrieval is to learn update weights that supply necessary personalized information without overriding the model's general knowledge, which is necessary for retrieving the personal concept in different contexts (e.g., "my dog Fido catching a frisbee" and "my dog Fido sitting on the couch"). We therefore propose a regularization scheme for POLAR. Differently than prior works that construct additional regularization examples to use during training [4, 7], we instead exploit the structure of our low-rank updates to directly minimize updates to the original representation. From Eq. (4), our parameter update alters the representation $V_{L}x$ by adding the term $B_{L,c}A_{L,c}x$ . With our use of rank $r = 1$ , we can interpret $A_{L,c}x$ as computing dot product similarity between the vectors $A_{L,c}^{T}$ and $x$ , which determines the scale of an added directional update $B_{L,c}$ . Our regularization comprises two components: first, we add a penalty on the size of the weights in $B_{L,c}$ to avoid unnecessary deviation from CLIP's existing representation, i.e., when the term $B_{L,c}A_{L,c}x = 0$ , the text embedding will be the same as CLIP's original representation $(\psi_T'(q) = \psi_T(q))$ . We modify our loss by adding a squared- $L_{2}$ regularization over the weights $B_{L,c}$ : + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {M S E}} + \lambda \mathcal {L} _ {\mathrm {r e g}}, \quad \mathcal {L} _ {\mathrm {r e g}} = | B _ {L, c} | ^ {2} \tag {6} +$$ + +where $\lambda$ is a tunable hyperparameter that determines the relative weight of regularization in the loss. Second, we impose the constraint $||A_{L,c}||_2 = 1$ , encouraging $A_{L,c}$ to learn to selectively identify when to apply personal context based on directional similarity with the incoming representation $x$ , while the magnitude of the personal update is controlled entirely by the regularized $B_{L,c}$ . + +# 3.4. Merging parameters for multi-concept queries + +For queries that reference multiple personal concepts (e.g., "my dog Fido is playing with Rex's favorite frisbee"), we propose merging the parameter updates for the concepts into one weight update that is applied during encoding to provide personal context for both concepts. Let $V_{L,c_1}'$ and $V_{L,c_2}'$ be the individually learned weight updates for concepts $c_1$ and $c_2$ . We construct a combined weight update as: + +$$ +V _ {L, c _ {1} + c _ {2}} ^ {\prime} = V _ {L, c _ {1}} ^ {\prime} + V _ {L, c _ {2}} ^ {\prime}. \tag {7} +$$ + +During the forward pass, this setup equates to adding the parameter updates for both personal concepts to the representation. This update is also equivalent to constructing a rank $r = 2$ update via matrix concatenation along the rank dimension as: + +$$ +V _ {L, c _ {1} + c _ {2}} ^ {\prime} = \left[ \begin{array}{l l} B _ {L, c _ {1}} & B _ {L, c _ {2}} \end{array} \right] \left[ \begin{array}{l} A _ {L, c _ {1}} \\ A _ {L, c _ {2}} \end{array} \right]. \tag {8} +$$ + +This approach also generalizes to merging greater than two concepts. We explore other merging strategies in Tab. 8. + +# 4. Experiments + +In this section, we evaluate POLAR and compare it to existing works. We also provide ablations to give insight into the design choices within our method. We include further experiment details, analysis of personalization time, and discussion of limitations in the supplemental material. + +# 4.1. Datasets + +DeepFashion2. Cohen et al. [7] define a personalized retrieval benchmark on the DeepFashion2 dataset [13], where the 50 personal concepts are different clothing items. The test set includes 221 captions and images for retrieval. ConCon-Chi. The recent ConCon-Chi dataset aims to more comprehensively evaluate unique personal concepts in a variety of contexts; we thus use it primarily for our analysis. It consists of 20 concepts including household objects and chimeric concepts (combinations of multiple objects). There are 1084 context queries (735 single-concept, 349 multi-concept), and 4008 retrieval images. For direct comparisons to the original baselines [28], we report on the full TEST set, which includes 3 validation concepts. We verify our gains hold on the TEST-UNSEEN split, which excludes these concepts, in the supplemental. + +# 4.2. Evaluation Protocol + +Context Queries. Context queries perform retrieval on a caption referencing the personal concept in a particular context (e.g., "my dog Fido catching a frisbee in the backyard"). The ground truth consists of the images labeled as matching this prompt. For DeepFashion2, there is one ground truth image per context query, while for ConCon-Chi there are 1-130 ground truth images (average $\approx 6$ ). + +Concept-only Queries. Following Yeh et al. [38], we also report retrieval accuracy on "concept-only" queries to evaluate the model's ability to recognize the personal concept independent of context. For each concept, we use the input query "An image of $\mathrm{V}^{*}$ " and compute retrieval metrics where the ground truth is all retrieval images that contain the concept. For ConCon-Chi, we include only single-concept images (2430 images) in the retrieval database. + +Metrics. We report retrieval accuracy using standard benchmark metrics for the datasets: mean reciprocal rank (mRR) – the average inverse rank of the first retrieved ground truth image; recall-at- $k$ ( $\mathbf{r}@\mathbf{k}$ ) – the average success rate within the top $k$ retrievals; and mean average precision (mAP) – the area under the precision-recall curve, averaged over all queries. We report mAP for settings with multiple ground truth images per query (concept-only queries on both datasets, and context queries on ConCon-Chi). + +
MethodArch.ContextConcept-only
mRRrecall@5mRRmAP
AdapterViT-B/325.9---
COLLIE [32]ViT-B/327.9---
Text OnlyViT-B/3217.6---
AvgIm + TextViT-B/3218.8---
PALAVRA [7]ViT-B/3228.439.2--
SEARLE [4]ViT-B/3221.9027.1512.7425.97
OursViT-B/3234.8244.8859.2628.75
SEARLE [4]ViT-L/1427.6234.1232.0716.17
OursViT-L/1440.7251.3165.9635.07
+ +Table 1. Comparison to prior work on the DeepFashion2 retrieval benchmark with 5 training images per concept. We report the mean over over 5 runs with 5 randomly chosen training images of the concept per run (see supplemental for standard error). + +
MethodmRRContextConcept-only
mAPrecall@1mRRmAP
Coarse (class name)24.2116.8314.48--
\(Discriminative^†\)43.1630.1631.92--
\(Rich^†\)40.5827.6529.98--
PALAVRA [7]35.9923.5926.75--
Pic2Word [31]38.6226.3927.68--
SEARLE [4]43.9330.7433.4996.6761.94
Ours46.3332.3336.16100.0068.71
+ +(a) Comparison to prior work on the ConCon-Chi benchmark. + +
MethodContext (Single-concept)Context (Multi-concept)
mRRmAPr@1mRRmAPr@1
SEARLE49.5035.2539.0532.0621.2221.78
Ours51.6436.7341.7735.1323.0524.36
+ +(b) Results for ConCon-Chi single-concept vs. multi-concept queries. + +Table 2. Our approach achieves state of the art results on the challenging ConCon-Chi benchmark on all metrics. We also break down the results of our method and SEARLE[4] by single-concept and multi-concept queries, demonstrating best results on both. † refer to ConCon-Chi's provided text descriptors for each concept, which serve as oracles since they use knowledge of all concepts to manually determine a differentiating description. + +Implementation Details. For our main method on ConCon-Chi, we use $\lambda = 0.35$ . We train for 500 iterations with learning rate 0.001 and the Adam optimizer. Our model converges within 50 epochs. Because we optimize minimal parameters and backpropagate through only the final layer, personalization is fast, taking under 1 second on a V100 GPU. On DeepFashion2 we append the classname to $\mathrm{V}^*$ (e.g., "sks dress") like in PALAVRA [7] and use use $\lambda = 0.1$ . We use the same template prompts as PALAVRA for training. We provide further details in the supplemental. + +# 4.3. Comparison to Prior Work + +We compare our method to prior work on DeepFashion2 in Tab. 1 and ConCon-Chi in Tab. 2. For DeepFashion2 we compare against PALAVRA and the baselines reported in its paper [7]. We also run SEARLE [4], a zero-shot com + +posed image retrieval network, using the publicly available checkpoints. Our method achieves state-of-the-art results in the standard setting using the CLIP ViT-B/32 architecture. We also see improvement over SEARLE when using the larger CLIP ViT-L/14 architecture. + +On ConCon-Chi, we compare against the baselines reported in the benchmark's paper [28]. All methods use the CLIP ViT-L/14 architecture. Differently from DeepFashion2, the zero-shot composed image retrieval methods (SEARLE and Pic2Word) outperform PALAVRA, which suggests that while they are effective in some settings, they struggle with differentiating between several similar concepts like the clothing items in DeepFashion2. This is qualitatively demonstrated in Fig. 3, which compares our method's retrievals with SEARLE. Our method, however, also achieves state-of-the-art results on ConCon-Chi, demonstrating flexibility across different benchmarks. Additionally, we achieve stronger mAP for the concept-only queries task than SEARLE. + +Evaluating General Knowledge in LoRA vs. Token Learning. Our primary hypothesis is that learning a small, regularized parameter update within the encoding process can more effectively allow CLIP to reason about personal and general information together. While concept-only queries assess personal knowledge and context queries assess the combination of personal and general knowledge, our setting of tuning the text encoder's parameters also allows us to measure the retention of general knowledge. We do so by inputting general queries, which do not reference the personal concept, to the text encoder with our parameter update for the concept still applied. To source non-personal queries, we use LLaVA [22] to caption each image in the retrieval set. We define a new metric VLM caption recall@10, which measures the success rate of retrieving the image for which the caption was generated in the top 10 retrievals. We choose 10 retrievals because ConCon-Chi has many similar images for which the same caption is valid. A drop from the original CLIP's performance on this metric (52.69) indicates forgetting of general knowledge. + +In Tab. 3, we compare our approach of learning internal parameter updates with learning input tokens in the same training setting. We consider Textual Inversion (TI), which learns a token that is integrated into input queries via a pseudo-word for the concept (e.g. "A photo of V* jumping), and Prompt Tuning [43], which bridges the gap between TI and parameter updates by prepending learned prompt tokens to all queries (both personal and general). In contrast to TI, this allows us to use our VLM caption metric to measure the interference of the prompt tokens on general queries. We include results with 1 learned token as well as 2 learned tokens, which is equivalent in size to our rank-1 parameter update. With prompt tuning, we can achieve strong results on concept-only queries, but our VLM cap + +![](images/ea0993fae84a031f1cd3caf3267b5a78d909615879868824a6f00abca322eda7.jpg) +Figure 3. We compare the top 3 retrievals of our method vs. SEARLE for personal queries in the ConCon-Chi and DeepFashion2 datasets, with green borders indicating correct retrievals. We observe that SEARLE struggles to differentiate between concepts of similar classes, such as different clothing items. Our method more consistently retrieves the correct concept in the correct context, demonstrating effective composition of personal and general knowledge. + +
MethodContext (Single)Concept-onlyVLM cap
mRRmAPmRRmAPr@10
Original CLIP29.3920.7610.756.4652.69
Text. Inv. (1 tok)42.4532.9397.5064.71N/A
Text. Inv. (2 tok)41.7327.94100.0063.88N/A
Prompt (1 tok)31.7720.7096.2558.9530.84
Prompt (2 tok)33.1420.93100.0064.4915.35
Text. Inv. + Ours39.2926.55100.0064.7252.57
Ours51.6436.73100.0068.7152.62
+ +Table 3. We compare our approach, which updates the weights of the text encoder, to tuning input tokens on ConCon-Chi. Rows 2-3 represent Textual Inversion, where a learned token is applied in place of the personal concept in queries that reference the concept. Rows 4-5 tune prompt tokens that are prepended to all text queries. Row 6 learns both our parameter update and the token. Tuning tokens can achieve competitive results on concept-only queries but struggles on contextual queries that require composing personal and general knowledge. This catastrophic forgetting of general knowledge is reflected by our VLM caption matching metric. We propose a new approach that achieves strong performance on contextual, concept-only, and general queries. + +tion metric shows that this strength comes at the cost of catastrophic forgetting of general knowledge. We also combine TI with our method by learning the input token in addition to our parameter update. We observe that learning the + +token reduces context performance, suggesting that the inserted learned token interferes with general knowledge. In contrast, our parameter updates achieve strong performance on personal queries while not interfering with CLIP's general knowledge, even when applied to non-personal queries. We hypothesize this is due to the minimally invasive structure of our updates, and that they are applied late in the encoding process in contrast to learned input tokens which influence the entire encoding process. We qualitatively illustrate this finding in Fig. 4; with the parameter update for a personal concept applied, the model successfully performs retrieval for both personalized queries and general queries. + +# 4.4. Ablation Study + +A key choice in developing our method is determining an effective, minimal set of parameters within the model for which to apply personalized updates. In this section, we empirically investigate rank size of the LoRA updates, which layers in the encoding process to apply LoRA updates, and on which parameters to learn LoRA updates. We additionally ablate our regularization and multi-concept merging strategies. All experiments use the CLIP ViT-L/14 architecture on the ConCon-Chi dataset with 5 training images, and we report our results on single-concept context queries + +![](images/f8975b6b5a8baeb0bd39df58f0a8361a4085aa6cb77cdb475075ab28549c7b2e.jpg) +Figure 4. Our parameter updates enable personalized retrieval without overriding the model's general knowledge. On the left we show the top 3 retrievals of our personalized model on context queries referencing the personal concept, with green indicating ground truth correct retrievals. On the right, we show the results of querying the same personalized model for general VLM-generated captions. There is not exhaustive ground truth for which images match each caption; however, qualitatively our model retrieves appropriate images in all cases. + +
LoRA rankContext (Single-Concept)Concept-onlyVLM cap
mRRmAPr@1mRRmAPr@10
r=252.3136.5942.04100.0066.0752.78
r=451.5236.6041.36100.0068.1352.61
r=851.4936.5841.36100.0068.1552.62
r=1651.6736.6641.50100.0067.9352.62
r=151.6436.7341.77100.0068.7152.62
+ +(we ablate the merging strategy for multi-concept queries in Sec. 4.4). We performed our model selection (design choices for our main method and regularization weight $\lambda$ ) on the validation split (3 concepts) and report our main results on the test split. + +LoRA Rank. We ablate the rank of the LoRA updates in Tab. 4, which controls the number of learned parameters for each update. We observe only small retrieval accuracy gains on single-concept queries as the rank increases, and achieve better concept-only and competitive multi-concept results with rank=1. The rank-1 concept update is also the most parameter efficient, storing only $2d$ parameters per concept, where $d$ is the encoder's internal dimension. + +Architecture Layers. Tab. 5 investigates which transformer layers to apply LoRA updates to, dictating how early or late into the encoder process personalized information is injected. We achieve the strongest results by applying + +Table 4. Ablation of LoRA rank on ConCon-Chi. + +
Layer(s)Context (Single-Concept)Concept-onlyVLM cap
mRRmAPr@1mRRmAPr@10
11,1250.1835.3439.73100.0064.0952.64
10,11,1251.5135.8141.63100.0065.8752.62
all layers43.2329.9332.5297.5063.7752.45
144.6931.1234.1597.5064.6652.18
1251.6436.7341.77100.0068.7152.62
+ +Table 5. Ablation of LoRA layers on ConCon-Chi. + +personalization in the final layer (layer 12). We hypothesize that updating later layers is better than earlier layers because it allows our approach to apply a small, targeted update to the developed text query representation to inject personal information. In contrast, earlier layer updates are more likely to alter the full representation - not just the parts semantically belonging to the personal concept. Additionally, our findings align with works that suggest that the later layers in transformer encoders are the most important in constructing the final representation [11]. We also do not observe an overall benefit by learning LoRA updates for earlier layers in addition to the final layer; the final layer alone achieves best performance while also requiring the fewest learned parameters per concept. + +Parameters. Tab. 6 ablates the component of the transformer layer on which the LoRA is learned. We consider the linear transforms within the attention mechanism (query, key, value, and output), the 2 MLP layers, as well as the + +
Param(s)Context (Single-Concept)Concept-onlyVLM cap
mRRmAPr@1mRRmAPr@10
Q16.6511.497.6232.6610.9151.84
K15.5511.456.5328.009.2852.12
O46.9931.0438.5097.5060.9852.52
Q,K,V,O47.5231.6938.7897.5060.9052.54
Q,V51.5035.6041.50100.0065.6052.65
MLP143.0427.9133.88100.0055.8752.05
MLP249.4032.8738.91100.0055.7051.65
final proj51.1534.9840.82100.0060.5851.58
V51.6436.7341.77100.0068.7152.62
+ +model's final projection layer. We see that the query and key transforms alone are not effective for personalization, indicating that directly transforming the output representation is crucial for personalization; re-weighting the existing tokens via altering the attention weight computation is not enough. The strongest results are achieved by updating the value transform. Interestingly, this setting outperforms updating the output transform and the following MLP layers. This result suggests that the value transform's placement in transforming the output of each attention head is optimal for personalization as opposed to later linear transforms that operate after the output of the attention heads is aggregated. We see negligible gains by pairing updates on the value transform with other parameters; in fact, this setting decreases the results on some metrics. Updating the final projection alone also produces competitive results, but lags on the concept-only metrics. Our analysis demonstrates that the value transform within the final layer is optimally positioned to learn a small, targeted update for personalization. Regularization. We ablate our regularization strategy in Tab. 7. Without regularization, we see drops on the contextual query performance and VLM caption metric, indicating forgetting of general knowledge. The concept-only mAP slightly increases, showing prioritization of personalization over retaining general knowledge. Our regularization strategies are complementary: using both together produces the best results. With our regularization scheme, we are able to produce strong results on contextual queries, avoid any degradation on general caption-matching performance, and still produce good concept-only performance. + +Multi-Concept Merging. We consider alternative strategies for merging the low-rank parameter updates of different concepts in Tab. 8. Our hypothesis is that because we learn a single, constrained update for each concept, the updates for different concepts will be sufficiently different to not interfere with each other. Our results validate this; we see applying each update is better than altering them via averaging or pooling them together. We also consider Orthogonal Adapation [26], which constrains updates for different concepts to be orthogonal to eachother, thus avoiding interference. In this method, the $A$ matrix in the low-rank update + +Table 6. Ablation of LoRA parameters on ConCon-Chi. + +
Reg.Context (Single-Concept)Concept-onlyVLM cap
ABmRRmAPr@1mRRmAPr@10
22.5114.3514.83100.0069.8952.52
✓33.7722.2925.44100.0069.6352.58
✓39.8426.4531.56100.0069.0152.57
✓✓51.6436.7341.77100.0068.7152.62
+ +Table 7. Ablation of our regularization scheme: A denotes imposing the constraint $||A_{L,c}||_2 = 1$ , and B denotes applying the squared L2 penalty to $B_{L,c}$ (Eq. 6). Without our regularization, the personalized parameter updates cause the model overfit to the concept, producing high concept-only metrics but catastrophically forgetting general knowledge, as reflected in the context and VLM caption metrics. Our regularization prevents this forgetting while still achieving high concept-only performance. + +
Merge strategymRRmAPr@1
Avg LoRAs25.0315.5115.76
Max LoRAs25.2715.5616.05
Orthogonal Adaptation [26]28.3819.1418.62
Add LoRAs35.1323.0524.36
+ +Table 8. Performance of different merging strategies for multi-concept queries in ConCon-Chi. + +is frozen and drawn from a shared orthogonal subspace, while the $B$ matrix is learned. The updates for different concepts are merged by adding them together. Interestingly, we find Orthogonal Adaptation to be less effective than our approach; we hypothesize this is due to the differences between where low rank updates are applied in our method vs. the original Orthogonal Adaptation method, which operates on text to image diffusion models. Whereas in personalized generation, parameter updates are applied throughout the full model, we find learning only a single parameter update is better suited for retrieval. Because we only learn this one parameter update late in the model, certain randomly selected $A$ matrices are not effective for personalization depending on how they interact with the incoming feature representation at that point in the model. Specifically, in computing $B_{L,c}A_{L,c}x$ if the randomly chosen $A_{L,c}$ is orthogonal to $x$ , this will eliminate the parameter update altogether. While personalized generation approaches avoid this by adapting the representation throughout all layers of the model, in our setting we find it is better to learn both $A$ and $B$ instead of imposing orthogonality. + +# 5. Conclusion + +In this work, we show that updating the internal representation of the CLIP text encoder serves as a better alternative to textual inversion for personalized search. By constraining the parameter update for each concept to a single rank-one update in the value transform of the final layer and strategically regularizing the parameters, we demonstrate that our approach effectively personalizes from a few image examples while maintaining the model's general knowledge. + +Acknowledgments The authors thank the members of the Hoffman Lab at Georgia Tech and Jitesh Jain for their feedback on this work. Fiona Ryan is supported by the NSF Graduate Research Fellowship under Grant No. DGE-2039655. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors(s) and do not necessarily reflect the views of the National Science Foundation. + +# References + +[1] Yuval Alaluf, Elad Richardson, Sergey Tulyakov, Kfir Aberman, and Daniel Cohen-Or. Myvlm: Personalizing vlms for user-specific queries. In European Conference on Computer Vision, pages 73–91. Springer, 2025. 2, 5 +[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4959-4968, 2022. 2 +[3] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21466-21474, 2022. 2 +[4] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15338-15347, 2023. 1, 2, 3, 4, 5 +[5] Adrian Bulat and Georgios Tzimiropoulos. Lasp: Text-to-text optimization for language-aware soft prompting of vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23232-23241, 2023. 3 +[6] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. arXiv preprint arXiv:2210.01253, 2022. 3 +[7] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. "this is my unicorn, fluffy": Personalizing frozen vision-language representations. In European conference on computer vision, pages 558-577. Springer, 2022. 1, 2, 3, 4, 5 +[8] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor Guilherme Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Variational prompt tuning improves generalization of vision-language models. 2022. 3 +[9] Ziyi Dong, Pengxu Wei, and Liang Lin. Dreamartist: Towards controllable one-shot text-to-image generation via positive-negative prompt-tuning. arXiv preprint arXiv:2211.11337, 2022. 2 +[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 + +[11] Yossi Gandelsman, Alexei A. Efros, and Jacob Steinhardt. Interpreting clip's image representation via text-based decomposition. In ICLR, 2023. 7 +[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3 +[13] Yuying Ge, Ruimao Zhang, Xiaogang Wang, Xiaou Tang, and Ping Luo. Deepfashion2: A versatile benchmark for detection, pose estimation, segmentation and re-identification of clothing images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5337–5345, 2019. 4 +[14] Cusuh Ham, Matthew Fisher, James Hays, Nicholas Kolkin, Yuchen Liu, Richard Zhang, and Tobias Hinz. Personalized residuals for concept-driven text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8186-8195, 2024. 2, 3, 4 +[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 2 +[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3 +[17] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 3 +[18] Bruno Korbar and Andrew Zisserman. Personalised clip or: how to find your vacation videos. 2022. 2, 5 +[19] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2, 3, 4 +[20] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1401-1411, 2023. 3 +[21] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 3 +[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5, 3 +[23] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 3 +[24] Joanna Materzynska, Josef Sivic, Eli Shechtman, Antonio Torralba, Richard Zhang, and Bryan Russell. Customizing + +motion in text-to-video diffusion models. arXiv preprint arXiv:2312.04966, 2023. 4 +[25] Thao Nguyen, Haotian Liu, Yuheng Li, Mu Cai, Utkarsh Ojha, and Yong Jae Lee. Yo'llava: Your personalized language and vision assistant. arXiv preprint arXiv:2406.09400, 2024. 2 +[26] Ryan Po, Guandao Yang, Kfir Aberman, and Gordon Wetzstein. Orthogonal adaptation for modular customization of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7964-7973, 2024. 2, 8 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2 +[28] Andrea Rosasco, Stefano Berti, Giulia Pasquale, Damiano Malafronte, Shogo Sato, Hiroyuki Segawa, Tetsugo Inada, and Lorenzo Natale. Concon-chi: Concept-context chimera benchmark for personalized vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22239-22248, 2024. 2, 4, 5, 1, 3 +[29] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2 +[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6527-6536, 2024. 2 +[31] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19305-19314, 2023. 2, 5, 3 +[32] Gabriel Skantze and Bram Willemsen. Collie: Continual learning of language grounding from language-image embeddings. Journal of Artificial Intelligence Research, 74: 1201-1223, 2022. 5, 2 +[33] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 conference proceedings, pages 1-11, 2023. 2 +[34] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 3 +[35] Nam Vo, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image + +for image retrieval—an empirical odyssey. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6439-6448, 2019. 2 +[36] Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guoqiang Liang, Peng Wang, and Yanning Zhang. Dual modality prompt tuning for vision-language pre-trained model. IEEE Transactions on Multimedia, 2023. 3 +[37] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6757-6767, 2023. 3 +[38] Chun-Hsiao Yeh, Bryan Russell, Josef Sivic, Fabian Caba Heilbron, and Simon Jenni. Meta-personalizing vision-language models to find named instances in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19123-19132, 2023. 2, 3, 4, 1 +[39] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the association for computational linguistics, 2:67-78, 2014. 4 +[40] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926, 2023. 2 +[41] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaptation of clip for few-shot classification. In European conference on computer vision, pages 493-510. Springer, 2022. 3 +[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 3 +[43] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 5 +[44] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 3 \ No newline at end of file diff --git a/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/images.zip b/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cd496a2fb2be53a27caa8fbdd47ec0ddc2209318 --- /dev/null +++ b/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcf44c74afe0bfe1dc22cff64c05b3c2b95072859fd4643bcda967ffe4cc1363 +size 664150 diff --git a/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/layout.json b/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a875552bb5e041d4ab89bcc9fb486d729eebcdc1 --- /dev/null +++ b/2025/Improving Personalized Search with Regularized Low-Rank Parameter Updates/layout.json @@ -0,0 +1,7711 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 553, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 553, + 121 + ], + "type": "text", + "content": "Improving Personalized Search with Regularized Low-Rank Parameter Updates" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "text", + "content": "Fiona Ryan" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "text", + "content": ", Josef Sivic" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "text", + "content": ", Fabian Caba Heilbron" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "text", + "content": ", Judy Hoffman" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "text", + "content": ", James M. Rehg" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "text", + "content": ", Bryan Russell" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "text", + "content": "Georgia Tech, " + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "text", + "content": "Adobe Research, " + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "text", + "content": "CIIRC CTU, " + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 142, + 563, + 172 + ], + "type": "text", + "content": "UIUC" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 211, + 200, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 211, + 200, + 224 + ], + "spans": [ + { + "bbox": [ + 151, + 211, + 200, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 236, + 297, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 236, + 297, + 513 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 297, + 513 + ], + "type": "text", + "content": "Personalized vision-language retrieval seeks to recognize new concepts (e.g., \"my dog Fido\") from only a few examples. This task is challenging because it requires not only learning a new concept from a few images, but also integrating the personal and general knowledge together to recognize the concept in different contexts. In this paper, we show how to effectively adapt the internal representation of a vision-language dual encoder model for personalized vision-language retrieval. We find that regularized low-rank adaption of a small set of parameters in the language encoder's final layer serves as a highly effective alternative to textual inversion for recognizing the personal concept while preserving general knowledge. Additionally, we explore strategies for combining parameters of multiple learned personal concepts, finding that parameter addition is effective. To evaluate how well general knowledge is preserved in a finetuned representation, we introduce a metric that measures image retrieval accuracy based on captions generated by a vision language model (VLM). Our approach achieves state-of-the-art accuracy on two benchmarks for personalized image retrieval with natural language queries - DeepFashion2 and ConCon-Chi - outperforming the prior art by " + }, + { + "bbox": [ + 54, + 236, + 297, + 513 + ], + "type": "inline_equation", + "content": "4\\% - 22\\%" + }, + { + "bbox": [ + 54, + 236, + 297, + 513 + ], + "type": "text", + "content": " on personal retrievals." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 549, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 296, + 693 + ], + "type": "text", + "content": "Personalizing a vision-language retrieval model (PerVL) aims to adapt a pretrained vision-language dual encoder model (e.g., CLIP [27]) to recognize new concepts (e.g., \"my dog Fido\") from just a few examples [7]. This task is important for search applications that need to identify concepts missing from the pretrained model's knowledge, such as searching one's personal photo library for a specific person, object, or pet. PerVL is challenging because it requires not only learning a new concept from a few visual examples, but also reasoning about the personal concept and general knowledge together to retrieve the concept in different contexts. For instance, searching for \"my dog Fido catching" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 200, + 557, + 361 + ], + "blocks": [ + { + "bbox": [ + 319, + 200, + 557, + 361 + ], + "lines": [ + { + "bbox": [ + 319, + 200, + 557, + 361 + ], + "spans": [ + { + "bbox": [ + 319, + 200, + 557, + 361 + ], + "type": "image", + "image_path": "b780b0f30021761ab92fc36ae0d2961150b87d45d007e156646ec29100987976.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 370, + 555, + 470 + ], + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 470 + ], + "type": "text", + "content": "Figure 1. Left: Prior works use a pretrained textual inversion network " + }, + { + "bbox": [ + 313, + 370, + 555, + 470 + ], + "type": "inline_equation", + "content": "(\\mathcal{F})" + }, + { + "bbox": [ + 313, + 370, + 555, + 470 + ], + "type": "text", + "content": " to compute a pseudo-token to represent a new concept, which may be further optimized during personalization. Right: We present POLAR, which represents new concepts as a small low-rank parameter update within the text encoder. Instead of inserting a learned pseudo-token to the text query, we use a fixed vocabulary token " + }, + { + "bbox": [ + 313, + 370, + 555, + 470 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 313, + 370, + 555, + 470 + ], + "type": "text", + "content": ". We show that our method is effective at recognizing the personal concept from a few examples while retaining the model's general knowledge, and does not require large scale pretraining." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 484, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 509 + ], + "type": "text", + "content": "a frisbee\" requires both personal knowledge (\"Fido\") and general knowledge (\"catching a frisbee\")." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 556, + 715 + ], + "type": "text", + "content": "Recent approaches for PerVL use textual inversion to learn a pseudo-text token (e.g., “" + }, + { + "bbox": [ + 312, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 312, + 510, + 556, + 715 + ], + "type": "text", + "content": "”) for each new personal concept by inverting from the target training images to the token [4, 7] (illustrated in Figure 1 (left)). These pseudo-text tokens require updating only a minimal set of parameters, and they can be used within a natural language query as input to the language encoder (e.g., “" + }, + { + "bbox": [ + 312, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 312, + 510, + 556, + 715 + ], + "type": "text", + "content": " catching a frisbee”). Typically, these approaches train a textual inversion network on large-scale data to predict a token representing a new concept, which may then be further optimized during personalization. While these approaches avoid updating the model's internal parameters, preventing the “overriding” of its general knowledge, their ability to represent the personal concept is limited to the single input text token. Furthermore, the token affects the entire text embedding process and can interfere with the language encoder's general knowledge. Consequently, textual inversion" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 261, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 261, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 261, + 713 + ], + "type": "text", + "content": "*Work partially done during internship at Adobe Research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19748" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "approaches often struggle to combine personalized and general knowledge across different datasets and can be time-consuming to optimize online due to the need for backpropagation through the full language encoder [7, 38], or require large-scale pretraining [4, 7, 38]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 133, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 133, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 133, + 294, + 251 + ], + "type": "text", + "content": "To address these shortcomings, we focus on updating the internal representation of CLIP's language encoder [27] using only a few training examples (illustrated in Figure 1 (right)). Updating the internal representation is challenging because the model must balance learning the personalized concept from limited examples while preventing catastrophic forgetting of its prior knowledge. Additionally, the training data does not inherently encourage the model to retain general knowledge. Our goal is to make this update efficiently, without relying on additional training data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 252, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 295, + 467 + ], + "type": "text", + "content": "Our contributions are fourfold. First, inspired by recent advances in text-to-image generation [14, 19, 29, 30], we show that CLIP's language encoder can be updated to learn a personalized concept from a few examples while retaining its general knowledge. We find that regularized low-rank adaptation (LoRA) finetuning [16] of the language encoder's last layer effectively balances the trade-off between learning the personalized concept and avoiding catastrophic forgetting. Our resulting method POLAR (PersOnalized Low-rank Adaptation for Retrieval) learns a low-rank parameter set that is minimal and comparable in size to the pseudo-text tokens used in textual inversion. Furthermore, we leverage LoRA's specific structure to introduce a regularization strategy that eliminates the need for additional training examples or regularization prompts. Unlike previous methods, our approach does not rely on any components pre-trained on large-scale data, allowing for seamless parameter updates using only a few training examples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 468, + 295, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 611 + ], + "type": "text", + "content": "Second, we introduce a new evaluation metric to assess how well general knowledge is preserved in our finetuned representation, using captions generated by a vision-language model (VLM). We find that our approach effectively maintains general knowledge. Third, we explore different strategies for combining learned representations for different personal concepts to support multi-concept queries. We find that adding LoRA representations is effective and outperforms orthogonal adaptation [26]. Finally, we demonstrate that POLAR achieves state-of-the-art accuracy on the DeepFashion2 [7] and ConCon-Chi [28] benchmarks, improving prior performance by " + }, + { + "bbox": [ + 55, + 468, + 295, + 611 + ], + "type": "inline_equation", + "content": "4\\% - 22\\%" + }, + { + "bbox": [ + 55, + 468, + 295, + 611 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Personalized Vision-Language Retrieval. Cohen et al. introduced the task of Personalized Vision-Language Retrieval [7] and proposed PALAVRA, a textual inversion approach for the task. PALAVRA first learns a textual inversion network on COCO, which takes a set of images of a concept and predicts an initial pseudo-word token to rep" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": "resent the concept in text queries. This token is then further optimized via backpropagation through CLIP's text encoder. Korbar et al. [18] propose a similar approach for the setting of retrieving specific people in videos. A few recent works explore learning embeddings to represent personal concepts for VLM captioning and QA tasks [1, 25]. Yeh et al. [38] build on the concept of textual inversion by meta-learning a basis for pseudo-tokens using large scale video data. However, they target a different setting where concepts are learned jointly, using the other concepts in the dataset as negative examples. Recently, Rosasco et al. introduced the Concept-Context Chimera dataset (ConConChi) for personalized retrieval, which provides a more rigorous benchmark for assessing retrieval of personal concepts across diverse contexts than prior datasets. In this work, we depart from prior work by representing new concepts as low rank parameter updates within the text encoder instead of pseudo-word tokens. We show that our approach more effectively composes personal and general knowledge, while requiring few parameters per-concept." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 315, + 553, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 553, + 639 + ], + "type": "text", + "content": "Personalized Generation. Personalized generation is a more studied related task that generates new images of a personal concept using text-to-image diffusion models. Some approaches use textual inversion to optimize pseudoword tokens to use within text prompts [9, 10], while others like Dreambooth [29] and Custom Diffusion [19] find that tuning the weights of the diffusion U-net generates personal concepts with better fidelity. Recent work has focused on selectively tuning certain parameters to promote parameter efficiency and speed [9, 15, 19, 40], with some leveraging low rank constraints [14, 26, 30, 33]. Most related to our work is Perfusion [33], which learns rank-one updates to the diffusion U-net with a key-locking mechanism to constrain updates to the concept's spatial location in the feature map. While personalization via parameter updates has become commonplace for personalized generation, it has not yet been explored for personalized retrieval, motivating our work. Due to the differences in the nature of the tasks (generative vs. discriminative) and models (text-to-image diffusion vs. dual encoder), we find that retrieval demands a different strategy for applying internal parameter updates; updating even sparse sets of parameters can result in catastrophic forgetting of the model's general knowledge. Instead of updating parameters throughout the full model, we apply a single rank-one parameter update to the final layer of CLIP's text encoder and directly regularize these parameters to avoid catastrophic forgetting of general knowledge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": "Composed Image Retrieval. Another related task is composed image retrieval [35], which takes an image and textual modification as inputs and performs image retrieval. Approaches have leveraged CLIP for composed image retrieval [2, 3], with Pic2Word [31] and SEARLE [4] learning textual inversion networks to predict a pseudo-word token" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19749" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 292, + 210 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 292, + 210 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 292, + 210 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 292, + 210 + ], + "type": "image", + "image_path": "af97af57bc75b134048dfd51e46fb3793fda95a4869c19f0fb143ae523521a6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 224, + 296, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 270 + ], + "type": "text", + "content": "Figure 2. Overview of POLAR. For a personal concept, POLAR learns a rank-1 update to the value transform in the final layer of the text encoder. To maintain general knowledge during personalization, we impose a regularization loss on the update." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 287, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 347 + ], + "type": "text", + "content": "for the input image. Importantly, composed image retrieval differs from personalized retrieval in that it does not require instance-level recognition of the same concept (e.g., retrieving the exact same person), and instead aims to retrieve images portraying a similar semantic class, layout, or style." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 504 + ], + "type": "text", + "content": "Few-Shot Adaptation of Vision-Language Models. Our task also relates to work on adapting models like CLIP for few-shot classification. Prior works primarily leverage prompt tuning [21], which learns input tokens to represent new classes [5, 6, 8, 23, 34, 36, 37, 42-44]. A few works consider tuning within the encoder by applying prompts to all layers [17, 20] or augmenting the encoder with adapter modules [12, 41]. An important distinction is that while CLIP's general knowledge may aid with learning classes from few examples, the general knowledge is not required for these downstream classification tasks. In contrast, personalized retrieval requires not just recognizing the personal concept, but composing it with retained general knowledge." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 295, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 544 + ], + "type": "text", + "content": "3. Personalized Low-Rank Adaptation for Retrieval (POLAR)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 180, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 180, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 180, + 563 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": "We follow the PerVL problem setup established by Cohen et al. [7]. Given a pretrained vision-language model " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": ", we aim to learn an adapted model " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\psi'" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " that is able to recognize a new personal concept " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " (e.g., \"my coffee mug\"). The pretrained model " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " consists of an image encoder " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\psi_I" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\psi_T" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": ". These encoders map image and textual inputs to a shared embedding space. At personalization time, we are given " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{I_i^c\\}_{i=1}^{N_c}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " depicting concept " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " as training data. Following Cohen et al. [7], we are also given the class name " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_c" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " of concept " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " (e.g., \"mug\"). At retrieval time, the input to the model consists of a textual query " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " and a set of " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "N_r" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{I_i^r\\}_{i=1}^{N_r}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " constituting the retrieval database. We" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 73, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 554, + 120 + ], + "type": "text", + "content": "compute the language embedding for an input query " + }, + { + "bbox": [ + 313, + 73, + 554, + 120 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 73, + 554, + 120 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 73, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\psi_T'(q)" + }, + { + "bbox": [ + 313, + 73, + 554, + 120 + ], + "type": "text", + "content": ", and perform retrieval by computing the cosine similarity between " + }, + { + "bbox": [ + 313, + 73, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\psi_T'(q)" + }, + { + "bbox": [ + 313, + 73, + 554, + 120 + ], + "type": "text", + "content": " and the image embedding for each retrieval image as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 341, + 125, + 555, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 125, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 341, + 125, + 555, + 152 + ], + "type": "interline_equation", + "content": "\\operatorname {s i m} \\left(\\psi_ {T} ^ {\\prime} (q), \\psi_ {I} ^ {\\prime} \\left(I _ {i} ^ {r}\\right)\\right) = \\frac {\\left\\langle \\psi_ {T} ^ {\\prime} (q) , \\psi_ {I} ^ {\\prime} \\left(I _ {i} ^ {r}\\right) \\right\\rangle}{\\left| \\left| \\psi_ {T} ^ {\\prime} (q) \\right| \\right| _ {2} \\left| \\left| \\psi_ {I} ^ {\\prime} \\left(I _ {i} ^ {r}\\right) \\right| \\right| _ {2}} \\tag {1}", + "image_path": "0502d91da87ab75b6e63016252ee4b646e3dce3424256243cb957a2ee14eb61a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 157, + 554, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 554, + 195 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 157, + 554, + 195 + ], + "type": "inline_equation", + "content": "\\langle \\cdot \\rangle" + }, + { + "bbox": [ + 313, + 157, + 554, + 195 + ], + "type": "text", + "content": " denotes the inner product. The top retrieval for the query is the image " + }, + { + "bbox": [ + 313, + 157, + 554, + 195 + ], + "type": "inline_equation", + "content": "I_{i}^{r}" + }, + { + "bbox": [ + 313, + 157, + 554, + 195 + ], + "type": "text", + "content": " corresponding to the embedding with the highest similarity to the query embedding " + }, + { + "bbox": [ + 313, + 157, + 554, + 195 + ], + "type": "inline_equation", + "content": "\\psi_T^\\prime (q)" + }, + { + "bbox": [ + 313, + 157, + 554, + 195 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 200, + 519, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 519, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 519, + 213 + ], + "type": "text", + "content": "3.2. Rank-One Personalized Value Updates" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 217, + 554, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 554, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 554, + 289 + ], + "type": "text", + "content": "In contrast to prior work that represents a concept as a learned input token for the text encoder [4, 7, 38], POLAR learns a low-rank parameter update to the text encoder for each concept (Fig. 2). We leverage LoRA [16]: for a weight " + }, + { + "bbox": [ + 313, + 217, + 554, + 289 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 313, + 217, + 554, + 289 + ], + "type": "text", + "content": " in a pretrained model, LoRA learns a low-rank update " + }, + { + "bbox": [ + 313, + 217, + 554, + 289 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 313, + 217, + 554, + 289 + ], + "type": "text", + "content": ", performing a modified forward pass as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 363, + 295, + 555, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 295, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 363, + 295, + 555, + 308 + ], + "type": "interline_equation", + "content": "y = (W + \\Delta W) x = W x + B A x \\tag {2}", + "image_path": "3a50675616197c34e4cfb0e8cc05bfc1db25fdebcad376e01a043b2d2fc6937e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{m \\times r}" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{r \\times n}" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "r < \\min(m, n)" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": " is the chosen rank of the weight update. By selecting " + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "r << \\min(m, n)" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": ", LoRA tunes minimal parameters in the original model " + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": ". For each concept " + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": ", we empirically choose to learn a rank-one (" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "r = 1" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": ") update to the value transform of the final attention layer " + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "\\psi_T" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": ". Our choice of " + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "r = 1" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": " reflects our goal to represent a single concept from very limited examples, while minimally interfering with the model's existing knowledge. Let " + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "Q_L, K_L, V_L \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": " be the pretrained text encoder's query, key, and value transforms for the attention mechanism in the final transformer layer " + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": " is the internal dimension of " + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "inline_equation", + "content": "\\psi_T" + }, + { + "bbox": [ + 313, + 314, + 554, + 471 + ], + "type": "text", + "content": ". For each attention head, the output of the multi-head attention layer is calculated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 485, + 554, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 485, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 331, + 485, + 554, + 513 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (x) = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {L} x \\left(K _ {L} x\\right) ^ {T}}{\\sqrt {d}}\\right) V _ {L, c} ^ {\\prime} x \\tag {3}", + "image_path": "d426167d4844898dcac8669d08011e9c4088731d80dd10c6f2d3c9ac5cc6554b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 517, + 546, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 546, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 546, + 531 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 517, + 546, + 531 + ], + "type": "inline_equation", + "content": "V_{L,c}^{\\prime}" + }, + { + "bbox": [ + 313, + 517, + 546, + 531 + ], + "type": "text", + "content": " is the value transform updated with LoRA as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 383, + 537, + 553, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 537, + 553, + 551 + ], + "spans": [ + { + "bbox": [ + 383, + 537, + 553, + 551 + ], + "type": "interline_equation", + "content": "V _ {L, c} ^ {\\prime} = V _ {L} + B _ {L, c} A _ {L, c}. \\tag {4}", + "image_path": "cae20207236b32d8051fec0d3cbb143fd90dbeedd482355cafbdb6f34a240f1d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 556, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 677 + ], + "type": "inline_equation", + "content": "V_{L,c}^{\\prime}" + }, + { + "bbox": [ + 313, + 556, + 554, + 677 + ], + "type": "text", + "content": " is learned separately for each concept " + }, + { + "bbox": [ + 313, + 556, + 554, + 677 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 556, + 554, + 677 + ], + "type": "text", + "content": ". Following prior work on personalized generation [14, 19], we choose a fixed token (e.g. \"sks\") in CLIP's vocabulary as a placeholder for the concept's place in the input queries. During training, we randomly select a template textual query (e.g. \"An image of sks\") to associate with each training image to form a set of text-image pairs " + }, + { + "bbox": [ + 313, + 556, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\{q_i^c, I_i^c\\}_{i=1}^{N_c}" + }, + { + "bbox": [ + 313, + 556, + 554, + 677 + ], + "type": "text", + "content": ". We supervise the learning of " + }, + { + "bbox": [ + 313, + 556, + 554, + 677 + ], + "type": "inline_equation", + "content": "V_{L,c}^{\\prime}" + }, + { + "bbox": [ + 313, + 556, + 554, + 677 + ], + "type": "text", + "content": " by using a mean-squared error (MSE) loss to push the normalized text and image embeddings for each training pair close together in the embedding space:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 328, + 681, + 555, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 681, + 555, + 716 + ], + "spans": [ + { + "bbox": [ + 328, + 681, + 555, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M S E}} = \\frac {1}{N _ {c}} \\sum_ {i = 1} ^ {N _ {c}} \\left(\\frac {\\psi_ {T , c} ^ {\\prime} (q _ {i})}{| | \\psi_ {T , c} ^ {\\prime} (q _ {i}) | | _ {2}} - \\frac {\\psi_ {I} \\left(I _ {i} ^ {c}\\right)}{| | \\psi_ {I} \\left(I _ {i} ^ {c}\\right) | | _ {2}}\\right) ^ {2} \\tag {5}", + "image_path": "3404e09a3ef40a3dc94a49cf499d7997f4ae8b909d74e32afd5ae3c4b3abfb68.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19750" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\psi_T^\\prime ,c" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " represents the text encoder with concept " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " 's LoRA update applied in the forward pass (Eq. (3))." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 149, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 149, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 149, + 115 + ], + "type": "text", + "content": "3.3. Regularization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "content": "A key challenge for personalized retrieval is to learn update weights that supply necessary personalized information without overriding the model's general knowledge, which is necessary for retrieving the personal concept in different contexts (e.g., \"my dog Fido catching a frisbee\" and \"my dog Fido sitting on the couch\"). We therefore propose a regularization scheme for POLAR. Differently than prior works that construct additional regularization examples to use during training [4, 7], we instead exploit the structure of our low-rank updates to directly minimize updates to the original representation. From Eq. (4), our parameter update alters the representation " + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "inline_equation", + "content": "V_{L}x" + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "content": " by adding the term " + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "inline_equation", + "content": "B_{L,c}A_{L,c}x" + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "content": ". With our use of rank " + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "inline_equation", + "content": "r = 1" + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "content": ", we can interpret " + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "inline_equation", + "content": "A_{L,c}x" + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "content": " as computing dot product similarity between the vectors " + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "inline_equation", + "content": "A_{L,c}^{T}" + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "content": ", which determines the scale of an added directional update " + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "inline_equation", + "content": "B_{L,c}" + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "content": ". Our regularization comprises two components: first, we add a penalty on the size of the weights in " + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "inline_equation", + "content": "B_{L,c}" + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "content": " to avoid unnecessary deviation from CLIP's existing representation, i.e., when the term " + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "inline_equation", + "content": "B_{L,c}A_{L,c}x = 0" + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "content": ", the text embedding will be the same as CLIP's original representation " + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "inline_equation", + "content": "(\\psi_T'(q) = \\psi_T(q))" + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "content": ". We modify our loss by adding a squared- " + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "content": " regularization over the weights " + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "inline_equation", + "content": "B_{L,c}" + }, + { + "bbox": [ + 55, + 120, + 296, + 384 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 389, + 295, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 389, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 99, + 389, + 295, + 404 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M S E}} + \\lambda \\mathcal {L} _ {\\mathrm {r e g}}, \\quad \\mathcal {L} _ {\\mathrm {r e g}} = | B _ {L, c} | ^ {2} \\tag {6}", + "image_path": "58f56fb6a4b087ad6c76ee4ff0f9e793f39b6a612d21401f576c596a8b792e5e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 409, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 409, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 409, + 296, + 495 + ], + "type": "text", + "content": " is a tunable hyperparameter that determines the relative weight of regularization in the loss. Second, we impose the constraint " + }, + { + "bbox": [ + 55, + 409, + 296, + 495 + ], + "type": "inline_equation", + "content": "||A_{L,c}||_2 = 1" + }, + { + "bbox": [ + 55, + 409, + 296, + 495 + ], + "type": "text", + "content": ", encouraging " + }, + { + "bbox": [ + 55, + 409, + 296, + 495 + ], + "type": "inline_equation", + "content": "A_{L,c}" + }, + { + "bbox": [ + 55, + 409, + 296, + 495 + ], + "type": "text", + "content": " to learn to selectively identify when to apply personal context based on directional similarity with the incoming representation " + }, + { + "bbox": [ + 55, + 409, + 296, + 495 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 409, + 296, + 495 + ], + "type": "text", + "content": ", while the magnitude of the personal update is controlled entirely by the regularized " + }, + { + "bbox": [ + 55, + 409, + 296, + 495 + ], + "type": "inline_equation", + "content": "B_{L,c}" + }, + { + "bbox": [ + 55, + 409, + 296, + 495 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 499, + 295, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 295, + 513 + ], + "type": "text", + "content": "3.4. Merging parameters for multi-concept queries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": "For queries that reference multiple personal concepts (e.g., \"my dog Fido is playing with Rex's favorite frisbee\"), we propose merging the parameter updates for the concepts into one weight update that is applied during encoding to provide personal context for both concepts. Let " + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "inline_equation", + "content": "V_{L,c_1}'" + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "inline_equation", + "content": "V_{L,c_2}'" + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": " be the individually learned weight updates for concepts " + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "inline_equation", + "content": "c_1" + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "inline_equation", + "content": "c_2" + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": ". We construct a combined weight update as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 120, + 606, + 295, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 606, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 120, + 606, + 295, + 622 + ], + "type": "interline_equation", + "content": "V _ {L, c _ {1} + c _ {2}} ^ {\\prime} = V _ {L, c _ {1}} ^ {\\prime} + V _ {L, c _ {2}} ^ {\\prime}. \\tag {7}", + "image_path": "898706a2441fd21011c637f4c1986d1463fc8a6f4c28476d105924eea0a834dc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 626, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 296, + 686 + ], + "type": "text", + "content": "During the forward pass, this setup equates to adding the parameter updates for both personal concepts to the representation. This update is also equivalent to constructing a rank " + }, + { + "bbox": [ + 55, + 626, + 296, + 686 + ], + "type": "inline_equation", + "content": "r = 2" + }, + { + "bbox": [ + 55, + 626, + 296, + 686 + ], + "type": "text", + "content": " update via matrix concatenation along the rank dimension as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 97, + 689, + 295, + 718 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 689, + 295, + 718 + ], + "spans": [ + { + "bbox": [ + 97, + 689, + 295, + 718 + ], + "type": "interline_equation", + "content": "V _ {L, c _ {1} + c _ {2}} ^ {\\prime} = \\left[ \\begin{array}{l l} B _ {L, c _ {1}} & B _ {L, c _ {2}} \\end{array} \\right] \\left[ \\begin{array}{l} A _ {L, c _ {1}} \\\\ A _ {L, c _ {2}} \\end{array} \\right]. \\tag {8}", + "image_path": "8276c1a95f5c0860ebacca69e4e1524f6f871832ac24b8a56237128b90c41f59.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "This approach also generalizes to merging greater than two concepts. We explore other merging strategies in Tab. 8." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 109, + 395, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 395, + 123 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 395, + 123 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 130, + 554, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 554, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 554, + 190 + ], + "type": "text", + "content": "In this section, we evaluate POLAR and compare it to existing works. We also provide ablations to give insight into the design choices within our method. We include further experiment details, analysis of personalization time, and discussion of limitations in the supplemental material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 199, + 377, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 377, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 377, + 211 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 217, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 397 + ], + "type": "text", + "content": "DeepFashion2. Cohen et al. [7] define a personalized retrieval benchmark on the DeepFashion2 dataset [13], where the 50 personal concepts are different clothing items. The test set includes 221 captions and images for retrieval. ConCon-Chi. The recent ConCon-Chi dataset aims to more comprehensively evaluate unique personal concepts in a variety of contexts; we thus use it primarily for our analysis. It consists of 20 concepts including household objects and chimeric concepts (combinations of multiple objects). There are 1084 context queries (735 single-concept, 349 multi-concept), and 4008 retrieval images. For direct comparisons to the original baselines [28], we report on the full TEST set, which includes 3 validation concepts. We verify our gains hold on the TEST-UNSEEN split, which excludes these concepts, in the supplemental." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 407, + 430, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 430, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 430, + 418 + ], + "type": "text", + "content": "4.2. Evaluation Protocol" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "text", + "content": "Context Queries. Context queries perform retrieval on a caption referencing the personal concept in a particular context (e.g., \"my dog Fido catching a frisbee in the backyard\"). The ground truth consists of the images labeled as matching this prompt. For DeepFashion2, there is one ground truth image per context query, while for ConCon-Chi there are 1-130 ground truth images (average " + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\approx 6" + }, + { + "bbox": [ + 313, + 425, + 554, + 509 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "type": "text", + "content": "Concept-only Queries. Following Yeh et al. [38], we also report retrieval accuracy on \"concept-only\" queries to evaluate the model's ability to recognize the personal concept independent of context. For each concept, we use the input query \"An image of " + }, + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{V}^{*}" + }, + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "type": "text", + "content": "\" and compute retrieval metrics where the ground truth is all retrieval images that contain the concept. For ConCon-Chi, we include only single-concept images (2430 images) in the retrieval database." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Metrics. We report retrieval accuracy using standard benchmark metrics for the datasets: mean reciprocal rank (mRR) – the average inverse rank of the first retrieved ground truth image; recall-at- " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{r}@\\mathbf{k}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ") – the average success rate within the top " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " retrievals; and mean average precision (mAP) – the area under the precision-recall curve, averaged over all queries. We report mAP for settings with multiple ground truth images per query (concept-only queries on both datasets, and context queries on ConCon-Chi)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19751" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 187 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 187 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 187 + ], + "type": "table", + "html": "
MethodArch.ContextConcept-only
mRRrecall@5mRRmAP
AdapterViT-B/325.9---
COLLIE [32]ViT-B/327.9---
Text OnlyViT-B/3217.6---
AvgIm + TextViT-B/3218.8---
PALAVRA [7]ViT-B/3228.439.2--
SEARLE [4]ViT-B/3221.9027.1512.7425.97
OursViT-B/3234.8244.8859.2628.75
SEARLE [4]ViT-L/1427.6234.1232.0716.17
OursViT-L/1440.7251.3165.9635.07
", + "image_path": "b6439906d51eec746452ef26710b441f2021b6c5a8bc20f997969ea0eece8fa0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 240, + 301, + 344 + ], + "blocks": [ + { + "bbox": [ + 55, + 191, + 295, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 295, + 236 + ], + "type": "text", + "content": "Table 1. Comparison to prior work on the DeepFashion2 retrieval benchmark with 5 training images per concept. We report the mean over over 5 runs with 5 randomly chosen training images of the concept per run (see supplemental for standard error)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 240, + 301, + 344 + ], + "lines": [ + { + "bbox": [ + 56, + 240, + 301, + 344 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 301, + 344 + ], + "type": "table", + "html": "
MethodmRRContextConcept-only
mAPrecall@1mRRmAP
Coarse (class name)24.2116.8314.48--
\\(Discriminative^†\\)43.1630.1631.92--
\\(Rich^†\\)40.5827.6529.98--
PALAVRA [7]35.9923.5926.75--
Pic2Word [31]38.6226.3927.68--
SEARLE [4]43.9330.7433.4996.6761.94
Ours46.3332.3336.16100.0068.71
", + "image_path": "1766cb38f9f5dfc5d67b1ac6f21c6ed562868a5a1ddc7e8d331307bd1c89cbb8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 62, + 356, + 288, + 407 + ], + "blocks": [ + { + "bbox": [ + 76, + 346, + 275, + 355 + ], + "lines": [ + { + "bbox": [ + 76, + 346, + 275, + 355 + ], + "spans": [ + { + "bbox": [ + 76, + 346, + 275, + 355 + ], + "type": "text", + "content": "(a) Comparison to prior work on the ConCon-Chi benchmark." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 356, + 288, + 407 + ], + "lines": [ + { + "bbox": [ + 62, + 356, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 356, + 288, + 407 + ], + "type": "table", + "html": "
MethodContext (Single-concept)Context (Multi-concept)
mRRmAPr@1mRRmAPr@1
SEARLE49.5035.2539.0532.0621.2221.78
Ours51.6436.7341.7735.1323.0524.36
", + "image_path": "6d6a00421002edd0195108ef38bdee46585765f9fcffae49feae279a874b0a37.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 408, + 288, + 418 + ], + "lines": [ + { + "bbox": [ + 63, + 408, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 63, + 408, + 288, + 418 + ], + "type": "text", + "content": "(b) Results for ConCon-Chi single-concept vs. multi-concept queries." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 423, + 295, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 501 + ], + "type": "text", + "content": "Table 2. Our approach achieves state of the art results on the challenging ConCon-Chi benchmark on all metrics. We also break down the results of our method and SEARLE[4] by single-concept and multi-concept queries, demonstrating best results on both. † refer to ConCon-Chi's provided text descriptors for each concept, which serve as oracles since they use knowledge of all concepts to manually determine a differentiating description." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 638 + ], + "type": "text", + "content": "Implementation Details. For our main method on ConCon-Chi, we use " + }, + { + "bbox": [ + 55, + 517, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\lambda = 0.35" + }, + { + "bbox": [ + 55, + 517, + 295, + 638 + ], + "type": "text", + "content": ". We train for 500 iterations with learning rate 0.001 and the Adam optimizer. Our model converges within 50 epochs. Because we optimize minimal parameters and backpropagate through only the final layer, personalization is fast, taking under 1 second on a V100 GPU. On DeepFashion2 we append the classname to " + }, + { + "bbox": [ + 55, + 517, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{V}^*" + }, + { + "bbox": [ + 55, + 517, + 295, + 638 + ], + "type": "text", + "content": " (e.g., \"sks dress\") like in PALAVRA [7] and use use " + }, + { + "bbox": [ + 55, + 517, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 55, + 517, + 295, + 638 + ], + "type": "text", + "content": ". We use the same template prompts as PALAVRA for training. We provide further details in the supplemental." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 647, + 205, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 205, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 205, + 659 + ], + "type": "text", + "content": "4.3. Comparison to Prior Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "We compare our method to prior work on DeepFashion2 in Tab. 1 and ConCon-Chi in Tab. 2. For DeepFashion2 we compare against PALAVRA and the baselines reported in its paper [7]. We also run SEARLE [4], a zero-shot com" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "posed image retrieval network, using the publicly available checkpoints. Our method achieves state-of-the-art results in the standard setting using the CLIP ViT-B/32 architecture. We also see improvement over SEARLE when using the larger CLIP ViT-L/14 architecture." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 134, + 555, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 134, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 555, + 302 + ], + "type": "text", + "content": "On ConCon-Chi, we compare against the baselines reported in the benchmark's paper [28]. All methods use the CLIP ViT-L/14 architecture. Differently from DeepFashion2, the zero-shot composed image retrieval methods (SEARLE and Pic2Word) outperform PALAVRA, which suggests that while they are effective in some settings, they struggle with differentiating between several similar concepts like the clothing items in DeepFashion2. This is qualitatively demonstrated in Fig. 3, which compares our method's retrievals with SEARLE. Our method, however, also achieves state-of-the-art results on ConCon-Chi, demonstrating flexibility across different benchmarks. Additionally, we achieve stronger mAP for the concept-only queries task than SEARLE." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 304, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 544 + ], + "type": "text", + "content": "Evaluating General Knowledge in LoRA vs. Token Learning. Our primary hypothesis is that learning a small, regularized parameter update within the encoding process can more effectively allow CLIP to reason about personal and general information together. While concept-only queries assess personal knowledge and context queries assess the combination of personal and general knowledge, our setting of tuning the text encoder's parameters also allows us to measure the retention of general knowledge. We do so by inputting general queries, which do not reference the personal concept, to the text encoder with our parameter update for the concept still applied. To source non-personal queries, we use LLaVA [22] to caption each image in the retrieval set. We define a new metric VLM caption recall@10, which measures the success rate of retrieving the image for which the caption was generated in the top 10 retrievals. We choose 10 retrievals because ConCon-Chi has many similar images for which the same caption is valid. A drop from the original CLIP's performance on this metric (52.69) indicates forgetting of general knowledge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 714 + ], + "type": "text", + "content": "In Tab. 3, we compare our approach of learning internal parameter updates with learning input tokens in the same training setting. We consider Textual Inversion (TI), which learns a token that is integrated into input queries via a pseudo-word for the concept (e.g. \"A photo of V* jumping), and Prompt Tuning [43], which bridges the gap between TI and parameter updates by prepending learned prompt tokens to all queries (both personal and general). In contrast to TI, this allows us to use our VLM caption metric to measure the interference of the prompt tokens on general queries. We include results with 1 learned token as well as 2 learned tokens, which is equivalent in size to our rank-1 parameter update. With prompt tuning, we can achieve strong results on concept-only queries, but our VLM cap" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19752" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 556, + 355 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 556, + 355 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 556, + 355 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 556, + 355 + ], + "type": "image", + "image_path": "ea0993fae84a031f1cd3caf3267b5a78d909615879868824a6f00abca322eda7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 361, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 555, + 407 + ], + "type": "text", + "content": "Figure 3. We compare the top 3 retrievals of our method vs. SEARLE for personal queries in the ConCon-Chi and DeepFashion2 datasets, with green borders indicating correct retrievals. We observe that SEARLE struggles to differentiate between concepts of similar classes, such as different clothing items. Our method more consistently retrieves the correct concept in the correct context, demonstrating effective composition of personal and general knowledge." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 415, + 298, + 518 + ], + "blocks": [ + { + "bbox": [ + 56, + 415, + 298, + 518 + ], + "lines": [ + { + "bbox": [ + 56, + 415, + 298, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 298, + 518 + ], + "type": "table", + "html": "
MethodContext (Single)Concept-onlyVLM cap
mRRmAPmRRmAPr@10
Original CLIP29.3920.7610.756.4652.69
Text. Inv. (1 tok)42.4532.9397.5064.71N/A
Text. Inv. (2 tok)41.7327.94100.0063.88N/A
Prompt (1 tok)31.7720.7096.2558.9530.84
Prompt (2 tok)33.1420.93100.0064.4915.35
Text. Inv. + Ours39.2926.55100.0064.7252.57
Ours51.6436.73100.0068.7152.62
", + "image_path": "d33c23654a81da96044055894cce9cce5804f1789dc5619d6feba79ebd5f5108.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 521, + 297, + 656 + ], + "lines": [ + { + "bbox": [ + 55, + 521, + 297, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 297, + 656 + ], + "type": "text", + "content": "Table 3. We compare our approach, which updates the weights of the text encoder, to tuning input tokens on ConCon-Chi. Rows 2-3 represent Textual Inversion, where a learned token is applied in place of the personal concept in queries that reference the concept. Rows 4-5 tune prompt tokens that are prepended to all text queries. Row 6 learns both our parameter update and the token. Tuning tokens can achieve competitive results on concept-only queries but struggles on contextual queries that require composing personal and general knowledge. This catastrophic forgetting of general knowledge is reflected by our VLM caption matching metric. We propose a new approach that achieves strong performance on contextual, concept-only, and general queries." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": "tion metric shows that this strength comes at the cost of catastrophic forgetting of general knowledge. We also combine TI with our method by learning the input token in addition to our parameter update. We observe that learning the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 417, + 556, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 556, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 556, + 562 + ], + "type": "text", + "content": "token reduces context performance, suggesting that the inserted learned token interferes with general knowledge. In contrast, our parameter updates achieve strong performance on personal queries while not interfering with CLIP's general knowledge, even when applied to non-personal queries. We hypothesize this is due to the minimally invasive structure of our updates, and that they are applied late in the encoding process in contrast to learned input tokens which influence the entire encoding process. We qualitatively illustrate this finding in Fig. 4; with the parameter update for a personal concept applied, the model successfully performs retrieval for both personalized queries and general queries." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 574, + 409, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 409, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 409, + 588 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "A key choice in developing our method is determining an effective, minimal set of parameters within the model for which to apply personalized updates. In this section, we empirically investigate rank size of the LoRA updates, which layers in the encoding process to apply LoRA updates, and on which parameters to learn LoRA updates. We additionally ablate our regularization and multi-concept merging strategies. All experiments use the CLIP ViT-L/14 architecture on the ConCon-Chi dataset with 5 training images, and we report our results on single-concept context queries" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19753" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 77, + 555, + 327 + ], + "blocks": [ + { + "bbox": [ + 63, + 77, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 63, + 77, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 63, + 77, + 555, + 327 + ], + "type": "image", + "image_path": "f8975b6b5a8baeb0bd39df58f0a8361a4085aa6cb77cdb475075ab28549c7b2e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 332, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 55, + 332, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 555, + 377 + ], + "type": "text", + "content": "Figure 4. Our parameter updates enable personalized retrieval without overriding the model's general knowledge. On the left we show the top 3 retrievals of our personalized model on context queries referencing the personal concept, with green indicating ground truth correct retrievals. On the right, we show the results of querying the same personalized model for general VLM-generated captions. There is not exhaustive ground truth for which images match each caption; however, qualitatively our model retrieves appropriate images in all cases." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 385, + 293, + 471 + ], + "blocks": [ + { + "bbox": [ + 57, + 385, + 293, + 471 + ], + "lines": [ + { + "bbox": [ + 57, + 385, + 293, + 471 + ], + "spans": [ + { + "bbox": [ + 57, + 385, + 293, + 471 + ], + "type": "table", + "html": "
LoRA rankContext (Single-Concept)Concept-onlyVLM cap
mRRmAPr@1mRRmAPr@10
r=252.3136.5942.04100.0066.0752.78
r=451.5236.6041.36100.0068.1352.61
r=851.4936.5841.36100.0068.1552.62
r=1651.6736.6641.50100.0067.9352.62
r=151.6436.7341.77100.0068.7152.62
", + "image_path": "8832cfdb9479672567f35e8a970872f0450ac3e9df2fc74e518d5b81d6ddb3b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 508, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 567 + ], + "type": "text", + "content": "(we ablate the merging strategy for multi-concept queries in Sec. 4.4). We performed our model selection (design choices for our main method and regularization weight " + }, + { + "bbox": [ + 55, + 508, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 508, + 295, + 567 + ], + "type": "text", + "content": ") on the validation split (3 concepts) and report our main results on the test split." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 568, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 663 + ], + "type": "text", + "content": "LoRA Rank. We ablate the rank of the LoRA updates in Tab. 4, which controls the number of learned parameters for each update. We observe only small retrieval accuracy gains on single-concept queries as the rank increases, and achieve better concept-only and competitive multi-concept results with rank=1. The rank-1 concept update is also the most parameter efficient, storing only " + }, + { + "bbox": [ + 55, + 568, + 295, + 663 + ], + "type": "inline_equation", + "content": "2d" + }, + { + "bbox": [ + 55, + 568, + 295, + 663 + ], + "type": "text", + "content": " parameters per concept, where " + }, + { + "bbox": [ + 55, + 568, + 295, + 663 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 568, + 295, + 663 + ], + "type": "text", + "content": " is the encoder's internal dimension." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Architecture Layers. Tab. 5 investigates which transformer layers to apply LoRA updates to, dictating how early or late into the encoder process personalized information is injected. We achieve the strongest results by applying" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 386, + 559, + 471 + ], + "blocks": [ + { + "bbox": [ + 85, + 472, + 265, + 483 + ], + "lines": [ + { + "bbox": [ + 85, + 472, + 265, + 483 + ], + "spans": [ + { + "bbox": [ + 85, + 472, + 265, + 483 + ], + "type": "text", + "content": "Table 4. Ablation of LoRA rank on ConCon-Chi." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 386, + 559, + 471 + ], + "lines": [ + { + "bbox": [ + 315, + 386, + 559, + 471 + ], + "spans": [ + { + "bbox": [ + 315, + 386, + 559, + 471 + ], + "type": "table", + "html": "
Layer(s)Context (Single-Concept)Concept-onlyVLM cap
mRRmAPr@1mRRmAPr@10
11,1250.1835.3439.73100.0064.0952.64
10,11,1251.5135.8141.63100.0065.8752.62
all layers43.2329.9332.5297.5063.7752.45
144.6931.1234.1597.5064.6652.18
1251.6436.7341.77100.0068.7152.62
", + "image_path": "7b89db7fde654055a5dc9aa399563e6ec24ce3d5a9a83b2d0bfc520305df6fba.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 340, + 474, + 526, + 485 + ], + "lines": [ + { + "bbox": [ + 340, + 474, + 526, + 485 + ], + "spans": [ + { + "bbox": [ + 340, + 474, + 526, + 485 + ], + "type": "text", + "content": "Table 5. Ablation of LoRA layers on ConCon-Chi." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 497, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 664 + ], + "type": "text", + "content": "personalization in the final layer (layer 12). We hypothesize that updating later layers is better than earlier layers because it allows our approach to apply a small, targeted update to the developed text query representation to inject personal information. In contrast, earlier layer updates are more likely to alter the full representation - not just the parts semantically belonging to the personal concept. Additionally, our findings align with works that suggest that the later layers in transformer encoders are the most important in constructing the final representation [11]. We also do not observe an overall benefit by learning LoRA updates for earlier layers in addition to the final layer; the final layer alone achieves best performance while also requiring the fewest learned parameters per concept." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Parameters. Tab. 6 ablates the component of the transformer layer on which the LoRA is learned. We consider the linear transforms within the attention mechanism (query, key, value, and output), the 2 MLP layers, as well as the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19754" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 300, + 193 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 300, + 193 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 300, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 300, + 193 + ], + "type": "table", + "html": "
Param(s)Context (Single-Concept)Concept-onlyVLM cap
mRRmAPr@1mRRmAPr@10
Q16.6511.497.6232.6610.9151.84
K15.5511.456.5328.009.2852.12
O46.9931.0438.5097.5060.9852.52
Q,K,V,O47.5231.6938.7897.5060.9052.54
Q,V51.5035.6041.50100.0065.6052.65
MLP143.0427.9133.88100.0055.8752.05
MLP249.4032.8738.91100.0055.7051.65
final proj51.1534.9840.82100.0060.5851.58
V51.6436.7341.77100.0068.7152.62
", + "image_path": "af2560d4465b3054ebedd08833a75e4fcbf018ed70cc34801411aa6181b06dde.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 222, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 295, + 582 + ], + "type": "text", + "content": "model's final projection layer. We see that the query and key transforms alone are not effective for personalization, indicating that directly transforming the output representation is crucial for personalization; re-weighting the existing tokens via altering the attention weight computation is not enough. The strongest results are achieved by updating the value transform. Interestingly, this setting outperforms updating the output transform and the following MLP layers. This result suggests that the value transform's placement in transforming the output of each attention head is optimal for personalization as opposed to later linear transforms that operate after the output of the attention heads is aggregated. We see negligible gains by pairing updates on the value transform with other parameters; in fact, this setting decreases the results on some metrics. Updating the final projection alone also produces competitive results, but lags on the concept-only metrics. Our analysis demonstrates that the value transform within the final layer is optimally positioned to learn a small, targeted update for personalization. Regularization. We ablate our regularization strategy in Tab. 7. Without regularization, we see drops on the contextual query performance and VLM caption metric, indicating forgetting of general knowledge. The concept-only mAP slightly increases, showing prioritization of personalization over retaining general knowledge. Our regularization strategies are complementary: using both together produces the best results. With our regularization scheme, we are able to produce strong results on contextual queries, avoid any degradation on general caption-matching performance, and still produce good concept-only performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Multi-Concept Merging. We consider alternative strategies for merging the low-rank parameter updates of different concepts in Tab. 8. Our hypothesis is that because we learn a single, constrained update for each concept, the updates for different concepts will be sufficiently different to not interfere with each other. Our results validate this; we see applying each update is better than altering them via averaging or pooling them together. We also consider Orthogonal Adapation [26], which constrains updates for different concepts to be orthogonal to eachother, thus avoiding interference. In this method, the " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " matrix in the low-rank update" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 566, + 140 + ], + "blocks": [ + { + "bbox": [ + 73, + 196, + 277, + 207 + ], + "lines": [ + { + "bbox": [ + 73, + 196, + 277, + 207 + ], + "spans": [ + { + "bbox": [ + 73, + 196, + 277, + 207 + ], + "type": "text", + "content": "Table 6. Ablation of LoRA parameters on ConCon-Chi." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 566, + 140 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 566, + 140 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 566, + 140 + ], + "type": "table", + "html": "
Reg.Context (Single-Concept)Concept-onlyVLM cap
ABmRRmAPr@1mRRmAPr@10
22.5114.3514.83100.0069.8952.52
✓33.7722.2925.44100.0069.6352.58
✓39.8426.4531.56100.0069.0152.57
✓✓51.6436.7341.77100.0068.7152.62
", + "image_path": "6b63374a7e447d2a4eab4f989134b2ba4e99f6fe90ec005f3747bb8d9311bd08.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 337, + 234, + 531, + 299 + ], + "blocks": [ + { + "bbox": [ + 313, + 144, + 555, + 233 + ], + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 233 + ], + "type": "text", + "content": "Table 7. Ablation of our regularization scheme: A denotes imposing the constraint " + }, + { + "bbox": [ + 313, + 144, + 555, + 233 + ], + "type": "inline_equation", + "content": "||A_{L,c}||_2 = 1" + }, + { + "bbox": [ + 313, + 144, + 555, + 233 + ], + "type": "text", + "content": ", and B denotes applying the squared L2 penalty to " + }, + { + "bbox": [ + 313, + 144, + 555, + 233 + ], + "type": "inline_equation", + "content": "B_{L,c}" + }, + { + "bbox": [ + 313, + 144, + 555, + 233 + ], + "type": "text", + "content": " (Eq. 6). Without our regularization, the personalized parameter updates cause the model overfit to the concept, producing high concept-only metrics but catastrophically forgetting general knowledge, as reflected in the context and VLM caption metrics. Our regularization prevents this forgetting while still achieving high concept-only performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 234, + 531, + 299 + ], + "lines": [ + { + "bbox": [ + 337, + 234, + 531, + 299 + ], + "spans": [ + { + "bbox": [ + 337, + 234, + 531, + 299 + ], + "type": "table", + "html": "
Merge strategymRRmAPr@1
Avg LoRAs25.0315.5115.76
Max LoRAs25.2715.5616.05
Orthogonal Adaptation [26]28.3819.1418.62
Add LoRAs35.1323.0524.36
", + "image_path": "407af02467bf46e27a12037a0ef040cf1193843859355a0eab6c71c2ff3cc6a7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 302, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 324 + ], + "type": "text", + "content": "Table 8. Performance of different merging strategies for multi-concept queries in ConCon-Chi." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "text", + "content": "is frozen and drawn from a shared orthogonal subspace, while the " + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "text", + "content": " matrix is learned. The updates for different concepts are merged by adding them together. Interestingly, we find Orthogonal Adaptation to be less effective than our approach; we hypothesize this is due to the differences between where low rank updates are applied in our method vs. the original Orthogonal Adaptation method, which operates on text to image diffusion models. Whereas in personalized generation, parameter updates are applied throughout the full model, we find learning only a single parameter update is better suited for retrieval. Because we only learn this one parameter update late in the model, certain randomly selected " + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "text", + "content": " matrices are not effective for personalization depending on how they interact with the incoming feature representation at that point in the model. Specifically, in computing " + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "inline_equation", + "content": "B_{L,c}A_{L,c}x" + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "text", + "content": " if the randomly chosen " + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "inline_equation", + "content": "A_{L,c}" + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "text", + "content": " is orthogonal to " + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "text", + "content": ", this will eliminate the parameter update altogether. While personalized generation approaches avoid this by adapting the representation throughout all layers of the model, in our setting we find it is better to learn both " + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 336, + 556, + 588 + ], + "type": "text", + "content": " instead of imposing orthogonality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "In this work, we show that updating the internal representation of the CLIP text encoder serves as a better alternative to textual inversion for personalized search. By constraining the parameter update for each concept to a single rank-one update in the value transform of the final layer and strategically regularizing the parameters, we demonstrate that our approach effectively personalizes from a few image examples while maintaining the model's general knowledge." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19755" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "type": "text", + "content": "Acknowledgments The authors thank the members of the Hoffman Lab at Georgia Tech and Jitesh Jain for their feedback on this work. Fiona Ryan is supported by the NSF Graduate Research Fellowship under Grant No. DGE-2039655. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors(s) and do not necessarily reflect the views of the National Science Foundation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 177, + 115, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 115, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 115, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 196, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 196, + 297, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 196, + 297, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 297, + 240 + ], + "type": "text", + "content": "[1] Yuval Alaluf, Elad Richardson, Sergey Tulyakov, Kfir Aberman, and Daniel Cohen-Or. Myvlm: Personalizing vlms for user-specific queries. In European Conference on Computer Vision, pages 73–91. Springer, 2025. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 241, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 241, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 61, + 241, + 296, + 296 + ], + "type": "text", + "content": "[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4959-4968, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 296, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 296, + 350 + ], + "type": "text", + "content": "[3] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21466-21474, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 351, + 296, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 296, + 405 + ], + "type": "text", + "content": "[4] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15338-15347, 2023. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 406, + 296, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 296, + 460 + ], + "type": "text", + "content": "[5] Adrian Bulat and Georgios Tzimiropoulos. Lasp: Text-to-text optimization for language-aware soft prompting of vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23232-23241, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 460, + 296, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 296, + 504 + ], + "type": "text", + "content": "[6] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. arXiv preprint arXiv:2210.01253, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "type": "text", + "content": "[7] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In European conference on computer vision, pages 558-577. Springer, 2022. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 560, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 560, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 560, + 296, + 613 + ], + "type": "text", + "content": "[8] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor Guilherme Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Variational prompt tuning improves generalization of vision-language models. 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 614, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 658 + ], + "type": "text", + "content": "[9] Ziyi Dong, Pengxu Wei, and Liang Lin. Dreamartist: Towards controllable one-shot text-to-image generation via positive-negative prompt-tuning. arXiv preprint arXiv:2211.11337, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[11] Yossi Gandelsman, Alexei A. Efros, and Jacob Steinhardt. Interpreting clip's image representation via text-based decomposition. In ICLR, 2023. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 106, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 553, + 160 + ], + "type": "text", + "content": "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 161, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 161, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 161, + 553, + 227 + ], + "type": "text", + "content": "[13] Yuying Ge, Ruimao Zhang, Xiaogang Wang, Xiaou Tang, and Ping Luo. Deepfashion2: A versatile benchmark for detection, pose estimation, segmentation and re-identification of clothing images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5337–5345, 2019. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "type": "text", + "content": "[14] Cusuh Ham, Matthew Fisher, James Hays, Nicholas Kolkin, Yuchen Liu, Richard Zhang, and Tobias Hinz. Personalized residuals for concept-driven text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8186-8195, 2024. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 293, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 348 + ], + "type": "text", + "content": "[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 349, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 392 + ], + "type": "text", + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "text", + "content": "[17] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 449, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 470 + ], + "type": "text", + "content": "[18] Bruno Korbar and Andrew Zisserman. Personalised clip or: how to find your vacation videos. 2022. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 471, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 525 + ], + "type": "text", + "content": "[19] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 525, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 525, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 525, + 553, + 581 + ], + "type": "text", + "content": "[20] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1401-1411, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 582, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 612 + ], + "type": "text", + "content": "[21] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 647 + ], + "type": "text", + "content": "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "type": "text", + "content": "[23] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[24] Joanna Materzynska, Josef Sivic, Eli Shechtman, Antonio Torralba, Richard Zhang, and Bryan Russell. Customizing" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19756" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "text", + "content": "motion in text-to-video diffusion models. arXiv preprint arXiv:2312.04966, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[25] Thao Nguyen, Haotian Liu, Yuheng Li, Mu Cai, Utkarsh Ojha, and Yong Jae Lee. Yo'llava: Your personalized language and vision assistant. arXiv preprint arXiv:2406.09400, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 294, + 196 + ], + "type": "text", + "content": "[26] Ryan Po, Guandao Yang, Kfir Aberman, and Gordon Wetzstein. Orthogonal adaptation for modular customization of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7964-7973, 2024. 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 275 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 352 + ], + "type": "text", + "content": "[28] Andrea Rosasco, Stefano Berti, Giulia Pasquale, Damiano Malafronte, Shogo Sato, Hiroyuki Segawa, Tetsugo Inada, and Lorenzo Natale. Concon-chi: Concept-context chimera benchmark for personalized vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22239-22248, 2024. 2, 4, 5, 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 419 + ], + "type": "text", + "content": "[29] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 421, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 487 + ], + "type": "text", + "content": "[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6527-6536, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 488, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 553 + ], + "type": "text", + "content": "[31] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19305-19314, 2023. 2, 5, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 555, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 599 + ], + "type": "text", + "content": "[32] Gabriel Skantze and Bram Willemsen. Collie: Continual learning of language grounding from language-image embeddings. Journal of Artificial Intelligence Research, 74: 1201-1223, 2022. 5, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 600, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 645 + ], + "type": "text", + "content": "[33] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 conference proceedings, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[34] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[35] Nam Vo, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 575 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "for image retrieval—an empirical odyssey. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6439-6448, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 150 + ], + "type": "text", + "content": "[36] Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guoqiang Liang, Peng Wang, and Yanning Zhang. Dual modality prompt tuning for vision-language pre-trained model. IEEE Transactions on Multimedia, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 205 + ], + "type": "text", + "content": "[37] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6757-6767, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 207, + 554, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 554, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 554, + 272 + ], + "type": "text", + "content": "[38] Chun-Hsiao Yeh, Bryan Russell, Josef Sivic, Fabian Caba Heilbron, and Simon Jenni. Meta-personalizing vision-language models to find named instances in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19123-19132, 2023. 2, 3, 4, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 274, + 554, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 554, + 329 + ], + "type": "text", + "content": "[39] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the association for computational linguistics, 2:67-78, 2014. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "type": "text", + "content": "[40] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 376, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 430 + ], + "type": "text", + "content": "[41] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaptation of clip for few-shot classification. In European conference on computer vision, pages 493-510. Springer, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 431, + 554, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 554, + 485 + ], + "type": "text", + "content": "[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 487, + 554, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 554, + 529 + ], + "type": "text", + "content": "[43] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 531, + 554, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 554, + 575 + ], + "type": "text", + "content": "[44] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19757" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/ca5cab4a-bdaf-48a0-9763-5efeeaa13237_content_list.json b/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/ca5cab4a-bdaf-48a0-9763-5efeeaa13237_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dbbfa8fe447d670c7eac8347fbb4e75c753b23b7 --- /dev/null +++ b/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/ca5cab4a-bdaf-48a0-9763-5efeeaa13237_content_list.json @@ -0,0 +1,1715 @@ +[ + { + "type": "text", + "text": "Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity", + "text_level": 1, + "bbox": [ + 199, + 130, + 799, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen-Yi Lu Kasra Derakhshandeh Somali Chaterji Purdue University", + "bbox": [ + 276, + 203, + 718, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lu842@purdue.edu, kderakhs@purdue.edu, schaterji@purdue.edu", + "bbox": [ + 233, + 241, + 756, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-supervised semantic segmentation with consistency regularization capitalizes on unlabeled images to enhance the accuracy of pixel-level segmentation. Current consistency learning methods primarily rely on the consistency loss between pseudo-labels and unlabeled images, neglecting the information within the feature representations of the backbone encoder. Preserving maximum information in feature embeddings requires achieving the alignment and uniformity objectives, as widely studied. To address this, we present SWSEG, a semi-supervised semantic segmentation algorithm that optimizes alignment and uniformity using the Sliced-Wasserstein Distance (SWD), and rigorously and empirically proves this connection. We further resolve the computational issues associated with conventional Monte Carlo-based SWD by implementing a Gaussian-approximated variant, which not only maintains the alignment and uniformity objectives but also improves training efficiency. We evaluate SWSEG on the PASCAL VOC 2012, Cityscapes, and ADE20K datasets, outshining supervised baselines in mIoU by up to $11.8\\%$ , $8.9\\%$ , and $8.2\\%$ , respectively, given an equivalent number of labeled samples. Further, SWSEG surpasses state-of-the-art methods in multiple settings across these three datasets. Our extensive ablation studies confirm the optimization of the uniformity and alignment objectives of the feature representations.", + "bbox": [ + 88, + 325, + 485, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 737, + 220, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation, the task of assigning pixel-level labels to images, is a cornerstone in domains such as autonomous driving and medical image analysis. Semantic segmentation models typically comprise an encoder and a decoder. The encoder extracts meaningful features from the input images, while the decoder assigns the feature embeddings to desired pixel-level labels. However, its effectiveness is often hampered by the need for costly pixel-level annotated data [9]. To mitigate this, semi-supervised learning", + "bbox": [ + 89, + 763, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d04dc6f268f1caeefd63cc33ccfe157d8fa1583edab6aafff1b48f678c66697.jpg", + "image_caption": [ + "Figure 1. SWSEG demonstrates marked superiority in mIoU across varying data splits of the ADE20K dataset." + ], + "image_footnote": [], + "bbox": [ + 517, + 292, + 901, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(SSL) was proposed [3]. SSL leverages a limited amount of labeled data in conjunction with a larger pool of readily available unlabeled data. The key lies in effectively harnessing the unlabeled data to improve the model's performance.", + "bbox": [ + 511, + 547, + 906, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pseudo-labeling and consistency regularization are two prominent techniques in SSL. Pseudo-labeling generates temporary labels from the model's predictions during training and using these to train the model in a supervised manner [2, 26]. The model then computes the loss between these pseudo-labels and the corresponding unlabeled images. This technique allows the model to leverage unlabeled data, thereby improving its performance. Consistency regularization, on the other hand, encourages the model to produce consistent predictions for different perturbations of the same input image, thereby promoting robustness and generalization [14, 32]. This technique is grounded in the smoothness assumption, which posits that data points with similar features should have similar labels [3]. In practice, random augmentations are applied to input images, and the model is trained to minimize discrepancies between its predictions on the original and augmented versions. In the realm of semantic segmentation, consistency regularization is often paired with pseudo-labeling, where weakly augmented unlabeled", + "bbox": [ + 511, + 613, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20233", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/363573a54e3ddec1f229574c2dd038224043414501d1f3e6b592354548ef1e97.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 189, + 155 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8f388bc10592902608f270c767326baf347696dd2bf645e34e6a86869ff5c962.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 202, + 88, + 290, + 155 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/01a54dbd302f93ff10838d9066ee8e876236316dd46e4aad4b20d3933c70f1bb.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 236, + 167, + 292, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/79eae512f77a09ccb44ea291b6e727636c74ebe06b85847c61b13223de5950cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 88, + 383, + 155 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/870460a032788f40fba01022108f754aa58b82ef9d06ffa3533388661a05e5bd.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 297, + 172, + 383, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/23dc3582923f82acc9376da1413e11d050b12f53684b9dc621631e10f8fb8c56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 88, + 478, + 156 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ec35b044e279de6d6c6f667d86e63bfa8a340e31273e7abee1694b029446d1af.jpg", + "image_caption": [ + "(c)", + "Figure 2. Grad-CAM visualization [33] and model predictions. (a) depicts results from a model trained without alignment optimization, leading to inadequate boundary definition. (b) shows moderate improvements using traditional cross-entropy loss, which partially improves boundary adherence. (c) illustrates SWSEG with SWD loss, significantly enhancing segmentation accuracy by ensuring uniformity and alignment of features." + ], + "image_footnote": [], + "bbox": [ + 392, + 172, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "data generate temporary \"pseudo\" labels for training [2, 34]. Concurrently, a parallel training stream employs strongly augmented data to calculate and minimize consistency losses, further enhancing model robustness [8, 23, 29, 46, 50].", + "bbox": [ + 89, + 400, + 483, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While consistency-based methods are effective, existing approaches focus solely on enforcing regularization at the decoder's output, neglecting the rich information embedded within the intermediate feature representations from the backbone encoder. This oversight neglects the potential of extending consistency regularization to the feature level, guided by the smoothness assumption. We argue that enforcing consistency within these feature representations, derived from CNN-based or Transformer-based backbone encoders [4, 43], is crucial for robust learning against image perturbations. In Fig. 2, we insert a cross entropy loss on the output of encoders between feature embeddings of weakly and strongly augmented images. The results show that it already improves the feature separability and segmentation performance.", + "bbox": [ + 89, + 462, + 483, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To explore the importance of superior feature representations, we draw parallels with the self-supervised contrastive learning literature [1, 6, 48]. These methods, which involve contrasting positive (similar) and negative (dissimilar) pairs, excel at promoting both alignment and uniformity [13, 40]. Alignment, which aligns with the smoothness assumption, posits that perturbations to the same input should yield proximate features. Uniformity, on the other hand, implies that feature representations should be maximally informative, a concept often overlooked in semi-supervised semantic segmentation. These metrics provide a comprehensive framework for evaluating and enhancing feature representations, ultimately leading to improved model performance. The attraction between positive pairs in contrastive learning fosters", + "bbox": [ + 89, + 688, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "alignment, while the repulsion of negative pairs encourages diverse and informative representations, enhancing uniformity. Inspired by the effectiveness of contrastive learning, we aim to bridge the gap in semi-supervised semantic segmentation by developing a method that explicitly optimizes both alignment and uniformity of feature representations. This raises a key question: How can we design an objective function that simultaneously achieves both uniformity and alignment of feature representations?", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce SWSEG, a semi-supervised semantic segmentation algorithm that directly addresses this question. We leverage the Sliced-Wasserstein distance (SWD), a metric that quantifies the distance between probability distributions, to simultaneously optimize alignment and uniformity of feature representations extracted from the backbone encoder model. SWD achieves this by projecting features onto a unit hypersphere and computing the Wasserstein distance between the resulting one-dimensional embeddings. Empirically, we demonstrate that minimizing SWD effectively optimizes both alignment and uniformity between feature representations. Furthermore, to address the computational burden associated with traditional Monte Carlo estimation of SWD for high-dimensional features, we introduce a variant that projects feature embeddings onto a Gaussian distribution. This not only maintains the uniformity objective, as a normalized Gaussian distribution is uniformly distributed on the unit hypersphere, but also enables efficient computation by leveraging the analytical solution for the quadratic Wasserstein distance between two Gaussian distributions. Additionally, we incorporate a regularization term that decorrelates feature representations, fulfilling the necessary condition for projecting features onto a Gaussian distribution. We evaluated our algorithm on the PASCAL VOC 2012 [12], Cityscapes [9], and ADE20K [51] datasets under various partitioning protocols. Our results consistently demonstrate SWSEG's superiority over supervised baselines, with performance improvements of up to $11.8\\%$ , $8.9\\%$ , and $8.2\\%$ on PASCAL VOC, CityScapes, and ADE20K, using the same amount of labeled data. Notably, SWSEG also achieves state-of-the-art (SOTA) results compared to existing methods under multiple data splitting protocols, as illustrated in Figure 1 [19, 23, 27]. Through extensive ablation studies, we confirm the critical role of uniformity in feature space for effective semi-supervised semantic segmentation. To the best of our knowledge, SWSEG is the first to establish a direct link between SSL and the uniformity metric, while also implementing the SWD to enhance feature representations. Our contributions can be summarized as follows:", + "bbox": [ + 511, + 228, + 908, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We introduce SWSEG, a semi-supervised semantic segmentation algorithm that explicitly optimizes feature representation uniformity and alignment using SWD, and back up our claim with extensive empirical studies.", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20234", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. We introduce an efficient variant of SWD estimation that projects feature embeddings onto a Gaussian distribution. This approach maintains uniformity while leveraging the analytical solution for quadratic Wasserstein distance between Gaussians, reducing computational complexity.", + "3. SWSEG achieves SOTA results on PASCAL VOC 2012, Cityscapes, and ADE20K datasets, outperforming supervised baselines and existing semi-supervised methods, with significant performance improvements of up to $11.8\\%$ on PASCAL VOC, $8.9\\%$ on CityScapes, and $8.2\\%$ on ADE20K compared to supervised methods." + ], + "bbox": [ + 86, + 90, + 482, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 273, + 240, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Semi-Supervised Learning", + "text_level": 1, + "bbox": [ + 89, + 299, + 331, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semi-supervised learning (SSL) has emerged as a promising solution to alleviate the burden of collecting labeled data, enabling the training of DNNs using only a fraction of labeled data while still achieving impressive results. Pseudo-labeling is a widely adopted technique which leverages the model's predictions on unlabeled data as surrogate labels [21]. This approach has been shown to improve model generalization by shifting the decision boundary towards low-density regions [3]. Additionally, many SSL algorithms incorporate consistency regularization, which regularizes the model to generate consistent outputs regardless of perturbations applied to the input data [20, 35, 38]. Laine et al. [20] combined cross-entropy loss with consistency loss for training, and incorporated model predictions into the pseudo-label set through an exponential moving average. Tarvainen and Valpola [38] posited that directly updating the model weights at each mini-batch accelerates the learning process while enhancing the model's performance. Sohn et al. [35] generated pseudo-labels from model predictions on weakly-augmented input images and computed the unsupervised loss between these pseudo-labels and the model predictions from strongly-augmented input images. However, these SSL methods were designed for image classification, which is not suitable for semantic segmentation where pixel-wise predictions are desired.", + "bbox": [ + 91, + 321, + 483, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Semi-Supervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 89, + 710, + 439, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A number of research papers have incorporated consistency training and pseudo-labeling in SSL for segmentation tasks [29, 37, 39, 42, 44-46, 50]. For instance, Wang et al. [42] stored pseudo-labels with lower confidence as negative samples and computed contrastive loss using these samples. Yang et al. [46] introduced an additional teacher model and used dropout as feature-wise perturbation for consistency-based pseudo-labeling. However, these methods typically enforce consistency regularization solely on the decoder output, neglecting the rich information present in the intermediate feature representations produced by the", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "backbone encoder. Recent works leveraging feature embedding in SSL segmentation include AllSpark [39], DDFP [41], and CorrMatch [36]. AllSpark balances the importance of labeled and unlabeled training streams by implementing a cross-attention mechanism in the feature embedding. DDFP injects perturbation in the feature space to regularize the decision boundary, with perturbation locations guided by a lightweight normalizing flow-based density estimator. CorrMatch utilizes the correlation map of features from strongly and weakly augmented inputs to regularize pseudo-label supervision in the decoder output. In contrast, SWSEG aims to fill this gap by explicitly focusing on training the backbone encoder to produce more meaningful feature representations for the segmentation head. By optimizing both the alignment and uniformity of these feature representations, SWSEG aims to improve overall segmentation performance. This approach goes beyond the limitations of existing methods that primarily focus on designing complex strategies such as cross-attention layers, normalizing flows, or heuristic label and region propagation techniques.", + "bbox": [ + 511, + 90, + 906, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Uniformity of Feature Representation", + "text_level": 1, + "bbox": [ + 511, + 402, + 838, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although existing SSL methods for semantic segmentation have demonstrated promising results through consistency learning, they typically impose consistency on the model's output. While few studies have explored consistency regularization on feature representations from the backbone model, this area remains under-explored compared to extensive research in self-supervised learning [1, 6, 48]. In self-supervised learning, where labeled data is scarce or absent, preventing representational collapse is paramount. This collapse occurs when the model produces uninformative feature representations, regardless of variations in the input. While the uniformity metric, which assesses how evenly features are distributed across the hypersphere, has been effectively used in self-supervised learning to mitigate collapse [40], its application in semi-supervised semantic segmentation has been absent. Recognizing the potential benefits of uniformity in preventing representational collapse and maximizing information preservation, we incorporate this metric into our semi-supervised approach. We further establish a connection between uniformity and the sliced-SWD metric, demonstrating that minimizing SWD between feature representations of differently augmented inputs can simultaneously maximize feature uniformity. This unified approach enhances model performance by ensuring that feature representations are both consistent and informative.", + "bbox": [ + 511, + 424, + 908, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 814, + 604, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section begins by introducing the conventional framework for consistency learning-based SSL. We then argue for the importance of enforcing consistency regularization in the feature embedding, discussing the roles of alignment and", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20235", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/20baeedbf32f3a51be025b25bef2c7f1f6db7308e3bcefa90bde067e91d38023.jpg", + "image_caption": [ + "Figure 3. SWSEG Framework for Joint Optimization of Alignment and Uniformity. The SWSEG framework employs both weak $(\\mathcal{A}^w)$ and strong perturbations $(\\mathcal{A}^s)$ applied to unlabeled input images $x^u$ . Both augmentations are passed through a shared backbone encoder to extract features, which are then mapped to a different space via non-linear MLPs, producing feature embeddings. To achieve both feature alignment and uniformity, SWSEG introduces an SWD loss $\\mathcal{L}_{swd}$ , computed between these intermediate feature embeddings. The goal is to encourage the feature embeddings from different augmentations to be consistent (alignment) and to be spread out uniformly in the feature space, maximizing information (uniformity). An unsupervised loss $\\mathcal{L}_u$ is computed to ensure consistency between segmentations generated from weakly and strongly augmented inputs." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 911, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "uniformity in achieving better feature representations. Next, we introduce a Gaussian-projected SWD metric to simultaneously achieve uniformity and alignment in the feature space. Finally, we detail the integration of these elements within the SWSEG framework.", + "bbox": [ + 88, + 477, + 485, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 561, + 220, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In semi-supervised learning, we are given labeled image sets, $D^{l} = \\{x_{i}^{l},y_{i}^{l}\\}$ , and unlabeled ones, $D^{u} = \\{x_{j}^{u}\\}$ . Typically, $D^{u}$ greatly outnumbers $D^{l}$ . Our goal is to leverage these unlabeled images to optimize the parameters $\\theta$ of a semantic segmentation model $f_{\\theta}(\\cdot)$ .", + "bbox": [ + 89, + 584, + 483, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In SWSEG, we adopt a typical framework from Fix-Match [34], training the model through two separate feedforward streams: supervised and unsupervised. In the supervised stream, we draw a batch of images from $D^{l}$ and compute the cross-entropy loss $\\mathcal{L}_s$ between the model's predictions $\\hat{y}_i^l$ and the labels $y_i^l$ . Conversely, the unsupervised stream draws images from $D^{u}$ , to which we apply weak or strong perturbations, $\\mathcal{A}^w(\\cdot)$ and $\\mathcal{A}^s(\\cdot)$ , respectively. The unsupervised loss $\\mathcal{L}_u$ is then minimized according to the equation (1):", + "bbox": [ + 89, + 662, + 483, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u} = \\frac {1}{| \\mathcal {D} ^ {u} |} \\sum_ {x \\in \\mathcal {D} ^ {u}} \\mathcal {H} \\left(f _ {\\theta} \\left(\\mathcal {A} ^ {s} \\left(x ^ {u}\\right)\\right), f _ {\\theta} \\left(\\mathcal {A} ^ {w} \\left(x ^ {u}\\right)\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 823, + 483, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We generate pseudo-labels $f_{\\theta}(\\mathcal{A}^{w}(x^{u}))$ by feeding weakly-augmented images into the model. We then compute", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the unsupervised loss using strongly-augmented samples $f_{\\theta}(\\mathcal{A}^{s}(x^{u}))$ , using $\\mathcal{H}$ , for cross-entropy loss. The pseudolabel generation does not contribute to gradient computation.", + "bbox": [ + 511, + 477, + 906, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Feature Alignment and Uniformity", + "text_level": 1, + "bbox": [ + 511, + 532, + 816, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In semantic segmentation, $f_{\\theta}(\\cdot)$ typically comprises a backbone encoder $p_{\\theta}$ and a segmentation head $q_{\\theta}$ [4, 43]. Existing methods [23, 27, 37, 46, 50] focus on enforcing consistency regularization on the segmentation head's output, often overlooking backbone encoder features. To further harness the potential of feature representations, we draw on self-supervised contrastive learning, which trains networks to generate meaningful embeddings without labels, emphasizing alignment and uniformity as key objectives for assessing feature quality.", + "bbox": [ + 511, + 555, + 906, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Alignment aims to minimize differences between embeddings from similar samples. Intuitively, model should produce similar feature embeddings when images with same classes are input. We achieve this by varying the magnitude of perturbations applied to the input images and minimizing the distance between embeddings using a distance metric dist:", + "bbox": [ + 511, + 707, + 906, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = \\operatorname {d i s t} \\left(p _ {\\theta} \\left(\\mathcal {A} ^ {s} \\left(x _ {i} ^ {u}\\right)\\right), p _ {\\theta} \\left(\\mathcal {A} ^ {w} \\left(x _ {i} ^ {u}\\right)\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 825, + 906, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Uniformity promotes even distributions of feature representations across a hypersphere, enhancing model robustness by preventing collapse to a lower-dimensional subspace.", + "bbox": [ + 511, + 854, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20236", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While often less emphasized in semi-supervised learning due to the stabilizing influence of labeled data, uniformity is critical in self-supervised learning to prevent the model from generating identical embeddings regardless of input. By ensuring uniformity, the model preserves a rich diversity of information within the feature space, which significantly improves the performance of the segmentation head.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Revisiting Sliced Wasserstein Distance", + "text_level": 1, + "bbox": [ + 89, + 205, + 419, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, the uniform distribution on a unit hypersphere exemplifies maximum uniformity. We need to establish an objective function that can achieve both uniformity and alignment. To this end, we implement the SWD as our distance metric. SWD calculates average distances from 1-D projections sampled from a unit hypersphere, thus inherently promoting both desired properties.", + "bbox": [ + 89, + 228, + 483, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The rationale behind SWD stems from its ability to reduce the computational complexity relative to the high-dimensional Wasserstein distances (WD). Unlike the traditional WD, which struggles with high dimensionality, SWD leverages the 1-dimensional Wasserstein distance, which can be computed in closed form: Let $\\mu_d$ and $\\nu_d$ be two probability measures in $\\mathcal{P}(\\Omega) \\in \\mathbb{R}^d$ :", + "bbox": [ + 89, + 334, + 483, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW _ {p} ^ {p} \\left(\\mu_ {d}, \\nu_ {d}\\right) = \\left(\\int_ {0} ^ {1} \\left| F _ {\\mu_ {d}} ^ {- 1} (z) - F _ {\\nu_ {d}} ^ {- 1} (z) \\right| ^ {p} d z\\right) ^ {1 / p}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 452, + 483, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p$ is the order of the distance, and $F_{\\mu_d}^{-1}$ and $F_{\\nu_d}^{-1}$ are the quantile functions of $\\mu_d$ and $\\nu_d$ , respectively. Computationally, this is achieved by sorting and computing the mean $l^2$ -distance between two 1-D vectors. SWD [30] involves projecting high-dimensional features onto a unit hypersphere and calculating the 1-D Wasserstein distance between these projections. Formally, let $\\mathbb{S}^{d-1}$ represent the unit sphere in $d$ dimensions, $\\sigma$ the uniform distribution on $\\mathbb{S}^{d-1}$ , and $P_{\\theta\\#}\\xi$ the push-forward measure of $\\xi \\in \\mu_d, \\nu_d$ . The SWD is defined as:", + "bbox": [ + 89, + 500, + 483, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS W _ {p} ^ {p} \\left(\\mu_ {d}, \\nu_ {d}\\right) = \\int_ {\\mathbb {S} ^ {d - 1}} W _ {p} ^ {p} \\left(P _ {\\theta \\#} \\mu_ {d}, P _ {\\theta \\#} \\nu_ {d}\\right) \\mathrm {d} \\sigma (\\theta). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 661, + 483, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To approximate SWD, we vectorize the input distributions $\\mu_{d}$ and $\\nu_{d}$ from $\\mathbb{R}^{n\\times n\\times d}$ to $\\mathbb{R}^{n^2 d}$ . We then employ a Monte Carlo scheme, sampling $L$ directions from the unit sphere $\\mathbb{S}^{d - 1}$ . Projecting $\\mu_{d}$ and $\\nu_{d}$ along these directions yields $L$ one-dimensional projections for each distribution. After sorting these projections, we calculate the 1D Wasserstein distance between them using Eq. (3). This method efficiently estimates the Wasserstein distance between feature embeddings, exploiting the closed form property of 1D calculations compared to high-dimensional alternatives.", + "bbox": [ + 89, + 702, + 483, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the unit sphere projection enables uniformity and minimizing 1-D Wasserstein distances facilitates alignment, training with Monte Carlo-approximated SWD", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "presents several challenges. First, the results are inherently non-deterministic and intractable due to the randomness inherent in the Monte Carlo method. Second, as the dimensions of $\\mu_d$ and $\\nu_d$ increase, both the computational cost and memory requirements for SWD escalate. Finally, to maintain accuracy in the face of increasing dimensions, the number of directions $L$ must also increase, complicating the Monte Carlo approximation further.", + "bbox": [ + 511, + 90, + 905, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Proposed Approach: SWSEG", + "text_level": 1, + "bbox": [ + 511, + 220, + 769, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance feature representation uniformity and alignment while alleviating the computational burden of Monte Carlo-based SWD, we implement an alternate method proposed by Nadjahi et al. [28] for approximating the SWD. Our method diverges by projecting these measures directly onto a Gaussian distribution. This modification offers significant computational benefits and supports key theoretical objectives in two primary ways: First, projecting onto Gaussian distributions allows the 2-Wasserstein distance between any two Gaussians to be expressed in a closed-form equation. Second, using normalized Gaussian variables ensures that the resulting distributions are uniformly distributed over the unit hypersphere, which we prove in the supplementary materials. This uniform distribution is essential for maintaining the model's uniformity objective, ensuring features are evenly distributed in the embedding space, maximizing its information for segmentation head.", + "bbox": [ + 511, + 241, + 906, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To formalize this idea, for any positive integer $d$ , let $\\{\\theta_i\\}_{i=1}^d$ be a sequence of i.i.d. 1-d standard Gaussian distributions and $\\{X_i\\}_{i=1}^d$ be a sequence of 1-d probability distributions. Several Central Limit Theorems [11, 31] ensure that the sequence of distributions formed by the inner product $d^{-1/2}\\langle \\theta_{1:d}, X_{1:d} \\rangle$ converges in probability to a Gaussian random variable when the distributions exhibit weak correlation. Furthermore, according to Proposition 2 from [28], as the dimensions of the distributions increase, the quadratic Wasserstein distance between two zero-mean Gaussian distributions with variance $d^{-1}m_2(\\bar{\\mu}_d)$ can be approximated by the SW distance between two Gaussian projected distributions. The Wasserstein distance between two Gaussian probability distribution is in closed form [10], which enables us to approximate the $l^2$ -SW distance by:", + "bbox": [ + 511, + 498, + 908, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s w d} = d ^ {- 1} \\left(m _ {2} \\left(\\bar {\\mu_ {d}}\\right) ^ {\\frac {1}{2}} - m _ {2} \\left(\\bar {\\nu_ {d}}\\right) ^ {\\frac {1}{2}}\\right) + d ^ {- 1} \\left| m _ {\\mu_ {d}} - m _ {\\nu_ {d}} \\right| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 731, + 906, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This equation comprises two terms: the first quantifies the difference in the second moments of the distributions, and the second captures the squared error in their means. For $\\xi_d \\in \\{\\bar{\\mu}_d, \\bar{\\nu}_d\\}$ , we define $m_2(\\xi_d) = n^{-1} \\sum_{j=1}^n ||x^j||^2$ , enabling a closed-form deterministic solution for $\\mathcal{L}_{swd}$ . Nadjahi et al. also proved in Proposition 1 that $\\mathcal{L}_{swd}$ with Gaussian projections equals the original $SW_p^p$ when $p = 2$ . Thus, we can use $\\mathcal{L}_{swd}$ as a computationally efficient Wasserstein metric for our consistency loss term. A PyTorch-styled", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20237", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/951ce2d6bcca966a7d0f1e9761d5111e4c0e7df916fc04d66c684c4876b88514.jpg", + "table_caption": [ + "Table 1. Results on ADE20k. The fractions represent the labeled image proportions sampled during training. The model is trained on SegFormer [43] with MiT-B1 backbone. We use mIOU as our evaluation metric." + ], + "table_footnote": [], + "table_body": "
Method1/32 (631)1/16 (1263)1/8 (2526)1/4 (5052)1/2 (10105)
Supervised Only15.222.325.929.532.6
CPS [8] [ICLR '21]18.822.327.932.436.9
U²PL [42] [CVPR '22]18.523.628.133.536.0
ReCo [22] [ICLR '22]21.123.727.730.235.5
GTA-Seg [19] [NeurIPS '22]19.523.126.927.829.9
DT [27] [NeurIPS '23]22.626.030.633.737.3
SWSEG (Ours)24.426.730.934.638.0
", + "bbox": [ + 163, + 127, + 834, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pseudo code is presented in Algorithm 1. Furthermore, to ensure a more accurate approximation of the Gaussian-SWD and comply with the relaxed Central Limit Theorems in [11, 31], minimizing strong correlations among input embeddings is essential. We address this by introducing a regularization term based on the input cross-covariance matrix into the SWD loss:", + "bbox": [ + 88, + 296, + 483, + 401 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e g} = | \\operatorname {C o v} (X) - \\operatorname {D i a g} (X) | _ {2} ^ {2} + | \\operatorname {C o v} (Y) - \\operatorname {D i a g} (Y) | _ {2} ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 410, + 483, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\mathrm{Cov}(\\cdot)$ calculates the covariance matrix of the input features, and $\\mathrm{Diag}(\\cdot)$ extracts its diagonal elements. Minimizing the off-diagonal elements decorrelates the features, thereby indirectly promoting the uniformity of the distribution of the features across the representation space [49].", + "bbox": [ + 89, + 441, + 483, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present a detailed overview of this training scheme in Fig. 3. We also insert an additional non-linear MLP projection layer before loss computation between representations, following recent self-supervised learning work [6, 7, 16]. The projection layer is solely incorporated during the training process, and it is subsequently discarded when the model is deployed for inference. Our objective is to retain more information in the features prior to the MLP layer. This will enable the segmentation head to utilize the preserved information more effectively for the segmentation task. The specific architecture of the projection layer used in this work consists of two Conv - BatchNorm - ReLu blocks followed by a single Conv layer. The overall training objective, managing the balance of various losses, is summarized in Eq. (7). The hyperparameters $\\lambda_1, \\lambda_2$ , and $\\lambda_3$ control the weight of each loss.", + "bbox": [ + 89, + 518, + 483, + 758 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o v e r a l l}} = \\mathcal {L} _ {s} + \\lambda_ {1} \\mathcal {L} _ {u} + \\lambda_ {2} \\mathcal {L} _ {s w d} + \\lambda_ {3} \\mathcal {L} _ {\\text {r e g}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 785, + 483, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 823, + 207, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 848, + 192, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ADE20K. [51] is a complex scene parsing dataset featuring 150 objects, with 25,574 training, 2,000 validation, and 3,000", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Pseudocode of our Gaussian-SWD loss in a PyTorch-like style." + ], + "code_body": "def gaussian_swd_loss(x, y):\n # Reshape the input tensors to (n, dim)\n x, y = x.view(x.shape[0], -1), y.view(y.shape[0], -1)\n n, dim = x.shape\n meanx = torch.mean(x, dim=0)\n xc = x - meanx # Zero mean\n m2x = torch.mean(torch.linalg(norm(xc, dim=1)) * 2) / dim\n meany = torch.mean(y, dim=0)\n yc = y - meany # Zero mean\n m2y = torch.mean(torch.linalg(norm(yc, dim=1)) * 2) / dim\n m_t = torch.linalg(norm(meanx - meany) * 2 / dim\n swd = m_t + (m2x * (1/2) - m2y * (1/2)) * 2\n return swd", + "guess_lang": "python", + "bbox": [ + 514, + 333, + 903, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "testing images. For a fair comparison, we used the data split from dual-teacher [27].", + "bbox": [ + 511, + 575, + 906, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PASCAL VOC 2012. [12] This dataset is a classic semantic segmentation dataset comprising 20 classes of objects and 1 background class, with 1464 training, 1449 validation, and 1546 testing high-quality images in the standard set. We incorporate the augmented set from [17], following previous works, increasing our labeled data to 10,582 images.", + "bbox": [ + 511, + 607, + 905, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cityscapes. [9] This is an autonomous driving semantic segmentation dataset that consists of 19 semantic classes, with 2975, 500, and 1525 training, validation, and testing images, respectively.", + "bbox": [ + 511, + 699, + 906, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation details.", + "text_level": 1, + "bbox": [ + 511, + 772, + 732, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For ADE20K, we implemented our method with SegFormer [43] architecture with MiT-B1 backbone. For PASCAL VOC 2012 and CityScapes, to have a fair comparison with previous benchmark semi-supervised semantic segmentation methods [27, 37, 42, 45, 46, 50], we implement our method using ResNet-101 [18] as the backbone encoder, initialized with ImageNet pre-trained weights, and", + "bbox": [ + 511, + 794, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20238", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3c8b59238dc1adbc7a4288d9af9d941bdcec621d2dbdb927835a680598be8dcf.jpg", + "table_caption": [ + "Table 2. Results on CityScapes dataset. The fractions represent the proportion of labeled images used during training. We use mIOU as our evaluation metric." + ], + "table_footnote": [], + "table_body": "
Method1/161/81/4
Supervised Only67.973.275.1
PCR [44] [NeurIPS '22]73.476.378.4
UniMatch [46] [CVPR '23]76.677.979.2
AugSeg [50] [CVPR '23]75.277.879.6
DT [27] [NeurIPS '23]76.878.479.5
DAW [37] [NeurIPS '23]76.678.479.8
CorrMatch [36] [CVPR '24]77.378.579.4
RankMatch [24] [CVPR '24]77.178.680.0
DDFP [41] [CVPR '24]77.178.279.9
SWSEG (Ours)76.878.679.5
", + "bbox": [ + 107, + 141, + 468, + 330 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fe236b55bddd7a61e3215bb42db06ce9df9eee17f81d54a6db8a36d70a011d1d.jpg", + "table_caption": [ + "Table 3. Results on PASCAL VOC 2012 dataset. The fractions represent the proportion of labeled images used during training. We use mIOU as our evaluation metric." + ], + "table_footnote": [], + "table_body": "
Method1/161/81/4
Supervised Only67.271.373.5
UniMatch [46] [CVPR '23]75.076.877.5
AugSeg [50] [CVPR '23]77.077.378.8
DAW [37] [NeurIPS '23]78.578.979.6
CorrMatch [36] [CVPR '24]78.479.379.6
RankMatch [24] [CVPR '24]78.979.280.0
DDFP [41] [CVPR '24]78.378.979.8
SWSEG (Ours)79.079.379.9
", + "bbox": [ + 107, + 385, + 468, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DeeplabV3+ [5] as the decoder. We use SGD optimizer with momentum of 0.9. The learning rate is set to 0.001, 0.02, and 5e-5 for PASCAL, CityScapes, and ADE20K, respectively. We utilize a polynomial scheduling strategy with 0.9 decay rate.", + "bbox": [ + 89, + 590, + 485, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data augmentations. For weak augmentations, horizontal flipping, random cropping, and random scaling are applied to the training images. For strong augmentations, we apply random color jittering, random grayscale, random blurring, and Cutmix [47] to the input training images. The training image size is set as $513 \\times 513$ , $769 \\times 769$ , and $512 \\times 512$ for PASCAL VOC, CityScapes, and ADE20K, respectively. The training epochs for VOC, CityScapes, and ADE20K is 80, 240, and 200 epochs.", + "bbox": [ + 89, + 676, + 485, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation. As in previous works, we employ mean Intersection-over-Union (mIOU) as our evaluation metric. For the PASCAL VOC 2012 and ADE20K, evaluations are on the center-cropped validation set, while for CityScapes, we employ a sliding window evaluation.", + "bbox": [ + 89, + 824, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/55f0cf25f709913130ba86b2259790fe8b91a130bcbac867c323c7d3b8588f23.jpg", + "table_caption": [ + "Table 4. Ablation analysis of feature alignment loss functions, $\\mathcal{L}_{reg}$ , and MLP layer. The results reveal that MSE, cross entropy, and Monte Carlo estimated SWD loss with $L = 128$ yield inferior results. The combination of Gaussian-SWD (GSWD) with $\\mathcal{L}_{reg}$ and MLP produces the best performance, with an mIoU of 79.0." + ], + "table_footnote": [], + "table_body": "
Loss FunctionsLregMLPmIoU
MSECESWDGSWD
✓✓✓70.2
✓✓✓77.1
✓✓✓77.6
✓77.4
✓✓78.1
✓✓✓79.0
", + "bbox": [ + 517, + 169, + 903, + 311 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/47e7fc9183529c562b574c92a315ff75e23164a0521f7f3a9ee06a3f2f386a13.jpg", + "table_caption": [ + "Table 5. Analysis of the Uniformity value $\\mathcal{L}_{\\mathrm{uniform}}$ across different data splits of the PASCAL VOC dataset demonstrates that training with the SWD loss $\\mathcal{L}_{swd}$ yields superior Uniformity values compared to training without this loss term." + ], + "table_footnote": [], + "table_body": "
Method1/81/21
Without Lswd-2.5896-2.7095-2.5574
With Lswd-2.6592-2.7104-2.6168
", + "bbox": [ + 547, + 381, + 874, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 511, + 469, + 818, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on ADE20K Dataset. Tab. 1 demonstrates the performance of SWSEG against the supervised baseline and existing methods. SWSEG significantly improves over the supervised baseline by $8.2\\%$ , $3.4\\%$ , $3.7\\%$ , $5.1\\%$ , and $5.4\\%$ mIoUs on the 1/32, 1/16, 1/8, 1/4, and 1/2 splits, respectively. Our method also outperforms the state-of-the-art across all data splits, showcasing the generalizability of SWSEG across different model architectures and datasets.", + "bbox": [ + 511, + 491, + 908, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Cityscapes Dataset. Tab. 2 presents our results in comparison to the supervised baseline and SOTA methods. SWSEG outperforms the supervised baseline by $8.9\\%$ , $5.3\\%$ , and $4.4\\%$ for the 1/16, 1/8, and 1/4 partitions, respectively. Additionally, SWSEG surpasses the previous state-of-the-art method, DAW, by $0.2\\%$ in the 1/16 and 1/8 partitions, and performs on par with contemporary methods.", + "bbox": [ + 511, + 613, + 908, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on PASCAL VOC 2012 Dataset. Tab. 3 compares SWSEG with existing state-of-the-arts on PASCAL VOC 2012 set. We follow the data partitions in CPS [8]. We outperform the supervised baselines by $11.8\\%$ , $8.0\\%$ and $6.4\\%$ mIoUs under 1/16, 1/8, and 1/4 partitioning protocols. Our method also beats the SOTA DDFP [37] by $0.7\\%$ , $0.4\\%$ , and $0.1\\%$ mIoUs under 1/16, 1/8, and 1/4 splits, respectively. We also visualized the predicted features in Fig. 4 to evaluate the effectiveness of our algorithm's clustering capabilities. Our method demonstrates enhanced discriminative clustering compared to the Supervised Only and UniMatch methods.", + "bbox": [ + 511, + 719, + 910, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20239", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2bcc5edd985fb3aefa80d72e8b9e171aeeb3194b10f367119670c9da78c2923a.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 122, + 90, + 215, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd06d584a69794e256f67beefe020733ea72ba308d5cf8662645803d4b0bc696.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 238, + 90, + 334, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/76af2cb30ff0c58380c0b74fe2fd4002beaf5cf1c78f08d1db7a3afedfa25dd4.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 357, + 90, + 452, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97d2962bbd79f7d564eabbe26f05508d9c01186fe8d3c61bd8a13d0476167de7.jpg", + "image_caption": [ + "Figure 4. UMAP visualization [25] of feature embeddings using models trained with: (a) Supervised Only, (b) UniMatch, and (c) SWSEG. These visualizations utilize the PASCAL VOC 2012 dataset. Our method demonstrates enhanced clustering capabilities, outperforming both the Supervised Only and UniMatch.", + "(a) Without $\\mathcal{L}_{swd}$", + "Figure 5. We visualize feature representations on a unit circle for models trained with and without $\\mathcal{L}_{swd}$ for uniformity analysis. Features from the model with $\\mathcal{L}_{swd}$ (b) are more uniformly distributed than those without $\\mathcal{L}_{swd}$ (a)." + ], + "image_footnote": [], + "bbox": [ + 94, + 263, + 284, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/14e7a1bf8bb5db89879cfcc0473cc58d49804bc28bb400328dc30a189120cfd2.jpg", + "image_caption": [ + "(b) With $\\mathcal{L}_{swd}$" + ], + "image_footnote": [], + "bbox": [ + 292, + 263, + 480, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 575, + 243, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Uniformity and Alignment. We conduct an empirical analysis of feature representation alignment and uniformity (Tab. 4). We compare different loss functions for enforcing consistency regularization on the feature representation. Our results indicate that, aside from MSE loss, additional loss terms generally enhance segmentation prediction performance, highlighting the critical role of feature alignment. Monte Carlo SWD with $L = 128$ outperforms MSE and cross-entropy but encounters out-of-memory issues at $L = 1000$ using an identical number of GPUs, highlighting its computational cost. Conversely, our Gaussian-based SWD approximation delivers optimal results with lower memory usage. For uniformity analysis, we reduce the dimension of feature embeddings using PCA [15], projecting them into $\\mathbb{R}^2$ and plotting their coordinates on a unit circle, as shown in Fig. 5. The visualization demonstrates that models trained with $\\mathcal{L}_{swd}$ achieve greater uniformity compared to those without it, confirming the effectiveness of our introduced SWD loss. To quantitatively validate this observation, we compute the uniformity metric $\\mathcal{L}_{\\text{uniform}}$ from [40, Equa", + "bbox": [ + 89, + 598, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25dc148c14a869e9ea05a85f48e481ba8c75618295ab9660bbc703ffc26c921f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 87, + 591, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3b7ee4544be51e46676b3ef4518189053923ec1551fa175a98b75608c252345.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 88, + 669, + 141 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2b32a8bd06fe0edd0414e9fb70129c6e10ee0964f69ce01b01272acdbd8c0033.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 142, + 669, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f2f8303d9a8d80399ee63bb7a0ca8e31c653bf468d363def389bd91d73544b9.jpg", + "image_caption": [ + "(b)", + "Figure 6. Qualitative results on PASCAL VOC 2012 validation set. All models are trained under 1/4 partition. (a) Input images. (b) Ground Truth labels. (c) Model output when trained with only labeled images. (d) Model output when training without SWD loss. (e) Predictions from SWSEG. When including SWD loss, the predictions are better in ambiguous regions and object borders." + ], + "image_footnote": [], + "bbox": [ + 594, + 196, + 669, + 304 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8d6ebeb99057817d16cdfde725180479b9a498e2fb1f138d5a16deec45eaa74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 88, + 746, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51d22d8443db3fe8e93042ee27abb3ecc681670ce7e4f92062dc16bca18aece2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 142, + 746, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/afedc0bf2e3222106dbca9c402d809adc6925b60a5419bb85d867d48b561206d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 196, + 746, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb590bec39f2b70943f64c5dcb2ac2ee822a43badc4a28db20bf28f58ee49b2f.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 671, + 250, + 746, + 304 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8deed34be5a158aa0f11ab540e16dd665e8f924da735642a9d3e1f9e51e0bd34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 88, + 825, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/381abb47076c5d9e82c1831b79481d75043d015f8164351aab7fb81202cefbac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 142, + 825, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e7865b2343e42292a4c971b63f11acbabdd87d48fbbaad421a82d1e6f79f37a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 196, + 825, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f4ca899082d2907cc60405d4f86794b7eaaa8f54dce06e65c1f67cf2c6d51dc6.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 750, + 250, + 825, + 304 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8db54dce7a8c91175a5397f7ec71e5786289681d6dff60421a99184c3d2cad29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 88, + 903, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4dc2fd916967a116d046a2a848d057195f5e25734615d3392fa1d58006179043.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 142, + 903, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/94a78afadb45ea89b6193d9630d2e8e3142520044ebc29a1b44226428d49ea0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 196, + 903, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/443007e41488d16f5bd391035ae13ebf1e11b4d24b1fc38643ed16ba14209dc8.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 828, + 250, + 903, + 304 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion 5.2] and compare the values for models trained with and without $\\mathcal{L}_{swd}$ in Tab. 5. The results show that including $\\mathcal{L}_{swd}$ decreases the uniformity loss, which aligns with our qualitative visualization in Fig. 5. We also conduct an ablation study on MLP layers and $\\mathcal{L}_{reg}$ in Tab. 4, showing that including all components yields the best outcome.", + "bbox": [ + 511, + 433, + 906, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative analysis. Fig. 6 shows the results of different training methods on the PASCAL VOC 2012 val set. When training with SWD loss, our method outperforms other methods, producing segmentation predictions that are closer to ground-truth labels. We can see from the results that our method performs exceptionally well on the border between different classes. Moreover, the segmentation results of our method have a cleaner prediction between the border and the backgrounds.", + "bbox": [ + 511, + 523, + 908, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced SWSEG, a semi-supervised semantic segmentation algorithm designed to enhance both alignment and uniformity of feature representations. We implement the Sliced-Wasserstein Distance (SWD) as an additional loss to jointly optimize these aspects, using a Gaussian-approximated variant to reduce computational overhead while preserving uniformity. Empirical evaluations on PASCAL VOC 2012, CityScapes, and ADE20K datasets show that SWSEG outperforms both supervised baselines and existing methods. Extensive ablation studies highlight the critical role of uniformity in the feature space for effective semi-supervised semantic segmentation. SWSEG is the first method to link semi-supervised learning with the uniformity metric using SWD, marking a significant advancement in the field.", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20240", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 89, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This material is based in part upon work supported by the National Science Foundation under Grant Numbers CNS-2333487 (NSF Frontier award) and CNS-2146449 (NSF CAREER award) and DEVCOM ARL Army Research Office under Contract number W911NF-2020-221. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors. The authors thank the reviewers for their enthusiastic comments and their valuable insights that improved our paper.", + "bbox": [ + 89, + 114, + 485, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 286, + 187, + 301 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. 2, 3", + "[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in Neural Information Processing Systems, 32, 2019. 1, 2", + "[3] Olivier Chapelle and Alexander Zien. Semi-supervised classification by low density separation. In International workshop on artificial intelligence and statistics, pages 57-64. PMLR, 2005. 1, 3", + "[4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 2, 4", + "[5] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 7", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 6", + "[7] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 6", + "[8] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 2, 6, 7", + "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 2, 6" + ], + "bbox": [ + 99, + 311, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] D.C. Dowson and B.V. Landau. The fréchet distance between multivariate normal distributions. Journal of multivariate analysis, 12(3):450-455, 1982. 5", + "[11] Lutz Dumbgen, Del Conte-Zerial, et al. On low-dimensional projections of high-dimensional distributions. In From Probability to Statistics and Back: High-Dimensional Models and Processes-A Festschrift in Honor of Jon A. Wellner, pages 91-105. Institute of Mathematical Statistics, 2013. 5, 6", + "[12] Mark Everingham and John Winn. The pascal visual object classes challenge 2012 (voc2012) development kit. Pattern Anal. Stat. Model. Comput. Learn., Tech. Rep, 2007:1-45, 2012. 2, 6", + "[13] Xianghong Fang, Jian Li, Qiang Sun, and Benyou Wang. Rethinking the uniformity metric in self-supervised learning. arXiv preprint arXiv:2403.00642, 2024. 2", + "[14] Geoff French, Samuli Laine, Timo Aila, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, varied perturbations. arXiv preprint arXiv:1906.01916, 2019. 1", + "[15] Karl Pearson F.R.S. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2 (11):559-572, 1901. 8", + "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284. Curran Associates, Inc., 2020. 6", + "[17] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991–998. IEEE, 2011. 6", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[19] Ying Jin, Jiaqi Wang, and Dahua Lin. Semi-supervised semantic segmentation via gentle teaching assistant. Advances in Neural Information Processing Systems, 35:2803-2816, 2022. 2, 6", + "[20] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 3", + "[21] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896, 2013. 3", + "[22] Shikun Liu, Shuaifeng Zhi, Edward Johns, and Andrew J Davison. Bootstrapping semantic segmentation with regional contrast. arXiv preprint arXiv:2104.04465, 2021. 6", + "[23] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022. 2, 4" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20241", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Huayu Mai, Rui Sun, Tianzhu Zhang, and Feng Wu. Rankmatch: Exploring the better consistency regularization for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3391-3401, 2024. 7", + "[25] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. Umap: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861, 2018. 8", + "[26] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 1", + "[27] Jaemin Na, Jung-Woo Ha, Hyung Jin Chang, Dongyoon Han, and Wonjun Hwang. Switching temporary teachers for semi-supervised semantic segmentation. Advances in Neural Information Processing Systems, 36, 2023. 2, 4, 6, 7", + "[28] Kimia Nadjahi, Alain Durmus, Pierre E Jacob, Roland Badeau, and Umut Simsekli. Fast approximation of the sliced-wasserstein distance using concentration of random projections. Advances in Neural Information Processing Systems, 34:12411-12424, 2021. 5", + "[29] Yassine Ouali, Celine Hudelot, and Myriam Tami. Semi-Supervised Semantic Segmentation With Cross-Consistency Training. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12671–12681, Seattle, WA, USA, 2020. IEEE. 2, 3", + "[30] Julien Rabin, Gabriel Peyré, Julie Delon, and Marc Bernot. Wasserstein barycenter and its application to texture mixing. In Scale Space and Variational Methods in Computer Vision: Third International Conference, SSVM 2011, Ein-Gedi, Israel, May 29-June 2, 2011, Revised Selected Papers 3, pages 435-446. Springer, 2012. 5", + "[31] Galen Reeves. Conditional central limit theorems for gaussian projections. In 2017 IEEE International Symposium on Information Theory (ISIT), pages 3045-3049. IEEE, 2017. 5, 6", + "[32] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 1", + "[33] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2", + "[34] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems, 33:596-608, 2020. 2, 4", + "[35] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence." + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Advances in Neural Information Processing Systems, 33:596-608, 2020. 3", + "[36] Boyuan Sun, Yuqi Yang, Le Zhang, Ming-Ming Cheng, and Qibin Hou. Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation. IEEE Computer Vision and Pattern Recognition (CVPR), 2024. 3, 7", + "[37] Rui Sun, Huayu Mai, Tianzhu Zhang, and Feng Wu. Daw: Exploring the better weighting function for semi-supervised semantic segmentation. Advances in Neural Information Processing Systems, 36, 2023. 3, 4, 6, 7", + "[38] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3", + "[39] Haonan Wang, Qixiang Zhang, Yi Li, and Xiaomeng Li. Allspark: Reborn labeled features from unlabeled in transformer for semi-supervised semantic segmentation. arXiv preprint arXiv:2403.01818, 2023. 3", + "[40] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 2, 3, 8", + "[41] Xiaoyang Wang, Huihui Bai, Limin Yu, Yao Zhao, and Jimin Xiao. Towards the uncharted: Density-descending feature perturbation for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3303-3312, 2024. 3, 7", + "[42] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudolabels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4248-4257, 2022. 3, 6", + "[43] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34:12077-12090, 2021. 2, 4, 6", + "[44] Haiming Xu, Lingqiao Liu, Qiuchen Bian, and Zhen Yang. Semi-supervised semantic segmentation with prototype-based consistency regularization. Advances in Neural Information Processing Systems, 35:26007-26020, 2022. 3, 7", + "[45] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4268-4277, 2022. 6", + "[46] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7236-7246, 2023. 2, 3, 4, 6, 7", + "[47] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 7" + ], + "bbox": [ + 516, + 90, + 908, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20242", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310–12320. PMLR, 2021. 2, 3", + "[49] Shaofeng Zhang, Feng Zhu, Junchi Yan, Rui Zhao, and Xiaokang Yang. Zero-cl: Instance and feature decorrelation for negative-free symmetric contrastive learning. In International Conference on Learning Representations, 2021. 6", + "[50] Zhen Zhao, Lihe Yang, Sifan Long, Jimin Pi, Luping Zhou, and Jingdong Wang. Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11350-11359, 2023. 2, 3, 4, 6, 7", + "[51] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 2, 6" + ], + "bbox": [ + 91, + 92, + 486, + 345 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20243", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/ca5cab4a-bdaf-48a0-9763-5efeeaa13237_model.json b/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/ca5cab4a-bdaf-48a0-9763-5efeeaa13237_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dd5053d2504e85abfa716c0cfa66facfd52a4bb8 --- /dev/null +++ b/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/ca5cab4a-bdaf-48a0-9763-5efeeaa13237_model.json @@ -0,0 +1,2389 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.2, + 0.131, + 0.8, + 0.177 + ], + "angle": 0, + "content": "Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.204, + 0.72, + 0.24 + ], + "angle": 0, + "content": "Chen-Yi Lu Kasra Derakhshandeh Somali Chaterji Purdue University" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.242, + 0.758, + 0.257 + ], + "angle": 0, + "content": "lu842@purdue.edu, kderakhs@purdue.edu, schaterji@purdue.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.486, + 0.705 + ], + "angle": 0, + "content": "Semi-supervised semantic segmentation with consistency regularization capitalizes on unlabeled images to enhance the accuracy of pixel-level segmentation. Current consistency learning methods primarily rely on the consistency loss between pseudo-labels and unlabeled images, neglecting the information within the feature representations of the backbone encoder. Preserving maximum information in feature embeddings requires achieving the alignment and uniformity objectives, as widely studied. To address this, we present SWSEG, a semi-supervised semantic segmentation algorithm that optimizes alignment and uniformity using the Sliced-Wasserstein Distance (SWD), and rigorously and empirically proves this connection. We further resolve the computational issues associated with conventional Monte Carlo-based SWD by implementing a Gaussian-approximated variant, which not only maintains the alignment and uniformity objectives but also improves training efficiency. We evaluate SWSEG on the PASCAL VOC 2012, Cityscapes, and ADE20K datasets, outshining supervised baselines in mIoU by up to \\(11.8\\%\\), \\(8.9\\%\\), and \\(8.2\\%\\), respectively, given an equivalent number of labeled samples. Further, SWSEG surpasses state-of-the-art methods in multiple settings across these three datasets. Our extensive ablation studies confirm the optimization of the uniformity and alignment objectives of the feature representations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.738, + 0.222, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Semantic segmentation, the task of assigning pixel-level labels to images, is a cornerstone in domains such as autonomous driving and medical image analysis. Semantic segmentation models typically comprise an encoder and a decoder. The encoder extracts meaningful features from the input images, while the decoder assigns the feature embeddings to desired pixel-level labels. However, its effectiveness is often hampered by the need for costly pixel-level annotated data [9]. To mitigate this, semi-supervised learning" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.294, + 0.903, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.48, + 0.907, + 0.508 + ], + "angle": 0, + "content": "Figure 1. SWSEG demonstrates marked superiority in mIoU across varying data splits of the ADE20K dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.908, + 0.61 + ], + "angle": 0, + "content": "(SSL) was proposed [3]. SSL leverages a limited amount of labeled data in conjunction with a larger pool of readily available unlabeled data. The key lies in effectively harnessing the unlabeled data to improve the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Pseudo-labeling and consistency regularization are two prominent techniques in SSL. Pseudo-labeling generates temporary labels from the model's predictions during training and using these to train the model in a supervised manner [2, 26]. The model then computes the loss between these pseudo-labels and the corresponding unlabeled images. This technique allows the model to leverage unlabeled data, thereby improving its performance. Consistency regularization, on the other hand, encourages the model to produce consistent predictions for different perturbations of the same input image, thereby promoting robustness and generalization [14, 32]. This technique is grounded in the smoothness assumption, which posits that data points with similar features should have similar labels [3]. In practice, random augmentations are applied to input images, and the model is trained to minimize discrepancies between its predictions on the original and augmented versions. In the realm of semantic segmentation, consistency regularization is often paired with pseudo-labeling, where weakly augmented unlabeled" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20233" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.191, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.158, + 0.165, + 0.17 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.203, + 0.089, + 0.292, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.168, + 0.293, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.245, + 0.256, + 0.257 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.089, + 0.385, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.173, + 0.385, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.243, + 0.351, + 0.256 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.089, + 0.48, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.174, + 0.482, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.244, + 0.446, + 0.256 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.27, + 0.485, + 0.368 + ], + "angle": 0, + "content": "Figure 2. Grad-CAM visualization [33] and model predictions. (a) depicts results from a model trained without alignment optimization, leading to inadequate boundary definition. (b) shows moderate improvements using traditional cross-entropy loss, which partially improves boundary adherence. (c) illustrates SWSEG with SWD loss, significantly enhancing segmentation accuracy by ensuring uniformity and alignment of features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.463 + ], + "angle": 0, + "content": "data generate temporary \"pseudo\" labels for training [2, 34]. Concurrently, a parallel training stream employs strongly augmented data to calculate and minimize consistency losses, further enhancing model robustness [8, 23, 29, 46, 50]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.688 + ], + "angle": 0, + "content": "While consistency-based methods are effective, existing approaches focus solely on enforcing regularization at the decoder's output, neglecting the rich information embedded within the intermediate feature representations from the backbone encoder. This oversight neglects the potential of extending consistency regularization to the feature level, guided by the smoothness assumption. We argue that enforcing consistency within these feature representations, derived from CNN-based or Transformer-based backbone encoders [4, 43], is crucial for robust learning against image perturbations. In Fig. 2, we insert a cross entropy loss on the output of encoders between feature embeddings of weakly and strongly augmented images. The results show that it already improves the feature separability and segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.901 + ], + "angle": 0, + "content": "To explore the importance of superior feature representations, we draw parallels with the self-supervised contrastive learning literature [1, 6, 48]. These methods, which involve contrasting positive (similar) and negative (dissimilar) pairs, excel at promoting both alignment and uniformity [13, 40]. Alignment, which aligns with the smoothness assumption, posits that perturbations to the same input should yield proximate features. Uniformity, on the other hand, implies that feature representations should be maximally informative, a concept often overlooked in semi-supervised semantic segmentation. These metrics provide a comprehensive framework for evaluating and enhancing feature representations, ultimately leading to improved model performance. The attraction between positive pairs in contrastive learning fosters" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.228 + ], + "angle": 0, + "content": "alignment, while the repulsion of negative pairs encourages diverse and informative representations, enhancing uniformity. Inspired by the effectiveness of contrastive learning, we aim to bridge the gap in semi-supervised semantic segmentation by developing a method that explicitly optimizes both alignment and uniformity of feature representations. This raises a key question: How can we design an objective function that simultaneously achieves both uniformity and alignment of feature representations?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.909, + 0.835 + ], + "angle": 0, + "content": "In this work, we introduce SWSEG, a semi-supervised semantic segmentation algorithm that directly addresses this question. We leverage the Sliced-Wasserstein distance (SWD), a metric that quantifies the distance between probability distributions, to simultaneously optimize alignment and uniformity of feature representations extracted from the backbone encoder model. SWD achieves this by projecting features onto a unit hypersphere and computing the Wasserstein distance between the resulting one-dimensional embeddings. Empirically, we demonstrate that minimizing SWD effectively optimizes both alignment and uniformity between feature representations. Furthermore, to address the computational burden associated with traditional Monte Carlo estimation of SWD for high-dimensional features, we introduce a variant that projects feature embeddings onto a Gaussian distribution. This not only maintains the uniformity objective, as a normalized Gaussian distribution is uniformly distributed on the unit hypersphere, but also enables efficient computation by leveraging the analytical solution for the quadratic Wasserstein distance between two Gaussian distributions. Additionally, we incorporate a regularization term that decorrelates feature representations, fulfilling the necessary condition for projecting features onto a Gaussian distribution. We evaluated our algorithm on the PASCAL VOC 2012 [12], Cityscapes [9], and ADE20K [51] datasets under various partitioning protocols. Our results consistently demonstrate SWSEG's superiority over supervised baselines, with performance improvements of up to \\(11.8\\%\\), \\(8.9\\%\\), and \\(8.2\\%\\) on PASCAL VOC, CityScapes, and ADE20K, using the same amount of labeled data. Notably, SWSEG also achieves state-of-the-art (SOTA) results compared to existing methods under multiple data splitting protocols, as illustrated in Figure 1 [19, 23, 27]. Through extensive ablation studies, we confirm the critical role of uniformity in feature space for effective semi-supervised semantic segmentation. To the best of our knowledge, SWSEG is the first to establish a direct link between SSL and the uniformity metric, while also implementing the SWD to enhance feature representations. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "1. We introduce SWSEG, a semi-supervised semantic segmentation algorithm that explicitly optimizes feature representation uniformity and alignment using SWD, and back up our claim with extensive empirical studies." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20234" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.088, + 0.092, + 0.483, + 0.167 + ], + "angle": 0, + "content": "2. We introduce an efficient variant of SWD estimation that projects feature embeddings onto a Gaussian distribution. This approach maintains uniformity while leveraging the analytical solution for quadratic Wasserstein distance between Gaussians, reducing computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.168, + 0.483, + 0.258 + ], + "angle": 0, + "content": "3. SWSEG achieves SOTA results on PASCAL VOC 2012, Cityscapes, and ADE20K datasets, outperforming supervised baselines and existing semi-supervised methods, with significant performance improvements of up to \\(11.8\\%\\) on PASCAL VOC, \\(8.9\\%\\) on CityScapes, and \\(8.2\\%\\) on ADE20K compared to supervised methods." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.483, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.274, + 0.241, + 0.29 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.3, + 0.332, + 0.316 + ], + "angle": 0, + "content": "2.1. Semi-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.323, + 0.485, + 0.699 + ], + "angle": 0, + "content": "Semi-supervised learning (SSL) has emerged as a promising solution to alleviate the burden of collecting labeled data, enabling the training of DNNs using only a fraction of labeled data while still achieving impressive results. Pseudo-labeling is a widely adopted technique which leverages the model's predictions on unlabeled data as surrogate labels [21]. This approach has been shown to improve model generalization by shifting the decision boundary towards low-density regions [3]. Additionally, many SSL algorithms incorporate consistency regularization, which regularizes the model to generate consistent outputs regardless of perturbations applied to the input data [20, 35, 38]. Laine et al. [20] combined cross-entropy loss with consistency loss for training, and incorporated model predictions into the pseudo-label set through an exponential moving average. Tarvainen and Valpola [38] posited that directly updating the model weights at each mini-batch accelerates the learning process while enhancing the model's performance. Sohn et al. [35] generated pseudo-labels from model predictions on weakly-augmented input images and computed the unsupervised loss between these pseudo-labels and the model predictions from strongly-augmented input images. However, these SSL methods were designed for image classification, which is not suitable for semantic segmentation where pixel-wise predictions are desired." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.441, + 0.729 + ], + "angle": 0, + "content": "2.2. Semi-Supervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "A number of research papers have incorporated consistency training and pseudo-labeling in SSL for segmentation tasks [29, 37, 39, 42, 44-46, 50]. For instance, Wang et al. [42] stored pseudo-labels with lower confidence as negative samples and computed contrastive loss using these samples. Yang et al. [46] introduced an additional teacher model and used dropout as feature-wise perturbation for consistency-based pseudo-labeling. However, these methods typically enforce consistency regularization solely on the decoder output, neglecting the rich information present in the intermediate feature representations produced by the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.395 + ], + "angle": 0, + "content": "backbone encoder. Recent works leveraging feature embedding in SSL segmentation include AllSpark [39], DDFP [41], and CorrMatch [36]. AllSpark balances the importance of labeled and unlabeled training streams by implementing a cross-attention mechanism in the feature embedding. DDFP injects perturbation in the feature space to regularize the decision boundary, with perturbation locations guided by a lightweight normalizing flow-based density estimator. CorrMatch utilizes the correlation map of features from strongly and weakly augmented inputs to regularize pseudo-label supervision in the decoder output. In contrast, SWSEG aims to fill this gap by explicitly focusing on training the backbone encoder to produce more meaningful feature representations for the segmentation head. By optimizing both the alignment and uniformity of these feature representations, SWSEG aims to improve overall segmentation performance. This approach goes beyond the limitations of existing methods that primarily focus on designing complex strategies such as cross-attention layers, normalizing flows, or heuristic label and region propagation techniques." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.403, + 0.839, + 0.419 + ], + "angle": 0, + "content": "2.3. Uniformity of Feature Representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.909, + 0.802 + ], + "angle": 0, + "content": "Although existing SSL methods for semantic segmentation have demonstrated promising results through consistency learning, they typically impose consistency on the model's output. While few studies have explored consistency regularization on feature representations from the backbone model, this area remains under-explored compared to extensive research in self-supervised learning [1, 6, 48]. In self-supervised learning, where labeled data is scarce or absent, preventing representational collapse is paramount. This collapse occurs when the model produces uninformative feature representations, regardless of variations in the input. While the uniformity metric, which assesses how evenly features are distributed across the hypersphere, has been effectively used in self-supervised learning to mitigate collapse [40], its application in semi-supervised semantic segmentation has been absent. Recognizing the potential benefits of uniformity in preventing representational collapse and maximizing information preservation, we incorporate this metric into our semi-supervised approach. We further establish a connection between uniformity and the sliced-SWD metric, demonstrating that minimizing SWD between feature representations of differently augmented inputs can simultaneously maximize feature uniformity. This unified approach enhances model performance by ensuring that feature representations are both consistent and informative." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.605, + 0.831 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "This section begins by introducing the conventional framework for consistency learning-based SSL. We then argue for the importance of enforcing consistency regularization in the feature embedding, discussing the roles of alignment and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20235" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.912, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.348, + 0.908, + 0.446 + ], + "angle": 0, + "content": "Figure 3. SWSEG Framework for Joint Optimization of Alignment and Uniformity. The SWSEG framework employs both weak \\((\\mathcal{A}^w)\\) and strong perturbations \\((\\mathcal{A}^s)\\) applied to unlabeled input images \\(x^u\\). Both augmentations are passed through a shared backbone encoder to extract features, which are then mapped to a different space via non-linear MLPs, producing feature embeddings. To achieve both feature alignment and uniformity, SWSEG introduces an SWD loss \\(\\mathcal{L}_{swd}\\), computed between these intermediate feature embeddings. The goal is to encourage the feature embeddings from different augmentations to be consistent (alignment) and to be spread out uniformly in the feature space, maximizing information (uniformity). An unsupervised loss \\(\\mathcal{L}_u\\) is computed to ensure consistency between segmentations generated from weakly and strongly augmented inputs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.486, + 0.554 + ], + "angle": 0, + "content": "uniformity in achieving better feature representations. Next, we introduce a Gaussian-projected SWD metric to simultaneously achieve uniformity and alignment in the feature space. Finally, we detail the integration of these elements within the SWSEG framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.563, + 0.221, + 0.579 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.663 + ], + "angle": 0, + "content": "In semi-supervised learning, we are given labeled image sets, \\( D^{l} = \\{x_{i}^{l},y_{i}^{l}\\} \\), and unlabeled ones, \\( D^{u} = \\{x_{j}^{u}\\} \\). Typically, \\( D^{u} \\) greatly outnumbers \\( D^{l} \\). Our goal is to leverage these unlabeled images to optimize the parameters \\( \\theta \\) of a semantic segmentation model \\( f_{\\theta}(\\cdot) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.484, + 0.814 + ], + "angle": 0, + "content": "In SWSEG, we adopt a typical framework from Fix-Match [34], training the model through two separate feedforward streams: supervised and unsupervised. In the supervised stream, we draw a batch of images from \\( D^{l} \\) and compute the cross-entropy loss \\( \\mathcal{L}_s \\) between the model's predictions \\( \\hat{y}_i^l \\) and the labels \\( y_i^l \\). Conversely, the unsupervised stream draws images from \\( D^{u} \\), to which we apply weak or strong perturbations, \\( \\mathcal{A}^w(\\cdot) \\) and \\( \\mathcal{A}^s(\\cdot) \\), respectively. The unsupervised loss \\( \\mathcal{L}_u \\) is then minimized according to the equation (1):" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.824, + 0.484, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u} = \\frac {1}{| \\mathcal {D} ^ {u} |} \\sum_ {x \\in \\mathcal {D} ^ {u}} \\mathcal {H} \\left(f _ {\\theta} \\left(\\mathcal {A} ^ {s} \\left(x ^ {u}\\right)\\right), f _ {\\theta} \\left(\\mathcal {A} ^ {w} \\left(x ^ {u}\\right)\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.87, + 0.485, + 0.902 + ], + "angle": 0, + "content": "We generate pseudo-labels \\( f_{\\theta}(\\mathcal{A}^{w}(x^{u})) \\) by feeding weakly-augmented images into the model. We then compute" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.908, + 0.525 + ], + "angle": 0, + "content": "the unsupervised loss using strongly-augmented samples \\( f_{\\theta}(\\mathcal{A}^{s}(x^{u})) \\), using \\( \\mathcal{H} \\), for cross-entropy loss. The pseudolabel generation does not contribute to gradient computation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.534, + 0.818, + 0.551 + ], + "angle": 0, + "content": "3.2. Feature Alignment and Uniformity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.908, + 0.707 + ], + "angle": 0, + "content": "In semantic segmentation, \\( f_{\\theta}(\\cdot) \\) typically comprises a backbone encoder \\( p_{\\theta} \\) and a segmentation head \\( q_{\\theta} \\) [4, 43]. Existing methods [23, 27, 37, 46, 50] focus on enforcing consistency regularization on the segmentation head's output, often overlooking backbone encoder features. To further harness the potential of feature representations, we draw on self-supervised contrastive learning, which trains networks to generate meaningful embeddings without labels, emphasizing alignment and uniformity as key objectives for assessing feature quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.907, + 0.812 + ], + "angle": 0, + "content": "Alignment aims to minimize differences between embeddings from similar samples. Intuitively, model should produce similar feature embeddings when images with same classes are input. We achieve this by varying the magnitude of perturbations applied to the input images and minimizing the distance between embeddings using a distance metric dist:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.826, + 0.907, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = \\operatorname {d i s t} \\left(p _ {\\theta} \\left(\\mathcal {A} ^ {s} \\left(x _ {i} ^ {u}\\right)\\right), p _ {\\theta} \\left(\\mathcal {A} ^ {w} \\left(x _ {i} ^ {u}\\right)\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Uniformity promotes even distributions of feature representations across a hypersphere, enhancing model robustness by preventing collapse to a lower-dimensional subspace." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20236" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "While often less emphasized in semi-supervised learning due to the stabilizing influence of labeled data, uniformity is critical in self-supervised learning to prevent the model from generating identical embeddings regardless of input. By ensuring uniformity, the model preserves a rich diversity of information within the feature space, which significantly improves the performance of the segmentation head." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.207, + 0.421, + 0.223 + ], + "angle": 0, + "content": "3.3. Revisiting Sliced Wasserstein Distance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.229, + 0.484, + 0.335 + ], + "angle": 0, + "content": "Intuitively, the uniform distribution on a unit hypersphere exemplifies maximum uniformity. We need to establish an objective function that can achieve both uniformity and alignment. To this end, we implement the SWD as our distance metric. SWD calculates average distances from 1-D projections sampled from a unit hypersphere, thus inherently promoting both desired properties." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.442 + ], + "angle": 0, + "content": "The rationale behind SWD stems from its ability to reduce the computational complexity relative to the high-dimensional Wasserstein distances (WD). Unlike the traditional WD, which struggles with high dimensionality, SWD leverages the 1-dimensional Wasserstein distance, which can be computed in closed form: Let \\(\\mu_d\\) and \\(\\nu_d\\) be two probability measures in \\(\\mathcal{P}(\\Omega) \\in \\mathbb{R}^d\\):" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.453, + 0.484, + 0.492 + ], + "angle": 0, + "content": "\\[\nW _ {p} ^ {p} \\left(\\mu_ {d}, \\nu_ {d}\\right) = \\left(\\int_ {0} ^ {1} \\left| F _ {\\mu_ {d}} ^ {- 1} (z) - F _ {\\nu_ {d}} ^ {- 1} (z) \\right| ^ {p} d z\\right) ^ {1 / p}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.653 + ], + "angle": 0, + "content": "where \\(p\\) is the order of the distance, and \\(F_{\\mu_d}^{-1}\\) and \\(F_{\\nu_d}^{-1}\\) are the quantile functions of \\(\\mu_d\\) and \\(\\nu_d\\), respectively. Computationally, this is achieved by sorting and computing the mean \\(l^2\\)-distance between two 1-D vectors. SWD [30] involves projecting high-dimensional features onto a unit hypersphere and calculating the 1-D Wasserstein distance between these projections. Formally, let \\(\\mathbb{S}^{d-1}\\) represent the unit sphere in \\(d\\) dimensions, \\(\\sigma\\) the uniform distribution on \\(\\mathbb{S}^{d-1}\\), and \\(P_{\\theta\\#}\\xi\\) the push-forward measure of \\(\\xi \\in \\mu_d, \\nu_d\\). The SWD is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.662, + 0.484, + 0.695 + ], + "angle": 0, + "content": "\\[\nS W _ {p} ^ {p} \\left(\\mu_ {d}, \\nu_ {d}\\right) = \\int_ {\\mathbb {S} ^ {d - 1}} W _ {p} ^ {p} \\left(P _ {\\theta \\#} \\mu_ {d}, P _ {\\theta \\#} \\nu_ {d}\\right) \\mathrm {d} \\sigma (\\theta). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.856 + ], + "angle": 0, + "content": "To approximate SWD, we vectorize the input distributions \\(\\mu_{d}\\) and \\(\\nu_{d}\\) from \\(\\mathbb{R}^{n\\times n\\times d}\\) to \\(\\mathbb{R}^{n^2 d}\\). We then employ a Monte Carlo scheme, sampling \\(L\\) directions from the unit sphere \\(\\mathbb{S}^{d - 1}\\). Projecting \\(\\mu_{d}\\) and \\(\\nu_{d}\\) along these directions yields \\(L\\) one-dimensional projections for each distribution. After sorting these projections, we calculate the 1D Wasserstein distance between them using Eq. (3). This method efficiently estimates the Wasserstein distance between feature embeddings, exploiting the closed form property of 1D calculations compared to high-dimensional alternatives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Although the unit sphere projection enables uniformity and minimizing 1-D Wasserstein distances facilitates alignment, training with Monte Carlo-approximated SWD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.213 + ], + "angle": 0, + "content": "presents several challenges. First, the results are inherently non-deterministic and intractable due to the randomness inherent in the Monte Carlo method. Second, as the dimensions of \\(\\mu_d\\) and \\(\\nu_d\\) increase, both the computational cost and memory requirements for SWD escalate. Finally, to maintain accuracy in the face of increasing dimensions, the number of directions \\(L\\) must also increase, complicating the Monte Carlo approximation further." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.221, + 0.771, + 0.236 + ], + "angle": 0, + "content": "3.4. Proposed Approach: SWSEG" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.242, + 0.907, + 0.499 + ], + "angle": 0, + "content": "To enhance feature representation uniformity and alignment while alleviating the computational burden of Monte Carlo-based SWD, we implement an alternate method proposed by Nadjahi et al. [28] for approximating the SWD. Our method diverges by projecting these measures directly onto a Gaussian distribution. This modification offers significant computational benefits and supports key theoretical objectives in two primary ways: First, projecting onto Gaussian distributions allows the 2-Wasserstein distance between any two Gaussians to be expressed in a closed-form equation. Second, using normalized Gaussian variables ensures that the resulting distributions are uniformly distributed over the unit hypersphere, which we prove in the supplementary materials. This uniform distribution is essential for maintaining the model's uniformity objective, ensuring features are evenly distributed in the embedding space, maximizing its information for segmentation head." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.499, + 0.909, + 0.727 + ], + "angle": 0, + "content": "To formalize this idea, for any positive integer \\(d\\), let \\(\\{\\theta_i\\}_{i=1}^d\\) be a sequence of i.i.d. 1-d standard Gaussian distributions and \\(\\{X_i\\}_{i=1}^d\\) be a sequence of 1-d probability distributions. Several Central Limit Theorems [11, 31] ensure that the sequence of distributions formed by the inner product \\(d^{-1/2}\\langle \\theta_{1:d}, X_{1:d} \\rangle\\) converges in probability to a Gaussian random variable when the distributions exhibit weak correlation. Furthermore, according to Proposition 2 from [28], as the dimensions of the distributions increase, the quadratic Wasserstein distance between two zero-mean Gaussian distributions with variance \\(d^{-1}m_2(\\bar{\\mu}_d)\\) can be approximated by the SW distance between two Gaussian projected distributions. The Wasserstein distance between two Gaussian probability distribution is in closed form [10], which enables us to approximate the \\(l^2\\)-SW distance by:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.732, + 0.907, + 0.766 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s w d} = d ^ {- 1} \\left(m _ {2} \\left(\\bar {\\mu_ {d}}\\right) ^ {\\frac {1}{2}} - m _ {2} \\left(\\bar {\\nu_ {d}}\\right) ^ {\\frac {1}{2}}\\right) + d ^ {- 1} \\left| m _ {\\mu_ {d}} - m _ {\\nu_ {d}} \\right| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "This equation comprises two terms: the first quantifies the difference in the second moments of the distributions, and the second captures the squared error in their means. For \\(\\xi_d \\in \\{\\bar{\\mu}_d, \\bar{\\nu}_d\\}\\), we define \\(m_2(\\xi_d) = n^{-1} \\sum_{j=1}^n ||x^j||^2\\), enabling a closed-form deterministic solution for \\(\\mathcal{L}_{swd}\\). Nadjahi et al. also proved in Proposition 1 that \\(\\mathcal{L}_{swd}\\) with Gaussian projections equals the original \\(SW_p^p\\) when \\(p = 2\\). Thus, we can use \\(\\mathcal{L}_{swd}\\) as a computationally efficient Wasserstein metric for our consistency loss term. A PyTorch-styled" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20237" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.095, + 0.908, + 0.123 + ], + "angle": 0, + "content": "Table 1. Results on ADE20k. The fractions represent the labeled image proportions sampled during training. The model is trained on SegFormer [43] with MiT-B1 backbone. We use mIOU as our evaluation metric." + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.128, + 0.835, + 0.273 + ], + "angle": 0, + "content": "
Method1/32 (631)1/16 (1263)1/8 (2526)1/4 (5052)1/2 (10105)
Supervised Only15.222.325.929.532.6
CPS [8] [ICLR '21]18.822.327.932.436.9
U²PL [42] [CVPR '22]18.523.628.133.536.0
ReCo [22] [ICLR '22]21.123.727.730.235.5
GTA-Seg [19] [NeurIPS '22]19.523.126.927.829.9
DT [27] [NeurIPS '23]22.626.030.633.737.3
SWSEG (Ours)24.426.730.934.638.0
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.297, + 0.484, + 0.402 + ], + "angle": 0, + "content": "pseudo code is presented in Algorithm 1. Furthermore, to ensure a more accurate approximation of the Gaussian-SWD and comply with the relaxed Central Limit Theorems in [11, 31], minimizing strong correlations among input embeddings is essential. We address this by introducing a regularization term based on the input cross-covariance matrix into the SWD loss:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.411, + 0.484, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e g} = | \\operatorname {C o v} (X) - \\operatorname {D i a g} (X) | _ {2} ^ {2} + | \\operatorname {C o v} (Y) - \\operatorname {D i a g} (Y) | _ {2} ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.485, + 0.518 + ], + "angle": 0, + "content": "Here, \\(\\mathrm{Cov}(\\cdot)\\) calculates the covariance matrix of the input features, and \\(\\mathrm{Diag}(\\cdot)\\) extracts its diagonal elements. Minimizing the off-diagonal elements decorrelates the features, thereby indirectly promoting the uniformity of the distribution of the features across the representation space [49]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.484, + 0.759 + ], + "angle": 0, + "content": "We present a detailed overview of this training scheme in Fig. 3. We also insert an additional non-linear MLP projection layer before loss computation between representations, following recent self-supervised learning work [6, 7, 16]. The projection layer is solely incorporated during the training process, and it is subsequently discarded when the model is deployed for inference. Our objective is to retain more information in the features prior to the MLP layer. This will enable the segmentation head to utilize the preserved information more effectively for the segmentation task. The specific architecture of the projection layer used in this work consists of two Conv - BatchNorm - ReLu blocks followed by a single Conv layer. The overall training objective, managing the balance of various losses, is summarized in Eq. (7). The hyperparameters \\(\\lambda_1, \\lambda_2\\), and \\(\\lambda_3\\) control the weight of each loss." + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.786, + 0.484, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o v e r a l l}} = \\mathcal {L} _ {s} + \\lambda_ {1} \\mathcal {L} _ {u} + \\lambda_ {2} \\mathcal {L} _ {s w d} + \\lambda_ {3} \\mathcal {L} _ {\\text {r e g}}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.208, + 0.839 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.194, + 0.863 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "ADE20K. [51] is a complex scene parsing dataset featuring 150 objects, with 25,574 training, 2,000 validation, and 3,000" + }, + { + "type": "code_caption", + "bbox": [ + 0.513, + 0.296, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Algorithm 1 Pseudocode of our Gaussian-SWD loss in a PyTorch-like style." + }, + { + "type": "code", + "bbox": [ + 0.516, + 0.334, + 0.905, + 0.548 + ], + "angle": 0, + "content": "def gaussian_swd_loss(x, y):\n # Reshape the input tensors to (n, dim)\n x, y = x.view(x.shape[0], -1), y.view(y.shape[0], -1)\n n, dim = x.shape\n meanx = torch.mean(x, dim=0)\n xc = x - meanx # Zero mean\n m2x = torch.mean(torch.linalg(norm(xc, dim=1)) * 2) / dim\n meany = torch.mean(y, dim=0)\n yc = y - meany # Zero mean\n m2y = torch.mean(torch.linalg(norm(yc, dim=1)) * 2) / dim\n m_t = torch.linalg(norm(meanx - meany) * 2 / dim\n swd = m_t + (m2x * (1/2) - m2y * (1/2)) * 2\n return swd" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.907, + 0.607 + ], + "angle": 0, + "content": "testing images. For a fair comparison, we used the data split from dual-teacher [27]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.906, + 0.699 + ], + "angle": 0, + "content": "PASCAL VOC 2012. [12] This dataset is a classic semantic segmentation dataset comprising 20 classes of objects and 1 background class, with 1464 training, 1449 validation, and 1546 testing high-quality images in the standard set. We incorporate the augmented set from [17], following previous works, increasing our labeled data to 10,582 images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.908, + 0.761 + ], + "angle": 0, + "content": "Cityscapes. [9] This is an autonomous driving semantic segmentation dataset that consists of 19 semantic classes, with 2975, 500, and 1525 training, validation, and testing images, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.733, + 0.789 + ], + "angle": 0, + "content": "4.2. Implementation details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.901 + ], + "angle": 0, + "content": "For ADE20K, we implemented our method with SegFormer [43] architecture with MiT-B1 backbone. For PASCAL VOC 2012 and CityScapes, to have a fair comparison with previous benchmark semi-supervised semantic segmentation methods [27, 37, 42, 45, 46, 50], we implement our method using ResNet-101 [18] as the backbone encoder, initialized with ImageNet pre-trained weights, and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20238" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.095, + 0.486, + 0.138 + ], + "angle": 0, + "content": "Table 2. Results on CityScapes dataset. The fractions represent the proportion of labeled images used during training. We use mIOU as our evaluation metric." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.142, + 0.47, + 0.332 + ], + "angle": 0, + "content": "
Method1/161/81/4
Supervised Only67.973.275.1
PCR [44] [NeurIPS '22]73.476.378.4
UniMatch [46] [CVPR '23]76.677.979.2
AugSeg [50] [CVPR '23]75.277.879.6
DT [27] [NeurIPS '23]76.878.479.5
DAW [37] [NeurIPS '23]76.678.479.8
CorrMatch [36] [CVPR '24]77.378.579.4
RankMatch [24] [CVPR '24]77.178.680.0
DDFP [41] [CVPR '24]77.178.279.9
SWSEG (Ours)76.878.679.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.34, + 0.485, + 0.382 + ], + "angle": 0, + "content": "Table 3. Results on PASCAL VOC 2012 dataset. The fractions represent the proportion of labeled images used during training. We use mIOU as our evaluation metric." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.386, + 0.47, + 0.548 + ], + "angle": 0, + "content": "
Method1/161/81/4
Supervised Only67.271.373.5
UniMatch [46] [CVPR '23]75.076.877.5
AugSeg [50] [CVPR '23]77.077.378.8
DAW [37] [NeurIPS '23]78.578.979.6
CorrMatch [36] [CVPR '24]78.479.379.6
RankMatch [24] [CVPR '24]78.979.280.0
DDFP [41] [CVPR '24]78.378.979.8
SWSEG (Ours)79.079.379.9
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.486, + 0.669 + ], + "angle": 0, + "content": "DeeplabV3+ [5] as the decoder. We use SGD optimizer with momentum of 0.9. The learning rate is set to 0.001, 0.02, and 5e-5 for PASCAL, CityScapes, and ADE20K, respectively. We utilize a polynomial scheduling strategy with 0.9 decay rate." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.678, + 0.486, + 0.816 + ], + "angle": 0, + "content": "Data augmentations. For weak augmentations, horizontal flipping, random cropping, and random scaling are applied to the training images. For strong augmentations, we apply random color jittering, random grayscale, random blurring, and Cutmix [47] to the input training images. The training image size is set as \\(513 \\times 513\\), \\(769 \\times 769\\), and \\(512 \\times 512\\) for PASCAL VOC, CityScapes, and ADE20K, respectively. The training epochs for VOC, CityScapes, and ADE20K is 80, 240, and 200 epochs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Evaluation. As in previous works, we employ mean Intersection-over-Union (mIOU) as our evaluation metric. For the PASCAL VOC 2012 and ADE20K, evaluations are on the center-cropped validation set, while for CityScapes, we employ a sliding window evaluation." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.095, + 0.907, + 0.166 + ], + "angle": 0, + "content": "Table 4. Ablation analysis of feature alignment loss functions, \\(\\mathcal{L}_{reg}\\), and MLP layer. The results reveal that MSE, cross entropy, and Monte Carlo estimated SWD loss with \\(L = 128\\) yield inferior results. The combination of Gaussian-SWD (GSWD) with \\(\\mathcal{L}_{reg}\\) and MLP produces the best performance, with an mIoU of 79.0." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.17, + 0.905, + 0.313 + ], + "angle": 0, + "content": "
Loss FunctionsLregMLPmIoU
MSECESWDGSWD
✓✓✓70.2
✓✓✓77.1
✓✓✓77.6
✓77.4
✓✓78.1
✓✓✓79.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.322, + 0.909, + 0.378 + ], + "angle": 0, + "content": "Table 5. Analysis of the Uniformity value \\(\\mathcal{L}_{\\mathrm{uniform}}\\) across different data splits of the PASCAL VOC dataset demonstrates that training with the SWD loss \\(\\mathcal{L}_{swd}\\) yields superior Uniformity values compared to training without this loss term." + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.382, + 0.875, + 0.446 + ], + "angle": 0, + "content": "
Method1/81/21
Without Lswd-2.5896-2.7095-2.5574
With Lswd-2.6592-2.7104-2.6168
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.47, + 0.819, + 0.487 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.909, + 0.612 + ], + "angle": 0, + "content": "Results on ADE20K Dataset. Tab. 1 demonstrates the performance of SWSEG against the supervised baseline and existing methods. SWSEG significantly improves over the supervised baseline by \\(8.2\\%\\), \\(3.4\\%\\), \\(3.7\\%\\), \\(5.1\\%\\), and \\(5.4\\%\\) mIoUs on the 1/32, 1/16, 1/8, 1/4, and 1/2 splits, respectively. Our method also outperforms the state-of-the-art across all data splits, showcasing the generalizability of SWSEG across different model architectures and datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.72 + ], + "angle": 0, + "content": "Results on Cityscapes Dataset. Tab. 2 presents our results in comparison to the supervised baseline and SOTA methods. SWSEG outperforms the supervised baseline by \\(8.9\\%\\), \\(5.3\\%\\), and \\(4.4\\%\\) for the 1/16, 1/8, and 1/4 partitions, respectively. Additionally, SWSEG surpasses the previous state-of-the-art method, DAW, by \\(0.2\\%\\) in the 1/16 and 1/8 partitions, and performs on par with contemporary methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Results on PASCAL VOC 2012 Dataset. Tab. 3 compares SWSEG with existing state-of-the-arts on PASCAL VOC 2012 set. We follow the data partitions in CPS [8]. We outperform the supervised baselines by \\(11.8\\%\\), \\(8.0\\%\\) and \\(6.4\\%\\) mIoUs under 1/16, 1/8, and 1/4 partitioning protocols. Our method also beats the SOTA DDFP [37] by \\(0.7\\%\\), \\(0.4\\%\\), and \\(0.1\\%\\) mIoUs under 1/16, 1/8, and 1/4 splits, respectively. We also visualized the predicted features in Fig. 4 to evaluate the effectiveness of our algorithm's clustering capabilities. Our method demonstrates enhanced discriminative clustering compared to the Supervised Only and UniMatch methods." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20239" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.091, + 0.216, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.165, + 0.178, + 0.175 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.091, + 0.335, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.165, + 0.296, + 0.176 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.091, + 0.453, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.165, + 0.413, + 0.176 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.185, + 0.484, + 0.255 + ], + "angle": 0, + "content": "Figure 4. UMAP visualization [25] of feature embeddings using models trained with: (a) Supervised Only, (b) UniMatch, and (c) SWSEG. These visualizations utilize the PASCAL VOC 2012 dataset. Our method demonstrates enhanced clustering capabilities, outperforming both the Supervised Only and UniMatch." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.264, + 0.285, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.465, + 0.237, + 0.477 + ], + "angle": 0, + "content": "(a) Without \\(\\mathcal{L}_{swd}\\)" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.264, + 0.482, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.465, + 0.426, + 0.477 + ], + "angle": 0, + "content": "(b) With \\(\\mathcal{L}_{swd}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.488, + 0.484, + 0.545 + ], + "angle": 0, + "content": "Figure 5. We visualize feature representations on a unit circle for models trained with and without \\(\\mathcal{L}_{swd}\\) for uniformity analysis. Features from the model with \\(\\mathcal{L}_{swd}\\) (b) are more uniformly distributed than those without \\(\\mathcal{L}_{swd}\\) (a)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.245, + 0.593 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Uniformity and Alignment. We conduct an empirical analysis of feature representation alignment and uniformity (Tab. 4). We compare different loss functions for enforcing consistency regularization on the feature representation. Our results indicate that, aside from MSE loss, additional loss terms generally enhance segmentation prediction performance, highlighting the critical role of feature alignment. Monte Carlo SWD with \\( L = 128 \\) outperforms MSE and cross-entropy but encounters out-of-memory issues at \\( L = 1000 \\) using an identical number of GPUs, highlighting its computational cost. Conversely, our Gaussian-based SWD approximation delivers optimal results with lower memory usage. For uniformity analysis, we reduce the dimension of feature embeddings using PCA [15], projecting them into \\( \\mathbb{R}^2 \\) and plotting their coordinates on a unit circle, as shown in Fig. 5. The visualization demonstrates that models trained with \\( \\mathcal{L}_{swd} \\) achieve greater uniformity compared to those without it, confirming the effectiveness of our introduced SWD loss. To quantitatively validate this observation, we compute the uniformity metric \\( \\mathcal{L}_{\\text{uniform}} \\) from [40, Equa" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.593, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.089, + 0.67, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.143, + 0.67, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.197, + 0.67, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.305, + 0.639, + 0.311 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.089, + 0.748, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.143, + 0.748, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.197, + 0.748, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.251, + 0.748, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.305, + 0.718, + 0.311 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.089, + 0.826, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.143, + 0.826, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.197, + 0.826, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.251, + 0.826, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.306, + 0.796, + 0.311 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.089, + 0.905, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.143, + 0.904, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.197, + 0.905, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.251, + 0.905, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.862, + 0.306, + 0.872, + 0.311 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.318, + 0.907, + 0.404 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on PASCAL VOC 2012 validation set. All models are trained under 1/4 partition. (a) Input images. (b) Ground Truth labels. (c) Model output when trained with only labeled images. (d) Model output when training without SWD loss. (e) Predictions from SWSEG. When including SWD loss, the predictions are better in ambiguous regions and object borders." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.524 + ], + "angle": 0, + "content": "tion 5.2] and compare the values for models trained with and without \\(\\mathcal{L}_{swd}\\) in Tab. 5. The results show that including \\(\\mathcal{L}_{swd}\\) decreases the uniformity loss, which aligns with our qualitative visualization in Fig. 5. We also conduct an ablation study on MLP layers and \\(\\mathcal{L}_{reg}\\) in Tab. 4, showing that including all components yields the best outcome." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.909, + 0.66 + ], + "angle": 0, + "content": "Qualitative analysis. Fig. 6 shows the results of different training methods on the PASCAL VOC 2012 val set. When training with SWD loss, our method outperforms other methods, producing segmentation predictions that are closer to ground-truth labels. We can see from the results that our method performs exceptionally well on the border between different classes. Moreover, the segmentation results of our method have a cleaner prediction between the border and the backgrounds." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We introduced SWSEG, a semi-supervised semantic segmentation algorithm designed to enhance both alignment and uniformity of feature representations. We implement the Sliced-Wasserstein Distance (SWD) as an additional loss to jointly optimize these aspects, using a Gaussian-approximated variant to reduce computational overhead while preserving uniformity. Empirical evaluations on PASCAL VOC 2012, CityScapes, and ADE20K datasets show that SWSEG outperforms both supervised baselines and existing methods. Extensive ablation studies highlight the critical role of uniformity in the feature space for effective semi-supervised semantic segmentation. SWSEG is the first method to link semi-supervised learning with the uniformity metric using SWD, marking a significant advancement in the field." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20240" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.256 + ], + "angle": 0, + "content": "This material is based in part upon work supported by the National Science Foundation under Grant Numbers CNS-2333487 (NSF Frontier award) and CNS-2146449 (NSF CAREER award) and DEVCOM ARL Army Research Office under Contract number W911NF-2020-221. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors. The authors thank the reviewers for their enthusiastic comments and their valuable insights that improved our paper." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.287, + 0.188, + 0.303 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.312, + 0.486, + 0.367 + ], + "angle": 0, + "content": "[1] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.369, + 0.484, + 0.425 + ], + "angle": 0, + "content": "[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.429, + 0.484, + 0.482 + ], + "angle": 0, + "content": "[3] Olivier Chapelle and Alexander Zien. Semi-supervised classification by low density separation. In International workshop on artificial intelligence and statistics, pages 57-64. PMLR, 2005. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.484, + 0.556 + ], + "angle": 0, + "content": "[4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.558, + 0.484, + 0.627 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[7] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[8] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.312, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[10] D.C. Dowson and B.V. Landau. The fréchet distance between multivariate normal distributions. Journal of multivariate analysis, 12(3):450-455, 1982. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.908, + 0.204 + ], + "angle": 0, + "content": "[11] Lutz Dumbgen, Del Conte-Zerial, et al. On low-dimensional projections of high-dimensional distributions. In From Probability to Statistics and Back: High-Dimensional Models and Processes-A Festschrift in Honor of Jon A. Wellner, pages 91-105. Institute of Mathematical Statistics, 2013. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.908, + 0.259 + ], + "angle": 0, + "content": "[12] Mark Everingham and John Winn. The pascal visual object classes challenge 2012 (voc2012) development kit. Pattern Anal. Stat. Model. Comput. Learn., Tech. Rep, 2007:1-45, 2012. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.907, + 0.302 + ], + "angle": 0, + "content": "[13] Xianghong Fang, Jian Li, Qiang Sun, and Benyou Wang. Rethinking the uniformity metric in self-supervised learning. arXiv preprint arXiv:2403.00642, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.907, + 0.357 + ], + "angle": 0, + "content": "[14] Geoff French, Samuli Laine, Timo Aila, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, varied perturbations. arXiv preprint arXiv:1906.01916, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[15] Karl Pearson F.R.S. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2 (11):559-572, 1901. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.415, + 0.907, + 0.523 + ], + "angle": 0, + "content": "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284. Curran Associates, Inc., 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.907, + 0.58 + ], + "angle": 0, + "content": "[17] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991–998. IEEE, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.582, + 0.907, + 0.635 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.907, + 0.691 + ], + "angle": 0, + "content": "[19] Ying Jin, Jiaqi Wang, and Dahua Lin. Semi-supervised semantic segmentation via gentle teaching assistant. Advances in Neural Information Processing Systems, 35:2803-2816, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.907, + 0.732 + ], + "angle": 0, + "content": "[20] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.907, + 0.789 + ], + "angle": 0, + "content": "[21] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[22] Shikun Liu, Shuaifeng Zhi, Edward Johns, and Andrew J Davison. Bootstrapping semantic segmentation with regional contrast. arXiv preprint arXiv:2104.04465, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[23] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20241" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.162 + ], + "angle": 0, + "content": "[24] Huayu Mai, Rui Sun, Tianzhu Zhang, and Feng Wu. Rankmatch: Exploring the better consistency regularization for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3391-3401, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.485, + 0.219 + ], + "angle": 0, + "content": "[25] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. Umap: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.485, + 0.289 + ], + "angle": 0, + "content": "[26] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.485, + 0.348 + ], + "angle": 0, + "content": "[27] Jaemin Na, Jung-Woo Ha, Hyung Jin Chang, Dongyoon Han, and Wonjun Hwang. Switching temporary teachers for semi-supervised semantic segmentation. Advances in Neural Information Processing Systems, 36, 2023. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.485, + 0.417 + ], + "angle": 0, + "content": "[28] Kimia Nadjahi, Alain Durmus, Pierre E Jacob, Roland Badeau, and Umut Simsekli. Fast approximation of the sliced-wasserstein distance using concentration of random projections. Advances in Neural Information Processing Systems, 34:12411-12424, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.485, + 0.489 + ], + "angle": 0, + "content": "[29] Yassine Ouali, Celine Hudelot, and Myriam Tami. Semi-Supervised Semantic Segmentation With Cross-Consistency Training. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12671–12681, Seattle, WA, USA, 2020. IEEE. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.485, + 0.574 + ], + "angle": 0, + "content": "[30] Julien Rabin, Gabriel Peyré, Julie Delon, and Marc Bernot. Wasserstein barycenter and its application to texture mixing. In Scale Space and Variational Methods in Computer Vision: Third International Conference, SSVM 2011, Ein-Gedi, Israel, May 29-June 2, 2011, Revised Selected Papers 3, pages 435-446. Springer, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.485, + 0.63 + ], + "angle": 0, + "content": "[31] Galen Reeves. Conditional central limit theorems for gaussian projections. In 2017 IEEE International Symposium on Information Theory (ISIT), pages 3045-3049. IEEE, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.485, + 0.688 + ], + "angle": 0, + "content": "[32] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.485, + 0.761 + ], + "angle": 0, + "content": "[33] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.485, + 0.843 + ], + "angle": 0, + "content": "[34] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems, 33:596-608, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.485, + 0.901 + ], + "angle": 0, + "content": "[35] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Advances in Neural Information Processing Systems, 33:596-608, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[36] Boyuan Sun, Yuqi Yang, Le Zhang, Ming-Ming Cheng, and Qibin Hou. Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation. IEEE Computer Vision and Pattern Recognition (CVPR), 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[37] Rui Sun, Huayu Mai, Tianzhu Zhang, and Feng Wu. Daw: Exploring the better weighting function for semi-supervised semantic segmentation. Advances in Neural Information Processing Systems, 36, 2023. 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.907, + 0.288 + ], + "angle": 0, + "content": "[38] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.907, + 0.343 + ], + "angle": 0, + "content": "[39] Haonan Wang, Qixiang Zhang, Yi Li, and Xiaomeng Li. Allspark: Reborn labeled features from unlabeled in transformer for semi-supervised semantic segmentation. arXiv preprint arXiv:2403.01818, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.907, + 0.4 + ], + "angle": 0, + "content": "[40] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.907, + 0.481 + ], + "angle": 0, + "content": "[41] Xiaoyang Wang, Huihui Bai, Limin Yu, Yao Zhao, and Jimin Xiao. Towards the uncharted: Density-descending feature perturbation for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3303-3312, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.908, + 0.566 + ], + "angle": 0, + "content": "[42] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudolabels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4248-4257, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.909, + 0.636 + ], + "angle": 0, + "content": "[43] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34:12077-12090, 2021. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.907, + 0.692 + ], + "angle": 0, + "content": "[44] Haiming Xu, Lingqiao Liu, Qiuchen Bian, and Zhen Yang. Semi-supervised semantic segmentation with prototype-based consistency regularization. Advances in Neural Information Processing Systems, 35:26007-26020, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.907, + 0.762 + ], + "angle": 0, + "content": "[45] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4268-4277, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.907, + 0.831 + ], + "angle": 0, + "content": "[46] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7236-7246, 2023. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[47] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.909, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20242" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.147 + ], + "angle": 0, + "content": "[48] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310–12320. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.486, + 0.205 + ], + "angle": 0, + "content": "[49] Shaofeng Zhang, Feng Zhu, Junchi Yan, Rui Zhao, and Xiaokang Yang. Zero-cl: Instance and feature decorrelation for negative-free symmetric contrastive learning. In International Conference on Learning Representations, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[50] Zhen Zhao, Lihe Yang, Sifan Long, Jimin Pi, Luping Zhou, and Jingdong Wang. Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11350-11359, 2023. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[51] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "20243" + } + ] +] \ No newline at end of file diff --git a/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/ca5cab4a-bdaf-48a0-9763-5efeeaa13237_origin.pdf b/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/ca5cab4a-bdaf-48a0-9763-5efeeaa13237_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5775bf1f3ceb99227db4537aa7eff1f6248cbb34 --- /dev/null +++ b/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/ca5cab4a-bdaf-48a0-9763-5efeeaa13237_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f6b1ffbb47d3ce880f63ff468c216dcac3dcef3ad1762496eac6fa4de9039dc +size 5012002 diff --git a/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/full.md b/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4f8b6996d23a4c917f61b4406b2b7c6a1cf774a3 --- /dev/null +++ b/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/full.md @@ -0,0 +1,351 @@ +# Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity + +Chen-Yi Lu Kasra Derakhshandeh Somali Chaterji Purdue University + +lu842@purdue.edu, kderakhs@purdue.edu, schaterji@purdue.edu + +# Abstract + +Semi-supervised semantic segmentation with consistency regularization capitalizes on unlabeled images to enhance the accuracy of pixel-level segmentation. Current consistency learning methods primarily rely on the consistency loss between pseudo-labels and unlabeled images, neglecting the information within the feature representations of the backbone encoder. Preserving maximum information in feature embeddings requires achieving the alignment and uniformity objectives, as widely studied. To address this, we present SWSEG, a semi-supervised semantic segmentation algorithm that optimizes alignment and uniformity using the Sliced-Wasserstein Distance (SWD), and rigorously and empirically proves this connection. We further resolve the computational issues associated with conventional Monte Carlo-based SWD by implementing a Gaussian-approximated variant, which not only maintains the alignment and uniformity objectives but also improves training efficiency. We evaluate SWSEG on the PASCAL VOC 2012, Cityscapes, and ADE20K datasets, outshining supervised baselines in mIoU by up to $11.8\%$ , $8.9\%$ , and $8.2\%$ , respectively, given an equivalent number of labeled samples. Further, SWSEG surpasses state-of-the-art methods in multiple settings across these three datasets. Our extensive ablation studies confirm the optimization of the uniformity and alignment objectives of the feature representations. + +# 1. Introduction + +Semantic segmentation, the task of assigning pixel-level labels to images, is a cornerstone in domains such as autonomous driving and medical image analysis. Semantic segmentation models typically comprise an encoder and a decoder. The encoder extracts meaningful features from the input images, while the decoder assigns the feature embeddings to desired pixel-level labels. However, its effectiveness is often hampered by the need for costly pixel-level annotated data [9]. To mitigate this, semi-supervised learning + +![](images/9d04dc6f268f1caeefd63cc33ccfe157d8fa1583edab6aafff1b48f678c66697.jpg) +Figure 1. SWSEG demonstrates marked superiority in mIoU across varying data splits of the ADE20K dataset. + +(SSL) was proposed [3]. SSL leverages a limited amount of labeled data in conjunction with a larger pool of readily available unlabeled data. The key lies in effectively harnessing the unlabeled data to improve the model's performance. + +Pseudo-labeling and consistency regularization are two prominent techniques in SSL. Pseudo-labeling generates temporary labels from the model's predictions during training and using these to train the model in a supervised manner [2, 26]. The model then computes the loss between these pseudo-labels and the corresponding unlabeled images. This technique allows the model to leverage unlabeled data, thereby improving its performance. Consistency regularization, on the other hand, encourages the model to produce consistent predictions for different perturbations of the same input image, thereby promoting robustness and generalization [14, 32]. This technique is grounded in the smoothness assumption, which posits that data points with similar features should have similar labels [3]. In practice, random augmentations are applied to input images, and the model is trained to minimize discrepancies between its predictions on the original and augmented versions. In the realm of semantic segmentation, consistency regularization is often paired with pseudo-labeling, where weakly augmented unlabeled + +![](images/363573a54e3ddec1f229574c2dd038224043414501d1f3e6b592354548ef1e97.jpg) +Input + +![](images/8f388bc10592902608f270c767326baf347696dd2bf645e34e6a86869ff5c962.jpg) + +![](images/01a54dbd302f93ff10838d9066ee8e876236316dd46e4aad4b20d3933c70f1bb.jpg) +(a) + +![](images/79eae512f77a09ccb44ea291b6e727636c74ebe06b85847c61b13223de5950cc.jpg) + +![](images/870460a032788f40fba01022108f754aa58b82ef9d06ffa3533388661a05e5bd.jpg) +(b) + +![](images/23dc3582923f82acc9376da1413e11d050b12f53684b9dc621631e10f8fb8c56.jpg) + +![](images/ec35b044e279de6d6c6f667d86e63bfa8a340e31273e7abee1694b029446d1af.jpg) +(c) +Figure 2. Grad-CAM visualization [33] and model predictions. (a) depicts results from a model trained without alignment optimization, leading to inadequate boundary definition. (b) shows moderate improvements using traditional cross-entropy loss, which partially improves boundary adherence. (c) illustrates SWSEG with SWD loss, significantly enhancing segmentation accuracy by ensuring uniformity and alignment of features. + +data generate temporary "pseudo" labels for training [2, 34]. Concurrently, a parallel training stream employs strongly augmented data to calculate and minimize consistency losses, further enhancing model robustness [8, 23, 29, 46, 50]. + +While consistency-based methods are effective, existing approaches focus solely on enforcing regularization at the decoder's output, neglecting the rich information embedded within the intermediate feature representations from the backbone encoder. This oversight neglects the potential of extending consistency regularization to the feature level, guided by the smoothness assumption. We argue that enforcing consistency within these feature representations, derived from CNN-based or Transformer-based backbone encoders [4, 43], is crucial for robust learning against image perturbations. In Fig. 2, we insert a cross entropy loss on the output of encoders between feature embeddings of weakly and strongly augmented images. The results show that it already improves the feature separability and segmentation performance. + +To explore the importance of superior feature representations, we draw parallels with the self-supervised contrastive learning literature [1, 6, 48]. These methods, which involve contrasting positive (similar) and negative (dissimilar) pairs, excel at promoting both alignment and uniformity [13, 40]. Alignment, which aligns with the smoothness assumption, posits that perturbations to the same input should yield proximate features. Uniformity, on the other hand, implies that feature representations should be maximally informative, a concept often overlooked in semi-supervised semantic segmentation. These metrics provide a comprehensive framework for evaluating and enhancing feature representations, ultimately leading to improved model performance. The attraction between positive pairs in contrastive learning fosters + +alignment, while the repulsion of negative pairs encourages diverse and informative representations, enhancing uniformity. Inspired by the effectiveness of contrastive learning, we aim to bridge the gap in semi-supervised semantic segmentation by developing a method that explicitly optimizes both alignment and uniformity of feature representations. This raises a key question: How can we design an objective function that simultaneously achieves both uniformity and alignment of feature representations? + +In this work, we introduce SWSEG, a semi-supervised semantic segmentation algorithm that directly addresses this question. We leverage the Sliced-Wasserstein distance (SWD), a metric that quantifies the distance between probability distributions, to simultaneously optimize alignment and uniformity of feature representations extracted from the backbone encoder model. SWD achieves this by projecting features onto a unit hypersphere and computing the Wasserstein distance between the resulting one-dimensional embeddings. Empirically, we demonstrate that minimizing SWD effectively optimizes both alignment and uniformity between feature representations. Furthermore, to address the computational burden associated with traditional Monte Carlo estimation of SWD for high-dimensional features, we introduce a variant that projects feature embeddings onto a Gaussian distribution. This not only maintains the uniformity objective, as a normalized Gaussian distribution is uniformly distributed on the unit hypersphere, but also enables efficient computation by leveraging the analytical solution for the quadratic Wasserstein distance between two Gaussian distributions. Additionally, we incorporate a regularization term that decorrelates feature representations, fulfilling the necessary condition for projecting features onto a Gaussian distribution. We evaluated our algorithm on the PASCAL VOC 2012 [12], Cityscapes [9], and ADE20K [51] datasets under various partitioning protocols. Our results consistently demonstrate SWSEG's superiority over supervised baselines, with performance improvements of up to $11.8\%$ , $8.9\%$ , and $8.2\%$ on PASCAL VOC, CityScapes, and ADE20K, using the same amount of labeled data. Notably, SWSEG also achieves state-of-the-art (SOTA) results compared to existing methods under multiple data splitting protocols, as illustrated in Figure 1 [19, 23, 27]. Through extensive ablation studies, we confirm the critical role of uniformity in feature space for effective semi-supervised semantic segmentation. To the best of our knowledge, SWSEG is the first to establish a direct link between SSL and the uniformity metric, while also implementing the SWD to enhance feature representations. Our contributions can be summarized as follows: + +1. We introduce SWSEG, a semi-supervised semantic segmentation algorithm that explicitly optimizes feature representation uniformity and alignment using SWD, and back up our claim with extensive empirical studies. + +2. We introduce an efficient variant of SWD estimation that projects feature embeddings onto a Gaussian distribution. This approach maintains uniformity while leveraging the analytical solution for quadratic Wasserstein distance between Gaussians, reducing computational complexity. +3. SWSEG achieves SOTA results on PASCAL VOC 2012, Cityscapes, and ADE20K datasets, outperforming supervised baselines and existing semi-supervised methods, with significant performance improvements of up to $11.8\%$ on PASCAL VOC, $8.9\%$ on CityScapes, and $8.2\%$ on ADE20K compared to supervised methods. + +# 2. Related Works + +# 2.1. Semi-Supervised Learning + +Semi-supervised learning (SSL) has emerged as a promising solution to alleviate the burden of collecting labeled data, enabling the training of DNNs using only a fraction of labeled data while still achieving impressive results. Pseudo-labeling is a widely adopted technique which leverages the model's predictions on unlabeled data as surrogate labels [21]. This approach has been shown to improve model generalization by shifting the decision boundary towards low-density regions [3]. Additionally, many SSL algorithms incorporate consistency regularization, which regularizes the model to generate consistent outputs regardless of perturbations applied to the input data [20, 35, 38]. Laine et al. [20] combined cross-entropy loss with consistency loss for training, and incorporated model predictions into the pseudo-label set through an exponential moving average. Tarvainen and Valpola [38] posited that directly updating the model weights at each mini-batch accelerates the learning process while enhancing the model's performance. Sohn et al. [35] generated pseudo-labels from model predictions on weakly-augmented input images and computed the unsupervised loss between these pseudo-labels and the model predictions from strongly-augmented input images. However, these SSL methods were designed for image classification, which is not suitable for semantic segmentation where pixel-wise predictions are desired. + +# 2.2. Semi-Supervised Semantic Segmentation + +A number of research papers have incorporated consistency training and pseudo-labeling in SSL for segmentation tasks [29, 37, 39, 42, 44-46, 50]. For instance, Wang et al. [42] stored pseudo-labels with lower confidence as negative samples and computed contrastive loss using these samples. Yang et al. [46] introduced an additional teacher model and used dropout as feature-wise perturbation for consistency-based pseudo-labeling. However, these methods typically enforce consistency regularization solely on the decoder output, neglecting the rich information present in the intermediate feature representations produced by the + +backbone encoder. Recent works leveraging feature embedding in SSL segmentation include AllSpark [39], DDFP [41], and CorrMatch [36]. AllSpark balances the importance of labeled and unlabeled training streams by implementing a cross-attention mechanism in the feature embedding. DDFP injects perturbation in the feature space to regularize the decision boundary, with perturbation locations guided by a lightweight normalizing flow-based density estimator. CorrMatch utilizes the correlation map of features from strongly and weakly augmented inputs to regularize pseudo-label supervision in the decoder output. In contrast, SWSEG aims to fill this gap by explicitly focusing on training the backbone encoder to produce more meaningful feature representations for the segmentation head. By optimizing both the alignment and uniformity of these feature representations, SWSEG aims to improve overall segmentation performance. This approach goes beyond the limitations of existing methods that primarily focus on designing complex strategies such as cross-attention layers, normalizing flows, or heuristic label and region propagation techniques. + +# 2.3. Uniformity of Feature Representation + +Although existing SSL methods for semantic segmentation have demonstrated promising results through consistency learning, they typically impose consistency on the model's output. While few studies have explored consistency regularization on feature representations from the backbone model, this area remains under-explored compared to extensive research in self-supervised learning [1, 6, 48]. In self-supervised learning, where labeled data is scarce or absent, preventing representational collapse is paramount. This collapse occurs when the model produces uninformative feature representations, regardless of variations in the input. While the uniformity metric, which assesses how evenly features are distributed across the hypersphere, has been effectively used in self-supervised learning to mitigate collapse [40], its application in semi-supervised semantic segmentation has been absent. Recognizing the potential benefits of uniformity in preventing representational collapse and maximizing information preservation, we incorporate this metric into our semi-supervised approach. We further establish a connection between uniformity and the sliced-SWD metric, demonstrating that minimizing SWD between feature representations of differently augmented inputs can simultaneously maximize feature uniformity. This unified approach enhances model performance by ensuring that feature representations are both consistent and informative. + +# 3. Method + +This section begins by introducing the conventional framework for consistency learning-based SSL. We then argue for the importance of enforcing consistency regularization in the feature embedding, discussing the roles of alignment and + +![](images/20baeedbf32f3a51be025b25bef2c7f1f6db7308e3bcefa90bde067e91d38023.jpg) +Figure 3. SWSEG Framework for Joint Optimization of Alignment and Uniformity. The SWSEG framework employs both weak $(\mathcal{A}^w)$ and strong perturbations $(\mathcal{A}^s)$ applied to unlabeled input images $x^u$ . Both augmentations are passed through a shared backbone encoder to extract features, which are then mapped to a different space via non-linear MLPs, producing feature embeddings. To achieve both feature alignment and uniformity, SWSEG introduces an SWD loss $\mathcal{L}_{swd}$ , computed between these intermediate feature embeddings. The goal is to encourage the feature embeddings from different augmentations to be consistent (alignment) and to be spread out uniformly in the feature space, maximizing information (uniformity). An unsupervised loss $\mathcal{L}_u$ is computed to ensure consistency between segmentations generated from weakly and strongly augmented inputs. + +uniformity in achieving better feature representations. Next, we introduce a Gaussian-projected SWD metric to simultaneously achieve uniformity and alignment in the feature space. Finally, we detail the integration of these elements within the SWSEG framework. + +# 3.1. Preliminary + +In semi-supervised learning, we are given labeled image sets, $D^{l} = \{x_{i}^{l},y_{i}^{l}\}$ , and unlabeled ones, $D^{u} = \{x_{j}^{u}\}$ . Typically, $D^{u}$ greatly outnumbers $D^{l}$ . Our goal is to leverage these unlabeled images to optimize the parameters $\theta$ of a semantic segmentation model $f_{\theta}(\cdot)$ . + +In SWSEG, we adopt a typical framework from Fix-Match [34], training the model through two separate feedforward streams: supervised and unsupervised. In the supervised stream, we draw a batch of images from $D^{l}$ and compute the cross-entropy loss $\mathcal{L}_s$ between the model's predictions $\hat{y}_i^l$ and the labels $y_i^l$ . Conversely, the unsupervised stream draws images from $D^{u}$ , to which we apply weak or strong perturbations, $\mathcal{A}^w(\cdot)$ and $\mathcal{A}^s(\cdot)$ , respectively. The unsupervised loss $\mathcal{L}_u$ is then minimized according to the equation (1): + +$$ +\mathcal {L} _ {u} = \frac {1}{| \mathcal {D} ^ {u} |} \sum_ {x \in \mathcal {D} ^ {u}} \mathcal {H} \left(f _ {\theta} \left(\mathcal {A} ^ {s} \left(x ^ {u}\right)\right), f _ {\theta} \left(\mathcal {A} ^ {w} \left(x ^ {u}\right)\right)\right), \tag {1} +$$ + +We generate pseudo-labels $f_{\theta}(\mathcal{A}^{w}(x^{u}))$ by feeding weakly-augmented images into the model. We then compute + +the unsupervised loss using strongly-augmented samples $f_{\theta}(\mathcal{A}^{s}(x^{u}))$ , using $\mathcal{H}$ , for cross-entropy loss. The pseudolabel generation does not contribute to gradient computation. + +# 3.2. Feature Alignment and Uniformity + +In semantic segmentation, $f_{\theta}(\cdot)$ typically comprises a backbone encoder $p_{\theta}$ and a segmentation head $q_{\theta}$ [4, 43]. Existing methods [23, 27, 37, 46, 50] focus on enforcing consistency regularization on the segmentation head's output, often overlooking backbone encoder features. To further harness the potential of feature representations, we draw on self-supervised contrastive learning, which trains networks to generate meaningful embeddings without labels, emphasizing alignment and uniformity as key objectives for assessing feature quality. + +Alignment aims to minimize differences between embeddings from similar samples. Intuitively, model should produce similar feature embeddings when images with same classes are input. We achieve this by varying the magnitude of perturbations applied to the input images and minimizing the distance between embeddings using a distance metric dist: + +$$ +\mathcal {L} _ {\text {a l i g n}} = \operatorname {d i s t} \left(p _ {\theta} \left(\mathcal {A} ^ {s} \left(x _ {i} ^ {u}\right)\right), p _ {\theta} \left(\mathcal {A} ^ {w} \left(x _ {i} ^ {u}\right)\right)\right). \tag {2} +$$ + +Uniformity promotes even distributions of feature representations across a hypersphere, enhancing model robustness by preventing collapse to a lower-dimensional subspace. + +While often less emphasized in semi-supervised learning due to the stabilizing influence of labeled data, uniformity is critical in self-supervised learning to prevent the model from generating identical embeddings regardless of input. By ensuring uniformity, the model preserves a rich diversity of information within the feature space, which significantly improves the performance of the segmentation head. + +# 3.3. Revisiting Sliced Wasserstein Distance + +Intuitively, the uniform distribution on a unit hypersphere exemplifies maximum uniformity. We need to establish an objective function that can achieve both uniformity and alignment. To this end, we implement the SWD as our distance metric. SWD calculates average distances from 1-D projections sampled from a unit hypersphere, thus inherently promoting both desired properties. + +The rationale behind SWD stems from its ability to reduce the computational complexity relative to the high-dimensional Wasserstein distances (WD). Unlike the traditional WD, which struggles with high dimensionality, SWD leverages the 1-dimensional Wasserstein distance, which can be computed in closed form: Let $\mu_d$ and $\nu_d$ be two probability measures in $\mathcal{P}(\Omega) \in \mathbb{R}^d$ : + +$$ +W _ {p} ^ {p} \left(\mu_ {d}, \nu_ {d}\right) = \left(\int_ {0} ^ {1} \left| F _ {\mu_ {d}} ^ {- 1} (z) - F _ {\nu_ {d}} ^ {- 1} (z) \right| ^ {p} d z\right) ^ {1 / p}, \tag {3} +$$ + +where $p$ is the order of the distance, and $F_{\mu_d}^{-1}$ and $F_{\nu_d}^{-1}$ are the quantile functions of $\mu_d$ and $\nu_d$ , respectively. Computationally, this is achieved by sorting and computing the mean $l^2$ -distance between two 1-D vectors. SWD [30] involves projecting high-dimensional features onto a unit hypersphere and calculating the 1-D Wasserstein distance between these projections. Formally, let $\mathbb{S}^{d-1}$ represent the unit sphere in $d$ dimensions, $\sigma$ the uniform distribution on $\mathbb{S}^{d-1}$ , and $P_{\theta\#}\xi$ the push-forward measure of $\xi \in \mu_d, \nu_d$ . The SWD is defined as: + +$$ +S W _ {p} ^ {p} \left(\mu_ {d}, \nu_ {d}\right) = \int_ {\mathbb {S} ^ {d - 1}} W _ {p} ^ {p} \left(P _ {\theta \#} \mu_ {d}, P _ {\theta \#} \nu_ {d}\right) \mathrm {d} \sigma (\theta). \tag {4} +$$ + +To approximate SWD, we vectorize the input distributions $\mu_{d}$ and $\nu_{d}$ from $\mathbb{R}^{n\times n\times d}$ to $\mathbb{R}^{n^2 d}$ . We then employ a Monte Carlo scheme, sampling $L$ directions from the unit sphere $\mathbb{S}^{d - 1}$ . Projecting $\mu_{d}$ and $\nu_{d}$ along these directions yields $L$ one-dimensional projections for each distribution. After sorting these projections, we calculate the 1D Wasserstein distance between them using Eq. (3). This method efficiently estimates the Wasserstein distance between feature embeddings, exploiting the closed form property of 1D calculations compared to high-dimensional alternatives. + +Although the unit sphere projection enables uniformity and minimizing 1-D Wasserstein distances facilitates alignment, training with Monte Carlo-approximated SWD + +presents several challenges. First, the results are inherently non-deterministic and intractable due to the randomness inherent in the Monte Carlo method. Second, as the dimensions of $\mu_d$ and $\nu_d$ increase, both the computational cost and memory requirements for SWD escalate. Finally, to maintain accuracy in the face of increasing dimensions, the number of directions $L$ must also increase, complicating the Monte Carlo approximation further. + +# 3.4. Proposed Approach: SWSEG + +To enhance feature representation uniformity and alignment while alleviating the computational burden of Monte Carlo-based SWD, we implement an alternate method proposed by Nadjahi et al. [28] for approximating the SWD. Our method diverges by projecting these measures directly onto a Gaussian distribution. This modification offers significant computational benefits and supports key theoretical objectives in two primary ways: First, projecting onto Gaussian distributions allows the 2-Wasserstein distance between any two Gaussians to be expressed in a closed-form equation. Second, using normalized Gaussian variables ensures that the resulting distributions are uniformly distributed over the unit hypersphere, which we prove in the supplementary materials. This uniform distribution is essential for maintaining the model's uniformity objective, ensuring features are evenly distributed in the embedding space, maximizing its information for segmentation head. + +To formalize this idea, for any positive integer $d$ , let $\{\theta_i\}_{i=1}^d$ be a sequence of i.i.d. 1-d standard Gaussian distributions and $\{X_i\}_{i=1}^d$ be a sequence of 1-d probability distributions. Several Central Limit Theorems [11, 31] ensure that the sequence of distributions formed by the inner product $d^{-1/2}\langle \theta_{1:d}, X_{1:d} \rangle$ converges in probability to a Gaussian random variable when the distributions exhibit weak correlation. Furthermore, according to Proposition 2 from [28], as the dimensions of the distributions increase, the quadratic Wasserstein distance between two zero-mean Gaussian distributions with variance $d^{-1}m_2(\bar{\mu}_d)$ can be approximated by the SW distance between two Gaussian projected distributions. The Wasserstein distance between two Gaussian probability distribution is in closed form [10], which enables us to approximate the $l^2$ -SW distance by: + +$$ +\mathcal {L} _ {s w d} = d ^ {- 1} \left(m _ {2} \left(\bar {\mu_ {d}}\right) ^ {\frac {1}{2}} - m _ {2} \left(\bar {\nu_ {d}}\right) ^ {\frac {1}{2}}\right) + d ^ {- 1} \left| m _ {\mu_ {d}} - m _ {\nu_ {d}} \right| _ {2} ^ {2}. \tag {5} +$$ + +This equation comprises two terms: the first quantifies the difference in the second moments of the distributions, and the second captures the squared error in their means. For $\xi_d \in \{\bar{\mu}_d, \bar{\nu}_d\}$ , we define $m_2(\xi_d) = n^{-1} \sum_{j=1}^n ||x^j||^2$ , enabling a closed-form deterministic solution for $\mathcal{L}_{swd}$ . Nadjahi et al. also proved in Proposition 1 that $\mathcal{L}_{swd}$ with Gaussian projections equals the original $SW_p^p$ when $p = 2$ . Thus, we can use $\mathcal{L}_{swd}$ as a computationally efficient Wasserstein metric for our consistency loss term. A PyTorch-styled + +Table 1. Results on ADE20k. The fractions represent the labeled image proportions sampled during training. The model is trained on SegFormer [43] with MiT-B1 backbone. We use mIOU as our evaluation metric. + +
Method1/32 (631)1/16 (1263)1/8 (2526)1/4 (5052)1/2 (10105)
Supervised Only15.222.325.929.532.6
CPS [8] [ICLR '21]18.822.327.932.436.9
U²PL [42] [CVPR '22]18.523.628.133.536.0
ReCo [22] [ICLR '22]21.123.727.730.235.5
GTA-Seg [19] [NeurIPS '22]19.523.126.927.829.9
DT [27] [NeurIPS '23]22.626.030.633.737.3
SWSEG (Ours)24.426.730.934.638.0
+ +pseudo code is presented in Algorithm 1. Furthermore, to ensure a more accurate approximation of the Gaussian-SWD and comply with the relaxed Central Limit Theorems in [11, 31], minimizing strong correlations among input embeddings is essential. We address this by introducing a regularization term based on the input cross-covariance matrix into the SWD loss: + +$$ +\mathcal {L} _ {r e g} = | \operatorname {C o v} (X) - \operatorname {D i a g} (X) | _ {2} ^ {2} + | \operatorname {C o v} (Y) - \operatorname {D i a g} (Y) | _ {2} ^ {2}. \tag {6} +$$ + +Here, $\mathrm{Cov}(\cdot)$ calculates the covariance matrix of the input features, and $\mathrm{Diag}(\cdot)$ extracts its diagonal elements. Minimizing the off-diagonal elements decorrelates the features, thereby indirectly promoting the uniformity of the distribution of the features across the representation space [49]. + +We present a detailed overview of this training scheme in Fig. 3. We also insert an additional non-linear MLP projection layer before loss computation between representations, following recent self-supervised learning work [6, 7, 16]. The projection layer is solely incorporated during the training process, and it is subsequently discarded when the model is deployed for inference. Our objective is to retain more information in the features prior to the MLP layer. This will enable the segmentation head to utilize the preserved information more effectively for the segmentation task. The specific architecture of the projection layer used in this work consists of two Conv - BatchNorm - ReLu blocks followed by a single Conv layer. The overall training objective, managing the balance of various losses, is summarized in Eq. (7). The hyperparameters $\lambda_1, \lambda_2$ , and $\lambda_3$ control the weight of each loss. + +$$ +\mathcal {L} _ {\text {o v e r a l l}} = \mathcal {L} _ {s} + \lambda_ {1} \mathcal {L} _ {u} + \lambda_ {2} \mathcal {L} _ {s w d} + \lambda_ {3} \mathcal {L} _ {\text {r e g}}. \tag {7} +$$ + +# 4. Evaluation + +# 4.1. Datasets + +ADE20K. [51] is a complex scene parsing dataset featuring 150 objects, with 25,574 training, 2,000 validation, and 3,000 + +Algorithm 1 Pseudocode of our Gaussian-SWD loss in a PyTorch-like style. +```python +def gaussian_swd_loss(x, y): + # Reshape the input tensors to (n, dim) + x, y = x.view(x.shape[0], -1), y.view(y.shape[0], -1) + n, dim = x.shape + meanx = torch.mean(x, dim=0) + xc = x - meanx # Zero mean + m2x = torch.mean(torch.linalg(norm(xc, dim=1)) * 2) / dim + meany = torch.mean(y, dim=0) + yc = y - meany # Zero mean + m2y = torch.mean(torch.linalg(norm(yc, dim=1)) * 2) / dim + m_t = torch.linalg(norm(meanx - meany) * 2 / dim + swd = m_t + (m2x * (1/2) - m2y * (1/2)) * 2 + return swd +``` + +testing images. For a fair comparison, we used the data split from dual-teacher [27]. + +PASCAL VOC 2012. [12] This dataset is a classic semantic segmentation dataset comprising 20 classes of objects and 1 background class, with 1464 training, 1449 validation, and 1546 testing high-quality images in the standard set. We incorporate the augmented set from [17], following previous works, increasing our labeled data to 10,582 images. + +Cityscapes. [9] This is an autonomous driving semantic segmentation dataset that consists of 19 semantic classes, with 2975, 500, and 1525 training, validation, and testing images, respectively. + +# 4.2. Implementation details. + +For ADE20K, we implemented our method with SegFormer [43] architecture with MiT-B1 backbone. For PASCAL VOC 2012 and CityScapes, to have a fair comparison with previous benchmark semi-supervised semantic segmentation methods [27, 37, 42, 45, 46, 50], we implement our method using ResNet-101 [18] as the backbone encoder, initialized with ImageNet pre-trained weights, and + +Table 2. Results on CityScapes dataset. The fractions represent the proportion of labeled images used during training. We use mIOU as our evaluation metric. + +
Method1/161/81/4
Supervised Only67.973.275.1
PCR [44] [NeurIPS '22]73.476.378.4
UniMatch [46] [CVPR '23]76.677.979.2
AugSeg [50] [CVPR '23]75.277.879.6
DT [27] [NeurIPS '23]76.878.479.5
DAW [37] [NeurIPS '23]76.678.479.8
CorrMatch [36] [CVPR '24]77.378.579.4
RankMatch [24] [CVPR '24]77.178.680.0
DDFP [41] [CVPR '24]77.178.279.9
SWSEG (Ours)76.878.679.5
+ +Table 3. Results on PASCAL VOC 2012 dataset. The fractions represent the proportion of labeled images used during training. We use mIOU as our evaluation metric. + +
Method1/161/81/4
Supervised Only67.271.373.5
UniMatch [46] [CVPR '23]75.076.877.5
AugSeg [50] [CVPR '23]77.077.378.8
DAW [37] [NeurIPS '23]78.578.979.6
CorrMatch [36] [CVPR '24]78.479.379.6
RankMatch [24] [CVPR '24]78.979.280.0
DDFP [41] [CVPR '24]78.378.979.8
SWSEG (Ours)79.079.379.9
+ +DeeplabV3+ [5] as the decoder. We use SGD optimizer with momentum of 0.9. The learning rate is set to 0.001, 0.02, and 5e-5 for PASCAL, CityScapes, and ADE20K, respectively. We utilize a polynomial scheduling strategy with 0.9 decay rate. + +Data augmentations. For weak augmentations, horizontal flipping, random cropping, and random scaling are applied to the training images. For strong augmentations, we apply random color jittering, random grayscale, random blurring, and Cutmix [47] to the input training images. The training image size is set as $513 \times 513$ , $769 \times 769$ , and $512 \times 512$ for PASCAL VOC, CityScapes, and ADE20K, respectively. The training epochs for VOC, CityScapes, and ADE20K is 80, 240, and 200 epochs. + +Evaluation. As in previous works, we employ mean Intersection-over-Union (mIOU) as our evaluation metric. For the PASCAL VOC 2012 and ADE20K, evaluations are on the center-cropped validation set, while for CityScapes, we employ a sliding window evaluation. + +Table 4. Ablation analysis of feature alignment loss functions, $\mathcal{L}_{reg}$ , and MLP layer. The results reveal that MSE, cross entropy, and Monte Carlo estimated SWD loss with $L = 128$ yield inferior results. The combination of Gaussian-SWD (GSWD) with $\mathcal{L}_{reg}$ and MLP produces the best performance, with an mIoU of 79.0. + +
Loss FunctionsLregMLPmIoU
MSECESWDGSWD
✓✓✓70.2
✓✓✓77.1
✓✓✓77.6
✓77.4
✓✓78.1
✓✓✓79.0
+ +Table 5. Analysis of the Uniformity value $\mathcal{L}_{\mathrm{uniform}}$ across different data splits of the PASCAL VOC dataset demonstrates that training with the SWD loss $\mathcal{L}_{swd}$ yields superior Uniformity values compared to training without this loss term. + +
Method1/81/21
Without Lswd-2.5896-2.7095-2.5574
With Lswd-2.6592-2.7104-2.6168
+ +# 4.3. Comparison with State-of-the-Arts + +Results on ADE20K Dataset. Tab. 1 demonstrates the performance of SWSEG against the supervised baseline and existing methods. SWSEG significantly improves over the supervised baseline by $8.2\%$ , $3.4\%$ , $3.7\%$ , $5.1\%$ , and $5.4\%$ mIoUs on the 1/32, 1/16, 1/8, 1/4, and 1/2 splits, respectively. Our method also outperforms the state-of-the-art across all data splits, showcasing the generalizability of SWSEG across different model architectures and datasets. + +Results on Cityscapes Dataset. Tab. 2 presents our results in comparison to the supervised baseline and SOTA methods. SWSEG outperforms the supervised baseline by $8.9\%$ , $5.3\%$ , and $4.4\%$ for the 1/16, 1/8, and 1/4 partitions, respectively. Additionally, SWSEG surpasses the previous state-of-the-art method, DAW, by $0.2\%$ in the 1/16 and 1/8 partitions, and performs on par with contemporary methods. + +Results on PASCAL VOC 2012 Dataset. Tab. 3 compares SWSEG with existing state-of-the-arts on PASCAL VOC 2012 set. We follow the data partitions in CPS [8]. We outperform the supervised baselines by $11.8\%$ , $8.0\%$ and $6.4\%$ mIoUs under 1/16, 1/8, and 1/4 partitioning protocols. Our method also beats the SOTA DDFP [37] by $0.7\%$ , $0.4\%$ , and $0.1\%$ mIoUs under 1/16, 1/8, and 1/4 splits, respectively. We also visualized the predicted features in Fig. 4 to evaluate the effectiveness of our algorithm's clustering capabilities. Our method demonstrates enhanced discriminative clustering compared to the Supervised Only and UniMatch methods. + +![](images/2bcc5edd985fb3aefa80d72e8b9e171aeeb3194b10f367119670c9da78c2923a.jpg) +(a) + +![](images/dd06d584a69794e256f67beefe020733ea72ba308d5cf8662645803d4b0bc696.jpg) +(b) + +![](images/76af2cb30ff0c58380c0b74fe2fd4002beaf5cf1c78f08d1db7a3afedfa25dd4.jpg) +(c) + +![](images/97d2962bbd79f7d564eabbe26f05508d9c01186fe8d3c61bd8a13d0476167de7.jpg) +Figure 4. UMAP visualization [25] of feature embeddings using models trained with: (a) Supervised Only, (b) UniMatch, and (c) SWSEG. These visualizations utilize the PASCAL VOC 2012 dataset. Our method demonstrates enhanced clustering capabilities, outperforming both the Supervised Only and UniMatch. +(a) Without $\mathcal{L}_{swd}$ +Figure 5. We visualize feature representations on a unit circle for models trained with and without $\mathcal{L}_{swd}$ for uniformity analysis. Features from the model with $\mathcal{L}_{swd}$ (b) are more uniformly distributed than those without $\mathcal{L}_{swd}$ (a). + +![](images/14e7a1bf8bb5db89879cfcc0473cc58d49804bc28bb400328dc30a189120cfd2.jpg) +(b) With $\mathcal{L}_{swd}$ + +# 4.4. Ablation Study + +Uniformity and Alignment. We conduct an empirical analysis of feature representation alignment and uniformity (Tab. 4). We compare different loss functions for enforcing consistency regularization on the feature representation. Our results indicate that, aside from MSE loss, additional loss terms generally enhance segmentation prediction performance, highlighting the critical role of feature alignment. Monte Carlo SWD with $L = 128$ outperforms MSE and cross-entropy but encounters out-of-memory issues at $L = 1000$ using an identical number of GPUs, highlighting its computational cost. Conversely, our Gaussian-based SWD approximation delivers optimal results with lower memory usage. For uniformity analysis, we reduce the dimension of feature embeddings using PCA [15], projecting them into $\mathbb{R}^2$ and plotting their coordinates on a unit circle, as shown in Fig. 5. The visualization demonstrates that models trained with $\mathcal{L}_{swd}$ achieve greater uniformity compared to those without it, confirming the effectiveness of our introduced SWD loss. To quantitatively validate this observation, we compute the uniformity metric $\mathcal{L}_{\text{uniform}}$ from [40, Equa + +![](images/25dc148c14a869e9ea05a85f48e481ba8c75618295ab9660bbc703ffc26c921f.jpg) + +![](images/f3b7ee4544be51e46676b3ef4518189053923ec1551fa175a98b75608c252345.jpg) + +![](images/2b32a8bd06fe0edd0414e9fb70129c6e10ee0964f69ce01b01272acdbd8c0033.jpg) + +![](images/1f2f8303d9a8d80399ee63bb7a0ca8e31c653bf468d363def389bd91d73544b9.jpg) +(b) +Figure 6. Qualitative results on PASCAL VOC 2012 validation set. All models are trained under 1/4 partition. (a) Input images. (b) Ground Truth labels. (c) Model output when trained with only labeled images. (d) Model output when training without SWD loss. (e) Predictions from SWSEG. When including SWD loss, the predictions are better in ambiguous regions and object borders. + +![](images/c8d6ebeb99057817d16cdfde725180479b9a498e2fb1f138d5a16deec45eaa74.jpg) + +![](images/51d22d8443db3fe8e93042ee27abb3ecc681670ce7e4f92062dc16bca18aece2.jpg) + +![](images/afedc0bf2e3222106dbca9c402d809adc6925b60a5419bb85d867d48b561206d.jpg) + +![](images/fb590bec39f2b70943f64c5dcb2ac2ee822a43badc4a28db20bf28f58ee49b2f.jpg) +(c) + +![](images/8deed34be5a158aa0f11ab540e16dd665e8f924da735642a9d3e1f9e51e0bd34.jpg) + +![](images/381abb47076c5d9e82c1831b79481d75043d015f8164351aab7fb81202cefbac.jpg) + +![](images/3e7865b2343e42292a4c971b63f11acbabdd87d48fbbaad421a82d1e6f79f37a.jpg) + +![](images/f4ca899082d2907cc60405d4f86794b7eaaa8f54dce06e65c1f67cf2c6d51dc6.jpg) +(d) + +![](images/8db54dce7a8c91175a5397f7ec71e5786289681d6dff60421a99184c3d2cad29.jpg) + +![](images/4dc2fd916967a116d046a2a848d057195f5e25734615d3392fa1d58006179043.jpg) + +![](images/94a78afadb45ea89b6193d9630d2e8e3142520044ebc29a1b44226428d49ea0d.jpg) + +![](images/443007e41488d16f5bd391035ae13ebf1e11b4d24b1fc38643ed16ba14209dc8.jpg) +(c) + +tion 5.2] and compare the values for models trained with and without $\mathcal{L}_{swd}$ in Tab. 5. The results show that including $\mathcal{L}_{swd}$ decreases the uniformity loss, which aligns with our qualitative visualization in Fig. 5. We also conduct an ablation study on MLP layers and $\mathcal{L}_{reg}$ in Tab. 4, showing that including all components yields the best outcome. + +Qualitative analysis. Fig. 6 shows the results of different training methods on the PASCAL VOC 2012 val set. When training with SWD loss, our method outperforms other methods, producing segmentation predictions that are closer to ground-truth labels. We can see from the results that our method performs exceptionally well on the border between different classes. Moreover, the segmentation results of our method have a cleaner prediction between the border and the backgrounds. + +# 5. Conclusion + +We introduced SWSEG, a semi-supervised semantic segmentation algorithm designed to enhance both alignment and uniformity of feature representations. We implement the Sliced-Wasserstein Distance (SWD) as an additional loss to jointly optimize these aspects, using a Gaussian-approximated variant to reduce computational overhead while preserving uniformity. Empirical evaluations on PASCAL VOC 2012, CityScapes, and ADE20K datasets show that SWSEG outperforms both supervised baselines and existing methods. Extensive ablation studies highlight the critical role of uniformity in the feature space for effective semi-supervised semantic segmentation. SWSEG is the first method to link semi-supervised learning with the uniformity metric using SWD, marking a significant advancement in the field. + +# 6. Acknowledgements + +This material is based in part upon work supported by the National Science Foundation under Grant Numbers CNS-2333487 (NSF Frontier award) and CNS-2146449 (NSF CAREER award) and DEVCOM ARL Army Research Office under Contract number W911NF-2020-221. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors. The authors thank the reviewers for their enthusiastic comments and their valuable insights that improved our paper. + +# References + +[1] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. 2, 3 +[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in Neural Information Processing Systems, 32, 2019. 1, 2 +[3] Olivier Chapelle and Alexander Zien. Semi-supervised classification by low density separation. In International workshop on artificial intelligence and statistics, pages 57-64. PMLR, 2005. 1, 3 +[4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 2, 4 +[5] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 7 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 6 +[7] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 6 +[8] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 2, 6, 7 +[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 2, 6 + +[10] D.C. Dowson and B.V. Landau. The fréchet distance between multivariate normal distributions. Journal of multivariate analysis, 12(3):450-455, 1982. 5 +[11] Lutz Dumbgen, Del Conte-Zerial, et al. On low-dimensional projections of high-dimensional distributions. In From Probability to Statistics and Back: High-Dimensional Models and Processes-A Festschrift in Honor of Jon A. Wellner, pages 91-105. Institute of Mathematical Statistics, 2013. 5, 6 +[12] Mark Everingham and John Winn. The pascal visual object classes challenge 2012 (voc2012) development kit. Pattern Anal. Stat. Model. Comput. Learn., Tech. Rep, 2007:1-45, 2012. 2, 6 +[13] Xianghong Fang, Jian Li, Qiang Sun, and Benyou Wang. Rethinking the uniformity metric in self-supervised learning. arXiv preprint arXiv:2403.00642, 2024. 2 +[14] Geoff French, Samuli Laine, Timo Aila, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, varied perturbations. arXiv preprint arXiv:1906.01916, 2019. 1 +[15] Karl Pearson F.R.S. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2 (11):559-572, 1901. 8 +[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284. Curran Associates, Inc., 2020. 6 +[17] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991–998. IEEE, 2011. 6 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[19] Ying Jin, Jiaqi Wang, and Dahua Lin. Semi-supervised semantic segmentation via gentle teaching assistant. Advances in Neural Information Processing Systems, 35:2803-2816, 2022. 2, 6 +[20] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 3 +[21] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896, 2013. 3 +[22] Shikun Liu, Shuaifeng Zhi, Edward Johns, and Andrew J Davison. Bootstrapping semantic segmentation with regional contrast. arXiv preprint arXiv:2104.04465, 2021. 6 +[23] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022. 2, 4 + +[24] Huayu Mai, Rui Sun, Tianzhu Zhang, and Feng Wu. Rankmatch: Exploring the better consistency regularization for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3391-3401, 2024. 7 +[25] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. Umap: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861, 2018. 8 +[26] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 1 +[27] Jaemin Na, Jung-Woo Ha, Hyung Jin Chang, Dongyoon Han, and Wonjun Hwang. Switching temporary teachers for semi-supervised semantic segmentation. Advances in Neural Information Processing Systems, 36, 2023. 2, 4, 6, 7 +[28] Kimia Nadjahi, Alain Durmus, Pierre E Jacob, Roland Badeau, and Umut Simsekli. Fast approximation of the sliced-wasserstein distance using concentration of random projections. Advances in Neural Information Processing Systems, 34:12411-12424, 2021. 5 +[29] Yassine Ouali, Celine Hudelot, and Myriam Tami. Semi-Supervised Semantic Segmentation With Cross-Consistency Training. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12671–12681, Seattle, WA, USA, 2020. IEEE. 2, 3 +[30] Julien Rabin, Gabriel Peyré, Julie Delon, and Marc Bernot. Wasserstein barycenter and its application to texture mixing. In Scale Space and Variational Methods in Computer Vision: Third International Conference, SSVM 2011, Ein-Gedi, Israel, May 29-June 2, 2011, Revised Selected Papers 3, pages 435-446. Springer, 2012. 5 +[31] Galen Reeves. Conditional central limit theorems for gaussian projections. In 2017 IEEE International Symposium on Information Theory (ISIT), pages 3045-3049. IEEE, 2017. 5, 6 +[32] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 1 +[33] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2 +[34] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems, 33:596-608, 2020. 2, 4 +[35] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. + +Advances in Neural Information Processing Systems, 33:596-608, 2020. 3 +[36] Boyuan Sun, Yuqi Yang, Le Zhang, Ming-Ming Cheng, and Qibin Hou. Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation. IEEE Computer Vision and Pattern Recognition (CVPR), 2024. 3, 7 +[37] Rui Sun, Huayu Mai, Tianzhu Zhang, and Feng Wu. Daw: Exploring the better weighting function for semi-supervised semantic segmentation. Advances in Neural Information Processing Systems, 36, 2023. 3, 4, 6, 7 +[38] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3 +[39] Haonan Wang, Qixiang Zhang, Yi Li, and Xiaomeng Li. Allspark: Reborn labeled features from unlabeled in transformer for semi-supervised semantic segmentation. arXiv preprint arXiv:2403.01818, 2023. 3 +[40] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 2, 3, 8 +[41] Xiaoyang Wang, Huihui Bai, Limin Yu, Yao Zhao, and Jimin Xiao. Towards the uncharted: Density-descending feature perturbation for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3303-3312, 2024. 3, 7 +[42] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudolabels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4248-4257, 2022. 3, 6 +[43] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34:12077-12090, 2021. 2, 4, 6 +[44] Haiming Xu, Lingqiao Liu, Qiuchen Bian, and Zhen Yang. Semi-supervised semantic segmentation with prototype-based consistency regularization. Advances in Neural Information Processing Systems, 35:26007-26020, 2022. 3, 7 +[45] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4268-4277, 2022. 6 +[46] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7236-7246, 2023. 2, 3, 4, 6, 7 +[47] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 7 + +[48] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310–12320. PMLR, 2021. 2, 3 +[49] Shaofeng Zhang, Feng Zhu, Junchi Yan, Rui Zhao, and Xiaokang Yang. Zero-cl: Instance and feature decorrelation for negative-free symmetric contrastive learning. In International Conference on Learning Representations, 2021. 6 +[50] Zhen Zhao, Lihe Yang, Sifan Long, Jimin Pi, Luping Zhou, and Jingdong Wang. Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11350-11359, 2023. 2, 3, 4, 6, 7 +[51] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 2, 6 \ No newline at end of file diff --git a/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/images.zip b/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..638f3d47e6f924d9de572f59a3a7dbf7469f85f9 --- /dev/null +++ b/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c177a7d38a30508152e210cba53d8fdc20d7c5db08d588ccd50cb3f1c413a15 +size 441469 diff --git a/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/layout.json b/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..031f28f89ff7dafff3b98151df4b214705338dc7 --- /dev/null +++ b/2025/Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity/layout.json @@ -0,0 +1,9814 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 122, + 103, + 489, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 103, + 489, + 140 + ], + "spans": [ + { + "bbox": [ + 122, + 103, + 489, + 140 + ], + "type": "text", + "content": "Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 169, + 161, + 440, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 161, + 440, + 190 + ], + "spans": [ + { + "bbox": [ + 169, + 161, + 440, + 190 + ], + "type": "text", + "content": "Chen-Yi Lu Kasra Derakhshandeh Somali Chaterji Purdue University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 191, + 463, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 191, + 463, + 203 + ], + "spans": [ + { + "bbox": [ + 143, + 191, + 463, + 203 + ], + "type": "text", + "content": "lu842@purdue.edu, kderakhs@purdue.edu, schaterji@purdue.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 258, + 297, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 297, + 558 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 297, + 558 + ], + "type": "text", + "content": "Semi-supervised semantic segmentation with consistency regularization capitalizes on unlabeled images to enhance the accuracy of pixel-level segmentation. Current consistency learning methods primarily rely on the consistency loss between pseudo-labels and unlabeled images, neglecting the information within the feature representations of the backbone encoder. Preserving maximum information in feature embeddings requires achieving the alignment and uniformity objectives, as widely studied. To address this, we present SWSEG, a semi-supervised semantic segmentation algorithm that optimizes alignment and uniformity using the Sliced-Wasserstein Distance (SWD), and rigorously and empirically proves this connection. We further resolve the computational issues associated with conventional Monte Carlo-based SWD by implementing a Gaussian-approximated variant, which not only maintains the alignment and uniformity objectives but also improves training efficiency. We evaluate SWSEG on the PASCAL VOC 2012, Cityscapes, and ADE20K datasets, outshining supervised baselines in mIoU by up to " + }, + { + "bbox": [ + 54, + 258, + 297, + 558 + ], + "type": "inline_equation", + "content": "11.8\\%" + }, + { + "bbox": [ + 54, + 258, + 297, + 558 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 258, + 297, + 558 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 54, + 258, + 297, + 558 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 258, + 297, + 558 + ], + "type": "inline_equation", + "content": "8.2\\%" + }, + { + "bbox": [ + 54, + 258, + 297, + 558 + ], + "type": "text", + "content": ", respectively, given an equivalent number of labeled samples. Further, SWSEG surpasses state-of-the-art methods in multiple settings across these three datasets. Our extensive ablation studies confirm the optimization of the uniformity and alignment objectives of the feature representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 584, + 135, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 135, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 135, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": "Semantic segmentation, the task of assigning pixel-level labels to images, is a cornerstone in domains such as autonomous driving and medical image analysis. Semantic segmentation models typically comprise an encoder and a decoder. The encoder extracts meaningful features from the input images, while the decoder assigns the feature embeddings to desired pixel-level labels. However, its effectiveness is often hampered by the need for costly pixel-level annotated data [9]. To mitigate this, semi-supervised learning" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 232, + 552, + 373 + ], + "blocks": [ + { + "bbox": [ + 317, + 232, + 552, + 373 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 552, + 373 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 552, + 373 + ], + "type": "image", + "image_path": "9d04dc6f268f1caeefd63cc33ccfe157d8fa1583edab6aafff1b48f678c66697.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 380, + 555, + 402 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 402 + ], + "type": "text", + "content": "Figure 1. SWSEG demonstrates marked superiority in mIoU across varying data splits of the ADE20K dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 434, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 483 + ], + "type": "text", + "content": "(SSL) was proposed [3]. SSL leverages a limited amount of labeled data in conjunction with a larger pool of readily available unlabeled data. The key lies in effectively harnessing the unlabeled data to improve the model's performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "type": "text", + "content": "Pseudo-labeling and consistency regularization are two prominent techniques in SSL. Pseudo-labeling generates temporary labels from the model's predictions during training and using these to train the model in a supervised manner [2, 26]. The model then computes the loss between these pseudo-labels and the corresponding unlabeled images. This technique allows the model to leverage unlabeled data, thereby improving its performance. Consistency regularization, on the other hand, encourages the model to produce consistent predictions for different perturbations of the same input image, thereby promoting robustness and generalization [14, 32]. This technique is grounded in the smoothness assumption, which posits that data points with similar features should have similar labels [3]. In practice, random augmentations are applied to input images, and the model is trained to minimize discrepancies between its predictions on the original and augmented versions. In the realm of semantic segmentation, consistency regularization is often paired with pseudo-labeling, where weakly augmented unlabeled" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20233" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 116, + 123 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 116, + 123 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 116, + 123 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 116, + 123 + ], + "type": "image", + "image_path": "363573a54e3ddec1f229574c2dd038224043414501d1f3e6b592354548ef1e97.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 125, + 100, + 134 + ], + "lines": [ + { + "bbox": [ + 78, + 125, + 100, + 134 + ], + "spans": [ + { + "bbox": [ + 78, + 125, + 100, + 134 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 124, + 70, + 178, + 123 + ], + "blocks": [ + { + "bbox": [ + 124, + 70, + 178, + 123 + ], + "lines": [ + { + "bbox": [ + 124, + 70, + 178, + 123 + ], + "spans": [ + { + "bbox": [ + 124, + 70, + 178, + 123 + ], + "type": "image", + "image_path": "8f388bc10592902608f270c767326baf347696dd2bf645e34e6a86869ff5c962.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 145, + 133, + 179, + 190 + ], + "blocks": [ + { + "bbox": [ + 145, + 133, + 179, + 190 + ], + "lines": [ + { + "bbox": [ + 145, + 133, + 179, + 190 + ], + "spans": [ + { + "bbox": [ + 145, + 133, + 179, + 190 + ], + "type": "image", + "image_path": "01a54dbd302f93ff10838d9066ee8e876236316dd46e4aad4b20d3933c70f1bb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 194, + 156, + 203 + ], + "lines": [ + { + "bbox": [ + 145, + 194, + 156, + 203 + ], + "spans": [ + { + "bbox": [ + 145, + 194, + 156, + 203 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 181, + 70, + 235, + 123 + ], + "blocks": [ + { + "bbox": [ + 181, + 70, + 235, + 123 + ], + "lines": [ + { + "bbox": [ + 181, + 70, + 235, + 123 + ], + "spans": [ + { + "bbox": [ + 181, + 70, + 235, + 123 + ], + "type": "image", + "image_path": "79eae512f77a09ccb44ea291b6e727636c74ebe06b85847c61b13223de5950cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 182, + 137, + 235, + 190 + ], + "blocks": [ + { + "bbox": [ + 182, + 137, + 235, + 190 + ], + "lines": [ + { + "bbox": [ + 182, + 137, + 235, + 190 + ], + "spans": [ + { + "bbox": [ + 182, + 137, + 235, + 190 + ], + "type": "image", + "image_path": "870460a032788f40fba01022108f754aa58b82ef9d06ffa3533388661a05e5bd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 192, + 214, + 202 + ], + "lines": [ + { + "bbox": [ + 203, + 192, + 214, + 202 + ], + "spans": [ + { + "bbox": [ + 203, + 192, + 214, + 202 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 239, + 70, + 293, + 124 + ], + "blocks": [ + { + "bbox": [ + 239, + 70, + 293, + 124 + ], + "lines": [ + { + "bbox": [ + 239, + 70, + 293, + 124 + ], + "spans": [ + { + "bbox": [ + 239, + 70, + 293, + 124 + ], + "type": "image", + "image_path": "23dc3582923f82acc9376da1413e11d050b12f53684b9dc621631e10f8fb8c56.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 240, + 137, + 294, + 192 + ], + "blocks": [ + { + "bbox": [ + 240, + 137, + 294, + 192 + ], + "lines": [ + { + "bbox": [ + 240, + 137, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 240, + 137, + 294, + 192 + ], + "type": "image", + "image_path": "ec35b044e279de6d6c6f667d86e63bfa8a340e31273e7abee1694b029446d1af.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 193, + 272, + 202 + ], + "lines": [ + { + "bbox": [ + 261, + 193, + 272, + 202 + ], + "spans": [ + { + "bbox": [ + 261, + 193, + 272, + 202 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 213, + 296, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 291 + ], + "type": "text", + "content": "Figure 2. Grad-CAM visualization [33] and model predictions. (a) depicts results from a model trained without alignment optimization, leading to inadequate boundary definition. (b) shows moderate improvements using traditional cross-entropy loss, which partially improves boundary adherence. (c) illustrates SWSEG with SWD loss, significantly enhancing segmentation accuracy by ensuring uniformity and alignment of features." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 317, + 296, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 366 + ], + "type": "text", + "content": "data generate temporary \"pseudo\" labels for training [2, 34]. Concurrently, a parallel training stream employs strongly augmented data to calculate and minimize consistency losses, further enhancing model robustness [8, 23, 29, 46, 50]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 366, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 544 + ], + "type": "text", + "content": "While consistency-based methods are effective, existing approaches focus solely on enforcing regularization at the decoder's output, neglecting the rich information embedded within the intermediate feature representations from the backbone encoder. This oversight neglects the potential of extending consistency regularization to the feature level, guided by the smoothness assumption. We argue that enforcing consistency within these feature representations, derived from CNN-based or Transformer-based backbone encoders [4, 43], is crucial for robust learning against image perturbations. In Fig. 2, we insert a cross entropy loss on the output of encoders between feature embeddings of weakly and strongly augmented images. The results show that it already improves the feature separability and segmentation performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": "To explore the importance of superior feature representations, we draw parallels with the self-supervised contrastive learning literature [1, 6, 48]. These methods, which involve contrasting positive (similar) and negative (dissimilar) pairs, excel at promoting both alignment and uniformity [13, 40]. Alignment, which aligns with the smoothness assumption, posits that perturbations to the same input should yield proximate features. Uniformity, on the other hand, implies that feature representations should be maximally informative, a concept often overlooked in semi-supervised semantic segmentation. These metrics provide a comprehensive framework for evaluating and enhancing feature representations, ultimately leading to improved model performance. The attraction between positive pairs in contrastive learning fosters" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "alignment, while the repulsion of negative pairs encourages diverse and informative representations, enhancing uniformity. Inspired by the effectiveness of contrastive learning, we aim to bridge the gap in semi-supervised semantic segmentation by developing a method that explicitly optimizes both alignment and uniformity of feature representations. This raises a key question: How can we design an objective function that simultaneously achieves both uniformity and alignment of feature representations?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 181, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 556, + 661 + ], + "type": "text", + "content": "In this work, we introduce SWSEG, a semi-supervised semantic segmentation algorithm that directly addresses this question. We leverage the Sliced-Wasserstein distance (SWD), a metric that quantifies the distance between probability distributions, to simultaneously optimize alignment and uniformity of feature representations extracted from the backbone encoder model. SWD achieves this by projecting features onto a unit hypersphere and computing the Wasserstein distance between the resulting one-dimensional embeddings. Empirically, we demonstrate that minimizing SWD effectively optimizes both alignment and uniformity between feature representations. Furthermore, to address the computational burden associated with traditional Monte Carlo estimation of SWD for high-dimensional features, we introduce a variant that projects feature embeddings onto a Gaussian distribution. This not only maintains the uniformity objective, as a normalized Gaussian distribution is uniformly distributed on the unit hypersphere, but also enables efficient computation by leveraging the analytical solution for the quadratic Wasserstein distance between two Gaussian distributions. Additionally, we incorporate a regularization term that decorrelates feature representations, fulfilling the necessary condition for projecting features onto a Gaussian distribution. We evaluated our algorithm on the PASCAL VOC 2012 [12], Cityscapes [9], and ADE20K [51] datasets under various partitioning protocols. Our results consistently demonstrate SWSEG's superiority over supervised baselines, with performance improvements of up to " + }, + { + "bbox": [ + 313, + 181, + 556, + 661 + ], + "type": "inline_equation", + "content": "11.8\\%" + }, + { + "bbox": [ + 313, + 181, + 556, + 661 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 181, + 556, + 661 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 313, + 181, + 556, + 661 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 181, + 556, + 661 + ], + "type": "inline_equation", + "content": "8.2\\%" + }, + { + "bbox": [ + 313, + 181, + 556, + 661 + ], + "type": "text", + "content": " on PASCAL VOC, CityScapes, and ADE20K, using the same amount of labeled data. Notably, SWSEG also achieves state-of-the-art (SOTA) results compared to existing methods under multiple data splitting protocols, as illustrated in Figure 1 [19, 23, 27]. Through extensive ablation studies, we confirm the critical role of uniformity in feature space for effective semi-supervised semantic segmentation. To the best of our knowledge, SWSEG is the first to establish a direct link between SSL and the uniformity metric, while also implementing the SWD to enhance feature representations. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "1. We introduce SWSEG, a semi-supervised semantic segmentation algorithm that explicitly optimizes feature representation uniformity and alignment using SWD, and back up our claim with extensive empirical studies." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20234" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 295, + 204 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 53, + 72, + 295, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 295, + 132 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 295, + 132 + ], + "type": "text", + "content": "2. We introduce an efficient variant of SWD estimation that projects feature embeddings onto a Gaussian distribution. This approach maintains uniformity while leveraging the analytical solution for quadratic Wasserstein distance between Gaussians, reducing computational complexity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 133, + 295, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 133, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 133, + 295, + 204 + ], + "type": "text", + "content": "3. SWSEG achieves SOTA results on PASCAL VOC 2012, Cityscapes, and ADE20K datasets, outperforming supervised baselines and existing semi-supervised methods, with significant performance improvements of up to " + }, + { + "bbox": [ + 53, + 133, + 295, + 204 + ], + "type": "inline_equation", + "content": "11.8\\%" + }, + { + "bbox": [ + 53, + 133, + 295, + 204 + ], + "type": "text", + "content": " on PASCAL VOC, " + }, + { + "bbox": [ + 53, + 133, + 295, + 204 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 53, + 133, + 295, + 204 + ], + "type": "text", + "content": " on CityScapes, and " + }, + { + "bbox": [ + 53, + 133, + 295, + 204 + ], + "type": "inline_equation", + "content": "8.2\\%" + }, + { + "bbox": [ + 53, + 133, + 295, + 204 + ], + "type": "text", + "content": " on ADE20K compared to supervised methods." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 217, + 147, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 147, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 147, + 229 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 237, + 203, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 203, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 203, + 250 + ], + "type": "text", + "content": "2.1. Semi-Supervised Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 255, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 296, + 553 + ], + "type": "text", + "content": "Semi-supervised learning (SSL) has emerged as a promising solution to alleviate the burden of collecting labeled data, enabling the training of DNNs using only a fraction of labeled data while still achieving impressive results. Pseudo-labeling is a widely adopted technique which leverages the model's predictions on unlabeled data as surrogate labels [21]. This approach has been shown to improve model generalization by shifting the decision boundary towards low-density regions [3]. Additionally, many SSL algorithms incorporate consistency regularization, which regularizes the model to generate consistent outputs regardless of perturbations applied to the input data [20, 35, 38]. Laine et al. [20] combined cross-entropy loss with consistency loss for training, and incorporated model predictions into the pseudo-label set through an exponential moving average. Tarvainen and Valpola [38] posited that directly updating the model weights at each mini-batch accelerates the learning process while enhancing the model's performance. Sohn et al. [35] generated pseudo-labels from model predictions on weakly-augmented input images and computed the unsupervised loss between these pseudo-labels and the model predictions from strongly-augmented input images. However, these SSL methods were designed for image classification, which is not suitable for semantic segmentation where pixel-wise predictions are desired." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 563, + 269, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 269, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 269, + 577 + ], + "type": "text", + "content": "2.2. Semi-Supervised Semantic Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "A number of research papers have incorporated consistency training and pseudo-labeling in SSL for segmentation tasks [29, 37, 39, 42, 44-46, 50]. For instance, Wang et al. [42] stored pseudo-labels with lower confidence as negative samples and computed contrastive loss using these samples. Yang et al. [46] introduced an additional teacher model and used dropout as feature-wise perturbation for consistency-based pseudo-labeling. However, these methods typically enforce consistency regularization solely on the decoder output, neglecting the rich information present in the intermediate feature representations produced by the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": "backbone encoder. Recent works leveraging feature embedding in SSL segmentation include AllSpark [39], DDFP [41], and CorrMatch [36]. AllSpark balances the importance of labeled and unlabeled training streams by implementing a cross-attention mechanism in the feature embedding. DDFP injects perturbation in the feature space to regularize the decision boundary, with perturbation locations guided by a lightweight normalizing flow-based density estimator. CorrMatch utilizes the correlation map of features from strongly and weakly augmented inputs to regularize pseudo-label supervision in the decoder output. In contrast, SWSEG aims to fill this gap by explicitly focusing on training the backbone encoder to produce more meaningful feature representations for the segmentation head. By optimizing both the alignment and uniformity of these feature representations, SWSEG aims to improve overall segmentation performance. This approach goes beyond the limitations of existing methods that primarily focus on designing complex strategies such as cross-attention layers, normalizing flows, or heuristic label and region propagation techniques." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 319, + 513, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 513, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 513, + 331 + ], + "type": "text", + "content": "2.3. Uniformity of Feature Representation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 336, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 556, + 635 + ], + "type": "text", + "content": "Although existing SSL methods for semantic segmentation have demonstrated promising results through consistency learning, they typically impose consistency on the model's output. While few studies have explored consistency regularization on feature representations from the backbone model, this area remains under-explored compared to extensive research in self-supervised learning [1, 6, 48]. In self-supervised learning, where labeled data is scarce or absent, preventing representational collapse is paramount. This collapse occurs when the model produces uninformative feature representations, regardless of variations in the input. While the uniformity metric, which assesses how evenly features are distributed across the hypersphere, has been effectively used in self-supervised learning to mitigate collapse [40], its application in semi-supervised semantic segmentation has been absent. Recognizing the potential benefits of uniformity in preventing representational collapse and maximizing information preservation, we incorporate this metric into our semi-supervised approach. We further establish a connection between uniformity and the sliced-SWD metric, demonstrating that minimizing SWD between feature representations of differently augmented inputs can simultaneously maximize feature uniformity. This unified approach enhances model performance by ensuring that feature representations are both consistent and informative." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 645, + 370, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 370, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 370, + 658 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "This section begins by introducing the conventional framework for consistency learning-based SSL. We then argue for the importance of enforcing consistency regularization in the feature embedding, discussing the roles of alignment and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20235" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 558, + 274 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 558, + 274 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 558, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 558, + 274 + ], + "type": "image", + "image_path": "20baeedbf32f3a51be025b25bef2c7f1f6db7308e3bcefa90bde067e91d38023.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 275, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 54, + 275, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 555, + 353 + ], + "type": "text", + "content": "Figure 3. SWSEG Framework for Joint Optimization of Alignment and Uniformity. The SWSEG framework employs both weak " + }, + { + "bbox": [ + 54, + 275, + 555, + 353 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}^w)" + }, + { + "bbox": [ + 54, + 275, + 555, + 353 + ], + "type": "text", + "content": " and strong perturbations " + }, + { + "bbox": [ + 54, + 275, + 555, + 353 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}^s)" + }, + { + "bbox": [ + 54, + 275, + 555, + 353 + ], + "type": "text", + "content": " applied to unlabeled input images " + }, + { + "bbox": [ + 54, + 275, + 555, + 353 + ], + "type": "inline_equation", + "content": "x^u" + }, + { + "bbox": [ + 54, + 275, + 555, + 353 + ], + "type": "text", + "content": ". Both augmentations are passed through a shared backbone encoder to extract features, which are then mapped to a different space via non-linear MLPs, producing feature embeddings. To achieve both feature alignment and uniformity, SWSEG introduces an SWD loss " + }, + { + "bbox": [ + 54, + 275, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{swd}" + }, + { + "bbox": [ + 54, + 275, + 555, + 353 + ], + "type": "text", + "content": ", computed between these intermediate feature embeddings. The goal is to encourage the feature embeddings from different augmentations to be consistent (alignment) and to be spread out uniformly in the feature space, maximizing information (uniformity). An unsupervised loss " + }, + { + "bbox": [ + 54, + 275, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 54, + 275, + 555, + 353 + ], + "type": "text", + "content": " is computed to ensure consistency between segmentations generated from weakly and strongly augmented inputs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 378, + 297, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 297, + 438 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 297, + 438 + ], + "type": "text", + "content": "uniformity in achieving better feature representations. Next, we introduce a Gaussian-projected SWD metric to simultaneously achieve uniformity and alignment in the feature space. Finally, we detail the integration of these elements within the SWSEG framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 135, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 135, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 135, + 458 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "text", + "content": "In semi-supervised learning, we are given labeled image sets, " + }, + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "inline_equation", + "content": "D^{l} = \\{x_{i}^{l},y_{i}^{l}\\}" + }, + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "text", + "content": ", and unlabeled ones, " + }, + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "inline_equation", + "content": "D^{u} = \\{x_{j}^{u}\\}" + }, + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "text", + "content": ". Typically, " + }, + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "inline_equation", + "content": "D^{u}" + }, + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "text", + "content": " greatly outnumbers " + }, + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "inline_equation", + "content": "D^{l}" + }, + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "text", + "content": ". Our goal is to leverage these unlabeled images to optimize the parameters " + }, + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "text", + "content": " of a semantic segmentation model " + }, + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 55, + 463, + 296, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "text", + "content": "In SWSEG, we adopt a typical framework from Fix-Match [34], training the model through two separate feedforward streams: supervised and unsupervised. In the supervised stream, we draw a batch of images from " + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "inline_equation", + "content": "D^{l}" + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "text", + "content": " and compute the cross-entropy loss " + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "text", + "content": " between the model's predictions " + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^l" + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "text", + "content": " and the labels " + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "inline_equation", + "content": "y_i^l" + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "text", + "content": ". Conversely, the unsupervised stream draws images from " + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "inline_equation", + "content": "D^{u}" + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "text", + "content": ", to which we apply weak or strong perturbations, " + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^w(\\cdot)" + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^s(\\cdot)" + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "text", + "content": ", respectively. The unsupervised loss " + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 55, + 525, + 296, + 644 + ], + "type": "text", + "content": " is then minimized according to the equation (1):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 652, + 296, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 652, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 66, + 652, + 296, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u} = \\frac {1}{| \\mathcal {D} ^ {u} |} \\sum_ {x \\in \\mathcal {D} ^ {u}} \\mathcal {H} \\left(f _ {\\theta} \\left(\\mathcal {A} ^ {s} \\left(x ^ {u}\\right)\\right), f _ {\\theta} \\left(\\mathcal {A} ^ {w} \\left(x ^ {u}\\right)\\right)\\right), \\tag {1}", + "image_path": "18aed766cb6728104d8c286aaa1c1991f9fc2fae8b35329a4a9c1fb3a0b210e7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "We generate pseudo-labels " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathcal{A}^{w}(x^{u}))" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " by feeding weakly-augmented images into the model. We then compute" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 378, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 415 + ], + "type": "text", + "content": "the unsupervised loss using strongly-augmented samples " + }, + { + "bbox": [ + 313, + 378, + 555, + 415 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathcal{A}^{s}(x^{u}))" + }, + { + "bbox": [ + 313, + 378, + 555, + 415 + ], + "type": "text", + "content": ", using " + }, + { + "bbox": [ + 313, + 378, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 378, + 555, + 415 + ], + "type": "text", + "content": ", for cross-entropy loss. The pseudolabel generation does not contribute to gradient computation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 422, + 500, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 500, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 500, + 436 + ], + "type": "text", + "content": "3.2. Feature Alignment and Uniformity" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 440, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 555, + 559 + ], + "type": "text", + "content": "In semantic segmentation, " + }, + { + "bbox": [ + 313, + 440, + 555, + 559 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 440, + 555, + 559 + ], + "type": "text", + "content": " typically comprises a backbone encoder " + }, + { + "bbox": [ + 313, + 440, + 555, + 559 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 313, + 440, + 555, + 559 + ], + "type": "text", + "content": " and a segmentation head " + }, + { + "bbox": [ + 313, + 440, + 555, + 559 + ], + "type": "inline_equation", + "content": "q_{\\theta}" + }, + { + "bbox": [ + 313, + 440, + 555, + 559 + ], + "type": "text", + "content": " [4, 43]. Existing methods [23, 27, 37, 46, 50] focus on enforcing consistency regularization on the segmentation head's output, often overlooking backbone encoder features. To further harness the potential of feature representations, we draw on self-supervised contrastive learning, which trains networks to generate meaningful embeddings without labels, emphasizing alignment and uniformity as key objectives for assessing feature quality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 560, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 555, + 643 + ], + "type": "text", + "content": "Alignment aims to minimize differences between embeddings from similar samples. Intuitively, model should produce similar feature embeddings when images with same classes are input. We achieve this by varying the magnitude of perturbations applied to the input images and minimizing the distance between embeddings using a distance metric dist:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 654, + 555, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 654, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 341, + 654, + 555, + 668 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = \\operatorname {d i s t} \\left(p _ {\\theta} \\left(\\mathcal {A} ^ {s} \\left(x _ {i} ^ {u}\\right)\\right), p _ {\\theta} \\left(\\mathcal {A} ^ {w} \\left(x _ {i} ^ {u}\\right)\\right)\\right). \\tag {2}", + "image_path": "f4c46aa0554b4378304f6677dbe58b7186e223bb658ef1073e60c4bde132d068.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "text", + "content": "Uniformity promotes even distributions of feature representations across a hypersphere, enhancing model robustness by preventing collapse to a lower-dimensional subspace." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20236" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "While often less emphasized in semi-supervised learning due to the stabilizing influence of labeled data, uniformity is critical in self-supervised learning to prevent the model from generating identical embeddings regardless of input. By ensuring uniformity, the model preserves a rich diversity of information within the feature space, which significantly improves the performance of the segmentation head." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 163, + 257, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 257, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 257, + 176 + ], + "type": "text", + "content": "3.3. Revisiting Sliced Wasserstein Distance" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 181, + 296, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 181, + 296, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 296, + 265 + ], + "type": "text", + "content": "Intuitively, the uniform distribution on a unit hypersphere exemplifies maximum uniformity. We need to establish an objective function that can achieve both uniformity and alignment. To this end, we implement the SWD as our distance metric. SWD calculates average distances from 1-D projections sampled from a unit hypersphere, thus inherently promoting both desired properties." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 265, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 350 + ], + "type": "text", + "content": "The rationale behind SWD stems from its ability to reduce the computational complexity relative to the high-dimensional Wasserstein distances (WD). Unlike the traditional WD, which struggles with high dimensionality, SWD leverages the 1-dimensional Wasserstein distance, which can be computed in closed form: Let " + }, + { + "bbox": [ + 55, + 265, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mu_d" + }, + { + "bbox": [ + 55, + 265, + 296, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 265, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\nu_d" + }, + { + "bbox": [ + 55, + 265, + 296, + 350 + ], + "type": "text", + "content": " be two probability measures in " + }, + { + "bbox": [ + 55, + 265, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\Omega) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 265, + 296, + 350 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 358, + 296, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 358, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 65, + 358, + 296, + 389 + ], + "type": "interline_equation", + "content": "W _ {p} ^ {p} \\left(\\mu_ {d}, \\nu_ {d}\\right) = \\left(\\int_ {0} ^ {1} \\left| F _ {\\mu_ {d}} ^ {- 1} (z) - F _ {\\nu_ {d}} ^ {- 1} (z) \\right| ^ {p} d z\\right) ^ {1 / p}, \\tag {3}", + "image_path": "b5baebc15fb754b7d3fbc35759d3c3ea388489eabe34585746bdbedf983bb219.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": " is the order of the distance, and " + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "inline_equation", + "content": "F_{\\mu_d}^{-1}" + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "inline_equation", + "content": "F_{\\nu_d}^{-1}" + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": " are the quantile functions of " + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mu_d" + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\nu_d" + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": ", respectively. Computationally, this is achieved by sorting and computing the mean " + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "inline_equation", + "content": "l^2" + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": "-distance between two 1-D vectors. SWD [30] involves projecting high-dimensional features onto a unit hypersphere and calculating the 1-D Wasserstein distance between these projections. Formally, let " + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^{d-1}" + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": " represent the unit sphere in " + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": " dimensions, " + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": " the uniform distribution on " + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^{d-1}" + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "inline_equation", + "content": "P_{\\theta\\#}\\xi" + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": " the push-forward measure of " + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\xi \\in \\mu_d, \\nu_d" + }, + { + "bbox": [ + 55, + 396, + 296, + 517 + ], + "type": "text", + "content": ". The SWD is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 524, + 296, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 524, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 66, + 524, + 296, + 550 + ], + "type": "interline_equation", + "content": "S W _ {p} ^ {p} \\left(\\mu_ {d}, \\nu_ {d}\\right) = \\int_ {\\mathbb {S} ^ {d - 1}} W _ {p} ^ {p} \\left(P _ {\\theta \\#} \\mu_ {d}, P _ {\\theta \\#} \\nu_ {d}\\right) \\mathrm {d} \\sigma (\\theta). \\tag {4}", + "image_path": "5da0bfb2b7769621f7823ca22da2a8b33489af913aa1584c7329a39ca0223ebd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "text", + "content": "To approximate SWD, we vectorize the input distributions " + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mu_{d}" + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\nu_{d}" + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n\\times n\\times d}" + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n^2 d}" + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "text", + "content": ". We then employ a Monte Carlo scheme, sampling " + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "text", + "content": " directions from the unit sphere " + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^{d - 1}" + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "text", + "content": ". Projecting " + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mu_{d}" + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\nu_{d}" + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "text", + "content": " along these directions yields " + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 556, + 296, + 677 + ], + "type": "text", + "content": " one-dimensional projections for each distribution. After sorting these projections, we calculate the 1D Wasserstein distance between them using Eq. (3). This method efficiently estimates the Wasserstein distance between feature embeddings, exploiting the closed form property of 1D calculations compared to high-dimensional alternatives." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Although the unit sphere projection enables uniformity and minimizing 1-D Wasserstein distances facilitates alignment, training with Monte Carlo-approximated SWD" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": "presents several challenges. First, the results are inherently non-deterministic and intractable due to the randomness inherent in the Monte Carlo method. Second, as the dimensions of " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "\\mu_d" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "\\nu_d" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " increase, both the computational cost and memory requirements for SWD escalate. Finally, to maintain accuracy in the face of increasing dimensions, the number of directions " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " must also increase, complicating the Monte Carlo approximation further." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 175, + 471, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 471, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 471, + 186 + ], + "type": "text", + "content": "3.4. Proposed Approach: SWSEG" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "content": "To enhance feature representation uniformity and alignment while alleviating the computational burden of Monte Carlo-based SWD, we implement an alternate method proposed by Nadjahi et al. [28] for approximating the SWD. Our method diverges by projecting these measures directly onto a Gaussian distribution. This modification offers significant computational benefits and supports key theoretical objectives in two primary ways: First, projecting onto Gaussian distributions allows the 2-Wasserstein distance between any two Gaussians to be expressed in a closed-form equation. Second, using normalized Gaussian variables ensures that the resulting distributions are uniformly distributed over the unit hypersphere, which we prove in the supplementary materials. This uniform distribution is essential for maintaining the model's uniformity objective, ensuring features are evenly distributed in the embedding space, maximizing its information for segmentation head." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "text", + "content": "To formalize this idea, for any positive integer " + }, + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\{\\theta_i\\}_{i=1}^d" + }, + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "text", + "content": " be a sequence of i.i.d. 1-d standard Gaussian distributions and " + }, + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\{X_i\\}_{i=1}^d" + }, + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "text", + "content": " be a sequence of 1-d probability distributions. Several Central Limit Theorems [11, 31] ensure that the sequence of distributions formed by the inner product " + }, + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "inline_equation", + "content": "d^{-1/2}\\langle \\theta_{1:d}, X_{1:d} \\rangle" + }, + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "text", + "content": " converges in probability to a Gaussian random variable when the distributions exhibit weak correlation. Furthermore, according to Proposition 2 from [28], as the dimensions of the distributions increase, the quadratic Wasserstein distance between two zero-mean Gaussian distributions with variance " + }, + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "inline_equation", + "content": "d^{-1}m_2(\\bar{\\mu}_d)" + }, + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "text", + "content": " can be approximated by the SW distance between two Gaussian projected distributions. The Wasserstein distance between two Gaussian probability distribution is in closed form [10], which enables us to approximate the " + }, + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "inline_equation", + "content": "l^2" + }, + { + "bbox": [ + 313, + 395, + 556, + 575 + ], + "type": "text", + "content": "-SW distance by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 579, + 555, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 579, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 315, + 579, + 555, + 606 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s w d} = d ^ {- 1} \\left(m _ {2} \\left(\\bar {\\mu_ {d}}\\right) ^ {\\frac {1}{2}} - m _ {2} \\left(\\bar {\\nu_ {d}}\\right) ^ {\\frac {1}{2}}\\right) + d ^ {- 1} \\left| m _ {\\mu_ {d}} - m _ {\\nu_ {d}} \\right| _ {2} ^ {2}. \\tag {5}", + "image_path": "c49199f375095d6e281b678582a2dbefa701e3f69b9beb1a940d4e4250ec6a9b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "This equation comprises two terms: the first quantifies the difference in the second moments of the distributions, and the second captures the squared error in their means. For " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\xi_d \\in \\{\\bar{\\mu}_d, \\bar{\\nu}_d\\}" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "m_2(\\xi_d) = n^{-1} \\sum_{j=1}^n ||x^j||^2" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": ", enabling a closed-form deterministic solution for " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{swd}" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": ". Nadjahi et al. also proved in Proposition 1 that " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{swd}" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": " with Gaussian projections equals the original " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "SW_p^p" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "p = 2" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": ". Thus, we can use " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{swd}" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": " as a computationally efficient Wasserstein metric for our consistency loss term. A PyTorch-styled" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20237" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 101, + 511, + 216 + ], + "blocks": [ + { + "bbox": [ + 55, + 75, + 555, + 97 + ], + "lines": [ + { + "bbox": [ + 55, + 75, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 75, + 555, + 97 + ], + "type": "text", + "content": "Table 1. Results on ADE20k. The fractions represent the labeled image proportions sampled during training. The model is trained on SegFormer [43] with MiT-B1 backbone. We use mIOU as our evaluation metric." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 101, + 511, + 216 + ], + "lines": [ + { + "bbox": [ + 100, + 101, + 511, + 216 + ], + "spans": [ + { + "bbox": [ + 100, + 101, + 511, + 216 + ], + "type": "table", + "html": "
Method1/32 (631)1/16 (1263)1/8 (2526)1/4 (5052)1/2 (10105)
Supervised Only15.222.325.929.532.6
CPS [8] [ICLR '21]18.822.327.932.436.9
U²PL [42] [CVPR '22]18.523.628.133.536.0
ReCo [22] [ICLR '22]21.123.727.730.235.5
GTA-Seg [19] [NeurIPS '22]19.523.126.927.829.9
DT [27] [NeurIPS '23]22.626.030.633.737.3
SWSEG (Ours)24.426.730.934.638.0
", + "image_path": "951ce2d6bcca966a7d0f1e9761d5111e4c0e7df916fc04d66c684c4876b88514.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 235, + 296, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 235, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 54, + 235, + 296, + 318 + ], + "type": "text", + "content": "pseudo code is presented in Algorithm 1. Furthermore, to ensure a more accurate approximation of the Gaussian-SWD and comply with the relaxed Central Limit Theorems in [11, 31], minimizing strong correlations among input embeddings is essential. We address this by introducing a regularization term based on the input cross-covariance matrix into the SWD loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 325, + 296, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 325, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 57, + 325, + 296, + 350 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e g} = | \\operatorname {C o v} (X) - \\operatorname {D i a g} (X) | _ {2} ^ {2} + | \\operatorname {C o v} (Y) - \\operatorname {D i a g} (Y) | _ {2} ^ {2}. \\tag {6}", + "image_path": "b90ed801f23cb72db13da5ed52a839e33c7131a66bded3ea934fae0e1ca78aff.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 350, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 410 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 350, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\mathrm{Cov}(\\cdot)" + }, + { + "bbox": [ + 55, + 350, + 296, + 410 + ], + "type": "text", + "content": " calculates the covariance matrix of the input features, and " + }, + { + "bbox": [ + 55, + 350, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\mathrm{Diag}(\\cdot)" + }, + { + "bbox": [ + 55, + 350, + 296, + 410 + ], + "type": "text", + "content": " extracts its diagonal elements. Minimizing the off-diagonal elements decorrelates the features, thereby indirectly promoting the uniformity of the distribution of the features across the representation space [49]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 411, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 601 + ], + "type": "text", + "content": "We present a detailed overview of this training scheme in Fig. 3. We also insert an additional non-linear MLP projection layer before loss computation between representations, following recent self-supervised learning work [6, 7, 16]. The projection layer is solely incorporated during the training process, and it is subsequently discarded when the model is deployed for inference. Our objective is to retain more information in the features prior to the MLP layer. This will enable the segmentation head to utilize the preserved information more effectively for the segmentation task. The specific architecture of the projection layer used in this work consists of two Conv - BatchNorm - ReLu blocks followed by a single Conv layer. The overall training objective, managing the balance of various losses, is summarized in Eq. (7). The hyperparameters " + }, + { + "bbox": [ + 55, + 411, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2" + }, + { + "bbox": [ + 55, + 411, + 296, + 601 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 411, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 55, + 411, + 296, + 601 + ], + "type": "text", + "content": " control the weight of each loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 622, + 296, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 622, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 83, + 622, + 296, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o v e r a l l}} = \\mathcal {L} _ {s} + \\lambda_ {1} \\mathcal {L} _ {u} + \\lambda_ {2} \\mathcal {L} _ {s w d} + \\lambda_ {3} \\mathcal {L} _ {\\text {r e g}}. \\tag {7}", + "image_path": "79125622238352848bae5e0fe4e5e77b04f72a634a54ec73c9fe512c6ab2d05b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 652, + 127, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 127, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 127, + 664 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 118, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 118, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 118, + 683 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "ADE20K. [51] is a complex scene parsing dataset featuring 150 objects, with 25,574 training, 2,000 validation, and 3,000" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 315, + 264, + 553, + 434 + ], + "blocks": [ + { + "bbox": [ + 313, + 234, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 257 + ], + "type": "text", + "content": "Algorithm 1 Pseudocode of our Gaussian-SWD loss in a PyTorch-like style." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 315, + 264, + 553, + 434 + ], + "lines": [ + { + "bbox": [ + 315, + 264, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 315, + 264, + 553, + 434 + ], + "type": "text", + "content": "def gaussian_swd_loss(x, y):\n # Reshape the input tensors to (n, dim)\n x, y = x.view(x.shape[0], -1), y.view(y.shape[0], -1)\n n, dim = x.shape\n meanx = torch.mean(x, dim=0)\n xc = x - meanx # Zero mean\n m2x = torch.mean(torch.linalg(norm(xc, dim=1)) * 2) / dim\n meany = torch.mean(y, dim=0)\n yc = y - meany # Zero mean\n m2y = torch.mean(torch.linalg(norm(yc, dim=1)) * 2) / dim\n m_t = torch.linalg(norm(meanx - meany) * 2 / dim\n swd = m_t + (m2x * (1/2) - m2y * (1/2)) * 2\n return swd" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 313, + 456, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 480 + ], + "type": "text", + "content": "testing images. For a fair comparison, we used the data split from dual-teacher [27]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "text", + "content": "PASCAL VOC 2012. [12] This dataset is a classic semantic segmentation dataset comprising 20 classes of objects and 1 background class, with 1464 training, 1449 validation, and 1546 testing high-quality images in the standard set. We incorporate the augmented set from [17], following previous works, increasing our labeled data to 10,582 images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 554, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 555, + 602 + ], + "type": "text", + "content": "Cityscapes. [9] This is an autonomous driving semantic segmentation dataset that consists of 19 semantic classes, with 2975, 500, and 1525 training, validation, and testing images, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 612, + 448, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 448, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 448, + 624 + ], + "type": "text", + "content": "4.2. Implementation details." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": "For ADE20K, we implemented our method with SegFormer [43] architecture with MiT-B1 backbone. For PASCAL VOC 2012 and CityScapes, to have a fair comparison with previous benchmark semi-supervised semantic segmentation methods [27, 37, 42, 45, 46, 50], we implement our method using ResNet-101 [18] as the backbone encoder, initialized with ImageNet pre-trained weights, and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20238" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 112, + 287, + 262 + ], + "blocks": [ + { + "bbox": [ + 55, + 75, + 297, + 109 + ], + "lines": [ + { + "bbox": [ + 55, + 75, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 75, + 297, + 109 + ], + "type": "text", + "content": "Table 2. Results on CityScapes dataset. The fractions represent the proportion of labeled images used during training. We use mIOU as our evaluation metric." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 112, + 287, + 262 + ], + "lines": [ + { + "bbox": [ + 66, + 112, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 66, + 112, + 287, + 262 + ], + "type": "table", + "html": "
Method1/161/81/4
Supervised Only67.973.275.1
PCR [44] [NeurIPS '22]73.476.378.4
UniMatch [46] [CVPR '23]76.677.979.2
AugSeg [50] [CVPR '23]75.277.879.6
DT [27] [NeurIPS '23]76.878.479.5
DAW [37] [NeurIPS '23]76.678.479.8
CorrMatch [36] [CVPR '24]77.378.579.4
RankMatch [24] [CVPR '24]77.178.680.0
DDFP [41] [CVPR '24]77.178.279.9
SWSEG (Ours)76.878.679.5
", + "image_path": "3c8b59238dc1adbc7a4288d9af9d941bdcec621d2dbdb927835a680598be8dcf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 305, + 287, + 434 + ], + "blocks": [ + { + "bbox": [ + 55, + 269, + 296, + 302 + ], + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 302 + ], + "type": "text", + "content": "Table 3. Results on PASCAL VOC 2012 dataset. The fractions represent the proportion of labeled images used during training. We use mIOU as our evaluation metric." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 305, + 287, + 434 + ], + "lines": [ + { + "bbox": [ + 66, + 305, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 66, + 305, + 287, + 434 + ], + "type": "table", + "html": "
Method1/161/81/4
Supervised Only67.271.373.5
UniMatch [46] [CVPR '23]75.076.877.5
AugSeg [50] [CVPR '23]77.077.378.8
DAW [37] [NeurIPS '23]78.578.979.6
CorrMatch [36] [CVPR '24]78.479.379.6
RankMatch [24] [CVPR '24]78.979.280.0
DDFP [41] [CVPR '24]78.378.979.8
SWSEG (Ours)79.079.379.9
", + "image_path": "fe236b55bddd7a61e3215bb42db06ce9df9eee17f81d54a6db8a36d70a011d1d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 297, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 297, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 297, + 529 + ], + "type": "text", + "content": "DeeplabV3+ [5] as the decoder. We use SGD optimizer with momentum of 0.9. The learning rate is set to 0.001, 0.02, and 5e-5 for PASCAL, CityScapes, and ADE20K, respectively. We utilize a polynomial scheduling strategy with 0.9 decay rate." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 536, + 297, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 297, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 297, + 646 + ], + "type": "text", + "content": "Data augmentations. For weak augmentations, horizontal flipping, random cropping, and random scaling are applied to the training images. For strong augmentations, we apply random color jittering, random grayscale, random blurring, and Cutmix [47] to the input training images. The training image size is set as " + }, + { + "bbox": [ + 55, + 536, + 297, + 646 + ], + "type": "inline_equation", + "content": "513 \\times 513" + }, + { + "bbox": [ + 55, + 536, + 297, + 646 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 536, + 297, + 646 + ], + "type": "inline_equation", + "content": "769 \\times 769" + }, + { + "bbox": [ + 55, + 536, + 297, + 646 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 536, + 297, + 646 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 536, + 297, + 646 + ], + "type": "text", + "content": " for PASCAL VOC, CityScapes, and ADE20K, respectively. The training epochs for VOC, CityScapes, and ADE20K is 80, 240, and 200 epochs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "content": "Evaluation. As in previous works, we employ mean Intersection-over-Union (mIOU) as our evaluation metric. For the PASCAL VOC 2012 and ADE20K, evaluations are on the center-cropped validation set, while for CityScapes, we employ a sliding window evaluation." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 134, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 313, + 75, + 555, + 131 + ], + "lines": [ + { + "bbox": [ + 313, + 75, + 555, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 75, + 555, + 131 + ], + "type": "text", + "content": "Table 4. Ablation analysis of feature alignment loss functions, " + }, + { + "bbox": [ + 313, + 75, + 555, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 313, + 75, + 555, + 131 + ], + "type": "text", + "content": ", and MLP layer. The results reveal that MSE, cross entropy, and Monte Carlo estimated SWD loss with " + }, + { + "bbox": [ + 313, + 75, + 555, + 131 + ], + "type": "inline_equation", + "content": "L = 128" + }, + { + "bbox": [ + 313, + 75, + 555, + 131 + ], + "type": "text", + "content": " yield inferior results. The combination of Gaussian-SWD (GSWD) with " + }, + { + "bbox": [ + 313, + 75, + 555, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 313, + 75, + 555, + 131 + ], + "type": "text", + "content": " and MLP produces the best performance, with an mIoU of 79.0." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 134, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 317, + 134, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 134, + 553, + 247 + ], + "type": "table", + "html": "
Loss FunctionsLregMLPmIoU
MSECESWDGSWD
✓✓✓70.2
✓✓✓77.1
✓✓✓77.6
✓77.4
✓✓78.1
✓✓✓79.0
", + "image_path": "55f0cf25f709913130ba86b2259790fe8b91a130bcbac867c323c7d3b8588f23.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 335, + 302, + 535, + 353 + ], + "blocks": [ + { + "bbox": [ + 313, + 255, + 556, + 299 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 556, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 556, + 299 + ], + "type": "text", + "content": "Table 5. Analysis of the Uniformity value " + }, + { + "bbox": [ + 313, + 255, + 556, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{uniform}}" + }, + { + "bbox": [ + 313, + 255, + 556, + 299 + ], + "type": "text", + "content": " across different data splits of the PASCAL VOC dataset demonstrates that training with the SWD loss " + }, + { + "bbox": [ + 313, + 255, + 556, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{swd}" + }, + { + "bbox": [ + 313, + 255, + 556, + 299 + ], + "type": "text", + "content": " yields superior Uniformity values compared to training without this loss term." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 302, + 535, + 353 + ], + "lines": [ + { + "bbox": [ + 335, + 302, + 535, + 353 + ], + "spans": [ + { + "bbox": [ + 335, + 302, + 535, + 353 + ], + "type": "table", + "html": "
Method1/81/21
Without Lswd-2.5896-2.7095-2.5574
With Lswd-2.6592-2.7104-2.6168
", + "image_path": "47e7fc9183529c562b574c92a315ff75e23164a0521f7f3a9ee06a3f2f386a13.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 372, + 501, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 501, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 501, + 385 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 389, + 556, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 556, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 556, + 484 + ], + "type": "text", + "content": "Results on ADE20K Dataset. Tab. 1 demonstrates the performance of SWSEG against the supervised baseline and existing methods. SWSEG significantly improves over the supervised baseline by " + }, + { + "bbox": [ + 313, + 389, + 556, + 484 + ], + "type": "inline_equation", + "content": "8.2\\%" + }, + { + "bbox": [ + 313, + 389, + 556, + 484 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 389, + 556, + 484 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 313, + 389, + 556, + 484 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 389, + 556, + 484 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 313, + 389, + 556, + 484 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 389, + 556, + 484 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 313, + 389, + 556, + 484 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 389, + 556, + 484 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 313, + 389, + 556, + 484 + ], + "type": "text", + "content": " mIoUs on the 1/32, 1/16, 1/8, 1/4, and 1/2 splits, respectively. Our method also outperforms the state-of-the-art across all data splits, showcasing the generalizability of SWSEG across different model architectures and datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 556, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 570 + ], + "type": "text", + "content": "Results on Cityscapes Dataset. Tab. 2 presents our results in comparison to the supervised baseline and SOTA methods. SWSEG outperforms the supervised baseline by " + }, + { + "bbox": [ + 313, + 486, + 556, + 570 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 313, + 486, + 556, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 486, + 556, + 570 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 313, + 486, + 556, + 570 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 486, + 556, + 570 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 313, + 486, + 556, + 570 + ], + "type": "text", + "content": " for the 1/16, 1/8, and 1/4 partitions, respectively. Additionally, SWSEG surpasses the previous state-of-the-art method, DAW, by " + }, + { + "bbox": [ + 313, + 486, + 556, + 570 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 313, + 486, + 556, + 570 + ], + "type": "text", + "content": " in the 1/16 and 1/8 partitions, and performs on par with contemporary methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "text", + "content": "Results on PASCAL VOC 2012 Dataset. Tab. 3 compares SWSEG with existing state-of-the-arts on PASCAL VOC 2012 set. We follow the data partitions in CPS [8]. We outperform the supervised baselines by " + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "inline_equation", + "content": "11.8\\%" + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "inline_equation", + "content": "8.0\\%" + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "text", + "content": " mIoUs under 1/16, 1/8, and 1/4 partitioning protocols. Our method also beats the SOTA DDFP [37] by " + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 313, + 570, + 557, + 713 + ], + "type": "text", + "content": " mIoUs under 1/16, 1/8, and 1/4 splits, respectively. We also visualized the predicted features in Fig. 4 to evaluate the effectiveness of our algorithm's clustering capabilities. Our method demonstrates enhanced discriminative clustering compared to the Supervised Only and UniMatch methods." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20239" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 72, + 132, + 129 + ], + "blocks": [ + { + "bbox": [ + 75, + 72, + 132, + 129 + ], + "lines": [ + { + "bbox": [ + 75, + 72, + 132, + 129 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 132, + 129 + ], + "type": "image", + "image_path": "2bcc5edd985fb3aefa80d72e8b9e171aeeb3194b10f367119670c9da78c2923a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 130, + 108, + 138 + ], + "lines": [ + { + "bbox": [ + 96, + 130, + 108, + 138 + ], + "spans": [ + { + "bbox": [ + 96, + 130, + 108, + 138 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 146, + 72, + 205, + 129 + ], + "blocks": [ + { + "bbox": [ + 146, + 72, + 205, + 129 + ], + "lines": [ + { + "bbox": [ + 146, + 72, + 205, + 129 + ], + "spans": [ + { + "bbox": [ + 146, + 72, + 205, + 129 + ], + "type": "image", + "image_path": "dd06d584a69794e256f67beefe020733ea72ba308d5cf8662645803d4b0bc696.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 130, + 181, + 139 + ], + "lines": [ + { + "bbox": [ + 171, + 130, + 181, + 139 + ], + "spans": [ + { + "bbox": [ + 171, + 130, + 181, + 139 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 219, + 72, + 277, + 129 + ], + "blocks": [ + { + "bbox": [ + 219, + 72, + 277, + 129 + ], + "lines": [ + { + "bbox": [ + 219, + 72, + 277, + 129 + ], + "spans": [ + { + "bbox": [ + 219, + 72, + 277, + 129 + ], + "type": "image", + "image_path": "76af2cb30ff0c58380c0b74fe2fd4002beaf5cf1c78f08d1db7a3afedfa25dd4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 130, + 252, + 139 + ], + "lines": [ + { + "bbox": [ + 243, + 130, + 252, + 139 + ], + "spans": [ + { + "bbox": [ + 243, + 130, + 252, + 139 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 209, + 174, + 365 + ], + "blocks": [ + { + "bbox": [ + 55, + 146, + 296, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 146, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 296, + 201 + ], + "type": "text", + "content": "Figure 4. UMAP visualization [25] of feature embeddings using models trained with: (a) Supervised Only, (b) UniMatch, and (c) SWSEG. These visualizations utilize the PASCAL VOC 2012 dataset. Our method demonstrates enhanced clustering capabilities, outperforming both the Supervised Only and UniMatch." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 209, + 174, + 365 + ], + "lines": [ + { + "bbox": [ + 58, + 209, + 174, + 365 + ], + "spans": [ + { + "bbox": [ + 58, + 209, + 174, + 365 + ], + "type": "image", + "image_path": "97d2962bbd79f7d564eabbe26f05508d9c01186fe8d3c61bd8a13d0476167de7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 368, + 145, + 377 + ], + "lines": [ + { + "bbox": [ + 86, + 368, + 145, + 377 + ], + "spans": [ + { + "bbox": [ + 86, + 368, + 145, + 377 + ], + "type": "text", + "content": "(a) Without " + }, + { + "bbox": [ + 86, + 368, + 145, + 377 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{swd}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 386, + 296, + 431 + ], + "lines": [ + { + "bbox": [ + 55, + 386, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 296, + 431 + ], + "type": "text", + "content": "Figure 5. We visualize feature representations on a unit circle for models trained with and without " + }, + { + "bbox": [ + 55, + 386, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{swd}" + }, + { + "bbox": [ + 55, + 386, + 296, + 431 + ], + "type": "text", + "content": " for uniformity analysis. Features from the model with " + }, + { + "bbox": [ + 55, + 386, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{swd}" + }, + { + "bbox": [ + 55, + 386, + 296, + 431 + ], + "type": "text", + "content": " (b) are more uniformly distributed than those without " + }, + { + "bbox": [ + 55, + 386, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{swd}" + }, + { + "bbox": [ + 55, + 386, + 296, + 431 + ], + "type": "text", + "content": " (a)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 179, + 209, + 294, + 365 + ], + "blocks": [ + { + "bbox": [ + 179, + 209, + 294, + 365 + ], + "lines": [ + { + "bbox": [ + 179, + 209, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 179, + 209, + 294, + 365 + ], + "type": "image", + "image_path": "14e7a1bf8bb5db89879cfcc0473cc58d49804bc28bb400328dc30a189120cfd2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 368, + 260, + 377 + ], + "lines": [ + { + "bbox": [ + 211, + 368, + 260, + 377 + ], + "spans": [ + { + "bbox": [ + 211, + 368, + 260, + 377 + ], + "type": "text", + "content": "(b) With " + }, + { + "bbox": [ + 211, + 368, + 260, + 377 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{swd}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 456, + 149, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 149, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 149, + 469 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "text", + "content": "Uniformity and Alignment. We conduct an empirical analysis of feature representation alignment and uniformity (Tab. 4). We compare different loss functions for enforcing consistency regularization on the feature representation. Our results indicate that, aside from MSE loss, additional loss terms generally enhance segmentation prediction performance, highlighting the critical role of feature alignment. Monte Carlo SWD with " + }, + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "inline_equation", + "content": "L = 128" + }, + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "text", + "content": " outperforms MSE and cross-entropy but encounters out-of-memory issues at " + }, + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "inline_equation", + "content": "L = 1000" + }, + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "text", + "content": " using an identical number of GPUs, highlighting its computational cost. Conversely, our Gaussian-based SWD approximation delivers optimal results with lower memory usage. For uniformity analysis, we reduce the dimension of feature embeddings using PCA [15], projecting them into " + }, + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "text", + "content": " and plotting their coordinates on a unit circle, as shown in Fig. 5. The visualization demonstrates that models trained with " + }, + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{swd}" + }, + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "text", + "content": " achieve greater uniformity compared to those without it, confirming the effectiveness of our introduced SWD loss. To quantitatively validate this observation, we compute the uniformity metric " + }, + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{uniform}}" + }, + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "text", + "content": " from [40, Equa" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 362, + 113 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 362, + 113 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 362, + 113 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 362, + 113 + ], + "type": "image", + "image_path": "25dc148c14a869e9ea05a85f48e481ba8c75618295ab9660bbc703ffc26c921f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 364, + 70, + 410, + 112 + ], + "blocks": [ + { + "bbox": [ + 364, + 70, + 410, + 112 + ], + "lines": [ + { + "bbox": [ + 364, + 70, + 410, + 112 + ], + "spans": [ + { + "bbox": [ + 364, + 70, + 410, + 112 + ], + "type": "image", + "image_path": "f3b7ee4544be51e46676b3ef4518189053923ec1551fa175a98b75608c252345.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 364, + 113, + 410, + 155 + ], + "blocks": [ + { + "bbox": [ + 364, + 113, + 410, + 155 + ], + "lines": [ + { + "bbox": [ + 364, + 113, + 410, + 155 + ], + "spans": [ + { + "bbox": [ + 364, + 113, + 410, + 155 + ], + "type": "image", + "image_path": "2b32a8bd06fe0edd0414e9fb70129c6e10ee0964f69ce01b01272acdbd8c0033.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 364, + 156, + 410, + 241 + ], + "blocks": [ + { + "bbox": [ + 364, + 156, + 410, + 241 + ], + "lines": [ + { + "bbox": [ + 364, + 156, + 410, + 241 + ], + "spans": [ + { + "bbox": [ + 364, + 156, + 410, + 241 + ], + "type": "image", + "image_path": "1f2f8303d9a8d80399ee63bb7a0ca8e31c653bf468d363def389bd91d73544b9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 241, + 391, + 246 + ], + "lines": [ + { + "bbox": [ + 364, + 241, + 391, + 246 + ], + "spans": [ + { + "bbox": [ + 364, + 241, + 391, + 246 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 251, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 319 + ], + "type": "text", + "content": "Figure 6. Qualitative results on PASCAL VOC 2012 validation set. All models are trained under 1/4 partition. (a) Input images. (b) Ground Truth labels. (c) Model output when trained with only labeled images. (d) Model output when training without SWD loss. (e) Predictions from SWSEG. When including SWD loss, the predictions are better in ambiguous regions and object borders." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 411, + 70, + 457, + 113 + ], + "blocks": [ + { + "bbox": [ + 411, + 70, + 457, + 113 + ], + "lines": [ + { + "bbox": [ + 411, + 70, + 457, + 113 + ], + "spans": [ + { + "bbox": [ + 411, + 70, + 457, + 113 + ], + "type": "image", + "image_path": "c8d6ebeb99057817d16cdfde725180479b9a498e2fb1f138d5a16deec45eaa74.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 411, + 113, + 457, + 156 + ], + "blocks": [ + { + "bbox": [ + 411, + 113, + 457, + 156 + ], + "lines": [ + { + "bbox": [ + 411, + 113, + 457, + 156 + ], + "spans": [ + { + "bbox": [ + 411, + 113, + 457, + 156 + ], + "type": "image", + "image_path": "51d22d8443db3fe8e93042ee27abb3ecc681670ce7e4f92062dc16bca18aece2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 411, + 156, + 457, + 198 + ], + "blocks": [ + { + "bbox": [ + 411, + 156, + 457, + 198 + ], + "lines": [ + { + "bbox": [ + 411, + 156, + 457, + 198 + ], + "spans": [ + { + "bbox": [ + 411, + 156, + 457, + 198 + ], + "type": "image", + "image_path": "afedc0bf2e3222106dbca9c402d809adc6925b60a5419bb85d867d48b561206d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 411, + 198, + 457, + 241 + ], + "blocks": [ + { + "bbox": [ + 411, + 198, + 457, + 241 + ], + "lines": [ + { + "bbox": [ + 411, + 198, + 457, + 241 + ], + "spans": [ + { + "bbox": [ + 411, + 198, + 457, + 241 + ], + "type": "image", + "image_path": "fb590bec39f2b70943f64c5dcb2ac2ee822a43badc4a28db20bf28f58ee49b2f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 241, + 439, + 246 + ], + "lines": [ + { + "bbox": [ + 411, + 241, + 439, + 246 + ], + "spans": [ + { + "bbox": [ + 411, + 241, + 439, + 246 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 459, + 70, + 505, + 113 + ], + "blocks": [ + { + "bbox": [ + 459, + 70, + 505, + 113 + ], + "lines": [ + { + "bbox": [ + 459, + 70, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 459, + 70, + 505, + 113 + ], + "type": "image", + "image_path": "8deed34be5a158aa0f11ab540e16dd665e8f924da735642a9d3e1f9e51e0bd34.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 459, + 113, + 505, + 156 + ], + "blocks": [ + { + "bbox": [ + 459, + 113, + 505, + 156 + ], + "lines": [ + { + "bbox": [ + 459, + 113, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 459, + 113, + 505, + 156 + ], + "type": "image", + "image_path": "381abb47076c5d9e82c1831b79481d75043d015f8164351aab7fb81202cefbac.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 459, + 156, + 505, + 198 + ], + "blocks": [ + { + "bbox": [ + 459, + 156, + 505, + 198 + ], + "lines": [ + { + "bbox": [ + 459, + 156, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 459, + 156, + 505, + 198 + ], + "type": "image", + "image_path": "3e7865b2343e42292a4c971b63f11acbabdd87d48fbbaad421a82d1e6f79f37a.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 459, + 198, + 505, + 241 + ], + "blocks": [ + { + "bbox": [ + 459, + 198, + 505, + 241 + ], + "lines": [ + { + "bbox": [ + 459, + 198, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 459, + 198, + 505, + 241 + ], + "type": "image", + "image_path": "f4ca899082d2907cc60405d4f86794b7eaaa8f54dce06e65c1f67cf2c6d51dc6.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 242, + 487, + 246 + ], + "lines": [ + { + "bbox": [ + 463, + 242, + 487, + 246 + ], + "spans": [ + { + "bbox": [ + 463, + 242, + 487, + 246 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 507, + 70, + 553, + 113 + ], + "blocks": [ + { + "bbox": [ + 507, + 70, + 553, + 113 + ], + "lines": [ + { + "bbox": [ + 507, + 70, + 553, + 113 + ], + "spans": [ + { + "bbox": [ + 507, + 70, + 553, + 113 + ], + "type": "image", + "image_path": "8db54dce7a8c91175a5397f7ec71e5786289681d6dff60421a99184c3d2cad29.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 507, + 113, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 507, + 113, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 507, + 113, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 507, + 113, + 553, + 156 + ], + "type": "image", + "image_path": "4dc2fd916967a116d046a2a848d057195f5e25734615d3392fa1d58006179043.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 507, + 156, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 507, + 156, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 507, + 156, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 507, + 156, + 553, + 198 + ], + "type": "image", + "image_path": "94a78afadb45ea89b6193d9630d2e8e3142520044ebc29a1b44226428d49ea0d.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 507, + 198, + 553, + 241 + ], + "blocks": [ + { + "bbox": [ + 507, + 198, + 553, + 241 + ], + "lines": [ + { + "bbox": [ + 507, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 507, + 198, + 553, + 241 + ], + "type": "image", + "image_path": "443007e41488d16f5bd391035ae13ebf1e11b4d24b1fc38643ed16ba14209dc8.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 527, + 242, + 533, + 246 + ], + "lines": [ + { + "bbox": [ + 527, + 242, + 533, + 246 + ], + "spans": [ + { + "bbox": [ + 527, + 242, + 533, + 246 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 343, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 415 + ], + "type": "text", + "content": "tion 5.2] and compare the values for models trained with and without " + }, + { + "bbox": [ + 313, + 343, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{swd}" + }, + { + "bbox": [ + 313, + 343, + 555, + 415 + ], + "type": "text", + "content": " in Tab. 5. The results show that including " + }, + { + "bbox": [ + 313, + 343, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{swd}" + }, + { + "bbox": [ + 313, + 343, + 555, + 415 + ], + "type": "text", + "content": " decreases the uniformity loss, which aligns with our qualitative visualization in Fig. 5. We also conduct an ablation study on MLP layers and " + }, + { + "bbox": [ + 313, + 343, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 313, + 343, + 555, + 415 + ], + "type": "text", + "content": " in Tab. 4, showing that including all components yields the best outcome." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 415, + 556, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 522 + ], + "type": "text", + "content": "Qualitative analysis. Fig. 6 shows the results of different training methods on the PASCAL VOC 2012 val set. When training with SWD loss, our method outperforms other methods, producing segmentation predictions that are closer to ground-truth labels. We can see from the results that our method performs exceptionally well on the border between different classes. Moreover, the segmentation results of our method have a cleaner prediction between the border and the backgrounds." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "We introduced SWSEG, a semi-supervised semantic segmentation algorithm designed to enhance both alignment and uniformity of feature representations. We implement the Sliced-Wasserstein Distance (SWD) as an additional loss to jointly optimize these aspects, using a Gaussian-approximated variant to reduce computational overhead while preserving uniformity. Empirical evaluations on PASCAL VOC 2012, CityScapes, and ADE20K datasets show that SWSEG outperforms both supervised baselines and existing methods. Extensive ablation studies highlight the critical role of uniformity in the feature space for effective semi-supervised semantic segmentation. SWSEG is the first method to link semi-supervised learning with the uniformity metric using SWD, marking a significant advancement in the field." + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20240" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 202 + ], + "type": "text", + "content": "This material is based in part upon work supported by the National Science Foundation under Grant Numbers CNS-2333487 (NSF Frontier award) and CNS-2146449 (NSF CAREER award) and DEVCOM ARL Army Research Office under Contract number W911NF-2020-221. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors. The authors thank the reviewers for their enthusiastic comments and their valuable insights that improved our paper." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 227, + 115, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 227, + 115, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 227, + 115, + 239 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 247, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 247, + 297, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 247, + 297, + 290 + ], + "spans": [ + { + "bbox": [ + 61, + 247, + 297, + 290 + ], + "type": "text", + "content": "[1] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 292, + 296, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 292, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 61, + 292, + 296, + 336 + ], + "type": "text", + "content": "[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 339, + 296, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 339, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 62, + 339, + 296, + 381 + ], + "type": "text", + "content": "[3] Olivier Chapelle and Alexander Zien. Semi-supervised classification by low density separation. In International workshop on artificial intelligence and statistics, pages 57-64. PMLR, 2005. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 385, + 296, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 296, + 440 + ], + "type": "text", + "content": "[4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 441, + 296, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 296, + 496 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 498, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 296, + 542 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 544, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 588 + ], + "type": "text", + "content": "[7] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 590, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 296, + 644 + ], + "type": "text", + "content": "[8] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "type": "text", + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 2, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[10] D.C. Dowson and B.V. Landau. The fréchet distance between multivariate normal distributions. Journal of multivariate analysis, 12(3):450-455, 1982. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 107, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 161 + ], + "type": "text", + "content": "[11] Lutz Dumbgen, Del Conte-Zerial, et al. On low-dimensional projections of high-dimensional distributions. In From Probability to Statistics and Back: High-Dimensional Models and Processes-A Festschrift in Honor of Jon A. Wellner, pages 91-105. Institute of Mathematical Statistics, 2013. 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 162, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 555, + 205 + ], + "type": "text", + "content": "[12] Mark Everingham and John Winn. The pascal visual object classes challenge 2012 (voc2012) development kit. Pattern Anal. Stat. Model. Comput. Learn., Tech. Rep, 2007:1-45, 2012. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 555, + 239 + ], + "type": "text", + "content": "[13] Xianghong Fang, Jian Li, Qiang Sun, and Benyou Wang. Rethinking the uniformity metric in self-supervised learning. arXiv preprint arXiv:2403.00642, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 555, + 282 + ], + "type": "text", + "content": "[14] Geoff French, Samuli Laine, Timo Aila, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, varied perturbations. arXiv preprint arXiv:1906.01916, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 284, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 554, + 327 + ], + "type": "text", + "content": "[15] Karl Pearson F.R.S. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2 (11):559-572, 1901. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 328, + 555, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 328, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 328, + 555, + 414 + ], + "type": "text", + "content": "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In Advances in Neural Information Processing Systems, pages 21271-21284. Curran Associates, Inc., 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 416, + 555, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 555, + 459 + ], + "type": "text", + "content": "[17] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991–998. IEEE, 2011. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 460, + 555, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 555, + 502 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 504, + 555, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 555, + 547 + ], + "type": "text", + "content": "[19] Ying Jin, Jiaqi Wang, and Dahua Lin. Semi-supervised semantic segmentation via gentle teaching assistant. Advances in Neural Information Processing Systems, 35:2803-2816, 2022. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 548, + 555, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 555, + 579 + ], + "type": "text", + "content": "[20] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 581, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 555, + 624 + ], + "type": "text", + "content": "[21] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896, 2013. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "type": "text", + "content": "[22] Shikun Liu, Shuaifeng Zhi, Edward Johns, and Andrew J Davison. Bootstrapping semantic segmentation with regional contrast. arXiv preprint arXiv:2104.04465, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 659, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 555, + 713 + ], + "type": "text", + "content": "[23] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022. 2, 4" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20241" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 128 + ], + "type": "text", + "content": "[24] Huayu Mai, Rui Sun, Tianzhu Zhang, and Feng Wu. Rankmatch: Exploring the better consistency regularization for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3391-3401, 2024. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "type": "text", + "content": "[25] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. Umap: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861, 2018. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 228 + ], + "type": "text", + "content": "[26] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 275 + ], + "type": "text", + "content": "[27] Jaemin Na, Jung-Woo Ha, Hyung Jin Chang, Dongyoon Han, and Wonjun Hwang. Switching temporary teachers for semi-supervised semantic segmentation. Advances in Neural Information Processing Systems, 36, 2023. 2, 4, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 330 + ], + "type": "text", + "content": "[28] Kimia Nadjahi, Alain Durmus, Pierre E Jacob, Roland Badeau, and Umut Simsekli. Fast approximation of the sliced-wasserstein distance using concentration of random projections. Advances in Neural Information Processing Systems, 34:12411-12424, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 296, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 296, + 387 + ], + "type": "text", + "content": "[29] Yassine Ouali, Celine Hudelot, and Myriam Tami. Semi-Supervised Semantic Segmentation With Cross-Consistency Training. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12671–12681, Seattle, WA, USA, 2020. IEEE. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 296, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 454 + ], + "type": "text", + "content": "[30] Julien Rabin, Gabriel Peyré, Julie Delon, and Marc Bernot. Wasserstein barycenter and its application to texture mixing. In Scale Space and Variational Methods in Computer Vision: Third International Conference, SSVM 2011, Ein-Gedi, Israel, May 29-June 2, 2011, Revised Selected Papers 3, pages 435-446. Springer, 2012. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 456, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 498 + ], + "type": "text", + "content": "[31] Galen Reeves. Conditional central limit theorems for gaussian projections. In 2017 IEEE International Symposium on Information Theory (ISIT), pages 3045-3049. IEEE, 2017. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 296, + 544 + ], + "type": "text", + "content": "[32] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 296, + 602 + ], + "type": "text", + "content": "[33] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 296, + 667 + ], + "type": "text", + "content": "[34] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems, 33:596-608, 2020. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "text", + "content": "[35] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 556, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "text", + "content": "Advances in Neural Information Processing Systems, 33:596-608, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "text", + "content": "[36] Boyuan Sun, Yuqi Yang, Le Zhang, Ming-Ming Cheng, and Qibin Hou. Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation. IEEE Computer Vision and Pattern Recognition (CVPR), 2024. 3, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "text", + "content": "[37] Rui Sun, Huayu Mai, Tianzhu Zhang, and Feng Wu. Daw: Exploring the better weighting function for semi-supervised semantic segmentation. Advances in Neural Information Processing Systems, 36, 2023. 3, 4, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 184, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 555, + 228 + ], + "type": "text", + "content": "[38] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "type": "text", + "content": "[39] Haonan Wang, Qixiang Zhang, Yi Li, and Xiaomeng Li. Allspark: Reborn labeled features from unlabeled in transformer for semi-supervised semantic segmentation. arXiv preprint arXiv:2403.01818, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 555, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 555, + 316 + ], + "type": "text", + "content": "[40] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 2, 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 555, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 555, + 380 + ], + "type": "text", + "content": "[41] Xiaoyang Wang, Huihui Bai, Limin Yu, Yao Zhao, and Jimin Xiao. Towards the uncharted: Density-descending feature perturbation for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3303-3312, 2024. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 555, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 555, + 448 + ], + "type": "text", + "content": "[42] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudolabels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4248-4257, 2022. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 556, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 556, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 556, + 503 + ], + "type": "text", + "content": "[43] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34:12077-12090, 2021. 2, 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 504, + 555, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 555, + 548 + ], + "type": "text", + "content": "[44] Haiming Xu, Lingqiao Liu, Qiuchen Bian, and Zhen Yang. Semi-supervised semantic segmentation with prototype-based consistency regularization. Advances in Neural Information Processing Systems, 35:26007-26020, 2022. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 548, + 555, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 555, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 555, + 603 + ], + "type": "text", + "content": "[45] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4268-4277, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 603, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 555, + 658 + ], + "type": "text", + "content": "[46] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7236-7246, 2023. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "text", + "content": "[47] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20242" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 274 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 298, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 298, + 116 + ], + "type": "text", + "content": "[48] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310–12320. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 297, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 297, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 297, + 162 + ], + "type": "text", + "content": "[49] Shaofeng Zhang, Feng Zhu, Junchi Yan, Rui Zhao, and Xiaokang Yang. Zero-cl: Instance and feature decorrelation for negative-free symmetric contrastive learning. In International Conference on Learning Representations, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 228 + ], + "type": "text", + "content": "[50] Zhen Zhao, Lihe Yang, Sifan Long, Jimin Pi, Luping Zhou, and Jingdong Wang. Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11350-11359, 2023. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 296, + 274 + ], + "type": "text", + "content": "[51] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 2, 6" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20243" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/feaa7f56-7770-404f-a305-b468e86df1e5_content_list.json b/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/feaa7f56-7770-404f-a305-b468e86df1e5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d52c86ce7c72784b268e8d74d451b83548a9a052 --- /dev/null +++ b/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/feaa7f56-7770-404f-a305-b468e86df1e5_content_list.json @@ -0,0 +1,1637 @@ +[ + { + "type": "text", + "text": "Improving Sound Source Localization with Joint Slot Attention on Image and Audio", + "text_level": 1, + "bbox": [ + 281, + 130, + 718, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inho Kim1 Youngkil Song2 Jicheol Park2 Won Hwa Kim1,2 Suha Kwak1,2", + "bbox": [ + 165, + 202, + 825, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dept. of CSE, POSTECH $^{2}$ Graduate School of AI, POSTECH", + "bbox": [ + 228, + 223, + 769, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{kimih, songyk, jicheol, wonhwa, suha.kwak}@postech.ac.kr", + "bbox": [ + 241, + 243, + 750, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 294, + 326, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sound source localization (SSL) is the task of locating the source of sound within an image. Due to the lack of localization labels, the de facto standard in SSL has been to represent an image and audio as a single embedding vector each, and use them to learn SSL via contrastive learning. To this end, previous work samples one of local image features as the image embedding and aggregates all local audio features to obtain the audio embedding, which is far from optimal due to the presence of noise and background irrelevant to the actual target in the input. We present a novel SSL method that addresses this chronic issue by joint slot attention on image and audio. To be specific, two slots competitively attend image and audio features to decompose them into target and off-target representations, and only target representations of image and audio are used for contrastive learning. Also, we introduce cross-modal attention matching to further align local features of image and audio. Our method achieved the best in almost all settings on three public benchmarks for SSL, and substantially outperformed all the prior work in cross-modal retrieval.", + "bbox": [ + 89, + 327, + 485, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 661, + 223, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans easily identify where a sound comes from within their sight. This capability is a key aspect of human perception facilitating interaction with surroundings, and is also demanded in a large variety of vision applications such as autonomous driving and robotics [10, 28, 34]. Motivated by this, a large body of research [1, 3, 6, 12, 14, 15, 21, 22, 25, 26, 29-33] has studied sound source localization (SSL), the task of locating the source of a sound within the paired image in the form of segmentation mask.", + "bbox": [ + 89, + 686, + 482, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to the prohibitively expensive cost of sound source annotation for image-audio pair data, existing methods [6, 21, 31, 33] have focused on learning SSL using only image-audio pairs without sound source labels. To this end, they consider the task as multiple instance learning [2, 13, 20]", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0edefc1447d60faec279218077c61fc689637c13d81c5c49cc69ea278f071b45.jpg", + "image_caption": [ + "(a) Previous work", + "Figure 1. Difference between previous work and ours. For contrastive learning of SSL, our method identifies and utilizes pairs of relevant image region and audio segment, while previous work exploits a local image feature and the global audio feature." + ], + "image_footnote": [], + "bbox": [ + 526, + 295, + 686, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cee2e6e0380f9fe5d0beeff613f7c05e3b77392cfb3ea91e72dbb29f5a6f7095.jpg", + "image_caption": [ + "(b) Ours" + ], + "image_footnote": [], + "bbox": [ + 686, + 295, + 893, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and assume that, given an image-audio pair, at least one patch of the image is relevant to the audio; then they represent the image with its local feature that most closely resembles the global audio feature, and learn SSL through contrastive learning [24, 27] with these features (Fig. 1(a)).", + "bbox": [ + 511, + 537, + 906, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While this approach has been proven to work, there is plenty of room for further improvement. First, audio may contain not only the sound of the target shared by the two modalities but also irrelevant sounds, such as noise or sounds from sources outside the image. In such cases, using the global audio feature for contrastive learning hinders accurate alignment between the image and audio features. Second, representing an image with only a single local feature in contrastive learning causes models to locate a small discriminative part of the sound source, not its whole body [35], which can degrade SSL accuracy.", + "bbox": [ + 511, + 613, + 908, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these limitations, we propose a novel SSL method based on slot attention [17], which is conceptually illustrated in Fig. 1(b). In slot attention, slots, which are randomly sampled or learnable embeddings, compete with each other for aggregating input features progressively, and thus they decompose the features into distinct and exclusive representations. Motivated by this, we employ slot attention with two learnable slots to jointly decompose image", + "bbox": [ + 511, + 779, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3121", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b4b6f866abc2ae0590ba4abe78b308ca4a7cf7e863535c21f098f90549188fa.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed method. First, image and audio features are extracted by the image and audio encoders, and serve as keys and values for slot attention. In slot attention, two learnable slots are used as queries to draw attentions on the input features while competing with each other, decomposing the features into the target and off-target representations per modality. The model is then trained by contrastive learning using only the target representations, and by cross-modal attention matching for further feature-level alignment between the two modalities. For localization in testing, we draw the attention map between the audio target representation and image features. Some details for learning slot attention, such as decoders and loss for slot reconstruction, have been omitted for clarity." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 767, + 304 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5103af55f0ec301d98f385465743772fba50039830256fc54a47baf0e4592ecd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 95, + 892, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and audio features into two distinct slot representations per modality, one for the target and the other for the rest. To be specific, we designate the first slot as the target slot, which is guided to focus on information related to the shared sound source across both modalities. The exclusive assignment property of slot attention directs the second slot, referred to as the off-target slot, to capture remaining elements, such as the image background or irrelevant sounds in the audio. Then we learn SSL by contrastive learning using the two target slot representations from the image and audio, respectively. Unlike the prior work using only a single global audio feature, our method allows to exploit only the relevant part (i.e., the target slot) of the audio for more accurate SSL. Furthermore, since the target slot of the image covers the sound source area rather than a single location, our method avoids the problem of localization being reduced to a small discriminative part.", + "bbox": [ + 88, + 428, + 485, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the target slot representations, SSL is conducted by drawing an attention map on the image feature map using the target slot representation of the audio as query (Fig. 2(b)). For accurate pixel-level localization, the target slot of the audio should be well aligned with the keys derived from relevant individual image features; in this regard, contrastive learning using only the target slots is insufficient as it does not relate the audio target slot with individual image features. We thus introduce an additional training strategy dubbed cross-modal attention matching. This strategy encourages the cross-modal attention map between the audio target slot and the input image features to approximate the intra-modal attention map between the image target slot and the input image features that are inherently well", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "aligned. In other words, it provides pseudo-supervision for SSL prediction (i.e., the cross-modal attention map) using the intra-modal attention map, which is drawn by the query and keys from the same modality and thus better localizes the target. To further facilitate its cross-modal alignment, we apply this strategy to the audio modality as well.", + "bbox": [ + 511, + 428, + 906, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our model was trained on two datasets, Flickr-SoundNet [4] and VGG-Sound [5], and evaluated on three benchmarks, Flickr-SoundNet-Test [4], VGG-SS [5], and AVSBench [36]. It achieved the state of the art in almost all settings across all the three evaluation benchmarks. Further, our model outperformed all the prior work by large margin in cross-modal retrieval, suggesting that it is not a simple saliency detector but properly learns to relate image and audio for SSL [31]. In summary, the contribution of our method is three-fold:", + "bbox": [ + 511, + 520, + 908, + 671 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel SSL model that identifies and extracts features relevant to sound source jointly from image and audio modalities through slot attention.", + "- We also propose cross-modal attention matching, which enables effective cross-modal alignment for individual input features and thus improves localization performance.", + "- Our model achieved the state of the art in almost all settings across all the three SSL benchmarks and improved cross-modal retrieval performance substantially." + ], + "bbox": [ + 511, + 676, + 906, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 828, + 653, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sound source localization aims to identify specific regions in an image that correspond to the sources of sounds from an audio input. Since constructing audio-visual datasets is", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3122", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "costly, sound source localization studies have evolved into self-supervised learning approaches that use only image-audio pairs for contrastive learning [1, 6, 12, 14, 16, 21, 22, 26, 29-31, 33].", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An early study, [29], addresses the task through contrastive learning between the audio-aware image features and the global audio feature. The audio-aware image features are obtained by calculating a correspondence map through the normalized inner product between the global audio feature and the image feature map. Chen et al. [6] propose method that uses a differentiable threshold on the correspondence map to divide audio-aware and other image features, treating the latter as hard negatives in contrastive learning to improve localization.", + "bbox": [ + 89, + 154, + 483, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent studies focus on fine-grained sound source localization through contrastive learning between the global audio feature and the most relevant feature selected from the image feature map [21, 22, 31, 33]. Mo and Morgado [21] propose a method that refines the audio-visual correspondence map through a post-processing technique leveraging a pretrained visual model. Sun et al. [33] introduce a regularizer that applies mean absolute error regularization to inter-modal and intra-modality similarity matrices to increase awareness of false negatives. Senocak et al. [31] propose a cross-modal alignment method that preserves localization performance and enhances cross-modal retrieval by leveraging multi-view augmentation and nearest neighbor search to construct multiple positive pairs for contrastive learning. These methods employ contrastive learning with the most relevant image feature to assist in coarsely identifying sound sources but often fall short of capturing a comprehensive representation, thereby hindering localization ability. Furthermore, relying solely on global audio features can mislead contrastive learning when noise exists, potentially highlighting incorrect image features.", + "bbox": [ + 91, + 306, + 483, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following a complementary direction, several studies employ external prior knowledge or even utilize off-the-shelf segmentation modules to achieve high-quality localization results [25, 26]. Also, several studies employ sound source localization with additional information such as optical flow [9], text inputs [19] or ground truth category [23]. By contrast to these studies, our method can effectively perform sound source localization without relying on external prior knowledge or off-the-shelf models.", + "bbox": [ + 89, + 626, + 482, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the above issues, we propose a sound source localization framework based on the slot attention mechanism. Slot attention [17] is proposed for unsupervised object-centric learning, effectively decomposing input features into distinct and mutually exclusive components via iterative attention and competition among slots. Inspired by this, we adopt slot attention in both image and audio features to decompose them into features related to the sound source and those irrelevant to it without supervision.", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 686, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our model consists of two parts: 1) image and audio encoders for feature extraction, and 2) joint feature decomposition module with joint slot attention mechanism. A pair of image and audio is first fed into each encoder to extract image and audio features (Sec. 3.1). The features of each modality are decomposed into two distinct slot representations by using joint slot attention: one slot for the target and the other for the rest. (Sec. 3.2). The slot representations are then used for training as input to contrastive learning and cross-modal attention matching (Sec. 3.3), while effectively handling false negatives (Sec. 3.4). Lastly, an attention on the image feature map using the target slot representation of the audio is computed to locate the source of sound (Sec. 3.5).", + "bbox": [ + 511, + 114, + 906, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Feature Extraction", + "text_level": 1, + "bbox": [ + 511, + 334, + 694, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $D = \\{(v_{i},a_{i})\\}_{i = 1}^{M}$ denote a training set of image-audio pairs, where $v_{i}\\in \\mathbb{R}^{3\\times H_{v}\\times W_{v}}$ is an image of size $H_{v}\\times W_{v}$ , $a_{i}\\in \\mathbb{R}^{1\\times F_{a}\\times T_{a}}$ is a corresponding audio spectrogram of frequency resolution $F_{a}$ and temporal length $T_{a}$ , and $M$ denotes the total number of such pairs in the dataset. For feature extraction, separate image and audio encoders $f_{v}$ and $f_{a}$ , are adopted. The image encoder extracts the image features $\\mathbf{v}_i\\coloneqq f_v(v_i)\\in \\mathbb{R}^{h\\times w\\times c}$ for $v_{i}$ , and the audio encoder extracts the audio features $\\mathbf{a}_i\\coloneqq f_a(a_i)\\in \\mathbb{R}^{t\\times c}$ for $a_{i}$ where $h$ and $w$ are the height and width of the image feature map respectively, $t$ is the length of the audio features, and $c$ is the image and audio feature dimension.", + "bbox": [ + 511, + 357, + 906, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Joint Feature Decomposition", + "text_level": 1, + "bbox": [ + 511, + 547, + 769, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Slot attention [17] is an attention mechanism where each slot progressively competes to aggregate input features, leading to the decomposition of input features into distinct and exclusive representations. Building on this idea, for joint feature decomposition, we propose joint slot attention (JSA), which jointly decomposes image and audio features into two distinct slot representations for each modality. To achieve this, we employ slot attention with two initial slots $S_0 \\in \\mathbb{R}^{2 \\times D}$ shared in the two modalities. Furthermore, we designate the first slot as target slot, which captures features relevant to sound source, and the other as off-target slot, which captures features for the rest. Since the overall pipeline of JSA is identical in the two modalities, we will describe the module in terms of image only for brevity.", + "bbox": [ + 511, + 571, + 906, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "JSA first projects an input image feature $\\mathbf{v}$ to derive its corresponding key $K^{\\nu}$ and value $V^{\\nu}$ by applying the following operations:", + "bbox": [ + 511, + 782, + 905, + 827 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nK ^ {v} = W ^ {k} (\\operatorname {L a y e r N o r m} (\\mathbf {v})) \\in \\mathbb {R} ^ {h w \\times d}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 835, + 903, + 861 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV ^ {v} = W ^ {v} \\left(\\operatorname {L a y e r N o r m} (\\mathbf {v})\\right) \\in \\mathbb {R} ^ {h w \\times d},\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 857, + 838, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W^{k}$ , $W^{v}$ are the linear projection layers to obtain", + "bbox": [ + 511, + 883, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3123", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "embeddings in $d$ -dimension. For each iteration $n \\in [1, N]$ , the query $Q_{n}^{v}$ is derived by applying linear projection to the slots with the initial slots $S_{0}^{v} \\coloneqq S_{0}$ , which is given by", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {n} ^ {v} = W ^ {q} \\left(\\operatorname {L a y e r N o r m} \\left(S _ {n - 1} ^ {v}\\right)\\right) \\in \\mathbb {R} ^ {2 \\times d}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 146, + 482, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W^q$ is a linear projection layer shared for all iterations. The slot attention matrix $A$ is computed by taking the dot product of the given query $Q_n^v$ and key $K^v$ , scaled by the square root of the dimension $d$ as follows:", + "bbox": [ + 89, + 174, + 483, + 236 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i, j} = \\frac {\\exp \\left(M _ {i , j}\\right)}{\\sum_ {l = 1} ^ {2} \\exp \\left(M _ {i , l}\\right)}, \\text {w h e r e} M = \\frac {K ^ {v} \\left(Q _ {n} ^ {v}\\right) ^ {\\top}}{\\sqrt {d}} \\in \\mathbb {R} ^ {h w \\times 2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 244, + 488, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i$ is key index, and $j$ is slot index. Subsequently, a weighted mean is utilized to the slot attention matrix $A$ :", + "bbox": [ + 89, + 294, + 483, + 325 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {A} _ {i, j} = \\frac {A _ {i , j}}{\\sum_ {l = 1} ^ {h w} A _ {l , j}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 333, + 482, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, the updated slot $\\bar{S}_n$ is computed by Gated Recurrent Unit (GRU) [7] with $S_{n - 1}$ as hidden state, and $\\hat{A}^{\\top}\\cdot V^{v}$ as input as follows:", + "bbox": [ + 89, + 378, + 483, + 425 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {S _ {n} ^ {v}} = \\operatorname {G R U} \\left(S _ {n - 1} ^ {v}, \\hat {A} ^ {\\top} \\cdot V ^ {v}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 434, + 482, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the slot is updated by a residual connection:", + "bbox": [ + 89, + 463, + 434, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {n} ^ {v} = \\bar {S} _ {n} ^ {v} + \\operatorname {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(\\bar {S} _ {n} ^ {v}\\right)\\right) \\in \\mathbb {R} ^ {2 \\times d}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 487, + 482, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall process encourages the input feature to bind to specific slots, ensuring that each slot forms a mutually exclusive representation. The final outputs of JSA are image slots $\\{\\mathbf{p}^v;\\mathbf{r}^v\\} := S_N^v\\in \\mathbb{R}^{2\\times d}$ and image query $Q_{N}^{v}\\in$ $\\mathbb{R}^{2\\times d}$ , where $\\mathbf{p}^v$ is a target slot and $\\mathbf{r}^v$ is an off-target slot. The same process is applied to the audio modality using audio feature $\\mathbf{a}$ , yielding audio slots $\\{\\mathbf{p}^a;\\mathbf{r}^a\\} := S_N^a\\in \\mathbb{R}^{2\\times d}$ and audio query $Q_{N}^{a}\\in \\mathbb{R}^{2\\times d}$ . Since this process begins from the shared initial slots, the initial slots are able to effectively learn joint features between two modalities. For brevity, we omit the subscript $N$ for the final query, $Q^{v}\\coloneqq Q_{N}^{v}$ and $Q^{a}\\coloneqq Q_{N}^{a}$ .", + "bbox": [ + 89, + 516, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Learning Objectives", + "text_level": 1, + "bbox": [ + 89, + 705, + 282, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive Learning with Target Slots. We adopt the InfoNCE loss [24, 27] for contrastive learning using the target slots, guiding the target slots to identify features relevant to the sound source jointly from image and audio. Let $s$ denote a similarity function, $s(x,y) = \\exp (\\cos (x,y) / \\tau)$ , where $\\tau$ is a temperature hyperparameter. The contrastive learning loss $\\mathcal{L}_{\\mathrm{cotr}}$ is given by", + "bbox": [ + 89, + 729, + 483, + 837 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o t r}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left\\{\\log \\frac {s \\left(\\mathbf {p} _ {i} ^ {v} , \\mathbf {p} _ {i} ^ {a}\\right)}{\\sum_ {j} ^ {B} s \\left(\\mathbf {p} _ {i} ^ {v} , \\mathbf {p} _ {j} ^ {a}\\right)} + \\log \\frac {s \\left(\\mathbf {p} _ {i} ^ {v} , \\mathbf {p} _ {i} ^ {a}\\right)}{\\sum_ {j} ^ {B} s \\left(\\mathbf {p} _ {j} ^ {v} , \\mathbf {p} _ {i} ^ {a}\\right)} \\right\\},\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 845, + 496, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B$ denotes the number of data in a mini-batch, and the negative pairs refer to the combinations of visual and audio features that do not belong to the same original pair within the mini-batch. Since the image target slot captures the whole features of the target region, it avoids focusing solely on a specific discriminative part of the image. Furthermore, the audio target slot captures only the relevant sound from the source, it prevents the issue of learning sounds that are unrelated to the source.", + "bbox": [ + 511, + 90, + 905, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-modal Attention Matching. As will be described in Sec. 3.5, the localization result is derived by utilizing cross-modal attention between the audio target slot and image features. For accurate sound source localization, the audio target slot should be properly aligned with keys computed from the relevant image features. Although contrastive learning encourages alignment of target slot representations between different modalities, it does not ensure alignment between the audio target slot and image features, which degrades the localization accuracy. We resolve this issue by cross-modal attention matching, which encourages the cross-modal attention map to resemble the intra-modal attention map for improving feature-level alignment across the two modalities. To this end, we first calculate the cross-modal and intra-modal attention matrices. Specifically, we compute the cross-modal attention matrix between the audio and visual modalities in the same manner as Eq. (3) and Eq. (4), with the audio queries $Q^a$ and visual keys $K^v$ . We denote this cross-modal attention matrix as $\\hat{A}^{a,v} \\in \\mathbb{R}^{hw \\times 2}$ , where $hw$ is the length of $K^v$ . Subsequently, to obtain the cross-modal attention that represents the attention weights between the audio target query and visual keys, we utilize only the first column of $\\hat{A}^{a,v}$ . This cross-modal attention is defined as follows:", + "bbox": [ + 511, + 227, + 906, + 589 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c a} ^ {a, v} := \\left[ \\hat {A} _ {i, 1} ^ {a, v} \\right] _ {i = 1} ^ {h w} \\in \\mathbb {R} ^ {h w}.\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 595, + 799, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, for intra-modal attention matrix within the visual modality, we compute the intra-attention matrix $\\hat{A}^{v,v}$ using the $Q^v$ and $K^v$ . Then, the intra-modal attention that represents the attention weights between the image target query and visual keys is defined as follows:", + "bbox": [ + 511, + 621, + 905, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {i a} ^ {v, v} := \\left[ \\hat {A} _ {i, 1} ^ {v, v} \\right] _ {i = 1} ^ {h w} \\in \\mathbb {R} ^ {h w}.\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 707, + 799, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further facilitate cross-modal alignment, we apply this strategy symmetrically to the audio modality as well. Finally we formulate the cross-modal attention matching loss as follows:", + "bbox": [ + 511, + 738, + 905, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a t c h}} = \\left\\| \\mathbf {c a} ^ {a, v} - \\operatorname {s g} \\left(\\mathbf {i a} ^ {v, v}\\right) \\right\\| _ {2} ^ {2} + \\left\\| \\mathbf {c a} ^ {v, a} - \\operatorname {s g} \\left(\\mathbf {i a} ^ {a, a}\\right) \\right\\| _ {2} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 810, + 906, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{sg}(\\cdot)$ denotes the stop-gradient operation. $\\mathcal{L}_{\\mathrm{match}}$ minimizes the discrepancy between cross-modal and intramodal attentions, promoting robust feature-level alignment between the two modalities for accurate localization.", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3124", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Slot Divergence. In slot attention, the diversity between the target slot and off-target slot should be large enough to ensure that each slot effectively captures distinct, mutually exclusive regions. Therefore, we introduce a slot divergence loss that encourages greater diversity among the slots. The slot divergence loss $\\mathcal{L}_{\\mathrm{div}}$ is given by", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d i v}} = \\max \\left\\{0, \\cos \\left(\\mathbf {p} ^ {v}, \\mathbf {r} ^ {v}\\right) \\right\\} + \\max \\left\\{0, \\cos \\left(\\mathbf {p} ^ {a}, \\mathbf {r} ^ {a}\\right) \\right\\}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 186, + 482, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It ensures that the different slot embeddings to be sufficiently dissimilar, enabling the target slot to more effectively localize the sound source.", + "bbox": [ + 89, + 209, + 480, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Slot Reconstruction. The successful feature reconstruction in slot attention implies that each slot successfully captures unique and independent parts of the input data, which means that the slot attention mechanism effectively decomposed the input into individual slot representations. It indicates that each slot accurately represents a specific object or region in the input. To encourage this condition, we apply the slot reconstruction loss $\\mathcal{L}_{\\mathrm{recon}}$ as follows:", + "bbox": [ + 89, + 255, + 482, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\| \\mathbf {v} - g _ {v} \\left(\\left\\{\\mathbf {p} ^ {v}; \\mathbf {r} ^ {v} \\right\\}\\right) \\| _ {2} ^ {2} + \\| \\mathbf {a} - g _ {a} \\left(\\left\\{\\mathbf {p} ^ {a}; \\mathbf {r} ^ {a} \\right\\}\\right) \\| _ {2} ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 380, + 482, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g_{v}$ and $g_{a}$ are the decoders for the image and audio features, respectively. These decoders are composed of a small number of MLP layers with ReLU activations. Note that the decoders are auxiliary modules used only in training and not required for inference.", + "bbox": [ + 89, + 402, + 482, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss is a combination of $\\mathcal{L}_{\\mathrm{cotr}}$ , $\\mathcal{L}_{\\mathrm{match}}$ , $\\mathcal{L}_{\\mathrm{div}}$ , and $\\mathcal{L}_{\\mathrm{recon}}$ , which is given by", + "bbox": [ + 89, + 479, + 482, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o t r}} + \\lambda_ {1} \\mathcal {L} _ {\\text {m a t c h}} + \\lambda_ {2} \\mathcal {L} _ {\\text {d i v}} + \\lambda_ {3} \\mathcal {L} _ {\\text {r e c o n}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 515, + 480, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ , $\\lambda_{2}$ , and $\\lambda_{3}$ are hyperparameters weighting the loss terms.", + "bbox": [ + 89, + 537, + 482, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Mitigating False Negatives", + "text_level": 1, + "bbox": [ + 89, + 574, + 330, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In general, unpaired images and audiios are all treated as negatives for contrastive training. However, some of them are false negatives that are semantically well matched, so treating these as negatives causes adverse effect [33]. Therefore, we propose a false negative mitigation strategy using $k$ -reciprocal nearest neighbors of target slots.", + "bbox": [ + 89, + 597, + 482, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the image target slot $\\mathbf{p}_i^v$ , which is derived from $v_{i}$ we define its set of $k$ -reciprocal nearest neighbors $R_{k}^{v}(i)$ as", + "bbox": [ + 89, + 686, + 482, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {k} ^ {v} (i) = \\left\\{\\left(v _ {j}, a _ {j}\\right) \\mid \\left(j \\in N _ {k} ^ {v} (i)\\right) \\cap \\left(i \\in N _ {k} ^ {v} (j)\\right) \\right\\}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 723, + 482, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{k}^{v}(i)$ is the $k$ -nearest neighbors of $\\mathbf{p}_i^v$ , computed using cosine similarity. Likewise, the set of $k$ -reciprocal nearest neighbors between audio target slots, i.e., $R_{k}^{a}(i)$ , can be derived. The predicted false negative set $R_{k}(i)$ is defined as", + "bbox": [ + 89, + 744, + 482, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {k} (i) = R _ {k} ^ {v} (i) \\cap R _ {k} ^ {a} (i). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 821, + 482, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the samples in $R_{k}(i)$ have both similar image and audio target slots with $(v_{i}, a_{i})$ , they are likely to belong to false negatives. We exclude these predicted false negatives from the contrastive learning in Eq. (7).", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Inference", + "text_level": 1, + "bbox": [ + 513, + 90, + 624, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For inference, $\\mathbf{ca}^{a,v}$ , which represents the attention between the audio target slot and the input image features, is utilized. Specifically, the attention is interpolated to the original image size, and pixels whose score exceeds the hyperparameter $\\theta$ are selected to generate the localization result.", + "bbox": [ + 511, + 114, + 903, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, unlike the object-guided localization [21], we propose a novel localization refinement strategy, dubbed Image-Query based Refinement (IQR), to refine the localization result without relying on external prior knowledge. IQR estimates the regions that are likely to contain the target by using the attention between the image target slot and the input image features. It can be used to identify prior regions in the image that favor the target, regardless of whether these regions are actually the source of the sound. The refined result derived by IQR, denoted as $R_{\\mathrm{IQR}}$ , is given as", + "bbox": [ + 511, + 189, + 906, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {\\mathrm {I Q R}} = \\alpha \\cdot \\mathbf {c a} ^ {a, v} + (1 - \\alpha) \\cdot \\mathbf {i a} ^ {v, v}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 353, + 903, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha \\in (0,1)$ balances the sound source localization result and potential sound source regions.", + "bbox": [ + 511, + 383, + 903, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 513, + 428, + 638, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 511, + 453, + 689, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Flickr-SoundNet [4] and VGG-Sound [5] are used to train. Flickr-SoundNet contains about 2 million unconstrained videos from Flickr, while VGG-Sound has around 200k video clips. For training, we use subsets of 10k and 144k from each dataset, respectively. For the evaluation, Flickr-SoundNet-Test [4], VGG-SS [5], and AVSBench [36] are used following the previous research [6, 21, 31, 33]. Flickr-SoundNet-Test has 250 audio-visual pairs with manually labeled bounding boxes. VGG-SS has about 5k audio-visual pairs over 220 categories. Since VGG-SS provides its category ground truth, it is also used to evaluate cross-modal retrieval performance. AVSBench has about 5k videos with pixel-wise binary segmentation maps. Following the previous study [6], the middle frame of the video clip is used for the image. For the audio, the middle 5 seconds are extracted and converted into log spectrograms with 257 frequency bins at a sampling rate of $16\\mathrm{kHz}$ .", + "bbox": [ + 511, + 476, + 906, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Architecture. ResNet-18 [11] is adopted for both the image encoder and the audio encoders. The image encoder is fine-tuned from an ImageNet [8] pretrained model, while the audio encoder is trained from scratch. Both encoders include an additional $1 \\times 1$ convolutional layer for projection, and the audio encoder also includes a MaxPooling layer for pooling along the frequency axis.", + "bbox": [ + 511, + 734, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hyperparameter Settings. We use the AdamW optimizer [18] with a learning rate of 5e-5 and a weight decay of 1e-2. The mini-batch size during training is set to 256, and the size of image and audio feature size $h$ , $w$ , $t$ and $c$", + "bbox": [ + 511, + 840, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3125", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2b4f21a720c1d8451c9596f8b0b675fb2a6c860ea995afe70f5afbc8ce968b0b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TrainsetMethodOGLVGG-SSFlickr-SoudNet
cIoUAUCcIoUAUC
Flickr SoundNet 10kAttention10k [29]--43.6044.90
CoursetoFine [26]--52.2049.60
AVObject [1]--54.6050.40
LVS [6]--58.2052.50
EZ-VSL [21]19.8630.9662.2454.74
FNAC [33]35.2738.0084.3363.26
Ours35.1138.6685.2065.26
EZ-VSL [21]✓37.6139.2181.9362.58
FNAC [33]✓40.9740.3884.7364.34
Ours✓39.1640.5187.6066.50
Flickr SoundNet 144kAttention10k [29]--66.0055.80
DMC [12]--67.1056.80
LVS [6]--69.9057.30
HardPos [30]--75.2059.70
EZ-VSL [21]30.2735.9272.6958.70
FNAC [33]33.9337.2978.7159.33
Alignment [31]--85.2062.20
Ours37.0139.4286.0065.16
EZ-VSL [21]✓41.0140.2383.1363.06
FNAC [33]✓41.1040.4483.9363.06
Alignment [31]✓--84.0064.16
Ours✓40.1540.7589.2064.50
", + "bbox": [ + 91, + 88, + 491, + 363 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a28be073f60b3512e3723eba657939c34b8bc8c060ad6821bc47a98ceec6a883.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TrainsetMethodOGLVGG-SSFlickr-SoudNet
cIoUAUCcIoUAUC
VGGSound 10kLVS [6]--61.8053.60
EZ-VSL [21]25.8433.6863.8554.44
FNAC [33]37.2938.9985.7463.66
Ours37.8439.8286.0064.90
EZ-VSL [21]✓38.7139.8078.7161.53
FNAC [33]✓40.6940.4282.1363.64
Ours✓41.0241.2987.6065.42
VGGSound 144kAttention10k [29]18.5030.20--
DMC [12]29.1034.80--
AVObject [1]29.7035.70--
LVS [6]34.4038.2073.5059.00
HardPos [30]34.6038.0076.8059.20
EZ-VSL [21]34.3837.7079.5161.17
SLAVC [22]37.22-83.20-
FNAC [33]39.5039.6684.7363.76
Alignment [31]39.9440.0279.6063.44
NoPrior [14]41.441.2--
Ours40.6241.6683.6064.68
EZ-VSL [21]✓38.8539.5483.9463.60
SLAVC [22]✓39.67-86.40-
FNAC [33]✓41.8540.8085.1464.30
Alignment [31]✓42.6441.4882.4064.60
Ours✓42.5642.4684.8064.82
", + "bbox": [ + 511, + 88, + 903, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative results of the model trained on Flickr-SoundNet-10k, 144k, and VGGSound-10k, 144k datasets. Note that the results of previous research are obtained from [14, 31, 33]. The results of our method represent the performance after applying Image-Query based Refinement (IQR). The result with OGL denotes the performance after applying object-guided localization [21].", + "bbox": [ + 89, + 372, + 906, + 415 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0960ed3c5dcdee521c0e1003cee3986a7f06a785f4ddb4a179cee306675f0003.jpg", + "image_caption": [ + "Figure 3. Sound localization results on Flickr-SoundNet-Test [4] and VGG-SS [5]. (a) Input image. (b) Ground-Truth. (c) Ours. (d) Alignment [31]. (e) FNAC [33]. (f) EZ-VSL [21]. The qualitative results are obtained by the model trained on Flickr-144k and the model trained on VGGSound-144k, respectively. Note that all visualizations are obtained without refinement." + ], + "image_footnote": [], + "bbox": [ + 117, + 428, + 875, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are set to 7, 7, 16, and 512 respectively. $\\tau$ in Eq. (7) is set to 0.03. $\\lambda_{1}, \\lambda_{2}$ , and $\\lambda_{3}$ in Eq. (11) are set to 100.0, 0.1, and 0.1, respectively. $\\alpha$ in Eq. (14) is set to 0.6. The value $k$ in Eq. (13) is set to 20. The number of iterations $N$ for slot attention mechanism is set to 5.", + "bbox": [ + 89, + 679, + 482, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metric. We measure the average precision at a Consensus Intersection over Union (cIoU) threshold of 0.5 to evaluate localization accuracy. Additionally, the AUC score is used to provide an overall performance measure across different thresholds.", + "bbox": [ + 89, + 756, + 482, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 845, + 284, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Flickr-SoundNet and VGGSound. In this section, we compare our method with previous self-supervised SSL", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "approaches. We trained our model on subsets of Flickr-SoundNet and VGGSound with dataset sizes of 10k and 144k, and evaluated the performance on Flickr-SoundNet-Test and VGG-SS. As shown in Table 1, our model achieves the state-of-the-art performance in most settings. Specifically, our method yields AUC improvements across all evaluated settings. Moreover, even without using Object-Guided Localization (OGL), our model performs competitively or better compared to previous methods, demonstrating the robustness of our architecture in focusing on the sound source without additional refinement steps.", + "bbox": [ + 511, + 679, + 906, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Heard 110 and Unheard 110. To assess the generalization ability of SSL, we evaluate the performance with Heard 110 and Unheard 110. Heard 110 and Unheard 110 are sub", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3126", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/09051436e92880583f1db94770ab26f2939c391da675d228d88b99f0ca84847b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodOGLHeard 110Unheard 110
cIoUAUCcIoUAUC
LVS [6]28.9036.2026.3034.70
EZ-VSL [21]31.8636.1932.6636.72
SLAVC [22]35.84-36.50-
Alignment [31]38.3139.0539.1139.80
Ours38.4640.1540.3541.25
EZ-VSL [21]✓37.2538.9739.5739.60
SLAVC [22]✓38.22-38.87-
FNAC [33]✓39.5439.8342.9141.17
Alignment [31]✓41.8540.9342.9441.54
Ours✓41.5041.6043.1142.04
", + "bbox": [ + 106, + 89, + 472, + 263 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d509a7b7f68994cf3721f5e6c7212740f93c159a1e92317664479ed0b18329c9.jpg", + "table_caption": [ + "Table 2. Quantitative results on Heard 110 and Unheard 110. Note that the results of previous research are obtained from [31, 33]. The results of our method represent the performance after applying Image-Query based Refinement (IQR). The result with OGL denotes the performance after applying object-guided localization." + ], + "table_footnote": [], + "table_body": "
TestsetMethodmIoUF-score
S4LVS [6]26.933.6
EZ-VSL [21]27.634.2
SLAVC [22]28.134.6
FNAC [33]27.1531.4
Alignment [31]29.635.9
\\( Alignment^† \\)29.335.6
Ours31.3344.30
", + "bbox": [ + 138, + 376, + 439, + 503 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1753d77bbaa50f35589e3f9547aa8fe2d30bc43dfc6fb39b3b3597262675ea27.jpg", + "table_caption": [ + "Table 3. Quantitative results on AVSBench. Note that the results of previous research are obtained from [31, 33]. $\\dagger$ denotes that the self-supervised pretrained encoder is utilized for their work." + ], + "table_footnote": [], + "table_body": "
MethodAudio → ImageImage → Audio
R@1R@5R@10R@1R@5R@10
LVS [6]3.8712.3520.734.9014.2921.37
EZ-VSL [21]5.0115.7324.8114.2033.5145.18
SSL-TIE [16]10.2930.6843.7612.7629.5839.72
SLAVC [22]4.7713.0819.106.1221.1632.12
Alignment [31]16.4736.9949.0020.0942.3853.66
\\( Alignment^† \\)14.3137.8149.1718.0038.3949.02
Ours30.6157.8769.3731.7058.4369.81
", + "bbox": [ + 91, + 575, + 480, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sets of the VGG-SS, each containing 110 non-overlapping categories. During training, the model is trained only to the category set of one group. The subset comprising categories seen during training is called Heard 110, while the subset with unseen categories is Unheard 110. As shown in Table 2, our model achieves the state-of-the-art performance on Unheard 110 both without additional refinement", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/840127ecbe826f612369e557772913daf03138f221b29631274c8ef3b01592c1.jpg", + "table_caption": [ + "Table 4. Cross-modal retrieval on the VGG-SS. All of the models are trained on VGGSound-144K. Note that the results of previous research are obtained from [31]. $\\dagger$ denotes that the self-supervised pretrained encoder is utilized for their work." + ], + "table_footnote": [], + "table_body": "
Trainset\\( {\\mathcal{L}}_{\\text{match }} \\)\\( {\\mathcal{L}}_{\\text{div }} \\)FN MitigateVGG-SS
cIoUAUC
VGGSound 144k✓14.9527.02
✓36.2339.13
18.6930.03
✓15.0328.18
✓✓40.5641.35
✓✓36.1439.31
✓✓20.3730.81
✓✓✓40.7141.62
", + "bbox": [ + 539, + 89, + 883, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/87249bb07f4072655853b5ddb81ff36c178389bf68b53f7fcc7fe00868a5f7c0.jpg", + "table_caption": [ + "Table 5. Detailed ablation study for losses." + ], + "table_footnote": [], + "table_body": "
TrainsetShared Initial SlotsVGG-SSFlickr
cIoUAUCcIoUAUC
Flickr-144k✓34.8038.7284.0061.70
36.6439.2486.8065.44
VGGSound-144k✓33.1137.9281.2064.36
40.7141.6284.4064.50
", + "bbox": [ + 514, + 263, + 901, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation study on shared initial slots.", + "bbox": [ + 563, + 363, + 854, + 377 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8ea648b19ac68aad20d33c35672dc060174bb30465a0e4af7489e6b0b5a6ea3.jpg", + "image_caption": [ + "Figure 4. Qualitative results to show the impact of $\\mathcal{L}_{\\mathrm{match}}$ on Flickr-SoundNet. (a) Input image. (b) Ground-Truth. (c) Results without cross-modal attention matching. (d) Results with cross-modal attention matching." + ], + "image_footnote": [], + "bbox": [ + 529, + 392, + 890, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and with OGL applied. These results demonstrate that our model has superior generalization ability compared to previous models.", + "bbox": [ + 511, + 718, + 906, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AVSBench. To more precisely evaluate the performance of the localization mask, we measured the performance on AVSBench, which provides segmentation masks. Our model was trained on VGGSound-144k and evaluated in a zero-shot manner on AVSBench. As shown in Table 3, we achieved the state-of-the-art results, specifically with improvements of $2.03\\%$ p in mIoU and $8.7\\%$ p in F-score.", + "bbox": [ + 511, + 763, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-modal Retrieval. As discussed in the previous study [31], the benchmark evaluation metrics, cIoU and", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3127", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0aea5b30ff7ce9d68e20812e204c773445d5e523bd4ef1ce6620ec75f50564a2.jpg", + "image_caption": [ + "Figure 5. Qualitative results of cross-modal retrieval on VGG-SS." + ], + "image_footnote": [], + "bbox": [ + 135, + 88, + 864, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AUC, do not guarantee better audio-visual semantic understanding. Therefore, we also evaluate cross-modal retrieval performance on VGG-SS. We extract the image and audio target slot representations and compute the cosine similarity between them. The similarities are ranked in descending order, and we calculate Recall@K performance using the category labels from the top-K retrieved audio or images. In Table 4, our model demonstrates a significant performance improvement over the previous state of the art in cross-modal retrieval. Specifically, our model improves Recall@1 by $16.3\\%$ p in audio-to-image retrieval and $13.7\\%$ p in image-to-audio retrieval. This result indicates that our model does not merely detect salient objects likely to produce sound within the image, but rather effectively locates the target specified by the audio input.", + "bbox": [ + 89, + 306, + 483, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 542, + 243, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Losses and False Negative Mitigating in Sec. 3.4. In Table 5, we conduct an ablation study to verify the effectiveness of each loss function and the strategy outlined in Sec. 3.4. The overall results show that the designed losses and the false negative mitigating strategy each contribute to improve the performance. Notably, the experiment with $\\mathcal{L}_{\\mathrm{match}}$ demonstrates a drastic increase in performance, highlighting the importance of cross-modality attention matching. Additionally, the ablation study on $\\mathcal{L}_{\\mathrm{div}}$ shows that explicitly separating the slots enables each slot to perform its designated role more effectively.", + "bbox": [ + 89, + 565, + 482, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Shared Initial Slots. We conduct an ablation study to examine the impact of shared initial slots by using different initial slots for each modality. As shown in Table 6, sharing initial slots resulted in better performance, indicating that this design choice facilitates the ability to jointly capture sound source-related information across different modalities.", + "bbox": [ + 89, + 732, + 482, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 847, + 272, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quality of SSL results. Fig. 3 visualizes the qualitative results of our model and previous studies on Flickr-", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SoundNet [4] and VGG-SS [5] test sets. These results demonstrate that our model has the ability to effectively localize sound source objects. Notably, it produces compact localization results even for challenging samples that other models struggle with, such as small birds, a gun, and cars surrounded by other objects in street scenes.", + "bbox": [ + 511, + 306, + 906, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Analysis of $\\mathcal{L}_{\\mathrm{match}}$ . Fig. 4 shows the qualitative result that shows the effect of $\\mathcal{L}_{\\mathrm{match}}$ on Flickr-SoundNet-Test. The attention maps is generated between the audio query of the audio target slot and the image keys derived from image features. Without $\\mathcal{L}_{\\mathrm{match}}$ , the alignment between the two is insufficient for compact localization, whereas with $\\mathcal{L}_{\\mathrm{match}}$ , the alignment improves, resulting in more accurate localization of the sound source.", + "bbox": [ + 511, + 397, + 906, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-modal Retrieval. Fig. 5 shows the qualitative results of our model for cross-modal retrieval. Consistent with the impressive performance shown in Table 4, our model demonstrates a strong ability to retrieve semantically similar samples across modalities that correspond to the sound source target.", + "bbox": [ + 511, + 517, + 906, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 621, + 633, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presented a novel self-supervised sound source localization method leveraging joint slot attention to align image and audio features. Our approach decomposes these features into target and off-target slots, using only target slots for contrastive learning to capture sound-related features. In addition, we introduced cross-modal attention matching and slot divergence loss which significantly improves localization accuracy and and we demonstrated the effectiveness of our model through comprehensive analysis. For future work, we aim to extend our framework to multisource sound source localization.", + "bbox": [ + 511, + 647, + 906, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the NRF grant (RS-2021-NR059830-40%) and the IITP grants (RS-2024-00509258-40%, RS-2021-II212068-10%, RS-2019-II191906-10%) funded by Ministry of Science and ICT, Korea.", + "bbox": [ + 511, + 813, + 906, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3128", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In Proc. European Conference on Computer Vision (ECCV), 2020. 1, 3, 6", + "[2] Stuart Andrews, Ioannis Tsochantaridis, and Thomas Hofmann. Support vector machines for multiple-instance learning. Advances in neural information processing systems, 15, 2002. 1", + "[3] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proc. European Conference on Computer Vision (ECCV), pages 435-451, 2018. 1", + "[4] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Sound-net: Learning sound representations from unlabeled video. Proc. Neural Information Processing Systems (NeurIPS), 29, 2016. 2, 5, 6, 8", + "[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2, 5, 6, 8", + "[6] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 5, 6, 7", + "[7] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724-1734, Doha, Qatar, 2014. Association for Computational Linguistics. 4", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: a large-scale hierarchical image database. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 5", + "[9] Dennis Fedorishin, Deen Dayal Mohan, Bhavin Jawade, Srirangaraj Setlur, and Venu Govindaraju. Hear the flow: Optical flow-based self-supervised visual sound source localization. In Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV), pages 2278-2287, 2023. 3", + "[10] Yury Furletov, Volker Willert, and Jürgen Adamy. Auditory scene understanding for autonomous driving. In 2021 IEEE Intelligent Vehicles Symposium (IV), pages 697-702. IEEE, 2021. 1", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5", + "[12] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 3, 6", + "[13] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In Proc. In-" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ternational Conference on Machine Learning (ICML), pages 2127-2136. PMLR, 2018. 1", + "[14] Dongjin Kim, Sung Jin Um, Sangmin Lee, and Jung Uk Kim. Learning to visually localize sound sources from mixtures without prior source knowledge. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 26467-26476, 2024. 1, 3, 6", + "[15] Chen Liu, Peike Patrick Li, Xingqun Qi, Hu Zhang, Lincheng Li, Dadong Wang, and Xin Yu. Audio-visual segmentation by exploring cross-modal mutual semantics. In Proc. ACM Multimedia Conference (ACMMM), pages 7590–7598, 2023. 1", + "[16] Jinxiang Liu, Chen Ju, Weidi Xie, and Ya Zhang. Exploiting transformation invariance and equivariance for self-supervised sound localisation. In Proc. ACM Multimedia Conference (ACMMM), pages 3742-3753, 2022. 3, 7", + "[17] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Proc. Neural Information Processing Systems (NeurIPS), 33:11525-11538, 2020. 1, 3", + "[18] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. International Conference on Learning Representations (ICLR), 2019. 5", + "[19] Tanvir Mahmud, Yapeng Tian, and Diana Marculescu. T-vsl: Text-guided visual sound source localization in mixtures. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 26742-26751, 2024. 3", + "[20] Oded Maron and Tomás Lozano-Pérez. A framework for multiple-instance learning. Advances in neural information processing systems, 10, 1997. 1", + "[21] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In Proc. European Conference on Computer Vision (ECCV), 2022. 1, 3, 5, 6, 7", + "[22] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. In Proc. Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 6, 7", + "[23] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10565–10574, 2023. 3", + "[24] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 4", + "[25] Sooyoung Park, Arda Senocak, and Joon Son Chung. Can clip help sound source localization? In Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV), 2024. 1, 3", + "[26] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In Proc. European Conference on Computer Vision (ECCV), 2020. 1, 3, 6", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3129", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4", + "[28] Yoko Sasaki, Ryo Tanabe, and Hiroshi Takernura. Online spatial sound perception using microphone array on mobile robot. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2478-2484. IEEE, 2018. 1", + "[29] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 3, 6", + "[30] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Learning sound localization better from semantically similar samples. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022. 6", + "[31] Arda Senocak, Hyeonggon Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 5, 6, 7", + "[32] Zengjie Song, Yuxi Wang, Junsong Fan, Tieniu Tan, and Zhaoxiang Zhang. Self-supervised predictive learning: A negative-free method for sound source localization in visual scenes. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3222-3231, 2022.", + "[33] Weixuan Sun, Jiayi Zhang, Jianyuan Wang, Zheyuan Liu, Yiran Zhong, Tianpeng Feng, Yandong Guo, Yanhao Zhang, and Nick Barnes. Learning audio-visual source localization via false negative aware contrastive learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 5, 6, 7", + "[34] Jun Yin, Stefano Damiano, Marian Verhelst, Toon van Watershoot, and Andre Guntoro. Real-time acoustic perception for automotive applications. In 2023 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1-6. IEEE, 2023. 1", + "[35] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep features for discriminative localization. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1", + "[36] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In Proc. European Conference on Computer Vision (ECCV), pages 386–403, 2022. 2, 5" + ], + "bbox": [ + 91, + 90, + 482, + 739 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3130", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/feaa7f56-7770-404f-a305-b468e86df1e5_model.json b/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/feaa7f56-7770-404f-a305-b468e86df1e5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1ba734ffed36922cea3729f96a5dd9abcdf544f9 --- /dev/null +++ b/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/feaa7f56-7770-404f-a305-b468e86df1e5_model.json @@ -0,0 +1,2090 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.282, + 0.131, + 0.72, + 0.177 + ], + "angle": 0, + "content": "Improving Sound Source Localization with Joint Slot Attention on Image and Audio" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.203, + 0.826, + 0.223 + ], + "angle": 0, + "content": "Inho Kim1 Youngkil Song2 Jicheol Park2 Won Hwa Kim1,2 Suha Kwak1,2" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.224, + 0.77, + 0.243 + ], + "angle": 0, + "content": "\\(^{1}\\)Dept. of CSE, POSTECH \\(^{2}\\)Graduate School of AI, POSTECH" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.244, + 0.751, + 0.262 + ], + "angle": 0, + "content": "{kimih, songyk, jicheol, wonhwa, suha.kwak}@postech.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.295, + 0.328, + 0.312 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.486, + 0.631 + ], + "angle": 0, + "content": "Sound source localization (SSL) is the task of locating the source of sound within an image. Due to the lack of localization labels, the de facto standard in SSL has been to represent an image and audio as a single embedding vector each, and use them to learn SSL via contrastive learning. To this end, previous work samples one of local image features as the image embedding and aggregates all local audio features to obtain the audio embedding, which is far from optimal due to the presence of noise and background irrelevant to the actual target in the input. We present a novel SSL method that addresses this chronic issue by joint slot attention on image and audio. To be specific, two slots competitively attend image and audio features to decompose them into target and off-target representations, and only target representations of image and audio are used for contrastive learning. Also, we introduce cross-modal attention matching to further align local features of image and audio. Our method achieved the best in almost all settings on three public benchmarks for SSL, and substantially outperformed all the prior work in cross-modal retrieval." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.662, + 0.224, + 0.679 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Humans easily identify where a sound comes from within their sight. This capability is a key aspect of human perception facilitating interaction with surroundings, and is also demanded in a large variety of vision applications such as autonomous driving and robotics [10, 28, 34]. Motivated by this, a large body of research [1, 3, 6, 12, 14, 15, 21, 22, 25, 26, 29-33] has studied sound source localization (SSL), the task of locating the source of a sound within the paired image in the form of segmentation mask." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Due to the prohibitively expensive cost of sound source annotation for image-audio pair data, existing methods [6, 21, 31, 33] have focused on learning SSL using only image-audio pairs without sound source labels. To this end, they consider the task as multiple instance learning [2, 13, 20]" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.296, + 0.687, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.435, + 0.645, + 0.446 + ], + "angle": 0, + "content": "(a) Previous work" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.296, + 0.895, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.435, + 0.815, + 0.446 + ], + "angle": 0, + "content": "(b) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.458, + 0.908, + 0.515 + ], + "angle": 0, + "content": "Figure 1. Difference between previous work and ours. For contrastive learning of SSL, our method identifies and utilizes pairs of relevant image region and audio segment, while previous work exploits a local image feature and the global audio feature." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.613 + ], + "angle": 0, + "content": "and assume that, given an image-audio pair, at least one patch of the image is relevant to the audio; then they represent the image with its local feature that most closely resembles the global audio feature, and learn SSL through contrastive learning [24, 27] with these features (Fig. 1(a))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.78 + ], + "angle": 0, + "content": "While this approach has been proven to work, there is plenty of room for further improvement. First, audio may contain not only the sound of the target shared by the two modalities but also irrelevant sounds, such as noise or sounds from sources outside the image. In such cases, using the global audio feature for contrastive learning hinders accurate alignment between the image and audio features. Second, representing an image with only a single local feature in contrastive learning causes models to locate a small discriminative part of the sound source, not its whole body [35], which can degrade SSL accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "To address these limitations, we propose a novel SSL method based on slot attention [17], which is conceptually illustrated in Fig. 1(b). In slot attention, slots, which are randomly sampled or learnable embeddings, compete with each other for aggregating input features progressively, and thus they decompose the features into distinct and exclusive representations. Motivated by this, we employ slot attention with two learnable slots to jointly decompose image" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3121" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.768, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.097, + 0.893, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.319, + 0.907, + 0.402 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed method. First, image and audio features are extracted by the image and audio encoders, and serve as keys and values for slot attention. In slot attention, two learnable slots are used as queries to draw attentions on the input features while competing with each other, decomposing the features into the target and off-target representations per modality. The model is then trained by contrastive learning using only the target representations, and by cross-modal attention matching for further feature-level alignment between the two modalities. For localization in testing, we draw the attention map between the audio target representation and image features. Some details for learning slot attention, such as decoders and loss for slot reconstruction, have been omitted for clarity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.429, + 0.486, + 0.686 + ], + "angle": 0, + "content": "and audio features into two distinct slot representations per modality, one for the target and the other for the rest. To be specific, we designate the first slot as the target slot, which is guided to focus on information related to the shared sound source across both modalities. The exclusive assignment property of slot attention directs the second slot, referred to as the off-target slot, to capture remaining elements, such as the image background or irrelevant sounds in the audio. Then we learn SSL by contrastive learning using the two target slot representations from the image and audio, respectively. Unlike the prior work using only a single global audio feature, our method allows to exploit only the relevant part (i.e., the target slot) of the audio for more accurate SSL. Furthermore, since the target slot of the image covers the sound source area rather than a single location, our method avoids the problem of localization being reduced to a small discriminative part." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Given the target slot representations, SSL is conducted by drawing an attention map on the image feature map using the target slot representation of the audio as query (Fig. 2(b)). For accurate pixel-level localization, the target slot of the audio should be well aligned with the keys derived from relevant individual image features; in this regard, contrastive learning using only the target slots is insufficient as it does not relate the audio target slot with individual image features. We thus introduce an additional training strategy dubbed cross-modal attention matching. This strategy encourages the cross-modal attention map between the audio target slot and the input image features to approximate the intra-modal attention map between the image target slot and the input image features that are inherently well" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.519 + ], + "angle": 0, + "content": "aligned. In other words, it provides pseudo-supervision for SSL prediction (i.e., the cross-modal attention map) using the intra-modal attention map, which is drawn by the query and keys from the same modality and thus better localizes the target. To further facilitate its cross-modal alignment, we apply this strategy to the audio modality as well." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.909, + 0.672 + ], + "angle": 0, + "content": "Our model was trained on two datasets, Flickr-SoundNet [4] and VGG-Sound [5], and evaluated on three benchmarks, Flickr-SoundNet-Test [4], VGG-SS [5], and AVSBench [36]. It achieved the state of the art in almost all settings across all the three evaluation benchmarks. Further, our model outperformed all the prior work by large margin in cross-modal retrieval, suggesting that it is not a simple saliency detector but properly learns to relate image and audio for SSL [31]. In summary, the contribution of our method is three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.677, + 0.907, + 0.72 + ], + "angle": 0, + "content": "- We propose a novel SSL model that identifies and extracts features relevant to sound source jointly from image and audio modalities through slot attention." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.722, + 0.906, + 0.767 + ], + "angle": 0, + "content": "- We also propose cross-modal attention matching, which enables effective cross-modal alignment for individual input features and thus improves localization performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.768, + 0.905, + 0.812 + ], + "angle": 0, + "content": "- Our model achieved the state of the art in almost all settings across all the three SSL benchmarks and improved cross-modal retrieval performance substantially." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.677, + 0.907, + 0.812 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.829, + 0.655, + 0.844 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Sound source localization aims to identify specific regions in an image that correspond to the sources of sounds from an audio input. Since constructing audio-visual datasets is" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3122" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "costly, sound source localization studies have evolved into self-supervised learning approaches that use only image-audio pairs for contrastive learning [1, 6, 12, 14, 16, 21, 22, 26, 29-31, 33]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.484, + 0.306 + ], + "angle": 0, + "content": "An early study, [29], addresses the task through contrastive learning between the audio-aware image features and the global audio feature. The audio-aware image features are obtained by calculating a correspondence map through the normalized inner product between the global audio feature and the image feature map. Chen et al. [6] propose method that uses a differentiable threshold on the correspondence map to divide audio-aware and other image features, treating the latter as hard negatives in contrastive learning to improve localization." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.308, + 0.484, + 0.625 + ], + "angle": 0, + "content": "Recent studies focus on fine-grained sound source localization through contrastive learning between the global audio feature and the most relevant feature selected from the image feature map [21, 22, 31, 33]. Mo and Morgado [21] propose a method that refines the audio-visual correspondence map through a post-processing technique leveraging a pretrained visual model. Sun et al. [33] introduce a regularizer that applies mean absolute error regularization to inter-modal and intra-modality similarity matrices to increase awareness of false negatives. Senocak et al. [31] propose a cross-modal alignment method that preserves localization performance and enhances cross-modal retrieval by leveraging multi-view augmentation and nearest neighbor search to construct multiple positive pairs for contrastive learning. These methods employ contrastive learning with the most relevant image feature to assist in coarsely identifying sound sources but often fall short of capturing a comprehensive representation, thereby hindering localization ability. Furthermore, relying solely on global audio features can mislead contrastive learning when noise exists, potentially highlighting incorrect image features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Following a complementary direction, several studies employ external prior knowledge or even utilize off-the-shelf segmentation modules to achieve high-quality localization results [25, 26]. Also, several studies employ sound source localization with additional information such as optical flow [9], text inputs [19] or ground truth category [23]. By contrast to these studies, our method can effectively perform sound source localization without relying on external prior knowledge or off-the-shelf models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To address the above issues, we propose a sound source localization framework based on the slot attention mechanism. Slot attention [17] is proposed for unsupervised object-centric learning, effectively decomposing input features into distinct and mutually exclusive components via iterative attention and competition among slots. Inspired by this, we adopt slot attention in both image and audio features to decompose them into features related to the sound source and those irrelevant to it without supervision." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.687, + 0.108 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.327 + ], + "angle": 0, + "content": "As shown in Fig. 2, our model consists of two parts: 1) image and audio encoders for feature extraction, and 2) joint feature decomposition module with joint slot attention mechanism. A pair of image and audio is first fed into each encoder to extract image and audio features (Sec. 3.1). The features of each modality are decomposed into two distinct slot representations by using joint slot attention: one slot for the target and the other for the rest. (Sec. 3.2). The slot representations are then used for training as input to contrastive learning and cross-modal attention matching (Sec. 3.3), while effectively handling false negatives (Sec. 3.4). Lastly, an attention on the image feature map using the target slot representation of the audio is computed to locate the source of sound (Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.335, + 0.696, + 0.35 + ], + "angle": 0, + "content": "3.1. Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Let \\( D = \\{(v_{i},a_{i})\\}_{i = 1}^{M} \\) denote a training set of image-audio pairs, where \\( v_{i}\\in \\mathbb{R}^{3\\times H_{v}\\times W_{v}} \\) is an image of size \\( H_{v}\\times W_{v} \\), \\( a_{i}\\in \\mathbb{R}^{1\\times F_{a}\\times T_{a}} \\) is a corresponding audio spectrogram of frequency resolution \\( F_{a} \\) and temporal length \\( T_{a} \\), and \\( M \\) denotes the total number of such pairs in the dataset. For feature extraction, separate image and audio encoders \\( f_{v} \\) and \\( f_{a} \\), are adopted. The image encoder extracts the image features \\( \\mathbf{v}_i\\coloneqq f_v(v_i)\\in \\mathbb{R}^{h\\times w\\times c} \\) for \\( v_{i} \\), and the audio encoder extracts the audio features \\( \\mathbf{a}_i\\coloneqq f_a(a_i)\\in \\mathbb{R}^{t\\times c} \\) for \\( a_{i} \\) where \\( h \\) and \\( w \\) are the height and width of the image feature map respectively, \\( t \\) is the length of the audio features, and \\( c \\) is the image and audio feature dimension." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.771, + 0.565 + ], + "angle": 0, + "content": "3.2. Joint Feature Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.907, + 0.782 + ], + "angle": 0, + "content": "Slot attention [17] is an attention mechanism where each slot progressively competes to aggregate input features, leading to the decomposition of input features into distinct and exclusive representations. Building on this idea, for joint feature decomposition, we propose joint slot attention (JSA), which jointly decomposes image and audio features into two distinct slot representations for each modality. To achieve this, we employ slot attention with two initial slots \\( S_0 \\in \\mathbb{R}^{2 \\times D} \\) shared in the two modalities. Furthermore, we designate the first slot as target slot, which captures features relevant to sound source, and the other as off-target slot, which captures features for the rest. Since the overall pipeline of JSA is identical in the two modalities, we will describe the module in terms of image only for brevity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.783, + 0.906, + 0.828 + ], + "angle": 0, + "content": "JSA first projects an input image feature \\(\\mathbf{v}\\) to derive its corresponding key \\(K^{\\nu}\\) and value \\(V^{\\nu}\\) by applying the following operations:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.836, + 0.905, + 0.862 + ], + "angle": 0, + "content": "\\[\nK ^ {v} = W ^ {k} (\\operatorname {L a y e r N o r m} (\\mathbf {v})) \\in \\mathbb {R} ^ {h w \\times d}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.858, + 0.839, + 0.875 + ], + "angle": 0, + "content": "\\[\nV ^ {v} = W ^ {v} \\left(\\operatorname {L a y e r N o r m} (\\mathbf {v})\\right) \\in \\mathbb {R} ^ {h w \\times d},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.885, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( W^{k} \\), \\( W^{v} \\) are the linear projection layers to obtain" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3123" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.138 + ], + "angle": 0, + "content": "embeddings in \\(d\\)-dimension. For each iteration \\(n \\in [1, N]\\), the query \\(Q_{n}^{v}\\) is derived by applying linear projection to the slots with the initial slots \\(S_{0}^{v} \\coloneqq S_{0}\\), which is given by" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.147, + 0.483, + 0.166 + ], + "angle": 0, + "content": "\\[\nQ _ {n} ^ {v} = W ^ {q} \\left(\\operatorname {L a y e r N o r m} \\left(S _ {n - 1} ^ {v}\\right)\\right) \\in \\mathbb {R} ^ {2 \\times d}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.175, + 0.484, + 0.237 + ], + "angle": 0, + "content": "where \\( W^q \\) is a linear projection layer shared for all iterations. The slot attention matrix \\( A \\) is computed by taking the dot product of the given query \\( Q_n^v \\) and key \\( K^v \\), scaled by the square root of the dimension \\( d \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.245, + 0.49, + 0.294 + ], + "angle": 0, + "content": "\\[\nA _ {i, j} = \\frac {\\exp \\left(M _ {i , j}\\right)}{\\sum_ {l = 1} ^ {2} \\exp \\left(M _ {i , l}\\right)}, \\text {w h e r e} M = \\frac {K ^ {v} \\left(Q _ {n} ^ {v}\\right) ^ {\\top}}{\\sqrt {d}} \\in \\mathbb {R} ^ {h w \\times 2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.295, + 0.484, + 0.326 + ], + "angle": 0, + "content": "where \\( i \\) is key index, and \\( j \\) is slot index. Subsequently, a weighted mean is utilized to the slot attention matrix \\( A \\):" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.334, + 0.483, + 0.371 + ], + "angle": 0, + "content": "\\[\n\\hat {A} _ {i, j} = \\frac {A _ {i , j}}{\\sum_ {l = 1} ^ {h w} A _ {l , j}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.426 + ], + "angle": 0, + "content": "Then, the updated slot \\(\\bar{S}_n\\) is computed by Gated Recurrent Unit (GRU) [7] with \\(S_{n - 1}\\) as hidden state, and \\(\\hat{A}^{\\top}\\cdot V^{v}\\) as input as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.435, + 0.483, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\bar {S _ {n} ^ {v}} = \\operatorname {G R U} \\left(S _ {n - 1} ^ {v}, \\hat {A} ^ {\\top} \\cdot V ^ {v}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.436, + 0.48 + ], + "angle": 0, + "content": "Finally, the slot is updated by a residual connection:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.488, + 0.483, + 0.508 + ], + "angle": 0, + "content": "\\[\nS _ {n} ^ {v} = \\bar {S} _ {n} ^ {v} + \\operatorname {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(\\bar {S} _ {n} ^ {v}\\right)\\right) \\in \\mathbb {R} ^ {2 \\times d}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.701 + ], + "angle": 0, + "content": "The overall process encourages the input feature to bind to specific slots, ensuring that each slot forms a mutually exclusive representation. The final outputs of JSA are image slots \\(\\{\\mathbf{p}^v;\\mathbf{r}^v\\} := S_N^v\\in \\mathbb{R}^{2\\times d}\\) and image query \\(Q_{N}^{v}\\in\\) \\(\\mathbb{R}^{2\\times d}\\), where \\(\\mathbf{p}^v\\) is a target slot and \\(\\mathbf{r}^v\\) is an off-target slot. The same process is applied to the audio modality using audio feature \\(\\mathbf{a}\\), yielding audio slots \\(\\{\\mathbf{p}^a;\\mathbf{r}^a\\} := S_N^a\\in \\mathbb{R}^{2\\times d}\\) and audio query \\(Q_{N}^{a}\\in \\mathbb{R}^{2\\times d}\\). Since this process begins from the shared initial slots, the initial slots are able to effectively learn joint features between two modalities. For brevity, we omit the subscript \\(N\\) for the final query, \\(Q^{v}\\coloneqq Q_{N}^{v}\\) and \\(Q^{a}\\coloneqq Q_{N}^{a}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.707, + 0.284, + 0.723 + ], + "angle": 0, + "content": "3.3. Learning Objectives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.484, + 0.838 + ], + "angle": 0, + "content": "Contrastive Learning with Target Slots. We adopt the InfoNCE loss [24, 27] for contrastive learning using the target slots, guiding the target slots to identify features relevant to the sound source jointly from image and audio. Let \\( s \\) denote a similarity function, \\( s(x,y) = \\exp (\\cos (x,y) / \\tau) \\), where \\( \\tau \\) is a temperature hyperparameter. The contrastive learning loss \\( \\mathcal{L}_{\\mathrm{cotr}} \\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.847, + 0.498, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o t r}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left\\{\\log \\frac {s \\left(\\mathbf {p} _ {i} ^ {v} , \\mathbf {p} _ {i} ^ {a}\\right)}{\\sum_ {j} ^ {B} s \\left(\\mathbf {p} _ {i} ^ {v} , \\mathbf {p} _ {j} ^ {a}\\right)} + \\log \\frac {s \\left(\\mathbf {p} _ {i} ^ {v} , \\mathbf {p} _ {i} ^ {a}\\right)}{\\sum_ {j} ^ {B} s \\left(\\mathbf {p} _ {j} ^ {v} , \\mathbf {p} _ {i} ^ {a}\\right)} \\right\\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.227 + ], + "angle": 0, + "content": "where \\( B \\) denotes the number of data in a mini-batch, and the negative pairs refer to the combinations of visual and audio features that do not belong to the same original pair within the mini-batch. Since the image target slot captures the whole features of the target region, it avoids focusing solely on a specific discriminative part of the image. Furthermore, the audio target slot captures only the relevant sound from the source, it prevents the issue of learning sounds that are unrelated to the source." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.907, + 0.59 + ], + "angle": 0, + "content": "Cross-modal Attention Matching. As will be described in Sec. 3.5, the localization result is derived by utilizing cross-modal attention between the audio target slot and image features. For accurate sound source localization, the audio target slot should be properly aligned with keys computed from the relevant image features. Although contrastive learning encourages alignment of target slot representations between different modalities, it does not ensure alignment between the audio target slot and image features, which degrades the localization accuracy. We resolve this issue by cross-modal attention matching, which encourages the cross-modal attention map to resemble the intra-modal attention map for improving feature-level alignment across the two modalities. To this end, we first calculate the cross-modal and intra-modal attention matrices. Specifically, we compute the cross-modal attention matrix between the audio and visual modalities in the same manner as Eq. (3) and Eq. (4), with the audio queries \\( Q^a \\) and visual keys \\( K^v \\). We denote this cross-modal attention matrix as \\( \\hat{A}^{a,v} \\in \\mathbb{R}^{hw \\times 2} \\), where \\( hw \\) is the length of \\( K^v \\). Subsequently, to obtain the cross-modal attention that represents the attention weights between the audio target query and visual keys, we utilize only the first column of \\( \\hat{A}^{a,v} \\). This cross-modal attention is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.597, + 0.8, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c a} ^ {a, v} := \\left[ \\hat {A} _ {i, 1} ^ {a, v} \\right] _ {i = 1} ^ {h w} \\in \\mathbb {R} ^ {h w}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.906, + 0.699 + ], + "angle": 0, + "content": "Similarly, for intra-modal attention matrix within the visual modality, we compute the intra-attention matrix \\(\\hat{A}^{v,v}\\) using the \\(Q^v\\) and \\(K^v\\). Then, the intra-modal attention that represents the attention weights between the image target query and visual keys is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.708, + 0.8, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathbf {i a} ^ {v, v} := \\left[ \\hat {A} _ {i, 1} ^ {v, v} \\right] _ {i = 1} ^ {h w} \\in \\mathbb {R} ^ {h w}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.739, + 0.906, + 0.799 + ], + "angle": 0, + "content": "To further facilitate cross-modal alignment, we apply this strategy symmetrically to the audio modality as well. Finally we formulate the cross-modal attention matching loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.811, + 0.907, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a t c h}} = \\left\\| \\mathbf {c a} ^ {a, v} - \\operatorname {s g} \\left(\\mathbf {i a} ^ {v, v}\\right) \\right\\| _ {2} ^ {2} + \\left\\| \\mathbf {c a} ^ {v, a} - \\operatorname {s g} \\left(\\mathbf {i a} ^ {a, a}\\right) \\right\\| _ {2} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathrm{sg}(\\cdot)\\) denotes the stop-gradient operation. \\(\\mathcal{L}_{\\mathrm{match}}\\) minimizes the discrepancy between cross-modal and intramodal attentions, promoting robust feature-level alignment between the two modalities for accurate localization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3124" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "Slot Divergence. In slot attention, the diversity between the target slot and off-target slot should be large enough to ensure that each slot effectively captures distinct, mutually exclusive regions. Therefore, we introduce a slot divergence loss that encourages greater diversity among the slots. The slot divergence loss \\(\\mathcal{L}_{\\mathrm{div}}\\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.188, + 0.483, + 0.205 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d i v}} = \\max \\left\\{0, \\cos \\left(\\mathbf {p} ^ {v}, \\mathbf {r} ^ {v}\\right) \\right\\} + \\max \\left\\{0, \\cos \\left(\\mathbf {p} ^ {a}, \\mathbf {r} ^ {a}\\right) \\right\\}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.21, + 0.482, + 0.255 + ], + "angle": 0, + "content": "It ensures that the different slot embeddings to be sufficiently dissimilar, enabling the target slot to more effectively localize the sound source." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.483, + 0.376 + ], + "angle": 0, + "content": "Slot Reconstruction. The successful feature reconstruction in slot attention implies that each slot successfully captures unique and independent parts of the input data, which means that the slot attention mechanism effectively decomposed the input into individual slot representations. It indicates that each slot accurately represents a specific object or region in the input. To encourage this condition, we apply the slot reconstruction loss \\(\\mathcal{L}_{\\mathrm{recon}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.381, + 0.483, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\| \\mathbf {v} - g _ {v} \\left(\\left\\{\\mathbf {p} ^ {v}; \\mathbf {r} ^ {v} \\right\\}\\right) \\| _ {2} ^ {2} + \\| \\mathbf {a} - g _ {a} \\left(\\left\\{\\mathbf {p} ^ {a}; \\mathbf {r} ^ {a} \\right\\}\\right) \\| _ {2} ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.483, + 0.478 + ], + "angle": 0, + "content": "where \\( g_{v} \\) and \\( g_{a} \\) are the decoders for the image and audio features, respectively. These decoders are composed of a small number of MLP layers with ReLU activations. Note that the decoders are auxiliary modules used only in training and not required for inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.483, + 0.51 + ], + "angle": 0, + "content": "The overall loss is a combination of \\(\\mathcal{L}_{\\mathrm{cotr}}\\), \\(\\mathcal{L}_{\\mathrm{match}}\\), \\(\\mathcal{L}_{\\mathrm{div}}\\), and \\(\\mathcal{L}_{\\mathrm{recon}}\\), which is given by" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.516, + 0.482, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o t r}} + \\lambda_ {1} \\mathcal {L} _ {\\text {m a t c h}} + \\lambda_ {2} \\mathcal {L} _ {\\text {d i v}} + \\lambda_ {3} \\mathcal {L} _ {\\text {r e c o n}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.566 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\), \\(\\lambda_{2}\\), and \\(\\lambda_{3}\\) are hyperparameters weighting the loss terms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.331, + 0.59 + ], + "angle": 0, + "content": "3.4. Mitigating False Negatives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.688 + ], + "angle": 0, + "content": "In general, unpaired images and audiios are all treated as negatives for contrastive training. However, some of them are false negatives that are semantically well matched, so treating these as negatives causes adverse effect [33]. Therefore, we propose a false negative mitigation strategy using \\( k \\)-reciprocal nearest neighbors of target slots." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.688, + 0.483, + 0.72 + ], + "angle": 0, + "content": "For the image target slot \\(\\mathbf{p}_i^v\\), which is derived from \\(v_{i}\\) we define its set of \\(k\\)-reciprocal nearest neighbors \\(R_{k}^{v}(i)\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.724, + 0.483, + 0.742 + ], + "angle": 0, + "content": "\\[\nR _ {k} ^ {v} (i) = \\left\\{\\left(v _ {j}, a _ {j}\\right) \\mid \\left(j \\in N _ {k} ^ {v} (i)\\right) \\cap \\left(i \\in N _ {k} ^ {v} (j)\\right) \\right\\}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.483, + 0.82 + ], + "angle": 0, + "content": "where \\(N_{k}^{v}(i)\\) is the \\(k\\)-nearest neighbors of \\(\\mathbf{p}_i^v\\), computed using cosine similarity. Likewise, the set of \\(k\\)-reciprocal nearest neighbors between audio target slots, i.e., \\(R_{k}^{a}(i)\\), can be derived. The predicted false negative set \\(R_{k}(i)\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.822, + 0.483, + 0.838 + ], + "angle": 0, + "content": "\\[\nR _ {k} (i) = R _ {k} ^ {v} (i) \\cap R _ {k} ^ {a} (i). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Since the samples in \\( R_{k}(i) \\) have both similar image and audio target slots with \\( (v_{i}, a_{i}) \\), they are likely to belong to false negatives. We exclude these predicted false negatives from the contrastive learning in Eq. (7)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.625, + 0.106 + ], + "angle": 0, + "content": "3.5. Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.905, + 0.189 + ], + "angle": 0, + "content": "For inference, \\(\\mathbf{ca}^{a,v}\\), which represents the attention between the audio target slot and the input image features, is utilized. Specifically, the attention is interpolated to the original image size, and pixels whose score exceeds the hyperparameter \\(\\theta\\) are selected to generate the localization result." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.19, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Furthermore, unlike the object-guided localization [21], we propose a novel localization refinement strategy, dubbed Image-Query based Refinement (IQR), to refine the localization result without relying on external prior knowledge. IQR estimates the regions that are likely to contain the target by using the attention between the image target slot and the input image features. It can be used to identify prior regions in the image that favor the target, regardless of whether these regions are actually the source of the sound. The refined result derived by IQR, denoted as \\( R_{\\mathrm{IQR}} \\), is given as" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.354, + 0.905, + 0.371 + ], + "angle": 0, + "content": "\\[\nR _ {\\mathrm {I Q R}} = \\alpha \\cdot \\mathbf {c a} ^ {a, v} + (1 - \\alpha) \\cdot \\mathbf {i a} ^ {v, v}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.905, + 0.414 + ], + "angle": 0, + "content": "where \\(\\alpha \\in (0,1)\\) balances the sound source localization result and potential sound source regions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.429, + 0.64, + 0.446 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.454, + 0.691, + 0.47 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Datasets. Flickr-SoundNet [4] and VGG-Sound [5] are used to train. Flickr-SoundNet contains about 2 million unconstrained videos from Flickr, while VGG-Sound has around 200k video clips. For training, we use subsets of 10k and 144k from each dataset, respectively. For the evaluation, Flickr-SoundNet-Test [4], VGG-SS [5], and AVSBench [36] are used following the previous research [6, 21, 31, 33]. Flickr-SoundNet-Test has 250 audio-visual pairs with manually labeled bounding boxes. VGG-SS has about 5k audio-visual pairs over 220 categories. Since VGG-SS provides its category ground truth, it is also used to evaluate cross-modal retrieval performance. AVSBench has about 5k videos with pixel-wise binary segmentation maps. Following the previous study [6], the middle frame of the video clip is used for the image. For the audio, the middle 5 seconds are extracted and converted into log spectrograms with 257 frequency bins at a sampling rate of \\(16\\mathrm{kHz}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Network Architecture. ResNet-18 [11] is adopted for both the image encoder and the audio encoders. The image encoder is fine-tuned from an ImageNet [8] pretrained model, while the audio encoder is trained from scratch. Both encoders include an additional \\( 1 \\times 1 \\) convolutional layer for projection, and the audio encoder also includes a MaxPooling layer for pooling along the frequency axis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Hyperparameter Settings. We use the AdamW optimizer [18] with a learning rate of 5e-5 and a weight decay of 1e-2. The mini-batch size during training is set to 256, and the size of image and audio feature size \\( h \\), \\( w \\), \\( t \\) and \\( c \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3125" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.493, + 0.364 + ], + "angle": 0, + "content": "
TrainsetMethodOGLVGG-SSFlickr-SoudNet
cIoUAUCcIoUAUC
Flickr SoundNet 10kAttention10k [29]--43.6044.90
CoursetoFine [26]--52.2049.60
AVObject [1]--54.6050.40
LVS [6]--58.2052.50
EZ-VSL [21]19.8630.9662.2454.74
FNAC [33]35.2738.0084.3363.26
Ours35.1138.6685.2065.26
EZ-VSL [21]✓37.6139.2181.9362.58
FNAC [33]✓40.9740.3884.7364.34
Ours✓39.1640.5187.6066.50
Flickr SoundNet 144kAttention10k [29]--66.0055.80
DMC [12]--67.1056.80
LVS [6]--69.9057.30
HardPos [30]--75.2059.70
EZ-VSL [21]30.2735.9272.6958.70
FNAC [33]33.9337.2978.7159.33
Alignment [31]--85.2062.20
Ours37.0139.4286.0065.16
EZ-VSL [21]✓41.0140.2383.1363.06
FNAC [33]✓41.1040.4483.9363.06
Alignment [31]✓--84.0064.16
Ours✓40.1540.7589.2064.50
" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.363 + ], + "angle": 0, + "content": "
TrainsetMethodOGLVGG-SSFlickr-SoudNet
cIoUAUCcIoUAUC
VGGSound 10kLVS [6]--61.8053.60
EZ-VSL [21]25.8433.6863.8554.44
FNAC [33]37.2938.9985.7463.66
Ours37.8439.8286.0064.90
EZ-VSL [21]✓38.7139.8078.7161.53
FNAC [33]✓40.6940.4282.1363.64
Ours✓41.0241.2987.6065.42
VGGSound 144kAttention10k [29]18.5030.20--
DMC [12]29.1034.80--
AVObject [1]29.7035.70--
LVS [6]34.4038.2073.5059.00
HardPos [30]34.6038.0076.8059.20
EZ-VSL [21]34.3837.7079.5161.17
SLAVC [22]37.22-83.20-
FNAC [33]39.5039.6684.7363.76
Alignment [31]39.9440.0279.6063.44
NoPrior [14]41.441.2--
Ours40.6241.6683.6064.68
EZ-VSL [21]✓38.8539.5483.9463.60
SLAVC [22]✓39.67-86.40-
FNAC [33]✓41.8540.8085.1464.30
Alignment [31]✓42.6441.4882.4064.60
Ours✓42.5642.4684.8064.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.373, + 0.907, + 0.416 + ], + "angle": 0, + "content": "Table 1. Quantitative results of the model trained on Flickr-SoundNet-10k, 144k, and VGGSound-10k, 144k datasets. Note that the results of previous research are obtained from [14, 31, 33]. The results of our method represent the performance after applying Image-Query based Refinement (IQR). The result with OGL denotes the performance after applying object-guided localization [21]." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.429, + 0.877, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.612, + 0.908, + 0.655 + ], + "angle": 0, + "content": "Figure 3. Sound localization results on Flickr-SoundNet-Test [4] and VGG-SS [5]. (a) Input image. (b) Ground-Truth. (c) Ours. (d) Alignment [31]. (e) FNAC [33]. (f) EZ-VSL [21]. The qualitative results are obtained by the model trained on Flickr-144k and the model trained on VGGSound-144k, respectively. Note that all visualizations are obtained without refinement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.483, + 0.755 + ], + "angle": 0, + "content": "are set to 7, 7, 16, and 512 respectively. \\(\\tau\\) in Eq. (7) is set to 0.03. \\(\\lambda_{1}, \\lambda_{2}\\), and \\(\\lambda_{3}\\) in Eq. (11) are set to 100.0, 0.1, and 0.1, respectively. \\(\\alpha\\) in Eq. (14) is set to 0.6. The value \\(k\\) in Eq. (13) is set to 20. The number of iterations \\(N\\) for slot attention mechanism is set to 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.832 + ], + "angle": 0, + "content": "Evaluation metric. We measure the average precision at a Consensus Intersection over Union (cIoU) threshold of 0.5 to evaluate localization accuracy. Additionally, the AUC score is used to provide an overall performance measure across different thresholds." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.285, + 0.862 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Flickr-SoundNet and VGGSound. In this section, we compare our method with previous self-supervised SSL" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.68, + 0.907, + 0.846 + ], + "angle": 0, + "content": "approaches. We trained our model on subsets of Flickr-SoundNet and VGGSound with dataset sizes of 10k and 144k, and evaluated the performance on Flickr-SoundNet-Test and VGG-SS. As shown in Table 1, our model achieves the state-of-the-art performance in most settings. Specifically, our method yields AUC improvements across all evaluated settings. Moreover, even without using Object-Guided Localization (OGL), our model performs competitively or better compared to previous methods, demonstrating the robustness of our architecture in focusing on the sound source without additional refinement steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Heard 110 and Unheard 110. To assess the generalization ability of SSL, we evaluate the performance with Heard 110 and Unheard 110. Heard 110 and Unheard 110 are sub" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3126" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.09, + 0.473, + 0.264 + ], + "angle": 0, + "content": "
MethodOGLHeard 110Unheard 110
cIoUAUCcIoUAUC
LVS [6]28.9036.2026.3034.70
EZ-VSL [21]31.8636.1932.6636.72
SLAVC [22]35.84-36.50-
Alignment [31]38.3139.0539.1139.80
Ours38.4640.1540.3541.25
EZ-VSL [21]✓37.2538.9739.5739.60
SLAVC [22]✓38.22-38.87-
FNAC [33]✓39.5439.8342.9141.17
Alignment [31]✓41.8540.9342.9441.54
Ours✓41.5041.6043.1142.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.275, + 0.483, + 0.358 + ], + "angle": 0, + "content": "Table 2. Quantitative results on Heard 110 and Unheard 110. Note that the results of previous research are obtained from [31, 33]. The results of our method represent the performance after applying Image-Query based Refinement (IQR). The result with OGL denotes the performance after applying object-guided localization." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.377, + 0.441, + 0.505 + ], + "angle": 0, + "content": "
TestsetMethodmIoUF-score
S4LVS [6]26.933.6
EZ-VSL [21]27.634.2
SLAVC [22]28.134.6
FNAC [33]27.1531.4
Alignment [31]29.635.9
\\( Alignment^† \\)29.335.6
Ours31.3344.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.515, + 0.483, + 0.558 + ], + "angle": 0, + "content": "Table 3. Quantitative results on AVSBench. Note that the results of previous research are obtained from [31, 33]. \\(\\dagger\\) denotes that the self-supervised pretrained encoder is utilized for their work." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.576, + 0.481, + 0.699 + ], + "angle": 0, + "content": "
MethodAudio → ImageImage → Audio
R@1R@5R@10R@1R@5R@10
LVS [6]3.8712.3520.734.9014.2921.37
EZ-VSL [21]5.0115.7324.8114.2033.5145.18
SSL-TIE [16]10.2930.6843.7612.7629.5839.72
SLAVC [22]4.7713.0819.106.1221.1632.12
Alignment [31]16.4736.9949.0020.0942.3853.66
\\( Alignment^† \\)14.3137.8149.1718.0038.3949.02
Ours30.6157.8769.3731.7058.4369.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.708, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Table 4. Cross-modal retrieval on the VGG-SS. All of the models are trained on VGGSound-144K. Note that the results of previous research are obtained from [31]. \\(\\dagger\\) denotes that the self-supervised pretrained encoder is utilized for their work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "sets of the VGG-SS, each containing 110 non-overlapping categories. During training, the model is trained only to the category set of one group. The subset comprising categories seen during training is called Heard 110, while the subset with unseen categories is Unheard 110. As shown in Table 2, our model achieves the state-of-the-art performance on Unheard 110 both without additional refinement" + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.09, + 0.885, + 0.226 + ], + "angle": 0, + "content": "
Trainset\\( {\\mathcal{L}}_{\\text{match }} \\)\\( {\\mathcal{L}}_{\\text{div }} \\)FN MitigateVGG-SS
cIoUAUC
VGGSound 144k✓14.9527.02
✓36.2339.13
18.6930.03
✓15.0328.18
✓✓40.5641.35
✓✓36.1439.31
✓✓20.3730.81
✓✓✓40.7141.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.577, + 0.235, + 0.843, + 0.249 + ], + "angle": 0, + "content": "Table 5. Detailed ablation study for losses." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.265, + 0.903, + 0.354 + ], + "angle": 0, + "content": "
TrainsetShared Initial SlotsVGG-SSFlickr
cIoUAUCcIoUAUC
Flickr-144k✓34.8038.7284.0061.70
36.6439.2486.8065.44
VGGSound-144k✓33.1137.9281.2064.36
40.7141.6284.4064.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.565, + 0.364, + 0.855, + 0.378 + ], + "angle": 0, + "content": "Table 6. Ablation study on shared initial slots." + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.393, + 0.892, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.635, + 0.907, + 0.692 + ], + "angle": 0, + "content": "Figure 4. Qualitative results to show the impact of \\(\\mathcal{L}_{\\mathrm{match}}\\) on Flickr-SoundNet. (a) Input image. (b) Ground-Truth. (c) Results without cross-modal attention matching. (d) Results with cross-modal attention matching." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.907, + 0.763 + ], + "angle": 0, + "content": "and with OGL applied. These results demonstrate that our model has superior generalization ability compared to previous models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.87 + ], + "angle": 0, + "content": "AVSBench. To more precisely evaluate the performance of the localization mask, we measured the performance on AVSBench, which provides segmentation masks. Our model was trained on VGGSound-144k and evaluated in a zero-shot manner on AVSBench. As shown in Table 3, we achieved the state-of-the-art results, specifically with improvements of \\(2.03\\%\\) p in mIoU and \\(8.7\\%\\) p in F-score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Cross-modal Retrieval. As discussed in the previous study [31], the benchmark evaluation metrics, cIoU and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3127" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.089, + 0.865, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.28, + 0.704, + 0.294 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of cross-modal retrieval on VGG-SS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.485, + 0.534 + ], + "angle": 0, + "content": "AUC, do not guarantee better audio-visual semantic understanding. Therefore, we also evaluate cross-modal retrieval performance on VGG-SS. We extract the image and audio target slot representations and compute the cosine similarity between them. The similarities are ranked in descending order, and we calculate Recall@K performance using the category labels from the top-K retrieved audio or images. In Table 4, our model demonstrates a significant performance improvement over the previous state of the art in cross-modal retrieval. Specifically, our model improves Recall@1 by \\(16.3\\%\\) p in audio-to-image retrieval and \\(13.7\\%\\) p in image-to-audio retrieval. This result indicates that our model does not merely detect salient objects likely to produce sound within the image, but rather effectively locates the target specified by the audio input." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.543, + 0.245, + 0.559 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Impact of Losses and False Negative Mitigating in Sec. 3.4. In Table 5, we conduct an ablation study to verify the effectiveness of each loss function and the strategy outlined in Sec. 3.4. The overall results show that the designed losses and the false negative mitigating strategy each contribute to improve the performance. Notably, the experiment with \\(\\mathcal{L}_{\\mathrm{match}}\\) demonstrates a drastic increase in performance, highlighting the importance of cross-modality attention matching. Additionally, the ablation study on \\(\\mathcal{L}_{\\mathrm{div}}\\) shows that explicitly separating the slots enables each slot to perform its designated role more effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Impact of Shared Initial Slots. We conduct an ablation study to examine the impact of shared initial slots by using different initial slots for each modality. As shown in Table 6, sharing initial slots resulted in better performance, indicating that this design choice facilitates the ability to jointly capture sound source-related information across different modalities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.274, + 0.863 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Quality of SSL results. Fig. 3 visualizes the qualitative results of our model and previous studies on Flickr-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.907, + 0.397 + ], + "angle": 0, + "content": "SoundNet [4] and VGG-SS [5] test sets. These results demonstrate that our model has the ability to effectively localize sound source objects. Notably, it produces compact localization results even for challenging samples that other models struggle with, such as small birds, a gun, and cars surrounded by other objects in street scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.518 + ], + "angle": 0, + "content": "Qualitative Analysis of \\(\\mathcal{L}_{\\mathrm{match}}\\). Fig. 4 shows the qualitative result that shows the effect of \\(\\mathcal{L}_{\\mathrm{match}}\\) on Flickr-SoundNet-Test. The attention maps is generated between the audio query of the audio target slot and the image keys derived from image features. Without \\(\\mathcal{L}_{\\mathrm{match}}\\), the alignment between the two is insufficient for compact localization, whereas with \\(\\mathcal{L}_{\\mathrm{match}}\\), the alignment improves, resulting in more accurate localization of the sound source." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.609 + ], + "angle": 0, + "content": "Cross-modal Retrieval. Fig. 5 shows the qualitative results of our model for cross-modal retrieval. Consistent with the impressive performance shown in Table 4, our model demonstrates a strong ability to retrieve semantically similar samples across modalities that correspond to the sound source target." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.622, + 0.634, + 0.637 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.907, + 0.812 + ], + "angle": 0, + "content": "This paper presented a novel self-supervised sound source localization method leveraging joint slot attention to align image and audio features. Our approach decomposes these features into target and off-target slots, using only target slots for contrastive learning to capture sound-related features. In addition, we introduced cross-modal attention matching and slot divergence loss which significantly improves localization accuracy and and we demonstrated the effectiveness of our model through comprehensive analysis. For future work, we aim to extend our framework to multisource sound source localization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.814, + 0.907, + 0.888 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the NRF grant (RS-2021-NR059830-40%) and the IITP grants (RS-2024-00509258-40%, RS-2021-II212068-10%, RS-2019-II191906-10%) funded by Ministry of Science and ICT, Korea." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3128" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In Proc. European Conference on Computer Vision (ECCV), 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Stuart Andrews, Ioannis Tsochantaridis, and Thomas Hofmann. Support vector machines for multiple-instance learning. Advances in neural information processing systems, 15, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.269 + ], + "angle": 0, + "content": "[3] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proc. European Conference on Computer Vision (ECCV), pages 435-451, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.482, + 0.324 + ], + "angle": 0, + "content": "[4] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Sound-net: Learning sound representations from unlabeled video. Proc. Neural Information Processing Systems (NeurIPS), 29, 2016. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.482, + 0.464 + ], + "angle": 0, + "content": "[6] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.482, + 0.576 + ], + "angle": 0, + "content": "[7] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724-1734, Doha, Qatar, 2014. Association for Computational Linguistics. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: a large-scale hierarchical image database. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[9] Dennis Fedorishin, Deen Dayal Mohan, Bhavin Jawade, Srirangaraj Setlur, and Venu Govindaraju. Hear the flow: Optical flow-based self-supervised visual sound source localization. In Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV), pages 2278-2287, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[10] Yury Furletov, Volker Willert, and Jürgen Adamy. Auditory scene understanding for autonomous driving. In 2021 IEEE Intelligent Vehicles Symposium (IV), pages 697-702. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[12] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In Proc. In-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ternational Conference on Machine Learning (ICML), pages 2127-2136. PMLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.124, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[14] Dongjin Kim, Sung Jin Um, Sangmin Lee, and Jung Uk Kim. Learning to visually localize sound sources from mixtures without prior source knowledge. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 26467-26476, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[15] Chen Liu, Peike Patrick Li, Xingqun Qi, Hu Zhang, Lincheng Li, Dadong Wang, and Xin Yu. Audio-visual segmentation by exploring cross-modal mutual semantics. In Proc. ACM Multimedia Conference (ACMMM), pages 7590–7598, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.266, + 0.905, + 0.321 + ], + "angle": 0, + "content": "[16] Jinxiang Liu, Chen Ju, Weidi Xie, and Ya Zhang. Exploiting transformation invariance and equivariance for self-supervised sound localisation. In Proc. ACM Multimedia Conference (ACMMM), pages 3742-3753, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.324, + 0.905, + 0.405 + ], + "angle": 0, + "content": "[17] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Proc. Neural Information Processing Systems (NeurIPS), 33:11525-11538, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.409, + 0.905, + 0.45 + ], + "angle": 0, + "content": "[18] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. International Conference on Learning Representations (ICLR), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.453, + 0.905, + 0.508 + ], + "angle": 0, + "content": "[19] Tanvir Mahmud, Yapeng Tian, and Diana Marculescu. T-vsl: Text-guided visual sound source localization in mixtures. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 26742-26751, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[20] Oded Maron and Tomás Lozano-Pérez. A framework for multiple-instance learning. Advances in neural information processing systems, 10, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.555, + 0.905, + 0.595 + ], + "angle": 0, + "content": "[21] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In Proc. European Conference on Computer Vision (ECCV), 2022. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.599, + 0.905, + 0.64 + ], + "angle": 0, + "content": "[22] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. In Proc. Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.643, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[23] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10565–10574, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.7, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[24] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.744, + 0.905, + 0.797 + ], + "angle": 0, + "content": "[25] Sooyoung Park, Arda Senocak, and Joon Son Chung. Can clip help sound source localization? In Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV), 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[26] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In Proc. European Conference on Computer Vision (ECCV), 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3129" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[28] Yoko Sasaki, Ryo Tanabe, and Hiroshi Takernura. Online spatial sound perception using microphone array on mobile robot. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2478-2484. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[29] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[30] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Learning sound localization better from semantically similar samples. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[31] Arda Senocak, Hyeonggon Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[32] Zengjie Song, Yuxi Wang, Junsong Fan, Tieniu Tan, and Zhaoxiang Zhang. Self-supervised predictive learning: A negative-free method for sound source localization in visual scenes. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3222-3231, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[33] Weixuan Sun, Jiayi Zhang, Jianyuan Wang, Zheyuan Liu, Yiran Zhong, Tianpeng Feng, Yandong Guo, Yanhao Zhang, and Nick Barnes. Learning audio-visual source localization via false negative aware contrastive learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[34] Jun Yin, Stefano Damiano, Marian Verhelst, Toon van Watershoot, and Andre Guntoro. Real-time acoustic perception for automotive applications. In 2023 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1-6. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.483, + 0.669 + ], + "angle": 0, + "content": "[35] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep features for discriminative localization. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[36] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In Proc. European Conference on Computer Vision (ECCV), pages 386–403, 2022. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3130" + } + ] +] \ No newline at end of file diff --git a/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/feaa7f56-7770-404f-a305-b468e86df1e5_origin.pdf b/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/feaa7f56-7770-404f-a305-b468e86df1e5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..312136992eda20aa2b48a9a75d5607fbb4c44b76 --- /dev/null +++ b/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/feaa7f56-7770-404f-a305-b468e86df1e5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c64a99bd84114d0a4c0f3435f4b55013eaf55249ec0376f4cc86d013adbe231f +size 2087523 diff --git a/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/full.md b/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/full.md new file mode 100644 index 0000000000000000000000000000000000000000..be53d22bad3af44e945134e904815aa69cf3e31d --- /dev/null +++ b/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/full.md @@ -0,0 +1,333 @@ +# Improving Sound Source Localization with Joint Slot Attention on Image and Audio + +Inho Kim1 Youngkil Song2 Jicheol Park2 Won Hwa Kim1,2 Suha Kwak1,2 + +$^{1}$ Dept. of CSE, POSTECH $^{2}$ Graduate School of AI, POSTECH + +{kimih, songyk, jicheol, wonhwa, suha.kwak}@postech.ac.kr + +# Abstract + +Sound source localization (SSL) is the task of locating the source of sound within an image. Due to the lack of localization labels, the de facto standard in SSL has been to represent an image and audio as a single embedding vector each, and use them to learn SSL via contrastive learning. To this end, previous work samples one of local image features as the image embedding and aggregates all local audio features to obtain the audio embedding, which is far from optimal due to the presence of noise and background irrelevant to the actual target in the input. We present a novel SSL method that addresses this chronic issue by joint slot attention on image and audio. To be specific, two slots competitively attend image and audio features to decompose them into target and off-target representations, and only target representations of image and audio are used for contrastive learning. Also, we introduce cross-modal attention matching to further align local features of image and audio. Our method achieved the best in almost all settings on three public benchmarks for SSL, and substantially outperformed all the prior work in cross-modal retrieval. + +# 1. Introduction + +Humans easily identify where a sound comes from within their sight. This capability is a key aspect of human perception facilitating interaction with surroundings, and is also demanded in a large variety of vision applications such as autonomous driving and robotics [10, 28, 34]. Motivated by this, a large body of research [1, 3, 6, 12, 14, 15, 21, 22, 25, 26, 29-33] has studied sound source localization (SSL), the task of locating the source of a sound within the paired image in the form of segmentation mask. + +Due to the prohibitively expensive cost of sound source annotation for image-audio pair data, existing methods [6, 21, 31, 33] have focused on learning SSL using only image-audio pairs without sound source labels. To this end, they consider the task as multiple instance learning [2, 13, 20] + +![](images/0edefc1447d60faec279218077c61fc689637c13d81c5c49cc69ea278f071b45.jpg) +(a) Previous work +Figure 1. Difference between previous work and ours. For contrastive learning of SSL, our method identifies and utilizes pairs of relevant image region and audio segment, while previous work exploits a local image feature and the global audio feature. + +![](images/cee2e6e0380f9fe5d0beeff613f7c05e3b77392cfb3ea91e72dbb29f5a6f7095.jpg) +(b) Ours + +and assume that, given an image-audio pair, at least one patch of the image is relevant to the audio; then they represent the image with its local feature that most closely resembles the global audio feature, and learn SSL through contrastive learning [24, 27] with these features (Fig. 1(a)). + +While this approach has been proven to work, there is plenty of room for further improvement. First, audio may contain not only the sound of the target shared by the two modalities but also irrelevant sounds, such as noise or sounds from sources outside the image. In such cases, using the global audio feature for contrastive learning hinders accurate alignment between the image and audio features. Second, representing an image with only a single local feature in contrastive learning causes models to locate a small discriminative part of the sound source, not its whole body [35], which can degrade SSL accuracy. + +To address these limitations, we propose a novel SSL method based on slot attention [17], which is conceptually illustrated in Fig. 1(b). In slot attention, slots, which are randomly sampled or learnable embeddings, compete with each other for aggregating input features progressively, and thus they decompose the features into distinct and exclusive representations. Motivated by this, we employ slot attention with two learnable slots to jointly decompose image + +![](images/9b4b6f866abc2ae0590ba4abe78b308ca4a7cf7e863535c21f098f90549188fa.jpg) +Figure 2. Overview of the proposed method. First, image and audio features are extracted by the image and audio encoders, and serve as keys and values for slot attention. In slot attention, two learnable slots are used as queries to draw attentions on the input features while competing with each other, decomposing the features into the target and off-target representations per modality. The model is then trained by contrastive learning using only the target representations, and by cross-modal attention matching for further feature-level alignment between the two modalities. For localization in testing, we draw the attention map between the audio target representation and image features. Some details for learning slot attention, such as decoders and loss for slot reconstruction, have been omitted for clarity. + +![](images/5103af55f0ec301d98f385465743772fba50039830256fc54a47baf0e4592ecd.jpg) + +and audio features into two distinct slot representations per modality, one for the target and the other for the rest. To be specific, we designate the first slot as the target slot, which is guided to focus on information related to the shared sound source across both modalities. The exclusive assignment property of slot attention directs the second slot, referred to as the off-target slot, to capture remaining elements, such as the image background or irrelevant sounds in the audio. Then we learn SSL by contrastive learning using the two target slot representations from the image and audio, respectively. Unlike the prior work using only a single global audio feature, our method allows to exploit only the relevant part (i.e., the target slot) of the audio for more accurate SSL. Furthermore, since the target slot of the image covers the sound source area rather than a single location, our method avoids the problem of localization being reduced to a small discriminative part. + +Given the target slot representations, SSL is conducted by drawing an attention map on the image feature map using the target slot representation of the audio as query (Fig. 2(b)). For accurate pixel-level localization, the target slot of the audio should be well aligned with the keys derived from relevant individual image features; in this regard, contrastive learning using only the target slots is insufficient as it does not relate the audio target slot with individual image features. We thus introduce an additional training strategy dubbed cross-modal attention matching. This strategy encourages the cross-modal attention map between the audio target slot and the input image features to approximate the intra-modal attention map between the image target slot and the input image features that are inherently well + +aligned. In other words, it provides pseudo-supervision for SSL prediction (i.e., the cross-modal attention map) using the intra-modal attention map, which is drawn by the query and keys from the same modality and thus better localizes the target. To further facilitate its cross-modal alignment, we apply this strategy to the audio modality as well. + +Our model was trained on two datasets, Flickr-SoundNet [4] and VGG-Sound [5], and evaluated on three benchmarks, Flickr-SoundNet-Test [4], VGG-SS [5], and AVSBench [36]. It achieved the state of the art in almost all settings across all the three evaluation benchmarks. Further, our model outperformed all the prior work by large margin in cross-modal retrieval, suggesting that it is not a simple saliency detector but properly learns to relate image and audio for SSL [31]. In summary, the contribution of our method is three-fold: + +- We propose a novel SSL model that identifies and extracts features relevant to sound source jointly from image and audio modalities through slot attention. +- We also propose cross-modal attention matching, which enables effective cross-modal alignment for individual input features and thus improves localization performance. +- Our model achieved the state of the art in almost all settings across all the three SSL benchmarks and improved cross-modal retrieval performance substantially. + +# 2. Related Work + +Sound source localization aims to identify specific regions in an image that correspond to the sources of sounds from an audio input. Since constructing audio-visual datasets is + +costly, sound source localization studies have evolved into self-supervised learning approaches that use only image-audio pairs for contrastive learning [1, 6, 12, 14, 16, 21, 22, 26, 29-31, 33]. + +An early study, [29], addresses the task through contrastive learning between the audio-aware image features and the global audio feature. The audio-aware image features are obtained by calculating a correspondence map through the normalized inner product between the global audio feature and the image feature map. Chen et al. [6] propose method that uses a differentiable threshold on the correspondence map to divide audio-aware and other image features, treating the latter as hard negatives in contrastive learning to improve localization. + +Recent studies focus on fine-grained sound source localization through contrastive learning between the global audio feature and the most relevant feature selected from the image feature map [21, 22, 31, 33]. Mo and Morgado [21] propose a method that refines the audio-visual correspondence map through a post-processing technique leveraging a pretrained visual model. Sun et al. [33] introduce a regularizer that applies mean absolute error regularization to inter-modal and intra-modality similarity matrices to increase awareness of false negatives. Senocak et al. [31] propose a cross-modal alignment method that preserves localization performance and enhances cross-modal retrieval by leveraging multi-view augmentation and nearest neighbor search to construct multiple positive pairs for contrastive learning. These methods employ contrastive learning with the most relevant image feature to assist in coarsely identifying sound sources but often fall short of capturing a comprehensive representation, thereby hindering localization ability. Furthermore, relying solely on global audio features can mislead contrastive learning when noise exists, potentially highlighting incorrect image features. + +Following a complementary direction, several studies employ external prior knowledge or even utilize off-the-shelf segmentation modules to achieve high-quality localization results [25, 26]. Also, several studies employ sound source localization with additional information such as optical flow [9], text inputs [19] or ground truth category [23]. By contrast to these studies, our method can effectively perform sound source localization without relying on external prior knowledge or off-the-shelf models. + +To address the above issues, we propose a sound source localization framework based on the slot attention mechanism. Slot attention [17] is proposed for unsupervised object-centric learning, effectively decomposing input features into distinct and mutually exclusive components via iterative attention and competition among slots. Inspired by this, we adopt slot attention in both image and audio features to decompose them into features related to the sound source and those irrelevant to it without supervision. + +# 3. Proposed Method + +As shown in Fig. 2, our model consists of two parts: 1) image and audio encoders for feature extraction, and 2) joint feature decomposition module with joint slot attention mechanism. A pair of image and audio is first fed into each encoder to extract image and audio features (Sec. 3.1). The features of each modality are decomposed into two distinct slot representations by using joint slot attention: one slot for the target and the other for the rest. (Sec. 3.2). The slot representations are then used for training as input to contrastive learning and cross-modal attention matching (Sec. 3.3), while effectively handling false negatives (Sec. 3.4). Lastly, an attention on the image feature map using the target slot representation of the audio is computed to locate the source of sound (Sec. 3.5). + +# 3.1. Feature Extraction + +Let $D = \{(v_{i},a_{i})\}_{i = 1}^{M}$ denote a training set of image-audio pairs, where $v_{i}\in \mathbb{R}^{3\times H_{v}\times W_{v}}$ is an image of size $H_{v}\times W_{v}$ , $a_{i}\in \mathbb{R}^{1\times F_{a}\times T_{a}}$ is a corresponding audio spectrogram of frequency resolution $F_{a}$ and temporal length $T_{a}$ , and $M$ denotes the total number of such pairs in the dataset. For feature extraction, separate image and audio encoders $f_{v}$ and $f_{a}$ , are adopted. The image encoder extracts the image features $\mathbf{v}_i\coloneqq f_v(v_i)\in \mathbb{R}^{h\times w\times c}$ for $v_{i}$ , and the audio encoder extracts the audio features $\mathbf{a}_i\coloneqq f_a(a_i)\in \mathbb{R}^{t\times c}$ for $a_{i}$ where $h$ and $w$ are the height and width of the image feature map respectively, $t$ is the length of the audio features, and $c$ is the image and audio feature dimension. + +# 3.2. Joint Feature Decomposition + +Slot attention [17] is an attention mechanism where each slot progressively competes to aggregate input features, leading to the decomposition of input features into distinct and exclusive representations. Building on this idea, for joint feature decomposition, we propose joint slot attention (JSA), which jointly decomposes image and audio features into two distinct slot representations for each modality. To achieve this, we employ slot attention with two initial slots $S_0 \in \mathbb{R}^{2 \times D}$ shared in the two modalities. Furthermore, we designate the first slot as target slot, which captures features relevant to sound source, and the other as off-target slot, which captures features for the rest. Since the overall pipeline of JSA is identical in the two modalities, we will describe the module in terms of image only for brevity. + +JSA first projects an input image feature $\mathbf{v}$ to derive its corresponding key $K^{\nu}$ and value $V^{\nu}$ by applying the following operations: + +$$ +K ^ {v} = W ^ {k} (\operatorname {L a y e r N o r m} (\mathbf {v})) \in \mathbb {R} ^ {h w \times d}, \tag {1} +$$ + +$$ +V ^ {v} = W ^ {v} \left(\operatorname {L a y e r N o r m} (\mathbf {v})\right) \in \mathbb {R} ^ {h w \times d}, +$$ + +where $W^{k}$ , $W^{v}$ are the linear projection layers to obtain + +embeddings in $d$ -dimension. For each iteration $n \in [1, N]$ , the query $Q_{n}^{v}$ is derived by applying linear projection to the slots with the initial slots $S_{0}^{v} \coloneqq S_{0}$ , which is given by + +$$ +Q _ {n} ^ {v} = W ^ {q} \left(\operatorname {L a y e r N o r m} \left(S _ {n - 1} ^ {v}\right)\right) \in \mathbb {R} ^ {2 \times d}, \tag {2} +$$ + +where $W^q$ is a linear projection layer shared for all iterations. The slot attention matrix $A$ is computed by taking the dot product of the given query $Q_n^v$ and key $K^v$ , scaled by the square root of the dimension $d$ as follows: + +$$ +A _ {i, j} = \frac {\exp \left(M _ {i , j}\right)}{\sum_ {l = 1} ^ {2} \exp \left(M _ {i , l}\right)}, \text {w h e r e} M = \frac {K ^ {v} \left(Q _ {n} ^ {v}\right) ^ {\top}}{\sqrt {d}} \in \mathbb {R} ^ {h w \times 2}, \tag {3} +$$ + +where $i$ is key index, and $j$ is slot index. Subsequently, a weighted mean is utilized to the slot attention matrix $A$ : + +$$ +\hat {A} _ {i, j} = \frac {A _ {i , j}}{\sum_ {l = 1} ^ {h w} A _ {l , j}}. \tag {4} +$$ + +Then, the updated slot $\bar{S}_n$ is computed by Gated Recurrent Unit (GRU) [7] with $S_{n - 1}$ as hidden state, and $\hat{A}^{\top}\cdot V^{v}$ as input as follows: + +$$ +\bar {S _ {n} ^ {v}} = \operatorname {G R U} \left(S _ {n - 1} ^ {v}, \hat {A} ^ {\top} \cdot V ^ {v}\right). \tag {5} +$$ + +Finally, the slot is updated by a residual connection: + +$$ +S _ {n} ^ {v} = \bar {S} _ {n} ^ {v} + \operatorname {M L P} \left(\operatorname {L a y e r N o r m} \left(\bar {S} _ {n} ^ {v}\right)\right) \in \mathbb {R} ^ {2 \times d}. \tag {6} +$$ + +The overall process encourages the input feature to bind to specific slots, ensuring that each slot forms a mutually exclusive representation. The final outputs of JSA are image slots $\{\mathbf{p}^v;\mathbf{r}^v\} := S_N^v\in \mathbb{R}^{2\times d}$ and image query $Q_{N}^{v}\in$ $\mathbb{R}^{2\times d}$ , where $\mathbf{p}^v$ is a target slot and $\mathbf{r}^v$ is an off-target slot. The same process is applied to the audio modality using audio feature $\mathbf{a}$ , yielding audio slots $\{\mathbf{p}^a;\mathbf{r}^a\} := S_N^a\in \mathbb{R}^{2\times d}$ and audio query $Q_{N}^{a}\in \mathbb{R}^{2\times d}$ . Since this process begins from the shared initial slots, the initial slots are able to effectively learn joint features between two modalities. For brevity, we omit the subscript $N$ for the final query, $Q^{v}\coloneqq Q_{N}^{v}$ and $Q^{a}\coloneqq Q_{N}^{a}$ . + +# 3.3. Learning Objectives + +Contrastive Learning with Target Slots. We adopt the InfoNCE loss [24, 27] for contrastive learning using the target slots, guiding the target slots to identify features relevant to the sound source jointly from image and audio. Let $s$ denote a similarity function, $s(x,y) = \exp (\cos (x,y) / \tau)$ , where $\tau$ is a temperature hyperparameter. The contrastive learning loss $\mathcal{L}_{\mathrm{cotr}}$ is given by + +$$ +\mathcal {L} _ {\mathrm {c o t r}} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \left\{\log \frac {s \left(\mathbf {p} _ {i} ^ {v} , \mathbf {p} _ {i} ^ {a}\right)}{\sum_ {j} ^ {B} s \left(\mathbf {p} _ {i} ^ {v} , \mathbf {p} _ {j} ^ {a}\right)} + \log \frac {s \left(\mathbf {p} _ {i} ^ {v} , \mathbf {p} _ {i} ^ {a}\right)}{\sum_ {j} ^ {B} s \left(\mathbf {p} _ {j} ^ {v} , \mathbf {p} _ {i} ^ {a}\right)} \right\}, +$$ + +where $B$ denotes the number of data in a mini-batch, and the negative pairs refer to the combinations of visual and audio features that do not belong to the same original pair within the mini-batch. Since the image target slot captures the whole features of the target region, it avoids focusing solely on a specific discriminative part of the image. Furthermore, the audio target slot captures only the relevant sound from the source, it prevents the issue of learning sounds that are unrelated to the source. + +Cross-modal Attention Matching. As will be described in Sec. 3.5, the localization result is derived by utilizing cross-modal attention between the audio target slot and image features. For accurate sound source localization, the audio target slot should be properly aligned with keys computed from the relevant image features. Although contrastive learning encourages alignment of target slot representations between different modalities, it does not ensure alignment between the audio target slot and image features, which degrades the localization accuracy. We resolve this issue by cross-modal attention matching, which encourages the cross-modal attention map to resemble the intra-modal attention map for improving feature-level alignment across the two modalities. To this end, we first calculate the cross-modal and intra-modal attention matrices. Specifically, we compute the cross-modal attention matrix between the audio and visual modalities in the same manner as Eq. (3) and Eq. (4), with the audio queries $Q^a$ and visual keys $K^v$ . We denote this cross-modal attention matrix as $\hat{A}^{a,v} \in \mathbb{R}^{hw \times 2}$ , where $hw$ is the length of $K^v$ . Subsequently, to obtain the cross-modal attention that represents the attention weights between the audio target query and visual keys, we utilize only the first column of $\hat{A}^{a,v}$ . This cross-modal attention is defined as follows: + +$$ +\mathbf {c a} ^ {a, v} := \left[ \hat {A} _ {i, 1} ^ {a, v} \right] _ {i = 1} ^ {h w} \in \mathbb {R} ^ {h w}. +$$ + +Similarly, for intra-modal attention matrix within the visual modality, we compute the intra-attention matrix $\hat{A}^{v,v}$ using the $Q^v$ and $K^v$ . Then, the intra-modal attention that represents the attention weights between the image target query and visual keys is defined as follows: + +$$ +\mathbf {i a} ^ {v, v} := \left[ \hat {A} _ {i, 1} ^ {v, v} \right] _ {i = 1} ^ {h w} \in \mathbb {R} ^ {h w}. +$$ + +To further facilitate cross-modal alignment, we apply this strategy symmetrically to the audio modality as well. Finally we formulate the cross-modal attention matching loss as follows: + +$$ +\mathcal {L} _ {\text {m a t c h}} = \left\| \mathbf {c a} ^ {a, v} - \operatorname {s g} \left(\mathbf {i a} ^ {v, v}\right) \right\| _ {2} ^ {2} + \left\| \mathbf {c a} ^ {v, a} - \operatorname {s g} \left(\mathbf {i a} ^ {a, a}\right) \right\| _ {2} ^ {2}, \tag {8} +$$ + +where $\mathrm{sg}(\cdot)$ denotes the stop-gradient operation. $\mathcal{L}_{\mathrm{match}}$ minimizes the discrepancy between cross-modal and intramodal attentions, promoting robust feature-level alignment between the two modalities for accurate localization. + +Slot Divergence. In slot attention, the diversity between the target slot and off-target slot should be large enough to ensure that each slot effectively captures distinct, mutually exclusive regions. Therefore, we introduce a slot divergence loss that encourages greater diversity among the slots. The slot divergence loss $\mathcal{L}_{\mathrm{div}}$ is given by + +$$ +\mathcal {L} _ {\mathrm {d i v}} = \max \left\{0, \cos \left(\mathbf {p} ^ {v}, \mathbf {r} ^ {v}\right) \right\} + \max \left\{0, \cos \left(\mathbf {p} ^ {a}, \mathbf {r} ^ {a}\right) \right\}. \tag {9} +$$ + +It ensures that the different slot embeddings to be sufficiently dissimilar, enabling the target slot to more effectively localize the sound source. + +Slot Reconstruction. The successful feature reconstruction in slot attention implies that each slot successfully captures unique and independent parts of the input data, which means that the slot attention mechanism effectively decomposed the input into individual slot representations. It indicates that each slot accurately represents a specific object or region in the input. To encourage this condition, we apply the slot reconstruction loss $\mathcal{L}_{\mathrm{recon}}$ as follows: + +$$ +\mathcal {L} _ {\text {r e c o n}} = \| \mathbf {v} - g _ {v} \left(\left\{\mathbf {p} ^ {v}; \mathbf {r} ^ {v} \right\}\right) \| _ {2} ^ {2} + \| \mathbf {a} - g _ {a} \left(\left\{\mathbf {p} ^ {a}; \mathbf {r} ^ {a} \right\}\right) \| _ {2} ^ {2}, \tag {10} +$$ + +where $g_{v}$ and $g_{a}$ are the decoders for the image and audio features, respectively. These decoders are composed of a small number of MLP layers with ReLU activations. Note that the decoders are auxiliary modules used only in training and not required for inference. + +The overall loss is a combination of $\mathcal{L}_{\mathrm{cotr}}$ , $\mathcal{L}_{\mathrm{match}}$ , $\mathcal{L}_{\mathrm{div}}$ , and $\mathcal{L}_{\mathrm{recon}}$ , which is given by + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c o t r}} + \lambda_ {1} \mathcal {L} _ {\text {m a t c h}} + \lambda_ {2} \mathcal {L} _ {\text {d i v}} + \lambda_ {3} \mathcal {L} _ {\text {r e c o n}}, \tag {11} +$$ + +where $\lambda_{1}$ , $\lambda_{2}$ , and $\lambda_{3}$ are hyperparameters weighting the loss terms. + +# 3.4. Mitigating False Negatives + +In general, unpaired images and audiios are all treated as negatives for contrastive training. However, some of them are false negatives that are semantically well matched, so treating these as negatives causes adverse effect [33]. Therefore, we propose a false negative mitigation strategy using $k$ -reciprocal nearest neighbors of target slots. + +For the image target slot $\mathbf{p}_i^v$ , which is derived from $v_{i}$ we define its set of $k$ -reciprocal nearest neighbors $R_{k}^{v}(i)$ as + +$$ +R _ {k} ^ {v} (i) = \left\{\left(v _ {j}, a _ {j}\right) \mid \left(j \in N _ {k} ^ {v} (i)\right) \cap \left(i \in N _ {k} ^ {v} (j)\right) \right\}, \tag {12} +$$ + +where $N_{k}^{v}(i)$ is the $k$ -nearest neighbors of $\mathbf{p}_i^v$ , computed using cosine similarity. Likewise, the set of $k$ -reciprocal nearest neighbors between audio target slots, i.e., $R_{k}^{a}(i)$ , can be derived. The predicted false negative set $R_{k}(i)$ is defined as + +$$ +R _ {k} (i) = R _ {k} ^ {v} (i) \cap R _ {k} ^ {a} (i). \tag {13} +$$ + +Since the samples in $R_{k}(i)$ have both similar image and audio target slots with $(v_{i}, a_{i})$ , they are likely to belong to false negatives. We exclude these predicted false negatives from the contrastive learning in Eq. (7). + +# 3.5. Inference + +For inference, $\mathbf{ca}^{a,v}$ , which represents the attention between the audio target slot and the input image features, is utilized. Specifically, the attention is interpolated to the original image size, and pixels whose score exceeds the hyperparameter $\theta$ are selected to generate the localization result. + +Furthermore, unlike the object-guided localization [21], we propose a novel localization refinement strategy, dubbed Image-Query based Refinement (IQR), to refine the localization result without relying on external prior knowledge. IQR estimates the regions that are likely to contain the target by using the attention between the image target slot and the input image features. It can be used to identify prior regions in the image that favor the target, regardless of whether these regions are actually the source of the sound. The refined result derived by IQR, denoted as $R_{\mathrm{IQR}}$ , is given as + +$$ +R _ {\mathrm {I Q R}} = \alpha \cdot \mathbf {c a} ^ {a, v} + (1 - \alpha) \cdot \mathbf {i a} ^ {v, v}, \tag {14} +$$ + +where $\alpha \in (0,1)$ balances the sound source localization result and potential sound source regions. + +# 4. Experiment + +# 4.1. Experiment Setup + +Datasets. Flickr-SoundNet [4] and VGG-Sound [5] are used to train. Flickr-SoundNet contains about 2 million unconstrained videos from Flickr, while VGG-Sound has around 200k video clips. For training, we use subsets of 10k and 144k from each dataset, respectively. For the evaluation, Flickr-SoundNet-Test [4], VGG-SS [5], and AVSBench [36] are used following the previous research [6, 21, 31, 33]. Flickr-SoundNet-Test has 250 audio-visual pairs with manually labeled bounding boxes. VGG-SS has about 5k audio-visual pairs over 220 categories. Since VGG-SS provides its category ground truth, it is also used to evaluate cross-modal retrieval performance. AVSBench has about 5k videos with pixel-wise binary segmentation maps. Following the previous study [6], the middle frame of the video clip is used for the image. For the audio, the middle 5 seconds are extracted and converted into log spectrograms with 257 frequency bins at a sampling rate of $16\mathrm{kHz}$ . + +Network Architecture. ResNet-18 [11] is adopted for both the image encoder and the audio encoders. The image encoder is fine-tuned from an ImageNet [8] pretrained model, while the audio encoder is trained from scratch. Both encoders include an additional $1 \times 1$ convolutional layer for projection, and the audio encoder also includes a MaxPooling layer for pooling along the frequency axis. + +Hyperparameter Settings. We use the AdamW optimizer [18] with a learning rate of 5e-5 and a weight decay of 1e-2. The mini-batch size during training is set to 256, and the size of image and audio feature size $h$ , $w$ , $t$ and $c$ + +
TrainsetMethodOGLVGG-SSFlickr-SoudNet
cIoUAUCcIoUAUC
Flickr SoundNet 10kAttention10k [29]--43.6044.90
CoursetoFine [26]--52.2049.60
AVObject [1]--54.6050.40
LVS [6]--58.2052.50
EZ-VSL [21]19.8630.9662.2454.74
FNAC [33]35.2738.0084.3363.26
Ours35.1138.6685.2065.26
EZ-VSL [21]✓37.6139.2181.9362.58
FNAC [33]✓40.9740.3884.7364.34
Ours✓39.1640.5187.6066.50
Flickr SoundNet 144kAttention10k [29]--66.0055.80
DMC [12]--67.1056.80
LVS [6]--69.9057.30
HardPos [30]--75.2059.70
EZ-VSL [21]30.2735.9272.6958.70
FNAC [33]33.9337.2978.7159.33
Alignment [31]--85.2062.20
Ours37.0139.4286.0065.16
EZ-VSL [21]✓41.0140.2383.1363.06
FNAC [33]✓41.1040.4483.9363.06
Alignment [31]✓--84.0064.16
Ours✓40.1540.7589.2064.50
+ +
TrainsetMethodOGLVGG-SSFlickr-SoudNet
cIoUAUCcIoUAUC
VGGSound 10kLVS [6]--61.8053.60
EZ-VSL [21]25.8433.6863.8554.44
FNAC [33]37.2938.9985.7463.66
Ours37.8439.8286.0064.90
EZ-VSL [21]✓38.7139.8078.7161.53
FNAC [33]✓40.6940.4282.1363.64
Ours✓41.0241.2987.6065.42
VGGSound 144kAttention10k [29]18.5030.20--
DMC [12]29.1034.80--
AVObject [1]29.7035.70--
LVS [6]34.4038.2073.5059.00
HardPos [30]34.6038.0076.8059.20
EZ-VSL [21]34.3837.7079.5161.17
SLAVC [22]37.22-83.20-
FNAC [33]39.5039.6684.7363.76
Alignment [31]39.9440.0279.6063.44
NoPrior [14]41.441.2--
Ours40.6241.6683.6064.68
EZ-VSL [21]✓38.8539.5483.9463.60
SLAVC [22]✓39.67-86.40-
FNAC [33]✓41.8540.8085.1464.30
Alignment [31]✓42.6441.4882.4064.60
Ours✓42.5642.4684.8064.82
+ +Table 1. Quantitative results of the model trained on Flickr-SoundNet-10k, 144k, and VGGSound-10k, 144k datasets. Note that the results of previous research are obtained from [14, 31, 33]. The results of our method represent the performance after applying Image-Query based Refinement (IQR). The result with OGL denotes the performance after applying object-guided localization [21]. + +![](images/0960ed3c5dcdee521c0e1003cee3986a7f06a785f4ddb4a179cee306675f0003.jpg) +Figure 3. Sound localization results on Flickr-SoundNet-Test [4] and VGG-SS [5]. (a) Input image. (b) Ground-Truth. (c) Ours. (d) Alignment [31]. (e) FNAC [33]. (f) EZ-VSL [21]. The qualitative results are obtained by the model trained on Flickr-144k and the model trained on VGGSound-144k, respectively. Note that all visualizations are obtained without refinement. + +are set to 7, 7, 16, and 512 respectively. $\tau$ in Eq. (7) is set to 0.03. $\lambda_{1}, \lambda_{2}$ , and $\lambda_{3}$ in Eq. (11) are set to 100.0, 0.1, and 0.1, respectively. $\alpha$ in Eq. (14) is set to 0.6. The value $k$ in Eq. (13) is set to 20. The number of iterations $N$ for slot attention mechanism is set to 5. + +Evaluation metric. We measure the average precision at a Consensus Intersection over Union (cIoU) threshold of 0.5 to evaluate localization accuracy. Additionally, the AUC score is used to provide an overall performance measure across different thresholds. + +# 4.2. Quantitative Results + +Flickr-SoundNet and VGGSound. In this section, we compare our method with previous self-supervised SSL + +approaches. We trained our model on subsets of Flickr-SoundNet and VGGSound with dataset sizes of 10k and 144k, and evaluated the performance on Flickr-SoundNet-Test and VGG-SS. As shown in Table 1, our model achieves the state-of-the-art performance in most settings. Specifically, our method yields AUC improvements across all evaluated settings. Moreover, even without using Object-Guided Localization (OGL), our model performs competitively or better compared to previous methods, demonstrating the robustness of our architecture in focusing on the sound source without additional refinement steps. + +Heard 110 and Unheard 110. To assess the generalization ability of SSL, we evaluate the performance with Heard 110 and Unheard 110. Heard 110 and Unheard 110 are sub + +
MethodOGLHeard 110Unheard 110
cIoUAUCcIoUAUC
LVS [6]28.9036.2026.3034.70
EZ-VSL [21]31.8636.1932.6636.72
SLAVC [22]35.84-36.50-
Alignment [31]38.3139.0539.1139.80
Ours38.4640.1540.3541.25
EZ-VSL [21]✓37.2538.9739.5739.60
SLAVC [22]✓38.22-38.87-
FNAC [33]✓39.5439.8342.9141.17
Alignment [31]✓41.8540.9342.9441.54
Ours✓41.5041.6043.1142.04
+ +Table 2. Quantitative results on Heard 110 and Unheard 110. Note that the results of previous research are obtained from [31, 33]. The results of our method represent the performance after applying Image-Query based Refinement (IQR). The result with OGL denotes the performance after applying object-guided localization. + +
TestsetMethodmIoUF-score
S4LVS [6]26.933.6
EZ-VSL [21]27.634.2
SLAVC [22]28.134.6
FNAC [33]27.1531.4
Alignment [31]29.635.9
\( Alignment^† \)29.335.6
Ours31.3344.30
+ +Table 3. Quantitative results on AVSBench. Note that the results of previous research are obtained from [31, 33]. $\dagger$ denotes that the self-supervised pretrained encoder is utilized for their work. + +
MethodAudio → ImageImage → Audio
R@1R@5R@10R@1R@5R@10
LVS [6]3.8712.3520.734.9014.2921.37
EZ-VSL [21]5.0115.7324.8114.2033.5145.18
SSL-TIE [16]10.2930.6843.7612.7629.5839.72
SLAVC [22]4.7713.0819.106.1221.1632.12
Alignment [31]16.4736.9949.0020.0942.3853.66
\( Alignment^† \)14.3137.8149.1718.0038.3949.02
Ours30.6157.8769.3731.7058.4369.81
+ +sets of the VGG-SS, each containing 110 non-overlapping categories. During training, the model is trained only to the category set of one group. The subset comprising categories seen during training is called Heard 110, while the subset with unseen categories is Unheard 110. As shown in Table 2, our model achieves the state-of-the-art performance on Unheard 110 both without additional refinement + +Table 4. Cross-modal retrieval on the VGG-SS. All of the models are trained on VGGSound-144K. Note that the results of previous research are obtained from [31]. $\dagger$ denotes that the self-supervised pretrained encoder is utilized for their work. + +
Trainset\( {\mathcal{L}}_{\text{match }} \)\( {\mathcal{L}}_{\text{div }} \)FN MitigateVGG-SS
cIoUAUC
VGGSound 144k✓14.9527.02
✓36.2339.13
18.6930.03
✓15.0328.18
✓✓40.5641.35
✓✓36.1439.31
✓✓20.3730.81
✓✓✓40.7141.62
+ +Table 5. Detailed ablation study for losses. + +
TrainsetShared Initial SlotsVGG-SSFlickr
cIoUAUCcIoUAUC
Flickr-144k✓34.8038.7284.0061.70
36.6439.2486.8065.44
VGGSound-144k✓33.1137.9281.2064.36
40.7141.6284.4064.50
+ +Table 6. Ablation study on shared initial slots. + +![](images/d8ea648b19ac68aad20d33c35672dc060174bb30465a0e4af7489e6b0b5a6ea3.jpg) +Figure 4. Qualitative results to show the impact of $\mathcal{L}_{\mathrm{match}}$ on Flickr-SoundNet. (a) Input image. (b) Ground-Truth. (c) Results without cross-modal attention matching. (d) Results with cross-modal attention matching. + +and with OGL applied. These results demonstrate that our model has superior generalization ability compared to previous models. + +AVSBench. To more precisely evaluate the performance of the localization mask, we measured the performance on AVSBench, which provides segmentation masks. Our model was trained on VGGSound-144k and evaluated in a zero-shot manner on AVSBench. As shown in Table 3, we achieved the state-of-the-art results, specifically with improvements of $2.03\%$ p in mIoU and $8.7\%$ p in F-score. + +Cross-modal Retrieval. As discussed in the previous study [31], the benchmark evaluation metrics, cIoU and + +![](images/0aea5b30ff7ce9d68e20812e204c773445d5e523bd4ef1ce6620ec75f50564a2.jpg) +Figure 5. Qualitative results of cross-modal retrieval on VGG-SS. + +AUC, do not guarantee better audio-visual semantic understanding. Therefore, we also evaluate cross-modal retrieval performance on VGG-SS. We extract the image and audio target slot representations and compute the cosine similarity between them. The similarities are ranked in descending order, and we calculate Recall@K performance using the category labels from the top-K retrieved audio or images. In Table 4, our model demonstrates a significant performance improvement over the previous state of the art in cross-modal retrieval. Specifically, our model improves Recall@1 by $16.3\%$ p in audio-to-image retrieval and $13.7\%$ p in image-to-audio retrieval. This result indicates that our model does not merely detect salient objects likely to produce sound within the image, but rather effectively locates the target specified by the audio input. + +# 4.3. Ablation Study + +Impact of Losses and False Negative Mitigating in Sec. 3.4. In Table 5, we conduct an ablation study to verify the effectiveness of each loss function and the strategy outlined in Sec. 3.4. The overall results show that the designed losses and the false negative mitigating strategy each contribute to improve the performance. Notably, the experiment with $\mathcal{L}_{\mathrm{match}}$ demonstrates a drastic increase in performance, highlighting the importance of cross-modality attention matching. Additionally, the ablation study on $\mathcal{L}_{\mathrm{div}}$ shows that explicitly separating the slots enables each slot to perform its designated role more effectively. + +Impact of Shared Initial Slots. We conduct an ablation study to examine the impact of shared initial slots by using different initial slots for each modality. As shown in Table 6, sharing initial slots resulted in better performance, indicating that this design choice facilitates the ability to jointly capture sound source-related information across different modalities. + +# 4.4. Qualitative Results + +Quality of SSL results. Fig. 3 visualizes the qualitative results of our model and previous studies on Flickr- + +SoundNet [4] and VGG-SS [5] test sets. These results demonstrate that our model has the ability to effectively localize sound source objects. Notably, it produces compact localization results even for challenging samples that other models struggle with, such as small birds, a gun, and cars surrounded by other objects in street scenes. + +Qualitative Analysis of $\mathcal{L}_{\mathrm{match}}$ . Fig. 4 shows the qualitative result that shows the effect of $\mathcal{L}_{\mathrm{match}}$ on Flickr-SoundNet-Test. The attention maps is generated between the audio query of the audio target slot and the image keys derived from image features. Without $\mathcal{L}_{\mathrm{match}}$ , the alignment between the two is insufficient for compact localization, whereas with $\mathcal{L}_{\mathrm{match}}$ , the alignment improves, resulting in more accurate localization of the sound source. + +Cross-modal Retrieval. Fig. 5 shows the qualitative results of our model for cross-modal retrieval. Consistent with the impressive performance shown in Table 4, our model demonstrates a strong ability to retrieve semantically similar samples across modalities that correspond to the sound source target. + +# 5. Conclusion + +This paper presented a novel self-supervised sound source localization method leveraging joint slot attention to align image and audio features. Our approach decomposes these features into target and off-target slots, using only target slots for contrastive learning to capture sound-related features. In addition, we introduced cross-modal attention matching and slot divergence loss which significantly improves localization accuracy and and we demonstrated the effectiveness of our model through comprehensive analysis. For future work, we aim to extend our framework to multisource sound source localization. + +Acknowledgement. This work was supported by the NRF grant (RS-2021-NR059830-40%) and the IITP grants (RS-2024-00509258-40%, RS-2021-II212068-10%, RS-2019-II191906-10%) funded by Ministry of Science and ICT, Korea. + +# References + +[1] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In Proc. European Conference on Computer Vision (ECCV), 2020. 1, 3, 6 +[2] Stuart Andrews, Ioannis Tsochantaridis, and Thomas Hofmann. Support vector machines for multiple-instance learning. Advances in neural information processing systems, 15, 2002. 1 +[3] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proc. European Conference on Computer Vision (ECCV), pages 435-451, 2018. 1 +[4] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Sound-net: Learning sound representations from unlabeled video. Proc. Neural Information Processing Systems (NeurIPS), 29, 2016. 2, 5, 6, 8 +[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2, 5, 6, 8 +[6] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 5, 6, 7 +[7] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724-1734, Doha, Qatar, 2014. Association for Computational Linguistics. 4 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: a large-scale hierarchical image database. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 5 +[9] Dennis Fedorishin, Deen Dayal Mohan, Bhavin Jawade, Srirangaraj Setlur, and Venu Govindaraju. Hear the flow: Optical flow-based self-supervised visual sound source localization. In Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV), pages 2278-2287, 2023. 3 +[10] Yury Furletov, Volker Willert, and Jürgen Adamy. Auditory scene understanding for autonomous driving. In 2021 IEEE Intelligent Vehicles Symposium (IV), pages 697-702. IEEE, 2021. 1 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5 +[12] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 3, 6 +[13] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In Proc. In- + +ternational Conference on Machine Learning (ICML), pages 2127-2136. PMLR, 2018. 1 +[14] Dongjin Kim, Sung Jin Um, Sangmin Lee, and Jung Uk Kim. Learning to visually localize sound sources from mixtures without prior source knowledge. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 26467-26476, 2024. 1, 3, 6 +[15] Chen Liu, Peike Patrick Li, Xingqun Qi, Hu Zhang, Lincheng Li, Dadong Wang, and Xin Yu. Audio-visual segmentation by exploring cross-modal mutual semantics. In Proc. ACM Multimedia Conference (ACMMM), pages 7590–7598, 2023. 1 +[16] Jinxiang Liu, Chen Ju, Weidi Xie, and Ya Zhang. Exploiting transformation invariance and equivariance for self-supervised sound localisation. In Proc. ACM Multimedia Conference (ACMMM), pages 3742-3753, 2022. 3, 7 +[17] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Proc. Neural Information Processing Systems (NeurIPS), 33:11525-11538, 2020. 1, 3 +[18] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. International Conference on Learning Representations (ICLR), 2019. 5 +[19] Tanvir Mahmud, Yapeng Tian, and Diana Marculescu. T-vsl: Text-guided visual sound source localization in mixtures. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 26742-26751, 2024. 3 +[20] Oded Maron and Tomás Lozano-Pérez. A framework for multiple-instance learning. Advances in neural information processing systems, 10, 1997. 1 +[21] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In Proc. European Conference on Computer Vision (ECCV), 2022. 1, 3, 5, 6, 7 +[22] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. In Proc. Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 6, 7 +[23] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10565–10574, 2023. 3 +[24] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 4 +[25] Sooyoung Park, Arda Senocak, and Joon Son Chung. Can clip help sound source localization? In Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV), 2024. 1, 3 +[26] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In Proc. European Conference on Computer Vision (ECCV), 2020. 1, 3, 6 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning + +transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4 +[28] Yoko Sasaki, Ryo Tanabe, and Hiroshi Takernura. Online spatial sound perception using microphone array on mobile robot. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2478-2484. IEEE, 2018. 1 +[29] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 3, 6 +[30] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Learning sound localization better from semantically similar samples. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022. 6 +[31] Arda Senocak, Hyeonggon Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 5, 6, 7 +[32] Zengjie Song, Yuxi Wang, Junsong Fan, Tieniu Tan, and Zhaoxiang Zhang. Self-supervised predictive learning: A negative-free method for sound source localization in visual scenes. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3222-3231, 2022. +[33] Weixuan Sun, Jiayi Zhang, Jianyuan Wang, Zheyuan Liu, Yiran Zhong, Tianpeng Feng, Yandong Guo, Yanhao Zhang, and Nick Barnes. Learning audio-visual source localization via false negative aware contrastive learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 5, 6, 7 +[34] Jun Yin, Stefano Damiano, Marian Verhelst, Toon van Watershoot, and Andre Guntoro. Real-time acoustic perception for automotive applications. In 2023 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1-6. IEEE, 2023. 1 +[35] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep features for discriminative localization. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1 +[36] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In Proc. European Conference on Computer Vision (ECCV), pages 386–403, 2022. 2, 5 \ No newline at end of file diff --git a/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/images.zip b/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a8944f2735c6a4b2ae47be5240b244a164746b7a --- /dev/null +++ b/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05334bd461d3573cdcd13c740ae668d820191610c3ac20bf509aee2fec1cc056 +size 766257 diff --git a/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/layout.json b/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..43b616956994f0169b7750edeae2757b3cf528bc --- /dev/null +++ b/2025/Improving Sound Source Localization with Joint Slot Attention on Image and Audio/layout.json @@ -0,0 +1,8806 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 172, + 103, + 440, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 103, + 440, + 140 + ], + "spans": [ + { + "bbox": [ + 172, + 103, + 440, + 140 + ], + "type": "text", + "content": "Improving Sound Source Localization with Joint Slot Attention on Image and Audio" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 160, + 505, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 160, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 101, + 160, + 505, + 176 + ], + "type": "text", + "content": "Inho Kim1 Youngkil Song2 Jicheol Park2 Won Hwa Kim1,2 Suha Kwak1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 177, + 471, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 177, + 471, + 192 + ], + "spans": [ + { + "bbox": [ + 140, + 177, + 471, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 177, + 471, + 192 + ], + "type": "text", + "content": "Dept. of CSE, POSTECH " + }, + { + "bbox": [ + 140, + 177, + 471, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 177, + 471, + 192 + ], + "type": "text", + "content": "Graduate School of AI, POSTECH" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 148, + 193, + 459, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 193, + 459, + 207 + ], + "spans": [ + { + "bbox": [ + 148, + 193, + 459, + 207 + ], + "type": "text", + "content": "{kimih, songyk, jicheol, wonhwa, suha.kwak}@postech.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 233, + 200, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 233, + 200, + 247 + ], + "spans": [ + { + "bbox": [ + 151, + 233, + 200, + 247 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 259, + 297, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 297, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 297, + 499 + ], + "type": "text", + "content": "Sound source localization (SSL) is the task of locating the source of sound within an image. Due to the lack of localization labels, the de facto standard in SSL has been to represent an image and audio as a single embedding vector each, and use them to learn SSL via contrastive learning. To this end, previous work samples one of local image features as the image embedding and aggregates all local audio features to obtain the audio embedding, which is far from optimal due to the presence of noise and background irrelevant to the actual target in the input. We present a novel SSL method that addresses this chronic issue by joint slot attention on image and audio. To be specific, two slots competitively attend image and audio features to decompose them into target and off-target representations, and only target representations of image and audio are used for contrastive learning. Also, we introduce cross-modal attention matching to further align local features of image and audio. Our method achieved the best in almost all settings on three public benchmarks for SSL, and substantially outperformed all the prior work in cross-modal retrieval." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 524, + 137, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 137, + 537 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 137, + 537 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": "Humans easily identify where a sound comes from within their sight. This capability is a key aspect of human perception facilitating interaction with surroundings, and is also demanded in a large variety of vision applications such as autonomous driving and robotics [10, 28, 34]. Motivated by this, a large body of research [1, 3, 6, 12, 14, 15, 21, 22, 25, 26, 29-33] has studied sound source localization (SSL), the task of locating the source of a sound within the paired image in the form of segmentation mask." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Due to the prohibitively expensive cost of sound source annotation for image-audio pair data, existing methods [6, 21, 31, 33] have focused on learning SSL using only image-audio pairs without sound source labels. To this end, they consider the task as multiple instance learning [2, 13, 20]" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 322, + 234, + 420, + 342 + ], + "blocks": [ + { + "bbox": [ + 322, + 234, + 420, + 342 + ], + "lines": [ + { + "bbox": [ + 322, + 234, + 420, + 342 + ], + "spans": [ + { + "bbox": [ + 322, + 234, + 420, + 342 + ], + "type": "image", + "image_path": "0edefc1447d60faec279218077c61fc689637c13d81c5c49cc69ea278f071b45.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 344, + 394, + 353 + ], + "lines": [ + { + "bbox": [ + 339, + 344, + 394, + 353 + ], + "spans": [ + { + "bbox": [ + 339, + 344, + 394, + 353 + ], + "type": "text", + "content": "(a) Previous work" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 362, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 407 + ], + "type": "text", + "content": "Figure 1. Difference between previous work and ours. For contrastive learning of SSL, our method identifies and utilizes pairs of relevant image region and audio segment, while previous work exploits a local image feature and the global audio feature." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 420, + 234, + 547, + 342 + ], + "blocks": [ + { + "bbox": [ + 420, + 234, + 547, + 342 + ], + "lines": [ + { + "bbox": [ + 420, + 234, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 420, + 234, + 547, + 342 + ], + "type": "image", + "image_path": "cee2e6e0380f9fe5d0beeff613f7c05e3b77392cfb3ea91e72dbb29f5a6f7095.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 344, + 498, + 353 + ], + "lines": [ + { + "bbox": [ + 471, + 344, + 498, + 353 + ], + "spans": [ + { + "bbox": [ + 471, + 344, + 498, + 353 + ], + "type": "text", + "content": "(b) Ours" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 485 + ], + "type": "text", + "content": "and assume that, given an image-audio pair, at least one patch of the image is relevant to the audio; then they represent the image with its local feature that most closely resembles the global audio feature, and learn SSL through contrastive learning [24, 27] with these features (Fig. 1(a))." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 486, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 617 + ], + "type": "text", + "content": "While this approach has been proven to work, there is plenty of room for further improvement. First, audio may contain not only the sound of the target shared by the two modalities but also irrelevant sounds, such as noise or sounds from sources outside the image. In such cases, using the global audio feature for contrastive learning hinders accurate alignment between the image and audio features. Second, representing an image with only a single local feature in contrastive learning causes models to locate a small discriminative part of the sound source, not its whole body [35], which can degrade SSL accuracy." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "content": "To address these limitations, we propose a novel SSL method based on slot attention [17], which is conceptually illustrated in Fig. 1(b). In slot attention, slots, which are randomly sampled or learnable embeddings, compete with each other for aggregating input features progressively, and thus they decompose the features into distinct and exclusive representations. Motivated by this, we employ slot attention with two learnable slots to jointly decompose image" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3121" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 470, + 241 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 470, + 241 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 470, + 241 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 470, + 241 + ], + "type": "image", + "image_path": "9b4b6f866abc2ae0590ba4abe78b308ca4a7cf7e863535c21f098f90549188fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 252, + 555, + 318 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 555, + 318 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed method. First, image and audio features are extracted by the image and audio encoders, and serve as keys and values for slot attention. In slot attention, two learnable slots are used as queries to draw attentions on the input features while competing with each other, decomposing the features into the target and off-target representations per modality. The model is then trained by contrastive learning using only the target representations, and by cross-modal attention matching for further feature-level alignment between the two modalities. For localization in testing, we draw the attention map between the audio target representation and image features. Some details for learning slot attention, such as decoders and loss for slot reconstruction, have been omitted for clarity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 474, + 76, + 546, + 239 + ], + "blocks": [ + { + "bbox": [ + 474, + 76, + 546, + 239 + ], + "lines": [ + { + "bbox": [ + 474, + 76, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 474, + 76, + 546, + 239 + ], + "type": "image", + "image_path": "5103af55f0ec301d98f385465743772fba50039830256fc54a47baf0e4592ecd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 339, + 297, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 339, + 297, + 543 + ], + "spans": [ + { + "bbox": [ + 54, + 339, + 297, + 543 + ], + "type": "text", + "content": "and audio features into two distinct slot representations per modality, one for the target and the other for the rest. To be specific, we designate the first slot as the target slot, which is guided to focus on information related to the shared sound source across both modalities. The exclusive assignment property of slot attention directs the second slot, referred to as the off-target slot, to capture remaining elements, such as the image background or irrelevant sounds in the audio. Then we learn SSL by contrastive learning using the two target slot representations from the image and audio, respectively. Unlike the prior work using only a single global audio feature, our method allows to exploit only the relevant part (i.e., the target slot) of the audio for more accurate SSL. Furthermore, since the target slot of the image covers the sound source area rather than a single location, our method avoids the problem of localization being reduced to a small discriminative part." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "Given the target slot representations, SSL is conducted by drawing an attention map on the image feature map using the target slot representation of the audio as query (Fig. 2(b)). For accurate pixel-level localization, the target slot of the audio should be well aligned with the keys derived from relevant individual image features; in this regard, contrastive learning using only the target slots is insufficient as it does not relate the audio target slot with individual image features. We thus introduce an additional training strategy dubbed cross-modal attention matching. This strategy encourages the cross-modal attention map between the audio target slot and the input image features to approximate the intra-modal attention map between the image target slot and the input image features that are inherently well" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 339, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 411 + ], + "type": "text", + "content": "aligned. In other words, it provides pseudo-supervision for SSL prediction (i.e., the cross-modal attention map) using the intra-modal attention map, which is drawn by the query and keys from the same modality and thus better localizes the target. To further facilitate its cross-modal alignment, we apply this strategy to the audio modality as well." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 412, + 556, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 556, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 556, + 532 + ], + "type": "text", + "content": "Our model was trained on two datasets, Flickr-SoundNet [4] and VGG-Sound [5], and evaluated on three benchmarks, Flickr-SoundNet-Test [4], VGG-SS [5], and AVSBench [36]. It achieved the state of the art in almost all settings across all the three evaluation benchmarks. Further, our model outperformed all the prior work by large margin in cross-modal retrieval, suggesting that it is not a simple saliency detector but properly learns to relate image and audio for SSL [31]. In summary, the contribution of our method is three-fold:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 536, + 555, + 643 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 536, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 570 + ], + "type": "text", + "content": "- We propose a novel SSL model that identifies and extracts features relevant to sound source jointly from image and audio modalities through slot attention." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 571, + 554, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 554, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 554, + 607 + ], + "type": "text", + "content": "- We also propose cross-modal attention matching, which enables effective cross-modal alignment for individual input features and thus improves localization performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 608, + 553, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 553, + 643 + ], + "type": "text", + "content": "- Our model achieved the state of the art in almost all settings across all the three SSL benchmarks and improved cross-modal retrieval performance substantially." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 656, + 400, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 656, + 400, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 656, + 400, + 668 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Sound source localization aims to identify specific regions in an image that correspond to the sources of sounds from an audio input. Since constructing audio-visual datasets is" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3122" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "costly, sound source localization studies have evolved into self-supervised learning approaches that use only image-audio pairs for contrastive learning [1, 6, 12, 14, 16, 21, 22, 26, 29-31, 33]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 296, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 296, + 242 + ], + "type": "text", + "content": "An early study, [29], addresses the task through contrastive learning between the audio-aware image features and the global audio feature. The audio-aware image features are obtained by calculating a correspondence map through the normalized inner product between the global audio feature and the image feature map. Chen et al. [6] propose method that uses a differentiable threshold on the correspondence map to divide audio-aware and other image features, treating the latter as hard negatives in contrastive learning to improve localization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 243, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 296, + 495 + ], + "type": "text", + "content": "Recent studies focus on fine-grained sound source localization through contrastive learning between the global audio feature and the most relevant feature selected from the image feature map [21, 22, 31, 33]. Mo and Morgado [21] propose a method that refines the audio-visual correspondence map through a post-processing technique leveraging a pretrained visual model. Sun et al. [33] introduce a regularizer that applies mean absolute error regularization to inter-modal and intra-modality similarity matrices to increase awareness of false negatives. Senocak et al. [31] propose a cross-modal alignment method that preserves localization performance and enhances cross-modal retrieval by leveraging multi-view augmentation and nearest neighbor search to construct multiple positive pairs for contrastive learning. These methods employ contrastive learning with the most relevant image feature to assist in coarsely identifying sound sources but often fall short of capturing a comprehensive representation, thereby hindering localization ability. Furthermore, relying solely on global audio features can mislead contrastive learning when noise exists, potentially highlighting incorrect image features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 496, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 604 + ], + "type": "text", + "content": "Following a complementary direction, several studies employ external prior knowledge or even utilize off-the-shelf segmentation modules to achieve high-quality localization results [25, 26]. Also, several studies employ sound source localization with additional information such as optical flow [9], text inputs [19] or ground truth category [23]. By contrast to these studies, our method can effectively perform sound source localization without relying on external prior knowledge or off-the-shelf models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "To address the above issues, we propose a sound source localization framework based on the slot attention mechanism. Slot attention [17] is proposed for unsupervised object-centric learning, effectively decomposing input features into distinct and mutually exclusive components via iterative attention and competition among slots. Inspired by this, we adopt slot attention in both image and audio features to decompose them into features related to the sound source and those irrelevant to it without supervision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 71, + 420, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 420, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 420, + 85 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 91, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 258 + ], + "type": "text", + "content": "As shown in Fig. 2, our model consists of two parts: 1) image and audio encoders for feature extraction, and 2) joint feature decomposition module with joint slot attention mechanism. A pair of image and audio is first fed into each encoder to extract image and audio features (Sec. 3.1). The features of each modality are decomposed into two distinct slot representations by using joint slot attention: one slot for the target and the other for the rest. (Sec. 3.2). The slot representations are then used for training as input to contrastive learning and cross-modal attention matching (Sec. 3.3), while effectively handling false negatives (Sec. 3.4). Lastly, an attention on the image feature map using the target slot representation of the audio is computed to locate the source of sound (Sec. 3.5)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 265, + 425, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 425, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 425, + 277 + ], + "type": "text", + "content": "3.1. Feature Extraction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "D = \\{(v_{i},a_{i})\\}_{i = 1}^{M}" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": " denote a training set of image-audio pairs, where " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "v_{i}\\in \\mathbb{R}^{3\\times H_{v}\\times W_{v}}" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": " is an image of size " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "H_{v}\\times W_{v}" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "a_{i}\\in \\mathbb{R}^{1\\times F_{a}\\times T_{a}}" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": " is a corresponding audio spectrogram of frequency resolution " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": " and temporal length " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "T_{a}" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": " denotes the total number of such pairs in the dataset. For feature extraction, separate image and audio encoders " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": ", are adopted. The image encoder extracts the image features " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i\\coloneqq f_v(v_i)\\in \\mathbb{R}^{h\\times w\\times c}" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": ", and the audio encoder extracts the audio features " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i\\coloneqq f_a(a_i)\\in \\mathbb{R}^{t\\times c}" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": " are the height and width of the image feature map respectively, " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": " is the length of the audio features, and " + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 283, + 555, + 427 + ], + "type": "text", + "content": " is the image and audio feature dimension." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 434, + 471, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 471, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 471, + 447 + ], + "type": "text", + "content": "3.2. Joint Feature Decomposition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 453, + 555, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 619 + ], + "type": "text", + "content": "Slot attention [17] is an attention mechanism where each slot progressively competes to aggregate input features, leading to the decomposition of input features into distinct and exclusive representations. Building on this idea, for joint feature decomposition, we propose joint slot attention (JSA), which jointly decomposes image and audio features into two distinct slot representations for each modality. To achieve this, we employ slot attention with two initial slots " + }, + { + "bbox": [ + 313, + 453, + 555, + 619 + ], + "type": "inline_equation", + "content": "S_0 \\in \\mathbb{R}^{2 \\times D}" + }, + { + "bbox": [ + 313, + 453, + 555, + 619 + ], + "type": "text", + "content": " shared in the two modalities. Furthermore, we designate the first slot as target slot, which captures features relevant to sound source, and the other as off-target slot, which captures features for the rest. Since the overall pipeline of JSA is identical in the two modalities, we will describe the module in terms of image only for brevity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 620, + 554, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 554, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 554, + 655 + ], + "type": "text", + "content": "JSA first projects an input image feature " + }, + { + "bbox": [ + 313, + 620, + 554, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 620, + 554, + 655 + ], + "type": "text", + "content": " to derive its corresponding key " + }, + { + "bbox": [ + 313, + 620, + 554, + 655 + ], + "type": "inline_equation", + "content": "K^{\\nu}" + }, + { + "bbox": [ + 313, + 620, + 554, + 655 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 313, + 620, + 554, + 655 + ], + "type": "inline_equation", + "content": "V^{\\nu}" + }, + { + "bbox": [ + 313, + 620, + 554, + 655 + ], + "type": "text", + "content": " by applying the following operations:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 662, + 553, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 662, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 354, + 662, + 553, + 682 + ], + "type": "interline_equation", + "content": "K ^ {v} = W ^ {k} (\\operatorname {L a y e r N o r m} (\\mathbf {v})) \\in \\mathbb {R} ^ {h w \\times d}, \\tag {1}", + "image_path": "75a630a4642073cadba7a23a3bdce0a4e7d98909c445245b2747d58d42fb915b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 679, + 513, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 679, + 513, + 693 + ], + "spans": [ + { + "bbox": [ + 358, + 679, + 513, + 693 + ], + "type": "interline_equation", + "content": "V ^ {v} = W ^ {v} \\left(\\operatorname {L a y e r N o r m} (\\mathbf {v})\\right) \\in \\mathbb {R} ^ {h w \\times d},", + "image_path": "2ebd58452f65b10eac66aab1e5bdca76c1812fc6d032063337d6df2c041bfc6f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 700, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 700, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 700, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 700, + 555, + 714 + ], + "type": "inline_equation", + "content": "W^{k}" + }, + { + "bbox": [ + 313, + 700, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 700, + 555, + 714 + ], + "type": "inline_equation", + "content": "W^{v}" + }, + { + "bbox": [ + 313, + 700, + 555, + 714 + ], + "type": "text", + "content": " are the linear projection layers to obtain" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3123" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "embeddings in " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "-dimension. For each iteration " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "n \\in [1, N]" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": ", the query " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "Q_{n}^{v}" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": " is derived by applying linear projection to the slots with the initial slots " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "S_{0}^{v} \\coloneqq S_{0}" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": ", which is given by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 92, + 116, + 295, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 116, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 92, + 116, + 295, + 131 + ], + "type": "interline_equation", + "content": "Q _ {n} ^ {v} = W ^ {q} \\left(\\operatorname {L a y e r N o r m} \\left(S _ {n - 1} ^ {v}\\right)\\right) \\in \\mathbb {R} ^ {2 \\times d}, \\tag {2}", + "image_path": "3b12830d01a85762926f7b7ccc0da768322c4f718d3a2626f4131702eda9641a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 138, + 296, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 296, + 187 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 138, + 296, + 187 + ], + "type": "inline_equation", + "content": "W^q" + }, + { + "bbox": [ + 55, + 138, + 296, + 187 + ], + "type": "text", + "content": " is a linear projection layer shared for all iterations. The slot attention matrix " + }, + { + "bbox": [ + 55, + 138, + 296, + 187 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 138, + 296, + 187 + ], + "type": "text", + "content": " is computed by taking the dot product of the given query " + }, + { + "bbox": [ + 55, + 138, + 296, + 187 + ], + "type": "inline_equation", + "content": "Q_n^v" + }, + { + "bbox": [ + 55, + 138, + 296, + 187 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 55, + 138, + 296, + 187 + ], + "type": "inline_equation", + "content": "K^v" + }, + { + "bbox": [ + 55, + 138, + 296, + 187 + ], + "type": "text", + "content": ", scaled by the square root of the dimension " + }, + { + "bbox": [ + 55, + 138, + 296, + 187 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 138, + 296, + 187 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 194, + 299, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 299, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 299, + 232 + ], + "type": "interline_equation", + "content": "A _ {i, j} = \\frac {\\exp \\left(M _ {i , j}\\right)}{\\sum_ {l = 1} ^ {2} \\exp \\left(M _ {i , l}\\right)}, \\text {w h e r e} M = \\frac {K ^ {v} \\left(Q _ {n} ^ {v}\\right) ^ {\\top}}{\\sqrt {d}} \\in \\mathbb {R} ^ {h w \\times 2}, \\tag {3}", + "image_path": "8199a89188c517b77eba8e68f1348b268e974ca31e7d3c04378a9d57bc34327a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 233, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 296, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 233, + 296, + 258 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 233, + 296, + 258 + ], + "type": "text", + "content": " is key index, and " + }, + { + "bbox": [ + 55, + 233, + 296, + 258 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 233, + 296, + 258 + ], + "type": "text", + "content": " is slot index. Subsequently, a weighted mean is utilized to the slot attention matrix " + }, + { + "bbox": [ + 55, + 233, + 296, + 258 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 233, + 296, + 258 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 264, + 295, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 264, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 135, + 264, + 295, + 293 + ], + "type": "interline_equation", + "content": "\\hat {A} _ {i, j} = \\frac {A _ {i , j}}{\\sum_ {l = 1} ^ {h w} A _ {l , j}}. \\tag {4}", + "image_path": "d67ae9bd3041a6355bbf4b77739887be10ada6fad87279480b32ac81ed3986fd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 300, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 337 + ], + "type": "text", + "content": "Then, the updated slot " + }, + { + "bbox": [ + 55, + 300, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\bar{S}_n" + }, + { + "bbox": [ + 55, + 300, + 296, + 337 + ], + "type": "text", + "content": " is computed by Gated Recurrent Unit (GRU) [7] with " + }, + { + "bbox": [ + 55, + 300, + 296, + 337 + ], + "type": "inline_equation", + "content": "S_{n - 1}" + }, + { + "bbox": [ + 55, + 300, + 296, + 337 + ], + "type": "text", + "content": " as hidden state, and " + }, + { + "bbox": [ + 55, + 300, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\hat{A}^{\\top}\\cdot V^{v}" + }, + { + "bbox": [ + 55, + 300, + 296, + 337 + ], + "type": "text", + "content": " as input as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 344, + 295, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 344, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 114, + 344, + 295, + 360 + ], + "type": "interline_equation", + "content": "\\bar {S _ {n} ^ {v}} = \\operatorname {G R U} \\left(S _ {n - 1} ^ {v}, \\hat {A} ^ {\\top} \\cdot V ^ {v}\\right). \\tag {5}", + "image_path": "7b88221f011ded8a1fa6766909540db3853a8654931ec7856e0d2daca5929923.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 367, + 266, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 266, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 266, + 380 + ], + "type": "text", + "content": "Finally, the slot is updated by a residual connection:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 84, + 386, + 295, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 386, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 84, + 386, + 295, + 402 + ], + "type": "interline_equation", + "content": "S _ {n} ^ {v} = \\bar {S} _ {n} ^ {v} + \\operatorname {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(\\bar {S} _ {n} ^ {v}\\right)\\right) \\in \\mathbb {R} ^ {2 \\times d}. \\tag {6}", + "image_path": "7efefb2fff728484c0e53217b9101ffd79890a9313fa6092b0a98eedc4e72e22.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "text", + "content": "The overall process encourages the input feature to bind to specific slots, ensuring that each slot forms a mutually exclusive representation. The final outputs of JSA are image slots " + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}^v;\\mathbf{r}^v\\} := S_N^v\\in \\mathbb{R}^{2\\times d}" + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "text", + "content": " and image query " + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "inline_equation", + "content": "Q_{N}^{v}\\in" + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{2\\times d}" + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^v" + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "text", + "content": " is a target slot and " + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{r}^v" + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "text", + "content": " is an off-target slot. The same process is applied to the audio modality using audio feature " + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "text", + "content": ", yielding audio slots " + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}^a;\\mathbf{r}^a\\} := S_N^a\\in \\mathbb{R}^{2\\times d}" + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "text", + "content": " and audio query " + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "inline_equation", + "content": "Q_{N}^{a}\\in \\mathbb{R}^{2\\times d}" + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "text", + "content": ". Since this process begins from the shared initial slots, the initial slots are able to effectively learn joint features between two modalities. For brevity, we omit the subscript " + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "text", + "content": " for the final query, " + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "inline_equation", + "content": "Q^{v}\\coloneqq Q_{N}^{v}" + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "inline_equation", + "content": "Q^{a}\\coloneqq Q_{N}^{a}" + }, + { + "bbox": [ + 55, + 409, + 296, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 559, + 173, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 173, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 173, + 572 + ], + "type": "text", + "content": "3.3. Learning Objectives" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "text", + "content": "Contrastive Learning with Target Slots. We adopt the InfoNCE loss [24, 27] for contrastive learning using the target slots, guiding the target slots to identify features relevant to the sound source jointly from image and audio. Let " + }, + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "text", + "content": " denote a similarity function, " + }, + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "inline_equation", + "content": "s(x,y) = \\exp (\\cos (x,y) / \\tau)" + }, + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "text", + "content": " is a temperature hyperparameter. The contrastive learning loss " + }, + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cotr}}" + }, + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 670, + 304, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 304, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 304, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o t r}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left\\{\\log \\frac {s \\left(\\mathbf {p} _ {i} ^ {v} , \\mathbf {p} _ {i} ^ {a}\\right)}{\\sum_ {j} ^ {B} s \\left(\\mathbf {p} _ {i} ^ {v} , \\mathbf {p} _ {j} ^ {a}\\right)} + \\log \\frac {s \\left(\\mathbf {p} _ {i} ^ {v} , \\mathbf {p} _ {i} ^ {a}\\right)}{\\sum_ {j} ^ {B} s \\left(\\mathbf {p} _ {j} ^ {v} , \\mathbf {p} _ {i} ^ {a}\\right)} \\right\\},", + "image_path": "4e2e64403299f6620761af904371dd162d600def5443ffe22b10ece8a1bde907.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 554, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 179 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 179 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 72, + 554, + 179 + ], + "type": "text", + "content": " denotes the number of data in a mini-batch, and the negative pairs refer to the combinations of visual and audio features that do not belong to the same original pair within the mini-batch. Since the image target slot captures the whole features of the target region, it avoids focusing solely on a specific discriminative part of the image. Furthermore, the audio target slot captures only the relevant sound from the source, it prevents the issue of learning sounds that are unrelated to the source." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "text", + "content": "Cross-modal Attention Matching. As will be described in Sec. 3.5, the localization result is derived by utilizing cross-modal attention between the audio target slot and image features. For accurate sound source localization, the audio target slot should be properly aligned with keys computed from the relevant image features. Although contrastive learning encourages alignment of target slot representations between different modalities, it does not ensure alignment between the audio target slot and image features, which degrades the localization accuracy. We resolve this issue by cross-modal attention matching, which encourages the cross-modal attention map to resemble the intra-modal attention map for improving feature-level alignment across the two modalities. To this end, we first calculate the cross-modal and intra-modal attention matrices. Specifically, we compute the cross-modal attention matrix between the audio and visual modalities in the same manner as Eq. (3) and Eq. (4), with the audio queries " + }, + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "inline_equation", + "content": "Q^a" + }, + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "text", + "content": " and visual keys " + }, + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "inline_equation", + "content": "K^v" + }, + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "text", + "content": ". We denote this cross-modal attention matrix as " + }, + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "inline_equation", + "content": "\\hat{A}^{a,v} \\in \\mathbb{R}^{hw \\times 2}" + }, + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "inline_equation", + "content": "hw" + }, + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "text", + "content": " is the length of " + }, + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "inline_equation", + "content": "K^v" + }, + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "text", + "content": ". Subsequently, to obtain the cross-modal attention that represents the attention weights between the audio target query and visual keys, we utilize only the first column of " + }, + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "inline_equation", + "content": "\\hat{A}^{a,v}" + }, + { + "bbox": [ + 313, + 180, + 555, + 467 + ], + "type": "text", + "content": ". This cross-modal attention is defined as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 378, + 472, + 489, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 472, + 489, + 488 + ], + "spans": [ + { + "bbox": [ + 378, + 472, + 489, + 488 + ], + "type": "interline_equation", + "content": "\\mathbf {c a} ^ {a, v} := \\left[ \\hat {A} _ {i, 1} ^ {a, v} \\right] _ {i = 1} ^ {h w} \\in \\mathbb {R} ^ {h w}.", + "image_path": "d3fe742734cf1161728e9396a0da8310b0ff6c248dcaf7f4cbff4255b59adf96.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 492, + 554, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 554, + 553 + ], + "type": "text", + "content": "Similarly, for intra-modal attention matrix within the visual modality, we compute the intra-attention matrix " + }, + { + "bbox": [ + 313, + 492, + 554, + 553 + ], + "type": "inline_equation", + "content": "\\hat{A}^{v,v}" + }, + { + "bbox": [ + 313, + 492, + 554, + 553 + ], + "type": "text", + "content": " using the " + }, + { + "bbox": [ + 313, + 492, + 554, + 553 + ], + "type": "inline_equation", + "content": "Q^v" + }, + { + "bbox": [ + 313, + 492, + 554, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 492, + 554, + 553 + ], + "type": "inline_equation", + "content": "K^v" + }, + { + "bbox": [ + 313, + 492, + 554, + 553 + ], + "type": "text", + "content": ". Then, the intra-modal attention that represents the attention weights between the image target query and visual keys is defined as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 378, + 560, + 489, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 560, + 489, + 576 + ], + "spans": [ + { + "bbox": [ + 378, + 560, + 489, + 576 + ], + "type": "interline_equation", + "content": "\\mathbf {i a} ^ {v, v} := \\left[ \\hat {A} _ {i, 1} ^ {v, v} \\right] _ {i = 1} ^ {h w} \\in \\mathbb {R} ^ {h w}.", + "image_path": "4ee52d01fc868cbd35a48a0c5a909bd528c48cd024e8dc1392fabe6a5944d7bc.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 585, + 554, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 554, + 632 + ], + "type": "text", + "content": "To further facilitate cross-modal alignment, we apply this strategy symmetrically to the audio modality as well. Finally we formulate the cross-modal attention matching loss as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 319, + 642, + 555, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 642, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 319, + 642, + 555, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a t c h}} = \\left\\| \\mathbf {c a} ^ {a, v} - \\operatorname {s g} \\left(\\mathbf {i a} ^ {v, v}\\right) \\right\\| _ {2} ^ {2} + \\left\\| \\mathbf {c a} ^ {v, a} - \\operatorname {s g} \\left(\\mathbf {i a} ^ {a, a}\\right) \\right\\| _ {2} ^ {2}, \\tag {8}", + "image_path": "6255620e40f96bf5cfca387242b862eeb8f85cbe4069635815a2c6bd73e29059.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}(\\cdot)" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " denotes the stop-gradient operation. " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " minimizes the discrepancy between cross-modal and intramodal attentions, promoting robust feature-level alignment between the two modalities for accurate localization." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3124" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "Slot Divergence. In slot attention, the diversity between the target slot and off-target slot should be large enough to ensure that each slot effectively captures distinct, mutually exclusive regions. Therefore, we introduce a slot divergence loss that encourages greater diversity among the slots. The slot divergence loss " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{div}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 148, + 295, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 148, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 63, + 148, + 295, + 162 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d i v}} = \\max \\left\\{0, \\cos \\left(\\mathbf {p} ^ {v}, \\mathbf {r} ^ {v}\\right) \\right\\} + \\max \\left\\{0, \\cos \\left(\\mathbf {p} ^ {a}, \\mathbf {r} ^ {a}\\right) \\right\\}. \\tag {9}", + "image_path": "62a9ab6c46eb18cc97cfb5c2a532109de7aaba09dea167497596f8e38e896b09.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 166, + 294, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 166, + 294, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 294, + 201 + ], + "type": "text", + "content": "It ensures that the different slot embeddings to be sufficiently dissimilar, enabling the target slot to more effectively localize the sound source." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 202, + 295, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 295, + 297 + ], + "type": "text", + "content": "Slot Reconstruction. The successful feature reconstruction in slot attention implies that each slot successfully captures unique and independent parts of the input data, which means that the slot attention mechanism effectively decomposed the input into individual slot representations. It indicates that each slot accurately represents a specific object or region in the input. To encourage this condition, we apply the slot reconstruction loss " + }, + { + "bbox": [ + 55, + 202, + 295, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 55, + 202, + 295, + 297 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 301, + 295, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 301, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 301, + 295, + 316 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\| \\mathbf {v} - g _ {v} \\left(\\left\\{\\mathbf {p} ^ {v}; \\mathbf {r} ^ {v} \\right\\}\\right) \\| _ {2} ^ {2} + \\| \\mathbf {a} - g _ {a} \\left(\\left\\{\\mathbf {p} ^ {a}; \\mathbf {r} ^ {a} \\right\\}\\right) \\| _ {2} ^ {2}, \\tag {10}", + "image_path": "c959de5d082c4e8daa455fffca91e45fdcbb44a0290a7c5753f8768c049986ee.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 319, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 319, + 295, + 378 + ], + "type": "inline_equation", + "content": "g_{v}" + }, + { + "bbox": [ + 55, + 319, + 295, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 295, + 378 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 55, + 319, + 295, + 378 + ], + "type": "text", + "content": " are the decoders for the image and audio features, respectively. These decoders are composed of a small number of MLP layers with ReLU activations. Note that the decoders are auxiliary modules used only in training and not required for inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "content": "The overall loss is a combination of " + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cotr}}" + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{div}}" + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "content": ", which is given by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 408, + 294, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 408, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 73, + 408, + 294, + 421 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o t r}} + \\lambda_ {1} \\mathcal {L} _ {\\text {m a t c h}} + \\lambda_ {2} \\mathcal {L} _ {\\text {d i v}} + \\lambda_ {3} \\mathcal {L} _ {\\text {r e c o n}}, \\tag {11}", + "image_path": "4f13ee1ff5b34a9b8a67a794b79150c377f4ba2b4f7444b76c03a21a08d8576e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 426, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 426, + 295, + 448 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 426, + 295, + 448 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 426, + 295, + 448 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 426, + 295, + 448 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 426, + 295, + 448 + ], + "type": "inline_equation", + "content": "\\lambda_{3}" + }, + { + "bbox": [ + 55, + 426, + 295, + 448 + ], + "type": "text", + "content": " are hyperparameters weighting the loss terms." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 455, + 202, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 202, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 202, + 467 + ], + "type": "text", + "content": "3.4. Mitigating False Negatives" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 473, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 544 + ], + "type": "text", + "content": "In general, unpaired images and audiios are all treated as negatives for contrastive training. However, some of them are false negatives that are semantically well matched, so treating these as negatives causes adverse effect [33]. Therefore, we propose a false negative mitigation strategy using " + }, + { + "bbox": [ + 55, + 473, + 295, + 544 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 473, + 295, + 544 + ], + "type": "text", + "content": "-reciprocal nearest neighbors of target slots." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 544, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 570 + ], + "type": "text", + "content": "For the image target slot " + }, + { + "bbox": [ + 55, + 544, + 295, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i^v" + }, + { + "bbox": [ + 55, + 544, + 295, + 570 + ], + "type": "text", + "content": ", which is derived from " + }, + { + "bbox": [ + 55, + 544, + 295, + 570 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 544, + 295, + 570 + ], + "type": "text", + "content": " we define its set of " + }, + { + "bbox": [ + 55, + 544, + 295, + 570 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 544, + 295, + 570 + ], + "type": "text", + "content": "-reciprocal nearest neighbors " + }, + { + "bbox": [ + 55, + 544, + 295, + 570 + ], + "type": "inline_equation", + "content": "R_{k}^{v}(i)" + }, + { + "bbox": [ + 55, + 544, + 295, + 570 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 573, + 295, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 573, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 67, + 573, + 295, + 587 + ], + "type": "interline_equation", + "content": "R _ {k} ^ {v} (i) = \\left\\{\\left(v _ {j}, a _ {j}\\right) \\mid \\left(j \\in N _ {k} ^ {v} (i)\\right) \\cap \\left(i \\in N _ {k} ^ {v} (j)\\right) \\right\\}, \\tag {12}", + "image_path": "78555520ac9a3493e3cbf718201e2ec7bfa478f761c4dc47dd83d2a733d7ead7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "inline_equation", + "content": "N_{k}^{v}(i)" + }, + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "text", + "content": "-nearest neighbors of " + }, + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i^v" + }, + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "text", + "content": ", computed using cosine similarity. Likewise, the set of " + }, + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "text", + "content": "-reciprocal nearest neighbors between audio target slots, i.e., " + }, + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "inline_equation", + "content": "R_{k}^{a}(i)" + }, + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "text", + "content": ", can be derived. The predicted false negative set " + }, + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "inline_equation", + "content": "R_{k}(i)" + }, + { + "bbox": [ + 55, + 590, + 295, + 649 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 651, + 295, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 651, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 125, + 651, + 295, + 663 + ], + "type": "interline_equation", + "content": "R _ {k} (i) = R _ {k} ^ {v} (i) \\cap R _ {k} ^ {a} (i). \\tag {13}", + "image_path": "cea5750ec3414caf20645c3e4e1f9afb1305f0a6e43acc6f2b189d33f15299f3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Since the samples in " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "R_{k}(i)" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": " have both similar image and audio target slots with " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "(v_{i}, a_{i})" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": ", they are likely to belong to false negatives. We exclude these predicted false negatives from the contrastive learning in Eq. (7)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 72, + 382, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 382, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 382, + 83 + ], + "type": "text", + "content": "3.5. Inference" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 91, + 553, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 149 + ], + "type": "text", + "content": "For inference, " + }, + { + "bbox": [ + 313, + 91, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\mathbf{ca}^{a,v}" + }, + { + "bbox": [ + 313, + 91, + 553, + 149 + ], + "type": "text", + "content": ", which represents the attention between the audio target slot and the input image features, is utilized. Specifically, the attention is interpolated to the original image size, and pixels whose score exceeds the hyperparameter " + }, + { + "bbox": [ + 313, + 91, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 91, + 553, + 149 + ], + "type": "text", + "content": " are selected to generate the localization result." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 150, + 555, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 555, + 270 + ], + "type": "text", + "content": "Furthermore, unlike the object-guided localization [21], we propose a novel localization refinement strategy, dubbed Image-Query based Refinement (IQR), to refine the localization result without relying on external prior knowledge. IQR estimates the regions that are likely to contain the target by using the attention between the image target slot and the input image features. It can be used to identify prior regions in the image that favor the target, regardless of whether these regions are actually the source of the sound. The refined result derived by IQR, denoted as " + }, + { + "bbox": [ + 313, + 150, + 555, + 270 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{IQR}}" + }, + { + "bbox": [ + 313, + 150, + 555, + 270 + ], + "type": "text", + "content": ", is given as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 362, + 280, + 553, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 280, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 362, + 280, + 553, + 293 + ], + "type": "interline_equation", + "content": "R _ {\\mathrm {I Q R}} = \\alpha \\cdot \\mathbf {c a} ^ {a, v} + (1 - \\alpha) \\cdot \\mathbf {i a} ^ {v, v}, \\tag {14}", + "image_path": "6edeba71e29f9c0481eb15b1f3be3b2a9f378731ab68cba7dbac5f7c4237d151.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 304, + 553, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 553, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 304, + 553, + 327 + ], + "type": "inline_equation", + "content": "\\alpha \\in (0,1)" + }, + { + "bbox": [ + 313, + 304, + 553, + 327 + ], + "type": "text", + "content": " balances the sound source localization result and potential sound source regions." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 339, + 391, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 339, + 391, + 353 + ], + "spans": [ + { + "bbox": [ + 314, + 339, + 391, + 353 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 359, + 422, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 422, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 422, + 372 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 377, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 581 + ], + "type": "text", + "content": "Datasets. Flickr-SoundNet [4] and VGG-Sound [5] are used to train. Flickr-SoundNet contains about 2 million unconstrained videos from Flickr, while VGG-Sound has around 200k video clips. For training, we use subsets of 10k and 144k from each dataset, respectively. For the evaluation, Flickr-SoundNet-Test [4], VGG-SS [5], and AVSBench [36] are used following the previous research [6, 21, 31, 33]. Flickr-SoundNet-Test has 250 audio-visual pairs with manually labeled bounding boxes. VGG-SS has about 5k audio-visual pairs over 220 categories. Since VGG-SS provides its category ground truth, it is also used to evaluate cross-modal retrieval performance. AVSBench has about 5k videos with pixel-wise binary segmentation maps. Following the previous study [6], the middle frame of the video clip is used for the image. For the audio, the middle 5 seconds are extracted and converted into log spectrograms with 257 frequency bins at a sampling rate of " + }, + { + "bbox": [ + 313, + 377, + 555, + 581 + ], + "type": "inline_equation", + "content": "16\\mathrm{kHz}" + }, + { + "bbox": [ + 313, + 377, + 555, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 582, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 665 + ], + "type": "text", + "content": "Network Architecture. ResNet-18 [11] is adopted for both the image encoder and the audio encoders. The image encoder is fine-tuned from an ImageNet [8] pretrained model, while the audio encoder is trained from scratch. Both encoders include an additional " + }, + { + "bbox": [ + 313, + 582, + 554, + 665 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 582, + 554, + 665 + ], + "type": "text", + "content": " convolutional layer for projection, and the audio encoder also includes a MaxPooling layer for pooling along the frequency axis." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": "Hyperparameter Settings. We use the AdamW optimizer [18] with a learning rate of 5e-5 and a weight decay of 1e-2. The mini-batch size during training is set to 256, and the size of image and audio feature size " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "c" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3125" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 301, + 288 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 301, + 288 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 301, + 288 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 301, + 288 + ], + "type": "table", + "html": "
TrainsetMethodOGLVGG-SSFlickr-SoudNet
cIoUAUCcIoUAUC
Flickr SoundNet 10kAttention10k [29]--43.6044.90
CoursetoFine [26]--52.2049.60
AVObject [1]--54.6050.40
LVS [6]--58.2052.50
EZ-VSL [21]19.8630.9662.2454.74
FNAC [33]35.2738.0084.3363.26
Ours35.1138.6685.2065.26
EZ-VSL [21]✓37.6139.2181.9362.58
FNAC [33]✓40.9740.3884.7364.34
Ours✓39.1640.5187.6066.50
Flickr SoundNet 144kAttention10k [29]--66.0055.80
DMC [12]--67.1056.80
LVS [6]--69.9057.30
HardPos [30]--75.2059.70
EZ-VSL [21]30.2735.9272.6958.70
FNAC [33]33.9337.2978.7159.33
Alignment [31]--85.2062.20
Ours37.0139.4286.0065.16
EZ-VSL [21]✓41.0140.2383.1363.06
FNAC [33]✓41.1040.4483.9363.06
Alignment [31]✓--84.0064.16
Ours✓40.1540.7589.2064.50
", + "image_path": "2b4f21a720c1d8451c9596f8b0b675fb2a6c860ea995afe70f5afbc8ce968b0b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 313, + 70, + 553, + 287 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 553, + 287 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 287 + ], + "type": "table", + "html": "
TrainsetMethodOGLVGG-SSFlickr-SoudNet
cIoUAUCcIoUAUC
VGGSound 10kLVS [6]--61.8053.60
EZ-VSL [21]25.8433.6863.8554.44
FNAC [33]37.2938.9985.7463.66
Ours37.8439.8286.0064.90
EZ-VSL [21]✓38.7139.8078.7161.53
FNAC [33]✓40.6940.4282.1363.64
Ours✓41.0241.2987.6065.42
VGGSound 144kAttention10k [29]18.5030.20--
DMC [12]29.1034.80--
AVObject [1]29.7035.70--
LVS [6]34.4038.2073.5059.00
HardPos [30]34.6038.0076.8059.20
EZ-VSL [21]34.3837.7079.5161.17
SLAVC [22]37.22-83.20-
FNAC [33]39.5039.6684.7363.76
Alignment [31]39.9440.0279.6063.44
NoPrior [14]41.441.2--
Ours40.6241.6683.6064.68
EZ-VSL [21]✓38.8539.5483.9463.60
SLAVC [22]✓39.67-86.40-
FNAC [33]✓41.8540.8085.1464.30
Alignment [31]✓42.6441.4882.4064.60
Ours✓42.5642.4684.8064.82
", + "image_path": "a28be073f60b3512e3723eba657939c34b8bc8c060ad6821bc47a98ceec6a883.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 295, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 55, + 295, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 555, + 329 + ], + "type": "text", + "content": "Table 1. Quantitative results of the model trained on Flickr-SoundNet-10k, 144k, and VGGSound-10k, 144k datasets. Note that the results of previous research are obtained from [14, 31, 33]. The results of our method represent the performance after applying Image-Query based Refinement (IQR). The result with OGL denotes the performance after applying object-guided localization [21]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 72, + 339, + 536, + 483 + ], + "blocks": [ + { + "bbox": [ + 72, + 339, + 536, + 483 + ], + "lines": [ + { + "bbox": [ + 72, + 339, + 536, + 483 + ], + "spans": [ + { + "bbox": [ + 72, + 339, + 536, + 483 + ], + "type": "image", + "image_path": "0960ed3c5dcdee521c0e1003cee3986a7f06a785f4ddb4a179cee306675f0003.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 484, + 555, + 518 + ], + "lines": [ + { + "bbox": [ + 55, + 484, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 555, + 518 + ], + "type": "text", + "content": "Figure 3. Sound localization results on Flickr-SoundNet-Test [4] and VGG-SS [5]. (a) Input image. (b) Ground-Truth. (c) Ours. (d) Alignment [31]. (e) FNAC [33]. (f) EZ-VSL [21]. The qualitative results are obtained by the model trained on Flickr-144k and the model trained on VGGSound-144k, respectively. Note that all visualizations are obtained without refinement." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "text", + "content": "are set to 7, 7, 16, and 512 respectively. " + }, + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "text", + "content": " in Eq. (7) is set to 0.03. " + }, + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "inline_equation", + "content": "\\lambda_{1}, \\lambda_{2}" + }, + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "inline_equation", + "content": "\\lambda_{3}" + }, + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "text", + "content": " in Eq. (11) are set to 100.0, 0.1, and 0.1, respectively. " + }, + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "text", + "content": " in Eq. (14) is set to 0.6. The value " + }, + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "text", + "content": " in Eq. (13) is set to 20. The number of iterations " + }, + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 538, + 295, + 597 + ], + "type": "text", + "content": " for slot attention mechanism is set to 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 599, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 658 + ], + "type": "text", + "content": "Evaluation metric. We measure the average precision at a Consensus Intersection over Union (cIoU) threshold of 0.5 to evaluate localization accuracy. Additionally, the AUC score is used to provide an overall performance measure across different thresholds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 670, + 174, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 174, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 174, + 682 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Flickr-SoundNet and VGGSound. In this section, we compare our method with previous self-supervised SSL" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 538, + 555, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 555, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 555, + 670 + ], + "type": "text", + "content": "approaches. We trained our model on subsets of Flickr-SoundNet and VGGSound with dataset sizes of 10k and 144k, and evaluated the performance on Flickr-SoundNet-Test and VGG-SS. As shown in Table 1, our model achieves the state-of-the-art performance in most settings. Specifically, our method yields AUC improvements across all evaluated settings. Moreover, even without using Object-Guided Localization (OGL), our model performs competitively or better compared to previous methods, demonstrating the robustness of our architecture in focusing on the sound source without additional refinement steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Heard 110 and Unheard 110. To assess the generalization ability of SSL, we evaluate the performance with Heard 110 and Unheard 110. Heard 110 and Unheard 110 are sub" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3126" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 71, + 289, + 209 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 289, + 209 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 289, + 209 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 289, + 209 + ], + "type": "table", + "html": "
MethodOGLHeard 110Unheard 110
cIoUAUCcIoUAUC
LVS [6]28.9036.2026.3034.70
EZ-VSL [21]31.8636.1932.6636.72
SLAVC [22]35.84-36.50-
Alignment [31]38.3139.0539.1139.80
Ours38.4640.1540.3541.25
EZ-VSL [21]✓37.2538.9739.5739.60
SLAVC [22]✓38.22-38.87-
FNAC [33]✓39.5439.8342.9141.17
Alignment [31]✓41.8540.9342.9441.54
Ours✓41.5041.6043.1142.04
", + "image_path": "09051436e92880583f1db94770ab26f2939c391da675d228d88b99f0ca84847b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 85, + 298, + 269, + 399 + ], + "blocks": [ + { + "bbox": [ + 55, + 217, + 295, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 283 + ], + "type": "text", + "content": "Table 2. Quantitative results on Heard 110 and Unheard 110. Note that the results of previous research are obtained from [31, 33]. The results of our method represent the performance after applying Image-Query based Refinement (IQR). The result with OGL denotes the performance after applying object-guided localization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 298, + 269, + 399 + ], + "lines": [ + { + "bbox": [ + 85, + 298, + 269, + 399 + ], + "spans": [ + { + "bbox": [ + 85, + 298, + 269, + 399 + ], + "type": "table", + "html": "
TestsetMethodmIoUF-score
S4LVS [6]26.933.6
EZ-VSL [21]27.634.2
SLAVC [22]28.134.6
FNAC [33]27.1531.4
Alignment [31]29.635.9
\\( Alignment^† \\)29.335.6
Ours31.3344.30
", + "image_path": "d509a7b7f68994cf3721f5e6c7212740f93c159a1e92317664479ed0b18329c9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 456, + 294, + 553 + ], + "blocks": [ + { + "bbox": [ + 55, + 407, + 295, + 441 + ], + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 441 + ], + "type": "text", + "content": "Table 3. Quantitative results on AVSBench. Note that the results of previous research are obtained from [31, 33]. " + }, + { + "bbox": [ + 55, + 407, + 295, + 441 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 407, + 295, + 441 + ], + "type": "text", + "content": " denotes that the self-supervised pretrained encoder is utilized for their work." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 456, + 294, + 553 + ], + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 553 + ], + "type": "table", + "html": "
MethodAudio → ImageImage → Audio
R@1R@5R@10R@1R@5R@10
LVS [6]3.8712.3520.734.9014.2921.37
EZ-VSL [21]5.0115.7324.8114.2033.5145.18
SSL-TIE [16]10.2930.6843.7612.7629.5839.72
SLAVC [22]4.7713.0819.106.1221.1632.12
Alignment [31]16.4736.9949.0020.0942.3853.66
\\( Alignment^† \\)14.3137.8149.1718.0038.3949.02
Ours30.6157.8769.3731.7058.4369.81
", + "image_path": "1753d77bbaa50f35589e3f9547aa8fe2d30bc43dfc6fb39b3b3597262675ea27.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "sets of the VGG-SS, each containing 110 non-overlapping categories. During training, the model is trained only to the category set of one group. The subset comprising categories seen during training is called Heard 110, while the subset with unseen categories is Unheard 110. As shown in Table 2, our model achieves the state-of-the-art performance on Unheard 110 both without additional refinement" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 330, + 71, + 541, + 178 + ], + "blocks": [ + { + "bbox": [ + 55, + 560, + 295, + 605 + ], + "lines": [ + { + "bbox": [ + 55, + 560, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 295, + 605 + ], + "type": "text", + "content": "Table 4. Cross-modal retrieval on the VGG-SS. All of the models are trained on VGGSound-144K. Note that the results of previous research are obtained from [31]. " + }, + { + "bbox": [ + 55, + 560, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 560, + 295, + 605 + ], + "type": "text", + "content": " denotes that the self-supervised pretrained encoder is utilized for their work." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 71, + 541, + 178 + ], + "lines": [ + { + "bbox": [ + 330, + 71, + 541, + 178 + ], + "spans": [ + { + "bbox": [ + 330, + 71, + 541, + 178 + ], + "type": "table", + "html": "
Trainset\\( {\\mathcal{L}}_{\\text{match }} \\)\\( {\\mathcal{L}}_{\\text{div }} \\)FN MitigateVGG-SS
cIoUAUC
VGGSound 144k✓14.9527.02
✓36.2339.13
18.6930.03
✓15.0328.18
✓✓40.5641.35
✓✓36.1439.31
✓✓20.3730.81
✓✓✓40.7141.62
", + "image_path": "840127ecbe826f612369e557772913daf03138f221b29631274c8ef3b01592c1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 209, + 552, + 280 + ], + "blocks": [ + { + "bbox": [ + 353, + 186, + 515, + 197 + ], + "lines": [ + { + "bbox": [ + 353, + 186, + 515, + 197 + ], + "spans": [ + { + "bbox": [ + 353, + 186, + 515, + 197 + ], + "type": "text", + "content": "Table 5. Detailed ablation study for losses." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 209, + 552, + 280 + ], + "lines": [ + { + "bbox": [ + 315, + 209, + 552, + 280 + ], + "spans": [ + { + "bbox": [ + 315, + 209, + 552, + 280 + ], + "type": "table", + "html": "
TrainsetShared Initial SlotsVGG-SSFlickr
cIoUAUCcIoUAUC
Flickr-144k✓34.8038.7284.0061.70
36.6439.2486.8065.44
VGGSound-144k✓33.1137.9281.2064.36
40.7141.6284.4064.50
", + "image_path": "87249bb07f4072655853b5ddb81ff36c178389bf68b53f7fcc7fe00868a5f7c0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 345, + 288, + 523, + 299 + ], + "lines": [ + { + "bbox": [ + 345, + 288, + 523, + 299 + ], + "spans": [ + { + "bbox": [ + 345, + 288, + 523, + 299 + ], + "type": "text", + "content": "Table 6. Ablation study on shared initial slots." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 324, + 311, + 545, + 497 + ], + "blocks": [ + { + "bbox": [ + 324, + 311, + 545, + 497 + ], + "lines": [ + { + "bbox": [ + 324, + 311, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 324, + 311, + 545, + 497 + ], + "type": "image", + "image_path": "d8ea648b19ac68aad20d33c35672dc060174bb30465a0e4af7489e6b0b5a6ea3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 502, + 555, + 548 + ], + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 548 + ], + "type": "text", + "content": "Figure 4. Qualitative results to show the impact of " + }, + { + "bbox": [ + 313, + 502, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 313, + 502, + 555, + 548 + ], + "type": "text", + "content": " on Flickr-SoundNet. (a) Input image. (b) Ground-Truth. (c) Results without cross-modal attention matching. (d) Results with cross-modal attention matching." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 569, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 555, + 604 + ], + "type": "text", + "content": "and with OGL applied. These results demonstrate that our model has superior generalization ability compared to previous models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 689 + ], + "type": "text", + "content": "AVSBench. To more precisely evaluate the performance of the localization mask, we measured the performance on AVSBench, which provides segmentation masks. Our model was trained on VGGSound-144k and evaluated in a zero-shot manner on AVSBench. As shown in Table 3, we achieved the state-of-the-art results, specifically with improvements of " + }, + { + "bbox": [ + 313, + 605, + 555, + 689 + ], + "type": "inline_equation", + "content": "2.03\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 689 + ], + "type": "text", + "content": " p in mIoU and " + }, + { + "bbox": [ + 313, + 605, + 555, + 689 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 689 + ], + "type": "text", + "content": " p in F-score." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Cross-modal Retrieval. As discussed in the previous study [31], the benchmark evaluation metrics, cIoU and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3127" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 70, + 529, + 214 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 529, + 214 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 529, + 214 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 529, + 214 + ], + "type": "image", + "image_path": "0aea5b30ff7ce9d68e20812e204c773445d5e523bd4ef1ce6620ec75f50564a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 221, + 430, + 232 + ], + "lines": [ + { + "bbox": [ + 179, + 221, + 430, + 232 + ], + "spans": [ + { + "bbox": [ + 179, + 221, + 430, + 232 + ], + "type": "text", + "content": "Figure 5. Qualitative results of cross-modal retrieval on VGG-SS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 243, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 422 + ], + "type": "text", + "content": "AUC, do not guarantee better audio-visual semantic understanding. Therefore, we also evaluate cross-modal retrieval performance on VGG-SS. We extract the image and audio target slot representations and compute the cosine similarity between them. The similarities are ranked in descending order, and we calculate Recall@K performance using the category labels from the top-K retrieved audio or images. In Table 4, our model demonstrates a significant performance improvement over the previous state of the art in cross-modal retrieval. Specifically, our model improves Recall@1 by " + }, + { + "bbox": [ + 55, + 243, + 296, + 422 + ], + "type": "inline_equation", + "content": "16.3\\%" + }, + { + "bbox": [ + 55, + 243, + 296, + 422 + ], + "type": "text", + "content": " p in audio-to-image retrieval and " + }, + { + "bbox": [ + 55, + 243, + 296, + 422 + ], + "type": "inline_equation", + "content": "13.7\\%" + }, + { + "bbox": [ + 55, + 243, + 296, + 422 + ], + "type": "text", + "content": " p in image-to-audio retrieval. This result indicates that our model does not merely detect salient objects likely to produce sound within the image, but rather effectively locates the target specified by the audio input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 430, + 149, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 149, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 149, + 442 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 448, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 579 + ], + "type": "text", + "content": "Impact of Losses and False Negative Mitigating in Sec. 3.4. In Table 5, we conduct an ablation study to verify the effectiveness of each loss function and the strategy outlined in Sec. 3.4. The overall results show that the designed losses and the false negative mitigating strategy each contribute to improve the performance. Notably, the experiment with " + }, + { + "bbox": [ + 55, + 448, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 55, + 448, + 295, + 579 + ], + "type": "text", + "content": " demonstrates a drastic increase in performance, highlighting the importance of cross-modality attention matching. Additionally, the ablation study on " + }, + { + "bbox": [ + 55, + 448, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{div}}" + }, + { + "bbox": [ + 55, + 448, + 295, + 579 + ], + "type": "text", + "content": " shows that explicitly separating the slots enables each slot to perform its designated role more effectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 580, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 662 + ], + "type": "text", + "content": "Impact of Shared Initial Slots. We conduct an ablation study to examine the impact of shared initial slots by using different initial slots for each modality. As shown in Table 6, sharing initial slots resulted in better performance, indicating that this design choice facilitates the ability to jointly capture sound source-related information across different modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 167, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 167, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 167, + 683 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Quality of SSL results. Fig. 3 visualizes the qualitative results of our model and previous studies on Flickr-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 243, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 314 + ], + "type": "text", + "content": "SoundNet [4] and VGG-SS [5] test sets. These results demonstrate that our model has the ability to effectively localize sound source objects. Notably, it produces compact localization results even for challenging samples that other models struggle with, such as small birds, a gun, and cars surrounded by other objects in street scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 315, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 410 + ], + "type": "text", + "content": "Qualitative Analysis of " + }, + { + "bbox": [ + 313, + 315, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 313, + 315, + 555, + 410 + ], + "type": "text", + "content": ". Fig. 4 shows the qualitative result that shows the effect of " + }, + { + "bbox": [ + 313, + 315, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 313, + 315, + 555, + 410 + ], + "type": "text", + "content": " on Flickr-SoundNet-Test. The attention maps is generated between the audio query of the audio target slot and the image keys derived from image features. Without " + }, + { + "bbox": [ + 313, + 315, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 313, + 315, + 555, + 410 + ], + "type": "text", + "content": ", the alignment between the two is insufficient for compact localization, whereas with " + }, + { + "bbox": [ + 313, + 315, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 313, + 315, + 555, + 410 + ], + "type": "text", + "content": ", the alignment improves, resulting in more accurate localization of the sound source." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 410, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 482 + ], + "type": "text", + "content": "Cross-modal Retrieval. Fig. 5 shows the qualitative results of our model for cross-modal retrieval. Consistent with the impressive performance shown in Table 4, our model demonstrates a strong ability to retrieve semantically similar samples across modalities that correspond to the sound source target." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 492, + 388, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 492, + 388, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 492, + 388, + 504 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 513, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 643 + ], + "type": "text", + "content": "This paper presented a novel self-supervised sound source localization method leveraging joint slot attention to align image and audio features. Our approach decomposes these features into target and off-target slots, using only target slots for contrastive learning to capture sound-related features. In addition, we introduced cross-modal attention matching and slot divergence loss which significantly improves localization accuracy and and we demonstrated the effectiveness of our model through comprehensive analysis. For future work, we aim to extend our framework to multisource sound source localization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 644, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 555, + 703 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the NRF grant (RS-2021-NR059830-40%) and the IITP grants (RS-2024-00509258-40%, RS-2021-II212068-10%, RS-2019-II191906-10%) funded by Ministry of Science and ICT, Korea." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3128" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In Proc. European Conference on Computer Vision (ECCV), 2020. 1, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] Stuart Andrews, Ioannis Tsochantaridis, and Thomas Hofmann. Support vector machines for multiple-instance learning. Advances in neural information processing systems, 15, 2002. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 213 + ], + "type": "text", + "content": "[3] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proc. European Conference on Computer Vision (ECCV), pages 435-451, 2018. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 294, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 256 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 256 + ], + "type": "text", + "content": "[4] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Sound-net: Learning sound representations from unlabeled video. Proc. Neural Information Processing Systems (NeurIPS), 29, 2016. 2, 5, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "text", + "content": "[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2, 5, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 294, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 294, + 367 + ], + "type": "text", + "content": "[6] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 456 + ], + "type": "text", + "content": "[7] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724-1734, Doha, Qatar, 2014. Association for Computational Linguistics. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 294, + 501 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: a large-scale hierarchical image database. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 556 + ], + "type": "text", + "content": "[9] Dennis Fedorishin, Deen Dayal Mohan, Bhavin Jawade, Srirangaraj Setlur, and Venu Govindaraju. Hear the flow: Optical flow-based self-supervised visual sound source localization. In Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV), pages 2278-2287, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 294, + 601 + ], + "type": "text", + "content": "[10] Yury Furletov, Volker Willert, and Jürgen Adamy. Auditory scene understanding for autonomous driving. In 2021 IEEE Intelligent Vehicles Symposium (IV), pages 697-702. IEEE, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[12] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[13] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In Proc. In-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "ternational Conference on Machine Learning (ICML), pages 2127-2136. PMLR, 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 98, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 98, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 98, + 553, + 151 + ], + "type": "text", + "content": "[14] Dongjin Kim, Sung Jin Um, Sangmin Lee, and Jung Uk Kim. Learning to visually localize sound sources from mixtures without prior source knowledge. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 26467-26476, 2024. 1, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 153, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 208 + ], + "type": "text", + "content": "[15] Chen Liu, Peike Patrick Li, Xingqun Qi, Hu Zhang, Lincheng Li, Dadong Wang, and Xin Yu. Audio-visual segmentation by exploring cross-modal mutual semantics. In Proc. ACM Multimedia Conference (ACMMM), pages 7590–7598, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 210, + 553, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 210, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 317, + 210, + 553, + 254 + ], + "type": "text", + "content": "[16] Jinxiang Liu, Chen Ju, Weidi Xie, and Ya Zhang. Exploiting transformation invariance and equivariance for self-supervised sound localisation. In Proc. ACM Multimedia Conference (ACMMM), pages 3742-3753, 2022. 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 256, + 553, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 256, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 317, + 256, + 553, + 320 + ], + "type": "text", + "content": "[17] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Proc. Neural Information Processing Systems (NeurIPS), 33:11525-11538, 2020. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 323, + 553, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 323, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 317, + 323, + 553, + 356 + ], + "type": "text", + "content": "[18] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. International Conference on Learning Representations (ICLR), 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 358, + 553, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 358, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 317, + 358, + 553, + 402 + ], + "type": "text", + "content": "[19] Tanvir Mahmud, Yapeng Tian, and Diana Marculescu. T-vsl: Text-guided visual sound source localization in mixtures. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 26742-26751, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 404, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 437 + ], + "type": "text", + "content": "[20] Oded Maron and Tomás Lozano-Pérez. A framework for multiple-instance learning. Advances in neural information processing systems, 10, 1997. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 439, + 553, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 439, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 317, + 439, + 553, + 471 + ], + "type": "text", + "content": "[21] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In Proc. European Conference on Computer Vision (ECCV), 2022. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 474, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 474, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 317, + 474, + 553, + 506 + ], + "type": "text", + "content": "[22] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. In Proc. Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 509, + 553, + 552 + ], + "type": "text", + "content": "[23] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10565–10574, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 554, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 554, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 317, + 554, + 553, + 586 + ], + "type": "text", + "content": "[24] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 589, + 553, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 589, + 553, + 631 + ], + "spans": [ + { + "bbox": [ + 317, + 589, + 553, + 631 + ], + "type": "text", + "content": "[25] Sooyoung Park, Arda Senocak, and Joon Son Chung. Can clip help sound source localization? In Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV), 2024. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 635, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 677 + ], + "type": "text", + "content": "[26] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In Proc. European Conference on Computer Vision (ECCV), 2020. 1, 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 680, + 553, + 713 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3129" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 586 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[28] Yoko Sasaki, Ryo Tanabe, and Hiroshi Takernura. Online spatial sound perception using microphone array on mobile robot. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2478-2484. IEEE, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "text", + "content": "[29] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "type": "text", + "content": "[30] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Learning sound localization better from semantically similar samples. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 306 + ], + "type": "text", + "content": "[31] Arda Senocak, Hyeonggon Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "type": "text", + "content": "[32] Zengjie Song, Yuxi Wang, Junsong Fan, Tieniu Tan, and Zhaoxiang Zhang. Self-supervised predictive learning: A negative-free method for sound source localization in visual scenes. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3222-3231, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 430 + ], + "type": "text", + "content": "[33] Weixuan Sun, Jiayi Zhang, Jianyuan Wang, Zheyuan Liu, Yiran Zhong, Tianpeng Feng, Yandong Guo, Yanhao Zhang, and Nick Barnes. Learning audio-visual source localization via false negative aware contrastive learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 295, + 485 + ], + "type": "text", + "content": "[34] Jun Yin, Stefano Damiano, Marian Verhelst, Toon van Watershoot, and Andre Guntoro. Real-time acoustic perception for automotive applications. In 2023 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1-6. IEEE, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 295, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 295, + 529 + ], + "type": "text", + "content": "[35] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep features for discriminative localization. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 532, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 295, + 586 + ], + "type": "text", + "content": "[36] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In Proc. European Conference on Computer Vision (ECCV), pages 386–403, 2022. 2, 5" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3130" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/5356bb72-29a6-442b-984c-adce3c009504_content_list.json b/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/5356bb72-29a6-442b-984c-adce3c009504_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a5c08e13cb8a79c68387f5825f4372de9e3100c7 --- /dev/null +++ b/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/5356bb72-29a6-442b-984c-adce3c009504_content_list.json @@ -0,0 +1,1605 @@ +[ + { + "type": "text", + "text": "Improving Transferable Targeted Attacks with Feature Tuning Mixup", + "text_level": 1, + "bbox": [ + 145, + 130, + 851, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kaisheng Liang Xuelong Dai Yanjie Li Dong Wang Bin Xiao* Department of Computing, The Hong Kong Polytechnic University", + "bbox": [ + 176, + 180, + 826, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cskliang@comp.polyu.edu.hk b.xiao@polyu.edu.hk \nelong.dai, yanjie.li, dong-comp.wang}@connect.polyu.hk", + "bbox": [ + 276, + 219, + 750, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) exhibit vulnerability to adversarial examples that can transfer across different DNN models. A particularly challenging problem is developing transferable targeted attacks that can mislead DNN models into predicting specific target classes. While various methods have been proposed to enhance attack transferability, they often incur substantial computational costs while yielding limited improvements. Recent clean feature mixup methods use random clean features to perturb the feature space but lack optimization for disrupting adversarial examples, overlooking the advantages of attack-specific perturbations. In this paper, we propose Feature Tuning Mixup (FTM), a novel method that enhances targeted attack transferability by combining both random and optimized noises in the feature space. FTM introduces learnable feature perturbations and employs an efficient stochastic update strategy for optimization. These learnable perturbations facilitate the generation of more robust adversarial examples with improved transferability. We further demonstrate that attack performance can be enhanced through an ensemble of multiple FTM-perturbed surrogate models. Extensive experiments on the ImageNet-compatible dataset across various DNN models demonstrate that our method achieves significant improvements over state-of-the-art methods while maintaining low computational cost.", + "bbox": [ + 88, + 319, + 483, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 729, + 220, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) are highly vulnerable to adversarial attacks [13, 23, 31], raising critical security concerns. In image classification tasks, adversarial attacks can deceive DNNs by adding imperceptible perturbations to benign images, causing the models to produce erroneous predictions. More concerningly, research [9, 10, 46] has demonstrated that adversarial examples possess trans", + "bbox": [ + 89, + 755, + 482, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b5de2d0290ae04e5720290995e182080e231572a745fdf24abe67249a955ba9f.jpg", + "image_caption": [ + "Figure 1. Effectiveness and efficiency of targeted attacks. Average attack success rates on 14 black-box models, along with the computation time required to generate an adversarial example. Our methods demonstrate superior performance with low computational cost, surpassing state-of-the-art methods." + ], + "image_footnote": [], + "bbox": [ + 526, + 290, + 880, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ferability properties: adversarial examples generated using one white-box model can successfully fool other black-box models without requiring access to their architectures or parameters. This cross-model transferability of adversarial examples poses severe security threats to real-world AI systems. Therefore, studying the transferability of adversarial examples is crucial for enhancing the robustness of DNNs against such security vulnerabilities.", + "bbox": [ + 511, + 547, + 906, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Targeted transfer-based attacks aim to mislead models into predicting specific target classes, which remains a significant challenge. Due to the lack of knowledge about the target model's training data and architecture, such attacks tend to generate targeted adversarial examples that overfit to surrogate models and fail to transfer to target models. Various transferable targeted attacks [2, 43, 49] have been proposed but still leave much room for improvement. Notably, data augmentation methods [29, 39, 40] are commonly employed for targeted attacks. But they primarily operate in the image space, while augmentation methods at the feature level remain relatively unexplored. Recently, the Clean Feature Mixup (CFM) method [3] has made important progress by using clean features from random images to disrupt adversarial examples, leading to better targeted", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 107, + 875, + 235, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Our code is available at https://github.com/uhiu/features-tuning-mixup", + "bbox": [ + 109, + 887, + 480, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25802", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/75bd2a809054658927670f6e784cc04f6e94caad810c2a60a151aaebf338570e.jpg", + "image_caption": [ + "Gradients from the Backward Process in Iteration i of the Attack", + "Figure 2. Overview of our FTM attack. In the forward pass, the learnable perturbations are added to the output features. Only a small portion of network layers are randomly selected for perturbation updates and clean feature mixup in each attack iteration. In the backward pass, the gradients of the selected perturbations and adversarial image are computed together for update." + ], + "image_footnote": [], + "bbox": [ + 132, + 89, + 867, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "attack success rates (Figure 1). However, CFM uses random clean features without optimizing them specifically for disrupting adversarial examples. We hypothesize that carefully optimized perturbations would be more effective in disrupting adversarial examples, potentially yielding substantial improvements in attack transferability.", + "bbox": [ + 88, + 395, + 480, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose Feature Tuning Mixup (FTM), a novel attack method that optimizes attack-specific feature perturbations to enhance the transferability of targeted adversarial examples. Our method fine-tunes the intermediate features of surrogate models during the attack process by introducing learnable feature perturbations. These feature perturbations are optimized by mixing them with clean features from random images to counteract the adversarial effects of targeted adversarial examples. Through this mechanism, adversarial examples are gradually compelled to become more robust, allowing them to overcome these feature-level disturbances. To address the computational overhead of optimizing feature perturbations, we propose a stochastic update strategy. This strategy initializes the learnable perturbations using values from previous attack iterations and selectively optimizes only a small subset of feature perturbations in each iteration. Ultimately, FTM significantly improves the transferability of targeted adversarial examples while maintaining computational efficiency.", + "bbox": [ + 91, + 489, + 482, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We provide an overview of the proposed FTM in Figure 2. FTM first introduces learnable perturbations into the deep layers of the surrogate model. The learnable perturbations are added element-wise to the original output features before passing to the next layer. In each attack iteration, a small subset of layers is randomly selected and combined with random clean features. Finally, we compute gradients for the learnable perturbations and adversarial im", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "age using only one forward and backward pass, similar to standard attacks. We update the selected learnable perturbations and adversarial image, while unselected perturbations remain unchanged. As a result, FTM fully leverages optimized attack-specific feature perturbations, prevents overfitting to surrogate models, and significantly enhances the transferability of targeted attacks.", + "bbox": [ + 511, + 395, + 905, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 531, + 503, + 834, + 516 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We revisit feature-level augmentation methods for transferable targeted attacks and identify that incorporating attack-specific feature perturbations, previously overlooked, can efficiently boost existing attacks.", + "- We introduce Feature Tuning Mixup (FTM), a novel approach that combines random and optimized feature perturbations to enhance targeted transfer-based attacks. Our method includes an efficient stochastic update strategy that maintains computational efficiency while improving attack transferability.", + "- We conduct extensive experiments across various source and target DNN models on the ImageNet-compatible dataset, demonstrating that our FTM achieves a significant improvement over state-of-the-art methods." + ], + "bbox": [ + 513, + 523, + 903, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 752, + 653, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial attacks [4, 12, 13, 19, 31] have been widely studied in the field of adversarial learning, including unrestricted attacks [7], physical attacks [26, 27], and black-box attacks [1, 9, 28]. Notably, black-box attacks can be categorized into query-based [1] and transfer-based attacks [9, 28]. The latter, which does not require access to the target DNN model's outputs, parameters or architecture, is more practical and is the focus of this paper.", + "bbox": [ + 511, + 779, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25803", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Non-targeted Transfer-based Attacks", + "text_level": 1, + "bbox": [ + 89, + 90, + 413, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Current non-targeted transfer attacks are primarily based on the Iterative-Fast Gradient Sign Method (I-FGSM) [23]. This approach iteratively computes the gradient of the classification loss with respect to the input and updates adversarial examples in the direction that maximizes the loss. However, it often leads to overfitting on white-box models, making the adversarial examples less effective against other black-box models. Numerous studies have explored solutions to the overfitting issue to enhance adversarial transferability. The Momentum Iterative FGSM (MIFGSM) [9] incorporates a momentum-based optimization method during iterative attacks. Translation-Invariant (TI-FGSM) method [10] enhances the transferability of adversarial examples by achieving translation invariance through gradient smoothing. Data augmentation techniques are widely used to mitigate overfitting. The Diverse Inputs (DI) method [46] employs random resizing and padding with a probability $p$ to generate diverse input, while the Resized Diverse Inputs (RDI) method [50] fixes $p = 1$ and rescales DI-enhanced images back to their original size. The Scale-Invariant (SI) method [29] generates adversarial perturbations using multiple scale transformations, and Admix [40] extends SI by blending images from other labels into the input. Variance Tuning (VT) method [39] randomly selects data points around a given data point and computes their gradients to avoid overfitting. These methods operate in the image space. Feature-level approaches have also been studied to improve adversarial transferability. For instance, the Feature Importance-aware Attack (FIA) [41] identifies and perturbs important features in the intermediate layers of a surrogate model. The Skip Gradient Method (SGM) [44] utilizes more gradients from skip connections. Ensemble-based attacks [9, 30] utilize multiple surrogate models to generate adversarial examples.", + "bbox": [ + 89, + 112, + 485, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Targeted Transfer-based Attacks", + "text_level": 1, + "bbox": [ + 89, + 636, + 380, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Targeted attacks are more challenging than non-targeted attacks because they not only aim to cause incorrect predictions by the models but also to mislead the models into predicting a specific target class. We can categorize targeted attacks into two types based on whether they require a training dataset. Previous works have utilized training datasets to generate auxiliary modules [21, 22] for transfer-based attacks or directly train a model to generate adversarial examples [32, 42, 47]. This paper primarily focuses on transfer-based attacks that do not require a training dataset.", + "bbox": [ + 89, + 657, + 482, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Targeted transfer-based attacks are widely studied in adversarial machine learning due to their practicality, as they do not require access to the target model's architecture and parameters, nor do they need training datasets. The Activation Attack (AA) [20] proposes using the distance between adversarial examples and targeted images in feature", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "space as a loss function, but its attack effectiveness may depend on the selection of targeted images. Po+Trip [24] employs Poincare distance to adapt the gradients' magnitude and uses triplet loss to address insufficient inter-class distances, but it only works well on target models with similar architectures. Logit Attack (Logit) [49] discovers that a simple logit loss can generate better transferable adversarial examples than CE loss, significantly outperforming previous methods. Self-Universality (SU) method [43] finds that highly universal adversarial perturbations tend to be more transferable for targeted attacks and proposes optimizing the perturbations in different local regions of the image. The Object-based Diverse Input (ODI) method [2] is an efficient data augmentation in image space for targeted attacks using random 3D object and rendering.", + "bbox": [ + 511, + 90, + 903, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, the Clean Feature Mixup (CFM) method [3] proposes a feature-level augmentation which achieves state-of-the-art performance in targeted transfer-based attacks. It stores clean features as random feature perturbations, which mitigates the overfitting to surrogate models. However, the proposed clean features in CFM lack specific optimization for targeted attacks, leaving significant room for improvement. We develop the Feature Tuning Mixup (FTM) method, which effectively and efficiently optimizes attack-specific feature perturbations and significantly improves transferable targeted attacks.", + "bbox": [ + 511, + 320, + 906, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Differences from intermediate-level attacks. This paper investigates feature-level augmentation for targeted attacks, which differs from existing intermediate-level attacks [18, 25, 41, 48]. Our FTM is based on the transfer-based attack framework [23] with logit loss [49], which is currently the most effective framework for targeted transfer-based attacks. In contrast, intermediate-level attacks use loss functions based on a specific intermediate layer, which is hard to apply to targeted attacks. For instance, FIA [41] and NAA [48] require identifying important features based on the true label, while targeted attacks focus on the target label instead of the true label. ILA [18] requires an existing adversarial example as guidance and cannot function as a standalone targeted attack. Therefore, we do not compare FTM with intermediate-level attacks in experiments.", + "bbox": [ + 511, + 489, + 908, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Comparison with adversarial learning methods. Our method learns feature perturbations in an adversarial manner without requiring additional training datasets. While some existing attacks [33, 45] also employ adversarial strategies, they operate in the image space and typically require access to training datasets. For instance, ATTA [45] improves transferability by training an adversarial transformation network to destroy adversarial examples. However, this approach demands substantial computational resources and training data. Like previous work [2, 3, 43, 49], we study targeted transfer attacks without extra data and do not compare with methods that need training data. RAP [33]", + "bbox": [ + 511, + 719, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25804", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "utilizes adversarial learning to identify worst-case perturbations in the image space, aiming to locate regions with flat loss landscapes. However, this process requires numerous additional backpropagation steps, leading to prohibitively high computational costs. In contrast, we propose a novel approach that optimizes feature perturbations directly in the multi-layer feature space, an unexplored direction in previous research. Through this innovation, our method achieves superior practicality, effectiveness, and efficiency compared to existing methods.", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 253, + 225, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 279, + 230, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem formulation. For targeted transfer-based attacks, the attacker is assumed to have a white-box surrogate model $F$ , which is a pre-trained image classifier in this paper. We assume that the attacker has no additional training dataset. Given a benign image $x$ , a surrogate model $F$ and a target label $y_{t}$ , which is different from the true label $y$ of $x$ . The goal of the attack is to generate an adversarial example $x^{\\mathrm{adv}}$ that is misclassified as $y_{t}$ by $F$ . The difference between $x$ and $x^{\\mathrm{adv}}$ should be subtle, and following previous work, we use the $\\ell_{\\infty}$ -norm constraint. The optimization objective can be formulated as:", + "bbox": [ + 89, + 301, + 483, + 467 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {x ^ {\\mathrm {a d v}}} \\mathcal {L} (F (x ^ {\\mathrm {a d v}}), y _ {t}), \\quad s. t. \\quad \\| x - x ^ {\\mathrm {a d v}} \\| _ {\\infty} \\leq \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 476, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ is the adversarial loss, e.g., the cross-entropy loss or logit-based loss, and $\\epsilon$ is the perturbation budget.", + "bbox": [ + 89, + 507, + 482, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MI-FGSM. To solve the problem above, existing methods usually adopt the iterative gradient-based technique, e.g., the MI-FGSM [9], formulated as follows:", + "bbox": [ + 89, + 537, + 483, + 583 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i + 1} = \\nabla_ {x} \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}\\right), y _ {t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 592, + 482, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {g} _ {i + 1} = \\mu \\cdot g _ {i} + g _ {i + 1} / \\| g _ {i + 1} \\| _ {1}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 612, + 482, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i + 1} ^ {\\mathrm {a d v}} = x _ {i} ^ {\\mathrm {a d v}} - \\eta \\cdot \\operatorname {s i g n} \\left(\\tilde {g} _ {i + 1}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 631, + 482, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i + 1} ^ {\\mathrm {a d v}} = \\operatorname {C l i p} _ {x, \\epsilon} \\left(x _ {i + 1} ^ {\\mathrm {a d v}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 652, + 482, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g_{i}$ is the gradient of the loss function with respect to the input image $x$ in the $i$ -th iteration. $\\mu$ is the momentum decay factor, and $\\eta$ is the step size. $x_{i+1}^{\\mathrm{adv}}$ is clipped within $\\epsilon$ bound. Initially, we set $x_{0}^{\\mathrm{adv}} = x$ and $g_{0} = 0$ .", + "bbox": [ + 89, + 679, + 482, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CFM. To improve the transferability of targeted attacks, the Clean Feature Mixup (CFM) method [3] leverages stored clean features to augment adversarial examples, achieving state-of-the-art performance among existing targeted transfer-based attacks. Specifically, it selects outputs from certain network layers as perturbation targets and uses output features from different benign images as feature perturbations. To integrate with MI-FGSM, the feature augmentation in CFM can be formulated as follows:", + "bbox": [ + 89, + 739, + 482, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {k, i} ^ {\\prime} = \\left(1 - \\alpha_ {k, i}\\right) \\odot z _ {k, i} + \\alpha_ {k, i} \\odot z _ {k, i} ^ {c}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise multiplication, $z_{k,i}$ is the input feature to the $k$ -th layer in the $i$ -th attack iteration of MI-FGSM, $z_{k,i}^{c}$ is a random stored clean feature, and $\\alpha_{k,i}$ is a vector representing channel-wise mixing ratios. The network layer $k$ is randomly selected from pre-defined candidate layers with probability $p$ .", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Analysis of CFM", + "text_level": 1, + "bbox": [ + 511, + 189, + 679, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CFM is a powerful feature-level augmentation method for targeted attacks. To understand the effectiveness of CFM, we formulate a generalized optimization step for MI-FGSM based on feature-level augmentation as follows:", + "bbox": [ + 511, + 210, + 905, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i + 1} ^ {\\prime} = \\nabla_ {x} \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}; \\Delta z _ {i}\\right), y _ {t}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 280, + 903, + 297 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta z _ {i} = \\left\\{\\Delta z _ {k, i}, k = 1, \\dots , n \\right\\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 300, + 903, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F(x_{i}^{\\mathrm{adv}};\\Delta z_{i})$ denotes the surrogate model $F$ with the perturbation $\\Delta z_{i}$ added to the network outputs in the $i$ -th attack iteration. Equation (7) improves upon Equation (2) by incorporating feature-level perturbations, which is the key idea of CFM. Notably, the perturbation $\\Delta z_{i}$ represents the collection of perturbations for all network layers. We use $\\Delta z_{k,i}$ to denote the perturbation for the input of the $k$ -th layer, where $n$ is the maximum layer index of $F$ . Initially, we set $\\Delta z_{i} = 0$ . For any layer $d$ that does not require perturbation, we simply set $\\Delta z_{d,i} \\equiv 0$ .", + "bbox": [ + 511, + 325, + 905, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CFM uses clean features to construct $\\Delta z_{i}$ , and it claims that the rationale is that clean features can suppress the adversarial effect of $x_{i}^{\\mathrm{adv}}$ [3]. We formulate this idea in the following equation:", + "bbox": [ + 511, + 477, + 905, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}; \\Delta z _ {i}\\right), y _ {t}\\right) > \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}\\right), y _ {t}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 544, + 903, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta z_{i}$ is supposed to appropriately disrupt adversarial loss in Equation (1) for obtaining robust $x^{\\mathrm{adv}}$ .", + "bbox": [ + 511, + 571, + 905, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, clean features are attack-agnostic in nature, as they are not explicitly optimized for the objective in Equation (9). Although CFM demonstrates promising improvements, we argue that there remains unexplored potential in feature-level augmentation for targeted attacks. We believe that integrating attack-specific optimization with attack-agnostic clean features can create a more powerful feature-level augmentation strategy.", + "bbox": [ + 511, + 602, + 905, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Feature Tuning Mixup (FTM)", + "text_level": 1, + "bbox": [ + 511, + 729, + 779, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose Feature Tuning Mixup (FTM), a novel method that uses attack-specific feature perturbations to enhance feature-level augmentation methods. We formulate the optimization problem for learning our feature perturbations and adversarial examples in each attack iteration as follows:", + "bbox": [ + 511, + 752, + 905, + 828 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {x _ {i} ^ {\\mathrm {a d v}}} \\max _ {\\Delta z _ {i}} \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}; \\Delta z _ {i}\\right), y _ {t}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 835, + 903, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the optimization problem above is challenging to solve from two perspectives: effectiveness and efficiency.", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25805", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regarding efficiency, traditional adversarial training [31] often requires multiple steps to optimize $\\Delta z_{i}$ in each attack iteration, resulting in substantial additional computational cost. As for effectiveness, since $\\Delta z_{i}$ involves a joint problem across multiple iterations and network layers, inappropriate perturbations can damage the attack effect, which will be shown in our experiments.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Momentum-based stochastic update. We propose a novel technique, momentum-based stochastic update, to optimize $\\Delta z_{i}$ . We employ one-step optimization with using perturbations in previous iterations as momentum, i.e., $\\Delta z_{i}$ will be initialized as $\\Delta z_{i-1}$ in Equation (10). Momentum can effectively improve the optimization effectiveness without increasing additional optimization steps. In addition, as we only require one-step optimization, the gradient for $\\Delta z_{i}$ can be jointly computed with the gradient for $x_{i}^{\\mathrm{adv}}$ , which can further speed up the optimization process. To further improve the optimization efficiency, we use a stochastic update strategy, where only the randomly selected layers are updated in each attack iteration. For each selected layer $k$ , it has probability $p$ to be updated in the $i$ -th attack iteration, otherwise, we set $\\Delta z_{k,i} = \\Delta z_{k,i-1}$ . We illustrate the detailed optimization process for FTM in the following.", + "bbox": [ + 89, + 198, + 482, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The forward process for FTM in the $i$ -th attack iteration is formulated as follows:", + "bbox": [ + 89, + 441, + 482, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {z} _ {k, i} = z _ {k, i} + \\beta \\| z _ {k, i} \\| \\cdot \\frac {\\Delta z _ {k , i}}{\\| \\Delta z _ {k , i} \\| + \\bar {\\epsilon}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 481, + 482, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {k, i} ^ {\\prime} = \\left\\{ \\begin{array}{l l} (1 - \\alpha_ {k, i}) \\odot \\bar {z} _ {k, i} + \\alpha_ {k, i} \\odot z _ {k, i} ^ {c}, & \\tau_ {k} < p, \\\\ \\bar {z} _ {k, i}, & \\text {e l s e ,} \\end{array} \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 517, + 480, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta \\| z_{k,i}\\|$ is the custom perturbation budget for our feature perturbation $\\Delta z_{k,i}$ . $\\beta$ is a scaling factor, and $\\bar{\\epsilon}$ is a small constant. We mix the output feature $z_{k,i}$ and our attack-specific perturbation $\\Delta z_{k,i}$ to obtain $\\bar{z}_{k,i}$ . Then, we use clean features to augment $\\bar{z}_{k,i}$ to get $z_{k,i}^{\\prime}$ , which is the final perturbed feature that will be fed into the $k$ -th network layer. These perturbed features are computed layer by layer during the forward process. $z_{k,i}^{c}$ is a random clean feature, and $\\alpha_{k,i}$ is a vector representing channel-wise mixing ratio. We randomly sample $\\alpha_{k,i}$ from $\\mathcal{U}(0,\\alpha_{max})$ . In each iteration, we randomly sample $\\tau_{k}$ from $\\mathcal{U}(0,1)$ to determine whether to update the $k$ -th perturbation $\\Delta z_{k,i}$ during the $i$ -th attack iteration. $p$ is the probability threshold.", + "bbox": [ + 89, + 571, + 482, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The backward process for FTM in the $i$ -th attack iteration is formulated as follows:", + "bbox": [ + 89, + 768, + 482, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta z _ {i} ^ {\\prime} = \\left\\{\\left(\\Delta z _ {k, i}, \\tau_ {k}\\right) \\mid k = 1, \\dots , n \\right\\}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 813, + 482, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i + 1} ^ {\\prime}, g ^ {\\Delta z ^ {\\prime}} = \\nabla_ {x _ {i} ^ {\\mathrm {a d v}}, \\Delta z _ {i} ^ {\\prime}} \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}; \\Delta z _ {i} ^ {\\prime}\\right), y _ {t}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 834, + 480, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta z _ {k, i + 1} = \\left\\{ \\begin{array}{l l} \\Delta z _ {k, i} + g _ {k} ^ {\\Delta z ^ {\\prime}}, & \\tau_ {k} < p, \\\\ \\Delta z _ {k, i}, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 857, + 482, + 897 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Feature Tuning Mixup (FTM) Algorithm", + "bbox": [ + 514, + 90, + 875, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Surrogate model $F$ , clean image $x$ , target label $y_{t}$ , iteration number $T$ , perturbation budget $\\epsilon$ , input transformation $\\phi(\\cdot)$ (RDI, etc.), step size $\\eta$ , decay factor $\\mu$", + "bbox": [ + 513, + 109, + 903, + 155 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Mixing range $\\alpha_{max}$ , probability $p$ , mixing ratio $\\beta$ \nOutput: An adversarial example $x^{\\mathrm{adv}}$", + "bbox": [ + 514, + 156, + 893, + 185 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $x_0^{\\mathrm{adv}} = x, g_0 = 0$ .", + "2: Initialize perturbation $\\Delta z_0 = 0$", + "3: for $i = 0 \\rightarrow T - 1$ do:", + "4: Extend $F(x_{j}^{\\mathrm{adv}})$ to $F(\\phi (x_{j}^{\\mathrm{adv}}))$", + "5: Select layers to update and obtain $\\Delta z_{i}^{\\prime}$ from $\\Delta z_{i}$", + "6: Forward and backward pass to get $g_{i+1}'$ and $g^{\\Delta z'}$ using Equations (12) and (14)", + "7: Obtain $\\Delta z_{i + 1}$ using Equation (15)", + "8: Calculate $\\tilde{g}_{i + 1} = \\mu \\cdot g_{i} + g_{i + 1}^{\\prime} / \\| g_{i + 1}^{\\prime}\\|_{1}$", + "9: Update example $x_{i+1}^{\\mathrm{adv}} = x_i^{\\mathrm{adv}} - \\eta \\cdot \\operatorname{sign}(\\tilde{g}_{i+1})$", + "10: Clip $x_{i+1}^{\\mathrm{adv}}$ to be $\\ell_{\\infty}$ -bounded by $\\epsilon$", + "11: end for", + "12: $x^{\\mathrm{adv}} = x_T^{\\mathrm{adv}}$", + "13: return $x^{\\mathrm{adv}}$" + ], + "bbox": [ + 519, + 185, + 901, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta z_{i}^{\\prime}$ represents a set containing both perturbations $\\Delta z_{k,i}$ and their corresponding random variables $\\tau_{k}$ from the forward process, which determines whether each perturbation needs to be updated. $g_{k}^{\\Delta z^{\\prime}}$ denotes the gradient with respect to the $k$ -th perturbation $\\Delta z_{k,i}$ . The threshold $p$ remains consistent with that used in the forward process. We summarize our FTM attack in Algorithm 1.", + "bbox": [ + 511, + 428, + 903, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Efficiency analysis. FTM does not require complex operations or additional forward and backward passes. Therefore, FTM is theoretically efficient on both CNNs and ViTs. In contrast, SI, VT, Admix, and SU require additional forward and backward passes, while ODI requires a computationally expensive input transformation.", + "bbox": [ + 511, + 534, + 903, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ensemble of surrogate variants (FTM-E). Observing that FTM can efficiently perturb the surrogate model, we propose to further enhance the attack effectiveness through an ensemble strategy. We create multiple copies of the surrogate model and independently apply FTM to each copy, then ensemble them as a new surrogate model. We empirically show that utilizing an ensemble of two perturbed surrogate models can substantially enhance the attack performance while maintaining reasonable computational costs.", + "bbox": [ + 511, + 625, + 906, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 777, + 645, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 801, + 720, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Following prior research [2, 3, 43, 49], we conduct experiments on the ImageNet-compatible dataset from the NIPS 2017 adversarial attack challenge. This dataset comprises 1,000 images, each annotated with both true and target class labels for targeted attack evaluation.", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25806", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/554becbfc9d188fa7686b9254f0c70a477310905e7cecb6f467be3544860b77a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SourceAttackTime (s)VGG-16RN-18RN-50DN-121XcepMB-v2EF-B0IR-v2Inc-v3Inc-v4Avg.
RN-50DI1.3763.256.498.974.75.527.829.05.69.09.437.9
RDI1.2366.370.798.281.913.246.447.217.129.923.349.4
RDI-SI5.9570.579.898.888.929.556.266.237.956.443.662.8
RDI-VT7.1668.878.798.282.527.954.556.132.845.837.958.3
RDI-Admix3.5574.280.798.786.820.959.456.126.742.734.158.0
RDI-SU2.2167.974.798.383.324.956.350.220.932.126.253.5
ODI4.3878.377.197.687.043.867.370.049.565.955.469.2
RDI-CFM1.3984.788.498.490.351.181.578.848.065.559.374.6
RDI-FTM1.5486.387.597.989.756.183.381.254.770.866.677.4
RDI-FTM-E2.9288.188.698.391.459.485.484.356.973.469.379.5
Inc-v3DI2.012.92.43.45.01.91.83.73.099.24.212.8
RDI1.763.53.84.07.03.13.05.96.398.77.114.2
RDI-SI8.114.05.25.711.06.34.68.211.698.812.116.8
RDI-VT10.505.98.99.413.27.45.99.812.398.714.718.6
RDI-Admix4.926.36.58.812.86.06.110.912.298.713.618.2
RDI-SU2.367.811.011.412.79.48.26.38.398.710.918.5
ODI6.4214.314.916.732.320.313.725.326.495.931.629.1
RDI-CFM2.1322.926.826.239.134.127.138.636.295.944.839.2
RDI-FTM2.3525.830.932.141.736.330.741.442.994.148.142.4
RDI-FTM-E4.3734.239.340.654.447.037.151.652.496.659.251.2
", + "bbox": [ + 107, + 88, + 890, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Targeted attack success rates (%) against ten target models. All methods are combined with MI-TI. The best results are shown in bold, and the second best results are underlined. Time (s) denotes the average computation time required to attack a single image.", + "bbox": [ + 89, + 392, + 906, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "General settings. We adopt experimental configurations consistent with recent literature [2, 3, 49]. Specifically, we use an $\\ell_{\\infty}$ -norm setting with $\\epsilon = 16/255$ . We set decay $\\mu = 1.0$ and step size $\\eta = 2/255$ . We perform 300 iterations ( $T = 300$ ) for iterative targeted attacks as per [3]. For optimization of adversarial examples, all methods employ the simple logit loss function [49].", + "bbox": [ + 89, + 436, + 482, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. We evaluate our method using 15 pretrained DNNs, consisting of 10 CNN-based models and 5 transformer-based models. For CNN-based architectures, we employ VGG-16 [35], ResNet-18 (RN-18) [15], ResNet-50 (RN-50) [15], DenseNet-121 (DN-121) [17], Xception (Xcep) [5], MobileNet-v2 (MB-v2) [34], EfficientNet-B0 (EF-B0) [38], Inception ResNet-v2 (IRv2) [37], Inception-v3 (Inc-v3) [36], and Inception-v4 (Inc-v4) [37]. For Transformer-based architectures, we utilize Vision Transformer (ViT) [11], LeViT [14], ConViT [8], Twins [6], and Pooling-based ViT (PiT) [16].", + "bbox": [ + 89, + 547, + 482, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline attacks. We mainly compare our FTM against ten existing techniques: DI [46], RDI [50], MI [9], TI [10], SI [29], VT [39], Admix [40], ODI [2], SU [43], and CFM [3]. We primarily reference CFM, the current state-of-the-art method for targeted transfer-based attacks that requires no additional training dataset. MI and TI techniques apply across all attack methods, so we omit the 'MI-TI' prefix in our references. RDI serves as a common baseline technique in most cases. For DI, RDI, TI, SI, Admix, and ODI, we follow the configurations specified in CFM. For Admix, we also follow CFM to set the mixing weight to $w = 0.2$ and use $N = 3$ mixed images (i.e., $m_2 = 3$ )", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in [40]). Given Admix's computational intensity, we set the number of scale images in Admix to 1 (i.e., $m_{1} = 1$ in [40]) unless otherwise specified. We report the results for ODI instead of RDI-ODI, as ODI performs similarly to RDI-ODI [2, 3]. For SU and CFM, we use the original settings from their papers.", + "bbox": [ + 511, + 436, + 906, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings for our FTM. We set the mixing ratio $\\beta$ to 0.01. For FTM-E, we set the number of perturbed models to 2 for all experiments due to the computational cost. Following CFM, we avoid perturbing shallow layers because their low-level features vary significantly with input transforms and may cause excessive disturbance during attacks, as explained in CFM's paper. Specifically, we perturb layers where the feature map size is smaller than or equal to $\\frac{1}{16}$ of the image size. For a pure ViT, we select the outputs of all fully-connected layers. We sample $\\alpha_{k,i}$ from $\\mathcal{U}(0,0.75)$ , i.e., $\\alpha_{max} = 0.75$ . We set the probability $p$ to 0.1.", + "bbox": [ + 511, + 529, + 908, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 710, + 718, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transfer success rates. Table 1 shows the transfer success rates of various attack methods on ten black-box models. When using Inc-v3 as the source model, RDI-FTM-E achieves an average success rate of $51.2\\%$ , surpassing the previous best method CFM by $12.0\\%$ . The improvement is particularly notable for challenging target models like IR-v2 and Inc-v4, where RDI-FTM-E increases the success rates by $16.2\\%$ and $14.4\\%$ , respectively. When using RN-50 as the source model, both RDI-FTM and RDI-FTM-E achieve higher success rates than CFM. The results show that the effective feature-level augmentation employed by FTM helps", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25807", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/16bfb60c454a0539358d4d95e468ff3bfbd761ac182f38c3e021b9dbce07a3ef.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SourceAttackTime (s)ViTLeViTConViTTwinsPiTAvg.
RN-50DI1.370.23.50.41.51.71.5
RDI1.230.713.21.76.17.05.7
RDI-SI5.952.929.46.315.517.914.4
RDI-VT7.162.928.15.215.014.013.0
RDI-Admix3.551.322.52.58.58.48.6
RDI-SU2.210.816.92.36.97.86.9
ODI4.385.137.010.720.129.120.4
RDI-CFM1.394.346.18.925.224.721.8
RDI-FTM1.545.952.910.832.431.526.7
RDI-FTM-E2.926.858.613.635.234.929.8
Inc-v3DI2.010.10.30.00.00.10.1
RDI1.760.21.80.20.40.70.7
RDI-SI8.110.34.10.90.73.21.8
RDI-VT10.50.45.20.81.61.82.0
RDI-Admix4.920.14.10.61.41.41.5
RDI-SU2.360.22.00.31.31.01.0
ODI6.420.812.41.73.56.75.0
RDI-CFM2.132.121.93.26.111.68.9
RDI-FTM2.352.425.04.510.015.311.5
RDI-FTM-E4.373.832.76.812.720.415.3
", + "bbox": [ + 91, + 88, + 480, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Targeted attack success rates $(\\%)$ against five transformer-based DNNs. All methods are combined with MI-TI. The best results are shown in bold, and the second best are underlined. Time (s) denotes the average time required to attack a single image.", + "bbox": [ + 89, + 335, + 482, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7dcdebbf824c61a03c1b9e359164d9d4514c8b925267b7536ef5ebe8672be52e.jpg", + "image_caption": [ + "Figure 3. Targeted attack success rates $(\\%)$ based on the number of iterations. The source model is RN-50 and Inc-v3, respectively." + ], + "image_footnote": [], + "bbox": [ + 104, + 407, + 287, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46cca3ae0e79edbf6d3e1d5057ac9b7cce7f1939e2bae319f364e9a265714cd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 407, + 475, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e8d221cfab650e8ce4d2b3f98383c5f07bcd38cdfd59d8d6b110fae4ccab8f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 506, + 465, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "generate transferable adversarial examples.", + "bbox": [ + 89, + 579, + 375, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows the transfer success rates of various attack methods on five transformer-based models. Our methods also achieve significant improvements over existing methods. For instance, when using RN-50 as the source model, RDI-FTM and RDI-FTM-E improve the targeted attack success rates of CFM by $4.9\\%$ and $8.0\\%$ , respectively. The results suggest that our method can effectively improve the transferability of adversarial examples across different architectures.", + "bbox": [ + 89, + 595, + 482, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 3 shows the targeted attack success rates based on the number of iterations. We evaluate our methods under different numbers of iterations using two experimental settings: RN-50 → Inc-v4 and Inc-v3 → IR-v2. The results show that FTM and FTM-E achieve stable performance across different iteration counts. For the RN-50 → Inc-v4 setting, SI exhibits superior performance at the beginning of the optimization process. However, as the number of iterations increases, FTM surpasses existing approaches. In the Inc-v3 → IR-v2 setting, FTM-E significantly outperforms other methods across the entire range of iterations.", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee2c2b1cc7aad49487060a0c2ff970f4a325062a6c184849f37f18c191a460a5.jpg", + "image_caption": [ + "Figure 4. Attack success rates with different mixing ratio $\\beta$ . Each color represents the attack effectiveness on a specific target model." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 707, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c3990ce1590421259b6ab0f586688f6222a8952e5bb6e1eb2915d5e46bc7d74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 90, + 898, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational overhead. We analyze the computational efficiency of different methods and report their costs in Table 2. We use an NVIDIA 3090 GPU to compute the time per image. The results demonstrate that our approach introduces reasonable computational overhead compared to CFM. Specifically, with RN-50 as the source model, RDI-FTM requires an average of 1.54 seconds per image, representing only a $10.8\\%$ increase over CFM's 1.39 seconds. In contrast, alternative methods like SI, VT, and ODI impose substantially higher computational burdens. For instance, VT requires 7.16 seconds per image, and SI requires 5.95 seconds per image. We also include relevant analysis in Figure 1. The result shows that FTM achieves strong attack performance while maintaining low computational cost.", + "bbox": [ + 511, + 280, + 906, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Combination with existing techniques. Table 3 presents transfer success rates of our method combined with ensemble-based method [30], SI, and Admix. Using the ensemble of RN-50 and Inc-v3 as the surrogate model, our RDI-SI-FTM-E achieves $74.5\\%$ average success rate, surpassing existing methods. FTM notably enhances performance on transformer-based models. For instance, RDI-SI-FTM achieves $64.3\\%$ on PiT, which is $5.0\\%$ higher than RDI-SI-CFM $(59.3\\%)$ with comparable computational cost. These results demonstrate that FTM is a generic technique that combines effectively with existing attack methods.", + "bbox": [ + 511, + 491, + 905, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 667, + 666, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For all ablation experiments, we use RN-50 to generate adversarial examples and evaluate the average success rates on all models from Section 4.1, unless stated otherwise.", + "bbox": [ + 511, + 688, + 903, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stability with respect to $\\beta$ . Figure 4 shows the transfer success rates of FTM across different mixing ratios $\\beta$ . The success rates on five target models are plotted in distinct colors. When $\\beta$ exceeds 0.04, the success rates decrease substantially across all target models. However, for $\\beta$ values below 0.02, we observe stable attack performance across all models. The results suggest that FTM maintains robust performance within a reasonable range of $\\beta$ values, showing its stability with respect to $\\beta$ .", + "bbox": [ + 511, + 734, + 905, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compare FTM and CFM using ensemble strategy. Figure 5 shows the attack performance of FTM and CFM", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25808", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f984a5b2379bff75e2aaa0d3f394b75d0d051d9e049fc046189a8fa2d9569616.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AttackRN-50 + Inc-v3 ⇒
Time (s)VGG-16RN-50*Inc-v3*DN-121IR-v2Inc-v4XcepViTLeViTConViTTwinsPiTAvg.
RDI3.2666.998.498.783.135.642.525.40.822.43.29.810.841.5
ODI7.1770.994.096.183.465.469.957.69.152.918.628.642.857.4
RDI-CFM4.3986.198.397.092.171.779.972.010.366.520.341.945.865.2
RDI-ODI-CFM7.7363.988.681.374.154.356.054.312.649.920.929.641.052.2
RDI-Admix-CFM13.3886.298.197.792.775.181.676.114.073.124.850.753.968.7
RDI-SI-CFM22.1387.598.197.392.778.081.678.020.876.332.955.059.371.5
RDI-FTM4.4787.597.896.792.475.481.774.014.571.025.249.354.468.3
RDI-FTM-E8.8188.298.197.193.476.783.477.118.474.930.153.458.070.7
RDI-Admix-FTM14.1086.096.796.592.278.383.877.320.175.829.955.758.570.9
RDI-SI-FTM22.6687.596.896.293.379.083.679.125.777.737.559.964.373.4
RDI-SI-FTM-E43.4387.897.396.793.979.183.578.227.379.242.962.965.674.5
", + "bbox": [ + 116, + 88, + 877, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Targeted attack success rates (%) of combination with different methods. The surrogate model is the ensemble of RN-50 and Inc-v3. All methods are combined with MI-TI. The best results are shown in bold, and the second best are underlined. Time (s) denotes the average time required to attack a single image.", + "bbox": [ + 89, + 260, + 906, + 303 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/441d4b9e6024e07ed96a24419f7036373577becc02ba2ba83906194d0418ada6.jpg", + "image_caption": [ + "Figure 5. Attack performance and computational costs with different numbers of ensemble models. Left: Our FTM vs. CFM using ensemble strategy. Right: Computational costs of our FTM and RAP. -ens2 means using 2 perturbed models, and so on." + ], + "image_footnote": [], + "bbox": [ + 94, + 313, + 285, + 429 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/725eedec4cde1e59b1a51d51f613603b65e030a1423161c064331a43b94cf52f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 314, + 478, + 429 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf1e993cd0797ceeeda2af3be640f32a8884d641ec7f44753186e6331a743310.jpg", + "image_caption": [ + "Figure 6. Left: Attack success rates vs. $p$ and $\\alpha_{max}$ . Right: Attack success rates vs. $p$ with/without clean features; The blue line is the success rates when the clean features from CFM are integrated with our FTM, while the orange line represents only using our learnable perturbations." + ], + "image_footnote": [], + "bbox": [ + 94, + 503, + 285, + 619 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0821458157fe6861f9f42b074e1ca1b9b4dda958babba4fc600b4d48cac09113.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 505, + 478, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "using different numbers of perturbed models. On the left, our FTM consistently outperforms CFM across all ensemble sizes. Notably, FTM with only 2 perturbed models achieves better attack performance than an ensemble of 5 CFM-perturbed models, underscoring FTM's superior effectiveness. It is also worth mentioning that single FTM and CFM have comparable time complexity (Table 2). On the right, we compare our FTM with RAP [33], and the results show that FTM significantly outperforms RAP in both effectiveness and computational efficiency.", + "bbox": [ + 89, + 715, + 482, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of probability $p$ and ratio $\\alpha_{\\text{max}}$ . Figure 6 presents an analysis of FTM's attack success rates under", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "varying values of $p$ and $\\alpha_{max}$ . The left plot shows that FTM achieves optimal performance with $p = 0.1$ and $\\alpha_{max} = 0.75$ . FTM maintains competitive performance under various parameter configurations, such as $p, \\alpha_{max} = 0.15, 0.5$ or $0.05, 1.0$ , indicating its parameter stability. The right plot illustrates the impact of clean features on FTM. With clean features enabled, FTM's performance declines significantly as $p$ increases from 0.1 to 1.0, while FTM without clean features shows stable performance across varying $p$ values. This suggests that although clean features are crucial to FTM's performance, they also add instability to the attack process. Additionally, setting $p = 1.0$ (where all layers are updated) causes excessive perturbations, making the attack overly aggressive and reducing its effectiveness.", + "bbox": [ + 511, + 314, + 906, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The drop of the blue line in Figure 6 occurs because the clean accuracy of the perturbed surrogate model becomes extremely low $(3.8\\%)$ at $p = 1.0$ when using clean feature mixup. Without clean feature mixup (orange line), the accuracy remains high at $65.3\\%$ . The low accuracy makes it impossible to generate effective adversarial examples, causing the attack success rate of the blue line to drop to near 0.", + "bbox": [ + 511, + 526, + 906, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 643, + 633, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we study the challenging problem of enhancing transferability in targeted adversarial attacks. Our investigation reveals that feature-level augmentation strategies show promising potential for boosting transferability effectively and efficiently. We introduce Feature Tuning Mixup (FTM), a novel method that incorporates both attack-specific feature perturbations and clean features in the feature space. To optimize the learnable perturbations efficiently, we develop a momentum-based stochastic update strategy. Comprehensive experiments on the ImageNet-compatible dataset demonstrate that FTM achieves substantial improvement over current state-of-the-art methods while maintaining low computational cost.", + "bbox": [ + 511, + 667, + 906, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by HK RGC GRF under Grant PolyU 15201323.", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25809", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. In ICLR, 2018. 2", + "[2] Junyoung Byun, Seungju Cho, Myung-Joon Kwon, Hee-Seon Kim, and Changick Kim. Improving the transferability of targeted adversarial examples through object-based diverse input. In CVPR, 2022. 1, 3, 5, 6", + "[3] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In CVPR, 2023. 1, 3, 4, 5, 6", + "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy (SP), 2017. 2", + "[5] Francois Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, 2017. 6", + "[6] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In NeurIPS, 2021. 6", + "[7] Xuelong Dai, Kaisheng Liang, and Bin Xiao. Advdiff: Generating unrestricted adversarial examples using diffusion models. In ECCV, 2024. 2", + "[8] Stéphane d'Ascoli, Hugo Touvron, Matthew L. Leavitt, Ari S. Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In ICML, 2021. 6", + "[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In CVPR, 2018. 1, 2, 3, 4, 6", + "[10] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In CVPR, 2019. 1, 3, 6", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 6", + "[12] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, and Dawn Song. Robust physical-world attacks on deep learning visual classification. In CVPR, 2018. 2", + "[13] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1, 2", + "[14] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In ICCV, 2021. 6", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6", + "[16] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, 2021. 6" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 6", + "[18] Qian Huang, Isay Katsman, Zeqi Gu, Horace He, Serge J. Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In ICCV, 2019. 3", + "[19] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. In NeurIPS, 2019. 2", + "[20] Nathan Inkawich, Wei Wen, Hai (Helen) Li, and Yiran Chen. Feature space perturbations yield more transferable adversarial examples. In CVPR, 2019. 3", + "[21] Nathan Inkawich, Kevin J. Liang, Lawrence Carin, and Yiran Chen. Transferable perturbations of deep feature distributions. In ICLR, 2020. 3", + "[22] Nathan Inkawich, Kevin J. Liang, Binghui Wang, Matthew Inkawich, Lawrence Carin, and Yiran Chen. Perturbing across the feature hierarchy to improve standard and strict blackbox attack transferability. In NeurIPS, 2020. 3", + "[23] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In ICLR (Workshop), 2017. 1, 3", + "[24] Maosen Li, Cheng Deng, Tengjiao Li, Junchi Yan, Xinbo Gao, and Heng Huang. Towards transferable targeted attack. In CVPR, 2020. 3", + "[25] Qizhang Li, Yiwen Guo, and Hao Chen. Yet another intermediate-level attack. In ECCV, 2020. 3", + "[26] Yanjie Li, Yiquan Li, Xuelong Dai, Songtao Guo, and Bin Xiao. Physical-world optical adversarial attacks on 3d face recognition. In CVPR, 2023. 2", + "[27] Yanjie Li, Kaisheng Liang, and Bin Xiao. UV-attack: Physical-world adversarial attacks on person detection via dynamic-NeRF-based UV mapping. In ICLR, 2025. 2", + "[28] Kaisheng Liang and Bin Xiao. Style: Boosting the transferability of adversarial examples. In CVPR, 2023. 2", + "[29] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In ICLR, 2020. 1, 3, 6", + "[30] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In ICLR, 2017. 3, 7", + "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 1, 2, 5", + "[32] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. On generating transferable targeted perturbations. In ICCV, 2021. 3", + "[33] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. In NeurIPS, 2022. 3, 8", + "[34] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25810", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 6", + "[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 6", + "[37] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A. Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In AAAI, 2017. 6", + "[38] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 6", + "[39] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In CVPR, 2021. 1, 3, 6", + "[40] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In ICCV, 2021. 1, 3, 6", + "[41] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In ICCV, 2021. 3", + "[42] Zhibo Wang, Hongshan Yang, Yunhe Feng, Peng Sun, Hengchang Guo, Zhifei Zhang, and Kui Ren. Towards transferable targeted adversarial examples. In CVPR, 2023. 3", + "[43] Zhipeng Wei, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Enhancing the self-universality for transferable targeted attacks. In CVPR, 2023. 1, 3, 5, 6", + "[44] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip connections matter: On the transferability of adversarial examples generated with resnets. In ICLR, 2020. 3", + "[45] Weibin Wu, Yuxin Su, Michael R Lyu, and Irwin King. Improving the transferability of adversarial samples with adversarial transformations. In CVPR, 2021. 3", + "[46] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019. 1, 3, 6", + "[47] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Boosting transferability of targeted adversarial examples via hierarchical generative networks. In ECCV, 2022. 3", + "[48] Jianping Zhang, Weibin Wu, Jen-tse Huang, Yizhan Huang, Wenxuan Wang, Yuxin Su, and Michael R Lyu. Improving adversarial transferability via neuron attribution-based attacks. In CVPR, 2022. 3", + "[49] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. On success and simplicity: A second look at transferable targeted attacks. In NeurIPS, 2021. 1, 3, 5, 6", + "[50] Junhua Zou, Zhisong Pan, Junyang Qiu, Xin Liu, Ting Rui, and Wei Li. Improving the transferability of adversarial examples with resized-diverse-inputs, diversity-ensemble and region fitting. In ECCV, 2020. 3, 6" + ], + "bbox": [ + 91, + 92, + 480, + 872 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25811", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/5356bb72-29a6-442b-984c-adce3c009504_model.json b/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/5356bb72-29a6-442b-984c-adce3c009504_model.json new file mode 100644 index 0000000000000000000000000000000000000000..737cec713dbda2da3d406b9aa440f8d67ec33ee4 --- /dev/null +++ b/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/5356bb72-29a6-442b-984c-adce3c009504_model.json @@ -0,0 +1,2288 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.131, + 0.852, + 0.154 + ], + "angle": 0, + "content": "Improving Transferable Targeted Attacks with Feature Tuning Mixup" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.181, + 0.827, + 0.219 + ], + "angle": 0, + "content": "Kaisheng Liang Xuelong Dai Yanjie Li Dong Wang Bin Xiao* Department of Computing, The Hong Kong Polytechnic University" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.22, + 0.751, + 0.252 + ], + "angle": 0, + "content": "cskliang@comp.polyu.edu.hk b.xiao@polyu.edu.hk \nelong.dai, yanjie.li, dong-comp.wang}@connect.polyu.hk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.32, + 0.485, + 0.699 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) exhibit vulnerability to adversarial examples that can transfer across different DNN models. A particularly challenging problem is developing transferable targeted attacks that can mislead DNN models into predicting specific target classes. While various methods have been proposed to enhance attack transferability, they often incur substantial computational costs while yielding limited improvements. Recent clean feature mixup methods use random clean features to perturb the feature space but lack optimization for disrupting adversarial examples, overlooking the advantages of attack-specific perturbations. In this paper, we propose Feature Tuning Mixup (FTM), a novel method that enhances targeted attack transferability by combining both random and optimized noises in the feature space. FTM introduces learnable feature perturbations and employs an efficient stochastic update strategy for optimization. These learnable perturbations facilitate the generation of more robust adversarial examples with improved transferability. We further demonstrate that attack performance can be enhanced through an ensemble of multiple FTM-perturbed surrogate models. Extensive experiments on the ImageNet-compatible dataset across various DNN models demonstrate that our method achieves significant improvements over state-of-the-art methods while maintaining low computational cost." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.73, + 0.222, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.483, + 0.862 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) are highly vulnerable to adversarial attacks [13, 23, 31], raising critical security concerns. In image classification tasks, adversarial attacks can deceive DNNs by adding imperceptible perturbations to benign images, causing the models to produce erroneous predictions. More concerningly, research [9, 10, 46] has demonstrated that adversarial examples possess trans" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.291, + 0.882, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.453, + 0.907, + 0.523 + ], + "angle": 0, + "content": "Figure 1. Effectiveness and efficiency of targeted attacks. Average attack success rates on 14 black-box models, along with the computation time required to generate an adversarial example. Our methods demonstrate superior performance with low computational cost, surpassing state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.67 + ], + "angle": 0, + "content": "ferability properties: adversarial examples generated using one white-box model can successfully fool other black-box models without requiring access to their architectures or parameters. This cross-model transferability of adversarial examples poses severe security threats to real-world AI systems. Therefore, studying the transferability of adversarial examples is crucial for enhancing the robustness of DNNs against such security vulnerabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Targeted transfer-based attacks aim to mislead models into predicting specific target classes, which remains a significant challenge. Due to the lack of knowledge about the target model's training data and architecture, such attacks tend to generate targeted adversarial examples that overfit to surrogate models and fail to transfer to target models. Various transferable targeted attacks [2, 43, 49] have been proposed but still leave much room for improvement. Notably, data augmentation methods [29, 39, 40] are commonly employed for targeted attacks. But they primarily operate in the image space, while augmentation methods at the feature level remain relatively unexplored. Recently, the Clean Feature Mixup (CFM) method [3] has made important progress by using clean features from random images to disrupt adversarial examples, leading to better targeted" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.482, + 0.901 + ], + "angle": 0, + "content": "1Our code is available at https://github.com/uhiu/features-tuning-mixup" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25802" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.133, + 0.09, + 0.868, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.445, + 0.311, + 0.733, + 0.321 + ], + "angle": 0, + "content": "Gradients from the Backward Process in Iteration i of the Attack" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.334, + 0.908, + 0.378 + ], + "angle": 0, + "content": "Figure 2. Overview of our FTM attack. In the forward pass, the learnable perturbations are added to the output features. Only a small portion of network layers are randomly selected for perturbation updates and clean feature mixup in each attack iteration. In the backward pass, the gradients of the selected perturbations and adversarial image are computed together for update." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.396, + 0.482, + 0.488 + ], + "angle": 0, + "content": "attack success rates (Figure 1). However, CFM uses random clean features without optimizing them specifically for disrupting adversarial examples. We hypothesize that carefully optimized perturbations would be more effective in disrupting adversarial examples, potentially yielding substantial improvements in attack transferability." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.49, + 0.483, + 0.777 + ], + "angle": 0, + "content": "In this work, we propose Feature Tuning Mixup (FTM), a novel attack method that optimizes attack-specific feature perturbations to enhance the transferability of targeted adversarial examples. Our method fine-tunes the intermediate features of surrogate models during the attack process by introducing learnable feature perturbations. These feature perturbations are optimized by mixing them with clean features from random images to counteract the adversarial effects of targeted adversarial examples. Through this mechanism, adversarial examples are gradually compelled to become more robust, allowing them to overcome these feature-level disturbances. To address the computational overhead of optimizing feature perturbations, we propose a stochastic update strategy. This strategy initializes the learnable perturbations using values from previous attack iterations and selectively optimizes only a small subset of feature perturbations in each iteration. Ultimately, FTM significantly improves the transferability of targeted adversarial examples while maintaining computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We provide an overview of the proposed FTM in Figure 2. FTM first introduces learnable perturbations into the deep layers of the surrogate model. The learnable perturbations are added element-wise to the original output features before passing to the next layer. In each attack iteration, a small subset of layers is randomly selected and combined with random clean features. Finally, we compute gradients for the learnable perturbations and adversarial im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.906, + 0.501 + ], + "angle": 0, + "content": "age using only one forward and backward pass, similar to standard attacks. We update the selected learnable perturbations and adversarial image, while unselected perturbations remain unchanged. As a result, FTM fully leverages optimized attack-specific feature perturbations, prevents overfitting to surrogate models, and significantly enhances the transferability of targeted attacks." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.504, + 0.836, + 0.517 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.524, + 0.905, + 0.583 + ], + "angle": 0, + "content": "- We revisit feature-level augmentation methods for transferable targeted attacks and identify that incorporating attack-specific feature perturbations, previously overlooked, can efficiently boost existing attacks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.584, + 0.905, + 0.674 + ], + "angle": 0, + "content": "- We introduce Feature Tuning Mixup (FTM), a novel approach that combines random and optimized feature perturbations to enhance targeted transfer-based attacks. Our method includes an efficient stochastic update strategy that maintains computational efficiency while improving attack transferability." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.675, + 0.905, + 0.735 + ], + "angle": 0, + "content": "- We conduct extensive experiments across various source and target DNN models on the ImageNet-compatible dataset, demonstrating that our FTM achieves a significant improvement over state-of-the-art methods." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.524, + 0.905, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.753, + 0.655, + 0.769 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Adversarial attacks [4, 12, 13, 19, 31] have been widely studied in the field of adversarial learning, including unrestricted attacks [7], physical attacks [26, 27], and black-box attacks [1, 9, 28]. Notably, black-box attacks can be categorized into query-based [1] and transfer-based attacks [9, 28]. The latter, which does not require access to the target DNN model's outputs, parameters or architecture, is more practical and is the focus of this paper." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25803" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.414, + 0.108 + ], + "angle": 0, + "content": "2.1. Non-targeted Transfer-based Attacks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.628 + ], + "angle": 0, + "content": "Current non-targeted transfer attacks are primarily based on the Iterative-Fast Gradient Sign Method (I-FGSM) [23]. This approach iteratively computes the gradient of the classification loss with respect to the input and updates adversarial examples in the direction that maximizes the loss. However, it often leads to overfitting on white-box models, making the adversarial examples less effective against other black-box models. Numerous studies have explored solutions to the overfitting issue to enhance adversarial transferability. The Momentum Iterative FGSM (MIFGSM) [9] incorporates a momentum-based optimization method during iterative attacks. Translation-Invariant (TI-FGSM) method [10] enhances the transferability of adversarial examples by achieving translation invariance through gradient smoothing. Data augmentation techniques are widely used to mitigate overfitting. The Diverse Inputs (DI) method [46] employs random resizing and padding with a probability \\( p \\) to generate diverse input, while the Resized Diverse Inputs (RDI) method [50] fixes \\( p = 1 \\) and rescales DI-enhanced images back to their original size. The Scale-Invariant (SI) method [29] generates adversarial perturbations using multiple scale transformations, and Admix [40] extends SI by blending images from other labels into the input. Variance Tuning (VT) method [39] randomly selects data points around a given data point and computes their gradients to avoid overfitting. These methods operate in the image space. Feature-level approaches have also been studied to improve adversarial transferability. For instance, the Feature Importance-aware Attack (FIA) [41] identifies and perturbs important features in the intermediate layers of a surrogate model. The Skip Gradient Method (SGM) [44] utilizes more gradients from skip connections. Ensemble-based attacks [9, 30] utilize multiple surrogate models to generate adversarial examples." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.381, + 0.654 + ], + "angle": 0, + "content": "2.2. Targeted Transfer-based Attacks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Targeted attacks are more challenging than non-targeted attacks because they not only aim to cause incorrect predictions by the models but also to mislead the models into predicting a specific target class. We can categorize targeted attacks into two types based on whether they require a training dataset. Previous works have utilized training datasets to generate auxiliary modules [21, 22] for transfer-based attacks or directly train a model to generate adversarial examples [32, 42, 47]. This paper primarily focuses on transfer-based attacks that do not require a training dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Targeted transfer-based attacks are widely studied in adversarial machine learning due to their practicality, as they do not require access to the target model's architecture and parameters, nor do they need training datasets. The Activation Attack (AA) [20] proposes using the distance between adversarial examples and targeted images in feature" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.319 + ], + "angle": 0, + "content": "space as a loss function, but its attack effectiveness may depend on the selection of targeted images. Po+Trip [24] employs Poincare distance to adapt the gradients' magnitude and uses triplet loss to address insufficient inter-class distances, but it only works well on target models with similar architectures. Logit Attack (Logit) [49] discovers that a simple logit loss can generate better transferable adversarial examples than CE loss, significantly outperforming previous methods. Self-Universality (SU) method [43] finds that highly universal adversarial perturbations tend to be more transferable for targeted attacks and proposes optimizing the perturbations in different local regions of the image. The Object-based Diverse Input (ODI) method [2] is an efficient data augmentation in image space for targeted attacks using random 3D object and rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.908, + 0.488 + ], + "angle": 0, + "content": "Recently, the Clean Feature Mixup (CFM) method [3] proposes a feature-level augmentation which achieves state-of-the-art performance in targeted transfer-based attacks. It stores clean features as random feature perturbations, which mitigates the overfitting to surrogate models. However, the proposed clean features in CFM lack specific optimization for targeted attacks, leaving significant room for improvement. We develop the Feature Tuning Mixup (FTM) method, which effectively and efficiently optimizes attack-specific feature perturbations and significantly improves transferable targeted attacks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.909, + 0.718 + ], + "angle": 0, + "content": "Differences from intermediate-level attacks. This paper investigates feature-level augmentation for targeted attacks, which differs from existing intermediate-level attacks [18, 25, 41, 48]. Our FTM is based on the transfer-based attack framework [23] with logit loss [49], which is currently the most effective framework for targeted transfer-based attacks. In contrast, intermediate-level attacks use loss functions based on a specific intermediate layer, which is hard to apply to targeted attacks. For instance, FIA [41] and NAA [48] require identifying important features based on the true label, while targeted attacks focus on the target label instead of the true label. ILA [18] requires an existing adversarial example as guidance and cannot function as a standalone targeted attack. Therefore, we do not compare FTM with intermediate-level attacks in experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Comparison with adversarial learning methods. Our method learns feature perturbations in an adversarial manner without requiring additional training datasets. While some existing attacks [33, 45] also employ adversarial strategies, they operate in the image space and typically require access to training datasets. For instance, ATTA [45] improves transferability by training an adversarial transformation network to destroy adversarial examples. However, this approach demands substantial computational resources and training data. Like previous work [2, 3, 43, 49], we study targeted transfer attacks without extra data and do not compare with methods that need training data. RAP [33]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25804" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.244 + ], + "angle": 0, + "content": "utilizes adversarial learning to identify worst-case perturbations in the image space, aiming to locate regions with flat loss landscapes. However, this process requires numerous additional backpropagation steps, leading to prohibitively high computational costs. In contrast, we propose a novel approach that optimizes feature perturbations directly in the multi-layer feature space, an unexplored direction in previous research. Through this innovation, our method achieves superior practicality, effectiveness, and efficiency compared to existing methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.255, + 0.226, + 0.273 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.28, + 0.232, + 0.294 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.485, + 0.468 + ], + "angle": 0, + "content": "Problem formulation. For targeted transfer-based attacks, the attacker is assumed to have a white-box surrogate model \\( F \\), which is a pre-trained image classifier in this paper. We assume that the attacker has no additional training dataset. Given a benign image \\( x \\), a surrogate model \\( F \\) and a target label \\( y_{t} \\), which is different from the true label \\( y \\) of \\( x \\). The goal of the attack is to generate an adversarial example \\( x^{\\mathrm{adv}} \\) that is misclassified as \\( y_{t} \\) by \\( F \\). The difference between \\( x \\) and \\( x^{\\mathrm{adv}} \\) should be subtle, and following previous work, we use the \\( \\ell_{\\infty} \\)-norm constraint. The optimization objective can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.477, + 0.484, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {x ^ {\\mathrm {a d v}}} \\mathcal {L} (F (x ^ {\\mathrm {a d v}}), y _ {t}), \\quad s. t. \\quad \\| x - x ^ {\\mathrm {a d v}} \\| _ {\\infty} \\leq \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.483, + 0.539 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) is the adversarial loss, e.g., the cross-entropy loss or logit-based loss, and \\(\\epsilon\\) is the perturbation budget." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.584 + ], + "angle": 0, + "content": "MI-FGSM. To solve the problem above, existing methods usually adopt the iterative gradient-based technique, e.g., the MI-FGSM [9], formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.593, + 0.483, + 0.611 + ], + "angle": 0, + "content": "\\[\ng _ {i + 1} = \\nabla_ {x} \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}\\right), y _ {t}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.613, + 0.483, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\tilde {g} _ {i + 1} = \\mu \\cdot g _ {i} + g _ {i + 1} / \\| g _ {i + 1} \\| _ {1}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.632, + 0.483, + 0.65 + ], + "angle": 0, + "content": "\\[\nx _ {i + 1} ^ {\\mathrm {a d v}} = x _ {i} ^ {\\mathrm {a d v}} - \\eta \\cdot \\operatorname {s i g n} \\left(\\tilde {g} _ {i + 1}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.653, + 0.483, + 0.671 + ], + "angle": 0, + "content": "\\[\nx _ {i + 1} ^ {\\mathrm {a d v}} = \\operatorname {C l i p} _ {x, \\epsilon} \\left(x _ {i + 1} ^ {\\mathrm {a d v}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.483, + 0.74 + ], + "angle": 0, + "content": "where \\( g_{i} \\) is the gradient of the loss function with respect to the input image \\( x \\) in the \\( i \\)-th iteration. \\( \\mu \\) is the momentum decay factor, and \\( \\eta \\) is the step size. \\( x_{i+1}^{\\mathrm{adv}} \\) is clipped within \\( \\epsilon \\) bound. Initially, we set \\( x_{0}^{\\mathrm{adv}} = x \\) and \\( g_{0} = 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.483, + 0.875 + ], + "angle": 0, + "content": "CFM. To improve the transferability of targeted attacks, the Clean Feature Mixup (CFM) method [3] leverages stored clean features to augment adversarial examples, achieving state-of-the-art performance among existing targeted transfer-based attacks. Specifically, it selects outputs from certain network layers as perturbation targets and uses output features from different benign images as feature perturbations. To integrate with MI-FGSM, the feature augmentation in CFM can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nz _ {k, i} ^ {\\prime} = \\left(1 - \\alpha_ {k, i}\\right) \\odot z _ {k, i} + \\alpha_ {k, i} \\odot z _ {k, i} ^ {c}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise multiplication, \\(z_{k,i}\\) is the input feature to the \\(k\\)-th layer in the \\(i\\)-th attack iteration of MI-FGSM, \\(z_{k,i}^{c}\\) is a random stored clean feature, and \\(\\alpha_{k,i}\\) is a vector representing channel-wise mixing ratios. The network layer \\(k\\) is randomly selected from pre-defined candidate layers with probability \\(p\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.19, + 0.681, + 0.205 + ], + "angle": 0, + "content": "3.2. Analysis of CFM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.212, + 0.906, + 0.273 + ], + "angle": 0, + "content": "CFM is a powerful feature-level augmentation method for targeted attacks. To understand the effectiveness of CFM, we formulate a generalized optimization step for MI-FGSM based on feature-level augmentation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.281, + 0.905, + 0.299 + ], + "angle": 0, + "content": "\\[\ng _ {i + 1} ^ {\\prime} = \\nabla_ {x} \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}; \\Delta z _ {i}\\right), y _ {t}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.301, + 0.905, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\Delta z _ {i} = \\left\\{\\Delta z _ {k, i}, k = 1, \\dots , n \\right\\}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.906, + 0.478 + ], + "angle": 0, + "content": "where \\( F(x_{i}^{\\mathrm{adv}};\\Delta z_{i}) \\) denotes the surrogate model \\( F \\) with the perturbation \\( \\Delta z_{i} \\) added to the network outputs in the \\( i \\)-th attack iteration. Equation (7) improves upon Equation (2) by incorporating feature-level perturbations, which is the key idea of CFM. Notably, the perturbation \\( \\Delta z_{i} \\) represents the collection of perturbations for all network layers. We use \\( \\Delta z_{k,i} \\) to denote the perturbation for the input of the \\( k \\)-th layer, where \\( n \\) is the maximum layer index of \\( F \\). Initially, we set \\( \\Delta z_{i} = 0 \\). For any layer \\( d \\) that does not require perturbation, we simply set \\( \\Delta z_{d,i} \\equiv 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.539 + ], + "angle": 0, + "content": "CFM uses clean features to construct \\(\\Delta z_{i}\\), and it claims that the rationale is that clean features can suppress the adversarial effect of \\(x_{i}^{\\mathrm{adv}}\\) [3]. We formulate this idea in the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.545, + 0.905, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}; \\Delta z _ {i}\\right), y _ {t}\\right) > \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}\\right), y _ {t}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.906, + 0.602 + ], + "angle": 0, + "content": "where \\(\\Delta z_{i}\\) is supposed to appropriately disrupt adversarial loss in Equation (1) for obtaining robust \\(x^{\\mathrm{adv}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.906, + 0.723 + ], + "angle": 0, + "content": "However, clean features are attack-agnostic in nature, as they are not explicitly optimized for the objective in Equation (9). Although CFM demonstrates promising improvements, we argue that there remains unexplored potential in feature-level augmentation for targeted attacks. We believe that integrating attack-specific optimization with attack-agnostic clean features can create a more powerful feature-level augmentation strategy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.731, + 0.78, + 0.747 + ], + "angle": 0, + "content": "3.3. Feature Tuning Mixup (FTM)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.753, + 0.906, + 0.829 + ], + "angle": 0, + "content": "We propose Feature Tuning Mixup (FTM), a novel method that uses attack-specific feature perturbations to enhance feature-level augmentation methods. We formulate the optimization problem for learning our feature perturbations and adversarial examples in each attack iteration as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.836, + 0.905, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\min _ {x _ {i} ^ {\\mathrm {a d v}}} \\max _ {\\Delta z _ {i}} \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}; \\Delta z _ {i}\\right), y _ {t}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "However, the optimization problem above is challenging to solve from two perspectives: effectiveness and efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25805" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "Regarding efficiency, traditional adversarial training [31] often requires multiple steps to optimize \\(\\Delta z_{i}\\) in each attack iteration, resulting in substantial additional computational cost. As for effectiveness, since \\(\\Delta z_{i}\\) involves a joint problem across multiple iterations and network layers, inappropriate perturbations can damage the attack effect, which will be shown in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.483, + 0.44 + ], + "angle": 0, + "content": "Momentum-based stochastic update. We propose a novel technique, momentum-based stochastic update, to optimize \\(\\Delta z_{i}\\). We employ one-step optimization with using perturbations in previous iterations as momentum, i.e., \\(\\Delta z_{i}\\) will be initialized as \\(\\Delta z_{i-1}\\) in Equation (10). Momentum can effectively improve the optimization effectiveness without increasing additional optimization steps. In addition, as we only require one-step optimization, the gradient for \\(\\Delta z_{i}\\) can be jointly computed with the gradient for \\(x_{i}^{\\mathrm{adv}}\\), which can further speed up the optimization process. To further improve the optimization efficiency, we use a stochastic update strategy, where only the randomly selected layers are updated in each attack iteration. For each selected layer \\(k\\), it has probability \\(p\\) to be updated in the \\(i\\)-th attack iteration, otherwise, we set \\(\\Delta z_{k,i} = \\Delta z_{k,i-1}\\). We illustrate the detailed optimization process for FTM in the following." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.483, + 0.471 + ], + "angle": 0, + "content": "The forward process for FTM in the \\(i\\)-th attack iteration is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.482, + 0.483, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\bar {z} _ {k, i} = z _ {k, i} + \\beta \\| z _ {k, i} \\| \\cdot \\frac {\\Delta z _ {k , i}}{\\| \\Delta z _ {k , i} \\| + \\bar {\\epsilon}}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.518, + 0.482, + 0.559 + ], + "angle": 0, + "content": "\\[\nz _ {k, i} ^ {\\prime} = \\left\\{ \\begin{array}{l l} (1 - \\alpha_ {k, i}) \\odot \\bar {z} _ {k, i} + \\alpha_ {k, i} \\odot z _ {k, i} ^ {c}, & \\tau_ {k} < p, \\\\ \\bar {z} _ {k, i}, & \\text {e l s e ,} \\end{array} \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.483, + 0.768 + ], + "angle": 0, + "content": "where \\(\\beta \\| z_{k,i}\\|\\) is the custom perturbation budget for our feature perturbation \\(\\Delta z_{k,i}\\). \\(\\beta\\) is a scaling factor, and \\(\\bar{\\epsilon}\\) is a small constant. We mix the output feature \\(z_{k,i}\\) and our attack-specific perturbation \\(\\Delta z_{k,i}\\) to obtain \\(\\bar{z}_{k,i}\\). Then, we use clean features to augment \\(\\bar{z}_{k,i}\\) to get \\(z_{k,i}^{\\prime}\\), which is the final perturbed feature that will be fed into the \\(k\\)-th network layer. These perturbed features are computed layer by layer during the forward process. \\(z_{k,i}^{c}\\) is a random clean feature, and \\(\\alpha_{k,i}\\) is a vector representing channel-wise mixing ratio. We randomly sample \\(\\alpha_{k,i}\\) from \\(\\mathcal{U}(0,\\alpha_{max})\\). In each iteration, we randomly sample \\(\\tau_{k}\\) from \\(\\mathcal{U}(0,1)\\) to determine whether to update the \\(k\\)-th perturbation \\(\\Delta z_{k,i}\\) during the \\(i\\)-th attack iteration. \\(p\\) is the probability threshold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.799 + ], + "angle": 0, + "content": "The backward process for FTM in the \\(i\\)-th attack iteration is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.814, + 0.483, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\Delta z _ {i} ^ {\\prime} = \\left\\{\\left(\\Delta z _ {k, i}, \\tau_ {k}\\right) \\mid k = 1, \\dots , n \\right\\}, \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.835, + 0.482, + 0.856 + ], + "angle": 0, + "content": "\\[\ng _ {i + 1} ^ {\\prime}, g ^ {\\Delta z ^ {\\prime}} = \\nabla_ {x _ {i} ^ {\\mathrm {a d v}}, \\Delta z _ {i} ^ {\\prime}} \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}; \\Delta z _ {i} ^ {\\prime}\\right), y _ {t}\\right), \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.858, + 0.483, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\Delta z _ {k, i + 1} = \\left\\{ \\begin{array}{l l} \\Delta z _ {k, i} + g _ {k} ^ {\\Delta z ^ {\\prime}}, & \\tau_ {k} < p, \\\\ \\Delta z _ {k, i}, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {15}\n\\]" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.092, + 0.876, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Feature Tuning Mixup (FTM) Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.11, + 0.905, + 0.156 + ], + "angle": 0, + "content": "Input: Surrogate model \\(F\\), clean image \\(x\\), target label \\(y_{t}\\), iteration number \\(T\\), perturbation budget \\(\\epsilon\\), input transformation \\(\\phi(\\cdot)\\) (RDI, etc.), step size \\(\\eta\\), decay factor \\(\\mu\\)" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.157, + 0.895, + 0.186 + ], + "angle": 0, + "content": "Input: Mixing range \\(\\alpha_{max}\\), probability \\(p\\), mixing ratio \\(\\beta\\) \nOutput: An adversarial example \\(x^{\\mathrm{adv}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.186, + 0.66, + 0.202 + ], + "angle": 0, + "content": "1: \\(x_0^{\\mathrm{adv}} = x, g_0 = 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.203, + 0.75, + 0.216 + ], + "angle": 0, + "content": "2: Initialize perturbation \\(\\Delta z_0 = 0\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.217, + 0.697, + 0.23 + ], + "angle": 0, + "content": "3: for \\( i = 0 \\rightarrow T - 1 \\) do:" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.231, + 0.767, + 0.246 + ], + "angle": 0, + "content": "4: Extend \\( F(x_{j}^{\\mathrm{adv}}) \\) to \\( F(\\phi (x_{j}^{\\mathrm{adv}})) \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.247, + 0.887, + 0.263 + ], + "angle": 0, + "content": "5: Select layers to update and obtain \\(\\Delta z_{i}^{\\prime}\\) from \\(\\Delta z_{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.263, + 0.903, + 0.293 + ], + "angle": 0, + "content": "6: Forward and backward pass to get \\( g_{i+1}' \\) and \\( g^{\\Delta z'} \\) using Equations (12) and (14)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.294, + 0.796, + 0.308 + ], + "angle": 0, + "content": "7: Obtain \\(\\Delta z_{i + 1}\\) using Equation (15)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.309, + 0.832, + 0.324 + ], + "angle": 0, + "content": "8: Calculate \\(\\tilde{g}_{i + 1} = \\mu \\cdot g_{i} + g_{i + 1}^{\\prime} / \\| g_{i + 1}^{\\prime}\\|_{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.324, + 0.869, + 0.339 + ], + "angle": 0, + "content": "9: Update example \\( x_{i+1}^{\\mathrm{adv}} = x_i^{\\mathrm{adv}} - \\eta \\cdot \\operatorname{sign}(\\tilde{g}_{i+1}) \\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.34, + 0.788, + 0.354 + ], + "angle": 0, + "content": "10: Clip \\( x_{i+1}^{\\mathrm{adv}} \\) to be \\( \\ell_{\\infty} \\)-bounded by \\( \\epsilon \\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.355, + 0.597, + 0.367 + ], + "angle": 0, + "content": "11: end for" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.368, + 0.621, + 0.385 + ], + "angle": 0, + "content": "12: \\(x^{\\mathrm{adv}} = x_T^{\\mathrm{adv}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.385, + 0.621, + 0.397 + ], + "angle": 0, + "content": "13: return \\( x^{\\mathrm{adv}} \\)" + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.186, + 0.903, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.905, + 0.534 + ], + "angle": 0, + "content": "where \\(\\Delta z_{i}^{\\prime}\\) represents a set containing both perturbations \\(\\Delta z_{k,i}\\) and their corresponding random variables \\(\\tau_{k}\\) from the forward process, which determines whether each perturbation needs to be updated. \\(g_{k}^{\\Delta z^{\\prime}}\\) denotes the gradient with respect to the \\(k\\)-th perturbation \\(\\Delta z_{k,i}\\). The threshold \\(p\\) remains consistent with that used in the forward process. We summarize our FTM attack in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.905, + 0.625 + ], + "angle": 0, + "content": "Efficiency analysis. FTM does not require complex operations or additional forward and backward passes. Therefore, FTM is theoretically efficient on both CNNs and ViTs. In contrast, SI, VT, Admix, and SU require additional forward and backward passes, while ODI requires a computationally expensive input transformation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Ensemble of surrogate variants (FTM-E). Observing that FTM can efficiently perturb the surrogate model, we propose to further enhance the attack effectiveness through an ensemble strategy. We create multiple copies of the surrogate model and independently apply FTM to each copy, then ensemble them as a new surrogate model. We empirically show that utilizing an ensemble of two perturbed surrogate models can substantially enhance the attack performance while maintaining reasonable computational costs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.778, + 0.646, + 0.795 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.721, + 0.82 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Datasets. Following prior research [2, 3, 43, 49], we conduct experiments on the ImageNet-compatible dataset from the NIPS 2017 adversarial attack challenge. This dataset comprises 1,000 images, each annotated with both true and target class labels for targeted attack evaluation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25806" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.891, + 0.382 + ], + "angle": 0, + "content": "
SourceAttackTime (s)VGG-16RN-18RN-50DN-121XcepMB-v2EF-B0IR-v2Inc-v3Inc-v4Avg.
RN-50DI1.3763.256.498.974.75.527.829.05.69.09.437.9
RDI1.2366.370.798.281.913.246.447.217.129.923.349.4
RDI-SI5.9570.579.898.888.929.556.266.237.956.443.662.8
RDI-VT7.1668.878.798.282.527.954.556.132.845.837.958.3
RDI-Admix3.5574.280.798.786.820.959.456.126.742.734.158.0
RDI-SU2.2167.974.798.383.324.956.350.220.932.126.253.5
ODI4.3878.377.197.687.043.867.370.049.565.955.469.2
RDI-CFM1.3984.788.498.490.351.181.578.848.065.559.374.6
RDI-FTM1.5486.387.597.989.756.183.381.254.770.866.677.4
RDI-FTM-E2.9288.188.698.391.459.485.484.356.973.469.379.5
Inc-v3DI2.012.92.43.45.01.91.83.73.099.24.212.8
RDI1.763.53.84.07.03.13.05.96.398.77.114.2
RDI-SI8.114.05.25.711.06.34.68.211.698.812.116.8
RDI-VT10.505.98.99.413.27.45.99.812.398.714.718.6
RDI-Admix4.926.36.58.812.86.06.110.912.298.713.618.2
RDI-SU2.367.811.011.412.79.48.26.38.398.710.918.5
ODI6.4214.314.916.732.320.313.725.326.495.931.629.1
RDI-CFM2.1322.926.826.239.134.127.138.636.295.944.839.2
RDI-FTM2.3525.830.932.141.736.330.741.442.994.148.142.4
RDI-FTM-E4.3734.239.340.654.447.037.151.652.496.659.251.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.393, + 0.908, + 0.423 + ], + "angle": 0, + "content": "Table 1. Targeted attack success rates (%) against ten target models. All methods are combined with MI-TI. The best results are shown in bold, and the second best results are underlined. Time (s) denotes the average computation time required to attack a single image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.483, + 0.543 + ], + "angle": 0, + "content": "General settings. We adopt experimental configurations consistent with recent literature [2, 3, 49]. Specifically, we use an \\(\\ell_{\\infty}\\)-norm setting with \\(\\epsilon = 16/255\\). We set decay \\(\\mu = 1.0\\) and step size \\(\\eta = 2/255\\). We perform 300 iterations (\\(T = 300\\)) for iterative targeted attacks as per [3]. For optimization of adversarial examples, all methods employ the simple logit loss function [49]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Models. We evaluate our method using 15 pretrained DNNs, consisting of 10 CNN-based models and 5 transformer-based models. For CNN-based architectures, we employ VGG-16 [35], ResNet-18 (RN-18) [15], ResNet-50 (RN-50) [15], DenseNet-121 (DN-121) [17], Xception (Xcep) [5], MobileNet-v2 (MB-v2) [34], EfficientNet-B0 (EF-B0) [38], Inception ResNet-v2 (IRv2) [37], Inception-v3 (Inc-v3) [36], and Inception-v4 (Inc-v4) [37]. For Transformer-based architectures, we utilize Vision Transformer (ViT) [11], LeViT [14], ConViT [8], Twins [6], and Pooling-based ViT (PiT) [16]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Baseline attacks. We mainly compare our FTM against ten existing techniques: DI [46], RDI [50], MI [9], TI [10], SI [29], VT [39], Admix [40], ODI [2], SU [43], and CFM [3]. We primarily reference CFM, the current state-of-the-art method for targeted transfer-based attacks that requires no additional training dataset. MI and TI techniques apply across all attack methods, so we omit the 'MI-TI' prefix in our references. RDI serves as a common baseline technique in most cases. For DI, RDI, TI, SI, Admix, and ODI, we follow the configurations specified in CFM. For Admix, we also follow CFM to set the mixing weight to \\( w = 0.2 \\) and use \\( N = 3 \\) mixed images (i.e., \\( m_2 = 3 \\))" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.907, + 0.529 + ], + "angle": 0, + "content": "in [40]). Given Admix's computational intensity, we set the number of scale images in Admix to 1 (i.e., \\( m_{1} = 1 \\) in [40]) unless otherwise specified. We report the results for ODI instead of RDI-ODI, as ODI performs similarly to RDI-ODI [2, 3]. For SU and CFM, we use the original settings from their papers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.909, + 0.698 + ], + "angle": 0, + "content": "Settings for our FTM. We set the mixing ratio \\(\\beta\\) to 0.01. For FTM-E, we set the number of perturbed models to 2 for all experiments due to the computational cost. Following CFM, we avoid perturbing shallow layers because their low-level features vary significantly with input transforms and may cause excessive disturbance during attacks, as explained in CFM's paper. Specifically, we perturb layers where the feature map size is smaller than or equal to \\(\\frac{1}{16}\\) of the image size. For a pure ViT, we select the outputs of all fully-connected layers. We sample \\(\\alpha_{k,i}\\) from \\(\\mathcal{U}(0,0.75)\\), i.e., \\(\\alpha_{max} = 0.75\\). We set the probability \\(p\\) to 0.1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.711, + 0.719, + 0.727 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Transfer success rates. Table 1 shows the transfer success rates of various attack methods on ten black-box models. When using Inc-v3 as the source model, RDI-FTM-E achieves an average success rate of \\(51.2\\%\\), surpassing the previous best method CFM by \\(12.0\\%\\). The improvement is particularly notable for challenging target models like IR-v2 and Inc-v4, where RDI-FTM-E increases the success rates by \\(16.2\\%\\) and \\(14.4\\%\\), respectively. When using RN-50 as the source model, both RDI-FTM and RDI-FTM-E achieve higher success rates than CFM. The results show that the effective feature-level augmentation employed by FTM helps" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25807" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.326 + ], + "angle": 0, + "content": "
SourceAttackTime (s)ViTLeViTConViTTwinsPiTAvg.
RN-50DI1.370.23.50.41.51.71.5
RDI1.230.713.21.76.17.05.7
RDI-SI5.952.929.46.315.517.914.4
RDI-VT7.162.928.15.215.014.013.0
RDI-Admix3.551.322.52.58.58.48.6
RDI-SU2.210.816.92.36.97.86.9
ODI4.385.137.010.720.129.120.4
RDI-CFM1.394.346.18.925.224.721.8
RDI-FTM1.545.952.910.832.431.526.7
RDI-FTM-E2.926.858.613.635.234.929.8
Inc-v3DI2.010.10.30.00.00.10.1
RDI1.760.21.80.20.40.70.7
RDI-SI8.110.34.10.90.73.21.8
RDI-VT10.50.45.20.81.61.82.0
RDI-Admix4.920.14.10.61.41.41.5
RDI-SU2.360.22.00.31.31.01.0
ODI6.420.812.41.73.56.75.0
RDI-CFM2.132.121.93.26.111.68.9
RDI-FTM2.352.425.04.510.015.311.5
RDI-FTM-E4.373.832.76.812.720.415.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.336, + 0.483, + 0.392 + ], + "angle": 0, + "content": "Table 2. Targeted attack success rates \\((\\%)\\) against five transformer-based DNNs. All methods are combined with MI-TI. The best results are shown in bold, and the second best are underlined. Time (s) denotes the average time required to attack a single image." + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.408, + 0.289, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.409, + 0.477, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.507, + 0.466, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.531, + 0.483, + 0.561 + ], + "angle": 0, + "content": "Figure 3. Targeted attack success rates \\((\\%)\\) based on the number of iterations. The source model is RN-50 and Inc-v3, respectively." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.58, + 0.377, + 0.595 + ], + "angle": 0, + "content": "generate transferable adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Table 2 shows the transfer success rates of various attack methods on five transformer-based models. Our methods also achieve significant improvements over existing methods. For instance, when using RN-50 as the source model, RDI-FTM and RDI-FTM-E improve the targeted attack success rates of CFM by \\(4.9\\%\\) and \\(8.0\\%\\), respectively. The results suggest that our method can effectively improve the transferability of adversarial examples across different architectures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Figure 3 shows the targeted attack success rates based on the number of iterations. We evaluate our methods under different numbers of iterations using two experimental settings: RN-50 → Inc-v4 and Inc-v3 → IR-v2. The results show that FTM and FTM-E achieve stable performance across different iteration counts. For the RN-50 → Inc-v4 setting, SI exhibits superior performance at the beginning of the optimization process. However, as the number of iterations increases, FTM surpasses existing approaches. In the Inc-v3 → IR-v2 setting, FTM-E significantly outperforms other methods across the entire range of iterations." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.709, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.092, + 0.9, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.233, + 0.905, + 0.261 + ], + "angle": 0, + "content": "Figure 4. Attack success rates with different mixing ratio \\(\\beta\\). Each color represents the attack effectiveness on a specific target model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.281, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Computational overhead. We analyze the computational efficiency of different methods and report their costs in Table 2. We use an NVIDIA 3090 GPU to compute the time per image. The results demonstrate that our approach introduces reasonable computational overhead compared to CFM. Specifically, with RN-50 as the source model, RDI-FTM requires an average of 1.54 seconds per image, representing only a \\(10.8\\%\\) increase over CFM's 1.39 seconds. In contrast, alternative methods like SI, VT, and ODI impose substantially higher computational burdens. For instance, VT requires 7.16 seconds per image, and SI requires 5.95 seconds per image. We also include relevant analysis in Figure 1. The result shows that FTM achieves strong attack performance while maintaining low computational cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Combination with existing techniques. Table 3 presents transfer success rates of our method combined with ensemble-based method [30], SI, and Admix. Using the ensemble of RN-50 and Inc-v3 as the surrogate model, our RDI-SI-FTM-E achieves \\(74.5\\%\\) average success rate, surpassing existing methods. FTM notably enhances performance on transformer-based models. For instance, RDI-SI-FTM achieves \\(64.3\\%\\) on PiT, which is \\(5.0\\%\\) higher than RDI-SI-CFM \\((59.3\\%)\\) with comparable computational cost. These results demonstrate that FTM is a generic technique that combines effectively with existing attack methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.667, + 0.684 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.735 + ], + "angle": 0, + "content": "For all ablation experiments, we use RN-50 to generate adversarial examples and evaluate the average success rates on all models from Section 4.1, unless stated otherwise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.871 + ], + "angle": 0, + "content": "Stability with respect to \\(\\beta\\). Figure 4 shows the transfer success rates of FTM across different mixing ratios \\(\\beta\\). The success rates on five target models are plotted in distinct colors. When \\(\\beta\\) exceeds 0.04, the success rates decrease substantially across all target models. However, for \\(\\beta\\) values below 0.02, we observe stable attack performance across all models. The results suggest that FTM maintains robust performance within a reasonable range of \\(\\beta\\) values, showing its stability with respect to \\(\\beta\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Compare FTM and CFM using ensemble strategy. Figure 5 shows the attack performance of FTM and CFM" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "25808" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.089, + 0.878, + 0.251 + ], + "angle": 0, + "content": "
AttackRN-50 + Inc-v3 ⇒
Time (s)VGG-16RN-50*Inc-v3*DN-121IR-v2Inc-v4XcepViTLeViTConViTTwinsPiTAvg.
RDI3.2666.998.498.783.135.642.525.40.822.43.29.810.841.5
ODI7.1770.994.096.183.465.469.957.69.152.918.628.642.857.4
RDI-CFM4.3986.198.397.092.171.779.972.010.366.520.341.945.865.2
RDI-ODI-CFM7.7363.988.681.374.154.356.054.312.649.920.929.641.052.2
RDI-Admix-CFM13.3886.298.197.792.775.181.676.114.073.124.850.753.968.7
RDI-SI-CFM22.1387.598.197.392.778.081.678.020.876.332.955.059.371.5
RDI-FTM4.4787.597.896.792.475.481.774.014.571.025.249.354.468.3
RDI-FTM-E8.8188.298.197.193.476.783.477.118.474.930.153.458.070.7
RDI-Admix-FTM14.1086.096.796.592.278.383.877.320.175.829.955.758.570.9
RDI-SI-FTM22.6687.596.896.293.379.083.679.125.777.737.559.964.373.4
RDI-SI-FTM-E43.4387.897.396.793.979.183.578.227.379.242.962.965.674.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.261, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Table 3. Targeted attack success rates (%) of combination with different methods. The surrogate model is the ensemble of RN-50 and Inc-v3. All methods are combined with MI-TI. The best results are shown in bold, and the second best are underlined. Time (s) denotes the average time required to attack a single image." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.314, + 0.287, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.315, + 0.48, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.435, + 0.483, + 0.492 + ], + "angle": 0, + "content": "Figure 5. Attack performance and computational costs with different numbers of ensemble models. Left: Our FTM vs. CFM using ensemble strategy. Right: Computational costs of our FTM and RAP. -ens2 means using 2 perturbed models, and so on." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.504, + 0.287, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.506, + 0.48, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.698 + ], + "angle": 0, + "content": "Figure 6. Left: Attack success rates vs. \\( p \\) and \\( \\alpha_{max} \\). Right: Attack success rates vs. \\( p \\) with/without clean features; The blue line is the success rates when the clean features from CFM are integrated with our FTM, while the orange line represents only using our learnable perturbations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.868 + ], + "angle": 0, + "content": "using different numbers of perturbed models. On the left, our FTM consistently outperforms CFM across all ensemble sizes. Notably, FTM with only 2 perturbed models achieves better attack performance than an ensemble of 5 CFM-perturbed models, underscoring FTM's superior effectiveness. It is also worth mentioning that single FTM and CFM have comparable time complexity (Table 2). On the right, we compare our FTM with RAP [33], and the results show that FTM significantly outperforms RAP in both effectiveness and computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Analysis of probability \\( p \\) and ratio \\( \\alpha_{\\text{max}} \\). Figure 6 presents an analysis of FTM's attack success rates under" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.907, + 0.526 + ], + "angle": 0, + "content": "varying values of \\( p \\) and \\( \\alpha_{max} \\). The left plot shows that FTM achieves optimal performance with \\( p = 0.1 \\) and \\( \\alpha_{max} = 0.75 \\). FTM maintains competitive performance under various parameter configurations, such as \\( p, \\alpha_{max} = 0.15, 0.5 \\) or \\( 0.05, 1.0 \\), indicating its parameter stability. The right plot illustrates the impact of clean features on FTM. With clean features enabled, FTM's performance declines significantly as \\( p \\) increases from 0.1 to 1.0, while FTM without clean features shows stable performance across varying \\( p \\) values. This suggests that although clean features are crucial to FTM's performance, they also add instability to the attack process. Additionally, setting \\( p = 1.0 \\) (where all layers are updated) causes excessive perturbations, making the attack overly aggressive and reducing its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.633 + ], + "angle": 0, + "content": "The drop of the blue line in Figure 6 occurs because the clean accuracy of the perturbed surrogate model becomes extremely low \\((3.8\\%)\\) at \\(p = 1.0\\) when using clean feature mixup. Without clean feature mixup (orange line), the accuracy remains high at \\(65.3\\%\\). The low accuracy makes it impossible to generate effective adversarial examples, causing the attack success rate of the blue line to drop to near 0." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.644, + 0.634, + 0.66 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.907, + 0.866 + ], + "angle": 0, + "content": "In this work, we study the challenging problem of enhancing transferability in targeted adversarial attacks. Our investigation reveals that feature-level augmentation strategies show promising potential for boosting transferability effectively and efficiently. We introduce Feature Tuning Mixup (FTM), a novel method that incorporates both attack-specific feature perturbations and clean features in the feature space. To optimize the learnable perturbations efficiently, we develop a momentum-based stochastic update strategy. Comprehensive experiments on the ImageNet-compatible dataset demonstrate that FTM achieves substantial improvement over current state-of-the-art methods while maintaining low computational cost." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by HK RGC GRF under Grant PolyU 15201323." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25809" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.483, + 0.156 + ], + "angle": 0, + "content": "[1] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.158, + 0.483, + 0.212 + ], + "angle": 0, + "content": "[2] Junyoung Byun, Seungju Cho, Myung-Joon Kwon, Hee-Seon Kim, and Changick Kim. Improving the transferability of targeted adversarial examples through object-based diverse input. In CVPR, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.483, + 0.269 + ], + "angle": 0, + "content": "[3] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In CVPR, 2023. 1, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.482, + 0.311 + ], + "angle": 0, + "content": "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy (SP), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.482, + 0.339 + ], + "angle": 0, + "content": "[5] Francois Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[6] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In NeurIPS, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[7] Xuelong Dai, Kaisheng Liang, and Bin Xiao. Advdiff: Generating unrestricted adversarial examples using diffusion models. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[8] Stéphane d'Ascoli, Hugo Touvron, Matthew L. Leavitt, Ari S. Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.482, + 0.535 + ], + "angle": 0, + "content": "[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In CVPR, 2018. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.538, + 0.482, + 0.578 + ], + "angle": 0, + "content": "[10] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In CVPR, 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.581, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.664, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[12] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, and Dawn Song. Robust physical-world attacks on deep learning visual classification. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[13] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[14] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[16] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[17] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[18] Qian Huang, Isay Katsman, Zeqi Gu, Horace He, Serge J. Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[19] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[20] Nathan Inkawich, Wei Wen, Hai (Helen) Li, and Yiran Chen. Feature space perturbations yield more transferable adversarial examples. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[21] Nathan Inkawich, Kevin J. Liang, Lawrence Carin, and Yiran Chen. Transferable perturbations of deep feature distributions. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[22] Nathan Inkawich, Kevin J. Liang, Binghui Wang, Matthew Inkawich, Lawrence Carin, and Yiran Chen. Perturbing across the feature hierarchy to improve standard and strict blackbox attack transferability. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[23] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In ICLR (Workshop), 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[24] Maosen Li, Cheng Deng, Tengjiao Li, Junchi Yan, Xinbo Gao, and Heng Huang. Towards transferable targeted attack. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[25] Qizhang Li, Yiwen Guo, and Hao Chen. Yet another intermediate-level attack. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[26] Yanjie Li, Yiquan Li, Xuelong Dai, Songtao Guo, and Bin Xiao. Physical-world optical adversarial attacks on 3d face recognition. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[27] Yanjie Li, Kaisheng Liang, and Bin Xiao. UV-attack: Physical-world adversarial attacks on person detection via dynamic-NeRF-based UV mapping. In ICLR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[28] Kaisheng Liang and Bin Xiao. Style: Boosting the transferability of adversarial examples. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[29] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In ICLR, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.662, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[30] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In ICLR, 2017. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[32] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. On generating transferable targeted perturbations. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[33] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. In NeurIPS, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[34] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25810" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[37] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A. Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In AAAI, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[38] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[39] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In CVPR, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[40] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In ICCV, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[41] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[42] Zhibo Wang, Hongshan Yang, Yunhe Feng, Peng Sun, Hengchang Guo, Zhifei Zhang, and Kui Ren. Towards transferable targeted adversarial examples. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[43] Zhipeng Wei, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Enhancing the self-universality for transferable targeted attacks. In CVPR, 2023. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[44] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip connections matter: On the transferability of adversarial examples generated with resnets. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[45] Weibin Wu, Yuxin Su, Michael R Lyu, and Irwin King. Improving the transferability of adversarial samples with adversarial transformations. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[46] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[47] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Boosting transferability of targeted adversarial examples via hierarchical generative networks. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[48] Jianping Zhang, Weibin Wu, Jen-tse Huang, Yizhan Huang, Wenxuan Wang, Yuxin Su, and Michael R Lyu. Improving adversarial transferability via neuron attribution-based attacks. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[49] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. On success and simplicity: A second look at transferable targeted attacks. In NeurIPS, 2021. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[50] Junhua Zou, Zhisong Pan, Junyang Qiu, Xin Liu, Ting Rui, and Wei Li. Improving the transferability of adversarial examples with resized-diverse-inputs, diversity-ensemble and region fitting. In ECCV, 2020. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.873 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25811" + } + ] +] \ No newline at end of file diff --git a/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/5356bb72-29a6-442b-984c-adce3c009504_origin.pdf b/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/5356bb72-29a6-442b-984c-adce3c009504_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c6a56959e0bb3c211a66aa40de5a1fd9a0f9b710 --- /dev/null +++ b/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/5356bb72-29a6-442b-984c-adce3c009504_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9cf057877f8d683587d8b002fc11cfbec2fde7c27a5b6800d4e10cbd4225986 +size 516506 diff --git a/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/full.md b/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9e74241c028c2fb953195ed849479373436e68f6 --- /dev/null +++ b/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/full.md @@ -0,0 +1,335 @@ +# Improving Transferable Targeted Attacks with Feature Tuning Mixup + +Kaisheng Liang Xuelong Dai Yanjie Li Dong Wang Bin Xiao* Department of Computing, The Hong Kong Polytechnic University + +cskliang@comp.polyu.edu.hk b.xiao@polyu.edu.hk +elong.dai, yanjie.li, dong-comp.wang}@connect.polyu.hk + +# Abstract + +Deep neural networks (DNNs) exhibit vulnerability to adversarial examples that can transfer across different DNN models. A particularly challenging problem is developing transferable targeted attacks that can mislead DNN models into predicting specific target classes. While various methods have been proposed to enhance attack transferability, they often incur substantial computational costs while yielding limited improvements. Recent clean feature mixup methods use random clean features to perturb the feature space but lack optimization for disrupting adversarial examples, overlooking the advantages of attack-specific perturbations. In this paper, we propose Feature Tuning Mixup (FTM), a novel method that enhances targeted attack transferability by combining both random and optimized noises in the feature space. FTM introduces learnable feature perturbations and employs an efficient stochastic update strategy for optimization. These learnable perturbations facilitate the generation of more robust adversarial examples with improved transferability. We further demonstrate that attack performance can be enhanced through an ensemble of multiple FTM-perturbed surrogate models. Extensive experiments on the ImageNet-compatible dataset across various DNN models demonstrate that our method achieves significant improvements over state-of-the-art methods while maintaining low computational cost. + +# 1. Introduction + +Deep neural networks (DNNs) are highly vulnerable to adversarial attacks [13, 23, 31], raising critical security concerns. In image classification tasks, adversarial attacks can deceive DNNs by adding imperceptible perturbations to benign images, causing the models to produce erroneous predictions. More concerningly, research [9, 10, 46] has demonstrated that adversarial examples possess trans + +![](images/b5de2d0290ae04e5720290995e182080e231572a745fdf24abe67249a955ba9f.jpg) +Figure 1. Effectiveness and efficiency of targeted attacks. Average attack success rates on 14 black-box models, along with the computation time required to generate an adversarial example. Our methods demonstrate superior performance with low computational cost, surpassing state-of-the-art methods. + +ferability properties: adversarial examples generated using one white-box model can successfully fool other black-box models without requiring access to their architectures or parameters. This cross-model transferability of adversarial examples poses severe security threats to real-world AI systems. Therefore, studying the transferability of adversarial examples is crucial for enhancing the robustness of DNNs against such security vulnerabilities. + +Targeted transfer-based attacks aim to mislead models into predicting specific target classes, which remains a significant challenge. Due to the lack of knowledge about the target model's training data and architecture, such attacks tend to generate targeted adversarial examples that overfit to surrogate models and fail to transfer to target models. Various transferable targeted attacks [2, 43, 49] have been proposed but still leave much room for improvement. Notably, data augmentation methods [29, 39, 40] are commonly employed for targeted attacks. But they primarily operate in the image space, while augmentation methods at the feature level remain relatively unexplored. Recently, the Clean Feature Mixup (CFM) method [3] has made important progress by using clean features from random images to disrupt adversarial examples, leading to better targeted + +![](images/75bd2a809054658927670f6e784cc04f6e94caad810c2a60a151aaebf338570e.jpg) +Gradients from the Backward Process in Iteration i of the Attack +Figure 2. Overview of our FTM attack. In the forward pass, the learnable perturbations are added to the output features. Only a small portion of network layers are randomly selected for perturbation updates and clean feature mixup in each attack iteration. In the backward pass, the gradients of the selected perturbations and adversarial image are computed together for update. + +attack success rates (Figure 1). However, CFM uses random clean features without optimizing them specifically for disrupting adversarial examples. We hypothesize that carefully optimized perturbations would be more effective in disrupting adversarial examples, potentially yielding substantial improvements in attack transferability. + +In this work, we propose Feature Tuning Mixup (FTM), a novel attack method that optimizes attack-specific feature perturbations to enhance the transferability of targeted adversarial examples. Our method fine-tunes the intermediate features of surrogate models during the attack process by introducing learnable feature perturbations. These feature perturbations are optimized by mixing them with clean features from random images to counteract the adversarial effects of targeted adversarial examples. Through this mechanism, adversarial examples are gradually compelled to become more robust, allowing them to overcome these feature-level disturbances. To address the computational overhead of optimizing feature perturbations, we propose a stochastic update strategy. This strategy initializes the learnable perturbations using values from previous attack iterations and selectively optimizes only a small subset of feature perturbations in each iteration. Ultimately, FTM significantly improves the transferability of targeted adversarial examples while maintaining computational efficiency. + +We provide an overview of the proposed FTM in Figure 2. FTM first introduces learnable perturbations into the deep layers of the surrogate model. The learnable perturbations are added element-wise to the original output features before passing to the next layer. In each attack iteration, a small subset of layers is randomly selected and combined with random clean features. Finally, we compute gradients for the learnable perturbations and adversarial im + +age using only one forward and backward pass, similar to standard attacks. We update the selected learnable perturbations and adversarial image, while unselected perturbations remain unchanged. As a result, FTM fully leverages optimized attack-specific feature perturbations, prevents overfitting to surrogate models, and significantly enhances the transferability of targeted attacks. + +Our contributions are summarized as follows: + +- We revisit feature-level augmentation methods for transferable targeted attacks and identify that incorporating attack-specific feature perturbations, previously overlooked, can efficiently boost existing attacks. +- We introduce Feature Tuning Mixup (FTM), a novel approach that combines random and optimized feature perturbations to enhance targeted transfer-based attacks. Our method includes an efficient stochastic update strategy that maintains computational efficiency while improving attack transferability. +- We conduct extensive experiments across various source and target DNN models on the ImageNet-compatible dataset, demonstrating that our FTM achieves a significant improvement over state-of-the-art methods. + +# 2. Related Work + +Adversarial attacks [4, 12, 13, 19, 31] have been widely studied in the field of adversarial learning, including unrestricted attacks [7], physical attacks [26, 27], and black-box attacks [1, 9, 28]. Notably, black-box attacks can be categorized into query-based [1] and transfer-based attacks [9, 28]. The latter, which does not require access to the target DNN model's outputs, parameters or architecture, is more practical and is the focus of this paper. + +# 2.1. Non-targeted Transfer-based Attacks + +Current non-targeted transfer attacks are primarily based on the Iterative-Fast Gradient Sign Method (I-FGSM) [23]. This approach iteratively computes the gradient of the classification loss with respect to the input and updates adversarial examples in the direction that maximizes the loss. However, it often leads to overfitting on white-box models, making the adversarial examples less effective against other black-box models. Numerous studies have explored solutions to the overfitting issue to enhance adversarial transferability. The Momentum Iterative FGSM (MIFGSM) [9] incorporates a momentum-based optimization method during iterative attacks. Translation-Invariant (TI-FGSM) method [10] enhances the transferability of adversarial examples by achieving translation invariance through gradient smoothing. Data augmentation techniques are widely used to mitigate overfitting. The Diverse Inputs (DI) method [46] employs random resizing and padding with a probability $p$ to generate diverse input, while the Resized Diverse Inputs (RDI) method [50] fixes $p = 1$ and rescales DI-enhanced images back to their original size. The Scale-Invariant (SI) method [29] generates adversarial perturbations using multiple scale transformations, and Admix [40] extends SI by blending images from other labels into the input. Variance Tuning (VT) method [39] randomly selects data points around a given data point and computes their gradients to avoid overfitting. These methods operate in the image space. Feature-level approaches have also been studied to improve adversarial transferability. For instance, the Feature Importance-aware Attack (FIA) [41] identifies and perturbs important features in the intermediate layers of a surrogate model. The Skip Gradient Method (SGM) [44] utilizes more gradients from skip connections. Ensemble-based attacks [9, 30] utilize multiple surrogate models to generate adversarial examples. + +# 2.2. Targeted Transfer-based Attacks + +Targeted attacks are more challenging than non-targeted attacks because they not only aim to cause incorrect predictions by the models but also to mislead the models into predicting a specific target class. We can categorize targeted attacks into two types based on whether they require a training dataset. Previous works have utilized training datasets to generate auxiliary modules [21, 22] for transfer-based attacks or directly train a model to generate adversarial examples [32, 42, 47]. This paper primarily focuses on transfer-based attacks that do not require a training dataset. + +Targeted transfer-based attacks are widely studied in adversarial machine learning due to their practicality, as they do not require access to the target model's architecture and parameters, nor do they need training datasets. The Activation Attack (AA) [20] proposes using the distance between adversarial examples and targeted images in feature + +space as a loss function, but its attack effectiveness may depend on the selection of targeted images. Po+Trip [24] employs Poincare distance to adapt the gradients' magnitude and uses triplet loss to address insufficient inter-class distances, but it only works well on target models with similar architectures. Logit Attack (Logit) [49] discovers that a simple logit loss can generate better transferable adversarial examples than CE loss, significantly outperforming previous methods. Self-Universality (SU) method [43] finds that highly universal adversarial perturbations tend to be more transferable for targeted attacks and proposes optimizing the perturbations in different local regions of the image. The Object-based Diverse Input (ODI) method [2] is an efficient data augmentation in image space for targeted attacks using random 3D object and rendering. + +Recently, the Clean Feature Mixup (CFM) method [3] proposes a feature-level augmentation which achieves state-of-the-art performance in targeted transfer-based attacks. It stores clean features as random feature perturbations, which mitigates the overfitting to surrogate models. However, the proposed clean features in CFM lack specific optimization for targeted attacks, leaving significant room for improvement. We develop the Feature Tuning Mixup (FTM) method, which effectively and efficiently optimizes attack-specific feature perturbations and significantly improves transferable targeted attacks. + +Differences from intermediate-level attacks. This paper investigates feature-level augmentation for targeted attacks, which differs from existing intermediate-level attacks [18, 25, 41, 48]. Our FTM is based on the transfer-based attack framework [23] with logit loss [49], which is currently the most effective framework for targeted transfer-based attacks. In contrast, intermediate-level attacks use loss functions based on a specific intermediate layer, which is hard to apply to targeted attacks. For instance, FIA [41] and NAA [48] require identifying important features based on the true label, while targeted attacks focus on the target label instead of the true label. ILA [18] requires an existing adversarial example as guidance and cannot function as a standalone targeted attack. Therefore, we do not compare FTM with intermediate-level attacks in experiments. + +Comparison with adversarial learning methods. Our method learns feature perturbations in an adversarial manner without requiring additional training datasets. While some existing attacks [33, 45] also employ adversarial strategies, they operate in the image space and typically require access to training datasets. For instance, ATTA [45] improves transferability by training an adversarial transformation network to destroy adversarial examples. However, this approach demands substantial computational resources and training data. Like previous work [2, 3, 43, 49], we study targeted transfer attacks without extra data and do not compare with methods that need training data. RAP [33] + +utilizes adversarial learning to identify worst-case perturbations in the image space, aiming to locate regions with flat loss landscapes. However, this process requires numerous additional backpropagation steps, leading to prohibitively high computational costs. In contrast, we propose a novel approach that optimizes feature perturbations directly in the multi-layer feature space, an unexplored direction in previous research. Through this innovation, our method achieves superior practicality, effectiveness, and efficiency compared to existing methods. + +# 3. Methodology + +# 3.1. Preliminaries + +Problem formulation. For targeted transfer-based attacks, the attacker is assumed to have a white-box surrogate model $F$ , which is a pre-trained image classifier in this paper. We assume that the attacker has no additional training dataset. Given a benign image $x$ , a surrogate model $F$ and a target label $y_{t}$ , which is different from the true label $y$ of $x$ . The goal of the attack is to generate an adversarial example $x^{\mathrm{adv}}$ that is misclassified as $y_{t}$ by $F$ . The difference between $x$ and $x^{\mathrm{adv}}$ should be subtle, and following previous work, we use the $\ell_{\infty}$ -norm constraint. The optimization objective can be formulated as: + +$$ +\arg \min _ {x ^ {\mathrm {a d v}}} \mathcal {L} (F (x ^ {\mathrm {a d v}}), y _ {t}), \quad s. t. \quad \| x - x ^ {\mathrm {a d v}} \| _ {\infty} \leq \epsilon , \tag {1} +$$ + +where $\mathcal{L}$ is the adversarial loss, e.g., the cross-entropy loss or logit-based loss, and $\epsilon$ is the perturbation budget. + +MI-FGSM. To solve the problem above, existing methods usually adopt the iterative gradient-based technique, e.g., the MI-FGSM [9], formulated as follows: + +$$ +g _ {i + 1} = \nabla_ {x} \mathcal {L} \left(F \left(x _ {i} ^ {\mathrm {a d v}}\right), y _ {t}\right), \tag {2} +$$ + +$$ +\tilde {g} _ {i + 1} = \mu \cdot g _ {i} + g _ {i + 1} / \| g _ {i + 1} \| _ {1}, \tag {3} +$$ + +$$ +x _ {i + 1} ^ {\mathrm {a d v}} = x _ {i} ^ {\mathrm {a d v}} - \eta \cdot \operatorname {s i g n} \left(\tilde {g} _ {i + 1}\right), \tag {4} +$$ + +$$ +x _ {i + 1} ^ {\mathrm {a d v}} = \operatorname {C l i p} _ {x, \epsilon} \left(x _ {i + 1} ^ {\mathrm {a d v}}\right), \tag {5} +$$ + +where $g_{i}$ is the gradient of the loss function with respect to the input image $x$ in the $i$ -th iteration. $\mu$ is the momentum decay factor, and $\eta$ is the step size. $x_{i+1}^{\mathrm{adv}}$ is clipped within $\epsilon$ bound. Initially, we set $x_{0}^{\mathrm{adv}} = x$ and $g_{0} = 0$ . + +CFM. To improve the transferability of targeted attacks, the Clean Feature Mixup (CFM) method [3] leverages stored clean features to augment adversarial examples, achieving state-of-the-art performance among existing targeted transfer-based attacks. Specifically, it selects outputs from certain network layers as perturbation targets and uses output features from different benign images as feature perturbations. To integrate with MI-FGSM, the feature augmentation in CFM can be formulated as follows: + +$$ +z _ {k, i} ^ {\prime} = \left(1 - \alpha_ {k, i}\right) \odot z _ {k, i} + \alpha_ {k, i} \odot z _ {k, i} ^ {c}, \tag {6} +$$ + +where $\odot$ denotes element-wise multiplication, $z_{k,i}$ is the input feature to the $k$ -th layer in the $i$ -th attack iteration of MI-FGSM, $z_{k,i}^{c}$ is a random stored clean feature, and $\alpha_{k,i}$ is a vector representing channel-wise mixing ratios. The network layer $k$ is randomly selected from pre-defined candidate layers with probability $p$ . + +# 3.2. Analysis of CFM + +CFM is a powerful feature-level augmentation method for targeted attacks. To understand the effectiveness of CFM, we formulate a generalized optimization step for MI-FGSM based on feature-level augmentation as follows: + +$$ +g _ {i + 1} ^ {\prime} = \nabla_ {x} \mathcal {L} \left(F \left(x _ {i} ^ {\mathrm {a d v}}; \Delta z _ {i}\right), y _ {t}\right), \tag {7} +$$ + +$$ +\Delta z _ {i} = \left\{\Delta z _ {k, i}, k = 1, \dots , n \right\}, \tag {8} +$$ + +where $F(x_{i}^{\mathrm{adv}};\Delta z_{i})$ denotes the surrogate model $F$ with the perturbation $\Delta z_{i}$ added to the network outputs in the $i$ -th attack iteration. Equation (7) improves upon Equation (2) by incorporating feature-level perturbations, which is the key idea of CFM. Notably, the perturbation $\Delta z_{i}$ represents the collection of perturbations for all network layers. We use $\Delta z_{k,i}$ to denote the perturbation for the input of the $k$ -th layer, where $n$ is the maximum layer index of $F$ . Initially, we set $\Delta z_{i} = 0$ . For any layer $d$ that does not require perturbation, we simply set $\Delta z_{d,i} \equiv 0$ . + +CFM uses clean features to construct $\Delta z_{i}$ , and it claims that the rationale is that clean features can suppress the adversarial effect of $x_{i}^{\mathrm{adv}}$ [3]. We formulate this idea in the following equation: + +$$ +\mathcal {L} \left(F \left(x _ {i} ^ {\mathrm {a d v}}; \Delta z _ {i}\right), y _ {t}\right) > \mathcal {L} \left(F \left(x _ {i} ^ {\mathrm {a d v}}\right), y _ {t}\right), \tag {9} +$$ + +where $\Delta z_{i}$ is supposed to appropriately disrupt adversarial loss in Equation (1) for obtaining robust $x^{\mathrm{adv}}$ . + +However, clean features are attack-agnostic in nature, as they are not explicitly optimized for the objective in Equation (9). Although CFM demonstrates promising improvements, we argue that there remains unexplored potential in feature-level augmentation for targeted attacks. We believe that integrating attack-specific optimization with attack-agnostic clean features can create a more powerful feature-level augmentation strategy. + +# 3.3. Feature Tuning Mixup (FTM) + +We propose Feature Tuning Mixup (FTM), a novel method that uses attack-specific feature perturbations to enhance feature-level augmentation methods. We formulate the optimization problem for learning our feature perturbations and adversarial examples in each attack iteration as follows: + +$$ +\min _ {x _ {i} ^ {\mathrm {a d v}}} \max _ {\Delta z _ {i}} \mathcal {L} \left(F \left(x _ {i} ^ {\mathrm {a d v}}; \Delta z _ {i}\right), y _ {t}\right). \tag {10} +$$ + +However, the optimization problem above is challenging to solve from two perspectives: effectiveness and efficiency. + +Regarding efficiency, traditional adversarial training [31] often requires multiple steps to optimize $\Delta z_{i}$ in each attack iteration, resulting in substantial additional computational cost. As for effectiveness, since $\Delta z_{i}$ involves a joint problem across multiple iterations and network layers, inappropriate perturbations can damage the attack effect, which will be shown in our experiments. + +Momentum-based stochastic update. We propose a novel technique, momentum-based stochastic update, to optimize $\Delta z_{i}$ . We employ one-step optimization with using perturbations in previous iterations as momentum, i.e., $\Delta z_{i}$ will be initialized as $\Delta z_{i-1}$ in Equation (10). Momentum can effectively improve the optimization effectiveness without increasing additional optimization steps. In addition, as we only require one-step optimization, the gradient for $\Delta z_{i}$ can be jointly computed with the gradient for $x_{i}^{\mathrm{adv}}$ , which can further speed up the optimization process. To further improve the optimization efficiency, we use a stochastic update strategy, where only the randomly selected layers are updated in each attack iteration. For each selected layer $k$ , it has probability $p$ to be updated in the $i$ -th attack iteration, otherwise, we set $\Delta z_{k,i} = \Delta z_{k,i-1}$ . We illustrate the detailed optimization process for FTM in the following. + +The forward process for FTM in the $i$ -th attack iteration is formulated as follows: + +$$ +\bar {z} _ {k, i} = z _ {k, i} + \beta \| z _ {k, i} \| \cdot \frac {\Delta z _ {k , i}}{\| \Delta z _ {k , i} \| + \bar {\epsilon}}, \tag {11} +$$ + +$$ +z _ {k, i} ^ {\prime} = \left\{ \begin{array}{l l} (1 - \alpha_ {k, i}) \odot \bar {z} _ {k, i} + \alpha_ {k, i} \odot z _ {k, i} ^ {c}, & \tau_ {k} < p, \\ \bar {z} _ {k, i}, & \text {e l s e ,} \end{array} \right. \tag {12} +$$ + +where $\beta \| z_{k,i}\|$ is the custom perturbation budget for our feature perturbation $\Delta z_{k,i}$ . $\beta$ is a scaling factor, and $\bar{\epsilon}$ is a small constant. We mix the output feature $z_{k,i}$ and our attack-specific perturbation $\Delta z_{k,i}$ to obtain $\bar{z}_{k,i}$ . Then, we use clean features to augment $\bar{z}_{k,i}$ to get $z_{k,i}^{\prime}$ , which is the final perturbed feature that will be fed into the $k$ -th network layer. These perturbed features are computed layer by layer during the forward process. $z_{k,i}^{c}$ is a random clean feature, and $\alpha_{k,i}$ is a vector representing channel-wise mixing ratio. We randomly sample $\alpha_{k,i}$ from $\mathcal{U}(0,\alpha_{max})$ . In each iteration, we randomly sample $\tau_{k}$ from $\mathcal{U}(0,1)$ to determine whether to update the $k$ -th perturbation $\Delta z_{k,i}$ during the $i$ -th attack iteration. $p$ is the probability threshold. + +The backward process for FTM in the $i$ -th attack iteration is formulated as follows: + +$$ +\Delta z _ {i} ^ {\prime} = \left\{\left(\Delta z _ {k, i}, \tau_ {k}\right) \mid k = 1, \dots , n \right\}, \tag {13} +$$ + +$$ +g _ {i + 1} ^ {\prime}, g ^ {\Delta z ^ {\prime}} = \nabla_ {x _ {i} ^ {\mathrm {a d v}}, \Delta z _ {i} ^ {\prime}} \mathcal {L} \left(F \left(x _ {i} ^ {\mathrm {a d v}}; \Delta z _ {i} ^ {\prime}\right), y _ {t}\right), \tag {14} +$$ + +$$ +\Delta z _ {k, i + 1} = \left\{ \begin{array}{l l} \Delta z _ {k, i} + g _ {k} ^ {\Delta z ^ {\prime}}, & \tau_ {k} < p, \\ \Delta z _ {k, i}, & \text {o t h e r w i s e ,} \end{array} \right. \tag {15} +$$ + +Algorithm 1 Feature Tuning Mixup (FTM) Algorithm + +Input: Surrogate model $F$ , clean image $x$ , target label $y_{t}$ , iteration number $T$ , perturbation budget $\epsilon$ , input transformation $\phi(\cdot)$ (RDI, etc.), step size $\eta$ , decay factor $\mu$ + +Input: Mixing range $\alpha_{max}$ , probability $p$ , mixing ratio $\beta$ +Output: An adversarial example $x^{\mathrm{adv}}$ + +1: $x_0^{\mathrm{adv}} = x, g_0 = 0$ . +2: Initialize perturbation $\Delta z_0 = 0$ +3: for $i = 0 \rightarrow T - 1$ do: +4: Extend $F(x_{j}^{\mathrm{adv}})$ to $F(\phi (x_{j}^{\mathrm{adv}}))$ +5: Select layers to update and obtain $\Delta z_{i}^{\prime}$ from $\Delta z_{i}$ +6: Forward and backward pass to get $g_{i+1}'$ and $g^{\Delta z'}$ using Equations (12) and (14) +7: Obtain $\Delta z_{i + 1}$ using Equation (15) +8: Calculate $\tilde{g}_{i + 1} = \mu \cdot g_{i} + g_{i + 1}^{\prime} / \| g_{i + 1}^{\prime}\|_{1}$ +9: Update example $x_{i+1}^{\mathrm{adv}} = x_i^{\mathrm{adv}} - \eta \cdot \operatorname{sign}(\tilde{g}_{i+1})$ +10: Clip $x_{i+1}^{\mathrm{adv}}$ to be $\ell_{\infty}$ -bounded by $\epsilon$ +11: end for +12: $x^{\mathrm{adv}} = x_T^{\mathrm{adv}}$ +13: return $x^{\mathrm{adv}}$ + +where $\Delta z_{i}^{\prime}$ represents a set containing both perturbations $\Delta z_{k,i}$ and their corresponding random variables $\tau_{k}$ from the forward process, which determines whether each perturbation needs to be updated. $g_{k}^{\Delta z^{\prime}}$ denotes the gradient with respect to the $k$ -th perturbation $\Delta z_{k,i}$ . The threshold $p$ remains consistent with that used in the forward process. We summarize our FTM attack in Algorithm 1. + +Efficiency analysis. FTM does not require complex operations or additional forward and backward passes. Therefore, FTM is theoretically efficient on both CNNs and ViTs. In contrast, SI, VT, Admix, and SU require additional forward and backward passes, while ODI requires a computationally expensive input transformation. + +Ensemble of surrogate variants (FTM-E). Observing that FTM can efficiently perturb the surrogate model, we propose to further enhance the attack effectiveness through an ensemble strategy. We create multiple copies of the surrogate model and independently apply FTM to each copy, then ensemble them as a new surrogate model. We empirically show that utilizing an ensemble of two perturbed surrogate models can substantially enhance the attack performance while maintaining reasonable computational costs. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. Following prior research [2, 3, 43, 49], we conduct experiments on the ImageNet-compatible dataset from the NIPS 2017 adversarial attack challenge. This dataset comprises 1,000 images, each annotated with both true and target class labels for targeted attack evaluation. + +
SourceAttackTime (s)VGG-16RN-18RN-50DN-121XcepMB-v2EF-B0IR-v2Inc-v3Inc-v4Avg.
RN-50DI1.3763.256.498.974.75.527.829.05.69.09.437.9
RDI1.2366.370.798.281.913.246.447.217.129.923.349.4
RDI-SI5.9570.579.898.888.929.556.266.237.956.443.662.8
RDI-VT7.1668.878.798.282.527.954.556.132.845.837.958.3
RDI-Admix3.5574.280.798.786.820.959.456.126.742.734.158.0
RDI-SU2.2167.974.798.383.324.956.350.220.932.126.253.5
ODI4.3878.377.197.687.043.867.370.049.565.955.469.2
RDI-CFM1.3984.788.498.490.351.181.578.848.065.559.374.6
RDI-FTM1.5486.387.597.989.756.183.381.254.770.866.677.4
RDI-FTM-E2.9288.188.698.391.459.485.484.356.973.469.379.5
Inc-v3DI2.012.92.43.45.01.91.83.73.099.24.212.8
RDI1.763.53.84.07.03.13.05.96.398.77.114.2
RDI-SI8.114.05.25.711.06.34.68.211.698.812.116.8
RDI-VT10.505.98.99.413.27.45.99.812.398.714.718.6
RDI-Admix4.926.36.58.812.86.06.110.912.298.713.618.2
RDI-SU2.367.811.011.412.79.48.26.38.398.710.918.5
ODI6.4214.314.916.732.320.313.725.326.495.931.629.1
RDI-CFM2.1322.926.826.239.134.127.138.636.295.944.839.2
RDI-FTM2.3525.830.932.141.736.330.741.442.994.148.142.4
RDI-FTM-E4.3734.239.340.654.447.037.151.652.496.659.251.2
+ +Table 1. Targeted attack success rates (%) against ten target models. All methods are combined with MI-TI. The best results are shown in bold, and the second best results are underlined. Time (s) denotes the average computation time required to attack a single image. + +General settings. We adopt experimental configurations consistent with recent literature [2, 3, 49]. Specifically, we use an $\ell_{\infty}$ -norm setting with $\epsilon = 16/255$ . We set decay $\mu = 1.0$ and step size $\eta = 2/255$ . We perform 300 iterations ( $T = 300$ ) for iterative targeted attacks as per [3]. For optimization of adversarial examples, all methods employ the simple logit loss function [49]. + +Models. We evaluate our method using 15 pretrained DNNs, consisting of 10 CNN-based models and 5 transformer-based models. For CNN-based architectures, we employ VGG-16 [35], ResNet-18 (RN-18) [15], ResNet-50 (RN-50) [15], DenseNet-121 (DN-121) [17], Xception (Xcep) [5], MobileNet-v2 (MB-v2) [34], EfficientNet-B0 (EF-B0) [38], Inception ResNet-v2 (IRv2) [37], Inception-v3 (Inc-v3) [36], and Inception-v4 (Inc-v4) [37]. For Transformer-based architectures, we utilize Vision Transformer (ViT) [11], LeViT [14], ConViT [8], Twins [6], and Pooling-based ViT (PiT) [16]. + +Baseline attacks. We mainly compare our FTM against ten existing techniques: DI [46], RDI [50], MI [9], TI [10], SI [29], VT [39], Admix [40], ODI [2], SU [43], and CFM [3]. We primarily reference CFM, the current state-of-the-art method for targeted transfer-based attacks that requires no additional training dataset. MI and TI techniques apply across all attack methods, so we omit the 'MI-TI' prefix in our references. RDI serves as a common baseline technique in most cases. For DI, RDI, TI, SI, Admix, and ODI, we follow the configurations specified in CFM. For Admix, we also follow CFM to set the mixing weight to $w = 0.2$ and use $N = 3$ mixed images (i.e., $m_2 = 3$ ) + +in [40]). Given Admix's computational intensity, we set the number of scale images in Admix to 1 (i.e., $m_{1} = 1$ in [40]) unless otherwise specified. We report the results for ODI instead of RDI-ODI, as ODI performs similarly to RDI-ODI [2, 3]. For SU and CFM, we use the original settings from their papers. + +Settings for our FTM. We set the mixing ratio $\beta$ to 0.01. For FTM-E, we set the number of perturbed models to 2 for all experiments due to the computational cost. Following CFM, we avoid perturbing shallow layers because their low-level features vary significantly with input transforms and may cause excessive disturbance during attacks, as explained in CFM's paper. Specifically, we perturb layers where the feature map size is smaller than or equal to $\frac{1}{16}$ of the image size. For a pure ViT, we select the outputs of all fully-connected layers. We sample $\alpha_{k,i}$ from $\mathcal{U}(0,0.75)$ , i.e., $\alpha_{max} = 0.75$ . We set the probability $p$ to 0.1. + +# 4.2. Experimental Results + +Transfer success rates. Table 1 shows the transfer success rates of various attack methods on ten black-box models. When using Inc-v3 as the source model, RDI-FTM-E achieves an average success rate of $51.2\%$ , surpassing the previous best method CFM by $12.0\%$ . The improvement is particularly notable for challenging target models like IR-v2 and Inc-v4, where RDI-FTM-E increases the success rates by $16.2\%$ and $14.4\%$ , respectively. When using RN-50 as the source model, both RDI-FTM and RDI-FTM-E achieve higher success rates than CFM. The results show that the effective feature-level augmentation employed by FTM helps + +
SourceAttackTime (s)ViTLeViTConViTTwinsPiTAvg.
RN-50DI1.370.23.50.41.51.71.5
RDI1.230.713.21.76.17.05.7
RDI-SI5.952.929.46.315.517.914.4
RDI-VT7.162.928.15.215.014.013.0
RDI-Admix3.551.322.52.58.58.48.6
RDI-SU2.210.816.92.36.97.86.9
ODI4.385.137.010.720.129.120.4
RDI-CFM1.394.346.18.925.224.721.8
RDI-FTM1.545.952.910.832.431.526.7
RDI-FTM-E2.926.858.613.635.234.929.8
Inc-v3DI2.010.10.30.00.00.10.1
RDI1.760.21.80.20.40.70.7
RDI-SI8.110.34.10.90.73.21.8
RDI-VT10.50.45.20.81.61.82.0
RDI-Admix4.920.14.10.61.41.41.5
RDI-SU2.360.22.00.31.31.01.0
ODI6.420.812.41.73.56.75.0
RDI-CFM2.132.121.93.26.111.68.9
RDI-FTM2.352.425.04.510.015.311.5
RDI-FTM-E4.373.832.76.812.720.415.3
+ +Table 2. Targeted attack success rates $(\%)$ against five transformer-based DNNs. All methods are combined with MI-TI. The best results are shown in bold, and the second best are underlined. Time (s) denotes the average time required to attack a single image. + +![](images/7dcdebbf824c61a03c1b9e359164d9d4514c8b925267b7536ef5ebe8672be52e.jpg) +Figure 3. Targeted attack success rates $(\%)$ based on the number of iterations. The source model is RN-50 and Inc-v3, respectively. + +![](images/46cca3ae0e79edbf6d3e1d5057ac9b7cce7f1939e2bae319f364e9a265714cd1.jpg) + +![](images/3e8d221cfab650e8ce4d2b3f98383c5f07bcd38cdfd59d8d6b110fae4ccab8f7.jpg) + +generate transferable adversarial examples. + +Table 2 shows the transfer success rates of various attack methods on five transformer-based models. Our methods also achieve significant improvements over existing methods. For instance, when using RN-50 as the source model, RDI-FTM and RDI-FTM-E improve the targeted attack success rates of CFM by $4.9\%$ and $8.0\%$ , respectively. The results suggest that our method can effectively improve the transferability of adversarial examples across different architectures. + +Figure 3 shows the targeted attack success rates based on the number of iterations. We evaluate our methods under different numbers of iterations using two experimental settings: RN-50 → Inc-v4 and Inc-v3 → IR-v2. The results show that FTM and FTM-E achieve stable performance across different iteration counts. For the RN-50 → Inc-v4 setting, SI exhibits superior performance at the beginning of the optimization process. However, as the number of iterations increases, FTM surpasses existing approaches. In the Inc-v3 → IR-v2 setting, FTM-E significantly outperforms other methods across the entire range of iterations. + +![](images/ee2c2b1cc7aad49487060a0c2ff970f4a325062a6c184849f37f18c191a460a5.jpg) +Figure 4. Attack success rates with different mixing ratio $\beta$ . Each color represents the attack effectiveness on a specific target model. + +![](images/1c3990ce1590421259b6ab0f586688f6222a8952e5bb6e1eb2915d5e46bc7d74.jpg) + +Computational overhead. We analyze the computational efficiency of different methods and report their costs in Table 2. We use an NVIDIA 3090 GPU to compute the time per image. The results demonstrate that our approach introduces reasonable computational overhead compared to CFM. Specifically, with RN-50 as the source model, RDI-FTM requires an average of 1.54 seconds per image, representing only a $10.8\%$ increase over CFM's 1.39 seconds. In contrast, alternative methods like SI, VT, and ODI impose substantially higher computational burdens. For instance, VT requires 7.16 seconds per image, and SI requires 5.95 seconds per image. We also include relevant analysis in Figure 1. The result shows that FTM achieves strong attack performance while maintaining low computational cost. + +Combination with existing techniques. Table 3 presents transfer success rates of our method combined with ensemble-based method [30], SI, and Admix. Using the ensemble of RN-50 and Inc-v3 as the surrogate model, our RDI-SI-FTM-E achieves $74.5\%$ average success rate, surpassing existing methods. FTM notably enhances performance on transformer-based models. For instance, RDI-SI-FTM achieves $64.3\%$ on PiT, which is $5.0\%$ higher than RDI-SI-CFM $(59.3\%)$ with comparable computational cost. These results demonstrate that FTM is a generic technique that combines effectively with existing attack methods. + +# 4.3. Ablation Study + +For all ablation experiments, we use RN-50 to generate adversarial examples and evaluate the average success rates on all models from Section 4.1, unless stated otherwise. + +Stability with respect to $\beta$ . Figure 4 shows the transfer success rates of FTM across different mixing ratios $\beta$ . The success rates on five target models are plotted in distinct colors. When $\beta$ exceeds 0.04, the success rates decrease substantially across all target models. However, for $\beta$ values below 0.02, we observe stable attack performance across all models. The results suggest that FTM maintains robust performance within a reasonable range of $\beta$ values, showing its stability with respect to $\beta$ . + +Compare FTM and CFM using ensemble strategy. Figure 5 shows the attack performance of FTM and CFM + +
AttackRN-50 + Inc-v3 ⇒
Time (s)VGG-16RN-50*Inc-v3*DN-121IR-v2Inc-v4XcepViTLeViTConViTTwinsPiTAvg.
RDI3.2666.998.498.783.135.642.525.40.822.43.29.810.841.5
ODI7.1770.994.096.183.465.469.957.69.152.918.628.642.857.4
RDI-CFM4.3986.198.397.092.171.779.972.010.366.520.341.945.865.2
RDI-ODI-CFM7.7363.988.681.374.154.356.054.312.649.920.929.641.052.2
RDI-Admix-CFM13.3886.298.197.792.775.181.676.114.073.124.850.753.968.7
RDI-SI-CFM22.1387.598.197.392.778.081.678.020.876.332.955.059.371.5
RDI-FTM4.4787.597.896.792.475.481.774.014.571.025.249.354.468.3
RDI-FTM-E8.8188.298.197.193.476.783.477.118.474.930.153.458.070.7
RDI-Admix-FTM14.1086.096.796.592.278.383.877.320.175.829.955.758.570.9
RDI-SI-FTM22.6687.596.896.293.379.083.679.125.777.737.559.964.373.4
RDI-SI-FTM-E43.4387.897.396.793.979.183.578.227.379.242.962.965.674.5
+ +Table 3. Targeted attack success rates (%) of combination with different methods. The surrogate model is the ensemble of RN-50 and Inc-v3. All methods are combined with MI-TI. The best results are shown in bold, and the second best are underlined. Time (s) denotes the average time required to attack a single image. + +![](images/441d4b9e6024e07ed96a24419f7036373577becc02ba2ba83906194d0418ada6.jpg) +Figure 5. Attack performance and computational costs with different numbers of ensemble models. Left: Our FTM vs. CFM using ensemble strategy. Right: Computational costs of our FTM and RAP. -ens2 means using 2 perturbed models, and so on. + +![](images/725eedec4cde1e59b1a51d51f613603b65e030a1423161c064331a43b94cf52f.jpg) + +![](images/bf1e993cd0797ceeeda2af3be640f32a8884d641ec7f44753186e6331a743310.jpg) +Figure 6. Left: Attack success rates vs. $p$ and $\alpha_{max}$ . Right: Attack success rates vs. $p$ with/without clean features; The blue line is the success rates when the clean features from CFM are integrated with our FTM, while the orange line represents only using our learnable perturbations. + +![](images/0821458157fe6861f9f42b074e1ca1b9b4dda958babba4fc600b4d48cac09113.jpg) + +using different numbers of perturbed models. On the left, our FTM consistently outperforms CFM across all ensemble sizes. Notably, FTM with only 2 perturbed models achieves better attack performance than an ensemble of 5 CFM-perturbed models, underscoring FTM's superior effectiveness. It is also worth mentioning that single FTM and CFM have comparable time complexity (Table 2). On the right, we compare our FTM with RAP [33], and the results show that FTM significantly outperforms RAP in both effectiveness and computational efficiency. + +Analysis of probability $p$ and ratio $\alpha_{\text{max}}$ . Figure 6 presents an analysis of FTM's attack success rates under + +varying values of $p$ and $\alpha_{max}$ . The left plot shows that FTM achieves optimal performance with $p = 0.1$ and $\alpha_{max} = 0.75$ . FTM maintains competitive performance under various parameter configurations, such as $p, \alpha_{max} = 0.15, 0.5$ or $0.05, 1.0$ , indicating its parameter stability. The right plot illustrates the impact of clean features on FTM. With clean features enabled, FTM's performance declines significantly as $p$ increases from 0.1 to 1.0, while FTM without clean features shows stable performance across varying $p$ values. This suggests that although clean features are crucial to FTM's performance, they also add instability to the attack process. Additionally, setting $p = 1.0$ (where all layers are updated) causes excessive perturbations, making the attack overly aggressive and reducing its effectiveness. + +The drop of the blue line in Figure 6 occurs because the clean accuracy of the perturbed surrogate model becomes extremely low $(3.8\%)$ at $p = 1.0$ when using clean feature mixup. Without clean feature mixup (orange line), the accuracy remains high at $65.3\%$ . The low accuracy makes it impossible to generate effective adversarial examples, causing the attack success rate of the blue line to drop to near 0. + +# 5. Conclusion + +In this work, we study the challenging problem of enhancing transferability in targeted adversarial attacks. Our investigation reveals that feature-level augmentation strategies show promising potential for boosting transferability effectively and efficiently. We introduce Feature Tuning Mixup (FTM), a novel method that incorporates both attack-specific feature perturbations and clean features in the feature space. To optimize the learnable perturbations efficiently, we develop a momentum-based stochastic update strategy. Comprehensive experiments on the ImageNet-compatible dataset demonstrate that FTM achieves substantial improvement over current state-of-the-art methods while maintaining low computational cost. + +Acknowledgments. This work was supported in part by HK RGC GRF under Grant PolyU 15201323. + +# References + +[1] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. In ICLR, 2018. 2 +[2] Junyoung Byun, Seungju Cho, Myung-Joon Kwon, Hee-Seon Kim, and Changick Kim. Improving the transferability of targeted adversarial examples through object-based diverse input. In CVPR, 2022. 1, 3, 5, 6 +[3] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In CVPR, 2023. 1, 3, 4, 5, 6 +[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy (SP), 2017. 2 +[5] Francois Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, 2017. 6 +[6] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In NeurIPS, 2021. 6 +[7] Xuelong Dai, Kaisheng Liang, and Bin Xiao. Advdiff: Generating unrestricted adversarial examples using diffusion models. In ECCV, 2024. 2 +[8] Stéphane d'Ascoli, Hugo Touvron, Matthew L. Leavitt, Ari S. Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In ICML, 2021. 6 +[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In CVPR, 2018. 1, 2, 3, 4, 6 +[10] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In CVPR, 2019. 1, 3, 6 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 6 +[12] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, and Dawn Song. Robust physical-world attacks on deep learning visual classification. In CVPR, 2018. 2 +[13] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1, 2 +[14] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In ICCV, 2021. 6 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6 +[16] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, 2021. 6 + +[17] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 6 +[18] Qian Huang, Isay Katsman, Zeqi Gu, Horace He, Serge J. Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In ICCV, 2019. 3 +[19] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. In NeurIPS, 2019. 2 +[20] Nathan Inkawich, Wei Wen, Hai (Helen) Li, and Yiran Chen. Feature space perturbations yield more transferable adversarial examples. In CVPR, 2019. 3 +[21] Nathan Inkawich, Kevin J. Liang, Lawrence Carin, and Yiran Chen. Transferable perturbations of deep feature distributions. In ICLR, 2020. 3 +[22] Nathan Inkawich, Kevin J. Liang, Binghui Wang, Matthew Inkawich, Lawrence Carin, and Yiran Chen. Perturbing across the feature hierarchy to improve standard and strict blackbox attack transferability. In NeurIPS, 2020. 3 +[23] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In ICLR (Workshop), 2017. 1, 3 +[24] Maosen Li, Cheng Deng, Tengjiao Li, Junchi Yan, Xinbo Gao, and Heng Huang. Towards transferable targeted attack. In CVPR, 2020. 3 +[25] Qizhang Li, Yiwen Guo, and Hao Chen. Yet another intermediate-level attack. In ECCV, 2020. 3 +[26] Yanjie Li, Yiquan Li, Xuelong Dai, Songtao Guo, and Bin Xiao. Physical-world optical adversarial attacks on 3d face recognition. In CVPR, 2023. 2 +[27] Yanjie Li, Kaisheng Liang, and Bin Xiao. UV-attack: Physical-world adversarial attacks on person detection via dynamic-NeRF-based UV mapping. In ICLR, 2025. 2 +[28] Kaisheng Liang and Bin Xiao. Style: Boosting the transferability of adversarial examples. In CVPR, 2023. 2 +[29] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In ICLR, 2020. 1, 3, 6 +[30] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In ICLR, 2017. 3, 7 +[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 1, 2, 5 +[32] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. On generating transferable targeted perturbations. In ICCV, 2021. 3 +[33] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. In NeurIPS, 2022. 3, 8 +[34] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 6 + +[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 6 +[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 6 +[37] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A. Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In AAAI, 2017. 6 +[38] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 6 +[39] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In CVPR, 2021. 1, 3, 6 +[40] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In ICCV, 2021. 1, 3, 6 +[41] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In ICCV, 2021. 3 +[42] Zhibo Wang, Hongshan Yang, Yunhe Feng, Peng Sun, Hengchang Guo, Zhifei Zhang, and Kui Ren. Towards transferable targeted adversarial examples. In CVPR, 2023. 3 +[43] Zhipeng Wei, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Enhancing the self-universality for transferable targeted attacks. In CVPR, 2023. 1, 3, 5, 6 +[44] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip connections matter: On the transferability of adversarial examples generated with resnets. In ICLR, 2020. 3 +[45] Weibin Wu, Yuxin Su, Michael R Lyu, and Irwin King. Improving the transferability of adversarial samples with adversarial transformations. In CVPR, 2021. 3 +[46] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019. 1, 3, 6 +[47] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Boosting transferability of targeted adversarial examples via hierarchical generative networks. In ECCV, 2022. 3 +[48] Jianping Zhang, Weibin Wu, Jen-tse Huang, Yizhan Huang, Wenxuan Wang, Yuxin Su, and Michael R Lyu. Improving adversarial transferability via neuron attribution-based attacks. In CVPR, 2022. 3 +[49] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. On success and simplicity: A second look at transferable targeted attacks. In NeurIPS, 2021. 1, 3, 5, 6 +[50] Junhua Zou, Zhisong Pan, Junyang Qiu, Xin Liu, Ting Rui, and Wei Li. Improving the transferability of adversarial examples with resized-diverse-inputs, diversity-ensemble and region fitting. In ECCV, 2020. 3, 6 \ No newline at end of file diff --git a/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/images.zip b/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..378b187f3f1844f2b6e4679cb17b83756e34a94f --- /dev/null +++ b/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1b8c8edd461eba30e9272253e8273b008056e359b9c3e151a51ba129016c9d0 +size 535126 diff --git a/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/layout.json b/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dd138fe5d744509bfe77cb3d3dc41c14ecde549c --- /dev/null +++ b/2025/Improving Transferable Targeted Attacks with Feature Tuning Mixup/layout.json @@ -0,0 +1,10272 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "type": "text", + "content": "Improving Transferable Targeted Attacks with Feature Tuning Mixup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 143, + 506, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 143, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 108, + 143, + 506, + 173 + ], + "type": "text", + "content": "Kaisheng Liang Xuelong Dai Yanjie Li Dong Wang Bin Xiao* Department of Computing, The Hong Kong Polytechnic University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 169, + 174, + 459, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 174, + 459, + 199 + ], + "spans": [ + { + "bbox": [ + 169, + 174, + 459, + 199 + ], + "type": "text", + "content": "cskliang@comp.polyu.edu.hk b.xiao@polyu.edu.hk \nelong.dai, yanjie.li, dong-comp.wang}@connect.polyu.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 253, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 253, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 54, + 253, + 296, + 553 + ], + "type": "text", + "content": "Deep neural networks (DNNs) exhibit vulnerability to adversarial examples that can transfer across different DNN models. A particularly challenging problem is developing transferable targeted attacks that can mislead DNN models into predicting specific target classes. While various methods have been proposed to enhance attack transferability, they often incur substantial computational costs while yielding limited improvements. Recent clean feature mixup methods use random clean features to perturb the feature space but lack optimization for disrupting adversarial examples, overlooking the advantages of attack-specific perturbations. In this paper, we propose Feature Tuning Mixup (FTM), a novel method that enhances targeted attack transferability by combining both random and optimized noises in the feature space. FTM introduces learnable feature perturbations and employs an efficient stochastic update strategy for optimization. These learnable perturbations facilitate the generation of more robust adversarial examples with improved transferability. We further demonstrate that attack performance can be enhanced through an ensemble of multiple FTM-perturbed surrogate models. Extensive experiments on the ImageNet-compatible dataset across various DNN models demonstrate that our method achieves significant improvements over state-of-the-art methods while maintaining low computational cost." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 598, + 295, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 295, + 682 + ], + "type": "text", + "content": "Deep neural networks (DNNs) are highly vulnerable to adversarial attacks [13, 23, 31], raising critical security concerns. In image classification tasks, adversarial attacks can deceive DNNs by adding imperceptible perturbations to benign images, causing the models to produce erroneous predictions. More concerningly, research [9, 10, 46] has demonstrated that adversarial examples possess trans" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 230, + 539, + 350 + ], + "blocks": [ + { + "bbox": [ + 322, + 230, + 539, + 350 + ], + "lines": [ + { + "bbox": [ + 322, + 230, + 539, + 350 + ], + "spans": [ + { + "bbox": [ + 322, + 230, + 539, + 350 + ], + "type": "image", + "image_path": "b5de2d0290ae04e5720290995e182080e231572a745fdf24abe67249a955ba9f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 358, + 555, + 414 + ], + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 414 + ], + "type": "text", + "content": "Figure 1. Effectiveness and efficiency of targeted attacks. Average attack success rates on 14 black-box models, along with the computation time required to generate an adversarial example. Our methods demonstrate superior performance with low computational cost, surpassing state-of-the-art methods." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 434, + 555, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 530 + ], + "type": "text", + "content": "ferability properties: adversarial examples generated using one white-box model can successfully fool other black-box models without requiring access to their architectures or parameters. This cross-model transferability of adversarial examples poses severe security threats to real-world AI systems. Therefore, studying the transferability of adversarial examples is crucial for enhancing the robustness of DNNs against such security vulnerabilities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Targeted transfer-based attacks aim to mislead models into predicting specific target classes, which remains a significant challenge. Due to the lack of knowledge about the target model's training data and architecture, such attacks tend to generate targeted adversarial examples that overfit to surrogate models and fail to transfer to target models. Various transferable targeted attacks [2, 43, 49] have been proposed but still leave much room for improvement. Notably, data augmentation methods [29, 39, 40] are commonly employed for targeted attacks. But they primarily operate in the image space, while augmentation methods at the feature level remain relatively unexplored. Recently, the Clean Feature Mixup (CFM) method [3] has made important progress by using clean features from random images to disrupt adversarial examples, leading to better targeted" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 144, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 144, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 144, + 703 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 294, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 294, + 713 + ], + "type": "text", + "content": "1Our code is available at https://github.com/uhiu/features-tuning-mixup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25802" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 71, + 531, + 245 + ], + "blocks": [ + { + "bbox": [ + 81, + 71, + 531, + 245 + ], + "lines": [ + { + "bbox": [ + 81, + 71, + 531, + 245 + ], + "spans": [ + { + "bbox": [ + 81, + 71, + 531, + 245 + ], + "type": "image", + "image_path": "75bd2a809054658927670f6e784cc04f6e94caad810c2a60a151aaebf338570e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 246, + 448, + 254 + ], + "lines": [ + { + "bbox": [ + 272, + 246, + 448, + 254 + ], + "spans": [ + { + "bbox": [ + 272, + 246, + 448, + 254 + ], + "type": "text", + "content": "Gradients from the Backward Process in Iteration i of the Attack" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 264, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 54, + 264, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 54, + 264, + 555, + 299 + ], + "type": "text", + "content": "Figure 2. Overview of our FTM attack. In the forward pass, the learnable perturbations are added to the output features. Only a small portion of network layers are randomly selected for perturbation updates and clean feature mixup in each attack iteration. In the backward pass, the gradients of the selected perturbations and adversarial image are computed together for update." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 313, + 294, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 313, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 54, + 313, + 294, + 386 + ], + "type": "text", + "content": "attack success rates (Figure 1). However, CFM uses random clean features without optimizing them specifically for disrupting adversarial examples. We hypothesize that carefully optimized perturbations would be more effective in disrupting adversarial examples, potentially yielding substantial improvements in attack transferability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 388, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 615 + ], + "type": "text", + "content": "In this work, we propose Feature Tuning Mixup (FTM), a novel attack method that optimizes attack-specific feature perturbations to enhance the transferability of targeted adversarial examples. Our method fine-tunes the intermediate features of surrogate models during the attack process by introducing learnable feature perturbations. These feature perturbations are optimized by mixing them with clean features from random images to counteract the adversarial effects of targeted adversarial examples. Through this mechanism, adversarial examples are gradually compelled to become more robust, allowing them to overcome these feature-level disturbances. To address the computational overhead of optimizing feature perturbations, we propose a stochastic update strategy. This strategy initializes the learnable perturbations using values from previous attack iterations and selectively optimizes only a small subset of feature perturbations in each iteration. Ultimately, FTM significantly improves the transferability of targeted adversarial examples while maintaining computational efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "We provide an overview of the proposed FTM in Figure 2. FTM first introduces learnable perturbations into the deep layers of the surrogate model. The learnable perturbations are added element-wise to the original output features before passing to the next layer. In each attack iteration, a small subset of layers is randomly selected and combined with random clean features. Finally, we compute gradients for the learnable perturbations and adversarial im" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 313, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 554, + 396 + ], + "type": "text", + "content": "age using only one forward and backward pass, similar to standard attacks. We update the selected learnable perturbations and adversarial image, while unselected perturbations remain unchanged. As a result, FTM fully leverages optimized attack-specific feature perturbations, prevents overfitting to surrogate models, and significantly enhances the transferability of targeted attacks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 325, + 399, + 511, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 399, + 511, + 409 + ], + "spans": [ + { + "bbox": [ + 325, + 399, + 511, + 409 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 415, + 553, + 582 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 415, + 553, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 415, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 314, + 415, + 553, + 461 + ], + "type": "text", + "content": "- We revisit feature-level augmentation methods for transferable targeted attacks and identify that incorporating attack-specific feature perturbations, previously overlooked, can efficiently boost existing attacks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 462, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 462, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 314, + 462, + 553, + 533 + ], + "type": "text", + "content": "- We introduce Feature Tuning Mixup (FTM), a novel approach that combines random and optimized feature perturbations to enhance targeted transfer-based attacks. Our method includes an efficient stochastic update strategy that maintains computational efficiency while improving attack transferability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 534, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 534, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 314, + 534, + 553, + 582 + ], + "type": "text", + "content": "- We conduct extensive experiments across various source and target DNN models on the ImageNet-compatible dataset, demonstrating that our FTM achieves a significant improvement over state-of-the-art methods." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 596, + 400, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 596, + 400, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 596, + 400, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "text", + "content": "Adversarial attacks [4, 12, 13, 19, 31] have been widely studied in the field of adversarial learning, including unrestricted attacks [7], physical attacks [26, 27], and black-box attacks [1, 9, 28]. Notably, black-box attacks can be categorized into query-based [1] and transfer-based attacks [9, 28]. The latter, which does not require access to the target DNN model's outputs, parameters or architecture, is more practical and is the focus of this paper." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25803" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 253, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 253, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 253, + 85 + ], + "type": "text", + "content": "2.1. Non-targeted Transfer-based Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 497 + ], + "type": "text", + "content": "Current non-targeted transfer attacks are primarily based on the Iterative-Fast Gradient Sign Method (I-FGSM) [23]. This approach iteratively computes the gradient of the classification loss with respect to the input and updates adversarial examples in the direction that maximizes the loss. However, it often leads to overfitting on white-box models, making the adversarial examples less effective against other black-box models. Numerous studies have explored solutions to the overfitting issue to enhance adversarial transferability. The Momentum Iterative FGSM (MIFGSM) [9] incorporates a momentum-based optimization method during iterative attacks. Translation-Invariant (TI-FGSM) method [10] enhances the transferability of adversarial examples by achieving translation invariance through gradient smoothing. Data augmentation techniques are widely used to mitigate overfitting. The Diverse Inputs (DI) method [46] employs random resizing and padding with a probability " + }, + { + "bbox": [ + 55, + 89, + 297, + 497 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 89, + 297, + 497 + ], + "type": "text", + "content": " to generate diverse input, while the Resized Diverse Inputs (RDI) method [50] fixes " + }, + { + "bbox": [ + 55, + 89, + 297, + 497 + ], + "type": "inline_equation", + "content": "p = 1" + }, + { + "bbox": [ + 55, + 89, + 297, + 497 + ], + "type": "text", + "content": " and rescales DI-enhanced images back to their original size. The Scale-Invariant (SI) method [29] generates adversarial perturbations using multiple scale transformations, and Admix [40] extends SI by blending images from other labels into the input. Variance Tuning (VT) method [39] randomly selects data points around a given data point and computes their gradients to avoid overfitting. These methods operate in the image space. Feature-level approaches have also been studied to improve adversarial transferability. For instance, the Feature Importance-aware Attack (FIA) [41] identifies and perturbs important features in the intermediate layers of a surrogate model. The Skip Gradient Method (SGM) [44] utilizes more gradients from skip connections. Ensemble-based attacks [9, 30] utilize multiple surrogate models to generate adversarial examples." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 504, + 233, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 233, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 233, + 517 + ], + "type": "text", + "content": "2.2. Targeted Transfer-based Attacks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 521, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 641 + ], + "type": "text", + "content": "Targeted attacks are more challenging than non-targeted attacks because they not only aim to cause incorrect predictions by the models but also to mislead the models into predicting a specific target class. We can categorize targeted attacks into two types based on whether they require a training dataset. Previous works have utilized training datasets to generate auxiliary modules [21, 22] for transfer-based attacks or directly train a model to generate adversarial examples [32, 42, 47]. This paper primarily focuses on transfer-based attacks that do not require a training dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Targeted transfer-based attacks are widely studied in adversarial machine learning due to their practicality, as they do not require access to the target model's architecture and parameters, nor do they need training datasets. The Activation Attack (AA) [20] proposes using the distance between adversarial examples and targeted images in feature" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": "space as a loss function, but its attack effectiveness may depend on the selection of targeted images. Po+Trip [24] employs Poincare distance to adapt the gradients' magnitude and uses triplet loss to address insufficient inter-class distances, but it only works well on target models with similar architectures. Logit Attack (Logit) [49] discovers that a simple logit loss can generate better transferable adversarial examples than CE loss, significantly outperforming previous methods. Self-Universality (SU) method [43] finds that highly universal adversarial perturbations tend to be more transferable for targeted attacks and proposes optimizing the perturbations in different local regions of the image. The Object-based Diverse Input (ODI) method [2] is an efficient data augmentation in image space for targeted attacks using random 3D object and rendering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 254, + 555, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 386 + ], + "type": "text", + "content": "Recently, the Clean Feature Mixup (CFM) method [3] proposes a feature-level augmentation which achieves state-of-the-art performance in targeted transfer-based attacks. It stores clean features as random feature perturbations, which mitigates the overfitting to surrogate models. However, the proposed clean features in CFM lack specific optimization for targeted attacks, leaving significant room for improvement. We develop the Feature Tuning Mixup (FTM) method, which effectively and efficiently optimizes attack-specific feature perturbations and significantly improves transferable targeted attacks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 388, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 568 + ], + "type": "text", + "content": "Differences from intermediate-level attacks. This paper investigates feature-level augmentation for targeted attacks, which differs from existing intermediate-level attacks [18, 25, 41, 48]. Our FTM is based on the transfer-based attack framework [23] with logit loss [49], which is currently the most effective framework for targeted transfer-based attacks. In contrast, intermediate-level attacks use loss functions based on a specific intermediate layer, which is hard to apply to targeted attacks. For instance, FIA [41] and NAA [48] require identifying important features based on the true label, while targeted attacks focus on the target label instead of the true label. ILA [18] requires an existing adversarial example as guidance and cannot function as a standalone targeted attack. Therefore, we do not compare FTM with intermediate-level attacks in experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "content": "Comparison with adversarial learning methods. Our method learns feature perturbations in an adversarial manner without requiring additional training datasets. While some existing attacks [33, 45] also employ adversarial strategies, they operate in the image space and typically require access to training datasets. For instance, ATTA [45] improves transferability by training an adversarial transformation network to destroy adversarial examples. However, this approach demands substantial computational resources and training data. Like previous work [2, 3, 43, 49], we study targeted transfer attacks without extra data and do not compare with methods that need training data. RAP [33]" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25804" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "utilizes adversarial learning to identify worst-case perturbations in the image space, aiming to locate regions with flat loss landscapes. However, this process requires numerous additional backpropagation steps, leading to prohibitively high computational costs. In contrast, we propose a novel approach that optimizes feature perturbations directly in the multi-layer feature space, an unexplored direction in previous research. Through this innovation, our method achieves superior practicality, effectiveness, and efficiency compared to existing methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 138, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 138, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 138, + 216 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 221, + 141, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 141, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 141, + 232 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "content": "Problem formulation. For targeted transfer-based attacks, the attacker is assumed to have a white-box surrogate model " + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "content": ", which is a pre-trained image classifier in this paper. We assume that the attacker has no additional training dataset. Given a benign image " + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "content": ", a surrogate model " + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "content": " and a target label " + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "content": ", which is different from the true label " + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "content": ". The goal of the attack is to generate an adversarial example " + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "content": " that is misclassified as " + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "content": ". The difference between " + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "content": " should be subtle, and following previous work, we use the " + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 239, + 296, + 370 + ], + "type": "text", + "content": "-norm constraint. The optimization objective can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 377, + 296, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 377, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 67, + 377, + 296, + 396 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {x ^ {\\mathrm {a d v}}} \\mathcal {L} (F (x ^ {\\mathrm {a d v}}), y _ {t}), \\quad s. t. \\quad \\| x - x ^ {\\mathrm {a d v}} \\| _ {\\infty} \\leq \\epsilon , \\tag {1}", + "image_path": "3673325d04b57af87fc32107baa13df78aa0794aca992a5f1ce7a5c74355d66e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 402, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 295, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 402, + 295, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 402, + 295, + 426 + ], + "type": "text", + "content": " is the adversarial loss, e.g., the cross-entropy loss or logit-based loss, and " + }, + { + "bbox": [ + 55, + 402, + 295, + 426 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 402, + 295, + 426 + ], + "type": "text", + "content": " is the perturbation budget." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 462 + ], + "type": "text", + "content": "MI-FGSM. To solve the problem above, existing methods usually adopt the iterative gradient-based technique, e.g., the MI-FGSM [9], formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 469, + 295, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 469, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 113, + 469, + 295, + 483 + ], + "type": "interline_equation", + "content": "g _ {i + 1} = \\nabla_ {x} \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}\\right), y _ {t}\\right), \\tag {2}", + "image_path": "2be9ccdbafc0bb30ee705e9fc9236388c0514cb2b9d8f87535f4add938e093b7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 485, + 295, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 485, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 113, + 485, + 295, + 498 + ], + "type": "interline_equation", + "content": "\\tilde {g} _ {i + 1} = \\mu \\cdot g _ {i} + g _ {i + 1} / \\| g _ {i + 1} \\| _ {1}, \\tag {3}", + "image_path": "436de3d5ee5be3b21750b18d9a56bc5cfa3e3ca70ac192fe9c04771c03af0619.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 500, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 500, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 113, + 500, + 295, + 514 + ], + "type": "interline_equation", + "content": "x _ {i + 1} ^ {\\mathrm {a d v}} = x _ {i} ^ {\\mathrm {a d v}} - \\eta \\cdot \\operatorname {s i g n} \\left(\\tilde {g} _ {i + 1}\\right), \\tag {4}", + "image_path": "467b9e6917f1e198b95294f11c00b43eb481dcbcfbfecc00994af67585aef714.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 517, + 295, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 517, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 113, + 517, + 295, + 531 + ], + "type": "interline_equation", + "content": "x _ {i + 1} ^ {\\mathrm {a d v}} = \\operatorname {C l i p} _ {x, \\epsilon} \\left(x _ {i + 1} ^ {\\mathrm {a d v}}\\right), \\tag {5}", + "image_path": "466e4ed9e7a43bf97098cf32e5533b9d123a3578699a1ae582bc36e23b3cc0ac.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "text", + "content": " is the gradient of the loss function with respect to the input image " + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "text", + "content": "-th iteration. " + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "text", + "content": " is the momentum decay factor, and " + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "text", + "content": " is the step size. " + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "inline_equation", + "content": "x_{i+1}^{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "text", + "content": " is clipped within " + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "text", + "content": " bound. Initially, we set " + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "inline_equation", + "content": "x_{0}^{\\mathrm{adv}} = x" + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "inline_equation", + "content": "g_{0} = 0" + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 586, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 693 + ], + "type": "text", + "content": "CFM. To improve the transferability of targeted attacks, the Clean Feature Mixup (CFM) method [3] leverages stored clean features to augment adversarial examples, achieving state-of-the-art performance among existing targeted transfer-based attacks. Specifically, it selects outputs from certain network layers as perturbation targets and uses output features from different benign images as feature perturbations. To integrate with MI-FGSM, the feature augmentation in CFM can be formulated as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 96, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 96, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "z _ {k, i} ^ {\\prime} = \\left(1 - \\alpha_ {k, i}\\right) \\odot z _ {k, i} + \\alpha_ {k, i} \\odot z _ {k, i} ^ {c}, \\tag {6}", + "image_path": "4fa9b2ae0174958eee3ead1508d3007cdd5969ef2f84523dbfcfc05642a0fa72.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " denotes element-wise multiplication, " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "z_{k,i}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " is the input feature to the " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "-th layer in the " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "-th attack iteration of MI-FGSM, " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "z_{k,i}^{c}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " is a random stored clean feature, and " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\alpha_{k,i}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " is a vector representing channel-wise mixing ratios. The network layer " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " is randomly selected from pre-defined candidate layers with probability " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 150, + 416, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 416, + 162 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 416, + 162 + ], + "type": "text", + "content": "3.2. Analysis of CFM" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 167, + 554, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 554, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 554, + 216 + ], + "type": "text", + "content": "CFM is a powerful feature-level augmentation method for targeted attacks. To understand the effectiveness of CFM, we formulate a generalized optimization step for MI-FGSM based on feature-level augmentation as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 222, + 553, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 222, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 367, + 222, + 553, + 236 + ], + "type": "interline_equation", + "content": "g _ {i + 1} ^ {\\prime} = \\nabla_ {x} \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}; \\Delta z _ {i}\\right), y _ {t}\\right), \\tag {7}", + "image_path": "e745a4cc0f5bb5d9cdf0c2038076fa25bfefe211bbd09304895c5ed6da28d7b3.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 370, + 238, + 553, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 238, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 370, + 238, + 553, + 251 + ], + "type": "interline_equation", + "content": "\\Delta z _ {i} = \\left\\{\\Delta z _ {k, i}, k = 1, \\dots , n \\right\\}, \\tag {8}", + "image_path": "45832ea07b2a8773350f5a8a09df03c66fd85bc2e04292c589fd56b69ee6322f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "inline_equation", + "content": "F(x_{i}^{\\mathrm{adv}};\\Delta z_{i})" + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "content": " denotes the surrogate model " + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "content": " with the perturbation " + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\Delta z_{i}" + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "content": " added to the network outputs in the " + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "content": "-th attack iteration. Equation (7) improves upon Equation (2) by incorporating feature-level perturbations, which is the key idea of CFM. Notably, the perturbation " + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\Delta z_{i}" + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "content": " represents the collection of perturbations for all network layers. We use " + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\Delta z_{k,i}" + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "content": " to denote the perturbation for the input of the " + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "content": "-th layer, where " + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "content": " is the maximum layer index of " + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "content": ". Initially, we set " + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\Delta z_{i} = 0" + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "content": ". For any layer " + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "content": " that does not require perturbation, we simply set " + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\Delta z_{d,i} \\equiv 0" + }, + { + "bbox": [ + 313, + 258, + 554, + 378 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 378, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 426 + ], + "type": "text", + "content": "CFM uses clean features to construct " + }, + { + "bbox": [ + 313, + 378, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\Delta z_{i}" + }, + { + "bbox": [ + 313, + 378, + 554, + 426 + ], + "type": "text", + "content": ", and it claims that the rationale is that clean features can suppress the adversarial effect of " + }, + { + "bbox": [ + 313, + 378, + 554, + 426 + ], + "type": "inline_equation", + "content": "x_{i}^{\\mathrm{adv}}" + }, + { + "bbox": [ + 313, + 378, + 554, + 426 + ], + "type": "text", + "content": " [3]. We formulate this idea in the following equation:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 351, + 431, + 553, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 431, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 351, + 431, + 553, + 447 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}; \\Delta z _ {i}\\right), y _ {t}\\right) > \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}\\right), y _ {t}\\right), \\tag {9}", + "image_path": "b674c777ab24b74155993e7c99c460b84e35575a75ea3fe69210d1fcc314fe78.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 453, + 554, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 476 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 453, + 554, + 476 + ], + "type": "inline_equation", + "content": "\\Delta z_{i}" + }, + { + "bbox": [ + 313, + 453, + 554, + 476 + ], + "type": "text", + "content": " is supposed to appropriately disrupt adversarial loss in Equation (1) for obtaining robust " + }, + { + "bbox": [ + 313, + 453, + 554, + 476 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{adv}}" + }, + { + "bbox": [ + 313, + 453, + 554, + 476 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 477, + 554, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 554, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 554, + 572 + ], + "type": "text", + "content": "However, clean features are attack-agnostic in nature, as they are not explicitly optimized for the objective in Equation (9). Although CFM demonstrates promising improvements, we argue that there remains unexplored potential in feature-level augmentation for targeted attacks. We believe that integrating attack-specific optimization with attack-agnostic clean features can create a more powerful feature-level augmentation strategy." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 578, + 477, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 477, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 477, + 591 + ], + "type": "text", + "content": "3.3. Feature Tuning Mixup (FTM)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 596, + 554, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 554, + 656 + ], + "type": "text", + "content": "We propose Feature Tuning Mixup (FTM), a novel method that uses attack-specific feature perturbations to enhance feature-level augmentation methods. We formulate the optimization problem for learning our feature perturbations and adversarial examples in each attack iteration as follows:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 370, + 662, + 553, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 662, + 553, + 684 + ], + "spans": [ + { + "bbox": [ + 370, + 662, + 553, + 684 + ], + "type": "interline_equation", + "content": "\\min _ {x _ {i} ^ {\\mathrm {a d v}}} \\max _ {\\Delta z _ {i}} \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}; \\Delta z _ {i}\\right), y _ {t}\\right). \\tag {10}", + "image_path": "aee834cc65eeb6a644803114e22e33916792f09856f574a3af8009d1334b18f7.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "However, the optimization problem above is challenging to solve from two perspectives: effectiveness and efficiency." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25805" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "Regarding efficiency, traditional adversarial training [31] often requires multiple steps to optimize " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\Delta z_{i}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " in each attack iteration, resulting in substantial additional computational cost. As for effectiveness, since " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\Delta z_{i}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " involves a joint problem across multiple iterations and network layers, inappropriate perturbations can damage the attack effect, which will be shown in our experiments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "text", + "content": "Momentum-based stochastic update. We propose a novel technique, momentum-based stochastic update, to optimize " + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "inline_equation", + "content": "\\Delta z_{i}" + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "text", + "content": ". We employ one-step optimization with using perturbations in previous iterations as momentum, i.e., " + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "inline_equation", + "content": "\\Delta z_{i}" + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "text", + "content": " will be initialized as " + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "inline_equation", + "content": "\\Delta z_{i-1}" + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "text", + "content": " in Equation (10). Momentum can effectively improve the optimization effectiveness without increasing additional optimization steps. In addition, as we only require one-step optimization, the gradient for " + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "inline_equation", + "content": "\\Delta z_{i}" + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "text", + "content": " can be jointly computed with the gradient for " + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "inline_equation", + "content": "x_{i}^{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "text", + "content": ", which can further speed up the optimization process. To further improve the optimization efficiency, we use a stochastic update strategy, where only the randomly selected layers are updated in each attack iteration. For each selected layer " + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "text", + "content": ", it has probability " + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "text", + "content": " to be updated in the " + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "text", + "content": "-th attack iteration, otherwise, we set " + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "inline_equation", + "content": "\\Delta z_{k,i} = \\Delta z_{k,i-1}" + }, + { + "bbox": [ + 55, + 157, + 295, + 348 + ], + "type": "text", + "content": ". We illustrate the detailed optimization process for FTM in the following." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 350, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 373 + ], + "type": "text", + "content": "The forward process for FTM in the " + }, + { + "bbox": [ + 55, + 350, + 295, + 373 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 350, + 295, + 373 + ], + "type": "text", + "content": "-th attack iteration is formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 381, + 295, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 381, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 63, + 381, + 295, + 407 + ], + "type": "interline_equation", + "content": "\\bar {z} _ {k, i} = z _ {k, i} + \\beta \\| z _ {k, i} \\| \\cdot \\frac {\\Delta z _ {k , i}}{\\| \\Delta z _ {k , i} \\| + \\bar {\\epsilon}}, \\tag {11}", + "image_path": "ca6be8875bae4addf910dd68b28fa39769c87488e8439f68737f4f16599e60f1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 410, + 294, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 410, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 63, + 410, + 294, + 442 + ], + "type": "interline_equation", + "content": "z _ {k, i} ^ {\\prime} = \\left\\{ \\begin{array}{l l} (1 - \\alpha_ {k, i}) \\odot \\bar {z} _ {k, i} + \\alpha_ {k, i} \\odot z _ {k, i} ^ {c}, & \\tau_ {k} < p, \\\\ \\bar {z} _ {k, i}, & \\text {e l s e ,} \\end{array} \\right. \\tag {12}", + "image_path": "c9f5a3a5630cc194184563236fe37b071ec78bbb903c45378bd2e323aa981329.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\beta \\| z_{k,i}\\|" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": " is the custom perturbation budget for our feature perturbation " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\Delta z_{k,i}" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": " is a scaling factor, and " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\bar{\\epsilon}" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": " is a small constant. We mix the output feature " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "z_{k,i}" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": " and our attack-specific perturbation " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\Delta z_{k,i}" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\bar{z}_{k,i}" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": ". Then, we use clean features to augment " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\bar{z}_{k,i}" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "z_{k,i}^{\\prime}" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": ", which is the final perturbed feature that will be fed into the " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": "-th network layer. These perturbed features are computed layer by layer during the forward process. " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "z_{k,i}^{c}" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": " is a random clean feature, and " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\alpha_{k,i}" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": " is a vector representing channel-wise mixing ratio. We randomly sample " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\alpha_{k,i}" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(0,\\alpha_{max})" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": ". In each iteration, we randomly sample " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\tau_{k}" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(0,1)" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": " to determine whether to update the " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": "-th perturbation " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\Delta z_{k,i}" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": " during the " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": "-th attack iteration. " + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 453, + 295, + 608 + ], + "type": "text", + "content": " is the probability threshold." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 609, + 295, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 632 + ], + "type": "text", + "content": "The backward process for FTM in the " + }, + { + "bbox": [ + 55, + 609, + 295, + 632 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 609, + 295, + 632 + ], + "type": "text", + "content": "-th attack iteration is formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 101, + 644, + 295, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 644, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 101, + 644, + 295, + 658 + ], + "type": "interline_equation", + "content": "\\Delta z _ {i} ^ {\\prime} = \\left\\{\\left(\\Delta z _ {k, i}, \\tau_ {k}\\right) \\mid k = 1, \\dots , n \\right\\}, \\tag {13}", + "image_path": "fb724c49e8edf0a8d3dfa9be4b5d59819c4d75d7c2d574f1266ec12e0863aa7b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 661, + 294, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 661, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 76, + 661, + 294, + 677 + ], + "type": "interline_equation", + "content": "g _ {i + 1} ^ {\\prime}, g ^ {\\Delta z ^ {\\prime}} = \\nabla_ {x _ {i} ^ {\\mathrm {a d v}}, \\Delta z _ {i} ^ {\\prime}} \\mathcal {L} \\left(F \\left(x _ {i} ^ {\\mathrm {a d v}}; \\Delta z _ {i} ^ {\\prime}\\right), y _ {t}\\right), \\tag {14}", + "image_path": "f7fa19a61341a771db78ada69320e6b07aa1f19a28f3504e0c62d624662b288a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 84, + 679, + 295, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 679, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 84, + 679, + 295, + 711 + ], + "type": "interline_equation", + "content": "\\Delta z _ {k, i + 1} = \\left\\{ \\begin{array}{l l} \\Delta z _ {k, i} + g _ {k} ^ {\\Delta z ^ {\\prime}}, & \\tau_ {k} < p, \\\\ \\Delta z _ {k, i}, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {15}", + "image_path": "f6415c6a825011b293e0494c7c3e7f3c29995c1e8dd667fff08e8d70799ddaa9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 72, + 536, + 84 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 536, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 536, + 84 + ], + "type": "text", + "content": "Algorithm 1 Feature Tuning Mixup (FTM) Algorithm" + } + ] + } + ], + "index": 10, + "type": "text" + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "spans": [ + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "text", + "content": "Input: Surrogate model " + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "text", + "content": ", clean image " + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "text", + "content": ", target label " + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "text", + "content": ", iteration number " + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "text", + "content": ", perturbation budget " + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "text", + "content": ", input transformation " + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "text", + "content": " (RDI, etc.), step size " + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "text", + "content": ", decay factor " + }, + { + "bbox": [ + 314, + 87, + 553, + 123 + ], + "type": "inline_equation", + "content": "\\mu" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 124, + 547, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 124, + 547, + 147 + ], + "spans": [ + { + "bbox": [ + 315, + 124, + 547, + 147 + ], + "type": "text", + "content": "Input: Mixing range " + }, + { + "bbox": [ + 315, + 124, + 547, + 147 + ], + "type": "inline_equation", + "content": "\\alpha_{max}" + }, + { + "bbox": [ + 315, + 124, + 547, + 147 + ], + "type": "text", + "content": ", probability " + }, + { + "bbox": [ + 315, + 124, + 547, + 147 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 315, + 124, + 547, + 147 + ], + "type": "text", + "content": ", mixing ratio " + }, + { + "bbox": [ + 315, + 124, + 547, + 147 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 315, + 124, + 547, + 147 + ], + "type": "text", + "content": " \nOutput: An adversarial example " + }, + { + "bbox": [ + 315, + 124, + 547, + 147 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{adv}}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 147, + 552, + 314 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 321, + 147, + 403, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 147, + 403, + 159 + ], + "spans": [ + { + "bbox": [ + 321, + 147, + 403, + 159 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 321, + 147, + 403, + 159 + ], + "type": "inline_equation", + "content": "x_0^{\\mathrm{adv}} = x, g_0 = 0" + }, + { + "bbox": [ + 321, + 147, + 403, + 159 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 160, + 459, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 160, + 459, + 171 + ], + "spans": [ + { + "bbox": [ + 321, + 160, + 459, + 171 + ], + "type": "text", + "content": "2: Initialize perturbation " + }, + { + "bbox": [ + 321, + 160, + 459, + 171 + ], + "type": "inline_equation", + "content": "\\Delta z_0 = 0" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 171, + 426, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 171, + 426, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 171, + 426, + 182 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 321, + 171, + 426, + 182 + ], + "type": "inline_equation", + "content": "i = 0 \\rightarrow T - 1" + }, + { + "bbox": [ + 321, + 171, + 426, + 182 + ], + "type": "text", + "content": " do:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 182, + 469, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 182, + 469, + 194 + ], + "spans": [ + { + "bbox": [ + 321, + 182, + 469, + 194 + ], + "type": "text", + "content": "4: Extend " + }, + { + "bbox": [ + 321, + 182, + 469, + 194 + ], + "type": "inline_equation", + "content": "F(x_{j}^{\\mathrm{adv}})" + }, + { + "bbox": [ + 321, + 182, + 469, + 194 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 321, + 182, + 469, + 194 + ], + "type": "inline_equation", + "content": "F(\\phi (x_{j}^{\\mathrm{adv}}))" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 195, + 542, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 195, + 542, + 208 + ], + "spans": [ + { + "bbox": [ + 321, + 195, + 542, + 208 + ], + "type": "text", + "content": "5: Select layers to update and obtain " + }, + { + "bbox": [ + 321, + 195, + 542, + 208 + ], + "type": "inline_equation", + "content": "\\Delta z_{i}^{\\prime}" + }, + { + "bbox": [ + 321, + 195, + 542, + 208 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 321, + 195, + 542, + 208 + ], + "type": "inline_equation", + "content": "\\Delta z_{i}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 208, + 552, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 208, + 552, + 232 + ], + "spans": [ + { + "bbox": [ + 321, + 208, + 552, + 232 + ], + "type": "text", + "content": "6: Forward and backward pass to get " + }, + { + "bbox": [ + 321, + 208, + 552, + 232 + ], + "type": "inline_equation", + "content": "g_{i+1}'" + }, + { + "bbox": [ + 321, + 208, + 552, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 321, + 208, + 552, + 232 + ], + "type": "inline_equation", + "content": "g^{\\Delta z'}" + }, + { + "bbox": [ + 321, + 208, + 552, + 232 + ], + "type": "text", + "content": " using Equations (12) and (14)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 232, + 487, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 232, + 487, + 243 + ], + "spans": [ + { + "bbox": [ + 321, + 232, + 487, + 243 + ], + "type": "text", + "content": "7: Obtain " + }, + { + "bbox": [ + 321, + 232, + 487, + 243 + ], + "type": "inline_equation", + "content": "\\Delta z_{i + 1}" + }, + { + "bbox": [ + 321, + 232, + 487, + 243 + ], + "type": "text", + "content": " using Equation (15)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 244, + 509, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 244, + 509, + 256 + ], + "spans": [ + { + "bbox": [ + 321, + 244, + 509, + 256 + ], + "type": "text", + "content": "8: Calculate " + }, + { + "bbox": [ + 321, + 244, + 509, + 256 + ], + "type": "inline_equation", + "content": "\\tilde{g}_{i + 1} = \\mu \\cdot g_{i} + g_{i + 1}^{\\prime} / \\| g_{i + 1}^{\\prime}\\|_{1}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 256, + 531, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 256, + 531, + 268 + ], + "spans": [ + { + "bbox": [ + 321, + 256, + 531, + 268 + ], + "type": "text", + "content": "9: Update example " + }, + { + "bbox": [ + 321, + 256, + 531, + 268 + ], + "type": "inline_equation", + "content": "x_{i+1}^{\\mathrm{adv}} = x_i^{\\mathrm{adv}} - \\eta \\cdot \\operatorname{sign}(\\tilde{g}_{i+1})" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 269, + 482, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 269, + 482, + 280 + ], + "spans": [ + { + "bbox": [ + 318, + 269, + 482, + 280 + ], + "type": "text", + "content": "10: Clip " + }, + { + "bbox": [ + 318, + 269, + 482, + 280 + ], + "type": "inline_equation", + "content": "x_{i+1}^{\\mathrm{adv}}" + }, + { + "bbox": [ + 318, + 269, + 482, + 280 + ], + "type": "text", + "content": " to be " + }, + { + "bbox": [ + 318, + 269, + 482, + 280 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 318, + 269, + 482, + 280 + ], + "type": "text", + "content": "-bounded by " + }, + { + "bbox": [ + 318, + 269, + 482, + 280 + ], + "type": "inline_equation", + "content": "\\epsilon" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 281, + 365, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 281, + 365, + 290 + ], + "spans": [ + { + "bbox": [ + 318, + 281, + 365, + 290 + ], + "type": "text", + "content": "11: end for" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 291, + 380, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 291, + 380, + 304 + ], + "spans": [ + { + "bbox": [ + 318, + 291, + 380, + 304 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 318, + 291, + 380, + 304 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{adv}} = x_T^{\\mathrm{adv}}" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 304, + 380, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 304, + 380, + 314 + ], + "spans": [ + { + "bbox": [ + 318, + 304, + 380, + 314 + ], + "type": "text", + "content": "13: return " + }, + { + "bbox": [ + 318, + 304, + 380, + 314 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{adv}}" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "inline_equation", + "content": "\\Delta z_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "text", + "content": " represents a set containing both perturbations " + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "inline_equation", + "content": "\\Delta z_{k,i}" + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "text", + "content": " and their corresponding random variables " + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "inline_equation", + "content": "\\tau_{k}" + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "text", + "content": " from the forward process, which determines whether each perturbation needs to be updated. " + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "inline_equation", + "content": "g_{k}^{\\Delta z^{\\prime}}" + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "text", + "content": " denotes the gradient with respect to the " + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "text", + "content": "-th perturbation " + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "inline_equation", + "content": "\\Delta z_{k,i}" + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "text", + "content": ". The threshold " + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 339, + 553, + 422 + ], + "type": "text", + "content": " remains consistent with that used in the forward process. We summarize our FTM attack in Algorithm 1." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 423, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 553, + 495 + ], + "type": "text", + "content": "Efficiency analysis. FTM does not require complex operations or additional forward and backward passes. Therefore, FTM is theoretically efficient on both CNNs and ViTs. In contrast, SI, VT, Admix, and SU require additional forward and backward passes, while ODI requires a computationally expensive input transformation." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 495, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 604 + ], + "type": "text", + "content": "Ensemble of surrogate variants (FTM-E). Observing that FTM can efficiently perturb the surrogate model, we propose to further enhance the attack effectiveness through an ensemble strategy. We create multiple copies of the surrogate model and independently apply FTM to each copy, then ensemble them as a new surrogate model. We empirically show that utilizing an ensemble of two perturbed surrogate models can substantially enhance the attack performance while maintaining reasonable computational costs." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 616, + 395, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 395, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 395, + 629 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 635, + 441, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 441, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 441, + 649 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Datasets. Following prior research [2, 3, 43, 49], we conduct experiments on the ImageNet-compatible dataset from the NIPS 2017 adversarial attack challenge. This dataset comprises 1,000 images, each annotated with both true and target class labels for targeted attack evaluation." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25806" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 545, + 302 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 545, + 302 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 545, + 302 + ], + "type": "table", + "html": "
SourceAttackTime (s)VGG-16RN-18RN-50DN-121XcepMB-v2EF-B0IR-v2Inc-v3Inc-v4Avg.
RN-50DI1.3763.256.498.974.75.527.829.05.69.09.437.9
RDI1.2366.370.798.281.913.246.447.217.129.923.349.4
RDI-SI5.9570.579.898.888.929.556.266.237.956.443.662.8
RDI-VT7.1668.878.798.282.527.954.556.132.845.837.958.3
RDI-Admix3.5574.280.798.786.820.959.456.126.742.734.158.0
RDI-SU2.2167.974.798.383.324.956.350.220.932.126.253.5
ODI4.3878.377.197.687.043.867.370.049.565.955.469.2
RDI-CFM1.3984.788.498.490.351.181.578.848.065.559.374.6
RDI-FTM1.5486.387.597.989.756.183.381.254.770.866.677.4
RDI-FTM-E2.9288.188.698.391.459.485.484.356.973.469.379.5
Inc-v3DI2.012.92.43.45.01.91.83.73.099.24.212.8
RDI1.763.53.84.07.03.13.05.96.398.77.114.2
RDI-SI8.114.05.25.711.06.34.68.211.698.812.116.8
RDI-VT10.505.98.99.413.27.45.99.812.398.714.718.6
RDI-Admix4.926.36.58.812.86.06.110.912.298.713.618.2
RDI-SU2.367.811.011.412.79.48.26.38.398.710.918.5
ODI6.4214.314.916.732.320.313.725.326.495.931.629.1
RDI-CFM2.1322.926.826.239.134.127.138.636.295.944.839.2
RDI-FTM2.3525.830.932.141.736.330.741.442.994.148.142.4
RDI-FTM-E4.3734.239.340.654.447.037.151.652.496.659.251.2
", + "image_path": "554becbfc9d188fa7686b9254f0c70a477310905e7cecb6f467be3544860b77a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 311, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 311, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 555, + 335 + ], + "type": "text", + "content": "Table 1. Targeted attack success rates (%) against ten target models. All methods are combined with MI-TI. The best results are shown in bold, and the second best results are underlined. Time (s) denotes the average computation time required to attack a single image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "text", + "content": "General settings. We adopt experimental configurations consistent with recent literature [2, 3, 49]. Specifically, we use an " + }, + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "text", + "content": "-norm setting with " + }, + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "inline_equation", + "content": "\\epsilon = 16/255" + }, + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "text", + "content": ". We set decay " + }, + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "inline_equation", + "content": "\\mu = 1.0" + }, + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "text", + "content": " and step size " + }, + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "inline_equation", + "content": "\\eta = 2/255" + }, + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "text", + "content": ". We perform 300 iterations (" + }, + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "inline_equation", + "content": "T = 300" + }, + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "text", + "content": ") for iterative targeted attacks as per [3]. For optimization of adversarial examples, all methods employ the simple logit loss function [49]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 567 + ], + "type": "text", + "content": "Models. We evaluate our method using 15 pretrained DNNs, consisting of 10 CNN-based models and 5 transformer-based models. For CNN-based architectures, we employ VGG-16 [35], ResNet-18 (RN-18) [15], ResNet-50 (RN-50) [15], DenseNet-121 (DN-121) [17], Xception (Xcep) [5], MobileNet-v2 (MB-v2) [34], EfficientNet-B0 (EF-B0) [38], Inception ResNet-v2 (IRv2) [37], Inception-v3 (Inc-v3) [36], and Inception-v4 (Inc-v4) [37]. For Transformer-based architectures, we utilize Vision Transformer (ViT) [11], LeViT [14], ConViT [8], Twins [6], and Pooling-based ViT (PiT) [16]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Baseline attacks. We mainly compare our FTM against ten existing techniques: DI [46], RDI [50], MI [9], TI [10], SI [29], VT [39], Admix [40], ODI [2], SU [43], and CFM [3]. We primarily reference CFM, the current state-of-the-art method for targeted transfer-based attacks that requires no additional training dataset. MI and TI techniques apply across all attack methods, so we omit the 'MI-TI' prefix in our references. RDI serves as a common baseline technique in most cases. For DI, RDI, TI, SI, Admix, and ODI, we follow the configurations specified in CFM. For Admix, we also follow CFM to set the mixing weight to " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "w = 0.2" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " and use " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " mixed images (i.e., " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "m_2 = 3" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 346, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 418 + ], + "type": "text", + "content": "in [40]). Given Admix's computational intensity, we set the number of scale images in Admix to 1 (i.e., " + }, + { + "bbox": [ + 313, + 346, + 555, + 418 + ], + "type": "inline_equation", + "content": "m_{1} = 1" + }, + { + "bbox": [ + 313, + 346, + 555, + 418 + ], + "type": "text", + "content": " in [40]) unless otherwise specified. We report the results for ODI instead of RDI-ODI, as ODI performs similarly to RDI-ODI [2, 3]. For SU and CFM, we use the original settings from their papers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "text", + "content": "Settings for our FTM. We set the mixing ratio " + }, + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "text", + "content": " to 0.01. For FTM-E, we set the number of perturbed models to 2 for all experiments due to the computational cost. Following CFM, we avoid perturbing shallow layers because their low-level features vary significantly with input transforms and may cause excessive disturbance during attacks, as explained in CFM's paper. Specifically, we perturb layers where the feature map size is smaller than or equal to " + }, + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "inline_equation", + "content": "\\frac{1}{16}" + }, + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "text", + "content": " of the image size. For a pure ViT, we select the outputs of all fully-connected layers. We sample " + }, + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "inline_equation", + "content": "\\alpha_{k,i}" + }, + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(0,0.75)" + }, + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "inline_equation", + "content": "\\alpha_{max} = 0.75" + }, + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "text", + "content": ". We set the probability " + }, + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 419, + 556, + 552 + ], + "type": "text", + "content": " to 0.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 563, + 440, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 440, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 440, + 575 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Transfer success rates. Table 1 shows the transfer success rates of various attack methods on ten black-box models. When using Inc-v3 as the source model, RDI-FTM-E achieves an average success rate of " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "51.2\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": ", surpassing the previous best method CFM by " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "12.0\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": ". The improvement is particularly notable for challenging target models like IR-v2 and Inc-v4, where RDI-FTM-E increases the success rates by " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "16.2\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "14.4\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": ", respectively. When using RN-50 as the source model, both RDI-FTM and RDI-FTM-E achieve higher success rates than CFM. The results show that the effective feature-level augmentation employed by FTM helps" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25807" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 258 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 258 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 258 + ], + "type": "table", + "html": "
SourceAttackTime (s)ViTLeViTConViTTwinsPiTAvg.
RN-50DI1.370.23.50.41.51.71.5
RDI1.230.713.21.76.17.05.7
RDI-SI5.952.929.46.315.517.914.4
RDI-VT7.162.928.15.215.014.013.0
RDI-Admix3.551.322.52.58.58.48.6
RDI-SU2.210.816.92.36.97.86.9
ODI4.385.137.010.720.129.120.4
RDI-CFM1.394.346.18.925.224.721.8
RDI-FTM1.545.952.910.832.431.526.7
RDI-FTM-E2.926.858.613.635.234.929.8
Inc-v3DI2.010.10.30.00.00.10.1
RDI1.760.21.80.20.40.70.7
RDI-SI8.110.34.10.90.73.21.8
RDI-VT10.50.45.20.81.61.82.0
RDI-Admix4.920.14.10.61.41.41.5
RDI-SU2.360.22.00.31.31.01.0
ODI6.420.812.41.73.56.75.0
RDI-CFM2.132.121.93.26.111.68.9
RDI-FTM2.352.425.04.510.015.311.5
RDI-FTM-E4.373.832.76.812.720.415.3
", + "image_path": "16bfb60c454a0539358d4d95e468ff3bfbd761ac182f38c3e021b9dbce07a3ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 266, + 295, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 310 + ], + "type": "text", + "content": "Table 2. Targeted attack success rates " + }, + { + "bbox": [ + 55, + 266, + 295, + 310 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 266, + 295, + 310 + ], + "type": "text", + "content": " against five transformer-based DNNs. All methods are combined with MI-TI. The best results are shown in bold, and the second best are underlined. Time (s) denotes the average time required to attack a single image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 64, + 323, + 176, + 396 + ], + "blocks": [ + { + "bbox": [ + 64, + 323, + 176, + 396 + ], + "lines": [ + { + "bbox": [ + 64, + 323, + 176, + 396 + ], + "spans": [ + { + "bbox": [ + 64, + 323, + 176, + 396 + ], + "type": "image", + "image_path": "7dcdebbf824c61a03c1b9e359164d9d4514c8b925267b7536ef5ebe8672be52e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 420, + 295, + 444 + ], + "lines": [ + { + "bbox": [ + 55, + 420, + 295, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 295, + 444 + ], + "type": "text", + "content": "Figure 3. Targeted attack success rates " + }, + { + "bbox": [ + 55, + 420, + 295, + 444 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 420, + 295, + 444 + ], + "type": "text", + "content": " based on the number of iterations. The source model is RN-50 and Inc-v3, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 323, + 291, + 396 + ], + "blocks": [ + { + "bbox": [ + 179, + 323, + 291, + 396 + ], + "lines": [ + { + "bbox": [ + 179, + 323, + 291, + 396 + ], + "spans": [ + { + "bbox": [ + 179, + 323, + 291, + 396 + ], + "type": "image", + "image_path": "46cca3ae0e79edbf6d3e1d5057ac9b7cce7f1939e2bae319f364e9a265714cd1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 70, + 401, + 285, + 409 + ], + "blocks": [ + { + "bbox": [ + 70, + 401, + 285, + 409 + ], + "lines": [ + { + "bbox": [ + 70, + 401, + 285, + 409 + ], + "spans": [ + { + "bbox": [ + 70, + 401, + 285, + 409 + ], + "type": "image", + "image_path": "3e8d221cfab650e8ce4d2b3f98383c5f07bcd38cdfd59d8d6b110fae4ccab8f7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 459, + 230, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 230, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 230, + 471 + ], + "type": "text", + "content": "generate transferable adversarial examples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 472, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 579 + ], + "type": "text", + "content": "Table 2 shows the transfer success rates of various attack methods on five transformer-based models. Our methods also achieve significant improvements over existing methods. For instance, when using RN-50 as the source model, RDI-FTM and RDI-FTM-E improve the targeted attack success rates of CFM by " + }, + { + "bbox": [ + 55, + 472, + 295, + 579 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 472, + 295, + 579 + ], + "type": "inline_equation", + "content": "8.0\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 579 + ], + "type": "text", + "content": ", respectively. The results suggest that our method can effectively improve the transferability of adversarial examples across different architectures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Figure 3 shows the targeted attack success rates based on the number of iterations. We evaluate our methods under different numbers of iterations using two experimental settings: RN-50 → Inc-v4 and Inc-v3 → IR-v2. The results show that FTM and FTM-E achieve stable performance across different iteration counts. For the RN-50 → Inc-v4 setting, SI exhibits superior performance at the beginning of the optimization process. However, as the number of iterations increases, FTM surpasses existing approaches. In the Inc-v3 → IR-v2 setting, FTM-E significantly outperforms other methods across the entire range of iterations." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 433, + 163 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 433, + 163 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 433, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 433, + 163 + ], + "type": "image", + "image_path": "ee2c2b1cc7aad49487060a0c2ff970f4a325062a6c184849f37f18c191a460a5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 184, + 553, + 206 + ], + "lines": [ + { + "bbox": [ + 313, + 184, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 553, + 206 + ], + "type": "text", + "content": "Figure 4. Attack success rates with different mixing ratio " + }, + { + "bbox": [ + 313, + 184, + 553, + 206 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 184, + 553, + 206 + ], + "type": "text", + "content": ". Each color represents the attack effectiveness on a specific target model." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 435, + 72, + 550, + 163 + ], + "blocks": [ + { + "bbox": [ + 435, + 72, + 550, + 163 + ], + "lines": [ + { + "bbox": [ + 435, + 72, + 550, + 163 + ], + "spans": [ + { + "bbox": [ + 435, + 72, + 550, + 163 + ], + "type": "image", + "image_path": "1c3990ce1590421259b6ab0f586688f6222a8952e5bb6e1eb2915d5e46bc7d74.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 222, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 555, + 389 + ], + "type": "text", + "content": "Computational overhead. We analyze the computational efficiency of different methods and report their costs in Table 2. We use an NVIDIA 3090 GPU to compute the time per image. The results demonstrate that our approach introduces reasonable computational overhead compared to CFM. Specifically, with RN-50 as the source model, RDI-FTM requires an average of 1.54 seconds per image, representing only a " + }, + { + "bbox": [ + 313, + 222, + 555, + 389 + ], + "type": "inline_equation", + "content": "10.8\\%" + }, + { + "bbox": [ + 313, + 222, + 555, + 389 + ], + "type": "text", + "content": " increase over CFM's 1.39 seconds. In contrast, alternative methods like SI, VT, and ODI impose substantially higher computational burdens. For instance, VT requires 7.16 seconds per image, and SI requires 5.95 seconds per image. We also include relevant analysis in Figure 1. The result shows that FTM achieves strong attack performance while maintaining low computational cost." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "text", + "content": "Combination with existing techniques. Table 3 presents transfer success rates of our method combined with ensemble-based method [30], SI, and Admix. Using the ensemble of RN-50 and Inc-v3 as the surrogate model, our RDI-SI-FTM-E achieves " + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "inline_equation", + "content": "74.5\\%" + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "text", + "content": " average success rate, surpassing existing methods. FTM notably enhances performance on transformer-based models. For instance, RDI-SI-FTM achieves " + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "inline_equation", + "content": "64.3\\%" + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "text", + "content": " on PiT, which is " + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "inline_equation", + "content": "5.0\\%" + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "text", + "content": " higher than RDI-SI-CFM " + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "inline_equation", + "content": "(59.3\\%)" + }, + { + "bbox": [ + 313, + 389, + 554, + 521 + ], + "type": "text", + "content": " with comparable computational cost. These results demonstrate that FTM is a generic technique that combines effectively with existing attack methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 529, + 408, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 408, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 408, + 541 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 545, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 582 + ], + "type": "text", + "content": "For all ablation experiments, we use RN-50 to generate adversarial examples and evaluate the average success rates on all models from Section 4.1, unless stated otherwise." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": "Stability with respect to " + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": ". Figure 4 shows the transfer success rates of FTM across different mixing ratios " + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": ". The success rates on five target models are plotted in distinct colors. When " + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": " exceeds 0.04, the success rates decrease substantially across all target models. However, for " + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": " values below 0.02, we observe stable attack performance across all models. The results suggest that FTM maintains robust performance within a reasonable range of " + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": " values, showing its stability with respect to " + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Compare FTM and CFM using ensemble strategy. Figure 5 shows the attack performance of FTM and CFM" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25808" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 70, + 537, + 198 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 537, + 198 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 537, + 198 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 537, + 198 + ], + "type": "table", + "html": "
AttackRN-50 + Inc-v3 ⇒
Time (s)VGG-16RN-50*Inc-v3*DN-121IR-v2Inc-v4XcepViTLeViTConViTTwinsPiTAvg.
RDI3.2666.998.498.783.135.642.525.40.822.43.29.810.841.5
ODI7.1770.994.096.183.465.469.957.69.152.918.628.642.857.4
RDI-CFM4.3986.198.397.092.171.779.972.010.366.520.341.945.865.2
RDI-ODI-CFM7.7363.988.681.374.154.356.054.312.649.920.929.641.052.2
RDI-Admix-CFM13.3886.298.197.792.775.181.676.114.073.124.850.753.968.7
RDI-SI-CFM22.1387.598.197.392.778.081.678.020.876.332.955.059.371.5
RDI-FTM4.4787.597.896.792.475.481.774.014.571.025.249.354.468.3
RDI-FTM-E8.8188.298.197.193.476.783.477.118.474.930.153.458.070.7
RDI-Admix-FTM14.1086.096.796.592.278.383.877.320.175.829.955.758.570.9
RDI-SI-FTM22.6687.596.896.293.379.083.679.125.777.737.559.964.373.4
RDI-SI-FTM-E43.4387.897.396.793.979.183.578.227.379.242.962.965.674.5
", + "image_path": "f984a5b2379bff75e2aaa0d3f394b75d0d051d9e049fc046189a8fa2d9569616.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 555, + 240 + ], + "type": "text", + "content": "Table 3. Targeted attack success rates (%) of combination with different methods. The surrogate model is the ensemble of RN-50 and Inc-v3. All methods are combined with MI-TI. The best results are shown in bold, and the second best are underlined. Time (s) denotes the average time required to attack a single image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 248, + 175, + 340 + ], + "blocks": [ + { + "bbox": [ + 58, + 248, + 175, + 340 + ], + "lines": [ + { + "bbox": [ + 58, + 248, + 175, + 340 + ], + "spans": [ + { + "bbox": [ + 58, + 248, + 175, + 340 + ], + "type": "image", + "image_path": "441d4b9e6024e07ed96a24419f7036373577becc02ba2ba83906194d0418ada6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 344, + 295, + 389 + ], + "lines": [ + { + "bbox": [ + 55, + 344, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 295, + 389 + ], + "type": "text", + "content": "Figure 5. Attack performance and computational costs with different numbers of ensemble models. Left: Our FTM vs. CFM using ensemble strategy. Right: Computational costs of our FTM and RAP. -ens2 means using 2 perturbed models, and so on." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 249, + 293, + 340 + ], + "blocks": [ + { + "bbox": [ + 175, + 249, + 293, + 340 + ], + "lines": [ + { + "bbox": [ + 175, + 249, + 293, + 340 + ], + "spans": [ + { + "bbox": [ + 175, + 249, + 293, + 340 + ], + "type": "image", + "image_path": "725eedec4cde1e59b1a51d51f613603b65e030a1423161c064331a43b94cf52f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 399, + 175, + 491 + ], + "blocks": [ + { + "bbox": [ + 58, + 399, + 175, + 491 + ], + "lines": [ + { + "bbox": [ + 58, + 399, + 175, + 491 + ], + "spans": [ + { + "bbox": [ + 58, + 399, + 175, + 491 + ], + "type": "image", + "image_path": "bf1e993cd0797ceeeda2af3be640f32a8884d641ec7f44753186e6331a743310.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 497, + 295, + 552 + ], + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 552 + ], + "type": "text", + "content": "Figure 6. Left: Attack success rates vs. " + }, + { + "bbox": [ + 55, + 497, + 295, + 552 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 497, + 295, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 497, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\alpha_{max}" + }, + { + "bbox": [ + 55, + 497, + 295, + 552 + ], + "type": "text", + "content": ". Right: Attack success rates vs. " + }, + { + "bbox": [ + 55, + 497, + 295, + 552 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 497, + 295, + 552 + ], + "type": "text", + "content": " with/without clean features; The blue line is the success rates when the clean features from CFM are integrated with our FTM, while the orange line represents only using our learnable perturbations." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 175, + 400, + 293, + 491 + ], + "blocks": [ + { + "bbox": [ + 175, + 400, + 293, + 491 + ], + "lines": [ + { + "bbox": [ + 175, + 400, + 293, + 491 + ], + "spans": [ + { + "bbox": [ + 175, + 400, + 293, + 491 + ], + "type": "image", + "image_path": "0821458157fe6861f9f42b074e1ca1b9b4dda958babba4fc600b4d48cac09113.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 687 + ], + "type": "text", + "content": "using different numbers of perturbed models. On the left, our FTM consistently outperforms CFM across all ensemble sizes. Notably, FTM with only 2 perturbed models achieves better attack performance than an ensemble of 5 CFM-perturbed models, underscoring FTM's superior effectiveness. It is also worth mentioning that single FTM and CFM have comparable time complexity (Table 2). On the right, we compare our FTM with RAP [33], and the results show that FTM significantly outperforms RAP in both effectiveness and computational efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Analysis of probability " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and ratio " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{\\text{max}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ". Figure 6 presents an analysis of FTM's attack success rates under" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "text", + "content": "varying values of " + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "inline_equation", + "content": "\\alpha_{max}" + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "text", + "content": ". The left plot shows that FTM achieves optimal performance with " + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "inline_equation", + "content": "p = 0.1" + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "inline_equation", + "content": "\\alpha_{max} = 0.75" + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "text", + "content": ". FTM maintains competitive performance under various parameter configurations, such as " + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "inline_equation", + "content": "p, \\alpha_{max} = 0.15, 0.5" + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "inline_equation", + "content": "0.05, 1.0" + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "text", + "content": ", indicating its parameter stability. The right plot illustrates the impact of clean features on FTM. With clean features enabled, FTM's performance declines significantly as " + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "text", + "content": " increases from 0.1 to 1.0, while FTM without clean features shows stable performance across varying " + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "text", + "content": " values. This suggests that although clean features are crucial to FTM's performance, they also add instability to the attack process. Additionally, setting " + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "inline_equation", + "content": "p = 1.0" + }, + { + "bbox": [ + 313, + 249, + 555, + 416 + ], + "type": "text", + "content": " (where all layers are updated) causes excessive perturbations, making the attack overly aggressive and reducing its effectiveness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 417, + 555, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 501 + ], + "type": "text", + "content": "The drop of the blue line in Figure 6 occurs because the clean accuracy of the perturbed surrogate model becomes extremely low " + }, + { + "bbox": [ + 313, + 417, + 555, + 501 + ], + "type": "inline_equation", + "content": "(3.8\\%)" + }, + { + "bbox": [ + 313, + 417, + 555, + 501 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 417, + 555, + 501 + ], + "type": "inline_equation", + "content": "p = 1.0" + }, + { + "bbox": [ + 313, + 417, + 555, + 501 + ], + "type": "text", + "content": " when using clean feature mixup. Without clean feature mixup (orange line), the accuracy remains high at " + }, + { + "bbox": [ + 313, + 417, + 555, + 501 + ], + "type": "inline_equation", + "content": "65.3\\%" + }, + { + "bbox": [ + 313, + 417, + 555, + 501 + ], + "type": "text", + "content": ". The low accuracy makes it impossible to generate effective adversarial examples, causing the attack success rate of the blue line to drop to near 0." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 388, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 388, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 388, + 522 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 529, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 685 + ], + "type": "text", + "content": "In this work, we study the challenging problem of enhancing transferability in targeted adversarial attacks. Our investigation reveals that feature-level augmentation strategies show promising potential for boosting transferability effectively and efficiently. We introduce Feature Tuning Mixup (FTM), a novel method that incorporates both attack-specific feature perturbations and clean features in the feature space. To optimize the learnable perturbations efficiently, we develop a momentum-based stochastic update strategy. Comprehensive experiments on the ImageNet-compatible dataset demonstrate that FTM achieves substantial improvement over current state-of-the-art methods while maintaining low computational cost." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by HK RGC GRF under Grant PolyU 15201323." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25809" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "text", + "content": "[1] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. In ICLR, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 167 + ], + "type": "text", + "content": "[2] Junyoung Byun, Seungju Cho, Myung-Joon Kwon, Hee-Seon Kim, and Changick Kim. Improving the transferability of targeted adversarial examples through object-based diverse input. In CVPR, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 295, + 213 + ], + "type": "text", + "content": "[3] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In CVPR, 2023. 1, 3, 4, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "type": "text", + "content": "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy (SP), 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 294, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 294, + 268 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 294, + 268 + ], + "type": "text", + "content": "[5] Francois Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 270, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 294, + 312 + ], + "type": "text", + "content": "[6] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In NeurIPS, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 315, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 294, + 345 + ], + "type": "text", + "content": "[7] Xuelong Dai, Kaisheng Liang, and Bin Xiao. Advdiff: Generating unrestricted adversarial examples using diffusion models. In ECCV, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 347, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 294, + 390 + ], + "type": "text", + "content": "[8] Stéphane d'Ascoli, Hugo Touvron, Matthew L. Leavitt, Ari S. Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In ICML, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 392, + 294, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 423 + ], + "type": "text", + "content": "[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In CVPR, 2018. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 426, + 294, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 426, + 294, + 457 + ], + "spans": [ + { + "bbox": [ + 57, + 426, + 294, + 457 + ], + "type": "text", + "content": "[10] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In CVPR, 2019. 1, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 460, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 460, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 57, + 460, + 294, + 523 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 525, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 294, + 568 + ], + "type": "text", + "content": "[12] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, and Dawn Song. Robust physical-world attacks on deep learning visual classification. In CVPR, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 570, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 294, + 601 + ], + "type": "text", + "content": "[13] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 603, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 294, + 645 + ], + "type": "text", + "content": "[14] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In ICCV, 2021. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[16] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, 2021. 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[17] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "text", + "content": "[18] Qian Huang, Isay Katsman, Zeqi Gu, Horace He, Serge J. Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In ICCV, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "type": "text", + "content": "[19] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "text", + "content": "[20] Nathan Inkawich, Wei Wen, Hai (Helen) Li, and Yiran Chen. Feature space perturbations yield more transferable adversarial examples. In CVPR, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "type": "text", + "content": "[21] Nathan Inkawich, Kevin J. Liang, Lawrence Carin, and Yiran Chen. Transferable perturbations of deep feature distributions. In ICLR, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 264, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 307 + ], + "type": "text", + "content": "[22] Nathan Inkawich, Kevin J. Liang, Binghui Wang, Matthew Inkawich, Lawrence Carin, and Yiran Chen. Perturbing across the feature hierarchy to improve standard and strict blackbox attack transferability. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "type": "text", + "content": "[23] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In ICLR (Workshop), 2017. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 343, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 374 + ], + "type": "text", + "content": "[24] Maosen Li, Cheng Deng, Tengjiao Li, Junchi Yan, Xinbo Gao, and Heng Huang. Towards transferable targeted attack. In CVPR, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 376, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 398 + ], + "type": "text", + "content": "[25] Qizhang Li, Yiwen Guo, and Hao Chen. Yet another intermediate-level attack. In ECCV, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 399, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 431 + ], + "type": "text", + "content": "[26] Yanjie Li, Yiquan Li, Xuelong Dai, Songtao Guo, and Bin Xiao. Physical-world optical adversarial attacks on 3d face recognition. In CVPR, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "type": "text", + "content": "[27] Yanjie Li, Kaisheng Liang, and Bin Xiao. UV-attack: Physical-world adversarial attacks on person detection via dynamic-NeRF-based UV mapping. In ICLR, 2025. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 468, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 488 + ], + "type": "text", + "content": "[28] Kaisheng Liang and Bin Xiao. Style: Boosting the transferability of adversarial examples. In CVPR, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 491, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 553, + 521 + ], + "type": "text", + "content": "[29] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In ICLR, 2020. 1, 3, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 524, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 524, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 524, + 553, + 555 + ], + "type": "text", + "content": "[30] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In ICLR, 2017. 3, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "type": "text", + "content": "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 1, 2, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 602, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 634 + ], + "type": "text", + "content": "[32] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. On generating transferable targeted perturbations. In ICCV, 2021. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 636, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 678 + ], + "type": "text", + "content": "[33] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. In NeurIPS, 2022. 3, 8" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[34] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 6" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25810" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 691 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[37] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A. Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In AAAI, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "type": "text", + "content": "[38] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "text", + "content": "[39] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In CVPR, 2021. 1, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[40] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In ICCV, 2021. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "text", + "content": "[41] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In ICCV, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "type": "text", + "content": "[42] Zhibo Wang, Hongshan Yang, Yunhe Feng, Peng Sun, Hengchang Guo, Zhifei Zhang, and Kui Ren. Towards transferable targeted adversarial examples. In CVPR, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 366, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 399 + ], + "type": "text", + "content": "[43] Zhipeng Wei, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Enhancing the self-universality for transferable targeted attacks. In CVPR, 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 400, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 442 + ], + "type": "text", + "content": "[44] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip connections matter: On the transferability of adversarial examples generated with resnets. In ICLR, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "type": "text", + "content": "[45] Weibin Wu, Yuxin Su, Michael R Lyu, and Irwin King. Improving the transferability of adversarial samples with adversarial transformations. In CVPR, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "type": "text", + "content": "[46] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019. 1, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "type": "text", + "content": "[47] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Boosting transferability of targeted adversarial examples via hierarchical generative networks. In ECCV, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 568, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 611 + ], + "type": "text", + "content": "[48] Jianping Zhang, Weibin Wu, Jen-tse Huang, Yizhan Huang, Wenxuan Wang, Yuxin Su, and Michael R Lyu. Improving adversarial transferability via neuron attribution-based attacks. In CVPR, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 613, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 645 + ], + "type": "text", + "content": "[49] Zhengyu Zhao, Zhuoran Liu, and Martha Larson. On success and simplicity: A second look at transferable targeted attacks. In NeurIPS, 2021. 1, 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "type": "text", + "content": "[50] Junhua Zou, Zhisong Pan, Junyang Qiu, Xin Liu, Ting Rui, and Wei Li. Improving the transferability of adversarial examples with resized-diverse-inputs, diversity-ensemble and region fitting. In ECCV, 2020. 3, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25811" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/0e31d729-0f45-4249-a70f-af6d4389a18f_content_list.json b/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/0e31d729-0f45-4249-a70f-af6d4389a18f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4564b9022220b147b2c5f12d7acd982e002d8a8a --- /dev/null +++ b/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/0e31d729-0f45-4249-a70f-af6d4389a18f_content_list.json @@ -0,0 +1,1704 @@ +[ + { + "type": "text", + "text": "Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration", + "text_level": 1, + "bbox": [ + 107, + 128, + 890, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuxuan Gu $^{1}$ , Haoxuan Wang $^{1}$ , Pengyang Ling $^{1,2}$ , Zhixiang Wei $^{1}$ ,", + "bbox": [ + 241, + 202, + 754, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huaian Chen $^{1\\dagger}$ , Yi Jin $^{1\\dagger}$ , and Enhong Chen $^{1}$", + "bbox": [ + 323, + 220, + 671, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ University of Science and Technology of China $^{2}$ Shanghai AI Laboratory", + "bbox": [ + 186, + 238, + 810, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we observe that the collaboration of various foundation models can perceive semantic and degraded information within images, thereby guiding the low-light enhancement process. Specifically, we propose a self-supervised low-light enhancement framework based on the multiple foundation models collaboration (dubbed FoCo), aimed at improving both the visual quality of enhanced images and the performance in high-level applications. At the feature level, FoCo leverages the rich features from various foundation models to enhance the model's semantic perception during training, thereby reducing the gap between enhanced results and high-quality images from a high-level perspective. At the task level, we exploit the robustness-gap between strong foundation models and weak models, applying high-level task guidance to the low-light enhancement training process. Through the collaboration of multiple foundation models, the proposed framework shows better enhancement performance and adapts better to high-level tasks. Extensive experiments across various enhancement and application benchmarks demonstrate the qualitative and quantitative superiority of the proposed method over numerous state-of-the-art techniques.", + "bbox": [ + 86, + 323, + 485, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 681, + 220, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Images taken under short exposure or low-light conditions frequently suffer from significant degradation, including loss of color information and the introduction of substantial noise. Such degradation severely impacts the performance of downstream tasks [19, 36]. Low-light image enhancement (LIE) aims to restore input images captured in low-light environments to high-quality, well-lit images. With the advancement of deep learning technology in computer vision, recent research [2, 25, 31, 33, 34, 39] has achieved low-light image enhancement through supervised training on paired datasets. However, collecting", + "bbox": [ + 89, + 707, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e934dfe43f4f9e781317759ad13a83d9c27db5c07d05b43b14c4df1c02f3717.jpg", + "image_caption": [ + "(a) Prior-based Framework" + ], + "image_footnote": [], + "bbox": [ + 516, + 290, + 903, + 445 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb14190bedc8083fdfe466f7f5b1bf49064a3c2525be1a386033979f7869ab65.jpg", + "image_caption": [ + "(b) Previous High-for-Low Framework", + "Figure 1. The proposed approach demonstrates substantial improvements over existing unsupervised methods in terms of enhancement tasks (PSNR, SSIM) and application tasks (classification/Acc, face detection/AP, and instance segmentation/mAP), attributed to the proposed Multi-VFM Collaboration Framework. Compared with previous unsupervised LIE learning paradigms (a) and (b), the proposed framework (c) can make full use of the foundation models from a collaborative perspective to improve visual quality and downstream task performance." + ], + "image_footnote": [], + "bbox": [ + 516, + 459, + 647, + 606 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c16ed0f454b6d418c106b0188de6dfaee2a297b5b123ca6679001322b239b424.jpg", + "image_caption": [ + "(c) Proposed Multi-Foundation model Collaboration Framework" + ], + "image_footnote": [], + "bbox": [ + 656, + 459, + 875, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the substantial amount of paired images required for these methods is both time-consuming and costly, which highlights the broad potential applications of self-supervised approaches [6, 7, 13, 18]. As shown in Fig. 1 (a), early self-supervised training paradigms mainly relied on manually defined priors, such as smoothness priors, brightness priors, or histogram equalization. The drawback of such prior-based frameworks is their lack of connection with semantic", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 107, + 886, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16071", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "information, and the efficacy of these prior functions can diminish under complex natural conditions. For instance, the smoothness prior may sometimes degrade semantic information, limiting these methods' performance in practical applications and impacting the visual quality of results.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recognizing these limitations, as illustrated in Figure 1 (b), some approaches [15, 16, 27, 29, 37] have introduced high-level models into the training of Low-light Image Enhancement (LIE) models, significantly improving the model's ability to retain semantic information. However, this line of research typically relies on a specific high-level model and does not fully explore the potential of different vision models, resulting in only limited visual improvements. In recent years, the emergence of various foundation vision models widely applied in real-world scenarios has led us to consider integrating multiple foundational models from a collaborative perspective, which aims to improve the visual quality and practical performance of LIE methods more effectively.", + "bbox": [ + 89, + 167, + 480, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we explore the use of multiple foundation models to strengthen the self-supervised enhancement process, as shown in Figure 1 (c). To this end, We address two key challenges: (1) how to effectively utilize pre-trained knowledge? (2) how to handle differences between models? To deal with these challenges, we propose FoCo, a Low-light Image Enhancement (LIE) framework based on the multiple foundation models ensemble, aligning the enhancement results with high-quality images from both feature-level and task-level perspectives. As the core strategies to implement FoCo, we introduce two optimization strategies: Foundation Feature-space Alignment (FoFA) and Gap-driven task-based Alignment (GaTA). FoFA focuses on creating a unified feature space for foundation models, using adversarial learning to align enhanced results with high-quality image sets. GaTA leverages robustness differences across models to model degradation scenarios, using robust pseudo-labels to directly incorporate high-level task training into the enhancement model's optimization process.", + "bbox": [ + 91, + 381, + 482, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of this paper are as follows:", + "bbox": [ + 89, + 684, + 482, + 712 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to explore the collaborative effects of multiple vision foundation models on the LIE task, and then propose FoCo, a framework designed to enhance both the human- and machine-understanding performance of unsupervised LIE methods by leveraging complementary interactions among multiple foundational vision models.", + "- We introduce two key strategies FoFA and GaTA to fully exploit the pre-trained knowledge of foundation models, which align enhancement results and high-quality images from the perspective of foundation models.", + "- We conduct extensive experiments on benchmarks for low-light enhancement, object detection, instance seg" + ], + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mentation, and classification. The experimental results show significant performance improvements, demonstrating the proposed method's effectiveness in enhancing visual quality and application performance.", + "bbox": [ + 526, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 167, + 660, + 184 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Low-light Image Enhancement", + "text_level": 1, + "bbox": [ + 511, + 193, + 785, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised methods, such as RetinexNet [25], KinD [33], and more recent approaches like Retinexformer [2], Diff-Retinex [31], and GLARE [39], utilize paired datasets and advanced model architectures to enhance image quality. While successful, these methods face challenges due to their dependence on large-scale paired datasets, which are difficult to obtain in real-world scenarios. Unsupervised methods, such as EnLightenGAN [11], and CoLIE [4], aim to overcome this limitation by eliminating the need for paired data, leveraging adversarial learning and multimodal approaches. However, these methods often neglect integration with high-level models, limiting their ability to preserve semantic information and adapt to downstream tasks.", + "bbox": [ + 511, + 215, + 903, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. High-Level guided Enhancement", + "text_level": 1, + "bbox": [ + 511, + 436, + 802, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works [15, 16, 27, 29, 37] have increasingly focused on leveraging high-level models to supervise low-light enhancement models, aiming to improve the fidelity of the enhancement results. For example, SGZ [37] incorporates the concept of focal loss to maximize the logits of a pre-trained semantic segmentation model, enhancing the model's ability to preserve semantic information. SKF [27] integrates features from semantic segmentation models into the enhancement model and proposes a semantic-based histogram equalization loss and adversarial loss based on segmentation results. Methods like NeRCo [29] and CLIP-LIT [16] employ pre-trained CLIP models to align enhancement results with high-quality images during adversarial training. The methods above have significant success, but a common limitation of these approaches is their reliance on a specific high-level model, and they often do not fully exploit the model's potential.", + "bbox": [ + 511, + 460, + 906, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 511, + 731, + 612, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. FoCo Framework", + "text_level": 1, + "bbox": [ + 511, + 756, + 687, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To implement the low-light enhancement process, we first define a simple enhancer $G$ for enhancing a low-light input image $I$ , and then use some priors to constrain the behavior of $G$ . By incorporating the smoothness and consistency losses $\\mathcal{L}_C$ borrowed from [21] and [6], we can construct a baseline low-light enhancement model that fully relies on simply defined hand-crafted prior. However, this dependency on these hand-crafted prior compromises", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16072", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c0ed6892d501810b49c536f7e15911915db94ad47e55d16f81fa4211f8a5c49.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed low-light enhancement method, which enhances a basic LIE model using multi-foundation model ensemble. Given the set of pre-trained foundation models, the enhancement model first performs basic enhancement, followed by supervision from the ensemble through two strategies: (1) Feature-Level: adversarial training aligns enhanced results with high-quality images in the unified feature space, and (2) Task-Level: robustness differences between models are used to align the enhancement results with high-quality images on high-level tasks." + ], + "image_footnote": [], + "bbox": [ + 151, + 70, + 843, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the model's visual quality and adaptability to downstream tasks. Mathematically, such a straightforward image-level enhancement approach can be formalized as:", + "bbox": [ + 88, + 431, + 486, + 477 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\theta} d (G (I), \\hat {I}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 484, + 482, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{I}$ represents the ideal normal-light image, $\\theta$ denotes the parameters of the enhancement model, $d(\\cdot)$ is the distance measurement function. To address the limitations of the prior-based approach, we introduce FoCo framework designed to enhance this model from the perspective of multiple foundation models collaboration. Specifically, we propose (1) Foundation Feature-space Alignment (FoFA), at the feature level (indicated in green), and (2) Gap-driven Task-based Alignment (GaTA), at the task level (indicated in blue). These strategies leverage the feature extraction process (denoted as $F^f$ ) and task decision process (denoted as $F^t$ ) of foundation models. Thus, FoCo's optimization objective can be expressed as:", + "bbox": [ + 88, + 516, + 486, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\arg \\min _ {\\theta} \\underbrace {d (G (I) , \\hat {I})} _ {\\mathcal {L} _ {C}} + \\underbrace {d \\left(F ^ {f} (G (I)) , F ^ {f} (\\hat {I})\\right)} _ {\\text {F o F A}} \\tag {2} \\\\ + \\underbrace {d (F ^ {t} (G (I)) , F ^ {t} (\\hat {I}))} _ {\\text {G a T A}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 720, + 483, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The corresponding framework is illustrated in Figure 2.", + "bbox": [ + 89, + 806, + 457, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Foundation Feature-space Alignment", + "text_level": 1, + "bbox": [ + 89, + 832, + 413, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observation. Figure 3 (a) displays a GAN-enhanced low-light image from the LIS [3] dataset, alongside its Mask-RCNN detect result. The GAN was trained in an unpaired", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "manner on the LIS dataset. It can be observed that the GAN-enhanced result suffers from significant semantic degradation, likely due to the discriminator's limited semantic awareness, which fails to guide the generator effectively. This observation leads us to a potential solution: could we enhance the discriminator by incorporating a feature extractor with strong semantic awareness? As shown in Figure 3 (b), we use CLIP, SAM and ResNet separately to extract feature maps from both images. Incorporating these base models into the discriminator training process, as shown in Figure 3 (a), led to significant improvements in detection performance. However, SAM produced darker results, CLIP introduced color bias, and ResNet maintained limited semantic integrity. We hypothesize that these differences stem from the varying focus of each foundational model: CLIP prioritizes the primary objects in the image and is less sensitive to color bias and subtle degradations, SAM captures fine-grained semantics but lacks sensitivity to lighting conditions, while ResNet offers weaker semantic understanding compared to the other models but is more susceptible to degradations and lighting variations, which still contributes to the discriminator's learning. Interestingly, ResNet's lack of robustness to degradations may become a key factor in complementing CLIP and SAM, while the discriminator needs to identify degradation and illumination. As shown in Figure 3 (c), where we observed that ResNet's features were more sensitive to lighting variations compared to the other models.", + "bbox": [ + 511, + 431, + 908, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This inspires us that CLIP, SAM, and ResNet may complement each other at the feature level: CLIP highlights key objects, SAM provides fine-grained guidance, and", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16073", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3a6c20a317ba9e5d8715c3b5c2a04d789db5794c45e7f989e3fb2320b078be63.jpg", + "image_caption": [ + "Figure 3. The perception capabilities of foundation models on unsupervised low-light enhancer training. (a) presents the discriminator feature maps, enhanced results, and instance segmentation outcomes using different foundation model-supervised discriminators. (b) illustrates the process of adversarial training supervised by the foundation model. (c) compares the robustness of CLIP and ResNet feature maps under varying lighting conditions. The results suggest that foundation models may collaborate at feature level." + ], + "image_footnote": [], + "bbox": [ + 109, + 68, + 885, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ResNet adds degradation-related information to guide semantic learning in the discriminator. However, the feature spaces of CLIP, SAM, and ResNet are unaligned, preventing the discriminator from directly utilizing their outputs. The challenge now is how to effectively integrate CLIP, SAM, and ResNet into the discriminator training process.", + "bbox": [ + 88, + 359, + 480, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Method. Based on this observation, Foundation Feature-space Alignment (FoFA) aims to leverage features from different foundation models to bring the enhanced results closer to a high-quality image set $\\tilde{I}$ . Let $F_{i,j}$ denote the feature extractor of the $i$ -th pre-trained foundation models, extending up to the $j$ -th layer, given the varied structures and training data of these models, our goal is to construct a unified feature space $\\hat{F}_j$ and find semantic and degradation information in it that is beneficial to the enhancement process. To this end, we use a series of convolutional projectors $P_{i,j}$ to map the features from different spaces into a unified feature space. To align the numerical scale of these features with the subsequent modules, we normalize the output features, defined as:", + "bbox": [ + 89, + 450, + 482, + 660 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {i, j} ^ {f} (x) = \\operatorname {N o r m} \\left(P _ {i, j} \\left(F _ {i, j} ^ {f} (x)\\right)\\right), s. t. x \\sim p (x), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 667, + 482, + 689 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {j} ^ {f} (x) = \\left\\{\\hat {F} _ {i, j} ^ {f} (x) \\right\\} _ {i = 0} ^ {n}, s. t. x \\sim p (x), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 700, + 480, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $n$ denotes the count number of foundation models, $P$ is the convolutional projectors and $F_{i,j}^{f}$ denotes feature extraction with foundation models. Then we adopt an adversarial learning approach by introducing a discriminator $D$ , which links the features from the foundation models with the enhancement model. The input to $D$ consists of images along with their corresponding semantic information derived from the unified feature space. Specifically, inspired by [27], we employ cross-attention across both spatial and channel dimensions between the image features and the semantic features to generate attention maps $A^c$ and $A^s$ :", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA ^ {c} = \\operatorname {S o f t m a x} \\left(P ^ {c, k} \\left(f _ {j} ^ {F}\\right) \\times P ^ {c, q} \\left(f _ {j} ^ {D}\\right) / \\sqrt {C}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 382, + 906, + 409 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA ^ {s} = \\operatorname {S o f t m a x} \\left(P ^ {s, k} \\left(f _ {j} ^ {F}\\right) \\cdot P ^ {s, q} \\left(f _ {j} ^ {D}\\right) / \\sqrt {H \\cdot W}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 436, + 906, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{j}^{F} = \\hat{F}_{j}^{f}(x)$ and $f_{j}^{D}$ is the feature map of $j$ -th layer in $D$ . These attention maps are then utilized to adjust the original image features to the output feature $f_{j}^{O}$ as follows:", + "bbox": [ + 511, + 474, + 905, + 523 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {j} ^ {O} = A ^ {c} \\times A ^ {s} + f _ {j} ^ {D}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 534, + 903, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The optimization objective of $D$ is to align its learned features with the outputs of the foundation model while simultaneously distinguishing between enhanced low-quality images and high-quality images. Throughout this process, it is crucial to ensure that the features input to the classification layer of $D$ are closely related to the semantic and degradation information. Based on the observation, the loss function for the discriminator $D$ can be expressed as follows:", + "bbox": [ + 511, + 559, + 906, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {D} = - \\mathbb {E} _ {x \\sim p _ {h i g h} (x)} [ \\log D (x) + \\left\\| D ^ {f} (x) - \\hat {F} _ {m} ^ {f} (x) \\right\\| ] \\\\ - \\mathbb {E} _ {x \\sim p l o w (x)} [ \\log (1 - D (G (x))) + \\\\ \\left\\| D ^ {f} (G (x)) - \\hat {F} _ {m} ^ {f} (G (x)) \\right\\| ], \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 696, + 903, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G} = - \\mathbb {E} _ {x \\sim p _ {\\text {l o w}} (x)} [ \\log D (G (x)) ], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 786, + 903, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m$ denotes the layer count of the foundation model, $D^{f}$ denotes feature extraction with discriminator. The discriminator $D$ effectively projects the foundation models' features into a unified feature space, providing valuable guidance for the enhancement model. Unlike traditional image-level adversarial training, FoFA focuses on selected", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16074", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/612497201aabd6991f1b5e21ef40aebc6e075b34e8b38d6f1b01b872e486ab60.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 285, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c03be5fbfaaa1b3d9c95d550950f886214dd553d87110430e16d3c72f6e842dc.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 287, + 88, + 468, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ea0c9e4b74b05934b9104f09644c89c74fe9b06918915a325d4fd82179b02218.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 119, + 200, + 187, + 234 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1ff69a92396fdf9be25bc22649b5b30fc8dd68a418326e433e0dc7343676ec9d.jpg", + "image_caption": [ + "CLIP: Ambulance (0.9067) ResNet:Police van (0.5195)", + "CLIP: Ambulance (0.9438) ResNet: Ambulance (0.7087)" + ], + "image_footnote": [], + "bbox": [ + 191, + 200, + 256, + 234 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2741c073732cfd083e50fff79a1626fcea3bf35433b889383c166e0d91d4bfc9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 200, + 325, + 234 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2c8b7680ba91aff6582fd8b3dfd5cda23e68441cef75e453264688574367a6d7.jpg", + "image_caption": [ + "CLIP: Ambulance (0.9131) \nResNet: Police van (0.6643)" + ], + "image_footnote": [], + "bbox": [ + 326, + 200, + 393, + 234 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0e08270354025a7607191e1b6f7b14b596325df1a926a40d715efd9259dcb482.jpg", + "image_caption": [ + "CLIP: Ambulance (0.9092) ResNet:Police van (0.8759)" + ], + "image_footnote": [], + "bbox": [ + 395, + 200, + 465, + 234 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3e92ccc07a0bd255bc1eab55626fdd4c761ef0875a1e1b5d92391d07480e7dcf.jpg", + "image_caption": [ + "(d)", + "CLIP: Vending (0.3428) \nResNet: Mail box (0.3349)", + "Figure 4. Robustness to illumination across models. (a) and (b) show the similarity of results between CLIP and ResNet under different lighting adjustments. (c) and (d) present specific image examples, along with the classification results and confidence scores. The goal of GaTA is to optimize the augmented results towards a direction that enhances correct recognition." + ], + "image_footnote": [], + "bbox": [ + 124, + 265, + 187, + 297 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/26d6e417045f58c0a10dbdee232ca86c6154558de6bcc93e4f00920f6e9a0776.jpg", + "image_caption": [ + "CLIP: Vending (0.4041) ResNet: Mail box (0.4871)" + ], + "image_footnote": [], + "bbox": [ + 187, + 265, + 256, + 296 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c59f63926f7d1d79ef71eab639e5bf62f9cb836e8684b747442382b31cee52be.jpg", + "image_caption": [ + "CLIP: Vending (0.4641) ResNet: Mail box (0.2925)" + ], + "image_footnote": [], + "bbox": [ + 256, + 265, + 325, + 296 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9139e1be7ce155ce8663412c56278c1366d6144b50febaa568bd8d5fa1661886.jpg", + "image_caption": [ + "CLIP: Vending (0.4731) ResNet: Mail box (0.2285)" + ], + "image_footnote": [], + "bbox": [ + 326, + 265, + 393, + 296 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/678fdbcbc1f287af8300e0475cc828dd3e182b8164570510446a43f5ca00711d.jpg", + "image_caption": [ + "CLIP: Vending (0.4685) ResNet: Vending (0.2813)" + ], + "image_footnote": [], + "bbox": [ + 395, + 265, + 465, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "semantic features, with the goal of aligning enhanced results with the high-quality image set.", + "bbox": [ + 89, + 414, + 482, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The strength of FoFA lies not only in its ability to preserve semantics but also in its optimization objective, which is designed to enhance compatibility with downstream tasks. In FoFA, the discriminator $D$ is specifically designed to inherit the strong semantic perception capabilities of the multi-foundation models. From another perspective, if the semantic information can no longer distinguish the enhanced result from the high-quality image, the high-level model's perception of the enhanced result will be indistinguishable from that of the high-quality image. Therefore, for downstream tasks, FoFA effectively transforms degraded images into the high-level models' familiar domain.", + "bbox": [ + 89, + 444, + 483, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Gap-driven Task-based Alignment", + "text_level": 1, + "bbox": [ + 89, + 635, + 390, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Observation. FoFA improves the similarity between augmented results and high-quality images at the feature level; however, a constraint is still needed to assess whether the augmented results are easy to recognize. In Section 3.2, we observed that ResNet is more sensitive to illumination compared to CLIP. This raises the question: can we also observe robustness differences between these models at the task level? To explore this, we generated images under different lighting conditions using two methods: (1) fusing low-light and normal-light image pairs from the LSRW dataset (Get the pixel value according to the fusion ratio $\\alpha$ ) (2) applying gamma transformations to images from ImageNet-1k. As shown in Figure 4, low-light conditions degrade the performance of both CLIP and ResNet, but ResNet is more significantly affected. Interestingly, their predictions are most similar under normal conditions, suggesting that CLIP's ro", + "bbox": [ + 89, + 659, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bustness stems from its training on large-scale data, while ResNet is more sensitive to illumination and degradation. This motivates us to leverage these robustness differences to guide the enhancement model to produce clean and well-lit results, aiming at \"detecting\" the degradation and brightness, which can be seen as obstacles in recognition.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Method. Gap-driven Task-based Alignment (GaTA) leverages the robustness differences between models to evaluate whether the augmented results are sufficiently good to be easily recognized. From another view, GaTA employs the less robust model as a \"degradation detector\" while the robust model functions as a ground truth provider. We found that tasks such as segmentation, classification, and detection can be applied to GaTA, with CLIP and ResNet serving as representative robust and degradation-sensitive models, respectively. Let $F^l$ represent the robust model and $F^s$ represent the degradation-sensitive model. Following the observation and robustness study in [22], we can apply CLIP and ResNet for $F^l$ and $F^s$ , respectively. For each enhanced image, we first use the robust model to generate pseudo-labels and obtain the logits from the sensitive model:", + "bbox": [ + 511, + 181, + 906, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY ^ {l} = F ^ {l} (G (x)), \\text {s . t .} x \\sim p _ {\\text {l o w}} (x), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 409, + 903, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY ^ {s} = F ^ {s} (G (x)), \\text {s . t .} x \\sim p _ {\\text {l o w}} (x). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 436, + 903, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A key issue to consider is that pseudo-labels are not always accurate. In extreme cases, if the generation quality of $G(x)$ is very poor, even robust models may fail to provide fully accurate labels. Fortunately, during iterative training, we can collect multiple enhanced results for each training image, constructing a memory bank of size $T$ . Leveraging this memory bank, we can apply a voting mechanism to correct the pseudo labels:", + "bbox": [ + 511, + 457, + 906, + 577 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Y} ^ {l} = \\arg \\max _ {c} \\sum_ {t = 1} ^ {T} \\mathbb {I} \\left[ \\arg \\max _ {c ^ {\\prime}} \\left(p \\left(Y _ {t} ^ {l} = c ^ {\\prime}\\right)\\right) = c \\right], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 583, + 903, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c'$ denotes the predicted class for each individual image while $c$ represents the final pseudo-label. Finally, we apply a high-level task loss to supervise GaTA:", + "bbox": [ + 511, + 626, + 903, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G a T A} = \\mathcal {L} _ {h} \\left(Y ^ {s}, \\hat {Y} ^ {l}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 678, + 903, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_h$ denotes the specific task loss for $Y^{s}$ and $\\hat{Y}^l$ . Please note that both $F_{t}^{l}$ and $F_{t}^{s}$ are frozen during GaTA training. By exploiting the robustness differences between models in a straightforward manner, GaTA effectively tests for degradation at the task level, ensuring the enhanced results exhibit both visual quality and strong performance in downstream applications.", + "bbox": [ + 511, + 704, + 906, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 821, + 645, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 847, + 730, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed method is implemented using PyTorch, running on an NVIDIA Titan RTX GPU alongside an In", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16075", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/582b49ed89bb421ed1d7796b8e894393b3db84e7afecace8f082b77a3dc861ad.jpg", + "table_caption": [ + "Table 1. Benchmarking results for low-light enhancement. Among unsupervised methods, we highlight the top-ranking scores in red and the second in blue." + ], + "table_footnote": [], + "table_body": "
DatasetsMetricsSupervisedUnsupervised
Retinexformer [2] ICCV'23LLflow [24] AAAI'22SKF [27] CVPR'23Kind++ [34] IJCV'21EnGAN [11] TIP'21SCI [21] CVPR'22RUAS [18] CVPR'21ZeroDCE [6] CVPR'20SCL-LLE [15] AAAI'22SGZ [37] WACV'22CoLIE [4] ECCV'24Ours
HuaweiPSNR↑17.9817.1819.3116.5217.8315.447.4116.3813.2116.9113.9618.61
SSIM↑0.610.630.610.500.590.550.600.470.490.560.470.72
LPIPS↓0.320.310.320.400.350.360.420.410.370.350.390.35
NikonPSNR↑16.8718.069.3515.0915.8617.527.7516.6815.1117.3714.9217.76
SSIM↑0.630.540.340.340.580.550.590.580.590.610.550.67
LPIPS↓0.330.370.530.530.370.380.460.390.380.380.390.37
FiveKPSNR↑23.2816.3216.5914.1112.8316.299.5213.5015.7010.6218.7119.58
SSIM↑0.840.770.770.700.780.830.660.730.770.430.820.83
LPIPS↓0.240.190.200.200.140.140.300.230.210.330.160.13
", + "bbox": [ + 94, + 102, + 906, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/faaae026970253fd47e5bc69e2d3aad76b56cb909cee594f1345c9c905874cd4.jpg", + "image_caption": [ + "Figure 5. Visual Comparisons on Low-Light Enhancement Datasets." + ], + "image_footnote": [], + "bbox": [ + 163, + 241, + 831, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tel(R) Xeon(R) Gold 6252 CPU @ 2.10GHz. Training is conducted with the Adam optimizer, using a learning rate of 1e-4. All images are resized to $448 \\times 448$ pixels, and the model is trained with a batch size of 32 over 50 epochs. For ResNet, we utilize the pre-trained weights obtained from training on ImageNet, following the origin setting in [9]. For CLIP and SAM, we employ the official weights trained on their respective datasets. For the training set of high-level applications, we randomly sample low-light images from ExDark [20] and normal-light images from COCO [17] as training set and test the obtained weights on the test sets of all benchmarks. More information can be found in Supplementary.", + "bbox": [ + 88, + 487, + 485, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Benchmarks", + "text_level": 1, + "bbox": [ + 89, + 695, + 225, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The performance of our method is evaluated across four metrics relevant to low-light enhancement: reference-based low-light enhancement, low-light image classification, low-light instance segmentation, and low-light face detection. Specifically, reference-based low-light enhancement is evaluated using the MIT-FiveK [1], Nikon, and Huawei [8] datasets, from which we select 500, 945, and 735 paired low-light and normal-light samples; Low-light image classification uses the CODaN [12] dataset, a 10-class dataset with a training set of 10,000 daytime images and a test set comprising 2,500 daytime and nighttime images. Low-light instance segmentation is performed using", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the LIS [3] dataset with 2230 images. For low-light face detection, we use the darkface [30] dataset, containing 6,000 images captured under low-light conditions, from which we select 1800 images for testing.", + "bbox": [ + 511, + 487, + 906, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To provide a comprehensive comparison, our method is benchmarked against 4 state-of-the-art supervised learning methods, including Retinexformer [2], LLFlow [24], SKF [27], and $\\mathrm{KinD} + +$ [34], as well as 7 unsupervised learning methods, such as Zero-DCE[6], RUAS [18], EnlightGAN [11], SGZ [37], SCL-LLE [15] CoLIE [4] and SCI [21] across all tasks.", + "bbox": [ + 511, + 547, + 908, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Enhancement Tasks", + "text_level": 1, + "bbox": [ + 511, + 665, + 705, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the effectiveness of our method in terms of generalization performance, we evaluated the proposed framework on several benchmark datasets, i.e. MITFiveK [1], Nikon and Huawei [8] datasets, using PSNR, SSIM, and LPIPS as performance metrics. The measurement results are presented in Table 1. It is evident that our proposed method demonstrates advantages compared to other self-supervised approaches, exhibiting competitive performance in terms of PSNR, SSIM, and LPIPS.", + "bbox": [ + 509, + 688, + 906, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 5 illustrates a visual comparison between our method and other approaches. While alternative methods may struggle with incomplete degradation removal or loss of semantic information, our method consistently demonstrates its ability to effectively recover image details. This", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16076", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5491dc1199429ac9b27d26c3c702d322c10865d15115d39100334c4c10832628.jpg", + "table_caption": [ + "Table 2. Performance comparison on different high-level benchmarks: CODaN for classification, Darkface for face detection, and LIS for instance segmentation. Highlighted in red are the top scores, and in blue are the second best methods." + ], + "table_footnote": [], + "table_body": "
DatasetsDetectorsMetricsRawSupervisedUnsupervised
Retformer ICCV'23LLflow AAAI'22SKF CVPR'23Kind++ IJCV'21EnGAN TIP'21SCI CVPR'22RUAS CVPR'21ZeroDCE CVPR'20SCL-LLE AAAI'22SGZ WACV'22CoLIE ECCV'24Ours
CoDANResNet18Acc0.5020.5510.5330.5250.5390.5310.5650.5560.5650.5720.5730.5640.588
EfficientNetAcc0.5870.6140.6530.6160.5590.6300.5730.5350.6050.6760.5910.6520.742
DarkfaceS3FDAP0.6030.6530.6360.6210.6460.6380.6610.6410.6480.6690.6650.6560.683
DSFDAP0.2740.3370.3480.3510.3670.3310.3490.3260.3670.3340.3580.3340.381
LISMaskRCNNmAPbox0.2280.3030.2860.2580.3090.2860.3010.2770.3170.3090.3170.3090.331
mAPseg0.1980.2520.2540.2370.2580.2440.2630.2380.2730.2650.2720.2580.281
", + "bbox": [ + 94, + 102, + 906, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6c9a69dfe25af52ba1697e54d2abb352a7495f164497be130420137958b299e.jpg", + "image_caption": [ + "Figure 6. Visual Comparisons and High-level Results on CODaN, Darkface, and LIS Dataset. The first row shows CODaN and the classification results by ResNet (color denotes incorrect/correct), the second row shows the Darkface detection results by S3FD, and the third row shows the LIS and detection results and Mask-RCNN." + ], + "image_footnote": [], + "bbox": [ + 130, + 215, + 862, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "further substantiates the advanced capabilities of the proposed framework.", + "bbox": [ + 89, + 465, + 482, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. High-level Tasks", + "text_level": 1, + "bbox": [ + 89, + 511, + 253, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Low-Light Classification. We evaluate the adaptability of various methods for low-light image classification tasks on the CODaN [12] dataset. For our experiments, we employ the official ResNet18 model, which was trained on the daytime portion of the dataset, as our primary detector. Additionally, to mitigate the impact of domain differences, we incorporate an EfficientNet [23] model trained on ImageNet, targeting classes corresponding to those in the CO-DaN dataset, as a supplementary detector. Using these two detectors, we assess Top-1 Accuracy as our evaluation metric. As shown in Table 2, our proposed method outperforms other low-light enhancement techniques, demonstrating exceptional performance. This superiority can be attributed to the more consistent illumination provided by the models enhanced with our framework, while the degraded results from other methods hinder classifier performance.", + "bbox": [ + 88, + 536, + 482, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Low-Light Face Detection. We evaluate the performance of two detectors, S3FD [32] and DSFD [14], trained on the WiderFace [28] dataset, and on the DarkFace [30] dataset to assess the performance of various enhancement methods for low-light face detection. It is noteworthy that face detection is more granular and sensitive to semantic information compared to classification tasks; thus, local blurriness can significantly impair performance. Con-", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sequently, the methods like EnGAN [11] and Retinexformer [25] demonstrate a marked decline relative to other tasks, whereas our proposed method consistently achieves excellent results across both detectors. This further underscores the advanced nature of our proposed framework.", + "bbox": [ + 511, + 465, + 906, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Low-Light Instance Segmentation. We evaluated various low-light enhancement methods for their adaptability to low-light instance segmentation using the LIS [3] dataset. As one of the most fine-grained high-level tasks, instance segmentation imposes significant demands on enhancement techniques. We conducted experiments using Mask R-CNN [10] as provided by the official LIS framework, which was initially trained on the COCO dataset and subsequently tested on low-light images from the LIS dataset. Our results demonstrate that the proposed methods achieve superior performance, further confirming that the framework effectively preserves semantic information and enhances the quality of the results.", + "bbox": [ + 511, + 542, + 908, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The above three tasks encompass three datasets, five detectors, and six evaluation metrics, providing a comprehensive assessment of the proposed method's robustness across different low-light scenarios and its applicability to high-level tasks.", + "bbox": [ + 511, + 741, + 906, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 830, + 679, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct extensive ablation studies to validate the efficacy of our proposed perspective and evaluate the contribution of each model component. All eval", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16077", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d9c1a2e2869dfd77e382a3e9dbce16813d9b44151612f1bf03ab5663fdfead0.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 184, + 135 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b6ecb9a356662260a582a28f16767bf05b5b22050fffe47c1aef3e153caa2902.jpg", + "image_caption": [ + "w/o GaTA" + ], + "image_footnote": [], + "bbox": [ + 194, + 88, + 284, + 135 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cefaf037e1235b52e8e32fe1885c1142b28d9042c8246cf1f883af9fac479f5c.jpg", + "image_caption": [ + "GaTA-Seg" + ], + "image_footnote": [], + "bbox": [ + 292, + 88, + 380, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56d563b3d4249538c07741abd4b7983c87f4f1c92075499c68a1bdfbd81edc38.jpg", + "image_caption": [ + "GaTA-Det" + ], + "image_footnote": [], + "bbox": [ + 390, + 88, + 480, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f903bb62ba83887890a6e20f392840afbf52f9e908fa7f960f1d62b33e934e1a.jpg", + "image_caption": [ + "w/o FoFA" + ], + "image_footnote": [], + "bbox": [ + 94, + 148, + 184, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57730f0bf6a3a85ca22f9c70a951f08f15951fc364b6d87f3d807caf92856f80.jpg", + "image_caption": [ + "w/o CLIP" + ], + "image_footnote": [], + "bbox": [ + 194, + 148, + 284, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5afe7dfe6a0977e6b52b5984df0663978765ae6f93f29218c1fa3580f3218162.jpg", + "image_caption": [ + "w/o SAM", + "Figure 7. Visual Comparisons of ablation studies." + ], + "image_footnote": [], + "bbox": [ + 292, + 148, + 380, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b4cabf75829a86d91e682f80cc4e52c2cf0e6cc82e2ddae022d76e61a520e4dc.jpg", + "image_caption": [ + "w/o ResNet" + ], + "image_footnote": [], + "bbox": [ + 390, + 148, + 480, + 194 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "uations are carried out on FiveK for the enhancement task and on three high-level task datasets for other tasks. We select ResNet18/Top-1 Acc, S3FD/AP, and MaskR-CNN/mAP segm as the high-level metrics.", + "bbox": [ + 89, + 234, + 482, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Studies on FoFA. We employed a combination of different models to examine the supervisory capabilities of various foundation models at the feature level in the enhancement process. As shown in Table 3, removing any single foundation model from the FoFA negatively impacts the performance of the enhancement model across various tasks, and removing all of them results in even poorer outcomes. Overall, excluding CLIP and SAM reduces the model's adaptability for high-level tasks, while the absence of ResNet leads to accuracy losses in image enhancement tasks. This is likely because, during FoFA training, CLIP and SAM primarily contribute robust semantic learning, which is critical for high-level tasks; in contrast, ResNet is more responsive to specific degradations and lighting variations, which helps the discriminator perceive visual quality in images. Our observations further validate this theory: as shown in Figure 7, removing ResNet results in significant lighting changes in the enhanced images, while omitting CLIP or SAM introduces blurred and distorted areas. If all foundation models are excluded, FoFA degrades to a conventional adversarial training process, where semantic losses in the enhancement results become more severe. In summary, each foundation model in FoFA contributes positively to the training of the enhancement model, and their integration yields the best enhancement outcomes.", + "bbox": [ + 91, + 296, + 483, + 672 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a52ad51443a35457edb627cba1956fbc4a61eae17caaa123cb204305160d8c6d.jpg", + "table_caption": [ + "Table 3. Quantitative analysis from ablation studies demonstrating the impact of foundation feature space alignment. The best results are marked in bold." + ], + "table_footnote": [], + "table_body": "
PSNRSSIMLPIPSAPmAP~segmTop-1 Acc
w FoFA19.580.830.130.680.280.59
w/o FoFA18.070.810.200.640.240.54
-CLIP18.670.810.150.650.230.53
-SAM18.950.820.170.640.260.54
-ResNet18.240.810.190.670.270.57
", + "bbox": [ + 94, + 723, + 482, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Studies on GaTA. To assess the role of GaTA in training the enhancement model, we first examined the impact of removing GaTA, as shown in Table 4. Removing GaTA negatively affected model performance in both the enhancement and high-level tasks. This is because GaTA primarily functions to detect incomplete degradation", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "elimination in the enhanced results, thereby improving the model's performance across various tasks. As illustrated in Figure 7, removing GaTA significantly increased the severity of degradation in the enhanced results, which not only impaired visual quality but also compromised the original semantic information.", + "bbox": [ + 511, + 90, + 903, + 180 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4cc131984994637129e193190e7bfae0e965cfa457012262f0d17263bb1f2476.jpg", + "table_caption": [ + "Table 4. Quantitative analysis from ablation studies demonstrating the Gap-driven Task-based Alignment. The best results are marked in bold." + ], + "table_footnote": [], + "table_body": "
PSNRSSIMLPIPSAPmAP~segmTop-1 Acc
w GaTA19.580.830.130.680.280.59
w/o GaTA18.010.810.160.670.250.52
+GaTA Seg18.500.790.190.670.300.49
+GaTA Det18.750.820.190.700.270.54
", + "bbox": [ + 517, + 231, + 903, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, in Section 3.3, we implement GaTA using a classification task, with CLIP and ResNet employed to achieve $F^l$ and $F^s$ , respectively. Since GaTA leverages robustness differences between models to perceive degradation, various tasks—such as object detection or semantic segmentation—could potentially be used for GaTA as well. To test this, we trained $F^l$ on a dataset with gamma transformations and noise augmentations, and $F^s$ on the high-quality subset of the dataset, filtered using advanced image quality assessment metric Q-Instruct [26]. In practice, we usedYOLOX [5], trained on COCO, and PSPNet [35], trained on ADE20k [38], to represent models for object detection and semantic segmentation, denote as Det and Seg respectively. Detailed experimental settings are provided in the supplementary materials. As shown in Table 4, both object detection and semantic segmentation tasks effectively implement GaTA, achieving performance comparable to ResNet and CLIP, further demonstrating the generalizability of the proposed method.", + "bbox": [ + 511, + 308, + 906, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 609, + 633, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a framework named FoCo designed to enhance the visual and applicational performance of low-light image enhancers by leveraging the collaborative potential of diverse foundation models. At the feature level, our approach uses Feature Semantic Alignment (FoFA) to integrate semantically relevant features into adversarial training, aligning the discriminator and the foundation model's output features within a unified feature space. This alignment promotes feature-level consistency between enhanced and high-quality images. At the task level, we introduce Gap-driven task-based Alignment (GaTA), which utilizes the robustness disparities across models by using the less robust model as a degradation detector, thereby further improving enhancement quality. The proposed framework exhibits outstanding performance, providing a novel training paradigm that integrates self-supervised low-light enhancement methods with high-level applications.", + "bbox": [ + 511, + 635, + 906, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16078", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This research was funded through Anhui Provincial Natural Science Foundation (Grant 2408085J024), National Natural Science Foundation of China (Grant 62401532), and Fundamental Research Funds for the Central Universities (Grant WK2090000065).", + "bbox": [ + 89, + 90, + 485, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 203 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Fredo Durand. Learning photographic global tonal adjustment with a database of input/output image pairs. In CVPR 2011, pages 97-104. IEEE, 2011. 6", + "[2] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. arXiv preprint arXiv:2303.06705, 2023. 1, 2, 6", + "[3] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. International Journal of Computer Vision, 131(8):2198-2218, 2023. 3, 6, 7", + "[4] Tomáš Chobola, Yu Liu, Hanyi Zhang, Julia A Schnabel, and Tingying Peng. Fast context-based low-light image enhancement via neural implicit representations. In European Conference on Computer Vision, pages 413-430. Springer, 2025. 2, 6", + "[5] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 8", + "[6] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1780-1789, 2020. 1, 2, 6", + "[7] Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE Transactions on image processing, 26(2):982-993, 2016. 1", + "[8] Jiang Hai, Zhu Xuan, Ren Yang, Yutong Hao, Fengzhu Zou, Fang Lin, and Songchen Han. R2rnet: Low-light image enhancement via real-low to real-normal network. Journal of Visual Communication and Image Representation, 90: 103712, 2023. 6", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7", + "[11] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. Enlightengan: Deep light enhancement without paired supervision. IEEE transactions on image processing, 30:2340-2349, 2021. 2, 6, 7", + "[12] Attila Lengyel, Sourav Garg, Michael Milford, and Jan C van Gemert. Zero-shot day-night domain adaptation with a physics prior. In Proceedings of the IEEE/CVF International" + ], + "bbox": [ + 93, + 212, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision, pages 4399-4409, 2021. 6, 7", + "[13] Chongyi Li, Chunle Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4225-4238, 2021. 1", + "[14] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd: dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 7", + "[15] Dong Liang, Ling Li, Mingqiang Wei, Shuo Yang, Liyan Zhang, Wenhan Yang, Yun Du, and Huiyu Zhou. Semantically contrastive learning for low-light image enhancement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1555-1563, 2022. 2, 6", + "[16] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 2", + "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[18] Risheng Liu, Long Ma, Jiao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10561-10570, 2021. 1, 6", + "[19] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 1", + "[20] Yuen Peng Loh and Chee Seng Chan. Getting to know low-light images with the exclusively dark dataset. Computer Vision and Image Understanding, 178:30-42, 2019. 6", + "[21] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 2, 6", + "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[23] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7", + "[24] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, Lap-Pui Chau, and Alex Kot. Low-light image enhancement with" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16079", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "normalizing flow. In Proceedings of the AAAI conference on artificial intelligence, pages 2604-2612, 2022. 6", + "[25] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 1, 2, 7", + "[26] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25490-25500, 2024. 8", + "[27] Yuhui Wu, Chen Pan, Guoqing Wang, Yang Yang, Jiwei Wei, Chongyi Li, and Heng Tao Shen. Learning semantic-aware knowledge guidance for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1662-1671, 2023. 2, 4, 6", + "[28] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5525-5533, 2016. 7", + "[29] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12918-12927, 2023. 2", + "[30] Wenhan Yang, Ye Yuan, Wenqi Ren, Jiaying Liu, Walter J Scheirer, Zhangyang Wang, Taiheng Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, et al. Advancing image understanding in poor visibility environments: A collective benchmark study. IEEE Transactions on Image Processing, 29: 5737-5752, 2020. 6, 7", + "[31] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2023. 1, 2", + "[32] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z Li. S3fd: Single shot scale-invariant face detector. In Proceedings of the IEEE international conference on computer vision, pages 192-201, 2017. 7", + "[33] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Proceedings of the 27th ACM international conference on multimedia, pages 1632-1640, 2019. 1, 2", + "[34] Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision, 129:1013-1037, 2021. 1, 6", + "[35] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 8", + "[36] Chuanjun Zheng, Daming Shi, and Wentian Shi. Adaptive unfolding total variation network for low-light image enhancement. pages 4439-4448, 2021. 1", + "[37] Shen Zheng and Gaurav Gupta. Semantic-guided zero-shot learning for low-light image/video enhancement. In Pro" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 581-590, 2022. 2, 6", + "[38] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 8", + "[39] Han Zhou, Wei Dong, Xiaohong Liu, Shuaicheng Liu, Xiongkuo Min, Guangtao Zhai, and Jun Chen. Glare: Low light image enhancement via generative latent feature based codebook retrieval. arXiv preprint arXiv:2407.12431, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 244 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16080", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/0e31d729-0f45-4249-a70f-af6d4389a18f_model.json b/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/0e31d729-0f45-4249-a70f-af6d4389a18f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f29dd7270cfa6f15da04ef20180d15d9f148fe20 --- /dev/null +++ b/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/0e31d729-0f45-4249-a70f-af6d4389a18f_model.json @@ -0,0 +1,2387 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.13, + 0.892, + 0.175 + ], + "angle": 0, + "content": "Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.203, + 0.756, + 0.222 + ], + "angle": 0, + "content": "Yuxuan Gu\\(^{1}\\), Haoxuan Wang\\(^{1}\\), Pengyang Ling\\(^{1,2}\\), Zhixiang Wei\\(^{1}\\)," + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.222, + 0.672, + 0.239 + ], + "angle": 0, + "content": "Huaian Chen\\(^{1\\dagger}\\), Yi Jin\\(^{1\\dagger}\\), and Enhong Chen\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.239, + 0.812, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)University of Science and Technology of China \\(^{2}\\)Shanghai AI Laboratory" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.324, + 0.486, + 0.658 + ], + "angle": 0, + "content": "In this paper, we observe that the collaboration of various foundation models can perceive semantic and degraded information within images, thereby guiding the low-light enhancement process. Specifically, we propose a self-supervised low-light enhancement framework based on the multiple foundation models collaboration (dubbed FoCo), aimed at improving both the visual quality of enhanced images and the performance in high-level applications. At the feature level, FoCo leverages the rich features from various foundation models to enhance the model's semantic perception during training, thereby reducing the gap between enhanced results and high-quality images from a high-level perspective. At the task level, we exploit the robustness-gap between strong foundation models and weak models, applying high-level task guidance to the low-light enhancement training process. Through the collaboration of multiple foundation models, the proposed framework shows better enhancement performance and adapts better to high-level tasks. Extensive experiments across various enhancement and application benchmarks demonstrate the qualitative and quantitative superiority of the proposed method over numerous state-of-the-art techniques." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.682, + 0.222, + 0.698 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Images taken under short exposure or low-light conditions frequently suffer from significant degradation, including loss of color information and the introduction of substantial noise. Such degradation severely impacts the performance of downstream tasks [19, 36]. Low-light image enhancement (LIE) aims to restore input images captured in low-light environments to high-quality, well-lit images. With the advancement of deep learning technology in computer vision, recent research [2, 25, 31, 33, 34, 39] has achieved low-light image enhancement through supervised training on paired datasets. However, collecting" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.442, + 0.613, + 0.46 + ], + "angle": 0, + "content": "(a) Prior-based Framework" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.46, + 0.648, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.608, + 0.643, + 0.63 + ], + "angle": 0, + "content": "(b) Previous High-for-Low Framework" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.46, + 0.876, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.608, + 0.853, + 0.63 + ], + "angle": 0, + "content": "(c) Proposed Multi-Foundation model Collaboration Framework" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.64, + 0.908, + 0.767 + ], + "angle": 0, + "content": "Figure 1. The proposed approach demonstrates substantial improvements over existing unsupervised methods in terms of enhancement tasks (PSNR, SSIM) and application tasks (classification/Acc, face detection/AP, and instance segmentation/mAP), attributed to the proposed Multi-VFM Collaboration Framework. Compared with previous unsupervised LIE learning paradigms (a) and (b), the proposed framework (c) can make full use of the foundation models from a collaborative perspective to improve visual quality and downstream task performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "the substantial amount of paired images required for these methods is both time-consuming and costly, which highlights the broad potential applications of self-supervised approaches [6, 7, 13, 18]. As shown in Fig. 1 (a), early self-supervised training paradigms mainly relied on manually defined priors, such as smoothness priors, brightness priors, or histogram equalization. The drawback of such prior-based frameworks is their lack of connection with semantic" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.245, + 0.901 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16071" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "information, and the efficacy of these prior functions can diminish under complex natural conditions. For instance, the smoothness prior may sometimes degrade semantic information, limiting these methods' performance in practical applications and impacting the visual quality of results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.482, + 0.38 + ], + "angle": 0, + "content": "Recognizing these limitations, as illustrated in Figure 1 (b), some approaches [15, 16, 27, 29, 37] have introduced high-level models into the training of Low-light Image Enhancement (LIE) models, significantly improving the model's ability to retain semantic information. However, this line of research typically relies on a specific high-level model and does not fully explore the potential of different vision models, resulting in only limited visual improvements. In recent years, the emergence of various foundation vision models widely applied in real-world scenarios has led us to consider integrating multiple foundational models from a collaborative perspective, which aims to improve the visual quality and practical performance of LIE methods more effectively." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.382, + 0.483, + 0.682 + ], + "angle": 0, + "content": "In this work, we explore the use of multiple foundation models to strengthen the self-supervised enhancement process, as shown in Figure 1 (c). To this end, We address two key challenges: (1) how to effectively utilize pre-trained knowledge? (2) how to handle differences between models? To deal with these challenges, we propose FoCo, a Low-light Image Enhancement (LIE) framework based on the multiple foundation models ensemble, aligning the enhancement results with high-quality images from both feature-level and task-level perspectives. As the core strategies to implement FoCo, we introduce two optimization strategies: Foundation Feature-space Alignment (FoFA) and Gap-driven task-based Alignment (GaTA). FoFA focuses on creating a unified feature space for foundation models, using adversarial learning to align enhanced results with high-quality image sets. GaTA leverages robustness differences across models to model degradation scenarios, using robust pseudo-labels to directly incorporate high-level task training into the enhancement model's optimization process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.713 + ], + "angle": 0, + "content": "In summary, the contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.72, + 0.483, + 0.81 + ], + "angle": 0, + "content": "- We are the first to explore the collaborative effects of multiple vision foundation models on the LIE task, and then propose FoCo, a framework designed to enhance both the human- and machine-understanding performance of unsupervised LIE methods by leveraging complementary interactions among multiple foundational vision models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.482, + 0.87 + ], + "angle": 0, + "content": "- We introduce two key strategies FoFA and GaTA to fully exploit the pre-trained knowledge of foundation models, which align enhancement results and high-quality images from the perspective of foundation models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "- We conduct extensive experiments on benchmarks for low-light enhancement, object detection, instance seg" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "mentation, and classification. The experimental results show significant performance improvements, demonstrating the proposed method's effectiveness in enhancing visual quality and application performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.169, + 0.661, + 0.185 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.194, + 0.787, + 0.21 + ], + "angle": 0, + "content": "2.1. Low-light Image Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.217, + 0.905, + 0.427 + ], + "angle": 0, + "content": "Supervised methods, such as RetinexNet [25], KinD [33], and more recent approaches like Retinexformer [2], Diff-Retinex [31], and GLARE [39], utilize paired datasets and advanced model architectures to enhance image quality. While successful, these methods face challenges due to their dependence on large-scale paired datasets, which are difficult to obtain in real-world scenarios. Unsupervised methods, such as EnLightenGAN [11], and CoLIE [4], aim to overcome this limitation by eliminating the need for paired data, leveraging adversarial learning and multimodal approaches. However, these methods often neglect integration with high-level models, limiting their ability to preserve semantic information and adapt to downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.803, + 0.454 + ], + "angle": 0, + "content": "2.2. High-Level guided Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Recent works [15, 16, 27, 29, 37] have increasingly focused on leveraging high-level models to supervise low-light enhancement models, aiming to improve the fidelity of the enhancement results. For example, SGZ [37] incorporates the concept of focal loss to maximize the logits of a pre-trained semantic segmentation model, enhancing the model's ability to preserve semantic information. SKF [27] integrates features from semantic segmentation models into the enhancement model and proposes a semantic-based histogram equalization loss and adversarial loss based on segmentation results. Methods like NeRCo [29] and CLIP-LIT [16] employ pre-trained CLIP models to align enhancement results with high-quality images during adversarial training. The methods above have significant success, but a common limitation of these approaches is their reliance on a specific high-level model, and they often do not fully exploit the model's potential." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.732, + 0.613, + 0.747 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.688, + 0.771 + ], + "angle": 0, + "content": "3.1. FoCo Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To implement the low-light enhancement process, we first define a simple enhancer \\( G \\) for enhancing a low-light input image \\( I \\), and then use some priors to constrain the behavior of \\( G \\). By incorporating the smoothness and consistency losses \\( \\mathcal{L}_C \\) borrowed from [21] and [6], we can construct a baseline low-light enhancement model that fully relies on simply defined hand-crafted prior. However, this dependency on these hand-crafted prior compromises" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16072" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.071, + 0.844, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.353, + 0.908, + 0.427 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed low-light enhancement method, which enhances a basic LIE model using multi-foundation model ensemble. Given the set of pre-trained foundation models, the enhancement model first performs basic enhancement, followed by supervision from the ensemble through two strategies: (1) Feature-Level: adversarial training aligns enhanced results with high-quality images in the unified feature space, and (2) Task-Level: robustness differences between models are used to align the enhancement results with high-quality images on high-level tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.432, + 0.487, + 0.478 + ], + "angle": 0, + "content": "the model's visual quality and adaptability to downstream tasks. Mathematically, such a straightforward image-level enhancement approach can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.485, + 0.483, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\theta} d (G (I), \\hat {I}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.517, + 0.487, + 0.716 + ], + "angle": 0, + "content": "where \\(\\hat{I}\\) represents the ideal normal-light image, \\(\\theta\\) denotes the parameters of the enhancement model, \\(d(\\cdot)\\) is the distance measurement function. To address the limitations of the prior-based approach, we introduce FoCo framework designed to enhance this model from the perspective of multiple foundation models collaboration. Specifically, we propose (1) Foundation Feature-space Alignment (FoFA), at the feature level (indicated in green), and (2) Gap-driven Task-based Alignment (GaTA), at the task level (indicated in blue). These strategies leverage the feature extraction process (denoted as \\(F^f\\)) and task decision process (denoted as \\(F^t\\)) of foundation models. Thus, FoCo's optimization objective can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.721, + 0.485, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\arg \\min _ {\\theta} \\underbrace {d (G (I) , \\hat {I})} _ {\\mathcal {L} _ {C}} + \\underbrace {d \\left(F ^ {f} (G (I)) , F ^ {f} (\\hat {I})\\right)} _ {\\text {F o F A}} \\tag {2} \\\\ + \\underbrace {d (F ^ {t} (G (I)) , F ^ {t} (\\hat {I}))} _ {\\text {G a T A}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.808, + 0.458, + 0.823 + ], + "angle": 0, + "content": "The corresponding framework is illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.414, + 0.849 + ], + "angle": 0, + "content": "3.2. Foundation Feature-space Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Observation. Figure 3 (a) displays a GAN-enhanced low-light image from the LIS [3] dataset, alongside its Mask-RCNN detect result. The GAN was trained in an unpaired" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.909, + 0.853 + ], + "angle": 0, + "content": "manner on the LIS dataset. It can be observed that the GAN-enhanced result suffers from significant semantic degradation, likely due to the discriminator's limited semantic awareness, which fails to guide the generator effectively. This observation leads us to a potential solution: could we enhance the discriminator by incorporating a feature extractor with strong semantic awareness? As shown in Figure 3 (b), we use CLIP, SAM and ResNet separately to extract feature maps from both images. Incorporating these base models into the discriminator training process, as shown in Figure 3 (a), led to significant improvements in detection performance. However, SAM produced darker results, CLIP introduced color bias, and ResNet maintained limited semantic integrity. We hypothesize that these differences stem from the varying focus of each foundational model: CLIP prioritizes the primary objects in the image and is less sensitive to color bias and subtle degradations, SAM captures fine-grained semantics but lacks sensitivity to lighting conditions, while ResNet offers weaker semantic understanding compared to the other models but is more susceptible to degradations and lighting variations, which still contributes to the discriminator's learning. Interestingly, ResNet's lack of robustness to degradations may become a key factor in complementing CLIP and SAM, while the discriminator needs to identify degradation and illumination. As shown in Figure 3 (c), where we observed that ResNet's features were more sensitive to lighting variations compared to the other models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This inspires us that CLIP, SAM, and ResNet may complement each other at the feature level: CLIP highlights key objects, SAM provides fine-grained guidance, and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16073" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.069, + 0.887, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.295, + 0.907, + 0.352 + ], + "angle": 0, + "content": "Figure 3. The perception capabilities of foundation models on unsupervised low-light enhancer training. (a) presents the discriminator feature maps, enhanced results, and instance segmentation outcomes using different foundation model-supervised discriminators. (b) illustrates the process of adversarial training supervised by the foundation model. (c) compares the robustness of CLIP and ResNet feature maps under varying lighting conditions. The results suggest that foundation models may collaborate at feature level." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.36, + 0.482, + 0.451 + ], + "angle": 0, + "content": "ResNet adds degradation-related information to guide semantic learning in the discriminator. However, the feature spaces of CLIP, SAM, and ResNet are unaligned, preventing the discriminator from directly utilizing their outputs. The challenge now is how to effectively integrate CLIP, SAM, and ResNet into the discriminator training process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.483, + 0.661 + ], + "angle": 0, + "content": "Method. Based on this observation, Foundation Feature-space Alignment (FoFA) aims to leverage features from different foundation models to bring the enhanced results closer to a high-quality image set \\(\\tilde{I}\\). Let \\(F_{i,j}\\) denote the feature extractor of the \\(i\\)-th pre-trained foundation models, extending up to the \\(j\\)-th layer, given the varied structures and training data of these models, our goal is to construct a unified feature space \\(\\hat{F}_j\\) and find semantic and degradation information in it that is beneficial to the enhancement process. To this end, we use a series of convolutional projectors \\(P_{i,j}\\) to map the features from different spaces into a unified feature space. To align the numerical scale of these features with the subsequent modules, we normalize the output features, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.668, + 0.483, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {i, j} ^ {f} (x) = \\operatorname {N o r m} \\left(P _ {i, j} \\left(F _ {i, j} ^ {f} (x)\\right)\\right), s. t. x \\sim p (x), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.702, + 0.482, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {j} ^ {f} (x) = \\left\\{\\hat {F} _ {i, j} ^ {f} (x) \\right\\} _ {i = 0} ^ {n}, s. t. x \\sim p (x), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(n\\) denotes the count number of foundation models, \\(P\\) is the convolutional projectors and \\(F_{i,j}^{f}\\) denotes feature extraction with foundation models. Then we adopt an adversarial learning approach by introducing a discriminator \\(D\\), which links the features from the foundation models with the enhancement model. The input to \\(D\\) consists of images along with their corresponding semantic information derived from the unified feature space. Specifically, inspired by [27], we employ cross-attention across both spatial and channel dimensions between the image features and the semantic features to generate attention maps \\(A^c\\) and \\(A^s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.383, + 0.907, + 0.41 + ], + "angle": 0, + "content": "\\[\nA ^ {c} = \\operatorname {S o f t m a x} \\left(P ^ {c, k} \\left(f _ {j} ^ {F}\\right) \\times P ^ {c, q} \\left(f _ {j} ^ {D}\\right) / \\sqrt {C}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.438, + 0.907, + 0.465 + ], + "angle": 0, + "content": "\\[\nA ^ {s} = \\operatorname {S o f t m a x} \\left(P ^ {s, k} \\left(f _ {j} ^ {F}\\right) \\cdot P ^ {s, q} \\left(f _ {j} ^ {D}\\right) / \\sqrt {H \\cdot W}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.906, + 0.525 + ], + "angle": 0, + "content": "where \\( f_{j}^{F} = \\hat{F}_{j}^{f}(x) \\) and \\( f_{j}^{D} \\) is the feature map of \\( j \\)-th layer in \\( D \\). These attention maps are then utilized to adjust the original image features to the output feature \\( f_{j}^{O} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.535, + 0.905, + 0.556 + ], + "angle": 0, + "content": "\\[\nf _ {j} ^ {O} = A ^ {c} \\times A ^ {s} + f _ {j} ^ {D}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.907, + 0.695 + ], + "angle": 0, + "content": "The optimization objective of \\( D \\) is to align its learned features with the outputs of the foundation model while simultaneously distinguishing between enhanced low-quality images and high-quality images. Throughout this process, it is crucial to ensure that the features input to the classification layer of \\( D \\) are closely related to the semantic and degradation information. Based on the observation, the loss function for the discriminator \\( D \\) can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.697, + 0.905, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {D} = - \\mathbb {E} _ {x \\sim p _ {h i g h} (x)} [ \\log D (x) + \\left\\| D ^ {f} (x) - \\hat {F} _ {m} ^ {f} (x) \\right\\| ] \\\\ - \\mathbb {E} _ {x \\sim p l o w (x)} [ \\log (1 - D (G (x))) + \\\\ \\left\\| D ^ {f} (G (x)) - \\hat {F} _ {m} ^ {f} (G (x)) \\right\\| ], \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.787, + 0.905, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G} = - \\mathbb {E} _ {x \\sim p _ {\\text {l o w}} (x)} [ \\log D (G (x)) ], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(m\\) denotes the layer count of the foundation model, \\(D^{f}\\) denotes feature extraction with discriminator. The discriminator \\(D\\) effectively projects the foundation models' features into a unified feature space, providing valuable guidance for the enhancement model. Unlike traditional image-level adversarial training, FoFA focuses on selected" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16074" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.287, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.189, + 0.209, + 0.198 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.089, + 0.47, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.189, + 0.391, + 0.198 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.228, + 0.117, + 0.238 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.202, + 0.188, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.202, + 0.257, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.236, + 0.257, + 0.263 + ], + "angle": 0, + "content": "CLIP: Ambulance (0.9067) ResNet:Police van (0.5195)" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.202, + 0.326, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.236, + 0.321, + 0.263 + ], + "angle": 0, + "content": "CLIP: Ambulance (0.9438) ResNet: Ambulance (0.7087)" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.202, + 0.394, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.236, + 0.389, + 0.263 + ], + "angle": 0, + "content": "CLIP: Ambulance (0.9131) \nResNet: Police van (0.6643)" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.202, + 0.466, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.401, + 0.236, + 0.459, + 0.263 + ], + "angle": 0, + "content": "CLIP: Ambulance (0.9092) ResNet:Police van (0.8759)" + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.294, + 0.116, + 0.303 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.266, + 0.188, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.3, + 0.179, + 0.326 + ], + "angle": 0, + "content": "CLIP: Vending (0.3428) \nResNet: Mail box (0.3349)" + }, + { + "type": "image", + "bbox": [ + 0.189, + 0.266, + 0.257, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.3, + 0.249, + 0.326 + ], + "angle": 0, + "content": "CLIP: Vending (0.4041) ResNet: Mail box (0.4871)" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.266, + 0.326, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.299, + 0.317, + 0.326 + ], + "angle": 0, + "content": "CLIP: Vending (0.4641) ResNet: Mail box (0.2925)" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.266, + 0.394, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.299, + 0.385, + 0.326 + ], + "angle": 0, + "content": "CLIP: Vending (0.4731) ResNet: Mail box (0.2285)" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.266, + 0.466, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.401, + 0.299, + 0.453, + 0.326 + ], + "angle": 0, + "content": "CLIP: Vending (0.4685) ResNet: Vending (0.2813)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.329, + 0.483, + 0.412 + ], + "angle": 0, + "content": "Figure 4. Robustness to illumination across models. (a) and (b) show the similarity of results between CLIP and ResNet under different lighting adjustments. (c) and (d) present specific image examples, along with the classification results and confidence scores. The goal of GaTA is to optimize the augmented results towards a direction that enhances correct recognition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.445 + ], + "angle": 0, + "content": "semantic features, with the goal of aligning enhanced results with the high-quality image set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.484, + 0.626 + ], + "angle": 0, + "content": "The strength of FoFA lies not only in its ability to preserve semantics but also in its optimization objective, which is designed to enhance compatibility with downstream tasks. In FoFA, the discriminator \\( D \\) is specifically designed to inherit the strong semantic perception capabilities of the multi-foundation models. From another perspective, if the semantic information can no longer distinguish the enhanced result from the high-quality image, the high-level model's perception of the enhanced result will be indistinguishable from that of the high-quality image. Therefore, for downstream tasks, FoFA effectively transforms degraded images into the high-level models' familiar domain." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.392, + 0.652 + ], + "angle": 0, + "content": "3.3. Gap-driven Task-based Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Observation. FoFA improves the similarity between augmented results and high-quality images at the feature level; however, a constraint is still needed to assess whether the augmented results are easy to recognize. In Section 3.2, we observed that ResNet is more sensitive to illumination compared to CLIP. This raises the question: can we also observe robustness differences between these models at the task level? To explore this, we generated images under different lighting conditions using two methods: (1) fusing low-light and normal-light image pairs from the LSRW dataset (Get the pixel value according to the fusion ratio \\(\\alpha\\)) (2) applying gamma transformations to images from ImageNet-1k. As shown in Figure 4, low-light conditions degrade the performance of both CLIP and ResNet, but ResNet is more significantly affected. Interestingly, their predictions are most similar under normal conditions, suggesting that CLIP's ro" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "bustness stems from its training on large-scale data, while ResNet is more sensitive to illumination and degradation. This motivates us to leverage these robustness differences to guide the enhancement model to produce clean and well-lit results, aiming at \"detecting\" the degradation and brightness, which can be seen as obstacles in recognition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.182, + 0.907, + 0.408 + ], + "angle": 0, + "content": "Method. Gap-driven Task-based Alignment (GaTA) leverages the robustness differences between models to evaluate whether the augmented results are sufficiently good to be easily recognized. From another view, GaTA employs the less robust model as a \"degradation detector\" while the robust model functions as a ground truth provider. We found that tasks such as segmentation, classification, and detection can be applied to GaTA, with CLIP and ResNet serving as representative robust and degradation-sensitive models, respectively. Let \\( F^l \\) represent the robust model and \\( F^s \\) represent the degradation-sensitive model. Following the observation and robustness study in [22], we can apply CLIP and ResNet for \\( F^l \\) and \\( F^s \\), respectively. For each enhanced image, we first use the robust model to generate pseudo-labels and obtain the logits from the sensitive model:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.41, + 0.905, + 0.428 + ], + "angle": 0, + "content": "\\[\nY ^ {l} = F ^ {l} (G (x)), \\text {s . t .} x \\sim p _ {\\text {l o w}} (x), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.438, + 0.905, + 0.455 + ], + "angle": 0, + "content": "\\[\nY ^ {s} = F ^ {s} (G (x)), \\text {s . t .} x \\sim p _ {\\text {l o w}} (x). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.578 + ], + "angle": 0, + "content": "A key issue to consider is that pseudo-labels are not always accurate. In extreme cases, if the generation quality of \\( G(x) \\) is very poor, even robust models may fail to provide fully accurate labels. Fortunately, during iterative training, we can collect multiple enhanced results for each training image, constructing a memory bank of size \\( T \\). Leveraging this memory bank, we can apply a voting mechanism to correct the pseudo labels:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.584, + 0.905, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\hat {Y} ^ {l} = \\arg \\max _ {c} \\sum_ {t = 1} ^ {T} \\mathbb {I} \\left[ \\arg \\max _ {c ^ {\\prime}} \\left(p \\left(Y _ {t} ^ {l} = c ^ {\\prime}\\right)\\right) = c \\right], \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.905, + 0.673 + ], + "angle": 0, + "content": "where \\( c' \\) denotes the predicted class for each individual image while \\( c \\) represents the final pseudo-label. Finally, we apply a high-level task loss to supervise GaTA:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.679, + 0.905, + 0.698 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G a T A} = \\mathcal {L} _ {h} \\left(Y ^ {s}, \\hat {Y} ^ {l}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.812 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_h\\) denotes the specific task loss for \\(Y^{s}\\) and \\(\\hat{Y}^l\\). Please note that both \\(F_{t}^{l}\\) and \\(F_{t}^{s}\\) are frozen during GaTA training. By exploiting the robustness differences between models in a straightforward manner, GaTA effectively tests for degradation at the task level, ensuring the enhanced results exhibit both visual quality and strong performance in downstream applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.823, + 0.646, + 0.839 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.731, + 0.864 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The proposed method is implemented using PyTorch, running on an NVIDIA Titan RTX GPU alongside an In" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16075" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.072, + 0.908, + 0.099 + ], + "angle": 0, + "content": "Table 1. Benchmarking results for low-light enhancement. Among unsupervised methods, we highlight the top-ranking scores in red and the second in blue." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.103, + 0.907, + 0.239 + ], + "angle": 0, + "content": "
DatasetsMetricsSupervisedUnsupervised
Retinexformer [2] ICCV'23LLflow [24] AAAI'22SKF [27] CVPR'23Kind++ [34] IJCV'21EnGAN [11] TIP'21SCI [21] CVPR'22RUAS [18] CVPR'21ZeroDCE [6] CVPR'20SCL-LLE [15] AAAI'22SGZ [37] WACV'22CoLIE [4] ECCV'24Ours
HuaweiPSNR↑17.9817.1819.3116.5217.8315.447.4116.3813.2116.9113.9618.61
SSIM↑0.610.630.610.500.590.550.600.470.490.560.470.72
LPIPS↓0.320.310.320.400.350.360.420.410.370.350.390.35
NikonPSNR↑16.8718.069.3515.0915.8617.527.7516.6815.1117.3714.9217.76
SSIM↑0.630.540.340.340.580.550.590.580.590.610.550.67
LPIPS↓0.330.370.530.530.370.380.460.390.380.380.390.37
FiveKPSNR↑23.2816.3216.5914.1112.8316.299.5213.5015.7010.6218.7119.58
SSIM↑0.840.770.770.700.780.830.660.730.770.430.820.83
LPIPS↓0.240.190.200.200.140.140.300.230.210.330.160.13
" + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.242, + 0.832, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.466, + 0.7, + 0.478 + ], + "angle": 0, + "content": "Figure 5. Visual Comparisons on Low-Light Enhancement Datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.488, + 0.486, + 0.685 + ], + "angle": 0, + "content": "tel(R) Xeon(R) Gold 6252 CPU @ 2.10GHz. Training is conducted with the Adam optimizer, using a learning rate of 1e-4. All images are resized to \\(448 \\times 448\\) pixels, and the model is trained with a batch size of 32 over 50 epochs. For ResNet, we utilize the pre-trained weights obtained from training on ImageNet, following the origin setting in [9]. For CLIP and SAM, we employ the official weights trained on their respective datasets. For the training set of high-level applications, we randomly sample low-light images from ExDark [20] and normal-light images from COCO [17] as training set and test the obtained weights on the test sets of all benchmarks. More information can be found in Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.226, + 0.71 + ], + "angle": 0, + "content": "4.2. Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.902 + ], + "angle": 0, + "content": "The performance of our method is evaluated across four metrics relevant to low-light enhancement: reference-based low-light enhancement, low-light image classification, low-light instance segmentation, and low-light face detection. Specifically, reference-based low-light enhancement is evaluated using the MIT-FiveK [1], Nikon, and Huawei [8] datasets, from which we select 500, 945, and 735 paired low-light and normal-light samples; Low-light image classification uses the CODaN [12] dataset, a 10-class dataset with a training set of 10,000 daytime images and a test set comprising 2,500 daytime and nighttime images. Low-light instance segmentation is performed using" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.548 + ], + "angle": 0, + "content": "the LIS [3] dataset with 2230 images. For low-light face detection, we use the darkface [30] dataset, containing 6,000 images captured under low-light conditions, from which we select 1800 images for testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.909, + 0.655 + ], + "angle": 0, + "content": "To provide a comprehensive comparison, our method is benchmarked against 4 state-of-the-art supervised learning methods, including Retinexformer [2], LLFlow [24], SKF [27], and \\(\\mathrm{KinD} + +\\) [34], as well as 7 unsupervised learning methods, such as Zero-DCE[6], RUAS [18], EnlightGAN [11], SGZ [37], SCL-LLE [15] CoLIE [4] and SCI [21] across all tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.707, + 0.68 + ], + "angle": 0, + "content": "4.3. Enhancement Tasks" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.907, + 0.825 + ], + "angle": 0, + "content": "To validate the effectiveness of our method in terms of generalization performance, we evaluated the proposed framework on several benchmark datasets, i.e. MITFiveK [1], Nikon and Huawei [8] datasets, using PSNR, SSIM, and LPIPS as performance metrics. The measurement results are presented in Table 1. It is evident that our proposed method demonstrates advantages compared to other self-supervised approaches, exhibiting competitive performance in terms of PSNR, SSIM, and LPIPS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Figure 5 illustrates a visual comparison between our method and other approaches. While alternative methods may struggle with incomplete degradation removal or loss of semantic information, our method consistently demonstrates its ability to effectively recover image details. This" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16076" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.072, + 0.908, + 0.1 + ], + "angle": 0, + "content": "Table 2. Performance comparison on different high-level benchmarks: CODaN for classification, Darkface for face detection, and LIS for instance segmentation. Highlighted in red are the top scores, and in blue are the second best methods." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.103, + 0.907, + 0.214 + ], + "angle": 0, + "content": "
DatasetsDetectorsMetricsRawSupervisedUnsupervised
Retformer ICCV'23LLflow AAAI'22SKF CVPR'23Kind++ IJCV'21EnGAN TIP'21SCI CVPR'22RUAS CVPR'21ZeroDCE CVPR'20SCL-LLE AAAI'22SGZ WACV'22CoLIE ECCV'24Ours
CoDANResNet18Acc0.5020.5510.5330.5250.5390.5310.5650.5560.5650.5720.5730.5640.588
EfficientNetAcc0.5870.6140.6530.6160.5590.6300.5730.5350.6050.6760.5910.6520.742
DarkfaceS3FDAP0.6030.6530.6360.6210.6460.6380.6610.6410.6480.6690.6650.6560.683
DSFDAP0.2740.3370.3480.3510.3670.3310.3490.3260.3670.3340.3580.3340.381
LISMaskRCNNmAPbox0.2280.3030.2860.2580.3090.2860.3010.2770.3170.3090.3170.3090.331
mAPseg0.1980.2520.2540.2370.2580.2440.2630.2380.2730.2650.2720.2580.281
" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.217, + 0.864, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.421, + 0.908, + 0.462 + ], + "angle": 0, + "content": "Figure 6. Visual Comparisons and High-level Results on CODaN, Darkface, and LIS Dataset. The first row shows CODaN and the classification results by ResNet (color denotes incorrect/correct), the second row shows the Darkface detection results by S3FD, and the third row shows the LIS and detection results and Mask-RCNN." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.483, + 0.499 + ], + "angle": 0, + "content": "further substantiates the advanced capabilities of the proposed framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.254, + 0.529 + ], + "angle": 0, + "content": "4.4. High-level Tasks" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.537, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Low-Light Classification. We evaluate the adaptability of various methods for low-light image classification tasks on the CODaN [12] dataset. For our experiments, we employ the official ResNet18 model, which was trained on the daytime portion of the dataset, as our primary detector. Additionally, to mitigate the impact of domain differences, we incorporate an EfficientNet [23] model trained on ImageNet, targeting classes corresponding to those in the CO-DaN dataset, as a supplementary detector. Using these two detectors, we assess Top-1 Accuracy as our evaluation metric. As shown in Table 2, our proposed method outperforms other low-light enhancement techniques, demonstrating exceptional performance. This superiority can be attributed to the more consistent illumination provided by the models enhanced with our framework, while the degraded results from other methods hinder classifier performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Low-Light Face Detection. We evaluate the performance of two detectors, S3FD [32] and DSFD [14], trained on the WiderFace [28] dataset, and on the DarkFace [30] dataset to assess the performance of various enhancement methods for low-light face detection. It is noteworthy that face detection is more granular and sensitive to semantic information compared to classification tasks; thus, local blurriness can significantly impair performance. Con-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.543 + ], + "angle": 0, + "content": "sequently, the methods like EnGAN [11] and Retinexformer [25] demonstrate a marked decline relative to other tasks, whereas our proposed method consistently achieves excellent results across both detectors. This further underscores the advanced nature of our proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.909, + 0.74 + ], + "angle": 0, + "content": "Low-Light Instance Segmentation. We evaluated various low-light enhancement methods for their adaptability to low-light instance segmentation using the LIS [3] dataset. As one of the most fine-grained high-level tasks, instance segmentation imposes significant demands on enhancement techniques. We conducted experiments using Mask R-CNN [10] as provided by the official LIS framework, which was initially trained on the COCO dataset and subsequently tested on low-light images from the LIS dataset. Our results demonstrate that the proposed methods achieve superior performance, further confirming that the framework effectively preserves semantic information and enhances the quality of the results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.742, + 0.907, + 0.817 + ], + "angle": 0, + "content": "The above three tasks encompass three datasets, five detectors, and six evaluation metrics, providing a comprehensive assessment of the proposed method's robustness across different low-light scenarios and its applicability to high-level tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.681, + 0.847 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "We conduct extensive ablation studies to validate the efficacy of our proposed perspective and evaluate the contribution of each model component. All eval" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16077" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.186, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.138, + 0.154, + 0.147 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.195, + 0.089, + 0.285, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.137, + 0.266, + 0.147 + ], + "angle": 0, + "content": "w/o GaTA" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.089, + 0.382, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.138, + 0.366, + 0.147 + ], + "angle": 0, + "content": "GaTA-Seg" + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.089, + 0.481, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.138, + 0.465, + 0.147 + ], + "angle": 0, + "content": "GaTA-Det" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.149, + 0.185, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.196, + 0.166, + 0.205 + ], + "angle": 0, + "content": "w/o FoFA" + }, + { + "type": "image", + "bbox": [ + 0.195, + 0.149, + 0.285, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.213, + 0.196, + 0.264, + 0.205 + ], + "angle": 0, + "content": "w/o CLIP" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.149, + 0.382, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.196, + 0.361, + 0.205 + ], + "angle": 0, + "content": "w/o SAM" + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.149, + 0.481, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.196, + 0.466, + 0.205 + ], + "angle": 0, + "content": "w/o ResNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.215, + 0.434, + 0.229 + ], + "angle": 0, + "content": "Figure 7. Visual Comparisons of ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.235, + 0.483, + 0.296 + ], + "angle": 0, + "content": "uations are carried out on FiveK for the enhancement task and on three high-level task datasets for other tasks. We select ResNet18/Top-1 Acc, S3FD/AP, and MaskR-CNN/mAP segm as the high-level metrics." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.297, + 0.485, + 0.673 + ], + "angle": 0, + "content": "Ablation Studies on FoFA. We employed a combination of different models to examine the supervisory capabilities of various foundation models at the feature level in the enhancement process. As shown in Table 3, removing any single foundation model from the FoFA negatively impacts the performance of the enhancement model across various tasks, and removing all of them results in even poorer outcomes. Overall, excluding CLIP and SAM reduces the model's adaptability for high-level tasks, while the absence of ResNet leads to accuracy losses in image enhancement tasks. This is likely because, during FoFA training, CLIP and SAM primarily contribute robust semantic learning, which is critical for high-level tasks; in contrast, ResNet is more responsive to specific degradations and lighting variations, which helps the discriminator perceive visual quality in images. Our observations further validate this theory: as shown in Figure 7, removing ResNet results in significant lighting changes in the enhanced images, while omitting CLIP or SAM introduces blurred and distorted areas. If all foundation models are excluded, FoFA degrades to a conventional adversarial training process, where semantic losses in the enhancement results become more severe. In summary, each foundation model in FoFA contributes positively to the training of the enhancement model, and their integration yields the best enhancement outcomes." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.678, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Table 3. Quantitative analysis from ablation studies demonstrating the impact of foundation feature space alignment. The best results are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.724, + 0.483, + 0.806 + ], + "angle": 0, + "content": "
PSNRSSIMLPIPSAPmAP~segmTop-1 Acc
w FoFA19.580.830.130.680.280.59
w/o FoFA18.070.810.200.640.240.54
-CLIP18.670.810.150.650.230.53
-SAM18.950.820.170.640.260.54
-ResNet18.240.810.190.670.270.57
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Ablation Studies on GaTA. To assess the role of GaTA in training the enhancement model, we first examined the impact of removing GaTA, as shown in Table 4. Removing GaTA negatively affected model performance in both the enhancement and high-level tasks. This is because GaTA primarily functions to detect incomplete degradation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "elimination in the enhanced results, thereby improving the model's performance across various tasks. As illustrated in Figure 7, removing GaTA significantly increased the severity of degradation in the enhanced results, which not only impaired visual quality but also compromised the original semantic information." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.186, + 0.907, + 0.227 + ], + "angle": 0, + "content": "Table 4. Quantitative analysis from ablation studies demonstrating the Gap-driven Task-based Alignment. The best results are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.232, + 0.905, + 0.302 + ], + "angle": 0, + "content": "
PSNRSSIMLPIPSAPmAP~segmTop-1 Acc
w GaTA19.580.830.130.680.280.59
w/o GaTA18.010.810.160.670.250.52
+GaTA Seg18.500.790.190.670.300.49
+GaTA Det18.750.820.190.700.270.54
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.907, + 0.597 + ], + "angle": 0, + "content": "Additionally, in Section 3.3, we implement GaTA using a classification task, with CLIP and ResNet employed to achieve \\( F^l \\) and \\( F^s \\), respectively. Since GaTA leverages robustness differences between models to perceive degradation, various tasks—such as object detection or semantic segmentation—could potentially be used for GaTA as well. To test this, we trained \\( F^l \\) on a dataset with gamma transformations and noise augmentations, and \\( F^s \\) on the high-quality subset of the dataset, filtered using advanced image quality assessment metric Q-Instruct [26]. In practice, we usedYOLOX [5], trained on COCO, and PSPNet [35], trained on ADE20k [38], to represent models for object detection and semantic segmentation, denote as Det and Seg respectively. Detailed experimental settings are provided in the supplementary materials. As shown in Table 4, both object detection and semantic segmentation tasks effectively implement GaTA, achieving performance comparable to ResNet and CLIP, further demonstrating the generalizability of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.61, + 0.634, + 0.626 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.907, + 0.892 + ], + "angle": 0, + "content": "We propose a framework named FoCo designed to enhance the visual and applicational performance of low-light image enhancers by leveraging the collaborative potential of diverse foundation models. At the feature level, our approach uses Feature Semantic Alignment (FoFA) to integrate semantically relevant features into adversarial training, aligning the discriminator and the foundation model's output features within a unified feature space. This alignment promotes feature-level consistency between enhanced and high-quality images. At the task level, we introduce Gap-driven task-based Alignment (GaTA), which utilizes the robustness disparities across models by using the less robust model as a degradation detector, thereby further improving enhancement quality. The proposed framework exhibits outstanding performance, providing a novel training paradigm that integrates self-supervised low-light enhancement methods with high-level applications." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16078" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.162 + ], + "angle": 0, + "content": "Acknowledgements: This research was funded through Anhui Provincial Natural Science Foundation (Grant 2408085J024), National Natural Science Foundation of China (Grant 62401532), and Fundamental Research Funds for the Central Universities (Grant WK2090000065)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.188, + 0.188, + 0.204 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.213, + 0.482, + 0.268 + ], + "angle": 0, + "content": "[1] Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Fredo Durand. Learning photographic global tonal adjustment with a database of input/output image pairs. In CVPR 2011, pages 97-104. IEEE, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.271, + 0.482, + 0.325 + ], + "angle": 0, + "content": "[2] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. arXiv preprint arXiv:2303.06705, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[3] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. International Journal of Computer Vision, 131(8):2198-2218, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.482, + 0.45 + ], + "angle": 0, + "content": "[4] Tomáš Chobola, Yu Liu, Hanyi Zhang, Julia A Schnabel, and Tingying Peng. Fast context-based low-light image enhancement via neural implicit representations. In European Conference on Computer Vision, pages 413-430. Springer, 2025. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[5] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.482, + 0.576 + ], + "angle": 0, + "content": "[6] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1780-1789, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[7] Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE Transactions on image processing, 26(2):982-993, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.482, + 0.689 + ], + "angle": 0, + "content": "[8] Jiang Hai, Zhu Xuan, Ren Yang, Yutong Hao, Fengzhu Zou, Fang Lin, and Songchen Han. R2rnet: Low-light image enhancement via real-low to real-normal network. Journal of Visual Communication and Image Representation, 90: 103712, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[11] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. Enlightengan: Deep light enhancement without paired supervision. IEEE transactions on image processing, 30:2340-2349, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Attila Lengyel, Sourav Garg, Michael Milford, and Jan C van Gemert. Zero-shot day-night domain adaptation with a physics prior. In Proceedings of the IEEE/CVF International" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.213, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Conference on Computer Vision, pages 4399-4409, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[13] Chongyi Li, Chunle Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4225-4238, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[14] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd: dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[15] Dong Liang, Ling Li, Mingqiang Wei, Shuo Yang, Liyan Zhang, Wenhan Yang, Yun Du, and Huiyu Zhou. Semantically contrastive learning for low-light image enhancement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1555-1563, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[16] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.387, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.471, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[18] Risheng Liu, Long Ma, Jiao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10561-10570, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[19] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[20] Yuen Peng Loh and Chee Seng Chan. Getting to know low-light images with the exclusively dark dataset. Computer Vision and Image Understanding, 178:30-42, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[21] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, Lap-Pui Chau, and Alex Kot. Low-light image enhancement with" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16079" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "normalizing flow. In Proceedings of the AAAI conference on artificial intelligence, pages 2604-2612, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[25] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[26] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25490-25500, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[27] Yuhui Wu, Chen Pan, Guoqing Wang, Yang Yang, Jiwei Wei, Chongyi Li, and Heng Tao Shen. Learning semantic-aware knowledge guidance for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1662-1671, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[28] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5525-5533, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[29] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12918-12927, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[30] Wenhan Yang, Ye Yuan, Wenqi Ren, Jiaying Liu, Walter J Scheirer, Zhangyang Wang, Taiheng Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, et al. Advancing image understanding in poor visibility environments: A collective benchmark study. IEEE Transactions on Image Processing, 29: 5737-5752, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[31] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[32] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z Li. S3fd: Single shot scale-invariant face detector. In Proceedings of the IEEE international conference on computer vision, pages 192-201, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.674, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[33] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Proceedings of the 27th ACM international conference on multimedia, pages 1632-1640, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[34] Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision, 129:1013-1037, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[35] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[36] Chuanjun Zheng, Daming Shi, and Wentian Shi. Adaptive unfolding total variation network for low-light image enhancement. pages 4439-4448, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Shen Zheng and Gaurav Gupta. Semantic-guided zero-shot learning for low-light image/video enhancement. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 581-590, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[38] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[39] Han Zhou, Wei Dong, Xiaohong Liu, Shuaicheng Liu, Xiongkuo Min, Guangtao Zhai, and Jun Chen. Glare: Low light image enhancement via generative latent feature based codebook retrieval. arXiv preprint arXiv:2407.12431, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16080" + } + ] +] \ No newline at end of file diff --git a/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/0e31d729-0f45-4249-a70f-af6d4389a18f_origin.pdf b/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/0e31d729-0f45-4249-a70f-af6d4389a18f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..360c0c9f9151dc54d41fe7d9fa7e756a75e0ce39 --- /dev/null +++ b/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/0e31d729-0f45-4249-a70f-af6d4389a18f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6cd756955224067176e23d00bb7d85e1019a837045f023cd51047723b6e1e1e +size 6510934 diff --git a/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/full.md b/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3d289a492a4d0830d0a59b26570e59f2cff7d93a --- /dev/null +++ b/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/full.md @@ -0,0 +1,350 @@ +# Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration + +Yuxuan Gu $^{1}$ , Haoxuan Wang $^{1}$ , Pengyang Ling $^{1,2}$ , Zhixiang Wei $^{1}$ , + +Huaian Chen $^{1\dagger}$ , Yi Jin $^{1\dagger}$ , and Enhong Chen $^{1}$ + +$^{1}$ University of Science and Technology of China $^{2}$ Shanghai AI Laboratory + +# Abstract + +In this paper, we observe that the collaboration of various foundation models can perceive semantic and degraded information within images, thereby guiding the low-light enhancement process. Specifically, we propose a self-supervised low-light enhancement framework based on the multiple foundation models collaboration (dubbed FoCo), aimed at improving both the visual quality of enhanced images and the performance in high-level applications. At the feature level, FoCo leverages the rich features from various foundation models to enhance the model's semantic perception during training, thereby reducing the gap between enhanced results and high-quality images from a high-level perspective. At the task level, we exploit the robustness-gap between strong foundation models and weak models, applying high-level task guidance to the low-light enhancement training process. Through the collaboration of multiple foundation models, the proposed framework shows better enhancement performance and adapts better to high-level tasks. Extensive experiments across various enhancement and application benchmarks demonstrate the qualitative and quantitative superiority of the proposed method over numerous state-of-the-art techniques. + +# 1. Introduction + +Images taken under short exposure or low-light conditions frequently suffer from significant degradation, including loss of color information and the introduction of substantial noise. Such degradation severely impacts the performance of downstream tasks [19, 36]. Low-light image enhancement (LIE) aims to restore input images captured in low-light environments to high-quality, well-lit images. With the advancement of deep learning technology in computer vision, recent research [2, 25, 31, 33, 34, 39] has achieved low-light image enhancement through supervised training on paired datasets. However, collecting + +![](images/0e934dfe43f4f9e781317759ad13a83d9c27db5c07d05b43b14c4df1c02f3717.jpg) +(a) Prior-based Framework + +![](images/fb14190bedc8083fdfe466f7f5b1bf49064a3c2525be1a386033979f7869ab65.jpg) +(b) Previous High-for-Low Framework +Figure 1. The proposed approach demonstrates substantial improvements over existing unsupervised methods in terms of enhancement tasks (PSNR, SSIM) and application tasks (classification/Acc, face detection/AP, and instance segmentation/mAP), attributed to the proposed Multi-VFM Collaboration Framework. Compared with previous unsupervised LIE learning paradigms (a) and (b), the proposed framework (c) can make full use of the foundation models from a collaborative perspective to improve visual quality and downstream task performance. + +![](images/c16ed0f454b6d418c106b0188de6dfaee2a297b5b123ca6679001322b239b424.jpg) +(c) Proposed Multi-Foundation model Collaboration Framework + +the substantial amount of paired images required for these methods is both time-consuming and costly, which highlights the broad potential applications of self-supervised approaches [6, 7, 13, 18]. As shown in Fig. 1 (a), early self-supervised training paradigms mainly relied on manually defined priors, such as smoothness priors, brightness priors, or histogram equalization. The drawback of such prior-based frameworks is their lack of connection with semantic + +information, and the efficacy of these prior functions can diminish under complex natural conditions. For instance, the smoothness prior may sometimes degrade semantic information, limiting these methods' performance in practical applications and impacting the visual quality of results. + +Recognizing these limitations, as illustrated in Figure 1 (b), some approaches [15, 16, 27, 29, 37] have introduced high-level models into the training of Low-light Image Enhancement (LIE) models, significantly improving the model's ability to retain semantic information. However, this line of research typically relies on a specific high-level model and does not fully explore the potential of different vision models, resulting in only limited visual improvements. In recent years, the emergence of various foundation vision models widely applied in real-world scenarios has led us to consider integrating multiple foundational models from a collaborative perspective, which aims to improve the visual quality and practical performance of LIE methods more effectively. + +In this work, we explore the use of multiple foundation models to strengthen the self-supervised enhancement process, as shown in Figure 1 (c). To this end, We address two key challenges: (1) how to effectively utilize pre-trained knowledge? (2) how to handle differences between models? To deal with these challenges, we propose FoCo, a Low-light Image Enhancement (LIE) framework based on the multiple foundation models ensemble, aligning the enhancement results with high-quality images from both feature-level and task-level perspectives. As the core strategies to implement FoCo, we introduce two optimization strategies: Foundation Feature-space Alignment (FoFA) and Gap-driven task-based Alignment (GaTA). FoFA focuses on creating a unified feature space for foundation models, using adversarial learning to align enhanced results with high-quality image sets. GaTA leverages robustness differences across models to model degradation scenarios, using robust pseudo-labels to directly incorporate high-level task training into the enhancement model's optimization process. + +In summary, the contributions of this paper are as follows: + +- We are the first to explore the collaborative effects of multiple vision foundation models on the LIE task, and then propose FoCo, a framework designed to enhance both the human- and machine-understanding performance of unsupervised LIE methods by leveraging complementary interactions among multiple foundational vision models. +- We introduce two key strategies FoFA and GaTA to fully exploit the pre-trained knowledge of foundation models, which align enhancement results and high-quality images from the perspective of foundation models. +- We conduct extensive experiments on benchmarks for low-light enhancement, object detection, instance seg + +mentation, and classification. The experimental results show significant performance improvements, demonstrating the proposed method's effectiveness in enhancing visual quality and application performance. + +# 2. Related works + +# 2.1. Low-light Image Enhancement + +Supervised methods, such as RetinexNet [25], KinD [33], and more recent approaches like Retinexformer [2], Diff-Retinex [31], and GLARE [39], utilize paired datasets and advanced model architectures to enhance image quality. While successful, these methods face challenges due to their dependence on large-scale paired datasets, which are difficult to obtain in real-world scenarios. Unsupervised methods, such as EnLightenGAN [11], and CoLIE [4], aim to overcome this limitation by eliminating the need for paired data, leveraging adversarial learning and multimodal approaches. However, these methods often neglect integration with high-level models, limiting their ability to preserve semantic information and adapt to downstream tasks. + +# 2.2. High-Level guided Enhancement + +Recent works [15, 16, 27, 29, 37] have increasingly focused on leveraging high-level models to supervise low-light enhancement models, aiming to improve the fidelity of the enhancement results. For example, SGZ [37] incorporates the concept of focal loss to maximize the logits of a pre-trained semantic segmentation model, enhancing the model's ability to preserve semantic information. SKF [27] integrates features from semantic segmentation models into the enhancement model and proposes a semantic-based histogram equalization loss and adversarial loss based on segmentation results. Methods like NeRCo [29] and CLIP-LIT [16] employ pre-trained CLIP models to align enhancement results with high-quality images during adversarial training. The methods above have significant success, but a common limitation of these approaches is their reliance on a specific high-level model, and they often do not fully exploit the model's potential. + +# 3. Methods + +# 3.1. FoCo Framework + +To implement the low-light enhancement process, we first define a simple enhancer $G$ for enhancing a low-light input image $I$ , and then use some priors to constrain the behavior of $G$ . By incorporating the smoothness and consistency losses $\mathcal{L}_C$ borrowed from [21] and [6], we can construct a baseline low-light enhancement model that fully relies on simply defined hand-crafted prior. However, this dependency on these hand-crafted prior compromises + +![](images/1c0ed6892d501810b49c536f7e15911915db94ad47e55d16f81fa4211f8a5c49.jpg) +Figure 2. Overview of the proposed low-light enhancement method, which enhances a basic LIE model using multi-foundation model ensemble. Given the set of pre-trained foundation models, the enhancement model first performs basic enhancement, followed by supervision from the ensemble through two strategies: (1) Feature-Level: adversarial training aligns enhanced results with high-quality images in the unified feature space, and (2) Task-Level: robustness differences between models are used to align the enhancement results with high-quality images on high-level tasks. + +the model's visual quality and adaptability to downstream tasks. Mathematically, such a straightforward image-level enhancement approach can be formalized as: + +$$ +\arg \min _ {\theta} d (G (I), \hat {I}), \tag {1} +$$ + +where $\hat{I}$ represents the ideal normal-light image, $\theta$ denotes the parameters of the enhancement model, $d(\cdot)$ is the distance measurement function. To address the limitations of the prior-based approach, we introduce FoCo framework designed to enhance this model from the perspective of multiple foundation models collaboration. Specifically, we propose (1) Foundation Feature-space Alignment (FoFA), at the feature level (indicated in green), and (2) Gap-driven Task-based Alignment (GaTA), at the task level (indicated in blue). These strategies leverage the feature extraction process (denoted as $F^f$ ) and task decision process (denoted as $F^t$ ) of foundation models. Thus, FoCo's optimization objective can be expressed as: + +$$ +\begin{array}{l} \arg \min _ {\theta} \underbrace {d (G (I) , \hat {I})} _ {\mathcal {L} _ {C}} + \underbrace {d \left(F ^ {f} (G (I)) , F ^ {f} (\hat {I})\right)} _ {\text {F o F A}} \tag {2} \\ + \underbrace {d (F ^ {t} (G (I)) , F ^ {t} (\hat {I}))} _ {\text {G a T A}}. \\ \end{array} +$$ + +The corresponding framework is illustrated in Figure 2. + +# 3.2. Foundation Feature-space Alignment + +Observation. Figure 3 (a) displays a GAN-enhanced low-light image from the LIS [3] dataset, alongside its Mask-RCNN detect result. The GAN was trained in an unpaired + +manner on the LIS dataset. It can be observed that the GAN-enhanced result suffers from significant semantic degradation, likely due to the discriminator's limited semantic awareness, which fails to guide the generator effectively. This observation leads us to a potential solution: could we enhance the discriminator by incorporating a feature extractor with strong semantic awareness? As shown in Figure 3 (b), we use CLIP, SAM and ResNet separately to extract feature maps from both images. Incorporating these base models into the discriminator training process, as shown in Figure 3 (a), led to significant improvements in detection performance. However, SAM produced darker results, CLIP introduced color bias, and ResNet maintained limited semantic integrity. We hypothesize that these differences stem from the varying focus of each foundational model: CLIP prioritizes the primary objects in the image and is less sensitive to color bias and subtle degradations, SAM captures fine-grained semantics but lacks sensitivity to lighting conditions, while ResNet offers weaker semantic understanding compared to the other models but is more susceptible to degradations and lighting variations, which still contributes to the discriminator's learning. Interestingly, ResNet's lack of robustness to degradations may become a key factor in complementing CLIP and SAM, while the discriminator needs to identify degradation and illumination. As shown in Figure 3 (c), where we observed that ResNet's features were more sensitive to lighting variations compared to the other models. + +This inspires us that CLIP, SAM, and ResNet may complement each other at the feature level: CLIP highlights key objects, SAM provides fine-grained guidance, and + +![](images/3a6c20a317ba9e5d8715c3b5c2a04d789db5794c45e7f989e3fb2320b078be63.jpg) +Figure 3. The perception capabilities of foundation models on unsupervised low-light enhancer training. (a) presents the discriminator feature maps, enhanced results, and instance segmentation outcomes using different foundation model-supervised discriminators. (b) illustrates the process of adversarial training supervised by the foundation model. (c) compares the robustness of CLIP and ResNet feature maps under varying lighting conditions. The results suggest that foundation models may collaborate at feature level. + +ResNet adds degradation-related information to guide semantic learning in the discriminator. However, the feature spaces of CLIP, SAM, and ResNet are unaligned, preventing the discriminator from directly utilizing their outputs. The challenge now is how to effectively integrate CLIP, SAM, and ResNet into the discriminator training process. + +Method. Based on this observation, Foundation Feature-space Alignment (FoFA) aims to leverage features from different foundation models to bring the enhanced results closer to a high-quality image set $\tilde{I}$ . Let $F_{i,j}$ denote the feature extractor of the $i$ -th pre-trained foundation models, extending up to the $j$ -th layer, given the varied structures and training data of these models, our goal is to construct a unified feature space $\hat{F}_j$ and find semantic and degradation information in it that is beneficial to the enhancement process. To this end, we use a series of convolutional projectors $P_{i,j}$ to map the features from different spaces into a unified feature space. To align the numerical scale of these features with the subsequent modules, we normalize the output features, defined as: + +$$ +\hat {F} _ {i, j} ^ {f} (x) = \operatorname {N o r m} \left(P _ {i, j} \left(F _ {i, j} ^ {f} (x)\right)\right), s. t. x \sim p (x), \tag {3} +$$ + +$$ +\hat {F} _ {j} ^ {f} (x) = \left\{\hat {F} _ {i, j} ^ {f} (x) \right\} _ {i = 0} ^ {n}, s. t. x \sim p (x), \tag {4} +$$ + +where $n$ denotes the count number of foundation models, $P$ is the convolutional projectors and $F_{i,j}^{f}$ denotes feature extraction with foundation models. Then we adopt an adversarial learning approach by introducing a discriminator $D$ , which links the features from the foundation models with the enhancement model. The input to $D$ consists of images along with their corresponding semantic information derived from the unified feature space. Specifically, inspired by [27], we employ cross-attention across both spatial and channel dimensions between the image features and the semantic features to generate attention maps $A^c$ and $A^s$ : + +$$ +A ^ {c} = \operatorname {S o f t m a x} \left(P ^ {c, k} \left(f _ {j} ^ {F}\right) \times P ^ {c, q} \left(f _ {j} ^ {D}\right) / \sqrt {C}\right), \tag {5} +$$ + +$$ +A ^ {s} = \operatorname {S o f t m a x} \left(P ^ {s, k} \left(f _ {j} ^ {F}\right) \cdot P ^ {s, q} \left(f _ {j} ^ {D}\right) / \sqrt {H \cdot W}\right), \tag {6} +$$ + +where $f_{j}^{F} = \hat{F}_{j}^{f}(x)$ and $f_{j}^{D}$ is the feature map of $j$ -th layer in $D$ . These attention maps are then utilized to adjust the original image features to the output feature $f_{j}^{O}$ as follows: + +$$ +f _ {j} ^ {O} = A ^ {c} \times A ^ {s} + f _ {j} ^ {D}. \tag {7} +$$ + +The optimization objective of $D$ is to align its learned features with the outputs of the foundation model while simultaneously distinguishing between enhanced low-quality images and high-quality images. Throughout this process, it is crucial to ensure that the features input to the classification layer of $D$ are closely related to the semantic and degradation information. Based on the observation, the loss function for the discriminator $D$ can be expressed as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {D} = - \mathbb {E} _ {x \sim p _ {h i g h} (x)} [ \log D (x) + \left\| D ^ {f} (x) - \hat {F} _ {m} ^ {f} (x) \right\| ] \\ - \mathbb {E} _ {x \sim p l o w (x)} [ \log (1 - D (G (x))) + \\ \left\| D ^ {f} (G (x)) - \hat {F} _ {m} ^ {f} (G (x)) \right\| ], \tag {8} \\ \end{array} +$$ + +$$ +\mathcal {L} _ {G} = - \mathbb {E} _ {x \sim p _ {\text {l o w}} (x)} [ \log D (G (x)) ], \tag {9} +$$ + +where $m$ denotes the layer count of the foundation model, $D^{f}$ denotes feature extraction with discriminator. The discriminator $D$ effectively projects the foundation models' features into a unified feature space, providing valuable guidance for the enhancement model. Unlike traditional image-level adversarial training, FoFA focuses on selected + +![](images/612497201aabd6991f1b5e21ef40aebc6e075b34e8b38d6f1b01b872e486ab60.jpg) +(a) + +![](images/c03be5fbfaaa1b3d9c95d550950f886214dd553d87110430e16d3c72f6e842dc.jpg) +(b) + +![](images/ea0c9e4b74b05934b9104f09644c89c74fe9b06918915a325d4fd82179b02218.jpg) +(c) + +![](images/1ff69a92396fdf9be25bc22649b5b30fc8dd68a418326e433e0dc7343676ec9d.jpg) +CLIP: Ambulance (0.9067) ResNet:Police van (0.5195) +CLIP: Ambulance (0.9438) ResNet: Ambulance (0.7087) + +![](images/2741c073732cfd083e50fff79a1626fcea3bf35433b889383c166e0d91d4bfc9.jpg) + +![](images/2c8b7680ba91aff6582fd8b3dfd5cda23e68441cef75e453264688574367a6d7.jpg) +CLIP: Ambulance (0.9131) +ResNet: Police van (0.6643) + +![](images/0e08270354025a7607191e1b6f7b14b596325df1a926a40d715efd9259dcb482.jpg) +CLIP: Ambulance (0.9092) ResNet:Police van (0.8759) + +![](images/3e92ccc07a0bd255bc1eab55626fdd4c761ef0875a1e1b5d92391d07480e7dcf.jpg) +(d) +CLIP: Vending (0.3428) +ResNet: Mail box (0.3349) +Figure 4. Robustness to illumination across models. (a) and (b) show the similarity of results between CLIP and ResNet under different lighting adjustments. (c) and (d) present specific image examples, along with the classification results and confidence scores. The goal of GaTA is to optimize the augmented results towards a direction that enhances correct recognition. + +![](images/26d6e417045f58c0a10dbdee232ca86c6154558de6bcc93e4f00920f6e9a0776.jpg) +CLIP: Vending (0.4041) ResNet: Mail box (0.4871) + +![](images/c59f63926f7d1d79ef71eab639e5bf62f9cb836e8684b747442382b31cee52be.jpg) +CLIP: Vending (0.4641) ResNet: Mail box (0.2925) + +![](images/9139e1be7ce155ce8663412c56278c1366d6144b50febaa568bd8d5fa1661886.jpg) +CLIP: Vending (0.4731) ResNet: Mail box (0.2285) + +![](images/678fdbcbc1f287af8300e0475cc828dd3e182b8164570510446a43f5ca00711d.jpg) +CLIP: Vending (0.4685) ResNet: Vending (0.2813) + +semantic features, with the goal of aligning enhanced results with the high-quality image set. + +The strength of FoFA lies not only in its ability to preserve semantics but also in its optimization objective, which is designed to enhance compatibility with downstream tasks. In FoFA, the discriminator $D$ is specifically designed to inherit the strong semantic perception capabilities of the multi-foundation models. From another perspective, if the semantic information can no longer distinguish the enhanced result from the high-quality image, the high-level model's perception of the enhanced result will be indistinguishable from that of the high-quality image. Therefore, for downstream tasks, FoFA effectively transforms degraded images into the high-level models' familiar domain. + +# 3.3. Gap-driven Task-based Alignment + +Observation. FoFA improves the similarity between augmented results and high-quality images at the feature level; however, a constraint is still needed to assess whether the augmented results are easy to recognize. In Section 3.2, we observed that ResNet is more sensitive to illumination compared to CLIP. This raises the question: can we also observe robustness differences between these models at the task level? To explore this, we generated images under different lighting conditions using two methods: (1) fusing low-light and normal-light image pairs from the LSRW dataset (Get the pixel value according to the fusion ratio $\alpha$ ) (2) applying gamma transformations to images from ImageNet-1k. As shown in Figure 4, low-light conditions degrade the performance of both CLIP and ResNet, but ResNet is more significantly affected. Interestingly, their predictions are most similar under normal conditions, suggesting that CLIP's ro + +bustness stems from its training on large-scale data, while ResNet is more sensitive to illumination and degradation. This motivates us to leverage these robustness differences to guide the enhancement model to produce clean and well-lit results, aiming at "detecting" the degradation and brightness, which can be seen as obstacles in recognition. + +Method. Gap-driven Task-based Alignment (GaTA) leverages the robustness differences between models to evaluate whether the augmented results are sufficiently good to be easily recognized. From another view, GaTA employs the less robust model as a "degradation detector" while the robust model functions as a ground truth provider. We found that tasks such as segmentation, classification, and detection can be applied to GaTA, with CLIP and ResNet serving as representative robust and degradation-sensitive models, respectively. Let $F^l$ represent the robust model and $F^s$ represent the degradation-sensitive model. Following the observation and robustness study in [22], we can apply CLIP and ResNet for $F^l$ and $F^s$ , respectively. For each enhanced image, we first use the robust model to generate pseudo-labels and obtain the logits from the sensitive model: + +$$ +Y ^ {l} = F ^ {l} (G (x)), \text {s . t .} x \sim p _ {\text {l o w}} (x), \tag {10} +$$ + +$$ +Y ^ {s} = F ^ {s} (G (x)), \text {s . t .} x \sim p _ {\text {l o w}} (x). \tag {11} +$$ + +A key issue to consider is that pseudo-labels are not always accurate. In extreme cases, if the generation quality of $G(x)$ is very poor, even robust models may fail to provide fully accurate labels. Fortunately, during iterative training, we can collect multiple enhanced results for each training image, constructing a memory bank of size $T$ . Leveraging this memory bank, we can apply a voting mechanism to correct the pseudo labels: + +$$ +\hat {Y} ^ {l} = \arg \max _ {c} \sum_ {t = 1} ^ {T} \mathbb {I} \left[ \arg \max _ {c ^ {\prime}} \left(p \left(Y _ {t} ^ {l} = c ^ {\prime}\right)\right) = c \right], \tag {12} +$$ + +where $c'$ denotes the predicted class for each individual image while $c$ represents the final pseudo-label. Finally, we apply a high-level task loss to supervise GaTA: + +$$ +\mathcal {L} _ {G a T A} = \mathcal {L} _ {h} \left(Y ^ {s}, \hat {Y} ^ {l}\right), \tag {13} +$$ + +where $\mathcal{L}_h$ denotes the specific task loss for $Y^{s}$ and $\hat{Y}^l$ . Please note that both $F_{t}^{l}$ and $F_{t}^{s}$ are frozen during GaTA training. By exploiting the robustness differences between models in a straightforward manner, GaTA effectively tests for degradation at the task level, ensuring the enhanced results exhibit both visual quality and strong performance in downstream applications. + +# 4. Experiments + +# 4.1. Implementation Details + +The proposed method is implemented using PyTorch, running on an NVIDIA Titan RTX GPU alongside an In + +Table 1. Benchmarking results for low-light enhancement. Among unsupervised methods, we highlight the top-ranking scores in red and the second in blue. + +
DatasetsMetricsSupervisedUnsupervised
Retinexformer [2] ICCV'23LLflow [24] AAAI'22SKF [27] CVPR'23Kind++ [34] IJCV'21EnGAN [11] TIP'21SCI [21] CVPR'22RUAS [18] CVPR'21ZeroDCE [6] CVPR'20SCL-LLE [15] AAAI'22SGZ [37] WACV'22CoLIE [4] ECCV'24Ours
HuaweiPSNR↑17.9817.1819.3116.5217.8315.447.4116.3813.2116.9113.9618.61
SSIM↑0.610.630.610.500.590.550.600.470.490.560.470.72
LPIPS↓0.320.310.320.400.350.360.420.410.370.350.390.35
NikonPSNR↑16.8718.069.3515.0915.8617.527.7516.6815.1117.3714.9217.76
SSIM↑0.630.540.340.340.580.550.590.580.590.610.550.67
LPIPS↓0.330.370.530.530.370.380.460.390.380.380.390.37
FiveKPSNR↑23.2816.3216.5914.1112.8316.299.5213.5015.7010.6218.7119.58
SSIM↑0.840.770.770.700.780.830.660.730.770.430.820.83
LPIPS↓0.240.190.200.200.140.140.300.230.210.330.160.13
+ +![](images/faaae026970253fd47e5bc69e2d3aad76b56cb909cee594f1345c9c905874cd4.jpg) +Figure 5. Visual Comparisons on Low-Light Enhancement Datasets. + +tel(R) Xeon(R) Gold 6252 CPU @ 2.10GHz. Training is conducted with the Adam optimizer, using a learning rate of 1e-4. All images are resized to $448 \times 448$ pixels, and the model is trained with a batch size of 32 over 50 epochs. For ResNet, we utilize the pre-trained weights obtained from training on ImageNet, following the origin setting in [9]. For CLIP and SAM, we employ the official weights trained on their respective datasets. For the training set of high-level applications, we randomly sample low-light images from ExDark [20] and normal-light images from COCO [17] as training set and test the obtained weights on the test sets of all benchmarks. More information can be found in Supplementary. + +# 4.2. Benchmarks + +The performance of our method is evaluated across four metrics relevant to low-light enhancement: reference-based low-light enhancement, low-light image classification, low-light instance segmentation, and low-light face detection. Specifically, reference-based low-light enhancement is evaluated using the MIT-FiveK [1], Nikon, and Huawei [8] datasets, from which we select 500, 945, and 735 paired low-light and normal-light samples; Low-light image classification uses the CODaN [12] dataset, a 10-class dataset with a training set of 10,000 daytime images and a test set comprising 2,500 daytime and nighttime images. Low-light instance segmentation is performed using + +the LIS [3] dataset with 2230 images. For low-light face detection, we use the darkface [30] dataset, containing 6,000 images captured under low-light conditions, from which we select 1800 images for testing. + +To provide a comprehensive comparison, our method is benchmarked against 4 state-of-the-art supervised learning methods, including Retinexformer [2], LLFlow [24], SKF [27], and $\mathrm{KinD} + +$ [34], as well as 7 unsupervised learning methods, such as Zero-DCE[6], RUAS [18], EnlightGAN [11], SGZ [37], SCL-LLE [15] CoLIE [4] and SCI [21] across all tasks. + +# 4.3. Enhancement Tasks + +To validate the effectiveness of our method in terms of generalization performance, we evaluated the proposed framework on several benchmark datasets, i.e. MITFiveK [1], Nikon and Huawei [8] datasets, using PSNR, SSIM, and LPIPS as performance metrics. The measurement results are presented in Table 1. It is evident that our proposed method demonstrates advantages compared to other self-supervised approaches, exhibiting competitive performance in terms of PSNR, SSIM, and LPIPS. + +Figure 5 illustrates a visual comparison between our method and other approaches. While alternative methods may struggle with incomplete degradation removal or loss of semantic information, our method consistently demonstrates its ability to effectively recover image details. This + +Table 2. Performance comparison on different high-level benchmarks: CODaN for classification, Darkface for face detection, and LIS for instance segmentation. Highlighted in red are the top scores, and in blue are the second best methods. + +
DatasetsDetectorsMetricsRawSupervisedUnsupervised
Retformer ICCV'23LLflow AAAI'22SKF CVPR'23Kind++ IJCV'21EnGAN TIP'21SCI CVPR'22RUAS CVPR'21ZeroDCE CVPR'20SCL-LLE AAAI'22SGZ WACV'22CoLIE ECCV'24Ours
CoDANResNet18Acc0.5020.5510.5330.5250.5390.5310.5650.5560.5650.5720.5730.5640.588
EfficientNetAcc0.5870.6140.6530.6160.5590.6300.5730.5350.6050.6760.5910.6520.742
DarkfaceS3FDAP0.6030.6530.6360.6210.6460.6380.6610.6410.6480.6690.6650.6560.683
DSFDAP0.2740.3370.3480.3510.3670.3310.3490.3260.3670.3340.3580.3340.381
LISMaskRCNNmAPbox0.2280.3030.2860.2580.3090.2860.3010.2770.3170.3090.3170.3090.331
mAPseg0.1980.2520.2540.2370.2580.2440.2630.2380.2730.2650.2720.2580.281
+ +![](images/e6c9a69dfe25af52ba1697e54d2abb352a7495f164497be130420137958b299e.jpg) +Figure 6. Visual Comparisons and High-level Results on CODaN, Darkface, and LIS Dataset. The first row shows CODaN and the classification results by ResNet (color denotes incorrect/correct), the second row shows the Darkface detection results by S3FD, and the third row shows the LIS and detection results and Mask-RCNN. + +further substantiates the advanced capabilities of the proposed framework. + +# 4.4. High-level Tasks + +Low-Light Classification. We evaluate the adaptability of various methods for low-light image classification tasks on the CODaN [12] dataset. For our experiments, we employ the official ResNet18 model, which was trained on the daytime portion of the dataset, as our primary detector. Additionally, to mitigate the impact of domain differences, we incorporate an EfficientNet [23] model trained on ImageNet, targeting classes corresponding to those in the CO-DaN dataset, as a supplementary detector. Using these two detectors, we assess Top-1 Accuracy as our evaluation metric. As shown in Table 2, our proposed method outperforms other low-light enhancement techniques, demonstrating exceptional performance. This superiority can be attributed to the more consistent illumination provided by the models enhanced with our framework, while the degraded results from other methods hinder classifier performance. + +Low-Light Face Detection. We evaluate the performance of two detectors, S3FD [32] and DSFD [14], trained on the WiderFace [28] dataset, and on the DarkFace [30] dataset to assess the performance of various enhancement methods for low-light face detection. It is noteworthy that face detection is more granular and sensitive to semantic information compared to classification tasks; thus, local blurriness can significantly impair performance. Con- + +sequently, the methods like EnGAN [11] and Retinexformer [25] demonstrate a marked decline relative to other tasks, whereas our proposed method consistently achieves excellent results across both detectors. This further underscores the advanced nature of our proposed framework. + +Low-Light Instance Segmentation. We evaluated various low-light enhancement methods for their adaptability to low-light instance segmentation using the LIS [3] dataset. As one of the most fine-grained high-level tasks, instance segmentation imposes significant demands on enhancement techniques. We conducted experiments using Mask R-CNN [10] as provided by the official LIS framework, which was initially trained on the COCO dataset and subsequently tested on low-light images from the LIS dataset. Our results demonstrate that the proposed methods achieve superior performance, further confirming that the framework effectively preserves semantic information and enhances the quality of the results. + +The above three tasks encompass three datasets, five detectors, and six evaluation metrics, providing a comprehensive assessment of the proposed method's robustness across different low-light scenarios and its applicability to high-level tasks. + +# 4.5. Ablation Studies + +We conduct extensive ablation studies to validate the efficacy of our proposed perspective and evaluate the contribution of each model component. All eval + +![](images/2d9c1a2e2869dfd77e382a3e9dbce16813d9b44151612f1bf03ab5663fdfead0.jpg) +Ours + +![](images/b6ecb9a356662260a582a28f16767bf05b5b22050fffe47c1aef3e153caa2902.jpg) +w/o GaTA + +![](images/cefaf037e1235b52e8e32fe1885c1142b28d9042c8246cf1f883af9fac479f5c.jpg) +GaTA-Seg + +![](images/56d563b3d4249538c07741abd4b7983c87f4f1c92075499c68a1bdfbd81edc38.jpg) +GaTA-Det + +![](images/f903bb62ba83887890a6e20f392840afbf52f9e908fa7f960f1d62b33e934e1a.jpg) +w/o FoFA + +![](images/57730f0bf6a3a85ca22f9c70a951f08f15951fc364b6d87f3d807caf92856f80.jpg) +w/o CLIP + +![](images/5afe7dfe6a0977e6b52b5984df0663978765ae6f93f29218c1fa3580f3218162.jpg) +w/o SAM +Figure 7. Visual Comparisons of ablation studies. + +![](images/b4cabf75829a86d91e682f80cc4e52c2cf0e6cc82e2ddae022d76e61a520e4dc.jpg) +w/o ResNet + +uations are carried out on FiveK for the enhancement task and on three high-level task datasets for other tasks. We select ResNet18/Top-1 Acc, S3FD/AP, and MaskR-CNN/mAP segm as the high-level metrics. + +Ablation Studies on FoFA. We employed a combination of different models to examine the supervisory capabilities of various foundation models at the feature level in the enhancement process. As shown in Table 3, removing any single foundation model from the FoFA negatively impacts the performance of the enhancement model across various tasks, and removing all of them results in even poorer outcomes. Overall, excluding CLIP and SAM reduces the model's adaptability for high-level tasks, while the absence of ResNet leads to accuracy losses in image enhancement tasks. This is likely because, during FoFA training, CLIP and SAM primarily contribute robust semantic learning, which is critical for high-level tasks; in contrast, ResNet is more responsive to specific degradations and lighting variations, which helps the discriminator perceive visual quality in images. Our observations further validate this theory: as shown in Figure 7, removing ResNet results in significant lighting changes in the enhanced images, while omitting CLIP or SAM introduces blurred and distorted areas. If all foundation models are excluded, FoFA degrades to a conventional adversarial training process, where semantic losses in the enhancement results become more severe. In summary, each foundation model in FoFA contributes positively to the training of the enhancement model, and their integration yields the best enhancement outcomes. + +Table 3. Quantitative analysis from ablation studies demonstrating the impact of foundation feature space alignment. The best results are marked in bold. + +
PSNRSSIMLPIPSAPmAP~segmTop-1 Acc
w FoFA19.580.830.130.680.280.59
w/o FoFA18.070.810.200.640.240.54
-CLIP18.670.810.150.650.230.53
-SAM18.950.820.170.640.260.54
-ResNet18.240.810.190.670.270.57
+ +Ablation Studies on GaTA. To assess the role of GaTA in training the enhancement model, we first examined the impact of removing GaTA, as shown in Table 4. Removing GaTA negatively affected model performance in both the enhancement and high-level tasks. This is because GaTA primarily functions to detect incomplete degradation + +elimination in the enhanced results, thereby improving the model's performance across various tasks. As illustrated in Figure 7, removing GaTA significantly increased the severity of degradation in the enhanced results, which not only impaired visual quality but also compromised the original semantic information. + +Table 4. Quantitative analysis from ablation studies demonstrating the Gap-driven Task-based Alignment. The best results are marked in bold. + +
PSNRSSIMLPIPSAPmAP~segmTop-1 Acc
w GaTA19.580.830.130.680.280.59
w/o GaTA18.010.810.160.670.250.52
+GaTA Seg18.500.790.190.670.300.49
+GaTA Det18.750.820.190.700.270.54
+ +Additionally, in Section 3.3, we implement GaTA using a classification task, with CLIP and ResNet employed to achieve $F^l$ and $F^s$ , respectively. Since GaTA leverages robustness differences between models to perceive degradation, various tasks—such as object detection or semantic segmentation—could potentially be used for GaTA as well. To test this, we trained $F^l$ on a dataset with gamma transformations and noise augmentations, and $F^s$ on the high-quality subset of the dataset, filtered using advanced image quality assessment metric Q-Instruct [26]. In practice, we usedYOLOX [5], trained on COCO, and PSPNet [35], trained on ADE20k [38], to represent models for object detection and semantic segmentation, denote as Det and Seg respectively. Detailed experimental settings are provided in the supplementary materials. As shown in Table 4, both object detection and semantic segmentation tasks effectively implement GaTA, achieving performance comparable to ResNet and CLIP, further demonstrating the generalizability of the proposed method. + +# 5. Conclusion + +We propose a framework named FoCo designed to enhance the visual and applicational performance of low-light image enhancers by leveraging the collaborative potential of diverse foundation models. At the feature level, our approach uses Feature Semantic Alignment (FoFA) to integrate semantically relevant features into adversarial training, aligning the discriminator and the foundation model's output features within a unified feature space. This alignment promotes feature-level consistency between enhanced and high-quality images. At the task level, we introduce Gap-driven task-based Alignment (GaTA), which utilizes the robustness disparities across models by using the less robust model as a degradation detector, thereby further improving enhancement quality. The proposed framework exhibits outstanding performance, providing a novel training paradigm that integrates self-supervised low-light enhancement methods with high-level applications. + +Acknowledgements: This research was funded through Anhui Provincial Natural Science Foundation (Grant 2408085J024), National Natural Science Foundation of China (Grant 62401532), and Fundamental Research Funds for the Central Universities (Grant WK2090000065). + +# References + +[1] Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Fredo Durand. Learning photographic global tonal adjustment with a database of input/output image pairs. In CVPR 2011, pages 97-104. IEEE, 2011. 6 +[2] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. arXiv preprint arXiv:2303.06705, 2023. 1, 2, 6 +[3] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. International Journal of Computer Vision, 131(8):2198-2218, 2023. 3, 6, 7 +[4] Tomáš Chobola, Yu Liu, Hanyi Zhang, Julia A Schnabel, and Tingying Peng. Fast context-based low-light image enhancement via neural implicit representations. In European Conference on Computer Vision, pages 413-430. Springer, 2025. 2, 6 +[5] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 8 +[6] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1780-1789, 2020. 1, 2, 6 +[7] Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE Transactions on image processing, 26(2):982-993, 2016. 1 +[8] Jiang Hai, Zhu Xuan, Ren Yang, Yutong Hao, Fengzhu Zou, Fang Lin, and Songchen Han. R2rnet: Low-light image enhancement via real-low to real-normal network. Journal of Visual Communication and Image Representation, 90: 103712, 2023. 6 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7 +[11] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. Enlightengan: Deep light enhancement without paired supervision. IEEE transactions on image processing, 30:2340-2349, 2021. 2, 6, 7 +[12] Attila Lengyel, Sourav Garg, Michael Milford, and Jan C van Gemert. Zero-shot day-night domain adaptation with a physics prior. In Proceedings of the IEEE/CVF International + +Conference on Computer Vision, pages 4399-4409, 2021. 6, 7 +[13] Chongyi Li, Chunle Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4225-4238, 2021. 1 +[14] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd: dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 7 +[15] Dong Liang, Ling Li, Mingqiang Wei, Shuo Yang, Liyan Zhang, Wenhan Yang, Yun Du, and Huiyu Zhou. Semantically contrastive learning for low-light image enhancement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1555-1563, 2022. 2, 6 +[16] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 2 +[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[18] Risheng Liu, Long Ma, Jiao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10561-10570, 2021. 1, 6 +[19] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 1 +[20] Yuen Peng Loh and Chee Seng Chan. Getting to know low-light images with the exclusively dark dataset. Computer Vision and Image Understanding, 178:30-42, 2019. 6 +[21] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 2, 6 +[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[23] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7 +[24] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, Lap-Pui Chau, and Alex Kot. Low-light image enhancement with + +normalizing flow. In Proceedings of the AAAI conference on artificial intelligence, pages 2604-2612, 2022. 6 +[25] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 1, 2, 7 +[26] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25490-25500, 2024. 8 +[27] Yuhui Wu, Chen Pan, Guoqing Wang, Yang Yang, Jiwei Wei, Chongyi Li, and Heng Tao Shen. Learning semantic-aware knowledge guidance for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1662-1671, 2023. 2, 4, 6 +[28] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5525-5533, 2016. 7 +[29] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12918-12927, 2023. 2 +[30] Wenhan Yang, Ye Yuan, Wenqi Ren, Jiaying Liu, Walter J Scheirer, Zhangyang Wang, Taiheng Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, et al. Advancing image understanding in poor visibility environments: A collective benchmark study. IEEE Transactions on Image Processing, 29: 5737-5752, 2020. 6, 7 +[31] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2023. 1, 2 +[32] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z Li. S3fd: Single shot scale-invariant face detector. In Proceedings of the IEEE international conference on computer vision, pages 192-201, 2017. 7 +[33] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Proceedings of the 27th ACM international conference on multimedia, pages 1632-1640, 2019. 1, 2 +[34] Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision, 129:1013-1037, 2021. 1, 6 +[35] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 8 +[36] Chuanjun Zheng, Daming Shi, and Wentian Shi. Adaptive unfolding total variation network for low-light image enhancement. pages 4439-4448, 2021. 1 +[37] Shen Zheng and Gaurav Gupta. Semantic-guided zero-shot learning for low-light image/video enhancement. In Pro + +ceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 581-590, 2022. 2, 6 +[38] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 8 +[39] Han Zhou, Wei Dong, Xiaohong Liu, Shuaicheng Liu, Xiongkuo Min, Guangtao Zhai, and Jun Chen. Glare: Low light image enhancement via generative latent feature based codebook retrieval. arXiv preprint arXiv:2407.12431, 2024. 1, 2 \ No newline at end of file diff --git a/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/images.zip b/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9c62508e122255ca85debe9dc6a8682a09082b10 --- /dev/null +++ b/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cad3cc9b3aaf08d9d1286db09e4170cfe58cca1e2d6e98f976a12490a992a957 +size 813279 diff --git a/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/layout.json b/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a686b91801a24f51773e58e5470f72e0bd541c84 --- /dev/null +++ b/2025/Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration/layout.json @@ -0,0 +1,8745 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 545, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 545, + 138 + ], + "type": "text", + "content": "Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 160, + 462, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 160, + 462, + 175 + ], + "spans": [ + { + "bbox": [ + 148, + 160, + 462, + 175 + ], + "type": "text", + "content": "Yuxuan Gu" + }, + { + "bbox": [ + 148, + 160, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 160, + 462, + 175 + ], + "type": "text", + "content": ", Haoxuan Wang" + }, + { + "bbox": [ + 148, + 160, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 160, + 462, + 175 + ], + "type": "text", + "content": ", Pengyang Ling" + }, + { + "bbox": [ + 148, + 160, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 148, + 160, + 462, + 175 + ], + "type": "text", + "content": ", Zhixiang Wei" + }, + { + "bbox": [ + 148, + 160, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 160, + 462, + 175 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 175, + 411, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 175, + 411, + 189 + ], + "spans": [ + { + "bbox": [ + 198, + 175, + 411, + 189 + ], + "type": "text", + "content": "Huaian Chen" + }, + { + "bbox": [ + 198, + 175, + 411, + 189 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 198, + 175, + 411, + 189 + ], + "type": "text", + "content": ", Yi Jin" + }, + { + "bbox": [ + 198, + 175, + 411, + 189 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 198, + 175, + 411, + 189 + ], + "type": "text", + "content": ", and Enhong Chen" + }, + { + "bbox": [ + 198, + 175, + 411, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 189, + 496, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 189, + 496, + 204 + ], + "spans": [ + { + "bbox": [ + 114, + 189, + 496, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 114, + 189, + 496, + 204 + ], + "type": "text", + "content": "University of Science and Technology of China " + }, + { + "bbox": [ + 114, + 189, + 496, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 114, + 189, + 496, + 204 + ], + "type": "text", + "content": "Shanghai AI Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 256, + 297, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 256, + 297, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 297, + 521 + ], + "type": "text", + "content": "In this paper, we observe that the collaboration of various foundation models can perceive semantic and degraded information within images, thereby guiding the low-light enhancement process. Specifically, we propose a self-supervised low-light enhancement framework based on the multiple foundation models collaboration (dubbed FoCo), aimed at improving both the visual quality of enhanced images and the performance in high-level applications. At the feature level, FoCo leverages the rich features from various foundation models to enhance the model's semantic perception during training, thereby reducing the gap between enhanced results and high-quality images from a high-level perspective. At the task level, we exploit the robustness-gap between strong foundation models and weak models, applying high-level task guidance to the low-light enhancement training process. Through the collaboration of multiple foundation models, the proposed framework shows better enhancement performance and adapts better to high-level tasks. Extensive experiments across various enhancement and application benchmarks demonstrate the qualitative and quantitative superiority of the proposed method over numerous state-of-the-art techniques." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 540, + 135, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 540, + 135, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 540, + 135, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": "Images taken under short exposure or low-light conditions frequently suffer from significant degradation, including loss of color information and the introduction of substantial noise. Such degradation severely impacts the performance of downstream tasks [19, 36]. Low-light image enhancement (LIE) aims to restore input images captured in low-light environments to high-quality, well-lit images. With the advancement of deep learning technology in computer vision, recent research [2, 25, 31, 33, 34, 39] has achieved low-light image enhancement through supervised training on paired datasets. However, collecting" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 230, + 553, + 353 + ], + "blocks": [ + { + "bbox": [ + 316, + 230, + 553, + 353 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 353 + ], + "type": "image", + "image_path": "0e934dfe43f4f9e781317759ad13a83d9c27db5c07d05b43b14c4df1c02f3717.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 350, + 375, + 364 + ], + "lines": [ + { + "bbox": [ + 333, + 350, + 375, + 364 + ], + "spans": [ + { + "bbox": [ + 333, + 350, + 375, + 364 + ], + "type": "text", + "content": "(a) Prior-based Framework" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 364, + 396, + 480 + ], + "blocks": [ + { + "bbox": [ + 316, + 364, + 396, + 480 + ], + "lines": [ + { + "bbox": [ + 316, + 364, + 396, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 396, + 480 + ], + "type": "image", + "image_path": "fb14190bedc8083fdfe466f7f5b1bf49064a3c2525be1a386033979f7869ab65.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 481, + 393, + 498 + ], + "lines": [ + { + "bbox": [ + 318, + 481, + 393, + 498 + ], + "spans": [ + { + "bbox": [ + 318, + 481, + 393, + 498 + ], + "type": "text", + "content": "(b) Previous High-for-Low Framework" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 506, + 555, + 607 + ], + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 607 + ], + "type": "text", + "content": "Figure 1. The proposed approach demonstrates substantial improvements over existing unsupervised methods in terms of enhancement tasks (PSNR, SSIM) and application tasks (classification/Acc, face detection/AP, and instance segmentation/mAP), attributed to the proposed Multi-VFM Collaboration Framework. Compared with previous unsupervised LIE learning paradigms (a) and (b), the proposed framework (c) can make full use of the foundation models from a collaborative perspective to improve visual quality and downstream task performance." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 402, + 364, + 536, + 480 + ], + "blocks": [ + { + "bbox": [ + 402, + 364, + 536, + 480 + ], + "lines": [ + { + "bbox": [ + 402, + 364, + 536, + 480 + ], + "spans": [ + { + "bbox": [ + 402, + 364, + 536, + 480 + ], + "type": "image", + "image_path": "c16ed0f454b6d418c106b0188de6dfaee2a297b5b123ca6679001322b239b424.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 481, + 522, + 498 + ], + "lines": [ + { + "bbox": [ + 417, + 481, + 522, + 498 + ], + "spans": [ + { + "bbox": [ + 417, + 481, + 522, + 498 + ], + "type": "text", + "content": "(c) Proposed Multi-Foundation model Collaboration Framework" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "the substantial amount of paired images required for these methods is both time-consuming and costly, which highlights the broad potential applications of self-supervised approaches [6, 7, 13, 18]. As shown in Fig. 1 (a), early self-supervised training paradigms mainly relied on manually defined priors, such as smoothness priors, brightness priors, or histogram equalization. The drawback of such prior-based frameworks is their lack of connection with semantic" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 149, + 713 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16071" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "information, and the efficacy of these prior functions can diminish under complex natural conditions. For instance, the smoothness prior may sometimes degrade semantic information, limiting these methods' performance in practical applications and impacting the visual quality of results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 294, + 300 + ], + "type": "text", + "content": "Recognizing these limitations, as illustrated in Figure 1 (b), some approaches [15, 16, 27, 29, 37] have introduced high-level models into the training of Low-light Image Enhancement (LIE) models, significantly improving the model's ability to retain semantic information. However, this line of research typically relies on a specific high-level model and does not fully explore the potential of different vision models, resulting in only limited visual improvements. In recent years, the emergence of various foundation vision models widely applied in real-world scenarios has led us to consider integrating multiple foundational models from a collaborative perspective, which aims to improve the visual quality and practical performance of LIE methods more effectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 302, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 302, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 302, + 295, + 540 + ], + "type": "text", + "content": "In this work, we explore the use of multiple foundation models to strengthen the self-supervised enhancement process, as shown in Figure 1 (c). To this end, We address two key challenges: (1) how to effectively utilize pre-trained knowledge? (2) how to handle differences between models? To deal with these challenges, we propose FoCo, a Low-light Image Enhancement (LIE) framework based on the multiple foundation models ensemble, aligning the enhancement results with high-quality images from both feature-level and task-level perspectives. As the core strategies to implement FoCo, we introduce two optimization strategies: Foundation Feature-space Alignment (FoFA) and Gap-driven task-based Alignment (GaTA). FoFA focuses on creating a unified feature space for foundation models, using adversarial learning to align enhanced results with high-quality image sets. GaTA leverages robustness differences across models to model degradation scenarios, using robust pseudo-labels to directly incorporate high-level task training into the enhancement model's optimization process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 542, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 564 + ], + "type": "text", + "content": "In summary, the contributions of this paper are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "content": "- We are the first to explore the collaborative effects of multiple vision foundation models on the LIE task, and then propose FoCo, a framework designed to enhance both the human- and machine-understanding performance of unsupervised LIE methods by leveraging complementary interactions among multiple foundational vision models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 294, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 294, + 689 + ], + "type": "text", + "content": "- We introduce two key strategies FoFA and GaTA to fully exploit the pre-trained knowledge of foundation models, which align enhancement results and high-quality images from the perspective of foundation models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "- We conduct extensive experiments on benchmarks for low-light enhancement, object detection, instance seg" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 120 + ], + "type": "text", + "content": "mentation, and classification. The experimental results show significant performance improvements, demonstrating the proposed method's effectiveness in enhancing visual quality and application performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 133, + 404, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 404, + 146 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 404, + 146 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 153, + 481, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 153, + 481, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 481, + 166 + ], + "type": "text", + "content": "2.1. Low-light Image Enhancement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 171, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 553, + 338 + ], + "type": "text", + "content": "Supervised methods, such as RetinexNet [25], KinD [33], and more recent approaches like Retinexformer [2], Diff-Retinex [31], and GLARE [39], utilize paired datasets and advanced model architectures to enhance image quality. While successful, these methods face challenges due to their dependence on large-scale paired datasets, which are difficult to obtain in real-world scenarios. Unsupervised methods, such as EnLightenGAN [11], and CoLIE [4], aim to overcome this limitation by eliminating the need for paired data, leveraging adversarial learning and multimodal approaches. However, these methods often neglect integration with high-level models, limiting their ability to preserve semantic information and adapt to downstream tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 346, + 491, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 491, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 491, + 359 + ], + "type": "text", + "content": "2.2. High-Level guided Enhancement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "text", + "content": "Recent works [15, 16, 27, 29, 37] have increasingly focused on leveraging high-level models to supervise low-light enhancement models, aiming to improve the fidelity of the enhancement results. For example, SGZ [37] incorporates the concept of focal loss to maximize the logits of a pre-trained semantic segmentation model, enhancing the model's ability to preserve semantic information. SKF [27] integrates features from semantic segmentation models into the enhancement model and proposes a semantic-based histogram equalization loss and adversarial loss based on segmentation results. Methods like NeRCo [29] and CLIP-LIT [16] employ pre-trained CLIP models to align enhancement results with high-quality images during adversarial training. The methods above have significant success, but a common limitation of these approaches is their reliance on a specific high-level model, and they often do not fully exploit the model's potential." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 579, + 375, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 375, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 375, + 591 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 599, + 421, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 421, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 421, + 610 + ], + "type": "text", + "content": "3.1. FoCo Framework" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "To implement the low-light enhancement process, we first define a simple enhancer " + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": " for enhancing a low-light input image " + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": ", and then use some priors to constrain the behavior of " + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": ". By incorporating the smoothness and consistency losses " + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": " borrowed from [21] and [6], we can construct a baseline low-light enhancement model that fully relies on simply defined hand-crafted prior. However, this dependency on these hand-crafted prior compromises" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16072" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 56, + 516, + 277 + ], + "blocks": [ + { + "bbox": [ + 93, + 56, + 516, + 277 + ], + "lines": [ + { + "bbox": [ + 93, + 56, + 516, + 277 + ], + "spans": [ + { + "bbox": [ + 93, + 56, + 516, + 277 + ], + "type": "image", + "image_path": "1c0ed6892d501810b49c536f7e15911915db94ad47e55d16f81fa4211f8a5c49.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 279, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 54, + 279, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 54, + 279, + 555, + 338 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed low-light enhancement method, which enhances a basic LIE model using multi-foundation model ensemble. Given the set of pre-trained foundation models, the enhancement model first performs basic enhancement, followed by supervision from the ensemble through two strategies: (1) Feature-Level: adversarial training aligns enhanced results with high-quality images in the unified feature space, and (2) Task-Level: robustness differences between models are used to align the enhancement results with high-quality images on high-level tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 342, + 298, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 342, + 298, + 378 + ], + "spans": [ + { + "bbox": [ + 54, + 342, + 298, + 378 + ], + "type": "text", + "content": "the model's visual quality and adaptability to downstream tasks. Mathematically, such a straightforward image-level enhancement approach can be formalized as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 384, + 295, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 384, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 133, + 384, + 295, + 402 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\theta} d (G (I), \\hat {I}), \\tag {1}", + "image_path": "afe030d33af5fcfee31977d9e8f50f17b3c8f6140b320c84e5bafde48740c047.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 409, + 298, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 409, + 298, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 409, + 298, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 409, + 298, + 567 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 54, + 409, + 298, + 567 + ], + "type": "text", + "content": " represents the ideal normal-light image, " + }, + { + "bbox": [ + 54, + 409, + 298, + 567 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 409, + 298, + 567 + ], + "type": "text", + "content": " denotes the parameters of the enhancement model, " + }, + { + "bbox": [ + 54, + 409, + 298, + 567 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 54, + 409, + 298, + 567 + ], + "type": "text", + "content": " is the distance measurement function. To address the limitations of the prior-based approach, we introduce FoCo framework designed to enhance this model from the perspective of multiple foundation models collaboration. Specifically, we propose (1) Foundation Feature-space Alignment (FoFA), at the feature level (indicated in green), and (2) Gap-driven Task-based Alignment (GaTA), at the task level (indicated in blue). These strategies leverage the feature extraction process (denoted as " + }, + { + "bbox": [ + 54, + 409, + 298, + 567 + ], + "type": "inline_equation", + "content": "F^f" + }, + { + "bbox": [ + 54, + 409, + 298, + 567 + ], + "type": "text", + "content": ") and task decision process (denoted as " + }, + { + "bbox": [ + 54, + 409, + 298, + 567 + ], + "type": "inline_equation", + "content": "F^t" + }, + { + "bbox": [ + 54, + 409, + 298, + 567 + ], + "type": "text", + "content": ") of foundation models. Thus, FoCo's optimization objective can be expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 571, + 296, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 571, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 85, + 571, + 296, + 632 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\arg \\min _ {\\theta} \\underbrace {d (G (I) , \\hat {I})} _ {\\mathcal {L} _ {C}} + \\underbrace {d \\left(F ^ {f} (G (I)) , F ^ {f} (\\hat {I})\\right)} _ {\\text {F o F A}} \\tag {2} \\\\ + \\underbrace {d (F ^ {t} (G (I)) , F ^ {t} (\\hat {I}))} _ {\\text {G a T A}}. \\\\ \\end{array}", + "image_path": "24ff74bab90a27154daff72f4e435740384c0ab1ed3b1ab794b3bb7900bd348c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 639, + 280, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 280, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 280, + 651 + ], + "type": "text", + "content": "The corresponding framework is illustrated in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 253, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 253, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 253, + 672 + ], + "type": "text", + "content": "3.2. Foundation Feature-space Alignment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Observation. Figure 3 (a) displays a GAN-enhanced low-light image from the LIS [3] dataset, alongside its Mask-RCNN detect result. The GAN was trained in an unpaired" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 342, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 675 + ], + "type": "text", + "content": "manner on the LIS dataset. It can be observed that the GAN-enhanced result suffers from significant semantic degradation, likely due to the discriminator's limited semantic awareness, which fails to guide the generator effectively. This observation leads us to a potential solution: could we enhance the discriminator by incorporating a feature extractor with strong semantic awareness? As shown in Figure 3 (b), we use CLIP, SAM and ResNet separately to extract feature maps from both images. Incorporating these base models into the discriminator training process, as shown in Figure 3 (a), led to significant improvements in detection performance. However, SAM produced darker results, CLIP introduced color bias, and ResNet maintained limited semantic integrity. We hypothesize that these differences stem from the varying focus of each foundational model: CLIP prioritizes the primary objects in the image and is less sensitive to color bias and subtle degradations, SAM captures fine-grained semantics but lacks sensitivity to lighting conditions, while ResNet offers weaker semantic understanding compared to the other models but is more susceptible to degradations and lighting variations, which still contributes to the discriminator's learning. Interestingly, ResNet's lack of robustness to degradations may become a key factor in complementing CLIP and SAM, while the discriminator needs to identify degradation and illumination. As shown in Figure 3 (c), where we observed that ResNet's features were more sensitive to lighting variations compared to the other models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "This inspires us that CLIP, SAM, and ResNet may complement each other at the feature level: CLIP highlights key objects, SAM provides fine-grained guidance, and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16073" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 54, + 542, + 232 + ], + "blocks": [ + { + "bbox": [ + 67, + 54, + 542, + 232 + ], + "lines": [ + { + "bbox": [ + 67, + 54, + 542, + 232 + ], + "spans": [ + { + "bbox": [ + 67, + 54, + 542, + 232 + ], + "type": "image", + "image_path": "3a6c20a317ba9e5d8715c3b5c2a04d789db5794c45e7f989e3fb2320b078be63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 233, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 54, + 233, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 54, + 233, + 555, + 278 + ], + "type": "text", + "content": "Figure 3. The perception capabilities of foundation models on unsupervised low-light enhancer training. (a) presents the discriminator feature maps, enhanced results, and instance segmentation outcomes using different foundation model-supervised discriminators. (b) illustrates the process of adversarial training supervised by the foundation model. (c) compares the robustness of CLIP and ResNet feature maps under varying lighting conditions. The results suggest that foundation models may collaborate at feature level." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 285, + 294, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 294, + 357 + ], + "type": "text", + "content": "ResNet adds degradation-related information to guide semantic learning in the discriminator. However, the feature spaces of CLIP, SAM, and ResNet are unaligned, preventing the discriminator from directly utilizing their outputs. The challenge now is how to effectively integrate CLIP, SAM, and ResNet into the discriminator training process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "text", + "content": "Method. Based on this observation, Foundation Feature-space Alignment (FoFA) aims to leverage features from different foundation models to bring the enhanced results closer to a high-quality image set " + }, + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "inline_equation", + "content": "\\tilde{I}" + }, + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "inline_equation", + "content": "F_{i,j}" + }, + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "text", + "content": " denote the feature extractor of the " + }, + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "text", + "content": "-th pre-trained foundation models, extending up to the " + }, + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "text", + "content": "-th layer, given the varied structures and training data of these models, our goal is to construct a unified feature space " + }, + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "inline_equation", + "content": "\\hat{F}_j" + }, + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "text", + "content": " and find semantic and degradation information in it that is beneficial to the enhancement process. To this end, we use a series of convolutional projectors " + }, + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "inline_equation", + "content": "P_{i,j}" + }, + { + "bbox": [ + 55, + 357, + 295, + 523 + ], + "type": "text", + "content": " to map the features from different spaces into a unified feature space. To align the numerical scale of these features with the subsequent modules, we normalize the output features, defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 529, + 295, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 529, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 79, + 529, + 295, + 546 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {i, j} ^ {f} (x) = \\operatorname {N o r m} \\left(P _ {i, j} \\left(F _ {i, j} ^ {f} (x)\\right)\\right), s. t. x \\sim p (x), \\tag {3}", + "image_path": "71e096702bb602385b96f94c6f24203a75b579a8e06642953221606c559f85bc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 555, + 294, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 555, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 96, + 555, + 294, + 578 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {j} ^ {f} (x) = \\left\\{\\hat {F} _ {i, j} ^ {f} (x) \\right\\} _ {i = 0} ^ {n}, s. t. x \\sim p (x), \\tag {4}", + "image_path": "ed58a6992dd9ef510bb979e18d69eb39c845998b2032273e9327a09203c81f51.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " denotes the count number of foundation models, " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " is the convolutional projectors and " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{i,j}^{f}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " denotes feature extraction with foundation models. Then we adopt an adversarial learning approach by introducing a discriminator " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ", which links the features from the foundation models with the enhancement model. The input to " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " consists of images along with their corresponding semantic information derived from the unified feature space. Specifically, inspired by [27], we employ cross-attention across both spatial and channel dimensions between the image features and the semantic features to generate attention maps " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "A^c" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "A^s" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 329, + 303, + 555, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 303, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 329, + 303, + 555, + 324 + ], + "type": "interline_equation", + "content": "A ^ {c} = \\operatorname {S o f t m a x} \\left(P ^ {c, k} \\left(f _ {j} ^ {F}\\right) \\times P ^ {c, q} \\left(f _ {j} ^ {D}\\right) / \\sqrt {C}\\right), \\tag {5}", + "image_path": "b4cac48bc74acd4417457820a1e7eaf8bee6e2e446975fef83fde849d0256442.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 346, + 555, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 346, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 321, + 346, + 555, + 368 + ], + "type": "interline_equation", + "content": "A ^ {s} = \\operatorname {S o f t m a x} \\left(P ^ {s, k} \\left(f _ {j} ^ {F}\\right) \\cdot P ^ {s, q} \\left(f _ {j} ^ {D}\\right) / \\sqrt {H \\cdot W}\\right), \\tag {6}", + "image_path": "5a0b6c3df17b4f899dc94d16ab81139438ff6a4f8ba83ba244c9a79321cd58ae.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 376, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 376, + 554, + 415 + ], + "type": "inline_equation", + "content": "f_{j}^{F} = \\hat{F}_{j}^{f}(x)" + }, + { + "bbox": [ + 313, + 376, + 554, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 376, + 554, + 415 + ], + "type": "inline_equation", + "content": "f_{j}^{D}" + }, + { + "bbox": [ + 313, + 376, + 554, + 415 + ], + "type": "text", + "content": " is the feature map of " + }, + { + "bbox": [ + 313, + 376, + 554, + 415 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 376, + 554, + 415 + ], + "type": "text", + "content": "-th layer in " + }, + { + "bbox": [ + 313, + 376, + 554, + 415 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 376, + 554, + 415 + ], + "type": "text", + "content": ". These attention maps are then utilized to adjust the original image features to the output feature " + }, + { + "bbox": [ + 313, + 376, + 554, + 415 + ], + "type": "inline_equation", + "content": "f_{j}^{O}" + }, + { + "bbox": [ + 313, + 376, + 554, + 415 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 387, + 423, + 553, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 423, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 387, + 423, + 553, + 440 + ], + "type": "interline_equation", + "content": "f _ {j} ^ {O} = A ^ {c} \\times A ^ {s} + f _ {j} ^ {D}. \\tag {7}", + "image_path": "787a33fcc1f9d867935591cef8d77f031b62ebc6fcb36fdc447100d61d068341.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 443, + 555, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 555, + 550 + ], + "type": "text", + "content": "The optimization objective of " + }, + { + "bbox": [ + 313, + 443, + 555, + 550 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 443, + 555, + 550 + ], + "type": "text", + "content": " is to align its learned features with the outputs of the foundation model while simultaneously distinguishing between enhanced low-quality images and high-quality images. Throughout this process, it is crucial to ensure that the features input to the classification layer of " + }, + { + "bbox": [ + 313, + 443, + 555, + 550 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 443, + 555, + 550 + ], + "type": "text", + "content": " are closely related to the semantic and degradation information. Based on the observation, the loss function for the discriminator " + }, + { + "bbox": [ + 313, + 443, + 555, + 550 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 443, + 555, + 550 + ], + "type": "text", + "content": " can be expressed as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 552, + 553, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 552, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 323, + 552, + 553, + 620 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {D} = - \\mathbb {E} _ {x \\sim p _ {h i g h} (x)} [ \\log D (x) + \\left\\| D ^ {f} (x) - \\hat {F} _ {m} ^ {f} (x) \\right\\| ] \\\\ - \\mathbb {E} _ {x \\sim p l o w (x)} [ \\log (1 - D (G (x))) + \\\\ \\left\\| D ^ {f} (G (x)) - \\hat {F} _ {m} ^ {f} (G (x)) \\right\\| ], \\tag {8} \\\\ \\end{array}", + "image_path": "972d4330f699726718e53b4a2fee1a326c349735d55555f60bbad027e4a177ad.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 363, + 623, + 553, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 623, + 553, + 637 + ], + "spans": [ + { + "bbox": [ + 363, + 623, + 553, + 637 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G} = - \\mathbb {E} _ {x \\sim p _ {\\text {l o w}} (x)} [ \\log D (G (x)) ], \\tag {9}", + "image_path": "ccfd8af51553e6bd0e04d480c41f64dc1ef8e512f3dec267d12575b65d7ece80.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " denotes the layer count of the foundation model, " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "D^{f}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " denotes feature extraction with discriminator. The discriminator " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " effectively projects the foundation models' features into a unified feature space, providing valuable guidance for the enhancement model. Unlike traditional image-level adversarial training, FoFA focuses on selected" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16074" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 175, + 148 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 175, + 148 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 175, + 148 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 175, + 148 + ], + "type": "image", + "image_path": "612497201aabd6991f1b5e21ef40aebc6e075b34e8b38d6f1b01b872e486ab60.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 149, + 127, + 156 + ], + "lines": [ + { + "bbox": [ + 118, + 149, + 127, + 156 + ], + "spans": [ + { + "bbox": [ + 118, + 149, + 127, + 156 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 70, + 287, + 148 + ], + "blocks": [ + { + "bbox": [ + 176, + 70, + 287, + 148 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 287, + 148 + ], + "type": "image", + "image_path": "c03be5fbfaaa1b3d9c95d550950f886214dd553d87110430e16d3c72f6e842dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 149, + 239, + 156 + ], + "lines": [ + { + "bbox": [ + 229, + 149, + 239, + 156 + ], + "spans": [ + { + "bbox": [ + 229, + 149, + 239, + 156 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 73, + 159, + 115, + 186 + ], + "blocks": [ + { + "bbox": [ + 61, + 180, + 71, + 188 + ], + "lines": [ + { + "bbox": [ + 61, + 180, + 71, + 188 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 71, + 188 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 73, + 159, + 115, + 186 + ], + "lines": [ + { + "bbox": [ + 73, + 159, + 115, + 186 + ], + "spans": [ + { + "bbox": [ + 73, + 159, + 115, + 186 + ], + "type": "image", + "image_path": "ea0c9e4b74b05934b9104f09644c89c74fe9b06918915a325d4fd82179b02218.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 117, + 159, + 157, + 186 + ], + "blocks": [ + { + "bbox": [ + 117, + 159, + 157, + 186 + ], + "lines": [ + { + "bbox": [ + 117, + 159, + 157, + 186 + ], + "spans": [ + { + "bbox": [ + 117, + 159, + 157, + 186 + ], + "type": "image", + "image_path": "1ff69a92396fdf9be25bc22649b5b30fc8dd68a418326e433e0dc7343676ec9d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 186, + 157, + 208 + ], + "lines": [ + { + "bbox": [ + 117, + 186, + 157, + 208 + ], + "spans": [ + { + "bbox": [ + 117, + 186, + 157, + 208 + ], + "type": "text", + "content": "CLIP: Ambulance (0.9067) ResNet:Police van (0.5195)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 186, + 196, + 208 + ], + "lines": [ + { + "bbox": [ + 157, + 186, + 196, + 208 + ], + "spans": [ + { + "bbox": [ + 157, + 186, + 196, + 208 + ], + "type": "text", + "content": "CLIP: Ambulance (0.9438) ResNet: Ambulance (0.7087)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 157, + 159, + 199, + 186 + ], + "blocks": [ + { + "bbox": [ + 157, + 159, + 199, + 186 + ], + "lines": [ + { + "bbox": [ + 157, + 159, + 199, + 186 + ], + "spans": [ + { + "bbox": [ + 157, + 159, + 199, + 186 + ], + "type": "image", + "image_path": "2741c073732cfd083e50fff79a1626fcea3bf35433b889383c166e0d91d4bfc9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 200, + 159, + 241, + 186 + ], + "blocks": [ + { + "bbox": [ + 200, + 159, + 241, + 186 + ], + "lines": [ + { + "bbox": [ + 200, + 159, + 241, + 186 + ], + "spans": [ + { + "bbox": [ + 200, + 159, + 241, + 186 + ], + "type": "image", + "image_path": "2c8b7680ba91aff6582fd8b3dfd5cda23e68441cef75e453264688574367a6d7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 186, + 238, + 208 + ], + "lines": [ + { + "bbox": [ + 202, + 186, + 238, + 208 + ], + "spans": [ + { + "bbox": [ + 202, + 186, + 238, + 208 + ], + "type": "text", + "content": "CLIP: Ambulance (0.9131) \nResNet: Police van (0.6643)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 242, + 159, + 285, + 186 + ], + "blocks": [ + { + "bbox": [ + 242, + 159, + 285, + 186 + ], + "lines": [ + { + "bbox": [ + 242, + 159, + 285, + 186 + ], + "spans": [ + { + "bbox": [ + 242, + 159, + 285, + 186 + ], + "type": "image", + "image_path": "0e08270354025a7607191e1b6f7b14b596325df1a926a40d715efd9259dcb482.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 245, + 186, + 280, + 208 + ], + "lines": [ + { + "bbox": [ + 245, + 186, + 280, + 208 + ], + "spans": [ + { + "bbox": [ + 245, + 186, + 280, + 208 + ], + "type": "text", + "content": "CLIP: Ambulance (0.9092) ResNet:Police van (0.8759)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 76, + 210, + 115, + 236 + ], + "blocks": [ + { + "bbox": [ + 61, + 232, + 70, + 239 + ], + "lines": [ + { + "bbox": [ + 61, + 232, + 70, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 70, + 239 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 210, + 115, + 236 + ], + "lines": [ + { + "bbox": [ + 76, + 210, + 115, + 236 + ], + "spans": [ + { + "bbox": [ + 76, + 210, + 115, + 236 + ], + "type": "image", + "image_path": "3e92ccc07a0bd255bc1eab55626fdd4c761ef0875a1e1b5d92391d07480e7dcf.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 237, + 109, + 258 + ], + "lines": [ + { + "bbox": [ + 78, + 237, + 109, + 258 + ], + "spans": [ + { + "bbox": [ + 78, + 237, + 109, + 258 + ], + "type": "text", + "content": "CLIP: Vending (0.3428) \nResNet: Mail box (0.3349)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 260, + 295, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 260, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 295, + 326 + ], + "type": "text", + "content": "Figure 4. Robustness to illumination across models. (a) and (b) show the similarity of results between CLIP and ResNet under different lighting adjustments. (c) and (d) present specific image examples, along with the classification results and confidence scores. The goal of GaTA is to optimize the augmented results towards a direction that enhances correct recognition." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 115, + 210, + 157, + 235 + ], + "blocks": [ + { + "bbox": [ + 115, + 210, + 157, + 235 + ], + "lines": [ + { + "bbox": [ + 115, + 210, + 157, + 235 + ], + "spans": [ + { + "bbox": [ + 115, + 210, + 157, + 235 + ], + "type": "image", + "image_path": "26d6e417045f58c0a10dbdee232ca86c6154558de6bcc93e4f00920f6e9a0776.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 237, + 152, + 258 + ], + "lines": [ + { + "bbox": [ + 118, + 237, + 152, + 258 + ], + "spans": [ + { + "bbox": [ + 118, + 237, + 152, + 258 + ], + "type": "text", + "content": "CLIP: Vending (0.4041) ResNet: Mail box (0.4871)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 157, + 210, + 199, + 235 + ], + "blocks": [ + { + "bbox": [ + 157, + 210, + 199, + 235 + ], + "lines": [ + { + "bbox": [ + 157, + 210, + 199, + 235 + ], + "spans": [ + { + "bbox": [ + 157, + 210, + 199, + 235 + ], + "type": "image", + "image_path": "c59f63926f7d1d79ef71eab639e5bf62f9cb836e8684b747442382b31cee52be.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 236, + 194, + 258 + ], + "lines": [ + { + "bbox": [ + 162, + 236, + 194, + 258 + ], + "spans": [ + { + "bbox": [ + 162, + 236, + 194, + 258 + ], + "type": "text", + "content": "CLIP: Vending (0.4641) ResNet: Mail box (0.2925)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 200, + 210, + 241, + 235 + ], + "blocks": [ + { + "bbox": [ + 200, + 210, + 241, + 235 + ], + "lines": [ + { + "bbox": [ + 200, + 210, + 241, + 235 + ], + "spans": [ + { + "bbox": [ + 200, + 210, + 241, + 235 + ], + "type": "image", + "image_path": "9139e1be7ce155ce8663412c56278c1366d6144b50febaa568bd8d5fa1661886.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 236, + 235, + 258 + ], + "lines": [ + { + "bbox": [ + 203, + 236, + 235, + 258 + ], + "spans": [ + { + "bbox": [ + 203, + 236, + 235, + 258 + ], + "type": "text", + "content": "CLIP: Vending (0.4731) ResNet: Mail box (0.2285)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 242, + 210, + 285, + 235 + ], + "blocks": [ + { + "bbox": [ + 242, + 210, + 285, + 235 + ], + "lines": [ + { + "bbox": [ + 242, + 210, + 285, + 235 + ], + "spans": [ + { + "bbox": [ + 242, + 210, + 285, + 235 + ], + "type": "image", + "image_path": "678fdbcbc1f287af8300e0475cc828dd3e182b8164570510446a43f5ca00711d.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 245, + 236, + 277, + 258 + ], + "lines": [ + { + "bbox": [ + 245, + 236, + 277, + 258 + ], + "spans": [ + { + "bbox": [ + 245, + 236, + 277, + 258 + ], + "type": "text", + "content": "CLIP: Vending (0.4685) ResNet: Vending (0.2813)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 328, + 295, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 352 + ], + "type": "text", + "content": "semantic features, with the goal of aligning enhanced results with the high-quality image set." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 352, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 495 + ], + "type": "text", + "content": "The strength of FoFA lies not only in its ability to preserve semantics but also in its optimization objective, which is designed to enhance compatibility with downstream tasks. In FoFA, the discriminator " + }, + { + "bbox": [ + 55, + 352, + 296, + 495 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 352, + 296, + 495 + ], + "type": "text", + "content": " is specifically designed to inherit the strong semantic perception capabilities of the multi-foundation models. From another perspective, if the semantic information can no longer distinguish the enhanced result from the high-quality image, the high-level model's perception of the enhanced result will be indistinguishable from that of the high-quality image. Therefore, for downstream tasks, FoFA effectively transforms degraded images into the high-level models' familiar domain." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 503, + 239, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 239, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 239, + 516 + ], + "type": "text", + "content": "3.3. Gap-driven Task-based Alignment" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": "Observation. FoFA improves the similarity between augmented results and high-quality images at the feature level; however, a constraint is still needed to assess whether the augmented results are easy to recognize. In Section 3.2, we observed that ResNet is more sensitive to illumination compared to CLIP. This raises the question: can we also observe robustness differences between these models at the task level? To explore this, we generated images under different lighting conditions using two methods: (1) fusing low-light and normal-light image pairs from the LSRW dataset (Get the pixel value according to the fusion ratio " + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": ") (2) applying gamma transformations to images from ImageNet-1k. As shown in Figure 4, low-light conditions degrade the performance of both CLIP and ResNet, but ResNet is more significantly affected. Interestingly, their predictions are most similar under normal conditions, suggesting that CLIP's ro" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "bustness stems from its training on large-scale data, while ResNet is more sensitive to illumination and degradation. This motivates us to leverage these robustness differences to guide the enhancement model to produce clean and well-lit results, aiming at \"detecting\" the degradation and brightness, which can be seen as obstacles in recognition." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 144, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 323 + ], + "type": "text", + "content": "Method. Gap-driven Task-based Alignment (GaTA) leverages the robustness differences between models to evaluate whether the augmented results are sufficiently good to be easily recognized. From another view, GaTA employs the less robust model as a \"degradation detector\" while the robust model functions as a ground truth provider. We found that tasks such as segmentation, classification, and detection can be applied to GaTA, with CLIP and ResNet serving as representative robust and degradation-sensitive models, respectively. Let " + }, + { + "bbox": [ + 313, + 144, + 555, + 323 + ], + "type": "inline_equation", + "content": "F^l" + }, + { + "bbox": [ + 313, + 144, + 555, + 323 + ], + "type": "text", + "content": " represent the robust model and " + }, + { + "bbox": [ + 313, + 144, + 555, + 323 + ], + "type": "inline_equation", + "content": "F^s" + }, + { + "bbox": [ + 313, + 144, + 555, + 323 + ], + "type": "text", + "content": " represent the degradation-sensitive model. Following the observation and robustness study in [22], we can apply CLIP and ResNet for " + }, + { + "bbox": [ + 313, + 144, + 555, + 323 + ], + "type": "inline_equation", + "content": "F^l" + }, + { + "bbox": [ + 313, + 144, + 555, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 144, + 555, + 323 + ], + "type": "inline_equation", + "content": "F^s" + }, + { + "bbox": [ + 313, + 144, + 555, + 323 + ], + "type": "text", + "content": ", respectively. For each enhanced image, we first use the robust model to generate pseudo-labels and obtain the logits from the sensitive model:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 364, + 324, + 553, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 324, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 364, + 324, + 553, + 338 + ], + "type": "interline_equation", + "content": "Y ^ {l} = F ^ {l} (G (x)), \\text {s . t .} x \\sim p _ {\\text {l o w}} (x), \\tag {10}", + "image_path": "e4c4a868f38048aa3354ffd73cc5c05db90769ce2a928a16ff86bbb9c38d01c2.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 364, + 346, + 553, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 346, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 364, + 346, + 553, + 360 + ], + "type": "interline_equation", + "content": "Y ^ {s} = F ^ {s} (G (x)), \\text {s . t .} x \\sim p _ {\\text {l o w}} (x). \\tag {11}", + "image_path": "300b5315ec675a9e0cb2ba8b0b303ab1baa019a037c9bc00d6b2fc18df0033e1.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 362, + 555, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 457 + ], + "type": "text", + "content": "A key issue to consider is that pseudo-labels are not always accurate. In extreme cases, if the generation quality of " + }, + { + "bbox": [ + 313, + 362, + 555, + 457 + ], + "type": "inline_equation", + "content": "G(x)" + }, + { + "bbox": [ + 313, + 362, + 555, + 457 + ], + "type": "text", + "content": " is very poor, even robust models may fail to provide fully accurate labels. Fortunately, during iterative training, we can collect multiple enhanced results for each training image, constructing a memory bank of size " + }, + { + "bbox": [ + 313, + 362, + 555, + 457 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 362, + 555, + 457 + ], + "type": "text", + "content": ". Leveraging this memory bank, we can apply a voting mechanism to correct the pseudo labels:" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 322, + 462, + 553, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 462, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 322, + 462, + 553, + 495 + ], + "type": "interline_equation", + "content": "\\hat {Y} ^ {l} = \\arg \\max _ {c} \\sum_ {t = 1} ^ {T} \\mathbb {I} \\left[ \\arg \\max _ {c ^ {\\prime}} \\left(p \\left(Y _ {t} ^ {l} = c ^ {\\prime}\\right)\\right) = c \\right], \\tag {12}", + "image_path": "3bf5287e74b225cce6f9894c24895098f09b86377159a7da6afc31196a5282bc.jpg" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "text", + "content": " denotes the predicted class for each individual image while " + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "text", + "content": " represents the final pseudo-label. Finally, we apply a high-level task loss to supervise GaTA:" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 386, + 537, + 553, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 537, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 386, + 537, + 553, + 552 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G a T A} = \\mathcal {L} _ {h} \\left(Y ^ {s}, \\hat {Y} ^ {l}\\right), \\tag {13}", + "image_path": "c55b00ae6fa5b7f21da20ff9dd0e50139cd75f6676da1a1fb5cca5dfbe450596.jpg" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 558, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 558, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + }, + { + "bbox": [ + 313, + 558, + 555, + 643 + ], + "type": "text", + "content": " denotes the specific task loss for " + }, + { + "bbox": [ + 313, + 558, + 555, + 643 + ], + "type": "inline_equation", + "content": "Y^{s}" + }, + { + "bbox": [ + 313, + 558, + 555, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 558, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\hat{Y}^l" + }, + { + "bbox": [ + 313, + 558, + 555, + 643 + ], + "type": "text", + "content": ". Please note that both " + }, + { + "bbox": [ + 313, + 558, + 555, + 643 + ], + "type": "inline_equation", + "content": "F_{t}^{l}" + }, + { + "bbox": [ + 313, + 558, + 555, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 558, + 555, + 643 + ], + "type": "inline_equation", + "content": "F_{t}^{s}" + }, + { + "bbox": [ + 313, + 558, + 555, + 643 + ], + "type": "text", + "content": " are frozen during GaTA training. By exploiting the robustness differences between models in a straightforward manner, GaTA effectively tests for degradation at the task level, ensuring the enhanced results exhibit both visual quality and strong performance in downstream applications." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 651, + 395, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 395, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 395, + 664 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 671, + 447, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 447, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 447, + 684 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "The proposed method is implemented using PyTorch, running on an NVIDIA Titan RTX GPU alongside an In" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16075" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 81, + 555, + 189 + ], + "blocks": [ + { + "bbox": [ + 55, + 57, + 555, + 78 + ], + "lines": [ + { + "bbox": [ + 55, + 57, + 555, + 78 + ], + "spans": [ + { + "bbox": [ + 55, + 57, + 555, + 78 + ], + "type": "text", + "content": "Table 1. Benchmarking results for low-light enhancement. Among unsupervised methods, we highlight the top-ranking scores in red and the second in blue." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 81, + 555, + 189 + ], + "lines": [ + { + "bbox": [ + 58, + 81, + 555, + 189 + ], + "spans": [ + { + "bbox": [ + 58, + 81, + 555, + 189 + ], + "type": "table", + "html": "
DatasetsMetricsSupervisedUnsupervised
Retinexformer [2] ICCV'23LLflow [24] AAAI'22SKF [27] CVPR'23Kind++ [34] IJCV'21EnGAN [11] TIP'21SCI [21] CVPR'22RUAS [18] CVPR'21ZeroDCE [6] CVPR'20SCL-LLE [15] AAAI'22SGZ [37] WACV'22CoLIE [4] ECCV'24Ours
HuaweiPSNR↑17.9817.1819.3116.5217.8315.447.4116.3813.2116.9113.9618.61
SSIM↑0.610.630.610.500.590.550.600.470.490.560.470.72
LPIPS↓0.320.310.320.400.350.360.420.410.370.350.390.35
NikonPSNR↑16.8718.069.3515.0915.8617.527.7516.6815.1117.3714.9217.76
SSIM↑0.630.540.340.340.580.550.590.580.590.610.550.67
LPIPS↓0.330.370.530.530.370.380.460.390.380.380.390.37
FiveKPSNR↑23.2816.3216.5914.1112.8316.299.5213.5015.7010.6218.7119.58
SSIM↑0.840.770.770.700.780.830.660.730.770.430.820.83
LPIPS↓0.240.190.200.200.140.140.300.230.210.330.160.13
", + "image_path": "582b49ed89bb421ed1d7796b8e894393b3db84e7afecace8f082b77a3dc861ad.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 100, + 191, + 509, + 369 + ], + "blocks": [ + { + "bbox": [ + 100, + 191, + 509, + 369 + ], + "lines": [ + { + "bbox": [ + 100, + 191, + 509, + 369 + ], + "spans": [ + { + "bbox": [ + 100, + 191, + 509, + 369 + ], + "type": "image", + "image_path": "faaae026970253fd47e5bc69e2d3aad76b56cb909cee594f1345c9c905874cd4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 369, + 428, + 378 + ], + "lines": [ + { + "bbox": [ + 180, + 369, + 428, + 378 + ], + "spans": [ + { + "bbox": [ + 180, + 369, + 428, + 378 + ], + "type": "text", + "content": "Figure 5. Visual Comparisons on Low-Light Enhancement Datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 386, + 297, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 386, + 297, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 386, + 297, + 542 + ], + "type": "text", + "content": "tel(R) Xeon(R) Gold 6252 CPU @ 2.10GHz. Training is conducted with the Adam optimizer, using a learning rate of 1e-4. All images are resized to " + }, + { + "bbox": [ + 54, + 386, + 297, + 542 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 54, + 386, + 297, + 542 + ], + "type": "text", + "content": " pixels, and the model is trained with a batch size of 32 over 50 epochs. For ResNet, we utilize the pre-trained weights obtained from training on ImageNet, following the origin setting in [9]. For CLIP and SAM, we employ the official weights trained on their respective datasets. For the training set of high-level applications, we randomly sample low-light images from ExDark [20] and normal-light images from COCO [17] as training set and test the obtained weights on the test sets of all benchmarks. More information can be found in Supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 551, + 138, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 138, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 138, + 562 + ], + "type": "text", + "content": "4.2. Benchmarks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "The performance of our method is evaluated across four metrics relevant to low-light enhancement: reference-based low-light enhancement, low-light image classification, low-light instance segmentation, and low-light face detection. Specifically, reference-based low-light enhancement is evaluated using the MIT-FiveK [1], Nikon, and Huawei [8] datasets, from which we select 500, 945, and 735 paired low-light and normal-light samples; Low-light image classification uses the CODaN [12] dataset, a 10-class dataset with a training set of 10,000 daytime images and a test set comprising 2,500 daytime and nighttime images. Low-light instance segmentation is performed using" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 386, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 434 + ], + "type": "text", + "content": "the LIS [3] dataset with 2230 images. For low-light face detection, we use the darkface [30] dataset, containing 6,000 images captured under low-light conditions, from which we select 1800 images for testing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 434, + 556, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 518 + ], + "type": "text", + "content": "To provide a comprehensive comparison, our method is benchmarked against 4 state-of-the-art supervised learning methods, including Retinexformer [2], LLFlow [24], SKF [27], and " + }, + { + "bbox": [ + 313, + 434, + 556, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{KinD} + +" + }, + { + "bbox": [ + 313, + 434, + 556, + 518 + ], + "type": "text", + "content": " [34], as well as 7 unsupervised learning methods, such as Zero-DCE[6], RUAS [18], EnlightGAN [11], SGZ [37], SCL-LLE [15] CoLIE [4] and SCI [21] across all tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 527, + 432, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 432, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 432, + 538 + ], + "type": "text", + "content": "4.3. Enhancement Tasks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 545, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 555, + 653 + ], + "type": "text", + "content": "To validate the effectiveness of our method in terms of generalization performance, we evaluated the proposed framework on several benchmark datasets, i.e. MITFiveK [1], Nikon and Huawei [8] datasets, using PSNR, SSIM, and LPIPS as performance metrics. The measurement results are presented in Table 1. It is evident that our proposed method demonstrates advantages compared to other self-supervised approaches, exhibiting competitive performance in terms of PSNR, SSIM, and LPIPS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Figure 5 illustrates a visual comparison between our method and other approaches. While alternative methods may struggle with incomplete degradation removal or loss of semantic information, our method consistently demonstrates its ability to effectively recover image details. This" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16076" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 81, + 555, + 169 + ], + "blocks": [ + { + "bbox": [ + 55, + 57, + 555, + 79 + ], + "lines": [ + { + "bbox": [ + 55, + 57, + 555, + 79 + ], + "spans": [ + { + "bbox": [ + 55, + 57, + 555, + 79 + ], + "type": "text", + "content": "Table 2. Performance comparison on different high-level benchmarks: CODaN for classification, Darkface for face detection, and LIS for instance segmentation. Highlighted in red are the top scores, and in blue are the second best methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 81, + 555, + 169 + ], + "lines": [ + { + "bbox": [ + 58, + 81, + 555, + 169 + ], + "spans": [ + { + "bbox": [ + 58, + 81, + 555, + 169 + ], + "type": "table", + "html": "
DatasetsDetectorsMetricsRawSupervisedUnsupervised
Retformer ICCV'23LLflow AAAI'22SKF CVPR'23Kind++ IJCV'21EnGAN TIP'21SCI CVPR'22RUAS CVPR'21ZeroDCE CVPR'20SCL-LLE AAAI'22SGZ WACV'22CoLIE ECCV'24Ours
CoDANResNet18Acc0.5020.5510.5330.5250.5390.5310.5650.5560.5650.5720.5730.5640.588
EfficientNetAcc0.5870.6140.6530.6160.5590.6300.5730.5350.6050.6760.5910.6520.742
DarkfaceS3FDAP0.6030.6530.6360.6210.6460.6380.6610.6410.6480.6690.6650.6560.683
DSFDAP0.2740.3370.3480.3510.3670.3310.3490.3260.3670.3340.3580.3340.381
LISMaskRCNNmAPbox0.2280.3030.2860.2580.3090.2860.3010.2770.3170.3090.3170.3090.331
mAPseg0.1980.2520.2540.2370.2580.2440.2630.2380.2730.2650.2720.2580.281
", + "image_path": "5491dc1199429ac9b27d26c3c702d322c10865d15115d39100334c4c10832628.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 80, + 171, + 528, + 331 + ], + "blocks": [ + { + "bbox": [ + 80, + 171, + 528, + 331 + ], + "lines": [ + { + "bbox": [ + 80, + 171, + 528, + 331 + ], + "spans": [ + { + "bbox": [ + 80, + 171, + 528, + 331 + ], + "type": "image", + "image_path": "e6c9a69dfe25af52ba1697e54d2abb352a7495f164497be130420137958b299e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 333, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 333, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 555, + 365 + ], + "type": "text", + "content": "Figure 6. Visual Comparisons and High-level Results on CODaN, Darkface, and LIS Dataset. The first row shows CODaN and the classification results by ResNet (color denotes incorrect/correct), the second row shows the Darkface detection results by S3FD, and the third row shows the LIS and detection results and Mask-RCNN." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 369, + 295, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 395 + ], + "type": "text", + "content": "further substantiates the advanced capabilities of the proposed framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 405, + 155, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 155, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 155, + 418 + ], + "type": "text", + "content": "4.4. High-level Tasks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 425, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 425, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 295, + 616 + ], + "type": "text", + "content": "Low-Light Classification. We evaluate the adaptability of various methods for low-light image classification tasks on the CODaN [12] dataset. For our experiments, we employ the official ResNet18 model, which was trained on the daytime portion of the dataset, as our primary detector. Additionally, to mitigate the impact of domain differences, we incorporate an EfficientNet [23] model trained on ImageNet, targeting classes corresponding to those in the CO-DaN dataset, as a supplementary detector. Using these two detectors, we assess Top-1 Accuracy as our evaluation metric. As shown in Table 2, our proposed method outperforms other low-light enhancement techniques, demonstrating exceptional performance. This superiority can be attributed to the more consistent illumination provided by the models enhanced with our framework, while the degraded results from other methods hinder classifier performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Low-Light Face Detection. We evaluate the performance of two detectors, S3FD [32] and DSFD [14], trained on the WiderFace [28] dataset, and on the DarkFace [30] dataset to assess the performance of various enhancement methods for low-light face detection. It is noteworthy that face detection is more granular and sensitive to semantic information compared to classification tasks; thus, local blurriness can significantly impair performance. Con-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "content": "sequently, the methods like EnGAN [11] and Retinexformer [25] demonstrate a marked decline relative to other tasks, whereas our proposed method consistently achieves excellent results across both detectors. This further underscores the advanced nature of our proposed framework." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 430, + 556, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 586 + ], + "type": "text", + "content": "Low-Light Instance Segmentation. We evaluated various low-light enhancement methods for their adaptability to low-light instance segmentation using the LIS [3] dataset. As one of the most fine-grained high-level tasks, instance segmentation imposes significant demands on enhancement techniques. We conducted experiments using Mask R-CNN [10] as provided by the official LIS framework, which was initially trained on the COCO dataset and subsequently tested on low-light images from the LIS dataset. Our results demonstrate that the proposed methods achieve superior performance, further confirming that the framework effectively preserves semantic information and enhances the quality of the results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 587, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 555, + 647 + ], + "type": "text", + "content": "The above three tasks encompass three datasets, five detectors, and six evaluation metrics, providing a comprehensive assessment of the proposed method's robustness across different low-light scenarios and its applicability to high-level tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 658, + 416, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 416, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 416, + 670 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "We conduct extensive ablation studies to validate the efficacy of our proposed perspective and evaluate the contribution of each model component. All eval" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16077" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 113, + 107 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 113, + 107 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 113, + 107 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 113, + 107 + ], + "type": "image", + "image_path": "2d9c1a2e2869dfd77e382a3e9dbce16813d9b44151612f1bf03ab5663fdfead0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 109, + 94, + 116 + ], + "lines": [ + { + "bbox": [ + 78, + 109, + 94, + 116 + ], + "spans": [ + { + "bbox": [ + 78, + 109, + 94, + 116 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 119, + 70, + 174, + 107 + ], + "blocks": [ + { + "bbox": [ + 119, + 70, + 174, + 107 + ], + "lines": [ + { + "bbox": [ + 119, + 70, + 174, + 107 + ], + "spans": [ + { + "bbox": [ + 119, + 70, + 174, + 107 + ], + "type": "image", + "image_path": "b6ecb9a356662260a582a28f16767bf05b5b22050fffe47c1aef3e153caa2902.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 108, + 162, + 116 + ], + "lines": [ + { + "bbox": [ + 129, + 108, + 162, + 116 + ], + "spans": [ + { + "bbox": [ + 129, + 108, + 162, + 116 + ], + "type": "text", + "content": "w/o GaTA" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 70, + 233, + 108 + ], + "blocks": [ + { + "bbox": [ + 179, + 70, + 233, + 108 + ], + "lines": [ + { + "bbox": [ + 179, + 70, + 233, + 108 + ], + "spans": [ + { + "bbox": [ + 179, + 70, + 233, + 108 + ], + "type": "image", + "image_path": "cefaf037e1235b52e8e32fe1885c1142b28d9042c8246cf1f883af9fac479f5c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 109, + 223, + 116 + ], + "lines": [ + { + "bbox": [ + 189, + 109, + 223, + 116 + ], + "spans": [ + { + "bbox": [ + 189, + 109, + 223, + 116 + ], + "type": "text", + "content": "GaTA-Seg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 239, + 70, + 294, + 108 + ], + "blocks": [ + { + "bbox": [ + 239, + 70, + 294, + 108 + ], + "lines": [ + { + "bbox": [ + 239, + 70, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 239, + 70, + 294, + 108 + ], + "type": "image", + "image_path": "56d563b3d4249538c07741abd4b7983c87f4f1c92075499c68a1bdfbd81edc38.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 109, + 284, + 116 + ], + "lines": [ + { + "bbox": [ + 250, + 109, + 284, + 116 + ], + "spans": [ + { + "bbox": [ + 250, + 109, + 284, + 116 + ], + "type": "text", + "content": "GaTA-Det" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 118, + 113, + 154 + ], + "blocks": [ + { + "bbox": [ + 58, + 118, + 113, + 154 + ], + "lines": [ + { + "bbox": [ + 58, + 118, + 113, + 154 + ], + "spans": [ + { + "bbox": [ + 58, + 118, + 113, + 154 + ], + "type": "image", + "image_path": "f903bb62ba83887890a6e20f392840afbf52f9e908fa7f960f1d62b33e934e1a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 155, + 101, + 162 + ], + "lines": [ + { + "bbox": [ + 70, + 155, + 101, + 162 + ], + "spans": [ + { + "bbox": [ + 70, + 155, + 101, + 162 + ], + "type": "text", + "content": "w/o FoFA" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 119, + 118, + 174, + 154 + ], + "blocks": [ + { + "bbox": [ + 119, + 118, + 174, + 154 + ], + "lines": [ + { + "bbox": [ + 119, + 118, + 174, + 154 + ], + "spans": [ + { + "bbox": [ + 119, + 118, + 174, + 154 + ], + "type": "image", + "image_path": "57730f0bf6a3a85ca22f9c70a951f08f15951fc364b6d87f3d807caf92856f80.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 155, + 161, + 162 + ], + "lines": [ + { + "bbox": [ + 130, + 155, + 161, + 162 + ], + "spans": [ + { + "bbox": [ + 130, + 155, + 161, + 162 + ], + "type": "text", + "content": "w/o CLIP" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 179, + 118, + 233, + 154 + ], + "blocks": [ + { + "bbox": [ + 179, + 118, + 233, + 154 + ], + "lines": [ + { + "bbox": [ + 179, + 118, + 233, + 154 + ], + "spans": [ + { + "bbox": [ + 179, + 118, + 233, + 154 + ], + "type": "image", + "image_path": "5afe7dfe6a0977e6b52b5984df0663978765ae6f93f29218c1fa3580f3218162.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 155, + 220, + 162 + ], + "lines": [ + { + "bbox": [ + 190, + 155, + 220, + 162 + ], + "spans": [ + { + "bbox": [ + 190, + 155, + 220, + 162 + ], + "type": "text", + "content": "w/o SAM" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 170, + 265, + 181 + ], + "lines": [ + { + "bbox": [ + 85, + 170, + 265, + 181 + ], + "spans": [ + { + "bbox": [ + 85, + 170, + 265, + 181 + ], + "type": "text", + "content": "Figure 7. Visual Comparisons of ablation studies." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 239, + 118, + 294, + 154 + ], + "blocks": [ + { + "bbox": [ + 239, + 118, + 294, + 154 + ], + "lines": [ + { + "bbox": [ + 239, + 118, + 294, + 154 + ], + "spans": [ + { + "bbox": [ + 239, + 118, + 294, + 154 + ], + "type": "image", + "image_path": "b4cabf75829a86d91e682f80cc4e52c2cf0e6cc82e2ddae022d76e61a520e4dc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 155, + 285, + 162 + ], + "lines": [ + { + "bbox": [ + 248, + 155, + 285, + 162 + ], + "spans": [ + { + "bbox": [ + 248, + 155, + 285, + 162 + ], + "type": "text", + "content": "w/o ResNet" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 186, + 295, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 295, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 295, + 234 + ], + "type": "text", + "content": "uations are carried out on FiveK for the enhancement task and on three high-level task datasets for other tasks. We select ResNet18/Top-1 Acc, S3FD/AP, and MaskR-CNN/mAP segm as the high-level metrics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 56, + 235, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 235, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 235, + 296, + 533 + ], + "type": "text", + "content": "Ablation Studies on FoFA. We employed a combination of different models to examine the supervisory capabilities of various foundation models at the feature level in the enhancement process. As shown in Table 3, removing any single foundation model from the FoFA negatively impacts the performance of the enhancement model across various tasks, and removing all of them results in even poorer outcomes. Overall, excluding CLIP and SAM reduces the model's adaptability for high-level tasks, while the absence of ResNet leads to accuracy losses in image enhancement tasks. This is likely because, during FoFA training, CLIP and SAM primarily contribute robust semantic learning, which is critical for high-level tasks; in contrast, ResNet is more responsive to specific degradations and lighting variations, which helps the discriminator perceive visual quality in images. Our observations further validate this theory: as shown in Figure 7, removing ResNet results in significant lighting changes in the enhanced images, while omitting CLIP or SAM introduces blurred and distorted areas. If all foundation models are excluded, FoFA degrades to a conventional adversarial training process, where semantic losses in the enhancement results become more severe. In summary, each foundation model in FoFA contributes positively to the training of the enhancement model, and their integration yields the best enhancement outcomes." + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 58, + 573, + 295, + 638 + ], + "blocks": [ + { + "bbox": [ + 55, + 536, + 295, + 569 + ], + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 569 + ], + "type": "text", + "content": "Table 3. Quantitative analysis from ablation studies demonstrating the impact of foundation feature space alignment. The best results are marked in bold." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 573, + 295, + 638 + ], + "lines": [ + { + "bbox": [ + 58, + 573, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 58, + 573, + 295, + 638 + ], + "type": "table", + "html": "
PSNRSSIMLPIPSAPmAP~segmTop-1 Acc
w FoFA19.580.830.130.680.280.59
w/o FoFA18.070.810.200.640.240.54
-CLIP18.670.810.150.650.230.53
-SAM18.950.820.170.640.260.54
-ResNet18.240.810.190.670.270.57
", + "image_path": "a52ad51443a35457edb627cba1956fbc4a61eae17caaa123cb204305160d8c6d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Ablation Studies on GaTA. To assess the role of GaTA in training the enhancement model, we first examined the impact of removing GaTA, as shown in Table 4. Removing GaTA negatively affected model performance in both the enhancement and high-level tasks. This is because GaTA primarily functions to detect incomplete degradation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": "elimination in the enhanced results, thereby improving the model's performance across various tasks. As illustrated in Figure 7, removing GaTA significantly increased the severity of degradation in the enhanced results, which not only impaired visual quality but also compromised the original semantic information." + } + ] + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 317, + 183, + 553, + 239 + ], + "blocks": [ + { + "bbox": [ + 313, + 147, + 555, + 179 + ], + "lines": [ + { + "bbox": [ + 313, + 147, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 555, + 179 + ], + "type": "text", + "content": "Table 4. Quantitative analysis from ablation studies demonstrating the Gap-driven Task-based Alignment. The best results are marked in bold." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 183, + 553, + 239 + ], + "lines": [ + { + "bbox": [ + 317, + 183, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 183, + 553, + 239 + ], + "type": "table", + "html": "
PSNRSSIMLPIPSAPmAP~segmTop-1 Acc
w GaTA19.580.830.130.680.280.59
w/o GaTA18.010.810.160.670.250.52
+GaTA Seg18.500.790.190.670.300.49
+GaTA Det18.750.820.190.700.270.54
", + "image_path": "4cc131984994637129e193190e7bfae0e965cfa457012262f0d17263bb1f2476.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 244, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 555, + 472 + ], + "type": "text", + "content": "Additionally, in Section 3.3, we implement GaTA using a classification task, with CLIP and ResNet employed to achieve " + }, + { + "bbox": [ + 313, + 244, + 555, + 472 + ], + "type": "inline_equation", + "content": "F^l" + }, + { + "bbox": [ + 313, + 244, + 555, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 244, + 555, + 472 + ], + "type": "inline_equation", + "content": "F^s" + }, + { + "bbox": [ + 313, + 244, + 555, + 472 + ], + "type": "text", + "content": ", respectively. Since GaTA leverages robustness differences between models to perceive degradation, various tasks—such as object detection or semantic segmentation—could potentially be used for GaTA as well. To test this, we trained " + }, + { + "bbox": [ + 313, + 244, + 555, + 472 + ], + "type": "inline_equation", + "content": "F^l" + }, + { + "bbox": [ + 313, + 244, + 555, + 472 + ], + "type": "text", + "content": " on a dataset with gamma transformations and noise augmentations, and " + }, + { + "bbox": [ + 313, + 244, + 555, + 472 + ], + "type": "inline_equation", + "content": "F^s" + }, + { + "bbox": [ + 313, + 244, + 555, + 472 + ], + "type": "text", + "content": " on the high-quality subset of the dataset, filtered using advanced image quality assessment metric Q-Instruct [26]. In practice, we usedYOLOX [5], trained on COCO, and PSPNet [35], trained on ADE20k [38], to represent models for object detection and semantic segmentation, denote as Det and Seg respectively. Detailed experimental settings are provided in the supplementary materials. As shown in Table 4, both object detection and semantic segmentation tasks effectively implement GaTA, achieving performance comparable to ResNet and CLIP, further demonstrating the generalizability of the proposed method." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 483, + 388, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 388, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 388, + 495 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 503, + 555, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 555, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 555, + 706 + ], + "type": "text", + "content": "We propose a framework named FoCo designed to enhance the visual and applicational performance of low-light image enhancers by leveraging the collaborative potential of diverse foundation models. At the feature level, our approach uses Feature Semantic Alignment (FoFA) to integrate semantically relevant features into adversarial training, aligning the discriminator and the foundation model's output features within a unified feature space. This alignment promotes feature-level consistency between enhanced and high-quality images. At the task level, we introduce Gap-driven task-based Alignment (GaTA), which utilizes the robustness disparities across models by using the less robust model as a degradation detector, thereby further improving enhancement quality. The proposed framework exhibits outstanding performance, providing a novel training paradigm that integrates self-supervised low-light enhancement methods with high-level applications." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16078" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 128 + ], + "type": "text", + "content": "Acknowledgements: This research was funded through Anhui Provincial Natural Science Foundation (Grant 2408085J024), National Natural Science Foundation of China (Grant 62401532), and Fundamental Research Funds for the Central Universities (Grant WK2090000065)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 168, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 168, + 294, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 168, + 294, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 168, + 294, + 212 + ], + "type": "text", + "content": "[1] Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Fredo Durand. Learning photographic global tonal adjustment with a database of input/output image pairs. In CVPR 2011, pages 97-104. IEEE, 2011. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 214, + 294, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 214, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 61, + 214, + 294, + 257 + ], + "type": "text", + "content": "[2] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. arXiv preprint arXiv:2303.06705, 2023. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "text", + "content": "[3] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. International Journal of Computer Vision, 131(8):2198-2218, 2023. 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 303, + 294, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 294, + 356 + ], + "type": "text", + "content": "[4] Tomáš Chobola, Yu Liu, Hanyi Zhang, Julia A Schnabel, and Tingying Peng. Fast context-based low-light image enhancement via neural implicit representations. In European Conference on Computer Vision, pages 413-430. Springer, 2025. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 358, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 294, + 390 + ], + "type": "text", + "content": "[5] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 456 + ], + "type": "text", + "content": "[6] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1780-1789, 2020. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 294, + 491 + ], + "type": "text", + "content": "[7] Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE Transactions on image processing, 26(2):982-993, 2016. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 492, + 294, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 294, + 545 + ], + "type": "text", + "content": "[8] Jiang Hai, Zhu Xuan, Ren Yang, Yutong Hao, Fengzhu Zou, Fang Lin, and Songchen Han. R2rnet: Low-light image enhancement via real-low to real-normal network. Journal of Visual Communication and Image Representation, 90: 103712, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 547, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 294, + 591 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "type": "text", + "content": "[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 678 + ], + "type": "text", + "content": "[11] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. Enlightengan: Deep light enhancement without paired supervision. IEEE transactions on image processing, 30:2340-2349, 2021. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[12] Attila Lengyel, Sourav Garg, Michael Milford, and Jan C van Gemert. Zero-shot day-night domain adaptation with a physics prior. In Proceedings of the IEEE/CVF International" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "Conference on Computer Vision, pages 4399-4409, 2021. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[13] Chongyi Li, Chunle Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4225-4238, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 140, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 555, + 194 + ], + "type": "text", + "content": "[14] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd: dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 195, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 249 + ], + "type": "text", + "content": "[15] Dong Liang, Ling Li, Mingqiang Wei, Shuo Yang, Liyan Zhang, Wenhan Yang, Yun Du, and Huiyu Zhou. Semantically contrastive learning for low-light image enhancement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1555-1563, 2022. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 251, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 304 + ], + "type": "text", + "content": "[16] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 306, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 306, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 317, + 306, + 553, + 371 + ], + "type": "text", + "content": "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 373, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 373, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 373, + 553, + 426 + ], + "type": "text", + "content": "[18] Risheng Liu, Long Ma, Jiao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10561-10570, 2021. 1, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 427, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 553, + 491 + ], + "type": "text", + "content": "[19] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 494, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 553, + 525 + ], + "type": "text", + "content": "[20] Yuen Peng Loh and Chee Seng Chan. Getting to know low-light images with the exclusively dark dataset. Computer Vision and Image Understanding, 178:30-42, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "type": "text", + "content": "[21] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 582, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 645 + ], + "type": "text", + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[23] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[24] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, Lap-Pui Chau, and Alex Kot. Low-light image enhancement with" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16079" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "normalizing flow. In Proceedings of the AAAI conference on artificial intelligence, pages 2604-2612, 2022. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[25] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 1, 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 196 + ], + "type": "text", + "content": "[26] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25490-25500, 2024. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "type": "text", + "content": "[27] Yuhui Wu, Chen Pan, Guoqing Wang, Yang Yang, Jiwei Wei, Chongyi Li, and Heng Tao Shen. Learning semantic-aware knowledge guidance for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1662-1671, 2023. 2, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "text", + "content": "[28] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5525-5533, 2016. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 364 + ], + "type": "text", + "content": "[29] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12918-12927, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 430 + ], + "type": "text", + "content": "[30] Wenhan Yang, Ye Yuan, Wenqi Ren, Jiaying Liu, Walter J Scheirer, Zhangyang Wang, Taiheng Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, et al. Advancing image understanding in poor visibility environments: A collective benchmark study. IEEE Transactions on Image Processing, 29: 5737-5752, 2020. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 432, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 294, + 487 + ], + "type": "text", + "content": "[31] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 488, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 532 + ], + "type": "text", + "content": "[32] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z Li. S3fd: Single shot scale-invariant face detector. In Proceedings of the IEEE international conference on computer vision, pages 192-201, 2017. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 533, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 577 + ], + "type": "text", + "content": "[33] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Proceedings of the 27th ACM international conference on multimedia, pages 1632-1640, 2019. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 578, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 610 + ], + "type": "text", + "content": "[34] Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision, 129:1013-1037, 2021. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "text", + "content": "[35] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[36] Chuanjun Zheng, Daming Shi, and Wentian Shi. Adaptive unfolding total variation network for low-light image enhancement. pages 4439-4448, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[37] Shen Zheng and Gaurav Gupta. Semantic-guided zero-shot learning for low-light image/video enhancement. In Pro" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 194 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 581-590, 2022. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[38] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "text", + "content": "[39] Han Zhou, Wei Dong, Xiaohong Liu, Shuaicheng Liu, Xiongkuo Min, Guangtao Zhai, and Jun Chen. Glare: Low light image enhancement via generative latent feature based codebook retrieval. arXiv preprint arXiv:2407.12431, 2024. 1, 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16080" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/9b5acaea-53bf-4055-87a9-1a5bf68ccd00_content_list.json b/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/9b5acaea-53bf-4055-87a9-1a5bf68ccd00_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c7c194495723307523341ed4710857862624ccd0 --- /dev/null +++ b/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/9b5acaea-53bf-4055-87a9-1a5bf68ccd00_content_list.json @@ -0,0 +1,1587 @@ +[ + { + "type": "text", + "text": "Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling", + "text_level": 1, + "bbox": [ + 200, + 130, + 795, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhaoyu Zhang $^{1}$ , Yang Hua $^{1}$ , Guanxiong Sun $^{2}$ , Hui Wang $^{1}$ , Seán McLoone $^{1}$ $^{1}$ Queen's University Belfast $^{2}$ Huawei UKRD", + "bbox": [ + 199, + 200, + 795, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zzhang55,Y.Hua,gsun02,h.wang,s.mcloone}@qub.ac.uk", + "bbox": [ + 274, + 241, + 718, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Data-Efficient Generative Adversarial Nets (DE-GANs) have become more and more popular in recent years. Existing methods apply data augmentation, noise injection and pre-trained models to maximally increase the number of training samples thus improving the training of DE-GANs. However, none of these methods considers the sample quality during training, which can also significantly influence the training of DE-GANs. Focusing on sample quality during training, in this paper, we are the first to incorporate discriminator rejection sampling (DRS) into the training process and introduce a novel method, called quality aware dynamic discriminator rejection sampling (QADDRS). Specifically, QADDRS consists of two steps: (1) the sample quality aware step, which aims to obtain the sorted critic scores, i.e., the ordered discriminator outputs, on real/fake samples in the current training stage; (2) the dynamic rejection step that obtains dynamic rejection number $N$ , where $N$ is controlled by the overfitting degree of discriminator $(D)$ during training. When updating the parameters of $D$ , the $N$ high critic score real samples and the $N$ low critic score fake samples in the minibatch are rejected dynamically based on the overfitting degree of $D$ . As a result, QADDRS can avoid $D$ becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of $D$ issue during training. Extensive experiments on several datasets demonstrate that integrating QADDRS into different DE-GANs can achieve better performance and deliver state-of-the-art results. Codes are available at https://github.com/zzhang05/QADDRS.", + "bbox": [ + 88, + 324, + 485, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 797, + 220, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past decade, Generative Adversarial Nets (GANs) [12] have demonstrated impressive performance in synthesizing images and videos [20, 21, 23, 59, 65]. However, this success heavily depends on access to large-scale datasets, which can be costly, time-consuming, and even impractical", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/03b43eb97c5800474cd86606b3972e4f89afc9c3bbf1e169222e656ee077a452.jpg", + "image_caption": [ + "Figure 1. Comparison of discriminator outputs on both real and fake samples with StyleGAN2 and StyleGAN2 + QADDRS using the 100-shot Obama dataset. Applying QADDRS to StyleGAN2 can effectively prevent $D$ from being overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of $D$ issue." + ], + "image_footnote": [], + "bbox": [ + 511, + 289, + 910, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to collect and clean in certain domains. As a result, data-efficient GANs (DE-GANs) [28, 38, 39] have garnered significant attention.", + "bbox": [ + 511, + 628, + 906, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies have applied techniques such as data augmentation [8, 9, 19, 22, 66], noise injection [55, 63] and pre-trained models [25, 44] to improve the DE-GANs training. All of these methods aim to maximally increase the number of training samples under data-efficient scenarios. However, none of these methods considers the sample quality during training, which can also significantly influence the DE-GANs training. For instance, as pointed out in several existing approaches [22, 55, 64], the highly undesirable quality samples produced by data augmentation or noise injection can be detrimental to DE-GANs training.", + "bbox": [ + 509, + 674, + 908, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Focusing on sample quality during training, in this paper, we are the first to introduce discriminator rejection sampling (DRS) [5] into the training process and propose a novel method, termed quality aware dynamic discrimina", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30682", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tor rejection sampling (QADDRS). Specifically, QADDRS comprises two steps: (1) the sample quality aware step that obtains the ranked critic scores, i.e., the sorted discriminator $(D)$ outputs, on real/fake samples to measure the quality of the real/fake samples in the current training stage and (2) the dynamic rejection step which is determined by the dynamic rejection number $N$ , where $N$ adapts based on the overfitting degree of $D$ . During updating the parameters of $D$ process, the $N$ real samples with the high critic scores and the $N$ fake samples with the low critic score in the minibatch are dynamically rejected according to the overfitting degree of $D$ . As illustrated in Figure 1, QADDRS can effectively prevent $D$ from becoming overly confident in distinguishing real and fake samples, thereby alleviating the overfitting of $D$ issue during training.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, the key contributions are shown as follows: (1) We introduce DRS into the GANs training process and propose a novel method named quality aware dynamic discriminator rejection sampling (QADDRS) to improve DEGANs training. As far as we are aware, QADDRS is the first approach that considers sample quality during DE-GANs training. (2) We present the theoretical analysis of QADDRS in GANs training, proving its robustness and rationality. Furthermore, we observe that applying QADDRS to DE-GANs can avoid $D$ becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of $D$ issue during training. (3) Experiments on low-shot [66] and FFHQ [23] datasets with different methods demonstrate that QADDRS can improve performance across various DE-GANs and deliver state-of-the-art performance over existing GANs and diffusion models.", + "bbox": [ + 89, + 325, + 482, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 608, + 232, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Generative Adversarial Nets (GANs)", + "text_level": 1, + "bbox": [ + 89, + 642, + 408, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the past ten years, generative adversarial nets (GANs) [12] have been among the most popular generative models [41, 50]. GANs consist of two networks: a generator $(G)$ and a discriminator $(D)$ , which engage in a min-max game. The $G$ seeks to generate realistic-looking synthetic samples from noise input, while the $D$ attempts to distinguish whether the input sample originates from the $G$ or from real samples. GANs are known to suffer from training instability, often resulting in low-quality and limited diversity in generated data [62]. To overcome this challenge, different approaches have been presented, including more advanced architectures [6, 34, 35, 42, 58, 61], more stable objective functions [4, 14, 15, 33, 43], and improved training strategies [10, 20, 28, 31, 32, 57, 60] to generate more realistic and diverse samples.", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Data-Efficient GANs", + "text_level": 1, + "bbox": [ + 513, + 90, + 709, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, training Data-Efficient GANs (DE-GANs) has gained increasing significance. Studies [8, 9, 19, 22, 66] have shown that DE-GAN training often suffers from discriminator $(D)$ overfitting. To address this issue, different approaches have been introduced, including novel data augmentation techniques [8, 9, 19, 22, 66], improved regularization methods for $D$ [11, 30, 48], noise injection [55, 63], and pre-trained models [25, 44].", + "bbox": [ + 511, + 113, + 903, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Rejection Sampling in Generative Models", + "text_level": 1, + "bbox": [ + 511, + 242, + 867, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, rejection sampling has played a significant role in increasing the generated image quality as a posttraining technique. Rejection sampling is commonly used in several kinds of generative models, such as GANs [5], diffusion models [37] and IMLE [51]. Among these methods, the most well-known method in GANs is discriminator rejection sampling (DRS). DRS has inspired the development of numerous techniques in GANs, such as MH-GAN [49], DDLS [7], DOT [46], DGflow [2] and LatentRS [18]. Although these studies have achieved great success in rejecting lower-quality generated samples, they are all posttraining methods and cannot directly benefit the training of DE-GANs. In this study, we are the first to incorporate the DRS in the training stage to improve the DE-GANs training by considering sample quality.", + "bbox": [ + 511, + 266, + 906, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 505, + 648, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Quality Aware Dynamic Discriminator Rejection Sampling (QADDRS)", + "text_level": 1, + "bbox": [ + 511, + 530, + 903, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As discussed in §2.3, existing discriminator rejection sampling (DRS) methods are rejecting low-quality generated samples in the post-training process to improve performance. However, it is infeasible to directly apply DRS to the training process, as it can cause the number of real and generated samples for the input of $D$ to become unbalanced during training.", + "bbox": [ + 511, + 568, + 905, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel approach called quality aware dynamic discriminator rejection sampling (QADDRS) to improve DE-GANs training by considering sample quality. The overview of applying QADDRS to update both $G$ and $D$ is shown in Figure 2. Specifically, when updating $G$ , all of the fake samples are used to update the parameters of $G$ , following the prior approaches [22, 66]. In contrast, when updating $D$ , the $N$ high critic score real samples and the $N$ low critic score fake samples in the minibatch are dynamically rejected, and the rest of the real and fake samples are used to update the parameters of $D$ . Consequently, QADDRS can prevent $D$ from becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of $D$ issue during training. The selection of high critic score real samples and low critic score", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30683", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0064759bf4c6f04c7a3d16640292945654e7af9dd25413af1f4f0bd16ac46857.jpg", + "image_caption": [ + "Figure 2. An overview of quality aware dynamic discriminator rejection sampling (QADDRS) for updating $G$ (left) and updating $D$ (right) in DE-GANs training. When updating $G$ , all of the generated samples are used to update the parameters of $G$ . When updating $D$ , the $N$ high critic score real samples and $N$ low critic score fake samples in the minibatch are dynamically rejected, and only the rest of the real and fake samples are used to update the parameters of $D$ . The selection of high critic score real samples and low critic score fake samples is guided by the discriminator outputs on both real and fake data during training. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 883, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fake samples is guided by the sample quality-aware step, while the rejection of $N$ real and fake samples is controlled by the dynamic rejection step.", + "bbox": [ + 88, + 489, + 480, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the sample quality aware step, the quality of real/fake samples can be represented by different critic scores on real/fake samples. Based on this, we sort real/fake samples via critic scores, which are based on the discriminator outputs on real/fake data in the current training stage.", + "bbox": [ + 88, + 536, + 482, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the dynamic rejection step, the dynamic rejection number $N$ is introduced to control sample rejection. It is obvious that rejecting real samples in QADDRS is highly undesirable, especially under data-efficient scenarios, which could cause a further reduction of training samples. To address this, we introduce a dynamic mechanism for the QADDRS by setting the rejection number $N$ to be adjusted dynamically according to the degree of overfitting without manual tuning. To achieve this, following ANDA [64], APA [19] and ADA [22], we first utilize an overfitting heuristic $\\eta$ which can quantify the $D$ 's overfitting degree as follows:", + "bbox": [ + 88, + 613, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\eta = \\mathbb {E} \\left(\\operatorname {s i g n} \\left(D _ {\\text {r e a l}}\\right)\\right), D _ {\\text {r e a l}} = \\operatorname {l o g i t} (D (x)), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 813, + 482, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{sign}(\\cdot)$ denotes the sign function, which returns $+1$ for a non-negative input; $-1$ , otherwise. Next, we follow ADA, APA and ANDA by using $\\eta$ to obtain the overfitting probability $p$ . The larger value of $p$ means the greater over", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/98089418bba55c5446e5d788e39e3711ea66ff7c92b0c78cbb58203d6c39a196.jpg", + "image_caption": [ + "Figure 3. The curve of the average number of real samples rejected by QADDRS $(\\lambda = 0.25)$ when updating $D$ for every 4K fake samples shown to $D$ during training. The experiments are conducted on the 100-shot Obama dataset using StyleGAN2." + ], + "image_footnote": [], + "bbox": [ + 511, + 487, + 911, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fitting degree. Next, we set the initial rejection number $N$ as 0, and adjust its value based on the overfitting probability $p$ as follows:", + "bbox": [ + 511, + 771, + 906, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nN = \\left\\{ \\begin{array}{l l} N + 1 & (p) \\\\ N - 1 & (1 - p) \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 830, + 903, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N$ is adjusted to $N + 1$ under the probability $p$ and is adjusted to $N - 1$ under the probability $1 - p$ . We introduce", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30684", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a hyperparameters $\\lambda$ and set the range of $N$ as $N_{\\min} = 0$ and $N_{\\max} = \\lambda \\times$ batchsize. To further demonstrate how the number $N$ evolves during the training process, the value of $N$ in every 4K real images shown to $D$ during training can be found in Figure 3.", + "bbox": [ + 89, + 90, + 483, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, employing QADDRS in DE-GANs can be formulated as:", + "bbox": [ + 89, + 167, + 483, + 198 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {D} (G, D) = \\mathbb {E} _ {x \\sim P _ {R}} [ \\tilde {g} _ {1} (D (x _ {m - N})) ] + \\mathbb {E} _ {x \\sim P _ {G}} [ \\tilde {g} _ {2} (D (x _ {m - N})) ],\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 215, + 485, + 233 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {G} (G, D) = - \\mathbb {E} _ {x \\sim P _ {G}} [ \\tilde {f} ((D (x _ {m}))) ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 234, + 483, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{m}$ is a batch of samples, $m$ denotes the total samples number in a batch, and $\\tilde{g}_{1}$ , $\\tilde{g}_{2}$ and $\\tilde{f}$ are scalar-to-scalar functions. For the StyleGAN2 [23] with non-saturating loss, $\\tilde{g}_{1}(x) = \\log (x)$ , $\\tilde{g}_{2}(x) = \\log (1 - x)$ and $\\tilde{f}(x) = \\log (x)$ . For the FastGAN [30] with hinge loss, $\\tilde{g}_{1}(x) = \\min (0, -1 + x)$ , $\\tilde{g}_{2}(x) = \\min (0, -1 - x)$ and $\\tilde{f}(x) = x$ .", + "bbox": [ + 89, + 266, + 483, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Theoretical Analysis", + "text_level": 1, + "bbox": [ + 89, + 378, + 284, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 QADDRS in GANs training", + "text_level": 1, + "bbox": [ + 89, + 402, + 339, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further demonstrate how QADDRS can influence GANs training in theory, we first revisit the hinge loss in GANs shown below:", + "bbox": [ + 89, + 426, + 483, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} V _ {D} (D, G) = \\mathbb {E} _ {x \\sim P _ {R}} [ \\min (0, - 1 + D (x)) ] \\\\ + \\mathbb {E} _ {x \\sim P _ {G}} [ \\min (0, - 1 - D (x)) ], \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 482, + 480, + 516 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {G} (D, G) = - \\mathbb {E} _ {x \\sim P _ {G}} [ D (x) ].\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 518, + 344, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For Eq.(4), $D(x) > 1$ on the real data or $D(G(z)) < -1$ on the fake data can lead to the loss value becoming 0, resulting in no gradient for updating $D$ . Therefore, the key essence of hinge loss in GANs training is to reject the very high critic score real samples (score $> 1$ ) and very low critic score fake samples (score $< -1$ ). Following this, our proposed QADDRS further rejects the real and fake samples based on the overfitting degree of $D$ to enhance the DE-GANs training. We then consider how QADDRS influences the distributions of real data $(P_R)$ and fake data $(P_G)$ in GANs training as follows.", + "bbox": [ + 89, + 537, + 483, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to Figure 2, QADDRS can be regarded as involving two actions, i.e., rejecting low critic score fake samples and rejecting high critic score real samples. Because rejecting real/fake samples in QADDRS is performed dynamically, the key operations of QADDRS on the distributions are re-weighting both real and fake samples for their data distributions, respectively. Consequently, in rejecting the low critic score fake samples action, applying QADDRS can transform the original fake data distribution $(P_{G})$ to the re-weighted fake data distribution $(w_{1}P_{G})$ . Similarly, in rejecting the high critic score real samples action, utilizing QADDRS can transform the original real data distribution $(P_{R})$ to the re-weighted real data distribution $(w_{2}P_{R})$ .", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Convergence analysis of QADDRS in StyleGAN2 and FastGAN", + "text_level": 1, + "bbox": [ + 513, + 90, + 906, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the analysis in §3.2.1, we then analyze the QADDRS on both StyleGAN2 with non-saturating loss and FastGAN with hinge loss as follows.", + "bbox": [ + 511, + 128, + 906, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1. For the StyleGAN2 with non-saturating loss, QADDRS can effectively introduce the advantages of hinge loss into the non-saturating loss via dynamically rejecting the $N$ high critic score real samples and the $N$ low critic score fake samples. Following §3.2.1 and [3], the optimal discriminator of QADDRS with non-saturating loss is defined as:", + "bbox": [ + 514, + 176, + 906, + 282 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD ^ {*} (x) = \\frac {w _ {2} (x) P _ {R} (x)}{w _ {2} (x) P _ {R} (x) + w _ {1} (x) P _ {G} (x)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 292, + 905, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With $D^{*}$ as the optimal discriminator, according to Theorem 2.5 in [3], training the $G$ with non-saturating loss can be expressed as:", + "bbox": [ + 532, + 330, + 906, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {G} (G, D ^ {*}) = \\mathbf {K L} \\left(w _ {1} P _ {G} \\mid \\mid w _ {2} P _ {R}\\right) - 2 \\mathbf {J S} \\left(w _ {1} P _ {G} \\mid \\mid w _ {2} P _ {R}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 388, + 903, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where KL and JS denote the Kullback-Leibler divergence and the Jensen-Shannon divergence, respectively. Eq.(6) shows that applying QADDRS to the non-saturating loss can lead to the minimization of KL-2JS divergence between the $w_{1}P_{G}$ and $w_{2}P_{R}$ . Then, according to [3], this minimization process depends on the supports of $w_{1}P_{G}$ and $w_{2}P_{R}$ . Previous studies [6, 13, 21, 45] have established that re-weighting operations applied to distributions do not alter their support. Therefore, the minimization of Eq.(6) corresponds to the minimization of KL-2JS divergence between the $P_{G}$ and $P_{R}$ , which demonstrates that applying QADDRS to non-saturating loss does not influence convergence.", + "bbox": [ + 532, + 417, + 905, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2. For the FastGAN with hinge loss, QADDRS can further take advantage by dynamically rejecting the $N$ high critic score real samples and the $N$ low critic score fake samples if all $D(x)$ and $D(G(z))$ are in $[-1, 1]$ . According to the §3.2.1 and [29, 35, 47], optimizing the hinge loss with QADDRS is equivalent to minimizing the reverse KL divergence term $\\mathbf{KL}(w_1P_G\\parallel w_2P_R)$ . With the fact that the supports of $w_1P_G$ and $w_2P_R$ are the same as the supports of $P_G$ and $P_R$ , applying QADDRS to hinge loss does not influence the convergence.", + "bbox": [ + 513, + 614, + 906, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3 Analysis of QADDRS in DE-GANs", + "text_level": 1, + "bbox": [ + 513, + 786, + 808, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to the theory presented in [64], data augmentation (DA) is commonly used in DE-GANs, which transforms the real and fake data distributions as $P_R^T$ and $P_G^T$ , respectively. As a result, applying QADDRS can result in the optimization between the $w_1P_G^T$ and $w_2P_R^T$ for both StyleGAN2 with non-saturating loss and FastGAN with hinge", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30685", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/f87d08160e1eb3c8386b46563d9d2e5f8226cdcfad7fb1d2245c7531eed97545.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneMAPre-training100-shotAnimal-Face
ObamaGrumpy CatPandaCatDog
EDM [24]DiffusionYesNo51.3036.9023.7048.60100.10
EDM + DA [24]DiffusionYesNo37.1029.9410.8136.8857.14
Patch Diffusion [54]DiffusionYesNo41.4730.8913.2543.7172.17
Scale/shift [40]GANNoYes50.7234.2021.3854.8383.04
MineGAN [53]GANNoYes50.6334.5414.8454.4593.03
TransferGAN [52]GANNoYes48.7334.0623.2052.6182.38
TransferGAN + DA [66]GANYesYes39.8529.7717.1249.1065.57
FreezeD [36]GANNoYes41.8731.2217.9547.7070.46
StyleGAN2 [23]GANNoNo80.2048.9034.2771.71131.90
FakeCLR [28]GANYesNo26.9519.568.4226.3442.02
FastGAN [30]GANYesNo35.8025.759.7033.8552.46
FreGAN [58]GANYesNo33.3924.938.9731.0547.85
AugSelf-StyleGAN2 [17]GANYesNo26.0019.818.3630.5348.19
CHAIN [39]GANYesNo20.9417.576.9319.5830.88
NICE [38]GANYesNo20.0815.638.1822.7028.65
Projected GAN [44]GANYesYes11.2115.803.9818.0117.88
StyleGAN2* [23]GANYesNo65.5739.9225.0851.6687.96
+ QADDRSGANYesNo53.5435.7720.5947.8773.29
StyleGAN2 + Diff-Augment [66]GANYesNo46.8727.0812.0642.4458.85
+ QADDRSGANYesNo41.2726.5310.8338.1551.83
StyleGAN2 + ADA [22]GANYesNo45.6926.6212.9040.7756.83
+ QADDRSGANYesNo40.8326.2711.9936.2553.01
InsGen + ADA [56]GANYesNo32.4222.019.8533.0144.93
+ QADDRSGANYesNo25.4119.017.9923.6935.01
Diffusion-Projected GAN (FastGAN) [55]GANYesYes10.5415.133.3917.8617.22
+ QADDRSGANYesYes9.9314.723.0816.9416.57
", + "bbox": [ + 102, + 88, + 895, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. FID [16] scores (lower is better) of different methods on ${256} \\times {256}$ low-shot datasets. MA donates Massive Augmentation [9], i.e., xflipping. For a fair comparison, FID is computed using $5\\mathrm{\\;K}$ generated samples,with the entire dataset serving as the reference distribution. The results are averaged over three evaluations, with relative standard deviations below 1%.", + "bbox": [ + 89, + 513, + 903, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "loss. As the supports of $w_{1}P_{G}^{T}$ and $w_{2}P_{R}^{T}$ are the same as the supports of $P_{R}^{T}$ and $P_{G}^{T}$ , utilizing QADDRS in the DE-GANs does not influence convergence. Furthermore, DE-GANs are known to suffer from the leaking of augmentations issue [22, 64]. In this case, QADDRS can also reweight the out-of-distribution samples for both real and fake data to alleviate the leaking issue.", + "bbox": [ + 89, + 582, + 482, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 700, + 223, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 726, + 410, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We select FFHQ [23] and low-shot [66] datasets for the experiments. The FFHQ dataset includes 70K high-resolution face images. In line with FakeCLR [28], we preprocess and resize the FFHQ dataset to $256 \\times 256$ and select different subsets consisting of 100, 1K, 2K, and 5K images for all experiments. Low-shot datasets $(256 \\times 256)$ consist of five sets (i.e., Obama, Grumpy Cat, Panda, Animal-Face Cat, and Animal-Face Dog) with 100, 100, 100, 160, and 389 training samples, respectively. The commonly used Fréchet Inception Distance (FID) [16], CLIP FID (FIDCLIP) [27],", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and Precision and Recall [26] are applied as the evaluation metrics. Following prior methods [22, 66], the complete dataset serves as the reference distribution for all metric evaluations.", + "bbox": [ + 511, + 583, + 903, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The implementation details are shown as follows. For the experiments on low-shot datasets with StyleGAN2, StyleGAN2 + Diff-Augment and StyleGAN2 + ADA, we set the batchsize as 4 and the overall training duration as 2000K images. For the experiments on low-shot datasets with InsGen + ADA, we set the batchsize as 64 and the overall training duration as 35000K images. For the experiments on low-shot datasets with Diffusion-Projected GAN, we set the batchsize as 64 and the overall training duration as 100000K images. For the experiments on FFHQ datasets, we set the batchsize as 64 and the overall training duration as 25000K images. The hyperparameter $\\lambda$ is set as 0.25 in QADDRS across all experiments.", + "bbox": [ + 511, + 643, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results on Low-shot Datasets", + "text_level": 1, + "bbox": [ + 511, + 847, + 774, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The FID results of different methods on the $256 \\times 256$ low-shot datasets are shown in Table 1. We show the effec", + "bbox": [ + 511, + 869, + 903, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30686", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ac981f9b37758324ef0ff0e4c19a165a4d865ffb1e936256e5a260da2d7ba097.jpg", + "image_caption": [ + "FID 65.57" + ], + "image_footnote": [], + "bbox": [ + 151, + 104, + 274, + 199 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c61b1b5d47bdf4c5ba543de382fd94edf86075edc4eba83cb69112abcbc59914.jpg", + "image_caption": [ + "FID 25.08" + ], + "image_footnote": [], + "bbox": [ + 294, + 104, + 418, + 199 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2190ea3082bc14db5cb2fe545f5b2d786398f67499b4c8ac3ad8846da896977c.jpg", + "image_caption": [ + "FID 39.92" + ], + "image_footnote": [], + "bbox": [ + 437, + 104, + 560, + 199 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/46fb38eb15da47f622e9bfbccd1b9f8919b1f85fb1f78e948884a5ae9d775742.jpg", + "image_caption": [ + "FID 51.66" + ], + "image_footnote": [], + "bbox": [ + 581, + 104, + 704, + 198 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/33c564d0e6838947465d03631665a01db1cc92ab82aea250a0651d9b927a4061.jpg", + "image_caption": [ + "FID 87.96" + ], + "image_footnote": [], + "bbox": [ + 723, + 104, + 844, + 198 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5080debdd0e6835c5d637a65469e5f2416ce9fc9aef6605184e35a48dd657c7d.jpg", + "image_caption": [ + "FID 53.54 (-12.03)" + ], + "image_footnote": [], + "bbox": [ + 151, + 224, + 274, + 319 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/730253ce8eb5570485746e38557d1fb36dd31cf0db7dfba5bd0d692aa238d74d.jpg", + "image_caption": [ + "FID 20.59 (-4.49)" + ], + "image_footnote": [], + "bbox": [ + 295, + 224, + 416, + 318 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76b4a2945e3ea29a2237edd468cce315b1f3441c12ed026311a7d514a86449f5.jpg", + "image_caption": [ + "FID 35.77 (-4.15)", + "Figure 4. The compared generated results between StyleGAN2 and StyleGAN2 + QADDRS on low-shot datasets. Top: Images generated by StyleGAN2. Bottom: Images generated by StyleGAN2 + QADDRS. The reduction of FID in red color shows the improvement of StyleGAN2 + QADDRS over the baseline StyleGAN2." + ], + "image_footnote": [], + "bbox": [ + 437, + 224, + 562, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/526620215b5210c465687773949aed9aab00809885a32e318e4bc623bb2cd7c0.jpg", + "image_caption": [ + "FID 47.87 (-3.79)" + ], + "image_footnote": [], + "bbox": [ + 581, + 224, + 702, + 319 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8ce54616d2867846ca0ed8081d3f842689e8578b5363b0cfd31526990b6ff65.jpg", + "image_caption": [ + "FID 73.29 (-14.67)" + ], + "image_footnote": [], + "bbox": [ + 723, + 224, + 846, + 320 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d3af9ef67d6107b79753cc4e80ef7227c149ddc41c949b6d74cfe36b5a8ca048.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneMAPre-trainingFFHQ
1001K2K5K
EDM [24]DiffusionYesNo79.10---
EDM + DA [24]DiffusionYesNo50.7330.7527.1724.51
Patch Diffusion [54]DiffusionYesNo44.4528.0325.3222.63
StyleGAN2 [23]GANNoNo179.00100.1654.3049.68
StyleGAN2 + Diff-Augment [66]GANYesNo61.9125.6624.3210.45
StyleGAN2 + ADA [22]GANYesNo85.821.2915.3910.96
APA [19]GANYesNo65.0018.8916.908.38
AugSelf-StyleGAN2+ [17]GANYesNo-20.39-9.15
FakeCLR [28]GANYesNo42.5615.929.907.25
Projected GAN [44]GANYesYes26.2511.128.256.85
InsGen [56]GANYesNo45.7518.2111.477.83
+ QADDRSGANYesNo43.6717.1610.737.51
Diffusion-Projected GAN (FastGAN) [55]GANYesYes25.478.767.996.59
+ QADDRSGANYesYes24.468.517.886.48
Diffusion-Projected GAN (FastGAN Lite) [55]GANYesYes20.657.025.474.74
+ QADDRSGANYesYes19.896.575.184.69
", + "bbox": [ + 96, + 398, + 901, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. FID scores (lower is better) of different methods on the ${256} \\times {256}\\mathrm{{FFHQ}}$ dataset. In line with FakeCLR [28], the experiments are performed on 100, 1K, 2K and 5K subsets of the FFHQ dataset. MA donates Massive Augmentation [9], i.e., xflipping. The results are averaged over five evaluations, with relative standard deviations below 1%.", + "bbox": [ + 89, + 686, + 906, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tiveness of the QADDRS with both StyleGAN2 and FastGAN using five different commonly-used DE-GANs, i.e., StyleGAN2, StyleGAN2 + Diff-Augment, StyleGAN2 + ADA, InsGen + ADA and Diffusion-Projected GAN. As shown in Table 1, the proposed QADDRS can increase the performance of these five different DE-GANs, demonstrating that QADDRS can be regarded as a general plugin training strategy for DE-GANs. As shown in Table 7, Diffusion-Projected GAN (FastGAN) achieves better per", + "bbox": [ + 89, + 756, + 482, + 893 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "formance than Diffusion-Projected GAN (FastGAN Lite), i.e., applying the lite version generator in FastGAN, on low-shot datasets. Therefore, we report the results of the Diffusion-Projected GAN (FastGAN) in Table 1. Although Diffusion-Projected GAN (FastGAN) has already utilized all widely-used approaches, i.e., DA, regularizations, noise injection and pre-trained models, to boost performance on the low-shot datasets, adding QADDRS can further improve its performance. In addition, following [1], we also report", + "bbox": [ + 511, + 756, + 906, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30687", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/afc20dd83b3c7eac1a3853366f18fa8aa67b70754222b05902808373297767c8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodObamaGrumpy CatPandaAF-CatAF-Dog
PRPRPRPRPR
Diffusion-Projected GAN (FastGAN) + QADDRS0.9990.2400.9960.1490.9990.4100.9970.1820.9980.203
0.9990.2420.9980.1510.9990.4180.9980.1860.9990.203
", + "bbox": [ + 114, + 88, + 883, + 160 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b46c7c9e26c1e94d432c24c3c5789d1ed59d3602aeef248355507e5e1d00d655.jpg", + "table_caption": [ + "Table 3. The comparison of Precision (P) and Recall (R) (higher is better) with Diffusion-Projected GAN (FastGAN) [55] and Diffusion-Projected GAN (FastGAN) + QADDRS on the $256 \\times 256$ low-shot datasets. The results are averaged over three evaluations, with relative standard deviations below $1\\%$ ." + ], + "table_footnote": [], + "table_body": "
MethodFFHQ-100FFHQ-1KFFHQ-2KFFHQ-5K
PRPRPRPR
Diffusion-Projected GAN (FastGAN Lite) + QADDRS0.7370.0180.7340.2690.7360.3210.7300.377
0.7390.0210.7340.2760.7390.3230.7330.379
", + "bbox": [ + 127, + 237, + 870, + 305 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f727e50ec87cec3fb32d60c4025fe074163735a1590c2863ea242f8563e892d7.jpg", + "table_caption": [ + "Table 4. The comparison of Precision (P) and Recall (R) (higher is better) with Diffusion-Projected GAN (FastGAN Lite) [55] and Diffusion-Projected GAN (FastGAN Lite) + QADDRS on the $256 \\times 256$ FFHQ dataset. The results are averaged over five evaluations, with relative standard deviations below $1\\%$ ." + ], + "table_footnote": [], + "table_body": "
Method100AF
OGCPCD
Diffusion Projected GAN (FastGAN)2.024.782.598.367.21
+ QADDRS1.984.712.538.207.09
", + "bbox": [ + 91, + 393, + 488, + 460 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/99968b0db0c30fa29ab6624447835d0ccdd1d2ea63e6db34adb68f0b17b28b65.jpg", + "table_caption": [ + "Table 5. Compared $\\mathrm{FID}_{\\mathrm{CLIP}}$ scores between Diffusion Projected GAN (FastGAN) and Diffusion Projected GAN (FastGAN) + QADDRS on low-shot datasets. We use the FID calculation model to calculate the $\\mathrm{FID}_{\\mathrm{CLIP}}$ . Here, AF denotes AnimalFace, O denotes Obama, GC denotes Grumpy Cat, P denotes Panda, C denotes Cat and D denotes Dog." + ], + "table_footnote": [], + "table_body": "
MethodFFHQ
1001K2K5K
Diffusion Projected GAN (FL)23.9111.459.618.33
+ QADDRS23.3211.149.428.19
", + "bbox": [ + 94, + 590, + 478, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the results of several SOTA diffusion models in Table 1. It is clear that GAN-based methods outperform diffusion-based methods on low-shot datasets. To further show the effectiveness of QADDRS, the compared Precision and Recall, and $\\mathrm{FID}_{\\mathrm{CLIP}}$ of adding QADDRS to Diffusion-Projected GAN (FastGAN) are shown in Tables 3 and 5, respectively. The compared generated images on StyleGAN2 and StyleGAN2 + QADDRS are shown in Figure 4.", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c8bdf90f7cb1868ee2cdf4f5379d462664af73250dc00c4736d2bc5924f0e83e.jpg", + "table_caption": [ + "Table 6. Compared $\\mathrm{FID}_{\\mathrm{CLIP}}$ scores between Diffusion Projected GAN (FastGAN Lite) and Diffusion Projected GAN (FastGAN Lite) + QADDRS on FFHQ datasets. We use the FID calculation model to calculate the $\\mathrm{FID}_{\\mathrm{CLIP}}$ . FL denotes the FastGAN Lite." + ], + "table_footnote": [], + "table_body": "
MethodFID
Diffusion Projected GAN (FastGAN)17.22
Diffusion Projected GAN (FastGAN Lite)20.91
", + "bbox": [ + 539, + 392, + 883, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Compared FID scores of Diffusion Projected GAN (FastGAN) and Diffusion Projected GAN (FastGAN Lite) on $256\\times 256$ AnimalFace-dog dataset. The results are averaged over three evaluations, with relative standard deviations below $1\\%$ .", + "bbox": [ + 511, + 469, + 906, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Results on FFHQ Dataset", + "text_level": 1, + "bbox": [ + 511, + 556, + 746, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "According to the results of five different DE-GANs on low-shot datasets, InsGen + ADA and Diffusion-Projected GAN achieve SOTA performance on StyleGAN2 and FastGAN, respectively. Therefore, we conduct experiments on the FFHQ dataset by selecting these two types of DEGANs. Following FakeCLR [28], we conduct experiments on the FFHQ 100, FFHQ 1K, FFHQ 2K and FFHQ 5K subsets. As shown in Table 2, QADDRS can improve performance with these two DE-GANs on the FFHQ dataset. More importantly, we report the results of both the Diffusion-Projected GAN (FastGAN) and the Diffusion-Projected GAN (FastGAN Lite) in Table 2. It is clear that Diffusion-Projected GAN (FastGAN Lite) achieves better performance than Diffusion-Projected GAN (FastGAN) on the FFHQ dataset. Additionally, the comparison of Precision and Recall, as well as $\\mathrm{FID}_{\\mathrm{CLIP}}$ with Diffusion-Projected GAN (FastGAN Lite) and Diffusion-Projected GAN (FastGAN Lite) + QADDRS are shown in Tables 4 and 6, respectively. Diffusion-Projected GAN (FastGAN Lite) + QADDRS outperforms Diffusion-Projected GAN (FastGAN Lite) by adding QADDRS, which demonstrates the", + "bbox": [ + 511, + 583, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30688", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/58d4291dedc611acc2652b3cdce33ebab699456fba537ac0b61d4f2228e5c6f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
λ0.000.250.500.75
FID46.8741.2742.4043.32
", + "bbox": [ + 106, + 88, + 467, + 121 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1ddb58cfaf9f65aafe10040bb4b6589b8db358d859dde04460e16d207a235723.jpg", + "table_caption": [ + "Table 8. FID scores on the ${256} \\times {256100}$ -shot Obama dataset with StyleGAN2 + Diff-Augment. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below 1%." + ], + "table_footnote": [], + "table_body": "
λ0.000.250.500.75
FID58.8551.8353.3154.45
", + "bbox": [ + 106, + 215, + 467, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9b2c118ad3edc4a7fcc36007ea73505eacaf18c5fa1379b8eb87ab8ac0346006.jpg", + "table_caption": [ + "Table 9. FID scores on the ${256} \\times {256}$ AnimalFace-Dog dataset with StyleGAN2 + Diff-Augment. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below 1%." + ], + "table_footnote": [], + "table_body": "
MethodObamaAF-Dog
StyleGAN2 + Diff-Augment46.8758.85
+ Fixed rejecting number N44.1254.85
+ QADDRS41.2751.83
", + "bbox": [ + 120, + 343, + 454, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "effectiveness of the proposed QADDRS. To further show the superiority of QADDRS, more generated images on the FFHQ dataset can be found in supplementary materials.", + "bbox": [ + 89, + 515, + 482, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 570, + 254, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the hyperparameter $\\lambda$ in proposed QADDRS. To demonstrate the effectiveness of the hyperparameter $\\lambda$ in the proposed QADDRS, we conduct experiments by selecting different values of $\\lambda$ on the 100-shot Obama and AnimalFace-Dog datasets using StyleGAN2 + Diff-Augment. The results, shown in Tables 8 and 9, indicate that setting $\\lambda$ to 0.25 achieves better performance.", + "bbox": [ + 89, + 594, + 482, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the dynamic rejection number $N$ in the proposed QADDRS. To show that the dynamic rejection number $N$ in QADDRS is effective, we conduct an ablation study by comparing QADDRS with a fixed rejection number $N = 0.25 \\times$ batchsize on the 100-shot Obama dataset with StyleGAN2 + Diff-Augment. The results are shown in Table 10. It is obvious that the dynamically determined rejection number $N$ can achieve better performance compared with the fixed rejection number.", + "bbox": [ + 89, + 700, + 482, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Comparisons of DRS and QADDRS", + "text_level": 1, + "bbox": [ + 89, + 845, + 401, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct the experiments by comparing QADDRS and the existing DRS method [5] using StyleGAN2 + Diff-", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/30995c6765630bda2920e6f8b6b2248bfc1866b38afdf9260ae41dd876602275.jpg", + "table_caption": [ + "Table 10. FID scores on the $256 \\times 256$ 100-shot Obama and AnimalFace-Dog datasets. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below $1\\%$ ." + ], + "table_footnote": [], + "table_body": "
MethodObamaAnimalFace-Dog
StyleGAN2 + Diff-Augment46.8758.85
+ DRS [5]45.2256.17
+ QADDRS41.2751.83
", + "bbox": [ + 519, + 87, + 898, + 148 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/67477d438ce87671e978de52c35d014fbcd436b385c2ea8c372e928e7babff7e.jpg", + "table_caption": [ + "Table 11. FID scores on the ${256} \\times {256}\\;{100}$ -shot Obama and Animal-Face Dog datasets. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below 1%." + ], + "table_footnote": [], + "table_body": "
MethodSeconds / 1K images
Diffusion-Projected GAN (FastGAN) + QADDRS17\n17.25
", + "bbox": [ + 514, + 242, + 915, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 12. The training time on the ${256} \\times {256}\\;{100}$ -shot Obama dataset with or without QADDRS using Diffusion-Projected GAN (FastGAN). The results are averaged over ten evaluations on two NVIDIA RTX A5000 GPUs with a batch size of 64. All standard deviations are below 1% relative to the mean.", + "bbox": [ + 511, + 313, + 906, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Augment on the 100-shot Obama and AnimalFace-dog datasets. The results are shown in Table 11. Based on ADA [22], the overfitting of $D$ during training finally results in the DE-GANs generating low-quality images under limited data settings. Although the existing DRS method can reject lower-quality generated samples during post-training, they can not benefit the training of DE-GANs. Thus, existing DRS methods can not prevent generating low-quality images in average, leaving limited gains under limited data settings. In contrast, QADDRS can alleviate the overfitting issue during training, thereby effectively avoiding producing low-quality images in average and achieving significant improvements. As a result, QADDRS outperforms the existing DRS method for DE-GANs.", + "bbox": [ + 511, + 407, + 906, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Computational Cost", + "text_level": 1, + "bbox": [ + 511, + 628, + 705, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results of the training time on the 100-shot Obama dataset with or without QADDRS using Diffusion-Projected GAN (FastGAN) have been shown in Table 12. The proposed QADDRS results in a negligible increase in computational cost.", + "bbox": [ + 511, + 651, + 905, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 739, + 633, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel method called quality aware dynamic discriminator rejection sampling (QADDRS), to improve DE-GANs training by considering sample quality. QADDRS can effectively avoid $D$ becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of $D$ issue. Extensive experiments on multiple datasets show that applying QADDRS to various DE-GANs can achieve better performance and deliver state-of-the-art results.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30689", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is partially supported by the High Performance Computing center of Queen's University Belfast and the Kelvin-2 supercomputer (EPSRC grant EP/T022175/1). First author is supported by a scholarship of the Queen's University Belfast and the Department for the Economy, Northern Ireland.", + "bbox": [ + 89, + 114, + 483, + 204 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 217, + 187, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mehran Aghabozorgi, Shichong Peng, and Ke Li. Adaptive imle for few-shot pretraining-free generative modelling. In ICML, 2023. 6", + "[2] Abdul Fatir Ansari, Ming Liang Ang, and Harold Soh. Refining deep generative models via discriminator gradient flow. arXiv preprint arXiv:2012.00780, 2020. 2", + "[3] Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks. In ICLR, 2017. 4", + "[4] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In ICML, 2017. 2", + "[5] Samaneh Azadi, Catherine Olsson, Trevor Darrell, Ian Goodfellow, and Augustus Odena. Discriminator rejection sampling. In ICLR, 2019. 1, 2, 8", + "[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 2, 4", + "[7] Tong Che, Ruixiang Zhang, Jascha Sohl-Dickstein, Hugo Larochelle, Liam Paull, Yuan Cao, and Yoshua Bengio. Your gan is secretly an energy-based model and you should use discriminator driven latent sampling. In NeurIPS, 2020. 2", + "[8] Tianlong Chen, Yu Cheng, Zhe Gan, Jingjing Liu, and Zhangyang Wang. Data-efficient GAN training beyond (just) augmentations: A lottery ticket perspective. In NeurIPS, 2021. 1, 2", + "[9] Kaiwen Cui, Jiaxing Huang, Zhipeng Luo, Gongjie Zhang, Fangneng Zhan, and Shijian Lu. Genco: Generative constraining for generative adversarial networks with limited data. In AAAI, 2022. 1, 2, 5, 6, 8", + "[10] Emily L Denton, Soumith Chintala, Rob Fergus, et al. Deep generative image models using a laplacian pyramid of adversarial networks. In NeurIPS, 2015. 2", + "[11] Tiantian Fang, Ruoyu Sun, and Alex Schwing. Diggan: Discriminator gradient gap regularization for GAN training with limited data. In NeurlPS, 2022. 2", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2", + "[13] Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Scholkopf, and Alexander Smola. A kernel two-sample test. The Journal of Machine Learning Research, 13(1):723-773, 2012. 4", + "[14] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein GANs. In NeurIPS, 2017. 2" + ], + "bbox": [ + 93, + 242, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Tianyu Guo, Chang Xu, Jiajun Huang, Yunhe Wang, Boxin Shi, Chao Xu, and Dacheng Tao. On positive-unlabeled classification in GAN. In CVPR, 2020. 2", + "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5", + "[17] Liang Hou, Qi Cao, Yige Yuan, Songtao Zhao, Chongyang Ma, Siyuan Pan, Pengfei Wan, Zhongyuan Wang, Huawei Shen, and Xueqi Cheng. Augmentation-aware self-supervision for data-efficient gan training. In NeurIPS, 2023. 5, 6", + "[18] Thibaut Issenhuth, Ugo Tanielian, David Picard, and Jeremie Mary. Latent reweighting, an almost free improvement for gans. In WACV, 2022. 2", + "[19] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Deceive d: Adaptive pseudo augmentation for GAN training with limited data. In NeurIPS, 2021. 1, 2, 3, 6", + "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 1, 2", + "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 4", + "[22] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, 2020. 1, 2, 3, 5, 6, 8", + "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of styleGAN. In CVPR, 2020. 1, 2, 4, 5, 6", + "[24] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 5, 6", + "[25] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for GAN training. In CVPR, 2022. 1, 2", + "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In NeurIPS, 2019. 5", + "[27] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fr\\`echet inception distance. In ICLR, 2023. 5", + "[28] Ziqiang Li, Chaoyue Wang, Heliang Zheng, Jing Zhang, and Bin Li. Fakeclr: Exploring contrastive learning for solving latent discontinuity in data-efficient gans. In ECCV, 2022. 1, 2, 5, 6, 7", + "[29] Jae Hyun Lim and Jong Chul Ye. Geometric GAN. arXiv preprint arXiv:1705.02894, 2017. 4", + "[30] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and Ahmed Elgammal. Towards faster and stabilized GAN training for high-fidelity few-shot image synthesis. In ICLR, 2021. 2, 4, 5", + "[31] Haozhe Liu, Wentian Zhang, Bing Li, Haoqian Wu, Nanjun He, Yawen Huang, Yuexiang Li, Bernard Ghanem, and Yefeng Zheng. Adaptivemix: Improving GAN training via feature space shrinkage. In CVPR, 2023. 2" + ], + "bbox": [ + 514, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30690", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Steven Liu, Tongzhou Wang, David Bau, Jun-Yan Zhu, and Antonio Torralba. Diverse image generation via self-conditioned GANs. In CVPR, 2020. 2", + "[33] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In ICCV, 2017. 2", + "[34] Takeru Miyato and Masanori Koyama. cGANs with projection discriminator. In ICLR, 2018. 2", + "[35] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 2, 4", + "[36] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze the discriminator: a simple baseline for fine-tuning GANs. arXiv preprint arXiv:2002.10964, 2020. 5", + "[37] Byeonghu Na, Yeongmin Kim, Minsang Park, Donghyeok Shin, Wanmo Kang, and Il-Chul Moon. Diffusion rejection sampling. In ICML, 2024. 2", + "[38] Yao Ni and Piotr Koniusz. Nice: Noise-modulated consistency regularization for data-efficient gans. In NeurIPS, 2023. 1, 5", + "[39] Yao Ni and Piotr Koniusz. Chain: Enhancing generalization in data-efficient gans via lipschitz continuity constrained normalization. In CVPR, 2024. 1, 5", + "[40] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In ICCV, 2019. 5", + "[41] Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In ICML, 2016. 2", + "[42] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. In ICLR, 2016. 2", + "[43] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. In NeurIPS, 2016. 2", + "[44] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected GANs converge faster. In NeurIPS, 2021. 1, 2, 5, 6", + "[45] Masashi Sugiyama, Matthias Krauledat, and Klaus-Robert Müller. Covariate shift adaptation by importance weighted cross validation. Journal of Machine Learning Research, 8 (5), 2007. 4", + "[46] Akinori Tanaka. Discriminator optimal transport. In NeurIPS, 2019. 2", + "[47] Dustin Tran, Rajesh Ranganath, and David M Blei. Deep and hierarchical implicit models. arXiv preprint arXiv:1702.08896, 7(3):13, 2017. 4", + "[48] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In CVPR, 2021. 2", + "[49] Ryan Turner, Jane Hung, Eric Frank, Yunus Saatchi, and Jason Yosinski. Metropolis-hastings generative adversarial networks. In ICML, 2019. 2", + "[50] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. In NeurIPS, 2016. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Chirag Vashist, Shichong Peng, and Ke Li. Rejection sampling imle: Designing priors for better few-shot image synthesis. In ECCV, 2024. 2", + "[52] Yaxing Wang, Chenshen Wu, Luis Herranz, Joost van de Weijer, Abel Gonzalez-Garcia, and Bogdan Raducanu. Transferring GANs: generating images from limited data. In ECCV, 2018. 5", + "[53] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, Luis Herranz, Fahad Shahbaz Khan, and Joost van de Weijer. MineGAN: effective knowledge transfer from GANs to target domains with few images. In CVPR, 2020. 5", + "[54] Zhendong Wang, Yifan Jiang, Huangjie Zheng, Peihao Wang, Pengcheng He, Zhangyang Wang, Weizhu Chen, and Mingyuan Zhou. Patch diffusion: Faster and more data-efficient training of diffusion models. In NeurIPS, 2023. 5, 6", + "[55] Zhendong Wang, Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Diffusion-gan: Training gans with diffusion. In ICLR, 2023. 1, 2, 5, 6, 7", + "[56] Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Data-efficient instance generation from instance discrimination. In NeurIPS, 2021. 5, 6", + "[57] Ceyuan Yang, Yujun Shen, Yinghao Xu, Deli Zhao, Bo Dai, and Bolei Zhou. Improving GANs with a dynamic discriminator. In NeurlPS, 2022. 2", + "[58] Mengping Yang, Zhe Wang, Ziqiu Chi, and Yanbing Zhang. Fegan: Exploiting frequency components for training gans under limited data. In NeurIPS, 2022. 2, 5", + "[59] Mengping Yang, Zhe Wang, Wenyi Feng, Qian Zhang, and Ting Xiao. Improving few-shot image generation by structural discrimination and textural modulation. In ACM MM, 2023. 1", + "[60] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In ICCV, 2017. 2", + "[61] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In ICML, 2019. 2", + "[62] Zhaoyu Zhang, Mengyan Li, and Jun Yu. On the convergence and mode collapse of GAN. In SIGGRAPH Asia 2018 Technical Briefs, 2018. 2", + "[63] Zhaoyu Zhang, Yang Hua, Guanxiong Sun, Hui Wang, and Seán McLoone. Improving the training of the gans with limited data via dual adaptive noise injection. In ACMMM, 2024. 1, 2", + "[64] Zhaoyu Zhang, Yang Hua, Guanxiong Sun, Hui Wang, and Seán McLoone. Improving the leaking of augmentations in data-efficient gans via adaptive negative data augmentation. In WACV, 2024. 1, 3, 4, 5", + "[65] Zhaoyu Zhang, Yang Hua, Hui Wang, and Seán McLoone. Improving the fairness of the min-max game in GANs training. In WACV, 2024. 1", + "[66] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient GAN training. In NeurIPS, 2020. 1, 2, 5, 6" + ], + "bbox": [ + 516, + 90, + 903, + 886 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30691", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/9b5acaea-53bf-4055-87a9-1a5bf68ccd00_model.json b/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/9b5acaea-53bf-4055-87a9-1a5bf68ccd00_model.json new file mode 100644 index 0000000000000000000000000000000000000000..80083e85e67706f7d2a0f1f604129d79f6c509f0 --- /dev/null +++ b/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/9b5acaea-53bf-4055-87a9-1a5bf68ccd00_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.202, + 0.131, + 0.797, + 0.177 + ], + "angle": 0, + "content": "Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.202, + 0.796, + 0.24 + ], + "angle": 0, + "content": "Zhaoyu Zhang\\(^{1}\\), Yang Hua\\(^{1}\\), Guanxiong Sun\\(^{2}\\), Hui Wang\\(^{1}\\), Seán McLoone\\(^{1}\\) \n\\(^{1}\\)Queen's University Belfast \\(^{2}\\)Huawei UKRD" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.242, + 0.719, + 0.258 + ], + "angle": 0, + "content": "{zzhang55,Y.Hua,gsun02,h.wang,s.mcloone}@qub.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.325, + 0.486, + 0.765 + ], + "angle": 0, + "content": "Data-Efficient Generative Adversarial Nets (DE-GANs) have become more and more popular in recent years. Existing methods apply data augmentation, noise injection and pre-trained models to maximally increase the number of training samples thus improving the training of DE-GANs. However, none of these methods considers the sample quality during training, which can also significantly influence the training of DE-GANs. Focusing on sample quality during training, in this paper, we are the first to incorporate discriminator rejection sampling (DRS) into the training process and introduce a novel method, called quality aware dynamic discriminator rejection sampling (QADDRS). Specifically, QADDRS consists of two steps: (1) the sample quality aware step, which aims to obtain the sorted critic scores, i.e., the ordered discriminator outputs, on real/fake samples in the current training stage; (2) the dynamic rejection step that obtains dynamic rejection number \\( N \\), where \\( N \\) is controlled by the overfitting degree of discriminator \\( (D) \\) during training. When updating the parameters of \\( D \\), the \\( N \\) high critic score real samples and the \\( N \\) low critic score fake samples in the minibatch are rejected dynamically based on the overfitting degree of \\( D \\). As a result, QADDRS can avoid \\( D \\) becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of \\( D \\) issue during training. Extensive experiments on several datasets demonstrate that integrating QADDRS into different DE-GANs can achieve better performance and deliver state-of-the-art results. Codes are available at https://github.com/zzhang05/QADDRS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.798, + 0.222, + 0.814 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Over the past decade, Generative Adversarial Nets (GANs) [12] have demonstrated impressive performance in synthesizing images and videos [20, 21, 23, 59, 65]. However, this success heavily depends on access to large-scale datasets, which can be costly, time-consuming, and even impractical" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.29, + 0.911, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.519, + 0.908, + 0.603 + ], + "angle": 0, + "content": "Figure 1. Comparison of discriminator outputs on both real and fake samples with StyleGAN2 and StyleGAN2 + QADDRS using the 100-shot Obama dataset. Applying QADDRS to StyleGAN2 can effectively prevent \\( D \\) from being overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of \\( D \\) issue." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.674 + ], + "angle": 0, + "content": "to collect and clean in certain domains. As a result, data-efficient GANs (DE-GANs) [28, 38, 39] have garnered significant attention." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Recent studies have applied techniques such as data augmentation [8, 9, 19, 22, 66], noise injection [55, 63] and pre-trained models [25, 44] to improve the DE-GANs training. All of these methods aim to maximally increase the number of training samples under data-efficient scenarios. However, none of these methods considers the sample quality during training, which can also significantly influence the DE-GANs training. For instance, as pointed out in several existing approaches [22, 55, 64], the highly undesirable quality samples produced by data augmentation or noise injection can be detrimental to DE-GANs training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Focusing on sample quality during training, in this paper, we are the first to introduce discriminator rejection sampling (DRS) [5] into the training process and propose a novel method, termed quality aware dynamic discrimina" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30682" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "tor rejection sampling (QADDRS). Specifically, QADDRS comprises two steps: (1) the sample quality aware step that obtains the ranked critic scores, i.e., the sorted discriminator \\((D)\\) outputs, on real/fake samples to measure the quality of the real/fake samples in the current training stage and (2) the dynamic rejection step which is determined by the dynamic rejection number \\(N\\), where \\(N\\) adapts based on the overfitting degree of \\(D\\). During updating the parameters of \\(D\\) process, the \\(N\\) real samples with the high critic scores and the \\(N\\) fake samples with the low critic score in the minibatch are dynamically rejected according to the overfitting degree of \\(D\\). As illustrated in Figure 1, QADDRS can effectively prevent \\(D\\) from becoming overly confident in distinguishing real and fake samples, thereby alleviating the overfitting of \\(D\\) issue during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.483, + 0.569 + ], + "angle": 0, + "content": "To sum up, the key contributions are shown as follows: (1) We introduce DRS into the GANs training process and propose a novel method named quality aware dynamic discriminator rejection sampling (QADDRS) to improve DEGANs training. As far as we are aware, QADDRS is the first approach that considers sample quality during DE-GANs training. (2) We present the theoretical analysis of QADDRS in GANs training, proving its robustness and rationality. Furthermore, we observe that applying QADDRS to DE-GANs can avoid \\(D\\) becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of \\(D\\) issue during training. (3) Experiments on low-shot [66] and FFHQ [23] datasets with different methods demonstrate that QADDRS can improve performance across various DE-GANs and deliver state-of-the-art performance over existing GANs and diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.609, + 0.233, + 0.625 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.643, + 0.409, + 0.658 + ], + "angle": 0, + "content": "2.1. Generative Adversarial Nets (GANs)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In the past ten years, generative adversarial nets (GANs) [12] have been among the most popular generative models [41, 50]. GANs consist of two networks: a generator \\((G)\\) and a discriminator \\((D)\\), which engage in a min-max game. The \\(G\\) seeks to generate realistic-looking synthetic samples from noise input, while the \\(D\\) attempts to distinguish whether the input sample originates from the \\(G\\) or from real samples. GANs are known to suffer from training instability, often resulting in low-quality and limited diversity in generated data [62]. To overcome this challenge, different approaches have been presented, including more advanced architectures [6, 34, 35, 42, 58, 61], more stable objective functions [4, 14, 15, 33, 43], and improved training strategies [10, 20, 28, 31, 32, 57, 60] to generate more realistic and diverse samples." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.71, + 0.106 + ], + "angle": 0, + "content": "2.2. Data-Efficient GANs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.114, + 0.905, + 0.236 + ], + "angle": 0, + "content": "Recently, training Data-Efficient GANs (DE-GANs) has gained increasing significance. Studies [8, 9, 19, 22, 66] have shown that DE-GAN training often suffers from discriminator \\((D)\\) overfitting. To address this issue, different approaches have been introduced, including novel data augmentation techniques [8, 9, 19, 22, 66], improved regularization methods for \\(D\\) [11, 30, 48], noise injection [55, 63], and pre-trained models [25, 44]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.243, + 0.869, + 0.26 + ], + "angle": 0, + "content": "2.3. Rejection Sampling in Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.908, + 0.493 + ], + "angle": 0, + "content": "In recent years, rejection sampling has played a significant role in increasing the generated image quality as a posttraining technique. Rejection sampling is commonly used in several kinds of generative models, such as GANs [5], diffusion models [37] and IMLE [51]. Among these methods, the most well-known method in GANs is discriminator rejection sampling (DRS). DRS has inspired the development of numerous techniques in GANs, such as MH-GAN [49], DDLS [7], DOT [46], DGflow [2] and LatentRS [18]. Although these studies have achieved great success in rejecting lower-quality generated samples, they are all posttraining methods and cannot directly benefit the training of DE-GANs. In this study, we are the first to incorporate the DRS in the training stage to improve the DE-GANs training by considering sample quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.506, + 0.649, + 0.523 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.905, + 0.562 + ], + "angle": 0, + "content": "3.1. Quality Aware Dynamic Discriminator Rejection Sampling (QADDRS)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.674 + ], + "angle": 0, + "content": "As discussed in §2.3, existing discriminator rejection sampling (DRS) methods are rejecting low-quality generated samples in the post-training process to improve performance. However, it is infeasible to directly apply DRS to the training process, as it can cause the number of real and generated samples for the input of \\(D\\) to become unbalanced during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel approach called quality aware dynamic discriminator rejection sampling (QADDRS) to improve DE-GANs training by considering sample quality. The overview of applying QADDRS to update both \\( G \\) and \\( D \\) is shown in Figure 2. Specifically, when updating \\( G \\), all of the fake samples are used to update the parameters of \\( G \\), following the prior approaches [22, 66]. In contrast, when updating \\( D \\), the \\( N \\) high critic score real samples and the \\( N \\) low critic score fake samples in the minibatch are dynamically rejected, and the rest of the real and fake samples are used to update the parameters of \\( D \\). Consequently, QADDRS can prevent \\( D \\) from becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of \\( D \\) issue during training. The selection of high critic score real samples and low critic score" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30683" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.885, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.381, + 0.908, + 0.455 + ], + "angle": 0, + "content": "Figure 2. An overview of quality aware dynamic discriminator rejection sampling (QADDRS) for updating \\( G \\) (left) and updating \\( D \\) (right) in DE-GANs training. When updating \\( G \\), all of the generated samples are used to update the parameters of \\( G \\). When updating \\( D \\), the \\( N \\) high critic score real samples and \\( N \\) low critic score fake samples in the minibatch are dynamically rejected, and only the rest of the real and fake samples are used to update the parameters of \\( D \\). The selection of high critic score real samples and low critic score fake samples is guided by the discriminator outputs on both real and fake data during training. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.491, + 0.482, + 0.536 + ], + "angle": 0, + "content": "fake samples is guided by the sample quality-aware step, while the rejection of \\( N \\) real and fake samples is controlled by the dynamic rejection step." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.537, + 0.483, + 0.613 + ], + "angle": 0, + "content": "For the sample quality aware step, the quality of real/fake samples can be represented by different critic scores on real/fake samples. Based on this, we sort real/fake samples via critic scores, which are based on the discriminator outputs on real/fake data in the current training stage." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.483, + 0.794 + ], + "angle": 0, + "content": "For the dynamic rejection step, the dynamic rejection number \\(N\\) is introduced to control sample rejection. It is obvious that rejecting real samples in QADDRS is highly undesirable, especially under data-efficient scenarios, which could cause a further reduction of training samples. To address this, we introduce a dynamic mechanism for the QADDRS by setting the rejection number \\(N\\) to be adjusted dynamically according to the degree of overfitting without manual tuning. To achieve this, following ANDA [64], APA [19] and ADA [22], we first utilize an overfitting heuristic \\(\\eta\\) which can quantify the \\(D\\)'s overfitting degree as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.814, + 0.483, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\eta = \\mathbb {E} \\left(\\operatorname {s i g n} \\left(D _ {\\text {r e a l}}\\right)\\right), D _ {\\text {r e a l}} = \\operatorname {l o g i t} (D (x)), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathrm{sign}(\\cdot)\\) denotes the sign function, which returns \\(+1\\) for a non-negative input; \\(-1\\), otherwise. Next, we follow ADA, APA and ANDA by using \\(\\eta\\) to obtain the overfitting probability \\(p\\). The larger value of \\(p\\) means the greater over" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.488, + 0.912, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.688, + 0.908, + 0.745 + ], + "angle": 0, + "content": "Figure 3. The curve of the average number of real samples rejected by QADDRS \\((\\lambda = 0.25)\\) when updating \\(D\\) for every 4K fake samples shown to \\(D\\) during training. The experiments are conducted on the 100-shot Obama dataset using StyleGAN2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.908, + 0.817 + ], + "angle": 0, + "content": "fitting degree. Next, we set the initial rejection number \\( N \\) as 0, and adjust its value based on the overfitting probability \\( p \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.831, + 0.905, + 0.864 + ], + "angle": 0, + "content": "\\[\nN = \\left\\{ \\begin{array}{l l} N + 1 & (p) \\\\ N - 1 & (1 - p) \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "where \\(N\\) is adjusted to \\(N + 1\\) under the probability \\(p\\) and is adjusted to \\(N - 1\\) under the probability \\(1 - p\\). We introduce" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "30684" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.166 + ], + "angle": 0, + "content": "a hyperparameters \\(\\lambda\\) and set the range of \\(N\\) as \\(N_{\\min} = 0\\) and \\(N_{\\max} = \\lambda \\times\\) batchsize. To further demonstrate how the number \\(N\\) evolves during the training process, the value of \\(N\\) in every 4K real images shown to \\(D\\) during training can be found in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.484, + 0.199 + ], + "angle": 0, + "content": "In summary, employing QADDRS in DE-GANs can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.217, + 0.486, + 0.234 + ], + "angle": 0, + "content": "\\[\nV _ {D} (G, D) = \\mathbb {E} _ {x \\sim P _ {R}} [ \\tilde {g} _ {1} (D (x _ {m - N})) ] + \\mathbb {E} _ {x \\sim P _ {G}} [ \\tilde {g} _ {2} (D (x _ {m - N})) ],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.236, + 0.484, + 0.267 + ], + "angle": 0, + "content": "\\[\nV _ {G} (G, D) = - \\mathbb {E} _ {x \\sim P _ {G}} [ \\tilde {f} ((D (x _ {m}))) ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.484, + 0.372 + ], + "angle": 0, + "content": "where \\( x_{m} \\) is a batch of samples, \\( m \\) denotes the total samples number in a batch, and \\( \\tilde{g}_{1} \\), \\( \\tilde{g}_{2} \\) and \\( \\tilde{f} \\) are scalar-to-scalar functions. For the StyleGAN2 [23] with non-saturating loss, \\( \\tilde{g}_{1}(x) = \\log (x) \\), \\( \\tilde{g}_{2}(x) = \\log (1 - x) \\) and \\( \\tilde{f}(x) = \\log (x) \\). For the FastGAN [30] with hinge loss, \\( \\tilde{g}_{1}(x) = \\min (0, -1 + x) \\), \\( \\tilde{g}_{2}(x) = \\min (0, -1 - x) \\) and \\( \\tilde{f}(x) = x \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.38, + 0.285, + 0.396 + ], + "angle": 0, + "content": "3.2. Theoretical Analysis" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.403, + 0.34, + 0.42 + ], + "angle": 0, + "content": "3.2.1 QADDRS in GANs training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.471 + ], + "angle": 0, + "content": "To further demonstrate how QADDRS can influence GANs training in theory, we first revisit the hinge loss in GANs shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.483, + 0.482, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} V _ {D} (D, G) = \\mathbb {E} _ {x \\sim P _ {R}} [ \\min (0, - 1 + D (x)) ] \\\\ + \\mathbb {E} _ {x \\sim P _ {G}} [ \\min (0, - 1 - D (x)) ], \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.519, + 0.345, + 0.535 + ], + "angle": 0, + "content": "\\[\nV _ {G} (D, G) = - \\mathbb {E} _ {x \\sim P _ {G}} [ D (x) ].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.704 + ], + "angle": 0, + "content": "For Eq.(4), \\( D(x) > 1 \\) on the real data or \\( D(G(z)) < -1 \\) on the fake data can lead to the loss value becoming 0, resulting in no gradient for updating \\( D \\). Therefore, the key essence of hinge loss in GANs training is to reject the very high critic score real samples (score \\( > 1 \\)) and very low critic score fake samples (score \\( < -1 \\)). Following this, our proposed QADDRS further rejects the real and fake samples based on the overfitting degree of \\( D \\) to enhance the DE-GANs training. We then consider how QADDRS influences the distributions of real data \\( (P_R) \\) and fake data \\( (P_G) \\) in GANs training as follows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "According to Figure 2, QADDRS can be regarded as involving two actions, i.e., rejecting low critic score fake samples and rejecting high critic score real samples. Because rejecting real/fake samples in QADDRS is performed dynamically, the key operations of QADDRS on the distributions are re-weighting both real and fake samples for their data distributions, respectively. Consequently, in rejecting the low critic score fake samples action, applying QADDRS can transform the original fake data distribution \\((P_{G})\\) to the re-weighted fake data distribution \\((w_{1}P_{G})\\). Similarly, in rejecting the high critic score real samples action, utilizing QADDRS can transform the original real data distribution \\((P_{R})\\) to the re-weighted real data distribution \\((w_{2}P_{R})\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "3.2.2 Convergence analysis of QADDRS in StyleGAN2 and FastGAN" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.13, + 0.907, + 0.176 + ], + "angle": 0, + "content": "Based on the analysis in §3.2.1, we then analyze the QADDRS on both StyleGAN2 with non-saturating loss and FastGAN with hinge loss as follows." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.178, + 0.907, + 0.283 + ], + "angle": 0, + "content": "1. For the StyleGAN2 with non-saturating loss, QADDRS can effectively introduce the advantages of hinge loss into the non-saturating loss via dynamically rejecting the \\(N\\) high critic score real samples and the \\(N\\) low critic score fake samples. Following §3.2.1 and [3], the optimal discriminator of QADDRS with non-saturating loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.294, + 0.906, + 0.328 + ], + "angle": 0, + "content": "\\[\nD ^ {*} (x) = \\frac {w _ {2} (x) P _ {R} (x)}{w _ {2} (x) P _ {R} (x) + w _ {1} (x) P _ {G} (x)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.332, + 0.907, + 0.378 + ], + "angle": 0, + "content": "With \\(D^{*}\\) as the optimal discriminator, according to Theorem 2.5 in [3], training the \\(G\\) with non-saturating loss can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.39, + 0.905, + 0.418 + ], + "angle": 0, + "content": "\\[\nV _ {G} (G, D ^ {*}) = \\mathbf {K L} \\left(w _ {1} P _ {G} \\mid \\mid w _ {2} P _ {R}\\right) - 2 \\mathbf {J S} \\left(w _ {1} P _ {G} \\mid \\mid w _ {2} P _ {R}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.419, + 0.906, + 0.614 + ], + "angle": 0, + "content": "where KL and JS denote the Kullback-Leibler divergence and the Jensen-Shannon divergence, respectively. Eq.(6) shows that applying QADDRS to the non-saturating loss can lead to the minimization of KL-2JS divergence between the \\( w_{1}P_{G} \\) and \\( w_{2}P_{R} \\). Then, according to [3], this minimization process depends on the supports of \\( w_{1}P_{G} \\) and \\( w_{2}P_{R} \\). Previous studies [6, 13, 21, 45] have established that re-weighting operations applied to distributions do not alter their support. Therefore, the minimization of Eq.(6) corresponds to the minimization of KL-2JS divergence between the \\( P_{G} \\) and \\( P_{R} \\), which demonstrates that applying QADDRS to non-saturating loss does not influence convergence." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.615, + 0.907, + 0.766 + ], + "angle": 0, + "content": "2. For the FastGAN with hinge loss, QADDRS can further take advantage by dynamically rejecting the \\(N\\) high critic score real samples and the \\(N\\) low critic score fake samples if all \\(D(x)\\) and \\(D(G(z))\\) are in \\([-1, 1]\\). According to the §3.2.1 and [29, 35, 47], optimizing the hinge loss with QADDRS is equivalent to minimizing the reverse KL divergence term \\(\\mathbf{KL}(w_1P_G\\parallel w_2P_R)\\). With the fact that the supports of \\(w_1P_G\\) and \\(w_2P_R\\) are the same as the supports of \\(P_G\\) and \\(P_R\\), applying QADDRS to hinge loss does not influence the convergence." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.787, + 0.81, + 0.802 + ], + "angle": 0, + "content": "3.2.3 Analysis of QADDRS in DE-GANs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "According to the theory presented in [64], data augmentation (DA) is commonly used in DE-GANs, which transforms the real and fake data distributions as \\( P_R^T \\) and \\( P_G^T \\), respectively. As a result, applying QADDRS can result in the optimization between the \\( w_1P_G^T \\) and \\( w_2P_R^T \\) for both StyleGAN2 with non-saturating loss and FastGAN with hinge" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30685" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.896, + 0.491 + ], + "angle": 0, + "content": "
MethodBackboneMAPre-training100-shotAnimal-Face
ObamaGrumpy CatPandaCatDog
EDM [24]DiffusionYesNo51.3036.9023.7048.60100.10
EDM + DA [24]DiffusionYesNo37.1029.9410.8136.8857.14
Patch Diffusion [54]DiffusionYesNo41.4730.8913.2543.7172.17
Scale/shift [40]GANNoYes50.7234.2021.3854.8383.04
MineGAN [53]GANNoYes50.6334.5414.8454.4593.03
TransferGAN [52]GANNoYes48.7334.0623.2052.6182.38
TransferGAN + DA [66]GANYesYes39.8529.7717.1249.1065.57
FreezeD [36]GANNoYes41.8731.2217.9547.7070.46
StyleGAN2 [23]GANNoNo80.2048.9034.2771.71131.90
FakeCLR [28]GANYesNo26.9519.568.4226.3442.02
FastGAN [30]GANYesNo35.8025.759.7033.8552.46
FreGAN [58]GANYesNo33.3924.938.9731.0547.85
AugSelf-StyleGAN2 [17]GANYesNo26.0019.818.3630.5348.19
CHAIN [39]GANYesNo20.9417.576.9319.5830.88
NICE [38]GANYesNo20.0815.638.1822.7028.65
Projected GAN [44]GANYesYes11.2115.803.9818.0117.88
StyleGAN2* [23]GANYesNo65.5739.9225.0851.6687.96
+ QADDRSGANYesNo53.5435.7720.5947.8773.29
StyleGAN2 + Diff-Augment [66]GANYesNo46.8727.0812.0642.4458.85
+ QADDRSGANYesNo41.2726.5310.8338.1551.83
StyleGAN2 + ADA [22]GANYesNo45.6926.6212.9040.7756.83
+ QADDRSGANYesNo40.8326.2711.9936.2553.01
InsGen + ADA [56]GANYesNo32.4222.019.8533.0144.93
+ QADDRSGANYesNo25.4119.017.9923.6935.01
Diffusion-Projected GAN (FastGAN) [55]GANYesYes10.5415.133.3917.8617.22
+ QADDRSGANYesYes9.9314.723.0816.9416.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.515, + 0.904, + 0.556 + ], + "angle": 0, + "content": "Table 1. FID [16] scores (lower is better) of different methods on \\( {256} \\times {256} \\) low-shot datasets. MA donates Massive Augmentation [9], i.e., xflipping. For a fair comparison, FID is computed using \\( 5\\mathrm{\\;K} \\) generated samples,with the entire dataset serving as the reference distribution. The results are averaged over three evaluations, with relative standard deviations below 1%." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.69 + ], + "angle": 0, + "content": "loss. As the supports of \\( w_{1}P_{G}^{T} \\) and \\( w_{2}P_{R}^{T} \\) are the same as the supports of \\( P_{R}^{T} \\) and \\( P_{G}^{T} \\), utilizing QADDRS in the DE-GANs does not influence convergence. Furthermore, DE-GANs are known to suffer from the leaking of augmentations issue [22, 64]. In this case, QADDRS can also reweight the out-of-distribution samples for both real and fake data to alleviate the leaking issue." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.702, + 0.225, + 0.719 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.411, + 0.743 + ], + "angle": 0, + "content": "4.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We select FFHQ [23] and low-shot [66] datasets for the experiments. The FFHQ dataset includes 70K high-resolution face images. In line with FakeCLR [28], we preprocess and resize the FFHQ dataset to \\(256 \\times 256\\) and select different subsets consisting of 100, 1K, 2K, and 5K images for all experiments. Low-shot datasets \\((256 \\times 256)\\) consist of five sets (i.e., Obama, Grumpy Cat, Panda, Animal-Face Cat, and Animal-Face Dog) with 100, 100, 100, 160, and 389 training samples, respectively. The commonly used Fréchet Inception Distance (FID) [16], CLIP FID (FIDCLIP) [27]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.905, + 0.642 + ], + "angle": 0, + "content": "and Precision and Recall [26] are applied as the evaluation metrics. Following prior methods [22, 66], the complete dataset serves as the reference distribution for all metric evaluations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.84 + ], + "angle": 0, + "content": "The implementation details are shown as follows. For the experiments on low-shot datasets with StyleGAN2, StyleGAN2 + Diff-Augment and StyleGAN2 + ADA, we set the batchsize as 4 and the overall training duration as 2000K images. For the experiments on low-shot datasets with InsGen + ADA, we set the batchsize as 64 and the overall training duration as 35000K images. For the experiments on low-shot datasets with Diffusion-Projected GAN, we set the batchsize as 64 and the overall training duration as 100000K images. For the experiments on FFHQ datasets, we set the batchsize as 64 and the overall training duration as 25000K images. The hyperparameter \\(\\lambda\\) is set as 0.25 in QADDRS across all experiments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.776, + 0.862 + ], + "angle": 0, + "content": "4.2. Results on Low-shot Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.9 + ], + "angle": 0, + "content": "The FID results of different methods on the \\(256 \\times 256\\) low-shot datasets are shown in Table 1. We show the effec" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30686" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.091, + 0.242, + 0.102 + ], + "angle": 0, + "content": "FID 65.57" + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.105, + 0.276, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.092, + 0.385, + 0.102 + ], + "angle": 0, + "content": "FID 25.08" + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.106, + 0.419, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.092, + 0.528, + 0.102 + ], + "angle": 0, + "content": "FID 39.92" + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.106, + 0.562, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.092, + 0.672, + 0.102 + ], + "angle": 0, + "content": "FID 51.66" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.106, + 0.705, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.092, + 0.813, + 0.102 + ], + "angle": 0, + "content": "FID 87.96" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.106, + 0.846, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.213, + 0.263, + 0.224 + ], + "angle": 0, + "content": "FID 53.54 (-12.03)" + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.226, + 0.276, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.214, + 0.403, + 0.225 + ], + "angle": 0, + "content": "FID 20.59 (-4.49)" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.226, + 0.418, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.214, + 0.547, + 0.225 + ], + "angle": 0, + "content": "FID 35.77 (-4.15)" + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.226, + 0.563, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.214, + 0.69, + 0.225 + ], + "angle": 0, + "content": "FID 47.87 (-3.79)" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.226, + 0.704, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.214, + 0.835, + 0.225 + ], + "angle": 0, + "content": "FID 73.29 (-14.67)" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.226, + 0.848, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.332, + 0.907, + 0.375 + ], + "angle": 0, + "content": "Figure 4. The compared generated results between StyleGAN2 and StyleGAN2 + QADDRS on low-shot datasets. Top: Images generated by StyleGAN2. Bottom: Images generated by StyleGAN2 + QADDRS. The reduction of FID in red color shows the improvement of StyleGAN2 + QADDRS over the baseline StyleGAN2." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.4, + 0.903, + 0.668 + ], + "angle": 0, + "content": "
MethodBackboneMAPre-trainingFFHQ
1001K2K5K
EDM [24]DiffusionYesNo79.10---
EDM + DA [24]DiffusionYesNo50.7330.7527.1724.51
Patch Diffusion [54]DiffusionYesNo44.4528.0325.3222.63
StyleGAN2 [23]GANNoNo179.00100.1654.3049.68
StyleGAN2 + Diff-Augment [66]GANYesNo61.9125.6624.3210.45
StyleGAN2 + ADA [22]GANYesNo85.821.2915.3910.96
APA [19]GANYesNo65.0018.8916.908.38
AugSelf-StyleGAN2+ [17]GANYesNo-20.39-9.15
FakeCLR [28]GANYesNo42.5615.929.907.25
Projected GAN [44]GANYesYes26.2511.128.256.85
InsGen [56]GANYesNo45.7518.2111.477.83
+ QADDRSGANYesNo43.6717.1610.737.51
Diffusion-Projected GAN (FastGAN) [55]GANYesYes25.478.767.996.59
+ QADDRSGANYesYes24.468.517.886.48
Diffusion-Projected GAN (FastGAN Lite) [55]GANYesYes20.657.025.474.74
+ QADDRSGANYesYes19.896.575.184.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.688, + 0.907, + 0.732 + ], + "angle": 0, + "content": "Table 2. FID scores (lower is better) of different methods on the \\( {256} \\times {256}\\mathrm{{FFHQ}} \\) dataset. In line with FakeCLR [28], the experiments are performed on 100, 1K, 2K and 5K subsets of the FFHQ dataset. MA donates Massive Augmentation [9], i.e., xflipping. The results are averaged over five evaluations, with relative standard deviations below 1%." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.894 + ], + "angle": 0, + "content": "tiveness of the QADDRS with both StyleGAN2 and FastGAN using five different commonly-used DE-GANs, i.e., StyleGAN2, StyleGAN2 + Diff-Augment, StyleGAN2 + ADA, InsGen + ADA and Diffusion-Projected GAN. As shown in Table 1, the proposed QADDRS can increase the performance of these five different DE-GANs, demonstrating that QADDRS can be regarded as a general plugin training strategy for DE-GANs. As shown in Table 7, Diffusion-Projected GAN (FastGAN) achieves better per" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.757, + 0.907, + 0.895 + ], + "angle": 0, + "content": "formance than Diffusion-Projected GAN (FastGAN Lite), i.e., applying the lite version generator in FastGAN, on low-shot datasets. Therefore, we report the results of the Diffusion-Projected GAN (FastGAN) in Table 1. Although Diffusion-Projected GAN (FastGAN) has already utilized all widely-used approaches, i.e., DA, regularizations, noise injection and pre-trained models, to boost performance on the low-shot datasets, adding QADDRS can further improve its performance. In addition, following [1], we also report" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30687" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.885, + 0.161 + ], + "angle": 0, + "content": "
MethodObamaGrumpy CatPandaAF-CatAF-Dog
PRPRPRPRPR
Diffusion-Projected GAN (FastGAN) + QADDRS0.9990.2400.9960.1490.9990.4100.9970.1820.9980.203
0.9990.2420.9980.1510.9990.4180.9980.1860.9990.203
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.17, + 0.908, + 0.212 + ], + "angle": 0, + "content": "Table 3. The comparison of Precision (P) and Recall (R) (higher is better) with Diffusion-Projected GAN (FastGAN) [55] and Diffusion-Projected GAN (FastGAN) + QADDRS on the \\(256 \\times 256\\) low-shot datasets. The results are averaged over three evaluations, with relative standard deviations below \\(1\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.238, + 0.872, + 0.306 + ], + "angle": 0, + "content": "
MethodFFHQ-100FFHQ-1KFFHQ-2KFFHQ-5K
PRPRPRPR
Diffusion-Projected GAN (FastGAN Lite) + QADDRS0.7370.0180.7340.2690.7360.3210.7300.377
0.7390.0210.7340.2760.7390.3230.7330.379
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.315, + 0.908, + 0.357 + ], + "angle": 0, + "content": "Table 4. The comparison of Precision (P) and Recall (R) (higher is better) with Diffusion-Projected GAN (FastGAN Lite) [55] and Diffusion-Projected GAN (FastGAN Lite) + QADDRS on the \\(256 \\times 256\\) FFHQ dataset. The results are averaged over five evaluations, with relative standard deviations below \\(1\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.395, + 0.49, + 0.461 + ], + "angle": 0, + "content": "
Method100AF
OGCPCD
Diffusion Projected GAN (FastGAN)2.024.782.598.367.21
+ QADDRS1.984.712.538.207.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.555 + ], + "angle": 0, + "content": "Table 5. Compared \\( \\mathrm{FID}_{\\mathrm{CLIP}} \\) scores between Diffusion Projected GAN (FastGAN) and Diffusion Projected GAN (FastGAN) + QADDRS on low-shot datasets. We use the FID calculation model to calculate the \\( \\mathrm{FID}_{\\mathrm{CLIP}} \\). Here, AF denotes AnimalFace, O denotes Obama, GC denotes Grumpy Cat, P denotes Panda, C denotes Cat and D denotes Dog." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.591, + 0.48, + 0.664 + ], + "angle": 0, + "content": "
MethodFFHQ
1001K2K5K
Diffusion Projected GAN (FL)23.9111.459.618.33
+ QADDRS23.3211.149.428.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.673, + 0.484, + 0.73 + ], + "angle": 0, + "content": "Table 6. Compared \\(\\mathrm{FID}_{\\mathrm{CLIP}}\\) scores between Diffusion Projected GAN (FastGAN Lite) and Diffusion Projected GAN (FastGAN Lite) + QADDRS on FFHQ datasets. We use the FID calculation model to calculate the \\(\\mathrm{FID}_{\\mathrm{CLIP}}\\). FL denotes the FastGAN Lite." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.903 + ], + "angle": 0, + "content": "the results of several SOTA diffusion models in Table 1. It is clear that GAN-based methods outperform diffusion-based methods on low-shot datasets. To further show the effectiveness of QADDRS, the compared Precision and Recall, and \\(\\mathrm{FID}_{\\mathrm{CLIP}}\\) of adding QADDRS to Diffusion-Projected GAN (FastGAN) are shown in Tables 3 and 5, respectively. The compared generated images on StyleGAN2 and StyleGAN2 + QADDRS are shown in Figure 4." + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.393, + 0.885, + 0.449 + ], + "angle": 0, + "content": "
MethodFID
Diffusion Projected GAN (FastGAN)17.22
Diffusion Projected GAN (FastGAN Lite)20.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.47, + 0.907, + 0.527 + ], + "angle": 0, + "content": "Table 7. Compared FID scores of Diffusion Projected GAN (FastGAN) and Diffusion Projected GAN (FastGAN Lite) on \\(256\\times 256\\) AnimalFace-dog dataset. The results are averaged over three evaluations, with relative standard deviations below \\(1\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.558, + 0.747, + 0.574 + ], + "angle": 0, + "content": "4.3. Results on FFHQ Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.903 + ], + "angle": 0, + "content": "According to the results of five different DE-GANs on low-shot datasets, InsGen + ADA and Diffusion-Projected GAN achieve SOTA performance on StyleGAN2 and FastGAN, respectively. Therefore, we conduct experiments on the FFHQ dataset by selecting these two types of DEGANs. Following FakeCLR [28], we conduct experiments on the FFHQ 100, FFHQ 1K, FFHQ 2K and FFHQ 5K subsets. As shown in Table 2, QADDRS can improve performance with these two DE-GANs on the FFHQ dataset. More importantly, we report the results of both the Diffusion-Projected GAN (FastGAN) and the Diffusion-Projected GAN (FastGAN Lite) in Table 2. It is clear that Diffusion-Projected GAN (FastGAN Lite) achieves better performance than Diffusion-Projected GAN (FastGAN) on the FFHQ dataset. Additionally, the comparison of Precision and Recall, as well as \\(\\mathrm{FID}_{\\mathrm{CLIP}}\\) with Diffusion-Projected GAN (FastGAN Lite) and Diffusion-Projected GAN (FastGAN Lite) + QADDRS are shown in Tables 4 and 6, respectively. Diffusion-Projected GAN (FastGAN Lite) + QADDRS outperforms Diffusion-Projected GAN (FastGAN Lite) by adding QADDRS, which demonstrates the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30688" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.468, + 0.122 + ], + "angle": 0, + "content": "
λ0.000.250.500.75
FID46.8741.2742.4043.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.146, + 0.483, + 0.203 + ], + "angle": 0, + "content": "Table 8. FID scores on the \\( {256} \\times {256100} \\) -shot Obama dataset with StyleGAN2 + Diff-Augment. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below 1%." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.216, + 0.468, + 0.251 + ], + "angle": 0, + "content": "
λ0.000.250.500.75
FID58.8551.8353.3154.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.33 + ], + "angle": 0, + "content": "Table 9. FID scores on the \\( {256} \\times {256} \\) AnimalFace-Dog dataset with StyleGAN2 + Diff-Augment. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below 1%." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.344, + 0.455, + 0.41 + ], + "angle": 0, + "content": "
MethodObamaAF-Dog
StyleGAN2 + Diff-Augment46.8758.85
+ Fixed rejecting number N44.1254.85
+ QADDRS41.2751.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.432, + 0.483, + 0.488 + ], + "angle": 0, + "content": "Table 10. FID scores on the \\(256 \\times 256\\) 100-shot Obama and AnimalFace-Dog datasets. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below \\(1\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.483, + 0.562 + ], + "angle": 0, + "content": "effectiveness of the proposed QADDRS. To further show the superiority of QADDRS, more generated images on the FFHQ dataset can be found in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.571, + 0.255, + 0.586 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.7 + ], + "angle": 0, + "content": "Effectiveness of the hyperparameter \\(\\lambda\\) in proposed QADDRS. To demonstrate the effectiveness of the hyperparameter \\(\\lambda\\) in the proposed QADDRS, we conduct experiments by selecting different values of \\(\\lambda\\) on the 100-shot Obama and AnimalFace-Dog datasets using StyleGAN2 + Diff-Augment. The results, shown in Tables 8 and 9, indicate that setting \\(\\lambda\\) to 0.25 achieves better performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Effectiveness of the dynamic rejection number \\(N\\) in the proposed QADDRS. To show that the dynamic rejection number \\(N\\) in QADDRS is effective, we conduct an ablation study by comparing QADDRS with a fixed rejection number \\(N = 0.25 \\times\\) batchsize on the 100-shot Obama dataset with StyleGAN2 + Diff-Augment. The results are shown in Table 10. It is obvious that the dynamically determined rejection number \\(N\\) can achieve better performance compared with the fixed rejection number." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.403, + 0.863 + ], + "angle": 0, + "content": "4.5. Comparisons of DRS and QADDRS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We conduct the experiments by comparing QADDRS and the existing DRS method [5] using StyleGAN2 + Diff-" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.088, + 0.9, + 0.15 + ], + "angle": 0, + "content": "
MethodObamaAnimalFace-Dog
StyleGAN2 + Diff-Augment46.8758.85
+ DRS [5]45.2256.17
+ QADDRS41.2751.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.173, + 0.907, + 0.229 + ], + "angle": 0, + "content": "Table 11. FID scores on the \\( {256} \\times {256}\\;{100} \\) -shot Obama and Animal-Face Dog datasets. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below 1%." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.243, + 0.916, + 0.291 + ], + "angle": 0, + "content": "
MethodSeconds / 1K images
Diffusion-Projected GAN (FastGAN) + QADDRS17\n17.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.314, + 0.907, + 0.383 + ], + "angle": 0, + "content": "Table 12. The training time on the \\( {256} \\times {256}\\;{100} \\) -shot Obama dataset with or without QADDRS using Diffusion-Projected GAN (FastGAN). The results are averaged over ten evaluations on two NVIDIA RTX A5000 GPUs with a batch size of 64. All standard deviations are below 1% relative to the mean." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.907, + 0.62 + ], + "angle": 0, + "content": "Augment on the 100-shot Obama and AnimalFace-dog datasets. The results are shown in Table 11. Based on ADA [22], the overfitting of \\( D \\) during training finally results in the DE-GANs generating low-quality images under limited data settings. Although the existing DRS method can reject lower-quality generated samples during post-training, they can not benefit the training of DE-GANs. Thus, existing DRS methods can not prevent generating low-quality images in average, leaving limited gains under limited data settings. In contrast, QADDRS can alleviate the overfitting issue during training, thereby effectively avoiding producing low-quality images in average and achieving significant improvements. As a result, QADDRS outperforms the existing DRS method for DE-GANs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.629, + 0.706, + 0.644 + ], + "angle": 0, + "content": "4.6. Computational Cost" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.906, + 0.727 + ], + "angle": 0, + "content": "The results of the training time on the 100-shot Obama dataset with or without QADDRS using Diffusion-Projected GAN (FastGAN) have been shown in Table 12. The proposed QADDRS results in a negligible increase in computational cost." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.634, + 0.755 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel method called quality aware dynamic discriminator rejection sampling (QADDRS), to improve DE-GANs training by considering sample quality. QADDRS can effectively avoid \\( D \\) becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of \\( D \\) issue. Extensive experiments on multiple datasets show that applying QADDRS to various DE-GANs can achieve better performance and deliver state-of-the-art results." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30689" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.205 + ], + "angle": 0, + "content": "This work is partially supported by the High Performance Computing center of Queen's University Belfast and the Kelvin-2 supercomputer (EPSRC grant EP/T022175/1). First author is supported by a scholarship of the Queen's University Belfast and the Department for the Economy, Northern Ireland." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.218, + 0.188, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.243, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[1] Mehran Aghabozorgi, Shichong Peng, and Ke Li. Adaptive imle for few-shot pretraining-free generative modelling. In ICML, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.286, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[2] Abdul Fatir Ansari, Ming Liang Ang, and Harold Soh. Refining deep generative models via discriminator gradient flow. arXiv preprint arXiv:2012.00780, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.328, + 0.483, + 0.368 + ], + "angle": 0, + "content": "[3] Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks. In ICLR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.482, + 0.409 + ], + "angle": 0, + "content": "[4] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In ICML, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.412, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[5] Samaneh Azadi, Catherine Olsson, Trevor Darrell, Ian Goodfellow, and Augustus Odena. Discriminator rejection sampling. In ICLR, 2019. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.483, + 0.494 + ], + "angle": 0, + "content": "[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.483, + 0.551 + ], + "angle": 0, + "content": "[7] Tong Che, Ruixiang Zhang, Jascha Sohl-Dickstein, Hugo Larochelle, Liam Paull, Yuan Cao, and Yoshua Bengio. Your gan is secretly an energy-based model and you should use discriminator driven latent sampling. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.552, + 0.483, + 0.606 + ], + "angle": 0, + "content": "[8] Tianlong Chen, Yu Cheng, Zhe Gan, Jingjing Liu, and Zhangyang Wang. Data-efficient GAN training beyond (just) augmentations: A lottery ticket perspective. In NeurIPS, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[9] Kaiwen Cui, Jiaxing Huang, Zhipeng Luo, Gongjie Zhang, Fangneng Zhan, and Shijian Lu. Genco: Generative constraining for generative adversarial networks with limited data. In AAAI, 2022. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.664, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[10] Emily L Denton, Soumith Chintala, Rob Fergus, et al. Deep generative image models using a laplacian pyramid of adversarial networks. In NeurIPS, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.706, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[11] Tiantian Fang, Ruoyu Sun, and Alex Schwing. Diggan: Discriminator gradient gap regularization for GAN training with limited data. In NeurlPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[13] Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Scholkopf, and Alexander Smola. A kernel two-sample test. The Journal of Machine Learning Research, 13(1):723-773, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[14] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein GANs. In NeurIPS, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.243, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[15] Tianyu Guo, Chang Xu, Jiajun Huang, Yunhe Wang, Boxin Shi, Chao Xu, and Dacheng Tao. On positive-unlabeled classification in GAN. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[17] Liang Hou, Qi Cao, Yige Yuan, Songtao Zhao, Chongyang Ma, Siyuan Pan, Pengfei Wan, Zhongyuan Wang, Huawei Shen, and Xueqi Cheng. Augmentation-aware self-supervision for data-efficient gan training. In NeurIPS, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[18] Thibaut Issenhuth, Ugo Tanielian, David Picard, and Jeremie Mary. Latent reweighting, an almost free improvement for gans. In WACV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[19] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Deceive d: Adaptive pseudo augmentation for GAN training with limited data. In NeurIPS, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[22] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, 2020. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of styleGAN. In CVPR, 2020. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[24] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[25] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for GAN training. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In NeurIPS, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[27] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fr\\`echet inception distance. In ICLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[28] Ziqiang Li, Chaoyue Wang, Heliang Zheng, Jing Zhang, and Bin Li. Fakeclr: Exploring contrastive learning for solving latent discontinuity in data-efficient gans. In ECCV, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.761, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[29] Jae Hyun Lim and Jong Chul Ye. Geometric GAN. arXiv preprint arXiv:1705.02894, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[30] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and Ahmed Elgammal. Towards faster and stabilized GAN training for high-fidelity few-shot image synthesis. In ICLR, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[31] Haozhe Liu, Wentian Zhang, Bing Li, Haoqian Wu, Nanjun He, Yawen Huang, Yuexiang Li, Bernard Ghanem, and Yefeng Zheng. Adaptivemix: Improving GAN training via feature space shrinkage. In CVPR, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30690" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[32] Steven Liu, Tongzhou Wang, David Bau, Jun-Yan Zhu, and Antonio Torralba. Diverse image generation via self-conditioned GANs. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[33] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.179, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[34] Takeru Miyato and Masanori Koyama. cGANs with projection discriminator. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.482, + 0.25 + ], + "angle": 0, + "content": "[35] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.252, + 0.483, + 0.294 + ], + "angle": 0, + "content": "[36] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze the discriminator: a simple baseline for fine-tuning GANs. arXiv preprint arXiv:2002.10964, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.482, + 0.338 + ], + "angle": 0, + "content": "[37] Byeonghu Na, Yeongmin Kim, Minsang Park, Donghyeok Shin, Wanmo Kang, and Il-Chul Moon. Diffusion rejection sampling. In ICML, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.34, + 0.482, + 0.38 + ], + "angle": 0, + "content": "[38] Yao Ni and Piotr Koniusz. Nice: Noise-modulated consistency regularization for data-efficient gans. In NeurIPS, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.382, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[39] Yao Ni and Piotr Koniusz. Chain: Enhancing generalization in data-efficient gans via lipschitz continuity constrained normalization. In CVPR, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[40] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In ICCV, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[41] Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In ICML, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[42] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. In ICLR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.556, + 0.482, + 0.598 + ], + "angle": 0, + "content": "[43] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. In NeurIPS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.6, + 0.482, + 0.64 + ], + "angle": 0, + "content": "[44] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected GANs converge faster. In NeurIPS, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.482, + 0.698 + ], + "angle": 0, + "content": "[45] Masashi Sugiyama, Matthias Krauledat, and Klaus-Robert Müller. Covariate shift adaptation by importance weighted cross validation. Journal of Machine Learning Research, 8 (5), 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.7, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[46] Akinori Tanaka. Discriminator optimal transport. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.729, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[47] Dustin Tran, Rajesh Ranganath, and David M Blei. Deep and hierarchical implicit models. arXiv preprint arXiv:1702.08896, 7(3):13, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[48] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[49] Ryan Turner, Jane Hung, Eric Frank, Yunus Saatchi, and Jason Yosinski. Metropolis-hastings generative adversarial networks. In ICML, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[50] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. In NeurIPS, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[51] Chirag Vashist, Shichong Peng, and Ke Li. Rejection sampling imle: Designing priors for better few-shot image synthesis. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[52] Yaxing Wang, Chenshen Wu, Luis Herranz, Joost van de Weijer, Abel Gonzalez-Garcia, and Bogdan Raducanu. Transferring GANs: generating images from limited data. In ECCV, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.192, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[53] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, Luis Herranz, Fahad Shahbaz Khan, and Joost van de Weijer. MineGAN: effective knowledge transfer from GANs to target domains with few images. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[54] Zhendong Wang, Yifan Jiang, Huangjie Zheng, Peihao Wang, Pengcheng He, Zhangyang Wang, Weizhu Chen, and Mingyuan Zhou. Patch diffusion: Faster and more data-efficient training of diffusion models. In NeurIPS, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[55] Zhendong Wang, Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Diffusion-gan: Training gans with diffusion. In ICLR, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.904, + 0.403 + ], + "angle": 0, + "content": "[56] Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Data-efficient instance generation from instance discrimination. In NeurIPS, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[57] Ceyuan Yang, Yujun Shen, Yinghao Xu, Deli Zhao, Bo Dai, and Bolei Zhou. Improving GANs with a dynamic discriminator. In NeurlPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[58] Mengping Yang, Zhe Wang, Ziqiu Chi, and Yanbing Zhang. Fegan: Exploiting frequency components for training gans under limited data. In NeurIPS, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[59] Mengping Yang, Zhe Wang, Wenyi Feng, Qian Zhang, and Ting Xiao. Improving few-shot image generation by structural discrimination and textural modulation. In ACM MM, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[60] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[61] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In ICML, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[62] Zhaoyu Zhang, Mengyan Li, and Jun Yu. On the convergence and mode collapse of GAN. In SIGGRAPH Asia 2018 Technical Briefs, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[63] Zhaoyu Zhang, Yang Hua, Guanxiong Sun, Hui Wang, and Seán McLoone. Improving the training of the gans with limited data via dual adaptive noise injection. In ACMMM, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[64] Zhaoyu Zhang, Yang Hua, Guanxiong Sun, Hui Wang, and Seán McLoone. Improving the leaking of augmentations in data-efficient gans via adaptive negative data augmentation. In WACV, 2024. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[65] Zhaoyu Zhang, Yang Hua, Hui Wang, and Seán McLoone. Improving the fairness of the min-max game in GANs training. In WACV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.887 + ], + "angle": 0, + "content": "[66] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient GAN training. In NeurIPS, 2020. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.887 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30691" + } + ] +] \ No newline at end of file diff --git a/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/9b5acaea-53bf-4055-87a9-1a5bf68ccd00_origin.pdf b/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/9b5acaea-53bf-4055-87a9-1a5bf68ccd00_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..822b29a51f1f86472b7eaede2308ac9d481f4f70 --- /dev/null +++ b/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/9b5acaea-53bf-4055-87a9-1a5bf68ccd00_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55b4a3b0526a32a0331e970e5407b2006a2c1b0420915ac1fdf9d8c35c1dcb8d +size 2046142 diff --git a/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/full.md b/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..65ffee67fe0233cd1a195ec414e9dc3bc32107c9 --- /dev/null +++ b/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/full.md @@ -0,0 +1,334 @@ +# Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling + +Zhaoyu Zhang $^{1}$ , Yang Hua $^{1}$ , Guanxiong Sun $^{2}$ , Hui Wang $^{1}$ , Seán McLoone $^{1}$ $^{1}$ Queen's University Belfast $^{2}$ Huawei UKRD + +{zzhang55,Y.Hua,gsun02,h.wang,s.mcloone}@qub.ac.uk + +# Abstract + +Data-Efficient Generative Adversarial Nets (DE-GANs) have become more and more popular in recent years. Existing methods apply data augmentation, noise injection and pre-trained models to maximally increase the number of training samples thus improving the training of DE-GANs. However, none of these methods considers the sample quality during training, which can also significantly influence the training of DE-GANs. Focusing on sample quality during training, in this paper, we are the first to incorporate discriminator rejection sampling (DRS) into the training process and introduce a novel method, called quality aware dynamic discriminator rejection sampling (QADDRS). Specifically, QADDRS consists of two steps: (1) the sample quality aware step, which aims to obtain the sorted critic scores, i.e., the ordered discriminator outputs, on real/fake samples in the current training stage; (2) the dynamic rejection step that obtains dynamic rejection number $N$ , where $N$ is controlled by the overfitting degree of discriminator $(D)$ during training. When updating the parameters of $D$ , the $N$ high critic score real samples and the $N$ low critic score fake samples in the minibatch are rejected dynamically based on the overfitting degree of $D$ . As a result, QADDRS can avoid $D$ becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of $D$ issue during training. Extensive experiments on several datasets demonstrate that integrating QADDRS into different DE-GANs can achieve better performance and deliver state-of-the-art results. Codes are available at https://github.com/zzhang05/QADDRS. + +# 1. Introduction + +Over the past decade, Generative Adversarial Nets (GANs) [12] have demonstrated impressive performance in synthesizing images and videos [20, 21, 23, 59, 65]. However, this success heavily depends on access to large-scale datasets, which can be costly, time-consuming, and even impractical + +![](images/03b43eb97c5800474cd86606b3972e4f89afc9c3bbf1e169222e656ee077a452.jpg) +Figure 1. Comparison of discriminator outputs on both real and fake samples with StyleGAN2 and StyleGAN2 + QADDRS using the 100-shot Obama dataset. Applying QADDRS to StyleGAN2 can effectively prevent $D$ from being overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of $D$ issue. + +to collect and clean in certain domains. As a result, data-efficient GANs (DE-GANs) [28, 38, 39] have garnered significant attention. + +Recent studies have applied techniques such as data augmentation [8, 9, 19, 22, 66], noise injection [55, 63] and pre-trained models [25, 44] to improve the DE-GANs training. All of these methods aim to maximally increase the number of training samples under data-efficient scenarios. However, none of these methods considers the sample quality during training, which can also significantly influence the DE-GANs training. For instance, as pointed out in several existing approaches [22, 55, 64], the highly undesirable quality samples produced by data augmentation or noise injection can be detrimental to DE-GANs training. + +Focusing on sample quality during training, in this paper, we are the first to introduce discriminator rejection sampling (DRS) [5] into the training process and propose a novel method, termed quality aware dynamic discrimina + +tor rejection sampling (QADDRS). Specifically, QADDRS comprises two steps: (1) the sample quality aware step that obtains the ranked critic scores, i.e., the sorted discriminator $(D)$ outputs, on real/fake samples to measure the quality of the real/fake samples in the current training stage and (2) the dynamic rejection step which is determined by the dynamic rejection number $N$ , where $N$ adapts based on the overfitting degree of $D$ . During updating the parameters of $D$ process, the $N$ real samples with the high critic scores and the $N$ fake samples with the low critic score in the minibatch are dynamically rejected according to the overfitting degree of $D$ . As illustrated in Figure 1, QADDRS can effectively prevent $D$ from becoming overly confident in distinguishing real and fake samples, thereby alleviating the overfitting of $D$ issue during training. + +To sum up, the key contributions are shown as follows: (1) We introduce DRS into the GANs training process and propose a novel method named quality aware dynamic discriminator rejection sampling (QADDRS) to improve DEGANs training. As far as we are aware, QADDRS is the first approach that considers sample quality during DE-GANs training. (2) We present the theoretical analysis of QADDRS in GANs training, proving its robustness and rationality. Furthermore, we observe that applying QADDRS to DE-GANs can avoid $D$ becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of $D$ issue during training. (3) Experiments on low-shot [66] and FFHQ [23] datasets with different methods demonstrate that QADDRS can improve performance across various DE-GANs and deliver state-of-the-art performance over existing GANs and diffusion models. + +# 2. Related Work + +# 2.1. Generative Adversarial Nets (GANs) + +In the past ten years, generative adversarial nets (GANs) [12] have been among the most popular generative models [41, 50]. GANs consist of two networks: a generator $(G)$ and a discriminator $(D)$ , which engage in a min-max game. The $G$ seeks to generate realistic-looking synthetic samples from noise input, while the $D$ attempts to distinguish whether the input sample originates from the $G$ or from real samples. GANs are known to suffer from training instability, often resulting in low-quality and limited diversity in generated data [62]. To overcome this challenge, different approaches have been presented, including more advanced architectures [6, 34, 35, 42, 58, 61], more stable objective functions [4, 14, 15, 33, 43], and improved training strategies [10, 20, 28, 31, 32, 57, 60] to generate more realistic and diverse samples. + +# 2.2. Data-Efficient GANs + +Recently, training Data-Efficient GANs (DE-GANs) has gained increasing significance. Studies [8, 9, 19, 22, 66] have shown that DE-GAN training often suffers from discriminator $(D)$ overfitting. To address this issue, different approaches have been introduced, including novel data augmentation techniques [8, 9, 19, 22, 66], improved regularization methods for $D$ [11, 30, 48], noise injection [55, 63], and pre-trained models [25, 44]. + +# 2.3. Rejection Sampling in Generative Models + +In recent years, rejection sampling has played a significant role in increasing the generated image quality as a posttraining technique. Rejection sampling is commonly used in several kinds of generative models, such as GANs [5], diffusion models [37] and IMLE [51]. Among these methods, the most well-known method in GANs is discriminator rejection sampling (DRS). DRS has inspired the development of numerous techniques in GANs, such as MH-GAN [49], DDLS [7], DOT [46], DGflow [2] and LatentRS [18]. Although these studies have achieved great success in rejecting lower-quality generated samples, they are all posttraining methods and cannot directly benefit the training of DE-GANs. In this study, we are the first to incorporate the DRS in the training stage to improve the DE-GANs training by considering sample quality. + +# 3. Methodology + +# 3.1. Quality Aware Dynamic Discriminator Rejection Sampling (QADDRS) + +As discussed in §2.3, existing discriminator rejection sampling (DRS) methods are rejecting low-quality generated samples in the post-training process to improve performance. However, it is infeasible to directly apply DRS to the training process, as it can cause the number of real and generated samples for the input of $D$ to become unbalanced during training. + +In this paper, we propose a novel approach called quality aware dynamic discriminator rejection sampling (QADDRS) to improve DE-GANs training by considering sample quality. The overview of applying QADDRS to update both $G$ and $D$ is shown in Figure 2. Specifically, when updating $G$ , all of the fake samples are used to update the parameters of $G$ , following the prior approaches [22, 66]. In contrast, when updating $D$ , the $N$ high critic score real samples and the $N$ low critic score fake samples in the minibatch are dynamically rejected, and the rest of the real and fake samples are used to update the parameters of $D$ . Consequently, QADDRS can prevent $D$ from becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of $D$ issue during training. The selection of high critic score real samples and low critic score + +![](images/0064759bf4c6f04c7a3d16640292945654e7af9dd25413af1f4f0bd16ac46857.jpg) +Figure 2. An overview of quality aware dynamic discriminator rejection sampling (QADDRS) for updating $G$ (left) and updating $D$ (right) in DE-GANs training. When updating $G$ , all of the generated samples are used to update the parameters of $G$ . When updating $D$ , the $N$ high critic score real samples and $N$ low critic score fake samples in the minibatch are dynamically rejected, and only the rest of the real and fake samples are used to update the parameters of $D$ . The selection of high critic score real samples and low critic score fake samples is guided by the discriminator outputs on both real and fake data during training. Best viewed in color. + +fake samples is guided by the sample quality-aware step, while the rejection of $N$ real and fake samples is controlled by the dynamic rejection step. + +For the sample quality aware step, the quality of real/fake samples can be represented by different critic scores on real/fake samples. Based on this, we sort real/fake samples via critic scores, which are based on the discriminator outputs on real/fake data in the current training stage. + +For the dynamic rejection step, the dynamic rejection number $N$ is introduced to control sample rejection. It is obvious that rejecting real samples in QADDRS is highly undesirable, especially under data-efficient scenarios, which could cause a further reduction of training samples. To address this, we introduce a dynamic mechanism for the QADDRS by setting the rejection number $N$ to be adjusted dynamically according to the degree of overfitting without manual tuning. To achieve this, following ANDA [64], APA [19] and ADA [22], we first utilize an overfitting heuristic $\eta$ which can quantify the $D$ 's overfitting degree as follows: + +$$ +\eta = \mathbb {E} \left(\operatorname {s i g n} \left(D _ {\text {r e a l}}\right)\right), D _ {\text {r e a l}} = \operatorname {l o g i t} (D (x)), \tag {1} +$$ + +where $\mathrm{sign}(\cdot)$ denotes the sign function, which returns $+1$ for a non-negative input; $-1$ , otherwise. Next, we follow ADA, APA and ANDA by using $\eta$ to obtain the overfitting probability $p$ . The larger value of $p$ means the greater over + +![](images/98089418bba55c5446e5d788e39e3711ea66ff7c92b0c78cbb58203d6c39a196.jpg) +Figure 3. The curve of the average number of real samples rejected by QADDRS $(\lambda = 0.25)$ when updating $D$ for every 4K fake samples shown to $D$ during training. The experiments are conducted on the 100-shot Obama dataset using StyleGAN2. + +fitting degree. Next, we set the initial rejection number $N$ as 0, and adjust its value based on the overfitting probability $p$ as follows: + +$$ +N = \left\{ \begin{array}{l l} N + 1 & (p) \\ N - 1 & (1 - p) \end{array} \right. \tag {2} +$$ + +where $N$ is adjusted to $N + 1$ under the probability $p$ and is adjusted to $N - 1$ under the probability $1 - p$ . We introduce + +a hyperparameters $\lambda$ and set the range of $N$ as $N_{\min} = 0$ and $N_{\max} = \lambda \times$ batchsize. To further demonstrate how the number $N$ evolves during the training process, the value of $N$ in every 4K real images shown to $D$ during training can be found in Figure 3. + +In summary, employing QADDRS in DE-GANs can be formulated as: + +$$ +V _ {D} (G, D) = \mathbb {E} _ {x \sim P _ {R}} [ \tilde {g} _ {1} (D (x _ {m - N})) ] + \mathbb {E} _ {x \sim P _ {G}} [ \tilde {g} _ {2} (D (x _ {m - N})) ], +$$ + +$$ +V _ {G} (G, D) = - \mathbb {E} _ {x \sim P _ {G}} [ \tilde {f} ((D (x _ {m}))) ], \tag {3} +$$ + +where $x_{m}$ is a batch of samples, $m$ denotes the total samples number in a batch, and $\tilde{g}_{1}$ , $\tilde{g}_{2}$ and $\tilde{f}$ are scalar-to-scalar functions. For the StyleGAN2 [23] with non-saturating loss, $\tilde{g}_{1}(x) = \log (x)$ , $\tilde{g}_{2}(x) = \log (1 - x)$ and $\tilde{f}(x) = \log (x)$ . For the FastGAN [30] with hinge loss, $\tilde{g}_{1}(x) = \min (0, -1 + x)$ , $\tilde{g}_{2}(x) = \min (0, -1 - x)$ and $\tilde{f}(x) = x$ . + +# 3.2. Theoretical Analysis + +# 3.2.1 QADDRS in GANs training + +To further demonstrate how QADDRS can influence GANs training in theory, we first revisit the hinge loss in GANs shown below: + +$$ +\begin{array}{l} V _ {D} (D, G) = \mathbb {E} _ {x \sim P _ {R}} [ \min (0, - 1 + D (x)) ] \\ + \mathbb {E} _ {x \sim P _ {G}} [ \min (0, - 1 - D (x)) ], \tag {4} \\ \end{array} +$$ + +$$ +V _ {G} (D, G) = - \mathbb {E} _ {x \sim P _ {G}} [ D (x) ]. +$$ + +For Eq.(4), $D(x) > 1$ on the real data or $D(G(z)) < -1$ on the fake data can lead to the loss value becoming 0, resulting in no gradient for updating $D$ . Therefore, the key essence of hinge loss in GANs training is to reject the very high critic score real samples (score $> 1$ ) and very low critic score fake samples (score $< -1$ ). Following this, our proposed QADDRS further rejects the real and fake samples based on the overfitting degree of $D$ to enhance the DE-GANs training. We then consider how QADDRS influences the distributions of real data $(P_R)$ and fake data $(P_G)$ in GANs training as follows. + +According to Figure 2, QADDRS can be regarded as involving two actions, i.e., rejecting low critic score fake samples and rejecting high critic score real samples. Because rejecting real/fake samples in QADDRS is performed dynamically, the key operations of QADDRS on the distributions are re-weighting both real and fake samples for their data distributions, respectively. Consequently, in rejecting the low critic score fake samples action, applying QADDRS can transform the original fake data distribution $(P_{G})$ to the re-weighted fake data distribution $(w_{1}P_{G})$ . Similarly, in rejecting the high critic score real samples action, utilizing QADDRS can transform the original real data distribution $(P_{R})$ to the re-weighted real data distribution $(w_{2}P_{R})$ . + +# 3.2.2 Convergence analysis of QADDRS in StyleGAN2 and FastGAN + +Based on the analysis in §3.2.1, we then analyze the QADDRS on both StyleGAN2 with non-saturating loss and FastGAN with hinge loss as follows. + +1. For the StyleGAN2 with non-saturating loss, QADDRS can effectively introduce the advantages of hinge loss into the non-saturating loss via dynamically rejecting the $N$ high critic score real samples and the $N$ low critic score fake samples. Following §3.2.1 and [3], the optimal discriminator of QADDRS with non-saturating loss is defined as: + +$$ +D ^ {*} (x) = \frac {w _ {2} (x) P _ {R} (x)}{w _ {2} (x) P _ {R} (x) + w _ {1} (x) P _ {G} (x)}. \tag {5} +$$ + +With $D^{*}$ as the optimal discriminator, according to Theorem 2.5 in [3], training the $G$ with non-saturating loss can be expressed as: + +$$ +V _ {G} (G, D ^ {*}) = \mathbf {K L} \left(w _ {1} P _ {G} \mid \mid w _ {2} P _ {R}\right) - 2 \mathbf {J S} \left(w _ {1} P _ {G} \mid \mid w _ {2} P _ {R}\right), \tag {6} +$$ + +where KL and JS denote the Kullback-Leibler divergence and the Jensen-Shannon divergence, respectively. Eq.(6) shows that applying QADDRS to the non-saturating loss can lead to the minimization of KL-2JS divergence between the $w_{1}P_{G}$ and $w_{2}P_{R}$ . Then, according to [3], this minimization process depends on the supports of $w_{1}P_{G}$ and $w_{2}P_{R}$ . Previous studies [6, 13, 21, 45] have established that re-weighting operations applied to distributions do not alter their support. Therefore, the minimization of Eq.(6) corresponds to the minimization of KL-2JS divergence between the $P_{G}$ and $P_{R}$ , which demonstrates that applying QADDRS to non-saturating loss does not influence convergence. + +2. For the FastGAN with hinge loss, QADDRS can further take advantage by dynamically rejecting the $N$ high critic score real samples and the $N$ low critic score fake samples if all $D(x)$ and $D(G(z))$ are in $[-1, 1]$ . According to the §3.2.1 and [29, 35, 47], optimizing the hinge loss with QADDRS is equivalent to minimizing the reverse KL divergence term $\mathbf{KL}(w_1P_G\parallel w_2P_R)$ . With the fact that the supports of $w_1P_G$ and $w_2P_R$ are the same as the supports of $P_G$ and $P_R$ , applying QADDRS to hinge loss does not influence the convergence. + +# 3.2.3 Analysis of QADDRS in DE-GANs + +According to the theory presented in [64], data augmentation (DA) is commonly used in DE-GANs, which transforms the real and fake data distributions as $P_R^T$ and $P_G^T$ , respectively. As a result, applying QADDRS can result in the optimization between the $w_1P_G^T$ and $w_2P_R^T$ for both StyleGAN2 with non-saturating loss and FastGAN with hinge + +
MethodBackboneMAPre-training100-shotAnimal-Face
ObamaGrumpy CatPandaCatDog
EDM [24]DiffusionYesNo51.3036.9023.7048.60100.10
EDM + DA [24]DiffusionYesNo37.1029.9410.8136.8857.14
Patch Diffusion [54]DiffusionYesNo41.4730.8913.2543.7172.17
Scale/shift [40]GANNoYes50.7234.2021.3854.8383.04
MineGAN [53]GANNoYes50.6334.5414.8454.4593.03
TransferGAN [52]GANNoYes48.7334.0623.2052.6182.38
TransferGAN + DA [66]GANYesYes39.8529.7717.1249.1065.57
FreezeD [36]GANNoYes41.8731.2217.9547.7070.46
StyleGAN2 [23]GANNoNo80.2048.9034.2771.71131.90
FakeCLR [28]GANYesNo26.9519.568.4226.3442.02
FastGAN [30]GANYesNo35.8025.759.7033.8552.46
FreGAN [58]GANYesNo33.3924.938.9731.0547.85
AugSelf-StyleGAN2 [17]GANYesNo26.0019.818.3630.5348.19
CHAIN [39]GANYesNo20.9417.576.9319.5830.88
NICE [38]GANYesNo20.0815.638.1822.7028.65
Projected GAN [44]GANYesYes11.2115.803.9818.0117.88
StyleGAN2* [23]GANYesNo65.5739.9225.0851.6687.96
+ QADDRSGANYesNo53.5435.7720.5947.8773.29
StyleGAN2 + Diff-Augment [66]GANYesNo46.8727.0812.0642.4458.85
+ QADDRSGANYesNo41.2726.5310.8338.1551.83
StyleGAN2 + ADA [22]GANYesNo45.6926.6212.9040.7756.83
+ QADDRSGANYesNo40.8326.2711.9936.2553.01
InsGen + ADA [56]GANYesNo32.4222.019.8533.0144.93
+ QADDRSGANYesNo25.4119.017.9923.6935.01
Diffusion-Projected GAN (FastGAN) [55]GANYesYes10.5415.133.3917.8617.22
+ QADDRSGANYesYes9.9314.723.0816.9416.57
+ +Table 1. FID [16] scores (lower is better) of different methods on ${256} \times {256}$ low-shot datasets. MA donates Massive Augmentation [9], i.e., xflipping. For a fair comparison, FID is computed using $5\mathrm{\;K}$ generated samples,with the entire dataset serving as the reference distribution. The results are averaged over three evaluations, with relative standard deviations below 1%. + +loss. As the supports of $w_{1}P_{G}^{T}$ and $w_{2}P_{R}^{T}$ are the same as the supports of $P_{R}^{T}$ and $P_{G}^{T}$ , utilizing QADDRS in the DE-GANs does not influence convergence. Furthermore, DE-GANs are known to suffer from the leaking of augmentations issue [22, 64]. In this case, QADDRS can also reweight the out-of-distribution samples for both real and fake data to alleviate the leaking issue. + +# 4. Experiments + +# 4.1. Datasets and Implementation Details + +We select FFHQ [23] and low-shot [66] datasets for the experiments. The FFHQ dataset includes 70K high-resolution face images. In line with FakeCLR [28], we preprocess and resize the FFHQ dataset to $256 \times 256$ and select different subsets consisting of 100, 1K, 2K, and 5K images for all experiments. Low-shot datasets $(256 \times 256)$ consist of five sets (i.e., Obama, Grumpy Cat, Panda, Animal-Face Cat, and Animal-Face Dog) with 100, 100, 100, 160, and 389 training samples, respectively. The commonly used Fréchet Inception Distance (FID) [16], CLIP FID (FIDCLIP) [27], + +and Precision and Recall [26] are applied as the evaluation metrics. Following prior methods [22, 66], the complete dataset serves as the reference distribution for all metric evaluations. + +The implementation details are shown as follows. For the experiments on low-shot datasets with StyleGAN2, StyleGAN2 + Diff-Augment and StyleGAN2 + ADA, we set the batchsize as 4 and the overall training duration as 2000K images. For the experiments on low-shot datasets with InsGen + ADA, we set the batchsize as 64 and the overall training duration as 35000K images. For the experiments on low-shot datasets with Diffusion-Projected GAN, we set the batchsize as 64 and the overall training duration as 100000K images. For the experiments on FFHQ datasets, we set the batchsize as 64 and the overall training duration as 25000K images. The hyperparameter $\lambda$ is set as 0.25 in QADDRS across all experiments. + +# 4.2. Results on Low-shot Datasets + +The FID results of different methods on the $256 \times 256$ low-shot datasets are shown in Table 1. We show the effec + +![](images/ac981f9b37758324ef0ff0e4c19a165a4d865ffb1e936256e5a260da2d7ba097.jpg) +FID 65.57 + +![](images/c61b1b5d47bdf4c5ba543de382fd94edf86075edc4eba83cb69112abcbc59914.jpg) +FID 25.08 + +![](images/2190ea3082bc14db5cb2fe545f5b2d786398f67499b4c8ac3ad8846da896977c.jpg) +FID 39.92 + +![](images/46fb38eb15da47f622e9bfbccd1b9f8919b1f85fb1f78e948884a5ae9d775742.jpg) +FID 51.66 + +![](images/33c564d0e6838947465d03631665a01db1cc92ab82aea250a0651d9b927a4061.jpg) +FID 87.96 + +![](images/5080debdd0e6835c5d637a65469e5f2416ce9fc9aef6605184e35a48dd657c7d.jpg) +FID 53.54 (-12.03) + +![](images/730253ce8eb5570485746e38557d1fb36dd31cf0db7dfba5bd0d692aa238d74d.jpg) +FID 20.59 (-4.49) + +![](images/76b4a2945e3ea29a2237edd468cce315b1f3441c12ed026311a7d514a86449f5.jpg) +FID 35.77 (-4.15) +Figure 4. The compared generated results between StyleGAN2 and StyleGAN2 + QADDRS on low-shot datasets. Top: Images generated by StyleGAN2. Bottom: Images generated by StyleGAN2 + QADDRS. The reduction of FID in red color shows the improvement of StyleGAN2 + QADDRS over the baseline StyleGAN2. + +![](images/526620215b5210c465687773949aed9aab00809885a32e318e4bc623bb2cd7c0.jpg) +FID 47.87 (-3.79) + +![](images/c8ce54616d2867846ca0ed8081d3f842689e8578b5363b0cfd31526990b6ff65.jpg) +FID 73.29 (-14.67) + +
MethodBackboneMAPre-trainingFFHQ
1001K2K5K
EDM [24]DiffusionYesNo79.10---
EDM + DA [24]DiffusionYesNo50.7330.7527.1724.51
Patch Diffusion [54]DiffusionYesNo44.4528.0325.3222.63
StyleGAN2 [23]GANNoNo179.00100.1654.3049.68
StyleGAN2 + Diff-Augment [66]GANYesNo61.9125.6624.3210.45
StyleGAN2 + ADA [22]GANYesNo85.821.2915.3910.96
APA [19]GANYesNo65.0018.8916.908.38
AugSelf-StyleGAN2+ [17]GANYesNo-20.39-9.15
FakeCLR [28]GANYesNo42.5615.929.907.25
Projected GAN [44]GANYesYes26.2511.128.256.85
InsGen [56]GANYesNo45.7518.2111.477.83
+ QADDRSGANYesNo43.6717.1610.737.51
Diffusion-Projected GAN (FastGAN) [55]GANYesYes25.478.767.996.59
+ QADDRSGANYesYes24.468.517.886.48
Diffusion-Projected GAN (FastGAN Lite) [55]GANYesYes20.657.025.474.74
+ QADDRSGANYesYes19.896.575.184.69
+ +Table 2. FID scores (lower is better) of different methods on the ${256} \times {256}\mathrm{{FFHQ}}$ dataset. In line with FakeCLR [28], the experiments are performed on 100, 1K, 2K and 5K subsets of the FFHQ dataset. MA donates Massive Augmentation [9], i.e., xflipping. The results are averaged over five evaluations, with relative standard deviations below 1%. + +tiveness of the QADDRS with both StyleGAN2 and FastGAN using five different commonly-used DE-GANs, i.e., StyleGAN2, StyleGAN2 + Diff-Augment, StyleGAN2 + ADA, InsGen + ADA and Diffusion-Projected GAN. As shown in Table 1, the proposed QADDRS can increase the performance of these five different DE-GANs, demonstrating that QADDRS can be regarded as a general plugin training strategy for DE-GANs. As shown in Table 7, Diffusion-Projected GAN (FastGAN) achieves better per + +formance than Diffusion-Projected GAN (FastGAN Lite), i.e., applying the lite version generator in FastGAN, on low-shot datasets. Therefore, we report the results of the Diffusion-Projected GAN (FastGAN) in Table 1. Although Diffusion-Projected GAN (FastGAN) has already utilized all widely-used approaches, i.e., DA, regularizations, noise injection and pre-trained models, to boost performance on the low-shot datasets, adding QADDRS can further improve its performance. In addition, following [1], we also report + +
MethodObamaGrumpy CatPandaAF-CatAF-Dog
PRPRPRPRPR
Diffusion-Projected GAN (FastGAN) + QADDRS0.9990.2400.9960.1490.9990.4100.9970.1820.9980.203
0.9990.2420.9980.1510.9990.4180.9980.1860.9990.203
+ +Table 3. The comparison of Precision (P) and Recall (R) (higher is better) with Diffusion-Projected GAN (FastGAN) [55] and Diffusion-Projected GAN (FastGAN) + QADDRS on the $256 \times 256$ low-shot datasets. The results are averaged over three evaluations, with relative standard deviations below $1\%$ . + +
MethodFFHQ-100FFHQ-1KFFHQ-2KFFHQ-5K
PRPRPRPR
Diffusion-Projected GAN (FastGAN Lite) + QADDRS0.7370.0180.7340.2690.7360.3210.7300.377
0.7390.0210.7340.2760.7390.3230.7330.379
+ +Table 4. The comparison of Precision (P) and Recall (R) (higher is better) with Diffusion-Projected GAN (FastGAN Lite) [55] and Diffusion-Projected GAN (FastGAN Lite) + QADDRS on the $256 \times 256$ FFHQ dataset. The results are averaged over five evaluations, with relative standard deviations below $1\%$ . + +
Method100AF
OGCPCD
Diffusion Projected GAN (FastGAN)2.024.782.598.367.21
+ QADDRS1.984.712.538.207.09
+ +Table 5. Compared $\mathrm{FID}_{\mathrm{CLIP}}$ scores between Diffusion Projected GAN (FastGAN) and Diffusion Projected GAN (FastGAN) + QADDRS on low-shot datasets. We use the FID calculation model to calculate the $\mathrm{FID}_{\mathrm{CLIP}}$ . Here, AF denotes AnimalFace, O denotes Obama, GC denotes Grumpy Cat, P denotes Panda, C denotes Cat and D denotes Dog. + +
MethodFFHQ
1001K2K5K
Diffusion Projected GAN (FL)23.9111.459.618.33
+ QADDRS23.3211.149.428.19
+ +the results of several SOTA diffusion models in Table 1. It is clear that GAN-based methods outperform diffusion-based methods on low-shot datasets. To further show the effectiveness of QADDRS, the compared Precision and Recall, and $\mathrm{FID}_{\mathrm{CLIP}}$ of adding QADDRS to Diffusion-Projected GAN (FastGAN) are shown in Tables 3 and 5, respectively. The compared generated images on StyleGAN2 and StyleGAN2 + QADDRS are shown in Figure 4. + +Table 6. Compared $\mathrm{FID}_{\mathrm{CLIP}}$ scores between Diffusion Projected GAN (FastGAN Lite) and Diffusion Projected GAN (FastGAN Lite) + QADDRS on FFHQ datasets. We use the FID calculation model to calculate the $\mathrm{FID}_{\mathrm{CLIP}}$ . FL denotes the FastGAN Lite. + +
MethodFID
Diffusion Projected GAN (FastGAN)17.22
Diffusion Projected GAN (FastGAN Lite)20.91
+ +Table 7. Compared FID scores of Diffusion Projected GAN (FastGAN) and Diffusion Projected GAN (FastGAN Lite) on $256\times 256$ AnimalFace-dog dataset. The results are averaged over three evaluations, with relative standard deviations below $1\%$ . + +# 4.3. Results on FFHQ Dataset + +According to the results of five different DE-GANs on low-shot datasets, InsGen + ADA and Diffusion-Projected GAN achieve SOTA performance on StyleGAN2 and FastGAN, respectively. Therefore, we conduct experiments on the FFHQ dataset by selecting these two types of DEGANs. Following FakeCLR [28], we conduct experiments on the FFHQ 100, FFHQ 1K, FFHQ 2K and FFHQ 5K subsets. As shown in Table 2, QADDRS can improve performance with these two DE-GANs on the FFHQ dataset. More importantly, we report the results of both the Diffusion-Projected GAN (FastGAN) and the Diffusion-Projected GAN (FastGAN Lite) in Table 2. It is clear that Diffusion-Projected GAN (FastGAN Lite) achieves better performance than Diffusion-Projected GAN (FastGAN) on the FFHQ dataset. Additionally, the comparison of Precision and Recall, as well as $\mathrm{FID}_{\mathrm{CLIP}}$ with Diffusion-Projected GAN (FastGAN Lite) and Diffusion-Projected GAN (FastGAN Lite) + QADDRS are shown in Tables 4 and 6, respectively. Diffusion-Projected GAN (FastGAN Lite) + QADDRS outperforms Diffusion-Projected GAN (FastGAN Lite) by adding QADDRS, which demonstrates the + +
λ0.000.250.500.75
FID46.8741.2742.4043.32
+ +Table 8. FID scores on the ${256} \times {256100}$ -shot Obama dataset with StyleGAN2 + Diff-Augment. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below 1%. + +
λ0.000.250.500.75
FID58.8551.8353.3154.45
+ +Table 9. FID scores on the ${256} \times {256}$ AnimalFace-Dog dataset with StyleGAN2 + Diff-Augment. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below 1%. + +
MethodObamaAF-Dog
StyleGAN2 + Diff-Augment46.8758.85
+ Fixed rejecting number N44.1254.85
+ QADDRS41.2751.83
+ +effectiveness of the proposed QADDRS. To further show the superiority of QADDRS, more generated images on the FFHQ dataset can be found in supplementary materials. + +# 4.4. Ablation Studies + +Effectiveness of the hyperparameter $\lambda$ in proposed QADDRS. To demonstrate the effectiveness of the hyperparameter $\lambda$ in the proposed QADDRS, we conduct experiments by selecting different values of $\lambda$ on the 100-shot Obama and AnimalFace-Dog datasets using StyleGAN2 + Diff-Augment. The results, shown in Tables 8 and 9, indicate that setting $\lambda$ to 0.25 achieves better performance. + +Effectiveness of the dynamic rejection number $N$ in the proposed QADDRS. To show that the dynamic rejection number $N$ in QADDRS is effective, we conduct an ablation study by comparing QADDRS with a fixed rejection number $N = 0.25 \times$ batchsize on the 100-shot Obama dataset with StyleGAN2 + Diff-Augment. The results are shown in Table 10. It is obvious that the dynamically determined rejection number $N$ can achieve better performance compared with the fixed rejection number. + +# 4.5. Comparisons of DRS and QADDRS + +We conduct the experiments by comparing QADDRS and the existing DRS method [5] using StyleGAN2 + Diff- + +Table 10. FID scores on the $256 \times 256$ 100-shot Obama and AnimalFace-Dog datasets. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below $1\%$ . + +
MethodObamaAnimalFace-Dog
StyleGAN2 + Diff-Augment46.8758.85
+ DRS [5]45.2256.17
+ QADDRS41.2751.83
+ +Table 11. FID scores on the ${256} \times {256}\;{100}$ -shot Obama and Animal-Face Dog datasets. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below 1%. + +
MethodSeconds / 1K images
Diffusion-Projected GAN (FastGAN) + QADDRS17 +17.25
+ +Table 12. The training time on the ${256} \times {256}\;{100}$ -shot Obama dataset with or without QADDRS using Diffusion-Projected GAN (FastGAN). The results are averaged over ten evaluations on two NVIDIA RTX A5000 GPUs with a batch size of 64. All standard deviations are below 1% relative to the mean. + +Augment on the 100-shot Obama and AnimalFace-dog datasets. The results are shown in Table 11. Based on ADA [22], the overfitting of $D$ during training finally results in the DE-GANs generating low-quality images under limited data settings. Although the existing DRS method can reject lower-quality generated samples during post-training, they can not benefit the training of DE-GANs. Thus, existing DRS methods can not prevent generating low-quality images in average, leaving limited gains under limited data settings. In contrast, QADDRS can alleviate the overfitting issue during training, thereby effectively avoiding producing low-quality images in average and achieving significant improvements. As a result, QADDRS outperforms the existing DRS method for DE-GANs. + +# 4.6. Computational Cost + +The results of the training time on the 100-shot Obama dataset with or without QADDRS using Diffusion-Projected GAN (FastGAN) have been shown in Table 12. The proposed QADDRS results in a negligible increase in computational cost. + +# 5. Conclusion + +In this paper, we propose a novel method called quality aware dynamic discriminator rejection sampling (QADDRS), to improve DE-GANs training by considering sample quality. QADDRS can effectively avoid $D$ becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of $D$ issue. Extensive experiments on multiple datasets show that applying QADDRS to various DE-GANs can achieve better performance and deliver state-of-the-art results. + +# 6. Acknowledgement + +This work is partially supported by the High Performance Computing center of Queen's University Belfast and the Kelvin-2 supercomputer (EPSRC grant EP/T022175/1). First author is supported by a scholarship of the Queen's University Belfast and the Department for the Economy, Northern Ireland. + +# References + +[1] Mehran Aghabozorgi, Shichong Peng, and Ke Li. Adaptive imle for few-shot pretraining-free generative modelling. In ICML, 2023. 6 +[2] Abdul Fatir Ansari, Ming Liang Ang, and Harold Soh. Refining deep generative models via discriminator gradient flow. arXiv preprint arXiv:2012.00780, 2020. 2 +[3] Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks. In ICLR, 2017. 4 +[4] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In ICML, 2017. 2 +[5] Samaneh Azadi, Catherine Olsson, Trevor Darrell, Ian Goodfellow, and Augustus Odena. Discriminator rejection sampling. In ICLR, 2019. 1, 2, 8 +[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 2, 4 +[7] Tong Che, Ruixiang Zhang, Jascha Sohl-Dickstein, Hugo Larochelle, Liam Paull, Yuan Cao, and Yoshua Bengio. Your gan is secretly an energy-based model and you should use discriminator driven latent sampling. In NeurIPS, 2020. 2 +[8] Tianlong Chen, Yu Cheng, Zhe Gan, Jingjing Liu, and Zhangyang Wang. Data-efficient GAN training beyond (just) augmentations: A lottery ticket perspective. In NeurIPS, 2021. 1, 2 +[9] Kaiwen Cui, Jiaxing Huang, Zhipeng Luo, Gongjie Zhang, Fangneng Zhan, and Shijian Lu. Genco: Generative constraining for generative adversarial networks with limited data. In AAAI, 2022. 1, 2, 5, 6, 8 +[10] Emily L Denton, Soumith Chintala, Rob Fergus, et al. Deep generative image models using a laplacian pyramid of adversarial networks. In NeurIPS, 2015. 2 +[11] Tiantian Fang, Ruoyu Sun, and Alex Schwing. Diggan: Discriminator gradient gap regularization for GAN training with limited data. In NeurlPS, 2022. 2 +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2 +[13] Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Scholkopf, and Alexander Smola. A kernel two-sample test. The Journal of Machine Learning Research, 13(1):723-773, 2012. 4 +[14] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein GANs. In NeurIPS, 2017. 2 + +[15] Tianyu Guo, Chang Xu, Jiajun Huang, Yunhe Wang, Boxin Shi, Chao Xu, and Dacheng Tao. On positive-unlabeled classification in GAN. In CVPR, 2020. 2 +[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5 +[17] Liang Hou, Qi Cao, Yige Yuan, Songtao Zhao, Chongyang Ma, Siyuan Pan, Pengfei Wan, Zhongyuan Wang, Huawei Shen, and Xueqi Cheng. Augmentation-aware self-supervision for data-efficient gan training. In NeurIPS, 2023. 5, 6 +[18] Thibaut Issenhuth, Ugo Tanielian, David Picard, and Jeremie Mary. Latent reweighting, an almost free improvement for gans. In WACV, 2022. 2 +[19] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Deceive d: Adaptive pseudo augmentation for GAN training with limited data. In NeurIPS, 2021. 1, 2, 3, 6 +[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 1, 2 +[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 4 +[22] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, 2020. 1, 2, 3, 5, 6, 8 +[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of styleGAN. In CVPR, 2020. 1, 2, 4, 5, 6 +[24] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 5, 6 +[25] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for GAN training. In CVPR, 2022. 1, 2 +[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In NeurIPS, 2019. 5 +[27] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fr\`echet inception distance. In ICLR, 2023. 5 +[28] Ziqiang Li, Chaoyue Wang, Heliang Zheng, Jing Zhang, and Bin Li. Fakeclr: Exploring contrastive learning for solving latent discontinuity in data-efficient gans. In ECCV, 2022. 1, 2, 5, 6, 7 +[29] Jae Hyun Lim and Jong Chul Ye. Geometric GAN. arXiv preprint arXiv:1705.02894, 2017. 4 +[30] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and Ahmed Elgammal. Towards faster and stabilized GAN training for high-fidelity few-shot image synthesis. In ICLR, 2021. 2, 4, 5 +[31] Haozhe Liu, Wentian Zhang, Bing Li, Haoqian Wu, Nanjun He, Yawen Huang, Yuexiang Li, Bernard Ghanem, and Yefeng Zheng. Adaptivemix: Improving GAN training via feature space shrinkage. In CVPR, 2023. 2 + +[32] Steven Liu, Tongzhou Wang, David Bau, Jun-Yan Zhu, and Antonio Torralba. Diverse image generation via self-conditioned GANs. In CVPR, 2020. 2 +[33] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In ICCV, 2017. 2 +[34] Takeru Miyato and Masanori Koyama. cGANs with projection discriminator. In ICLR, 2018. 2 +[35] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 2, 4 +[36] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze the discriminator: a simple baseline for fine-tuning GANs. arXiv preprint arXiv:2002.10964, 2020. 5 +[37] Byeonghu Na, Yeongmin Kim, Minsang Park, Donghyeok Shin, Wanmo Kang, and Il-Chul Moon. Diffusion rejection sampling. In ICML, 2024. 2 +[38] Yao Ni and Piotr Koniusz. Nice: Noise-modulated consistency regularization for data-efficient gans. In NeurIPS, 2023. 1, 5 +[39] Yao Ni and Piotr Koniusz. Chain: Enhancing generalization in data-efficient gans via lipschitz continuity constrained normalization. In CVPR, 2024. 1, 5 +[40] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In ICCV, 2019. 5 +[41] Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In ICML, 2016. 2 +[42] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. In ICLR, 2016. 2 +[43] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. In NeurIPS, 2016. 2 +[44] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected GANs converge faster. In NeurIPS, 2021. 1, 2, 5, 6 +[45] Masashi Sugiyama, Matthias Krauledat, and Klaus-Robert Müller. Covariate shift adaptation by importance weighted cross validation. Journal of Machine Learning Research, 8 (5), 2007. 4 +[46] Akinori Tanaka. Discriminator optimal transport. In NeurIPS, 2019. 2 +[47] Dustin Tran, Rajesh Ranganath, and David M Blei. Deep and hierarchical implicit models. arXiv preprint arXiv:1702.08896, 7(3):13, 2017. 4 +[48] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In CVPR, 2021. 2 +[49] Ryan Turner, Jane Hung, Eric Frank, Yunus Saatchi, and Jason Yosinski. Metropolis-hastings generative adversarial networks. In ICML, 2019. 2 +[50] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. In NeurIPS, 2016. 2 + +[51] Chirag Vashist, Shichong Peng, and Ke Li. Rejection sampling imle: Designing priors for better few-shot image synthesis. In ECCV, 2024. 2 +[52] Yaxing Wang, Chenshen Wu, Luis Herranz, Joost van de Weijer, Abel Gonzalez-Garcia, and Bogdan Raducanu. Transferring GANs: generating images from limited data. In ECCV, 2018. 5 +[53] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, Luis Herranz, Fahad Shahbaz Khan, and Joost van de Weijer. MineGAN: effective knowledge transfer from GANs to target domains with few images. In CVPR, 2020. 5 +[54] Zhendong Wang, Yifan Jiang, Huangjie Zheng, Peihao Wang, Pengcheng He, Zhangyang Wang, Weizhu Chen, and Mingyuan Zhou. Patch diffusion: Faster and more data-efficient training of diffusion models. In NeurIPS, 2023. 5, 6 +[55] Zhendong Wang, Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Diffusion-gan: Training gans with diffusion. In ICLR, 2023. 1, 2, 5, 6, 7 +[56] Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Data-efficient instance generation from instance discrimination. In NeurIPS, 2021. 5, 6 +[57] Ceyuan Yang, Yujun Shen, Yinghao Xu, Deli Zhao, Bo Dai, and Bolei Zhou. Improving GANs with a dynamic discriminator. In NeurlPS, 2022. 2 +[58] Mengping Yang, Zhe Wang, Ziqiu Chi, and Yanbing Zhang. Fegan: Exploiting frequency components for training gans under limited data. In NeurIPS, 2022. 2, 5 +[59] Mengping Yang, Zhe Wang, Wenyi Feng, Qian Zhang, and Ting Xiao. Improving few-shot image generation by structural discrimination and textural modulation. In ACM MM, 2023. 1 +[60] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In ICCV, 2017. 2 +[61] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In ICML, 2019. 2 +[62] Zhaoyu Zhang, Mengyan Li, and Jun Yu. On the convergence and mode collapse of GAN. In SIGGRAPH Asia 2018 Technical Briefs, 2018. 2 +[63] Zhaoyu Zhang, Yang Hua, Guanxiong Sun, Hui Wang, and Seán McLoone. Improving the training of the gans with limited data via dual adaptive noise injection. In ACMMM, 2024. 1, 2 +[64] Zhaoyu Zhang, Yang Hua, Guanxiong Sun, Hui Wang, and Seán McLoone. Improving the leaking of augmentations in data-efficient gans via adaptive negative data augmentation. In WACV, 2024. 1, 3, 4, 5 +[65] Zhaoyu Zhang, Yang Hua, Hui Wang, and Seán McLoone. Improving the fairness of the min-max game in GANs training. In WACV, 2024. 1 +[66] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient GAN training. In NeurIPS, 2020. 1, 2, 5, 6 \ No newline at end of file diff --git a/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/images.zip b/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..42e28ff60d0031c0ad041599f6e3551ae616f937 --- /dev/null +++ b/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b919586f0015612ca6a1d41134bdb0219cd755a442a3e75de6d3803681e5ec4 +size 755051 diff --git a/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/layout.json b/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf984e88df38ac83d5ddfe463fb57f95d43e8030 --- /dev/null +++ b/2025/Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling/layout.json @@ -0,0 +1,10799 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 487, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 487, + 140 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 487, + 140 + ], + "type": "text", + "content": "Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "spans": [ + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "text", + "content": "Zhaoyu Zhang" + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "text", + "content": ", Yang Hua" + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "text", + "content": ", Guanxiong Sun" + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "text", + "content": ", Hui Wang" + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "text", + "content": ", Seán McLoone" + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "text", + "content": "Queen's University Belfast " + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 159, + 487, + 190 + ], + "type": "text", + "content": "Huawei UKRD" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 168, + 191, + 440, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 191, + 440, + 204 + ], + "spans": [ + { + "bbox": [ + 168, + 191, + 440, + 204 + ], + "type": "text", + "content": "{zzhang55,Y.Hua,gsun02,h.wang,s.mcloone}@qub.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "text", + "content": "Data-Efficient Generative Adversarial Nets (DE-GANs) have become more and more popular in recent years. Existing methods apply data augmentation, noise injection and pre-trained models to maximally increase the number of training samples thus improving the training of DE-GANs. However, none of these methods considers the sample quality during training, which can also significantly influence the training of DE-GANs. Focusing on sample quality during training, in this paper, we are the first to incorporate discriminator rejection sampling (DRS) into the training process and introduce a novel method, called quality aware dynamic discriminator rejection sampling (QADDRS). Specifically, QADDRS consists of two steps: (1) the sample quality aware step, which aims to obtain the sorted critic scores, i.e., the ordered discriminator outputs, on real/fake samples in the current training stage; (2) the dynamic rejection step that obtains dynamic rejection number " + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "text", + "content": " is controlled by the overfitting degree of discriminator " + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "text", + "content": " during training. When updating the parameters of " + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "text", + "content": " high critic score real samples and the " + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "text", + "content": " low critic score fake samples in the minibatch are rejected dynamically based on the overfitting degree of " + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "text", + "content": ". As a result, QADDRS can avoid " + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "text", + "content": " becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of " + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "text", + "content": " issue during training. Extensive experiments on several datasets demonstrate that integrating QADDRS into different DE-GANs can achieve better performance and deliver state-of-the-art results. Codes are available at https://github.com/zzhang05/QADDRS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 632, + 135, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 135, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 135, + 644 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Over the past decade, Generative Adversarial Nets (GANs) [12] have demonstrated impressive performance in synthesizing images and videos [20, 21, 23, 59, 65]. However, this success heavily depends on access to large-scale datasets, which can be costly, time-consuming, and even impractical" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 229, + 557, + 403 + ], + "blocks": [ + { + "bbox": [ + 313, + 229, + 557, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 229, + 557, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 557, + 403 + ], + "type": "image", + "image_path": "03b43eb97c5800474cd86606b3972e4f89afc9c3bbf1e169222e656ee077a452.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 411, + 555, + 477 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 477 + ], + "type": "text", + "content": "Figure 1. Comparison of discriminator outputs on both real and fake samples with StyleGAN2 and StyleGAN2 + QADDRS using the 100-shot Obama dataset. Applying QADDRS to StyleGAN2 can effectively prevent " + }, + { + "bbox": [ + 313, + 411, + 555, + 477 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 411, + 555, + 477 + ], + "type": "text", + "content": " from being overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of " + }, + { + "bbox": [ + 313, + 411, + 555, + 477 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 411, + 555, + 477 + ], + "type": "text", + "content": " issue." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 533 + ], + "type": "text", + "content": "to collect and clean in certain domains. As a result, data-efficient GANs (DE-GANs) [28, 38, 39] have garnered significant attention." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 534, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 556, + 665 + ], + "type": "text", + "content": "Recent studies have applied techniques such as data augmentation [8, 9, 19, 22, 66], noise injection [55, 63] and pre-trained models [25, 44] to improve the DE-GANs training. All of these methods aim to maximally increase the number of training samples under data-efficient scenarios. However, none of these methods considers the sample quality during training, which can also significantly influence the DE-GANs training. For instance, as pointed out in several existing approaches [22, 55, 64], the highly undesirable quality samples produced by data augmentation or noise injection can be detrimental to DE-GANs training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Focusing on sample quality during training, in this paper, we are the first to introduce discriminator rejection sampling (DRS) [5] into the training process and propose a novel method, termed quality aware dynamic discrimina" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30682" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "tor rejection sampling (QADDRS). Specifically, QADDRS comprises two steps: (1) the sample quality aware step that obtains the ranked critic scores, i.e., the sorted discriminator " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " outputs, on real/fake samples to measure the quality of the real/fake samples in the current training stage and (2) the dynamic rejection step which is determined by the dynamic rejection number " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " adapts based on the overfitting degree of " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": ". During updating the parameters of " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " process, the " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " real samples with the high critic scores and the " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " fake samples with the low critic score in the minibatch are dynamically rejected according to the overfitting degree of " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": ". As illustrated in Figure 1, QADDRS can effectively prevent " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " from becoming overly confident in distinguishing real and fake samples, thereby alleviating the overfitting of " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " issue during training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 450 + ], + "type": "text", + "content": "To sum up, the key contributions are shown as follows: (1) We introduce DRS into the GANs training process and propose a novel method named quality aware dynamic discriminator rejection sampling (QADDRS) to improve DEGANs training. As far as we are aware, QADDRS is the first approach that considers sample quality during DE-GANs training. (2) We present the theoretical analysis of QADDRS in GANs training, proving its robustness and rationality. Furthermore, we observe that applying QADDRS to DE-GANs can avoid " + }, + { + "bbox": [ + 55, + 258, + 295, + 450 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 258, + 295, + 450 + ], + "type": "text", + "content": " becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of " + }, + { + "bbox": [ + 55, + 258, + 295, + 450 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 258, + 295, + 450 + ], + "type": "text", + "content": " issue during training. (3) Experiments on low-shot [66] and FFHQ [23] datasets with different methods demonstrate that QADDRS can improve performance across various DE-GANs and deliver state-of-the-art performance over existing GANs and diffusion models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 482, + 142, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 142, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 142, + 495 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 509, + 250, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 250, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 250, + 521 + ], + "type": "text", + "content": "2.1. Generative Adversarial Nets (GANs)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "In the past ten years, generative adversarial nets (GANs) [12] have been among the most popular generative models [41, 50]. GANs consist of two networks: a generator " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "(G)" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " and a discriminator " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": ", which engage in a min-max game. The " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " seeks to generate realistic-looking synthetic samples from noise input, while the " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " attempts to distinguish whether the input sample originates from the " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " or from real samples. GANs are known to suffer from training instability, often resulting in low-quality and limited diversity in generated data [62]. To overcome this challenge, different approaches have been presented, including more advanced architectures [6, 34, 35, 42, 58, 61], more stable objective functions [4, 14, 15, 33, 43], and improved training strategies [10, 20, 28, 31, 32, 57, 60] to generate more realistic and diverse samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 434, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 434, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 434, + 83 + ], + "type": "text", + "content": "2.2. Data-Efficient GANs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 90, + 553, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 553, + 186 + ], + "type": "text", + "content": "Recently, training Data-Efficient GANs (DE-GANs) has gained increasing significance. Studies [8, 9, 19, 22, 66] have shown that DE-GAN training often suffers from discriminator " + }, + { + "bbox": [ + 313, + 90, + 553, + 186 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 313, + 90, + 553, + 186 + ], + "type": "text", + "content": " overfitting. To address this issue, different approaches have been introduced, including novel data augmentation techniques [8, 9, 19, 22, 66], improved regularization methods for " + }, + { + "bbox": [ + 313, + 90, + 553, + 186 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 90, + 553, + 186 + ], + "type": "text", + "content": " [11, 30, 48], noise injection [55, 63], and pre-trained models [25, 44]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 192, + 531, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 531, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 531, + 205 + ], + "type": "text", + "content": "2.3. Rejection Sampling in Generative Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 211, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 390 + ], + "type": "text", + "content": "In recent years, rejection sampling has played a significant role in increasing the generated image quality as a posttraining technique. Rejection sampling is commonly used in several kinds of generative models, such as GANs [5], diffusion models [37] and IMLE [51]. Among these methods, the most well-known method in GANs is discriminator rejection sampling (DRS). DRS has inspired the development of numerous techniques in GANs, such as MH-GAN [49], DDLS [7], DOT [46], DGflow [2] and LatentRS [18]. Although these studies have achieved great success in rejecting lower-quality generated samples, they are all posttraining methods and cannot directly benefit the training of DE-GANs. In this study, we are the first to incorporate the DRS in the training stage to improve the DE-GANs training by considering sample quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 400, + 397, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 397, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 397, + 414 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 420, + 553, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 553, + 445 + ], + "type": "text", + "content": "3.1. Quality Aware Dynamic Discriminator Rejection Sampling (QADDRS)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "content": "As discussed in §2.3, existing discriminator rejection sampling (DRS) methods are rejecting low-quality generated samples in the post-training process to improve performance. However, it is infeasible to directly apply DRS to the training process, as it can cause the number of real and generated samples for the input of " + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "content": " to become unbalanced during training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel approach called quality aware dynamic discriminator rejection sampling (QADDRS) to improve DE-GANs training by considering sample quality. The overview of applying QADDRS to update both " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " is shown in Figure 2. Specifically, when updating " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", all of the fake samples are used to update the parameters of " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", following the prior approaches [22, 66]. In contrast, when updating " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " high critic score real samples and the " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " low critic score fake samples in the minibatch are dynamically rejected, and the rest of the real and fake samples are used to update the parameters of " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ". Consequently, QADDRS can prevent " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " from becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " issue during training. The selection of high critic score real samples and low critic score" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30683" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 541, + 296 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 541, + 296 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 541, + 296 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 541, + 296 + ], + "type": "image", + "image_path": "0064759bf4c6f04c7a3d16640292945654e7af9dd25413af1f4f0bd16ac46857.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "text", + "content": "Figure 2. An overview of quality aware dynamic discriminator rejection sampling (QADDRS) for updating " + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "text", + "content": " (left) and updating " + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "text", + "content": " (right) in DE-GANs training. When updating " + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "text", + "content": ", all of the generated samples are used to update the parameters of " + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "text", + "content": ". When updating " + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "text", + "content": " high critic score real samples and " + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "text", + "content": " low critic score fake samples in the minibatch are dynamically rejected, and only the rest of the real and fake samples are used to update the parameters of " + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 301, + 555, + 360 + ], + "type": "text", + "content": ". The selection of high critic score real samples and low critic score fake samples is guided by the discriminator outputs on both real and fake data during training. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 388, + 294, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 294, + 424 + ], + "type": "text", + "content": "fake samples is guided by the sample quality-aware step, while the rejection of " + }, + { + "bbox": [ + 54, + 388, + 294, + 424 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 388, + 294, + 424 + ], + "type": "text", + "content": " real and fake samples is controlled by the dynamic rejection step." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 425, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 425, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 295, + 485 + ], + "type": "text", + "content": "For the sample quality aware step, the quality of real/fake samples can be represented by different critic scores on real/fake samples. Based on this, we sort real/fake samples via critic scores, which are based on the discriminator outputs on real/fake data in the current training stage." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 486, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 295, + 628 + ], + "type": "text", + "content": "For the dynamic rejection step, the dynamic rejection number " + }, + { + "bbox": [ + 54, + 486, + 295, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 486, + 295, + 628 + ], + "type": "text", + "content": " is introduced to control sample rejection. It is obvious that rejecting real samples in QADDRS is highly undesirable, especially under data-efficient scenarios, which could cause a further reduction of training samples. To address this, we introduce a dynamic mechanism for the QADDRS by setting the rejection number " + }, + { + "bbox": [ + 54, + 486, + 295, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 486, + 295, + 628 + ], + "type": "text", + "content": " to be adjusted dynamically according to the degree of overfitting without manual tuning. To achieve this, following ANDA [64], APA [19] and ADA [22], we first utilize an overfitting heuristic " + }, + { + "bbox": [ + 54, + 486, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 54, + 486, + 295, + 628 + ], + "type": "text", + "content": " which can quantify the " + }, + { + "bbox": [ + 54, + 486, + 295, + 628 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 486, + 295, + 628 + ], + "type": "text", + "content": "'s overfitting degree as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 644, + 295, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 644, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 88, + 644, + 295, + 658 + ], + "type": "interline_equation", + "content": "\\eta = \\mathbb {E} \\left(\\operatorname {s i g n} \\left(D _ {\\text {r e a l}}\\right)\\right), D _ {\\text {r e a l}} = \\operatorname {l o g i t} (D (x)), \\tag {1}", + "image_path": "5693fba7ad44fe1af2006cd86abca116d11641f665547b1b3e466e9601332087.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{sign}(\\cdot)" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " denotes the sign function, which returns " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "+1" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " for a non-negative input; " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", otherwise. Next, we follow ADA, APA and ANDA by using " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " to obtain the overfitting probability " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". The larger value of " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " means the greater over" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 386, + 558, + 536 + ], + "blocks": [ + { + "bbox": [ + 313, + 386, + 558, + 536 + ], + "lines": [ + { + "bbox": [ + 313, + 386, + 558, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 558, + 536 + ], + "type": "image", + "image_path": "98089418bba55c5446e5d788e39e3711ea66ff7c92b0c78cbb58203d6c39a196.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 544, + 555, + 590 + ], + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 590 + ], + "type": "text", + "content": "Figure 3. The curve of the average number of real samples rejected by QADDRS " + }, + { + "bbox": [ + 313, + 544, + 555, + 590 + ], + "type": "inline_equation", + "content": "(\\lambda = 0.25)" + }, + { + "bbox": [ + 313, + 544, + 555, + 590 + ], + "type": "text", + "content": " when updating " + }, + { + "bbox": [ + 313, + 544, + 555, + 590 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 544, + 555, + 590 + ], + "type": "text", + "content": " for every 4K fake samples shown to " + }, + { + "bbox": [ + 313, + 544, + 555, + 590 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 544, + 555, + 590 + ], + "type": "text", + "content": " during training. The experiments are conducted on the 100-shot Obama dataset using StyleGAN2." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 611, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 555, + 647 + ], + "type": "text", + "content": "fitting degree. Next, we set the initial rejection number " + }, + { + "bbox": [ + 313, + 611, + 555, + 647 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 611, + 555, + 647 + ], + "type": "text", + "content": " as 0, and adjust its value based on the overfitting probability " + }, + { + "bbox": [ + 313, + 611, + 555, + 647 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 611, + 555, + 647 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 386, + 658, + 553, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 658, + 553, + 684 + ], + "spans": [ + { + "bbox": [ + 386, + 658, + 553, + 684 + ], + "type": "interline_equation", + "content": "N = \\left\\{ \\begin{array}{l l} N + 1 & (p) \\\\ N - 1 & (1 - p) \\end{array} \\right. \\tag {2}", + "image_path": "739417ee355d5f1582d182d6f8dc42420e0ce40f258c2f0fe7b8f55f9fd36767.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is adjusted to " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "N + 1" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " under the probability " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " and is adjusted to " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " under the probability " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "1 - p" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": ". We introduce" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30684" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "text", + "content": "a hyperparameters " + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "text", + "content": " and set the range of " + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "inline_equation", + "content": "N_{\\min} = 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "inline_equation", + "content": "N_{\\max} = \\lambda \\times" + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "text", + "content": " batchsize. To further demonstrate how the number " + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "text", + "content": " evolves during the training process, the value of " + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "text", + "content": " in every 4K real images shown to " + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "text", + "content": " during training can be found in Figure 3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 296, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 296, + 157 + ], + "type": "text", + "content": "In summary, employing QADDRS in DE-GANs can be formulated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 171, + 297, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 297, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 297, + 185 + ], + "type": "interline_equation", + "content": "V _ {D} (G, D) = \\mathbb {E} _ {x \\sim P _ {R}} [ \\tilde {g} _ {1} (D (x _ {m - N})) ] + \\mathbb {E} _ {x \\sim P _ {G}} [ \\tilde {g} _ {2} (D (x _ {m - N})) ],", + "image_path": "b4ffe96738567409e66e413a6b8b5dc5d15e9a973213082dd4888492834dfe52.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 211 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 211 + ], + "type": "interline_equation", + "content": "V _ {G} (G, D) = - \\mathbb {E} _ {x \\sim P _ {G}} [ \\tilde {f} ((D (x _ {m}))) ], \\tag {3}", + "image_path": "7383ee612e24670bed5a2e3b4983ba87fcd60ddc4785758a512607fe6190d306.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": " is a batch of samples, " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": " denotes the total samples number in a batch, and " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\tilde{g}_{1}" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\tilde{g}_{2}" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": " are scalar-to-scalar functions. For the StyleGAN2 [23] with non-saturating loss, " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\tilde{g}_{1}(x) = \\log (x)" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\tilde{g}_{2}(x) = \\log (1 - x)" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\tilde{f}(x) = \\log (x)" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": ". For the FastGAN [30] with hinge loss, " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\tilde{g}_{1}(x) = \\min (0, -1 + x)" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\tilde{g}_{2}(x) = \\min (0, -1 - x)" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\tilde{f}(x) = x" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 300, + 174, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 174, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 174, + 313 + ], + "type": "text", + "content": "3.2. Theoretical Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 319, + 208, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 208, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 208, + 332 + ], + "type": "text", + "content": "3.2.1 QADDRS in GANs training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "text", + "content": "To further demonstrate how QADDRS can influence GANs training in theory, we first revisit the hinge loss in GANs shown below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 382, + 294, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 382, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 94, + 382, + 294, + 409 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} V _ {D} (D, G) = \\mathbb {E} _ {x \\sim P _ {R}} [ \\min (0, - 1 + D (x)) ] \\\\ + \\mathbb {E} _ {x \\sim P _ {G}} [ \\min (0, - 1 - D (x)) ], \\tag {4} \\\\ \\end{array}", + "image_path": "f71ee807532f5b6d0372728fada26af952997f1aeb4d1397beeb27807c7c856d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 96, + 411, + 211, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 411, + 211, + 423 + ], + "spans": [ + { + "bbox": [ + 96, + 411, + 211, + 423 + ], + "type": "interline_equation", + "content": "V _ {G} (D, G) = - \\mathbb {E} _ {x \\sim P _ {G}} [ D (x) ].", + "image_path": "649c3d943278efa82ca7e6048d17f366f7d5dc775b9e808a618b75aad5e03efc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "text", + "content": "For Eq.(4), " + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "inline_equation", + "content": "D(x) > 1" + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "text", + "content": " on the real data or " + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "inline_equation", + "content": "D(G(z)) < -1" + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "text", + "content": " on the fake data can lead to the loss value becoming 0, resulting in no gradient for updating " + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "text", + "content": ". Therefore, the key essence of hinge loss in GANs training is to reject the very high critic score real samples (score " + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "inline_equation", + "content": "> 1" + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "text", + "content": ") and very low critic score fake samples (score " + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "inline_equation", + "content": "< -1" + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "text", + "content": "). Following this, our proposed QADDRS further rejects the real and fake samples based on the overfitting degree of " + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "text", + "content": " to enhance the DE-GANs training. We then consider how QADDRS influences the distributions of real data " + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "inline_equation", + "content": "(P_R)" + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "text", + "content": " and fake data " + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "inline_equation", + "content": "(P_G)" + }, + { + "bbox": [ + 55, + 426, + 296, + 557 + ], + "type": "text", + "content": " in GANs training as follows." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "According to Figure 2, QADDRS can be regarded as involving two actions, i.e., rejecting low critic score fake samples and rejecting high critic score real samples. Because rejecting real/fake samples in QADDRS is performed dynamically, the key operations of QADDRS on the distributions are re-weighting both real and fake samples for their data distributions, respectively. Consequently, in rejecting the low critic score fake samples action, applying QADDRS can transform the original fake data distribution " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "(P_{G})" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " to the re-weighted fake data distribution " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "(w_{1}P_{G})" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ". Similarly, in rejecting the high critic score real samples action, utilizing QADDRS can transform the original real data distribution " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "(P_{R})" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " to the re-weighted real data distribution " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "(w_{2}P_{R})" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 72, + 555, + 95 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 555, + 95 + ], + "type": "text", + "content": "3.2.2 Convergence analysis of QADDRS in StyleGAN2 and FastGAN" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 102, + 555, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 102, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 555, + 139 + ], + "type": "text", + "content": "Based on the analysis in §3.2.1, we then analyze the QADDRS on both StyleGAN2 with non-saturating loss and FastGAN with hinge loss as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 140, + 555, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 140, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 315, + 140, + 555, + 224 + ], + "type": "text", + "content": "1. For the StyleGAN2 with non-saturating loss, QADDRS can effectively introduce the advantages of hinge loss into the non-saturating loss via dynamically rejecting the " + }, + { + "bbox": [ + 315, + 140, + 555, + 224 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 315, + 140, + 555, + 224 + ], + "type": "text", + "content": " high critic score real samples and the " + }, + { + "bbox": [ + 315, + 140, + 555, + 224 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 315, + 140, + 555, + 224 + ], + "type": "text", + "content": " low critic score fake samples. Following §3.2.1 and [3], the optimal discriminator of QADDRS with non-saturating loss is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 358, + 232, + 554, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 232, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 358, + 232, + 554, + 259 + ], + "type": "interline_equation", + "content": "D ^ {*} (x) = \\frac {w _ {2} (x) P _ {R} (x)}{w _ {2} (x) P _ {R} (x) + w _ {1} (x) P _ {G} (x)}. \\tag {5}", + "image_path": "69536d8b83499e9c52245d0f7d19b797a26b732fa68bda2c18ac7be1832f24bb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 262, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 262, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 326, + 262, + 555, + 299 + ], + "type": "text", + "content": "With " + }, + { + "bbox": [ + 326, + 262, + 555, + 299 + ], + "type": "inline_equation", + "content": "D^{*}" + }, + { + "bbox": [ + 326, + 262, + 555, + 299 + ], + "type": "text", + "content": " as the optimal discriminator, according to Theorem 2.5 in [3], training the " + }, + { + "bbox": [ + 326, + 262, + 555, + 299 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 326, + 262, + 555, + 299 + ], + "type": "text", + "content": " with non-saturating loss can be expressed as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 330, + 308, + 553, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 308, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 330, + 308, + 553, + 331 + ], + "type": "interline_equation", + "content": "V _ {G} (G, D ^ {*}) = \\mathbf {K L} \\left(w _ {1} P _ {G} \\mid \\mid w _ {2} P _ {R}\\right) - 2 \\mathbf {J S} \\left(w _ {1} P _ {G} \\mid \\mid w _ {2} P _ {R}\\right), \\tag {6}", + "image_path": "1a92b61387e65ea7b83ad9cabda36ab671e3cb861de3226f8311e3517f2fd903.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "text", + "content": "where KL and JS denote the Kullback-Leibler divergence and the Jensen-Shannon divergence, respectively. Eq.(6) shows that applying QADDRS to the non-saturating loss can lead to the minimization of KL-2JS divergence between the " + }, + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "inline_equation", + "content": "w_{1}P_{G}" + }, + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "inline_equation", + "content": "w_{2}P_{R}" + }, + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "text", + "content": ". Then, according to [3], this minimization process depends on the supports of " + }, + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "inline_equation", + "content": "w_{1}P_{G}" + }, + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "inline_equation", + "content": "w_{2}P_{R}" + }, + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "text", + "content": ". Previous studies [6, 13, 21, 45] have established that re-weighting operations applied to distributions do not alter their support. Therefore, the minimization of Eq.(6) corresponds to the minimization of KL-2JS divergence between the " + }, + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "inline_equation", + "content": "P_{G}" + }, + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "inline_equation", + "content": "P_{R}" + }, + { + "bbox": [ + 326, + 331, + 554, + 486 + ], + "type": "text", + "content": ", which demonstrates that applying QADDRS to non-saturating loss does not influence convergence." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "text", + "content": "2. For the FastGAN with hinge loss, QADDRS can further take advantage by dynamically rejecting the " + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "text", + "content": " high critic score real samples and the " + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "text", + "content": " low critic score fake samples if all " + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "inline_equation", + "content": "D(x)" + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "inline_equation", + "content": "D(G(z))" + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "text", + "content": " are in " + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "text", + "content": ". According to the §3.2.1 and [29, 35, 47], optimizing the hinge loss with QADDRS is equivalent to minimizing the reverse KL divergence term " + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{KL}(w_1P_G\\parallel w_2P_R)" + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "text", + "content": ". With the fact that the supports of " + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "inline_equation", + "content": "w_1P_G" + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "inline_equation", + "content": "w_2P_R" + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "text", + "content": " are the same as the supports of " + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "inline_equation", + "content": "P_G" + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "inline_equation", + "content": "P_R" + }, + { + "bbox": [ + 314, + 487, + 555, + 606 + ], + "type": "text", + "content": ", applying QADDRS to hinge loss does not influence the convergence." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 623, + 495, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 623, + 495, + 635 + ], + "spans": [ + { + "bbox": [ + 314, + 623, + 495, + 635 + ], + "type": "text", + "content": "3.2.3 Analysis of QADDRS in DE-GANs" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "According to the theory presented in [64], data augmentation (DA) is commonly used in DE-GANs, which transforms the real and fake data distributions as " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_R^T" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_G^T" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": ", respectively. As a result, applying QADDRS can result in the optimization between the " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "w_1P_G^T" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "w_2P_R^T" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " for both StyleGAN2 with non-saturating loss and FastGAN with hinge" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30685" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 548, + 388 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 548, + 388 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 548, + 388 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 548, + 388 + ], + "type": "table", + "html": "
MethodBackboneMAPre-training100-shotAnimal-Face
ObamaGrumpy CatPandaCatDog
EDM [24]DiffusionYesNo51.3036.9023.7048.60100.10
EDM + DA [24]DiffusionYesNo37.1029.9410.8136.8857.14
Patch Diffusion [54]DiffusionYesNo41.4730.8913.2543.7172.17
Scale/shift [40]GANNoYes50.7234.2021.3854.8383.04
MineGAN [53]GANNoYes50.6334.5414.8454.4593.03
TransferGAN [52]GANNoYes48.7334.0623.2052.6182.38
TransferGAN + DA [66]GANYesYes39.8529.7717.1249.1065.57
FreezeD [36]GANNoYes41.8731.2217.9547.7070.46
StyleGAN2 [23]GANNoNo80.2048.9034.2771.71131.90
FakeCLR [28]GANYesNo26.9519.568.4226.3442.02
FastGAN [30]GANYesNo35.8025.759.7033.8552.46
FreGAN [58]GANYesNo33.3924.938.9731.0547.85
AugSelf-StyleGAN2 [17]GANYesNo26.0019.818.3630.5348.19
CHAIN [39]GANYesNo20.9417.576.9319.5830.88
NICE [38]GANYesNo20.0815.638.1822.7028.65
Projected GAN [44]GANYesYes11.2115.803.9818.0117.88
StyleGAN2* [23]GANYesNo65.5739.9225.0851.6687.96
+ QADDRSGANYesNo53.5435.7720.5947.8773.29
StyleGAN2 + Diff-Augment [66]GANYesNo46.8727.0812.0642.4458.85
+ QADDRSGANYesNo41.2726.5310.8338.1551.83
StyleGAN2 + ADA [22]GANYesNo45.6926.6212.9040.7756.83
+ QADDRSGANYesNo40.8326.2711.9936.2553.01
InsGen + ADA [56]GANYesNo32.4222.019.8533.0144.93
+ QADDRSGANYesNo25.4119.017.9923.6935.01
Diffusion-Projected GAN (FastGAN) [55]GANYesYes10.5415.133.3917.8617.22
+ QADDRSGANYesYes9.9314.723.0816.9416.57
", + "image_path": "f87d08160e1eb3c8386b46563d9d2e5f8226cdcfad7fb1d2245c7531eed97545.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 407, + 553, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 407, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 553, + 440 + ], + "type": "text", + "content": "Table 1. FID [16] scores (lower is better) of different methods on " + }, + { + "bbox": [ + 55, + 407, + 553, + 440 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 55, + 407, + 553, + 440 + ], + "type": "text", + "content": " low-shot datasets. MA donates Massive Augmentation [9], i.e., xflipping. For a fair comparison, FID is computed using " + }, + { + "bbox": [ + 55, + 407, + 553, + 440 + ], + "type": "inline_equation", + "content": "5\\mathrm{\\;K}" + }, + { + "bbox": [ + 55, + 407, + 553, + 440 + ], + "type": "text", + "content": " generated samples,with the entire dataset serving as the reference distribution. The results are averaged over three evaluations, with relative standard deviations below 1%." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 461, + 295, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 546 + ], + "type": "text", + "content": "loss. As the supports of " + }, + { + "bbox": [ + 55, + 461, + 295, + 546 + ], + "type": "inline_equation", + "content": "w_{1}P_{G}^{T}" + }, + { + "bbox": [ + 55, + 461, + 295, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 461, + 295, + 546 + ], + "type": "inline_equation", + "content": "w_{2}P_{R}^{T}" + }, + { + "bbox": [ + 55, + 461, + 295, + 546 + ], + "type": "text", + "content": " are the same as the supports of " + }, + { + "bbox": [ + 55, + 461, + 295, + 546 + ], + "type": "inline_equation", + "content": "P_{R}^{T}" + }, + { + "bbox": [ + 55, + 461, + 295, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 461, + 295, + 546 + ], + "type": "inline_equation", + "content": "P_{G}^{T}" + }, + { + "bbox": [ + 55, + 461, + 295, + 546 + ], + "type": "text", + "content": ", utilizing QADDRS in the DE-GANs does not influence convergence. Furthermore, DE-GANs are known to suffer from the leaking of augmentations issue [22, 64]. In this case, QADDRS can also reweight the out-of-distribution samples for both real and fake data to alleviate the leaking issue." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 555, + 137, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 137, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 137, + 569 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 575, + 251, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 251, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 251, + 588 + ], + "type": "text", + "content": "4.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "We select FFHQ [23] and low-shot [66] datasets for the experiments. The FFHQ dataset includes 70K high-resolution face images. In line with FakeCLR [28], we preprocess and resize the FFHQ dataset to " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " and select different subsets consisting of 100, 1K, 2K, and 5K images for all experiments. Low-shot datasets " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "(256 \\times 256)" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " consist of five sets (i.e., Obama, Grumpy Cat, Panda, Animal-Face Cat, and Animal-Face Dog) with 100, 100, 100, 160, and 389 training samples, respectively. The commonly used Fréchet Inception Distance (FID) [16], CLIP FID (FIDCLIP) [27]," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 462, + 553, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 553, + 508 + ], + "type": "text", + "content": "and Precision and Recall [26] are applied as the evaluation metrics. Following prior methods [22, 66], the complete dataset serves as the reference distribution for all metric evaluations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "type": "text", + "content": "The implementation details are shown as follows. For the experiments on low-shot datasets with StyleGAN2, StyleGAN2 + Diff-Augment and StyleGAN2 + ADA, we set the batchsize as 4 and the overall training duration as 2000K images. For the experiments on low-shot datasets with InsGen + ADA, we set the batchsize as 64 and the overall training duration as 35000K images. For the experiments on low-shot datasets with Diffusion-Projected GAN, we set the batchsize as 64 and the overall training duration as 100000K images. For the experiments on FFHQ datasets, we set the batchsize as 64 and the overall training duration as 25000K images. The hyperparameter " + }, + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "type": "text", + "content": " is set as 0.25 in QADDRS across all experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 671, + 474, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 474, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 474, + 682 + ], + "type": "text", + "content": "4.2. Results on Low-shot Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": "The FID results of different methods on the " + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": " low-shot datasets are shown in Table 1. We show the effec" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "30686" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 83, + 168, + 158 + ], + "blocks": [ + { + "bbox": [ + 113, + 72, + 148, + 80 + ], + "lines": [ + { + "bbox": [ + 113, + 72, + 148, + 80 + ], + "spans": [ + { + "bbox": [ + 113, + 72, + 148, + 80 + ], + "type": "text", + "content": "FID 65.57" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 83, + 168, + 158 + ], + "lines": [ + { + "bbox": [ + 93, + 83, + 168, + 158 + ], + "spans": [ + { + "bbox": [ + 93, + 83, + 168, + 158 + ], + "type": "image", + "image_path": "ac981f9b37758324ef0ff0e4c19a165a4d865ffb1e936256e5a260da2d7ba097.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 180, + 83, + 256, + 158 + ], + "blocks": [ + { + "bbox": [ + 201, + 72, + 235, + 80 + ], + "lines": [ + { + "bbox": [ + 201, + 72, + 235, + 80 + ], + "spans": [ + { + "bbox": [ + 201, + 72, + 235, + 80 + ], + "type": "text", + "content": "FID 25.08" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 180, + 83, + 256, + 158 + ], + "lines": [ + { + "bbox": [ + 180, + 83, + 256, + 158 + ], + "spans": [ + { + "bbox": [ + 180, + 83, + 256, + 158 + ], + "type": "image", + "image_path": "c61b1b5d47bdf4c5ba543de382fd94edf86075edc4eba83cb69112abcbc59914.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 268, + 83, + 343, + 158 + ], + "blocks": [ + { + "bbox": [ + 288, + 72, + 323, + 80 + ], + "lines": [ + { + "bbox": [ + 288, + 72, + 323, + 80 + ], + "spans": [ + { + "bbox": [ + 288, + 72, + 323, + 80 + ], + "type": "text", + "content": "FID 39.92" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 268, + 83, + 343, + 158 + ], + "lines": [ + { + "bbox": [ + 268, + 83, + 343, + 158 + ], + "spans": [ + { + "bbox": [ + 268, + 83, + 343, + 158 + ], + "type": "image", + "image_path": "2190ea3082bc14db5cb2fe545f5b2d786398f67499b4c8ac3ad8846da896977c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 356, + 83, + 431, + 157 + ], + "blocks": [ + { + "bbox": [ + 376, + 72, + 411, + 80 + ], + "lines": [ + { + "bbox": [ + 376, + 72, + 411, + 80 + ], + "spans": [ + { + "bbox": [ + 376, + 72, + 411, + 80 + ], + "type": "text", + "content": "FID 51.66" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 83, + 431, + 157 + ], + "lines": [ + { + "bbox": [ + 356, + 83, + 431, + 157 + ], + "spans": [ + { + "bbox": [ + 356, + 83, + 431, + 157 + ], + "type": "image", + "image_path": "46fb38eb15da47f622e9bfbccd1b9f8919b1f85fb1f78e948884a5ae9d775742.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 443, + 83, + 517, + 157 + ], + "blocks": [ + { + "bbox": [ + 463, + 72, + 497, + 80 + ], + "lines": [ + { + "bbox": [ + 463, + 72, + 497, + 80 + ], + "spans": [ + { + "bbox": [ + 463, + 72, + 497, + 80 + ], + "type": "text", + "content": "FID 87.96" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 443, + 83, + 517, + 157 + ], + "lines": [ + { + "bbox": [ + 443, + 83, + 517, + 157 + ], + "spans": [ + { + "bbox": [ + 443, + 83, + 517, + 157 + ], + "type": "image", + "image_path": "33c564d0e6838947465d03631665a01db1cc92ab82aea250a0651d9b927a4061.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 93, + 178, + 168, + 253 + ], + "blocks": [ + { + "bbox": [ + 100, + 168, + 160, + 177 + ], + "lines": [ + { + "bbox": [ + 100, + 168, + 160, + 177 + ], + "spans": [ + { + "bbox": [ + 100, + 168, + 160, + 177 + ], + "type": "text", + "content": "FID 53.54 (-12.03)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 178, + 168, + 253 + ], + "lines": [ + { + "bbox": [ + 93, + 178, + 168, + 253 + ], + "spans": [ + { + "bbox": [ + 93, + 178, + 168, + 253 + ], + "type": "image", + "image_path": "5080debdd0e6835c5d637a65469e5f2416ce9fc9aef6605184e35a48dd657c7d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 181, + 178, + 255, + 252 + ], + "blocks": [ + { + "bbox": [ + 190, + 169, + 246, + 178 + ], + "lines": [ + { + "bbox": [ + 190, + 169, + 246, + 178 + ], + "spans": [ + { + "bbox": [ + 190, + 169, + 246, + 178 + ], + "type": "text", + "content": "FID 20.59 (-4.49)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 178, + 255, + 252 + ], + "lines": [ + { + "bbox": [ + 181, + 178, + 255, + 252 + ], + "spans": [ + { + "bbox": [ + 181, + 178, + 255, + 252 + ], + "type": "image", + "image_path": "730253ce8eb5570485746e38557d1fb36dd31cf0db7dfba5bd0d692aa238d74d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 268, + 178, + 344, + 255 + ], + "blocks": [ + { + "bbox": [ + 278, + 169, + 334, + 178 + ], + "lines": [ + { + "bbox": [ + 278, + 169, + 334, + 178 + ], + "spans": [ + { + "bbox": [ + 278, + 169, + 334, + 178 + ], + "type": "text", + "content": "FID 35.77 (-4.15)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 268, + 178, + 344, + 255 + ], + "lines": [ + { + "bbox": [ + 268, + 178, + 344, + 255 + ], + "spans": [ + { + "bbox": [ + 268, + 178, + 344, + 255 + ], + "type": "image", + "image_path": "76b4a2945e3ea29a2237edd468cce315b1f3441c12ed026311a7d514a86449f5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 555, + 297 + ], + "type": "text", + "content": "Figure 4. The compared generated results between StyleGAN2 and StyleGAN2 + QADDRS on low-shot datasets. Top: Images generated by StyleGAN2. Bottom: Images generated by StyleGAN2 + QADDRS. The reduction of FID in red color shows the improvement of StyleGAN2 + QADDRS over the baseline StyleGAN2." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 356, + 178, + 430, + 253 + ], + "blocks": [ + { + "bbox": [ + 365, + 169, + 422, + 178 + ], + "lines": [ + { + "bbox": [ + 365, + 169, + 422, + 178 + ], + "spans": [ + { + "bbox": [ + 365, + 169, + 422, + 178 + ], + "type": "text", + "content": "FID 47.87 (-3.79)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 178, + 430, + 253 + ], + "lines": [ + { + "bbox": [ + 356, + 178, + 430, + 253 + ], + "spans": [ + { + "bbox": [ + 356, + 178, + 430, + 253 + ], + "type": "image", + "image_path": "526620215b5210c465687773949aed9aab00809885a32e318e4bc623bb2cd7c0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 443, + 178, + 518, + 254 + ], + "blocks": [ + { + "bbox": [ + 451, + 169, + 511, + 178 + ], + "lines": [ + { + "bbox": [ + 451, + 169, + 511, + 178 + ], + "spans": [ + { + "bbox": [ + 451, + 169, + 511, + 178 + ], + "type": "text", + "content": "FID 73.29 (-14.67)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 443, + 178, + 518, + 254 + ], + "lines": [ + { + "bbox": [ + 443, + 178, + 518, + 254 + ], + "spans": [ + { + "bbox": [ + 443, + 178, + 518, + 254 + ], + "type": "image", + "image_path": "c8ce54616d2867846ca0ed8081d3f842689e8578b5363b0cfd31526990b6ff65.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 59, + 316, + 552, + 529 + ], + "blocks": [ + { + "bbox": [ + 59, + 316, + 552, + 529 + ], + "lines": [ + { + "bbox": [ + 59, + 316, + 552, + 529 + ], + "spans": [ + { + "bbox": [ + 59, + 316, + 552, + 529 + ], + "type": "table", + "html": "
MethodBackboneMAPre-trainingFFHQ
1001K2K5K
EDM [24]DiffusionYesNo79.10---
EDM + DA [24]DiffusionYesNo50.7330.7527.1724.51
Patch Diffusion [54]DiffusionYesNo44.4528.0325.3222.63
StyleGAN2 [23]GANNoNo179.00100.1654.3049.68
StyleGAN2 + Diff-Augment [66]GANYesNo61.9125.6624.3210.45
StyleGAN2 + ADA [22]GANYesNo85.821.2915.3910.96
APA [19]GANYesNo65.0018.8916.908.38
AugSelf-StyleGAN2+ [17]GANYesNo-20.39-9.15
FakeCLR [28]GANYesNo42.5615.929.907.25
Projected GAN [44]GANYesYes26.2511.128.256.85
InsGen [56]GANYesNo45.7518.2111.477.83
+ QADDRSGANYesNo43.6717.1610.737.51
Diffusion-Projected GAN (FastGAN) [55]GANYesYes25.478.767.996.59
+ QADDRSGANYesYes24.468.517.886.48
Diffusion-Projected GAN (FastGAN Lite) [55]GANYesYes20.657.025.474.74
+ QADDRSGANYesYes19.896.575.184.69
", + "image_path": "d3af9ef67d6107b79753cc4e80ef7227c149ddc41c949b6d74cfe36b5a8ca048.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 544, + 555, + 579 + ], + "lines": [ + { + "bbox": [ + 55, + 544, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 555, + 579 + ], + "type": "text", + "content": "Table 2. FID scores (lower is better) of different methods on the " + }, + { + "bbox": [ + 55, + 544, + 555, + 579 + ], + "type": "inline_equation", + "content": "{256} \\times {256}\\mathrm{{FFHQ}}" + }, + { + "bbox": [ + 55, + 544, + 555, + 579 + ], + "type": "text", + "content": " dataset. In line with FakeCLR [28], the experiments are performed on 100, 1K, 2K and 5K subsets of the FFHQ dataset. MA donates Massive Augmentation [9], i.e., xflipping. The results are averaged over five evaluations, with relative standard deviations below 1%." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 599, + 295, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 708 + ], + "type": "text", + "content": "tiveness of the QADDRS with both StyleGAN2 and FastGAN using five different commonly-used DE-GANs, i.e., StyleGAN2, StyleGAN2 + Diff-Augment, StyleGAN2 + ADA, InsGen + ADA and Diffusion-Projected GAN. As shown in Table 1, the proposed QADDRS can increase the performance of these five different DE-GANs, demonstrating that QADDRS can be regarded as a general plugin training strategy for DE-GANs. As shown in Table 7, Diffusion-Projected GAN (FastGAN) achieves better per" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 599, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 555, + 708 + ], + "type": "text", + "content": "formance than Diffusion-Projected GAN (FastGAN Lite), i.e., applying the lite version generator in FastGAN, on low-shot datasets. Therefore, we report the results of the Diffusion-Projected GAN (FastGAN) in Table 1. Although Diffusion-Projected GAN (FastGAN) has already utilized all widely-used approaches, i.e., DA, regularizations, noise injection and pre-trained models, to boost performance on the low-shot datasets, adding QADDRS can further improve its performance. In addition, following [1], we also report" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30687" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 541, + 127 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 541, + 127 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 541, + 127 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 541, + 127 + ], + "type": "table", + "html": "
MethodObamaGrumpy CatPandaAF-CatAF-Dog
PRPRPRPRPR
Diffusion-Projected GAN (FastGAN) + QADDRS0.9990.2400.9960.1490.9990.4100.9970.1820.9980.203
0.9990.2420.9980.1510.9990.4180.9980.1860.9990.203
", + "image_path": "afc20dd83b3c7eac1a3853366f18fa8aa67b70754222b05902808373297767c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 78, + 188, + 533, + 242 + ], + "blocks": [ + { + "bbox": [ + 55, + 134, + 555, + 167 + ], + "lines": [ + { + "bbox": [ + 55, + 134, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 555, + 167 + ], + "type": "text", + "content": "Table 3. The comparison of Precision (P) and Recall (R) (higher is better) with Diffusion-Projected GAN (FastGAN) [55] and Diffusion-Projected GAN (FastGAN) + QADDRS on the " + }, + { + "bbox": [ + 55, + 134, + 555, + 167 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 134, + 555, + 167 + ], + "type": "text", + "content": " low-shot datasets. The results are averaged over three evaluations, with relative standard deviations below " + }, + { + "bbox": [ + 55, + 134, + 555, + 167 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 134, + 555, + 167 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 188, + 533, + 242 + ], + "lines": [ + { + "bbox": [ + 78, + 188, + 533, + 242 + ], + "spans": [ + { + "bbox": [ + 78, + 188, + 533, + 242 + ], + "type": "table", + "html": "
MethodFFHQ-100FFHQ-1KFFHQ-2KFFHQ-5K
PRPRPRPR
Diffusion-Projected GAN (FastGAN Lite) + QADDRS0.7370.0180.7340.2690.7360.3210.7300.377
0.7390.0210.7340.2760.7390.3230.7330.379
", + "image_path": "b46c7c9e26c1e94d432c24c3c5789d1ed59d3602aeef248355507e5e1d00d655.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 312, + 299, + 365 + ], + "blocks": [ + { + "bbox": [ + 55, + 249, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 555, + 282 + ], + "type": "text", + "content": "Table 4. The comparison of Precision (P) and Recall (R) (higher is better) with Diffusion-Projected GAN (FastGAN Lite) [55] and Diffusion-Projected GAN (FastGAN Lite) + QADDRS on the " + }, + { + "bbox": [ + 55, + 249, + 555, + 282 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 249, + 555, + 282 + ], + "type": "text", + "content": " FFHQ dataset. The results are averaged over five evaluations, with relative standard deviations below " + }, + { + "bbox": [ + 55, + 249, + 555, + 282 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 249, + 555, + 282 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 312, + 299, + 365 + ], + "lines": [ + { + "bbox": [ + 56, + 312, + 299, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 299, + 365 + ], + "type": "table", + "html": "
Method100AF
OGCPCD
Diffusion Projected GAN (FastGAN)2.024.782.598.367.21
+ QADDRS1.984.712.538.207.09
", + "image_path": "f727e50ec87cec3fb32d60c4025fe074163735a1590c2863ea242f8563e892d7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 468, + 293, + 525 + ], + "blocks": [ + { + "bbox": [ + 55, + 373, + 296, + 439 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 439 + ], + "type": "text", + "content": "Table 5. Compared " + }, + { + "bbox": [ + 55, + 373, + 296, + 439 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 373, + 296, + 439 + ], + "type": "text", + "content": " scores between Diffusion Projected GAN (FastGAN) and Diffusion Projected GAN (FastGAN) + QADDRS on low-shot datasets. We use the FID calculation model to calculate the " + }, + { + "bbox": [ + 55, + 373, + 296, + 439 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 373, + 296, + 439 + ], + "type": "text", + "content": ". Here, AF denotes AnimalFace, O denotes Obama, GC denotes Grumpy Cat, P denotes Panda, C denotes Cat and D denotes Dog." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 468, + 293, + 525 + ], + "lines": [ + { + "bbox": [ + 58, + 468, + 293, + 525 + ], + "spans": [ + { + "bbox": [ + 58, + 468, + 293, + 525 + ], + "type": "table", + "html": "
MethodFFHQ
1001K2K5K
Diffusion Projected GAN (FL)23.9111.459.618.33
+ QADDRS23.3211.149.428.19
", + "image_path": "99968b0db0c30fa29ab6624447835d0ccdd1d2ea63e6db34adb68f0b17b28b65.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "the results of several SOTA diffusion models in Table 1. It is clear that GAN-based methods outperform diffusion-based methods on low-shot datasets. To further show the effectiveness of QADDRS, the compared Precision and Recall, and " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": " of adding QADDRS to Diffusion-Projected GAN (FastGAN) are shown in Tables 3 and 5, respectively. The compared generated images on StyleGAN2 and StyleGAN2 + QADDRS are shown in Figure 4." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 330, + 311, + 541, + 355 + ], + "blocks": [ + { + "bbox": [ + 55, + 533, + 296, + 578 + ], + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 578 + ], + "type": "text", + "content": "Table 6. Compared " + }, + { + "bbox": [ + 55, + 533, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 533, + 296, + 578 + ], + "type": "text", + "content": " scores between Diffusion Projected GAN (FastGAN Lite) and Diffusion Projected GAN (FastGAN Lite) + QADDRS on FFHQ datasets. We use the FID calculation model to calculate the " + }, + { + "bbox": [ + 55, + 533, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 533, + 296, + 578 + ], + "type": "text", + "content": ". FL denotes the FastGAN Lite." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 311, + 541, + 355 + ], + "lines": [ + { + "bbox": [ + 330, + 311, + 541, + 355 + ], + "spans": [ + { + "bbox": [ + 330, + 311, + 541, + 355 + ], + "type": "table", + "html": "
MethodFID
Diffusion Projected GAN (FastGAN)17.22
Diffusion Projected GAN (FastGAN Lite)20.91
", + "image_path": "c8bdf90f7cb1868ee2cdf4f5379d462664af73250dc00c4736d2bc5924f0e83e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 372, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 417 + ], + "type": "text", + "content": "Table 7. Compared FID scores of Diffusion Projected GAN (FastGAN) and Diffusion Projected GAN (FastGAN Lite) on " + }, + { + "bbox": [ + 313, + 372, + 555, + 417 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 313, + 372, + 555, + 417 + ], + "type": "text", + "content": " AnimalFace-dog dataset. The results are averaged over three evaluations, with relative standard deviations below " + }, + { + "bbox": [ + 313, + 372, + 555, + 417 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 372, + 555, + 417 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 441, + 457, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 457, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 457, + 454 + ], + "type": "text", + "content": "4.3. Results on FFHQ Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": "According to the results of five different DE-GANs on low-shot datasets, InsGen + ADA and Diffusion-Projected GAN achieve SOTA performance on StyleGAN2 and FastGAN, respectively. Therefore, we conduct experiments on the FFHQ dataset by selecting these two types of DEGANs. Following FakeCLR [28], we conduct experiments on the FFHQ 100, FFHQ 1K, FFHQ 2K and FFHQ 5K subsets. As shown in Table 2, QADDRS can improve performance with these two DE-GANs on the FFHQ dataset. More importantly, we report the results of both the Diffusion-Projected GAN (FastGAN) and the Diffusion-Projected GAN (FastGAN Lite) in Table 2. It is clear that Diffusion-Projected GAN (FastGAN Lite) achieves better performance than Diffusion-Projected GAN (FastGAN) on the FFHQ dataset. Additionally, the comparison of Precision and Recall, as well as " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " with Diffusion-Projected GAN (FastGAN Lite) and Diffusion-Projected GAN (FastGAN Lite) + QADDRS are shown in Tables 4 and 6, respectively. Diffusion-Projected GAN (FastGAN Lite) + QADDRS outperforms Diffusion-Projected GAN (FastGAN Lite) by adding QADDRS, which demonstrates the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30688" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 286, + 96 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 286, + 96 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 286, + 96 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 286, + 96 + ], + "type": "table", + "html": "
λ0.000.250.500.75
FID46.8741.2742.4043.32
", + "image_path": "58d4291dedc611acc2652b3cdce33ebab699456fba537ac0b61d4f2228e5c6f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 171, + 286, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 115, + 295, + 160 + ], + "lines": [ + { + "bbox": [ + 55, + 115, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 115, + 295, + 160 + ], + "type": "text", + "content": "Table 8. FID scores on the " + }, + { + "bbox": [ + 55, + 115, + 295, + 160 + ], + "type": "inline_equation", + "content": "{256} \\times {256100}" + }, + { + "bbox": [ + 55, + 115, + 295, + 160 + ], + "type": "text", + "content": " -shot Obama dataset with StyleGAN2 + Diff-Augment. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below 1%." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 171, + 286, + 198 + ], + "lines": [ + { + "bbox": [ + 65, + 171, + 286, + 198 + ], + "spans": [ + { + "bbox": [ + 65, + 171, + 286, + 198 + ], + "type": "table", + "html": "
λ0.000.250.500.75
FID58.8551.8353.3154.45
", + "image_path": "1ddb58cfaf9f65aafe10040bb4b6589b8db358d859dde04460e16d207a235723.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 74, + 272, + 278, + 324 + ], + "blocks": [ + { + "bbox": [ + 55, + 217, + 295, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 261 + ], + "type": "text", + "content": "Table 9. FID scores on the " + }, + { + "bbox": [ + 55, + 217, + 295, + 261 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 55, + 217, + 295, + 261 + ], + "type": "text", + "content": " AnimalFace-Dog dataset with StyleGAN2 + Diff-Augment. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below 1%." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 272, + 278, + 324 + ], + "lines": [ + { + "bbox": [ + 74, + 272, + 278, + 324 + ], + "spans": [ + { + "bbox": [ + 74, + 272, + 278, + 324 + ], + "type": "table", + "html": "
MethodObamaAF-Dog
StyleGAN2 + Diff-Augment46.8758.85
+ Fixed rejecting number N44.1254.85
+ QADDRS41.2751.83
", + "image_path": "9b2c118ad3edc4a7fcc36007ea73505eacaf18c5fa1379b8eb87ab8ac0346006.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 408, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 295, + 445 + ], + "type": "text", + "content": "effectiveness of the proposed QADDRS. To further show the superiority of QADDRS, more generated images on the FFHQ dataset can be found in supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 452, + 156, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 156, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 156, + 464 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 471, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 554 + ], + "type": "text", + "content": "Effectiveness of the hyperparameter " + }, + { + "bbox": [ + 55, + 471, + 295, + 554 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 471, + 295, + 554 + ], + "type": "text", + "content": " in proposed QADDRS. To demonstrate the effectiveness of the hyperparameter " + }, + { + "bbox": [ + 55, + 471, + 295, + 554 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 471, + 295, + 554 + ], + "type": "text", + "content": " in the proposed QADDRS, we conduct experiments by selecting different values of " + }, + { + "bbox": [ + 55, + 471, + 295, + 554 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 471, + 295, + 554 + ], + "type": "text", + "content": " on the 100-shot Obama and AnimalFace-Dog datasets using StyleGAN2 + Diff-Augment. The results, shown in Tables 8 and 9, indicate that setting " + }, + { + "bbox": [ + 55, + 471, + 295, + 554 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 471, + 295, + 554 + ], + "type": "text", + "content": " to 0.25 achieves better performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 555, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 662 + ], + "type": "text", + "content": "Effectiveness of the dynamic rejection number " + }, + { + "bbox": [ + 55, + 555, + 295, + 662 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 555, + 295, + 662 + ], + "type": "text", + "content": " in the proposed QADDRS. To show that the dynamic rejection number " + }, + { + "bbox": [ + 55, + 555, + 295, + 662 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 555, + 295, + 662 + ], + "type": "text", + "content": " in QADDRS is effective, we conduct an ablation study by comparing QADDRS with a fixed rejection number " + }, + { + "bbox": [ + 55, + 555, + 295, + 662 + ], + "type": "inline_equation", + "content": "N = 0.25 \\times" + }, + { + "bbox": [ + 55, + 555, + 295, + 662 + ], + "type": "text", + "content": " batchsize on the 100-shot Obama dataset with StyleGAN2 + Diff-Augment. The results are shown in Table 10. It is obvious that the dynamically determined rejection number " + }, + { + "bbox": [ + 55, + 555, + 295, + 662 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 555, + 295, + 662 + ], + "type": "text", + "content": " can achieve better performance compared with the fixed rejection number." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 670, + 246, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 246, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 246, + 683 + ], + "type": "text", + "content": "4.5. Comparisons of DRS and QADDRS" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "We conduct the experiments by comparing QADDRS and the existing DRS method [5] using StyleGAN2 + Diff-" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 318, + 69, + 550, + 118 + ], + "blocks": [ + { + "bbox": [ + 55, + 342, + 295, + 386 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 386 + ], + "type": "text", + "content": "Table 10. FID scores on the " + }, + { + "bbox": [ + 55, + 342, + 295, + 386 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 342, + 295, + 386 + ], + "type": "text", + "content": " 100-shot Obama and AnimalFace-Dog datasets. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below " + }, + { + "bbox": [ + 55, + 342, + 295, + 386 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 342, + 295, + 386 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 69, + 550, + 118 + ], + "lines": [ + { + "bbox": [ + 318, + 69, + 550, + 118 + ], + "spans": [ + { + "bbox": [ + 318, + 69, + 550, + 118 + ], + "type": "table", + "html": "
MethodObamaAnimalFace-Dog
StyleGAN2 + Diff-Augment46.8758.85
+ DRS [5]45.2256.17
+ QADDRS41.2751.83
", + "image_path": "30995c6765630bda2920e6f8b6b2248bfc1866b38afdf9260ae41dd876602275.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 315, + 192, + 560, + 230 + ], + "blocks": [ + { + "bbox": [ + 313, + 137, + 555, + 181 + ], + "lines": [ + { + "bbox": [ + 313, + 137, + 555, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 555, + 181 + ], + "type": "text", + "content": "Table 11. FID scores on the " + }, + { + "bbox": [ + 313, + 137, + 555, + 181 + ], + "type": "inline_equation", + "content": "{256} \\times {256}\\;{100}" + }, + { + "bbox": [ + 313, + 137, + 555, + 181 + ], + "type": "text", + "content": " -shot Obama and Animal-Face Dog datasets. Massive Augmentation [9], i.e., xflipping, is applied to all methods. The results are averaged over three evaluations, with relative standard deviations below 1%." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 192, + 560, + 230 + ], + "lines": [ + { + "bbox": [ + 315, + 192, + 560, + 230 + ], + "spans": [ + { + "bbox": [ + 315, + 192, + 560, + 230 + ], + "type": "table", + "html": "
MethodSeconds / 1K images
Diffusion-Projected GAN (FastGAN) + QADDRS17\n17.25
", + "image_path": "67477d438ce87671e978de52c35d014fbcd436b385c2ea8c372e928e7babff7e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 248, + 555, + 303 + ], + "lines": [ + { + "bbox": [ + 313, + 248, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 555, + 303 + ], + "type": "text", + "content": "Table 12. The training time on the " + }, + { + "bbox": [ + 313, + 248, + 555, + 303 + ], + "type": "inline_equation", + "content": "{256} \\times {256}\\;{100}" + }, + { + "bbox": [ + 313, + 248, + 555, + 303 + ], + "type": "text", + "content": " -shot Obama dataset with or without QADDRS using Diffusion-Projected GAN (FastGAN). The results are averaged over ten evaluations on two NVIDIA RTX A5000 GPUs with a batch size of 64. All standard deviations are below 1% relative to the mean." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 323, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 491 + ], + "type": "text", + "content": "Augment on the 100-shot Obama and AnimalFace-dog datasets. The results are shown in Table 11. Based on ADA [22], the overfitting of " + }, + { + "bbox": [ + 313, + 323, + 555, + 491 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 323, + 555, + 491 + ], + "type": "text", + "content": " during training finally results in the DE-GANs generating low-quality images under limited data settings. Although the existing DRS method can reject lower-quality generated samples during post-training, they can not benefit the training of DE-GANs. Thus, existing DRS methods can not prevent generating low-quality images in average, leaving limited gains under limited data settings. In contrast, QADDRS can alleviate the overfitting issue during training, thereby effectively avoiding producing low-quality images in average and achieving significant improvements. As a result, QADDRS outperforms the existing DRS method for DE-GANs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 432, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 432, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 432, + 510 + ], + "type": "text", + "content": "4.6. Computational Cost" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "type": "text", + "content": "The results of the training time on the 100-shot Obama dataset with or without QADDRS using Diffusion-Projected GAN (FastGAN) have been shown in Table 12. The proposed QADDRS results in a negligible increase in computational cost." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 586, + 388, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 388, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 388, + 597 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel method called quality aware dynamic discriminator rejection sampling (QADDRS), to improve DE-GANs training by considering sample quality. QADDRS can effectively avoid " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " becoming overly confident in distinguishing both real and fake samples, thereby alleviating the overfitting of " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " issue. Extensive experiments on multiple datasets show that applying QADDRS to various DE-GANs can achieve better performance and deliver state-of-the-art results." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30689" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 162 + ], + "type": "text", + "content": "This work is partially supported by the High Performance Computing center of Queen's University Belfast and the Kelvin-2 supercomputer (EPSRC grant EP/T022175/1). First author is supported by a scholarship of the Queen's University Belfast and the Department for the Economy, Northern Ireland." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 192, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 192, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 192, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 192, + 295, + 224 + ], + "type": "text", + "content": "[1] Mehran Aghabozorgi, Shichong Peng, and Ke Li. Adaptive imle for few-shot pretraining-free generative modelling. In ICML, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 226, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 226, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 226, + 295, + 258 + ], + "type": "text", + "content": "[2] Abdul Fatir Ansari, Ming Liang Ang, and Harold Soh. Refining deep generative models via discriminator gradient flow. arXiv preprint arXiv:2012.00780, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 259, + 295, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 259, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 259, + 295, + 291 + ], + "type": "text", + "content": "[3] Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks. In ICLR, 2017. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 293, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 294, + 323 + ], + "type": "text", + "content": "[4] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In ICML, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 326, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 326, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 326, + 295, + 358 + ], + "type": "text", + "content": "[5] Samaneh Azadi, Catherine Olsson, Trevor Darrell, Ian Goodfellow, and Augustus Odena. Discriminator rejection sampling. In ICLR, 2019. 1, 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 360, + 295, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 295, + 391 + ], + "type": "text", + "content": "[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 392, + 295, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 436 + ], + "type": "text", + "content": "[7] Tong Che, Ruixiang Zhang, Jascha Sohl-Dickstein, Hugo Larochelle, Liam Paull, Yuan Cao, and Yoshua Bengio. Your gan is secretly an energy-based model and you should use discriminator driven latent sampling. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 437, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 295, + 479 + ], + "type": "text", + "content": "[8] Tianlong Chen, Yu Cheng, Zhe Gan, Jingjing Liu, and Zhangyang Wang. Data-efficient GAN training beyond (just) augmentations: A lottery ticket perspective. In NeurIPS, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 481, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 295, + 524 + ], + "type": "text", + "content": "[9] Kaiwen Cui, Jiaxing Huang, Zhipeng Luo, Gongjie Zhang, Fangneng Zhan, and Shijian Lu. Genco: Generative constraining for generative adversarial networks with limited data. In AAAI, 2022. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 525, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 557 + ], + "type": "text", + "content": "[10] Emily L Denton, Soumith Chintala, Rob Fergus, et al. Deep generative image models using a laplacian pyramid of adversarial networks. In NeurIPS, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 559, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 295, + 590 + ], + "type": "text", + "content": "[11] Tiantian Fang, Ruoyu Sun, and Alex Schwing. Diggan: Discriminator gradient gap regularization for GAN training with limited data. In NeurlPS, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "text", + "content": "[13] Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Scholkopf, and Alexander Smola. A kernel two-sample test. The Journal of Machine Learning Research, 13(1):723-773, 2012. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "text", + "content": "[14] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein GANs. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Tianyu Guo, Chang Xu, Jiajun Huang, Yunhe Wang, Boxin Shi, Chao Xu, and Dacheng Tao. On positive-unlabeled classification in GAN. In CVPR, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "text", + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 205 + ], + "type": "text", + "content": "[17] Liang Hou, Qi Cao, Yige Yuan, Songtao Zhao, Chongyang Ma, Siyuan Pan, Pengfei Wan, Zhongyuan Wang, Huawei Shen, and Xueqi Cheng. Augmentation-aware self-supervision for data-efficient gan training. In NeurIPS, 2023. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 208, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 239 + ], + "type": "text", + "content": "[18] Thibaut Issenhuth, Ugo Tanielian, David Picard, and Jeremie Mary. Latent reweighting, an almost free improvement for gans. In WACV, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[19] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Deceive d: Adaptive pseudo augmentation for GAN training with limited data. In NeurIPS, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 276, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 307 + ], + "type": "text", + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In ICLR, 2018. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "type": "text", + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "type": "text", + "content": "[22] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, 2020. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of styleGAN. In CVPR, 2020. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "text", + "content": "[24] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "text", + "content": "[25] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for GAN training. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 489, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 521 + ], + "type": "text", + "content": "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In NeurIPS, 2019. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[27] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fr\\`echet inception distance. In ICLR, 2023. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "text", + "content": "[28] Ziqiang Li, Chaoyue Wang, Heliang Zheng, Jing Zhang, and Bin Li. Fakeclr: Exploring contrastive learning for solving latent discontinuity in data-efficient gans. In ECCV, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 315, + 602, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 602, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 315, + 602, + 553, + 624 + ], + "type": "text", + "content": "[29] Jae Hyun Lim and Jong Chul Ye. Geometric GAN. arXiv preprint arXiv:1705.02894, 2017. 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 625, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 666 + ], + "type": "text", + "content": "[30] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and Ahmed Elgammal. Towards faster and stabilized GAN training for high-fidelity few-shot image synthesis. In ICLR, 2021. 2, 4, 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[31] Haozhe Liu, Wentian Zhang, Bing Li, Haoqian Wu, Nanjun He, Yawen Huang, Yuexiang Li, Bernard Ghanem, and Yefeng Zheng. Adaptivemix: Improving GAN training via feature space shrinkage. In CVPR, 2023. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30690" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[32] Steven Liu, Tongzhou Wang, David Bau, Jun-Yan Zhu, and Antonio Torralba. Diverse image generation via self-conditioned GANs. In CVPR, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[33] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In ICCV, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 141, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 141, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 57, + 141, + 294, + 163 + ], + "type": "text", + "content": "[34] Takeru Miyato and Masanori Koyama. cGANs with projection discriminator. In ICLR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 165, + 294, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 294, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 294, + 198 + ], + "type": "text", + "content": "[35] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 199, + 295, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 199, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 57, + 199, + 295, + 232 + ], + "type": "text", + "content": "[36] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze the discriminator: a simple baseline for fine-tuning GANs. arXiv preprint arXiv:2002.10964, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 234, + 294, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 294, + 267 + ], + "type": "text", + "content": "[37] Byeonghu Na, Yeongmin Kim, Minsang Park, Donghyeok Shin, Wanmo Kang, and Il-Chul Moon. Diffusion rejection sampling. In ICML, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 269, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 269, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 269, + 294, + 300 + ], + "type": "text", + "content": "[38] Yao Ni and Piotr Koniusz. Nice: Noise-modulated consistency regularization for data-efficient gans. In NeurIPS, 2023. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 302, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 302, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 302, + 294, + 335 + ], + "type": "text", + "content": "[39] Yao Ni and Piotr Koniusz. Chain: Enhancing generalization in data-efficient gans via lipschitz continuity constrained normalization. In CVPR, 2024. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 337, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 294, + 369 + ], + "type": "text", + "content": "[40] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In ICCV, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "type": "text", + "content": "[41] Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In ICML, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 405, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 294, + 437 + ], + "type": "text", + "content": "[42] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. In ICLR, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 440, + 294, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 440, + 294, + 473 + ], + "spans": [ + { + "bbox": [ + 56, + 440, + 294, + 473 + ], + "type": "text", + "content": "[43] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. In NeurIPS, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 475, + 294, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 506 + ], + "type": "text", + "content": "[44] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected GANs converge faster. In NeurIPS, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "type": "text", + "content": "[45] Masashi Sugiyama, Matthias Krauledat, and Klaus-Robert Müller. Covariate shift adaptation by importance weighted cross validation. Journal of Machine Learning Research, 8 (5), 2007. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 554, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 294, + 575 + ], + "type": "text", + "content": "[46] Akinori Tanaka. Discriminator optimal transport. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 577, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 294, + 610 + ], + "type": "text", + "content": "[47] Dustin Tran, Rajesh Ranganath, and David M Blei. Deep and hierarchical implicit models. arXiv preprint arXiv:1702.08896, 7(3):13, 2017. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "type": "text", + "content": "[48] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In CVPR, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 646, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 678 + ], + "type": "text", + "content": "[49] Ryan Turner, Jane Hung, Eric Frank, Yunus Saatchi, and Jason Yosinski. Metropolis-hastings generative adversarial networks. In ICML, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[50] Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. In NeurIPS, 2016. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 702 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[51] Chirag Vashist, Shichong Peng, and Ke Li. Rejection sampling imle: Designing priors for better few-shot image synthesis. In ECCV, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "text", + "content": "[52] Yaxing Wang, Chenshen Wu, Luis Herranz, Joost van de Weijer, Abel Gonzalez-Garcia, and Bogdan Raducanu. Transferring GANs: generating images from limited data. In ECCV, 2018. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "type": "text", + "content": "[53] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, Luis Herranz, Fahad Shahbaz Khan, and Joost van de Weijer. MineGAN: effective knowledge transfer from GANs to target domains with few images. In CVPR, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[54] Zhendong Wang, Yifan Jiang, Huangjie Zheng, Peihao Wang, Pengcheng He, Zhangyang Wang, Weizhu Chen, and Mingyuan Zhou. Patch diffusion: Faster and more data-efficient training of diffusion models. In NeurIPS, 2023. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "text", + "content": "[55] Zhendong Wang, Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Diffusion-gan: Training gans with diffusion. In ICLR, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[56] Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Data-efficient instance generation from instance discrimination. In NeurIPS, 2021. 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 320, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 353 + ], + "type": "text", + "content": "[57] Ceyuan Yang, Yujun Shen, Yinghao Xu, Deli Zhao, Bo Dai, and Bolei Zhou. Improving GANs with a dynamic discriminator. In NeurlPS, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[58] Mengping Yang, Zhe Wang, Ziqiu Chi, and Yanbing Zhang. Fegan: Exploiting frequency components for training gans under limited data. In NeurIPS, 2022. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[59] Mengping Yang, Zhe Wang, Wenyi Feng, Qian Zhang, and Ting Xiao. Improving few-shot image generation by structural discrimination and textural modulation. In ACM MM, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "text", + "content": "[60] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In ICCV, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "text", + "content": "[61] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In ICML, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 511, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 544 + ], + "type": "text", + "content": "[62] Zhaoyu Zhang, Mengyan Li, and Jun Yu. On the convergence and mode collapse of GAN. In SIGGRAPH Asia 2018 Technical Briefs, 2018. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "text", + "content": "[63] Zhaoyu Zhang, Yang Hua, Guanxiong Sun, Hui Wang, and Seán McLoone. Improving the training of the gans with limited data via dual adaptive noise injection. In ACMMM, 2024. 1, 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "text", + "content": "[64] Zhaoyu Zhang, Yang Hua, Guanxiong Sun, Hui Wang, and Seán McLoone. Improving the leaking of augmentations in data-efficient gans via adaptive negative data augmentation. In WACV, 2024. 1, 3, 4, 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "text", + "content": "[65] Zhaoyu Zhang, Yang Hua, Hui Wang, and Seán McLoone. Improving the fairness of the min-max game in GANs training. In WACV, 2024. 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 669, + 553, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 702 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 702 + ], + "type": "text", + "content": "[66] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient GAN training. In NeurIPS, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30691" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/aa0474ee-612f-4595-88a1-c6ba9ec15b27_content_list.json b/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/aa0474ee-612f-4595-88a1-c6ba9ec15b27_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..390c905e3f8a378315a420abac7f7ccd9875e02b --- /dev/null +++ b/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/aa0474ee-612f-4595-88a1-c6ba9ec15b27_content_list.json @@ -0,0 +1,1856 @@ +[ + { + "type": "text", + "text": "Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation", + "text_level": 1, + "bbox": [ + 101, + 128, + 898, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fengfan Zhou $^{1}$ , Bangjie Yin $^{2}$ , Hefei Ling $^{1*}$ , Qianyu Zhou $^{3}$ , Wenxuan Wang $^{4}$", + "bbox": [ + 196, + 202, + 799, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Technology, Huazhong University of Science and Technology; $^{2}$ Shanghai Shizhuang Information Technology Co., Ltd;", + "bbox": [ + 117, + 220, + 880, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Computer Science and Engineering, Shanghai Jiao Tong University; $^{4}$ School of Computer Science, Northwestern Polytechnical University.", + "bbox": [ + 163, + 256, + 831, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ {ffzhou, lhefei}@hust.edu.cn, $^{2}$ jamesyin10@gmail.com,", + "bbox": [ + 271, + 292, + 725, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 zhouqianyu@sjtu.edu.cn, 4wxwang@nwpu.edu.cn", + "bbox": [ + 300, + 309, + 681, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 361, + 326, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face Recognition (FR) models are vulnerable to adversarial examples that subtly manipulate benign face images, underscoring the urgent need to improve the transferability of adversarial attacks in order to expose the blind spots of these systems. Existing adversarial attack methods often overlook the potential benefits of augmenting the surrogate model with diverse initializations, which limits the transferability of the generated adversarial examples. To address this gap, we propose a novel method called Diverse Parameters Augmentation (DPA) attack method, which enhances surrogate models by incorporating diverse parameter initializations, resulting in a broader and more diverse set of surrogate models. Specifically, DPA consists of two key stages: Diverse Parameters Optimization (DPO) and Hard Model Aggregation (HMA). In the DPO stage, we initialize the parameters of the surrogate model using both pre-trained and random parameters. Subsequently, we save the models in the intermediate training process to obtain a diverse set of surrogate models. During the HMA stage, we enhance the feature maps of the diversified surrogate models by incorporating beneficial perturbations, thereby further improving the transferability. Experimental results demonstrate that our proposed attack method can effectively enhance the transferability of the crafted adversarial face examples.", + "bbox": [ + 89, + 393, + 485, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 771, + 223, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Owing to the relentless progress in deep learning, Face Recognition (FR) has achieved substantial advancements [1, 2, 8, 20, 34, 44]. However, the inherent vulnerabilities in existing FR models pose significant security risks [7, 31, 53, 74-77], with adversarial attacks emerging as one", + "bbox": [ + 89, + 797, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c47edcb32e1651a36b4d9a872f7d0104cd56232907865e641704ce2e3c7d03a5.jpg", + "image_caption": [ + "Pre-trained Parameters" + ], + "image_footnote": [], + "bbox": [ + 535, + 364, + 575, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9cb077fee36cf08dad4b10a9a539c89f900eb48b4939f09f4bf6f099d310babe.jpg", + "image_caption": [ + "Randomly Initialized Parameters" + ], + "image_footnote": [], + "bbox": [ + 697, + 364, + 736, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f55d9096223cfc97a092623c4d7879f0699f3b92d3bf6a0bf488615dd7d3faed.jpg", + "image_caption": [ + "(a) Input-based Augmentation" + ], + "image_footnote": [], + "bbox": [ + 532, + 392, + 596, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3cb750f5feb3254ee84e78fb0f946d9be731f090879f34e46d2c40945800c4d.jpg", + "image_caption": [ + "(b) Feature-based Augmentation" + ], + "image_footnote": [], + "bbox": [ + 630, + 392, + 694, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e0e9fe7450a975fe6a6ca4c145b3444df2f280e33fc694b7b0f7c7b991c33268.jpg", + "image_caption": [ + "(c) Parameter-based Augmentation" + ], + "image_footnote": [], + "bbox": [ + 730, + 392, + 790, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9cdc84cef9e805b9334097ac1a1b3127cede10b57bcec2ac9acf2a56a5e3d614.jpg", + "image_caption": [ + "Figure 1. Top: comparison between traditional augmentation-based adversarial attack methods and our proposed method. The black pattern filling on the left and right sides of the blue line represents input-based and parameter-based augmentation, respectively. The orange pattern filling indicates feature-based augmentation. Bottom: comparison of performance among 4 types of augmentations." + ], + "image_footnote": [], + "bbox": [ + 529, + 474, + 643, + 560 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/68b5289689ff9abd4610bb641df32913a4a9fe22d441a4413b70160745f92192.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 469, + 771, + 560 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/951477ee640f009ee37f96b5390bfe8ef40e7f2135e45af4812f55966cee0f5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 474, + 895, + 560 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of the critical concerns. Therefore, there is an urgent need to bolster the resilience of FR models against adversarial face examples to reveal and address the vulnerabilities.", + "bbox": [ + 511, + 684, + 906, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a result, numerous research efforts have been focused on this area. Several adversarial attacks have been developed to create adversarial face examples with features like stealthiness [6, 15, 33, 35, 61], transferability [22, 69-71], and the capacity for physical attacks [21, 62, 63]. These initiatives are aimed at improving the effectiveness of adversarial attacks on FR. Nevertheless, the transferability of these adversarial attacks remains limited. To enhance the transferability of adversarial attacks, augmentation emerges as one of the most effective methods. As illustrated in Figure 1, augmentation-based adversarial attack consists of three aug", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 107, + 886, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3516", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mentation types: input-based [27, 46, 52, 59], feature-based [69, 71], and parameter-based [12, 57].", + "bbox": [ + 89, + 90, + 485, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most augmentation-based adversarial attack methods focus on input-based augmentation to improve transferability. Previous research highlights a symmetry between surrogate models used in adversarial attacks and input data in training tasks [9, 24, 71]. In training, data augmentation has proven effective in enhancing model generalization. Drawing on this symmetry, augmenting models for crafting adversarial examples can yield examples with greater transferability. Input-based augmentation can be seen as methods that only augment models at the input layer, demonstrating significant effectiveness in enhancing transferability. However, augmenting surrogate models in deeper layers (i.e., feature-based and parameter-based augmentation) offers a more direct form of augmentation. Nevertheless, few studies explore deep-layer augmentation [12, 57, 71]. As a typical parameter-based augmentation method in deep layers, LGV [12] uses a pre-trained surrogate model and collects multiple parameter sets through additional training epochs with a high, constant learning rate, thereby enhancing the transferability of adversarial examples. Although these parameter-based augmentation methods [12, 57] show promising effectiveness, they face two problems: (1) Static surrogate model initializations: these methods solely augment the surrogate model from pre-trained parameters, limiting the parameter diversity of the surrogate models and thereby hindering the transferability of the crafted adversarial examples. (2) Unavailability of the FR head: modern FR training procedures typically involve training both a backbone model and a head model [8, 44]. Following training, inference is conducted solely with the backbone models, while the head models are often not released as open-source. Consequently, in the majority of instances, we are unable to access the pre-trained parameters for the head models. Traditional parameter-based augmentation adversarial attacks augment models from pretrained parameters [12, 57]. This limitation makes it challenging to craft adversarial examples on FR models with numerous open-sourced FR models.", + "bbox": [ + 91, + 123, + 485, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these problems, we introduce a novel adversarial attack called Diverse Parameters Augmentation (DPA) to enhance the transferability of crafted adversarial face examples. Unlike existing parameter-based augmentation adversarial attack methods that overlook the use of diverse parameter initializations [12, 57], we diversify parameter initializations with both random and pre-trained values, thereby improving black-box attack capacity.", + "bbox": [ + 89, + 686, + 485, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Technically, our proposed attack method comprises two stages: Diverse Parameters Optimization (DPO) and Hard Model Aggregation (HMA). During the DPO stage, we initiate a subset of the optimization parameters with random noise, preserving the refined parameters in the intermediate training process to yield a diverse set of surrogate models,", + "bbox": [ + 89, + 810, + 485, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which is instrumental in bolstering the transferability of the adversarial examples. In the HMA stage, we add beneficial perturbations [56] with optimization directions opposite to those of the adversarial perturbations onto the feature maps of the parameter-augmented surrogate models, achieving the effect of hard model augmentation [71] to further enhances the transferability of the crafted adversarial examples. Our proposed attack method effectively addresses the challenges posed by the lack of diverse parameters and the absence of the FR head. The comparison between traditional adversarial attack methods and our proposed method is illustrated in Figure 1. By using diverse initializations for surrogate models, we can expand the parameter set of these models, thereby enhancing the transferability of the adversarial examples.", + "bbox": [ + 511, + 90, + 906, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 532, + 303, + 872, + 316 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel perspective that parameter-based augmentation adversarial attacks should augment the parameters of surrogate models by incorporating diverse initialized parameters. To the best of our knowledge, this is the first adversarial attack on FR that utilizes parameter augmentation to enhance transferability.", + "- We introduce a new adversarial attack method on FR, called DPA, which comprises DPO and HMA stages. The DPO stage uses both pre-trained and random initializations to optimize the model and save the models in the intermediate training process to diversify the surrogate model set. The HMA stage adds beneficial perturbations with optimization directions opposite to those of adversarial perturbations onto the feature maps of the diversified surrogate models to further enhance transferability.", + "- Extensive experiments reveal that our proposed method attains superior performance when compared with the state-of-the-art adversarial attack methods." + ], + "bbox": [ + 511, + 316, + 908, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 603, + 655, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial Attacks. The primary objective of adversarial attacks is to introduce subtle perturbations into benign images, thereby deceiving machine learning systems and inducing them to produce incorrect predictions [11, 42]. The presence of adversarial examples constitutes a substantial security threat to contemporary machine learning systems. Consequently, significant research efforts have been expended to investigate adversarial attacks, with the aim of bolstering system robustness [10, 23, 27, 28, 30, 36, 66, 78]. To enhance the potency of black-box adversarial attacks, DI [59] incorporates random transformations into adversarial examples at each iteration, effectively achieving data augmentation. VMI-FGSM [50] harnesses gradient variance to stabilize the update process, thereby enhancing black-box attack performance. SSA [27] translates adversarial examples into the frequency domain and applies spectral manipulation for augmentation. SIA [52] introduces random image transformations on each image block, producing di", + "bbox": [ + 511, + 628, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3517", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4032ae9fb5acf39f50ad550aa33cbc7ed4053d3efee011085844182784053871.jpg", + "image_caption": [ + "Figure 2. The framework of the Diverse Parameters Optimization (DPO). We enhance the diversity of the surrogate model parameters by integrating both pre-trained and random initializations. The method yields a diverse set of surrogate model parameters, which enhances the parameter diversity of the surrogate FR models and consequently improves transferability of the crafted adversarial examples." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 890, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "verse variations for gradient estimation. BSR [46] segments the input image into multiple blocks, randomly shuffling and rotating them to generate a set of new images for gradient computation. Despite these advancements, these methods overlook the potential benefit of augmenting the surrogate model with diverse initializations, which limits the transferability of the crafted adversarial examples. In contrast, our proposed method augments surrogate models using both pre-trained and randomly initialized parameters, producing a broader and more diverse set of surrogate models.", + "bbox": [ + 88, + 440, + 483, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Attacks on Face Recognition. Adversarial attacks on FR models can be classified based on the constraints imposed on the adversarial perturbations. These attacks are broadly categorized into two types: restricted attacks [4, 9, 25, 26, 28, 60, 72, 79] and unrestricted attacks [3, 5, 37, 40, 43, 54, 55, 64]. Restricted attacks create adversarial examples that adhere to a specified perturbation limit, such as an $L_{p}$ norm bound. Our proposed attack method is a restricted attack. Consequently, we will delve into the specifics of restricted attacks in detail. To enhance the transferability of adversarial attacks on FR, Zhong and Deng [69] introduced DFANet, which employs dropout on the feature maps of convolutional layers to achieve an ensemble-like effect. Zhou et al. [71] proposed BPFA, enhancing attack transferability by integrating beneficial perturbations [56] onto the feature maps of FR models, resulting in the effect of hard model augmentation. Li et al. [22] leveraged additional information from FR-related tasks and applied a multi-task optimization framework to further improve the transferability of adversarial examples. Unrestricted adversarial attacks", + "bbox": [ + 91, + 598, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "conversely generate adversarial examples without the limitations of a predefined perturbation bound. These attacks are primarily focused on physical attacks [21, 58, 62], attribute editing [18, 33], and adversarial example generation through makeup transfer [15, 35, 39, 63]. Both restricted and unrestricted adversarial attacks on FR models have substantially advanced the capabilities of these attacks. However, existing methods for generating adversarial attacks on FR often rely on surrogate models with fixed parameters, which limits the transferability of the crafted adversarial examples. In contrast, our proposed attack method addresses this limitation by diversifying the surrogate model parameters, using models with varying parameters across different epochs.", + "bbox": [ + 511, + 440, + 908, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 655, + 650, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation and Framework", + "text_level": 1, + "bbox": [ + 511, + 681, + 846, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. Let $\\mathcal{F}^{vct}(\\mathbf{x}') \\in \\mathbb{R}^r$ denote the FR model employed by the victim to extract the embedding from a face image $\\mathbf{x}'$ . We denote $\\mathbf{x}^s$ and $\\mathbf{x}^t$ as the source and target images, respectively. The objective of the adversarial attacks explored in our research is to manipulate $\\mathcal{F}^{vct}(\\mathbf{x})$ to misclassify the adversarial example $\\mathbf{x}^{adv}$ as the target image $\\mathbf{x}^t$ , while ensuring that $\\mathbf{x}^{adv}$ bears a close visual resemblance to $\\mathbf{x}^s$ . The objective of adversarial attack as delineated in this paper is as follows:", + "bbox": [ + 511, + 705, + 906, + 843 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {x} ^ {a d v} = \\underset {\\mathbf {x} ^ {a d v}} {\\arg \\min } \\left(\\mathcal {D} \\left(\\mathcal {F} ^ {v c t} \\left(\\mathbf {x} ^ {a d v}\\right), \\mathcal {F} ^ {v c t} \\left(\\mathbf {x} ^ {t}\\right)\\right)\\right) \\tag {1} \\\\ \\mathrm {s . t .} \\| \\mathbf {x} ^ {a d v} - \\mathbf {x} ^ {s} \\| _ {p} \\leqslant \\epsilon \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 854, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3518", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the symbol $\\mathcal{D}$ denotes a predefined distance metric employed for the optimization of adversarial face examples.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, the attacker typically cannot access the model owned by the victim. As a result, black-box attacks play a crucial role in adversarial attacks [32, 48, 49]. A widely adopted approach for black-box attacks involves using a surrogate model $\\mathcal{F}$ to generate adversarial examples and then transfer the crafted adversarial examples to the victim model [50, 71]. Consequently, the transferability is important and constitutes the key problem studied in this research.", + "bbox": [ + 89, + 121, + 483, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Framework Overview. To enhance the transferability of adversarial examples, we propose a novel method called Diverse Parameters Augmentation (DPA). DPA augments surrogate models by incorporating both pre-trained and randomly initialized parameters, resulting in a more diverse and expansive set of surrogate models. Specifically, DPA consists of two key stages: Diverse Parameters Optimization (DPO) and Hard Model Aggregation (HMA). In the DPO stage, we diversify the surrogate model parameters by combining pre-trained and random initializations, as illustrated in Figure 2. In the HMA stage, we apply beneficial perturbations to the feature maps of the diversified surrogate models and combine them to achieve a higher degree of augmentation, as shown in Figure 3. In the following, we will introduce our proposed DPA attack method in detail.", + "bbox": [ + 89, + 242, + 483, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Diverse Parameters Optimization", + "text_level": 1, + "bbox": [ + 89, + 477, + 383, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The parameter diversity of surrogate models is crucial for the transferability of crafted adversarial examples. Previous parameter-based augmentation adversarial attack methods solely initialize parameters with pre-trained values, thereby limiting the parameter diversity of surrogate models [12, 57]. In contrast, the key innovation of the stage of our proposed attack method is the diversification of initialized parameters using both pre-trained and randomly initialized parameters. Following initialization, we augment the parameters of the surrogate models using the parameters in the intermediate training process to diversify the surrogate models, thereby improving the transferability of the adversarial examples.", + "bbox": [ + 89, + 500, + 483, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\mathbf{x}$ be a batch of face images, $\\mathbf{w} \\in \\mathbb{R}^{s \\times r}$ be the parameters of the FR head, $b$ be the batch size of $\\mathbf{x}$ , and $s$ be the class number in the training dataset. To calculate the loss function, we should first calculate the cosine similarity matrix $\\mathbf{r} \\in \\mathbb{R}^{b \\times s}$ between $\\mathcal{F}(\\mathbf{x}) \\in \\mathbb{R}^{b \\times r}$ and the parameters of the FR head $\\mathbf{w}^{\\top}$ using the following formula:", + "bbox": [ + 89, + 680, + 483, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} = \\cos \\mathbf {a} = \\bar {\\mathbf {e}} \\bar {\\mathbf {w}} ^ {\\top} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 779, + 482, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\overline{\\mathbf{e}} = \\frac{\\mathcal{F}(\\mathbf{x})}{\\|\\mathcal{F}(\\mathbf{x})\\|}, \\overline{\\mathbf{w}} = \\frac{\\mathbf{w}}{\\|\\mathbf{w}\\|}$ and $\\mathbf{a}$ is the angle between $\\mathcal{F}(\\mathbf{x})$ and $\\mathbf{w}^{\\top}$ . After obtaining the cosine similarity matrix by Equation (2), we can get the sine similarity between $\\mathcal{F}(\\mathbf{x})$ and $\\mathbf{w}$ using the following equation:", + "bbox": [ + 89, + 806, + 483, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sin \\mathbf {a} = \\sqrt {1 . 0 - \\cos^ {2} \\mathbf {a}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 883, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using the cosine and sine similarity matrix, the formula to get the additive angular margin cosine similarity matrix can be expressed as:", + "bbox": [ + 511, + 90, + 906, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\cos (\\mathbf {a} + m) = \\cos \\mathbf {a} \\cos m - \\sin \\mathbf {a} \\sin m \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 146, + 906, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m$ is the margin value for increasing the discrimination of the FR model [8]. Let $\\mathbf{y} \\in \\mathbb{R}^b$ be the labels of the current batch. By utilizing $\\mathbf{y}$ , we can calculate its one-hot encoded matrix using the following formula:", + "bbox": [ + 511, + 171, + 906, + 232 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathbf {h}} = \\Psi (\\mathbf {y}) \\in \\mathbb {R} ^ {b \\times s} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 239, + 906, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Psi (\\cdot)$ is the one-hot encode operation. Next, we use the following function to calculate the output of the head:", + "bbox": [ + 511, + 267, + 905, + 297 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {q} = d \\left(\\hat {\\mathbf {h}} \\odot \\mathbf {p} + (1 - \\hat {\\mathbf {h}}) \\odot \\mathbf {r}\\right) \\\\ \\mathbf {p} = \\left\\{ \\begin{array}{l l} \\cos (\\mathbf {a} + m) & \\text {s . t .} \\quad \\mathbf {a} < \\pi - m, \\\\ \\cos \\mathbf {a} - m \\sin m & \\text {s . t .} \\quad \\mathbf {a} \\geqslant \\pi - m, \\end{array} \\right. \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 305, + 905, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d$ is a pre-defined scale factor. Using the output of the head, the formula to calculate the loss function can be expressed as the following:", + "bbox": [ + 511, + 375, + 906, + 420 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathbf {x}, \\mathbf {y}) = - \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\log \\frac {e ^ {\\mathbf {q} _ {i , \\mathbf {y} _ {i}}}}{\\sum_ {j = 1} ^ {s} e ^ {\\mathbf {q} _ {i , j}}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 429, + 906, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $c$ and $\\mathbf{v}$ be the number of the training epoch and the parameters of the FR backbone, respectively. Utilizing Equation (7), we can get the parameters set [8]:", + "bbox": [ + 511, + 476, + 906, + 522 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {i}, \\mathbf {w} _ {i} = \\varphi \\left(\\underset {\\mathbf {v}, \\mathbf {w}} {\\arg \\min } \\mathcal {L} \\left(\\mathbf {x}, \\mathbf {y}\\right), i\\right) \\quad i \\in \\{1, 2, \\ldots , c \\}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 530, + 890, + 564 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} _ {c} = \\left\\{\\mathbf {v} _ {1}, \\mathbf {v} _ {2}, \\dots , \\mathbf {v} _ {c - 1}, \\mathbf {v} _ {c} \\right\\} \\quad \\mathbf {W} _ {c} = \\left\\{\\mathbf {w} _ {1}, \\mathbf {w} _ {2}, \\dots , \\mathbf {w} _ {c - 1}, \\mathbf {w} _ {c} \\right\\} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 566, + 923, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varphi (\\cdot ,i)$ is the operation that obtain the $\\mathbf{v}$ and $\\mathbf{w}$ parameters in the $i$ th epoch. We denote $\\mathbf{v}_0$ and $\\mathbf{w}_0$ as the initialized parameters of the FR model and head, respectively. If $\\mathbf{v}_0$ and $\\mathbf{w}_0$ are fixed, $\\mathbf{V}_c$ and $\\mathbf{W}_c$ become deterministic aside from minor random factors in the training process of Equation (8). Therefore, there exists a mapping between $\\{\\mathbf{V}_c,\\mathbf{W}_c\\}$ and $\\{\\mathbf{v}_0,\\mathbf{w}_0\\}$ that can be expressed as:", + "bbox": [ + 511, + 597, + 906, + 703 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {V} _ {c}, \\mathbf {W} _ {c} \\right\\} = F \\left(\\left\\{\\mathbf {v} _ {0}, \\mathbf {w} _ {0} \\right\\}, i\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 712, + 906, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing adversarial attack methods based on parameter augmentation typically initialize the parameters $\\{\\mathbf{v}_0,\\mathbf{w}_0\\}$ solely with pre-trained values [12, 57]. This constraint diminishes the diversity within the augmented surrogate model set. In contrast, we opt to initialize $\\{\\mathbf{v}_0,\\mathbf{w}_0\\}$ with diverse parameters to enhance the diversity of the augmented surrogate model set, thereby improving transferability. The $\\{\\mathbf{V}_c,\\mathbf{W}_c\\}$ of our proposed attack can be expressed as:", + "bbox": [ + 511, + 738, + 906, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left\\{\\mathbf {V} _ {c} ^ {p}, \\mathbf {W} _ {c} ^ {p} \\right\\} = F \\left(\\left\\{\\mathbf {v} _ {0} ^ {p}, \\mathbf {w} _ {0} ^ {p} \\right\\}, i\\right) \\tag {10} \\\\ \\left\\{\\mathbf {V} _ {c} ^ {a}, \\mathbf {W} _ {c} ^ {a} \\right\\} = F \\left(\\left\\{\\mathbf {v} _ {0} ^ {a}, \\mathbf {w} _ {0} ^ {a} \\right\\}, i\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 867, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3519", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db00f76aa6e1fab04667b020fc5c4f89d78f35a29b43e361205fb5377649a297.jpg", + "image_caption": [ + "Figure 3. The framework of the Hard Model Aggregation (HMA). After acquiring a surrogate model set with diverse parameters (i.e., $\\mathbf{V}_c^q$ ), we introduce beneficial perturbations with the optimization direction opposite to that of adversarial perturbations onto the feature maps of these diversified surrogate models, transforming them into hard models and aggregate the hard models to increase the transferability." + ], + "image_footnote": [], + "bbox": [ + 107, + 89, + 893, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_0^p$ represents the pre-trained parameters, while $\\mathbf{w}_0^p$ , $\\mathbf{v}_0^a$ and $\\mathbf{w}_0^a$ are randomly initialized. Using Equation (10), we can obtain the more diverse parameters for the FR model and head. After obtaining $\\mathbf{V}_c^p$ and $\\mathbf{V}_c^a$ , we use the following set of the parameters to craft the adversarial examples:", + "bbox": [ + 89, + 277, + 483, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {V} _ {c} ^ {q} = \\left\\{\\mathbf {v} _ {0} ^ {p}, \\mathbf {v} _ {1} ^ {p}, \\mathbf {v} _ {j _ {1}} ^ {p}, \\dots , \\mathbf {v} _ {c} ^ {p} \\right\\} \\bigcup \\left\\{\\mathbf {v} _ {1} ^ {a}, \\mathbf {v} _ {j _ {1}} ^ {a}, \\mathbf {v} _ {j _ {2}} ^ {a}, \\dots , \\mathbf {v} _ {c} ^ {a} \\right\\} \\tag {11} \\\\ \\mathrm {s . t .} j \\in \\{1, 2, \\dots , c | j \\bmod \\kappa = 1 \\} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 367, + 483, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\kappa = \\lfloor \\sqrt{c} \\rfloor$ determines the epoch interval to select the parameters. The overall framework and pseudo-code of the DPO stage in our proposed method is illustrated in Figure 2 and Algorithm 1, respectively.", + "bbox": [ + 89, + 420, + 483, + 481 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Diverse Parameters Optimization (DPO)" + ], + "code_body": "Input: Diverse initial parameters $\\mathbf{v}_0^p$ , $\\mathbf{w}_0^p$ , $\\mathbf{v}_0^a$ , and $\\mathbf{w}_0^a$ , the number of epochs $c$ , the optimizer $\\mathbf{T}$ , the FR training dataset $\\zeta$ , loss function $\\mathcal{L}$ .", + "guess_lang": "latex", + "bbox": [ + 91, + 515, + 483, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Augmented parameter set $\\mathbf{V}_c^q$ .", + "bbox": [ + 91, + 563, + 356, + 577 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $\\mathbf{P} = \\{\\mathbf{v}_0^p, \\mathbf{w}_0^p\\}$ , $\\mathbf{A} = \\{\\mathbf{v}_0^a, \\mathbf{w}_0^a\\}$ , $\\mathbf{V}_c^q = \\{\\mathbf{v}_0^p\\}$", + "2: for $\\mathbf{J} \\in \\{\\mathbf{P}, \\mathbf{A}\\}$ do", + "3: $\\{\\mathbf{v},\\mathbf{w}\\} = \\mathbf{J}$", + "4: $\\mathcal{F} = \\Psi (\\mathbf{v})\\quad \\triangleright$ Map the parameters to the models.", + "5: for $i = 1,\\dots,c$ do", + "6: for $\\mathbf{x}\\in \\zeta$ do", + "7: Calculate $\\mathcal{L}$ using Equation (7).", + "8: Backward $(\\mathbf{T},\\mathcal{L})\\triangleright$ Backpropagation using $\\mathbf{T}$", + "9: Update the parameters $\\mathbf{v},\\mathbf{w}$", + "0: end for", + "11: if the $\\mathbf{v}$ satisfies the condition outlined in Equation (11) then", + "12: Add $(\\mathbf{V}_{c}^{q},\\mathbf{v})$ Incorporate $\\mathbf{v}$ into $\\mathbf{V}_c^q$", + "13: end if", + "14: end for", + "15: end for" + ], + "bbox": [ + 98, + 577, + 483, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Hard Model Aggregation", + "text_level": 1, + "bbox": [ + 89, + 847, + 320, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hard models [71] refer to models with higher loss values compared to normal models, designed to simulate the effect", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of hard samples, which are hard to the optimization process of the adversarial examples. Since hard samples can improve the generalization of models, we leverage multiple hard models to enhance the transferability of the surrogate models with diverse parameters by adding beneficial perturbations [56] onto the pre-defined feature maps of the surrogate models.", + "bbox": [ + 511, + 280, + 906, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the following, we will provide a detailed introduction to the HMA stage. Let $\\Psi$ be the mapping from the parameters to the corresponding models that can be expressed as:", + "bbox": [ + 511, + 369, + 905, + 415 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} = \\Psi \\left(\\mathbf {V} _ {c} ^ {q}\\right) = \\left\\{\\mathcal {F} _ {1}, \\mathcal {F} _ {2}, \\dots , \\mathcal {F} _ {g} \\right\\} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 422, + 906, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g = |\\mathbf{V}_c^q|$ . Hard sample augmentation has demonstrated significant effectiveness in enhancing model generalization [41, 67]. Based on the relationship between adversarial attack tasks and training tasks [9, 24], using hard models for augmentation can lead to improved transferability [71]. Consequently, by transforming vanilla parameter-augmented surrogate models into hard parameter-augmented surrogate models, we can achieve more transferable results. To accomplish this transformation, we use the following formula to calculate the loss function:", + "bbox": [ + 511, + 446, + 908, + 598 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widetilde {\\mathcal {L}} _ {t} = \\frac {1}{g} \\sum_ {i = 0} ^ {g} \\left\\| \\phi \\left(\\mathcal {H} _ {i} \\left(\\mathcal {T} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right)\\right)\\right) - \\phi \\left(\\mathcal {F} _ {i} \\left(\\mathbf {x} ^ {t}\\right)\\right) \\right\\| \\tag {13} \\\\ \\begin{array}{l l} \\text {s . t .} & \\mathcal {F} _ {i} \\in \\mathbf {F} \\quad t \\in \\{1, 2, \\dots , n \\} \\end{array} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 603, + 905, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_t^{adv}$ denotes the adversarial example generated during the $t$ th iteration, the variable $n$ represents the maximum number of iterations allocated for crafting the adversarial examples, $\\phi(\\cdot)$ denotes the normalization operation, $\\mathcal{T}$ signifies the input transformation, and $\\mathcal{H}$ is the corresponding hard model [71] of $\\mathcal{F}$ , which can be expressed as:", + "bbox": [ + 511, + 669, + 908, + 760 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right) = \\left\\{ \\begin{array}{l l} \\mathcal {F} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right) & \\text {s . t .} \\quad t = 1, \\\\ \\chi \\left(\\mathcal {F} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right), \\Omega\\right) & \\text {s . t .} \\quad t > 1, \\end{array} \\right. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 766, + 906, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\chi$ is the mapping for adding beneficial perturbations on the feature maps in the forward propagation whose formula can be expressed as:", + "bbox": [ + 511, + 806, + 906, + 851 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\chi \\left(\\mathcal {F} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right), \\Omega\\right): \\omega = \\omega + \\eta \\text {s i g n} \\left(\\nabla_ {\\omega} \\tilde {\\mathcal {L}} _ {t - 1}\\right) \\tag {15} \\\\ \\begin{array}{c c} \\text {s . t .} & \\omega \\in \\Omega \\end{array} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 858, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3520", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Omega$ is the pre-defined set of layers for adding beneficial perturbations. Let $\\epsilon$ be the maximum allowable perturbation. Utilizing $\\tilde{\\mathcal{L}}_t$ , we can craft the adversarial face examples using following formula:", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t + 1} ^ {a d v} = \\prod_ {\\mathbf {x} ^ {s}, \\epsilon} \\left(\\mathbf {x} _ {t} ^ {a d v} - \\beta \\operatorname {s i g n} \\left(\\nabla_ {\\mathbf {x} _ {t} ^ {a d v}} \\tilde {\\mathcal {L}} _ {t}\\right)\\right) \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 161, + 483, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\beta$ denotes the step size used to construct the adversarial examples, while $\\prod$ represents the clipping operation that confines the pixel values of the generated adversarial examples within the range $[\\mathbf{x}^s -\\epsilon ,\\mathbf{x}^s +\\epsilon ]$ . The framework of the HMA is illustrated in Figure 3.", + "bbox": [ + 89, + 205, + 483, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the following, we present the pseudo-code for the HMA. For the sake of clarity, we focus on neural networks with a single branch in their computational graphs. In the case of neural networks with multiple branches, the HMA algorithm remains largely unchanged, except for the incorporation of mechanisms to handle multiple branches. Before proceeding, we introduce some necessary definitions. Let $\\Phi$ denote the index set of pre-selected layers to which beneficial perturbations will be added:", + "bbox": [ + 89, + 281, + 483, + 416 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi = \\varkappa (\\Omega) \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 428, + 482, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\varkappa$ represents the mapping from the pre-defined set of layers for adding beneficial perturbations $\\Omega$ , to their corresponding layer index set $\\Phi$ . Let $f^i$ denote the $i$ th layer within the model $\\mathcal{F}$ , and let $z$ represent the total number of layers in $\\mathcal{F}$ . We define the segment of $\\mathcal{F}$ from layer $f^i$ to layer $f^j$ as follows:", + "bbox": [ + 89, + 455, + 483, + 546 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} ^ {i, j} = f ^ {i} \\circ f ^ {i + 1} \\circ \\dots \\circ f ^ {j - 1} \\circ f ^ {j} \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 559, + 482, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The pseudo-code for the HMA stage of our proposed method is presented in Algorithm 2.", + "bbox": [ + 89, + 584, + 482, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 627, + 220, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 89, + 652, + 290, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We opt to use the LFW [16], and CelebA-HQ [19] for our experiments to verify the effectiveness of our proposed attack method. LFW serves as an unconstrained face dataset for FR. CelebA-HQ consists of face images with high visual quality. The LFW and CelebA-HQ utilized in our experiments are identical to those employed in [70, 71, 73]. For the parameter augmentation process, we select BUPT-Balancedface [47] as our training dataset.", + "bbox": [ + 89, + 672, + 483, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Face Recognition Models. The normal trained FR models employed in our experiments encompass IR152 [13], IRSE50 [14], FaceNet [34], MobileFace [8], CurricularFace [17], MagFace [29], ArcFace [8], CircleLoss [38], MV-Softmax [51], and NPCFace [65]. Specifically, IR152, FaceNet, IRSE50, and MobileFace are the same models utilized in [15, 63, 70, 71]. CurricularFace, MagFace, ArcFace,", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 2 Hard Model Aggregation (HMA)", + "bbox": [ + 514, + 90, + 823, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input: The source image $\\mathbf{x}^s$ , the target image $\\mathbf{x}^t$ , the mapping from the parameters to the corresponding models $\\Psi$ , the maximum number of iterations $n$ , the index set of pre-selected layers to be added beneficial perturbations $\\Phi$ , the step size of the beneficial perturbations $\\eta$ , the step size of the adversarial perturbations $\\beta$ , the maximum permissible magnitude of perturbation $\\epsilon$ , the parameter set $\\mathbf{V}_c^q$ , the total number of layers in a single surrogate model $z$ , normalization operation $\\phi$ .", + "bbox": [ + 514, + 111, + 906, + 246 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Output: An adversarial example $\\mathbf{x}_n^{adv}$", + "bbox": [ + 514, + 247, + 771, + 263 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $\\mathbf{x}_0^{adv} = \\mathbf{x}^s$ , $u = |\\Phi|, s_1 = 1, g = |\\mathbf{V}_c^q|$", + "2: $\\mathbf{F} = \\Psi (\\mathbf{V}_c^q)$ $\\triangleright$ Acquire the diversified models set.", + "3: for $t = 1, \\dots, n$ do", + "4: for $i = 1,\\dots,g$ do", + "5: $\\mathcal{F} = \\mathbf{F}_i$ Derive the parameter-augmented model.", + "6: $\\omega^{t,0} = \\mathcal{T}\\left(\\mathbf{x}_{t - 1}^{adv}\\right)$", + "7: for $j = 1,\\dots,u$ do", + "8: $s_2 = \\Phi_j$", + "9: $\\omega^{t,j} = \\mathcal{F}^{s_1,s_2}\\left(\\omega^{t,j - 1}\\right)$", + "10: $s_1 = s_2$", + "11: if $t \\neq 1$ then", + "12: $\\omega^{t,j} = \\omega^{t,j} + \\eta \\mathrm{sign}\\left(\\nabla_{\\omega^{t - 1,j}}\\widetilde{\\mathcal{L}}_{t - 1}\\right)$", + "13: end if", + "14: end for", + "15: $\\tilde{\\mathcal{L}}^i = \\| \\phi (\\mathcal{F}^{\\Phi_u,z}(\\omega^{t,u})) - \\phi (\\mathcal{F}(\\mathbf{x}^t))\\| _2^2$", + "16: end for", + "17: $\\widetilde{\\mathcal{L}}_t = \\frac{1}{g}\\sum_{i = 1}^g\\widetilde{\\mathcal{L}}^i$ Compute the loss function utilizing the parameter-augmented models.", + "18: $\\mathbf{x}_t^{adv} = \\prod_{\\mathbf{x}^s,\\epsilon}\\left(\\mathbf{x}_{t - 1}^{adv} - \\beta \\mathrm{sign}\\left(\\nabla_{\\mathbf{x}_{t - 1}^{adv}}\\widetilde{\\mathcal{L}}_t\\right)\\right)$", + "19: end for" + ], + "bbox": [ + 519, + 263, + 903, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CircleLoss, MV-Softmax, and NPCFace are the official models provided by FaceX-ZOO [45]. Furthermore, we integrate adversarial robust FR models into our experiments, denoted as IR152 $^{adv}$ , IRSE50 $^{adv}$ , FaceNet $^{adv}$ , and MobileFace $^{adv}$ which correspond to the models used in [71].", + "bbox": [ + 511, + 626, + 906, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attack Setting. We set the maximum allowable perturbation magnitude $\\epsilon$ to 10, based on the $L_{\\infty}$ norm bound, without any specific emphasis. In addition, we specify the maximum number of iterative steps as 200. More detailed attack settings are provided in the supplementary.", + "bbox": [ + 511, + 702, + 906, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We utilize Attack Success Rate (ASR) as the metric to assess the efficacy of various adversarial attacks. The ASR represents the ratio of adversarial examples that successfully evade the victim model to the total number of adversarial examples generated. In calculating the ASR, we determine the threshold based on a FAR@0.001 on the entire LFW dataset for each victim model.", + "bbox": [ + 511, + 779, + 906, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. Our proposed attack is a form of re", + "bbox": [ + 511, + 885, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3521", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "stricted adversarial attack designed to expose vulnerabilities within FR systems. It would be inequitable to juxtapose our method with unrestricted attacks, which do not impose limits on the magnitude of adversarial perturbations. Consequently, we opt to benchmark our approach against other restricted attacks on FR systems that are explicitly malicious at attacking the systems [69] [71], as well as state-of-the-art transfer attacks [59] [27] [12] [52] [46].", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison Studies", + "text_level": 1, + "bbox": [ + 89, + 220, + 282, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DPA achieves the best black-box attack results on both normally trained and adversarial robust models. To verify the effectiveness of our proposed attack method, we craft the adversarial examples on the LFW and CelebA-HQ dataset. The black-box performance with MobileFace as the surrogate models on the LFW dataset are demonstrated in Table 1. Some of the adversarial examples are demonstrated in Figure 4. More experimental results on the LFW dataset, as well as those on the CelebA-HQ dataset, are provided in the supplementary. These results demonstrate that our proposing method consistently outperforms the baseline attack, thereby highlighting its effectiveness in improving the transferability of adversarial face examples.", + "bbox": [ + 89, + 242, + 483, + 439 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de1ea6c5002e68905201547293464742956703050fb5bf07b6a141d0a27e1736.jpg", + "image_caption": [ + "Figure 4. The illustration of adversarial examples crafted by various attacks. First column: some source images. Last column: the corresponding target images. The second to fifth columns exhibit the corresponding adversarial face examples crafted by SIA [52], BPFA [71], BSR [46], and Our proposed attack, respectively." + ], + "image_footnote": [], + "bbox": [ + 91, + 450, + 483, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DPA achieves superior black-box performance under JPEG compression. JPEG compression is a widely adopted method for image compression during transmission, and it also serves as a defense mechanism against adversarial examples. To evaluate the effectiveness of our proposed attack under JPEG compression, we employ MobileFace as the surrogate model and IRSE50 as the victim model. We assess the attack performance on the LFW and CelebA-HQ datasets, with the results presented in Figure 5. These results indicate", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fd445ac5a1daa5bd8dcd450c5f09db095b7b43c16cc25a5e3188aecd35899ecf.jpg", + "table_caption": [ + "Table 1. Comparisons of black-box ASR $(\\%)$ results for attacks using MobileFace as the surrogate model on the LFW dataset. I, S, F, M denote IR152, IRSE50, FaceNet, and MobileFace, respectively." + ], + "table_footnote": [], + "table_body": "
AttacksISFIadvSadvFadvMadv
FIM [68]5.373.47.52.54.52.810.9
DI [59]18.497.332.910.618.210.238.6
DFANet [69]7.086.411.93.66.43.916.8
VMI [50]13.696.020.27.612.87.632.3
SSA [27]13.896.419.65.513.17.231.5
SIA [52]15.796.826.78.114.59.135.3
BPFA [71]15.895.517.85.911.05.128.4
BSR [46]5.474.29.52.95.94.911.4
Ours67.798.290.855.466.442.671.6
", + "bbox": [ + 517, + 136, + 901, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/81d0a75afe082434297617af1f5a3e3084a098981cf16c2e169eb8b3acea8c0b.jpg", + "image_caption": [ + "Figure 5. Performance of ASR across various JPEG Q values: (a) Results on the LFW dataset. (b) Results on the CelebA-HQ dataset." + ], + "image_footnote": [], + "bbox": [ + 526, + 285, + 710, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f3d27c33027bb07124314b3e6202b6e9ab2a38f2f7cd273e4e5c53a701b1ae2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 285, + 893, + 407 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/53554cbcb6f1e29c827a0c5c31f4d1f0f8770c6707da1f3838b4a3d10c38ddb4.jpg", + "table_caption": [ + "Table 2. Comparison of black-box ASR (%) results using the parameter-based augmented adversarial attack method as the baseline on the LFW dataset. $\\mathrm{ASR}^{adv}$ represents the average attack success rate on adversarial robust models." + ], + "table_footnote": [], + "table_body": "
IR152IRSE50FaceNetASRadv
Baseline4.819.615.77.2
Ours67.798.290.859.0
", + "bbox": [ + 529, + 517, + 888, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that our proposed attack method consistently outperforms the baseline attack methods across varying levels of JPEG compression, thereby underscoring the robustness of our proposed attack method under such conditions.", + "bbox": [ + 511, + 597, + 906, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DPA demonstrates better black-box performance compared to the parameter-based augmented adversarial attack. The LGV method is an effective parameter-based augmented adversarial attack technique designed to enhance transferability. Our proposed attack incorporates parameter augmentation, making LGV an appropriate baseline. We selected MobileFace as the surrogate model and generated adversarial examples on the LFW dataset. The results are presented in Table 2. Table 2 clearly shows that our proposed attack outperforms the baseline, further validating the effectiveness of our method. The primary limitation of LGV stems from its random selection of one model per iteration, which introduces instability and adversely impacts blackbox ASR [9, 50]. Furthermore, LGV suffers from a lack of diverse initialization parameters. These shortcomings significantly impair the overall ASR performance of LGV.", + "bbox": [ + 511, + 659, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3522", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b4d847a842e07f84ed23c476d9cf9f43c589da7dc0cb22b0cb16c91bd5fc462.jpg", + "image_caption": [ + "Figure 6. ASR on victim models trained with various algorithms, using FaceNet as the surrogate model on the LFW dataset." + ], + "image_footnote": [], + "bbox": [ + 106, + 90, + 472, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DPA achieves superior black-box performance on FR models trained with various algorithms. In order to substantiate the effectiveness of our proposed attack method across various FR models, we conducted additional experiments using BSR [46] as the Baseline and FaceNet as the surrogate model. These experiments adhered to the same settings detailed in Table 7 in the supplementary. The ASR across multiple FR models is illustrated in Figure 6. As shown in Figure 6, the ASR of the adversarial examples generated by our proposed method exceeds that of the Baseline, thereby reinforcing the efficacy of our proposed method.", + "bbox": [ + 89, + 305, + 485, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 484, + 254, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ablation studies on each stage of our proposed attack method. Our proposed attack method consists of two stages. To verify the effectiveness of each stage, we conduct ablation studies on the stages. Initially, we craft adversarial examples using only the surrogate model, which we denote as 'Vanilla'. Next, we generate adversarial examples using the ensemble of surrogate models obtained in the DPO stage (i.e. F), denoted as DPO. After incorporating the HMA stage, the complete attack method is denoted as 'DPO + HMA'. The results are presented in Table 3. Table 3 demonstrates that the addition of the DPO stage results in an increase in black-box ASR, showcasing the effectiveness of the DPO stage in enhancing transferability. Further incorporation of the HMA stage leads to an additional improvement in attack performance, underscoring the effectiveness of the HMA stage in boosting black-box performance. These results collectively demonstrate the effectiveness of each stage in our proposed attack method in improving the transferability.", + "bbox": [ + 89, + 507, + 485, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ablation studies on the effectiveness of multiple surrogate models at different epochs in improving transferability. Since our proposed attack method utilizes models from intermediate training epochs to craft adversarial examples, it is essential to verify the effectiveness of this approach compared to using only models from the final training epoch. We conduct ablation experiments on these two approaches using the LFW dataset and MobileFace as the surrogate", + "bbox": [ + 89, + 780, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6d9cc361770d106b3458acb5ec3b96a7f937f67e17cc421d53f5b06243ca7ce5.jpg", + "table_caption": [ + "Table 3. Ablation study of black-box ASR (\\%) results on the LFW dataset using MobileFace as the surrogate model. $\\mathrm{ASR}^{adv}$ denotes the average attack success rate on adversarial robust models." + ], + "table_footnote": [], + "table_body": "
AttacksIR152IRSE50FaceNetASRadv
Vanilla5.373.47.55.2
DPO35.991.367.932.9
DPO + HMA67.798.290.859.0
", + "bbox": [ + 514, + 136, + 906, + 202 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c6b9838490e15f14baa3dbea71f526b79d6197b440a75a527dc890342d3cdcd4.jpg", + "table_caption": [ + "Table 4. Ablation study of black-box ASR (\\%) results to verify the effectiveness of multiple surrogate models at different epochs in improving transferability. $\\mathrm{ASR}^{adv}$ denotes the average attack success rate on adversarial robust models." + ], + "table_footnote": [], + "table_body": "
AttacksIR152IRSE50FaceNetASRadv
Fm20.386.733.617.8
F (Ours)35.991.367.932.9
", + "bbox": [ + 529, + 276, + 890, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model. We employ the DPO process to obtain the parameter sets $\\mathbf{V}_c^q$ and $\\mathbf{V}_c^m = \\{\\mathbf{v}_0^p, \\mathbf{v}_c^p, \\mathbf{v}_c^a\\}$ . Next, we map $\\mathbf{V}_c^m$ into its corresponding model set $\\mathbf{F}^m = \\Psi(\\mathbf{V}_c^m)$ . We then craft adversarial examples by ensembling the models from $\\mathbf{V}_c^q$ (i.e. $\\mathbf{F}$ ) and $\\mathbf{F}^m$ , respectively. The results are presented in Table 4. Table 4 shows that the performance of the ensemble of $\\mathbf{F}$ surpasses that of the ensemble of $\\mathbf{F}^m$ , demonstrating the effectiveness of using models from intermediate training epochs to enhance the transferability.", + "bbox": [ + 511, + 351, + 906, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The hyper-parameter analysis studies of our proposed method are in the supplementary.", + "bbox": [ + 511, + 487, + 906, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 531, + 633, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present an innovative advancement in the field of adversarial attacks on FR through the introduction of the Diverse Parameters Augmentation (DPA) attack method. By addressing the problems of traditional adversarial attacks, particularly the lack of diverse parameters and the exclusion of the FR head, DPA enhances the transferability of adversarial face examples. In addition, DPA emphasizes the importance of integrating both pre-trained and randomly initialized parameters, providing a fresh perspective on augmentation-based adversarial attacks. The extensive experimental results demonstrate the effectiveness of our proposed DPA attack across multiple datasets and target models. We anticipate that our proposed method will illuminate potential directions for advancing research in adversarial attacks on FR.", + "bbox": [ + 511, + 556, + 908, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 777, + 684, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the Natural Science Foundation of China under Grant 62372203 and 62302186, in part by the Major Scientific and Technological Project of Shenzhen (202316021), in part by the National key research and development program of China (2022YFB2601802), in part by the Major Scientific and Technological Project of Hubei Province (2022BAA046, 2022BAA042).", + "bbox": [ + 511, + 801, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3523", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 121, + 187, + 136 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, and Ying Fu. Partial FC: training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 1445-1449, 2021. 1", + "[2] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19593-19604, 2023. 1", + "[3] Tom B Brown, Nicholas Carlini, Chiyuan Zhang, Catherine Olsson, Paul Christiano, and Ian Goodfellow. Unrestricted adversarial examples. arXiv preprint arXiv:1809.08352, 2018. 3", + "[4] Bin Chen, Jia-Li Yin, Shukai Chen, Bohao Chen, and Ximeng Liu. An adaptive model ensemble adversarial attack for boosting adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4466-4475, 2023. 3", + "[5] Zhaoyu Chen, Bo Li, Shuang Wu, Kaixun Jiang, Shouhong Ding, and Wenqiang Zhang. Content-based unrestricted adversarial attack. In Advances in Neural Information Processing Systems, 2023. 3", + "[6] Valeriia Cherepanova, Micah Goldblum, Harrison Foley, Shiyuan Duan, John P Dickerson, Gavin Taylor, and Tom Goldstein. Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In International Conference on Learning Representations, 2021. 1", + "[7] Sowmen Das, Selim Seferbekov, Arup Datta, Md Saiful Islam, and Md Ruhul Amin. Towards solving the deepfake problem: An analysis on improving deepfake detection using dynamic face augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3776-3785, 2021. 1", + "[8] Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5962-5979, 2022. 1, 2, 4, 6", + "[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 2, 3, 5, 7", + "[10] Zhijin Ge, Fanhua Shang, Hongying Liu, Yuanyuan Liu, Liang Wan, Wei Feng, and Xiaosen Wang. Improving the transferability of adversarial examples with arbitrary style transfer. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4440-4449, 2023. 2", + "[11] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. 2" + ], + "bbox": [ + 93, + 147, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Martin Gubri, Maxime Cordy, Mike Papadakis, Yves Le Traon, and Koushik Sen. Lgv: Boosting adversarial example transferability from large geometric vicinity. In European Conference on Computer Vision, pages 603-618. Springer, 2022. 2, 4, 7", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6", + "[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 6", + "[15] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14994–15003, 2022. 1, 3, 6", + "[16] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, 2007. 6", + "[17] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: Adaptive curriculum learning loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 6", + "[18] Shuai Jia, Bangjie Yin, Taiping Yao, Shouhong Ding, Chunhua Shen, Xiaokang Yang, and Chao Ma. Adv-attribute: Inconspicuous and transferable adversarial attack on face recognition. In Advances in Neural Information Processing Systems, 2022. 3", + "[19] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representation, 2018. 6", + "[20] Jingzhi Li, Zidong Guo, Hui Li, Seungju Han, Ji-Won Baek, Min Yang, Ran Yang, and Sungjoo Suh. Rethinking feature-based knowledge distillation for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20156-20165, 2023. 1", + "[21] Yanjie Li, Yiquan Li, Xuelong Dai, Songtao Guo, and Bin Xiao. Physical-world optical adversarial attacks on 3d face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24699-24708, 2023. 1, 3", + "[22] Zexin Li, Bangjie Yin, Taiping Yao, Junfeng Guo, Shouhong Ding, Simin Chen, and Cong Liu. Sibling-attack: Rethinking transferable adversarial attacks against face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24626–24637, 2023. 1, 3", + "[23] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3524", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "from diffusion models via adversarial examples. In International Conference on Machine Learning, pages 20763-20786, 2023. 2", + "[24] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representation, 2020. 2, 5", + "[25] Xuannan Liu, Yaoyao Zhong, Yuhang Zhang, Lixiong Qin, and Weihong Deng. Enhancing generalization of universal adversarial perturbation through gradient aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4412-4421, 2023. 3", + "[26] Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, and Di Ming. TRM-UAP: enhancing the transferability of data-free universal adversarial perturbation via truncated ratio maximization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4739-4748, 2023. 3", + "[27] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xi-anglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In European Conference on Computer Vision, pages 549–566, 2022. 2, 7, 1", + "[28] Dong Lu, Zhiqiang Wang, Teng Wang, Weili Guan, Hongchang Gao, and Feng Zheng. Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 102-111, 2023. 2, 3", + "[29] Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14225-14234, 2021. 6", + "[30] Duan Mingxing, Kenli Li, Lingxi Xie, Qi Tian, and Bin Xiao. Towards multiple black-boxes attack via adversarial example generation network. In Proceedings of the 29th ACM International Conference on Multimedia, pages 264-272, 2021. 2", + "[31] Kartik Narayan, Harsh Agarwal, Kartik Thakral, Surbhi Mittal, Mayank Vatsa, and Richa Singh. Df-platter: Multiface heterogeneous deepfake dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9739-9748, 2023. 1", + "[32] Xuelin Qian, Wenxuan Wang, Yu-Gang Jiang, Xiangyang Xue, and Yanwei Fu. Dynamic routing and knowledge relearning for data-free black-box attack. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4", + "[33] Haonan Qiu, Chaowei Xiao, Lei Yang, Xinchen Yan, Honglak Lee, and Bo Li. Semanticadv: Generating adversarial examples via attribute-conditioned image editing. In European Conference on Computer Vision, pages 19-37, 2020. 1, 3", + "[34] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 815-823, 2015. 1, 6", + "[35] Fahad Shamshad, Muzammal Naseer, and Karthik Nandakumar. Clip2protect: Protecting facial privacy using text-guided" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "makeup via adversarial latent search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20595-20605, 2023. 1, 3", + "[36] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In International Conference on Learning Representations, 2024. 2", + "[37] Florian Stimberg, Ayan Chakrabarti, Chun-Ta Lu, Hussein Hazimeh, Otilia Stretcu, Wei Qiao, Yintao Liu, Merve Kaya, Cyrus Rashtchian, Ariel Fuxman, Mehmet Tek, and Sven Gowal. Benchmarking robustness to adversarial image obfuscations. In Advances in Neural Information Processing Systems, 2023. 3", + "[38] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 6", + "[39] Yuhao Sun, Lingyun Yu, Hongtao Xie, Jiaming Li, and Yongdong Zhang. Diffam: Diffusion-based adversarial makeup transfer for facial privacy protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24584-24594, 2024. 3", + "[40] Naufal Suryanto, Yongsu Kim, Harashta Tatimma Larasati, Hyoeun Kang, Thi-Thu-Huong Le, Yoonyoung Hong, Hunmin Yang, Se-Yoon Oh, and Howon Kim. ACTIVE: towards highly transferable 3d physical camouflage for universal and robust vehicle evasion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4282-4291, 2023. 3", + "[41] Teppei Suzuki. Teachaugment: Data augmentation optimization using teacher knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10894-10904, 2022. 5", + "[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 2", + "[43] Donghua Wang, Wen Yao, Tingsong Jiang, Chao Li, and Xiaogian Chen. RFLA: A stealthy reflected light adversarial attack in the physical world. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4442, 2023. 3", + "[44] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 2", + "[45] Jun Wang, Yinglu Liu, Yibo Hu, Hailin Shi, and Tao Mei. Facex-zoo: A python toolbox for face recognition. 2021. 6", + "[46] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting Adversarial Transferability by Block Shuffle and Rotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 7, 8, 1", + "[47] Mei Wang, Weihong Deng, Jiani Hu, Xunqiang Tao, and Yaohai Huang. Racial faces in the wild: Reducing racial" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3525", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "bias by information maximization adaptation network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 6", + "[48] Wenxuan Wang, Bangjie Yin, Taiping Yao, Li Zhang, Yanwei Fu, Shouhong Ding, Jilin Li, Feiyue Huang, and Xiangyang Xue. Delving into data: Effectively substitute training for black-box attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4761-4770, 2021. 4", + "[49] Wenxuan Wang, Xuelin Qian, Yanwei Fu, and Xiangyang Xue. Dst: Dynamic substitute training for data-free black-box attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14361-14370, 2022. 4", + "[50] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1924-1933, 2021. 2, 4, 7, 1", + "[51] Xiaobo Wang, Shifeng Zhang, Shuo Wang, Tianyu Fu, Hailin Shi, and Tao Mei. Mis-classified vector guided softmax loss for face recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12241-12248, 2020. 6", + "[52] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4607-4619, 2023. 2, 7, 1", + "[53] Xudong Wang, Ke-Yue Zhang, Taiping Yao, Qianyu Zhou, Shouhong Ding, Pingyang Dai, and Rongrong Ji. Tf-fas: Twofold-element fine-grained semantic guidance for generalizable face anti-spoofing. In European Conference on Computer Vision, 2024. 1", + "[54] Xingxing Wei, Yao Huang, Yitong Sun, and Jie Yu. Unified adversarial patch for cross-modal attacks in the physical world. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4422-4431, 2023. 3", + "[55] Xingxing Wei, Jie Yu, and Yao Huang. Physically adversarial infrared patches with learnable shapes and locations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12334-12342, 2023. 3", + "[56] Shixian Wen and Laurent Itti. Beneficial perturbations network for defending adversarial examples. arXiv preprint arXiv:2009.12724, 2020. 2, 3, 5", + "[57] Han Wu, Guanyan Ou, Weibin Wu, and Zibin Zheng. Improving transferable targeted adversarial attacks with model self-enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24615-24624, 2024. 2, 4", + "[58] Zihao Xiao, Xianfeng Gao, Chilin Fu, Yinpeng Dong, Wei Gao, Xiaolu Zhang, Jun Zhou, and Jun Zhu. Improving transferability of adversarial patches on face recognition with generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11840-11849, 2021. 3", + "[59] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019. 2, 7, 1" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Zhuoer Xu, Zhangxuan Gu, Jianping Zhang, Shiwen Cui, Changhua Meng, and Weiqiang Wang. Backpropagation path search on adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4640-4650, 2023. 3", + "[61] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3877-3887, 2021. 1", + "[62] Xiao Yang, Chang Liu, Longlong Xu, Yikai Wang, Yinpeng Dong, Ning Chen, Hang Su, and Jun Zhu. Towards effective adversarial textured 3d meshes on physical face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4119-4128, 2023. 1, 3", + "[63] Bangjie Yin, Wenxuan Wang, Taiping Yao, Junfeng Guo, Zelun Kong, Shouhong Ding, Jilin Li, and Cong Liu. Advmakeup: A new imperceptible and transferable attack on face recognition. In Proceedings of the 30th International Joint Conference on Artificial Intelligence, pages 1252-1258, 2021. 1, 3, 6", + "[64] Shengming Yuan, Qilong Zhang, Lianli Gao, Yaya Cheng, and Jingkuan Song. Natural color fool: Towards boosting black-box unrestricted attacks. In Advances in Neural Information Processing Systems, 2022. 3", + "[65] Dan Zeng, Hailin Shi, Hang Du, Jun Wang, Zhen Lei, and Tao Mei. Npcface: Negative-positive collaborative training for large-scale face recognition. 2020. 6", + "[66] Jiaming Zhang, Qi Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5005-5013, 2022. 2", + "[67] Xinyu Zhang, Qiang Wang, Jian Zhang, and Zhao Zhong. Adversarial autoaugment. In International Conference on Learning Representations, 2020. 5", + "[68] Yaoyao Zhong and Weihong Deng. Adversarial learning with margin-based triplet embedding regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6548-6557, 2019. 7, 1, 2", + "[69] Yaoyao Zhong and Weihong Deng. Towards transferable adversarial attack against deep face recognition. IEEE Transactions on Information Forensics and Security, 16:1452-1466, 2021. 1, 2, 3, 7", + "[70] Fengfan Zhou, Hefei Ling, Yuxuan Shi, Jiazhong Chen, and Ping Li. Improving visual quality and transferability of adversarial attacks on face recognition simultaneously with adversarial restoration. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4540-4544, 2024. 6", + "[71] Fengfan Zhou, Hefei Ling, Yuxuan Shi, Jiazhong Chen, Zongyi Li, and Ping Li. Improving the transferability of adversarial attacks on face recognition with beneficial perturbation feature augmentation. IEEE Transactions on Computational Social Systems, 11(6):8130-8142, 2024. 1, 2, 3, 4, 5, 6, 7", + "[72] Fengfan Zhou, Qianyu Zhou, Xiangtai Li, Xuequan Lu, Lizhuang Ma, and Hefei Ling. Adversarial attacks on both face recognition and face anti-spoofing models. arXiv preprint arXiv:2405.16940, 2024. 3" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3526", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Fengfan Zhou, Qianyu Zhou, Bangjie Yin, Hui Zheng, Xuequan Lu, Lizhuang Ma, and Hefei Ling. Rethinking impersonation and dodging attacks on face recognition systems. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2487-2496, 2024. 6", + "[74] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Shouhong Ding, and Lizhuang Ma. Adaptive mixture of experts learning for generalizable face anti-spoofing. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6009-6018, 2022. 1", + "[75] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Kekai Sheng, Shouhong Ding, and Lizhuang Ma. Generative domain adaptation for face anti-spoofing. In European Conference on Computer Vision, pages 335-356, 2022.", + "[76] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Xuequan Lu, Ran Yi, Shouhong Ding, and Lizhuang Ma. Instance-aware domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20453-20463, 2023.", + "[77] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Xuequan Lu, Shouhong Ding, and Lizhuang Ma. Test-time domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 175-187, 2024. 1", + "[78] Ziqi Zhou, Shengshan Hu, Minghui Li, Hangtao Zhang, Yechao Zhang, and Hai Jin. Advclip: Downstream-agnostic adversarial examples in multimodal contrastive learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6311–6320, 2023. 2", + "[79] Hegui Zhu, Yuchen Ren, Xiaoyan Sui, Lianping Yang, and Wuming Jiang. Boosting adversarial transferability via gradient relevance attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4718-4727, 2023. 3" + ], + "bbox": [ + 91, + 92, + 486, + 569 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "3527", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/aa0474ee-612f-4595-88a1-c6ba9ec15b27_model.json b/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/aa0474ee-612f-4595-88a1-c6ba9ec15b27_model.json new file mode 100644 index 0000000000000000000000000000000000000000..93da68c2706a67e81b67fe92749215699c63ce31 --- /dev/null +++ b/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/aa0474ee-612f-4595-88a1-c6ba9ec15b27_model.json @@ -0,0 +1,3117 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.13, + 0.9, + 0.177 + ], + "angle": 0, + "content": "Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.203, + 0.8, + 0.222 + ], + "angle": 0, + "content": "Fengfan Zhou\\(^{1}\\), Bangjie Yin\\(^{2}\\), Hefei Ling\\(^{1*}\\), Qianyu Zhou\\(^{3}\\), Wenxuan Wang\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.222, + 0.882, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Technology, Huazhong University of Science and Technology; \\(^{2}\\)Shanghai Shizhuang Information Technology Co., Ltd;" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.257, + 0.833, + 0.293 + ], + "angle": 0, + "content": "\\(^{3}\\) Department of Computer Science and Engineering, Shanghai Jiao Tong University; \\(^{4}\\) School of Computer Science, Northwestern Polytechnical University." + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.293, + 0.727, + 0.309 + ], + "angle": 0, + "content": "\\(^{1}\\) {ffzhou, lhefei}@hust.edu.cn, \\(^{2}\\) jamesyin10@gmail.com," + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.31, + 0.682, + 0.327 + ], + "angle": 0, + "content": "3 zhouqianyu@sjtu.edu.cn, 4wxwang@nwpu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.362, + 0.327, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.486, + 0.759 + ], + "angle": 0, + "content": "Face Recognition (FR) models are vulnerable to adversarial examples that subtly manipulate benign face images, underscoring the urgent need to improve the transferability of adversarial attacks in order to expose the blind spots of these systems. Existing adversarial attack methods often overlook the potential benefits of augmenting the surrogate model with diverse initializations, which limits the transferability of the generated adversarial examples. To address this gap, we propose a novel method called Diverse Parameters Augmentation (DPA) attack method, which enhances surrogate models by incorporating diverse parameter initializations, resulting in a broader and more diverse set of surrogate models. Specifically, DPA consists of two key stages: Diverse Parameters Optimization (DPO) and Hard Model Aggregation (HMA). In the DPO stage, we initialize the parameters of the surrogate model using both pre-trained and random parameters. Subsequently, we save the models in the intermediate training process to obtain a diverse set of surrogate models. During the HMA stage, we enhance the feature maps of the diversified surrogate models by incorporating beneficial perturbations, thereby further improving the transferability. Experimental results demonstrate that our proposed attack method can effectively enhance the transferability of the crafted adversarial face examples." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.772, + 0.224, + 0.788 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.798, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Owing to the relentless progress in deep learning, Face Recognition (FR) has achieved substantial advancements [1, 2, 8, 20, 34, 44]. However, the inherent vulnerabilities in existing FR models pose significant security risks [7, 31, 53, 74-77], with adversarial attacks emerging as one" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.365, + 0.576, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.366, + 0.685, + 0.377 + ], + "angle": 0, + "content": "Pre-trained Parameters" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.366, + 0.738, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.366, + 0.893, + 0.378 + ], + "angle": 0, + "content": "Randomly Initialized Parameters" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.393, + 0.597, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.441, + 0.603, + 0.462 + ], + "angle": 0, + "content": "(a) Input-based Augmentation" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.393, + 0.695, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.441, + 0.703, + 0.462 + ], + "angle": 0, + "content": "(b) Feature-based Augmentation" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.393, + 0.792, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.441, + 0.808, + 0.462 + ], + "angle": 0, + "content": "(c) Parameter-based Augmentation" + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.476, + 0.645, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.47, + 0.772, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.476, + 0.896, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.572, + 0.908, + 0.656 + ], + "angle": 0, + "content": "Figure 1. Top: comparison between traditional augmentation-based adversarial attack methods and our proposed method. The black pattern filling on the left and right sides of the blue line represents input-based and parameter-based augmentation, respectively. The orange pattern filling indicates feature-based augmentation. Bottom: comparison of performance among 4 types of augmentations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.907, + 0.731 + ], + "angle": 0, + "content": "of the critical concerns. Therefore, there is an urgent need to bolster the resilience of FR models against adversarial face examples to reveal and address the vulnerabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "As a result, numerous research efforts have been focused on this area. Several adversarial attacks have been developed to create adversarial face examples with features like stealthiness [6, 15, 33, 35, 61], transferability [22, 69-71], and the capacity for physical attacks [21, 62, 63]. These initiatives are aimed at improving the effectiveness of adversarial attacks on FR. Nevertheless, the transferability of these adversarial attacks remains limited. To enhance the transferability of adversarial attacks, augmentation emerges as one of the most effective methods. As illustrated in Figure 1, augmentation-based adversarial attack consists of three aug" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.238, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3516" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.122 + ], + "angle": 0, + "content": "mentation types: input-based [27, 46, 52, 59], feature-based [69, 71], and parameter-based [12, 57]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.124, + 0.486, + 0.685 + ], + "angle": 0, + "content": "Most augmentation-based adversarial attack methods focus on input-based augmentation to improve transferability. Previous research highlights a symmetry between surrogate models used in adversarial attacks and input data in training tasks [9, 24, 71]. In training, data augmentation has proven effective in enhancing model generalization. Drawing on this symmetry, augmenting models for crafting adversarial examples can yield examples with greater transferability. Input-based augmentation can be seen as methods that only augment models at the input layer, demonstrating significant effectiveness in enhancing transferability. However, augmenting surrogate models in deeper layers (i.e., feature-based and parameter-based augmentation) offers a more direct form of augmentation. Nevertheless, few studies explore deep-layer augmentation [12, 57, 71]. As a typical parameter-based augmentation method in deep layers, LGV [12] uses a pre-trained surrogate model and collects multiple parameter sets through additional training epochs with a high, constant learning rate, thereby enhancing the transferability of adversarial examples. Although these parameter-based augmentation methods [12, 57] show promising effectiveness, they face two problems: (1) Static surrogate model initializations: these methods solely augment the surrogate model from pre-trained parameters, limiting the parameter diversity of the surrogate models and thereby hindering the transferability of the crafted adversarial examples. (2) Unavailability of the FR head: modern FR training procedures typically involve training both a backbone model and a head model [8, 44]. Following training, inference is conducted solely with the backbone models, while the head models are often not released as open-source. Consequently, in the majority of instances, we are unable to access the pre-trained parameters for the head models. Traditional parameter-based augmentation adversarial attacks augment models from pretrained parameters [12, 57]. This limitation makes it challenging to craft adversarial examples on FR models with numerous open-sourced FR models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.486, + 0.809 + ], + "angle": 0, + "content": "To address these problems, we introduce a novel adversarial attack called Diverse Parameters Augmentation (DPA) to enhance the transferability of crafted adversarial face examples. Unlike existing parameter-based augmentation adversarial attack methods that overlook the use of diverse parameter initializations [12, 57], we diversify parameter initializations with both random and pre-trained values, thereby improving black-box attack capacity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Technically, our proposed attack method comprises two stages: Diverse Parameters Optimization (DPO) and Hard Model Aggregation (HMA). During the DPO stage, we initiate a subset of the optimization parameters with random noise, preserving the refined parameters in the intermediate training process to yield a diverse set of surrogate models," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.303 + ], + "angle": 0, + "content": "which is instrumental in bolstering the transferability of the adversarial examples. In the HMA stage, we add beneficial perturbations [56] with optimization directions opposite to those of the adversarial perturbations onto the feature maps of the parameter-augmented surrogate models, achieving the effect of hard model augmentation [71] to further enhances the transferability of the crafted adversarial examples. Our proposed attack method effectively addresses the challenges posed by the lack of diverse parameters and the absence of the FR head. The comparison between traditional adversarial attack methods and our proposed method is illustrated in Figure 1. By using diverse initializations for surrogate models, we can expand the parameter set of these models, thereby enhancing the transferability of the adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.304, + 0.874, + 0.317 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.318, + 0.907, + 0.409 + ], + "angle": 0, + "content": "- We introduce a novel perspective that parameter-based augmentation adversarial attacks should augment the parameters of surrogate models by incorporating diverse initialized parameters. To the best of our knowledge, this is the first adversarial attack on FR that utilizes parameter augmentation to enhance transferability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.409, + 0.909, + 0.544 + ], + "angle": 0, + "content": "- We introduce a new adversarial attack method on FR, called DPA, which comprises DPO and HMA stages. The DPO stage uses both pre-trained and random initializations to optimize the model and save the models in the intermediate training process to diversify the surrogate model set. The HMA stage adds beneficial perturbations with optimization directions opposite to those of adversarial perturbations onto the feature maps of the diversified surrogate models to further enhance transferability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.545, + 0.909, + 0.589 + ], + "angle": 0, + "content": "- Extensive experiments reveal that our proposed method attains superior performance when compared with the state-of-the-art adversarial attack methods." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.318, + 0.909, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.604, + 0.656, + 0.62 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Adversarial Attacks. The primary objective of adversarial attacks is to introduce subtle perturbations into benign images, thereby deceiving machine learning systems and inducing them to produce incorrect predictions [11, 42]. The presence of adversarial examples constitutes a substantial security threat to contemporary machine learning systems. Consequently, significant research efforts have been expended to investigate adversarial attacks, with the aim of bolstering system robustness [10, 23, 27, 28, 30, 36, 66, 78]. To enhance the potency of black-box adversarial attacks, DI [59] incorporates random transformations into adversarial examples at each iteration, effectively achieving data augmentation. VMI-FGSM [50] harnesses gradient variance to stabilize the update process, thereby enhancing black-box attack performance. SSA [27] translates adversarial examples into the frequency domain and applies spectral manipulation for augmentation. SIA [52] introduces random image transformations on each image block, producing di" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3517" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.891, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.377, + 0.908, + 0.421 + ], + "angle": 0, + "content": "Figure 2. The framework of the Diverse Parameters Optimization (DPO). We enhance the diversity of the surrogate model parameters by integrating both pre-trained and random initializations. The method yields a diverse set of surrogate model parameters, which enhances the parameter diversity of the surrogate FR models and consequently improves transferability of the crafted adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.441, + 0.484, + 0.593 + ], + "angle": 0, + "content": "verse variations for gradient estimation. BSR [46] segments the input image into multiple blocks, randomly shuffling and rotating them to generate a set of new images for gradient computation. Despite these advancements, these methods overlook the potential benefit of augmenting the surrogate model with diverse initializations, which limits the transferability of the crafted adversarial examples. In contrast, our proposed method augments surrogate models using both pre-trained and randomly initialized parameters, producing a broader and more diverse set of surrogate models." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.599, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Adversarial Attacks on Face Recognition. Adversarial attacks on FR models can be classified based on the constraints imposed on the adversarial perturbations. These attacks are broadly categorized into two types: restricted attacks [4, 9, 25, 26, 28, 60, 72, 79] and unrestricted attacks [3, 5, 37, 40, 43, 54, 55, 64]. Restricted attacks create adversarial examples that adhere to a specified perturbation limit, such as an \\( L_{p} \\) norm bound. Our proposed attack method is a restricted attack. Consequently, we will delve into the specifics of restricted attacks in detail. To enhance the transferability of adversarial attacks on FR, Zhong and Deng [69] introduced DFANet, which employs dropout on the feature maps of convolutional layers to achieve an ensemble-like effect. Zhou et al. [71] proposed BPFA, enhancing attack transferability by integrating beneficial perturbations [56] onto the feature maps of FR models, resulting in the effect of hard model augmentation. Li et al. [22] leveraged additional information from FR-related tasks and applied a multi-task optimization framework to further improve the transferability of adversarial examples. Unrestricted adversarial attacks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.909, + 0.639 + ], + "angle": 0, + "content": "conversely generate adversarial examples without the limitations of a predefined perturbation bound. These attacks are primarily focused on physical attacks [21, 58, 62], attribute editing [18, 33], and adversarial example generation through makeup transfer [15, 35, 39, 63]. Both restricted and unrestricted adversarial attacks on FR models have substantially advanced the capabilities of these attacks. However, existing methods for generating adversarial attacks on FR often rely on surrogate models with fixed parameters, which limits the transferability of the crafted adversarial examples. In contrast, our proposed attack method addresses this limitation by diversifying the surrogate model parameters, using models with varying parameters across different epochs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.656, + 0.651, + 0.674 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.847, + 0.699 + ], + "angle": 0, + "content": "3.1. Problem Formulation and Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.907, + 0.844 + ], + "angle": 0, + "content": "Problem Formulation. Let \\(\\mathcal{F}^{vct}(\\mathbf{x}') \\in \\mathbb{R}^r\\) denote the FR model employed by the victim to extract the embedding from a face image \\(\\mathbf{x}'\\). We denote \\(\\mathbf{x}^s\\) and \\(\\mathbf{x}^t\\) as the source and target images, respectively. The objective of the adversarial attacks explored in our research is to manipulate \\(\\mathcal{F}^{vct}(\\mathbf{x})\\) to misclassify the adversarial example \\(\\mathbf{x}^{adv}\\) as the target image \\(\\mathbf{x}^t\\), while ensuring that \\(\\mathbf{x}^{adv}\\) bears a close visual resemblance to \\(\\mathbf{x}^s\\). The objective of adversarial attack as delineated in this paper is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.856, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {x} ^ {a d v} = \\underset {\\mathbf {x} ^ {a d v}} {\\arg \\min } \\left(\\mathcal {D} \\left(\\mathcal {F} ^ {v c t} \\left(\\mathbf {x} ^ {a d v}\\right), \\mathcal {F} ^ {v c t} \\left(\\mathbf {x} ^ {t}\\right)\\right)\\right) \\tag {1} \\\\ \\mathrm {s . t .} \\| \\mathbf {x} ^ {a d v} - \\mathbf {x} ^ {s} \\| _ {p} \\leqslant \\epsilon \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3518" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "where the symbol \\(\\mathcal{D}\\) denotes a predefined distance metric employed for the optimization of adversarial face examples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.242 + ], + "angle": 0, + "content": "In practice, the attacker typically cannot access the model owned by the victim. As a result, black-box attacks play a crucial role in adversarial attacks [32, 48, 49]. A widely adopted approach for black-box attacks involves using a surrogate model \\(\\mathcal{F}\\) to generate adversarial examples and then transfer the crafted adversarial examples to the victim model [50, 71]. Consequently, the transferability is important and constitutes the key problem studied in this research." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.485, + 0.469 + ], + "angle": 0, + "content": "Framework Overview. To enhance the transferability of adversarial examples, we propose a novel method called Diverse Parameters Augmentation (DPA). DPA augments surrogate models by incorporating both pre-trained and randomly initialized parameters, resulting in a more diverse and expansive set of surrogate models. Specifically, DPA consists of two key stages: Diverse Parameters Optimization (DPO) and Hard Model Aggregation (HMA). In the DPO stage, we diversify the surrogate model parameters by combining pre-trained and random initializations, as illustrated in Figure 2. In the HMA stage, we apply beneficial perturbations to the feature maps of the diversified surrogate models and combine them to achieve a higher degree of augmentation, as shown in Figure 3. In the following, we will introduce our proposed DPA attack method in detail." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.478, + 0.384, + 0.493 + ], + "angle": 0, + "content": "3.2. Diverse Parameters Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.484, + 0.68 + ], + "angle": 0, + "content": "The parameter diversity of surrogate models is crucial for the transferability of crafted adversarial examples. Previous parameter-based augmentation adversarial attack methods solely initialize parameters with pre-trained values, thereby limiting the parameter diversity of surrogate models [12, 57]. In contrast, the key innovation of the stage of our proposed attack method is the diversification of initialized parameters using both pre-trained and randomly initialized parameters. Following initialization, we augment the parameters of the surrogate models using the parameters in the intermediate training process to diversify the surrogate models, thereby improving the transferability of the adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.484, + 0.772 + ], + "angle": 0, + "content": "Let \\(\\mathbf{x}\\) be a batch of face images, \\(\\mathbf{w} \\in \\mathbb{R}^{s \\times r}\\) be the parameters of the FR head, \\(b\\) be the batch size of \\(\\mathbf{x}\\), and \\(s\\) be the class number in the training dataset. To calculate the loss function, we should first calculate the cosine similarity matrix \\(\\mathbf{r} \\in \\mathbb{R}^{b \\times s}\\) between \\(\\mathcal{F}(\\mathbf{x}) \\in \\mathbb{R}^{b \\times r}\\) and the parameters of the FR head \\(\\mathbf{w}^{\\top}\\) using the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.78, + 0.483, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} = \\cos \\mathbf {a} = \\bar {\\mathbf {e}} \\bar {\\mathbf {w}} ^ {\\top} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.808, + 0.484, + 0.875 + ], + "angle": 0, + "content": "where \\(\\overline{\\mathbf{e}} = \\frac{\\mathcal{F}(\\mathbf{x})}{\\|\\mathcal{F}(\\mathbf{x})\\|}, \\overline{\\mathbf{w}} = \\frac{\\mathbf{w}}{\\|\\mathbf{w}\\|}\\) and \\(\\mathbf{a}\\) is the angle between \\(\\mathcal{F}(\\mathbf{x})\\) and \\(\\mathbf{w}^{\\top}\\). After obtaining the cosine similarity matrix by Equation (2), we can get the sine similarity between \\(\\mathcal{F}(\\mathbf{x})\\) and \\(\\mathbf{w}\\) using the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.884, + 0.484, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\sin \\mathbf {a} = \\sqrt {1 . 0 - \\cos^ {2} \\mathbf {a}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.136 + ], + "angle": 0, + "content": "Using the cosine and sine similarity matrix, the formula to get the additive angular margin cosine similarity matrix can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.147, + 0.907, + 0.163 + ], + "angle": 0, + "content": "\\[\n\\cos (\\mathbf {a} + m) = \\cos \\mathbf {a} \\cos m - \\sin \\mathbf {a} \\sin m \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.172, + 0.907, + 0.233 + ], + "angle": 0, + "content": "where \\(m\\) is the margin value for increasing the discrimination of the FR model [8]. Let \\(\\mathbf{y} \\in \\mathbb{R}^b\\) be the labels of the current batch. By utilizing \\(\\mathbf{y}\\), we can calculate its one-hot encoded matrix using the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.24, + 0.907, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathbf {h}} = \\Psi (\\mathbf {y}) \\in \\mathbb {R} ^ {b \\times s} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.268, + 0.906, + 0.299 + ], + "angle": 0, + "content": "where \\(\\Psi (\\cdot)\\) is the one-hot encode operation. Next, we use the following function to calculate the output of the head:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.306, + 0.906, + 0.367 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {q} = d \\left(\\hat {\\mathbf {h}} \\odot \\mathbf {p} + (1 - \\hat {\\mathbf {h}}) \\odot \\mathbf {r}\\right) \\\\ \\mathbf {p} = \\left\\{ \\begin{array}{l l} \\cos (\\mathbf {a} + m) & \\text {s . t .} \\quad \\mathbf {a} < \\pi - m, \\\\ \\cos \\mathbf {a} - m \\sin m & \\text {s . t .} \\quad \\mathbf {a} \\geqslant \\pi - m, \\end{array} \\right. \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.421 + ], + "angle": 0, + "content": "where \\(d\\) is a pre-defined scale factor. Using the output of the head, the formula to calculate the loss function can be expressed as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.43, + 0.907, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathbf {x}, \\mathbf {y}) = - \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\log \\frac {e ^ {\\mathbf {q} _ {i , \\mathbf {y} _ {i}}}}{\\sum_ {j = 1} ^ {s} e ^ {\\mathbf {q} _ {i , j}}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.477, + 0.907, + 0.523 + ], + "angle": 0, + "content": "Let \\( c \\) and \\( \\mathbf{v} \\) be the number of the training epoch and the parameters of the FR backbone, respectively. Utilizing Equation (7), we can get the parameters set [8]:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.531, + 0.892, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {i}, \\mathbf {w} _ {i} = \\varphi \\left(\\underset {\\mathbf {v}, \\mathbf {w}} {\\arg \\min } \\mathcal {L} \\left(\\mathbf {x}, \\mathbf {y}\\right), i\\right) \\quad i \\in \\{1, 2, \\ldots , c \\}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.568, + 0.924, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} _ {c} = \\left\\{\\mathbf {v} _ {1}, \\mathbf {v} _ {2}, \\dots , \\mathbf {v} _ {c - 1}, \\mathbf {v} _ {c} \\right\\} \\quad \\mathbf {W} _ {c} = \\left\\{\\mathbf {w} _ {1}, \\mathbf {w} _ {2}, \\dots , \\mathbf {w} _ {c - 1}, \\mathbf {w} _ {c} \\right\\} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.598, + 0.907, + 0.704 + ], + "angle": 0, + "content": "where \\(\\varphi (\\cdot ,i)\\) is the operation that obtain the \\(\\mathbf{v}\\) and \\(\\mathbf{w}\\) parameters in the \\(i\\)th epoch. We denote \\(\\mathbf{v}_0\\) and \\(\\mathbf{w}_0\\) as the initialized parameters of the FR model and head, respectively. If \\(\\mathbf{v}_0\\) and \\(\\mathbf{w}_0\\) are fixed, \\(\\mathbf{V}_c\\) and \\(\\mathbf{W}_c\\) become deterministic aside from minor random factors in the training process of Equation (8). Therefore, there exists a mapping between \\(\\{\\mathbf{V}_c,\\mathbf{W}_c\\}\\) and \\(\\{\\mathbf{v}_0,\\mathbf{w}_0\\}\\) that can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.713, + 0.907, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {V} _ {c}, \\mathbf {W} _ {c} \\right\\} = F \\left(\\left\\{\\mathbf {v} _ {0}, \\mathbf {w} _ {0} \\right\\}, i\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.739, + 0.907, + 0.86 + ], + "angle": 0, + "content": "Existing adversarial attack methods based on parameter augmentation typically initialize the parameters \\(\\{\\mathbf{v}_0,\\mathbf{w}_0\\}\\) solely with pre-trained values [12, 57]. This constraint diminishes the diversity within the augmented surrogate model set. In contrast, we opt to initialize \\(\\{\\mathbf{v}_0,\\mathbf{w}_0\\}\\) with diverse parameters to enhance the diversity of the augmented surrogate model set, thereby improving transferability. The \\(\\{\\mathbf{V}_c,\\mathbf{W}_c\\}\\) of our proposed attack can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.868, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left\\{\\mathbf {V} _ {c} ^ {p}, \\mathbf {W} _ {c} ^ {p} \\right\\} = F \\left(\\left\\{\\mathbf {v} _ {0} ^ {p}, \\mathbf {w} _ {0} ^ {p} \\right\\}, i\\right) \\tag {10} \\\\ \\left\\{\\mathbf {V} _ {c} ^ {a}, \\mathbf {W} _ {c} ^ {a} \\right\\} = F \\left(\\left\\{\\mathbf {v} _ {0} ^ {a}, \\mathbf {w} _ {0} ^ {a} \\right\\}, i\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3519" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.09, + 0.895, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.217, + 0.908, + 0.26 + ], + "angle": 0, + "content": "Figure 3. The framework of the Hard Model Aggregation (HMA). After acquiring a surrogate model set with diverse parameters (i.e., \\(\\mathbf{V}_c^q\\)), we introduce beneficial perturbations with the optimization direction opposite to that of adversarial perturbations onto the feature maps of these diversified surrogate models, transforming them into hard models and aggregate the hard models to increase the transferability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.279, + 0.484, + 0.356 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_0^p\\) represents the pre-trained parameters, while \\(\\mathbf{w}_0^p\\), \\(\\mathbf{v}_0^a\\) and \\(\\mathbf{w}_0^a\\) are randomly initialized. Using Equation (10), we can obtain the more diverse parameters for the FR model and head. After obtaining \\(\\mathbf{V}_c^p\\) and \\(\\mathbf{V}_c^a\\), we use the following set of the parameters to craft the adversarial examples:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.368, + 0.484, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {V} _ {c} ^ {q} = \\left\\{\\mathbf {v} _ {0} ^ {p}, \\mathbf {v} _ {1} ^ {p}, \\mathbf {v} _ {j _ {1}} ^ {p}, \\dots , \\mathbf {v} _ {c} ^ {p} \\right\\} \\bigcup \\left\\{\\mathbf {v} _ {1} ^ {a}, \\mathbf {v} _ {j _ {1}} ^ {a}, \\mathbf {v} _ {j _ {2}} ^ {a}, \\dots , \\mathbf {v} _ {c} ^ {a} \\right\\} \\tag {11} \\\\ \\mathrm {s . t .} j \\in \\{1, 2, \\dots , c | j \\bmod \\kappa = 1 \\} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.484, + 0.482 + ], + "angle": 0, + "content": "where \\(\\kappa = \\lfloor \\sqrt{c} \\rfloor\\) determines the epoch interval to select the parameters. The overall framework and pseudo-code of the DPO stage in our proposed method is illustrated in Figure 2 and Algorithm 1, respectively." + }, + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.497, + 0.449, + 0.513 + ], + "angle": 0, + "content": "Algorithm 1 Diverse Parameters Optimization (DPO)" + }, + { + "type": "code", + "bbox": [ + 0.093, + 0.516, + 0.484, + 0.564 + ], + "angle": 0, + "content": "Input: Diverse initial parameters \\(\\mathbf{v}_0^p\\), \\(\\mathbf{w}_0^p\\), \\(\\mathbf{v}_0^a\\), and \\(\\mathbf{w}_0^a\\), the number of epochs \\(c\\), the optimizer \\(\\mathbf{T}\\), the FR training dataset \\(\\zeta\\), loss function \\(\\mathcal{L}\\)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.564, + 0.357, + 0.578 + ], + "angle": 0, + "content": "Output: Augmented parameter set \\(\\mathbf{V}_c^q\\)." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.578, + 0.407, + 0.595 + ], + "angle": 0, + "content": "1: \\(\\mathbf{P} = \\{\\mathbf{v}_0^p, \\mathbf{w}_0^p\\}\\), \\(\\mathbf{A} = \\{\\mathbf{v}_0^a, \\mathbf{w}_0^a\\}\\), \\(\\mathbf{V}_c^q = \\{\\mathbf{v}_0^p\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.594, + 0.246, + 0.609 + ], + "angle": 0, + "content": "2: for \\(\\mathbf{J} \\in \\{\\mathbf{P}, \\mathbf{A}\\}\\) do" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.608, + 0.217, + 0.624 + ], + "angle": 0, + "content": "3: \\(\\{\\mathbf{v},\\mathbf{w}\\} = \\mathbf{J}\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.624, + 0.476, + 0.639 + ], + "angle": 0, + "content": "4: \\(\\mathcal{F} = \\Psi (\\mathbf{v})\\quad \\triangleright\\) Map the parameters to the models." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.639, + 0.256, + 0.653 + ], + "angle": 0, + "content": "5: for \\(i = 1,\\dots,c\\) do" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.655, + 0.239, + 0.668 + ], + "angle": 0, + "content": "6: for \\(\\mathbf{x}\\in \\zeta\\) do" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.669, + 0.378, + 0.684 + ], + "angle": 0, + "content": "7: Calculate \\(\\mathcal{L}\\) using Equation (7)." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.684, + 0.476, + 0.699 + ], + "angle": 0, + "content": "8: Backward \\((\\mathbf{T},\\mathcal{L})\\triangleright\\) Backpropagation using \\(\\mathbf{T}\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.7, + 0.358, + 0.715 + ], + "angle": 0, + "content": "9: Update the parameters \\(\\mathbf{v},\\mathbf{w}\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.715, + 0.206, + 0.728 + ], + "angle": 0, + "content": "0: end for" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.729, + 0.484, + 0.759 + ], + "angle": 0, + "content": "11: if the \\(\\mathbf{v}\\) satisfies the condition outlined in Equation (11) then" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.759, + 0.484, + 0.775 + ], + "angle": 0, + "content": "12: Add \\((\\mathbf{V}_{c}^{q},\\mathbf{v})\\) Incorporate \\(\\mathbf{v}\\) into \\(\\mathbf{V}_c^q\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.776, + 0.197, + 0.789 + ], + "angle": 0, + "content": "13: end if" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.79, + 0.19, + 0.804 + ], + "angle": 0, + "content": "14: end for" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.806, + 0.174, + 0.818 + ], + "angle": 0, + "content": "15: end for" + }, + { + "type": "list", + "bbox": [ + 0.099, + 0.578, + 0.484, + 0.818 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.321, + 0.866 + ], + "angle": 0, + "content": "3.3. Hard Model Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Hard models [71] refer to models with higher loss values compared to normal models, designed to simulate the effect" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.281, + 0.907, + 0.371 + ], + "angle": 0, + "content": "of hard samples, which are hard to the optimization process of the adversarial examples. Since hard samples can improve the generalization of models, we leverage multiple hard models to enhance the transferability of the surrogate models with diverse parameters by adding beneficial perturbations [56] onto the pre-defined feature maps of the surrogate models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.906, + 0.416 + ], + "angle": 0, + "content": "In the following, we will provide a detailed introduction to the HMA stage. Let \\(\\Psi\\) be the mapping from the parameters to the corresponding models that can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.424, + 0.907, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} = \\Psi \\left(\\mathbf {V} _ {c} ^ {q}\\right) = \\left\\{\\mathcal {F} _ {1}, \\mathcal {F} _ {2}, \\dots , \\mathcal {F} _ {g} \\right\\} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.909, + 0.599 + ], + "angle": 0, + "content": "where \\( g = |\\mathbf{V}_c^q| \\). Hard sample augmentation has demonstrated significant effectiveness in enhancing model generalization [41, 67]. Based on the relationship between adversarial attack tasks and training tasks [9, 24], using hard models for augmentation can lead to improved transferability [71]. Consequently, by transforming vanilla parameter-augmented surrogate models into hard parameter-augmented surrogate models, we can achieve more transferable results. To accomplish this transformation, we use the following formula to calculate the loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.604, + 0.906, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widetilde {\\mathcal {L}} _ {t} = \\frac {1}{g} \\sum_ {i = 0} ^ {g} \\left\\| \\phi \\left(\\mathcal {H} _ {i} \\left(\\mathcal {T} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right)\\right)\\right) - \\phi \\left(\\mathcal {F} _ {i} \\left(\\mathbf {x} ^ {t}\\right)\\right) \\right\\| \\tag {13} \\\\ \\begin{array}{l l} \\text {s . t .} & \\mathcal {F} _ {i} \\in \\mathbf {F} \\quad t \\in \\{1, 2, \\dots , n \\} \\end{array} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.909, + 0.761 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_t^{adv}\\) denotes the adversarial example generated during the \\(t\\)th iteration, the variable \\(n\\) represents the maximum number of iterations allocated for crafting the adversarial examples, \\(\\phi(\\cdot)\\) denotes the normalization operation, \\(\\mathcal{T}\\) signifies the input transformation, and \\(\\mathcal{H}\\) is the corresponding hard model [71] of \\(\\mathcal{F}\\), which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.767, + 0.907, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right) = \\left\\{ \\begin{array}{l l} \\mathcal {F} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right) & \\text {s . t .} \\quad t = 1, \\\\ \\chi \\left(\\mathcal {F} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right), \\Omega\\right) & \\text {s . t .} \\quad t > 1, \\end{array} \\right. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.908, + 0.852 + ], + "angle": 0, + "content": "where \\(\\chi\\) is the mapping for adding beneficial perturbations on the feature maps in the forward propagation whose formula can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.859, + 0.907, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\chi \\left(\\mathcal {F} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right), \\Omega\\right): \\omega = \\omega + \\eta \\text {s i g n} \\left(\\nabla_ {\\omega} \\tilde {\\mathcal {L}} _ {t - 1}\\right) \\tag {15} \\\\ \\begin{array}{c c} \\text {s . t .} & \\omega \\in \\Omega \\end{array} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3520" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "where \\(\\Omega\\) is the pre-defined set of layers for adding beneficial perturbations. Let \\(\\epsilon\\) be the maximum allowable perturbation. Utilizing \\(\\tilde{\\mathcal{L}}_t\\), we can craft the adversarial face examples using following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.162, + 0.484, + 0.196 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t + 1} ^ {a d v} = \\prod_ {\\mathbf {x} ^ {s}, \\epsilon} \\left(\\mathbf {x} _ {t} ^ {a d v} - \\beta \\operatorname {s i g n} \\left(\\nabla_ {\\mathbf {x} _ {t} ^ {a d v}} \\tilde {\\mathcal {L}} _ {t}\\right)\\right) \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.206, + 0.484, + 0.281 + ], + "angle": 0, + "content": "where \\(\\beta\\) denotes the step size used to construct the adversarial examples, while \\(\\prod\\) represents the clipping operation that confines the pixel values of the generated adversarial examples within the range \\([\\mathbf{x}^s -\\epsilon ,\\mathbf{x}^s +\\epsilon ]\\). The framework of the HMA is illustrated in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.484, + 0.417 + ], + "angle": 0, + "content": "In the following, we present the pseudo-code for the HMA. For the sake of clarity, we focus on neural networks with a single branch in their computational graphs. In the case of neural networks with multiple branches, the HMA algorithm remains largely unchanged, except for the incorporation of mechanisms to handle multiple branches. Before proceeding, we introduce some necessary definitions. Let \\(\\Phi\\) denote the index set of pre-selected layers to which beneficial perturbations will be added:" + }, + { + "type": "equation", + "bbox": [ + 0.248, + 0.429, + 0.483, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\Phi = \\varkappa (\\Omega) \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.547 + ], + "angle": 0, + "content": "where \\(\\varkappa\\) represents the mapping from the pre-defined set of layers for adding beneficial perturbations \\(\\Omega\\), to their corresponding layer index set \\(\\Phi\\). Let \\(f^i\\) denote the \\(i\\)th layer within the model \\(\\mathcal{F}\\), and let \\(z\\) represent the total number of layers in \\(\\mathcal{F}\\). We define the segment of \\(\\mathcal{F}\\) from layer \\(f^i\\) to layer \\(f^j\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.56, + 0.483, + 0.578 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} ^ {i, j} = f ^ {i} \\circ f ^ {i + 1} \\circ \\dots \\circ f ^ {j - 1} \\circ f ^ {j} \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.615 + ], + "angle": 0, + "content": "The pseudo-code for the HMA stage of our proposed method is presented in Algorithm 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.628, + 0.222, + 0.645 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.291, + 0.669 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.795 + ], + "angle": 0, + "content": "Datasets. We opt to use the LFW [16], and CelebA-HQ [19] for our experiments to verify the effectiveness of our proposed attack method. LFW serves as an unconstrained face dataset for FR. CelebA-HQ consists of face images with high visual quality. The LFW and CelebA-HQ utilized in our experiments are identical to those employed in [70, 71, 73]. For the parameter augmentation process, we select BUPT-Balancedface [47] as our training dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Face Recognition Models. The normal trained FR models employed in our experiments encompass IR152 [13], IRSE50 [14], FaceNet [34], MobileFace [8], CurricularFace [17], MagFace [29], ArcFace [8], CircleLoss [38], MV-Softmax [51], and NPCFace [65]. Specifically, IR152, FaceNet, IRSE50, and MobileFace are the same models utilized in [15, 63, 70, 71]. CurricularFace, MagFace, ArcFace," + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.092, + 0.825, + 0.107 + ], + "angle": 0, + "content": "Algorithm 2 Hard Model Aggregation (HMA)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.112, + 0.907, + 0.247 + ], + "angle": 0, + "content": "Input: The source image \\(\\mathbf{x}^s\\), the target image \\(\\mathbf{x}^t\\), the mapping from the parameters to the corresponding models \\(\\Psi\\), the maximum number of iterations \\(n\\), the index set of pre-selected layers to be added beneficial perturbations \\(\\Phi\\), the step size of the beneficial perturbations \\(\\eta\\), the step size of the adversarial perturbations \\(\\beta\\), the maximum permissible magnitude of perturbation \\(\\epsilon\\), the parameter set \\(\\mathbf{V}_c^q\\), the total number of layers in a single surrogate model \\(z\\), normalization operation \\(\\phi\\)." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.248, + 0.772, + 0.264 + ], + "angle": 0, + "content": "Output: An adversarial example \\(\\mathbf{x}_n^{adv}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.264, + 0.798, + 0.279 + ], + "angle": 0, + "content": "1: \\(\\mathbf{x}_0^{adv} = \\mathbf{x}^s\\), \\(u = |\\Phi|, s_1 = 1, g = |\\mathbf{V}_c^q|\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.279, + 0.9, + 0.293 + ], + "angle": 0, + "content": "2: \\(\\mathbf{F} = \\Psi (\\mathbf{V}_c^q)\\) \\(\\triangleright\\) Acquire the diversified models set." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.294, + 0.666, + 0.307 + ], + "angle": 0, + "content": "3: for \\( t = 1, \\dots, n \\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.309, + 0.68, + 0.323 + ], + "angle": 0, + "content": "4: for \\(i = 1,\\dots,g\\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.324, + 0.905, + 0.351 + ], + "angle": 0, + "content": "5: \\(\\mathcal{F} = \\mathbf{F}_i\\) Derive the parameter-augmented model." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.353, + 0.688, + 0.37 + ], + "angle": 0, + "content": "6: \\(\\omega^{t,0} = \\mathcal{T}\\left(\\mathbf{x}_{t - 1}^{adv}\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.371, + 0.7, + 0.383 + ], + "angle": 0, + "content": "7: for \\(j = 1,\\dots,u\\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.385, + 0.68, + 0.399 + ], + "angle": 0, + "content": "8: \\(s_2 = \\Phi_j\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.399, + 0.745, + 0.416 + ], + "angle": 0, + "content": "9: \\(\\omega^{t,j} = \\mathcal{F}^{s_1,s_2}\\left(\\omega^{t,j - 1}\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.417, + 0.646, + 0.429 + ], + "angle": 0, + "content": "10: \\(s_1 = s_2\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.43, + 0.68, + 0.443 + ], + "angle": 0, + "content": "11: if \\( t \\neq 1 \\) then" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.444, + 0.847, + 0.466 + ], + "angle": 0, + "content": "12: \\(\\omega^{t,j} = \\omega^{t,j} + \\eta \\mathrm{sign}\\left(\\nabla_{\\omega^{t - 1,j}}\\widetilde{\\mathcal{L}}_{t - 1}\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.467, + 0.636, + 0.478 + ], + "angle": 0, + "content": "13: end if" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.481, + 0.63, + 0.492 + ], + "angle": 0, + "content": "14: end for" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.493, + 0.842, + 0.511 + ], + "angle": 0, + "content": "15: \\(\\tilde{\\mathcal{L}}^i = \\| \\phi (\\mathcal{F}^{\\Phi_u,z}(\\omega^{t,u})) - \\phi (\\mathcal{F}(\\mathbf{x}^t))\\| _2^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.511, + 0.612, + 0.523 + ], + "angle": 0, + "content": "16: end for" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.523, + 0.904, + 0.554 + ], + "angle": 0, + "content": "17: \\(\\widetilde{\\mathcal{L}}_t = \\frac{1}{g}\\sum_{i = 1}^g\\widetilde{\\mathcal{L}}^i\\) Compute the loss function utilizing the parameter-augmented models." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.554, + 0.83, + 0.582 + ], + "angle": 0, + "content": "18: \\(\\mathbf{x}_t^{adv} = \\prod_{\\mathbf{x}^s,\\epsilon}\\left(\\mathbf{x}_{t - 1}^{adv} - \\beta \\mathrm{sign}\\left(\\nabla_{\\mathbf{x}_{t - 1}^{adv}}\\widetilde{\\mathcal{L}}_t\\right)\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.583, + 0.596, + 0.595 + ], + "angle": 0, + "content": "19: end for" + }, + { + "type": "list", + "bbox": [ + 0.521, + 0.264, + 0.905, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.702 + ], + "angle": 0, + "content": "CircleLoss, MV-Softmax, and NPCFace are the official models provided by FaceX-ZOO [45]. Furthermore, we integrate adversarial robust FR models into our experiments, denoted as IR152\\(^{adv}\\), IRSE50\\(^{adv}\\), FaceNet\\(^{adv}\\), and MobileFace\\(^{adv}\\) which correspond to the models used in [71]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Attack Setting. We set the maximum allowable perturbation magnitude \\(\\epsilon\\) to 10, based on the \\(L_{\\infty}\\) norm bound, without any specific emphasis. In addition, we specify the maximum number of iterative steps as 200. More detailed attack settings are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.884 + ], + "angle": 0, + "content": "Evaluation Metrics. We utilize Attack Success Rate (ASR) as the metric to assess the efficacy of various adversarial attacks. The ASR represents the ratio of adversarial examples that successfully evade the victim model to the total number of adversarial examples generated. In calculating the ASR, we determine the threshold based on a FAR@0.001 on the entire LFW dataset for each victim model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Compared methods. Our proposed attack is a form of re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3521" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "stricted adversarial attack designed to expose vulnerabilities within FR systems. It would be inequitable to juxtapose our method with unrestricted attacks, which do not impose limits on the magnitude of adversarial perturbations. Consequently, we opt to benchmark our approach against other restricted attacks on FR systems that are explicitly malicious at attacking the systems [69] [71], as well as state-of-the-art transfer attacks [59] [27] [12] [52] [46]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.221, + 0.283, + 0.237 + ], + "angle": 0, + "content": "4.2. Comparison Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.485, + 0.44 + ], + "angle": 0, + "content": "DPA achieves the best black-box attack results on both normally trained and adversarial robust models. To verify the effectiveness of our proposed attack method, we craft the adversarial examples on the LFW and CelebA-HQ dataset. The black-box performance with MobileFace as the surrogate models on the LFW dataset are demonstrated in Table 1. Some of the adversarial examples are demonstrated in Figure 4. More experimental results on the LFW dataset, as well as those on the CelebA-HQ dataset, are provided in the supplementary. These results demonstrate that our proposing method consistently outperforms the baseline attack, thereby highlighting its effectiveness in improving the transferability of adversarial face examples." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.452, + 0.484, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.685, + 0.484, + 0.756 + ], + "angle": 0, + "content": "Figure 4. The illustration of adversarial examples crafted by various attacks. First column: some source images. Last column: the corresponding target images. The second to fifth columns exhibit the corresponding adversarial face examples crafted by SIA [52], BPFA [71], BSR [46], and Our proposed attack, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "DPA achieves superior black-box performance under JPEG compression. JPEG compression is a widely adopted method for image compression during transmission, and it also serves as a defense mechanism against adversarial examples. To evaluate the effectiveness of our proposed attack under JPEG compression, we employ MobileFace as the surrogate model and IRSE50 as the victim model. We assess the attack performance on the LFW and CelebA-HQ datasets, with the results presented in Figure 5. These results indicate" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.908, + 0.132 + ], + "angle": 0, + "content": "Table 1. Comparisons of black-box ASR \\((\\%)\\) results for attacks using MobileFace as the surrogate model on the LFW dataset. I, S, F, M denote IR152, IRSE50, FaceNet, and MobileFace, respectively." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.137, + 0.903, + 0.268 + ], + "angle": 0, + "content": "
AttacksISFIadvSadvFadvMadv
FIM [68]5.373.47.52.54.52.810.9
DI [59]18.497.332.910.618.210.238.6
DFANet [69]7.086.411.93.66.43.916.8
VMI [50]13.696.020.27.612.87.632.3
SSA [27]13.896.419.65.513.17.231.5
SIA [52]15.796.826.78.114.59.135.3
BPFA [71]15.895.517.85.911.05.128.4
BSR [46]5.474.29.52.95.94.911.4
Ours67.798.290.855.466.442.671.6
" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.286, + 0.712, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.286, + 0.895, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.416, + 0.907, + 0.444 + ], + "angle": 0, + "content": "Figure 5. Performance of ASR across various JPEG Q values: (a) Results on the LFW dataset. (b) Results on the CelebA-HQ dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.457, + 0.908, + 0.513 + ], + "angle": 0, + "content": "Table 2. Comparison of black-box ASR (%) results using the parameter-based augmented adversarial attack method as the baseline on the LFW dataset. \\( \\mathrm{ASR}^{adv} \\) represents the average attack success rate on adversarial robust models." + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.518, + 0.89, + 0.569 + ], + "angle": 0, + "content": "
IR152IRSE50FaceNetASRadv
Baseline4.819.615.77.2
Ours67.798.290.859.0
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.658 + ], + "angle": 0, + "content": "that our proposed attack method consistently outperforms the baseline attack methods across varying levels of JPEG compression, thereby underscoring the robustness of our proposed attack method under such conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.901 + ], + "angle": 0, + "content": "DPA demonstrates better black-box performance compared to the parameter-based augmented adversarial attack. The LGV method is an effective parameter-based augmented adversarial attack technique designed to enhance transferability. Our proposed attack incorporates parameter augmentation, making LGV an appropriate baseline. We selected MobileFace as the surrogate model and generated adversarial examples on the LFW dataset. The results are presented in Table 2. Table 2 clearly shows that our proposed attack outperforms the baseline, further validating the effectiveness of our method. The primary limitation of LGV stems from its random selection of one model per iteration, which introduces instability and adversely impacts blackbox ASR [9, 50]. Furthermore, LGV suffers from a lack of diverse initialization parameters. These shortcomings significantly impair the overall ASR performance of LGV." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3522" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.092, + 0.473, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.255, + 0.483, + 0.283 + ], + "angle": 0, + "content": "Figure 6. ASR on victim models trained with various algorithms, using FaceNet as the surrogate model on the LFW dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.486, + 0.473 + ], + "angle": 0, + "content": "DPA achieves superior black-box performance on FR models trained with various algorithms. In order to substantiate the effectiveness of our proposed attack method across various FR models, we conducted additional experiments using BSR [46] as the Baseline and FaceNet as the surrogate model. These experiments adhered to the same settings detailed in Table 7 in the supplementary. The ASR across multiple FR models is illustrated in Figure 6. As shown in Figure 6, the ASR of the adversarial examples generated by our proposed method exceeds that of the Baseline, thereby reinforcing the efficacy of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.485, + 0.255, + 0.5 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.486, + 0.78 + ], + "angle": 0, + "content": "The ablation studies on each stage of our proposed attack method. Our proposed attack method consists of two stages. To verify the effectiveness of each stage, we conduct ablation studies on the stages. Initially, we craft adversarial examples using only the surrogate model, which we denote as 'Vanilla'. Next, we generate adversarial examples using the ensemble of surrogate models obtained in the DPO stage (i.e. F), denoted as DPO. After incorporating the HMA stage, the complete attack method is denoted as 'DPO + HMA'. The results are presented in Table 3. Table 3 demonstrates that the addition of the DPO stage results in an increase in black-box ASR, showcasing the effectiveness of the DPO stage in enhancing transferability. Further incorporation of the HMA stage leads to an additional improvement in attack performance, underscoring the effectiveness of the HMA stage in boosting black-box performance. These results collectively demonstrate the effectiveness of each stage in our proposed attack method in improving the transferability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The ablation studies on the effectiveness of multiple surrogate models at different epochs in improving transferability. Since our proposed attack method utilizes models from intermediate training epochs to craft adversarial examples, it is essential to verify the effectiveness of this approach compared to using only models from the final training epoch. We conduct ablation experiments on these two approaches using the LFW dataset and MobileFace as the surrogate" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 3. Ablation study of black-box ASR (\\%) results on the LFW dataset using MobileFace as the surrogate model. \\( \\mathrm{ASR}^{adv} \\) denotes the average attack success rate on adversarial robust models." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.137, + 0.907, + 0.203 + ], + "angle": 0, + "content": "
AttacksIR152IRSE50FaceNetASRadv
Vanilla5.373.47.55.2
DPO35.991.367.932.9
DPO + HMA67.798.290.859.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.215, + 0.907, + 0.271 + ], + "angle": 0, + "content": "Table 4. Ablation study of black-box ASR (\\%) results to verify the effectiveness of multiple surrogate models at different epochs in improving transferability. \\(\\mathrm{ASR}^{adv}\\) denotes the average attack success rate on adversarial robust models." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.277, + 0.892, + 0.328 + ], + "angle": 0, + "content": "
AttacksIR152IRSE50FaceNetASRadv
Fm20.386.733.617.8
F (Ours)35.991.367.932.9
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.488 + ], + "angle": 0, + "content": "model. We employ the DPO process to obtain the parameter sets \\(\\mathbf{V}_c^q\\) and \\(\\mathbf{V}_c^m = \\{\\mathbf{v}_0^p, \\mathbf{v}_c^p, \\mathbf{v}_c^a\\}\\). Next, we map \\(\\mathbf{V}_c^m\\) into its corresponding model set \\(\\mathbf{F}^m = \\Psi(\\mathbf{V}_c^m)\\). We then craft adversarial examples by ensembling the models from \\(\\mathbf{V}_c^q\\) (i.e. \\(\\mathbf{F}\\)) and \\(\\mathbf{F}^m\\), respectively. The results are presented in Table 4. Table 4 shows that the performance of the ensemble of \\(\\mathbf{F}\\) surpasses that of the ensemble of \\(\\mathbf{F}^m\\), demonstrating the effectiveness of using models from intermediate training epochs to enhance the transferability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.488, + 0.907, + 0.518 + ], + "angle": 0, + "content": "The hyper-parameter analysis studies of our proposed method are in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.532, + 0.634, + 0.547 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.909, + 0.77 + ], + "angle": 0, + "content": "We present an innovative advancement in the field of adversarial attacks on FR through the introduction of the Diverse Parameters Augmentation (DPA) attack method. By addressing the problems of traditional adversarial attacks, particularly the lack of diverse parameters and the exclusion of the FR head, DPA enhances the transferability of adversarial face examples. In addition, DPA emphasizes the importance of integrating both pre-trained and randomly initialized parameters, providing a fresh perspective on augmentation-based adversarial attacks. The extensive experimental results demonstrate the effectiveness of our proposed DPA attack across multiple datasets and target models. We anticipate that our proposed method will illuminate potential directions for advancing research in adversarial attacks on FR." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.779, + 0.686, + 0.796 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.803, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This work was supported in part by the Natural Science Foundation of China under Grant 62372203 and 62302186, in part by the Major Scientific and Technological Project of Shenzhen (202316021), in part by the National key research and development program of China (2022YFB2601802), in part by the Major Scientific and Technological Project of Hubei Province (2022BAA046, 2022BAA042)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3523" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.122, + 0.188, + 0.137 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.148, + 0.486, + 0.23 + ], + "angle": 0, + "content": "[1] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, and Ying Fu. Partial FC: training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 1445-1449, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.233, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[2] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19593-19604, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.304, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[3] Tom B Brown, Nicholas Carlini, Chiyuan Zhang, Catherine Olsson, Paul Christiano, and Ian Goodfellow. Unrestricted adversarial examples. arXiv preprint arXiv:1809.08352, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.484, + 0.431 + ], + "angle": 0, + "content": "[4] Bin Chen, Jia-Li Yin, Shukai Chen, Bohao Chen, and Ximeng Liu. An adaptive model ensemble adversarial attack for boosting adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4466-4475, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.484, + 0.487 + ], + "angle": 0, + "content": "[5] Zhaoyu Chen, Bo Li, Shuang Wu, Kaixun Jiang, Shouhong Ding, and Wenqiang Zhang. Content-based unrestricted adversarial attack. In Advances in Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[6] Valeriia Cherepanova, Micah Goldblum, Harrison Foley, Shiyuan Duan, John P Dickerson, Gavin Taylor, and Tom Goldstein. Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[7] Sowmen Das, Selim Seferbekov, Arup Datta, Md Saiful Islam, and Md Ruhul Amin. Towards solving the deepfake problem: An analysis on improving deepfake detection using dynamic face augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3776-3785, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.484, + 0.714 + ], + "angle": 0, + "content": "[8] Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5962-5979, 2022. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.484, + 0.785 + ], + "angle": 0, + "content": "[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[10] Zhijin Ge, Fanhua Shang, Hongying Liu, Yuanyuan Liu, Liang Wan, Wei Feng, and Xiaosen Wang. Improving the transferability of adversarial examples with arbitrary style transfer. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4440-4449, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.148, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[12] Martin Gubri, Maxime Cordy, Mike Papadakis, Yves Le Traon, and Koushik Sen. Lgv: Boosting adversarial example transferability from large geometric vicinity. In European Conference on Computer Vision, pages 603-618. Springer, 2022. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.165, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[15] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14994–15003, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.907, + 0.431 + ], + "angle": 0, + "content": "[16] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[17] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: Adaptive curriculum learning loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.907, + 0.573 + ], + "angle": 0, + "content": "[18] Shuai Jia, Bangjie Yin, Taiping Yao, Shouhong Ding, Chunhua Shen, Xiaokang Yang, and Chao Ma. Adv-attribute: Inconspicuous and transferable adversarial attack on face recognition. In Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.907, + 0.63 + ], + "angle": 0, + "content": "[19] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representation, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.907, + 0.702 + ], + "angle": 0, + "content": "[20] Jingzhi Li, Zidong Guo, Hui Li, Seungju Han, Ji-Won Baek, Min Yang, Ran Yang, and Sungjoo Suh. Rethinking feature-based knowledge distillation for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20156-20165, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.907, + 0.771 + ], + "angle": 0, + "content": "[21] Yanjie Li, Yiquan Li, Xuelong Dai, Songtao Guo, and Bin Xiao. Physical-world optical adversarial attacks on 3d face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24699-24708, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.907, + 0.856 + ], + "angle": 0, + "content": "[22] Zexin Li, Bangjie Yin, Taiping Yao, Junfeng Guo, Shouhong Ding, Simin Chen, and Cong Liu. Sibling-attack: Rethinking transferable adversarial attacks against face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24626–24637, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[23] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3524" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "from diffusion models via adversarial examples. In International Conference on Machine Learning, pages 20763-20786, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[24] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representation, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[25] Xuannan Liu, Yaoyao Zhong, Yuhang Zhang, Lixiong Qin, and Weihong Deng. Enhancing generalization of universal adversarial perturbation through gradient aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4412-4421, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[26] Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, and Di Ming. TRM-UAP: enhancing the transferability of data-free universal adversarial perturbation via truncated ratio maximization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4739-4748, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[27] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xi-anglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In European Conference on Computer Vision, pages 549–566, 2022. 2, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.484, + 0.487 + ], + "angle": 0, + "content": "[28] Dong Lu, Zhiqiang Wang, Teng Wang, Weili Guan, Hongchang Gao, and Feng Zheng. Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 102-111, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[29] Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14225-14234, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.484, + 0.629 + ], + "angle": 0, + "content": "[30] Duan Mingxing, Kenli Li, Lingxi Xie, Qi Tian, and Bin Xiao. Towards multiple black-boxes attack via adversarial example generation network. In Proceedings of the 29th ACM International Conference on Multimedia, pages 264-272, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.484, + 0.7 + ], + "angle": 0, + "content": "[31] Kartik Narayan, Harsh Agarwal, Kartik Thakral, Surbhi Mittal, Mayank Vatsa, and Richa Singh. Df-platter: Multiface heterogeneous deepfake dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9739-9748, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[32] Xuelin Qian, Wenxuan Wang, Yu-Gang Jiang, Xiangyang Xue, and Yanwei Fu. Dynamic routing and knowledge relearning for data-free black-box attack. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[33] Haonan Qiu, Chaowei Xiao, Lei Yang, Xinchen Yan, Honglak Lee, and Bo Li. Semanticadv: Generating adversarial examples via attribute-conditioned image editing. In European Conference on Computer Vision, pages 19-37, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[34] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 815-823, 2015. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[35] Fahad Shamshad, Muzammal Naseer, and Karthik Nandakumar. Clip2protect: Protecting facial privacy using text-guided" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.134 + ], + "angle": 0, + "content": "makeup via adversarial latent search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20595-20605, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.908, + 0.191 + ], + "angle": 0, + "content": "[36] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[37] Florian Stimberg, Ayan Chakrabarti, Chun-Ta Lu, Hussein Hazimeh, Otilia Stretcu, Wei Qiao, Yintao Liu, Merve Kaya, Cyrus Rashtchian, Ariel Fuxman, Mehmet Tek, and Sven Gowal. Benchmarking robustness to adversarial image obfuscations. In Advances in Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.907, + 0.346 + ], + "angle": 0, + "content": "[38] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.908, + 0.417 + ], + "angle": 0, + "content": "[39] Yuhao Sun, Lingyun Yu, Hongtao Xie, Jiaming Li, and Yongdong Zhang. Diffam: Diffusion-based adversarial makeup transfer for facial privacy protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24584-24594, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.908, + 0.514 + ], + "angle": 0, + "content": "[40] Naufal Suryanto, Yongsu Kim, Harashta Tatimma Larasati, Hyoeun Kang, Thi-Thu-Huong Le, Yoonyoung Hong, Hunmin Yang, Se-Yoon Oh, and Howon Kim. ACTIVE: towards highly transferable 3d physical camouflage for universal and robust vehicle evasion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4282-4291, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.908, + 0.572 + ], + "angle": 0, + "content": "[41] Teppei Suzuki. Teachaugment: Data augmentation optimization using teacher knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10894-10904, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.907, + 0.63 + ], + "angle": 0, + "content": "[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.908, + 0.699 + ], + "angle": 0, + "content": "[43] Donghua Wang, Wen Yao, Tingsong Jiang, Chao Li, and Xiaogian Chen. RFLA: A stealthy reflected light adversarial attack in the physical world. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4442, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.908, + 0.77 + ], + "angle": 0, + "content": "[44] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.908, + 0.801 + ], + "angle": 0, + "content": "[45] Jun Wang, Yinglu Liu, Yibo Hu, Hailin Shi, and Tao Mei. Facex-zoo: A python toolbox for face recognition. 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.908, + 0.87 + ], + "angle": 0, + "content": "[46] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting Adversarial Transferability by Block Shuffle and Rotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[47] Mei Wang, Weihong Deng, Jiani Hu, Xunqiang Tao, and Yaohai Huang. Racial faces in the wild: Reducing racial" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3525" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "bias by information maximization adaptation network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.484, + 0.217 + ], + "angle": 0, + "content": "[48] Wenxuan Wang, Bangjie Yin, Taiping Yao, Li Zhang, Yanwei Fu, Shouhong Ding, Jilin Li, Feiyue Huang, and Xiangyang Xue. Delving into data: Effectively substitute training for black-box attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4761-4770, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.484, + 0.286 + ], + "angle": 0, + "content": "[49] Wenxuan Wang, Xuelin Qian, Yanwei Fu, and Xiangyang Xue. Dst: Dynamic substitute training for data-free black-box attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14361-14370, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.483, + 0.344 + ], + "angle": 0, + "content": "[50] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1924-1933, 2021. 2, 4, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[51] Xiaobo Wang, Shifeng Zhang, Shuo Wang, Tianyu Fu, Hailin Shi, and Tao Mei. Mis-classified vector guided softmax loss for face recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12241-12248, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[52] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4607-4619, 2023. 2, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.484, + 0.525 + ], + "angle": 0, + "content": "[53] Xudong Wang, Ke-Yue Zhang, Taiping Yao, Qianyu Zhou, Shouhong Ding, Pingyang Dai, and Rongrong Ji. Tf-fas: Twofold-element fine-grained semantic guidance for generalizable face anti-spoofing. In European Conference on Computer Vision, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[54] Xingxing Wei, Yao Huang, Yitong Sun, and Jie Yu. Unified adversarial patch for cross-modal attacks in the physical world. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4422-4431, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.484, + 0.637 + ], + "angle": 0, + "content": "[55] Xingxing Wei, Jie Yu, and Yao Huang. Physically adversarial infrared patches with learnable shapes and locations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12334-12342, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[56] Shixian Wen and Laurent Itti. Beneficial perturbations network for defending adversarial examples. arXiv preprint arXiv:2009.12724, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[57] Han Wu, Guanyan Ou, Weibin Wu, and Zibin Zheng. Improving transferable targeted adversarial attacks with model self-enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24615-24624, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[58] Zihao Xiao, Xianfeng Gao, Chilin Fu, Yinpeng Dong, Wei Gao, Xiaolu Zhang, Jun Zhou, and Jun Zhu. Improving transferability of adversarial patches on face recognition with generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11840-11849, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[59] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019. 2, 7, 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[60] Zhuoer Xu, Zhangxuan Gu, Jianping Zhang, Shiwen Cui, Changhua Meng, and Weiqiang Wang. Backpropagation path search on adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4640-4650, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[61] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3877-3887, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.232, + 0.907, + 0.302 + ], + "angle": 0, + "content": "[62] Xiao Yang, Chang Liu, Longlong Xu, Yikai Wang, Yinpeng Dong, Ning Chen, Hang Su, and Jun Zhu. Towards effective adversarial textured 3d meshes on physical face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4119-4128, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.907, + 0.383 + ], + "angle": 0, + "content": "[63] Bangjie Yin, Wenxuan Wang, Taiping Yao, Junfeng Guo, Zelun Kong, Shouhong Ding, Jilin Li, and Cong Liu. Advmakeup: A new imperceptible and transferable attack on face recognition. In Proceedings of the 30th International Joint Conference on Artificial Intelligence, pages 1252-1258, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.907, + 0.441 + ], + "angle": 0, + "content": "[64] Shengming Yuan, Qilong Zhang, Lianli Gao, Yaya Cheng, and Jingkuan Song. Natural color fool: Towards boosting black-box unrestricted attacks. In Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.907, + 0.483 + ], + "angle": 0, + "content": "[65] Dan Zeng, Hailin Shi, Hang Du, Jun Wang, Zhen Lei, and Tao Mei. Npcface: Negative-positive collaborative training for large-scale face recognition. 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.907, + 0.538 + ], + "angle": 0, + "content": "[66] Jiaming Zhang, Qi Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5005-5013, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.907, + 0.58 + ], + "angle": 0, + "content": "[67] Xinyu Zhang, Qiang Wang, Jian Zhang, and Zhao Zhong. Adversarial autoaugment. In International Conference on Learning Representations, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.58, + 0.907, + 0.637 + ], + "angle": 0, + "content": "[68] Yaoyao Zhong and Weihong Deng. Adversarial learning with margin-based triplet embedding regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6548-6557, 2019. 7, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.907, + 0.691 + ], + "angle": 0, + "content": "[69] Yaoyao Zhong and Weihong Deng. Towards transferable adversarial attack against deep face recognition. IEEE Transactions on Information Forensics and Security, 16:1452-1466, 2021. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.907, + 0.774 + ], + "angle": 0, + "content": "[70] Fengfan Zhou, Hefei Ling, Yuxuan Shi, Jiazhong Chen, and Ping Li. Improving visual quality and transferability of adversarial attacks on face recognition simultaneously with adversarial restoration. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4540-4544, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.907, + 0.845 + ], + "angle": 0, + "content": "[71] Fengfan Zhou, Hefei Ling, Yuxuan Shi, Jiazhong Chen, Zongyi Li, and Ping Li. Improving the transferability of adversarial attacks on face recognition with beneficial perturbation feature augmentation. IEEE Transactions on Computational Social Systems, 11(6):8130-8142, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[72] Fengfan Zhou, Qianyu Zhou, Xiangtai Li, Xuequan Lu, Lizhuang Ma, and Hefei Ling. Adversarial attacks on both face recognition and face anti-spoofing models. arXiv preprint arXiv:2405.16940, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3526" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.161 + ], + "angle": 0, + "content": "[73] Fengfan Zhou, Qianyu Zhou, Bangjie Yin, Hui Zheng, Xuequan Lu, Lizhuang Ma, and Hefei Ling. Rethinking impersonation and dodging attacks on face recognition systems. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2487-2496, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[74] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Shouhong Ding, and Lizhuang Ma. Adaptive mixture of experts learning for generalizable face anti-spoofing. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6009-6018, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.484, + 0.289 + ], + "angle": 0, + "content": "[75] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Kekai Sheng, Shouhong Ding, and Lizhuang Ma. Generative domain adaptation for face anti-spoofing. In European Conference on Computer Vision, pages 335-356, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.359 + ], + "angle": 0, + "content": "[76] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Xuequan Lu, Ran Yi, Shouhong Ding, and Lizhuang Ma. Instance-aware domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20453-20463, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.484, + 0.431 + ], + "angle": 0, + "content": "[77] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Xuequan Lu, Shouhong Ding, and Lizhuang Ma. Test-time domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 175-187, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[78] Ziqi Zhou, Shengshan Hu, Minghui Li, Hangtao Zhang, Yechao Zhang, and Hai Jin. Advclip: Downstream-agnostic adversarial examples in multimodal contrastive learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6311–6320, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.502, + 0.484, + 0.57 + ], + "angle": 0, + "content": "[79] Hegui Zhu, Yuchen Ren, Xiaoyan Sui, Lianping Yang, and Wuming Jiang. Boosting adversarial transferability via gradient relevance attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4718-4727, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3527" + } + ] +] \ No newline at end of file diff --git a/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/aa0474ee-612f-4595-88a1-c6ba9ec15b27_origin.pdf b/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/aa0474ee-612f-4595-88a1-c6ba9ec15b27_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e017e69c609b95b36d2d310901041c3ee871a1d6 --- /dev/null +++ b/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/aa0474ee-612f-4595-88a1-c6ba9ec15b27_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fe3a3f83e7db5201564d5e90b0df8eeb4fa730a06b1080e2b95bbc8b679e47a +size 2392795 diff --git a/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/full.md b/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f8f2d0838e83b2620ba46503f4345b8d34f75b50 --- /dev/null +++ b/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/full.md @@ -0,0 +1,430 @@ +# Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation + +Fengfan Zhou $^{1}$ , Bangjie Yin $^{2}$ , Hefei Ling $^{1*}$ , Qianyu Zhou $^{3}$ , Wenxuan Wang $^{4}$ + +$^{1}$ School of Computer Science and Technology, Huazhong University of Science and Technology; $^{2}$ Shanghai Shizhuang Information Technology Co., Ltd; + +$^{3}$ Department of Computer Science and Engineering, Shanghai Jiao Tong University; $^{4}$ School of Computer Science, Northwestern Polytechnical University. + +$^{1}$ {ffzhou, lhefei}@hust.edu.cn, $^{2}$ jamesyin10@gmail.com, + +3 zhouqianyu@sjtu.edu.cn, 4wxwang@nwpu.edu.cn + +# Abstract + +Face Recognition (FR) models are vulnerable to adversarial examples that subtly manipulate benign face images, underscoring the urgent need to improve the transferability of adversarial attacks in order to expose the blind spots of these systems. Existing adversarial attack methods often overlook the potential benefits of augmenting the surrogate model with diverse initializations, which limits the transferability of the generated adversarial examples. To address this gap, we propose a novel method called Diverse Parameters Augmentation (DPA) attack method, which enhances surrogate models by incorporating diverse parameter initializations, resulting in a broader and more diverse set of surrogate models. Specifically, DPA consists of two key stages: Diverse Parameters Optimization (DPO) and Hard Model Aggregation (HMA). In the DPO stage, we initialize the parameters of the surrogate model using both pre-trained and random parameters. Subsequently, we save the models in the intermediate training process to obtain a diverse set of surrogate models. During the HMA stage, we enhance the feature maps of the diversified surrogate models by incorporating beneficial perturbations, thereby further improving the transferability. Experimental results demonstrate that our proposed attack method can effectively enhance the transferability of the crafted adversarial face examples. + +# 1. Introduction + +Owing to the relentless progress in deep learning, Face Recognition (FR) has achieved substantial advancements [1, 2, 8, 20, 34, 44]. However, the inherent vulnerabilities in existing FR models pose significant security risks [7, 31, 53, 74-77], with adversarial attacks emerging as one + +![](images/c47edcb32e1651a36b4d9a872f7d0104cd56232907865e641704ce2e3c7d03a5.jpg) +Pre-trained Parameters + +![](images/9cb077fee36cf08dad4b10a9a539c89f900eb48b4939f09f4bf6f099d310babe.jpg) +Randomly Initialized Parameters + +![](images/f55d9096223cfc97a092623c4d7879f0699f3b92d3bf6a0bf488615dd7d3faed.jpg) +(a) Input-based Augmentation + +![](images/e3cb750f5feb3254ee84e78fb0f946d9be731f090879f34e46d2c40945800c4d.jpg) +(b) Feature-based Augmentation + +![](images/e0e9fe7450a975fe6a6ca4c145b3444df2f280e33fc694b7b0f7c7b991c33268.jpg) +(c) Parameter-based Augmentation + +![](images/9cdc84cef9e805b9334097ac1a1b3127cede10b57bcec2ac9acf2a56a5e3d614.jpg) +Figure 1. Top: comparison between traditional augmentation-based adversarial attack methods and our proposed method. The black pattern filling on the left and right sides of the blue line represents input-based and parameter-based augmentation, respectively. The orange pattern filling indicates feature-based augmentation. Bottom: comparison of performance among 4 types of augmentations. + +![](images/68b5289689ff9abd4610bb641df32913a4a9fe22d441a4413b70160745f92192.jpg) + +![](images/951477ee640f009ee37f96b5390bfe8ef40e7f2135e45af4812f55966cee0f5b.jpg) + +of the critical concerns. Therefore, there is an urgent need to bolster the resilience of FR models against adversarial face examples to reveal and address the vulnerabilities. + +As a result, numerous research efforts have been focused on this area. Several adversarial attacks have been developed to create adversarial face examples with features like stealthiness [6, 15, 33, 35, 61], transferability [22, 69-71], and the capacity for physical attacks [21, 62, 63]. These initiatives are aimed at improving the effectiveness of adversarial attacks on FR. Nevertheless, the transferability of these adversarial attacks remains limited. To enhance the transferability of adversarial attacks, augmentation emerges as one of the most effective methods. As illustrated in Figure 1, augmentation-based adversarial attack consists of three aug + +mentation types: input-based [27, 46, 52, 59], feature-based [69, 71], and parameter-based [12, 57]. + +Most augmentation-based adversarial attack methods focus on input-based augmentation to improve transferability. Previous research highlights a symmetry between surrogate models used in adversarial attacks and input data in training tasks [9, 24, 71]. In training, data augmentation has proven effective in enhancing model generalization. Drawing on this symmetry, augmenting models for crafting adversarial examples can yield examples with greater transferability. Input-based augmentation can be seen as methods that only augment models at the input layer, demonstrating significant effectiveness in enhancing transferability. However, augmenting surrogate models in deeper layers (i.e., feature-based and parameter-based augmentation) offers a more direct form of augmentation. Nevertheless, few studies explore deep-layer augmentation [12, 57, 71]. As a typical parameter-based augmentation method in deep layers, LGV [12] uses a pre-trained surrogate model and collects multiple parameter sets through additional training epochs with a high, constant learning rate, thereby enhancing the transferability of adversarial examples. Although these parameter-based augmentation methods [12, 57] show promising effectiveness, they face two problems: (1) Static surrogate model initializations: these methods solely augment the surrogate model from pre-trained parameters, limiting the parameter diversity of the surrogate models and thereby hindering the transferability of the crafted adversarial examples. (2) Unavailability of the FR head: modern FR training procedures typically involve training both a backbone model and a head model [8, 44]. Following training, inference is conducted solely with the backbone models, while the head models are often not released as open-source. Consequently, in the majority of instances, we are unable to access the pre-trained parameters for the head models. Traditional parameter-based augmentation adversarial attacks augment models from pretrained parameters [12, 57]. This limitation makes it challenging to craft adversarial examples on FR models with numerous open-sourced FR models. + +To address these problems, we introduce a novel adversarial attack called Diverse Parameters Augmentation (DPA) to enhance the transferability of crafted adversarial face examples. Unlike existing parameter-based augmentation adversarial attack methods that overlook the use of diverse parameter initializations [12, 57], we diversify parameter initializations with both random and pre-trained values, thereby improving black-box attack capacity. + +Technically, our proposed attack method comprises two stages: Diverse Parameters Optimization (DPO) and Hard Model Aggregation (HMA). During the DPO stage, we initiate a subset of the optimization parameters with random noise, preserving the refined parameters in the intermediate training process to yield a diverse set of surrogate models, + +which is instrumental in bolstering the transferability of the adversarial examples. In the HMA stage, we add beneficial perturbations [56] with optimization directions opposite to those of the adversarial perturbations onto the feature maps of the parameter-augmented surrogate models, achieving the effect of hard model augmentation [71] to further enhances the transferability of the crafted adversarial examples. Our proposed attack method effectively addresses the challenges posed by the lack of diverse parameters and the absence of the FR head. The comparison between traditional adversarial attack methods and our proposed method is illustrated in Figure 1. By using diverse initializations for surrogate models, we can expand the parameter set of these models, thereby enhancing the transferability of the adversarial examples. + +Our main contributions are summarized as follows: + +- We introduce a novel perspective that parameter-based augmentation adversarial attacks should augment the parameters of surrogate models by incorporating diverse initialized parameters. To the best of our knowledge, this is the first adversarial attack on FR that utilizes parameter augmentation to enhance transferability. +- We introduce a new adversarial attack method on FR, called DPA, which comprises DPO and HMA stages. The DPO stage uses both pre-trained and random initializations to optimize the model and save the models in the intermediate training process to diversify the surrogate model set. The HMA stage adds beneficial perturbations with optimization directions opposite to those of adversarial perturbations onto the feature maps of the diversified surrogate models to further enhance transferability. +- Extensive experiments reveal that our proposed method attains superior performance when compared with the state-of-the-art adversarial attack methods. + +# 2. Related Work + +Adversarial Attacks. The primary objective of adversarial attacks is to introduce subtle perturbations into benign images, thereby deceiving machine learning systems and inducing them to produce incorrect predictions [11, 42]. The presence of adversarial examples constitutes a substantial security threat to contemporary machine learning systems. Consequently, significant research efforts have been expended to investigate adversarial attacks, with the aim of bolstering system robustness [10, 23, 27, 28, 30, 36, 66, 78]. To enhance the potency of black-box adversarial attacks, DI [59] incorporates random transformations into adversarial examples at each iteration, effectively achieving data augmentation. VMI-FGSM [50] harnesses gradient variance to stabilize the update process, thereby enhancing black-box attack performance. SSA [27] translates adversarial examples into the frequency domain and applies spectral manipulation for augmentation. SIA [52] introduces random image transformations on each image block, producing di + +![](images/4032ae9fb5acf39f50ad550aa33cbc7ed4053d3efee011085844182784053871.jpg) +Figure 2. The framework of the Diverse Parameters Optimization (DPO). We enhance the diversity of the surrogate model parameters by integrating both pre-trained and random initializations. The method yields a diverse set of surrogate model parameters, which enhances the parameter diversity of the surrogate FR models and consequently improves transferability of the crafted adversarial examples. + +verse variations for gradient estimation. BSR [46] segments the input image into multiple blocks, randomly shuffling and rotating them to generate a set of new images for gradient computation. Despite these advancements, these methods overlook the potential benefit of augmenting the surrogate model with diverse initializations, which limits the transferability of the crafted adversarial examples. In contrast, our proposed method augments surrogate models using both pre-trained and randomly initialized parameters, producing a broader and more diverse set of surrogate models. + +Adversarial Attacks on Face Recognition. Adversarial attacks on FR models can be classified based on the constraints imposed on the adversarial perturbations. These attacks are broadly categorized into two types: restricted attacks [4, 9, 25, 26, 28, 60, 72, 79] and unrestricted attacks [3, 5, 37, 40, 43, 54, 55, 64]. Restricted attacks create adversarial examples that adhere to a specified perturbation limit, such as an $L_{p}$ norm bound. Our proposed attack method is a restricted attack. Consequently, we will delve into the specifics of restricted attacks in detail. To enhance the transferability of adversarial attacks on FR, Zhong and Deng [69] introduced DFANet, which employs dropout on the feature maps of convolutional layers to achieve an ensemble-like effect. Zhou et al. [71] proposed BPFA, enhancing attack transferability by integrating beneficial perturbations [56] onto the feature maps of FR models, resulting in the effect of hard model augmentation. Li et al. [22] leveraged additional information from FR-related tasks and applied a multi-task optimization framework to further improve the transferability of adversarial examples. Unrestricted adversarial attacks + +conversely generate adversarial examples without the limitations of a predefined perturbation bound. These attacks are primarily focused on physical attacks [21, 58, 62], attribute editing [18, 33], and adversarial example generation through makeup transfer [15, 35, 39, 63]. Both restricted and unrestricted adversarial attacks on FR models have substantially advanced the capabilities of these attacks. However, existing methods for generating adversarial attacks on FR often rely on surrogate models with fixed parameters, which limits the transferability of the crafted adversarial examples. In contrast, our proposed attack method addresses this limitation by diversifying the surrogate model parameters, using models with varying parameters across different epochs. + +# 3. Methodology + +# 3.1. Problem Formulation and Framework + +Problem Formulation. Let $\mathcal{F}^{vct}(\mathbf{x}') \in \mathbb{R}^r$ denote the FR model employed by the victim to extract the embedding from a face image $\mathbf{x}'$ . We denote $\mathbf{x}^s$ and $\mathbf{x}^t$ as the source and target images, respectively. The objective of the adversarial attacks explored in our research is to manipulate $\mathcal{F}^{vct}(\mathbf{x})$ to misclassify the adversarial example $\mathbf{x}^{adv}$ as the target image $\mathbf{x}^t$ , while ensuring that $\mathbf{x}^{adv}$ bears a close visual resemblance to $\mathbf{x}^s$ . The objective of adversarial attack as delineated in this paper is as follows: + +$$ +\begin{array}{l} \mathbf {x} ^ {a d v} = \underset {\mathbf {x} ^ {a d v}} {\arg \min } \left(\mathcal {D} \left(\mathcal {F} ^ {v c t} \left(\mathbf {x} ^ {a d v}\right), \mathcal {F} ^ {v c t} \left(\mathbf {x} ^ {t}\right)\right)\right) \tag {1} \\ \mathrm {s . t .} \| \mathbf {x} ^ {a d v} - \mathbf {x} ^ {s} \| _ {p} \leqslant \epsilon \\ \end{array} +$$ + +where the symbol $\mathcal{D}$ denotes a predefined distance metric employed for the optimization of adversarial face examples. + +In practice, the attacker typically cannot access the model owned by the victim. As a result, black-box attacks play a crucial role in adversarial attacks [32, 48, 49]. A widely adopted approach for black-box attacks involves using a surrogate model $\mathcal{F}$ to generate adversarial examples and then transfer the crafted adversarial examples to the victim model [50, 71]. Consequently, the transferability is important and constitutes the key problem studied in this research. + +Framework Overview. To enhance the transferability of adversarial examples, we propose a novel method called Diverse Parameters Augmentation (DPA). DPA augments surrogate models by incorporating both pre-trained and randomly initialized parameters, resulting in a more diverse and expansive set of surrogate models. Specifically, DPA consists of two key stages: Diverse Parameters Optimization (DPO) and Hard Model Aggregation (HMA). In the DPO stage, we diversify the surrogate model parameters by combining pre-trained and random initializations, as illustrated in Figure 2. In the HMA stage, we apply beneficial perturbations to the feature maps of the diversified surrogate models and combine them to achieve a higher degree of augmentation, as shown in Figure 3. In the following, we will introduce our proposed DPA attack method in detail. + +# 3.2. Diverse Parameters Optimization + +The parameter diversity of surrogate models is crucial for the transferability of crafted adversarial examples. Previous parameter-based augmentation adversarial attack methods solely initialize parameters with pre-trained values, thereby limiting the parameter diversity of surrogate models [12, 57]. In contrast, the key innovation of the stage of our proposed attack method is the diversification of initialized parameters using both pre-trained and randomly initialized parameters. Following initialization, we augment the parameters of the surrogate models using the parameters in the intermediate training process to diversify the surrogate models, thereby improving the transferability of the adversarial examples. + +Let $\mathbf{x}$ be a batch of face images, $\mathbf{w} \in \mathbb{R}^{s \times r}$ be the parameters of the FR head, $b$ be the batch size of $\mathbf{x}$ , and $s$ be the class number in the training dataset. To calculate the loss function, we should first calculate the cosine similarity matrix $\mathbf{r} \in \mathbb{R}^{b \times s}$ between $\mathcal{F}(\mathbf{x}) \in \mathbb{R}^{b \times r}$ and the parameters of the FR head $\mathbf{w}^{\top}$ using the following formula: + +$$ +\mathbf {r} = \cos \mathbf {a} = \bar {\mathbf {e}} \bar {\mathbf {w}} ^ {\top} \tag {2} +$$ + +where $\overline{\mathbf{e}} = \frac{\mathcal{F}(\mathbf{x})}{\|\mathcal{F}(\mathbf{x})\|}, \overline{\mathbf{w}} = \frac{\mathbf{w}}{\|\mathbf{w}\|}$ and $\mathbf{a}$ is the angle between $\mathcal{F}(\mathbf{x})$ and $\mathbf{w}^{\top}$ . After obtaining the cosine similarity matrix by Equation (2), we can get the sine similarity between $\mathcal{F}(\mathbf{x})$ and $\mathbf{w}$ using the following equation: + +$$ +\sin \mathbf {a} = \sqrt {1 . 0 - \cos^ {2} \mathbf {a}} \tag {3} +$$ + +Using the cosine and sine similarity matrix, the formula to get the additive angular margin cosine similarity matrix can be expressed as: + +$$ +\cos (\mathbf {a} + m) = \cos \mathbf {a} \cos m - \sin \mathbf {a} \sin m \tag {4} +$$ + +where $m$ is the margin value for increasing the discrimination of the FR model [8]. Let $\mathbf{y} \in \mathbb{R}^b$ be the labels of the current batch. By utilizing $\mathbf{y}$ , we can calculate its one-hot encoded matrix using the following formula: + +$$ +\widehat {\mathbf {h}} = \Psi (\mathbf {y}) \in \mathbb {R} ^ {b \times s} \tag {5} +$$ + +where $\Psi (\cdot)$ is the one-hot encode operation. Next, we use the following function to calculate the output of the head: + +$$ +\begin{array}{l} \mathbf {q} = d \left(\hat {\mathbf {h}} \odot \mathbf {p} + (1 - \hat {\mathbf {h}}) \odot \mathbf {r}\right) \\ \mathbf {p} = \left\{ \begin{array}{l l} \cos (\mathbf {a} + m) & \text {s . t .} \quad \mathbf {a} < \pi - m, \\ \cos \mathbf {a} - m \sin m & \text {s . t .} \quad \mathbf {a} \geqslant \pi - m, \end{array} \right. \tag {6} \\ \end{array} +$$ + +where $d$ is a pre-defined scale factor. Using the output of the head, the formula to calculate the loss function can be expressed as the following: + +$$ +\mathcal {L} (\mathbf {x}, \mathbf {y}) = - \frac {1}{b} \sum_ {i = 1} ^ {b} \log \frac {e ^ {\mathbf {q} _ {i , \mathbf {y} _ {i}}}}{\sum_ {j = 1} ^ {s} e ^ {\mathbf {q} _ {i , j}}} \tag {7} +$$ + +Let $c$ and $\mathbf{v}$ be the number of the training epoch and the parameters of the FR backbone, respectively. Utilizing Equation (7), we can get the parameters set [8]: + +$$ +\mathbf {v} _ {i}, \mathbf {w} _ {i} = \varphi \left(\underset {\mathbf {v}, \mathbf {w}} {\arg \min } \mathcal {L} \left(\mathbf {x}, \mathbf {y}\right), i\right) \quad i \in \{1, 2, \ldots , c \} +$$ + +$$ +\mathbf {V} _ {c} = \left\{\mathbf {v} _ {1}, \mathbf {v} _ {2}, \dots , \mathbf {v} _ {c - 1}, \mathbf {v} _ {c} \right\} \quad \mathbf {W} _ {c} = \left\{\mathbf {w} _ {1}, \mathbf {w} _ {2}, \dots , \mathbf {w} _ {c - 1}, \mathbf {w} _ {c} \right\} \tag {8} +$$ + +where $\varphi (\cdot ,i)$ is the operation that obtain the $\mathbf{v}$ and $\mathbf{w}$ parameters in the $i$ th epoch. We denote $\mathbf{v}_0$ and $\mathbf{w}_0$ as the initialized parameters of the FR model and head, respectively. If $\mathbf{v}_0$ and $\mathbf{w}_0$ are fixed, $\mathbf{V}_c$ and $\mathbf{W}_c$ become deterministic aside from minor random factors in the training process of Equation (8). Therefore, there exists a mapping between $\{\mathbf{V}_c,\mathbf{W}_c\}$ and $\{\mathbf{v}_0,\mathbf{w}_0\}$ that can be expressed as: + +$$ +\left\{\mathbf {V} _ {c}, \mathbf {W} _ {c} \right\} = F \left(\left\{\mathbf {v} _ {0}, \mathbf {w} _ {0} \right\}, i\right) \tag {9} +$$ + +Existing adversarial attack methods based on parameter augmentation typically initialize the parameters $\{\mathbf{v}_0,\mathbf{w}_0\}$ solely with pre-trained values [12, 57]. This constraint diminishes the diversity within the augmented surrogate model set. In contrast, we opt to initialize $\{\mathbf{v}_0,\mathbf{w}_0\}$ with diverse parameters to enhance the diversity of the augmented surrogate model set, thereby improving transferability. The $\{\mathbf{V}_c,\mathbf{W}_c\}$ of our proposed attack can be expressed as: + +$$ +\begin{array}{l} \left\{\mathbf {V} _ {c} ^ {p}, \mathbf {W} _ {c} ^ {p} \right\} = F \left(\left\{\mathbf {v} _ {0} ^ {p}, \mathbf {w} _ {0} ^ {p} \right\}, i\right) \tag {10} \\ \left\{\mathbf {V} _ {c} ^ {a}, \mathbf {W} _ {c} ^ {a} \right\} = F \left(\left\{\mathbf {v} _ {0} ^ {a}, \mathbf {w} _ {0} ^ {a} \right\}, i\right) \\ \end{array} +$$ + +![](images/db00f76aa6e1fab04667b020fc5c4f89d78f35a29b43e361205fb5377649a297.jpg) +Figure 3. The framework of the Hard Model Aggregation (HMA). After acquiring a surrogate model set with diverse parameters (i.e., $\mathbf{V}_c^q$ ), we introduce beneficial perturbations with the optimization direction opposite to that of adversarial perturbations onto the feature maps of these diversified surrogate models, transforming them into hard models and aggregate the hard models to increase the transferability. + +where $\mathbf{v}_0^p$ represents the pre-trained parameters, while $\mathbf{w}_0^p$ , $\mathbf{v}_0^a$ and $\mathbf{w}_0^a$ are randomly initialized. Using Equation (10), we can obtain the more diverse parameters for the FR model and head. After obtaining $\mathbf{V}_c^p$ and $\mathbf{V}_c^a$ , we use the following set of the parameters to craft the adversarial examples: + +$$ +\begin{array}{l} \mathbf {V} _ {c} ^ {q} = \left\{\mathbf {v} _ {0} ^ {p}, \mathbf {v} _ {1} ^ {p}, \mathbf {v} _ {j _ {1}} ^ {p}, \dots , \mathbf {v} _ {c} ^ {p} \right\} \bigcup \left\{\mathbf {v} _ {1} ^ {a}, \mathbf {v} _ {j _ {1}} ^ {a}, \mathbf {v} _ {j _ {2}} ^ {a}, \dots , \mathbf {v} _ {c} ^ {a} \right\} \tag {11} \\ \mathrm {s . t .} j \in \{1, 2, \dots , c | j \bmod \kappa = 1 \} \\ \end{array} +$$ + +where $\kappa = \lfloor \sqrt{c} \rfloor$ determines the epoch interval to select the parameters. The overall framework and pseudo-code of the DPO stage in our proposed method is illustrated in Figure 2 and Algorithm 1, respectively. + +Algorithm 1 Diverse Parameters Optimization (DPO) +```latex +Input: Diverse initial parameters $\mathbf{v}_0^p$ , $\mathbf{w}_0^p$ , $\mathbf{v}_0^a$ , and $\mathbf{w}_0^a$ , the number of epochs $c$ , the optimizer $\mathbf{T}$ , the FR training dataset $\zeta$ , loss function $\mathcal{L}$ . +``` + +Output: Augmented parameter set $\mathbf{V}_c^q$ . + +1: $\mathbf{P} = \{\mathbf{v}_0^p, \mathbf{w}_0^p\}$ , $\mathbf{A} = \{\mathbf{v}_0^a, \mathbf{w}_0^a\}$ , $\mathbf{V}_c^q = \{\mathbf{v}_0^p\}$ +2: for $\mathbf{J} \in \{\mathbf{P}, \mathbf{A}\}$ do +3: $\{\mathbf{v},\mathbf{w}\} = \mathbf{J}$ +4: $\mathcal{F} = \Psi (\mathbf{v})\quad \triangleright$ Map the parameters to the models. +5: for $i = 1,\dots,c$ do +6: for $\mathbf{x}\in \zeta$ do +7: Calculate $\mathcal{L}$ using Equation (7). +8: Backward $(\mathbf{T},\mathcal{L})\triangleright$ Backpropagation using $\mathbf{T}$ +9: Update the parameters $\mathbf{v},\mathbf{w}$ +0: end for +11: if the $\mathbf{v}$ satisfies the condition outlined in Equation (11) then +12: Add $(\mathbf{V}_{c}^{q},\mathbf{v})$ Incorporate $\mathbf{v}$ into $\mathbf{V}_c^q$ +13: end if +14: end for +15: end for + +# 3.3. Hard Model Aggregation + +Hard models [71] refer to models with higher loss values compared to normal models, designed to simulate the effect + +of hard samples, which are hard to the optimization process of the adversarial examples. Since hard samples can improve the generalization of models, we leverage multiple hard models to enhance the transferability of the surrogate models with diverse parameters by adding beneficial perturbations [56] onto the pre-defined feature maps of the surrogate models. + +In the following, we will provide a detailed introduction to the HMA stage. Let $\Psi$ be the mapping from the parameters to the corresponding models that can be expressed as: + +$$ +\mathbf {F} = \Psi \left(\mathbf {V} _ {c} ^ {q}\right) = \left\{\mathcal {F} _ {1}, \mathcal {F} _ {2}, \dots , \mathcal {F} _ {g} \right\} \tag {12} +$$ + +where $g = |\mathbf{V}_c^q|$ . Hard sample augmentation has demonstrated significant effectiveness in enhancing model generalization [41, 67]. Based on the relationship between adversarial attack tasks and training tasks [9, 24], using hard models for augmentation can lead to improved transferability [71]. Consequently, by transforming vanilla parameter-augmented surrogate models into hard parameter-augmented surrogate models, we can achieve more transferable results. To accomplish this transformation, we use the following formula to calculate the loss function: + +$$ +\begin{array}{l} \widetilde {\mathcal {L}} _ {t} = \frac {1}{g} \sum_ {i = 0} ^ {g} \left\| \phi \left(\mathcal {H} _ {i} \left(\mathcal {T} \left(\mathbf {x} _ {t} ^ {a d v}\right)\right)\right) - \phi \left(\mathcal {F} _ {i} \left(\mathbf {x} ^ {t}\right)\right) \right\| \tag {13} \\ \begin{array}{l l} \text {s . t .} & \mathcal {F} _ {i} \in \mathbf {F} \quad t \in \{1, 2, \dots , n \} \end{array} \\ \end{array} +$$ + +where $\mathbf{x}_t^{adv}$ denotes the adversarial example generated during the $t$ th iteration, the variable $n$ represents the maximum number of iterations allocated for crafting the adversarial examples, $\phi(\cdot)$ denotes the normalization operation, $\mathcal{T}$ signifies the input transformation, and $\mathcal{H}$ is the corresponding hard model [71] of $\mathcal{F}$ , which can be expressed as: + +$$ +\mathcal {H} \left(\mathbf {x} _ {t} ^ {a d v}\right) = \left\{ \begin{array}{l l} \mathcal {F} \left(\mathbf {x} _ {t} ^ {a d v}\right) & \text {s . t .} \quad t = 1, \\ \chi \left(\mathcal {F} \left(\mathbf {x} _ {t} ^ {a d v}\right), \Omega\right) & \text {s . t .} \quad t > 1, \end{array} \right. \tag {14} +$$ + +where $\chi$ is the mapping for adding beneficial perturbations on the feature maps in the forward propagation whose formula can be expressed as: + +$$ +\begin{array}{l} \chi \left(\mathcal {F} \left(\mathbf {x} _ {t} ^ {a d v}\right), \Omega\right): \omega = \omega + \eta \text {s i g n} \left(\nabla_ {\omega} \tilde {\mathcal {L}} _ {t - 1}\right) \tag {15} \\ \begin{array}{c c} \text {s . t .} & \omega \in \Omega \end{array} \\ \end{array} +$$ + +where $\Omega$ is the pre-defined set of layers for adding beneficial perturbations. Let $\epsilon$ be the maximum allowable perturbation. Utilizing $\tilde{\mathcal{L}}_t$ , we can craft the adversarial face examples using following formula: + +$$ +\mathbf {x} _ {t + 1} ^ {a d v} = \prod_ {\mathbf {x} ^ {s}, \epsilon} \left(\mathbf {x} _ {t} ^ {a d v} - \beta \operatorname {s i g n} \left(\nabla_ {\mathbf {x} _ {t} ^ {a d v}} \tilde {\mathcal {L}} _ {t}\right)\right) \tag {16} +$$ + +where $\beta$ denotes the step size used to construct the adversarial examples, while $\prod$ represents the clipping operation that confines the pixel values of the generated adversarial examples within the range $[\mathbf{x}^s -\epsilon ,\mathbf{x}^s +\epsilon ]$ . The framework of the HMA is illustrated in Figure 3. + +In the following, we present the pseudo-code for the HMA. For the sake of clarity, we focus on neural networks with a single branch in their computational graphs. In the case of neural networks with multiple branches, the HMA algorithm remains largely unchanged, except for the incorporation of mechanisms to handle multiple branches. Before proceeding, we introduce some necessary definitions. Let $\Phi$ denote the index set of pre-selected layers to which beneficial perturbations will be added: + +$$ +\Phi = \varkappa (\Omega) \tag {17} +$$ + +where $\varkappa$ represents the mapping from the pre-defined set of layers for adding beneficial perturbations $\Omega$ , to their corresponding layer index set $\Phi$ . Let $f^i$ denote the $i$ th layer within the model $\mathcal{F}$ , and let $z$ represent the total number of layers in $\mathcal{F}$ . We define the segment of $\mathcal{F}$ from layer $f^i$ to layer $f^j$ as follows: + +$$ +\mathcal {F} ^ {i, j} = f ^ {i} \circ f ^ {i + 1} \circ \dots \circ f ^ {j - 1} \circ f ^ {j} \tag {18} +$$ + +The pseudo-code for the HMA stage of our proposed method is presented in Algorithm 2. + +# 4. Experiments + +# 4.1. Experimental Setting + +Datasets. We opt to use the LFW [16], and CelebA-HQ [19] for our experiments to verify the effectiveness of our proposed attack method. LFW serves as an unconstrained face dataset for FR. CelebA-HQ consists of face images with high visual quality. The LFW and CelebA-HQ utilized in our experiments are identical to those employed in [70, 71, 73]. For the parameter augmentation process, we select BUPT-Balancedface [47] as our training dataset. + +Face Recognition Models. The normal trained FR models employed in our experiments encompass IR152 [13], IRSE50 [14], FaceNet [34], MobileFace [8], CurricularFace [17], MagFace [29], ArcFace [8], CircleLoss [38], MV-Softmax [51], and NPCFace [65]. Specifically, IR152, FaceNet, IRSE50, and MobileFace are the same models utilized in [15, 63, 70, 71]. CurricularFace, MagFace, ArcFace, + +Algorithm 2 Hard Model Aggregation (HMA) + +Input: The source image $\mathbf{x}^s$ , the target image $\mathbf{x}^t$ , the mapping from the parameters to the corresponding models $\Psi$ , the maximum number of iterations $n$ , the index set of pre-selected layers to be added beneficial perturbations $\Phi$ , the step size of the beneficial perturbations $\eta$ , the step size of the adversarial perturbations $\beta$ , the maximum permissible magnitude of perturbation $\epsilon$ , the parameter set $\mathbf{V}_c^q$ , the total number of layers in a single surrogate model $z$ , normalization operation $\phi$ . + +Output: An adversarial example $\mathbf{x}_n^{adv}$ + +1: $\mathbf{x}_0^{adv} = \mathbf{x}^s$ , $u = |\Phi|, s_1 = 1, g = |\mathbf{V}_c^q|$ +2: $\mathbf{F} = \Psi (\mathbf{V}_c^q)$ $\triangleright$ Acquire the diversified models set. +3: for $t = 1, \dots, n$ do +4: for $i = 1,\dots,g$ do +5: $\mathcal{F} = \mathbf{F}_i$ Derive the parameter-augmented model. +6: $\omega^{t,0} = \mathcal{T}\left(\mathbf{x}_{t - 1}^{adv}\right)$ +7: for $j = 1,\dots,u$ do +8: $s_2 = \Phi_j$ +9: $\omega^{t,j} = \mathcal{F}^{s_1,s_2}\left(\omega^{t,j - 1}\right)$ +10: $s_1 = s_2$ +11: if $t \neq 1$ then +12: $\omega^{t,j} = \omega^{t,j} + \eta \mathrm{sign}\left(\nabla_{\omega^{t - 1,j}}\widetilde{\mathcal{L}}_{t - 1}\right)$ +13: end if +14: end for +15: $\tilde{\mathcal{L}}^i = \| \phi (\mathcal{F}^{\Phi_u,z}(\omega^{t,u})) - \phi (\mathcal{F}(\mathbf{x}^t))\| _2^2$ +16: end for +17: $\widetilde{\mathcal{L}}_t = \frac{1}{g}\sum_{i = 1}^g\widetilde{\mathcal{L}}^i$ Compute the loss function utilizing the parameter-augmented models. +18: $\mathbf{x}_t^{adv} = \prod_{\mathbf{x}^s,\epsilon}\left(\mathbf{x}_{t - 1}^{adv} - \beta \mathrm{sign}\left(\nabla_{\mathbf{x}_{t - 1}^{adv}}\widetilde{\mathcal{L}}_t\right)\right)$ +19: end for + +CircleLoss, MV-Softmax, and NPCFace are the official models provided by FaceX-ZOO [45]. Furthermore, we integrate adversarial robust FR models into our experiments, denoted as IR152 $^{adv}$ , IRSE50 $^{adv}$ , FaceNet $^{adv}$ , and MobileFace $^{adv}$ which correspond to the models used in [71]. + +Attack Setting. We set the maximum allowable perturbation magnitude $\epsilon$ to 10, based on the $L_{\infty}$ norm bound, without any specific emphasis. In addition, we specify the maximum number of iterative steps as 200. More detailed attack settings are provided in the supplementary. + +Evaluation Metrics. We utilize Attack Success Rate (ASR) as the metric to assess the efficacy of various adversarial attacks. The ASR represents the ratio of adversarial examples that successfully evade the victim model to the total number of adversarial examples generated. In calculating the ASR, we determine the threshold based on a FAR@0.001 on the entire LFW dataset for each victim model. + +Compared methods. Our proposed attack is a form of re + +stricted adversarial attack designed to expose vulnerabilities within FR systems. It would be inequitable to juxtapose our method with unrestricted attacks, which do not impose limits on the magnitude of adversarial perturbations. Consequently, we opt to benchmark our approach against other restricted attacks on FR systems that are explicitly malicious at attacking the systems [69] [71], as well as state-of-the-art transfer attacks [59] [27] [12] [52] [46]. + +# 4.2. Comparison Studies + +DPA achieves the best black-box attack results on both normally trained and adversarial robust models. To verify the effectiveness of our proposed attack method, we craft the adversarial examples on the LFW and CelebA-HQ dataset. The black-box performance with MobileFace as the surrogate models on the LFW dataset are demonstrated in Table 1. Some of the adversarial examples are demonstrated in Figure 4. More experimental results on the LFW dataset, as well as those on the CelebA-HQ dataset, are provided in the supplementary. These results demonstrate that our proposing method consistently outperforms the baseline attack, thereby highlighting its effectiveness in improving the transferability of adversarial face examples. + +![](images/de1ea6c5002e68905201547293464742956703050fb5bf07b6a141d0a27e1736.jpg) +Figure 4. The illustration of adversarial examples crafted by various attacks. First column: some source images. Last column: the corresponding target images. The second to fifth columns exhibit the corresponding adversarial face examples crafted by SIA [52], BPFA [71], BSR [46], and Our proposed attack, respectively. + +DPA achieves superior black-box performance under JPEG compression. JPEG compression is a widely adopted method for image compression during transmission, and it also serves as a defense mechanism against adversarial examples. To evaluate the effectiveness of our proposed attack under JPEG compression, we employ MobileFace as the surrogate model and IRSE50 as the victim model. We assess the attack performance on the LFW and CelebA-HQ datasets, with the results presented in Figure 5. These results indicate + +Table 1. Comparisons of black-box ASR $(\%)$ results for attacks using MobileFace as the surrogate model on the LFW dataset. I, S, F, M denote IR152, IRSE50, FaceNet, and MobileFace, respectively. + +
AttacksISFIadvSadvFadvMadv
FIM [68]5.373.47.52.54.52.810.9
DI [59]18.497.332.910.618.210.238.6
DFANet [69]7.086.411.93.66.43.916.8
VMI [50]13.696.020.27.612.87.632.3
SSA [27]13.896.419.65.513.17.231.5
SIA [52]15.796.826.78.114.59.135.3
BPFA [71]15.895.517.85.911.05.128.4
BSR [46]5.474.29.52.95.94.911.4
Ours67.798.290.855.466.442.671.6
+ +![](images/81d0a75afe082434297617af1f5a3e3084a098981cf16c2e169eb8b3acea8c0b.jpg) +Figure 5. Performance of ASR across various JPEG Q values: (a) Results on the LFW dataset. (b) Results on the CelebA-HQ dataset. + +![](images/9f3d27c33027bb07124314b3e6202b6e9ab2a38f2f7cd273e4e5c53a701b1ae2.jpg) + +Table 2. Comparison of black-box ASR (%) results using the parameter-based augmented adversarial attack method as the baseline on the LFW dataset. $\mathrm{ASR}^{adv}$ represents the average attack success rate on adversarial robust models. + +
IR152IRSE50FaceNetASRadv
Baseline4.819.615.77.2
Ours67.798.290.859.0
+ +that our proposed attack method consistently outperforms the baseline attack methods across varying levels of JPEG compression, thereby underscoring the robustness of our proposed attack method under such conditions. + +DPA demonstrates better black-box performance compared to the parameter-based augmented adversarial attack. The LGV method is an effective parameter-based augmented adversarial attack technique designed to enhance transferability. Our proposed attack incorporates parameter augmentation, making LGV an appropriate baseline. We selected MobileFace as the surrogate model and generated adversarial examples on the LFW dataset. The results are presented in Table 2. Table 2 clearly shows that our proposed attack outperforms the baseline, further validating the effectiveness of our method. The primary limitation of LGV stems from its random selection of one model per iteration, which introduces instability and adversely impacts blackbox ASR [9, 50]. Furthermore, LGV suffers from a lack of diverse initialization parameters. These shortcomings significantly impair the overall ASR performance of LGV. + +![](images/4b4d847a842e07f84ed23c476d9cf9f43c589da7dc0cb22b0cb16c91bd5fc462.jpg) +Figure 6. ASR on victim models trained with various algorithms, using FaceNet as the surrogate model on the LFW dataset. + +DPA achieves superior black-box performance on FR models trained with various algorithms. In order to substantiate the effectiveness of our proposed attack method across various FR models, we conducted additional experiments using BSR [46] as the Baseline and FaceNet as the surrogate model. These experiments adhered to the same settings detailed in Table 7 in the supplementary. The ASR across multiple FR models is illustrated in Figure 6. As shown in Figure 6, the ASR of the adversarial examples generated by our proposed method exceeds that of the Baseline, thereby reinforcing the efficacy of our proposed method. + +# 4.3. Ablation Studies + +The ablation studies on each stage of our proposed attack method. Our proposed attack method consists of two stages. To verify the effectiveness of each stage, we conduct ablation studies on the stages. Initially, we craft adversarial examples using only the surrogate model, which we denote as 'Vanilla'. Next, we generate adversarial examples using the ensemble of surrogate models obtained in the DPO stage (i.e. F), denoted as DPO. After incorporating the HMA stage, the complete attack method is denoted as 'DPO + HMA'. The results are presented in Table 3. Table 3 demonstrates that the addition of the DPO stage results in an increase in black-box ASR, showcasing the effectiveness of the DPO stage in enhancing transferability. Further incorporation of the HMA stage leads to an additional improvement in attack performance, underscoring the effectiveness of the HMA stage in boosting black-box performance. These results collectively demonstrate the effectiveness of each stage in our proposed attack method in improving the transferability. + +The ablation studies on the effectiveness of multiple surrogate models at different epochs in improving transferability. Since our proposed attack method utilizes models from intermediate training epochs to craft adversarial examples, it is essential to verify the effectiveness of this approach compared to using only models from the final training epoch. We conduct ablation experiments on these two approaches using the LFW dataset and MobileFace as the surrogate + +Table 3. Ablation study of black-box ASR (\%) results on the LFW dataset using MobileFace as the surrogate model. $\mathrm{ASR}^{adv}$ denotes the average attack success rate on adversarial robust models. + +
AttacksIR152IRSE50FaceNetASRadv
Vanilla5.373.47.55.2
DPO35.991.367.932.9
DPO + HMA67.798.290.859.0
+ +Table 4. Ablation study of black-box ASR (\%) results to verify the effectiveness of multiple surrogate models at different epochs in improving transferability. $\mathrm{ASR}^{adv}$ denotes the average attack success rate on adversarial robust models. + +
AttacksIR152IRSE50FaceNetASRadv
Fm20.386.733.617.8
F (Ours)35.991.367.932.9
+ +model. We employ the DPO process to obtain the parameter sets $\mathbf{V}_c^q$ and $\mathbf{V}_c^m = \{\mathbf{v}_0^p, \mathbf{v}_c^p, \mathbf{v}_c^a\}$ . Next, we map $\mathbf{V}_c^m$ into its corresponding model set $\mathbf{F}^m = \Psi(\mathbf{V}_c^m)$ . We then craft adversarial examples by ensembling the models from $\mathbf{V}_c^q$ (i.e. $\mathbf{F}$ ) and $\mathbf{F}^m$ , respectively. The results are presented in Table 4. Table 4 shows that the performance of the ensemble of $\mathbf{F}$ surpasses that of the ensemble of $\mathbf{F}^m$ , demonstrating the effectiveness of using models from intermediate training epochs to enhance the transferability. + +The hyper-parameter analysis studies of our proposed method are in the supplementary. + +# 5. Conclusion + +We present an innovative advancement in the field of adversarial attacks on FR through the introduction of the Diverse Parameters Augmentation (DPA) attack method. By addressing the problems of traditional adversarial attacks, particularly the lack of diverse parameters and the exclusion of the FR head, DPA enhances the transferability of adversarial face examples. In addition, DPA emphasizes the importance of integrating both pre-trained and randomly initialized parameters, providing a fresh perspective on augmentation-based adversarial attacks. The extensive experimental results demonstrate the effectiveness of our proposed DPA attack across multiple datasets and target models. We anticipate that our proposed method will illuminate potential directions for advancing research in adversarial attacks on FR. + +# 6. Acknowledgment + +This work was supported in part by the Natural Science Foundation of China under Grant 62372203 and 62302186, in part by the Major Scientific and Technological Project of Shenzhen (202316021), in part by the National key research and development program of China (2022YFB2601802), in part by the Major Scientific and Technological Project of Hubei Province (2022BAA046, 2022BAA042). + +# References + +[1] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, and Ying Fu. Partial FC: training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 1445-1449, 2021. 1 +[2] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19593-19604, 2023. 1 +[3] Tom B Brown, Nicholas Carlini, Chiyuan Zhang, Catherine Olsson, Paul Christiano, and Ian Goodfellow. Unrestricted adversarial examples. arXiv preprint arXiv:1809.08352, 2018. 3 +[4] Bin Chen, Jia-Li Yin, Shukai Chen, Bohao Chen, and Ximeng Liu. An adaptive model ensemble adversarial attack for boosting adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4466-4475, 2023. 3 +[5] Zhaoyu Chen, Bo Li, Shuang Wu, Kaixun Jiang, Shouhong Ding, and Wenqiang Zhang. Content-based unrestricted adversarial attack. In Advances in Neural Information Processing Systems, 2023. 3 +[6] Valeriia Cherepanova, Micah Goldblum, Harrison Foley, Shiyuan Duan, John P Dickerson, Gavin Taylor, and Tom Goldstein. Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In International Conference on Learning Representations, 2021. 1 +[7] Sowmen Das, Selim Seferbekov, Arup Datta, Md Saiful Islam, and Md Ruhul Amin. Towards solving the deepfake problem: An analysis on improving deepfake detection using dynamic face augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3776-3785, 2021. 1 +[8] Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5962-5979, 2022. 1, 2, 4, 6 +[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 2, 3, 5, 7 +[10] Zhijin Ge, Fanhua Shang, Hongying Liu, Yuanyuan Liu, Liang Wan, Wei Feng, and Xiaosen Wang. Improving the transferability of adversarial examples with arbitrary style transfer. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4440-4449, 2023. 2 +[11] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. 2 + +[12] Martin Gubri, Maxime Cordy, Mike Papadakis, Yves Le Traon, and Koushik Sen. Lgv: Boosting adversarial example transferability from large geometric vicinity. In European Conference on Computer Vision, pages 603-618. Springer, 2022. 2, 4, 7 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6 +[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 6 +[15] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14994–15003, 2022. 1, 3, 6 +[16] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, 2007. 6 +[17] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: Adaptive curriculum learning loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 6 +[18] Shuai Jia, Bangjie Yin, Taiping Yao, Shouhong Ding, Chunhua Shen, Xiaokang Yang, and Chao Ma. Adv-attribute: Inconspicuous and transferable adversarial attack on face recognition. In Advances in Neural Information Processing Systems, 2022. 3 +[19] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representation, 2018. 6 +[20] Jingzhi Li, Zidong Guo, Hui Li, Seungju Han, Ji-Won Baek, Min Yang, Ran Yang, and Sungjoo Suh. Rethinking feature-based knowledge distillation for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20156-20165, 2023. 1 +[21] Yanjie Li, Yiquan Li, Xuelong Dai, Songtao Guo, and Bin Xiao. Physical-world optical adversarial attacks on 3d face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24699-24708, 2023. 1, 3 +[22] Zexin Li, Bangjie Yin, Taiping Yao, Junfeng Guo, Shouhong Ding, Simin Chen, and Cong Liu. Sibling-attack: Rethinking transferable adversarial attacks against face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24626–24637, 2023. 1, 3 +[23] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation + +from diffusion models via adversarial examples. In International Conference on Machine Learning, pages 20763-20786, 2023. 2 +[24] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representation, 2020. 2, 5 +[25] Xuannan Liu, Yaoyao Zhong, Yuhang Zhang, Lixiong Qin, and Weihong Deng. Enhancing generalization of universal adversarial perturbation through gradient aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4412-4421, 2023. 3 +[26] Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, and Di Ming. TRM-UAP: enhancing the transferability of data-free universal adversarial perturbation via truncated ratio maximization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4739-4748, 2023. 3 +[27] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xi-anglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In European Conference on Computer Vision, pages 549–566, 2022. 2, 7, 1 +[28] Dong Lu, Zhiqiang Wang, Teng Wang, Weili Guan, Hongchang Gao, and Feng Zheng. Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 102-111, 2023. 2, 3 +[29] Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14225-14234, 2021. 6 +[30] Duan Mingxing, Kenli Li, Lingxi Xie, Qi Tian, and Bin Xiao. Towards multiple black-boxes attack via adversarial example generation network. In Proceedings of the 29th ACM International Conference on Multimedia, pages 264-272, 2021. 2 +[31] Kartik Narayan, Harsh Agarwal, Kartik Thakral, Surbhi Mittal, Mayank Vatsa, and Richa Singh. Df-platter: Multiface heterogeneous deepfake dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9739-9748, 2023. 1 +[32] Xuelin Qian, Wenxuan Wang, Yu-Gang Jiang, Xiangyang Xue, and Yanwei Fu. Dynamic routing and knowledge relearning for data-free black-box attack. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4 +[33] Haonan Qiu, Chaowei Xiao, Lei Yang, Xinchen Yan, Honglak Lee, and Bo Li. Semanticadv: Generating adversarial examples via attribute-conditioned image editing. In European Conference on Computer Vision, pages 19-37, 2020. 1, 3 +[34] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 815-823, 2015. 1, 6 +[35] Fahad Shamshad, Muzammal Naseer, and Karthik Nandakumar. Clip2protect: Protecting facial privacy using text-guided + +makeup via adversarial latent search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20595-20605, 2023. 1, 3 +[36] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In International Conference on Learning Representations, 2024. 2 +[37] Florian Stimberg, Ayan Chakrabarti, Chun-Ta Lu, Hussein Hazimeh, Otilia Stretcu, Wei Qiao, Yintao Liu, Merve Kaya, Cyrus Rashtchian, Ariel Fuxman, Mehmet Tek, and Sven Gowal. Benchmarking robustness to adversarial image obfuscations. In Advances in Neural Information Processing Systems, 2023. 3 +[38] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 6 +[39] Yuhao Sun, Lingyun Yu, Hongtao Xie, Jiaming Li, and Yongdong Zhang. Diffam: Diffusion-based adversarial makeup transfer for facial privacy protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24584-24594, 2024. 3 +[40] Naufal Suryanto, Yongsu Kim, Harashta Tatimma Larasati, Hyoeun Kang, Thi-Thu-Huong Le, Yoonyoung Hong, Hunmin Yang, Se-Yoon Oh, and Howon Kim. ACTIVE: towards highly transferable 3d physical camouflage for universal and robust vehicle evasion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4282-4291, 2023. 3 +[41] Teppei Suzuki. Teachaugment: Data augmentation optimization using teacher knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10894-10904, 2022. 5 +[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 2 +[43] Donghua Wang, Wen Yao, Tingsong Jiang, Chao Li, and Xiaogian Chen. RFLA: A stealthy reflected light adversarial attack in the physical world. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4442, 2023. 3 +[44] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 2 +[45] Jun Wang, Yinglu Liu, Yibo Hu, Hailin Shi, and Tao Mei. Facex-zoo: A python toolbox for face recognition. 2021. 6 +[46] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting Adversarial Transferability by Block Shuffle and Rotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 7, 8, 1 +[47] Mei Wang, Weihong Deng, Jiani Hu, Xunqiang Tao, and Yaohai Huang. Racial faces in the wild: Reducing racial + +bias by information maximization adaptation network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 6 +[48] Wenxuan Wang, Bangjie Yin, Taiping Yao, Li Zhang, Yanwei Fu, Shouhong Ding, Jilin Li, Feiyue Huang, and Xiangyang Xue. Delving into data: Effectively substitute training for black-box attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4761-4770, 2021. 4 +[49] Wenxuan Wang, Xuelin Qian, Yanwei Fu, and Xiangyang Xue. Dst: Dynamic substitute training for data-free black-box attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14361-14370, 2022. 4 +[50] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1924-1933, 2021. 2, 4, 7, 1 +[51] Xiaobo Wang, Shifeng Zhang, Shuo Wang, Tianyu Fu, Hailin Shi, and Tao Mei. Mis-classified vector guided softmax loss for face recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12241-12248, 2020. 6 +[52] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4607-4619, 2023. 2, 7, 1 +[53] Xudong Wang, Ke-Yue Zhang, Taiping Yao, Qianyu Zhou, Shouhong Ding, Pingyang Dai, and Rongrong Ji. Tf-fas: Twofold-element fine-grained semantic guidance for generalizable face anti-spoofing. In European Conference on Computer Vision, 2024. 1 +[54] Xingxing Wei, Yao Huang, Yitong Sun, and Jie Yu. Unified adversarial patch for cross-modal attacks in the physical world. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4422-4431, 2023. 3 +[55] Xingxing Wei, Jie Yu, and Yao Huang. Physically adversarial infrared patches with learnable shapes and locations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12334-12342, 2023. 3 +[56] Shixian Wen and Laurent Itti. Beneficial perturbations network for defending adversarial examples. arXiv preprint arXiv:2009.12724, 2020. 2, 3, 5 +[57] Han Wu, Guanyan Ou, Weibin Wu, and Zibin Zheng. Improving transferable targeted adversarial attacks with model self-enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24615-24624, 2024. 2, 4 +[58] Zihao Xiao, Xianfeng Gao, Chilin Fu, Yinpeng Dong, Wei Gao, Xiaolu Zhang, Jun Zhou, and Jun Zhu. Improving transferability of adversarial patches on face recognition with generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11840-11849, 2021. 3 +[59] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019. 2, 7, 1 + +[60] Zhuoer Xu, Zhangxuan Gu, Jianping Zhang, Shiwen Cui, Changhua Meng, and Weiqiang Wang. Backpropagation path search on adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4640-4650, 2023. 3 +[61] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3877-3887, 2021. 1 +[62] Xiao Yang, Chang Liu, Longlong Xu, Yikai Wang, Yinpeng Dong, Ning Chen, Hang Su, and Jun Zhu. Towards effective adversarial textured 3d meshes on physical face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4119-4128, 2023. 1, 3 +[63] Bangjie Yin, Wenxuan Wang, Taiping Yao, Junfeng Guo, Zelun Kong, Shouhong Ding, Jilin Li, and Cong Liu. Advmakeup: A new imperceptible and transferable attack on face recognition. In Proceedings of the 30th International Joint Conference on Artificial Intelligence, pages 1252-1258, 2021. 1, 3, 6 +[64] Shengming Yuan, Qilong Zhang, Lianli Gao, Yaya Cheng, and Jingkuan Song. Natural color fool: Towards boosting black-box unrestricted attacks. In Advances in Neural Information Processing Systems, 2022. 3 +[65] Dan Zeng, Hailin Shi, Hang Du, Jun Wang, Zhen Lei, and Tao Mei. Npcface: Negative-positive collaborative training for large-scale face recognition. 2020. 6 +[66] Jiaming Zhang, Qi Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5005-5013, 2022. 2 +[67] Xinyu Zhang, Qiang Wang, Jian Zhang, and Zhao Zhong. Adversarial autoaugment. In International Conference on Learning Representations, 2020. 5 +[68] Yaoyao Zhong and Weihong Deng. Adversarial learning with margin-based triplet embedding regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6548-6557, 2019. 7, 1, 2 +[69] Yaoyao Zhong and Weihong Deng. Towards transferable adversarial attack against deep face recognition. IEEE Transactions on Information Forensics and Security, 16:1452-1466, 2021. 1, 2, 3, 7 +[70] Fengfan Zhou, Hefei Ling, Yuxuan Shi, Jiazhong Chen, and Ping Li. Improving visual quality and transferability of adversarial attacks on face recognition simultaneously with adversarial restoration. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4540-4544, 2024. 6 +[71] Fengfan Zhou, Hefei Ling, Yuxuan Shi, Jiazhong Chen, Zongyi Li, and Ping Li. Improving the transferability of adversarial attacks on face recognition with beneficial perturbation feature augmentation. IEEE Transactions on Computational Social Systems, 11(6):8130-8142, 2024. 1, 2, 3, 4, 5, 6, 7 +[72] Fengfan Zhou, Qianyu Zhou, Xiangtai Li, Xuequan Lu, Lizhuang Ma, and Hefei Ling. Adversarial attacks on both face recognition and face anti-spoofing models. arXiv preprint arXiv:2405.16940, 2024. 3 + +[73] Fengfan Zhou, Qianyu Zhou, Bangjie Yin, Hui Zheng, Xuequan Lu, Lizhuang Ma, and Hefei Ling. Rethinking impersonation and dodging attacks on face recognition systems. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2487-2496, 2024. 6 +[74] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Shouhong Ding, and Lizhuang Ma. Adaptive mixture of experts learning for generalizable face anti-spoofing. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6009-6018, 2022. 1 +[75] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Kekai Sheng, Shouhong Ding, and Lizhuang Ma. Generative domain adaptation for face anti-spoofing. In European Conference on Computer Vision, pages 335-356, 2022. +[76] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Xuequan Lu, Ran Yi, Shouhong Ding, and Lizhuang Ma. Instance-aware domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20453-20463, 2023. +[77] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Xuequan Lu, Shouhong Ding, and Lizhuang Ma. Test-time domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 175-187, 2024. 1 +[78] Ziqi Zhou, Shengshan Hu, Minghui Li, Hangtao Zhang, Yechao Zhang, and Hai Jin. Advclip: Downstream-agnostic adversarial examples in multimodal contrastive learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6311–6320, 2023. 2 +[79] Hegui Zhu, Yuchen Ren, Xiaoyan Sui, Lianping Yang, and Wuming Jiang. Boosting adversarial transferability via gradient relevance attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4718-4727, 2023. 3 \ No newline at end of file diff --git a/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/images.zip b/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7adfa2ac3656b791bf5a40026a0a0ef73ad6eba7 --- /dev/null +++ b/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:503b7dd179d07f13e2e82a2c21381667a2142d08db913d11d70d77f507aa44ec +size 454284 diff --git a/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/layout.json b/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b4fdf2e0ab7c051f770409d3ce835a9cd82af7cc --- /dev/null +++ b/2025/Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation/layout.json @@ -0,0 +1,12644 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 102, + 550, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 102, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 550, + 140 + ], + "type": "text", + "content": "Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 160, + 489, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 160, + 489, + 175 + ], + "spans": [ + { + "bbox": [ + 120, + 160, + 489, + 175 + ], + "type": "text", + "content": "Fengfan Zhou" + }, + { + "bbox": [ + 120, + 160, + 489, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 160, + 489, + 175 + ], + "type": "text", + "content": ", Bangjie Yin" + }, + { + "bbox": [ + 120, + 160, + 489, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 160, + 489, + 175 + ], + "type": "text", + "content": ", Hefei Ling" + }, + { + "bbox": [ + 120, + 160, + 489, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 120, + 160, + 489, + 175 + ], + "type": "text", + "content": ", Qianyu Zhou" + }, + { + "bbox": [ + 120, + 160, + 489, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 120, + 160, + 489, + 175 + ], + "type": "text", + "content": ", Wenxuan Wang" + }, + { + "bbox": [ + 120, + 160, + 489, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 175, + 539, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 175, + 539, + 203 + ], + "spans": [ + { + "bbox": [ + 72, + 175, + 539, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 175, + 539, + 203 + ], + "type": "text", + "content": "School of Computer Science and Technology, Huazhong University of Science and Technology; " + }, + { + "bbox": [ + 72, + 175, + 539, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 175, + 539, + 203 + ], + "type": "text", + "content": "Shanghai Shizhuang Information Technology Co., Ltd;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 203, + 509, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 203, + 509, + 232 + ], + "spans": [ + { + "bbox": [ + 100, + 203, + 509, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 100, + 203, + 509, + 232 + ], + "type": "text", + "content": " Department of Computer Science and Engineering, Shanghai Jiao Tong University; " + }, + { + "bbox": [ + 100, + 203, + 509, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 100, + 203, + 509, + 232 + ], + "type": "text", + "content": " School of Computer Science, Northwestern Polytechnical University." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 166, + 232, + 444, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 232, + 444, + 244 + ], + "spans": [ + { + "bbox": [ + 166, + 232, + 444, + 244 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 166, + 232, + 444, + 244 + ], + "type": "text", + "content": " {ffzhou, lhefei}@hust.edu.cn, " + }, + { + "bbox": [ + 166, + 232, + 444, + 244 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 166, + 232, + 444, + 244 + ], + "type": "text", + "content": " jamesyin10@gmail.com," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 184, + 245, + 417, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 245, + 417, + 258 + ], + "spans": [ + { + "bbox": [ + 184, + 245, + 417, + 258 + ], + "type": "text", + "content": "3 zhouqianyu@sjtu.edu.cn, 4wxwang@nwpu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 286, + 200, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 286, + 200, + 300 + ], + "spans": [ + { + "bbox": [ + 152, + 286, + 200, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 312, + 297, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 297, + 601 + ], + "type": "text", + "content": "Face Recognition (FR) models are vulnerable to adversarial examples that subtly manipulate benign face images, underscoring the urgent need to improve the transferability of adversarial attacks in order to expose the blind spots of these systems. Existing adversarial attack methods often overlook the potential benefits of augmenting the surrogate model with diverse initializations, which limits the transferability of the generated adversarial examples. To address this gap, we propose a novel method called Diverse Parameters Augmentation (DPA) attack method, which enhances surrogate models by incorporating diverse parameter initializations, resulting in a broader and more diverse set of surrogate models. Specifically, DPA consists of two key stages: Diverse Parameters Optimization (DPO) and Hard Model Aggregation (HMA). In the DPO stage, we initialize the parameters of the surrogate model using both pre-trained and random parameters. Subsequently, we save the models in the intermediate training process to obtain a diverse set of surrogate models. During the HMA stage, we enhance the feature maps of the diversified surrogate models by incorporating beneficial perturbations, thereby further improving the transferability. Experimental results demonstrate that our proposed attack method can effectively enhance the transferability of the crafted adversarial face examples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 611, + 137, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 611, + 137, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 611, + 137, + 624 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": "Owing to the relentless progress in deep learning, Face Recognition (FR) has achieved substantial advancements [1, 2, 8, 20, 34, 44]. However, the inherent vulnerabilities in existing FR models pose significant security risks [7, 31, 53, 74-77], with adversarial attacks emerging as one" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 328, + 289, + 352, + 300 + ], + "blocks": [ + { + "bbox": [ + 328, + 289, + 352, + 300 + ], + "lines": [ + { + "bbox": [ + 328, + 289, + 352, + 300 + ], + "spans": [ + { + "bbox": [ + 328, + 289, + 352, + 300 + ], + "type": "image", + "image_path": "c47edcb32e1651a36b4d9a872f7d0104cd56232907865e641704ce2e3c7d03a5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 289, + 419, + 298 + ], + "lines": [ + { + "bbox": [ + 354, + 289, + 419, + 298 + ], + "spans": [ + { + "bbox": [ + 354, + 289, + 419, + 298 + ], + "type": "text", + "content": "Pre-trained Parameters" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 289, + 451, + 299 + ], + "blocks": [ + { + "bbox": [ + 427, + 289, + 451, + 299 + ], + "lines": [ + { + "bbox": [ + 427, + 289, + 451, + 299 + ], + "spans": [ + { + "bbox": [ + 427, + 289, + 451, + 299 + ], + "type": "image", + "image_path": "9cb077fee36cf08dad4b10a9a539c89f900eb48b4939f09f4bf6f099d310babe.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 289, + 546, + 299 + ], + "lines": [ + { + "bbox": [ + 452, + 289, + 546, + 299 + ], + "spans": [ + { + "bbox": [ + 452, + 289, + 546, + 299 + ], + "type": "text", + "content": "Randomly Initialized Parameters" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 326, + 311, + 365, + 340 + ], + "blocks": [ + { + "bbox": [ + 326, + 311, + 365, + 340 + ], + "lines": [ + { + "bbox": [ + 326, + 311, + 365, + 340 + ], + "spans": [ + { + "bbox": [ + 326, + 311, + 365, + 340 + ], + "type": "image", + "image_path": "f55d9096223cfc97a092623c4d7879f0699f3b92d3bf6a0bf488615dd7d3faed.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 349, + 369, + 365 + ], + "lines": [ + { + "bbox": [ + 324, + 349, + 369, + 365 + ], + "spans": [ + { + "bbox": [ + 324, + 349, + 369, + 365 + ], + "type": "text", + "content": "(a) Input-based Augmentation" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 386, + 311, + 425, + 340 + ], + "blocks": [ + { + "bbox": [ + 386, + 311, + 425, + 340 + ], + "lines": [ + { + "bbox": [ + 386, + 311, + 425, + 340 + ], + "spans": [ + { + "bbox": [ + 386, + 311, + 425, + 340 + ], + "type": "image", + "image_path": "e3cb750f5feb3254ee84e78fb0f946d9be731f090879f34e46d2c40945800c4d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 349, + 430, + 365 + ], + "lines": [ + { + "bbox": [ + 380, + 349, + 430, + 365 + ], + "spans": [ + { + "bbox": [ + 380, + 349, + 430, + 365 + ], + "type": "text", + "content": "(b) Feature-based Augmentation" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 447, + 311, + 484, + 340 + ], + "blocks": [ + { + "bbox": [ + 447, + 311, + 484, + 340 + ], + "lines": [ + { + "bbox": [ + 447, + 311, + 484, + 340 + ], + "spans": [ + { + "bbox": [ + 447, + 311, + 484, + 340 + ], + "type": "image", + "image_path": "e0e9fe7450a975fe6a6ca4c145b3444df2f280e33fc694b7b0f7c7b991c33268.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 349, + 494, + 365 + ], + "lines": [ + { + "bbox": [ + 436, + 349, + 494, + 365 + ], + "spans": [ + { + "bbox": [ + 436, + 349, + 494, + 365 + ], + "type": "text", + "content": "(c) Parameter-based Augmentation" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 324, + 376, + 394, + 444 + ], + "blocks": [ + { + "bbox": [ + 324, + 376, + 394, + 444 + ], + "lines": [ + { + "bbox": [ + 324, + 376, + 394, + 444 + ], + "spans": [ + { + "bbox": [ + 324, + 376, + 394, + 444 + ], + "type": "image", + "image_path": "9cdc84cef9e805b9334097ac1a1b3127cede10b57bcec2ac9acf2a56a5e3d614.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 453, + 555, + 519 + ], + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 519 + ], + "type": "text", + "content": "Figure 1. Top: comparison between traditional augmentation-based adversarial attack methods and our proposed method. The black pattern filling on the left and right sides of the blue line represents input-based and parameter-based augmentation, respectively. The orange pattern filling indicates feature-based augmentation. Bottom: comparison of performance among 4 types of augmentations." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 406, + 372, + 472, + 444 + ], + "blocks": [ + { + "bbox": [ + 406, + 372, + 472, + 444 + ], + "lines": [ + { + "bbox": [ + 406, + 372, + 472, + 444 + ], + "spans": [ + { + "bbox": [ + 406, + 372, + 472, + 444 + ], + "type": "image", + "image_path": "68b5289689ff9abd4610bb641df32913a4a9fe22d441a4413b70160745f92192.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 484, + 376, + 548, + 444 + ], + "blocks": [ + { + "bbox": [ + 484, + 376, + 548, + 444 + ], + "lines": [ + { + "bbox": [ + 484, + 376, + 548, + 444 + ], + "spans": [ + { + "bbox": [ + 484, + 376, + 548, + 444 + ], + "type": "image", + "image_path": "951477ee640f009ee37f96b5390bfe8ef40e7f2135e45af4812f55966cee0f5b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 542, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 578 + ], + "type": "text", + "content": "of the critical concerns. Therefore, there is an urgent need to bolster the resilience of FR models against adversarial face examples to reveal and address the vulnerabilities." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "As a result, numerous research efforts have been focused on this area. Several adversarial attacks have been developed to create adversarial face examples with features like stealthiness [6, 15, 33, 35, 61], transferability [22, 69-71], and the capacity for physical attacks [21, 62, 63]. These initiatives are aimed at improving the effectiveness of adversarial attacks on FR. Nevertheless, the transferability of these adversarial attacks remains limited. To enhance the transferability of adversarial attacks, augmentation emerges as one of the most effective methods. As illustrated in Figure 1, augmentation-based adversarial attack consists of three aug" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 145, + 713 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3516" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "content": "mentation types: input-based [27, 46, 52, 59], feature-based [69, 71], and parameter-based [12, 57]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 98, + 297, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 98, + 297, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 98, + 297, + 542 + ], + "type": "text", + "content": "Most augmentation-based adversarial attack methods focus on input-based augmentation to improve transferability. Previous research highlights a symmetry between surrogate models used in adversarial attacks and input data in training tasks [9, 24, 71]. In training, data augmentation has proven effective in enhancing model generalization. Drawing on this symmetry, augmenting models for crafting adversarial examples can yield examples with greater transferability. Input-based augmentation can be seen as methods that only augment models at the input layer, demonstrating significant effectiveness in enhancing transferability. However, augmenting surrogate models in deeper layers (i.e., feature-based and parameter-based augmentation) offers a more direct form of augmentation. Nevertheless, few studies explore deep-layer augmentation [12, 57, 71]. As a typical parameter-based augmentation method in deep layers, LGV [12] uses a pre-trained surrogate model and collects multiple parameter sets through additional training epochs with a high, constant learning rate, thereby enhancing the transferability of adversarial examples. Although these parameter-based augmentation methods [12, 57] show promising effectiveness, they face two problems: (1) Static surrogate model initializations: these methods solely augment the surrogate model from pre-trained parameters, limiting the parameter diversity of the surrogate models and thereby hindering the transferability of the crafted adversarial examples. (2) Unavailability of the FR head: modern FR training procedures typically involve training both a backbone model and a head model [8, 44]. Following training, inference is conducted solely with the backbone models, while the head models are often not released as open-source. Consequently, in the majority of instances, we are unable to access the pre-trained parameters for the head models. Traditional parameter-based augmentation adversarial attacks augment models from pretrained parameters [12, 57]. This limitation makes it challenging to craft adversarial examples on FR models with numerous open-sourced FR models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 544, + 297, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 297, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 297, + 640 + ], + "type": "text", + "content": "To address these problems, we introduce a novel adversarial attack called Diverse Parameters Augmentation (DPA) to enhance the transferability of crafted adversarial face examples. Unlike existing parameter-based augmentation adversarial attack methods that overlook the use of diverse parameter initializations [12, 57], we diversify parameter initializations with both random and pre-trained values, thereby improving black-box attack capacity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "text", + "content": "Technically, our proposed attack method comprises two stages: Diverse Parameters Optimization (DPO) and Hard Model Aggregation (HMA). During the DPO stage, we initiate a subset of the optimization parameters with random noise, preserving the refined parameters in the intermediate training process to yield a diverse set of surrogate models," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "content": "which is instrumental in bolstering the transferability of the adversarial examples. In the HMA stage, we add beneficial perturbations [56] with optimization directions opposite to those of the adversarial perturbations onto the feature maps of the parameter-augmented surrogate models, achieving the effect of hard model augmentation [71] to further enhances the transferability of the crafted adversarial examples. Our proposed attack method effectively addresses the challenges posed by the lack of diverse parameters and the absence of the FR head. The comparison between traditional adversarial attack methods and our proposed method is illustrated in Figure 1. By using diverse initializations for surrogate models, we can expand the parameter set of these models, thereby enhancing the transferability of the adversarial examples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 240, + 534, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 240, + 534, + 251 + ], + "spans": [ + { + "bbox": [ + 326, + 240, + 534, + 251 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 251, + 556, + 466 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 251, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 555, + 323 + ], + "type": "text", + "content": "- We introduce a novel perspective that parameter-based augmentation adversarial attacks should augment the parameters of surrogate models by incorporating diverse initialized parameters. To the best of our knowledge, this is the first adversarial attack on FR that utilizes parameter augmentation to enhance transferability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 323, + 556, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 556, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 556, + 430 + ], + "type": "text", + "content": "- We introduce a new adversarial attack method on FR, called DPA, which comprises DPO and HMA stages. The DPO stage uses both pre-trained and random initializations to optimize the model and save the models in the intermediate training process to diversify the surrogate model set. The HMA stage adds beneficial perturbations with optimization directions opposite to those of adversarial perturbations onto the feature maps of the diversified surrogate models to further enhance transferability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 431, + 556, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 556, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 556, + 466 + ], + "type": "text", + "content": "- Extensive experiments reveal that our proposed method attains superior performance when compared with the state-of-the-art adversarial attack methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 478, + 401, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 478, + 401, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 478, + 401, + 491 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 557, + 715 + ], + "type": "text", + "content": "Adversarial Attacks. The primary objective of adversarial attacks is to introduce subtle perturbations into benign images, thereby deceiving machine learning systems and inducing them to produce incorrect predictions [11, 42]. The presence of adversarial examples constitutes a substantial security threat to contemporary machine learning systems. Consequently, significant research efforts have been expended to investigate adversarial attacks, with the aim of bolstering system robustness [10, 23, 27, 28, 30, 36, 66, 78]. To enhance the potency of black-box adversarial attacks, DI [59] incorporates random transformations into adversarial examples at each iteration, effectively achieving data augmentation. VMI-FGSM [50] harnesses gradient variance to stabilize the update process, thereby enhancing black-box attack performance. SSA [27] translates adversarial examples into the frequency domain and applies spectral manipulation for augmentation. SIA [52] introduces random image transformations on each image block, producing di" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3517" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 545, + 293 + ], + "type": "image", + "image_path": "4032ae9fb5acf39f50ad550aa33cbc7ed4053d3efee011085844182784053871.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 298, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 54, + 298, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 555, + 333 + ], + "type": "text", + "content": "Figure 2. The framework of the Diverse Parameters Optimization (DPO). We enhance the diversity of the surrogate model parameters by integrating both pre-trained and random initializations. The method yields a diverse set of surrogate model parameters, which enhances the parameter diversity of the surrogate FR models and consequently improves transferability of the crafted adversarial examples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 349, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 349, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 54, + 349, + 296, + 469 + ], + "type": "text", + "content": "verse variations for gradient estimation. BSR [46] segments the input image into multiple blocks, randomly shuffling and rotating them to generate a set of new images for gradient computation. Despite these advancements, these methods overlook the potential benefit of augmenting the surrogate model with diverse initializations, which limits the transferability of the crafted adversarial examples. In contrast, our proposed method augments surrogate models using both pre-trained and randomly initialized parameters, producing a broader and more diverse set of surrogate models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "content": "Adversarial Attacks on Face Recognition. Adversarial attacks on FR models can be classified based on the constraints imposed on the adversarial perturbations. These attacks are broadly categorized into two types: restricted attacks [4, 9, 25, 26, 28, 60, 72, 79] and unrestricted attacks [3, 5, 37, 40, 43, 54, 55, 64]. Restricted attacks create adversarial examples that adhere to a specified perturbation limit, such as an " + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "content": " norm bound. Our proposed attack method is a restricted attack. Consequently, we will delve into the specifics of restricted attacks in detail. To enhance the transferability of adversarial attacks on FR, Zhong and Deng [69] introduced DFANet, which employs dropout on the feature maps of convolutional layers to achieve an ensemble-like effect. Zhou et al. [71] proposed BPFA, enhancing attack transferability by integrating beneficial perturbations [56] onto the feature maps of FR models, resulting in the effect of hard model augmentation. Li et al. [22] leveraged additional information from FR-related tasks and applied a multi-task optimization framework to further improve the transferability of adversarial examples. Unrestricted adversarial attacks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 349, + 556, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 556, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 556, + 506 + ], + "type": "text", + "content": "conversely generate adversarial examples without the limitations of a predefined perturbation bound. These attacks are primarily focused on physical attacks [21, 58, 62], attribute editing [18, 33], and adversarial example generation through makeup transfer [15, 35, 39, 63]. Both restricted and unrestricted adversarial attacks on FR models have substantially advanced the capabilities of these attacks. However, existing methods for generating adversarial attacks on FR often rely on surrogate models with fixed parameters, which limits the transferability of the crafted adversarial examples. In contrast, our proposed attack method addresses this limitation by diversifying the surrogate model parameters, using models with varying parameters across different epochs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 519, + 398, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 398, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 398, + 533 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 540, + 518, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 518, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 518, + 553 + ], + "type": "text", + "content": "3.1. Problem Formulation and Framework" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "text", + "content": "Problem Formulation. Let " + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{vct}(\\mathbf{x}') \\in \\mathbb{R}^r" + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "text", + "content": " denote the FR model employed by the victim to extract the embedding from a face image " + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "text", + "content": ". We denote " + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^s" + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^t" + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "text", + "content": " as the source and target images, respectively. The objective of the adversarial attacks explored in our research is to manipulate " + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{vct}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "text", + "content": " to misclassify the adversarial example " + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{adv}" + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "text", + "content": " as the target image " + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^t" + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "text", + "content": ", while ensuring that " + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{adv}" + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "text", + "content": " bears a close visual resemblance to " + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^s" + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "text", + "content": ". The objective of adversarial attack as delineated in this paper is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 677, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 326, + 677, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {x} ^ {a d v} = \\underset {\\mathbf {x} ^ {a d v}} {\\arg \\min } \\left(\\mathcal {D} \\left(\\mathcal {F} ^ {v c t} \\left(\\mathbf {x} ^ {a d v}\\right), \\mathcal {F} ^ {v c t} \\left(\\mathbf {x} ^ {t}\\right)\\right)\\right) \\tag {1} \\\\ \\mathrm {s . t .} \\| \\mathbf {x} ^ {a d v} - \\mathbf {x} ^ {s} \\| _ {p} \\leqslant \\epsilon \\\\ \\end{array}", + "image_path": "e41d308987e5283b4629046073ce3df5cd71cdd2c82ce7e5e76aa846d0f01f4f.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3518" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "where the symbol " + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": " denotes a predefined distance metric employed for the optimization of adversarial face examples." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 191 + ], + "type": "text", + "content": "In practice, the attacker typically cannot access the model owned by the victim. As a result, black-box attacks play a crucial role in adversarial attacks [32, 48, 49]. A widely adopted approach for black-box attacks involves using a surrogate model " + }, + { + "bbox": [ + 55, + 96, + 296, + 191 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 96, + 296, + 191 + ], + "type": "text", + "content": " to generate adversarial examples and then transfer the crafted adversarial examples to the victim model [50, 71]. Consequently, the transferability is important and constitutes the key problem studied in this research." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 192, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 371 + ], + "type": "text", + "content": "Framework Overview. To enhance the transferability of adversarial examples, we propose a novel method called Diverse Parameters Augmentation (DPA). DPA augments surrogate models by incorporating both pre-trained and randomly initialized parameters, resulting in a more diverse and expansive set of surrogate models. Specifically, DPA consists of two key stages: Diverse Parameters Optimization (DPO) and Hard Model Aggregation (HMA). In the DPO stage, we diversify the surrogate model parameters by combining pre-trained and random initializations, as illustrated in Figure 2. In the HMA stage, we apply beneficial perturbations to the feature maps of the diversified surrogate models and combine them to achieve a higher degree of augmentation, as shown in Figure 3. In the following, we will introduce our proposed DPA attack method in detail." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 378, + 235, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 235, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 235, + 390 + ], + "type": "text", + "content": "3.2. Diverse Parameters Optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 538 + ], + "type": "text", + "content": "The parameter diversity of surrogate models is crucial for the transferability of crafted adversarial examples. Previous parameter-based augmentation adversarial attack methods solely initialize parameters with pre-trained values, thereby limiting the parameter diversity of surrogate models [12, 57]. In contrast, the key innovation of the stage of our proposed attack method is the diversification of initialized parameters using both pre-trained and randomly initialized parameters. Following initialization, we augment the parameters of the surrogate models using the parameters in the intermediate training process to diversify the surrogate models, thereby improving the transferability of the adversarial examples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": " be a batch of face images, " + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\mathbb{R}^{s \\times r}" + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": " be the parameters of the FR head, " + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": " be the batch size of " + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": " be the class number in the training dataset. To calculate the loss function, we should first calculate the cosine similarity matrix " + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{r} \\in \\mathbb{R}^{b \\times s}" + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{x}) \\in \\mathbb{R}^{b \\times r}" + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": " and the parameters of the FR head " + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{\\top}" + }, + { + "bbox": [ + 55, + 539, + 296, + 611 + ], + "type": "text", + "content": " using the following formula:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 617, + 295, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 617, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 137, + 617, + 295, + 631 + ], + "type": "interline_equation", + "content": "\\mathbf {r} = \\cos \\mathbf {a} = \\bar {\\mathbf {e}} \\bar {\\mathbf {w}} ^ {\\top} \\tag {2}", + "image_path": "4a3192ea9734b0603c95bb8d585bda33ed14cd33cd91d06c509f1febd4b3da9d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{e}} = \\frac{\\mathcal{F}(\\mathbf{x})}{\\|\\mathcal{F}(\\mathbf{x})\\|}, \\overline{\\mathbf{w}} = \\frac{\\mathbf{w}}{\\|\\mathbf{w}\\|}" + }, + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "text", + "content": " is the angle between " + }, + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{\\top}" + }, + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "text", + "content": ". After obtaining the cosine similarity matrix by Equation (2), we can get the sine similarity between " + }, + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 639, + 296, + 693 + ], + "type": "text", + "content": " using the following equation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 700, + 296, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 700, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 127, + 700, + 296, + 714 + ], + "type": "interline_equation", + "content": "\\sin \\mathbf {a} = \\sqrt {1 . 0 - \\cos^ {2} \\mathbf {a}} \\tag {3}", + "image_path": "fa92cdb0eb2237e2479770d1077f4665a44d441c8e95e6af663cd5ed7bc1f4c4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "content": "Using the cosine and sine similarity matrix, the formula to get the additive angular margin cosine similarity matrix can be expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 350, + 116, + 555, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 116, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 350, + 116, + 555, + 129 + ], + "type": "interline_equation", + "content": "\\cos (\\mathbf {a} + m) = \\cos \\mathbf {a} \\cos m - \\sin \\mathbf {a} \\sin m \\tag {4}", + "image_path": "91478d4ec4db62b993e811f4d6f166d1d5f8794ae64c2120aa40c03fcb184be4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 136, + 555, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 136, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 313, + 136, + 555, + 184 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 136, + 555, + 184 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 136, + 555, + 184 + ], + "type": "text", + "content": " is the margin value for increasing the discrimination of the FR model [8]. Let " + }, + { + "bbox": [ + 313, + 136, + 555, + 184 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathbb{R}^b" + }, + { + "bbox": [ + 313, + 136, + 555, + 184 + ], + "type": "text", + "content": " be the labels of the current batch. By utilizing " + }, + { + "bbox": [ + 313, + 136, + 555, + 184 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 136, + 555, + 184 + ], + "type": "text", + "content": ", we can calculate its one-hot encoded matrix using the following formula:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 394, + 190, + 555, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 190, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 394, + 190, + 555, + 205 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathbf {h}} = \\Psi (\\mathbf {y}) \\in \\mathbb {R} ^ {b \\times s} \\tag {5}", + "image_path": "d3e1822c0e4b19a0ef5e980f6cede5eab1ecb3bd672b87131e2ab3563cc1509c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 212, + 554, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 554, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 554, + 236 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 212, + 554, + 236 + ], + "type": "inline_equation", + "content": "\\Psi (\\cdot)" + }, + { + "bbox": [ + 313, + 212, + 554, + 236 + ], + "type": "text", + "content": " is the one-hot encode operation. Next, we use the following function to calculate the output of the head:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 341, + 242, + 554, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 242, + 554, + 290 + ], + "spans": [ + { + "bbox": [ + 341, + 242, + 554, + 290 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {q} = d \\left(\\hat {\\mathbf {h}} \\odot \\mathbf {p} + (1 - \\hat {\\mathbf {h}}) \\odot \\mathbf {r}\\right) \\\\ \\mathbf {p} = \\left\\{ \\begin{array}{l l} \\cos (\\mathbf {a} + m) & \\text {s . t .} \\quad \\mathbf {a} < \\pi - m, \\\\ \\cos \\mathbf {a} - m \\sin m & \\text {s . t .} \\quad \\mathbf {a} \\geqslant \\pi - m, \\end{array} \\right. \\tag {6} \\\\ \\end{array}", + "image_path": "c5d024ec9140d3d067c0f85718bfbdf0b6394e27ebe83bb6a8de0fab7e7e26c0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 297, + 555, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 333 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 297, + 555, + 333 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 297, + 555, + 333 + ], + "type": "text", + "content": " is a pre-defined scale factor. Using the output of the head, the formula to calculate the loss function can be expressed as the following:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 363, + 340, + 555, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 340, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 363, + 340, + 555, + 371 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathbf {x}, \\mathbf {y}) = - \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\log \\frac {e ^ {\\mathbf {q} _ {i , \\mathbf {y} _ {i}}}}{\\sum_ {j = 1} ^ {s} e ^ {\\mathbf {q} _ {i , j}}} \\tag {7}", + "image_path": "038b489032bcf9ceeb385fdca04156b0c6e5e06cc72191f13032502c79105b6f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 377, + 555, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 414 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 377, + 555, + 414 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 377, + 555, + 414 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 377, + 555, + 414 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 377, + 555, + 414 + ], + "type": "text", + "content": " be the number of the training epoch and the parameters of the FR backbone, respectively. Utilizing Equation (7), we can get the parameters set [8]:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 333, + 420, + 545, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 420, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 333, + 420, + 545, + 447 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {i}, \\mathbf {w} _ {i} = \\varphi \\left(\\underset {\\mathbf {v}, \\mathbf {w}} {\\arg \\min } \\mathcal {L} \\left(\\mathbf {x}, \\mathbf {y}\\right), i\\right) \\quad i \\in \\{1, 2, \\ldots , c \\}", + "image_path": "a18dbd56a4e52cc1c15a2c227787d976d0d8eeda85d47d6fd6f8bc82ac5a3225.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 449, + 565, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 449, + 565, + 472 + ], + "spans": [ + { + "bbox": [ + 315, + 449, + 565, + 472 + ], + "type": "interline_equation", + "content": "\\mathbf {V} _ {c} = \\left\\{\\mathbf {v} _ {1}, \\mathbf {v} _ {2}, \\dots , \\mathbf {v} _ {c - 1}, \\mathbf {v} _ {c} \\right\\} \\quad \\mathbf {W} _ {c} = \\left\\{\\mathbf {w} _ {1}, \\mathbf {w} _ {2}, \\dots , \\mathbf {w} _ {c - 1}, \\mathbf {w} _ {c} \\right\\} \\tag {8}", + "image_path": "c2d5f754ccec8ff1a57da78347ef8f7739cefe8ed62f8877386654adf2741dec.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\varphi (\\cdot ,i)" + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "content": " is the operation that obtain the " + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "content": " parameters in the " + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "content": "th epoch. We denote " + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0" + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_0" + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "content": " as the initialized parameters of the FR model and head, respectively. If " + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0" + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_0" + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "content": " are fixed, " + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_c" + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_c" + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "content": " become deterministic aside from minor random factors in the training process of Equation (8). Therefore, there exists a mapping between " + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}_c,\\mathbf{W}_c\\}" + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_0,\\mathbf{w}_0\\}" + }, + { + "bbox": [ + 313, + 473, + 555, + 557 + ], + "type": "text", + "content": " that can be expressed as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 373, + 564, + 555, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 564, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 373, + 564, + 555, + 578 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {V} _ {c}, \\mathbf {W} _ {c} \\right\\} = F \\left(\\left\\{\\mathbf {v} _ {0}, \\mathbf {w} _ {0} \\right\\}, i\\right) \\tag {9}", + "image_path": "3958b8ed8a0272b16d9dd673c1974652690308d2f9500e1ec0f06413358ad087.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 585, + 555, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 555, + 681 + ], + "type": "text", + "content": "Existing adversarial attack methods based on parameter augmentation typically initialize the parameters " + }, + { + "bbox": [ + 313, + 585, + 555, + 681 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_0,\\mathbf{w}_0\\}" + }, + { + "bbox": [ + 313, + 585, + 555, + 681 + ], + "type": "text", + "content": " solely with pre-trained values [12, 57]. This constraint diminishes the diversity within the augmented surrogate model set. In contrast, we opt to initialize " + }, + { + "bbox": [ + 313, + 585, + 555, + 681 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_0,\\mathbf{w}_0\\}" + }, + { + "bbox": [ + 313, + 585, + 555, + 681 + ], + "type": "text", + "content": " with diverse parameters to enhance the diversity of the augmented surrogate model set, thereby improving transferability. The " + }, + { + "bbox": [ + 313, + 585, + 555, + 681 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}_c,\\mathbf{W}_c\\}" + }, + { + "bbox": [ + 313, + 585, + 555, + 681 + ], + "type": "text", + "content": " of our proposed attack can be expressed as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 372, + 687, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 687, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 372, + 687, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left\\{\\mathbf {V} _ {c} ^ {p}, \\mathbf {W} _ {c} ^ {p} \\right\\} = F \\left(\\left\\{\\mathbf {v} _ {0} ^ {p}, \\mathbf {w} _ {0} ^ {p} \\right\\}, i\\right) \\tag {10} \\\\ \\left\\{\\mathbf {V} _ {c} ^ {a}, \\mathbf {W} _ {c} ^ {a} \\right\\} = F \\left(\\left\\{\\mathbf {v} _ {0} ^ {a}, \\mathbf {w} _ {0} ^ {a} \\right\\}, i\\right) \\\\ \\end{array}", + "image_path": "ac0de3433ac12187c23cccf99b960606cee926aa665ab0a33a2a6b8e58cf106a.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3519" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 71, + 547, + 168 + ], + "blocks": [ + { + "bbox": [ + 66, + 71, + 547, + 168 + ], + "lines": [ + { + "bbox": [ + 66, + 71, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 547, + 168 + ], + "type": "image", + "image_path": "db00f76aa6e1fab04667b020fc5c4f89d78f35a29b43e361205fb5377649a297.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "type": "text", + "content": "Figure 3. The framework of the Hard Model Aggregation (HMA). After acquiring a surrogate model set with diverse parameters (i.e., " + }, + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_c^q" + }, + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "type": "text", + "content": "), we introduce beneficial perturbations with the optimization direction opposite to that of adversarial perturbations onto the feature maps of these diversified surrogate models, transforming them into hard models and aggregate the hard models to increase the transferability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0^p" + }, + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "text", + "content": " represents the pre-trained parameters, while " + }, + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_0^p" + }, + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0^a" + }, + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_0^a" + }, + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "text", + "content": " are randomly initialized. Using Equation (10), we can obtain the more diverse parameters for the FR model and head. After obtaining " + }, + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_c^p" + }, + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_c^a" + }, + { + "bbox": [ + 55, + 220, + 296, + 281 + ], + "type": "text", + "content": ", we use the following set of the parameters to craft the adversarial examples:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 291, + 296, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 291, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 61, + 291, + 296, + 323 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {V} _ {c} ^ {q} = \\left\\{\\mathbf {v} _ {0} ^ {p}, \\mathbf {v} _ {1} ^ {p}, \\mathbf {v} _ {j _ {1}} ^ {p}, \\dots , \\mathbf {v} _ {c} ^ {p} \\right\\} \\bigcup \\left\\{\\mathbf {v} _ {1} ^ {a}, \\mathbf {v} _ {j _ {1}} ^ {a}, \\mathbf {v} _ {j _ {2}} ^ {a}, \\dots , \\mathbf {v} _ {c} ^ {a} \\right\\} \\tag {11} \\\\ \\mathrm {s . t .} j \\in \\{1, 2, \\dots , c | j \\bmod \\kappa = 1 \\} \\\\ \\end{array}", + "image_path": "451e3e8a9319f32ba07eedfdd3f19292ce1e4265630a1692c21dfe4bd12b1266.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 333, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 296, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 333, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\kappa = \\lfloor \\sqrt{c} \\rfloor" + }, + { + "bbox": [ + 55, + 333, + 296, + 381 + ], + "type": "text", + "content": " determines the epoch interval to select the parameters. The overall framework and pseudo-code of the DPO stage in our proposed method is illustrated in Figure 2 and Algorithm 1, respectively." + } + ] + } + ], + "index": 4 + }, + { + "type": "code", + "bbox": [ + 56, + 408, + 296, + 446 + ], + "blocks": [ + { + "bbox": [ + 56, + 393, + 274, + 406 + ], + "lines": [ + { + "bbox": [ + 56, + 393, + 274, + 406 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 274, + 406 + ], + "type": "text", + "content": "Algorithm 1 Diverse Parameters Optimization (DPO)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "lines": [ + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "text", + "content": "Input: Diverse initial parameters " + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0^p" + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_0^p" + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0^a" + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_0^a" + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "text", + "content": ", the number of epochs " + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "text", + "content": ", the optimizer " + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "text", + "content": ", the FR training dataset " + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "text", + "content": ", loss function " + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 56, + 408, + 296, + 446 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_body" + } + ], + "index": 6, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "bbox": [ + 56, + 446, + 218, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 446, + 218, + 457 + ], + "spans": [ + { + "bbox": [ + 56, + 446, + 218, + 457 + ], + "type": "text", + "content": "Output: Augmented parameter set " + }, + { + "bbox": [ + 56, + 446, + 218, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_c^q" + }, + { + "bbox": [ + 56, + 446, + 218, + 457 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 60, + 457, + 296, + 647 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 63, + 457, + 249, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 457, + 249, + 471 + ], + "spans": [ + { + "bbox": [ + 63, + 457, + 249, + 471 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 63, + 457, + 249, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\mathbf{v}_0^p, \\mathbf{w}_0^p\\}" + }, + { + "bbox": [ + 63, + 457, + 249, + 471 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 63, + 457, + 249, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = \\{\\mathbf{v}_0^a, \\mathbf{w}_0^a\\}" + }, + { + "bbox": [ + 63, + 457, + 249, + 471 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 63, + 457, + 249, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_c^q = \\{\\mathbf{v}_0^p\\}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 470, + 150, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 470, + 150, + 482 + ], + "spans": [ + { + "bbox": [ + 63, + 470, + 150, + 482 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 63, + 470, + 150, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{J} \\in \\{\\mathbf{P}, \\mathbf{A}\\}" + }, + { + "bbox": [ + 63, + 470, + 150, + 482 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 481, + 132, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 481, + 132, + 494 + ], + "spans": [ + { + "bbox": [ + 63, + 481, + 132, + 494 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 63, + 481, + 132, + 494 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v},\\mathbf{w}\\} = \\mathbf{J}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 494, + 291, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 494, + 291, + 506 + ], + "spans": [ + { + "bbox": [ + 63, + 494, + 291, + 506 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 63, + 494, + 291, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\Psi (\\mathbf{v})\\quad \\triangleright" + }, + { + "bbox": [ + 63, + 494, + 291, + 506 + ], + "type": "text", + "content": " Map the parameters to the models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 506, + 156, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 506, + 156, + 517 + ], + "spans": [ + { + "bbox": [ + 63, + 506, + 156, + 517 + ], + "type": "text", + "content": "5: for " + }, + { + "bbox": [ + 63, + 506, + 156, + 517 + ], + "type": "inline_equation", + "content": "i = 1,\\dots,c" + }, + { + "bbox": [ + 63, + 506, + 156, + 517 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 63, + 518, + 146, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 518, + 146, + 529 + ], + "spans": [ + { + "bbox": [ + 63, + 518, + 146, + 529 + ], + "type": "text", + "content": "6: for " + }, + { + "bbox": [ + 63, + 518, + 146, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\zeta" + }, + { + "bbox": [ + 63, + 518, + 146, + 529 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 63, + 529, + 231, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 529, + 231, + 541 + ], + "spans": [ + { + "bbox": [ + 63, + 529, + 231, + 541 + ], + "type": "text", + "content": "7: Calculate " + }, + { + "bbox": [ + 63, + 529, + 231, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 63, + 529, + 231, + 541 + ], + "type": "text", + "content": " using Equation (7)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 63, + 541, + 291, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 541, + 291, + 553 + ], + "spans": [ + { + "bbox": [ + 63, + 541, + 291, + 553 + ], + "type": "text", + "content": "8: Backward " + }, + { + "bbox": [ + 63, + 541, + 291, + 553 + ], + "type": "inline_equation", + "content": "(\\mathbf{T},\\mathcal{L})\\triangleright" + }, + { + "bbox": [ + 63, + 541, + 291, + 553 + ], + "type": "text", + "content": " Backpropagation using " + }, + { + "bbox": [ + 63, + 541, + 291, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 63, + 554, + 219, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 554, + 219, + 566 + ], + "spans": [ + { + "bbox": [ + 63, + 554, + 219, + 566 + ], + "type": "text", + "content": "9: Update the parameters " + }, + { + "bbox": [ + 63, + 554, + 219, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{v},\\mathbf{w}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 63, + 566, + 126, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 566, + 126, + 576 + ], + "spans": [ + { + "bbox": [ + 63, + 566, + 126, + 576 + ], + "type": "text", + "content": "0: end for" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 61, + 577, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 577, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 61, + 577, + 296, + 601 + ], + "type": "text", + "content": "11: if the " + }, + { + "bbox": [ + 61, + 577, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 61, + 577, + 296, + 601 + ], + "type": "text", + "content": " satisfies the condition outlined in Equation (11) then" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 60, + 601, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 601, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 60, + 601, + 296, + 613 + ], + "type": "text", + "content": "12: Add " + }, + { + "bbox": [ + 60, + 601, + 296, + 613 + ], + "type": "inline_equation", + "content": "(\\mathbf{V}_{c}^{q},\\mathbf{v})" + }, + { + "bbox": [ + 60, + 601, + 296, + 613 + ], + "type": "text", + "content": " Incorporate " + }, + { + "bbox": [ + 60, + 601, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 60, + 601, + 296, + 613 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 60, + 601, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_c^q" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 60, + 614, + 120, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 614, + 120, + 624 + ], + "spans": [ + { + "bbox": [ + 60, + 614, + 120, + 624 + ], + "type": "text", + "content": "13: end if" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 60, + 625, + 116, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 625, + 116, + 636 + ], + "spans": [ + { + "bbox": [ + 60, + 625, + 116, + 636 + ], + "type": "text", + "content": "14: end for" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 60, + 638, + 106, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 638, + 106, + 647 + ], + "spans": [ + { + "bbox": [ + 60, + 638, + 106, + 647 + ], + "type": "text", + "content": "15: end for" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 671, + 196, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 196, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 196, + 685 + ], + "type": "text", + "content": "3.3. Hard Model Aggregation" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Hard models [71] refer to models with higher loss values compared to normal models, designed to simulate the effect" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 222, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 555, + 293 + ], + "type": "text", + "content": "of hard samples, which are hard to the optimization process of the adversarial examples. Since hard samples can improve the generalization of models, we leverage multiple hard models to enhance the transferability of the surrogate models with diverse parameters by adding beneficial perturbations [56] onto the pre-defined feature maps of the surrogate models." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 293, + 554, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 554, + 329 + ], + "type": "text", + "content": "In the following, we will provide a detailed introduction to the HMA stage. Let " + }, + { + "bbox": [ + 313, + 293, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 293, + 554, + 329 + ], + "type": "text", + "content": " be the mapping from the parameters to the corresponding models that can be expressed as:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 367, + 335, + 555, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 335, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 367, + 335, + 555, + 350 + ], + "type": "interline_equation", + "content": "\\mathbf {F} = \\Psi \\left(\\mathbf {V} _ {c} ^ {q}\\right) = \\left\\{\\mathcal {F} _ {1}, \\mathcal {F} _ {2}, \\dots , \\mathcal {F} _ {g} \\right\\} \\tag {12}", + "image_path": "fffe83993e2573ad062be116fe0a684166286c3fcc20b97ca28b8a254e9f2bdc.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 354, + 556, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 354, + 556, + 474 + ], + "type": "inline_equation", + "content": "g = |\\mathbf{V}_c^q|" + }, + { + "bbox": [ + 313, + 354, + 556, + 474 + ], + "type": "text", + "content": ". Hard sample augmentation has demonstrated significant effectiveness in enhancing model generalization [41, 67]. Based on the relationship between adversarial attack tasks and training tasks [9, 24], using hard models for augmentation can lead to improved transferability [71]. Consequently, by transforming vanilla parameter-augmented surrogate models into hard parameter-augmented surrogate models, we can achieve more transferable results. To accomplish this transformation, we use the following formula to calculate the loss function:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 328, + 478, + 554, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 478, + 554, + 523 + ], + "spans": [ + { + "bbox": [ + 328, + 478, + 554, + 523 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widetilde {\\mathcal {L}} _ {t} = \\frac {1}{g} \\sum_ {i = 0} ^ {g} \\left\\| \\phi \\left(\\mathcal {H} _ {i} \\left(\\mathcal {T} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right)\\right)\\right) - \\phi \\left(\\mathcal {F} _ {i} \\left(\\mathbf {x} ^ {t}\\right)\\right) \\right\\| \\tag {13} \\\\ \\begin{array}{l l} \\text {s . t .} & \\mathcal {F} _ {i} \\in \\mathbf {F} \\quad t \\in \\{1, 2, \\dots , n \\} \\end{array} \\\\ \\end{array}", + "image_path": "7dd06476f18186afff3be7be6e09a9c09b8368a60f062cba84e980abaf0e8f0f.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^{adv}" + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "text", + "content": " denotes the adversarial example generated during the " + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "text", + "content": "th iteration, the variable " + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "text", + "content": " represents the maximum number of iterations allocated for crafting the adversarial examples, " + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "text", + "content": " denotes the normalization operation, " + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "text", + "content": " signifies the input transformation, and " + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "text", + "content": " is the corresponding hard model [71] of " + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 530, + 556, + 602 + ], + "type": "text", + "content": ", which can be expressed as:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 325, + 607, + 555, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 607, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 325, + 607, + 555, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {H} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right) = \\left\\{ \\begin{array}{l l} \\mathcal {F} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right) & \\text {s . t .} \\quad t = 1, \\\\ \\chi \\left(\\mathcal {F} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right), \\Omega\\right) & \\text {s . t .} \\quad t > 1, \\end{array} \\right. \\tag {14}", + "image_path": "ce982ef1d098d769cfbfc3ce0c10018d1a4254f1cb284ead50524b9d4a63075b.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 639, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 555, + 674 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 639, + 555, + 674 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 313, + 639, + 555, + 674 + ], + "type": "text", + "content": " is the mapping for adding beneficial perturbations on the feature maps in the forward propagation whose formula can be expressed as:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 331, + 680, + 555, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 680, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 331, + 680, + 555, + 713 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\chi \\left(\\mathcal {F} \\left(\\mathbf {x} _ {t} ^ {a d v}\\right), \\Omega\\right): \\omega = \\omega + \\eta \\text {s i g n} \\left(\\nabla_ {\\omega} \\tilde {\\mathcal {L}} _ {t - 1}\\right) \\tag {15} \\\\ \\begin{array}{c c} \\text {s . t .} & \\omega \\in \\Omega \\end{array} \\\\ \\end{array}", + "image_path": "226040f7ed3e1344cc1f7331cc1b1e12adbccbbd9539600022aba4af828775a2.jpg" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3520" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " is the pre-defined set of layers for adding beneficial perturbations. Let " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " be the maximum allowable perturbation. Utilizing " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_t" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ", we can craft the adversarial face examples using following formula:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 90, + 128, + 296, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 128, + 296, + 155 + ], + "spans": [ + { + "bbox": [ + 90, + 128, + 296, + 155 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t + 1} ^ {a d v} = \\prod_ {\\mathbf {x} ^ {s}, \\epsilon} \\left(\\mathbf {x} _ {t} ^ {a d v} - \\beta \\operatorname {s i g n} \\left(\\nabla_ {\\mathbf {x} _ {t} ^ {a d v}} \\tilde {\\mathcal {L}} _ {t}\\right)\\right) \\tag {16}", + "image_path": "10bb6162cd2502beed783e28cfc1dcfef61b568f7ded67ad207858ecd47c23fb.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "text", + "content": " denotes the step size used to construct the adversarial examples, while " + }, + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "inline_equation", + "content": "\\prod" + }, + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "text", + "content": " represents the clipping operation that confines the pixel values of the generated adversarial examples within the range " + }, + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "inline_equation", + "content": "[\\mathbf{x}^s -\\epsilon ,\\mathbf{x}^s +\\epsilon ]" + }, + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "text", + "content": ". The framework of the HMA is illustrated in Figure 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 223, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 330 + ], + "type": "text", + "content": "In the following, we present the pseudo-code for the HMA. For the sake of clarity, we focus on neural networks with a single branch in their computational graphs. In the case of neural networks with multiple branches, the HMA algorithm remains largely unchanged, except for the incorporation of mechanisms to handle multiple branches. Before proceeding, we introduce some necessary definitions. Let " + }, + { + "bbox": [ + 55, + 223, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 223, + 296, + 330 + ], + "type": "text", + "content": " denote the index set of pre-selected layers to which beneficial perturbations will be added:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 339, + 295, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 339, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 151, + 339, + 295, + 353 + ], + "type": "interline_equation", + "content": "\\Phi = \\varkappa (\\Omega) \\tag {17}", + "image_path": "179a25cc94352b25aaddabefe14333def0aba0d7ef217e49afddc075398b9f39.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\varkappa" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": " represents the mapping from the pre-defined set of layers for adding beneficial perturbations " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": ", to their corresponding layer index set " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "f^i" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": "th layer within the model " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": ", and let " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": " represent the total number of layers in " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": ". We define the segment of " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": " from layer " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "f^i" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": " to layer " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "f^j" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 443, + 295, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 443, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 295, + 457 + ], + "type": "interline_equation", + "content": "\\mathcal {F} ^ {i, j} = f ^ {i} \\circ f ^ {i + 1} \\circ \\dots \\circ f ^ {j - 1} \\circ f ^ {j} \\tag {18}", + "image_path": "85ec06268a88def29b5a52d663f69bf3522f620dc1c8723b7123311f90537b50.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 463, + 295, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 487 + ], + "type": "text", + "content": "The pseudo-code for the HMA stage of our proposed method is presented in Algorithm 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 497, + 135, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 135, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 135, + 510 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 517, + 178, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 178, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 178, + 529 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 533, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 629 + ], + "type": "text", + "content": "Datasets. We opt to use the LFW [16], and CelebA-HQ [19] for our experiments to verify the effectiveness of our proposed attack method. LFW serves as an unconstrained face dataset for FR. CelebA-HQ consists of face images with high visual quality. The LFW and CelebA-HQ utilized in our experiments are identical to those employed in [70, 71, 73]. For the parameter augmentation process, we select BUPT-Balancedface [47] as our training dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Face Recognition Models. The normal trained FR models employed in our experiments encompass IR152 [13], IRSE50 [14], FaceNet [34], MobileFace [8], CurricularFace [17], MagFace [29], ArcFace [8], CircleLoss [38], MV-Softmax [51], and NPCFace [65]. Specifically, IR152, FaceNet, IRSE50, and MobileFace are the same models utilized in [15, 63, 70, 71]. CurricularFace, MagFace, ArcFace," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 72, + 504, + 84 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 504, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 504, + 84 + ], + "type": "text", + "content": "Algorithm 2 Hard Model Aggregation (HMA)" + } + ] + } + ], + "index": 12, + "type": "text" + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "text", + "content": "Input: The source image " + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^s" + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "text", + "content": ", the target image " + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^t" + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "text", + "content": ", the mapping from the parameters to the corresponding models " + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "text", + "content": ", the maximum number of iterations " + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "text", + "content": ", the index set of pre-selected layers to be added beneficial perturbations " + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "text", + "content": ", the step size of the beneficial perturbations " + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "text", + "content": ", the step size of the adversarial perturbations " + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "text", + "content": ", the maximum permissible magnitude of perturbation " + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "text", + "content": ", the parameter set " + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_c^q" + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "text", + "content": ", the total number of layers in a single surrogate model " + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "text", + "content": ", normalization operation " + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 315, + 88, + 555, + 195 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 196, + 472, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 196, + 472, + 209 + ], + "spans": [ + { + "bbox": [ + 315, + 196, + 472, + 209 + ], + "type": "text", + "content": "Output: An adversarial example " + }, + { + "bbox": [ + 315, + 196, + 472, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_n^{adv}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 209, + 553, + 471 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 321, + 209, + 488, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 209, + 488, + 220 + ], + "spans": [ + { + "bbox": [ + 321, + 209, + 488, + 220 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 321, + 209, + 488, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{adv} = \\mathbf{x}^s" + }, + { + "bbox": [ + 321, + 209, + 488, + 220 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 321, + 209, + 488, + 220 + ], + "type": "inline_equation", + "content": "u = |\\Phi|, s_1 = 1, g = |\\mathbf{V}_c^q|" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 220, + 550, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 220, + 550, + 232 + ], + "spans": [ + { + "bbox": [ + 321, + 220, + 550, + 232 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 321, + 220, + 550, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\Psi (\\mathbf{V}_c^q)" + }, + { + "bbox": [ + 321, + 220, + 550, + 232 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 321, + 220, + 550, + 232 + ], + "type": "text", + "content": " Acquire the diversified models set." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 232, + 407, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 232, + 407, + 243 + ], + "spans": [ + { + "bbox": [ + 321, + 232, + 407, + 243 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 321, + 232, + 407, + 243 + ], + "type": "inline_equation", + "content": "t = 1, \\dots, n" + }, + { + "bbox": [ + 321, + 232, + 407, + 243 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 244, + 416, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 244, + 416, + 255 + ], + "spans": [ + { + "bbox": [ + 321, + 244, + 416, + 255 + ], + "type": "text", + "content": "4: for " + }, + { + "bbox": [ + 321, + 244, + 416, + 255 + ], + "type": "inline_equation", + "content": "i = 1,\\dots,g" + }, + { + "bbox": [ + 321, + 244, + 416, + 255 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 256, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 256, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 321, + 256, + 553, + 277 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 321, + 256, + 553, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\mathbf{F}_i" + }, + { + "bbox": [ + 321, + 256, + 553, + 277 + ], + "type": "text", + "content": " Derive the parameter-augmented model." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 279, + 421, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 279, + 421, + 293 + ], + "spans": [ + { + "bbox": [ + 321, + 279, + 421, + 293 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 321, + 279, + 421, + 293 + ], + "type": "inline_equation", + "content": "\\omega^{t,0} = \\mathcal{T}\\left(\\mathbf{x}_{t - 1}^{adv}\\right)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 293, + 428, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 293, + 428, + 303 + ], + "spans": [ + { + "bbox": [ + 321, + 293, + 428, + 303 + ], + "type": "text", + "content": "7: for " + }, + { + "bbox": [ + 321, + 293, + 428, + 303 + ], + "type": "inline_equation", + "content": "j = 1,\\dots,u" + }, + { + "bbox": [ + 321, + 293, + 428, + 303 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 304, + 416, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 304, + 416, + 316 + ], + "spans": [ + { + "bbox": [ + 321, + 304, + 416, + 316 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 321, + 304, + 416, + 316 + ], + "type": "inline_equation", + "content": "s_2 = \\Phi_j" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 316, + 455, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 316, + 455, + 329 + ], + "spans": [ + { + "bbox": [ + 321, + 316, + 455, + 329 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 321, + 316, + 455, + 329 + ], + "type": "inline_equation", + "content": "\\omega^{t,j} = \\mathcal{F}^{s_1,s_2}\\left(\\omega^{t,j - 1}\\right)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 330, + 395, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 330, + 395, + 339 + ], + "spans": [ + { + "bbox": [ + 318, + 330, + 395, + 339 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 318, + 330, + 395, + 339 + ], + "type": "inline_equation", + "content": "s_1 = s_2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 340, + 416, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 340, + 416, + 350 + ], + "spans": [ + { + "bbox": [ + 318, + 340, + 416, + 350 + ], + "type": "text", + "content": "11: if " + }, + { + "bbox": [ + 318, + 340, + 416, + 350 + ], + "type": "inline_equation", + "content": "t \\neq 1" + }, + { + "bbox": [ + 318, + 340, + 416, + 350 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 351, + 518, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 351, + 518, + 369 + ], + "spans": [ + { + "bbox": [ + 318, + 351, + 518, + 369 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 318, + 351, + 518, + 369 + ], + "type": "inline_equation", + "content": "\\omega^{t,j} = \\omega^{t,j} + \\eta \\mathrm{sign}\\left(\\nabla_{\\omega^{t - 1,j}}\\widetilde{\\mathcal{L}}_{t - 1}\\right)" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 369, + 389, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 369, + 389, + 378 + ], + "spans": [ + { + "bbox": [ + 318, + 369, + 389, + 378 + ], + "type": "text", + "content": "13: end if" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 380, + 385, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 380, + 385, + 389 + ], + "spans": [ + { + "bbox": [ + 318, + 380, + 385, + 389 + ], + "type": "text", + "content": "14: end for" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 318, + 390, + 515, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 390, + 515, + 404 + ], + "spans": [ + { + "bbox": [ + 318, + 390, + 515, + 404 + ], + "type": "text", + "content": "15: " + }, + { + "bbox": [ + 318, + 390, + 515, + 404 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}^i = \\| \\phi (\\mathcal{F}^{\\Phi_u,z}(\\omega^{t,u})) - \\phi (\\mathcal{F}(\\mathbf{x}^t))\\| _2^2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 318, + 404, + 374, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 404, + 374, + 414 + ], + "spans": [ + { + "bbox": [ + 318, + 404, + 374, + 414 + ], + "type": "text", + "content": "16: end for" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 318, + 414, + 553, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 414, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 318, + 414, + 553, + 438 + ], + "type": "text", + "content": "17: " + }, + { + "bbox": [ + 318, + 414, + 553, + 438 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{L}}_t = \\frac{1}{g}\\sum_{i = 1}^g\\widetilde{\\mathcal{L}}^i" + }, + { + "bbox": [ + 318, + 414, + 553, + 438 + ], + "type": "text", + "content": " Compute the loss function utilizing the parameter-augmented models." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 318, + 438, + 507, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 438, + 507, + 460 + ], + "spans": [ + { + "bbox": [ + 318, + 438, + 507, + 460 + ], + "type": "text", + "content": "18: " + }, + { + "bbox": [ + 318, + 438, + 507, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^{adv} = \\prod_{\\mathbf{x}^s,\\epsilon}\\left(\\mathbf{x}_{t - 1}^{adv} - \\beta \\mathrm{sign}\\left(\\nabla_{\\mathbf{x}_{t - 1}^{adv}}\\widetilde{\\mathcal{L}}_t\\right)\\right)" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 318, + 461, + 364, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 461, + 364, + 471 + ], + "spans": [ + { + "bbox": [ + 318, + 461, + 364, + 471 + ], + "type": "text", + "content": "19: end for" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "type": "text", + "content": "CircleLoss, MV-Softmax, and NPCFace are the official models provided by FaceX-ZOO [45]. Furthermore, we integrate adversarial robust FR models into our experiments, denoted as IR152" + }, + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "type": "inline_equation", + "content": "^{adv}" + }, + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "type": "text", + "content": ", IRSE50" + }, + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "type": "inline_equation", + "content": "^{adv}" + }, + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "type": "text", + "content": ", FaceNet" + }, + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "type": "inline_equation", + "content": "^{adv}" + }, + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "type": "text", + "content": ", and MobileFace" + }, + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "type": "inline_equation", + "content": "^{adv}" + }, + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "type": "text", + "content": " which correspond to the models used in [71]." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 556, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 616 + ], + "type": "text", + "content": "Attack Setting. We set the maximum allowable perturbation magnitude " + }, + { + "bbox": [ + 313, + 556, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 556, + 555, + 616 + ], + "type": "text", + "content": " to 10, based on the " + }, + { + "bbox": [ + 313, + 556, + 555, + 616 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 313, + 556, + 555, + 616 + ], + "type": "text", + "content": " norm bound, without any specific emphasis. In addition, we specify the maximum number of iterative steps as 200. More detailed attack settings are provided in the supplementary." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 617, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 700 + ], + "type": "text", + "content": "Evaluation Metrics. We utilize Attack Success Rate (ASR) as the metric to assess the efficacy of various adversarial attacks. The ASR represents the ratio of adversarial examples that successfully evade the victim model to the total number of adversarial examples generated. In calculating the ASR, we determine the threshold based on a FAR@0.001 on the entire LFW dataset for each victim model." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": "Compared methods. Our proposed attack is a form of re" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3521" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "stricted adversarial attack designed to expose vulnerabilities within FR systems. It would be inequitable to juxtapose our method with unrestricted attacks, which do not impose limits on the magnitude of adversarial perturbations. Consequently, we opt to benchmark our approach against other restricted attacks on FR systems that are explicitly malicious at attacking the systems [69] [71], as well as state-of-the-art transfer attacks [59] [27] [12] [52] [46]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 175, + 173, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 173, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 173, + 187 + ], + "type": "text", + "content": "4.2. Comparison Studies" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 192, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 348 + ], + "type": "text", + "content": "DPA achieves the best black-box attack results on both normally trained and adversarial robust models. To verify the effectiveness of our proposed attack method, we craft the adversarial examples on the LFW and CelebA-HQ dataset. The black-box performance with MobileFace as the surrogate models on the LFW dataset are demonstrated in Table 1. Some of the adversarial examples are demonstrated in Figure 4. More experimental results on the LFW dataset, as well as those on the CelebA-HQ dataset, are provided in the supplementary. These results demonstrate that our proposing method consistently outperforms the baseline attack, thereby highlighting its effectiveness in improving the transferability of adversarial face examples." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 357, + 296, + 538 + ], + "blocks": [ + { + "bbox": [ + 56, + 357, + 296, + 538 + ], + "lines": [ + { + "bbox": [ + 56, + 357, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 56, + 357, + 296, + 538 + ], + "type": "image", + "image_path": "de1ea6c5002e68905201547293464742956703050fb5bf07b6a141d0a27e1736.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 542, + 296, + 598 + ], + "lines": [ + { + "bbox": [ + 55, + 542, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 296, + 598 + ], + "type": "text", + "content": "Figure 4. The illustration of adversarial examples crafted by various attacks. First column: some source images. Last column: the corresponding target images. The second to fifth columns exhibit the corresponding adversarial face examples crafted by SIA [52], BPFA [71], BSR [46], and Our proposed attack, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "DPA achieves superior black-box performance under JPEG compression. JPEG compression is a widely adopted method for image compression during transmission, and it also serves as a defense mechanism against adversarial examples. To evaluate the effectiveness of our proposed attack under JPEG compression, we employ MobileFace as the surrogate model and IRSE50 as the victim model. We assess the attack performance on the LFW and CelebA-HQ datasets, with the results presented in Figure 5. These results indicate" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 108, + 552, + 212 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "type": "text", + "content": "Table 1. Comparisons of black-box ASR " + }, + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "type": "text", + "content": " results for attacks using MobileFace as the surrogate model on the LFW dataset. I, S, F, M denote IR152, IRSE50, FaceNet, and MobileFace, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 108, + 552, + 212 + ], + "lines": [ + { + "bbox": [ + 317, + 108, + 552, + 212 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 552, + 212 + ], + "type": "table", + "html": "
AttacksISFIadvSadvFadvMadv
FIM [68]5.373.47.52.54.52.810.9
DI [59]18.497.332.910.618.210.238.6
DFANet [69]7.086.411.93.66.43.916.8
VMI [50]13.696.020.27.612.87.632.3
SSA [27]13.896.419.65.513.17.231.5
SIA [52]15.796.826.78.114.59.135.3
BPFA [71]15.895.517.85.911.05.128.4
BSR [46]5.474.29.52.95.94.911.4
Ours67.798.290.855.466.442.671.6
", + "image_path": "fd445ac5a1daa5bd8dcd450c5f09db095b7b43c16cc25a5e3188aecd35899ecf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 226, + 435, + 322 + ], + "blocks": [ + { + "bbox": [ + 322, + 226, + 435, + 322 + ], + "lines": [ + { + "bbox": [ + 322, + 226, + 435, + 322 + ], + "spans": [ + { + "bbox": [ + 322, + 226, + 435, + 322 + ], + "type": "image", + "image_path": "81d0a75afe082434297617af1f5a3e3084a098981cf16c2e169eb8b3acea8c0b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 329, + 555, + 351 + ], + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 351 + ], + "type": "text", + "content": "Figure 5. Performance of ASR across various JPEG Q values: (a) Results on the LFW dataset. (b) Results on the CelebA-HQ dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 438, + 226, + 547, + 323 + ], + "blocks": [ + { + "bbox": [ + 438, + 226, + 547, + 323 + ], + "lines": [ + { + "bbox": [ + 438, + 226, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 438, + 226, + 547, + 323 + ], + "type": "image", + "image_path": "9f3d27c33027bb07124314b3e6202b6e9ab2a38f2f7cd273e4e5c53a701b1ae2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 324, + 410, + 544, + 450 + ], + "blocks": [ + { + "bbox": [ + 313, + 361, + 555, + 406 + ], + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 406 + ], + "type": "text", + "content": "Table 2. Comparison of black-box ASR (%) results using the parameter-based augmented adversarial attack method as the baseline on the LFW dataset. " + }, + { + "bbox": [ + 313, + 361, + 555, + 406 + ], + "type": "inline_equation", + "content": "\\mathrm{ASR}^{adv}" + }, + { + "bbox": [ + 313, + 361, + 555, + 406 + ], + "type": "text", + "content": " represents the average attack success rate on adversarial robust models." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 410, + 544, + 450 + ], + "lines": [ + { + "bbox": [ + 324, + 410, + 544, + 450 + ], + "spans": [ + { + "bbox": [ + 324, + 410, + 544, + 450 + ], + "type": "table", + "html": "
IR152IRSE50FaceNetASRadv
Baseline4.819.615.77.2
Ours67.798.290.859.0
", + "image_path": "53554cbcb6f1e29c827a0c5c31f4d1f0f8770c6707da1f3838b4a3d10c38ddb4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 473, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 521 + ], + "type": "text", + "content": "that our proposed attack method consistently outperforms the baseline attack methods across varying levels of JPEG compression, thereby underscoring the robustness of our proposed attack method under such conditions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "content": "DPA demonstrates better black-box performance compared to the parameter-based augmented adversarial attack. The LGV method is an effective parameter-based augmented adversarial attack technique designed to enhance transferability. Our proposed attack incorporates parameter augmentation, making LGV an appropriate baseline. We selected MobileFace as the surrogate model and generated adversarial examples on the LFW dataset. The results are presented in Table 2. Table 2 clearly shows that our proposed attack outperforms the baseline, further validating the effectiveness of our method. The primary limitation of LGV stems from its random selection of one model per iteration, which introduces instability and adversely impacts blackbox ASR [9, 50]. Furthermore, LGV suffers from a lack of diverse initialization parameters. These shortcomings significantly impair the overall ASR performance of LGV." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3522" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 72, + 289, + 194 + ], + "blocks": [ + { + "bbox": [ + 65, + 72, + 289, + 194 + ], + "lines": [ + { + "bbox": [ + 65, + 72, + 289, + 194 + ], + "spans": [ + { + "bbox": [ + 65, + 72, + 289, + 194 + ], + "type": "image", + "image_path": "4b4d847a842e07f84ed23c476d9cf9f43c589da7dc0cb22b0cb16c91bd5fc462.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 295, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 224 + ], + "type": "text", + "content": "Figure 6. ASR on victim models trained with various algorithms, using FaceNet as the surrogate model on the LFW dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 242, + 297, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 297, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 297, + 374 + ], + "type": "text", + "content": "DPA achieves superior black-box performance on FR models trained with various algorithms. In order to substantiate the effectiveness of our proposed attack method across various FR models, we conducted additional experiments using BSR [46] as the Baseline and FaceNet as the surrogate model. These experiments adhered to the same settings detailed in Table 7 in the supplementary. The ASR across multiple FR models is illustrated in Figure 6. As shown in Figure 6, the ASR of the adversarial examples generated by our proposed method exceeds that of the Baseline, thereby reinforcing the efficacy of our proposed method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 156, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 156, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 156, + 396 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 402, + 297, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 297, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 297, + 617 + ], + "type": "text", + "content": "The ablation studies on each stage of our proposed attack method. Our proposed attack method consists of two stages. To verify the effectiveness of each stage, we conduct ablation studies on the stages. Initially, we craft adversarial examples using only the surrogate model, which we denote as 'Vanilla'. Next, we generate adversarial examples using the ensemble of surrogate models obtained in the DPO stage (i.e. F), denoted as DPO. After incorporating the HMA stage, the complete attack method is denoted as 'DPO + HMA'. The results are presented in Table 3. Table 3 demonstrates that the addition of the DPO stage results in an increase in black-box ASR, showcasing the effectiveness of the DPO stage in enhancing transferability. Further incorporation of the HMA stage leads to an additional improvement in attack performance, underscoring the effectiveness of the HMA stage in boosting black-box performance. These results collectively demonstrate the effectiveness of each stage in our proposed attack method in improving the transferability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": "The ablation studies on the effectiveness of multiple surrogate models at different epochs in improving transferability. Since our proposed attack method utilizes models from intermediate training epochs to craft adversarial examples, it is essential to verify the effectiveness of this approach compared to using only models from the final training epoch. We conduct ablation experiments on these two approaches using the LFW dataset and MobileFace as the surrogate" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 108, + 555, + 160 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": "Table 3. Ablation study of black-box ASR (\\%) results on the LFW dataset using MobileFace as the surrogate model. " + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "inline_equation", + "content": "\\mathrm{ASR}^{adv}" + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": " denotes the average attack success rate on adversarial robust models." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 108, + 555, + 160 + ], + "lines": [ + { + "bbox": [ + 315, + 108, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 315, + 108, + 555, + 160 + ], + "type": "table", + "html": "
AttacksIR152IRSE50FaceNetASRadv
Vanilla5.373.47.55.2
DPO35.991.367.932.9
DPO + HMA67.798.290.859.0
", + "image_path": "6d9cc361770d106b3458acb5ec3b96a7f937f67e17cc421d53f5b06243ca7ce5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 324, + 219, + 545, + 259 + ], + "blocks": [ + { + "bbox": [ + 313, + 170, + 555, + 214 + ], + "lines": [ + { + "bbox": [ + 313, + 170, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 555, + 214 + ], + "type": "text", + "content": "Table 4. Ablation study of black-box ASR (\\%) results to verify the effectiveness of multiple surrogate models at different epochs in improving transferability. " + }, + { + "bbox": [ + 313, + 170, + 555, + 214 + ], + "type": "inline_equation", + "content": "\\mathrm{ASR}^{adv}" + }, + { + "bbox": [ + 313, + 170, + 555, + 214 + ], + "type": "text", + "content": " denotes the average attack success rate on adversarial robust models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 219, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 324, + 219, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 324, + 219, + 545, + 259 + ], + "type": "table", + "html": "
AttacksIR152IRSE50FaceNetASRadv
Fm20.386.733.617.8
F (Ours)35.991.367.932.9
", + "image_path": "c6b9838490e15f14baa3dbea71f526b79d6197b440a75a527dc890342d3cdcd4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "text", + "content": "model. We employ the DPO process to obtain the parameter sets " + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_c^q" + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_c^m = \\{\\mathbf{v}_0^p, \\mathbf{v}_c^p, \\mathbf{v}_c^a\\}" + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "text", + "content": ". Next, we map " + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_c^m" + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "text", + "content": " into its corresponding model set " + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^m = \\Psi(\\mathbf{V}_c^m)" + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "text", + "content": ". We then craft adversarial examples by ensembling the models from " + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_c^q" + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "text", + "content": " (i.e. " + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^m" + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "text", + "content": ", respectively. The results are presented in Table 4. Table 4 shows that the performance of the ensemble of " + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "text", + "content": " surpasses that of the ensemble of " + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^m" + }, + { + "bbox": [ + 313, + 278, + 555, + 386 + ], + "type": "text", + "content": ", demonstrating the effectiveness of using models from intermediate training epochs to enhance the transferability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 386, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 410 + ], + "type": "text", + "content": "The hyper-parameter analysis studies of our proposed method are in the supplementary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 421, + 388, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 421, + 388, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 388, + 433 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 441, + 556, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 556, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 556, + 609 + ], + "type": "text", + "content": "We present an innovative advancement in the field of adversarial attacks on FR through the introduction of the Diverse Parameters Augmentation (DPA) attack method. By addressing the problems of traditional adversarial attacks, particularly the lack of diverse parameters and the exclusion of the FR head, DPA enhances the transferability of adversarial face examples. In addition, DPA emphasizes the importance of integrating both pre-trained and randomly initialized parameters, providing a fresh perspective on augmentation-based adversarial attacks. The extensive experimental results demonstrate the effectiveness of our proposed DPA attack across multiple datasets and target models. We anticipate that our proposed method will illuminate potential directions for advancing research in adversarial attacks on FR." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 616, + 419, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 419, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 419, + 630 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 635, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 556, + 715 + ], + "type": "text", + "content": "This work was supported in part by the Natural Science Foundation of China under Grant 62372203 and 62302186, in part by the Major Scientific and Technological Project of Shenzhen (202316021), in part by the National key research and development program of China (2022YFB2601802), in part by the Major Scientific and Technological Project of Hubei Province (2022BAA046, 2022BAA042)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3523" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 96, + 115, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 115, + 108 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 115, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 117, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 117, + 297, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 117, + 297, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 117, + 297, + 182 + ], + "type": "text", + "content": "[1] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, and Ying Fu. Partial FC: training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 1445-1449, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "type": "text", + "content": "[2] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19593-19604, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 240, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 240, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 296, + 282 + ], + "type": "text", + "content": "[3] Tom B Brown, Nicholas Carlini, Chiyuan Zhang, Catherine Olsson, Paul Christiano, and Ian Goodfellow. Unrestricted adversarial examples. arXiv preprint arXiv:1809.08352, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 285, + 296, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 296, + 341 + ], + "type": "text", + "content": "[4] Bin Chen, Jia-Li Yin, Shukai Chen, Bohao Chen, and Ximeng Liu. An adaptive model ensemble adversarial attack for boosting adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4466-4475, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 342, + 296, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 296, + 385 + ], + "type": "text", + "content": "[5] Zhaoyu Chen, Bo Li, Shuang Wu, Kaixun Jiang, Shouhong Ding, and Wenqiang Zhang. Content-based unrestricted adversarial attack. In Advances in Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "type": "text", + "content": "[6] Valeriia Cherepanova, Micah Goldblum, Harrison Foley, Shiyuan Duan, John P Dickerson, Gavin Taylor, and Tom Goldstein. Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 445, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 509 + ], + "type": "text", + "content": "[7] Sowmen Das, Selim Seferbekov, Arup Datta, Md Saiful Islam, and Md Ruhul Amin. Towards solving the deepfake problem: An analysis on improving deepfake detection using dynamic face augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3776-3785, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 511, + 296, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 296, + 565 + ], + "type": "text", + "content": "[8] Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5962-5979, 2022. 1, 2, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 567, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 621 + ], + "type": "text", + "content": "[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 2, 3, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "type": "text", + "content": "[10] Zhijin Ge, Fanhua Shang, Hongying Liu, Yuanyuan Liu, Liang Wan, Wei Feng, and Xiaosen Wang. Improving the transferability of adversarial examples with arbitrary style transfer. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4440-4449, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[11] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "text", + "content": "[12] Martin Gubri, Maxime Cordy, Mike Papadakis, Yves Le Traon, and Koushik Sen. Lgv: Boosting adversarial example transferability from large geometric vicinity. In European Conference on Computer Vision, pages 603-618. Springer, 2022. 2, 4, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 130, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 130, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 130, + 555, + 173 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 175, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 555, + 217 + ], + "type": "text", + "content": "[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 220, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 555, + 285 + ], + "type": "text", + "content": "[15] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14994–15003, 2022. 1, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 287, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 555, + 341 + ], + "type": "text", + "content": "[16] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, 2007. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 343, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 555, + 398 + ], + "type": "text", + "content": "[17] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: Adaptive curriculum learning loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "type": "text", + "content": "[18] Shuai Jia, Bangjie Yin, Taiping Yao, Shouhong Ding, Chunhua Shen, Xiaokang Yang, and Chao Ma. Adv-attribute: Inconspicuous and transferable adversarial attack on face recognition. In Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 456, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 555, + 498 + ], + "type": "text", + "content": "[19] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representation, 2018. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 501, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 555, + 555 + ], + "type": "text", + "content": "[20] Jingzhi Li, Zidong Guo, Hui Li, Seungju Han, Ji-Won Baek, Min Yang, Ran Yang, and Sungjoo Suh. Rethinking feature-based knowledge distillation for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20156-20165, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 557, + 555, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 555, + 610 + ], + "type": "text", + "content": "[21] Yanjie Li, Yiquan Li, Xuelong Dai, Songtao Guo, and Bin Xiao. Physical-world optical adversarial attacks on 3d face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24699-24708, 2023. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 613, + 555, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 555, + 677 + ], + "type": "text", + "content": "[22] Zexin Li, Bangjie Yin, Taiping Yao, Junfeng Guo, Shouhong Ding, Simin Chen, and Cong Liu. Sibling-attack: Rethinking transferable adversarial attacks against face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24626–24637, 2023. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "text", + "content": "[23] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3524" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "text", + "content": "from diffusion models via adversarial examples. In International Conference on Machine Learning, pages 20763-20786, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "type": "text", + "content": "[24] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representation, 2020. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "text", + "content": "[25] Xuannan Liu, Yaoyao Zhong, Yuhang Zhang, Lixiong Qin, and Weihong Deng. Enhancing generalization of universal adversarial perturbation through gradient aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4412-4421, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "text", + "content": "[26] Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, and Di Ming. TRM-UAP: enhancing the transferability of data-free universal adversarial perturbation via truncated ratio maximization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4739-4748, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "text", + "content": "[27] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xi-anglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In European Conference on Computer Vision, pages 549–566, 2022. 2, 7, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 296, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 385 + ], + "type": "text", + "content": "[28] Dong Lu, Zhiqiang Wang, Teng Wang, Weili Guan, Hongchang Gao, and Feng Zheng. Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 102-111, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 442 + ], + "type": "text", + "content": "[29] Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14225-14234, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 296, + 498 + ], + "type": "text", + "content": "[30] Duan Mingxing, Kenli Li, Lingxi Xie, Qi Tian, and Bin Xiao. Towards multiple black-boxes attack via adversarial example generation network. In Proceedings of the 29th ACM International Conference on Multimedia, pages 264-272, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 296, + 554 + ], + "type": "text", + "content": "[31] Kartik Narayan, Harsh Agarwal, Kartik Thakral, Surbhi Mittal, Mayank Vatsa, and Richa Singh. Df-platter: Multiface heterogeneous deepfake dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9739-9748, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 555, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 296, + 600 + ], + "type": "text", + "content": "[32] Xuelin Qian, Wenxuan Wang, Yu-Gang Jiang, Xiangyang Xue, and Yanwei Fu. Dynamic routing and knowledge relearning for data-free black-box attack. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 296, + 644 + ], + "type": "text", + "content": "[33] Haonan Qiu, Chaowei Xiao, Lei Yang, Xinchen Yan, Honglak Lee, and Bo Li. Semanticadv: Generating adversarial examples via attribute-conditioned image editing. In European Conference on Computer Vision, pages 19-37, 2020. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "text", + "content": "[34] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 815-823, 2015. 1, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[35] Fahad Shamshad, Muzammal Naseer, and Karthik Nandakumar. Clip2protect: Protecting facial privacy using text-guided" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "makeup via adversarial latent search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20595-20605, 2023. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 151 + ], + "type": "text", + "content": "[36] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 217 + ], + "type": "text", + "content": "[37] Florian Stimberg, Ayan Chakrabarti, Chun-Ta Lu, Hussein Hazimeh, Otilia Stretcu, Wei Qiao, Yintao Liu, Merve Kaya, Cyrus Rashtchian, Ariel Fuxman, Mehmet Tek, and Sven Gowal. Benchmarking robustness to adversarial image obfuscations. In Advances in Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 274 + ], + "type": "text", + "content": "[38] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "text", + "content": "[39] Yuhao Sun, Lingyun Yu, Hongtao Xie, Jiaming Li, and Yongdong Zhang. Diffam: Diffusion-based adversarial makeup transfer for facial privacy protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24584-24594, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 555, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 407 + ], + "type": "text", + "content": "[40] Naufal Suryanto, Yongsu Kim, Harashta Tatimma Larasati, Hyoeun Kang, Thi-Thu-Huong Le, Yoonyoung Hong, Hunmin Yang, Se-Yoon Oh, and Howon Kim. ACTIVE: towards highly transferable 3d physical camouflage for universal and robust vehicle evasion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4282-4291, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 410, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 555, + 453 + ], + "type": "text", + "content": "[41] Teppei Suzuki. Teachaugment: Data augmentation optimization using teacher knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10894-10904, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 455, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 555, + 498 + ], + "type": "text", + "content": "[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 500, + 555, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 555, + 553 + ], + "type": "text", + "content": "[43] Donghua Wang, Wen Yao, Tingsong Jiang, Chao Li, and Xiaogian Chen. RFLA: A stealthy reflected light adversarial attack in the physical world. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4442, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 556, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 555, + 609 + ], + "type": "text", + "content": "[44] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 612, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 555, + 634 + ], + "type": "text", + "content": "[45] Jun Wang, Yinglu Liu, Yibo Hu, Hailin Shi, and Tao Mei. Facex-zoo: A python toolbox for face recognition. 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "type": "text", + "content": "[46] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting Adversarial Transferability by Block Shuffle and Rotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 7, 8, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[47] Mei Wang, Weihong Deng, Jiani Hu, Xunqiang Tao, and Yaohai Huang. Racial faces in the wild: Reducing racial" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3525" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "bias by information maximization adaptation network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 171 + ], + "type": "text", + "content": "[48] Wenxuan Wang, Bangjie Yin, Taiping Yao, Li Zhang, Yanwei Fu, Shouhong Ding, Jilin Li, Feiyue Huang, and Xiangyang Xue. Delving into data: Effectively substitute training for black-box attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4761-4770, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 296, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 226 + ], + "type": "text", + "content": "[49] Wenxuan Wang, Xuelin Qian, Yanwei Fu, and Xiangyang Xue. Dst: Dynamic substitute training for data-free black-box attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14361-14370, 2022. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 295, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 272 + ], + "type": "text", + "content": "[50] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1924-1933, 2021. 2, 4, 7, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "type": "text", + "content": "[51] Xiaobo Wang, Shifeng Zhang, Shuo Wang, Tianyu Fu, Hailin Shi, and Tao Mei. Mis-classified vector guided softmax loss for face recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12241-12248, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "type": "text", + "content": "[52] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4607-4619, 2023. 2, 7, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "type": "text", + "content": "[53] Xudong Wang, Ke-Yue Zhang, Taiping Yao, Qianyu Zhou, Shouhong Ding, Pingyang Dai, and Rongrong Ji. Tf-fas: Twofold-element fine-grained semantic guidance for generalizable face anti-spoofing. In European Conference on Computer Vision, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 296, + 459 + ], + "type": "text", + "content": "[54] Xingxing Wei, Yao Huang, Yitong Sun, and Jie Yu. Unified adversarial patch for cross-modal attacks in the physical world. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4422-4431, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 459, + 296, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 296, + 504 + ], + "type": "text", + "content": "[55] Xingxing Wei, Jie Yu, and Yao Huang. Physically adversarial infrared patches with learnable shapes and locations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12334-12342, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 296, + 536 + ], + "type": "text", + "content": "[56] Shixian Wen and Laurent Itti. Beneficial perturbations network for defending adversarial examples. arXiv preprint arXiv:2009.12724, 2020. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "text", + "content": "[57] Han Wu, Guanyan Ou, Weibin Wu, and Zibin Zheng. Improving transferable targeted adversarial attacks with model self-enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24615-24624, 2024. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 657 + ], + "type": "text", + "content": "[58] Zihao Xiao, Xianfeng Gao, Chilin Fu, Yinpeng Dong, Wei Gao, Xiaolu Zhang, Jun Zhou, and Jun Zhu. Improving transferability of adversarial patches on face recognition with generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11840-11849, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "text", + "content": "[59] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019. 2, 7, 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "text", + "content": "[60] Zhuoer Xu, Zhangxuan Gu, Jianping Zhang, Shiwen Cui, Changhua Meng, and Weiqiang Wang. Backpropagation path search on adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4640-4650, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "text", + "content": "[61] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3877-3887, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 183, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 555, + 239 + ], + "type": "text", + "content": "[62] Xiao Yang, Chang Liu, Longlong Xu, Yikai Wang, Yinpeng Dong, Ning Chen, Hang Su, and Jun Zhu. Towards effective adversarial textured 3d meshes on physical face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4119-4128, 2023. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 555, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 555, + 303 + ], + "type": "text", + "content": "[63] Bangjie Yin, Wenxuan Wang, Taiping Yao, Junfeng Guo, Zelun Kong, Shouhong Ding, Jilin Li, and Cong Liu. Advmakeup: A new imperceptible and transferable attack on face recognition. In Proceedings of the 30th International Joint Conference on Artificial Intelligence, pages 1252-1258, 2021. 1, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 304, + 555, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 555, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 555, + 349 + ], + "type": "text", + "content": "[64] Shengming Yuan, Qilong Zhang, Lianli Gao, Yaya Cheng, and Jingkuan Song. Natural color fool: Towards boosting black-box unrestricted attacks. In Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 349, + 555, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 555, + 382 + ], + "type": "text", + "content": "[65] Dan Zeng, Hailin Shi, Hang Du, Jun Wang, Zhen Lei, and Tao Mei. Npcface: Negative-positive collaborative training for large-scale face recognition. 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 555, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 555, + 426 + ], + "type": "text", + "content": "[66] Jiaming Zhang, Qi Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5005-5013, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 426, + 555, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 555, + 459 + ], + "type": "text", + "content": "[67] Xinyu Zhang, Qiang Wang, Jian Zhang, and Zhao Zhong. Adversarial autoaugment. In International Conference on Learning Representations, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 459, + 555, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 459, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 555, + 504 + ], + "type": "text", + "content": "[68] Yaoyao Zhong and Weihong Deng. Adversarial learning with margin-based triplet embedding regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6548-6557, 2019. 7, 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 504, + 555, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 555, + 547 + ], + "type": "text", + "content": "[69] Yaoyao Zhong and Weihong Deng. Towards transferable adversarial attack against deep face recognition. IEEE Transactions on Information Forensics and Security, 16:1452-1466, 2021. 1, 2, 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 555, + 613 + ], + "type": "text", + "content": "[70] Fengfan Zhou, Hefei Ling, Yuxuan Shi, Jiazhong Chen, and Ping Li. Improving visual quality and transferability of adversarial attacks on face recognition simultaneously with adversarial restoration. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4540-4544, 2024. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 669 + ], + "type": "text", + "content": "[71] Fengfan Zhou, Hefei Ling, Yuxuan Shi, Jiazhong Chen, Zongyi Li, and Ping Li. Improving the transferability of adversarial attacks on face recognition with beneficial perturbation feature augmentation. IEEE Transactions on Computational Social Systems, 11(6):8130-8142, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[72] Fengfan Zhou, Qianyu Zhou, Xiangtai Li, Xuequan Lu, Lizhuang Ma, and Hefei Ling. Adversarial attacks on both face recognition and face anti-spoofing models. arXiv preprint arXiv:2405.16940, 2024. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3526" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 451 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 298, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 298, + 127 + ], + "type": "text", + "content": "[73] Fengfan Zhou, Qianyu Zhou, Bangjie Yin, Hui Zheng, Xuequan Lu, Lizhuang Ma, and Hefei Ling. Rethinking impersonation and dodging attacks on face recognition systems. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2487-2496, 2024. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 184 + ], + "type": "text", + "content": "[74] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Shouhong Ding, and Lizhuang Ma. Adaptive mixture of experts learning for generalizable face anti-spoofing. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6009-6018, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "type": "text", + "content": "[75] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Kekai Sheng, Shouhong Ding, and Lizhuang Ma. Generative domain adaptation for face anti-spoofing. In European Conference on Computer Vision, pages 335-356, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 284 + ], + "type": "text", + "content": "[76] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Xuequan Lu, Ran Yi, Shouhong Ding, and Lizhuang Ma. Instance-aware domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20453-20463, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 296, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 296, + 341 + ], + "type": "text", + "content": "[77] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Xuequan Lu, Shouhong Ding, and Lizhuang Ma. Test-time domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 175-187, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 296, + 396 + ], + "type": "text", + "content": "[78] Ziqi Zhou, Shengshan Hu, Minghui Li, Hangtao Zhang, Yechao Zhang, and Hai Jin. Advclip: Downstream-agnostic adversarial examples in multimodal contrastive learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6311–6320, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 397, + 296, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 397, + 296, + 451 + ], + "spans": [ + { + "bbox": [ + 56, + 397, + 296, + 451 + ], + "type": "text", + "content": "[79] Hegui Zhu, Yuchen Ren, Xiaoyan Sui, Lianping Yang, and Wuming Jiang. Boosting adversarial transferability via gradient relevance attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4718-4727, 2023. 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3527" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/380dd383-b00d-4370-9caa-3f9395391322_content_list.json b/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/380dd383-b00d-4370-9caa-3f9395391322_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..994538f372605acc15ab7d80971152c4a400a1d7 --- /dev/null +++ b/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/380dd383-b00d-4370-9caa-3f9395391322_content_list.json @@ -0,0 +1,2174 @@ +[ + { + "type": "text", + "text": "Imputation-free and Alignment-free: Incomplete Multi-view Clustering Driven by Consensus Semantic Learning", + "text_level": 1, + "bbox": [ + 86, + 128, + 880, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuzhuo Dai $^{1}$ , Jiaqi Jin $^{1}$ , Zhibin Dong $^{1}$ , Siwei Wang $^{2,*}$ , Xinwang Liu $^{1,*}$ , En Zhu $^{1,*}$ , Xihong Yang $^{1}$ , Xinbiao Gan $^{1}$ , Yu Feng $^{1}$", + "bbox": [ + 165, + 202, + 803, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National University of Defense Technology, Changsha, China", + "bbox": [ + 238, + 239, + 730, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Intelligent Game and Decision Lab, Beijing, China", + "bbox": [ + 279, + 256, + 689, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yzdai24, wangsiwei13}@nudt.edu.cn", + "bbox": [ + 331, + 277, + 630, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 312, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In incomplete multi-view clustering (IMVC), missing data induce prototype shifts within views and semantic inconsistencies across views. A feasible solution is to explore cross-view consistency in paired complete observations, further imputing and aligning the similarity relationships inherently shared across views. Nevertheless, existing methods are constrained by two-tiered limitations: (1) Neither instance- nor cluster-level consistency learning construct a semantic space shared across views to learn consensus semantics. The former enforces cross-view instances alignment, and wrongly regards unpaired observations with semantic consistency as negative pairs; the latter focuses on cross-view cluster counterparts while coarsely handling fine-grained intra-cluster relationships within views. (2) Excessive reliance on consistency results in unreliable imputation and alignment without incorporating view-specific cluster information. Thus, we propose an IMVC framework, imputation- and alignment-free for consensus semantics learning (FreeCSL). To bridge semantic gaps across all observations, we learn consensus prototypes from available data to discover a shared space, where semantically similar observations are pulled closer for consensus semantics learning. To capture semantic relationships within specific views, we design a heuristic graph clustering based on modularity to recover cluster structure with intra-cluster compactness and inter-cluster separation for cluster semantics enhancement. Extensive experiments demonstrate, compared to state-of-the-art competitors, FreeCSL achieves more confident and robust assignments on IMVC task.", + "bbox": [ + 76, + 357, + 472, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 823, + 209, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thanks to representation learning enhanced by data observed from different perspectives, multi-view clustering", + "bbox": [ + 76, + 848, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bea63629f3bb2620e981da766fc86802231dca321c1d7f9cda640ac0e6fe1dc4.jpg", + "image_caption": [ + "(a) Consistency Learning Paradigms", + "Figure 1. Research Motivation for Consensus Semantic Learning. (a) two existing paradigms, instance- and cluster-level, either treat cross-view unpaired observations with similar semantics as false negatives or neglect to treat within-view observations with similar semantics as true positives; (b) we propose a novel semantic-level paradigm based on contrastive clustering with a set of consensus prototypes to foster semantic consistency across all view data." + ], + "image_footnote": [], + "bbox": [ + 503, + 325, + 671, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c8ea22c0cc0aadeabdc2d019f4a15670337047bdc505b9147defac3acb1cf5b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 325, + 700, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0ce666c53306ccfcf037916107d8bdd9aefda76c0681a3391b7f5b6fd854c981.jpg", + "image_caption": [ + "(b) Prototype-based Consensus Learning" + ], + "image_footnote": [], + "bbox": [ + 705, + 325, + 890, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(MVC) has achieved significant breakthroughs in the field of unsupervised learning [3, 6, 11, 17, 21, 34, 46, 57]. However, in practical applications, the assumption of data completeness is often difficult to satisfy that incomplete multiview clustering (IMVC) is introduced [7, 25, 29, 33, 35, 37, 54]. In IMVC, missing data causes prototypes shifts within views and semantic misalignment across views, due to the discrepancy between the distributions of complete and incomplete instances [14, 19]. More and more studies [4, 16] have noted that variations in complete instances across different views further exacerbate prototype misalignment. It is challenging to achieve semantic consistency on cluster assignments across all view data.", + "bbox": [ + 496, + 579, + 892, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To alleviate prototype shifts and misalignment while promoting semantic consensus in cluster assignments, existing IMVC methods explore consistency information from complete instances for imputation and alignment [20, 24, 32]. Unfortunately, they still suffer from several significant drawbacks in practical applications. In terms of consistency learning, as shown in Fig. 1 (a), one widely used paradigm is instance-level [13, 36], which pulls paired ob", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5071", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "observations (the same instance in different views) closer in the representation space by enforcing highly similar representations, but may inadvertently discard view-specific information [26, 31, 48]. More importantly, it tends to introduce false-negative noise, where unpaired observations with the same semantics are incorrectly treated as negative pairs [2, 8, 27]. [49],[52] and [27] attempted to optimize this issue by proposing a cluster-level paradigm [47] that encourages observations to find their cluster counterparts across different views. It learns a cluster space shared across views but not applicable within specific views, as there are no clustering interactions among intra-view observations.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the above consistency learning paradigms fail to account for semantic consistency across all view data, they cannot mitigate semantic gaps among intra-cluster observations [22, 30, 56]. Therefore, imputation of missing data is required to restore the original data distribution, including neighborhood-based recovery via cross-view graph structure transfer [39], adversarial generation or contrastive prediction through cross-view mutual information interaction [15, 37, 42, 53], and prototype-based imputation via cross-view sample-prototype relationship inheritance [19]. Meanwhile, cross-view alignment of assignments [22, 32], prototypes [16, 19, 38], or distributions [4, 45] is also a crucial approach for further enhancing consistent learning. More related works are enumerated in Appendix A. Both imputation and alignment are limited by the consistent information from cross-view paired data and cannot fully exploit within-view unpaired data to mine view-specific cluster information, i.e., within-cluster compactness and between-cluster separation [9, 19]. Particularly, once the amount of missing data is too excessive to provide sufficient consistent information, model performance may even decline rapidly due to the noise introduced by improper imputation and alignment.", + "bbox": [ + 75, + 276, + 472, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Realizing the above issues, we ponder: imputation and alignment aim to restore the similarity relationships inherited from other complete views for the clustering task. Can we directly bridge the semantic gaps while exploring the semantic relationships among all data in consistency learning, thereby avoiding imputation and alignment operations with uncertainty noise? Thus, we propose an IMVC framework, a novel semantic-level paradigm, as shown in Fig. 1 (b), that is driven by imputation- and alignment-free consensus semantic learning (FreeCSL). Notably, our proposed consensus learning involves concurrent interaction among all data, rather than being limited to within or across views. To bridge semantic gaps among view observations and learn cluster semantics information, FreeCSL employs contrastive clustering based on consensus prototypes to discover a shared semantic space, where observations converge toward their semantic prototypes respectively. In practice, we set missing statistical weights for observations to facilitate view collaboration in consensus representations that in", + "bbox": [ + 75, + 613, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tegrate detailed information of all views while adapting the impact of different views with different missing instances. Based on consensus representation, our model can construct robust consensus prototypes for semantic-level clustering without imputation and alignment. To discover view-specific semantic relationships and further enrich consensus representations, FreeCSL exploit graph clustering to capture the cluster structure for each view, which maximizes graph modularity within views to enhance intra-cluster connections and reduce inter-cluster interactions. In short, our model encodes data correlation, discovered by semantic learning, into a shared space to obtain consensus semantic representations for instance clustering. Our prominent contributions can be summarized as follows:", + "bbox": [ + 496, + 90, + 893, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- In terms of bridging semantic gaps, we design the consensus semantic learning module, a novel semantic-level paradigm based prototypical contrastive clustering, to discover a common semantic space where all observations are embedded as representations with consistent semantics, avoiding additional imputation and alignment.", + "- In terms of exploring semantic relationships, we employ the cluster semantics enhancement module, a heuristic graph clustering method with modularity-based learning objective, to mine the inherent cluster structures that reveal the semantic correlations within views.", + "- Extensive experiments show our model surpasses state-of-the-art (SOTA) competitors in complex tasks with high missing rates, multiple clusters and large-scale data." + ], + "bbox": [ + 500, + 303, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Method", + "text_level": 1, + "bbox": [ + 500, + 526, + 589, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, FreeCSL, a deep IMVC method without imputation or alignment, is proposed to learn consensus semantic representations for clustering. The framework in Fig. 2 coordinates reconstruction (REC) module, cross-view consensus semantic learning (CSL) module, and within-view cluster semantic enhancement (CSE) module.", + "bbox": [ + 496, + 551, + 890, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Problem Statement", + "text_level": 1, + "bbox": [ + 500, + 648, + 684, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notations. Given a multi-view dataset $\\mathcal{X} = \\{\\mathbf{X}^v\\in$ $\\mathbb{R}^{N\\times D_v}\\}_{v = 1}^V$ with $N$ instances across $V$ views, $\\tilde{\\mathbf{X}}^v =$ $\\{\\tilde{\\mathbf{x}}_i^v\\in \\mathbb{R}^{D_v}\\}_{i = 1}^{N_v}$ is an incomplete subset of the $v$ -th view with $N_{v}$ observations, and $\\overline{\\mathbf{X}}^{m,n} = \\{(\\overline{\\mathbf{x}}_i^m,\\overline{\\mathbf{x}}_i^n)\\}_{i = 1}^{N_{mn}}$ is a pair-observed subset with $N_{mn}$ instances observed in both the $m$ -th and $n$ -th view. The task is to partition $N$ instances into $K$ clusters.", + "bbox": [ + 496, + 672, + 890, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 1. Instance-level Consistency (IC): $\\forall m \\neq n$ , $\\mathbf{x}_i^m$ and $\\mathbf{x}_j^n$ are instance-level consistent across views if $i = j$ (they are cross-view observations of the same instance $\\mathbf{x}$ ), expressed as $I(\\mathbf{x}_i^m, \\mathbf{x}_j^n) = 1$ and $O$ otherwise.", + "bbox": [ + 496, + 779, + 890, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 2. Cluster-level Consistency (CC): $\\forall m \\neq n$ , $\\mathbf{x}_i^m$ and $\\mathbf{x}_j^n$ are cluster-level consistent across views if they belong to the same cluster $k$ , expressed as $C(\\mathbf{x}_i^m, \\mathbf{x}_j^n) = 1$ and 0 otherwise.", + "bbox": [ + 496, + 840, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5072", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 3. Semantic-level Consensus (SC): $\\forall m$ and $n$ , $\\mathbf{x}_i^m$ and $\\mathbf{x}_j^n$ achieve semantic-level consensus in MVC task if all observations share a set of cluster prototypes $\\mathbf{C} = \\{\\mathbf{c}_k\\}_{k=1}^K$ and $\\arg \\max_k \\rho(\\mathbf{x}_i^m, \\mathbf{c}_k) = \\arg \\max_k \\rho(\\mathbf{x}_j^n, \\mathbf{c}_k)$ , expressed as $S(\\mathbf{x}_i^m, \\mathbf{x}_j^n) = 1$ and $0$ otherwise.", + "bbox": [ + 76, + 90, + 468, + 172 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1. Consensus semantic learning yields more confident and robust cluster assignments than instance- and cluster-level paradigms. (Proof is provided in Appendix B.)", + "bbox": [ + 73, + 179, + 470, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 2. Since paired observations $(\\overline{\\mathbf{x}}_i^m, \\overline{\\mathbf{x}}_i^n)$ inherently satisfy instance- and cluster-level consistency, they can achieve semantic consensus via a shared set of prototypes C. (Proof is provided in Appendix B.)", + "bbox": [ + 75, + 229, + 468, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Within-view Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 306, + 328, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To avoid clustering instability from similarity measures on manifold structures in high-dimensional spaces, we use autoencoders to encode view observations $\\tilde{\\mathbf{X}}^v$ into clustering-friendly low-dimensional representations $\\tilde{\\mathbf{Z}}^v$ . Considering multiple views are mostly heterogeneous and differently distributed, we provide an independent encoder-decoder $\\{\\mathcal{E}_v,\\mathcal{D}_v\\}_{v = 1}^V$ for each view. The encoder embeds the latent representation and the decoder recovers the original data from it, which jointly minimize reconstruction loss:", + "bbox": [ + 76, + 330, + 468, + 465 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {r e c} = \\sum_ {v = 1} ^ {V} \\left\\| \\tilde {\\mathbf {X}} ^ {v} - \\mathcal {D} _ {v} \\left(\\tilde {\\mathbf {Z}} ^ {v}\\right) \\right\\| _ {F} ^ {2} \\tag {1} \\\\ = \\sum_ {v = 1} ^ {V} \\sum_ {i = 1} ^ {N _ {v}} \\| \\tilde {\\mathbf {x}} _ {i} ^ {v} - \\mathcal {D} _ {v} \\left(\\mathcal {E} _ {v} \\left(\\tilde {\\mathbf {x}} _ {i} ^ {v}\\right)\\right) \\| _ {2} ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 470, + 468, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{E}_v(\\tilde{\\mathbf{X}}^v;\\theta_v):\\tilde{\\mathbf{X}}^v\\in \\mathbb{R}^{N_v\\times D_v}\\to \\tilde{\\mathbf{Z}}^v\\in \\mathbb{R}^{N_v\\times d}$ and $\\mathcal{D}_v(\\tilde{\\mathbf{Z}}^v;\\phi_v):\\tilde{\\mathbf{Z}}^v\\in \\mathbb{R}^{N_v\\times d}\\to \\hat{\\mathbf{X}}^v\\in \\mathbb{R}^{N_v\\times D_v}$ .", + "bbox": [ + 76, + 563, + 468, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Cross-view Consensus Semantic Learning", + "text_level": 1, + "bbox": [ + 76, + 604, + 433, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on Theorem 1, we design prototypical contrastive clustering for consensus semantic learning. To reach semantic-level consensus on assignments, consensus prototypes with all view information, are introduced into contrastive clustering to explore a shared semantic space, where all observations with similar semantics are pulled closer.", + "bbox": [ + 75, + 628, + 468, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consensus Representations and Consensus Prototypes. Due to the assumptions of consistency and complementarity in MVC, inseparable clusters in one view will become linearly separable by introducing complementary information from other views [10, 43, 55]. We utilize a fusion manner, denoted as $\\mathbb{T}(\\cdot)$ , to map the latent representations $\\mathbf{Z}^v$ into a linearly weighted representation space for consensus representations $\\mathbf{Z} \\in \\mathbb{R}^{N \\times d}$ that integrates consistency and complementary information from multiple views:", + "bbox": [ + 75, + 719, + 468, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} = \\mathbb {T} \\left(\\left\\{\\mathbf {Z} ^ {v} \\right\\} _ {v = 1} ^ {V}\\right) = \\sum_ {v = 1} ^ {V} \\mathbf {w} ^ {v} \\mathbf {Z} ^ {v} = \\left\\{\\sum_ {v = 1} ^ {V} w _ {i} ^ {v} \\mathbf {z} _ {i} ^ {v} \\right\\} _ {i = 1} ^ {N}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 862, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w_{i}^{v}$ is the instance-level fine-grained fusion weight:", + "bbox": [ + 500, + 90, + 880, + 107 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} ^ {v} = \\frac {\\mathbb {I} \\left(\\mathbf {z} _ {i} ^ {v} \\neq N a N\\right)}{\\sum_ {v ^ {\\prime}} ^ {V} \\mathbb {I} \\left(\\mathbf {z} _ {i} ^ {v ^ {\\prime}} \\neq N a N\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 117, + 890, + 155 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{I}(\\cdot)$ is the indicative function that takes 1 when $\\mathbf{z}_i^v$ is the representation of $i$ -th instance $\\mathbf{X}_i$ observed in view $v$ , and 0 otherwise. We set a completeness statistical weight $w_i^v$ based on the number of observations in $V$ views for instance $\\mathbf{X}_i$ as the fusion weight, which makes use of view-specific complementary information without discarding unpaired observations and mitigating the negative impact of missing noise by adapting to the differences in missing instances across different views.", + "bbox": [ + 498, + 165, + 890, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We derive a set of consensus semantic prototypes $\\mathbf{C} = \\{\\mathbf{c}_k\\in \\mathbb{R}^d\\}_{k = 1}^K$ via $k$ -means on consensus representations $\\mathbf{Z}$ . As \"a representative embedding for semantically similar observations\", $\\mathbf{C}$ comprehensively captures the semantic information of all data [50] and is continuously refined throughout consensus representation learning.", + "bbox": [ + 498, + 301, + 890, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prototype-based Contrastive Clustering. We perform semantic-level contrastive clustering in a shared vector space spanned by consensus prototypes, where the latent representation $\\mathbf{z}_i^v$ encoded as the semantic representation $\\mathbf{h}_i^v$ and $\\mathbf{h}_i^v$ is assigned to the prototype $\\mathbf{c}_k$ with the longest projected distance to obtain optimal assignments, in two steps:", + "bbox": [ + 498, + 393, + 890, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Semantic similarity measure: Encode the semantic representation $\\mathbf{h}_i^v$ and project it onto each prototype, then calculate the probability $p_{i,k}^{v}$ of belonging to cluster $k$ :", + "bbox": [ + 500, + 484, + 890, + 531 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {i, k} ^ {v} = \\frac {\\exp \\left(\\mathbf {h} _ {i} ^ {v \\top} \\mathbf {c} _ {k} / \\tau\\right)}{\\sum_ {k ^ {\\prime}} ^ {K} \\exp \\left(\\mathbf {h} _ {i} ^ {v \\top} \\mathbf {c} _ {k ^ {\\prime}} / \\tau\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 542, + 890, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is a temperature parameter [40]. $\\mathbf{p}_i^v = \\{p_{i,k}^v\\}_{k = 1}^K$ is the soft assignments of observation $\\mathbf{x}_i^v$", + "bbox": [ + 511, + 594, + 888, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Swapped knowledge distillation: Based on Theorem 2, $(\\overline{\\mathbf{x}}_i^m,\\overline{\\mathbf{x}}_i^n)$ share the same cluster semantics on consensus prototypes $\\mathbf{C}$ , thus their true labels $\\mathbf{y}_i^m,\\mathbf{y}_i^n\\in \\mathbb{R}^K$ should be are distributionally consistent. To this end, \"swapped\" knowledge distillation (KD) utilizes pseudo-labels $\\mathbf{q}_i^m$ , $\\mathbf{q}_i^n$ as mutual supervised signals to prompt their cluster assignments $\\mathbf{p}_i^m,\\mathbf{p}_i^n$ as identical as possible:", + "bbox": [ + 500, + 626, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {c c} ^ {m, n} = \\ell_ {k d} \\left(\\mathbf {H} ^ {m}, \\mathbf {Q} ^ {n}\\right) + \\ell_ {k d} \\left(\\mathbf {H} ^ {n}, \\mathbf {Q} ^ {m}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 742, + 890, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\ell_{kd}(\\mathbf{H}^m,\\mathbf{Q}^n) = -\\frac{1}{N_{nm}}\\sum_{i = 1}^{N_{mn}}\\sum_{k = 1}^{K}\\mathbf{q}_i^n\\log \\mathbf{p}_i^m$ We extend Eq. (5) to more than two views, fostering greater view collaboration to enhance semantic learning:", + "bbox": [ + 511, + 771, + 890, + 842 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c c} = \\sum_ {m = 1} ^ {V} \\sum_ {\\substack {n = 1 \\\\ n \\neq m}} ^ {V} \\ell_ {c c} ^ {m, n}. \\tag{6}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 852, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5073", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/982238dac916e28d963632dc3904412fd56538aa13f450a290c06bafe53add61.jpg", + "image_caption": [ + "Figure 2. The framework of FreeCSL. (a) Reconstruction module, encodes observations into clustering-friendly representations for each view; (b) Consensus semantic learning module, learns semantic representations through cross-view contrastive clustering in a shared semantic space, where paired observations are assigned to their nearest semantic prototype for consistent assignments. (c) Cluster semantic enhancement module, enriches semantic representations with cluster structure information through within-view graph clustering, which applies GCN to aggregate view-specific semantic information and maximizes spectral modularity to recover cluster structure with greater separation and lower entropy. Ultimately, perform $k$ -means on the consensus semantic representations to predict cluster labels." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 867, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Consensus label solution: To obtain pseudo-labels $\\mathbf{Q}^m = \\{\\mathbf{q}_i^m\\}_{i=1}^{N_{mn}}$ , project semantic representations $\\mathbf{H}^m$ onto consensus prototypes $\\mathbf{C}$ and maximize similarity between the representation $\\mathbf{h}_i^m$ and its assigned prototype $\\mathbf{c}_k$ :", + "bbox": [ + 76, + 527, + 470, + 589 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} ^ {*} = \\arg \\max _ {\\mathbf {Q} ^ {m} \\in \\mathcal {Q}} \\operatorname {T r} \\left\\{\\left(\\mathbf {C Q} ^ {m}\\right) ^ {\\top} \\mathbf {H} ^ {m}\\right) \\} + \\alpha \\mathcal {R} (\\mathbf {Q} ^ {m}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 592, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{R}(\\mathbf{Q}^m) = -\\sum_i^{N_{mn}}\\mathbf{q}_i^m\\log \\mathbf{q}_i^m$ , an entropy regularizer prevents trivial solutions via the smoothness $\\alpha$", + "bbox": [ + 89, + 625, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4. Within-view Cluster Semantic Enhancement", + "text_level": 1, + "bbox": [ + 76, + 662, + 455, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering that graph pooling can discover data implicit correlation and reduce missing noise via aggregating the information of neighboring nodes, we design a graph clustering to enhance representations with cluster semantic information. It combines a heuristic learning objective based on modularity, an evaluation metric for cluster structure quality, to effectively recover the cluster structure[51].", + "bbox": [ + 75, + 686, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Modularity. Modularity [28] quantifies the deviation between the actual cluster structure and the expected structure generated by random combinations in the null model. Modularity matrix $\\mathbf{B}$ is defined as $\\mathbf{B} = \\mathbf{A} - \\frac{\\mathbf{d}\\mathbf{d}^{\\top}}{2m}$ , where $\\mathbf{A}$ is the adjacency matrix, $\\mathbf{d}$ is the degree vector and $m$ is the total number of edges in the graph. The entry $b_{ij} \\in \\mathbf{B}$ measures connection strength between nodes $i$ and $j$ .", + "bbox": [ + 75, + 792, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Modularity-based Graph clustering. Graph clustering relies on graph pooling and node assignment. For each view, we construct graph structures with $k$ -nearest neighbors (KNN), then deploy graph convolutional network (GCN) $\\mathcal{G}_v(\\cdot)$ to aggregate graph embeddings:", + "bbox": [ + 496, + 527, + 892, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {v} \\left(\\mathbf {Z} ^ {v}; \\mathbf {A} ^ {v}\\right) = \\sigma \\left(\\mathbf {D} ^ {v - \\frac {1}{2}} \\mathbf {A} ^ {v} \\mathbf {D} ^ {v - \\frac {1}{2}} \\mathbf {Z} ^ {v} \\mathbf {W} ^ {v} + \\mathbf {Z} ^ {v} \\mathbf {W} _ {s} ^ {v}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 609, + 890, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ is an activation function used to reinforce the nonlinear aggregation capability of GCN. We replace a skip connection matrix $\\mathbf{W}_s^v$ with the self-loop $\\mathbf{A}^v +\\mathbf{I}$ to alleviate the over-smoothing of node embeddings. Graph pooling operates on the latent representations $\\mathbf{Z}^v$ , while the adjacency matrix $\\mathbf{A}^v$ is constructed from the original observations $\\mathbf{X}^v$ :", + "bbox": [ + 496, + 637, + 893, + 728 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\na _ {i, j} ^ {v} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\left(\\mathbf {x} _ {i} ^ {v} \\text {a n d} \\mathbf {x} _ {j} ^ {v} \\neq N a N\\right) \\& \\\\ & \\left(\\mathbf {x} _ {i} ^ {v} \\in \\mathcal {N} _ {k} \\left(\\mathbf {x} _ {j} ^ {v}\\right) \\text {o r} \\mathbf {x} _ {j} ^ {v} \\in \\mathcal {N} _ {k} \\left(\\mathbf {x} _ {i} ^ {v}\\right)\\right), \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 736, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{N}_k(\\mathbf{x}_i^v)$ is a set including the $k$ -nearest neighbors of $\\mathbf{x}_i^v$ . It brings neighbors closer in the embedding space without introducing missing noise into the graph structure.", + "bbox": [ + 496, + 797, + 892, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The softmax classifier offers the soft assignment $\\mathbf{P}^v$ for node embeddings encoded by $\\mathcal{G}_v(\\mathbf{Z}^v;\\mathbf{A}^v)$ :", + "bbox": [ + 498, + 844, + 890, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} ^ {v} = \\operatorname {S o f t m a x} \\left(\\mathcal {G} _ {v} \\left(\\mathbf {Z} ^ {v}; \\mathbf {A} ^ {v}\\right)\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 882, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5074", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain a well-separated cluster assignment $\\mathbf{P}^v$ , we maximize the spectral modularity $\\mathrm{Tr}\\left((\\mathbf{P}^v)^\\top \\mathbf{B}^v\\mathbf{P}^v\\right)$ [1, 18], which captures intra-cluster connections and intercluster margins by mapping $\\mathbf{P}^v$ to modularity $\\mathbf{B}^v$ ; To further improve the confidence and robustness of $\\mathbf{P}^v$ , we apply a self-supervised signal $\\mathbf{L}^v$ , learned from contrastive clustering as detailed in Sec. 2.3, to guide $\\mathbf{P}^v$ in chasing $\\mathbf{L}^v$ via self-knowledge distillation. Specifically, the following objective is designed to optimize $\\mathbf{P}^v$ :", + "bbox": [ + 76, + 90, + 472, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {m} ^ {v} \\left(\\mathbf {P} ^ {v}; \\mathbf {L} ^ {v}\\right) = - \\frac {1}{2 m} \\operatorname {T r} \\left(\\left(\\mathbf {P} ^ {v}\\right) ^ {\\top} \\mathbf {B} ^ {v} \\mathbf {P} ^ {v}\\right) + \\lambda \\operatorname {K L} \\left(\\mathbf {L} ^ {v} \\| \\mathbf {P} ^ {v}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 237, + 468, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where Kullback-Leibler divergence $\\mathrm{KL}(\\cdot)$ serves as a robust regularizer, leveraging $\\lambda$ to regulate the information flow of $\\mathbf{L}^v\\in \\mathbb{R}^{N\\times K}$ to ensure stable cluster performance. We project semantic representations $\\mathbf{H}^v$ learned through contrastive clustering onto their $k$ -means prototypes and introduce Student's t-distribution [41] as kernel to predict high-confidence labels $\\mathbf{L}^v$ with a nearly uniform distribution:", + "bbox": [ + 75, + 279, + 472, + 385 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {i, k} ^ {v} = \\frac {\\left(1 + \\frac {\\| \\mathbf {h} _ {i} ^ {v} - \\mathbf {c} _ {k} ^ {v} \\| ^ {2}}{\\gamma}\\right) ^ {- \\frac {\\gamma + 1}{2}}}{\\sum_ {k ^ {\\prime}} \\left(1 + \\frac {\\| \\mathbf {h} _ {i} ^ {v} - \\mathbf {c} _ {k ^ {\\prime}} ^ {v} \\| ^ {2}}{\\gamma}\\right) ^ {- \\frac {\\gamma + 1}{2}}} \\in \\mathbf {L} ^ {v}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 396, + 468, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct graph clustering on each view separately, then optimize assignments across all views concurrently:", + "bbox": [ + 76, + 468, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g c} = \\sum_ {v = 1} ^ {V} \\ell_ {m} ^ {v} \\left(\\mathbf {P} ^ {v}, \\mathbf {L} ^ {v}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 510, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.5. Clustering Driven by Consensus Semantics", + "text_level": 1, + "bbox": [ + 76, + 561, + 442, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our FreeCSL learns the semantic representations $\\mathbf{H}^v$ end-to-end by jointly minimizing the reconstruction loss, contrastive clustering loss and graph clustering loss:", + "bbox": [ + 76, + 585, + 468, + 632 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {r e c} + \\mathcal {L} _ {c c} + \\mathcal {L} _ {g c}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 643, + 468, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Without searching for the optimal balancing weights of three loss terms, outstanding performance is easily achieved by applying $k$ -means to the consensus semantic representations $\\mathbf{H}$ learned via the fusion manner $\\mathbb{T}\\left(\\{\\mathbf{H}^v\\}_{v=1}^V\\right)$ .", + "bbox": [ + 76, + 671, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The implementation for FreeCSL consists of two stages: warm-up training for the encoder-decoder, followed by fine-tuning for semantic representation learning and clustering. For details, refer to Algorithm 1.", + "bbox": [ + 76, + 733, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Experiment", + "text_level": 1, + "bbox": [ + 76, + 806, + 202, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 830, + 284, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We select four representative datasets namely Caltech-5V[44], ALOI-100[5], YoutubeFace10[12] and NoisyMNIST[23] to demonstrate our model's performance", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 FreeCSL for Learning and Clustering", + "bbox": [ + 500, + 90, + 839, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Complete, incomplete and pair-observed multiview dataset $\\{\\mathbf{X}^v\\}_{v = 1}^V$ $\\{\\tilde{\\mathbf{X}}^v\\}_{v = 1}^V$ $\\{\\overline{\\mathbf{X}}^{m,n}\\}_{m\\neq n}^V$ networks $\\{\\mathcal{E}_v,\\mathcal{D}_v,\\mathcal{G}_v\\}_{v = 1}^V$ ; warm-up and fine-tuning epochs $e,E$", + "bbox": [ + 500, + 108, + 890, + 155 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for $t = 1$ to e do", + "bbox": [ + 500, + 155, + 619, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On $\\{\\tilde{\\mathbf{X}}^v\\}_{v = 1}^V$ , warming up $\\{\\mathcal{E}_v,\\mathcal{D}_v\\}_{v = 1}^V$ with Eq.(1).", + "bbox": [ + 500, + 167, + 872, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for $t = 1$ to $E$ do", + "bbox": [ + 500, + 189, + 620, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Obtain latent representations $\\{\\mathbf{Z}^v |\\mathcal{E}_v:\\mathbf{X}^v\\to \\mathbf{Z}^v\\}_{v = 1}^V$ Construct consensus prototypes $\\mathbf{C} = \\{\\mathbf{c}_k\\}_{k = 1}^K$ as elaborated in Sec. 2.3.", + "bbox": [ + 524, + 203, + 888, + 248 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compute the reconstruction loss $\\mathcal{L}_{rec}$ with Eq.(1).", + "bbox": [ + 524, + 250, + 856, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "// Cross-view Contrastive Clustering", + "bbox": [ + 524, + 265, + 880, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for $m,n = 1$ to $V(m\\neq n)$ do", + "bbox": [ + 524, + 281, + 733, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compute the cluster probability $\\mathbf{p}_i^m,\\mathbf{p}_i^n$ with Eq.(4);", + "bbox": [ + 550, + 295, + 890, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Solve consensus pseudo-label $\\mathbf{q}_i^m$ with Eq.(7);", + "bbox": [ + 558, + 311, + 866, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Get swapped loss $\\ell_{cc}^{m,n}$ with Eq.(5) on $\\overline{\\mathbf{X}}^{m,n}$ .", + "bbox": [ + 558, + 325, + 857, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Calculate contrastive clustering loss $\\mathcal{L}_{cc}$ with Eq.(6)", + "bbox": [ + 513, + 345, + 869, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "// Within-view Graph Clustering", + "bbox": [ + 535, + 362, + 841, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for $v = 1$ to $V$ do", + "bbox": [ + 535, + 376, + 656, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Construct adjacency matrix $\\mathbf{A}^v$ with Eq.(9).", + "bbox": [ + 558, + 390, + 849, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Solve node assignment $\\mathbf{P}^v$ with Eq.(10)", + "bbox": [ + 558, + 406, + 823, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Solve pseudo-labels $\\mathbf{L}^v$ with Eq.(12);", + "bbox": [ + 558, + 421, + 807, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compute KL-modularity loss $\\ell_m^v$ with Eq.(11)", + "bbox": [ + 558, + 436, + 864, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Calculate graph clustering loss $\\mathcal{L}_{gc}$ with Eq.(13)", + "bbox": [ + 524, + 454, + 844, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "// Semantic Representation Learning", + "bbox": [ + 534, + 470, + 879, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Calculate the overall loss $\\mathcal{L}$ with Eq.(14);", + "bbox": [ + 534, + 484, + 808, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimize $\\{\\mathcal{E}_v,\\mathcal{D}_v,\\mathcal{G}_v\\}_{v = 1}^V$ to minimize $\\mathcal{L}$", + "bbox": [ + 534, + 500, + 813, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learn latent and consensus representations $\\{\\mathbf{Z}^v\\}_{v = 1}^V$", + "bbox": [ + 534, + 513, + 890, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $\\mathbf{Z}$ , semantic representations $\\{\\mathbf{H}^v\\}_{v=1}^V$ , consensus", + "bbox": [ + 534, + 529, + 888, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "semantic representations $\\mathbf{H}$ , consensus prototypes $\\mathbf{C}$ .", + "bbox": [ + 534, + 545, + 885, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: $\\hat{\\mathbf{Y}} = \\{\\hat{y}_i\\}_{i=1}^N$ predicted by $k$ -means on $\\mathbf{H}$ .", + "bbox": [ + 500, + 561, + 841, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in comparison with seven SOTA methods summarized in Table 1. To comprehensively evaluate experimental results, three metrics are adopted: accuracy (ACC), normalized mutual information (NMI), and adjusted rand index (ARI).", + "bbox": [ + 496, + 607, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/665ea7542b9b044bdf8b7df407e386328f225e2af6f11a64c5171df005bafdcd.jpg", + "table_caption": [ + "Table 1. SOTA methods categorized by the types of techniques for consistency, imputation, and alignment." + ], + "table_footnote": [], + "table_body": "
CompetitorsConsistencyImputationAlignment
CPM-Nets (TPAMI'20)instance-levelmutual information interaction
COMPLETER (CVPR'21)instance-levelmutual information interaction
DIMVC (AAAI'22)instance-level\\assignment-based
SURE (TPAMI'23)cluster-levelgraph structure transfer
ProLmp (IICA'23)instance-levelsample-prototype relationship inheritanceprototype-based
ICMVC (AAAI'24)instance-levelgraph structure transferassignment-based
DIVIDE (AAAI'24)cluster-levelmutual information interaction
", + "bbox": [ + 503, + 719, + 901, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 816, + 712, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models are trained on an NVIDIA 3090 GPU using Adam optimizer in PyTorch 2.1.0. The warming up and finetuning learning rates are 0.0003 and 0.0005, with a batch size of 512. The hyperparameters are set as follows for", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "aside_text", + "text": "1 \n2 \n3 \n4", + "bbox": [ + 486, + 296, + 500, + 467 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5075", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "different datasets: smoothness $\\alpha = 0.5$ , temperature $\\tau \\in \\{0.1, 0.2\\}$ , neighbors $\\zeta = 3$ , and regularizer weight $\\lambda \\in \\{0.05, 0.1, 0.2, 0.3\\}$ . For all datasets, 4-layer autoencoders and 2-layer GCNs with same MLP structures are used for each view. The layers of contrastive clustering and graph clustering are shared across all views via a single FC layer.", + "bbox": [ + 76, + 90, + 472, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Competitiveness of FreeCSL", + "text_level": 1, + "bbox": [ + 76, + 191, + 333, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FreeCSL is compared with seven SOTA methods across three metrics in Table 2. The results indicate that FreeCSL excellently handles challenges of high missing rates, multiple clusters, and large-scale issues in IMVC:", + "bbox": [ + 75, + 215, + 468, + 275 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- From the perspective of effectiveness, FreeBSD surpasses most SOTA models, particularly on challenging datasets like ALOI-100 (100 clusters) and YouTubeFace10 (30,000 samples). It achieves significant improvements in ACC, with gains of $15.12\\%$ , $26.38\\%$ , $23.91\\%$ , $26.21\\%$ for ALOI-100, and $1.16\\%$ , $8.81\\%$ , $6.32\\%$ , $1.68\\%$ for YouTubeFace10, at missing rate $r = 0.1$ , $0.3$ , $0.5$ , $0.7$ .", + "- From the perspective of robustness, Fig. 3 (visualized from Table 2) shows our model's stability as $r$ increases. FreeCSL declines gradually, unlike other models with sharp ACC drops, highlighting its robustness in IMVC task. On the small dataset Caltech-5V with $r = 0.7$ , its Acc remains at $84.56\\%$ , while on the larger dataset NoisyMNIST, also achieves $92.19\\%$ ." + ], + "bbox": [ + 76, + 277, + 470, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The above phenomenon can be explained as follows: As $r$ increases, paired observations become scarce. Methods like instance-level consistency learning or cross-view imputation and alignment (e.g., CPM-Net, COMPLETER, ICMVC), constrained by their dependence on consistency information, will introduce biases such as false negatives and semantic inconsistencies, resulting in degraded performance. For the complex task ALOI-100, cluster-level methods like SURE and DIVIDE surpass instance-level by exploiting cross-view cluster consistency. But, their reliance on cross-view graphs or mutual information for imputation, while neglecting within-view cluster affiliations, restricts performance. ProImp acknowledges the need to integrate within-view and cross-view information for semantically consistent imputation. However, its prototype alignment accumulates errors, particularly with a large number of clusters, achieving only $20\\%$ ACC on ALOI-100.", + "bbox": [ + 75, + 488, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The superiority of FreeCSL stems from the synergistic effect of the CSL and CSE modules, which eliminate semantic gaps and ensure consistent assignments for all observations in a shared semantic space.", + "bbox": [ + 75, + 744, + 468, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Understanding FreeCSL", + "text_level": 1, + "bbox": [ + 76, + 816, + 303, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation Study. As shown in Table 3, ablation studies on FreeCSL's three components reveal that the CSL module contributes the most. On ALOI-100, it demonstrates exceptional performance and strong stability, with ACC drop", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44a05a75f84b556842986767080c87927e6efd088f3c383155cc3e63a819a247.jpg", + "image_caption": [ + "(a) Caltech-5V" + ], + "image_footnote": [], + "bbox": [ + 511, + 88, + 692, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/85c9df279d182798f03f30987165d0a467e7d1e18294a69fd7a931481cba0a26.jpg", + "image_caption": [ + "(b) ALOI-100" + ], + "image_footnote": [], + "bbox": [ + 702, + 89, + 877, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1b84ce01963f7f0afc1aa5b276652d554cd1c87c0f719c30d7cb83dd5f3c2dc7.jpg", + "image_caption": [ + "(c) YoutubeFace10" + ], + "image_footnote": [], + "bbox": [ + 511, + 207, + 687, + 308 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5601e511d580dbfd6206497a847821e276a337df2824b369ad4151970a346a98.jpg", + "image_caption": [ + "(d) NoisyMNIST", + "Figure 3. Visualization for Table 2 based on metric ACC." + ], + "image_footnote": [], + "bbox": [ + 700, + 208, + 877, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ping only $12.4\\%$ as $r$ rises from 0.1 to 0.7. This is attributed to CSL's establishment of a shared semantic space via prototype-based semantic contrast, reducing semantic gaps within clusters. The CSE module also plays a positive role in discovering tighter cluster structures by optimizing modularity. Thanks to its enhanced semantics, both the REC and CSL modules achieve notable improvements.", + "bbox": [ + 496, + 357, + 890, + 463 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aa3562998a24fb9c685041adb1552e65e604fac63fc3adb17144c7ffe5597d73.jpg", + "image_caption": [ + "(a) $\\mathbf{Z}^{(1)}$" + ], + "image_footnote": [], + "bbox": [ + 501, + 474, + 581, + 536 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1190bef9176bc25e375f1a402aa26b8cd0f07c1dce21c97684398f3d64088821.jpg", + "image_caption": [ + "(b) $\\mathbf{Z}^{(2)}$" + ], + "image_footnote": [], + "bbox": [ + 584, + 474, + 661, + 536 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be049ac7d7bc6a93db89c7810cc3d9e8e27367c31c872d04f860331de60a9e1d.jpg", + "image_caption": [ + "(c) $\\mathbf{Z}^{(3)}$" + ], + "image_footnote": [], + "bbox": [ + 665, + 474, + 745, + 536 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a86e48e7bf1e44e8176ce1bdd11e0d5747ed4113a8f31bdecaaceb3d3d270140.jpg", + "image_caption": [ + "(d) $\\mathbf{Z}^{(4)}$" + ], + "image_footnote": [], + "bbox": [ + 750, + 474, + 826, + 536 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da8924045196a8f8977ee07e95801095fe486f5a6a6c74bf5f90b34fe48dc386.jpg", + "image_caption": [ + "(e) Z" + ], + "image_footnote": [], + "bbox": [ + 828, + 474, + 908, + 537 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3a549c7483c4570036923ec7730acc478c8dd1c092cf33f18f2a8d076fb27e8e.jpg", + "image_caption": [ + "Figure 4. Similarity matrices of $\\{\\mathbf{Z}^v\\}_{v=1}^4$ , $\\mathbf{Z}$ without consensus semantic learning on ALOI-100 with $r = 0.5$ .", + "(a) $\\mathbf{H}^{(1)}$", + "Figure 5. Similarity matrices of $\\{\\mathbf{H}^v\\}_{v=1}^4$ , $\\mathbf{H}$ with consensus semantic learning on ALOI-100 with $r = 0.5$ ." + ], + "image_footnote": [], + "bbox": [ + 501, + 606, + 580, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ba7702b5c0b82ff079233308ebc3fb4d6a656d385f2cec563eae387b3752eef4.jpg", + "image_caption": [ + "(b) $\\mathbf{H}^{(2)}$" + ], + "image_footnote": [], + "bbox": [ + 584, + 606, + 661, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be1186ed0f33521b45d1520fe5a77984b6479338f23eda584f699ce81140c2fb.jpg", + "image_caption": [ + "(c) $\\mathbf{H}^{(3)}$" + ], + "image_footnote": [], + "bbox": [ + 665, + 606, + 743, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf7a82ff40da53536f810e53225dc1e4c9d88a560faeec609abf1c63cdfd95b4.jpg", + "image_caption": [ + "(d) $\\mathbf{H}^{(4)}$" + ], + "image_footnote": [], + "bbox": [ + 750, + 606, + 826, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ccfa1adbbe01951b9dc76e113e1d6d5ec3aa3901859807a138ad0a9d6e337f2c.jpg", + "image_caption": [ + "(e) H" + ], + "image_footnote": [], + "bbox": [ + 828, + 606, + 908, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Imputation- and Alignment-free CSL. To verify FreeCSL can reduce semantic gaps and capture semantic relationships without requiring imputation or alignment, we set up two imputation experiments: one imputation for latent representations $\\mathbf{Z}^v$ without consensus semantic learning named ILR, and the second for semantic representations $\\mathbf{H}^v$ learned from consensus semantic learning named ISR. They transfer complete $k$ -nn graph structures from other views to incomplete views and utilize corresponding $k$ neighbors for imputation. The control groups perform $k$ -means on the sum of all view matrices to predict cluster", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5076", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fe63b0e6f33c0188f20ebd865948b55ddc8ee551cad69ff87b63be1c56ab8ebb.jpg", + "table_caption": [ + "Table 2. Performance comparison on four multi-view benchmarks. The best and second-best results are highlighted in red and blue." + ], + "table_footnote": [], + "table_body": "
Missing ratesr=0.1r=0.3r=0.5r=0.7
MetricsACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)
Caltech-5VCPM-Nets85.7274.6271.9285.4372.9070.8481.0868.6166.0879.5065.5262.70
COMPLETETER58.6867.6942.9865.0768.7547.1064.4666.0942.8069.0067.4651.49
DIMVC85.8680.0174.6378.9078.7669.5580.0170.0866.3569.8660.1648.70
SURE75.6469.1262.8377.8967.8961.1577.2265.9960.9973.9660.8953.66
ProImp66.6257.6647.5472.1761.7052.4965.2855.7746.2260.3649.2339.63
ICMVC76.8673.9267.3776.9373.3966.3079.2972.6766.2172.2166.6358.18
DIVIDE57.2947.0033.2368.1458.9344.1681.5771.4559.2074.3668.6260.14
Ours91.5785.3283.1488.8681.2678.6888.3680.0178.0283.6471.7968.06
ALOI-100CPM-Nets63.9579.1451.2652.3070.7439.7942.7462.4628.4030.5254.0317.18
COMPLETETER21.8646.9412.1022.0446.4511.6219.3844.8110.7817.1843.689.60
DIMVC56.9075.0035.5950.5572.5128.1052.5471.4532.5648.9770.8026.16
SURE64.5977.7452.6160.2274.7647.2160.6574.8447.3950.2368.9137.48
ProImp26.1552.9431.4024.5253.0728.1827.9758.1033.4220.3350.1024.28
ICMVC55.4874.0442.9148.8367.4636.5035.4256.9824.1926.8447.4316.63
DIVIDE76.0188.5970.5961.9080.5353.2952.5673.1740.5742.6667.8230.93
Ours91.1395.2287.5988.2893.3483.9584.5690.5378.9076.4485.0067.68
YouTubeFace10CPM-Nets75.4878.5666.5067.5075.8261.1566.6075.1962.0167.3274.1062.54
COMPLETETER46.7252.0421.0350.2452.8127.7736.4135.2415.1840.9839.9220.82
DIMVC81.7781.3274.5971.9676.4262.2973.8775.4463.9774.9479.3468.96
SURE26.6523.6011.4126.4322.4010.5531.8428.7715.2729.3727.4812.85
ProImp48.9753.2033.8343.3049.0426.5339.6750.1826.6849.5251.4729.12
ICMVC68.9074.1760.0662.7665.4453.0344.5652.5132.5127.4427.0915.33
DIVIDE31.8529.6411.3535.8630.8112.3132.7430.8212.7435.1733.4715.93
Ours82.9383.5574.7680.7781.4671.6280.1981.0771.3776.6281.3173.22
NoisyMNISTCPM-Nets48.6444.0832.0042.8837.0126.9250.5243.5433.5646.4637.1827.04
COMPLETETER76.5587.9577.1878.1686.1176.8364.7965.7754.6841.1540.6027.58
DIMVC55.8854.7841.2956.7758.2943.5162.2862.6648.0165.9467.4354.58
SURE98.6195.8996.9697.1193.7591.6494.8589.0687.8984.0078.2576.02
ProImp98.3095.3196.3593.2986.0585.3878.1669.4564.3750.0441.3333.34
ICMVC98.7896.3697.3597.7593.7195.1181.6479.2475.4753.9150.9144.56
DIVIDE94.7291.4489.4395.8589.7191.0657.0957.2946.2828.5725.6111.14
Ours99.1397.2398.1097.6893.9494.9496.0489.8191.4892.1982.5083.56
", + "bbox": [ + 80, + 109, + 893, + 469 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b8dba52c9e56866444244d9b6d87505e359cba356da2b20ffddd55438aa29903.jpg", + "table_caption": [ + "Table 3. Ablation study on Caltech-5V and ALOI-100. $\\checkmark$ denotes FreeCSL with the component and the best results are highlighted in red." + ], + "table_footnote": [], + "table_body": "
Componentsr=0.1r=0.3r=0.5r=0.7
LrecLccLgcACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)
Caltech-5V✓75.0065.6458.8969.0057.8451.4952.2144.1836.7151.9343.4235.75
✓✓85.5078.0772.2882.0074.5367.6785.3674.3271.1079.1467.8162.78
✓✓81.0770.4566.1380.8669.6167.4974.1459.3955.9070.7153.2549.95
✓✓✓91.5785.3283.1488.8681.2678.6888.3680.0178.0283.6471.7968.06
ALOI-100✓63.8177.9249.3844.6965.6431.1132.2055.8318.9423.7848.9011.31
✓✓87.6993.6484.1285.2591.0279.4580.2388.0474.1575.2983.3865.72
✓✓81.0685.9671.2164.1375.8751.4947.5265.5934.2932.7356.3519.23
✓✓✓91.1395.2287.5988.2893.3483.9584.5690.5378.9076.4485.0067.68
", + "bbox": [ + 80, + 502, + 890, + 626 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ffea6045aaa997e612660d27dac77827397ab7f2c8ae96856a1ac3b2655d8b32.jpg", + "table_caption": [ + "Table 4. Imputation- and alignment-free study on Caltech-5V and ALOI-100. ILR and ISR are filled with K-NN imputation via cross-view graph for latent representations and semantic representations $\\mathbf{Z}^{(v)}$ , $\\mathbf{H}^{(v)}$ . The best results are highlighted in red." + ], + "table_footnote": [], + "table_body": "
Missing ratesr=0.1r=0.3r=0.5r=0.7
MetricsACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)
Caltech-5VILR91.7185.7683.4489.1481.6979.0788.7980.7978.3777.8665.6455.12
ISR91.6485.5583.3089.4382.2679.6288.8680.7878.7481.2966.6863.59
FreeCSL91.5785.3283.1488.8681.2678.6888.3680.0178.0283.6471.7968.06
ALGI-100ILR58.9279.6741.5453.4675.8836.8945.8369.4727.4138.0858.9621.35
ISR90.1695.3887.8688.9194.0885.4882.5388.7371.5465.3677.8546.06
FreeCSL91.1395.2287.5988.2893.3483.9584.5690.5378.9076.4485.0067.68
", + "bbox": [ + 84, + 672, + 887, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "labels. Table 4 shows FreeBSD's robustness advantage at high $r$ on Caltech-5V, with notable performance disparity on ALOI-100. ILR, lacking consensus semantic learning, struggles with multi-cluster tasks due to cross-view semantic gaps. ISR avoids semantic confusion via consensus semantic learning but underperforms FreeBSD due to biased", + "bbox": [ + 75, + 803, + 470, + 893 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "imputation as consistency information decreases at higher $r$ . FreeCSL, by contrast, integrates consistency and complementary information in consensus semantic representations for confident decisions.", + "bbox": [ + 498, + 803, + 890, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further illustrate how consensus semantic learning \"free up\" imputation and alignment, we construct cosine", + "bbox": [ + 500, + 869, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5077", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "similarity matrices on ALOI-100 for view-specific latent representations $\\{\\mathbf{Z}^v\\}_{v = 1}^4$ and their consensus $\\mathbf{Z}$ , as well as view-specific semantic representations $\\{\\mathbf{H}^v\\}_{v = 1}^4$ and their consensus $\\mathbf{H}$ in Fig. 5, then analyze their information entropy. The similarity matrices of $\\{\\mathbf{Z}^v\\}_{v = 1}^4$ are chaotic, with high uncertainty in intra- and inter- cluster relationships. The fusion manner $\\mathbb{T}(\\cdot)$ with Eq. (2) alleviates high entropy by incorporating view-specific complementary information. Compared to Fig. 4, the similarity matrices of $\\{\\mathbf{H}^v\\}_{v = 1}^4$ shows a clear block structure, with high intracluster similarity and distinct inter-cluster differences. This confirms $\\mathbf{H}$ , enhanced by view-specific information, reduces cross-view semantic gaps, while consensus semantic learning, capturing cluster semantic relationship, promotes high-confidence assignments as stated in Theorem 1.", + "bbox": [ + 76, + 90, + 472, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5. Analysis on FreeCSL", + "text_level": 1, + "bbox": [ + 76, + 328, + 274, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Convergence and Robustness Analysis. In Fig. 6, we plot metrics and losses over training iterations, with error bands from 5 random Caltech-5V experiments to assess robustness. Both converge to stable values with minimal fluctuations and reach stability simultaneously. The gradual decrease in reconstruction loss indicates that the CSL and CSE modules are reasonable and beneficial, as they don't introduce significant discrepancies between semantic and latent representations, further underscoring the benefits of pre-training. Thanks to the harmonious collaboration of the three modules, our FreeCSL achieved satisfactory results within 30 iterations at minimal computational cost.", + "bbox": [ + 75, + 352, + 468, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Parameter Sensitivity Analysis. Two hyperparameters in FreeCSL warrant investigation, namely, graph neighbors $\\zeta$ in Eq. (9) and regularizer coefficient $\\lambda$ in Eq. (11). We expanded $\\zeta$ and $\\lambda$ to the range of 0.05 to 0.5 and 3 to 32 respectively. In Fig. 7, our model is insensitive to $\\zeta$ but more affected by variations in $\\lambda$ . This is because the robust regularizer leverages semantic knowledge learned from the CSL module to guide the CSE model toward faster and more stable learning. A small $\\lambda$ cannot provide effective consistency constraints while a large one makes CSL to dominate the learn direction, hindering CSE from exploring cluster structure information. Therefore, we set $\\lambda \\in [0.05, 0.2]$ to balance cooperative relationship between CSE and CSL. Without sacrificing ACC and NMI, we prefer to set the minimum $\\zeta = 3$ to reduce computational complexity.", + "bbox": [ + 75, + 534, + 468, + 761 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.6. Visualization of Consensus Semantic Clusters", + "text_level": 1, + "bbox": [ + 76, + 770, + 464, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further examine the quality of cluster structures formed by consensus semantic representations, t-SNE visualization based on true labels are plotted in Fig. 8 and clearly show minimal misclassification and clusters with strong intracluster cohesion while distinct inter-cluster separation. This indicates FreeCSL nearly recovers true cluster structures by learning consensus semantic information from all data.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02329fcf383e903b6352efbb8b9fd3ea543fc078ad344ffb05e1ba3122356704.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 89, + 874, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2062e21a39ba687681e800e9a1b0c9e0812579a95be44574e4bea859edbec620.jpg", + "image_caption": [ + "Figure 6. Convergence analysis on Caltech-5V with $r = 0.5$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 157, + 874, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c77d24e6d6d40a3792fe17dc3a5c26fc5d74651030a5090cf99044f1c516b22.jpg", + "image_caption": [ + "(a) Caltech-5V" + ], + "image_footnote": [], + "bbox": [ + 514, + 266, + 691, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2747f88cb4ba3cc24c3cd73a34ae0d1b17a3d4e36003973fce81f362debab36.jpg", + "image_caption": [ + "(b) ALOI-100" + ], + "image_footnote": [], + "bbox": [ + 696, + 266, + 877, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a184b06c80493552a4a7dec02d99654557a186dec520c0d0c59456030409d828.jpg", + "image_caption": [ + "(a) Pre-training $(\\mathrm{NMI} = 44.18\\%)$" + ], + "image_footnote": [], + "bbox": [ + 506, + 411, + 676, + 508 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4864853a6c5b8ece8aea7eda096bea427ceb7de01d819a16b259cd0ce1295390.jpg", + "image_caption": [ + "Figure 7. Parameter analyses for $\\zeta$ and $\\lambda$ with $r = 0.5$ .", + "(b) Training $(\\mathrm{NMI} = 79.61\\%)$" + ], + "image_footnote": [], + "bbox": [ + 709, + 410, + 890, + 511 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b7e775dda1051fb260d6174348c8a73dc5a5f9b93a95534842b4db00fbdc4f8.jpg", + "image_caption": [ + "(c) Pre-training $(\\mathrm{NMI} = 33.74\\%)$" + ], + "image_footnote": [], + "bbox": [ + 503, + 534, + 676, + 633 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc926b38bf6e88ca2189db92557430db3e98a2292d60f720f102c51885195282.jpg", + "image_caption": [ + "(d) Training $(\\mathrm{NMI} = 89.81\\%)$", + "Figure 8. Visualization for Caltech-5V and NoisyMNIST." + ], + "image_footnote": [], + "bbox": [ + 709, + 532, + 885, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 686, + 617, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose FreeBSD, a novel semantic learning paradigm free from imputation and alignment compared to existing consistency learning. We design prototype-based contrastive clustering to discover a shared semantic space, where observations converge toward their respective semantic prototype and are encoded as consensus semantics representations for clustering. Furthermore, we employ modularity-inspired graph clustering to enrich semantic representation with view-specific cluster information. The effective synergy of consensus semantic learning and cluster semantic enhancement makes FreeBSD excel in most complex IMVC tasks.", + "bbox": [ + 496, + 710, + 890, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5078", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 76, + 90, + 235, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Key R&D Program of China under Grant No.2022ZD0209103, the National Natural Science Foundation of China (project no. 62325604, 62276271, 62406329, 62476281, 62441618), National Natural Science Foundation of China Joint Found under Grant No. U24A20323.", + "bbox": [ + 76, + 114, + 468, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 209, + 174, + 224 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ulrik Brandes, Daniel Delling, Marco Gaertler, Robert Görke, Martin Hoefer, Zoran Nikoloski, and Dorothea Wagner. Maximizing modularity is hard. arXiv preprint physics/0608255, 2006. 5", + "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2", + "[3] Chenhang Cui, Yazhou Ren, Jingyu Pu, Jiawei Li, Xiaorong Pu, Tianyi Wu, Yutao Shi, and Lifang He. A novel approach for effective multi-view clustering with information-theoretic perspective. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[4] Zhibin Dong, Jiaqi Jin, Yuyang Xiao, Bin Xiao, Siwei Wang, Xinwang Liu, and En Zhu. Subgraph propagation and contrastive calibration for incomplete multiview data clustering. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1, 2", + "[5] Guowang Du, Lihua Zhou, Yudi Yang, Kevin Lu, and Lizhen Wang. Deep multiple auto-encoder-based multi-view clustering. Data Science and Engineering, 6(3):323-338, 2021. 5", + "[6] Uno Fang, Man Li, Jianxin Li, Longxiang Gao, Tao Jia, and Yanchun Zhang. A comprehensive survey on multi-view clustering. IEEE Transactions on Knowledge and Data Engineering, 35(12):12350-12368, 2023. 1", + "[7] Wei Feng, Guoshuai Sheng, Qianqian Wang, Quanxue Gao, Zhiqiang Tao, and Bo Dong. Partial multi-view clustering via self-supervised network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11988-11995, 2024. 1", + "[8] Ruiming Guo, Mouxing Yang, Yijie Lin, Xi Peng, and Peng Hu. Robust contrastive multi-view clustering against dual noisy correspondence. Advances in Neural Information Processing Systems, 37:121401-121421, 2024. 2", + "[9] Changhao He, Hongyuan Zhu, Peng Hu, and Xi Peng. Robust variational contrastive learning for partially view-unaligned clustering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4167-4176, 2024. 2", + "[10] Junlin Hu, Jiwen Lu, and Yap-Peng Tan. Sharable and individual multi-view metric learning. IEEE transactions on pattern analysis and machine intelligence, 40(9):2281-2288, 2017. 3", + "[11] Peng Hu, Liangli Zhen, Xi Peng, Hongyuan Zhu, Jie Lin, Xu Wang, and Dezhong Peng. Deep supervised multi-view learning with graph priors. IEEE Transactions on Image Processing, 33:123-133, 2023. 1" + ], + "bbox": [ + 78, + 234, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Dong Huang, Chang-Dong Wang, and Jian-Huang Lai. Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity. IEEE Transactions on Knowledge and Data Engineering, 35(11):11388-11402, 2023. 5", + "[13] Jiabo Huang, Shaogang Gong, and Xiatian Zhu. Deep semantic clustering by partition confidence maximisation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8849-8858, 2020. 1", + "[14] Zhenyu Huang, Peng Hu, Joey Tianyi Zhou, Jiancheng Lv, and Xi Peng. Partially view-aligned clustering. Advances in Neural Information Processing Systems, 33:2892-2902, 2020. 1", + "[15] Yangbangyan Jiang, Qianqian Xu, Zhiyong Yang, Xiaochun Cao, and Qingming Huang. Dm2c: Deep mixed-modal clustering. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[16] Jiaqi Jin, Siwei Wang, Zhibin Dong, Xinwang Liu, and En Zhu. Deep incomplete multi-view clustering with cross-view partial sample and prototype alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11600-11609, 2023. 1, 2", + "[17] Guanzhou Ke, Bo Wang, Xiaoli Wang, and Shengfeng He. Rethinking multi-view representation learning via distilled disentangling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26774-26783, 2024. 1", + "[18] Brian W Kernighan and Shen Lin. An efficient heuristic procedure for partitioning graphs. The Bell system technical journal, 49(2):291-307, 1970. 5", + "[19] Haobin Li, Yunfan Li, Mouxing Yang, Peng Hu, Dezhong Peng, and Xi Peng. Incomplete multi-view clustering via prototype-based imputation. arXiv preprint arXiv:2301.11045, 2023. 1, 2", + "[20] Xingfeng Li, Yinghui Sun, Quansen Sun, Zhenwen Ren, and Yuan Sun. Cross-view graph matching guided anchor alignment for incomplete multi-view clustering. Information Fusion, 100:101941, 2023. 1", + "[21] Yingming Li, Ming Yang, and Zhongfei Zhang. A survey of multi-view representation learning. IEEE transactions on knowledge and data engineering, 31(10):1863-1883, 2018. 1", + "[22] Yunfan Li, Peng Hu, Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8547-8555, 2021. 2", + "[23] Yijie Lin, Yuanbiao Gou, Zitao Liu, Boyun Li, Jiancheng Lv, and Xi Peng. Completer: Incomplete multi-view clustering via contrastive prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11174-11183, 2021. 5", + "[24] Yijie Lin, Yuanbiao Gou, Xiaotian Liu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4): 4447-4461, 2022. 1", + "[25] Xinwang Liu, Xinzhong Zhu, Miaomiao Li, Lei Wang, Chang Tang, Jianping Yin, Dinggang Shen, Huaimin Wang," + ], + "bbox": [ + 501, + 92, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5079", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Wen Gao. Late fusion incomplete multi-view clustering. IEEE transactions on pattern analysis and machine intelligence, 41(10):2410-2423, 2018. 1", + "[26] Yiding Lu, Haobin Li, Yunfan Li, Yijie Lin, and Xi Peng. A survey on deep clustering: from the prior perspective. *Vicinagearth*, 1(1):4, 2024. 2", + "[27] Yiding Lu, Yijie Lin, Mouxing Yang, Dezhong Peng, Peng Hu, and Xi Peng. Decoupled contrastive multi-view clustering with high-order random walks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14193-14201, 2024. 2", + "[28] Mark EJ Newman. Modularity and community structure in networks. Proceedings of the national academy of sciences, 103(23):8577-8582, 2006. 4", + "[29] Jingyu Pu, Chenhang Cui, Xinyue Chen, Yazhou Ren, Xiaorong Pu, Zhifeng Hao, S Yu Philip, and Lifang He. Adaptive feature imputation with latent graph for deep incomplete multi-view clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14633-14641, 2024. 1", + "[30] Yuming Shen, Ziyi Shen, Menghan Wang, Jie Qin, Philip Torr, and Ling Shao. You never cluster alone. Advances in Neural Information Processing Systems, 34:27734-27746, 2021. 2", + "[31] Yuan Sun, Yang Qin, Yongxiang Li, Dezhong Peng, Xi Peng, and Peng Hu. Robust multi-view clustering with noisy correspondence. IEEE Transactions on Knowledge and Data Engineering, 2024. 2", + "[32] Huayi Tang and Yong Liu. Deep safe multi-view clustering: Reducing the risk of clustering performance degradation caused by view increase. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 202-211, 2022. 1, 2", + "[33] Jingjing Tang, Qingqing Yi, Saiji Fu, and Yingjie Tian. Incomplete multi-view learning: Review, analysis, and prospects. Applied Soft Computing, page 111278, 2024. 1", + "[34] Xinhang Wan, Jiyuan Liu, Xinbiao Gan, Xinwang Liu, Siwei Wang, Yi Wen, Tianjiao Wan, and En Zhu. One-step multiview clustering with diverse representation. IEEE Transactions on Neural Networks and Learning Systems, pages 1-13, 2024. 1", + "[35] Xinhang Wan, Bin Xiao, Xinwang Liu, Jiyuan Liu, Weixuan Liang, and En Zhu. Fast continual multi-view clustering with incomplete views. IEEE Transactions on Image Processing, 33:2995-3008, 2024. 1", + "[36] Haoqing Wang, Xun Guo, Zhi-Hong Deng, and Yan Lu. Rethinking minimal sufficient representation in contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16041-16050, 2022. 1", + "[37] Qianqian Wang, Zhengming Ding, Zhiqiang Tao, Quanxue Gao, and Yun Fu. Generative partial multi-view clustering with adaptive fusion and cycle consistency. IEEE Transactions on Image Processing, 30:1771-1783, 2021. 1, 2", + "[38] Siwei Wang, Xinwang Liu, Suyuan Liu, Jiaqi Jin, Wenxuan Tu, Xinzhong Zhu, and En Zhu. Align then fusion: Generalized large-scale multi-view clustering with anchor matching correspondences. Advances in Neural Information Processing Systems, 35:5882-5895, 2022. 2" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Yiming Wang, Dongxia Chang, Zhiqiang Fu, Jie Wen, and Yao Zhao. Incomplete multi-view clustering via cross-view relation transfer. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 2", + "[40] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 3", + "[41] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487. PMLR, 2016. 5", + "[42] Cai Xu, Ziyu Guan, Wei Zhao, Hongchang Wu, Yunfei Niu, and Beilei Ling. Adversarial incomplete multi-view clustering. In IJCAI, pages 3933-3939, 2019. 2", + "[43] Jie Xu, Chao Li, Yazhou Ren, Liang Peng, Yujie Mo, Xiaoshuang Shi, and Xiaofeng Zhu. Deep incomplete multiview clustering via mining cluster complementarity. In Proceedings of the AAAI conference on artificial intelligence, pages 8761-8769, 2022. 3", + "[44] Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiaofeng Zhu, and Lifang He. Multi-level feature learning for contrastive multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16051-16060, 2022. 5", + "[45] Jie Xu, Chao Li, Liang Peng, Yazhou Ren, Xiaoshuang Shi, Heng Tao Shen, and Xiaofeng Zhu. Adaptive feature projection with distribution alignment for deep incomplete multiview clustering. IEEE Transactions on Image Processing, 32:1354-1366, 2023. 2", + "[46] Xiaoqiang Yan, Zhixiang Jin, Fengshou Han, and Yangdong Ye. Differentiable information bottleneck for deterministic multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27435-27444, 2024. 1", + "[47] Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, and Xi Peng. Partially view-aligned representation learning with noise-robust contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1134–1143, 2021. 2", + "[48] Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, and Xi Peng. Partially view-aligned representation learning with noise-robust contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1134–1143, 2021. 2", + "[49] Mouxing Yang, Yunfan Li, Peng Hu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Robust multi-view clustering with incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1055-1069, 2022. 2", + "[50] Shengju Yu, Suyuan Liu, Siwei Wang, et al. Sparse low-rank multi-view subspace clustering with consensus anchors and unified bipartite graph. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3", + "[51] Shengju Yu, Siwei Wang, Zhibin Dong, et al. A nonparametric graph clustering framework for multi-view data. In Proceedings of the AAAI conference on artificial intelligence, pages 16558-16567, 2024. 4" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5080", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Pengxin Zeng, Mouxing Yang, Yiding Lu, Changqing Zhang, Peng Hu, and Xi Peng. Semantic invariant multiview clustering with fully incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2139-2150, 2023. 2", + "[53] Changqing Zhang, Yajie Cui, Zongbo Han, Joey Tianyi Zhou, Huazhu Fu, and Qinghua Hu. Deep partial multi-view learning. IEEE transactions on pattern analysis and machine intelligence, 44(5):2402-2415, 2020. 2", + "[54] Yi Zhang, Xinwang Liu, Siwei Wang, Jiyuan Liu, Sisi Dai, and En Zhu. One-stage incomplete multi-view clustering via late fusion. In Proceedings of the 29th ACM international conference on multimedia, pages 2717-2725, 2021. 1", + "[55] Yi Zhang, Fengyu Tian, Chuan Ma, Miaomiao Li, Hengfu Yang, Zhe Liu, En Zhu, and Xinwang Liu. Regularized instance weighting multiview clustering via late fusion alignment. IEEE Transactions on Neural Networks and Learning Systems, 2024. 3", + "[56] Huasong Zhong, Chong Chen, Zhongming Jin, and Xian-Sheng Hua. Deep robust clustering by contrastive learning. arXiv preprint arXiv:2008.03030, 2020. 2", + "[57] Lihua Zhou, Guowang Du, Kevin Lu, Lizheng Wang, and Jingwei Du. A survey and an empirical evaluation of multiview clustering approaches. ACM Computing Surveys, 56 (7):1-38, 2024. 1" + ], + "bbox": [ + 78, + 90, + 468, + 444 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5081", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/380dd383-b00d-4370-9caa-3f9395391322_model.json b/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/380dd383-b00d-4370-9caa-3f9395391322_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a5f53563e46b850fa914cc1c61883c611a450993 --- /dev/null +++ b/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/380dd383-b00d-4370-9caa-3f9395391322_model.json @@ -0,0 +1,3027 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.13, + 0.882, + 0.177 + ], + "angle": 0, + "content": "Imputation-free and Alignment-free: Incomplete Multi-view Clustering Driven by Consensus Semantic Learning" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.203, + 0.804, + 0.24 + ], + "angle": 0, + "content": "Yuzhuo Dai\\(^{1}\\), Jiaqi Jin\\(^{1}\\), Zhibin Dong\\(^{1}\\), Siwei Wang\\(^{2,*}\\), Xinwang Liu\\(^{1,*}\\), En Zhu\\(^{1,*}\\), Xihong Yang\\(^{1}\\), Xinbiao Gan\\(^{1}\\), Yu Feng\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.24, + 0.732, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)National University of Defense Technology, Changsha, China" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.257, + 0.69, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)Intelligent Game and Decision Lab, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.278, + 0.632, + 0.292 + ], + "angle": 0, + "content": "{yzdai24, wangsiwei13}@nudt.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.313, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.358, + 0.473, + 0.798 + ], + "angle": 0, + "content": "In incomplete multi-view clustering (IMVC), missing data induce prototype shifts within views and semantic inconsistencies across views. A feasible solution is to explore cross-view consistency in paired complete observations, further imputing and aligning the similarity relationships inherently shared across views. Nevertheless, existing methods are constrained by two-tiered limitations: (1) Neither instance- nor cluster-level consistency learning construct a semantic space shared across views to learn consensus semantics. The former enforces cross-view instances alignment, and wrongly regards unpaired observations with semantic consistency as negative pairs; the latter focuses on cross-view cluster counterparts while coarsely handling fine-grained intra-cluster relationships within views. (2) Excessive reliance on consistency results in unreliable imputation and alignment without incorporating view-specific cluster information. Thus, we propose an IMVC framework, imputation- and alignment-free for consensus semantics learning (FreeCSL). To bridge semantic gaps across all observations, we learn consensus prototypes from available data to discover a shared space, where semantically similar observations are pulled closer for consensus semantics learning. To capture semantic relationships within specific views, we design a heuristic graph clustering based on modularity to recover cluster structure with intra-cluster compactness and inter-cluster separation for cluster semantics enhancement. Extensive experiments demonstrate, compared to state-of-the-art competitors, FreeCSL achieves more confident and robust assignments on IMVC task." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.824, + 0.21, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.849, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Thanks to representation learning enhanced by data observed from different perspectives, multi-view clustering" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.327, + 0.673, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.445, + 0.674, + 0.454 + ], + "angle": 0, + "content": "(a) Consistency Learning Paradigms" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.327, + 0.701, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.327, + 0.892, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.445, + 0.88, + 0.454 + ], + "angle": 0, + "content": "(b) Prototype-based Consensus Learning" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Figure 1. Research Motivation for Consensus Semantic Learning. (a) two existing paradigms, instance- and cluster-level, either treat cross-view unpaired observations with similar semantics as false negatives or neglect to treat within-view observations with similar semantics as true positives; (b) we propose a novel semantic-level paradigm based on contrastive clustering with a set of consensus prototypes to foster semantic consistency across all view data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.893, + 0.776 + ], + "angle": 0, + "content": "(MVC) has achieved significant breakthroughs in the field of unsupervised learning [3, 6, 11, 17, 21, 34, 46, 57]. However, in practical applications, the assumption of data completeness is often difficult to satisfy that incomplete multiview clustering (IMVC) is introduced [7, 25, 29, 33, 35, 37, 54]. In IMVC, missing data causes prototypes shifts within views and semantic misalignment across views, due to the discrepancy between the distributions of complete and incomplete instances [14, 19]. More and more studies [4, 16] have noted that variations in complete instances across different views further exacerbate prototype misalignment. It is challenging to achieve semantic consistency on cluster assignments across all view data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To alleviate prototype shifts and misalignment while promoting semantic consensus in cluster assignments, existing IMVC methods explore consistency information from complete instances for imputation and alignment [20, 24, 32]. Unfortunately, they still suffer from several significant drawbacks in practical applications. In terms of consistency learning, as shown in Fig. 1 (a), one widely used paradigm is instance-level [13, 36], which pulls paired ob" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5071" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "observations (the same instance in different views) closer in the representation space by enforcing highly similar representations, but may inadvertently discard view-specific information [26, 31, 48]. More importantly, it tends to introduce false-negative noise, where unpaired observations with the same semantics are incorrectly treated as negative pairs [2, 8, 27]. [49],[52] and [27] attempted to optimize this issue by proposing a cluster-level paradigm [47] that encourages observations to find their cluster counterparts across different views. It learns a cluster space shared across views but not applicable within specific views, as there are no clustering interactions among intra-view observations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.473, + 0.611 + ], + "angle": 0, + "content": "Since the above consistency learning paradigms fail to account for semantic consistency across all view data, they cannot mitigate semantic gaps among intra-cluster observations [22, 30, 56]. Therefore, imputation of missing data is required to restore the original data distribution, including neighborhood-based recovery via cross-view graph structure transfer [39], adversarial generation or contrastive prediction through cross-view mutual information interaction [15, 37, 42, 53], and prototype-based imputation via cross-view sample-prototype relationship inheritance [19]. Meanwhile, cross-view alignment of assignments [22, 32], prototypes [16, 19, 38], or distributions [4, 45] is also a crucial approach for further enhancing consistent learning. More related works are enumerated in Appendix A. Both imputation and alignment are limited by the consistent information from cross-view paired data and cannot fully exploit within-view unpaired data to mine view-specific cluster information, i.e., within-cluster compactness and between-cluster separation [9, 19]. Particularly, once the amount of missing data is too excessive to provide sufficient consistent information, model performance may even decline rapidly due to the noise introduced by improper imputation and alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Realizing the above issues, we ponder: imputation and alignment aim to restore the similarity relationships inherited from other complete views for the clustering task. Can we directly bridge the semantic gaps while exploring the semantic relationships among all data in consistency learning, thereby avoiding imputation and alignment operations with uncertainty noise? Thus, we propose an IMVC framework, a novel semantic-level paradigm, as shown in Fig. 1 (b), that is driven by imputation- and alignment-free consensus semantic learning (FreeCSL). Notably, our proposed consensus learning involves concurrent interaction among all data, rather than being limited to within or across views. To bridge semantic gaps among view observations and learn cluster semantics information, FreeCSL employs contrastive clustering based on consensus prototypes to discover a shared semantic space, where observations converge toward their semantic prototypes respectively. In practice, we set missing statistical weights for observations to facilitate view collaboration in consensus representations that in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.302 + ], + "angle": 0, + "content": "tegrate detailed information of all views while adapting the impact of different views with different missing instances. Based on consensus representation, our model can construct robust consensus prototypes for semantic-level clustering without imputation and alignment. To discover view-specific semantic relationships and further enrich consensus representations, FreeCSL exploit graph clustering to capture the cluster structure for each view, which maximizes graph modularity within views to enhance intra-cluster connections and reduce inter-cluster interactions. In short, our model encodes data correlation, discovered by semantic learning, into a shared space to obtain consensus semantic representations for instance clustering. Our prominent contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.304, + 0.892, + 0.395 + ], + "angle": 0, + "content": "- In terms of bridging semantic gaps, we design the consensus semantic learning module, a novel semantic-level paradigm based prototypical contrastive clustering, to discover a common semantic space where all observations are embedded as representations with consistent semantics, avoiding additional imputation and alignment." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.395, + 0.892, + 0.469 + ], + "angle": 0, + "content": "- In terms of exploring semantic relationships, we employ the cluster semantics enhancement module, a heuristic graph clustering method with modularity-based learning objective, to mine the inherent cluster structures that reveal the semantic correlations within views." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.471, + 0.892, + 0.515 + ], + "angle": 0, + "content": "- Extensive experiments show our model surpasses state-of-the-art (SOTA) competitors in complex tasks with high missing rates, multiple clusters and large-scale data." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.304, + 0.892, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.527, + 0.591, + 0.542 + ], + "angle": 0, + "content": "2. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.644 + ], + "angle": 0, + "content": "In this section, FreeCSL, a deep IMVC method without imputation or alignment, is proposed to learn consensus semantic representations for clustering. The framework in Fig. 2 coordinates reconstruction (REC) module, cross-view consensus semantic learning (CSL) module, and within-view cluster semantic enhancement (CSE) module." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.65, + 0.685, + 0.665 + ], + "angle": 0, + "content": "2.1. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Notations. Given a multi-view dataset \\(\\mathcal{X} = \\{\\mathbf{X}^v\\in\\) \\(\\mathbb{R}^{N\\times D_v}\\}_{v = 1}^V\\) with \\(N\\) instances across \\(V\\) views, \\(\\tilde{\\mathbf{X}}^v =\\) \\(\\{\\tilde{\\mathbf{x}}_i^v\\in \\mathbb{R}^{D_v}\\}_{i = 1}^{N_v}\\) is an incomplete subset of the \\(v\\) -th view with \\(N_{v}\\) observations, and \\(\\overline{\\mathbf{X}}^{m,n} = \\{(\\overline{\\mathbf{x}}_i^m,\\overline{\\mathbf{x}}_i^n)\\}_{i = 1}^{N_{mn}}\\) is a pair-observed subset with \\(N_{mn}\\) instances observed in both the \\(m\\) -th and \\(n\\) -th view. The task is to partition \\(N\\) instances into \\(K\\) clusters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.841 + ], + "angle": 0, + "content": "Definition 1. Instance-level Consistency (IC): \\(\\forall m \\neq n\\), \\(\\mathbf{x}_i^m\\) and \\(\\mathbf{x}_j^n\\) are instance-level consistent across views if \\(i = j\\) (they are cross-view observations of the same instance \\(\\mathbf{x}\\)), expressed as \\(I(\\mathbf{x}_i^m, \\mathbf{x}_j^n) = 1\\) and \\(O\\) otherwise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Definition 2. Cluster-level Consistency (CC): \\(\\forall m \\neq n\\), \\(\\mathbf{x}_i^m\\) and \\(\\mathbf{x}_j^n\\) are cluster-level consistent across views if they belong to the same cluster \\(k\\), expressed as \\(C(\\mathbf{x}_i^m, \\mathbf{x}_j^n) = 1\\) and 0 otherwise." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5072" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "Definition 3. Semantic-level Consensus (SC): \\(\\forall m\\) and \\(n\\), \\(\\mathbf{x}_i^m\\) and \\(\\mathbf{x}_j^n\\) achieve semantic-level consensus in MVC task if all observations share a set of cluster prototypes \\(\\mathbf{C} = \\{\\mathbf{c}_k\\}_{k=1}^K\\) and \\(\\arg \\max_k \\rho(\\mathbf{x}_i^m, \\mathbf{c}_k) = \\arg \\max_k \\rho(\\mathbf{x}_j^n, \\mathbf{c}_k)\\), expressed as \\(S(\\mathbf{x}_i^m, \\mathbf{x}_j^n) = 1\\) and \\(0\\) otherwise." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.18, + 0.471, + 0.226 + ], + "angle": 0, + "content": "Theorem 1. Consensus semantic learning yields more confident and robust cluster assignments than instance- and cluster-level paradigms. (Proof is provided in Appendix B.)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.231, + 0.469, + 0.293 + ], + "angle": 0, + "content": "Theorem 2. Since paired observations \\((\\overline{\\mathbf{x}}_i^m, \\overline{\\mathbf{x}}_i^n)\\) inherently satisfy instance- and cluster-level consistency, they can achieve semantic consensus via a shared set of prototypes C. (Proof is provided in Appendix B.)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.308, + 0.33, + 0.323 + ], + "angle": 0, + "content": "2.2. Within-view Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.331, + 0.469, + 0.466 + ], + "angle": 0, + "content": "To avoid clustering instability from similarity measures on manifold structures in high-dimensional spaces, we use autoencoders to encode view observations \\(\\tilde{\\mathbf{X}}^v\\) into clustering-friendly low-dimensional representations \\(\\tilde{\\mathbf{Z}}^v\\). Considering multiple views are mostly heterogeneous and differently distributed, we provide an independent encoder-decoder \\(\\{\\mathcal{E}_v,\\mathcal{D}_v\\}_{v = 1}^V\\) for each view. The encoder embeds the latent representation and the decoder recovers the original data from it, which jointly minimize reconstruction loss:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.471, + 0.469, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {r e c} = \\sum_ {v = 1} ^ {V} \\left\\| \\tilde {\\mathbf {X}} ^ {v} - \\mathcal {D} _ {v} \\left(\\tilde {\\mathbf {Z}} ^ {v}\\right) \\right\\| _ {F} ^ {2} \\tag {1} \\\\ = \\sum_ {v = 1} ^ {V} \\sum_ {i = 1} ^ {N _ {v}} \\| \\tilde {\\mathbf {x}} _ {i} ^ {v} - \\mathcal {D} _ {v} \\left(\\mathcal {E} _ {v} \\left(\\tilde {\\mathbf {x}} _ {i} ^ {v}\\right)\\right) \\| _ {2} ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.564, + 0.469, + 0.598 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}_v(\\tilde{\\mathbf{X}}^v;\\theta_v):\\tilde{\\mathbf{X}}^v\\in \\mathbb{R}^{N_v\\times D_v}\\to \\tilde{\\mathbf{Z}}^v\\in \\mathbb{R}^{N_v\\times d}\\) and \\(\\mathcal{D}_v(\\tilde{\\mathbf{Z}}^v;\\phi_v):\\tilde{\\mathbf{Z}}^v\\in \\mathbb{R}^{N_v\\times d}\\to \\hat{\\mathbf{X}}^v\\in \\mathbb{R}^{N_v\\times D_v}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.434, + 0.623 + ], + "angle": 0, + "content": "2.3. Cross-view Consensus Semantic Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.719 + ], + "angle": 0, + "content": "Based on Theorem 1, we design prototypical contrastive clustering for consensus semantic learning. To reach semantic-level consensus on assignments, consensus prototypes with all view information, are introduced into contrastive clustering to explore a shared semantic space, where all observations with similar semantics are pulled closer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.855 + ], + "angle": 0, + "content": "Consensus Representations and Consensus Prototypes. Due to the assumptions of consistency and complementarity in MVC, inseparable clusters in one view will become linearly separable by introducing complementary information from other views [10, 43, 55]. We utilize a fusion manner, denoted as \\(\\mathbb{T}(\\cdot)\\), to map the latent representations \\(\\mathbf{Z}^v\\) into a linearly weighted representation space for consensus representations \\(\\mathbf{Z} \\in \\mathbb{R}^{N \\times d}\\) that integrates consistency and complementary information from multiple views:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.863, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} = \\mathbb {T} \\left(\\left\\{\\mathbf {Z} ^ {v} \\right\\} _ {v = 1} ^ {V}\\right) = \\sum_ {v = 1} ^ {V} \\mathbf {w} ^ {v} \\mathbf {Z} ^ {v} = \\left\\{\\sum_ {v = 1} ^ {V} w _ {i} ^ {v} \\mathbf {z} _ {i} ^ {v} \\right\\} _ {i = 1} ^ {N}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.882, + 0.108 + ], + "angle": 0, + "content": "where \\(w_{i}^{v}\\) is the instance-level fine-grained fusion weight:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.118, + 0.892, + 0.156 + ], + "angle": 0, + "content": "\\[\nw _ {i} ^ {v} = \\frac {\\mathbb {I} \\left(\\mathbf {z} _ {i} ^ {v} \\neq N a N\\right)}{\\sum_ {v ^ {\\prime}} ^ {V} \\mathbb {I} \\left(\\mathbf {z} _ {i} ^ {v ^ {\\prime}} \\neq N a N\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.166, + 0.892, + 0.302 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}(\\cdot)\\) is the indicative function that takes 1 when \\(\\mathbf{z}_i^v\\) is the representation of \\(i\\)-th instance \\(\\mathbf{X}_i\\) observed in view \\(v\\), and 0 otherwise. We set a completeness statistical weight \\(w_i^v\\) based on the number of observations in \\(V\\) views for instance \\(\\mathbf{X}_i\\) as the fusion weight, which makes use of view-specific complementary information without discarding unpaired observations and mitigating the negative impact of missing noise by adapting to the differences in missing instances across different views." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.303, + 0.892, + 0.393 + ], + "angle": 0, + "content": "We derive a set of consensus semantic prototypes \\(\\mathbf{C} = \\{\\mathbf{c}_k\\in \\mathbb{R}^d\\}_{k = 1}^K\\) via \\(k\\)-means on consensus representations \\(\\mathbf{Z}\\). As \"a representative embedding for semantically similar observations\", \\(\\mathbf{C}\\) comprehensively captures the semantic information of all data [50] and is continuously refined throughout consensus representation learning." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.394, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Prototype-based Contrastive Clustering. We perform semantic-level contrastive clustering in a shared vector space spanned by consensus prototypes, where the latent representation \\(\\mathbf{z}_i^v\\) encoded as the semantic representation \\(\\mathbf{h}_i^v\\) and \\(\\mathbf{h}_i^v\\) is assigned to the prototype \\(\\mathbf{c}_k\\) with the longest projected distance to obtain optimal assignments, in two steps:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.485, + 0.892, + 0.532 + ], + "angle": 0, + "content": "- Semantic similarity measure: Encode the semantic representation \\(\\mathbf{h}_i^v\\) and project it onto each prototype, then calculate the probability \\(p_{i,k}^{v}\\) of belonging to cluster \\(k\\):" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.544, + 0.892, + 0.583 + ], + "angle": 0, + "content": "\\[\np _ {i, k} ^ {v} = \\frac {\\exp \\left(\\mathbf {h} _ {i} ^ {v \\top} \\mathbf {c} _ {k} / \\tau\\right)}{\\sum_ {k ^ {\\prime}} ^ {K} \\exp \\left(\\mathbf {h} _ {i} ^ {v \\top} \\mathbf {c} _ {k ^ {\\prime}} / \\tau\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.89, + 0.626 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a temperature parameter [40]. \\(\\mathbf{p}_i^v = \\{p_{i,k}^v\\}_{k = 1}^K\\) is the soft assignments of observation \\(\\mathbf{x}_i^v\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.627, + 0.892, + 0.732 + ], + "angle": 0, + "content": "- Swapped knowledge distillation: Based on Theorem 2, \\((\\overline{\\mathbf{x}}_i^m,\\overline{\\mathbf{x}}_i^n)\\) share the same cluster semantics on consensus prototypes \\(\\mathbf{C}\\), thus their true labels \\(\\mathbf{y}_i^m,\\mathbf{y}_i^n\\in \\mathbb{R}^K\\) should be are distributionally consistent. To this end, \"swapped\" knowledge distillation (KD) utilizes pseudo-labels \\(\\mathbf{q}_i^m\\), \\(\\mathbf{q}_i^n\\) as mutual supervised signals to prompt their cluster assignments \\(\\mathbf{p}_i^m,\\mathbf{p}_i^n\\) as identical as possible:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.743, + 0.892, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\ell_ {c c} ^ {m, n} = \\ell_ {k d} \\left(\\mathbf {H} ^ {m}, \\mathbf {Q} ^ {n}\\right) + \\ell_ {k d} \\left(\\mathbf {H} ^ {n}, \\mathbf {Q} ^ {m}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.772, + 0.892, + 0.843 + ], + "angle": 0, + "content": "where \\(\\ell_{kd}(\\mathbf{H}^m,\\mathbf{Q}^n) = -\\frac{1}{N_{nm}}\\sum_{i = 1}^{N_{mn}}\\sum_{k = 1}^{K}\\mathbf{q}_i^n\\log \\mathbf{p}_i^m\\) We extend Eq. (5) to more than two views, fostering greater view collaboration to enhance semantic learning:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.853, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c c} = \\sum_ {m = 1} ^ {V} \\sum_ {\\substack {n = 1 \\\\ n \\neq m}} ^ {V} \\ell_ {c c} ^ {m, n}. \\tag{6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5073" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.868, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.425, + 0.896, + 0.51 + ], + "angle": 0, + "content": "Figure 2. The framework of FreeCSL. (a) Reconstruction module, encodes observations into clustering-friendly representations for each view; (b) Consensus semantic learning module, learns semantic representations through cross-view contrastive clustering in a shared semantic space, where paired observations are assigned to their nearest semantic prototype for consistent assignments. (c) Cluster semantic enhancement module, enriches semantic representations with cluster structure information through within-view graph clustering, which applies GCN to aggregate view-specific semantic information and maximizes spectral modularity to recover cluster structure with greater separation and lower entropy. Ultimately, perform \\( k \\)-means on the consensus semantic representations to predict cluster labels." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.528, + 0.472, + 0.59 + ], + "angle": 0, + "content": "- Consensus label solution: To obtain pseudo-labels \\(\\mathbf{Q}^m = \\{\\mathbf{q}_i^m\\}_{i=1}^{N_{mn}}\\), project semantic representations \\(\\mathbf{H}^m\\) onto consensus prototypes \\(\\mathbf{C}\\) and maximize similarity between the representation \\(\\mathbf{h}_i^m\\) and its assigned prototype \\(\\mathbf{c}_k\\):" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.593, + 0.47, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} ^ {*} = \\arg \\max _ {\\mathbf {Q} ^ {m} \\in \\mathcal {Q}} \\operatorname {T r} \\left\\{\\left(\\mathbf {C Q} ^ {m}\\right) ^ {\\top} \\mathbf {H} ^ {m}\\right) \\} + \\alpha \\mathcal {R} (\\mathbf {Q} ^ {m}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.47, + 0.658 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}(\\mathbf{Q}^m) = -\\sum_i^{N_{mn}}\\mathbf{q}_i^m\\log \\mathbf{q}_i^m\\) , an entropy regularizer prevents trivial solutions via the smoothness \\(\\alpha\\)" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.664, + 0.456, + 0.68 + ], + "angle": 0, + "content": "2.4. Within-view Cluster Semantic Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Considering that graph pooling can discover data implicit correlation and reduce missing noise via aggregating the information of neighboring nodes, we design a graph clustering to enhance representations with cluster semantic information. It combines a heuristic learning objective based on modularity, an evaluation metric for cluster structure quality, to effectively recover the cluster structure[51]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Modularity. Modularity [28] quantifies the deviation between the actual cluster structure and the expected structure generated by random combinations in the null model. Modularity matrix \\(\\mathbf{B}\\) is defined as \\(\\mathbf{B} = \\mathbf{A} - \\frac{\\mathbf{d}\\mathbf{d}^{\\top}}{2m}\\), where \\(\\mathbf{A}\\) is the adjacency matrix, \\(\\mathbf{d}\\) is the degree vector and \\(m\\) is the total number of edges in the graph. The entry \\(b_{ij} \\in \\mathbf{B}\\) measures connection strength between nodes \\(i\\) and \\(j\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.893, + 0.605 + ], + "angle": 0, + "content": "Modularity-based Graph clustering. Graph clustering relies on graph pooling and node assignment. For each view, we construct graph structures with \\( k \\)-nearest neighbors (KNN), then deploy graph convolutional network (GCN) \\( \\mathcal{G}_v(\\cdot) \\) to aggregate graph embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.61, + 0.892, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {v} \\left(\\mathbf {Z} ^ {v}; \\mathbf {A} ^ {v}\\right) = \\sigma \\left(\\mathbf {D} ^ {v - \\frac {1}{2}} \\mathbf {A} ^ {v} \\mathbf {D} ^ {v - \\frac {1}{2}} \\mathbf {Z} ^ {v} \\mathbf {W} ^ {v} + \\mathbf {Z} ^ {v} \\mathbf {W} _ {s} ^ {v}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.894, + 0.729 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) is an activation function used to reinforce the nonlinear aggregation capability of GCN. We replace a skip connection matrix \\(\\mathbf{W}_s^v\\) with the self-loop \\(\\mathbf{A}^v +\\mathbf{I}\\) to alleviate the over-smoothing of node embeddings. Graph pooling operates on the latent representations \\(\\mathbf{Z}^v\\), while the adjacency matrix \\(\\mathbf{A}^v\\) is constructed from the original observations \\(\\mathbf{X}^v\\):" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.737, + 0.892, + 0.793 + ], + "angle": 0, + "content": "\\[\na _ {i, j} ^ {v} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\left(\\mathbf {x} _ {i} ^ {v} \\text {a n d} \\mathbf {x} _ {j} ^ {v} \\neq N a N\\right) \\& \\\\ & \\left(\\mathbf {x} _ {i} ^ {v} \\in \\mathcal {N} _ {k} \\left(\\mathbf {x} _ {j} ^ {v}\\right) \\text {o r} \\mathbf {x} _ {j} ^ {v} \\in \\mathcal {N} _ {k} \\left(\\mathbf {x} _ {i} ^ {v}\\right)\\right), \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.799, + 0.893, + 0.845 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}_k(\\mathbf{x}_i^v)\\) is a set including the \\(k\\)-nearest neighbors of \\(\\mathbf{x}_i^v\\). It brings neighbors closer in the embedding space without introducing missing noise into the graph structure." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.845, + 0.892, + 0.875 + ], + "angle": 0, + "content": "The softmax classifier offers the soft assignment \\(\\mathbf{P}^v\\) for node embeddings encoded by \\(\\mathcal{G}_v(\\mathbf{Z}^v;\\mathbf{A}^v)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.883, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} ^ {v} = \\operatorname {S o f t m a x} \\left(\\mathcal {G} _ {v} \\left(\\mathbf {Z} ^ {v}; \\mathbf {A} ^ {v}\\right)\\right). \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5074" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "To obtain a well-separated cluster assignment \\(\\mathbf{P}^v\\), we maximize the spectral modularity \\(\\mathrm{Tr}\\left((\\mathbf{P}^v)^\\top \\mathbf{B}^v\\mathbf{P}^v\\right)\\) [1, 18], which captures intra-cluster connections and intercluster margins by mapping \\(\\mathbf{P}^v\\) to modularity \\(\\mathbf{B}^v\\); To further improve the confidence and robustness of \\(\\mathbf{P}^v\\), we apply a self-supervised signal \\(\\mathbf{L}^v\\), learned from contrastive clustering as detailed in Sec. 2.3, to guide \\(\\mathbf{P}^v\\) in chasing \\(\\mathbf{L}^v\\) via self-knowledge distillation. Specifically, the following objective is designed to optimize \\(\\mathbf{P}^v\\):" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.238, + 0.47, + 0.279 + ], + "angle": 0, + "content": "\\[\n\\ell_ {m} ^ {v} \\left(\\mathbf {P} ^ {v}; \\mathbf {L} ^ {v}\\right) = - \\frac {1}{2 m} \\operatorname {T r} \\left(\\left(\\mathbf {P} ^ {v}\\right) ^ {\\top} \\mathbf {B} ^ {v} \\mathbf {P} ^ {v}\\right) + \\lambda \\operatorname {K L} \\left(\\mathbf {L} ^ {v} \\| \\mathbf {P} ^ {v}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.28, + 0.473, + 0.386 + ], + "angle": 0, + "content": "where Kullback-Leibler divergence \\(\\mathrm{KL}(\\cdot)\\) serves as a robust regularizer, leveraging \\(\\lambda\\) to regulate the information flow of \\(\\mathbf{L}^v\\in \\mathbb{R}^{N\\times K}\\) to ensure stable cluster performance. We project semantic representations \\(\\mathbf{H}^v\\) learned through contrastive clustering onto their \\(k\\)-means prototypes and introduce Student's t-distribution [41] as kernel to predict high-confidence labels \\(\\mathbf{L}^v\\) with a nearly uniform distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.397, + 0.47, + 0.459 + ], + "angle": 0, + "content": "\\[\nl _ {i, k} ^ {v} = \\frac {\\left(1 + \\frac {\\| \\mathbf {h} _ {i} ^ {v} - \\mathbf {c} _ {k} ^ {v} \\| ^ {2}}{\\gamma}\\right) ^ {- \\frac {\\gamma + 1}{2}}}{\\sum_ {k ^ {\\prime}} \\left(1 + \\frac {\\| \\mathbf {h} _ {i} ^ {v} - \\mathbf {c} _ {k ^ {\\prime}} ^ {v} \\| ^ {2}}{\\gamma}\\right) ^ {- \\frac {\\gamma + 1}{2}}} \\in \\mathbf {L} ^ {v}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.469, + 0.47, + 0.501 + ], + "angle": 0, + "content": "We conduct graph clustering on each view separately, then optimize assignments across all views concurrently:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.511, + 0.47, + 0.552 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g c} = \\sum_ {v = 1} ^ {V} \\ell_ {m} ^ {v} \\left(\\mathbf {P} ^ {v}, \\mathbf {L} ^ {v}\\right). \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.563, + 0.444, + 0.58 + ], + "angle": 0, + "content": "2.5. Clustering Driven by Consensus Semantics" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.586, + 0.47, + 0.633 + ], + "angle": 0, + "content": "Our FreeCSL learns the semantic representations \\(\\mathbf{H}^v\\) end-to-end by jointly minimizing the reconstruction loss, contrastive clustering loss and graph clustering loss:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.644, + 0.47, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {r e c} + \\mathcal {L} _ {c c} + \\mathcal {L} _ {g c}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.672, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Without searching for the optimal balancing weights of three loss terms, outstanding performance is easily achieved by applying \\( k \\)-means to the consensus semantic representations \\( \\mathbf{H} \\) learned via the fusion manner \\( \\mathbb{T}\\left(\\{\\mathbf{H}^v\\}_{v=1}^V\\right) \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.734, + 0.469, + 0.794 + ], + "angle": 0, + "content": "The implementation for FreeCSL consists of two stages: warm-up training for the encoder-decoder, followed by fine-tuning for semantic representation learning and clustering. For details, refer to Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.808, + 0.203, + 0.825 + ], + "angle": 0, + "content": "3. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.832, + 0.285, + 0.849 + ], + "angle": 0, + "content": "3.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We select four representative datasets namely Caltech-5V[44], ALOI-100[5], YoutubeFace10[12] and NoisyMNIST[23] to demonstrate our model's performance" + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.84, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 FreeCSL for Learning and Clustering" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.109, + 0.892, + 0.156 + ], + "angle": 0, + "content": "Input: Complete, incomplete and pair-observed multiview dataset \\(\\{\\mathbf{X}^v\\}_{v = 1}^V\\) \\(\\{\\tilde{\\mathbf{X}}^v\\}_{v = 1}^V\\) \\(\\{\\overline{\\mathbf{X}}^{m,n}\\}_{m\\neq n}^V\\) networks \\(\\{\\mathcal{E}_v,\\mathcal{D}_v,\\mathcal{G}_v\\}_{v = 1}^V\\) ; warm-up and fine-tuning epochs \\(e,E\\)" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.156, + 0.62, + 0.169 + ], + "angle": 0, + "content": "for \\(t = 1\\) to e do" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.169, + 0.873, + 0.188 + ], + "angle": 0, + "content": "On \\(\\{\\tilde{\\mathbf{X}}^v\\}_{v = 1}^V\\) , warming up \\(\\{\\mathcal{E}_v,\\mathcal{D}_v\\}_{v = 1}^V\\) with Eq.(1)." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.19, + 0.622, + 0.203 + ], + "angle": 0, + "content": "for \\(t = 1\\) to \\(E\\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.204, + 0.89, + 0.249 + ], + "angle": 0, + "content": "Obtain latent representations \\(\\{\\mathbf{Z}^v |\\mathcal{E}_v:\\mathbf{X}^v\\to \\mathbf{Z}^v\\}_{v = 1}^V\\) Construct consensus prototypes \\(\\mathbf{C} = \\{\\mathbf{c}_k\\}_{k = 1}^K\\) as elaborated in Sec. 2.3." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.25, + 0.857, + 0.266 + ], + "angle": 0, + "content": "Compute the reconstruction loss \\(\\mathcal{L}_{rec}\\) with Eq.(1)." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.266, + 0.882, + 0.281 + ], + "angle": 0, + "content": "// Cross-view Contrastive Clustering" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.282, + 0.735, + 0.296 + ], + "angle": 0, + "content": "for \\(m,n = 1\\) to \\(V(m\\neq n)\\) do" + }, + { + "type": "text", + "bbox": [ + 0.551, + 0.296, + 0.891, + 0.312 + ], + "angle": 0, + "content": "Compute the cluster probability \\(\\mathbf{p}_i^m,\\mathbf{p}_i^n\\) with Eq.(4);" + }, + { + "type": "text", + "bbox": [ + 0.559, + 0.312, + 0.867, + 0.327 + ], + "angle": 0, + "content": "Solve consensus pseudo-label \\(\\mathbf{q}_i^m\\) with Eq.(7);" + }, + { + "type": "text", + "bbox": [ + 0.559, + 0.327, + 0.858, + 0.343 + ], + "angle": 0, + "content": "Get swapped loss \\(\\ell_{cc}^{m,n}\\) with Eq.(5) on \\(\\overline{\\mathbf{X}}^{m,n}\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.347, + 0.87, + 0.362 + ], + "angle": 0, + "content": "Calculate contrastive clustering loss \\(\\mathcal{L}_{cc}\\) with Eq.(6)" + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.363, + 0.842, + 0.376 + ], + "angle": 0, + "content": "// Within-view Graph Clustering" + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.377, + 0.657, + 0.389 + ], + "angle": 0, + "content": "for \\(v = 1\\) to \\(V\\) do" + }, + { + "type": "aside_text", + "bbox": [ + 0.488, + 0.297, + 0.5, + 0.468 + ], + "angle": 0, + "content": "1 \n2 \n3 \n4" + }, + { + "type": "text", + "bbox": [ + 0.559, + 0.391, + 0.85, + 0.407 + ], + "angle": 0, + "content": "Construct adjacency matrix \\(\\mathbf{A}^v\\) with Eq.(9)." + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.407, + 0.825, + 0.421 + ], + "angle": 0, + "content": "Solve node assignment \\(\\mathbf{P}^v\\) with Eq.(10)" + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.422, + 0.808, + 0.436 + ], + "angle": 0, + "content": "Solve pseudo-labels \\(\\mathbf{L}^v\\) with Eq.(12);" + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.437, + 0.865, + 0.453 + ], + "angle": 0, + "content": "Compute KL-modularity loss \\(\\ell_m^v\\) with Eq.(11)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.455, + 0.845, + 0.471 + ], + "angle": 0, + "content": "Calculate graph clustering loss \\(\\mathcal{L}_{gc}\\) with Eq.(13)" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.472, + 0.88, + 0.486 + ], + "angle": 0, + "content": "// Semantic Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.486, + 0.809, + 0.5 + ], + "angle": 0, + "content": "Calculate the overall loss \\(\\mathcal{L}\\) with Eq.(14);" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.5, + 0.814, + 0.515 + ], + "angle": 0, + "content": "Optimize \\(\\{\\mathcal{E}_v,\\mathcal{D}_v,\\mathcal{G}_v\\}_{v = 1}^V\\) to minimize \\(\\mathcal{L}\\)" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.515, + 0.891, + 0.53 + ], + "angle": 0, + "content": "Learn latent and consensus representations \\(\\{\\mathbf{Z}^v\\}_{v = 1}^V\\)" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.53, + 0.89, + 0.546 + ], + "angle": 0, + "content": "and \\(\\mathbf{Z}\\), semantic representations \\(\\{\\mathbf{H}^v\\}_{v=1}^V\\), consensus" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.546, + 0.886, + 0.561 + ], + "angle": 0, + "content": "semantic representations \\(\\mathbf{H}\\), consensus prototypes \\(\\mathbf{C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.563, + 0.842, + 0.582 + ], + "angle": 0, + "content": "Output: \\(\\hat{\\mathbf{Y}} = \\{\\hat{y}_i\\}_{i=1}^N\\) predicted by \\(k\\)-means on \\(\\mathbf{H}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.669 + ], + "angle": 0, + "content": "in comparison with seven SOTA methods summarized in Table 1. To comprehensively evaluate experimental results, three metrics are adopted: accuracy (ACC), normalized mutual information (NMI), and adjusted rand index (ARI)." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.68, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Table 1. SOTA methods categorized by the types of techniques for consistency, imputation, and alignment." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.72, + 0.902, + 0.796 + ], + "angle": 0, + "content": "
CompetitorsConsistencyImputationAlignment
CPM-Nets (TPAMI'20)instance-levelmutual information interaction
COMPLETER (CVPR'21)instance-levelmutual information interaction
DIMVC (AAAI'22)instance-level\\assignment-based
SURE (TPAMI'23)cluster-levelgraph structure transfer
ProLmp (IICA'23)instance-levelsample-prototype relationship inheritanceprototype-based
ICMVC (AAAI'24)instance-levelgraph structure transferassignment-based
DIVIDE (AAAI'24)cluster-levelmutual information interaction
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.714, + 0.833 + ], + "angle": 0, + "content": "3.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Models are trained on an NVIDIA 3090 GPU using Adam optimizer in PyTorch 2.1.0. The warming up and finetuning learning rates are 0.0003 and 0.0005, with a batch size of 512. The hyperparameters are set as follows for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5075" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "different datasets: smoothness \\(\\alpha = 0.5\\), temperature \\(\\tau \\in \\{0.1, 0.2\\}\\), neighbors \\(\\zeta = 3\\), and regularizer weight \\(\\lambda \\in \\{0.05, 0.1, 0.2, 0.3\\}\\). For all datasets, 4-layer autoencoders and 2-layer GCNs with same MLP structures are used for each view. The layers of contrastive clustering and graph clustering are shared across all views via a single FC layer." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.193, + 0.334, + 0.209 + ], + "angle": 0, + "content": "3.3. Competitiveness of FreeCSL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.216, + 0.47, + 0.276 + ], + "angle": 0, + "content": "FreeCSL is compared with seven SOTA methods across three metrics in Table 2. The results indicate that FreeCSL excellently handles challenges of high missing rates, multiple clusters, and large-scale issues in IMVC:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.278, + 0.47, + 0.383 + ], + "angle": 0, + "content": "- From the perspective of effectiveness, FreeBSD surpasses most SOTA models, particularly on challenging datasets like ALOI-100 (100 clusters) and YouTubeFace10 (30,000 samples). It achieves significant improvements in ACC, with gains of \\(15.12\\%\\), \\(26.38\\%\\), \\(23.91\\%\\), \\(26.21\\%\\) for ALOI-100, and \\(1.16\\%\\), \\(8.81\\%\\), \\(6.32\\%\\), \\(1.68\\%\\) for YouTubeFace10, at missing rate \\(r = 0.1\\), \\(0.3\\), \\(0.5\\), \\(0.7\\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.384, + 0.471, + 0.488 + ], + "angle": 0, + "content": "- From the perspective of robustness, Fig. 3 (visualized from Table 2) shows our model's stability as \\( r \\) increases. FreeCSL declines gradually, unlike other models with sharp ACC drops, highlighting its robustness in IMVC task. On the small dataset Caltech-5V with \\( r = 0.7 \\), its Acc remains at \\( 84.56\\% \\), while on the larger dataset NoisyMNIST, also achieves \\( 92.19\\% \\)." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.278, + 0.471, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.745 + ], + "angle": 0, + "content": "The above phenomenon can be explained as follows: As \\( r \\) increases, paired observations become scarce. Methods like instance-level consistency learning or cross-view imputation and alignment (e.g., CPM-Net, COMPLETER, ICMVC), constrained by their dependence on consistency information, will introduce biases such as false negatives and semantic inconsistencies, resulting in degraded performance. For the complex task ALOI-100, cluster-level methods like SURE and DIVIDE surpass instance-level by exploiting cross-view cluster consistency. But, their reliance on cross-view graphs or mutual information for imputation, while neglecting within-view cluster affiliations, restricts performance. ProImp acknowledges the need to integrate within-view and cross-view information for semantically consistent imputation. However, its prototype alignment accumulates errors, particularly with a large number of clusters, achieving only \\( 20\\% \\) ACC on ALOI-100." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.47, + 0.807 + ], + "angle": 0, + "content": "The superiority of FreeCSL stems from the synergistic effect of the CSL and CSE modules, which eliminate semantic gaps and ensure consistent assignments for all observations in a shared semantic space." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.304, + 0.833 + ], + "angle": 0, + "content": "3.4. Understanding FreeCSL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Ablation Study. As shown in Table 3, ablation studies on FreeCSL's three components reveal that the CSL module contributes the most. On ALOI-100, it demonstrates exceptional performance and strong stability, with ACC drop" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.089, + 0.693, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.195, + 0.642, + 0.206 + ], + "angle": 0, + "content": "(a) Caltech-5V" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.09, + 0.878, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.195, + 0.83, + 0.206 + ], + "angle": 0, + "content": "(b) ALOI-100" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.208, + 0.688, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.312, + 0.651, + 0.323 + ], + "angle": 0, + "content": "(c) YoutubeFace10" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.209, + 0.878, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.312, + 0.836, + 0.323 + ], + "angle": 0, + "content": "(d) NoisyMNIST" + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.333, + 0.866, + 0.347 + ], + "angle": 0, + "content": "Figure 3. Visualization for Table 2 based on metric ACC." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.464 + ], + "angle": 0, + "content": "ping only \\(12.4\\%\\) as \\(r\\) rises from 0.1 to 0.7. This is attributed to CSL's establishment of a shared semantic space via prototype-based semantic contrast, reducing semantic gaps within clusters. The CSE module also plays a positive role in discovering tighter cluster structures by optimizing modularity. Thanks to its enhanced semantics, both the REC and CSL modules achieve notable improvements." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.476, + 0.582, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.539, + 0.564, + 0.551 + ], + "angle": 0, + "content": "(a) \\(\\mathbf{Z}^{(1)}\\)" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.476, + 0.663, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.539, + 0.646, + 0.551 + ], + "angle": 0, + "content": "(b) \\(\\mathbf{Z}^{(2)}\\)" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.476, + 0.746, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.539, + 0.728, + 0.551 + ], + "angle": 0, + "content": "(c) \\(\\mathbf{Z}^{(3)}\\)" + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.476, + 0.827, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.539, + 0.81, + 0.551 + ], + "angle": 0, + "content": "(d) \\(\\mathbf{Z}^{(4)}\\)" + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.476, + 0.91, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.856, + 0.54, + 0.885, + 0.552 + ], + "angle": 0, + "content": "(e) Z" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.565, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Figure 4. Similarity matrices of \\(\\{\\mathbf{Z}^v\\}_{v=1}^4\\), \\(\\mathbf{Z}\\) without consensus semantic learning on ALOI-100 with \\(r = 0.5\\)." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.607, + 0.581, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.67, + 0.566, + 0.683 + ], + "angle": 0, + "content": "(a) \\(\\mathbf{H}^{(1)}\\)" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.607, + 0.663, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.67, + 0.648, + 0.683 + ], + "angle": 0, + "content": "(b) \\(\\mathbf{H}^{(2)}\\)" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.607, + 0.745, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.67, + 0.728, + 0.683 + ], + "angle": 0, + "content": "(c) \\(\\mathbf{H}^{(3)}\\)" + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.607, + 0.827, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.67, + 0.811, + 0.683 + ], + "angle": 0, + "content": "(d) \\(\\mathbf{H}^{(4)}\\)" + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.607, + 0.91, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.854, + 0.671, + 0.886, + 0.683 + ], + "angle": 0, + "content": "(e) H" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.69, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Figure 5. Similarity matrices of \\(\\{\\mathbf{H}^v\\}_{v=1}^4\\), \\(\\mathbf{H}\\) with consensus semantic learning on ALOI-100 with \\(r = 0.5\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Imputation- and Alignment-free CSL. To verify FreeCSL can reduce semantic gaps and capture semantic relationships without requiring imputation or alignment, we set up two imputation experiments: one imputation for latent representations \\(\\mathbf{Z}^v\\) without consensus semantic learning named ILR, and the second for semantic representations \\(\\mathbf{H}^v\\) learned from consensus semantic learning named ISR. They transfer complete \\(k\\)-nn graph structures from other views to incomplete views and utilize corresponding \\(k\\) neighbors for imputation. The control groups perform \\(k\\)-means on the sum of all view matrices to predict cluster" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5076" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.096, + 0.09, + 0.873, + 0.104 + ], + "angle": 0, + "content": "Table 2. Performance comparison on four multi-view benchmarks. The best and second-best results are highlighted in red and blue." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.11, + 0.894, + 0.47 + ], + "angle": 0, + "content": "
Missing ratesr=0.1r=0.3r=0.5r=0.7
MetricsACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)
Caltech-5VCPM-Nets85.7274.6271.9285.4372.9070.8481.0868.6166.0879.5065.5262.70
COMPLETETER58.6867.6942.9865.0768.7547.1064.4666.0942.8069.0067.4651.49
DIMVC85.8680.0174.6378.9078.7669.5580.0170.0866.3569.8660.1648.70
SURE75.6469.1262.8377.8967.8961.1577.2265.9960.9973.9660.8953.66
ProImp66.6257.6647.5472.1761.7052.4965.2855.7746.2260.3649.2339.63
ICMVC76.8673.9267.3776.9373.3966.3079.2972.6766.2172.2166.6358.18
DIVIDE57.2947.0033.2368.1458.9344.1681.5771.4559.2074.3668.6260.14
Ours91.5785.3283.1488.8681.2678.6888.3680.0178.0283.6471.7968.06
ALOI-100CPM-Nets63.9579.1451.2652.3070.7439.7942.7462.4628.4030.5254.0317.18
COMPLETETER21.8646.9412.1022.0446.4511.6219.3844.8110.7817.1843.689.60
DIMVC56.9075.0035.5950.5572.5128.1052.5471.4532.5648.9770.8026.16
SURE64.5977.7452.6160.2274.7647.2160.6574.8447.3950.2368.9137.48
ProImp26.1552.9431.4024.5253.0728.1827.9758.1033.4220.3350.1024.28
ICMVC55.4874.0442.9148.8367.4636.5035.4256.9824.1926.8447.4316.63
DIVIDE76.0188.5970.5961.9080.5353.2952.5673.1740.5742.6667.8230.93
Ours91.1395.2287.5988.2893.3483.9584.5690.5378.9076.4485.0067.68
YouTubeFace10CPM-Nets75.4878.5666.5067.5075.8261.1566.6075.1962.0167.3274.1062.54
COMPLETETER46.7252.0421.0350.2452.8127.7736.4135.2415.1840.9839.9220.82
DIMVC81.7781.3274.5971.9676.4262.2973.8775.4463.9774.9479.3468.96
SURE26.6523.6011.4126.4322.4010.5531.8428.7715.2729.3727.4812.85
ProImp48.9753.2033.8343.3049.0426.5339.6750.1826.6849.5251.4729.12
ICMVC68.9074.1760.0662.7665.4453.0344.5652.5132.5127.4427.0915.33
DIVIDE31.8529.6411.3535.8630.8112.3132.7430.8212.7435.1733.4715.93
Ours82.9383.5574.7680.7781.4671.6280.1981.0771.3776.6281.3173.22
NoisyMNISTCPM-Nets48.6444.0832.0042.8837.0126.9250.5243.5433.5646.4637.1827.04
COMPLETETER76.5587.9577.1878.1686.1176.8364.7965.7754.6841.1540.6027.58
DIMVC55.8854.7841.2956.7758.2943.5162.2862.6648.0165.9467.4354.58
SURE98.6195.8996.9697.1193.7591.6494.8589.0687.8984.0078.2576.02
ProImp98.3095.3196.3593.2986.0585.3878.1669.4564.3750.0441.3333.34
ICMVC98.7896.3697.3597.7593.7195.1181.6479.2475.4753.9150.9144.56
DIVIDE94.7291.4489.4395.8589.7191.0657.0957.2946.2828.5725.6111.14
Ours99.1397.2398.1097.6893.9494.9496.0489.8191.4892.1982.5083.56
" + }, + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.484, + 0.891, + 0.499 + ], + "angle": 0, + "content": "Table 3. Ablation study on Caltech-5V and ALOI-100. \\( \\checkmark \\) denotes FreeCSL with the component and the best results are highlighted in red." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.503, + 0.892, + 0.627 + ], + "angle": 0, + "content": "
Componentsr=0.1r=0.3r=0.5r=0.7
LrecLccLgcACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)
Caltech-5V✓75.0065.6458.8969.0057.8451.4952.2144.1836.7151.9343.4235.75
✓✓85.5078.0772.2882.0074.5367.6785.3674.3271.1079.1467.8162.78
✓✓81.0770.4566.1380.8669.6167.4974.1459.3955.9070.7153.2549.95
✓✓✓91.5785.3283.1488.8681.2678.6888.3680.0178.0283.6471.7968.06
ALOI-100✓63.8177.9249.3844.6965.6431.1132.2055.8318.9423.7848.9011.31
✓✓87.6993.6484.1285.2591.0279.4580.2388.0474.1575.2983.3865.72
✓✓81.0685.9671.2164.1375.8751.4947.5265.5934.2932.7356.3519.23
✓✓✓91.1395.2287.5988.2893.3483.9584.5690.5378.9076.4485.0067.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.64, + 0.893, + 0.671 + ], + "angle": 0, + "content": "Table 4. Imputation- and alignment-free study on Caltech-5V and ALOI-100. ILR and ISR are filled with K-NN imputation via cross-view graph for latent representations and semantic representations \\(\\mathbf{Z}^{(v)}\\), \\(\\mathbf{H}^{(v)}\\). The best results are highlighted in red." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.674, + 0.888, + 0.78 + ], + "angle": 0, + "content": "
Missing ratesr=0.1r=0.3r=0.5r=0.7
MetricsACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)
Caltech-5VILR91.7185.7683.4489.1481.6979.0788.7980.7978.3777.8665.6455.12
ISR91.6485.5583.3089.4382.2679.6288.8680.7878.7481.2966.6863.59
FreeCSL91.5785.3283.1488.8681.2678.6888.3680.0178.0283.6471.7968.06
ALGI-100ILR58.9279.6741.5453.4675.8836.8945.8369.4727.4138.0858.9621.35
ISR90.1695.3887.8688.9194.0885.4882.5388.7371.5465.3677.8546.06
FreeCSL91.1395.2287.5988.2893.3483.9584.5690.5378.9076.4485.0067.68
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.804, + 0.471, + 0.895 + ], + "angle": 0, + "content": "labels. Table 4 shows FreeBSD's robustness advantage at high \\( r \\) on Caltech-5V, with notable performance disparity on ALOI-100. ILR, lacking consensus semantic learning, struggles with multi-cluster tasks due to cross-view semantic gaps. ISR avoids semantic confusion via consensus semantic learning but underperforms FreeBSD due to biased" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.804, + 0.892, + 0.864 + ], + "angle": 0, + "content": "imputation as consistency information decreases at higher \\( r \\). FreeCSL, by contrast, integrates consistency and complementary information in consensus semantic representations for confident decisions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.903 + ], + "angle": 0, + "content": "To further illustrate how consensus semantic learning \"free up\" imputation and alignment, we construct cosine" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5077" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "similarity matrices on ALOI-100 for view-specific latent representations \\(\\{\\mathbf{Z}^v\\}_{v = 1}^4\\) and their consensus \\(\\mathbf{Z}\\), as well as view-specific semantic representations \\(\\{\\mathbf{H}^v\\}_{v = 1}^4\\) and their consensus \\(\\mathbf{H}\\) in Fig. 5, then analyze their information entropy. The similarity matrices of \\(\\{\\mathbf{Z}^v\\}_{v = 1}^4\\) are chaotic, with high uncertainty in intra- and inter- cluster relationships. The fusion manner \\(\\mathbb{T}(\\cdot)\\) with Eq. (2) alleviates high entropy by incorporating view-specific complementary information. Compared to Fig. 4, the similarity matrices of \\(\\{\\mathbf{H}^v\\}_{v = 1}^4\\) shows a clear block structure, with high intracluster similarity and distinct inter-cluster differences. This confirms \\(\\mathbf{H}\\), enhanced by view-specific information, reduces cross-view semantic gaps, while consensus semantic learning, capturing cluster semantic relationship, promotes high-confidence assignments as stated in Theorem 1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.329, + 0.276, + 0.344 + ], + "angle": 0, + "content": "3.5. Analysis on FreeCSL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.534 + ], + "angle": 0, + "content": "Convergence and Robustness Analysis. In Fig. 6, we plot metrics and losses over training iterations, with error bands from 5 random Caltech-5V experiments to assess robustness. Both converge to stable values with minimal fluctuations and reach stability simultaneously. The gradual decrease in reconstruction loss indicates that the CSL and CSE modules are reasonable and beneficial, as they don't introduce significant discrepancies between semantic and latent representations, further underscoring the benefits of pre-training. Thanks to the harmonious collaboration of the three modules, our FreeCSL achieved satisfactory results within 30 iterations at minimal computational cost." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.762 + ], + "angle": 0, + "content": "Parameter Sensitivity Analysis. Two hyperparameters in FreeCSL warrant investigation, namely, graph neighbors \\(\\zeta\\) in Eq. (9) and regularizer coefficient \\(\\lambda\\) in Eq. (11). We expanded \\(\\zeta\\) and \\(\\lambda\\) to the range of 0.05 to 0.5 and 3 to 32 respectively. In Fig. 7, our model is insensitive to \\(\\zeta\\) but more affected by variations in \\(\\lambda\\). This is because the robust regularizer leverages semantic knowledge learned from the CSL module to guide the CSE model toward faster and more stable learning. A small \\(\\lambda\\) cannot provide effective consistency constraints while a large one makes CSL to dominate the learn direction, hindering CSE from exploring cluster structure information. Therefore, we set \\(\\lambda \\in [0.05, 0.2]\\) to balance cooperative relationship between CSE and CSL. Without sacrificing ACC and NMI, we prefer to set the minimum \\(\\zeta = 3\\) to reduce computational complexity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.465, + 0.787 + ], + "angle": 0, + "content": "3.6. Visualization of Consensus Semantic Clusters" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To further examine the quality of cluster structures formed by consensus semantic representations, t-SNE visualization based on true labels are plotted in Fig. 8 and clearly show minimal misclassification and clusters with strong intracluster cohesion while distinct inter-cluster separation. This indicates FreeCSL nearly recovers true cluster structures by learning consensus semantic information from all data." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.875, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.159, + 0.875, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.239, + 0.875, + 0.253 + ], + "angle": 0, + "content": "Figure 6. Convergence analysis on Caltech-5V with \\( r = 0.5 \\)." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.267, + 0.692, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.362, + 0.645, + 0.374 + ], + "angle": 0, + "content": "(a) Caltech-5V" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.267, + 0.878, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.362, + 0.828, + 0.374 + ], + "angle": 0, + "content": "(b) ALOI-100" + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.38, + 0.86, + 0.395 + ], + "angle": 0, + "content": "Figure 7. Parameter analyses for \\(\\zeta\\) and \\(\\lambda\\) with \\(r = 0.5\\)." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.412, + 0.677, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.515, + 0.675, + 0.527 + ], + "angle": 0, + "content": "(a) Pre-training \\((\\mathrm{NMI} = 44.18\\%)\\)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.411, + 0.891, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.515, + 0.878, + 0.527 + ], + "angle": 0, + "content": "(b) Training \\((\\mathrm{NMI} = 79.61\\%)\\)" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.535, + 0.678, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.639, + 0.675, + 0.652 + ], + "angle": 0, + "content": "(c) Pre-training \\((\\mathrm{NMI} = 33.74\\%)\\)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.534, + 0.887, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.639, + 0.875, + 0.652 + ], + "angle": 0, + "content": "(d) Training \\((\\mathrm{NMI} = 89.81\\%)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.66, + 0.868, + 0.674 + ], + "angle": 0, + "content": "Figure 8. Visualization for Caltech-5V and NoisyMNIST." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.687, + 0.619, + 0.702 + ], + "angle": 0, + "content": "4. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.893 + ], + "angle": 0, + "content": "In this paper, we propose FreeBSD, a novel semantic learning paradigm free from imputation and alignment compared to existing consistency learning. We design prototype-based contrastive clustering to discover a shared semantic space, where observations converge toward their respective semantic prototype and are encoded as consensus semantics representations for clustering. Furthermore, we employ modularity-inspired graph clustering to enrich semantic representation with view-specific cluster information. The effective synergy of consensus semantic learning and cluster semantic enhancement makes FreeBSD excel in most complex IMVC tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5078" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.236, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "This work is supported by the National Key R&D Program of China under Grant No.2022ZD0209103, the National Natural Science Foundation of China (project no. 62325604, 62276271, 62406329, 62476281, 62441618), National Natural Science Foundation of China Joint Found under Grant No. U24A20323." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.21, + 0.175, + 0.226 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.235, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[1] Ulrik Brandes, Daniel Delling, Marco Gaertler, Robert Görke, Martin Hoefer, Zoran Nikoloski, and Dorothea Wagner. Maximizing modularity is hard. arXiv preprint physics/0608255, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.472, + 0.359 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.47, + 0.429 + ], + "angle": 0, + "content": "[3] Chenhang Cui, Yazhou Ren, Jingyu Pu, Jiawei Li, Xiaorong Pu, Tianyi Wu, Yutao Shi, and Lifang He. A novel approach for effective multi-view clustering with information-theoretic perspective. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.47, + 0.498 + ], + "angle": 0, + "content": "[4] Zhibin Dong, Jiaqi Jin, Yuyang Xiao, Bin Xiao, Siwei Wang, Xinwang Liu, and En Zhu. Subgraph propagation and contrastive calibration for incomplete multiview data clustering. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.499, + 0.47, + 0.553 + ], + "angle": 0, + "content": "[5] Guowang Du, Lihua Zhou, Yudi Yang, Kevin Lu, and Lizhen Wang. Deep multiple auto-encoder-based multi-view clustering. Data Science and Engineering, 6(3):323-338, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.554, + 0.47, + 0.61 + ], + "angle": 0, + "content": "[6] Uno Fang, Man Li, Jianxin Li, Longxiang Gao, Tao Jia, and Yanchun Zhang. A comprehensive survey on multi-view clustering. IEEE Transactions on Knowledge and Data Engineering, 35(12):12350-12368, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.611, + 0.47, + 0.665 + ], + "angle": 0, + "content": "[7] Wei Feng, Guoshuai Sheng, Qianqian Wang, Quanxue Gao, Zhiqiang Tao, and Bo Dong. Partial multi-view clustering via self-supervised network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11988-11995, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.665, + 0.47, + 0.72 + ], + "angle": 0, + "content": "[8] Ruiming Guo, Mouxing Yang, Yijie Lin, Xi Peng, and Peng Hu. Robust contrastive multi-view clustering against dual noisy correspondence. Advances in Neural Information Processing Systems, 37:121401-121421, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.721, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[9] Changhao He, Hongyuan Zhu, Peng Hu, and Xi Peng. Robust variational contrastive learning for partially view-unaligned clustering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4167-4176, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.79, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[10] Junlin Hu, Jiwen Lu, and Yap-Peng Tan. Sharable and individual multi-view metric learning. IEEE transactions on pattern analysis and machine intelligence, 40(9):2281-2288, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Peng Hu, Liangli Zhen, Xi Peng, Hongyuan Zhu, Jie Lin, Xu Wang, and Dezhong Peng. Deep supervised multi-view learning with graph priors. IEEE Transactions on Image Processing, 33:123-133, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.235, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[12] Dong Huang, Chang-Dong Wang, and Jian-Huang Lai. Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity. IEEE Transactions on Knowledge and Data Engineering, 35(11):11388-11402, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[13] Jiabo Huang, Shaogang Gong, and Xiatian Zhu. Deep semantic clustering by partition confidence maximisation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8849-8858, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[14] Zhenyu Huang, Peng Hu, Joey Tianyi Zhou, Jiancheng Lv, and Xi Peng. Partially view-aligned clustering. Advances in Neural Information Processing Systems, 33:2892-2902, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[15] Yangbangyan Jiang, Qianqian Xu, Zhiyong Yang, Xiaochun Cao, and Qingming Huang. Dm2c: Deep mixed-modal clustering. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[16] Jiaqi Jin, Siwei Wang, Zhibin Dong, Xinwang Liu, and En Zhu. Deep incomplete multi-view clustering with cross-view partial sample and prototype alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11600-11609, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[17] Guanzhou Ke, Bo Wang, Xiaoli Wang, and Shengfeng He. Rethinking multi-view representation learning via distilled disentangling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26774-26783, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[18] Brian W Kernighan and Shen Lin. An efficient heuristic procedure for partitioning graphs. The Bell system technical journal, 49(2):291-307, 1970. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[19] Haobin Li, Yunfan Li, Mouxing Yang, Peng Hu, Dezhong Peng, and Xi Peng. Incomplete multi-view clustering via prototype-based imputation. arXiv preprint arXiv:2301.11045, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[20] Xingfeng Li, Yinghui Sun, Quansen Sun, Zhenwen Ren, and Yuan Sun. Cross-view graph matching guided anchor alignment for incomplete multi-view clustering. Information Fusion, 100:101941, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[21] Yingming Li, Ming Yang, and Zhongfei Zhang. A survey of multi-view representation learning. IEEE transactions on knowledge and data engineering, 31(10):1863-1883, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.894, + 0.731 + ], + "angle": 0, + "content": "[22] Yunfan Li, Peng Hu, Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8547-8555, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[23] Yijie Lin, Yuanbiao Gou, Zitao Liu, Boyun Li, Jiancheng Lv, and Xi Peng. Completer: Incomplete multi-view clustering via contrastive prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11174-11183, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Yijie Lin, Yuanbiao Gou, Xiaotian Liu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4): 4447-4461, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[25] Xinwang Liu, Xinzhong Zhu, Miaomiao Li, Lei Wang, Chang Tang, Jianping Yin, Dinggang Shen, Huaimin Wang," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5079" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "and Wen Gao. Late fusion incomplete multi-view clustering. IEEE transactions on pattern analysis and machine intelligence, 41(10):2410-2423, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.468, + 0.176 + ], + "angle": 0, + "content": "[26] Yiding Lu, Haobin Li, Yunfan Li, Yijie Lin, and Xi Peng. A survey on deep clustering: from the prior perspective. *Vicinagearth*, 1(1):4, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.468, + 0.244 + ], + "angle": 0, + "content": "[27] Yiding Lu, Yijie Lin, Mouxing Yang, Dezhong Peng, Peng Hu, and Xi Peng. Decoupled contrastive multi-view clustering with high-order random walks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14193-14201, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.468, + 0.286 + ], + "angle": 0, + "content": "[28] Mark EJ Newman. Modularity and community structure in networks. Proceedings of the national academy of sciences, 103(23):8577-8582, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.468, + 0.357 + ], + "angle": 0, + "content": "[29] Jingyu Pu, Chenhang Cui, Xinyue Chen, Yazhou Ren, Xiaorong Pu, Zhifeng Hao, S Yu Philip, and Lifang He. Adaptive feature imputation with latent graph for deep incomplete multi-view clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14633-14641, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.468, + 0.411 + ], + "angle": 0, + "content": "[30] Yuming Shen, Ziyi Shen, Menghan Wang, Jie Qin, Philip Torr, and Ling Shao. You never cluster alone. Advances in Neural Information Processing Systems, 34:27734-27746, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.468, + 0.468 + ], + "angle": 0, + "content": "[31] Yuan Sun, Yang Qin, Yongxiang Li, Dezhong Peng, Xi Peng, and Peng Hu. Robust multi-view clustering with noisy correspondence. IEEE Transactions on Knowledge and Data Engineering, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.468, + 0.538 + ], + "angle": 0, + "content": "[32] Huayi Tang and Yong Liu. Deep safe multi-view clustering: Reducing the risk of clustering performance degradation caused by view increase. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 202-211, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.468, + 0.58 + ], + "angle": 0, + "content": "[33] Jingjing Tang, Qingqing Yi, Saiji Fu, and Yingjie Tian. Incomplete multi-view learning: Review, analysis, and prospects. Applied Soft Computing, page 111278, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.468, + 0.649 + ], + "angle": 0, + "content": "[34] Xinhang Wan, Jiyuan Liu, Xinbiao Gan, Xinwang Liu, Siwei Wang, Yi Wen, Tianjiao Wan, and En Zhu. One-step multiview clustering with diverse representation. IEEE Transactions on Neural Networks and Learning Systems, pages 1-13, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[35] Xinhang Wan, Bin Xiao, Xinwang Liu, Jiyuan Liu, Weixuan Liang, and En Zhu. Fast continual multi-view clustering with incomplete views. IEEE Transactions on Image Processing, 33:2995-3008, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[36] Haoqing Wang, Xun Guo, Zhi-Hong Deng, and Yan Lu. Rethinking minimal sufficient representation in contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16041-16050, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.83 + ], + "angle": 0, + "content": "[37] Qianqian Wang, Zhengming Ding, Zhiqiang Tao, Quanxue Gao, and Yun Fu. Generative partial multi-view clustering with adaptive fusion and cycle consistency. IEEE Transactions on Image Processing, 30:1771-1783, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[38] Siwei Wang, Xinwang Liu, Suyuan Liu, Jiaqi Jin, Wenxuan Tu, Xinzhong Zhu, and En Zhu. Align then fusion: Generalized large-scale multi-view clustering with anchor matching correspondences. Advances in Neural Information Processing Systems, 35:5882-5895, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[39] Yiming Wang, Dongxia Chang, Zhiqiang Fu, Jie Wen, and Yao Zhao. Incomplete multi-view clustering via cross-view relation transfer. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[40] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[41] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487. PMLR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[42] Cai Xu, Ziyu Guan, Wei Zhao, Hongchang Wu, Yunfei Niu, and Beilei Ling. Adversarial incomplete multi-view clustering. In IJCAI, pages 3933-3939, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[43] Jie Xu, Chao Li, Yazhou Ren, Liang Peng, Yujie Mo, Xiaoshuang Shi, and Xiaofeng Zhu. Deep incomplete multiview clustering via mining cluster complementarity. In Proceedings of the AAAI conference on artificial intelligence, pages 8761-8769, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[44] Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiaofeng Zhu, and Lifang He. Multi-level feature learning for contrastive multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16051-16060, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[45] Jie Xu, Chao Li, Liang Peng, Yazhou Ren, Xiaoshuang Shi, Heng Tao Shen, and Xiaofeng Zhu. Adaptive feature projection with distribution alignment for deep incomplete multiview clustering. IEEE Transactions on Image Processing, 32:1354-1366, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[46] Xiaoqiang Yan, Zhixiang Jin, Fengshou Han, and Yangdong Ye. Differentiable information bottleneck for deterministic multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27435-27444, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[47] Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, and Xi Peng. Partially view-aligned representation learning with noise-robust contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1134–1143, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[48] Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, and Xi Peng. Partially view-aligned representation learning with noise-robust contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1134–1143, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[49] Mouxing Yang, Yunfan Li, Peng Hu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Robust multi-view clustering with incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1055-1069, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[50] Shengju Yu, Suyuan Liu, Siwei Wang, et al. Sparse low-rank multi-view subspace clustering with consensus anchors and unified bipartite graph. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[51] Shengju Yu, Siwei Wang, Zhibin Dong, et al. A nonparametric graph clustering framework for multi-view data. In Proceedings of the AAAI conference on artificial intelligence, pages 16558-16567, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5080" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[52] Pengxin Zeng, Mouxing Yang, Yiding Lu, Changqing Zhang, Peng Hu, and Xi Peng. Semantic invariant multiview clustering with fully incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2139-2150, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[53] Changqing Zhang, Yajie Cui, Zongbo Han, Joey Tianyi Zhou, Huazhu Fu, and Qinghua Hu. Deep partial multi-view learning. IEEE transactions on pattern analysis and machine intelligence, 44(5):2402-2415, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[54] Yi Zhang, Xinwang Liu, Siwei Wang, Jiyuan Liu, Sisi Dai, and En Zhu. One-stage incomplete multi-view clustering via late fusion. In Proceedings of the 29th ACM international conference on multimedia, pages 2717-2725, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.346 + ], + "angle": 0, + "content": "[55] Yi Zhang, Fengyu Tian, Chuan Ma, Miaomiao Li, Hengfu Yang, Zhe Liu, En Zhu, and Xinwang Liu. Regularized instance weighting multiview clustering via late fusion alignment. IEEE Transactions on Neural Networks and Learning Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[56] Huasong Zhong, Chong Chen, Zhongming Jin, and Xian-Sheng Hua. Deep robust clustering by contrastive learning. arXiv preprint arXiv:2008.03030, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[57] Lihua Zhou, Guowang Du, Kevin Lu, Lizheng Wang, and Jingwei Du. A survey and an empirical evaluation of multiview clustering approaches. ACM Computing Surveys, 56 (7):1-38, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "5081" + } + ] +] \ No newline at end of file diff --git a/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/380dd383-b00d-4370-9caa-3f9395391322_origin.pdf b/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/380dd383-b00d-4370-9caa-3f9395391322_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e6d9076f25b5607b49740478950075ea81b917c0 --- /dev/null +++ b/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/380dd383-b00d-4370-9caa-3f9395391322_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0f373b72e2959c95c38ffa90870a49b56deb4ba9122b57258a29250dc29698f +size 4248598 diff --git a/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/full.md b/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7e45ed2292e378aeb5f8403c6cbc38b1033cfa8b --- /dev/null +++ b/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/full.md @@ -0,0 +1,443 @@ +# Imputation-free and Alignment-free: Incomplete Multi-view Clustering Driven by Consensus Semantic Learning + +Yuzhuo Dai $^{1}$ , Jiaqi Jin $^{1}$ , Zhibin Dong $^{1}$ , Siwei Wang $^{2,*}$ , Xinwang Liu $^{1,*}$ , En Zhu $^{1,*}$ , Xihong Yang $^{1}$ , Xinbiao Gan $^{1}$ , Yu Feng $^{1}$ + +$^{1}$ National University of Defense Technology, Changsha, China + +$^{2}$ Intelligent Game and Decision Lab, Beijing, China + +{yzdai24, wangsiwei13}@nudt.edu.cn + +# Abstract + +In incomplete multi-view clustering (IMVC), missing data induce prototype shifts within views and semantic inconsistencies across views. A feasible solution is to explore cross-view consistency in paired complete observations, further imputing and aligning the similarity relationships inherently shared across views. Nevertheless, existing methods are constrained by two-tiered limitations: (1) Neither instance- nor cluster-level consistency learning construct a semantic space shared across views to learn consensus semantics. The former enforces cross-view instances alignment, and wrongly regards unpaired observations with semantic consistency as negative pairs; the latter focuses on cross-view cluster counterparts while coarsely handling fine-grained intra-cluster relationships within views. (2) Excessive reliance on consistency results in unreliable imputation and alignment without incorporating view-specific cluster information. Thus, we propose an IMVC framework, imputation- and alignment-free for consensus semantics learning (FreeCSL). To bridge semantic gaps across all observations, we learn consensus prototypes from available data to discover a shared space, where semantically similar observations are pulled closer for consensus semantics learning. To capture semantic relationships within specific views, we design a heuristic graph clustering based on modularity to recover cluster structure with intra-cluster compactness and inter-cluster separation for cluster semantics enhancement. Extensive experiments demonstrate, compared to state-of-the-art competitors, FreeCSL achieves more confident and robust assignments on IMVC task. + +# 1. Introduction + +Thanks to representation learning enhanced by data observed from different perspectives, multi-view clustering + +![](images/bea63629f3bb2620e981da766fc86802231dca321c1d7f9cda640ac0e6fe1dc4.jpg) +(a) Consistency Learning Paradigms +Figure 1. Research Motivation for Consensus Semantic Learning. (a) two existing paradigms, instance- and cluster-level, either treat cross-view unpaired observations with similar semantics as false negatives or neglect to treat within-view observations with similar semantics as true positives; (b) we propose a novel semantic-level paradigm based on contrastive clustering with a set of consensus prototypes to foster semantic consistency across all view data. + +![](images/c8ea22c0cc0aadeabdc2d019f4a15670337047bdc505b9147defac3acb1cf5b5.jpg) + +![](images/0ce666c53306ccfcf037916107d8bdd9aefda76c0681a3391b7f5b6fd854c981.jpg) +(b) Prototype-based Consensus Learning + +(MVC) has achieved significant breakthroughs in the field of unsupervised learning [3, 6, 11, 17, 21, 34, 46, 57]. However, in practical applications, the assumption of data completeness is often difficult to satisfy that incomplete multiview clustering (IMVC) is introduced [7, 25, 29, 33, 35, 37, 54]. In IMVC, missing data causes prototypes shifts within views and semantic misalignment across views, due to the discrepancy between the distributions of complete and incomplete instances [14, 19]. More and more studies [4, 16] have noted that variations in complete instances across different views further exacerbate prototype misalignment. It is challenging to achieve semantic consistency on cluster assignments across all view data. + +To alleviate prototype shifts and misalignment while promoting semantic consensus in cluster assignments, existing IMVC methods explore consistency information from complete instances for imputation and alignment [20, 24, 32]. Unfortunately, they still suffer from several significant drawbacks in practical applications. In terms of consistency learning, as shown in Fig. 1 (a), one widely used paradigm is instance-level [13, 36], which pulls paired ob + +observations (the same instance in different views) closer in the representation space by enforcing highly similar representations, but may inadvertently discard view-specific information [26, 31, 48]. More importantly, it tends to introduce false-negative noise, where unpaired observations with the same semantics are incorrectly treated as negative pairs [2, 8, 27]. [49],[52] and [27] attempted to optimize this issue by proposing a cluster-level paradigm [47] that encourages observations to find their cluster counterparts across different views. It learns a cluster space shared across views but not applicable within specific views, as there are no clustering interactions among intra-view observations. + +Since the above consistency learning paradigms fail to account for semantic consistency across all view data, they cannot mitigate semantic gaps among intra-cluster observations [22, 30, 56]. Therefore, imputation of missing data is required to restore the original data distribution, including neighborhood-based recovery via cross-view graph structure transfer [39], adversarial generation or contrastive prediction through cross-view mutual information interaction [15, 37, 42, 53], and prototype-based imputation via cross-view sample-prototype relationship inheritance [19]. Meanwhile, cross-view alignment of assignments [22, 32], prototypes [16, 19, 38], or distributions [4, 45] is also a crucial approach for further enhancing consistent learning. More related works are enumerated in Appendix A. Both imputation and alignment are limited by the consistent information from cross-view paired data and cannot fully exploit within-view unpaired data to mine view-specific cluster information, i.e., within-cluster compactness and between-cluster separation [9, 19]. Particularly, once the amount of missing data is too excessive to provide sufficient consistent information, model performance may even decline rapidly due to the noise introduced by improper imputation and alignment. + +Realizing the above issues, we ponder: imputation and alignment aim to restore the similarity relationships inherited from other complete views for the clustering task. Can we directly bridge the semantic gaps while exploring the semantic relationships among all data in consistency learning, thereby avoiding imputation and alignment operations with uncertainty noise? Thus, we propose an IMVC framework, a novel semantic-level paradigm, as shown in Fig. 1 (b), that is driven by imputation- and alignment-free consensus semantic learning (FreeCSL). Notably, our proposed consensus learning involves concurrent interaction among all data, rather than being limited to within or across views. To bridge semantic gaps among view observations and learn cluster semantics information, FreeCSL employs contrastive clustering based on consensus prototypes to discover a shared semantic space, where observations converge toward their semantic prototypes respectively. In practice, we set missing statistical weights for observations to facilitate view collaboration in consensus representations that in + +tegrate detailed information of all views while adapting the impact of different views with different missing instances. Based on consensus representation, our model can construct robust consensus prototypes for semantic-level clustering without imputation and alignment. To discover view-specific semantic relationships and further enrich consensus representations, FreeCSL exploit graph clustering to capture the cluster structure for each view, which maximizes graph modularity within views to enhance intra-cluster connections and reduce inter-cluster interactions. In short, our model encodes data correlation, discovered by semantic learning, into a shared space to obtain consensus semantic representations for instance clustering. Our prominent contributions can be summarized as follows: + +- In terms of bridging semantic gaps, we design the consensus semantic learning module, a novel semantic-level paradigm based prototypical contrastive clustering, to discover a common semantic space where all observations are embedded as representations with consistent semantics, avoiding additional imputation and alignment. +- In terms of exploring semantic relationships, we employ the cluster semantics enhancement module, a heuristic graph clustering method with modularity-based learning objective, to mine the inherent cluster structures that reveal the semantic correlations within views. +- Extensive experiments show our model surpasses state-of-the-art (SOTA) competitors in complex tasks with high missing rates, multiple clusters and large-scale data. + +# 2. Method + +In this section, FreeCSL, a deep IMVC method without imputation or alignment, is proposed to learn consensus semantic representations for clustering. The framework in Fig. 2 coordinates reconstruction (REC) module, cross-view consensus semantic learning (CSL) module, and within-view cluster semantic enhancement (CSE) module. + +# 2.1. Problem Statement + +Notations. Given a multi-view dataset $\mathcal{X} = \{\mathbf{X}^v\in$ $\mathbb{R}^{N\times D_v}\}_{v = 1}^V$ with $N$ instances across $V$ views, $\tilde{\mathbf{X}}^v =$ $\{\tilde{\mathbf{x}}_i^v\in \mathbb{R}^{D_v}\}_{i = 1}^{N_v}$ is an incomplete subset of the $v$ -th view with $N_{v}$ observations, and $\overline{\mathbf{X}}^{m,n} = \{(\overline{\mathbf{x}}_i^m,\overline{\mathbf{x}}_i^n)\}_{i = 1}^{N_{mn}}$ is a pair-observed subset with $N_{mn}$ instances observed in both the $m$ -th and $n$ -th view. The task is to partition $N$ instances into $K$ clusters. + +Definition 1. Instance-level Consistency (IC): $\forall m \neq n$ , $\mathbf{x}_i^m$ and $\mathbf{x}_j^n$ are instance-level consistent across views if $i = j$ (they are cross-view observations of the same instance $\mathbf{x}$ ), expressed as $I(\mathbf{x}_i^m, \mathbf{x}_j^n) = 1$ and $O$ otherwise. + +Definition 2. Cluster-level Consistency (CC): $\forall m \neq n$ , $\mathbf{x}_i^m$ and $\mathbf{x}_j^n$ are cluster-level consistent across views if they belong to the same cluster $k$ , expressed as $C(\mathbf{x}_i^m, \mathbf{x}_j^n) = 1$ and 0 otherwise. + +Definition 3. Semantic-level Consensus (SC): $\forall m$ and $n$ , $\mathbf{x}_i^m$ and $\mathbf{x}_j^n$ achieve semantic-level consensus in MVC task if all observations share a set of cluster prototypes $\mathbf{C} = \{\mathbf{c}_k\}_{k=1}^K$ and $\arg \max_k \rho(\mathbf{x}_i^m, \mathbf{c}_k) = \arg \max_k \rho(\mathbf{x}_j^n, \mathbf{c}_k)$ , expressed as $S(\mathbf{x}_i^m, \mathbf{x}_j^n) = 1$ and $0$ otherwise. + +Theorem 1. Consensus semantic learning yields more confident and robust cluster assignments than instance- and cluster-level paradigms. (Proof is provided in Appendix B.) + +Theorem 2. Since paired observations $(\overline{\mathbf{x}}_i^m, \overline{\mathbf{x}}_i^n)$ inherently satisfy instance- and cluster-level consistency, they can achieve semantic consensus via a shared set of prototypes C. (Proof is provided in Appendix B.) + +# 2.2. Within-view Reconstruction + +To avoid clustering instability from similarity measures on manifold structures in high-dimensional spaces, we use autoencoders to encode view observations $\tilde{\mathbf{X}}^v$ into clustering-friendly low-dimensional representations $\tilde{\mathbf{Z}}^v$ . Considering multiple views are mostly heterogeneous and differently distributed, we provide an independent encoder-decoder $\{\mathcal{E}_v,\mathcal{D}_v\}_{v = 1}^V$ for each view. The encoder embeds the latent representation and the decoder recovers the original data from it, which jointly minimize reconstruction loss: + +$$ +\begin{array}{l} \mathcal {L} _ {r e c} = \sum_ {v = 1} ^ {V} \left\| \tilde {\mathbf {X}} ^ {v} - \mathcal {D} _ {v} \left(\tilde {\mathbf {Z}} ^ {v}\right) \right\| _ {F} ^ {2} \tag {1} \\ = \sum_ {v = 1} ^ {V} \sum_ {i = 1} ^ {N _ {v}} \| \tilde {\mathbf {x}} _ {i} ^ {v} - \mathcal {D} _ {v} \left(\mathcal {E} _ {v} \left(\tilde {\mathbf {x}} _ {i} ^ {v}\right)\right) \| _ {2} ^ {2}, \\ \end{array} +$$ + +where $\mathcal{E}_v(\tilde{\mathbf{X}}^v;\theta_v):\tilde{\mathbf{X}}^v\in \mathbb{R}^{N_v\times D_v}\to \tilde{\mathbf{Z}}^v\in \mathbb{R}^{N_v\times d}$ and $\mathcal{D}_v(\tilde{\mathbf{Z}}^v;\phi_v):\tilde{\mathbf{Z}}^v\in \mathbb{R}^{N_v\times d}\to \hat{\mathbf{X}}^v\in \mathbb{R}^{N_v\times D_v}$ . + +# 2.3. Cross-view Consensus Semantic Learning + +Based on Theorem 1, we design prototypical contrastive clustering for consensus semantic learning. To reach semantic-level consensus on assignments, consensus prototypes with all view information, are introduced into contrastive clustering to explore a shared semantic space, where all observations with similar semantics are pulled closer. + +Consensus Representations and Consensus Prototypes. Due to the assumptions of consistency and complementarity in MVC, inseparable clusters in one view will become linearly separable by introducing complementary information from other views [10, 43, 55]. We utilize a fusion manner, denoted as $\mathbb{T}(\cdot)$ , to map the latent representations $\mathbf{Z}^v$ into a linearly weighted representation space for consensus representations $\mathbf{Z} \in \mathbb{R}^{N \times d}$ that integrates consistency and complementary information from multiple views: + +$$ +\mathbf {Z} = \mathbb {T} \left(\left\{\mathbf {Z} ^ {v} \right\} _ {v = 1} ^ {V}\right) = \sum_ {v = 1} ^ {V} \mathbf {w} ^ {v} \mathbf {Z} ^ {v} = \left\{\sum_ {v = 1} ^ {V} w _ {i} ^ {v} \mathbf {z} _ {i} ^ {v} \right\} _ {i = 1} ^ {N}, \tag {2} +$$ + +where $w_{i}^{v}$ is the instance-level fine-grained fusion weight: + +$$ +w _ {i} ^ {v} = \frac {\mathbb {I} \left(\mathbf {z} _ {i} ^ {v} \neq N a N\right)}{\sum_ {v ^ {\prime}} ^ {V} \mathbb {I} \left(\mathbf {z} _ {i} ^ {v ^ {\prime}} \neq N a N\right)}, \tag {3} +$$ + +where $\mathbb{I}(\cdot)$ is the indicative function that takes 1 when $\mathbf{z}_i^v$ is the representation of $i$ -th instance $\mathbf{X}_i$ observed in view $v$ , and 0 otherwise. We set a completeness statistical weight $w_i^v$ based on the number of observations in $V$ views for instance $\mathbf{X}_i$ as the fusion weight, which makes use of view-specific complementary information without discarding unpaired observations and mitigating the negative impact of missing noise by adapting to the differences in missing instances across different views. + +We derive a set of consensus semantic prototypes $\mathbf{C} = \{\mathbf{c}_k\in \mathbb{R}^d\}_{k = 1}^K$ via $k$ -means on consensus representations $\mathbf{Z}$ . As "a representative embedding for semantically similar observations", $\mathbf{C}$ comprehensively captures the semantic information of all data [50] and is continuously refined throughout consensus representation learning. + +Prototype-based Contrastive Clustering. We perform semantic-level contrastive clustering in a shared vector space spanned by consensus prototypes, where the latent representation $\mathbf{z}_i^v$ encoded as the semantic representation $\mathbf{h}_i^v$ and $\mathbf{h}_i^v$ is assigned to the prototype $\mathbf{c}_k$ with the longest projected distance to obtain optimal assignments, in two steps: + +- Semantic similarity measure: Encode the semantic representation $\mathbf{h}_i^v$ and project it onto each prototype, then calculate the probability $p_{i,k}^{v}$ of belonging to cluster $k$ : + +$$ +p _ {i, k} ^ {v} = \frac {\exp \left(\mathbf {h} _ {i} ^ {v \top} \mathbf {c} _ {k} / \tau\right)}{\sum_ {k ^ {\prime}} ^ {K} \exp \left(\mathbf {h} _ {i} ^ {v \top} \mathbf {c} _ {k ^ {\prime}} / \tau\right)}, \tag {4} +$$ + +where $\tau$ is a temperature parameter [40]. $\mathbf{p}_i^v = \{p_{i,k}^v\}_{k = 1}^K$ is the soft assignments of observation $\mathbf{x}_i^v$ + +- Swapped knowledge distillation: Based on Theorem 2, $(\overline{\mathbf{x}}_i^m,\overline{\mathbf{x}}_i^n)$ share the same cluster semantics on consensus prototypes $\mathbf{C}$ , thus their true labels $\mathbf{y}_i^m,\mathbf{y}_i^n\in \mathbb{R}^K$ should be are distributionally consistent. To this end, "swapped" knowledge distillation (KD) utilizes pseudo-labels $\mathbf{q}_i^m$ , $\mathbf{q}_i^n$ as mutual supervised signals to prompt their cluster assignments $\mathbf{p}_i^m,\mathbf{p}_i^n$ as identical as possible: + +$$ +\ell_ {c c} ^ {m, n} = \ell_ {k d} \left(\mathbf {H} ^ {m}, \mathbf {Q} ^ {n}\right) + \ell_ {k d} \left(\mathbf {H} ^ {n}, \mathbf {Q} ^ {m}\right), \tag {5} +$$ + +where $\ell_{kd}(\mathbf{H}^m,\mathbf{Q}^n) = -\frac{1}{N_{nm}}\sum_{i = 1}^{N_{mn}}\sum_{k = 1}^{K}\mathbf{q}_i^n\log \mathbf{p}_i^m$ We extend Eq. (5) to more than two views, fostering greater view collaboration to enhance semantic learning: + +$$ +\mathcal {L} _ {c c} = \sum_ {m = 1} ^ {V} \sum_ {\substack {n = 1 \\ n \neq m}} ^ {V} \ell_ {c c} ^ {m, n}. \tag{6} +$$ + +![](images/982238dac916e28d963632dc3904412fd56538aa13f450a290c06bafe53add61.jpg) +Figure 2. The framework of FreeCSL. (a) Reconstruction module, encodes observations into clustering-friendly representations for each view; (b) Consensus semantic learning module, learns semantic representations through cross-view contrastive clustering in a shared semantic space, where paired observations are assigned to their nearest semantic prototype for consistent assignments. (c) Cluster semantic enhancement module, enriches semantic representations with cluster structure information through within-view graph clustering, which applies GCN to aggregate view-specific semantic information and maximizes spectral modularity to recover cluster structure with greater separation and lower entropy. Ultimately, perform $k$ -means on the consensus semantic representations to predict cluster labels. + +- Consensus label solution: To obtain pseudo-labels $\mathbf{Q}^m = \{\mathbf{q}_i^m\}_{i=1}^{N_{mn}}$ , project semantic representations $\mathbf{H}^m$ onto consensus prototypes $\mathbf{C}$ and maximize similarity between the representation $\mathbf{h}_i^m$ and its assigned prototype $\mathbf{c}_k$ : + +$$ +\mathbf {Q} ^ {*} = \arg \max _ {\mathbf {Q} ^ {m} \in \mathcal {Q}} \operatorname {T r} \left\{\left(\mathbf {C Q} ^ {m}\right) ^ {\top} \mathbf {H} ^ {m}\right) \} + \alpha \mathcal {R} (\mathbf {Q} ^ {m}), \tag {7} +$$ + +where $\mathcal{R}(\mathbf{Q}^m) = -\sum_i^{N_{mn}}\mathbf{q}_i^m\log \mathbf{q}_i^m$ , an entropy regularizer prevents trivial solutions via the smoothness $\alpha$ + +# 2.4. Within-view Cluster Semantic Enhancement + +Considering that graph pooling can discover data implicit correlation and reduce missing noise via aggregating the information of neighboring nodes, we design a graph clustering to enhance representations with cluster semantic information. It combines a heuristic learning objective based on modularity, an evaluation metric for cluster structure quality, to effectively recover the cluster structure[51]. + +Modularity. Modularity [28] quantifies the deviation between the actual cluster structure and the expected structure generated by random combinations in the null model. Modularity matrix $\mathbf{B}$ is defined as $\mathbf{B} = \mathbf{A} - \frac{\mathbf{d}\mathbf{d}^{\top}}{2m}$ , where $\mathbf{A}$ is the adjacency matrix, $\mathbf{d}$ is the degree vector and $m$ is the total number of edges in the graph. The entry $b_{ij} \in \mathbf{B}$ measures connection strength between nodes $i$ and $j$ . + +Modularity-based Graph clustering. Graph clustering relies on graph pooling and node assignment. For each view, we construct graph structures with $k$ -nearest neighbors (KNN), then deploy graph convolutional network (GCN) $\mathcal{G}_v(\cdot)$ to aggregate graph embeddings: + +$$ +\mathcal {G} _ {v} \left(\mathbf {Z} ^ {v}; \mathbf {A} ^ {v}\right) = \sigma \left(\mathbf {D} ^ {v - \frac {1}{2}} \mathbf {A} ^ {v} \mathbf {D} ^ {v - \frac {1}{2}} \mathbf {Z} ^ {v} \mathbf {W} ^ {v} + \mathbf {Z} ^ {v} \mathbf {W} _ {s} ^ {v}\right), \tag {8} +$$ + +where $\sigma (\cdot)$ is an activation function used to reinforce the nonlinear aggregation capability of GCN. We replace a skip connection matrix $\mathbf{W}_s^v$ with the self-loop $\mathbf{A}^v +\mathbf{I}$ to alleviate the over-smoothing of node embeddings. Graph pooling operates on the latent representations $\mathbf{Z}^v$ , while the adjacency matrix $\mathbf{A}^v$ is constructed from the original observations $\mathbf{X}^v$ : + +$$ +a _ {i, j} ^ {v} = \left\{ \begin{array}{l l} 1, & \text {i f} \left(\mathbf {x} _ {i} ^ {v} \text {a n d} \mathbf {x} _ {j} ^ {v} \neq N a N\right) \& \\ & \left(\mathbf {x} _ {i} ^ {v} \in \mathcal {N} _ {k} \left(\mathbf {x} _ {j} ^ {v}\right) \text {o r} \mathbf {x} _ {j} ^ {v} \in \mathcal {N} _ {k} \left(\mathbf {x} _ {i} ^ {v}\right)\right), \\ 0, & \text {o t h e r w i s e}, \end{array} \right. \tag {9} +$$ + +where $\mathcal{N}_k(\mathbf{x}_i^v)$ is a set including the $k$ -nearest neighbors of $\mathbf{x}_i^v$ . It brings neighbors closer in the embedding space without introducing missing noise into the graph structure. + +The softmax classifier offers the soft assignment $\mathbf{P}^v$ for node embeddings encoded by $\mathcal{G}_v(\mathbf{Z}^v;\mathbf{A}^v)$ : + +$$ +\mathbf {P} ^ {v} = \operatorname {S o f t m a x} \left(\mathcal {G} _ {v} \left(\mathbf {Z} ^ {v}; \mathbf {A} ^ {v}\right)\right). \tag {10} +$$ + +To obtain a well-separated cluster assignment $\mathbf{P}^v$ , we maximize the spectral modularity $\mathrm{Tr}\left((\mathbf{P}^v)^\top \mathbf{B}^v\mathbf{P}^v\right)$ [1, 18], which captures intra-cluster connections and intercluster margins by mapping $\mathbf{P}^v$ to modularity $\mathbf{B}^v$ ; To further improve the confidence and robustness of $\mathbf{P}^v$ , we apply a self-supervised signal $\mathbf{L}^v$ , learned from contrastive clustering as detailed in Sec. 2.3, to guide $\mathbf{P}^v$ in chasing $\mathbf{L}^v$ via self-knowledge distillation. Specifically, the following objective is designed to optimize $\mathbf{P}^v$ : + +$$ +\ell_ {m} ^ {v} \left(\mathbf {P} ^ {v}; \mathbf {L} ^ {v}\right) = - \frac {1}{2 m} \operatorname {T r} \left(\left(\mathbf {P} ^ {v}\right) ^ {\top} \mathbf {B} ^ {v} \mathbf {P} ^ {v}\right) + \lambda \operatorname {K L} \left(\mathbf {L} ^ {v} \| \mathbf {P} ^ {v}\right), \tag {11} +$$ + +where Kullback-Leibler divergence $\mathrm{KL}(\cdot)$ serves as a robust regularizer, leveraging $\lambda$ to regulate the information flow of $\mathbf{L}^v\in \mathbb{R}^{N\times K}$ to ensure stable cluster performance. We project semantic representations $\mathbf{H}^v$ learned through contrastive clustering onto their $k$ -means prototypes and introduce Student's t-distribution [41] as kernel to predict high-confidence labels $\mathbf{L}^v$ with a nearly uniform distribution: + +$$ +l _ {i, k} ^ {v} = \frac {\left(1 + \frac {\| \mathbf {h} _ {i} ^ {v} - \mathbf {c} _ {k} ^ {v} \| ^ {2}}{\gamma}\right) ^ {- \frac {\gamma + 1}{2}}}{\sum_ {k ^ {\prime}} \left(1 + \frac {\| \mathbf {h} _ {i} ^ {v} - \mathbf {c} _ {k ^ {\prime}} ^ {v} \| ^ {2}}{\gamma}\right) ^ {- \frac {\gamma + 1}{2}}} \in \mathbf {L} ^ {v}. \tag {12} +$$ + +We conduct graph clustering on each view separately, then optimize assignments across all views concurrently: + +$$ +\mathcal {L} _ {g c} = \sum_ {v = 1} ^ {V} \ell_ {m} ^ {v} \left(\mathbf {P} ^ {v}, \mathbf {L} ^ {v}\right). \tag {13} +$$ + +# 2.5. Clustering Driven by Consensus Semantics + +Our FreeCSL learns the semantic representations $\mathbf{H}^v$ end-to-end by jointly minimizing the reconstruction loss, contrastive clustering loss and graph clustering loss: + +$$ +\mathcal {L} = \mathcal {L} _ {r e c} + \mathcal {L} _ {c c} + \mathcal {L} _ {g c}. \tag {14} +$$ + +Without searching for the optimal balancing weights of three loss terms, outstanding performance is easily achieved by applying $k$ -means to the consensus semantic representations $\mathbf{H}$ learned via the fusion manner $\mathbb{T}\left(\{\mathbf{H}^v\}_{v=1}^V\right)$ . + +The implementation for FreeCSL consists of two stages: warm-up training for the encoder-decoder, followed by fine-tuning for semantic representation learning and clustering. For details, refer to Algorithm 1. + +# 3. Experiment + +# 3.1. Experimental Settings + +We select four representative datasets namely Caltech-5V[44], ALOI-100[5], YoutubeFace10[12] and NoisyMNIST[23] to demonstrate our model's performance + +Algorithm 1 FreeCSL for Learning and Clustering + +Input: Complete, incomplete and pair-observed multiview dataset $\{\mathbf{X}^v\}_{v = 1}^V$ $\{\tilde{\mathbf{X}}^v\}_{v = 1}^V$ $\{\overline{\mathbf{X}}^{m,n}\}_{m\neq n}^V$ networks $\{\mathcal{E}_v,\mathcal{D}_v,\mathcal{G}_v\}_{v = 1}^V$ ; warm-up and fine-tuning epochs $e,E$ + +for $t = 1$ to e do + +On $\{\tilde{\mathbf{X}}^v\}_{v = 1}^V$ , warming up $\{\mathcal{E}_v,\mathcal{D}_v\}_{v = 1}^V$ with Eq.(1). + +for $t = 1$ to $E$ do + +Obtain latent representations $\{\mathbf{Z}^v |\mathcal{E}_v:\mathbf{X}^v\to \mathbf{Z}^v\}_{v = 1}^V$ Construct consensus prototypes $\mathbf{C} = \{\mathbf{c}_k\}_{k = 1}^K$ as elaborated in Sec. 2.3. + +Compute the reconstruction loss $\mathcal{L}_{rec}$ with Eq.(1). + +// Cross-view Contrastive Clustering + +for $m,n = 1$ to $V(m\neq n)$ do + +Compute the cluster probability $\mathbf{p}_i^m,\mathbf{p}_i^n$ with Eq.(4); + +Solve consensus pseudo-label $\mathbf{q}_i^m$ with Eq.(7); + +Get swapped loss $\ell_{cc}^{m,n}$ with Eq.(5) on $\overline{\mathbf{X}}^{m,n}$ . + +Calculate contrastive clustering loss $\mathcal{L}_{cc}$ with Eq.(6) + +// Within-view Graph Clustering + +for $v = 1$ to $V$ do + +Construct adjacency matrix $\mathbf{A}^v$ with Eq.(9). + +Solve node assignment $\mathbf{P}^v$ with Eq.(10) + +Solve pseudo-labels $\mathbf{L}^v$ with Eq.(12); + +Compute KL-modularity loss $\ell_m^v$ with Eq.(11) + +Calculate graph clustering loss $\mathcal{L}_{gc}$ with Eq.(13) + +// Semantic Representation Learning + +Calculate the overall loss $\mathcal{L}$ with Eq.(14); + +Optimize $\{\mathcal{E}_v,\mathcal{D}_v,\mathcal{G}_v\}_{v = 1}^V$ to minimize $\mathcal{L}$ + +Learn latent and consensus representations $\{\mathbf{Z}^v\}_{v = 1}^V$ + +and $\mathbf{Z}$ , semantic representations $\{\mathbf{H}^v\}_{v=1}^V$ , consensus + +semantic representations $\mathbf{H}$ , consensus prototypes $\mathbf{C}$ . + +Output: $\hat{\mathbf{Y}} = \{\hat{y}_i\}_{i=1}^N$ predicted by $k$ -means on $\mathbf{H}$ . + +in comparison with seven SOTA methods summarized in Table 1. To comprehensively evaluate experimental results, three metrics are adopted: accuracy (ACC), normalized mutual information (NMI), and adjusted rand index (ARI). + +Table 1. SOTA methods categorized by the types of techniques for consistency, imputation, and alignment. + +
CompetitorsConsistencyImputationAlignment
CPM-Nets (TPAMI'20)instance-levelmutual information interaction
COMPLETER (CVPR'21)instance-levelmutual information interaction
DIMVC (AAAI'22)instance-level\assignment-based
SURE (TPAMI'23)cluster-levelgraph structure transfer
ProLmp (IICA'23)instance-levelsample-prototype relationship inheritanceprototype-based
ICMVC (AAAI'24)instance-levelgraph structure transferassignment-based
DIVIDE (AAAI'24)cluster-levelmutual information interaction
+ +# 3.2. Implementation details + +Models are trained on an NVIDIA 3090 GPU using Adam optimizer in PyTorch 2.1.0. The warming up and finetuning learning rates are 0.0003 and 0.0005, with a batch size of 512. The hyperparameters are set as follows for + +different datasets: smoothness $\alpha = 0.5$ , temperature $\tau \in \{0.1, 0.2\}$ , neighbors $\zeta = 3$ , and regularizer weight $\lambda \in \{0.05, 0.1, 0.2, 0.3\}$ . For all datasets, 4-layer autoencoders and 2-layer GCNs with same MLP structures are used for each view. The layers of contrastive clustering and graph clustering are shared across all views via a single FC layer. + +# 3.3. Competitiveness of FreeCSL + +FreeCSL is compared with seven SOTA methods across three metrics in Table 2. The results indicate that FreeCSL excellently handles challenges of high missing rates, multiple clusters, and large-scale issues in IMVC: + +- From the perspective of effectiveness, FreeBSD surpasses most SOTA models, particularly on challenging datasets like ALOI-100 (100 clusters) and YouTubeFace10 (30,000 samples). It achieves significant improvements in ACC, with gains of $15.12\%$ , $26.38\%$ , $23.91\%$ , $26.21\%$ for ALOI-100, and $1.16\%$ , $8.81\%$ , $6.32\%$ , $1.68\%$ for YouTubeFace10, at missing rate $r = 0.1$ , $0.3$ , $0.5$ , $0.7$ . +- From the perspective of robustness, Fig. 3 (visualized from Table 2) shows our model's stability as $r$ increases. FreeCSL declines gradually, unlike other models with sharp ACC drops, highlighting its robustness in IMVC task. On the small dataset Caltech-5V with $r = 0.7$ , its Acc remains at $84.56\%$ , while on the larger dataset NoisyMNIST, also achieves $92.19\%$ . + +The above phenomenon can be explained as follows: As $r$ increases, paired observations become scarce. Methods like instance-level consistency learning or cross-view imputation and alignment (e.g., CPM-Net, COMPLETER, ICMVC), constrained by their dependence on consistency information, will introduce biases such as false negatives and semantic inconsistencies, resulting in degraded performance. For the complex task ALOI-100, cluster-level methods like SURE and DIVIDE surpass instance-level by exploiting cross-view cluster consistency. But, their reliance on cross-view graphs or mutual information for imputation, while neglecting within-view cluster affiliations, restricts performance. ProImp acknowledges the need to integrate within-view and cross-view information for semantically consistent imputation. However, its prototype alignment accumulates errors, particularly with a large number of clusters, achieving only $20\%$ ACC on ALOI-100. + +The superiority of FreeCSL stems from the synergistic effect of the CSL and CSE modules, which eliminate semantic gaps and ensure consistent assignments for all observations in a shared semantic space. + +# 3.4. Understanding FreeCSL + +Ablation Study. As shown in Table 3, ablation studies on FreeCSL's three components reveal that the CSL module contributes the most. On ALOI-100, it demonstrates exceptional performance and strong stability, with ACC drop + +![](images/44a05a75f84b556842986767080c87927e6efd088f3c383155cc3e63a819a247.jpg) +(a) Caltech-5V + +![](images/85c9df279d182798f03f30987165d0a467e7d1e18294a69fd7a931481cba0a26.jpg) +(b) ALOI-100 + +![](images/1b84ce01963f7f0afc1aa5b276652d554cd1c87c0f719c30d7cb83dd5f3c2dc7.jpg) +(c) YoutubeFace10 + +![](images/5601e511d580dbfd6206497a847821e276a337df2824b369ad4151970a346a98.jpg) +(d) NoisyMNIST +Figure 3. Visualization for Table 2 based on metric ACC. + +ping only $12.4\%$ as $r$ rises from 0.1 to 0.7. This is attributed to CSL's establishment of a shared semantic space via prototype-based semantic contrast, reducing semantic gaps within clusters. The CSE module also plays a positive role in discovering tighter cluster structures by optimizing modularity. Thanks to its enhanced semantics, both the REC and CSL modules achieve notable improvements. + +![](images/aa3562998a24fb9c685041adb1552e65e604fac63fc3adb17144c7ffe5597d73.jpg) +(a) $\mathbf{Z}^{(1)}$ + +![](images/1190bef9176bc25e375f1a402aa26b8cd0f07c1dce21c97684398f3d64088821.jpg) +(b) $\mathbf{Z}^{(2)}$ + +![](images/be049ac7d7bc6a93db89c7810cc3d9e8e27367c31c872d04f860331de60a9e1d.jpg) +(c) $\mathbf{Z}^{(3)}$ + +![](images/a86e48e7bf1e44e8176ce1bdd11e0d5747ed4113a8f31bdecaaceb3d3d270140.jpg) +(d) $\mathbf{Z}^{(4)}$ + +![](images/da8924045196a8f8977ee07e95801095fe486f5a6a6c74bf5f90b34fe48dc386.jpg) +(e) Z + +![](images/3a549c7483c4570036923ec7730acc478c8dd1c092cf33f18f2a8d076fb27e8e.jpg) +Figure 4. Similarity matrices of $\{\mathbf{Z}^v\}_{v=1}^4$ , $\mathbf{Z}$ without consensus semantic learning on ALOI-100 with $r = 0.5$ . +(a) $\mathbf{H}^{(1)}$ +Figure 5. Similarity matrices of $\{\mathbf{H}^v\}_{v=1}^4$ , $\mathbf{H}$ with consensus semantic learning on ALOI-100 with $r = 0.5$ . + +![](images/ba7702b5c0b82ff079233308ebc3fb4d6a656d385f2cec563eae387b3752eef4.jpg) +(b) $\mathbf{H}^{(2)}$ + +![](images/be1186ed0f33521b45d1520fe5a77984b6479338f23eda584f699ce81140c2fb.jpg) +(c) $\mathbf{H}^{(3)}$ + +![](images/bf7a82ff40da53536f810e53225dc1e4c9d88a560faeec609abf1c63cdfd95b4.jpg) +(d) $\mathbf{H}^{(4)}$ + +![](images/ccfa1adbbe01951b9dc76e113e1d6d5ec3aa3901859807a138ad0a9d6e337f2c.jpg) +(e) H + +Imputation- and Alignment-free CSL. To verify FreeCSL can reduce semantic gaps and capture semantic relationships without requiring imputation or alignment, we set up two imputation experiments: one imputation for latent representations $\mathbf{Z}^v$ without consensus semantic learning named ILR, and the second for semantic representations $\mathbf{H}^v$ learned from consensus semantic learning named ISR. They transfer complete $k$ -nn graph structures from other views to incomplete views and utilize corresponding $k$ neighbors for imputation. The control groups perform $k$ -means on the sum of all view matrices to predict cluster + +Table 2. Performance comparison on four multi-view benchmarks. The best and second-best results are highlighted in red and blue. + +
Missing ratesr=0.1r=0.3r=0.5r=0.7
MetricsACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)
Caltech-5VCPM-Nets85.7274.6271.9285.4372.9070.8481.0868.6166.0879.5065.5262.70
COMPLETETER58.6867.6942.9865.0768.7547.1064.4666.0942.8069.0067.4651.49
DIMVC85.8680.0174.6378.9078.7669.5580.0170.0866.3569.8660.1648.70
SURE75.6469.1262.8377.8967.8961.1577.2265.9960.9973.9660.8953.66
ProImp66.6257.6647.5472.1761.7052.4965.2855.7746.2260.3649.2339.63
ICMVC76.8673.9267.3776.9373.3966.3079.2972.6766.2172.2166.6358.18
DIVIDE57.2947.0033.2368.1458.9344.1681.5771.4559.2074.3668.6260.14
Ours91.5785.3283.1488.8681.2678.6888.3680.0178.0283.6471.7968.06
ALOI-100CPM-Nets63.9579.1451.2652.3070.7439.7942.7462.4628.4030.5254.0317.18
COMPLETETER21.8646.9412.1022.0446.4511.6219.3844.8110.7817.1843.689.60
DIMVC56.9075.0035.5950.5572.5128.1052.5471.4532.5648.9770.8026.16
SURE64.5977.7452.6160.2274.7647.2160.6574.8447.3950.2368.9137.48
ProImp26.1552.9431.4024.5253.0728.1827.9758.1033.4220.3350.1024.28
ICMVC55.4874.0442.9148.8367.4636.5035.4256.9824.1926.8447.4316.63
DIVIDE76.0188.5970.5961.9080.5353.2952.5673.1740.5742.6667.8230.93
Ours91.1395.2287.5988.2893.3483.9584.5690.5378.9076.4485.0067.68
YouTubeFace10CPM-Nets75.4878.5666.5067.5075.8261.1566.6075.1962.0167.3274.1062.54
COMPLETETER46.7252.0421.0350.2452.8127.7736.4135.2415.1840.9839.9220.82
DIMVC81.7781.3274.5971.9676.4262.2973.8775.4463.9774.9479.3468.96
SURE26.6523.6011.4126.4322.4010.5531.8428.7715.2729.3727.4812.85
ProImp48.9753.2033.8343.3049.0426.5339.6750.1826.6849.5251.4729.12
ICMVC68.9074.1760.0662.7665.4453.0344.5652.5132.5127.4427.0915.33
DIVIDE31.8529.6411.3535.8630.8112.3132.7430.8212.7435.1733.4715.93
Ours82.9383.5574.7680.7781.4671.6280.1981.0771.3776.6281.3173.22
NoisyMNISTCPM-Nets48.6444.0832.0042.8837.0126.9250.5243.5433.5646.4637.1827.04
COMPLETETER76.5587.9577.1878.1686.1176.8364.7965.7754.6841.1540.6027.58
DIMVC55.8854.7841.2956.7758.2943.5162.2862.6648.0165.9467.4354.58
SURE98.6195.8996.9697.1193.7591.6494.8589.0687.8984.0078.2576.02
ProImp98.3095.3196.3593.2986.0585.3878.1669.4564.3750.0441.3333.34
ICMVC98.7896.3697.3597.7593.7195.1181.6479.2475.4753.9150.9144.56
DIVIDE94.7291.4489.4395.8589.7191.0657.0957.2946.2828.5725.6111.14
Ours99.1397.2398.1097.6893.9494.9496.0489.8191.4892.1982.5083.56
+ +Table 3. Ablation study on Caltech-5V and ALOI-100. $\checkmark$ denotes FreeCSL with the component and the best results are highlighted in red. + +
Componentsr=0.1r=0.3r=0.5r=0.7
LrecLccLgcACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)
Caltech-5V✓75.0065.6458.8969.0057.8451.4952.2144.1836.7151.9343.4235.75
✓✓85.5078.0772.2882.0074.5367.6785.3674.3271.1079.1467.8162.78
✓✓81.0770.4566.1380.8669.6167.4974.1459.3955.9070.7153.2549.95
✓✓✓91.5785.3283.1488.8681.2678.6888.3680.0178.0283.6471.7968.06
ALOI-100✓63.8177.9249.3844.6965.6431.1132.2055.8318.9423.7848.9011.31
✓✓87.6993.6484.1285.2591.0279.4580.2388.0474.1575.2983.3865.72
✓✓81.0685.9671.2164.1375.8751.4947.5265.5934.2932.7356.3519.23
✓✓✓91.1395.2287.5988.2893.3483.9584.5690.5378.9076.4485.0067.68
+ +Table 4. Imputation- and alignment-free study on Caltech-5V and ALOI-100. ILR and ISR are filled with K-NN imputation via cross-view graph for latent representations and semantic representations $\mathbf{Z}^{(v)}$ , $\mathbf{H}^{(v)}$ . The best results are highlighted in red. + +
Missing ratesr=0.1r=0.3r=0.5r=0.7
MetricsACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)
Caltech-5VILR91.7185.7683.4489.1481.6979.0788.7980.7978.3777.8665.6455.12
ISR91.6485.5583.3089.4382.2679.6288.8680.7878.7481.2966.6863.59
FreeCSL91.5785.3283.1488.8681.2678.6888.3680.0178.0283.6471.7968.06
ALGI-100ILR58.9279.6741.5453.4675.8836.8945.8369.4727.4138.0858.9621.35
ISR90.1695.3887.8688.9194.0885.4882.5388.7371.5465.3677.8546.06
FreeCSL91.1395.2287.5988.2893.3483.9584.5690.5378.9076.4485.0067.68
+ +labels. Table 4 shows FreeBSD's robustness advantage at high $r$ on Caltech-5V, with notable performance disparity on ALOI-100. ILR, lacking consensus semantic learning, struggles with multi-cluster tasks due to cross-view semantic gaps. ISR avoids semantic confusion via consensus semantic learning but underperforms FreeBSD due to biased + +imputation as consistency information decreases at higher $r$ . FreeCSL, by contrast, integrates consistency and complementary information in consensus semantic representations for confident decisions. + +To further illustrate how consensus semantic learning "free up" imputation and alignment, we construct cosine + +similarity matrices on ALOI-100 for view-specific latent representations $\{\mathbf{Z}^v\}_{v = 1}^4$ and their consensus $\mathbf{Z}$ , as well as view-specific semantic representations $\{\mathbf{H}^v\}_{v = 1}^4$ and their consensus $\mathbf{H}$ in Fig. 5, then analyze their information entropy. The similarity matrices of $\{\mathbf{Z}^v\}_{v = 1}^4$ are chaotic, with high uncertainty in intra- and inter- cluster relationships. The fusion manner $\mathbb{T}(\cdot)$ with Eq. (2) alleviates high entropy by incorporating view-specific complementary information. Compared to Fig. 4, the similarity matrices of $\{\mathbf{H}^v\}_{v = 1}^4$ shows a clear block structure, with high intracluster similarity and distinct inter-cluster differences. This confirms $\mathbf{H}$ , enhanced by view-specific information, reduces cross-view semantic gaps, while consensus semantic learning, capturing cluster semantic relationship, promotes high-confidence assignments as stated in Theorem 1. + +# 3.5. Analysis on FreeCSL + +Convergence and Robustness Analysis. In Fig. 6, we plot metrics and losses over training iterations, with error bands from 5 random Caltech-5V experiments to assess robustness. Both converge to stable values with minimal fluctuations and reach stability simultaneously. The gradual decrease in reconstruction loss indicates that the CSL and CSE modules are reasonable and beneficial, as they don't introduce significant discrepancies between semantic and latent representations, further underscoring the benefits of pre-training. Thanks to the harmonious collaboration of the three modules, our FreeCSL achieved satisfactory results within 30 iterations at minimal computational cost. + +Parameter Sensitivity Analysis. Two hyperparameters in FreeCSL warrant investigation, namely, graph neighbors $\zeta$ in Eq. (9) and regularizer coefficient $\lambda$ in Eq. (11). We expanded $\zeta$ and $\lambda$ to the range of 0.05 to 0.5 and 3 to 32 respectively. In Fig. 7, our model is insensitive to $\zeta$ but more affected by variations in $\lambda$ . This is because the robust regularizer leverages semantic knowledge learned from the CSL module to guide the CSE model toward faster and more stable learning. A small $\lambda$ cannot provide effective consistency constraints while a large one makes CSL to dominate the learn direction, hindering CSE from exploring cluster structure information. Therefore, we set $\lambda \in [0.05, 0.2]$ to balance cooperative relationship between CSE and CSL. Without sacrificing ACC and NMI, we prefer to set the minimum $\zeta = 3$ to reduce computational complexity. + +# 3.6. Visualization of Consensus Semantic Clusters + +To further examine the quality of cluster structures formed by consensus semantic representations, t-SNE visualization based on true labels are plotted in Fig. 8 and clearly show minimal misclassification and clusters with strong intracluster cohesion while distinct inter-cluster separation. This indicates FreeCSL nearly recovers true cluster structures by learning consensus semantic information from all data. + +![](images/02329fcf383e903b6352efbb8b9fd3ea543fc078ad344ffb05e1ba3122356704.jpg) + +![](images/2062e21a39ba687681e800e9a1b0c9e0812579a95be44574e4bea859edbec620.jpg) +Figure 6. Convergence analysis on Caltech-5V with $r = 0.5$ . + +![](images/8c77d24e6d6d40a3792fe17dc3a5c26fc5d74651030a5090cf99044f1c516b22.jpg) +(a) Caltech-5V + +![](images/e2747f88cb4ba3cc24c3cd73a34ae0d1b17a3d4e36003973fce81f362debab36.jpg) +(b) ALOI-100 + +![](images/a184b06c80493552a4a7dec02d99654557a186dec520c0d0c59456030409d828.jpg) +(a) Pre-training $(\mathrm{NMI} = 44.18\%)$ + +![](images/4864853a6c5b8ece8aea7eda096bea427ceb7de01d819a16b259cd0ce1295390.jpg) +Figure 7. Parameter analyses for $\zeta$ and $\lambda$ with $r = 0.5$ . +(b) Training $(\mathrm{NMI} = 79.61\%)$ + +![](images/9b7e775dda1051fb260d6174348c8a73dc5a5f9b93a95534842b4db00fbdc4f8.jpg) +(c) Pre-training $(\mathrm{NMI} = 33.74\%)$ + +![](images/dc926b38bf6e88ca2189db92557430db3e98a2292d60f720f102c51885195282.jpg) +(d) Training $(\mathrm{NMI} = 89.81\%)$ +Figure 8. Visualization for Caltech-5V and NoisyMNIST. + +# 4. Conclusion + +In this paper, we propose FreeBSD, a novel semantic learning paradigm free from imputation and alignment compared to existing consistency learning. We design prototype-based contrastive clustering to discover a shared semantic space, where observations converge toward their respective semantic prototype and are encoded as consensus semantics representations for clustering. Furthermore, we employ modularity-inspired graph clustering to enrich semantic representation with view-specific cluster information. The effective synergy of consensus semantic learning and cluster semantic enhancement makes FreeBSD excel in most complex IMVC tasks. + +# Acknowledgments + +This work is supported by the National Key R&D Program of China under Grant No.2022ZD0209103, the National Natural Science Foundation of China (project no. 62325604, 62276271, 62406329, 62476281, 62441618), National Natural Science Foundation of China Joint Found under Grant No. U24A20323. + +# References + +[1] Ulrik Brandes, Daniel Delling, Marco Gaertler, Robert Görke, Martin Hoefer, Zoran Nikoloski, and Dorothea Wagner. Maximizing modularity is hard. arXiv preprint physics/0608255, 2006. 5 +[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2 +[3] Chenhang Cui, Yazhou Ren, Jingyu Pu, Jiawei Li, Xiaorong Pu, Tianyi Wu, Yutao Shi, and Lifang He. A novel approach for effective multi-view clustering with information-theoretic perspective. Advances in Neural Information Processing Systems, 36, 2024. 1 +[4] Zhibin Dong, Jiaqi Jin, Yuyang Xiao, Bin Xiao, Siwei Wang, Xinwang Liu, and En Zhu. Subgraph propagation and contrastive calibration for incomplete multiview data clustering. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1, 2 +[5] Guowang Du, Lihua Zhou, Yudi Yang, Kevin Lu, and Lizhen Wang. Deep multiple auto-encoder-based multi-view clustering. Data Science and Engineering, 6(3):323-338, 2021. 5 +[6] Uno Fang, Man Li, Jianxin Li, Longxiang Gao, Tao Jia, and Yanchun Zhang. A comprehensive survey on multi-view clustering. IEEE Transactions on Knowledge and Data Engineering, 35(12):12350-12368, 2023. 1 +[7] Wei Feng, Guoshuai Sheng, Qianqian Wang, Quanxue Gao, Zhiqiang Tao, and Bo Dong. Partial multi-view clustering via self-supervised network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11988-11995, 2024. 1 +[8] Ruiming Guo, Mouxing Yang, Yijie Lin, Xi Peng, and Peng Hu. Robust contrastive multi-view clustering against dual noisy correspondence. Advances in Neural Information Processing Systems, 37:121401-121421, 2024. 2 +[9] Changhao He, Hongyuan Zhu, Peng Hu, and Xi Peng. Robust variational contrastive learning for partially view-unaligned clustering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4167-4176, 2024. 2 +[10] Junlin Hu, Jiwen Lu, and Yap-Peng Tan. Sharable and individual multi-view metric learning. IEEE transactions on pattern analysis and machine intelligence, 40(9):2281-2288, 2017. 3 +[11] Peng Hu, Liangli Zhen, Xi Peng, Hongyuan Zhu, Jie Lin, Xu Wang, and Dezhong Peng. Deep supervised multi-view learning with graph priors. IEEE Transactions on Image Processing, 33:123-133, 2023. 1 + +[12] Dong Huang, Chang-Dong Wang, and Jian-Huang Lai. Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity. IEEE Transactions on Knowledge and Data Engineering, 35(11):11388-11402, 2023. 5 +[13] Jiabo Huang, Shaogang Gong, and Xiatian Zhu. Deep semantic clustering by partition confidence maximisation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8849-8858, 2020. 1 +[14] Zhenyu Huang, Peng Hu, Joey Tianyi Zhou, Jiancheng Lv, and Xi Peng. Partially view-aligned clustering. Advances in Neural Information Processing Systems, 33:2892-2902, 2020. 1 +[15] Yangbangyan Jiang, Qianqian Xu, Zhiyong Yang, Xiaochun Cao, and Qingming Huang. Dm2c: Deep mixed-modal clustering. Advances in Neural Information Processing Systems, 32, 2019. 2 +[16] Jiaqi Jin, Siwei Wang, Zhibin Dong, Xinwang Liu, and En Zhu. Deep incomplete multi-view clustering with cross-view partial sample and prototype alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11600-11609, 2023. 1, 2 +[17] Guanzhou Ke, Bo Wang, Xiaoli Wang, and Shengfeng He. Rethinking multi-view representation learning via distilled disentangling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26774-26783, 2024. 1 +[18] Brian W Kernighan and Shen Lin. An efficient heuristic procedure for partitioning graphs. The Bell system technical journal, 49(2):291-307, 1970. 5 +[19] Haobin Li, Yunfan Li, Mouxing Yang, Peng Hu, Dezhong Peng, and Xi Peng. Incomplete multi-view clustering via prototype-based imputation. arXiv preprint arXiv:2301.11045, 2023. 1, 2 +[20] Xingfeng Li, Yinghui Sun, Quansen Sun, Zhenwen Ren, and Yuan Sun. Cross-view graph matching guided anchor alignment for incomplete multi-view clustering. Information Fusion, 100:101941, 2023. 1 +[21] Yingming Li, Ming Yang, and Zhongfei Zhang. A survey of multi-view representation learning. IEEE transactions on knowledge and data engineering, 31(10):1863-1883, 2018. 1 +[22] Yunfan Li, Peng Hu, Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8547-8555, 2021. 2 +[23] Yijie Lin, Yuanbiao Gou, Zitao Liu, Boyun Li, Jiancheng Lv, and Xi Peng. Completer: Incomplete multi-view clustering via contrastive prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11174-11183, 2021. 5 +[24] Yijie Lin, Yuanbiao Gou, Xiaotian Liu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4): 4447-4461, 2022. 1 +[25] Xinwang Liu, Xinzhong Zhu, Miaomiao Li, Lei Wang, Chang Tang, Jianping Yin, Dinggang Shen, Huaimin Wang, + +and Wen Gao. Late fusion incomplete multi-view clustering. IEEE transactions on pattern analysis and machine intelligence, 41(10):2410-2423, 2018. 1 +[26] Yiding Lu, Haobin Li, Yunfan Li, Yijie Lin, and Xi Peng. A survey on deep clustering: from the prior perspective. *Vicinagearth*, 1(1):4, 2024. 2 +[27] Yiding Lu, Yijie Lin, Mouxing Yang, Dezhong Peng, Peng Hu, and Xi Peng. Decoupled contrastive multi-view clustering with high-order random walks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14193-14201, 2024. 2 +[28] Mark EJ Newman. Modularity and community structure in networks. Proceedings of the national academy of sciences, 103(23):8577-8582, 2006. 4 +[29] Jingyu Pu, Chenhang Cui, Xinyue Chen, Yazhou Ren, Xiaorong Pu, Zhifeng Hao, S Yu Philip, and Lifang He. Adaptive feature imputation with latent graph for deep incomplete multi-view clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14633-14641, 2024. 1 +[30] Yuming Shen, Ziyi Shen, Menghan Wang, Jie Qin, Philip Torr, and Ling Shao. You never cluster alone. Advances in Neural Information Processing Systems, 34:27734-27746, 2021. 2 +[31] Yuan Sun, Yang Qin, Yongxiang Li, Dezhong Peng, Xi Peng, and Peng Hu. Robust multi-view clustering with noisy correspondence. IEEE Transactions on Knowledge and Data Engineering, 2024. 2 +[32] Huayi Tang and Yong Liu. Deep safe multi-view clustering: Reducing the risk of clustering performance degradation caused by view increase. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 202-211, 2022. 1, 2 +[33] Jingjing Tang, Qingqing Yi, Saiji Fu, and Yingjie Tian. Incomplete multi-view learning: Review, analysis, and prospects. Applied Soft Computing, page 111278, 2024. 1 +[34] Xinhang Wan, Jiyuan Liu, Xinbiao Gan, Xinwang Liu, Siwei Wang, Yi Wen, Tianjiao Wan, and En Zhu. One-step multiview clustering with diverse representation. IEEE Transactions on Neural Networks and Learning Systems, pages 1-13, 2024. 1 +[35] Xinhang Wan, Bin Xiao, Xinwang Liu, Jiyuan Liu, Weixuan Liang, and En Zhu. Fast continual multi-view clustering with incomplete views. IEEE Transactions on Image Processing, 33:2995-3008, 2024. 1 +[36] Haoqing Wang, Xun Guo, Zhi-Hong Deng, and Yan Lu. Rethinking minimal sufficient representation in contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16041-16050, 2022. 1 +[37] Qianqian Wang, Zhengming Ding, Zhiqiang Tao, Quanxue Gao, and Yun Fu. Generative partial multi-view clustering with adaptive fusion and cycle consistency. IEEE Transactions on Image Processing, 30:1771-1783, 2021. 1, 2 +[38] Siwei Wang, Xinwang Liu, Suyuan Liu, Jiaqi Jin, Wenxuan Tu, Xinzhong Zhu, and En Zhu. Align then fusion: Generalized large-scale multi-view clustering with anchor matching correspondences. Advances in Neural Information Processing Systems, 35:5882-5895, 2022. 2 + +[39] Yiming Wang, Dongxia Chang, Zhiqiang Fu, Jie Wen, and Yao Zhao. Incomplete multi-view clustering via cross-view relation transfer. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 2 +[40] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 3 +[41] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487. PMLR, 2016. 5 +[42] Cai Xu, Ziyu Guan, Wei Zhao, Hongchang Wu, Yunfei Niu, and Beilei Ling. Adversarial incomplete multi-view clustering. In IJCAI, pages 3933-3939, 2019. 2 +[43] Jie Xu, Chao Li, Yazhou Ren, Liang Peng, Yujie Mo, Xiaoshuang Shi, and Xiaofeng Zhu. Deep incomplete multiview clustering via mining cluster complementarity. In Proceedings of the AAAI conference on artificial intelligence, pages 8761-8769, 2022. 3 +[44] Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiaofeng Zhu, and Lifang He. Multi-level feature learning for contrastive multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16051-16060, 2022. 5 +[45] Jie Xu, Chao Li, Liang Peng, Yazhou Ren, Xiaoshuang Shi, Heng Tao Shen, and Xiaofeng Zhu. Adaptive feature projection with distribution alignment for deep incomplete multiview clustering. IEEE Transactions on Image Processing, 32:1354-1366, 2023. 2 +[46] Xiaoqiang Yan, Zhixiang Jin, Fengshou Han, and Yangdong Ye. Differentiable information bottleneck for deterministic multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27435-27444, 2024. 1 +[47] Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, and Xi Peng. Partially view-aligned representation learning with noise-robust contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1134–1143, 2021. 2 +[48] Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, and Xi Peng. Partially view-aligned representation learning with noise-robust contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1134–1143, 2021. 2 +[49] Mouxing Yang, Yunfan Li, Peng Hu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Robust multi-view clustering with incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1055-1069, 2022. 2 +[50] Shengju Yu, Suyuan Liu, Siwei Wang, et al. Sparse low-rank multi-view subspace clustering with consensus anchors and unified bipartite graph. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3 +[51] Shengju Yu, Siwei Wang, Zhibin Dong, et al. A nonparametric graph clustering framework for multi-view data. In Proceedings of the AAAI conference on artificial intelligence, pages 16558-16567, 2024. 4 + +[52] Pengxin Zeng, Mouxing Yang, Yiding Lu, Changqing Zhang, Peng Hu, and Xi Peng. Semantic invariant multiview clustering with fully incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2139-2150, 2023. 2 +[53] Changqing Zhang, Yajie Cui, Zongbo Han, Joey Tianyi Zhou, Huazhu Fu, and Qinghua Hu. Deep partial multi-view learning. IEEE transactions on pattern analysis and machine intelligence, 44(5):2402-2415, 2020. 2 +[54] Yi Zhang, Xinwang Liu, Siwei Wang, Jiyuan Liu, Sisi Dai, and En Zhu. One-stage incomplete multi-view clustering via late fusion. In Proceedings of the 29th ACM international conference on multimedia, pages 2717-2725, 2021. 1 +[55] Yi Zhang, Fengyu Tian, Chuan Ma, Miaomiao Li, Hengfu Yang, Zhe Liu, En Zhu, and Xinwang Liu. Regularized instance weighting multiview clustering via late fusion alignment. IEEE Transactions on Neural Networks and Learning Systems, 2024. 3 +[56] Huasong Zhong, Chong Chen, Zhongming Jin, and Xian-Sheng Hua. Deep robust clustering by contrastive learning. arXiv preprint arXiv:2008.03030, 2020. 2 +[57] Lihua Zhou, Guowang Du, Kevin Lu, Lizheng Wang, and Jingwei Du. A survey and an empirical evaluation of multiview clustering approaches. ACM Computing Surveys, 56 (7):1-38, 2024. 1 \ No newline at end of file diff --git a/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/images.zip b/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bfa282dea8daac8d7f328012009a64d35f95976d --- /dev/null +++ b/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2599b225cb6def9def336236f95c3830f9f6172b71f527ca80c3793aa91fe0b6 +size 886234 diff --git a/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/layout.json b/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d928d1104489dd502bf536312f03c65d8eca2dbd --- /dev/null +++ b/2025/Imputation-free and Alignment-free_ Incomplete Multi-view Clustering Driven by Consensus Semantic Learning/layout.json @@ -0,0 +1,14001 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 102, + 539, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 102, + 539, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 539, + 140 + ], + "type": "text", + "content": "Imputation-free and Alignment-free: Incomplete Multi-view Clustering Driven by Consensus Semantic Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "spans": [ + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "text", + "content": "Yuzhuo Dai" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Jiaqi Jin" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Zhibin Dong" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Siwei Wang" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Xinwang Liu" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "text", + "content": ", En Zhu" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Xihong Yang" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Xinbiao Gan" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Yu Feng" + }, + { + "bbox": [ + 101, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 190, + 447, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 190, + 447, + 203 + ], + "spans": [ + { + "bbox": [ + 146, + 190, + 447, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 190, + 447, + 203 + ], + "type": "text", + "content": "National University of Defense Technology, Changsha, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 171, + 203, + 422, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 203, + 422, + 217 + ], + "spans": [ + { + "bbox": [ + 171, + 203, + 422, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 171, + 203, + 422, + 217 + ], + "type": "text", + "content": "Intelligent Game and Decision Lab, Beijing, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 203, + 220, + 386, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 220, + 386, + 231 + ], + "spans": [ + { + "bbox": [ + 203, + 220, + 386, + 231 + ], + "type": "text", + "content": "{yzdai24, wangsiwei13}@nudt.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 258, + 191, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 191, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 191, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 283, + 289, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 289, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 289, + 632 + ], + "type": "text", + "content": "In incomplete multi-view clustering (IMVC), missing data induce prototype shifts within views and semantic inconsistencies across views. A feasible solution is to explore cross-view consistency in paired complete observations, further imputing and aligning the similarity relationships inherently shared across views. Nevertheless, existing methods are constrained by two-tiered limitations: (1) Neither instance- nor cluster-level consistency learning construct a semantic space shared across views to learn consensus semantics. The former enforces cross-view instances alignment, and wrongly regards unpaired observations with semantic consistency as negative pairs; the latter focuses on cross-view cluster counterparts while coarsely handling fine-grained intra-cluster relationships within views. (2) Excessive reliance on consistency results in unreliable imputation and alignment without incorporating view-specific cluster information. Thus, we propose an IMVC framework, imputation- and alignment-free for consensus semantics learning (FreeCSL). To bridge semantic gaps across all observations, we learn consensus prototypes from available data to discover a shared space, where semantically similar observations are pulled closer for consensus semantics learning. To capture semantic relationships within specific views, we design a heuristic graph clustering based on modularity to recover cluster structure with intra-cluster compactness and inter-cluster separation for cluster semantics enhancement. Extensive experiments demonstrate, compared to state-of-the-art competitors, FreeCSL achieves more confident and robust assignments on IMVC task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "content": "Thanks to representation learning enhanced by data observed from different perspectives, multi-view clustering" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 258, + 411, + 352 + ], + "blocks": [ + { + "bbox": [ + 308, + 258, + 411, + 352 + ], + "lines": [ + { + "bbox": [ + 308, + 258, + 411, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 411, + 352 + ], + "type": "image", + "image_path": "bea63629f3bb2620e981da766fc86802231dca321c1d7f9cda640ac0e6fe1dc4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 352, + 412, + 359 + ], + "lines": [ + { + "bbox": [ + 326, + 352, + 412, + 359 + ], + "spans": [ + { + "bbox": [ + 326, + 352, + 412, + 359 + ], + "type": "text", + "content": "(a) Consistency Learning Paradigms" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 365, + 545, + 443 + ], + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 443 + ], + "type": "text", + "content": "Figure 1. Research Motivation for Consensus Semantic Learning. (a) two existing paradigms, instance- and cluster-level, either treat cross-view unpaired observations with similar semantics as false negatives or neglect to treat within-view observations with similar semantics as true positives; (b) we propose a novel semantic-level paradigm based on contrastive clustering with a set of consensus prototypes to foster semantic consistency across all view data." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 411, + 258, + 429, + 352 + ], + "blocks": [ + { + "bbox": [ + 411, + 258, + 429, + 352 + ], + "lines": [ + { + "bbox": [ + 411, + 258, + 429, + 352 + ], + "spans": [ + { + "bbox": [ + 411, + 258, + 429, + 352 + ], + "type": "image", + "image_path": "c8ea22c0cc0aadeabdc2d019f4a15670337047bdc505b9147defac3acb1cf5b5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 432, + 258, + 545, + 352 + ], + "blocks": [ + { + "bbox": [ + 432, + 258, + 545, + 352 + ], + "lines": [ + { + "bbox": [ + 432, + 258, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 432, + 258, + 545, + 352 + ], + "type": "image", + "image_path": "0ce666c53306ccfcf037916107d8bdd9aefda76c0681a3391b7f5b6fd854c981.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 352, + 538, + 359 + ], + "lines": [ + { + "bbox": [ + 440, + 352, + 538, + 359 + ], + "spans": [ + { + "bbox": [ + 440, + 352, + 538, + 359 + ], + "type": "text", + "content": "(b) Prototype-based Consensus Learning" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 459, + 546, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 546, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 546, + 614 + ], + "type": "text", + "content": "(MVC) has achieved significant breakthroughs in the field of unsupervised learning [3, 6, 11, 17, 21, 34, 46, 57]. However, in practical applications, the assumption of data completeness is often difficult to satisfy that incomplete multiview clustering (IMVC) is introduced [7, 25, 29, 33, 35, 37, 54]. In IMVC, missing data causes prototypes shifts within views and semantic misalignment across views, due to the discrepancy between the distributions of complete and incomplete instances [14, 19]. More and more studies [4, 16] have noted that variations in complete instances across different views further exacerbate prototype misalignment. It is challenging to achieve semantic consistency on cluster assignments across all view data." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "To alleviate prototype shifts and misalignment while promoting semantic consensus in cluster assignments, existing IMVC methods explore consistency information from complete instances for imputation and alignment [20, 24, 32]. Unfortunately, they still suffer from several significant drawbacks in practical applications. In terms of consistency learning, as shown in Fig. 1 (a), one widely used paradigm is instance-level [13, 36], which pulls paired ob" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5071" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "observations (the same instance in different views) closer in the representation space by enforcing highly similar representations, but may inadvertently discard view-specific information [26, 31, 48]. More importantly, it tends to introduce false-negative noise, where unpaired observations with the same semantics are incorrectly treated as negative pairs [2, 8, 27]. [49],[52] and [27] attempted to optimize this issue by proposing a cluster-level paradigm [47] that encourages observations to find their cluster counterparts across different views. It learns a cluster space shared across views but not applicable within specific views, as there are no clustering interactions among intra-view observations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 219, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 289, + 483 + ], + "type": "text", + "content": "Since the above consistency learning paradigms fail to account for semantic consistency across all view data, they cannot mitigate semantic gaps among intra-cluster observations [22, 30, 56]. Therefore, imputation of missing data is required to restore the original data distribution, including neighborhood-based recovery via cross-view graph structure transfer [39], adversarial generation or contrastive prediction through cross-view mutual information interaction [15, 37, 42, 53], and prototype-based imputation via cross-view sample-prototype relationship inheritance [19]. Meanwhile, cross-view alignment of assignments [22, 32], prototypes [16, 19, 38], or distributions [4, 45] is also a crucial approach for further enhancing consistent learning. More related works are enumerated in Appendix A. Both imputation and alignment are limited by the consistent information from cross-view paired data and cannot fully exploit within-view unpaired data to mine view-specific cluster information, i.e., within-cluster compactness and between-cluster separation [9, 19]. Particularly, once the amount of missing data is too excessive to provide sufficient consistent information, model performance may even decline rapidly due to the noise introduced by improper imputation and alignment." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": "Realizing the above issues, we ponder: imputation and alignment aim to restore the similarity relationships inherited from other complete views for the clustering task. Can we directly bridge the semantic gaps while exploring the semantic relationships among all data in consistency learning, thereby avoiding imputation and alignment operations with uncertainty noise? Thus, we propose an IMVC framework, a novel semantic-level paradigm, as shown in Fig. 1 (b), that is driven by imputation- and alignment-free consensus semantic learning (FreeCSL). Notably, our proposed consensus learning involves concurrent interaction among all data, rather than being limited to within or across views. To bridge semantic gaps among view observations and learn cluster semantics information, FreeCSL employs contrastive clustering based on consensus prototypes to discover a shared semantic space, where observations converge toward their semantic prototypes respectively. In practice, we set missing statistical weights for observations to facilitate view collaboration in consensus representations that in" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "content": "tegrate detailed information of all views while adapting the impact of different views with different missing instances. Based on consensus representation, our model can construct robust consensus prototypes for semantic-level clustering without imputation and alignment. To discover view-specific semantic relationships and further enrich consensus representations, FreeCSL exploit graph clustering to capture the cluster structure for each view, which maximizes graph modularity within views to enhance intra-cluster connections and reduce inter-cluster interactions. In short, our model encodes data correlation, discovered by semantic learning, into a shared space to obtain consensus semantic representations for instance clustering. Our prominent contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 240, + 545, + 407 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 240, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 240, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 240, + 545, + 312 + ], + "type": "text", + "content": "- In terms of bridging semantic gaps, we design the consensus semantic learning module, a novel semantic-level paradigm based prototypical contrastive clustering, to discover a common semantic space where all observations are embedded as representations with consistent semantics, avoiding additional imputation and alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "type": "text", + "content": "- In terms of exploring semantic relationships, we employ the cluster semantics enhancement module, a heuristic graph clustering method with modularity-based learning objective, to mine the inherent cluster structures that reveal the semantic correlations within views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 373, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 545, + 407 + ], + "type": "text", + "content": "- Extensive experiments show our model surpasses state-of-the-art (SOTA) competitors in complex tasks with high missing rates, multiple clusters and large-scale data." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 417, + 361, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 417, + 361, + 429 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 361, + 429 + ], + "type": "text", + "content": "2. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "type": "text", + "content": "In this section, FreeCSL, a deep IMVC method without imputation or alignment, is proposed to learn consensus semantic representations for clustering. The framework in Fig. 2 coordinates reconstruction (REC) module, cross-view consensus semantic learning (CSL) module, and within-view cluster semantic enhancement (CSE) module." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 514, + 419, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 419, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 419, + 526 + ], + "type": "text", + "content": "2.1. Problem Statement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": "Notations. Given a multi-view dataset " + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\mathbf{X}^v\\in" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N\\times D_v}\\}_{v = 1}^V" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": " instances across " + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": " views, " + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}^v =" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{x}}_i^v\\in \\mathbb{R}^{D_v}\\}_{i = 1}^{N_v}" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": " is an incomplete subset of the " + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": " -th view with " + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": " observations, and " + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{X}}^{m,n} = \\{(\\overline{\\mathbf{x}}_i^m,\\overline{\\mathbf{x}}_i^n)\\}_{i = 1}^{N_{mn}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": " is a pair-observed subset with " + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "N_{mn}" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": " instances observed in both the " + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": " -th and " + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": " -th view. The task is to partition " + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": " instances into " + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 533, + 545, + 616 + ], + "type": "text", + "content": " clusters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "content": "Definition 1. Instance-level Consistency (IC): " + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\forall m \\neq n" + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^m" + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j^n" + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "content": " are instance-level consistent across views if " + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "content": " (they are cross-view observations of the same instance " + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "content": "), expressed as " + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "inline_equation", + "content": "I(\\mathbf{x}_i^m, \\mathbf{x}_j^n) = 1" + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "content": " otherwise." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "content": "Definition 2. Cluster-level Consistency (CC): " + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\forall m \\neq n" + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^m" + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j^n" + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "content": " are cluster-level consistent across views if they belong to the same cluster " + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "content": ", expressed as " + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "inline_equation", + "content": "C(\\mathbf{x}_i^m, \\mathbf{x}_j^n) = 1" + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "content": " and 0 otherwise." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5072" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "content": "Definition 3. Semantic-level Consensus (SC): " + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "inline_equation", + "content": "\\forall m" + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^m" + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j^n" + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "content": " achieve semantic-level consensus in MVC task if all observations share a set of cluster prototypes " + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\{\\mathbf{c}_k\\}_{k=1}^K" + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "inline_equation", + "content": "\\arg \\max_k \\rho(\\mathbf{x}_i^m, \\mathbf{c}_k) = \\arg \\max_k \\rho(\\mathbf{x}_j^n, \\mathbf{c}_k)" + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "content": ", expressed as " + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}_i^m, \\mathbf{x}_j^n) = 1" + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "inline_equation", + "content": "0" + }, + { + "bbox": [ + 47, + 72, + 287, + 137 + ], + "type": "text", + "content": " otherwise." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 142, + 288, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 142, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 45, + 142, + 288, + 178 + ], + "type": "text", + "content": "Theorem 1. Consensus semantic learning yields more confident and robust cluster assignments than instance- and cluster-level paradigms. (Proof is provided in Appendix B.)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 182, + 287, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 232 + ], + "type": "text", + "content": "Theorem 2. Since paired observations " + }, + { + "bbox": [ + 46, + 182, + 287, + 232 + ], + "type": "inline_equation", + "content": "(\\overline{\\mathbf{x}}_i^m, \\overline{\\mathbf{x}}_i^n)" + }, + { + "bbox": [ + 46, + 182, + 287, + 232 + ], + "type": "text", + "content": " inherently satisfy instance- and cluster-level consistency, they can achieve semantic consensus via a shared set of prototypes C. (Proof is provided in Appendix B.)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 243, + 201, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 201, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 201, + 255 + ], + "type": "text", + "content": "2.2. Within-view Reconstruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 262, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 287, + 369 + ], + "type": "text", + "content": "To avoid clustering instability from similarity measures on manifold structures in high-dimensional spaces, we use autoencoders to encode view observations " + }, + { + "bbox": [ + 47, + 262, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}^v" + }, + { + "bbox": [ + 47, + 262, + 287, + 369 + ], + "type": "text", + "content": " into clustering-friendly low-dimensional representations " + }, + { + "bbox": [ + 47, + 262, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Z}}^v" + }, + { + "bbox": [ + 47, + 262, + 287, + 369 + ], + "type": "text", + "content": ". Considering multiple views are mostly heterogeneous and differently distributed, we provide an independent encoder-decoder " + }, + { + "bbox": [ + 47, + 262, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}_v,\\mathcal{D}_v\\}_{v = 1}^V" + }, + { + "bbox": [ + 47, + 262, + 287, + 369 + ], + "type": "text", + "content": " for each view. The encoder embeds the latent representation and the decoder recovers the original data from it, which jointly minimize reconstruction loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 87, + 373, + 287, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 373, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 87, + 373, + 287, + 440 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {r e c} = \\sum_ {v = 1} ^ {V} \\left\\| \\tilde {\\mathbf {X}} ^ {v} - \\mathcal {D} _ {v} \\left(\\tilde {\\mathbf {Z}} ^ {v}\\right) \\right\\| _ {F} ^ {2} \\tag {1} \\\\ = \\sum_ {v = 1} ^ {V} \\sum_ {i = 1} ^ {N _ {v}} \\| \\tilde {\\mathbf {x}} _ {i} ^ {v} - \\mathcal {D} _ {v} \\left(\\mathcal {E} _ {v} \\left(\\tilde {\\mathbf {x}} _ {i} ^ {v}\\right)\\right) \\| _ {2} ^ {2}, \\\\ \\end{array}", + "image_path": "766c5923b5b608804fd5ac0b7bd6a517e74fc5149e17272e8ac3b554764338cd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 446, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 446, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 47, + 446, + 287, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 446, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_v(\\tilde{\\mathbf{X}}^v;\\theta_v):\\tilde{\\mathbf{X}}^v\\in \\mathbb{R}^{N_v\\times D_v}\\to \\tilde{\\mathbf{Z}}^v\\in \\mathbb{R}^{N_v\\times d}" + }, + { + "bbox": [ + 47, + 446, + 287, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 446, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v(\\tilde{\\mathbf{Z}}^v;\\phi_v):\\tilde{\\mathbf{Z}}^v\\in \\mathbb{R}^{N_v\\times d}\\to \\hat{\\mathbf{X}}^v\\in \\mathbb{R}^{N_v\\times D_v}" + }, + { + "bbox": [ + 47, + 446, + 287, + 473 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 479, + 265, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 265, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 265, + 493 + ], + "type": "text", + "content": "2.3. Cross-view Consensus Semantic Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "text", + "content": "Based on Theorem 1, we design prototypical contrastive clustering for consensus semantic learning. To reach semantic-level consensus on assignments, consensus prototypes with all view information, are introduced into contrastive clustering to explore a shared semantic space, where all observations with similar semantics are pulled closer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "content": "Consensus Representations and Consensus Prototypes. Due to the assumptions of consistency and complementarity in MVC, inseparable clusters in one view will become linearly separable by introducing complementary information from other views [10, 43, 55]. We utilize a fusion manner, denoted as " + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathbb{T}(\\cdot)" + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "content": ", to map the latent representations " + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^v" + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "content": " into a linearly weighted representation space for consensus representations " + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "content": " that integrates consistency and complementary information from multiple views:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 683, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} = \\mathbb {T} \\left(\\left\\{\\mathbf {Z} ^ {v} \\right\\} _ {v = 1} ^ {V}\\right) = \\sum_ {v = 1} ^ {V} \\mathbf {w} ^ {v} \\mathbf {Z} ^ {v} = \\left\\{\\sum_ {v = 1} ^ {V} w _ {i} ^ {v} \\mathbf {z} _ {i} ^ {v} \\right\\} _ {i = 1} ^ {N}, \\tag {2}", + "image_path": "214338308bcdf793d4427ec88439e7696564de46523515746e2d7f9e331ba35a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 539, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 539, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 539, + 85 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 72, + 539, + 85 + ], + "type": "inline_equation", + "content": "w_{i}^{v}" + }, + { + "bbox": [ + 306, + 72, + 539, + 85 + ], + "type": "text", + "content": " is the instance-level fine-grained fusion weight:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 367, + 93, + 545, + 123 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 93, + 545, + 123 + ], + "spans": [ + { + "bbox": [ + 367, + 93, + 545, + 123 + ], + "type": "interline_equation", + "content": "w _ {i} ^ {v} = \\frac {\\mathbb {I} \\left(\\mathbf {z} _ {i} ^ {v} \\neq N a N\\right)}{\\sum_ {v ^ {\\prime}} ^ {V} \\mathbb {I} \\left(\\mathbf {z} _ {i} ^ {v ^ {\\prime}} \\neq N a N\\right)}, \\tag {3}", + "image_path": "841e9973781ee35c103592ac1388a0bd63bab8fc5e8625b9ca494c5bd23813f7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "text", + "content": " is the indicative function that takes 1 when " + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^v" + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "text", + "content": " is the representation of " + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "text", + "content": "-th instance " + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "text", + "content": " observed in view " + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "text", + "content": ", and 0 otherwise. We set a completeness statistical weight " + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "inline_equation", + "content": "w_i^v" + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "text", + "content": " based on the number of observations in " + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "text", + "content": " views for instance " + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 305, + 131, + 545, + 239 + ], + "type": "text", + "content": " as the fusion weight, which makes use of view-specific complementary information without discarding unpaired observations and mitigating the negative impact of missing noise by adapting to the differences in missing instances across different views." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 239, + 545, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 545, + 311 + ], + "type": "text", + "content": "We derive a set of consensus semantic prototypes " + }, + { + "bbox": [ + 305, + 239, + 545, + 311 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\{\\mathbf{c}_k\\in \\mathbb{R}^d\\}_{k = 1}^K" + }, + { + "bbox": [ + 305, + 239, + 545, + 311 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 305, + 239, + 545, + 311 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 239, + 545, + 311 + ], + "type": "text", + "content": "-means on consensus representations " + }, + { + "bbox": [ + 305, + 239, + 545, + 311 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 305, + 239, + 545, + 311 + ], + "type": "text", + "content": ". As \"a representative embedding for semantically similar observations\", " + }, + { + "bbox": [ + 305, + 239, + 545, + 311 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 305, + 239, + 545, + 311 + ], + "type": "text", + "content": " comprehensively captures the semantic information of all data [50] and is continuously refined throughout consensus representation learning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "text", + "content": "Prototype-based Contrastive Clustering. We perform semantic-level contrastive clustering in a shared vector space spanned by consensus prototypes, where the latent representation " + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^v" + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "text", + "content": " encoded as the semantic representation " + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^v" + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^v" + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "text", + "content": " is assigned to the prototype " + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_k" + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "text", + "content": " with the longest projected distance to obtain optimal assignments, in two steps:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 384, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 545, + 421 + ], + "type": "text", + "content": "- Semantic similarity measure: Encode the semantic representation " + }, + { + "bbox": [ + 306, + 384, + 545, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^v" + }, + { + "bbox": [ + 306, + 384, + 545, + 421 + ], + "type": "text", + "content": " and project it onto each prototype, then calculate the probability " + }, + { + "bbox": [ + 306, + 384, + 545, + 421 + ], + "type": "inline_equation", + "content": "p_{i,k}^{v}" + }, + { + "bbox": [ + 306, + 384, + 545, + 421 + ], + "type": "text", + "content": " of belonging to cluster " + }, + { + "bbox": [ + 306, + 384, + 545, + 421 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 384, + 545, + 421 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 430, + 545, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 430, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 367, + 430, + 545, + 461 + ], + "type": "interline_equation", + "content": "p _ {i, k} ^ {v} = \\frac {\\exp \\left(\\mathbf {h} _ {i} ^ {v \\top} \\mathbf {c} _ {k} / \\tau\\right)}{\\sum_ {k ^ {\\prime}} ^ {K} \\exp \\left(\\mathbf {h} _ {i} ^ {v \\top} \\mathbf {c} _ {k ^ {\\prime}} / \\tau\\right)}, \\tag {4}", + "image_path": "e387828c7c8ec4ab8d943a9ad3d55c66708e8c1684773347618a68b9e297f3f3.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 471, + 544, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 544, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 544, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 471, + 544, + 495 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 471, + 544, + 495 + ], + "type": "text", + "content": " is a temperature parameter [40]. " + }, + { + "bbox": [ + 313, + 471, + 544, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i^v = \\{p_{i,k}^v\\}_{k = 1}^K" + }, + { + "bbox": [ + 313, + 471, + 544, + 495 + ], + "type": "text", + "content": " is the soft assignments of observation " + }, + { + "bbox": [ + 313, + 471, + 544, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^v" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "text", + "content": "- Swapped knowledge distillation: Based on Theorem 2, " + }, + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "inline_equation", + "content": "(\\overline{\\mathbf{x}}_i^m,\\overline{\\mathbf{x}}_i^n)" + }, + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "text", + "content": " share the same cluster semantics on consensus prototypes " + }, + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "text", + "content": ", thus their true labels " + }, + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^m,\\mathbf{y}_i^n\\in \\mathbb{R}^K" + }, + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "text", + "content": " should be are distributionally consistent. To this end, \"swapped\" knowledge distillation (KD) utilizes pseudo-labels " + }, + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i^m" + }, + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i^n" + }, + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "text", + "content": " as mutual supervised signals to prompt their cluster assignments " + }, + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i^m,\\mathbf{p}_i^n" + }, + { + "bbox": [ + 306, + 496, + 545, + 579 + ], + "type": "text", + "content": " as identical as possible:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 348, + 588, + 545, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 588, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 348, + 588, + 545, + 602 + ], + "type": "interline_equation", + "content": "\\ell_ {c c} ^ {m, n} = \\ell_ {k d} \\left(\\mathbf {H} ^ {m}, \\mathbf {Q} ^ {n}\\right) + \\ell_ {k d} \\left(\\mathbf {H} ^ {n}, \\mathbf {Q} ^ {m}\\right), \\tag {5}", + "image_path": "518bbb1b42dffcc36385441fd4ad5ec918da1eace84e2def0b6ce437ddc4c9f5.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 611, + 545, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 545, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 611, + 545, + 667 + ], + "type": "inline_equation", + "content": "\\ell_{kd}(\\mathbf{H}^m,\\mathbf{Q}^n) = -\\frac{1}{N_{nm}}\\sum_{i = 1}^{N_{mn}}\\sum_{k = 1}^{K}\\mathbf{q}_i^n\\log \\mathbf{p}_i^m" + }, + { + "bbox": [ + 313, + 611, + 545, + 667 + ], + "type": "text", + "content": " We extend Eq. (5) to more than two views, fostering greater view collaboration to enhance semantic learning:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 384, + 675, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 675, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 384, + 675, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c c} = \\sum_ {m = 1} ^ {V} \\sum_ {\\substack {n = 1 \\\\ n \\neq m}} ^ {V} \\ell_ {c c} ^ {m, n}. \\tag{6}", + "image_path": "052aef6333b4b5f1d928e0f71216b9670bf74637d725dd022abf0d188ffef1b8.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5073" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 531, + 329 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 531, + 329 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 531, + 329 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 531, + 329 + ], + "type": "image", + "image_path": "982238dac916e28d963632dc3904412fd56538aa13f450a290c06bafe53add61.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 336, + 548, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 336, + 548, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 548, + 403 + ], + "type": "text", + "content": "Figure 2. The framework of FreeCSL. (a) Reconstruction module, encodes observations into clustering-friendly representations for each view; (b) Consensus semantic learning module, learns semantic representations through cross-view contrastive clustering in a shared semantic space, where paired observations are assigned to their nearest semantic prototype for consistent assignments. (c) Cluster semantic enhancement module, enriches semantic representations with cluster structure information through within-view graph clustering, which applies GCN to aggregate view-specific semantic information and maximizes spectral modularity to recover cluster structure with greater separation and lower entropy. Ultimately, perform " + }, + { + "bbox": [ + 46, + 336, + 548, + 403 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 336, + 548, + 403 + ], + "type": "text", + "content": "-means on the consensus semantic representations to predict cluster labels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 418, + 288, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 288, + 467 + ], + "type": "text", + "content": "- Consensus label solution: To obtain pseudo-labels " + }, + { + "bbox": [ + 47, + 418, + 288, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^m = \\{\\mathbf{q}_i^m\\}_{i=1}^{N_{mn}}" + }, + { + "bbox": [ + 47, + 418, + 288, + 467 + ], + "type": "text", + "content": ", project semantic representations " + }, + { + "bbox": [ + 47, + 418, + 288, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^m" + }, + { + "bbox": [ + 47, + 418, + 288, + 467 + ], + "type": "text", + "content": " onto consensus prototypes " + }, + { + "bbox": [ + 47, + 418, + 288, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 47, + 418, + 288, + 467 + ], + "type": "text", + "content": " and maximize similarity between the representation " + }, + { + "bbox": [ + 47, + 418, + 288, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^m" + }, + { + "bbox": [ + 47, + 418, + 288, + 467 + ], + "type": "text", + "content": " and its assigned prototype " + }, + { + "bbox": [ + 47, + 418, + 288, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_k" + }, + { + "bbox": [ + 47, + 418, + 288, + 467 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 469, + 287, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 287, + 491 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} ^ {*} = \\arg \\max _ {\\mathbf {Q} ^ {m} \\in \\mathcal {Q}} \\operatorname {T r} \\left\\{\\left(\\mathbf {C Q} ^ {m}\\right) ^ {\\top} \\mathbf {H} ^ {m}\\right) \\} + \\alpha \\mathcal {R} (\\mathbf {Q} ^ {m}), \\tag {7}", + "image_path": "c69640c11743f750d94c0ac694477dbab27a885055a4e0be609c43a32ba28075.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 495, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 287, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 495, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\mathbf{Q}^m) = -\\sum_i^{N_{mn}}\\mathbf{q}_i^m\\log \\mathbf{q}_i^m" + }, + { + "bbox": [ + 55, + 495, + 287, + 521 + ], + "type": "text", + "content": " , an entropy regularizer prevents trivial solutions via the smoothness " + }, + { + "bbox": [ + 55, + 495, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 525, + 279, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 279, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 279, + 538 + ], + "type": "text", + "content": "2.4. Within-view Cluster Semantic Enhancement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": "Considering that graph pooling can discover data implicit correlation and reduce missing noise via aggregating the information of neighboring nodes, we design a graph clustering to enhance representations with cluster semantic information. It combines a heuristic learning objective based on modularity, an evaluation metric for cluster structure quality, to effectively recover the cluster structure[51]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": "Modularity. Modularity [28] quantifies the deviation between the actual cluster structure and the expected structure generated by random combinations in the null model. Modularity matrix " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{B} = \\mathbf{A} - \\frac{\\mathbf{d}\\mathbf{d}^{\\top}}{2m}" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": " is the adjacency matrix, " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": " is the degree vector and " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": " is the total number of edges in the graph. The entry " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "b_{ij} \\in \\mathbf{B}" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": " measures connection strength between nodes " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 418, + 546, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 546, + 479 + ], + "type": "text", + "content": "Modularity-based Graph clustering. Graph clustering relies on graph pooling and node assignment. For each view, we construct graph structures with " + }, + { + "bbox": [ + 304, + 418, + 546, + 479 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 418, + 546, + 479 + ], + "type": "text", + "content": "-nearest neighbors (KNN), then deploy graph convolutional network (GCN) " + }, + { + "bbox": [ + 304, + 418, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_v(\\cdot)" + }, + { + "bbox": [ + 304, + 418, + 546, + 479 + ], + "type": "text", + "content": " to aggregate graph embeddings:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 483, + 545, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 483, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 311, + 483, + 545, + 500 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {v} \\left(\\mathbf {Z} ^ {v}; \\mathbf {A} ^ {v}\\right) = \\sigma \\left(\\mathbf {D} ^ {v - \\frac {1}{2}} \\mathbf {A} ^ {v} \\mathbf {D} ^ {v - \\frac {1}{2}} \\mathbf {Z} ^ {v} \\mathbf {W} ^ {v} + \\mathbf {Z} ^ {v} \\mathbf {W} _ {s} ^ {v}\\right), \\tag {8}", + "image_path": "249a3b2b2fccc96431f89d129bd9c30456d5f2cb7f237a15ea21f670ef95d812.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "text", + "content": " is an activation function used to reinforce the nonlinear aggregation capability of GCN. We replace a skip connection matrix " + }, + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_s^v" + }, + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "text", + "content": " with the self-loop " + }, + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^v +\\mathbf{I}" + }, + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "text", + "content": " to alleviate the over-smoothing of node embeddings. Graph pooling operates on the latent representations " + }, + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^v" + }, + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "text", + "content": ", while the adjacency matrix " + }, + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^v" + }, + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "text", + "content": " is constructed from the original observations " + }, + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^v" + }, + { + "bbox": [ + 304, + 505, + 547, + 577 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 329, + 583, + 545, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 583, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 329, + 583, + 545, + 628 + ], + "type": "interline_equation", + "content": "a _ {i, j} ^ {v} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\left(\\mathbf {x} _ {i} ^ {v} \\text {a n d} \\mathbf {x} _ {j} ^ {v} \\neq N a N\\right) \\& \\\\ & \\left(\\mathbf {x} _ {i} ^ {v} \\in \\mathcal {N} _ {k} \\left(\\mathbf {x} _ {j} ^ {v}\\right) \\text {o r} \\mathbf {x} _ {j} ^ {v} \\in \\mathcal {N} _ {k} \\left(\\mathbf {x} _ {i} ^ {v}\\right)\\right), \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {9}", + "image_path": "a3f8a1204d0cd8f2739c9e3a573a90bb855ab4e020679fe0b7a3eb767d9feb68.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 632, + 546, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 546, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 546, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 632, + 546, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_k(\\mathbf{x}_i^v)" + }, + { + "bbox": [ + 304, + 632, + 546, + 669 + ], + "type": "text", + "content": " is a set including the " + }, + { + "bbox": [ + 304, + 632, + 546, + 669 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 632, + 546, + 669 + ], + "type": "text", + "content": "-nearest neighbors of " + }, + { + "bbox": [ + 304, + 632, + 546, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^v" + }, + { + "bbox": [ + 304, + 632, + 546, + 669 + ], + "type": "text", + "content": ". It brings neighbors closer in the embedding space without introducing missing noise into the graph structure." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "text", + "content": "The softmax classifier offers the soft assignment " + }, + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^v" + }, + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "text", + "content": " for node embeddings encoded by " + }, + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_v(\\mathbf{Z}^v;\\mathbf{A}^v)" + }, + { + "bbox": [ + 305, + 669, + 545, + 693 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 699, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 699, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 365, + 699, + 545, + 713 + ], + "type": "interline_equation", + "content": "\\mathbf {P} ^ {v} = \\operatorname {S o f t m a x} \\left(\\mathcal {G} _ {v} \\left(\\mathbf {Z} ^ {v}; \\mathbf {A} ^ {v}\\right)\\right). \\tag {10}", + "image_path": "e7d54d6dd64add64398c39255bfa08b9181915f5547c0c9cafaa1b2c1e923413.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5074" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": "To obtain a well-separated cluster assignment " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^v" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": ", we maximize the spectral modularity " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathrm{Tr}\\left((\\mathbf{P}^v)^\\top \\mathbf{B}^v\\mathbf{P}^v\\right)" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " [1, 18], which captures intra-cluster connections and intercluster margins by mapping " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^v" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " to modularity " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^v" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": "; To further improve the confidence and robustness of " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^v" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": ", we apply a self-supervised signal " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^v" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": ", learned from contrastive clustering as detailed in Sec. 2.3, to guide " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^v" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " in chasing " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^v" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " via self-knowledge distillation. Specifically, the following objective is designed to optimize " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^v" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 188, + 287, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 188, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 188, + 287, + 220 + ], + "type": "interline_equation", + "content": "\\ell_ {m} ^ {v} \\left(\\mathbf {P} ^ {v}; \\mathbf {L} ^ {v}\\right) = - \\frac {1}{2 m} \\operatorname {T r} \\left(\\left(\\mathbf {P} ^ {v}\\right) ^ {\\top} \\mathbf {B} ^ {v} \\mathbf {P} ^ {v}\\right) + \\lambda \\operatorname {K L} \\left(\\mathbf {L} ^ {v} \\| \\mathbf {P} ^ {v}\\right), \\tag {11}", + "image_path": "719ecf53865b708c29230fa5ff0ca53c0da1a3d5a2d733ef43f731d048adabd0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "text", + "content": "where Kullback-Leibler divergence " + }, + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}(\\cdot)" + }, + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "text", + "content": " serves as a robust regularizer, leveraging " + }, + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "text", + "content": " to regulate the information flow of " + }, + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^v\\in \\mathbb{R}^{N\\times K}" + }, + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "text", + "content": " to ensure stable cluster performance. We project semantic representations " + }, + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^v" + }, + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "text", + "content": " learned through contrastive clustering onto their " + }, + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "text", + "content": "-means prototypes and introduce Student's t-distribution [41] as kernel to predict high-confidence labels " + }, + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^v" + }, + { + "bbox": [ + 46, + 221, + 289, + 305 + ], + "type": "text", + "content": " with a nearly uniform distribution:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 314, + 287, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 314, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 81, + 314, + 287, + 363 + ], + "type": "interline_equation", + "content": "l _ {i, k} ^ {v} = \\frac {\\left(1 + \\frac {\\| \\mathbf {h} _ {i} ^ {v} - \\mathbf {c} _ {k} ^ {v} \\| ^ {2}}{\\gamma}\\right) ^ {- \\frac {\\gamma + 1}{2}}}{\\sum_ {k ^ {\\prime}} \\left(1 + \\frac {\\| \\mathbf {h} _ {i} ^ {v} - \\mathbf {c} _ {k ^ {\\prime}} ^ {v} \\| ^ {2}}{\\gamma}\\right) ^ {- \\frac {\\gamma + 1}{2}}} \\in \\mathbf {L} ^ {v}. \\tag {12}", + "image_path": "19ccdb5b2f29d1797d4c9a5465af99467aa4c4c2128e3cbbfdce2175fb7784df.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 371, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 287, + 396 + ], + "type": "text", + "content": "We conduct graph clustering on each view separately, then optimize assignments across all views concurrently:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 404, + 287, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 404, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 117, + 404, + 287, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g c} = \\sum_ {v = 1} ^ {V} \\ell_ {m} ^ {v} \\left(\\mathbf {P} ^ {v}, \\mathbf {L} ^ {v}\\right). \\tag {13}", + "image_path": "986ec1fd472511c4d65d8a479564073c9a2d8f0dba7bb4ef1023cc64df64fde8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 445, + 271, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 271, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 271, + 459 + ], + "type": "text", + "content": "2.5. Clustering Driven by Consensus Semantics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 464, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 287, + 501 + ], + "type": "text", + "content": "Our FreeCSL learns the semantic representations " + }, + { + "bbox": [ + 47, + 464, + 287, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^v" + }, + { + "bbox": [ + 47, + 464, + 287, + 501 + ], + "type": "text", + "content": " end-to-end by jointly minimizing the reconstruction loss, contrastive clustering loss and graph clustering loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 510, + 287, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 510, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 117, + 510, + 287, + 523 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {r e c} + \\mathcal {L} _ {c c} + \\mathcal {L} _ {g c}. \\tag {14}", + "image_path": "238bc39265139b8b21272870454ed8b0e49209f8342174c93cb1e79836484e56.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 532, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 287, + 581 + ], + "type": "text", + "content": "Without searching for the optimal balancing weights of three loss terms, outstanding performance is easily achieved by applying " + }, + { + "bbox": [ + 47, + 532, + 287, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 532, + 287, + 581 + ], + "type": "text", + "content": "-means to the consensus semantic representations " + }, + { + "bbox": [ + 47, + 532, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 47, + 532, + 287, + 581 + ], + "type": "text", + "content": " learned via the fusion manner " + }, + { + "bbox": [ + 47, + 532, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbb{T}\\left(\\{\\mathbf{H}^v\\}_{v=1}^V\\right)" + }, + { + "bbox": [ + 47, + 532, + 287, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "text", + "content": "The implementation for FreeCSL consists of two stages: warm-up training for the encoder-decoder, followed by fine-tuning for semantic representation learning and clustering. For details, refer to Algorithm 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 639, + 124, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 124, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 124, + 653 + ], + "type": "text", + "content": "3. Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 658, + 174, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 174, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 174, + 672 + ], + "type": "text", + "content": "3.1. Experimental Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "We select four representative datasets namely Caltech-5V[44], ALOI-100[5], YoutubeFace10[12] and NoisyMNIST[23] to demonstrate our model's performance" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 514, + 84 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 514, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 514, + 84 + ], + "type": "text", + "content": "Algorithm 1 FreeCSL for Learning and Clustering" + } + ] + } + ], + "index": 14, + "type": "text" + }, + { + "bbox": [ + 306, + 86, + 545, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 86, + 545, + 123 + ], + "spans": [ + { + "bbox": [ + 306, + 86, + 545, + 123 + ], + "type": "text", + "content": "Input: Complete, incomplete and pair-observed multiview dataset " + }, + { + "bbox": [ + 306, + 86, + 545, + 123 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}^v\\}_{v = 1}^V" + }, + { + "bbox": [ + 306, + 86, + 545, + 123 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{X}}^v\\}_{v = 1}^V" + }, + { + "bbox": [ + 306, + 86, + 545, + 123 + ], + "type": "inline_equation", + "content": "\\{\\overline{\\mathbf{X}}^{m,n}\\}_{m\\neq n}^V" + }, + { + "bbox": [ + 306, + 86, + 545, + 123 + ], + "type": "text", + "content": " networks " + }, + { + "bbox": [ + 306, + 86, + 545, + 123 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}_v,\\mathcal{D}_v,\\mathcal{G}_v\\}_{v = 1}^V" + }, + { + "bbox": [ + 306, + 86, + 545, + 123 + ], + "type": "text", + "content": " ; warm-up and fine-tuning epochs " + }, + { + "bbox": [ + 306, + 86, + 545, + 123 + ], + "type": "inline_equation", + "content": "e,E" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 123, + 379, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 123, + 379, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 123, + 379, + 133 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 306, + 123, + 379, + 133 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 306, + 123, + 379, + 133 + ], + "type": "text", + "content": " to e do" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 133, + 534, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 534, + 148 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 534, + 148 + ], + "type": "text", + "content": "On " + }, + { + "bbox": [ + 306, + 133, + 534, + 148 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{X}}^v\\}_{v = 1}^V" + }, + { + "bbox": [ + 306, + 133, + 534, + 148 + ], + "type": "text", + "content": " , warming up " + }, + { + "bbox": [ + 306, + 133, + 534, + 148 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}_v,\\mathcal{D}_v\\}_{v = 1}^V" + }, + { + "bbox": [ + 306, + 133, + 534, + 148 + ], + "type": "text", + "content": " with Eq.(1)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 150, + 380, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 150, + 380, + 160 + ], + "spans": [ + { + "bbox": [ + 306, + 150, + 380, + 160 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 306, + 150, + 380, + 160 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 306, + 150, + 380, + 160 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 150, + 380, + 160 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 306, + 150, + 380, + 160 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 161, + 544, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 161, + 544, + 197 + ], + "spans": [ + { + "bbox": [ + 321, + 161, + 544, + 197 + ], + "type": "text", + "content": "Obtain latent representations " + }, + { + "bbox": [ + 321, + 161, + 544, + 197 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}^v |\\mathcal{E}_v:\\mathbf{X}^v\\to \\mathbf{Z}^v\\}_{v = 1}^V" + }, + { + "bbox": [ + 321, + 161, + 544, + 197 + ], + "type": "text", + "content": " Construct consensus prototypes " + }, + { + "bbox": [ + 321, + 161, + 544, + 197 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\{\\mathbf{c}_k\\}_{k = 1}^K" + }, + { + "bbox": [ + 321, + 161, + 544, + 197 + ], + "type": "text", + "content": " as elaborated in Sec. 2.3." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 198, + 524, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 198, + 524, + 210 + ], + "spans": [ + { + "bbox": [ + 321, + 198, + 524, + 210 + ], + "type": "text", + "content": "Compute the reconstruction loss " + }, + { + "bbox": [ + 321, + 198, + 524, + 210 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 321, + 198, + 524, + 210 + ], + "type": "text", + "content": " with Eq.(1)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 210, + 539, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 210, + 539, + 222 + ], + "spans": [ + { + "bbox": [ + 321, + 210, + 539, + 222 + ], + "type": "text", + "content": "// Cross-view Contrastive Clustering" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 223, + 449, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 223, + 449, + 234 + ], + "spans": [ + { + "bbox": [ + 321, + 223, + 449, + 234 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 321, + 223, + 449, + 234 + ], + "type": "inline_equation", + "content": "m,n = 1" + }, + { + "bbox": [ + 321, + 223, + 449, + 234 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 321, + 223, + 449, + 234 + ], + "type": "inline_equation", + "content": "V(m\\neq n)" + }, + { + "bbox": [ + 321, + 223, + 449, + 234 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 337, + 234, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 234, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 337, + 234, + 545, + 247 + ], + "type": "text", + "content": "Compute the cluster probability " + }, + { + "bbox": [ + 337, + 234, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i^m,\\mathbf{p}_i^n" + }, + { + "bbox": [ + 337, + 234, + 545, + 247 + ], + "type": "text", + "content": " with Eq.(4);" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 342, + 247, + 530, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 247, + 530, + 258 + ], + "spans": [ + { + "bbox": [ + 342, + 247, + 530, + 258 + ], + "type": "text", + "content": "Solve consensus pseudo-label " + }, + { + "bbox": [ + 342, + 247, + 530, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i^m" + }, + { + "bbox": [ + 342, + 247, + 530, + 258 + ], + "type": "text", + "content": " with Eq.(7);" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 342, + 258, + 525, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 258, + 525, + 271 + ], + "spans": [ + { + "bbox": [ + 342, + 258, + 525, + 271 + ], + "type": "text", + "content": "Get swapped loss " + }, + { + "bbox": [ + 342, + 258, + 525, + 271 + ], + "type": "inline_equation", + "content": "\\ell_{cc}^{m,n}" + }, + { + "bbox": [ + 342, + 258, + 525, + 271 + ], + "type": "text", + "content": " with Eq.(5) on " + }, + { + "bbox": [ + 342, + 258, + 525, + 271 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{X}}^{m,n}" + }, + { + "bbox": [ + 342, + 258, + 525, + 271 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 274, + 532, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 532, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 532, + 286 + ], + "type": "text", + "content": "Calculate contrastive clustering loss " + }, + { + "bbox": [ + 314, + 274, + 532, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cc}" + }, + { + "bbox": [ + 314, + 274, + 532, + 286 + ], + "type": "text", + "content": " with Eq.(6)" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 328, + 287, + 515, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 287, + 515, + 297 + ], + "spans": [ + { + "bbox": [ + 328, + 287, + 515, + 297 + ], + "type": "text", + "content": "// Within-view Graph Clustering" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 328, + 298, + 402, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 298, + 402, + 308 + ], + "spans": [ + { + "bbox": [ + 328, + 298, + 402, + 308 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 328, + 298, + 402, + 308 + ], + "type": "inline_equation", + "content": "v = 1" + }, + { + "bbox": [ + 328, + 298, + 402, + 308 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 328, + 298, + 402, + 308 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 328, + 298, + 402, + 308 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 342, + 309, + 520, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 309, + 520, + 322 + ], + "spans": [ + { + "bbox": [ + 342, + 309, + 520, + 322 + ], + "type": "text", + "content": "Construct adjacency matrix " + }, + { + "bbox": [ + 342, + 309, + 520, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^v" + }, + { + "bbox": [ + 342, + 309, + 520, + 322 + ], + "type": "text", + "content": " with Eq.(9)." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 342, + 322, + 504, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 322, + 504, + 333 + ], + "spans": [ + { + "bbox": [ + 342, + 322, + 504, + 333 + ], + "type": "text", + "content": "Solve node assignment " + }, + { + "bbox": [ + 342, + 322, + 504, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^v" + }, + { + "bbox": [ + 342, + 322, + 504, + 333 + ], + "type": "text", + "content": " with Eq.(10)" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 342, + 334, + 494, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 334, + 494, + 345 + ], + "spans": [ + { + "bbox": [ + 342, + 334, + 494, + 345 + ], + "type": "text", + "content": "Solve pseudo-labels " + }, + { + "bbox": [ + 342, + 334, + 494, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^v" + }, + { + "bbox": [ + 342, + 334, + 494, + 345 + ], + "type": "text", + "content": " with Eq.(12);" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 342, + 346, + 529, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 346, + 529, + 358 + ], + "spans": [ + { + "bbox": [ + 342, + 346, + 529, + 358 + ], + "type": "text", + "content": "Compute KL-modularity loss " + }, + { + "bbox": [ + 342, + 346, + 529, + 358 + ], + "type": "inline_equation", + "content": "\\ell_m^v" + }, + { + "bbox": [ + 342, + 346, + 529, + 358 + ], + "type": "text", + "content": " with Eq.(11)" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 321, + 360, + 517, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 360, + 517, + 373 + ], + "spans": [ + { + "bbox": [ + 321, + 360, + 517, + 373 + ], + "type": "text", + "content": "Calculate graph clustering loss " + }, + { + "bbox": [ + 321, + 360, + 517, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gc}" + }, + { + "bbox": [ + 321, + 360, + 517, + 373 + ], + "type": "text", + "content": " with Eq.(13)" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 327, + 373, + 538, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 373, + 538, + 384 + ], + "spans": [ + { + "bbox": [ + 327, + 373, + 538, + 384 + ], + "type": "text", + "content": "// Semantic Representation Learning" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 327, + 384, + 495, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 384, + 495, + 396 + ], + "spans": [ + { + "bbox": [ + 327, + 384, + 495, + 396 + ], + "type": "text", + "content": "Calculate the overall loss " + }, + { + "bbox": [ + 327, + 384, + 495, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 327, + 384, + 495, + 396 + ], + "type": "text", + "content": " with Eq.(14);" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 327, + 396, + 498, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 396, + 498, + 407 + ], + "spans": [ + { + "bbox": [ + 327, + 396, + 498, + 407 + ], + "type": "text", + "content": "Optimize " + }, + { + "bbox": [ + 327, + 396, + 498, + 407 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}_v,\\mathcal{D}_v,\\mathcal{G}_v\\}_{v = 1}^V" + }, + { + "bbox": [ + 327, + 396, + 498, + 407 + ], + "type": "text", + "content": " to minimize " + }, + { + "bbox": [ + 327, + 396, + 498, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 327, + 407, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 407, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 327, + 407, + 545, + 419 + ], + "type": "text", + "content": "Learn latent and consensus representations " + }, + { + "bbox": [ + 327, + 407, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}^v\\}_{v = 1}^V" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 327, + 419, + 544, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 419, + 544, + 432 + ], + "spans": [ + { + "bbox": [ + 327, + 419, + 544, + 432 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 327, + 419, + 544, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 327, + 419, + 544, + 432 + ], + "type": "text", + "content": ", semantic representations " + }, + { + "bbox": [ + 327, + 419, + 544, + 432 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{H}^v\\}_{v=1}^V" + }, + { + "bbox": [ + 327, + 419, + 544, + 432 + ], + "type": "text", + "content": ", consensus" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 327, + 432, + 542, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 432, + 542, + 444 + ], + "spans": [ + { + "bbox": [ + 327, + 432, + 542, + 444 + ], + "type": "text", + "content": "semantic representations " + }, + { + "bbox": [ + 327, + 432, + 542, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 327, + 432, + 542, + 444 + ], + "type": "text", + "content": ", consensus prototypes " + }, + { + "bbox": [ + 327, + 432, + 542, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 327, + 432, + 542, + 444 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 306, + 445, + 515, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 515, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 515, + 460 + ], + "type": "text", + "content": "Output: " + }, + { + "bbox": [ + 306, + 445, + 515, + 460 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}} = \\{\\hat{y}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 306, + 445, + 515, + 460 + ], + "type": "text", + "content": " predicted by " + }, + { + "bbox": [ + 306, + 445, + 515, + 460 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 445, + 515, + 460 + ], + "type": "text", + "content": "-means on " + }, + { + "bbox": [ + 306, + 445, + 515, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 306, + 445, + 515, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 481, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 529 + ], + "type": "text", + "content": "in comparison with seven SOTA methods summarized in Table 1. To comprehensively evaluate experimental results, three metrics are adopted: accuracy (ACC), normalized mutual information (NMI), and adjusted rand index (ARI)." + } + ] + } + ], + "index": 42 + }, + { + "type": "table", + "bbox": [ + 308, + 570, + 552, + 630 + ], + "blocks": [ + { + "bbox": [ + 305, + 538, + 545, + 561 + ], + "lines": [ + { + "bbox": [ + 305, + 538, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 305, + 538, + 545, + 561 + ], + "type": "text", + "content": "Table 1. SOTA methods categorized by the types of techniques for consistency, imputation, and alignment." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 570, + 552, + 630 + ], + "lines": [ + { + "bbox": [ + 308, + 570, + 552, + 630 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 552, + 630 + ], + "type": "table", + "html": "
CompetitorsConsistencyImputationAlignment
CPM-Nets (TPAMI'20)instance-levelmutual information interaction
COMPLETER (CVPR'21)instance-levelmutual information interaction
DIMVC (AAAI'22)instance-level\\assignment-based
SURE (TPAMI'23)cluster-levelgraph structure transfer
ProLmp (IICA'23)instance-levelsample-prototype relationship inheritanceprototype-based
ICMVC (AAAI'24)instance-levelgraph structure transferassignment-based
DIVIDE (AAAI'24)cluster-levelmutual information interaction
", + "image_path": "665ea7542b9b044bdf8b7df407e386328f225e2af6f11a64c5171df005bafdcd.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "table_body" + } + ], + "index": 44 + }, + { + "bbox": [ + 306, + 647, + 436, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 436, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 436, + 659 + ], + "type": "text", + "content": "3.2. Implementation details" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Models are trained on an NVIDIA 3090 GPU using Adam optimizer in PyTorch 2.1.0. The warming up and finetuning learning rates are 0.0003 and 0.0005, with a batch size of 512. The hyperparameters are set as follows for" + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 235, + 306, + 370 + ], + "type": "aside_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 235, + 306, + 370 + ], + "spans": [ + { + "bbox": [ + 298, + 235, + 306, + 370 + ], + "type": "text", + "content": "1 \n2 \n3 \n4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5075" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": "different datasets: smoothness " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": ", temperature " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\tau \\in \\{0.1, 0.2\\}" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": ", neighbors " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\zeta = 3" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": ", and regularizer weight " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\{0.05, 0.1, 0.2, 0.3\\}" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": ". For all datasets, 4-layer autoencoders and 2-layer GCNs with same MLP structures are used for each view. The layers of contrastive clustering and graph clustering are shared across all views via a single FC layer." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 152, + 204, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 204, + 165 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 204, + 165 + ], + "type": "text", + "content": "3.3. Competitiveness of FreeCSL" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 171, + 287, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 218 + ], + "type": "text", + "content": "FreeCSL is compared with seven SOTA methods across three metrics in Table 2. The results indicate that FreeCSL excellently handles challenges of high missing rates, multiple clusters, and large-scale issues in IMVC:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 220, + 288, + 386 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "content": "- From the perspective of effectiveness, FreeBSD surpasses most SOTA models, particularly on challenging datasets like ALOI-100 (100 clusters) and YouTubeFace10 (30,000 samples). It achieves significant improvements in ACC, with gains of " + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "inline_equation", + "content": "15.12\\%" + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "inline_equation", + "content": "26.38\\%" + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "inline_equation", + "content": "23.91\\%" + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "inline_equation", + "content": "26.21\\%" + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "content": " for ALOI-100, and " + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "inline_equation", + "content": "1.16\\%" + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "inline_equation", + "content": "8.81\\%" + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "inline_equation", + "content": "6.32\\%" + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "inline_equation", + "content": "1.68\\%" + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "content": " for YouTubeFace10, at missing rate " + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "inline_equation", + "content": "r = 0.1" + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "inline_equation", + "content": "0.3" + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "inline_equation", + "content": "0.5" + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "inline_equation", + "content": "0.7" + }, + { + "bbox": [ + 47, + 220, + 287, + 303 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 304, + 288, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 288, + 386 + ], + "type": "text", + "content": "- From the perspective of robustness, Fig. 3 (visualized from Table 2) shows our model's stability as " + }, + { + "bbox": [ + 47, + 304, + 288, + 386 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 304, + 288, + 386 + ], + "type": "text", + "content": " increases. FreeCSL declines gradually, unlike other models with sharp ACC drops, highlighting its robustness in IMVC task. On the small dataset Caltech-5V with " + }, + { + "bbox": [ + 47, + 304, + 288, + 386 + ], + "type": "inline_equation", + "content": "r = 0.7" + }, + { + "bbox": [ + 47, + 304, + 288, + 386 + ], + "type": "text", + "content": ", its Acc remains at " + }, + { + "bbox": [ + 47, + 304, + 288, + 386 + ], + "type": "inline_equation", + "content": "84.56\\%" + }, + { + "bbox": [ + 47, + 304, + 288, + 386 + ], + "type": "text", + "content": ", while on the larger dataset NoisyMNIST, also achieves " + }, + { + "bbox": [ + 47, + 304, + 288, + 386 + ], + "type": "inline_equation", + "content": "92.19\\%" + }, + { + "bbox": [ + 47, + 304, + 288, + 386 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 387, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 590 + ], + "type": "text", + "content": "The above phenomenon can be explained as follows: As " + }, + { + "bbox": [ + 46, + 387, + 287, + 590 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 387, + 287, + 590 + ], + "type": "text", + "content": " increases, paired observations become scarce. Methods like instance-level consistency learning or cross-view imputation and alignment (e.g., CPM-Net, COMPLETER, ICMVC), constrained by their dependence on consistency information, will introduce biases such as false negatives and semantic inconsistencies, resulting in degraded performance. For the complex task ALOI-100, cluster-level methods like SURE and DIVIDE surpass instance-level by exploiting cross-view cluster consistency. But, their reliance on cross-view graphs or mutual information for imputation, while neglecting within-view cluster affiliations, restricts performance. ProImp acknowledges the need to integrate within-view and cross-view information for semantically consistent imputation. However, its prototype alignment accumulates errors, particularly with a large number of clusters, achieving only " + }, + { + "bbox": [ + 46, + 387, + 287, + 590 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 387, + 287, + 590 + ], + "type": "text", + "content": " ACC on ALOI-100." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": "The superiority of FreeCSL stems from the synergistic effect of the CSL and CSE modules, which eliminate semantic gaps and ensure consistent assignments for all observations in a shared semantic space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 647, + 186, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 186, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 186, + 659 + ], + "type": "text", + "content": "3.4. Understanding FreeCSL" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Ablation Study. As shown in Table 3, ablation studies on FreeCSL's three components reveal that the CSL module contributes the most. On ALOI-100, it demonstrates exceptional performance and strong stability, with ACC drop" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 70, + 424, + 152 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 424, + 152 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 424, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 424, + 152 + ], + "type": "image", + "image_path": "44a05a75f84b556842986767080c87927e6efd088f3c383155cc3e63a819a247.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 154, + 392, + 163 + ], + "lines": [ + { + "bbox": [ + 343, + 154, + 392, + 163 + ], + "spans": [ + { + "bbox": [ + 343, + 154, + 392, + 163 + ], + "type": "text", + "content": "(a) Caltech-5V" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 430, + 71, + 537, + 152 + ], + "blocks": [ + { + "bbox": [ + 430, + 71, + 537, + 152 + ], + "lines": [ + { + "bbox": [ + 430, + 71, + 537, + 152 + ], + "spans": [ + { + "bbox": [ + 430, + 71, + 537, + 152 + ], + "type": "image", + "image_path": "85c9df279d182798f03f30987165d0a467e7d1e18294a69fd7a931481cba0a26.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 154, + 507, + 163 + ], + "lines": [ + { + "bbox": [ + 461, + 154, + 507, + 163 + ], + "spans": [ + { + "bbox": [ + 461, + 154, + 507, + 163 + ], + "type": "text", + "content": "(b) ALOI-100" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 313, + 164, + 421, + 244 + ], + "blocks": [ + { + "bbox": [ + 313, + 164, + 421, + 244 + ], + "lines": [ + { + "bbox": [ + 313, + 164, + 421, + 244 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 421, + 244 + ], + "type": "image", + "image_path": "1b84ce01963f7f0afc1aa5b276652d554cd1c87c0f719c30d7cb83dd5f3c2dc7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 247, + 398, + 255 + ], + "lines": [ + { + "bbox": [ + 335, + 247, + 398, + 255 + ], + "spans": [ + { + "bbox": [ + 335, + 247, + 398, + 255 + ], + "type": "text", + "content": "(c) YoutubeFace10" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 429, + 165, + 537, + 245 + ], + "blocks": [ + { + "bbox": [ + 429, + 165, + 537, + 245 + ], + "lines": [ + { + "bbox": [ + 429, + 165, + 537, + 245 + ], + "spans": [ + { + "bbox": [ + 429, + 165, + 537, + 245 + ], + "type": "image", + "image_path": "5601e511d580dbfd6206497a847821e276a337df2824b369ad4151970a346a98.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 247, + 511, + 255 + ], + "lines": [ + { + "bbox": [ + 454, + 247, + 511, + 255 + ], + "spans": [ + { + "bbox": [ + 454, + 247, + 511, + 255 + ], + "type": "text", + "content": "(d) NoisyMNIST" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 263, + 529, + 274 + ], + "lines": [ + { + "bbox": [ + 321, + 263, + 529, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 263, + 529, + 274 + ], + "type": "text", + "content": "Figure 3. Visualization for Table 2 based on metric ACC." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 283, + 545, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 367 + ], + "type": "text", + "content": "ping only " + }, + { + "bbox": [ + 304, + 283, + 545, + 367 + ], + "type": "inline_equation", + "content": "12.4\\%" + }, + { + "bbox": [ + 304, + 283, + 545, + 367 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 283, + 545, + 367 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 283, + 545, + 367 + ], + "type": "text", + "content": " rises from 0.1 to 0.7. This is attributed to CSL's establishment of a shared semantic space via prototype-based semantic contrast, reducing semantic gaps within clusters. The CSE module also plays a positive role in discovering tighter cluster structures by optimizing modularity. Thanks to its enhanced semantics, both the REC and CSL modules achieve notable improvements." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 307, + 376, + 356, + 425 + ], + "blocks": [ + { + "bbox": [ + 307, + 376, + 356, + 425 + ], + "lines": [ + { + "bbox": [ + 307, + 376, + 356, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 356, + 425 + ], + "type": "image", + "image_path": "aa3562998a24fb9c685041adb1552e65e604fac63fc3adb17144c7ffe5597d73.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 426, + 345, + 436 + ], + "lines": [ + { + "bbox": [ + 317, + 426, + 345, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 345, + 436 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 317, + 426, + 345, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(1)}" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 358, + 376, + 405, + 425 + ], + "blocks": [ + { + "bbox": [ + 358, + 376, + 405, + 425 + ], + "lines": [ + { + "bbox": [ + 358, + 376, + 405, + 425 + ], + "spans": [ + { + "bbox": [ + 358, + 376, + 405, + 425 + ], + "type": "image", + "image_path": "1190bef9176bc25e375f1a402aa26b8cd0f07c1dce21c97684398f3d64088821.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 426, + 395, + 436 + ], + "lines": [ + { + "bbox": [ + 367, + 426, + 395, + 436 + ], + "spans": [ + { + "bbox": [ + 367, + 426, + 395, + 436 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 367, + 426, + 395, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(2)}" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 407, + 376, + 456, + 425 + ], + "blocks": [ + { + "bbox": [ + 407, + 376, + 456, + 425 + ], + "lines": [ + { + "bbox": [ + 407, + 376, + 456, + 425 + ], + "spans": [ + { + "bbox": [ + 407, + 376, + 456, + 425 + ], + "type": "image", + "image_path": "be049ac7d7bc6a93db89c7810cc3d9e8e27367c31c872d04f860331de60a9e1d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 426, + 445, + 436 + ], + "lines": [ + { + "bbox": [ + 417, + 426, + 445, + 436 + ], + "spans": [ + { + "bbox": [ + 417, + 426, + 445, + 436 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 417, + 426, + 445, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(3)}" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 459, + 376, + 506, + 425 + ], + "blocks": [ + { + "bbox": [ + 459, + 376, + 506, + 425 + ], + "lines": [ + { + "bbox": [ + 459, + 376, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 459, + 376, + 506, + 425 + ], + "type": "image", + "image_path": "a86e48e7bf1e44e8176ce1bdd11e0d5747ed4113a8f31bdecaaceb3d3d270140.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 426, + 495, + 436 + ], + "lines": [ + { + "bbox": [ + 468, + 426, + 495, + 436 + ], + "spans": [ + { + "bbox": [ + 468, + 426, + 495, + 436 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 468, + 426, + 495, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(4)}" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 507, + 376, + 556, + 426 + ], + "blocks": [ + { + "bbox": [ + 507, + 376, + 556, + 426 + ], + "lines": [ + { + "bbox": [ + 507, + 376, + 556, + 426 + ], + "spans": [ + { + "bbox": [ + 507, + 376, + 556, + 426 + ], + "type": "image", + "image_path": "da8924045196a8f8977ee07e95801095fe486f5a6a6c74bf5f90b34fe48dc386.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 523, + 427, + 541, + 437 + ], + "lines": [ + { + "bbox": [ + 523, + 427, + 541, + 437 + ], + "spans": [ + { + "bbox": [ + 523, + 427, + 541, + 437 + ], + "type": "text", + "content": "(e) Z" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 307, + 480, + 355, + 529 + ], + "blocks": [ + { + "bbox": [ + 306, + 447, + 545, + 469 + ], + "lines": [ + { + "bbox": [ + 306, + 447, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 447, + 545, + 469 + ], + "type": "text", + "content": "Figure 4. Similarity matrices of " + }, + { + "bbox": [ + 306, + 447, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}^v\\}_{v=1}^4" + }, + { + "bbox": [ + 306, + 447, + 545, + 469 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 447, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 306, + 447, + 545, + 469 + ], + "type": "text", + "content": " without consensus semantic learning on ALOI-100 with " + }, + { + "bbox": [ + 306, + 447, + 545, + 469 + ], + "type": "inline_equation", + "content": "r = 0.5" + }, + { + "bbox": [ + 306, + 447, + 545, + 469 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 480, + 355, + 529 + ], + "lines": [ + { + "bbox": [ + 307, + 480, + 355, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 480, + 355, + 529 + ], + "type": "image", + "image_path": "3a549c7483c4570036923ec7730acc478c8dd1c092cf33f18f2a8d076fb27e8e.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 530, + 346, + 540 + ], + "lines": [ + { + "bbox": [ + 317, + 530, + 346, + 540 + ], + "spans": [ + { + "bbox": [ + 317, + 530, + 346, + 540 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 317, + 530, + 346, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{(1)}" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 546, + 545, + 568 + ], + "lines": [ + { + "bbox": [ + 306, + 546, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 306, + 546, + 545, + 568 + ], + "type": "text", + "content": "Figure 5. Similarity matrices of " + }, + { + "bbox": [ + 306, + 546, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{H}^v\\}_{v=1}^4" + }, + { + "bbox": [ + 306, + 546, + 545, + 568 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 546, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 306, + 546, + 545, + 568 + ], + "type": "text", + "content": " with consensus semantic learning on ALOI-100 with " + }, + { + "bbox": [ + 306, + 546, + 545, + 568 + ], + "type": "inline_equation", + "content": "r = 0.5" + }, + { + "bbox": [ + 306, + 546, + 545, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 358, + 480, + 405, + 529 + ], + "blocks": [ + { + "bbox": [ + 358, + 480, + 405, + 529 + ], + "lines": [ + { + "bbox": [ + 358, + 480, + 405, + 529 + ], + "spans": [ + { + "bbox": [ + 358, + 480, + 405, + 529 + ], + "type": "image", + "image_path": "ba7702b5c0b82ff079233308ebc3fb4d6a656d385f2cec563eae387b3752eef4.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 530, + 396, + 540 + ], + "lines": [ + { + "bbox": [ + 367, + 530, + 396, + 540 + ], + "spans": [ + { + "bbox": [ + 367, + 530, + 396, + 540 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 367, + 530, + 396, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{(2)}" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 407, + 480, + 455, + 529 + ], + "blocks": [ + { + "bbox": [ + 407, + 480, + 455, + 529 + ], + "lines": [ + { + "bbox": [ + 407, + 480, + 455, + 529 + ], + "spans": [ + { + "bbox": [ + 407, + 480, + 455, + 529 + ], + "type": "image", + "image_path": "be1186ed0f33521b45d1520fe5a77984b6479338f23eda584f699ce81140c2fb.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 530, + 445, + 540 + ], + "lines": [ + { + "bbox": [ + 417, + 530, + 445, + 540 + ], + "spans": [ + { + "bbox": [ + 417, + 530, + 445, + 540 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 417, + 530, + 445, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{(3)}" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 459, + 480, + 506, + 529 + ], + "blocks": [ + { + "bbox": [ + 459, + 480, + 506, + 529 + ], + "lines": [ + { + "bbox": [ + 459, + 480, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 459, + 480, + 506, + 529 + ], + "type": "image", + "image_path": "bf7a82ff40da53536f810e53225dc1e4c9d88a560faeec609abf1c63cdfd95b4.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 467, + 530, + 496, + 540 + ], + "lines": [ + { + "bbox": [ + 467, + 530, + 496, + 540 + ], + "spans": [ + { + "bbox": [ + 467, + 530, + 496, + 540 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 467, + 530, + 496, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{(4)}" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 507, + 480, + 556, + 529 + ], + "blocks": [ + { + "bbox": [ + 507, + 480, + 556, + 529 + ], + "lines": [ + { + "bbox": [ + 507, + 480, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 507, + 480, + 556, + 529 + ], + "type": "image", + "image_path": "ccfa1adbbe01951b9dc76e113e1d6d5ec3aa3901859807a138ad0a9d6e337f2c.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 522, + 531, + 542, + 540 + ], + "lines": [ + { + "bbox": [ + 522, + 531, + 542, + 540 + ], + "spans": [ + { + "bbox": [ + 522, + 531, + 542, + 540 + ], + "type": "text", + "content": "(e) H" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "Imputation- and Alignment-free CSL. To verify FreeCSL can reduce semantic gaps and capture semantic relationships without requiring imputation or alignment, we set up two imputation experiments: one imputation for latent representations " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^v" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " without consensus semantic learning named ILR, and the second for semantic representations " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^v" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " learned from consensus semantic learning named ISR. They transfer complete " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "-nn graph structures from other views to incomplete views and utilize corresponding " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " neighbors for imputation. The control groups perform " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "-means on the sum of all view matrices to predict cluster" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5076" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 87, + 547, + 372 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 534, + 82 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 534, + 82 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 534, + 82 + ], + "type": "text", + "content": "Table 2. Performance comparison on four multi-view benchmarks. The best and second-best results are highlighted in red and blue." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 87, + 547, + 372 + ], + "lines": [ + { + "bbox": [ + 49, + 87, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 49, + 87, + 547, + 372 + ], + "type": "table", + "html": "
Missing ratesr=0.1r=0.3r=0.5r=0.7
MetricsACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)
Caltech-5VCPM-Nets85.7274.6271.9285.4372.9070.8481.0868.6166.0879.5065.5262.70
COMPLETETER58.6867.6942.9865.0768.7547.1064.4666.0942.8069.0067.4651.49
DIMVC85.8680.0174.6378.9078.7669.5580.0170.0866.3569.8660.1648.70
SURE75.6469.1262.8377.8967.8961.1577.2265.9960.9973.9660.8953.66
ProImp66.6257.6647.5472.1761.7052.4965.2855.7746.2260.3649.2339.63
ICMVC76.8673.9267.3776.9373.3966.3079.2972.6766.2172.2166.6358.18
DIVIDE57.2947.0033.2368.1458.9344.1681.5771.4559.2074.3668.6260.14
Ours91.5785.3283.1488.8681.2678.6888.3680.0178.0283.6471.7968.06
ALOI-100CPM-Nets63.9579.1451.2652.3070.7439.7942.7462.4628.4030.5254.0317.18
COMPLETETER21.8646.9412.1022.0446.4511.6219.3844.8110.7817.1843.689.60
DIMVC56.9075.0035.5950.5572.5128.1052.5471.4532.5648.9770.8026.16
SURE64.5977.7452.6160.2274.7647.2160.6574.8447.3950.2368.9137.48
ProImp26.1552.9431.4024.5253.0728.1827.9758.1033.4220.3350.1024.28
ICMVC55.4874.0442.9148.8367.4636.5035.4256.9824.1926.8447.4316.63
DIVIDE76.0188.5970.5961.9080.5353.2952.5673.1740.5742.6667.8230.93
Ours91.1395.2287.5988.2893.3483.9584.5690.5378.9076.4485.0067.68
YouTubeFace10CPM-Nets75.4878.5666.5067.5075.8261.1566.6075.1962.0167.3274.1062.54
COMPLETETER46.7252.0421.0350.2452.8127.7736.4135.2415.1840.9839.9220.82
DIMVC81.7781.3274.5971.9676.4262.2973.8775.4463.9774.9479.3468.96
SURE26.6523.6011.4126.4322.4010.5531.8428.7715.2729.3727.4812.85
ProImp48.9753.2033.8343.3049.0426.5339.6750.1826.6849.5251.4729.12
ICMVC68.9074.1760.0662.7665.4453.0344.5652.5132.5127.4427.0915.33
DIVIDE31.8529.6411.3535.8630.8112.3132.7430.8212.7435.1733.4715.93
Ours82.9383.5574.7680.7781.4671.6280.1981.0771.3776.6281.3173.22
NoisyMNISTCPM-Nets48.6444.0832.0042.8837.0126.9250.5243.5433.5646.4637.1827.04
COMPLETETER76.5587.9577.1878.1686.1176.8364.7965.7754.6841.1540.6027.58
DIMVC55.8854.7841.2956.7758.2943.5162.2862.6648.0165.9467.4354.58
SURE98.6195.8996.9697.1193.7591.6494.8589.0687.8984.0078.2576.02
ProImp98.3095.3196.3593.2986.0585.3878.1669.4564.3750.0441.3333.34
ICMVC98.7896.3697.3597.7593.7195.1181.6479.2475.4753.9150.9144.56
DIVIDE94.7291.4489.4395.8589.7191.0657.0957.2946.2828.5725.6111.14
Ours99.1397.2398.1097.6893.9494.9496.0489.8191.4892.1982.5083.56
", + "image_path": "fe63b0e6f33c0188f20ebd865948b55ddc8ee551cad69ff87b63be1c56ab8ebb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 398, + 545, + 496 + ], + "blocks": [ + { + "bbox": [ + 48, + 383, + 545, + 395 + ], + "lines": [ + { + "bbox": [ + 48, + 383, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 545, + 395 + ], + "type": "text", + "content": "Table 3. Ablation study on Caltech-5V and ALOI-100. " + }, + { + "bbox": [ + 48, + 383, + 545, + 395 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 48, + 383, + 545, + 395 + ], + "type": "text", + "content": " denotes FreeCSL with the component and the best results are highlighted in red." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 398, + 545, + 496 + ], + "lines": [ + { + "bbox": [ + 49, + 398, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 49, + 398, + 545, + 496 + ], + "type": "table", + "html": "
Componentsr=0.1r=0.3r=0.5r=0.7
LrecLccLgcACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)
Caltech-5V✓75.0065.6458.8969.0057.8451.4952.2144.1836.7151.9343.4235.75
✓✓85.5078.0772.2882.0074.5367.6785.3674.3271.1079.1467.8162.78
✓✓81.0770.4566.1380.8669.6167.4974.1459.3955.9070.7153.2549.95
✓✓✓91.5785.3283.1488.8681.2678.6888.3680.0178.0283.6471.7968.06
ALOI-100✓63.8177.9249.3844.6965.6431.1132.2055.8318.9423.7848.9011.31
✓✓87.6993.6484.1285.2591.0279.4580.2388.0474.1575.2983.3865.72
✓✓81.0685.9671.2164.1375.8751.4947.5265.5934.2932.7356.3519.23
✓✓✓91.1395.2287.5988.2893.3483.9584.5690.5378.9076.4485.0067.68
", + "image_path": "b8dba52c9e56866444244d9b6d87505e359cba356da2b20ffddd55438aa29903.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 52, + 533, + 543, + 617 + ], + "blocks": [ + { + "bbox": [ + 47, + 506, + 546, + 531 + ], + "lines": [ + { + "bbox": [ + 47, + 506, + 546, + 531 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 546, + 531 + ], + "type": "text", + "content": "Table 4. Imputation- and alignment-free study on Caltech-5V and ALOI-100. ILR and ISR are filled with K-NN imputation via cross-view graph for latent representations and semantic representations " + }, + { + "bbox": [ + 47, + 506, + 546, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(v)}" + }, + { + "bbox": [ + 47, + 506, + 546, + 531 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 506, + 546, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{(v)}" + }, + { + "bbox": [ + 47, + 506, + 546, + 531 + ], + "type": "text", + "content": ". The best results are highlighted in red." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 533, + 543, + 617 + ], + "lines": [ + { + "bbox": [ + 52, + 533, + 543, + 617 + ], + "spans": [ + { + "bbox": [ + 52, + 533, + 543, + 617 + ], + "type": "table", + "html": "
Missing ratesr=0.1r=0.3r=0.5r=0.7
MetricsACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)ACC (%)NMI (%)ARI (%)
Caltech-5VILR91.7185.7683.4489.1481.6979.0788.7980.7978.3777.8665.6455.12
ISR91.6485.5583.3089.4382.2679.6288.8680.7878.7481.2966.6863.59
FreeCSL91.5785.3283.1488.8681.2678.6888.3680.0178.0283.6471.7968.06
ALGI-100ILR58.9279.6741.5453.4675.8836.8945.8369.4727.4138.0858.9621.35
ISR90.1695.3887.8688.9194.0885.4882.5388.7371.5465.3677.8546.06
FreeCSL91.1395.2287.5988.2893.3483.9584.5690.5378.9076.4485.0067.68
", + "image_path": "ffea6045aaa997e612660d27dac77827397ab7f2c8ae96856a1ac3b2655d8b32.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 636, + 288, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 636, + 288, + 708 + ], + "spans": [ + { + "bbox": [ + 46, + 636, + 288, + 708 + ], + "type": "text", + "content": "labels. Table 4 shows FreeBSD's robustness advantage at high " + }, + { + "bbox": [ + 46, + 636, + 288, + 708 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 636, + 288, + 708 + ], + "type": "text", + "content": " on Caltech-5V, with notable performance disparity on ALOI-100. ILR, lacking consensus semantic learning, struggles with multi-cluster tasks due to cross-view semantic gaps. ISR avoids semantic confusion via consensus semantic learning but underperforms FreeBSD due to biased" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 636, + 545, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 636, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 305, + 636, + 545, + 684 + ], + "type": "text", + "content": "imputation as consistency information decreases at higher " + }, + { + "bbox": [ + 305, + 636, + 545, + 684 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 305, + 636, + 545, + 684 + ], + "type": "text", + "content": ". FreeCSL, by contrast, integrates consistency and complementary information in consensus semantic representations for confident decisions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 689, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 715 + ], + "type": "text", + "content": "To further illustrate how consensus semantic learning \"free up\" imputation and alignment, we construct cosine" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5077" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": "similarity matrices on ALOI-100 for view-specific latent representations " + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}^v\\}_{v = 1}^4" + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": " and their consensus " + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": ", as well as view-specific semantic representations " + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{H}^v\\}_{v = 1}^4" + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": " and their consensus " + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": " in Fig. 5, then analyze their information entropy. The similarity matrices of " + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}^v\\}_{v = 1}^4" + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": " are chaotic, with high uncertainty in intra- and inter- cluster relationships. The fusion manner " + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\mathbb{T}(\\cdot)" + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": " with Eq. (2) alleviates high entropy by incorporating view-specific complementary information. Compared to Fig. 4, the similarity matrices of " + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{H}^v\\}_{v = 1}^4" + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": " shows a clear block structure, with high intracluster similarity and distinct inter-cluster differences. This confirms " + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 47, + 72, + 289, + 251 + ], + "type": "text", + "content": ", enhanced by view-specific information, reduces cross-view semantic gaps, while consensus semantic learning, capturing cluster semantic relationship, promotes high-confidence assignments as stated in Theorem 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 260, + 168, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 260, + 168, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 168, + 272 + ], + "type": "text", + "content": "3.5. Analysis on FreeCSL" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 279, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 422 + ], + "type": "text", + "content": "Convergence and Robustness Analysis. In Fig. 6, we plot metrics and losses over training iterations, with error bands from 5 random Caltech-5V experiments to assess robustness. Both converge to stable values with minimal fluctuations and reach stability simultaneously. The gradual decrease in reconstruction loss indicates that the CSL and CSE modules are reasonable and beneficial, as they don't introduce significant discrepancies between semantic and latent representations, further underscoring the benefits of pre-training. Thanks to the harmonious collaboration of the three modules, our FreeCSL achieved satisfactory results within 30 iterations at minimal computational cost." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "text", + "content": "Parameter Sensitivity Analysis. Two hyperparameters in FreeCSL warrant investigation, namely, graph neighbors " + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "text", + "content": " in Eq. (9) and regularizer coefficient " + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "text", + "content": " in Eq. (11). We expanded " + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "text", + "content": " to the range of 0.05 to 0.5 and 3 to 32 respectively. In Fig. 7, our model is insensitive to " + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "text", + "content": " but more affected by variations in " + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "text", + "content": ". This is because the robust regularizer leverages semantic knowledge learned from the CSL module to guide the CSE model toward faster and more stable learning. A small " + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "text", + "content": " cannot provide effective consistency constraints while a large one makes CSL to dominate the learn direction, hindering CSE from exploring cluster structure information. Therefore, we set " + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\lambda \\in [0.05, 0.2]" + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "text", + "content": " to balance cooperative relationship between CSE and CSL. Without sacrificing ACC and NMI, we prefer to set the minimum " + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\zeta = 3" + }, + { + "bbox": [ + 46, + 423, + 287, + 603 + ], + "type": "text", + "content": " to reduce computational complexity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 610, + 284, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 284, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 284, + 623 + ], + "type": "text", + "content": "3.6. Visualization of Consensus Semantic Clusters" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "To further examine the quality of cluster structures formed by consensus semantic representations, t-SNE visualization based on true labels are plotted in Fig. 8 and clearly show minimal misclassification and clusters with strong intracluster cohesion while distinct inter-cluster separation. This indicates FreeCSL nearly recovers true cluster structures by learning consensus semantic information from all data." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 535, + 125 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 535, + 125 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 535, + 125 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 535, + 125 + ], + "type": "image", + "image_path": "02329fcf383e903b6352efbb8b9fd3ea543fc078ad344ffb05e1ba3122356704.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 125, + 535, + 186 + ], + "blocks": [ + { + "bbox": [ + 317, + 125, + 535, + 186 + ], + "lines": [ + { + "bbox": [ + 317, + 125, + 535, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 125, + 535, + 186 + ], + "type": "image", + "image_path": "2062e21a39ba687681e800e9a1b0c9e0812579a95be44574e4bea859edbec620.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 189, + 535, + 200 + ], + "lines": [ + { + "bbox": [ + 317, + 189, + 535, + 200 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 535, + 200 + ], + "type": "text", + "content": "Figure 6. Convergence analysis on Caltech-5V with " + }, + { + "bbox": [ + 317, + 189, + 535, + 200 + ], + "type": "inline_equation", + "content": "r = 0.5" + }, + { + "bbox": [ + 317, + 189, + 535, + 200 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 211, + 423, + 285 + ], + "blocks": [ + { + "bbox": [ + 315, + 211, + 423, + 285 + ], + "lines": [ + { + "bbox": [ + 315, + 211, + 423, + 285 + ], + "spans": [ + { + "bbox": [ + 315, + 211, + 423, + 285 + ], + "type": "image", + "image_path": "8c77d24e6d6d40a3792fe17dc3a5c26fc5d74651030a5090cf99044f1c516b22.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 286, + 394, + 296 + ], + "lines": [ + { + "bbox": [ + 345, + 286, + 394, + 296 + ], + "spans": [ + { + "bbox": [ + 345, + 286, + 394, + 296 + ], + "type": "text", + "content": "(a) Caltech-5V" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 426, + 211, + 537, + 285 + ], + "blocks": [ + { + "bbox": [ + 426, + 211, + 537, + 285 + ], + "lines": [ + { + "bbox": [ + 426, + 211, + 537, + 285 + ], + "spans": [ + { + "bbox": [ + 426, + 211, + 537, + 285 + ], + "type": "image", + "image_path": "e2747f88cb4ba3cc24c3cd73a34ae0d1b17a3d4e36003973fce81f362debab36.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 286, + 506, + 296 + ], + "lines": [ + { + "bbox": [ + 460, + 286, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 460, + 286, + 506, + 296 + ], + "type": "text", + "content": "(b) ALOI-100" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 326, + 414, + 403 + ], + "blocks": [ + { + "bbox": [ + 310, + 326, + 414, + 403 + ], + "lines": [ + { + "bbox": [ + 310, + 326, + 414, + 403 + ], + "spans": [ + { + "bbox": [ + 310, + 326, + 414, + 403 + ], + "type": "image", + "image_path": "a184b06c80493552a4a7dec02d99654557a186dec520c0d0c59456030409d828.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 407, + 413, + 417 + ], + "lines": [ + { + "bbox": [ + 310, + 407, + 413, + 417 + ], + "spans": [ + { + "bbox": [ + 310, + 407, + 413, + 417 + ], + "type": "text", + "content": "(a) Pre-training " + }, + { + "bbox": [ + 310, + 407, + 413, + 417 + ], + "type": "inline_equation", + "content": "(\\mathrm{NMI} = 44.18\\%)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 434, + 325, + 545, + 405 + ], + "blocks": [ + { + "bbox": [ + 325, + 300, + 526, + 312 + ], + "lines": [ + { + "bbox": [ + 325, + 300, + 526, + 312 + ], + "spans": [ + { + "bbox": [ + 325, + 300, + 526, + 312 + ], + "type": "text", + "content": "Figure 7. Parameter analyses for " + }, + { + "bbox": [ + 325, + 300, + 526, + 312 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 325, + 300, + 526, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 325, + 300, + 526, + 312 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 325, + 300, + 526, + 312 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 325, + 300, + 526, + 312 + ], + "type": "inline_equation", + "content": "r = 0.5" + }, + { + "bbox": [ + 325, + 300, + 526, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 434, + 325, + 545, + 405 + ], + "lines": [ + { + "bbox": [ + 434, + 325, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 434, + 325, + 545, + 405 + ], + "type": "image", + "image_path": "4864853a6c5b8ece8aea7eda096bea427ceb7de01d819a16b259cd0ce1295390.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 407, + 537, + 417 + ], + "lines": [ + { + "bbox": [ + 445, + 407, + 537, + 417 + ], + "spans": [ + { + "bbox": [ + 445, + 407, + 537, + 417 + ], + "type": "text", + "content": "(b) Training " + }, + { + "bbox": [ + 445, + 407, + 537, + 417 + ], + "type": "inline_equation", + "content": "(\\mathrm{NMI} = 79.61\\%)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 308, + 423, + 414, + 502 + ], + "blocks": [ + { + "bbox": [ + 308, + 423, + 414, + 502 + ], + "lines": [ + { + "bbox": [ + 308, + 423, + 414, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 423, + 414, + 502 + ], + "type": "image", + "image_path": "9b7e775dda1051fb260d6174348c8a73dc5a5f9b93a95534842b4db00fbdc4f8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 506, + 413, + 516 + ], + "lines": [ + { + "bbox": [ + 310, + 506, + 413, + 516 + ], + "spans": [ + { + "bbox": [ + 310, + 506, + 413, + 516 + ], + "type": "text", + "content": "(c) Pre-training " + }, + { + "bbox": [ + 310, + 506, + 413, + 516 + ], + "type": "inline_equation", + "content": "(\\mathrm{NMI} = 33.74\\%)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 434, + 422, + 542, + 504 + ], + "blocks": [ + { + "bbox": [ + 434, + 422, + 542, + 504 + ], + "lines": [ + { + "bbox": [ + 434, + 422, + 542, + 504 + ], + "spans": [ + { + "bbox": [ + 434, + 422, + 542, + 504 + ], + "type": "image", + "image_path": "dc926b38bf6e88ca2189db92557430db3e98a2292d60f720f102c51885195282.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 506, + 535, + 516 + ], + "lines": [ + { + "bbox": [ + 443, + 506, + 535, + 516 + ], + "spans": [ + { + "bbox": [ + 443, + 506, + 535, + 516 + ], + "type": "text", + "content": "(d) Training " + }, + { + "bbox": [ + 443, + 506, + 535, + 516 + ], + "type": "inline_equation", + "content": "(\\mathrm{NMI} = 89.81\\%)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 522, + 531, + 533 + ], + "lines": [ + { + "bbox": [ + 320, + 522, + 531, + 533 + ], + "spans": [ + { + "bbox": [ + 320, + 522, + 531, + 533 + ], + "type": "text", + "content": "Figure 8. Visualization for Caltech-5V and NoisyMNIST." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 544, + 378, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 378, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 378, + 555 + ], + "type": "text", + "content": "4. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 563, + 545, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 707 + ], + "type": "text", + "content": "In this paper, we propose FreeBSD, a novel semantic learning paradigm free from imputation and alignment compared to existing consistency learning. We design prototype-based contrastive clustering to discover a shared semantic space, where observations converge toward their respective semantic prototype and are encoded as consensus semantics representations for clustering. Furthermore, we employ modularity-inspired graph clustering to enrich semantic representation with view-specific cluster information. The effective synergy of consensus semantic learning and cluster semantic enhancement makes FreeBSD excel in most complex IMVC tasks." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5078" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 144, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 144, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 144, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 156 + ], + "type": "text", + "content": "This work is supported by the National Key R&D Program of China under Grant No.2022ZD0209103, the National Natural Science Foundation of China (project no. 62325604, 62276271, 62406329, 62476281, 62441618), National Natural Science Foundation of China Joint Found under Grant No. U24A20323." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 166, + 107, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 166, + 107, + 178 + ], + "spans": [ + { + "bbox": [ + 48, + 166, + 107, + 178 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 186, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 287, + 229 + ], + "type": "text", + "content": "[1] Ulrik Brandes, Daniel Delling, Marco Gaertler, Robert Görke, Martin Hoefer, Zoran Nikoloski, and Dorothea Wagner. Maximizing modularity is hard. arXiv preprint physics/0608255, 2006. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 230, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 288, + 284 + ], + "type": "text", + "content": "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 339 + ], + "type": "text", + "content": "[3] Chenhang Cui, Yazhou Ren, Jingyu Pu, Jiawei Li, Xiaorong Pu, Tianyi Wu, Yutao Shi, and Lifang He. A novel approach for effective multi-view clustering with information-theoretic perspective. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 340, + 287, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 394 + ], + "type": "text", + "content": "[4] Zhibin Dong, Jiaqi Jin, Yuyang Xiao, Bin Xiao, Siwei Wang, Xinwang Liu, and En Zhu. Subgraph propagation and contrastive calibration for incomplete multiview data clustering. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 395, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 395, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 395, + 287, + 437 + ], + "type": "text", + "content": "[5] Guowang Du, Lihua Zhou, Yudi Yang, Kevin Lu, and Lizhen Wang. Deep multiple auto-encoder-based multi-view clustering. Data Science and Engineering, 6(3):323-338, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 438, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 438, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 53, + 438, + 287, + 483 + ], + "type": "text", + "content": "[6] Uno Fang, Man Li, Jianxin Li, Longxiang Gao, Tao Jia, and Yanchun Zhang. A comprehensive survey on multi-view clustering. IEEE Transactions on Knowledge and Data Engineering, 35(12):12350-12368, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 483, + 287, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 483, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 53, + 483, + 287, + 526 + ], + "type": "text", + "content": "[7] Wei Feng, Guoshuai Sheng, Qianqian Wang, Quanxue Gao, Zhiqiang Tao, and Bo Dong. Partial multi-view clustering via self-supervised network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11988-11995, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 526, + 287, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 526, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 53, + 526, + 287, + 570 + ], + "type": "text", + "content": "[8] Ruiming Guo, Mouxing Yang, Yijie Lin, Xi Peng, and Peng Hu. Robust contrastive multi-view clustering against dual noisy correspondence. Advances in Neural Information Processing Systems, 37:121401-121421, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 571, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 571, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 571, + 287, + 624 + ], + "type": "text", + "content": "[9] Changhao He, Hongyuan Zhu, Peng Hu, and Xi Peng. Robust variational contrastive learning for partially view-unaligned clustering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4167-4176, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "text", + "content": "[10] Junlin Hu, Jiwen Lu, and Yap-Peng Tan. Sharable and individual multi-view metric learning. IEEE transactions on pattern analysis and machine intelligence, 40(9):2281-2288, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[11] Peng Hu, Liangli Zhen, Xi Peng, Hongyuan Zhu, Jie Lin, Xu Wang, and Dezhong Peng. Deep supervised multi-view learning with graph priors. IEEE Transactions on Image Processing, 33:123-133, 2023. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[12] Dong Huang, Chang-Dong Wang, and Jian-Huang Lai. Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity. IEEE Transactions on Knowledge and Data Engineering, 35(11):11388-11402, 2023. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "type": "text", + "content": "[13] Jiabo Huang, Shaogang Gong, and Xiatian Zhu. Deep semantic clustering by partition confidence maximisation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8849-8858, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[14] Zhenyu Huang, Peng Hu, Joey Tianyi Zhou, Jiancheng Lv, and Xi Peng. Partially view-aligned clustering. Advances in Neural Information Processing Systems, 33:2892-2902, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "text", + "content": "[15] Yangbangyan Jiang, Qianqian Xu, Zhiyong Yang, Xiaochun Cao, and Qingming Huang. Dm2c: Deep mixed-modal clustering. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 253, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 308 + ], + "type": "text", + "content": "[16] Jiaqi Jin, Siwei Wang, Zhibin Dong, Xinwang Liu, and En Zhu. Deep incomplete multi-view clustering with cross-view partial sample and prototype alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11600-11609, 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "type": "text", + "content": "[17] Guanzhou Ke, Bo Wang, Xiaoli Wang, and Shengfeng He. Rethinking multi-view representation learning via distilled disentangling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26774-26783, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[18] Brian W Kernighan and Shen Lin. An efficient heuristic procedure for partitioning graphs. The Bell system technical journal, 49(2):291-307, 1970. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[19] Haobin Li, Yunfan Li, Mouxing Yang, Peng Hu, Dezhong Peng, and Xi Peng. Incomplete multi-view clustering via prototype-based imputation. arXiv preprint arXiv:2301.11045, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "text", + "content": "[20] Xingfeng Li, Yinghui Sun, Quansen Sun, Zhenwen Ren, and Yuan Sun. Cross-view graph matching guided anchor alignment for incomplete multi-view clustering. Information Fusion, 100:101941, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "text", + "content": "[21] Yingming Li, Ming Yang, and Zhongfei Zhang. A survey of multi-view representation learning. IEEE transactions on knowledge and data engineering, 31(10):1863-1883, 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 534, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 547, + 578 + ], + "type": "text", + "content": "[22] Yunfan Li, Peng Hu, Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8547-8555, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "type": "text", + "content": "[23] Yijie Lin, Yuanbiao Gou, Zitao Liu, Boyun Li, Jiancheng Lv, and Xi Peng. Completer: Incomplete multi-view clustering via contrastive prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11174-11183, 2021. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[24] Yijie Lin, Yuanbiao Gou, Xiaotian Liu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4): 4447-4461, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 690, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 715 + ], + "type": "text", + "content": "[25] Xinwang Liu, Xinzhong Zhu, Miaomiao Li, Lei Wang, Chang Tang, Jianping Yin, Dinggang Shen, Huaimin Wang," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5079" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 106 + ], + "type": "text", + "content": "and Wen Gao. Late fusion incomplete multi-view clustering. IEEE transactions on pattern analysis and machine intelligence, 41(10):2410-2423, 2018. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 286, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 286, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 286, + 139 + ], + "type": "text", + "content": "[26] Yiding Lu, Haobin Li, Yunfan Li, Yijie Lin, and Xi Peng. A survey on deep clustering: from the prior perspective. *Vicinagearth*, 1(1):4, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 286, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 286, + 193 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 286, + 193 + ], + "type": "text", + "content": "[27] Yiding Lu, Yijie Lin, Mouxing Yang, Dezhong Peng, Peng Hu, and Xi Peng. Decoupled contrastive multi-view clustering with high-order random walks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14193-14201, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 286, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 286, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 286, + 226 + ], + "type": "text", + "content": "[28] Mark EJ Newman. Modularity and community structure in networks. Proceedings of the national academy of sciences, 103(23):8577-8582, 2006. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 286, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 286, + 282 + ], + "type": "text", + "content": "[29] Jingyu Pu, Chenhang Cui, Xinyue Chen, Yazhou Ren, Xiaorong Pu, Zhifeng Hao, S Yu Philip, and Lifang He. Adaptive feature imputation with latent graph for deep incomplete multi-view clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14633-14641, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 286, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 286, + 325 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 286, + 325 + ], + "type": "text", + "content": "[30] Yuming Shen, Ziyi Shen, Menghan Wang, Jie Qin, Philip Torr, and Ling Shao. You never cluster alone. Advances in Neural Information Processing Systems, 34:27734-27746, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 286, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 286, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 286, + 370 + ], + "type": "text", + "content": "[31] Yuan Sun, Yang Qin, Yongxiang Li, Dezhong Peng, Xi Peng, and Peng Hu. Robust multi-view clustering with noisy correspondence. IEEE Transactions on Knowledge and Data Engineering, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 372, + 286, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 286, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 286, + 426 + ], + "type": "text", + "content": "[32] Huayi Tang and Yong Liu. Deep safe multi-view clustering: Reducing the risk of clustering performance degradation caused by view increase. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 202-211, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 286, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 286, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 286, + 459 + ], + "type": "text", + "content": "[33] Jingjing Tang, Qingqing Yi, Saiji Fu, and Yingjie Tian. Incomplete multi-view learning: Review, analysis, and prospects. Applied Soft Computing, page 111278, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 286, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 286, + 514 + ], + "type": "text", + "content": "[34] Xinhang Wan, Jiyuan Liu, Xinbiao Gan, Xinwang Liu, Siwei Wang, Yi Wen, Tianjiao Wan, and En Zhu. One-step multiview clustering with diverse representation. IEEE Transactions on Neural Networks and Learning Systems, pages 1-13, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 286, + 557 + ], + "type": "text", + "content": "[35] Xinhang Wan, Bin Xiao, Xinwang Liu, Jiyuan Liu, Weixuan Liang, and En Zhu. Fast continual multi-view clustering with incomplete views. IEEE Transactions on Image Processing, 33:2995-3008, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "type": "text", + "content": "[36] Haoqing Wang, Xun Guo, Zhi-Hong Deng, and Yan Lu. Rethinking minimal sufficient representation in contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16041-16050, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 286, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 657 + ], + "type": "text", + "content": "[37] Qianqian Wang, Zhengming Ding, Zhiqiang Tao, Quanxue Gao, and Yun Fu. Generative partial multi-view clustering with adaptive fusion and cycle consistency. IEEE Transactions on Image Processing, 30:1771-1783, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[38] Siwei Wang, Xinwang Liu, Suyuan Liu, Jiaqi Jin, Wenxuan Tu, Xinzhong Zhu, and En Zhu. Align then fusion: Generalized large-scale multi-view clustering with anchor matching correspondences. Advances in Neural Information Processing Systems, 35:5882-5895, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[39] Yiming Wang, Dongxia Chang, Zhiqiang Fu, Jie Wen, and Yao Zhao. Incomplete multi-view clustering via cross-view relation transfer. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[40] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 215 + ], + "type": "text", + "content": "[41] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487. PMLR, 2016. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 217, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 249 + ], + "type": "text", + "content": "[42] Cai Xu, Ziyu Guan, Wei Zhao, Hongchang Wu, Yunfei Niu, and Beilei Ling. Adversarial incomplete multi-view clustering. In IJCAI, pages 3933-3939, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 250, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 304 + ], + "type": "text", + "content": "[43] Jie Xu, Chao Li, Yazhou Ren, Liang Peng, Yujie Mo, Xiaoshuang Shi, and Xiaofeng Zhu. Deep incomplete multiview clustering via mining cluster complementarity. In Proceedings of the AAAI conference on artificial intelligence, pages 8761-8769, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 304, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 358 + ], + "type": "text", + "content": "[44] Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiaofeng Zhu, and Lifang He. Multi-level feature learning for contrastive multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16051-16060, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 360, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 414 + ], + "type": "text", + "content": "[45] Jie Xu, Chao Li, Liang Peng, Yazhou Ren, Xiaoshuang Shi, Heng Tao Shen, and Xiaofeng Zhu. Adaptive feature projection with distribution alignment for deep incomplete multiview clustering. IEEE Transactions on Image Processing, 32:1354-1366, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 415, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 469 + ], + "type": "text", + "content": "[46] Xiaoqiang Yan, Zhixiang Jin, Fengshou Han, and Yangdong Ye. Differentiable information bottleneck for deterministic multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27435-27444, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 525 + ], + "type": "text", + "content": "[47] Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, and Xi Peng. Partially view-aligned representation learning with noise-robust contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1134–1143, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 525, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 580 + ], + "type": "text", + "content": "[48] Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, and Xi Peng. Partially view-aligned representation learning with noise-robust contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1134–1143, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 581, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 545, + 624 + ], + "type": "text", + "content": "[49] Mouxing Yang, Yunfan Li, Peng Hu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Robust multi-view clustering with incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1055-1069, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "text", + "content": "[50] Shengju Yu, Suyuan Liu, Siwei Wang, et al. Sparse low-rank multi-view subspace clustering with consensus anchors and unified bipartite graph. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[51] Shengju Yu, Siwei Wang, Zhibin Dong, et al. A nonparametric graph clustering framework for multi-view data. In Proceedings of the AAAI conference on artificial intelligence, pages 16558-16567, 2024. 4" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5080" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 352 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[52] Pengxin Zeng, Mouxing Yang, Yiding Lu, Changqing Zhang, Peng Hu, and Xi Peng. Semantic invariant multiview clustering with fully incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2139-2150, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[53] Changqing Zhang, Yajie Cui, Zongbo Han, Joey Tianyi Zhou, Huazhu Fu, and Qinghua Hu. Deep partial multi-view learning. IEEE transactions on pattern analysis and machine intelligence, 44(5):2402-2415, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "text", + "content": "[54] Yi Zhang, Xinwang Liu, Siwei Wang, Jiyuan Liu, Sisi Dai, and En Zhu. One-stage incomplete multi-view clustering via late fusion. In Proceedings of the 29th ACM international conference on multimedia, pages 2717-2725, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 274 + ], + "type": "text", + "content": "[55] Yi Zhang, Fengyu Tian, Chuan Ma, Miaomiao Li, Hengfu Yang, Zhe Liu, En Zhu, and Xinwang Liu. Regularized instance weighting multiview clustering via late fusion alignment. IEEE Transactions on Neural Networks and Learning Systems, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 308 + ], + "type": "text", + "content": "[56] Huasong Zhong, Chong Chen, Zhongming Jin, and Xian-Sheng Hua. Deep robust clustering by contrastive learning. arXiv preprint arXiv:2008.03030, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "text", + "content": "[57] Lihua Zhou, Guowang Du, Kevin Lu, Lizheng Wang, and Jingwei Du. A survey and an empirical evaluation of multiview clustering approaches. ACM Computing Surveys, 56 (7):1-38, 2024. 1" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5081" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/14854e80-d5af-4094-8f56-d57259d0057b_content_list.json b/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/14854e80-d5af-4094-8f56-d57259d0057b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..81e11a26364b6d3342e3fa626e72327512aa5fd8 --- /dev/null +++ b/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/14854e80-d5af-4094-8f56-d57259d0057b_content_list.json @@ -0,0 +1,1588 @@ +[ + { + "type": "text", + "text": "InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment", + "text_level": 1, + "bbox": [ + 127, + 128, + 870, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunhong Lu $^{1}$ Qichao Wang $^{1}$ Hengyuan Cao $^{1}$ Xierui Wang $^{1}$ Xiaoyin Xu $^{1}$ Min Zhang $^{1,2*}$ $^{1}$ Zhejiang University $^{2}$ Shanghai Institute for Advanced Study-Zhejiang University \n{yunhonglu,qichaowang,caohy,sherrywang,xiaoyinxu,min_zhang}@zju.edu.cn", + "bbox": [ + 101, + 202, + 915, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 277, + 291, + 354, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Without using explicit reward, direct preference optimization (DPO) employs paired human preference data to fine-tune generative models, a method that has garnered considerable attention in large language models (LLMs). However, exploration of aligning text-to-image (T2I) diffusion models with human preferences remains limited. In comparison to supervised fine-tuning, existing methods that align diffusion model suffer from low training efficiency and subpar generation quality due to the long Markov chain process and the intractability of the reverse process. To address these limitations, we introduce DDIM-InPO, an efficient method for direct preference alignment of diffusion models. Our approach conceptualizes diffusion model as a single-step generative model, allowing us to fine-tune the outputs of specific latent variables selectively. In order to accomplish this objective, we first assign implicit rewards to any latent variable directly via a reparameterization technique. Then we construct an Inversion technique to estimate appropriate latent variables for preference optimization. This modification process enables the diffusion model to only fine-tune the outputs of latent variables that have a strong correlation with the preference dataset. Experimental results indicate that our DDIM-InPO achieves state-of-the-art performance with just 400 steps of fine-tuning, surpassing all preference aligning baselines for T2I diffusion models in human preference evaluation tasks.", + "bbox": [ + 88, + 323, + 485, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 744, + 220, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the domain of T2I generation, tailoring image outputs to reflect user preferences is still in an early stage of development. In contrast, large language models (LLMs) [1] have made notable progress in generating human-aligned text, following a two-stage process: first, pre-training on extensively collected large-scale datasets and then, aligning on smaller datasets that capture specific human needs.", + "bbox": [ + 89, + 771, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This alignment aims to produce outputs that meet human preferences while retaining model's learned capabilities. In recent years, T2I diffusion models [18, 43, 52] have been at the forefront of image generation. Nevertheless, achieving alignment between model outputs and human preferences remains a significant challenge.", + "bbox": [ + 511, + 292, + 906, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, researchers have developed several methods tailoring diffusion models to human preferences, yet these approaches generally lack effectiveness. Some methods adopt a two-stage training approach, where a reward function is first trained and then followed by updating the diffusion model using gradient-based back-propagation [12, 45]. These approaches tend to be unstable and are susceptible to reward leakage [75]. Alternatively, RL (reinforcement learning)-based methods model the denoising process as a Markov decision process, with online evaluation of generated images based on reward feedback [6]. However, their effectiveness is constrained by the length of the Markov chain. Other approaches such as [32, 63], inspired by direct preference optimization (DPO) [48], reduce the RL objective to an optimal policy training goal. Nevertheless, due to the influence of sparse rewards [23, 69], these methods' performance is limited, especially when handling out-of-distribution inputs.", + "bbox": [ + 511, + 385, + 908, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these issues, we introduce DDIM-InPO, an efficient approach for direct preference optimization of diffusion models. In response to the challenge of the implicit rewards allocation in the long-chain denoising process, our key idea is to treat the diffusion model as a single-step generation framework, only fine-tuning the outputs of latent variables directly linked to the target image. To achieve this, we first link the latent variables to the target image space in a single step, using a reparameterization method. It allows us to allocate implicit rewards to latent variables at any timestep directly. Secondly, to mitigate the effects of sparse rewards, we use an Inversion technique in the target image space to identify images that are highly correlated with the target and obtain the corresponding latent variables. Finally, we focus on optimizing the outputs of this limited set of latent variables and thus achieve rapid computations, enabling", + "bbox": [ + 511, + 659, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 114, + 887, + 228, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28629", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2b1019296cc1256e9ede8ffa9871198271fdce86b8b048dba91404764f84bbf.jpg", + "image_caption": [ + "Figure 1. We develop DDIM-InPO, an efficient method to align diffusion models with human preference. It suffices to directly optimize the outputs of a small set of variables using human feedback data. This figure illustrates the results after 400 fine-tuning steps of SDXL-base-1.0 using our method, showing that the generated images exhibit strong visual appeal and align well with human preferences." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 903, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "high-quality optimization of the diffusion model. Our main contributions are three-folds:", + "bbox": [ + 89, + 545, + 482, + 573 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel DPO implicit rewards allocation method for the diffusion model, treating the diffusion model as a single-step generative model.", + "- We introduce DDIM-InPO, an efficient framework for direct preference optimization of diffusion models. By applying an Inversion technique in the target image space, we estimate images highly correlated with the preference dataset and derive the corresponding latent variables sufficient for fine-tuning.", + "- In our experiments, we benchmark our method against state-of-the-art fine-tuning approaches for human preference, demonstrating enhanced training efficiency (refer to Fig. 4) and improved generation quality (examples shown in Fig. 1 and Fig. 3)." + ], + "bbox": [ + 89, + 575, + 482, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 800, + 240, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Image Generative Models. T2I generation models aim to produce high-quality, high-fidelity images from a given text prompt. Additionally, these models have contributed to advancements in related fields such as image editing [29], video generation [9], and 3D modeling [44].", + "bbox": [ + 89, + 825, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Among T2I generative models [14, 49-51, 70], diffusion models have recently gained significant attention due to their open-source nature and superior generation performance. However, high-quality diffusion models are typically trained on large, noisy datasets, which often leads to deviations from human preferences, an area that still requires further refinement.", + "bbox": [ + 511, + 545, + 906, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models Alignment. In diffusion models, supervised fine-tuning (SFT) is commonly employed to align them with human preferences. Inspired by RL-based fine-tuning methods used in LLMs [2, 19, 28, 56, 60], various reward models such as PickScore [31] and HPSv2 [65] have been developed to simulate human preferences. With reward models, approaches like DRaFT [12] and AlignProp [45] use differentiable rewards and back-propagate through the diffusion sampling process, which demands substantial memory and would result in reward leakage. DPOK [20] and DDPO [6] model the sampling process of diffusion models as a Markov decision process and apply policy gradient methods for fine-tuning. Diffusion-DPO [63] and D3PO [68] adopt direct preference optimization by leveraging the denoising steps of diffusion models to optimize", + "bbox": [ + 511, + 674, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28630", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model parameters at each step. In the DPO framework, DenseReward [69] allocates higher weights to the initial steps. Going further, Diffusion-KTO [32] requires only binary feedback signals for each image to fine-tune the model. However, all these methods are limited to expanding the decision space by incorporating the reward function over all denoising steps, which results in sparse rewards and inefficient training.", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 224, + 230, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 248, + 259, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models involve a forward process $\\{q(\\pmb{x}_t)\\}_{t\\in [0,T]}$ that incrementally adds noise to data $\\pmb{x}_0\\sim q(\\pmb{x}_0)$ and a learned reverse process $\\{p(\\pmb{x}_t)\\}_{t\\in [0,T]}$ that aims to denoise the data. The forward process is defined as $q(\\pmb{x}_t|\\pmb{x}_0)\\coloneqq \\mathcal{N}(\\sqrt{\\alpha_t}\\pmb{x}_0,(1 - \\alpha_t)\\mathbf{I})$ and $q(\\pmb{x}_t)\\coloneqq \\int q(\\pmb{x}_t|\\pmb{x}_0)q(\\pmb{x}_0)\\mathrm{d}\\pmb{x}_0,$ where $\\alpha_{t}$ is a noise schedule (as defined in [61]). Starting from $p(\\pmb{x}_T)\\coloneqq \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , the reverse process is defined by a parameterized denoiser $\\epsilon_{\\theta}^{t}(\\pmb{x}_{t})$ , which aims to predict noise added to $\\pmb{x}_0$ . The denoiser is optimized by minimizing:", + "bbox": [ + 89, + 271, + 483, + 407 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D M}} := \\mathbb {E} _ {x _ {0}, t, \\epsilon} [ w (t) \\| \\epsilon_ {\\theta} ^ {t} (\\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon) - \\epsilon \\| _ {2} ^ {2} ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 415, + 480, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{x}_0\\sim q(\\pmb{x}_0),t\\sim \\mathcal{U}(0,T),\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and $w(t)$ is a pre-specified weight function.", + "bbox": [ + 89, + 450, + 482, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising Diffusion Implicit Models (DDIM) generate samples by deterministically reversing the diffusion process. By reparameterizing noisy image $\\bar{\\pmb{x}}_t = \\pmb{x}_t / \\sqrt{\\alpha_t}$ , the denoising process can be viewed as an ODE:", + "bbox": [ + 89, + 479, + 483, + 541 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\mathrm {d} \\bar {\\boldsymbol {x}} _ {t}}{\\mathrm {d} t} = \\epsilon_ {\\theta} ^ {t} \\left(\\frac {\\bar {\\boldsymbol {x}} _ {t}}{\\sqrt {\\sigma_ {t} ^ {2} + 1}}\\right) \\frac {\\mathrm {d} \\sigma_ {t}}{\\mathrm {d} t} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 549, + 482, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma_{t} = \\sqrt{1 - \\alpha_{t}} / \\sqrt{\\alpha_{t}}$ and $\\bar{x}_T$ is sampled from Gaussian distribution. The DDIM ODE can also be integrated in reverse to estimate $\\pmb{x}_t$ at any timestep $t$ starting from a clean image $\\pmb{x}_0$ , a process known as DDIM inversion [39].", + "bbox": [ + 89, + 599, + 483, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Human Preference Optimization", + "text_level": 1, + "bbox": [ + 89, + 667, + 377, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preference optimization involves fine-tuning generative models to produce samples more closely align with human preferences. Assuming that human preferences are represented by an implicit reward function $r^*$ , there is a dataset $\\mathcal{D} = \\{(x_0^w, x_0^l, c)\\}$ , where $c$ denotes the condition, and $x_0^w$ and $x_0^l$ represent the winner and loser samples, respectively.", + "bbox": [ + 89, + 691, + 483, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RLHF. RLHF employs maximum likelihood estimation to fit the Bradley-Terry (BT) model on the dataset $\\mathcal{D}$ , thereby obtaining an explicit parameterized reward model $r_{\\phi}$ .", + "bbox": [ + 89, + 799, + 483, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {B T}} (\\phi) := - \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right) \\sim \\mathcal {D}} [ \\log \\sigma \\left(r _ {\\phi} \\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {c}\\right) - r _ {\\phi} \\left(\\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right)\\right) ] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 868, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the reward, the generative distribution $p_{\\theta}$ can then be optimized as a policy using RL that maximizes reward feedback, with an regularization term for the KL-divergence from the reference distribution $p_{\\mathrm{ref}}$ :", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {p _ {\\theta}} \\mathbb {E} _ {\\substack {\\boldsymbol {c} \\sim \\mathcal {D} _ {\\boldsymbol {c}} \\\\ \\boldsymbol {x} _ {0} \\sim p _ {\\theta} (\\boldsymbol {x} _ {0} | \\boldsymbol {c})}} [ r (\\boldsymbol {x} _ {0}, \\boldsymbol {c}) ] - \\beta \\mathbb {D} _ {\\mathrm {K L}} [ p _ {\\theta} (\\boldsymbol {x} _ {0} | \\boldsymbol {c}) | | p _ {\\text {ref}} (\\boldsymbol {x} _ {0} | \\boldsymbol {c}) ] \\tag{4}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 160, + 906, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta$ is a hyperparameter that controls regularization.", + "bbox": [ + 513, + 199, + 887, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Direct Preference Optimization. Direct Preference Optimization (DPO) considers the unique global optimal policy $p_{\\theta}^{*}$ under the given reward function $r$ in Eq. (4):", + "bbox": [ + 511, + 232, + 905, + 277 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} ^ {*} \\left(\\boldsymbol {x} _ {0} \\mid \\boldsymbol {c}\\right) = \\frac {p _ {\\text {r e f}} \\left(\\boldsymbol {x} _ {0} \\mid \\boldsymbol {c}\\right) \\exp \\left(r \\left(\\boldsymbol {x} _ {0} , \\boldsymbol {c}\\right) / \\beta\\right)}{\\sum_ {\\boldsymbol {x} _ {0}} p _ {\\text {r e f}} \\left(\\boldsymbol {x} _ {0} \\mid \\boldsymbol {c}\\right) \\exp \\left(r \\left(\\boldsymbol {x} _ {0} , \\boldsymbol {c}\\right) / \\beta\\right)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 286, + 905, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, the reward function $r(\\pmb{x}_0, \\pmb{c})$ is rewritten according to Eq. (5) and inserted into the BT model from Eq. (3), resulting in the DPO objective function:", + "bbox": [ + 511, + 330, + 905, + 368 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D P O}} := - \\underset {\\substack {(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l} \\\\ , \\boldsymbol {c}) \\sim \\mathcal {D}}} {\\mathbb {E}} \\log \\sigma \\left(\\beta \\log \\frac {p _ {\\theta} (\\boldsymbol {x} _ {0} ^ {w} | \\boldsymbol {c})}{p _ {\\text {ref}} (\\boldsymbol {x} _ {0} ^ {w} | \\boldsymbol {c})} - \\beta \\log \\frac {p _ {\\theta} (\\boldsymbol {x} _ {0} ^ {l} | \\boldsymbol {c})}{p _ {\\text {ref}} (\\boldsymbol {x} _ {0} ^ {l} | \\boldsymbol {c})}\\right) \\tag{6}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 372, + 903, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on Eqs. (4) and (6), Wallace et al. [63] redistribute rewards $r(\\pmb{x}_0, c)$ across all possible diffusion paths $p_{\\theta}(\\pmb{x}_{1:T} | \\pmb{x}_0, c)$ and minimize the KL-divergence of the joint probability density $\\mathbb{D}_{\\mathrm{KL}}(p_{\\theta}(\\pmb{x}_{0:T} | c) || p_{\\mathrm{ref}}(\\pmb{x}_{0:T} | c))$ , thus designing the DPO objective for diffusion models:", + "bbox": [ + 511, + 419, + 905, + 494 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D P O - D i f f u s i o n}} := - \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right) \\sim \\mathcal {D}} \\log \\sigma\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 503, + 784, + 522 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\left(\\beta \\mathbb {E} _ {\\substack {\\boldsymbol {x} _ {1: T} ^ {w} \\sim p _ {\\theta} ^ {c} (\\boldsymbol {x} _ {1: T} ^ {w} \\mid \\boldsymbol {x} _ {0} ^ {w}) \\\\ \\boldsymbol {x} _ {1: T} ^ {l} \\sim p _ {\\theta} ^ {c} (\\boldsymbol {x} _ {1: T} ^ {l} \\mid \\boldsymbol {x} _ {0} ^ {l})}} \\left[ \\log \\frac {p _ {\\theta} ^ {c} (\\boldsymbol {x} _ {0 : T} ^ {w})}{p _ {\\text {ref}} ^ {c} (\\boldsymbol {x} _ {0 : T} ^ {w})} - \\log \\frac {p _ {\\theta} ^ {c} (\\boldsymbol {x} _ {0 : T} ^ {l})}{p _ {\\text {ref}} ^ {c} (\\boldsymbol {x} _ {0 : T} ^ {l})} \\right]\\right) \\right. \\tag{7}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 525, + 905, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p_{\\theta}^{c}(\\cdot)$ represents $p_{\\theta}(\\cdot | c)$ for compactness. Without the need for RL algorithms, the DPO-based method directly optimizes the optimal distribution $p_{\\theta}$ .", + "bbox": [ + 511, + 590, + 905, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Inversion Preference Optimization", + "text_level": 1, + "bbox": [ + 511, + 648, + 826, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a fixed dataset $\\mathcal{D} = \\{(\\pmb{x}_0^w, \\pmb{x}_0^l, \\pmb{c})\\}$ , where each sample includes a prompt $c$ and an image pair generated by reference model $p_{\\mathrm{ref}}$ labeled as $\\pmb{x}_0^w \\succ \\pmb{x}_0^l$ , our goal is to train a model $p_\\theta$ that aligns with human preferences. The primary challenge lies in how to match the parameterized distribution $p_\\theta^c(\\pmb{x}_0)$ according to Eq. (6). Our insight is that the diffusion model can be conceptualized as a timestep-aware single-step generative model $p_\\theta^c(\\pmb{x}_0|\\pmb{x}_t)$ . Specifically, there exists some latent variables $\\pmb{x}_t$ that can generate $\\pmb{x}_0$ in a single step for any timestep $t$ , which can be written mathematically as: $p_\\theta^c(\\pmb{x}_0) = \\int p_\\theta^c(\\pmb{x}_0|\\pmb{x}_t)p_\\theta^c(\\pmb{x}_t)\\mathrm{d}\\pmb{x}_t$ . Thus, the first step is to determine the relationship between $\\pmb{x}_t$ and $\\pmb{x}_0$ in Sec. 4.1, and the second step is to identify appropriate latent variables $\\pmb{x}_t$ for any timestep in Sec. 4.2, within the DPO framework.", + "bbox": [ + 511, + 672, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28631", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Reparametrized DDIM for DPO", + "text_level": 1, + "bbox": [ + 89, + 90, + 375, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We establish a relationship between $\\boldsymbol{x}_t$ and $\\boldsymbol{x}_0$ to be used within the DPO framework. Instead of viewing DDIM as a denoising process within the noisy image $\\boldsymbol{x}_t$ space, we reparameterize it with 'initial' variable:", + "bbox": [ + 89, + 112, + 483, + 172 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {0} (t) := \\bar {\\boldsymbol {x}} _ {t} - \\sigma_ {t} \\epsilon_ {\\theta} ^ {t} (\\boldsymbol {x} _ {t}, \\boldsymbol {c}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 183, + 482, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where noisy image $\\bar{\\pmb{x}}_t = \\pmb {x}_t / \\sqrt{\\alpha_t}$ . In this context, $\\pmb {x}_0(t)$ is the image obtained after the single-step denoising at timestep $t$ and is in $\\pmb {x}_0$ space. Therefore, $p_{\\theta}^{c}(\\pmb {x}_0,\\pmb {x}_t)$ is well-defined, allowing us to present a formal definition of 'initial' reward $r(\\pmb {x}_0,\\pmb {c})$ with the joint' reward $r_t^c (\\pmb {x}_0,\\pmb {x}_t)$ ..", + "bbox": [ + 89, + 210, + 483, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr \\left(\\boldsymbol {x} _ {0}, \\boldsymbol {c}\\right) = \\mathbb {E} _ {p _ {\\theta} ^ {c} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right)} \\left[ r _ {t} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}\\right) \\right] \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 301, + 482, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to [63], we minimize the joint KL divergence $\\mathbb{D}_{\\mathrm{KL}}[p_{\\theta}^{c}(\\boldsymbol{x}_0,\\boldsymbol{x}_t)||p_{\\mathrm{ref}}^{c}(\\boldsymbol{x}_0,\\boldsymbol{x}_t)]$ to serve as an upper bound for the KL-divergence term in Eq. (4). With Eq. (9) and KL-divergence bound, we can reformulate Eq. (4) into the objective below:", + "bbox": [ + 89, + 323, + 483, + 398 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {p _ {\\theta} ^ {\\mathcal {C}}} \\mathbb {E} _ {\\substack {t, (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) \\sim \\\\ p _ {\\theta} ^ {\\mathcal {C}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t})}} [ r _ {t} ^ {\\mathcal {C}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) ] - \\beta \\mathbb {D} _ {\\mathrm {K L}} [ p _ {\\theta} ^ {\\mathcal {C}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) | | p _ {\\text {ref}} ^ {\\mathcal {C}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) ] \\tag{10}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 410, + 501, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this context, we consider all timesteps. This objective enables the effective allocation of $x_0$ 's reward across any timestep within the model, enhancing training efficiency. In a manner analogous to Eq. (3) through Eq. (6), we derive the following optimization objective:", + "bbox": [ + 89, + 450, + 483, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) := - \\mathbb {E} _ {t, \\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right) \\sim \\mathcal {D}} \\log \\sigma\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 537, + 302, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\beta \\mathbb {E} _ {\\substack {\\boldsymbol {x} _ {t} ^ {w} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {t} ^ {w} | \\boldsymbol {x} _ {0} ^ {w}) \\\\ \\boldsymbol {x} _ {t} ^ {l} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {t} ^ {l} | \\boldsymbol {x} _ {0} ^ {l})}} \\left[ \\log \\frac {p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {0} ^ {w} , \\boldsymbol {x} _ {t} ^ {w})}{p _ {\\text {ref}} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {0} ^ {w} , \\boldsymbol {x} _ {t} ^ {w})} - \\log \\frac {p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {0} ^ {l} , \\boldsymbol {x} _ {t} ^ {l})}{p _ {\\text {ref}} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {0} ^ {l} , \\boldsymbol {x} _ {t} ^ {l})} \\right] \\right. \\tag{11}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 558, + 483, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To optimize Eq. (11), we need to sample $\\pmb{x}_t \\sim p_\\theta^c(\\pmb{x}_t | \\pmb{x}_0)$ , where $\\pmb{x}_t$ should generate an approximation of $\\pmb{x}_0$ in a single step.", + "bbox": [ + 89, + 626, + 482, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Preference Optimization via Inversion", + "text_level": 1, + "bbox": [ + 89, + 680, + 419, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "So how can we determine whether $\\pmb{x}_t$ can generate an approximation of $\\pmb{x}_0$ in a single step? Given $\\pmb{x}_0(t)$ , it remains unclear how to calculate $\\pmb{x}_t$ . From Eq. (8), $\\pmb{x}_t$ should satisfy the following nonlinear equation:", + "bbox": [ + 89, + 703, + 483, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t} = \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} (t) + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} ^ {t} (\\boldsymbol {x} _ {t}, c) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 773, + 482, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and we can rewrite Eq. (12) with respect to noise:", + "bbox": [ + 89, + 801, + 419, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon = \\epsilon_ {\\theta} ^ {t} \\left(\\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} (t) + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\boldsymbol {c}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 827, + 482, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon = (\\pmb{x}_t - \\sqrt{\\alpha_t}\\pmb{x}_0(t)) / \\sqrt{1 - \\alpha_t}$ . Thus, we can define $\\delta_t(\\pmb{x}_0(t)) = \\epsilon$ as the solution when given $\\pmb{x}_0(t)$ . In this context, $\\pmb{x}_0(t)$ serves as a reliable approximation of $\\pmb{x}_0$ at", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/00fd003baa5f2ab149e49f73aa49efdbde4e2cca564c325d41ba134260c24d8d.jpg", + "image_caption": [ + "Figure 2. Illustration of Inversion for Preference Optimization." + ], + "image_footnote": [], + "bbox": [ + 522, + 88, + 901, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "timestep $t$ , that is, $x_0(t) \\approx x_0$ . Additionally, we rewrite Eq. (2) by applying the definition of $x_0(t)$ (Eq. (8)) by employing some algebra:", + "bbox": [ + 511, + 386, + 903, + 431 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\mathrm {d} \\boldsymbol {x} _ {0} (t)}{\\mathrm {d} t} = - \\sigma_ {t} \\frac {\\mathrm {d}}{\\mathrm {d} t} \\epsilon_ {\\theta} ^ {t} \\left(\\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} (t) + \\sqrt {1 - \\alpha_ {t}} \\delta_ {t} (\\boldsymbol {x} _ {0} (t)), \\boldsymbol {c}\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 444, + 903, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the work of [36, 39], we suggest to derive $\\delta_t(\\pmb{x}_0(t))$ by DDIM inversion (It should be noted that $\\delta_t(\\pmb{x}_0(t))$ can be estimated using various methods, including but not limited to, inversion.) and obtain $\\pmb{x}_0(t)$ by inverting Eq. (14) consequently. To clarify the process, we consider the consecutive time points $t$ and $t + \\kappa > t$ , and the discretized form of inversion is given as follows:", + "bbox": [ + 511, + 487, + 906, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\pmb {x} _ {0} (t + \\kappa) = \\pmb {x} _ {0} (t) - \\sigma_ {t + \\kappa} [ \\epsilon_ {\\theta} ^ {t + \\kappa} (\\sqrt {\\alpha_ {t + \\kappa}} \\pmb {x} _ {0} (t) \\\\ \\left. + \\sqrt {1 - \\alpha_ {t + \\kappa}} \\delta_ {t} (\\boldsymbol {x} _ {0} (t))) - \\delta_ {t} (\\boldsymbol {x} _ {0} (t)) \\right] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 604, + 849, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to Eq. (12), $\\pmb{x}_t$ is well-defined as follows, and so is $p_{\\theta}^{c}(\\pmb{x}_{t}|\\pmb{x}_{0})$ , ensuring that Eq. (11) can be optimized.", + "bbox": [ + 511, + 657, + 903, + 688 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t} = \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} (t) + \\sqrt {1 - \\alpha_ {t}} \\delta_ {t} (\\boldsymbol {x} _ {0} (t)) \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 702, + 903, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the sake of simplicity, we approximate $p_{\\mathrm{ref}}^{\\mathbf{c}}(\\pmb{x}_0,\\pmb{x}_t)$ with $p_{\\mathrm{ref}}^{\\mathbf{c}}(\\pmb{x}_0|\\pmb{x}_t)p_\\theta^{\\mathbf{c}}(\\pmb{x}_t)$ . Using Jensen's inequality and some algebra, we approximate Eq. (11) as:", + "bbox": [ + 511, + 733, + 903, + 779 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\theta) = - \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right) \\sim \\mathcal {D}, t \\sim \\mathcal {U} (0, T),} \\quad \\log \\sigma (- \\beta ( \\\\ \\boldsymbol {x} _ {t} ^ {w} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {t} ^ {w} \\mid \\boldsymbol {x} _ {0} ^ {w}\\right), \\boldsymbol {x} _ {t} ^ {l} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {t} ^ {l} \\mid \\boldsymbol {x} _ {0} ^ {l}\\right) \\\\ + \\mathbb {D} _ {\\mathrm {K L}} \\left(q ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {w} \\mid \\boldsymbol {x} _ {t} ^ {w}, \\boldsymbol {x} _ {0} ^ {w} (t)\\right) \\| p _ {\\theta} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {w} \\mid \\boldsymbol {x} _ {t} ^ {w}\\right)\\right) \\\\ - \\mathbb {D} _ {\\mathrm {K L}} \\left(q ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {w} \\mid \\boldsymbol {x} _ {t} ^ {w}, \\boldsymbol {x} _ {0} ^ {w} (t)\\right) \\| p _ {\\text {r e f}} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {w} \\mid \\boldsymbol {x} _ {t} ^ {w}\\right)\\right) \\tag {17} \\\\ - \\mathbb {D} _ {\\mathrm {K L}} \\left(q ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {l} \\mid \\boldsymbol {x} _ {t} ^ {l}, \\boldsymbol {x} _ {0} ^ {l} (t)\\right) \\| p _ {\\theta} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {l} \\mid \\boldsymbol {x} _ {t} ^ {l}\\right)\\right) \\\\ + \\mathbb {D} _ {\\mathrm {K L}} (q ^ {\\boldsymbol {c}} (\\pmb {x} _ {0} ^ {l} | \\pmb {x} _ {t} ^ {l}, \\pmb {x} _ {0} ^ {l} (t)) \\| p _ {\\mathrm {r e f}} ^ {\\boldsymbol {c}} (\\pmb {x} _ {0} ^ {l} | \\pmb {x} _ {t} ^ {l}))) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 792, + 903, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28632", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a7bf779cbd38521276e37c76d6876ffbb3abc28e8526cc30d4330880a42756f6.jpg", + "image_caption": [ + "Figure 3. Top, in human evaluations, InPO-SDXL shows a marked improvement over both DPO-SDXL and Base-SDXL. Bottom, qualitative comparisons among baselines. InPO-SDXL achieves superior prompt alignment and produces images of higher quality." + ], + "image_footnote": [], + "bbox": [ + 89, + 87, + 903, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q^c (\\cdot |\\cdot)$ is a Gaussian transition kernel. Using Eq.(16), the above loss can be simplified as follows:", + "bbox": [ + 89, + 561, + 482, + 593 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\theta) = - \\mathbb {E} _ {(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}) \\sim \\mathcal {D}, t \\sim \\mathcal {U} (0, T),} \\quad \\log \\sigma (- \\beta w (t) ( \\\\ \\boldsymbol {x} _ {t} ^ {w} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {t} ^ {w} | \\boldsymbol {x} _ {0} ^ {w}), \\boldsymbol {x} _ {t} ^ {l} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {t} ^ {l} | \\boldsymbol {x} _ {0} ^ {l}) \\\\ \\left\\| \\tau_ {t} ^ {w} - \\epsilon_ {\\theta} ^ {t} \\left(\\boldsymbol {x} _ {t} ^ {w}, \\boldsymbol {c}\\right) \\right\\| _ {2} ^ {2} - \\left\\| \\tau_ {t} ^ {w} - \\epsilon_ {\\mathrm {r e f}} ^ {t} \\left(\\boldsymbol {x} _ {t} ^ {w}, \\boldsymbol {c}\\right) \\right\\| _ {2} ^ {2} \\\\ - \\left\\| \\tau_ {t} ^ {l} - \\epsilon_ {\\theta} ^ {t} \\left(\\boldsymbol {x} _ {t} ^ {l}, \\boldsymbol {c}\\right) \\right\\| _ {2} ^ {2} + \\left\\| \\tau_ {t} ^ {l} - \\epsilon_ {\\text {r e f}} ^ {t} \\left(\\boldsymbol {x} _ {t} ^ {l}, \\boldsymbol {c}\\right) \\right\\| _ {2} ^ {2}) \\tag {18} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 604, + 480, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau_{t}^{*} = (\\pmb{x}_{0}^{*}(t) - \\pmb{x}_{0}^{*}) / \\sigma_{t} + \\delta_{t}(\\pmb{x}_{0}^{*}(t))$ (refer to Fig. 2) and $w(t)$ is a weight function defined consistent with [63].", + "bbox": [ + 89, + 708, + 482, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Link to Diffusion-DPO. Diffusion-DPO defines a reward function on the denoising trajectory, which expands the decision space because it requires combining the reward functions across all denoise steps. This problem is well known to be closely related to delayed feedback/sparse rewards in reinforcement learning, where effective feedback is only available after generating the entire trajectory. Within our framework, Diffusion-DPO utilizes a rough approximation of $\\delta_t$ which uses i.i.d random noise $\\delta_t^{\\mathrm{DPO}}(\\pmb{x}_0(t))\\sim$", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{N}(0,I)$ , matching diffusion incorporated in the forward process of noise addition.", + "bbox": [ + 511, + 561, + 906, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 606, + 645, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Setup", + "text_level": 1, + "bbox": [ + 511, + 631, + 594, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Models. We conduct all experiments on the Pick-a-Pic v2 [31] dataset, which is collected from the Pick-a-Pic web application (refer to [31] for details). Consistent with the setting in [63], approximately $12\\%$ of the data is excluded, where preferences cannot be determined. As a result, we obtain 851,293 data pairs and 58,960 unique prompts. Our method is applied to fine-tune the base models Stable Diffusion 1.5 (SD1.5) [52] and Stable Diffusion XL-base-1.0 (SDXL) [43].", + "bbox": [ + 511, + 652, + 906, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Basic Hyperparameters. To ensure a fair comparison of training efficiency, we use hyperparameters largely consistent with those in [63]. We employ AdamW [35] as the optimizer for SD1.5 and Adafactor [57] for SDXL. All experiments are conducted on eight NVIDIA H800 GPUs, with each GPU handling a batch size of one pair and gradient", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28633", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ac1d8b979b7185068161181af2fe64712756b3c394a7ee8534caba9121cbed45.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelBaselinesWin-rate (HPDv2)Win-rate (Parti-Prompts)
AestheticPickScoreHPSv2CLIPAestheticPickScoreHPSv2CLIP
InPO-SDXLvs. Base-SDXL56.3779.2585.3853.3761.7072.8978.3150.55
vs. SFT-SDXL66.6688.3189.9159.0069.7088.5488.9156.99
vs. DPO-SDXL55.8759.2865.8146.9157.1756.9260.1142.28
InPO-SD1.5vs. Base-SD1.580.1985.8490.2266.4474.6373.1681.8661.89
vs. SFT-SD1.556.6366.8458.9157.8752.2762.5058.0357.54
vs. DPO-SD1.570.5074.9484.0661.0968.0163.5475.3157.54
vs. KTO-SD1.559.8167.4160.9460.0359.0162.2558.7657.60
", + "bbox": [ + 94, + 88, + 903, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Win-rate comparison between DDIM-InPO and baselines. We use prompts from HPDv2 and Parti-Prompts, evaluated across different evaluators. SDXL and SD1.5 are used as base models. For simplicity, we denote models using a 'Method-Base' format, such as InPO-SDXL to represent the SDXL model fine-tuned with the DDIM-InPO method. Win-rates exceeding $50\\%$ are highlighted in bold.", + "bbox": [ + 89, + 257, + 906, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "accumulation over 128 steps, resulting in an effective batch size of 1024 pairs (half the batch size compared to [63]). The learning rate is set to $\\frac{2000}{\\beta} 2.048^{-8}$ , with a linear warm-up phase. For SD1.5 training, $\\beta$ is set to 2000, while for SDXL training, it is set to 5000.", + "bbox": [ + 89, + 325, + 485, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. We compare DDIM-InPO with existing baselines using user studies and automatic preference metrics. We use the following four evaluators to evaluate image quality: the LAION aesthetic classifier [55] to assess image aesthetic scores, CLIP [47] to measure text-image alignment, and PickScore [31] and HPSv2 [66] to predict human preferences for given image-text pairs. In test stage, we generated images using both the baseline and our model on the Parti-Prompts (1632 prompts) and HPDv2 (3200 prompts) test sets. The median and win-rate (how frequently does the evaluator prefer the generations from DDIM-InPO over those from baselines) of the aforementioned four evaluators are used as automatic preference metrics for comparison. Additionally, we conduct a user study (16 participants) to compare DDIM-InPO with the baselines. We design three questions as [63]: (1) Which image is your overall preferred choice? (2) Which image is more visually attractive? (3) Which image better matches the text description? We also compare the traininggpu hours (on H800) with baselines. In this work, we adopt the following baselines: base model (SD1.5 and SDXL), supervised fine-tuning (SFT), Diffusion-DPO, and Diffusion-KTO (for SD1.5).", + "bbox": [ + 91, + 426, + 483, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Primary Results", + "text_level": 1, + "bbox": [ + 89, + 771, + 253, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. Firstly, we use the quantitative evaluation method for model preferences outlined in [32, 63]. Tab. 1 presents the win-rates of the InPO-aligned diffusion models in comparison to the baselines. Overall, following fine-tuning with DDIM-InPO, both SD1.5 and SDXL achieve superior performance compared to the baselines across nearly all evaluators and test datasets, thereby vali", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a15ed25ff296813fa49bcd76607c730c5c108ed63ffd44b055f7e679d985b203.jpg", + "image_caption": [ + "Figure 4. Comparison of the trade-off between the quality of generated images and training efficiency following human preference optimization of SD1.5 on the HPDv2 test set. Sizes of the circles represent the volume of training data used. Our DDIM-InPO achieves superior performance, with a training speed that is 18.4 and 3.6 times faster than Diffusion-KTO [32] and Diffusion-DPO [63], respectively, while producing images of higher quality." + ], + "image_footnote": [], + "bbox": [ + 526, + 342, + 870, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dating the effectiveness of our method. Specifically, HPSv2 evaluation shows that, InPO-SDXL obtains a win-rate of $65.81\\%$ against DPO-SDXL, while InPO-SD1.5 achieves a win-rate of $60.94\\%$ compared to KTO-SD1.5 on the HPDv2 test set. Moreover, it significantly improves the alignment with human preferences for the base model. According to HPSv2, it achieves win-rate of up to $85.38\\%$ and $90.22\\%$ when compared to the SD1.5 and SDXL base models, respectively, demonstrating the efficacy of DDIM-InPO.", + "bbox": [ + 511, + 654, + 906, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It should be noted that our DDIM-InPO significantly advances training efficiency, with a training speed that is 18.4 and 3.6 times faster than Diffusion-KTO and Diffusion-DPO, respectively, as shown in Fig. 4, while producing higher-quality images. Additionally, Fig. 3 (top) presents the results of our user study, where we find that human evaluators prefer the model aligned with DDIM-InPO. The", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28634", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/140ec73286f04a450c31a1805c74c853302bd21870aeb76d4d4b1ff8fe998ff4.jpg", + "image_caption": [ + "Figure 5. Advantages of DDIM-InPO: Compared with Refinement technique and Diffusion-DPO, we find that images generated by DDIM-InPO exhibit enhanced light and spatial structure, better realism and detail capture, greater imaginative design, stable color consistency, optimized multi-instance layout and text integration in visuals. These are some hidden advantages aligned with human preferences. Prompts from left to right: (1) A small bird sitting in a metal wheel. (2) Trees seen through a car window on a rainy day. (3) Description, An artistic rendering of a cosmic portal with a beach at dusk on the other side. (4) A person in full samurai armor at the beach. (5) A man standing in front of a bunch of doughnuts. (6) A towel with the word 'cat' printed on it, simple and clear text." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 903, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "evaluation results clearly show that our method provides a more efficient way for fine-tuning diffusion models with human preferences, demonstrating the effectiveness of our proposed paradigm for learning specific latent variables.", + "bbox": [ + 88, + 545, + 486, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. We display qualitative comparison results of InPO-SDXL against several baselines (Base-SDXL, DPO-SDXL) in Fig. 3 (bottom). The images generated by InPO-SDXL demonstrate enhanced performance in both textual alignment and visual attractiveness. In addition, Fig. 5 illustrates the advantages of our model, which serves as a method for directly learning human preferences for diffusion models. The generated images showcase improved light and spatial structures, detailed capture, greater creativity, consistent color, organized multi-instance layouts, and text-visual integration. These advantages align closely with the preferences of users.", + "bbox": [ + 88, + 626, + 486, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. AI Preference Optimization", + "text_level": 1, + "bbox": [ + 89, + 818, + 341, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The high cost of manual annotation has driven researchers to seek alternative approaches. Bai et al. [3] introduce RL from AI feedback as one such solution. It follows naturally that PickScore, and HPSv2 can be employed as AI evalua", + "bbox": [ + 88, + 839, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/049d4e9ac21d043c961e5bd66e05ff8de445b28ed0b2d486efd000a2e3702d8b.jpg", + "image_caption": [ + "Figure 6. Median of PickScore and HPSv2 comparisons for all baselines and test sets on SD1.5. (Detailed analysis in Sec. 5.3)" + ], + "image_footnote": [], + "bbox": [ + 514, + 542, + 908, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tor to simulate human preferences. We relabel the Pick-a-Pic v2 dataset using AI evaluators and fine-tuned the SD1.5 model. As shown in Fig. 6, we observe a further improvement in these two metrics with AI feedback.", + "bbox": [ + 511, + 758, + 908, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablations", + "text_level": 1, + "bbox": [ + 511, + 832, + 627, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 presents the results of our ablation experiments on SD1.5 and the median values of the preference evaluators on the HPDv2 test set. For a clearer comparison of our", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28635", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/68c4aa89f9c464422e7690ece567c1cf5a738044927aea5410b9c02891f0a932.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Median (HPDv2)GPU hours
AestheticPickScoreHPSv2CLIP
Base-SD1.55.349120.71926.64734.276-
DPO-SD1.55.692221.27427.42835.902204.8
Base initialized5.650421.76328.23736.23857.6
β = 30005.773021.80828.50536.78857.6
β = 40005.763721.79628.44836.68557.6
β = 50005.742821.77128.47336.66957.6
Inv-steps n=305.822521.95728.56837.019136
Inv-steps n=55.771521.84528.43236.76943.2
Inv-steps n=35.748921.83328.50836.54236
CFG win=15.738121.83528.51636.45162.4
CFG win=55.732021.71228.46236.426104
CFG win=7.55.749521.74228.47336.469104
InPO-SD1.55.773421.89428.52336.87657.6
", + "bbox": [ + 91, + 88, + 498, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation studies of our DDIM-InPO on fine-tuning SD1.5. In each column of preference evaluators, the highest value is highlighted in bold, and the second highest is underlined. For detailed analysis, refer to Sec. 5.4", + "bbox": [ + 89, + 325, + 483, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "method's effectiveness, we present the results for the Base-SD1.5 and DPO-SD1.5 baselines in rows 3 and 4 of Tab. 2.", + "bbox": [ + 89, + 409, + 483, + 440 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proposed improvements. The fifth row of Tab. 2 presents the results of initializing $p_{\\theta}^{c}$ in Eq. (11) with the SD1.5 base model. After applying our core contribution, DDIM inversion, we observe a significant improvement in quality compared to the median values of the evaluators for the baselines, indicating that enhancement primarily stems from our method. As instructed by [48], practically an available SFT model can be used to initialize the reference model $p_{\\mathrm{ref}}^{c}$ . This strategy aids in reducing the issue of distribution shift.", + "bbox": [ + 89, + 458, + 483, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Parameters Choices. We set $\\beta$ to 2000 for SD1.5. Rows 6 to 8 of Tab. 2 indicate that if $\\beta$ is set too high, the KL divergence penalty term severely limits the model's flexibility in adjustments. It also maintains a consistent learning rate with other baselines, ensuring a fair comparison in terms of efficiency. Rows 9 to 11 of Tab. 2 show the steps required for DDIM inversion. We observe that setting inversion steps to 30 yields the best performance, further confirming the effectiveness of our method. However, this also increases the training time. Thus we select inversion steps $n = 10$ as it strikes a good balance between generation quality and training speed. The work of [38, 39] indicates that DDIM inversion can lead to considerable numerical errors when Classifier Free Guidance (CFG) [26] $w_{\\mathrm{inv}} > 0$ . Rows 12 to 14 of Tab. 2 demonstrate that a CFG $w_{\\mathrm{inv}} > 0$ during the inversion process adversely impacts the fine-tuning of human preferences. Consequently, we utilize unconditional DDIM inversion, setting CFG $w_{\\mathrm{inv}} = 0$ . We optimize parameters to balance generation quality and training efficiency.", + "bbox": [ + 89, + 613, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/841e6584132d563e24d019930ca2435c9ad0be7b699f7ceb5052fe6ec916c943.jpg", + "image_caption": [ + "Figure 7. Qualitative evaluation of InPO-SDXL in comparison with Base-SDXL and DPO-SDXL on conditional generation tasks (From top to bottom: depth map, canny edge, and inpainting). Prompts: (1) A dolphin in an astronaut suit on Saturn. (2) The Great Pyramid of Giza. (3) A cat riding on a horse." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 905, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Conditional generation", + "text_level": 1, + "bbox": [ + 511, + 436, + 725, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After fine-tuning the model using the DDIM-InPO method, it can be directly applied to conditional generation tasks without the need for further training. We add conditional control to InPO-SDXL using ControlNet [74] which originally trained with Base-SDXL. Specifically, we use depth map and canny edge as conditions for text-to-image generation tasks. Furthermore, we evaluate our model on image inpainting [37] tasks. As shown in Fig. 7, the advantages of our model are seamlessly transferred to conditional generation tasks, demonstrating superior generation quality compared with Base-SDXL and DPO-SDXL.", + "bbox": [ + 511, + 460, + 906, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 647, + 633, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce DDIM-InPO, an efficient method for aligning diffusion models with human preferences, focusing on directly fine-tuning specific latent variables that are strongly correlated with the target image. By fine-tuning the T2I diffusion models using the Pick-a-Pic v2 dataset, our experiments show that DDIM-InPO significantly enhances training efficiency, achieving state-of-the-art performance in just 400 training steps and surpassing all baseline models across nearly all evaluators. DDIM-InPO is based on offline fine-tuning and is sensitive to dataset quality; the presence of harmful, violent, and pornographic content can affect the results, leading to the generation of inappropriate outputs, such as overly feminine images. In addition, we consider the online training paradigm to be a promising and reliable approach for the future.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28636", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Major Science and Technology Projects (the grant number 2022ZD0117000) and the National Natural Science Foundation of China (grant number 62202426).", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 188, + 187, + 204 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[2] Mohammad Gheshlaghi Azar, Mark Rowland, Bilal Piot, Daniel Guo, Daniele Calandriello, Michal Valko, and Rémi Munos. A general theoretical paradigm to understand learning from human preferences. ArXiv, abs/2310.12036, 2023. 2, 1", + "[3] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 7", + "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. *ArXiv*, abs/2211.01324, 2022. 1", + "[5] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 1", + "[6] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. ArXiv, abs/2305.13301, 2023. 1, 2", + "[7] A. Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, and Dominik Lorenz. Stable video diffusion: Scaling latent video diffusion models to large datasets. ArXiv, abs/2311.15127, 2023. 1", + "[8] A. Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22563-22575, 2023. 1", + "[9] Haoxin Chen, Menghan Xia, Yin-Yin He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao-Liang Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation. ArXiv, abs/2310.19512, 2023. 2", + "[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ArXiv, abs/2310.00426, 2023. 1" + ], + "bbox": [ + 93, + 213, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Paul Francis Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. ArXiv, abs/1706.03741, 2017. 1", + "[12] Kevin Clark, Paul Vicol, Kevin Swersky, and David J. Fleet. Directly fine-tuning diffusion models on differentiable rewards. ArXiv, abs/2309.17400, 2023. 1, 2", + "[13] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam S. Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Kumar Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yiqian Wen, Yi-Zhe Song, Roshan Sumbaly, Vignesh Ramanathan, Zijian He, Péter Vajda, and Devi Parikh. Emu: Enhancing image generation models using photogenic needles in a haystack. ArXiv, abs/2309.15807, 2023. 1", + "[14] Quan Dao, Hao Phung, Binh Duc Nguyen, and A. Tran. Flow matching in latent space. ArXiv, abs/2307.08698, 2023. 2", + "[15] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. ArXiv, abs/2105.05233, 2021. 1", + "[16] Hanze Dong, Wei Xiong, Deepanshu Goyal, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and T. Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. ArXiv, abs/2304.06767, 2023. 1", + "[17] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7312-7322, 2023. 1", + "[18] Patrick Esser, Sumith Kulal, A. Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. ArXiv, abs/2403.03206, 2024. 1", + "[19] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. ArXiv, abs/2402.01306, 2024. 2, 1", + "[20] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, P. Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. ArXiv, abs/2305.16381, 2023. 2", + "[21] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising. ArXiv, abs/2403.14602, 2024. 1", + "[22] Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. ArXiv, abs/2310.11513, 2023. 1", + "[23] Han Guo, Bowen Tan, Zhengzhong Liu, Eric P. Xing, and Zhiting Hu. Efficient (soft) q-learning for text generation with limited good data. In Conference on Empirical Methods in Natural Language Processing, 2021. 1", + "[24] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28637", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Photorealistic video generation with diffusion models. In European Conference on Computer Vision, 2023. 1", + "[25] Amir Hertz, Ron Mokady, Jay M. Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. ArXiv, abs/2208.01626, 2022. 1", + "[26] Jonathan Ho. Classifier-free diffusion guidance. ArXiv, abs/2207.12598, 2022. 8, 1", + "[27] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. ArXiv, abs/2210.02303, 2022. 1", + "[28] Jiwoo Hong, Noah Lee, and James Thorne. Orpo: Monolithic preference optimization without reference model. ArXiv, abs/2403.07691, 2024. 2, 1", + "[29] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. ArXiv, abs/2402.17525, 2024. 2, 1", + "[30] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. ArXiv, abs/2206.00364, 2022. 1", + "[31] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 36652-36663, 2023. 2, 5, 6", + "[32] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. arXiv preprint arXiv:2404.04465, 2024. 1, 3, 6", + "[33] Zhanhao Liang, Yuhui Yuan, Shuyang Gu, Bohan Chen, Tiankai Hang, Ji Li, and Liang Zheng. Step-aware preference optimization: Aligning preference with denoising performance at each step. ArXiv, abs/2406.04314, 2024. 2", + "[34] Tianqi Liu, Yao Zhao, Rishabh Joshi, Misha Khalman, Mohammad Saleh, Peter J. Liu, and Jialu Liu. Statistical rejection sampling improves preference optimization. ArXiv, abs/2309.06657, 2023. 1", + "[35] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. ArXiv, abs/1711.05101, 2017. 5", + "[36] Artem Lukoianov, Haitz Sáez de Ocariz Borde, Kristjan Greenewald, Vitor Campagnolo Guizilini, Timur Bagautdinov, Vincent Sitzmann, and Justin Solomon. Score distillation via reparametrized ddim. arXiv preprint arXiv:2405.15891, 2024. 4, 3", + "[37] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 8, 1", + "[38] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 8, 1" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3, 4, 8", + "[40] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. ArXiv, abs/2203.02155, 2022. 1", + "[41] William S. Peebles and Saining Xie. Scalable diffusion models with transformers. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4172-4182, 2022. 1", + "[42] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. 2023. 1", + "[43] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 5", + "[44] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. ArXiv, abs/2209.14988, 2022. 2", + "[45] Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. ArXiv, abs/2310.03739, 2023. 1, 2", + "[46] Mihir Prabhudesai, Russell Mendonca, Zheyang Qin, Katerina Fragkiadaki, and Deepak Pathak. Video diffusion alignment via reward gradients. ArXiv, abs/2407.08737, 2024. 1", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 6", + "[48] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 1, 8", + "[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 2", + "[50] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022.", + "[51] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International Conference on Machine Learning, 2016. 2, 1" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28638", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5", + "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1", + "[54] Axel Sauer, Dominik Lorenz, A. Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, 2023. 1", + "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv, abs/2210.08402, 2022.6", + "[56] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. ArXiv, abs/1707.06347, 2017. 2, 1", + "[57] Noam M. Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. ArXiv, abs/1804.04235, 2018. 5", + "[58] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8871-8879, 2023. 1", + "[59] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 1", + "[60] Feifan Song, Yu Bowen, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. Preference ranking optimization for human alignment. ArXiv, abs/2306.17492, 2023. 2, 1", + "[61] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3", + "[62] Yang Song, Jascha Narain Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. ArXiv, abs/2011.13456, 2020. 1", + "[63] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 1, 2, 3, 4, 5, 6", + "[64] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. ArXiv, abs/2311.03079, 2023. 1", + "[65] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. ArXiv, abs/2306.09341, 2023. 2", + "[66] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2096-2105, 2023. 6", + "[67] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. ArXiv, abs/2304.05977, 2023. 2", + "[68] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Qimai Li, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8941-8951, 2023. 2", + "[69] Shentao Yang, Tianqi Chen, and Mingyuan Zhou. A dense reward view on aligning text-to-image diffusion with preference. ArXiv, abs/2402.08265, 2024. 1, 3", + "[70] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Trans. Mach. Learn. Res., 2022, 2022. 2", + "[71] Hangjie Yuan, Shiwei Zhang, Xiang Wang, Yujie Wei, Tao Feng, Yining Pan, Yingya Zhang, Ziwei Liu, Samuel Albanie, and Dong Ni. Instructvideo: Instructing video diffusion models with human feedback. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6463-6474, 2023. 1", + "[72] Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Feiran Huang. Rrhf: Rank responses to align language models with human feedback without tears. ArXiv, abs/2304.05302, 2023. 1", + "[73] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1204-1213, 2021. 1", + "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3813-3824, 2023. 8, 1", + "[75] Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. arXiv preprint arXiv:2401.12244, 4, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 840 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28639", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/14854e80-d5af-4094-8f56-d57259d0057b_model.json b/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/14854e80-d5af-4094-8f56-d57259d0057b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..eb847424edb49e4f86deeef7ccb8621e89eae6fe --- /dev/null +++ b/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/14854e80-d5af-4094-8f56-d57259d0057b_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.13, + 0.872, + 0.177 + ], + "angle": 0, + "content": "InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.203, + 0.916, + 0.258 + ], + "angle": 0, + "content": "Yunhong Lu\\(^{1}\\) Qichao Wang\\(^{1}\\) Hengyuan Cao\\(^{1}\\) Xierui Wang\\(^{1}\\) Xiaoyin Xu\\(^{1}\\) Min Zhang\\(^{1,2*}\\) \n\\(^{1}\\) Zhejiang University \\(^{2}\\) Shanghai Institute for Advanced Study-Zhejiang University \n{yunhonglu,qichaowang,caohy,sherrywang,xiaoyinxu,min_zhang}@zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.278, + 0.292, + 0.356, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.324, + 0.486, + 0.718 + ], + "angle": 0, + "content": "Without using explicit reward, direct preference optimization (DPO) employs paired human preference data to fine-tune generative models, a method that has garnered considerable attention in large language models (LLMs). However, exploration of aligning text-to-image (T2I) diffusion models with human preferences remains limited. In comparison to supervised fine-tuning, existing methods that align diffusion model suffer from low training efficiency and subpar generation quality due to the long Markov chain process and the intractability of the reverse process. To address these limitations, we introduce DDIM-InPO, an efficient method for direct preference alignment of diffusion models. Our approach conceptualizes diffusion model as a single-step generative model, allowing us to fine-tune the outputs of specific latent variables selectively. In order to accomplish this objective, we first assign implicit rewards to any latent variable directly via a reparameterization technique. Then we construct an Inversion technique to estimate appropriate latent variables for preference optimization. This modification process enables the diffusion model to only fine-tune the outputs of latent variables that have a strong correlation with the preference dataset. Experimental results indicate that our DDIM-InPO achieves state-of-the-art performance with just 400 steps of fine-tuning, surpassing all preference aligning baselines for T2I diffusion models in human preference evaluation tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.222, + 0.761 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.879 + ], + "angle": 0, + "content": "In the domain of T2I generation, tailoring image outputs to reflect user preferences is still in an early stage of development. In contrast, large language models (LLMs) [1] have made notable progress in generating human-aligned text, following a two-stage process: first, pre-training on extensively collected large-scale datasets and then, aligning on smaller datasets that capture specific human needs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.384 + ], + "angle": 0, + "content": "This alignment aims to produce outputs that meet human preferences while retaining model's learned capabilities. In recent years, T2I diffusion models [18, 43, 52] have been at the forefront of image generation. Nevertheless, achieving alignment between model outputs and human preferences remains a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.909, + 0.657 + ], + "angle": 0, + "content": "Recently, researchers have developed several methods tailoring diffusion models to human preferences, yet these approaches generally lack effectiveness. Some methods adopt a two-stage training approach, where a reward function is first trained and then followed by updating the diffusion model using gradient-based back-propagation [12, 45]. These approaches tend to be unstable and are susceptible to reward leakage [75]. Alternatively, RL (reinforcement learning)-based methods model the denoising process as a Markov decision process, with online evaluation of generated images based on reward feedback [6]. However, their effectiveness is constrained by the length of the Markov chain. Other approaches such as [32, 63], inspired by direct preference optimization (DPO) [48], reduce the RL objective to an optimal policy training goal. Nevertheless, due to the influence of sparse rewards [23, 69], these methods' performance is limited, especially when handling out-of-distribution inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.91, + 0.903 + ], + "angle": 0, + "content": "To address these issues, we introduce DDIM-InPO, an efficient approach for direct preference optimization of diffusion models. In response to the challenge of the implicit rewards allocation in the long-chain denoising process, our key idea is to treat the diffusion model as a single-step generation framework, only fine-tuning the outputs of latent variables directly linked to the target image. To achieve this, we first link the latent variables to the target image space in a single step, using a reparameterization method. It allows us to allocate implicit rewards to latent variables at any timestep directly. Secondly, to mitigate the effects of sparse rewards, we use an Inversion technique in the target image space to identify images that are highly correlated with the target and obtain the corresponding latent variables. Finally, we focus on optimizing the outputs of this limited set of latent variables and thus achieve rapid computations, enabling" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.229, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28629" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.905, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.477, + 0.908, + 0.521 + ], + "angle": 0, + "content": "Figure 1. We develop DDIM-InPO, an efficient method to align diffusion models with human preference. It suffices to directly optimize the outputs of a small set of variables using human feedback data. This figure illustrates the results after 400 fine-tuning steps of SDXL-base-1.0 using our method, showing that the generated images exhibit strong visual appeal and align well with human preferences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.483, + 0.574 + ], + "angle": 0, + "content": "high-quality optimization of the diffusion model. Our main contributions are three-folds:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.576, + 0.483, + 0.62 + ], + "angle": 0, + "content": "- We propose a novel DPO implicit rewards allocation method for the diffusion model, treating the diffusion model as a single-step generative model." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.622, + 0.483, + 0.711 + ], + "angle": 0, + "content": "- We introduce DDIM-InPO, an efficient framework for direct preference optimization of diffusion models. By applying an Inversion technique in the target image space, we estimate images highly correlated with the preference dataset and derive the corresponding latent variables sufficient for fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.712, + 0.483, + 0.787 + ], + "angle": 0, + "content": "- In our experiments, we benchmark our method against state-of-the-art fine-tuning approaches for human preference, demonstrating enhanced training efficiency (refer to Fig. 4) and improved generation quality (examples shown in Fig. 1 and Fig. 3)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.576, + 0.483, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.241, + 0.816 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Text-to-Image Generative Models. T2I generation models aim to produce high-quality, high-fidelity images from a given text prompt. Additionally, these models have contributed to advancements in related fields such as image editing [29], video generation [9], and 3D modeling [44]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.907, + 0.651 + ], + "angle": 0, + "content": "Among T2I generative models [14, 49-51, 70], diffusion models have recently gained significant attention due to their open-source nature and superior generation performance. However, high-quality diffusion models are typically trained on large, noisy datasets, which often leads to deviations from human preferences, an area that still requires further refinement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Diffusion Models Alignment. In diffusion models, supervised fine-tuning (SFT) is commonly employed to align them with human preferences. Inspired by RL-based fine-tuning methods used in LLMs [2, 19, 28, 56, 60], various reward models such as PickScore [31] and HPSv2 [65] have been developed to simulate human preferences. With reward models, approaches like DRaFT [12] and AlignProp [45] use differentiable rewards and back-propagate through the diffusion sampling process, which demands substantial memory and would result in reward leakage. DPOK [20] and DDPO [6] model the sampling process of diffusion models as a Markov decision process and apply policy gradient methods for fine-tuning. Diffusion-DPO [63] and D3PO [68] adopt direct preference optimization by leveraging the denoising steps of diffusion models to optimize" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28630" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "model parameters at each step. In the DPO framework, DenseReward [69] allocates higher weights to the initial steps. Going further, Diffusion-KTO [32] requires only binary feedback signals for each image to fine-tune the model. However, all these methods are limited to expanding the decision space by incorporating the reward function over all denoising steps, which results in sparse rewards and inefficient training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.225, + 0.231, + 0.24 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.249, + 0.26, + 0.264 + ], + "angle": 0, + "content": "3.1. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.485, + 0.409 + ], + "angle": 0, + "content": "Diffusion models involve a forward process \\(\\{q(\\pmb{x}_t)\\}_{t\\in [0,T]}\\) that incrementally adds noise to data \\(\\pmb{x}_0\\sim q(\\pmb{x}_0)\\) and a learned reverse process \\(\\{p(\\pmb{x}_t)\\}_{t\\in [0,T]}\\) that aims to denoise the data. The forward process is defined as \\(q(\\pmb{x}_t|\\pmb{x}_0)\\coloneqq \\mathcal{N}(\\sqrt{\\alpha_t}\\pmb{x}_0,(1 - \\alpha_t)\\mathbf{I})\\) and \\(q(\\pmb{x}_t)\\coloneqq \\int q(\\pmb{x}_t|\\pmb{x}_0)q(\\pmb{x}_0)\\mathrm{d}\\pmb{x}_0,\\) where \\(\\alpha_{t}\\) is a noise schedule (as defined in [61]). Starting from \\(p(\\pmb{x}_T)\\coloneqq \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) , the reverse process is defined by a parameterized denoiser \\(\\epsilon_{\\theta}^{t}(\\pmb{x}_{t})\\) , which aims to predict noise added to \\(\\pmb{x}_0\\) . The denoiser is optimized by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.416, + 0.482, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D M}} := \\mathbb {E} _ {x _ {0}, t, \\epsilon} [ w (t) \\| \\epsilon_ {\\theta} ^ {t} (\\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon) - \\epsilon \\| _ {2} ^ {2} ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.483, + 0.48 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_0\\sim q(\\pmb{x}_0),t\\sim \\mathcal{U}(0,T),\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) and \\(w(t)\\) is a pre-specified weight function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.484, + 0.542 + ], + "angle": 0, + "content": "Denoising Diffusion Implicit Models (DDIM) generate samples by deterministically reversing the diffusion process. By reparameterizing noisy image \\(\\bar{\\pmb{x}}_t = \\pmb{x}_t / \\sqrt{\\alpha_t}\\), the denoising process can be viewed as an ODE:" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.55, + 0.483, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\frac {\\mathrm {d} \\bar {\\boldsymbol {x}} _ {t}}{\\mathrm {d} t} = \\epsilon_ {\\theta} ^ {t} \\left(\\frac {\\bar {\\boldsymbol {x}} _ {t}}{\\sqrt {\\sigma_ {t} ^ {2} + 1}}\\right) \\frac {\\mathrm {d} \\sigma_ {t}}{\\mathrm {d} t} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.484, + 0.662 + ], + "angle": 0, + "content": "where \\(\\sigma_{t} = \\sqrt{1 - \\alpha_{t}} / \\sqrt{\\alpha_{t}}\\) and \\(\\bar{x}_T\\) is sampled from Gaussian distribution. The DDIM ODE can also be integrated in reverse to estimate \\(\\pmb{x}_t\\) at any timestep \\(t\\) starting from a clean image \\(\\pmb{x}_0\\), a process known as DDIM inversion [39]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.669, + 0.379, + 0.685 + ], + "angle": 0, + "content": "3.2. Human Preference Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.484, + 0.783 + ], + "angle": 0, + "content": "Preference optimization involves fine-tuning generative models to produce samples more closely align with human preferences. Assuming that human preferences are represented by an implicit reward function \\( r^* \\), there is a dataset \\( \\mathcal{D} = \\{(x_0^w, x_0^l, c)\\} \\), where \\( c \\) denotes the condition, and \\( x_0^w \\) and \\( x_0^l \\) represent the winner and loser samples, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.8, + 0.484, + 0.861 + ], + "angle": 0, + "content": "RLHF. RLHF employs maximum likelihood estimation to fit the Bradley-Terry (BT) model on the dataset \\(\\mathcal{D}\\), thereby obtaining an explicit parameterized reward model \\(r_{\\phi}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.869, + 0.484, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {B T}} (\\phi) := - \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right) \\sim \\mathcal {D}} [ \\log \\sigma \\left(r _ {\\phi} \\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {c}\\right) - r _ {\\phi} \\left(\\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right)\\right) ] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "With the reward, the generative distribution \\( p_{\\theta} \\) can then be optimized as a policy using RL that maximizes reward feedback, with an regularization term for the KL-divergence from the reference distribution \\( p_{\\mathrm{ref}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.161, + 0.907, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\max _ {p _ {\\theta}} \\mathbb {E} _ {\\substack {\\boldsymbol {c} \\sim \\mathcal {D} _ {\\boldsymbol {c}} \\\\ \\boldsymbol {x} _ {0} \\sim p _ {\\theta} (\\boldsymbol {x} _ {0} | \\boldsymbol {c})}} [ r (\\boldsymbol {x} _ {0}, \\boldsymbol {c}) ] - \\beta \\mathbb {D} _ {\\mathrm {K L}} [ p _ {\\theta} (\\boldsymbol {x} _ {0} | \\boldsymbol {c}) | | p _ {\\text {ref}} (\\boldsymbol {x} _ {0} | \\boldsymbol {c}) ] \\tag{4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.2, + 0.888, + 0.216 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a hyperparameter that controls regularization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.233, + 0.906, + 0.279 + ], + "angle": 0, + "content": "Direct Preference Optimization. Direct Preference Optimization (DPO) considers the unique global optimal policy \\( p_{\\theta}^{*} \\) under the given reward function \\( r \\) in Eq. (4):" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.287, + 0.906, + 0.323 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} ^ {*} \\left(\\boldsymbol {x} _ {0} \\mid \\boldsymbol {c}\\right) = \\frac {p _ {\\text {r e f}} \\left(\\boldsymbol {x} _ {0} \\mid \\boldsymbol {c}\\right) \\exp \\left(r \\left(\\boldsymbol {x} _ {0} , \\boldsymbol {c}\\right) / \\beta\\right)}{\\sum_ {\\boldsymbol {x} _ {0}} p _ {\\text {r e f}} \\left(\\boldsymbol {x} _ {0} \\mid \\boldsymbol {c}\\right) \\exp \\left(r \\left(\\boldsymbol {x} _ {0} , \\boldsymbol {c}\\right) / \\beta\\right)} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.331, + 0.906, + 0.369 + ], + "angle": 0, + "content": "Then, the reward function \\( r(\\pmb{x}_0, \\pmb{c}) \\) is rewritten according to Eq. (5) and inserted into the BT model from Eq. (3), resulting in the DPO objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.373, + 0.905, + 0.419 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D P O}} := - \\underset {\\substack {(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l} \\\\ , \\boldsymbol {c}) \\sim \\mathcal {D}}} {\\mathbb {E}} \\log \\sigma \\left(\\beta \\log \\frac {p _ {\\theta} (\\boldsymbol {x} _ {0} ^ {w} | \\boldsymbol {c})}{p _ {\\text {ref}} (\\boldsymbol {x} _ {0} ^ {w} | \\boldsymbol {c})} - \\beta \\log \\frac {p _ {\\theta} (\\boldsymbol {x} _ {0} ^ {l} | \\boldsymbol {c})}{p _ {\\text {ref}} (\\boldsymbol {x} _ {0} ^ {l} | \\boldsymbol {c})}\\right) \\tag{6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.42, + 0.906, + 0.495 + ], + "angle": 0, + "content": "Based on Eqs. (4) and (6), Wallace et al. [63] redistribute rewards \\( r(\\pmb{x}_0, c) \\) across all possible diffusion paths \\( p_{\\theta}(\\pmb{x}_{1:T} | \\pmb{x}_0, c) \\) and minimize the KL-divergence of the joint probability density \\( \\mathbb{D}_{\\mathrm{KL}}(p_{\\theta}(\\pmb{x}_{0:T} | c) || p_{\\mathrm{ref}}(\\pmb{x}_{0:T} | c)) \\), thus designing the DPO objective for diffusion models:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.505, + 0.785, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D P O - D i f f u s i o n}} := - \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right) \\sim \\mathcal {D}} \\log \\sigma\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.526, + 0.906, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\left. \\left(\\beta \\mathbb {E} _ {\\substack {\\boldsymbol {x} _ {1: T} ^ {w} \\sim p _ {\\theta} ^ {c} (\\boldsymbol {x} _ {1: T} ^ {w} \\mid \\boldsymbol {x} _ {0} ^ {w}) \\\\ \\boldsymbol {x} _ {1: T} ^ {l} \\sim p _ {\\theta} ^ {c} (\\boldsymbol {x} _ {1: T} ^ {l} \\mid \\boldsymbol {x} _ {0} ^ {l})}} \\left[ \\log \\frac {p _ {\\theta} ^ {c} (\\boldsymbol {x} _ {0 : T} ^ {w})}{p _ {\\text {ref}} ^ {c} (\\boldsymbol {x} _ {0 : T} ^ {w})} - \\log \\frac {p _ {\\theta} ^ {c} (\\boldsymbol {x} _ {0 : T} ^ {l})}{p _ {\\text {ref}} ^ {c} (\\boldsymbol {x} _ {0 : T} ^ {l})} \\right]\\right) \\right. \\tag{7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.591, + 0.906, + 0.637 + ], + "angle": 0, + "content": "where \\( p_{\\theta}^{c}(\\cdot) \\) represents \\( p_{\\theta}(\\cdot | c) \\) for compactness. Without the need for RL algorithms, the DPO-based method directly optimizes the optimal distribution \\( p_{\\theta} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.828, + 0.666 + ], + "angle": 0, + "content": "4. Inversion Preference Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Given a fixed dataset \\(\\mathcal{D} = \\{(\\pmb{x}_0^w, \\pmb{x}_0^l, \\pmb{c})\\}\\), where each sample includes a prompt \\(c\\) and an image pair generated by reference model \\(p_{\\mathrm{ref}}\\) labeled as \\(\\pmb{x}_0^w \\succ \\pmb{x}_0^l\\), our goal is to train a model \\(p_\\theta\\) that aligns with human preferences. The primary challenge lies in how to match the parameterized distribution \\(p_\\theta^c(\\pmb{x}_0)\\) according to Eq. (6). Our insight is that the diffusion model can be conceptualized as a timestep-aware single-step generative model \\(p_\\theta^c(\\pmb{x}_0|\\pmb{x}_t)\\). Specifically, there exists some latent variables \\(\\pmb{x}_t\\) that can generate \\(\\pmb{x}_0\\) in a single step for any timestep \\(t\\), which can be written mathematically as: \\(p_\\theta^c(\\pmb{x}_0) = \\int p_\\theta^c(\\pmb{x}_0|\\pmb{x}_t)p_\\theta^c(\\pmb{x}_t)\\mathrm{d}\\pmb{x}_t\\). Thus, the first step is to determine the relationship between \\(\\pmb{x}_t\\) and \\(\\pmb{x}_0\\) in Sec. 4.1, and the second step is to identify appropriate latent variables \\(\\pmb{x}_t\\) for any timestep in Sec. 4.2, within the DPO framework." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28631" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.376, + 0.106 + ], + "angle": 0, + "content": "4.1. Reparametrized DDIM for DPO" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.173 + ], + "angle": 0, + "content": "We establish a relationship between \\( \\boldsymbol{x}_t \\) and \\( \\boldsymbol{x}_0 \\) to be used within the DPO framework. Instead of viewing DDIM as a denoising process within the noisy image \\( \\boldsymbol{x}_t \\) space, we reparameterize it with 'initial' variable:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.184, + 0.483, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {0} (t) := \\bar {\\boldsymbol {x}} _ {t} - \\sigma_ {t} \\epsilon_ {\\theta} ^ {t} (\\boldsymbol {x} _ {t}, \\boldsymbol {c}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.485, + 0.288 + ], + "angle": 0, + "content": "where noisy image \\(\\bar{\\pmb{x}}_t = \\pmb {x}_t / \\sqrt{\\alpha_t}\\) . In this context, \\(\\pmb {x}_0(t)\\) is the image obtained after the single-step denoising at timestep \\(t\\) and is in \\(\\pmb {x}_0\\) space. Therefore, \\(p_{\\theta}^{c}(\\pmb {x}_0,\\pmb {x}_t)\\) is well-defined, allowing us to present a formal definition of 'initial' reward \\(r(\\pmb {x}_0,\\pmb {c})\\) with the joint' reward \\(r_t^c (\\pmb {x}_0,\\pmb {x}_t)\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.302, + 0.483, + 0.32 + ], + "angle": 0, + "content": "\\[\nr \\left(\\boldsymbol {x} _ {0}, \\boldsymbol {c}\\right) = \\mathbb {E} _ {p _ {\\theta} ^ {c} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right)} \\left[ r _ {t} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}\\right) \\right] \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.484, + 0.4 + ], + "angle": 0, + "content": "Similar to [63], we minimize the joint KL divergence \\(\\mathbb{D}_{\\mathrm{KL}}[p_{\\theta}^{c}(\\boldsymbol{x}_0,\\boldsymbol{x}_t)||p_{\\mathrm{ref}}^{c}(\\boldsymbol{x}_0,\\boldsymbol{x}_t)]\\) to serve as an upper bound for the KL-divergence term in Eq. (4). With Eq. (9) and KL-divergence bound, we can reformulate Eq. (4) into the objective below:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.411, + 0.502, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\max _ {p _ {\\theta} ^ {\\mathcal {C}}} \\mathbb {E} _ {\\substack {t, (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) \\sim \\\\ p _ {\\theta} ^ {\\mathcal {C}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t})}} [ r _ {t} ^ {\\mathcal {C}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) ] - \\beta \\mathbb {D} _ {\\mathrm {K L}} [ p _ {\\theta} ^ {\\mathcal {C}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) | | p _ {\\text {ref}} ^ {\\mathcal {C}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) ] \\tag{10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.484, + 0.528 + ], + "angle": 0, + "content": "In this context, we consider all timesteps. This objective enables the effective allocation of \\( x_0 \\)'s reward across any timestep within the model, enhancing training efficiency. In a manner analogous to Eq. (3) through Eq. (6), we derive the following optimization objective:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.538, + 0.303, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) := - \\mathbb {E} _ {t, \\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right) \\sim \\mathcal {D}} \\log \\sigma\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.559, + 0.484, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\left(\\beta \\mathbb {E} _ {\\substack {\\boldsymbol {x} _ {t} ^ {w} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {t} ^ {w} | \\boldsymbol {x} _ {0} ^ {w}) \\\\ \\boldsymbol {x} _ {t} ^ {l} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {t} ^ {l} | \\boldsymbol {x} _ {0} ^ {l})}} \\left[ \\log \\frac {p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {0} ^ {w} , \\boldsymbol {x} _ {t} ^ {w})}{p _ {\\text {ref}} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {0} ^ {w} , \\boldsymbol {x} _ {t} ^ {w})} - \\log \\frac {p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {0} ^ {l} , \\boldsymbol {x} _ {t} ^ {l})}{p _ {\\text {ref}} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {0} ^ {l} , \\boldsymbol {x} _ {t} ^ {l})} \\right] \\right. \\tag{11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.627, + 0.483, + 0.672 + ], + "angle": 0, + "content": "To optimize Eq. (11), we need to sample \\( \\pmb{x}_t \\sim p_\\theta^c(\\pmb{x}_t | \\pmb{x}_0) \\), where \\( \\pmb{x}_t \\) should generate an approximation of \\( \\pmb{x}_0 \\) in a single step." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.42, + 0.697 + ], + "angle": 0, + "content": "4.2. Preference Optimization via Inversion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.764 + ], + "angle": 0, + "content": "So how can we determine whether \\( \\pmb{x}_t \\) can generate an approximation of \\( \\pmb{x}_0 \\) in a single step? Given \\( \\pmb{x}_0(t) \\), it remains unclear how to calculate \\( \\pmb{x}_t \\). From Eq. (8), \\( \\pmb{x}_t \\) should satisfy the following nonlinear equation:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.774, + 0.483, + 0.792 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t} = \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} (t) + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} ^ {t} (\\boldsymbol {x} _ {t}, c) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.803, + 0.42, + 0.817 + ], + "angle": 0, + "content": "and we can rewrite Eq. (12) with respect to noise:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.828, + 0.483, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\epsilon = \\epsilon_ {\\theta} ^ {t} \\left(\\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} (t) + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\boldsymbol {c}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\epsilon = (\\pmb{x}_t - \\sqrt{\\alpha_t}\\pmb{x}_0(t)) / \\sqrt{1 - \\alpha_t}\\). Thus, we can define \\(\\delta_t(\\pmb{x}_0(t)) = \\epsilon\\) as the solution when given \\(\\pmb{x}_0(t)\\). In this context, \\(\\pmb{x}_0(t)\\) serves as a reliable approximation of \\(\\pmb{x}_0\\) at" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.089, + 0.902, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.344, + 0.895, + 0.358 + ], + "angle": 0, + "content": "Figure 2. Illustration of Inversion for Preference Optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.905, + 0.433 + ], + "angle": 0, + "content": "timestep \\( t \\), that is, \\( x_0(t) \\approx x_0 \\). Additionally, we rewrite Eq. (2) by applying the definition of \\( x_0(t) \\) (Eq. (8)) by employing some algebra:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.445, + 0.905, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\frac {\\mathrm {d} \\boldsymbol {x} _ {0} (t)}{\\mathrm {d} t} = - \\sigma_ {t} \\frac {\\mathrm {d}}{\\mathrm {d} t} \\epsilon_ {\\theta} ^ {t} \\left(\\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} (t) + \\sqrt {1 - \\alpha_ {t}} \\delta_ {t} (\\boldsymbol {x} _ {0} (t)), \\boldsymbol {c}\\right) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.488, + 0.907, + 0.593 + ], + "angle": 0, + "content": "Inspired by the work of [36, 39], we suggest to derive \\(\\delta_t(\\pmb{x}_0(t))\\) by DDIM inversion (It should be noted that \\(\\delta_t(\\pmb{x}_0(t))\\) can be estimated using various methods, including but not limited to, inversion.) and obtain \\(\\pmb{x}_0(t)\\) by inverting Eq. (14) consequently. To clarify the process, we consider the consecutive time points \\(t\\) and \\(t + \\kappa > t\\), and the discretized form of inversion is given as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.606, + 0.85, + 0.646 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\pmb {x} _ {0} (t + \\kappa) = \\pmb {x} _ {0} (t) - \\sigma_ {t + \\kappa} [ \\epsilon_ {\\theta} ^ {t + \\kappa} (\\sqrt {\\alpha_ {t + \\kappa}} \\pmb {x} _ {0} (t) \\\\ \\left. + \\sqrt {1 - \\alpha_ {t + \\kappa}} \\delta_ {t} (\\boldsymbol {x} _ {0} (t))) - \\delta_ {t} (\\boldsymbol {x} _ {0} (t)) \\right] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.905, + 0.689 + ], + "angle": 0, + "content": "According to Eq. (12), \\( \\pmb{x}_t \\) is well-defined as follows, and so is \\( p_{\\theta}^{c}(\\pmb{x}_{t}|\\pmb{x}_{0}) \\), ensuring that Eq. (11) can be optimized." + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.703, + 0.905, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t} = \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} (t) + \\sqrt {1 - \\alpha_ {t}} \\delta_ {t} (\\boldsymbol {x} _ {0} (t)) \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.905, + 0.78 + ], + "angle": 0, + "content": "For the sake of simplicity, we approximate \\( p_{\\mathrm{ref}}^{\\mathbf{c}}(\\pmb{x}_0,\\pmb{x}_t) \\) with \\( p_{\\mathrm{ref}}^{\\mathbf{c}}(\\pmb{x}_0|\\pmb{x}_t)p_\\theta^{\\mathbf{c}}(\\pmb{x}_t) \\). Using Jensen's inequality and some algebra, we approximate Eq. (11) as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.793, + 0.904, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\theta) = - \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right) \\sim \\mathcal {D}, t \\sim \\mathcal {U} (0, T),} \\quad \\log \\sigma (- \\beta ( \\\\ \\boldsymbol {x} _ {t} ^ {w} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {t} ^ {w} \\mid \\boldsymbol {x} _ {0} ^ {w}\\right), \\boldsymbol {x} _ {t} ^ {l} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {t} ^ {l} \\mid \\boldsymbol {x} _ {0} ^ {l}\\right) \\\\ + \\mathbb {D} _ {\\mathrm {K L}} \\left(q ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {w} \\mid \\boldsymbol {x} _ {t} ^ {w}, \\boldsymbol {x} _ {0} ^ {w} (t)\\right) \\| p _ {\\theta} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {w} \\mid \\boldsymbol {x} _ {t} ^ {w}\\right)\\right) \\\\ - \\mathbb {D} _ {\\mathrm {K L}} \\left(q ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {w} \\mid \\boldsymbol {x} _ {t} ^ {w}, \\boldsymbol {x} _ {0} ^ {w} (t)\\right) \\| p _ {\\text {r e f}} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {w} \\mid \\boldsymbol {x} _ {t} ^ {w}\\right)\\right) \\tag {17} \\\\ - \\mathbb {D} _ {\\mathrm {K L}} \\left(q ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {l} \\mid \\boldsymbol {x} _ {t} ^ {l}, \\boldsymbol {x} _ {0} ^ {l} (t)\\right) \\| p _ {\\theta} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {l} \\mid \\boldsymbol {x} _ {t} ^ {l}\\right)\\right) \\\\ + \\mathbb {D} _ {\\mathrm {K L}} (q ^ {\\boldsymbol {c}} (\\pmb {x} _ {0} ^ {l} | \\pmb {x} _ {t} ^ {l}, \\pmb {x} _ {0} ^ {l} (t)) \\| p _ {\\mathrm {r e f}} ^ {\\boldsymbol {c}} (\\pmb {x} _ {0} ^ {l} | \\pmb {x} _ {t} ^ {l}))) \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28632" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.088, + 0.905, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.507, + 0.908, + 0.537 + ], + "angle": 0, + "content": "Figure 3. Top, in human evaluations, InPO-SDXL shows a marked improvement over both DPO-SDXL and Base-SDXL. Bottom, qualitative comparisons among baselines. InPO-SDXL achieves superior prompt alignment and produces images of higher quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.483, + 0.594 + ], + "angle": 0, + "content": "where \\(q^c (\\cdot |\\cdot)\\) is a Gaussian transition kernel. Using Eq.(16), the above loss can be simplified as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.606, + 0.482, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\theta) = - \\mathbb {E} _ {(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}) \\sim \\mathcal {D}, t \\sim \\mathcal {U} (0, T),} \\quad \\log \\sigma (- \\beta w (t) ( \\\\ \\boldsymbol {x} _ {t} ^ {w} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {t} ^ {w} | \\boldsymbol {x} _ {0} ^ {w}), \\boldsymbol {x} _ {t} ^ {l} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {t} ^ {l} | \\boldsymbol {x} _ {0} ^ {l}) \\\\ \\left\\| \\tau_ {t} ^ {w} - \\epsilon_ {\\theta} ^ {t} \\left(\\boldsymbol {x} _ {t} ^ {w}, \\boldsymbol {c}\\right) \\right\\| _ {2} ^ {2} - \\left\\| \\tau_ {t} ^ {w} - \\epsilon_ {\\mathrm {r e f}} ^ {t} \\left(\\boldsymbol {x} _ {t} ^ {w}, \\boldsymbol {c}\\right) \\right\\| _ {2} ^ {2} \\\\ - \\left\\| \\tau_ {t} ^ {l} - \\epsilon_ {\\theta} ^ {t} \\left(\\boldsymbol {x} _ {t} ^ {l}, \\boldsymbol {c}\\right) \\right\\| _ {2} ^ {2} + \\left\\| \\tau_ {t} ^ {l} - \\epsilon_ {\\text {r e f}} ^ {t} \\left(\\boldsymbol {x} _ {t} ^ {l}, \\boldsymbol {c}\\right) \\right\\| _ {2} ^ {2}) \\tag {18} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.483, + 0.74 + ], + "angle": 0, + "content": "where \\(\\tau_{t}^{*} = (\\pmb{x}_{0}^{*}(t) - \\pmb{x}_{0}^{*}) / \\sigma_{t} + \\delta_{t}(\\pmb{x}_{0}^{*}(t))\\) (refer to Fig. 2) and \\(w(t)\\) is a weight function defined consistent with [63]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Link to Diffusion-DPO. Diffusion-DPO defines a reward function on the denoising trajectory, which expands the decision space because it requires combining the reward functions across all denoise steps. This problem is well known to be closely related to delayed feedback/sparse rewards in reinforcement learning, where effective feedback is only available after generating the entire trajectory. Within our framework, Diffusion-DPO utilizes a rough approximation of \\(\\delta_t\\) which uses i.i.d random noise \\(\\delta_t^{\\mathrm{DPO}}(\\pmb{x}_0(t))\\sim\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.593 + ], + "angle": 0, + "content": "\\(\\mathcal{N}(0,I)\\), matching diffusion incorporated in the forward process of noise addition." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.646, + 0.625 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.632, + 0.596, + 0.648 + ], + "angle": 0, + "content": "5.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.907, + 0.791 + ], + "angle": 0, + "content": "Datasets and Models. We conduct all experiments on the Pick-a-Pic v2 [31] dataset, which is collected from the Pick-a-Pic web application (refer to [31] for details). Consistent with the setting in [63], approximately \\(12\\%\\) of the data is excluded, where preferences cannot be determined. As a result, we obtain 851,293 data pairs and 58,960 unique prompts. Our method is applied to fine-tune the base models Stable Diffusion 1.5 (SD1.5) [52] and Stable Diffusion XL-base-1.0 (SDXL) [43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Basic Hyperparameters. To ensure a fair comparison of training efficiency, we use hyperparameters largely consistent with those in [63]. We employ AdamW [35] as the optimizer for SD1.5 and Adafactor [57] for SDXL. All experiments are conducted on eight NVIDIA H800 GPUs, with each GPU handling a batch size of one pair and gradient" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28633" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.248 + ], + "angle": 0, + "content": "
ModelBaselinesWin-rate (HPDv2)Win-rate (Parti-Prompts)
AestheticPickScoreHPSv2CLIPAestheticPickScoreHPSv2CLIP
InPO-SDXLvs. Base-SDXL56.3779.2585.3853.3761.7072.8978.3150.55
vs. SFT-SDXL66.6688.3189.9159.0069.7088.5488.9156.99
vs. DPO-SDXL55.8759.2865.8146.9157.1756.9260.1142.28
InPO-SD1.5vs. Base-SD1.580.1985.8490.2266.4474.6373.1681.8661.89
vs. SFT-SD1.556.6366.8458.9157.8752.2762.5058.0357.54
vs. DPO-SD1.570.5074.9484.0661.0968.0163.5475.3157.54
vs. KTO-SD1.559.8167.4160.9460.0359.0162.2558.7657.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.258, + 0.907, + 0.302 + ], + "angle": 0, + "content": "Table 1. Win-rate comparison between DDIM-InPO and baselines. We use prompts from HPDv2 and Parti-Prompts, evaluated across different evaluators. SDXL and SD1.5 are used as base models. For simplicity, we denote models using a 'Method-Base' format, such as InPO-SDXL to represent the SDXL model fine-tuned with the DDIM-InPO method. Win-rates exceeding \\(50\\%\\) are highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.486, + 0.405 + ], + "angle": 0, + "content": "accumulation over 128 steps, resulting in an effective batch size of 1024 pairs (half the batch size compared to [63]). The learning rate is set to \\( \\frac{2000}{\\beta} 2.048^{-8} \\), with a linear warm-up phase. For SD1.5 training, \\( \\beta \\) is set to 2000, while for SDXL training, it is set to 5000." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.427, + 0.485, + 0.76 + ], + "angle": 0, + "content": "Evaluation. We compare DDIM-InPO with existing baselines using user studies and automatic preference metrics. We use the following four evaluators to evaluate image quality: the LAION aesthetic classifier [55] to assess image aesthetic scores, CLIP [47] to measure text-image alignment, and PickScore [31] and HPSv2 [66] to predict human preferences for given image-text pairs. In test stage, we generated images using both the baseline and our model on the Parti-Prompts (1632 prompts) and HPDv2 (3200 prompts) test sets. The median and win-rate (how frequently does the evaluator prefer the generations from DDIM-InPO over those from baselines) of the aforementioned four evaluators are used as automatic preference metrics for comparison. Additionally, we conduct a user study (16 participants) to compare DDIM-InPO with the baselines. We design three questions as [63]: (1) Which image is your overall preferred choice? (2) Which image is more visually attractive? (3) Which image better matches the text description? We also compare the traininggpu hours (on H800) with baselines. In this work, we adopt the following baselines: base model (SD1.5 and SDXL), supervised fine-tuning (SFT), Diffusion-DPO, and Diffusion-KTO (for SD1.5)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.254, + 0.789 + ], + "angle": 0, + "content": "5.2. Primary Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Quantitative Results. Firstly, we use the quantitative evaluation method for model preferences outlined in [32, 63]. Tab. 1 presents the win-rates of the InPO-aligned diffusion models in comparison to the baselines. Overall, following fine-tuning with DDIM-InPO, both SD1.5 and SDXL achieve superior performance compared to the baselines across nearly all evaluators and test datasets, thereby vali" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.343, + 0.871, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.524, + 0.907, + 0.624 + ], + "angle": 0, + "content": "Figure 4. Comparison of the trade-off between the quality of generated images and training efficiency following human preference optimization of SD1.5 on the HPDv2 test set. Sizes of the circles represent the volume of training data used. Our DDIM-InPO achieves superior performance, with a training speed that is 18.4 and 3.6 times faster than Diffusion-KTO [32] and Diffusion-DPO [63], respectively, while producing images of higher quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.793 + ], + "angle": 0, + "content": "dating the effectiveness of our method. Specifically, HPSv2 evaluation shows that, InPO-SDXL obtains a win-rate of \\(65.81\\%\\) against DPO-SDXL, while InPO-SD1.5 achieves a win-rate of \\(60.94\\%\\) compared to KTO-SD1.5 on the HPDv2 test set. Moreover, it significantly improves the alignment with human preferences for the base model. According to HPSv2, it achieves win-rate of up to \\(85.38\\%\\) and \\(90.22\\%\\) when compared to the SD1.5 and SDXL base models, respectively, demonstrating the efficacy of DDIM-InPO." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "It should be noted that our DDIM-InPO significantly advances training efficiency, with a training speed that is 18.4 and 3.6 times faster than Diffusion-KTO and Diffusion-DPO, respectively, as shown in Fig. 4, while producing higher-quality images. Additionally, Fig. 3 (top) presents the results of our user study, where we find that human evaluators prefer the model aligned with DDIM-InPO. The" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28634" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.905, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.436, + 0.908, + 0.52 + ], + "angle": 0, + "content": "Figure 5. Advantages of DDIM-InPO: Compared with Refinement technique and Diffusion-DPO, we find that images generated by DDIM-InPO exhibit enhanced light and spatial structure, better realism and detail capture, greater imaginative design, stable color consistency, optimized multi-instance layout and text integration in visuals. These are some hidden advantages aligned with human preferences. Prompts from left to right: (1) A small bird sitting in a metal wheel. (2) Trees seen through a car window on a rainy day. (3) Description, An artistic rendering of a cosmic portal with a beach at dusk on the other side. (4) A person in full samurai armor at the beach. (5) A man standing in front of a bunch of doughnuts. (6) A towel with the word 'cat' printed on it, simple and clear text." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.546, + 0.487, + 0.609 + ], + "angle": 0, + "content": "evaluation results clearly show that our method provides a more efficient way for fine-tuning diffusion models with human preferences, demonstrating the effectiveness of our proposed paradigm for learning specific latent variables." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.627, + 0.487, + 0.809 + ], + "angle": 0, + "content": "Qualitative results. We display qualitative comparison results of InPO-SDXL against several baselines (Base-SDXL, DPO-SDXL) in Fig. 3 (bottom). The images generated by InPO-SDXL demonstrate enhanced performance in both textual alignment and visual attractiveness. In addition, Fig. 5 illustrates the advantages of our model, which serves as a method for directly learning human preferences for diffusion models. The generated images showcase improved light and spatial structures, detailed capture, greater creativity, consistent color, organized multi-instance layouts, and text-visual integration. These advantages align closely with the preferences of users." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.819, + 0.342, + 0.836 + ], + "angle": 0, + "content": "5.3. AI Preference Optimization" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.486, + 0.903 + ], + "angle": 0, + "content": "The high cost of manual annotation has driven researchers to seek alternative approaches. Bai et al. [3] introduce RL from AI feedback as one such solution. It follows naturally that PickScore, and HPSv2 can be employed as AI evalua" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.543, + 0.909, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.701, + 0.909, + 0.731 + ], + "angle": 0, + "content": "Figure 6. Median of PickScore and HPSv2 comparisons for all baselines and test sets on SD1.5. (Detailed analysis in Sec. 5.3)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.759, + 0.909, + 0.819 + ], + "angle": 0, + "content": "tor to simulate human preferences. We relabel the Pick-a-Pic v2 dataset using AI evaluators and fine-tuned the SD1.5 model. As shown in Fig. 6, we observe a further improvement in these two metrics with AI feedback." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.628, + 0.848 + ], + "angle": 0, + "content": "5.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Tab. 2 presents the results of our ablation experiments on SD1.5 and the median values of the preference evaluators on the HPDv2 test set. For a clearer comparison of our" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28635" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.499, + 0.317 + ], + "angle": 0, + "content": "
Median (HPDv2)GPU hours
AestheticPickScoreHPSv2CLIP
Base-SD1.55.349120.71926.64734.276-
DPO-SD1.55.692221.27427.42835.902204.8
Base initialized5.650421.76328.23736.23857.6
β = 30005.773021.80828.50536.78857.6
β = 40005.763721.79628.44836.68557.6
β = 50005.742821.77128.47336.66957.6
Inv-steps n=305.822521.95728.56837.019136
Inv-steps n=55.771521.84528.43236.76943.2
Inv-steps n=35.748921.83328.50836.54236
CFG win=15.738121.83528.51636.45162.4
CFG win=55.732021.71228.46236.426104
CFG win=7.55.749521.74228.47336.469104
InPO-SD1.55.773421.89428.52336.87657.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.385 + ], + "angle": 0, + "content": "Table 2. Ablation studies of our DDIM-InPO on fine-tuning SD1.5. In each column of preference evaluators, the highest value is highlighted in bold, and the second highest is underlined. For detailed analysis, refer to Sec. 5.4" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.441 + ], + "angle": 0, + "content": "method's effectiveness, we present the results for the Base-SD1.5 and DPO-SD1.5 baselines in rows 3 and 4 of Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.596 + ], + "angle": 0, + "content": "Proposed improvements. The fifth row of Tab. 2 presents the results of initializing \\( p_{\\theta}^{c} \\) in Eq. (11) with the SD1.5 base model. After applying our core contribution, DDIM inversion, we observe a significant improvement in quality compared to the median values of the evaluators for the baselines, indicating that enhancement primarily stems from our method. As instructed by [48], practically an available SFT model can be used to initialize the reference model \\( p_{\\mathrm{ref}}^{c} \\). This strategy aids in reducing the issue of distribution shift." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Parameters Choices. We set \\(\\beta\\) to 2000 for SD1.5. Rows 6 to 8 of Tab. 2 indicate that if \\(\\beta\\) is set too high, the KL divergence penalty term severely limits the model's flexibility in adjustments. It also maintains a consistent learning rate with other baselines, ensuring a fair comparison in terms of efficiency. Rows 9 to 11 of Tab. 2 show the steps required for DDIM inversion. We observe that setting inversion steps to 30 yields the best performance, further confirming the effectiveness of our method. However, this also increases the training time. Thus we select inversion steps \\(n = 10\\) as it strikes a good balance between generation quality and training speed. The work of [38, 39] indicates that DDIM inversion can lead to considerable numerical errors when Classifier Free Guidance (CFG) [26] \\(w_{\\mathrm{inv}} > 0\\). Rows 12 to 14 of Tab. 2 demonstrate that a CFG \\(w_{\\mathrm{inv}} > 0\\) during the inversion process adversely impacts the fine-tuning of human preferences. Consequently, we utilize unconditional DDIM inversion, setting CFG \\(w_{\\mathrm{inv}} = 0\\). We optimize parameters to balance generation quality and training efficiency." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.906, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.339, + 0.907, + 0.41 + ], + "angle": 0, + "content": "Figure 7. Qualitative evaluation of InPO-SDXL in comparison with Base-SDXL and DPO-SDXL on conditional generation tasks (From top to bottom: depth map, canny edge, and inpainting). Prompts: (1) A dolphin in an astronaut suit on Saturn. (2) The Great Pyramid of Giza. (3) A cat riding on a horse." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.727, + 0.455 + ], + "angle": 0, + "content": "5.5. Conditional generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.629 + ], + "angle": 0, + "content": "After fine-tuning the model using the DDIM-InPO method, it can be directly applied to conditional generation tasks without the need for further training. We add conditional control to InPO-SDXL using ControlNet [74] which originally trained with Base-SDXL. Specifically, we use depth map and canny edge as conditions for text-to-image generation tasks. Furthermore, we evaluate our model on image inpainting [37] tasks. As shown in Fig. 7, the advantages of our model are seamlessly transferred to conditional generation tasks, demonstrating superior generation quality compared with Base-SDXL and DPO-SDXL." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.634, + 0.663 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We introduce DDIM-InPO, an efficient method for aligning diffusion models with human preferences, focusing on directly fine-tuning specific latent variables that are strongly correlated with the target image. By fine-tuning the T2I diffusion models using the Pick-a-Pic v2 dataset, our experiments show that DDIM-InPO significantly enhances training efficiency, achieving state-of-the-art performance in just 400 training steps and surpassing all baseline models across nearly all evaluators. DDIM-InPO is based on offline fine-tuning and is sensitive to dataset quality; the presence of harmful, violent, and pornographic content can affect the results, leading to the generation of inappropriate outputs, such as overly feminine images. In addition, we consider the online training paradigm to be a promising and reliable approach for the future." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28636" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.177 + ], + "angle": 0, + "content": "This work was supported by the National Major Science and Technology Projects (the grant number 2022ZD0117000) and the National Natural Science Foundation of China (grant number 62202426)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.189, + 0.188, + 0.205 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.284, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[2] Mohammad Gheshlaghi Azar, Mark Rowland, Bilal Piot, Daniel Guo, Daniele Calandriello, Michal Valko, and Rémi Munos. A general theoretical paradigm to understand learning from human preferences. ArXiv, abs/2310.12036, 2023. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.354, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[3] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.424, + 0.484, + 0.507 + ], + "angle": 0, + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. *ArXiv*, abs/2211.01324, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.484, + 0.579 + ], + "angle": 0, + "content": "[5] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[6] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. ArXiv, abs/2305.13301, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[7] A. Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, and Dominik Lorenz. Stable video diffusion: Scaling latent video diffusion models to large datasets. ArXiv, abs/2311.15127, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[8] A. Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22563-22575, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[9] Haoxin Chen, Menghan Xia, Yin-Yin He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao-Liang Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation. ArXiv, abs/2310.19512, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ArXiv, abs/2310.00426, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.214, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[11] Paul Francis Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. ArXiv, abs/1706.03741, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[12] Kevin Clark, Paul Vicol, Kevin Swersky, and David J. Fleet. Directly fine-tuning diffusion models on differentiable rewards. ArXiv, abs/2309.17400, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.318 + ], + "angle": 0, + "content": "[13] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam S. Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Kumar Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yiqian Wen, Yi-Zhe Song, Roshan Sumbaly, Vignesh Ramanathan, Zijian He, Péter Vajda, and Devi Parikh. Emu: Enhancing image generation models using photogenic needles in a haystack. ArXiv, abs/2309.15807, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[14] Quan Dao, Hao Phung, Binh Duc Nguyen, and A. Tran. Flow matching in latent space. ArXiv, abs/2307.08698, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.906, + 0.376 + ], + "angle": 0, + "content": "[15] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. ArXiv, abs/2105.05233, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[16] Hanze Dong, Wei Xiong, Deepanshu Goyal, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and T. Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. ArXiv, abs/2304.06767, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[17] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7312-7322, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.6 + ], + "angle": 0, + "content": "[18] Patrick Esser, Sumith Kulal, A. Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. ArXiv, abs/2403.03206, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[19] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. ArXiv, abs/2402.01306, 2024. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.906, + 0.715 + ], + "angle": 0, + "content": "[20] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, P. Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. ArXiv, abs/2305.16381, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.906, + 0.771 + ], + "angle": 0, + "content": "[21] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising. ArXiv, abs/2403.14602, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.772, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[22] Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. ArXiv, abs/2310.11513, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[23] Han Guo, Bowen Tan, Zhengzhong Liu, Eric P. Xing, and Zhiting Hu. Efficient (soft) q-learning for text generation with limited good data. In Conference on Empirical Methods in Natural Language Processing, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28637" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Photorealistic video generation with diffusion models. In European Conference on Computer Vision, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[25] Amir Hertz, Ron Mokady, Jay M. Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. ArXiv, abs/2208.01626, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[26] Jonathan Ho. Classifier-free diffusion guidance. ArXiv, abs/2207.12598, 2022. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.211, + 0.482, + 0.279 + ], + "angle": 0, + "content": "[27] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. ArXiv, abs/2210.02303, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.282, + 0.482, + 0.323 + ], + "angle": 0, + "content": "[28] Jiwoo Hong, Noah Lee, and James Thorne. Orpo: Monolithic preference optimization without reference model. ArXiv, abs/2403.07691, 2024. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.326, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[29] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. ArXiv, abs/2402.17525, 2024. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.384, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[30] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. ArXiv, abs/2206.00364, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[31] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 36652-36663, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.499, + 0.482, + 0.554 + ], + "angle": 0, + "content": "[32] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. arXiv preprint arXiv:2404.04465, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.557, + 0.482, + 0.612 + ], + "angle": 0, + "content": "[33] Zhanhao Liang, Yuhui Yuan, Shuyang Gu, Bohan Chen, Tiankai Hang, Ji Li, and Liang Zheng. Step-aware preference optimization: Aligning preference with denoising performance at each step. ArXiv, abs/2406.04314, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[34] Tianqi Liu, Yao Zhao, Rishabh Joshi, Misha Khalman, Mohammad Saleh, Peter J. Liu, and Jialu Liu. Statistical rejection sampling improves preference optimization. ArXiv, abs/2309.06657, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[35] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. ArXiv, abs/1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[36] Artem Lukoianov, Haitz Sáez de Ocariz Borde, Kristjan Greenewald, Vitor Campagnolo Guizilini, Timur Bagautdinov, Vincent Sitzmann, and Justin Solomon. Score distillation via reparametrized ddim. arXiv preprint arXiv:2405.15891, 2024. 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[37] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 8, 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[39] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[40] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. ArXiv, abs/2203.02155, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[41] William S. Peebles and Saining Xie. Scalable diffusion models with transformers. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4172-4182, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[42] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[43] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[44] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. ArXiv, abs/2209.14988, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[45] Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. ArXiv, abs/2310.03739, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[46] Mihir Prabhudesai, Russell Mendonca, Zheyang Qin, Katerina Fragkiadaki, and Deepak Pathak. Video diffusion alignment via reward gradients. ArXiv, abs/2407.08737, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.906, + 0.743 + ], + "angle": 0, + "content": "[48] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[50] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[51] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International Conference on Machine Learning, 2016. 2, 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28638" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[54] Axel Sauer, Dominik Lorenz, A. Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.411 + ], + "angle": 0, + "content": "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv, abs/2210.08402, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[56] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. ArXiv, abs/1707.06347, 2017. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[57] Noam M. Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. ArXiv, abs/1804.04235, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[58] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8871-8879, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[59] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[60] Feifan Song, Yu Bowen, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. Preference ranking optimization for human alignment. ArXiv, abs/2306.17492, 2023. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.483, + 0.733 + ], + "angle": 0, + "content": "[61] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[62] Yang Song, Jascha Narain Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. ArXiv, abs/2011.13456, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[63] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[64] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. ArXiv, abs/2311.03079, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[65] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. ArXiv, abs/2306.09341, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[66] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2096-2105, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[67] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. ArXiv, abs/2304.05977, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[68] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Qimai Li, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8941-8951, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[69] Shentao Yang, Tianqi Chen, and Mingyuan Zhou. A dense reward view on aligning text-to-image diffusion with preference. ArXiv, abs/2402.08265, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[70] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Trans. Mach. Learn. Res., 2022, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[71] Hangjie Yuan, Shiwei Zhang, Xiang Wang, Yujie Wei, Tao Feng, Yining Pan, Yingya Zhang, Ziwei Liu, Samuel Albanie, and Dong Ni. Instructvideo: Instructing video diffusion models with human feedback. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6463-6474, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[72] Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Feiran Huang. Rrhf: Rank responses to align language models with human feedback without tears. ArXiv, abs/2304.05302, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.685, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[73] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1204-1213, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.742, + 0.905, + 0.797 + ], + "angle": 0, + "content": "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3813-3824, 2023. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.799, + 0.905, + 0.841 + ], + "angle": 0, + "content": "[75] Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. arXiv preprint arXiv:2401.12244, 4, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.841 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28639" + } + ] +] \ No newline at end of file diff --git a/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/14854e80-d5af-4094-8f56-d57259d0057b_origin.pdf b/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/14854e80-d5af-4094-8f56-d57259d0057b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fd56f1ecd514ad74c556fa303af4ff610591399e --- /dev/null +++ b/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/14854e80-d5af-4094-8f56-d57259d0057b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e04411ab429fc47f7c461b3abb1c33c565e297074788afa163601f8aebf8f5f +size 3706029 diff --git a/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/full.md b/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6de0b3972890f13a0ca2ab35f58660ba2720bbe1 --- /dev/null +++ b/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/full.md @@ -0,0 +1,350 @@ +# InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment + +Yunhong Lu $^{1}$ Qichao Wang $^{1}$ Hengyuan Cao $^{1}$ Xierui Wang $^{1}$ Xiaoyin Xu $^{1}$ Min Zhang $^{1,2*}$ $^{1}$ Zhejiang University $^{2}$ Shanghai Institute for Advanced Study-Zhejiang University +{yunhonglu,qichaowang,caohy,sherrywang,xiaoyinxu,min_zhang}@zju.edu.cn + +# Abstract + +Without using explicit reward, direct preference optimization (DPO) employs paired human preference data to fine-tune generative models, a method that has garnered considerable attention in large language models (LLMs). However, exploration of aligning text-to-image (T2I) diffusion models with human preferences remains limited. In comparison to supervised fine-tuning, existing methods that align diffusion model suffer from low training efficiency and subpar generation quality due to the long Markov chain process and the intractability of the reverse process. To address these limitations, we introduce DDIM-InPO, an efficient method for direct preference alignment of diffusion models. Our approach conceptualizes diffusion model as a single-step generative model, allowing us to fine-tune the outputs of specific latent variables selectively. In order to accomplish this objective, we first assign implicit rewards to any latent variable directly via a reparameterization technique. Then we construct an Inversion technique to estimate appropriate latent variables for preference optimization. This modification process enables the diffusion model to only fine-tune the outputs of latent variables that have a strong correlation with the preference dataset. Experimental results indicate that our DDIM-InPO achieves state-of-the-art performance with just 400 steps of fine-tuning, surpassing all preference aligning baselines for T2I diffusion models in human preference evaluation tasks. + +# 1. Introduction + +In the domain of T2I generation, tailoring image outputs to reflect user preferences is still in an early stage of development. In contrast, large language models (LLMs) [1] have made notable progress in generating human-aligned text, following a two-stage process: first, pre-training on extensively collected large-scale datasets and then, aligning on smaller datasets that capture specific human needs. + +This alignment aims to produce outputs that meet human preferences while retaining model's learned capabilities. In recent years, T2I diffusion models [18, 43, 52] have been at the forefront of image generation. Nevertheless, achieving alignment between model outputs and human preferences remains a significant challenge. + +Recently, researchers have developed several methods tailoring diffusion models to human preferences, yet these approaches generally lack effectiveness. Some methods adopt a two-stage training approach, where a reward function is first trained and then followed by updating the diffusion model using gradient-based back-propagation [12, 45]. These approaches tend to be unstable and are susceptible to reward leakage [75]. Alternatively, RL (reinforcement learning)-based methods model the denoising process as a Markov decision process, with online evaluation of generated images based on reward feedback [6]. However, their effectiveness is constrained by the length of the Markov chain. Other approaches such as [32, 63], inspired by direct preference optimization (DPO) [48], reduce the RL objective to an optimal policy training goal. Nevertheless, due to the influence of sparse rewards [23, 69], these methods' performance is limited, especially when handling out-of-distribution inputs. + +To address these issues, we introduce DDIM-InPO, an efficient approach for direct preference optimization of diffusion models. In response to the challenge of the implicit rewards allocation in the long-chain denoising process, our key idea is to treat the diffusion model as a single-step generation framework, only fine-tuning the outputs of latent variables directly linked to the target image. To achieve this, we first link the latent variables to the target image space in a single step, using a reparameterization method. It allows us to allocate implicit rewards to latent variables at any timestep directly. Secondly, to mitigate the effects of sparse rewards, we use an Inversion technique in the target image space to identify images that are highly correlated with the target and obtain the corresponding latent variables. Finally, we focus on optimizing the outputs of this limited set of latent variables and thus achieve rapid computations, enabling + +![](images/a2b1019296cc1256e9ede8ffa9871198271fdce86b8b048dba91404764f84bbf.jpg) +Figure 1. We develop DDIM-InPO, an efficient method to align diffusion models with human preference. It suffices to directly optimize the outputs of a small set of variables using human feedback data. This figure illustrates the results after 400 fine-tuning steps of SDXL-base-1.0 using our method, showing that the generated images exhibit strong visual appeal and align well with human preferences. + +high-quality optimization of the diffusion model. Our main contributions are three-folds: + +- We propose a novel DPO implicit rewards allocation method for the diffusion model, treating the diffusion model as a single-step generative model. +- We introduce DDIM-InPO, an efficient framework for direct preference optimization of diffusion models. By applying an Inversion technique in the target image space, we estimate images highly correlated with the preference dataset and derive the corresponding latent variables sufficient for fine-tuning. +- In our experiments, we benchmark our method against state-of-the-art fine-tuning approaches for human preference, demonstrating enhanced training efficiency (refer to Fig. 4) and improved generation quality (examples shown in Fig. 1 and Fig. 3). + +# 2. Related Works + +Text-to-Image Generative Models. T2I generation models aim to produce high-quality, high-fidelity images from a given text prompt. Additionally, these models have contributed to advancements in related fields such as image editing [29], video generation [9], and 3D modeling [44]. + +Among T2I generative models [14, 49-51, 70], diffusion models have recently gained significant attention due to their open-source nature and superior generation performance. However, high-quality diffusion models are typically trained on large, noisy datasets, which often leads to deviations from human preferences, an area that still requires further refinement. + +Diffusion Models Alignment. In diffusion models, supervised fine-tuning (SFT) is commonly employed to align them with human preferences. Inspired by RL-based fine-tuning methods used in LLMs [2, 19, 28, 56, 60], various reward models such as PickScore [31] and HPSv2 [65] have been developed to simulate human preferences. With reward models, approaches like DRaFT [12] and AlignProp [45] use differentiable rewards and back-propagate through the diffusion sampling process, which demands substantial memory and would result in reward leakage. DPOK [20] and DDPO [6] model the sampling process of diffusion models as a Markov decision process and apply policy gradient methods for fine-tuning. Diffusion-DPO [63] and D3PO [68] adopt direct preference optimization by leveraging the denoising steps of diffusion models to optimize + +model parameters at each step. In the DPO framework, DenseReward [69] allocates higher weights to the initial steps. Going further, Diffusion-KTO [32] requires only binary feedback signals for each image to fine-tune the model. However, all these methods are limited to expanding the decision space by incorporating the reward function over all denoising steps, which results in sparse rewards and inefficient training. + +# 3. Preliminaries + +# 3.1. Diffusion Models + +Diffusion models involve a forward process $\{q(\pmb{x}_t)\}_{t\in [0,T]}$ that incrementally adds noise to data $\pmb{x}_0\sim q(\pmb{x}_0)$ and a learned reverse process $\{p(\pmb{x}_t)\}_{t\in [0,T]}$ that aims to denoise the data. The forward process is defined as $q(\pmb{x}_t|\pmb{x}_0)\coloneqq \mathcal{N}(\sqrt{\alpha_t}\pmb{x}_0,(1 - \alpha_t)\mathbf{I})$ and $q(\pmb{x}_t)\coloneqq \int q(\pmb{x}_t|\pmb{x}_0)q(\pmb{x}_0)\mathrm{d}\pmb{x}_0,$ where $\alpha_{t}$ is a noise schedule (as defined in [61]). Starting from $p(\pmb{x}_T)\coloneqq \mathcal{N}(\mathbf{0},\mathbf{I})$ , the reverse process is defined by a parameterized denoiser $\epsilon_{\theta}^{t}(\pmb{x}_{t})$ , which aims to predict noise added to $\pmb{x}_0$ . The denoiser is optimized by minimizing: + +$$ +\mathcal {L} _ {\mathrm {D M}} := \mathbb {E} _ {x _ {0}, t, \epsilon} [ w (t) \| \epsilon_ {\theta} ^ {t} (\sqrt {\alpha_ {t}} \boldsymbol {x} _ {0} + \sqrt {1 - \alpha_ {t}} \epsilon) - \epsilon \| _ {2} ^ {2} ] \tag {1} +$$ + +where $\pmb{x}_0\sim q(\pmb{x}_0),t\sim \mathcal{U}(0,T),\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ and $w(t)$ is a pre-specified weight function. + +Denoising Diffusion Implicit Models (DDIM) generate samples by deterministically reversing the diffusion process. By reparameterizing noisy image $\bar{\pmb{x}}_t = \pmb{x}_t / \sqrt{\alpha_t}$ , the denoising process can be viewed as an ODE: + +$$ +\frac {\mathrm {d} \bar {\boldsymbol {x}} _ {t}}{\mathrm {d} t} = \epsilon_ {\theta} ^ {t} \left(\frac {\bar {\boldsymbol {x}} _ {t}}{\sqrt {\sigma_ {t} ^ {2} + 1}}\right) \frac {\mathrm {d} \sigma_ {t}}{\mathrm {d} t} \tag {2} +$$ + +where $\sigma_{t} = \sqrt{1 - \alpha_{t}} / \sqrt{\alpha_{t}}$ and $\bar{x}_T$ is sampled from Gaussian distribution. The DDIM ODE can also be integrated in reverse to estimate $\pmb{x}_t$ at any timestep $t$ starting from a clean image $\pmb{x}_0$ , a process known as DDIM inversion [39]. + +# 3.2. Human Preference Optimization + +Preference optimization involves fine-tuning generative models to produce samples more closely align with human preferences. Assuming that human preferences are represented by an implicit reward function $r^*$ , there is a dataset $\mathcal{D} = \{(x_0^w, x_0^l, c)\}$ , where $c$ denotes the condition, and $x_0^w$ and $x_0^l$ represent the winner and loser samples, respectively. + +RLHF. RLHF employs maximum likelihood estimation to fit the Bradley-Terry (BT) model on the dataset $\mathcal{D}$ , thereby obtaining an explicit parameterized reward model $r_{\phi}$ . + +$$ +\mathcal {L} _ {\mathrm {B T}} (\phi) := - \mathbb {E} _ {\left(\boldsymbol {x} _ {0} ^ {w}, \boldsymbol {x} _ {0} ^ {l}, \boldsymbol {c}\right) \sim \mathcal {D}} [ \log \sigma \left(r _ {\phi} \left(\boldsymbol {x} _ {0} ^ {w}, \boldsymbol {c}\right) - r _ {\phi} \left(\boldsymbol {x} _ {0} ^ {l}, \boldsymbol {c}\right)\right) ] \tag {3} +$$ + +With the reward, the generative distribution $p_{\theta}$ can then be optimized as a policy using RL that maximizes reward feedback, with an regularization term for the KL-divergence from the reference distribution $p_{\mathrm{ref}}$ : + +$$ +\max _ {p _ {\theta}} \mathbb {E} _ {\substack {\boldsymbol {c} \sim \mathcal {D} _ {\boldsymbol {c}} \\ \boldsymbol {x} _ {0} \sim p _ {\theta} (\boldsymbol {x} _ {0} | \boldsymbol {c})}} [ r (\boldsymbol {x} _ {0}, \boldsymbol {c}) ] - \beta \mathbb {D} _ {\mathrm {K L}} [ p _ {\theta} (\boldsymbol {x} _ {0} | \boldsymbol {c}) | | p _ {\text {ref}} (\boldsymbol {x} _ {0} | \boldsymbol {c}) ] \tag{4} +$$ + +where $\beta$ is a hyperparameter that controls regularization. + +Direct Preference Optimization. Direct Preference Optimization (DPO) considers the unique global optimal policy $p_{\theta}^{*}$ under the given reward function $r$ in Eq. (4): + +$$ +p _ {\theta} ^ {*} \left(\boldsymbol {x} _ {0} \mid \boldsymbol {c}\right) = \frac {p _ {\text {r e f}} \left(\boldsymbol {x} _ {0} \mid \boldsymbol {c}\right) \exp \left(r \left(\boldsymbol {x} _ {0} , \boldsymbol {c}\right) / \beta\right)}{\sum_ {\boldsymbol {x} _ {0}} p _ {\text {r e f}} \left(\boldsymbol {x} _ {0} \mid \boldsymbol {c}\right) \exp \left(r \left(\boldsymbol {x} _ {0} , \boldsymbol {c}\right) / \beta\right)} \tag {5} +$$ + +Then, the reward function $r(\pmb{x}_0, \pmb{c})$ is rewritten according to Eq. (5) and inserted into the BT model from Eq. (3), resulting in the DPO objective function: + +$$ +\mathcal {L} _ {\mathrm {D P O}} := - \underset {\substack {(\boldsymbol {x} _ {0} ^ {w}, \boldsymbol {x} _ {0} ^ {l} \\ , \boldsymbol {c}) \sim \mathcal {D}}} {\mathbb {E}} \log \sigma \left(\beta \log \frac {p _ {\theta} (\boldsymbol {x} _ {0} ^ {w} | \boldsymbol {c})}{p _ {\text {ref}} (\boldsymbol {x} _ {0} ^ {w} | \boldsymbol {c})} - \beta \log \frac {p _ {\theta} (\boldsymbol {x} _ {0} ^ {l} | \boldsymbol {c})}{p _ {\text {ref}} (\boldsymbol {x} _ {0} ^ {l} | \boldsymbol {c})}\right) \tag{6} +$$ + +Based on Eqs. (4) and (6), Wallace et al. [63] redistribute rewards $r(\pmb{x}_0, c)$ across all possible diffusion paths $p_{\theta}(\pmb{x}_{1:T} | \pmb{x}_0, c)$ and minimize the KL-divergence of the joint probability density $\mathbb{D}_{\mathrm{KL}}(p_{\theta}(\pmb{x}_{0:T} | c) || p_{\mathrm{ref}}(\pmb{x}_{0:T} | c))$ , thus designing the DPO objective for diffusion models: + +$$ +\mathcal {L} _ {\mathrm {D P O - D i f f u s i o n}} := - \mathbb {E} _ {\left(\boldsymbol {x} _ {0} ^ {w}, \boldsymbol {x} _ {0} ^ {l}, \boldsymbol {c}\right) \sim \mathcal {D}} \log \sigma +$$ + +$$ +\left. \left(\beta \mathbb {E} _ {\substack {\boldsymbol {x} _ {1: T} ^ {w} \sim p _ {\theta} ^ {c} (\boldsymbol {x} _ {1: T} ^ {w} \mid \boldsymbol {x} _ {0} ^ {w}) \\ \boldsymbol {x} _ {1: T} ^ {l} \sim p _ {\theta} ^ {c} (\boldsymbol {x} _ {1: T} ^ {l} \mid \boldsymbol {x} _ {0} ^ {l})}} \left[ \log \frac {p _ {\theta} ^ {c} (\boldsymbol {x} _ {0 : T} ^ {w})}{p _ {\text {ref}} ^ {c} (\boldsymbol {x} _ {0 : T} ^ {w})} - \log \frac {p _ {\theta} ^ {c} (\boldsymbol {x} _ {0 : T} ^ {l})}{p _ {\text {ref}} ^ {c} (\boldsymbol {x} _ {0 : T} ^ {l})} \right]\right) \right. \tag{7} +$$ + +where $p_{\theta}^{c}(\cdot)$ represents $p_{\theta}(\cdot | c)$ for compactness. Without the need for RL algorithms, the DPO-based method directly optimizes the optimal distribution $p_{\theta}$ . + +# 4. Inversion Preference Optimization + +Given a fixed dataset $\mathcal{D} = \{(\pmb{x}_0^w, \pmb{x}_0^l, \pmb{c})\}$ , where each sample includes a prompt $c$ and an image pair generated by reference model $p_{\mathrm{ref}}$ labeled as $\pmb{x}_0^w \succ \pmb{x}_0^l$ , our goal is to train a model $p_\theta$ that aligns with human preferences. The primary challenge lies in how to match the parameterized distribution $p_\theta^c(\pmb{x}_0)$ according to Eq. (6). Our insight is that the diffusion model can be conceptualized as a timestep-aware single-step generative model $p_\theta^c(\pmb{x}_0|\pmb{x}_t)$ . Specifically, there exists some latent variables $\pmb{x}_t$ that can generate $\pmb{x}_0$ in a single step for any timestep $t$ , which can be written mathematically as: $p_\theta^c(\pmb{x}_0) = \int p_\theta^c(\pmb{x}_0|\pmb{x}_t)p_\theta^c(\pmb{x}_t)\mathrm{d}\pmb{x}_t$ . Thus, the first step is to determine the relationship between $\pmb{x}_t$ and $\pmb{x}_0$ in Sec. 4.1, and the second step is to identify appropriate latent variables $\pmb{x}_t$ for any timestep in Sec. 4.2, within the DPO framework. + +# 4.1. Reparametrized DDIM for DPO + +We establish a relationship between $\boldsymbol{x}_t$ and $\boldsymbol{x}_0$ to be used within the DPO framework. Instead of viewing DDIM as a denoising process within the noisy image $\boldsymbol{x}_t$ space, we reparameterize it with 'initial' variable: + +$$ +\boldsymbol {x} _ {0} (t) := \bar {\boldsymbol {x}} _ {t} - \sigma_ {t} \epsilon_ {\theta} ^ {t} (\boldsymbol {x} _ {t}, \boldsymbol {c}) \tag {8} +$$ + +where noisy image $\bar{\pmb{x}}_t = \pmb {x}_t / \sqrt{\alpha_t}$ . In this context, $\pmb {x}_0(t)$ is the image obtained after the single-step denoising at timestep $t$ and is in $\pmb {x}_0$ space. Therefore, $p_{\theta}^{c}(\pmb {x}_0,\pmb {x}_t)$ is well-defined, allowing us to present a formal definition of 'initial' reward $r(\pmb {x}_0,\pmb {c})$ with the joint' reward $r_t^c (\pmb {x}_0,\pmb {x}_t)$ .. + +$$ +r \left(\boldsymbol {x} _ {0}, \boldsymbol {c}\right) = \mathbb {E} _ {p _ {\theta} ^ {c} \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {0}\right)} \left[ r _ {t} ^ {\boldsymbol {c}} \left(\boldsymbol {x} _ {0}, \boldsymbol {x} _ {t}\right) \right] \tag {9} +$$ + +Similar to [63], we minimize the joint KL divergence $\mathbb{D}_{\mathrm{KL}}[p_{\theta}^{c}(\boldsymbol{x}_0,\boldsymbol{x}_t)||p_{\mathrm{ref}}^{c}(\boldsymbol{x}_0,\boldsymbol{x}_t)]$ to serve as an upper bound for the KL-divergence term in Eq. (4). With Eq. (9) and KL-divergence bound, we can reformulate Eq. (4) into the objective below: + +$$ +\max _ {p _ {\theta} ^ {\mathcal {C}}} \mathbb {E} _ {\substack {t, (\boldsymbol {x} _ {0}, \boldsymbol {x} _ {t}) \sim \\ p _ {\theta} ^ {\mathcal {C}} (\boldsymbol {x} _ {0}, \boldsymbol {x} _ {t})}} [ r _ {t} ^ {\mathcal {C}} (\boldsymbol {x} _ {0}, \boldsymbol {x} _ {t}) ] - \beta \mathbb {D} _ {\mathrm {K L}} [ p _ {\theta} ^ {\mathcal {C}} (\boldsymbol {x} _ {0}, \boldsymbol {x} _ {t}) | | p _ {\text {ref}} ^ {\mathcal {C}} (\boldsymbol {x} _ {0}, \boldsymbol {x} _ {t}) ] \tag{10} +$$ + +In this context, we consider all timesteps. This objective enables the effective allocation of $x_0$ 's reward across any timestep within the model, enhancing training efficiency. In a manner analogous to Eq. (3) through Eq. (6), we derive the following optimization objective: + +$$ +\mathcal {L} (\theta) := - \mathbb {E} _ {t, \left(\boldsymbol {x} _ {0} ^ {w}, \boldsymbol {x} _ {0} ^ {l}, \boldsymbol {c}\right) \sim \mathcal {D}} \log \sigma +$$ + +$$ +\left(\beta \mathbb {E} _ {\substack {\boldsymbol {x} _ {t} ^ {w} \sim p _ {\theta} ^ {\boldsymbol {c}} (\boldsymbol {x} _ {t} ^ {w} | \boldsymbol {x} _ {0} ^ {w}) \\ \boldsymbol {x} _ {t} ^ {l} \sim p _ {\theta} ^ {\boldsymbol {c}} (\boldsymbol {x} _ {t} ^ {l} | \boldsymbol {x} _ {0} ^ {l})}} \left[ \log \frac {p _ {\theta} ^ {\boldsymbol {c}} (\boldsymbol {x} _ {0} ^ {w} , \boldsymbol {x} _ {t} ^ {w})}{p _ {\text {ref}} ^ {\boldsymbol {c}} (\boldsymbol {x} _ {0} ^ {w} , \boldsymbol {x} _ {t} ^ {w})} - \log \frac {p _ {\theta} ^ {\boldsymbol {c}} (\boldsymbol {x} _ {0} ^ {l} , \boldsymbol {x} _ {t} ^ {l})}{p _ {\text {ref}} ^ {\boldsymbol {c}} (\boldsymbol {x} _ {0} ^ {l} , \boldsymbol {x} _ {t} ^ {l})} \right] \right. \tag{11} +$$ + +To optimize Eq. (11), we need to sample $\pmb{x}_t \sim p_\theta^c(\pmb{x}_t | \pmb{x}_0)$ , where $\pmb{x}_t$ should generate an approximation of $\pmb{x}_0$ in a single step. + +# 4.2. Preference Optimization via Inversion + +So how can we determine whether $\pmb{x}_t$ can generate an approximation of $\pmb{x}_0$ in a single step? Given $\pmb{x}_0(t)$ , it remains unclear how to calculate $\pmb{x}_t$ . From Eq. (8), $\pmb{x}_t$ should satisfy the following nonlinear equation: + +$$ +\boldsymbol {x} _ {t} = \sqrt {\alpha_ {t}} \boldsymbol {x} _ {0} (t) + \sqrt {1 - \alpha_ {t}} \epsilon_ {\theta} ^ {t} (\boldsymbol {x} _ {t}, c) \tag {12} +$$ + +and we can rewrite Eq. (12) with respect to noise: + +$$ +\epsilon = \epsilon_ {\theta} ^ {t} \left(\sqrt {\alpha_ {t}} \boldsymbol {x} _ {0} (t) + \sqrt {1 - \alpha_ {t}} \epsilon , \boldsymbol {c}\right) \tag {13} +$$ + +where $\epsilon = (\pmb{x}_t - \sqrt{\alpha_t}\pmb{x}_0(t)) / \sqrt{1 - \alpha_t}$ . Thus, we can define $\delta_t(\pmb{x}_0(t)) = \epsilon$ as the solution when given $\pmb{x}_0(t)$ . In this context, $\pmb{x}_0(t)$ serves as a reliable approximation of $\pmb{x}_0$ at + +![](images/00fd003baa5f2ab149e49f73aa49efdbde4e2cca564c325d41ba134260c24d8d.jpg) +Figure 2. Illustration of Inversion for Preference Optimization. + +timestep $t$ , that is, $x_0(t) \approx x_0$ . Additionally, we rewrite Eq. (2) by applying the definition of $x_0(t)$ (Eq. (8)) by employing some algebra: + +$$ +\frac {\mathrm {d} \boldsymbol {x} _ {0} (t)}{\mathrm {d} t} = - \sigma_ {t} \frac {\mathrm {d}}{\mathrm {d} t} \epsilon_ {\theta} ^ {t} \left(\sqrt {\alpha_ {t}} \boldsymbol {x} _ {0} (t) + \sqrt {1 - \alpha_ {t}} \delta_ {t} (\boldsymbol {x} _ {0} (t)), \boldsymbol {c}\right) \tag {14} +$$ + +Inspired by the work of [36, 39], we suggest to derive $\delta_t(\pmb{x}_0(t))$ by DDIM inversion (It should be noted that $\delta_t(\pmb{x}_0(t))$ can be estimated using various methods, including but not limited to, inversion.) and obtain $\pmb{x}_0(t)$ by inverting Eq. (14) consequently. To clarify the process, we consider the consecutive time points $t$ and $t + \kappa > t$ , and the discretized form of inversion is given as follows: + +$$ +\begin{array}{l} \pmb {x} _ {0} (t + \kappa) = \pmb {x} _ {0} (t) - \sigma_ {t + \kappa} [ \epsilon_ {\theta} ^ {t + \kappa} (\sqrt {\alpha_ {t + \kappa}} \pmb {x} _ {0} (t) \\ \left. + \sqrt {1 - \alpha_ {t + \kappa}} \delta_ {t} (\boldsymbol {x} _ {0} (t))) - \delta_ {t} (\boldsymbol {x} _ {0} (t)) \right] \\ \end{array} +$$ + +According to Eq. (12), $\pmb{x}_t$ is well-defined as follows, and so is $p_{\theta}^{c}(\pmb{x}_{t}|\pmb{x}_{0})$ , ensuring that Eq. (11) can be optimized. + +$$ +\boldsymbol {x} _ {t} = \sqrt {\alpha_ {t}} \boldsymbol {x} _ {0} (t) + \sqrt {1 - \alpha_ {t}} \delta_ {t} (\boldsymbol {x} _ {0} (t)) \tag {16} +$$ + +For the sake of simplicity, we approximate $p_{\mathrm{ref}}^{\mathbf{c}}(\pmb{x}_0,\pmb{x}_t)$ with $p_{\mathrm{ref}}^{\mathbf{c}}(\pmb{x}_0|\pmb{x}_t)p_\theta^{\mathbf{c}}(\pmb{x}_t)$ . Using Jensen's inequality and some algebra, we approximate Eq. (11) as: + +$$ +\begin{array}{l} \mathcal {L} (\theta) = - \mathbb {E} _ {\left(\boldsymbol {x} _ {0} ^ {w}, \boldsymbol {x} _ {0} ^ {l}, \boldsymbol {c}\right) \sim \mathcal {D}, t \sim \mathcal {U} (0, T),} \quad \log \sigma (- \beta ( \\ \boldsymbol {x} _ {t} ^ {w} \sim p _ {\theta} ^ {\boldsymbol {c}} \left(\boldsymbol {x} _ {t} ^ {w} \mid \boldsymbol {x} _ {0} ^ {w}\right), \boldsymbol {x} _ {t} ^ {l} \sim p _ {\theta} ^ {\boldsymbol {c}} \left(\boldsymbol {x} _ {t} ^ {l} \mid \boldsymbol {x} _ {0} ^ {l}\right) \\ + \mathbb {D} _ {\mathrm {K L}} \left(q ^ {\boldsymbol {c}} \left(\boldsymbol {x} _ {0} ^ {w} \mid \boldsymbol {x} _ {t} ^ {w}, \boldsymbol {x} _ {0} ^ {w} (t)\right) \| p _ {\theta} ^ {\boldsymbol {c}} \left(\boldsymbol {x} _ {0} ^ {w} \mid \boldsymbol {x} _ {t} ^ {w}\right)\right) \\ - \mathbb {D} _ {\mathrm {K L}} \left(q ^ {\boldsymbol {c}} \left(\boldsymbol {x} _ {0} ^ {w} \mid \boldsymbol {x} _ {t} ^ {w}, \boldsymbol {x} _ {0} ^ {w} (t)\right) \| p _ {\text {r e f}} ^ {\boldsymbol {c}} \left(\boldsymbol {x} _ {0} ^ {w} \mid \boldsymbol {x} _ {t} ^ {w}\right)\right) \tag {17} \\ - \mathbb {D} _ {\mathrm {K L}} \left(q ^ {\boldsymbol {c}} \left(\boldsymbol {x} _ {0} ^ {l} \mid \boldsymbol {x} _ {t} ^ {l}, \boldsymbol {x} _ {0} ^ {l} (t)\right) \| p _ {\theta} ^ {\boldsymbol {c}} \left(\boldsymbol {x} _ {0} ^ {l} \mid \boldsymbol {x} _ {t} ^ {l}\right)\right) \\ + \mathbb {D} _ {\mathrm {K L}} (q ^ {\boldsymbol {c}} (\pmb {x} _ {0} ^ {l} | \pmb {x} _ {t} ^ {l}, \pmb {x} _ {0} ^ {l} (t)) \| p _ {\mathrm {r e f}} ^ {\boldsymbol {c}} (\pmb {x} _ {0} ^ {l} | \pmb {x} _ {t} ^ {l}))) \\ \end{array} +$$ + +![](images/a7bf779cbd38521276e37c76d6876ffbb3abc28e8526cc30d4330880a42756f6.jpg) +Figure 3. Top, in human evaluations, InPO-SDXL shows a marked improvement over both DPO-SDXL and Base-SDXL. Bottom, qualitative comparisons among baselines. InPO-SDXL achieves superior prompt alignment and produces images of higher quality. + +where $q^c (\cdot |\cdot)$ is a Gaussian transition kernel. Using Eq.(16), the above loss can be simplified as follows: + +$$ +\begin{array}{l} \mathcal {L} (\theta) = - \mathbb {E} _ {(\boldsymbol {x} _ {0} ^ {w}, \boldsymbol {x} _ {0} ^ {l}, \boldsymbol {c}) \sim \mathcal {D}, t \sim \mathcal {U} (0, T),} \quad \log \sigma (- \beta w (t) ( \\ \boldsymbol {x} _ {t} ^ {w} \sim p _ {\theta} ^ {\boldsymbol {c}} (\boldsymbol {x} _ {t} ^ {w} | \boldsymbol {x} _ {0} ^ {w}), \boldsymbol {x} _ {t} ^ {l} \sim p _ {\theta} ^ {\boldsymbol {c}} (\boldsymbol {x} _ {t} ^ {l} | \boldsymbol {x} _ {0} ^ {l}) \\ \left\| \tau_ {t} ^ {w} - \epsilon_ {\theta} ^ {t} \left(\boldsymbol {x} _ {t} ^ {w}, \boldsymbol {c}\right) \right\| _ {2} ^ {2} - \left\| \tau_ {t} ^ {w} - \epsilon_ {\mathrm {r e f}} ^ {t} \left(\boldsymbol {x} _ {t} ^ {w}, \boldsymbol {c}\right) \right\| _ {2} ^ {2} \\ - \left\| \tau_ {t} ^ {l} - \epsilon_ {\theta} ^ {t} \left(\boldsymbol {x} _ {t} ^ {l}, \boldsymbol {c}\right) \right\| _ {2} ^ {2} + \left\| \tau_ {t} ^ {l} - \epsilon_ {\text {r e f}} ^ {t} \left(\boldsymbol {x} _ {t} ^ {l}, \boldsymbol {c}\right) \right\| _ {2} ^ {2}) \tag {18} \\ \end{array} +$$ + +where $\tau_{t}^{*} = (\pmb{x}_{0}^{*}(t) - \pmb{x}_{0}^{*}) / \sigma_{t} + \delta_{t}(\pmb{x}_{0}^{*}(t))$ (refer to Fig. 2) and $w(t)$ is a weight function defined consistent with [63]. + +Link to Diffusion-DPO. Diffusion-DPO defines a reward function on the denoising trajectory, which expands the decision space because it requires combining the reward functions across all denoise steps. This problem is well known to be closely related to delayed feedback/sparse rewards in reinforcement learning, where effective feedback is only available after generating the entire trajectory. Within our framework, Diffusion-DPO utilizes a rough approximation of $\delta_t$ which uses i.i.d random noise $\delta_t^{\mathrm{DPO}}(\pmb{x}_0(t))\sim$ + +$\mathcal{N}(0,I)$ , matching diffusion incorporated in the forward process of noise addition. + +# 5. Experiments + +# 5.1. Setup + +Datasets and Models. We conduct all experiments on the Pick-a-Pic v2 [31] dataset, which is collected from the Pick-a-Pic web application (refer to [31] for details). Consistent with the setting in [63], approximately $12\%$ of the data is excluded, where preferences cannot be determined. As a result, we obtain 851,293 data pairs and 58,960 unique prompts. Our method is applied to fine-tune the base models Stable Diffusion 1.5 (SD1.5) [52] and Stable Diffusion XL-base-1.0 (SDXL) [43]. + +Basic Hyperparameters. To ensure a fair comparison of training efficiency, we use hyperparameters largely consistent with those in [63]. We employ AdamW [35] as the optimizer for SD1.5 and Adafactor [57] for SDXL. All experiments are conducted on eight NVIDIA H800 GPUs, with each GPU handling a batch size of one pair and gradient + +
ModelBaselinesWin-rate (HPDv2)Win-rate (Parti-Prompts)
AestheticPickScoreHPSv2CLIPAestheticPickScoreHPSv2CLIP
InPO-SDXLvs. Base-SDXL56.3779.2585.3853.3761.7072.8978.3150.55
vs. SFT-SDXL66.6688.3189.9159.0069.7088.5488.9156.99
vs. DPO-SDXL55.8759.2865.8146.9157.1756.9260.1142.28
InPO-SD1.5vs. Base-SD1.580.1985.8490.2266.4474.6373.1681.8661.89
vs. SFT-SD1.556.6366.8458.9157.8752.2762.5058.0357.54
vs. DPO-SD1.570.5074.9484.0661.0968.0163.5475.3157.54
vs. KTO-SD1.559.8167.4160.9460.0359.0162.2558.7657.60
+ +Table 1. Win-rate comparison between DDIM-InPO and baselines. We use prompts from HPDv2 and Parti-Prompts, evaluated across different evaluators. SDXL and SD1.5 are used as base models. For simplicity, we denote models using a 'Method-Base' format, such as InPO-SDXL to represent the SDXL model fine-tuned with the DDIM-InPO method. Win-rates exceeding $50\%$ are highlighted in bold. + +accumulation over 128 steps, resulting in an effective batch size of 1024 pairs (half the batch size compared to [63]). The learning rate is set to $\frac{2000}{\beta} 2.048^{-8}$ , with a linear warm-up phase. For SD1.5 training, $\beta$ is set to 2000, while for SDXL training, it is set to 5000. + +Evaluation. We compare DDIM-InPO with existing baselines using user studies and automatic preference metrics. We use the following four evaluators to evaluate image quality: the LAION aesthetic classifier [55] to assess image aesthetic scores, CLIP [47] to measure text-image alignment, and PickScore [31] and HPSv2 [66] to predict human preferences for given image-text pairs. In test stage, we generated images using both the baseline and our model on the Parti-Prompts (1632 prompts) and HPDv2 (3200 prompts) test sets. The median and win-rate (how frequently does the evaluator prefer the generations from DDIM-InPO over those from baselines) of the aforementioned four evaluators are used as automatic preference metrics for comparison. Additionally, we conduct a user study (16 participants) to compare DDIM-InPO with the baselines. We design three questions as [63]: (1) Which image is your overall preferred choice? (2) Which image is more visually attractive? (3) Which image better matches the text description? We also compare the traininggpu hours (on H800) with baselines. In this work, we adopt the following baselines: base model (SD1.5 and SDXL), supervised fine-tuning (SFT), Diffusion-DPO, and Diffusion-KTO (for SD1.5). + +# 5.2. Primary Results + +Quantitative Results. Firstly, we use the quantitative evaluation method for model preferences outlined in [32, 63]. Tab. 1 presents the win-rates of the InPO-aligned diffusion models in comparison to the baselines. Overall, following fine-tuning with DDIM-InPO, both SD1.5 and SDXL achieve superior performance compared to the baselines across nearly all evaluators and test datasets, thereby vali + +![](images/a15ed25ff296813fa49bcd76607c730c5c108ed63ffd44b055f7e679d985b203.jpg) +Figure 4. Comparison of the trade-off between the quality of generated images and training efficiency following human preference optimization of SD1.5 on the HPDv2 test set. Sizes of the circles represent the volume of training data used. Our DDIM-InPO achieves superior performance, with a training speed that is 18.4 and 3.6 times faster than Diffusion-KTO [32] and Diffusion-DPO [63], respectively, while producing images of higher quality. + +dating the effectiveness of our method. Specifically, HPSv2 evaluation shows that, InPO-SDXL obtains a win-rate of $65.81\%$ against DPO-SDXL, while InPO-SD1.5 achieves a win-rate of $60.94\%$ compared to KTO-SD1.5 on the HPDv2 test set. Moreover, it significantly improves the alignment with human preferences for the base model. According to HPSv2, it achieves win-rate of up to $85.38\%$ and $90.22\%$ when compared to the SD1.5 and SDXL base models, respectively, demonstrating the efficacy of DDIM-InPO. + +It should be noted that our DDIM-InPO significantly advances training efficiency, with a training speed that is 18.4 and 3.6 times faster than Diffusion-KTO and Diffusion-DPO, respectively, as shown in Fig. 4, while producing higher-quality images. Additionally, Fig. 3 (top) presents the results of our user study, where we find that human evaluators prefer the model aligned with DDIM-InPO. The + +![](images/140ec73286f04a450c31a1805c74c853302bd21870aeb76d4d4b1ff8fe998ff4.jpg) +Figure 5. Advantages of DDIM-InPO: Compared with Refinement technique and Diffusion-DPO, we find that images generated by DDIM-InPO exhibit enhanced light and spatial structure, better realism and detail capture, greater imaginative design, stable color consistency, optimized multi-instance layout and text integration in visuals. These are some hidden advantages aligned with human preferences. Prompts from left to right: (1) A small bird sitting in a metal wheel. (2) Trees seen through a car window on a rainy day. (3) Description, An artistic rendering of a cosmic portal with a beach at dusk on the other side. (4) A person in full samurai armor at the beach. (5) A man standing in front of a bunch of doughnuts. (6) A towel with the word 'cat' printed on it, simple and clear text. + +evaluation results clearly show that our method provides a more efficient way for fine-tuning diffusion models with human preferences, demonstrating the effectiveness of our proposed paradigm for learning specific latent variables. + +Qualitative results. We display qualitative comparison results of InPO-SDXL against several baselines (Base-SDXL, DPO-SDXL) in Fig. 3 (bottom). The images generated by InPO-SDXL demonstrate enhanced performance in both textual alignment and visual attractiveness. In addition, Fig. 5 illustrates the advantages of our model, which serves as a method for directly learning human preferences for diffusion models. The generated images showcase improved light and spatial structures, detailed capture, greater creativity, consistent color, organized multi-instance layouts, and text-visual integration. These advantages align closely with the preferences of users. + +# 5.3. AI Preference Optimization + +The high cost of manual annotation has driven researchers to seek alternative approaches. Bai et al. [3] introduce RL from AI feedback as one such solution. It follows naturally that PickScore, and HPSv2 can be employed as AI evalua + +![](images/049d4e9ac21d043c961e5bd66e05ff8de445b28ed0b2d486efd000a2e3702d8b.jpg) +Figure 6. Median of PickScore and HPSv2 comparisons for all baselines and test sets on SD1.5. (Detailed analysis in Sec. 5.3) + +tor to simulate human preferences. We relabel the Pick-a-Pic v2 dataset using AI evaluators and fine-tuned the SD1.5 model. As shown in Fig. 6, we observe a further improvement in these two metrics with AI feedback. + +# 5.4. Ablations + +Tab. 2 presents the results of our ablation experiments on SD1.5 and the median values of the preference evaluators on the HPDv2 test set. For a clearer comparison of our + +
Median (HPDv2)GPU hours
AestheticPickScoreHPSv2CLIP
Base-SD1.55.349120.71926.64734.276-
DPO-SD1.55.692221.27427.42835.902204.8
Base initialized5.650421.76328.23736.23857.6
β = 30005.773021.80828.50536.78857.6
β = 40005.763721.79628.44836.68557.6
β = 50005.742821.77128.47336.66957.6
Inv-steps n=305.822521.95728.56837.019136
Inv-steps n=55.771521.84528.43236.76943.2
Inv-steps n=35.748921.83328.50836.54236
CFG win=15.738121.83528.51636.45162.4
CFG win=55.732021.71228.46236.426104
CFG win=7.55.749521.74228.47336.469104
InPO-SD1.55.773421.89428.52336.87657.6
+ +Table 2. Ablation studies of our DDIM-InPO on fine-tuning SD1.5. In each column of preference evaluators, the highest value is highlighted in bold, and the second highest is underlined. For detailed analysis, refer to Sec. 5.4 + +method's effectiveness, we present the results for the Base-SD1.5 and DPO-SD1.5 baselines in rows 3 and 4 of Tab. 2. + +Proposed improvements. The fifth row of Tab. 2 presents the results of initializing $p_{\theta}^{c}$ in Eq. (11) with the SD1.5 base model. After applying our core contribution, DDIM inversion, we observe a significant improvement in quality compared to the median values of the evaluators for the baselines, indicating that enhancement primarily stems from our method. As instructed by [48], practically an available SFT model can be used to initialize the reference model $p_{\mathrm{ref}}^{c}$ . This strategy aids in reducing the issue of distribution shift. + +Parameters Choices. We set $\beta$ to 2000 for SD1.5. Rows 6 to 8 of Tab. 2 indicate that if $\beta$ is set too high, the KL divergence penalty term severely limits the model's flexibility in adjustments. It also maintains a consistent learning rate with other baselines, ensuring a fair comparison in terms of efficiency. Rows 9 to 11 of Tab. 2 show the steps required for DDIM inversion. We observe that setting inversion steps to 30 yields the best performance, further confirming the effectiveness of our method. However, this also increases the training time. Thus we select inversion steps $n = 10$ as it strikes a good balance between generation quality and training speed. The work of [38, 39] indicates that DDIM inversion can lead to considerable numerical errors when Classifier Free Guidance (CFG) [26] $w_{\mathrm{inv}} > 0$ . Rows 12 to 14 of Tab. 2 demonstrate that a CFG $w_{\mathrm{inv}} > 0$ during the inversion process adversely impacts the fine-tuning of human preferences. Consequently, we utilize unconditional DDIM inversion, setting CFG $w_{\mathrm{inv}} = 0$ . We optimize parameters to balance generation quality and training efficiency. + +![](images/841e6584132d563e24d019930ca2435c9ad0be7b699f7ceb5052fe6ec916c943.jpg) +Figure 7. Qualitative evaluation of InPO-SDXL in comparison with Base-SDXL and DPO-SDXL on conditional generation tasks (From top to bottom: depth map, canny edge, and inpainting). Prompts: (1) A dolphin in an astronaut suit on Saturn. (2) The Great Pyramid of Giza. (3) A cat riding on a horse. + +# 5.5. Conditional generation + +After fine-tuning the model using the DDIM-InPO method, it can be directly applied to conditional generation tasks without the need for further training. We add conditional control to InPO-SDXL using ControlNet [74] which originally trained with Base-SDXL. Specifically, we use depth map and canny edge as conditions for text-to-image generation tasks. Furthermore, we evaluate our model on image inpainting [37] tasks. As shown in Fig. 7, the advantages of our model are seamlessly transferred to conditional generation tasks, demonstrating superior generation quality compared with Base-SDXL and DPO-SDXL. + +# 6. Conclusion + +We introduce DDIM-InPO, an efficient method for aligning diffusion models with human preferences, focusing on directly fine-tuning specific latent variables that are strongly correlated with the target image. By fine-tuning the T2I diffusion models using the Pick-a-Pic v2 dataset, our experiments show that DDIM-InPO significantly enhances training efficiency, achieving state-of-the-art performance in just 400 training steps and surpassing all baseline models across nearly all evaluators. DDIM-InPO is based on offline fine-tuning and is sensitive to dataset quality; the presence of harmful, violent, and pornographic content can affect the results, leading to the generation of inappropriate outputs, such as overly feminine images. In addition, we consider the online training paradigm to be a promising and reliable approach for the future. + +# Acknowledgments + +This work was supported by the National Major Science and Technology Projects (the grant number 2022ZD0117000) and the National Natural Science Foundation of China (grant number 62202426). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] Mohammad Gheshlaghi Azar, Mark Rowland, Bilal Piot, Daniel Guo, Daniele Calandriello, Michal Valko, and Rémi Munos. A general theoretical paradigm to understand learning from human preferences. ArXiv, abs/2310.12036, 2023. 2, 1 +[3] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 7 +[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. *ArXiv*, abs/2211.01324, 2022. 1 +[5] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 1 +[6] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. ArXiv, abs/2305.13301, 2023. 1, 2 +[7] A. Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, and Dominik Lorenz. Stable video diffusion: Scaling latent video diffusion models to large datasets. ArXiv, abs/2311.15127, 2023. 1 +[8] A. Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22563-22575, 2023. 1 +[9] Haoxin Chen, Menghan Xia, Yin-Yin He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao-Liang Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation. ArXiv, abs/2310.19512, 2023. 2 +[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ArXiv, abs/2310.00426, 2023. 1 + +[11] Paul Francis Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. ArXiv, abs/1706.03741, 2017. 1 +[12] Kevin Clark, Paul Vicol, Kevin Swersky, and David J. Fleet. Directly fine-tuning diffusion models on differentiable rewards. ArXiv, abs/2309.17400, 2023. 1, 2 +[13] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam S. Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Kumar Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yiqian Wen, Yi-Zhe Song, Roshan Sumbaly, Vignesh Ramanathan, Zijian He, Péter Vajda, and Devi Parikh. Emu: Enhancing image generation models using photogenic needles in a haystack. ArXiv, abs/2309.15807, 2023. 1 +[14] Quan Dao, Hao Phung, Binh Duc Nguyen, and A. Tran. Flow matching in latent space. ArXiv, abs/2307.08698, 2023. 2 +[15] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. ArXiv, abs/2105.05233, 2021. 1 +[16] Hanze Dong, Wei Xiong, Deepanshu Goyal, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and T. Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. ArXiv, abs/2304.06767, 2023. 1 +[17] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7312-7322, 2023. 1 +[18] Patrick Esser, Sumith Kulal, A. Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. ArXiv, abs/2403.03206, 2024. 1 +[19] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. ArXiv, abs/2402.01306, 2024. 2, 1 +[20] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, P. Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. ArXiv, abs/2305.16381, 2023. 2 +[21] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising. ArXiv, abs/2403.14602, 2024. 1 +[22] Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. ArXiv, abs/2310.11513, 2023. 1 +[23] Han Guo, Bowen Tan, Zhengzhong Liu, Eric P. Xing, and Zhiting Hu. Efficient (soft) q-learning for text generation with limited good data. In Conference on Empirical Methods in Natural Language Processing, 2021. 1 +[24] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. + +Photorealistic video generation with diffusion models. In European Conference on Computer Vision, 2023. 1 +[25] Amir Hertz, Ron Mokady, Jay M. Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. ArXiv, abs/2208.01626, 2022. 1 +[26] Jonathan Ho. Classifier-free diffusion guidance. ArXiv, abs/2207.12598, 2022. 8, 1 +[27] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. ArXiv, abs/2210.02303, 2022. 1 +[28] Jiwoo Hong, Noah Lee, and James Thorne. Orpo: Monolithic preference optimization without reference model. ArXiv, abs/2403.07691, 2024. 2, 1 +[29] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. ArXiv, abs/2402.17525, 2024. 2, 1 +[30] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. ArXiv, abs/2206.00364, 2022. 1 +[31] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 36652-36663, 2023. 2, 5, 6 +[32] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. arXiv preprint arXiv:2404.04465, 2024. 1, 3, 6 +[33] Zhanhao Liang, Yuhui Yuan, Shuyang Gu, Bohan Chen, Tiankai Hang, Ji Li, and Liang Zheng. Step-aware preference optimization: Aligning preference with denoising performance at each step. ArXiv, abs/2406.04314, 2024. 2 +[34] Tianqi Liu, Yao Zhao, Rishabh Joshi, Misha Khalman, Mohammad Saleh, Peter J. Liu, and Jialu Liu. Statistical rejection sampling improves preference optimization. ArXiv, abs/2309.06657, 2023. 1 +[35] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. ArXiv, abs/1711.05101, 2017. 5 +[36] Artem Lukoianov, Haitz Sáez de Ocariz Borde, Kristjan Greenewald, Vitor Campagnolo Guizilini, Timur Bagautdinov, Vincent Sitzmann, and Justin Solomon. Score distillation via reparametrized ddim. arXiv preprint arXiv:2405.15891, 2024. 4, 3 +[37] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 8, 1 +[38] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 8, 1 + +[39] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3, 4, 8 +[40] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. ArXiv, abs/2203.02155, 2022. 1 +[41] William S. Peebles and Saining Xie. Scalable diffusion models with transformers. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4172-4182, 2022. 1 +[42] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. 2023. 1 +[43] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 5 +[44] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. ArXiv, abs/2209.14988, 2022. 2 +[45] Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. ArXiv, abs/2310.03739, 2023. 1, 2 +[46] Mihir Prabhudesai, Russell Mendonca, Zheyang Qin, Katerina Fragkiadaki, and Deepak Pathak. Video diffusion alignment via reward gradients. ArXiv, abs/2407.08737, 2024. 1 +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 6 +[48] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 1, 8 +[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 2 +[50] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. +[51] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International Conference on Machine Learning, 2016. 2, 1 + +[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5 +[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1 +[54] Axel Sauer, Dominik Lorenz, A. Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, 2023. 1 +[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv, abs/2210.08402, 2022.6 +[56] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. ArXiv, abs/1707.06347, 2017. 2, 1 +[57] Noam M. Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. ArXiv, abs/1804.04235, 2018. 5 +[58] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8871-8879, 2023. 1 +[59] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 1 +[60] Feifan Song, Yu Bowen, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. Preference ranking optimization for human alignment. ArXiv, abs/2306.17492, 2023. 2, 1 +[61] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3 +[62] Yang Song, Jascha Narain Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. ArXiv, abs/2011.13456, 2020. 1 +[63] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 1, 2, 3, 4, 5, 6 +[64] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan + +Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. ArXiv, abs/2311.03079, 2023. 1 +[65] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. ArXiv, abs/2306.09341, 2023. 2 +[66] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2096-2105, 2023. 6 +[67] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. ArXiv, abs/2304.05977, 2023. 2 +[68] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Qimai Li, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8941-8951, 2023. 2 +[69] Shentao Yang, Tianqi Chen, and Mingyuan Zhou. A dense reward view on aligning text-to-image diffusion with preference. ArXiv, abs/2402.08265, 2024. 1, 3 +[70] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Trans. Mach. Learn. Res., 2022, 2022. 2 +[71] Hangjie Yuan, Shiwei Zhang, Xiang Wang, Yujie Wei, Tao Feng, Yining Pan, Yingya Zhang, Ziwei Liu, Samuel Albanie, and Dong Ni. Instructvideo: Instructing video diffusion models with human feedback. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6463-6474, 2023. 1 +[72] Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Feiran Huang. Rrhf: Rank responses to align language models with human feedback without tears. ArXiv, abs/2304.05302, 2023. 1 +[73] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1204-1213, 2021. 1 +[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3813-3824, 2023. 8, 1 +[75] Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. arXiv preprint arXiv:2401.12244, 4, 2024. 1 \ No newline at end of file diff --git a/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/images.zip b/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a4a37f4aca1589b71cd65d4d918b91e3258229ed --- /dev/null +++ b/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4ad409ee5f8a63291f31b5926903e77c76c2f50cfe45a81de7c97cf07af6698 +size 1125751 diff --git a/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/layout.json b/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..308925c186cdacb2390dc57b5280e51a3bc4ba3d --- /dev/null +++ b/2025/InPO_ Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment/layout.json @@ -0,0 +1,9562 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 102, + 533, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 102, + 533, + 140 + ], + "spans": [ + { + "bbox": [ + 78, + 102, + 533, + 140 + ], + "type": "text", + "content": "InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "text", + "content": "Yunhong Lu" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "text", + "content": " Qichao Wang" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "text", + "content": " Hengyuan Cao" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "text", + "content": " Xierui Wang" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "text", + "content": " Xiaoyin Xu" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "text", + "content": " Min Zhang" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "text", + "content": " Zhejiang University " + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 160, + 560, + 204 + ], + "type": "text", + "content": " Shanghai Institute for Advanced Study-Zhejiang University \n{yunhonglu,qichaowang,caohy,sherrywang,xiaoyinxu,min_zhang}@zju.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 170, + 231, + 217, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 231, + 217, + 243 + ], + "spans": [ + { + "bbox": [ + 170, + 231, + 217, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 256, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 256, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 297, + 568 + ], + "type": "text", + "content": "Without using explicit reward, direct preference optimization (DPO) employs paired human preference data to fine-tune generative models, a method that has garnered considerable attention in large language models (LLMs). However, exploration of aligning text-to-image (T2I) diffusion models with human preferences remains limited. In comparison to supervised fine-tuning, existing methods that align diffusion model suffer from low training efficiency and subpar generation quality due to the long Markov chain process and the intractability of the reverse process. To address these limitations, we introduce DDIM-InPO, an efficient method for direct preference alignment of diffusion models. Our approach conceptualizes diffusion model as a single-step generative model, allowing us to fine-tune the outputs of specific latent variables selectively. In order to accomplish this objective, we first assign implicit rewards to any latent variable directly via a reparameterization technique. Then we construct an Inversion technique to estimate appropriate latent variables for preference optimization. This modification process enables the diffusion model to only fine-tune the outputs of latent variables that have a strong correlation with the preference dataset. Experimental results indicate that our DDIM-InPO achieves state-of-the-art performance with just 400 steps of fine-tuning, surpassing all preference aligning baselines for T2I diffusion models in human preference evaluation tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 590, + 135, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 135, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 135, + 602 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "content": "In the domain of T2I generation, tailoring image outputs to reflect user preferences is still in an early stage of development. In contrast, large language models (LLMs) [1] have made notable progress in generating human-aligned text, following a two-stage process: first, pre-training on extensively collected large-scale datasets and then, aligning on smaller datasets that capture specific human needs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 232, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 304 + ], + "type": "text", + "content": "This alignment aims to produce outputs that meet human preferences while retaining model's learned capabilities. In recent years, T2I diffusion models [18, 43, 52] have been at the forefront of image generation. Nevertheless, achieving alignment between model outputs and human preferences remains a significant challenge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 305, + 556, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 556, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 556, + 520 + ], + "type": "text", + "content": "Recently, researchers have developed several methods tailoring diffusion models to human preferences, yet these approaches generally lack effectiveness. Some methods adopt a two-stage training approach, where a reward function is first trained and then followed by updating the diffusion model using gradient-based back-propagation [12, 45]. These approaches tend to be unstable and are susceptible to reward leakage [75]. Alternatively, RL (reinforcement learning)-based methods model the denoising process as a Markov decision process, with online evaluation of generated images based on reward feedback [6]. However, their effectiveness is constrained by the length of the Markov chain. Other approaches such as [32, 63], inspired by direct preference optimization (DPO) [48], reduce the RL objective to an optimal policy training goal. Nevertheless, due to the influence of sparse rewards [23, 69], these methods' performance is limited, especially when handling out-of-distribution inputs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "text", + "content": "To address these issues, we introduce DDIM-InPO, an efficient approach for direct preference optimization of diffusion models. In response to the challenge of the implicit rewards allocation in the long-chain denoising process, our key idea is to treat the diffusion model as a single-step generation framework, only fine-tuning the outputs of latent variables directly linked to the target image. To achieve this, we first link the latent variables to the target image space in a single step, using a reparameterization method. It allows us to allocate implicit rewards to latent variables at any timestep directly. Secondly, to mitigate the effects of sparse rewards, we use an Inversion technique in the target image space to identify images that are highly correlated with the target and obtain the corresponding latent variables. Finally, we focus on optimizing the outputs of this limited set of latent variables and thus achieve rapid computations, enabling" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 140, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 140, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 140, + 713 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28629" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 553, + 370 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 370 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 370 + ], + "type": "image", + "image_path": "a2b1019296cc1256e9ede8ffa9871198271fdce86b8b048dba91404764f84bbf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 377, + 555, + 412 + ], + "lines": [ + { + "bbox": [ + 55, + 377, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 555, + 412 + ], + "type": "text", + "content": "Figure 1. We develop DDIM-InPO, an efficient method to align diffusion models with human preference. It suffices to directly optimize the outputs of a small set of variables using human feedback data. This figure illustrates the results after 400 fine-tuning steps of SDXL-base-1.0 using our method, showing that the generated images exhibit strong visual appeal and align well with human preferences." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 432, + 295, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 454 + ], + "type": "text", + "content": "high-quality optimization of the diffusion model. Our main contributions are three-folds:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 456, + 295, + 623 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 456, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 491 + ], + "type": "text", + "content": "- We propose a novel DPO implicit rewards allocation method for the diffusion model, treating the diffusion model as a single-step generative model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 492, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 295, + 563 + ], + "type": "text", + "content": "- We introduce DDIM-InPO, an efficient framework for direct preference optimization of diffusion models. By applying an Inversion technique in the target image space, we estimate images highly correlated with the preference dataset and derive the corresponding latent variables sufficient for fine-tuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 563, + 295, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 623 + ], + "type": "text", + "content": "- In our experiments, we benchmark our method against state-of-the-art fine-tuning approaches for human preference, demonstrating enhanced training efficiency (refer to Fig. 4) and improved generation quality (examples shown in Fig. 1 and Fig. 3)." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 634, + 147, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 147, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 147, + 646 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "content": "Text-to-Image Generative Models. T2I generation models aim to produce high-quality, high-fidelity images from a given text prompt. Additionally, these models have contributed to advancements in related fields such as image editing [29], video generation [9], and 3D modeling [44]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 432, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 515 + ], + "type": "text", + "content": "Among T2I generative models [14, 49-51, 70], diffusion models have recently gained significant attention due to their open-source nature and superior generation performance. However, high-quality diffusion models are typically trained on large, noisy datasets, which often leads to deviations from human preferences, an area that still requires further refinement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": "Diffusion Models Alignment. In diffusion models, supervised fine-tuning (SFT) is commonly employed to align them with human preferences. Inspired by RL-based fine-tuning methods used in LLMs [2, 19, 28, 56, 60], various reward models such as PickScore [31] and HPSv2 [65] have been developed to simulate human preferences. With reward models, approaches like DRaFT [12] and AlignProp [45] use differentiable rewards and back-propagate through the diffusion sampling process, which demands substantial memory and would result in reward leakage. DPOK [20] and DDPO [6] model the sampling process of diffusion models as a Markov decision process and apply policy gradient methods for fine-tuning. Diffusion-DPO [63] and D3PO [68] adopt direct preference optimization by leveraging the denoising steps of diffusion models to optimize" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28630" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "model parameters at each step. In the DPO framework, DenseReward [69] allocates higher weights to the initial steps. Going further, Diffusion-KTO [32] requires only binary feedback signals for each image to fine-tune the model. However, all these methods are limited to expanding the decision space by incorporating the reward function over all denoising steps, which results in sparse rewards and inefficient training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 178, + 141, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 141, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 141, + 190 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 197, + 159, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 159, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 159, + 209 + ], + "type": "text", + "content": "3.1. Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "text", + "content": "Diffusion models involve a forward process " + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\{q(\\pmb{x}_t)\\}_{t\\in [0,T]}" + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "text", + "content": " that incrementally adds noise to data " + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0\\sim q(\\pmb{x}_0)" + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "text", + "content": " and a learned reverse process " + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\{p(\\pmb{x}_t)\\}_{t\\in [0,T]}" + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "text", + "content": " that aims to denoise the data. The forward process is defined as " + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_t|\\pmb{x}_0)\\coloneqq \\mathcal{N}(\\sqrt{\\alpha_t}\\pmb{x}_0,(1 - \\alpha_t)\\mathbf{I})" + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_t)\\coloneqq \\int q(\\pmb{x}_t|\\pmb{x}_0)q(\\pmb{x}_0)\\mathrm{d}\\pmb{x}_0," + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "text", + "content": " is a noise schedule (as defined in [61]). Starting from " + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_T)\\coloneqq \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "text", + "content": " , the reverse process is defined by a parameterized denoiser " + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{t}(\\pmb{x}_{t})" + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "text", + "content": " , which aims to predict noise added to " + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "text", + "content": " . The denoiser is optimized by minimizing:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 329, + 294, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 329, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 63, + 329, + 294, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D M}} := \\mathbb {E} _ {x _ {0}, t, \\epsilon} [ w (t) \\| \\epsilon_ {\\theta} ^ {t} (\\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon) - \\epsilon \\| _ {2} ^ {2} ] \\tag {1}", + "image_path": "09ce255542b633648e805407845e8bfcef47de5ea2231bbfa989f523deb37b00.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0\\sim q(\\pmb{x}_0),t\\sim \\mathcal{U}(0,T),\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "content": " is a pre-specified weight function." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 380, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 429 + ], + "type": "text", + "content": "Denoising Diffusion Implicit Models (DDIM) generate samples by deterministically reversing the diffusion process. By reparameterizing noisy image " + }, + { + "bbox": [ + 55, + 380, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{x}}_t = \\pmb{x}_t / \\sqrt{\\alpha_t}" + }, + { + "bbox": [ + 55, + 380, + 296, + 429 + ], + "type": "text", + "content": ", the denoising process can be viewed as an ODE:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 435, + 295, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 435, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 115, + 435, + 295, + 468 + ], + "type": "interline_equation", + "content": "\\frac {\\mathrm {d} \\bar {\\boldsymbol {x}} _ {t}}{\\mathrm {d} t} = \\epsilon_ {\\theta} ^ {t} \\left(\\frac {\\bar {\\boldsymbol {x}} _ {t}}{\\sqrt {\\sigma_ {t} ^ {2} + 1}}\\right) \\frac {\\mathrm {d} \\sigma_ {t}}{\\mathrm {d} t} \\tag {2}", + "image_path": "e27fd3241182dc5142ac6b9024d8c066db486a26b492089306dbc9d90b61c28a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 475, + 296, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 524 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 475, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\sigma_{t} = \\sqrt{1 - \\alpha_{t}} / \\sqrt{\\alpha_{t}}" + }, + { + "bbox": [ + 55, + 475, + 296, + 524 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 475, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\bar{x}_T" + }, + { + "bbox": [ + 55, + 475, + 296, + 524 + ], + "type": "text", + "content": " is sampled from Gaussian distribution. The DDIM ODE can also be integrated in reverse to estimate " + }, + { + "bbox": [ + 55, + 475, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 55, + 475, + 296, + 524 + ], + "type": "text", + "content": " at any timestep " + }, + { + "bbox": [ + 55, + 475, + 296, + 524 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 475, + 296, + 524 + ], + "type": "text", + "content": " starting from a clean image " + }, + { + "bbox": [ + 55, + 475, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 55, + 475, + 296, + 524 + ], + "type": "text", + "content": ", a process known as DDIM inversion [39]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 529, + 231, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 231, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 231, + 542 + ], + "type": "text", + "content": "3.2. Human Preference Optimization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "content": "Preference optimization involves fine-tuning generative models to produce samples more closely align with human preferences. Assuming that human preferences are represented by an implicit reward function " + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "inline_equation", + "content": "r^*" + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "content": ", there is a dataset " + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_0^w, x_0^l, c)\\}" + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "content": " denotes the condition, and " + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "inline_equation", + "content": "x_0^w" + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "inline_equation", + "content": "x_0^l" + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "content": " represent the winner and loser samples, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "text", + "content": "RLHF. RLHF employs maximum likelihood estimation to fit the Bradley-Terry (BT) model on the dataset " + }, + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "text", + "content": ", thereby obtaining an explicit parameterized reward model " + }, + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "inline_equation", + "content": "r_{\\phi}" + }, + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 688, + 296, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 688, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 688, + 296, + 713 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {B T}} (\\phi) := - \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right) \\sim \\mathcal {D}} [ \\log \\sigma \\left(r _ {\\phi} \\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {c}\\right) - r _ {\\phi} \\left(\\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right)\\right) ] \\tag {3}", + "image_path": "6207df10a6975cb3cf2cae1b8c206c95be0cc317029d7df5d3b85c11b63c10bc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "With the reward, the generative distribution " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " can then be optimized as a policy using RL that maximizes reward feedback, with an regularization term for the KL-divergence from the reference distribution " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 127, + 555, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 127, + 555, + 158 + ], + "spans": [ + { + "bbox": [ + 314, + 127, + 555, + 158 + ], + "type": "interline_equation", + "content": "\\max _ {p _ {\\theta}} \\mathbb {E} _ {\\substack {\\boldsymbol {c} \\sim \\mathcal {D} _ {\\boldsymbol {c}} \\\\ \\boldsymbol {x} _ {0} \\sim p _ {\\theta} (\\boldsymbol {x} _ {0} | \\boldsymbol {c})}} [ r (\\boldsymbol {x} _ {0}, \\boldsymbol {c}) ] - \\beta \\mathbb {D} _ {\\mathrm {K L}} [ p _ {\\theta} (\\boldsymbol {x} _ {0} | \\boldsymbol {c}) | | p _ {\\text {ref}} (\\boldsymbol {x} _ {0} | \\boldsymbol {c}) ] \\tag{4}", + "image_path": "67c7e182db4ac83dfd92065704149eb7b7b474e5899fb2981797cb80ba125447.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 158, + 543, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 158, + 543, + 171 + ], + "spans": [ + { + "bbox": [ + 314, + 158, + 543, + 171 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 314, + 158, + 543, + 171 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 314, + 158, + 543, + 171 + ], + "type": "text", + "content": " is a hyperparameter that controls regularization." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 184, + 554, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 554, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 554, + 220 + ], + "type": "text", + "content": "Direct Preference Optimization. Direct Preference Optimization (DPO) considers the unique global optimal policy " + }, + { + "bbox": [ + 313, + 184, + 554, + 220 + ], + "type": "inline_equation", + "content": "p_{\\theta}^{*}" + }, + { + "bbox": [ + 313, + 184, + 554, + 220 + ], + "type": "text", + "content": " under the given reward function " + }, + { + "bbox": [ + 313, + 184, + 554, + 220 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 184, + 554, + 220 + ], + "type": "text", + "content": " in Eq. (4):" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 339, + 227, + 554, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 227, + 554, + 255 + ], + "spans": [ + { + "bbox": [ + 339, + 227, + 554, + 255 + ], + "type": "interline_equation", + "content": "p _ {\\theta} ^ {*} \\left(\\boldsymbol {x} _ {0} \\mid \\boldsymbol {c}\\right) = \\frac {p _ {\\text {r e f}} \\left(\\boldsymbol {x} _ {0} \\mid \\boldsymbol {c}\\right) \\exp \\left(r \\left(\\boldsymbol {x} _ {0} , \\boldsymbol {c}\\right) / \\beta\\right)}{\\sum_ {\\boldsymbol {x} _ {0}} p _ {\\text {r e f}} \\left(\\boldsymbol {x} _ {0} \\mid \\boldsymbol {c}\\right) \\exp \\left(r \\left(\\boldsymbol {x} _ {0} , \\boldsymbol {c}\\right) / \\beta\\right)} \\tag {5}", + "image_path": "8abf1fa0d1cc993f83c069a1616b01ca2e14281463167934400fa4fd069dd2f9.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 262, + 554, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 554, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 554, + 292 + ], + "type": "text", + "content": "Then, the reward function " + }, + { + "bbox": [ + 313, + 262, + 554, + 292 + ], + "type": "inline_equation", + "content": "r(\\pmb{x}_0, \\pmb{c})" + }, + { + "bbox": [ + 313, + 262, + 554, + 292 + ], + "type": "text", + "content": " is rewritten according to Eq. (5) and inserted into the BT model from Eq. (3), resulting in the DPO objective function:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 295, + 553, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 295, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 315, + 295, + 553, + 331 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D P O}} := - \\underset {\\substack {(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l} \\\\ , \\boldsymbol {c}) \\sim \\mathcal {D}}} {\\mathbb {E}} \\log \\sigma \\left(\\beta \\log \\frac {p _ {\\theta} (\\boldsymbol {x} _ {0} ^ {w} | \\boldsymbol {c})}{p _ {\\text {ref}} (\\boldsymbol {x} _ {0} ^ {w} | \\boldsymbol {c})} - \\beta \\log \\frac {p _ {\\theta} (\\boldsymbol {x} _ {0} ^ {l} | \\boldsymbol {c})}{p _ {\\text {ref}} (\\boldsymbol {x} _ {0} ^ {l} | \\boldsymbol {c})}\\right) \\tag{6}", + "image_path": "6b1a2ef91b7d0a1f0e502797dbfa5d47b22f52dd030229d68c718ce873fa357b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "text", + "content": "Based on Eqs. (4) and (6), Wallace et al. [63] redistribute rewards " + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "inline_equation", + "content": "r(\\pmb{x}_0, c)" + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "text", + "content": " across all possible diffusion paths " + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}_{1:T} | \\pmb{x}_0, c)" + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "text", + "content": " and minimize the KL-divergence of the joint probability density " + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{KL}}(p_{\\theta}(\\pmb{x}_{0:T} | c) || p_{\\mathrm{ref}}(\\pmb{x}_{0:T} | c))" + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "text", + "content": ", thus designing the DPO objective for diffusion models:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 399, + 480, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 399, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 315, + 399, + 480, + 414 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D P O - D i f f u s i o n}} := - \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right) \\sim \\mathcal {D}} \\log \\sigma", + "image_path": "f1b8ac5c4632f197351e006dc69848a9d448237e84175f85d1c6d2d018b09837.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 416, + 554, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 554, + 460 + ], + "type": "interline_equation", + "content": "\\left. \\left(\\beta \\mathbb {E} _ {\\substack {\\boldsymbol {x} _ {1: T} ^ {w} \\sim p _ {\\theta} ^ {c} (\\boldsymbol {x} _ {1: T} ^ {w} \\mid \\boldsymbol {x} _ {0} ^ {w}) \\\\ \\boldsymbol {x} _ {1: T} ^ {l} \\sim p _ {\\theta} ^ {c} (\\boldsymbol {x} _ {1: T} ^ {l} \\mid \\boldsymbol {x} _ {0} ^ {l})}} \\left[ \\log \\frac {p _ {\\theta} ^ {c} (\\boldsymbol {x} _ {0 : T} ^ {w})}{p _ {\\text {ref}} ^ {c} (\\boldsymbol {x} _ {0 : T} ^ {w})} - \\log \\frac {p _ {\\theta} ^ {c} (\\boldsymbol {x} _ {0 : T} ^ {l})}{p _ {\\text {ref}} ^ {c} (\\boldsymbol {x} _ {0 : T} ^ {l})} \\right]\\right) \\right. \\tag{7}", + "image_path": "920a79dc6a086ad5e3e6da9b5567996313a4c24c1f51f82a5dce40966fdc970b.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 468, + 554, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 504 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 468, + 554, + 504 + ], + "type": "inline_equation", + "content": "p_{\\theta}^{c}(\\cdot)" + }, + { + "bbox": [ + 313, + 468, + 554, + 504 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 313, + 468, + 554, + 504 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\cdot | c)" + }, + { + "bbox": [ + 313, + 468, + 554, + 504 + ], + "type": "text", + "content": " for compactness. Without the need for RL algorithms, the DPO-based method directly optimizes the optimal distribution " + }, + { + "bbox": [ + 313, + 468, + 554, + 504 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 313, + 468, + 554, + 504 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 514, + 506, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 506, + 527 + ], + "type": "text", + "content": "4. Inversion Preference Optimization" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": "Given a fixed dataset " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(\\pmb{x}_0^w, \\pmb{x}_0^l, \\pmb{c})\\}" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": ", where each sample includes a prompt " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": " and an image pair generated by reference model " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": " labeled as " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^w \\succ \\pmb{x}_0^l" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": ", our goal is to train a model " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "p_\\theta" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": " that aligns with human preferences. The primary challenge lies in how to match the parameterized distribution " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "p_\\theta^c(\\pmb{x}_0)" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": " according to Eq. (6). Our insight is that the diffusion model can be conceptualized as a timestep-aware single-step generative model " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "p_\\theta^c(\\pmb{x}_0|\\pmb{x}_t)" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": ". Specifically, there exists some latent variables " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": " that can generate " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": " in a single step for any timestep " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": ", which can be written mathematically as: " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "p_\\theta^c(\\pmb{x}_0) = \\int p_\\theta^c(\\pmb{x}_0|\\pmb{x}_t)p_\\theta^c(\\pmb{x}_t)\\mathrm{d}\\pmb{x}_t" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": ". Thus, the first step is to determine the relationship between " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": " in Sec. 4.1, and the second step is to identify appropriate latent variables " + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": " for any timestep in Sec. 4.2, within the DPO framework." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28631" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 230, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 230, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 230, + 83 + ], + "type": "text", + "content": "4.1. Reparametrized DDIM for DPO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "content": "We establish a relationship between " + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0" + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "content": " to be used within the DPO framework. Instead of viewing DDIM as a denoising process within the noisy image " + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "content": " space, we reparameterize it with 'initial' variable:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 121, + 145, + 295, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 145, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 121, + 145, + 295, + 159 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {0} (t) := \\bar {\\boldsymbol {x}} _ {t} - \\sigma_ {t} \\epsilon_ {\\theta} ^ {t} (\\boldsymbol {x} _ {t}, \\boldsymbol {c}) \\tag {8}", + "image_path": "5b1409075d9c22bb61eb19b07fb8faa3652064e00668fd1b2fbb3282d25f1a18.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": "where noisy image " + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{x}}_t = \\pmb {x}_t / \\sqrt{\\alpha_t}" + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": " . In this context, " + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\pmb {x}_0(t)" + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": " is the image obtained after the single-step denoising at timestep " + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": " and is in " + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\pmb {x}_0" + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": " space. Therefore, " + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "inline_equation", + "content": "p_{\\theta}^{c}(\\pmb {x}_0,\\pmb {x}_t)" + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": " is well-defined, allowing us to present a formal definition of 'initial' reward " + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "inline_equation", + "content": "r(\\pmb {x}_0,\\pmb {c})" + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": " with the joint' reward " + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "inline_equation", + "content": "r_t^c (\\pmb {x}_0,\\pmb {x}_t)" + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 239, + 295, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 239, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 295, + 253 + ], + "type": "interline_equation", + "content": "r \\left(\\boldsymbol {x} _ {0}, \\boldsymbol {c}\\right) = \\mathbb {E} _ {p _ {\\theta} ^ {c} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right)} \\left[ r _ {t} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}\\right) \\right] \\tag {9}", + "image_path": "0155adc19d66071a0983ff3aa1a7a85c8b1c33e5b297105386a87505efdb00e8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 316 + ], + "type": "text", + "content": "Similar to [63], we minimize the joint KL divergence " + }, + { + "bbox": [ + 55, + 256, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\mathrm{KL}}[p_{\\theta}^{c}(\\boldsymbol{x}_0,\\boldsymbol{x}_t)||p_{\\mathrm{ref}}^{c}(\\boldsymbol{x}_0,\\boldsymbol{x}_t)]" + }, + { + "bbox": [ + 55, + 256, + 296, + 316 + ], + "type": "text", + "content": " to serve as an upper bound for the KL-divergence term in Eq. (4). With Eq. (9) and KL-divergence bound, we can reformulate Eq. (4) into the objective below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 325, + 307, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 307, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 307, + 356 + ], + "type": "interline_equation", + "content": "\\max _ {p _ {\\theta} ^ {\\mathcal {C}}} \\mathbb {E} _ {\\substack {t, (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) \\sim \\\\ p _ {\\theta} ^ {\\mathcal {C}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t})}} [ r _ {t} ^ {\\mathcal {C}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) ] - \\beta \\mathbb {D} _ {\\mathrm {K L}} [ p _ {\\theta} ^ {\\mathcal {C}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) | | p _ {\\text {ref}} ^ {\\mathcal {C}} (\\boldsymbol {x} _ {0}, \\boldsymbol {x} _ {t}) ] \\tag{10}", + "image_path": "c39cdd48adabdf42242b4b4a639cc2b3f89c39218b88fb8e061bc46a1b874eaf.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "type": "text", + "content": "In this context, we consider all timesteps. This objective enables the effective allocation of " + }, + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 357, + 296, + 418 + ], + "type": "text", + "content": "'s reward across any timestep within the model, enhancing training efficiency. In a manner analogous to Eq. (3) through Eq. (6), we derive the following optimization objective:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 426, + 185, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 185, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 185, + 441 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) := - \\mathbb {E} _ {t, \\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right) \\sim \\mathcal {D}} \\log \\sigma", + "image_path": "7585a5dcbaa39c60942d3ee4b3fc98ffe98d0023cc905e8ab6e84b72fd5625ed.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 442, + 296, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 442, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 59, + 442, + 296, + 487 + ], + "type": "interline_equation", + "content": "\\left(\\beta \\mathbb {E} _ {\\substack {\\boldsymbol {x} _ {t} ^ {w} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {t} ^ {w} | \\boldsymbol {x} _ {0} ^ {w}) \\\\ \\boldsymbol {x} _ {t} ^ {l} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {t} ^ {l} | \\boldsymbol {x} _ {0} ^ {l})}} \\left[ \\log \\frac {p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {0} ^ {w} , \\boldsymbol {x} _ {t} ^ {w})}{p _ {\\text {ref}} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {0} ^ {w} , \\boldsymbol {x} _ {t} ^ {w})} - \\log \\frac {p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {0} ^ {l} , \\boldsymbol {x} _ {t} ^ {l})}{p _ {\\text {ref}} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {0} ^ {l} , \\boldsymbol {x} _ {t} ^ {l})} \\right] \\right. \\tag{11}", + "image_path": "c2067874be03a6de5189b68f5dd1c614fc9b40bc7923b1d90f484878a7852c7d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 496, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 532 + ], + "type": "text", + "content": "To optimize Eq. (11), we need to sample " + }, + { + "bbox": [ + 55, + 496, + 295, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t \\sim p_\\theta^c(\\pmb{x}_t | \\pmb{x}_0)" + }, + { + "bbox": [ + 55, + 496, + 295, + 532 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 496, + 295, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 55, + 496, + 295, + 532 + ], + "type": "text", + "content": " should generate an approximation of " + }, + { + "bbox": [ + 55, + 496, + 295, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 55, + 496, + 295, + 532 + ], + "type": "text", + "content": " in a single step." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 539, + 257, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 257, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 257, + 552 + ], + "type": "text", + "content": "4.2. Preference Optimization via Inversion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "text", + "content": "So how can we determine whether " + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "text", + "content": " can generate an approximation of " + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "text", + "content": " in a single step? Given " + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0(t)" + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "text", + "content": ", it remains unclear how to calculate " + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "text", + "content": ". From Eq. (8), " + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 55, + 557, + 296, + 605 + ], + "type": "text", + "content": " should satisfy the following nonlinear equation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 99, + 613, + 295, + 627 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 613, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 99, + 613, + 295, + 627 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t} = \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} (t) + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} ^ {t} (\\boldsymbol {x} _ {t}, c) \\tag {12}", + "image_path": "755976fee2a5b7fdee6ddfddb57802f9e197e2deedd5e47761f3ffe78f1f98fb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 635, + 257, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 257, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 257, + 647 + ], + "type": "text", + "content": "and we can rewrite Eq. (12) with respect to noise:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 655, + 295, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 655, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 295, + 669 + ], + "type": "interline_equation", + "content": "\\epsilon = \\epsilon_ {\\theta} ^ {t} \\left(\\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} (t) + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\boldsymbol {c}\\right) \\tag {13}", + "image_path": "e25cf12b04d655964d48c3ee33e406d38d97e2b046cae81ae4d4877c25295d8b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon = (\\pmb{x}_t - \\sqrt{\\alpha_t}\\pmb{x}_0(t)) / \\sqrt{1 - \\alpha_t}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ". Thus, we can define " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\delta_t(\\pmb{x}_0(t)) = \\epsilon" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " as the solution when given " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0(t)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ". In this context, " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0(t)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " serves as a reliable approximation of " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " at" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 320, + 70, + 552, + 265 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 552, + 265 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 552, + 265 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 552, + 265 + ], + "type": "image", + "image_path": "00fd003baa5f2ab149e49f73aa49efdbde4e2cca564c325d41ba134260c24d8d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 272, + 547, + 283 + ], + "lines": [ + { + "bbox": [ + 320, + 272, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 320, + 272, + 547, + 283 + ], + "type": "text", + "content": "Figure 2. Illustration of Inversion for Preference Optimization." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "type": "text", + "content": "timestep " + }, + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "type": "text", + "content": ", that is, " + }, + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "type": "inline_equation", + "content": "x_0(t) \\approx x_0" + }, + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "type": "text", + "content": ". Additionally, we rewrite Eq. (2) by applying the definition of " + }, + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "type": "inline_equation", + "content": "x_0(t)" + }, + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "type": "text", + "content": " (Eq. (8)) by employing some algebra:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 352, + 553, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 352, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 321, + 352, + 553, + 386 + ], + "type": "interline_equation", + "content": "\\frac {\\mathrm {d} \\boldsymbol {x} _ {0} (t)}{\\mathrm {d} t} = - \\sigma_ {t} \\frac {\\mathrm {d}}{\\mathrm {d} t} \\epsilon_ {\\theta} ^ {t} \\left(\\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} (t) + \\sqrt {1 - \\alpha_ {t}} \\delta_ {t} (\\boldsymbol {x} _ {0} (t)), \\boldsymbol {c}\\right) \\tag {14}", + "image_path": "f943ff8beda31235076d32429e4ef51db10f8605d52555fb5220a519b07a66c1.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 386, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 469 + ], + "type": "text", + "content": "Inspired by the work of [36, 39], we suggest to derive " + }, + { + "bbox": [ + 313, + 386, + 555, + 469 + ], + "type": "inline_equation", + "content": "\\delta_t(\\pmb{x}_0(t))" + }, + { + "bbox": [ + 313, + 386, + 555, + 469 + ], + "type": "text", + "content": " by DDIM inversion (It should be noted that " + }, + { + "bbox": [ + 313, + 386, + 555, + 469 + ], + "type": "inline_equation", + "content": "\\delta_t(\\pmb{x}_0(t))" + }, + { + "bbox": [ + 313, + 386, + 555, + 469 + ], + "type": "text", + "content": " can be estimated using various methods, including but not limited to, inversion.) and obtain " + }, + { + "bbox": [ + 313, + 386, + 555, + 469 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0(t)" + }, + { + "bbox": [ + 313, + 386, + 555, + 469 + ], + "type": "text", + "content": " by inverting Eq. (14) consequently. To clarify the process, we consider the consecutive time points " + }, + { + "bbox": [ + 313, + 386, + 555, + 469 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 386, + 555, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 386, + 555, + 469 + ], + "type": "inline_equation", + "content": "t + \\kappa > t" + }, + { + "bbox": [ + 313, + 386, + 555, + 469 + ], + "type": "text", + "content": ", and the discretized form of inversion is given as follows:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 331, + 479, + 520, + 511 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 479, + 520, + 511 + ], + "spans": [ + { + "bbox": [ + 331, + 479, + 520, + 511 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\pmb {x} _ {0} (t + \\kappa) = \\pmb {x} _ {0} (t) - \\sigma_ {t + \\kappa} [ \\epsilon_ {\\theta} ^ {t + \\kappa} (\\sqrt {\\alpha_ {t + \\kappa}} \\pmb {x} _ {0} (t) \\\\ \\left. + \\sqrt {1 - \\alpha_ {t + \\kappa}} \\delta_ {t} (\\boldsymbol {x} _ {0} (t))) - \\delta_ {t} (\\boldsymbol {x} _ {0} (t)) \\right] \\\\ \\end{array}", + "image_path": "1b96e98d6a5b3d56999743cdd2a96049b1d2d0f12972ba97b1e7999d65528267.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": "According to Eq. (12), " + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": " is well-defined as follows, and so is " + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "inline_equation", + "content": "p_{\\theta}^{c}(\\pmb{x}_{t}|\\pmb{x}_{0})" + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": ", ensuring that Eq. (11) can be optimized." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 358, + 556, + 553, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 556, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 358, + 556, + 553, + 570 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t} = \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {0} (t) + \\sqrt {1 - \\alpha_ {t}} \\delta_ {t} (\\boldsymbol {x} _ {0} (t)) \\tag {16}", + "image_path": "5270cb3c7a849778dcf4091a1547b64898ee00a5baa2cb123282e50c95546486.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 581, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 553, + 617 + ], + "type": "text", + "content": "For the sake of simplicity, we approximate " + }, + { + "bbox": [ + 313, + 581, + 553, + 617 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{ref}}^{\\mathbf{c}}(\\pmb{x}_0,\\pmb{x}_t)" + }, + { + "bbox": [ + 313, + 581, + 553, + 617 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 581, + 553, + 617 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{ref}}^{\\mathbf{c}}(\\pmb{x}_0|\\pmb{x}_t)p_\\theta^{\\mathbf{c}}(\\pmb{x}_t)" + }, + { + "bbox": [ + 313, + 581, + 553, + 617 + ], + "type": "text", + "content": ". Using Jensen's inequality and some algebra, we approximate Eq. (11) as:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 324, + 628, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 628, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 324, + 628, + 553, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\theta) = - \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}\\right) \\sim \\mathcal {D}, t \\sim \\mathcal {U} (0, T),} \\quad \\log \\sigma (- \\beta ( \\\\ \\boldsymbol {x} _ {t} ^ {w} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {t} ^ {w} \\mid \\boldsymbol {x} _ {0} ^ {w}\\right), \\boldsymbol {x} _ {t} ^ {l} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {t} ^ {l} \\mid \\boldsymbol {x} _ {0} ^ {l}\\right) \\\\ + \\mathbb {D} _ {\\mathrm {K L}} \\left(q ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {w} \\mid \\boldsymbol {x} _ {t} ^ {w}, \\boldsymbol {x} _ {0} ^ {w} (t)\\right) \\| p _ {\\theta} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {w} \\mid \\boldsymbol {x} _ {t} ^ {w}\\right)\\right) \\\\ - \\mathbb {D} _ {\\mathrm {K L}} \\left(q ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {w} \\mid \\boldsymbol {x} _ {t} ^ {w}, \\boldsymbol {x} _ {0} ^ {w} (t)\\right) \\| p _ {\\text {r e f}} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {w} \\mid \\boldsymbol {x} _ {t} ^ {w}\\right)\\right) \\tag {17} \\\\ - \\mathbb {D} _ {\\mathrm {K L}} \\left(q ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {l} \\mid \\boldsymbol {x} _ {t} ^ {l}, \\boldsymbol {x} _ {0} ^ {l} (t)\\right) \\| p _ {\\theta} ^ {\\boldsymbol {c}} \\left(\\boldsymbol {x} _ {0} ^ {l} \\mid \\boldsymbol {x} _ {t} ^ {l}\\right)\\right) \\\\ + \\mathbb {D} _ {\\mathrm {K L}} (q ^ {\\boldsymbol {c}} (\\pmb {x} _ {0} ^ {l} | \\pmb {x} _ {t} ^ {l}, \\pmb {x} _ {0} ^ {l} (t)) \\| p _ {\\mathrm {r e f}} ^ {\\boldsymbol {c}} (\\pmb {x} _ {0} ^ {l} | \\pmb {x} _ {t} ^ {l}))) \\\\ \\end{array}", + "image_path": "a7b5589599a6dbdfb8c4b0acab74b4c5bbc6d090e2718cb744cd397a5facb600.jpg" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28632" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 69, + 553, + 395 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 553, + 395 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 553, + 395 + ], + "type": "image", + "image_path": "a7bf779cbd38521276e37c76d6876ffbb3abc28e8526cc30d4330880a42756f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 401, + 555, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 401, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 555, + 425 + ], + "type": "text", + "content": "Figure 3. Top, in human evaluations, InPO-SDXL shows a marked improvement over both DPO-SDXL and Base-SDXL. Bottom, qualitative comparisons among baselines. InPO-SDXL achieves superior prompt alignment and produces images of higher quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 445, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 445, + 295, + 470 + ], + "type": "inline_equation", + "content": "q^c (\\cdot |\\cdot)" + }, + { + "bbox": [ + 55, + 445, + 295, + 470 + ], + "type": "text", + "content": " is a Gaussian transition kernel. Using Eq.(16), the above loss can be simplified as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 479, + 294, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 479, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 64, + 479, + 294, + 548 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\theta) = - \\mathbb {E} _ {(\\boldsymbol {x} _ {0} ^ {w}, \\boldsymbol {x} _ {0} ^ {l}, \\boldsymbol {c}) \\sim \\mathcal {D}, t \\sim \\mathcal {U} (0, T),} \\quad \\log \\sigma (- \\beta w (t) ( \\\\ \\boldsymbol {x} _ {t} ^ {w} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {t} ^ {w} | \\boldsymbol {x} _ {0} ^ {w}), \\boldsymbol {x} _ {t} ^ {l} \\sim p _ {\\theta} ^ {\\boldsymbol {c}} (\\boldsymbol {x} _ {t} ^ {l} | \\boldsymbol {x} _ {0} ^ {l}) \\\\ \\left\\| \\tau_ {t} ^ {w} - \\epsilon_ {\\theta} ^ {t} \\left(\\boldsymbol {x} _ {t} ^ {w}, \\boldsymbol {c}\\right) \\right\\| _ {2} ^ {2} - \\left\\| \\tau_ {t} ^ {w} - \\epsilon_ {\\mathrm {r e f}} ^ {t} \\left(\\boldsymbol {x} _ {t} ^ {w}, \\boldsymbol {c}\\right) \\right\\| _ {2} ^ {2} \\\\ - \\left\\| \\tau_ {t} ^ {l} - \\epsilon_ {\\theta} ^ {t} \\left(\\boldsymbol {x} _ {t} ^ {l}, \\boldsymbol {c}\\right) \\right\\| _ {2} ^ {2} + \\left\\| \\tau_ {t} ^ {l} - \\epsilon_ {\\text {r e f}} ^ {t} \\left(\\boldsymbol {x} _ {t} ^ {l}, \\boldsymbol {c}\\right) \\right\\| _ {2} ^ {2}) \\tag {18} \\\\ \\end{array}", + "image_path": "4c75173b2337451c66cdca15d6f08ea70a4c9f9d924f5a7a2a091038ff1a90bd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 561, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 295, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 561, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\tau_{t}^{*} = (\\pmb{x}_{0}^{*}(t) - \\pmb{x}_{0}^{*}) / \\sigma_{t} + \\delta_{t}(\\pmb{x}_{0}^{*}(t))" + }, + { + "bbox": [ + 55, + 561, + 295, + 586 + ], + "type": "text", + "content": " (refer to Fig. 2) and " + }, + { + "bbox": [ + 55, + 561, + 295, + 586 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 55, + 561, + 295, + 586 + ], + "type": "text", + "content": " is a weight function defined consistent with [63]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "Link to Diffusion-DPO. Diffusion-DPO defines a reward function on the denoising trajectory, which expands the decision space because it requires combining the reward functions across all denoise steps. This problem is well known to be closely related to delayed feedback/sparse rewards in reinforcement learning, where effective feedback is only available after generating the entire trajectory. Within our framework, Diffusion-DPO utilizes a rough approximation of " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\delta_t" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " which uses i.i.d random noise " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\delta_t^{\\mathrm{DPO}}(\\pmb{x}_0(t))\\sim" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 445, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 445, + 555, + 469 + ], + "type": "text", + "content": ", matching diffusion incorporated in the forward process of noise addition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 480, + 395, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 395, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 395, + 495 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 500, + 364, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 364, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 364, + 513 + ], + "type": "text", + "content": "5.1. Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 517, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 626 + ], + "type": "text", + "content": "Datasets and Models. We conduct all experiments on the Pick-a-Pic v2 [31] dataset, which is collected from the Pick-a-Pic web application (refer to [31] for details). Consistent with the setting in [63], approximately " + }, + { + "bbox": [ + 313, + 517, + 555, + 626 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 313, + 517, + 555, + 626 + ], + "type": "text", + "content": " of the data is excluded, where preferences cannot be determined. As a result, we obtain 851,293 data pairs and 58,960 unique prompts. Our method is applied to fine-tune the base models Stable Diffusion 1.5 (SD1.5) [52] and Stable Diffusion XL-base-1.0 (SDXL) [43]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Basic Hyperparameters. To ensure a fair comparison of training efficiency, we use hyperparameters largely consistent with those in [63]. We employ AdamW [35] as the optimizer for SD1.5 and Adafactor [57] for SDXL. All experiments are conducted on eight NVIDIA H800 GPUs, with each GPU handling a batch size of one pair and gradient" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28633" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 196 + ], + "type": "table", + "html": "
ModelBaselinesWin-rate (HPDv2)Win-rate (Parti-Prompts)
AestheticPickScoreHPSv2CLIPAestheticPickScoreHPSv2CLIP
InPO-SDXLvs. Base-SDXL56.3779.2585.3853.3761.7072.8978.3150.55
vs. SFT-SDXL66.6688.3189.9159.0069.7088.5488.9156.99
vs. DPO-SDXL55.8759.2865.8146.9157.1756.9260.1142.28
InPO-SD1.5vs. Base-SD1.580.1985.8490.2266.4474.6373.1681.8661.89
vs. SFT-SD1.556.6366.8458.9157.8752.2762.5058.0357.54
vs. DPO-SD1.570.5074.9484.0661.0968.0163.5475.3157.54
vs. KTO-SD1.559.8167.4160.9460.0359.0162.2558.7657.60
", + "image_path": "ac1d8b979b7185068161181af2fe64712756b3c394a7ee8534caba9121cbed45.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 555, + 239 + ], + "type": "text", + "content": "Table 1. Win-rate comparison between DDIM-InPO and baselines. We use prompts from HPDv2 and Parti-Prompts, evaluated across different evaluators. SDXL and SD1.5 are used as base models. For simplicity, we denote models using a 'Method-Base' format, such as InPO-SDXL to represent the SDXL model fine-tuned with the DDIM-InPO method. Win-rates exceeding " + }, + { + "bbox": [ + 55, + 204, + 555, + 239 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 204, + 555, + 239 + ], + "type": "text", + "content": " are highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 258, + 297, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 297, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 297, + 320 + ], + "type": "text", + "content": "accumulation over 128 steps, resulting in an effective batch size of 1024 pairs (half the batch size compared to [63]). The learning rate is set to " + }, + { + "bbox": [ + 55, + 258, + 297, + 320 + ], + "type": "inline_equation", + "content": "\\frac{2000}{\\beta} 2.048^{-8}" + }, + { + "bbox": [ + 55, + 258, + 297, + 320 + ], + "type": "text", + "content": ", with a linear warm-up phase. For SD1.5 training, " + }, + { + "bbox": [ + 55, + 258, + 297, + 320 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 258, + 297, + 320 + ], + "type": "text", + "content": " is set to 2000, while for SDXL training, it is set to 5000." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 338, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 296, + 601 + ], + "type": "text", + "content": "Evaluation. We compare DDIM-InPO with existing baselines using user studies and automatic preference metrics. We use the following four evaluators to evaluate image quality: the LAION aesthetic classifier [55] to assess image aesthetic scores, CLIP [47] to measure text-image alignment, and PickScore [31] and HPSv2 [66] to predict human preferences for given image-text pairs. In test stage, we generated images using both the baseline and our model on the Parti-Prompts (1632 prompts) and HPDv2 (3200 prompts) test sets. The median and win-rate (how frequently does the evaluator prefer the generations from DDIM-InPO over those from baselines) of the aforementioned four evaluators are used as automatic preference metrics for comparison. Additionally, we conduct a user study (16 participants) to compare DDIM-InPO with the baselines. We design three questions as [63]: (1) Which image is your overall preferred choice? (2) Which image is more visually attractive? (3) Which image better matches the text description? We also compare the traininggpu hours (on H800) with baselines. In this work, we adopt the following baselines: base model (SD1.5 and SDXL), supervised fine-tuning (SFT), Diffusion-DPO, and Diffusion-KTO (for SD1.5)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 611, + 155, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 155, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 155, + 624 + ], + "type": "text", + "content": "5.2. Primary Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Quantitative Results. Firstly, we use the quantitative evaluation method for model preferences outlined in [32, 63]. Tab. 1 presents the win-rates of the InPO-aligned diffusion models in comparison to the baselines. Overall, following fine-tuning with DDIM-InPO, both SD1.5 and SDXL achieve superior performance compared to the baselines across nearly all evaluators and test datasets, thereby vali" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 322, + 271, + 533, + 406 + ], + "blocks": [ + { + "bbox": [ + 322, + 271, + 533, + 406 + ], + "lines": [ + { + "bbox": [ + 322, + 271, + 533, + 406 + ], + "spans": [ + { + "bbox": [ + 322, + 271, + 533, + 406 + ], + "type": "image", + "image_path": "a15ed25ff296813fa49bcd76607c730c5c108ed63ffd44b055f7e679d985b203.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 415, + 555, + 494 + ], + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 494 + ], + "type": "text", + "content": "Figure 4. Comparison of the trade-off between the quality of generated images and training efficiency following human preference optimization of SD1.5 on the HPDv2 test set. Sizes of the circles represent the volume of training data used. Our DDIM-InPO achieves superior performance, with a training speed that is 18.4 and 3.6 times faster than Diffusion-KTO [32] and Diffusion-DPO [63], respectively, while producing images of higher quality." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 518, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 628 + ], + "type": "text", + "content": "dating the effectiveness of our method. Specifically, HPSv2 evaluation shows that, InPO-SDXL obtains a win-rate of " + }, + { + "bbox": [ + 313, + 518, + 555, + 628 + ], + "type": "inline_equation", + "content": "65.81\\%" + }, + { + "bbox": [ + 313, + 518, + 555, + 628 + ], + "type": "text", + "content": " against DPO-SDXL, while InPO-SD1.5 achieves a win-rate of " + }, + { + "bbox": [ + 313, + 518, + 555, + 628 + ], + "type": "inline_equation", + "content": "60.94\\%" + }, + { + "bbox": [ + 313, + 518, + 555, + 628 + ], + "type": "text", + "content": " compared to KTO-SD1.5 on the HPDv2 test set. Moreover, it significantly improves the alignment with human preferences for the base model. According to HPSv2, it achieves win-rate of up to " + }, + { + "bbox": [ + 313, + 518, + 555, + 628 + ], + "type": "inline_equation", + "content": "85.38\\%" + }, + { + "bbox": [ + 313, + 518, + 555, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 518, + 555, + 628 + ], + "type": "inline_equation", + "content": "90.22\\%" + }, + { + "bbox": [ + 313, + 518, + 555, + 628 + ], + "type": "text", + "content": " when compared to the SD1.5 and SDXL base models, respectively, demonstrating the efficacy of DDIM-InPO." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "It should be noted that our DDIM-InPO significantly advances training efficiency, with a training speed that is 18.4 and 3.6 times faster than Diffusion-KTO and Diffusion-DPO, respectively, as shown in Fig. 4, while producing higher-quality images. Additionally, Fig. 3 (top) presents the results of our user study, where we find that human evaluators prefer the model aligned with DDIM-InPO. The" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28634" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 553, + 338 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 553, + 338 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 553, + 338 + ], + "type": "image", + "image_path": "140ec73286f04a450c31a1805c74c853302bd21870aeb76d4d4b1ff8fe998ff4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 345, + 555, + 411 + ], + "lines": [ + { + "bbox": [ + 54, + 345, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 345, + 555, + 411 + ], + "type": "text", + "content": "Figure 5. Advantages of DDIM-InPO: Compared with Refinement technique and Diffusion-DPO, we find that images generated by DDIM-InPO exhibit enhanced light and spatial structure, better realism and detail capture, greater imaginative design, stable color consistency, optimized multi-instance layout and text integration in visuals. These are some hidden advantages aligned with human preferences. Prompts from left to right: (1) A small bird sitting in a metal wheel. (2) Trees seen through a car window on a rainy day. (3) Description, An artistic rendering of a cosmic portal with a beach at dusk on the other side. (4) A person in full samurai armor at the beach. (5) A man standing in front of a bunch of doughnuts. (6) A towel with the word 'cat' printed on it, simple and clear text." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 432, + 298, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 432, + 298, + 482 + ], + "spans": [ + { + "bbox": [ + 54, + 432, + 298, + 482 + ], + "type": "text", + "content": "evaluation results clearly show that our method provides a more efficient way for fine-tuning diffusion models with human preferences, demonstrating the effectiveness of our proposed paradigm for learning specific latent variables." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 496, + 298, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 496, + 298, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 496, + 298, + 640 + ], + "type": "text", + "content": "Qualitative results. We display qualitative comparison results of InPO-SDXL against several baselines (Base-SDXL, DPO-SDXL) in Fig. 3 (bottom). The images generated by InPO-SDXL demonstrate enhanced performance in both textual alignment and visual attractiveness. In addition, Fig. 5 illustrates the advantages of our model, which serves as a method for directly learning human preferences for diffusion models. The generated images showcase improved light and spatial structures, detailed capture, greater creativity, consistent color, organized multi-instance layouts, and text-visual integration. These advantages align closely with the preferences of users." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 648, + 209, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 209, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 209, + 662 + ], + "type": "text", + "content": "5.3. AI Preference Optimization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 665, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 297, + 715 + ], + "type": "text", + "content": "The high cost of manual annotation has driven researchers to seek alternative approaches. Bai et al. [3] introduce RL from AI feedback as one such solution. It follows naturally that PickScore, and HPSv2 can be employed as AI evalua" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 430, + 556, + 547 + ], + "blocks": [ + { + "bbox": [ + 315, + 430, + 556, + 547 + ], + "lines": [ + { + "bbox": [ + 315, + 430, + 556, + 547 + ], + "spans": [ + { + "bbox": [ + 315, + 430, + 556, + 547 + ], + "type": "image", + "image_path": "049d4e9ac21d043c961e5bd66e05ff8de445b28ed0b2d486efd000a2e3702d8b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 555, + 556, + 578 + ], + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 578 + ], + "type": "text", + "content": "Figure 6. Median of PickScore and HPSv2 comparisons for all baselines and test sets on SD1.5. (Detailed analysis in Sec. 5.3)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 601, + 556, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 556, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 556, + 648 + ], + "type": "text", + "content": "tor to simulate human preferences. We relabel the Pick-a-Pic v2 dataset using AI evaluators and fine-tuned the SD1.5 model. As shown in Fig. 6, we observe a further improvement in these two metrics with AI feedback." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 659, + 384, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 384, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 384, + 671 + ], + "type": "text", + "content": "5.4. Ablations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Tab. 2 presents the results of our ablation experiments on SD1.5 and the median values of the preference evaluators on the HPDv2 test set. For a clearer comparison of our" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28635" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 305, + 251 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 305, + 251 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 305, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 305, + 251 + ], + "type": "table", + "html": "
Median (HPDv2)GPU hours
AestheticPickScoreHPSv2CLIP
Base-SD1.55.349120.71926.64734.276-
DPO-SD1.55.692221.27427.42835.902204.8
Base initialized5.650421.76328.23736.23857.6
β = 30005.773021.80828.50536.78857.6
β = 40005.763721.79628.44836.68557.6
β = 50005.742821.77128.47336.66957.6
Inv-steps n=305.822521.95728.56837.019136
Inv-steps n=55.771521.84528.43236.76943.2
Inv-steps n=35.748921.83328.50836.54236
CFG win=15.738121.83528.51636.45162.4
CFG win=55.732021.71228.46236.426104
CFG win=7.55.749521.74228.47336.469104
InPO-SD1.55.773421.89428.52336.87657.6
", + "image_path": "68c4aa89f9c464422e7690ece567c1cf5a738044927aea5410b9c02891f0a932.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 296, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 304 + ], + "type": "text", + "content": "Table 2. Ablation studies of our DDIM-InPO on fine-tuning SD1.5. In each column of preference evaluators, the highest value is highlighted in bold, and the second highest is underlined. For detailed analysis, refer to Sec. 5.4" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 324, + 296, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 349 + ], + "type": "text", + "content": "method's effectiveness, we present the results for the Base-SD1.5 and DPO-SD1.5 baselines in rows 3 and 4 of Tab. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 363, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 472 + ], + "type": "text", + "content": "Proposed improvements. The fifth row of Tab. 2 presents the results of initializing " + }, + { + "bbox": [ + 55, + 363, + 296, + 472 + ], + "type": "inline_equation", + "content": "p_{\\theta}^{c}" + }, + { + "bbox": [ + 55, + 363, + 296, + 472 + ], + "type": "text", + "content": " in Eq. (11) with the SD1.5 base model. After applying our core contribution, DDIM inversion, we observe a significant improvement in quality compared to the median values of the evaluators for the baselines, indicating that enhancement primarily stems from our method. As instructed by [48], practically an available SFT model can be used to initialize the reference model " + }, + { + "bbox": [ + 55, + 363, + 296, + 472 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{ref}}^{c}" + }, + { + "bbox": [ + 55, + 363, + 296, + 472 + ], + "type": "text", + "content": ". This strategy aids in reducing the issue of distribution shift." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "text", + "content": "Parameters Choices. We set " + }, + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "text", + "content": " to 2000 for SD1.5. Rows 6 to 8 of Tab. 2 indicate that if " + }, + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "text", + "content": " is set too high, the KL divergence penalty term severely limits the model's flexibility in adjustments. It also maintains a consistent learning rate with other baselines, ensuring a fair comparison in terms of efficiency. Rows 9 to 11 of Tab. 2 show the steps required for DDIM inversion. We observe that setting inversion steps to 30 yields the best performance, further confirming the effectiveness of our method. However, this also increases the training time. Thus we select inversion steps " + }, + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "inline_equation", + "content": "n = 10" + }, + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "text", + "content": " as it strikes a good balance between generation quality and training speed. The work of [38, 39] indicates that DDIM inversion can lead to considerable numerical errors when Classifier Free Guidance (CFG) [26] " + }, + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{inv}} > 0" + }, + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "text", + "content": ". Rows 12 to 14 of Tab. 2 demonstrate that a CFG " + }, + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{inv}} > 0" + }, + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "text", + "content": " during the inversion process adversely impacts the fine-tuning of human preferences. Consequently, we utilize unconditional DDIM inversion, setting CFG " + }, + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{inv}} = 0" + }, + { + "bbox": [ + 55, + 486, + 296, + 715 + ], + "type": "text", + "content": ". We optimize parameters to balance generation quality and training efficiency." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 554, + 260 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 260 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 260 + ], + "type": "image", + "image_path": "841e6584132d563e24d019930ca2435c9ad0be7b699f7ceb5052fe6ec916c943.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 268, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 324 + ], + "type": "text", + "content": "Figure 7. Qualitative evaluation of InPO-SDXL in comparison with Base-SDXL and DPO-SDXL on conditional generation tasks (From top to bottom: depth map, canny edge, and inpainting). Prompts: (1) A dolphin in an astronaut suit on Saturn. (2) The Great Pyramid of Giza. (3) A cat riding on a horse." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 346, + 444, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 444, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 444, + 360 + ], + "type": "text", + "content": "5.5. Conditional generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 365, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 498 + ], + "type": "text", + "content": "After fine-tuning the model using the DDIM-InPO method, it can be directly applied to conditional generation tasks without the need for further training. We add conditional control to InPO-SDXL using ControlNet [74] which originally trained with Base-SDXL. Specifically, we use depth map and canny edge as conditions for text-to-image generation tasks. Furthermore, we evaluate our model on image inpainting [37] tasks. As shown in Fig. 7, the advantages of our model are seamlessly transferred to conditional generation tasks, demonstrating superior generation quality compared with Base-SDXL and DPO-SDXL." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "We introduce DDIM-InPO, an efficient method for aligning diffusion models with human preferences, focusing on directly fine-tuning specific latent variables that are strongly correlated with the target image. By fine-tuning the T2I diffusion models using the Pick-a-Pic v2 dataset, our experiments show that DDIM-InPO significantly enhances training efficiency, achieving state-of-the-art performance in just 400 training steps and surpassing all baseline models across nearly all evaluators. DDIM-InPO is based on offline fine-tuning and is sensitive to dataset quality; the presence of harmful, violent, and pornographic content can affect the results, leading to the generation of inappropriate outputs, such as overly feminine images. In addition, we consider the online training paradigm to be a promising and reliable approach for the future." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28636" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "This work was supported by the National Major Science and Technology Projects (the grant number 2022ZD0117000) and the National Natural Science Foundation of China (grant number 62202426)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 169, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 169, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 296, + 224 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 224, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 296, + 280 + ], + "type": "text", + "content": "[2] Mohammad Gheshlaghi Azar, Mark Rowland, Bilal Piot, Daniel Guo, Daniele Calandriello, Michal Valko, and Rémi Munos. A general theoretical paradigm to understand learning from human preferences. ArXiv, abs/2310.12036, 2023. 2, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 280, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 296, + 335 + ], + "type": "text", + "content": "[3] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 335, + 296, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 335, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 61, + 335, + 296, + 401 + ], + "type": "text", + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. *ArXiv*, abs/2211.01324, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 296, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 458 + ], + "type": "text", + "content": "[5] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 296, + 491 + ], + "type": "text", + "content": "[6] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. ArXiv, abs/2305.13301, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 296, + 536 + ], + "type": "text", + "content": "[7] A. Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, and Dominik Lorenz. Stable video diffusion: Scaling latent video diffusion models to large datasets. ArXiv, abs/2311.15127, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 536, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 602 + ], + "type": "text", + "content": "[8] A. Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22563-22575, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 602, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 296, + 658 + ], + "type": "text", + "content": "[9] Haoxin Chen, Menghan Xia, Yin-Yin He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao-Liang Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation. ArXiv, abs/2310.19512, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[10] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ArXiv, abs/2310.00426, 2023. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[11] Paul Francis Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. ArXiv, abs/1706.03741, 2017. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 150 + ], + "type": "text", + "content": "[12] Kevin Clark, Paul Vicol, Kevin Swersky, and David J. Fleet. Directly fine-tuning diffusion models on differentiable rewards. ArXiv, abs/2309.17400, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 251 + ], + "type": "text", + "content": "[13] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam S. Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Kumar Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yiqian Wen, Yi-Zhe Song, Roshan Sumbaly, Vignesh Ramanathan, Zijian He, Péter Vajda, and Devi Parikh. Emu: Enhancing image generation models using photogenic needles in a haystack. ArXiv, abs/2309.15807, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 252, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 554, + 274 + ], + "type": "text", + "content": "[14] Quan Dao, Hao Phung, Binh Duc Nguyen, and A. Tran. Flow matching in latent space. ArXiv, abs/2307.08698, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 554, + 297 + ], + "type": "text", + "content": "[15] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. ArXiv, abs/2105.05233, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 342 + ], + "type": "text", + "content": "[16] Hanze Dong, Wei Xiong, Deepanshu Goyal, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and T. Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. ArXiv, abs/2304.06767, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 554, + 398 + ], + "type": "text", + "content": "[17] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7312-7322, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 475 + ], + "type": "text", + "content": "[18] Patrick Esser, Sumith Kulal, A. Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. ArXiv, abs/2403.03206, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 476, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 554, + 510 + ], + "type": "text", + "content": "[19] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. ArXiv, abs/2402.01306, 2024. 2, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 510, + 554, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 554, + 566 + ], + "type": "text", + "content": "[20] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, P. Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. ArXiv, abs/2305.16381, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 567, + 554, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 554, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 554, + 610 + ], + "type": "text", + "content": "[21] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising. ArXiv, abs/2403.14602, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 611, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 611, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 611, + 554, + 645 + ], + "type": "text", + "content": "[22] Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. ArXiv, abs/2310.11513, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 646, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 554, + 690 + ], + "type": "text", + "content": "[23] Han Guo, Bowen Tan, Zhengzhong Liu, Eric P. Xing, and Zhiting Hu. Efficient (soft) q-learning for text generation with limited good data. In Conference on Empirical Methods in Natural Language Processing, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 690, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 554, + 713 + ], + "type": "text", + "content": "[24] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28637" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Photorealistic video generation with diffusion models. In European Conference on Computer Vision, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "text", + "content": "[25] Amir Hertz, Ron Mokady, Jay M. Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. ArXiv, abs/2208.01626, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 294, + 163 + ], + "type": "text", + "content": "[26] Jonathan Ho. Classifier-free diffusion guidance. ArXiv, abs/2207.12598, 2022. 8, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 167, + 294, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 294, + 220 + ], + "type": "text", + "content": "[27] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. ArXiv, abs/2210.02303, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 223, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 294, + 255 + ], + "type": "text", + "content": "[28] Jiwoo Hong, Noah Lee, and James Thorne. Orpo: Monolithic preference optimization without reference model. ArXiv, abs/2403.07691, 2024. 2, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 258, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 294, + 301 + ], + "type": "text", + "content": "[29] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. ArXiv, abs/2402.17525, 2024. 2, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 304, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 335 + ], + "type": "text", + "content": "[30] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. ArXiv, abs/2206.00364, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "text", + "content": "[31] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 36652-36663, 2023. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 395, + 294, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 395, + 294, + 438 + ], + "spans": [ + { + "bbox": [ + 56, + 395, + 294, + 438 + ], + "type": "text", + "content": "[32] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. arXiv preprint arXiv:2404.04465, 2024. 1, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 441, + 294, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 294, + 484 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 294, + 484 + ], + "type": "text", + "content": "[33] Zhanhao Liang, Yuhui Yuan, Shuyang Gu, Bohan Chen, Tiankai Hang, Ji Li, and Liang Zheng. Step-aware preference optimization: Aligning preference with denoising performance at each step. ArXiv, abs/2406.04314, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "type": "text", + "content": "[34] Tianqi Liu, Yao Zhao, Rishabh Joshi, Misha Khalman, Mohammad Saleh, Peter J. Liu, and Jialu Liu. Statistical rejection sampling improves preference optimization. ArXiv, abs/2309.06657, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 533, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 553 + ], + "type": "text", + "content": "[35] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. ArXiv, abs/1711.05101, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 556, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 609 + ], + "type": "text", + "content": "[36] Artem Lukoianov, Haitz Sáez de Ocariz Borde, Kristjan Greenewald, Vitor Campagnolo Guizilini, Timur Bagautdinov, Vincent Sitzmann, and Justin Solomon. Score distillation via reparametrized ddim. arXiv preprint arXiv:2405.15891, 2024. 4, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "text", + "content": "[37] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 8, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[38] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. arXiv preprint arXiv:2305.16807, 2023. 8, 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "text", + "content": "[39] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3, 4, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 217 + ], + "type": "text", + "content": "[40] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. ArXiv, abs/2203.02155, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "text", + "content": "[41] William S. Peebles and Saining Xie. Scalable diffusion models with transformers. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4172-4182, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 252, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 295 + ], + "type": "text", + "content": "[42] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "text", + "content": "[43] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 385 + ], + "type": "text", + "content": "[44] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. ArXiv, abs/2209.14988, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[45] Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. ArXiv, abs/2310.03739, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "text", + "content": "[46] Mihir Prabhudesai, Russell Mendonca, Zheyang Qin, Katerina Fragkiadaki, and Deepak Pathak. Video diffusion alignment via reward gradients. ArXiv, abs/2407.08737, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 467, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 554, + 533 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 534, + 554, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 554, + 588 + ], + "type": "text", + "content": "[48] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 1, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 633 + ], + "type": "text", + "content": "[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 667 + ], + "type": "text", + "content": "[50] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[51] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International Conference on Machine Learning, 2016. 2, 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28638" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 295, + 205 + ], + "type": "text", + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 205, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 295, + 239 + ], + "type": "text", + "content": "[54] Axel Sauer, Dominik Lorenz, A. Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 295, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 325 + ], + "type": "text", + "content": "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv, abs/2210.08402, 2022.6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 327, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 358 + ], + "type": "text", + "content": "[56] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. ArXiv, abs/1707.06347, 2017. 2, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 360, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 295, + 392 + ], + "type": "text", + "content": "[57] Noam M. Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. ArXiv, abs/1804.04235, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 448 + ], + "type": "text", + "content": "[58] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8871-8879, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "text", + "content": "[59] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 503, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 295, + 546 + ], + "type": "text", + "content": "[60] Feifan Song, Yu Bowen, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. Preference ranking optimization for human alignment. ArXiv, abs/2306.17492, 2023. 2, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 295, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 580 + ], + "type": "text", + "content": "[61] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 581, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 295, + 624 + ], + "type": "text", + "content": "[62] Yang Song, Jascha Narain Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. ArXiv, abs/2011.13456, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 690 + ], + "type": "text", + "content": "[63] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[64] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 666 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. ArXiv, abs/2311.03079, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[65] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. ArXiv, abs/2306.09341, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "text", + "content": "[66] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2096-2105, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "text", + "content": "[67] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. ArXiv, abs/2304.05977, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 317 + ], + "type": "text", + "content": "[68] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Qimai Li, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8941-8951, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "text", + "content": "[69] Shentao Yang, Tianqi Chen, and Mingyuan Zhou. A dense reward view on aligning text-to-image diffusion with preference. ArXiv, abs/2402.08265, 2024. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 353, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 430 + ], + "type": "text", + "content": "[70] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Trans. Mach. Learn. Res., 2022, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 431, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 496 + ], + "type": "text", + "content": "[71] Hangjie Yuan, Shiwei Zhang, Xiang Wang, Yujie Wei, Tao Feng, Yining Pan, Yingya Zhang, Ziwei Liu, Samuel Albanie, and Dong Ni. Instructvideo: Instructing video diffusion models with human feedback. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6463-6474, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "text", + "content": "[72] Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Feiran Huang. Rrhf: Rank responses to align language models with human feedback without tears. ArXiv, abs/2304.05302, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 542, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 542, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 542, + 553, + 586 + ], + "type": "text", + "content": "[73] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1204-1213, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 587, + 553, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 587, + 553, + 631 + ], + "spans": [ + { + "bbox": [ + 316, + 587, + 553, + 631 + ], + "type": "text", + "content": "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3813-3824, 2023. 8, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 632, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 632, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 632, + 553, + 666 + ], + "type": "text", + "content": "[75] Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. arXiv preprint arXiv:2401.12244, 4, 2024. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28639" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/c5e31ebc-2f5f-4333-a4a5-98739dcdbce4_content_list.json b/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/c5e31ebc-2f5f-4333-a4a5-98739dcdbce4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5866e1a3e71c32bc6c80d0f2dcbe091f81b95913 --- /dev/null +++ b/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/c5e31ebc-2f5f-4333-a4a5-98739dcdbce4_content_list.json @@ -0,0 +1,1436 @@ +[ + { + "type": "text", + "text": "IncEventGS: Pose-Free Gaussian Splating from a Single Event Camera", + "text_level": 1, + "bbox": [ + 142, + 130, + 869, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jian Huang $^{1,2}$ Chengrui Dong $^{1,2}$ Xuanhua Chen $^{2,3}$ Peidong Liu $^{2*}$", + "bbox": [ + 196, + 191, + 810, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Zhejiang University $^{2}$ Westlake University $^{3}$ Northeastern University $\\{$ huangjian39, dongchengrui, chenxuanhua, liupeidong\\}@westlake.edu.cn https://github.com/WU-CVGL/IncEventGS", + "bbox": [ + 199, + 222, + 810, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 251, + 313, + 328, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Implicit neural representation and explicit 3D Gaussian Splating (3D-GS) for novel view synthesis have achieved remarkable progress with frame-based camera (e.g. RGB and RGB-D cameras) recently. Compared to frame-based camera, a novel type of bio-inspired visual sensor, i.e. event camera, has demonstrated advantages in high temporal resolution, high dynamic range, low power consumption, and low latency, which make it favored for many robotic applications. In this work, we present IncEventGS, an incremental 3D Gaussian Splating reconstruction algorithm with a single event camera, without the assumption of known camera poses. To recover the 3D scene representation incrementally, we exploit the tracking and mapping paradigm of conventional SLAM pipelines for IncEventGS. Given the incoming event stream, the tracker first estimates an initial camera motion based on prior reconstructed 3D-GS scene representation. The mapper then jointly refines both the 3D scene representation and camera motion based on the previously estimated motion trajectory from the tracker. The experimental results demonstrate that IncEventGS delivers superior performance compared to prior NeRF-based methods and other related baselines, even if we do not have the ground-truth camera poses. Furthermore, our method can also deliver better performance compared to state-of-the-art event visual odometry methods in terms of camera motion estimation.", + "bbox": [ + 89, + 344, + 493, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 750, + 220, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing accurate 3D scene representations from 2D images has been a long-standing challenge in computer vision and robotics, driving substantial efforts over the past few decades. Among those pioneering works, Neural Radiance Fields (NeRF) [19] and 3D Gaussian Splatting (3D-GS) [11], stand out for their utilization of differentiable rendering techniques, and have garnered significant attention due to their capability to recover high-quality 3D scene representation", + "bbox": [ + 89, + 775, + 490, + 896 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from 2D images. Commonly used sensors for 3D scene reconstruction are usually frame-based cameras, such as the RGB and RGB-D cameras. They usually capture full-brightness intensity images within a short exposure time at a regular frequency. Due to the characteristic of this data-capturing process, they often suffer from motion blur or fail to capture accurate and informative intensity information under fast motion and low-light conditions, which would further affect the performance of downstream applications.", + "bbox": [ + 517, + 314, + 921, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The event camera, a bio-inspired sensor, has gained significant attention in recent years for its potential to address the limitations of frame-based cameras under challenging conditions. Unlike conventional cameras, event cameras record brightness changes asynchronously at each pixel, emitting events when a predefined threshold is surpassed. This unique operation offers several advantages over conventional cameras, in terms of high temporal resolution, high dynamic range, low latency, and power consumption. Although event cameras have attractive characteristics for challenging environments, they cannot be directly integrated into existing frame-based 3D reconstruction algorithms that rely on processing dense 2D brightness intensity images.", + "bbox": [ + 517, + 450, + 921, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several pioneering works have been proposed to exploit event stream [3, 12, 23] to recover the motion trajectory and scene representation. While existing methods deliver impressive performance, they usually exploit 2.5D semi-dense depth maps to represent the 3D scene, and bundle adjustment (BA) is hardly performed, due to the asynchronous and sparse characteristics of event data stream. Klenk et al. [15] recently proposed to convert event stream into event voxel grids, and then adapt a previous frame-based deep visual odometry pipeline [30] for accurate camera motion estimation. As NeRF exhibited impressive scene representation capability recently, several works [8, 14, 17, 25] explore to recover the underlying dense 3D scene NeRF representation from event stream, by assuming ground-truth poses are available.", + "bbox": [ + 517, + 647, + 921, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast to those works, we propose IncEventGS, an incremental dense 3D scene reconstruction algorithm from a single event camera, by exploiting Gaussian Splatting as the underlying scene representation. Different from prior event-", + "bbox": [ + 517, + 858, + 921, + 919 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 114, + 905, + 241, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26933", + "bbox": [ + 478, + 953, + 517, + 965 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "based NeRF reconstruction methods, IncEventGS does not require any ground-truth camera poses, which is more challenging and provides more flexibility for real-world robotic application scenarios. To overcome the challenges brought by unknown poses, IncEventGS adopts the tracking and mapping paradigm of conventional SLAM pipelines [20]. In particular, IncEventGS exploits prior explored and reconstructed 3D scenes for camera motion estimation of incoming event stream during the tracking stage. Both the 3D-GS scene representation and camera motions are then jointly optimized (i.e. event-based bundle adjustment) during the mapping stage, for more accurate scene representation and motion estimation. The 3D scene is progressively expanded and densified. The experimental results on both synthetic and real datasets demonstrate that IncEventGS can recover the underlying 3D scene representation and camera motion trajectory accurately. In particular, IncEventGS outperforms prior NeRF-based methods and other related baselines in terms of scene representation recovery, even without ground-truth poses. Furthermore, our method also delivers better camera motion estimation accuracy than the most recent state-of-the-art visual odometry algorithm, in terms of the Absolute Trajectory Error (ATE) metric. The recovered 3D scene representation can be further used to render novel brightness images. Our main contributions can be summarized as follows:", + "bbox": [ + 89, + 90, + 491, + 467 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present an incremental 3D Gaussian Splitting reconstruction algorithm from a single event camera, without requiring the ground-truth camera poses.", + "- We propose a novel initialization strategy tailored to the event data stream, which is vital to the success of the algorithm.", + "- The experimental results on both the synthetic and real datasets demonstrate the superior performance of our method over prior NeRF-based methods and related baselines in terms of novel view synthesis and better performance over state-of-the-art event-based visual odometry algorithm in terms of camera motion estimation." + ], + "bbox": [ + 89, + 468, + 491, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 662, + 240, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We review two main areas of prior works: event-based neural radiance fields and 3D Gaussian Splatting, which are the most related to our work.", + "bbox": [ + 89, + 688, + 488, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event-based Neural Radiance Fields. Prior works [8, 14, 25] propose to exploit event stream to recover the neural radiance fields with known camera motion trajectory. Low et al. [17] further improves the reconstruction algorithm to handle sparse and noisy events under non-uniform motion. The recovered neural radiance fields can then be used to render novel view brightness images. The ground-truth poses are usually computed from corresponding brightness images via COLMAP [26] or provided by the indoor motion-capturing system. Recently, Qu et al. [22] proposed to integrate event measurements into an RGB-D implicit neural SLAM framework and achieved robust performance in mo", + "bbox": [ + 89, + 737, + 491, + 919 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion blur scenarios. Li et al. [16] also proposed exploiting event measurements and a single blurry image to recover the underlying neural 3D scene representation. In contrast to those works, IncEventGS conducts incremental 3D scene reconstruction without requiring any prior ground-truth poses, which is more challenging and provides more flexibility for practical robotic application scenarios. The method further exploits 3D Gaussian Splatting as the underlying scene representation, which demonstrates better image rendering quality and efficiency, compared to the NeRF-based representation.", + "bbox": [ + 517, + 90, + 921, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussian Splatting. 3D Gaussian Splatting [11] proposes a novel explicit 3D representation to further improve both the training and rendering efficiency compared to Neural Radiance Fields. Due to its impressive efficient scene representation capability, several pioneering works have been proposed to exploit 3D-GS for incremental 3D reconstruction. For example, Keetha et al. [10] propose an RGBD-based 3D-GS SLAM, employing an online tracking and mapping system tailored to the underlying Gaussian representation. Yan et al. [36] implement a coarse-to-fine camera tracking approach based on the sparse selection of Gaussians. Matsuki et al. [18] propose to apply 3D Gaussian Splatting to do incremental 3D reconstruction using a single moving monocular or RGB-D camera. Huang et al. [7] exploit ORB-SLAM3 to compute accurate camera poses and feed it into a 3D-GS algorithm for dense mapping. Fu et al. [2] use monocular depth estimation with 3D-GS. Yugay et al. [37] combine DROID-SLAM [29] based camera tracking with active and inactive 3D-GS sub-maps. Wang et al. [33] integrate bundle-adjustment and 3DGS [38] to estimated camera trajectory within the exposure time. Hu et al. [6] propose a novel depth uncertainty model to ensure the selection of valuable Gaussian primitives during optimization. While those methods deliver impressive performance in terms of 3D scene recovery and motion estimation, they usually assume the usage of frame-based images (i.e. either RGB or RGB-D date). On the contrary, we propose to exploit pure event measurements for incremental 3D-GS reconstruction. Several concurrent studies have recently explored the use of 3D Gaussians for event-based reconstruction, including EvGGS [31], Event3DGS [35], and E2GS [1]. EvGGS and Event3DGS rely solely on event data, whereas E2GS incorporates both event data and blurry images. However, the key difference between these methods and ours is that they all rely on ground-truth poses.", + "bbox": [ + 517, + 244, + 921, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 519, + 772, + 612, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The overview of our IncEventGS is shown in Fig. 1. Given only a single event camera, IncEventGS incrementally performs tracking and dense mapping under the framework of 3D Gaussian Splatting, to recover both the camera motion trajectory and 3D scene representation simultaneously. Since event data are asynchronous, they cannot be directly integrated with the 3D-GS representation. We therefore process the event data stream into chunks according to a fixed time window.", + "bbox": [ + 517, + 797, + 921, + 919 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26934", + "bbox": [ + 478, + 953, + 519, + 965 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3d633c7e1ccb14850406539ba2de87c11d49c5b276f733782eece3bafe0e1918.jpg", + "image_caption": [ + "Figure 1. The pipeline of IncEventGS. IncEventGS processes incoming event stream by dividing it into chunks and representing the camera trajectory as a continuous model. It randomly samples two close consecutive timestamps to integrate the corresponding event streams. Two brightness images are rendered from 3D-GS at the corresponding poses, and we minimize the photometric loss between the synthesized and measured brightness change. During initialization, a pre-trained depth estimation model estimates depth from the rendered images to bootstrap the system." + ], + "image_footnote": [], + "bbox": [ + 155, + 89, + 854, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We associate each chunk with a continuous time trajectory parameterization in the $\\mathfrak{se}(3)$ space. Two close consecutive timestamps (i.e., $t_k$ and $t_{k + \\Delta t}$ , where $\\Delta t$ is a small time interval) can be randomly sampled, and the measured brightness change $\\mathbf{E}(x)$ can be computed from the corresponding event stream. Based on the parameterized trajectory, the corresponding camera poses (i.e., $\\mathbf{T}_k$ , $\\mathbf{T}_{k + \\Delta t}$ ) can be determined and the brightness images (i.e., $\\bar{\\mathbf{I}}_k$ , $\\bar{\\mathbf{I}}_{k + \\Delta t}$ ) can be further rendered from the 3D-GS. The synthesized brightness change $\\hat{\\mathbf{E}}(x)$ can be computed for event loss computation.", + "bbox": [ + 88, + 479, + 490, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the tracking stage, we optimize only the camera motion trajectory of the newly accumulated event chunk and exploit the recovered trajectory to initialize the dense bundle adjustment (BA) algorithm for the mapping stage. During the mapping stage, we continuously densify 3D Gaussians for newly explored areas and prune transparent 3D Gaussians. For computational efficiency, we exploit a sliding window of the latest chunks and perform BA only within this window for both 3D-GS reconstruction and motion trajectory estimation. We will detail each component as follows.", + "bbox": [ + 89, + 633, + 491, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D Scene Representation", + "text_level": 1, + "bbox": [ + 89, + 801, + 318, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following 3D-GS [11], the scene is represented by a set of 3D Gaussian primitives, each of which contains mean position $\\pmb{\\mu} \\in \\mathbb{R}^3$ in the world coordinate, 3D covariance $\\pmb{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}$ , opacity $\\mathbf{o} \\in \\mathbb{R}$ , and color $\\mathbf{c} \\in \\mathbb{R}^3$ . To ensure that the covariance matrix remains positive semi-definite throughout the gradient descent, the covariance $\\pmb{\\Sigma}$ is parameterized using a", + "bbox": [ + 89, + 828, + 491, + 919 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scale vector $\\mathbf{s} \\in \\mathbb{R}^3$ and rotation matrix $\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}$ :", + "bbox": [ + 519, + 479, + 861, + 494 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 505, + 919, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where scale matrix $\\mathbf{S} = \\text{diag}([s])$ is derived from the scale vector $\\mathbf{s} \\in \\mathbb{R}^3$ .", + "bbox": [ + 517, + 534, + 919, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to enable rendering, 3D-GS projects 3D Gaussian primitives to the 2D image plane from a given camera pose $\\mathbf{T}_c = \\{\\mathbf{R}_c\\in \\mathbb{R}^{3\\times 3},\\mathbf{t}_c\\in \\mathbb{R}^3\\}$ using following equation:", + "bbox": [ + 519, + 565, + 919, + 612 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {R} _ {c} \\boldsymbol {\\Sigma} \\mathbf {R} _ {c} ^ {T} \\mathbf {J} ^ {T}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 622, + 919, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{\\Sigma}^{\\prime}\\in \\mathbb{R}^{2\\times 2}$ is the 2D covariance matrix, $J\\in \\mathbb{R}^{2\\times 3}$ is the Jacobian of the affine approximation of the projective transformation. After projecting 3D Gaussians onto the image plane, the color of each pixel is determined by sorting the Gaussians according to their depth and then applying near-to-far $\\alpha$ -blending rendering via the following equation:", + "bbox": [ + 517, + 651, + 921, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} = \\sum_ {i} ^ {N} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 753, + 919, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{c}_i$ is the learnable color of each Gaussian, and $\\alpha_{i}$ is the alpha value computed by evaluating the 2D covariance $\\boldsymbol{\\Sigma}^{\\prime}$ multiplied with the learned Gaussian opacity $\\mathbf{o}$ :", + "bbox": [ + 517, + 806, + 919, + 853 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\mathbf {o} _ {i} \\cdot \\exp (- \\sigma_ {i}), \\quad \\sigma_ {i} = \\frac {1}{2} \\Delta_ {i} ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime - 1} \\Delta_ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 863, + 919, + 892 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Delta_i\\in \\mathbb{R}^2$ is the offset between the pixel center and the", + "bbox": [ + 519, + 902, + 919, + 919 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26935", + "bbox": [ + 478, + 953, + 517, + 965 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D Gaussian center. Depth is rendered by:", + "bbox": [ + 89, + 90, + 372, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} = \\sum_ {i} ^ {N} \\mathbf {d} _ {i} \\alpha_ {i} \\prod_ {j} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 116, + 488, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{i}$ denotes the z-depth of the center of the i-th 3D Gaussian to the camera. We also render alpha map to determine visibility:", + "bbox": [ + 89, + 171, + 490, + 215 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} = \\sum_ {i} ^ {N} \\alpha_ {i} \\prod_ {j} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 215, + 488, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The derivations presented above demonstrate that the rendered pixel color, denoted as I in Eq. (3), is a function that is differentiable with respect to the learnable attributes of 3D-GS, and the camera poses $\\mathbf{T}_c$ . This facilitates our bundle adjustment formulation, accommodating a set of event chunks and inaccurate camera motion trajectories within the framework of 3D-GS.", + "bbox": [ + 89, + 263, + 491, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.Event Data Formation Model", + "text_level": 1, + "bbox": [ + 89, + 378, + 352, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An event camera records changes in brightness as a stream of events asynchronously. To relate 3D-GS representation with the event stream, we sample two close consecutive timestamps (i.e., $t_k$ and $t_{k + \\Delta t}$ , where $\\Delta t$ is a small time interval), and the measured brightness change between $\\Delta t$ is:", + "bbox": [ + 89, + 397, + 490, + 472 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} (\\mathbf {x}) = C \\left\\{e _ {i} \\left(\\mathbf {x}, t _ {i}, p _ {i}\\right) \\right\\} _ {t _ {k} < t _ {i} < t _ {k} + \\Delta t}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 484, + 488, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $e(\\mathbf{x}, t_i, p_i)$ is the $i^{th}$ event within the defined time interval corresponding to pixel $\\mathbf{x}$ and $C$ is the fixed contrast threshold. The corresponding camera poses $T_k$ and $T_{k + \\Delta t}$ can be interpolated from the camera motion trajectory parameterization, and two corresponding brightness images (i.e. $\\hat{\\mathbf{I}}_k$ and $\\hat{\\mathbf{I}}_{k + \\Delta t}$ ) can be rendered from 3D-GS. The synthesized brightness change $\\hat{\\mathbf{E}}$ is modelled as:", + "bbox": [ + 89, + 512, + 490, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {E}} (\\mathbf {x}) = \\log \\left(\\hat {\\mathbf {I}} _ {k + \\Delta t} (\\mathbf {x})\\right) - \\log \\left(\\hat {\\mathbf {I}} _ {k} (\\mathbf {x})\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 628, + 488, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{E}} (\\mathbf{x})$ depends on the parameters of both the motion trajectory parameters and 3D-GS, and is differentiable with respect to them.", + "bbox": [ + 89, + 659, + 488, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Both in tracking and mapping, inspired by the work of [25], we segment the current event chunks into $n_{seg}$ equal segments according to the number of events, obtaining $n_{seg}$ timestamps that correspond to the end of each segment. We then randomly select one timestamp from these $n_{seg}$ timestamps to serve as $t_{k + \\Delta t}$ , and we randomly sample an integer $n_{win}$ between the integer bounds $n_{low}$ and $n_{up}$ . The index of $t_k$ is equal to the index of $t_{k + \\Delta t}$ subtract $n_{win}$ . $n_{seg}$ , $n_{low}$ and $n_{up}$ are hyperparameters. This sampling strategy enables the model to capture both local and global information.", + "bbox": [ + 89, + 705, + 488, + 857 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Camera Motion Trajectory Modeling", + "text_level": 1, + "bbox": [ + 89, + 866, + 413, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since each event chunk usually contains too many events, we sample a portion of them according to the total number of", + "bbox": [ + 89, + 888, + 490, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "events during optimization. Following [32, 38], we formulate the corresponding poses (i.e. $\\mathbf{T}_k$ and $\\mathbf{T}_{k + \\Delta t}$ ) at the beginning and end of the sampled event portion within each chunk, by employing a camera motion trajectory. The trajectory is represented through linear interpolation between two camera poses, one at the beginning of the chunk $\\mathbf{T}_{\\mathrm{start}} \\in \\mathbf{SE}(3)$ and the other at the end $\\mathbf{T}_{\\mathrm{end}} \\in \\mathbf{SE}(3)$ . The camera pose at time $t_k$ can thus be expressed as follows:", + "bbox": [ + 517, + 90, + 919, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {k} = \\mathbf {T} _ {\\text {s t a r t}} \\cdot \\exp \\left(\\frac {t _ {k} - t _ {\\text {s t a r t}}}{t _ {\\text {e n d}} - t _ {\\text {s t a r t}}} \\cdot \\log \\left(\\mathbf {T} _ {\\text {s t a r t}} ^ {- 1} \\cdot \\mathbf {T} _ {\\text {e n d}}\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 222, + 919, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t_{start}$ and $t_{end}$ represent the timestamps corresponding to the boundary of the event chunk. It follows that $\\mathbf{T}_k$ is differentiable with respect to both $\\mathbf{T}_{start}$ and $\\mathbf{T}_{end}$ . The objective of IncEventGS is thus to estimate both $\\mathbf{T}_{start}$ and $\\mathbf{T}_{end}$ for each event chunk, along with the learnable parameters of 3D-GS $\\mathbf{G}_{\\theta}$ .", + "bbox": [ + 517, + 262, + 919, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Incremental Tracking and Mapping", + "text_level": 1, + "bbox": [ + 519, + 362, + 831, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For both tracking and mapping, we aim to minimize the difference between the synthesized and measured brightness changes. In particular, we compute the loss of the latest event chunk only for the tracking stage and minimize the following loss function:", + "bbox": [ + 517, + 383, + 919, + 458 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {\\text {s t a r t}} ^ {*}, \\mathbf {T} _ {\\text {e n d}} ^ {*} = \\underset {\\mathbf {T} _ {\\text {s t a r t}}, \\mathbf {T} _ {\\text {e n d}}} {\\operatorname {a r g m i n}} \\left\\| \\mathbf {E} (\\mathbf {x}) - \\hat {\\mathbf {E}} (\\mathbf {x}) \\right\\| _ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 465, + 919, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{E}} (\\mathbf{x})$ and $\\mathbf{E}(\\mathbf{x})$ are the synthesized and measured brightness changes respectively, corresponding to a randomly sampled event portion within the latest event chunk.", + "bbox": [ + 517, + 510, + 919, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once the tracking is done, we insert the latest event chunk to the mapper and exploit the estimated $\\mathbf{T}_{start}^{*}$ and $\\mathbf{T}_{end}^{*}$ as the initial value of the chunk to perform dense bundle adjustment. For computational consideration, we exploit a sliding window BA of the latest $n_w$ chunks, and $n_w$ is a hyperparameter. In particular, we optimize both the motion trajectories and the 3D-GS jointly by minimizing the following loss functions:", + "bbox": [ + 517, + 556, + 921, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {\\text {e v e n t}} + \\lambda \\mathcal {L} _ {\\text {s s i m}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 674, + 919, + 689 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e v e n t}} = \\left\\| \\mathbf {E} _ {i} (\\mathbf {x}) - \\hat {\\mathbf {E}} _ {i} (\\mathbf {x}) \\right\\| _ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 693, + 919, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s s i m} = S S I M \\left(\\mathbf {E} _ {i} (\\mathbf {x}), \\hat {\\mathbf {E}} _ {i} (\\mathbf {x})\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 723, + 919, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter, SSIM is the structural dissimilarity loss [34]. As the event data streams in, we alternatively perform tracking and mapping.", + "bbox": [ + 517, + 751, + 919, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. System Initialization and Boot-strapping", + "text_level": 1, + "bbox": [ + 519, + 806, + 867, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conventional frame-based 3D-GS methods usually require a good initial point cloud and camera poses computed via COLMAP [27] for initialization. However, they are usually not easy to obtain if we are only given a single event camera. We therefore initialize the 3D-GS by sampling point cloud randomly within a bounding box. The first $m$ event chunks", + "bbox": [ + 517, + 828, + 919, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26936", + "bbox": [ + 478, + 953, + 519, + 965 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(where $m$ is a hyperparameter) are selected for initialization, and all corresponding camera poses are randomly initialized to be near the identity matrix. We then minimize the loss computed by Eq. (11) with respect to the attributes of 3D-GS and the parameters of camera motion trajectories jointly.", + "bbox": [ + 88, + 90, + 491, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through experiments, we found that the above initialization procedure consistently produces satisfactory brightness images. However, the 3D structure remains of low quality due to the short baselines of the moving event camera. We further find that it could potentially affect the performance of the whole pipeline without a good initial 3D structure as more event data is received. Therefore, we utilize a monocular depth estimation network [9] to predict a dense depth map from the rendered brightness image after the pipeline is trained for certain iterations. This depth map is then used to re-initialize the centers of the 3D Gaussians by unprojecting the pixel depths, after which we repeat the minimization of Eq. (11) for system bootstrapping. More details about system re-initialization can be found in our supplementary material.", + "bbox": [ + 89, + 167, + 491, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D-GS Incrementally Growing. As the camera moves, new Gaussians is periodically introduced to cover newly explored regions. After tracking, we obtain an accurate camera pose estimate for each new event chunk. The centers of new Gaussians are determined by:", + "bbox": [ + 89, + 382, + 491, + 457 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} = \\mathbf {T} \\cdot \\pi^ {- 1} \\left(\\mathbf {u}, d _ {u}\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 468, + 488, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{u} \\in \\mathbb{R}^2$ is pixel coordinate in the image plane, $d_u$ is depth of the 3D point $\\mathbf{p}$ projecting onto the image plane, which is rendered by Eq. (5), $\\pi^{-1}$ denotes camera inverse projection, $\\mathbf{T}$ is the camera pose from tracking. To ensure that new Gaussians are only added in previously unmapped areas, a visibility mask is computed to guide the expansion of the Gaussian splattering process, as following:", + "bbox": [ + 89, + 497, + 490, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} (p) = V < \\lambda_ {V}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 617, + 488, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{V}$ is the rendered alpha map and $\\lambda_V$ is a hyperparameter.", + "bbox": [ + 89, + 645, + 488, + 675 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 690, + 223, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setups.", + "text_level": 1, + "bbox": [ + 89, + 714, + 290, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. All experiments are conducted on a desktop PC equipped with a 5.73GHz AMD Ryzen 9 7900x CPU and an NVIDIA RTX 3090 GPU. The first $m = 3$ event chunks are used for initialization. During the mapping stage, a sliding window size of $n_w = 20$ is employed for the bundle adjustment algorithm. The hyperparameters are set as follows: $\\lambda = 0.05$ , $\\lambda_V = 0.8$ , and $n_{seg} = 100$ . For the synthetic dataset, $n_{low} = 400k$ and $n_{up} = 500k$ , while for the real dataset, $n_{low} = 60k$ and $n_{up} = 80k$ . Each event chunk has a time interval of $50~\\mathrm{ms}$ . The learning rate of the camera poses is set to 1e-4 and that for the attributes of 3D-GS are set the same as the original 3D-GS work. The number", + "bbox": [ + 89, + 737, + 491, + 919 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of optimization steps for initialization is 4500, and that for tracking and mapping are set to 200 and 1500 respectively. The contrast threshold $C$ of the event camera is set to 0.1 for synthetic datasets and 0.2 for real datasets empirically.", + "bbox": [ + 517, + 90, + 921, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines and Evaluation Metrics. To the best of our knowledge, there are no existing event-only NeRF or 3D-GS methods that do not rely on ground-truth poses, making direct comparisons challenging. Therefore, we conduct a thorough comparison of our method with several event-based NeRF approaches, including E-NeRF [14], EventNeRF [25], and Robust e-NeRF [17], as well as our custom implemented two-stage method (i.e. E2VID [24] + COLMAP [27] + 3DGS [11]). E-NeRF, EventNeRF, and Robust e-NeRF leverage implicit neural radiance fields for 3D scene representation, requiring ground-truth camera poses for accurate NeRF reconstruction. For E2VID + COLMAP + 3DGS, event data is first converted into brightness images using E2VID. The camera poses are then estimated from these images using COLMAP, and 3D-GS is trained with the generated images and poses. Both the quantitative and qualitative comparisons are performed on the synthetic dataset. Since there are no paired ground truth images for the real dataset, we only perform qualitative comparisons on the real dataset. In terms of motion trajectory evaluations, we use the publicly available state-of-the-art event-only visual odometry methods, i.e. DEVO (mono) [15] and ESVO2(stereo) [21], for comparison.", + "bbox": [ + 517, + 159, + 921, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The metrics used for novel view synthesis (NVS) include the commonly employed PSNR, SSIM, and LPIPS. For motion trajectory evaluations, we utilize Absolute Trajectory Error (ATE). To ensure fair comparisons, we employ the evaluation code provided by EventNeRF to compute the NVS metrics, which applies a linear color transformation between predictions and ground truth. Additionally, we use the public EVO toolbox [5] to compute the trajectory metrics.", + "bbox": [ + 517, + 494, + 921, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark Datasets. To properly evaluate the performance of NVS and motion trajectory estimation, we synthesized event data using the 3D scene models from the Replica dataset [28]. In particular, we exploit the room0, room2, office0, office2, and office3 scenes. We render high frame rate RGB images at $1000\\mathrm{Hz}$ with a resolution of $768\\times 480$ pixels. These images are then converted to grayscale, and the event data is generated via the events simulator [4]. The contrast threshold is set to 0.1. To simulate real-world camera motions, we exploit the same motion trajectories as that of NICE-SLAM [39] for data generation.", + "bbox": [ + 517, + 657, + 921, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the event dataset provided by TUM-VIE [13] for real data evaluations, which is also used by E-NeRF and Robust e-NeRF. TUM-VIE captures the event datasets by a pair of Prophesee Gen4 HD event cameras with a resolution of $1280 \\times 720$ pixels. We only use the left-event camera data for our experiment.", + "bbox": [ + 517, + 828, + 921, + 919 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26937", + "bbox": [ + 478, + 953, + 517, + 965 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ec53771c8d5d065b90240b3c1440f322de20fa4c99f213696b8ee4ff704a4364.jpg", + "image_caption": [ + "Figure 2. Qualitative evaluation of novel view image synthesis on the Replica dataset. The experimental results demonstrate that our method renders higher-quality images with fewer artifacts compared to event-based NeRF and two-stage approaches." + ], + "image_footnote": [], + "bbox": [ + 99, + 87, + 903, + 566 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/56afb1d11992bf19ff7eb43305d1c389df28f36034040148ce777d5937dff71f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
room0room2office0office2office3
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
E-NeRF13.990.580.5115.560.470.5818.910.510.5713.050.650.4414.010.620.48
EventNeRF17.290.620.3916.020.540.6418.900.430.6215.180.660.4516.770.730.33
Robust e-NeRF17.260.840.1816.430.500.5218.930.520.5616.810.810.2519.220.840.18
E2VID+COLMAP+3DGS14.450.440.5215.740.510.5518.910.310.6814.030.570.4813.250.470.53
Ours24.310.850.1723.750.790.2325.640.540.3021.740.820.2321.180.880.13
", + "bbox": [ + 94, + 608, + 918, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. NVs performance comparison on Replica dataset. All results are calculated using the same code and ground-truth images. The results demonstrate that our method outperforms NeRF-based and two-stage methods.", + "bbox": [ + 89, + 736, + 919, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 789, + 243, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct ablation studies to confirm our design choices. In particular, we study the effect of a monocular depth estimation network for system bootstrapping and event slicing hyperparameters $n_{low}$ , $n_{up}$ . The experiments are conducted with the Replica dataset, and the results are shown in Table 3 and Table 4, respectively.", + "bbox": [ + 89, + 811, + 490, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We found that depth initialization significantly impacts", + "bbox": [ + 109, + 902, + 488, + 919 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pose estimation, reducing the Average Trajectory Error (ATE) from $1.534\\mathrm{cm}$ to $0.064\\mathrm{cm}$ . Additionally, this improvement in pose estimation leads to a slight enhancement in Novel View Synthesis (NVS) performance. These results verify the importance of using depth initialization during the bootstrapping stage.", + "bbox": [ + 517, + 790, + 921, + 881 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare several combinations of hyperparameters $n_{\\mathrm{low}}$ and $n_{\\mathrm{up}}$ , which refer to the range of event slicing window", + "bbox": [ + 519, + 888, + 921, + 919 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26938", + "bbox": [ + 478, + 953, + 519, + 965 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6a39b90025b518decbb9ec15db062d3021531a14f0ea1c08f9fb383b68e5b1e2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
room0room2office0office2office31d3d6dofdeskdesk2
DEVO0.2890.2660.1380.2810.1560.1470.3032.930.7320.201
E2VID+COLMAP17.9359.96105.1918.41417.284.26816.909.8821.5710.13
ESVO2-----0.3371.0660.5871.1472.506
Ours0.0460.0670.0450.0460.0540.1150.2980.2510.2310.129
", + "bbox": [ + 158, + 88, + 854, + 167 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2dd37915808dbe380d06c56b0ff7983954ea747e75aea35bc72d8321d1f6e01a.jpg", + "table_caption": [ + "Table 2. Pose accuracy (ATE, cm) on Replica and TUM-VIE datasets. The results demonstrate that our method delivers better performance in terms of camera motion estimation. Since ESVO2 is a stereo event camera method, it cannot run on the Replica dataset." + ], + "table_footnote": [], + "table_body": "
SettingPSNR↑SSIM↑LPIPS↓ATE
full21.740.820.230.046
w/o17.800.760.261.534
", + "bbox": [ + 148, + 229, + 431, + 286 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8a791a93fc4c5e0ab86c7d35530d43d92be63ce3acb925b55ea6ba7af276c183.jpg", + "table_caption": [ + "Table 3. Ablation Study about Depth Initialization. The unit of ATE is cm. The experimental results demonstrate the effectiveness of the initialization strategy. It not only improves the quality of rendered images, but also improves the accuracy of the camera motion estimation significantly." + ], + "table_footnote": [], + "table_body": "
SettingPSNR↑SSIM↑LPIPS↓ATE
1k-10k16.070.640.460.167
10k-50k18.410.720.330.079
80k-200k20.990.790.250.079
400k-500k21.740.820.230.046
500k-600k20.950.790.230.050
600k-700k18.060.750.280.214
", + "bbox": [ + 138, + 381, + 444, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation Study on Event Slice Window Size (Hyperparameters $n_{\\mathrm{low}}$ and $n_{\\mathrm{up}}$ ). The unit of ATE is cm.", + "bbox": [ + 89, + 496, + 491, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "size. Table 4 demonstrates that both too small and too large window sizes negatively impact the performance of Novel View Synthesis (NVS) and pose estimation. Consequently, we select $n_{\\mathrm{low}} = 400k$ and $n_{\\mathrm{up}} = 500k$ for our experiments on the Replica dataset.", + "bbox": [ + 89, + 532, + 490, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Quantitative Evaluations.", + "text_level": 1, + "bbox": [ + 89, + 621, + 323, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct quantitative evaluations against event NeRF methods (E-NeRF, EventNeRF, and Robust e-NeRF) and our custom implemented two-stage method (i.e. E2VID + COLMAP + 3DGS) in terms of the quality of NVS and pose estimation performance.", + "bbox": [ + 89, + 643, + 488, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The NVS performance is evaluated on Replica-dataset and the results are presented in Table 1. It is important to note that the metrics are lower than those typically observed in standard NeRF/3D-GS methods for RGB images, primarily due to the lack of absolute brightness supervision. Even though NeRF-based methods use ground truth poses for training, IncEventGS still significantly outperforms them, highlighting the advantages of our approach utilizing a 3D Gaussian representation. Additionally, our method greatly surpasses two-stage method that also employs 3D Gaussian representation, demonstrating superior pose estimation and the effectiveness of our bundle adjustment technique.", + "bbox": [ + 89, + 720, + 491, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate pose estimation performance using the ATE", + "bbox": [ + 109, + 902, + 488, + 919 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "metric on both synthetic and real datasets, comparing our method with DEVO, ESVO2 and E2VID + COLMAP. The results, presented in Table 2, show that our method outperforms both baselines, validating the effectiveness of our incremental tracking and mapping technique.", + "bbox": [ + 517, + 233, + 921, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative Evaluations.", + "text_level": 1, + "bbox": [ + 519, + 330, + 741, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our method against event NeRF methods and two-stage method qualitatively in terms of novel view image synthesis, both on synthetic and real data. The results are presented in both Fig. 2 and Fig. 3. It demonstrates that our method can deliver better novel view images, while event NeRF methods and two-stage method render images with additional artifacts. Compared to NeRF-based methods, our approach demonstrates the advantage of IncEventGS by leveraging 3D Gaussian Splatting as the underlying scene representation. In contrast to two-stage method, our dense bundle adjustment optimizes both 3D Gaussian Splatting and camera pose using event data, whereas two-stage approaches tend to accumulate errors over time, as confirmed by the experimental results. We also provide representative visualization of ATE error mapped onto trajectories in Fig. 4, both on synthetic and real dataset. It demonstrates that IncEventGS is able to recover more accurate motion trajectories.", + "bbox": [ + 517, + 356, + 921, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 519, + 638, + 638, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present the first pose-free 3D Gaussian splattering reconstruction model based on event camera, i.e. IncEventGS, and is useful for real world applications. We adopt the tracking and mapping paradigm in conventional SLAM pipeline to do incremental motion estimation and 3D scene reconstruction simultaneously. To handle the continuous and asynchronous characteristics of event stream, we exploit a continuous trajectory model to model the event data formation process. The experimental results on both synthetic and real datasets demonstrate the superior performance of IncEventGS over prior state-of-the-art methods in terms of high-quality novel image synthesis and camera pose estimation.", + "bbox": [ + 517, + 669, + 921, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by NSFC under Grant 62202389, in part by a grant from the Westlake University-Muyuan Joint Research Institute, and in part by the Westlake Education Foundation.", + "bbox": [ + 517, + 858, + 919, + 919 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26939", + "bbox": [ + 478, + 953, + 519, + 965 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43889b42d33be00152edcc0f79ae10f5bde5061a44ffc0b75c25a0585944cf18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 83, + 903, + 513 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2cee0c49b5ec01484f8577bd27e78f92f0bfa55c347f1ec4985960b6176b495.jpg", + "image_caption": [ + "Figure 3. Qualitative evaluation for novel view image synthesis on real dataset. It demonstrates that our method is able to render better images with fewer artifacts than event NeRF methods and two-stage methods. Note that there are no GT images aligned with the event camera, and we choose the closest images from the RGB camera and crop them to the same size as the rendered images for visual comparisons.", + "Figure 4. Representative visualization of ATE error mapped onto trajectories for the synthetic (office0) and real (6 dof) datasets, generated by the EVO toolbox using the same ground truth poses, demonstrating the superior performance of our method in pose estimation." + ], + "image_footnote": [], + "bbox": [ + 96, + 564, + 901, + 867 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26940", + "bbox": [ + 478, + 953, + 517, + 964 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hiroyuki Deguchi, Mana Masuda, Takuya Nakabayashi, and Hideo Saito. E2gs: Event enhanced gaussian splatting. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1676-1682. IEEE, 2024. 2", + "[2] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. arXiv preprint arXiv:2312.07504, 2023. 2", + "[3] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth and optical flow estimation. In CVPR, 2018. 1", + "[4] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 5", + "[5] Michael Grupp. evo: Python package for the evaluation of odometry and slam. https://github.com/MichaelGrupp/evo, 2017.5", + "[6] Jiarui Hu, Xianhao Chen, Boyin Feng, Guanglin Li, Liangjing Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field. arXiv preprint arXiv:2403.16095, 2024. 2", + "[7] Huajian Huang, Longwei Li, Hui Cheng, and Sai-Kit Yeung. Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular, stereo, and rgb-d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[8] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-nerf: Event based neural radiance field. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 837–847, 2023. 1, 2", + "[9] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5", + "[10] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[11] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 2023. 1, 2, 3, 5", + "[12] Hanme Kim, Stefan Leutenegger, and Andrew J. Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In ECCV, 2016. 1", + "[13] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. Tum-vie: The tum stereo visual-inertial event dataset. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8601-8608. IEEE, 2021. 5", + "[14] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 8(3):1587-1594, 2023. 1, 2, 5" + ], + "bbox": [ + 91, + 114, + 491, + 917 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Simon Klenk, Marvin Motzet, Lukas Koestler, and Daniel Cremers. Deep event visual odometry. In 2024 International Conference on 3D Vision (3DV), pages 739-749. IEEE, 2024. 1, 5", + "[16] Wenpu Li, Pian Wan, Peng Wang, Jinghang Li, Yi Zhou, and Peidong Liu. Benerf: Neural radiance fields from a single blurry image and event stream. In European Conference on Computer Vision (ECCV), 2024. 2", + "[17] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse and noisy events under non-uniform motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 5", + "[18] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 2", + "[19] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1", + "[20] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE Transactions on Robotics, 33(5):1255-1262, 2017. 2", + "[21] Junkai Niu, Sheng Zhong, Xiuyuan Lu, Shaojie Shen, Guillermo Gallego, and Yi Zhou. Esvo2: Direct visual-inertial odometry with stereo event cameras. IEEE Transactions on Robotics, 2025. 5", + "[22] Delin Qu, Chi Yan, Dong Wang, Jie Yin, Dan Xu, Bin Zhao, and Xuelong Li. Implicit event-rgbd neural slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[23] Henri Rebecq, Timo Horstschaefer, Guillermo Gallego, and Davide Scaramuzza. Evo: A geometric approach to event-based 6-dof parallel tracking and mapping in real time. IEEE Robotics and Automation Letters, 2017. 1", + "[24] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 5", + "[25] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4992-5002, 2023. 1, 2, 4, 5", + "[26] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion Revisited. In CVPR, pages 4104-4113, 2016. 2", + "[27] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 5", + "[28] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica" + ], + "bbox": [ + 522, + 92, + 921, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26941", + "bbox": [ + 478, + 953, + 517, + 965 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5", + "[29] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 2", + "[30] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. Advances in Neural Information Processing Systems, 2023. 1", + "[31] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. arXiv preprint arXiv:2405.14959, 2024. 2", + "[32] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4170–4179, 2023. 4", + "[33] Peng Wang, Lingzhe Zhao, Yin Zhang, Shiyu Zhao, and Peidong Liu. Mba-slam: Motion blur aware dense visual slam with radiance fields representation. arXiv preprint arXiv:2411.08279, 2024. 2", + "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4", + "[35] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher A Metzler. Event3dgs: Event-based 3d gaussian splattering for fast egomotion. arXiv preprint arXiv:2406.02972, 2024. 2", + "[36] Chi Yan, Delin Qu, Dong Wang, Dan Xu, Zhigang Wang, Bin Zhao, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[37] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2", + "[38] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. In European Conference on Computer Vision (ECCV), 2024. 2, 4", + "[39] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 5" + ], + "bbox": [ + 91, + 90, + 488, + 700 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26942", + "bbox": [ + 478, + 954, + 517, + 965 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/c5e31ebc-2f5f-4333-a4a5-98739dcdbce4_model.json b/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/c5e31ebc-2f5f-4333-a4a5-98739dcdbce4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bd88535bd51a8512761695ab771b8aa4c6752058 --- /dev/null +++ b/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/c5e31ebc-2f5f-4333-a4a5-98739dcdbce4_model.json @@ -0,0 +1,1881 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.87, + 0.153 + ], + "angle": 0, + "content": "IncEventGS: Pose-Free Gaussian Splating from a Single Event Camera" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.193, + 0.812, + 0.213 + ], + "angle": 0, + "content": "Jian Huang\\(^{1,2}\\) Chengrui Dong\\(^{1,2}\\) Xuanhua Chen\\(^{2,3}\\) Peidong Liu\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.223, + 0.812, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\)Zhejiang University \\(^{2}\\)Westlake University \\(^{3}\\)Northeastern University \\(\\{\\)huangjian39, dongchengrui, chenxuanhua, liupeidong\\}@westlake.edu.cn https://github.com/WU-CVGL/IncEventGS" + }, + { + "type": "title", + "bbox": [ + 0.252, + 0.314, + 0.33, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.494, + 0.722 + ], + "angle": 0, + "content": "Implicit neural representation and explicit 3D Gaussian Splating (3D-GS) for novel view synthesis have achieved remarkable progress with frame-based camera (e.g. RGB and RGB-D cameras) recently. Compared to frame-based camera, a novel type of bio-inspired visual sensor, i.e. event camera, has demonstrated advantages in high temporal resolution, high dynamic range, low power consumption, and low latency, which make it favored for many robotic applications. In this work, we present IncEventGS, an incremental 3D Gaussian Splating reconstruction algorithm with a single event camera, without the assumption of known camera poses. To recover the 3D scene representation incrementally, we exploit the tracking and mapping paradigm of conventional SLAM pipelines for IncEventGS. Given the incoming event stream, the tracker first estimates an initial camera motion based on prior reconstructed 3D-GS scene representation. The mapper then jointly refines both the 3D scene representation and camera motion based on the previously estimated motion trajectory from the tracker. The experimental results demonstrate that IncEventGS delivers superior performance compared to prior NeRF-based methods and other related baselines, even if we do not have the ground-truth camera poses. Furthermore, our method can also deliver better performance compared to state-of-the-art event visual odometry methods in terms of camera motion estimation." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.751, + 0.222, + 0.767 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.776, + 0.491, + 0.897 + ], + "angle": 0, + "content": "Reconstructing accurate 3D scene representations from 2D images has been a long-standing challenge in computer vision and robotics, driving substantial efforts over the past few decades. Among those pioneering works, Neural Radiance Fields (NeRF) [19] and 3D Gaussian Splatting (3D-GS) [11], stand out for their utilization of differentiable rendering techniques, and have garnered significant attention due to their capability to recover high-quality 3D scene representation" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.315, + 0.923, + 0.451 + ], + "angle": 0, + "content": "from 2D images. Commonly used sensors for 3D scene reconstruction are usually frame-based cameras, such as the RGB and RGB-D cameras. They usually capture full-brightness intensity images within a short exposure time at a regular frequency. Due to the characteristic of this data-capturing process, they often suffer from motion blur or fail to capture accurate and informative intensity information under fast motion and low-light conditions, which would further affect the performance of downstream applications." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.452, + 0.923, + 0.648 + ], + "angle": 0, + "content": "The event camera, a bio-inspired sensor, has gained significant attention in recent years for its potential to address the limitations of frame-based cameras under challenging conditions. Unlike conventional cameras, event cameras record brightness changes asynchronously at each pixel, emitting events when a predefined threshold is surpassed. This unique operation offers several advantages over conventional cameras, in terms of high temporal resolution, high dynamic range, low latency, and power consumption. Although event cameras have attractive characteristics for challenging environments, they cannot be directly integrated into existing frame-based 3D reconstruction algorithms that rely on processing dense 2D brightness intensity images." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.648, + 0.923, + 0.859 + ], + "angle": 0, + "content": "Several pioneering works have been proposed to exploit event stream [3, 12, 23] to recover the motion trajectory and scene representation. While existing methods deliver impressive performance, they usually exploit 2.5D semi-dense depth maps to represent the 3D scene, and bundle adjustment (BA) is hardly performed, due to the asynchronous and sparse characteristics of event data stream. Klenk et al. [15] recently proposed to convert event stream into event voxel grids, and then adapt a previous frame-based deep visual odometry pipeline [30] for accurate camera motion estimation. As NeRF exhibited impressive scene representation capability recently, several works [8, 14, 17, 25] explore to recover the underlying dense 3D scene NeRF representation from event stream, by assuming ground-truth poses are available." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.859, + 0.922, + 0.92 + ], + "angle": 0, + "content": "In contrast to those works, we propose IncEventGS, an incremental dense 3D scene reconstruction algorithm from a single event camera, by exploiting Gaussian Splatting as the underlying scene representation. Different from prior event-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.906, + 0.242, + 0.918 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.954, + 0.519, + 0.966 + ], + "angle": 0, + "content": "26933" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.493, + 0.468 + ], + "angle": 0, + "content": "based NeRF reconstruction methods, IncEventGS does not require any ground-truth camera poses, which is more challenging and provides more flexibility for real-world robotic application scenarios. To overcome the challenges brought by unknown poses, IncEventGS adopts the tracking and mapping paradigm of conventional SLAM pipelines [20]. In particular, IncEventGS exploits prior explored and reconstructed 3D scenes for camera motion estimation of incoming event stream during the tracking stage. Both the 3D-GS scene representation and camera motions are then jointly optimized (i.e. event-based bundle adjustment) during the mapping stage, for more accurate scene representation and motion estimation. The 3D scene is progressively expanded and densified. The experimental results on both synthetic and real datasets demonstrate that IncEventGS can recover the underlying 3D scene representation and camera motion trajectory accurately. In particular, IncEventGS outperforms prior NeRF-based methods and other related baselines in terms of scene representation recovery, even without ground-truth poses. Furthermore, our method also delivers better camera motion estimation accuracy than the most recent state-of-the-art visual odometry algorithm, in terms of the Absolute Trajectory Error (ATE) metric. The recovered 3D scene representation can be further used to render novel brightness images. Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.469, + 0.49, + 0.514 + ], + "angle": 0, + "content": "- We present an incremental 3D Gaussian Splitting reconstruction algorithm from a single event camera, without requiring the ground-truth camera poses." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.515, + 0.492, + 0.558 + ], + "angle": 0, + "content": "- We propose a novel initialization strategy tailored to the event data stream, which is vital to the success of the algorithm." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.56, + 0.492, + 0.651 + ], + "angle": 0, + "content": "- The experimental results on both the synthetic and real datasets demonstrate the superior performance of our method over prior NeRF-based methods and related baselines in terms of novel view synthesis and better performance over state-of-the-art event-based visual odometry algorithm in terms of camera motion estimation." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.469, + 0.492, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.241, + 0.68 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.49, + 0.735 + ], + "angle": 0, + "content": "We review two main areas of prior works: event-based neural radiance fields and 3D Gaussian Splatting, which are the most related to our work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.493, + 0.92 + ], + "angle": 0, + "content": "Event-based Neural Radiance Fields. Prior works [8, 14, 25] propose to exploit event stream to recover the neural radiance fields with known camera motion trajectory. Low et al. [17] further improves the reconstruction algorithm to handle sparse and noisy events under non-uniform motion. The recovered neural radiance fields can then be used to render novel view brightness images. The ground-truth poses are usually computed from corresponding brightness images via COLMAP [26] or provided by the indoor motion-capturing system. Recently, Qu et al. [22] proposed to integrate event measurements into an RGB-D implicit neural SLAM framework and achieved robust performance in mo" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.092, + 0.923, + 0.243 + ], + "angle": 0, + "content": "tion blur scenarios. Li et al. [16] also proposed exploiting event measurements and a single blurry image to recover the underlying neural 3D scene representation. In contrast to those works, IncEventGS conducts incremental 3D scene reconstruction without requiring any prior ground-truth poses, which is more challenging and provides more flexibility for practical robotic application scenarios. The method further exploits 3D Gaussian Splatting as the underlying scene representation, which demonstrates better image rendering quality and efficiency, compared to the NeRF-based representation." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.246, + 0.923, + 0.761 + ], + "angle": 0, + "content": "3D Gaussian Splatting. 3D Gaussian Splatting [11] proposes a novel explicit 3D representation to further improve both the training and rendering efficiency compared to Neural Radiance Fields. Due to its impressive efficient scene representation capability, several pioneering works have been proposed to exploit 3D-GS for incremental 3D reconstruction. For example, Keetha et al. [10] propose an RGBD-based 3D-GS SLAM, employing an online tracking and mapping system tailored to the underlying Gaussian representation. Yan et al. [36] implement a coarse-to-fine camera tracking approach based on the sparse selection of Gaussians. Matsuki et al. [18] propose to apply 3D Gaussian Splatting to do incremental 3D reconstruction using a single moving monocular or RGB-D camera. Huang et al. [7] exploit ORB-SLAM3 to compute accurate camera poses and feed it into a 3D-GS algorithm for dense mapping. Fu et al. [2] use monocular depth estimation with 3D-GS. Yugay et al. [37] combine DROID-SLAM [29] based camera tracking with active and inactive 3D-GS sub-maps. Wang et al. [33] integrate bundle-adjustment and 3DGS [38] to estimated camera trajectory within the exposure time. Hu et al. [6] propose a novel depth uncertainty model to ensure the selection of valuable Gaussian primitives during optimization. While those methods deliver impressive performance in terms of 3D scene recovery and motion estimation, they usually assume the usage of frame-based images (i.e. either RGB or RGB-D date). On the contrary, we propose to exploit pure event measurements for incremental 3D-GS reconstruction. Several concurrent studies have recently explored the use of 3D Gaussians for event-based reconstruction, including EvGGS [31], Event3DGS [35], and E2GS [1]. EvGGS and Event3DGS rely solely on event data, whereas E2GS incorporates both event data and blurry images. However, the key difference between these methods and ours is that they all rely on ground-truth poses." + }, + { + "type": "title", + "bbox": [ + 0.521, + 0.773, + 0.613, + 0.789 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.799, + 0.922, + 0.92 + ], + "angle": 0, + "content": "The overview of our IncEventGS is shown in Fig. 1. Given only a single event camera, IncEventGS incrementally performs tracking and dense mapping under the framework of 3D Gaussian Splatting, to recover both the camera motion trajectory and 3D scene representation simultaneously. Since event data are asynchronous, they cannot be directly integrated with the 3D-GS representation. We therefore process the event data stream into chunks according to a fixed time window." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.954, + 0.521, + 0.966 + ], + "angle": 0, + "content": "26934" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.09, + 0.856, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.407, + 0.922, + 0.476 + ], + "angle": 0, + "content": "Figure 1. The pipeline of IncEventGS. IncEventGS processes incoming event stream by dividing it into chunks and representing the camera trajectory as a continuous model. It randomly samples two close consecutive timestamps to integrate the corresponding event streams. Two brightness images are rendered from 3D-GS at the corresponding poses, and we minimize the photometric loss between the synthesized and measured brightness change. During initialization, a pre-trained depth estimation model estimates depth from the rendered images to bootstrap the system." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.48, + 0.491, + 0.632 + ], + "angle": 0, + "content": "We associate each chunk with a continuous time trajectory parameterization in the \\(\\mathfrak{se}(3)\\) space. Two close consecutive timestamps (i.e., \\(t_k\\) and \\(t_{k + \\Delta t}\\), where \\(\\Delta t\\) is a small time interval) can be randomly sampled, and the measured brightness change \\(\\mathbf{E}(x)\\) can be computed from the corresponding event stream. Based on the parameterized trajectory, the corresponding camera poses (i.e., \\(\\mathbf{T}_k\\), \\(\\mathbf{T}_{k + \\Delta t}\\)) can be determined and the brightness images (i.e., \\(\\bar{\\mathbf{I}}_k\\), \\(\\bar{\\mathbf{I}}_{k + \\Delta t}\\)) can be further rendered from the 3D-GS. The synthesized brightness change \\(\\hat{\\mathbf{E}}(x)\\) can be computed for event loss computation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.493, + 0.786 + ], + "angle": 0, + "content": "During the tracking stage, we optimize only the camera motion trajectory of the newly accumulated event chunk and exploit the recovered trajectory to initialize the dense bundle adjustment (BA) algorithm for the mapping stage. During the mapping stage, we continuously densify 3D Gaussians for newly explored areas and prune transparent 3D Gaussians. For computational efficiency, we exploit a sliding window of the latest chunks and perform BA only within this window for both 3D-GS reconstruction and motion trajectory estimation. We will detail each component as follows." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.32, + 0.82 + ], + "angle": 0, + "content": "3.1. 3D Scene Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.829, + 0.492, + 0.92 + ], + "angle": 0, + "content": "Following 3D-GS [11], the scene is represented by a set of 3D Gaussian primitives, each of which contains mean position \\(\\pmb{\\mu} \\in \\mathbb{R}^3\\) in the world coordinate, 3D covariance \\(\\pmb{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}\\), opacity \\(\\mathbf{o} \\in \\mathbb{R}\\), and color \\(\\mathbf{c} \\in \\mathbb{R}^3\\). To ensure that the covariance matrix remains positive semi-definite throughout the gradient descent, the covariance \\(\\pmb{\\Sigma}\\) is parameterized using a" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.48, + 0.862, + 0.495 + ], + "angle": 0, + "content": "scale vector \\(\\mathbf{s} \\in \\mathbb{R}^3\\) and rotation matrix \\(\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.506, + 0.921, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.535, + 0.92, + 0.564 + ], + "angle": 0, + "content": "where scale matrix \\(\\mathbf{S} = \\text{diag}([s])\\) is derived from the scale vector \\(\\mathbf{s} \\in \\mathbb{R}^3\\)." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.566, + 0.921, + 0.613 + ], + "angle": 0, + "content": "In order to enable rendering, 3D-GS projects 3D Gaussian primitives to the 2D image plane from a given camera pose \\(\\mathbf{T}_c = \\{\\mathbf{R}_c\\in \\mathbb{R}^{3\\times 3},\\mathbf{t}_c\\in \\mathbb{R}^3\\}\\) using following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.623, + 0.921, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {R} _ {c} \\boldsymbol {\\Sigma} \\mathbf {R} _ {c} ^ {T} \\mathbf {J} ^ {T}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.652, + 0.922, + 0.745 + ], + "angle": 0, + "content": "where \\(\\pmb{\\Sigma}^{\\prime}\\in \\mathbb{R}^{2\\times 2}\\) is the 2D covariance matrix, \\(J\\in \\mathbb{R}^{2\\times 3}\\) is the Jacobian of the affine approximation of the projective transformation. After projecting 3D Gaussians onto the image plane, the color of each pixel is determined by sorting the Gaussians according to their depth and then applying near-to-far \\(\\alpha\\) -blending rendering via the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.755, + 0.921, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} = \\sum_ {i} ^ {N} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.808, + 0.921, + 0.854 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}_i\\) is the learnable color of each Gaussian, and \\(\\alpha_{i}\\) is the alpha value computed by evaluating the 2D covariance \\(\\boldsymbol{\\Sigma}^{\\prime}\\) multiplied with the learned Gaussian opacity \\(\\mathbf{o}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.864, + 0.921, + 0.893 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\mathbf {o} _ {i} \\cdot \\exp (- \\sigma_ {i}), \\quad \\sigma_ {i} = \\frac {1}{2} \\Delta_ {i} ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime - 1} \\Delta_ {i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.904, + 0.921, + 0.92 + ], + "angle": 0, + "content": "where \\(\\Delta_i\\in \\mathbb{R}^2\\) is the offset between the pixel center and the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.954, + 0.519, + 0.966 + ], + "angle": 0, + "content": "26935" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.373, + 0.106 + ], + "angle": 0, + "content": "2D Gaussian center. Depth is rendered by:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.117, + 0.49, + 0.161 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} = \\sum_ {i} ^ {N} \\mathbf {d} _ {i} \\alpha_ {i} \\prod_ {j} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.172, + 0.491, + 0.216 + ], + "angle": 0, + "content": "where \\(d_{i}\\) denotes the z-depth of the center of the i-th 3D Gaussian to the camera. We also render alpha map to determine visibility:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.217, + 0.49, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} = \\sum_ {i} ^ {N} \\alpha_ {i} \\prod_ {j} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.265, + 0.492, + 0.369 + ], + "angle": 0, + "content": "The derivations presented above demonstrate that the rendered pixel color, denoted as I in Eq. (3), is a function that is differentiable with respect to the learnable attributes of 3D-GS, and the camera poses \\(\\mathbf{T}_c\\). This facilitates our bundle adjustment formulation, accommodating a set of event chunks and inaccurate camera motion trajectories within the framework of 3D-GS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.38, + 0.353, + 0.394 + ], + "angle": 0, + "content": "3.2.Event Data Formation Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.491, + 0.473 + ], + "angle": 0, + "content": "An event camera records changes in brightness as a stream of events asynchronously. To relate 3D-GS representation with the event stream, we sample two close consecutive timestamps (i.e., \\( t_k \\) and \\( t_{k + \\Delta t} \\), where \\( \\Delta t \\) is a small time interval), and the measured brightness change between \\( \\Delta t \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.486, + 0.49, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} (\\mathbf {x}) = C \\left\\{e _ {i} \\left(\\mathbf {x}, t _ {i}, p _ {i}\\right) \\right\\} _ {t _ {k} < t _ {i} < t _ {k} + \\Delta t}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.491, + 0.62 + ], + "angle": 0, + "content": "where \\( e(\\mathbf{x}, t_i, p_i) \\) is the \\( i^{th} \\) event within the defined time interval corresponding to pixel \\( \\mathbf{x} \\) and \\( C \\) is the fixed contrast threshold. The corresponding camera poses \\( T_k \\) and \\( T_{k + \\Delta t} \\) can be interpolated from the camera motion trajectory parameterization, and two corresponding brightness images (i.e. \\( \\hat{\\mathbf{I}}_k \\) and \\( \\hat{\\mathbf{I}}_{k + \\Delta t} \\)) can be rendered from 3D-GS. The synthesized brightness change \\( \\hat{\\mathbf{E}} \\) is modelled as:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.63, + 0.49, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {E}} (\\mathbf {x}) = \\log \\left(\\hat {\\mathbf {I}} _ {k + \\Delta t} (\\mathbf {x})\\right) - \\log \\left(\\hat {\\mathbf {I}} _ {k} (\\mathbf {x})\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.489, + 0.706 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{E}} (\\mathbf{x})\\) depends on the parameters of both the motion trajectory parameters and 3D-GS, and is differentiable with respect to them." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.707, + 0.49, + 0.858 + ], + "angle": 0, + "content": "Both in tracking and mapping, inspired by the work of [25], we segment the current event chunks into \\( n_{seg} \\) equal segments according to the number of events, obtaining \\( n_{seg} \\) timestamps that correspond to the end of each segment. We then randomly select one timestamp from these \\( n_{seg} \\) timestamps to serve as \\( t_{k + \\Delta t} \\), and we randomly sample an integer \\( n_{win} \\) between the integer bounds \\( n_{low} \\) and \\( n_{up} \\). The index of \\( t_k \\) is equal to the index of \\( t_{k + \\Delta t} \\) subtract \\( n_{win} \\). \\( n_{seg} \\), \\( n_{low} \\) and \\( n_{up} \\) are hyperparameters. This sampling strategy enables the model to capture both local and global information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.867, + 0.414, + 0.884 + ], + "angle": 0, + "content": "3.3. Camera Motion Trajectory Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.89, + 0.491, + 0.919 + ], + "angle": 0, + "content": "Since each event chunk usually contains too many events, we sample a portion of them according to the total number of" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.092, + 0.921, + 0.213 + ], + "angle": 0, + "content": "events during optimization. Following [32, 38], we formulate the corresponding poses (i.e. \\(\\mathbf{T}_k\\) and \\(\\mathbf{T}_{k + \\Delta t}\\)) at the beginning and end of the sampled event portion within each chunk, by employing a camera motion trajectory. The trajectory is represented through linear interpolation between two camera poses, one at the beginning of the chunk \\(\\mathbf{T}_{\\mathrm{start}} \\in \\mathbf{SE}(3)\\) and the other at the end \\(\\mathbf{T}_{\\mathrm{end}} \\in \\mathbf{SE}(3)\\). The camera pose at time \\(t_k\\) can thus be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.223, + 0.92, + 0.254 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {k} = \\mathbf {T} _ {\\text {s t a r t}} \\cdot \\exp \\left(\\frac {t _ {k} - t _ {\\text {s t a r t}}}{t _ {\\text {e n d}} - t _ {\\text {s t a r t}}} \\cdot \\log \\left(\\mathbf {T} _ {\\text {s t a r t}} ^ {- 1} \\cdot \\mathbf {T} _ {\\text {e n d}}\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.263, + 0.921, + 0.352 + ], + "angle": 0, + "content": "where \\( t_{start} \\) and \\( t_{end} \\) represent the timestamps corresponding to the boundary of the event chunk. It follows that \\( \\mathbf{T}_k \\) is differentiable with respect to both \\( \\mathbf{T}_{start} \\) and \\( \\mathbf{T}_{end} \\). The objective of IncEventGS is thus to estimate both \\( \\mathbf{T}_{start} \\) and \\( \\mathbf{T}_{end} \\) for each event chunk, along with the learnable parameters of 3D-GS \\( \\mathbf{G}_{\\theta} \\)." + }, + { + "type": "title", + "bbox": [ + 0.521, + 0.363, + 0.833, + 0.379 + ], + "angle": 0, + "content": "3.4. Incremental Tracking and Mapping" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.385, + 0.921, + 0.459 + ], + "angle": 0, + "content": "For both tracking and mapping, we aim to minimize the difference between the synthesized and measured brightness changes. In particular, we compute the loss of the latest event chunk only for the tracking stage and minimize the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.467, + 0.92, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {\\text {s t a r t}} ^ {*}, \\mathbf {T} _ {\\text {e n d}} ^ {*} = \\underset {\\mathbf {T} _ {\\text {s t a r t}}, \\mathbf {T} _ {\\text {e n d}}} {\\operatorname {a r g m i n}} \\left\\| \\mathbf {E} (\\mathbf {x}) - \\hat {\\mathbf {E}} (\\mathbf {x}) \\right\\| _ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.511, + 0.92, + 0.557 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{E}} (\\mathbf{x})\\) and \\(\\mathbf{E}(\\mathbf{x})\\) are the synthesized and measured brightness changes respectively, corresponding to a randomly sampled event portion within the latest event chunk." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.557, + 0.922, + 0.663 + ], + "angle": 0, + "content": "Once the tracking is done, we insert the latest event chunk to the mapper and exploit the estimated \\(\\mathbf{T}_{start}^{*}\\) and \\(\\mathbf{T}_{end}^{*}\\) as the initial value of the chunk to perform dense bundle adjustment. For computational consideration, we exploit a sliding window BA of the latest \\(n_w\\) chunks, and \\(n_w\\) is a hyperparameter. In particular, we optimize both the motion trajectories and the 3D-GS jointly by minimizing the following loss functions:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.675, + 0.92, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {\\text {e v e n t}} + \\lambda \\mathcal {L} _ {\\text {s s i m}}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.694, + 0.92, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e v e n t}} = \\left\\| \\mathbf {E} _ {i} (\\mathbf {x}) - \\hat {\\mathbf {E}} _ {i} (\\mathbf {x}) \\right\\| _ {2}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.724, + 0.92, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s s i m} = S S I M \\left(\\mathbf {E} _ {i} (\\mathbf {x}), \\hat {\\mathbf {E}} _ {i} (\\mathbf {x})\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.752, + 0.921, + 0.799 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter, SSIM is the structural dissimilarity loss [34]. As the event data streams in, we alternatively perform tracking and mapping." + }, + { + "type": "title", + "bbox": [ + 0.52, + 0.807, + 0.869, + 0.823 + ], + "angle": 0, + "content": "3.5. System Initialization and Boot-strapping" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.829, + 0.921, + 0.919 + ], + "angle": 0, + "content": "Conventional frame-based 3D-GS methods usually require a good initial point cloud and camera poses computed via COLMAP [27] for initialization. However, they are usually not easy to obtain if we are only given a single event camera. We therefore initialize the 3D-GS by sampling point cloud randomly within a bounding box. The first \\( m \\) event chunks" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.954, + 0.52, + 0.966 + ], + "angle": 0, + "content": "26936" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.492, + 0.168 + ], + "angle": 0, + "content": "(where \\( m \\) is a hyperparameter) are selected for initialization, and all corresponding camera poses are randomly initialized to be near the identity matrix. We then minimize the loss computed by Eq. (11) with respect to the attributes of 3D-GS and the parameters of camera motion trajectories jointly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.493, + 0.38 + ], + "angle": 0, + "content": "Through experiments, we found that the above initialization procedure consistently produces satisfactory brightness images. However, the 3D structure remains of low quality due to the short baselines of the moving event camera. We further find that it could potentially affect the performance of the whole pipeline without a good initial 3D structure as more event data is received. Therefore, we utilize a monocular depth estimation network [9] to predict a dense depth map from the rendered brightness image after the pipeline is trained for certain iterations. This depth map is then used to re-initialize the centers of the 3D Gaussians by unprojecting the pixel depths, after which we repeat the minimization of Eq. (11) for system bootstrapping. More details about system re-initialization can be found in our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.492, + 0.458 + ], + "angle": 0, + "content": "3D-GS Incrementally Growing. As the camera moves, new Gaussians is periodically introduced to cover newly explored regions. After tracking, we obtain an accurate camera pose estimate for each new event chunk. The centers of new Gaussians are determined by:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.469, + 0.49, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} = \\mathbf {T} \\cdot \\pi^ {- 1} \\left(\\mathbf {u}, d _ {u}\\right) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.491, + 0.606 + ], + "angle": 0, + "content": "where \\(\\mathbf{u} \\in \\mathbb{R}^2\\) is pixel coordinate in the image plane, \\(d_u\\) is depth of the 3D point \\(\\mathbf{p}\\) projecting onto the image plane, which is rendered by Eq. (5), \\(\\pi^{-1}\\) denotes camera inverse projection, \\(\\mathbf{T}\\) is the camera pose from tracking. To ensure that new Gaussians are only added in previously unmapped areas, a visibility mask is computed to guide the expansion of the Gaussian splattering process, as following:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.618, + 0.49, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} (p) = V < \\lambda_ {V}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.49, + 0.676 + ], + "angle": 0, + "content": "where \\(\\mathbf{V}\\) is the rendered alpha map and \\(\\lambda_V\\) is a hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.691, + 0.224, + 0.708 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.715, + 0.292, + 0.733 + ], + "angle": 0, + "content": "4.1. Experimental Setups." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.493, + 0.92 + ], + "angle": 0, + "content": "Implementation Details. All experiments are conducted on a desktop PC equipped with a 5.73GHz AMD Ryzen 9 7900x CPU and an NVIDIA RTX 3090 GPU. The first \\( m = 3 \\) event chunks are used for initialization. During the mapping stage, a sliding window size of \\( n_w = 20 \\) is employed for the bundle adjustment algorithm. The hyperparameters are set as follows: \\( \\lambda = 0.05 \\), \\( \\lambda_V = 0.8 \\), and \\( n_{seg} = 100 \\). For the synthetic dataset, \\( n_{low} = 400k \\) and \\( n_{up} = 500k \\), while for the real dataset, \\( n_{low} = 60k \\) and \\( n_{up} = 80k \\). Each event chunk has a time interval of \\( 50~\\mathrm{ms} \\). The learning rate of the camera poses is set to 1e-4 and that for the attributes of 3D-GS are set the same as the original 3D-GS work. The number" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.092, + 0.923, + 0.153 + ], + "angle": 0, + "content": "of optimization steps for initialization is 4500, and that for tracking and mapping are set to 200 and 1500 respectively. The contrast threshold \\( C \\) of the event camera is set to 0.1 for synthetic datasets and 0.2 for real datasets empirically." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.16, + 0.923, + 0.493 + ], + "angle": 0, + "content": "Baselines and Evaluation Metrics. To the best of our knowledge, there are no existing event-only NeRF or 3D-GS methods that do not rely on ground-truth poses, making direct comparisons challenging. Therefore, we conduct a thorough comparison of our method with several event-based NeRF approaches, including E-NeRF [14], EventNeRF [25], and Robust e-NeRF [17], as well as our custom implemented two-stage method (i.e. E2VID [24] + COLMAP [27] + 3DGS [11]). E-NeRF, EventNeRF, and Robust e-NeRF leverage implicit neural radiance fields for 3D scene representation, requiring ground-truth camera poses for accurate NeRF reconstruction. For E2VID + COLMAP + 3DGS, event data is first converted into brightness images using E2VID. The camera poses are then estimated from these images using COLMAP, and 3D-GS is trained with the generated images and poses. Both the quantitative and qualitative comparisons are performed on the synthetic dataset. Since there are no paired ground truth images for the real dataset, we only perform qualitative comparisons on the real dataset. In terms of motion trajectory evaluations, we use the publicly available state-of-the-art event-only visual odometry methods, i.e. DEVO (mono) [15] and ESVO2(stereo) [21], for comparison." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.496, + 0.922, + 0.618 + ], + "angle": 0, + "content": "The metrics used for novel view synthesis (NVS) include the commonly employed PSNR, SSIM, and LPIPS. For motion trajectory evaluations, we utilize Absolute Trajectory Error (ATE). To ensure fair comparisons, we employ the evaluation code provided by EventNeRF to compute the NVS metrics, which applies a linear color transformation between predictions and ground truth. Additionally, we use the public EVO toolbox [5] to compute the trajectory metrics." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.658, + 0.922, + 0.825 + ], + "angle": 0, + "content": "Benchmark Datasets. To properly evaluate the performance of NVS and motion trajectory estimation, we synthesized event data using the 3D scene models from the Replica dataset [28]. In particular, we exploit the room0, room2, office0, office2, and office3 scenes. We render high frame rate RGB images at \\(1000\\mathrm{Hz}\\) with a resolution of \\(768\\times 480\\) pixels. These images are then converted to grayscale, and the event data is generated via the events simulator [4]. The contrast threshold is set to 0.1. To simulate real-world camera motions, we exploit the same motion trajectories as that of NICE-SLAM [39] for data generation." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.829, + 0.922, + 0.92 + ], + "angle": 0, + "content": "We use the event dataset provided by TUM-VIE [13] for real data evaluations, which is also used by E-NeRF and Robust e-NeRF. TUM-VIE captures the event datasets by a pair of Prophesee Gen4 HD event cameras with a resolution of \\(1280 \\times 720\\) pixels. We only use the left-event camera data for our experiment." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.954, + 0.519, + 0.966 + ], + "angle": 0, + "content": "26937" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.088, + 0.905, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.573, + 0.92, + 0.601 + ], + "angle": 0, + "content": "Figure 2. Qualitative evaluation of novel view image synthesis on the Replica dataset. The experimental results demonstrate that our method renders higher-quality images with fewer artifacts compared to event-based NeRF and two-stage approaches." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.609, + 0.919, + 0.727 + ], + "angle": 0, + "content": "
room0room2office0office2office3
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
E-NeRF13.990.580.5115.560.470.5818.910.510.5713.050.650.4414.010.620.48
EventNeRF17.290.620.3916.020.540.6418.900.430.6215.180.660.4516.770.730.33
Robust e-NeRF17.260.840.1816.430.500.5218.930.520.5616.810.810.2519.220.840.18
E2VID+COLMAP+3DGS14.450.440.5215.740.510.5518.910.310.6814.030.570.4813.250.470.53
Ours24.310.850.1723.750.790.2325.640.540.3021.740.820.2321.180.880.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.737, + 0.92, + 0.765 + ], + "angle": 0, + "content": "Table 1. NVs performance comparison on Replica dataset. All results are calculated using the same code and ground-truth images. The results demonstrate that our method outperforms NeRF-based and two-stage methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.79, + 0.245, + 0.807 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.813, + 0.491, + 0.904 + ], + "angle": 0, + "content": "We conduct ablation studies to confirm our design choices. In particular, we study the effect of a monocular depth estimation network for system bootstrapping and event slicing hyperparameters \\( n_{low} \\), \\( n_{up} \\). The experiments are conducted with the Replica dataset, and the results are shown in Table 3 and Table 4, respectively." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.904, + 0.49, + 0.92 + ], + "angle": 0, + "content": "We found that depth initialization significantly impacts" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.791, + 0.922, + 0.882 + ], + "angle": 0, + "content": "pose estimation, reducing the Average Trajectory Error (ATE) from \\(1.534\\mathrm{cm}\\) to \\(0.064\\mathrm{cm}\\). Additionally, this improvement in pose estimation leads to a slight enhancement in Novel View Synthesis (NVS) performance. These results verify the importance of using depth initialization during the bootstrapping stage." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.889, + 0.922, + 0.92 + ], + "angle": 0, + "content": "We compare several combinations of hyperparameters \\( n_{\\mathrm{low}} \\) and \\( n_{\\mathrm{up}} \\), which refer to the range of event slicing window" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.954, + 0.521, + 0.966 + ], + "angle": 0, + "content": "26938" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.16, + 0.089, + 0.855, + 0.169 + ], + "angle": 0, + "content": "
room0room2office0office2office31d3d6dofdeskdesk2
DEVO0.2890.2660.1380.2810.1560.1470.3032.930.7320.201
E2VID+COLMAP17.9359.96105.1918.41417.284.26816.909.8821.5710.13
ESVO2-----0.3371.0660.5871.1472.506
Ours0.0460.0670.0450.0460.0540.1150.2980.2510.2310.129
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.179, + 0.921, + 0.208 + ], + "angle": 0, + "content": "Table 2. Pose accuracy (ATE, cm) on Replica and TUM-VIE datasets. The results demonstrate that our method delivers better performance in terms of camera motion estimation. Since ESVO2 is a stereo event camera method, it cannot run on the Replica dataset." + }, + { + "type": "table", + "bbox": [ + 0.15, + 0.231, + 0.432, + 0.287 + ], + "angle": 0, + "content": "
SettingPSNR↑SSIM↑LPIPS↓ATE
full21.740.820.230.046
w/o17.800.760.261.534
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.296, + 0.491, + 0.366 + ], + "angle": 0, + "content": "Table 3. Ablation Study about Depth Initialization. The unit of ATE is cm. The experimental results demonstrate the effectiveness of the initialization strategy. It not only improves the quality of rendered images, but also improves the accuracy of the camera motion estimation significantly." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.382, + 0.445, + 0.487 + ], + "angle": 0, + "content": "
SettingPSNR↑SSIM↑LPIPS↓ATE
1k-10k16.070.640.460.167
10k-50k18.410.720.330.079
80k-200k20.990.790.250.079
400k-500k21.740.820.230.046
500k-600k20.950.790.230.050
600k-700k18.060.750.280.214
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.497, + 0.493, + 0.526 + ], + "angle": 0, + "content": "Table 4. Ablation Study on Event Slice Window Size (Hyperparameters \\( n_{\\mathrm{low}} \\) and \\( n_{\\mathrm{up}} \\)). The unit of ATE is cm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.491, + 0.61 + ], + "angle": 0, + "content": "size. Table 4 demonstrates that both too small and too large window sizes negatively impact the performance of Novel View Synthesis (NVS) and pose estimation. Consequently, we select \\( n_{\\mathrm{low}} = 400k \\) and \\( n_{\\mathrm{up}} = 500k \\) for our experiments on the Replica dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.622, + 0.325, + 0.639 + ], + "angle": 0, + "content": "4.3. Quantitative Evaluations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.49, + 0.72 + ], + "angle": 0, + "content": "We conduct quantitative evaluations against event NeRF methods (E-NeRF, EventNeRF, and Robust e-NeRF) and our custom implemented two-stage method (i.e. E2VID + COLMAP + 3DGS) in terms of the quality of NVS and pose estimation performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.722, + 0.492, + 0.902 + ], + "angle": 0, + "content": "The NVS performance is evaluated on Replica-dataset and the results are presented in Table 1. It is important to note that the metrics are lower than those typically observed in standard NeRF/3D-GS methods for RGB images, primarily due to the lack of absolute brightness supervision. Even though NeRF-based methods use ground truth poses for training, IncEventGS still significantly outperforms them, highlighting the advantages of our approach utilizing a 3D Gaussian representation. Additionally, our method greatly surpasses two-stage method that also employs 3D Gaussian representation, demonstrating superior pose estimation and the effectiveness of our bundle adjustment technique." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.904, + 0.49, + 0.92 + ], + "angle": 0, + "content": "We evaluate pose estimation performance using the ATE" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.234, + 0.923, + 0.311 + ], + "angle": 0, + "content": "metric on both synthetic and real datasets, comparing our method with DEVO, ESVO2 and E2VID + COLMAP. The results, presented in Table 2, show that our method outperforms both baselines, validating the effectiveness of our incremental tracking and mapping technique." + }, + { + "type": "title", + "bbox": [ + 0.52, + 0.332, + 0.742, + 0.348 + ], + "angle": 0, + "content": "4.4. Qualitative Evaluations." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.357, + 0.923, + 0.616 + ], + "angle": 0, + "content": "We evaluate our method against event NeRF methods and two-stage method qualitatively in terms of novel view image synthesis, both on synthetic and real data. The results are presented in both Fig. 2 and Fig. 3. It demonstrates that our method can deliver better novel view images, while event NeRF methods and two-stage method render images with additional artifacts. Compared to NeRF-based methods, our approach demonstrates the advantage of IncEventGS by leveraging 3D Gaussian Splatting as the underlying scene representation. In contrast to two-stage method, our dense bundle adjustment optimizes both 3D Gaussian Splatting and camera pose using event data, whereas two-stage approaches tend to accumulate errors over time, as confirmed by the experimental results. We also provide representative visualization of ATE error mapped onto trajectories in Fig. 4, both on synthetic and real dataset. It demonstrates that IncEventGS is able to recover more accurate motion trajectories." + }, + { + "type": "title", + "bbox": [ + 0.521, + 0.64, + 0.64, + 0.657 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.67, + 0.922, + 0.852 + ], + "angle": 0, + "content": "We present the first pose-free 3D Gaussian splattering reconstruction model based on event camera, i.e. IncEventGS, and is useful for real world applications. We adopt the tracking and mapping paradigm in conventional SLAM pipeline to do incremental motion estimation and 3D scene reconstruction simultaneously. To handle the continuous and asynchronous characteristics of event stream, we exploit a continuous trajectory model to model the event data formation process. The experimental results on both synthetic and real datasets demonstrate the superior performance of IncEventGS over prior state-of-the-art methods in terms of high-quality novel image synthesis and camera pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.859, + 0.921, + 0.92 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by NSFC under Grant 62202389, in part by a grant from the Westlake University-Muyuan Joint Research Institute, and in part by the Westlake Education Foundation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.954, + 0.521, + 0.966 + ], + "angle": 0, + "content": "26939" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.084, + 0.905, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.522, + 0.92, + 0.565 + ], + "angle": 0, + "content": "Figure 3. Qualitative evaluation for novel view image synthesis on real dataset. It demonstrates that our method is able to render better images with fewer artifacts than event NeRF methods and two-stage methods. Note that there are no GT images aligned with the event camera, and we choose the closest images from the RGB camera and crop them to the same size as the rendered images for visual comparisons." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.565, + 0.902, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.881, + 0.92, + 0.91 + ], + "angle": 0, + "content": "Figure 4. Representative visualization of ATE error mapped onto trajectories for the synthetic (office0) and real (6 dof) datasets, generated by the EVO toolbox using the same ground truth poses, demonstrating the superior performance of our method in pose estimation." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.954, + 0.518, + 0.965 + ], + "angle": 0, + "content": "26940" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.492, + 0.17 + ], + "angle": 0, + "content": "[1] Hiroyuki Deguchi, Mana Masuda, Takuya Nakabayashi, and Hideo Saito. E2gs: Event enhanced gaussian splatting. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1676-1682. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.492, + 0.212 + ], + "angle": 0, + "content": "[2] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. arXiv preprint arXiv:2312.07504, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.492, + 0.267 + ], + "angle": 0, + "content": "[3] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth and optical flow estimation. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.268, + 0.492, + 0.336 + ], + "angle": 0, + "content": "[4] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.337, + 0.492, + 0.379 + ], + "angle": 0, + "content": "[5] Michael Grupp. evo: Python package for the evaluation of odometry and slam. https://github.com/MichaelGrupp/evo, 2017.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.379, + 0.492, + 0.447 + ], + "angle": 0, + "content": "[6] Jiarui Hu, Xianhao Chen, Boyin Feng, Guanglin Li, Liangjing Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field. arXiv preprint arXiv:2403.16095, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.448, + 0.492, + 0.517 + ], + "angle": 0, + "content": "[7] Huajian Huang, Longwei Li, Hui Cheng, and Sai-Kit Yeung. Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular, stereo, and rgb-d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.492, + 0.572 + ], + "angle": 0, + "content": "[8] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-nerf: Event based neural radiance field. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 837–847, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.573, + 0.492, + 0.642 + ], + "angle": 0, + "content": "[9] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.643, + 0.491, + 0.711 + ], + "angle": 0, + "content": "[10] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.712, + 0.492, + 0.766 + ], + "angle": 0, + "content": "[11] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.767, + 0.492, + 0.807 + ], + "angle": 0, + "content": "[12] Hanme Kim, Stefan Leutenegger, and Andrew J. Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In ECCV, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.808, + 0.492, + 0.864 + ], + "angle": 0, + "content": "[13] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. Tum-vie: The tum stereo visual-inertial event dataset. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8601-8608. IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.865, + 0.492, + 0.919 + ], + "angle": 0, + "content": "[14] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 8(3):1587-1594, 2023. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.115, + 0.492, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.922, + 0.147 + ], + "angle": 0, + "content": "[15] Simon Klenk, Marvin Motzet, Lukas Koestler, and Daniel Cremers. Deep event visual odometry. In 2024 International Conference on 3D Vision (3DV), pages 739-749. IEEE, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.149, + 0.921, + 0.206 + ], + "angle": 0, + "content": "[16] Wenpu Li, Pian Wan, Peng Wang, Jinghang Li, Yi Zhou, and Peidong Liu. Benerf: Neural radiance fields from a single blurry image and event stream. In European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.207, + 0.921, + 0.262 + ], + "angle": 0, + "content": "[17] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse and noisy events under non-uniform motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.264, + 0.921, + 0.32 + ], + "angle": 0, + "content": "[18] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.321, + 0.921, + 0.377 + ], + "angle": 0, + "content": "[19] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.379, + 0.921, + 0.42 + ], + "angle": 0, + "content": "[20] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE Transactions on Robotics, 33(5):1255-1262, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.422, + 0.921, + 0.476 + ], + "angle": 0, + "content": "[21] Junkai Niu, Sheng Zhong, Xiuyuan Lu, Shaojie Shen, Guillermo Gallego, and Yi Zhou. Esvo2: Direct visual-inertial odometry with stereo event cameras. IEEE Transactions on Robotics, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.479, + 0.921, + 0.535 + ], + "angle": 0, + "content": "[22] Delin Qu, Chi Yan, Dong Wang, Jie Yin, Dan Xu, Bin Zhao, and Xuelong Li. Implicit event-rgbd neural slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.536, + 0.921, + 0.591 + ], + "angle": 0, + "content": "[23] Henri Rebecq, Timo Horstschaefer, Guillermo Gallego, and Davide Scaramuzza. Evo: A geometric approach to event-based 6-dof parallel tracking and mapping in real time. IEEE Robotics and Automation Letters, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.593, + 0.921, + 0.649 + ], + "angle": 0, + "content": "[24] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.65, + 0.921, + 0.72 + ], + "angle": 0, + "content": "[25] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4992-5002, 2023. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.721, + 0.921, + 0.762 + ], + "angle": 0, + "content": "[26] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion Revisited. In CVPR, pages 4104-4113, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.765, + 0.921, + 0.807 + ], + "angle": 0, + "content": "[27] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.808, + 0.921, + 0.919 + ], + "angle": 0, + "content": "[28] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.922, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.954, + 0.518, + 0.966 + ], + "angle": 0, + "content": "26941" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.49, + 0.12 + ], + "angle": 0, + "content": "dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.49, + 0.163 + ], + "angle": 0, + "content": "[29] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.49, + 0.205 + ], + "angle": 0, + "content": "[30] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. Advances in Neural Information Processing Systems, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.49, + 0.262 + ], + "angle": 0, + "content": "[31] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. arXiv preprint arXiv:2405.14959, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.49, + 0.319 + ], + "angle": 0, + "content": "[32] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4170–4179, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.49, + 0.375 + ], + "angle": 0, + "content": "[33] Peng Wang, Lingzhe Zhao, Yin Zhang, Shiyu Zhao, and Peidong Liu. Mba-slam: Motion blur aware dense visual slam with radiance fields representation. arXiv preprint arXiv:2411.08279, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.49, + 0.432 + ], + "angle": 0, + "content": "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.49, + 0.489 + ], + "angle": 0, + "content": "[35] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher A Metzler. Event3dgs: Event-based 3d gaussian splattering for fast egomotion. arXiv preprint arXiv:2406.02972, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.49, + 0.545 + ], + "angle": 0, + "content": "[36] Chi Yan, Delin Qu, Dong Wang, Dan Xu, Zhigang Wang, Bin Zhao, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.49, + 0.589 + ], + "angle": 0, + "content": "[37] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.49, + 0.631 + ], + "angle": 0, + "content": "[38] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. In European Conference on Computer Vision (ECCV), 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.49, + 0.702 + ], + "angle": 0, + "content": "[39] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.49, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.955, + 0.519, + 0.966 + ], + "angle": 0, + "content": "26942" + } + ] +] \ No newline at end of file diff --git a/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/c5e31ebc-2f5f-4333-a4a5-98739dcdbce4_origin.pdf b/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/c5e31ebc-2f5f-4333-a4a5-98739dcdbce4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1e826137b775312989555ec79a8720c1a0276c18 --- /dev/null +++ b/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/c5e31ebc-2f5f-4333-a4a5-98739dcdbce4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cda3fb8256eeccda8fa80ca108ed19f3d389fa568c4e998a146ae56efde5821f +size 9757826 diff --git a/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/full.md b/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5ff46597db1b1465abec76194e20aaecb25b6bc3 --- /dev/null +++ b/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/full.md @@ -0,0 +1,290 @@ +# IncEventGS: Pose-Free Gaussian Splating from a Single Event Camera + +Jian Huang $^{1,2}$ Chengrui Dong $^{1,2}$ Xuanhua Chen $^{2,3}$ Peidong Liu $^{2*}$ + +$^{1}$ Zhejiang University $^{2}$ Westlake University $^{3}$ Northeastern University $\{$ huangjian39, dongchengrui, chenxuanhua, liupeidong\}@westlake.edu.cn https://github.com/WU-CVGL/IncEventGS + +# Abstract + +Implicit neural representation and explicit 3D Gaussian Splating (3D-GS) for novel view synthesis have achieved remarkable progress with frame-based camera (e.g. RGB and RGB-D cameras) recently. Compared to frame-based camera, a novel type of bio-inspired visual sensor, i.e. event camera, has demonstrated advantages in high temporal resolution, high dynamic range, low power consumption, and low latency, which make it favored for many robotic applications. In this work, we present IncEventGS, an incremental 3D Gaussian Splating reconstruction algorithm with a single event camera, without the assumption of known camera poses. To recover the 3D scene representation incrementally, we exploit the tracking and mapping paradigm of conventional SLAM pipelines for IncEventGS. Given the incoming event stream, the tracker first estimates an initial camera motion based on prior reconstructed 3D-GS scene representation. The mapper then jointly refines both the 3D scene representation and camera motion based on the previously estimated motion trajectory from the tracker. The experimental results demonstrate that IncEventGS delivers superior performance compared to prior NeRF-based methods and other related baselines, even if we do not have the ground-truth camera poses. Furthermore, our method can also deliver better performance compared to state-of-the-art event visual odometry methods in terms of camera motion estimation. + +# 1. Introduction + +Reconstructing accurate 3D scene representations from 2D images has been a long-standing challenge in computer vision and robotics, driving substantial efforts over the past few decades. Among those pioneering works, Neural Radiance Fields (NeRF) [19] and 3D Gaussian Splatting (3D-GS) [11], stand out for their utilization of differentiable rendering techniques, and have garnered significant attention due to their capability to recover high-quality 3D scene representation + +from 2D images. Commonly used sensors for 3D scene reconstruction are usually frame-based cameras, such as the RGB and RGB-D cameras. They usually capture full-brightness intensity images within a short exposure time at a regular frequency. Due to the characteristic of this data-capturing process, they often suffer from motion blur or fail to capture accurate and informative intensity information under fast motion and low-light conditions, which would further affect the performance of downstream applications. + +The event camera, a bio-inspired sensor, has gained significant attention in recent years for its potential to address the limitations of frame-based cameras under challenging conditions. Unlike conventional cameras, event cameras record brightness changes asynchronously at each pixel, emitting events when a predefined threshold is surpassed. This unique operation offers several advantages over conventional cameras, in terms of high temporal resolution, high dynamic range, low latency, and power consumption. Although event cameras have attractive characteristics for challenging environments, they cannot be directly integrated into existing frame-based 3D reconstruction algorithms that rely on processing dense 2D brightness intensity images. + +Several pioneering works have been proposed to exploit event stream [3, 12, 23] to recover the motion trajectory and scene representation. While existing methods deliver impressive performance, they usually exploit 2.5D semi-dense depth maps to represent the 3D scene, and bundle adjustment (BA) is hardly performed, due to the asynchronous and sparse characteristics of event data stream. Klenk et al. [15] recently proposed to convert event stream into event voxel grids, and then adapt a previous frame-based deep visual odometry pipeline [30] for accurate camera motion estimation. As NeRF exhibited impressive scene representation capability recently, several works [8, 14, 17, 25] explore to recover the underlying dense 3D scene NeRF representation from event stream, by assuming ground-truth poses are available. + +In contrast to those works, we propose IncEventGS, an incremental dense 3D scene reconstruction algorithm from a single event camera, by exploiting Gaussian Splatting as the underlying scene representation. Different from prior event- + +based NeRF reconstruction methods, IncEventGS does not require any ground-truth camera poses, which is more challenging and provides more flexibility for real-world robotic application scenarios. To overcome the challenges brought by unknown poses, IncEventGS adopts the tracking and mapping paradigm of conventional SLAM pipelines [20]. In particular, IncEventGS exploits prior explored and reconstructed 3D scenes for camera motion estimation of incoming event stream during the tracking stage. Both the 3D-GS scene representation and camera motions are then jointly optimized (i.e. event-based bundle adjustment) during the mapping stage, for more accurate scene representation and motion estimation. The 3D scene is progressively expanded and densified. The experimental results on both synthetic and real datasets demonstrate that IncEventGS can recover the underlying 3D scene representation and camera motion trajectory accurately. In particular, IncEventGS outperforms prior NeRF-based methods and other related baselines in terms of scene representation recovery, even without ground-truth poses. Furthermore, our method also delivers better camera motion estimation accuracy than the most recent state-of-the-art visual odometry algorithm, in terms of the Absolute Trajectory Error (ATE) metric. The recovered 3D scene representation can be further used to render novel brightness images. Our main contributions can be summarized as follows: + +- We present an incremental 3D Gaussian Splitting reconstruction algorithm from a single event camera, without requiring the ground-truth camera poses. +- We propose a novel initialization strategy tailored to the event data stream, which is vital to the success of the algorithm. +- The experimental results on both the synthetic and real datasets demonstrate the superior performance of our method over prior NeRF-based methods and related baselines in terms of novel view synthesis and better performance over state-of-the-art event-based visual odometry algorithm in terms of camera motion estimation. + +# 2. Related Works + +We review two main areas of prior works: event-based neural radiance fields and 3D Gaussian Splatting, which are the most related to our work. + +Event-based Neural Radiance Fields. Prior works [8, 14, 25] propose to exploit event stream to recover the neural radiance fields with known camera motion trajectory. Low et al. [17] further improves the reconstruction algorithm to handle sparse and noisy events under non-uniform motion. The recovered neural radiance fields can then be used to render novel view brightness images. The ground-truth poses are usually computed from corresponding brightness images via COLMAP [26] or provided by the indoor motion-capturing system. Recently, Qu et al. [22] proposed to integrate event measurements into an RGB-D implicit neural SLAM framework and achieved robust performance in mo + +tion blur scenarios. Li et al. [16] also proposed exploiting event measurements and a single blurry image to recover the underlying neural 3D scene representation. In contrast to those works, IncEventGS conducts incremental 3D scene reconstruction without requiring any prior ground-truth poses, which is more challenging and provides more flexibility for practical robotic application scenarios. The method further exploits 3D Gaussian Splatting as the underlying scene representation, which demonstrates better image rendering quality and efficiency, compared to the NeRF-based representation. + +3D Gaussian Splatting. 3D Gaussian Splatting [11] proposes a novel explicit 3D representation to further improve both the training and rendering efficiency compared to Neural Radiance Fields. Due to its impressive efficient scene representation capability, several pioneering works have been proposed to exploit 3D-GS for incremental 3D reconstruction. For example, Keetha et al. [10] propose an RGBD-based 3D-GS SLAM, employing an online tracking and mapping system tailored to the underlying Gaussian representation. Yan et al. [36] implement a coarse-to-fine camera tracking approach based on the sparse selection of Gaussians. Matsuki et al. [18] propose to apply 3D Gaussian Splatting to do incremental 3D reconstruction using a single moving monocular or RGB-D camera. Huang et al. [7] exploit ORB-SLAM3 to compute accurate camera poses and feed it into a 3D-GS algorithm for dense mapping. Fu et al. [2] use monocular depth estimation with 3D-GS. Yugay et al. [37] combine DROID-SLAM [29] based camera tracking with active and inactive 3D-GS sub-maps. Wang et al. [33] integrate bundle-adjustment and 3DGS [38] to estimated camera trajectory within the exposure time. Hu et al. [6] propose a novel depth uncertainty model to ensure the selection of valuable Gaussian primitives during optimization. While those methods deliver impressive performance in terms of 3D scene recovery and motion estimation, they usually assume the usage of frame-based images (i.e. either RGB or RGB-D date). On the contrary, we propose to exploit pure event measurements for incremental 3D-GS reconstruction. Several concurrent studies have recently explored the use of 3D Gaussians for event-based reconstruction, including EvGGS [31], Event3DGS [35], and E2GS [1]. EvGGS and Event3DGS rely solely on event data, whereas E2GS incorporates both event data and blurry images. However, the key difference between these methods and ours is that they all rely on ground-truth poses. + +# 3. Method + +The overview of our IncEventGS is shown in Fig. 1. Given only a single event camera, IncEventGS incrementally performs tracking and dense mapping under the framework of 3D Gaussian Splatting, to recover both the camera motion trajectory and 3D scene representation simultaneously. Since event data are asynchronous, they cannot be directly integrated with the 3D-GS representation. We therefore process the event data stream into chunks according to a fixed time window. + +![](images/3d633c7e1ccb14850406539ba2de87c11d49c5b276f733782eece3bafe0e1918.jpg) +Figure 1. The pipeline of IncEventGS. IncEventGS processes incoming event stream by dividing it into chunks and representing the camera trajectory as a continuous model. It randomly samples two close consecutive timestamps to integrate the corresponding event streams. Two brightness images are rendered from 3D-GS at the corresponding poses, and we minimize the photometric loss between the synthesized and measured brightness change. During initialization, a pre-trained depth estimation model estimates depth from the rendered images to bootstrap the system. + +We associate each chunk with a continuous time trajectory parameterization in the $\mathfrak{se}(3)$ space. Two close consecutive timestamps (i.e., $t_k$ and $t_{k + \Delta t}$ , where $\Delta t$ is a small time interval) can be randomly sampled, and the measured brightness change $\mathbf{E}(x)$ can be computed from the corresponding event stream. Based on the parameterized trajectory, the corresponding camera poses (i.e., $\mathbf{T}_k$ , $\mathbf{T}_{k + \Delta t}$ ) can be determined and the brightness images (i.e., $\bar{\mathbf{I}}_k$ , $\bar{\mathbf{I}}_{k + \Delta t}$ ) can be further rendered from the 3D-GS. The synthesized brightness change $\hat{\mathbf{E}}(x)$ can be computed for event loss computation. + +During the tracking stage, we optimize only the camera motion trajectory of the newly accumulated event chunk and exploit the recovered trajectory to initialize the dense bundle adjustment (BA) algorithm for the mapping stage. During the mapping stage, we continuously densify 3D Gaussians for newly explored areas and prune transparent 3D Gaussians. For computational efficiency, we exploit a sliding window of the latest chunks and perform BA only within this window for both 3D-GS reconstruction and motion trajectory estimation. We will detail each component as follows. + +# 3.1. 3D Scene Representation + +Following 3D-GS [11], the scene is represented by a set of 3D Gaussian primitives, each of which contains mean position $\pmb{\mu} \in \mathbb{R}^3$ in the world coordinate, 3D covariance $\pmb{\Sigma} \in \mathbb{R}^{3 \times 3}$ , opacity $\mathbf{o} \in \mathbb{R}$ , and color $\mathbf{c} \in \mathbb{R}^3$ . To ensure that the covariance matrix remains positive semi-definite throughout the gradient descent, the covariance $\pmb{\Sigma}$ is parameterized using a + +scale vector $\mathbf{s} \in \mathbb{R}^3$ and rotation matrix $\mathbf{R} \in \mathbb{R}^{3 \times 3}$ : + +$$ +\boldsymbol {\Sigma} = \mathbf {R} \mathbf {S} \mathbf {S} ^ {T} \mathbf {R} ^ {T}, \tag {1} +$$ + +where scale matrix $\mathbf{S} = \text{diag}([s])$ is derived from the scale vector $\mathbf{s} \in \mathbb{R}^3$ . + +In order to enable rendering, 3D-GS projects 3D Gaussian primitives to the 2D image plane from a given camera pose $\mathbf{T}_c = \{\mathbf{R}_c\in \mathbb{R}^{3\times 3},\mathbf{t}_c\in \mathbb{R}^3\}$ using following equation: + +$$ +\boldsymbol {\Sigma} ^ {\prime} = \mathbf {J} \mathbf {R} _ {c} \boldsymbol {\Sigma} \mathbf {R} _ {c} ^ {T} \mathbf {J} ^ {T}, \tag {2} +$$ + +where $\pmb{\Sigma}^{\prime}\in \mathbb{R}^{2\times 2}$ is the 2D covariance matrix, $J\in \mathbb{R}^{2\times 3}$ is the Jacobian of the affine approximation of the projective transformation. After projecting 3D Gaussians onto the image plane, the color of each pixel is determined by sorting the Gaussians according to their depth and then applying near-to-far $\alpha$ -blending rendering via the following equation: + +$$ +\mathbf {I} = \sum_ {i} ^ {N} \mathbf {c} _ {i} \alpha_ {i} \prod_ {j} ^ {i - 1} (1 - \alpha_ {j}), \tag {3} +$$ + +where $\mathbf{c}_i$ is the learnable color of each Gaussian, and $\alpha_{i}$ is the alpha value computed by evaluating the 2D covariance $\boldsymbol{\Sigma}^{\prime}$ multiplied with the learned Gaussian opacity $\mathbf{o}$ : + +$$ +\alpha_ {i} = \mathbf {o} _ {i} \cdot \exp (- \sigma_ {i}), \quad \sigma_ {i} = \frac {1}{2} \Delta_ {i} ^ {T} \boldsymbol {\Sigma} ^ {\prime - 1} \Delta_ {i}, \tag {4} +$$ + +where $\Delta_i\in \mathbb{R}^2$ is the offset between the pixel center and the + +2D Gaussian center. Depth is rendered by: + +$$ +\mathbf {D} = \sum_ {i} ^ {N} \mathbf {d} _ {i} \alpha_ {i} \prod_ {j} ^ {i - 1} (1 - \alpha_ {j}), \tag {5} +$$ + +where $d_{i}$ denotes the z-depth of the center of the i-th 3D Gaussian to the camera. We also render alpha map to determine visibility: + +$$ +\mathbf {V} = \sum_ {i} ^ {N} \alpha_ {i} \prod_ {j} ^ {i - 1} (1 - \alpha_ {j}), \tag {6} +$$ + +The derivations presented above demonstrate that the rendered pixel color, denoted as I in Eq. (3), is a function that is differentiable with respect to the learnable attributes of 3D-GS, and the camera poses $\mathbf{T}_c$ . This facilitates our bundle adjustment formulation, accommodating a set of event chunks and inaccurate camera motion trajectories within the framework of 3D-GS. + +# 3.2.Event Data Formation Model + +An event camera records changes in brightness as a stream of events asynchronously. To relate 3D-GS representation with the event stream, we sample two close consecutive timestamps (i.e., $t_k$ and $t_{k + \Delta t}$ , where $\Delta t$ is a small time interval), and the measured brightness change between $\Delta t$ is: + +$$ +\mathbf {E} (\mathbf {x}) = C \left\{e _ {i} \left(\mathbf {x}, t _ {i}, p _ {i}\right) \right\} _ {t _ {k} < t _ {i} < t _ {k} + \Delta t}, \tag {7} +$$ + +where $e(\mathbf{x}, t_i, p_i)$ is the $i^{th}$ event within the defined time interval corresponding to pixel $\mathbf{x}$ and $C$ is the fixed contrast threshold. The corresponding camera poses $T_k$ and $T_{k + \Delta t}$ can be interpolated from the camera motion trajectory parameterization, and two corresponding brightness images (i.e. $\hat{\mathbf{I}}_k$ and $\hat{\mathbf{I}}_{k + \Delta t}$ ) can be rendered from 3D-GS. The synthesized brightness change $\hat{\mathbf{E}}$ is modelled as: + +$$ +\hat {\mathbf {E}} (\mathbf {x}) = \log \left(\hat {\mathbf {I}} _ {k + \Delta t} (\mathbf {x})\right) - \log \left(\hat {\mathbf {I}} _ {k} (\mathbf {x})\right), \tag {8} +$$ + +where $\hat{\mathbf{E}} (\mathbf{x})$ depends on the parameters of both the motion trajectory parameters and 3D-GS, and is differentiable with respect to them. + +Both in tracking and mapping, inspired by the work of [25], we segment the current event chunks into $n_{seg}$ equal segments according to the number of events, obtaining $n_{seg}$ timestamps that correspond to the end of each segment. We then randomly select one timestamp from these $n_{seg}$ timestamps to serve as $t_{k + \Delta t}$ , and we randomly sample an integer $n_{win}$ between the integer bounds $n_{low}$ and $n_{up}$ . The index of $t_k$ is equal to the index of $t_{k + \Delta t}$ subtract $n_{win}$ . $n_{seg}$ , $n_{low}$ and $n_{up}$ are hyperparameters. This sampling strategy enables the model to capture both local and global information. + +# 3.3. Camera Motion Trajectory Modeling + +Since each event chunk usually contains too many events, we sample a portion of them according to the total number of + +events during optimization. Following [32, 38], we formulate the corresponding poses (i.e. $\mathbf{T}_k$ and $\mathbf{T}_{k + \Delta t}$ ) at the beginning and end of the sampled event portion within each chunk, by employing a camera motion trajectory. The trajectory is represented through linear interpolation between two camera poses, one at the beginning of the chunk $\mathbf{T}_{\mathrm{start}} \in \mathbf{SE}(3)$ and the other at the end $\mathbf{T}_{\mathrm{end}} \in \mathbf{SE}(3)$ . The camera pose at time $t_k$ can thus be expressed as follows: + +$$ +\mathbf {T} _ {k} = \mathbf {T} _ {\text {s t a r t}} \cdot \exp \left(\frac {t _ {k} - t _ {\text {s t a r t}}}{t _ {\text {e n d}} - t _ {\text {s t a r t}}} \cdot \log \left(\mathbf {T} _ {\text {s t a r t}} ^ {- 1} \cdot \mathbf {T} _ {\text {e n d}}\right)\right), \tag {9} +$$ + +where $t_{start}$ and $t_{end}$ represent the timestamps corresponding to the boundary of the event chunk. It follows that $\mathbf{T}_k$ is differentiable with respect to both $\mathbf{T}_{start}$ and $\mathbf{T}_{end}$ . The objective of IncEventGS is thus to estimate both $\mathbf{T}_{start}$ and $\mathbf{T}_{end}$ for each event chunk, along with the learnable parameters of 3D-GS $\mathbf{G}_{\theta}$ . + +# 3.4. Incremental Tracking and Mapping + +For both tracking and mapping, we aim to minimize the difference between the synthesized and measured brightness changes. In particular, we compute the loss of the latest event chunk only for the tracking stage and minimize the following loss function: + +$$ +\mathbf {T} _ {\text {s t a r t}} ^ {*}, \mathbf {T} _ {\text {e n d}} ^ {*} = \underset {\mathbf {T} _ {\text {s t a r t}}, \mathbf {T} _ {\text {e n d}}} {\operatorname {a r g m i n}} \left\| \mathbf {E} (\mathbf {x}) - \hat {\mathbf {E}} (\mathbf {x}) \right\| _ {2}, \tag {10} +$$ + +where $\hat{\mathbf{E}} (\mathbf{x})$ and $\mathbf{E}(\mathbf{x})$ are the synthesized and measured brightness changes respectively, corresponding to a randomly sampled event portion within the latest event chunk. + +Once the tracking is done, we insert the latest event chunk to the mapper and exploit the estimated $\mathbf{T}_{start}^{*}$ and $\mathbf{T}_{end}^{*}$ as the initial value of the chunk to perform dense bundle adjustment. For computational consideration, we exploit a sliding window BA of the latest $n_w$ chunks, and $n_w$ is a hyperparameter. In particular, we optimize both the motion trajectories and the 3D-GS jointly by minimizing the following loss functions: + +$$ +\mathcal {L} = (1 - \lambda) \mathcal {L} _ {\text {e v e n t}} + \lambda \mathcal {L} _ {\text {s s i m}}, \tag {11} +$$ + +$$ +\mathcal {L} _ {\text {e v e n t}} = \left\| \mathbf {E} _ {i} (\mathbf {x}) - \hat {\mathbf {E}} _ {i} (\mathbf {x}) \right\| _ {2}, \tag {12} +$$ + +$$ +\mathcal {L} _ {s s i m} = S S I M \left(\mathbf {E} _ {i} (\mathbf {x}), \hat {\mathbf {E}} _ {i} (\mathbf {x})\right), \tag {13} +$$ + +where $\lambda$ is a hyperparameter, SSIM is the structural dissimilarity loss [34]. As the event data streams in, we alternatively perform tracking and mapping. + +# 3.5. System Initialization and Boot-strapping + +Conventional frame-based 3D-GS methods usually require a good initial point cloud and camera poses computed via COLMAP [27] for initialization. However, they are usually not easy to obtain if we are only given a single event camera. We therefore initialize the 3D-GS by sampling point cloud randomly within a bounding box. The first $m$ event chunks + +(where $m$ is a hyperparameter) are selected for initialization, and all corresponding camera poses are randomly initialized to be near the identity matrix. We then minimize the loss computed by Eq. (11) with respect to the attributes of 3D-GS and the parameters of camera motion trajectories jointly. + +Through experiments, we found that the above initialization procedure consistently produces satisfactory brightness images. However, the 3D structure remains of low quality due to the short baselines of the moving event camera. We further find that it could potentially affect the performance of the whole pipeline without a good initial 3D structure as more event data is received. Therefore, we utilize a monocular depth estimation network [9] to predict a dense depth map from the rendered brightness image after the pipeline is trained for certain iterations. This depth map is then used to re-initialize the centers of the 3D Gaussians by unprojecting the pixel depths, after which we repeat the minimization of Eq. (11) for system bootstrapping. More details about system re-initialization can be found in our supplementary material. + +3D-GS Incrementally Growing. As the camera moves, new Gaussians is periodically introduced to cover newly explored regions. After tracking, we obtain an accurate camera pose estimate for each new event chunk. The centers of new Gaussians are determined by: + +$$ +\mathbf {p} = \mathbf {T} \cdot \pi^ {- 1} \left(\mathbf {u}, d _ {u}\right) \tag {14} +$$ + +where $\mathbf{u} \in \mathbb{R}^2$ is pixel coordinate in the image plane, $d_u$ is depth of the 3D point $\mathbf{p}$ projecting onto the image plane, which is rendered by Eq. (5), $\pi^{-1}$ denotes camera inverse projection, $\mathbf{T}$ is the camera pose from tracking. To ensure that new Gaussians are only added in previously unmapped areas, a visibility mask is computed to guide the expansion of the Gaussian splattering process, as following: + +$$ +\mathbf {M} (p) = V < \lambda_ {V}, \tag {15} +$$ + +where $\mathbf{V}$ is the rendered alpha map and $\lambda_V$ is a hyperparameter. + +# 4. Experiments + +# 4.1. Experimental Setups. + +Implementation Details. All experiments are conducted on a desktop PC equipped with a 5.73GHz AMD Ryzen 9 7900x CPU and an NVIDIA RTX 3090 GPU. The first $m = 3$ event chunks are used for initialization. During the mapping stage, a sliding window size of $n_w = 20$ is employed for the bundle adjustment algorithm. The hyperparameters are set as follows: $\lambda = 0.05$ , $\lambda_V = 0.8$ , and $n_{seg} = 100$ . For the synthetic dataset, $n_{low} = 400k$ and $n_{up} = 500k$ , while for the real dataset, $n_{low} = 60k$ and $n_{up} = 80k$ . Each event chunk has a time interval of $50~\mathrm{ms}$ . The learning rate of the camera poses is set to 1e-4 and that for the attributes of 3D-GS are set the same as the original 3D-GS work. The number + +of optimization steps for initialization is 4500, and that for tracking and mapping are set to 200 and 1500 respectively. The contrast threshold $C$ of the event camera is set to 0.1 for synthetic datasets and 0.2 for real datasets empirically. + +Baselines and Evaluation Metrics. To the best of our knowledge, there are no existing event-only NeRF or 3D-GS methods that do not rely on ground-truth poses, making direct comparisons challenging. Therefore, we conduct a thorough comparison of our method with several event-based NeRF approaches, including E-NeRF [14], EventNeRF [25], and Robust e-NeRF [17], as well as our custom implemented two-stage method (i.e. E2VID [24] + COLMAP [27] + 3DGS [11]). E-NeRF, EventNeRF, and Robust e-NeRF leverage implicit neural radiance fields for 3D scene representation, requiring ground-truth camera poses for accurate NeRF reconstruction. For E2VID + COLMAP + 3DGS, event data is first converted into brightness images using E2VID. The camera poses are then estimated from these images using COLMAP, and 3D-GS is trained with the generated images and poses. Both the quantitative and qualitative comparisons are performed on the synthetic dataset. Since there are no paired ground truth images for the real dataset, we only perform qualitative comparisons on the real dataset. In terms of motion trajectory evaluations, we use the publicly available state-of-the-art event-only visual odometry methods, i.e. DEVO (mono) [15] and ESVO2(stereo) [21], for comparison. + +The metrics used for novel view synthesis (NVS) include the commonly employed PSNR, SSIM, and LPIPS. For motion trajectory evaluations, we utilize Absolute Trajectory Error (ATE). To ensure fair comparisons, we employ the evaluation code provided by EventNeRF to compute the NVS metrics, which applies a linear color transformation between predictions and ground truth. Additionally, we use the public EVO toolbox [5] to compute the trajectory metrics. + +Benchmark Datasets. To properly evaluate the performance of NVS and motion trajectory estimation, we synthesized event data using the 3D scene models from the Replica dataset [28]. In particular, we exploit the room0, room2, office0, office2, and office3 scenes. We render high frame rate RGB images at $1000\mathrm{Hz}$ with a resolution of $768\times 480$ pixels. These images are then converted to grayscale, and the event data is generated via the events simulator [4]. The contrast threshold is set to 0.1. To simulate real-world camera motions, we exploit the same motion trajectories as that of NICE-SLAM [39] for data generation. + +We use the event dataset provided by TUM-VIE [13] for real data evaluations, which is also used by E-NeRF and Robust e-NeRF. TUM-VIE captures the event datasets by a pair of Prophesee Gen4 HD event cameras with a resolution of $1280 \times 720$ pixels. We only use the left-event camera data for our experiment. + +![](images/ec53771c8d5d065b90240b3c1440f322de20fa4c99f213696b8ee4ff704a4364.jpg) +Figure 2. Qualitative evaluation of novel view image synthesis on the Replica dataset. The experimental results demonstrate that our method renders higher-quality images with fewer artifacts compared to event-based NeRF and two-stage approaches. + +
room0room2office0office2office3
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
E-NeRF13.990.580.5115.560.470.5818.910.510.5713.050.650.4414.010.620.48
EventNeRF17.290.620.3916.020.540.6418.900.430.6215.180.660.4516.770.730.33
Robust e-NeRF17.260.840.1816.430.500.5218.930.520.5616.810.810.2519.220.840.18
E2VID+COLMAP+3DGS14.450.440.5215.740.510.5518.910.310.6814.030.570.4813.250.470.53
Ours24.310.850.1723.750.790.2325.640.540.3021.740.820.2321.180.880.13
+ +Table 1. NVs performance comparison on Replica dataset. All results are calculated using the same code and ground-truth images. The results demonstrate that our method outperforms NeRF-based and two-stage methods. + +# 4.2. Ablation Study + +We conduct ablation studies to confirm our design choices. In particular, we study the effect of a monocular depth estimation network for system bootstrapping and event slicing hyperparameters $n_{low}$ , $n_{up}$ . The experiments are conducted with the Replica dataset, and the results are shown in Table 3 and Table 4, respectively. + +We found that depth initialization significantly impacts + +pose estimation, reducing the Average Trajectory Error (ATE) from $1.534\mathrm{cm}$ to $0.064\mathrm{cm}$ . Additionally, this improvement in pose estimation leads to a slight enhancement in Novel View Synthesis (NVS) performance. These results verify the importance of using depth initialization during the bootstrapping stage. + +We compare several combinations of hyperparameters $n_{\mathrm{low}}$ and $n_{\mathrm{up}}$ , which refer to the range of event slicing window + +
room0room2office0office2office31d3d6dofdeskdesk2
DEVO0.2890.2660.1380.2810.1560.1470.3032.930.7320.201
E2VID+COLMAP17.9359.96105.1918.41417.284.26816.909.8821.5710.13
ESVO2-----0.3371.0660.5871.1472.506
Ours0.0460.0670.0450.0460.0540.1150.2980.2510.2310.129
+ +Table 2. Pose accuracy (ATE, cm) on Replica and TUM-VIE datasets. The results demonstrate that our method delivers better performance in terms of camera motion estimation. Since ESVO2 is a stereo event camera method, it cannot run on the Replica dataset. + +
SettingPSNR↑SSIM↑LPIPS↓ATE
full21.740.820.230.046
w/o17.800.760.261.534
+ +Table 3. Ablation Study about Depth Initialization. The unit of ATE is cm. The experimental results demonstrate the effectiveness of the initialization strategy. It not only improves the quality of rendered images, but also improves the accuracy of the camera motion estimation significantly. + +
SettingPSNR↑SSIM↑LPIPS↓ATE
1k-10k16.070.640.460.167
10k-50k18.410.720.330.079
80k-200k20.990.790.250.079
400k-500k21.740.820.230.046
500k-600k20.950.790.230.050
600k-700k18.060.750.280.214
+ +Table 4. Ablation Study on Event Slice Window Size (Hyperparameters $n_{\mathrm{low}}$ and $n_{\mathrm{up}}$ ). The unit of ATE is cm. + +size. Table 4 demonstrates that both too small and too large window sizes negatively impact the performance of Novel View Synthesis (NVS) and pose estimation. Consequently, we select $n_{\mathrm{low}} = 400k$ and $n_{\mathrm{up}} = 500k$ for our experiments on the Replica dataset. + +# 4.3. Quantitative Evaluations. + +We conduct quantitative evaluations against event NeRF methods (E-NeRF, EventNeRF, and Robust e-NeRF) and our custom implemented two-stage method (i.e. E2VID + COLMAP + 3DGS) in terms of the quality of NVS and pose estimation performance. + +The NVS performance is evaluated on Replica-dataset and the results are presented in Table 1. It is important to note that the metrics are lower than those typically observed in standard NeRF/3D-GS methods for RGB images, primarily due to the lack of absolute brightness supervision. Even though NeRF-based methods use ground truth poses for training, IncEventGS still significantly outperforms them, highlighting the advantages of our approach utilizing a 3D Gaussian representation. Additionally, our method greatly surpasses two-stage method that also employs 3D Gaussian representation, demonstrating superior pose estimation and the effectiveness of our bundle adjustment technique. + +We evaluate pose estimation performance using the ATE + +metric on both synthetic and real datasets, comparing our method with DEVO, ESVO2 and E2VID + COLMAP. The results, presented in Table 2, show that our method outperforms both baselines, validating the effectiveness of our incremental tracking and mapping technique. + +# 4.4. Qualitative Evaluations. + +We evaluate our method against event NeRF methods and two-stage method qualitatively in terms of novel view image synthesis, both on synthetic and real data. The results are presented in both Fig. 2 and Fig. 3. It demonstrates that our method can deliver better novel view images, while event NeRF methods and two-stage method render images with additional artifacts. Compared to NeRF-based methods, our approach demonstrates the advantage of IncEventGS by leveraging 3D Gaussian Splatting as the underlying scene representation. In contrast to two-stage method, our dense bundle adjustment optimizes both 3D Gaussian Splatting and camera pose using event data, whereas two-stage approaches tend to accumulate errors over time, as confirmed by the experimental results. We also provide representative visualization of ATE error mapped onto trajectories in Fig. 4, both on synthetic and real dataset. It demonstrates that IncEventGS is able to recover more accurate motion trajectories. + +# 5. Conclusion + +We present the first pose-free 3D Gaussian splattering reconstruction model based on event camera, i.e. IncEventGS, and is useful for real world applications. We adopt the tracking and mapping paradigm in conventional SLAM pipeline to do incremental motion estimation and 3D scene reconstruction simultaneously. To handle the continuous and asynchronous characteristics of event stream, we exploit a continuous trajectory model to model the event data formation process. The experimental results on both synthetic and real datasets demonstrate the superior performance of IncEventGS over prior state-of-the-art methods in terms of high-quality novel image synthesis and camera pose estimation. + +Acknowledgements. This work was supported in part by NSFC under Grant 62202389, in part by a grant from the Westlake University-Muyuan Joint Research Institute, and in part by the Westlake Education Foundation. + +![](images/43889b42d33be00152edcc0f79ae10f5bde5061a44ffc0b75c25a0585944cf18.jpg) + +![](images/e2cee0c49b5ec01484f8577bd27e78f92f0bfa55c347f1ec4985960b6176b495.jpg) +Figure 3. Qualitative evaluation for novel view image synthesis on real dataset. It demonstrates that our method is able to render better images with fewer artifacts than event NeRF methods and two-stage methods. Note that there are no GT images aligned with the event camera, and we choose the closest images from the RGB camera and crop them to the same size as the rendered images for visual comparisons. +Figure 4. Representative visualization of ATE error mapped onto trajectories for the synthetic (office0) and real (6 dof) datasets, generated by the EVO toolbox using the same ground truth poses, demonstrating the superior performance of our method in pose estimation. + +# References + +[1] Hiroyuki Deguchi, Mana Masuda, Takuya Nakabayashi, and Hideo Saito. E2gs: Event enhanced gaussian splatting. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1676-1682. IEEE, 2024. 2 +[2] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. arXiv preprint arXiv:2312.07504, 2023. 2 +[3] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth and optical flow estimation. In CVPR, 2018. 1 +[4] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 5 +[5] Michael Grupp. evo: Python package for the evaluation of odometry and slam. https://github.com/MichaelGrupp/evo, 2017.5 +[6] Jiarui Hu, Xianhao Chen, Boyin Feng, Guanglin Li, Liangjing Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field. arXiv preprint arXiv:2403.16095, 2024. 2 +[7] Huajian Huang, Longwei Li, Hui Cheng, and Sai-Kit Yeung. Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular, stereo, and rgb-d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[8] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-nerf: Event based neural radiance field. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 837–847, 2023. 1, 2 +[9] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5 +[10] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[11] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 2023. 1, 2, 3, 5 +[12] Hanme Kim, Stefan Leutenegger, and Andrew J. Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In ECCV, 2016. 1 +[13] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. Tum-vie: The tum stereo visual-inertial event dataset. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8601-8608. IEEE, 2021. 5 +[14] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 8(3):1587-1594, 2023. 1, 2, 5 + +[15] Simon Klenk, Marvin Motzet, Lukas Koestler, and Daniel Cremers. Deep event visual odometry. In 2024 International Conference on 3D Vision (3DV), pages 739-749. IEEE, 2024. 1, 5 +[16] Wenpu Li, Pian Wan, Peng Wang, Jinghang Li, Yi Zhou, and Peidong Liu. Benerf: Neural radiance fields from a single blurry image and event stream. In European Conference on Computer Vision (ECCV), 2024. 2 +[17] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse and noisy events under non-uniform motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 5 +[18] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 2 +[19] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1 +[20] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE Transactions on Robotics, 33(5):1255-1262, 2017. 2 +[21] Junkai Niu, Sheng Zhong, Xiuyuan Lu, Shaojie Shen, Guillermo Gallego, and Yi Zhou. Esvo2: Direct visual-inertial odometry with stereo event cameras. IEEE Transactions on Robotics, 2025. 5 +[22] Delin Qu, Chi Yan, Dong Wang, Jie Yin, Dan Xu, Bin Zhao, and Xuelong Li. Implicit event-rgbd neural slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[23] Henri Rebecq, Timo Horstschaefer, Guillermo Gallego, and Davide Scaramuzza. Evo: A geometric approach to event-based 6-dof parallel tracking and mapping in real time. IEEE Robotics and Automation Letters, 2017. 1 +[24] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 5 +[25] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4992-5002, 2023. 1, 2, 4, 5 +[26] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion Revisited. In CVPR, pages 4104-4113, 2016. 2 +[27] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 5 +[28] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica + +dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5 +[29] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 2 +[30] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. Advances in Neural Information Processing Systems, 2023. 1 +[31] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. arXiv preprint arXiv:2405.14959, 2024. 2 +[32] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4170–4179, 2023. 4 +[33] Peng Wang, Lingzhe Zhao, Yin Zhang, Shiyu Zhao, and Peidong Liu. Mba-slam: Motion blur aware dense visual slam with radiance fields representation. arXiv preprint arXiv:2411.08279, 2024. 2 +[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4 +[35] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher A Metzler. Event3dgs: Event-based 3d gaussian splattering for fast egomotion. arXiv preprint arXiv:2406.02972, 2024. 2 +[36] Chi Yan, Delin Qu, Dong Wang, Dan Xu, Zhigang Wang, Bin Zhao, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[37] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2 +[38] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. In European Conference on Computer Vision (ECCV), 2024. 2, 4 +[39] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 5 \ No newline at end of file diff --git a/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/images.zip b/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9a15c75d7abd7b42614fa39e8830cecbf0e06eb3 --- /dev/null +++ b/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e0bd7cfe243b1cbd4925ce32750b823e6b3d349c0b61602fda4b3b4b0e30d68 +size 801557 diff --git a/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/layout.json b/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5e8759f75a7f5b220fecc1118cfdc85918452d39 --- /dev/null +++ b/2025/IncEventGS_ Pose-Free Gaussian Splatting from a Single Event Camera/layout.json @@ -0,0 +1,8072 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 532, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 532, + 121 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 532, + 121 + ], + "type": "text", + "content": "IncEventGS: Pose-Free Gaussian Splating from a Single Event Camera" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 152, + 496, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 152, + 496, + 168 + ], + "spans": [ + { + "bbox": [ + 120, + 152, + 496, + 168 + ], + "type": "text", + "content": "Jian Huang" + }, + { + "bbox": [ + 120, + 152, + 496, + 168 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 120, + 152, + 496, + 168 + ], + "type": "text", + "content": " Chengrui Dong" + }, + { + "bbox": [ + 120, + 152, + 496, + 168 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 120, + 152, + 496, + 168 + ], + "type": "text", + "content": " Xuanhua Chen" + }, + { + "bbox": [ + 120, + 152, + 496, + 168 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 120, + 152, + 496, + 168 + ], + "type": "text", + "content": " Peidong Liu" + }, + { + "bbox": [ + 120, + 152, + 496, + 168 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 122, + 176, + 496, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 176, + 496, + 213 + ], + "spans": [ + { + "bbox": [ + 122, + 176, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 176, + 496, + 213 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 122, + 176, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 176, + 496, + 213 + ], + "type": "text", + "content": "Westlake University " + }, + { + "bbox": [ + 122, + 176, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 122, + 176, + 496, + 213 + ], + "type": "text", + "content": "Northeastern University " + }, + { + "bbox": [ + 122, + 176, + 496, + 213 + ], + "type": "inline_equation", + "content": "\\{" + }, + { + "bbox": [ + 122, + 176, + 496, + 213 + ], + "type": "text", + "content": "huangjian39, dongchengrui, chenxuanhua, liupeidong\\}@westlake.edu.cn https://github.com/WU-CVGL/IncEventGS" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 154, + 248, + 201, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 248, + 201, + 260 + ], + "spans": [ + { + "bbox": [ + 154, + 248, + 201, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 273, + 302, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 302, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 302, + 571 + ], + "type": "text", + "content": "Implicit neural representation and explicit 3D Gaussian Splating (3D-GS) for novel view synthesis have achieved remarkable progress with frame-based camera (e.g. RGB and RGB-D cameras) recently. Compared to frame-based camera, a novel type of bio-inspired visual sensor, i.e. event camera, has demonstrated advantages in high temporal resolution, high dynamic range, low power consumption, and low latency, which make it favored for many robotic applications. In this work, we present IncEventGS, an incremental 3D Gaussian Splating reconstruction algorithm with a single event camera, without the assumption of known camera poses. To recover the 3D scene representation incrementally, we exploit the tracking and mapping paradigm of conventional SLAM pipelines for IncEventGS. Given the incoming event stream, the tracker first estimates an initial camera motion based on prior reconstructed 3D-GS scene representation. The mapper then jointly refines both the 3D scene representation and camera motion based on the previously estimated motion trajectory from the tracker. The experimental results demonstrate that IncEventGS delivers superior performance compared to prior NeRF-based methods and other related baselines, even if we do not have the ground-truth camera poses. Furthermore, our method can also deliver better performance compared to state-of-the-art event visual odometry methods in terms of camera motion estimation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 594, + 135, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 594, + 135, + 607 + ], + "spans": [ + { + "bbox": [ + 56, + 594, + 135, + 607 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 614, + 300, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 300, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 300, + 710 + ], + "type": "text", + "content": "Reconstructing accurate 3D scene representations from 2D images has been a long-standing challenge in computer vision and robotics, driving substantial efforts over the past few decades. Among those pioneering works, Neural Radiance Fields (NeRF) [19] and 3D Gaussian Splatting (3D-GS) [11], stand out for their utilization of differentiable rendering techniques, and have garnered significant attention due to their capability to recover high-quality 3D scene representation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 249, + 564, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 249, + 564, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 249, + 564, + 357 + ], + "type": "text", + "content": "from 2D images. Commonly used sensors for 3D scene reconstruction are usually frame-based cameras, such as the RGB and RGB-D cameras. They usually capture full-brightness intensity images within a short exposure time at a regular frequency. Due to the characteristic of this data-capturing process, they often suffer from motion blur or fail to capture accurate and informative intensity information under fast motion and low-light conditions, which would further affect the performance of downstream applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 357, + 564, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 357, + 564, + 513 + ], + "spans": [ + { + "bbox": [ + 317, + 357, + 564, + 513 + ], + "type": "text", + "content": "The event camera, a bio-inspired sensor, has gained significant attention in recent years for its potential to address the limitations of frame-based cameras under challenging conditions. Unlike conventional cameras, event cameras record brightness changes asynchronously at each pixel, emitting events when a predefined threshold is surpassed. This unique operation offers several advantages over conventional cameras, in terms of high temporal resolution, high dynamic range, low latency, and power consumption. Although event cameras have attractive characteristics for challenging environments, they cannot be directly integrated into existing frame-based 3D reconstruction algorithms that rely on processing dense 2D brightness intensity images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 513, + 564, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 564, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 564, + 680 + ], + "type": "text", + "content": "Several pioneering works have been proposed to exploit event stream [3, 12, 23] to recover the motion trajectory and scene representation. While existing methods deliver impressive performance, they usually exploit 2.5D semi-dense depth maps to represent the 3D scene, and bundle adjustment (BA) is hardly performed, due to the asynchronous and sparse characteristics of event data stream. Klenk et al. [15] recently proposed to convert event stream into event voxel grids, and then adapt a previous frame-based deep visual odometry pipeline [30] for accurate camera motion estimation. As NeRF exhibited impressive scene representation capability recently, several works [8, 14, 17, 25] explore to recover the underlying dense 3D scene NeRF representation from event stream, by assuming ground-truth poses are available." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 680, + 564, + 728 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 680, + 564, + 728 + ], + "spans": [ + { + "bbox": [ + 317, + 680, + 564, + 728 + ], + "type": "text", + "content": "In contrast to those works, we propose IncEventGS, an incremental dense 3D scene reconstruction algorithm from a single event camera, by exploiting Gaussian Splatting as the underlying scene representation. Different from prior event-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 717, + 148, + 727 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 717, + 148, + 727 + ], + "spans": [ + { + "bbox": [ + 70, + 717, + 148, + 727 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 755, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 755, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 293, + 755, + 317, + 765 + ], + "type": "text", + "content": "26933" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 301, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 301, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 301, + 370 + ], + "type": "text", + "content": "based NeRF reconstruction methods, IncEventGS does not require any ground-truth camera poses, which is more challenging and provides more flexibility for real-world robotic application scenarios. To overcome the challenges brought by unknown poses, IncEventGS adopts the tracking and mapping paradigm of conventional SLAM pipelines [20]. In particular, IncEventGS exploits prior explored and reconstructed 3D scenes for camera motion estimation of incoming event stream during the tracking stage. Both the 3D-GS scene representation and camera motions are then jointly optimized (i.e. event-based bundle adjustment) during the mapping stage, for more accurate scene representation and motion estimation. The 3D scene is progressively expanded and densified. The experimental results on both synthetic and real datasets demonstrate that IncEventGS can recover the underlying 3D scene representation and camera motion trajectory accurately. In particular, IncEventGS outperforms prior NeRF-based methods and other related baselines in terms of scene representation recovery, even without ground-truth poses. Furthermore, our method also delivers better camera motion estimation accuracy than the most recent state-of-the-art visual odometry algorithm, in terms of the Absolute Trajectory Error (ATE) metric. The recovered 3D scene representation can be further used to render novel brightness images. Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 371, + 301, + 515 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 371, + 299, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 299, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 299, + 407 + ], + "type": "text", + "content": "- We present an incremental 3D Gaussian Splitting reconstruction algorithm from a single event camera, without requiring the ground-truth camera poses." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 407, + 301, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 407, + 301, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 407, + 301, + 441 + ], + "type": "text", + "content": "- We propose a novel initialization strategy tailored to the event data stream, which is vital to the success of the algorithm." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 443, + 301, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 301, + 515 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 301, + 515 + ], + "type": "text", + "content": "- The experimental results on both the synthetic and real datasets demonstrate the superior performance of our method over prior NeRF-based methods and related baselines in terms of novel view synthesis and better performance over state-of-the-art event-based visual odometry algorithm in terms of camera motion estimation." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 525, + 147, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 147, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 147, + 538 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 299, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 299, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 299, + 582 + ], + "type": "text", + "content": "We review two main areas of prior works: event-based neural radiance fields and 3D Gaussian Splatting, which are the most related to our work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "content": "Event-based Neural Radiance Fields. Prior works [8, 14, 25] propose to exploit event stream to recover the neural radiance fields with known camera motion trajectory. Low et al. [17] further improves the reconstruction algorithm to handle sparse and noisy events under non-uniform motion. The recovered neural radiance fields can then be used to render novel view brightness images. The ground-truth poses are usually computed from corresponding brightness images via COLMAP [26] or provided by the indoor motion-capturing system. Recently, Qu et al. [22] proposed to integrate event measurements into an RGB-D implicit neural SLAM framework and achieved robust performance in mo" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 72, + 564, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 564, + 192 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 564, + 192 + ], + "type": "text", + "content": "tion blur scenarios. Li et al. [16] also proposed exploiting event measurements and a single blurry image to recover the underlying neural 3D scene representation. In contrast to those works, IncEventGS conducts incremental 3D scene reconstruction without requiring any prior ground-truth poses, which is more challenging and provides more flexibility for practical robotic application scenarios. The method further exploits 3D Gaussian Splatting as the underlying scene representation, which demonstrates better image rendering quality and efficiency, compared to the NeRF-based representation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 194, + 564, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 194, + 564, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 194, + 564, + 602 + ], + "type": "text", + "content": "3D Gaussian Splatting. 3D Gaussian Splatting [11] proposes a novel explicit 3D representation to further improve both the training and rendering efficiency compared to Neural Radiance Fields. Due to its impressive efficient scene representation capability, several pioneering works have been proposed to exploit 3D-GS for incremental 3D reconstruction. For example, Keetha et al. [10] propose an RGBD-based 3D-GS SLAM, employing an online tracking and mapping system tailored to the underlying Gaussian representation. Yan et al. [36] implement a coarse-to-fine camera tracking approach based on the sparse selection of Gaussians. Matsuki et al. [18] propose to apply 3D Gaussian Splatting to do incremental 3D reconstruction using a single moving monocular or RGB-D camera. Huang et al. [7] exploit ORB-SLAM3 to compute accurate camera poses and feed it into a 3D-GS algorithm for dense mapping. Fu et al. [2] use monocular depth estimation with 3D-GS. Yugay et al. [37] combine DROID-SLAM [29] based camera tracking with active and inactive 3D-GS sub-maps. Wang et al. [33] integrate bundle-adjustment and 3DGS [38] to estimated camera trajectory within the exposure time. Hu et al. [6] propose a novel depth uncertainty model to ensure the selection of valuable Gaussian primitives during optimization. While those methods deliver impressive performance in terms of 3D scene recovery and motion estimation, they usually assume the usage of frame-based images (i.e. either RGB or RGB-D date). On the contrary, we propose to exploit pure event measurements for incremental 3D-GS reconstruction. Several concurrent studies have recently explored the use of 3D Gaussians for event-based reconstruction, including EvGGS [31], Event3DGS [35], and E2GS [1]. EvGGS and Event3DGS rely solely on event data, whereas E2GS incorporates both event data and blurry images. However, the key difference between these methods and ours is that they all rely on ground-truth poses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 612, + 375, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 612, + 375, + 624 + ], + "spans": [ + { + "bbox": [ + 318, + 612, + 375, + 624 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 632, + 564, + 728 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 632, + 564, + 728 + ], + "spans": [ + { + "bbox": [ + 317, + 632, + 564, + 728 + ], + "type": "text", + "content": "The overview of our IncEventGS is shown in Fig. 1. Given only a single event camera, IncEventGS incrementally performs tracking and dense mapping under the framework of 3D Gaussian Splatting, to recover both the camera motion trajectory and 3D scene representation simultaneously. Since event data are asynchronous, they cannot be directly integrated with the 3D-GS representation. We therefore process the event data stream into chunks according to a fixed time window." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 755, + 318, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 755, + 318, + 765 + ], + "spans": [ + { + "bbox": [ + 293, + 755, + 318, + 765 + ], + "type": "text", + "content": "26934" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 71, + 523, + 314 + ], + "blocks": [ + { + "bbox": [ + 95, + 71, + 523, + 314 + ], + "lines": [ + { + "bbox": [ + 95, + 71, + 523, + 314 + ], + "spans": [ + { + "bbox": [ + 95, + 71, + 523, + 314 + ], + "type": "image", + "image_path": "3d633c7e1ccb14850406539ba2de87c11d49c5b276f733782eece3bafe0e1918.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 322, + 564, + 376 + ], + "lines": [ + { + "bbox": [ + 54, + 322, + 564, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 322, + 564, + 376 + ], + "type": "text", + "content": "Figure 1. The pipeline of IncEventGS. IncEventGS processes incoming event stream by dividing it into chunks and representing the camera trajectory as a continuous model. It randomly samples two close consecutive timestamps to integrate the corresponding event streams. Two brightness images are rendered from 3D-GS at the corresponding poses, and we minimize the photometric loss between the synthesized and measured brightness change. During initialization, a pre-trained depth estimation model estimates depth from the rendered images to bootstrap the system." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "spans": [ + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "text", + "content": "We associate each chunk with a continuous time trajectory parameterization in the " + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "inline_equation", + "content": "\\mathfrak{se}(3)" + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "text", + "content": " space. Two close consecutive timestamps (i.e., " + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "inline_equation", + "content": "t_{k + \\Delta t}" + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "text", + "content": " is a small time interval) can be randomly sampled, and the measured brightness change " + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(x)" + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "text", + "content": " can be computed from the corresponding event stream. Based on the parameterized trajectory, the corresponding camera poses (i.e., " + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_k" + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{k + \\Delta t}" + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "text", + "content": ") can be determined and the brightness images (i.e., " + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{I}}_k" + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{I}}_{k + \\Delta t}" + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "text", + "content": ") can be further rendered from the 3D-GS. The synthesized brightness change " + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{E}}(x)" + }, + { + "bbox": [ + 54, + 380, + 300, + 500 + ], + "type": "text", + "content": " can be computed for event loss computation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 502, + 301, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 301, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 301, + 622 + ], + "type": "text", + "content": "During the tracking stage, we optimize only the camera motion trajectory of the newly accumulated event chunk and exploit the recovered trajectory to initialize the dense bundle adjustment (BA) algorithm for the mapping stage. During the mapping stage, we continuously densify 3D Gaussians for newly explored areas and prune transparent 3D Gaussians. For computational efficiency, we exploit a sliding window of the latest chunks and perform BA only within this window for both 3D-GS reconstruction and motion trajectory estimation. We will detail each component as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 635, + 195, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 195, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 195, + 649 + ], + "type": "text", + "content": "3.1. 3D Scene Representation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 656, + 301, + 728 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 301, + 728 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 301, + 728 + ], + "type": "text", + "content": "Following 3D-GS [11], the scene is represented by a set of 3D Gaussian primitives, each of which contains mean position " + }, + { + "bbox": [ + 55, + 656, + 301, + 728 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 656, + 301, + 728 + ], + "type": "text", + "content": " in the world coordinate, 3D covariance " + }, + { + "bbox": [ + 55, + 656, + 301, + 728 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 656, + 301, + 728 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 55, + 656, + 301, + 728 + ], + "type": "inline_equation", + "content": "\\mathbf{o} \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 656, + 301, + 728 + ], + "type": "text", + "content": ", and color " + }, + { + "bbox": [ + 55, + 656, + 301, + 728 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 656, + 301, + 728 + ], + "type": "text", + "content": ". To ensure that the covariance matrix remains positive semi-definite throughout the gradient descent, the covariance " + }, + { + "bbox": [ + 55, + 656, + 301, + 728 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 55, + 656, + 301, + 728 + ], + "type": "text", + "content": " is parameterized using a" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 380, + 527, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 380, + 527, + 392 + ], + "spans": [ + { + "bbox": [ + 318, + 380, + 527, + 392 + ], + "type": "text", + "content": "scale vector " + }, + { + "bbox": [ + 318, + 380, + 527, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 318, + 380, + 527, + 392 + ], + "type": "text", + "content": " and rotation matrix " + }, + { + "bbox": [ + 318, + 380, + 527, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 318, + 380, + 527, + 392 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 405, + 400, + 563, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 400, + 563, + 414 + ], + "spans": [ + { + "bbox": [ + 405, + 400, + 563, + 414 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {1}", + "image_path": "03709b21f0bc2c9622bf8c5fba139c597e37ca6a0c9b7e6062a1896faf6e3924.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 423, + 563, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 423, + 563, + 446 + ], + "spans": [ + { + "bbox": [ + 317, + 423, + 563, + 446 + ], + "type": "text", + "content": "where scale matrix " + }, + { + "bbox": [ + 317, + 423, + 563, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\text{diag}([s])" + }, + { + "bbox": [ + 317, + 423, + 563, + 446 + ], + "type": "text", + "content": " is derived from the scale vector " + }, + { + "bbox": [ + 317, + 423, + 563, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 317, + 423, + 563, + 446 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 448, + 563, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 448, + 563, + 485 + ], + "spans": [ + { + "bbox": [ + 318, + 448, + 563, + 485 + ], + "type": "text", + "content": "In order to enable rendering, 3D-GS projects 3D Gaussian primitives to the 2D image plane from a given camera pose " + }, + { + "bbox": [ + 318, + 448, + 563, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_c = \\{\\mathbf{R}_c\\in \\mathbb{R}^{3\\times 3},\\mathbf{t}_c\\in \\mathbb{R}^3\\}" + }, + { + "bbox": [ + 318, + 448, + 563, + 485 + ], + "type": "text", + "content": " using following equation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 398, + 493, + 563, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 493, + 563, + 509 + ], + "spans": [ + { + "bbox": [ + 398, + 493, + 563, + 509 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {R} _ {c} \\boldsymbol {\\Sigma} \\mathbf {R} _ {c} ^ {T} \\mathbf {J} ^ {T}, \\tag {2}", + "image_path": "a22f003ded798591b12c116a9d28f97deab0b04642a0e2f655770893b9889cf1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 516, + 564, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 516, + 564, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 516, + 564, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 317, + 516, + 564, + 590 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^{\\prime}\\in \\mathbb{R}^{2\\times 2}" + }, + { + "bbox": [ + 317, + 516, + 564, + 590 + ], + "type": "text", + "content": " is the 2D covariance matrix, " + }, + { + "bbox": [ + 317, + 516, + 564, + 590 + ], + "type": "inline_equation", + "content": "J\\in \\mathbb{R}^{2\\times 3}" + }, + { + "bbox": [ + 317, + 516, + 564, + 590 + ], + "type": "text", + "content": " is the Jacobian of the affine approximation of the projective transformation. After projecting 3D Gaussians onto the image plane, the color of each pixel is determined by sorting the Gaussians according to their depth and then applying near-to-far " + }, + { + "bbox": [ + 317, + 516, + 564, + 590 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 317, + 516, + 564, + 590 + ], + "type": "text", + "content": " -blending rendering via the following equation:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 384, + 597, + 563, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 597, + 563, + 632 + ], + "spans": [ + { + "bbox": [ + 384, + 597, + 563, + 632 + ], + "type": "interline_equation", + "content": "\\mathbf {I} = \\sum_ {i} ^ {N} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {3}", + "image_path": "0f4e9309caa31f2cd65bbe0e199251e36b9d7267a910d5ffb5544b4d383c32df.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 639, + 563, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 639, + 563, + 676 + ], + "spans": [ + { + "bbox": [ + 317, + 639, + 563, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 317, + 639, + 563, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 317, + 639, + 563, + 676 + ], + "type": "text", + "content": " is the learnable color of each Gaussian, and " + }, + { + "bbox": [ + 317, + 639, + 563, + 676 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 317, + 639, + 563, + 676 + ], + "type": "text", + "content": " is the alpha value computed by evaluating the 2D covariance " + }, + { + "bbox": [ + 317, + 639, + 563, + 676 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}^{\\prime}" + }, + { + "bbox": [ + 317, + 639, + 563, + 676 + ], + "type": "text", + "content": " multiplied with the learned Gaussian opacity " + }, + { + "bbox": [ + 317, + 639, + 563, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 317, + 639, + 563, + 676 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 350, + 684, + 563, + 707 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 684, + 563, + 707 + ], + "spans": [ + { + "bbox": [ + 350, + 684, + 563, + 707 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\mathbf {o} _ {i} \\cdot \\exp (- \\sigma_ {i}), \\quad \\sigma_ {i} = \\frac {1}{2} \\Delta_ {i} ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime - 1} \\Delta_ {i}, \\tag {4}", + "image_path": "9f2a126e6eda45155b7c8f51a3023da78f4d7cef3b7d9d18374f74afd8adcffd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 715, + 563, + 728 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 715, + 563, + 728 + ], + "spans": [ + { + "bbox": [ + 318, + 715, + 563, + 728 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 318, + 715, + 563, + 728 + ], + "type": "inline_equation", + "content": "\\Delta_i\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 318, + 715, + 563, + 728 + ], + "type": "text", + "content": " is the offset between the pixel center and the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 755, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 755, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 293, + 755, + 317, + 765 + ], + "type": "text", + "content": "26935" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 228, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 228, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 228, + 83 + ], + "type": "text", + "content": "2D Gaussian center. Depth is rendered by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 92, + 299, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 92, + 299, + 127 + ], + "spans": [ + { + "bbox": [ + 118, + 92, + 299, + 127 + ], + "type": "interline_equation", + "content": "\\mathbf {D} = \\sum_ {i} ^ {N} \\mathbf {d} _ {i} \\alpha_ {i} \\prod_ {j} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {5}", + "image_path": "ecb565eb9a6d54cc983ee421dac7ba86011094a1651031a066e6ca1a64abbc5b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 136, + 300, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 136, + 300, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 300, + 171 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 136, + 300, + 171 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 55, + 136, + 300, + 171 + ], + "type": "text", + "content": " denotes the z-depth of the center of the i-th 3D Gaussian to the camera. We also render alpha map to determine visibility:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 171, + 299, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 171, + 299, + 205 + ], + "spans": [ + { + "bbox": [ + 123, + 171, + 299, + 205 + ], + "type": "interline_equation", + "content": "\\mathbf {V} = \\sum_ {i} ^ {N} \\alpha_ {i} \\prod_ {j} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {6}", + "image_path": "29f3d584240c0d137908d6444f3ac7199ecf57e4a7319a8f7fb78ac22c7445ae.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 209, + 301, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 301, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 301, + 292 + ], + "type": "text", + "content": "The derivations presented above demonstrate that the rendered pixel color, denoted as I in Eq. (3), is a function that is differentiable with respect to the learnable attributes of 3D-GS, and the camera poses " + }, + { + "bbox": [ + 55, + 209, + 301, + 292 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_c" + }, + { + "bbox": [ + 55, + 209, + 301, + 292 + ], + "type": "text", + "content": ". This facilitates our bundle adjustment formulation, accommodating a set of event chunks and inaccurate camera motion trajectories within the framework of 3D-GS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 300, + 216, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 216, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 216, + 312 + ], + "type": "text", + "content": "3.2.Event Data Formation Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 315, + 300, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 300, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 300, + 374 + ], + "type": "text", + "content": "An event camera records changes in brightness as a stream of events asynchronously. To relate 3D-GS representation with the event stream, we sample two close consecutive timestamps (i.e., " + }, + { + "bbox": [ + 55, + 315, + 300, + 374 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 55, + 315, + 300, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 315, + 300, + 374 + ], + "type": "inline_equation", + "content": "t_{k + \\Delta t}" + }, + { + "bbox": [ + 55, + 315, + 300, + 374 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 315, + 300, + 374 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 55, + 315, + 300, + 374 + ], + "type": "text", + "content": " is a small time interval), and the measured brightness change between " + }, + { + "bbox": [ + 55, + 315, + 300, + 374 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 55, + 315, + 300, + 374 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 384, + 299, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 384, + 299, + 397 + ], + "spans": [ + { + "bbox": [ + 100, + 384, + 299, + 397 + ], + "type": "interline_equation", + "content": "\\mathbf {E} (\\mathbf {x}) = C \\left\\{e _ {i} \\left(\\mathbf {x}, t _ {i}, p _ {i}\\right) \\right\\} _ {t _ {k} < t _ {i} < t _ {k} + \\Delta t}, \\tag {7}", + "image_path": "15e1f61794b6a8b5a64d2521ba4fc1744b14c91d91d30cef200fefab0b2ae430.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "inline_equation", + "content": "e(\\mathbf{x}, t_i, p_i)" + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "text", + "content": " event within the defined time interval corresponding to pixel " + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "text", + "content": " is the fixed contrast threshold. The corresponding camera poses " + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "inline_equation", + "content": "T_k" + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "inline_equation", + "content": "T_{k + \\Delta t}" + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "text", + "content": " can be interpolated from the camera motion trajectory parameterization, and two corresponding brightness images (i.e. " + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_k" + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{k + \\Delta t}" + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "text", + "content": ") can be rendered from 3D-GS. The synthesized brightness change " + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{E}}" + }, + { + "bbox": [ + 55, + 406, + 300, + 491 + ], + "type": "text", + "content": " is modelled as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 99, + 498, + 299, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 498, + 299, + 514 + ], + "spans": [ + { + "bbox": [ + 99, + 498, + 299, + 514 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {E}} (\\mathbf {x}) = \\log \\left(\\hat {\\mathbf {I}} _ {k + \\Delta t} (\\mathbf {x})\\right) - \\log \\left(\\hat {\\mathbf {I}} _ {k} (\\mathbf {x})\\right), \\tag {8}", + "image_path": "45fe7db2ef6abf79a557bf7a96ffd7d1253cc512a09f152beed74da70552e21a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 522, + 299, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 299, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 299, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 522, + 299, + 559 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{E}} (\\mathbf{x})" + }, + { + "bbox": [ + 55, + 522, + 299, + 559 + ], + "type": "text", + "content": " depends on the parameters of both the motion trajectory parameters and 3D-GS, and is differentiable with respect to them." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": "Both in tracking and mapping, inspired by the work of [25], we segment the current event chunks into " + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "inline_equation", + "content": "n_{seg}" + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": " equal segments according to the number of events, obtaining " + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "inline_equation", + "content": "n_{seg}" + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": " timestamps that correspond to the end of each segment. We then randomly select one timestamp from these " + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "inline_equation", + "content": "n_{seg}" + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": " timestamps to serve as " + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "inline_equation", + "content": "t_{k + \\Delta t}" + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": ", and we randomly sample an integer " + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "inline_equation", + "content": "n_{win}" + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": " between the integer bounds " + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "inline_equation", + "content": "n_{low}" + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "inline_equation", + "content": "n_{up}" + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": ". The index of " + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": " is equal to the index of " + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "inline_equation", + "content": "t_{k + \\Delta t}" + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": " subtract " + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "inline_equation", + "content": "n_{win}" + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "inline_equation", + "content": "n_{seg}" + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "inline_equation", + "content": "n_{low}" + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "inline_equation", + "content": "n_{up}" + }, + { + "bbox": [ + 55, + 559, + 299, + 679 + ], + "type": "text", + "content": " are hyperparameters. This sampling strategy enables the model to capture both local and global information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 686, + 253, + 700 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 686, + 253, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 686, + 253, + 700 + ], + "type": "text", + "content": "3.3. Camera Motion Trajectory Modeling" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 704, + 300, + 727 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 704, + 300, + 727 + ], + "spans": [ + { + "bbox": [ + 55, + 704, + 300, + 727 + ], + "type": "text", + "content": "Since each event chunk usually contains too many events, we sample a portion of them according to the total number of" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 72, + 563, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 563, + 168 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 563, + 168 + ], + "type": "text", + "content": "events during optimization. Following [32, 38], we formulate the corresponding poses (i.e. " + }, + { + "bbox": [ + 317, + 72, + 563, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_k" + }, + { + "bbox": [ + 317, + 72, + 563, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 72, + 563, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{k + \\Delta t}" + }, + { + "bbox": [ + 317, + 72, + 563, + 168 + ], + "type": "text", + "content": ") at the beginning and end of the sampled event portion within each chunk, by employing a camera motion trajectory. The trajectory is represented through linear interpolation between two camera poses, one at the beginning of the chunk " + }, + { + "bbox": [ + 317, + 72, + 563, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{start}} \\in \\mathbf{SE}(3)" + }, + { + "bbox": [ + 317, + 72, + 563, + 168 + ], + "type": "text", + "content": " and the other at the end " + }, + { + "bbox": [ + 317, + 72, + 563, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{end}} \\in \\mathbf{SE}(3)" + }, + { + "bbox": [ + 317, + 72, + 563, + 168 + ], + "type": "text", + "content": ". The camera pose at time " + }, + { + "bbox": [ + 317, + 72, + 563, + 168 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 317, + 72, + 563, + 168 + ], + "type": "text", + "content": " can thus be expressed as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 323, + 176, + 563, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 176, + 563, + 201 + ], + "spans": [ + { + "bbox": [ + 323, + 176, + 563, + 201 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {k} = \\mathbf {T} _ {\\text {s t a r t}} \\cdot \\exp \\left(\\frac {t _ {k} - t _ {\\text {s t a r t}}}{t _ {\\text {e n d}} - t _ {\\text {s t a r t}}} \\cdot \\log \\left(\\mathbf {T} _ {\\text {s t a r t}} ^ {- 1} \\cdot \\mathbf {T} _ {\\text {e n d}}\\right)\\right), \\tag {9}", + "image_path": "fdcb707b281dc75acefdf8179e8189b43a3eb3605ca5a157e4d6b63a9f8a7e46.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "spans": [ + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "inline_equation", + "content": "t_{start}" + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "inline_equation", + "content": "t_{end}" + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "text", + "content": " represent the timestamps corresponding to the boundary of the event chunk. It follows that " + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_k" + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "text", + "content": " is differentiable with respect to both " + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{start}" + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{end}" + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "text", + "content": ". The objective of IncEventGS is thus to estimate both " + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{start}" + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{end}" + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "text", + "content": " for each event chunk, along with the learnable parameters of 3D-GS " + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\theta}" + }, + { + "bbox": [ + 317, + 208, + 563, + 278 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 287, + 509, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 287, + 509, + 300 + ], + "spans": [ + { + "bbox": [ + 318, + 287, + 509, + 300 + ], + "type": "text", + "content": "3.4. Incremental Tracking and Mapping" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 304, + 563, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 304, + 563, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 304, + 563, + 363 + ], + "type": "text", + "content": "For both tracking and mapping, we aim to minimize the difference between the synthesized and measured brightness changes. In particular, we compute the loss of the latest event chunk only for the tracking stage and minimize the following loss function:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 337, + 369, + 563, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 369, + 563, + 395 + ], + "spans": [ + { + "bbox": [ + 337, + 369, + 563, + 395 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {\\text {s t a r t}} ^ {*}, \\mathbf {T} _ {\\text {e n d}} ^ {*} = \\underset {\\mathbf {T} _ {\\text {s t a r t}}, \\mathbf {T} _ {\\text {e n d}}} {\\operatorname {a r g m i n}} \\left\\| \\mathbf {E} (\\mathbf {x}) - \\hat {\\mathbf {E}} (\\mathbf {x}) \\right\\| _ {2}, \\tag {10}", + "image_path": "72e48643487727e3eedd936670c028a06fbbcc1957d958949e79ee972d3fdc0a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 404, + 563, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 563, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 563, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 317, + 404, + 563, + 441 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{E}} (\\mathbf{x})" + }, + { + "bbox": [ + 317, + 404, + 563, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 404, + 563, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(\\mathbf{x})" + }, + { + "bbox": [ + 317, + 404, + 563, + 441 + ], + "type": "text", + "content": " are the synthesized and measured brightness changes respectively, corresponding to a randomly sampled event portion within the latest event chunk." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 441, + 564, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 441, + 564, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 441, + 564, + 525 + ], + "type": "text", + "content": "Once the tracking is done, we insert the latest event chunk to the mapper and exploit the estimated " + }, + { + "bbox": [ + 317, + 441, + 564, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{start}^{*}" + }, + { + "bbox": [ + 317, + 441, + 564, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 441, + 564, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{end}^{*}" + }, + { + "bbox": [ + 317, + 441, + 564, + 525 + ], + "type": "text", + "content": " as the initial value of the chunk to perform dense bundle adjustment. For computational consideration, we exploit a sliding window BA of the latest " + }, + { + "bbox": [ + 317, + 441, + 564, + 525 + ], + "type": "inline_equation", + "content": "n_w" + }, + { + "bbox": [ + 317, + 441, + 564, + 525 + ], + "type": "text", + "content": " chunks, and " + }, + { + "bbox": [ + 317, + 441, + 564, + 525 + ], + "type": "inline_equation", + "content": "n_w" + }, + { + "bbox": [ + 317, + 441, + 564, + 525 + ], + "type": "text", + "content": " is a hyperparameter. In particular, we optimize both the motion trajectories and the 3D-GS jointly by minimizing the following loss functions:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 388, + 534, + 563, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 534, + 563, + 546 + ], + "spans": [ + { + "bbox": [ + 388, + 534, + 563, + 546 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {\\text {e v e n t}} + \\lambda \\mathcal {L} _ {\\text {s s i m}}, \\tag {11}", + "image_path": "e6b6784b2191d025e572702b3afadab89915f1de300f3ce57785777c1ea06b0c.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 367, + 549, + 563, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 549, + 563, + 570 + ], + "spans": [ + { + "bbox": [ + 367, + 549, + 563, + 570 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e v e n t}} = \\left\\| \\mathbf {E} _ {i} (\\mathbf {x}) - \\hat {\\mathbf {E}} _ {i} (\\mathbf {x}) \\right\\| _ {2}, \\tag {12}", + "image_path": "8723e40c1507cefa27480f456c46636455b1ce048c4b734390831e8e7dcc1460.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 370, + 573, + 563, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 573, + 563, + 586 + ], + "spans": [ + { + "bbox": [ + 370, + 573, + 563, + 586 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s s i m} = S S I M \\left(\\mathbf {E} _ {i} (\\mathbf {x}), \\hat {\\mathbf {E}} _ {i} (\\mathbf {x})\\right), \\tag {13}", + "image_path": "89ac415effd6ee8a216c8222abe42ff77042e5e25f2a598149aba1c039f22c24.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 595, + 563, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 595, + 563, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 595, + 563, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 317, + 595, + 563, + 632 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 317, + 595, + 563, + 632 + ], + "type": "text", + "content": " is a hyperparameter, SSIM is the structural dissimilarity loss [34]. As the event data streams in, we alternatively perform tracking and mapping." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 639, + 531, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 639, + 531, + 651 + ], + "spans": [ + { + "bbox": [ + 318, + 639, + 531, + 651 + ], + "type": "text", + "content": "3.5. System Initialization and Boot-strapping" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 656, + 563, + 727 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 656, + 563, + 727 + ], + "spans": [ + { + "bbox": [ + 317, + 656, + 563, + 727 + ], + "type": "text", + "content": "Conventional frame-based 3D-GS methods usually require a good initial point cloud and camera poses computed via COLMAP [27] for initialization. However, they are usually not easy to obtain if we are only given a single event camera. We therefore initialize the 3D-GS by sampling point cloud randomly within a bounding box. The first " + }, + { + "bbox": [ + 317, + 656, + 563, + 727 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 317, + 656, + 563, + 727 + ], + "type": "text", + "content": " event chunks" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 755, + 318, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 755, + 318, + 765 + ], + "spans": [ + { + "bbox": [ + 293, + 755, + 318, + 765 + ], + "type": "text", + "content": "26936" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 301, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 301, + 133 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 301, + 133 + ], + "type": "text", + "content": "(where " + }, + { + "bbox": [ + 54, + 72, + 301, + 133 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 72, + 301, + 133 + ], + "type": "text", + "content": " is a hyperparameter) are selected for initialization, and all corresponding camera poses are randomly initialized to be near the identity matrix. We then minimize the loss computed by Eq. (11) with respect to the attributes of 3D-GS and the parameters of camera motion trajectories jointly." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 301, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 301, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 301, + 300 + ], + "type": "text", + "content": "Through experiments, we found that the above initialization procedure consistently produces satisfactory brightness images. However, the 3D structure remains of low quality due to the short baselines of the moving event camera. We further find that it could potentially affect the performance of the whole pipeline without a good initial 3D structure as more event data is received. Therefore, we utilize a monocular depth estimation network [9] to predict a dense depth map from the rendered brightness image after the pipeline is trained for certain iterations. This depth map is then used to re-initialize the centers of the 3D Gaussians by unprojecting the pixel depths, after which we repeat the minimization of Eq. (11) for system bootstrapping. More details about system re-initialization can be found in our supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 303, + 301, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 301, + 362 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 301, + 362 + ], + "type": "text", + "content": "3D-GS Incrementally Growing. As the camera moves, new Gaussians is periodically introduced to cover newly explored regions. After tracking, we obtain an accurate camera pose estimate for each new event chunk. The centers of new Gaussians are determined by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 371, + 299, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 371, + 299, + 386 + ], + "spans": [ + { + "bbox": [ + 135, + 371, + 299, + 386 + ], + "type": "interline_equation", + "content": "\\mathbf {p} = \\mathbf {T} \\cdot \\pi^ {- 1} \\left(\\mathbf {u}, d _ {u}\\right) \\tag {14}", + "image_path": "24b58c38a30ba3cbb8bd3c84a94814aaa5108c7fc0c73db984d24361489025d3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 394, + 300, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 300, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 300, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 394, + 300, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{u} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 394, + 300, + 479 + ], + "type": "text", + "content": " is pixel coordinate in the image plane, " + }, + { + "bbox": [ + 55, + 394, + 300, + 479 + ], + "type": "inline_equation", + "content": "d_u" + }, + { + "bbox": [ + 55, + 394, + 300, + 479 + ], + "type": "text", + "content": " is depth of the 3D point " + }, + { + "bbox": [ + 55, + 394, + 300, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 394, + 300, + 479 + ], + "type": "text", + "content": " projecting onto the image plane, which is rendered by Eq. (5), " + }, + { + "bbox": [ + 55, + 394, + 300, + 479 + ], + "type": "inline_equation", + "content": "\\pi^{-1}" + }, + { + "bbox": [ + 55, + 394, + 300, + 479 + ], + "type": "text", + "content": " denotes camera inverse projection, " + }, + { + "bbox": [ + 55, + 394, + 300, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 394, + 300, + 479 + ], + "type": "text", + "content": " is the camera pose from tracking. To ensure that new Gaussians are only added in previously unmapped areas, a visibility mask is computed to guide the expansion of the Gaussian splattering process, as following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 489, + 299, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 489, + 299, + 502 + ], + "spans": [ + { + "bbox": [ + 139, + 489, + 299, + 502 + ], + "type": "interline_equation", + "content": "\\mathbf {M} (p) = V < \\lambda_ {V}, \\tag {15}", + "image_path": "ff447207dcca2f3c9a51715622cce79413ba313dcc92850b344b40d3f899942f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 511, + 299, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 299, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 299, + 535 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 511, + 299, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 511, + 299, + 535 + ], + "type": "text", + "content": " is the rendered alpha map and " + }, + { + "bbox": [ + 55, + 511, + 299, + 535 + ], + "type": "inline_equation", + "content": "\\lambda_V" + }, + { + "bbox": [ + 55, + 511, + 299, + 535 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 547, + 137, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 137, + 560 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 137, + 560 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 566, + 178, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 178, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 178, + 580 + ], + "type": "text", + "content": "4.1. Experimental Setups." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "content": "Implementation Details. All experiments are conducted on a desktop PC equipped with a 5.73GHz AMD Ryzen 9 7900x CPU and an NVIDIA RTX 3090 GPU. The first " + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "inline_equation", + "content": "m = 3" + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "content": " event chunks are used for initialization. During the mapping stage, a sliding window size of " + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "inline_equation", + "content": "n_w = 20" + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "content": " is employed for the bundle adjustment algorithm. The hyperparameters are set as follows: " + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "inline_equation", + "content": "\\lambda = 0.05" + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "inline_equation", + "content": "\\lambda_V = 0.8" + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "inline_equation", + "content": "n_{seg} = 100" + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "content": ". For the synthetic dataset, " + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "inline_equation", + "content": "n_{low} = 400k" + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "inline_equation", + "content": "n_{up} = 500k" + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "content": ", while for the real dataset, " + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "inline_equation", + "content": "n_{low} = 60k" + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "inline_equation", + "content": "n_{up} = 80k" + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "content": ". Each event chunk has a time interval of " + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "inline_equation", + "content": "50~\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 584, + 301, + 728 + ], + "type": "text", + "content": ". The learning rate of the camera poses is set to 1e-4 and that for the attributes of 3D-GS are set the same as the original 3D-GS work. The number" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 72, + 564, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 564, + 121 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 564, + 121 + ], + "type": "text", + "content": "of optimization steps for initialization is 4500, and that for tracking and mapping are set to 200 and 1500 respectively. The contrast threshold " + }, + { + "bbox": [ + 317, + 72, + 564, + 121 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 317, + 72, + 564, + 121 + ], + "type": "text", + "content": " of the event camera is set to 0.1 for synthetic datasets and 0.2 for real datasets empirically." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 126, + 564, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 126, + 564, + 390 + ], + "spans": [ + { + "bbox": [ + 317, + 126, + 564, + 390 + ], + "type": "text", + "content": "Baselines and Evaluation Metrics. To the best of our knowledge, there are no existing event-only NeRF or 3D-GS methods that do not rely on ground-truth poses, making direct comparisons challenging. Therefore, we conduct a thorough comparison of our method with several event-based NeRF approaches, including E-NeRF [14], EventNeRF [25], and Robust e-NeRF [17], as well as our custom implemented two-stage method (i.e. E2VID [24] + COLMAP [27] + 3DGS [11]). E-NeRF, EventNeRF, and Robust e-NeRF leverage implicit neural radiance fields for 3D scene representation, requiring ground-truth camera poses for accurate NeRF reconstruction. For E2VID + COLMAP + 3DGS, event data is first converted into brightness images using E2VID. The camera poses are then estimated from these images using COLMAP, and 3D-GS is trained with the generated images and poses. Both the quantitative and qualitative comparisons are performed on the synthetic dataset. Since there are no paired ground truth images for the real dataset, we only perform qualitative comparisons on the real dataset. In terms of motion trajectory evaluations, we use the publicly available state-of-the-art event-only visual odometry methods, i.e. DEVO (mono) [15] and ESVO2(stereo) [21], for comparison." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 392, + 564, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 392, + 564, + 489 + ], + "spans": [ + { + "bbox": [ + 317, + 392, + 564, + 489 + ], + "type": "text", + "content": "The metrics used for novel view synthesis (NVS) include the commonly employed PSNR, SSIM, and LPIPS. For motion trajectory evaluations, we utilize Absolute Trajectory Error (ATE). To ensure fair comparisons, we employ the evaluation code provided by EventNeRF to compute the NVS metrics, which applies a linear color transformation between predictions and ground truth. Additionally, we use the public EVO toolbox [5] to compute the trajectory metrics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 521, + 564, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 521, + 564, + 653 + ], + "spans": [ + { + "bbox": [ + 317, + 521, + 564, + 653 + ], + "type": "text", + "content": "Benchmark Datasets. To properly evaluate the performance of NVS and motion trajectory estimation, we synthesized event data using the 3D scene models from the Replica dataset [28]. In particular, we exploit the room0, room2, office0, office2, and office3 scenes. We render high frame rate RGB images at " + }, + { + "bbox": [ + 317, + 521, + 564, + 653 + ], + "type": "inline_equation", + "content": "1000\\mathrm{Hz}" + }, + { + "bbox": [ + 317, + 521, + 564, + 653 + ], + "type": "text", + "content": " with a resolution of " + }, + { + "bbox": [ + 317, + 521, + 564, + 653 + ], + "type": "inline_equation", + "content": "768\\times 480" + }, + { + "bbox": [ + 317, + 521, + 564, + 653 + ], + "type": "text", + "content": " pixels. These images are then converted to grayscale, and the event data is generated via the events simulator [4]. The contrast threshold is set to 0.1. To simulate real-world camera motions, we exploit the same motion trajectories as that of NICE-SLAM [39] for data generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 656, + 564, + 728 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 656, + 564, + 728 + ], + "spans": [ + { + "bbox": [ + 317, + 656, + 564, + 728 + ], + "type": "text", + "content": "We use the event dataset provided by TUM-VIE [13] for real data evaluations, which is also used by E-NeRF and Robust e-NeRF. TUM-VIE captures the event datasets by a pair of Prophesee Gen4 HD event cameras with a resolution of " + }, + { + "bbox": [ + 317, + 656, + 564, + 728 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 317, + 656, + 564, + 728 + ], + "type": "text", + "content": " pixels. We only use the left-event camera data for our experiment." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 755, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 755, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 293, + 755, + 317, + 765 + ], + "type": "text", + "content": "26937" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 69, + 553, + 449 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 553, + 449 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 553, + 449 + ], + "type": "image", + "image_path": "ec53771c8d5d065b90240b3c1440f322de20fa4c99f213696b8ee4ff704a4364.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 453, + 563, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 453, + 563, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 563, + 475 + ], + "type": "text", + "content": "Figure 2. Qualitative evaluation of novel view image synthesis on the Replica dataset. The experimental results demonstrate that our method renders higher-quality images with fewer artifacts compared to event-based NeRF and two-stage approaches." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 482, + 562, + 575 + ], + "blocks": [ + { + "bbox": [ + 58, + 482, + 562, + 575 + ], + "lines": [ + { + "bbox": [ + 58, + 482, + 562, + 575 + ], + "spans": [ + { + "bbox": [ + 58, + 482, + 562, + 575 + ], + "type": "table", + "html": "
room0room2office0office2office3
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
E-NeRF13.990.580.5115.560.470.5818.910.510.5713.050.650.4414.010.620.48
EventNeRF17.290.620.3916.020.540.6418.900.430.6215.180.660.4516.770.730.33
Robust e-NeRF17.260.840.1816.430.500.5218.930.520.5616.810.810.2519.220.840.18
E2VID+COLMAP+3DGS14.450.440.5215.740.510.5518.910.310.6814.030.570.4813.250.470.53
Ours24.310.850.1723.750.790.2325.640.540.3021.740.820.2321.180.880.13
", + "image_path": "56afb1d11992bf19ff7eb43305d1c389df28f36034040148ce777d5937dff71f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 583, + 563, + 605 + ], + "lines": [ + { + "bbox": [ + 55, + 583, + 563, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 583, + 563, + 605 + ], + "type": "text", + "content": "Table 1. NVs performance comparison on Replica dataset. All results are calculated using the same code and ground-truth images. The results demonstrate that our method outperforms NeRF-based and two-stage methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 625, + 149, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 149, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 149, + 639 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 643, + 300, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 300, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 300, + 715 + ], + "type": "text", + "content": "We conduct ablation studies to confirm our design choices. In particular, we study the effect of a monocular depth estimation network for system bootstrapping and event slicing hyperparameters " + }, + { + "bbox": [ + 55, + 643, + 300, + 715 + ], + "type": "inline_equation", + "content": "n_{low}" + }, + { + "bbox": [ + 55, + 643, + 300, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 643, + 300, + 715 + ], + "type": "inline_equation", + "content": "n_{up}" + }, + { + "bbox": [ + 55, + 643, + 300, + 715 + ], + "type": "text", + "content": ". The experiments are conducted with the Replica dataset, and the results are shown in Table 3 and Table 4, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 715, + 299, + 728 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 715, + 299, + 728 + ], + "spans": [ + { + "bbox": [ + 67, + 715, + 299, + 728 + ], + "type": "text", + "content": "We found that depth initialization significantly impacts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 626, + 564, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 626, + 564, + 698 + ], + "spans": [ + { + "bbox": [ + 317, + 626, + 564, + 698 + ], + "type": "text", + "content": "pose estimation, reducing the Average Trajectory Error (ATE) from " + }, + { + "bbox": [ + 317, + 626, + 564, + 698 + ], + "type": "inline_equation", + "content": "1.534\\mathrm{cm}" + }, + { + "bbox": [ + 317, + 626, + 564, + 698 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 317, + 626, + 564, + 698 + ], + "type": "inline_equation", + "content": "0.064\\mathrm{cm}" + }, + { + "bbox": [ + 317, + 626, + 564, + 698 + ], + "type": "text", + "content": ". Additionally, this improvement in pose estimation leads to a slight enhancement in Novel View Synthesis (NVS) performance. These results verify the importance of using depth initialization during the bootstrapping stage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 704, + 564, + 728 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 704, + 564, + 728 + ], + "spans": [ + { + "bbox": [ + 318, + 704, + 564, + 728 + ], + "type": "text", + "content": "We compare several combinations of hyperparameters " + }, + { + "bbox": [ + 318, + 704, + 564, + 728 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{low}}" + }, + { + "bbox": [ + 318, + 704, + 564, + 728 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 318, + 704, + 564, + 728 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{up}}" + }, + { + "bbox": [ + 318, + 704, + 564, + 728 + ], + "type": "text", + "content": ", which refer to the range of event slicing window" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 755, + 318, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 755, + 318, + 765 + ], + "spans": [ + { + "bbox": [ + 293, + 755, + 318, + 765 + ], + "type": "text", + "content": "26938" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 523, + 133 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 523, + 133 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 523, + 133 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 523, + 133 + ], + "type": "table", + "html": "
room0room2office0office2office31d3d6dofdeskdesk2
DEVO0.2890.2660.1380.2810.1560.1470.3032.930.7320.201
E2VID+COLMAP17.9359.96105.1918.41417.284.26816.909.8821.5710.13
ESVO2-----0.3371.0660.5871.1472.506
Ours0.0460.0670.0450.0460.0540.1150.2980.2510.2310.129
", + "image_path": "6a39b90025b518decbb9ec15db062d3021531a14f0ea1c08f9fb383b68e5b1e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 91, + 182, + 264, + 227 + ], + "blocks": [ + { + "bbox": [ + 55, + 141, + 563, + 164 + ], + "lines": [ + { + "bbox": [ + 55, + 141, + 563, + 164 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 563, + 164 + ], + "type": "text", + "content": "Table 2. Pose accuracy (ATE, cm) on Replica and TUM-VIE datasets. The results demonstrate that our method delivers better performance in terms of camera motion estimation. Since ESVO2 is a stereo event camera method, it cannot run on the Replica dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 91, + 182, + 264, + 227 + ], + "lines": [ + { + "bbox": [ + 91, + 182, + 264, + 227 + ], + "spans": [ + { + "bbox": [ + 91, + 182, + 264, + 227 + ], + "type": "table", + "html": "
SettingPSNR↑SSIM↑LPIPS↓ATE
full21.740.820.230.046
w/o17.800.760.261.534
", + "image_path": "2dd37915808dbe380d06c56b0ff7983954ea747e75aea35bc72d8321d1f6e01a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 85, + 302, + 272, + 385 + ], + "blocks": [ + { + "bbox": [ + 55, + 234, + 300, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 300, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 300, + 289 + ], + "type": "text", + "content": "Table 3. Ablation Study about Depth Initialization. The unit of ATE is cm. The experimental results demonstrate the effectiveness of the initialization strategy. It not only improves the quality of rendered images, but also improves the accuracy of the camera motion estimation significantly." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 302, + 272, + 385 + ], + "lines": [ + { + "bbox": [ + 85, + 302, + 272, + 385 + ], + "spans": [ + { + "bbox": [ + 85, + 302, + 272, + 385 + ], + "type": "table", + "html": "
SettingPSNR↑SSIM↑LPIPS↓ATE
1k-10k16.070.640.460.167
10k-50k18.410.720.330.079
80k-200k20.990.790.250.079
400k-500k21.740.820.230.046
500k-600k20.950.790.230.050
600k-700k18.060.750.280.214
", + "image_path": "8a791a93fc4c5e0ab86c7d35530d43d92be63ce3acb925b55ea6ba7af276c183.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 393, + 301, + 416 + ], + "lines": [ + { + "bbox": [ + 55, + 393, + 301, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 301, + 416 + ], + "type": "text", + "content": "Table 4. Ablation Study on Event Slice Window Size (Hyperparameters " + }, + { + "bbox": [ + 55, + 393, + 301, + 416 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{low}}" + }, + { + "bbox": [ + 55, + 393, + 301, + 416 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 393, + 301, + 416 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{up}}" + }, + { + "bbox": [ + 55, + 393, + 301, + 416 + ], + "type": "text", + "content": "). The unit of ATE is cm." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 422, + 300, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 300, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 300, + 483 + ], + "type": "text", + "content": "size. Table 4 demonstrates that both too small and too large window sizes negatively impact the performance of Novel View Synthesis (NVS) and pose estimation. Consequently, we select " + }, + { + "bbox": [ + 55, + 422, + 300, + 483 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{low}} = 400k" + }, + { + "bbox": [ + 55, + 422, + 300, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 422, + 300, + 483 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{up}} = 500k" + }, + { + "bbox": [ + 55, + 422, + 300, + 483 + ], + "type": "text", + "content": " for our experiments on the Replica dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 492, + 198, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 198, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 198, + 506 + ], + "type": "text", + "content": "4.3. Quantitative Evaluations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 299, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 299, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 299, + 570 + ], + "type": "text", + "content": "We conduct quantitative evaluations against event NeRF methods (E-NeRF, EventNeRF, and Robust e-NeRF) and our custom implemented two-stage method (i.e. E2VID + COLMAP + 3DGS) in terms of the quality of NVS and pose estimation performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 571, + 301, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 301, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 301, + 714 + ], + "type": "text", + "content": "The NVS performance is evaluated on Replica-dataset and the results are presented in Table 1. It is important to note that the metrics are lower than those typically observed in standard NeRF/3D-GS methods for RGB images, primarily due to the lack of absolute brightness supervision. Even though NeRF-based methods use ground truth poses for training, IncEventGS still significantly outperforms them, highlighting the advantages of our approach utilizing a 3D Gaussian representation. Additionally, our method greatly surpasses two-stage method that also employs 3D Gaussian representation, demonstrating superior pose estimation and the effectiveness of our bundle adjustment technique." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 715, + 299, + 728 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 715, + 299, + 728 + ], + "spans": [ + { + "bbox": [ + 67, + 715, + 299, + 728 + ], + "type": "text", + "content": "We evaluate pose estimation performance using the ATE" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 185, + 564, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 564, + 246 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 564, + 246 + ], + "type": "text", + "content": "metric on both synthetic and real datasets, comparing our method with DEVO, ESVO2 and E2VID + COLMAP. The results, presented in Table 2, show that our method outperforms both baselines, validating the effectiveness of our incremental tracking and mapping technique." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 262, + 454, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 262, + 454, + 275 + ], + "spans": [ + { + "bbox": [ + 318, + 262, + 454, + 275 + ], + "type": "text", + "content": "4.4. Qualitative Evaluations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 282, + 564, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 282, + 564, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 282, + 564, + 487 + ], + "type": "text", + "content": "We evaluate our method against event NeRF methods and two-stage method qualitatively in terms of novel view image synthesis, both on synthetic and real data. The results are presented in both Fig. 2 and Fig. 3. It demonstrates that our method can deliver better novel view images, while event NeRF methods and two-stage method render images with additional artifacts. Compared to NeRF-based methods, our approach demonstrates the advantage of IncEventGS by leveraging 3D Gaussian Splatting as the underlying scene representation. In contrast to two-stage method, our dense bundle adjustment optimizes both 3D Gaussian Splatting and camera pose using event data, whereas two-stage approaches tend to accumulate errors over time, as confirmed by the experimental results. We also provide representative visualization of ATE error mapped onto trajectories in Fig. 4, both on synthetic and real dataset. It demonstrates that IncEventGS is able to recover more accurate motion trajectories." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 506, + 391, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 506, + 391, + 520 + ], + "spans": [ + { + "bbox": [ + 318, + 506, + 391, + 520 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 530, + 564, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 530, + 564, + 674 + ], + "spans": [ + { + "bbox": [ + 317, + 530, + 564, + 674 + ], + "type": "text", + "content": "We present the first pose-free 3D Gaussian splattering reconstruction model based on event camera, i.e. IncEventGS, and is useful for real world applications. We adopt the tracking and mapping paradigm in conventional SLAM pipeline to do incremental motion estimation and 3D scene reconstruction simultaneously. To handle the continuous and asynchronous characteristics of event stream, we exploit a continuous trajectory model to model the event data formation process. The experimental results on both synthetic and real datasets demonstrate the superior performance of IncEventGS over prior state-of-the-art methods in terms of high-quality novel image synthesis and camera pose estimation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 680, + 563, + 728 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 680, + 563, + 728 + ], + "spans": [ + { + "bbox": [ + 317, + 680, + 563, + 728 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by NSFC under Grant 62202389, in part by a grant from the Westlake University-Muyuan Joint Research Institute, and in part by the Westlake Education Foundation." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 755, + 318, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 755, + 318, + 765 + ], + "spans": [ + { + "bbox": [ + 293, + 755, + 318, + 765 + ], + "type": "text", + "content": "26939" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 66, + 553, + 407 + ], + "blocks": [ + { + "bbox": [ + 61, + 66, + 553, + 407 + ], + "lines": [ + { + "bbox": [ + 61, + 66, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 61, + 66, + 553, + 407 + ], + "type": "image", + "image_path": "43889b42d33be00152edcc0f79ae10f5bde5061a44ffc0b75c25a0585944cf18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 447, + 552, + 687 + ], + "blocks": [ + { + "bbox": [ + 55, + 413, + 563, + 447 + ], + "lines": [ + { + "bbox": [ + 55, + 413, + 563, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 563, + 447 + ], + "type": "text", + "content": "Figure 3. Qualitative evaluation for novel view image synthesis on real dataset. It demonstrates that our method is able to render better images with fewer artifacts than event NeRF methods and two-stage methods. Note that there are no GT images aligned with the event camera, and we choose the closest images from the RGB camera and crop them to the same size as the rendered images for visual comparisons." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 447, + 552, + 687 + ], + "lines": [ + { + "bbox": [ + 59, + 447, + 552, + 687 + ], + "spans": [ + { + "bbox": [ + 59, + 447, + 552, + 687 + ], + "type": "image", + "image_path": "e2cee0c49b5ec01484f8577bd27e78f92f0bfa55c347f1ec4985960b6176b495.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 697, + 563, + 720 + ], + "lines": [ + { + "bbox": [ + 55, + 697, + 563, + 720 + ], + "spans": [ + { + "bbox": [ + 55, + 697, + 563, + 720 + ], + "type": "text", + "content": "Figure 4. Representative visualization of ATE error mapped onto trajectories for the synthetic (office0) and real (6 dof) datasets, generated by the EVO toolbox using the same ground truth poses, demonstrating the superior performance of our method in pose estimation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 755, + 317, + 764 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 755, + 317, + 764 + ], + "spans": [ + { + "bbox": [ + 293, + 755, + 317, + 764 + ], + "type": "text", + "content": "26940" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 91, + 301, + 727 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 301, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 301, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 301, + 134 + ], + "type": "text", + "content": "[1] Hiroyuki Deguchi, Mana Masuda, Takuya Nakabayashi, and Hideo Saito. E2gs: Event enhanced gaussian splatting. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1676-1682. IEEE, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 301, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 301, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 301, + 167 + ], + "type": "text", + "content": "[2] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. arXiv preprint arXiv:2312.07504, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 301, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 301, + 211 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 301, + 211 + ], + "type": "text", + "content": "[3] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth and optical flow estimation. In CVPR, 2018. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 212, + 301, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 212, + 301, + 266 + ], + "spans": [ + { + "bbox": [ + 62, + 212, + 301, + 266 + ], + "type": "text", + "content": "[4] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 266, + 301, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 266, + 301, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 266, + 301, + 300 + ], + "type": "text", + "content": "[5] Michael Grupp. evo: Python package for the evaluation of odometry and slam. https://github.com/MichaelGrupp/evo, 2017.5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 300, + 301, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 301, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 301, + 354 + ], + "type": "text", + "content": "[6] Jiarui Hu, Xianhao Chen, Boyin Feng, Guanglin Li, Liangjing Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field. arXiv preprint arXiv:2403.16095, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 354, + 301, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 301, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 301, + 409 + ], + "type": "text", + "content": "[7] Huajian Huang, Longwei Li, Hui Cheng, and Sai-Kit Yeung. Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular, stereo, and rgb-d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 410, + 301, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 301, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 301, + 453 + ], + "type": "text", + "content": "[8] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-nerf: Event based neural radiance field. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 837–847, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 453, + 301, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 301, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 301, + 508 + ], + "type": "text", + "content": "[9] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 509, + 300, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 509, + 300, + 563 + ], + "spans": [ + { + "bbox": [ + 57, + 509, + 300, + 563 + ], + "type": "text", + "content": "[10] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 563, + 301, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 563, + 301, + 606 + ], + "spans": [ + { + "bbox": [ + 57, + 563, + 301, + 606 + ], + "type": "text", + "content": "[11] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 607, + 301, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 607, + 301, + 639 + ], + "spans": [ + { + "bbox": [ + 57, + 607, + 301, + 639 + ], + "type": "text", + "content": "[12] Hanme Kim, Stefan Leutenegger, and Andrew J. Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In ECCV, 2016. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 639, + 301, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 301, + 684 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 301, + 684 + ], + "type": "text", + "content": "[13] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. Tum-vie: The tum stereo visual-inertial event dataset. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8601-8608. IEEE, 2021. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 685, + 301, + 727 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 685, + 301, + 727 + ], + "spans": [ + { + "bbox": [ + 57, + 685, + 301, + 727 + ], + "type": "text", + "content": "[14] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 8(3):1587-1594, 2023. 1, 2, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 564, + 727 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 73, + 564, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 564, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 564, + 116 + ], + "type": "text", + "content": "[15] Simon Klenk, Marvin Motzet, Lukas Koestler, and Daniel Cremers. Deep event visual odometry. In 2024 International Conference on 3D Vision (3DV), pages 739-749. IEEE, 2024. 1, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 118, + 563, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 563, + 163 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 563, + 163 + ], + "type": "text", + "content": "[16] Wenpu Li, Pian Wan, Peng Wang, Jinghang Li, Yi Zhou, and Peidong Liu. Benerf: Neural radiance fields from a single blurry image and event stream. In European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 163, + 563, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 163, + 563, + 207 + ], + "spans": [ + { + "bbox": [ + 320, + 163, + 563, + 207 + ], + "type": "text", + "content": "[17] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse and noisy events under non-uniform motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 209, + 563, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 209, + 563, + 253 + ], + "spans": [ + { + "bbox": [ + 320, + 209, + 563, + 253 + ], + "type": "text", + "content": "[18] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 254, + 563, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 254, + 563, + 298 + ], + "spans": [ + { + "bbox": [ + 320, + 254, + 563, + 298 + ], + "type": "text", + "content": "[19] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 300, + 563, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 300, + 563, + 332 + ], + "spans": [ + { + "bbox": [ + 320, + 300, + 563, + 332 + ], + "type": "text", + "content": "[20] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE Transactions on Robotics, 33(5):1255-1262, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 334, + 563, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 334, + 563, + 376 + ], + "spans": [ + { + "bbox": [ + 320, + 334, + 563, + 376 + ], + "type": "text", + "content": "[21] Junkai Niu, Sheng Zhong, Xiuyuan Lu, Shaojie Shen, Guillermo Gallego, and Yi Zhou. Esvo2: Direct visual-inertial odometry with stereo event cameras. IEEE Transactions on Robotics, 2025. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 379, + 563, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 379, + 563, + 423 + ], + "spans": [ + { + "bbox": [ + 320, + 379, + 563, + 423 + ], + "type": "text", + "content": "[22] Delin Qu, Chi Yan, Dong Wang, Jie Yin, Dan Xu, Bin Zhao, and Xuelong Li. Implicit event-rgbd neural slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 424, + 563, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 424, + 563, + 468 + ], + "spans": [ + { + "bbox": [ + 320, + 424, + 563, + 468 + ], + "type": "text", + "content": "[23] Henri Rebecq, Timo Horstschaefer, Guillermo Gallego, and Davide Scaramuzza. Evo: A geometric approach to event-based 6-dof parallel tracking and mapping in real time. IEEE Robotics and Automation Letters, 2017. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 469, + 563, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 469, + 563, + 514 + ], + "spans": [ + { + "bbox": [ + 320, + 469, + 563, + 514 + ], + "type": "text", + "content": "[24] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 514, + 563, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 514, + 563, + 570 + ], + "spans": [ + { + "bbox": [ + 320, + 514, + 563, + 570 + ], + "type": "text", + "content": "[25] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4992-5002, 2023. 1, 2, 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 571, + 563, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 571, + 563, + 603 + ], + "spans": [ + { + "bbox": [ + 320, + 571, + 563, + 603 + ], + "type": "text", + "content": "[26] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion Revisited. In CVPR, pages 4104-4113, 2016. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 605, + 563, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 605, + 563, + 639 + ], + "spans": [ + { + "bbox": [ + 320, + 605, + 563, + 639 + ], + "type": "text", + "content": "[27] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 639, + 563, + 727 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 639, + 563, + 727 + ], + "spans": [ + { + "bbox": [ + 320, + 639, + 563, + 727 + ], + "type": "text", + "content": "[28] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 755, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 755, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 293, + 755, + 317, + 765 + ], + "type": "text", + "content": "26941" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 299, + 555 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 299, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 299, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 299, + 95 + ], + "type": "text", + "content": "dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 299, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 299, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 299, + 129 + ], + "type": "text", + "content": "[29] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. Advances in neural information processing systems, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 299, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 299, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 299, + 162 + ], + "type": "text", + "content": "[30] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. Advances in Neural Information Processing Systems, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 299, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 299, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 299, + 207 + ], + "type": "text", + "content": "[31] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. arXiv preprint arXiv:2405.14959, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 299, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 299, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 299, + 252 + ], + "type": "text", + "content": "[32] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4170–4179, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 299, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 299, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 299, + 297 + ], + "type": "text", + "content": "[33] Peng Wang, Lingzhe Zhao, Yin Zhang, Shiyu Zhao, and Peidong Liu. Mba-slam: Motion blur aware dense visual slam with radiance fields representation. arXiv preprint arXiv:2411.08279, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 299, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 299, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 299, + 342 + ], + "type": "text", + "content": "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 299, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 299, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 299, + 387 + ], + "type": "text", + "content": "[35] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher A Metzler. Event3dgs: Event-based 3d gaussian splattering for fast egomotion. arXiv preprint arXiv:2406.02972, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 299, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 299, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 299, + 431 + ], + "type": "text", + "content": "[36] Chi Yan, Delin Qu, Dong Wang, Dan Xu, Zhigang Wang, Bin Zhao, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 434, + 299, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 299, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 299, + 466 + ], + "type": "text", + "content": "[37] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 468, + 299, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 299, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 299, + 499 + ], + "type": "text", + "content": "[38] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. In European Conference on Computer Vision (ECCV), 2024. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 501, + 299, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 299, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 299, + 555 + ], + "type": "text", + "content": "[39] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 756, + 317, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 756, + 317, + 765 + ], + "spans": [ + { + "bbox": [ + 293, + 756, + 317, + 765 + ], + "type": "text", + "content": "26942" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/419f8c28-377b-4f29-8c8b-a2cba093a85b_content_list.json b/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/419f8c28-377b-4f29-8c8b-a2cba093a85b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8b2f44a968cc8b295dd61c949052bc6c72c7cdef --- /dev/null +++ b/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/419f8c28-377b-4f29-8c8b-a2cba093a85b_content_list.json @@ -0,0 +1,1677 @@ +[ + { + "type": "text", + "text": "Incomplete Multi-View Multi-Label Learning via Disentangled Representation and Label Semantic Embedding", + "text_level": 1, + "bbox": [ + 102, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xu Yan $^{1}$ , Jun Yin $^{1,*}$ , Jie Wen $^{2}$ $^{1}$ Shanghai Maritime University, $^{2}$ Harbin Institute of Technology, Shenzhen yanxu@stu.shmtu.edu.cn, junyin@shmtu.edu.cn, jiewen_pr@126.com", + "bbox": [ + 202, + 202, + 794, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In incomplete multi-view multi-label learning scenarios, it is crucial to use the incomplete multi-view data to extract consistent and specific representations from different data sources and to fully exploit the missing label information. However, most previous approaches ignore the separation problem between view-shared and specific information. To address this problem, in this paper, we propose a method that can separate view-consistent features from view-specific features under the Variational Autoencoder (VAE) framework. Specifically, we first introduce cross-view reconstruction to capture view-consistent features and extract shared information from different views through unsupervised pre-training. Subsequently, we develop a disentangling module to learn specific features by minimizing the variational upper bound of mutual information between consistent and specific features. Finally, we utilize prior label relevance information derived from training data to guide the learning of the distribution of label semantic embeddings, aggregating relevant semantic embeddings and maintaining the label relevance topology in the semantic space. In extensive experiments, our model outperforms existing state-of-the-art algorithms on several real-world datasets, which fully validates its strong adaptability to missing views and labels.", + "bbox": [ + 91, + 323, + 483, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 715, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, with the increasing diversity and complexity of data, multi-view learning has gained widespread attention as an effective data processing technique. In many real-world applications, such as image categorization, video analysis, and text mining, data is often presented in the form of multiple views, which makes multi-view learning methods an effective tool for solving complex problems [28, 35, 36]. In the past few years, many multi-view learning methods have been proposed based on the assumption", + "bbox": [ + 89, + 739, + 482, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of view completeness [4, 30, 38]. However, it is not uncommon to have missing views when dealing with multi-view data, which poses a challenge to the learning effectiveness of the model [19, 39].", + "bbox": [ + 511, + 292, + 903, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meanwhile, multi-label learning is an equally important research area. Unlike traditional single-label classification tasks, multi-label learning aims to deal with situations where instances may belong to more than one category at the same time. The key to multi-label learning is how to deal with the correlation between labels, which is helpful in improving the classification performance [12, 42]. Multi-label datasets are often incomplete due to various uncertainties involved in manual labeling. So research on incomplete multi-label learning is particularly important.", + "bbox": [ + 511, + 354, + 905, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional multi-label classification relies on features extracted from a single view for classification, which leads to insufficient feature information. Using multi-view data for multi-label classification can effectively alleviate this problem. Currently, there are many models dealing with multi-view multi-label classification (MVMLC) tasks that show promising performance [33, 37, 41]. These models typically assume that the data is complete, and currently, only a few methods are capable of handling both missing views and missing labels simultaneously.", + "bbox": [ + 511, + 507, + 905, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To deal with the Double Incomplete Multi-View Multi-Label Classification problem (DIMVMLC), several deep learning models based on autoencoders have been proposed [22, 23, 31]. However, there are still only a few methods for disentangling consistent features and specific features. SIMM [33] identifies consistent features by minimizing confusing adversarial loss and multi-label loss, uses orthogonal constraints to eliminate shared information, and extracts specific information from raw information. However, SIMM cannot be applied to the missing data case, and the forced orthogonality constraint eliminates linear correlation, and there may still be a nonlinear relationship between shared and specific features. This relationship may be difficult to capture by simple orthogonality constraints, affecting the expressiveness of the features. MTD [23] simultaneously learns consistent and specific features, sepa", + "bbox": [ + 511, + 659, + 905, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30722", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rating them using multi-view cross-channel contrastive loss. Optimizing both tasks simultaneously may lead to a complex optimization process where the model needs to trade off between two objectives, increasing the difficulty of convergence.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above complex double incomplete multi-view multi-label learning problem, we propose a Disentangled Representation and Label Semantic Embedding based method (DRLS), as shown in Fig. 1. Our approach consists of three main parts: (1) Learning multi-view consistent features by minimizing the KL divergence between the approximate posterior of each view's consistent features and the multi-view joint posterior. (2) Learning the disentangled specific features by minimizing the variational upper bound of mutual information between the consistent features and the specific features. (3) We model the distribution of label semantic embeddings by leveraging prior label relevance information from the training data and utilizing graph neural networks. Our model is built upon the Variational Autoencoder (VAE) framework. Numerous experiments demonstrate the excellent performance of DRLS. Finally, the contributions of this paper are summarized as follows:", + "bbox": [ + 91, + 167, + 482, + 436 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel approach based on disentangled representation and label semantic embedding for solving the DIMVMLC problem. It can effectively utilize the consistent and specific information of multi-view data, and separate them through a disentangled representation learning module for effective multi-label classification tasks.", + "- We introduce a pre-training strategy that learns view-consistent representations via pre-training and then minimizes the variational upper bound of mutual information between consistent and specific features to separate them effectively.", + "- We design a graph neural network-based label semantic embedding learning module that simultaneously aggregates embeddings of related labels and maintains the label relevance topology in the semantic space.", + "- Our approach achieves leading performance on five publicly available datasets and can be applied to both complete and arbitrarily missing data." + ], + "bbox": [ + 89, + 439, + 482, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 89, + 724, + 235, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many MVMLC methods have been proposed. For example, CDMM [40] builds independent classifiers based on neural networks constructed with nonlinear kernel mapping to ensure that each view learns consistent labeling results. The diversity across views is addressed by integrating the Hilbert-Schmidt independence criterion [7] with an exploration of the variations among different views. MMFA [3] initially extracts both common and private features from different views, then computes the similarity between the original features and the common features. The attention", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "weight matrix is derived by multiplying this similarity with the synergistic feature matrix, which is subsequently used for multi-label prediction.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For DIMVMLC, iMvWL [27] uses Nonnegative Matrix Factorization (NMF) to learn the shared subspace of multiple views and introduces label information and local label correlation matrices to improve the discriminative power of the subspace. NAIML [20] exploits both consistency among multiple views and the complex global and local structure among labels to cope with the lack of available labels. Both of these methods are shallow machine learning approaches, and current deep neural network-based methods show better performance in this area, such as DDINet [31], DICNet [22], and SIP [21]. DDINet is the first to use a deep neural network to deal with DIMVMLC. DICNet leverages contrastive loss to learn shared features across views, aligns cross-view instances, and improves feature discrimination through a multi-view fusion module, effectively handling missing views and labels. The SIP model utilizes the principles of the information bottleneck theory to reduce the non-shared information while preserving the relevant features, and captures the associations between labels through prototype instances in the latent space.", + "bbox": [ + 511, + 136, + 906, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 453, + 604, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Problem definition", + "text_level": 1, + "bbox": [ + 511, + 479, + 692, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a dataset $\\{x^{(v)}\\}_{v=1}^m$ consisting of $n$ samples, each with $m$ views, where $x^{(v)} \\in \\mathbb{R}^{d_v}$ . Here $d_v$ denotes the feature dimension of the $v$ th view. The label vector is $y \\in \\{0,1\\}^k$ , where $k$ denotes the number of categories. If $y_i = 0$ , the sample does not have the $i$ -th label; if $y_i = 1$ , it does. To handle missing data, we define the set $\\mathcal{W}(|\\mathcal{W}| \\leq m)$ to denote the set of observable views of a sample, and the observable multi-view data can be defined as $\\{x^{(v)}\\}_{v \\in \\mathcal{W}}$ , abbreviated as $\\{x\\}$ . We also define a set of observable labels for a sample $\\mathcal{U}(|\\mathcal{U}| \\leq k)$ . Our goal is to train a model to predict the labels of unseen samples. Our loss and objective functions are defined in the case of a single sample.", + "bbox": [ + 511, + 502, + 906, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2. View-consistent feature", + "text_level": 1, + "bbox": [ + 511, + 695, + 730, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In traditional multi-view learning, consistent features are often learned separately without cross-view interaction [33], making it hard to fully capture the common semantic information across views. In this subsection, we first introduce how to learn multi-view consistent features through cross-view reconstruction.", + "bbox": [ + 511, + 719, + 903, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, we define $q_{\\phi_v}(c \\mid x^{(v)})$ with learnable parameters $\\phi_v$ to approximate the posterior $p(c \\mid x^{(v)})$ , where $c$ represents the consistent feature. The consistent features of each view should contain the public information of all views, inspired by [2], we learn the consistent features by constraining the approximate posterior of each view to be similar to", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30723", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the multi-view joint posterior:", + "bbox": [ + 89, + 90, + 292, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ D _ {K L} \\left(q _ {\\phi_ {v}} (c \\mid x ^ {(v)}) \\| p (c \\mid \\{x \\})\\right) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 112, + 482, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D_{KL}(\\cdot \\|\\cdot)$ represents the KL divergence, $|\\mathcal{W}|$ denotes the number of available views, and $p(c\\mid \\{x\\})$ denotes the multi-view joint posterior.", + "bbox": [ + 89, + 154, + 482, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Evidence Lower Bound (ELBO) is derived from Eq. (1) by further applying Bayes' theorem to decompose the multi-view joint posterior distribution $p(c \\mid \\{x\\})$ . Due to space constraints, the exact derivation is given in the supplementary material Sec. 6.1.", + "bbox": [ + 89, + 199, + 483, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The final result is as follows:", + "bbox": [ + 109, + 276, + 303, + 289 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\log p (\\{x \\}) - \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ D _ {K L} \\left(q _ {\\phi_ {v}} (c | x ^ {(v)}) \\| p (c | \\{x \\})\\right) \\right] \\\\ = \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ \\mathbb {E} _ {q _ {\\phi_ {v}} (c | x ^ {(v)})} [ \\log p (\\{x \\} | c) ] \\right. \\\\ \\left. - D _ {K L} \\left(q _ {\\phi_ {v}} (c \\mid x ^ {(v)}) \\| p (c)\\right) \\right] \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 292, + 480, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This establishes the Evidence Lower Bound (ELBO). Thus, by maximizing this lower bound, we are effectively solving the minimization problem in Eq. (1).", + "bbox": [ + 89, + 409, + 483, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By assuming each view is conditionally independent from the others given $c$ , we can efficiently decompose the data likelihood as $p(\\{x\\} |c) = \\prod_{i\\in \\mathcal{W}}p\\left(x^{(i)}|c\\right)\\approx \\prod_{i\\in \\mathcal{W}}p_{\\theta_i}\\left(x^{(i)}|c\\right)$ , where $p_{\\theta_i}\\left(x^{(i)}|c\\right)$ with learnable parameters $\\theta_{i}$ approximates $p(x^{(i)}|c)$ . The objective in Eq. (2) involves extracting consistent features from different views. To facilitate subsequent analysis, we introduce an identifier for the consistent features of each view to distinguish them, i.e., $\\{c^{(v)}\\}_{v = 1}^m$ . Introducing this identifier does not alter the model architecture. Additionally, we introduce a trade-off parameter $\\alpha$ [10], which controls the trade-off between reconstruction quality and regularization, resulting in the following lower bound:", + "bbox": [ + 89, + 455, + 483, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n s i s t}} = \\\\ \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ \\mathbb {E} _ {q _ {\\phi_ {v}} (c ^ {(v)} | x ^ {(v)})} \\left[ \\sum_ {i \\in \\mathcal {W}} \\log p _ {\\theta_ {i}} (x ^ {(i)} | c ^ {(v)}) \\right] \\right. \\\\ \\left. - \\alpha D _ {K L} \\left(q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) \\| p (c)\\right) \\right] \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 672, + 480, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "here $q_{\\phi_v}(c^{(v)} \\mid x^{(v)})$ can be regarded as an encoder, and we assume that $q_{\\phi_v}(c^{(v)} \\mid x^{(v)}) \\sim \\mathcal{N}(\\mu_{\\phi_v}(x^{(v)}), \\sigma_{\\phi_v}^2(x^{(v)}) I)$ where $\\mu_{\\phi_v}(x^{(v)})$ and $\\sigma_{\\phi_v}^2(x^{(v)})$ denote the mean and variance encoder of the $v$ th view, and $I$ denotes the identity matrix. Due to the non-differentiable nature of the sampling process, we introduce the reparameterization trick [16] to obtain the features:", + "bbox": [ + 89, + 768, + 483, + 876 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nc ^ {(v)} = \\mu_ {\\phi_ {v}} \\left(x ^ {(v)}\\right) + \\sigma_ {\\phi_ {v}} \\left(x ^ {(v)}\\right) \\odot \\delta \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 883, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta \\sim \\mathcal{N}(0,I)$ , $\\odot$ denotes the Hadamard product, and $c^{(v)}\\in \\mathbb{R}^{d_e}$ , where $d_{e}$ represents the feature dimension. The distribution $p_{\\theta_i}(x^{(i)}|c^{(v)})$ in Eq. (3) is modeled as an MLP decoder, assuming $p_{\\theta_i}(x^{(i)}|c^{(v)})\\sim \\mathcal{N}(\\mu_{\\theta_i}(c^{(v)}),I)$ , yielding the reconstructed feature $\\hat{x}^{(i,v)} = \\mu_{\\theta_i}(c^{(v)})$ .", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A brief analysis of Eq. (3) is as follows: The first term represents both self- and cross-view reconstruction. It allows $c^{(v)}$ from each view to reconstruct the original data of different views, which helps $c^{(v)}$ capture shared information. The second term, the KL divergence, aligns the posterior $q_{\\phi_v}(c^{(v)} \\mid x^{(v)})$ of each view with the shared prior $p(c) \\sim \\mathcal{N}(0,I)$ . These two terms simultaneously encourage $c^{(v)}$ to encode the shared information between views.", + "bbox": [ + 511, + 167, + 905, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, in order to obtain a fused consistent feature shared by multiple views for the classification task, we compute the multi-view joint posterior $p(c|\\{x\\})$ using the Product-of-Expert (PoE) approach [11]. This method is flexible and effective in dealing with the missing views by directly removing the corresponding parameter when a view is missing. According to the setting in [32], we compute $p(c|\\{x\\})$ as follows:", + "bbox": [ + 511, + 287, + 905, + 409 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p (c | \\{x \\}) \\propto p (c) \\prod_ {v \\in \\mathcal {W}} p (c \\mid x ^ {(v)}) \\\\ := p (c) \\prod_ {v \\in \\mathcal {W}} q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 415, + 903, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The same reparameterization trick is used to obtain the fused multi-view consistent feature $\\bar{c} \\in \\mathbb{R}^{d_e}$ , which is sampled from $p(c|\\{x\\})$ .", + "bbox": [ + 511, + 489, + 905, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Disentangled view-specific feature", + "text_level": 1, + "bbox": [ + 511, + 544, + 808, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce view-specific features $\\{s^{(v)}\\}_{v = 1}^{m}$ that encode information specific to each view beyond the consistent features. We consider minimizing the mutual information $I(c^{(v)};s^{(v)})$ between the consistent and specific features for each view, which allows us to maximally disentangle the specific features from the consistent features. A natural idea is to do the learning of consistent features first, followed by disentangling the specific features from them [15]. Thus, we first pre-train with $\\mathcal{L}_{\\text{consist}}$ for a certain number of epochs, depending on the dataset size. In the second stage, we decrease the learning rate for the encoders of consistent features $\\{q_{\\phi_v}(c^{(v)}\\mid x^{(v)})\\}_{v = 1}^m$ to stabilize the pre-trained features during further disentanglement. Our disentangling objective function is as follows:", + "bbox": [ + 511, + 566, + 906, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} I \\left(c ^ {(v)}; s ^ {(v)}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 785, + 903, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We rewrite $I(c^{(v)}; s^{(v)})$ as follows, more details are provided in the supplementary material Sec. 6.2:", + "bbox": [ + 511, + 830, + 903, + 861 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(c ^ {(v)}; s ^ {(v)}\\right) = I \\left(x ^ {(v)}; c ^ {(v)}\\right) + I \\left(x ^ {(v)}; s ^ {(v)}\\right) - I \\left(x ^ {(v)}; c ^ {(v)}, s ^ {(v)}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 867, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30724", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/20f4de535dc3f2dea301c5a8e6a84c11cf76cfb33fab2ca7e619a873cab5b8a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 89, + 468, + 276 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/06eb3429ae8cb37d63d1552a95323cfe190c63b0f879a9644f76d77c6b4c21cd.jpg", + "image_caption": [ + "Figure 1. The main framework of DRLS. The upper-left corner of the figure represents the view-consistent feature learning module, the upper-right corner shows the disentangled view-specific feature learning module, and the lower part depicts the label semantic embedding learning module." + ], + "image_footnote": [], + "bbox": [ + 99, + 277, + 228, + 398 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/011e2c2d4a035e7f5163749b37ce9d34224ec98559f2a5996647c8faf2dcfa35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 277, + 341, + 383 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1cf3ddb39043e7ebe6d16790b34b36cf72f6819ee4ae77f783c06b509e72fbf3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 277, + 488, + 396 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/65044f6f5d93d7f6168d0e82a77bafbc7facf670a5c8779c5b35b3e9034cb6b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 89, + 903, + 284 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd8d930529901c6b889138194a7ce6a7a683230b041bfada56d8dc8ea74cbbef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 282, + 903, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Clearly, directly optimizing Eq. (7) is challenging. Therefore, we introduce the variational upper bounds for $I(x^{(v)};c^{(v)})$ and $I(x^{(v)};s^{(v)})$ , as well as the variational lower bound for $I(x^{(v)};c^{(v)},s^{(v)})$ [14], resulting in:", + "bbox": [ + 89, + 462, + 482, + 523 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I \\left(c ^ {(v)}; s ^ {(v)}\\right) \\leq D _ {K L} \\left(q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) \\| p \\left(c ^ {(v)}\\right)\\right) \\\\ + D _ {K L} \\left(q _ {\\varphi_ {v}} \\left(s ^ {(v)} \\mid x ^ {(v)}\\right) \\| p \\left(s ^ {(v)}\\right)\\right) \\\\ - \\mathbb {E} _ {q _ {\\phi_ {v}} (c ^ {(v)} | x ^ {(v)}) q _ {\\varphi_ {v}} (s ^ {(v)} | x ^ {(v)})} [ \\log p _ {\\vartheta_ {v}} (x ^ {(v)} | c ^ {(v)}, s ^ {(v)}) ] \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 527, + 480, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that minimizing this variational upper bound on mutual information trains a decoder $p_{\\vartheta_v}(x^{(v)}|c^{(v)}, s^{(v)})$ and a specific feature encoder $q_{\\varphi_v}(s^{(v)}|x^{(v)})$ . Here, we concatenate $c^{(v)}$ and $s^{(v)}$ as input to the MLP decoder $p_{\\vartheta_v}(x^{(v)}|c^{(v)}, s^{(v)})$ to produce the reconstructed feature $\\hat{x}^{(v)}$ . Similar to Sec. 3.2, $q_{\\varphi_v}(s^{(v)}|x^{(v)}) \\sim \\mathcal{N}(\\mu_{\\varphi_v}(x^{(v)}), \\sigma_{\\varphi_v}^2(x^{(v)})I)$ parameterizes the distribution of the specific features. The feature $s^{(v)} \\in \\mathbb{R}^{d_e}$ is derived using the reparameterization trick as shown in Eq. (4). We assume that $p(c^{(v)})$ and $p(s^{(v)})$ follow standard Gaussian distributions. Finally, we obtain the disentanglement loss:", + "bbox": [ + 89, + 604, + 483, + 772 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i s e n t}} = \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ D _ {K L} \\left(q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) \\| p \\left(c ^ {(v)}\\right)\\right) \\right. \\\\ + D _ {K L} \\left(q _ {\\varphi_ {v}} \\left(s ^ {(v)} \\mid x ^ {(v)}\\right) \\| p \\left(s ^ {(v)}\\right)\\right) \\\\ \\left. - \\mathbb {E} _ {q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) q _ {\\varphi_ {v}} \\left(s ^ {(v)} \\mid x ^ {(v)}\\right)} \\left[ \\log p _ {\\vartheta_ {v}} \\left(x ^ {(v)} \\mid c ^ {(v)}, s ^ {(v)}\\right) \\right] \\right] \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 777, + 480, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comparing Eq. (3) and Eq. (9), the consistent objective function requires each consistent feature $c^{(v)}$ to reconstruct", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the original features across views, whereas in the disentanglement loss, we use the concatenation of consistent and specific features exclusively for self-view reconstruction. Since we pre-train a set of encoders $\\{q_{\\phi_v}(c^{(v)}|x^{(v)})\\}_{v = 1}^m$ in the second stage, $c^{(v)}$ and $s^{(v)}$ are concatenated to reconstruct $x^{(v)}$ . As $c^{(v)}$ already contains the information shared between $x^{(v)}$ and other views, $s^{(v)}$ only needs to encode the complementary information that is not captured by $c^{(v)}$ . This pre-training strategy enables more efficient disentanglement. We use a simple and effective method to obtain the fused multi-view specific feature $\\bar{s}\\in \\mathbb{R}^{d_e}$ , for subsequent classification tasks:", + "bbox": [ + 511, + 462, + 906, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {s} = \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} s ^ {(v)} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 650, + 906, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we do not use PoE to fuse the specific features as in Sec. 3.2, because it is more suitable for merging shared information. PoE is sensitive to consensus among different views [18]. Specifically, if one view's probability distribution is overly sharp and biased, the joint distribution will have a biased mean prediction [1, 18, 26]. Given that specific features from different views often have differences, we choose a simple average fusion method to mitigate extreme variations and retain the overall trend.", + "bbox": [ + 511, + 694, + 906, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we fuse the multi-view consistent and specific information. Following the method in [23], we use specific features to enhance the consistent features:", + "bbox": [ + 511, + 830, + 905, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz = \\omega (\\bar {s}) \\odot \\bar {c} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 885, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30725", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$z \\in \\mathbb{R}^{d_e}$ represents the final fused representation, $\\omega(\\cdot)$ is the sigmoid activation function, and $\\odot$ denotes the Hadamard product.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Label semantic embedding", + "text_level": 1, + "bbox": [ + 89, + 145, + 334, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To fully utilize the incomplete supervisory information, we introduce label semantic embedding learning guided by prior information. Here, we construct our model within the framework of the Variational Graph Autoencoder (VGAE) [17], which employs variational inference to model the latent distribution of label embeddings during the learning process.", + "bbox": [ + 89, + 167, + 483, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we define a relationship graph among labels, where the graph $G = (V,E)$ represents the label relationship graph, with $V$ denoting the nodes (categories) and $E$ representing the edges connecting the nodes. We model the adjacency matrix $A \\in \\mathbb{R}^{k \\times k}$ of graph $G$ in terms of conditional probabilities, specifically $P(l_j|l_i)$ , which denotes the probability that label $l_j$ occurs when label $l_i$ occurs, where $l_i$ represents the $i$ th label. This is computed based on the label matrix $Y \\in \\mathbb{R}^{n_{train} \\times k}$ from the training set. Since $P(l_j|l_i)$ is not equal to $P(l_i|l_j)$ , the adjacency matrix $A$ is asymmetric, which more accurately reflects real-world scenarios. The specific calculation method is as follows:", + "bbox": [ + 89, + 273, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i j} = \\left\\{ \\begin{array}{l l} P \\left(l _ {j} \\mid l _ {i}\\right) = \\frac {C _ {i j}}{d _ {i}}, & \\text {i f} i \\neq j \\\\ 0, & \\text {i f} i = j \\end{array} \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 465, + 483, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C = Y^{\\top}Y$ , $C \\in \\mathbb{R}^{k \\times k}$ is the label co-occurrence matrix, $C_{ij}$ is the co-occurrence count of labels $l_i$ and $l_j$ , $d_i = \\sum_{a=1}^{n_{train}} Y_{ai} = Y_{:i}^{\\top} Y_{:i}$ for $i = 1,2,\\ldots,k$ , which represents the total occurrence count of label $l_i$ . It should be noted that we set the diagonal of the adjacency matrix $A$ to 0.", + "bbox": [ + 89, + 513, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Graph Isomorphism Network (GIN) provides stronger feature aggregation capabilities through learnable weights and information from self-nodes [34]. Therefore, we use GIN to model the posterior distribution of label semantic embeddings $H \\in \\mathbb{R}^{k \\times d_e}$ :", + "bbox": [ + 89, + 606, + 483, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(H \\mid H ^ {0}, A\\right) = \\prod_ {i = 1} ^ {k} q \\left(h _ {i} \\mid H ^ {0}, A\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 691, + 482, + 733 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {w i t h} q (h _ {i} | H ^ {0}, A) \\sim \\mathcal {N} (\\mu_ {i}, \\sigma_ {i} ^ {2} I)\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 734, + 395, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "here, $h_i \\in \\mathbb{R}^{d_e}$ is the $i$ -th row of $H$ , representing the semantic embedding of the $i$ -th label, and $H^0 \\in \\mathbb{R}^{k \\times k}$ is an identity matrix, which means that the embedding for label $i$ is initialized as a one-hot vector with the $i$ -th position set to 1. This approach ensures that the initial distance between different labels is large, which helps to better distinguish the semantic differences between labels during the subsequent learning process. $\\pmb{\\mu} = \\mathrm{GIN}_{\\pmb{\\mu}}(H^0, A)$ , $\\pmb{\\mu} \\in \\mathbb{R}^{k \\times d_e}$ is the mean matrix, where the $i$ -th row corresponds to the mean", + "bbox": [ + 89, + 763, + 485, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "vector $\\mu_{i}$ for the $i$ -th label; similarly, $\\sigma^2 = \\mathrm{GIN}_{\\sigma^2}(H^0,A)$ . Here, we use a three-layer GIN, where one layer can be defined as $H^{(t + 1)} = f^{(t + 1)}[(1 + \\epsilon^{(t + 1)})H^{(t)} + AH^{(t)}]$ , $t$ represents the current number of network layers, $\\epsilon^{(t + 1)}$ represents a learnable parameter, and $f^{(t + 1)}$ represents a two-layer fully connected network. $H^{(t + 1)}\\in \\mathbb{R}^{k\\times d_{t + 1}}$ represents the output of the intermediate layer.", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, we also use the reparameterization trick to sample the label semantic embeddings, with a formula similar to Eq. (4), i.e., $h_i = \\mu_i + \\sigma_i \\odot \\delta, \\delta \\sim \\mathcal{N}(0, I)$ . Finally, we obtain the label semantic embedding matrix $H$ .", + "bbox": [ + 511, + 196, + 905, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a simple and efficient inner product between variables as a generative model:", + "bbox": [ + 511, + 257, + 903, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np (A | H) = \\prod_ {i = 1} ^ {k} \\prod_ {j = 1} ^ {k} p \\left(A _ {i j} \\mid h _ {i}, h _ {j}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 299, + 903, + 342 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w i t h} p \\left(A _ {i j} \\mid h _ {i}, h _ {j}\\right) = \\omega \\left(h _ {i} ^ {\\top} h _ {j}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 344, + 820, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega(\\cdot)$ denotes the sigmoid activation function. By using the label correlation matrix as the adjacency matrix of a graph, we enable GIN to aggregate semantically related embeddings to nearby positions in the label semantic space. However, it cannot guarantee that unrelated labels are effectively pushed apart. Therefore, we introduce the generative model from VGAE to reconstruct the adjacency matrix $A$ , ensuring that the label semantic embeddings effectively preserve the topology of label correlations in the semantic space. Finally, our variational lower bound is:", + "bbox": [ + 511, + 375, + 906, + 526 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l e} = \\mathbb {E} _ {q (H | H ^ {0}, A)} [ \\log p (A | H) ] - D _ {K L} (q (H | H ^ {0}, A) \\| p (H)) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 537, + 903, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further adopt a Gaussian prior $p(H) = \\prod_{i}p(h_{i}) = \\prod_{i}\\mathcal{N}(h_{i}|0,I)$ . We use binary cross-entropy to compute the first term of $\\mathcal{L}_{le}$ as the reconstruction loss and add an identity matrix to $A$ before computing the cross-entropy loss to avoid enforcing $\\omega (h_i^\\top h_i)\\approx 0$ due to its zero diagonal.", + "bbox": [ + 511, + 569, + 905, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Overall loss function", + "text_level": 1, + "bbox": [ + 511, + 654, + 709, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we utilize label semantic embeddings for feature selection, choosing discriminative and specific features for each label. First, the learned label semantic embedding $h_i$ is mapped to a feature importance vector [9] via a fully connected layer $f(\\cdot)$ , specifically $r_i = \\omega(f(h_i))$ , where $r_i \\in \\mathbb{R}^{d_e}$ and $\\omega(\\cdot)$ is the sigmoid activation function. Subsequently, we use $r_i$ to select the features and input them into a simple neural network to obtain predictions:", + "bbox": [ + 511, + 676, + 906, + 799 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} = \\omega \\left(f _ {\\text {c l s}} \\left(z \\odot r _ {i}\\right)\\right) \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 811, + 903, + 828 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ denotes the Hadamard product, $f_{cls}$ is a fully connected network, and $p_i$ is the predicted probability of the sample being labeled as the $i$ -th class. We use the commonly used binary cross-entropy function to calculate the", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30726", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e68080dc55af55025e06d0db6a806e6235ef2636e113b61ebd4f1ff15bfc8737.jpg", + "table_caption": [ + "Table 1. Statistics for five multi-view multi-label datasets, where $m$ denotes the number of views, $n$ denotes the number of samples, $k$ denotes the number of categories, and $n/k$ denotes the average number of labels per sample." + ], + "table_footnote": [], + "table_body": "
Datasetmnkn/k
Corel5k649992603.396
Pascal0769963201.465
Espgame6207702684.686
IAPRTC126196272915.719
Mirflickr625000384.716
", + "bbox": [ + 140, + 157, + 436, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "multi-label classification loss:", + "bbox": [ + 89, + 280, + 294, + 294 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B C E} = - \\frac {1}{| \\mathcal {U} |} \\sum_ {i \\in \\mathcal {U}} \\left(y _ {i} \\log p _ {i} + \\left(1 - y _ {i}\\right) \\log \\left(1 - p _ {i}\\right)\\right) \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 303, + 480, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $y_{i}$ denotes the ground truth and we exclude missing labels from loss computation to filter out the impact of missing labels. Our final loss function is:", + "bbox": [ + 89, + 354, + 483, + 398 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {B C E} + \\beta \\mathcal {L} _ {\\text {d i s e n t}} - \\gamma \\mathcal {L} _ {l e} \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 411, + 482, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\beta$ and $\\gamma$ are trade-off parameters used to balance the different loss terms.", + "bbox": [ + 89, + 440, + 483, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 484, + 223, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 510, + 194, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments based on the settings from [21, 23, 31] to evaluate the performance of our proposed model on five datasets. They are Corel5k [5], Pascal07 [6], Espgame [29], IAPRTC12 [8], and Mirflickr [13]. In these five datasets, we utilize six different types of features, which are DenseHue, DenseSift, GIST, HSV, RGB, and LAB, with dimensions of 100, 1000, 512, 4096, 4096, and 4096, respectively. More detailed information can be found in Tab. 1.", + "bbox": [ + 89, + 534, + 483, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Compared methods", + "text_level": 1, + "bbox": [ + 89, + 680, + 279, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiments, we select eight methods for comparison with our model: DM2L [24], CDMM [40], iMvWL [27], NAIML [20], DDINet [31], DICNet [22], MTD [23], and SIP [21]. It is important to note that iMvWL, NAIML, DDINet, DICNet, MTD, and SIP are incomplete multi-view multi-label methods. CDMM can only handle complete multi-view multi-label data; therefore, we follow the settings from [21] and use the mean of available instances from the corresponding view to fill in the unavailable instances for missing views. To handle incomplete labels, we treat the missing labels as negative labels. DM2L is a single-view incomplete multi-label method, and we record the best result of each view in our experiments.", + "bbox": [ + 89, + 704, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation details", + "text_level": 1, + "bbox": [ + 513, + 90, + 728, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the datasets, in order to simulate real-world data missingness, we generate incomplete datasets following the settings in [27] and [21]. For each view, we randomly exclude $50\\%$ of the instances, ensuring at least one view per sample. Additionally, we randomly remove $50\\%$ of both positive and negative labels. We fill the missing features and labels with zeros. We use $70\\%$ of the samples for training and the remaining $30\\%$ for validation and testing. To ensure the stability of the results, we repeat the experiment 10 times and report the mean and variance of the results. Additional details are in the supplementary material Sec. 7.", + "bbox": [ + 511, + 113, + 906, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Evaluation metrics", + "text_level": 1, + "bbox": [ + 513, + 287, + 696, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consistent with some previous studies [21, 23, 31], in our experiments, we evaluate the performance of the model using six commonly used multi-label classification metrics, namely, Average Precision (AP), Area Under the Receiver Operating Characteristic Curve (AUC), Ranking Loss (RL), Hamming Loss (HL), OneError (OE), and Coverage (Cov). To better visualize the results, we use 1-RL, 1-HL, 1-OE, and 1-Cov to record the final results, where larger values indicate better performance.", + "bbox": [ + 511, + 310, + 906, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Experimental results", + "text_level": 1, + "bbox": [ + 513, + 454, + 710, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 2, we present the experimental results of nine methods on five public multi-view multi-label datasets, which include six evaluation metrics. The results in Tab. 2 are obtained under experimental settings with a $50\\%$ view missing rate, a $50\\%$ label missing rate, and $70\\%$ of training samples. The results in Tab. 2 show that our method outperforms the other eight competitors on each metric in most datasets. For example, on the AP metric of the IAPRTC12 dataset, our method is $2.4\\%$ higher than the second-ranked MTD. Specifically, Ave.R represents the average ranking of each method across the six metrics, and DRLS ranks first on all five datasets, which fully demonstrates the effectiveness of our method in addressing the DIMVMLC problem.", + "bbox": [ + 511, + 477, + 906, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, further analysis of the experimental results reveals that CDMM, DDINet, DICNet, MTD, and SIP outperform the other three methods on all five datasets. These are deep learning methods, indicating the great potential of deep learning in addressing the DIMVMLC problem.", + "bbox": [ + 511, + 674, + 906, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides solving the DIMVMLC problem, we also further validated the performance of our method in addressing the MVMLC problem. In Tab. 3, we present the experimental results of four methods on five datasets, with experiments conducted under the settings of $70\\%$ training samples, full views, and full labels. We select three representative evaluation metrics. From the results in Tab. 3, it is evident that our model still demonstrates competitive performance under this setup, indicating that our method can effectively address the MVMLC problem.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30727", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e82db5108a5de085a2796b0fe3a52d5cba837be74635fef79986ad7c5209b840.jpg", + "table_caption": [ + "Table 2. Experimental results of nine methods on five datasets with $50\\%$ missing views, $50\\%$ missing labels, and $70\\%$ training samples. Ave.R denotes the average ranking of each method on the six metrics, and the numbers at the bottom right are the standard deviations." + ], + "table_footnote": [], + "table_body": "
DatasetMetricDM2LCDMMiMvWLNAIMLDDINetDICNetMTDSIPDRLS
Corel5kAP0.2620.0050.3540.0040.2830.0080.3090.0040.3600.0090.3780.0040.4130.0070.4160.0150.4330.008
1-HL0.9870.0000.9870.0000.9780.0000.9870.0000.9870.0000.9870.0000.9880.0000.9880.0000.9880.000
1-RL0.8430.0020.8840.0030.8650.0050.8780.0020.8650.0050.8770.0040.8920.0040.9100.0030.9160.002
AUC0.8450.0020.8880.0030.8680.0050.8810.0020.8680.0050.8810.0030.8950.0040.9120.0030.9180.002
1-OE0.2950.0140.4100.0070.3110.0150.3500.0090.4370.0120.4640.0120.4910.0100.4920.0180.5090.019
1-Cov0.6470.0050.7230.0070.7020.0080.7250.0050.6890.0120.7140.0100.7480.0090.7860.0070.8040.006
Ave.R8.1674.8337.6675.3336.0004.8332.6671.8331.000
Pascal07AP0.4710.0080.5080.0050.4370.0180.4880.0030.5320.0100.5020.0070.5500.0040.5500.0090.5670.008
1-HL0.9280.0010.9310.0010.8820.0040.9280.0010.9320.0010.9300.0010.9320.0010.9310.0020.9340.001
1-RL0.7610.0050.8120.0040.7360.0150.7830.0010.8080.0050.7810.0070.8300.0030.8250.0060.8430.004
AUC0.7790.0040.8380.0030.7670.0150.8110.0010.8290.0040.8050.0060.8490.0040.8450.0050.8640.003
1-OE0.4200.0110.4190.0080.3620.0230.4210.0060.4480.0150.4260.0130.4570.0080.4630.0120.4770.011
1-Cov0.6920.0040.7590.0030.6770.0150.7270.0020.7570.0050.7280.0070.7830.0040.7770.0050.7980.004
Ave.R7.6674.8339.0006.5004.1676.1672.1672.8331.000
EspaugeAP0.2120.0020.2890.0030.2440.0050.2460.0020.2860.0040.2990.0040.3060.0030.3100.0040.3260.005
1-HL0.9820.0000.9830.0000.9720.0000.9830.0000.9830.0000.9830.0000.9830.0000.9830.0000.9830.000
1-RL0.7810.0010.8320.0010.8080.0020.8180.0020.8150.0030.8330.0030.8370.0010.8490.0020.8580.002
AUC0.7850.0010.8360.0010.8130.0020.8240.0020.8190.0030.8370.0020.8420.0010.8530.0020.8620.002
1-OE0.2940.0060.3960.0050.3430.0130.3390.0030.4270.0100.4370.0100.4480.0060.4510.0120.4730.001
1-Cov0.4880.0030.5740.0040.5480.0040.5710.0030.5530.0050.5980.0060.6010.0040.6280.0040.6520.003
Ave.R8.8334.5008.0005.6675.5003.5002.6671.8331.000
IAPRTC12AP0.2340.0030.3050.0040.2370.0030.2610.0010.3020.0050.3270.0050.3320.0020.3310.0070.3560.006
1-HL0.9800.0000.9810.0000.9690.0000.9800.0000.9800.0000.9800.0000.9810.0000.9810.0000.9810.000
1-RL0.8230.0020.8620.0020.8330.0020.8480.0010.8530.0020.8720.0020.8750.0010.8870.0040.8960.003
AUC0.8250.0010.8640.0020.8350.0010.8500.0010.8550.0030.8730.0010.8760.0010.8880.0030.8980.002
1-OE0.3400.0060.4320.0080.3520.0080.3900.0050.4350.0090.4650.0130.4710.0060.4660.0010.4900.012
1-Cov0.5290.0040.5970.0040.5640.0050.5920.0040.5940.0070.6480.0050.6490.0020.6790.0080.7070.007
Ave.R8.3334.5008.1676.6675.6674.1672.3332.1671.000
MirfickrAP0.5140.0060.5700.0020.4900.0120.5510.0020.5880.0030.5860.0050.6080.0040.6150.0040.6300.005
1-HL0.8780.0010.8860.0010.8390.0020.8820.0010.8880.0010.8880.0010.8910.0010.8910.0010.8950.001
1-RL0.8310.0030.8560.0010.8030.0080.8440.0010.8650.0020.8610.0040.8750.0010.8780.0020.8850.002
AUC0.8280.0030.8460.0010.7870.0120.8370.0010.8530.0020.8480.0040.8610.0020.8640.0020.8720.003
1-OE0.5100.0080.6310.0040.5110.0220.5850.0030.6360.0080.6420.0060.6560.0040.6640.0060.6860.009
1-Cov0.6040.0050.6400.0010.5720.0130.6310.0020.6540.0030.6460.0060.6770.0020.6760.0040.6920.003
Ave.R8.1676.0008.8337.0004.1674.6672.6672.1671.000
", + "bbox": [ + 91, + 119, + 903, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Parameter analysis", + "text_level": 1, + "bbox": [ + 89, + 680, + 274, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our DRLS model, there are three hyperparameters that need to be tuned: $\\alpha$ , $\\beta$ , and $\\gamma$ . To evaluate the sensitivity of our model to these hyperparameters, we conduct experiments on the Corel5k and Pascal07 datasets, with experiments carried out under conditions where $50\\%$ of the views are missing, $50\\%$ of the labels are missing, and $70\\%$ of the training samples are used.", + "bbox": [ + 89, + 703, + 482, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figs. 2a and 2b show the AP values of our model under different combinations of $\\beta$ and $\\gamma$ , while Figs. 2c and 2d show the AP and AUC values for different values of $\\alpha$ . To ensure the validity of the experiments, the other unrelated hyperparameters are fixed. Based on Figs. 2a and 2b, for the Corel5k dataset, $\\beta$ and $\\gamma$ are taken in the range of", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$[1e - 3, 1e - 1]$ and $[1e - 5, 1e - 3]$ , respectively, and for the Pascal07 dataset, $\\beta$ and $\\gamma$ are taken in the range of $[1e - 3, 1e - 1]$ and $[1e - 5, 1e - 3]$ , respectively. Under these conditions, our model demonstrates good performance. Regarding the hyperparameter $\\alpha$ , from Figs. 2c and 2d, it can be seen that the variation in AP and AUC values is minimal for both datasets, indicating that our model is not sensitive to $\\alpha$ . We set $\\alpha$ within the range of $[1e - 1, 3e - 1]$ .", + "bbox": [ + 511, + 680, + 906, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 815, + 663, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model consists of multiple modules and objective functions. To further analyze the contribution of each module and objective function, we conduct a series of ablation experiments on the Corel5k and Pascal07 datasets under a set-", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30728", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ea11110148074d252ae667c62ff0fa31c427203206fee4eaaaa10aa2b1487c68.jpg", + "table_caption": [ + "Table 3. Experimental results of four methods on five datasets with full views, full labels, and $70\\%$ training samples." + ], + "table_footnote": [], + "table_body": "
DatasetMetricDICNetMTDSIPDRLS
Corel5kAP0.5040.4860.5200.551
1-RL0.9250.9210.9370.947
AUC0.9280.9240.9390.949
Pascal07AP0.6090.6520.6450.655
1-RL0.8600.8900.8870.889
AUC0.8750.8980.8990.901
EspgameAP0.3880.3970.3560.399
1-RL0.8720.8800.8750.893
AUC0.8750.8830.8780.896
IAPRTC12AP0.4280.4410.3660.447
1-RL0.9100.9160.9040.926
AUC0.9110.9160.9060.927
MirflickrAP0.6610.6820.6760.698
1-RL0.8970.9060.9070.915
AUC0.8790.8880.8900.898
", + "bbox": [ + 112, + 123, + 460, + 371 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fc8debaba44af15c26aaa3f890c0f917762a371925cebef1ba970b6200fa11e2.jpg", + "table_caption": [ + "Table 4. Ablation results on two datasets with $50\\%$ missing views, $50\\%$ missing labels, and $70\\%$ training samples. 'w/o' denotes 'without'. 'A' denotes the consistent feature extraction module, 'B' denotes the disentangled specific feature extraction module, and DRLS denotes the full model. $\\mathcal{L}_{con.kl}$ represents the KL divergence term of $\\mathcal{L}_{consist}$ ." + ], + "table_footnote": [], + "table_body": "
MethodCorel5kPascal07
AP1-RLAUCAP1-RLAUC
A0.4160.9090.9120.5590.8360.858
A+B0.4310.9130.9160.5630.8400.863
DRLS0.4330.9160.9180.5670.8430.864
DRLS w/o Lcon.kl0.4150.9080.9110.5410.8230.846
DRLS w/o Ldisent0.4140.9130.9160.5570.8340.857
DRLS w/o Lle0.4300.9130.9150.5650.8420.863
", + "bbox": [ + 91, + 474, + 488, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ting of $50\\%$ missing views and $50\\%$ missing labels.", + "bbox": [ + 89, + 627, + 431, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results of the ablation experiments are presented in Tab. 4. Specifically, we first test the performance of the model using only the consistent feature extraction module. Then, we add the disentangled specific feature extraction module, and finally, we incorporate the label semantic embedding learning module. Through these three experimental setups, we can observe that each module contributes to the model's performance. Additionally, by observing the results in Tab. 4, we can see that the disentangled specific feature extraction module plays a key role in improving model performance. Furthermore, the label semantic embedding learning module also helps to enhance the model's classification ability.", + "bbox": [ + 88, + 643, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition, we also validate the impact of different objective functions on model performance. We find that removing $\\mathcal{L}_{con.kl}$ or $\\mathcal{L}_{disent}$ leads to a significant decline in model performance. Among them, $\\mathcal{L}_{con.kl}$ is the objective", + "bbox": [ + 88, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a15c22535848be4de32af0f5b392c3fcbba72a069d899999dee3c8e8037a0753.jpg", + "image_caption": [ + "(a) Corel5k" + ], + "image_footnote": [], + "bbox": [ + 516, + 74, + 712, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/762165a4b4618d75db47e24feee5dcc0a03738ab1540b1150f19b633ec65a022.jpg", + "image_caption": [ + "(b) Pascal07" + ], + "image_footnote": [], + "bbox": [ + 709, + 74, + 906, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b51f6f29bec8663730269e1b9945b830dfb92baf6f8690da363ee5c9aa7fa0b.jpg", + "image_caption": [ + "(c) Corel5k" + ], + "image_footnote": [], + "bbox": [ + 514, + 198, + 712, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00b786215a61f27137b84620256e54a56d597211d3ddddacd89a440dbf6d4e01.jpg", + "image_caption": [ + "(d) Pascal07", + "Figure 2. The AP values under different combinations of $\\beta$ and $\\gamma$ on the Corel5k (a) and Pascal07 (b) datasets; the AP and AUC values for different values of $\\alpha$ on the Corel5k (c) and Pascal07 (d) datasets." + ], + "image_footnote": [], + "bbox": [ + 709, + 198, + 903, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "function used in the pre-training stage. The experimental results show that $\\mathcal{L}_{con.kl}$ plays a regularization role, reducing redundant information in the shared features. The disentangling loss $\\mathcal{L}_{disent}$ is responsible for separating the consistent features from the specific features, which improves the quality of the specific features.", + "bbox": [ + 511, + 424, + 906, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 549, + 633, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel incomplete multi-view multi-label learning framework. Based on the assumptions of consistency and complementarity of multi-view data, we establish a method using Disentangled Representation and Label Semantic embedding (DRLS). Firstly, view-consistent features are learned using cross-view reconstruction, and the difficulty of disentangling specific features is reduced by unsupervised pre-training. Next, we achieve disentangled learning of specific features by minimizing the variational upper bound of mutual information between consistent and specific features. Meanwhile, we guide the learning of label semantic embeddings with the help of a priori label relevance to aggregate relevant semantic embeddings and maintain the topology of label relevance in the semantic space. Finally, we conduct a large number of experiments to demonstrate the effectiveness of our method in addressing the issues of missing views and labels.", + "bbox": [ + 511, + 580, + 908, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. by National Natural China (62476166) and (22PJD029).", + "bbox": [ + 511, + 844, + 691, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported Science Foundation of Shanghai Pujiang Program", + "bbox": [ + 700, + 844, + 906, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30729", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ana Lawry Aguila and Andre Altmann. A tutorial on multi-view autoencoders using the multi-view-ae library. arXiv preprint arXiv:2403.07456, 2024. 4", + "[2] Luigi Antelmi, Nicholas Ayache, Philippe Robert, and Marco Lorenzi. Sparse multi-channel variational autoencoder for the joint analysis of heterogeneous data. In International Conference on Machine Learning, pages 302-311. PMLR, 2019. 2", + "[3] Yusheng Cheng, Qingyan Li, Yibin Wang, and Weijie Zheng. Multi-view multi-label learning with view feature attention allocation. Neurocomputing, 501:857-874, 2022. 2", + "[4] Bassoma Diallo, Jie Hu, Tianrui Li, Ghufran Ahmad Khan, Xinyan Liang, and Hongjun Wang. Auto-attention mechanism for multi-view deep embedding clustering. Pattern Recognition, 143:109764, 2023. 1", + "[5] Pinar Duygulu, Kobus Barnard, Joao FG de Freitas, and David A Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In Computer Vision—ECCV 2002: 7th European Conference on Computer Vision Copenhagen, Denmark, May 28–31, 2002 Proceedings, Part IV 7, pages 97–112. Springer, 2002. 6", + "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6", + "[7] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 2", + "[8] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International workshop ontoImage, 2006. 6", + "[9] Jun-Yi Hang and Min-Ling Zhang. Collaborative learning of label semantics and deep label-specific features for multi-label classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9860-9871, 2021. 5", + "[10] Irina Higgins, Loic Matthew, Arka Pal, Christopher P Burgess, Xavier Glorot, Matthew M Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. *ICLR (Poster)*, 3, 2017. 3", + "[11] Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence. Neural computation, 14(8): 1771-1800, 2002. 3", + "[12] Sheng-Jun Huang and Zhi-Hua Zhou. Multi-label learning by exploiting label correlations locally. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 949-955, 2012. 1", + "[13] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In Proceedings of the 1st ACM international conference on Multimedia information retrieval, pages 39-43, 2008. 6", + "[14] HyeongJoo Hwang, Geon-Hyeong Kim, Seunghoon Hong, and Kee-Eung Kim. Variational interaction information maximization for cross-domain disentanglement. Advances in" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Neural Information Processing Systems, 33:22479-22491, 2020. 4, 1", + "[15] Guanzhou Ke, Bo Wang, Xiaoli Wang, and Shengfeng He. Rethinking multi-view representation learning via distilled disentangling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26774-26783, 2024. 3", + "[16] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[17] Thomas N Kipf and Max Welling. Variational graph autoencoders. arXiv preprint arXiv:1611.07308, 2016. 5", + "[18] Ana Lawry Aguila, James Chapman, and Andre Altmann. Multi-modal variational autoencoders for normative modelling across multiple imaging modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 425-434. Springer, 2023. 4", + "[19] Haobin Li, Yunfan Li, Mouxing Yang, Peng Hu, Dezhong Peng, and Xi Peng. Incomplete multi-view clustering via prototype-based imputation. arXiv preprint arXiv:2301.11045, 2023. 1", + "[20] Xiang Li and Songcan Chen. A concise yet effective model for non-aligned incomplete multi-view and missing multilabel learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5918-5932, 2021. 2, 6", + "[21] Chengliang Liu, Gehui Xu, Jie Wen, Yabo Liu, Chao Huang, and Yong Xu. Partial multi-view multi-label classification via semantic invariance learning and prototype modeling. In *Forty-first International Conference on Machine Learning*, 2, 6", + "[22] Chengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, and Yong Xu. Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In Proceedings of the AAAI conference on artificial intelligence, pages 8807-8815, 2023. 1, 2, 6", + "[23] Chengliang Liu, Jie Wen, Yabo Liu, Chao Huang, Zhihao Wu, Xiaoling Luo, and Yong Xu. Masked two-channel decoupling framework for incomplete multi-view weak multi-label learning. Advances in Neural Information Processing Systems, 36, 2024. 1, 4, 6", + "[24] Zhongchen Ma and Songcan Chen. Expand globally, shrink locally: Discriminant multi-label learning with missing labels. Pattern Recognition, 111:107675, 2021. 6", + "[25] William McGill. Multivariate information transmission. Transactions of the IRE Professional Group on Information Theory, 4(4):93-111, 1954. 1", + "[26] Yuge Shi, Brooks Paige, Philip Torr, et al. Variational mixture-of-experts autoencoders for multi-modal deep generative models. Advances in neural information processing systems, 32, 2019. 4", + "[27] Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zili Zhang. Incomplete multi-view weak-label learning. In Ijcai, pages 2703-2709, 2018. 2, 6", + "[28] Chang Tang, Zhenglai Li, Jun Wang, Xinwang Liu, Wei Zhang, and En Zhu. Unified one-step multi-view spectral clustering. IEEE Transactions on Knowledge and Data Engineering, 35(6):6449-6460, 2022. 1" + ], + "bbox": [ + 516, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30730", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Luis Von Ahn and Laura Dabbish. Labeling images with a computer game. In Proceedings of the SIGCHI conference on Human factors in computing systems, pages 319-326, 2004. 6", + "[30] Xu Wang, Dezhong Peng, Peng Hu, and Yongsheng Sang. Adversarial correlated autoencoder for unsupervised multiview representation learning. Knowledge-Based Systems, 168:109-120, 2019. 1", + "[31] Jie Wen, Chengliang Liu, Shijie Deng, Yicheng Liu, Lunke Fei, Ke Yan, and Yong Xu. Deep double incomplete multiview multi-label learning with incomplete labels and missing views. IEEE transactions on neural networks and learning systems, 2023. 1, 2, 6", + "[32] Mike Wu and Noah Goodman. Multimodal generative models for scalable weakly-supervised learning. Advances in neural information processing systems, 31, 2018. 3", + "[33] Xuan Wu, Qing-Guo Chen, Yao Hu, Dengbao Wang, Xiaodong Chang, Xiaobo Wang, and Min-Ling Zhang. Multiview multi-label learning with view-specific information extraction. In IJCAI, pages 3884-3890, 2019. 1, 2", + "[34] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018. 5", + "[35] Xiaoqiang Yan, Shizhe Hu, Yiqiao Mao, Yangdong Ye, and Hui Yu. Deep multi-view learning methods: A review. Neurocomputing, 448:106-129, 2021. 1", + "[36] Jun Yin and Shiliang Sun. Incomplete multi-view clustering with reconstructed views. IEEE Transactions on Knowledge and Data Engineering, 35(3):2671-2682, 2021. 1", + "[37] Jun Yin and Wentao Zhang. Multi-view multi-label learning with double orders manifold preserving. Applied Intelligence, 53(12):14703-14716, 2023. 1", + "[38] Jun Yin, Shiliang Sun, Lai Wei, and Pei Wang. Discriminatively fuzzy multi-view k-means clustering with local structure preserving. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16478-16485, 2024. 1", + "[39] Honglin Yuan, Shiyun Lai, Xingfeng Li, Jian Dai, Yuan Sun, and Zhenwen Ren. Robust prototype completion for incomplete multi-view clustering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10402-10411, 2024. 1", + "[40] Dawei Zhao, Qingwei Gao, Yixiang Lu, Dong Sun, and Yusheng Cheng. Consistency and diversity neural network multi-view multi-label learning. Knowledge-Based Systems, 218:106841, 2021. 2, 6", + "[41] Dawei Zhao, Qingwei Gao, Yixiang Lu, Dong Sun, and Yusheng Cheng. Consistency and diversity neural network multi-view multi-label learning. Knowledge-Based Systems, 218:106841, 2021. 1", + "[42] Yue Zhu, James T Kwok, and Zhi-Hua Zhou. Multi-label learning with global and local label correlation. IEEE Transactions on Knowledge and Data Engineering, 30(6):1081-1094, 2017. 1" + ], + "bbox": [ + 91, + 92, + 480, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30731", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/419f8c28-377b-4f29-8c8b-a2cba093a85b_model.json b/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/419f8c28-377b-4f29-8c8b-a2cba093a85b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d932e40f9debb451dea95d149cc0b8db112b2739 --- /dev/null +++ b/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/419f8c28-377b-4f29-8c8b-a2cba093a85b_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.13, + 0.895, + 0.177 + ], + "angle": 0, + "content": "Incomplete Multi-View Multi-Label Learning via Disentangled Representation and Label Semantic Embedding" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.203, + 0.795, + 0.258 + ], + "angle": 0, + "content": "Xu Yan\\(^{1}\\), Jun Yin\\(^{1,*}\\), Jie Wen\\(^{2}\\) \n\\(^{1}\\)Shanghai Maritime University, \\(^{2}\\)Harbin Institute of Technology, Shenzhen yanxu@stu.shmtu.edu.cn, junyin@shmtu.edu.cn, jiewen_pr@126.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.324, + 0.485, + 0.686 + ], + "angle": 0, + "content": "In incomplete multi-view multi-label learning scenarios, it is crucial to use the incomplete multi-view data to extract consistent and specific representations from different data sources and to fully exploit the missing label information. However, most previous approaches ignore the separation problem between view-shared and specific information. To address this problem, in this paper, we propose a method that can separate view-consistent features from view-specific features under the Variational Autoencoder (VAE) framework. Specifically, we first introduce cross-view reconstruction to capture view-consistent features and extract shared information from different views through unsupervised pre-training. Subsequently, we develop a disentangling module to learn specific features by minimizing the variational upper bound of mutual information between consistent and specific features. Finally, we utilize prior label relevance information derived from training data to guide the learning of the distribution of label semantic embeddings, aggregating relevant semantic embeddings and maintaining the label relevance topology in the semantic space. In extensive experiments, our model outperforms existing state-of-the-art algorithms on several real-world datasets, which fully validates its strong adaptability to missing views and labels." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.716, + 0.222, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.483, + 0.878 + ], + "angle": 0, + "content": "In recent years, with the increasing diversity and complexity of data, multi-view learning has gained widespread attention as an effective data processing technique. In many real-world applications, such as image categorization, video analysis, and text mining, data is often presented in the form of multiple views, which makes multi-view learning methods an effective tool for solving complex problems [28, 35, 36]. In the past few years, many multi-view learning methods have been proposed based on the assumption" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.905, + 0.354 + ], + "angle": 0, + "content": "of view completeness [4, 30, 38]. However, it is not uncommon to have missing views when dealing with multi-view data, which poses a challenge to the learning effectiveness of the model [19, 39]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.906, + 0.506 + ], + "angle": 0, + "content": "Meanwhile, multi-label learning is an equally important research area. Unlike traditional single-label classification tasks, multi-label learning aims to deal with situations where instances may belong to more than one category at the same time. The key to multi-label learning is how to deal with the correlation between labels, which is helpful in improving the classification performance [12, 42]. Multi-label datasets are often incomplete due to various uncertainties involved in manual labeling. So research on incomplete multi-label learning is particularly important." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Traditional multi-label classification relies on features extracted from a single view for classification, which leads to insufficient feature information. Using multi-view data for multi-label classification can effectively alleviate this problem. Currently, there are many models dealing with multi-view multi-label classification (MVMLC) tasks that show promising performance [33, 37, 41]. These models typically assume that the data is complete, and currently, only a few methods are capable of handling both missing views and missing labels simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To deal with the Double Incomplete Multi-View Multi-Label Classification problem (DIMVMLC), several deep learning models based on autoencoders have been proposed [22, 23, 31]. However, there are still only a few methods for disentangling consistent features and specific features. SIMM [33] identifies consistent features by minimizing confusing adversarial loss and multi-label loss, uses orthogonal constraints to eliminate shared information, and extracts specific information from raw information. However, SIMM cannot be applied to the missing data case, and the forced orthogonality constraint eliminates linear correlation, and there may still be a nonlinear relationship between shared and specific features. This relationship may be difficult to capture by simple orthogonality constraints, affecting the expressiveness of the features. MTD [23] simultaneously learns consistent and specific features, sepa" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30722" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "rating them using multi-view cross-channel contrastive loss. Optimizing both tasks simultaneously may lead to a complex optimization process where the model needs to trade off between two objectives, increasing the difficulty of convergence." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.168, + 0.483, + 0.437 + ], + "angle": 0, + "content": "To address the above complex double incomplete multi-view multi-label learning problem, we propose a Disentangled Representation and Label Semantic Embedding based method (DRLS), as shown in Fig. 1. Our approach consists of three main parts: (1) Learning multi-view consistent features by minimizing the KL divergence between the approximate posterior of each view's consistent features and the multi-view joint posterior. (2) Learning the disentangled specific features by minimizing the variational upper bound of mutual information between the consistent features and the specific features. (3) We model the distribution of label semantic embeddings by leveraging prior label relevance information from the training data and utilizing graph neural networks. Our model is built upon the Variational Autoencoder (VAE) framework. Numerous experiments demonstrate the excellent performance of DRLS. Finally, the contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.44, + 0.482, + 0.528 + ], + "angle": 0, + "content": "- We propose a novel approach based on disentangled representation and label semantic embedding for solving the DIMVMLC problem. It can effectively utilize the consistent and specific information of multi-view data, and separate them through a disentangled representation learning module for effective multi-label classification tasks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.53, + 0.483, + 0.605 + ], + "angle": 0, + "content": "- We introduce a pre-training strategy that learns view-consistent representations via pre-training and then minimizes the variational upper bound of mutual information between consistent and specific features to separate them effectively." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.606, + 0.483, + 0.665 + ], + "angle": 0, + "content": "- We design a graph neural network-based label semantic embedding learning module that simultaneously aggregates embeddings of related labels and maintains the label relevance topology in the semantic space." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.666, + 0.483, + 0.711 + ], + "angle": 0, + "content": "- Our approach achieves leading performance on five publicly available datasets and can be applied to both complete and arbitrarily missing data." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.44, + 0.483, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.236, + 0.74 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Many MVMLC methods have been proposed. For example, CDMM [40] builds independent classifiers based on neural networks constructed with nonlinear kernel mapping to ensure that each view learns consistent labeling results. The diversity across views is addressed by integrating the Hilbert-Schmidt independence criterion [7] with an exploration of the variations among different views. MMFA [3] initially extracts both common and private features from different views, then computes the similarity between the original features and the common features. The attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "weight matrix is derived by multiplying this similarity with the synergistic feature matrix, which is subsequently used for multi-label prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.44 + ], + "angle": 0, + "content": "For DIMVMLC, iMvWL [27] uses Nonnegative Matrix Factorization (NMF) to learn the shared subspace of multiple views and introduces label information and local label correlation matrices to improve the discriminative power of the subspace. NAIML [20] exploits both consistency among multiple views and the complex global and local structure among labels to cope with the lack of available labels. Both of these methods are shallow machine learning approaches, and current deep neural network-based methods show better performance in this area, such as DDINet [31], DICNet [22], and SIP [21]. DDINet is the first to use a deep neural network to deal with DIMVMLC. DICNet leverages contrastive loss to learn shared features across views, aligns cross-view instances, and improves feature discrimination through a multi-view fusion module, effectively handling missing views and labels. The SIP model utilizes the principles of the information bottleneck theory to reduce the non-shared information while preserving the relevant features, and captures the associations between labels through prototype instances in the latent space." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.454, + 0.605, + 0.47 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.48, + 0.694, + 0.495 + ], + "angle": 0, + "content": "3.1. Problem definition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.685 + ], + "angle": 0, + "content": "Given a dataset \\(\\{x^{(v)}\\}_{v=1}^m\\) consisting of \\(n\\) samples, each with \\(m\\) views, where \\(x^{(v)} \\in \\mathbb{R}^{d_v}\\). Here \\(d_v\\) denotes the feature dimension of the \\(v\\)th view. The label vector is \\(y \\in \\{0,1\\}^k\\), where \\(k\\) denotes the number of categories. If \\(y_i = 0\\), the sample does not have the \\(i\\)-th label; if \\(y_i = 1\\), it does. To handle missing data, we define the set \\(\\mathcal{W}(|\\mathcal{W}| \\leq m)\\) to denote the set of observable views of a sample, and the observable multi-view data can be defined as \\(\\{x^{(v)}\\}_{v \\in \\mathcal{W}}\\), abbreviated as \\(\\{x\\}\\). We also define a set of observable labels for a sample \\(\\mathcal{U}(|\\mathcal{U}| \\leq k)\\). Our goal is to train a model to predict the labels of unseen samples. Our loss and objective functions are defined in the case of a single sample." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.731, + 0.71 + ], + "angle": 0, + "content": "3.2. View-consistent feature" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.809 + ], + "angle": 0, + "content": "In traditional multi-view learning, consistent features are often learned separately without cross-view interaction [33], making it hard to fully capture the common semantic information across views. In this subsection, we first introduce how to learn multi-view consistent features through cross-view reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "First, we define \\( q_{\\phi_v}(c \\mid x^{(v)}) \\) with learnable parameters \\( \\phi_v \\) to approximate the posterior \\( p(c \\mid x^{(v)}) \\), where \\( c \\) represents the consistent feature. The consistent features of each view should contain the public information of all views, inspired by [2], we learn the consistent features by constraining the approximate posterior of each view to be similar to" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30723" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.294, + 0.107 + ], + "angle": 0, + "content": "the multi-view joint posterior:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.113, + 0.483, + 0.149 + ], + "angle": 0, + "content": "\\[\n\\min \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ D _ {K L} \\left(q _ {\\phi_ {v}} (c \\mid x ^ {(v)}) \\| p (c \\mid \\{x \\})\\right) \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.2 + ], + "angle": 0, + "content": "where \\(D_{KL}(\\cdot \\|\\cdot)\\) represents the KL divergence, \\(|\\mathcal{W}|\\) denotes the number of available views, and \\(p(c\\mid \\{x\\})\\) denotes the multi-view joint posterior." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.484, + 0.276 + ], + "angle": 0, + "content": "The Evidence Lower Bound (ELBO) is derived from Eq. (1) by further applying Bayes' theorem to decompose the multi-view joint posterior distribution \\( p(c \\mid \\{x\\}) \\). Due to space constraints, the exact derivation is given in the supplementary material Sec. 6.1." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.277, + 0.305, + 0.29 + ], + "angle": 0, + "content": "The final result is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.293, + 0.482, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\log p (\\{x \\}) - \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ D _ {K L} \\left(q _ {\\phi_ {v}} (c | x ^ {(v)}) \\| p (c | \\{x \\})\\right) \\right] \\\\ = \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ \\mathbb {E} _ {q _ {\\phi_ {v}} (c | x ^ {(v)})} [ \\log p (\\{x \\} | c) ] \\right. \\\\ \\left. - D _ {K L} \\left(q _ {\\phi_ {v}} (c \\mid x ^ {(v)}) \\| p (c)\\right) \\right] \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.455 + ], + "angle": 0, + "content": "This establishes the Evidence Lower Bound (ELBO). Thus, by maximizing this lower bound, we are effectively solving the minimization problem in Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.655 + ], + "angle": 0, + "content": "By assuming each view is conditionally independent from the others given \\( c \\), we can efficiently decompose the data likelihood as \\( p(\\{x\\} |c) = \\prod_{i\\in \\mathcal{W}}p\\left(x^{(i)}|c\\right)\\approx \\prod_{i\\in \\mathcal{W}}p_{\\theta_i}\\left(x^{(i)}|c\\right) \\), where \\( p_{\\theta_i}\\left(x^{(i)}|c\\right) \\) with learnable parameters \\( \\theta_{i} \\) approximates \\( p(x^{(i)}|c) \\). The objective in Eq. (2) involves extracting consistent features from different views. To facilitate subsequent analysis, we introduce an identifier for the consistent features of each view to distinguish them, i.e., \\( \\{c^{(v)}\\}_{v = 1}^m \\). Introducing this identifier does not alter the model architecture. Additionally, we introduce a trade-off parameter \\( \\alpha \\) [10], which controls the trade-off between reconstruction quality and regularization, resulting in the following lower bound:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.674, + 0.482, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n s i s t}} = \\\\ \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ \\mathbb {E} _ {q _ {\\phi_ {v}} (c ^ {(v)} | x ^ {(v)})} \\left[ \\sum_ {i \\in \\mathcal {W}} \\log p _ {\\theta_ {i}} (x ^ {(i)} | c ^ {(v)}) \\right] \\right. \\\\ \\left. - \\alpha D _ {K L} \\left(q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) \\| p (c)\\right) \\right] \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.484, + 0.877 + ], + "angle": 0, + "content": "here \\(q_{\\phi_v}(c^{(v)} \\mid x^{(v)})\\) can be regarded as an encoder, and we assume that \\(q_{\\phi_v}(c^{(v)} \\mid x^{(v)}) \\sim \\mathcal{N}(\\mu_{\\phi_v}(x^{(v)}), \\sigma_{\\phi_v}^2(x^{(v)}) I)\\) where \\(\\mu_{\\phi_v}(x^{(v)})\\) and \\(\\sigma_{\\phi_v}^2(x^{(v)})\\) denote the mean and variance encoder of the \\(v\\)th view, and \\(I\\) denotes the identity matrix. Due to the non-differentiable nature of the sampling process, we introduce the reparameterization trick [16] to obtain the features:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nc ^ {(v)} = \\mu_ {\\phi_ {v}} \\left(x ^ {(v)}\\right) + \\sigma_ {\\phi_ {v}} \\left(x ^ {(v)}\\right) \\odot \\delta \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "where \\(\\delta \\sim \\mathcal{N}(0,I)\\), \\(\\odot\\) denotes the Hadamard product, and \\(c^{(v)}\\in \\mathbb{R}^{d_e}\\), where \\(d_{e}\\) represents the feature dimension. The distribution \\(p_{\\theta_i}(x^{(i)}|c^{(v)})\\) in Eq. (3) is modeled as an MLP decoder, assuming \\(p_{\\theta_i}(x^{(i)}|c^{(v)})\\sim \\mathcal{N}(\\mu_{\\theta_i}(c^{(v)}),I)\\), yielding the reconstructed feature \\(\\hat{x}^{(i,v)} = \\mu_{\\theta_i}(c^{(v)})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.906, + 0.287 + ], + "angle": 0, + "content": "A brief analysis of Eq. (3) is as follows: The first term represents both self- and cross-view reconstruction. It allows \\( c^{(v)} \\) from each view to reconstruct the original data of different views, which helps \\( c^{(v)} \\) capture shared information. The second term, the KL divergence, aligns the posterior \\( q_{\\phi_v}(c^{(v)} \\mid x^{(v)}) \\) of each view with the shared prior \\( p(c) \\sim \\mathcal{N}(0,I) \\). These two terms simultaneously encourage \\( c^{(v)} \\) to encode the shared information between views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.906, + 0.41 + ], + "angle": 0, + "content": "Finally, in order to obtain a fused consistent feature shared by multiple views for the classification task, we compute the multi-view joint posterior \\( p(c|\\{x\\}) \\) using the Product-of-Expert (PoE) approach [11]. This method is flexible and effective in dealing with the missing views by directly removing the corresponding parameter when a view is missing. According to the setting in [32], we compute \\( p(c|\\{x\\}) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.416, + 0.905, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p (c | \\{x \\}) \\propto p (c) \\prod_ {v \\in \\mathcal {W}} p (c \\mid x ^ {(v)}) \\\\ := p (c) \\prod_ {v \\in \\mathcal {W}} q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.906, + 0.537 + ], + "angle": 0, + "content": "The same reparameterization trick is used to obtain the fused multi-view consistent feature \\(\\bar{c} \\in \\mathbb{R}^{d_e}\\), which is sampled from \\(p(c|\\{x\\})\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.81, + 0.56 + ], + "angle": 0, + "content": "3.3. Disentangled view-specific feature" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.779 + ], + "angle": 0, + "content": "In this section, we introduce view-specific features \\(\\{s^{(v)}\\}_{v = 1}^{m}\\) that encode information specific to each view beyond the consistent features. We consider minimizing the mutual information \\(I(c^{(v)};s^{(v)})\\) between the consistent and specific features for each view, which allows us to maximally disentangle the specific features from the consistent features. A natural idea is to do the learning of consistent features first, followed by disentangling the specific features from them [15]. Thus, we first pre-train with \\(\\mathcal{L}_{\\text{consist}}\\) for a certain number of epochs, depending on the dataset size. In the second stage, we decrease the learning rate for the encoders of consistent features \\(\\{q_{\\phi_v}(c^{(v)}\\mid x^{(v)})\\}_{v = 1}^m\\) to stabilize the pre-trained features during further disentanglement. Our disentangling objective function is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.786, + 0.905, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\min \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} I \\left(c ^ {(v)}; s ^ {(v)}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.905, + 0.862 + ], + "angle": 0, + "content": "We rewrite \\( I(c^{(v)}; s^{(v)}) \\) as follows, more details are provided in the supplementary material Sec. 6.2:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.868, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\[\nI \\left(c ^ {(v)}; s ^ {(v)}\\right) = I \\left(x ^ {(v)}; c ^ {(v)}\\right) + I \\left(x ^ {(v)}; s ^ {(v)}\\right) - I \\left(x ^ {(v)}; c ^ {(v)}, s ^ {(v)}\\right) \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30724" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.47, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.278, + 0.23, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.279, + 0.342, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.279, + 0.489, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.09, + 0.904, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.284, + 0.904, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.412, + 0.907, + 0.455 + ], + "angle": 0, + "content": "Figure 1. The main framework of DRLS. The upper-left corner of the figure represents the view-consistent feature learning module, the upper-right corner shows the disentangled view-specific feature learning module, and the lower part depicts the label semantic embedding learning module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.525 + ], + "angle": 0, + "content": "Clearly, directly optimizing Eq. (7) is challenging. Therefore, we introduce the variational upper bounds for \\( I(x^{(v)};c^{(v)}) \\) and \\( I(x^{(v)};s^{(v)}) \\), as well as the variational lower bound for \\( I(x^{(v)};c^{(v)},s^{(v)}) \\) [14], resulting in:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.528, + 0.482, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I \\left(c ^ {(v)}; s ^ {(v)}\\right) \\leq D _ {K L} \\left(q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) \\| p \\left(c ^ {(v)}\\right)\\right) \\\\ + D _ {K L} \\left(q _ {\\varphi_ {v}} \\left(s ^ {(v)} \\mid x ^ {(v)}\\right) \\| p \\left(s ^ {(v)}\\right)\\right) \\\\ - \\mathbb {E} _ {q _ {\\phi_ {v}} (c ^ {(v)} | x ^ {(v)}) q _ {\\varphi_ {v}} (s ^ {(v)} | x ^ {(v)})} [ \\log p _ {\\vartheta_ {v}} (x ^ {(v)} | c ^ {(v)}, s ^ {(v)}) ] \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.773 + ], + "angle": 0, + "content": "We observe that minimizing this variational upper bound on mutual information trains a decoder \\( p_{\\vartheta_v}(x^{(v)}|c^{(v)}, s^{(v)}) \\) and a specific feature encoder \\( q_{\\varphi_v}(s^{(v)}|x^{(v)}) \\). Here, we concatenate \\( c^{(v)} \\) and \\( s^{(v)} \\) as input to the MLP decoder \\( p_{\\vartheta_v}(x^{(v)}|c^{(v)}, s^{(v)}) \\) to produce the reconstructed feature \\( \\hat{x}^{(v)} \\). Similar to Sec. 3.2, \\( q_{\\varphi_v}(s^{(v)}|x^{(v)}) \\sim \\mathcal{N}(\\mu_{\\varphi_v}(x^{(v)}), \\sigma_{\\varphi_v}^2(x^{(v)})I) \\) parameterizes the distribution of the specific features. The feature \\( s^{(v)} \\in \\mathbb{R}^{d_e} \\) is derived using the reparameterization trick as shown in Eq. (4). We assume that \\( p(c^{(v)}) \\) and \\( p(s^{(v)}) \\) follow standard Gaussian distributions. Finally, we obtain the disentanglement loss:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.778, + 0.482, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i s e n t}} = \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ D _ {K L} \\left(q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) \\| p \\left(c ^ {(v)}\\right)\\right) \\right. \\\\ + D _ {K L} \\left(q _ {\\varphi_ {v}} \\left(s ^ {(v)} \\mid x ^ {(v)}\\right) \\| p \\left(s ^ {(v)}\\right)\\right) \\\\ \\left. - \\mathbb {E} _ {q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) q _ {\\varphi_ {v}} \\left(s ^ {(v)} \\mid x ^ {(v)}\\right)} \\left[ \\log p _ {\\vartheta_ {v}} \\left(x ^ {(v)} \\mid c ^ {(v)}, s ^ {(v)}\\right) \\right] \\right] \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Comparing Eq. (3) and Eq. (9), the consistent objective function requires each consistent feature \\( c^{(v)} \\) to reconstruct" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.645 + ], + "angle": 0, + "content": "the original features across views, whereas in the disentanglement loss, we use the concatenation of consistent and specific features exclusively for self-view reconstruction. Since we pre-train a set of encoders \\(\\{q_{\\phi_v}(c^{(v)}|x^{(v)})\\}_{v = 1}^m\\) in the second stage, \\(c^{(v)}\\) and \\(s^{(v)}\\) are concatenated to reconstruct \\(x^{(v)}\\). As \\(c^{(v)}\\) already contains the information shared between \\(x^{(v)}\\) and other views, \\(s^{(v)}\\) only needs to encode the complementary information that is not captured by \\(c^{(v)}\\). This pre-training strategy enables more efficient disentanglement. We use a simple and effective method to obtain the fused multi-view specific feature \\(\\bar{s}\\in \\mathbb{R}^{d_e}\\), for subsequent classification tasks:" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.651, + 0.907, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\bar {s} = \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} s ^ {(v)} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.907, + 0.83 + ], + "angle": 0, + "content": "Here, we do not use PoE to fuse the specific features as in Sec. 3.2, because it is more suitable for merging shared information. PoE is sensitive to consensus among different views [18]. Specifically, if one view's probability distribution is overly sharp and biased, the joint distribution will have a biased mean prediction [1, 18, 26]. Given that specific features from different views often have differences, we choose a simple average fusion method to mitigate extreme variations and retain the overall trend." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.906, + 0.875 + ], + "angle": 0, + "content": "Finally, we fuse the multi-view consistent and specific information. Following the method in [23], we use specific features to enhance the consistent features:" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\[\nz = \\omega (\\bar {s}) \\odot \\bar {c} \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30725" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.138 + ], + "angle": 0, + "content": "\\(z \\in \\mathbb{R}^{d_e}\\) represents the final fused representation, \\(\\omega(\\cdot)\\) is the sigmoid activation function, and \\(\\odot\\) denotes the Hadamard product." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.146, + 0.336, + 0.163 + ], + "angle": 0, + "content": "3.4. Label semantic embedding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.485, + 0.274 + ], + "angle": 0, + "content": "To fully utilize the incomplete supervisory information, we introduce label semantic embedding learning guided by prior information. Here, we construct our model within the framework of the Variational Graph Autoencoder (VGAE) [17], which employs variational inference to model the latent distribution of label embeddings during the learning process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.275, + 0.485, + 0.456 + ], + "angle": 0, + "content": "First, we define a relationship graph among labels, where the graph \\( G = (V,E) \\) represents the label relationship graph, with \\( V \\) denoting the nodes (categories) and \\( E \\) representing the edges connecting the nodes. We model the adjacency matrix \\( A \\in \\mathbb{R}^{k \\times k} \\) of graph \\( G \\) in terms of conditional probabilities, specifically \\( P(l_j|l_i) \\), which denotes the probability that label \\( l_j \\) occurs when label \\( l_i \\) occurs, where \\( l_i \\) represents the \\( i \\)th label. This is computed based on the label matrix \\( Y \\in \\mathbb{R}^{n_{train} \\times k} \\) from the training set. Since \\( P(l_j|l_i) \\) is not equal to \\( P(l_i|l_j) \\), the adjacency matrix \\( A \\) is asymmetric, which more accurately reflects real-world scenarios. The specific calculation method is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.466, + 0.484, + 0.507 + ], + "angle": 0, + "content": "\\[\nA _ {i j} = \\left\\{ \\begin{array}{l l} P \\left(l _ {j} \\mid l _ {i}\\right) = \\frac {C _ {i j}}{d _ {i}}, & \\text {i f} i \\neq j \\\\ 0, & \\text {i f} i = j \\end{array} \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.605 + ], + "angle": 0, + "content": "where \\(C = Y^{\\top}Y\\), \\(C \\in \\mathbb{R}^{k \\times k}\\) is the label co-occurrence matrix, \\(C_{ij}\\) is the co-occurrence count of labels \\(l_i\\) and \\(l_j\\), \\(d_i = \\sum_{a=1}^{n_{train}} Y_{ai} = Y_{:i}^{\\top} Y_{:i}\\) for \\(i = 1,2,\\ldots,k\\), which represents the total occurrence count of label \\(l_i\\). It should be noted that we set the diagonal of the adjacency matrix \\(A\\) to 0." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.485, + 0.682 + ], + "angle": 0, + "content": "Graph Isomorphism Network (GIN) provides stronger feature aggregation capabilities through learnable weights and information from self-nodes [34]. Therefore, we use GIN to model the posterior distribution of label semantic embeddings \\( H \\in \\mathbb{R}^{k \\times d_e} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.693, + 0.483, + 0.734 + ], + "angle": 0, + "content": "\\[\nq \\left(H \\mid H ^ {0}, A\\right) = \\prod_ {i = 1} ^ {k} q \\left(h _ {i} \\mid H ^ {0}, A\\right), \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.736, + 0.396, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathrm {w i t h} q (h _ {i} | H ^ {0}, A) \\sim \\mathcal {N} (\\mu_ {i}, \\sigma_ {i} ^ {2} I)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.901 + ], + "angle": 0, + "content": "here, \\( h_i \\in \\mathbb{R}^{d_e} \\) is the \\( i \\)-th row of \\( H \\), representing the semantic embedding of the \\( i \\)-th label, and \\( H^0 \\in \\mathbb{R}^{k \\times k} \\) is an identity matrix, which means that the embedding for label \\( i \\) is initialized as a one-hot vector with the \\( i \\)-th position set to 1. This approach ensures that the initial distance between different labels is large, which helps to better distinguish the semantic differences between labels during the subsequent learning process. \\( \\pmb{\\mu} = \\mathrm{GIN}_{\\pmb{\\mu}}(H^0, A) \\), \\( \\pmb{\\mu} \\in \\mathbb{R}^{k \\times d_e} \\) is the mean matrix, where the \\( i \\)-th row corresponds to the mean" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.906, + 0.198 + ], + "angle": 0, + "content": "vector \\(\\mu_{i}\\) for the \\(i\\)-th label; similarly, \\(\\sigma^2 = \\mathrm{GIN}_{\\sigma^2}(H^0,A)\\). Here, we use a three-layer GIN, where one layer can be defined as \\(H^{(t + 1)} = f^{(t + 1)}[(1 + \\epsilon^{(t + 1)})H^{(t)} + AH^{(t)}]\\), \\(t\\) represents the current number of network layers, \\(\\epsilon^{(t + 1)}\\) represents a learnable parameter, and \\(f^{(t + 1)}\\) represents a two-layer fully connected network. \\(H^{(t + 1)}\\in \\mathbb{R}^{k\\times d_{t + 1}}\\) represents the output of the intermediate layer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.906, + 0.258 + ], + "angle": 0, + "content": "Subsequently, we also use the reparameterization trick to sample the label semantic embeddings, with a formula similar to Eq. (4), i.e., \\( h_i = \\mu_i + \\sigma_i \\odot \\delta, \\delta \\sim \\mathcal{N}(0, I) \\). Finally, we obtain the label semantic embedding matrix \\( H \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.905, + 0.289 + ], + "angle": 0, + "content": "We use a simple and efficient inner product between variables as a generative model:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.3, + 0.905, + 0.343 + ], + "angle": 0, + "content": "\\[\np (A | H) = \\prod_ {i = 1} ^ {k} \\prod_ {j = 1} ^ {k} p \\left(A _ {i j} \\mid h _ {i}, h _ {j}\\right), \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.345, + 0.821, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\text {w i t h} p \\left(A _ {i j} \\mid h _ {i}, h _ {j}\\right) = \\omega \\left(h _ {i} ^ {\\top} h _ {j}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.527 + ], + "angle": 0, + "content": "where \\(\\omega(\\cdot)\\) denotes the sigmoid activation function. By using the label correlation matrix as the adjacency matrix of a graph, we enable GIN to aggregate semantically related embeddings to nearby positions in the label semantic space. However, it cannot guarantee that unrelated labels are effectively pushed apart. Therefore, we introduce the generative model from VGAE to reconstruct the adjacency matrix \\(A\\), ensuring that the label semantic embeddings effectively preserve the topology of label correlations in the semantic space. Finally, our variational lower bound is:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.538, + 0.905, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l e} = \\mathbb {E} _ {q (H | H ^ {0}, A)} [ \\log p (A | H) ] - D _ {K L} (q (H | H ^ {0}, A) \\| p (H)) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.57, + 0.906, + 0.646 + ], + "angle": 0, + "content": "We further adopt a Gaussian prior \\( p(H) = \\prod_{i}p(h_{i}) = \\prod_{i}\\mathcal{N}(h_{i}|0,I) \\). We use binary cross-entropy to compute the first term of \\( \\mathcal{L}_{le} \\) as the reconstruction loss and add an identity matrix to \\( A \\) before computing the cross-entropy loss to avoid enforcing \\( \\omega (h_i^\\top h_i)\\approx 0 \\) due to its zero diagonal." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.655, + 0.71, + 0.67 + ], + "angle": 0, + "content": "3.5. Overall loss function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.907, + 0.8 + ], + "angle": 0, + "content": "Finally, we utilize label semantic embeddings for feature selection, choosing discriminative and specific features for each label. First, the learned label semantic embedding \\( h_i \\) is mapped to a feature importance vector [9] via a fully connected layer \\( f(\\cdot) \\), specifically \\( r_i = \\omega(f(h_i)) \\), where \\( r_i \\in \\mathbb{R}^{d_e} \\) and \\( \\omega(\\cdot) \\) is the sigmoid activation function. Subsequently, we use \\( r_i \\) to select the features and input them into a simple neural network to obtain predictions:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.812, + 0.905, + 0.829 + ], + "angle": 0, + "content": "\\[\np _ {i} = \\omega \\left(f _ {\\text {c l s}} \\left(z \\odot r _ {i}\\right)\\right) \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes the Hadamard product, \\(f_{cls}\\) is a fully connected network, and \\(p_i\\) is the predicted probability of the sample being labeled as the \\(i\\)-th class. We use the commonly used binary cross-entropy function to calculate the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30726" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.148 + ], + "angle": 0, + "content": "Table 1. Statistics for five multi-view multi-label datasets, where \\( m \\) denotes the number of views, \\( n \\) denotes the number of samples, \\( k \\) denotes the number of categories, and \\( n/k \\) denotes the average number of labels per sample." + }, + { + "type": "table", + "bbox": [ + 0.142, + 0.158, + 0.437, + 0.257 + ], + "angle": 0, + "content": "
Datasetmnkn/k
Corel5k649992603.396
Pascal0769963201.465
Espgame6207702684.686
IAPRTC126196272915.719
Mirflickr625000384.716
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.281, + 0.295, + 0.295 + ], + "angle": 0, + "content": "multi-label classification loss:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.304, + 0.482, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B C E} = - \\frac {1}{| \\mathcal {U} |} \\sum_ {i \\in \\mathcal {U}} \\left(y _ {i} \\log p _ {i} + \\left(1 - y _ {i}\\right) \\log \\left(1 - p _ {i}\\right)\\right) \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.485, + 0.4 + ], + "angle": 0, + "content": "where \\( y_{i} \\) denotes the ground truth and we exclude missing labels from loss computation to filter out the impact of missing labels. Our final loss function is:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.412, + 0.483, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {B C E} + \\beta \\mathcal {L} _ {\\text {d i s e n t}} - \\gamma \\mathcal {L} _ {l e} \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.471 + ], + "angle": 0, + "content": "where \\(\\beta\\) and \\(\\gamma\\) are trade-off parameters used to balance the different loss terms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.486, + 0.224, + 0.503 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.511, + 0.196, + 0.525 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.671 + ], + "angle": 0, + "content": "We conduct experiments based on the settings from [21, 23, 31] to evaluate the performance of our proposed model on five datasets. They are Corel5k [5], Pascal07 [6], Espgame [29], IAPRTC12 [8], and Mirflickr [13]. In these five datasets, we utilize six different types of features, which are DenseHue, DenseSift, GIST, HSV, RGB, and LAB, with dimensions of 100, 1000, 512, 4096, 4096, and 4096, respectively. More detailed information can be found in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.28, + 0.697 + ], + "angle": 0, + "content": "4.2. Compared methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.903 + ], + "angle": 0, + "content": "In our experiments, we select eight methods for comparison with our model: DM2L [24], CDMM [40], iMvWL [27], NAIML [20], DDINet [31], DICNet [22], MTD [23], and SIP [21]. It is important to note that iMvWL, NAIML, DDINet, DICNet, MTD, and SIP are incomplete multi-view multi-label methods. CDMM can only handle complete multi-view multi-label data; therefore, we follow the settings from [21] and use the mean of available instances from the corresponding view to fill in the unavailable instances for missing views. To handle incomplete labels, we treat the missing labels as negative labels. DM2L is a single-view incomplete multi-label method, and we record the best result of each view in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.729, + 0.107 + ], + "angle": 0, + "content": "4.3. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.114, + 0.907, + 0.283 + ], + "angle": 0, + "content": "For the datasets, in order to simulate real-world data missingness, we generate incomplete datasets following the settings in [27] and [21]. For each view, we randomly exclude \\(50\\%\\) of the instances, ensuring at least one view per sample. Additionally, we randomly remove \\(50\\%\\) of both positive and negative labels. We fill the missing features and labels with zeros. We use \\(70\\%\\) of the samples for training and the remaining \\(30\\%\\) for validation and testing. To ensure the stability of the results, we repeat the experiment 10 times and report the mean and variance of the results. Additional details are in the supplementary material Sec. 7." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.288, + 0.697, + 0.303 + ], + "angle": 0, + "content": "4.4. Evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.311, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Consistent with some previous studies [21, 23, 31], in our experiments, we evaluate the performance of the model using six commonly used multi-label classification metrics, namely, Average Precision (AP), Area Under the Receiver Operating Characteristic Curve (AUC), Ranking Loss (RL), Hamming Loss (HL), OneError (OE), and Coverage (Cov). To better visualize the results, we use 1-RL, 1-HL, 1-OE, and 1-Cov to record the final results, where larger values indicate better performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.455, + 0.712, + 0.47 + ], + "angle": 0, + "content": "4.5. Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.907, + 0.675 + ], + "angle": 0, + "content": "In Tab. 2, we present the experimental results of nine methods on five public multi-view multi-label datasets, which include six evaluation metrics. The results in Tab. 2 are obtained under experimental settings with a \\(50\\%\\) view missing rate, a \\(50\\%\\) label missing rate, and \\(70\\%\\) of training samples. The results in Tab. 2 show that our method outperforms the other eight competitors on each metric in most datasets. For example, on the AP metric of the IAPRTC12 dataset, our method is \\(2.4\\%\\) higher than the second-ranked MTD. Specifically, Ave.R represents the average ranking of each method across the six metrics, and DRLS ranks first on all five datasets, which fully demonstrates the effectiveness of our method in addressing the DIMVMLC problem." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.907, + 0.75 + ], + "angle": 0, + "content": "In addition, further analysis of the experimental results reveals that CDMM, DDINet, DICNet, MTD, and SIP outperform the other three methods on all five datasets. These are deep learning methods, indicating the great potential of deep learning in addressing the DIMVMLC problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Besides solving the DIMVMLC problem, we also further validated the performance of our method in addressing the MVMLC problem. In Tab. 3, we present the experimental results of four methods on five datasets, with experiments conducted under the settings of \\(70\\%\\) training samples, full views, and full labels. We select three representative evaluation metrics. From the results in Tab. 3, it is evident that our model still demonstrates competitive performance under this setup, indicating that our method can effectively address the MVMLC problem." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "30727" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.905, + 0.118 + ], + "angle": 0, + "content": "Table 2. Experimental results of nine methods on five datasets with \\(50\\%\\) missing views, \\(50\\%\\) missing labels, and \\(70\\%\\) training samples. Ave.R denotes the average ranking of each method on the six metrics, and the numbers at the bottom right are the standard deviations." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.12, + 0.905, + 0.673 + ], + "angle": 0, + "content": "
DatasetMetricDM2LCDMMiMvWLNAIMLDDINetDICNetMTDSIPDRLS
Corel5kAP0.2620.0050.3540.0040.2830.0080.3090.0040.3600.0090.3780.0040.4130.0070.4160.0150.4330.008
1-HL0.9870.0000.9870.0000.9780.0000.9870.0000.9870.0000.9870.0000.9880.0000.9880.0000.9880.000
1-RL0.8430.0020.8840.0030.8650.0050.8780.0020.8650.0050.8770.0040.8920.0040.9100.0030.9160.002
AUC0.8450.0020.8880.0030.8680.0050.8810.0020.8680.0050.8810.0030.8950.0040.9120.0030.9180.002
1-OE0.2950.0140.4100.0070.3110.0150.3500.0090.4370.0120.4640.0120.4910.0100.4920.0180.5090.019
1-Cov0.6470.0050.7230.0070.7020.0080.7250.0050.6890.0120.7140.0100.7480.0090.7860.0070.8040.006
Ave.R8.1674.8337.6675.3336.0004.8332.6671.8331.000
Pascal07AP0.4710.0080.5080.0050.4370.0180.4880.0030.5320.0100.5020.0070.5500.0040.5500.0090.5670.008
1-HL0.9280.0010.9310.0010.8820.0040.9280.0010.9320.0010.9300.0010.9320.0010.9310.0020.9340.001
1-RL0.7610.0050.8120.0040.7360.0150.7830.0010.8080.0050.7810.0070.8300.0030.8250.0060.8430.004
AUC0.7790.0040.8380.0030.7670.0150.8110.0010.8290.0040.8050.0060.8490.0040.8450.0050.8640.003
1-OE0.4200.0110.4190.0080.3620.0230.4210.0060.4480.0150.4260.0130.4570.0080.4630.0120.4770.011
1-Cov0.6920.0040.7590.0030.6770.0150.7270.0020.7570.0050.7280.0070.7830.0040.7770.0050.7980.004
Ave.R7.6674.8339.0006.5004.1676.1672.1672.8331.000
EspaugeAP0.2120.0020.2890.0030.2440.0050.2460.0020.2860.0040.2990.0040.3060.0030.3100.0040.3260.005
1-HL0.9820.0000.9830.0000.9720.0000.9830.0000.9830.0000.9830.0000.9830.0000.9830.0000.9830.000
1-RL0.7810.0010.8320.0010.8080.0020.8180.0020.8150.0030.8330.0030.8370.0010.8490.0020.8580.002
AUC0.7850.0010.8360.0010.8130.0020.8240.0020.8190.0030.8370.0020.8420.0010.8530.0020.8620.002
1-OE0.2940.0060.3960.0050.3430.0130.3390.0030.4270.0100.4370.0100.4480.0060.4510.0120.4730.001
1-Cov0.4880.0030.5740.0040.5480.0040.5710.0030.5530.0050.5980.0060.6010.0040.6280.0040.6520.003
Ave.R8.8334.5008.0005.6675.5003.5002.6671.8331.000
IAPRTC12AP0.2340.0030.3050.0040.2370.0030.2610.0010.3020.0050.3270.0050.3320.0020.3310.0070.3560.006
1-HL0.9800.0000.9810.0000.9690.0000.9800.0000.9800.0000.9800.0000.9810.0000.9810.0000.9810.000
1-RL0.8230.0020.8620.0020.8330.0020.8480.0010.8530.0020.8720.0020.8750.0010.8870.0040.8960.003
AUC0.8250.0010.8640.0020.8350.0010.8500.0010.8550.0030.8730.0010.8760.0010.8880.0030.8980.002
1-OE0.3400.0060.4320.0080.3520.0080.3900.0050.4350.0090.4650.0130.4710.0060.4660.0010.4900.012
1-Cov0.5290.0040.5970.0040.5640.0050.5920.0040.5940.0070.6480.0050.6490.0020.6790.0080.7070.007
Ave.R8.3334.5008.1676.6675.6674.1672.3332.1671.000
MirfickrAP0.5140.0060.5700.0020.4900.0120.5510.0020.5880.0030.5860.0050.6080.0040.6150.0040.6300.005
1-HL0.8780.0010.8860.0010.8390.0020.8820.0010.8880.0010.8880.0010.8910.0010.8910.0010.8950.001
1-RL0.8310.0030.8560.0010.8030.0080.8440.0010.8650.0020.8610.0040.8750.0010.8780.0020.8850.002
AUC0.8280.0030.8460.0010.7870.0120.8370.0010.8530.0020.8480.0040.8610.0020.8640.0020.8720.003
1-OE0.5100.0080.6310.0040.5110.0220.5850.0030.6360.0080.6420.0060.6560.0040.6640.0060.6860.009
1-Cov0.6040.0050.6400.0010.5720.0130.6310.0020.6540.0030.6460.0060.6770.0020.6760.0040.6920.003
Ave.R8.1676.0008.8337.0004.1674.6672.6672.1671.000
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.275, + 0.697 + ], + "angle": 0, + "content": "4.6. Parameter analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.81 + ], + "angle": 0, + "content": "In our DRLS model, there are three hyperparameters that need to be tuned: \\(\\alpha\\), \\(\\beta\\), and \\(\\gamma\\). To evaluate the sensitivity of our model to these hyperparameters, we conduct experiments on the Corel5k and Pascal07 datasets, with experiments carried out under conditions where \\(50\\%\\) of the views are missing, \\(50\\%\\) of the labels are missing, and \\(70\\%\\) of the training samples are used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Figs. 2a and 2b show the AP values of our model under different combinations of \\(\\beta\\) and \\(\\gamma\\), while Figs. 2c and 2d show the AP and AUC values for different values of \\(\\alpha\\). To ensure the validity of the experiments, the other unrelated hyperparameters are fixed. Based on Figs. 2a and 2b, for the Corel5k dataset, \\(\\beta\\) and \\(\\gamma\\) are taken in the range of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.907, + 0.804 + ], + "angle": 0, + "content": "\\([1e - 3, 1e - 1]\\) and \\([1e - 5, 1e - 3]\\), respectively, and for the Pascal07 dataset, \\(\\beta\\) and \\(\\gamma\\) are taken in the range of \\([1e - 3, 1e - 1]\\) and \\([1e - 5, 1e - 3]\\), respectively. Under these conditions, our model demonstrates good performance. Regarding the hyperparameter \\(\\alpha\\), from Figs. 2c and 2d, it can be seen that the variation in AP and AUC values is minimal for both datasets, indicating that our model is not sensitive to \\(\\alpha\\). We set \\(\\alpha\\) within the range of \\([1e - 1, 3e - 1]\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.664, + 0.832 + ], + "angle": 0, + "content": "4.7. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Our model consists of multiple modules and objective functions. To further analyze the contribution of each module and objective function, we conduct a series of ablation experiments on the Corel5k and Pascal07 datasets under a set-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30728" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.119 + ], + "angle": 0, + "content": "Table 3. Experimental results of four methods on five datasets with full views, full labels, and \\(70\\%\\) training samples." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.124, + 0.462, + 0.372 + ], + "angle": 0, + "content": "
DatasetMetricDICNetMTDSIPDRLS
Corel5kAP0.5040.4860.5200.551
1-RL0.9250.9210.9370.947
AUC0.9280.9240.9390.949
Pascal07AP0.6090.6520.6450.655
1-RL0.8600.8900.8870.889
AUC0.8750.8980.8990.901
EspgameAP0.3880.3970.3560.399
1-RL0.8720.8800.8750.893
AUC0.8750.8830.8780.896
IAPRTC12AP0.4280.4410.3660.447
1-RL0.9100.9160.9040.926
AUC0.9110.9160.9060.927
MirflickrAP0.6610.6820.6760.698
1-RL0.8970.9060.9070.915
AUC0.8790.8880.8900.898
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.387, + 0.485, + 0.472 + ], + "angle": 0, + "content": "Table 4. Ablation results on two datasets with \\(50\\%\\) missing views, \\(50\\%\\) missing labels, and \\(70\\%\\) training samples. 'w/o' denotes 'without'. 'A' denotes the consistent feature extraction module, 'B' denotes the disentangled specific feature extraction module, and DRLS denotes the full model. \\(\\mathcal{L}_{con.kl}\\) represents the KL divergence term of \\(\\mathcal{L}_{consist}\\)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.476, + 0.49, + 0.603 + ], + "angle": 0, + "content": "
MethodCorel5kPascal07
AP1-RLAUCAP1-RLAUC
A0.4160.9090.9120.5590.8360.858
A+B0.4310.9130.9160.5630.8400.863
DRLS0.4330.9160.9180.5670.8430.864
DRLS w/o Lcon.kl0.4150.9080.9110.5410.8230.846
DRLS w/o Ldisent0.4140.9130.9160.5570.8340.857
DRLS w/o Lle0.4300.9130.9150.5650.8420.863
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.433, + 0.643 + ], + "angle": 0, + "content": "ting of \\(50\\%\\) missing views and \\(50\\%\\) missing labels." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.839 + ], + "angle": 0, + "content": "The results of the ablation experiments are presented in Tab. 4. Specifically, we first test the performance of the model using only the consistent feature extraction module. Then, we add the disentangled specific feature extraction module, and finally, we incorporate the label semantic embedding learning module. Through these three experimental setups, we can observe that each module contributes to the model's performance. Additionally, by observing the results in Tab. 4, we can see that the disentangled specific feature extraction module plays a key role in improving model performance. Furthermore, the label semantic embedding learning module also helps to enhance the model's classification ability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In addition, we also validate the impact of different objective functions on model performance. We find that removing \\(\\mathcal{L}_{con.kl}\\) or \\(\\mathcal{L}_{disent}\\) leads to a significant decline in model performance. Among them, \\(\\mathcal{L}_{con.kl}\\) is the objective" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.075, + 0.714, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.185, + 0.645, + 0.196 + ], + "angle": 0, + "content": "(a) Corel5k" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.075, + 0.907, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.185, + 0.84, + 0.197 + ], + "angle": 0, + "content": "(b) Pascal07" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.199, + 0.714, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.325, + 0.646, + 0.337 + ], + "angle": 0, + "content": "(c) Corel5k" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.199, + 0.905, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.325, + 0.84, + 0.337 + ], + "angle": 0, + "content": "(d) Pascal07" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.346, + 0.908, + 0.403 + ], + "angle": 0, + "content": "Figure 2. The AP values under different combinations of \\(\\beta\\) and \\(\\gamma\\) on the Corel5k (a) and Pascal07 (b) datasets; the AP and AUC values for different values of \\(\\alpha\\) on the Corel5k (c) and Pascal07 (d) datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.907, + 0.519 + ], + "angle": 0, + "content": "function used in the pre-training stage. The experimental results show that \\(\\mathcal{L}_{con.kl}\\) plays a regularization role, reducing redundant information in the shared features. The disentangling loss \\(\\mathcal{L}_{disent}\\) is responsible for separating the consistent features from the specific features, which improves the quality of the specific features." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.55, + 0.634, + 0.567 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.909, + 0.84 + ], + "angle": 0, + "content": "In this paper, we propose a novel incomplete multi-view multi-label learning framework. Based on the assumptions of consistency and complementarity of multi-view data, we establish a method using Disentangled Representation and Label Semantic embedding (DRLS). Firstly, view-consistent features are learned using cross-view reconstruction, and the difficulty of disentangling specific features is reduced by unsupervised pre-training. Next, we achieve disentangled learning of specific features by minimizing the variational upper bound of mutual information between consistent and specific features. Meanwhile, we guide the learning of label semantic embeddings with the help of a priori label relevance to aggregate relevant semantic embeddings and maintain the topology of label relevance in the semantic space. Finally, we conduct a large number of experiments to demonstrate the effectiveness of our method in addressing the issues of missing views and labels." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.845, + 0.692, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. by National Natural China (62476166) and (22PJD029)." + }, + { + "type": "text", + "bbox": [ + 0.702, + 0.845, + 0.908, + 0.889 + ], + "angle": 0, + "content": "This work is supported Science Foundation of Shanghai Pujiang Program" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30729" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.157 + ], + "angle": 0, + "content": "[1] Ana Lawry Aguila and Andre Altmann. A tutorial on multi-view autoencoders using the multi-view-ae library. arXiv preprint arXiv:2403.07456, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Luigi Antelmi, Nicholas Ayache, Philippe Robert, and Marco Lorenzi. Sparse multi-channel variational autoencoder for the joint analysis of heterogeneous data. In International Conference on Machine Learning, pages 302-311. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.269 + ], + "angle": 0, + "content": "[3] Yusheng Cheng, Qingyan Li, Yibin Wang, and Weijie Zheng. Multi-view multi-label learning with view feature attention allocation. Neurocomputing, 501:857-874, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[4] Bassoma Diallo, Jie Hu, Tianrui Li, Ghufran Ahmad Khan, Xinyan Liang, and Hongjun Wang. Auto-attention mechanism for multi-view deep embedding clustering. Pattern Recognition, 143:109764, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.409 + ], + "angle": 0, + "content": "[5] Pinar Duygulu, Kobus Barnard, Joao FG de Freitas, and David A Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In Computer Vision—ECCV 2002: 7th European Conference on Computer Vision Copenhagen, Denmark, May 28–31, 2002 Proceedings, Part IV 7, pages 97–112. Springer, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[7] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.578 + ], + "angle": 0, + "content": "[8] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International workshop ontoImage, 2006. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.483, + 0.634 + ], + "angle": 0, + "content": "[9] Jun-Yi Hang and Min-Ling Zhang. Collaborative learning of label semantics and deep label-specific features for multi-label classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9860-9871, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.636, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[10] Irina Higgins, Loic Matthew, Arka Pal, Christopher P Burgess, Xavier Glorot, Matthew M Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. *ICLR (Poster)*, 3, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[11] Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence. Neural computation, 14(8): 1771-1800, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[12] Sheng-Jun Huang and Zhi-Hua Zhou. Multi-label learning by exploiting label correlations locally. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 949-955, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[13] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In Proceedings of the 1st ACM international conference on Multimedia information retrieval, pages 39-43, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] HyeongJoo Hwang, Geon-Hyeong Kim, Seunghoon Hong, and Kee-Eung Kim. Variational interaction information maximization for cross-domain disentanglement. Advances in" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Neural Information Processing Systems, 33:22479-22491, 2020. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[15] Guanzhou Ke, Bo Wang, Xiaoli Wang, and Shengfeng He. Rethinking multi-view representation learning via distilled disentangling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26774-26783, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.905, + 0.223 + ], + "angle": 0, + "content": "[16] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.225, + 0.905, + 0.252 + ], + "angle": 0, + "content": "[17] Thomas N Kipf and Max Welling. Variational graph autoencoders. arXiv preprint arXiv:1611.07308, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.256, + 0.905, + 0.323 + ], + "angle": 0, + "content": "[18] Ana Lawry Aguila, James Chapman, and Andre Altmann. Multi-modal variational autoencoders for normative modelling across multiple imaging modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 425-434. Springer, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.327, + 0.905, + 0.38 + ], + "angle": 0, + "content": "[19] Haobin Li, Yunfan Li, Mouxing Yang, Peng Hu, Dezhong Peng, and Xi Peng. Incomplete multi-view clustering via prototype-based imputation. arXiv preprint arXiv:2301.11045, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.384, + 0.905, + 0.438 + ], + "angle": 0, + "content": "[20] Xiang Li and Songcan Chen. A concise yet effective model for non-aligned incomplete multi-view and missing multilabel learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5918-5932, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[21] Chengliang Liu, Gehui Xu, Jie Wen, Yabo Liu, Chao Huang, and Yong Xu. Partial multi-view multi-label classification via semantic invariance learning and prototype modeling. In *Forty-first International Conference on Machine Learning*, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.513, + 0.905, + 0.582 + ], + "angle": 0, + "content": "[22] Chengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, and Yong Xu. Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In Proceedings of the AAAI conference on artificial intelligence, pages 8807-8815, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.585, + 0.905, + 0.653 + ], + "angle": 0, + "content": "[23] Chengliang Liu, Jie Wen, Yabo Liu, Chao Huang, Zhihao Wu, Xiaoling Luo, and Yong Xu. Masked two-channel decoupling framework for incomplete multi-view weak multi-label learning. Advances in Neural Information Processing Systems, 36, 2024. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.657, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[24] Zhongchen Ma and Songcan Chen. Expand globally, shrink locally: Discriminant multi-label learning with missing labels. Pattern Recognition, 111:107675, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.701, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[25] William McGill. Multivariate information transmission. Transactions of the IRE Professional Group on Information Theory, 4(4):93-111, 1954. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.744, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[26] Yuge Shi, Brooks Paige, Philip Torr, et al. Variational mixture-of-experts autoencoders for multi-modal deep generative models. Advances in neural information processing systems, 32, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[27] Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zili Zhang. Incomplete multi-view weak-label learning. In Ijcai, pages 2703-2709, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Chang Tang, Zhenglai Li, Jun Wang, Xinwang Liu, Wei Zhang, and En Zhu. Unified one-step multi-view spectral clustering. IEEE Transactions on Knowledge and Data Engineering, 35(6):6449-6460, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30730" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[29] Luis Von Ahn and Laura Dabbish. Labeling images with a computer game. In Proceedings of the SIGCHI conference on Human factors in computing systems, pages 319-326, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[30] Xu Wang, Dezhong Peng, Peng Hu, and Yongsheng Sang. Adversarial correlated autoencoder for unsupervised multiview representation learning. Knowledge-Based Systems, 168:109-120, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[31] Jie Wen, Chengliang Liu, Shijie Deng, Yicheng Liu, Lunke Fei, Ke Yan, and Yong Xu. Deep double incomplete multiview multi-label learning with incomplete labels and missing views. IEEE transactions on neural networks and learning systems, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[32] Mike Wu and Noah Goodman. Multimodal generative models for scalable weakly-supervised learning. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[33] Xuan Wu, Qing-Guo Chen, Yao Hu, Dengbao Wang, Xiaodong Chang, Xiaobo Wang, and Min-Ling Zhang. Multiview multi-label learning with view-specific information extraction. In IJCAI, pages 3884-3890, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[34] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[35] Xiaoqiang Yan, Shizhe Hu, Yiqiao Mao, Yangdong Ye, and Hui Yu. Deep multi-view learning methods: A review. Neurocomputing, 448:106-129, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[36] Jun Yin and Shiliang Sun. Incomplete multi-view clustering with reconstructed views. IEEE Transactions on Knowledge and Data Engineering, 35(3):2671-2682, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[37] Jun Yin and Wentao Zhang. Multi-view multi-label learning with double orders manifold preserving. Applied Intelligence, 53(12):14703-14716, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[38] Jun Yin, Shiliang Sun, Lai Wei, and Pei Wang. Discriminatively fuzzy multi-view k-means clustering with local structure preserving. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16478-16485, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[39] Honglin Yuan, Shiyun Lai, Xingfeng Li, Jian Dai, Yuan Sun, and Zhenwen Ren. Robust prototype completion for incomplete multi-view clustering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10402-10411, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.674, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[40] Dawei Zhao, Qingwei Gao, Yixiang Lu, Dong Sun, and Yusheng Cheng. Consistency and diversity neural network multi-view multi-label learning. Knowledge-Based Systems, 218:106841, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[41] Dawei Zhao, Qingwei Gao, Yixiang Lu, Dong Sun, and Yusheng Cheng. Consistency and diversity neural network multi-view multi-label learning. Knowledge-Based Systems, 218:106841, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.787, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[42] Yue Zhu, James T Kwok, and Zhi-Hua Zhou. Multi-label learning with global and local label correlation. IEEE Transactions on Knowledge and Data Engineering, 30(6):1081-1094, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.842 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "30731" + } + ] +] \ No newline at end of file diff --git a/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/419f8c28-377b-4f29-8c8b-a2cba093a85b_origin.pdf b/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/419f8c28-377b-4f29-8c8b-a2cba093a85b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6696194d8a7f0ab02c52aaa5f0d959f0b50e4894 --- /dev/null +++ b/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/419f8c28-377b-4f29-8c8b-a2cba093a85b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9083950f473f46f44610cb7a3a2fc85c771c3bc7b0fecae21902d5770079112 +size 512883 diff --git a/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/full.md b/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7cc5e6772112a5631f6cd5cd612893911b6b3528 --- /dev/null +++ b/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/full.md @@ -0,0 +1,348 @@ +# Incomplete Multi-View Multi-Label Learning via Disentangled Representation and Label Semantic Embedding + +Xu Yan $^{1}$ , Jun Yin $^{1,*}$ , Jie Wen $^{2}$ $^{1}$ Shanghai Maritime University, $^{2}$ Harbin Institute of Technology, Shenzhen yanxu@stu.shmtu.edu.cn, junyin@shmtu.edu.cn, jiewen_pr@126.com + +# Abstract + +In incomplete multi-view multi-label learning scenarios, it is crucial to use the incomplete multi-view data to extract consistent and specific representations from different data sources and to fully exploit the missing label information. However, most previous approaches ignore the separation problem between view-shared and specific information. To address this problem, in this paper, we propose a method that can separate view-consistent features from view-specific features under the Variational Autoencoder (VAE) framework. Specifically, we first introduce cross-view reconstruction to capture view-consistent features and extract shared information from different views through unsupervised pre-training. Subsequently, we develop a disentangling module to learn specific features by minimizing the variational upper bound of mutual information between consistent and specific features. Finally, we utilize prior label relevance information derived from training data to guide the learning of the distribution of label semantic embeddings, aggregating relevant semantic embeddings and maintaining the label relevance topology in the semantic space. In extensive experiments, our model outperforms existing state-of-the-art algorithms on several real-world datasets, which fully validates its strong adaptability to missing views and labels. + +# 1. Introduction + +In recent years, with the increasing diversity and complexity of data, multi-view learning has gained widespread attention as an effective data processing technique. In many real-world applications, such as image categorization, video analysis, and text mining, data is often presented in the form of multiple views, which makes multi-view learning methods an effective tool for solving complex problems [28, 35, 36]. In the past few years, many multi-view learning methods have been proposed based on the assumption + +of view completeness [4, 30, 38]. However, it is not uncommon to have missing views when dealing with multi-view data, which poses a challenge to the learning effectiveness of the model [19, 39]. + +Meanwhile, multi-label learning is an equally important research area. Unlike traditional single-label classification tasks, multi-label learning aims to deal with situations where instances may belong to more than one category at the same time. The key to multi-label learning is how to deal with the correlation between labels, which is helpful in improving the classification performance [12, 42]. Multi-label datasets are often incomplete due to various uncertainties involved in manual labeling. So research on incomplete multi-label learning is particularly important. + +Traditional multi-label classification relies on features extracted from a single view for classification, which leads to insufficient feature information. Using multi-view data for multi-label classification can effectively alleviate this problem. Currently, there are many models dealing with multi-view multi-label classification (MVMLC) tasks that show promising performance [33, 37, 41]. These models typically assume that the data is complete, and currently, only a few methods are capable of handling both missing views and missing labels simultaneously. + +To deal with the Double Incomplete Multi-View Multi-Label Classification problem (DIMVMLC), several deep learning models based on autoencoders have been proposed [22, 23, 31]. However, there are still only a few methods for disentangling consistent features and specific features. SIMM [33] identifies consistent features by minimizing confusing adversarial loss and multi-label loss, uses orthogonal constraints to eliminate shared information, and extracts specific information from raw information. However, SIMM cannot be applied to the missing data case, and the forced orthogonality constraint eliminates linear correlation, and there may still be a nonlinear relationship between shared and specific features. This relationship may be difficult to capture by simple orthogonality constraints, affecting the expressiveness of the features. MTD [23] simultaneously learns consistent and specific features, sepa + +rating them using multi-view cross-channel contrastive loss. Optimizing both tasks simultaneously may lead to a complex optimization process where the model needs to trade off between two objectives, increasing the difficulty of convergence. + +To address the above complex double incomplete multi-view multi-label learning problem, we propose a Disentangled Representation and Label Semantic Embedding based method (DRLS), as shown in Fig. 1. Our approach consists of three main parts: (1) Learning multi-view consistent features by minimizing the KL divergence between the approximate posterior of each view's consistent features and the multi-view joint posterior. (2) Learning the disentangled specific features by minimizing the variational upper bound of mutual information between the consistent features and the specific features. (3) We model the distribution of label semantic embeddings by leveraging prior label relevance information from the training data and utilizing graph neural networks. Our model is built upon the Variational Autoencoder (VAE) framework. Numerous experiments demonstrate the excellent performance of DRLS. Finally, the contributions of this paper are summarized as follows: + +- We propose a novel approach based on disentangled representation and label semantic embedding for solving the DIMVMLC problem. It can effectively utilize the consistent and specific information of multi-view data, and separate them through a disentangled representation learning module for effective multi-label classification tasks. +- We introduce a pre-training strategy that learns view-consistent representations via pre-training and then minimizes the variational upper bound of mutual information between consistent and specific features to separate them effectively. +- We design a graph neural network-based label semantic embedding learning module that simultaneously aggregates embeddings of related labels and maintains the label relevance topology in the semantic space. +- Our approach achieves leading performance on five publicly available datasets and can be applied to both complete and arbitrarily missing data. + +# 2. Related works + +Many MVMLC methods have been proposed. For example, CDMM [40] builds independent classifiers based on neural networks constructed with nonlinear kernel mapping to ensure that each view learns consistent labeling results. The diversity across views is addressed by integrating the Hilbert-Schmidt independence criterion [7] with an exploration of the variations among different views. MMFA [3] initially extracts both common and private features from different views, then computes the similarity between the original features and the common features. The attention + +weight matrix is derived by multiplying this similarity with the synergistic feature matrix, which is subsequently used for multi-label prediction. + +For DIMVMLC, iMvWL [27] uses Nonnegative Matrix Factorization (NMF) to learn the shared subspace of multiple views and introduces label information and local label correlation matrices to improve the discriminative power of the subspace. NAIML [20] exploits both consistency among multiple views and the complex global and local structure among labels to cope with the lack of available labels. Both of these methods are shallow machine learning approaches, and current deep neural network-based methods show better performance in this area, such as DDINet [31], DICNet [22], and SIP [21]. DDINet is the first to use a deep neural network to deal with DIMVMLC. DICNet leverages contrastive loss to learn shared features across views, aligns cross-view instances, and improves feature discrimination through a multi-view fusion module, effectively handling missing views and labels. The SIP model utilizes the principles of the information bottleneck theory to reduce the non-shared information while preserving the relevant features, and captures the associations between labels through prototype instances in the latent space. + +# 3. Method + +# 3.1. Problem definition + +Given a dataset $\{x^{(v)}\}_{v=1}^m$ consisting of $n$ samples, each with $m$ views, where $x^{(v)} \in \mathbb{R}^{d_v}$ . Here $d_v$ denotes the feature dimension of the $v$ th view. The label vector is $y \in \{0,1\}^k$ , where $k$ denotes the number of categories. If $y_i = 0$ , the sample does not have the $i$ -th label; if $y_i = 1$ , it does. To handle missing data, we define the set $\mathcal{W}(|\mathcal{W}| \leq m)$ to denote the set of observable views of a sample, and the observable multi-view data can be defined as $\{x^{(v)}\}_{v \in \mathcal{W}}$ , abbreviated as $\{x\}$ . We also define a set of observable labels for a sample $\mathcal{U}(|\mathcal{U}| \leq k)$ . Our goal is to train a model to predict the labels of unseen samples. Our loss and objective functions are defined in the case of a single sample. + +# 3.2. View-consistent feature + +In traditional multi-view learning, consistent features are often learned separately without cross-view interaction [33], making it hard to fully capture the common semantic information across views. In this subsection, we first introduce how to learn multi-view consistent features through cross-view reconstruction. + +First, we define $q_{\phi_v}(c \mid x^{(v)})$ with learnable parameters $\phi_v$ to approximate the posterior $p(c \mid x^{(v)})$ , where $c$ represents the consistent feature. The consistent features of each view should contain the public information of all views, inspired by [2], we learn the consistent features by constraining the approximate posterior of each view to be similar to + +the multi-view joint posterior: + +$$ +\min \frac {1}{| \mathcal {W} |} \sum_ {v \in \mathcal {W}} \left[ D _ {K L} \left(q _ {\phi_ {v}} (c \mid x ^ {(v)}) \| p (c \mid \{x \})\right) \right] \tag {1} +$$ + +where $D_{KL}(\cdot \|\cdot)$ represents the KL divergence, $|\mathcal{W}|$ denotes the number of available views, and $p(c\mid \{x\})$ denotes the multi-view joint posterior. + +The Evidence Lower Bound (ELBO) is derived from Eq. (1) by further applying Bayes' theorem to decompose the multi-view joint posterior distribution $p(c \mid \{x\})$ . Due to space constraints, the exact derivation is given in the supplementary material Sec. 6.1. + +The final result is as follows: + +$$ +\begin{array}{l} \log p (\{x \}) - \frac {1}{| \mathcal {W} |} \sum_ {v \in \mathcal {W}} \left[ D _ {K L} \left(q _ {\phi_ {v}} (c | x ^ {(v)}) \| p (c | \{x \})\right) \right] \\ = \frac {1}{| \mathcal {W} |} \sum_ {v \in \mathcal {W}} \left[ \mathbb {E} _ {q _ {\phi_ {v}} (c | x ^ {(v)})} [ \log p (\{x \} | c) ] \right. \\ \left. - D _ {K L} \left(q _ {\phi_ {v}} (c \mid x ^ {(v)}) \| p (c)\right) \right] \tag {2} \\ \end{array} +$$ + +This establishes the Evidence Lower Bound (ELBO). Thus, by maximizing this lower bound, we are effectively solving the minimization problem in Eq. (1). + +By assuming each view is conditionally independent from the others given $c$ , we can efficiently decompose the data likelihood as $p(\{x\} |c) = \prod_{i\in \mathcal{W}}p\left(x^{(i)}|c\right)\approx \prod_{i\in \mathcal{W}}p_{\theta_i}\left(x^{(i)}|c\right)$ , where $p_{\theta_i}\left(x^{(i)}|c\right)$ with learnable parameters $\theta_{i}$ approximates $p(x^{(i)}|c)$ . The objective in Eq. (2) involves extracting consistent features from different views. To facilitate subsequent analysis, we introduce an identifier for the consistent features of each view to distinguish them, i.e., $\{c^{(v)}\}_{v = 1}^m$ . Introducing this identifier does not alter the model architecture. Additionally, we introduce a trade-off parameter $\alpha$ [10], which controls the trade-off between reconstruction quality and regularization, resulting in the following lower bound: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {c o n s i s t}} = \\ \frac {1}{| \mathcal {W} |} \sum_ {v \in \mathcal {W}} \left[ \mathbb {E} _ {q _ {\phi_ {v}} (c ^ {(v)} | x ^ {(v)})} \left[ \sum_ {i \in \mathcal {W}} \log p _ {\theta_ {i}} (x ^ {(i)} | c ^ {(v)}) \right] \right. \\ \left. - \alpha D _ {K L} \left(q _ {\phi_ {v}} \left(c ^ {(v)} \mid x ^ {(v)}\right) \| p (c)\right) \right] \tag {3} \\ \end{array} +$$ + +here $q_{\phi_v}(c^{(v)} \mid x^{(v)})$ can be regarded as an encoder, and we assume that $q_{\phi_v}(c^{(v)} \mid x^{(v)}) \sim \mathcal{N}(\mu_{\phi_v}(x^{(v)}), \sigma_{\phi_v}^2(x^{(v)}) I)$ where $\mu_{\phi_v}(x^{(v)})$ and $\sigma_{\phi_v}^2(x^{(v)})$ denote the mean and variance encoder of the $v$ th view, and $I$ denotes the identity matrix. Due to the non-differentiable nature of the sampling process, we introduce the reparameterization trick [16] to obtain the features: + +$$ +c ^ {(v)} = \mu_ {\phi_ {v}} \left(x ^ {(v)}\right) + \sigma_ {\phi_ {v}} \left(x ^ {(v)}\right) \odot \delta \tag {4} +$$ + +where $\delta \sim \mathcal{N}(0,I)$ , $\odot$ denotes the Hadamard product, and $c^{(v)}\in \mathbb{R}^{d_e}$ , where $d_{e}$ represents the feature dimension. The distribution $p_{\theta_i}(x^{(i)}|c^{(v)})$ in Eq. (3) is modeled as an MLP decoder, assuming $p_{\theta_i}(x^{(i)}|c^{(v)})\sim \mathcal{N}(\mu_{\theta_i}(c^{(v)}),I)$ , yielding the reconstructed feature $\hat{x}^{(i,v)} = \mu_{\theta_i}(c^{(v)})$ . + +A brief analysis of Eq. (3) is as follows: The first term represents both self- and cross-view reconstruction. It allows $c^{(v)}$ from each view to reconstruct the original data of different views, which helps $c^{(v)}$ capture shared information. The second term, the KL divergence, aligns the posterior $q_{\phi_v}(c^{(v)} \mid x^{(v)})$ of each view with the shared prior $p(c) \sim \mathcal{N}(0,I)$ . These two terms simultaneously encourage $c^{(v)}$ to encode the shared information between views. + +Finally, in order to obtain a fused consistent feature shared by multiple views for the classification task, we compute the multi-view joint posterior $p(c|\{x\})$ using the Product-of-Expert (PoE) approach [11]. This method is flexible and effective in dealing with the missing views by directly removing the corresponding parameter when a view is missing. According to the setting in [32], we compute $p(c|\{x\})$ as follows: + +$$ +\begin{array}{l} p (c | \{x \}) \propto p (c) \prod_ {v \in \mathcal {W}} p (c \mid x ^ {(v)}) \\ := p (c) \prod_ {v \in \mathcal {W}} q _ {\phi_ {v}} \left(c ^ {(v)} \mid x ^ {(v)}\right) \tag {5} \\ \end{array} +$$ + +The same reparameterization trick is used to obtain the fused multi-view consistent feature $\bar{c} \in \mathbb{R}^{d_e}$ , which is sampled from $p(c|\{x\})$ . + +# 3.3. Disentangled view-specific feature + +In this section, we introduce view-specific features $\{s^{(v)}\}_{v = 1}^{m}$ that encode information specific to each view beyond the consistent features. We consider minimizing the mutual information $I(c^{(v)};s^{(v)})$ between the consistent and specific features for each view, which allows us to maximally disentangle the specific features from the consistent features. A natural idea is to do the learning of consistent features first, followed by disentangling the specific features from them [15]. Thus, we first pre-train with $\mathcal{L}_{\text{consist}}$ for a certain number of epochs, depending on the dataset size. In the second stage, we decrease the learning rate for the encoders of consistent features $\{q_{\phi_v}(c^{(v)}\mid x^{(v)})\}_{v = 1}^m$ to stabilize the pre-trained features during further disentanglement. Our disentangling objective function is as follows: + +$$ +\min \frac {1}{| \mathcal {W} |} \sum_ {v \in \mathcal {W}} I \left(c ^ {(v)}; s ^ {(v)}\right) \tag {6} +$$ + +We rewrite $I(c^{(v)}; s^{(v)})$ as follows, more details are provided in the supplementary material Sec. 6.2: + +$$ +I \left(c ^ {(v)}; s ^ {(v)}\right) = I \left(x ^ {(v)}; c ^ {(v)}\right) + I \left(x ^ {(v)}; s ^ {(v)}\right) - I \left(x ^ {(v)}; c ^ {(v)}, s ^ {(v)}\right) \tag {7} +$$ + +![](images/20f4de535dc3f2dea301c5a8e6a84c11cf76cfb33fab2ca7e619a873cab5b8a4.jpg) + +![](images/06eb3429ae8cb37d63d1552a95323cfe190c63b0f879a9644f76d77c6b4c21cd.jpg) +Figure 1. The main framework of DRLS. The upper-left corner of the figure represents the view-consistent feature learning module, the upper-right corner shows the disentangled view-specific feature learning module, and the lower part depicts the label semantic embedding learning module. + +![](images/011e2c2d4a035e7f5163749b37ce9d34224ec98559f2a5996647c8faf2dcfa35.jpg) + +![](images/1cf3ddb39043e7ebe6d16790b34b36cf72f6819ee4ae77f783c06b509e72fbf3.jpg) + +![](images/65044f6f5d93d7f6168d0e82a77bafbc7facf670a5c8779c5b35b3e9034cb6b2.jpg) + +![](images/cd8d930529901c6b889138194a7ce6a7a683230b041bfada56d8dc8ea74cbbef.jpg) + +Clearly, directly optimizing Eq. (7) is challenging. Therefore, we introduce the variational upper bounds for $I(x^{(v)};c^{(v)})$ and $I(x^{(v)};s^{(v)})$ , as well as the variational lower bound for $I(x^{(v)};c^{(v)},s^{(v)})$ [14], resulting in: + +$$ +\begin{array}{l} I \left(c ^ {(v)}; s ^ {(v)}\right) \leq D _ {K L} \left(q _ {\phi_ {v}} \left(c ^ {(v)} \mid x ^ {(v)}\right) \| p \left(c ^ {(v)}\right)\right) \\ + D _ {K L} \left(q _ {\varphi_ {v}} \left(s ^ {(v)} \mid x ^ {(v)}\right) \| p \left(s ^ {(v)}\right)\right) \\ - \mathbb {E} _ {q _ {\phi_ {v}} (c ^ {(v)} | x ^ {(v)}) q _ {\varphi_ {v}} (s ^ {(v)} | x ^ {(v)})} [ \log p _ {\vartheta_ {v}} (x ^ {(v)} | c ^ {(v)}, s ^ {(v)}) ] \tag {8} \\ \end{array} +$$ + +We observe that minimizing this variational upper bound on mutual information trains a decoder $p_{\vartheta_v}(x^{(v)}|c^{(v)}, s^{(v)})$ and a specific feature encoder $q_{\varphi_v}(s^{(v)}|x^{(v)})$ . Here, we concatenate $c^{(v)}$ and $s^{(v)}$ as input to the MLP decoder $p_{\vartheta_v}(x^{(v)}|c^{(v)}, s^{(v)})$ to produce the reconstructed feature $\hat{x}^{(v)}$ . Similar to Sec. 3.2, $q_{\varphi_v}(s^{(v)}|x^{(v)}) \sim \mathcal{N}(\mu_{\varphi_v}(x^{(v)}), \sigma_{\varphi_v}^2(x^{(v)})I)$ parameterizes the distribution of the specific features. The feature $s^{(v)} \in \mathbb{R}^{d_e}$ is derived using the reparameterization trick as shown in Eq. (4). We assume that $p(c^{(v)})$ and $p(s^{(v)})$ follow standard Gaussian distributions. Finally, we obtain the disentanglement loss: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {d i s e n t}} = \frac {1}{| \mathcal {W} |} \sum_ {v \in \mathcal {W}} \left[ D _ {K L} \left(q _ {\phi_ {v}} \left(c ^ {(v)} \mid x ^ {(v)}\right) \| p \left(c ^ {(v)}\right)\right) \right. \\ + D _ {K L} \left(q _ {\varphi_ {v}} \left(s ^ {(v)} \mid x ^ {(v)}\right) \| p \left(s ^ {(v)}\right)\right) \\ \left. - \mathbb {E} _ {q _ {\phi_ {v}} \left(c ^ {(v)} \mid x ^ {(v)}\right) q _ {\varphi_ {v}} \left(s ^ {(v)} \mid x ^ {(v)}\right)} \left[ \log p _ {\vartheta_ {v}} \left(x ^ {(v)} \mid c ^ {(v)}, s ^ {(v)}\right) \right] \right] \tag {9} \\ \end{array} +$$ + +Comparing Eq. (3) and Eq. (9), the consistent objective function requires each consistent feature $c^{(v)}$ to reconstruct + +the original features across views, whereas in the disentanglement loss, we use the concatenation of consistent and specific features exclusively for self-view reconstruction. Since we pre-train a set of encoders $\{q_{\phi_v}(c^{(v)}|x^{(v)})\}_{v = 1}^m$ in the second stage, $c^{(v)}$ and $s^{(v)}$ are concatenated to reconstruct $x^{(v)}$ . As $c^{(v)}$ already contains the information shared between $x^{(v)}$ and other views, $s^{(v)}$ only needs to encode the complementary information that is not captured by $c^{(v)}$ . This pre-training strategy enables more efficient disentanglement. We use a simple and effective method to obtain the fused multi-view specific feature $\bar{s}\in \mathbb{R}^{d_e}$ , for subsequent classification tasks: + +$$ +\bar {s} = \frac {1}{| \mathcal {W} |} \sum_ {v \in \mathcal {W}} s ^ {(v)} \tag {10} +$$ + +Here, we do not use PoE to fuse the specific features as in Sec. 3.2, because it is more suitable for merging shared information. PoE is sensitive to consensus among different views [18]. Specifically, if one view's probability distribution is overly sharp and biased, the joint distribution will have a biased mean prediction [1, 18, 26]. Given that specific features from different views often have differences, we choose a simple average fusion method to mitigate extreme variations and retain the overall trend. + +Finally, we fuse the multi-view consistent and specific information. Following the method in [23], we use specific features to enhance the consistent features: + +$$ +z = \omega (\bar {s}) \odot \bar {c} \tag {11} +$$ + +$z \in \mathbb{R}^{d_e}$ represents the final fused representation, $\omega(\cdot)$ is the sigmoid activation function, and $\odot$ denotes the Hadamard product. + +# 3.4. Label semantic embedding + +To fully utilize the incomplete supervisory information, we introduce label semantic embedding learning guided by prior information. Here, we construct our model within the framework of the Variational Graph Autoencoder (VGAE) [17], which employs variational inference to model the latent distribution of label embeddings during the learning process. + +First, we define a relationship graph among labels, where the graph $G = (V,E)$ represents the label relationship graph, with $V$ denoting the nodes (categories) and $E$ representing the edges connecting the nodes. We model the adjacency matrix $A \in \mathbb{R}^{k \times k}$ of graph $G$ in terms of conditional probabilities, specifically $P(l_j|l_i)$ , which denotes the probability that label $l_j$ occurs when label $l_i$ occurs, where $l_i$ represents the $i$ th label. This is computed based on the label matrix $Y \in \mathbb{R}^{n_{train} \times k}$ from the training set. Since $P(l_j|l_i)$ is not equal to $P(l_i|l_j)$ , the adjacency matrix $A$ is asymmetric, which more accurately reflects real-world scenarios. The specific calculation method is as follows: + +$$ +A _ {i j} = \left\{ \begin{array}{l l} P \left(l _ {j} \mid l _ {i}\right) = \frac {C _ {i j}}{d _ {i}}, & \text {i f} i \neq j \\ 0, & \text {i f} i = j \end{array} \right. \tag {12} +$$ + +where $C = Y^{\top}Y$ , $C \in \mathbb{R}^{k \times k}$ is the label co-occurrence matrix, $C_{ij}$ is the co-occurrence count of labels $l_i$ and $l_j$ , $d_i = \sum_{a=1}^{n_{train}} Y_{ai} = Y_{:i}^{\top} Y_{:i}$ for $i = 1,2,\ldots,k$ , which represents the total occurrence count of label $l_i$ . It should be noted that we set the diagonal of the adjacency matrix $A$ to 0. + +Graph Isomorphism Network (GIN) provides stronger feature aggregation capabilities through learnable weights and information from self-nodes [34]. Therefore, we use GIN to model the posterior distribution of label semantic embeddings $H \in \mathbb{R}^{k \times d_e}$ : + +$$ +q \left(H \mid H ^ {0}, A\right) = \prod_ {i = 1} ^ {k} q \left(h _ {i} \mid H ^ {0}, A\right), \tag {13} +$$ + +$$ +\mathrm {w i t h} q (h _ {i} | H ^ {0}, A) \sim \mathcal {N} (\mu_ {i}, \sigma_ {i} ^ {2} I) +$$ + +here, $h_i \in \mathbb{R}^{d_e}$ is the $i$ -th row of $H$ , representing the semantic embedding of the $i$ -th label, and $H^0 \in \mathbb{R}^{k \times k}$ is an identity matrix, which means that the embedding for label $i$ is initialized as a one-hot vector with the $i$ -th position set to 1. This approach ensures that the initial distance between different labels is large, which helps to better distinguish the semantic differences between labels during the subsequent learning process. $\pmb{\mu} = \mathrm{GIN}_{\pmb{\mu}}(H^0, A)$ , $\pmb{\mu} \in \mathbb{R}^{k \times d_e}$ is the mean matrix, where the $i$ -th row corresponds to the mean + +vector $\mu_{i}$ for the $i$ -th label; similarly, $\sigma^2 = \mathrm{GIN}_{\sigma^2}(H^0,A)$ . Here, we use a three-layer GIN, where one layer can be defined as $H^{(t + 1)} = f^{(t + 1)}[(1 + \epsilon^{(t + 1)})H^{(t)} + AH^{(t)}]$ , $t$ represents the current number of network layers, $\epsilon^{(t + 1)}$ represents a learnable parameter, and $f^{(t + 1)}$ represents a two-layer fully connected network. $H^{(t + 1)}\in \mathbb{R}^{k\times d_{t + 1}}$ represents the output of the intermediate layer. + +Subsequently, we also use the reparameterization trick to sample the label semantic embeddings, with a formula similar to Eq. (4), i.e., $h_i = \mu_i + \sigma_i \odot \delta, \delta \sim \mathcal{N}(0, I)$ . Finally, we obtain the label semantic embedding matrix $H$ . + +We use a simple and efficient inner product between variables as a generative model: + +$$ +p (A | H) = \prod_ {i = 1} ^ {k} \prod_ {j = 1} ^ {k} p \left(A _ {i j} \mid h _ {i}, h _ {j}\right), \tag {14} +$$ + +$$ +\text {w i t h} p \left(A _ {i j} \mid h _ {i}, h _ {j}\right) = \omega \left(h _ {i} ^ {\top} h _ {j}\right) +$$ + +where $\omega(\cdot)$ denotes the sigmoid activation function. By using the label correlation matrix as the adjacency matrix of a graph, we enable GIN to aggregate semantically related embeddings to nearby positions in the label semantic space. However, it cannot guarantee that unrelated labels are effectively pushed apart. Therefore, we introduce the generative model from VGAE to reconstruct the adjacency matrix $A$ , ensuring that the label semantic embeddings effectively preserve the topology of label correlations in the semantic space. Finally, our variational lower bound is: + +$$ +\mathcal {L} _ {l e} = \mathbb {E} _ {q (H | H ^ {0}, A)} [ \log p (A | H) ] - D _ {K L} (q (H | H ^ {0}, A) \| p (H)) \tag {15} +$$ + +We further adopt a Gaussian prior $p(H) = \prod_{i}p(h_{i}) = \prod_{i}\mathcal{N}(h_{i}|0,I)$ . We use binary cross-entropy to compute the first term of $\mathcal{L}_{le}$ as the reconstruction loss and add an identity matrix to $A$ before computing the cross-entropy loss to avoid enforcing $\omega (h_i^\top h_i)\approx 0$ due to its zero diagonal. + +# 3.5. Overall loss function + +Finally, we utilize label semantic embeddings for feature selection, choosing discriminative and specific features for each label. First, the learned label semantic embedding $h_i$ is mapped to a feature importance vector [9] via a fully connected layer $f(\cdot)$ , specifically $r_i = \omega(f(h_i))$ , where $r_i \in \mathbb{R}^{d_e}$ and $\omega(\cdot)$ is the sigmoid activation function. Subsequently, we use $r_i$ to select the features and input them into a simple neural network to obtain predictions: + +$$ +p _ {i} = \omega \left(f _ {\text {c l s}} \left(z \odot r _ {i}\right)\right) \tag {16} +$$ + +where $\odot$ denotes the Hadamard product, $f_{cls}$ is a fully connected network, and $p_i$ is the predicted probability of the sample being labeled as the $i$ -th class. We use the commonly used binary cross-entropy function to calculate the + +Table 1. Statistics for five multi-view multi-label datasets, where $m$ denotes the number of views, $n$ denotes the number of samples, $k$ denotes the number of categories, and $n/k$ denotes the average number of labels per sample. + +
Datasetmnkn/k
Corel5k649992603.396
Pascal0769963201.465
Espgame6207702684.686
IAPRTC126196272915.719
Mirflickr625000384.716
+ +multi-label classification loss: + +$$ +\mathcal {L} _ {B C E} = - \frac {1}{| \mathcal {U} |} \sum_ {i \in \mathcal {U}} \left(y _ {i} \log p _ {i} + \left(1 - y _ {i}\right) \log \left(1 - p _ {i}\right)\right) \tag {17} +$$ + +where $y_{i}$ denotes the ground truth and we exclude missing labels from loss computation to filter out the impact of missing labels. Our final loss function is: + +$$ +\mathcal {L} = \mathcal {L} _ {B C E} + \beta \mathcal {L} _ {\text {d i s e n t}} - \gamma \mathcal {L} _ {l e} \tag {18} +$$ + +where $\beta$ and $\gamma$ are trade-off parameters used to balance the different loss terms. + +# 4. Experiments + +# 4.1. Datasets + +We conduct experiments based on the settings from [21, 23, 31] to evaluate the performance of our proposed model on five datasets. They are Corel5k [5], Pascal07 [6], Espgame [29], IAPRTC12 [8], and Mirflickr [13]. In these five datasets, we utilize six different types of features, which are DenseHue, DenseSift, GIST, HSV, RGB, and LAB, with dimensions of 100, 1000, 512, 4096, 4096, and 4096, respectively. More detailed information can be found in Tab. 1. + +# 4.2. Compared methods + +In our experiments, we select eight methods for comparison with our model: DM2L [24], CDMM [40], iMvWL [27], NAIML [20], DDINet [31], DICNet [22], MTD [23], and SIP [21]. It is important to note that iMvWL, NAIML, DDINet, DICNet, MTD, and SIP are incomplete multi-view multi-label methods. CDMM can only handle complete multi-view multi-label data; therefore, we follow the settings from [21] and use the mean of available instances from the corresponding view to fill in the unavailable instances for missing views. To handle incomplete labels, we treat the missing labels as negative labels. DM2L is a single-view incomplete multi-label method, and we record the best result of each view in our experiments. + +# 4.3. Implementation details + +For the datasets, in order to simulate real-world data missingness, we generate incomplete datasets following the settings in [27] and [21]. For each view, we randomly exclude $50\%$ of the instances, ensuring at least one view per sample. Additionally, we randomly remove $50\%$ of both positive and negative labels. We fill the missing features and labels with zeros. We use $70\%$ of the samples for training and the remaining $30\%$ for validation and testing. To ensure the stability of the results, we repeat the experiment 10 times and report the mean and variance of the results. Additional details are in the supplementary material Sec. 7. + +# 4.4. Evaluation metrics + +Consistent with some previous studies [21, 23, 31], in our experiments, we evaluate the performance of the model using six commonly used multi-label classification metrics, namely, Average Precision (AP), Area Under the Receiver Operating Characteristic Curve (AUC), Ranking Loss (RL), Hamming Loss (HL), OneError (OE), and Coverage (Cov). To better visualize the results, we use 1-RL, 1-HL, 1-OE, and 1-Cov to record the final results, where larger values indicate better performance. + +# 4.5. Experimental results + +In Tab. 2, we present the experimental results of nine methods on five public multi-view multi-label datasets, which include six evaluation metrics. The results in Tab. 2 are obtained under experimental settings with a $50\%$ view missing rate, a $50\%$ label missing rate, and $70\%$ of training samples. The results in Tab. 2 show that our method outperforms the other eight competitors on each metric in most datasets. For example, on the AP metric of the IAPRTC12 dataset, our method is $2.4\%$ higher than the second-ranked MTD. Specifically, Ave.R represents the average ranking of each method across the six metrics, and DRLS ranks first on all five datasets, which fully demonstrates the effectiveness of our method in addressing the DIMVMLC problem. + +In addition, further analysis of the experimental results reveals that CDMM, DDINet, DICNet, MTD, and SIP outperform the other three methods on all five datasets. These are deep learning methods, indicating the great potential of deep learning in addressing the DIMVMLC problem. + +Besides solving the DIMVMLC problem, we also further validated the performance of our method in addressing the MVMLC problem. In Tab. 3, we present the experimental results of four methods on five datasets, with experiments conducted under the settings of $70\%$ training samples, full views, and full labels. We select three representative evaluation metrics. From the results in Tab. 3, it is evident that our model still demonstrates competitive performance under this setup, indicating that our method can effectively address the MVMLC problem. + +Table 2. Experimental results of nine methods on five datasets with $50\%$ missing views, $50\%$ missing labels, and $70\%$ training samples. Ave.R denotes the average ranking of each method on the six metrics, and the numbers at the bottom right are the standard deviations. + +
DatasetMetricDM2LCDMMiMvWLNAIMLDDINetDICNetMTDSIPDRLS
Corel5kAP0.2620.0050.3540.0040.2830.0080.3090.0040.3600.0090.3780.0040.4130.0070.4160.0150.4330.008
1-HL0.9870.0000.9870.0000.9780.0000.9870.0000.9870.0000.9870.0000.9880.0000.9880.0000.9880.000
1-RL0.8430.0020.8840.0030.8650.0050.8780.0020.8650.0050.8770.0040.8920.0040.9100.0030.9160.002
AUC0.8450.0020.8880.0030.8680.0050.8810.0020.8680.0050.8810.0030.8950.0040.9120.0030.9180.002
1-OE0.2950.0140.4100.0070.3110.0150.3500.0090.4370.0120.4640.0120.4910.0100.4920.0180.5090.019
1-Cov0.6470.0050.7230.0070.7020.0080.7250.0050.6890.0120.7140.0100.7480.0090.7860.0070.8040.006
Ave.R8.1674.8337.6675.3336.0004.8332.6671.8331.000
Pascal07AP0.4710.0080.5080.0050.4370.0180.4880.0030.5320.0100.5020.0070.5500.0040.5500.0090.5670.008
1-HL0.9280.0010.9310.0010.8820.0040.9280.0010.9320.0010.9300.0010.9320.0010.9310.0020.9340.001
1-RL0.7610.0050.8120.0040.7360.0150.7830.0010.8080.0050.7810.0070.8300.0030.8250.0060.8430.004
AUC0.7790.0040.8380.0030.7670.0150.8110.0010.8290.0040.8050.0060.8490.0040.8450.0050.8640.003
1-OE0.4200.0110.4190.0080.3620.0230.4210.0060.4480.0150.4260.0130.4570.0080.4630.0120.4770.011
1-Cov0.6920.0040.7590.0030.6770.0150.7270.0020.7570.0050.7280.0070.7830.0040.7770.0050.7980.004
Ave.R7.6674.8339.0006.5004.1676.1672.1672.8331.000
EspaugeAP0.2120.0020.2890.0030.2440.0050.2460.0020.2860.0040.2990.0040.3060.0030.3100.0040.3260.005
1-HL0.9820.0000.9830.0000.9720.0000.9830.0000.9830.0000.9830.0000.9830.0000.9830.0000.9830.000
1-RL0.7810.0010.8320.0010.8080.0020.8180.0020.8150.0030.8330.0030.8370.0010.8490.0020.8580.002
AUC0.7850.0010.8360.0010.8130.0020.8240.0020.8190.0030.8370.0020.8420.0010.8530.0020.8620.002
1-OE0.2940.0060.3960.0050.3430.0130.3390.0030.4270.0100.4370.0100.4480.0060.4510.0120.4730.001
1-Cov0.4880.0030.5740.0040.5480.0040.5710.0030.5530.0050.5980.0060.6010.0040.6280.0040.6520.003
Ave.R8.8334.5008.0005.6675.5003.5002.6671.8331.000
IAPRTC12AP0.2340.0030.3050.0040.2370.0030.2610.0010.3020.0050.3270.0050.3320.0020.3310.0070.3560.006
1-HL0.9800.0000.9810.0000.9690.0000.9800.0000.9800.0000.9800.0000.9810.0000.9810.0000.9810.000
1-RL0.8230.0020.8620.0020.8330.0020.8480.0010.8530.0020.8720.0020.8750.0010.8870.0040.8960.003
AUC0.8250.0010.8640.0020.8350.0010.8500.0010.8550.0030.8730.0010.8760.0010.8880.0030.8980.002
1-OE0.3400.0060.4320.0080.3520.0080.3900.0050.4350.0090.4650.0130.4710.0060.4660.0010.4900.012
1-Cov0.5290.0040.5970.0040.5640.0050.5920.0040.5940.0070.6480.0050.6490.0020.6790.0080.7070.007
Ave.R8.3334.5008.1676.6675.6674.1672.3332.1671.000
MirfickrAP0.5140.0060.5700.0020.4900.0120.5510.0020.5880.0030.5860.0050.6080.0040.6150.0040.6300.005
1-HL0.8780.0010.8860.0010.8390.0020.8820.0010.8880.0010.8880.0010.8910.0010.8910.0010.8950.001
1-RL0.8310.0030.8560.0010.8030.0080.8440.0010.8650.0020.8610.0040.8750.0010.8780.0020.8850.002
AUC0.8280.0030.8460.0010.7870.0120.8370.0010.8530.0020.8480.0040.8610.0020.8640.0020.8720.003
1-OE0.5100.0080.6310.0040.5110.0220.5850.0030.6360.0080.6420.0060.6560.0040.6640.0060.6860.009
1-Cov0.6040.0050.6400.0010.5720.0130.6310.0020.6540.0030.6460.0060.6770.0020.6760.0040.6920.003
Ave.R8.1676.0008.8337.0004.1674.6672.6672.1671.000
+ +# 4.6. Parameter analysis + +In our DRLS model, there are three hyperparameters that need to be tuned: $\alpha$ , $\beta$ , and $\gamma$ . To evaluate the sensitivity of our model to these hyperparameters, we conduct experiments on the Corel5k and Pascal07 datasets, with experiments carried out under conditions where $50\%$ of the views are missing, $50\%$ of the labels are missing, and $70\%$ of the training samples are used. + +Figs. 2a and 2b show the AP values of our model under different combinations of $\beta$ and $\gamma$ , while Figs. 2c and 2d show the AP and AUC values for different values of $\alpha$ . To ensure the validity of the experiments, the other unrelated hyperparameters are fixed. Based on Figs. 2a and 2b, for the Corel5k dataset, $\beta$ and $\gamma$ are taken in the range of + +$[1e - 3, 1e - 1]$ and $[1e - 5, 1e - 3]$ , respectively, and for the Pascal07 dataset, $\beta$ and $\gamma$ are taken in the range of $[1e - 3, 1e - 1]$ and $[1e - 5, 1e - 3]$ , respectively. Under these conditions, our model demonstrates good performance. Regarding the hyperparameter $\alpha$ , from Figs. 2c and 2d, it can be seen that the variation in AP and AUC values is minimal for both datasets, indicating that our model is not sensitive to $\alpha$ . We set $\alpha$ within the range of $[1e - 1, 3e - 1]$ . + +# 4.7. Ablation study + +Our model consists of multiple modules and objective functions. To further analyze the contribution of each module and objective function, we conduct a series of ablation experiments on the Corel5k and Pascal07 datasets under a set- + +Table 3. Experimental results of four methods on five datasets with full views, full labels, and $70\%$ training samples. + +
DatasetMetricDICNetMTDSIPDRLS
Corel5kAP0.5040.4860.5200.551
1-RL0.9250.9210.9370.947
AUC0.9280.9240.9390.949
Pascal07AP0.6090.6520.6450.655
1-RL0.8600.8900.8870.889
AUC0.8750.8980.8990.901
EspgameAP0.3880.3970.3560.399
1-RL0.8720.8800.8750.893
AUC0.8750.8830.8780.896
IAPRTC12AP0.4280.4410.3660.447
1-RL0.9100.9160.9040.926
AUC0.9110.9160.9060.927
MirflickrAP0.6610.6820.6760.698
1-RL0.8970.9060.9070.915
AUC0.8790.8880.8900.898
+ +Table 4. Ablation results on two datasets with $50\%$ missing views, $50\%$ missing labels, and $70\%$ training samples. 'w/o' denotes 'without'. 'A' denotes the consistent feature extraction module, 'B' denotes the disentangled specific feature extraction module, and DRLS denotes the full model. $\mathcal{L}_{con.kl}$ represents the KL divergence term of $\mathcal{L}_{consist}$ . + +
MethodCorel5kPascal07
AP1-RLAUCAP1-RLAUC
A0.4160.9090.9120.5590.8360.858
A+B0.4310.9130.9160.5630.8400.863
DRLS0.4330.9160.9180.5670.8430.864
DRLS w/o Lcon.kl0.4150.9080.9110.5410.8230.846
DRLS w/o Ldisent0.4140.9130.9160.5570.8340.857
DRLS w/o Lle0.4300.9130.9150.5650.8420.863
+ +ting of $50\%$ missing views and $50\%$ missing labels. + +The results of the ablation experiments are presented in Tab. 4. Specifically, we first test the performance of the model using only the consistent feature extraction module. Then, we add the disentangled specific feature extraction module, and finally, we incorporate the label semantic embedding learning module. Through these three experimental setups, we can observe that each module contributes to the model's performance. Additionally, by observing the results in Tab. 4, we can see that the disentangled specific feature extraction module plays a key role in improving model performance. Furthermore, the label semantic embedding learning module also helps to enhance the model's classification ability. + +In addition, we also validate the impact of different objective functions on model performance. We find that removing $\mathcal{L}_{con.kl}$ or $\mathcal{L}_{disent}$ leads to a significant decline in model performance. Among them, $\mathcal{L}_{con.kl}$ is the objective + +![](images/a15c22535848be4de32af0f5b392c3fcbba72a069d899999dee3c8e8037a0753.jpg) +(a) Corel5k + +![](images/762165a4b4618d75db47e24feee5dcc0a03738ab1540b1150f19b633ec65a022.jpg) +(b) Pascal07 + +![](images/6b51f6f29bec8663730269e1b9945b830dfb92baf6f8690da363ee5c9aa7fa0b.jpg) +(c) Corel5k + +![](images/00b786215a61f27137b84620256e54a56d597211d3ddddacd89a440dbf6d4e01.jpg) +(d) Pascal07 +Figure 2. The AP values under different combinations of $\beta$ and $\gamma$ on the Corel5k (a) and Pascal07 (b) datasets; the AP and AUC values for different values of $\alpha$ on the Corel5k (c) and Pascal07 (d) datasets. + +function used in the pre-training stage. The experimental results show that $\mathcal{L}_{con.kl}$ plays a regularization role, reducing redundant information in the shared features. The disentangling loss $\mathcal{L}_{disent}$ is responsible for separating the consistent features from the specific features, which improves the quality of the specific features. + +# 5. Conclusion + +In this paper, we propose a novel incomplete multi-view multi-label learning framework. Based on the assumptions of consistency and complementarity of multi-view data, we establish a method using Disentangled Representation and Label Semantic embedding (DRLS). Firstly, view-consistent features are learned using cross-view reconstruction, and the difficulty of disentangling specific features is reduced by unsupervised pre-training. Next, we achieve disentangled learning of specific features by minimizing the variational upper bound of mutual information between consistent and specific features. Meanwhile, we guide the learning of label semantic embeddings with the help of a priori label relevance to aggregate relevant semantic embeddings and maintain the topology of label relevance in the semantic space. Finally, we conduct a large number of experiments to demonstrate the effectiveness of our method in addressing the issues of missing views and labels. + +Acknowledgment. by National Natural China (62476166) and (22PJD029). + +This work is supported Science Foundation of Shanghai Pujiang Program + +# References + +[1] Ana Lawry Aguila and Andre Altmann. A tutorial on multi-view autoencoders using the multi-view-ae library. arXiv preprint arXiv:2403.07456, 2024. 4 +[2] Luigi Antelmi, Nicholas Ayache, Philippe Robert, and Marco Lorenzi. Sparse multi-channel variational autoencoder for the joint analysis of heterogeneous data. In International Conference on Machine Learning, pages 302-311. PMLR, 2019. 2 +[3] Yusheng Cheng, Qingyan Li, Yibin Wang, and Weijie Zheng. Multi-view multi-label learning with view feature attention allocation. Neurocomputing, 501:857-874, 2022. 2 +[4] Bassoma Diallo, Jie Hu, Tianrui Li, Ghufran Ahmad Khan, Xinyan Liang, and Hongjun Wang. Auto-attention mechanism for multi-view deep embedding clustering. Pattern Recognition, 143:109764, 2023. 1 +[5] Pinar Duygulu, Kobus Barnard, Joao FG de Freitas, and David A Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In Computer Vision—ECCV 2002: 7th European Conference on Computer Vision Copenhagen, Denmark, May 28–31, 2002 Proceedings, Part IV 7, pages 97–112. Springer, 2002. 6 +[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6 +[7] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 2 +[8] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International workshop ontoImage, 2006. 6 +[9] Jun-Yi Hang and Min-Ling Zhang. Collaborative learning of label semantics and deep label-specific features for multi-label classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9860-9871, 2021. 5 +[10] Irina Higgins, Loic Matthew, Arka Pal, Christopher P Burgess, Xavier Glorot, Matthew M Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. *ICLR (Poster)*, 3, 2017. 3 +[11] Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence. Neural computation, 14(8): 1771-1800, 2002. 3 +[12] Sheng-Jun Huang and Zhi-Hua Zhou. Multi-label learning by exploiting label correlations locally. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 949-955, 2012. 1 +[13] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In Proceedings of the 1st ACM international conference on Multimedia information retrieval, pages 39-43, 2008. 6 +[14] HyeongJoo Hwang, Geon-Hyeong Kim, Seunghoon Hong, and Kee-Eung Kim. Variational interaction information maximization for cross-domain disentanglement. Advances in + +Neural Information Processing Systems, 33:22479-22491, 2020. 4, 1 +[15] Guanzhou Ke, Bo Wang, Xiaoli Wang, and Shengfeng He. Rethinking multi-view representation learning via distilled disentangling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26774-26783, 2024. 3 +[16] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[17] Thomas N Kipf and Max Welling. Variational graph autoencoders. arXiv preprint arXiv:1611.07308, 2016. 5 +[18] Ana Lawry Aguila, James Chapman, and Andre Altmann. Multi-modal variational autoencoders for normative modelling across multiple imaging modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 425-434. Springer, 2023. 4 +[19] Haobin Li, Yunfan Li, Mouxing Yang, Peng Hu, Dezhong Peng, and Xi Peng. Incomplete multi-view clustering via prototype-based imputation. arXiv preprint arXiv:2301.11045, 2023. 1 +[20] Xiang Li and Songcan Chen. A concise yet effective model for non-aligned incomplete multi-view and missing multilabel learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5918-5932, 2021. 2, 6 +[21] Chengliang Liu, Gehui Xu, Jie Wen, Yabo Liu, Chao Huang, and Yong Xu. Partial multi-view multi-label classification via semantic invariance learning and prototype modeling. In *Forty-first International Conference on Machine Learning*, 2, 6 +[22] Chengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, and Yong Xu. Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In Proceedings of the AAAI conference on artificial intelligence, pages 8807-8815, 2023. 1, 2, 6 +[23] Chengliang Liu, Jie Wen, Yabo Liu, Chao Huang, Zhihao Wu, Xiaoling Luo, and Yong Xu. Masked two-channel decoupling framework for incomplete multi-view weak multi-label learning. Advances in Neural Information Processing Systems, 36, 2024. 1, 4, 6 +[24] Zhongchen Ma and Songcan Chen. Expand globally, shrink locally: Discriminant multi-label learning with missing labels. Pattern Recognition, 111:107675, 2021. 6 +[25] William McGill. Multivariate information transmission. Transactions of the IRE Professional Group on Information Theory, 4(4):93-111, 1954. 1 +[26] Yuge Shi, Brooks Paige, Philip Torr, et al. Variational mixture-of-experts autoencoders for multi-modal deep generative models. Advances in neural information processing systems, 32, 2019. 4 +[27] Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zili Zhang. Incomplete multi-view weak-label learning. In Ijcai, pages 2703-2709, 2018. 2, 6 +[28] Chang Tang, Zhenglai Li, Jun Wang, Xinwang Liu, Wei Zhang, and En Zhu. Unified one-step multi-view spectral clustering. IEEE Transactions on Knowledge and Data Engineering, 35(6):6449-6460, 2022. 1 + +[29] Luis Von Ahn and Laura Dabbish. Labeling images with a computer game. In Proceedings of the SIGCHI conference on Human factors in computing systems, pages 319-326, 2004. 6 +[30] Xu Wang, Dezhong Peng, Peng Hu, and Yongsheng Sang. Adversarial correlated autoencoder for unsupervised multiview representation learning. Knowledge-Based Systems, 168:109-120, 2019. 1 +[31] Jie Wen, Chengliang Liu, Shijie Deng, Yicheng Liu, Lunke Fei, Ke Yan, and Yong Xu. Deep double incomplete multiview multi-label learning with incomplete labels and missing views. IEEE transactions on neural networks and learning systems, 2023. 1, 2, 6 +[32] Mike Wu and Noah Goodman. Multimodal generative models for scalable weakly-supervised learning. Advances in neural information processing systems, 31, 2018. 3 +[33] Xuan Wu, Qing-Guo Chen, Yao Hu, Dengbao Wang, Xiaodong Chang, Xiaobo Wang, and Min-Ling Zhang. Multiview multi-label learning with view-specific information extraction. In IJCAI, pages 3884-3890, 2019. 1, 2 +[34] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018. 5 +[35] Xiaoqiang Yan, Shizhe Hu, Yiqiao Mao, Yangdong Ye, and Hui Yu. Deep multi-view learning methods: A review. Neurocomputing, 448:106-129, 2021. 1 +[36] Jun Yin and Shiliang Sun. Incomplete multi-view clustering with reconstructed views. IEEE Transactions on Knowledge and Data Engineering, 35(3):2671-2682, 2021. 1 +[37] Jun Yin and Wentao Zhang. Multi-view multi-label learning with double orders manifold preserving. Applied Intelligence, 53(12):14703-14716, 2023. 1 +[38] Jun Yin, Shiliang Sun, Lai Wei, and Pei Wang. Discriminatively fuzzy multi-view k-means clustering with local structure preserving. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16478-16485, 2024. 1 +[39] Honglin Yuan, Shiyun Lai, Xingfeng Li, Jian Dai, Yuan Sun, and Zhenwen Ren. Robust prototype completion for incomplete multi-view clustering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10402-10411, 2024. 1 +[40] Dawei Zhao, Qingwei Gao, Yixiang Lu, Dong Sun, and Yusheng Cheng. Consistency and diversity neural network multi-view multi-label learning. Knowledge-Based Systems, 218:106841, 2021. 2, 6 +[41] Dawei Zhao, Qingwei Gao, Yixiang Lu, Dong Sun, and Yusheng Cheng. Consistency and diversity neural network multi-view multi-label learning. Knowledge-Based Systems, 218:106841, 2021. 1 +[42] Yue Zhu, James T Kwok, and Zhi-Hua Zhou. Multi-label learning with global and local label correlation. IEEE Transactions on Knowledge and Data Engineering, 30(6):1081-1094, 2017. 1 \ No newline at end of file diff --git a/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/images.zip b/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a3c8e76e1566817985a77d9d3e51b36468571b04 --- /dev/null +++ b/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ddd28fa851934c08e36d46ba4dc1bc402b2e8ab81a0e138c1999fdbec709a8f +size 897546 diff --git a/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/layout.json b/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4371fb7855ab5f6b37fc1369fd662d189cb74352 --- /dev/null +++ b/2025/Incomplete Multi-View Multi-label Learning via Disentangled Representation and Label Semantic Embedding/layout.json @@ -0,0 +1,10750 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "type": "text", + "content": "Incomplete Multi-View Multi-Label Learning via Disentangled Representation and Label Semantic Embedding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 160, + 486, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 160, + 486, + 204 + ], + "spans": [ + { + "bbox": [ + 124, + 160, + 486, + 204 + ], + "type": "text", + "content": "Xu Yan" + }, + { + "bbox": [ + 124, + 160, + 486, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 160, + 486, + 204 + ], + "type": "text", + "content": ", Jun Yin" + }, + { + "bbox": [ + 124, + 160, + 486, + 204 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 124, + 160, + 486, + 204 + ], + "type": "text", + "content": ", Jie Wen" + }, + { + "bbox": [ + 124, + 160, + 486, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 124, + 160, + 486, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 160, + 486, + 204 + ], + "type": "text", + "content": "Shanghai Maritime University, " + }, + { + "bbox": [ + 124, + 160, + 486, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 124, + 160, + 486, + 204 + ], + "type": "text", + "content": "Harbin Institute of Technology, Shenzhen yanxu@stu.shmtu.edu.cn, junyin@shmtu.edu.cn, jiewen_pr@126.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 256, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 256, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 296, + 543 + ], + "type": "text", + "content": "In incomplete multi-view multi-label learning scenarios, it is crucial to use the incomplete multi-view data to extract consistent and specific representations from different data sources and to fully exploit the missing label information. However, most previous approaches ignore the separation problem between view-shared and specific information. To address this problem, in this paper, we propose a method that can separate view-consistent features from view-specific features under the Variational Autoencoder (VAE) framework. Specifically, we first introduce cross-view reconstruction to capture view-consistent features and extract shared information from different views through unsupervised pre-training. Subsequently, we develop a disentangling module to learn specific features by minimizing the variational upper bound of mutual information between consistent and specific features. Finally, we utilize prior label relevance information derived from training data to guide the learning of the distribution of label semantic embeddings, aggregating relevant semantic embeddings and maintaining the label relevance topology in the semantic space. In extensive experiments, our model outperforms existing state-of-the-art algorithms on several real-world datasets, which fully validates its strong adaptability to missing views and labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 295, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 695 + ], + "type": "text", + "content": "In recent years, with the increasing diversity and complexity of data, multi-view learning has gained widespread attention as an effective data processing technique. In many real-world applications, such as image categorization, video analysis, and text mining, data is often presented in the form of multiple views, which makes multi-view learning methods an effective tool for solving complex problems [28, 35, 36]. In the past few years, many multi-view learning methods have been proposed based on the assumption" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 232, + 553, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 280 + ], + "type": "text", + "content": "of view completeness [4, 30, 38]. However, it is not uncommon to have missing views when dealing with multi-view data, which poses a challenge to the learning effectiveness of the model [19, 39]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 281, + 554, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 554, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 554, + 400 + ], + "type": "text", + "content": "Meanwhile, multi-label learning is an equally important research area. Unlike traditional single-label classification tasks, multi-label learning aims to deal with situations where instances may belong to more than one category at the same time. The key to multi-label learning is how to deal with the correlation between labels, which is helpful in improving the classification performance [12, 42]. Multi-label datasets are often incomplete due to various uncertainties involved in manual labeling. So research on incomplete multi-label learning is particularly important." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "text", + "content": "Traditional multi-label classification relies on features extracted from a single view for classification, which leads to insufficient feature information. Using multi-view data for multi-label classification can effectively alleviate this problem. Currently, there are many models dealing with multi-view multi-label classification (MVMLC) tasks that show promising performance [33, 37, 41]. These models typically assume that the data is complete, and currently, only a few methods are capable of handling both missing views and missing labels simultaneously." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 522, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 713 + ], + "type": "text", + "content": "To deal with the Double Incomplete Multi-View Multi-Label Classification problem (DIMVMLC), several deep learning models based on autoencoders have been proposed [22, 23, 31]. However, there are still only a few methods for disentangling consistent features and specific features. SIMM [33] identifies consistent features by minimizing confusing adversarial loss and multi-label loss, uses orthogonal constraints to eliminate shared information, and extracts specific information from raw information. However, SIMM cannot be applied to the missing data case, and the forced orthogonality constraint eliminates linear correlation, and there may still be a nonlinear relationship between shared and specific features. This relationship may be difficult to capture by simple orthogonality constraints, affecting the expressiveness of the features. MTD [23] simultaneously learns consistent and specific features, sepa" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30722" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "rating them using multi-view cross-channel contrastive loss. Optimizing both tasks simultaneously may lead to a complex optimization process where the model needs to trade off between two objectives, increasing the difficulty of convergence." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 133, + 295, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 295, + 346 + ], + "type": "text", + "content": "To address the above complex double incomplete multi-view multi-label learning problem, we propose a Disentangled Representation and Label Semantic Embedding based method (DRLS), as shown in Fig. 1. Our approach consists of three main parts: (1) Learning multi-view consistent features by minimizing the KL divergence between the approximate posterior of each view's consistent features and the multi-view joint posterior. (2) Learning the disentangled specific features by minimizing the variational upper bound of mutual information between the consistent features and the specific features. (3) We model the distribution of label semantic embeddings by leveraging prior label relevance information from the training data and utilizing graph neural networks. Our model is built upon the Variational Autoencoder (VAE) framework. Numerous experiments demonstrate the excellent performance of DRLS. Finally, the contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 348, + 295, + 563 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 348, + 294, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 294, + 418 + ], + "type": "text", + "content": "- We propose a novel approach based on disentangled representation and label semantic embedding for solving the DIMVMLC problem. It can effectively utilize the consistent and specific information of multi-view data, and separate them through a disentangled representation learning module for effective multi-label classification tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 419, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 419, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 419, + 295, + 479 + ], + "type": "text", + "content": "- We introduce a pre-training strategy that learns view-consistent representations via pre-training and then minimizes the variational upper bound of mutual information between consistent and specific features to separate them effectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 479, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 295, + 526 + ], + "type": "text", + "content": "- We design a graph neural network-based label semantic embedding learning module that simultaneously aggregates embeddings of related labels and maintains the label relevance topology in the semantic space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 527, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 527, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 527, + 295, + 563 + ], + "type": "text", + "content": "- Our approach achieves leading performance on five publicly available datasets and can be applied to both complete and arbitrarily missing data." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 574, + 144, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 144, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 144, + 586 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Many MVMLC methods have been proposed. For example, CDMM [40] builds independent classifiers based on neural networks constructed with nonlinear kernel mapping to ensure that each view learns consistent labeling results. The diversity across views is addressed by integrating the Hilbert-Schmidt independence criterion [7] with an exploration of the variations among different views. MMFA [3] initially extracts both common and private features from different views, then computes the similarity between the original features and the common features. The attention" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "weight matrix is derived by multiplying this similarity with the synergistic feature matrix, which is subsequently used for multi-label prediction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 108, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 348 + ], + "type": "text", + "content": "For DIMVMLC, iMvWL [27] uses Nonnegative Matrix Factorization (NMF) to learn the shared subspace of multiple views and introduces label information and local label correlation matrices to improve the discriminative power of the subspace. NAIML [20] exploits both consistency among multiple views and the complex global and local structure among labels to cope with the lack of available labels. Both of these methods are shallow machine learning approaches, and current deep neural network-based methods show better performance in this area, such as DDINet [31], DICNet [22], and SIP [21]. DDINet is the first to use a deep neural network to deal with DIMVMLC. DICNet leverages contrastive loss to learn shared features across views, aligns cross-view instances, and improves feature discrimination through a multi-view fusion module, effectively handling missing views and labels. The SIP model utilizes the principles of the information bottleneck theory to reduce the non-shared information while preserving the relevant features, and captures the associations between labels through prototype instances in the latent space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 359, + 370, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 359, + 370, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 359, + 370, + 372 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 380, + 424, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 424, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 424, + 392 + ], + "type": "text", + "content": "3.1. Problem definition" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": "Given a dataset " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\{x^{(v)}\\}_{v=1}^m" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": " samples, each with " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": " views, where " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "x^{(v)} \\in \\mathbb{R}^{d_v}" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "d_v" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": " denotes the feature dimension of the " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": "th view. The label vector is " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "y \\in \\{0,1\\}^k" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": " denotes the number of categories. If " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "y_i = 0" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": ", the sample does not have the " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": "-th label; if " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "y_i = 1" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": ", it does. To handle missing data, we define the set " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(|\\mathcal{W}| \\leq m)" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": " to denote the set of observable views of a sample, and the observable multi-view data can be defined as " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\{x^{(v)}\\}_{v \\in \\mathcal{W}}" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": ", abbreviated as " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\{x\\}" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": ". We also define a set of observable labels for a sample " + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(|\\mathcal{U}| \\leq k)" + }, + { + "bbox": [ + 313, + 398, + 555, + 542 + ], + "type": "text", + "content": ". Our goal is to train a model to predict the labels of unseen samples. Our loss and objective functions are defined in the case of a single sample." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 551, + 447, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 447, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 447, + 562 + ], + "type": "text", + "content": "3.2. View-consistent feature" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 640 + ], + "type": "text", + "content": "In traditional multi-view learning, consistent features are often learned separately without cross-view interaction [33], making it hard to fully capture the common semantic information across views. In this subsection, we first introduce how to learn multi-view consistent features through cross-view reconstruction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "First, we define " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "q_{\\phi_v}(c \\mid x^{(v)})" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " with learnable parameters " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\phi_v" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " to approximate the posterior " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "p(c \\mid x^{(v)})" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " represents the consistent feature. The consistent features of each view should contain the public information of all views, inspired by [2], we learn the consistent features by constraining the approximate posterior of each view to be similar to" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30723" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 179, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 179, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 179, + 84 + ], + "type": "text", + "content": "the multi-view joint posterior:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 89, + 295, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 89, + 295, + 118 + ], + "spans": [ + { + "bbox": [ + 71, + 89, + 295, + 118 + ], + "type": "interline_equation", + "content": "\\min \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ D _ {K L} \\left(q _ {\\phi_ {v}} (c \\mid x ^ {(v)}) \\| p (c \\mid \\{x \\})\\right) \\right] \\tag {1}", + "image_path": "e5e4a71f9907a329733c7571e7b8c30ece35684d30076e91e0398cac6a1883a0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 122, + 295, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 158 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 122, + 295, + 158 + ], + "type": "inline_equation", + "content": "D_{KL}(\\cdot \\|\\cdot)" + }, + { + "bbox": [ + 55, + 122, + 295, + 158 + ], + "type": "text", + "content": " represents the KL divergence, " + }, + { + "bbox": [ + 55, + 122, + 295, + 158 + ], + "type": "inline_equation", + "content": "|\\mathcal{W}|" + }, + { + "bbox": [ + 55, + 122, + 295, + 158 + ], + "type": "text", + "content": " denotes the number of available views, and " + }, + { + "bbox": [ + 55, + 122, + 295, + 158 + ], + "type": "inline_equation", + "content": "p(c\\mid \\{x\\})" + }, + { + "bbox": [ + 55, + 122, + 295, + 158 + ], + "type": "text", + "content": " denotes the multi-view joint posterior." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 158, + 296, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 296, + 218 + ], + "type": "text", + "content": "The Evidence Lower Bound (ELBO) is derived from Eq. (1) by further applying Bayes' theorem to decompose the multi-view joint posterior distribution " + }, + { + "bbox": [ + 55, + 158, + 296, + 218 + ], + "type": "inline_equation", + "content": "p(c \\mid \\{x\\})" + }, + { + "bbox": [ + 55, + 158, + 296, + 218 + ], + "type": "text", + "content": ". Due to space constraints, the exact derivation is given in the supplementary material Sec. 6.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 219, + 186, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 219, + 186, + 229 + ], + "spans": [ + { + "bbox": [ + 67, + 219, + 186, + 229 + ], + "type": "text", + "content": "The final result is as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 232, + 294, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 232, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 64, + 232, + 294, + 323 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\log p (\\{x \\}) - \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ D _ {K L} \\left(q _ {\\phi_ {v}} (c | x ^ {(v)}) \\| p (c | \\{x \\})\\right) \\right] \\\\ = \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ \\mathbb {E} _ {q _ {\\phi_ {v}} (c | x ^ {(v)})} [ \\log p (\\{x \\} | c) ] \\right. \\\\ \\left. - D _ {K L} \\left(q _ {\\phi_ {v}} (c \\mid x ^ {(v)}) \\| p (c)\\right) \\right] \\tag {2} \\\\ \\end{array}", + "image_path": "23fc09435093631ac20e1b70cd808ac2aa15a93c50fa6500ea94869e99162fbe.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 324, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 360 + ], + "type": "text", + "content": "This establishes the Evidence Lower Bound (ELBO). Thus, by maximizing this lower bound, we are effectively solving the minimization problem in Eq. (1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "text", + "content": "By assuming each view is conditionally independent from the others given " + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "text", + "content": ", we can efficiently decompose the data likelihood as " + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "inline_equation", + "content": "p(\\{x\\} |c) = \\prod_{i\\in \\mathcal{W}}p\\left(x^{(i)}|c\\right)\\approx \\prod_{i\\in \\mathcal{W}}p_{\\theta_i}\\left(x^{(i)}|c\\right)" + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "inline_equation", + "content": "p_{\\theta_i}\\left(x^{(i)}|c\\right)" + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "text", + "content": " with learnable parameters " + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "text", + "content": " approximates " + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "inline_equation", + "content": "p(x^{(i)}|c)" + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "text", + "content": ". The objective in Eq. (2) involves extracting consistent features from different views. To facilitate subsequent analysis, we introduce an identifier for the consistent features of each view to distinguish them, i.e., " + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\{c^{(v)}\\}_{v = 1}^m" + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "text", + "content": ". Introducing this identifier does not alter the model architecture. Additionally, we introduce a trade-off parameter " + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "text", + "content": " [10], which controls the trade-off between reconstruction quality and regularization, resulting in the following lower bound:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 533, + 294, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 533, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 64, + 533, + 294, + 609 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {c o n s i s t}} = \\\\ \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ \\mathbb {E} _ {q _ {\\phi_ {v}} (c ^ {(v)} | x ^ {(v)})} \\left[ \\sum_ {i \\in \\mathcal {W}} \\log p _ {\\theta_ {i}} (x ^ {(i)} | c ^ {(v)}) \\right] \\right. \\\\ \\left. - \\alpha D _ {K L} \\left(q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) \\| p (c)\\right) \\right] \\tag {3} \\\\ \\end{array}", + "image_path": "e302b8fd9c3a8d8efc9cdd1e477450f6386decb4be5761ed3f7dfac4a69779d2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "text", + "content": "here " + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "inline_equation", + "content": "q_{\\phi_v}(c^{(v)} \\mid x^{(v)})" + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "text", + "content": " can be regarded as an encoder, and we assume that " + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "inline_equation", + "content": "q_{\\phi_v}(c^{(v)} \\mid x^{(v)}) \\sim \\mathcal{N}(\\mu_{\\phi_v}(x^{(v)}), \\sigma_{\\phi_v}^2(x^{(v)}) I)" + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\mu_{\\phi_v}(x^{(v)})" + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\sigma_{\\phi_v}^2(x^{(v)})" + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "text", + "content": " denote the mean and variance encoder of the " + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "text", + "content": "th view, and " + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 609, + 296, + 694 + ], + "type": "text", + "content": " denotes the identity matrix. Due to the non-differentiable nature of the sampling process, we introduce the reparameterization trick [16] to obtain the features:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 104, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "c ^ {(v)} = \\mu_ {\\phi_ {v}} \\left(x ^ {(v)}\\right) + \\sigma_ {\\phi_ {v}} \\left(x ^ {(v)}\\right) \\odot \\delta \\tag {4}", + "image_path": "12fc09faf0a710c175003002ea0881e3a6d0877f06749cb2c9b2e1c6dfbd368a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\delta \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " denotes the Hadamard product, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "c^{(v)}\\in \\mathbb{R}^{d_e}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "d_{e}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " represents the feature dimension. The distribution " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "p_{\\theta_i}(x^{(i)}|c^{(v)})" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " in Eq. (3) is modeled as an MLP decoder, assuming " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "p_{\\theta_i}(x^{(i)}|c^{(v)})\\sim \\mathcal{N}(\\mu_{\\theta_i}(c^{(v)}),I)" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", yielding the reconstructed feature " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\hat{x}^{(i,v)} = \\mu_{\\theta_i}(c^{(v)})" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "text", + "content": "A brief analysis of Eq. (3) is as follows: The first term represents both self- and cross-view reconstruction. It allows " + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "inline_equation", + "content": "c^{(v)}" + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "text", + "content": " from each view to reconstruct the original data of different views, which helps " + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "inline_equation", + "content": "c^{(v)}" + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "text", + "content": " capture shared information. The second term, the KL divergence, aligns the posterior " + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "inline_equation", + "content": "q_{\\phi_v}(c^{(v)} \\mid x^{(v)})" + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "text", + "content": " of each view with the shared prior " + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "inline_equation", + "content": "p(c) \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "text", + "content": ". These two terms simultaneously encourage " + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "inline_equation", + "content": "c^{(v)}" + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "text", + "content": " to encode the shared information between views." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 228, + 554, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 554, + 324 + ], + "type": "text", + "content": "Finally, in order to obtain a fused consistent feature shared by multiple views for the classification task, we compute the multi-view joint posterior " + }, + { + "bbox": [ + 313, + 228, + 554, + 324 + ], + "type": "inline_equation", + "content": "p(c|\\{x\\})" + }, + { + "bbox": [ + 313, + 228, + 554, + 324 + ], + "type": "text", + "content": " using the Product-of-Expert (PoE) approach [11]. This method is flexible and effective in dealing with the missing views by directly removing the corresponding parameter when a view is missing. According to the setting in [32], we compute " + }, + { + "bbox": [ + 313, + 228, + 554, + 324 + ], + "type": "inline_equation", + "content": "p(c|\\{x\\})" + }, + { + "bbox": [ + 313, + 228, + 554, + 324 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 355, + 329, + 553, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 329, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 355, + 329, + 553, + 383 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p (c | \\{x \\}) \\propto p (c) \\prod_ {v \\in \\mathcal {W}} p (c \\mid x ^ {(v)}) \\\\ := p (c) \\prod_ {v \\in \\mathcal {W}} q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) \\tag {5} \\\\ \\end{array}", + "image_path": "47d6aa297981a15badf67982801f4ede4ee65183675b628b8ff7f3730471c0e4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "type": "text", + "content": "The same reparameterization trick is used to obtain the fused multi-view consistent feature " + }, + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "type": "inline_equation", + "content": "\\bar{c} \\in \\mathbb{R}^{d_e}" + }, + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "type": "text", + "content": ", which is sampled from " + }, + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "type": "inline_equation", + "content": "p(c|\\{x\\})" + }, + { + "bbox": [ + 313, + 388, + 554, + 425 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 431, + 495, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 495, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 495, + 443 + ], + "type": "text", + "content": "3.3. Disentangled view-specific feature" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "type": "text", + "content": "In this section, we introduce view-specific features " + }, + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\{s^{(v)}\\}_{v = 1}^{m}" + }, + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "type": "text", + "content": " that encode information specific to each view beyond the consistent features. We consider minimizing the mutual information " + }, + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "type": "inline_equation", + "content": "I(c^{(v)};s^{(v)})" + }, + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "type": "text", + "content": " between the consistent and specific features for each view, which allows us to maximally disentangle the specific features from the consistent features. A natural idea is to do the learning of consistent features first, followed by disentangling the specific features from them [15]. Thus, we first pre-train with " + }, + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{consist}}" + }, + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "type": "text", + "content": " for a certain number of epochs, depending on the dataset size. In the second stage, we decrease the learning rate for the encoders of consistent features " + }, + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\{q_{\\phi_v}(c^{(v)}\\mid x^{(v)})\\}_{v = 1}^m" + }, + { + "bbox": [ + 313, + 449, + 555, + 616 + ], + "type": "text", + "content": " to stabilize the pre-trained features during further disentanglement. Our disentangling objective function is as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 378, + 622, + 553, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 622, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 378, + 622, + 553, + 651 + ], + "type": "interline_equation", + "content": "\\min \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} I \\left(c ^ {(v)}; s ^ {(v)}\\right) \\tag {6}", + "image_path": "6b3205397aa918574c71d4b04a15293010e36e3a3e582d6f6abc2885236203dd.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 658, + 553, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 553, + 682 + ], + "type": "text", + "content": "We rewrite " + }, + { + "bbox": [ + 313, + 658, + 553, + 682 + ], + "type": "inline_equation", + "content": "I(c^{(v)}; s^{(v)})" + }, + { + "bbox": [ + 313, + 658, + 553, + 682 + ], + "type": "text", + "content": " as follows, more details are provided in the supplementary material Sec. 6.2:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 687, + 553, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 687, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 687, + 553, + 713 + ], + "type": "interline_equation", + "content": "I \\left(c ^ {(v)}; s ^ {(v)}\\right) = I \\left(x ^ {(v)}; c ^ {(v)}\\right) + I \\left(x ^ {(v)}; s ^ {(v)}\\right) - I \\left(x ^ {(v)}; c ^ {(v)}, s ^ {(v)}\\right) \\tag {7}", + "image_path": "2ca916e10a68094513e86be4fd2c9658c17a9ce298a712cb207fbfd6045f2b16.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30724" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 287, + 219 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 287, + 219 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 287, + 219 + ], + "type": "image", + "image_path": "20f4de535dc3f2dea301c5a8e6a84c11cf76cfb33fab2ca7e619a873cab5b8a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 220, + 140, + 316 + ], + "blocks": [ + { + "bbox": [ + 61, + 220, + 140, + 316 + ], + "lines": [ + { + "bbox": [ + 61, + 220, + 140, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 140, + 316 + ], + "type": "image", + "image_path": "06eb3429ae8cb37d63d1552a95323cfe190c63b0f879a9644f76d77c6b4c21cd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 326, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 326, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 555, + 360 + ], + "type": "text", + "content": "Figure 1. The main framework of DRLS. The upper-left corner of the figure represents the view-consistent feature learning module, the upper-right corner shows the disentangled view-specific feature learning module, and the lower part depicts the label semantic embedding learning module." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 145, + 220, + 209, + 304 + ], + "blocks": [ + { + "bbox": [ + 145, + 220, + 209, + 304 + ], + "lines": [ + { + "bbox": [ + 145, + 220, + 209, + 304 + ], + "spans": [ + { + "bbox": [ + 145, + 220, + 209, + 304 + ], + "type": "image", + "image_path": "011e2c2d4a035e7f5163749b37ce9d34224ec98559f2a5996647c8faf2dcfa35.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 213, + 220, + 299, + 314 + ], + "blocks": [ + { + "bbox": [ + 213, + 220, + 299, + 314 + ], + "lines": [ + { + "bbox": [ + 213, + 220, + 299, + 314 + ], + "spans": [ + { + "bbox": [ + 213, + 220, + 299, + 314 + ], + "type": "image", + "image_path": "1cf3ddb39043e7ebe6d16790b34b36cf72f6819ee4ae77f783c06b509e72fbf3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 299, + 71, + 553, + 225 + ], + "blocks": [ + { + "bbox": [ + 299, + 71, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 299, + 71, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 299, + 71, + 553, + 225 + ], + "type": "image", + "image_path": "65044f6f5d93d7f6168d0e82a77bafbc7facf670a5c8779c5b35b3e9034cb6b2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 299, + 224, + 553, + 315 + ], + "blocks": [ + { + "bbox": [ + 299, + 224, + 553, + 315 + ], + "lines": [ + { + "bbox": [ + 299, + 224, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 299, + 224, + 553, + 315 + ], + "type": "image", + "image_path": "cd8d930529901c6b889138194a7ce6a7a683230b041bfada56d8dc8ea74cbbef.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 366, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 415 + ], + "type": "text", + "content": "Clearly, directly optimizing Eq. (7) is challenging. Therefore, we introduce the variational upper bounds for " + }, + { + "bbox": [ + 55, + 366, + 295, + 415 + ], + "type": "inline_equation", + "content": "I(x^{(v)};c^{(v)})" + }, + { + "bbox": [ + 55, + 366, + 295, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 366, + 295, + 415 + ], + "type": "inline_equation", + "content": "I(x^{(v)};s^{(v)})" + }, + { + "bbox": [ + 55, + 366, + 295, + 415 + ], + "type": "text", + "content": ", as well as the variational lower bound for " + }, + { + "bbox": [ + 55, + 366, + 295, + 415 + ], + "type": "inline_equation", + "content": "I(x^{(v)};c^{(v)},s^{(v)})" + }, + { + "bbox": [ + 55, + 366, + 295, + 415 + ], + "type": "text", + "content": " [14], resulting in:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 418, + 294, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 294, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 294, + 478 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I \\left(c ^ {(v)}; s ^ {(v)}\\right) \\leq D _ {K L} \\left(q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) \\| p \\left(c ^ {(v)}\\right)\\right) \\\\ + D _ {K L} \\left(q _ {\\varphi_ {v}} \\left(s ^ {(v)} \\mid x ^ {(v)}\\right) \\| p \\left(s ^ {(v)}\\right)\\right) \\\\ - \\mathbb {E} _ {q _ {\\phi_ {v}} (c ^ {(v)} | x ^ {(v)}) q _ {\\varphi_ {v}} (s ^ {(v)} | x ^ {(v)})} [ \\log p _ {\\vartheta_ {v}} (x ^ {(v)} | c ^ {(v)}, s ^ {(v)}) ] \\tag {8} \\\\ \\end{array}", + "image_path": "01a8786a15a268f9994117e7361826e462197a69d7b58476b5cd607e63c7d5db.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": "We observe that minimizing this variational upper bound on mutual information trains a decoder " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "p_{\\vartheta_v}(x^{(v)}|c^{(v)}, s^{(v)})" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": " and a specific feature encoder " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "q_{\\varphi_v}(s^{(v)}|x^{(v)})" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": ". Here, we concatenate " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "c^{(v)}" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "s^{(v)}" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": " as input to the MLP decoder " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "p_{\\vartheta_v}(x^{(v)}|c^{(v)}, s^{(v)})" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": " to produce the reconstructed feature " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\hat{x}^{(v)}" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": ". Similar to Sec. 3.2, " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "q_{\\varphi_v}(s^{(v)}|x^{(v)}) \\sim \\mathcal{N}(\\mu_{\\varphi_v}(x^{(v)}), \\sigma_{\\varphi_v}^2(x^{(v)})I)" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": " parameterizes the distribution of the specific features. The feature " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "s^{(v)} \\in \\mathbb{R}^{d_e}" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": " is derived using the reparameterization trick as shown in Eq. (4). We assume that " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "p(c^{(v)})" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "p(s^{(v)})" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": " follow standard Gaussian distributions. Finally, we obtain the disentanglement loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 616, + 294, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 616, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 59, + 616, + 294, + 689 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {d i s e n t}} = \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} \\left[ D _ {K L} \\left(q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) \\| p \\left(c ^ {(v)}\\right)\\right) \\right. \\\\ + D _ {K L} \\left(q _ {\\varphi_ {v}} \\left(s ^ {(v)} \\mid x ^ {(v)}\\right) \\| p \\left(s ^ {(v)}\\right)\\right) \\\\ \\left. - \\mathbb {E} _ {q _ {\\phi_ {v}} \\left(c ^ {(v)} \\mid x ^ {(v)}\\right) q _ {\\varphi_ {v}} \\left(s ^ {(v)} \\mid x ^ {(v)}\\right)} \\left[ \\log p _ {\\vartheta_ {v}} \\left(x ^ {(v)} \\mid c ^ {(v)}, s ^ {(v)}\\right) \\right] \\right] \\tag {9} \\\\ \\end{array}", + "image_path": "95851929244ea3175d479f64146d7d78dc3ab782cd651059628ab411c385bd7a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Comparing Eq. (3) and Eq. (9), the consistent objective function requires each consistent feature " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "c^{(v)}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " to reconstruct" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": "the original features across views, whereas in the disentanglement loss, we use the concatenation of consistent and specific features exclusively for self-view reconstruction. Since we pre-train a set of encoders " + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\{q_{\\phi_v}(c^{(v)}|x^{(v)})\\}_{v = 1}^m" + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": " in the second stage, " + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "inline_equation", + "content": "c^{(v)}" + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "inline_equation", + "content": "s^{(v)}" + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": " are concatenated to reconstruct " + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "inline_equation", + "content": "x^{(v)}" + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "inline_equation", + "content": "c^{(v)}" + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": " already contains the information shared between " + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "inline_equation", + "content": "x^{(v)}" + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": " and other views, " + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "inline_equation", + "content": "s^{(v)}" + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": " only needs to encode the complementary information that is not captured by " + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "inline_equation", + "content": "c^{(v)}" + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": ". This pre-training strategy enables more efficient disentanglement. We use a simple and effective method to obtain the fused multi-view specific feature " + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\bar{s}\\in \\mathbb{R}^{d_e}" + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": ", for subsequent classification tasks:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 395, + 515, + 555, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 515, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 395, + 515, + 555, + 544 + ], + "type": "interline_equation", + "content": "\\bar {s} = \\frac {1}{| \\mathcal {W} |} \\sum_ {v \\in \\mathcal {W}} s ^ {(v)} \\tag {10}", + "image_path": "fd999ac2e45df7f902df1ead1331a34da4d1958b627717b7fdc0685e7eaf41e6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 550, + 555, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 555, + 657 + ], + "type": "text", + "content": "Here, we do not use PoE to fuse the specific features as in Sec. 3.2, because it is more suitable for merging shared information. PoE is sensitive to consensus among different views [18]. Specifically, if one view's probability distribution is overly sharp and biased, the joint distribution will have a biased mean prediction [1, 18, 26]. Given that specific features from different views often have differences, we choose a simple average fusion method to mitigate extreme variations and retain the overall trend." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 658, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 554, + 693 + ], + "type": "text", + "content": "Finally, we fuse the multi-view consistent and specific information. Following the method in [23], we use specific features to enhance the consistent features:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 405, + 701, + 555, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 405, + 701, + 555, + 714 + ], + "type": "interline_equation", + "content": "z = \\omega (\\bar {s}) \\odot \\bar {c} \\tag {11}", + "image_path": "528509b40acec766369bff3d690eb885c4a67f2f4d741500a971ecf4f27122ca.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30725" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{d_e}" + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": " represents the final fused representation, " + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "inline_equation", + "content": "\\omega(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": " is the sigmoid activation function, and " + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": " denotes the Hadamard product." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 115, + 205, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 115, + 205, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 115, + 205, + 129 + ], + "type": "text", + "content": "3.4. Label semantic embedding" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 296, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 296, + 217 + ], + "type": "text", + "content": "To fully utilize the incomplete supervisory information, we introduce label semantic embedding learning guided by prior information. Here, we construct our model within the framework of the Variational Graph Autoencoder (VGAE) [17], which employs variational inference to model the latent distribution of label embeddings during the learning process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": "First, we define a relationship graph among labels, where the graph " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "G = (V,E)" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": " represents the label relationship graph, with " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": " denoting the nodes (categories) and " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": " representing the edges connecting the nodes. We model the adjacency matrix " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{k \\times k}" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": " of graph " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": " in terms of conditional probabilities, specifically " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "P(l_j|l_i)" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": ", which denotes the probability that label " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "l_j" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": " occurs when label " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": " occurs, where " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": "th label. This is computed based on the label matrix " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "Y \\in \\mathbb{R}^{n_{train} \\times k}" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": " from the training set. Since " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "P(l_j|l_i)" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": " is not equal to " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "P(l_i|l_j)" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": ", the adjacency matrix " + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 217, + 296, + 361 + ], + "type": "text", + "content": " is asymmetric, which more accurately reflects real-world scenarios. The specific calculation method is as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 369, + 296, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 369, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 104, + 369, + 296, + 401 + ], + "type": "interline_equation", + "content": "A _ {i j} = \\left\\{ \\begin{array}{l l} P \\left(l _ {j} \\mid l _ {i}\\right) = \\frac {C _ {i j}}{d _ {i}}, & \\text {i f} i \\neq j \\\\ 0, & \\text {i f} i = j \\end{array} \\right. \\tag {12}", + "image_path": "e5f47ddba56a857e6c11ed4a72febf8a4f36d5333d92ea9519347ffbac7f11ef.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "inline_equation", + "content": "C = Y^{\\top}Y" + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "inline_equation", + "content": "C \\in \\mathbb{R}^{k \\times k}" + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": " is the label co-occurrence matrix, " + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "inline_equation", + "content": "C_{ij}" + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": " is the co-occurrence count of labels " + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "inline_equation", + "content": "l_j" + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "inline_equation", + "content": "d_i = \\sum_{a=1}^{n_{train}} Y_{ai} = Y_{:i}^{\\top} Y_{:i}" + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "inline_equation", + "content": "i = 1,2,\\ldots,k" + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": ", which represents the total occurrence count of label " + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": ". It should be noted that we set the diagonal of the adjacency matrix " + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": " to 0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 480, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 540 + ], + "type": "text", + "content": "Graph Isomorphism Network (GIN) provides stronger feature aggregation capabilities through learnable weights and information from self-nodes [34]. Therefore, we use GIN to model the posterior distribution of label semantic embeddings " + }, + { + "bbox": [ + 55, + 480, + 296, + 540 + ], + "type": "inline_equation", + "content": "H \\in \\mathbb{R}^{k \\times d_e}" + }, + { + "bbox": [ + 55, + 480, + 296, + 540 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 548, + 295, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 548, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 107, + 548, + 295, + 581 + ], + "type": "interline_equation", + "content": "q \\left(H \\mid H ^ {0}, A\\right) = \\prod_ {i = 1} ^ {k} q \\left(h _ {i} \\mid H ^ {0}, A\\right), \\tag {13}", + "image_path": "045dc683d1a906b05c7e70d813249379768ee430b39b7b06ecbf5ba7bd4ef286.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 109, + 582, + 242, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 582, + 242, + 597 + ], + "spans": [ + { + "bbox": [ + 109, + 582, + 242, + 597 + ], + "type": "interline_equation", + "content": "\\mathrm {w i t h} q (h _ {i} | H ^ {0}, A) \\sim \\mathcal {N} (\\mu_ {i}, \\sigma_ {i} ^ {2} I)", + "image_path": "52e9b46b57b130f4789b7f31e80f6b8b103a29d6e104ec38362ede2f6a12899e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": "here, " + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "inline_equation", + "content": "h_i \\in \\mathbb{R}^{d_e}" + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": "-th row of " + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": ", representing the semantic embedding of the " + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": "-th label, and " + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "inline_equation", + "content": "H^0 \\in \\mathbb{R}^{k \\times k}" + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": " is an identity matrix, which means that the embedding for label " + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": " is initialized as a one-hot vector with the " + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": "-th position set to 1. This approach ensures that the initial distance between different labels is large, which helps to better distinguish the semantic differences between labels during the subsequent learning process. " + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} = \\mathrm{GIN}_{\\pmb{\\mu}}(H^0, A)" + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^{k \\times d_e}" + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": " is the mean matrix, where the " + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": "-th row corresponds to the mean" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "vector " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "-th label; similarly, " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\sigma^2 = \\mathrm{GIN}_{\\sigma^2}(H^0,A)" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": ". Here, we use a three-layer GIN, where one layer can be defined as " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "H^{(t + 1)} = f^{(t + 1)}[(1 + \\epsilon^{(t + 1)})H^{(t)} + AH^{(t)}]" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " represents the current number of network layers, " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\epsilon^{(t + 1)}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " represents a learnable parameter, and " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "f^{(t + 1)}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " represents a two-layer fully connected network. " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "H^{(t + 1)}\\in \\mathbb{R}^{k\\times d_{t + 1}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " represents the output of the intermediate layer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 156, + 554, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 554, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 554, + 204 + ], + "type": "text", + "content": "Subsequently, we also use the reparameterization trick to sample the label semantic embeddings, with a formula similar to Eq. (4), i.e., " + }, + { + "bbox": [ + 313, + 156, + 554, + 204 + ], + "type": "inline_equation", + "content": "h_i = \\mu_i + \\sigma_i \\odot \\delta, \\delta \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 156, + 554, + 204 + ], + "type": "text", + "content": ". Finally, we obtain the label semantic embedding matrix " + }, + { + "bbox": [ + 313, + 156, + 554, + 204 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 156, + 554, + 204 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 204, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 228 + ], + "type": "text", + "content": "We use a simple and efficient inner product between variables as a generative model:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 237, + 553, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 237, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 365, + 237, + 553, + 271 + ], + "type": "interline_equation", + "content": "p (A | H) = \\prod_ {i = 1} ^ {k} \\prod_ {j = 1} ^ {k} p \\left(A _ {i j} \\mid h _ {i}, h _ {j}\\right), \\tag {14}", + "image_path": "ba869412e6c792fd1b0bc374660c0905a24a9d9a01b9f3abcdbd476cebeba642.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 375, + 273, + 502, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 273, + 502, + 289 + ], + "spans": [ + { + "bbox": [ + 375, + 273, + 502, + 289 + ], + "type": "interline_equation", + "content": "\\text {w i t h} p \\left(A _ {i j} \\mid h _ {i}, h _ {j}\\right) = \\omega \\left(h _ {i} ^ {\\top} h _ {j}\\right)", + "image_path": "6362a4d0d696c7fac1d71d0f62a17a10c24bd3620fd372b0bddffcc48de754c8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 297, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 297, + 555, + 417 + ], + "type": "inline_equation", + "content": "\\omega(\\cdot)" + }, + { + "bbox": [ + 313, + 297, + 555, + 417 + ], + "type": "text", + "content": " denotes the sigmoid activation function. By using the label correlation matrix as the adjacency matrix of a graph, we enable GIN to aggregate semantically related embeddings to nearby positions in the label semantic space. However, it cannot guarantee that unrelated labels are effectively pushed apart. Therefore, we introduce the generative model from VGAE to reconstruct the adjacency matrix " + }, + { + "bbox": [ + 313, + 297, + 555, + 417 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 297, + 555, + 417 + ], + "type": "text", + "content": ", ensuring that the label semantic embeddings effectively preserve the topology of label correlations in the semantic space. Finally, our variational lower bound is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 426, + 553, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 553, + 450 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l e} = \\mathbb {E} _ {q (H | H ^ {0}, A)} [ \\log p (A | H) ] - D _ {K L} (q (H | H ^ {0}, A) \\| p (H)) \\tag {15}", + "image_path": "6067d944e64c6520c9c39bcd43412bfd4862eadd2988596a6eb10da1f816daeb.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "type": "text", + "content": "We further adopt a Gaussian prior " + }, + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "type": "inline_equation", + "content": "p(H) = \\prod_{i}p(h_{i}) = \\prod_{i}\\mathcal{N}(h_{i}|0,I)" + }, + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "type": "text", + "content": ". We use binary cross-entropy to compute the first term of " + }, + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{le}" + }, + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "type": "text", + "content": " as the reconstruction loss and add an identity matrix to " + }, + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "type": "text", + "content": " before computing the cross-entropy loss to avoid enforcing " + }, + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "type": "inline_equation", + "content": "\\omega (h_i^\\top h_i)\\approx 0" + }, + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "type": "text", + "content": " due to its zero diagonal." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 518, + 434, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 434, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 434, + 530 + ], + "type": "text", + "content": "3.5. Overall loss function" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "text", + "content": "Finally, we utilize label semantic embeddings for feature selection, choosing discriminative and specific features for each label. First, the learned label semantic embedding " + }, + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "text", + "content": " is mapped to a feature importance vector [9] via a fully connected layer " + }, + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "text", + "content": ", specifically " + }, + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "inline_equation", + "content": "r_i = \\omega(f(h_i))" + }, + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "inline_equation", + "content": "r_i \\in \\mathbb{R}^{d_e}" + }, + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\omega(\\cdot)" + }, + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "text", + "content": " is the sigmoid activation function. Subsequently, we use " + }, + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 313, + 536, + 555, + 633 + ], + "type": "text", + "content": " to select the features and input them into a simple neural network to obtain predictions:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 388, + 643, + 553, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 643, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 388, + 643, + 553, + 656 + ], + "type": "interline_equation", + "content": "p _ {i} = \\omega \\left(f _ {\\text {c l s}} \\left(z \\odot r _ {i}\\right)\\right) \\tag {16}", + "image_path": "12c827084c7775b0ccc7fd7d1c77634bf43f6a9efa62b6cf7a1cc933fd9fd87b.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " denotes the Hadamard product, " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "f_{cls}" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " is a fully connected network, and " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " is the predicted probability of the sample being labeled as the " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "-th class. We use the commonly used binary cross-entropy function to calculate the" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30726" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 125, + 267, + 203 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": "Table 1. Statistics for five multi-view multi-label datasets, where " + }, + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": " denotes the number of views, " + }, + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": " denotes the number of samples, " + }, + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": " denotes the number of categories, and " + }, + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "inline_equation", + "content": "n/k" + }, + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": " denotes the average number of labels per sample." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 125, + 267, + 203 + ], + "lines": [ + { + "bbox": [ + 86, + 125, + 267, + 203 + ], + "spans": [ + { + "bbox": [ + 86, + 125, + 267, + 203 + ], + "type": "table", + "html": "
Datasetmnkn/k
Corel5k649992603.396
Pascal0769963201.465
Espgame6207702684.686
IAPRTC126196272915.719
Mirflickr625000384.716
", + "image_path": "e68080dc55af55025e06d0db6a806e6235ef2636e113b61ebd4f1ff15bfc8737.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 222, + 180, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 180, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 180, + 233 + ], + "type": "text", + "content": "multi-label classification loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 240, + 294, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 240, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 67, + 240, + 294, + 280 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B C E} = - \\frac {1}{| \\mathcal {U} |} \\sum_ {i \\in \\mathcal {U}} \\left(y _ {i} \\log p _ {i} + \\left(1 - y _ {i}\\right) \\log \\left(1 - p _ {i}\\right)\\right) \\tag {17}", + "image_path": "9be7c9720a3e50b8dcf97f73e8bcc0b5f2dc29e6a0140052654ed4c979084ba8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 281, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 281, + 296, + 316 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 281, + 296, + 316 + ], + "type": "text", + "content": " denotes the ground truth and we exclude missing labels from loss computation to filter out the impact of missing labels. Our final loss function is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 326, + 295, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 326, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 110, + 326, + 295, + 340 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {B C E} + \\beta \\mathcal {L} _ {\\text {d i s e n t}} - \\gamma \\mathcal {L} _ {l e} \\tag {18}", + "image_path": "3b7b5e2af901ec7b9ac1e528186166a8b08446a5f7d4ede37ed19ac62b6a24f2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 349, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 349, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 349, + 296, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 349, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 349, + 296, + 373 + ], + "type": "text", + "content": " are trade-off parameters used to balance the different loss terms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 384, + 137, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 137, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 137, + 398 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 404, + 119, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 119, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 119, + 415 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 423, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 531 + ], + "type": "text", + "content": "We conduct experiments based on the settings from [21, 23, 31] to evaluate the performance of our proposed model on five datasets. They are Corel5k [5], Pascal07 [6], Espgame [29], IAPRTC12 [8], and Mirflickr [13]. In these five datasets, we utilize six different types of features, which are DenseHue, DenseSift, GIST, HSV, RGB, and LAB, with dimensions of 100, 1000, 512, 4096, 4096, and 4096, respectively. More detailed information can be found in Tab. 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 539, + 171, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 171, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 171, + 552 + ], + "type": "text", + "content": "4.2. Compared methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 558, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 715 + ], + "type": "text", + "content": "In our experiments, we select eight methods for comparison with our model: DM2L [24], CDMM [40], iMvWL [27], NAIML [20], DDINet [31], DICNet [22], MTD [23], and SIP [21]. It is important to note that iMvWL, NAIML, DDINet, DICNet, MTD, and SIP are incomplete multi-view multi-label methods. CDMM can only handle complete multi-view multi-label data; therefore, we follow the settings from [21] and use the mean of available instances from the corresponding view to fill in the unavailable instances for missing views. To handle incomplete labels, we treat the missing labels as negative labels. DM2L is a single-view incomplete multi-label method, and we record the best result of each view in our experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 72, + 446, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 446, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 446, + 84 + ], + "type": "text", + "content": "4.3. Implementation details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 90, + 555, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 555, + 224 + ], + "type": "text", + "content": "For the datasets, in order to simulate real-world data missingness, we generate incomplete datasets following the settings in [27] and [21]. For each view, we randomly exclude " + }, + { + "bbox": [ + 313, + 90, + 555, + 224 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 90, + 555, + 224 + ], + "type": "text", + "content": " of the instances, ensuring at least one view per sample. Additionally, we randomly remove " + }, + { + "bbox": [ + 313, + 90, + 555, + 224 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 90, + 555, + 224 + ], + "type": "text", + "content": " of both positive and negative labels. We fill the missing features and labels with zeros. We use " + }, + { + "bbox": [ + 313, + 90, + 555, + 224 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 90, + 555, + 224 + ], + "type": "text", + "content": " of the samples for training and the remaining " + }, + { + "bbox": [ + 313, + 90, + 555, + 224 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 90, + 555, + 224 + ], + "type": "text", + "content": " for validation and testing. To ensure the stability of the results, we repeat the experiment 10 times and report the mean and variance of the results. Additional details are in the supplementary material Sec. 7." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 228, + 426, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 426, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 426, + 239 + ], + "type": "text", + "content": "4.4. Evaluation metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 246, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 354 + ], + "type": "text", + "content": "Consistent with some previous studies [21, 23, 31], in our experiments, we evaluate the performance of the model using six commonly used multi-label classification metrics, namely, Average Precision (AP), Area Under the Receiver Operating Characteristic Curve (AUC), Ranking Loss (RL), Hamming Loss (HL), OneError (OE), and Coverage (Cov). To better visualize the results, we use 1-RL, 1-HL, 1-OE, and 1-Cov to record the final results, where larger values indicate better performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 360, + 435, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 360, + 435, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 360, + 435, + 372 + ], + "type": "text", + "content": "4.5. Experimental results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 378, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 534 + ], + "type": "text", + "content": "In Tab. 2, we present the experimental results of nine methods on five public multi-view multi-label datasets, which include six evaluation metrics. The results in Tab. 2 are obtained under experimental settings with a " + }, + { + "bbox": [ + 313, + 378, + 555, + 534 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 378, + 555, + 534 + ], + "type": "text", + "content": " view missing rate, a " + }, + { + "bbox": [ + 313, + 378, + 555, + 534 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 378, + 555, + 534 + ], + "type": "text", + "content": " label missing rate, and " + }, + { + "bbox": [ + 313, + 378, + 555, + 534 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 378, + 555, + 534 + ], + "type": "text", + "content": " of training samples. The results in Tab. 2 show that our method outperforms the other eight competitors on each metric in most datasets. For example, on the AP metric of the IAPRTC12 dataset, our method is " + }, + { + "bbox": [ + 313, + 378, + 555, + 534 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 313, + 378, + 555, + 534 + ], + "type": "text", + "content": " higher than the second-ranked MTD. Specifically, Ave.R represents the average ranking of each method across the six metrics, and DRLS ranks first on all five datasets, which fully demonstrates the effectiveness of our method in addressing the DIMVMLC problem." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 534, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 594 + ], + "type": "text", + "content": "In addition, further analysis of the experimental results reveals that CDMM, DDINet, DICNet, MTD, and SIP outperform the other three methods on all five datasets. These are deep learning methods, indicating the great potential of deep learning in addressing the DIMVMLC problem." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Besides solving the DIMVMLC problem, we also further validated the performance of our method in addressing the MVMLC problem. In Tab. 3, we present the experimental results of four methods on five datasets, with experiments conducted under the settings of " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " training samples, full views, and full labels. We select three representative evaluation metrics. From the results in Tab. 3, it is evident that our model still demonstrates competitive performance under this setup, indicating that our method can effectively address the MVMLC problem." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30727" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 95, + 553, + 533 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 553, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 553, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 553, + 93 + ], + "type": "text", + "content": "Table 2. Experimental results of nine methods on five datasets with " + }, + { + "bbox": [ + 55, + 70, + 553, + 93 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 70, + 553, + 93 + ], + "type": "text", + "content": " missing views, " + }, + { + "bbox": [ + 55, + 70, + 553, + 93 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 70, + 553, + 93 + ], + "type": "text", + "content": " missing labels, and " + }, + { + "bbox": [ + 55, + 70, + 553, + 93 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 70, + 553, + 93 + ], + "type": "text", + "content": " training samples. Ave.R denotes the average ranking of each method on the six metrics, and the numbers at the bottom right are the standard deviations." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 95, + 553, + 533 + ], + "lines": [ + { + "bbox": [ + 56, + 95, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 553, + 533 + ], + "type": "table", + "html": "
DatasetMetricDM2LCDMMiMvWLNAIMLDDINetDICNetMTDSIPDRLS
Corel5kAP0.2620.0050.3540.0040.2830.0080.3090.0040.3600.0090.3780.0040.4130.0070.4160.0150.4330.008
1-HL0.9870.0000.9870.0000.9780.0000.9870.0000.9870.0000.9870.0000.9880.0000.9880.0000.9880.000
1-RL0.8430.0020.8840.0030.8650.0050.8780.0020.8650.0050.8770.0040.8920.0040.9100.0030.9160.002
AUC0.8450.0020.8880.0030.8680.0050.8810.0020.8680.0050.8810.0030.8950.0040.9120.0030.9180.002
1-OE0.2950.0140.4100.0070.3110.0150.3500.0090.4370.0120.4640.0120.4910.0100.4920.0180.5090.019
1-Cov0.6470.0050.7230.0070.7020.0080.7250.0050.6890.0120.7140.0100.7480.0090.7860.0070.8040.006
Ave.R8.1674.8337.6675.3336.0004.8332.6671.8331.000
Pascal07AP0.4710.0080.5080.0050.4370.0180.4880.0030.5320.0100.5020.0070.5500.0040.5500.0090.5670.008
1-HL0.9280.0010.9310.0010.8820.0040.9280.0010.9320.0010.9300.0010.9320.0010.9310.0020.9340.001
1-RL0.7610.0050.8120.0040.7360.0150.7830.0010.8080.0050.7810.0070.8300.0030.8250.0060.8430.004
AUC0.7790.0040.8380.0030.7670.0150.8110.0010.8290.0040.8050.0060.8490.0040.8450.0050.8640.003
1-OE0.4200.0110.4190.0080.3620.0230.4210.0060.4480.0150.4260.0130.4570.0080.4630.0120.4770.011
1-Cov0.6920.0040.7590.0030.6770.0150.7270.0020.7570.0050.7280.0070.7830.0040.7770.0050.7980.004
Ave.R7.6674.8339.0006.5004.1676.1672.1672.8331.000
EspaugeAP0.2120.0020.2890.0030.2440.0050.2460.0020.2860.0040.2990.0040.3060.0030.3100.0040.3260.005
1-HL0.9820.0000.9830.0000.9720.0000.9830.0000.9830.0000.9830.0000.9830.0000.9830.0000.9830.000
1-RL0.7810.0010.8320.0010.8080.0020.8180.0020.8150.0030.8330.0030.8370.0010.8490.0020.8580.002
AUC0.7850.0010.8360.0010.8130.0020.8240.0020.8190.0030.8370.0020.8420.0010.8530.0020.8620.002
1-OE0.2940.0060.3960.0050.3430.0130.3390.0030.4270.0100.4370.0100.4480.0060.4510.0120.4730.001
1-Cov0.4880.0030.5740.0040.5480.0040.5710.0030.5530.0050.5980.0060.6010.0040.6280.0040.6520.003
Ave.R8.8334.5008.0005.6675.5003.5002.6671.8331.000
IAPRTC12AP0.2340.0030.3050.0040.2370.0030.2610.0010.3020.0050.3270.0050.3320.0020.3310.0070.3560.006
1-HL0.9800.0000.9810.0000.9690.0000.9800.0000.9800.0000.9800.0000.9810.0000.9810.0000.9810.000
1-RL0.8230.0020.8620.0020.8330.0020.8480.0010.8530.0020.8720.0020.8750.0010.8870.0040.8960.003
AUC0.8250.0010.8640.0020.8350.0010.8500.0010.8550.0030.8730.0010.8760.0010.8880.0030.8980.002
1-OE0.3400.0060.4320.0080.3520.0080.3900.0050.4350.0090.4650.0130.4710.0060.4660.0010.4900.012
1-Cov0.5290.0040.5970.0040.5640.0050.5920.0040.5940.0070.6480.0050.6490.0020.6790.0080.7070.007
Ave.R8.3334.5008.1676.6675.6674.1672.3332.1671.000
MirfickrAP0.5140.0060.5700.0020.4900.0120.5510.0020.5880.0030.5860.0050.6080.0040.6150.0040.6300.005
1-HL0.8780.0010.8860.0010.8390.0020.8820.0010.8880.0010.8880.0010.8910.0010.8910.0010.8950.001
1-RL0.8310.0030.8560.0010.8030.0080.8440.0010.8650.0020.8610.0040.8750.0010.8780.0020.8850.002
AUC0.8280.0030.8460.0010.7870.0120.8370.0010.8530.0020.8480.0040.8610.0020.8640.0020.8720.003
1-OE0.5100.0080.6310.0040.5110.0220.5850.0030.6360.0080.6420.0060.6560.0040.6640.0060.6860.009
1-Cov0.6040.0050.6400.0010.5720.0130.6310.0020.6540.0030.6460.0060.6770.0020.6760.0040.6920.003
Ave.R8.1676.0008.8337.0004.1674.6672.6672.1671.000
", + "image_path": "e82db5108a5de085a2796b0fe3a52d5cba837be74635fef79986ad7c5209b840.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 539, + 168, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 168, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 168, + 552 + ], + "type": "text", + "content": "4.6. Parameter analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "content": "In our DRLS model, there are three hyperparameters that need to be tuned: " + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "content": ". To evaluate the sensitivity of our model to these hyperparameters, we conduct experiments on the Corel5k and Pascal07 datasets, with experiments carried out under conditions where " + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "content": " of the views are missing, " + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "content": " of the labels are missing, and " + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "content": " of the training samples are used." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Figs. 2a and 2b show the AP values of our model under different combinations of " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ", while Figs. 2c and 2d show the AP and AUC values for different values of " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ". To ensure the validity of the experiments, the other unrelated hyperparameters are fixed. Based on Figs. 2a and 2b, for the Corel5k dataset, " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " are taken in the range of" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "inline_equation", + "content": "[1e - 3, 1e - 1]" + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "inline_equation", + "content": "[1e - 5, 1e - 3]" + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "text", + "content": ", respectively, and for the Pascal07 dataset, " + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "text", + "content": " are taken in the range of " + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "inline_equation", + "content": "[1e - 3, 1e - 1]" + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "inline_equation", + "content": "[1e - 5, 1e - 3]" + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "text", + "content": ", respectively. Under these conditions, our model demonstrates good performance. Regarding the hyperparameter " + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "text", + "content": ", from Figs. 2c and 2d, it can be seen that the variation in AP and AUC values is minimal for both datasets, indicating that our model is not sensitive to " + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "text", + "content": " within the range of " + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "inline_equation", + "content": "[1e - 1, 3e - 1]" + }, + { + "bbox": [ + 313, + 539, + 555, + 636 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 646, + 406, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 406, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 406, + 658 + ], + "type": "text", + "content": "4.7. Ablation study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "Our model consists of multiple modules and objective functions. To further analyze the contribution of each module and objective function, we conduct a series of ablation experiments on the Corel5k and Pascal07 datasets under a set-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30728" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 98, + 282, + 294 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "text", + "content": "Table 3. Experimental results of four methods on five datasets with full views, full labels, and " + }, + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "text", + "content": " training samples." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 98, + 282, + 294 + ], + "lines": [ + { + "bbox": [ + 69, + 98, + 282, + 294 + ], + "spans": [ + { + "bbox": [ + 69, + 98, + 282, + 294 + ], + "type": "table", + "html": "
DatasetMetricDICNetMTDSIPDRLS
Corel5kAP0.5040.4860.5200.551
1-RL0.9250.9210.9370.947
AUC0.9280.9240.9390.949
Pascal07AP0.6090.6520.6450.655
1-RL0.8600.8900.8870.889
AUC0.8750.8980.8990.901
EspgameAP0.3880.3970.3560.399
1-RL0.8720.8800.8750.893
AUC0.8750.8830.8780.896
IAPRTC12AP0.4280.4410.3660.447
1-RL0.9100.9160.9040.926
AUC0.9110.9160.9060.927
MirflickrAP0.6610.6820.6760.698
1-RL0.8970.9060.9070.915
AUC0.8790.8880.8900.898
", + "image_path": "ea11110148074d252ae667c62ff0fa31c427203206fee4eaaaa10aa2b1487c68.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 376, + 299, + 477 + ], + "blocks": [ + { + "bbox": [ + 55, + 306, + 296, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 373 + ], + "type": "text", + "content": "Table 4. Ablation results on two datasets with " + }, + { + "bbox": [ + 55, + 306, + 296, + 373 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 306, + 296, + 373 + ], + "type": "text", + "content": " missing views, " + }, + { + "bbox": [ + 55, + 306, + 296, + 373 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 306, + 296, + 373 + ], + "type": "text", + "content": " missing labels, and " + }, + { + "bbox": [ + 55, + 306, + 296, + 373 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 306, + 296, + 373 + ], + "type": "text", + "content": " training samples. 'w/o' denotes 'without'. 'A' denotes the consistent feature extraction module, 'B' denotes the disentangled specific feature extraction module, and DRLS denotes the full model. " + }, + { + "bbox": [ + 55, + 306, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con.kl}" + }, + { + "bbox": [ + 55, + 306, + 296, + 373 + ], + "type": "text", + "content": " represents the KL divergence term of " + }, + { + "bbox": [ + 55, + 306, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{consist}" + }, + { + "bbox": [ + 55, + 306, + 296, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 376, + 299, + 477 + ], + "lines": [ + { + "bbox": [ + 56, + 376, + 299, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 299, + 477 + ], + "type": "table", + "html": "
MethodCorel5kPascal07
AP1-RLAUCAP1-RLAUC
A0.4160.9090.9120.5590.8360.858
A+B0.4310.9130.9160.5630.8400.863
DRLS0.4330.9160.9180.5670.8430.864
DRLS w/o Lcon.kl0.4150.9080.9110.5410.8230.846
DRLS w/o Ldisent0.4140.9130.9160.5570.8340.857
DRLS w/o Lle0.4300.9130.9150.5650.8420.863
", + "image_path": "fc8debaba44af15c26aaa3f890c0f917762a371925cebef1ba970b6200fa11e2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 497, + 264, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 264, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 264, + 509 + ], + "type": "text", + "content": "ting of " + }, + { + "bbox": [ + 55, + 497, + 264, + 509 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 497, + 264, + 509 + ], + "type": "text", + "content": " missing views and " + }, + { + "bbox": [ + 55, + 497, + 264, + 509 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 497, + 264, + 509 + ], + "type": "text", + "content": " missing labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "text", + "content": "The results of the ablation experiments are presented in Tab. 4. Specifically, we first test the performance of the model using only the consistent feature extraction module. Then, we add the disentangled specific feature extraction module, and finally, we incorporate the label semantic embedding learning module. Through these three experimental setups, we can observe that each module contributes to the model's performance. Additionally, by observing the results in Tab. 4, we can see that the disentangled specific feature extraction module plays a key role in improving model performance. Furthermore, the label semantic embedding learning module also helps to enhance the model's classification ability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": "In addition, we also validate the impact of different objective functions on model performance. We find that removing " + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con.kl}" + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{disent}" + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": " leads to a significant decline in model performance. Among them, " + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con.kl}" + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": " is the objective" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 59, + 436, + 144 + ], + "blocks": [ + { + "bbox": [ + 316, + 59, + 436, + 144 + ], + "lines": [ + { + "bbox": [ + 316, + 59, + 436, + 144 + ], + "spans": [ + { + "bbox": [ + 316, + 59, + 436, + 144 + ], + "type": "image", + "image_path": "a15c22535848be4de32af0f5b392c3fcbba72a069d899999dee3c8e8037a0753.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 146, + 394, + 155 + ], + "lines": [ + { + "bbox": [ + 356, + 146, + 394, + 155 + ], + "spans": [ + { + "bbox": [ + 356, + 146, + 394, + 155 + ], + "type": "text", + "content": "(a) Corel5k" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 434, + 59, + 555, + 144 + ], + "blocks": [ + { + "bbox": [ + 434, + 59, + 555, + 144 + ], + "lines": [ + { + "bbox": [ + 434, + 59, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 434, + 59, + 555, + 144 + ], + "type": "image", + "image_path": "762165a4b4618d75db47e24feee5dcc0a03738ab1540b1150f19b633ec65a022.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 146, + 514, + 156 + ], + "lines": [ + { + "bbox": [ + 473, + 146, + 514, + 156 + ], + "spans": [ + { + "bbox": [ + 473, + 146, + 514, + 156 + ], + "type": "text", + "content": "(b) Pascal07" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 157, + 436, + 256 + ], + "blocks": [ + { + "bbox": [ + 315, + 157, + 436, + 256 + ], + "lines": [ + { + "bbox": [ + 315, + 157, + 436, + 256 + ], + "spans": [ + { + "bbox": [ + 315, + 157, + 436, + 256 + ], + "type": "image", + "image_path": "6b51f6f29bec8663730269e1b9945b830dfb92baf6f8690da363ee5c9aa7fa0b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 257, + 395, + 266 + ], + "lines": [ + { + "bbox": [ + 356, + 257, + 395, + 266 + ], + "spans": [ + { + "bbox": [ + 356, + 257, + 395, + 266 + ], + "type": "text", + "content": "(c) Corel5k" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 434, + 157, + 553, + 256 + ], + "blocks": [ + { + "bbox": [ + 434, + 157, + 553, + 256 + ], + "lines": [ + { + "bbox": [ + 434, + 157, + 553, + 256 + ], + "spans": [ + { + "bbox": [ + 434, + 157, + 553, + 256 + ], + "type": "image", + "image_path": "00b786215a61f27137b84620256e54a56d597211d3ddddacd89a440dbf6d4e01.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 257, + 514, + 266 + ], + "lines": [ + { + "bbox": [ + 473, + 257, + 514, + 266 + ], + "spans": [ + { + "bbox": [ + 473, + 257, + 514, + 266 + ], + "type": "text", + "content": "(d) Pascal07" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 274, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 319 + ], + "type": "text", + "content": "Figure 2. The AP values under different combinations of " + }, + { + "bbox": [ + 313, + 274, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 274, + 555, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 274, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 274, + 555, + 319 + ], + "type": "text", + "content": " on the Corel5k (a) and Pascal07 (b) datasets; the AP and AUC values for different values of " + }, + { + "bbox": [ + 313, + 274, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 274, + 555, + 319 + ], + "type": "text", + "content": " on the Corel5k (c) and Pascal07 (d) datasets." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 336, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 411 + ], + "type": "text", + "content": "function used in the pre-training stage. The experimental results show that " + }, + { + "bbox": [ + 313, + 336, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con.kl}" + }, + { + "bbox": [ + 313, + 336, + 555, + 411 + ], + "type": "text", + "content": " plays a regularization role, reducing redundant information in the shared features. The disentangling loss " + }, + { + "bbox": [ + 313, + 336, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{disent}" + }, + { + "bbox": [ + 313, + 336, + 555, + 411 + ], + "type": "text", + "content": " is responsible for separating the consistent features from the specific features, which improves the quality of the specific features." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 435, + 388, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 435, + 388, + 449 + ], + "spans": [ + { + "bbox": [ + 314, + 435, + 388, + 449 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 460, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 665 + ], + "type": "text", + "content": "In this paper, we propose a novel incomplete multi-view multi-label learning framework. Based on the assumptions of consistency and complementarity of multi-view data, we establish a method using Disentangled Representation and Label Semantic embedding (DRLS). Firstly, view-consistent features are learned using cross-view reconstruction, and the difficulty of disentangling specific features is reduced by unsupervised pre-training. Next, we achieve disentangled learning of specific features by minimizing the variational upper bound of mutual information between consistent and specific features. Meanwhile, we guide the learning of label semantic embeddings with the help of a priori label relevance to aggregate relevant semantic embeddings and maintain the topology of label relevance in the semantic space. Finally, we conduct a large number of experiments to demonstrate the effectiveness of our method in addressing the issues of missing views and labels." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 669, + 423, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 423, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 423, + 713 + ], + "type": "text", + "content": "Acknowledgment. by National Natural China (62476166) and (22PJD029)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 429, + 669, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 429, + 669, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 429, + 669, + 555, + 704 + ], + "type": "text", + "content": "This work is supported Science Foundation of Shanghai Pujiang Program" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30729" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "text", + "content": "[1] Ana Lawry Aguila and Andre Altmann. A tutorial on multi-view autoencoders using the multi-view-ae library. arXiv preprint arXiv:2403.07456, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "type": "text", + "content": "[2] Luigi Antelmi, Nicholas Ayache, Philippe Robert, and Marco Lorenzi. Sparse multi-channel variational autoencoder for the joint analysis of heterogeneous data. In International Conference on Machine Learning, pages 302-311. PMLR, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 213 + ], + "type": "text", + "content": "[3] Yusheng Cheng, Qingyan Li, Yibin Wang, and Weijie Zheng. Multi-view multi-label learning with view feature attention allocation. Neurocomputing, 501:857-874, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "text", + "content": "[4] Bassoma Diallo, Jie Hu, Tianrui Li, Ghufran Ahmad Khan, Xinyan Liang, and Hongjun Wang. Auto-attention mechanism for multi-view deep embedding clustering. Pattern Recognition, 143:109764, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 323 + ], + "type": "text", + "content": "[5] Pinar Duygulu, Kobus Barnard, Joao FG de Freitas, and David A Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In Computer Vision—ECCV 2002: 7th European Conference on Computer Vision Copenhagen, Denmark, May 28–31, 2002 Proceedings, Part IV 7, pages 97–112. Springer, 2002. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 325, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 367 + ], + "type": "text", + "content": "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "type": "text", + "content": "[7] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 457 + ], + "type": "text", + "content": "[8] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International workshop ontoImage, 2006. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 459, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 295, + 502 + ], + "type": "text", + "content": "[9] Jun-Yi Hang and Min-Ling Zhang. Collaborative learning of label semantics and deep label-specific features for multi-label classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9860-9871, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "type": "text", + "content": "[10] Irina Higgins, Loic Matthew, Arka Pal, Christopher P Burgess, Xavier Glorot, Matthew M Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. *ICLR (Poster)*, 3, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 558, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 590 + ], + "type": "text", + "content": "[11] Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence. Neural computation, 14(8): 1771-1800, 2002. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 592, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 634 + ], + "type": "text", + "content": "[12] Sheng-Jun Huang and Zhi-Hua Zhou. Multi-label learning by exploiting label correlations locally. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 949-955, 2012. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "text", + "content": "[13] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In Proceedings of the 1st ACM international conference on Multimedia information retrieval, pages 39-43, 2008. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[14] HyeongJoo Hwang, Geon-Hyeong Kim, Seunghoon Hong, and Kee-Eung Kim. Variational interaction information maximization for cross-domain disentanglement. Advances in" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "Neural Information Processing Systems, 33:22479-22491, 2020. 4, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 98, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 151 + ], + "type": "text", + "content": "[15] Guanzhou Ke, Bo Wang, Xiaoli Wang, and Shengfeng He. Rethinking multi-view representation learning via distilled disentangling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26774-26783, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 154, + 553, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 553, + 176 + ], + "type": "text", + "content": "[16] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 178, + 553, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 178, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 316, + 178, + 553, + 199 + ], + "type": "text", + "content": "[17] Thomas N Kipf and Max Welling. Variational graph autoencoders. arXiv preprint arXiv:1611.07308, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 202, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 202, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 202, + 553, + 255 + ], + "type": "text", + "content": "[18] Ana Lawry Aguila, James Chapman, and Andre Altmann. Multi-modal variational autoencoders for normative modelling across multiple imaging modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 425-434. Springer, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 258, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 258, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 553, + 300 + ], + "type": "text", + "content": "[19] Haobin Li, Yunfan Li, Mouxing Yang, Peng Hu, Dezhong Peng, and Xi Peng. Incomplete multi-view clustering via prototype-based imputation. arXiv preprint arXiv:2301.11045, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 304, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 346 + ], + "type": "text", + "content": "[20] Xiang Li and Songcan Chen. A concise yet effective model for non-aligned incomplete multi-view and missing multilabel learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5918-5932, 2021. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 403 + ], + "type": "text", + "content": "[21] Chengliang Liu, Gehui Xu, Jie Wen, Yabo Liu, Chao Huang, and Yong Xu. Partial multi-view multi-label classification via semantic invariance learning and prototype modeling. In *Forty-first International Conference on Machine Learning*, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 406, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 406, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 317, + 406, + 553, + 460 + ], + "type": "text", + "content": "[22] Chengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, and Yong Xu. Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In Proceedings of the AAAI conference on artificial intelligence, pages 8807-8815, 2023. 1, 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 463, + 553, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 463, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 317, + 463, + 553, + 517 + ], + "type": "text", + "content": "[23] Chengliang Liu, Jie Wen, Yabo Liu, Chao Huang, Zhihao Wu, Xiaoling Luo, and Yong Xu. Masked two-channel decoupling framework for incomplete multi-view weak multi-label learning. Advances in Neural Information Processing Systems, 36, 2024. 1, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 520, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 520, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 520, + 553, + 552 + ], + "type": "text", + "content": "[24] Zhongchen Ma and Songcan Chen. Expand globally, shrink locally: Discriminant multi-label learning with missing labels. Pattern Recognition, 111:107675, 2021. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 555, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 553, + 586 + ], + "type": "text", + "content": "[25] William McGill. Multivariate information transmission. Transactions of the IRE Professional Group on Information Theory, 4(4):93-111, 1954. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 589, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 589, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 589, + 553, + 632 + ], + "type": "text", + "content": "[26] Yuge Shi, Brooks Paige, Philip Torr, et al. Variational mixture-of-experts autoencoders for multi-modal deep generative models. Advances in neural information processing systems, 32, 2019. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "text", + "content": "[27] Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zili Zhang. Incomplete multi-view weak-label learning. In Ijcai, pages 2703-2709, 2018. 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[28] Chang Tang, Zhenglai Li, Jun Wang, Xinwang Liu, Wei Zhang, and En Zhu. Unified one-step multi-view spectral clustering. IEEE Transactions on Knowledge and Data Engineering, 35(6):6449-6460, 2022. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30730" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 666 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[29] Luis Von Ahn and Laura Dabbish. Labeling images with a computer game. In Proceedings of the SIGCHI conference on Human factors in computing systems, pages 319-326, 2004. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[30] Xu Wang, Dezhong Peng, Peng Hu, and Yongsheng Sang. Adversarial correlated autoencoder for unsupervised multiview representation learning. Knowledge-Based Systems, 168:109-120, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[31] Jie Wen, Chengliang Liu, Shijie Deng, Yicheng Liu, Lunke Fei, Ke Yan, and Yong Xu. Deep double incomplete multiview multi-label learning with incomplete labels and missing views. IEEE transactions on neural networks and learning systems, 2023. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "text", + "content": "[32] Mike Wu and Noah Goodman. Multimodal generative models for scalable weakly-supervised learning. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "text", + "content": "[33] Xuan Wu, Qing-Guo Chen, Yao Hu, Dengbao Wang, Xiaodong Chang, Xiaobo Wang, and Min-Ling Zhang. Multiview multi-label learning with view-specific information extraction. In IJCAI, pages 3884-3890, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 329 + ], + "type": "text", + "content": "[34] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "text", + "content": "[35] Xiaoqiang Yan, Shizhe Hu, Yiqiao Mao, Yangdong Ye, and Hui Yu. Deep multi-view learning methods: A review. Neurocomputing, 448:106-129, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 397 + ], + "type": "text", + "content": "[36] Jun Yin and Shiliang Sun. Incomplete multi-view clustering with reconstructed views. IEEE Transactions on Knowledge and Data Engineering, 35(3):2671-2682, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "type": "text", + "content": "[37] Jun Yin and Wentao Zhang. Multi-view multi-label learning with double orders manifold preserving. Applied Intelligence, 53(12):14703-14716, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "text", + "content": "[38] Jun Yin, Shiliang Sun, Lai Wei, and Pei Wang. Discriminatively fuzzy multi-view k-means clustering with local structure preserving. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16478-16485, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 478, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 531 + ], + "type": "text", + "content": "[39] Honglin Yuan, Shiyun Lai, Xingfeng Li, Jian Dai, Yuan Sun, and Zhenwen Ren. Robust prototype completion for incomplete multi-view clustering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10402-10411, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "type": "text", + "content": "[40] Dawei Zhao, Qingwei Gao, Yixiang Lu, Dong Sun, and Yusheng Cheng. Consistency and diversity neural network multi-view multi-label learning. Knowledge-Based Systems, 218:106841, 2021. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 578, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 620 + ], + "type": "text", + "content": "[41] Dawei Zhao, Qingwei Gao, Yixiang Lu, Dong Sun, and Yusheng Cheng. Consistency and diversity neural network multi-view multi-label learning. Knowledge-Based Systems, 218:106841, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 623, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 294, + 666 + ], + "type": "text", + "content": "[42] Yue Zhu, James T Kwok, and Zhi-Hua Zhou. Multi-label learning with global and local label correlation. IEEE Transactions on Knowledge and Data Engineering, 30(6):1081-1094, 2017. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30731" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/58ab88d4-b760-40c9-8213-6c50aa676cf3_content_list.json b/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/58ab88d4-b760-40c9-8213-6c50aa676cf3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..54d0506cb4ecaac133794ff9dc9f34663a11d864 --- /dev/null +++ b/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/58ab88d4-b760-40c9-8213-6c50aa676cf3_content_list.json @@ -0,0 +1,1733 @@ +[ + { + "type": "text", + "text": "Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model", + "text_level": 1, + "bbox": [ + 233, + 130, + 763, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zheyu Zhang $^{1,2}$ Yayuan Lu $^{1,2}$ Feipeng Ma $^{1,2}$ Yueyi Zhang $^{1,2}$ Huanjing Yue $^{3}$ Xiaoyan Sun $^{1,2,*}$ $^{1}$ MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, 230026, China \n $^{2}$ Anhui Province Key Laboratory of Biomedical Imaging and Intelligent Processing, Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, Hefei 230088, China \n $^{3}$ School of Electrical and Information Engineering, Tianjin University, Tianjin, 300072, China \nzhangzy0@mail.ustc.edu.cn sunxiaoyan@ustc.edu.cn", + "bbox": [ + 104, + 200, + 911, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 361, + 326, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Brain tumor segmentation plays a crucial role in clinical diagnosis, yet the frequent unavailability of certain MRI modalities poses a significant challenge. In this paper, we introduce the Learnable Sorting State Space Model (LS3M), a novel framework designed to maximize the utilization of available modalities for brain tumor segmentation. LS3M excels at efficiently modeling long-range dependencies based on the Mamba design, while incorporating differentiable permutation matrices that reorder input sequences based on modality-specific characteristics. This dynamic reordering ensures that critical spatial inductive biases and long-range semantic correlations inherent in 3D brain MRI are preserved, which is crucial for incomplete multi-modal brain tumor segmentation. Once the input sequences are reordered using the generated permutation matrix, the Series State Space Model (S3M) block models the relationships between them, capturing both local and long-range dependencies. This enables effective representation of intra-modal and inter-modal relationships, significantly improving segmentation accuracy. Extensive experiments on the BraTS2018 and BraTS2020 datasets demonstrate that LS3M outperforms existing methods, offering a robust solution for brain tumor segmentation, particularly in scenarios with missing modalities.", + "bbox": [ + 89, + 393, + 485, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 787, + 220, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate brain tumor segmentation in clinical settings relies on the integration of various MRI modalities (FLAIR, T1ce, T1, T2) that provide complementary information for diagnosis. However, the frequent unavailability of cer-", + "bbox": [ + 89, + 814, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tain modalities, like T1ce due to contrast agent allergies, presents a challenge. Existing methods addressing missing modalities fall into two categories: (1) bridging the data gap between complete and missing modalities through modality synthesis [26, 27, 32, 51, 54, 62] and knowledge distillation [2, 39, 47, 48] and (2) optimizing architectures to maximize information extraction from available modalities, which is the focus of our work. Some methods [8, 63] adopt multi-scale convolutions to capture critical structural details by focusing on spatial inductive bias, such as local texture. However, the limited receptive field restricts their ability to represent tumor tissue effectively, especially when other areas may appear visually similar in localized views, making it difficult to differentiate tumor tissue [35, 40]. Some methods explore models such as Transformers [55] and graph neural networks [51, 59] to build long-range semantic correlation to mitigate the mentioned problem, but they suffer from quadratic computational complexity. The computational inefficiency is further exacerbated by the large sequence lengths typical of 3D brain MRI.", + "bbox": [ + 511, + 362, + 906, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in selective state space models (SSMs), such as Mamba [7, 13], offer the potential to obtain long-range representations for available modalities with dynamic weighting parameters in an efficient way. Mamba flattens the 3D brain MRI into a 1D sequence and iteratively scans this sequence to compress hidden states, which are then used to enhance the current token by combining it with previously scanned tokens. While Mamba can efficiently build long-range dependencies, the influence of distant tokens on the current token diminishes significantly. This becomes problematic for 3D brain MRI, as flattening the sequence destroys the 3D spatial structure. Follow-up methods [33, 50, 58, 65] propose bi-directional, four-directional, and omni-directional scanning to mitigate distortion to spatial structures, as shown in Fig. 1. However, these methods", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25982", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "either neglect local inductive bias or fail to capture long-range semantic correlations of 3D brain MRI, hindering their ability to fully capture contextual relationships essential for brain tumor segmentation, particularly in scenarios with missing modalities.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose the Learnable Sorting State Space Model (LS3M), specifically designed for brain tumor segmentation with missing modalities. Firstly, LS3M incorporates the Sinkhorn operator with the Gumbel-Softmax reparameterization, which converts the traditional non-differentiable sorting into a differentiable permutation matrix. This matrix is used to reorder the 1D sequence of available modalities, considering spatial inductive biases and long-range semantic correlations. Importantly, the permutation matrix does not directly establish relationships; rather, it prepares the sequence for further processing by the Series State Space Model (S3M), which is capable of modeling intra-modal and inter-modal relationships by leveraging both local and long-range correlations. Next, LS3M scales the intra-modal permutation matrix at larger resolutions, allowing for the extraction of multi-scale information. These modality-specific permutation matrices are also combined for multi-modal fusion to capture correlations across modalities. The reordered sequences, now enriched with spatial and semantic information, are then processed by the S3M block, which builds relationships at both intra-modal and inter-modal levels, maximizing the extraction of information from the available modalities. Lastly, LS3M departs from existing models that rely on cropped input volumes by adopting a global input strategy, supplemented with a relative position mask. This approach not only provides richer contextual information but also informs the model about the location of the cropped volume within the larger global context, significantly enhancing spatial awareness and diagnostic accuracy. In summary, our key contributions are as follows:", + "bbox": [ + 91, + 171, + 483, + 638 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose LS3M for brain tumor segmentation with missing modalities. LS3M improves SSM by dynamically reordering data, capturing patient-specific characteristics, and enhancing the ability to extract local and long-range correlations from available modalities.", + "- LS3M scales and integrates the permutation matrix to capture multi-scale and multi-modal information. The reordered sequences are then processed by the S3M block to model complex intra-modal and inter-modal relationships, improving segmentation accuracy.", + "- LS3M incorporates a global input strategy with relative position embeddings, significantly enhancing contextual awareness towards comprehensive utilization of available modalities.", + "- Extensive evaluations on BraTS2018 and BraTS2020 demonstrate the superiority and effectiveness of LS3M for brain tumor segmentation, particularly in clinically" + ], + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/226b16430542e9eda8c561eef97e71061ea237dd5a3199b0540dc10da8b1fc52.jpg", + "image_caption": [ + "Long-range Semantic Correlation", + "(a) Requirements for identifying brain tumor" + ], + "image_footnote": [], + "bbox": [ + 539, + 89, + 622, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c5c93f5ad2c5433c5c6b438dd2c15828fc95d70ef809b7cd509095b79163af68.jpg", + "image_caption": [ + "(b) Horizontal Scanning" + ], + "image_footnote": [], + "bbox": [ + 630, + 92, + 710, + 125 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0b5e33c4b176ad89c13e443450f34cab9c2344e525547128ae9867a0bfbd5fc0.jpg", + "image_caption": [ + "(d) Local Scanning", + "Figure 1. (a) Brain tumor segmentation requires local inductive bias to capture small but critical structural details, as well as long-range semantic correlations to understand spatial relationships between the tumor and surrounding structures. (b)-(e) Different scanning strategies." + ], + "image_footnote": [], + "bbox": [ + 632, + 135, + 710, + 167 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/56b3633d1f10860e505a4efae85e5e7c96f13c06306be2496aeb275343e432c3.jpg", + "image_caption": [ + "(c) Vertical Scanning" + ], + "image_footnote": [], + "bbox": [ + 722, + 92, + 802, + 125 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/57487da9e89fdbb6754875089c890ead9ca48d169035c5e7e1ff80fc17f5fd9a.jpg", + "image_caption": [ + "(e) Our Sorted Scanning" + ], + "image_footnote": [], + "bbox": [ + 725, + 135, + 825, + 167 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/789d2c2627058b7a85e096435041360b9d66210021dc5559f3556af0cb939e8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 813, + 92, + 883, + 125 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8fc4edbd42798f53e0efb5f24a8a81e4d12f3014148c6a611cc750d3b73ea556.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 852, + 126, + 883, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "relevant scenarios with missing modalities.", + "bbox": [ + 527, + 289, + 812, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 319, + 653, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Medical Image Segmentation. Medical image segmentation traditionally relies on U-Net-like structures, which formulate the task as dense pixel classification. Enhancements to segmentation structures include cascaded networks [24, 61], refined skip connections [64], hierarchical pyramids [9], attention mechanisms [4, 6, 12, 16, 17, 29, 42, 46, 53], and large depth-wise kernels [28]. Despite achieving notable performance, these methods often fail when certain modalities are absent.", + "bbox": [ + 511, + 344, + 906, + 481 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-modal Learning for Missing Modalities. Research in multi-modal learning with missing modalities generally seeks to either bridge the gap between scenarios with complete and missing modalities or to optimize architectural design to maximize the use of available modal information. The former approach includes modality synthesis [26, 27, 32, 51, 54, 62] to generate absent modalities and knowledge distillation techniques [2, 30, 39, 47, 48] that transfer insights from complete to incomplete scenarios. The latter approach utilizes techniques like modality disentanglement [5, 52] to separate modality-specific information in the frequency domain and modality fusion strategies involving hierarchical pyramids [8, 63] and attention mechanisms [25, 55-57]. However, combining attention mechanisms with other techniques often results in significant computational overhead, which can impede effective modeling long-range dependencies. In our work, we leverage SSM, specifically the Mamba, to efficiently build long-range correlations from all available modalities.", + "bbox": [ + 511, + 501, + 906, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State Space Model. Drawing inspiration from control theory, the State Space Model (SSM) recurrently updates sequences through latent states [14, 43], as demonstrated by the Structured State Space Sequence model (S4), which utilizes linear state space equations to efficiently scale with sequence length. Recently, Gu and Dao [13] propose a", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25983", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generic modal, called Mamba, with selective projection, gated mechanism and efficient hardware design, outperforming Transformers in linear-time inference on natural language processing. This success has spurred adaptations of Mamba in computer vision. Since visual tokens lack a sequential arrangement, methods propose diverse scanning strategies like bi-/four-/omni-directional [33, 50, 58, 65], local [22], and atrous scanning [38], enhancing contextual data integration for tasks such as medical image segmentation [34, 41, 41], image restoration [15, 19, 60], point cloud analysis [31], image generation [11, 49], and graph neural networks [3, 44].", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 282, + 230, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "State Space Model (SSM) for Sequence Modeling. The SSM defines a linear Ordinary Differential Equation (ODE) to map an input sequence $x(t) \\in \\mathbb{R}$ to an output sequence $y(t) \\in \\mathbb{R}$ through a hidden state $h(t) \\in \\mathbb{R}^N$ , formulated as:", + "bbox": [ + 89, + 309, + 483, + 369 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), \\quad y (t) = \\mathbf {C} h (t), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 375, + 480, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{A} \\in \\mathbb{R}^{N \\times N}$ , $\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}$ , and $\\mathbf{C} \\in \\mathbb{R}^{1 \\times N}$ are learned parameters. Since images are processed into discrete token sequences, the discrete version of Eq. (1) is defined as follows:", + "bbox": [ + 89, + 397, + 483, + 457 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {k} = \\overline {{\\mathbf {A}}} h _ {k - 1} + \\overline {{\\mathbf {B}}} x _ {k}, \\quad y _ {k} = \\mathbf {C} h _ {k}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 464, + 480, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\overline{\\mathbf{A}}$ and $\\overline{\\mathbf{B}}$ are discretized using the discretization step size $\\Delta$ in the zero-order hold (ZOH) rule. The long-range dependencies are constructed via recurrent scanning. However, the recurrent SSM is not practical. Eq. (2) is equivalent to a global convolution kernel $\\overline{\\mathbf{K}}\\in \\mathbb{R}^{L}$ , i.e.,", + "bbox": [ + 89, + 488, + 483, + 566 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {K}}} = \\left(\\mathbf {C} \\overline {{\\mathbf {B}}}, \\mathbf {C} \\overline {{\\mathbf {A B}}}, \\dots , \\mathbf {C} \\overline {{\\mathbf {A}}} ^ {\\mathbf {L} - 1} \\overline {{\\mathbf {B}}}\\right), \\quad \\mathbf {y} = \\mathbf {x} * \\overline {{\\mathbf {K}}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 570, + 480, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Structured state space sequence models (S4) imposes a structure on the state matrix $\\mathbf{A}$ based on HIPPO matrices, which significantly improves the efficiency for obtaining the SSM kernel $\\overline{\\mathbf{K}}$ . Further, Mamba [13] adds selective projection via making $\\mathbf{B}, \\mathbf{C}, \\Delta$ input-dependent.", + "bbox": [ + 89, + 599, + 483, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sinkhorn Operator. The Sinkhorn operator $S(\\cdot)$ is an iterative procedure used to transform a non-negative matrix $\\mathbf{P} \\in \\mathbb{R}^{L \\times L}$ into a doubly stochastic matrix, which is a square matrix with non-negative entries such that each row and each column sums to one [1, 36]. The process can be formulated as:", + "bbox": [ + 89, + 691, + 483, + 782 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l}S ^ {0} (\\mathbf {P}) = \\exp (\\mathbf {P}), \\quad S ^ {n} (\\mathbf {P}) = \\mathcal {T} _ {c} \\left(\\mathcal {T} _ {r} \\left(S ^ {n - 1} (\\mathbf {P})\\right)\\right),\\\\S (\\mathbf {P}) = \\lim _ {n \\rightarrow \\infty} S ^ {n} (\\mathbf {P}),\\end{array}\\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 785, + 482, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $n$ is iterative times. $\\mathcal{T}_r(\\cdot)$ and $\\mathcal{T}_c(\\cdot)$ are the row-wise and column-wise normalizations, defined as follows:", + "bbox": [ + 89, + 832, + 482, + 863 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {r} (\\mathbf {P}) = \\mathbf {P} \\oslash \\left(\\mathbf {P} \\mathbf {1} _ {\\mathbf {N}} \\mathbf {1} _ {\\mathbf {N}} ^ {\\top}\\right), \\mathcal {T} _ {c} (\\mathbf {P}) = \\mathbf {P} \\oslash \\left(\\mathbf {1} _ {\\mathbf {N}} \\mathbf {1} _ {\\mathbf {N}} ^ {\\top} \\mathbf {P}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 868, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\oslash$ denotes the element-wise division and $1_{N}$ is a column vector of ones. We employ Sinkhorn $(\\mathbf{P},n)$ to represent Eq. (4).", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gumbel-Softmax Reparameterization. The Gumbel-Softmax provides a differentiable approximation of the discrete categorical distribution [23]. The process of sample $z$ from a categorical distribution with class probabilities $x$ can be defined as follows:", + "bbox": [ + 511, + 154, + 906, + 229 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz = \\operatorname {o n e - h o t} (\\underset {i} {\\arg \\max } [ g _ {i} + \\log x _ {i} ]), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 234, + 906, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $g_{i}$ is the injected standard i.i.d Gumbel noise. Jang et al. [23] use the softmax function as a continuous, differentiable approximation to the arg max function, i.e.,", + "bbox": [ + 511, + 267, + 906, + 313 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s o f t m a x} (y / \\tau) \\equiv \\frac {\\exp \\left(y _ {i} / \\tau\\right)}{\\sum_ {i = 1} ^ {k} \\exp \\left(y _ {i} / \\tau\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 320, + 905, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y_{i} = \\log x_{i} + g_{i}$ . When temperature $\\tau \\rightarrow 0$ , the softmax distribution converges to a one-hot representation identical to samples from a categorical distribution. We use Gumbel-Softmax(X, $\\tau$ ) to represent Eq. (7), where $\\mathbf{X} = [x_0, x_1, \\dots, x_{|\\mathbf{X}| - 1}]$ .", + "bbox": [ + 511, + 364, + 905, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 452, + 604, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first explain the motivation and challenges with Mamba. We then introduce the learnable sorting process to address these challenges, followed by expanding the permutation matrix for longer and multi-modal sequences. Next, we describe the S3M block and the global input strategy. Finally, we present the LS3M architecture and its optimization loss.", + "bbox": [ + 511, + 477, + 905, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Motivations", + "text_level": 1, + "bbox": [ + 513, + 590, + 640, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Brain tumor segmentation requires complementary information from multiple modalities. Missing modalities lead to insufficient information, inspiring us to maximize information extraction from available modalities. We try to do this from two aspects: building long-range dependencies and incorporating broader context for available modalities.", + "bbox": [ + 511, + 613, + 905, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, brain tumors often exhibit complex anatomy, where local inductive biases alone may be insufficient to identify tumor types and boundaries. Long-range dependencies are essential for understanding spatial relationships between the tumor and surrounding structures. While Transformers can model such dependencies, their quadratic computational complexity makes them impractical for 3D MRI data due to large token sequences and high computational costs, as illustrated in the Appendix. Mamba, however, captures long-range dependencies more efficiently, making it a better fit for this task.", + "bbox": [ + 511, + 704, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Second, most segmentation methods focus on cropped volumes to reduce computational cost, neglecting the", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25984", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a4fd3450080f08915800b6a0a20969c752829cb4af17862d5781c2d2c9531aab.jpg", + "image_caption": [ + "(a) Our LS3M framework" + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 591, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f1232b9edc89770ad4783313e3bb5aa6365e714acce722389cc2bfb575da4a05.jpg", + "image_caption": [ + "(b) $\\mathrm{SortP_M}$", + "(c) $\\mathrm{SortP_L}$", + "(d) $\\mathrm{SortP_{Mul}}$" + ], + "image_footnote": [], + "bbox": [ + 598, + 90, + 895, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5313ba132fb3d5d77cfeb1dd084eb9ff3473d6972524f05dfe8740dc4d12ffe9.jpg", + "image_caption": [ + "Figure 2. (a) The comprehensive LS3M framework integrates the global input strategy with adaptive pooling, the SortP for generating permutation matrices, and the S3M for feature enhancement. $(\\cdot)_M$ denotes that we perform $(\\cdot)$ process separately for each modality. (b) SortP $M$ means the learnable sorting process applied to each modality. (c) SortP $L$ denotes obtaining the larger permutation matrix for longer sequences. (d) SortP $Mul$ refers to the generation of the multi-modal permutation matrix.", + "Figure 3. The structure of the proposed S3M block." + ], + "image_footnote": [], + "bbox": [ + 96, + 352, + 475, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "broader context crucial for accurately understanding complex conditions like brain tumors. This is especially important in the case of intra-tumor heterogeneity, which requires a wider context to characterize the tumor type [35, 40]. Moreover, in scenarios with missing modalities, it is critical to utilize all available data from the entire image to compensate for the missing information. Thus, incorporating broader context without significant computational overhead is essential.", + "bbox": [ + 89, + 474, + 483, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Challenges of Existing Mamba Approaches", + "text_level": 1, + "bbox": [ + 89, + 619, + 457, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mamba is primarily designed for 1D sequences, with characteristics that include:", + "bbox": [ + 89, + 642, + 482, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mamba Characteristic 1: the current output token $y_{k}$ is limited to modeling relationships solely with previously scanned tokens $x_{i}, i \\in \\{0,1,\\dots ,k - 1,k\\}$ , as detailed in Eq. (2).", + "bbox": [ + 89, + 672, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mamba Characteristic 2: the influence of distant tokens diminishes significantly due to the weight $\\overline{\\mathbf{A}}^{i - 1}$ , as explained in Eq. (3).", + "bbox": [ + 89, + 734, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate these limitations, existing Mamba-based methods [33, 38, 50, 65] introduce forward and backward scanning strategies, which enable the current token to model relationships with both preceding and subsequent tokens. These methods are more suitable for vision applications, where local dependencies can often provide sufficient context. However, in the case of missing modalities, it is insufficient to rely solely on local context for tumor identi", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fication. Tumor heterogeneity often requires capturing information from distant regions of the MRI to fully understand the tumor's characteristics, as highlighted by prior studies [35, 40]. For example, in the segmentation of the whole tumor (WT) with only T1ce modality, local features may be insufficient, and distant correlations are necessary to improve segmentation accuracy.", + "bbox": [ + 511, + 352, + 903, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to Mamba Characteristic 2, the influence of distant tokens diminishes over sequence steps, which is problematic in 3D MRI segmentation tasks. Therefore, a dynamic mechanism is required to ensure that the most relevant tokens—both spatially close and semantically related over long distances—are positioned closer to the target token, facilitating more effective modeling.", + "bbox": [ + 511, + 458, + 905, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although LocalMamba [22] uses a Z-scan to ensure local relationships in 2D, this approach cannot guarantee the preservation of 3D neighborhood relationships, such as the 27-neighborhood in 3D space. Even with axial scanning added to horizontal and vertical scanning, the relationships between distant tokens in 3D space are still not adequately preserved. Consequently, there is a need for a solution that dynamically reorders the tokens for each 3D MRI scan, ensuring that all potentially relevant tokens—whether spatially close or far but semantically related—are placed before the current token. This dynamic reordering will allow the subsequent S3M block to model these relationships more effectively, improving the segmentation accuracy.", + "bbox": [ + 511, + 565, + 906, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Learnable Sorting Process", + "text_level": 1, + "bbox": [ + 511, + 772, + 751, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Traditional Non-differentiable Sorting Process. A pivotal aspect of LS3M involves addressing the non-differentiability of traditional sorting mechanisms. The patient-specific dynamic sorting process is initially driven by a predictive score $P \\in \\mathbb{R}^{L}$ , assigned to each token in the sequence $\\mathbf{X}$ . These scores dictate the token positions in the sorted sequence $\\mathbf{X}_{\\mathbf{S}}$ , subsequently processed by the", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25985", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mamba block. However, the traditional sorting function, which maps these continuous input scores to discrete token indices, interrupts the gradient flow necessary for end-to-end training of the network.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Learnable Sorting Process. The learnable sorting process is represented as:", + "bbox": [ + 89, + 169, + 482, + 200 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {P}} = \\operatorname {S o r t N e t} (\\mathbf {f}),\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 210, + 256, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} = \\text {S i n k h o r n} (\\text {G u m b e l - S o f t m a x} (\\hat {\\mathbf {P}}, \\tau), n), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 233, + 482, + 250 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} = \\text {f l a t t e n} (\\mathbf {f}), \\quad \\mathbf {X} _ {\\mathbf {S}} = \\mathbf {P X}.\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 253, + 372, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This enable end-to-end differentiable sorting. As shown in Fig. 2(b), firstly, our SortNet generates an estimated matrix $\\hat{\\mathbf{P}}\\in \\mathbb{R}^{L\\times L}$ tailored to individual patients. It employs a convolutional layer with a kernel size of $C\\times C\\times 3\\times 3\\times 3$ to capture spatial inductive biases, extracting information from the 26-neighborhood 3D spatial structure. After convolution, a SiLU activation function [20] is applied to enhance the network's representational capacity. The resulting feature map is split into $\\mathbf{f}_1$ and $\\mathbf{f}_2$ , with $\\mathbf{f}_2$ being processed to compute a similarity map that captures long-range semantic correlations. A ReLU activation is then applied to ensure the non-negativity of $\\hat{\\mathbf{P}}$ , preparing it for subsequent processing by the Sinkhorn operator.", + "bbox": [ + 89, + 280, + 483, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Secondly, we use the Gumbel-Softmax and Sinkhorn operators to generate a permutation matrix $\\mathbf{P}$ , where each row and column contains exactly one 1 with all other entries being 0s. The estimated matrix $\\hat{\\mathbf{P}}$ is sliced row-wise, with each row being passed through the Gumbel-Softmax function. By adjusting the temperature parameter $\\tau$ , the output satisfies the one-hot condition for each row. The resulting rows are then concatenated and passed through the Sinkhorn operator, which iteratively adjusts the rows and columns of the matrix, yielding the final permutation matrix $\\mathbf{P}$ .", + "bbox": [ + 89, + 477, + 483, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the feature map $\\mathbf{f}$ is flattened into a sequence $\\mathbf{X}$ , and $\\mathbf{P}$ is applied to $\\mathbf{X}$ to perform the sorting operation. This process can be viewed as permuting the rows of an $L\\times L$ identity matrix according to the sorting procedure. Since all of these operations are differentiable, the entire sorting process is also differentiable, which ensures that each token is optimally positioned in the sequence for subsequent processing by our S3M block.", + "bbox": [ + 89, + 628, + 483, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Handling Longer and Multi-modal Sequences", + "text_level": 1, + "bbox": [ + 89, + 757, + 478, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Handling Longer Sequence. Extracting multi-scale information from available modalities requires generating larger permutation matrices to accommodate longer sequences corresponding to high-resolution feature maps. However, this process introduces quadratic computational complexity due to the matrix multiplications involved. To address this, we propose an efficient alternative: bilinear interpolation to scale the permutation matrix, eliminating", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the need for SortNet. The scaled matrix is then processed through ReLU activation, Gumbel-Softmax, and Sinkhorn to generate the permutation matrix $\\mathbf{P}_{\\mathrm{Large}}$ for the extended sequence, as represented in Fig. 2(c).", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Handling Multi-modal Sequences. To effectively leverage inter-modal correlations, we aim to generate a permutation matrix that accounts for scenarios with missing modalities. We dynamically fuse the individual permutation matrices of each modality— $\\mathbf{P}_{\\mathrm{FLAIR}}$ , $\\mathbf{P}_{\\mathrm{T1ce}}$ , $\\mathbf{P}_{\\mathrm{T1}}$ , $\\mathbf{P}_{\\mathrm{T2}}$ with a multi-modal code, represented by $h = \\{h_1, h_2, h_3, h_4\\}$ , where $h_i \\in \\{0, 1\\}$ indicates whether the corresponding modality is available. This fusion produces a comprehensive multi-modal estimated matrix, $\\hat{\\mathbf{P}}_{\\mathrm{Mul}}$ . We then apply a ReLU activation followed by Gumbel-Softmax and Sinkhorn operations to generate the final multi-modal permutation matrix $\\mathbf{P}_{\\mathrm{Mul}}$ , as shown in Fig. 2(d). Before reordering the multi-modal sequence, each modal sequence is initially projected into a unified representation space. The sequences are then sorted according to $\\mathbf{P}_{\\mathrm{Mul}}$ , ensuring alignment based on inter-modal relevance and coherence.", + "bbox": [ + 511, + 189, + 906, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Series State Space Model (S3M) Block", + "text_level": 1, + "bbox": [ + 511, + 450, + 841, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the insufficient information of missing modalities, it becomes crucial to exhaustively mine information from all available modalities. We introduce the S3M model for local and long-range dependency modeling, inspired by the success of Mamba.", + "bbox": [ + 511, + 476, + 905, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3, our S3M block connects two unidirectional Mamba-like blocks, $F_{Fwd}$ and $F_{Bwd}$ , in a series connection but in opposite directions, together with a Channel Attention (FCA) block [21], which can be expressed as follows:", + "bbox": [ + 511, + 556, + 905, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\mathbf {S 3 M}} = \\mathrm {F} _ {\\mathrm {C A}} \\left(\\mathrm {F} _ {\\mathrm {B w d}} \\left(\\mathrm {F} _ {\\mathrm {F w d}} \\left(\\mathbf {f} _ {\\mathbf {I n}}, \\mathbf {P}\\right)\\right)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 650, + 903, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Before processing, the feature $\\mathbf{f}_{\\mathrm{In}}$ is reshaped into a sequence and sorted by multiplying it with a soft permutation matrix $\\mathbf{P}$ generated through the learnable sorting process.", + "bbox": [ + 511, + 685, + 905, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each Mamba-like block comprises three branches: a basic SSM branch, a gated weighting branch, and a residual connection branch. In the SSM branch, the input feature $\\mathbf{f}_{\\mathrm{In}}$ undergoes a linear projection to expand its channels, followed by depth-wise convolution and SiLU activation, producing $\\mathbf{f}_{\\mathrm{Mid}}$ and the SSM parameters $\\{\\overline{\\mathbf{A}},\\overline{\\mathbf{B}},\\mathbf{C}\\}$ . They are sent to the Forward/Backward SSM module, followed by layer normalization to obtain $\\mathbf{f}_{\\mathrm{SSM}}$ . The input feature $\\mathbf{f}_{\\mathrm{In}}$ is processed through a linear layer and SiLU activation to generate a gated weight feature map $\\mathbf{f}_{\\mathrm{Gate}}$ . Combining these with the residual connection branch, the output of the", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25986", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mamba-like block is:", + "bbox": [ + 91, + 90, + 235, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {f} _ {\\mathbf {S S M}} = \\mathrm {L N} (\\mathrm {S S M} (\\mathrm {S i L U} (\\mathrm {D W - C o n v} (\\mathrm {L i n e a r} (\\mathrm {L N} (\\mathbf {f} _ {\\mathbf {I n}})))))), \\\\ \\mathbf {f} _ {\\text {G a t e}} = \\operatorname {S i L U} \\left(\\operatorname {L i n e a r} \\left(\\mathbf {f} _ {\\text {I n}}\\right)\\right), \\\\ \\mathbf {f} _ {\\text {O u t}} = \\operatorname {L i n e a r} \\left(\\mathbf {f} _ {\\text {S S M}} \\odot \\mathbf {f} _ {\\text {G a t e}}\\right) + \\mathbf {f} _ {\\text {I n}}. \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 123, + 485, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For further channel-wise interactions, we introduce the Channel Attention block. Our S3M block, when compared to the recent Vim block [65] and SS2D block [33], offers two main advantages: (1) The series connection in S3M ensures that information from the forward SSM block is integrated into subsequent tokens, which is then relayed back to earlier tokens in the backward SSM block. This contrasts with parallel connections in Vim block and SS2D block, where token sequences receive the same information level. (2) The integration of the permutation matrix in S3M preserves both spatial proximity and long-range semantic correlations.", + "bbox": [ + 89, + 195, + 483, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For multi-modal fusion, each modal sequence is first projected into a unified representation space and then concatenated along the channel dimension. The multi-modal sequences $\\mathbf{X}_{\\mathrm{Mul}}$ are sorted by $\\mathbf{P}_{\\mathrm{Mul}}$ , which are further delivered into our S3M for building inter-modal correlations. Specifically, the S3M utilizes the Mamba-like block where the information from each token in $\\mathbf{f}_{\\mathrm{Mul}}$ is compressed into SSM parameters— $\\overline{\\mathbf{A}}_{Mul}$ , $\\overline{\\mathbf{B}}_{Mul}$ , $\\mathbf{C}_{Mul}$ , which contain the compact multi-modal information along their hidden state dimension. Then, the Forward/Backward SSM module interact across both the channel dimensions of the multi-modal sequences and the state dimensions of the SSM parameters, representing the modality interaction between the origin format and the compact format, thereby, resulting in an efficient fusion that markedly improves segmentation.", + "bbox": [ + 89, + 380, + 483, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.6. Global Input Strategy", + "text_level": 1, + "bbox": [ + 89, + 628, + 294, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We introduce a global input strategy that integrates the full volumes information into the analysis of cropped volumes without incurring excessive computational costs, as detailed in Fig. 2. The full volumes, augmented with a relative position mask, undergoes a series of transformations through convolutional layers and non-linear activations. They are then resized to match the dimensions of the cropped volumes using adaptive average pooling. The resulting feature map, which combines coarse global information with the fine details of the cropped volume, is then fed into our LS3M for segmentation.", + "bbox": [ + 89, + 654, + 482, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The relative position mask, specifically designed for the full volumes, is encoded at positions corresponding to the cropped region. This design emphasizes the relevant area within the larger volume, guiding the network's attention to the regions slated for segmentation.", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.7. Network Architecture and Loss", + "text_level": 1, + "bbox": [ + 513, + 90, + 790, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our framework, depicted in Fig. 2, processes inputs of four modalities. Each modality includes both a cropped volume for segmentation and its corresponding global volume. The global volumes are first processed through our global input strategy. The yielded feature maps are then concatenated with their respective cropped volumes and passed through individual encoders $E(\\cdot)_M$ to extract hierarchical intra-modal features.", + "bbox": [ + 511, + 112, + 903, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The features from each modality are transformed into sequences and fed into $\\mathrm{SortP}_M$ , the learnable sorting process, which generates a modality-specific permutation matrix, $\\mathbf{P}_{\\mathbf{M}}$ . This matrix reorders the sequences before they are input into the S3M tailored for each modality.", + "bbox": [ + 511, + 234, + 903, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At higher resolutions, for each modality, a larger permutation matrix, $\\mathbf{P}_{\\mathrm{Large}}$ , is obtained by bilinear upsampling. For multi-modal integration, the comprehensive multi-modal permutation matrix, $\\mathbf{P}_{\\mathrm{Mul}}$ , is generated by combining the modality-specific permutation matrix $\\mathbf{P}_{\\mathrm{M}}$ with a multi-modal code $h$ . These permutation matrices play a crucial role in integrating the sequences within the S3M, ultimately enhancing segmentation accuracy.", + "bbox": [ + 511, + 310, + 903, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The final segmentation results are obtained after hierarchical fusion and processing of the multi-modal features. Simulated missing modalities are represented by substituting zero-value feature maps prior to concatenating individual modal features.", + "bbox": [ + 511, + 431, + 903, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [8, 55], we employ the weighted cross-entropy loss $Loss_{WCE}$ and the Dice loss $Loss_{Dice}$ . The total loss is $Loss = Loss_{WCE} + Loss_{Dice}$ .", + "bbox": [ + 511, + 506, + 903, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 566, + 643, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset and implementation details", + "text_level": 1, + "bbox": [ + 511, + 590, + 820, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform our experiments on BraTS2018 and BraTS2020 from the Multi-modal Brain Tumor Segmentation Challenge [37], which align with that of previous studies [32, 45]. The two datasets contain 285 and 369 cases, respectively, with publicly available ground truth. Consistent with [8], BraTS2020 is randomly split into training, validation, and test sets with a ratio of $219:50:100$ . BraTS2018 is split into $199:29:57$ and includes a three-fold validation. These splits are consistently used across all experiments, ensuring that all methods are evaluated under the same conditions.", + "bbox": [ + 511, + 613, + 905, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Each case of the datasets has four different modalities, i.e., FLAIR, T1ce, T1, and T2 modalities. These modalities are characterized by a volume size of $240 \\times 240 \\times 155$ , and they capture various properties of brain tumor subregions: GD-enhancing tumor (ET), peritumoral edema (ED), and the necrotic and non-enhancing tumor core (NCR/NET). These subregions of brain tumors are grouped into three nested subregions: the whole tumor (WT), the tumor core", + "bbox": [ + 511, + 780, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25987", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b82111d5aebc7e81ac1b261bd70d1874763f1bbce24cedbd0c0a6e84d7623d57.jpg", + "image_caption": [ + "Figure 4. Segmentation results of different methods with various available modalities on BraTS2020." + ], + "image_footnote": [], + "bbox": [ + 158, + 99, + 834, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(TC), and the enhancing tumor (ET). We use Dice coefficient, $95\\%$ Hausdorff distance, sensitivity, and specificity to evaluate our method.", + "bbox": [ + 88, + 369, + 482, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct our experiments using the PyTorch framework version 1.13.0. The training process is carried out on four NVIDIA A800 GPUs, spanning 500 epochs, with a batch size of four volumes. We utilize the Adam optimizer with an initial learning rate set to $2 \\times 10^{-4}$ and a weight decay of $1 \\times 10^{-5}$ . During the training phase, the input images are randomly cropped to dimensions of $80 \\times 80 \\times 80$ . Data augmentation techniques include random rotations, flips, and intensity adjustments.", + "bbox": [ + 88, + 414, + 483, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparisons with the State-of-the-art", + "text_level": 1, + "bbox": [ + 89, + 560, + 419, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our method, we compare our method with nine state-of-the-art methods on different cases with missing modalities. The involved methods contain HeMiS [18], U-HVED [10], RFNet [8], UNet-MFI [59], mmFormer [55], MMANet [48], ShaSpec [45], M3AE [32], and GSS [39]. For a fair comparison, all methods are trained under their recommended hyper-parameters within the same dataset split.", + "bbox": [ + 88, + 582, + 482, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Tab. 1, our method achieves better performance for most combinations of missing modalities. We achieve improvements of $1.22\\%$ , $1.31\\%$ , and $2.3\\%$ over the second-ranked method on the average DSC for WT, TC, and ET. We also provide a visualization comparison in Fig. 4, illustrating that our method yields more accurate segmentation results in different combinations of modalities. More comparisons, including statistical results, additional metrics and model complexity, are provided in the Appendix.", + "bbox": [ + 88, + 703, + 482, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 848, + 243, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments on BraTS2020 to assess the effectiveness of specific designs. The baseline is served", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f4a497b3db952e301a3611662629cffe949df24fc7bf7e324ee03d792af0b11a.jpg", + "image_caption": [ + "Figure 5. Effect of hyper-parameters $\\tau$ and $n$ ." + ], + "image_footnote": [], + "bbox": [ + 545, + 369, + 870, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as a simplified version of LS3M without GIS, SortP, and S3M components.", + "bbox": [ + 511, + 564, + 906, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Scanning Strategies. Relative to our baseline, which employs no specific scanning strategies, adding bidirectional, and fourth-directional (cross scan) methods demonstrated enhanced effectiveness, attributed to their ability to maintain directional spatial coherence. Notably, our sorted scanning strategy, as detailed in Tab. 2, achieved the highest Dice scores by preserving spatial inductive biases and semantic correlations.", + "bbox": [ + 511, + 619, + 906, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Proposed Components. As shown in Tab. 3, integrating a global input significantly improves baseline performance by providing larger spatial contexts. The addition of a relative position mask further enhances results by focusing on key areas within the broader input context. The average performance improves $1.84\\%$ with S3M. The complete LS3M architecture, after including SortP, achieves superior performance, effectively exploring intra-modal and inter-modal relations.", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25988", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e35e2c95cc0556d590dddee5a6cfe3d4a6656c4d7bde8a0fbb91919108fe36fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MFLAIR●○○○●●●○○○●●●○●AVG
T1ce○●○○●○○●○○●●○●●AVG
T1○○●○○○○●○●●○●●●●
T2○○○●○○○●○●○●●●●●
WTHeMiS71.6067.7168.9668.1969.1768.6769.8369.0169.7869.4070.2171.2870.7371.5872.0669.88
U-HVED69.8546.8246.7754.0361.4558.2564.5062.9165.7664.2966.9969.7068.3870.3571.4162.76
RFNet86.4277.3476.4686.2189.5589.3089.3581.0087.4587.9590.3990.2090.4288.5990.7786.76
UNet-MFI82.2773.1872.1082.4583.6484.3484.8577.3083.4483.5285.4585.7085.8584.2084.9382.21
mmFormer82.4074.2574.3783.0784.5484.6185.8277.9884.0584.0085.3486.1186.2284.6486.3882.92
MMANet83.2164.4664.3582.8886.7487.6487.4770.9885.6485.3387.9289.3787.9786.3889.2582.64
ShaSpec84.1070.3670.2383.9588.6589.0788.7076.0286.2187.1589.4789.4390.0787.6290.1184.74
M3AE86.5373.8576.7186.0989.4889.3889.2578.1187.3787.2089.9990.1890.4288.6190.5686.25
GSS*86.3672.4573.3386.7689.5190.6590.9178.3888.8790.0591.1493.2590.6789.4393.2887.00
Ours88.6678.2479.3287.2991.0090.5791.4582.5889.1888.8891.5291.8791.4589.7592.1588.22
TCHeMiS53.4351.4151.5651.1151.7051.0851.8551.8852.3551.5152.9553.7652.9754.3855.0352.46
U-HVED34.6235.5127.3037.6742.1538.2643.4144.9347.5344.9749.1351.3049.4052.7254.1743.53
RFNet65.0482.3764.3168.4784.6971.4572.6283.1584.0672.1184.7184.7074.2884.1184.7477.39
UNet-MFI63.9477.6359.3868.0579.9268.2370.7277.6180.0970.2180.0380.9471.4080.7581.2874.01
mmFormer66.1977.9661.1769.1880.3669.5871.5579.9380.7970.9080.1881.3172.0281.1281.2274.90
MMANet65.7974.0557.4770.2185.2172.7372.6580.9985.2471.8885.8785.6874.4886.3986.6376.99
ShaSpec66.0677.3459.5768.2985.0272.8772.9580.8484.4672.2885.8385.5774.7485.4786.5077.19
M3AE68.0481.3966.0070.2782.0173.8274.9582.3983.0172.5482.4483.0675.0984.0684.4077.56
GSS*67.3378.4361.6770.1084.8873.4574.1782.2086.1274.1886.5088.3974.3386.2888.6778.45
Ours71.9883.6166.6772.3086.3275.2576.8784.0985.1874.3286.0585.8476.8785.8385.8679.76
ETHeMiS43.7742.4141.5941.4541.8340.2941.1942.0842.3941.0043.6744.1642.9545.2746.3342.69
U-HVED12.8824.947.2724.2630.0221.9529.4033.6436.1832.1239.3940.9138.0943.1845.3330.64
RFNet40.4774.2737.5143.5976.4543.8146.9975.2273.9446.3777.0176.3848.9576.3876.6460.93
UNet-MFI39.7069.4229.3846.0070.1340.0648.6969.2572.3245.7171.2870.8846.5572.0071.4157.52
mmFormer40.4768.9133.9745.6169.8143.6348.0971.1070.7245.9270.0871.6048.3870.6571.3658.02
MMANet36.4066.0728.2241.9170.4341.5044.2368.8771.8643.5871.9471.9844.6071.8172.4356.39
ShaSpec38.9266.9331.7542.5870.8543.2245.6269.9271.0444.2172.0871.0946.2271.8172.0257.22
M3AE40.4972.4339.9345.9774.6643.2047.3075.4276.8146.6375.9477.0848.1977.4078.0061.30
GSS*42.0369.4635.3045.6074.2047.7849.3374.4174.7848.9476.1276.3250.0276.9875.9561.15
Ours40.3477.6836.5745.3079.1745.0751.5379.4978.6947.0480.7580.5151.5381.5380.9463.60
", + "bbox": [ + 133, + 88, + 866, + 450 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ffb759749bc774e375d76a0d5d30a1b0de41b02b722dac7454dba5f7e94912bb.jpg", + "table_caption": [ + "Table 1. Performance comparison (DSC%) with SOTA methods on BraTS2020. Available and missing modalities are denoted by $\\bullet$ and $\\circ$ , respectively. \\* demonstrates that its codes and results are reproduced by ourselves." + ], + "table_footnote": [], + "table_body": "
DiceBaseline+ Uni-D+ Bi-D+ Fourth-D+ Random+ Our sorted
WT85.4385.2085.5586.1384.2386.94
TC76.1775.6976.6777.5474.0878.01
ET58.0257.9858.4360.7156.7962.06
", + "bbox": [ + 102, + 512, + 470, + 556 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/74fd5a7d189e124feb288696c564304f197e8bb0b2b74122478edd0bfa90bd7f.jpg", + "table_caption": [ + "Table 2. Effect of scanning strategies, '-D' denotes directional scanning." + ], + "table_footnote": [], + "table_body": "
DiceBaseline+GIS (w/o mask)+GIS+GIS + S3M+GIS + SortP + S3M
WT85.4385.9986.2387.1788.22
TC76.1776.7877.0679.0679.76
ET58.0259.0160.3562.9463.60
", + "bbox": [ + 102, + 609, + 470, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Effect of proposed components, 'GIS' denotes Global Input Strategy.", + "bbox": [ + 89, + 662, + 482, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of hyper-parameters of temperature $\\tau$ and iterative times $n$ . Adjusting $\\tau$ and $n$ reveals their critical roles in influencing the performance of our LS3M, as shown in Fig. 5. The parameter $\\tau$ modulates the smoothness of the permutation matrix, with $\\tau = 1$ yielding optimal results when $n$ is fixed. Conversely, increasing $n$ enhances the precision of the permutation matrix, thus improving segmentation outcomes. Balancing computational efficiency and precision, we select $n = 10$ .", + "bbox": [ + 89, + 718, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional ablation experiments are provided in the appendix. These experiments further validate the importance of the proposed components and their contribution to the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model's performance.", + "bbox": [ + 513, + 515, + 660, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 566, + 633, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce the Learnable Sorting State Space Model (LS3M), a novel framework for brain tumor segmentation with missing modalities, which aims to maximizing mining information from available modalities. LS3M replaces traditional non-differentiable sorting with a differentiable permutation matrix, preserving 3D spatial information and enabling more effective modeling. The framework adeptly scales these matrices to accommodate longer sequences, capturing multi-scale information. It also dynamically integrates these matrices for robust multi-modal sequence fusion, effectively managing various scenarios of missing modalities. The S3M block captures both intra-modal and inter-modal correlations, enhancing segmentation accuracy. With a global input strategy, LS3M provides comprehensive contextual insights, surpassing existing segmentation methods. Experimental results on BraTS2018 and BraTS2020 demonstrate that LS3M outperforms state-of-the-art. Our results highlight LS3M's potential in advancing medical imaging and healthcare diagnostics.", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25989", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was in part supported by the National Natural Science Foundation of China under grants 62032006 and 62021001, and the Anhui Province Key Laboratory of Biomedical Imaging and Intelligent Processing under grants 23YGXT001.", + "bbox": [ + 89, + 114, + 483, + 186 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 215, + 187, + 231 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ryan Prescott Adams and Richard S Zemel. Ranking via sinkhorn propagation. arXiv preprint arXiv:1106.1925, 2011.3", + "[2] Reza Azad, Nika Khosravi, and Dorit Merhof. Smu-net: Style matching u-net for brain tumor segmentation with missing modalities. In International Conference on Medical Imaging with Deep Learning, pages 48-62. PMLR, 2022. 1, 2", + "[3] Ali Behrouz and Farnoosh Hashemi. Graph mamba: Towards learning on graphs with state space models. arXiv preprint arXiv:2402.08678, 2024. 3", + "[4] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In Computer Vision-ECCV 2022 Workshops, pages 205-218. Springer, 2023. 2", + "[5] Cheng Chen, Qi Dou, Yueming Jin, Hao Chen, Jing Qin, and Pheng-Ann Heng. Robust multimodal brain tumor segmentation via feature disentanglement and gated fusion. In Medical Image Computing and Computer-Assisted Intervention-MICCAI, pages 447-456. Springer, 2019. 2", + "[6] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 2", + "[7] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 1", + "[8] Yuhang Ding, Xin Yu, and Yi Yang. Rfnet: Region-aware fusion network for incomplete multi-modal brain tumor segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3975-3984, 2021. 1, 2, 6, 7", + "[9] Yuhang Ding, Liulei Li, Wenguan Wang, and Yi Yang. Clustering propagation for universal medical image segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2024. 2", + "[10] Reuben Dorent, Samuel Joutard, Marc Modat, Sébastien Ourselin, and Tom Vercauteren. Hetero-modal variational encoder-decoder for joint modality completion and segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI, pages 74-82. Springer, 2019. 7", + "[11] Zhengcong Fei, Mingyuan Fan, Changqian Yu, and Junshi Huang. Scalable diffusion models with state space backbone. arXiv preprint arXiv:2402.05608, 2024. 3" + ], + "bbox": [ + 93, + 241, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Jiannan Ge, Lingxi Xie, Hongtao Xie, Pandeng Li, Xiaopeng Zhang, Yongdong Zhang, and Qi Tian. Alignzeg: Mitigating objective misalignment for zero-shot semantic segmentation. In European Conference on Computer Vision, pages 142-161. Springer, 2024. 2", + "[13] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3", + "[14] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In International Conference on Learning Representations, 2021. 2", + "[15] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024.3", + "[16] Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In *Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: 7th International Workshop, BrainLes* 2021, Held in Conjunction with MICCAI 2021, Virtual Event, pages 272-284. Springer, 2022. 2", + "[17] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 574-584, 2022. 2", + "[18] Mohammad Havaei, Nicolas Guizard, Nicolas Chapados, and Yoshua Bengio. Hemis: Hetero-modal image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI, pages 469-477. Springer, 2016. 7", + "[19] Xuanhua He, Ke Cao, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Pan-mamba: Effective pan-sharpening with state space model. arXiv preprint arXiv:2402.12192, 2024.3", + "[20] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5", + "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 5", + "[22] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3, 4", + "[23] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In International Conference on Learning Representations, 2016. 3", + "[24] Zeyu Jiang, Changxing Ding, Minfeng Liu, and Dacheng Tao. Two-stage cascaded u-net: 1st place solution to brats challenge 2019 segmentation task. In *Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries*, pages 231-241. Springer, 2020. 2", + "[25] Sanaz Karimijafarbigloo, Reza Azad, Amirhossein Kazerouni, Saeed Ebadollahi, and Dorit Merhof. Mmcformer: Missing modality compensation transformer for brain tumor segmentation. In Medical Imaging with Deep Learning, pages 1144-1162. PMLR, 2024. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25990", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Jonghun Kim and Hyunjin Park. Adaptive latent diffusion model for 3d medical image to image translation: Multimodal magnetic resonance imaging study. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7604-7613, 2024. 1, 2", + "[27] Dongwook Lee, Won-Jin Moon, and Jong Chul Ye. Assessing the importance of magnetic resonance contrasts using collaborative generative adversarial networks. Nature Machine Intelligence, 2(1):34-42, 2020. 1, 2", + "[28] Ho Hin Lee, Shunxing Bao, Yuankai Huo, and Bennett A Landman. 3d ux-net: A large kernel volumetric convnet modernizing hierarchical transformer for medical image segmentation. arXiv preprint arXiv:2209.15076, 2022. 2", + "[29] Jiangyun Li, Wenxuan Wang, Chen Chen, Tianxiang Zhang, Sen Zha, Hong Yu, and Jing Wang. Transbtsv2: Wider instead of deeper transformer for medical image segmentation. arXiv preprint arXiv:2201.12785, 2022. 2", + "[30] Mingcheng Li, Dingkang Yang, Xiao Zhao, Shuaibing Wang, Yan Wang, Kun Yang, Mingyang Sun, Dongliang Kou, Ziyun Qian, and Lihua Zhang. Correlation-decoupled knowledge distillation for multimodal sentiment analysis with incomplete modalities. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2024. 2", + "[31] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 3", + "[32] Hong Liu, Dong Wei, Donghuan Lu, Jinghan Sun, Liansheng Wang, and Yefeng Zheng. M3ae: Multimodal representation learning for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1657-1665, 2023. 1, 2, 6, 7", + "[33] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 1, 3, 4, 6", + "[34] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 3", + "[35] Andriy Marusyk, Vanessa Almendro, and Kornelia Polyak. Intra-tumour heterogeneity: a looking glass for cancer? Nature reviews cancer, 12(5):323-334, 2012. 1, 4", + "[36] Gonzalo Mena, David Belanger, Scott Linderman, and Jasper Snoek. Learning latent permutations with gumbel-sinkhorn networks. In International Conference on Learning Representations, 2018. 3", + "[37] Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging, 34(10):1993-2024, 2014. 6", + "[38] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977, 2024. 3, 4", + "[39] Yansheng Qiu, Delin Chen, Hongdou Yao, Yongchao Xu, and Zheng Wang. Scratch each other's back: Incomplete" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "multi-modal brain tumor segmentation via category aware group self-support learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21317-21326, 2023. 1, 2, 7", + "[40] Santiago Ramón y Cajal, Marta Sesé, Claudia Capdevila, Trond Aasen, Leticia De Mattos-Arruda, Salvador J Diaz-Cano, Javier Hernández-Losa, and Josep Castellvi. Clinical implications of intratumor heterogeneity: challenges and opportunities. Journal of Molecular Medicine, 98:161-177, 2020. 1, 4", + "[41] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 3", + "[42] Dong She, Yueyi Zhang, Zheyu Zhang, Hebei Li, Zihan Yan, and Xiaoyan Sun. Eoformer: Edge-oriented transformer for brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 333-343. Springer, 2023. 2", + "[43] Jimmy TH Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling. In International Conference on Learning Representations, 2022. 2", + "[44] Chloe Wang, Oleksii Tsepa, Jun Ma, and Bo Wang. Graphmamba: Towards long-range graph sequence modeling with selective state spaces. arXiv preprint arXiv:2402.00789, 2024.3", + "[45] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15878-15887, 2023. 6, 7", + "[46] Wenxuan Wang, Chen Chen, Meng Ding, Hong Yu, Sen Zha, and Jiangyun Li. Transbts: Multimodal brain tumor segmentation using transformer. In Medical Image Computing and Computer Assisted Intervention-MICCAI, pages 109-119. Springer, 2021. 2", + "[47] Yixin Wang, Yang Zhang, Yang Liu, Zihao Lin, Jiang Tian, Cheng Zhong, Zhongchao Shi, Jianping Fan, and Zhiqiang He. Acn: adversarial co-training network for brain tumor segmentation with missing modalities. In Medical Image Computing and Computer Assisted Intervention-MICCAI, pages 410-420. Springer, 2021. 1, 2", + "[48] Shicai Wei, Chunbo Luo, and Yang Luo. Mmanet: Margin-aware distillation and modality-aware regularization for incomplete multimodal learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20039-20049, 2023. 1, 2, 7", + "[49] Jing Nathan Yan, Jiatao Gu, and Alexander M Rush. Diffusion models without attention. arXiv preprint arXiv:2311.18257, 2023. 3", + "[50] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. arXiv preprint arXiv:2403.17695, 2024. 1, 3, 4", + "[51] Heran Yang, Jian Sun, and Zongben Xu. Learning unified hyper-network for multi-modal mr image synthesis and tu" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25991", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "mor segmentation with missing modalities. IEEE Transactions on Medical Imaging, 2023. 1, 2", + "[52] Qiushi Yang, Xiaoqing Guo, Zhen Chen, Peter YM Woo, and Yixuan Yuan. D 2-net: Dual disentanglement network for brain tumor segmentation with missing modalities. IEEE Transactions on Medical Imaging, 41(10):2953-2964, 2022. 2", + "[53] Chenyu You, Ruihan Zhao, Fenglin Liu, Siyuan Dong, Sandeep Chinchali, Ufuk Topcu, Lawrence Staib, and James Duncan. Class-aware adversarial transformers for medical image segmentation. pages 29582-29596, 2022. 2", + "[54] Ziqi Yu, Xiaoyang Han, Shengjie Zhang, Jianfeng Feng, Tingying Peng, and Xiao-Yong Zhang. Mousegan++: Unsupervised disentanglement and contrastive representation for multiple mri modalities synthesis and structural segmentation of mouse brain. IEEE Transactions on Medical Imaging, 2022, 1, 2", + "[55] Yao Zhang, Nanjun He, Jiawei Yang, Yuexiang Li, Dong Wei, Yawen Huang, Yang Zhang, Zhiqiang He, and Yefeng Zheng. mmformer: Multimodal medical transformer for incomplete multimodal learning of brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 107-117. Springer, 2022. 1, 2, 6, 7", + "[56] Zheyu Zhang, Xinzhao Liu, Zheng Chen, Yueyi Zhang, Huanjing Yue, Yunwei Ou, and Xiaoyan Sun. Anatomical consistency distillation and inconsistency synthesis for brain tumor segmentation with missing modalities. arXiv preprint arXiv:2408.13733, 2024.", + "[57] Zheyu Zhang, Gang Yang, Yueyi Zhang, Huanjing Yue, Aiping Liu, Yunwei Ou, Jian Gong, and Xiaoyan Sun. Tmformer: Token merging transformer for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7414-7422, 2024. 2", + "[58] Sijie Zhao, Hao Chen, Xueliang Zhang, Pengfeng Xiao, Lei Bai, and Wanli Ouyang. Rs-mamba for large remote sensing image dense prediction. arXiv preprint arXiv:2404.02668, 2024. 1, 3", + "[59] Zechen Zhao, Heran Yang, and Jian Sun. Modality-adaptive feature interaction for brain tumor segmentation with missing modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 183-192. Springer, 2022. 1, 7", + "[60] Zhuoran Zheng and Jun Zhang. Fd-vision mamba for endoscopic exposure correction. arXiv preprint arXiv:2402.06378, 2024. 3", + "[61] Chenhong Zhou, Changxing Ding, Xinchao Wang, Zhentai Lu, and Dacheng Tao. One-pass multi-task networks with cross-task guided attention for brain tumor segmentation. IEEE Transactions on Image Processing, 29:4516-4529, 2020. 2", + "[62] Tao Zhou, Huazhu Fu, Geng Chen, Jianbing Shen, and Ling Shao. Hi-net: hybrid-fusion network for multi-modal mr image synthesis. IEEE Transactions on Medical Imaging, 39 (9):2772-2781, 2020. 1, 2", + "[63] Tongxue Zhou, Stéphane Canu, Pierre Vera, and Su Ruan. Latent correlation representation learning for brain tumor" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "segmentation with missing mri modalities. IEEE Transactions on Image Processing, 30:4263-4274, 2021. 1, 2", + "[64] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: Redesigning skip connections to exploit multiscale features in image segmentation. IEEE Transactions on Medical Imaging, 39(6):1856-1867, 2019. 2", + "[65] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. 2024. 1, 3, 4, 6" + ], + "bbox": [ + 516, + 92, + 903, + 246 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25992", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/58ab88d4-b760-40c9-8213-6c50aa676cf3_model.json b/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/58ab88d4-b760-40c9-8213-6c50aa676cf3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a71c12bdb79bc8e3147f87f736f07e6e44922d01 --- /dev/null +++ b/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/58ab88d4-b760-40c9-8213-6c50aa676cf3_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.764, + 0.177 + ], + "angle": 0, + "content": "Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.202, + 0.912, + 0.327 + ], + "angle": 0, + "content": "Zheyu Zhang\\(^{1,2}\\) Yayuan Lu\\(^{1,2}\\) Feipeng Ma\\(^{1,2}\\) Yueyi Zhang\\(^{1,2}\\) Huanjing Yue\\(^{3}\\) Xiaoyan Sun\\(^{1,2,*}\\) \n\\(^{1}\\)MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, 230026, China \n\\(^{2}\\)Anhui Province Key Laboratory of Biomedical Imaging and Intelligent Processing, Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, Hefei 230088, China \n\\(^{3}\\)School of Electrical and Information Engineering, Tianjin University, Tianjin, 300072, China \nzhangzy0@mail.ustc.edu.cn sunxiaoyan@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.362, + 0.327, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.486, + 0.759 + ], + "angle": 0, + "content": "Brain tumor segmentation plays a crucial role in clinical diagnosis, yet the frequent unavailability of certain MRI modalities poses a significant challenge. In this paper, we introduce the Learnable Sorting State Space Model (LS3M), a novel framework designed to maximize the utilization of available modalities for brain tumor segmentation. LS3M excels at efficiently modeling long-range dependencies based on the Mamba design, while incorporating differentiable permutation matrices that reorder input sequences based on modality-specific characteristics. This dynamic reordering ensures that critical spatial inductive biases and long-range semantic correlations inherent in 3D brain MRI are preserved, which is crucial for incomplete multi-modal brain tumor segmentation. Once the input sequences are reordered using the generated permutation matrix, the Series State Space Model (S3M) block models the relationships between them, capturing both local and long-range dependencies. This enables effective representation of intra-modal and inter-modal relationships, significantly improving segmentation accuracy. Extensive experiments on the BraTS2018 and BraTS2020 datasets demonstrate that LS3M outperforms existing methods, offering a robust solution for brain tumor segmentation, particularly in scenarios with missing modalities." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.788, + 0.222, + 0.804 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.485, + 0.876 + ], + "angle": 0, + "content": "Accurate brain tumor segmentation in clinical settings relies on the integration of various MRI modalities (FLAIR, T1ce, T1, T2) that provide complementary information for diagnosis. However, the frequent unavailability of cer-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.908, + 0.667 + ], + "angle": 0, + "content": "tain modalities, like T1ce due to contrast agent allergies, presents a challenge. Existing methods addressing missing modalities fall into two categories: (1) bridging the data gap between complete and missing modalities through modality synthesis [26, 27, 32, 51, 54, 62] and knowledge distillation [2, 39, 47, 48] and (2) optimizing architectures to maximize information extraction from available modalities, which is the focus of our work. Some methods [8, 63] adopt multi-scale convolutions to capture critical structural details by focusing on spatial inductive bias, such as local texture. However, the limited receptive field restricts their ability to represent tumor tissue effectively, especially when other areas may appear visually similar in localized views, making it difficult to differentiate tumor tissue [35, 40]. Some methods explore models such as Transformers [55] and graph neural networks [51, 59] to build long-range semantic correlation to mitigate the mentioned problem, but they suffer from quadratic computational complexity. The computational inefficiency is further exacerbated by the large sequence lengths typical of 3D brain MRI." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recent advancements in selective state space models (SSMs), such as Mamba [7, 13], offer the potential to obtain long-range representations for available modalities with dynamic weighting parameters in an efficient way. Mamba flattens the 3D brain MRI into a 1D sequence and iteratively scans this sequence to compress hidden states, which are then used to enhance the current token by combining it with previously scanned tokens. While Mamba can efficiently build long-range dependencies, the influence of distant tokens on the current token diminishes significantly. This becomes problematic for 3D brain MRI, as flattening the sequence destroys the 3D spatial structure. Follow-up methods [33, 50, 58, 65] propose bi-directional, four-directional, and omni-directional scanning to mitigate distortion to spatial structures, as shown in Fig. 1. However, these methods" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25982" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "either neglect local inductive bias or fail to capture long-range semantic correlations of 3D brain MRI, hindering their ability to fully capture contextual relationships essential for brain tumor segmentation, particularly in scenarios with missing modalities." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.172, + 0.485, + 0.64 + ], + "angle": 0, + "content": "To address these challenges, we propose the Learnable Sorting State Space Model (LS3M), specifically designed for brain tumor segmentation with missing modalities. Firstly, LS3M incorporates the Sinkhorn operator with the Gumbel-Softmax reparameterization, which converts the traditional non-differentiable sorting into a differentiable permutation matrix. This matrix is used to reorder the 1D sequence of available modalities, considering spatial inductive biases and long-range semantic correlations. Importantly, the permutation matrix does not directly establish relationships; rather, it prepares the sequence for further processing by the Series State Space Model (S3M), which is capable of modeling intra-modal and inter-modal relationships by leveraging both local and long-range correlations. Next, LS3M scales the intra-modal permutation matrix at larger resolutions, allowing for the extraction of multi-scale information. These modality-specific permutation matrices are also combined for multi-modal fusion to capture correlations across modalities. The reordered sequences, now enriched with spatial and semantic information, are then processed by the S3M block, which builds relationships at both intra-modal and inter-modal levels, maximizing the extraction of information from the available modalities. Lastly, LS3M departs from existing models that rely on cropped input volumes by adopting a global input strategy, supplemented with a relative position mask. This approach not only provides richer contextual information but also informs the model about the location of the cropped volume within the larger global context, significantly enhancing spatial awareness and diagnostic accuracy. In summary, our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.645, + 0.483, + 0.719 + ], + "angle": 0, + "content": "- We propose LS3M for brain tumor segmentation with missing modalities. LS3M improves SSM by dynamically reordering data, capturing patient-specific characteristics, and enhancing the ability to extract local and long-range correlations from available modalities." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.72, + 0.483, + 0.796 + ], + "angle": 0, + "content": "- LS3M scales and integrates the permutation matrix to capture multi-scale and multi-modal information. The reordered sequences are then processed by the S3M block to model complex intra-modal and inter-modal relationships, improving segmentation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.796, + 0.483, + 0.854 + ], + "angle": 0, + "content": "- LS3M incorporates a global input strategy with relative position embeddings, significantly enhancing contextual awareness towards comprehensive utilization of available modalities." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- Extensive evaluations on BraTS2018 and BraTS2020 demonstrate the superiority and effectiveness of LS3M for brain tumor segmentation, particularly in clinically" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.645, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.09, + 0.623, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.154, + 0.623, + 0.162 + ], + "angle": 0, + "content": "Long-range Semantic Correlation" + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.165, + 0.61, + 0.178 + ], + "angle": 0, + "content": "(a) Requirements for identifying brain tumor" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.093, + 0.712, + 0.125 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.127, + 0.706, + 0.134 + ], + "angle": 0, + "content": "(b) Horizontal Scanning" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.136, + 0.712, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.17, + 0.7, + 0.177 + ], + "angle": 0, + "content": "(d) Local Scanning" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.093, + 0.803, + 0.125 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.127, + 0.791, + 0.134 + ], + "angle": 0, + "content": "(c) Vertical Scanning" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.136, + 0.826, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.17, + 0.833, + 0.177 + ], + "angle": 0, + "content": "(e) Our Sorted Scanning" + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.093, + 0.884, + 0.125 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.854, + 0.127, + 0.884, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.192, + 0.907, + 0.262 + ], + "angle": 0, + "content": "Figure 1. (a) Brain tumor segmentation requires local inductive bias to capture small but critical structural details, as well as long-range semantic correlations to understand spatial relationships between the tumor and surrounding structures. (b)-(e) Different scanning strategies." + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.29, + 0.813, + 0.305 + ], + "angle": 0, + "content": "relevant scenarios with missing modalities." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.32, + 0.655, + 0.335 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Medical Image Segmentation. Medical image segmentation traditionally relies on U-Net-like structures, which formulate the task as dense pixel classification. Enhancements to segmentation structures include cascaded networks [24, 61], refined skip connections [64], hierarchical pyramids [9], attention mechanisms [4, 6, 12, 16, 17, 29, 42, 46, 53], and large depth-wise kernels [28]. Despite achieving notable performance, these methods often fail when certain modalities are absent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.79 + ], + "angle": 0, + "content": "Multi-modal Learning for Missing Modalities. Research in multi-modal learning with missing modalities generally seeks to either bridge the gap between scenarios with complete and missing modalities or to optimize architectural design to maximize the use of available modal information. The former approach includes modality synthesis [26, 27, 32, 51, 54, 62] to generate absent modalities and knowledge distillation techniques [2, 30, 39, 47, 48] that transfer insights from complete to incomplete scenarios. The latter approach utilizes techniques like modality disentanglement [5, 52] to separate modality-specific information in the frequency domain and modality fusion strategies involving hierarchical pyramids [8, 63] and attention mechanisms [25, 55-57]. However, combining attention mechanisms with other techniques often results in significant computational overhead, which can impede effective modeling long-range dependencies. In our work, we leverage SSM, specifically the Mamba, to efficiently build long-range correlations from all available modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "State Space Model. Drawing inspiration from control theory, the State Space Model (SSM) recurrently updates sequences through latent states [14, 43], as demonstrated by the Structured State Space Sequence model (S4), which utilizes linear state space equations to efficiently scale with sequence length. Recently, Gu and Dao [13] propose a" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25983" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "generic modal, called Mamba, with selective projection, gated mechanism and efficient hardware design, outperforming Transformers in linear-time inference on natural language processing. This success has spurred adaptations of Mamba in computer vision. Since visual tokens lack a sequential arrangement, methods propose diverse scanning strategies like bi-/four-/omni-directional [33, 50, 58, 65], local [22], and atrous scanning [38], enhancing contextual data integration for tasks such as medical image segmentation [34, 41, 41], image restoration [15, 19, 60], point cloud analysis [31], image generation [11, 49], and graph neural networks [3, 44]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.284, + 0.231, + 0.3 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.371 + ], + "angle": 0, + "content": "State Space Model (SSM) for Sequence Modeling. The SSM defines a linear Ordinary Differential Equation (ODE) to map an input sequence \\( x(t) \\in \\mathbb{R} \\) to an output sequence \\( y(t) \\in \\mathbb{R} \\) through a hidden state \\( h(t) \\in \\mathbb{R}^N \\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.376, + 0.482, + 0.394 + ], + "angle": 0, + "content": "\\[\nh ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), \\quad y (t) = \\mathbf {C} h (t), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.458 + ], + "angle": 0, + "content": "where \\(\\mathbf{A} \\in \\mathbb{R}^{N \\times N}\\), \\(\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}\\), and \\(\\mathbf{C} \\in \\mathbb{R}^{1 \\times N}\\) are learned parameters. Since images are processed into discrete token sequences, the discrete version of Eq. (1) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.465, + 0.482, + 0.483 + ], + "angle": 0, + "content": "\\[\nh _ {k} = \\overline {{\\mathbf {A}}} h _ {k - 1} + \\overline {{\\mathbf {B}}} x _ {k}, \\quad y _ {k} = \\mathbf {C} h _ {k}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.567 + ], + "angle": 0, + "content": "where \\(\\overline{\\mathbf{A}}\\) and \\(\\overline{\\mathbf{B}}\\) are discretized using the discretization step size \\(\\Delta\\) in the zero-order hold (ZOH) rule. The long-range dependencies are constructed via recurrent scanning. However, the recurrent SSM is not practical. Eq. (2) is equivalent to a global convolution kernel \\(\\overline{\\mathbf{K}}\\in \\mathbb{R}^{L}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.571, + 0.482, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {K}}} = \\left(\\mathbf {C} \\overline {{\\mathbf {B}}}, \\mathbf {C} \\overline {{\\mathbf {A B}}}, \\dots , \\mathbf {C} \\overline {{\\mathbf {A}}} ^ {\\mathbf {L} - 1} \\overline {{\\mathbf {B}}}\\right), \\quad \\mathbf {y} = \\mathbf {x} * \\overline {{\\mathbf {K}}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.484, + 0.678 + ], + "angle": 0, + "content": "Structured state space sequence models (S4) imposes a structure on the state matrix \\(\\mathbf{A}\\) based on HIPPO matrices, which significantly improves the efficiency for obtaining the SSM kernel \\(\\overline{\\mathbf{K}}\\). Further, Mamba [13] adds selective projection via making \\(\\mathbf{B}, \\mathbf{C}, \\Delta\\) input-dependent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.484, + 0.783 + ], + "angle": 0, + "content": "Sinkhorn Operator. The Sinkhorn operator \\( S(\\cdot) \\) is an iterative procedure used to transform a non-negative matrix \\( \\mathbf{P} \\in \\mathbb{R}^{L \\times L} \\) into a doubly stochastic matrix, which is a square matrix with non-negative entries such that each row and each column sums to one [1, 36]. The process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.786, + 0.483, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l}S ^ {0} (\\mathbf {P}) = \\exp (\\mathbf {P}), \\quad S ^ {n} (\\mathbf {P}) = \\mathcal {T} _ {c} \\left(\\mathcal {T} _ {r} \\left(S ^ {n - 1} (\\mathbf {P})\\right)\\right),\\\\S (\\mathbf {P}) = \\lim _ {n \\rightarrow \\infty} S ^ {n} (\\mathbf {P}),\\end{array}\\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.483, + 0.864 + ], + "angle": 0, + "content": "where \\(n\\) is iterative times. \\(\\mathcal{T}_r(\\cdot)\\) and \\(\\mathcal{T}_c(\\cdot)\\) are the row-wise and column-wise normalizations, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.869, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {r} (\\mathbf {P}) = \\mathbf {P} \\oslash \\left(\\mathbf {P} \\mathbf {1} _ {\\mathbf {N}} \\mathbf {1} _ {\\mathbf {N}} ^ {\\top}\\right), \\mathcal {T} _ {c} (\\mathbf {P}) = \\mathbf {P} \\oslash \\left(\\mathbf {1} _ {\\mathbf {N}} \\mathbf {1} _ {\\mathbf {N}} ^ {\\top} \\mathbf {P}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "where \\(\\oslash\\) denotes the element-wise division and \\(1_{N}\\) is a column vector of ones. We employ Sinkhorn \\((\\mathbf{P},n)\\) to represent Eq. (4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.155, + 0.907, + 0.23 + ], + "angle": 0, + "content": "Gumbel-Softmax Reparameterization. The Gumbel-Softmax provides a differentiable approximation of the discrete categorical distribution [23]. The process of sample \\( z \\) from a categorical distribution with class probabilities \\( x \\) can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.236, + 0.907, + 0.261 + ], + "angle": 0, + "content": "\\[\nz = \\operatorname {o n e - h o t} (\\underset {i} {\\arg \\max } [ g _ {i} + \\log x _ {i} ]), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.268, + 0.907, + 0.314 + ], + "angle": 0, + "content": "where \\( g_{i} \\) is the injected standard i.i.d Gumbel noise. Jang et al. [23] use the softmax function as a continuous, differentiable approximation to the arg max function, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.321, + 0.906, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s o f t m a x} (y / \\tau) \\equiv \\frac {\\exp \\left(y _ {i} / \\tau\\right)}{\\sum_ {i = 1} ^ {k} \\exp \\left(y _ {i} / \\tau\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.365, + 0.906, + 0.442 + ], + "angle": 0, + "content": "where \\( y_{i} = \\log x_{i} + g_{i} \\). When temperature \\( \\tau \\rightarrow 0 \\), the softmax distribution converges to a one-hot representation identical to samples from a categorical distribution. We use Gumbel-Softmax(X, \\( \\tau \\)) to represent Eq. (7), where \\( \\mathbf{X} = [x_0, x_1, \\dots, x_{|\\mathbf{X}| - 1}] \\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.453, + 0.605, + 0.468 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.584 + ], + "angle": 0, + "content": "In this section, we first explain the motivation and challenges with Mamba. We then introduce the learnable sorting process to address these challenges, followed by expanding the permutation matrix for longer and multi-modal sequences. Next, we describe the S3M block and the global input strategy. Finally, we present the LS3M architecture and its optimization loss." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.592, + 0.642, + 0.607 + ], + "angle": 0, + "content": "4.1. Motivations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Brain tumor segmentation requires complementary information from multiple modalities. Missing modalities lead to insufficient information, inspiring us to maximize information extraction from available modalities. We try to do this from two aspects: building long-range dependencies and incorporating broader context for available modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.87 + ], + "angle": 0, + "content": "First, brain tumors often exhibit complex anatomy, where local inductive biases alone may be insufficient to identify tumor types and boundaries. Long-range dependencies are essential for understanding spatial relationships between the tumor and surrounding structures. While Transformers can model such dependencies, their quadratic computational complexity makes them impractical for 3D MRI data due to large token sequences and high computational costs, as illustrated in the Appendix. Mamba, however, captures long-range dependencies more efficiently, making it a better fit for this task." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Second, most segmentation methods focus on cropped volumes to reduce computational cost, neglecting the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25984" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.593, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.244, + 0.42, + 0.257 + ], + "angle": 0, + "content": "(a) Our LS3M framework" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.091, + 0.897, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.244, + 0.676, + 0.258 + ], + "angle": 0, + "content": "(b) \\(\\mathrm{SortP_M}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.244, + 0.765, + 0.257 + ], + "angle": 0, + "content": "(c) \\(\\mathrm{SortP_L}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.242, + 0.867, + 0.256 + ], + "angle": 0, + "content": "(d) \\(\\mathrm{SortP_{Mul}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.27, + 0.908, + 0.327 + ], + "angle": 0, + "content": "Figure 2. (a) The comprehensive LS3M framework integrates the global input strategy with adaptive pooling, the SortP for generating permutation matrices, and the S3M for feature enhancement. \\((\\cdot)_M\\) denotes that we perform \\((\\cdot)\\) process separately for each modality. (b) SortP \\(M\\) means the learnable sorting process applied to each modality. (c) SortP \\(L\\) denotes obtaining the larger permutation matrix for longer sequences. (d) SortP \\(Mul\\) refers to the generation of the multi-modal permutation matrix." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.353, + 0.476, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.433, + 0.441, + 0.447 + ], + "angle": 0, + "content": "Figure 3. The structure of the proposed S3M block." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.484, + 0.61 + ], + "angle": 0, + "content": "broader context crucial for accurately understanding complex conditions like brain tumors. This is especially important in the case of intra-tumor heterogeneity, which requires a wider context to characterize the tumor type [35, 40]. Moreover, in scenarios with missing modalities, it is critical to utilize all available data from the entire image to compensate for the missing information. Thus, incorporating broader context without significant computational overhead is essential." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.458, + 0.637 + ], + "angle": 0, + "content": "4.2. Challenges of Existing Mamba Approaches" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.672 + ], + "angle": 0, + "content": "Mamba is primarily designed for 1D sequences, with characteristics that include:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Mamba Characteristic 1: the current output token \\( y_{k} \\) is limited to modeling relationships solely with previously scanned tokens \\( x_{i}, i \\in \\{0,1,\\dots ,k - 1,k\\} \\), as detailed in Eq. (2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Mamba Characteristic 2: the influence of distant tokens diminishes significantly due to the weight \\(\\overline{\\mathbf{A}}^{i - 1}\\), as explained in Eq. (3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To mitigate these limitations, existing Mamba-based methods [33, 38, 50, 65] introduce forward and backward scanning strategies, which enable the current token to model relationships with both preceding and subsequent tokens. These methods are more suitable for vision applications, where local dependencies can often provide sufficient context. However, in the case of missing modalities, it is insufficient to rely solely on local context for tumor identi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.905, + 0.458 + ], + "angle": 0, + "content": "fication. Tumor heterogeneity often requires capturing information from distant regions of the MRI to fully understand the tumor's characteristics, as highlighted by prior studies [35, 40]. For example, in the segmentation of the whole tumor (WT) with only T1ce modality, local features may be insufficient, and distant correlations are necessary to improve segmentation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.906, + 0.565 + ], + "angle": 0, + "content": "Due to Mamba Characteristic 2, the influence of distant tokens diminishes over sequence steps, which is problematic in 3D MRI segmentation tasks. Therefore, a dynamic mechanism is required to ensure that the most relevant tokens—both spatially close and semantically related over long distances—are positioned closer to the target token, facilitating more effective modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Although LocalMamba [22] uses a Z-scan to ensure local relationships in 2D, this approach cannot guarantee the preservation of 3D neighborhood relationships, such as the 27-neighborhood in 3D space. Even with axial scanning added to horizontal and vertical scanning, the relationships between distant tokens in 3D space are still not adequately preserved. Consequently, there is a need for a solution that dynamically reorders the tokens for each 3D MRI scan, ensuring that all potentially relevant tokens—whether spatially close or far but semantically related—are placed before the current token. This dynamic reordering will allow the subsequent S3M block to model these relationships more effectively, improving the segmentation accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.753, + 0.789 + ], + "angle": 0, + "content": "4.3. Learnable Sorting Process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Traditional Non-differentiable Sorting Process. A pivotal aspect of LS3M involves addressing the non-differentiability of traditional sorting mechanisms. The patient-specific dynamic sorting process is initially driven by a predictive score \\( P \\in \\mathbb{R}^{L} \\), assigned to each token in the sequence \\( \\mathbf{X} \\). These scores dictate the token positions in the sorted sequence \\( \\mathbf{X}_{\\mathbf{S}} \\), subsequently processed by the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25985" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "Mamba block. However, the traditional sorting function, which maps these continuous input scores to discrete token indices, interrupts the gradient flow necessary for end-to-end training of the network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.202 + ], + "angle": 0, + "content": "The Learnable Sorting Process. The learnable sorting process is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.212, + 0.258, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {P}} = \\operatorname {S o r t N e t} (\\mathbf {f}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.234, + 0.483, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} = \\text {S i n k h o r n} (\\text {G u m b e l - S o f t m a x} (\\hat {\\mathbf {P}}, \\tau), n), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.254, + 0.374, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} = \\text {f l a t t e n} (\\mathbf {f}), \\quad \\mathbf {X} _ {\\mathbf {S}} = \\mathbf {P X}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.484, + 0.477 + ], + "angle": 0, + "content": "This enable end-to-end differentiable sorting. As shown in Fig. 2(b), firstly, our SortNet generates an estimated matrix \\(\\hat{\\mathbf{P}}\\in \\mathbb{R}^{L\\times L}\\) tailored to individual patients. It employs a convolutional layer with a kernel size of \\(C\\times C\\times 3\\times 3\\times 3\\) to capture spatial inductive biases, extracting information from the 26-neighborhood 3D spatial structure. After convolution, a SiLU activation function [20] is applied to enhance the network's representational capacity. The resulting feature map is split into \\(\\mathbf{f}_1\\) and \\(\\mathbf{f}_2\\), with \\(\\mathbf{f}_2\\) being processed to compute a similarity map that captures long-range semantic correlations. A ReLU activation is then applied to ensure the non-negativity of \\(\\hat{\\mathbf{P}}\\), preparing it for subsequent processing by the Sinkhorn operator." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.628 + ], + "angle": 0, + "content": "Secondly, we use the Gumbel-Softmax and Sinkhorn operators to generate a permutation matrix \\(\\mathbf{P}\\), where each row and column contains exactly one 1 with all other entries being 0s. The estimated matrix \\(\\hat{\\mathbf{P}}\\) is sliced row-wise, with each row being passed through the Gumbel-Softmax function. By adjusting the temperature parameter \\(\\tau\\), the output satisfies the one-hot condition for each row. The resulting rows are then concatenated and passed through the Sinkhorn operator, which iteratively adjusts the rows and columns of the matrix, yielding the final permutation matrix \\(\\mathbf{P}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.749 + ], + "angle": 0, + "content": "Finally, the feature map \\(\\mathbf{f}\\) is flattened into a sequence \\(\\mathbf{X}\\), and \\(\\mathbf{P}\\) is applied to \\(\\mathbf{X}\\) to perform the sorting operation. This process can be viewed as permuting the rows of an \\(L\\times L\\) identity matrix according to the sorting procedure. Since all of these operations are differentiable, the entire sorting process is also differentiable, which ensures that each token is optimally positioned in the sequence for subsequent processing by our S3M block." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.479, + 0.774 + ], + "angle": 0, + "content": "4.4. Handling Longer and Multi-modal Sequences" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Handling Longer Sequence. Extracting multi-scale information from available modalities requires generating larger permutation matrices to accommodate longer sequences corresponding to high-resolution feature maps. However, this process introduces quadratic computational complexity due to the matrix multiplications involved. To address this, we propose an efficient alternative: bilinear interpolation to scale the permutation matrix, eliminating" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "the need for SortNet. The scaled matrix is then processed through ReLU activation, Gumbel-Softmax, and Sinkhorn to generate the permutation matrix \\(\\mathbf{P}_{\\mathrm{Large}}\\) for the extended sequence, as represented in Fig. 2(c)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.19, + 0.907, + 0.432 + ], + "angle": 0, + "content": "Handling Multi-modal Sequences. To effectively leverage inter-modal correlations, we aim to generate a permutation matrix that accounts for scenarios with missing modalities. We dynamically fuse the individual permutation matrices of each modality—\\(\\mathbf{P}_{\\mathrm{FLAIR}}\\), \\(\\mathbf{P}_{\\mathrm{T1ce}}\\), \\(\\mathbf{P}_{\\mathrm{T1}}\\), \\(\\mathbf{P}_{\\mathrm{T2}}\\) with a multi-modal code, represented by \\(h = \\{h_1, h_2, h_3, h_4\\}\\), where \\(h_i \\in \\{0, 1\\}\\) indicates whether the corresponding modality is available. This fusion produces a comprehensive multi-modal estimated matrix, \\(\\hat{\\mathbf{P}}_{\\mathrm{Mul}}\\). We then apply a ReLU activation followed by Gumbel-Softmax and Sinkhorn operations to generate the final multi-modal permutation matrix \\(\\mathbf{P}_{\\mathrm{Mul}}\\), as shown in Fig. 2(d). Before reordering the multi-modal sequence, each modal sequence is initially projected into a unified representation space. The sequences are then sorted according to \\(\\mathbf{P}_{\\mathrm{Mul}}\\), ensuring alignment based on inter-modal relevance and coherence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.452, + 0.843, + 0.468 + ], + "angle": 0, + "content": "4.5. Series State Space Model (S3M) Block" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Due to the insufficient information of missing modalities, it becomes crucial to exhaustively mine information from all available modalities. We introduce the S3M model for local and long-range dependency modeling, inspired by the success of Mamba." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.906, + 0.631 + ], + "angle": 0, + "content": "As shown in Fig. 3, our S3M block connects two unidirectional Mamba-like blocks, \\( F_{Fwd} \\) and \\( F_{Bwd} \\), in a series connection but in opposite directions, together with a Channel Attention (FCA) block [21], which can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.651, + 0.905, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\mathbf {S 3 M}} = \\mathrm {F} _ {\\mathrm {C A}} \\left(\\mathrm {F} _ {\\mathrm {B w d}} \\left(\\mathrm {F} _ {\\mathrm {F w d}} \\left(\\mathbf {f} _ {\\mathbf {I n}}, \\mathbf {P}\\right)\\right)\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.906, + 0.732 + ], + "angle": 0, + "content": "Before processing, the feature \\(\\mathbf{f}_{\\mathrm{In}}\\) is reshaped into a sequence and sorted by multiplying it with a soft permutation matrix \\(\\mathbf{P}\\) generated through the learnable sorting process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Each Mamba-like block comprises three branches: a basic SSM branch, a gated weighting branch, and a residual connection branch. In the SSM branch, the input feature \\(\\mathbf{f}_{\\mathrm{In}}\\) undergoes a linear projection to expand its channels, followed by depth-wise convolution and SiLU activation, producing \\(\\mathbf{f}_{\\mathrm{Mid}}\\) and the SSM parameters \\(\\{\\overline{\\mathbf{A}},\\overline{\\mathbf{B}},\\mathbf{C}\\}\\). They are sent to the Forward/Backward SSM module, followed by layer normalization to obtain \\(\\mathbf{f}_{\\mathrm{SSM}}\\). The input feature \\(\\mathbf{f}_{\\mathrm{In}}\\) is processed through a linear layer and SiLU activation to generate a gated weight feature map \\(\\mathbf{f}_{\\mathrm{Gate}}\\). Combining these with the residual connection branch, the output of the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25986" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.092, + 0.092, + 0.236, + 0.105 + ], + "angle": 0, + "content": "Mamba-like block is:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.124, + 0.486, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {f} _ {\\mathbf {S S M}} = \\mathrm {L N} (\\mathrm {S S M} (\\mathrm {S i L U} (\\mathrm {D W - C o n v} (\\mathrm {L i n e a r} (\\mathrm {L N} (\\mathbf {f} _ {\\mathbf {I n}})))))), \\\\ \\mathbf {f} _ {\\text {G a t e}} = \\operatorname {S i L U} \\left(\\operatorname {L i n e a r} \\left(\\mathbf {f} _ {\\text {I n}}\\right)\\right), \\\\ \\mathbf {f} _ {\\text {O u t}} = \\operatorname {L i n e a r} \\left(\\mathbf {f} _ {\\text {S S M}} \\odot \\mathbf {f} _ {\\text {G a t e}}\\right) + \\mathbf {f} _ {\\text {I n}}. \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.196, + 0.484, + 0.376 + ], + "angle": 0, + "content": "For further channel-wise interactions, we introduce the Channel Attention block. Our S3M block, when compared to the recent Vim block [65] and SS2D block [33], offers two main advantages: (1) The series connection in S3M ensures that information from the forward SSM block is integrated into subsequent tokens, which is then relayed back to earlier tokens in the backward SSM block. This contrasts with parallel connections in Vim block and SS2D block, where token sequences receive the same information level. (2) The integration of the permutation matrix in S3M preserves both spatial proximity and long-range semantic correlations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.484, + 0.609 + ], + "angle": 0, + "content": "For multi-modal fusion, each modal sequence is first projected into a unified representation space and then concatenated along the channel dimension. The multi-modal sequences \\(\\mathbf{X}_{\\mathrm{Mul}}\\) are sorted by \\(\\mathbf{P}_{\\mathrm{Mul}}\\), which are further delivered into our S3M for building inter-modal correlations. Specifically, the S3M utilizes the Mamba-like block where the information from each token in \\(\\mathbf{f}_{\\mathrm{Mul}}\\) is compressed into SSM parameters—\\(\\overline{\\mathbf{A}}_{Mul}\\), \\(\\overline{\\mathbf{B}}_{Mul}\\), \\(\\mathbf{C}_{Mul}\\), which contain the compact multi-modal information along their hidden state dimension. Then, the Forward/Backward SSM module interact across both the channel dimensions of the multi-modal sequences and the state dimensions of the SSM parameters, representing the modality interaction between the origin format and the compact format, thereby, resulting in an efficient fusion that markedly improves segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.63, + 0.295, + 0.646 + ], + "angle": 0, + "content": "4.6. Global Input Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.822 + ], + "angle": 0, + "content": "We introduce a global input strategy that integrates the full volumes information into the analysis of cropped volumes without incurring excessive computational costs, as detailed in Fig. 2. The full volumes, augmented with a relative position mask, undergoes a series of transformations through convolutional layers and non-linear activations. They are then resized to match the dimensions of the cropped volumes using adaptive average pooling. The resulting feature map, which combines coarse global information with the fine details of the cropped volume, is then fed into our LS3M for segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The relative position mask, specifically designed for the full volumes, is encoded at positions corresponding to the cropped region. This design emphasizes the relevant area within the larger volume, guiding the network's attention to the regions slated for segmentation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.791, + 0.105 + ], + "angle": 0, + "content": "4.7. Network Architecture and Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.905, + 0.233 + ], + "angle": 0, + "content": "Our framework, depicted in Fig. 2, processes inputs of four modalities. Each modality includes both a cropped volume for segmentation and its corresponding global volume. The global volumes are first processed through our global input strategy. The yielded feature maps are then concatenated with their respective cropped volumes and passed through individual encoders \\( E(\\cdot)_M \\) to extract hierarchical intra-modal features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.235, + 0.905, + 0.31 + ], + "angle": 0, + "content": "The features from each modality are transformed into sequences and fed into \\(\\mathrm{SortP}_M\\), the learnable sorting process, which generates a modality-specific permutation matrix, \\(\\mathbf{P}_{\\mathbf{M}}\\). This matrix reorders the sequences before they are input into the S3M tailored for each modality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.905, + 0.431 + ], + "angle": 0, + "content": "At higher resolutions, for each modality, a larger permutation matrix, \\(\\mathbf{P}_{\\mathrm{Large}}\\), is obtained by bilinear upsampling. For multi-modal integration, the comprehensive multi-modal permutation matrix, \\(\\mathbf{P}_{\\mathrm{Mul}}\\), is generated by combining the modality-specific permutation matrix \\(\\mathbf{P}_{\\mathrm{M}}\\) with a multi-modal code \\(h\\). These permutation matrices play a crucial role in integrating the sequences within the S3M, ultimately enhancing segmentation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.905, + 0.505 + ], + "angle": 0, + "content": "The final segmentation results are obtained after hierarchical fusion and processing of the multi-modal features. Simulated missing modalities are represented by substituting zero-value feature maps prior to concatenating individual modal features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.905, + 0.553 + ], + "angle": 0, + "content": "Following [8, 55], we employ the weighted cross-entropy loss \\( Loss_{WCE} \\) and the Dice loss \\( Loss_{Dice} \\). The total loss is \\( Loss = Loss_{WCE} + Loss_{Dice} \\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.567, + 0.645, + 0.584 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.821, + 0.607 + ], + "angle": 0, + "content": "5.1. Dataset and implementation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.779 + ], + "angle": 0, + "content": "We perform our experiments on BraTS2018 and BraTS2020 from the Multi-modal Brain Tumor Segmentation Challenge [37], which align with that of previous studies [32, 45]. The two datasets contain 285 and 369 cases, respectively, with publicly available ground truth. Consistent with [8], BraTS2020 is randomly split into training, validation, and test sets with a ratio of \\(219:50:100\\). BraTS2018 is split into \\(199:29:57\\) and includes a three-fold validation. These splits are consistently used across all experiments, ensuring that all methods are evaluated under the same conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Each case of the datasets has four different modalities, i.e., FLAIR, T1ce, T1, and T2 modalities. These modalities are characterized by a volume size of \\(240 \\times 240 \\times 155\\), and they capture various properties of brain tumor subregions: GD-enhancing tumor (ET), peritumoral edema (ED), and the necrotic and non-enhancing tumor core (NCR/NET). These subregions of brain tumors are grouped into three nested subregions: the whole tumor (WT), the tumor core" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25987" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.1, + 0.835, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.198, + 0.33, + 0.798, + 0.344 + ], + "angle": 0, + "content": "Figure 4. Segmentation results of different methods with various available modalities on BraTS2020." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.37, + 0.483, + 0.414 + ], + "angle": 0, + "content": "(TC), and the enhancing tumor (ET). We use Dice coefficient, \\(95\\%\\) Hausdorff distance, sensitivity, and specificity to evaluate our method." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.415, + 0.484, + 0.552 + ], + "angle": 0, + "content": "We conduct our experiments using the PyTorch framework version 1.13.0. The training process is carried out on four NVIDIA A800 GPUs, spanning 500 epochs, with a batch size of four volumes. We utilize the Adam optimizer with an initial learning rate set to \\(2 \\times 10^{-4}\\) and a weight decay of \\(1 \\times 10^{-5}\\). During the training phase, the input images are randomly cropped to dimensions of \\(80 \\times 80 \\times 80\\). Data augmentation techniques include random rotations, flips, and intensity adjustments." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.561, + 0.421, + 0.577 + ], + "angle": 0, + "content": "5.2. Comparisons with the State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.583, + 0.483, + 0.702 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our method, we compare our method with nine state-of-the-art methods on different cases with missing modalities. The involved methods contain HeMiS [18], U-HVED [10], RFNet [8], UNet-MFI [59], mmFormer [55], MMANet [48], ShaSpec [45], M3AE [32], and GSS [39]. For a fair comparison, all methods are trained under their recommended hyper-parameters within the same dataset split." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.704, + 0.483, + 0.84 + ], + "angle": 0, + "content": "As illustrated in Tab. 1, our method achieves better performance for most combinations of missing modalities. We achieve improvements of \\(1.22\\%\\), \\(1.31\\%\\), and \\(2.3\\%\\) over the second-ranked method on the average DSC for WT, TC, and ET. We also provide a visualization comparison in Fig. 4, illustrating that our method yields more accurate segmentation results in different combinations of modalities. More comparisons, including statistical results, additional metrics and model complexity, are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.245, + 0.866 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We conduct ablation experiments on BraTS2020 to assess the effectiveness of specific designs. The baseline is served" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.371, + 0.871, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.522, + 0.846, + 0.537 + ], + "angle": 0, + "content": "Figure 5. Effect of hyper-parameters \\(\\tau\\) and \\(n\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.565, + 0.907, + 0.597 + ], + "angle": 0, + "content": "as a simplified version of LS3M without GIS, SortP, and S3M components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.74 + ], + "angle": 0, + "content": "Effect of Scanning Strategies. Relative to our baseline, which employs no specific scanning strategies, adding bidirectional, and fourth-directional (cross scan) methods demonstrated enhanced effectiveness, attributed to their ability to maintain directional spatial coherence. Notably, our sorted scanning strategy, as detailed in Tab. 2, achieved the highest Dice scores by preserving spatial inductive biases and semantic correlations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Effect of Proposed Components. As shown in Tab. 3, integrating a global input significantly improves baseline performance by providing larger spatial contexts. The addition of a relative position mask further enhances results by focusing on key areas within the broader input context. The average performance improves \\(1.84\\%\\) with S3M. The complete LS3M architecture, after including SortP, achieves superior performance, effectively exploring intra-modal and inter-modal relations." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25988" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.089, + 0.867, + 0.451 + ], + "angle": 0, + "content": "
MFLAIR●○○○●●●○○○●●●○●AVG
T1ce○●○○●○○●○○●●○●●AVG
T1○○●○○○○●○●●○●●●●
T2○○○●○○○●○●○●●●●●
WTHeMiS71.6067.7168.9668.1969.1768.6769.8369.0169.7869.4070.2171.2870.7371.5872.0669.88
U-HVED69.8546.8246.7754.0361.4558.2564.5062.9165.7664.2966.9969.7068.3870.3571.4162.76
RFNet86.4277.3476.4686.2189.5589.3089.3581.0087.4587.9590.3990.2090.4288.5990.7786.76
UNet-MFI82.2773.1872.1082.4583.6484.3484.8577.3083.4483.5285.4585.7085.8584.2084.9382.21
mmFormer82.4074.2574.3783.0784.5484.6185.8277.9884.0584.0085.3486.1186.2284.6486.3882.92
MMANet83.2164.4664.3582.8886.7487.6487.4770.9885.6485.3387.9289.3787.9786.3889.2582.64
ShaSpec84.1070.3670.2383.9588.6589.0788.7076.0286.2187.1589.4789.4390.0787.6290.1184.74
M3AE86.5373.8576.7186.0989.4889.3889.2578.1187.3787.2089.9990.1890.4288.6190.5686.25
GSS*86.3672.4573.3386.7689.5190.6590.9178.3888.8790.0591.1493.2590.6789.4393.2887.00
Ours88.6678.2479.3287.2991.0090.5791.4582.5889.1888.8891.5291.8791.4589.7592.1588.22
TCHeMiS53.4351.4151.5651.1151.7051.0851.8551.8852.3551.5152.9553.7652.9754.3855.0352.46
U-HVED34.6235.5127.3037.6742.1538.2643.4144.9347.5344.9749.1351.3049.4052.7254.1743.53
RFNet65.0482.3764.3168.4784.6971.4572.6283.1584.0672.1184.7184.7074.2884.1184.7477.39
UNet-MFI63.9477.6359.3868.0579.9268.2370.7277.6180.0970.2180.0380.9471.4080.7581.2874.01
mmFormer66.1977.9661.1769.1880.3669.5871.5579.9380.7970.9080.1881.3172.0281.1281.2274.90
MMANet65.7974.0557.4770.2185.2172.7372.6580.9985.2471.8885.8785.6874.4886.3986.6376.99
ShaSpec66.0677.3459.5768.2985.0272.8772.9580.8484.4672.2885.8385.5774.7485.4786.5077.19
M3AE68.0481.3966.0070.2782.0173.8274.9582.3983.0172.5482.4483.0675.0984.0684.4077.56
GSS*67.3378.4361.6770.1084.8873.4574.1782.2086.1274.1886.5088.3974.3386.2888.6778.45
Ours71.9883.6166.6772.3086.3275.2576.8784.0985.1874.3286.0585.8476.8785.8385.8679.76
ETHeMiS43.7742.4141.5941.4541.8340.2941.1942.0842.3941.0043.6744.1642.9545.2746.3342.69
U-HVED12.8824.947.2724.2630.0221.9529.4033.6436.1832.1239.3940.9138.0943.1845.3330.64
RFNet40.4774.2737.5143.5976.4543.8146.9975.2273.9446.3777.0176.3848.9576.3876.6460.93
UNet-MFI39.7069.4229.3846.0070.1340.0648.6969.2572.3245.7171.2870.8846.5572.0071.4157.52
mmFormer40.4768.9133.9745.6169.8143.6348.0971.1070.7245.9270.0871.6048.3870.6571.3658.02
MMANet36.4066.0728.2241.9170.4341.5044.2368.8771.8643.5871.9471.9844.6071.8172.4356.39
ShaSpec38.9266.9331.7542.5870.8543.2245.6269.9271.0444.2172.0871.0946.2271.8172.0257.22
M3AE40.4972.4339.9345.9774.6643.2047.3075.4276.8146.6375.9477.0848.1977.4078.0061.30
GSS*42.0369.4635.3045.6074.2047.7849.3374.4174.7848.9476.1276.3250.0276.9875.9561.15
Ours40.3477.6836.5745.3079.1745.0751.5379.4978.6947.0480.7580.5151.5381.5380.9463.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.461, + 0.904, + 0.49 + ], + "angle": 0, + "content": "Table 1. Performance comparison (DSC%) with SOTA methods on BraTS2020. Available and missing modalities are denoted by \\(\\bullet\\) and \\(\\circ\\), respectively. \\* demonstrates that its codes and results are reproduced by ourselves." + }, + { + "type": "table", + "bbox": [ + 0.103, + 0.513, + 0.472, + 0.557 + ], + "angle": 0, + "content": "
DiceBaseline+ Uni-D+ Bi-D+ Fourth-D+ Random+ Our sorted
WT85.4385.2085.5586.1384.2386.94
TC76.1775.6976.6777.5474.0878.01
ET58.0257.9858.4360.7156.7962.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.595 + ], + "angle": 0, + "content": "Table 2. Effect of scanning strategies, '-D' denotes directional scanning." + }, + { + "type": "table", + "bbox": [ + 0.103, + 0.611, + 0.472, + 0.653 + ], + "angle": 0, + "content": "
DiceBaseline+GIS (w/o mask)+GIS+GIS + S3M+GIS + SortP + S3M
WT85.4385.9986.2387.1788.22
TC76.1776.7877.0679.0679.76
ET58.0259.0160.3562.9463.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.692 + ], + "angle": 0, + "content": "Table 3. Effect of proposed components, 'GIS' denotes Global Input Strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Effects of hyper-parameters of temperature \\(\\tau\\) and iterative times \\(n\\). Adjusting \\(\\tau\\) and \\(n\\) reveals their critical roles in influencing the performance of our LS3M, as shown in Fig. 5. The parameter \\(\\tau\\) modulates the smoothness of the permutation matrix, with \\(\\tau = 1\\) yielding optimal results when \\(n\\) is fixed. Conversely, increasing \\(n\\) enhances the precision of the permutation matrix, thus improving segmentation outcomes. Balancing computational efficiency and precision, we select \\(n = 10\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Additional ablation experiments are provided in the appendix. These experiments further validate the importance of the proposed components and their contribution to the" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.516, + 0.661, + 0.53 + ], + "angle": 0, + "content": "model's performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.567, + 0.634, + 0.582 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduce the Learnable Sorting State Space Model (LS3M), a novel framework for brain tumor segmentation with missing modalities, which aims to maximizing mining information from available modalities. LS3M replaces traditional non-differentiable sorting with a differentiable permutation matrix, preserving 3D spatial information and enabling more effective modeling. The framework adeptly scales these matrices to accommodate longer sequences, capturing multi-scale information. It also dynamically integrates these matrices for robust multi-modal sequence fusion, effectively managing various scenarios of missing modalities. The S3M block captures both intra-modal and inter-modal correlations, enhancing segmentation accuracy. With a global input strategy, LS3M provides comprehensive contextual insights, surpassing existing segmentation methods. Experimental results on BraTS2018 and BraTS2020 demonstrate that LS3M outperforms state-of-the-art. Our results highlight LS3M's potential in advancing medical imaging and healthcare diagnostics." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25989" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.187 + ], + "angle": 0, + "content": "This work was in part supported by the National Natural Science Foundation of China under grants 62032006 and 62021001, and the Anhui Province Key Laboratory of Biomedical Imaging and Intelligent Processing under grants 23YGXT001." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.216, + 0.188, + 0.232 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.242, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[1] Ryan Prescott Adams and Richard S Zemel. Ranking via sinkhorn propagation. arXiv preprint arXiv:1106.1925, 2011.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.285, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[2] Reza Azad, Nika Khosravi, and Dorit Merhof. Smu-net: Style matching u-net for brain tumor segmentation with missing modalities. In International Conference on Medical Imaging with Deep Learning, pages 48-62. PMLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.357, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[3] Ali Behrouz and Farnoosh Hashemi. Graph mamba: Towards learning on graphs with state space models. arXiv preprint arXiv:2402.08678, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.401, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[4] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In Computer Vision-ECCV 2022 Workshops, pages 205-218. Springer, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.472, + 0.484, + 0.541 + ], + "angle": 0, + "content": "[5] Cheng Chen, Qi Dou, Yueming Jin, Hao Chen, Jing Qin, and Pheng-Ann Heng. Robust multimodal brain tumor segmentation via feature disentanglement and gated fusion. In Medical Image Computing and Computer-Assisted Intervention-MICCAI, pages 447-456. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.543, + 0.484, + 0.611 + ], + "angle": 0, + "content": "[6] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.615, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[7] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[8] Yuhang Ding, Xin Yu, and Yi Yang. Rfnet: Region-aware fusion network for incomplete multi-modal brain tumor segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3975-3984, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.484, + 0.785 + ], + "angle": 0, + "content": "[9] Yuhang Ding, Liulei Li, Wenguan Wang, and Yi Yang. Clustering propagation for universal medical image segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[10] Reuben Dorent, Samuel Joutard, Marc Modat, Sébastien Ourselin, and Tom Vercauteren. Hetero-modal variational encoder-decoder for joint modality completion and segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI, pages 74-82. Springer, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Zhengcong Fei, Mingyuan Fan, Changqian Yu, and Junshi Huang. Scalable diffusion models with state space backbone. arXiv preprint arXiv:2402.05608, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.242, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[12] Jiannan Ge, Lingxi Xie, Hongtao Xie, Pandeng Li, Xiaopeng Zhang, Yongdong Zhang, and Qi Tian. Alignzeg: Mitigating objective misalignment for zero-shot semantic segmentation. In European Conference on Computer Vision, pages 142-161. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.203 + ], + "angle": 0, + "content": "[13] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[14] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[15] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[16] Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In *Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: 7th International Workshop, BrainLes* 2021, Held in Conjunction with MICCAI 2021, Virtual Event, pages 272-284. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.906, + 0.481 + ], + "angle": 0, + "content": "[17] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 574-584, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[18] Mohammad Havaei, Nicolas Guizard, Nicolas Chapados, and Yoshua Bengio. Hemis: Hetero-modal image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI, pages 469-477. Springer, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.906, + 0.593 + ], + "angle": 0, + "content": "[19] Xuanhua He, Ke Cao, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Pan-mamba: Effective pan-sharpening with state space model. arXiv preprint arXiv:2402.12192, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.906, + 0.621 + ], + "angle": 0, + "content": "[20] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.906, + 0.664 + ], + "angle": 0, + "content": "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.906, + 0.718 + ], + "angle": 0, + "content": "[22] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.906, + 0.762 + ], + "angle": 0, + "content": "[23] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In International Conference on Learning Representations, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[24] Zeyu Jiang, Changxing Ding, Minfeng Liu, and Dacheng Tao. Two-stage cascaded u-net: 1st place solution to brats challenge 2019 segmentation task. In *Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries*, pages 231-241. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Sanaz Karimijafarbigloo, Reza Azad, Amirhossein Kazerouni, Saeed Ebadollahi, and Dorit Merhof. Mmcformer: Missing modality compensation transformer for brain tumor segmentation. In Medical Imaging with Deep Learning, pages 1144-1162. PMLR, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25990" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[26] Jonghun Kim and Hyunjin Park. Adaptive latent diffusion model for 3d medical image to image translation: Multimodal magnetic resonance imaging study. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7604-7613, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[27] Dongwook Lee, Won-Jin Moon, and Jong Chul Ye. Assessing the importance of magnetic resonance contrasts using collaborative generative adversarial networks. Nature Machine Intelligence, 2(1):34-42, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[28] Ho Hin Lee, Shunxing Bao, Yuankai Huo, and Bennett A Landman. 3d ux-net: A large kernel volumetric convnet modernizing hierarchical transformer for medical image segmentation. arXiv preprint arXiv:2209.15076, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[29] Jiangyun Li, Wenxuan Wang, Chen Chen, Tianxiang Zhang, Sen Zha, Hong Yu, and Jing Wang. Transbtsv2: Wider instead of deeper transformer for medical image segmentation. arXiv preprint arXiv:2201.12785, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[30] Mingcheng Li, Dingkang Yang, Xiao Zhao, Shuaibing Wang, Yan Wang, Kun Yang, Mingyang Sun, Dongliang Kou, Ziyun Qian, and Lihua Zhang. Correlation-decoupled knowledge distillation for multimodal sentiment analysis with incomplete modalities. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[31] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[32] Hong Liu, Dong Wei, Donghuan Lu, Jinghan Sun, Liansheng Wang, and Yefeng Zheng. M3ae: Multimodal representation learning for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1657-1665, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.483, + 0.6 + ], + "angle": 0, + "content": "[33] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[34] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[35] Andriy Marusyk, Vanessa Almendro, and Kornelia Polyak. Intra-tumour heterogeneity: a looking glass for cancer? Nature reviews cancer, 12(5):323-334, 2012. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[36] Gonzalo Mena, David Belanger, Scott Linderman, and Jasper Snoek. Learning latent permutations with gumbel-sinkhorn networks. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[37] Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging, 34(10):1993-2024, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[38] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[39] Yansheng Qiu, Delin Chen, Hongdou Yao, Yongchao Xu, and Zheng Wang. Scratch each other's back: Incomplete" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "multi-modal brain tumor segmentation via category aware group self-support learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21317-21326, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[40] Santiago Ramón y Cajal, Marta Sesé, Claudia Capdevila, Trond Aasen, Leticia De Mattos-Arruda, Salvador J Diaz-Cano, Javier Hernández-Losa, and Josep Castellvi. Clinical implications of intratumor heterogeneity: challenges and opportunities. Journal of Molecular Medicine, 98:161-177, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[41] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[42] Dong She, Yueyi Zhang, Zheyu Zhang, Hebei Li, Zihan Yan, and Xiaoyan Sun. Eoformer: Edge-oriented transformer for brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 333-343. Springer, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[43] Jimmy TH Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[44] Chloe Wang, Oleksii Tsepa, Jun Ma, and Bo Wang. Graphmamba: Towards long-range graph sequence modeling with selective state spaces. arXiv preprint arXiv:2402.00789, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[45] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15878-15887, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[46] Wenxuan Wang, Chen Chen, Meng Ding, Hong Yu, Sen Zha, and Jiangyun Li. Transbts: Multimodal brain tumor segmentation using transformer. In Medical Image Computing and Computer Assisted Intervention-MICCAI, pages 109-119. Springer, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[47] Yixin Wang, Yang Zhang, Yang Liu, Zihao Lin, Jiang Tian, Cheng Zhong, Zhongchao Shi, Jianping Fan, and Zhiqiang He. Acn: adversarial co-training network for brain tumor segmentation with missing modalities. In Medical Image Computing and Computer Assisted Intervention-MICCAI, pages 410-420. Springer, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[48] Shicai Wei, Chunbo Luo, and Yang Luo. Mmanet: Margin-aware distillation and modality-aware regularization for incomplete multimodal learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20039-20049, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[49] Jing Nathan Yan, Jiatao Gu, and Alexander M Rush. Diffusion models without attention. arXiv preprint arXiv:2311.18257, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[50] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. arXiv preprint arXiv:2403.17695, 2024. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Heran Yang, Jian Sun, and Zongben Xu. Learning unified hyper-network for multi-modal mr image synthesis and tu" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25991" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "mor segmentation with missing modalities. IEEE Transactions on Medical Imaging, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[52] Qiushi Yang, Xiaoqing Guo, Zhen Chen, Peter YM Woo, and Yixuan Yuan. D 2-net: Dual disentanglement network for brain tumor segmentation with missing modalities. IEEE Transactions on Medical Imaging, 41(10):2953-2964, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[53] Chenyu You, Ruihan Zhao, Fenglin Liu, Siyuan Dong, Sandeep Chinchali, Ufuk Topcu, Lawrence Staib, and James Duncan. Class-aware adversarial transformers for medical image segmentation. pages 29582-29596, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[54] Ziqi Yu, Xiaoyang Han, Shengjie Zhang, Jianfeng Feng, Tingying Peng, and Xiao-Yong Zhang. Mousegan++: Unsupervised disentanglement and contrastive representation for multiple mri modalities synthesis and structural segmentation of mouse brain. IEEE Transactions on Medical Imaging, 2022, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[55] Yao Zhang, Nanjun He, Jiawei Yang, Yuexiang Li, Dong Wei, Yawen Huang, Yang Zhang, Zhiqiang He, and Yefeng Zheng. mmformer: Multimodal medical transformer for incomplete multimodal learning of brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 107-117. Springer, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[56] Zheyu Zhang, Xinzhao Liu, Zheng Chen, Yueyi Zhang, Huanjing Yue, Yunwei Ou, and Xiaoyan Sun. Anatomical consistency distillation and inconsistency synthesis for brain tumor segmentation with missing modalities. arXiv preprint arXiv:2408.13733, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[57] Zheyu Zhang, Gang Yang, Yueyi Zhang, Huanjing Yue, Aiping Liu, Yunwei Ou, Jian Gong, and Xiaoyan Sun. Tmformer: Token merging transformer for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7414-7422, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[58] Sijie Zhao, Hao Chen, Xueliang Zhang, Pengfeng Xiao, Lei Bai, and Wanli Ouyang. Rs-mamba for large remote sensing image dense prediction. arXiv preprint arXiv:2404.02668, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[59] Zechen Zhao, Heran Yang, and Jian Sun. Modality-adaptive feature interaction for brain tumor segmentation with missing modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 183-192. Springer, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[60] Zhuoran Zheng and Jun Zhang. Fd-vision mamba for endoscopic exposure correction. arXiv preprint arXiv:2402.06378, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[61] Chenhong Zhou, Changxing Ding, Xinchao Wang, Zhentai Lu, and Dacheng Tao. One-pass multi-task networks with cross-task guided attention for brain tumor segmentation. IEEE Transactions on Image Processing, 29:4516-4529, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[62] Tao Zhou, Huazhu Fu, Geng Chen, Jianbing Shen, and Ling Shao. Hi-net: hybrid-fusion network for multi-modal mr image synthesis. IEEE Transactions on Medical Imaging, 39 (9):2772-2781, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[63] Tongxue Zhou, Stéphane Canu, Pierre Vera, and Su Ruan. Latent correlation representation learning for brain tumor" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "segmentation with missing mri modalities. IEEE Transactions on Image Processing, 30:4263-4274, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[64] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: Redesigning skip connections to exploit multiscale features in image segmentation. IEEE Transactions on Medical Imaging, 39(6):1856-1867, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[65] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. 2024. 1, 3, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25992" + } + ] +] \ No newline at end of file diff --git a/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/58ab88d4-b760-40c9-8213-6c50aa676cf3_origin.pdf b/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/58ab88d4-b760-40c9-8213-6c50aa676cf3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c85417c4ae651495420535fa92c9ecb71f74f19d --- /dev/null +++ b/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/58ab88d4-b760-40c9-8213-6c50aa676cf3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d62737b99fa79c49a5cbfe743a9cda2e5ee8915ce5b8aadd9442169c52de1a54 +size 4196285 diff --git a/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/full.md b/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c95dadf7a3bf4be9b9a191c5359c405c4f85354a --- /dev/null +++ b/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/full.md @@ -0,0 +1,364 @@ +# Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model + +Zheyu Zhang $^{1,2}$ Yayuan Lu $^{1,2}$ Feipeng Ma $^{1,2}$ Yueyi Zhang $^{1,2}$ Huanjing Yue $^{3}$ Xiaoyan Sun $^{1,2,*}$ $^{1}$ MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, 230026, China + $^{2}$ Anhui Province Key Laboratory of Biomedical Imaging and Intelligent Processing, Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, Hefei 230088, China + $^{3}$ School of Electrical and Information Engineering, Tianjin University, Tianjin, 300072, China +zhangzy0@mail.ustc.edu.cn sunxiaoyan@ustc.edu.cn + +# Abstract + +Brain tumor segmentation plays a crucial role in clinical diagnosis, yet the frequent unavailability of certain MRI modalities poses a significant challenge. In this paper, we introduce the Learnable Sorting State Space Model (LS3M), a novel framework designed to maximize the utilization of available modalities for brain tumor segmentation. LS3M excels at efficiently modeling long-range dependencies based on the Mamba design, while incorporating differentiable permutation matrices that reorder input sequences based on modality-specific characteristics. This dynamic reordering ensures that critical spatial inductive biases and long-range semantic correlations inherent in 3D brain MRI are preserved, which is crucial for incomplete multi-modal brain tumor segmentation. Once the input sequences are reordered using the generated permutation matrix, the Series State Space Model (S3M) block models the relationships between them, capturing both local and long-range dependencies. This enables effective representation of intra-modal and inter-modal relationships, significantly improving segmentation accuracy. Extensive experiments on the BraTS2018 and BraTS2020 datasets demonstrate that LS3M outperforms existing methods, offering a robust solution for brain tumor segmentation, particularly in scenarios with missing modalities. + +# 1. Introduction + +Accurate brain tumor segmentation in clinical settings relies on the integration of various MRI modalities (FLAIR, T1ce, T1, T2) that provide complementary information for diagnosis. However, the frequent unavailability of cer- + +tain modalities, like T1ce due to contrast agent allergies, presents a challenge. Existing methods addressing missing modalities fall into two categories: (1) bridging the data gap between complete and missing modalities through modality synthesis [26, 27, 32, 51, 54, 62] and knowledge distillation [2, 39, 47, 48] and (2) optimizing architectures to maximize information extraction from available modalities, which is the focus of our work. Some methods [8, 63] adopt multi-scale convolutions to capture critical structural details by focusing on spatial inductive bias, such as local texture. However, the limited receptive field restricts their ability to represent tumor tissue effectively, especially when other areas may appear visually similar in localized views, making it difficult to differentiate tumor tissue [35, 40]. Some methods explore models such as Transformers [55] and graph neural networks [51, 59] to build long-range semantic correlation to mitigate the mentioned problem, but they suffer from quadratic computational complexity. The computational inefficiency is further exacerbated by the large sequence lengths typical of 3D brain MRI. + +Recent advancements in selective state space models (SSMs), such as Mamba [7, 13], offer the potential to obtain long-range representations for available modalities with dynamic weighting parameters in an efficient way. Mamba flattens the 3D brain MRI into a 1D sequence and iteratively scans this sequence to compress hidden states, which are then used to enhance the current token by combining it with previously scanned tokens. While Mamba can efficiently build long-range dependencies, the influence of distant tokens on the current token diminishes significantly. This becomes problematic for 3D brain MRI, as flattening the sequence destroys the 3D spatial structure. Follow-up methods [33, 50, 58, 65] propose bi-directional, four-directional, and omni-directional scanning to mitigate distortion to spatial structures, as shown in Fig. 1. However, these methods + +either neglect local inductive bias or fail to capture long-range semantic correlations of 3D brain MRI, hindering their ability to fully capture contextual relationships essential for brain tumor segmentation, particularly in scenarios with missing modalities. + +To address these challenges, we propose the Learnable Sorting State Space Model (LS3M), specifically designed for brain tumor segmentation with missing modalities. Firstly, LS3M incorporates the Sinkhorn operator with the Gumbel-Softmax reparameterization, which converts the traditional non-differentiable sorting into a differentiable permutation matrix. This matrix is used to reorder the 1D sequence of available modalities, considering spatial inductive biases and long-range semantic correlations. Importantly, the permutation matrix does not directly establish relationships; rather, it prepares the sequence for further processing by the Series State Space Model (S3M), which is capable of modeling intra-modal and inter-modal relationships by leveraging both local and long-range correlations. Next, LS3M scales the intra-modal permutation matrix at larger resolutions, allowing for the extraction of multi-scale information. These modality-specific permutation matrices are also combined for multi-modal fusion to capture correlations across modalities. The reordered sequences, now enriched with spatial and semantic information, are then processed by the S3M block, which builds relationships at both intra-modal and inter-modal levels, maximizing the extraction of information from the available modalities. Lastly, LS3M departs from existing models that rely on cropped input volumes by adopting a global input strategy, supplemented with a relative position mask. This approach not only provides richer contextual information but also informs the model about the location of the cropped volume within the larger global context, significantly enhancing spatial awareness and diagnostic accuracy. In summary, our key contributions are as follows: + +- We propose LS3M for brain tumor segmentation with missing modalities. LS3M improves SSM by dynamically reordering data, capturing patient-specific characteristics, and enhancing the ability to extract local and long-range correlations from available modalities. +- LS3M scales and integrates the permutation matrix to capture multi-scale and multi-modal information. The reordered sequences are then processed by the S3M block to model complex intra-modal and inter-modal relationships, improving segmentation accuracy. +- LS3M incorporates a global input strategy with relative position embeddings, significantly enhancing contextual awareness towards comprehensive utilization of available modalities. +- Extensive evaluations on BraTS2018 and BraTS2020 demonstrate the superiority and effectiveness of LS3M for brain tumor segmentation, particularly in clinically + +![](images/226b16430542e9eda8c561eef97e71061ea237dd5a3199b0540dc10da8b1fc52.jpg) +Long-range Semantic Correlation +(a) Requirements for identifying brain tumor + +![](images/c5c93f5ad2c5433c5c6b438dd2c15828fc95d70ef809b7cd509095b79163af68.jpg) +(b) Horizontal Scanning + +![](images/0b5e33c4b176ad89c13e443450f34cab9c2344e525547128ae9867a0bfbd5fc0.jpg) +(d) Local Scanning +Figure 1. (a) Brain tumor segmentation requires local inductive bias to capture small but critical structural details, as well as long-range semantic correlations to understand spatial relationships between the tumor and surrounding structures. (b)-(e) Different scanning strategies. + +![](images/56b3633d1f10860e505a4efae85e5e7c96f13c06306be2496aeb275343e432c3.jpg) +(c) Vertical Scanning + +![](images/57487da9e89fdbb6754875089c890ead9ca48d169035c5e7e1ff80fc17f5fd9a.jpg) +(e) Our Sorted Scanning + +![](images/789d2c2627058b7a85e096435041360b9d66210021dc5559f3556af0cb939e8b.jpg) + +![](images/8fc4edbd42798f53e0efb5f24a8a81e4d12f3014148c6a611cc750d3b73ea556.jpg) + +relevant scenarios with missing modalities. + +# 2. Related Work + +Medical Image Segmentation. Medical image segmentation traditionally relies on U-Net-like structures, which formulate the task as dense pixel classification. Enhancements to segmentation structures include cascaded networks [24, 61], refined skip connections [64], hierarchical pyramids [9], attention mechanisms [4, 6, 12, 16, 17, 29, 42, 46, 53], and large depth-wise kernels [28]. Despite achieving notable performance, these methods often fail when certain modalities are absent. + +Multi-modal Learning for Missing Modalities. Research in multi-modal learning with missing modalities generally seeks to either bridge the gap between scenarios with complete and missing modalities or to optimize architectural design to maximize the use of available modal information. The former approach includes modality synthesis [26, 27, 32, 51, 54, 62] to generate absent modalities and knowledge distillation techniques [2, 30, 39, 47, 48] that transfer insights from complete to incomplete scenarios. The latter approach utilizes techniques like modality disentanglement [5, 52] to separate modality-specific information in the frequency domain and modality fusion strategies involving hierarchical pyramids [8, 63] and attention mechanisms [25, 55-57]. However, combining attention mechanisms with other techniques often results in significant computational overhead, which can impede effective modeling long-range dependencies. In our work, we leverage SSM, specifically the Mamba, to efficiently build long-range correlations from all available modalities. + +State Space Model. Drawing inspiration from control theory, the State Space Model (SSM) recurrently updates sequences through latent states [14, 43], as demonstrated by the Structured State Space Sequence model (S4), which utilizes linear state space equations to efficiently scale with sequence length. Recently, Gu and Dao [13] propose a + +generic modal, called Mamba, with selective projection, gated mechanism and efficient hardware design, outperforming Transformers in linear-time inference on natural language processing. This success has spurred adaptations of Mamba in computer vision. Since visual tokens lack a sequential arrangement, methods propose diverse scanning strategies like bi-/four-/omni-directional [33, 50, 58, 65], local [22], and atrous scanning [38], enhancing contextual data integration for tasks such as medical image segmentation [34, 41, 41], image restoration [15, 19, 60], point cloud analysis [31], image generation [11, 49], and graph neural networks [3, 44]. + +# 3. Preliminaries + +State Space Model (SSM) for Sequence Modeling. The SSM defines a linear Ordinary Differential Equation (ODE) to map an input sequence $x(t) \in \mathbb{R}$ to an output sequence $y(t) \in \mathbb{R}$ through a hidden state $h(t) \in \mathbb{R}^N$ , formulated as: + +$$ +h ^ {\prime} (t) = \mathbf {A} h (t) + \mathbf {B} x (t), \quad y (t) = \mathbf {C} h (t), \tag {1} +$$ + +where $\mathbf{A} \in \mathbb{R}^{N \times N}$ , $\mathbf{B} \in \mathbb{R}^{N \times 1}$ , and $\mathbf{C} \in \mathbb{R}^{1 \times N}$ are learned parameters. Since images are processed into discrete token sequences, the discrete version of Eq. (1) is defined as follows: + +$$ +h _ {k} = \overline {{\mathbf {A}}} h _ {k - 1} + \overline {{\mathbf {B}}} x _ {k}, \quad y _ {k} = \mathbf {C} h _ {k}, \tag {2} +$$ + +where $\overline{\mathbf{A}}$ and $\overline{\mathbf{B}}$ are discretized using the discretization step size $\Delta$ in the zero-order hold (ZOH) rule. The long-range dependencies are constructed via recurrent scanning. However, the recurrent SSM is not practical. Eq. (2) is equivalent to a global convolution kernel $\overline{\mathbf{K}}\in \mathbb{R}^{L}$ , i.e., + +$$ +\overline {{\mathbf {K}}} = \left(\mathbf {C} \overline {{\mathbf {B}}}, \mathbf {C} \overline {{\mathbf {A B}}}, \dots , \mathbf {C} \overline {{\mathbf {A}}} ^ {\mathbf {L} - 1} \overline {{\mathbf {B}}}\right), \quad \mathbf {y} = \mathbf {x} * \overline {{\mathbf {K}}}. \tag {3} +$$ + +Structured state space sequence models (S4) imposes a structure on the state matrix $\mathbf{A}$ based on HIPPO matrices, which significantly improves the efficiency for obtaining the SSM kernel $\overline{\mathbf{K}}$ . Further, Mamba [13] adds selective projection via making $\mathbf{B}, \mathbf{C}, \Delta$ input-dependent. + +Sinkhorn Operator. The Sinkhorn operator $S(\cdot)$ is an iterative procedure used to transform a non-negative matrix $\mathbf{P} \in \mathbb{R}^{L \times L}$ into a doubly stochastic matrix, which is a square matrix with non-negative entries such that each row and each column sums to one [1, 36]. The process can be formulated as: + +$$ +\begin{array}{l}S ^ {0} (\mathbf {P}) = \exp (\mathbf {P}), \quad S ^ {n} (\mathbf {P}) = \mathcal {T} _ {c} \left(\mathcal {T} _ {r} \left(S ^ {n - 1} (\mathbf {P})\right)\right),\\S (\mathbf {P}) = \lim _ {n \rightarrow \infty} S ^ {n} (\mathbf {P}),\end{array}\tag {4} +$$ + +where $n$ is iterative times. $\mathcal{T}_r(\cdot)$ and $\mathcal{T}_c(\cdot)$ are the row-wise and column-wise normalizations, defined as follows: + +$$ +\mathcal {T} _ {r} (\mathbf {P}) = \mathbf {P} \oslash \left(\mathbf {P} \mathbf {1} _ {\mathbf {N}} \mathbf {1} _ {\mathbf {N}} ^ {\top}\right), \mathcal {T} _ {c} (\mathbf {P}) = \mathbf {P} \oslash \left(\mathbf {1} _ {\mathbf {N}} \mathbf {1} _ {\mathbf {N}} ^ {\top} \mathbf {P}\right), \tag {5} +$$ + +where $\oslash$ denotes the element-wise division and $1_{N}$ is a column vector of ones. We employ Sinkhorn $(\mathbf{P},n)$ to represent Eq. (4). + +Gumbel-Softmax Reparameterization. The Gumbel-Softmax provides a differentiable approximation of the discrete categorical distribution [23]. The process of sample $z$ from a categorical distribution with class probabilities $x$ can be defined as follows: + +$$ +z = \operatorname {o n e - h o t} (\underset {i} {\arg \max } [ g _ {i} + \log x _ {i} ]), \tag {6} +$$ + +where $g_{i}$ is the injected standard i.i.d Gumbel noise. Jang et al. [23] use the softmax function as a continuous, differentiable approximation to the arg max function, i.e., + +$$ +\operatorname {s o f t m a x} (y / \tau) \equiv \frac {\exp \left(y _ {i} / \tau\right)}{\sum_ {i = 1} ^ {k} \exp \left(y _ {i} / \tau\right)}, \tag {7} +$$ + +where $y_{i} = \log x_{i} + g_{i}$ . When temperature $\tau \rightarrow 0$ , the softmax distribution converges to a one-hot representation identical to samples from a categorical distribution. We use Gumbel-Softmax(X, $\tau$ ) to represent Eq. (7), where $\mathbf{X} = [x_0, x_1, \dots, x_{|\mathbf{X}| - 1}]$ . + +# 4. Method + +In this section, we first explain the motivation and challenges with Mamba. We then introduce the learnable sorting process to address these challenges, followed by expanding the permutation matrix for longer and multi-modal sequences. Next, we describe the S3M block and the global input strategy. Finally, we present the LS3M architecture and its optimization loss. + +# 4.1. Motivations + +Brain tumor segmentation requires complementary information from multiple modalities. Missing modalities lead to insufficient information, inspiring us to maximize information extraction from available modalities. We try to do this from two aspects: building long-range dependencies and incorporating broader context for available modalities. + +First, brain tumors often exhibit complex anatomy, where local inductive biases alone may be insufficient to identify tumor types and boundaries. Long-range dependencies are essential for understanding spatial relationships between the tumor and surrounding structures. While Transformers can model such dependencies, their quadratic computational complexity makes them impractical for 3D MRI data due to large token sequences and high computational costs, as illustrated in the Appendix. Mamba, however, captures long-range dependencies more efficiently, making it a better fit for this task. + +Second, most segmentation methods focus on cropped volumes to reduce computational cost, neglecting the + +![](images/a4fd3450080f08915800b6a0a20969c752829cb4af17862d5781c2d2c9531aab.jpg) +(a) Our LS3M framework + +![](images/f1232b9edc89770ad4783313e3bb5aa6365e714acce722389cc2bfb575da4a05.jpg) +(b) $\mathrm{SortP_M}$ +(c) $\mathrm{SortP_L}$ +(d) $\mathrm{SortP_{Mul}}$ + +![](images/5313ba132fb3d5d77cfeb1dd084eb9ff3473d6972524f05dfe8740dc4d12ffe9.jpg) +Figure 2. (a) The comprehensive LS3M framework integrates the global input strategy with adaptive pooling, the SortP for generating permutation matrices, and the S3M for feature enhancement. $(\cdot)_M$ denotes that we perform $(\cdot)$ process separately for each modality. (b) SortP $M$ means the learnable sorting process applied to each modality. (c) SortP $L$ denotes obtaining the larger permutation matrix for longer sequences. (d) SortP $Mul$ refers to the generation of the multi-modal permutation matrix. +Figure 3. The structure of the proposed S3M block. + +broader context crucial for accurately understanding complex conditions like brain tumors. This is especially important in the case of intra-tumor heterogeneity, which requires a wider context to characterize the tumor type [35, 40]. Moreover, in scenarios with missing modalities, it is critical to utilize all available data from the entire image to compensate for the missing information. Thus, incorporating broader context without significant computational overhead is essential. + +# 4.2. Challenges of Existing Mamba Approaches + +Mamba is primarily designed for 1D sequences, with characteristics that include: + +Mamba Characteristic 1: the current output token $y_{k}$ is limited to modeling relationships solely with previously scanned tokens $x_{i}, i \in \{0,1,\dots ,k - 1,k\}$ , as detailed in Eq. (2). + +Mamba Characteristic 2: the influence of distant tokens diminishes significantly due to the weight $\overline{\mathbf{A}}^{i - 1}$ , as explained in Eq. (3). + +To mitigate these limitations, existing Mamba-based methods [33, 38, 50, 65] introduce forward and backward scanning strategies, which enable the current token to model relationships with both preceding and subsequent tokens. These methods are more suitable for vision applications, where local dependencies can often provide sufficient context. However, in the case of missing modalities, it is insufficient to rely solely on local context for tumor identi + +fication. Tumor heterogeneity often requires capturing information from distant regions of the MRI to fully understand the tumor's characteristics, as highlighted by prior studies [35, 40]. For example, in the segmentation of the whole tumor (WT) with only T1ce modality, local features may be insufficient, and distant correlations are necessary to improve segmentation accuracy. + +Due to Mamba Characteristic 2, the influence of distant tokens diminishes over sequence steps, which is problematic in 3D MRI segmentation tasks. Therefore, a dynamic mechanism is required to ensure that the most relevant tokens—both spatially close and semantically related over long distances—are positioned closer to the target token, facilitating more effective modeling. + +Although LocalMamba [22] uses a Z-scan to ensure local relationships in 2D, this approach cannot guarantee the preservation of 3D neighborhood relationships, such as the 27-neighborhood in 3D space. Even with axial scanning added to horizontal and vertical scanning, the relationships between distant tokens in 3D space are still not adequately preserved. Consequently, there is a need for a solution that dynamically reorders the tokens for each 3D MRI scan, ensuring that all potentially relevant tokens—whether spatially close or far but semantically related—are placed before the current token. This dynamic reordering will allow the subsequent S3M block to model these relationships more effectively, improving the segmentation accuracy. + +# 4.3. Learnable Sorting Process + +Traditional Non-differentiable Sorting Process. A pivotal aspect of LS3M involves addressing the non-differentiability of traditional sorting mechanisms. The patient-specific dynamic sorting process is initially driven by a predictive score $P \in \mathbb{R}^{L}$ , assigned to each token in the sequence $\mathbf{X}$ . These scores dictate the token positions in the sorted sequence $\mathbf{X}_{\mathbf{S}}$ , subsequently processed by the + +Mamba block. However, the traditional sorting function, which maps these continuous input scores to discrete token indices, interrupts the gradient flow necessary for end-to-end training of the network. + +The Learnable Sorting Process. The learnable sorting process is represented as: + +$$ +\hat {\mathbf {P}} = \operatorname {S o r t N e t} (\mathbf {f}), +$$ + +$$ +\mathbf {P} = \text {S i n k h o r n} (\text {G u m b e l - S o f t m a x} (\hat {\mathbf {P}}, \tau), n), \tag {8} +$$ + +$$ +\mathbf {X} = \text {f l a t t e n} (\mathbf {f}), \quad \mathbf {X} _ {\mathbf {S}} = \mathbf {P X}. +$$ + +This enable end-to-end differentiable sorting. As shown in Fig. 2(b), firstly, our SortNet generates an estimated matrix $\hat{\mathbf{P}}\in \mathbb{R}^{L\times L}$ tailored to individual patients. It employs a convolutional layer with a kernel size of $C\times C\times 3\times 3\times 3$ to capture spatial inductive biases, extracting information from the 26-neighborhood 3D spatial structure. After convolution, a SiLU activation function [20] is applied to enhance the network's representational capacity. The resulting feature map is split into $\mathbf{f}_1$ and $\mathbf{f}_2$ , with $\mathbf{f}_2$ being processed to compute a similarity map that captures long-range semantic correlations. A ReLU activation is then applied to ensure the non-negativity of $\hat{\mathbf{P}}$ , preparing it for subsequent processing by the Sinkhorn operator. + +Secondly, we use the Gumbel-Softmax and Sinkhorn operators to generate a permutation matrix $\mathbf{P}$ , where each row and column contains exactly one 1 with all other entries being 0s. The estimated matrix $\hat{\mathbf{P}}$ is sliced row-wise, with each row being passed through the Gumbel-Softmax function. By adjusting the temperature parameter $\tau$ , the output satisfies the one-hot condition for each row. The resulting rows are then concatenated and passed through the Sinkhorn operator, which iteratively adjusts the rows and columns of the matrix, yielding the final permutation matrix $\mathbf{P}$ . + +Finally, the feature map $\mathbf{f}$ is flattened into a sequence $\mathbf{X}$ , and $\mathbf{P}$ is applied to $\mathbf{X}$ to perform the sorting operation. This process can be viewed as permuting the rows of an $L\times L$ identity matrix according to the sorting procedure. Since all of these operations are differentiable, the entire sorting process is also differentiable, which ensures that each token is optimally positioned in the sequence for subsequent processing by our S3M block. + +# 4.4. Handling Longer and Multi-modal Sequences + +Handling Longer Sequence. Extracting multi-scale information from available modalities requires generating larger permutation matrices to accommodate longer sequences corresponding to high-resolution feature maps. However, this process introduces quadratic computational complexity due to the matrix multiplications involved. To address this, we propose an efficient alternative: bilinear interpolation to scale the permutation matrix, eliminating + +the need for SortNet. The scaled matrix is then processed through ReLU activation, Gumbel-Softmax, and Sinkhorn to generate the permutation matrix $\mathbf{P}_{\mathrm{Large}}$ for the extended sequence, as represented in Fig. 2(c). + +Handling Multi-modal Sequences. To effectively leverage inter-modal correlations, we aim to generate a permutation matrix that accounts for scenarios with missing modalities. We dynamically fuse the individual permutation matrices of each modality— $\mathbf{P}_{\mathrm{FLAIR}}$ , $\mathbf{P}_{\mathrm{T1ce}}$ , $\mathbf{P}_{\mathrm{T1}}$ , $\mathbf{P}_{\mathrm{T2}}$ with a multi-modal code, represented by $h = \{h_1, h_2, h_3, h_4\}$ , where $h_i \in \{0, 1\}$ indicates whether the corresponding modality is available. This fusion produces a comprehensive multi-modal estimated matrix, $\hat{\mathbf{P}}_{\mathrm{Mul}}$ . We then apply a ReLU activation followed by Gumbel-Softmax and Sinkhorn operations to generate the final multi-modal permutation matrix $\mathbf{P}_{\mathrm{Mul}}$ , as shown in Fig. 2(d). Before reordering the multi-modal sequence, each modal sequence is initially projected into a unified representation space. The sequences are then sorted according to $\mathbf{P}_{\mathrm{Mul}}$ , ensuring alignment based on inter-modal relevance and coherence. + +# 4.5. Series State Space Model (S3M) Block + +Due to the insufficient information of missing modalities, it becomes crucial to exhaustively mine information from all available modalities. We introduce the S3M model for local and long-range dependency modeling, inspired by the success of Mamba. + +As shown in Fig. 3, our S3M block connects two unidirectional Mamba-like blocks, $F_{Fwd}$ and $F_{Bwd}$ , in a series connection but in opposite directions, together with a Channel Attention (FCA) block [21], which can be expressed as follows: + +$$ +\mathbf {f} _ {\mathbf {S 3 M}} = \mathrm {F} _ {\mathrm {C A}} \left(\mathrm {F} _ {\mathrm {B w d}} \left(\mathrm {F} _ {\mathrm {F w d}} \left(\mathbf {f} _ {\mathbf {I n}}, \mathbf {P}\right)\right)\right). \tag {9} +$$ + +Before processing, the feature $\mathbf{f}_{\mathrm{In}}$ is reshaped into a sequence and sorted by multiplying it with a soft permutation matrix $\mathbf{P}$ generated through the learnable sorting process. + +Each Mamba-like block comprises three branches: a basic SSM branch, a gated weighting branch, and a residual connection branch. In the SSM branch, the input feature $\mathbf{f}_{\mathrm{In}}$ undergoes a linear projection to expand its channels, followed by depth-wise convolution and SiLU activation, producing $\mathbf{f}_{\mathrm{Mid}}$ and the SSM parameters $\{\overline{\mathbf{A}},\overline{\mathbf{B}},\mathbf{C}\}$ . They are sent to the Forward/Backward SSM module, followed by layer normalization to obtain $\mathbf{f}_{\mathrm{SSM}}$ . The input feature $\mathbf{f}_{\mathrm{In}}$ is processed through a linear layer and SiLU activation to generate a gated weight feature map $\mathbf{f}_{\mathrm{Gate}}$ . Combining these with the residual connection branch, the output of the + +Mamba-like block is: + +$$ +\begin{array}{l} \mathbf {f} _ {\mathbf {S S M}} = \mathrm {L N} (\mathrm {S S M} (\mathrm {S i L U} (\mathrm {D W - C o n v} (\mathrm {L i n e a r} (\mathrm {L N} (\mathbf {f} _ {\mathbf {I n}})))))), \\ \mathbf {f} _ {\text {G a t e}} = \operatorname {S i L U} \left(\operatorname {L i n e a r} \left(\mathbf {f} _ {\text {I n}}\right)\right), \\ \mathbf {f} _ {\text {O u t}} = \operatorname {L i n e a r} \left(\mathbf {f} _ {\text {S S M}} \odot \mathbf {f} _ {\text {G a t e}}\right) + \mathbf {f} _ {\text {I n}}. \tag {10} \\ \end{array} +$$ + +For further channel-wise interactions, we introduce the Channel Attention block. Our S3M block, when compared to the recent Vim block [65] and SS2D block [33], offers two main advantages: (1) The series connection in S3M ensures that information from the forward SSM block is integrated into subsequent tokens, which is then relayed back to earlier tokens in the backward SSM block. This contrasts with parallel connections in Vim block and SS2D block, where token sequences receive the same information level. (2) The integration of the permutation matrix in S3M preserves both spatial proximity and long-range semantic correlations. + +For multi-modal fusion, each modal sequence is first projected into a unified representation space and then concatenated along the channel dimension. The multi-modal sequences $\mathbf{X}_{\mathrm{Mul}}$ are sorted by $\mathbf{P}_{\mathrm{Mul}}$ , which are further delivered into our S3M for building inter-modal correlations. Specifically, the S3M utilizes the Mamba-like block where the information from each token in $\mathbf{f}_{\mathrm{Mul}}$ is compressed into SSM parameters— $\overline{\mathbf{A}}_{Mul}$ , $\overline{\mathbf{B}}_{Mul}$ , $\mathbf{C}_{Mul}$ , which contain the compact multi-modal information along their hidden state dimension. Then, the Forward/Backward SSM module interact across both the channel dimensions of the multi-modal sequences and the state dimensions of the SSM parameters, representing the modality interaction between the origin format and the compact format, thereby, resulting in an efficient fusion that markedly improves segmentation. + +# 4.6. Global Input Strategy + +We introduce a global input strategy that integrates the full volumes information into the analysis of cropped volumes without incurring excessive computational costs, as detailed in Fig. 2. The full volumes, augmented with a relative position mask, undergoes a series of transformations through convolutional layers and non-linear activations. They are then resized to match the dimensions of the cropped volumes using adaptive average pooling. The resulting feature map, which combines coarse global information with the fine details of the cropped volume, is then fed into our LS3M for segmentation. + +The relative position mask, specifically designed for the full volumes, is encoded at positions corresponding to the cropped region. This design emphasizes the relevant area within the larger volume, guiding the network's attention to the regions slated for segmentation. + +# 4.7. Network Architecture and Loss + +Our framework, depicted in Fig. 2, processes inputs of four modalities. Each modality includes both a cropped volume for segmentation and its corresponding global volume. The global volumes are first processed through our global input strategy. The yielded feature maps are then concatenated with their respective cropped volumes and passed through individual encoders $E(\cdot)_M$ to extract hierarchical intra-modal features. + +The features from each modality are transformed into sequences and fed into $\mathrm{SortP}_M$ , the learnable sorting process, which generates a modality-specific permutation matrix, $\mathbf{P}_{\mathbf{M}}$ . This matrix reorders the sequences before they are input into the S3M tailored for each modality. + +At higher resolutions, for each modality, a larger permutation matrix, $\mathbf{P}_{\mathrm{Large}}$ , is obtained by bilinear upsampling. For multi-modal integration, the comprehensive multi-modal permutation matrix, $\mathbf{P}_{\mathrm{Mul}}$ , is generated by combining the modality-specific permutation matrix $\mathbf{P}_{\mathrm{M}}$ with a multi-modal code $h$ . These permutation matrices play a crucial role in integrating the sequences within the S3M, ultimately enhancing segmentation accuracy. + +The final segmentation results are obtained after hierarchical fusion and processing of the multi-modal features. Simulated missing modalities are represented by substituting zero-value feature maps prior to concatenating individual modal features. + +Following [8, 55], we employ the weighted cross-entropy loss $Loss_{WCE}$ and the Dice loss $Loss_{Dice}$ . The total loss is $Loss = Loss_{WCE} + Loss_{Dice}$ . + +# 5. Experiments + +# 5.1. Dataset and implementation details + +We perform our experiments on BraTS2018 and BraTS2020 from the Multi-modal Brain Tumor Segmentation Challenge [37], which align with that of previous studies [32, 45]. The two datasets contain 285 and 369 cases, respectively, with publicly available ground truth. Consistent with [8], BraTS2020 is randomly split into training, validation, and test sets with a ratio of $219:50:100$ . BraTS2018 is split into $199:29:57$ and includes a three-fold validation. These splits are consistently used across all experiments, ensuring that all methods are evaluated under the same conditions. + +Each case of the datasets has four different modalities, i.e., FLAIR, T1ce, T1, and T2 modalities. These modalities are characterized by a volume size of $240 \times 240 \times 155$ , and they capture various properties of brain tumor subregions: GD-enhancing tumor (ET), peritumoral edema (ED), and the necrotic and non-enhancing tumor core (NCR/NET). These subregions of brain tumors are grouped into three nested subregions: the whole tumor (WT), the tumor core + +![](images/b82111d5aebc7e81ac1b261bd70d1874763f1bbce24cedbd0c0a6e84d7623d57.jpg) +Figure 4. Segmentation results of different methods with various available modalities on BraTS2020. + +(TC), and the enhancing tumor (ET). We use Dice coefficient, $95\%$ Hausdorff distance, sensitivity, and specificity to evaluate our method. + +We conduct our experiments using the PyTorch framework version 1.13.0. The training process is carried out on four NVIDIA A800 GPUs, spanning 500 epochs, with a batch size of four volumes. We utilize the Adam optimizer with an initial learning rate set to $2 \times 10^{-4}$ and a weight decay of $1 \times 10^{-5}$ . During the training phase, the input images are randomly cropped to dimensions of $80 \times 80 \times 80$ . Data augmentation techniques include random rotations, flips, and intensity adjustments. + +# 5.2. Comparisons with the State-of-the-art + +To evaluate the effectiveness of our method, we compare our method with nine state-of-the-art methods on different cases with missing modalities. The involved methods contain HeMiS [18], U-HVED [10], RFNet [8], UNet-MFI [59], mmFormer [55], MMANet [48], ShaSpec [45], M3AE [32], and GSS [39]. For a fair comparison, all methods are trained under their recommended hyper-parameters within the same dataset split. + +As illustrated in Tab. 1, our method achieves better performance for most combinations of missing modalities. We achieve improvements of $1.22\%$ , $1.31\%$ , and $2.3\%$ over the second-ranked method on the average DSC for WT, TC, and ET. We also provide a visualization comparison in Fig. 4, illustrating that our method yields more accurate segmentation results in different combinations of modalities. More comparisons, including statistical results, additional metrics and model complexity, are provided in the Appendix. + +# 5.3. Ablation Study + +We conduct ablation experiments on BraTS2020 to assess the effectiveness of specific designs. The baseline is served + +![](images/f4a497b3db952e301a3611662629cffe949df24fc7bf7e324ee03d792af0b11a.jpg) +Figure 5. Effect of hyper-parameters $\tau$ and $n$ . + +as a simplified version of LS3M without GIS, SortP, and S3M components. + +Effect of Scanning Strategies. Relative to our baseline, which employs no specific scanning strategies, adding bidirectional, and fourth-directional (cross scan) methods demonstrated enhanced effectiveness, attributed to their ability to maintain directional spatial coherence. Notably, our sorted scanning strategy, as detailed in Tab. 2, achieved the highest Dice scores by preserving spatial inductive biases and semantic correlations. + +Effect of Proposed Components. As shown in Tab. 3, integrating a global input significantly improves baseline performance by providing larger spatial contexts. The addition of a relative position mask further enhances results by focusing on key areas within the broader input context. The average performance improves $1.84\%$ with S3M. The complete LS3M architecture, after including SortP, achieves superior performance, effectively exploring intra-modal and inter-modal relations. + +
MFLAIR●○○○●●●○○○●●●○●AVG
T1ce○●○○●○○●○○●●○●●AVG
T1○○●○○○○●○●●○●●●●
T2○○○●○○○●○●○●●●●●
WTHeMiS71.6067.7168.9668.1969.1768.6769.8369.0169.7869.4070.2171.2870.7371.5872.0669.88
U-HVED69.8546.8246.7754.0361.4558.2564.5062.9165.7664.2966.9969.7068.3870.3571.4162.76
RFNet86.4277.3476.4686.2189.5589.3089.3581.0087.4587.9590.3990.2090.4288.5990.7786.76
UNet-MFI82.2773.1872.1082.4583.6484.3484.8577.3083.4483.5285.4585.7085.8584.2084.9382.21
mmFormer82.4074.2574.3783.0784.5484.6185.8277.9884.0584.0085.3486.1186.2284.6486.3882.92
MMANet83.2164.4664.3582.8886.7487.6487.4770.9885.6485.3387.9289.3787.9786.3889.2582.64
ShaSpec84.1070.3670.2383.9588.6589.0788.7076.0286.2187.1589.4789.4390.0787.6290.1184.74
M3AE86.5373.8576.7186.0989.4889.3889.2578.1187.3787.2089.9990.1890.4288.6190.5686.25
GSS*86.3672.4573.3386.7689.5190.6590.9178.3888.8790.0591.1493.2590.6789.4393.2887.00
Ours88.6678.2479.3287.2991.0090.5791.4582.5889.1888.8891.5291.8791.4589.7592.1588.22
TCHeMiS53.4351.4151.5651.1151.7051.0851.8551.8852.3551.5152.9553.7652.9754.3855.0352.46
U-HVED34.6235.5127.3037.6742.1538.2643.4144.9347.5344.9749.1351.3049.4052.7254.1743.53
RFNet65.0482.3764.3168.4784.6971.4572.6283.1584.0672.1184.7184.7074.2884.1184.7477.39
UNet-MFI63.9477.6359.3868.0579.9268.2370.7277.6180.0970.2180.0380.9471.4080.7581.2874.01
mmFormer66.1977.9661.1769.1880.3669.5871.5579.9380.7970.9080.1881.3172.0281.1281.2274.90
MMANet65.7974.0557.4770.2185.2172.7372.6580.9985.2471.8885.8785.6874.4886.3986.6376.99
ShaSpec66.0677.3459.5768.2985.0272.8772.9580.8484.4672.2885.8385.5774.7485.4786.5077.19
M3AE68.0481.3966.0070.2782.0173.8274.9582.3983.0172.5482.4483.0675.0984.0684.4077.56
GSS*67.3378.4361.6770.1084.8873.4574.1782.2086.1274.1886.5088.3974.3386.2888.6778.45
Ours71.9883.6166.6772.3086.3275.2576.8784.0985.1874.3286.0585.8476.8785.8385.8679.76
ETHeMiS43.7742.4141.5941.4541.8340.2941.1942.0842.3941.0043.6744.1642.9545.2746.3342.69
U-HVED12.8824.947.2724.2630.0221.9529.4033.6436.1832.1239.3940.9138.0943.1845.3330.64
RFNet40.4774.2737.5143.5976.4543.8146.9975.2273.9446.3777.0176.3848.9576.3876.6460.93
UNet-MFI39.7069.4229.3846.0070.1340.0648.6969.2572.3245.7171.2870.8846.5572.0071.4157.52
mmFormer40.4768.9133.9745.6169.8143.6348.0971.1070.7245.9270.0871.6048.3870.6571.3658.02
MMANet36.4066.0728.2241.9170.4341.5044.2368.8771.8643.5871.9471.9844.6071.8172.4356.39
ShaSpec38.9266.9331.7542.5870.8543.2245.6269.9271.0444.2172.0871.0946.2271.8172.0257.22
M3AE40.4972.4339.9345.9774.6643.2047.3075.4276.8146.6375.9477.0848.1977.4078.0061.30
GSS*42.0369.4635.3045.6074.2047.7849.3374.4174.7848.9476.1276.3250.0276.9875.9561.15
Ours40.3477.6836.5745.3079.1745.0751.5379.4978.6947.0480.7580.5151.5381.5380.9463.60
+ +Table 1. Performance comparison (DSC%) with SOTA methods on BraTS2020. Available and missing modalities are denoted by $\bullet$ and $\circ$ , respectively. \* demonstrates that its codes and results are reproduced by ourselves. + +
DiceBaseline+ Uni-D+ Bi-D+ Fourth-D+ Random+ Our sorted
WT85.4385.2085.5586.1384.2386.94
TC76.1775.6976.6777.5474.0878.01
ET58.0257.9858.4360.7156.7962.06
+ +Table 2. Effect of scanning strategies, '-D' denotes directional scanning. + +
DiceBaseline+GIS (w/o mask)+GIS+GIS + S3M+GIS + SortP + S3M
WT85.4385.9986.2387.1788.22
TC76.1776.7877.0679.0679.76
ET58.0259.0160.3562.9463.60
+ +Table 3. Effect of proposed components, 'GIS' denotes Global Input Strategy. + +Effects of hyper-parameters of temperature $\tau$ and iterative times $n$ . Adjusting $\tau$ and $n$ reveals their critical roles in influencing the performance of our LS3M, as shown in Fig. 5. The parameter $\tau$ modulates the smoothness of the permutation matrix, with $\tau = 1$ yielding optimal results when $n$ is fixed. Conversely, increasing $n$ enhances the precision of the permutation matrix, thus improving segmentation outcomes. Balancing computational efficiency and precision, we select $n = 10$ . + +Additional ablation experiments are provided in the appendix. These experiments further validate the importance of the proposed components and their contribution to the + +model's performance. + +# 6. Conclusion + +In this work, we introduce the Learnable Sorting State Space Model (LS3M), a novel framework for brain tumor segmentation with missing modalities, which aims to maximizing mining information from available modalities. LS3M replaces traditional non-differentiable sorting with a differentiable permutation matrix, preserving 3D spatial information and enabling more effective modeling. The framework adeptly scales these matrices to accommodate longer sequences, capturing multi-scale information. It also dynamically integrates these matrices for robust multi-modal sequence fusion, effectively managing various scenarios of missing modalities. The S3M block captures both intra-modal and inter-modal correlations, enhancing segmentation accuracy. With a global input strategy, LS3M provides comprehensive contextual insights, surpassing existing segmentation methods. Experimental results on BraTS2018 and BraTS2020 demonstrate that LS3M outperforms state-of-the-art. Our results highlight LS3M's potential in advancing medical imaging and healthcare diagnostics. + +# 7. Acknowledgement + +This work was in part supported by the National Natural Science Foundation of China under grants 62032006 and 62021001, and the Anhui Province Key Laboratory of Biomedical Imaging and Intelligent Processing under grants 23YGXT001. + +# References + +[1] Ryan Prescott Adams and Richard S Zemel. Ranking via sinkhorn propagation. arXiv preprint arXiv:1106.1925, 2011.3 +[2] Reza Azad, Nika Khosravi, and Dorit Merhof. Smu-net: Style matching u-net for brain tumor segmentation with missing modalities. In International Conference on Medical Imaging with Deep Learning, pages 48-62. PMLR, 2022. 1, 2 +[3] Ali Behrouz and Farnoosh Hashemi. Graph mamba: Towards learning on graphs with state space models. arXiv preprint arXiv:2402.08678, 2024. 3 +[4] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In Computer Vision-ECCV 2022 Workshops, pages 205-218. Springer, 2023. 2 +[5] Cheng Chen, Qi Dou, Yueming Jin, Hao Chen, Jing Qin, and Pheng-Ann Heng. Robust multimodal brain tumor segmentation via feature disentanglement and gated fusion. In Medical Image Computing and Computer-Assisted Intervention-MICCAI, pages 447-456. Springer, 2019. 2 +[6] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 2 +[7] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 1 +[8] Yuhang Ding, Xin Yu, and Yi Yang. Rfnet: Region-aware fusion network for incomplete multi-modal brain tumor segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3975-3984, 2021. 1, 2, 6, 7 +[9] Yuhang Ding, Liulei Li, Wenguan Wang, and Yi Yang. Clustering propagation for universal medical image segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2024. 2 +[10] Reuben Dorent, Samuel Joutard, Marc Modat, Sébastien Ourselin, and Tom Vercauteren. Hetero-modal variational encoder-decoder for joint modality completion and segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI, pages 74-82. Springer, 2019. 7 +[11] Zhengcong Fei, Mingyuan Fan, Changqian Yu, and Junshi Huang. Scalable diffusion models with state space backbone. arXiv preprint arXiv:2402.05608, 2024. 3 + +[12] Jiannan Ge, Lingxi Xie, Hongtao Xie, Pandeng Li, Xiaopeng Zhang, Yongdong Zhang, and Qi Tian. Alignzeg: Mitigating objective misalignment for zero-shot semantic segmentation. In European Conference on Computer Vision, pages 142-161. Springer, 2024. 2 +[13] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3 +[14] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In International Conference on Learning Representations, 2021. 2 +[15] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024.3 +[16] Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In *Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: 7th International Workshop, BrainLes* 2021, Held in Conjunction with MICCAI 2021, Virtual Event, pages 272-284. Springer, 2022. 2 +[17] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 574-584, 2022. 2 +[18] Mohammad Havaei, Nicolas Guizard, Nicolas Chapados, and Yoshua Bengio. Hemis: Hetero-modal image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI, pages 469-477. Springer, 2016. 7 +[19] Xuanhua He, Ke Cao, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Pan-mamba: Effective pan-sharpening with state space model. arXiv preprint arXiv:2402.12192, 2024.3 +[20] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5 +[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 5 +[22] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3, 4 +[23] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In International Conference on Learning Representations, 2016. 3 +[24] Zeyu Jiang, Changxing Ding, Minfeng Liu, and Dacheng Tao. Two-stage cascaded u-net: 1st place solution to brats challenge 2019 segmentation task. In *Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries*, pages 231-241. Springer, 2020. 2 +[25] Sanaz Karimijafarbigloo, Reza Azad, Amirhossein Kazerouni, Saeed Ebadollahi, and Dorit Merhof. Mmcformer: Missing modality compensation transformer for brain tumor segmentation. In Medical Imaging with Deep Learning, pages 1144-1162. PMLR, 2024. 2 + +[26] Jonghun Kim and Hyunjin Park. Adaptive latent diffusion model for 3d medical image to image translation: Multimodal magnetic resonance imaging study. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7604-7613, 2024. 1, 2 +[27] Dongwook Lee, Won-Jin Moon, and Jong Chul Ye. Assessing the importance of magnetic resonance contrasts using collaborative generative adversarial networks. Nature Machine Intelligence, 2(1):34-42, 2020. 1, 2 +[28] Ho Hin Lee, Shunxing Bao, Yuankai Huo, and Bennett A Landman. 3d ux-net: A large kernel volumetric convnet modernizing hierarchical transformer for medical image segmentation. arXiv preprint arXiv:2209.15076, 2022. 2 +[29] Jiangyun Li, Wenxuan Wang, Chen Chen, Tianxiang Zhang, Sen Zha, Hong Yu, and Jing Wang. Transbtsv2: Wider instead of deeper transformer for medical image segmentation. arXiv preprint arXiv:2201.12785, 2022. 2 +[30] Mingcheng Li, Dingkang Yang, Xiao Zhao, Shuaibing Wang, Yan Wang, Kun Yang, Mingyang Sun, Dongliang Kou, Ziyun Qian, and Lihua Zhang. Correlation-decoupled knowledge distillation for multimodal sentiment analysis with incomplete modalities. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2024. 2 +[31] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 3 +[32] Hong Liu, Dong Wei, Donghuan Lu, Jinghan Sun, Liansheng Wang, and Yefeng Zheng. M3ae: Multimodal representation learning for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1657-1665, 2023. 1, 2, 6, 7 +[33] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 1, 3, 4, 6 +[34] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 3 +[35] Andriy Marusyk, Vanessa Almendro, and Kornelia Polyak. Intra-tumour heterogeneity: a looking glass for cancer? Nature reviews cancer, 12(5):323-334, 2012. 1, 4 +[36] Gonzalo Mena, David Belanger, Scott Linderman, and Jasper Snoek. Learning latent permutations with gumbel-sinkhorn networks. In International Conference on Learning Representations, 2018. 3 +[37] Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging, 34(10):1993-2024, 2014. 6 +[38] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977, 2024. 3, 4 +[39] Yansheng Qiu, Delin Chen, Hongdou Yao, Yongchao Xu, and Zheng Wang. Scratch each other's back: Incomplete + +multi-modal brain tumor segmentation via category aware group self-support learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21317-21326, 2023. 1, 2, 7 +[40] Santiago Ramón y Cajal, Marta Sesé, Claudia Capdevila, Trond Aasen, Leticia De Mattos-Arruda, Salvador J Diaz-Cano, Javier Hernández-Losa, and Josep Castellvi. Clinical implications of intratumor heterogeneity: challenges and opportunities. Journal of Molecular Medicine, 98:161-177, 2020. 1, 4 +[41] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 3 +[42] Dong She, Yueyi Zhang, Zheyu Zhang, Hebei Li, Zihan Yan, and Xiaoyan Sun. Eoformer: Edge-oriented transformer for brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 333-343. Springer, 2023. 2 +[43] Jimmy TH Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling. In International Conference on Learning Representations, 2022. 2 +[44] Chloe Wang, Oleksii Tsepa, Jun Ma, and Bo Wang. Graphmamba: Towards long-range graph sequence modeling with selective state spaces. arXiv preprint arXiv:2402.00789, 2024.3 +[45] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15878-15887, 2023. 6, 7 +[46] Wenxuan Wang, Chen Chen, Meng Ding, Hong Yu, Sen Zha, and Jiangyun Li. Transbts: Multimodal brain tumor segmentation using transformer. In Medical Image Computing and Computer Assisted Intervention-MICCAI, pages 109-119. Springer, 2021. 2 +[47] Yixin Wang, Yang Zhang, Yang Liu, Zihao Lin, Jiang Tian, Cheng Zhong, Zhongchao Shi, Jianping Fan, and Zhiqiang He. Acn: adversarial co-training network for brain tumor segmentation with missing modalities. In Medical Image Computing and Computer Assisted Intervention-MICCAI, pages 410-420. Springer, 2021. 1, 2 +[48] Shicai Wei, Chunbo Luo, and Yang Luo. Mmanet: Margin-aware distillation and modality-aware regularization for incomplete multimodal learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20039-20049, 2023. 1, 2, 7 +[49] Jing Nathan Yan, Jiatao Gu, and Alexander M Rush. Diffusion models without attention. arXiv preprint arXiv:2311.18257, 2023. 3 +[50] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. arXiv preprint arXiv:2403.17695, 2024. 1, 3, 4 +[51] Heran Yang, Jian Sun, and Zongben Xu. Learning unified hyper-network for multi-modal mr image synthesis and tu + +mor segmentation with missing modalities. IEEE Transactions on Medical Imaging, 2023. 1, 2 +[52] Qiushi Yang, Xiaoqing Guo, Zhen Chen, Peter YM Woo, and Yixuan Yuan. D 2-net: Dual disentanglement network for brain tumor segmentation with missing modalities. IEEE Transactions on Medical Imaging, 41(10):2953-2964, 2022. 2 +[53] Chenyu You, Ruihan Zhao, Fenglin Liu, Siyuan Dong, Sandeep Chinchali, Ufuk Topcu, Lawrence Staib, and James Duncan. Class-aware adversarial transformers for medical image segmentation. pages 29582-29596, 2022. 2 +[54] Ziqi Yu, Xiaoyang Han, Shengjie Zhang, Jianfeng Feng, Tingying Peng, and Xiao-Yong Zhang. Mousegan++: Unsupervised disentanglement and contrastive representation for multiple mri modalities synthesis and structural segmentation of mouse brain. IEEE Transactions on Medical Imaging, 2022, 1, 2 +[55] Yao Zhang, Nanjun He, Jiawei Yang, Yuexiang Li, Dong Wei, Yawen Huang, Yang Zhang, Zhiqiang He, and Yefeng Zheng. mmformer: Multimodal medical transformer for incomplete multimodal learning of brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 107-117. Springer, 2022. 1, 2, 6, 7 +[56] Zheyu Zhang, Xinzhao Liu, Zheng Chen, Yueyi Zhang, Huanjing Yue, Yunwei Ou, and Xiaoyan Sun. Anatomical consistency distillation and inconsistency synthesis for brain tumor segmentation with missing modalities. arXiv preprint arXiv:2408.13733, 2024. +[57] Zheyu Zhang, Gang Yang, Yueyi Zhang, Huanjing Yue, Aiping Liu, Yunwei Ou, Jian Gong, and Xiaoyan Sun. Tmformer: Token merging transformer for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7414-7422, 2024. 2 +[58] Sijie Zhao, Hao Chen, Xueliang Zhang, Pengfeng Xiao, Lei Bai, and Wanli Ouyang. Rs-mamba for large remote sensing image dense prediction. arXiv preprint arXiv:2404.02668, 2024. 1, 3 +[59] Zechen Zhao, Heran Yang, and Jian Sun. Modality-adaptive feature interaction for brain tumor segmentation with missing modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 183-192. Springer, 2022. 1, 7 +[60] Zhuoran Zheng and Jun Zhang. Fd-vision mamba for endoscopic exposure correction. arXiv preprint arXiv:2402.06378, 2024. 3 +[61] Chenhong Zhou, Changxing Ding, Xinchao Wang, Zhentai Lu, and Dacheng Tao. One-pass multi-task networks with cross-task guided attention for brain tumor segmentation. IEEE Transactions on Image Processing, 29:4516-4529, 2020. 2 +[62] Tao Zhou, Huazhu Fu, Geng Chen, Jianbing Shen, and Ling Shao. Hi-net: hybrid-fusion network for multi-modal mr image synthesis. IEEE Transactions on Medical Imaging, 39 (9):2772-2781, 2020. 1, 2 +[63] Tongxue Zhou, Stéphane Canu, Pierre Vera, and Su Ruan. Latent correlation representation learning for brain tumor + +segmentation with missing mri modalities. IEEE Transactions on Image Processing, 30:4263-4274, 2021. 1, 2 +[64] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: Redesigning skip connections to exploit multiscale features in image segmentation. IEEE Transactions on Medical Imaging, 39(6):1856-1867, 2019. 2 +[65] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. 2024. 1, 3, 4, 6 \ No newline at end of file diff --git a/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/images.zip b/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ad76765ede31de2c49ae41564bbdf552f56d2f2c --- /dev/null +++ b/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ed69700f33580501ccaa8613597f20120115853002862d386af8ba20ba86fb1 +size 647519 diff --git a/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/layout.json b/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..07ef1a54ac0cc8f973c75208d2c9272a2233b235 --- /dev/null +++ b/2025/Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model/layout.json @@ -0,0 +1,10259 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 467, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 467, + 140 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 467, + 140 + ], + "type": "text", + "content": "Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "spans": [ + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "text", + "content": "Zheyu Zhang" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "text", + "content": " Yayuan Lu" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "text", + "content": " Feipeng Ma" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "text", + "content": " Yueyi Zhang" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "text", + "content": " Huanjing Yue" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "text", + "content": " Xiaoyan Sun" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "text", + "content": "MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, 230026, China \n" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "text", + "content": "Anhui Province Key Laboratory of Biomedical Imaging and Intelligent Processing, Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, Hefei 230088, China \n" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 159, + 558, + 258 + ], + "type": "text", + "content": "School of Electrical and Information Engineering, Tianjin University, Tianjin, 300072, China \nzhangzy0@mail.ustc.edu.cn sunxiaoyan@ustc.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 312, + 297, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 297, + 601 + ], + "type": "text", + "content": "Brain tumor segmentation plays a crucial role in clinical diagnosis, yet the frequent unavailability of certain MRI modalities poses a significant challenge. In this paper, we introduce the Learnable Sorting State Space Model (LS3M), a novel framework designed to maximize the utilization of available modalities for brain tumor segmentation. LS3M excels at efficiently modeling long-range dependencies based on the Mamba design, while incorporating differentiable permutation matrices that reorder input sequences based on modality-specific characteristics. This dynamic reordering ensures that critical spatial inductive biases and long-range semantic correlations inherent in 3D brain MRI are preserved, which is crucial for incomplete multi-modal brain tumor segmentation. Once the input sequences are reordered using the generated permutation matrix, the Series State Space Model (S3M) block models the relationships between them, capturing both local and long-range dependencies. This enables effective representation of intra-modal and inter-modal relationships, significantly improving segmentation accuracy. Extensive experiments on the BraTS2018 and BraTS2020 datasets demonstrate that LS3M outperforms existing methods, offering a robust solution for brain tumor segmentation, particularly in scenarios with missing modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 624, + 135, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 135, + 636 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 135, + 636 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "type": "text", + "content": "Accurate brain tumor segmentation in clinical settings relies on the integration of various MRI modalities (FLAIR, T1ce, T1, T2) that provide complementary information for diagnosis. However, the frequent unavailability of cer-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 287, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 528 + ], + "type": "text", + "content": "tain modalities, like T1ce due to contrast agent allergies, presents a challenge. Existing methods addressing missing modalities fall into two categories: (1) bridging the data gap between complete and missing modalities through modality synthesis [26, 27, 32, 51, 54, 62] and knowledge distillation [2, 39, 47, 48] and (2) optimizing architectures to maximize information extraction from available modalities, which is the focus of our work. Some methods [8, 63] adopt multi-scale convolutions to capture critical structural details by focusing on spatial inductive bias, such as local texture. However, the limited receptive field restricts their ability to represent tumor tissue effectively, especially when other areas may appear visually similar in localized views, making it difficult to differentiate tumor tissue [35, 40]. Some methods explore models such as Transformers [55] and graph neural networks [51, 59] to build long-range semantic correlation to mitigate the mentioned problem, but they suffer from quadratic computational complexity. The computational inefficiency is further exacerbated by the large sequence lengths typical of 3D brain MRI." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Recent advancements in selective state space models (SSMs), such as Mamba [7, 13], offer the potential to obtain long-range representations for available modalities with dynamic weighting parameters in an efficient way. Mamba flattens the 3D brain MRI into a 1D sequence and iteratively scans this sequence to compress hidden states, which are then used to enhance the current token by combining it with previously scanned tokens. While Mamba can efficiently build long-range dependencies, the influence of distant tokens on the current token diminishes significantly. This becomes problematic for 3D brain MRI, as flattening the sequence destroys the 3D spatial structure. Follow-up methods [33, 50, 58, 65] propose bi-directional, four-directional, and omni-directional scanning to mitigate distortion to spatial structures, as shown in Fig. 1. However, these methods" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25982" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "either neglect local inductive bias or fail to capture long-range semantic correlations of 3D brain MRI, hindering their ability to fully capture contextual relationships essential for brain tumor segmentation, particularly in scenarios with missing modalities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 136, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 136, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 136, + 296, + 506 + ], + "type": "text", + "content": "To address these challenges, we propose the Learnable Sorting State Space Model (LS3M), specifically designed for brain tumor segmentation with missing modalities. Firstly, LS3M incorporates the Sinkhorn operator with the Gumbel-Softmax reparameterization, which converts the traditional non-differentiable sorting into a differentiable permutation matrix. This matrix is used to reorder the 1D sequence of available modalities, considering spatial inductive biases and long-range semantic correlations. Importantly, the permutation matrix does not directly establish relationships; rather, it prepares the sequence for further processing by the Series State Space Model (S3M), which is capable of modeling intra-modal and inter-modal relationships by leveraging both local and long-range correlations. Next, LS3M scales the intra-modal permutation matrix at larger resolutions, allowing for the extraction of multi-scale information. These modality-specific permutation matrices are also combined for multi-modal fusion to capture correlations across modalities. The reordered sequences, now enriched with spatial and semantic information, are then processed by the S3M block, which builds relationships at both intra-modal and inter-modal levels, maximizing the extraction of information from the available modalities. Lastly, LS3M departs from existing models that rely on cropped input volumes by adopting a global input strategy, supplemented with a relative position mask. This approach not only provides richer contextual information but also informs the model about the location of the cropped volume within the larger global context, significantly enhancing spatial awareness and diagnostic accuracy. In summary, our key contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "content": "- We propose LS3M for brain tumor segmentation with missing modalities. LS3M improves SSM by dynamically reordering data, capturing patient-specific characteristics, and enhancing the ability to extract local and long-range correlations from available modalities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 570, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 630 + ], + "type": "text", + "content": "- LS3M scales and integrates the permutation matrix to capture multi-scale and multi-modal information. The reordered sequences are then processed by the S3M block to model complex intra-modal and inter-modal relationships, improving segmentation accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 630, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 630, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 56, + 630, + 295, + 676 + ], + "type": "text", + "content": "- LS3M incorporates a global input strategy with relative position embeddings, significantly enhancing contextual awareness towards comprehensive utilization of available modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 296, + 714 + ], + "type": "text", + "content": "- Extensive evaluations on BraTS2018 and BraTS2020 demonstrate the superiority and effectiveness of LS3M for brain tumor segmentation, particularly in clinically" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 330, + 71, + 381, + 121 + ], + "blocks": [ + { + "bbox": [ + 330, + 71, + 381, + 121 + ], + "lines": [ + { + "bbox": [ + 330, + 71, + 381, + 121 + ], + "spans": [ + { + "bbox": [ + 330, + 71, + 381, + 121 + ], + "type": "image", + "image_path": "226b16430542e9eda8c561eef97e71061ea237dd5a3199b0540dc10da8b1fc52.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 330, + 121, + 381, + 128 + ], + "lines": [ + { + "bbox": [ + 330, + 121, + 381, + 128 + ], + "spans": [ + { + "bbox": [ + 330, + 121, + 381, + 128 + ], + "type": "text", + "content": "Long-range Semantic Correlation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 333, + 130, + 373, + 140 + ], + "lines": [ + { + "bbox": [ + 333, + 130, + 373, + 140 + ], + "spans": [ + { + "bbox": [ + 333, + 130, + 373, + 140 + ], + "type": "text", + "content": "(a) Requirements for identifying brain tumor" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 386, + 73, + 435, + 99 + ], + "blocks": [ + { + "bbox": [ + 386, + 73, + 435, + 99 + ], + "lines": [ + { + "bbox": [ + 386, + 73, + 435, + 99 + ], + "spans": [ + { + "bbox": [ + 386, + 73, + 435, + 99 + ], + "type": "image", + "image_path": "c5c93f5ad2c5433c5c6b438dd2c15828fc95d70ef809b7cd509095b79163af68.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 100, + 432, + 106 + ], + "lines": [ + { + "bbox": [ + 389, + 100, + 432, + 106 + ], + "spans": [ + { + "bbox": [ + 389, + 100, + 432, + 106 + ], + "type": "text", + "content": "(b) Horizontal Scanning" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 387, + 107, + 435, + 133 + ], + "blocks": [ + { + "bbox": [ + 387, + 107, + 435, + 133 + ], + "lines": [ + { + "bbox": [ + 387, + 107, + 435, + 133 + ], + "spans": [ + { + "bbox": [ + 387, + 107, + 435, + 133 + ], + "type": "image", + "image_path": "0b5e33c4b176ad89c13e443450f34cab9c2344e525547128ae9867a0bfbd5fc0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 134, + 428, + 140 + ], + "lines": [ + { + "bbox": [ + 395, + 134, + 428, + 140 + ], + "spans": [ + { + "bbox": [ + 395, + 134, + 428, + 140 + ], + "type": "text", + "content": "(d) Local Scanning" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 152, + 555, + 207 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 207 + ], + "type": "text", + "content": "Figure 1. (a) Brain tumor segmentation requires local inductive bias to capture small but critical structural details, as well as long-range semantic correlations to understand spatial relationships between the tumor and surrounding structures. (b)-(e) Different scanning strategies." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 442, + 73, + 491, + 99 + ], + "blocks": [ + { + "bbox": [ + 442, + 73, + 491, + 99 + ], + "lines": [ + { + "bbox": [ + 442, + 73, + 491, + 99 + ], + "spans": [ + { + "bbox": [ + 442, + 73, + 491, + 99 + ], + "type": "image", + "image_path": "56b3633d1f10860e505a4efae85e5e7c96f13c06306be2496aeb275343e432c3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 100, + 484, + 106 + ], + "lines": [ + { + "bbox": [ + 447, + 100, + 484, + 106 + ], + "spans": [ + { + "bbox": [ + 447, + 100, + 484, + 106 + ], + "type": "text", + "content": "(c) Vertical Scanning" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 444, + 107, + 505, + 133 + ], + "blocks": [ + { + "bbox": [ + 444, + 107, + 505, + 133 + ], + "lines": [ + { + "bbox": [ + 444, + 107, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 444, + 107, + 505, + 133 + ], + "type": "image", + "image_path": "57487da9e89fdbb6754875089c890ead9ca48d169035c5e7e1ff80fc17f5fd9a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 134, + 509, + 140 + ], + "lines": [ + { + "bbox": [ + 468, + 134, + 509, + 140 + ], + "spans": [ + { + "bbox": [ + 468, + 134, + 509, + 140 + ], + "type": "text", + "content": "(e) Our Sorted Scanning" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 498, + 73, + 541, + 99 + ], + "blocks": [ + { + "bbox": [ + 498, + 73, + 541, + 99 + ], + "lines": [ + { + "bbox": [ + 498, + 73, + 541, + 99 + ], + "spans": [ + { + "bbox": [ + 498, + 73, + 541, + 99 + ], + "type": "image", + "image_path": "789d2c2627058b7a85e096435041360b9d66210021dc5559f3556af0cb939e8b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 522, + 100, + 541, + 133 + ], + "blocks": [ + { + "bbox": [ + 522, + 100, + 541, + 133 + ], + "lines": [ + { + "bbox": [ + 522, + 100, + 541, + 133 + ], + "spans": [ + { + "bbox": [ + 522, + 100, + 541, + 133 + ], + "type": "image", + "image_path": "8fc4edbd42798f53e0efb5f24a8a81e4d12f3014148c6a611cc750d3b73ea556.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 323, + 229, + 497, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 229, + 497, + 241 + ], + "spans": [ + { + "bbox": [ + 323, + 229, + 497, + 241 + ], + "type": "text", + "content": "relevant scenarios with missing modalities." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 253, + 400, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 253, + 400, + 265 + ], + "spans": [ + { + "bbox": [ + 314, + 253, + 400, + 265 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 273, + 555, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 381 + ], + "type": "text", + "content": "Medical Image Segmentation. Medical image segmentation traditionally relies on U-Net-like structures, which formulate the task as dense pixel classification. Enhancements to segmentation structures include cascaded networks [24, 61], refined skip connections [64], hierarchical pyramids [9], attention mechanisms [4, 6, 12, 16, 17, 29, 42, 46, 53], and large depth-wise kernels [28]. Despite achieving notable performance, these methods often fail when certain modalities are absent." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 397, + 555, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 625 + ], + "type": "text", + "content": "Multi-modal Learning for Missing Modalities. Research in multi-modal learning with missing modalities generally seeks to either bridge the gap between scenarios with complete and missing modalities or to optimize architectural design to maximize the use of available modal information. The former approach includes modality synthesis [26, 27, 32, 51, 54, 62] to generate absent modalities and knowledge distillation techniques [2, 30, 39, 47, 48] that transfer insights from complete to incomplete scenarios. The latter approach utilizes techniques like modality disentanglement [5, 52] to separate modality-specific information in the frequency domain and modality fusion strategies involving hierarchical pyramids [8, 63] and attention mechanisms [25, 55-57]. However, combining attention mechanisms with other techniques often results in significant computational overhead, which can impede effective modeling long-range dependencies. In our work, we leverage SSM, specifically the Mamba, to efficiently build long-range correlations from all available modalities." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "State Space Model. Drawing inspiration from control theory, the State Space Model (SSM) recurrently updates sequences through latent states [14, 43], as demonstrated by the Structured State Space Sequence model (S4), which utilizes linear state space equations to efficiently scale with sequence length. Recently, Gu and Dao [13] propose a" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25983" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "generic modal, called Mamba, with selective projection, gated mechanism and efficient hardware design, outperforming Transformers in linear-time inference on natural language processing. This success has spurred adaptations of Mamba in computer vision. Since visual tokens lack a sequential arrangement, methods propose diverse scanning strategies like bi-/four-/omni-directional [33, 50, 58, 65], local [22], and atrous scanning [38], enhancing contextual data integration for tasks such as medical image segmentation [34, 41, 41], image restoration [15, 19, 60], point cloud analysis [31], image generation [11, 49], and graph neural networks [3, 44]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 224, + 141, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 141, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 141, + 237 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 245, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 293 + ], + "type": "text", + "content": "State Space Model (SSM) for Sequence Modeling. The SSM defines a linear Ordinary Differential Equation (ODE) to map an input sequence " + }, + { + "bbox": [ + 55, + 245, + 296, + 293 + ], + "type": "inline_equation", + "content": "x(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 245, + 296, + 293 + ], + "type": "text", + "content": " to an output sequence " + }, + { + "bbox": [ + 55, + 245, + 296, + 293 + ], + "type": "inline_equation", + "content": "y(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 245, + 296, + 293 + ], + "type": "text", + "content": " through a hidden state " + }, + { + "bbox": [ + 55, + 245, + 296, + 293 + ], + "type": "inline_equation", + "content": "h(t) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 245, + 296, + 293 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 297, + 294, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 297, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 79, + 297, + 294, + 312 + ], + "type": "interline_equation", + "content": "h ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), \\quad y (t) = \\mathbf {C} h (t), \\tag {1}", + "image_path": "71786e06f3e9dc27c06d94a481d426bdc994de7037acf7953770db5e9019239e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 315, + 296, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 362 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 315, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 55, + 315, + 296, + 362 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 315, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 55, + 315, + 296, + 362 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 315, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{1 \\times N}" + }, + { + "bbox": [ + 55, + 315, + 296, + 362 + ], + "type": "text", + "content": " are learned parameters. Since images are processed into discrete token sequences, the discrete version of Eq. (1) is defined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 92, + 368, + 294, + 382 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 368, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 92, + 368, + 294, + 382 + ], + "type": "interline_equation", + "content": "h _ {k} = \\overline {{\\mathbf {A}}} h _ {k - 1} + \\overline {{\\mathbf {B}}} x _ {k}, \\quad y _ {k} = \\mathbf {C} h _ {k}, \\tag {2}", + "image_path": "be74112b212ac26576a6f742be95a8f00946ca81a6cba3fe17ecbf230e25ff38.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 387, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 387, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}}" + }, + { + "bbox": [ + 55, + 387, + 296, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 387, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{B}}" + }, + { + "bbox": [ + 55, + 387, + 296, + 449 + ], + "type": "text", + "content": " are discretized using the discretization step size " + }, + { + "bbox": [ + 55, + 387, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 387, + 296, + 449 + ], + "type": "text", + "content": " in the zero-order hold (ZOH) rule. The long-range dependencies are constructed via recurrent scanning. However, the recurrent SSM is not practical. Eq. (2) is equivalent to a global convolution kernel " + }, + { + "bbox": [ + 55, + 387, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{K}}\\in \\mathbb{R}^{L}" + }, + { + "bbox": [ + 55, + 387, + 296, + 449 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 60, + 452, + 294, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 452, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 60, + 452, + 294, + 472 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {K}}} = \\left(\\mathbf {C} \\overline {{\\mathbf {B}}}, \\mathbf {C} \\overline {{\\mathbf {A B}}}, \\dots , \\mathbf {C} \\overline {{\\mathbf {A}}} ^ {\\mathbf {L} - 1} \\overline {{\\mathbf {B}}}\\right), \\quad \\mathbf {y} = \\mathbf {x} * \\overline {{\\mathbf {K}}}. \\tag {3}", + "image_path": "393eee7775ee7b557e5405dd205f4c1d999d86966a9fbb8f7fe1b087ba9c864f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 475, + 296, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 536 + ], + "type": "text", + "content": "Structured state space sequence models (S4) imposes a structure on the state matrix " + }, + { + "bbox": [ + 55, + 475, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 475, + 296, + 536 + ], + "type": "text", + "content": " based on HIPPO matrices, which significantly improves the efficiency for obtaining the SSM kernel " + }, + { + "bbox": [ + 55, + 475, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{K}}" + }, + { + "bbox": [ + 55, + 475, + 296, + 536 + ], + "type": "text", + "content": ". Further, Mamba [13] adds selective projection via making " + }, + { + "bbox": [ + 55, + 475, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{B}, \\mathbf{C}, \\Delta" + }, + { + "bbox": [ + 55, + 475, + 296, + 536 + ], + "type": "text", + "content": " input-dependent." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "content": "Sinkhorn Operator. The Sinkhorn operator " + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "content": " is an iterative procedure used to transform a non-negative matrix " + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{L \\times L}" + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "content": " into a doubly stochastic matrix, which is a square matrix with non-negative entries such that each row and each column sums to one [1, 36]. The process can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 622, + 295, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 622, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 64, + 622, + 295, + 656 + ], + "type": "interline_equation", + "content": "\\begin{array}{l}S ^ {0} (\\mathbf {P}) = \\exp (\\mathbf {P}), \\quad S ^ {n} (\\mathbf {P}) = \\mathcal {T} _ {c} \\left(\\mathcal {T} _ {r} \\left(S ^ {n - 1} (\\mathbf {P})\\right)\\right),\\\\S (\\mathbf {P}) = \\lim _ {n \\rightarrow \\infty} S ^ {n} (\\mathbf {P}),\\end{array}\\tag {4}", + "image_path": "021c2e8abd5fca155cbc864b8f2f815f29f0c0d613bacd985641be2c29c5284e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 659, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 295, + 684 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 659, + 295, + 684 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 659, + 295, + 684 + ], + "type": "text", + "content": " is iterative times. " + }, + { + "bbox": [ + 55, + 659, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_r(\\cdot)" + }, + { + "bbox": [ + 55, + 659, + 295, + 684 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 659, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c(\\cdot)" + }, + { + "bbox": [ + 55, + 659, + 295, + 684 + ], + "type": "text", + "content": " are the row-wise and column-wise normalizations, defined as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 64, + 688, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 688, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 64, + 688, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {r} (\\mathbf {P}) = \\mathbf {P} \\oslash \\left(\\mathbf {P} \\mathbf {1} _ {\\mathbf {N}} \\mathbf {1} _ {\\mathbf {N}} ^ {\\top}\\right), \\mathcal {T} _ {c} (\\mathbf {P}) = \\mathbf {P} \\oslash \\left(\\mathbf {1} _ {\\mathbf {N}} \\mathbf {1} _ {\\mathbf {N}} ^ {\\top} \\mathbf {P}\\right), \\tag {5}", + "image_path": "4ad559a86a92ad1413a17063a4d85f6544e4be9ff534143b7f78a4bddf6c7342.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\oslash" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " denotes the element-wise division and " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "1_{N}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " is a column vector of ones. We employ Sinkhorn " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "(\\mathbf{P},n)" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " to represent Eq. (4)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 122, + 555, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 555, + 182 + ], + "type": "text", + "content": "Gumbel-Softmax Reparameterization. The Gumbel-Softmax provides a differentiable approximation of the discrete categorical distribution [23]. The process of sample " + }, + { + "bbox": [ + 313, + 122, + 555, + 182 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 122, + 555, + 182 + ], + "type": "text", + "content": " from a categorical distribution with class probabilities " + }, + { + "bbox": [ + 313, + 122, + 555, + 182 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 122, + 555, + 182 + ], + "type": "text", + "content": " can be defined as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 186, + 555, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 186, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 359, + 186, + 555, + 206 + ], + "type": "interline_equation", + "content": "z = \\operatorname {o n e - h o t} (\\underset {i} {\\arg \\max } [ g _ {i} + \\log x _ {i} ]), \\tag {6}", + "image_path": "0d9ad9fa2adac3504a826dd76ad3188aba36966bf1eaf29f85e70296608f92da.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 212, + 555, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 555, + 248 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 212, + 555, + 248 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 313, + 212, + 555, + 248 + ], + "type": "text", + "content": " is the injected standard i.i.d Gumbel noise. Jang et al. [23] use the softmax function as a continuous, differentiable approximation to the arg max function, i.e.," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 254, + 554, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 254, + 554, + 283 + ], + "spans": [ + { + "bbox": [ + 359, + 254, + 554, + 283 + ], + "type": "interline_equation", + "content": "\\operatorname {s o f t m a x} (y / \\tau) \\equiv \\frac {\\exp \\left(y _ {i} / \\tau\\right)}{\\sum_ {i = 1} ^ {k} \\exp \\left(y _ {i} / \\tau\\right)}, \\tag {7}", + "image_path": "8d01556ecaa8ff6955a78337736a0c931782e8f541ae8f57a9e08e4829798847.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 289, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 289, + 554, + 350 + ], + "type": "inline_equation", + "content": "y_{i} = \\log x_{i} + g_{i}" + }, + { + "bbox": [ + 313, + 289, + 554, + 350 + ], + "type": "text", + "content": ". When temperature " + }, + { + "bbox": [ + 313, + 289, + 554, + 350 + ], + "type": "inline_equation", + "content": "\\tau \\rightarrow 0" + }, + { + "bbox": [ + 313, + 289, + 554, + 350 + ], + "type": "text", + "content": ", the softmax distribution converges to a one-hot representation identical to samples from a categorical distribution. We use Gumbel-Softmax(X, " + }, + { + "bbox": [ + 313, + 289, + 554, + 350 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 289, + 554, + 350 + ], + "type": "text", + "content": ") to represent Eq. (7), where " + }, + { + "bbox": [ + 313, + 289, + 554, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = [x_0, x_1, \\dots, x_{|\\mathbf{X}| - 1}]" + }, + { + "bbox": [ + 313, + 289, + 554, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 358, + 370, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 358, + 370, + 370 + ], + "spans": [ + { + "bbox": [ + 314, + 358, + 370, + 370 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 378, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 462 + ], + "type": "text", + "content": "In this section, we first explain the motivation and challenges with Mamba. We then introduce the learnable sorting process to address these challenges, followed by expanding the permutation matrix for longer and multi-modal sequences. Next, we describe the S3M block and the global input strategy. Finally, we present the LS3M architecture and its optimization loss." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 468, + 392, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 392, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 392, + 480 + ], + "type": "text", + "content": "4.1. Motivations" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": "Brain tumor segmentation requires complementary information from multiple modalities. Missing modalities lead to insufficient information, inspiring us to maximize information extraction from available modalities. We try to do this from two aspects: building long-range dependencies and incorporating broader context for available modalities." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": "First, brain tumors often exhibit complex anatomy, where local inductive biases alone may be insufficient to identify tumor types and boundaries. Long-range dependencies are essential for understanding spatial relationships between the tumor and surrounding structures. While Transformers can model such dependencies, their quadratic computational complexity makes them impractical for 3D MRI data due to large token sequences and high computational costs, as illustrated in the Appendix. Mamba, however, captures long-range dependencies more efficiently, making it a better fit for this task." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Second, most segmentation methods focus on cropped volumes to reduce computational cost, neglecting the" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25984" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 362, + 190 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 362, + 190 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 362, + 190 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 362, + 190 + ], + "type": "image", + "image_path": "a4fd3450080f08915800b6a0a20969c752829cb4af17862d5781c2d2c9531aab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 193, + 257, + 203 + ], + "lines": [ + { + "bbox": [ + 163, + 193, + 257, + 203 + ], + "spans": [ + { + "bbox": [ + 163, + 193, + 257, + 203 + ], + "type": "text", + "content": "(a) Our LS3M framework" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 366, + 72, + 548, + 192 + ], + "blocks": [ + { + "bbox": [ + 366, + 72, + 548, + 192 + ], + "lines": [ + { + "bbox": [ + 366, + 72, + 548, + 192 + ], + "spans": [ + { + "bbox": [ + 366, + 72, + 548, + 192 + ], + "type": "image", + "image_path": "f1232b9edc89770ad4783313e3bb5aa6365e714acce722389cc2bfb575da4a05.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 193, + 413, + 204 + ], + "lines": [ + { + "bbox": [ + 374, + 193, + 413, + 204 + ], + "spans": [ + { + "bbox": [ + 374, + 193, + 413, + 204 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 374, + 193, + 413, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{SortP_M}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 193, + 468, + 203 + ], + "lines": [ + { + "bbox": [ + 430, + 193, + 468, + 203 + ], + "spans": [ + { + "bbox": [ + 430, + 193, + 468, + 203 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 430, + 193, + 468, + 203 + ], + "type": "inline_equation", + "content": "\\mathrm{SortP_L}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 486, + 191, + 530, + 202 + ], + "lines": [ + { + "bbox": [ + 486, + 191, + 530, + 202 + ], + "spans": [ + { + "bbox": [ + 486, + 191, + 530, + 202 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 486, + 191, + 530, + 202 + ], + "type": "inline_equation", + "content": "\\mathrm{SortP_{Mul}}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 279, + 291, + 335 + ], + "blocks": [ + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "text", + "content": "Figure 2. (a) The comprehensive LS3M framework integrates the global input strategy with adaptive pooling, the SortP for generating permutation matrices, and the S3M for feature enhancement. " + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "inline_equation", + "content": "(\\cdot)_M" + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "text", + "content": " denotes that we perform " + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "text", + "content": " process separately for each modality. (b) SortP " + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "text", + "content": " means the learnable sorting process applied to each modality. (c) SortP " + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "text", + "content": " denotes obtaining the larger permutation matrix for longer sequences. (d) SortP " + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "inline_equation", + "content": "Mul" + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "text", + "content": " refers to the generation of the multi-modal permutation matrix." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 279, + 291, + 335 + ], + "lines": [ + { + "bbox": [ + 59, + 279, + 291, + 335 + ], + "spans": [ + { + "bbox": [ + 59, + 279, + 291, + 335 + ], + "type": "image", + "image_path": "5313ba132fb3d5d77cfeb1dd084eb9ff3473d6972524f05dfe8740dc4d12ffe9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 342, + 269, + 354 + ], + "lines": [ + { + "bbox": [ + 81, + 342, + 269, + 354 + ], + "spans": [ + { + "bbox": [ + 81, + 342, + 269, + 354 + ], + "type": "text", + "content": "Figure 3. The structure of the proposed S3M block." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 376, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 483 + ], + "type": "text", + "content": "broader context crucial for accurately understanding complex conditions like brain tumors. This is especially important in the case of intra-tumor heterogeneity, which requires a wider context to characterize the tumor type [35, 40]. Moreover, in scenarios with missing modalities, it is critical to utilize all available data from the entire image to compensate for the missing information. Thus, incorporating broader context without significant computational overhead is essential." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 491, + 280, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 280, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 280, + 504 + ], + "type": "text", + "content": "4.2. Challenges of Existing Mamba Approaches" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 509, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 532 + ], + "type": "text", + "content": "Mamba is primarily designed for 1D sequences, with characteristics that include:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 533, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 581 + ], + "type": "text", + "content": "Mamba Characteristic 1: the current output token " + }, + { + "bbox": [ + 55, + 533, + 295, + 581 + ], + "type": "inline_equation", + "content": "y_{k}" + }, + { + "bbox": [ + 55, + 533, + 295, + 581 + ], + "type": "text", + "content": " is limited to modeling relationships solely with previously scanned tokens " + }, + { + "bbox": [ + 55, + 533, + 295, + 581 + ], + "type": "inline_equation", + "content": "x_{i}, i \\in \\{0,1,\\dots ,k - 1,k\\}" + }, + { + "bbox": [ + 55, + 533, + 295, + 581 + ], + "type": "text", + "content": ", as detailed in Eq. (2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 582, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 616 + ], + "type": "text", + "content": "Mamba Characteristic 2: the influence of distant tokens diminishes significantly due to the weight " + }, + { + "bbox": [ + 55, + 582, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}}^{i - 1}" + }, + { + "bbox": [ + 55, + 582, + 295, + 616 + ], + "type": "text", + "content": ", as explained in Eq. (3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "To mitigate these limitations, existing Mamba-based methods [33, 38, 50, 65] introduce forward and backward scanning strategies, which enable the current token to model relationships with both preceding and subsequent tokens. These methods are more suitable for vision applications, where local dependencies can often provide sufficient context. However, in the case of missing modalities, it is insufficient to rely solely on local context for tumor identi" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 279, + 553, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 553, + 362 + ], + "type": "text", + "content": "fication. Tumor heterogeneity often requires capturing information from distant regions of the MRI to fully understand the tumor's characteristics, as highlighted by prior studies [35, 40]. For example, in the segmentation of the whole tumor (WT) with only T1ce modality, local features may be insufficient, and distant correlations are necessary to improve segmentation accuracy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 363, + 554, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 554, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 554, + 447 + ], + "type": "text", + "content": "Due to Mamba Characteristic 2, the influence of distant tokens diminishes over sequence steps, which is problematic in 3D MRI segmentation tasks. Therefore, a dynamic mechanism is required to ensure that the most relevant tokens—both spatially close and semantically related over long distances—are positioned closer to the target token, facilitating more effective modeling." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 448, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 604 + ], + "type": "text", + "content": "Although LocalMamba [22] uses a Z-scan to ensure local relationships in 2D, this approach cannot guarantee the preservation of 3D neighborhood relationships, such as the 27-neighborhood in 3D space. Even with axial scanning added to horizontal and vertical scanning, the relationships between distant tokens in 3D space are still not adequately preserved. Consequently, there is a need for a solution that dynamically reorders the tokens for each 3D MRI scan, ensuring that all potentially relevant tokens—whether spatially close or far but semantically related—are placed before the current token. This dynamic reordering will allow the subsequent S3M block to model these relationships more effectively, improving the segmentation accuracy." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 612, + 460, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 460, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 460, + 624 + ], + "type": "text", + "content": "4.3. Learnable Sorting Process" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "Traditional Non-differentiable Sorting Process. A pivotal aspect of LS3M involves addressing the non-differentiability of traditional sorting mechanisms. The patient-specific dynamic sorting process is initially driven by a predictive score " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{L}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": ", assigned to each token in the sequence " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": ". These scores dictate the token positions in the sorted sequence " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathbf{S}}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": ", subsequently processed by the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25985" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "Mamba block. However, the traditional sorting function, which maps these continuous input scores to discrete token indices, interrupts the gradient flow necessary for end-to-end training of the network." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 159 + ], + "type": "text", + "content": "The Learnable Sorting Process. The learnable sorting process is represented as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 86, + 167, + 157, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 167, + 157, + 180 + ], + "spans": [ + { + "bbox": [ + 86, + 167, + 157, + 180 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {P}} = \\operatorname {S o r t N e t} (\\mathbf {f}),", + "image_path": "40499e637035c3c7e8a1046652c6ef9791633a20b2dcd6ea5633d94ee377f08a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 185, + 295, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 185, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 86, + 185, + 295, + 198 + ], + "type": "interline_equation", + "content": "\\mathbf {P} = \\text {S i n k h o r n} (\\text {G u m b e l - S o f t m a x} (\\hat {\\mathbf {P}}, \\tau), n), \\tag {8}", + "image_path": "8f99ff87b0466df1157c3516054b2bc77a2c2336b6d5e516d3cb0a1ef02bed43.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 201, + 228, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 201, + 228, + 213 + ], + "spans": [ + { + "bbox": [ + 86, + 201, + 228, + 213 + ], + "type": "interline_equation", + "content": "\\mathbf {X} = \\text {f l a t t e n} (\\mathbf {f}), \\quad \\mathbf {X} _ {\\mathbf {S}} = \\mathbf {P X}.", + "image_path": "283e1b077537cbdd189122d1ca19472da4a32bfaf892300bf001a49b909e568b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "text", + "content": "This enable end-to-end differentiable sorting. As shown in Fig. 2(b), firstly, our SortNet generates an estimated matrix " + }, + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}\\in \\mathbb{R}^{L\\times L}" + }, + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "text", + "content": " tailored to individual patients. It employs a convolutional layer with a kernel size of " + }, + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "inline_equation", + "content": "C\\times C\\times 3\\times 3\\times 3" + }, + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "text", + "content": " to capture spatial inductive biases, extracting information from the 26-neighborhood 3D spatial structure. After convolution, a SiLU activation function [20] is applied to enhance the network's representational capacity. The resulting feature map is split into " + }, + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_1" + }, + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_2" + }, + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_2" + }, + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "text", + "content": " being processed to compute a similarity map that captures long-range semantic correlations. A ReLU activation is then applied to ensure the non-negativity of " + }, + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}" + }, + { + "bbox": [ + 55, + 222, + 296, + 377 + ], + "type": "text", + "content": ", preparing it for subsequent processing by the Sinkhorn operator." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 378, + 296, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 497 + ], + "type": "text", + "content": "Secondly, we use the Gumbel-Softmax and Sinkhorn operators to generate a permutation matrix " + }, + { + "bbox": [ + 55, + 378, + 296, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 378, + 296, + 497 + ], + "type": "text", + "content": ", where each row and column contains exactly one 1 with all other entries being 0s. The estimated matrix " + }, + { + "bbox": [ + 55, + 378, + 296, + 497 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 497 + ], + "type": "text", + "content": " is sliced row-wise, with each row being passed through the Gumbel-Softmax function. By adjusting the temperature parameter " + }, + { + "bbox": [ + 55, + 378, + 296, + 497 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 378, + 296, + 497 + ], + "type": "text", + "content": ", the output satisfies the one-hot condition for each row. The resulting rows are then concatenated and passed through the Sinkhorn operator, which iteratively adjusts the rows and columns of the matrix, yielding the final permutation matrix " + }, + { + "bbox": [ + 55, + 378, + 296, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 378, + 296, + 497 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "content": "Finally, the feature map " + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "content": " is flattened into a sequence " + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "content": " is applied to " + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "content": " to perform the sorting operation. This process can be viewed as permuting the rows of an " + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "inline_equation", + "content": "L\\times L" + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "content": " identity matrix according to the sorting procedure. Since all of these operations are differentiable, the entire sorting process is also differentiable, which ensures that each token is optimally positioned in the sequence for subsequent processing by our S3M block." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 600, + 293, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 293, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 293, + 613 + ], + "type": "text", + "content": "4.4. Handling Longer and Multi-modal Sequences" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "Handling Longer Sequence. Extracting multi-scale information from available modalities requires generating larger permutation matrices to accommodate longer sequences corresponding to high-resolution feature maps. However, this process introduces quadratic computational complexity due to the matrix multiplications involved. To address this, we propose an efficient alternative: bilinear interpolation to scale the permutation matrix, eliminating" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "the need for SortNet. The scaled matrix is then processed through ReLU activation, Gumbel-Softmax, and Sinkhorn to generate the permutation matrix " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{Large}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " for the extended sequence, as represented in Fig. 2(c)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "text", + "content": "Handling Multi-modal Sequences. To effectively leverage inter-modal correlations, we aim to generate a permutation matrix that accounts for scenarios with missing modalities. We dynamically fuse the individual permutation matrices of each modality—" + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{FLAIR}}" + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{T1ce}}" + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{T1}}" + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{T2}}" + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "text", + "content": " with a multi-modal code, represented by " + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "inline_equation", + "content": "h = \\{h_1, h_2, h_3, h_4\\}" + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "inline_equation", + "content": "h_i \\in \\{0, 1\\}" + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "text", + "content": " indicates whether the corresponding modality is available. This fusion produces a comprehensive multi-modal estimated matrix, " + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_{\\mathrm{Mul}}" + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "text", + "content": ". We then apply a ReLU activation followed by Gumbel-Softmax and Sinkhorn operations to generate the final multi-modal permutation matrix " + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{Mul}}" + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "text", + "content": ", as shown in Fig. 2(d). Before reordering the multi-modal sequence, each modal sequence is initially projected into a unified representation space. The sequences are then sorted according to " + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{Mul}}" + }, + { + "bbox": [ + 313, + 150, + 555, + 342 + ], + "type": "text", + "content": ", ensuring alignment based on inter-modal relevance and coherence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 357, + 515, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 515, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 515, + 370 + ], + "type": "text", + "content": "4.5. Series State Space Model (S3M) Block" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "content": "Due to the insufficient information of missing modalities, it becomes crucial to exhaustively mine information from all available modalities. We introduce the S3M model for local and long-range dependency modeling, inspired by the success of Mamba." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 441, + 554, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 499 + ], + "type": "text", + "content": "As shown in Fig. 3, our S3M block connects two unidirectional Mamba-like blocks, " + }, + { + "bbox": [ + 313, + 441, + 554, + 499 + ], + "type": "inline_equation", + "content": "F_{Fwd}" + }, + { + "bbox": [ + 313, + 441, + 554, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 441, + 554, + 499 + ], + "type": "inline_equation", + "content": "F_{Bwd}" + }, + { + "bbox": [ + 313, + 441, + 554, + 499 + ], + "type": "text", + "content": ", in a series connection but in opposite directions, together with a Channel Attention (FCA) block [21], which can be expressed as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 358, + 515, + 553, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 515, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 358, + 515, + 553, + 529 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\mathbf {S 3 M}} = \\mathrm {F} _ {\\mathrm {C A}} \\left(\\mathrm {F} _ {\\mathrm {B w d}} \\left(\\mathrm {F} _ {\\mathrm {F w d}} \\left(\\mathbf {f} _ {\\mathbf {I n}}, \\mathbf {P}\\right)\\right)\\right). \\tag {9}", + "image_path": "f4483f270e2ab541a99c64e77010b268183aa30a9c165b40214981586fe57d02.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 543, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 554, + 579 + ], + "type": "text", + "content": "Before processing, the feature " + }, + { + "bbox": [ + 313, + 543, + 554, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{In}}" + }, + { + "bbox": [ + 313, + 543, + 554, + 579 + ], + "type": "text", + "content": " is reshaped into a sequence and sorted by multiplying it with a soft permutation matrix " + }, + { + "bbox": [ + 313, + 543, + 554, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 543, + 554, + 579 + ], + "type": "text", + "content": " generated through the learnable sorting process." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Each Mamba-like block comprises three branches: a basic SSM branch, a gated weighting branch, and a residual connection branch. In the SSM branch, the input feature " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{In}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " undergoes a linear projection to expand its channels, followed by depth-wise convolution and SiLU activation, producing " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{Mid}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " and the SSM parameters " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\overline{\\mathbf{A}},\\overline{\\mathbf{B}},\\mathbf{C}\\}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ". They are sent to the Forward/Backward SSM module, followed by layer normalization to obtain " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{SSM}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ". The input feature " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{In}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " is processed through a linear layer and SiLU activation to generate a gated weight feature map " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{Gate}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ". Combining these with the residual connection branch, the output of the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25986" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 144, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 144, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 144, + 83 + ], + "type": "text", + "content": "Mamba-like block is:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 98, + 297, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 98, + 297, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 98, + 297, + 152 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {f} _ {\\mathbf {S S M}} = \\mathrm {L N} (\\mathrm {S S M} (\\mathrm {S i L U} (\\mathrm {D W - C o n v} (\\mathrm {L i n e a r} (\\mathrm {L N} (\\mathbf {f} _ {\\mathbf {I n}})))))), \\\\ \\mathbf {f} _ {\\text {G a t e}} = \\operatorname {S i L U} \\left(\\operatorname {L i n e a r} \\left(\\mathbf {f} _ {\\text {I n}}\\right)\\right), \\\\ \\mathbf {f} _ {\\text {O u t}} = \\operatorname {L i n e a r} \\left(\\mathbf {f} _ {\\text {S S M}} \\odot \\mathbf {f} _ {\\text {G a t e}}\\right) + \\mathbf {f} _ {\\text {I n}}. \\tag {10} \\\\ \\end{array}", + "image_path": "8e780604c557f11b1b04d104aa459de2242eb2ed889a8c0352578e38ce6e66fa.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 155, + 296, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 296, + 297 + ], + "type": "text", + "content": "For further channel-wise interactions, we introduce the Channel Attention block. Our S3M block, when compared to the recent Vim block [65] and SS2D block [33], offers two main advantages: (1) The series connection in S3M ensures that information from the forward SSM block is integrated into subsequent tokens, which is then relayed back to earlier tokens in the backward SSM block. This contrasts with parallel connections in Vim block and SS2D block, where token sequences receive the same information level. (2) The integration of the permutation matrix in S3M preserves both spatial proximity and long-range semantic correlations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "text", + "content": "For multi-modal fusion, each modal sequence is first projected into a unified representation space and then concatenated along the channel dimension. The multi-modal sequences " + }, + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{Mul}}" + }, + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "text", + "content": " are sorted by " + }, + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{Mul}}" + }, + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "text", + "content": ", which are further delivered into our S3M for building inter-modal correlations. Specifically, the S3M utilizes the Mamba-like block where the information from each token in " + }, + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{Mul}}" + }, + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "text", + "content": " is compressed into SSM parameters—" + }, + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}}_{Mul}" + }, + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{B}}_{Mul}" + }, + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{Mul}" + }, + { + "bbox": [ + 55, + 301, + 296, + 482 + ], + "type": "text", + "content": ", which contain the compact multi-modal information along their hidden state dimension. Then, the Forward/Backward SSM module interact across both the channel dimensions of the multi-modal sequences and the state dimensions of the SSM parameters, representing the modality interaction between the origin format and the compact format, thereby, resulting in an efficient fusion that markedly improves segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 180, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 180, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 180, + 511 + ], + "type": "text", + "content": "4.6. Global Input Strategy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 518, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 651 + ], + "type": "text", + "content": "We introduce a global input strategy that integrates the full volumes information into the analysis of cropped volumes without incurring excessive computational costs, as detailed in Fig. 2. The full volumes, augmented with a relative position mask, undergoes a series of transformations through convolutional layers and non-linear activations. They are then resized to match the dimensions of the cropped volumes using adaptive average pooling. The resulting feature map, which combines coarse global information with the fine details of the cropped volume, is then fed into our LS3M for segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "The relative position mask, specifically designed for the full volumes, is encoded at positions corresponding to the cropped region. This design emphasizes the relevant area within the larger volume, guiding the network's attention to the regions slated for segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 484, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 484, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 484, + 83 + ], + "type": "text", + "content": "4.7. Network Architecture and Loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 89, + 553, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 184 + ], + "type": "text", + "content": "Our framework, depicted in Fig. 2, processes inputs of four modalities. Each modality includes both a cropped volume for segmentation and its corresponding global volume. The global volumes are first processed through our global input strategy. The yielded feature maps are then concatenated with their respective cropped volumes and passed through individual encoders " + }, + { + "bbox": [ + 313, + 89, + 553, + 184 + ], + "type": "inline_equation", + "content": "E(\\cdot)_M" + }, + { + "bbox": [ + 313, + 89, + 553, + 184 + ], + "type": "text", + "content": " to extract hierarchical intra-modal features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 186, + 553, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 245 + ], + "type": "text", + "content": "The features from each modality are transformed into sequences and fed into " + }, + { + "bbox": [ + 313, + 186, + 553, + 245 + ], + "type": "inline_equation", + "content": "\\mathrm{SortP}_M" + }, + { + "bbox": [ + 313, + 186, + 553, + 245 + ], + "type": "text", + "content": ", the learnable sorting process, which generates a modality-specific permutation matrix, " + }, + { + "bbox": [ + 313, + 186, + 553, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{M}}" + }, + { + "bbox": [ + 313, + 186, + 553, + 245 + ], + "type": "text", + "content": ". This matrix reorders the sequences before they are input into the S3M tailored for each modality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 246, + 553, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 341 + ], + "type": "text", + "content": "At higher resolutions, for each modality, a larger permutation matrix, " + }, + { + "bbox": [ + 313, + 246, + 553, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{Large}}" + }, + { + "bbox": [ + 313, + 246, + 553, + 341 + ], + "type": "text", + "content": ", is obtained by bilinear upsampling. For multi-modal integration, the comprehensive multi-modal permutation matrix, " + }, + { + "bbox": [ + 313, + 246, + 553, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{Mul}}" + }, + { + "bbox": [ + 313, + 246, + 553, + 341 + ], + "type": "text", + "content": ", is generated by combining the modality-specific permutation matrix " + }, + { + "bbox": [ + 313, + 246, + 553, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{M}}" + }, + { + "bbox": [ + 313, + 246, + 553, + 341 + ], + "type": "text", + "content": " with a multi-modal code " + }, + { + "bbox": [ + 313, + 246, + 553, + 341 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 246, + 553, + 341 + ], + "type": "text", + "content": ". These permutation matrices play a crucial role in integrating the sequences within the S3M, ultimately enhancing segmentation accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 399 + ], + "type": "text", + "content": "The final segmentation results are obtained after hierarchical fusion and processing of the multi-modal features. Simulated missing modalities are represented by substituting zero-value feature maps prior to concatenating individual modal features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 401, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 553, + 437 + ], + "type": "text", + "content": "Following [8, 55], we employ the weighted cross-entropy loss " + }, + { + "bbox": [ + 313, + 401, + 553, + 437 + ], + "type": "inline_equation", + "content": "Loss_{WCE}" + }, + { + "bbox": [ + 313, + 401, + 553, + 437 + ], + "type": "text", + "content": " and the Dice loss " + }, + { + "bbox": [ + 313, + 401, + 553, + 437 + ], + "type": "inline_equation", + "content": "Loss_{Dice}" + }, + { + "bbox": [ + 313, + 401, + 553, + 437 + ], + "type": "text", + "content": ". The total loss is " + }, + { + "bbox": [ + 313, + 401, + 553, + 437 + ], + "type": "inline_equation", + "content": "Loss = Loss_{WCE} + Loss_{Dice}" + }, + { + "bbox": [ + 313, + 401, + 553, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 449, + 394, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 449, + 394, + 462 + ], + "spans": [ + { + "bbox": [ + 314, + 449, + 394, + 462 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 468, + 502, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 502, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 502, + 480 + ], + "type": "text", + "content": "5.1. Dataset and implementation details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 486, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 616 + ], + "type": "text", + "content": "We perform our experiments on BraTS2018 and BraTS2020 from the Multi-modal Brain Tumor Segmentation Challenge [37], which align with that of previous studies [32, 45]. The two datasets contain 285 and 369 cases, respectively, with publicly available ground truth. Consistent with [8], BraTS2020 is randomly split into training, validation, and test sets with a ratio of " + }, + { + "bbox": [ + 313, + 486, + 554, + 616 + ], + "type": "inline_equation", + "content": "219:50:100" + }, + { + "bbox": [ + 313, + 486, + 554, + 616 + ], + "type": "text", + "content": ". BraTS2018 is split into " + }, + { + "bbox": [ + 313, + 486, + 554, + 616 + ], + "type": "inline_equation", + "content": "199:29:57" + }, + { + "bbox": [ + 313, + 486, + 554, + 616 + ], + "type": "text", + "content": " and includes a three-fold validation. These splits are consistently used across all experiments, ensuring that all methods are evaluated under the same conditions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": "Each case of the datasets has four different modalities, i.e., FLAIR, T1ce, T1, and T2 modalities. These modalities are characterized by a volume size of " + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "inline_equation", + "content": "240 \\times 240 \\times 155" + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": ", and they capture various properties of brain tumor subregions: GD-enhancing tumor (ET), peritumoral edema (ED), and the necrotic and non-enhancing tumor core (NCR/NET). These subregions of brain tumors are grouped into three nested subregions: the whole tumor (WT), the tumor core" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25987" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 79, + 511, + 258 + ], + "blocks": [ + { + "bbox": [ + 97, + 79, + 511, + 258 + ], + "lines": [ + { + "bbox": [ + 97, + 79, + 511, + 258 + ], + "spans": [ + { + "bbox": [ + 97, + 79, + 511, + 258 + ], + "type": "image", + "image_path": "b82111d5aebc7e81ac1b261bd70d1874763f1bbce24cedbd0c0a6e84d7623d57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 261, + 488, + 272 + ], + "lines": [ + { + "bbox": [ + 121, + 261, + 488, + 272 + ], + "spans": [ + { + "bbox": [ + 121, + 261, + 488, + 272 + ], + "type": "text", + "content": "Figure 4. Segmentation results of different methods with various available modalities on BraTS2020." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 293, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 295, + 327 + ], + "type": "text", + "content": "(TC), and the enhancing tumor (ET). We use Dice coefficient, " + }, + { + "bbox": [ + 54, + 293, + 295, + 327 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 54, + 293, + 295, + 327 + ], + "type": "text", + "content": " Hausdorff distance, sensitivity, and specificity to evaluate our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 328, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 328, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 296, + 437 + ], + "type": "text", + "content": "We conduct our experiments using the PyTorch framework version 1.13.0. The training process is carried out on four NVIDIA A800 GPUs, spanning 500 epochs, with a batch size of four volumes. We utilize the Adam optimizer with an initial learning rate set to " + }, + { + "bbox": [ + 54, + 328, + 296, + 437 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 54, + 328, + 296, + 437 + ], + "type": "text", + "content": " and a weight decay of " + }, + { + "bbox": [ + 54, + 328, + 296, + 437 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 54, + 328, + 296, + 437 + ], + "type": "text", + "content": ". During the training phase, the input images are randomly cropped to dimensions of " + }, + { + "bbox": [ + 54, + 328, + 296, + 437 + ], + "type": "inline_equation", + "content": "80 \\times 80 \\times 80" + }, + { + "bbox": [ + 54, + 328, + 296, + 437 + ], + "type": "text", + "content": ". Data augmentation techniques include random rotations, flips, and intensity adjustments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 444, + 257, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 257, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 257, + 456 + ], + "type": "text", + "content": "5.2. Comparisons with the State-of-the-art" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 461, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 461, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 461, + 295, + 555 + ], + "type": "text", + "content": "To evaluate the effectiveness of our method, we compare our method with nine state-of-the-art methods on different cases with missing modalities. The involved methods contain HeMiS [18], U-HVED [10], RFNet [8], UNet-MFI [59], mmFormer [55], MMANet [48], ShaSpec [45], M3AE [32], and GSS [39]. For a fair comparison, all methods are trained under their recommended hyper-parameters within the same dataset split." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 557, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 557, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 54, + 557, + 295, + 665 + ], + "type": "text", + "content": "As illustrated in Tab. 1, our method achieves better performance for most combinations of missing modalities. We achieve improvements of " + }, + { + "bbox": [ + 54, + 557, + 295, + 665 + ], + "type": "inline_equation", + "content": "1.22\\%" + }, + { + "bbox": [ + 54, + 557, + 295, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 557, + 295, + 665 + ], + "type": "inline_equation", + "content": "1.31\\%" + }, + { + "bbox": [ + 54, + 557, + 295, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 557, + 295, + 665 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 54, + 557, + 295, + 665 + ], + "type": "text", + "content": " over the second-ranked method on the average DSC for WT, TC, and ET. We also provide a visualization comparison in Fig. 4, illustrating that our method yields more accurate segmentation results in different combinations of modalities. More comparisons, including statistical results, additional metrics and model complexity, are provided in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 672, + 149, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 149, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 149, + 685 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "We conduct ablation experiments on BraTS2020 to assess the effectiveness of specific designs. The baseline is served" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 334, + 293, + 533, + 407 + ], + "blocks": [ + { + "bbox": [ + 334, + 293, + 533, + 407 + ], + "lines": [ + { + "bbox": [ + 334, + 293, + 533, + 407 + ], + "spans": [ + { + "bbox": [ + 334, + 293, + 533, + 407 + ], + "type": "image", + "image_path": "f4a497b3db952e301a3611662629cffe949df24fc7bf7e324ee03d792af0b11a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 413, + 517, + 425 + ], + "lines": [ + { + "bbox": [ + 350, + 413, + 517, + 425 + ], + "spans": [ + { + "bbox": [ + 350, + 413, + 517, + 425 + ], + "type": "text", + "content": "Figure 5. Effect of hyper-parameters " + }, + { + "bbox": [ + 350, + 413, + 517, + 425 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 350, + 413, + 517, + 425 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 350, + 413, + 517, + 425 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 350, + 413, + 517, + 425 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 447, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 472 + ], + "type": "text", + "content": "as a simplified version of LS3M without GIS, SortP, and S3M components." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 491, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 586 + ], + "type": "text", + "content": "Effect of Scanning Strategies. Relative to our baseline, which employs no specific scanning strategies, adding bidirectional, and fourth-directional (cross scan) methods demonstrated enhanced effectiveness, attributed to their ability to maintain directional spatial coherence. Notably, our sorted scanning strategy, as detailed in Tab. 2, achieved the highest Dice scores by preserving spatial inductive biases and semantic correlations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "Effect of Proposed Components. As shown in Tab. 3, integrating a global input significantly improves baseline performance by providing larger spatial contexts. The addition of a relative position mask further enhances results by focusing on key areas within the broader input context. The average performance improves " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "1.84\\%" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": " with S3M. The complete LS3M architecture, after including SortP, achieves superior performance, effectively exploring intra-modal and inter-modal relations." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25988" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 530, + 357 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 530, + 357 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 530, + 357 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 530, + 357 + ], + "type": "table", + "html": "
MFLAIR●○○○●●●○○○●●●○●AVG
T1ce○●○○●○○●○○●●○●●AVG
T1○○●○○○○●○●●○●●●●
T2○○○●○○○●○●○●●●●●
WTHeMiS71.6067.7168.9668.1969.1768.6769.8369.0169.7869.4070.2171.2870.7371.5872.0669.88
U-HVED69.8546.8246.7754.0361.4558.2564.5062.9165.7664.2966.9969.7068.3870.3571.4162.76
RFNet86.4277.3476.4686.2189.5589.3089.3581.0087.4587.9590.3990.2090.4288.5990.7786.76
UNet-MFI82.2773.1872.1082.4583.6484.3484.8577.3083.4483.5285.4585.7085.8584.2084.9382.21
mmFormer82.4074.2574.3783.0784.5484.6185.8277.9884.0584.0085.3486.1186.2284.6486.3882.92
MMANet83.2164.4664.3582.8886.7487.6487.4770.9885.6485.3387.9289.3787.9786.3889.2582.64
ShaSpec84.1070.3670.2383.9588.6589.0788.7076.0286.2187.1589.4789.4390.0787.6290.1184.74
M3AE86.5373.8576.7186.0989.4889.3889.2578.1187.3787.2089.9990.1890.4288.6190.5686.25
GSS*86.3672.4573.3386.7689.5190.6590.9178.3888.8790.0591.1493.2590.6789.4393.2887.00
Ours88.6678.2479.3287.2991.0090.5791.4582.5889.1888.8891.5291.8791.4589.7592.1588.22
TCHeMiS53.4351.4151.5651.1151.7051.0851.8551.8852.3551.5152.9553.7652.9754.3855.0352.46
U-HVED34.6235.5127.3037.6742.1538.2643.4144.9347.5344.9749.1351.3049.4052.7254.1743.53
RFNet65.0482.3764.3168.4784.6971.4572.6283.1584.0672.1184.7184.7074.2884.1184.7477.39
UNet-MFI63.9477.6359.3868.0579.9268.2370.7277.6180.0970.2180.0380.9471.4080.7581.2874.01
mmFormer66.1977.9661.1769.1880.3669.5871.5579.9380.7970.9080.1881.3172.0281.1281.2274.90
MMANet65.7974.0557.4770.2185.2172.7372.6580.9985.2471.8885.8785.6874.4886.3986.6376.99
ShaSpec66.0677.3459.5768.2985.0272.8772.9580.8484.4672.2885.8385.5774.7485.4786.5077.19
M3AE68.0481.3966.0070.2782.0173.8274.9582.3983.0172.5482.4483.0675.0984.0684.4077.56
GSS*67.3378.4361.6770.1084.8873.4574.1782.2086.1274.1886.5088.3974.3386.2888.6778.45
Ours71.9883.6166.6772.3086.3275.2576.8784.0985.1874.3286.0585.8476.8785.8385.8679.76
ETHeMiS43.7742.4141.5941.4541.8340.2941.1942.0842.3941.0043.6744.1642.9545.2746.3342.69
U-HVED12.8824.947.2724.2630.0221.9529.4033.6436.1832.1239.3940.9138.0943.1845.3330.64
RFNet40.4774.2737.5143.5976.4543.8146.9975.2273.9446.3777.0176.3848.9576.3876.6460.93
UNet-MFI39.7069.4229.3846.0070.1340.0648.6969.2572.3245.7171.2870.8846.5572.0071.4157.52
mmFormer40.4768.9133.9745.6169.8143.6348.0971.1070.7245.9270.0871.6048.3870.6571.3658.02
MMANet36.4066.0728.2241.9170.4341.5044.2368.8771.8643.5871.9471.9844.6071.8172.4356.39
ShaSpec38.9266.9331.7542.5870.8543.2245.6269.9271.0444.2172.0871.0946.2271.8172.0257.22
M3AE40.4972.4339.9345.9774.6643.2047.3075.4276.8146.6375.9477.0848.1977.4078.0061.30
GSS*42.0369.4635.3045.6074.2047.7849.3374.4174.7848.9476.1276.3250.0276.9875.9561.15
Ours40.3477.6836.5745.3079.1745.0751.5379.4978.6947.0480.7580.5151.5381.5380.9463.60
", + "image_path": "e35e2c95cc0556d590dddee5a6cfe3d4a6656c4d7bde8a0fbb91919108fe36fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 406, + 288, + 441 + ], + "blocks": [ + { + "bbox": [ + 55, + 365, + 553, + 388 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 553, + 388 + ], + "type": "text", + "content": "Table 1. Performance comparison (DSC%) with SOTA methods on BraTS2020. Available and missing modalities are denoted by " + }, + { + "bbox": [ + 55, + 365, + 553, + 388 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 365, + 553, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 365, + 553, + 388 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 55, + 365, + 553, + 388 + ], + "type": "text", + "content": ", respectively. \\* demonstrates that its codes and results are reproduced by ourselves." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 406, + 288, + 441 + ], + "lines": [ + { + "bbox": [ + 63, + 406, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 63, + 406, + 288, + 441 + ], + "type": "table", + "html": "
DiceBaseline+ Uni-D+ Bi-D+ Fourth-D+ Random+ Our sorted
WT85.4385.2085.5586.1384.2386.94
TC76.1775.6976.6777.5474.0878.01
ET58.0257.9858.4360.7156.7962.06
", + "image_path": "ffb759749bc774e375d76a0d5d30a1b0de41b02b722dac7454dba5f7e94912bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 63, + 483, + 288, + 517 + ], + "blocks": [ + { + "bbox": [ + 55, + 449, + 295, + 471 + ], + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 471 + ], + "type": "text", + "content": "Table 2. Effect of scanning strategies, '-D' denotes directional scanning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 483, + 288, + 517 + ], + "lines": [ + { + "bbox": [ + 63, + 483, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 63, + 483, + 288, + 517 + ], + "type": "table", + "html": "
DiceBaseline+GIS (w/o mask)+GIS+GIS + S3M+GIS + SortP + S3M
WT85.4385.9986.2387.1788.22
TC76.1776.7877.0679.0679.76
ET58.0259.0160.3562.9463.60
", + "image_path": "74fd5a7d189e124feb288696c564304f197e8bb0b2b74122478edd0bfa90bd7f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "text", + "content": "Table 3. Effect of proposed components, 'GIS' denotes Global Input Strategy." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "text", + "content": "Effects of hyper-parameters of temperature " + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "text", + "content": " and iterative times " + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "text", + "content": ". Adjusting " + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "text", + "content": " reveals their critical roles in influencing the performance of our LS3M, as shown in Fig. 5. The parameter " + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "text", + "content": " modulates the smoothness of the permutation matrix, with " + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\tau = 1" + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "text", + "content": " yielding optimal results when " + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "text", + "content": " is fixed. Conversely, increasing " + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "text", + "content": " enhances the precision of the permutation matrix, thus improving segmentation outcomes. Balancing computational efficiency and precision, we select " + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "inline_equation", + "content": "n = 10" + }, + { + "bbox": [ + 55, + 569, + 295, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Additional ablation experiments are provided in the appendix. These experiments further validate the importance of the proposed components and their contribution to the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 408, + 404, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 408, + 404, + 419 + ], + "spans": [ + { + "bbox": [ + 314, + 408, + 404, + 419 + ], + "type": "text", + "content": "model's performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 449, + 388, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 449, + 388, + 460 + ], + "spans": [ + { + "bbox": [ + 314, + 449, + 388, + 460 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "In this work, we introduce the Learnable Sorting State Space Model (LS3M), a novel framework for brain tumor segmentation with missing modalities, which aims to maximizing mining information from available modalities. LS3M replaces traditional non-differentiable sorting with a differentiable permutation matrix, preserving 3D spatial information and enabling more effective modeling. The framework adeptly scales these matrices to accommodate longer sequences, capturing multi-scale information. It also dynamically integrates these matrices for robust multi-modal sequence fusion, effectively managing various scenarios of missing modalities. The S3M block captures both intra-modal and inter-modal correlations, enhancing segmentation accuracy. With a global input strategy, LS3M provides comprehensive contextual insights, surpassing existing segmentation methods. Experimental results on BraTS2018 and BraTS2020 demonstrate that LS3M outperforms state-of-the-art. Our results highlight LS3M's potential in advancing medical imaging and healthcare diagnostics." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25989" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 148 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 148 + ], + "type": "text", + "content": "This work was in part supported by the National Natural Science Foundation of China under grants 62032006 and 62021001, and the Anhui Province Key Laboratory of Biomedical Imaging and Intelligent Processing under grants 23YGXT001." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 171, + 115, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 171, + 115, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 115, + 183 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 191, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 191, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 191, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 61, + 191, + 295, + 223 + ], + "type": "text", + "content": "[1] Ryan Prescott Adams and Richard S Zemel. Ranking via sinkhorn propagation. arXiv preprint arXiv:1106.1925, 2011.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 225, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 225, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 225, + 296, + 280 + ], + "type": "text", + "content": "[2] Reza Azad, Nika Khosravi, and Dorit Merhof. Smu-net: Style matching u-net for brain tumor segmentation with missing modalities. In International Conference on Medical Imaging with Deep Learning, pages 48-62. PMLR, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 282, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 282, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 61, + 282, + 295, + 315 + ], + "type": "text", + "content": "[3] Ali Behrouz and Farnoosh Hashemi. Graph mamba: Towards learning on graphs with state space models. arXiv preprint arXiv:2402.08678, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 317, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 317, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 61, + 317, + 295, + 372 + ], + "type": "text", + "content": "[4] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In Computer Vision-ECCV 2022 Workshops, pages 205-218. Springer, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 373, + 296, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 296, + 428 + ], + "type": "text", + "content": "[5] Cheng Chen, Qi Dou, Yueming Jin, Hao Chen, Jing Qin, and Pheng-Ann Heng. Robust multimodal brain tumor segmentation via feature disentanglement and gated fusion. In Medical Image Computing and Computer-Assisted Intervention-MICCAI, pages 447-456. Springer, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 430, + 296, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 296, + 483 + ], + "type": "text", + "content": "[6] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 520 + ], + "type": "text", + "content": "[7] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 521, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 296, + 576 + ], + "type": "text", + "content": "[8] Yuhang Ding, Xin Yu, and Yi Yang. Rfnet: Region-aware fusion network for incomplete multi-modal brain tumor segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3975-3984, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 578, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 296, + 621 + ], + "type": "text", + "content": "[9] Yuhang Ding, Liulei Li, Wenguan Wang, and Yi Yang. Clustering propagation for universal medical image segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "text", + "content": "[10] Reuben Dorent, Samuel Joutard, Marc Modat, Sébastien Ourselin, and Tom Vercauteren. Hetero-modal variational encoder-decoder for joint modality completion and segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI, pages 74-82. Springer, 2019. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[11] Zhengcong Fei, Mingyuan Fan, Changqian Yu, and Junshi Huang. Scalable diffusion models with state space backbone. arXiv preprint arXiv:2402.05608, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[12] Jiannan Ge, Lingxi Xie, Hongtao Xie, Pandeng Li, Xiaopeng Zhang, Yongdong Zhang, and Qi Tian. Alignzeg: Mitigating objective misalignment for zero-shot semantic segmentation. In European Conference on Computer Vision, pages 142-161. Springer, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 554, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 160 + ], + "type": "text", + "content": "[13] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "text", + "content": "[14] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "text", + "content": "[15] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 316 + ], + "type": "text", + "content": "[16] Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In *Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: 7th International Workshop, BrainLes* 2021, Held in Conjunction with MICCAI 2021, Virtual Event, pages 272-284. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 317, + 554, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 554, + 380 + ], + "type": "text", + "content": "[17] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 574-584, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "type": "text", + "content": "[18] Mohammad Havaei, Nicolas Guizard, Nicolas Chapados, and Yoshua Bengio. Hemis: Hetero-modal image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI, pages 469-477. Springer, 2016. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 427, + 554, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 554, + 469 + ], + "type": "text", + "content": "[19] Xuanhua He, Ke Cao, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Pan-mamba: Effective pan-sharpening with state space model. arXiv preprint arXiv:2402.12192, 2024.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 471, + 554, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 554, + 491 + ], + "type": "text", + "content": "[20] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 494, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 554, + 525 + ], + "type": "text", + "content": "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 526, + 554, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 554, + 568 + ], + "type": "text", + "content": "[22] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 570, + 554, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 554, + 603 + ], + "type": "text", + "content": "[23] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In International Conference on Learning Representations, 2016. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 604, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 554, + 658 + ], + "type": "text", + "content": "[24] Zeyu Jiang, Changxing Ding, Minfeng Liu, and Dacheng Tao. Two-stage cascaded u-net: 1st place solution to brats challenge 2019 segmentation task. In *Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries*, pages 231-241. Springer, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 659, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 554, + 713 + ], + "type": "text", + "content": "[25] Sanaz Karimijafarbigloo, Reza Azad, Amirhossein Kazerouni, Saeed Ebadollahi, and Dorit Merhof. Mmcformer: Missing modality compensation transformer for brain tumor segmentation. In Medical Imaging with Deep Learning, pages 1144-1162. PMLR, 2024. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25990" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[26] Jonghun Kim and Hyunjin Park. Adaptive latent diffusion model for 3d medical image to image translation: Multimodal magnetic resonance imaging study. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7604-7613, 2024. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "text", + "content": "[27] Dongwook Lee, Won-Jin Moon, and Jong Chul Ye. Assessing the importance of magnetic resonance contrasts using collaborative generative adversarial networks. Nature Machine Intelligence, 2(1):34-42, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "text", + "content": "[28] Ho Hin Lee, Shunxing Bao, Yuankai Huo, and Bennett A Landman. 3d ux-net: A large kernel volumetric convnet modernizing hierarchical transformer for medical image segmentation. arXiv preprint arXiv:2209.15076, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[29] Jiangyun Li, Wenxuan Wang, Chen Chen, Tianxiang Zhang, Sen Zha, Hong Yu, and Jing Wang. Transbtsv2: Wider instead of deeper transformer for medical image segmentation. arXiv preprint arXiv:2201.12785, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 330 + ], + "type": "text", + "content": "[30] Mingcheng Li, Dingkang Yang, Xiao Zhao, Shuaibing Wang, Yan Wang, Kun Yang, Mingyang Sun, Dongliang Kou, Ziyun Qian, and Lihua Zhang. Correlation-decoupled knowledge distillation for multimodal sentiment analysis with incomplete modalities. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "text", + "content": "[31] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 431 + ], + "type": "text", + "content": "[32] Hong Liu, Dong Wei, Donghuan Lu, Jinghan Sun, Liansheng Wang, and Yefeng Zheng. M3ae: Multimodal representation learning for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1657-1665, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 295, + 475 + ], + "type": "text", + "content": "[33] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 1, 3, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 477, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 295, + 510 + ], + "type": "text", + "content": "[34] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 511, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 295, + 544 + ], + "type": "text", + "content": "[35] Andriy Marusyk, Vanessa Almendro, and Kornelia Polyak. Intra-tumour heterogeneity: a looking glass for cancer? Nature reviews cancer, 12(5):323-334, 2012. 1, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 588 + ], + "type": "text", + "content": "[36] Gonzalo Mena, David Belanger, Scott Linderman, and Jasper Snoek. Learning latent permutations with gumbel-sinkhorn networks. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 655 + ], + "type": "text", + "content": "[37] Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging, 34(10):1993-2024, 2014. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 689 + ], + "type": "text", + "content": "[38] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977, 2024. 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[39] Yansheng Qiu, Delin Chen, Hongdou Yao, Yongchao Xu, and Zheng Wang. Scratch each other's back: Incomplete" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "multi-modal brain tumor segmentation via category aware group self-support learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21317-21326, 2023. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 183 + ], + "type": "text", + "content": "[40] Santiago Ramón y Cajal, Marta Sesé, Claudia Capdevila, Trond Aasen, Leticia De Mattos-Arruda, Salvador J Diaz-Cano, Javier Hernández-Losa, and Josep Castellvi. Clinical implications of intratumor heterogeneity: challenges and opportunities. Journal of Molecular Medicine, 98:161-177, 2020. 1, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 217 + ], + "type": "text", + "content": "[41] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "text", + "content": "[42] Dong She, Yueyi Zhang, Zheyu Zhang, Hebei Li, Zihan Yan, and Xiaoyan Sun. Eoformer: Edge-oriented transformer for brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 333-343. Springer, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 318 + ], + "type": "text", + "content": "[43] Jimmy TH Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "text", + "content": "[44] Chloe Wang, Oleksii Tsepa, Jun Ma, and Bo Wang. Graphmamba: Towards long-range graph sequence modeling with selective state spaces. arXiv preprint arXiv:2402.00789, 2024.3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 430 + ], + "type": "text", + "content": "[45] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15878-15887, 2023. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "text", + "content": "[46] Wenxuan Wang, Chen Chen, Meng Ding, Hong Yu, Sen Zha, and Jiangyun Li. Transbts: Multimodal brain tumor segmentation using transformer. In Medical Image Computing and Computer Assisted Intervention-MICCAI, pages 109-119. Springer, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 489, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 555 + ], + "type": "text", + "content": "[47] Yixin Wang, Yang Zhang, Yang Liu, Zihao Lin, Jiang Tian, Cheng Zhong, Zhongchao Shi, Jianping Fan, and Zhiqiang He. Acn: adversarial co-training network for brain tumor segmentation with missing modalities. In Medical Image Computing and Computer Assisted Intervention-MICCAI, pages 410-420. Springer, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 556, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 609 + ], + "type": "text", + "content": "[48] Shicai Wei, Chunbo Luo, and Yang Luo. Mmanet: Margin-aware distillation and modality-aware regularization for incomplete multimodal learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20039-20049, 2023. 1, 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 612, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 644 + ], + "type": "text", + "content": "[49] Jing Nathan Yan, Jiatao Gu, and Alexander M Rush. Diffusion models without attention. arXiv preprint arXiv:2311.18257, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[50] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. arXiv preprint arXiv:2403.17695, 2024. 1, 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[51] Heran Yang, Jian Sun, and Zongben Xu. Learning unified hyper-network for multi-modal mr image synthesis and tu" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25991" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "mor segmentation with missing modalities. IEEE Transactions on Medical Imaging, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 148 + ], + "type": "text", + "content": "[52] Qiushi Yang, Xiaoqing Guo, Zhen Chen, Peter YM Woo, and Yixuan Yuan. D 2-net: Dual disentanglement network for brain tumor segmentation with missing modalities. IEEE Transactions on Medical Imaging, 41(10):2953-2964, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[53] Chenyu You, Ruihan Zhao, Fenglin Liu, Siyuan Dong, Sandeep Chinchali, Ufuk Topcu, Lawrence Staib, and James Duncan. Class-aware adversarial transformers for medical image segmentation. pages 29582-29596, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 258 + ], + "type": "text", + "content": "[54] Ziqi Yu, Xiaoyang Han, Shengjie Zhang, Jianfeng Feng, Tingying Peng, and Xiao-Yong Zhang. Mousegan++: Unsupervised disentanglement and contrastive representation for multiple mri modalities synthesis and structural segmentation of mouse brain. IEEE Transactions on Medical Imaging, 2022, 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 337 + ], + "type": "text", + "content": "[55] Yao Zhang, Nanjun He, Jiawei Yang, Yuexiang Li, Dong Wei, Yawen Huang, Yang Zhang, Zhiqiang He, and Yefeng Zheng. mmformer: Multimodal medical transformer for incomplete multimodal learning of brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 107-117. Springer, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "text", + "content": "[56] Zheyu Zhang, Xinzhao Liu, Zheng Chen, Yueyi Zhang, Huanjing Yue, Yunwei Ou, and Xiaoyan Sun. Anatomical consistency distillation and inconsistency synthesis for brain tumor segmentation with missing modalities. arXiv preprint arXiv:2408.13733, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 458 + ], + "type": "text", + "content": "[57] Zheyu Zhang, Gang Yang, Yueyi Zhang, Huanjing Yue, Aiping Liu, Yunwei Ou, Jian Gong, and Xiaoyan Sun. Tmformer: Token merging transformer for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7414-7422, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[58] Sijie Zhao, Hao Chen, Xueliang Zhang, Pengfeng Xiao, Lei Bai, and Wanli Ouyang. Rs-mamba for large remote sensing image dense prediction. arXiv preprint arXiv:2404.02668, 2024. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "text", + "content": "[59] Zechen Zhao, Heran Yang, and Jian Sun. Modality-adaptive feature interaction for brain tumor segmentation with missing modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 183-192. Springer, 2022. 1, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 590 + ], + "type": "text", + "content": "[60] Zhuoran Zheng and Jun Zhang. Fd-vision mamba for endoscopic exposure correction. arXiv preprint arXiv:2402.06378, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 593, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 645 + ], + "type": "text", + "content": "[61] Chenhong Zhou, Changxing Ding, Xinchao Wang, Zhentai Lu, and Dacheng Tao. One-pass multi-task networks with cross-task guided attention for brain tumor segmentation. IEEE Transactions on Image Processing, 29:4516-4529, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[62] Tao Zhou, Huazhu Fu, Geng Chen, Jianbing Shen, and Ling Shao. Hi-net: hybrid-fusion network for multi-modal mr image synthesis. IEEE Transactions on Medical Imaging, 39 (9):2772-2781, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[63] Tongxue Zhou, Stéphane Canu, Pierre Vera, and Su Ruan. Latent correlation representation learning for brain tumor" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 195 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "segmentation with missing mri modalities. IEEE Transactions on Image Processing, 30:4263-4274, 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[64] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: Redesigning skip connections to exploit multiscale features in image segmentation. IEEE Transactions on Medical Imaging, 39(6):1856-1867, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[65] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. 2024. 1, 3, 4, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25992" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/f0305c72-f982-4ece-ab22-ea79e8e3955f_content_list.json b/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/f0305c72-f982-4ece-ab22-ea79e8e3955f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d0a14831e6d9e1dbd6faa6808adef4f70dcbb5fc --- /dev/null +++ b/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/f0305c72-f982-4ece-ab22-ea79e8e3955f_content_list.json @@ -0,0 +1,1477 @@ +[ + { + "type": "text", + "text": "Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models", + "text_level": 1, + "bbox": [ + 256, + 130, + 743, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhao Cui $^{1}$ Xinxing Zu $^{1}$ Wenhua Zhang $^{1,2}$ Zhongzhou Zhao $^{1}$ Jinyang Gao $^{3}$ $^{1}$ Alibaba Cloud Computing, China \n $^{2}$ Dalian University of Technology, China \n $^{3}$ Tongyi Lab, China \ncyhbrilliant@gmail.com, patrickfollow@gmail.com, 1062894314zwh@mail.dlut.edu.cn, flyhdq@gmail.com, jinyang.gjy@alibaba-inc.com", + "bbox": [ + 132, + 202, + 864, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leveraging Large Language Models (LLMs) for text representation has achieved significant success, but the exploration of using Multimodal LLMs (MLLMs) for multimodal representation remains limited. Previous MLLM-based representation studies have primarily focused on unifying the embedding space while neglecting the importance of multimodal alignment. As a result, their cross-modal retrieval performance falls markedly behind that of the CLIP series models. To address this, in our work, we 1) construct DeKon5M, a contrastive learning dataset enriched with dense multimodal knowledge, which efficiently enhances multimodal alignment capabilities in representation tasks. 2) design a framework for training unified representation on MLLMs. Building upon this unified representation framework and the dense knowledge dataset DeKon5M, we developed the Dense Knowledge Representation model DeKR on Qwen2VL. Through extensive quantitative and qualitative experiments, our results demonstrate that DeKR not only aligns text, image, video, and text-image combinations within a unified embedding space but also achieves cross-modal retrieval performance comparable to SoTA CLIP series models. This fully validates the effectiveness of our approach and provides new insights for multimodal representation research.", + "bbox": [ + 89, + 342, + 483, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 737, + 220, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal retrieval has consistently been a crucial component in applications such as search systems and retrieval-augmented generation. Due to the outstanding cross-modal retrieval performance of CLIP [45], models from the CLIP family [9, 15, 26, 27, 33, 49, 50, 54, 56, 61] are widely employed in multimodal retrieval tasks. In light of the tremendous success Large Language Models (LLMs) have achieved in text representation tasks [2, 7, 19, 20, 28, 36, 37, 52, 64], recent explorations have begun to investigate", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/581ad1c71b1322191eaebc0f314032668ac3ee9b063c6ef9129978390c692dfe.jpg", + "image_caption": [ + "Figure 1. Comparison of performance with SoTA models in different retrieval tasks. \"ZS\" stands for Zero-shot." + ], + "image_footnote": [], + "bbox": [ + 516, + 306, + 901, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "representation tasks based on Multimodal Large Language Models (MLLMs). To obtain combined representations of image and text modalities, Zhou et al. [67] encoded images into visual features using a Vision Transformer (ViT) [11] and then input them together with text tokens into a text encoder. This process can be approximated as an MLLM. Jiang et al. [17] were the first to utilize an MLLM for multimodal representation learning directly. Building upon LLaVA-NeXT [31], they trained using text-pair data to embed image and text modalities within a unified embedding space. Their retrieval performance demonstrates the feasibility of the MLLM-based representation model. However, previous MLLM-based representations shared a common issue: their cross-modal performance significantly lagged behind the SOTA CLIP series models. We believe this gap arises because their training data didn't activate the multimodal alignment capabilities within MLLMs.", + "bbox": [ + 511, + 609, + 906, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As illustrated in Figure 2(a), we observed that early multimodal pretraining studies, such as Ernie-VIL [60] and", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29733", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ba0ce7b38a17f73c96dd4e8d29ec1b51ee54938dca36721b2d5bd90ac1deeda.jpg", + "image_caption": [ + "Figure 2. Various uses of multimodal knowledge." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 472, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rosita [10], addressed the alignment problem by increasing training on multimodal knowledge, i.e., visual-centric words (objects, colors, actions, etc.). And as shown in Figure 2(b), training MLLMs typically uses dense image descriptions rich in visual-centric words to align ViT outputs with LLM inputs. Thus, we hypothesize that the key to multimodal alignment is training more knowledge that links vision and text. To enhance representation alignment, as shown in Figure 2(c), we introduce a denser multimodal knowledge data into contrastive learning. We were surprised to find that performance after training with dense knowledge data showed a significant improvement compared to training with only sparse knowledge data. To further enhance uniformity, we added video-text retrieval data and composed modality retrieval data, resulting in DeKon5M. Based on this dataset, we constructed the dense knowledge representation model DeKR on Qwen2VL. To demonstrate its superiority, as shown in Figure 1, we tested on four cross-modal retrieval tasks, four composed retrieval tasks, and two ranking tasks, achieving SoTA results on all.", + "bbox": [ + 89, + 369, + 483, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 686, + 233, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. LLM-based Representation Models", + "text_level": 1, + "bbox": [ + 89, + 710, + 398, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Wang et al. [52], building upon the LLM model MISTrial [16], constructed text representation tasks using only a small amount of synthetic data and achieved outstanding rankings on the MTEB [39]. Since then, numerous studies have delved into exploiting the representational potential of LLMs. Muennighoff et al. [40] jointly trained representation and generation tasks, enabling a single model to encompass all stages of retrieval-augmented generation tasks. Recently, Li et al. [28] and Chen et al. [7] have respectively upgraded previous works to versions tailored for large language model representations based on Qwen2 and", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lemma2. In summary, large language models have already been assumed a dominant position in text representation tasks.", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multimodal Representation Models", + "text_level": 1, + "bbox": [ + 511, + 130, + 823, + 145 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the introduction of CLIP [45], multimodal representation tasks have garnered widespread attention. Schuhmann et al. [47, 48] utilized the CLIP model for large-scale data cleaning, presenting the first open-source billion-level image-text pair dataset, LAION-5B. Building on LAION-5B, Gabriel et al. [15] developed OpenCLIP, an open-source multimodal representation model that surpassed CLIP in performance for the first time. Sun et al. [49, 50] explored the scaling laws of representation models, expanding the model sizes to larger scales of 5B, 8B, and 18B parameters. Zhai et al. [61] investigated the advantages of sigmoid loss in contrastive learning, achieving SoTA with only a 400M-parameter model; however, their data details were not publicly disclosed. Despite ongoing research into cross-modal retrieval tasks, the increasing complexity of retrieval applications has rendered existing approaches insufficient for accommodating the diverse range of retrieval scenarios.", + "bbox": [ + 511, + 151, + 906, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To accommodate composed retrieval tasks, Zhou et al. [67], drawing inspiration from the style of MLLMs, encoded visual information using a ViT and then inputted it alongside text tokens into the text encoder BGEm3 [7]. This improved fusion approach led to significant performance enhancements. Jiang et al. [17] were the first to propose using MLLMs for unified modality representation. They fine-tuned text data based on LLaVA-NeXT [31] to enable the model to represent both image and text within a unified embedding space. However, the cross-modal retrieval performance of these MLLM-based representation models still lags considerably behind SoTA CLIP series models. In this work, we introduce DeKR, a powerful unified representation model capable of representing text, image, video, and text-image combination. Using these representations, we outperform previous works on image-text retrieval, videotext retrieval, and composed retrieval tasks.", + "bbox": [ + 511, + 409, + 908, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows: 1) We are the first to demonstrate that MLLM-based models can surpass the CLIP series in cross-modal retrieval tasks. 2) To enhance alignment, we introduced a novel method of using dense knowledge data for contrastive learning and dedicated significant effort to building the DeKon5M dataset. 3) Leveraging DeKon5M, we trained DeKR, the first model capable of representing text, images, videos, and composed modalities, achieving SoTA results on 14 tasks using only one-thousandth of the training data used by the CLIP series.", + "bbox": [ + 511, + 666, + 908, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Learning Unified Multimodal Embeddings", + "text_level": 1, + "bbox": [ + 511, + 830, + 893, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first introduce the framework for training DeKR, then analyze how dense multimodal knowledge activates multimodal alignment capabilities, and finally present", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29734", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c2d98911dcf53b2aea5311c37d15893db265bd4778a1e5869e90c65f5a11216d.jpg", + "image_caption": [ + "Figure 3. Overview of model architecture. To reduce computational overhead, the Vision Encoder is kept frozen during training, while the parameters of the Projector and the LLM are trained." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 480, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the construction details of the DeKon5M dataset.", + "bbox": [ + 89, + 364, + 413, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Modeling", + "text_level": 1, + "bbox": [ + 89, + 393, + 202, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal large language models. Recent MLLMs tend toward a unified structure, comprising three core components: a visual encoder $V$ , which processes input from a single image or multiple video frames, typically utilizing a pretrained Vision Transformer; a projector $P$ that maps visual information into the text token space, such as a Q-former [24] or MLP layers [30]; and a pretrained large language model $L$ , which contains extensive world knowledge and has fine-grained token-level understanding capabilities. An MLLM can be represented as $\\mathrm{MLLM}(i,v,t) = L(P(V(i,v))\\odot t)$ , where $\\odot$ denotes the concatenation operation, $i$ stands for images, $v$ for videos, and $t$ for text tokens. Unified representation using MLLMs. The CLIP model employs separate visual and text encoders to process images and texts into image embedding $e_i \\in \\mathbb{R}^d$ and text embedding $e_t \\in \\mathbb{R}^d$ . In contrast, since an MLLM is capable of handling both visual and textual inputs simultaneously, we utilize a single, parameter-shared MLLM as a unified encoder. For cross-modal and unimodal embedding tasks, each input consists of one element among $i$ , $v$ or $t$ , such as $\\mathrm{MLLM}(i)$ or $\\mathrm{MLLM}(t)$ . In the case of composed modal embedding, combinations such as $\\mathrm{MLLM}(i,t)$ can be inputted simultaneously. As illustrated in Figure 3, although the input forms differ for each task, the output consistently comprises a token sequence, $y = \\mathrm{MLLM}(x)$ , where $y \\in \\mathbb{R}^{n\\times d}$ . We use the output corresponding to the last token as the representation of the entire input, followed by a linear layer that maps it to the specified embedding dimension.", + "bbox": [ + 89, + 417, + 483, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive learning. To accelerate convergence, we made some minor modifications to the contrastive learning loss. Specifically, the standard contrastive learning loss treats", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "other samples within the current batch as negative samples, which may result in the inability to find hard negative samples when the batch size is not sufficiently large. We manually introduce hard negative samples for training and minimize the following objectives:", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n}} (x, y, \\tilde {y}) = \\frac {1}{\\mathcal {B}} \\sum_ {i = 1} ^ {\\mathcal {B}} \\log \\frac {e ^ {\\tau x _ {i} y _ {i}}}{\\sum_ {j , k = 1} ^ {\\mathcal {B}} \\left(e ^ {\\tau x _ {i} y _ {j}} + e ^ {\\tau x _ {i} \\tilde {y} _ {k}}\\right)} \\tag {1} \\\\ \\mathcal {L} (x, y, \\tilde {x}, \\tilde {y}) = \\frac {1}{2} \\left(\\mathcal {L} _ {c o n} (x, y, \\tilde {y}) + \\mathcal {L} _ {c o n} (y, x, \\tilde {x})\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 172, + 903, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is a learnable scaling parameter, $x$ and $y$ are the embedding vectors to be contrasted, $\\tilde{x}$ and $\\tilde{y}$ are the hard negative samples corresponding to $y$ and $x$ , respectively, and $\\mathcal{B}$ is the batch size. In addition, we use caching operations to increase the batch size, similar to previous work [15].", + "bbox": [ + 511, + 250, + 906, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Integrating ranking task. To enhance accuracy, ranking models [8, 22-25, 57, 63] typically compute similarity scores online by simultaneously inputting the content to be matched, whereas representation models computes and stores vectors separately, then calculates scores offline. Leveraging the flexible input capabilities of MLLMs, we can easily extend representation models into ranking models. Structurally, all parameters are shared except for the final linear layer. To prevent knowledge confusion, we introduce specific instructions: the instruction for representation tasks is \"summarize the content.\" and the instruction for ranking tasks is \"Do they match?\".", + "bbox": [ + 511, + 325, + 908, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dense Multimodal Knowledge", + "text_level": 1, + "bbox": [ + 511, + 513, + 782, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous MLLM-based representation models exhibited relatively weak cross-modal retrieval performance, reflecting deficiencies in their multimodal alignment capabilities. Therefore, enhancing the efficiency of multimodal alignment is crucial for MLLM-based representation models.", + "bbox": [ + 511, + 535, + 905, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 4(a) illustrates our analysis of cross-modal retrieval data, revealing that low-confidence negative samples typically exhibit three types of errors: Action Errors, such as riding a motorcycle as shown in the figure; Object Errors, such as the condition of oranges; and Description Errors, such as the color of the bus. These errors are concentrated on visual-related words, underscoring the importance of these words in addressing cross-modal retrieval challenges. Following Ernie-VIL [60] and Rosita [10], we refer to such words that align the image modality with the text modality as multimodal knowledge. Generally, assuming that the descriptive text accurately corresponds to the image content, the knowledge can be mapped to certain parts of speech: action knowledge to verbs, object knowledge to common nouns1, and description knowledge to adjectives.", + "bbox": [ + 511, + 611, + 905, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 4(b), we present a comparison between dense knowledge data and sparse knowledge data.", + "bbox": [ + 511, + 837, + 903, + 868 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Nouns except proper nouns. Proper nouns, such as names and locations, are considered to have weaker visual associations.", + "bbox": [ + 511, + 875, + 903, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29735", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2d32fde97154de93ae05c6ec83b342540a16eca06e4fefbb223ed92badf4a4dc.jpg", + "image_caption": [ + "Action Error" + ], + "image_footnote": [], + "bbox": [ + 99, + 104, + 189, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A man wearing a helmet is .... sitting on his blue motorcycle", + "bbox": [ + 192, + 127, + 377, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A man is riding a blue motorcycle outside", + "bbox": [ + 192, + 156, + 320, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A yellow beverage with a large ice cube in a glass", + "bbox": [ + 112, + 224, + 287, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A glass full of juice with an orange slice sticking out of", + "bbox": [ + 112, + 252, + 271, + 277 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/570269aa70ecf0196b9adc85d05a63340b06386b5c9c9722fe09af18a44dc64b.jpg", + "image_caption": [ + "Object Error" + ], + "image_footnote": [], + "bbox": [ + 290, + 205, + 377, + 287 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5694f179a3df24b0e9c89dd33c1042731a7abd6c05b8903c91f549027dc876ba.jpg", + "image_caption": [ + "Description Error", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 102, + 301, + 215, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A red and black bus being towed in by a flatbed", + "bbox": [ + 218, + 323, + 377, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Blue and white tow truck pulling a blue bus behind", + "bbox": [ + 218, + 353, + 364, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2c0347ba63139ca34dac291981429906c146664ccc2940c8a118721809770030.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 390, + 104, + 727, + 232 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/731cd7b728e028fd3084f7a1bf28f90b0df7918e713b537b5a29c2faff77f9a4.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 390, + 251, + 743, + 393 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e75efa9cec3ec57cf44a43d22c6e0e86f8956841591d5f2ab572d436116aa4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 88, + 895, + 189 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cf53e82068a7ea56f4b32d6130223be5d3251fd7fb947cb9022601dbbc9ce106.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 191, + 897, + 287 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c4e07bc720f012132e657516827c9b6c7e3eeeab79259fbee3a1e9f23bdc6afb.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 743, + 290, + 893, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It can be observed that the dense knowledge data contains significantly more multimodal knowledge than the sparse knowledge data, enabling a more comprehensive and detailed description of almost all the content within an image. Our collected Dense Knowledge (DK) data typically includes more than 50 knowledge points per sample. Such dense descriptive data can significantly enhance the efficiency of alignment between images and text.", + "bbox": [ + 89, + 462, + 483, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further analyze the differences between DK data and other data, as illustrated in Figure 4(c), we quantified the average amount of this knowledge for each sample across various datasets. We found that the alt-text data used to train the CLIP series models contain very sparse knowledge. For instance, the average amount of knowledge in COYO [4] is only one-third that of Flickr30k [59], a downstream cross-modal retrieval dataset. This sparsity may contribute to the large data requirements for training the CLIP series models. Subsequently, we quantified the average amount of knowledge in the DK data. The average amount of knowledge in DK data is four times that of Flickr30k, indicating that DK data possesses higher image-text alignment efficiency. Furthermore, as illustrated in Figure 4(d), from the perspective of knowledge distribution, we found that the proportion of knowledge within sentences exceeds $70\\%$ in both DK data and Flickr30k, whereas it is only approximately $35\\%$ in COYO. This metric demonstrates that DK data maintains dense knowledge while ensuring efficient learning of such knowledge during model training.", + "bbox": [ + 89, + 598, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6c279653cec252d2c8d8598a8de6da6916325a517ce912e202ec4da5c83867ea.jpg", + "image_caption": [ + "Figure 4. Analysis and statistics of dense knowledge in the data. Subfigure (a) illustrates common cross-modal retrieval errors. Subfigure (b) contrasts dense knowledge data with sparse knowledge data. Subfigure (c) presents statistics on the average amount of knowledge per sample in the data. Subfigure (d) shows the distribution of knowledge across three types of data.", + "Figure 5. Detailed data distribution of Dekon5M." + ], + "image_footnote": [], + "bbox": [ + 514, + 460, + 908, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Data Curation", + "text_level": 1, + "bbox": [ + 511, + 633, + 661, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dense knowledge data. The validated high-quality dense knowledge data primarily originates from the first stage of training MLLMs. This data is mostly generated by trained dense descriptors [42, 43] to ensure that the majority of the content in the images is well-described. We collected a total of 2.3 million image-text pairs from four datasets [6, 13, 30, 51]. To further enhance the quality of the data, we mainly followed the Data Juicer processing pipeline to clean the text part. Then we used a CLIP score of 0.2 followed by a BLIP score of 0.45 to ensure alignment. In the end, we refined the dataset to 1.2 million image-text pairs.", + "bbox": [ + 511, + 657, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data statistics. As shown in Figure 5, the multimodal unified representation training dataset, DeKon5M, is divided into four parts: 1.5 million image-text retrieval data, 1.7 million video-text retrieval data, 1.6 million multimodal composed retrieval data, and 0.3 million unimodal retrieval", + "bbox": [ + 511, + 825, + 906, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29736", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9e7586eae04b73efd2415d023ca9e486fa7cd76df1154f5a53aa33a7dd03ebb1.jpg", + "image_caption": [ + "Figure 6. The image retrieval performance of Dense Knowledge at different proportions, given a total data volume of 1 million." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 289, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/298ced7767608a427042373d3ad4acbf8493dfbba05e6a82935a27abe0d93f2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 85, + 485, + 220 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2535dcae271b9e1780d0082d7b88079c89a739adc6058ff5e8a0e6e8caa76c21.jpg", + "table_caption": [ + "Table 1. Taking the model trained on 1 million COYO samples as the baseline, we compared the performance changes after incorporating different types of modality data. Specifically, I. refers to image-text paired data, V. refers to video-text paired data, and C. refers to composed modality retrieval data." + ], + "table_footnote": [], + "table_body": "
Training dataCOCO I2TCOCO T2IFlickr I2TFlickr T2I
COYO48.6833.6274.9061.80
DeKon-I.64.3044.9686.7071.74
DeKon-I.+V.69.4449.8891.3077.60
DeKon-I.+V.+C.70.4852.3891.7079.68
", + "bbox": [ + 91, + 335, + 501, + 420 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3a9e211a0cb0c989da1e06d2ed73b18e31de80e58f3c45fb9a369afe60ded9f2.jpg", + "table_caption": [ + "Table 2. The DeKR were trained separately using Causal Attention and Bidirectional Attention and were evaluated on image retrieval and video retrieval tasks." + ], + "table_footnote": [], + "table_body": "
Attn. typeCOCO I2TCOCO T2IMsrvtt V2TMsrvtt T2V
Causal69.2650.7843.0042.90
Bidirectional70.4852.3842.0042.40
", + "bbox": [ + 91, + 477, + 500, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "data. The image-text retrieval data primarily consists of cleaned dense knowledge data; however, these data tend to have excessively long text. To ensure that the model can generalize across different length distributions, we incorporated the cleaned SBU data [44] as short text data. The specific performance changes brought about by this adjustment will be analyzed in Section 4.1. Moreover, for the multimodal composed retrieval data, in addition to using some filtered VISTA dataset [67], we synthesized more data [3, 13, 14, 21, 30, 51, 62] for each retrieval form to ensure that the dataset size is roughly equivalent to that of the image-text retrieval data. For the video-text retrieval data, we did not perform extensive optimization; instead, we randomly selected a comparable number of samples from ordinary public video retrieval datasets [1, 54] to maintain a balanced ratio.", + "bbox": [ + 88, + 547, + 482, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 811, + 220, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section is divided into four parts. The first two sections present ablation studies on the data and the model. The third section compares DeKR with SoTA retrieval models, and the final section conducts a qualitative analysis.", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f05f2dbc618d7355236129127b969e5a7a8fd6dae14e1d82397f402ca1a29fe3.jpg", + "image_caption": [ + "Figure 7. Exploring training scaling in model size, training data, and input resolution, and comparing training efficiency with CLIP series models." + ], + "image_footnote": [], + "bbox": [ + 514, + 89, + 905, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Ablation Studies of Data", + "text_level": 1, + "bbox": [ + 511, + 375, + 736, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dense knowledge vs Sparse knowledge. As illustrated in Figure 6, we analyzed the proportion of dense knowledge within the training data. A dense knowledge ratio of 0 indicates the exclusive use of sparse knowledge data, while a ratio of 1 signifies the exclusive use of dense knowledge data. The sparse knowledge data primarily comprises COYO [4] and SBU [44], whereas the dense knowledge data mainly consists of SAM [13], LLaVA-Pretrain [30], and ALLaVA [6]. As shown in the figure, retrieval performance continuously improves with an increasing proportion of dense knowledge until it plateaus at an $80\\%$ ratio. We observed a significant performance decline between $80\\%$ and $100\\%$ , which we attribute to issues with data generalization. The dense knowledge data tends to be longer, whereas the downstream task primarily consists of short texts. Therefore, it is necessary to incorporate a small proportion of short text data into the training dataset to ensure that the model maintains generalization capabilities regarding text length.", + "bbox": [ + 511, + 402, + 906, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Retrieval task mixing. As shown in Table 1, we validated the effectiveness of multi-task training. We found that incorporating video retrieval tasks and composed modality retrieval tasks improved image retrieval performance by approximately 5 points and 2 points, respectively. We believe this improvement is due to multimodal alignment being performed from different dimensions. The image retrieval task primarily focuses on object alignment, the video retrieval task focuses on temporal action alignment, and the composed retrieval task focuses on detail variations and the unification of task formats. Additionally, models trained with dense knowledge data achieved an average improvement of 12.1 points in retrieval performance compared to models trained only with sparse knowledge in COYO, demonstrating the effectiveness of our dense knowledge data.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29737", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6464efa73a4b44c6a540a2927674551cdf6e3df24ef186e0e437fe7421b90cbc.jpg", + "table_caption": [ + "Table 3. Comparison on Zero-Shot Image-Text Retrieval. R* denotes Recall@*. The models under comparison are divided into three sections: the top section comprises the CLIP series models, the middle section consists of the previously MLLM-based models, and the bottom section includes our DeKR models of various sizes." + ], + "table_footnote": [], + "table_body": "
COCO-5K I2TCOCO-5K T2IFlickr I2TFlickr T2IAvg.
R1R5R10R1R5R10R1R5R10R1R5R10
CLIP-L58.481.588.137.862.472.288.098.799.468.790.695.278.4
OpenCLIP-L62.183.490.346.170.779.488.798.499.275.092.595.681.8
OpenCLIP-bigG67.386.892.751.474.983.092.999.399.879.595.097.185.0
Eva02-L64.185.290.847.971.780.089.298.999.677.994.296.883.0
Eva02-E+68.787.792.851.275.082.793.999.499.878.894.296.885.1
SigLip-so400M70.2--52.0---------
InternVL-C70.689.093.554.177.384.694.799.699.981.796.098.286.6
VISTA25.549.762.025.749.059.849.879.186.152.078.386.058.6
E5-V(Llava-8B)62.083.689.752.076.584.788.298.799.479.595.097.683.9
DeKR-2B@44870.589.493.952.476.984.891.798.899.479.795.297.985.9
DeKR-7B@44873.391.095.053.878.485.993.199.499.980.395.998.387.1
", + "bbox": [ + 91, + 133, + 906, + 380 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a25cd1af08973ffa033d2e7349cc56b7e03b9ad077528669a2010f4487a6d329.jpg", + "table_caption": [ + "Table 4. Comparison on Zero-Shot Video-Text Retrieval. Models marked with a \"+\" denote those trained exclusively for the image retrieval task, which use multi-frame averaging for inference." + ], + "table_footnote": [], + "table_body": "
Msrvtt V2TMsrvtt T2VDiDemo V2TDiDemo T2VAvg.
CLIP-L*25.829.019.111.521.4
SigLip-so400M*35.836.330.435.834.6
CLIP4Clip27.232.0---
ViCLIP41.342.427.918.432.5
InternVideo-L39.640.733.531.536.3
DeKR-2B@44842.042.437.435.439.3
DeKR-7B@44843.742.939.138.241.0
", + "bbox": [ + 91, + 433, + 478, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Studies of Model", + "text_level": 1, + "bbox": [ + 89, + 602, + 326, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Causal vs Bidirectional. In previous work on LLM-based text representation model, two forms of attention—Causal Attention [38, 41, 65] and Bidirectional Attention [2, 28, 40]—have been employed. As shown in Table 2, we explored both Causal and Bidirectional Attention in our DeKR model. We found that Bidirectional Attention performs better for image retrieval, while Causal Attention is more effective for video retrieval. We believe this is because Causal Attention focuses more on temporal changes, whereas Bidirectional Attention emphasizes detailed information.", + "bbox": [ + 89, + 628, + 482, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training scaling. As illustrated in Figure 7, we conducted scaling experiments across three dimensions: data volume, model size, and input resolution. For data scaling, we observed that increasing the number of training samples from 625K to 5M resulted in a steady improvement in model performance, exhibiting a convergent trend. A further horizontal comparison revealed that, in contrast to the CLIP series models which typically train over 10 billion samples, our", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1f4ceec2e112a16c694b9c4966e1d5638a9824688c03be45c97669fbce2e20c0.jpg", + "table_caption": [ + "Table 5. Comparison on Zero-Shot Composed Image Retrieval. \"val-all\" refers to using queries from the validation set to retrieve from a corpus that includes the training, validation, and test sets." + ], + "table_footnote": [], + "table_body": "
FashionIQCIRR
DressShirtTopTeeAvg.val-allval-all
R10R10R10R10R5R5
CLIP-L----3.613.4
Pic2Word20.026.227.924.78.223.4
LinCIR20.929.128.826.3--
VISTA16.221.821.119.37.522.5
E5-V23.836.435.331.8--
DeKR-2B21.129.829.426.89.527.1
DeKR-7B26.635.433.932.012.533.7
", + "bbox": [ + 514, + 435, + 903, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DeKR model achieved SoTA performance with data that is three orders of magnitude smaller. This demonstrates that dense knowledge can efficiently enhance the multimodal alignment capabilities of MLLMs.", + "bbox": [ + 511, + 627, + 906, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regarding model size scaling, we found that upgrading the model from 2B to 7B parameters led to a performance increase of 2.8 points. This indicates that DeKR effectively inherits knowledge from larger models. Additionally, we explored the impact of input resolution on performance. Due to the limitations of the ViT patch size, the input resolution for CLIP series models is generally restricted to below 336. However, the pooling operation in the Projector component of DeKR overcomes this limitation. We tested input resolutions ranging from 224 to 560 and found that performance growth plateaued beyond a resolution of 448. This is attributed to the fact that most training data have an original resolution of around 500, and higher-resolution images are relatively scarce.", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29738", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/098c02c7d3bd57bc0e0228387a9798172139a725e722122090bbc3a5907e18a9.jpg", + "table_caption": [ + "Table 6. Comparison on Composed Text Retrieval." + ], + "table_footnote": [], + "table_body": "
ReMuQWebQA
R5R10T2ITT2T
Fine-tuned
ReViz-ICT76.283.3--
GeMKR90.392.7--
Zero-shot
CLIP-L0.81.4--
DPR43.448.8--
ReViz-ICT39.446.8--
VISTA84.7-65.169.5
DeKR-2B81.183.659.270.4
DeKR-7B86.087.969.480.6
", + "bbox": [ + 102, + 106, + 467, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison", + "text_level": 1, + "bbox": [ + 89, + 325, + 222, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot image-text retrieval. As shown in Table 3, we compared the CLIP series models [9, 15, 45, 49, 61], MLLM-based retrieval models [17, 67], and our DeKR on the Image-Text Retrieval task. We report their performances on the COCO-5K [29] and Flickr-1K [59] datasets. In terms of average scores, our DeKR-7B model surpasses the SoTA InternVL-C, which has 13B parameters, by 0.5 points and outperforms E5-V by 3.2 points. Notably, our work is the first to demonstrate that MLLM-based retrieval models can exceed the performance of the CLIP series models.", + "bbox": [ + 89, + 347, + 482, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot video-text retrieval. As shown in Table 4, in addition to image retrieval, we also evaluated the multi-frame representation capabilities. We compared embedding models [33, 53, 54] trained for visual and textual alignment on the MSR-VTT [58] and DiDeMo [12] datasets. Besides comparing models trained for video retrieval tasks, we also used a frame averaging method to compare models trained solely for image retrieval tasks as a reference. We found that image retrieval models can generalize well to video retrieval tasks. In terms of average scores, our DeKR achieved SoTA performance at both the 2B and 7B scales. Our model surpasses InternVideo-L by 4.7 points, demonstrating that our unified embeddings can simultaneously represent image details and video temporal information.", + "bbox": [ + 89, + 500, + 482, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot composed modality retrieval. Composed modality retrieval tasks require models to possess the capability of joint representation for both images and text. However, CLIP series models are difficult to support such retrieval formats due to the absence of modality fusion layers. Composed modality retrieval tasks are divided into two categories: composed image retrieval and composed text retrieval. Composed image retrieval refers to adding text-related conditional controls on top of image retrieval, with primary tasks including FashionIQ [55] and CIRR [32] in the $\\mathrm{I + T\\to I}$ format. Composed text retrieval involves com", + "bbox": [ + 89, + 710, + 482, + 876 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a48f2bd064c04c93ccad7dba8bcee3a7e5d423085fd2f83088c015eebc436165.jpg", + "table_caption": [ + "Table 7. Comparison on Zero-Shot Multimodal Ranking. \"-PE\" refers to testing conducted solely through prompt engineering." + ], + "table_footnote": [], + "table_body": "
COCO-5KFlickr
I2TT2II2TT2I
R1R10R1R10R1R10R1R10
AlBEAF----94.199.782.898.1
Qwen2VL-PE56.168.931.158.678.987.351.971.0
Qwen2VL-Rank72.895.139.077.494.299.870.296.2
DeKR-Rank79.396.160.288.196.999.984.898.4
", + "bbox": [ + 514, + 122, + 908, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bining image information to assist text-based retrieval, with main tasks such as WebQA [5] in the $\\mathrm{T}\\rightarrow \\mathrm{I} + \\mathrm{T}$ format and ReMuQ [34] in the $\\mathrm{I} + \\mathrm{T}\\rightarrow \\mathrm{T}$ format. As shown in Table 5, our performance on composed image retrieval surpasses that of Pic2Word [46], a model trained with specifically designed pseudo-word tasks, exceeding it by 7.3 points on the FashionIQ validation set and by 4.3 points on the Val-All task, which utilizes a larger corpus. Additionally, most of our trained composed modality retrieval data were filtered from Vista's training data, yet our composed image retrieval performance is 12.7 points higher than Vista's [67]. This indicates that better multimodal alignment capabilities can also enhance composed retrieval performance. In comparison with E5-V [17], our DeKR outperforms E5-V's 8B-sized model with a smaller 7B model, achieving SoTA performance on composed image retrieval tasks.", + "bbox": [ + 511, + 250, + 906, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 6, our DeKR outperforms Vista and other zero-shot proprietary models [18, 35] on two composed text retrieval tasks, achieving SOTA performance under zero-shot settings. It is noteworthy that since Vista does not finetune the text encoder component, it can be regarded as BGEm3 [7] in text retrieval tasks, serving as a strong baseline for text retrieval. We found that DeKR exceeds BGEm3 by 11.1 points in the WebQA text retrieval task, indicating that DeKR not only possesses unified multimodal representation capabilities but also exhibits excellent text representation abilities.", + "bbox": [ + 511, + 492, + 908, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multimodal ranking. Multimodal ranking tasks typically refer to image-text matching tasks, where both image and text are inputted to produce a matching score based on their correlation. Often, due to differences in input between representation tasks and ranking tasks, separate models are used for representation and ranking. The DeKR model, however, leverages MLLMs to handle combined multimodal inputs, allowing it to perform both representation and ranking tasks within a unified architecture.", + "bbox": [ + 511, + 657, + 908, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify whether DeKR's representation capabilities can positively influence ranking tasks, as shown in Table 7, we compared the ranking performance of three approaches: directly using prompt engineering with Qwen2VL, training ranking data based on Qwen2VL, and training ranking data based on DeKR. The ranking prompt for Qwen2VL was: {\"Image}{Text}Do they match? A.Yes B.No\". The match-", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{2}$ InternVideo2 is not discussed here as it is trained for audio alignment", + "bbox": [ + 107, + 886, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29739", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4d94ff8dd3f5405cd30324a7035f3c98f9e499ef9e0979c9466aed513e48661b.jpg", + "table_caption": [ + "Table 8. Comparison on Fine-tuned Multimodal Ranking." + ], + "table_footnote": [], + "table_body": "
Flickr I2TFlickr T2I
R1R5R10R1R5R10
Uniter87.398.099.275.694.196.8
AlBEAF95.999.8100.085.697.598.9
mPlug297.2100.0100.088.197.699.1
BLIP97.499.899.987.697.799.0
BLIP2-L96.9100.0100.088.697.698.9
BLIP2-g97.6100.0100.089.798.199.0
Qwen2VL95.199.499.870.292.096.5
DeKR98.4100.0100.090.198.299.2
", + "bbox": [ + 91, + 109, + 488, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing score was obtained by normalizing the probabilities of options A and B from the output vocabulary. For the zero-shot setting ranking training data, we randomly selected 500k image-text pairs from DeKon5M and concatenated them into a unified input format as positive samples. Additionally, we selected one negative sample for each modality and followed the multi-range difficulty selection strategy used in mGTE [66].", + "bbox": [ + 89, + 277, + 482, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Surprisingly, we found that Qwen2VL-PE, without any data training, already possessed ranking capabilities, indicating that MLLMs inherently have strong image-text alignment abilities. Furthermore, when comparing the two models trained on ranking data—Qwen2VL+Rank and DeKR+Rank—we observed that the latter outperformed the former by 14.6 points on the Text-to-Image task. This demonstrates that DeKR's representation capabilities have a crucial impact on ranking tasks.", + "bbox": [ + 89, + 398, + 483, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, as shown in Table 8, we report the performance of DeKR after fine-tuning the downstream ranking dataset Flickr. We compared it with previous ranking models and found that DeKR achieved SoTA results on both Image-to-Text and Text-to-Image ranking. This indicates that DeKR successfully unifies the architecture for both representation and ranking tasks. Furthermore, when we finetuned Qwen2VL on the Flickr data, we observed a performance gap of up to 20 points on the T2I task, further emphasizing the necessity of DeKR's representation training.", + "bbox": [ + 89, + 536, + 483, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 696, + 282, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-modal retrieval. In Figure 8(a), we compare DeKR and SigLip on Top-1 recall for short and long text cross-modal retrieval. Both models succeed with short texts, with DeKR preferring detailed descriptions and SigLip favoring broader summaries. For long texts, SigLip retrieves wrong answers when accurate on objects but misses color and position details, suggesting CLIP models focus on object alignment but neglect action and descriptive nuances.", + "bbox": [ + 89, + 719, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Composed modality retrieval. In Figure 8(b), we explore DeKR's performance on more complex multimodal retrieval tasks. The image editing task builds upon the image retrieval task. As shown at the top of the figure, DeKR", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7501d5b1152066c09772a22cf1400413ae31415f2661a6d38db823765e5574ff.jpg", + "image_caption": [ + "Figure 8. Case analysis of cross and composed modality retrieval." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 915, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "first interprets the attributes of the two gorillas in the left image, such as size, posture, and orientation. It then processes the text instruction to only reduce the gorillas' size while maintaining other attributes, successfully retrieving the correct result depicted in the right image. This demonstrates that fine-grained alignment remains crucial in complex retrieval tasks. Additionally, we explored the VQA retrieval task, where DeKR must perform VQA reasoning internally and represent the outcome to retrieve the correct answer, indicating that DeKR retains the original MLLM's reasoning capabilities well. Finally, we assessed DeKR's ability to recall multimodal information and found that it could recognize and retrieve data with the text \"I AM A MAN\" from the image, showcasing its OCR capabilities, which are also inherent in the original MLLM. These cases suggest that if a retrieval model can effectively inherit the original capabilities of MLLM, there is significant potential for retrieval tasks, paving the way for further exploration of DeKR.", + "bbox": [ + 511, + 419, + 906, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 708, + 638, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce DeKR, a unified multimodal retrieval model trained with MLLMs. DeKR integrates text, images, videos, and composed image-text into a unified embedding space and consolidates representation and ranking tasks within one framework. To enhance multimodal alignment, we develop DeKon5M, a contrastive learning dataset rich in dense knowledge. Extensive experiments demonstrate that DeKR achieves SoTA performance in cross-modal retrieval, composed image-text retrieval, and ranking tasks. This shows DeKR's advancement, and DeKon5M contributes to future academic research.", + "bbox": [ + 511, + 732, + 906, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29740", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 5", + "[2] Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. Llm2vec: Large language models are secretly powerful text encoders. CoRR, abs/2404.05961, 2024. 1, 6", + "[3] Rumeysa Bodur, Erhan Gundogdu, Binod Bhattarai, Tae-Kyun Kim, Michael Donoser, and Loris Bazzani. iedit: Localised text-guided image editing with weak supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024 - Workshops, Seattle, WA, USA, June 17-18, 2024, pages 7426-7435. IEEE, 2024. 5", + "[4] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.4.5", + "[5] Yingshan Chang, Guihong Cao, Mridu Narang, Jianfeng Gao, Hisami Suzuki, and Yonatan Bisk. Webqa: Multihop and multimodal QA. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16474-16483. IEEE, 2022. 7", + "[6] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for A lite vision-language model. CoRR, abs/2402.11684, 2024. 4, 5", + "[7] Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. Bge m3-embedding: Multilingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation, 2024. 1, 2, 7", + "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 3", + "[9] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1, 7", + "[10] Yuhao Cui, Zhou Yu, Chunqi Wang, Zhongzhou Zhao, Ji Zhang, Meng Wang, and Jun Yu. Rosita: Enhancing vision-and-language semantic alignments via cross- and intra-modal knowledge integration. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 2, 3", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan C. Russell. Localizing moments in video with natural language. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 5804-5813. IEEE Computer Society, 2017. 7", + "[13] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13405-13417, 2024. 4, 5", + "[14] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. 5", + "[15] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 1, 2, 3, 7", + "[16] Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. Mistral 7b. CoRR, abs/2310.06825, 2023. 2", + "[17] Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-V: universal embeddings with multimodal large language models. CoRR, abs/2407.12580, 2024. 1, 2, 7", + "[18] Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wentau Yih. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781, Online, 2020. Association for Computational Linguistics. 7", + "[19] Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin Kim, Minkyung Cho, Jy yong Sohn, and Chanyeol Choi. Linq-embed-mistral: elevating text retrieval with improved gpt data through task-specific control and quality refinement. Linq AI Research Blog, 2024. 1", + "[20] Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nv-embed: Improved techniques for training llms as generalist embedding models. CoRR, abs/2405.17428, 2024. 1", + "[21] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. CoRR, abs/2408.03326, 2024. 5", + "[22] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 3" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29741", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022.", + "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 19730-19742. PMLR, 2023. 3", + "[25] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. ECCV 2020, 2020. 3", + "[26] Xianhang Li, Zeyu Wang, and Cihang Xie. An inverse scaling law for CLIP training. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 1", + "[27] Xianhang Li, Zeyu Wang, and Cihang Xie. Clipa-v2: Scaling CLIP training with $81.1\\%$ zero-shot imagenet accuracy within a $10,000 budget; an extra $4,000 unlocks 81.8% accuracy. CoRR, abs/2306.15658, 2023. 1", + "[28] Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281, 2023. 1, 2, 6", + "[29] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V, pages 740-755. Springer, 2014. 7", + "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3, 4, 5", + "[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 1, 2", + "[32] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre-trained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2125–2134, 2021. 7", + "[33] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 1, 7", + "[34] Man Luo, Zhiyuan Fang, Tejas Gokhale, Yezhou Yang, and Chitta Baral. End-to-end knowledge retrieval with multimodal queries. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 8573-8589. Association for Computational Linguistics, 2023. 7", + "[35] Man Luo, Zhiyuan Fang, Tejas Gokhale, Yezhou Yang, and Chitta Baral. End-to-end knowledge retrieval with multi-" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "modal queries. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 8573-8589. Association for Computational Linguistics, 2023. 7", + "[36] Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. Sfr-embedding-mistral:enhance text retrieval with transfer learning. Salesforce AI Research Blog, 2024. 1", + "[37] Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. Sfr-embedding-2: Advanced text embedding with multi-stage training, 2024. 1", + "[38] Niklas Muennighoff. Sgpt: Gpt sentence embeddings for semantic search. arXiv preprint arXiv:2202.08904, 2022. 6", + "[39] Niklas Muennighoff, Nouamane Tazi, Loïc Magne, and Nils Reimers. MTEB: massive text embedding benchmark. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2023, Dubrovnik, Croatia, May 2-6, 2023, pages 2006-2029. Association for Computational Linguistics, 2023. 2", + "[40] Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. Generative representational instruction tuning, 2024. 2, 6", + "[41] Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas Tezak, Jong Wook Kim, Chris Hallacy, Johannes Heidecke, Pranav Shyam, Boris Power, Tyna Eloundou Nekoul, Girish Sastry, Gretchen Krueger, David Schnurr, Felipe Petroski Such, Kenny Hsu, Madeleine Thompson, Tabarak Khan, Toki Sherbakov, Joanne Jang, Peter Welinder, and Lilian Weng. Text and code embeddings by contrastive pretraining. CoRR, abs/2201.10005, 2022. 6", + "[42] OpenAI. Gpt-4v(ision) system card. 2023. 4", + "[43] OpenAI. Gpt-4o system card. 2024. 4", + "[44] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2011. 5", + "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 1, 2, 7", + "[46] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 19305-19314. IEEE, 2023. 7", + "[47] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: open dataset of clip-filtered 400 million image-text pairs. CoRR, abs/2111.02114, 2021. 2" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29742", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2", + "[49] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 1, 2, 7", + "[50] Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. EVA-CLIP-18B: scaling CLIP to 18 billion parameters. CoRR, abs/2402.04252, 2024. 1, 2", + "[51] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024. 4, 5", + "[52] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 11897-11916. Association for Computational Linguistics, 2024. 1, 2", + "[53] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 7", + "[54] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 1, 5, 7", + "[55] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The fashion iq dataset: Retrieving images by combining side information and relative natural language feedback. CVPR, 2021. 7", + "[56] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), Online, 2021. Association for Computational Linguistics. 1", + "[57] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang," + ], + "bbox": [ + 91, + 90, + 485, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Guohai Xu, Ji Zhang, Songfang Huang, Fei Huang, and Jingren Zhou. mplug-2: A modularized multi-modal foundation model across text, image and video. ArXiv, abs/2302.00402, 2023. 3", + "[58] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 5288-5296. IEEE Computer Society, 2016. 7", + "[59] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2:67-78, 2014. 4, 7", + "[60] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 3208-3216. AAAI Press, 2021. 1, 3", + "[61] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 11941-11952. IEEE, 2023. 1, 2, 7", + "[62] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 5", + "[63] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021. 3", + "[64] Peitian Zhang, Shitao Xiao, Zheng Liu, Zhicheng Dou, and Jian-Yun Nie. Retrieve anything to augment large language models. CoRR, abs/2310.07554, 2023. 1", + "[65] Xin Zhang, Zehan Li, Yanzhao Zhang, Dingkun Long, Pengjun Xie, Meishan Zhang, and Min Zhang. Language models are universal embeddeders. CoRR, abs/2310.08232, 2023. 6", + "[66] Xin Zhang, Yanzhao Zhang, Dingkun Long, Wen Xie, Ziqi Dai, Jialong Tang, Huan Lin, Baosong Yang, Pengjun Xie, Fei Huang, Meishan Zhang, Wenjie Li, and Min Zhang. mgte: Generalized long-context text representation and reranking models for multilingual text retrieval. CoRR, abs/2407.19669, 2024. 8", + "[67] Junjie Zhou, Zheng Liu, Shitao Xiao, Bo Zhao, and Yongping Xiong. VISTA: visualized text embedding for universal multi-modal retrieval. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 3185-3200. Association for Computational Linguistics, 2024. 1, 2, 5, 7" + ], + "bbox": [ + 516, + 90, + 906, + 880 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29743", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/f0305c72-f982-4ece-ab22-ea79e8e3955f_model.json b/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/f0305c72-f982-4ece-ab22-ea79e8e3955f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4b9c45e5b50c1b226ab9554ddfec91871e57171e --- /dev/null +++ b/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/f0305c72-f982-4ece-ab22-ea79e8e3955f_model.json @@ -0,0 +1,2312 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.257, + 0.131, + 0.744, + 0.177 + ], + "angle": 0, + "content": "Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.203, + 0.866, + 0.276 + ], + "angle": 0, + "content": "Yuhao Cui\\(^{1}\\) Xinxing Zu\\(^{1}\\) Wenhua Zhang\\(^{1,2}\\) Zhongzhou Zhao\\(^{1}\\) Jinyang Gao\\(^{3}\\) \n\\(^{1}\\)Alibaba Cloud Computing, China \n\\(^{2}\\)Dalian University of Technology, China \n\\(^{3}\\)Tongyi Lab, China \ncyhbrilliant@gmail.com, patrickfollow@gmail.com, 1062894314zwh@mail.dlut.edu.cn, flyhdq@gmail.com, jinyang.gjy@alibaba-inc.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.485, + 0.706 + ], + "angle": 0, + "content": "Leveraging Large Language Models (LLMs) for text representation has achieved significant success, but the exploration of using Multimodal LLMs (MLLMs) for multimodal representation remains limited. Previous MLLM-based representation studies have primarily focused on unifying the embedding space while neglecting the importance of multimodal alignment. As a result, their cross-modal retrieval performance falls markedly behind that of the CLIP series models. To address this, in our work, we 1) construct DeKon5M, a contrastive learning dataset enriched with dense multimodal knowledge, which efficiently enhances multimodal alignment capabilities in representation tasks. 2) design a framework for training unified representation on MLLMs. Building upon this unified representation framework and the dense knowledge dataset DeKon5M, we developed the Dense Knowledge Representation model DeKR on Qwen2VL. Through extensive quantitative and qualitative experiments, our results demonstrate that DeKR not only aligns text, image, video, and text-image combinations within a unified embedding space but also achieves cross-modal retrieval performance comparable to SoTA CLIP series models. This fully validates the effectiveness of our approach and provides new insights for multimodal representation research." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.738, + 0.222, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Multimodal retrieval has consistently been a crucial component in applications such as search systems and retrieval-augmented generation. Due to the outstanding cross-modal retrieval performance of CLIP [45], models from the CLIP family [9, 15, 26, 27, 33, 49, 50, 54, 56, 61] are widely employed in multimodal retrieval tasks. In light of the tremendous success Large Language Models (LLMs) have achieved in text representation tasks [2, 7, 19, 20, 28, 36, 37, 52, 64], recent explorations have begun to investigate" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.308, + 0.903, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.574, + 0.907, + 0.601 + ], + "angle": 0, + "content": "Figure 1. Comparison of performance with SoTA models in different retrieval tasks. \"ZS\" stands for Zero-shot." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.867 + ], + "angle": 0, + "content": "representation tasks based on Multimodal Large Language Models (MLLMs). To obtain combined representations of image and text modalities, Zhou et al. [67] encoded images into visual features using a Vision Transformer (ViT) [11] and then input them together with text tokens into a text encoder. This process can be approximated as an MLLM. Jiang et al. [17] were the first to utilize an MLLM for multimodal representation learning directly. Building upon LLaVA-NeXT [31], they trained using text-pair data to embed image and text modalities within a unified embedding space. Their retrieval performance demonstrates the feasibility of the MLLM-based representation model. However, previous MLLM-based representations shared a common issue: their cross-modal performance significantly lagged behind the SOTA CLIP series models. We believe this gap arises because their training data didn't activate the multimodal alignment capabilities within MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "As illustrated in Figure 2(a), we observed that early multimodal pretraining studies, such as Ernie-VIL [60] and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29733" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.086, + 0.473, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.352, + 0.434, + 0.366 + ], + "angle": 0, + "content": "Figure 2. Various uses of multimodal knowledge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.485, + 0.675 + ], + "angle": 0, + "content": "Rosita [10], addressed the alignment problem by increasing training on multimodal knowledge, i.e., visual-centric words (objects, colors, actions, etc.). And as shown in Figure 2(b), training MLLMs typically uses dense image descriptions rich in visual-centric words to align ViT outputs with LLM inputs. Thus, we hypothesize that the key to multimodal alignment is training more knowledge that links vision and text. To enhance representation alignment, as shown in Figure 2(c), we introduce a denser multimodal knowledge data into contrastive learning. We were surprised to find that performance after training with dense knowledge data showed a significant improvement compared to training with only sparse knowledge data. To further enhance uniformity, we added video-text retrieval data and composed modality retrieval data, resulting in DeKon5M. Based on this dataset, we constructed the dense knowledge representation model DeKR on Qwen2VL. To demonstrate its superiority, as shown in Figure 1, we tested on four cross-modal retrieval tasks, four composed retrieval tasks, and two ranking tasks, achieving SoTA results on all." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.687, + 0.234, + 0.703 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.4, + 0.729 + ], + "angle": 0, + "content": "2.1. LLM-based Representation Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Wang et al. [52], building upon the LLM model MISTrial [16], constructed text representation tasks using only a small amount of synthetic data and achieved outstanding rankings on the MTEB [39]. Since then, numerous studies have delved into exploiting the representational potential of LLMs. Muennighoff et al. [40] jointly trained representation and generation tasks, enabling a single model to encompass all stages of retrieval-augmented generation tasks. Recently, Li et al. [28] and Chen et al. [7] have respectively upgraded previous works to versions tailored for large language model representations based on Qwen2 and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "Lemma2. In summary, large language models have already been assumed a dominant position in text representation tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.131, + 0.824, + 0.146 + ], + "angle": 0, + "content": "2.2. Multimodal Representation Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.907, + 0.41 + ], + "angle": 0, + "content": "Since the introduction of CLIP [45], multimodal representation tasks have garnered widespread attention. Schuhmann et al. [47, 48] utilized the CLIP model for large-scale data cleaning, presenting the first open-source billion-level image-text pair dataset, LAION-5B. Building on LAION-5B, Gabriel et al. [15] developed OpenCLIP, an open-source multimodal representation model that surpassed CLIP in performance for the first time. Sun et al. [49, 50] explored the scaling laws of representation models, expanding the model sizes to larger scales of 5B, 8B, and 18B parameters. Zhai et al. [61] investigated the advantages of sigmoid loss in contrastive learning, achieving SoTA with only a 400M-parameter model; however, their data details were not publicly disclosed. Despite ongoing research into cross-modal retrieval tasks, the increasing complexity of retrieval applications has rendered existing approaches insufficient for accommodating the diverse range of retrieval scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.909, + 0.666 + ], + "angle": 0, + "content": "To accommodate composed retrieval tasks, Zhou et al. [67], drawing inspiration from the style of MLLMs, encoded visual information using a ViT and then inputted it alongside text tokens into the text encoder BGEm3 [7]. This improved fusion approach led to significant performance enhancements. Jiang et al. [17] were the first to propose using MLLMs for unified modality representation. They fine-tuned text data based on LLaVA-NeXT [31] to enable the model to represent both image and text within a unified embedding space. However, the cross-modal retrieval performance of these MLLM-based representation models still lags considerably behind SoTA CLIP series models. In this work, we introduce DeKR, a powerful unified representation model capable of representing text, image, video, and text-image combination. Using these representations, we outperform previous works on image-text retrieval, videotext retrieval, and composed retrieval tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.909, + 0.817 + ], + "angle": 0, + "content": "Our contributions are as follows: 1) We are the first to demonstrate that MLLM-based models can surpass the CLIP series in cross-modal retrieval tasks. 2) To enhance alignment, we introduced a novel method of using dense knowledge data for contrastive learning and dedicated significant effort to building the DeKon5M dataset. 3) Leveraging DeKon5M, we trained DeKR, the first model capable of representing text, images, videos, and composed modalities, achieving SoTA results on 14 tasks using only one-thousandth of the training data used by the CLIP series." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.895, + 0.848 + ], + "angle": 0, + "content": "3. Learning Unified Multimodal Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.902 + ], + "angle": 0, + "content": "In this section, we first introduce the framework for training DeKR, then analyze how dense multimodal knowledge activates multimodal alignment capabilities, and finally present" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29734" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.482, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.307, + 0.483, + 0.351 + ], + "angle": 0, + "content": "Figure 3. Overview of model architecture. To reduce computational overhead, the Vision Encoder is kept frozen during training, while the parameters of the Projector and the LLM are trained." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.365, + 0.415, + 0.38 + ], + "angle": 0, + "content": "the construction details of the DeKon5M dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.394, + 0.203, + 0.41 + ], + "angle": 0, + "content": "3.1. Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.853 + ], + "angle": 0, + "content": "Multimodal large language models. Recent MLLMs tend toward a unified structure, comprising three core components: a visual encoder \\( V \\), which processes input from a single image or multiple video frames, typically utilizing a pretrained Vision Transformer; a projector \\( P \\) that maps visual information into the text token space, such as a Q-former [24] or MLP layers [30]; and a pretrained large language model \\( L \\), which contains extensive world knowledge and has fine-grained token-level understanding capabilities. An MLLM can be represented as \\( \\mathrm{MLLM}(i,v,t) = L(P(V(i,v))\\odot t) \\), where \\( \\odot \\) denotes the concatenation operation, \\( i \\) stands for images, \\( v \\) for videos, and \\( t \\) for text tokens. Unified representation using MLLMs. The CLIP model employs separate visual and text encoders to process images and texts into image embedding \\( e_i \\in \\mathbb{R}^d \\) and text embedding \\( e_t \\in \\mathbb{R}^d \\). In contrast, since an MLLM is capable of handling both visual and textual inputs simultaneously, we utilize a single, parameter-shared MLLM as a unified encoder. For cross-modal and unimodal embedding tasks, each input consists of one element among \\( i \\), \\( v \\) or \\( t \\), such as \\( \\mathrm{MLLM}(i) \\) or \\( \\mathrm{MLLM}(t) \\). In the case of composed modal embedding, combinations such as \\( \\mathrm{MLLM}(i,t) \\) can be inputted simultaneously. As illustrated in Figure 3, although the input forms differ for each task, the output consistently comprises a token sequence, \\( y = \\mathrm{MLLM}(x) \\), where \\( y \\in \\mathbb{R}^{n\\times d} \\). We use the output corresponding to the last token as the representation of the entire input, followed by a linear layer that maps it to the specified embedding dimension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Contrastive learning. To accelerate convergence, we made some minor modifications to the contrastive learning loss. Specifically, the standard contrastive learning loss treats" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "other samples within the current batch as negative samples, which may result in the inability to find hard negative samples when the batch size is not sufficiently large. We manually introduce hard negative samples for training and minimize the following objectives:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.173, + 0.905, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n}} (x, y, \\tilde {y}) = \\frac {1}{\\mathcal {B}} \\sum_ {i = 1} ^ {\\mathcal {B}} \\log \\frac {e ^ {\\tau x _ {i} y _ {i}}}{\\sum_ {j , k = 1} ^ {\\mathcal {B}} \\left(e ^ {\\tau x _ {i} y _ {j}} + e ^ {\\tau x _ {i} \\tilde {y} _ {k}}\\right)} \\tag {1} \\\\ \\mathcal {L} (x, y, \\tilde {x}, \\tilde {y}) = \\frac {1}{2} \\left(\\mathcal {L} _ {c o n} (x, y, \\tilde {y}) + \\mathcal {L} _ {c o n} (y, x, \\tilde {x})\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.907, + 0.326 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a learnable scaling parameter, \\(x\\) and \\(y\\) are the embedding vectors to be contrasted, \\(\\tilde{x}\\) and \\(\\tilde{y}\\) are the hard negative samples corresponding to \\(y\\) and \\(x\\), respectively, and \\(\\mathcal{B}\\) is the batch size. In addition, we use caching operations to increase the batch size, similar to previous work [15]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.909, + 0.508 + ], + "angle": 0, + "content": "Integrating ranking task. To enhance accuracy, ranking models [8, 22-25, 57, 63] typically compute similarity scores online by simultaneously inputting the content to be matched, whereas representation models computes and stores vectors separately, then calculates scores offline. Leveraging the flexible input capabilities of MLLMs, we can easily extend representation models into ranking models. Structurally, all parameters are shared except for the final linear layer. To prevent knowledge confusion, we introduce specific instructions: the instruction for representation tasks is \"summarize the content.\" and the instruction for ranking tasks is \"Do they match?\"." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.783, + 0.531 + ], + "angle": 0, + "content": "3.2. Dense Multimodal Knowledge" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.906, + 0.612 + ], + "angle": 0, + "content": "Previous MLLM-based representation models exhibited relatively weak cross-modal retrieval performance, reflecting deficiencies in their multimodal alignment capabilities. Therefore, enhancing the efficiency of multimodal alignment is crucial for MLLM-based representation models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.838 + ], + "angle": 0, + "content": "Figure 4(a) illustrates our analysis of cross-modal retrieval data, revealing that low-confidence negative samples typically exhibit three types of errors: Action Errors, such as riding a motorcycle as shown in the figure; Object Errors, such as the condition of oranges; and Description Errors, such as the color of the bus. These errors are concentrated on visual-related words, underscoring the importance of these words in addressing cross-modal retrieval challenges. Following Ernie-VIL [60] and Rosita [10], we refer to such words that align the image modality with the text modality as multimodal knowledge. Generally, assuming that the descriptive text accurately corresponds to the image content, the knowledge can be mapped to certain parts of speech: action knowledge to verbs, object knowledge to common nouns1, and description knowledge to adjectives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.838, + 0.905, + 0.869 + ], + "angle": 0, + "content": "As shown in Figure 4(b), we present a comparison between dense knowledge data and sparse knowledge data." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Nouns except proper nouns. Proper nouns, such as names and locations, are considered to have weaker visual associations." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29735" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.105, + 0.191, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.111, + 0.277, + 0.125 + ], + "angle": 0, + "content": "Action Error" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.128, + 0.379, + 0.154 + ], + "angle": 0, + "content": "A man wearing a helmet is .... sitting on his blue motorcycle" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.157, + 0.321, + 0.184 + ], + "angle": 0, + "content": "A man is riding a blue motorcycle outside" + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.209, + 0.286, + 0.222 + ], + "angle": 0, + "content": "Object Error" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.225, + 0.288, + 0.25 + ], + "angle": 0, + "content": "A yellow beverage with a large ice cube in a glass" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.253, + 0.272, + 0.279 + ], + "angle": 0, + "content": "A glass full of juice with an orange slice sticking out of" + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.206, + 0.378, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.302, + 0.216, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.308, + 0.332, + 0.321 + ], + "angle": 0, + "content": "Description Error" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.324, + 0.379, + 0.35 + ], + "angle": 0, + "content": "A red and black bus being towed in by a flatbed" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.354, + 0.365, + 0.38 + ], + "angle": 0, + "content": "Blue and white tow truck pulling a blue bus behind" + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.397, + 0.253, + 0.409 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.105, + 0.728, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.236, + 0.592, + 0.25 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.252, + 0.744, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.397, + 0.591, + 0.41 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.089, + 0.896, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.192, + 0.898, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.291, + 0.895, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.808, + 0.395, + 0.83, + 0.409 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.411, + 0.907, + 0.455 + ], + "angle": 0, + "content": "Figure 4. Analysis and statistics of dense knowledge in the data. Subfigure (a) illustrates common cross-modal retrieval errors. Subfigure (b) contrasts dense knowledge data with sparse knowledge data. Subfigure (c) presents statistics on the average amount of knowledge per sample in the data. Subfigure (d) shows the distribution of knowledge across three types of data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.584 + ], + "angle": 0, + "content": "It can be observed that the dense knowledge data contains significantly more multimodal knowledge than the sparse knowledge data, enabling a more comprehensive and detailed description of almost all the content within an image. Our collected Dense Knowledge (DK) data typically includes more than 50 knowledge points per sample. Such dense descriptive data can significantly enhance the efficiency of alignment between images and text." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To further analyze the differences between DK data and other data, as illustrated in Figure 4(c), we quantified the average amount of this knowledge for each sample across various datasets. We found that the alt-text data used to train the CLIP series models contain very sparse knowledge. For instance, the average amount of knowledge in COYO [4] is only one-third that of Flickr30k [59], a downstream cross-modal retrieval dataset. This sparsity may contribute to the large data requirements for training the CLIP series models. Subsequently, we quantified the average amount of knowledge in the DK data. The average amount of knowledge in DK data is four times that of Flickr30k, indicating that DK data possesses higher image-text alignment efficiency. Furthermore, as illustrated in Figure 4(d), from the perspective of knowledge distribution, we found that the proportion of knowledge within sentences exceeds \\(70\\%\\) in both DK data and Flickr30k, whereas it is only approximately \\(35\\%\\) in COYO. This metric demonstrates that DK data maintains dense knowledge while ensuring efficient learning of such knowledge during model training." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.462, + 0.909, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.601, + 0.856, + 0.615 + ], + "angle": 0, + "content": "Figure 5. Detailed data distribution of Dekon5M." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.662, + 0.649 + ], + "angle": 0, + "content": "3.3. Data Curation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Dense knowledge data. The validated high-quality dense knowledge data primarily originates from the first stage of training MLLMs. This data is mostly generated by trained dense descriptors [42, 43] to ensure that the majority of the content in the images is well-described. We collected a total of 2.3 million image-text pairs from four datasets [6, 13, 30, 51]. To further enhance the quality of the data, we mainly followed the Data Juicer processing pipeline to clean the text part. Then we used a CLIP score of 0.2 followed by a BLIP score of 0.45 to ensure alignment. In the end, we refined the dataset to 1.2 million image-text pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Data statistics. As shown in Figure 5, the multimodal unified representation training dataset, DeKon5M, is divided into four parts: 1.5 million image-text retrieval data, 1.7 million video-text retrieval data, 1.6 million multimodal composed retrieval data, and 0.3 million unimodal retrieval" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29736" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.29, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.087, + 0.486, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.223, + 0.483, + 0.251 + ], + "angle": 0, + "content": "Figure 6. The image retrieval performance of Dense Knowledge at different proportions, given a total data volume of 1 million." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.331 + ], + "angle": 0, + "content": "Table 1. Taking the model trained on 1 million COYO samples as the baseline, we compared the performance changes after incorporating different types of modality data. Specifically, I. refers to image-text paired data, V. refers to video-text paired data, and C. refers to composed modality retrieval data." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.336, + 0.502, + 0.421 + ], + "angle": 0, + "content": "
Training dataCOCO I2TCOCO T2IFlickr I2TFlickr T2I
COYO48.6833.6274.9061.80
DeKon-I.64.3044.9686.7071.74
DeKon-I.+V.69.4449.8891.3077.60
DeKon-I.+V.+C.70.4852.3891.7079.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.473 + ], + "angle": 0, + "content": "Table 2. The DeKR were trained separately using Causal Attention and Bidirectional Attention and were evaluated on image retrieval and video retrieval tasks." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.478, + 0.501, + 0.535 + ], + "angle": 0, + "content": "
Attn. typeCOCO I2TCOCO T2IMsrvtt V2TMsrvtt T2V
Causal69.2650.7843.0042.90
Bidirectional70.4852.3842.0042.40
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.549, + 0.483, + 0.79 + ], + "angle": 0, + "content": "data. The image-text retrieval data primarily consists of cleaned dense knowledge data; however, these data tend to have excessively long text. To ensure that the model can generalize across different length distributions, we incorporated the cleaned SBU data [44] as short text data. The specific performance changes brought about by this adjustment will be analyzed in Section 4.1. Moreover, for the multimodal composed retrieval data, in addition to using some filtered VISTA dataset [67], we synthesized more data [3, 13, 14, 21, 30, 51, 62] for each retrieval form to ensure that the dataset size is roughly equivalent to that of the image-text retrieval data. For the video-text retrieval data, we did not perform extensive optimization; instead, we randomly selected a comparable number of samples from ordinary public video retrieval datasets [1, 54] to maintain a balanced ratio." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.813, + 0.222, + 0.83 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "This section is divided into four parts. The first two sections present ablation studies on the data and the model. The third section compares DeKR with SoTA retrieval models, and the final section conducts a qualitative analysis." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.09, + 0.906, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.322, + 0.907, + 0.363 + ], + "angle": 0, + "content": "Figure 7. Exploring training scaling in model size, training data, and input resolution, and comparing training efficiency with CLIP series models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.376, + 0.738, + 0.391 + ], + "angle": 0, + "content": "4.1. Ablation Studies of Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.675 + ], + "angle": 0, + "content": "Dense knowledge vs Sparse knowledge. As illustrated in Figure 6, we analyzed the proportion of dense knowledge within the training data. A dense knowledge ratio of 0 indicates the exclusive use of sparse knowledge data, while a ratio of 1 signifies the exclusive use of dense knowledge data. The sparse knowledge data primarily comprises COYO [4] and SBU [44], whereas the dense knowledge data mainly consists of SAM [13], LLaVA-Pretrain [30], and ALLaVA [6]. As shown in the figure, retrieval performance continuously improves with an increasing proportion of dense knowledge until it plateaus at an \\(80\\%\\) ratio. We observed a significant performance decline between \\(80\\%\\) and \\(100\\%\\), which we attribute to issues with data generalization. The dense knowledge data tends to be longer, whereas the downstream task primarily consists of short texts. Therefore, it is necessary to incorporate a small proportion of short text data into the training dataset to ensure that the model maintains generalization capabilities regarding text length." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Retrieval task mixing. As shown in Table 1, we validated the effectiveness of multi-task training. We found that incorporating video retrieval tasks and composed modality retrieval tasks improved image retrieval performance by approximately 5 points and 2 points, respectively. We believe this improvement is due to multimodal alignment being performed from different dimensions. The image retrieval task primarily focuses on object alignment, the video retrieval task focuses on temporal action alignment, and the composed retrieval task focuses on detail variations and the unification of task formats. Additionally, models trained with dense knowledge data achieved an average improvement of 12.1 points in retrieval performance compared to models trained only with sparse knowledge in COYO, demonstrating the effectiveness of our dense knowledge data." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29737" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 3. Comparison on Zero-Shot Image-Text Retrieval. R* denotes Recall@*. The models under comparison are divided into three sections: the top section comprises the CLIP series models, the middle section consists of the previously MLLM-based models, and the bottom section includes our DeKR models of various sizes." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.135, + 0.907, + 0.381 + ], + "angle": 0, + "content": "
COCO-5K I2TCOCO-5K T2IFlickr I2TFlickr T2IAvg.
R1R5R10R1R5R10R1R5R10R1R5R10
CLIP-L58.481.588.137.862.472.288.098.799.468.790.695.278.4
OpenCLIP-L62.183.490.346.170.779.488.798.499.275.092.595.681.8
OpenCLIP-bigG67.386.892.751.474.983.092.999.399.879.595.097.185.0
Eva02-L64.185.290.847.971.780.089.298.999.677.994.296.883.0
Eva02-E+68.787.792.851.275.082.793.999.499.878.894.296.885.1
SigLip-so400M70.2--52.0---------
InternVL-C70.689.093.554.177.384.694.799.699.981.796.098.286.6
VISTA25.549.762.025.749.059.849.879.186.152.078.386.058.6
E5-V(Llava-8B)62.083.689.752.076.584.788.298.799.479.595.097.683.9
DeKR-2B@44870.589.493.952.476.984.891.798.899.479.795.297.985.9
DeKR-7B@44873.391.095.053.878.485.993.199.499.980.395.998.387.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.389, + 0.482, + 0.432 + ], + "angle": 0, + "content": "Table 4. Comparison on Zero-Shot Video-Text Retrieval. Models marked with a \"+\" denote those trained exclusively for the image retrieval task, which use multi-frame averaging for inference." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.434, + 0.48, + 0.581 + ], + "angle": 0, + "content": "
Msrvtt V2TMsrvtt T2VDiDemo V2TDiDemo T2VAvg.
CLIP-L*25.829.019.111.521.4
SigLip-so400M*35.836.330.435.834.6
CLIP4Clip27.232.0---
ViCLIP41.342.427.918.432.5
InternVideo-L39.640.733.531.536.3
DeKR-2B@44842.042.437.435.439.3
DeKR-7B@44843.742.939.138.241.0
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.603, + 0.328, + 0.618 + ], + "angle": 0, + "content": "4.2. Ablation Studies of Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Causal vs Bidirectional. In previous work on LLM-based text representation model, two forms of attention—Causal Attention [38, 41, 65] and Bidirectional Attention [2, 28, 40]—have been employed. As shown in Table 2, we explored both Causal and Bidirectional Attention in our DeKR model. We found that Bidirectional Attention performs better for image retrieval, while Causal Attention is more effective for video retrieval. We believe this is because Causal Attention focuses more on temporal changes, whereas Bidirectional Attention emphasizes detailed information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Training scaling. As illustrated in Figure 7, we conducted scaling experiments across three dimensions: data volume, model size, and input resolution. For data scaling, we observed that increasing the number of training samples from 625K to 5M resulted in a steady improvement in model performance, exhibiting a convergent trend. A further horizontal comparison revealed that, in contrast to the CLIP series models which typically train over 10 billion samples, our" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.389, + 0.907, + 0.432 + ], + "angle": 0, + "content": "Table 5. Comparison on Zero-Shot Composed Image Retrieval. \"val-all\" refers to using queries from the validation set to retrieve from a corpus that includes the training, validation, and test sets." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.436, + 0.905, + 0.611 + ], + "angle": 0, + "content": "
FashionIQCIRR
DressShirtTopTeeAvg.val-allval-all
R10R10R10R10R5R5
CLIP-L----3.613.4
Pic2Word20.026.227.924.78.223.4
LinCIR20.929.128.826.3--
VISTA16.221.821.119.37.522.5
E5-V23.836.435.331.8--
DeKR-2B21.129.829.426.89.527.1
DeKR-7B26.635.433.932.012.533.7
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.689 + ], + "angle": 0, + "content": "DeKR model achieved SoTA performance with data that is three orders of magnitude smaller. This demonstrates that dense knowledge can efficiently enhance the multimodal alignment capabilities of MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Regarding model size scaling, we found that upgrading the model from 2B to 7B parameters led to a performance increase of 2.8 points. This indicates that DeKR effectively inherits knowledge from larger models. Additionally, we explored the impact of input resolution on performance. Due to the limitations of the ViT patch size, the input resolution for CLIP series models is generally restricted to below 336. However, the pooling operation in the Projector component of DeKR overcomes this limitation. We tested input resolutions ranging from 224 to 560 and found that performance growth plateaued beyond a resolution of 448. This is attributed to the fact that most training data have an original resolution of around 500, and higher-resolution images are relatively scarce." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29738" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.136, + 0.09, + 0.437, + 0.103 + ], + "angle": 0, + "content": "Table 6. Comparison on Composed Text Retrieval." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.107, + 0.468, + 0.314 + ], + "angle": 0, + "content": "
ReMuQWebQA
R5R10T2ITT2T
Fine-tuned
ReViz-ICT76.283.3--
GeMKR90.392.7--
Zero-shot
CLIP-L0.81.4--
DPR43.448.8--
ReViz-ICT39.446.8--
VISTA84.7-65.169.5
DeKR-2B81.183.659.270.4
DeKR-7B86.087.969.480.6
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.326, + 0.223, + 0.342 + ], + "angle": 0, + "content": "4.3. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.483, + 0.498 + ], + "angle": 0, + "content": "Zero-shot image-text retrieval. As shown in Table 3, we compared the CLIP series models [9, 15, 45, 49, 61], MLLM-based retrieval models [17, 67], and our DeKR on the Image-Text Retrieval task. We report their performances on the COCO-5K [29] and Flickr-1K [59] datasets. In terms of average scores, our DeKR-7B model surpasses the SoTA InternVL-C, which has 13B parameters, by 0.5 points and outperforms E5-V by 3.2 points. Notably, our work is the first to demonstrate that MLLM-based retrieval models can exceed the performance of the CLIP series models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.71 + ], + "angle": 0, + "content": "Zero-shot video-text retrieval. As shown in Table 4, in addition to image retrieval, we also evaluated the multi-frame representation capabilities. We compared embedding models [33, 53, 54] trained for visual and textual alignment on the MSR-VTT [58] and DiDeMo [12] datasets. Besides comparing models trained for video retrieval tasks, we also used a frame averaging method to compare models trained solely for image retrieval tasks as a reference. We found that image retrieval models can generalize well to video retrieval tasks. In terms of average scores, our DeKR achieved SoTA performance at both the 2B and 7B scales. Our model surpasses InternVideo-L by 4.7 points, demonstrating that our unified embeddings can simultaneously represent image details and video temporal information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Zero-shot composed modality retrieval. Composed modality retrieval tasks require models to possess the capability of joint representation for both images and text. However, CLIP series models are difficult to support such retrieval formats due to the absence of modality fusion layers. Composed modality retrieval tasks are divided into two categories: composed image retrieval and composed text retrieval. Composed image retrieval refers to adding text-related conditional controls on top of image retrieval, with primary tasks including FashionIQ [55] and CIRR [32] in the \\(\\mathrm{I + T\\to I}\\) format. Composed text retrieval involves com" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 7. Comparison on Zero-Shot Multimodal Ranking. \"-PE\" refers to testing conducted solely through prompt engineering." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.123, + 0.909, + 0.248 + ], + "angle": 0, + "content": "
COCO-5KFlickr
I2TT2II2TT2I
R1R10R1R10R1R10R1R10
AlBEAF----94.199.782.898.1
Qwen2VL-PE56.168.931.158.678.987.351.971.0
Qwen2VL-Rank72.895.139.077.494.299.870.296.2
DeKR-Rank79.396.160.288.196.999.984.898.4
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.251, + 0.907, + 0.492 + ], + "angle": 0, + "content": "bining image information to assist text-based retrieval, with main tasks such as WebQA [5] in the \\(\\mathrm{T}\\rightarrow \\mathrm{I} + \\mathrm{T}\\) format and ReMuQ [34] in the \\(\\mathrm{I} + \\mathrm{T}\\rightarrow \\mathrm{T}\\) format. As shown in Table 5, our performance on composed image retrieval surpasses that of Pic2Word [46], a model trained with specifically designed pseudo-word tasks, exceeding it by 7.3 points on the FashionIQ validation set and by 4.3 points on the Val-All task, which utilizes a larger corpus. Additionally, most of our trained composed modality retrieval data were filtered from Vista's training data, yet our composed image retrieval performance is 12.7 points higher than Vista's [67]. This indicates that better multimodal alignment capabilities can also enhance composed retrieval performance. In comparison with E5-V [17], our DeKR outperforms E5-V's 8B-sized model with a smaller 7B model, achieving SoTA performance on composed image retrieval tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.909, + 0.658 + ], + "angle": 0, + "content": "As shown in Table 6, our DeKR outperforms Vista and other zero-shot proprietary models [18, 35] on two composed text retrieval tasks, achieving SOTA performance under zero-shot settings. It is noteworthy that since Vista does not finetune the text encoder component, it can be regarded as BGEm3 [7] in text retrieval tasks, serving as a strong baseline for text retrieval. We found that DeKR exceeds BGEm3 by 11.1 points in the WebQA text retrieval task, indicating that DeKR not only possesses unified multimodal representation capabilities but also exhibits excellent text representation abilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Multimodal ranking. Multimodal ranking tasks typically refer to image-text matching tasks, where both image and text are inputted to produce a matching score based on their correlation. Often, due to differences in input between representation tasks and ranking tasks, separate models are used for representation and ranking. The DeKR model, however, leverages MLLMs to handle combined multimodal inputs, allowing it to perform both representation and ranking tasks within a unified architecture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To verify whether DeKR's representation capabilities can positively influence ranking tasks, as shown in Table 7, we compared the ranking performance of three approaches: directly using prompt engineering with Qwen2VL, training ranking data based on Qwen2VL, and training ranking data based on DeKR. The ranking prompt for Qwen2VL was: {\"Image}{Text}Do they match? A.Yes B.No\". The match-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)InternVideo2 is not discussed here as it is trained for audio alignment" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29739" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.09, + 0.458, + 0.104 + ], + "angle": 0, + "content": "Table 8. Comparison on Fine-tuned Multimodal Ranking." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.11, + 0.49, + 0.27 + ], + "angle": 0, + "content": "
Flickr I2TFlickr T2I
R1R5R10R1R5R10
Uniter87.398.099.275.694.196.8
AlBEAF95.999.8100.085.697.598.9
mPlug297.2100.0100.088.197.699.1
BLIP97.499.899.987.697.799.0
BLIP2-L96.9100.0100.088.697.698.9
BLIP2-g97.6100.0100.089.798.199.0
Qwen2VL95.199.499.870.292.096.5
DeKR98.4100.0100.090.198.299.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.279, + 0.483, + 0.4 + ], + "angle": 0, + "content": "ing score was obtained by normalizing the probabilities of options A and B from the output vocabulary. For the zero-shot setting ranking training data, we randomly selected 500k image-text pairs from DeKon5M and concatenated them into a unified input format as positive samples. Additionally, we selected one negative sample for each modality and followed the multi-range difficulty selection strategy used in mGTE [66]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.536 + ], + "angle": 0, + "content": "Surprisingly, we found that Qwen2VL-PE, without any data training, already possessed ranking capabilities, indicating that MLLMs inherently have strong image-text alignment abilities. Furthermore, when comparing the two models trained on ranking data—Qwen2VL+Rank and DeKR+Rank—we observed that the latter outperformed the former by 14.6 points on the Text-to-Image task. This demonstrates that DeKR's representation capabilities have a crucial impact on ranking tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.689 + ], + "angle": 0, + "content": "Additionally, as shown in Table 8, we report the performance of DeKR after fine-tuning the downstream ranking dataset Flickr. We compared it with previous ranking models and found that DeKR achieved SoTA results on both Image-to-Text and Text-to-Image ranking. This indicates that DeKR successfully unifies the architecture for both representation and ranking tasks. Furthermore, when we finetuned Qwen2VL on the Flickr data, we observed a performance gap of up to 20 points on the T2I task, further emphasizing the necessity of DeKR's representation training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.283, + 0.714 + ], + "angle": 0, + "content": "4.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Cross-modal retrieval. In Figure 8(a), we compare DeKR and SigLip on Top-1 recall for short and long text cross-modal retrieval. Both models succeed with short texts, with DeKR preferring detailed descriptions and SigLip favoring broader summaries. For long texts, SigLip retrieves wrong answers when accurate on objects but misses color and position details, suggesting CLIP models focus on object alignment but neglect action and descriptive nuances." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Composed modality retrieval. In Figure 8(b), we explore DeKR's performance on more complex multimodal retrieval tasks. The image editing task builds upon the image retrieval task. As shown at the top of the figure, DeKR" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.916, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.417 + ], + "angle": 0, + "content": "Figure 8. Case analysis of cross and composed modality retrieval." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.907, + 0.693 + ], + "angle": 0, + "content": "first interprets the attributes of the two gorillas in the left image, such as size, posture, and orientation. It then processes the text instruction to only reduce the gorillas' size while maintaining other attributes, successfully retrieving the correct result depicted in the right image. This demonstrates that fine-grained alignment remains crucial in complex retrieval tasks. Additionally, we explored the VQA retrieval task, where DeKR must perform VQA reasoning internally and represent the outcome to retrieve the correct answer, indicating that DeKR retains the original MLLM's reasoning capabilities well. Finally, we assessed DeKR's ability to recall multimodal information and found that it could recognize and retrieve data with the text \"I AM A MAN\" from the image, showcasing its OCR capabilities, which are also inherent in the original MLLM. These cases suggest that if a retrieval model can effectively inherit the original capabilities of MLLM, there is significant potential for retrieval tasks, paving the way for further exploration of DeKR." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.64, + 0.724 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.907, + 0.887 + ], + "angle": 0, + "content": "In this paper, we introduce DeKR, a unified multimodal retrieval model trained with MLLMs. DeKR integrates text, images, videos, and composed image-text into a unified embedding space and consolidates representation and ranking tasks within one framework. To enhance multimodal alignment, we develop DeKon5M, a contrastive learning dataset rich in dense knowledge. Extensive experiments demonstrate that DeKR achieves SoTA performance in cross-modal retrieval, composed image-text retrieval, and ranking tasks. This shows DeKR's advancement, and DeKon5M contributes to future academic research." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29740" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. Llm2vec: Large language models are secretly powerful text encoders. CoRR, abs/2404.05961, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[3] Rumeysa Bodur, Erhan Gundogdu, Binod Bhattarai, Tae-Kyun Kim, Michael Donoser, and Loris Bazzani. iedit: Localised text-guided image editing with weak supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024 - Workshops, Seattle, WA, USA, June 17-18, 2024, pages 7426-7435. IEEE, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[4] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.4.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.483, + 0.456 + ], + "angle": 0, + "content": "[5] Yingshan Chang, Guihong Cao, Mridu Narang, Jianfeng Gao, Hisami Suzuki, and Yonatan Bisk. Webqa: Multihop and multimodal QA. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16474-16483. IEEE, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.483, + 0.528 + ], + "angle": 0, + "content": "[6] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for A lite vision-language model. CoRR, abs/2402.11684, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[7] Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. Bge m3-embedding: Multilingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.483, + 0.643 + ], + "angle": 0, + "content": "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[9] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[10] Yuhao Cui, Zhou Yu, Chunqi Wang, Zhongzhou Zhao, Ji Zhang, Meng Wang, and Jun Yu. Rosita: Enhancing vision-and-language semantic alignments via cross- and intra-modal knowledge integration. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[12] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan C. Russell. Localizing moments in video with natural language. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 5804-5813. IEEE Computer Society, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.906, + 0.247 + ], + "angle": 0, + "content": "[13] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13405-13417, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[14] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[15] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[16] Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. Mistral 7b. CoRR, abs/2310.06825, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[17] Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-V: universal embeddings with multimodal large language models. CoRR, abs/2407.12580, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[18] Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wentau Yih. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781, Online, 2020. Association for Computational Linguistics. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.66, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[19] Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin Kim, Minkyung Cho, Jy yong Sohn, and Chanyeol Choi. Linq-embed-mistral: elevating text retrieval with improved gpt data through task-specific control and quality refinement. Linq AI Research Blog, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[20] Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nv-embed: Improved techniques for training llms as generalist embedding models. CoRR, abs/2405.17428, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[21] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. CoRR, abs/2408.03326, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[22] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29741" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.482, + 0.231 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 19730-19742. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[25] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. ECCV 2020, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[26] Xianhang Li, Zeyu Wang, and Cihang Xie. An inverse scaling law for CLIP training. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[27] Xianhang Li, Zeyu Wang, and Cihang Xie. Clipa-v2: Scaling CLIP training with \\(81.1\\%\\) zero-shot imagenet accuracy within a $10,000 budget; an extra $4,000 unlocks 81.8% accuracy. CoRR, abs/2306.15658, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[28] Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V, pages 740-755. Springer, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[32] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre-trained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2125–2134, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[33] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[34] Man Luo, Zhiyuan Fang, Tejas Gokhale, Yezhou Yang, and Chitta Baral. End-to-end knowledge retrieval with multimodal queries. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 8573-8589. Association for Computational Linguistics, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Man Luo, Zhiyuan Fang, Tejas Gokhale, Yezhou Yang, and Chitta Baral. End-to-end knowledge retrieval with multi-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "modal queries. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 8573-8589. Association for Computational Linguistics, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[36] Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. Sfr-embedding-mistral:enhance text retrieval with transfer learning. Salesforce AI Research Blog, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[37] Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. Sfr-embedding-2: Advanced text embedding with multi-stage training, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[38] Niklas Muennighoff. Sgpt: Gpt sentence embeddings for semantic search. arXiv preprint arXiv:2202.08904, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[39] Niklas Muennighoff, Nouamane Tazi, Loïc Magne, and Nils Reimers. MTEB: massive text embedding benchmark. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2023, Dubrovnik, Croatia, May 2-6, 2023, pages 2006-2029. Association for Computational Linguistics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.42 + ], + "angle": 0, + "content": "[40] Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. Generative representational instruction tuning, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[41] Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas Tezak, Jong Wook Kim, Chris Hallacy, Johannes Heidecke, Pranav Shyam, Boris Power, Tyna Eloundou Nekoul, Girish Sastry, Gretchen Krueger, David Schnurr, Felipe Petroski Such, Kenny Hsu, Madeleine Thompson, Tabarak Khan, Toki Sherbakov, Joanne Jang, Peter Welinder, and Lilian Weng. Text and code embeddings by contrastive pretraining. CoRR, abs/2201.10005, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.815, + 0.561 + ], + "angle": 0, + "content": "[42] OpenAI. Gpt-4v(ision) system card. 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.777, + 0.577 + ], + "angle": 0, + "content": "[43] OpenAI. Gpt-4o system card. 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.578, + 0.905, + 0.632 + ], + "angle": 0, + "content": "[44] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.634, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[46] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 19305-19314. IEEE, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: open dataset of clip-filtered 400 million image-text pairs. CoRR, abs/2111.02114, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29742" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.218 + ], + "angle": 0, + "content": "[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.485, + 0.261 + ], + "angle": 0, + "content": "[49] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[50] Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. EVA-CLIP-18B: scaling CLIP to 18 billion parameters. CoRR, abs/2402.04252, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[51] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.485, + 0.503 + ], + "angle": 0, + "content": "[52] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 11897-11916. Association for Computational Linguistics, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.485, + 0.588 + ], + "angle": 0, + "content": "[53] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.485, + 0.7 + ], + "angle": 0, + "content": "[54] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.485, + 0.759 + ], + "angle": 0, + "content": "[55] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The fashion iq dataset: Retrieving images by combining side information and relative natural language feedback. CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.485, + 0.871 + ], + "angle": 0, + "content": "[56] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), Online, 2021. Association for Computational Linguistics. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.485, + 0.903 + ], + "angle": 0, + "content": "[57] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.907, + 0.147 + ], + "angle": 0, + "content": "Guohai Xu, Ji Zhang, Songfang Huang, Fei Huang, and Jingren Zhou. mplug-2: A modularized multi-modal foundation model across text, image and video. ArXiv, abs/2302.00402, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[58] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 5288-5296. IEEE Computer Society, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.906, + 0.289 + ], + "angle": 0, + "content": "[59] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2:67-78, 2014. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.907, + 0.415 + ], + "angle": 0, + "content": "[60] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 3208-3216. AAAI Press, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.416, + 0.906, + 0.485 + ], + "angle": 0, + "content": "[61] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 11941-11952. IEEE, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.486, + 0.906, + 0.542 + ], + "angle": 0, + "content": "[62] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.543, + 0.906, + 0.599 + ], + "angle": 0, + "content": "[63] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.6, + 0.906, + 0.641 + ], + "angle": 0, + "content": "[64] Peitian Zhang, Shitao Xiao, Zheng Liu, Zhicheng Dou, and Jian-Yun Nie. Retrieve anything to augment large language models. CoRR, abs/2310.07554, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.642, + 0.906, + 0.697 + ], + "angle": 0, + "content": "[65] Xin Zhang, Zehan Li, Yanzhao Zhang, Dingkun Long, Pengjun Xie, Meishan Zhang, and Min Zhang. Language models are universal embeddeders. CoRR, abs/2310.08232, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.698, + 0.906, + 0.782 + ], + "angle": 0, + "content": "[66] Xin Zhang, Yanzhao Zhang, Dingkun Long, Wen Xie, Ziqi Dai, Jialong Tang, Huan Lin, Baosong Yang, Pengjun Xie, Fei Huang, Meishan Zhang, Wenjie Li, and Min Zhang. mgte: Generalized long-context text representation and reranking models for multilingual text retrieval. CoRR, abs/2407.19669, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.782, + 0.907, + 0.881 + ], + "angle": 0, + "content": "[67] Junjie Zhou, Zheng Liu, Shitao Xiao, Bo Zhao, and Yongping Xiong. VISTA: visualized text embedding for universal multi-modal retrieval. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 3185-3200. Association for Computational Linguistics, 2024. 1, 2, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.881 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29743" + } + ] +] \ No newline at end of file diff --git a/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/f0305c72-f982-4ece-ab22-ea79e8e3955f_origin.pdf b/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/f0305c72-f982-4ece-ab22-ea79e8e3955f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8a1e61a33f909d327740b875f5dd55217fbb37b7 --- /dev/null +++ b/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/f0305c72-f982-4ece-ab22-ea79e8e3955f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5d30c90a081b2a8791e48bf5929219f47c41a305b05d3cb781329e961980076 +size 3520299 diff --git a/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/full.md b/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..63666691a178698acb0f565be4b1f8b7c82a86f8 --- /dev/null +++ b/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/full.md @@ -0,0 +1,300 @@ +# Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models + +Yuhao Cui $^{1}$ Xinxing Zu $^{1}$ Wenhua Zhang $^{1,2}$ Zhongzhou Zhao $^{1}$ Jinyang Gao $^{3}$ $^{1}$ Alibaba Cloud Computing, China + $^{2}$ Dalian University of Technology, China + $^{3}$ Tongyi Lab, China +cyhbrilliant@gmail.com, patrickfollow@gmail.com, 1062894314zwh@mail.dlut.edu.cn, flyhdq@gmail.com, jinyang.gjy@alibaba-inc.com + +# Abstract + +Leveraging Large Language Models (LLMs) for text representation has achieved significant success, but the exploration of using Multimodal LLMs (MLLMs) for multimodal representation remains limited. Previous MLLM-based representation studies have primarily focused on unifying the embedding space while neglecting the importance of multimodal alignment. As a result, their cross-modal retrieval performance falls markedly behind that of the CLIP series models. To address this, in our work, we 1) construct DeKon5M, a contrastive learning dataset enriched with dense multimodal knowledge, which efficiently enhances multimodal alignment capabilities in representation tasks. 2) design a framework for training unified representation on MLLMs. Building upon this unified representation framework and the dense knowledge dataset DeKon5M, we developed the Dense Knowledge Representation model DeKR on Qwen2VL. Through extensive quantitative and qualitative experiments, our results demonstrate that DeKR not only aligns text, image, video, and text-image combinations within a unified embedding space but also achieves cross-modal retrieval performance comparable to SoTA CLIP series models. This fully validates the effectiveness of our approach and provides new insights for multimodal representation research. + +# 1. Introduction + +Multimodal retrieval has consistently been a crucial component in applications such as search systems and retrieval-augmented generation. Due to the outstanding cross-modal retrieval performance of CLIP [45], models from the CLIP family [9, 15, 26, 27, 33, 49, 50, 54, 56, 61] are widely employed in multimodal retrieval tasks. In light of the tremendous success Large Language Models (LLMs) have achieved in text representation tasks [2, 7, 19, 20, 28, 36, 37, 52, 64], recent explorations have begun to investigate + +![](images/581ad1c71b1322191eaebc0f314032668ac3ee9b063c6ef9129978390c692dfe.jpg) +Figure 1. Comparison of performance with SoTA models in different retrieval tasks. "ZS" stands for Zero-shot. + +representation tasks based on Multimodal Large Language Models (MLLMs). To obtain combined representations of image and text modalities, Zhou et al. [67] encoded images into visual features using a Vision Transformer (ViT) [11] and then input them together with text tokens into a text encoder. This process can be approximated as an MLLM. Jiang et al. [17] were the first to utilize an MLLM for multimodal representation learning directly. Building upon LLaVA-NeXT [31], they trained using text-pair data to embed image and text modalities within a unified embedding space. Their retrieval performance demonstrates the feasibility of the MLLM-based representation model. However, previous MLLM-based representations shared a common issue: their cross-modal performance significantly lagged behind the SOTA CLIP series models. We believe this gap arises because their training data didn't activate the multimodal alignment capabilities within MLLMs. + +As illustrated in Figure 2(a), we observed that early multimodal pretraining studies, such as Ernie-VIL [60] and + +![](images/6ba0ce7b38a17f73c96dd4e8d29ec1b51ee54938dca36721b2d5bd90ac1deeda.jpg) +Figure 2. Various uses of multimodal knowledge. + +Rosita [10], addressed the alignment problem by increasing training on multimodal knowledge, i.e., visual-centric words (objects, colors, actions, etc.). And as shown in Figure 2(b), training MLLMs typically uses dense image descriptions rich in visual-centric words to align ViT outputs with LLM inputs. Thus, we hypothesize that the key to multimodal alignment is training more knowledge that links vision and text. To enhance representation alignment, as shown in Figure 2(c), we introduce a denser multimodal knowledge data into contrastive learning. We were surprised to find that performance after training with dense knowledge data showed a significant improvement compared to training with only sparse knowledge data. To further enhance uniformity, we added video-text retrieval data and composed modality retrieval data, resulting in DeKon5M. Based on this dataset, we constructed the dense knowledge representation model DeKR on Qwen2VL. To demonstrate its superiority, as shown in Figure 1, we tested on four cross-modal retrieval tasks, four composed retrieval tasks, and two ranking tasks, achieving SoTA results on all. + +# 2. Related Work + +# 2.1. LLM-based Representation Models + +Wang et al. [52], building upon the LLM model MISTrial [16], constructed text representation tasks using only a small amount of synthetic data and achieved outstanding rankings on the MTEB [39]. Since then, numerous studies have delved into exploiting the representational potential of LLMs. Muennighoff et al. [40] jointly trained representation and generation tasks, enabling a single model to encompass all stages of retrieval-augmented generation tasks. Recently, Li et al. [28] and Chen et al. [7] have respectively upgraded previous works to versions tailored for large language model representations based on Qwen2 and + +Lemma2. In summary, large language models have already been assumed a dominant position in text representation tasks. + +# 2.2. Multimodal Representation Models + +Since the introduction of CLIP [45], multimodal representation tasks have garnered widespread attention. Schuhmann et al. [47, 48] utilized the CLIP model for large-scale data cleaning, presenting the first open-source billion-level image-text pair dataset, LAION-5B. Building on LAION-5B, Gabriel et al. [15] developed OpenCLIP, an open-source multimodal representation model that surpassed CLIP in performance for the first time. Sun et al. [49, 50] explored the scaling laws of representation models, expanding the model sizes to larger scales of 5B, 8B, and 18B parameters. Zhai et al. [61] investigated the advantages of sigmoid loss in contrastive learning, achieving SoTA with only a 400M-parameter model; however, their data details were not publicly disclosed. Despite ongoing research into cross-modal retrieval tasks, the increasing complexity of retrieval applications has rendered existing approaches insufficient for accommodating the diverse range of retrieval scenarios. + +To accommodate composed retrieval tasks, Zhou et al. [67], drawing inspiration from the style of MLLMs, encoded visual information using a ViT and then inputted it alongside text tokens into the text encoder BGEm3 [7]. This improved fusion approach led to significant performance enhancements. Jiang et al. [17] were the first to propose using MLLMs for unified modality representation. They fine-tuned text data based on LLaVA-NeXT [31] to enable the model to represent both image and text within a unified embedding space. However, the cross-modal retrieval performance of these MLLM-based representation models still lags considerably behind SoTA CLIP series models. In this work, we introduce DeKR, a powerful unified representation model capable of representing text, image, video, and text-image combination. Using these representations, we outperform previous works on image-text retrieval, videotext retrieval, and composed retrieval tasks. + +Our contributions are as follows: 1) We are the first to demonstrate that MLLM-based models can surpass the CLIP series in cross-modal retrieval tasks. 2) To enhance alignment, we introduced a novel method of using dense knowledge data for contrastive learning and dedicated significant effort to building the DeKon5M dataset. 3) Leveraging DeKon5M, we trained DeKR, the first model capable of representing text, images, videos, and composed modalities, achieving SoTA results on 14 tasks using only one-thousandth of the training data used by the CLIP series. + +# 3. Learning Unified Multimodal Embeddings + +In this section, we first introduce the framework for training DeKR, then analyze how dense multimodal knowledge activates multimodal alignment capabilities, and finally present + +![](images/c2d98911dcf53b2aea5311c37d15893db265bd4778a1e5869e90c65f5a11216d.jpg) +Figure 3. Overview of model architecture. To reduce computational overhead, the Vision Encoder is kept frozen during training, while the parameters of the Projector and the LLM are trained. + +the construction details of the DeKon5M dataset. + +# 3.1. Modeling + +Multimodal large language models. Recent MLLMs tend toward a unified structure, comprising three core components: a visual encoder $V$ , which processes input from a single image or multiple video frames, typically utilizing a pretrained Vision Transformer; a projector $P$ that maps visual information into the text token space, such as a Q-former [24] or MLP layers [30]; and a pretrained large language model $L$ , which contains extensive world knowledge and has fine-grained token-level understanding capabilities. An MLLM can be represented as $\mathrm{MLLM}(i,v,t) = L(P(V(i,v))\odot t)$ , where $\odot$ denotes the concatenation operation, $i$ stands for images, $v$ for videos, and $t$ for text tokens. Unified representation using MLLMs. The CLIP model employs separate visual and text encoders to process images and texts into image embedding $e_i \in \mathbb{R}^d$ and text embedding $e_t \in \mathbb{R}^d$ . In contrast, since an MLLM is capable of handling both visual and textual inputs simultaneously, we utilize a single, parameter-shared MLLM as a unified encoder. For cross-modal and unimodal embedding tasks, each input consists of one element among $i$ , $v$ or $t$ , such as $\mathrm{MLLM}(i)$ or $\mathrm{MLLM}(t)$ . In the case of composed modal embedding, combinations such as $\mathrm{MLLM}(i,t)$ can be inputted simultaneously. As illustrated in Figure 3, although the input forms differ for each task, the output consistently comprises a token sequence, $y = \mathrm{MLLM}(x)$ , where $y \in \mathbb{R}^{n\times d}$ . We use the output corresponding to the last token as the representation of the entire input, followed by a linear layer that maps it to the specified embedding dimension. + +Contrastive learning. To accelerate convergence, we made some minor modifications to the contrastive learning loss. Specifically, the standard contrastive learning loss treats + +other samples within the current batch as negative samples, which may result in the inability to find hard negative samples when the batch size is not sufficiently large. We manually introduce hard negative samples for training and minimize the following objectives: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {c o n}} (x, y, \tilde {y}) = \frac {1}{\mathcal {B}} \sum_ {i = 1} ^ {\mathcal {B}} \log \frac {e ^ {\tau x _ {i} y _ {i}}}{\sum_ {j , k = 1} ^ {\mathcal {B}} \left(e ^ {\tau x _ {i} y _ {j}} + e ^ {\tau x _ {i} \tilde {y} _ {k}}\right)} \tag {1} \\ \mathcal {L} (x, y, \tilde {x}, \tilde {y}) = \frac {1}{2} \left(\mathcal {L} _ {c o n} (x, y, \tilde {y}) + \mathcal {L} _ {c o n} (y, x, \tilde {x})\right) \\ \end{array} +$$ + +where $\tau$ is a learnable scaling parameter, $x$ and $y$ are the embedding vectors to be contrasted, $\tilde{x}$ and $\tilde{y}$ are the hard negative samples corresponding to $y$ and $x$ , respectively, and $\mathcal{B}$ is the batch size. In addition, we use caching operations to increase the batch size, similar to previous work [15]. + +Integrating ranking task. To enhance accuracy, ranking models [8, 22-25, 57, 63] typically compute similarity scores online by simultaneously inputting the content to be matched, whereas representation models computes and stores vectors separately, then calculates scores offline. Leveraging the flexible input capabilities of MLLMs, we can easily extend representation models into ranking models. Structurally, all parameters are shared except for the final linear layer. To prevent knowledge confusion, we introduce specific instructions: the instruction for representation tasks is "summarize the content." and the instruction for ranking tasks is "Do they match?". + +# 3.2. Dense Multimodal Knowledge + +Previous MLLM-based representation models exhibited relatively weak cross-modal retrieval performance, reflecting deficiencies in their multimodal alignment capabilities. Therefore, enhancing the efficiency of multimodal alignment is crucial for MLLM-based representation models. + +Figure 4(a) illustrates our analysis of cross-modal retrieval data, revealing that low-confidence negative samples typically exhibit three types of errors: Action Errors, such as riding a motorcycle as shown in the figure; Object Errors, such as the condition of oranges; and Description Errors, such as the color of the bus. These errors are concentrated on visual-related words, underscoring the importance of these words in addressing cross-modal retrieval challenges. Following Ernie-VIL [60] and Rosita [10], we refer to such words that align the image modality with the text modality as multimodal knowledge. Generally, assuming that the descriptive text accurately corresponds to the image content, the knowledge can be mapped to certain parts of speech: action knowledge to verbs, object knowledge to common nouns1, and description knowledge to adjectives. + +As shown in Figure 4(b), we present a comparison between dense knowledge data and sparse knowledge data. + +![](images/2d32fde97154de93ae05c6ec83b342540a16eca06e4fefbb223ed92badf4a4dc.jpg) +Action Error + +A man wearing a helmet is .... sitting on his blue motorcycle + +A man is riding a blue motorcycle outside + +A yellow beverage with a large ice cube in a glass + +A glass full of juice with an orange slice sticking out of + +![](images/570269aa70ecf0196b9adc85d05a63340b06386b5c9c9722fe09af18a44dc64b.jpg) +Object Error + +![](images/5694f179a3df24b0e9c89dd33c1042731a7abd6c05b8903c91f549027dc876ba.jpg) +Description Error +(a) + +A red and black bus being towed in by a flatbed + +Blue and white tow truck pulling a blue bus behind + +![](images/2c0347ba63139ca34dac291981429906c146664ccc2940c8a118721809770030.jpg) +(b) + +![](images/731cd7b728e028fd3084f7a1bf28f90b0df7918e713b537b5a29c2faff77f9a4.jpg) +(c) + +![](images/1e75efa9cec3ec57cf44a43d22c6e0e86f8956841591d5f2ab572d436116aa4b.jpg) + +![](images/cf53e82068a7ea56f4b32d6130223be5d3251fd7fb947cb9022601dbbc9ce106.jpg) + +![](images/c4e07bc720f012132e657516827c9b6c7e3eeeab79259fbee3a1e9f23bdc6afb.jpg) +(d) + +It can be observed that the dense knowledge data contains significantly more multimodal knowledge than the sparse knowledge data, enabling a more comprehensive and detailed description of almost all the content within an image. Our collected Dense Knowledge (DK) data typically includes more than 50 knowledge points per sample. Such dense descriptive data can significantly enhance the efficiency of alignment between images and text. + +To further analyze the differences between DK data and other data, as illustrated in Figure 4(c), we quantified the average amount of this knowledge for each sample across various datasets. We found that the alt-text data used to train the CLIP series models contain very sparse knowledge. For instance, the average amount of knowledge in COYO [4] is only one-third that of Flickr30k [59], a downstream cross-modal retrieval dataset. This sparsity may contribute to the large data requirements for training the CLIP series models. Subsequently, we quantified the average amount of knowledge in the DK data. The average amount of knowledge in DK data is four times that of Flickr30k, indicating that DK data possesses higher image-text alignment efficiency. Furthermore, as illustrated in Figure 4(d), from the perspective of knowledge distribution, we found that the proportion of knowledge within sentences exceeds $70\%$ in both DK data and Flickr30k, whereas it is only approximately $35\%$ in COYO. This metric demonstrates that DK data maintains dense knowledge while ensuring efficient learning of such knowledge during model training. + +![](images/6c279653cec252d2c8d8598a8de6da6916325a517ce912e202ec4da5c83867ea.jpg) +Figure 4. Analysis and statistics of dense knowledge in the data. Subfigure (a) illustrates common cross-modal retrieval errors. Subfigure (b) contrasts dense knowledge data with sparse knowledge data. Subfigure (c) presents statistics on the average amount of knowledge per sample in the data. Subfigure (d) shows the distribution of knowledge across three types of data. +Figure 5. Detailed data distribution of Dekon5M. + +# 3.3. Data Curation + +Dense knowledge data. The validated high-quality dense knowledge data primarily originates from the first stage of training MLLMs. This data is mostly generated by trained dense descriptors [42, 43] to ensure that the majority of the content in the images is well-described. We collected a total of 2.3 million image-text pairs from four datasets [6, 13, 30, 51]. To further enhance the quality of the data, we mainly followed the Data Juicer processing pipeline to clean the text part. Then we used a CLIP score of 0.2 followed by a BLIP score of 0.45 to ensure alignment. In the end, we refined the dataset to 1.2 million image-text pairs. + +Data statistics. As shown in Figure 5, the multimodal unified representation training dataset, DeKon5M, is divided into four parts: 1.5 million image-text retrieval data, 1.7 million video-text retrieval data, 1.6 million multimodal composed retrieval data, and 0.3 million unimodal retrieval + +![](images/9e7586eae04b73efd2415d023ca9e486fa7cd76df1154f5a53aa33a7dd03ebb1.jpg) +Figure 6. The image retrieval performance of Dense Knowledge at different proportions, given a total data volume of 1 million. + +![](images/298ced7767608a427042373d3ad4acbf8493dfbba05e6a82935a27abe0d93f2b.jpg) + +Table 1. Taking the model trained on 1 million COYO samples as the baseline, we compared the performance changes after incorporating different types of modality data. Specifically, I. refers to image-text paired data, V. refers to video-text paired data, and C. refers to composed modality retrieval data. + +
Training dataCOCO I2TCOCO T2IFlickr I2TFlickr T2I
COYO48.6833.6274.9061.80
DeKon-I.64.3044.9686.7071.74
DeKon-I.+V.69.4449.8891.3077.60
DeKon-I.+V.+C.70.4852.3891.7079.68
+ +Table 2. The DeKR were trained separately using Causal Attention and Bidirectional Attention and were evaluated on image retrieval and video retrieval tasks. + +
Attn. typeCOCO I2TCOCO T2IMsrvtt V2TMsrvtt T2V
Causal69.2650.7843.0042.90
Bidirectional70.4852.3842.0042.40
+ +data. The image-text retrieval data primarily consists of cleaned dense knowledge data; however, these data tend to have excessively long text. To ensure that the model can generalize across different length distributions, we incorporated the cleaned SBU data [44] as short text data. The specific performance changes brought about by this adjustment will be analyzed in Section 4.1. Moreover, for the multimodal composed retrieval data, in addition to using some filtered VISTA dataset [67], we synthesized more data [3, 13, 14, 21, 30, 51, 62] for each retrieval form to ensure that the dataset size is roughly equivalent to that of the image-text retrieval data. For the video-text retrieval data, we did not perform extensive optimization; instead, we randomly selected a comparable number of samples from ordinary public video retrieval datasets [1, 54] to maintain a balanced ratio. + +# 4. Experiments + +This section is divided into four parts. The first two sections present ablation studies on the data and the model. The third section compares DeKR with SoTA retrieval models, and the final section conducts a qualitative analysis. + +![](images/f05f2dbc618d7355236129127b969e5a7a8fd6dae14e1d82397f402ca1a29fe3.jpg) +Figure 7. Exploring training scaling in model size, training data, and input resolution, and comparing training efficiency with CLIP series models. + +# 4.1. Ablation Studies of Data + +Dense knowledge vs Sparse knowledge. As illustrated in Figure 6, we analyzed the proportion of dense knowledge within the training data. A dense knowledge ratio of 0 indicates the exclusive use of sparse knowledge data, while a ratio of 1 signifies the exclusive use of dense knowledge data. The sparse knowledge data primarily comprises COYO [4] and SBU [44], whereas the dense knowledge data mainly consists of SAM [13], LLaVA-Pretrain [30], and ALLaVA [6]. As shown in the figure, retrieval performance continuously improves with an increasing proportion of dense knowledge until it plateaus at an $80\%$ ratio. We observed a significant performance decline between $80\%$ and $100\%$ , which we attribute to issues with data generalization. The dense knowledge data tends to be longer, whereas the downstream task primarily consists of short texts. Therefore, it is necessary to incorporate a small proportion of short text data into the training dataset to ensure that the model maintains generalization capabilities regarding text length. + +Retrieval task mixing. As shown in Table 1, we validated the effectiveness of multi-task training. We found that incorporating video retrieval tasks and composed modality retrieval tasks improved image retrieval performance by approximately 5 points and 2 points, respectively. We believe this improvement is due to multimodal alignment being performed from different dimensions. The image retrieval task primarily focuses on object alignment, the video retrieval task focuses on temporal action alignment, and the composed retrieval task focuses on detail variations and the unification of task formats. Additionally, models trained with dense knowledge data achieved an average improvement of 12.1 points in retrieval performance compared to models trained only with sparse knowledge in COYO, demonstrating the effectiveness of our dense knowledge data. + +Table 3. Comparison on Zero-Shot Image-Text Retrieval. R* denotes Recall@*. The models under comparison are divided into three sections: the top section comprises the CLIP series models, the middle section consists of the previously MLLM-based models, and the bottom section includes our DeKR models of various sizes. + +
COCO-5K I2TCOCO-5K T2IFlickr I2TFlickr T2IAvg.
R1R5R10R1R5R10R1R5R10R1R5R10
CLIP-L58.481.588.137.862.472.288.098.799.468.790.695.278.4
OpenCLIP-L62.183.490.346.170.779.488.798.499.275.092.595.681.8
OpenCLIP-bigG67.386.892.751.474.983.092.999.399.879.595.097.185.0
Eva02-L64.185.290.847.971.780.089.298.999.677.994.296.883.0
Eva02-E+68.787.792.851.275.082.793.999.499.878.894.296.885.1
SigLip-so400M70.2--52.0---------
InternVL-C70.689.093.554.177.384.694.799.699.981.796.098.286.6
VISTA25.549.762.025.749.059.849.879.186.152.078.386.058.6
E5-V(Llava-8B)62.083.689.752.076.584.788.298.799.479.595.097.683.9
DeKR-2B@44870.589.493.952.476.984.891.798.899.479.795.297.985.9
DeKR-7B@44873.391.095.053.878.485.993.199.499.980.395.998.387.1
+ +Table 4. Comparison on Zero-Shot Video-Text Retrieval. Models marked with a "+" denote those trained exclusively for the image retrieval task, which use multi-frame averaging for inference. + +
Msrvtt V2TMsrvtt T2VDiDemo V2TDiDemo T2VAvg.
CLIP-L*25.829.019.111.521.4
SigLip-so400M*35.836.330.435.834.6
CLIP4Clip27.232.0---
ViCLIP41.342.427.918.432.5
InternVideo-L39.640.733.531.536.3
DeKR-2B@44842.042.437.435.439.3
DeKR-7B@44843.742.939.138.241.0
+ +# 4.2. Ablation Studies of Model + +Causal vs Bidirectional. In previous work on LLM-based text representation model, two forms of attention—Causal Attention [38, 41, 65] and Bidirectional Attention [2, 28, 40]—have been employed. As shown in Table 2, we explored both Causal and Bidirectional Attention in our DeKR model. We found that Bidirectional Attention performs better for image retrieval, while Causal Attention is more effective for video retrieval. We believe this is because Causal Attention focuses more on temporal changes, whereas Bidirectional Attention emphasizes detailed information. + +Training scaling. As illustrated in Figure 7, we conducted scaling experiments across three dimensions: data volume, model size, and input resolution. For data scaling, we observed that increasing the number of training samples from 625K to 5M resulted in a steady improvement in model performance, exhibiting a convergent trend. A further horizontal comparison revealed that, in contrast to the CLIP series models which typically train over 10 billion samples, our + +Table 5. Comparison on Zero-Shot Composed Image Retrieval. "val-all" refers to using queries from the validation set to retrieve from a corpus that includes the training, validation, and test sets. + +
FashionIQCIRR
DressShirtTopTeeAvg.val-allval-all
R10R10R10R10R5R5
CLIP-L----3.613.4
Pic2Word20.026.227.924.78.223.4
LinCIR20.929.128.826.3--
VISTA16.221.821.119.37.522.5
E5-V23.836.435.331.8--
DeKR-2B21.129.829.426.89.527.1
DeKR-7B26.635.433.932.012.533.7
+ +DeKR model achieved SoTA performance with data that is three orders of magnitude smaller. This demonstrates that dense knowledge can efficiently enhance the multimodal alignment capabilities of MLLMs. + +Regarding model size scaling, we found that upgrading the model from 2B to 7B parameters led to a performance increase of 2.8 points. This indicates that DeKR effectively inherits knowledge from larger models. Additionally, we explored the impact of input resolution on performance. Due to the limitations of the ViT patch size, the input resolution for CLIP series models is generally restricted to below 336. However, the pooling operation in the Projector component of DeKR overcomes this limitation. We tested input resolutions ranging from 224 to 560 and found that performance growth plateaued beyond a resolution of 448. This is attributed to the fact that most training data have an original resolution of around 500, and higher-resolution images are relatively scarce. + +Table 6. Comparison on Composed Text Retrieval. + +
ReMuQWebQA
R5R10T2ITT2T
Fine-tuned
ReViz-ICT76.283.3--
GeMKR90.392.7--
Zero-shot
CLIP-L0.81.4--
DPR43.448.8--
ReViz-ICT39.446.8--
VISTA84.7-65.169.5
DeKR-2B81.183.659.270.4
DeKR-7B86.087.969.480.6
+ +# 4.3. Comparison + +Zero-shot image-text retrieval. As shown in Table 3, we compared the CLIP series models [9, 15, 45, 49, 61], MLLM-based retrieval models [17, 67], and our DeKR on the Image-Text Retrieval task. We report their performances on the COCO-5K [29] and Flickr-1K [59] datasets. In terms of average scores, our DeKR-7B model surpasses the SoTA InternVL-C, which has 13B parameters, by 0.5 points and outperforms E5-V by 3.2 points. Notably, our work is the first to demonstrate that MLLM-based retrieval models can exceed the performance of the CLIP series models. + +Zero-shot video-text retrieval. As shown in Table 4, in addition to image retrieval, we also evaluated the multi-frame representation capabilities. We compared embedding models [33, 53, 54] trained for visual and textual alignment on the MSR-VTT [58] and DiDeMo [12] datasets. Besides comparing models trained for video retrieval tasks, we also used a frame averaging method to compare models trained solely for image retrieval tasks as a reference. We found that image retrieval models can generalize well to video retrieval tasks. In terms of average scores, our DeKR achieved SoTA performance at both the 2B and 7B scales. Our model surpasses InternVideo-L by 4.7 points, demonstrating that our unified embeddings can simultaneously represent image details and video temporal information. + +Zero-shot composed modality retrieval. Composed modality retrieval tasks require models to possess the capability of joint representation for both images and text. However, CLIP series models are difficult to support such retrieval formats due to the absence of modality fusion layers. Composed modality retrieval tasks are divided into two categories: composed image retrieval and composed text retrieval. Composed image retrieval refers to adding text-related conditional controls on top of image retrieval, with primary tasks including FashionIQ [55] and CIRR [32] in the $\mathrm{I + T\to I}$ format. Composed text retrieval involves com + +Table 7. Comparison on Zero-Shot Multimodal Ranking. "-PE" refers to testing conducted solely through prompt engineering. + +
COCO-5KFlickr
I2TT2II2TT2I
R1R10R1R10R1R10R1R10
AlBEAF----94.199.782.898.1
Qwen2VL-PE56.168.931.158.678.987.351.971.0
Qwen2VL-Rank72.895.139.077.494.299.870.296.2
DeKR-Rank79.396.160.288.196.999.984.898.4
+ +bining image information to assist text-based retrieval, with main tasks such as WebQA [5] in the $\mathrm{T}\rightarrow \mathrm{I} + \mathrm{T}$ format and ReMuQ [34] in the $\mathrm{I} + \mathrm{T}\rightarrow \mathrm{T}$ format. As shown in Table 5, our performance on composed image retrieval surpasses that of Pic2Word [46], a model trained with specifically designed pseudo-word tasks, exceeding it by 7.3 points on the FashionIQ validation set and by 4.3 points on the Val-All task, which utilizes a larger corpus. Additionally, most of our trained composed modality retrieval data were filtered from Vista's training data, yet our composed image retrieval performance is 12.7 points higher than Vista's [67]. This indicates that better multimodal alignment capabilities can also enhance composed retrieval performance. In comparison with E5-V [17], our DeKR outperforms E5-V's 8B-sized model with a smaller 7B model, achieving SoTA performance on composed image retrieval tasks. + +As shown in Table 6, our DeKR outperforms Vista and other zero-shot proprietary models [18, 35] on two composed text retrieval tasks, achieving SOTA performance under zero-shot settings. It is noteworthy that since Vista does not finetune the text encoder component, it can be regarded as BGEm3 [7] in text retrieval tasks, serving as a strong baseline for text retrieval. We found that DeKR exceeds BGEm3 by 11.1 points in the WebQA text retrieval task, indicating that DeKR not only possesses unified multimodal representation capabilities but also exhibits excellent text representation abilities. + +Multimodal ranking. Multimodal ranking tasks typically refer to image-text matching tasks, where both image and text are inputted to produce a matching score based on their correlation. Often, due to differences in input between representation tasks and ranking tasks, separate models are used for representation and ranking. The DeKR model, however, leverages MLLMs to handle combined multimodal inputs, allowing it to perform both representation and ranking tasks within a unified architecture. + +To verify whether DeKR's representation capabilities can positively influence ranking tasks, as shown in Table 7, we compared the ranking performance of three approaches: directly using prompt engineering with Qwen2VL, training ranking data based on Qwen2VL, and training ranking data based on DeKR. The ranking prompt for Qwen2VL was: {"Image}{Text}Do they match? A.Yes B.No". The match- + +Table 8. Comparison on Fine-tuned Multimodal Ranking. + +
Flickr I2TFlickr T2I
R1R5R10R1R5R10
Uniter87.398.099.275.694.196.8
AlBEAF95.999.8100.085.697.598.9
mPlug297.2100.0100.088.197.699.1
BLIP97.499.899.987.697.799.0
BLIP2-L96.9100.0100.088.697.698.9
BLIP2-g97.6100.0100.089.798.199.0
Qwen2VL95.199.499.870.292.096.5
DeKR98.4100.0100.090.198.299.2
+ +ing score was obtained by normalizing the probabilities of options A and B from the output vocabulary. For the zero-shot setting ranking training data, we randomly selected 500k image-text pairs from DeKon5M and concatenated them into a unified input format as positive samples. Additionally, we selected one negative sample for each modality and followed the multi-range difficulty selection strategy used in mGTE [66]. + +Surprisingly, we found that Qwen2VL-PE, without any data training, already possessed ranking capabilities, indicating that MLLMs inherently have strong image-text alignment abilities. Furthermore, when comparing the two models trained on ranking data—Qwen2VL+Rank and DeKR+Rank—we observed that the latter outperformed the former by 14.6 points on the Text-to-Image task. This demonstrates that DeKR's representation capabilities have a crucial impact on ranking tasks. + +Additionally, as shown in Table 8, we report the performance of DeKR after fine-tuning the downstream ranking dataset Flickr. We compared it with previous ranking models and found that DeKR achieved SoTA results on both Image-to-Text and Text-to-Image ranking. This indicates that DeKR successfully unifies the architecture for both representation and ranking tasks. Furthermore, when we finetuned Qwen2VL on the Flickr data, we observed a performance gap of up to 20 points on the T2I task, further emphasizing the necessity of DeKR's representation training. + +# 4.4. Qualitative Analysis + +Cross-modal retrieval. In Figure 8(a), we compare DeKR and SigLip on Top-1 recall for short and long text cross-modal retrieval. Both models succeed with short texts, with DeKR preferring detailed descriptions and SigLip favoring broader summaries. For long texts, SigLip retrieves wrong answers when accurate on objects but misses color and position details, suggesting CLIP models focus on object alignment but neglect action and descriptive nuances. + +Composed modality retrieval. In Figure 8(b), we explore DeKR's performance on more complex multimodal retrieval tasks. The image editing task builds upon the image retrieval task. As shown at the top of the figure, DeKR + +![](images/7501d5b1152066c09772a22cf1400413ae31415f2661a6d38db823765e5574ff.jpg) +Figure 8. Case analysis of cross and composed modality retrieval. + +first interprets the attributes of the two gorillas in the left image, such as size, posture, and orientation. It then processes the text instruction to only reduce the gorillas' size while maintaining other attributes, successfully retrieving the correct result depicted in the right image. This demonstrates that fine-grained alignment remains crucial in complex retrieval tasks. Additionally, we explored the VQA retrieval task, where DeKR must perform VQA reasoning internally and represent the outcome to retrieve the correct answer, indicating that DeKR retains the original MLLM's reasoning capabilities well. Finally, we assessed DeKR's ability to recall multimodal information and found that it could recognize and retrieve data with the text "I AM A MAN" from the image, showcasing its OCR capabilities, which are also inherent in the original MLLM. These cases suggest that if a retrieval model can effectively inherit the original capabilities of MLLM, there is significant potential for retrieval tasks, paving the way for further exploration of DeKR. + +# 5. Conclusions + +In this paper, we introduce DeKR, a unified multimodal retrieval model trained with MLLMs. DeKR integrates text, images, videos, and composed image-text into a unified embedding space and consolidates representation and ranking tasks within one framework. To enhance multimodal alignment, we develop DeKon5M, a contrastive learning dataset rich in dense knowledge. Extensive experiments demonstrate that DeKR achieves SoTA performance in cross-modal retrieval, composed image-text retrieval, and ranking tasks. This shows DeKR's advancement, and DeKon5M contributes to future academic research. + +# References + +[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 5 +[2] Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. Llm2vec: Large language models are secretly powerful text encoders. CoRR, abs/2404.05961, 2024. 1, 6 +[3] Rumeysa Bodur, Erhan Gundogdu, Binod Bhattarai, Tae-Kyun Kim, Michael Donoser, and Loris Bazzani. iedit: Localised text-guided image editing with weak supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024 - Workshops, Seattle, WA, USA, June 17-18, 2024, pages 7426-7435. IEEE, 2024. 5 +[4] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.4.5 +[5] Yingshan Chang, Guihong Cao, Mridu Narang, Jianfeng Gao, Hisami Suzuki, and Yonatan Bisk. Webqa: Multihop and multimodal QA. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16474-16483. IEEE, 2022. 7 +[6] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for A lite vision-language model. CoRR, abs/2402.11684, 2024. 4, 5 +[7] Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. Bge m3-embedding: Multilingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation, 2024. 1, 2, 7 +[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 3 +[9] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1, 7 +[10] Yuhao Cui, Zhou Yu, Chunqi Wang, Zhongzhou Zhao, Ji Zhang, Meng Wang, and Jun Yu. Rosita: Enhancing vision-and-language semantic alignments via cross- and intra-modal knowledge integration. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 2, 3 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1 + +[12] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan C. Russell. Localizing moments in video with natural language. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 5804-5813. IEEE Computer Society, 2017. 7 +[13] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13405-13417, 2024. 4, 5 +[14] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. 5 +[15] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 1, 2, 3, 7 +[16] Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. Mistral 7b. CoRR, abs/2310.06825, 2023. 2 +[17] Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-V: universal embeddings with multimodal large language models. CoRR, abs/2407.12580, 2024. 1, 2, 7 +[18] Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wentau Yih. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781, Online, 2020. Association for Computational Linguistics. 7 +[19] Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin Kim, Minkyung Cho, Jy yong Sohn, and Chanyeol Choi. Linq-embed-mistral: elevating text retrieval with improved gpt data through task-specific control and quality refinement. Linq AI Research Blog, 2024. 1 +[20] Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nv-embed: Improved techniques for training llms as generalist embedding models. CoRR, abs/2405.17428, 2024. 1 +[21] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. CoRR, abs/2408.03326, 2024. 5 +[22] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 3 + +[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. +[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 19730-19742. PMLR, 2023. 3 +[25] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. ECCV 2020, 2020. 3 +[26] Xianhang Li, Zeyu Wang, and Cihang Xie. An inverse scaling law for CLIP training. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 1 +[27] Xianhang Li, Zeyu Wang, and Cihang Xie. Clipa-v2: Scaling CLIP training with $81.1\%$ zero-shot imagenet accuracy within a $10,000 budget; an extra $4,000 unlocks 81.8% accuracy. CoRR, abs/2306.15658, 2023. 1 +[28] Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281, 2023. 1, 2, 6 +[29] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V, pages 740-755. Springer, 2014. 7 +[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3, 4, 5 +[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 1, 2 +[32] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre-trained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2125–2134, 2021. 7 +[33] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 1, 7 +[34] Man Luo, Zhiyuan Fang, Tejas Gokhale, Yezhou Yang, and Chitta Baral. End-to-end knowledge retrieval with multimodal queries. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 8573-8589. Association for Computational Linguistics, 2023. 7 +[35] Man Luo, Zhiyuan Fang, Tejas Gokhale, Yezhou Yang, and Chitta Baral. End-to-end knowledge retrieval with multi- + +modal queries. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 8573-8589. Association for Computational Linguistics, 2023. 7 +[36] Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. Sfr-embedding-mistral:enhance text retrieval with transfer learning. Salesforce AI Research Blog, 2024. 1 +[37] Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. Sfr-embedding-2: Advanced text embedding with multi-stage training, 2024. 1 +[38] Niklas Muennighoff. Sgpt: Gpt sentence embeddings for semantic search. arXiv preprint arXiv:2202.08904, 2022. 6 +[39] Niklas Muennighoff, Nouamane Tazi, Loïc Magne, and Nils Reimers. MTEB: massive text embedding benchmark. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2023, Dubrovnik, Croatia, May 2-6, 2023, pages 2006-2029. Association for Computational Linguistics, 2023. 2 +[40] Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. Generative representational instruction tuning, 2024. 2, 6 +[41] Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas Tezak, Jong Wook Kim, Chris Hallacy, Johannes Heidecke, Pranav Shyam, Boris Power, Tyna Eloundou Nekoul, Girish Sastry, Gretchen Krueger, David Schnurr, Felipe Petroski Such, Kenny Hsu, Madeleine Thompson, Tabarak Khan, Toki Sherbakov, Joanne Jang, Peter Welinder, and Lilian Weng. Text and code embeddings by contrastive pretraining. CoRR, abs/2201.10005, 2022. 6 +[42] OpenAI. Gpt-4v(ision) system card. 2023. 4 +[43] OpenAI. Gpt-4o system card. 2024. 4 +[44] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2011. 5 +[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 1, 2, 7 +[46] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 19305-19314. IEEE, 2023. 7 +[47] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: open dataset of clip-filtered 400 million image-text pairs. CoRR, abs/2111.02114, 2021. 2 + +[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2 +[49] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 1, 2, 7 +[50] Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. EVA-CLIP-18B: scaling CLIP to 18 billion parameters. CoRR, abs/2402.04252, 2024. 1, 2 +[51] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024. 4, 5 +[52] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 11897-11916. Association for Computational Linguistics, 2024. 1, 2 +[53] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 7 +[54] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 1, 5, 7 +[55] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The fashion iq dataset: Retrieving images by combining side information and relative natural language feedback. CVPR, 2021. 7 +[56] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), Online, 2021. Association for Computational Linguistics. 1 +[57] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, + +Guohai Xu, Ji Zhang, Songfang Huang, Fei Huang, and Jingren Zhou. mplug-2: A modularized multi-modal foundation model across text, image and video. ArXiv, abs/2302.00402, 2023. 3 +[58] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 5288-5296. IEEE Computer Society, 2016. 7 +[59] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2:67-78, 2014. 4, 7 +[60] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 3208-3216. AAAI Press, 2021. 1, 3 +[61] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 11941-11952. IEEE, 2023. 1, 2, 7 +[62] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 5 +[63] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021. 3 +[64] Peitian Zhang, Shitao Xiao, Zheng Liu, Zhicheng Dou, and Jian-Yun Nie. Retrieve anything to augment large language models. CoRR, abs/2310.07554, 2023. 1 +[65] Xin Zhang, Zehan Li, Yanzhao Zhang, Dingkun Long, Pengjun Xie, Meishan Zhang, and Min Zhang. Language models are universal embeddeders. CoRR, abs/2310.08232, 2023. 6 +[66] Xin Zhang, Yanzhao Zhang, Dingkun Long, Wen Xie, Ziqi Dai, Jialong Tang, Huan Lin, Baosong Yang, Pengjun Xie, Fei Huang, Meishan Zhang, Wenjie Li, and Min Zhang. mgte: Generalized long-context text representation and reranking models for multilingual text retrieval. CoRR, abs/2407.19669, 2024. 8 +[67] Junjie Zhou, Zheng Liu, Shitao Xiao, Bo Zhao, and Yongping Xiong. VISTA: visualized text embedding for universal multi-modal retrieval. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 3185-3200. Association for Computational Linguistics, 2024. 1, 2, 5, 7 \ No newline at end of file diff --git a/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/images.zip b/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2f22e352534c6a25e77d88fa36016a9e4315b19d --- /dev/null +++ b/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f64721ed12cdf5c64113a5557b84cd723b5fcad1ee6874a9fdc44afa1150a03 +size 794174 diff --git a/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/layout.json b/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5b9cf833c0de9f1e3fd31f15d25925d51b14187e --- /dev/null +++ b/2025/Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models/layout.json @@ -0,0 +1,8044 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 103, + 455, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 103, + 455, + 140 + ], + "spans": [ + { + "bbox": [ + 157, + 103, + 455, + 140 + ], + "type": "text", + "content": "Incorporating Dense Knowledge Alignment into Unified Multimodal Representation Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "spans": [ + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "text", + "content": "Yuhao Cui" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "text", + "content": " Xinxing Zu" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "text", + "content": " Wenhua Zhang" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "text", + "content": " Zhongzhou Zhao" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "text", + "content": " Jinyang Gao" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "text", + "content": "Alibaba Cloud Computing, China \n" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "text", + "content": "Dalian University of Technology, China \n" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 81, + 160, + 529, + 218 + ], + "type": "text", + "content": "Tongyi Lab, China \ncyhbrilliant@gmail.com, patrickfollow@gmail.com, 1062894314zwh@mail.dlut.edu.cn, flyhdq@gmail.com, jinyang.gjy@alibaba-inc.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 271, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 559 + ], + "type": "text", + "content": "Leveraging Large Language Models (LLMs) for text representation has achieved significant success, but the exploration of using Multimodal LLMs (MLLMs) for multimodal representation remains limited. Previous MLLM-based representation studies have primarily focused on unifying the embedding space while neglecting the importance of multimodal alignment. As a result, their cross-modal retrieval performance falls markedly behind that of the CLIP series models. To address this, in our work, we 1) construct DeKon5M, a contrastive learning dataset enriched with dense multimodal knowledge, which efficiently enhances multimodal alignment capabilities in representation tasks. 2) design a framework for training unified representation on MLLMs. Building upon this unified representation framework and the dense knowledge dataset DeKon5M, we developed the Dense Knowledge Representation model DeKR on Qwen2VL. Through extensive quantitative and qualitative experiments, our results demonstrate that DeKR not only aligns text, image, video, and text-image combinations within a unified embedding space but also achieves cross-modal retrieval performance comparable to SoTA CLIP series models. This fully validates the effectiveness of our approach and provides new insights for multimodal representation research." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "spans": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Multimodal retrieval has consistently been a crucial component in applications such as search systems and retrieval-augmented generation. Due to the outstanding cross-modal retrieval performance of CLIP [45], models from the CLIP family [9, 15, 26, 27, 33, 49, 50, 54, 56, 61] are widely employed in multimodal retrieval tasks. In light of the tremendous success Large Language Models (LLMs) have achieved in text representation tasks [2, 7, 19, 20, 28, 36, 37, 52, 64], recent explorations have begun to investigate" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 243, + 552, + 452 + ], + "blocks": [ + { + "bbox": [ + 316, + 243, + 552, + 452 + ], + "lines": [ + { + "bbox": [ + 316, + 243, + 552, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 552, + 452 + ], + "type": "image", + "image_path": "581ad1c71b1322191eaebc0f314032668ac3ee9b063c6ef9129978390c692dfe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 454, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 475 + ], + "type": "text", + "content": "Figure 1. Comparison of performance with SoTA models in different retrieval tasks. \"ZS\" stands for Zero-shot." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 483, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 686 + ], + "type": "text", + "content": "representation tasks based on Multimodal Large Language Models (MLLMs). To obtain combined representations of image and text modalities, Zhou et al. [67] encoded images into visual features using a Vision Transformer (ViT) [11] and then input them together with text tokens into a text encoder. This process can be approximated as an MLLM. Jiang et al. [17] were the first to utilize an MLLM for multimodal representation learning directly. Building upon LLaVA-NeXT [31], they trained using text-pair data to embed image and text modalities within a unified embedding space. Their retrieval performance demonstrates the feasibility of the MLLM-based representation model. However, previous MLLM-based representations shared a common issue: their cross-modal performance significantly lagged behind the SOTA CLIP series models. We believe this gap arises because their training data didn't activate the multimodal alignment capabilities within MLLMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "As illustrated in Figure 2(a), we observed that early multimodal pretraining studies, such as Ernie-VIL [60] and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29733" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 289, + 277 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 289, + 277 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 289, + 277 + ], + "type": "image", + "image_path": "6ba0ce7b38a17f73c96dd4e8d29ec1b51ee54938dca36721b2d5bd90ac1deeda.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 278, + 265, + 289 + ], + "lines": [ + { + "bbox": [ + 85, + 278, + 265, + 289 + ], + "spans": [ + { + "bbox": [ + 85, + 278, + 265, + 289 + ], + "type": "text", + "content": "Figure 2. Various uses of multimodal knowledge." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 293, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 534 + ], + "type": "text", + "content": "Rosita [10], addressed the alignment problem by increasing training on multimodal knowledge, i.e., visual-centric words (objects, colors, actions, etc.). And as shown in Figure 2(b), training MLLMs typically uses dense image descriptions rich in visual-centric words to align ViT outputs with LLM inputs. Thus, we hypothesize that the key to multimodal alignment is training more knowledge that links vision and text. To enhance representation alignment, as shown in Figure 2(c), we introduce a denser multimodal knowledge data into contrastive learning. We were surprised to find that performance after training with dense knowledge data showed a significant improvement compared to training with only sparse knowledge data. To further enhance uniformity, we added video-text retrieval data and composed modality retrieval data, resulting in DeKon5M. Based on this dataset, we constructed the dense knowledge representation model DeKR on Qwen2VL. To demonstrate its superiority, as shown in Figure 1, we tested on four cross-modal retrieval tasks, four composed retrieval tasks, and two ranking tasks, achieving SoTA results on all." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 544, + 143, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 143, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 143, + 556 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 563, + 244, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 244, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 244, + 577 + ], + "type": "text", + "content": "2.1. LLM-based Representation Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Wang et al. [52], building upon the LLM model MISTrial [16], constructed text representation tasks using only a small amount of synthetic data and achieved outstanding rankings on the MTEB [39]. Since then, numerous studies have delved into exploiting the representational potential of LLMs. Muennighoff et al. [40] jointly trained representation and generation tasks, enabling a single model to encompass all stages of retrieval-augmented generation tasks. Recently, Li et al. [28] and Chen et al. [7] have respectively upgraded previous works to versions tailored for large language model representations based on Qwen2 and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "Lemma2. In summary, large language models have already been assumed a dominant position in text representation tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 103, + 504, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 103, + 504, + 115 + ], + "spans": [ + { + "bbox": [ + 313, + 103, + 504, + 115 + ], + "type": "text", + "content": "2.2. Multimodal Representation Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 120, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 555, + 324 + ], + "type": "text", + "content": "Since the introduction of CLIP [45], multimodal representation tasks have garnered widespread attention. Schuhmann et al. [47, 48] utilized the CLIP model for large-scale data cleaning, presenting the first open-source billion-level image-text pair dataset, LAION-5B. Building on LAION-5B, Gabriel et al. [15] developed OpenCLIP, an open-source multimodal representation model that surpassed CLIP in performance for the first time. Sun et al. [49, 50] explored the scaling laws of representation models, expanding the model sizes to larger scales of 5B, 8B, and 18B parameters. Zhai et al. [61] investigated the advantages of sigmoid loss in contrastive learning, achieving SoTA with only a 400M-parameter model; however, their data details were not publicly disclosed. Despite ongoing research into cross-modal retrieval tasks, the increasing complexity of retrieval applications has rendered existing approaches insufficient for accommodating the diverse range of retrieval scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 324, + 556, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 556, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 556, + 527 + ], + "type": "text", + "content": "To accommodate composed retrieval tasks, Zhou et al. [67], drawing inspiration from the style of MLLMs, encoded visual information using a ViT and then inputted it alongside text tokens into the text encoder BGEm3 [7]. This improved fusion approach led to significant performance enhancements. Jiang et al. [17] were the first to propose using MLLMs for unified modality representation. They fine-tuned text data based on LLaVA-NeXT [31] to enable the model to represent both image and text within a unified embedding space. However, the cross-modal retrieval performance of these MLLM-based representation models still lags considerably behind SoTA CLIP series models. In this work, we introduce DeKR, a powerful unified representation model capable of representing text, image, video, and text-image combination. Using these representations, we outperform previous works on image-text retrieval, videotext retrieval, and composed retrieval tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 528, + 556, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 556, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 556, + 647 + ], + "type": "text", + "content": "Our contributions are as follows: 1) We are the first to demonstrate that MLLM-based models can surpass the CLIP series in cross-modal retrieval tasks. 2) To enhance alignment, we introduced a novel method of using dense knowledge data for contrastive learning and dedicated significant effort to building the DeKon5M dataset. 3) Leveraging DeKon5M, we trained DeKR, the first model capable of representing text, images, videos, and composed modalities, achieving SoTA results on 14 tasks using only one-thousandth of the training data used by the CLIP series." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 658, + 547, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 547, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 547, + 671 + ], + "type": "text", + "content": "3. Learning Unified Multimodal Embeddings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "In this section, we first introduce the framework for training DeKR, then analyze how dense multimodal knowledge activates multimodal alignment capabilities, and finally present" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29734" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 294, + 238 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 294, + 238 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 294, + 238 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 294, + 238 + ], + "type": "image", + "image_path": "c2d98911dcf53b2aea5311c37d15893db265bd4778a1e5869e90c65f5a11216d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 243, + 295, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 295, + 277 + ], + "type": "text", + "content": "Figure 3. Overview of model architecture. To reduce computational overhead, the Vision Encoder is kept frozen during training, while the parameters of the Projector and the LLM are trained." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 253, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 253, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 253, + 300 + ], + "type": "text", + "content": "the construction details of the DeKon5M dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 124, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 124, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 124, + 324 + ], + "type": "text", + "content": "3.1. Modeling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": "Multimodal large language models. Recent MLLMs tend toward a unified structure, comprising three core components: a visual encoder " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": ", which processes input from a single image or multiple video frames, typically utilizing a pretrained Vision Transformer; a projector " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": " that maps visual information into the text token space, such as a Q-former [24] or MLP layers [30]; and a pretrained large language model " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": ", which contains extensive world knowledge and has fine-grained token-level understanding capabilities. An MLLM can be represented as " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "\\mathrm{MLLM}(i,v,t) = L(P(V(i,v))\\odot t)" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": " denotes the concatenation operation, " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": " stands for images, " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": " for videos, and " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": " for text tokens. Unified representation using MLLMs. The CLIP model employs separate visual and text encoders to process images and texts into image embedding " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "e_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": " and text embedding " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "e_t \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": ". In contrast, since an MLLM is capable of handling both visual and textual inputs simultaneously, we utilize a single, parameter-shared MLLM as a unified encoder. For cross-modal and unimodal embedding tasks, each input consists of one element among " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": ", such as " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "\\mathrm{MLLM}(i)" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "\\mathrm{MLLM}(t)" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": ". In the case of composed modal embedding, combinations such as " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "\\mathrm{MLLM}(i,t)" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": " can be inputted simultaneously. As illustrated in Figure 3, although the input forms differ for each task, the output consistently comprises a token sequence, " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "y = \\mathrm{MLLM}(x)" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "inline_equation", + "content": "y \\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 55, + 331, + 296, + 675 + ], + "type": "text", + "content": ". We use the output corresponding to the last token as the representation of the entire input, followed by a linear layer that maps it to the specified embedding dimension." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Contrastive learning. To accelerate convergence, we made some minor modifications to the contrastive learning loss. Specifically, the standard contrastive learning loss treats" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "other samples within the current batch as negative samples, which may result in the inability to find hard negative samples when the batch size is not sufficiently large. We manually introduce hard negative samples for training and minimize the following objectives:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 320, + 137, + 553, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 137, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 320, + 137, + 553, + 194 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {c o n}} (x, y, \\tilde {y}) = \\frac {1}{\\mathcal {B}} \\sum_ {i = 1} ^ {\\mathcal {B}} \\log \\frac {e ^ {\\tau x _ {i} y _ {i}}}{\\sum_ {j , k = 1} ^ {\\mathcal {B}} \\left(e ^ {\\tau x _ {i} y _ {j}} + e ^ {\\tau x _ {i} \\tilde {y} _ {k}}\\right)} \\tag {1} \\\\ \\mathcal {L} (x, y, \\tilde {x}, \\tilde {y}) = \\frac {1}{2} \\left(\\mathcal {L} _ {c o n} (x, y, \\tilde {y}) + \\mathcal {L} _ {c o n} (y, x, \\tilde {x})\\right) \\\\ \\end{array}", + "image_path": "538d748f03754bd085732b17874fd9a51a8f9a7589896768d1790c343a0e4e59.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": " is a learnable scaling parameter, " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": " are the embedding vectors to be contrasted, " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": " are the hard negative samples corresponding to " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": " is the batch size. In addition, we use caching operations to increase the batch size, similar to previous work [15]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 258, + 556, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 556, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 556, + 402 + ], + "type": "text", + "content": "Integrating ranking task. To enhance accuracy, ranking models [8, 22-25, 57, 63] typically compute similarity scores online by simultaneously inputting the content to be matched, whereas representation models computes and stores vectors separately, then calculates scores offline. Leveraging the flexible input capabilities of MLLMs, we can easily extend representation models into ranking models. Structurally, all parameters are shared except for the final linear layer. To prevent knowledge confusion, we introduce specific instructions: the instruction for representation tasks is \"summarize the content.\" and the instruction for ranking tasks is \"Do they match?\"." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 407, + 479, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 479, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 479, + 420 + ], + "type": "text", + "content": "3.2. Dense Multimodal Knowledge" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 424, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 554, + 484 + ], + "type": "text", + "content": "Previous MLLM-based representation models exhibited relatively weak cross-modal retrieval performance, reflecting deficiencies in their multimodal alignment capabilities. Therefore, enhancing the efficiency of multimodal alignment is crucial for MLLM-based representation models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 554, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 663 + ], + "type": "text", + "content": "Figure 4(a) illustrates our analysis of cross-modal retrieval data, revealing that low-confidence negative samples typically exhibit three types of errors: Action Errors, such as riding a motorcycle as shown in the figure; Object Errors, such as the condition of oranges; and Description Errors, such as the color of the bus. These errors are concentrated on visual-related words, underscoring the importance of these words in addressing cross-modal retrieval challenges. Following Ernie-VIL [60] and Rosita [10], we refer to such words that align the image modality with the text modality as multimodal knowledge. Generally, assuming that the descriptive text accurately corresponds to the image content, the knowledge can be mapped to certain parts of speech: action knowledge to verbs, object knowledge to common nouns1, and description knowledge to adjectives." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 663, + 553, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 663, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 663, + 553, + 688 + ], + "type": "text", + "content": "As shown in Figure 4(b), we present a comparison between dense knowledge data and sparse knowledge data." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "text", + "content": "Nouns except proper nouns. Proper nouns, such as names and locations, are considered to have weaker visual associations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29735" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 83, + 116, + 150 + ], + "blocks": [ + { + "bbox": [ + 61, + 83, + 116, + 150 + ], + "lines": [ + { + "bbox": [ + 61, + 83, + 116, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 83, + 116, + 150 + ], + "type": "image", + "image_path": "2d32fde97154de93ae05c6ec83b342540a16eca06e4fefbb223ed92badf4a4dc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 87, + 169, + 99 + ], + "lines": [ + { + "bbox": [ + 118, + 87, + 169, + 99 + ], + "spans": [ + { + "bbox": [ + 118, + 87, + 169, + 99 + ], + "type": "text", + "content": "Action Error" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 101, + 231, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 101, + 231, + 121 + ], + "spans": [ + { + "bbox": [ + 118, + 101, + 231, + 121 + ], + "type": "text", + "content": "A man wearing a helmet is .... sitting on his blue motorcycle" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 124, + 196, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 124, + 196, + 145 + ], + "spans": [ + { + "bbox": [ + 118, + 124, + 196, + 145 + ], + "type": "text", + "content": "A man is riding a blue motorcycle outside" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 178, + 176, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 178, + 176, + 198 + ], + "spans": [ + { + "bbox": [ + 69, + 178, + 176, + 198 + ], + "type": "text", + "content": "A yellow beverage with a large ice cube in a glass" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 200, + 166, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 200, + 166, + 220 + ], + "spans": [ + { + "bbox": [ + 69, + 200, + 166, + 220 + ], + "type": "text", + "content": "A glass full of juice with an orange slice sticking out of" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 178, + 163, + 231, + 228 + ], + "blocks": [ + { + "bbox": [ + 125, + 165, + 175, + 175 + ], + "lines": [ + { + "bbox": [ + 125, + 165, + 175, + 175 + ], + "spans": [ + { + "bbox": [ + 125, + 165, + 175, + 175 + ], + "type": "text", + "content": "Object Error" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 163, + 231, + 228 + ], + "lines": [ + { + "bbox": [ + 178, + 163, + 231, + 228 + ], + "spans": [ + { + "bbox": [ + 178, + 163, + 231, + 228 + ], + "type": "image", + "image_path": "570269aa70ecf0196b9adc85d05a63340b06386b5c9c9722fe09af18a44dc64b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 63, + 239, + 132, + 301 + ], + "blocks": [ + { + "bbox": [ + 63, + 239, + 132, + 301 + ], + "lines": [ + { + "bbox": [ + 63, + 239, + 132, + 301 + ], + "spans": [ + { + "bbox": [ + 63, + 239, + 132, + 301 + ], + "type": "image", + "image_path": "5694f179a3df24b0e9c89dd33c1042731a7abd6c05b8903c91f549027dc876ba.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 243, + 203, + 254 + ], + "lines": [ + { + "bbox": [ + 134, + 243, + 203, + 254 + ], + "spans": [ + { + "bbox": [ + 134, + 243, + 203, + 254 + ], + "type": "text", + "content": "Description Error" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 314, + 154, + 323 + ], + "lines": [ + { + "bbox": [ + 141, + 314, + 154, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 314, + 154, + 323 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 256, + 231, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 256, + 231, + 277 + ], + "spans": [ + { + "bbox": [ + 134, + 256, + 231, + 277 + ], + "type": "text", + "content": "A red and black bus being towed in by a flatbed" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 280, + 223, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 280, + 223, + 300 + ], + "spans": [ + { + "bbox": [ + 134, + 280, + 223, + 300 + ], + "type": "text", + "content": "Blue and white tow truck pulling a blue bus behind" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 239, + 83, + 445, + 184 + ], + "blocks": [ + { + "bbox": [ + 239, + 83, + 445, + 184 + ], + "lines": [ + { + "bbox": [ + 239, + 83, + 445, + 184 + ], + "spans": [ + { + "bbox": [ + 239, + 83, + 445, + 184 + ], + "type": "image", + "image_path": "2c0347ba63139ca34dac291981429906c146664ccc2940c8a118721809770030.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 186, + 362, + 198 + ], + "lines": [ + { + "bbox": [ + 348, + 186, + 362, + 198 + ], + "spans": [ + { + "bbox": [ + 348, + 186, + 362, + 198 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 239, + 199, + 455, + 312 + ], + "blocks": [ + { + "bbox": [ + 239, + 199, + 455, + 312 + ], + "lines": [ + { + "bbox": [ + 239, + 199, + 455, + 312 + ], + "spans": [ + { + "bbox": [ + 239, + 199, + 455, + 312 + ], + "type": "image", + "image_path": "731cd7b728e028fd3084f7a1bf28f90b0df7918e713b537b5a29c2faff77f9a4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 314, + 361, + 324 + ], + "lines": [ + { + "bbox": [ + 349, + 314, + 361, + 324 + ], + "spans": [ + { + "bbox": [ + 349, + 314, + 361, + 324 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 453, + 70, + 548, + 150 + ], + "blocks": [ + { + "bbox": [ + 453, + 70, + 548, + 150 + ], + "lines": [ + { + "bbox": [ + 453, + 70, + 548, + 150 + ], + "spans": [ + { + "bbox": [ + 453, + 70, + 548, + 150 + ], + "type": "image", + "image_path": "1e75efa9cec3ec57cf44a43d22c6e0e86f8956841591d5f2ab572d436116aa4b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 451, + 152, + 549, + 228 + ], + "blocks": [ + { + "bbox": [ + 451, + 152, + 549, + 228 + ], + "lines": [ + { + "bbox": [ + 451, + 152, + 549, + 228 + ], + "spans": [ + { + "bbox": [ + 451, + 152, + 549, + 228 + ], + "type": "image", + "image_path": "cf53e82068a7ea56f4b32d6130223be5d3251fd7fb947cb9022601dbbc9ce106.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 455, + 230, + 547, + 311 + ], + "blocks": [ + { + "bbox": [ + 455, + 230, + 547, + 311 + ], + "lines": [ + { + "bbox": [ + 455, + 230, + 547, + 311 + ], + "spans": [ + { + "bbox": [ + 455, + 230, + 547, + 311 + ], + "type": "image", + "image_path": "c4e07bc720f012132e657516827c9b6c7e3eeeab79259fbee3a1e9f23bdc6afb.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 494, + 312, + 507, + 323 + ], + "lines": [ + { + "bbox": [ + 494, + 312, + 507, + 323 + ], + "spans": [ + { + "bbox": [ + 494, + 312, + 507, + 323 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 366, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 462 + ], + "type": "text", + "content": "It can be observed that the dense knowledge data contains significantly more multimodal knowledge than the sparse knowledge data, enabling a more comprehensive and detailed description of almost all the content within an image. Our collected Dense Knowledge (DK) data typically includes more than 50 knowledge points per sample. Such dense descriptive data can significantly enhance the efficiency of alignment between images and text." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 474, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 715 + ], + "type": "text", + "content": "To further analyze the differences between DK data and other data, as illustrated in Figure 4(c), we quantified the average amount of this knowledge for each sample across various datasets. We found that the alt-text data used to train the CLIP series models contain very sparse knowledge. For instance, the average amount of knowledge in COYO [4] is only one-third that of Flickr30k [59], a downstream cross-modal retrieval dataset. This sparsity may contribute to the large data requirements for training the CLIP series models. Subsequently, we quantified the average amount of knowledge in the DK data. The average amount of knowledge in DK data is four times that of Flickr30k, indicating that DK data possesses higher image-text alignment efficiency. Furthermore, as illustrated in Figure 4(d), from the perspective of knowledge distribution, we found that the proportion of knowledge within sentences exceeds " + }, + { + "bbox": [ + 55, + 474, + 296, + 715 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 474, + 296, + 715 + ], + "type": "text", + "content": " in both DK data and Flickr30k, whereas it is only approximately " + }, + { + "bbox": [ + 55, + 474, + 296, + 715 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 55, + 474, + 296, + 715 + ], + "type": "text", + "content": " in COYO. This metric demonstrates that DK data maintains dense knowledge while ensuring efficient learning of such knowledge during model training." + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 315, + 365, + 556, + 472 + ], + "blocks": [ + { + "bbox": [ + 55, + 325, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 325, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 555, + 360 + ], + "type": "text", + "content": "Figure 4. Analysis and statistics of dense knowledge in the data. Subfigure (a) illustrates common cross-modal retrieval errors. Subfigure (b) contrasts dense knowledge data with sparse knowledge data. Subfigure (c) presents statistics on the average amount of knowledge per sample in the data. Subfigure (d) shows the distribution of knowledge across three types of data." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 365, + 556, + 472 + ], + "lines": [ + { + "bbox": [ + 315, + 365, + 556, + 472 + ], + "spans": [ + { + "bbox": [ + 315, + 365, + 556, + 472 + ], + "type": "image", + "image_path": "6c279653cec252d2c8d8598a8de6da6916325a517ce912e202ec4da5c83867ea.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 475, + 523, + 487 + ], + "lines": [ + { + "bbox": [ + 343, + 475, + 523, + 487 + ], + "spans": [ + { + "bbox": [ + 343, + 475, + 523, + 487 + ], + "type": "text", + "content": "Figure 5. Detailed data distribution of Dekon5M." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 502, + 405, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 405, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 405, + 514 + ], + "type": "text", + "content": "3.3. Data Curation" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "text", + "content": "Dense knowledge data. The validated high-quality dense knowledge data primarily originates from the first stage of training MLLMs. This data is mostly generated by trained dense descriptors [42, 43] to ensure that the majority of the content in the images is well-described. We collected a total of 2.3 million image-text pairs from four datasets [6, 13, 30, 51]. To further enhance the quality of the data, we mainly followed the Data Juicer processing pipeline to clean the text part. Then we used a CLIP score of 0.2 followed by a BLIP score of 0.45 to ensure alignment. In the end, we refined the dataset to 1.2 million image-text pairs." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": "Data statistics. As shown in Figure 5, the multimodal unified representation training dataset, DeKon5M, is divided into four parts: 1.5 million image-text retrieval data, 1.7 million video-text retrieval data, 1.6 million multimodal composed retrieval data, and 0.3 million unimodal retrieval" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29736" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 177, + 175 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 177, + 175 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 177, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 177, + 175 + ], + "type": "image", + "image_path": "9e7586eae04b73efd2415d023ca9e486fa7cd76df1154f5a53aa33a7dd03ebb1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 176, + 295, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 176, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 295, + 198 + ], + "type": "text", + "content": "Figure 6. The image retrieval performance of Dense Knowledge at different proportions, given a total data volume of 1 million." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 68, + 297, + 175 + ], + "blocks": [ + { + "bbox": [ + 179, + 68, + 297, + 175 + ], + "lines": [ + { + "bbox": [ + 179, + 68, + 297, + 175 + ], + "spans": [ + { + "bbox": [ + 179, + 68, + 297, + 175 + ], + "type": "image", + "image_path": "298ced7767608a427042373d3ad4acbf8493dfbba05e6a82935a27abe0d93f2b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 266, + 307, + 333 + ], + "blocks": [ + { + "bbox": [ + 55, + 206, + 295, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 262 + ], + "type": "text", + "content": "Table 1. Taking the model trained on 1 million COYO samples as the baseline, we compared the performance changes after incorporating different types of modality data. Specifically, I. refers to image-text paired data, V. refers to video-text paired data, and C. refers to composed modality retrieval data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 266, + 307, + 333 + ], + "lines": [ + { + "bbox": [ + 56, + 266, + 307, + 333 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 307, + 333 + ], + "type": "table", + "html": "
Training dataCOCO I2TCOCO T2IFlickr I2TFlickr T2I
COYO48.6833.6274.9061.80
DeKon-I.64.3044.9686.7071.74
DeKon-I.+V.69.4449.8891.3077.60
DeKon-I.+V.+C.70.4852.3891.7079.68
", + "image_path": "2535dcae271b9e1780d0082d7b88079c89a739adc6058ff5e8a0e6e8caa76c21.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 378, + 306, + 423 + ], + "blocks": [ + { + "bbox": [ + 55, + 342, + 295, + 374 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 374 + ], + "type": "text", + "content": "Table 2. The DeKR were trained separately using Causal Attention and Bidirectional Attention and were evaluated on image retrieval and video retrieval tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 378, + 306, + 423 + ], + "lines": [ + { + "bbox": [ + 56, + 378, + 306, + 423 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 306, + 423 + ], + "type": "table", + "html": "
Attn. typeCOCO I2TCOCO T2IMsrvtt V2TMsrvtt T2V
Causal69.2650.7843.0042.90
Bidirectional70.4852.3842.0042.40
", + "image_path": "3a9e211a0cb0c989da1e06d2ed73b18e31de80e58f3c45fb9a369afe60ded9f2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 434, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 434, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 54, + 434, + 295, + 625 + ], + "type": "text", + "content": "data. The image-text retrieval data primarily consists of cleaned dense knowledge data; however, these data tend to have excessively long text. To ensure that the model can generalize across different length distributions, we incorporated the cleaned SBU data [44] as short text data. The specific performance changes brought about by this adjustment will be analyzed in Section 4.1. Moreover, for the multimodal composed retrieval data, in addition to using some filtered VISTA dataset [67], we synthesized more data [3, 13, 14, 21, 30, 51, 62] for each retrieval form to ensure that the dataset size is roughly equivalent to that of the image-text retrieval data. For the video-text retrieval data, we did not perform extensive optimization; instead, we randomly selected a comparable number of samples from ordinary public video retrieval datasets [1, 54] to maintain a balanced ratio." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 643, + 135, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 135, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 135, + 657 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "This section is divided into four parts. The first two sections present ablation studies on the data and the model. The third section compares DeKR with SoTA retrieval models, and the final section conducts a qualitative analysis." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 554, + 252 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 554, + 252 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 554, + 252 + ], + "type": "image", + "image_path": "f05f2dbc618d7355236129127b969e5a7a8fd6dae14e1d82397f402ca1a29fe3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 255, + 555, + 287 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 287 + ], + "type": "text", + "content": "Figure 7. Exploring training scaling in model size, training data, and input resolution, and comparing training efficiency with CLIP series models." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 297, + 451, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 451, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 451, + 309 + ], + "type": "text", + "content": "4.1. Ablation Studies of Data" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 534 + ], + "type": "text", + "content": "Dense knowledge vs Sparse knowledge. As illustrated in Figure 6, we analyzed the proportion of dense knowledge within the training data. A dense knowledge ratio of 0 indicates the exclusive use of sparse knowledge data, while a ratio of 1 signifies the exclusive use of dense knowledge data. The sparse knowledge data primarily comprises COYO [4] and SBU [44], whereas the dense knowledge data mainly consists of SAM [13], LLaVA-Pretrain [30], and ALLaVA [6]. As shown in the figure, retrieval performance continuously improves with an increasing proportion of dense knowledge until it plateaus at an " + }, + { + "bbox": [ + 313, + 319, + 555, + 534 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 319, + 555, + 534 + ], + "type": "text", + "content": " ratio. We observed a significant performance decline between " + }, + { + "bbox": [ + 313, + 319, + 555, + 534 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 319, + 555, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 319, + 555, + 534 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 319, + 555, + 534 + ], + "type": "text", + "content": ", which we attribute to issues with data generalization. The dense knowledge data tends to be longer, whereas the downstream task primarily consists of short texts. Therefore, it is necessary to incorporate a small proportion of short text data into the training dataset to ensure that the model maintains generalization capabilities regarding text length." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "Retrieval task mixing. As shown in Table 1, we validated the effectiveness of multi-task training. We found that incorporating video retrieval tasks and composed modality retrieval tasks improved image retrieval performance by approximately 5 points and 2 points, respectively. We believe this improvement is due to multimodal alignment being performed from different dimensions. The image retrieval task primarily focuses on object alignment, the video retrieval task focuses on temporal action alignment, and the composed retrieval task focuses on detail variations and the unification of task formats. Additionally, models trained with dense knowledge data achieved an average improvement of 12.1 points in retrieval performance compared to models trained only with sparse knowledge in COYO, demonstrating the effectiveness of our dense knowledge data." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29737" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 106, + 555, + 301 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 3. Comparison on Zero-Shot Image-Text Retrieval. R* denotes Recall@*. The models under comparison are divided into three sections: the top section comprises the CLIP series models, the middle section consists of the previously MLLM-based models, and the bottom section includes our DeKR models of various sizes." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 106, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 56, + 106, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 555, + 301 + ], + "type": "table", + "html": "
COCO-5K I2TCOCO-5K T2IFlickr I2TFlickr T2IAvg.
R1R5R10R1R5R10R1R5R10R1R5R10
CLIP-L58.481.588.137.862.472.288.098.799.468.790.695.278.4
OpenCLIP-L62.183.490.346.170.779.488.798.499.275.092.595.681.8
OpenCLIP-bigG67.386.892.751.474.983.092.999.399.879.595.097.185.0
Eva02-L64.185.290.847.971.780.089.298.999.677.994.296.883.0
Eva02-E+68.787.792.851.275.082.793.999.499.878.894.296.885.1
SigLip-so400M70.2--52.0---------
InternVL-C70.689.093.554.177.384.694.799.699.981.796.098.286.6
VISTA25.549.762.025.749.059.849.879.186.152.078.386.058.6
E5-V(Llava-8B)62.083.689.752.076.584.788.298.799.479.595.097.683.9
DeKR-2B@44870.589.493.952.476.984.891.798.899.479.795.297.985.9
DeKR-7B@44873.391.095.053.878.485.993.199.499.980.395.998.387.1
", + "image_path": "6464efa73a4b44c6a540a2927674551cdf6e3df24ef186e0e437fe7421b90cbc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 343, + 293, + 460 + ], + "blocks": [ + { + "bbox": [ + 55, + 308, + 294, + 342 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 294, + 342 + ], + "type": "text", + "content": "Table 4. Comparison on Zero-Shot Video-Text Retrieval. Models marked with a \"+\" denote those trained exclusively for the image retrieval task, which use multi-frame averaging for inference." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 343, + 293, + 460 + ], + "lines": [ + { + "bbox": [ + 56, + 343, + 293, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 293, + 460 + ], + "type": "table", + "html": "
Msrvtt V2TMsrvtt T2VDiDemo V2TDiDemo T2VAvg.
CLIP-L*25.829.019.111.521.4
SigLip-so400M*35.836.330.435.834.6
CLIP4Clip27.232.0---
ViCLIP41.342.427.918.432.5
InternVideo-L39.640.733.531.536.3
DeKR-2B@44842.042.437.435.439.3
DeKR-7B@44843.742.939.138.241.0
", + "image_path": "a25cd1af08973ffa033d2e7349cc56b7e03b9ad077528669a2010f4487a6d329.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 477, + 200, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 200, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 200, + 489 + ], + "type": "text", + "content": "4.2. Ablation Studies of Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 617 + ], + "type": "text", + "content": "Causal vs Bidirectional. In previous work on LLM-based text representation model, two forms of attention—Causal Attention [38, 41, 65] and Bidirectional Attention [2, 28, 40]—have been employed. As shown in Table 2, we explored both Causal and Bidirectional Attention in our DeKR model. We found that Bidirectional Attention performs better for image retrieval, while Causal Attention is more effective for video retrieval. We believe this is because Causal Attention focuses more on temporal changes, whereas Bidirectional Attention emphasizes detailed information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Training scaling. As illustrated in Figure 7, we conducted scaling experiments across three dimensions: data volume, model size, and input resolution. For data scaling, we observed that increasing the number of training samples from 625K to 5M resulted in a steady improvement in model performance, exhibiting a convergent trend. A further horizontal comparison revealed that, in contrast to the CLIP series models which typically train over 10 billion samples, our" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 345, + 553, + 483 + ], + "blocks": [ + { + "bbox": [ + 313, + 308, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 342 + ], + "type": "text", + "content": "Table 5. Comparison on Zero-Shot Composed Image Retrieval. \"val-all\" refers to using queries from the validation set to retrieve from a corpus that includes the training, validation, and test sets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 345, + 553, + 483 + ], + "lines": [ + { + "bbox": [ + 315, + 345, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 315, + 345, + 553, + 483 + ], + "type": "table", + "html": "
FashionIQCIRR
DressShirtTopTeeAvg.val-allval-all
R10R10R10R10R5R5
CLIP-L----3.613.4
Pic2Word20.026.227.924.78.223.4
LinCIR20.929.128.826.3--
VISTA16.221.821.119.37.522.5
E5-V23.836.435.331.8--
DeKR-2B21.129.829.426.89.527.1
DeKR-7B26.635.433.932.012.533.7
", + "image_path": "1f4ceec2e112a16c694b9c4966e1d5638a9824688c03be45c97669fbce2e20c0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 497, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 545 + ], + "type": "text", + "content": "DeKR model achieved SoTA performance with data that is three orders of magnitude smaller. This demonstrates that dense knowledge can efficiently enhance the multimodal alignment capabilities of MLLMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "Regarding model size scaling, we found that upgrading the model from 2B to 7B parameters led to a performance increase of 2.8 points. This indicates that DeKR effectively inherits knowledge from larger models. Additionally, we explored the impact of input resolution on performance. Due to the limitations of the ViT patch size, the input resolution for CLIP series models is generally restricted to below 336. However, the pooling operation in the Projector component of DeKR overcomes this limitation. We tested input resolutions ranging from 224 to 560 and found that performance growth plateaued beyond a resolution of 448. This is attributed to the fact that most training data have an original resolution of around 500, and higher-resolution images are relatively scarce." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29738" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 84, + 286, + 248 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 267, + 81 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 267, + 81 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 267, + 81 + ], + "type": "text", + "content": "Table 6. Comparison on Composed Text Retrieval." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 84, + 286, + 248 + ], + "lines": [ + { + "bbox": [ + 63, + 84, + 286, + 248 + ], + "spans": [ + { + "bbox": [ + 63, + 84, + 286, + 248 + ], + "type": "table", + "html": "
ReMuQWebQA
R5R10T2ITT2T
Fine-tuned
ReViz-ICT76.283.3--
GeMKR90.392.7--
Zero-shot
CLIP-L0.81.4--
DPR43.448.8--
ReViz-ICT39.446.8--
VISTA84.7-65.169.5
DeKR-2B81.183.659.270.4
DeKR-7B86.087.969.480.6
", + "image_path": "098c02c7d3bd57bc0e0228387a9798172139a725e722122090bbc3a5907e18a9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 258, + 136, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 136, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 136, + 270 + ], + "type": "text", + "content": "4.3. Comparison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 275, + 295, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 394 + ], + "type": "text", + "content": "Zero-shot image-text retrieval. As shown in Table 3, we compared the CLIP series models [9, 15, 45, 49, 61], MLLM-based retrieval models [17, 67], and our DeKR on the Image-Text Retrieval task. We report their performances on the COCO-5K [29] and Flickr-1K [59] datasets. In terms of average scores, our DeKR-7B model surpasses the SoTA InternVL-C, which has 13B parameters, by 0.5 points and outperforms E5-V by 3.2 points. Notably, our work is the first to demonstrate that MLLM-based retrieval models can exceed the performance of the CLIP series models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 562 + ], + "type": "text", + "content": "Zero-shot video-text retrieval. As shown in Table 4, in addition to image retrieval, we also evaluated the multi-frame representation capabilities. We compared embedding models [33, 53, 54] trained for visual and textual alignment on the MSR-VTT [58] and DiDeMo [12] datasets. Besides comparing models trained for video retrieval tasks, we also used a frame averaging method to compare models trained solely for image retrieval tasks as a reference. We found that image retrieval models can generalize well to video retrieval tasks. In terms of average scores, our DeKR achieved SoTA performance at both the 2B and 7B scales. Our model surpasses InternVideo-L by 4.7 points, demonstrating that our unified embeddings can simultaneously represent image details and video temporal information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 563, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 694 + ], + "type": "text", + "content": "Zero-shot composed modality retrieval. Composed modality retrieval tasks require models to possess the capability of joint representation for both images and text. However, CLIP series models are difficult to support such retrieval formats due to the absence of modality fusion layers. Composed modality retrieval tasks are divided into two categories: composed image retrieval and composed text retrieval. Composed image retrieval refers to adding text-related conditional controls on top of image retrieval, with primary tasks including FashionIQ [55] and CIRR [32] in the " + }, + { + "bbox": [ + 55, + 563, + 295, + 694 + ], + "type": "inline_equation", + "content": "\\mathrm{I + T\\to I}" + }, + { + "bbox": [ + 55, + 563, + 295, + 694 + ], + "type": "text", + "content": " format. Composed text retrieval involves com" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 97, + 556, + 196 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 7. Comparison on Zero-Shot Multimodal Ranking. \"-PE\" refers to testing conducted solely through prompt engineering." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 97, + 556, + 196 + ], + "lines": [ + { + "bbox": [ + 315, + 97, + 556, + 196 + ], + "spans": [ + { + "bbox": [ + 315, + 97, + 556, + 196 + ], + "type": "table", + "html": "
COCO-5KFlickr
I2TT2II2TT2I
R1R10R1R10R1R10R1R10
AlBEAF----94.199.782.898.1
Qwen2VL-PE56.168.931.158.678.987.351.971.0
Qwen2VL-Rank72.895.139.077.494.299.870.296.2
DeKR-Rank79.396.160.288.196.999.984.898.4
", + "image_path": "a48f2bd064c04c93ccad7dba8bcee3a7e5d423085fd2f83088c015eebc436165.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 198, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 389 + ], + "type": "text", + "content": "bining image information to assist text-based retrieval, with main tasks such as WebQA [5] in the " + }, + { + "bbox": [ + 313, + 198, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mathrm{T}\\rightarrow \\mathrm{I} + \\mathrm{T}" + }, + { + "bbox": [ + 313, + 198, + 555, + 389 + ], + "type": "text", + "content": " format and ReMuQ [34] in the " + }, + { + "bbox": [ + 313, + 198, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mathrm{I} + \\mathrm{T}\\rightarrow \\mathrm{T}" + }, + { + "bbox": [ + 313, + 198, + 555, + 389 + ], + "type": "text", + "content": " format. As shown in Table 5, our performance on composed image retrieval surpasses that of Pic2Word [46], a model trained with specifically designed pseudo-word tasks, exceeding it by 7.3 points on the FashionIQ validation set and by 4.3 points on the Val-All task, which utilizes a larger corpus. Additionally, most of our trained composed modality retrieval data were filtered from Vista's training data, yet our composed image retrieval performance is 12.7 points higher than Vista's [67]. This indicates that better multimodal alignment capabilities can also enhance composed retrieval performance. In comparison with E5-V [17], our DeKR outperforms E5-V's 8B-sized model with a smaller 7B model, achieving SoTA performance on composed image retrieval tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 390, + 556, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 556, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 556, + 521 + ], + "type": "text", + "content": "As shown in Table 6, our DeKR outperforms Vista and other zero-shot proprietary models [18, 35] on two composed text retrieval tasks, achieving SOTA performance under zero-shot settings. It is noteworthy that since Vista does not finetune the text encoder component, it can be regarded as BGEm3 [7] in text retrieval tasks, serving as a strong baseline for text retrieval. We found that DeKR exceeds BGEm3 by 11.1 points in the WebQA text retrieval task, indicating that DeKR not only possesses unified multimodal representation capabilities but also exhibits excellent text representation abilities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 521, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 629 + ], + "type": "text", + "content": "Multimodal ranking. Multimodal ranking tasks typically refer to image-text matching tasks, where both image and text are inputted to produce a matching score based on their correlation. Often, due to differences in input between representation tasks and ranking tasks, separate models are used for representation and ranking. The DeKR model, however, leverages MLLMs to handle combined multimodal inputs, allowing it to perform both representation and ranking tasks within a unified architecture." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "To verify whether DeKR's representation capabilities can positively influence ranking tasks, as shown in Table 7, we compared the ranking performance of three approaches: directly using prompt engineering with Qwen2VL, training ranking data based on Qwen2VL, and training ranking data based on DeKR. The ranking prompt for Qwen2VL was: {\"Image}{Text}Do they match? A.Yes B.No\". The match-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "type": "text", + "content": "InternVideo2 is not discussed here as it is trained for audio alignment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29739" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 87, + 299, + 213 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 280, + 82 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 280, + 82 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 280, + 82 + ], + "type": "text", + "content": "Table 8. Comparison on Fine-tuned Multimodal Ranking." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 87, + 299, + 213 + ], + "lines": [ + { + "bbox": [ + 56, + 87, + 299, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 299, + 213 + ], + "type": "table", + "html": "
Flickr I2TFlickr T2I
R1R5R10R1R5R10
Uniter87.398.099.275.694.196.8
AlBEAF95.999.8100.085.697.598.9
mPlug297.2100.0100.088.197.699.1
BLIP97.499.899.987.697.799.0
BLIP2-L96.9100.0100.088.697.698.9
BLIP2-g97.6100.0100.089.798.199.0
Qwen2VL95.199.499.870.292.096.5
DeKR98.4100.0100.090.198.299.2
", + "image_path": "4d94ff8dd3f5405cd30324a7035f3c98f9e499ef9e0979c9466aed513e48661b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 220, + 295, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 295, + 316 + ], + "type": "text", + "content": "ing score was obtained by normalizing the probabilities of options A and B from the output vocabulary. For the zero-shot setting ranking training data, we randomly selected 500k image-text pairs from DeKon5M and concatenated them into a unified input format as positive samples. Additionally, we selected one negative sample for each modality and followed the multi-range difficulty selection strategy used in mGTE [66]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 424 + ], + "type": "text", + "content": "Surprisingly, we found that Qwen2VL-PE, without any data training, already possessed ranking capabilities, indicating that MLLMs inherently have strong image-text alignment abilities. Furthermore, when comparing the two models trained on ranking data—Qwen2VL+Rank and DeKR+Rank—we observed that the latter outperformed the former by 14.6 points on the Text-to-Image task. This demonstrates that DeKR's representation capabilities have a crucial impact on ranking tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 425, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 545 + ], + "type": "text", + "content": "Additionally, as shown in Table 8, we report the performance of DeKR after fine-tuning the downstream ranking dataset Flickr. We compared it with previous ranking models and found that DeKR achieved SoTA results on both Image-to-Text and Text-to-Image ranking. This indicates that DeKR successfully unifies the architecture for both representation and ranking tasks. Furthermore, when we finetuned Qwen2VL on the Flickr data, we observed a performance gap of up to 20 points on the T2I task, further emphasizing the necessity of DeKR's representation training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 173, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 173, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 173, + 565 + ], + "type": "text", + "content": "4.4. Qualitative Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": "Cross-modal retrieval. In Figure 8(a), we compare DeKR and SigLip on Top-1 recall for short and long text cross-modal retrieval. Both models succeed with short texts, with DeKR preferring detailed descriptions and SigLip favoring broader summaries. For long texts, SigLip retrieves wrong answers when accurate on objects but misses color and position details, suggesting CLIP models focus on object alignment but neglect action and descriptive nuances." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Composed modality retrieval. In Figure 8(b), we explore DeKR's performance on more complex multimodal retrieval tasks. The image editing task builds upon the image retrieval task. As shown at the top of the figure, DeKR" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 560, + 319 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 560, + 319 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 560, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 560, + 319 + ], + "type": "image", + "image_path": "7501d5b1152066c09772a22cf1400413ae31415f2661a6d38db823765e5574ff.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 319, + 555, + 330 + ], + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 330 + ], + "type": "text", + "content": "Figure 8. Case analysis of cross and composed modality retrieval." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 332, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 548 + ], + "type": "text", + "content": "first interprets the attributes of the two gorillas in the left image, such as size, posture, and orientation. It then processes the text instruction to only reduce the gorillas' size while maintaining other attributes, successfully retrieving the correct result depicted in the right image. This demonstrates that fine-grained alignment remains crucial in complex retrieval tasks. Additionally, we explored the VQA retrieval task, where DeKR must perform VQA reasoning internally and represent the outcome to retrieve the correct answer, indicating that DeKR retains the original MLLM's reasoning capabilities well. Finally, we assessed DeKR's ability to recall multimodal information and found that it could recognize and retrieve data with the text \"I AM A MAN\" from the image, showcasing its OCR capabilities, which are also inherent in the original MLLM. These cases suggest that if a retrieval model can effectively inherit the original capabilities of MLLM, there is significant potential for retrieval tasks, paving the way for further exploration of DeKR." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 561, + 391, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 391, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 391, + 573 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 580, + 555, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 702 + ], + "type": "text", + "content": "In this paper, we introduce DeKR, a unified multimodal retrieval model trained with MLLMs. DeKR integrates text, images, videos, and composed image-text into a unified embedding space and consolidates representation and ranking tasks within one framework. To enhance multimodal alignment, we develop DeKon5M, a contrastive learning dataset rich in dense knowledge. Extensive experiments demonstrate that DeKR achieves SoTA performance in cross-modal retrieval, composed image-text retrieval, and ranking tasks. This shows DeKR's advancement, and DeKon5M contributes to future academic research." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29740" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "text", + "content": "[2] Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. Llm2vec: Large language models are secretly powerful text encoders. CoRR, abs/2404.05961, 2024. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 248 + ], + "type": "text", + "content": "[3] Rumeysa Bodur, Erhan Gundogdu, Binod Bhattarai, Tae-Kyun Kim, Michael Donoser, and Loris Bazzani. iedit: Localised text-guided image editing with weak supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024 - Workshops, Seattle, WA, USA, June 17-18, 2024, pages 7426-7435. IEEE, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 251, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 294 + ], + "type": "text", + "content": "[4] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.4.5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 296, + 295, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 295, + 361 + ], + "type": "text", + "content": "[5] Yingshan Chang, Guihong Cao, Mridu Narang, Jianfeng Gao, Hisami Suzuki, and Yonatan Bisk. Webqa: Multihop and multimodal QA. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16474-16483. IEEE, 2022. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 364, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 295, + 418 + ], + "type": "text", + "content": "[6] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for A lite vision-language model. CoRR, abs/2402.11684, 2024. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 464 + ], + "type": "text", + "content": "[7] Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. Bge m3-embedding: Multilingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation, 2024. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 466, + 295, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 295, + 509 + ], + "type": "text", + "content": "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 512, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 295, + 578 + ], + "type": "text", + "content": "[9] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 580, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 644 + ], + "type": "text", + "content": "[10] Yuhao Cui, Zhou Yu, Chunqi Wang, Zhongzhou Zhao, Ji Zhang, Meng Wang, and Jun Yu. Rosita: Enhancing vision-and-language semantic alignments via cross- and intra-modal knowledge integration. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "text", + "content": "[12] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan C. Russell. Localizing moments in video with natural language. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 5804-5813. IEEE Computer Society, 2017. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 141, + 554, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 554, + 195 + ], + "type": "text", + "content": "[13] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13405-13417, 2024. 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "text", + "content": "[14] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 243, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 307 + ], + "type": "text", + "content": "[15] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 1, 2, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 309, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 385 + ], + "type": "text", + "content": "[16] Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. Mistral 7b. CoRR, abs/2310.06825, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "type": "text", + "content": "[17] Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-V: universal embeddings with multimodal large language models. CoRR, abs/2407.12580, 2024. 1, 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 445, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 520 + ], + "type": "text", + "content": "[18] Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wentau Yih. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781, Online, 2020. Association for Computational Linguistics. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 522, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 522, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 553, + 577 + ], + "type": "text", + "content": "[19] Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin Kim, Minkyung Cho, Jy yong Sohn, and Chanyeol Choi. Linq-embed-mistral: elevating text retrieval with improved gpt data through task-specific control and quality refinement. Linq AI Research Blog, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 579, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 622 + ], + "type": "text", + "content": "[20] Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nv-embed: Improved techniques for training llms as generalist embedding models. CoRR, abs/2405.17428, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "type": "text", + "content": "[21] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. CoRR, abs/2408.03326, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[22] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29741" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 294, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 294, + 182 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 19730-19742. PMLR, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "type": "text", + "content": "[25] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. ECCV 2020, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "type": "text", + "content": "[26] Xianhang Li, Zeyu Wang, and Cihang Xie. An inverse scaling law for CLIP training. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "text", + "content": "[27] Xianhang Li, Zeyu Wang, and Cihang Xie. Clipa-v2: Scaling CLIP training with " + }, + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "inline_equation", + "content": "81.1\\%" + }, + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "text", + "content": " zero-shot imagenet accuracy within a $10,000 budget; an extra $4,000 unlocks 81.8% accuracy. CoRR, abs/2306.15658, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[28] Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281, 2023. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 458 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V, pages 740-755. Springer, 2014. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 459, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 294, + 481 + ], + "type": "text", + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 569 + ], + "type": "text", + "content": "[32] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre-trained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2125–2134, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[33] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 1, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 689 + ], + "type": "text", + "content": "[34] Man Luo, Zhiyuan Fang, Tejas Gokhale, Yezhou Yang, and Chitta Baral. End-to-end knowledge retrieval with multimodal queries. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 8573-8589. Association for Computational Linguistics, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "type": "text", + "content": "[35] Man Luo, Zhiyuan Fang, Tejas Gokhale, Yezhou Yang, and Chitta Baral. End-to-end knowledge retrieval with multi-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 127 + ], + "type": "text", + "content": "modal queries. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 8573-8589. Association for Computational Linguistics, 2023. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[36] Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. Sfr-embedding-mistral:enhance text retrieval with transfer learning. Salesforce AI Research Blog, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[37] Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. Sfr-embedding-2: Advanced text embedding with multi-stage training, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 230 + ], + "type": "text", + "content": "[38] Niklas Muennighoff. Sgpt: Gpt sentence embeddings for semantic search. arXiv preprint arXiv:2202.08904, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 232, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 297 + ], + "type": "text", + "content": "[39] Niklas Muennighoff, Nouamane Tazi, Loïc Magne, and Nils Reimers. MTEB: massive text embedding benchmark. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2023, Dubrovnik, Croatia, May 2-6, 2023, pages 2006-2029. Association for Computational Linguistics, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 553, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 332 + ], + "type": "text", + "content": "[40] Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. Generative representational instruction tuning, 2024. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 333, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 431 + ], + "type": "text", + "content": "[41] Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas Tezak, Jong Wook Kim, Chris Hallacy, Johannes Heidecke, Pranav Shyam, Boris Power, Tyna Eloundou Nekoul, Girish Sastry, Gretchen Krueger, David Schnurr, Felipe Petroski Such, Kenny Hsu, Madeleine Thompson, Tabarak Khan, Toki Sherbakov, Joanne Jang, Peter Welinder, and Lilian Weng. Text and code embeddings by contrastive pretraining. CoRR, abs/2201.10005, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 433, + 498, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 498, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 498, + 444 + ], + "type": "text", + "content": "[42] OpenAI. Gpt-4v(ision) system card. 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 445, + 475, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 475, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 475, + 456 + ], + "type": "text", + "content": "[43] OpenAI. Gpt-4o system card. 2024. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 457, + 553, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 457, + 553, + 500 + ], + "spans": [ + { + "bbox": [ + 316, + 457, + 553, + 500 + ], + "type": "text", + "content": "[44] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2011. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 502, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 553, + 590 + ], + "type": "text", + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 1, 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 591, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 657 + ], + "type": "text", + "content": "[46] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 19305-19314. IEEE, 2023. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[47] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: open dataset of clip-filtered 400 million image-text pairs. CoRR, abs/2111.02114, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29742" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 172 + ], + "type": "text", + "content": "[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 173, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 296, + 206 + ], + "type": "text", + "content": "[49] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 1, 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 207, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 296, + 252 + ], + "type": "text", + "content": "[50] Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. EVA-CLIP-18B: scaling CLIP to 18 billion parameters. CoRR, abs/2402.04252, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 253, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 319 + ], + "type": "text", + "content": "[51] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024. 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 320, + 296, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 398 + ], + "type": "text", + "content": "[52] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 11897-11916. Association for Computational Linguistics, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 399, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 296, + 465 + ], + "type": "text", + "content": "[53] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 466, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 296, + 554 + ], + "type": "text", + "content": "[54] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 1, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 555, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 296, + 601 + ], + "type": "text", + "content": "[55] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The fashion iq dataset: Retrieving images by combining side information and relative natural language feedback. CVPR, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 601, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 296, + 689 + ], + "type": "text", + "content": "[56] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), Online, 2021. Association for Computational Linguistics. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 690, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 296, + 715 + ], + "type": "text", + "content": "[57] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang," + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 697 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 116 + ], + "type": "text", + "content": "Guohai Xu, Ji Zhang, Songfang Huang, Fei Huang, and Jingren Zhou. mplug-2: A modularized multi-modal foundation model across text, image and video. ArXiv, abs/2302.00402, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 118, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 183 + ], + "type": "text", + "content": "[58] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 5288-5296. IEEE Computer Society, 2016. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 184, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 554, + 228 + ], + "type": "text", + "content": "[59] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2:67-78, 2014. 4, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 229, + 555, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 555, + 328 + ], + "type": "text", + "content": "[60] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 3208-3216. AAAI Press, 2021. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 329, + 554, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 329, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 329, + 554, + 384 + ], + "type": "text", + "content": "[61] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 11941-11952. IEEE, 2023. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 384, + 554, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 554, + 429 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 554, + 429 + ], + "type": "text", + "content": "[62] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 430, + 554, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 554, + 474 + ], + "type": "text", + "content": "[63] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 475, + 554, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 554, + 507 + ], + "type": "text", + "content": "[64] Peitian Zhang, Shitao Xiao, Zheng Liu, Zhicheng Dou, and Jian-Yun Nie. Retrieve anything to augment large language models. CoRR, abs/2310.07554, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 508, + 554, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 508, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 508, + 554, + 552 + ], + "type": "text", + "content": "[65] Xin Zhang, Zehan Li, Yanzhao Zhang, Dingkun Long, Pengjun Xie, Meishan Zhang, and Min Zhang. Language models are universal embeddeders. CoRR, abs/2310.08232, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 552, + 554, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 552, + 554, + 619 + ], + "spans": [ + { + "bbox": [ + 316, + 552, + 554, + 619 + ], + "type": "text", + "content": "[66] Xin Zhang, Yanzhao Zhang, Dingkun Long, Wen Xie, Ziqi Dai, Jialong Tang, Huan Lin, Baosong Yang, Pengjun Xie, Fei Huang, Meishan Zhang, Wenjie Li, and Min Zhang. mgte: Generalized long-context text representation and reranking models for multilingual text retrieval. CoRR, abs/2407.19669, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 619, + 555, + 697 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 619, + 555, + 697 + ], + "spans": [ + { + "bbox": [ + 316, + 619, + 555, + 697 + ], + "type": "text", + "content": "[67] Junjie Zhou, Zheng Liu, Shitao Xiao, Bo Zhao, and Yongping Xiong. VISTA: visualized text embedding for universal multi-modal retrieval. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 3185-3200. Association for Computational Linguistics, 2024. 1, 2, 5, 7" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29743" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Incremental Object Keypoint Learning/d3b12539-b476-45cb-a752-bae4990c3622_content_list.json b/2025/Incremental Object Keypoint Learning/d3b12539-b476-45cb-a752-bae4990c3622_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a98a17b38405484c03368e658af253faf402a200 --- /dev/null +++ b/2025/Incremental Object Keypoint Learning/d3b12539-b476-45cb-a752-bae4990c3622_content_list.json @@ -0,0 +1,1369 @@ +[ + { + "type": "text", + "text": "Incremental Object Keypoint Learning", + "text_level": 1, + "bbox": [ + 300, + 130, + 697, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingfu Liang $^{1}$ Jiahuan Zhou $^{2,*}$ Xu Zou $^{3}$ Ying Wu $^{1}$", + "bbox": [ + 264, + 179, + 728, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Northwestern University $^{2}$ Wangxuan Institute of Computer Technology, Peking University $^{3}$ Huazhong University of Science and Technology", + "bbox": [ + 122, + 198, + 872, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing progress in object keypoint estimation primarily benefits from the conventional supervised learning paradigm based on numerous data labeled with pre-defined keypoints. However, these well-trained models can hardly detect the undefined new keypoints in test time, which largely hinders their feasibility for diverse downstream tasks. To handle this, various solutions are explored but still suffer from either limited generalizability or transferability. Therefore, in this paper, we explore a novel keypoint learning paradigm in that we only annotate new keypoints in the new data and incrementally train the model, without retaining any old data, called Incremental object Keypoint Learning (IKL). A two-stage learning scheme as a novel baseline tailored to IKL is developed. In the first Knowledge Association stage, given the data labeled with only new keypoints, an auxiliary KA-Net is trained to automatically associate the old keypoints to these new ones based on their spatial and intrinsic anatomical relations. In the second Mutual Promotion stage, based on a keypoint-oriented spatial distillation loss, we jointly leverage the auxiliary KA-Net and the old model for knowledge consolidation to mutually promote the estimation of all old and new keypoints. Owing to the investigation of the correlations between new and old keypoints, our proposed method can not just effectively mitigate the catastrophic forgetting of old keypoints, but may even further improve the estimation of the old ones and achieve a positive transfer beyond antiforgetting. Such an observation has been solidly verified by extensive experiments on different keypoint datasets, where our method exhibits superiority in alleviating the forgetting issue and boosting performance while enjoying labeling efficiency even under the low-shot data regime.", + "bbox": [ + 88, + 301, + 485, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 816, + 220, + 832 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a fundamental task in computer vision, estimating the visual keypoint locations of an object serves as the indis-", + "bbox": [ + 89, + 843, + 482, + 873 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59d36260c6f937c2d3c3a51708fd622dfb077665f6fefc60bc8f6272a6cc47ed.jpg", + "image_caption": [ + "Figure 1. Medical analysis can frequently change and require new keypoints essentially [6], while the labeling is time-consuming." + ], + "image_footnote": [], + "bbox": [ + 558, + 267, + 854, + 385 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pensable primitive to support numerous down-stream tasks, e.g., object pose detection [29, 77, 80], tracking [28], action recognition [7], generation [57] and animation [28, 58], etc. Over a long period of time, supervised keypoint learning (SKL) [19, 36, 39, 68] has significantly advanced the progress on keypoint estimation. By exploiting large-scale pure 2D keypoint datasets [5, 44] with extreme variance, SKL aims to train a deep neural network that can robustly detect a set of pre-defined keypoints in test images.", + "bbox": [ + 511, + 429, + 906, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the SKL models can not estimate newly-added undefined keypoints of an object, while new demands from downstream tasks will inevitably require new keypoints. For instance, until now, models trained by existing keypoint dataset [67] can only detect 19 keypoints to support Downs analysis in the Cephalometric analysis, while they are infeasible for important analysis like Steiner analysis [53], as shown in Fig. 1. Such an issue motivated a recent MIC-CAI challenge [6] on annotating much more keypoints to support increasing analysis. However, keypoint labeling is time-consuming [18, 45, 50], especially when keypoints are defined sequentially. Another alternative solution is to train separate new keypoint estimator [12, 31] for each set of new keypoints as in Fig. 2. However, the number of estimators will increase linearly, and the separate training strategy will make each estimator easily overfit the new keypoints [12] and hardly capture the intrinsic relation between old and new keypoints, leading to sub-optimal performance.", + "bbox": [ + 511, + 566, + 908, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To mitigate the above issues, unsupervised keypoint learning (UKL) [22, 23, 49, 55, 84] is proposed for new keypoint estimation. The models are firstly pretrained in an unsupervised manner to predict numbers of keypoints ran", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*: Corresponding author", + "bbox": [ + 114, + 887, + 246, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25399", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e57ab3b3387667f941eb381fbf19203dfa588a21d2a90fe22206936958567688.jpg", + "image_caption": [ + "Figure 2. Separate Estimators need multiple estimators in test time. Unsupervised keypoint learning (UKL) and category-agnostic pose estimation (CAPE) exploit a fixed pretrained keypoint estimator on unseen new keypoints. While our Incremental object Keypoint Keaning (IKL) continually updates the same model on the new data labeled only on new keypoints without retaining any old training data." + ], + "image_footnote": [], + "bbox": [ + 89, + 87, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "domly, then a linear transformation is learned to transfer the existing keypoints to the new undefined ones. However, UKL is merely effective on objects with relatively rigid motion or video data with consistent object instances and local changes in appearance and motion. As a result, it is hard to perform UKL on large-scale pure 2D datasets [5, 44] due to their extreme inter-image differences in the background, appearance, and motions of distinct object instances. Recently, a supervised metric learning-based [59] paradigm named category-agnostic pose estimation (CAPE) [76] is proposed to learn a class-agnostic pose estimator that can estimate unseen keypoints in test time with a few labeled support samples. However, as the model is frozen after pretraining, CAPE is limited by its restricted generalization and transferability for newly-defined keypoints (Tab. 4).", + "bbox": [ + 88, + 345, + 485, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, in this paper, we propose to solve the new keypoint estimation problem creatively in an incremental learning (IL) manner. As shown in Fig. 2, we propose an Incremental object Keypoint Learning (IKL) paradigm to incrementally train a keypoint estimator on the new data labeled only with new keypoints. Even without retaining any old data, the IKL model should not forget the old keypoints catastrophically. Compared to the naive solutions that either label both new and old keypoints or learn separate estimators, our IKL only labels the new keypoints and maintains just one model in the lifetime, which is computationally efficient. Moreover, different from the UKL and CAPE whose performances are largely limited by their restricted transferability depending on a fixed pretrained model, the IKL alleviates such an issue by continually training the model to sufficiently expand its knowledge on newly-defined keypoints and may even improve the old keypoint estimation when learning more relevant new keypoints. Lastly, as most works in IL mainly study the classification task [13, 70], to our best knowledge, IKL as an incremental keypoint regression task has rarely been stud", + "bbox": [ + 89, + 583, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ied before, which also provides novel insights for IL.", + "bbox": [ + 511, + 345, + 862, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, IKL also poses a new challenge known as label non-co-occurrence (LNCO). As only new keypoints are labeled in the new data, the kinematic and anatomical constraints between old and new keypoints are not explicitly presented in the label space, making it hard for the model to capture such a physical prior during incremental training. Existing IL methods (Tab. 1) do not explicitly model such inter-keypoint relations in IKL, thus the old and new keypoint predictions can not mutually support each other and may readily bias toward the new ones, exacerbating the forgetting of the old keypoints (Tab. 1).", + "bbox": [ + 511, + 366, + 906, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the challenge of IKL, we propose a novel two-stage learning scheme called KAMP as a new baseline tailored to IKL. In the first Knowledge Association stage, we train an auxiliary KA-Net to associate old and new keypoints based on their physiological connection, represented by their spatial adjacency. Specifically, KA-Net learns to predict the selected old keypoint given the related new keypoints to acquire their intrinsic anatomical relevance. In the second Mutual Promotion stage, we train the new model on all keypoints to mutually promote their estimation, where the old ones are updated by distilling from the KA-Net and the old model with a newly designed keypoint-oriented spatial distillation loss for better knowledge preservation. To verify our effectiveness, we simulate IKL based on four keypoints datasets on both medical and natural images, i.e., Cephalometric [6], Chest [27], MPII [5] and ATRW [37]. Extensive results demonstrate that our KAMP can not just effectively alleviate the catastrophic forgetting of old keypoints, but even further boost their performance and thus outperform the existing exemplar-free IL methods by a significant margin (Tab. 1) and can also work for low-shot scenarios (Tab. 3 and 4). Our analysis further reveals that IKL is complementary to learning paradigms like CAPE (Tab. 4) and labeling-efficient for new keypoint estimation.", + "bbox": [ + 511, + 537, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25400", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are three-fold: (1) We establish a novel paradigm, Incremental object Keypoint Learning (IKL), to tackle the challenging demands of new keypoints estimation. (2) We propose a two-stage exemplar-free IKL method to explicitly model and exploit the relation between old and new keypoints to help alleviate the label non-co-occurrence problem tailored to IKL, which may further boost the performance of old keypoints beyond antiforgetting. (3) As a proof-of-concept, we empirically show that IKL is practical and labeling-efficient to scale up a pretrained keypoint estimator on new keypoints estimation, which is much better than other alternative paradigms.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 286, + 238, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Keypoint Estimation (KE). Existing research of KE focuses on estimating a fixed number of keypoints for a specific category [5, 29, 44, 54, 78-80]. Most works design new methods [20, 34, 35, 38, 89] or network architectures [39, 48, 61, 62, 68, 77] to improve the supervised learning over large-scale pure 2D keypoint datasets [5, 44]. Estimating dense keypoints [29, 82] of the human body has been proposed but requires extreme labeling cost. Recent works explore semi-supervised [24, 50, 66] or unsupervised learning [63, 84] to reduce the laborious labeling consumption, but the unsupervised methods still hardly achieve a precise estimation of semantically meaningful keypoints [49]. Recently, Category-agnostic keypoint estimation (CAPE) [10, 56, 76] is proposed to learn a class-agnostic model to detect keypoints of unseen categories by a few labeled support images without retraining the model. However, as the model is frozen after pretraining, CAPE is limited by restricted generalization and transferability due to the intra-class appearance variation, self-occlusion, and appearance ambiguity, as detailed in [76]. Different from the above paradigm that pre-defined a set of keypoints or assumed the pretrained model was sufficient enough for different downstream tasks, our proposed IKL paradigm continuously updates the model only on the new keypoints to increasingly expand its knowledge, which is much more flexible and labeling-efficient.", + "bbox": [ + 91, + 311, + 483, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Incremental Learning (IL). Existing IL methods can be categorized as exemplar-free and exemplar-based ones [3, 25, 26, 32, 41, 52], given whether old data can be retained. However, saving exemplars not just introduces increasing costs on memory, but also has privacy issues in real-world applications. For exemplar-free methods, existing approaches focus on adding regularization over parameters [2, 33, 47, 81, 83] or the network's output in different positions [15, 40, 73, 85]. Regarding the IL settings, adding new classes [13] or domains [41, 46, 65] have been studied popularly and mainly tested on classification benchmarks [13, 14]. In contrast, keypoint estimation is a location regression task [20, 34, 48], and our proposed IKL aims", + "bbox": [ + 89, + 705, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to detect newly defined keypoints incrementally. The incremental animal pose estimation [51] (IAPE) is a related setting to us. However, IAPE assumes the label space of keypoints is fixed for each animal, and it only learns new animals' poses, i.e., only the input distribution is changed with new animal domains. While for our IKL, we instead consider adding new keypoints to the label space for an object category, and the old and new data distribution may be relatively correlated.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Positive Transfer (PT) in IL. Most IL approaches focus on alleviating the catastrophic forgetting problem [13], while only a few works are dedicated to achieving positive transfer (PT) in IL [30, 43, 47], i.e., improving the old tasks while learning a new task. Those methods encourage PT by modifying the gradient updates based on the new and old tasks' correlations which are measured by saving the data [47] from the old task or based on formal analyses [30, 43]. Differently, our KAMP achieves PT by explicitly modeling the relation between old and new keypoints based on their intrinsic anatomical relevance, and we exploits the relation to design novel distillation to further boost the estimation of the old keypoint, which is tailored to the IKL and keypoint estimation.", + "bbox": [ + 511, + 227, + 903, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Incremental Object Keypoint Learning", + "text_level": 1, + "bbox": [ + 511, + 452, + 864, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 477, + 718, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We denote $t = 0$ as the initial step and $t = 1, \\dots, T$ as the incremental steps. Then the training set for the $t$ -th incremental step is $D_{t}^{\\mathrm{train}} = \\left\\{x_{t,j}^{\\mathrm{train}}, y_{t,j}^{\\mathrm{train}}\\right\\}_{j=1}^{N_{t}}$ , where $x, y, N_{t}$ denote the inputs, the keypoint labels and the size of $D_{t}^{\\mathrm{train}}$ . As we only label new keypoints in IKL, thus the newly introduced keypoints must be disjointed with the old ones, i.e., $y_{t}^{\\mathrm{train}} \\cap \\left(\\cup_{i=0}^{t-1} y_{i}^{\\mathrm{train}}\\right) = \\emptyset$ . Different from the classification task, the common practice for keypoint estimation is to regress a 2D heatmap for a keypoint, i.e., a Gaussian peak around the ground-truth keypoint location [71]. Thus the model $m_{t}$ in IKL comprises the feature extractor $f_{t}$ and a stack of convolutional keypoint estimation heads $\\{G_{i}\\}_{i=1}^{t}$ that parameterized by $\\theta_{t}$ and $\\{\\phi_{i}\\}_{i=1}^{t}$ , respectively. Each estimation head ends with a convolutional layer to generate the 2D heatmap. $G_{t}$ denotes the estimation head for the new keypoints in the $t$ -th step. $f_{t}$ and $\\{G_{i}\\}_{i=1}^{t-1}$ in model $m_{t}$ are initialized by the same weights in the old model $m_{t-1}$ .", + "bbox": [ + 511, + 500, + 906, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. KAMP: A Novel Baseline for IKL", + "text_level": 1, + "bbox": [ + 511, + 772, + 808, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our two-stage method KAMP as a novel baseline tailored to IKL. We conjecture that distilling the relation between related old and new keypoints into the model may help it implicitly capture the intrinsic physical prior to alleviate the label non-co-occurrence issue. To do so, for each incremental step, in Stage-I, we design an auxiliary prediction task to associate the related", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25401", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f7f48df20e868dfced7c09b3499eecef0acd55c0be93d4ff6dcbcfa97def3a39.jpg", + "image_caption": [ + "Figure 3. Overview of KAMP using the human body for illustration. In Stage-I, we learn an auxiliary KA-Net to associate the related old and new keypoints based on their local anatomical constraint. In Stage-II, we jointly leverage the old model and the KA-Net as an auxiliary teacher to consolidate all old keypoints' prediction and also learn the new keypoints simultaneously to achieve mutual promotions." + ], + "image_footnote": [], + "bbox": [ + 153, + 87, + 614, + 272 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a549ac296f9aa6e810114de445b96bbc40672e9f07549f4fec5dac317a0c04b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 88, + 843, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "old keypoint's with the new ones by the KA-Net. In Stage-II, the old model $m_{t - 1}$ and KA-Net are both frozen to distill knowledge into the new model $m_t$ for all old keypoints, and the new keypoints are learned concurrently to achieve mutual promotion. Fig. 3 provides a schematic view of KAMP.", + "bbox": [ + 88, + 334, + 480, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Stage-I: Knowledge Association (KA)", + "text_level": 1, + "bbox": [ + 89, + 416, + 397, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned in Sec. 1, the old and new keypoints in IKL are not jointly labeled, making it difficult for the model to learn relationships between keypoints, such as their structural and anatomical relevance, using only the labels of new keypoints. To mitigate this issue, we propose to model the constraint among the spatially and anatomically related old and new keypoints as an implicit function in the current incremental step $t$ . Existing analyses [11, 12, 64] show that keypoints association can be well-modeled by the triangulation constraint. Motivated by this, we define a triangulation constraint $F$ on the distribution of three related keypoints:", + "bbox": [ + 88, + 434, + 482, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF \\left(P \\left(K _ {i}\\right), P \\left(K _ {j}\\right), P \\left(K _ {k}\\right)\\right) = 0, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 603, + 480, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P(K_{i})$ denotes the distribution of the keypoint $i$ , and the tuple may include one or two new keypoints. However, as there may be multiple valid solutions (constraints) that satisfy the implicit function $F$ , it is challenging to derive an analytical solution for $F$ in IKL.", + "bbox": [ + 89, + 628, + 482, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we consider an alternative formulation, e.g., $P(K_{\\mathrm{o}_j}) = F(P(K_{\\mathrm{n}_1}),P(K_{\\mathrm{n}_2}))$ , which is a special case of the Eqn. 1 that we condition the distribution of an old keypoint $P(K_{\\mathrm{o}_j})$ on its related new keypoints $P(K_{\\mathrm{n}_1})$ and $P(K_{\\mathrm{n}_2})$ . We focus on two new keypoints since only they have ground-truth labels, and the ground-truth supervision can help reduce the uncertainty in learning the constraints. Moreover, in practice, new tasks typically require more than one new keypoint for support. Nonetheless, we verify in our Supplementary that our approach performs effectively even with a single new keypoint. To enhance this constraint, we model the conditional probability between the related old and new keypoints as an auxiliary prediction task:", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(K _ {\\mathrm {o} j} \\mid K _ {\\mathrm {n} 1}, K _ {\\mathrm {n} 2}, v\\right) = \\mathrm {K A} \\left(P \\left(K _ {\\mathrm {n} 1}\\right), P \\left(K _ {\\mathrm {n} 2}\\right), v\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 333, + 903, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where KA stands for the Knowledge Association Network (KA-Net). The insight of this auxiliary task is inspired by the fact that spatially adjacent and visually correlated keypoints can mutually predict each other as shown in many existing works [11, 62, 64].", + "bbox": [ + 511, + 354, + 903, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "KA-Net. The input of the KA-Net is the ground-truth heatmap $(y_{i,\\mathrm{n}_1}^{gt}$ and $y_{i,\\mathrm{n}_2}^{gt})$ of new keypoints $K_{\\mathrm{n_1}}$ and $K_{\\mathrm{n_2}}$ of an image $i$ , and the output is the predicted heatmap $\\tilde{y}_{i,j}$ of the old keypoint $K_{\\mathrm{o_j}}$ . $v_{i}$ denotes the holistic visual features $v$ of the image $i$ comprised of the intermediate features extracted from the frozen feature $f_{t - 1}$ and are re-scaled to the same spatial size as $y_{i,\\mathrm{n_1}}^{gt}$ and $y_{i,\\mathrm{n_2}}^{gt}$ . The visual features $v_{i}$ are incorporated alongside spatial information ( $y_{i,\\mathrm{n_1}}^{gt}$ and $y_{i,\\mathrm{n_2}}^{gt}$ ) because, in a 2D image, spatial coordinates alone are insufficient to determine a keypoint's location due to uncertainty stemming from factors such as camera angle, object appearance, and motion. By incorporating visual features, we provide additional contextual information that helps account for this uncertainty in keypoint associations.", + "bbox": [ + 511, + 431, + 906, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the present paper, we explore KA-Net's simplest construction to minimize its training cost: for each new keypoint, we first perform the element-wise multiplication between its ground-truth heatmap $(y_{i,\\mathrm{n}_1}^{gt}$ and $y_{i,\\mathrm{n}_2}^{gt})$ and the holistic features $v_{i}$ to obtain the keypoint-oriented spatial features. Then we concatenate the spatial features and feed them forward over three convolutional layers accompanied by the Batch Normalization (BN) and ReLU to predict the selected old keypoint $K_{\\mathrm{o}_j}$ . Note that the trained KA-Net will only be used to distill the keypoint association in the Stage-II and will not be used in test time after IKL.", + "bbox": [ + 511, + 648, + 906, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training of KA-Net. We train KA-Net by the new data $D_{t}^{\\mathrm{train}}$ . As $D_{t}^{\\mathrm{train}}$ does not have ground truth labels for old keypoints, we use the pseudo-labels predicted by the old model $m_{t - 1}$ to supervise KA-Net for keypoint regression:", + "bbox": [ + 511, + 814, + 906, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {K A} = \\frac {1}{N _ {t}} \\sum_ {i} ^ {N t} \\sum_ {j \\in \\mathcal {K} _ {K A}} \\left\\| \\hat {y} _ {i, j} ^ {t - 1} - \\tilde {y} _ {i, j} \\right\\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 875, + 903, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25402", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{K}_{KA}$ denotes the selected old keypoints used for the auxiliary prediction task, $\\hat{y}_{i,j}^{t-1}$ denotes the prediction of the $j$ -th keypoint by $m_{t-1}$ given image $i$ , and $\\tilde{y}_{i,j}$ denotes the prediction by the KA-Net. We will show in Sec. 4 that our KAMP can effectively reduce the forgetting of old keypoints and even improve them, which largely reduce the accumulative error to use pseudo-labels for KA-Net.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Auxiliary Task Creation. To select the old keypoint for training KA-Net, in each incremental step, we first locate all the old and new keypoints on a general object anatomy diagram by their semantic definition. Then we measure their relative proximity by simple distance calculations (e.g., Euclidean distance) to identify two new keypoints $(K_{\\mathrm{n_1}}$ and $K_{\\mathrm{n_2}})$ that are close to an old one $K_{\\mathrm{o_j}}$ . This process can also be automated by multi-modality large language models like GPT-4o [1, 42, 75] as shown in Fig. 3. If we identify several tuples of old and new keypoints satisfied the requirement, we randomly choose one tuple to create the auxiliary prediction task to avoid training too many KA-Net. For all results of our KAMP in Section 4, we show that even by creating only one auxiliary task in each step, we can still bring sufficient improvement on all keypoints. Note that task creation does not need any training, and we only need to perform once before each step, which is highly efficient. Such a design provides an interpretable way to incorporate the physical knowledge of the object to guide the IKL.", + "bbox": [ + 91, + 198, + 486, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Stage-II: Mutual Promotion (MP)", + "text_level": 1, + "bbox": [ + 89, + 494, + 372, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Stage-II, we jointly optimize all the new and old keypoints on the new model $m_t$ by the loss $l_{MP}$ :", + "bbox": [ + 89, + 515, + 482, + 546 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {M P} = \\ell_ {G T} + \\alpha \\left(\\ell_ {K S D} + \\ell_ {K A}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 553, + 482, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\ell_{GT}$ denotes the L2 loss between the ground truth of the new keypoints and their predictions by the new model to acquire the knowledge of new keypoints. $\\alpha$ is a hyperparameter to balance the new keypoints acquisition and old keypoints forgetting. For $\\ell_{KA}$ , we use the frozen KA-Net as an auxiliary teacher to supervised the old keypoint selected in Stage-I to distill the keypoint association knowledge. Since $\\ell_{KA}$ is only applied to the selected old keypoint for knowledge transfer instead of mitigating the forgetting, we further consolidate the knowledge of all old keypoints by the loss $\\ell_{KSD}$ using the frozen old model $m_{t - 1}$ . As the distribution of new and old data in IKL do not change dramatically and may even have a strong correlation, LWF [40] can be a baseline for $\\ell_{KSD}$ in such a scenario as analyzed in [13]. In general, $\\ell_{KSD}$ represents the negative log-likelihood $\\frac{1}{N_t}\\sum_i^{N_t}s\\left(\\hat{y}_i^{t - 1}\\right)\\cdot \\log s\\left(\\hat{y}_i^t\\right)$ between the predictions of all old keypoints by the new model, $\\hat{y}_i^{t - 1}\\in \\mathbb{R}^{C\\times H\\times W}$ , and predictions by the old model, $\\hat{y}_i^t\\in \\mathbb{R}^{C\\times H\\times W}$ . $s(\\cdot)$ denotes the Softmax operator. $C$ denotes the numbers of old keypoints learned before step $t$ , $H$ and $W$ are the height and width of each keypoint's heatmap.", + "bbox": [ + 89, + 583, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, since LWF and its variants all focus on the classification task, they perform the Softmax across different classes by default, i.e., over the $C$ dimension, to obtain the normalized class prediction score. For IKL, this means normalizing across all old keypoints' predictions, which can not explicitly regularize the discrepancy of each old keypoint between the old and new model. To better preserve keypoint-specific knowledge, we adopt a spatial softmax operation over height $(H)$ and width $(W)$ dimensions of the keypoint prediction, $s_{\\mathrm{sp}}^{\\mathrm{d}}(\\cdot), \\mathrm{d} \\in \\{H, W\\}$ , and combine them to encourage spatial-oriented knowledge distillation. We discuss in our Supplementary the difference between each softmax alternative. Thus, our $\\ell_{KSD}$ is defined as", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{N _ {t}} \\sum_ {i} ^ {N t} \\sum_ {j} ^ {C} \\sum_ {\\mathrm {d}} - s _ {\\mathrm {s p}} ^ {\\mathrm {d}} (\\hat {y} _ {i, j} ^ {t - 1}) \\cdot \\log s _ {\\mathrm {s p}} ^ {\\mathrm {d}} (\\hat {y} _ {i, j} ^ {t}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 290, + 903, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 330, + 637, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. As the IKL has rarely been studied before, there is no specific dataset or benchmark for IKL. Motivated by the real-world application of IKL in medical analysis, we leverage the large-scale Cephalometric keypoint dataset [6] proposed in the 2023 MICCAI Challenge and also the commonly benchmarked Chest dataset [27] to further create the IKL protocols to validate our proposed method. Note that the Cephalometric keypoint dataset [6] differs from the head dataset [67] used in previous works [78, 79] commonly, as [6] has much larger number of keypoints and re-collected from different hospitals to enlarge the variance of images, making it much more challenging than [67]. Thus we term this dataset [6] as Head-2023 for clarity. Note that both Head-2023 and Chest datasets have less than or equal to 400 images in total. Moreover, though there is not a large need for IKL on human and animal, the widely used human and animal keypoint datasets, i.e., MPII [5] and ATRW [37], have extreme variations and non-rigid motions. Thus we also choose them for experiments to show the generality of our KAMP under different domains and challenging scenarios. We detail full dataset statistics in our Supplementary.", + "bbox": [ + 511, + 356, + 906, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared Methods. For keypoint estimation, common network structures for classification like ResNet [21] need specific modifications, e.g., adding the deconvolution layers and convolution layers to generate the keypoint location heatmap. These adjustments make many IL methods inapplicable for IKL, e.g., prototype-based methods [87, 88] like PASS [88] that leverages the feature mean prototype before the linear classification layer. Furthermore, the IKL requires that the old data can not be retained, and it is hard to identify specific methods for IKL. Thus we choose several general and representative exemplar-free methods that can adapt to different IL settings based on their methodologies, including EWC [33], LWF [40], MAS [72], RW [9], AFEC [69], CPR [8] for comparison. They are regularization-based methods and", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25403", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a3cdef58c9ba509fd022ad003e7076eec357de074f0904fea6e92df4d71a2f70.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSplit ChestSplit Head-2023Split MPIISplit ATRW
A-MRE1↓AT1↑MT1↑A-MRE4↓AT4↑MT4↑AAA4↑AT4↑MT4↑AAA3↑AT3↑MT3↑
Joint Training5.43--2.12--88.50--94.69--
Finetune43.1--51.3--37.41--13.24--
EWC [33]13.28-8.23-3.6710.97-6.37-4.7638.64-51.84-12.2114.38-59.75-2.08
RW [9]9.48-7.12-4.156.49-4.23-2.8838.47-18.83-7.1384.15-10.870.00
MAS [4]7.36-1.86-0.195.31-2.15-1.3367.03-7.560.3485.68-5.80-1.13
LWF [40]6.35-1.340.184.31-1.260.5775.75-3.860.4187.31-5.10-0.64
AFEC [69]8.04-2.670.155.77-3.45-1.4637.24-22.85-15.4233.03-40.25-8.02
CPR [8]6.17-0.870.293.71-1.180.1675.52-3.240.7589.34-2.764.49
SFD [17]7.68-0.540.134.76-0.430.0271.49-0.930.2186.11-1.130.41
WF [74]7.31-0.310.164.580.030.1172.87-0.460.3886.69-0.970.62
GBD [16]6.420.060.214.340.120.4775.62-0.180.3587.42-0.890.65
KAMP (Ours)5.670.290.622.320.410.8479.931.804.2393.16-0.845.13
", + "bbox": [ + 117, + 63, + 879, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Result of 4 datasets after 2-Step, 5-Step, 5-Step and 4-Step IKL for Chest, Head-2023, MPII, and ATRW respectively. All comparison methods are started from the same Step-0 trained model. A-MRE: smaller the better; AAA/AT/MT: larger the better.", + "bbox": [ + 89, + 250, + 903, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "can be easily applied to the IKL without trial and error. We further adapt methods from Continual Semantic Segmentation (CSS) and Class Incremental Learning (CIL) to ISL for more comparisons, i.e., spatial feature distillation (SFD) [17], weight fusion (WF) [74], and gradient balanced distillation (GBD) [16]. We also report the native baseline, i.e., directly finetune the model during IKL, and the upper bound that trains the model (e.g., HRNet [60]) with all the data, where we denote the former as \"Finetune\" and the latter as \"Joint Training\".", + "bbox": [ + 88, + 292, + 480, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. To assess the keypoint regression task, we employ the widely used mean radial error (MRE) for Head-2023 and Chest datasets as in [78, 79], and Probability of Correct Keypoint (PCK) [50, 56, 66, 77] for MPII and ATRW and use their defaulted $\\sigma$ as in [50, 60] to compute PCK. To measure the performance of incremental learning (IL), we use Average Accuracy (AAA), calculating the accuracy (\\%) across all keypoints post-incremental step $i$ under PCK, and extend this approach to MRE, denoting it as A-MRE. Additionally, we introduce two metrics for knowledge transfer in IL: (1) Average Transfer (AT), also known as backward transfer [47], which averages the performance improvement of keypoints over all previous steps after step $i$ , and (2) Maximal Transfer (MT), measuring the largest performance improvement in any old keypoint post-step $i$ . Notably, when calculating AT and MT with MRE, we invert the sign of the change in error. This adjustment ensures consistency, since a decrease in MRE signifies improvement, in contrast to the direct correlation of increased accuracy with improvement in the PCK metric. Detailed explanations for the calculation of each metric are provided in our supplementary materials.", + "bbox": [ + 91, + 445, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Design. We randomly split the keypoints of Chest, Head-2023, MPII, and ATRW into different incremental steps to create the Spilt Head-2023, Chest, MPII and ATRW protocols respectively. As the Chest dataset only has 6 keypoints in total, we split them into two groups where the first group has 4 keypoints with 150 training images, and the second group for IKL has 2 keypoints with only 50 training images. For Split Head-2023, we split the 38 keypoints into", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 groups, where the first group has 19 keypoints with the same definition as [67] with 100 training images, and we split the rest of keypoint into 4 groups randomly to simulate four incremental steps with only 50 training images for each step. Similarly, the 16 MPII keypoints and 15 ATRW keypoints were split into 5 and 4 groups, respectively, with an equal distribution of training images per step. We repeat five times under different orders of keypoint groups and report the mean value in the main paper. The standard deviation (std), details about the keypoints group, and results of other experimental setups are in our Supplementary.", + "bbox": [ + 511, + 292, + 903, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We use the HRNet [60] as the backbone for all methods as it is widely compared in keypoint estimation [19, 36, 39, 50, 66, 79]. For a fair comparison, we initialize all methods with the same initial step (Step-0) model for incremental learning. As the first benchmark for the IKL, we use the Continual Hyperparameter Framework (CHF) [13, 41] to identify the training parameters like training epoch (100), initial learning rate (2e-3 for Split Chest and Head-2023, 1e-2 for Split MPII, 1e-3 for Split-ATRW), momentum (0.9), weight decay (1e-4), and also the hyperparameters of each compared methods, all included in our Supplementary. For our KAMP, the hyperparameter $\\alpha$ is set as 1e2 for the Split Head-2023 and Split-MPII and 1e4 for the Split Chest and Split-ATRW in all experiments. The analysis of the $\\alpha$ and other implementation details are included in our Supplementary.", + "bbox": [ + 511, + 459, + 903, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with SOTA for IKL", + "text_level": 1, + "bbox": [ + 511, + 712, + 797, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 1, our KAMP consistently improves accuracy and reduces error across all old and new keypoints after each incremental step, achieving the highest performance compared to all other methods on all datasets. For example, KAMP outperforms the second-best method by $1.39\\%$ in $\\mathrm{A - MRE_4}$ after four incremental steps on Split Head-2023, and by $4.18\\%$ in $\\mathrm{AAA_4}$ on Split MPII. Moreover, KAMP is the only method to consistently yield the highest average transfer (AT) and maximal transfer (MT) scores across all datasets. This result highlights that our two-stage learning scheme better facilitates knowledge transfer, thereby en", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25404", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "hancing accuracy on old keypoints while learning new ones. When AT is positive, this indicates no forgetting on average, underscoring the effectiveness of using pseudo-labels from the old model to train KA-Net, as described in Sec. 3.2.1. In Split-ATRW, the AT is negative for all methods, indicating that forgetting of old keypoints outweighs knowledge transfer. However, KAMP still achieves the highest AT among the methods. Notably, the absolute value of KAMP's AT is nearly zero, suggesting minimal negative transfer. Additionally, by examining the maximal transfer metric, we observe a significant positive transfer in specific cases, such as $5.13\\%$ for $\\mathrm{MT}_3$ , indicating that some old keypoints experience a substantial positive transfer in each incremental step. This positive transfer helps offset the forgetting of other old keypoints, resulting in a minimal negative average transfer. We further validate this effect by reporting per-keypoint performance in our Supplementary. Qualitative results in Fig. 4 further verify our superiority in achieving more structurally correct prediction than competitive methods.", + "bbox": [ + 89, + 90, + 480, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lastly, we observe that methods adapted from CSS and CIL, i.e., SFD, WF, and GBD, achieve strong performance on AT and MT metrics compared to competitive methods like LWF and CPR, but perform poorly on overall metrics like A-MRE and AAA. This is because these methods prioritize reducing forgetting, resulting in an overly rigid model that struggles to effectively incorporate new keypoints in IKL. This observation highlights a crucial limitation of existing CSS and CIL methods that they focus heavily on regularizing forgetting but fail to balance this with knowledge acquisition for new keypoints, which is especially problematic in our newly defined IKL setting. Consequently, this emphasizes the need to study IKL as a distinct incremental learning scenario in this paper, as traditional incremental learning methods are too general to excel here. Our proposed novel baseline, KAMP, effectively addresses this gap for IKL, offering a well-balanced approach that goes beyond anti-forgetting to deliver robust performance across both accuracy and transfer metrics.", + "bbox": [ + 91, + 378, + 482, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 680, + 243, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the effectiveness of our method, we further analyze how each component may influence the result. Since our method contains two-stage training and we consider a keypoint-oriented spatial distillation loss $(\\ell_{KSD})$ , thus we compared our method with three alternatives: (1) the competitive baseline, i.e., the LWF, which performs the Softmax operation across old keypoints. (2) only use our $\\ell_{KSD}$ to train the model without the KA-Net. (3) Constructing the KA-Net randomly without using physical knowledge. Results are shown in Tab. 2, and we can observe that: (1) our adapted $\\ell_{KSD}$ can effectively outperform the LWF with $1.18\\%$ , which demonstrates the essence of our keypoint-oriented adaptation. (2) With proper auxiliary task creation", + "bbox": [ + 89, + 703, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9bc349fce77a40bc20027e1f385974a09571eb4f8f16e91bde38e4730d31ce2a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAAA4↑AT4↑MT4↑
LWF [40]75.75-3.860.41
KAMP (only ℓKSD)76.93-2.240.65
KAMP (Random KA-Net)77.13-0.481.24
KAMP (Ours)79.931.804.23
", + "bbox": [ + 527, + 88, + 893, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "based on physical prior, we can achieve more positive average transfer than other counterparts and finally achieve the largest improvement over all the keypoints on $\\mathrm{AAA}_4$ . This shows that our proposed two-stage learning scheme for IKL can not just provide better knowledge consolidation on the old keypoints than the competitive baseline, but the auxiliary prediction task can also bring large improvement to the old keypoints, showing that our KAMP will be a novel and strong baseline for the IKL paradigm. More ablation studies like using backbones other than HRNet and more datasets are in our Supplementary.", + "bbox": [ + 511, + 210, + 903, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Compare IKL to other low-data paradigms", + "text_level": 1, + "bbox": [ + 511, + 383, + 880, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Now we provide more insights into our IKL by comparing it with other learning paradigms with low-shot data. Discussion of limitations are in our Supplementary.", + "bbox": [ + 511, + 406, + 903, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As mentioned before, existing learning paradigms like UKL [22, 23, 78] and CAPE [10, 56, 76] leverage large-scale self-supervised and multi-dataset training to enable novel keypoint detection with a few labeled samples in test time. To fairly compare with UKL and CAPE, we extend KAMP to low-data regime following EGT [78], where we also pretrain an auxiliary self-supervised model at Step-0 to provide pseudo-labels for old keypoints during IKL, and the details are in the Supplementary. For medical datasets like Head-2023, we compare to the SOTA one-shot methods CC2D [78] and EGT [79]. For 2D datasets like MPII, we compare to the SOTA Autolink [22] and MetaPoint+ [10] with official implementation available. Note that methods like [86] require extra expensive 3D CAD models to identify keypoint proposals for novel keypoint detection and are inferior to the SOTA [10, 56], thus we only compare to the MetaPoint+ [10]. For KAMP, we use the low-data adaptation for 1, 5, and 10 shot and do not use it for 50-shot since 50-shot is large enough for us to bypass the adaptation. Full experimental details are in our Supplementary.", + "bbox": [ + 511, + 452, + 906, + 753 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4870fbf0d37e554cc68359876e3eb8d2c08c358f1b3785e60e0e57d55bcf6964.jpg", + "table_caption": [ + "Table 2. Ablation Study on Split MPII." + ], + "table_footnote": [], + "table_body": "
Method/(MRE ↓)1-shot5-shot10-shot50-shot
CC2D [78]5.144.834.083.47
EGT [79]5.014.583.873.21
KAMP (Ours)4.353.703.032.32
", + "bbox": [ + 524, + 763, + 895, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Compare to low-shot methods on Split Head-2023.", + "bbox": [ + 529, + 853, + 885, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Tab. 3 we observe that even on the extreme 1-shot setting, KAMP still outperforms the CC2D and EGT, while", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25405", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a26e20ec19212a20f8a594ec6b9c5df4390b1e52d27f7db6e2b1890108d71f5.jpg", + "image_caption": [ + "Figure 4. Qualitative results on Split Head-2023 and MPII. All methods start from the same Step-0 model, whose prediction is in the second column. GT: ground truth. The red circles denote the keypoints learned in Step 0, and the green ones denote all the new keypoints learned in later incremental steps. We observe that after IKL, the compared methods (LWF and CPR) may acquire the new keypoints as ours, but they have obvious miss-detection and wrong estimation (e.g., out of the body). While our method can consistently associate the new and old keypoints and achieve structurally accurate keypoint predictions. More results are included in our Supplementary." + ], + "image_footnote": [], + "bbox": [ + 96, + 93, + 903, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "all the compared methods do not have good performances. This is because the Head-2023 dataset is collected from multiple sources with large discrepancy, and only one image is hard to represent all the variations of the data. However, when more annotations are available, KAMP can scale much better than CC2D and EGT since our two-stage learning scheme can capture the relation of old and new keypoints and distill them back to the model to improve the old keypoints, which has not been handled in CC2D and EGT.", + "bbox": [ + 89, + 371, + 483, + 507 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c2bc7d004ba1dd51fb702b092fc69ca7a5eac2756a74dedc54b5c3d0a12aa75d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method/(PCK [10] ↑)1-shot5-shot10-shot50-shot
UKL [22]54.9559.1162.5364.36
MetaPoint+ [10]65.7166.8767.2668.98
KAMP (Ours)70.0972.2373.1876.97
KAMP (Ours) + [10]73.4975.1077.7679.18
", + "bbox": [ + 91, + 517, + 488, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Compare to low-shot methods on Split MPII.", + "bbox": [ + 125, + 619, + 446, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 4, we observe that while the CAPE [10, 56, 76] and IKL are two disentangled paradigms, they can complement each other effectively. CAPE relies on pretraining with varied datasets, while IKL focuses on incremental learning, making their training costs incomparable. For example, CAPE methods like MetaPoint+ [10] require a substantial training cost (4 GPUs, 1.5 days), whereas the 'pretrain+incremental' approach of IKL is more efficient (1 GPU, 5 hours). To demonstrate that CAPE and IKL are orthogonal and complementary, we applied IKL using our KAMP method on MetaPoint+. As shown in Tab. 4, this combined approach results in significant improvements, even in low-data scenarios, highlighting that IKL can substantially enhance keypoint estimation when applied on a CAPE-pretrained keypoint detector.", + "bbox": [ + 89, + 643, + 483, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The goal of comparing these paradigms for novel key-point estimation within the same object category is to es", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tabish the necessity of IKL. Although CAPE does not involve continual learning and appears more deploy-friendly, its performance plateaus and does not scale effectively with additional data. In contrast, our KAMP achieves higher accuracy with limited data and demonstrates superior scalability with increased data, as evidenced in Tab. 4.", + "bbox": [ + 511, + 371, + 906, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lastly, as the first study to explore the IKL paradigm, we focus on scaling up existing keypoint estimators for novel keypoints within the same object category, laying the foundation for future research. While CAPE can perform keypoint estimation on novel object categories with a few support images, its generalization remains limited. In the future, we will explore extending the IKL paradigm to novel object categories by applying IKL to keypoint estimators pretrained with CAPE, aiming to combine the strengths of both paradigms to develop a keypoint estimator capable of continually learning new keypoints across diverse object categories without forgetting and with effective scalability.", + "bbox": [ + 511, + 462, + 908, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, the comparison of Tab. 3 and 4 show that our proposed IKL paradigm is highly label-efficient for acquiring new keypoints. This characteristic is advantageous for real-world applications where obtaining sufficient labels is time-consuming, such as in medical analysis [6].", + "bbox": [ + 511, + 645, + 908, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 737, + 633, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We explore learning the newly defined keypoint incrementally without retaining any old data, called Incremental object Keypoint Learning (IKL). We propose a two-stage learning scheme as a novel baseline tailored to the IKL. Extensive experiments show that our method can effectively alleviate the forgetting issue and may even improve the old keypoints' estimation during IKL. Our further analysis reveals that the IKL is label efficient in acquiring the new keypoints, which is promising for real-world applications.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25406", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported in part by National Science Foundation grant IIS-2007613. This work was also supported by the National Natural Science Foundation of China (62376011).", + "bbox": [ + 89, + 90, + 485, + 148 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 193 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hello gpt-4o, howpublished = https://openai.com/index/hello-gpt-4o/.5", + "[2] Hongjoon Ahn, Sungmin Cha, Donggyu Lee, and Taesup Moon. Uncertainty-based continual learning with adaptive regularization. In Advances in Neural Information Processing Systems, pages 4392-4402, 2019. 3", + "[3] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3", + "[4] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 6", + "[5] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 1, 2, 3, 5", + "[6] Jun Cao, Juan Dai, Xuguang Li, Bingsheng Huang, Ching-Wei Wang, and Hongyuan Zhang. Cephalometric landmark detection in lateral x-ray images, 2023. 1, 2, 5, 8", + "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1", + "[8] Sungmin Cha, Hsiang Hsu, Taebaek Hwang, Flavio Calmon, and Taesup Moon. {CPR}: Classifier-projection regularization for continual learning. In International Conference on Learning Representations, 2021. 5, 6", + "[9] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajanthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European Conference on Computer Vision (ECCV), pages 532-547, 2018. 5, 6", + "[10] Junjie Chen, Jiebin Yan, Yuming Fang, and Li Niu. Metapoint learning and refining for category-agnostic pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23534–23543, 2024. 3, 7, 8", + "[11] Xiao Chu, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Structured feature learning for pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4715-4723, 2016. 4", + "[12] Shengyang Dai, Ming Yang, Ying Wu, and Aggelos Katsaggelos. Detector ensemble. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 1, 4" + ], + "bbox": [ + 93, + 202, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 3, 5, 6", + "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 3", + "[15] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5138-5146, 2019. 3", + "[16] Jiahua Dong, Wenqi Liang, Yang Cong, and Gan Sun. Heterogeneous forgetting compensation for class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11742-11751, 2023. 6", + "[17] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4040-4050, 2021. 6", + "[18] Ana Paula Reis Durão, Aline Morosolli, Pisha Pittayapat, Napat Bolstad, Afonso P Ferreira, and Reinhilde Jacobs. Cephalometric landmark variability among orthodontists and dentomaxillofacial radiologists: a comparative study. Imaging science in dentistry, 45(4):213-220, 2015. 1", + "[19] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14676-14686, 2021. 1, 6", + "[20] Kerui Gu, Linlin Yang, and Angela Yao. Removing the bias of integral pose regression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11067-11076, 2021. 3", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5", + "[22] Xingzhe He, Bastian Wandt, and Helge Rhodin. Autolink: Self-supervised learning of human skeletons and object outlines by linking keypoints. In Advances in Neural Information Processing Systems, 2022. 1, 7, 8", + "[23] Xingzhe He, Gaurav Bharaj, David Ferman, Helge Rhodin, and Pablo Garrido. Few-shot geometry-aware keypoint localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21337-21348, 2023. 1, 7", + "[24] Sina Honari, Pavlo Molchanov, Stephen Tyree, Pascal Vincent, Christopher Pal, and Jan Kautz. Improving landmark localization with semi-supervised learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1546-1555, 2018. 3", + "[25] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25407", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rebalancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 3", + "[26] Zhongzhan Huang, Mingfu Liang, Senwei Liang, and Shanshan Zhong. Flat local minima for continual learning on semantic segmentation. In International Conference on Multimedia Modeling, pages 388-400. Springer, 2025. 3", + "[27] Stefan Jaeger, Alexandros Karargyris, Sema Candemir, Les Folio, Jenifer Siegelman, Fiona Callaghan, Zhiyun Xue, Kannappan Palaniappan, Rahul K Singh, Sameer Antani, George Thoma, Yi-Xiang Wang, Pu-Xuan Lu, and Clement J McDonald. Automatic tuberculosis screening using chest radiographs. IEEE Trans. Med. Imaging, 33(2):233-245, 2014. 2, 5", + "[28] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. arXiv preprint arXiv:2207.13784, 2022. 1", + "[29] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In European Conference on Computer Vision, pages 196-214. Springer, 2020. 1, 3", + "[30] Ta-Chu Kao, Kristopher Jensen, Gido van de Ven, Alberto Bernacchia, and Guillaume Hennequin. Natural continual learning: success is a journey, not (just) a destination. Advances in Neural Information Processing Systems, 34: 28067-28079, 2021. 3", + "[31] Leonid Karlinsky and Shimon Ullman. Using linking features in learning non-parametric part models. In European Conference on Computer Vision, pages 326-339. Springer, 2012. 1", + "[32] Chris Dongjoo Kim, Jinseo Jeong, and Gunhee Kim. Imbalanced continual learning with partitioning reservoir sampling. In European Conference on Computer Vision, pages 411-428. Springer, 2020. 3", + "[33] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 3, 5, 6", + "[34] Jiefeng Li, Siyuan Bian, Ailing Zeng, Can Wang, Bo Pang, Wentao Liu, and Cewu Lu. Human pose regression with residual log-likelihood estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11025–11034, 2021. 3", + "[35] Jiefeng Li, Tong Chen, Ruiqi Shi, Yujing Lou, Yong-Lu Li, and Cewu Lu. Localization with sampling-argmax. Advances in Neural Information Processing Systems, 34: 27236-27248, 2021. 3", + "[36] Ke Li, Shijie Wang, Xiang Zhang, Yifan Xu, Weijian Xu, and Zhuowen Tu. Pose recognition with cascade transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1944-1953, 2021. 1, 6", + "[37] Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, and Weiyao Lin. Atrw: A benchmark for amur tiger re-identification in" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2590-2598, 2020. 2, 5", + "[38] Yanjie Li, Sen Yang, Shoukui Zhang, Zhicheng Wang, Wankou Yang, Shu-Tao Xia, and Erjin Zhou. Is 2d heatmap representation even necessary for human pose estimation? arXiv preprint arXiv:2107.03332, 2021. 3", + "[39] Yanjie Li, Shoukui Zhang, Zhicheng Wang, Sen Yang, Wankou Yang, Shu-Tao Xia, and Erjin Zhou. Tokenpose: Learning keypoint tokens for human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11313-11322, 2021. 1, 3, 6", + "[40] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, 2017. 3, 5, 6, 7", + "[41] Mingfu Liang, Jiahuan Zhou, Wei Wei, and Ying Wu. Balancing between forgetting and acquisition in incremental subpopulation learning. In European Conference on Computer Vision, pages 364-380. Springer, 2022. 3, 6", + "[42] Mingfu Liang, Jong-Chyi Su, Samuel Schulter, Sparsh Garg, Shiyu Zhao, Ying Wu, and Manmohan Chandraker. Aide: An automatic data engine for object detection in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14695-14706, 2024. 5", + "[43] Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. Beyond not-forgetting: Continual learning with backward knowledge transfer. In Advances in Neural Information Processing Systems, 2022. 3", + "[44] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 2, 3", + "[45] Buyu Liu and Vittorio Ferrari. Active learning for human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 4363-4372, 2017. 1", + "[46] Vincenzo Lomonaco and Davide Maltoni. Core50: a new dataset and benchmark for continuous object recognition. In Conference on Robot Learning, pages 17-26. PMLR, 2017. 3", + "[47] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Advances in Neural Information Processing Systems, pages 6467-6476, 2017. 3, 6", + "[48] Zhengxiong Luo, Zhicheng Wang, Yan Huang, Liang Wang, Tieniu Tan, and Erjin Zhou. Rethinking the heatmap regression for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13264-13273, 2021. 3", + "[49] Dimitrios Mallis, Enrique Sanchez, Matthew Bell, and Georgios Tzimiropoulos. Unsupervised learning of object landmarks via self-training correspondence. Advances in Neural Information Processing Systems, 33:4709-4720, 2020. 1, 3", + "[50] Olga Moskvyak, Frederic Maire, Feras Dayoub, and Mahsa Baktashmotlagh. Semi-supervised keypoint localization. In International Conference on Learning Representations, 2021. 1, 3, 6" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25408", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Gaurav Kumar Nayak, Het Shah, and Anirban Chakraborty. Incremental learning for animal pose estimation using rbf k-dpp. arXiv preprint arXiv:2110.13598, 2021. 3", + "[52] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In European Conference on Computer Vision, pages 524-540. Springer, 2020. 3", + "[53] Ricketts Robert. Orthodontic Diagnosis and Planning: Their Roles in Preventive and Rehabilitative Dentistry. Denver Colo; Rocky Mountain/Orthodontics, 1982. 1", + "[54] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In Proceedings of the IEEE international conference on computer vision workshops, pages 397-403, 2013. 3", + "[55] Enrique Sanchez and Georgios Tzimiropoulos. Object landmark discovery through unsupervised adaptation. Advances in Neural Information Processing Systems, 32, 2019. 1", + "[56] Min Shi, Zihao Huang, Xianzheng Ma, Xiaowei Hu, and Zhiguo Cao. Matching is not enough: A two-stage framework for category-agnostic pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7308–7317, 2023. 3, 6, 7, 8", + "[57] Aliaksandr Siarohin, Enver Sangineto, Stéphane Lathuiliere, and Nicu Sebe. Deformable gans for pose-based human image generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3408-3416, 2018. 1", + "[58] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2377-2386, 2019. 1", + "[59] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 2", + "[60] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693-5703, 2019. 6", + "[61] Wei Tang and Ying Wu. Does learning specific features for related parts help human pose estimation? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1107-1116, 2019. 3", + "[62] Wei Tang, Pei Yu, and Ying Wu. Deeply learned compositional models for human pose estimation. In Proceedings of the European conference on computer vision (ECCV), pages 190-206, 2018. 3, 4", + "[63] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial embeddings. In Proceedings of the IEEE international conference on computer vision, pages 5916-5925, 2017. 3", + "[64] Jonathan J. Thompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems, 27, 2014. 4" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Riccardo Volpi, Diane Larlus, and Grégory Rogez. Continual adaptation of visual representations via domain randomization and meta-learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4443-4453, 2021. 3", + "[66] Can Wang, Sheng Jin, Yingda Guan, Wentao Liu, Chen Qian, Ping Luo, and Wanli Ouyang. Pseudo-labeled autocurriculum learning for semi-supervised keypoint localization. In International Conference on Learning Representations, 2022. 3, 6", + "[67] Ching-Wei Wang, Cheng-Ta Huang, Jia-Hong Lee, Chung-Hsing Li, Sheng-Wei Chang, Ming-Jhih Siao, Tat-Ming Lai, Bulat Ibragimov, Tomaz Vrtovec, Olaf Ronneberger, Philipp Fischer, Tim F Cootes, and Claudia Lindner. A benchmark for comparison of dental radiography analysis algorithms. Med. Image Anal., 31:63-76, 2016. 1, 5, 6", + "[68] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 1, 3", + "[69] Liyuan Wang, Mingtian Zhang, Zhongfan Jia, Qian Li, Chenglong Bao, Kaisheng Ma, Jun Zhu, and Yi Zhong. Afec: Active forgetting of negative transfer in continual learning. Advances in Neural Information Processing Systems, 34: 22379-22391, 2021. 5, 6", + "[70] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[71] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 4724–4732, 2016. 3", + "[72] Chenshen Wu, Luis Herranz, Xialei Liu, Joost van de Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. In Advances in Neural Information Processing Systems, pages 5962-5972, 2018. 5", + "[73] Guile Wu, Shaogang Gong, and Pan Li. Striking a balance between stability and plasticity for class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1124-1133, 2021. 3", + "[74] Jia-Wen Xiao, Chang-Bin Zhang, Jiekang Feng, Xialei Liu, Joost van de Weijer, and Ming-Ming Cheng. Endpoints weight fusion for class incremental semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7204-7213, 2023. 6", + "[75] Xiaoying Xing, Mingfu Liang, and Ying Wu. Toa: Task-oriented active vqa. Advances in Neural Information Processing Systems, 36:54061-54074, 2023. 5", + "[76] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In Computer Vision-ECCV 2022: 17th European Conference," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25409", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 398-416. Springer, 2022. 2, 3, 7, 8", + "[77] Sen Yang, Zhibin Quan, Mu Nie, and Wankou Yang. Transpose: Keypoint localization via transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11802-11812, 2021. 1, 3, 6", + "[78] Qingsong Yao, Quan Quan, Li Xiao, and S Kevin Zhou. One-shot medical landmark detection. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24, pages 177-188. Springer, 2021. 3, 5, 6, 7", + "[79] Zihao Yin, Ping Gong, Chunyu Wang, Yizhou Yu, and Yizhou Wang. One-shot medical landmark localization by edge-guided transform and noisy landmark refinement. In European Conference on Computer Vision, pages 473-489. Springer, 2022. 5, 6, 7", + "[80] Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. AP-10k: A benchmark for animal pose estimation in the wild. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 1, 3", + "[81] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6982-6991, 2020. 3", + "[82] Duncan Zauss, Sven Kreiss, and Alexandre Alahi. Keypoint communities. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11057-11066, 2021. 3", + "[83] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. Proceedings of Machine Learning Research, 70:3987, 2017. 3", + "[84] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2694-2703, 2018. 1, 3", + "[85] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13208-13217, 2020. 3", + "[86] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018. 7", + "[87] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34:14306-14318, 2021. 5", + "[88] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 5" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 11 + }, + { + "type": "ref_text", + "text": "[89] Xu Zou, Sheng Zhong, Luxin Yan, Xiangyun Zhao, Jiahuan Zhou, and Ying Wu. Learning robust facial landmark detection via hierarchical structured ensemble. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 141-150, 2019. 3", + "bbox": [ + 516, + 90, + 903, + 161 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "25410", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Incremental Object Keypoint Learning/d3b12539-b476-45cb-a752-bae4990c3622_model.json b/2025/Incremental Object Keypoint Learning/d3b12539-b476-45cb-a752-bae4990c3622_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f4e5b057c682d6179b313119819220b278123120 --- /dev/null +++ b/2025/Incremental Object Keypoint Learning/d3b12539-b476-45cb-a752-bae4990c3622_model.json @@ -0,0 +1,2292 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.301, + 0.131, + 0.698, + 0.154 + ], + "angle": 0, + "content": "Incremental Object Keypoint Learning" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.18, + 0.73, + 0.199 + ], + "angle": 0, + "content": "Mingfu Liang\\(^{1}\\) Jiahuan Zhou\\(^{2,*}\\) Xu Zou\\(^{3}\\) Ying Wu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.199, + 0.873, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\) Northwestern University \\(^{2}\\) Wangxuan Institute of Computer Technology, Peking University \\(^{3}\\) Huazhong University of Science and Technology" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.302, + 0.486, + 0.786 + ], + "angle": 0, + "content": "Existing progress in object keypoint estimation primarily benefits from the conventional supervised learning paradigm based on numerous data labeled with pre-defined keypoints. However, these well-trained models can hardly detect the undefined new keypoints in test time, which largely hinders their feasibility for diverse downstream tasks. To handle this, various solutions are explored but still suffer from either limited generalizability or transferability. Therefore, in this paper, we explore a novel keypoint learning paradigm in that we only annotate new keypoints in the new data and incrementally train the model, without retaining any old data, called Incremental object Keypoint Learning (IKL). A two-stage learning scheme as a novel baseline tailored to IKL is developed. In the first Knowledge Association stage, given the data labeled with only new keypoints, an auxiliary KA-Net is trained to automatically associate the old keypoints to these new ones based on their spatial and intrinsic anatomical relations. In the second Mutual Promotion stage, based on a keypoint-oriented spatial distillation loss, we jointly leverage the auxiliary KA-Net and the old model for knowledge consolidation to mutually promote the estimation of all old and new keypoints. Owing to the investigation of the correlations between new and old keypoints, our proposed method can not just effectively mitigate the catastrophic forgetting of old keypoints, but may even further improve the estimation of the old ones and achieve a positive transfer beyond antiforgetting. Such an observation has been solidly verified by extensive experiments on different keypoint datasets, where our method exhibits superiority in alleviating the forgetting issue and boosting performance while enjoying labeling efficiency even under the low-shot data regime." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.817, + 0.222, + 0.833 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.844, + 0.483, + 0.874 + ], + "angle": 0, + "content": "As a fundamental task in computer vision, estimating the visual keypoint locations of an object serves as the indis-" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.268, + 0.856, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.385, + 0.905, + 0.412 + ], + "angle": 0, + "content": "Figure 1. Medical analysis can frequently change and require new keypoints essentially [6], while the labeling is time-consuming." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.566 + ], + "angle": 0, + "content": "pensable primitive to support numerous down-stream tasks, e.g., object pose detection [29, 77, 80], tracking [28], action recognition [7], generation [57] and animation [28, 58], etc. Over a long period of time, supervised keypoint learning (SKL) [19, 36, 39, 68] has significantly advanced the progress on keypoint estimation. By exploiting large-scale pure 2D keypoint datasets [5, 44] with extreme variance, SKL aims to train a deep neural network that can robustly detect a set of pre-defined keypoints in test images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.839 + ], + "angle": 0, + "content": "However, the SKL models can not estimate newly-added undefined keypoints of an object, while new demands from downstream tasks will inevitably require new keypoints. For instance, until now, models trained by existing keypoint dataset [67] can only detect 19 keypoints to support Downs analysis in the Cephalometric analysis, while they are infeasible for important analysis like Steiner analysis [53], as shown in Fig. 1. Such an issue motivated a recent MIC-CAI challenge [6] on annotating much more keypoints to support increasing analysis. However, keypoint labeling is time-consuming [18, 45, 50], especially when keypoints are defined sequentially. Another alternative solution is to train separate new keypoint estimator [12, 31] for each set of new keypoints as in Fig. 2. However, the number of estimators will increase linearly, and the separate training strategy will make each estimator easily overfit the new keypoints [12] and hardly capture the intrinsic relation between old and new keypoints, leading to sub-optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To mitigate the above issues, unsupervised keypoint learning (UKL) [22, 23, 49, 55, 84] is proposed for new keypoint estimation. The models are firstly pretrained in an unsupervised manner to predict numbers of keypoints ran" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.248, + 0.9 + ], + "angle": 0, + "content": "*: Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25399" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.088, + 0.905, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.296, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 2. Separate Estimators need multiple estimators in test time. Unsupervised keypoint learning (UKL) and category-agnostic pose estimation (CAPE) exploit a fixed pretrained keypoint estimator on unseen new keypoints. While our Incremental object Keypoint Keaning (IKL) continually updates the same model on the new data labeled only on new keypoints without retaining any old training data." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.346, + 0.486, + 0.574 + ], + "angle": 0, + "content": "domly, then a linear transformation is learned to transfer the existing keypoints to the new undefined ones. However, UKL is merely effective on objects with relatively rigid motion or video data with consistent object instances and local changes in appearance and motion. As a result, it is hard to perform UKL on large-scale pure 2D datasets [5, 44] due to their extreme inter-image differences in the background, appearance, and motions of distinct object instances. Recently, a supervised metric learning-based [59] paradigm named category-agnostic pose estimation (CAPE) [76] is proposed to learn a class-agnostic pose estimator that can estimate unseen keypoints in test time with a few labeled support samples. However, as the model is frozen after pretraining, CAPE is limited by its restricted generalization and transferability for newly-defined keypoints (Tab. 4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Therefore, in this paper, we propose to solve the new keypoint estimation problem creatively in an incremental learning (IL) manner. As shown in Fig. 2, we propose an Incremental object Keypoint Learning (IKL) paradigm to incrementally train a keypoint estimator on the new data labeled only with new keypoints. Even without retaining any old data, the IKL model should not forget the old keypoints catastrophically. Compared to the naive solutions that either label both new and old keypoints or learn separate estimators, our IKL only labels the new keypoints and maintains just one model in the lifetime, which is computationally efficient. Moreover, different from the UKL and CAPE whose performances are largely limited by their restricted transferability depending on a fixed pretrained model, the IKL alleviates such an issue by continually training the model to sufficiently expand its knowledge on newly-defined keypoints and may even improve the old keypoint estimation when learning more relevant new keypoints. Lastly, as most works in IL mainly study the classification task [13, 70], to our best knowledge, IKL as an incremental keypoint regression task has rarely been stud" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.346, + 0.864, + 0.362 + ], + "angle": 0, + "content": "ied before, which also provides novel insights for IL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.907, + 0.534 + ], + "angle": 0, + "content": "However, IKL also poses a new challenge known as label non-co-occurrence (LNCO). As only new keypoints are labeled in the new data, the kinematic and anatomical constraints between old and new keypoints are not explicitly presented in the label space, making it hard for the model to capture such a physical prior during incremental training. Existing IL methods (Tab. 1) do not explicitly model such inter-keypoint relations in IKL, thus the old and new keypoint predictions can not mutually support each other and may readily bias toward the new ones, exacerbating the forgetting of the old keypoints (Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To tackle the challenge of IKL, we propose a novel two-stage learning scheme called KAMP as a new baseline tailored to IKL. In the first Knowledge Association stage, we train an auxiliary KA-Net to associate old and new keypoints based on their physiological connection, represented by their spatial adjacency. Specifically, KA-Net learns to predict the selected old keypoint given the related new keypoints to acquire their intrinsic anatomical relevance. In the second Mutual Promotion stage, we train the new model on all keypoints to mutually promote their estimation, where the old ones are updated by distilling from the KA-Net and the old model with a newly designed keypoint-oriented spatial distillation loss for better knowledge preservation. To verify our effectiveness, we simulate IKL based on four keypoints datasets on both medical and natural images, i.e., Cephalometric [6], Chest [27], MPII [5] and ATRW [37]. Extensive results demonstrate that our KAMP can not just effectively alleviate the catastrophic forgetting of old keypoints, but even further boost their performance and thus outperform the existing exemplar-free IL methods by a significant margin (Tab. 1) and can also work for low-shot scenarios (Tab. 3 and 4). Our analysis further reveals that IKL is complementary to learning paradigms like CAPE (Tab. 4) and labeling-efficient for new keypoint estimation." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "25400" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "To summarize, our contributions are three-fold: (1) We establish a novel paradigm, Incremental object Keypoint Learning (IKL), to tackle the challenging demands of new keypoints estimation. (2) We propose a two-stage exemplar-free IKL method to explicitly model and exploit the relation between old and new keypoints to help alleviate the label non-co-occurrence problem tailored to IKL, which may further boost the performance of old keypoints beyond antiforgetting. (3) As a proof-of-concept, we empirically show that IKL is practical and labeling-efficient to scale up a pretrained keypoint estimator on new keypoints estimation, which is much better than other alternative paradigms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.287, + 0.24, + 0.303 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.313, + 0.485, + 0.705 + ], + "angle": 0, + "content": "Keypoint Estimation (KE). Existing research of KE focuses on estimating a fixed number of keypoints for a specific category [5, 29, 44, 54, 78-80]. Most works design new methods [20, 34, 35, 38, 89] or network architectures [39, 48, 61, 62, 68, 77] to improve the supervised learning over large-scale pure 2D keypoint datasets [5, 44]. Estimating dense keypoints [29, 82] of the human body has been proposed but requires extreme labeling cost. Recent works explore semi-supervised [24, 50, 66] or unsupervised learning [63, 84] to reduce the laborious labeling consumption, but the unsupervised methods still hardly achieve a precise estimation of semantically meaningful keypoints [49]. Recently, Category-agnostic keypoint estimation (CAPE) [10, 56, 76] is proposed to learn a class-agnostic model to detect keypoints of unseen categories by a few labeled support images without retraining the model. However, as the model is frozen after pretraining, CAPE is limited by restricted generalization and transferability due to the intra-class appearance variation, self-occlusion, and appearance ambiguity, as detailed in [76]. Different from the above paradigm that pre-defined a set of keypoints or assumed the pretrained model was sufficient enough for different downstream tasks, our proposed IKL paradigm continuously updates the model only on the new keypoints to increasingly expand its knowledge, which is much more flexible and labeling-efficient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Incremental Learning (IL). Existing IL methods can be categorized as exemplar-free and exemplar-based ones [3, 25, 26, 32, 41, 52], given whether old data can be retained. However, saving exemplars not just introduces increasing costs on memory, but also has privacy issues in real-world applications. For exemplar-free methods, existing approaches focus on adding regularization over parameters [2, 33, 47, 81, 83] or the network's output in different positions [15, 40, 73, 85]. Regarding the IL settings, adding new classes [13] or domains [41, 46, 65] have been studied popularly and mainly tested on classification benchmarks [13, 14]. In contrast, keypoint estimation is a location regression task [20, 34, 48], and our proposed IKL aims" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "to detect newly defined keypoints incrementally. The incremental animal pose estimation [51] (IAPE) is a related setting to us. However, IAPE assumes the label space of keypoints is fixed for each animal, and it only learns new animals' poses, i.e., only the input distribution is changed with new animal domains. While for our IKL, we instead consider adding new keypoints to the label space for an object category, and the old and new data distribution may be relatively correlated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.44 + ], + "angle": 0, + "content": "Positive Transfer (PT) in IL. Most IL approaches focus on alleviating the catastrophic forgetting problem [13], while only a few works are dedicated to achieving positive transfer (PT) in IL [30, 43, 47], i.e., improving the old tasks while learning a new task. Those methods encourage PT by modifying the gradient updates based on the new and old tasks' correlations which are measured by saving the data [47] from the old task or based on formal analyses [30, 43]. Differently, our KAMP achieves PT by explicitly modeling the relation between old and new keypoints based on their intrinsic anatomical relevance, and we exploits the relation to design novel distillation to further boost the estimation of the old keypoint, which is tailored to the IKL and keypoint estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.866, + 0.471 + ], + "angle": 0, + "content": "3. Incremental Object Keypoint Learning" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.478, + 0.719, + 0.493 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.907, + 0.765 + ], + "angle": 0, + "content": "We denote \\( t = 0 \\) as the initial step and \\( t = 1, \\dots, T \\) as the incremental steps. Then the training set for the \\( t \\)-th incremental step is \\( D_{t}^{\\mathrm{train}} = \\left\\{x_{t,j}^{\\mathrm{train}}, y_{t,j}^{\\mathrm{train}}\\right\\}_{j=1}^{N_{t}} \\), where \\( x, y, N_{t} \\) denote the inputs, the keypoint labels and the size of \\( D_{t}^{\\mathrm{train}} \\). As we only label new keypoints in IKL, thus the newly introduced keypoints must be disjointed with the old ones, i.e., \\( y_{t}^{\\mathrm{train}} \\cap \\left(\\cup_{i=0}^{t-1} y_{i}^{\\mathrm{train}}\\right) = \\emptyset \\). Different from the classification task, the common practice for keypoint estimation is to regress a 2D heatmap for a keypoint, i.e., a Gaussian peak around the ground-truth keypoint location [71]. Thus the model \\( m_{t} \\) in IKL comprises the feature extractor \\( f_{t} \\) and a stack of convolutional keypoint estimation heads \\( \\{G_{i}\\}_{i=1}^{t} \\) that parameterized by \\( \\theta_{t} \\) and \\( \\{\\phi_{i}\\}_{i=1}^{t} \\), respectively. Each estimation head ends with a convolutional layer to generate the 2D heatmap. \\( G_{t} \\) denotes the estimation head for the new keypoints in the \\( t \\)-th step. \\( f_{t} \\) and \\( \\{G_{i}\\}_{i=1}^{t-1} \\) in model \\( m_{t} \\) are initialized by the same weights in the old model \\( m_{t-1} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.81, + 0.788 + ], + "angle": 0, + "content": "3.2. KAMP: A Novel Baseline for IKL" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In this section, we present our two-stage method KAMP as a novel baseline tailored to IKL. We conjecture that distilling the relation between related old and new keypoints into the model may help it implicitly capture the intrinsic physical prior to alleviate the label non-co-occurrence issue. To do so, for each incremental step, in Stage-I, we design an auxiliary prediction task to associate the related" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25401" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.154, + 0.088, + 0.615, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.089, + 0.844, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.284, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Figure 3. Overview of KAMP using the human body for illustration. In Stage-I, we learn an auxiliary KA-Net to associate the related old and new keypoints based on their local anatomical constraint. In Stage-II, we jointly leverage the old model and the KA-Net as an auxiliary teacher to consolidate all old keypoints' prediction and also learn the new keypoints simultaneously to achieve mutual promotions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.335, + 0.482, + 0.41 + ], + "angle": 0, + "content": "old keypoint's with the new ones by the KA-Net. In Stage-II, the old model \\( m_{t - 1} \\) and KA-Net are both frozen to distill knowledge into the new model \\( m_t \\) for all old keypoints, and the new keypoints are learned concurrently to achieve mutual promotion. Fig. 3 provides a schematic view of KAMP." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.417, + 0.398, + 0.431 + ], + "angle": 0, + "content": "3.2.1. Stage-I: Knowledge Association (KA)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.435, + 0.483, + 0.601 + ], + "angle": 0, + "content": "As mentioned in Sec. 1, the old and new keypoints in IKL are not jointly labeled, making it difficult for the model to learn relationships between keypoints, such as their structural and anatomical relevance, using only the labels of new keypoints. To mitigate this issue, we propose to model the constraint among the spatially and anatomically related old and new keypoints as an implicit function in the current incremental step \\( t \\). Existing analyses [11, 12, 64] show that keypoints association can be well-modeled by the triangulation constraint. Motivated by this, we define a triangulation constraint \\( F \\) on the distribution of three related keypoints:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.604, + 0.482, + 0.621 + ], + "angle": 0, + "content": "\\[\nF \\left(P \\left(K _ {i}\\right), P \\left(K _ {j}\\right), P \\left(K _ {k}\\right)\\right) = 0, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.704 + ], + "angle": 0, + "content": "where \\( P(K_{i}) \\) denotes the distribution of the keypoint \\( i \\), and the tuple may include one or two new keypoints. However, as there may be multiple valid solutions (constraints) that satisfy the implicit function \\( F \\), it is challenging to derive an analytical solution for \\( F \\) in IKL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Therefore, we consider an alternative formulation, e.g., \\( P(K_{\\mathrm{o}_j}) = F(P(K_{\\mathrm{n}_1}),P(K_{\\mathrm{n}_2})) \\), which is a special case of the Eqn. 1 that we condition the distribution of an old keypoint \\( P(K_{\\mathrm{o}_j}) \\) on its related new keypoints \\( P(K_{\\mathrm{n}_1}) \\) and \\( P(K_{\\mathrm{n}_2}) \\). We focus on two new keypoints since only they have ground-truth labels, and the ground-truth supervision can help reduce the uncertainty in learning the constraints. Moreover, in practice, new tasks typically require more than one new keypoint for support. Nonetheless, we verify in our Supplementary that our approach performs effectively even with a single new keypoint. To enhance this constraint, we model the conditional probability between the related old and new keypoints as an auxiliary prediction task:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.334, + 0.905, + 0.351 + ], + "angle": 0, + "content": "\\[\nP \\left(K _ {\\mathrm {o} j} \\mid K _ {\\mathrm {n} 1}, K _ {\\mathrm {n} 2}, v\\right) = \\mathrm {K A} \\left(P \\left(K _ {\\mathrm {n} 1}\\right), P \\left(K _ {\\mathrm {n} 2}\\right), v\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.905, + 0.432 + ], + "angle": 0, + "content": "where KA stands for the Knowledge Association Network (KA-Net). The insight of this auxiliary task is inspired by the fact that spatially adjacent and visually correlated keypoints can mutually predict each other as shown in many existing works [11, 62, 64]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.649 + ], + "angle": 0, + "content": "KA-Net. The input of the KA-Net is the ground-truth heatmap \\((y_{i,\\mathrm{n}_1}^{gt}\\) and \\(y_{i,\\mathrm{n}_2}^{gt})\\) of new keypoints \\(K_{\\mathrm{n_1}}\\) and \\(K_{\\mathrm{n_2}}\\) of an image \\(i\\), and the output is the predicted heatmap \\(\\tilde{y}_{i,j}\\) of the old keypoint \\(K_{\\mathrm{o_j}}\\). \\(v_{i}\\) denotes the holistic visual features \\(v\\) of the image \\(i\\) comprised of the intermediate features extracted from the frozen feature \\(f_{t - 1}\\) and are re-scaled to the same spatial size as \\(y_{i,\\mathrm{n_1}}^{gt}\\) and \\(y_{i,\\mathrm{n_2}}^{gt}\\). The visual features \\(v_{i}\\) are incorporated alongside spatial information (\\(y_{i,\\mathrm{n_1}}^{gt}\\) and \\(y_{i,\\mathrm{n_2}}^{gt}\\)) because, in a 2D image, spatial coordinates alone are insufficient to determine a keypoint's location due to uncertainty stemming from factors such as camera angle, object appearance, and motion. By incorporating visual features, we provide additional contextual information that helps account for this uncertainty in keypoint associations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.814 + ], + "angle": 0, + "content": "In the present paper, we explore KA-Net's simplest construction to minimize its training cost: for each new keypoint, we first perform the element-wise multiplication between its ground-truth heatmap \\((y_{i,\\mathrm{n}_1}^{gt}\\) and \\(y_{i,\\mathrm{n}_2}^{gt})\\) and the holistic features \\(v_{i}\\) to obtain the keypoint-oriented spatial features. Then we concatenate the spatial features and feed them forward over three convolutional layers accompanied by the Batch Normalization (BN) and ReLU to predict the selected old keypoint \\(K_{\\mathrm{o}_j}\\). Note that the trained KA-Net will only be used to distill the keypoint association in the Stage-II and will not be used in test time after IKL." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.815, + 0.907, + 0.875 + ], + "angle": 0, + "content": "Training of KA-Net. We train KA-Net by the new data \\( D_{t}^{\\mathrm{train}} \\). As \\( D_{t}^{\\mathrm{train}} \\) does not have ground truth labels for old keypoints, we use the pseudo-labels predicted by the old model \\( m_{t - 1} \\) to supervise KA-Net for keypoint regression:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.875, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\ell_ {K A} = \\frac {1}{N _ {t}} \\sum_ {i} ^ {N t} \\sum_ {j \\in \\mathcal {K} _ {K A}} \\left\\| \\hat {y} _ {i, j} ^ {t - 1} - \\tilde {y} _ {i, j} \\right\\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25402" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "where \\(\\mathcal{K}_{KA}\\) denotes the selected old keypoints used for the auxiliary prediction task, \\(\\hat{y}_{i,j}^{t-1}\\) denotes the prediction of the \\(j\\)-th keypoint by \\(m_{t-1}\\) given image \\(i\\), and \\(\\tilde{y}_{i,j}\\) denotes the prediction by the KA-Net. We will show in Sec. 4 that our KAMP can effectively reduce the forgetting of old keypoints and even improve them, which largely reduce the accumulative error to use pseudo-labels for KA-Net." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.199, + 0.487, + 0.486 + ], + "angle": 0, + "content": "Auxiliary Task Creation. To select the old keypoint for training KA-Net, in each incremental step, we first locate all the old and new keypoints on a general object anatomy diagram by their semantic definition. Then we measure their relative proximity by simple distance calculations (e.g., Euclidean distance) to identify two new keypoints \\((K_{\\mathrm{n_1}}\\) and \\(K_{\\mathrm{n_2}})\\) that are close to an old one \\(K_{\\mathrm{o_j}}\\). This process can also be automated by multi-modality large language models like GPT-4o [1, 42, 75] as shown in Fig. 3. If we identify several tuples of old and new keypoints satisfied the requirement, we randomly choose one tuple to create the auxiliary prediction task to avoid training too many KA-Net. For all results of our KAMP in Section 4, we show that even by creating only one auxiliary task in each step, we can still bring sufficient improvement on all keypoints. Note that task creation does not need any training, and we only need to perform once before each step, which is highly efficient. Such a design provides an interpretable way to incorporate the physical knowledge of the object to guide the IKL." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.496, + 0.373, + 0.511 + ], + "angle": 0, + "content": "3.2.2. Stage-II: Mutual Promotion (MP)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.483, + 0.547 + ], + "angle": 0, + "content": "In Stage-II, we jointly optimize all the new and old keypoints on the new model \\( m_t \\) by the loss \\( l_{MP} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.554, + 0.483, + 0.571 + ], + "angle": 0, + "content": "\\[\nl _ {M P} = \\ell_ {G T} + \\alpha \\left(\\ell_ {K S D} + \\ell_ {K A}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.901 + ], + "angle": 0, + "content": "where \\(\\ell_{GT}\\) denotes the L2 loss between the ground truth of the new keypoints and their predictions by the new model to acquire the knowledge of new keypoints. \\(\\alpha\\) is a hyperparameter to balance the new keypoints acquisition and old keypoints forgetting. For \\(\\ell_{KA}\\), we use the frozen KA-Net as an auxiliary teacher to supervised the old keypoint selected in Stage-I to distill the keypoint association knowledge. Since \\(\\ell_{KA}\\) is only applied to the selected old keypoint for knowledge transfer instead of mitigating the forgetting, we further consolidate the knowledge of all old keypoints by the loss \\(\\ell_{KSD}\\) using the frozen old model \\(m_{t - 1}\\). As the distribution of new and old data in IKL do not change dramatically and may even have a strong correlation, LWF [40] can be a baseline for \\(\\ell_{KSD}\\) in such a scenario as analyzed in [13]. In general, \\(\\ell_{KSD}\\) represents the negative log-likelihood \\(\\frac{1}{N_t}\\sum_i^{N_t}s\\left(\\hat{y}_i^{t - 1}\\right)\\cdot \\log s\\left(\\hat{y}_i^t\\right)\\) between the predictions of all old keypoints by the new model, \\(\\hat{y}_i^{t - 1}\\in \\mathbb{R}^{C\\times H\\times W}\\), and predictions by the old model, \\(\\hat{y}_i^t\\in \\mathbb{R}^{C\\times H\\times W}\\). \\(s(\\cdot)\\) denotes the Softmax operator. \\(C\\) denotes the numbers of old keypoints learned before step \\(t\\), \\(H\\) and \\(W\\) are the height and width of each keypoint's heatmap." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "However, since LWF and its variants all focus on the classification task, they perform the Softmax across different classes by default, i.e., over the \\(C\\) dimension, to obtain the normalized class prediction score. For IKL, this means normalizing across all old keypoints' predictions, which can not explicitly regularize the discrepancy of each old keypoint between the old and new model. To better preserve keypoint-specific knowledge, we adopt a spatial softmax operation over height \\((H)\\) and width \\((W)\\) dimensions of the keypoint prediction, \\(s_{\\mathrm{sp}}^{\\mathrm{d}}(\\cdot), \\mathrm{d} \\in \\{H, W\\}\\), and combine them to encourage spatial-oriented knowledge distillation. We discuss in our Supplementary the difference between each softmax alternative. Thus, our \\(\\ell_{KSD}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.291, + 0.905, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{N _ {t}} \\sum_ {i} ^ {N t} \\sum_ {j} ^ {C} \\sum_ {\\mathrm {d}} - s _ {\\mathrm {s p}} ^ {\\mathrm {d}} (\\hat {y} _ {i, j} ^ {t - 1}) \\cdot \\log s _ {\\mathrm {s p}} ^ {\\mathrm {d}} (\\hat {y} _ {i, j} ^ {t}). \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.332, + 0.638, + 0.349 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.675 + ], + "angle": 0, + "content": "Datasets. As the IKL has rarely been studied before, there is no specific dataset or benchmark for IKL. Motivated by the real-world application of IKL in medical analysis, we leverage the large-scale Cephalometric keypoint dataset [6] proposed in the 2023 MICCAI Challenge and also the commonly benchmarked Chest dataset [27] to further create the IKL protocols to validate our proposed method. Note that the Cephalometric keypoint dataset [6] differs from the head dataset [67] used in previous works [78, 79] commonly, as [6] has much larger number of keypoints and re-collected from different hospitals to enlarge the variance of images, making it much more challenging than [67]. Thus we term this dataset [6] as Head-2023 for clarity. Note that both Head-2023 and Chest datasets have less than or equal to 400 images in total. Moreover, though there is not a large need for IKL on human and animal, the widely used human and animal keypoint datasets, i.e., MPII [5] and ATRW [37], have extreme variations and non-rigid motions. Thus we also choose them for experiments to show the generality of our KAMP under different domains and challenging scenarios. We detail full dataset statistics in our Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Compared Methods. For keypoint estimation, common network structures for classification like ResNet [21] need specific modifications, e.g., adding the deconvolution layers and convolution layers to generate the keypoint location heatmap. These adjustments make many IL methods inapplicable for IKL, e.g., prototype-based methods [87, 88] like PASS [88] that leverages the feature mean prototype before the linear classification layer. Furthermore, the IKL requires that the old data can not be retained, and it is hard to identify specific methods for IKL. Thus we choose several general and representative exemplar-free methods that can adapt to different IL settings based on their methodologies, including EWC [33], LWF [40], MAS [72], RW [9], AFEC [69], CPR [8] for comparison. They are regularization-based methods and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25403" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.064, + 0.88, + 0.239 + ], + "angle": 0, + "content": "
MethodSplit ChestSplit Head-2023Split MPIISplit ATRW
A-MRE1↓AT1↑MT1↑A-MRE4↓AT4↑MT4↑AAA4↑AT4↑MT4↑AAA3↑AT3↑MT3↑
Joint Training5.43--2.12--88.50--94.69--
Finetune43.1--51.3--37.41--13.24--
EWC [33]13.28-8.23-3.6710.97-6.37-4.7638.64-51.84-12.2114.38-59.75-2.08
RW [9]9.48-7.12-4.156.49-4.23-2.8838.47-18.83-7.1384.15-10.870.00
MAS [4]7.36-1.86-0.195.31-2.15-1.3367.03-7.560.3485.68-5.80-1.13
LWF [40]6.35-1.340.184.31-1.260.5775.75-3.860.4187.31-5.10-0.64
AFEC [69]8.04-2.670.155.77-3.45-1.4637.24-22.85-15.4233.03-40.25-8.02
CPR [8]6.17-0.870.293.71-1.180.1675.52-3.240.7589.34-2.764.49
SFD [17]7.68-0.540.134.76-0.430.0271.49-0.930.2186.11-1.130.41
WF [74]7.31-0.310.164.580.030.1172.87-0.460.3886.69-0.970.62
GBD [16]6.420.060.214.340.120.4775.62-0.180.3587.42-0.890.65
KAMP (Ours)5.670.290.622.320.410.8479.931.804.2393.16-0.845.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.251, + 0.905, + 0.279 + ], + "angle": 0, + "content": "Table 1. Result of 4 datasets after 2-Step, 5-Step, 5-Step and 4-Step IKL for Chest, Head-2023, MPII, and ATRW respectively. All comparison methods are started from the same Step-0 trained model. A-MRE: smaller the better; AAA/AT/MT: larger the better." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.293, + 0.482, + 0.444 + ], + "angle": 0, + "content": "can be easily applied to the IKL without trial and error. We further adapt methods from Continual Semantic Segmentation (CSS) and Class Incremental Learning (CIL) to ISL for more comparisons, i.e., spatial feature distillation (SFD) [17], weight fusion (WF) [74], and gradient balanced distillation (GBD) [16]. We also report the native baseline, i.e., directly finetune the model during IKL, and the upper bound that trains the model (e.g., HRNet [60]) with all the data, where we denote the former as \"Finetune\" and the latter as \"Joint Training\"." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.446, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Evaluation Metrics. To assess the keypoint regression task, we employ the widely used mean radial error (MRE) for Head-2023 and Chest datasets as in [78, 79], and Probability of Correct Keypoint (PCK) [50, 56, 66, 77] for MPII and ATRW and use their defaulted \\(\\sigma\\) as in [50, 60] to compute PCK. To measure the performance of incremental learning (IL), we use Average Accuracy (AAA), calculating the accuracy (\\%) across all keypoints post-incremental step \\(i\\) under PCK, and extend this approach to MRE, denoting it as A-MRE. Additionally, we introduce two metrics for knowledge transfer in IL: (1) Average Transfer (AT), also known as backward transfer [47], which averages the performance improvement of keypoints over all previous steps after step \\(i\\), and (2) Maximal Transfer (MT), measuring the largest performance improvement in any old keypoint post-step \\(i\\). Notably, when calculating AT and MT with MRE, we invert the sign of the change in error. This adjustment ensures consistency, since a decrease in MRE signifies improvement, in contrast to the direct correlation of increased accuracy with improvement in the PCK metric. Detailed explanations for the calculation of each metric are provided in our supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Experimental Design. We randomly split the keypoints of Chest, Head-2023, MPII, and ATRW into different incremental steps to create the Spilt Head-2023, Chest, MPII and ATRW protocols respectively. As the Chest dataset only has 6 keypoints in total, we split them into two groups where the first group has 4 keypoints with 150 training images, and the second group for IKL has 2 keypoints with only 50 training images. For Split Head-2023, we split the 38 keypoints into" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.905, + 0.459 + ], + "angle": 0, + "content": "5 groups, where the first group has 19 keypoints with the same definition as [67] with 100 training images, and we split the rest of keypoint into 4 groups randomly to simulate four incremental steps with only 50 training images for each step. Similarly, the 16 MPII keypoints and 15 ATRW keypoints were split into 5 and 4 groups, respectively, with an equal distribution of training images per step. We repeat five times under different orders of keypoint groups and report the mean value in the main paper. The standard deviation (std), details about the keypoints group, and results of other experimental setups are in our Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.905, + 0.702 + ], + "angle": 0, + "content": "Implementation Details. We use the HRNet [60] as the backbone for all methods as it is widely compared in keypoint estimation [19, 36, 39, 50, 66, 79]. For a fair comparison, we initialize all methods with the same initial step (Step-0) model for incremental learning. As the first benchmark for the IKL, we use the Continual Hyperparameter Framework (CHF) [13, 41] to identify the training parameters like training epoch (100), initial learning rate (2e-3 for Split Chest and Head-2023, 1e-2 for Split MPII, 1e-3 for Split-ATRW), momentum (0.9), weight decay (1e-4), and also the hyperparameters of each compared methods, all included in our Supplementary. For our KAMP, the hyperparameter \\(\\alpha\\) is set as 1e2 for the Split Head-2023 and Split-MPII and 1e4 for the Split Chest and Split-ATRW in all experiments. The analysis of the \\(\\alpha\\) and other implementation details are included in our Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.799, + 0.728 + ], + "angle": 0, + "content": "4.1. Comparison with SOTA for IKL" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 1, our KAMP consistently improves accuracy and reduces error across all old and new keypoints after each incremental step, achieving the highest performance compared to all other methods on all datasets. For example, KAMP outperforms the second-best method by \\(1.39\\%\\) in \\(\\mathrm{A - MRE_4}\\) after four incremental steps on Split Head-2023, and by \\(4.18\\%\\) in \\(\\mathrm{AAA_4}\\) on Split MPII. Moreover, KAMP is the only method to consistently yield the highest average transfer (AT) and maximal transfer (MT) scores across all datasets. This result highlights that our two-stage learning scheme better facilitates knowledge transfer, thereby en" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25404" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.379 + ], + "angle": 0, + "content": "hancing accuracy on old keypoints while learning new ones. When AT is positive, this indicates no forgetting on average, underscoring the effectiveness of using pseudo-labels from the old model to train KA-Net, as described in Sec. 3.2.1. In Split-ATRW, the AT is negative for all methods, indicating that forgetting of old keypoints outweighs knowledge transfer. However, KAMP still achieves the highest AT among the methods. Notably, the absolute value of KAMP's AT is nearly zero, suggesting minimal negative transfer. Additionally, by examining the maximal transfer metric, we observe a significant positive transfer in specific cases, such as \\(5.13\\%\\) for \\(\\mathrm{MT}_3\\), indicating that some old keypoints experience a substantial positive transfer in each incremental step. This positive transfer helps offset the forgetting of other old keypoints, resulting in a minimal negative average transfer. We further validate this effect by reporting per-keypoint performance in our Supplementary. Qualitative results in Fig. 4 further verify our superiority in achieving more structurally correct prediction than competitive methods." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.38, + 0.483, + 0.668 + ], + "angle": 0, + "content": "Lastly, we observe that methods adapted from CSS and CIL, i.e., SFD, WF, and GBD, achieve strong performance on AT and MT metrics compared to competitive methods like LWF and CPR, but perform poorly on overall metrics like A-MRE and AAA. This is because these methods prioritize reducing forgetting, resulting in an overly rigid model that struggles to effectively incorporate new keypoints in IKL. This observation highlights a crucial limitation of existing CSS and CIL methods that they focus heavily on regularizing forgetting but fail to balance this with knowledge acquisition for new keypoints, which is especially problematic in our newly defined IKL setting. Consequently, this emphasizes the need to study IKL as a distinct incremental learning scenario in this paper, as traditional incremental learning methods are too general to excel here. Our proposed novel baseline, KAMP, effectively addresses this gap for IKL, offering a well-balanced approach that goes beyond anti-forgetting to deliver robust performance across both accuracy and transfer metrics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.245, + 0.697 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To verify the effectiveness of our method, we further analyze how each component may influence the result. Since our method contains two-stage training and we consider a keypoint-oriented spatial distillation loss \\((\\ell_{KSD})\\), thus we compared our method with three alternatives: (1) the competitive baseline, i.e., the LWF, which performs the Softmax operation across old keypoints. (2) only use our \\(\\ell_{KSD}\\) to train the model without the KA-Net. (3) Constructing the KA-Net randomly without using physical knowledge. Results are shown in Tab. 2, and we can observe that: (1) our adapted \\(\\ell_{KSD}\\) can effectively outperform the LWF with \\(1.18\\%\\), which demonstrates the essence of our keypoint-oriented adaptation. (2) With proper auxiliary task creation" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.089, + 0.895, + 0.182 + ], + "angle": 0, + "content": "
MethodAAA4↑AT4↑MT4↑
LWF [40]75.75-3.860.41
KAMP (only ℓKSD)76.93-2.240.65
KAMP (Random KA-Net)77.13-0.481.24
KAMP (Ours)79.931.804.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.592, + 0.192, + 0.825, + 0.206 + ], + "angle": 0, + "content": "Table 2. Ablation Study on Split MPII." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.211, + 0.905, + 0.378 + ], + "angle": 0, + "content": "based on physical prior, we can achieve more positive average transfer than other counterparts and finally achieve the largest improvement over all the keypoints on \\(\\mathrm{AAA}_4\\). This shows that our proposed two-stage learning scheme for IKL can not just provide better knowledge consolidation on the old keypoints than the competitive baseline, but the auxiliary prediction task can also bring large improvement to the old keypoints, showing that our KAMP will be a novel and strong baseline for the IKL paradigm. More ablation studies like using backbones other than HRNet and more datasets are in our Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.385, + 0.881, + 0.402 + ], + "angle": 0, + "content": "4.3. Compare IKL to other low-data paradigms" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.905, + 0.452 + ], + "angle": 0, + "content": "Now we provide more insights into our IKL by comparing it with other learning paradigms with low-shot data. Discussion of limitations are in our Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.755 + ], + "angle": 0, + "content": "As mentioned before, existing learning paradigms like UKL [22, 23, 78] and CAPE [10, 56, 76] leverage large-scale self-supervised and multi-dataset training to enable novel keypoint detection with a few labeled samples in test time. To fairly compare with UKL and CAPE, we extend KAMP to low-data regime following EGT [78], where we also pretrain an auxiliary self-supervised model at Step-0 to provide pseudo-labels for old keypoints during IKL, and the details are in the Supplementary. For medical datasets like Head-2023, we compare to the SOTA one-shot methods CC2D [78] and EGT [79]. For 2D datasets like MPII, we compare to the SOTA Autolink [22] and MetaPoint+ [10] with official implementation available. Note that methods like [86] require extra expensive 3D CAD models to identify keypoint proposals for novel keypoint detection and are inferior to the SOTA [10, 56], thus we only compare to the MetaPoint+ [10]. For KAMP, we use the low-data adaptation for 1, 5, and 10 shot and do not use it for 50-shot since 50-shot is large enough for us to bypass the adaptation. Full experimental details are in our Supplementary." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.765, + 0.897, + 0.844 + ], + "angle": 0, + "content": "
Method/(MRE ↓)1-shot5-shot10-shot50-shot
CC2D [78]5.144.834.083.47
EGT [79]5.014.583.873.21
KAMP (Ours)4.353.703.032.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.531, + 0.854, + 0.887, + 0.868 + ], + "angle": 0, + "content": "Table 3. Compare to low-shot methods on Split Head-2023." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "From Tab. 3 we observe that even on the extreme 1-shot setting, KAMP still outperforms the CC2D and EGT, while" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25405" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.094, + 0.904, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.294, + 0.908, + 0.366 + ], + "angle": 0, + "content": "Figure 4. Qualitative results on Split Head-2023 and MPII. All methods start from the same Step-0 model, whose prediction is in the second column. GT: ground truth. The red circles denote the keypoints learned in Step 0, and the green ones denote all the new keypoints learned in later incremental steps. We observe that after IKL, the compared methods (LWF and CPR) may acquire the new keypoints as ours, but they have obvious miss-detection and wrong estimation (e.g., out of the body). While our method can consistently associate the new and old keypoints and achieve structurally accurate keypoint predictions. More results are included in our Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.484, + 0.508 + ], + "angle": 0, + "content": "all the compared methods do not have good performances. This is because the Head-2023 dataset is collected from multiple sources with large discrepancy, and only one image is hard to represent all the variations of the data. However, when more annotations are available, KAMP can scale much better than CC2D and EGT since our two-stage learning scheme can capture the relation of old and new keypoints and distill them back to the model to improve the old keypoints, which has not been handled in CC2D and EGT." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.518, + 0.489, + 0.611 + ], + "angle": 0, + "content": "
Method/(PCK [10] ↑)1-shot5-shot10-shot50-shot
UKL [22]54.9559.1162.5364.36
MetaPoint+ [10]65.7166.8767.2668.98
KAMP (Ours)70.0972.2373.1876.97
KAMP (Ours) + [10]73.4975.1077.7679.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.126, + 0.621, + 0.447, + 0.635 + ], + "angle": 0, + "content": "Table 4. Compare to low-shot methods on Split MPII." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.87 + ], + "angle": 0, + "content": "As shown in Tab. 4, we observe that while the CAPE [10, 56, 76] and IKL are two disentangled paradigms, they can complement each other effectively. CAPE relies on pretraining with varied datasets, while IKL focuses on incremental learning, making their training costs incomparable. For example, CAPE methods like MetaPoint+ [10] require a substantial training cost (4 GPUs, 1.5 days), whereas the 'pretrain+incremental' approach of IKL is more efficient (1 GPU, 5 hours). To demonstrate that CAPE and IKL are orthogonal and complementary, we applied IKL using our KAMP method on MetaPoint+. As shown in Tab. 4, this combined approach results in significant improvements, even in low-data scenarios, highlighting that IKL can substantially enhance keypoint estimation when applied on a CAPE-pretrained keypoint detector." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The goal of comparing these paradigms for novel key-point estimation within the same object category is to es" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.462 + ], + "angle": 0, + "content": "tabish the necessity of IKL. Although CAPE does not involve continual learning and appears more deploy-friendly, its performance plateaus and does not scale effectively with additional data. In contrast, our KAMP achieves higher accuracy with limited data and demonstrates superior scalability with increased data, as evidenced in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.909, + 0.646 + ], + "angle": 0, + "content": "Lastly, as the first study to explore the IKL paradigm, we focus on scaling up existing keypoint estimators for novel keypoints within the same object category, laying the foundation for future research. While CAPE can perform keypoint estimation on novel object categories with a few support images, its generalization remains limited. In the future, we will explore extending the IKL paradigm to novel object categories by applying IKL to keypoint estimators pretrained with CAPE, aiming to combine the strengths of both paradigms to develop a keypoint estimator capable of continually learning new keypoints across diverse object categories without forgetting and with effective scalability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.909, + 0.723 + ], + "angle": 0, + "content": "In summary, the comparison of Tab. 3 and 4 show that our proposed IKL paradigm is highly label-efficient for acquiring new keypoints. This characteristic is advantageous for real-world applications where obtaining sufficient labels is time-consuming, such as in medical analysis [6]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.738, + 0.634, + 0.755 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We explore learning the newly defined keypoint incrementally without retaining any old data, called Incremental object Keypoint Learning (IKL). We propose a two-stage learning scheme as a novel baseline tailored to the IKL. Extensive experiments show that our method can effectively alleviate the forgetting issue and may even improve the old keypoints' estimation during IKL. Our further analysis reveals that the IKL is label efficient in acquiring the new keypoints, which is promising for real-world applications." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25406" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.15 + ], + "angle": 0, + "content": "Acknowledgements This work was supported in part by National Science Foundation grant IIS-2007613. This work was also supported by the National Natural Science Foundation of China (62376011)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.178, + 0.188, + 0.194 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.203, + 0.482, + 0.23 + ], + "angle": 0, + "content": "[1] Hello gpt-4o, howpublished = https://openai.com/index/hello-gpt-4o/.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.232, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[2] Hongjoon Ahn, Sungmin Cha, Donggyu Lee, and Taesup Moon. Uncertainty-based continual learning with adaptive regularization. In Advances in Neural Information Processing Systems, pages 4392-4402, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.29, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[3] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[4] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[5] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[6] Jun Cao, Juan Dai, Xuguang Li, Bingsheng Huang, Ching-Wei Wang, and Hongyuan Zhang. Cephalometric landmark detection in lateral x-ray images, 2023. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[8] Sungmin Cha, Hsiang Hsu, Taebaek Hwang, Flavio Calmon, and Taesup Moon. {CPR}: Classifier-projection regularization for continual learning. In International Conference on Learning Representations, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[9] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajanthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European Conference on Computer Vision (ECCV), pages 532-547, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[10] Junjie Chen, Jiebin Yan, Yuming Fang, and Li Niu. Metapoint learning and refining for category-agnostic pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23534–23543, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[11] Xiao Chu, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Structured feature learning for pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4715-4723, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Shengyang Dai, Ming Yang, Ying Wu, and Aggelos Katsaggelos. Detector ensemble. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 1, 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.203, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[13] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[15] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5138-5146, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[16] Jiahua Dong, Wenqi Liang, Yang Cong, and Gan Sun. Heterogeneous forgetting compensation for class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11742-11751, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[17] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4040-4050, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[18] Ana Paula Reis Durão, Aline Morosolli, Pisha Pittayapat, Napat Bolstad, Afonso P Ferreira, and Reinhilde Jacobs. Cephalometric landmark variability among orthodontists and dentomaxillofacial radiologists: a comparative study. Imaging science in dentistry, 45(4):213-220, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[19] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14676-14686, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[20] Kerui Gu, Linlin Yang, and Angela Yao. Removing the bias of integral pose regression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11067-11076, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[22] Xingzhe He, Bastian Wandt, and Helge Rhodin. Autolink: Self-supervised learning of human skeletons and object outlines by linking keypoints. In Advances in Neural Information Processing Systems, 2022. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[23] Xingzhe He, Gaurav Bharaj, David Ferman, Helge Rhodin, and Pablo Garrido. Few-shot geometry-aware keypoint localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21337-21348, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Sina Honari, Pavlo Molchanov, Stephen Tyree, Pascal Vincent, Christopher Pal, and Jan Kautz. Improving landmark localization with semi-supervised learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1546-1555, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25407" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "rebalancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.192 + ], + "angle": 0, + "content": "[26] Zhongzhan Huang, Mingfu Liang, Senwei Liang, and Shanshan Zhong. Flat local minima for continual learning on semantic segmentation. In International Conference on Multimedia Modeling, pages 388-400. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[27] Stefan Jaeger, Alexandros Karargyris, Sema Candemir, Les Folio, Jenifer Siegelman, Fiona Callaghan, Zhiyun Xue, Kannappan Palaniappan, Rahul K Singh, Sameer Antani, George Thoma, Yi-Xiang Wang, Pu-Xuan Lu, and Clement J McDonald. Automatic tuberculosis screening using chest radiographs. IEEE Trans. Med. Imaging, 33(2):233-245, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[28] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. arXiv preprint arXiv:2207.13784, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[29] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In European Conference on Computer Vision, pages 196-214. Springer, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[30] Ta-Chu Kao, Kristopher Jensen, Gido van de Ven, Alberto Bernacchia, and Guillaume Hennequin. Natural continual learning: success is a journey, not (just) a destination. Advances in Neural Information Processing Systems, 34: 28067-28079, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[31] Leonid Karlinsky and Shimon Ullman. Using linking features in learning non-parametric part models. In European Conference on Computer Vision, pages 326-339. Springer, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[32] Chris Dongjoo Kim, Jinseo Jeong, and Gunhee Kim. Imbalanced continual learning with partitioning reservoir sampling. In European Conference on Computer Vision, pages 411-428. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[33] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[34] Jiefeng Li, Siyuan Bian, Ailing Zeng, Can Wang, Bo Pang, Wentao Liu, and Cewu Lu. Human pose regression with residual log-likelihood estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11025–11034, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[35] Jiefeng Li, Tong Chen, Ruiqi Shi, Yujing Lou, Yong-Lu Li, and Cewu Lu. Localization with sampling-argmax. Advances in Neural Information Processing Systems, 34: 27236-27248, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[36] Ke Li, Shijie Wang, Xiang Zhang, Yifan Xu, Weijian Xu, and Zhuowen Tu. Pose recognition with cascade transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1944-1953, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[37] Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, and Weiyao Lin. Atrw: A benchmark for amur tiger re-identification in" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2590-2598, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[38] Yanjie Li, Sen Yang, Shoukui Zhang, Zhicheng Wang, Wankou Yang, Shu-Tao Xia, and Erjin Zhou. Is 2d heatmap representation even necessary for human pose estimation? arXiv preprint arXiv:2107.03332, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[39] Yanjie Li, Shoukui Zhang, Zhicheng Wang, Sen Yang, Wankou Yang, Shu-Tao Xia, and Erjin Zhou. Tokenpose: Learning keypoint tokens for human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11313-11322, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[40] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, 2017. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[41] Mingfu Liang, Jiahuan Zhou, Wei Wei, and Ying Wu. Balancing between forgetting and acquisition in incremental subpopulation learning. In European Conference on Computer Vision, pages 364-380. Springer, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[42] Mingfu Liang, Jong-Chyi Su, Samuel Schulter, Sparsh Garg, Shiyu Zhao, Ying Wu, and Manmohan Chandraker. Aide: An automatic data engine for object detection in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14695-14706, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[43] Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. Beyond not-forgetting: Continual learning with backward knowledge transfer. In Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[44] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[45] Buyu Liu and Vittorio Ferrari. Active learning for human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 4363-4372, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[46] Vincenzo Lomonaco and Davide Maltoni. Core50: a new dataset and benchmark for continuous object recognition. In Conference on Robot Learning, pages 17-26. PMLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[47] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Advances in Neural Information Processing Systems, pages 6467-6476, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[48] Zhengxiong Luo, Zhicheng Wang, Yan Huang, Liang Wang, Tieniu Tan, and Erjin Zhou. Rethinking the heatmap regression for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13264-13273, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[49] Dimitrios Mallis, Enrique Sanchez, Matthew Bell, and Georgios Tzimiropoulos. Unsupervised learning of object landmarks via self-training correspondence. Advances in Neural Information Processing Systems, 33:4709-4720, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Olga Moskvyak, Frederic Maire, Feras Dayoub, and Mahsa Baktashmotlagh. Semi-supervised keypoint localization. In International Conference on Learning Representations, 2021. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25408" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[51] Gaurav Kumar Nayak, Het Shah, and Anirban Chakraborty. Incremental learning for animal pose estimation using rbf k-dpp. arXiv preprint arXiv:2110.13598, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[52] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In European Conference on Computer Vision, pages 524-540. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[53] Ricketts Robert. Orthodontic Diagnosis and Planning: Their Roles in Preventive and Rehabilitative Dentistry. Denver Colo; Rocky Mountain/Orthodontics, 1982. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[54] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In Proceedings of the IEEE international conference on computer vision workshops, pages 397-403, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.344 + ], + "angle": 0, + "content": "[55] Enrique Sanchez and Georgios Tzimiropoulos. Object landmark discovery through unsupervised adaptation. Advances in Neural Information Processing Systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[56] Min Shi, Zihao Huang, Xianzheng Ma, Xiaowei Hu, and Zhiguo Cao. Matching is not enough: A two-stage framework for category-agnostic pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7308–7317, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[57] Aliaksandr Siarohin, Enver Sangineto, Stéphane Lathuiliere, and Nicu Sebe. Deformable gans for pose-based human image generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3408-3416, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.496, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[58] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2377-2386, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.566, + 0.483, + 0.608 + ], + "angle": 0, + "content": "[59] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[60] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693-5703, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[61] Wei Tang and Ying Wu. Does learning specific features for related parts help human pose estimation? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1107-1116, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[62] Wei Tang, Pei Yu, and Ying Wu. Deeply learned compositional models for human pose estimation. In Proceedings of the European conference on computer vision (ECCV), pages 190-206, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[63] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial embeddings. In Proceedings of the IEEE international conference on computer vision, pages 5916-5925, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[64] Jonathan J. Thompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems, 27, 2014. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[65] Riccardo Volpi, Diane Larlus, and Grégory Rogez. Continual adaptation of visual representations via domain randomization and meta-learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4443-4453, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[66] Can Wang, Sheng Jin, Yingda Guan, Wentao Liu, Chen Qian, Ping Luo, and Wanli Ouyang. Pseudo-labeled autocurriculum learning for semi-supervised keypoint localization. In International Conference on Learning Representations, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[67] Ching-Wei Wang, Cheng-Ta Huang, Jia-Hong Lee, Chung-Hsing Li, Sheng-Wei Chang, Ming-Jhih Siao, Tat-Ming Lai, Bulat Ibragimov, Tomaz Vrtovec, Olaf Ronneberger, Philipp Fischer, Tim F Cootes, and Claudia Lindner. A benchmark for comparison of dental radiography analysis algorithms. Med. Image Anal., 31:63-76, 2016. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[68] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[69] Liyuan Wang, Mingtian Zhang, Zhongfan Jia, Qian Li, Chenglong Bao, Kaisheng Ma, Jun Zhu, and Yi Zhong. Afec: Active forgetting of negative transfer in continual learning. Advances in Neural Information Processing Systems, 34: 22379-22391, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.473, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[70] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[71] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 4724–4732, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[72] Chenshen Wu, Luis Herranz, Xialei Liu, Joost van de Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. In Advances in Neural Information Processing Systems, pages 5962-5972, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[73] Guile Wu, Shaogang Gong, and Pan Li. Striking a balance between stability and plasticity for class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1124-1133, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[74] Jia-Wen Xiao, Chang-Bin Zhang, Jiekang Feng, Xialei Liu, Joost van de Weijer, and Ming-Ming Cheng. Endpoints weight fusion for class incremental semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7204-7213, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.8, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[75] Xiaoying Xing, Mingfu Liang, and Ying Wu. Toa: Task-oriented active vqa. Advances in Neural Information Processing Systems, 36:54061-54074, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.844, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[76] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In Computer Vision-ECCV 2022: 17th European Conference," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25409" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 398-416. Springer, 2022. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[77] Sen Yang, Zhibin Quan, Mu Nie, and Wankou Yang. Transpose: Keypoint localization via transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11802-11812, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[78] Qingsong Yao, Quan Quan, Li Xiao, and S Kevin Zhou. One-shot medical landmark detection. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24, pages 177-188. Springer, 2021. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.329 + ], + "angle": 0, + "content": "[79] Zihao Yin, Ping Gong, Chunyu Wang, Yizhou Yu, and Yizhou Wang. One-shot medical landmark localization by edge-guided transform and noisy landmark refinement. In European Conference on Computer Vision, pages 473-489. Springer, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[80] Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. AP-10k: A benchmark for animal pose estimation in the wild. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[81] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6982-6991, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.536 + ], + "angle": 0, + "content": "[82] Duncan Zauss, Sven Kreiss, and Alexandre Alahi. Keypoint communities. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11057-11066, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[83] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. Proceedings of Machine Learning Research, 70:3987, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[84] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2694-2703, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[85] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13208-13217, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[86] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.777, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[87] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34:14306-14318, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[88] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[89] Xu Zou, Sheng Zhong, Luxin Yan, Xiangyun Zhao, Jiahuan Zhou, and Ying Wu. Learning robust facial landmark detection via hierarchical structured ensemble. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 141-150, 2019. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25410" + } + ] +] \ No newline at end of file diff --git a/2025/Incremental Object Keypoint Learning/d3b12539-b476-45cb-a752-bae4990c3622_origin.pdf b/2025/Incremental Object Keypoint Learning/d3b12539-b476-45cb-a752-bae4990c3622_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4345739a94f3a841ace5ad1d8e0ac91d6d9a7e80 --- /dev/null +++ b/2025/Incremental Object Keypoint Learning/d3b12539-b476-45cb-a752-bae4990c3622_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7992cae0be60b0602a69e4e722260551e71010cc57b620ca2e05135b76609390 +size 2599714 diff --git a/2025/Incremental Object Keypoint Learning/full.md b/2025/Incremental Object Keypoint Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..901fe42f587043b115114242ee112a200456e86b --- /dev/null +++ b/2025/Incremental Object Keypoint Learning/full.md @@ -0,0 +1,289 @@ +# Incremental Object Keypoint Learning + +Mingfu Liang $^{1}$ Jiahuan Zhou $^{2,*}$ Xu Zou $^{3}$ Ying Wu $^{1}$ + +$^{1}$ Northwestern University $^{2}$ Wangxuan Institute of Computer Technology, Peking University $^{3}$ Huazhong University of Science and Technology + +# Abstract + +Existing progress in object keypoint estimation primarily benefits from the conventional supervised learning paradigm based on numerous data labeled with pre-defined keypoints. However, these well-trained models can hardly detect the undefined new keypoints in test time, which largely hinders their feasibility for diverse downstream tasks. To handle this, various solutions are explored but still suffer from either limited generalizability or transferability. Therefore, in this paper, we explore a novel keypoint learning paradigm in that we only annotate new keypoints in the new data and incrementally train the model, without retaining any old data, called Incremental object Keypoint Learning (IKL). A two-stage learning scheme as a novel baseline tailored to IKL is developed. In the first Knowledge Association stage, given the data labeled with only new keypoints, an auxiliary KA-Net is trained to automatically associate the old keypoints to these new ones based on their spatial and intrinsic anatomical relations. In the second Mutual Promotion stage, based on a keypoint-oriented spatial distillation loss, we jointly leverage the auxiliary KA-Net and the old model for knowledge consolidation to mutually promote the estimation of all old and new keypoints. Owing to the investigation of the correlations between new and old keypoints, our proposed method can not just effectively mitigate the catastrophic forgetting of old keypoints, but may even further improve the estimation of the old ones and achieve a positive transfer beyond antiforgetting. Such an observation has been solidly verified by extensive experiments on different keypoint datasets, where our method exhibits superiority in alleviating the forgetting issue and boosting performance while enjoying labeling efficiency even under the low-shot data regime. + +# 1. Introduction + +As a fundamental task in computer vision, estimating the visual keypoint locations of an object serves as the indis- + +![](images/59d36260c6f937c2d3c3a51708fd622dfb077665f6fefc60bc8f6272a6cc47ed.jpg) +Figure 1. Medical analysis can frequently change and require new keypoints essentially [6], while the labeling is time-consuming. + +pensable primitive to support numerous down-stream tasks, e.g., object pose detection [29, 77, 80], tracking [28], action recognition [7], generation [57] and animation [28, 58], etc. Over a long period of time, supervised keypoint learning (SKL) [19, 36, 39, 68] has significantly advanced the progress on keypoint estimation. By exploiting large-scale pure 2D keypoint datasets [5, 44] with extreme variance, SKL aims to train a deep neural network that can robustly detect a set of pre-defined keypoints in test images. + +However, the SKL models can not estimate newly-added undefined keypoints of an object, while new demands from downstream tasks will inevitably require new keypoints. For instance, until now, models trained by existing keypoint dataset [67] can only detect 19 keypoints to support Downs analysis in the Cephalometric analysis, while they are infeasible for important analysis like Steiner analysis [53], as shown in Fig. 1. Such an issue motivated a recent MIC-CAI challenge [6] on annotating much more keypoints to support increasing analysis. However, keypoint labeling is time-consuming [18, 45, 50], especially when keypoints are defined sequentially. Another alternative solution is to train separate new keypoint estimator [12, 31] for each set of new keypoints as in Fig. 2. However, the number of estimators will increase linearly, and the separate training strategy will make each estimator easily overfit the new keypoints [12] and hardly capture the intrinsic relation between old and new keypoints, leading to sub-optimal performance. + +To mitigate the above issues, unsupervised keypoint learning (UKL) [22, 23, 49, 55, 84] is proposed for new keypoint estimation. The models are firstly pretrained in an unsupervised manner to predict numbers of keypoints ran + +![](images/e57ab3b3387667f941eb381fbf19203dfa588a21d2a90fe22206936958567688.jpg) +Figure 2. Separate Estimators need multiple estimators in test time. Unsupervised keypoint learning (UKL) and category-agnostic pose estimation (CAPE) exploit a fixed pretrained keypoint estimator on unseen new keypoints. While our Incremental object Keypoint Keaning (IKL) continually updates the same model on the new data labeled only on new keypoints without retaining any old training data. + +domly, then a linear transformation is learned to transfer the existing keypoints to the new undefined ones. However, UKL is merely effective on objects with relatively rigid motion or video data with consistent object instances and local changes in appearance and motion. As a result, it is hard to perform UKL on large-scale pure 2D datasets [5, 44] due to their extreme inter-image differences in the background, appearance, and motions of distinct object instances. Recently, a supervised metric learning-based [59] paradigm named category-agnostic pose estimation (CAPE) [76] is proposed to learn a class-agnostic pose estimator that can estimate unseen keypoints in test time with a few labeled support samples. However, as the model is frozen after pretraining, CAPE is limited by its restricted generalization and transferability for newly-defined keypoints (Tab. 4). + +Therefore, in this paper, we propose to solve the new keypoint estimation problem creatively in an incremental learning (IL) manner. As shown in Fig. 2, we propose an Incremental object Keypoint Learning (IKL) paradigm to incrementally train a keypoint estimator on the new data labeled only with new keypoints. Even without retaining any old data, the IKL model should not forget the old keypoints catastrophically. Compared to the naive solutions that either label both new and old keypoints or learn separate estimators, our IKL only labels the new keypoints and maintains just one model in the lifetime, which is computationally efficient. Moreover, different from the UKL and CAPE whose performances are largely limited by their restricted transferability depending on a fixed pretrained model, the IKL alleviates such an issue by continually training the model to sufficiently expand its knowledge on newly-defined keypoints and may even improve the old keypoint estimation when learning more relevant new keypoints. Lastly, as most works in IL mainly study the classification task [13, 70], to our best knowledge, IKL as an incremental keypoint regression task has rarely been stud + +ied before, which also provides novel insights for IL. + +However, IKL also poses a new challenge known as label non-co-occurrence (LNCO). As only new keypoints are labeled in the new data, the kinematic and anatomical constraints between old and new keypoints are not explicitly presented in the label space, making it hard for the model to capture such a physical prior during incremental training. Existing IL methods (Tab. 1) do not explicitly model such inter-keypoint relations in IKL, thus the old and new keypoint predictions can not mutually support each other and may readily bias toward the new ones, exacerbating the forgetting of the old keypoints (Tab. 1). + +To tackle the challenge of IKL, we propose a novel two-stage learning scheme called KAMP as a new baseline tailored to IKL. In the first Knowledge Association stage, we train an auxiliary KA-Net to associate old and new keypoints based on their physiological connection, represented by their spatial adjacency. Specifically, KA-Net learns to predict the selected old keypoint given the related new keypoints to acquire their intrinsic anatomical relevance. In the second Mutual Promotion stage, we train the new model on all keypoints to mutually promote their estimation, where the old ones are updated by distilling from the KA-Net and the old model with a newly designed keypoint-oriented spatial distillation loss for better knowledge preservation. To verify our effectiveness, we simulate IKL based on four keypoints datasets on both medical and natural images, i.e., Cephalometric [6], Chest [27], MPII [5] and ATRW [37]. Extensive results demonstrate that our KAMP can not just effectively alleviate the catastrophic forgetting of old keypoints, but even further boost their performance and thus outperform the existing exemplar-free IL methods by a significant margin (Tab. 1) and can also work for low-shot scenarios (Tab. 3 and 4). Our analysis further reveals that IKL is complementary to learning paradigms like CAPE (Tab. 4) and labeling-efficient for new keypoint estimation. + +To summarize, our contributions are three-fold: (1) We establish a novel paradigm, Incremental object Keypoint Learning (IKL), to tackle the challenging demands of new keypoints estimation. (2) We propose a two-stage exemplar-free IKL method to explicitly model and exploit the relation between old and new keypoints to help alleviate the label non-co-occurrence problem tailored to IKL, which may further boost the performance of old keypoints beyond antiforgetting. (3) As a proof-of-concept, we empirically show that IKL is practical and labeling-efficient to scale up a pretrained keypoint estimator on new keypoints estimation, which is much better than other alternative paradigms. + +# 2. Related Works + +Keypoint Estimation (KE). Existing research of KE focuses on estimating a fixed number of keypoints for a specific category [5, 29, 44, 54, 78-80]. Most works design new methods [20, 34, 35, 38, 89] or network architectures [39, 48, 61, 62, 68, 77] to improve the supervised learning over large-scale pure 2D keypoint datasets [5, 44]. Estimating dense keypoints [29, 82] of the human body has been proposed but requires extreme labeling cost. Recent works explore semi-supervised [24, 50, 66] or unsupervised learning [63, 84] to reduce the laborious labeling consumption, but the unsupervised methods still hardly achieve a precise estimation of semantically meaningful keypoints [49]. Recently, Category-agnostic keypoint estimation (CAPE) [10, 56, 76] is proposed to learn a class-agnostic model to detect keypoints of unseen categories by a few labeled support images without retraining the model. However, as the model is frozen after pretraining, CAPE is limited by restricted generalization and transferability due to the intra-class appearance variation, self-occlusion, and appearance ambiguity, as detailed in [76]. Different from the above paradigm that pre-defined a set of keypoints or assumed the pretrained model was sufficient enough for different downstream tasks, our proposed IKL paradigm continuously updates the model only on the new keypoints to increasingly expand its knowledge, which is much more flexible and labeling-efficient. + +Incremental Learning (IL). Existing IL methods can be categorized as exemplar-free and exemplar-based ones [3, 25, 26, 32, 41, 52], given whether old data can be retained. However, saving exemplars not just introduces increasing costs on memory, but also has privacy issues in real-world applications. For exemplar-free methods, existing approaches focus on adding regularization over parameters [2, 33, 47, 81, 83] or the network's output in different positions [15, 40, 73, 85]. Regarding the IL settings, adding new classes [13] or domains [41, 46, 65] have been studied popularly and mainly tested on classification benchmarks [13, 14]. In contrast, keypoint estimation is a location regression task [20, 34, 48], and our proposed IKL aims + +to detect newly defined keypoints incrementally. The incremental animal pose estimation [51] (IAPE) is a related setting to us. However, IAPE assumes the label space of keypoints is fixed for each animal, and it only learns new animals' poses, i.e., only the input distribution is changed with new animal domains. While for our IKL, we instead consider adding new keypoints to the label space for an object category, and the old and new data distribution may be relatively correlated. + +Positive Transfer (PT) in IL. Most IL approaches focus on alleviating the catastrophic forgetting problem [13], while only a few works are dedicated to achieving positive transfer (PT) in IL [30, 43, 47], i.e., improving the old tasks while learning a new task. Those methods encourage PT by modifying the gradient updates based on the new and old tasks' correlations which are measured by saving the data [47] from the old task or based on formal analyses [30, 43]. Differently, our KAMP achieves PT by explicitly modeling the relation between old and new keypoints based on their intrinsic anatomical relevance, and we exploits the relation to design novel distillation to further boost the estimation of the old keypoint, which is tailored to the IKL and keypoint estimation. + +# 3. Incremental Object Keypoint Learning + +# 3.1. Problem Formulation + +We denote $t = 0$ as the initial step and $t = 1, \dots, T$ as the incremental steps. Then the training set for the $t$ -th incremental step is $D_{t}^{\mathrm{train}} = \left\{x_{t,j}^{\mathrm{train}}, y_{t,j}^{\mathrm{train}}\right\}_{j=1}^{N_{t}}$ , where $x, y, N_{t}$ denote the inputs, the keypoint labels and the size of $D_{t}^{\mathrm{train}}$ . As we only label new keypoints in IKL, thus the newly introduced keypoints must be disjointed with the old ones, i.e., $y_{t}^{\mathrm{train}} \cap \left(\cup_{i=0}^{t-1} y_{i}^{\mathrm{train}}\right) = \emptyset$ . Different from the classification task, the common practice for keypoint estimation is to regress a 2D heatmap for a keypoint, i.e., a Gaussian peak around the ground-truth keypoint location [71]. Thus the model $m_{t}$ in IKL comprises the feature extractor $f_{t}$ and a stack of convolutional keypoint estimation heads $\{G_{i}\}_{i=1}^{t}$ that parameterized by $\theta_{t}$ and $\{\phi_{i}\}_{i=1}^{t}$ , respectively. Each estimation head ends with a convolutional layer to generate the 2D heatmap. $G_{t}$ denotes the estimation head for the new keypoints in the $t$ -th step. $f_{t}$ and $\{G_{i}\}_{i=1}^{t-1}$ in model $m_{t}$ are initialized by the same weights in the old model $m_{t-1}$ . + +# 3.2. KAMP: A Novel Baseline for IKL + +In this section, we present our two-stage method KAMP as a novel baseline tailored to IKL. We conjecture that distilling the relation between related old and new keypoints into the model may help it implicitly capture the intrinsic physical prior to alleviate the label non-co-occurrence issue. To do so, for each incremental step, in Stage-I, we design an auxiliary prediction task to associate the related + +![](images/f7f48df20e868dfced7c09b3499eecef0acd55c0be93d4ff6dcbcfa97def3a39.jpg) +Figure 3. Overview of KAMP using the human body for illustration. In Stage-I, we learn an auxiliary KA-Net to associate the related old and new keypoints based on their local anatomical constraint. In Stage-II, we jointly leverage the old model and the KA-Net as an auxiliary teacher to consolidate all old keypoints' prediction and also learn the new keypoints simultaneously to achieve mutual promotions. + +![](images/a549ac296f9aa6e810114de445b96bbc40672e9f07549f4fec5dac317a0c04b5.jpg) + +old keypoint's with the new ones by the KA-Net. In Stage-II, the old model $m_{t - 1}$ and KA-Net are both frozen to distill knowledge into the new model $m_t$ for all old keypoints, and the new keypoints are learned concurrently to achieve mutual promotion. Fig. 3 provides a schematic view of KAMP. + +# 3.2.1. Stage-I: Knowledge Association (KA) + +As mentioned in Sec. 1, the old and new keypoints in IKL are not jointly labeled, making it difficult for the model to learn relationships between keypoints, such as their structural and anatomical relevance, using only the labels of new keypoints. To mitigate this issue, we propose to model the constraint among the spatially and anatomically related old and new keypoints as an implicit function in the current incremental step $t$ . Existing analyses [11, 12, 64] show that keypoints association can be well-modeled by the triangulation constraint. Motivated by this, we define a triangulation constraint $F$ on the distribution of three related keypoints: + +$$ +F \left(P \left(K _ {i}\right), P \left(K _ {j}\right), P \left(K _ {k}\right)\right) = 0, \tag {1} +$$ + +where $P(K_{i})$ denotes the distribution of the keypoint $i$ , and the tuple may include one or two new keypoints. However, as there may be multiple valid solutions (constraints) that satisfy the implicit function $F$ , it is challenging to derive an analytical solution for $F$ in IKL. + +Therefore, we consider an alternative formulation, e.g., $P(K_{\mathrm{o}_j}) = F(P(K_{\mathrm{n}_1}),P(K_{\mathrm{n}_2}))$ , which is a special case of the Eqn. 1 that we condition the distribution of an old keypoint $P(K_{\mathrm{o}_j})$ on its related new keypoints $P(K_{\mathrm{n}_1})$ and $P(K_{\mathrm{n}_2})$ . We focus on two new keypoints since only they have ground-truth labels, and the ground-truth supervision can help reduce the uncertainty in learning the constraints. Moreover, in practice, new tasks typically require more than one new keypoint for support. Nonetheless, we verify in our Supplementary that our approach performs effectively even with a single new keypoint. To enhance this constraint, we model the conditional probability between the related old and new keypoints as an auxiliary prediction task: + +$$ +P \left(K _ {\mathrm {o} j} \mid K _ {\mathrm {n} 1}, K _ {\mathrm {n} 2}, v\right) = \mathrm {K A} \left(P \left(K _ {\mathrm {n} 1}\right), P \left(K _ {\mathrm {n} 2}\right), v\right), \tag {2} +$$ + +where KA stands for the Knowledge Association Network (KA-Net). The insight of this auxiliary task is inspired by the fact that spatially adjacent and visually correlated keypoints can mutually predict each other as shown in many existing works [11, 62, 64]. + +KA-Net. The input of the KA-Net is the ground-truth heatmap $(y_{i,\mathrm{n}_1}^{gt}$ and $y_{i,\mathrm{n}_2}^{gt})$ of new keypoints $K_{\mathrm{n_1}}$ and $K_{\mathrm{n_2}}$ of an image $i$ , and the output is the predicted heatmap $\tilde{y}_{i,j}$ of the old keypoint $K_{\mathrm{o_j}}$ . $v_{i}$ denotes the holistic visual features $v$ of the image $i$ comprised of the intermediate features extracted from the frozen feature $f_{t - 1}$ and are re-scaled to the same spatial size as $y_{i,\mathrm{n_1}}^{gt}$ and $y_{i,\mathrm{n_2}}^{gt}$ . The visual features $v_{i}$ are incorporated alongside spatial information ( $y_{i,\mathrm{n_1}}^{gt}$ and $y_{i,\mathrm{n_2}}^{gt}$ ) because, in a 2D image, spatial coordinates alone are insufficient to determine a keypoint's location due to uncertainty stemming from factors such as camera angle, object appearance, and motion. By incorporating visual features, we provide additional contextual information that helps account for this uncertainty in keypoint associations. + +In the present paper, we explore KA-Net's simplest construction to minimize its training cost: for each new keypoint, we first perform the element-wise multiplication between its ground-truth heatmap $(y_{i,\mathrm{n}_1}^{gt}$ and $y_{i,\mathrm{n}_2}^{gt})$ and the holistic features $v_{i}$ to obtain the keypoint-oriented spatial features. Then we concatenate the spatial features and feed them forward over three convolutional layers accompanied by the Batch Normalization (BN) and ReLU to predict the selected old keypoint $K_{\mathrm{o}_j}$ . Note that the trained KA-Net will only be used to distill the keypoint association in the Stage-II and will not be used in test time after IKL. + +Training of KA-Net. We train KA-Net by the new data $D_{t}^{\mathrm{train}}$ . As $D_{t}^{\mathrm{train}}$ does not have ground truth labels for old keypoints, we use the pseudo-labels predicted by the old model $m_{t - 1}$ to supervise KA-Net for keypoint regression: + +$$ +\ell_ {K A} = \frac {1}{N _ {t}} \sum_ {i} ^ {N t} \sum_ {j \in \mathcal {K} _ {K A}} \left\| \hat {y} _ {i, j} ^ {t - 1} - \tilde {y} _ {i, j} \right\| _ {2} ^ {2}, \tag {3} +$$ + +where $\mathcal{K}_{KA}$ denotes the selected old keypoints used for the auxiliary prediction task, $\hat{y}_{i,j}^{t-1}$ denotes the prediction of the $j$ -th keypoint by $m_{t-1}$ given image $i$ , and $\tilde{y}_{i,j}$ denotes the prediction by the KA-Net. We will show in Sec. 4 that our KAMP can effectively reduce the forgetting of old keypoints and even improve them, which largely reduce the accumulative error to use pseudo-labels for KA-Net. + +Auxiliary Task Creation. To select the old keypoint for training KA-Net, in each incremental step, we first locate all the old and new keypoints on a general object anatomy diagram by their semantic definition. Then we measure their relative proximity by simple distance calculations (e.g., Euclidean distance) to identify two new keypoints $(K_{\mathrm{n_1}}$ and $K_{\mathrm{n_2}})$ that are close to an old one $K_{\mathrm{o_j}}$ . This process can also be automated by multi-modality large language models like GPT-4o [1, 42, 75] as shown in Fig. 3. If we identify several tuples of old and new keypoints satisfied the requirement, we randomly choose one tuple to create the auxiliary prediction task to avoid training too many KA-Net. For all results of our KAMP in Section 4, we show that even by creating only one auxiliary task in each step, we can still bring sufficient improvement on all keypoints. Note that task creation does not need any training, and we only need to perform once before each step, which is highly efficient. Such a design provides an interpretable way to incorporate the physical knowledge of the object to guide the IKL. + +# 3.2.2. Stage-II: Mutual Promotion (MP) + +In Stage-II, we jointly optimize all the new and old keypoints on the new model $m_t$ by the loss $l_{MP}$ : + +$$ +l _ {M P} = \ell_ {G T} + \alpha \left(\ell_ {K S D} + \ell_ {K A}\right), \tag {4} +$$ + +where $\ell_{GT}$ denotes the L2 loss between the ground truth of the new keypoints and their predictions by the new model to acquire the knowledge of new keypoints. $\alpha$ is a hyperparameter to balance the new keypoints acquisition and old keypoints forgetting. For $\ell_{KA}$ , we use the frozen KA-Net as an auxiliary teacher to supervised the old keypoint selected in Stage-I to distill the keypoint association knowledge. Since $\ell_{KA}$ is only applied to the selected old keypoint for knowledge transfer instead of mitigating the forgetting, we further consolidate the knowledge of all old keypoints by the loss $\ell_{KSD}$ using the frozen old model $m_{t - 1}$ . As the distribution of new and old data in IKL do not change dramatically and may even have a strong correlation, LWF [40] can be a baseline for $\ell_{KSD}$ in such a scenario as analyzed in [13]. In general, $\ell_{KSD}$ represents the negative log-likelihood $\frac{1}{N_t}\sum_i^{N_t}s\left(\hat{y}_i^{t - 1}\right)\cdot \log s\left(\hat{y}_i^t\right)$ between the predictions of all old keypoints by the new model, $\hat{y}_i^{t - 1}\in \mathbb{R}^{C\times H\times W}$ , and predictions by the old model, $\hat{y}_i^t\in \mathbb{R}^{C\times H\times W}$ . $s(\cdot)$ denotes the Softmax operator. $C$ denotes the numbers of old keypoints learned before step $t$ , $H$ and $W$ are the height and width of each keypoint's heatmap. + +However, since LWF and its variants all focus on the classification task, they perform the Softmax across different classes by default, i.e., over the $C$ dimension, to obtain the normalized class prediction score. For IKL, this means normalizing across all old keypoints' predictions, which can not explicitly regularize the discrepancy of each old keypoint between the old and new model. To better preserve keypoint-specific knowledge, we adopt a spatial softmax operation over height $(H)$ and width $(W)$ dimensions of the keypoint prediction, $s_{\mathrm{sp}}^{\mathrm{d}}(\cdot), \mathrm{d} \in \{H, W\}$ , and combine them to encourage spatial-oriented knowledge distillation. We discuss in our Supplementary the difference between each softmax alternative. Thus, our $\ell_{KSD}$ is defined as + +$$ +\frac {1}{N _ {t}} \sum_ {i} ^ {N t} \sum_ {j} ^ {C} \sum_ {\mathrm {d}} - s _ {\mathrm {s p}} ^ {\mathrm {d}} (\hat {y} _ {i, j} ^ {t - 1}) \cdot \log s _ {\mathrm {s p}} ^ {\mathrm {d}} (\hat {y} _ {i, j} ^ {t}). \tag {5} +$$ + +# 4. Experiment + +Datasets. As the IKL has rarely been studied before, there is no specific dataset or benchmark for IKL. Motivated by the real-world application of IKL in medical analysis, we leverage the large-scale Cephalometric keypoint dataset [6] proposed in the 2023 MICCAI Challenge and also the commonly benchmarked Chest dataset [27] to further create the IKL protocols to validate our proposed method. Note that the Cephalometric keypoint dataset [6] differs from the head dataset [67] used in previous works [78, 79] commonly, as [6] has much larger number of keypoints and re-collected from different hospitals to enlarge the variance of images, making it much more challenging than [67]. Thus we term this dataset [6] as Head-2023 for clarity. Note that both Head-2023 and Chest datasets have less than or equal to 400 images in total. Moreover, though there is not a large need for IKL on human and animal, the widely used human and animal keypoint datasets, i.e., MPII [5] and ATRW [37], have extreme variations and non-rigid motions. Thus we also choose them for experiments to show the generality of our KAMP under different domains and challenging scenarios. We detail full dataset statistics in our Supplementary. + +Compared Methods. For keypoint estimation, common network structures for classification like ResNet [21] need specific modifications, e.g., adding the deconvolution layers and convolution layers to generate the keypoint location heatmap. These adjustments make many IL methods inapplicable for IKL, e.g., prototype-based methods [87, 88] like PASS [88] that leverages the feature mean prototype before the linear classification layer. Furthermore, the IKL requires that the old data can not be retained, and it is hard to identify specific methods for IKL. Thus we choose several general and representative exemplar-free methods that can adapt to different IL settings based on their methodologies, including EWC [33], LWF [40], MAS [72], RW [9], AFEC [69], CPR [8] for comparison. They are regularization-based methods and + +
MethodSplit ChestSplit Head-2023Split MPIISplit ATRW
A-MRE1↓AT1↑MT1↑A-MRE4↓AT4↑MT4↑AAA4↑AT4↑MT4↑AAA3↑AT3↑MT3↑
Joint Training5.43--2.12--88.50--94.69--
Finetune43.1--51.3--37.41--13.24--
EWC [33]13.28-8.23-3.6710.97-6.37-4.7638.64-51.84-12.2114.38-59.75-2.08
RW [9]9.48-7.12-4.156.49-4.23-2.8838.47-18.83-7.1384.15-10.870.00
MAS [4]7.36-1.86-0.195.31-2.15-1.3367.03-7.560.3485.68-5.80-1.13
LWF [40]6.35-1.340.184.31-1.260.5775.75-3.860.4187.31-5.10-0.64
AFEC [69]8.04-2.670.155.77-3.45-1.4637.24-22.85-15.4233.03-40.25-8.02
CPR [8]6.17-0.870.293.71-1.180.1675.52-3.240.7589.34-2.764.49
SFD [17]7.68-0.540.134.76-0.430.0271.49-0.930.2186.11-1.130.41
WF [74]7.31-0.310.164.580.030.1172.87-0.460.3886.69-0.970.62
GBD [16]6.420.060.214.340.120.4775.62-0.180.3587.42-0.890.65
KAMP (Ours)5.670.290.622.320.410.8479.931.804.2393.16-0.845.13
+ +Table 1. Result of 4 datasets after 2-Step, 5-Step, 5-Step and 4-Step IKL for Chest, Head-2023, MPII, and ATRW respectively. All comparison methods are started from the same Step-0 trained model. A-MRE: smaller the better; AAA/AT/MT: larger the better. + +can be easily applied to the IKL without trial and error. We further adapt methods from Continual Semantic Segmentation (CSS) and Class Incremental Learning (CIL) to ISL for more comparisons, i.e., spatial feature distillation (SFD) [17], weight fusion (WF) [74], and gradient balanced distillation (GBD) [16]. We also report the native baseline, i.e., directly finetune the model during IKL, and the upper bound that trains the model (e.g., HRNet [60]) with all the data, where we denote the former as "Finetune" and the latter as "Joint Training". + +Evaluation Metrics. To assess the keypoint regression task, we employ the widely used mean radial error (MRE) for Head-2023 and Chest datasets as in [78, 79], and Probability of Correct Keypoint (PCK) [50, 56, 66, 77] for MPII and ATRW and use their defaulted $\sigma$ as in [50, 60] to compute PCK. To measure the performance of incremental learning (IL), we use Average Accuracy (AAA), calculating the accuracy (\%) across all keypoints post-incremental step $i$ under PCK, and extend this approach to MRE, denoting it as A-MRE. Additionally, we introduce two metrics for knowledge transfer in IL: (1) Average Transfer (AT), also known as backward transfer [47], which averages the performance improvement of keypoints over all previous steps after step $i$ , and (2) Maximal Transfer (MT), measuring the largest performance improvement in any old keypoint post-step $i$ . Notably, when calculating AT and MT with MRE, we invert the sign of the change in error. This adjustment ensures consistency, since a decrease in MRE signifies improvement, in contrast to the direct correlation of increased accuracy with improvement in the PCK metric. Detailed explanations for the calculation of each metric are provided in our supplementary materials. + +Experimental Design. We randomly split the keypoints of Chest, Head-2023, MPII, and ATRW into different incremental steps to create the Spilt Head-2023, Chest, MPII and ATRW protocols respectively. As the Chest dataset only has 6 keypoints in total, we split them into two groups where the first group has 4 keypoints with 150 training images, and the second group for IKL has 2 keypoints with only 50 training images. For Split Head-2023, we split the 38 keypoints into + +5 groups, where the first group has 19 keypoints with the same definition as [67] with 100 training images, and we split the rest of keypoint into 4 groups randomly to simulate four incremental steps with only 50 training images for each step. Similarly, the 16 MPII keypoints and 15 ATRW keypoints were split into 5 and 4 groups, respectively, with an equal distribution of training images per step. We repeat five times under different orders of keypoint groups and report the mean value in the main paper. The standard deviation (std), details about the keypoints group, and results of other experimental setups are in our Supplementary. + +Implementation Details. We use the HRNet [60] as the backbone for all methods as it is widely compared in keypoint estimation [19, 36, 39, 50, 66, 79]. For a fair comparison, we initialize all methods with the same initial step (Step-0) model for incremental learning. As the first benchmark for the IKL, we use the Continual Hyperparameter Framework (CHF) [13, 41] to identify the training parameters like training epoch (100), initial learning rate (2e-3 for Split Chest and Head-2023, 1e-2 for Split MPII, 1e-3 for Split-ATRW), momentum (0.9), weight decay (1e-4), and also the hyperparameters of each compared methods, all included in our Supplementary. For our KAMP, the hyperparameter $\alpha$ is set as 1e2 for the Split Head-2023 and Split-MPII and 1e4 for the Split Chest and Split-ATRW in all experiments. The analysis of the $\alpha$ and other implementation details are included in our Supplementary. + +# 4.1. Comparison with SOTA for IKL + +As shown in Tab. 1, our KAMP consistently improves accuracy and reduces error across all old and new keypoints after each incremental step, achieving the highest performance compared to all other methods on all datasets. For example, KAMP outperforms the second-best method by $1.39\%$ in $\mathrm{A - MRE_4}$ after four incremental steps on Split Head-2023, and by $4.18\%$ in $\mathrm{AAA_4}$ on Split MPII. Moreover, KAMP is the only method to consistently yield the highest average transfer (AT) and maximal transfer (MT) scores across all datasets. This result highlights that our two-stage learning scheme better facilitates knowledge transfer, thereby en + +hancing accuracy on old keypoints while learning new ones. When AT is positive, this indicates no forgetting on average, underscoring the effectiveness of using pseudo-labels from the old model to train KA-Net, as described in Sec. 3.2.1. In Split-ATRW, the AT is negative for all methods, indicating that forgetting of old keypoints outweighs knowledge transfer. However, KAMP still achieves the highest AT among the methods. Notably, the absolute value of KAMP's AT is nearly zero, suggesting minimal negative transfer. Additionally, by examining the maximal transfer metric, we observe a significant positive transfer in specific cases, such as $5.13\%$ for $\mathrm{MT}_3$ , indicating that some old keypoints experience a substantial positive transfer in each incremental step. This positive transfer helps offset the forgetting of other old keypoints, resulting in a minimal negative average transfer. We further validate this effect by reporting per-keypoint performance in our Supplementary. Qualitative results in Fig. 4 further verify our superiority in achieving more structurally correct prediction than competitive methods. + +Lastly, we observe that methods adapted from CSS and CIL, i.e., SFD, WF, and GBD, achieve strong performance on AT and MT metrics compared to competitive methods like LWF and CPR, but perform poorly on overall metrics like A-MRE and AAA. This is because these methods prioritize reducing forgetting, resulting in an overly rigid model that struggles to effectively incorporate new keypoints in IKL. This observation highlights a crucial limitation of existing CSS and CIL methods that they focus heavily on regularizing forgetting but fail to balance this with knowledge acquisition for new keypoints, which is especially problematic in our newly defined IKL setting. Consequently, this emphasizes the need to study IKL as a distinct incremental learning scenario in this paper, as traditional incremental learning methods are too general to excel here. Our proposed novel baseline, KAMP, effectively addresses this gap for IKL, offering a well-balanced approach that goes beyond anti-forgetting to deliver robust performance across both accuracy and transfer metrics. + +# 4.2. Ablation Study + +To verify the effectiveness of our method, we further analyze how each component may influence the result. Since our method contains two-stage training and we consider a keypoint-oriented spatial distillation loss $(\ell_{KSD})$ , thus we compared our method with three alternatives: (1) the competitive baseline, i.e., the LWF, which performs the Softmax operation across old keypoints. (2) only use our $\ell_{KSD}$ to train the model without the KA-Net. (3) Constructing the KA-Net randomly without using physical knowledge. Results are shown in Tab. 2, and we can observe that: (1) our adapted $\ell_{KSD}$ can effectively outperform the LWF with $1.18\%$ , which demonstrates the essence of our keypoint-oriented adaptation. (2) With proper auxiliary task creation + +
MethodAAA4↑AT4↑MT4↑
LWF [40]75.75-3.860.41
KAMP (only ℓKSD)76.93-2.240.65
KAMP (Random KA-Net)77.13-0.481.24
KAMP (Ours)79.931.804.23
+ +based on physical prior, we can achieve more positive average transfer than other counterparts and finally achieve the largest improvement over all the keypoints on $\mathrm{AAA}_4$ . This shows that our proposed two-stage learning scheme for IKL can not just provide better knowledge consolidation on the old keypoints than the competitive baseline, but the auxiliary prediction task can also bring large improvement to the old keypoints, showing that our KAMP will be a novel and strong baseline for the IKL paradigm. More ablation studies like using backbones other than HRNet and more datasets are in our Supplementary. + +# 4.3. Compare IKL to other low-data paradigms + +Now we provide more insights into our IKL by comparing it with other learning paradigms with low-shot data. Discussion of limitations are in our Supplementary. + +As mentioned before, existing learning paradigms like UKL [22, 23, 78] and CAPE [10, 56, 76] leverage large-scale self-supervised and multi-dataset training to enable novel keypoint detection with a few labeled samples in test time. To fairly compare with UKL and CAPE, we extend KAMP to low-data regime following EGT [78], where we also pretrain an auxiliary self-supervised model at Step-0 to provide pseudo-labels for old keypoints during IKL, and the details are in the Supplementary. For medical datasets like Head-2023, we compare to the SOTA one-shot methods CC2D [78] and EGT [79]. For 2D datasets like MPII, we compare to the SOTA Autolink [22] and MetaPoint+ [10] with official implementation available. Note that methods like [86] require extra expensive 3D CAD models to identify keypoint proposals for novel keypoint detection and are inferior to the SOTA [10, 56], thus we only compare to the MetaPoint+ [10]. For KAMP, we use the low-data adaptation for 1, 5, and 10 shot and do not use it for 50-shot since 50-shot is large enough for us to bypass the adaptation. Full experimental details are in our Supplementary. + +Table 2. Ablation Study on Split MPII. + +
Method/(MRE ↓)1-shot5-shot10-shot50-shot
CC2D [78]5.144.834.083.47
EGT [79]5.014.583.873.21
KAMP (Ours)4.353.703.032.32
+ +Table 3. Compare to low-shot methods on Split Head-2023. + +From Tab. 3 we observe that even on the extreme 1-shot setting, KAMP still outperforms the CC2D and EGT, while + +![](images/4a26e20ec19212a20f8a594ec6b9c5df4390b1e52d27f7db6e2b1890108d71f5.jpg) +Figure 4. Qualitative results on Split Head-2023 and MPII. All methods start from the same Step-0 model, whose prediction is in the second column. GT: ground truth. The red circles denote the keypoints learned in Step 0, and the green ones denote all the new keypoints learned in later incremental steps. We observe that after IKL, the compared methods (LWF and CPR) may acquire the new keypoints as ours, but they have obvious miss-detection and wrong estimation (e.g., out of the body). While our method can consistently associate the new and old keypoints and achieve structurally accurate keypoint predictions. More results are included in our Supplementary. + +all the compared methods do not have good performances. This is because the Head-2023 dataset is collected from multiple sources with large discrepancy, and only one image is hard to represent all the variations of the data. However, when more annotations are available, KAMP can scale much better than CC2D and EGT since our two-stage learning scheme can capture the relation of old and new keypoints and distill them back to the model to improve the old keypoints, which has not been handled in CC2D and EGT. + +
Method/(PCK [10] ↑)1-shot5-shot10-shot50-shot
UKL [22]54.9559.1162.5364.36
MetaPoint+ [10]65.7166.8767.2668.98
KAMP (Ours)70.0972.2373.1876.97
KAMP (Ours) + [10]73.4975.1077.7679.18
+ +Table 4. Compare to low-shot methods on Split MPII. + +As shown in Tab. 4, we observe that while the CAPE [10, 56, 76] and IKL are two disentangled paradigms, they can complement each other effectively. CAPE relies on pretraining with varied datasets, while IKL focuses on incremental learning, making their training costs incomparable. For example, CAPE methods like MetaPoint+ [10] require a substantial training cost (4 GPUs, 1.5 days), whereas the 'pretrain+incremental' approach of IKL is more efficient (1 GPU, 5 hours). To demonstrate that CAPE and IKL are orthogonal and complementary, we applied IKL using our KAMP method on MetaPoint+. As shown in Tab. 4, this combined approach results in significant improvements, even in low-data scenarios, highlighting that IKL can substantially enhance keypoint estimation when applied on a CAPE-pretrained keypoint detector. + +The goal of comparing these paradigms for novel key-point estimation within the same object category is to es + +tabish the necessity of IKL. Although CAPE does not involve continual learning and appears more deploy-friendly, its performance plateaus and does not scale effectively with additional data. In contrast, our KAMP achieves higher accuracy with limited data and demonstrates superior scalability with increased data, as evidenced in Tab. 4. + +Lastly, as the first study to explore the IKL paradigm, we focus on scaling up existing keypoint estimators for novel keypoints within the same object category, laying the foundation for future research. While CAPE can perform keypoint estimation on novel object categories with a few support images, its generalization remains limited. In the future, we will explore extending the IKL paradigm to novel object categories by applying IKL to keypoint estimators pretrained with CAPE, aiming to combine the strengths of both paradigms to develop a keypoint estimator capable of continually learning new keypoints across diverse object categories without forgetting and with effective scalability. + +In summary, the comparison of Tab. 3 and 4 show that our proposed IKL paradigm is highly label-efficient for acquiring new keypoints. This characteristic is advantageous for real-world applications where obtaining sufficient labels is time-consuming, such as in medical analysis [6]. + +# 5. Conclusion + +We explore learning the newly defined keypoint incrementally without retaining any old data, called Incremental object Keypoint Learning (IKL). We propose a two-stage learning scheme as a novel baseline tailored to the IKL. Extensive experiments show that our method can effectively alleviate the forgetting issue and may even improve the old keypoints' estimation during IKL. Our further analysis reveals that the IKL is label efficient in acquiring the new keypoints, which is promising for real-world applications. + +Acknowledgements This work was supported in part by National Science Foundation grant IIS-2007613. This work was also supported by the National Natural Science Foundation of China (62376011). + +# References + +[1] Hello gpt-4o, howpublished = https://openai.com/index/hello-gpt-4o/.5 +[2] Hongjoon Ahn, Sungmin Cha, Donggyu Lee, and Taesup Moon. Uncertainty-based continual learning with adaptive regularization. In Advances in Neural Information Processing Systems, pages 4392-4402, 2019. 3 +[3] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3 +[4] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 6 +[5] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 1, 2, 3, 5 +[6] Jun Cao, Juan Dai, Xuguang Li, Bingsheng Huang, Ching-Wei Wang, and Hongyuan Zhang. Cephalometric landmark detection in lateral x-ray images, 2023. 1, 2, 5, 8 +[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1 +[8] Sungmin Cha, Hsiang Hsu, Taebaek Hwang, Flavio Calmon, and Taesup Moon. {CPR}: Classifier-projection regularization for continual learning. In International Conference on Learning Representations, 2021. 5, 6 +[9] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajanthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European Conference on Computer Vision (ECCV), pages 532-547, 2018. 5, 6 +[10] Junjie Chen, Jiebin Yan, Yuming Fang, and Li Niu. Metapoint learning and refining for category-agnostic pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23534–23543, 2024. 3, 7, 8 +[11] Xiao Chu, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Structured feature learning for pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4715-4723, 2016. 4 +[12] Shengyang Dai, Ming Yang, Ying Wu, and Aggelos Katsaggelos. Detector ensemble. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 1, 4 + +[13] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 3, 5, 6 +[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 3 +[15] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5138-5146, 2019. 3 +[16] Jiahua Dong, Wenqi Liang, Yang Cong, and Gan Sun. Heterogeneous forgetting compensation for class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11742-11751, 2023. 6 +[17] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4040-4050, 2021. 6 +[18] Ana Paula Reis Durão, Aline Morosolli, Pisha Pittayapat, Napat Bolstad, Afonso P Ferreira, and Reinhilde Jacobs. Cephalometric landmark variability among orthodontists and dentomaxillofacial radiologists: a comparative study. Imaging science in dentistry, 45(4):213-220, 2015. 1 +[19] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14676-14686, 2021. 1, 6 +[20] Kerui Gu, Linlin Yang, and Angela Yao. Removing the bias of integral pose regression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11067-11076, 2021. 3 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5 +[22] Xingzhe He, Bastian Wandt, and Helge Rhodin. Autolink: Self-supervised learning of human skeletons and object outlines by linking keypoints. In Advances in Neural Information Processing Systems, 2022. 1, 7, 8 +[23] Xingzhe He, Gaurav Bharaj, David Ferman, Helge Rhodin, and Pablo Garrido. Few-shot geometry-aware keypoint localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21337-21348, 2023. 1, 7 +[24] Sina Honari, Pavlo Molchanov, Stephen Tyree, Pascal Vincent, Christopher Pal, and Jan Kautz. Improving landmark localization with semi-supervised learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1546-1555, 2018. 3 +[25] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via + +rebalancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 3 +[26] Zhongzhan Huang, Mingfu Liang, Senwei Liang, and Shanshan Zhong. Flat local minima for continual learning on semantic segmentation. In International Conference on Multimedia Modeling, pages 388-400. Springer, 2025. 3 +[27] Stefan Jaeger, Alexandros Karargyris, Sema Candemir, Les Folio, Jenifer Siegelman, Fiona Callaghan, Zhiyun Xue, Kannappan Palaniappan, Rahul K Singh, Sameer Antani, George Thoma, Yi-Xiang Wang, Pu-Xuan Lu, and Clement J McDonald. Automatic tuberculosis screening using chest radiographs. IEEE Trans. Med. Imaging, 33(2):233-245, 2014. 2, 5 +[28] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. arXiv preprint arXiv:2207.13784, 2022. 1 +[29] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In European Conference on Computer Vision, pages 196-214. Springer, 2020. 1, 3 +[30] Ta-Chu Kao, Kristopher Jensen, Gido van de Ven, Alberto Bernacchia, and Guillaume Hennequin. Natural continual learning: success is a journey, not (just) a destination. Advances in Neural Information Processing Systems, 34: 28067-28079, 2021. 3 +[31] Leonid Karlinsky and Shimon Ullman. Using linking features in learning non-parametric part models. In European Conference on Computer Vision, pages 326-339. Springer, 2012. 1 +[32] Chris Dongjoo Kim, Jinseo Jeong, and Gunhee Kim. Imbalanced continual learning with partitioning reservoir sampling. In European Conference on Computer Vision, pages 411-428. Springer, 2020. 3 +[33] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 3, 5, 6 +[34] Jiefeng Li, Siyuan Bian, Ailing Zeng, Can Wang, Bo Pang, Wentao Liu, and Cewu Lu. Human pose regression with residual log-likelihood estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11025–11034, 2021. 3 +[35] Jiefeng Li, Tong Chen, Ruiqi Shi, Yujing Lou, Yong-Lu Li, and Cewu Lu. Localization with sampling-argmax. Advances in Neural Information Processing Systems, 34: 27236-27248, 2021. 3 +[36] Ke Li, Shijie Wang, Xiang Zhang, Yifan Xu, Weijian Xu, and Zhuowen Tu. Pose recognition with cascade transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1944-1953, 2021. 1, 6 +[37] Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, and Weiyao Lin. Atrw: A benchmark for amur tiger re-identification in + +the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2590-2598, 2020. 2, 5 +[38] Yanjie Li, Sen Yang, Shoukui Zhang, Zhicheng Wang, Wankou Yang, Shu-Tao Xia, and Erjin Zhou. Is 2d heatmap representation even necessary for human pose estimation? arXiv preprint arXiv:2107.03332, 2021. 3 +[39] Yanjie Li, Shoukui Zhang, Zhicheng Wang, Sen Yang, Wankou Yang, Shu-Tao Xia, and Erjin Zhou. Tokenpose: Learning keypoint tokens for human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11313-11322, 2021. 1, 3, 6 +[40] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, 2017. 3, 5, 6, 7 +[41] Mingfu Liang, Jiahuan Zhou, Wei Wei, and Ying Wu. Balancing between forgetting and acquisition in incremental subpopulation learning. In European Conference on Computer Vision, pages 364-380. Springer, 2022. 3, 6 +[42] Mingfu Liang, Jong-Chyi Su, Samuel Schulter, Sparsh Garg, Shiyu Zhao, Ying Wu, and Manmohan Chandraker. Aide: An automatic data engine for object detection in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14695-14706, 2024. 5 +[43] Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. Beyond not-forgetting: Continual learning with backward knowledge transfer. In Advances in Neural Information Processing Systems, 2022. 3 +[44] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 2, 3 +[45] Buyu Liu and Vittorio Ferrari. Active learning for human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 4363-4372, 2017. 1 +[46] Vincenzo Lomonaco and Davide Maltoni. Core50: a new dataset and benchmark for continuous object recognition. In Conference on Robot Learning, pages 17-26. PMLR, 2017. 3 +[47] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Advances in Neural Information Processing Systems, pages 6467-6476, 2017. 3, 6 +[48] Zhengxiong Luo, Zhicheng Wang, Yan Huang, Liang Wang, Tieniu Tan, and Erjin Zhou. Rethinking the heatmap regression for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13264-13273, 2021. 3 +[49] Dimitrios Mallis, Enrique Sanchez, Matthew Bell, and Georgios Tzimiropoulos. Unsupervised learning of object landmarks via self-training correspondence. Advances in Neural Information Processing Systems, 33:4709-4720, 2020. 1, 3 +[50] Olga Moskvyak, Frederic Maire, Feras Dayoub, and Mahsa Baktashmotlagh. Semi-supervised keypoint localization. In International Conference on Learning Representations, 2021. 1, 3, 6 + +[51] Gaurav Kumar Nayak, Het Shah, and Anirban Chakraborty. Incremental learning for animal pose estimation using rbf k-dpp. arXiv preprint arXiv:2110.13598, 2021. 3 +[52] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In European Conference on Computer Vision, pages 524-540. Springer, 2020. 3 +[53] Ricketts Robert. Orthodontic Diagnosis and Planning: Their Roles in Preventive and Rehabilitative Dentistry. Denver Colo; Rocky Mountain/Orthodontics, 1982. 1 +[54] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In Proceedings of the IEEE international conference on computer vision workshops, pages 397-403, 2013. 3 +[55] Enrique Sanchez and Georgios Tzimiropoulos. Object landmark discovery through unsupervised adaptation. Advances in Neural Information Processing Systems, 32, 2019. 1 +[56] Min Shi, Zihao Huang, Xianzheng Ma, Xiaowei Hu, and Zhiguo Cao. Matching is not enough: A two-stage framework for category-agnostic pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7308–7317, 2023. 3, 6, 7, 8 +[57] Aliaksandr Siarohin, Enver Sangineto, Stéphane Lathuiliere, and Nicu Sebe. Deformable gans for pose-based human image generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3408-3416, 2018. 1 +[58] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2377-2386, 2019. 1 +[59] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 2 +[60] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693-5703, 2019. 6 +[61] Wei Tang and Ying Wu. Does learning specific features for related parts help human pose estimation? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1107-1116, 2019. 3 +[62] Wei Tang, Pei Yu, and Ying Wu. Deeply learned compositional models for human pose estimation. In Proceedings of the European conference on computer vision (ECCV), pages 190-206, 2018. 3, 4 +[63] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial embeddings. In Proceedings of the IEEE international conference on computer vision, pages 5916-5925, 2017. 3 +[64] Jonathan J. Thompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems, 27, 2014. 4 + +[65] Riccardo Volpi, Diane Larlus, and Grégory Rogez. Continual adaptation of visual representations via domain randomization and meta-learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4443-4453, 2021. 3 +[66] Can Wang, Sheng Jin, Yingda Guan, Wentao Liu, Chen Qian, Ping Luo, and Wanli Ouyang. Pseudo-labeled autocurriculum learning for semi-supervised keypoint localization. In International Conference on Learning Representations, 2022. 3, 6 +[67] Ching-Wei Wang, Cheng-Ta Huang, Jia-Hong Lee, Chung-Hsing Li, Sheng-Wei Chang, Ming-Jhih Siao, Tat-Ming Lai, Bulat Ibragimov, Tomaz Vrtovec, Olaf Ronneberger, Philipp Fischer, Tim F Cootes, and Claudia Lindner. A benchmark for comparison of dental radiography analysis algorithms. Med. Image Anal., 31:63-76, 2016. 1, 5, 6 +[68] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 1, 3 +[69] Liyuan Wang, Mingtian Zhang, Zhongfan Jia, Qian Li, Chenglong Bao, Kaisheng Ma, Jun Zhu, and Yi Zhong. Afec: Active forgetting of negative transfer in continual learning. Advances in Neural Information Processing Systems, 34: 22379-22391, 2021. 5, 6 +[70] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[71] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 4724–4732, 2016. 3 +[72] Chenshen Wu, Luis Herranz, Xialei Liu, Joost van de Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. In Advances in Neural Information Processing Systems, pages 5962-5972, 2018. 5 +[73] Guile Wu, Shaogang Gong, and Pan Li. Striking a balance between stability and plasticity for class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1124-1133, 2021. 3 +[74] Jia-Wen Xiao, Chang-Bin Zhang, Jiekang Feng, Xialei Liu, Joost van de Weijer, and Ming-Ming Cheng. Endpoints weight fusion for class incremental semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7204-7213, 2023. 6 +[75] Xiaoying Xing, Mingfu Liang, and Ying Wu. Toa: Task-oriented active vqa. Advances in Neural Information Processing Systems, 36:54061-54074, 2023. 5 +[76] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, + +Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 398-416. Springer, 2022. 2, 3, 7, 8 +[77] Sen Yang, Zhibin Quan, Mu Nie, and Wankou Yang. Transpose: Keypoint localization via transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11802-11812, 2021. 1, 3, 6 +[78] Qingsong Yao, Quan Quan, Li Xiao, and S Kevin Zhou. One-shot medical landmark detection. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24, pages 177-188. Springer, 2021. 3, 5, 6, 7 +[79] Zihao Yin, Ping Gong, Chunyu Wang, Yizhou Yu, and Yizhou Wang. One-shot medical landmark localization by edge-guided transform and noisy landmark refinement. In European Conference on Computer Vision, pages 473-489. Springer, 2022. 5, 6, 7 +[80] Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. AP-10k: A benchmark for animal pose estimation in the wild. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 1, 3 +[81] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6982-6991, 2020. 3 +[82] Duncan Zauss, Sven Kreiss, and Alexandre Alahi. Keypoint communities. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11057-11066, 2021. 3 +[83] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. Proceedings of Machine Learning Research, 70:3987, 2017. 3 +[84] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2694-2703, 2018. 1, 3 +[85] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13208-13217, 2020. 3 +[86] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018. 7 +[87] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34:14306-14318, 2021. 5 +[88] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 5 + +[89] Xu Zou, Sheng Zhong, Luxin Yan, Xiangyun Zhao, Jiahuan Zhou, and Ying Wu. Learning robust facial landmark detection via hierarchical structured ensemble. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 141-150, 2019. 3 \ No newline at end of file diff --git a/2025/Incremental Object Keypoint Learning/images.zip b/2025/Incremental Object Keypoint Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e451ecc6b4f845d0a76c668674cc6e4c89d75a95 --- /dev/null +++ b/2025/Incremental Object Keypoint Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c53b17aa1f4964a1b79e514c098842cd18c26e2ce935c1f8ac466a8b0858134 +size 466997 diff --git a/2025/Incremental Object Keypoint Learning/layout.json b/2025/Incremental Object Keypoint Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..312f18ca355bb6f8760399a9dbf308a42504036c --- /dev/null +++ b/2025/Incremental Object Keypoint Learning/layout.json @@ -0,0 +1,9080 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 184, + 103, + 427, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 103, + 427, + 121 + ], + "spans": [ + { + "bbox": [ + 184, + 103, + 427, + 121 + ], + "type": "text", + "content": "Incremental Object Keypoint Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 142, + 446, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 142, + 446, + 157 + ], + "spans": [ + { + "bbox": [ + 162, + 142, + 446, + 157 + ], + "type": "text", + "content": "Mingfu Liang" + }, + { + "bbox": [ + 162, + 142, + 446, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 142, + 446, + 157 + ], + "type": "text", + "content": " Jiahuan Zhou" + }, + { + "bbox": [ + 162, + 142, + 446, + 157 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 162, + 142, + 446, + 157 + ], + "type": "text", + "content": " Xu Zou" + }, + { + "bbox": [ + 162, + 142, + 446, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 162, + 142, + 446, + 157 + ], + "type": "text", + "content": " Ying Wu" + }, + { + "bbox": [ + 162, + 142, + 446, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 157, + 534, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 157, + 534, + 186 + ], + "spans": [ + { + "bbox": [ + 75, + 157, + 534, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 157, + 534, + 186 + ], + "type": "text", + "content": " Northwestern University " + }, + { + "bbox": [ + 75, + 157, + 534, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 157, + 534, + 186 + ], + "type": "text", + "content": " Wangxuan Institute of Computer Technology, Peking University " + }, + { + "bbox": [ + 75, + 157, + 534, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 75, + 157, + 534, + 186 + ], + "type": "text", + "content": " Huazhong University of Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 239, + 297, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 622 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 622 + ], + "type": "text", + "content": "Existing progress in object keypoint estimation primarily benefits from the conventional supervised learning paradigm based on numerous data labeled with pre-defined keypoints. However, these well-trained models can hardly detect the undefined new keypoints in test time, which largely hinders their feasibility for diverse downstream tasks. To handle this, various solutions are explored but still suffer from either limited generalizability or transferability. Therefore, in this paper, we explore a novel keypoint learning paradigm in that we only annotate new keypoints in the new data and incrementally train the model, without retaining any old data, called Incremental object Keypoint Learning (IKL). A two-stage learning scheme as a novel baseline tailored to IKL is developed. In the first Knowledge Association stage, given the data labeled with only new keypoints, an auxiliary KA-Net is trained to automatically associate the old keypoints to these new ones based on their spatial and intrinsic anatomical relations. In the second Mutual Promotion stage, based on a keypoint-oriented spatial distillation loss, we jointly leverage the auxiliary KA-Net and the old model for knowledge consolidation to mutually promote the estimation of all old and new keypoints. Owing to the investigation of the correlations between new and old keypoints, our proposed method can not just effectively mitigate the catastrophic forgetting of old keypoints, but may even further improve the estimation of the old ones and achieve a positive transfer beyond antiforgetting. Such an observation has been solidly verified by extensive experiments on different keypoint datasets, where our method exhibits superiority in alleviating the forgetting issue and boosting performance while enjoying labeling efficiency even under the low-shot data regime." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 647, + 135, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 135, + 659 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 135, + 659 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 668, + 295, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 295, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 295, + 692 + ], + "type": "text", + "content": "As a fundamental task in computer vision, estimating the visual keypoint locations of an object serves as the indis-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 342, + 212, + 523, + 305 + ], + "blocks": [ + { + "bbox": [ + 342, + 212, + 523, + 305 + ], + "lines": [ + { + "bbox": [ + 342, + 212, + 523, + 305 + ], + "spans": [ + { + "bbox": [ + 342, + 212, + 523, + 305 + ], + "type": "image", + "image_path": "59d36260c6f937c2d3c3a51708fd622dfb077665f6fefc60bc8f6272a6cc47ed.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 304, + 553, + 326 + ], + "lines": [ + { + "bbox": [ + 313, + 304, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 553, + 326 + ], + "type": "text", + "content": "Figure 1. Medical analysis can frequently change and require new keypoints essentially [6], while the labeling is time-consuming." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 340, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 448 + ], + "type": "text", + "content": "pensable primitive to support numerous down-stream tasks, e.g., object pose detection [29, 77, 80], tracking [28], action recognition [7], generation [57] and animation [28, 58], etc. Over a long period of time, supervised keypoint learning (SKL) [19, 36, 39, 68] has significantly advanced the progress on keypoint estimation. By exploiting large-scale pure 2D keypoint datasets [5, 44] with extreme variance, SKL aims to train a deep neural network that can robustly detect a set of pre-defined keypoints in test images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 664 + ], + "type": "text", + "content": "However, the SKL models can not estimate newly-added undefined keypoints of an object, while new demands from downstream tasks will inevitably require new keypoints. For instance, until now, models trained by existing keypoint dataset [67] can only detect 19 keypoints to support Downs analysis in the Cephalometric analysis, while they are infeasible for important analysis like Steiner analysis [53], as shown in Fig. 1. Such an issue motivated a recent MIC-CAI challenge [6] on annotating much more keypoints to support increasing analysis. However, keypoint labeling is time-consuming [18, 45, 50], especially when keypoints are defined sequentially. Another alternative solution is to train separate new keypoint estimator [12, 31] for each set of new keypoints as in Fig. 2. However, the number of estimators will increase linearly, and the separate training strategy will make each estimator easily overfit the new keypoints [12] and hardly capture the intrinsic relation between old and new keypoints, leading to sub-optimal performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "To mitigate the above issues, unsupervised keypoint learning (UKL) [22, 23, 49, 55, 84] is proposed for new keypoint estimation. The models are firstly pretrained in an unsupervised manner to predict numbers of keypoints ran" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 151, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 151, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 151, + 712 + ], + "type": "text", + "content": "*: Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25399" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 69, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 553, + 228 + ], + "type": "image", + "image_path": "e57ab3b3387667f941eb381fbf19203dfa588a21d2a90fe22206936958567688.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 234, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 555, + 269 + ], + "type": "text", + "content": "Figure 2. Separate Estimators need multiple estimators in test time. Unsupervised keypoint learning (UKL) and category-agnostic pose estimation (CAPE) exploit a fixed pretrained keypoint estimator on unseen new keypoints. While our Incremental object Keypoint Keaning (IKL) continually updates the same model on the new data labeled only on new keypoints without retaining any old training data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 274, + 297, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 274, + 297, + 454 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 297, + 454 + ], + "type": "text", + "content": "domly, then a linear transformation is learned to transfer the existing keypoints to the new undefined ones. However, UKL is merely effective on objects with relatively rigid motion or video data with consistent object instances and local changes in appearance and motion. As a result, it is hard to perform UKL on large-scale pure 2D datasets [5, 44] due to their extreme inter-image differences in the background, appearance, and motions of distinct object instances. Recently, a supervised metric learning-based [59] paradigm named category-agnostic pose estimation (CAPE) [76] is proposed to learn a class-agnostic pose estimator that can estimate unseen keypoints in test time with a few labeled support samples. However, as the model is frozen after pretraining, CAPE is limited by its restricted generalization and transferability for newly-defined keypoints (Tab. 4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 462, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 714 + ], + "type": "text", + "content": "Therefore, in this paper, we propose to solve the new keypoint estimation problem creatively in an incremental learning (IL) manner. As shown in Fig. 2, we propose an Incremental object Keypoint Learning (IKL) paradigm to incrementally train a keypoint estimator on the new data labeled only with new keypoints. Even without retaining any old data, the IKL model should not forget the old keypoints catastrophically. Compared to the naive solutions that either label both new and old keypoints or learn separate estimators, our IKL only labels the new keypoints and maintains just one model in the lifetime, which is computationally efficient. Moreover, different from the UKL and CAPE whose performances are largely limited by their restricted transferability depending on a fixed pretrained model, the IKL alleviates such an issue by continually training the model to sufficiently expand its knowledge on newly-defined keypoints and may even improve the old keypoint estimation when learning more relevant new keypoints. Lastly, as most works in IL mainly study the classification task [13, 70], to our best knowledge, IKL as an incremental keypoint regression task has rarely been stud" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 274, + 528, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 528, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 528, + 286 + ], + "type": "text", + "content": "ied before, which also provides novel insights for IL." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 290, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 422 + ], + "type": "text", + "content": "However, IKL also poses a new challenge known as label non-co-occurrence (LNCO). As only new keypoints are labeled in the new data, the kinematic and anatomical constraints between old and new keypoints are not explicitly presented in the label space, making it hard for the model to capture such a physical prior during incremental training. Existing IL methods (Tab. 1) do not explicitly model such inter-keypoint relations in IKL, thus the old and new keypoint predictions can not mutually support each other and may readily bias toward the new ones, exacerbating the forgetting of the old keypoints (Tab. 1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "To tackle the challenge of IKL, we propose a novel two-stage learning scheme called KAMP as a new baseline tailored to IKL. In the first Knowledge Association stage, we train an auxiliary KA-Net to associate old and new keypoints based on their physiological connection, represented by their spatial adjacency. Specifically, KA-Net learns to predict the selected old keypoint given the related new keypoints to acquire their intrinsic anatomical relevance. In the second Mutual Promotion stage, we train the new model on all keypoints to mutually promote their estimation, where the old ones are updated by distilling from the KA-Net and the old model with a newly designed keypoint-oriented spatial distillation loss for better knowledge preservation. To verify our effectiveness, we simulate IKL based on four keypoints datasets on both medical and natural images, i.e., Cephalometric [6], Chest [27], MPII [5] and ATRW [37]. Extensive results demonstrate that our KAMP can not just effectively alleviate the catastrophic forgetting of old keypoints, but even further boost their performance and thus outperform the existing exemplar-free IL methods by a significant margin (Tab. 1) and can also work for low-shot scenarios (Tab. 3 and 4). Our analysis further reveals that IKL is complementary to learning paradigms like CAPE (Tab. 4) and labeling-efficient for new keypoint estimation." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25400" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "To summarize, our contributions are three-fold: (1) We establish a novel paradigm, Incremental object Keypoint Learning (IKL), to tackle the challenging demands of new keypoints estimation. (2) We propose a two-stage exemplar-free IKL method to explicitly model and exploit the relation between old and new keypoints to help alleviate the label non-co-occurrence problem tailored to IKL, which may further boost the performance of old keypoints beyond antiforgetting. (3) As a proof-of-concept, we empirically show that IKL is practical and labeling-efficient to scale up a pretrained keypoint estimator on new keypoints estimation, which is much better than other alternative paradigms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 227, + 146, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 146, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 146, + 239 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 247, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 247, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 247, + 296, + 558 + ], + "type": "text", + "content": "Keypoint Estimation (KE). Existing research of KE focuses on estimating a fixed number of keypoints for a specific category [5, 29, 44, 54, 78-80]. Most works design new methods [20, 34, 35, 38, 89] or network architectures [39, 48, 61, 62, 68, 77] to improve the supervised learning over large-scale pure 2D keypoint datasets [5, 44]. Estimating dense keypoints [29, 82] of the human body has been proposed but requires extreme labeling cost. Recent works explore semi-supervised [24, 50, 66] or unsupervised learning [63, 84] to reduce the laborious labeling consumption, but the unsupervised methods still hardly achieve a precise estimation of semantically meaningful keypoints [49]. Recently, Category-agnostic keypoint estimation (CAPE) [10, 56, 76] is proposed to learn a class-agnostic model to detect keypoints of unseen categories by a few labeled support images without retraining the model. However, as the model is frozen after pretraining, CAPE is limited by restricted generalization and transferability due to the intra-class appearance variation, self-occlusion, and appearance ambiguity, as detailed in [76]. Different from the above paradigm that pre-defined a set of keypoints or assumed the pretrained model was sufficient enough for different downstream tasks, our proposed IKL paradigm continuously updates the model only on the new keypoints to increasingly expand its knowledge, which is much more flexible and labeling-efficient." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 559, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 713 + ], + "type": "text", + "content": "Incremental Learning (IL). Existing IL methods can be categorized as exemplar-free and exemplar-based ones [3, 25, 26, 32, 41, 52], given whether old data can be retained. However, saving exemplars not just introduces increasing costs on memory, but also has privacy issues in real-world applications. For exemplar-free methods, existing approaches focus on adding regularization over parameters [2, 33, 47, 81, 83] or the network's output in different positions [15, 40, 73, 85]. Regarding the IL settings, adding new classes [13] or domains [41, 46, 65] have been studied popularly and mainly tested on classification benchmarks [13, 14]. In contrast, keypoint estimation is a location regression task [20, 34, 48], and our proposed IKL aims" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "to detect newly defined keypoints incrementally. The incremental animal pose estimation [51] (IAPE) is a related setting to us. However, IAPE assumes the label space of keypoints is fixed for each animal, and it only learns new animals' poses, i.e., only the input distribution is changed with new animal domains. While for our IKL, we instead consider adding new keypoints to the label space for an object category, and the old and new data distribution may be relatively correlated." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 180, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 348 + ], + "type": "text", + "content": "Positive Transfer (PT) in IL. Most IL approaches focus on alleviating the catastrophic forgetting problem [13], while only a few works are dedicated to achieving positive transfer (PT) in IL [30, 43, 47], i.e., improving the old tasks while learning a new task. Those methods encourage PT by modifying the gradient updates based on the new and old tasks' correlations which are measured by saving the data [47] from the old task or based on formal analyses [30, 43]. Differently, our KAMP achieves PT by explicitly modeling the relation between old and new keypoints based on their intrinsic anatomical relevance, and we exploits the relation to design novel distillation to further boost the estimation of the old keypoint, which is tailored to the IKL and keypoint estimation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 358, + 529, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 529, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 529, + 373 + ], + "type": "text", + "content": "3. Incremental Object Keypoint Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 378, + 440, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 440, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 440, + 390 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": "We denote " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": " as the initial step and " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "t = 1, \\dots, T" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": " as the incremental steps. Then the training set for the " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": "-th incremental step is " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "D_{t}^{\\mathrm{train}} = \\left\\{x_{t,j}^{\\mathrm{train}}, y_{t,j}^{\\mathrm{train}}\\right\\}_{j=1}^{N_{t}}" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "x, y, N_{t}" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": " denote the inputs, the keypoint labels and the size of " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "D_{t}^{\\mathrm{train}}" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": ". As we only label new keypoints in IKL, thus the newly introduced keypoints must be disjointed with the old ones, i.e., " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "y_{t}^{\\mathrm{train}} \\cap \\left(\\cup_{i=0}^{t-1} y_{i}^{\\mathrm{train}}\\right) = \\emptyset" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": ". Different from the classification task, the common practice for keypoint estimation is to regress a 2D heatmap for a keypoint, i.e., a Gaussian peak around the ground-truth keypoint location [71]. Thus the model " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "m_{t}" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": " in IKL comprises the feature extractor " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": " and a stack of convolutional keypoint estimation heads " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\{G_{i}\\}_{i=1}^{t}" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": " that parameterized by " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\{\\phi_{i}\\}_{i=1}^{t}" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": ", respectively. Each estimation head ends with a convolutional layer to generate the 2D heatmap. " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": " denotes the estimation head for the new keypoints in the " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": "-th step. " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\{G_{i}\\}_{i=1}^{t-1}" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": " in model " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "m_{t}" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": " are initialized by the same weights in the old model " + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "inline_equation", + "content": "m_{t-1}" + }, + { + "bbox": [ + 313, + 396, + 555, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 612, + 495, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 495, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 495, + 624 + ], + "type": "text", + "content": "3.2. KAMP: A Novel Baseline for IKL" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "In this section, we present our two-stage method KAMP as a novel baseline tailored to IKL. We conjecture that distilling the relation between related old and new keypoints into the model may help it implicitly capture the intrinsic physical prior to alleviate the label non-co-occurrence issue. To do so, for each incremental step, in Stage-I, we design an auxiliary prediction task to associate the related" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25401" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 69, + 376, + 216 + ], + "blocks": [ + { + "bbox": [ + 94, + 69, + 376, + 216 + ], + "lines": [ + { + "bbox": [ + 94, + 69, + 376, + 216 + ], + "spans": [ + { + "bbox": [ + 94, + 69, + 376, + 216 + ], + "type": "image", + "image_path": "f7f48df20e868dfced7c09b3499eecef0acd55c0be93d4ff6dcbcfa97def3a39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 224, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 224, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 555, + 258 + ], + "type": "text", + "content": "Figure 3. Overview of KAMP using the human body for illustration. In Stage-I, we learn an auxiliary KA-Net to associate the related old and new keypoints based on their local anatomical constraint. In Stage-II, we jointly leverage the old model and the KA-Net as an auxiliary teacher to consolidate all old keypoints' prediction and also learn the new keypoints simultaneously to achieve mutual promotions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 385, + 70, + 516, + 216 + ], + "blocks": [ + { + "bbox": [ + 385, + 70, + 516, + 216 + ], + "lines": [ + { + "bbox": [ + 385, + 70, + 516, + 216 + ], + "spans": [ + { + "bbox": [ + 385, + 70, + 516, + 216 + ], + "type": "image", + "image_path": "a549ac296f9aa6e810114de445b96bbc40672e9f07549f4fec5dac317a0c04b5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 265, + 294, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 265, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 54, + 265, + 294, + 324 + ], + "type": "text", + "content": "old keypoint's with the new ones by the KA-Net. In Stage-II, the old model " + }, + { + "bbox": [ + 54, + 265, + 294, + 324 + ], + "type": "inline_equation", + "content": "m_{t - 1}" + }, + { + "bbox": [ + 54, + 265, + 294, + 324 + ], + "type": "text", + "content": " and KA-Net are both frozen to distill knowledge into the new model " + }, + { + "bbox": [ + 54, + 265, + 294, + 324 + ], + "type": "inline_equation", + "content": "m_t" + }, + { + "bbox": [ + 54, + 265, + 294, + 324 + ], + "type": "text", + "content": " for all old keypoints, and the new keypoints are learned concurrently to achieve mutual promotion. Fig. 3 provides a schematic view of KAMP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 330, + 243, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 243, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 243, + 341 + ], + "type": "text", + "content": "3.2.1. Stage-I: Knowledge Association (KA)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 344, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 344, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 54, + 344, + 295, + 475 + ], + "type": "text", + "content": "As mentioned in Sec. 1, the old and new keypoints in IKL are not jointly labeled, making it difficult for the model to learn relationships between keypoints, such as their structural and anatomical relevance, using only the labels of new keypoints. To mitigate this issue, we propose to model the constraint among the spatially and anatomically related old and new keypoints as an implicit function in the current incremental step " + }, + { + "bbox": [ + 54, + 344, + 295, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 344, + 295, + 475 + ], + "type": "text", + "content": ". Existing analyses [11, 12, 64] show that keypoints association can be well-modeled by the triangulation constraint. Motivated by this, we define a triangulation constraint " + }, + { + "bbox": [ + 54, + 344, + 295, + 475 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 54, + 344, + 295, + 475 + ], + "type": "text", + "content": " on the distribution of three related keypoints:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 478, + 294, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 478, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 108, + 478, + 294, + 491 + ], + "type": "interline_equation", + "content": "F \\left(P \\left(K _ {i}\\right), P \\left(K _ {j}\\right), P \\left(K _ {k}\\right)\\right) = 0, \\tag {1}", + "image_path": "2431a5c0b12bbc5dda1a16c7623a285a5a962c246adf7b4015feeff54aa6e343.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 498, + 295, + 557 + ], + "type": "inline_equation", + "content": "P(K_{i})" + }, + { + "bbox": [ + 55, + 498, + 295, + 557 + ], + "type": "text", + "content": " denotes the distribution of the keypoint " + }, + { + "bbox": [ + 55, + 498, + 295, + 557 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 498, + 295, + 557 + ], + "type": "text", + "content": ", and the tuple may include one or two new keypoints. However, as there may be multiple valid solutions (constraints) that satisfy the implicit function " + }, + { + "bbox": [ + 55, + 498, + 295, + 557 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 498, + 295, + 557 + ], + "type": "text", + "content": ", it is challenging to derive an analytical solution for " + }, + { + "bbox": [ + 55, + 498, + 295, + 557 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 498, + 295, + 557 + ], + "type": "text", + "content": " in IKL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "Therefore, we consider an alternative formulation, e.g., " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "P(K_{\\mathrm{o}_j}) = F(P(K_{\\mathrm{n}_1}),P(K_{\\mathrm{n}_2}))" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": ", which is a special case of the Eqn. 1 that we condition the distribution of an old keypoint " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "P(K_{\\mathrm{o}_j})" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " on its related new keypoints " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "P(K_{\\mathrm{n}_1})" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "P(K_{\\mathrm{n}_2})" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": ". We focus on two new keypoints since only they have ground-truth labels, and the ground-truth supervision can help reduce the uncertainty in learning the constraints. Moreover, in practice, new tasks typically require more than one new keypoint for support. Nonetheless, we verify in our Supplementary that our approach performs effectively even with a single new keypoint. To enhance this constraint, we model the conditional probability between the related old and new keypoints as an auxiliary prediction task:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 324, + 264, + 553, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 264, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 324, + 264, + 553, + 277 + ], + "type": "interline_equation", + "content": "P \\left(K _ {\\mathrm {o} j} \\mid K _ {\\mathrm {n} 1}, K _ {\\mathrm {n} 2}, v\\right) = \\mathrm {K A} \\left(P \\left(K _ {\\mathrm {n} 1}\\right), P \\left(K _ {\\mathrm {n} 2}\\right), v\\right), \\tag {2}", + "image_path": "72d0df6754ddc9d0e2ad16ba3592fcd0099ee50d9e81c422ebf7dd665c4d25d0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 281, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 342 + ], + "type": "text", + "content": "where KA stands for the Knowledge Association Network (KA-Net). The insight of this auxiliary task is inspired by the fact that spatially adjacent and visually correlated keypoints can mutually predict each other as shown in many existing works [11, 62, 64]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": "KA-Net. The input of the KA-Net is the ground-truth heatmap " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "(y_{i,\\mathrm{n}_1}^{gt}" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "y_{i,\\mathrm{n}_2}^{gt})" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": " of new keypoints " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{n_1}}" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{n_2}}" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": " of an image " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": ", and the output is the predicted heatmap " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{i,j}" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": " of the old keypoint " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{o_j}}" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": " denotes the holistic visual features " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": " of the image " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": " comprised of the intermediate features extracted from the frozen feature " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "f_{t - 1}" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": " and are re-scaled to the same spatial size as " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "y_{i,\\mathrm{n_1}}^{gt}" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "y_{i,\\mathrm{n_2}}^{gt}" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": ". The visual features " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": " are incorporated alongside spatial information (" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "y_{i,\\mathrm{n_1}}^{gt}" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "inline_equation", + "content": "y_{i,\\mathrm{n_2}}^{gt}" + }, + { + "bbox": [ + 313, + 342, + 555, + 514 + ], + "type": "text", + "content": ") because, in a 2D image, spatial coordinates alone are insufficient to determine a keypoint's location due to uncertainty stemming from factors such as camera angle, object appearance, and motion. By incorporating visual features, we provide additional contextual information that helps account for this uncertainty in keypoint associations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 514, + 555, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 644 + ], + "type": "text", + "content": "In the present paper, we explore KA-Net's simplest construction to minimize its training cost: for each new keypoint, we first perform the element-wise multiplication between its ground-truth heatmap " + }, + { + "bbox": [ + 313, + 514, + 555, + 644 + ], + "type": "inline_equation", + "content": "(y_{i,\\mathrm{n}_1}^{gt}" + }, + { + "bbox": [ + 313, + 514, + 555, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 514, + 555, + 644 + ], + "type": "inline_equation", + "content": "y_{i,\\mathrm{n}_2}^{gt})" + }, + { + "bbox": [ + 313, + 514, + 555, + 644 + ], + "type": "text", + "content": " and the holistic features " + }, + { + "bbox": [ + 313, + 514, + 555, + 644 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 514, + 555, + 644 + ], + "type": "text", + "content": " to obtain the keypoint-oriented spatial features. Then we concatenate the spatial features and feed them forward over three convolutional layers accompanied by the Batch Normalization (BN) and ReLU to predict the selected old keypoint " + }, + { + "bbox": [ + 313, + 514, + 555, + 644 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{o}_j}" + }, + { + "bbox": [ + 313, + 514, + 555, + 644 + ], + "type": "text", + "content": ". Note that the trained KA-Net will only be used to distill the keypoint association in the Stage-II and will not be used in test time after IKL." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 645, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 555, + 693 + ], + "type": "text", + "content": "Training of KA-Net. We train KA-Net by the new data " + }, + { + "bbox": [ + 313, + 645, + 555, + 693 + ], + "type": "inline_equation", + "content": "D_{t}^{\\mathrm{train}}" + }, + { + "bbox": [ + 313, + 645, + 555, + 693 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 313, + 645, + 555, + 693 + ], + "type": "inline_equation", + "content": "D_{t}^{\\mathrm{train}}" + }, + { + "bbox": [ + 313, + 645, + 555, + 693 + ], + "type": "text", + "content": " does not have ground truth labels for old keypoints, we use the pseudo-labels predicted by the old model " + }, + { + "bbox": [ + 313, + 645, + 555, + 693 + ], + "type": "inline_equation", + "content": "m_{t - 1}" + }, + { + "bbox": [ + 313, + 645, + 555, + 693 + ], + "type": "text", + "content": " to supervise KA-Net for keypoint regression:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 341, + 693, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 693, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 341, + 693, + 553, + 717 + ], + "type": "interline_equation", + "content": "\\ell_ {K A} = \\frac {1}{N _ {t}} \\sum_ {i} ^ {N t} \\sum_ {j \\in \\mathcal {K} _ {K A}} \\left\\| \\hat {y} _ {i, j} ^ {t - 1} - \\tilde {y} _ {i, j} \\right\\| _ {2} ^ {2}, \\tag {3}", + "image_path": "1c1b5fb7c083f250c3d54bdfd87425e828d2a31061e53db55272db22407bf000.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25402" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_{KA}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " denotes the selected old keypoints used for the auxiliary prediction task, " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i,j}^{t-1}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " denotes the prediction of the " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "-th keypoint by " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "m_{t-1}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " given image " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{i,j}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " denotes the prediction by the KA-Net. We will show in Sec. 4 that our KAMP can effectively reduce the forgetting of old keypoints and even improve them, which largely reduce the accumulative error to use pseudo-labels for KA-Net." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": "Auxiliary Task Creation. To select the old keypoint for training KA-Net, in each incremental step, we first locate all the old and new keypoints on a general object anatomy diagram by their semantic definition. Then we measure their relative proximity by simple distance calculations (e.g., Euclidean distance) to identify two new keypoints " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "(K_{\\mathrm{n_1}}" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{n_2}})" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": " that are close to an old one " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{o_j}}" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": ". This process can also be automated by multi-modality large language models like GPT-4o [1, 42, 75] as shown in Fig. 3. If we identify several tuples of old and new keypoints satisfied the requirement, we randomly choose one tuple to create the auxiliary prediction task to avoid training too many KA-Net. For all results of our KAMP in Section 4, we show that even by creating only one auxiliary task in each step, we can still bring sufficient improvement on all keypoints. Note that task creation does not need any training, and we only need to perform once before each step, which is highly efficient. Such a design provides an interpretable way to incorporate the physical knowledge of the object to guide the IKL." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 392, + 228, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 228, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 228, + 404 + ], + "type": "text", + "content": "3.2.2. Stage-II: Mutual Promotion (MP)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 408, + 295, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 295, + 433 + ], + "type": "text", + "content": "In Stage-II, we jointly optimize all the new and old keypoints on the new model " + }, + { + "bbox": [ + 55, + 408, + 295, + 433 + ], + "type": "inline_equation", + "content": "m_t" + }, + { + "bbox": [ + 55, + 408, + 295, + 433 + ], + "type": "text", + "content": " by the loss " + }, + { + "bbox": [ + 55, + 408, + 295, + 433 + ], + "type": "inline_equation", + "content": "l_{MP}" + }, + { + "bbox": [ + 55, + 408, + 295, + 433 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 438, + 295, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 438, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 295, + 452 + ], + "type": "interline_equation", + "content": "l _ {M P} = \\ell_ {G T} + \\alpha \\left(\\ell_ {K S D} + \\ell_ {K A}\\right), \\tag {4}", + "image_path": "d7f90061ad97b63a64e08937f8eaef8357feaa3079cb15a8fccdc470eaa649d8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\ell_{GT}" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " denotes the L2 loss between the ground truth of the new keypoints and their predictions by the new model to acquire the knowledge of new keypoints. " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " is a hyperparameter to balance the new keypoints acquisition and old keypoints forgetting. For " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\ell_{KA}" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": ", we use the frozen KA-Net as an auxiliary teacher to supervised the old keypoint selected in Stage-I to distill the keypoint association knowledge. Since " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\ell_{KA}" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " is only applied to the selected old keypoint for knowledge transfer instead of mitigating the forgetting, we further consolidate the knowledge of all old keypoints by the loss " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\ell_{KSD}" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " using the frozen old model " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "m_{t - 1}" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": ". As the distribution of new and old data in IKL do not change dramatically and may even have a strong correlation, LWF [40] can be a baseline for " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\ell_{KSD}" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " in such a scenario as analyzed in [13]. In general, " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\ell_{KSD}" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " represents the negative log-likelihood " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\frac{1}{N_t}\\sum_i^{N_t}s\\left(\\hat{y}_i^{t - 1}\\right)\\cdot \\log s\\left(\\hat{y}_i^t\\right)" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " between the predictions of all old keypoints by the new model, " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^{t - 1}\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": ", and predictions by the old model, " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^t\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "s(\\cdot)" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " denotes the Softmax operator. " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " denotes the numbers of old keypoints learned before step " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " are the height and width of each keypoint's heatmap." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "However, since LWF and its variants all focus on the classification task, they perform the Softmax across different classes by default, i.e., over the " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " dimension, to obtain the normalized class prediction score. For IKL, this means normalizing across all old keypoints' predictions, which can not explicitly regularize the discrepancy of each old keypoint between the old and new model. To better preserve keypoint-specific knowledge, we adopt a spatial softmax operation over height " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "(H)" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "(W)" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " dimensions of the keypoint prediction, " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{sp}}^{\\mathrm{d}}(\\cdot), \\mathrm{d} \\in \\{H, W\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": ", and combine them to encourage spatial-oriented knowledge distillation. We discuss in our Supplementary the difference between each softmax alternative. Thus, our " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\ell_{KSD}" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 329, + 230, + 553, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 230, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 329, + 230, + 553, + 255 + ], + "type": "interline_equation", + "content": "\\frac {1}{N _ {t}} \\sum_ {i} ^ {N t} \\sum_ {j} ^ {C} \\sum_ {\\mathrm {d}} - s _ {\\mathrm {s p}} ^ {\\mathrm {d}} (\\hat {y} _ {i, j} ^ {t - 1}) \\cdot \\log s _ {\\mathrm {s p}} ^ {\\mathrm {d}} (\\hat {y} _ {i, j} ^ {t}). \\tag {5}", + "image_path": "6baf7534edbd6810c61076e099ec871c4f9ae7fb201bbb3271602bdfd156cbcc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 262, + 390, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 390, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 390, + 276 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 282, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 534 + ], + "type": "text", + "content": "Datasets. As the IKL has rarely been studied before, there is no specific dataset or benchmark for IKL. Motivated by the real-world application of IKL in medical analysis, we leverage the large-scale Cephalometric keypoint dataset [6] proposed in the 2023 MICCAI Challenge and also the commonly benchmarked Chest dataset [27] to further create the IKL protocols to validate our proposed method. Note that the Cephalometric keypoint dataset [6] differs from the head dataset [67] used in previous works [78, 79] commonly, as [6] has much larger number of keypoints and re-collected from different hospitals to enlarge the variance of images, making it much more challenging than [67]. Thus we term this dataset [6] as Head-2023 for clarity. Note that both Head-2023 and Chest datasets have less than or equal to 400 images in total. Moreover, though there is not a large need for IKL on human and animal, the widely used human and animal keypoint datasets, i.e., MPII [5] and ATRW [37], have extreme variations and non-rigid motions. Thus we also choose them for experiments to show the generality of our KAMP under different domains and challenging scenarios. We detail full dataset statistics in our Supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "Compared Methods. For keypoint estimation, common network structures for classification like ResNet [21] need specific modifications, e.g., adding the deconvolution layers and convolution layers to generate the keypoint location heatmap. These adjustments make many IL methods inapplicable for IKL, e.g., prototype-based methods [87, 88] like PASS [88] that leverages the feature mean prototype before the linear classification layer. Furthermore, the IKL requires that the old data can not be retained, and it is hard to identify specific methods for IKL. Thus we choose several general and representative exemplar-free methods that can adapt to different IL settings based on their methodologies, including EWC [33], LWF [40], MAS [72], RW [9], AFEC [69], CPR [8] for comparison. They are regularization-based methods and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25403" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 50, + 538, + 189 + ], + "blocks": [ + { + "bbox": [ + 72, + 50, + 538, + 189 + ], + "lines": [ + { + "bbox": [ + 72, + 50, + 538, + 189 + ], + "spans": [ + { + "bbox": [ + 72, + 50, + 538, + 189 + ], + "type": "table", + "html": "
MethodSplit ChestSplit Head-2023Split MPIISplit ATRW
A-MRE1↓AT1↑MT1↑A-MRE4↓AT4↑MT4↑AAA4↑AT4↑MT4↑AAA3↑AT3↑MT3↑
Joint Training5.43--2.12--88.50--94.69--
Finetune43.1--51.3--37.41--13.24--
EWC [33]13.28-8.23-3.6710.97-6.37-4.7638.64-51.84-12.2114.38-59.75-2.08
RW [9]9.48-7.12-4.156.49-4.23-2.8838.47-18.83-7.1384.15-10.870.00
MAS [4]7.36-1.86-0.195.31-2.15-1.3367.03-7.560.3485.68-5.80-1.13
LWF [40]6.35-1.340.184.31-1.260.5775.75-3.860.4187.31-5.10-0.64
AFEC [69]8.04-2.670.155.77-3.45-1.4637.24-22.85-15.4233.03-40.25-8.02
CPR [8]6.17-0.870.293.71-1.180.1675.52-3.240.7589.34-2.764.49
SFD [17]7.68-0.540.134.76-0.430.0271.49-0.930.2186.11-1.130.41
WF [74]7.31-0.310.164.580.030.1172.87-0.460.3886.69-0.970.62
GBD [16]6.420.060.214.340.120.4775.62-0.180.3587.42-0.890.65
KAMP (Ours)5.670.290.622.320.410.8479.931.804.2393.16-0.845.13
", + "image_path": "a3cdef58c9ba509fd022ad003e7076eec357de074f0904fea6e92df4d71a2f70.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 198, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 553, + 220 + ], + "type": "text", + "content": "Table 1. Result of 4 datasets after 2-Step, 5-Step, 5-Step and 4-Step IKL for Chest, Head-2023, MPII, and ATRW respectively. All comparison methods are started from the same Step-0 trained model. A-MRE: smaller the better; AAA/AT/MT: larger the better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 232, + 294, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 232, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 54, + 232, + 294, + 351 + ], + "type": "text", + "content": "can be easily applied to the IKL without trial and error. We further adapt methods from Continual Semantic Segmentation (CSS) and Class Incremental Learning (CIL) to ISL for more comparisons, i.e., spatial feature distillation (SFD) [17], weight fusion (WF) [74], and gradient balanced distillation (GBD) [16]. We also report the native baseline, i.e., directly finetune the model during IKL, and the upper bound that trains the model (e.g., HRNet [60]) with all the data, where we denote the former as \"Finetune\" and the latter as \"Joint Training\"." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 353, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 295, + 616 + ], + "type": "text", + "content": "Evaluation Metrics. To assess the keypoint regression task, we employ the widely used mean radial error (MRE) for Head-2023 and Chest datasets as in [78, 79], and Probability of Correct Keypoint (PCK) [50, 56, 66, 77] for MPII and ATRW and use their defaulted " + }, + { + "bbox": [ + 56, + 353, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 56, + 353, + 295, + 616 + ], + "type": "text", + "content": " as in [50, 60] to compute PCK. To measure the performance of incremental learning (IL), we use Average Accuracy (AAA), calculating the accuracy (\\%) across all keypoints post-incremental step " + }, + { + "bbox": [ + 56, + 353, + 295, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 56, + 353, + 295, + 616 + ], + "type": "text", + "content": " under PCK, and extend this approach to MRE, denoting it as A-MRE. Additionally, we introduce two metrics for knowledge transfer in IL: (1) Average Transfer (AT), also known as backward transfer [47], which averages the performance improvement of keypoints over all previous steps after step " + }, + { + "bbox": [ + 56, + 353, + 295, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 56, + 353, + 295, + 616 + ], + "type": "text", + "content": ", and (2) Maximal Transfer (MT), measuring the largest performance improvement in any old keypoint post-step " + }, + { + "bbox": [ + 56, + 353, + 295, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 56, + 353, + 295, + 616 + ], + "type": "text", + "content": ". Notably, when calculating AT and MT with MRE, we invert the sign of the change in error. This adjustment ensures consistency, since a decrease in MRE signifies improvement, in contrast to the direct correlation of increased accuracy with improvement in the PCK metric. Detailed explanations for the calculation of each metric are provided in our supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Experimental Design. We randomly split the keypoints of Chest, Head-2023, MPII, and ATRW into different incremental steps to create the Spilt Head-2023, Chest, MPII and ATRW protocols respectively. As the Chest dataset only has 6 keypoints in total, we split them into two groups where the first group has 4 keypoints with 150 training images, and the second group for IKL has 2 keypoints with only 50 training images. For Split Head-2023, we split the 38 keypoints into" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 232, + 553, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 363 + ], + "type": "text", + "content": "5 groups, where the first group has 19 keypoints with the same definition as [67] with 100 training images, and we split the rest of keypoint into 4 groups randomly to simulate four incremental steps with only 50 training images for each step. Similarly, the 16 MPII keypoints and 15 ATRW keypoints were split into 5 and 4 groups, respectively, with an equal distribution of training images per step. We repeat five times under different orders of keypoint groups and report the mean value in the main paper. The standard deviation (std), details about the keypoints group, and results of other experimental setups are in our Supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 364, + 553, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 553, + 555 + ], + "type": "text", + "content": "Implementation Details. We use the HRNet [60] as the backbone for all methods as it is widely compared in keypoint estimation [19, 36, 39, 50, 66, 79]. For a fair comparison, we initialize all methods with the same initial step (Step-0) model for incremental learning. As the first benchmark for the IKL, we use the Continual Hyperparameter Framework (CHF) [13, 41] to identify the training parameters like training epoch (100), initial learning rate (2e-3 for Split Chest and Head-2023, 1e-2 for Split MPII, 1e-3 for Split-ATRW), momentum (0.9), weight decay (1e-4), and also the hyperparameters of each compared methods, all included in our Supplementary. For our KAMP, the hyperparameter " + }, + { + "bbox": [ + 313, + 364, + 553, + 555 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 364, + 553, + 555 + ], + "type": "text", + "content": " is set as 1e2 for the Split Head-2023 and Split-MPII and 1e4 for the Split Chest and Split-ATRW in all experiments. The analysis of the " + }, + { + "bbox": [ + 313, + 364, + 553, + 555 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 364, + 553, + 555 + ], + "type": "text", + "content": " and other implementation details are included in our Supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 564, + 488, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 488, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 488, + 576 + ], + "type": "text", + "content": "4.1. Comparison with SOTA for IKL" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "As shown in Tab. 1, our KAMP consistently improves accuracy and reduces error across all old and new keypoints after each incremental step, achieving the highest performance compared to all other methods on all datasets. For example, KAMP outperforms the second-best method by " + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "inline_equation", + "content": "1.39\\%" + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{A - MRE_4}" + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": " after four incremental steps on Split Head-2023, and by " + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "inline_equation", + "content": "4.18\\%" + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AAA_4}" + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": " on Split MPII. Moreover, KAMP is the only method to consistently yield the highest average transfer (AT) and maximal transfer (MT) scores across all datasets. This result highlights that our two-stage learning scheme better facilitates knowledge transfer, thereby en" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25404" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "content": "hancing accuracy on old keypoints while learning new ones. When AT is positive, this indicates no forgetting on average, underscoring the effectiveness of using pseudo-labels from the old model to train KA-Net, as described in Sec. 3.2.1. In Split-ATRW, the AT is negative for all methods, indicating that forgetting of old keypoints outweighs knowledge transfer. However, KAMP still achieves the highest AT among the methods. Notably, the absolute value of KAMP's AT is nearly zero, suggesting minimal negative transfer. Additionally, by examining the maximal transfer metric, we observe a significant positive transfer in specific cases, such as " + }, + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "inline_equation", + "content": "5.13\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{MT}_3" + }, + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "content": ", indicating that some old keypoints experience a substantial positive transfer in each incremental step. This positive transfer helps offset the forgetting of other old keypoints, resulting in a minimal negative average transfer. We further validate this effect by reporting per-keypoint performance in our Supplementary. Qualitative results in Fig. 4 further verify our superiority in achieving more structurally correct prediction than competitive methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 300, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 295, + 529 + ], + "type": "text", + "content": "Lastly, we observe that methods adapted from CSS and CIL, i.e., SFD, WF, and GBD, achieve strong performance on AT and MT metrics compared to competitive methods like LWF and CPR, but perform poorly on overall metrics like A-MRE and AAA. This is because these methods prioritize reducing forgetting, resulting in an overly rigid model that struggles to effectively incorporate new keypoints in IKL. This observation highlights a crucial limitation of existing CSS and CIL methods that they focus heavily on regularizing forgetting but fail to balance this with knowledge acquisition for new keypoints, which is especially problematic in our newly defined IKL setting. Consequently, this emphasizes the need to study IKL as a distinct incremental learning scenario in this paper, as traditional incremental learning methods are too general to excel here. Our proposed novel baseline, KAMP, effectively addresses this gap for IKL, offering a well-balanced approach that goes beyond anti-forgetting to deliver robust performance across both accuracy and transfer metrics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 539, + 149, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 149, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 149, + 552 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "content": "To verify the effectiveness of our method, we further analyze how each component may influence the result. Since our method contains two-stage training and we consider a keypoint-oriented spatial distillation loss " + }, + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "inline_equation", + "content": "(\\ell_{KSD})" + }, + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "content": ", thus we compared our method with three alternatives: (1) the competitive baseline, i.e., the LWF, which performs the Softmax operation across old keypoints. (2) only use our " + }, + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\ell_{KSD}" + }, + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "content": " to train the model without the KA-Net. (3) Constructing the KA-Net randomly without using physical knowledge. Results are shown in Tab. 2, and we can observe that: (1) our adapted " + }, + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\ell_{KSD}" + }, + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "content": " can effectively outperform the LWF with " + }, + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.18\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "content": ", which demonstrates the essence of our keypoint-oriented adaptation. (2) With proper auxiliary task creation" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 323, + 70, + 547, + 144 + ], + "blocks": [ + { + "bbox": [ + 323, + 70, + 547, + 144 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 547, + 144 + ], + "type": "table", + "html": "
MethodAAA4↑AT4↑MT4↑
LWF [40]75.75-3.860.41
KAMP (only ℓKSD)76.93-2.240.65
KAMP (Random KA-Net)77.13-0.481.24
KAMP (Ours)79.931.804.23
", + "image_path": "9bc349fce77a40bc20027e1f385974a09571eb4f8f16e91bde38e4730d31ce2a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 167, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 553, + 299 + ], + "type": "text", + "content": "based on physical prior, we can achieve more positive average transfer than other counterparts and finally achieve the largest improvement over all the keypoints on " + }, + { + "bbox": [ + 313, + 167, + 553, + 299 + ], + "type": "inline_equation", + "content": "\\mathrm{AAA}_4" + }, + { + "bbox": [ + 313, + 167, + 553, + 299 + ], + "type": "text", + "content": ". This shows that our proposed two-stage learning scheme for IKL can not just provide better knowledge consolidation on the old keypoints than the competitive baseline, but the auxiliary prediction task can also bring large improvement to the old keypoints, showing that our KAMP will be a novel and strong baseline for the IKL paradigm. More ablation studies like using backbones other than HRNet and more datasets are in our Supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 304, + 539, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 539, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 539, + 318 + ], + "type": "text", + "content": "4.3. Compare IKL to other low-data paradigms" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 322, + 553, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 553, + 357 + ], + "type": "text", + "content": "Now we provide more insights into our IKL by comparing it with other learning paradigms with low-shot data. Discussion of limitations are in our Supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 358, + 555, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 597 + ], + "type": "text", + "content": "As mentioned before, existing learning paradigms like UKL [22, 23, 78] and CAPE [10, 56, 76] leverage large-scale self-supervised and multi-dataset training to enable novel keypoint detection with a few labeled samples in test time. To fairly compare with UKL and CAPE, we extend KAMP to low-data regime following EGT [78], where we also pretrain an auxiliary self-supervised model at Step-0 to provide pseudo-labels for old keypoints during IKL, and the details are in the Supplementary. For medical datasets like Head-2023, we compare to the SOTA one-shot methods CC2D [78] and EGT [79]. For 2D datasets like MPII, we compare to the SOTA Autolink [22] and MetaPoint+ [10] with official implementation available. Note that methods like [86] require extra expensive 3D CAD models to identify keypoint proposals for novel keypoint detection and are inferior to the SOTA [10, 56], thus we only compare to the MetaPoint+ [10]. For KAMP, we use the low-data adaptation for 1, 5, and 10 shot and do not use it for 50-shot since 50-shot is large enough for us to bypass the adaptation. Full experimental details are in our Supplementary." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 321, + 605, + 548, + 668 + ], + "blocks": [ + { + "bbox": [ + 362, + 152, + 504, + 163 + ], + "lines": [ + { + "bbox": [ + 362, + 152, + 504, + 163 + ], + "spans": [ + { + "bbox": [ + 362, + 152, + 504, + 163 + ], + "type": "text", + "content": "Table 2. Ablation Study on Split MPII." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 605, + 548, + 668 + ], + "lines": [ + { + "bbox": [ + 321, + 605, + 548, + 668 + ], + "spans": [ + { + "bbox": [ + 321, + 605, + 548, + 668 + ], + "type": "table", + "html": "
Method/(MRE ↓)1-shot5-shot10-shot50-shot
CC2D [78]5.144.834.083.47
EGT [79]5.014.583.873.21
KAMP (Ours)4.353.703.032.32
", + "image_path": "4870fbf0d37e554cc68359876e3eb8d2c08c358f1b3785e60e0e57d55bcf6964.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 676, + 542, + 687 + ], + "lines": [ + { + "bbox": [ + 324, + 676, + 542, + 687 + ], + "spans": [ + { + "bbox": [ + 324, + 676, + 542, + 687 + ], + "type": "text", + "content": "Table 3. Compare to low-shot methods on Split Head-2023." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "From Tab. 3 we observe that even on the extreme 1-shot setting, KAMP still outperforms the CC2D and EGT, while" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25405" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 74, + 553, + 224 + ], + "blocks": [ + { + "bbox": [ + 59, + 74, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 59, + 74, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 59, + 74, + 553, + 224 + ], + "type": "image", + "image_path": "4a26e20ec19212a20f8a594ec6b9c5df4390b1e52d27f7db6e2b1890108d71f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 289 + ], + "type": "text", + "content": "Figure 4. Qualitative results on Split Head-2023 and MPII. All methods start from the same Step-0 model, whose prediction is in the second column. GT: ground truth. The red circles denote the keypoints learned in Step 0, and the green ones denote all the new keypoints learned in later incremental steps. We observe that after IKL, the compared methods (LWF and CPR) may acquire the new keypoints as ours, but they have obvious miss-detection and wrong estimation (e.g., out of the body). While our method can consistently associate the new and old keypoints and achieve structurally accurate keypoint predictions. More results are included in our Supplementary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 294, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 296, + 402 + ], + "type": "text", + "content": "all the compared methods do not have good performances. This is because the Head-2023 dataset is collected from multiple sources with large discrepancy, and only one image is hard to represent all the variations of the data. However, when more annotations are available, KAMP can scale much better than CC2D and EGT since our two-stage learning scheme can capture the relation of old and new keypoints and distill them back to the model to improve the old keypoints, which has not been handled in CC2D and EGT." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 410, + 299, + 483 + ], + "blocks": [ + { + "bbox": [ + 56, + 410, + 299, + 483 + ], + "lines": [ + { + "bbox": [ + 56, + 410, + 299, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 299, + 483 + ], + "type": "table", + "html": "
Method/(PCK [10] ↑)1-shot5-shot10-shot50-shot
UKL [22]54.9559.1162.5364.36
MetaPoint+ [10]65.7166.8767.2668.98
KAMP (Ours)70.0972.2373.1876.97
KAMP (Ours) + [10]73.4975.1077.7679.18
", + "image_path": "c2bc7d004ba1dd51fb702b092fc69ca7a5eac2756a74dedc54b5c3d0a12aa75d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 491, + 273, + 502 + ], + "lines": [ + { + "bbox": [ + 77, + 491, + 273, + 502 + ], + "spans": [ + { + "bbox": [ + 77, + 491, + 273, + 502 + ], + "type": "text", + "content": "Table 4. Compare to low-shot methods on Split MPII." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 510, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 689 + ], + "type": "text", + "content": "As shown in Tab. 4, we observe that while the CAPE [10, 56, 76] and IKL are two disentangled paradigms, they can complement each other effectively. CAPE relies on pretraining with varied datasets, while IKL focuses on incremental learning, making their training costs incomparable. For example, CAPE methods like MetaPoint+ [10] require a substantial training cost (4 GPUs, 1.5 days), whereas the 'pretrain+incremental' approach of IKL is more efficient (1 GPU, 5 hours). To demonstrate that CAPE and IKL are orthogonal and complementary, we applied IKL using our KAMP method on MetaPoint+. As shown in Tab. 4, this combined approach results in significant improvements, even in low-data scenarios, highlighting that IKL can substantially enhance keypoint estimation when applied on a CAPE-pretrained keypoint detector." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "The goal of comparing these paradigms for novel key-point estimation within the same object category is to es" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 294, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 365 + ], + "type": "text", + "content": "tabish the necessity of IKL. Although CAPE does not involve continual learning and appears more deploy-friendly, its performance plateaus and does not scale effectively with additional data. In contrast, our KAMP achieves higher accuracy with limited data and demonstrates superior scalability with increased data, as evidenced in Tab. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 366, + 556, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 556, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 556, + 511 + ], + "type": "text", + "content": "Lastly, as the first study to explore the IKL paradigm, we focus on scaling up existing keypoint estimators for novel keypoints within the same object category, laying the foundation for future research. While CAPE can perform keypoint estimation on novel object categories with a few support images, its generalization remains limited. In the future, we will explore extending the IKL paradigm to novel object categories by applying IKL to keypoint estimators pretrained with CAPE, aiming to combine the strengths of both paradigms to develop a keypoint estimator capable of continually learning new keypoints across diverse object categories without forgetting and with effective scalability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 511, + 556, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 556, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 556, + 572 + ], + "type": "text", + "content": "In summary, the comparison of Tab. 3 and 4 show that our proposed IKL paradigm is highly label-efficient for acquiring new keypoints. This characteristic is advantageous for real-world applications where obtaining sufficient labels is time-consuming, such as in medical analysis [6]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 584, + 388, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 388, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 388, + 597 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "We explore learning the newly defined keypoint incrementally without retaining any old data, called Incremental object Keypoint Learning (IKL). We propose a two-stage learning scheme as a novel baseline tailored to the IKL. Extensive experiments show that our method can effectively alleviate the forgetting issue and may even improve the old keypoints' estimation during IKL. Our further analysis reveals that the IKL is label efficient in acquiring the new keypoints, which is promising for real-world applications." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25406" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 118 + ], + "type": "text", + "content": "Acknowledgements This work was supported in part by National Science Foundation grant IIS-2007613. This work was also supported by the National Natural Science Foundation of China (62376011)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 160, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 160, + 294, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 294, + 182 + ], + "type": "text", + "content": "[1] Hello gpt-4o, howpublished = https://openai.com/index/hello-gpt-4o/.5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 183, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 183, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 183, + 295, + 228 + ], + "type": "text", + "content": "[2] Hongjoon Ahn, Sungmin Cha, Donggyu Lee, and Taesup Moon. Uncertainty-based continual learning with adaptive regularization. In Advances in Neural Information Processing Systems, pages 4392-4402, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 229, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 229, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 229, + 294, + 272 + ], + "type": "text", + "content": "[3] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 328 + ], + "type": "text", + "content": "[4] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 331, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 295, + 384 + ], + "type": "text", + "content": "[5] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 1, 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 387, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 294, + 419 + ], + "type": "text", + "content": "[6] Jun Cao, Juan Dai, Xuguang Li, Bingsheng Huang, Ching-Wei Wang, and Hongyuan Zhang. Cephalometric landmark detection in lateral x-ray images, 2023. 1, 2, 5, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 295, + 464 + ], + "type": "text", + "content": "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 294, + 510 + ], + "type": "text", + "content": "[8] Sungmin Cha, Hsiang Hsu, Taebaek Hwang, Flavio Calmon, and Taesup Moon. {CPR}: Classifier-projection regularization for continual learning. In International Conference on Learning Representations, 2021. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 512, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 294, + 566 + ], + "type": "text", + "content": "[9] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajanthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European Conference on Computer Vision (ECCV), pages 532-547, 2018. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 568, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 295, + 621 + ], + "type": "text", + "content": "[10] Junjie Chen, Jiebin Yan, Yuming Fang, and Li Niu. Metapoint learning and refining for category-agnostic pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23534–23543, 2024. 3, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "text", + "content": "[11] Xiao Chu, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Structured feature learning for pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4715-4723, 2016. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[12] Shengyang Dai, Ming Yang, Ying Wu, and Aggelos Katsaggelos. Detector ensemble. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 1, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[13] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "text", + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "type": "text", + "content": "[15] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5138-5146, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 220, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 272 + ], + "type": "text", + "content": "[16] Jiahua Dong, Wenqi Liang, Yang Cong, and Gan Sun. Heterogeneous forgetting compensation for class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11742-11751, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 276, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 330 + ], + "type": "text", + "content": "[17] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4040-4050, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 332, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 386 + ], + "type": "text", + "content": "[18] Ana Paula Reis Durão, Aline Morosolli, Pisha Pittayapat, Napat Bolstad, Afonso P Ferreira, and Reinhilde Jacobs. Cephalometric landmark variability among orthodontists and dentomaxillofacial radiologists: a comparative study. Imaging science in dentistry, 45(4):213-220, 2015. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "type": "text", + "content": "[19] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14676-14686, 2021. 1, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 445, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 486 + ], + "type": "text", + "content": "[20] Kerui Gu, Linlin Yang, and Angela Yao. Removing the bias of integral pose regression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11067-11076, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 489, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 533 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "text", + "content": "[22] Xingzhe He, Bastian Wandt, and Helge Rhodin. Autolink: Self-supervised learning of human skeletons and object outlines by linking keypoints. In Advances in Neural Information Processing Systems, 2022. 1, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "type": "text", + "content": "[23] Xingzhe He, Gaurav Bharaj, David Ferman, Helge Rhodin, and Pablo Garrido. Few-shot geometry-aware keypoint localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21337-21348, 2023. 1, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[24] Sina Honari, Pavlo Molchanov, Stephen Tyree, Pascal Vincent, Christopher Pal, and Jan Kautz. Improving landmark localization with semi-supervised learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1546-1555, 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25407" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "text", + "content": "rebalancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-839, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 152 + ], + "type": "text", + "content": "[26] Zhongzhan Huang, Mingfu Liang, Senwei Liang, and Shanshan Zhong. Flat local minima for continual learning on semantic segmentation. In International Conference on Multimedia Modeling, pages 388-400. Springer, 2025. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 296, + 228 + ], + "type": "text", + "content": "[27] Stefan Jaeger, Alexandros Karargyris, Sema Candemir, Les Folio, Jenifer Siegelman, Fiona Callaghan, Zhiyun Xue, Kannappan Palaniappan, Rahul K Singh, Sameer Antani, George Thoma, Yi-Xiang Wang, Pu-Xuan Lu, and Clement J McDonald. Automatic tuberculosis screening using chest radiographs. IEEE Trans. Med. Imaging, 33(2):233-245, 2014. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 274 + ], + "type": "text", + "content": "[28] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. arXiv preprint arXiv:2207.13784, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "text", + "content": "[29] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In European Conference on Computer Vision, pages 196-214. Springer, 2020. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 374 + ], + "type": "text", + "content": "[30] Ta-Chu Kao, Kristopher Jensen, Gido van de Ven, Alberto Bernacchia, and Guillaume Hennequin. Natural continual learning: success is a journey, not (just) a destination. Advances in Neural Information Processing Systems, 34: 28067-28079, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 419 + ], + "type": "text", + "content": "[31] Leonid Karlinsky and Shimon Ullman. Using linking features in learning non-parametric part models. In European Conference on Computer Vision, pages 326-339. Springer, 2012. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "type": "text", + "content": "[32] Chris Dongjoo Kim, Jinseo Jeong, and Gunhee Kim. Imbalanced continual learning with partitioning reservoir sampling. In European Conference on Computer Vision, pages 411-428. Springer, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 295, + 533 + ], + "type": "text", + "content": "[33] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 589 + ], + "type": "text", + "content": "[34] Jiefeng Li, Siyuan Bian, Ailing Zeng, Can Wang, Bo Pang, Wentao Liu, and Cewu Lu. Human pose regression with residual log-likelihood estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11025–11034, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 633 + ], + "type": "text", + "content": "[35] Jiefeng Li, Tong Chen, Ruiqi Shi, Yujing Lou, Yong-Lu Li, and Cewu Lu. Localization with sampling-argmax. Advances in Neural Information Processing Systems, 34: 27236-27248, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "text", + "content": "[36] Ke Li, Shijie Wang, Xiang Zhang, Yifan Xu, Weijian Xu, and Zhuowen Tu. Pose recognition with cascade transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1944-1953, 2021. 1, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[37] Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, and Weiyao Lin. Atrw: A benchmark for amur tiger re-identification in" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2590-2598, 2020. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "text", + "content": "[38] Yanjie Li, Sen Yang, Shoukui Zhang, Zhicheng Wang, Wankou Yang, Shu-Tao Xia, and Erjin Zhou. Is 2d heatmap representation even necessary for human pose estimation? arXiv preprint arXiv:2107.03332, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "type": "text", + "content": "[39] Yanjie Li, Shoukui Zhang, Zhicheng Wang, Sen Yang, Wankou Yang, Shu-Tao Xia, and Erjin Zhou. Tokenpose: Learning keypoint tokens for human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11313-11322, 2021. 1, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "type": "text", + "content": "[40] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, 2017. 3, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "type": "text", + "content": "[41] Mingfu Liang, Jiahuan Zhou, Wei Wei, and Ying Wu. Balancing between forgetting and acquisition in incremental subpopulation learning. In European Conference on Computer Vision, pages 364-380. Springer, 2022. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 277, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 341 + ], + "type": "text", + "content": "[42] Mingfu Liang, Jong-Chyi Su, Samuel Schulter, Sparsh Garg, Shiyu Zhao, Ying Wu, and Manmohan Chandraker. Aide: An automatic data engine for object detection in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14695-14706, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "text", + "content": "[43] Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. Beyond not-forgetting: Continual learning with backward knowledge transfer. In Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "text", + "content": "[44] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 445, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 477 + ], + "type": "text", + "content": "[45] Buyu Liu and Vittorio Ferrari. Active learning for human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 4363-4372, 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 479, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 520 + ], + "type": "text", + "content": "[46] Vincenzo Lomonaco and Davide Maltoni. Core50: a new dataset and benchmark for continuous object recognition. In Conference on Robot Learning, pages 17-26. PMLR, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[47] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Advances in Neural Information Processing Systems, pages 6467-6476, 2017. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "type": "text", + "content": "[48] Zhengxiong Luo, Zhicheng Wang, Yan Huang, Liang Wang, Tieniu Tan, and Erjin Zhou. Rethinking the heatmap regression for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13264-13273, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[49] Dimitrios Mallis, Enrique Sanchez, Matthew Bell, and Georgios Tzimiropoulos. Unsupervised learning of object landmarks via self-training correspondence. Advances in Neural Information Processing Systems, 33:4709-4720, 2020. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[50] Olga Moskvyak, Frederic Maire, Feras Dayoub, and Mahsa Baktashmotlagh. Semi-supervised keypoint localization. In International Conference on Learning Representations, 2021. 1, 3, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25408" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[51] Gaurav Kumar Nayak, Het Shah, and Anirban Chakraborty. Incremental learning for animal pose estimation using rbf k-dpp. arXiv preprint arXiv:2110.13598, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "type": "text", + "content": "[52] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In European Conference on Computer Vision, pages 524-540. Springer, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 295, + 183 + ], + "type": "text", + "content": "[53] Ricketts Robert. Orthodontic Diagnosis and Planning: Their Roles in Preventive and Rehabilitative Dentistry. Denver Colo; Rocky Mountain/Orthodontics, 1982. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 183, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 295, + 239 + ], + "type": "text", + "content": "[54] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In Proceedings of the IEEE international conference on computer vision workshops, pages 397-403, 2013. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 295, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 272 + ], + "type": "text", + "content": "[55] Enrique Sanchez and Georgios Tzimiropoulos. Object landmark discovery through unsupervised adaptation. Advances in Neural Information Processing Systems, 32, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 337 + ], + "type": "text", + "content": "[56] Min Shi, Zihao Huang, Xianzheng Ma, Xiaowei Hu, and Zhiguo Cao. Matching is not enough: A two-stage framework for category-agnostic pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7308–7317, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 337, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 295, + 392 + ], + "type": "text", + "content": "[57] Aliaksandr Siarohin, Enver Sangineto, Stéphane Lathuiliere, and Nicu Sebe. Deformable gans for pose-based human image generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3408-3416, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 392, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 392, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 295, + 448 + ], + "type": "text", + "content": "[58] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2377-2386, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 448, + 295, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 448, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 448, + 295, + 481 + ], + "type": "text", + "content": "[59] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 295, + 536 + ], + "type": "text", + "content": "[60] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693-5703, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 536, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 295, + 581 + ], + "type": "text", + "content": "[61] Wei Tang and Ying Wu. Does learning specific features for related parts help human pose estimation? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1107-1116, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 581, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 295, + 624 + ], + "type": "text", + "content": "[62] Wei Tang, Pei Yu, and Ying Wu. Deeply learned compositional models for human pose estimation. In Proceedings of the European conference on computer vision (ECCV), pages 190-206, 2018. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 669 + ], + "type": "text", + "content": "[63] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial embeddings. In Proceedings of the IEEE international conference on computer vision, pages 5916-5925, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "text", + "content": "[64] Jonathan J. Thompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems, 27, 2014. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[65] Riccardo Volpi, Diane Larlus, and Grégory Rogez. Continual adaptation of visual representations via domain randomization and meta-learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4443-4453, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "text", + "content": "[66] Can Wang, Sheng Jin, Yingda Guan, Wentao Liu, Chen Qian, Ping Luo, and Wanli Ouyang. Pseudo-labeled autocurriculum learning for semi-supervised keypoint localization. In International Conference on Learning Representations, 2022. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 251 + ], + "type": "text", + "content": "[67] Ching-Wei Wang, Cheng-Ta Huang, Jia-Hong Lee, Chung-Hsing Li, Sheng-Wei Chang, Ming-Jhih Siao, Tat-Ming Lai, Bulat Ibragimov, Tomaz Vrtovec, Olaf Ronneberger, Philipp Fischer, Tim F Cootes, and Claudia Lindner. A benchmark for comparison of dental radiography analysis algorithms. Med. Image Anal., 31:63-76, 2016. 1, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 252, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 319 + ], + "type": "text", + "content": "[68] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 374 + ], + "type": "text", + "content": "[69] Liyuan Wang, Mingtian Zhang, Zhongfan Jia, Qian Li, Chenglong Bao, Kaisheng Ma, Jun Zhu, and Yi Zhong. Afec: Active forgetting of negative transfer in continual learning. Advances in Neural Information Processing Systems, 34: 22379-22391, 2021. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 374, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 374, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 553, + 420 + ], + "type": "text", + "content": "[70] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 420, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 465 + ], + "type": "text", + "content": "[71] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 4724–4732, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "type": "text", + "content": "[72] Chenshen Wu, Luis Herranz, Xialei Liu, Joost van de Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. In Advances in Neural Information Processing Systems, pages 5962-5972, 2018. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 521, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 577 + ], + "type": "text", + "content": "[73] Guile Wu, Shaogang Gong, and Pan Li. Striking a balance between stability and plasticity for class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1124-1133, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 577, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 633 + ], + "type": "text", + "content": "[74] Jia-Wen Xiao, Chang-Bin Zhang, Jiekang Feng, Xialei Liu, Joost van de Weijer, and Ming-Ming Cheng. Endpoints weight fusion for class incremental semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7204-7213, 2023. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 633, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 633, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 633, + 553, + 668 + ], + "type": "text", + "content": "[75] Xiaoying Xing, Mingfu Liang, and Ying Wu. Toa: Task-oriented active vqa. Advances in Neural Information Processing Systems, 36:54061-54074, 2023. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 668, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 668, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 668, + 553, + 713 + ], + "type": "text", + "content": "[76] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In Computer Vision-ECCV 2022: 17th European Conference," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25409" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 398-416. Springer, 2022. 2, 3, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "text", + "content": "[77] Sen Yang, Zhibin Quan, Mu Nie, and Wankou Yang. Transpose: Keypoint localization via transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11802-11812, 2021. 1, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "text", + "content": "[78] Qingsong Yao, Quan Quan, Li Xiao, and S Kevin Zhou. One-shot medical landmark detection. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24, pages 177-188. Springer, 2021. 3, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "type": "text", + "content": "[79] Zihao Yin, Ping Gong, Chunyu Wang, Yizhou Yu, and Yizhou Wang. One-shot medical landmark localization by edge-guided transform and noisy landmark refinement. In European Conference on Computer Vision, pages 473-489. Springer, 2022. 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "text", + "content": "[80] Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. AP-10k: A benchmark for animal pose estimation in the wild. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 317, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 381 + ], + "type": "text", + "content": "[81] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6982-6991, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 424 + ], + "type": "text", + "content": "[82] Duncan Zauss, Sven Kreiss, and Alexandre Alahi. Keypoint communities. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11057-11066, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 459 + ], + "type": "text", + "content": "[83] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. Proceedings of Machine Learning Research, 70:3987, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[84] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2694-2703, 2018. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "text", + "content": "[85] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13208-13217, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "text", + "content": "[86] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 615, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 615, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 615, + 294, + 657 + ], + "type": "text", + "content": "[87] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34:14306-14318, 2021. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[88] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[89] Xu Zou, Sheng Zhong, Luxin Yan, Xiangyun Zhao, Jiahuan Zhou, and Ying Wu. Learning robust facial landmark detection via hierarchical structured ensemble. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 141-150, 2019. 3" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25410" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/848b3be8-e7e7-4fb9-82b5-d4028969c01d_content_list.json b/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/848b3be8-e7e7-4fb9-82b5-d4028969c01d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..052edc561aa97689efddcb577138ba0746b4bc46 --- /dev/null +++ b/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/848b3be8-e7e7-4fb9-82b5-d4028969c01d_content_list.json @@ -0,0 +1,1687 @@ +[ + { + "type": "text", + "text": "IndoorGS: Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction", + "text_level": 1, + "bbox": [ + 204, + 130, + 794, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cong Ruan Yuesong Wang* Tao Guan", + "bbox": [ + 308, + 203, + 684, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computer Science & Technology, Huazhong University of Science and Technology, China {rc22, yuesongwang, qd.gt}@hust.edu.cn", + "bbox": [ + 99, + 220, + 895, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Zhang Shenzhen Smart City Technology Development Group Co., Ltd., China", + "bbox": [ + 217, + 263, + 779, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhangbin@smartcitysz.com", + "bbox": [ + 390, + 301, + 607, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lili Ju \nDepartment of Mathematics, University of South Carolina, USA", + "bbox": [ + 243, + 324, + 751, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ju@math.sc.edu", + "bbox": [ + 433, + 363, + 558, + 376 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86286184a81fa774ef4661008d3ba522052961a5bf96479b596ef63ebb4bad3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 440, + 308, + 539 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a383e9358108d975821e845dc09ce6b7c15b1cb5e20e79316dc07a49954a2c72.jpg", + "image_caption": [ + "GOF" + ], + "image_footnote": [], + "bbox": [ + 101, + 542, + 307, + 637 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/afc5f02bb84e4364acfbc1b7e5ca1bf4c1edaa9334d1bd29d36a12ec3f8dd5de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 443, + 503, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f514a39c5ccc3b8f36aee53f46497a711199eb45ba54e6c3cd0aa41c287f74ec.jpg", + "image_caption": [ + "RaDeGS" + ], + "image_footnote": [], + "bbox": [ + 318, + 542, + 503, + 638 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/82d714b2771d7b9c57300e2b8531079a8e47abb6743c3c278306faedc403b6a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 443, + 697, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6dd97cc05486b5981e971fff87df30dee8bf49442b5deb965f7db0ca011eb623.jpg", + "image_caption": [ + "PGSR", + "Figure 1. Our IndoorGS achieves state-of-the-art (SOTA) performance in indoor scene reconstruction. Compared to other methods that produce poor geometric results, our approach generates well-structural geometries while preserving all details. We can achieve superior rendering quality thanks to our precise geometry estimation." + ], + "image_footnote": [], + "bbox": [ + 514, + 542, + 697, + 638 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c81823ce147160df633083428e050a13e4b4d44d151a356c65c6aa5e24376c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 443, + 893, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f8e8c1c2b0debe1cd53f932432771077d304be39ce79c5f28ac3a4d7e1fcbd9c.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 707, + 542, + 893, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 718, + 326, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) has shown impressive performance in scene reconstruction, offering high rendering quality and rapid rendering speed with short training time. However, it often yields unsatisfactory results when applied to indoor scenes due to its poor ability to learn geometries without enough textural information. In this paper, we propose a new 3DGS-based method \"IndoorGS\", that leverages the commonly found yet important geometric", + "bbox": [ + 88, + 751, + 482, + 873 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cues in indoor scenes to improve the reconstruction quality. Specifically, we first extract 2D lines from input images and fuse them into 3D line cues via feature-based matching, which can provide a structural understanding of the target scene. We then apply the statistical outlier removal to refine Structure-from-Motion (SfM) points, ensuring robust cues in texture-rich areas. Based on these two types of cues, we further extract reliable 3D plane-like cues for textureless regions. Such geometric information will be utilized not only for initialization but also in the realization of a geometric-cue-guided adaptive density control (ADC) strategy. The proposed ADC approach is grounded in the principle of", + "bbox": [ + 511, + 720, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. Contact him at yuesongwang@hust.edu.cn.", + "bbox": [ + 107, + 888, + 465, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "844", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "divide-and-conquer and optimizes the use of each type of geometric cues to enhance overall reconstruction performance. Extensive experiments on multiple indoor datasets show that our method can deliver much more accurate geometry and higher rendering quality for indoor scenes than existing 3DGS approaches.", + "bbox": [ + 89, + 90, + 486, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 210, + 222, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Indoor scene reconstruction is one of common tasks in computer vision, with broad applications in virtual reality, gaming, autonomous driving, robotics, and so on [25, 51]. Recently, 3D Gaussian Splitting (3DGS) [15] has emerged as a promising scene reconstruction method, which introduces a hybrid optimization method with Gaussian primitives similar to point clouds and combines differentiable backpropagation with adaptive density control. This method has demonstrated superior efficacy, exhibiting expedited training process, accelerated rendering speed, and high-quality view synthesis, particularly in highly textured outdoor environments. However, it often performs poorly when applied to indoor environments, which mainly attributes to the fact that 3DGS is not capable of reconstructing correct geometry when facing textureless regions, as there are too many potential combinations of geometry and appearance that can satisfy the training images.", + "bbox": [ + 89, + 236, + 485, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although there has been a large amount of research works [5, 9, 10, 14, 47, 48] devoted to improving the geometry inference ability of 3DGS-based methods, their reconstruction quality is still unsatisfactory when dealing with indoor scenes, resulting in holes or floaters in the rendering images, as shown in Fig. 1. There are many local optima during the training process, and without strong constraints and correct guidance, 3DGS could be easily trapped by these local optima. What's even worse is that the initialization of 3DGS relies entirely on SfM points, which are inherently sparse and noisy in regions with weak textures, and the lack of reliable seeds can lead to blind exploration of 3DGS in these regions, further exacerbating the likelihood of being trapped by the local optima. To improve the performance of 3DGS for indoor scene reconstruction, we need to, on the one hand, provide better reliable seeds and, on the other hand, add strong geometric constraints to help the optimization move towards the global optima.", + "bbox": [ + 89, + 493, + 483, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To obtain better initial seeds, we propose to extract more reliable geometric cues commonly found in the indoor scene and use them to help initialize the 3DGS. Specifically, we first extract 2D line segments and perform feature-based matching on these 2D lines to obtain 3D line cues. Matching only line segments can drastically reduce the computational complexity and the matching ambiguity, allowing us to get accurate and noise-free 3D line cues to help 3DGS understand the scene's construction. While SfM [26, 32, 34]", + "bbox": [ + 89, + 765, + 485, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "points can provide good geometric priors in texture-rich regions, they usually behave as outliers in textureless or repeating-textured areas and can also trouble the reconstruction. Therefore, we filter the SfM points via Statistical Outlier Removal (SOR) and use the remaining points as reliable geometric cues in texture-rich regions. We further extend these cues to the weakly textured planar-like regions commonly existing in indoor scenes, such as walls, floors, ceilings, and furniture surfaces. We first extract the weakly textured regions from the 2D images and back-project them into 3D space using aligned monocular depths. We then analyze the spatial relationship between the extracted 3D lines and these back-projected planes, allowing us to extract reliable 3D plane-like cues. The above three types of geometric cues are fed to 3DGS for initialization.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the distinctive characteristics of these geometric cues, we further divide the Gaussians into three types, namely line Gaussians, plane-like Gaussians, and SfM Gaussians, and design a geometric-cue-guided adaptive density control strategy to enhance the optimization of 3DGS marching towards global optima. As 3D lines are typically located at object edges with high spatial accuracy, we apply a tangent space densification method to these line Gaussians, encouraging them to expand to complete the object. For plane-like Gaussians, since the initial seeds from 3D plane-like region cues may introduce spatial offsets and redundancies, we disable the splitting and cloning operations for these Gaussians, focusing solely on optimizing Gaussian parameters and pruning processes. For SfM geometric cues, because SfM points are often noisier than 3D lines, we implement a densification strategy that prioritizes line Gaussian over SfM Gaussian.", + "bbox": [ + 511, + 316, + 903, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 529, + 574, + 870, + 588 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a geometric cue extraction method for obtaining different types of reliable geometric priors from indoor scenes to guide the reconstruction.", + "- We classify Gaussian primitives according to their geometric types and design a geometric-cue-guided adaptive density control strategy to better utilize the unique properties of each type of geometric cue.", + "- Our proposed IndoorGS demonstrates superior performance in both the rendering quality and the accuracy of extracted geometry across various indoor datasets." + ], + "bbox": [ + 511, + 590, + 903, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 753, + 653, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3DGS has been as a prominent 3D reconstruction technique in the past couple of years due to its training efficiency, rendering performance, and impressive results across various 3D tasks [24, 33, 37, 41]. The core of 3DGS lies in a carefully designed rasterization [53] method that efficiently leverages GPU computation by representing scenes as collections of Gaussian primitives. Despite the remarkable success of 3DGS and its subsequent modifica", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "845", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4eb2bc40b68ad2f362d4009f638cfb33739f4dd9bd934f9f926831cda4a6b27c.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed IndoorGS. (a) The left part illustrates the extraction process of geometric cues, i.e., line cues, plane-like cues, and SfM cues, which are then used to initialize 3DGS. (b) The right part depicts the optimization process, where the 3DGS is further optimized using geometric-cue-guided adaptive density control and supervised by a series of loss functions." + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 903, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions [23, 35, 46, 50] in novel view synthesis, it faces significant challenges in surface reconstruction, mainly due to the discrete and unstructured nature of its Gaussian primitives. Recent research has sought to enhance 3DGS-based surface reconstruction [5, 9, 10, 14, 47, 48], yielding improvements in geometrically rich environments with strong textures. However, these methods still struggle in weakly textured settings, such as indoor scenes, which has driven growing interest in 3DGS reconstruction for indoor environments. GaussianRoom [42] integrates neural signed distance functions (SDF) with 3DGS, incorporating monocular normal and edge priors to enhance indoor scene details. Kim et al.[16] proposed an approach that employs 3D Gaussians for object reconstruction while using meshes to capture the room layout. 360-GS [1] leverages four panoramic images to initialize 3D Gaussians based on room layout and depth maps derived from panoramas for comprehensive indoor reconstruction. FreeSplat [39] applies 2D backbones and a cost volume for multi-view aggregation, while MGSLAM [21] utilizes the Manhattan World hypothesis to improve geometric accuracy and completeness in real-time SLAM [2-4] reconstructions. Despite the contributions, these methods usually require specialized inputs or are inefficient, exhibiting limitations in texture-poor environments. Indoor scenes are rich in geometric cues like lines and planes, which are useful for 3DGS-based reconstructions. However, SfM often encounters issues in these environments due to large textureless surfaces and repetitive elements. This is where 3D line map reconstruction approaches [12, 13, 20] show natural advantages in overcoming weak and repetitive textures common in indoor settings. By fully leveraging these geometric cues, 3DGS is expected to better handle the unique challenges of indoor", + "bbox": [ + 91, + 396, + 483, + 895 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scene reconstruction.", + "bbox": [ + 513, + 398, + 656, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 422, + 653, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 511, + 448, + 723, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3DGS [15] explicitly represents a 3D scene with a set of Gaussian primitives. Each of them is parameterized via a 3D covariance matrix $\\pmb{\\Sigma}$ and mean $\\pmb{\\mu}$ :", + "bbox": [ + 511, + 470, + 906, + 516 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (\\boldsymbol {p}) = \\exp \\left(- \\frac {1}{2} \\left(\\boldsymbol {p} - \\boldsymbol {\\mu}\\right) ^ {\\intercal} \\boldsymbol {\\Sigma} ^ {- 1} \\left(\\boldsymbol {p} - \\boldsymbol {\\mu}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 521, + 906, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{\\Sigma}$ can be further factorized into $\\mathbf{RSS}^{\\intercal}\\mathbf{R}^{\\intercal}$ with a scaling matrix $\\mathbf{S}$ and a rotation matrix $\\mathbf{R}$ . $\\alpha_{i}$ is the blending coefficient for a Gaussian with center $\\pmb{\\mu}_{i}$ in screen space:", + "bbox": [ + 511, + 556, + 906, + 603 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = o _ {i} \\cdot \\exp \\left(- \\frac {1}{2} (\\pmb {p} - \\pmb {\\mu} _ {i}) ^ {\\intercal} (\\pmb {\\Sigma} _ {i} ^ {\\prime}) ^ {- 1} (\\pmb {p} - \\pmb {\\mu} _ {i})\\right). (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 609, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $o_i$ is the Gaussian opacity, the 3D Gaussian is transformed into the image coordinates with world-to-camera transform matrix $\\mathbf{W}$ and projected onto pixel space with the Jacobian of the affine approximation for projection matrix $\\mathbf{J}$ , then the 2D covariance matrix $\\boldsymbol{\\Sigma}^{\\prime}$ is computed as $\\boldsymbol{\\Sigma}^{\\prime} = \\mathbf{J}\\mathbf{W}\\boldsymbol{\\Sigma}\\mathbf{W}^{\\mathrm{T}}\\mathbf{J}^{\\mathrm{T}}$ . The rendering result is computed using point-based alpha-blending on each ray:", + "bbox": [ + 511, + 648, + 908, + 755 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} T _ {i}, \\quad \\text {w h e r e} T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 761, + 906, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T_{i}$ is the accumulated transmittance at pixel location $\\pmb{p}$ , $\\mathrm{c}_i$ is the view-dependent color defined by spherical harmonics (SH) coefficients, and $N$ is the number of Gaussian primitives. A photometric loss function optimizes the Gaussian parameters to reduce the discrepancies between the rendered images and the actual observational data.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "846", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Rasterizing Depth and Normal", + "text_level": 1, + "bbox": [ + 89, + 90, + 362, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following RaDeGS [48], the center $\\mu_{i}$ of a Gaussian $G_{i}$ is initially projected into the camera space as $\\mu_{i}^{\\prime}$ . The depth of each pixel correspondent to the Gaussian is computed as:", + "bbox": [ + 89, + 112, + 483, + 159 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd = z _ {i} ^ {\\prime} + \\boldsymbol {p} \\left( \\begin{array}{c} \\Delta x \\\\ \\Delta y \\end{array} \\right), \\quad \\boldsymbol {\\mu} _ {i} ^ {\\prime} = \\left[ \\begin{array}{c} x _ {i} ^ {\\prime} \\\\ y _ {i} ^ {\\prime} \\\\ z _ {i} ^ {\\prime} \\end{array} \\right] = \\boldsymbol {W} \\boldsymbol {\\mu} _ {i} + t, (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 167, + 482, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{i}^{\\prime}$ represents the depth of the Gaussian center, $\\Delta x = x_{i}^{\\prime} - x$ and $\\Delta y = y_{i}^{\\prime} - y$ denote the relative pixel positions. The vector $\\pmb{p}$ is determined by the Gaussian parameters $[\\pmb{W},\\pmb{t}] \\in \\mathbb{R}^{3\\times 4}$ . Per-pixel z-depth estimates $\\hat{D}$ are rendered using the discrete volume rendering approximation similar to color values:", + "bbox": [ + 89, + 229, + 483, + 319 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {D} = \\sum_ {i \\in N} d _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 329, + 482, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{i}$ is the $i^{\\mathrm{th}}$ Gaussian's depth according to the Eq. (4). The normal direction of each projected Gaussian is computed as follows:", + "bbox": [ + 89, + 383, + 483, + 429 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {n}} = - \\boldsymbol {J} ^ {\\intercal} \\left(\\frac {z _ {i} ^ {\\prime}}{z _ {i}} \\boldsymbol {p}, 1\\right) ^ {\\intercal} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 440, + 482, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Normals are then alpha-blended, resulting in a single, unified per-pixel normal estimate:", + "bbox": [ + 89, + 484, + 483, + 516 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {N}} = \\sum_ {i \\in N} \\hat {\\boldsymbol {n}} _ {i} \\alpha_ {i} T _ {i}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 526, + 482, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 570, + 264, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 gives an overview of our methods. We first extract geometric cues, i.e., line, plane, and SfM point cues, from the input images (Section 4.1). Then, we divide Gaussians into three types according to the geometric cues and treat them differently in our geometric-cue-guided adaptive density control (Section 4.2). The training process is supervised by color and geometric loss (Section 4.3).", + "bbox": [ + 89, + 595, + 483, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Geometric Cues Extraction", + "text_level": 1, + "bbox": [ + 89, + 712, + 336, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SfM points are often too sparse to guide the reconstruction of indoor scenes and sometimes contain geometric errors, which can cause lots of floaters in the subsequent 3DGS reconstruction process. Therefore, we need to provide the 3DGS with sufficient and accurate geometric cues. As depicted in Fig. 2, there are three types of geometric cues that we use to guide the reconstruction.", + "bbox": [ + 89, + 734, + 483, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometric Cues from 3D lines. In contrast to sparse points from SfM, which is structureless, 3D line segments can concisely encode the high-level layout, as they often delineate the main structural elements of the scene. We follow", + "bbox": [ + 89, + 840, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the idea of LIMAP [20] to extract 3D lines, mainly consisting of 2D line extraction and line matching. Although rich 2D line segment features can be extracted from images using learning-based methods such as DeepLSD [29] or Sold2 [28], when encountering large continuous weakly textured or untextured areas, such as wall corner joint between white walls, they still have difficulties in obtaining distinctive 2D line segment features using only color information. However, we can easily extract distinctive line segment features from textureless regions on their normal maps, as shown in Fig. 3.", + "bbox": [ + 511, + 90, + 906, + 258 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b2d5a0b6a063fea8388cbb78d2651ef015aab20ee61a7efc3e522f286f6fb59d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 268, + 709, + 369 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f227e2694a21010988a00558ba27d9ee217a96b0b7e73d06c4ea8a26fce85f06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 268, + 895, + 369 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/21a41c1c872ff3abc3f23c0b2c51b29b9061e7b89a35800ea4f2906dab2be1bd.jpg", + "image_caption": [ + "RGB" + ], + "image_footnote": [], + "bbox": [ + 524, + 375, + 709, + 473 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1c19b040043e13a978e48120f3ca13047609a5966a376950486752d425422482.jpg", + "image_caption": [ + "Normal", + "Figure 3. The complementary effects of extracting structured lines from RGB images and normal maps. Top row: The line segments can be easily seen from the normal map, but extracting the corresponding line segments from the RGB image is difficult. Bottom row: The opposite case." + ], + "image_footnote": [], + "bbox": [ + 710, + 376, + 893, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, at the beginning of the reconstruction, we do not know the scene's geometry and can not render the normal maps. Fortunately, nowadays, foundation models for monocular normal estimation can give us a good prior. Thus, we extracted 2D line segments from RGB images and monocular normal maps using the DeepLSD [29] method. Subsequently, the GlueStick [30] method was used to extract the line segments and match them.", + "bbox": [ + 511, + 584, + 905, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we obtain the 3D line segments for the scene. To capture color information along these segments, we extract the corresponding 2D line colors from all visible views of each 3D line, then perform multiview color averaging to achieve a consistent color representation, and sample 3D points on 3D line segments in fixed steps.", + "bbox": [ + 511, + 704, + 905, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometric Cues from SfM. The sparse points of SfM in indoor scenes often contain noise or outliers to repetitive structures, reflective surfaces, and limited textures. Although sparse 3D feature points are generally accurate in strong-texture regions, they are often misaligned in weak-texture regions, as 2D points are more challenging to match correctly across multiple views. Therefore, filtering SfM", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "847", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "points is crucial, as erroneous inputs can be amplified during adaptive density control (ADC) in 3DGS, causing floater artifacts. To address this, we apply a statistical outlier removal filter (SOR) [31], which calculates the average distance between each point and its K nearest neighbors, removing points with abnormally high average distances. The remaining SfM points serve as reliable geometric cues in texture-rich regions.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Geometric Cues from Plane-like Regions. The plane-like regions in indoor scenes are usually weakly textured walls, ceilings, etc. For detecting these weakly textured regions in images, we use the Segment Anything Model (SAM) [17] for image segmentation, then determine texture strength by counting the density of SIFT [22] points in each segmented region.", + "bbox": [ + 89, + 212, + 480, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is difficult to calculate the depth of these regions due to the lack of texture information. However, the positions of the structural line clues on their edges are accurate enough. Thus, we can locate these plane-like cues using the 3D line cues with the help of the monocular depth estimation model. Specifically, we first obtain the monocular depth estimation from Metric3D [44]. While the monocular depth estimation can reflect the relative positions of lines and planes, it remains inconsistent with absolute depths. Thus, we estimate a relative per-image depth scale by aligning monocular depth with line cues and SfM cues using linear regression, which is then used to adjust the monocular depths. Based on the adjusted monocular depths, the weakly textured region is back-projected into 3D space to form a point cloud. To ensure the reliability of the back-projected point cloud, we then calculate the distances between sampling points from 3D line cues and the back-projected point cloud, based on which we filter out points exceeding a preset threshold. The retained points form the plane-like cues, giving reliable priors for weakly textured regions in 3D space. The whole process is described as Algorithm 1 in the supplementary material.", + "bbox": [ + 91, + 319, + 483, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Geometric-cue-guided ADC Strategies", + "text_level": 1, + "bbox": [ + 89, + 662, + 421, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides extracting points from geometric cues for initialization, we further let these cues guide the reconstruction process. To achieve this, we design Geometric-cue-guided Adaptive Densification Control (ADC) strategies for the three Gaussian types based on the positional accuracy and significance of their respective 3D spatial locations. In the following, we denote the Gaussians from SfM cues, line cues, and plane-like cues as $G_{M}$ , $G_{L}$ , and $G_{S}$ , respectively.", + "bbox": [ + 89, + 685, + 482, + 806 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Line Gaussian $G_{L}$ : They typically locate at object edges with high spatial accuracy. Thus, as discussed in detail below, we specifically design a tangent space densification method for them.", + "- SfM Gaussian $G_{M}$ : The ADC process for $G_{L}$ takes precedence over $G_{M}$ since $G_{M}$ are often noisier than $G_{L}$ ." + ], + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we allow $G_{M}$ to split and clone freely between 500-700 iterations to ensure it occupies strong texture regions correctly. Between iterations 700 and 3500, no ADC process is applied to $G_{M}$ ; only the tangent space densification of $G_{L}$ occurs during this period. After 3500 iterations, the ADC process for $G_{M}$ resumes.", + "bbox": [ + 524, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Plane-like Gaussian $G_{S}$ : Due to potential spatial offsets and redundancy of $G_{S}$ , we disable splitting and cloning operations for $G_{S}$ . Instead, only the parameters optimization and pruning processes are conducted for them.", + "bbox": [ + 513, + 181, + 903, + 242 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8600296d216d3b058c1cfc1f4f5c9aa590730db32304e9983ec2bed6169184b0.jpg", + "image_caption": [ + "Figure 4. Densification in the tangential space. Left: Cloning. When the Gaussian $G_{i}$ is cloned, an average gradient is added to its spatial location $\\mu_{i}$ . The projection of this gradient in the normal direction (the correction factor) is then subtracted, resulting in the new cloned green Gaussian $G_{i+1}$ . Right: Splitting. When the Gaussian $G_{i}$ is split, a small random vector $v$ is generated and added to its position. The component of $v$ along the normal is set to 0, yielding the two newly split Gaussians $G_{i+1}$ ." + ], + "image_footnote": [], + "bbox": [ + 519, + 256, + 709, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d617d2d7634461cd87bc46f3de155e2962cf08eb6c7fe8a72453f7d3d0c00111.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 256, + 901, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Densification in the tangential space. The vanilla 3D Gaussian Splatting (3DGS) is traditionally densified by randomly splitting or cloning Gaussian primitives within their Gaussian fields. While this method is flexible and easy to implement, it encounters difficulties in indoor environments as this unconstrained densification strategy can result in artifacts like floaters in weakly textured regions. Inspired by GeoGS [19], we propose a similar tangent space densification scheme. However, unlike GeoGS, whose initial point cloud is derived from PlanarSLAM [18] and whose normal vectors can be directly obtained from a spatial fit, we do not simply reduce the third scale of the Gaussian primitive to a very small value as we do not know the initial normal vectors. Instead, we introduce a soft constraint that flattens the Gaussian ellipsoid, akin to the approach used in Neusg [6], we directly minimize the minimum scale factor $S_{i} = \\text{diag}(s_{1}, s_{2}, s_{3})$ for each Gaussian that originates from 3D line points:", + "bbox": [ + 511, + 488, + 906, + 760 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\left\\| \\min \\left(s _ {1}, s _ {2}, s _ {3}\\right) \\right\\| _ {1}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 767, + 903, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this case, Gaussian primitive becomes flat, forming dislike surfel, and the smallest scaling axis approximates the normal direction. Simultaneously, we align the Gaussians' normal with the monocular surface normal estimated by Metrci3D using the following normal constraints:", + "bbox": [ + 511, + 789, + 903, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n} = \\frac {1}{| \\hat {N} |} \\sum \\left\\| \\hat {N} _ {i j} - N _ {\\text {m o n o}, i j} \\right\\| _ {1}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 869, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "848", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{N}_{ij}$ represents estimated normal values at pixel position $(i,j)$ , and $N_{\\mathrm{mono},ij}$ represents monocular normal values at pixel position $(i,j)$ . We then define the geometric normal for a flat Gaussian primitive as follows:", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {n}} _ {f l a t} = \\boldsymbol {R} \\cdot \\operatorname {O n e H o t} (\\arg \\min (s _ {1}, s _ {2}, s _ {3})), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 166, + 482, + 184 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where rotation matrix $\\mathbf{R} \\in SO(3)$ which is obtained from its quaternion $\\mathbf{q}$ , OneHot(.) ∈ $\\mathbb{R}^3$ returns a unit vector with zeros everywhere except at the position where the scaling $s_i = (s_1, s_2, s_3)$ is minimum. However, the orientation of the shortest axis can be ambiguous, as it may point either outward or inward relative to the surface. We leverage the relationship between the observation direction and the Gaussian normal to resolve this ambiguity. Specifically, the angle between the observation direction and the Gaussian normal should always be greater than 90 degrees. If this condition is not met, we simply invert the Gaussian normal vector by multiplying it by -1, ensuring the correct orientation. Then, we implement a tangent space densification strategy for these flat Gaussians, as shown in Fig. 4.", + "bbox": [ + 89, + 196, + 483, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When a cloning operation is performed, the position of the newly cloned Gaussian is determined by adding the average gradient to the position of the original Gaussian, along with a correction factor. The correction is calculated as follows:", + "bbox": [ + 89, + 409, + 483, + 482 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} ^ {i + 1} = \\boldsymbol {\\mu} ^ {i} + \\nabla \\boldsymbol {\\mu} ^ {i} - \\hat {\\mathbf {n}} _ {f l a t} \\cdot \\nabla \\boldsymbol {\\mu} ^ {i}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 486, + 482, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{\\mu}^i$ is the position of the original Gaussian $G^i$ , $\\nabla \\pmb{\\mu}^i$ is the accumulated average gradient of the original Gaussian $G_i$ , $\\hat{\\mathbf{n}}_{flat}$ is the normal of the original Gaussian $G_i$ .", + "bbox": [ + 89, + 513, + 483, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When a splitting operation is performed, the position of the newly split Gaussian is determined by adding an offset to the original Gaussian's position. This offset is drawn from a normal distribution with a mean of 0 and the three scales of the Gaussian as the standard deviation. We set the offset along the axis corresponding to the most miniature scale to zero, ensuring no movement occurs along that direction.", + "bbox": [ + 89, + 561, + 483, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The advantage of this approach is that the tangent space densification of $G_{L}$ enables better splitting and cloning along surfaces, effectively suppressing the issue of \"floaters\" caused by excessive splitting and cloning of erroneous seeds in $G_{S}$ .", + "bbox": [ + 89, + 683, + 483, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Loss Functions Definition", + "text_level": 1, + "bbox": [ + 89, + 771, + 323, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite adding geometric guidance to adaptive density control, Gaussians may still drift away from the correct locations during parameter optimization due to the weak constraints from color consistency loss at textureless regions, resulting in floaters or holes in the test views. Therefore, we add two geometric regularization losses to ensure the optimization proceeds in the right direction.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Monocular depth regularization loss. Monocular depth estimation faces challenges such as scale ambiguity, uneven scale distribution, and a lack of strict geometric consistency between depth maps. A simple linear transformation of monocular depth may destabilize 3D points and introduce distortions during alignment. To address this, the Pearson correlation coefficient (PCC) [8] can be used on small blocks of the depth map to extract robust geometric priors from monocular depth rather than imposing strict consistency. Specifically, we randomly sample $N$ non-overlapping patches to compute the monocular depth regularization loss as follows:", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\frac {1}{N} \\sum_ {i} ^ {N} 1 - P C C \\left(\\hat {\\boldsymbol {P}} _ {i}, \\boldsymbol {P} _ {\\text {m o n o , i}}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 282, + 906, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{P}_i\\in \\mathbb{R}^{S^2}$ denotes the $i$ th patch of $\\hat{D}$ , and $P_{\\mathrm{mono,i}}\\in \\mathbb{R}^{S^2}$ denotes the $i$ th patch of $D_{mono}$ , with the patch size $S$ being a hyperparameter. PCC, similar to normalized cross-correlation, measures the similarity between the two depth patches. This loss function encourages local consistency between rendered and monocular depth maps, improving alignment over small regions while handling local variations, thus enhancing depth estimation accuracy.", + "bbox": [ + 511, + 335, + 906, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Normal consistency loss. Following RaDeGS [48], we measure the consistency between the normal directions computed from the Gaussians and those derived from the depth map:", + "bbox": [ + 511, + 459, + 905, + 520 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n c} = \\sum \\left(1 - \\hat {\\boldsymbol {N}} _ {i j} \\cdot \\vec {\\boldsymbol {N}} _ {i j}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 529, + 905, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{N}_{ij}$ is the rendered normal at pixel(i, j) from Eq. (7), $\\vec{N}_{ij}$ is the normal estimated by the gradient of the depth map at pixel(i, j).", + "bbox": [ + 511, + 564, + 905, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Together with $\\mathcal{L}_{\\mathrm{n}}$ and $\\mathcal{L}_{\\mathrm{s}}$ used in the tangent space densification, our final loss function is defined as follows:", + "bbox": [ + 511, + 612, + 905, + 642 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\mathrm {n}} + \\lambda_ {\\mathrm {s}} \\mathcal {L} _ {\\mathrm {s}} + \\lambda_ {\\mathrm {n c}} \\mathcal {L} _ {\\mathrm {n c}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 652, + 903, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{rgb}}$ is the photometric loss proposed in 3DGS [15]. We set $\\lambda_{\\mathrm{d}} = 0.1$ , $\\lambda_{\\mathrm{n}} = 0.5$ , $\\lambda_{\\mathrm{s}} = 100.0$ and $\\lambda_{\\mathrm{nc}} = 0.1$ in our experiments.", + "bbox": [ + 511, + 680, + 905, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 739, + 645, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first demonstrate our IndoorGS's SOTA performance. We compare Gaussian-based methods, known for their high-quality rendering and fast training and rendering speed. We compare our method with 3DGS [15], GOF [47], RaDeGS [48] and PGSR [5]. And PGSR is the state-of-the-art surface reconstruction approach. We then analyze the contribution of each proposed component.", + "bbox": [ + 511, + 763, + 905, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We evaluate the performance of our approach on geometry reconstruction and rendering quality using 14", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "849", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b134ebe4a2a7e3aafd730e077a87fb78f811b87a76b78ed11547bffd359093d8.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison for geometry on Replica(the left two columns), Scannet++(the middle two columns), and Deep Blending(the right two columns) datasets. The experimental results demonstrate IndoorGS's robustness performance in less textured areas and ability to capture fine details as well." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 906, + 431 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8901497c2772f66a5b991e6044bfb29c3de601ed0f321f823cd7d23d5f57660a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ReplicaScannet++Deep BlendingAverage on all scenes
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓Time
3DGS39.520.9680.12129.460.9180.15229.410.9030.24332.790.9290.17210m
GOF40.980.9800.07429.820.9190.15329.210.9040.24633.330.9340.1582h
RaDeGS41.040.9800.07430.080.9210.14729.460.9070.24333.530.9360.15515m
PGSR40.020.9740.10129.990.9200.15429.220.8930.25533.080.9290.17025m
Ours41.110.9800.07330.450.9280.13130.060.9100.23533.880.9390.14635m
", + "bbox": [ + 114, + 498, + 883, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "indoor scenes from publicly available datasets. This includes 8 scenes from Replica [36], 4 scenes from ScanNet++ [43] and 2 scenes from Deep Blending [11].", + "bbox": [ + 89, + 648, + 482, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details. We choose RaDeGS [48] as our baseline. Since our initialization points are denser than vanilla 3DGS, the densification stops earlier at 12,000 iterations. To extract the final mesh, we apply Truncated Signed Distance Function (TSDF) [27] surface reconstruction, then use Marching Cube [40] to extract mesh. All experiments are conducted on a single Nvidia RTX4080 GPU.", + "bbox": [ + 89, + 699, + 482, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We chose three widely used image evaluation metrics for rendering evaluation: PSNR, SSIM, and LPIPS [49]. For mesh evaluation, we follow the evaluation protocol from monosdf [45] and Go-surf [38], report Chamfer- $L_{1}$ distance $(\\mathrm{C - }L_{1})$ , F-score (F1) with a threshold of 5cm and Normal Consistency (NC).", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4fa0322421ceae312c4ff668a8f0f5675f9c9d154433e461220fef843885f275.jpg", + "table_caption": [ + "Table 1. Quantitative comparison results of rendering quality for novel view synthesis on the Replica, Scannet++, and Deep Blending datasets. \"Red\" denote the best results, \"yellow\" denote the second best results. Test results on multiple datasets show that IndoorGS outperforms the other three methods." + ], + "table_footnote": [], + "table_body": "
ReplicaScannet++Average on all scenes
C-L1↓F1↑NC↑C-L1↓F1↑NC↑C-L1↓F1↑NC↑
GOF0.0860.5760.7970.1050.4790.7020.0950.5280.749
RaDeGS0.0790.6000.8620.0960.4970.7610.0870.5490.811
PGSR0.0490.8180.8950.0790.6250.7780.0640.7220.837
Ours0.0360.8250.9400.0400.7760.8770.0380.8010.909
", + "bbox": [ + 517, + 646, + 911, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison results of Chamfer distance, F-Score and Normal Consistency for reconstruction on the Replica and Scannet++ datasets. IndoorGS achieves the highest reconstruction accuracy across all metrics.", + "bbox": [ + 511, + 751, + 906, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Mesh Comparison", + "text_level": 1, + "bbox": [ + 511, + 830, + 692, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "IndoorGS demonstrates superior performance in indoor scene reconstruction, achieving the highest scores across all evaluated metrics—Chamfer Distance, F-Score, and Nor", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "850", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f329bb9377e6770e9ec158f5f507355916c51ce680c7e1c107333e5ebb24191e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGeometric CuesGeometric-cue-guided ADCLossMeshNovel View Synthesis
LineSfMPlaneLineSfMPlane\\(L_d\\)\\(L_n\\)\\(L_{nc}\\)C-L1↓F1↑NC↑PSNR↑SSIM↑LPIPS↓
w/o Line & Plane & SfM cues------✓✓✓0.0550.6810.85930.3080.9240.137
w/o Line & Plane cues-✓--✓-✓✓✓0.0510.7170.78930.5130.9270.135
w/o Plane cues✓✓-✓✓-✓✓✓0.0500.7280.86230.5830.9270.134
w/o All customized ADC✓✓✓---✓✓✓0.0520.7110.83930.4990.9260.135
w/o Line ADC✓✓✓-✓✓✓✓✓0.0510.7200.84830.5950.9260.135
w/o SfM ADC✓✓✓✓-✓✓✓✓0.0400.7710.87530.5730.9270.133
w/o Plane ADC✓✓✓✓✓-✓✓✓0.0460.7410.87130.3900.9250.137
w/o \\(L_d\\)✓✓✓✓✓✓-✓✓0.0440.7620.86430.5530.9260.134
w/o \\(L_n\\)✓✓✓✓✓✓✓-✓0.0510.7250.84830.5100.9260.137
w/o \\(L_{nc}\\)✓✓✓✓✓✓✓✓-0.0440.7440.80830.5050.9250.136
Full model✓✓✓✓✓✓✓✓✓0.0400.7760.87730.4530.9280.131
", + "bbox": [ + 96, + 88, + 903, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study results. Geometric cues significantly improve reconstruction quality. We report mesh and novel view synthesis metrics on the Scannet++ dataset. Adding the SfM cues means using the filtered SfM points instead, and using $L_{n}$ implies incorporating both $L_{n}$ and $L_{s}$ .", + "bbox": [ + 89, + 287, + 906, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mal Consistency, on all of the Replica, Scannet++, and Deep Blending datasets as shown in Tab. 2 and Fig. 5. This consistent excellence highlights its effectiveness in accurately reconstructing complex indoor environments. Since the Deep Blending dataset does not include directly provided mesh ground truth (GT), we do not perform quantitative metrics calculations. In contrast to other methods, which often result in uneven surfaces in weakly textured regions and overly smooth surfaces along edges, IndoorGS achieves a more accurate and detailed reconstruction.", + "bbox": [ + 88, + 338, + 480, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Rendering Comparison", + "text_level": 1, + "bbox": [ + 89, + 500, + 307, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the rendering quality of our method, we compared the test-view rendering results of our approach with those from vanilla 3DGS, GOF, RaDeGS, and PGSR on Replica, ScanNet++, and Deep Blending datasets. As shown in Tab. 1, Our method outperforms all of the above Gaussian-based reconstruction methods in terms of novel view synthesis, achieving an average PSNR that is 0.35 dB higher than the second-best method across the three datasets. Notably, following vanilla 3DGS, our test set is obtained by sequential sampling from the original dataset. The test views are positioned close to the training views in indoor scenes. In this case, despite the model overfitting the training views and the geometry inaccurately estimated, the PSNR of the rendered images from the test views can remain relatively high, which undermines the effectiveness of the test set in evaluating the true capability of the model, as it fails to challenge the model with sufficiently diverse perspectives. We provide a detailed experimental analysis of this issue in the appendix. Additionally, our method takes only 10 minutes longer than PGSR despite including the 3D line map construction and planar cues extraction process.", + "bbox": [ + 91, + 521, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 848, + 254, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted an ablation study on the ScanNet++ dataset to test the impact of each component we introduced.", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, we analyzed the effects of geometric cues, geometric-cue-guided ADC strategies, depth regularization, normal regularization, and normal consistency. The results are presented in Tab. 3. These results show that our full model performs best across all three mesh metrics and the final two NVS metrics (SSIM and LPIPS). Notably, as shown in the first three rows of the table, the step-by-step addition of geometric cues and the corresponding ADC strategy produces incremental performance improvement. Rows 4 to 7 further illustrate that introducing various ADC strategies after adding geometric cues results in further performance enhancement, and line-cue-guided ADC has the most significant impact among the three strategies. The final three rows show the positive effects of the three loss functions, with normal regularization providing the greatest enhancement.", + "bbox": [ + 511, + 338, + 906, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 599, + 633, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose IndoorGS, an indoor scene reconstruction method based on 3DGS. Our IndoorGS utilizes geometric cues such as filtered SfM points, line segments, and weakly textured planar surfaces, commonly found in indoor environments, as input data and takes a geometric-cues-guided ADC strategy for Gaussian scene optimization. We validate our method's rendering and reconstruction quality on the Replica, ScanNet++, and Deep Blending datasets. Experimental results demonstrate that our IndoorGS outperforms existing Gaussian-based techniques in geometric reconstruction accuracy and rendering quality.", + "bbox": [ + 511, + 627, + 906, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 813, + 665, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Project supported by the National Natural Science Foundation of China (Grant No. 62302174). The computation is completed in the HPC Platform of Huazhong University of Science and Technology.", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "851", + "bbox": [ + 485, + 944, + 509, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jiayang Bai, Letian Huang, Jie Guo, Wen Gong, Yuanqi Li, and Yanwen Guo. 360-gs: Layout-guided panoramic gaussian splatting for indoor roaming. arXiv preprint arXiv:2402.00763, 2024. 3", + "[2] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 3", + "[3] Danpeng Chen, Nan Wang, Runsen Xu, Weijian Xie, Hujun Bao, and Guofeng Zhang. Rnin-vio: Robust neural inertial navigation aided visual-inertial odometry in challenging scenes. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 275-283. IEEE, 2021.", + "[4] Danpeng Chen, Shuai Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Hujun Bao, and Guofeng Zhang. Vip-slam: An efficient tightly-coupled rgb-d visual inertial planar slam. In 2022 International Conference on Robotics and Automation (ICRA), pages 5615-5621. IEEE, 2022. 3", + "[5] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 2, 3, 6", + "[6] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 5", + "[7] Yutong Chen, Marko Mihajlovic, Xiyi Chen, Yiming Wang, Sergey Prokudin, and Siyu Tang. Splatformer: Point transformer for robust 3d gaussian splatting, 2024. 2", + "[8] Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. Pearson correlation coefficient. Noise reduction in speech processing, pages 1-4, 2009. 6", + "[9] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3", + "[10] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2, 3", + "[11] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 7", + "[12] Manuel Hofer, Michael Maurer, and Horst Bischof. Line3d: Efficient 3d scene abstraction for the built environment. In Pattern Recognition: 37th German Conference, GCPR 2015, Aachen, Germany, October 7-10, 2015, Proceedings 37, pages 237-248. Springer, 2015. 3", + "[13] Manuel Hofer, Michael Maurer, and Horst Bischof. Efficient 3d scene abstraction using line segments. Computer Vision and Image Understanding, 157:167-178, 2017. 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3", + "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 2, 3, 6", + "[16] Jiyeop Kim and Jongwoo Lim. Integrating meshes and 3d gaussians for indoor scene reconstruction with sam mask guidance. arXiv preprint arXiv:2407.16173, 2024. 3", + "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 5", + "[18] Yanyan Li, Raza Yunus, Nikolas Brasch, Nassir Navab, and Federico Tombari. Rgb-d slam with structural regularities. In 2021 IEEE international conference on Robotics and automation (ICRA), pages 11581-11587. IEEE, 2021. 5", + "[19] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering. arXiv preprint arXiv:2403.11324, 2024. 5", + "[20] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 3, 4", + "[21] Shuhong Liu, Heng Zhou, Liuzhuozheng Li, Yun Liu, Tianchen Deng, Yiming Zhou, and Mingrui Li. Structure gaussian slam with Manhattan world hypothesis. arXiv preprint arXiv:2405.20031, 2024. 3", + "[22] G Lowe. Sift-the scale invariant feature transform. Int. J, 2 (91-110):2, 2004. 5", + "[23] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3", + "[24] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2", + "[25] Run Luo, Zikai Song, Lintao Ma, Jinlin Wei, Wei Yang, and Min Yang. Diffusiontrack: Diffusion model for multi-object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3991-3999, 2024. 2", + "[26] Pierre Moulon, Pascal Monasse, Romuald Perrot, and Renaud Marlet. Openmvg: Open multiple view geometry. In Reproducible Research in Pattern Recognition: First International Workshop, RRPR 2016, Cancún, Mexico, December 4, 2016, Revised Selected Papers 1, pages 60-74. Springer, 2017. 2", + "[27] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and track" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "852", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 7", + "[28] Rémi Pautrat, Juan-Ting Lin, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Sold2: Self-supervised occlusion-aware line description and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11368-11378, 2021. 4", + "[29] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17327-17336, 2023. 4", + "[30] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 4", + "[31] Radu Bogdan Rusu and Steve Cousins. 3d is here: Point cloud library (pcl). In 2011 IEEE international conference on robotics and automation, pages 1-4. IEEE, 2011. 5", + "[32] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2", + "[33] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint arXiv:2312.05133, 2023. 2", + "[34] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International journal of computer vision, 80:189-210, 2008. 2", + "[35] Xiaowei Song, Jv Zheng, Shiran Yuan, Huan-ang Gao, Jingwei Zhao, Xiang He, Weihao Gu, and Hao Zhao. Saags: Scale-adaptive gaussian splatting for training-free antialIASing. arXiv preprint arXiv:2403.19615, 2024. 3", + "[36] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 7", + "[37] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2", + "[38] Jingwen Wang, Tymoteusz Bleja, and Lourdes Agapito. Gosurf: Neural feature grid optimization for fast, high-fidelity rgb-d surface reconstruction. In 2022 International Conference on 3D Vision (3DV), pages 433-442. IEEE, 2022. 7", + "[39] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free-view synthesis of indoor scenes. arXiv preprint arXiv:2405.17958, 2024. 3", + "[40] LORENSEN WE. Marching cubes: A high resolution 3d surface construction algorithm. Computer graphics, 21(1): 7-12, 1987. 7", + "[41] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang." + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2", + "[42] Haodong Xiang, Xinghui Li, Xiansong Lai, Wanting Zhang, Zhichao Liao, Kai Cheng, and Xueping Liu. Gaussianroom: Improving 3d gaussian splatting with sdf guidance and monocular cues for indoor scene reconstruction. arXiv preprint arXiv:2405.19671, 2024. 3", + "[43] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 7", + "[44] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9043-9053, 2023. 5", + "[45] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 7", + "[46] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 3", + "[47] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2, 3, 6", + "[48] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2, 3, 4, 6, 7", + "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7", + "[50] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 3", + "[51] Hang Zhou, Jiale Cai, Yuteng Ye, Yonghui Feng, Chenxing Gao, Junqing Yu, Zikai Song, and Wei Yang. Video anomaly detection with motion and appearance guided patch diffusion model. arXiv preprint arXiv:2412.09026, 2024. 2", + "[52] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 2", + "[53] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 2" + ], + "bbox": [ + 516, + 92, + 903, + 896 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "853", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/848b3be8-e7e7-4fb9-82b5-d4028969c01d_model.json b/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/848b3be8-e7e7-4fb9-82b5-d4028969c01d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..87b324661df6f23c00ac30f41fcb41b3e64ce18d --- /dev/null +++ b/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/848b3be8-e7e7-4fb9-82b5-d4028969c01d_model.json @@ -0,0 +1,2332 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.131, + 0.795, + 0.175 + ], + "angle": 0, + "content": "IndoorGS: Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.204, + 0.686, + 0.222 + ], + "angle": 0, + "content": "Cong Ruan Yuesong Wang* Tao Guan" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.222, + 0.896, + 0.257 + ], + "angle": 0, + "content": "School of Computer Science & Technology, Huazhong University of Science and Technology, China {rc22, yuesongwang, qd.gt}@hust.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.265, + 0.781, + 0.301 + ], + "angle": 0, + "content": "Bin Zhang Shenzhen Smart City Technology Development Group Co., Ltd., China" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.303, + 0.608, + 0.317 + ], + "angle": 0, + "content": "zhangbin@smartcitysz.com" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.325, + 0.753, + 0.361 + ], + "angle": 0, + "content": "Lili Ju \nDepartment of Mathematics, University of South Carolina, USA" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.364, + 0.56, + 0.377 + ], + "angle": 0, + "content": "ju@math.sc.edu" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.441, + 0.309, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.544, + 0.308, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.639, + 0.237, + 0.652 + ], + "angle": 0, + "content": "GOF" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.444, + 0.504, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.544, + 0.504, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.639, + 0.447, + 0.653 + ], + "angle": 0, + "content": "RaDeGS" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.444, + 0.699, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.544, + 0.699, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.639, + 0.634, + 0.653 + ], + "angle": 0, + "content": "PGSR" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.444, + 0.895, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.544, + 0.895, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.639, + 0.824, + 0.652 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.662, + 0.908, + 0.7 + ], + "angle": 0, + "content": "Figure 1. Our IndoorGS achieves state-of-the-art (SOTA) performance in indoor scene reconstruction. Compared to other methods that produce poor geometric results, our approach generates well-structural geometries while preserving all details. We can achieve superior rendering quality thanks to our precise geometry estimation." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.719, + 0.327, + 0.735 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.752, + 0.483, + 0.874 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) has shown impressive performance in scene reconstruction, offering high rendering quality and rapid rendering speed with short training time. However, it often yields unsatisfactory results when applied to indoor scenes due to its poor ability to learn geometries without enough textural information. In this paper, we propose a new 3DGS-based method \"IndoorGS\", that leverages the commonly found yet important geometric" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.721, + 0.909, + 0.904 + ], + "angle": 0, + "content": "cues in indoor scenes to improve the reconstruction quality. Specifically, we first extract 2D lines from input images and fuse them into 3D line cues via feature-based matching, which can provide a structural understanding of the target scene. We then apply the statistical outlier removal to refine Structure-from-Motion (SfM) points, ensuring robust cues in texture-rich areas. Based on these two types of cues, we further extract reliable 3D plane-like cues for textureless regions. Such geometric information will be utilized not only for initialization but also in the realization of a geometric-cue-guided adaptive density control (ADC) strategy. The proposed ADC approach is grounded in the principle of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.889, + 0.466, + 0.902 + ], + "angle": 0, + "content": "*Corresponding author. Contact him at yuesongwang@hust.edu.cn." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "844" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.488, + 0.184 + ], + "angle": 0, + "content": "divide-and-conquer and optimizes the use of each type of geometric cues to enhance overall reconstruction performance. Extensive experiments on multiple indoor datasets show that our method can deliver much more accurate geometry and higher rendering quality for indoor scenes than existing 3DGS approaches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.211, + 0.223, + 0.227 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.237, + 0.486, + 0.493 + ], + "angle": 0, + "content": "Indoor scene reconstruction is one of common tasks in computer vision, with broad applications in virtual reality, gaming, autonomous driving, robotics, and so on [25, 51]. Recently, 3D Gaussian Splitting (3DGS) [15] has emerged as a promising scene reconstruction method, which introduces a hybrid optimization method with Gaussian primitives similar to point clouds and combines differentiable backpropagation with adaptive density control. This method has demonstrated superior efficacy, exhibiting expedited training process, accelerated rendering speed, and high-quality view synthesis, particularly in highly textured outdoor environments. However, it often performs poorly when applied to indoor environments, which mainly attributes to the fact that 3DGS is not capable of reconstructing correct geometry when facing textureless regions, as there are too many potential combinations of geometry and appearance that can satisfy the training images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.485, + 0.765 + ], + "angle": 0, + "content": "Although there has been a large amount of research works [5, 9, 10, 14, 47, 48] devoted to improving the geometry inference ability of 3DGS-based methods, their reconstruction quality is still unsatisfactory when dealing with indoor scenes, resulting in holes or floaters in the rendering images, as shown in Fig. 1. There are many local optima during the training process, and without strong constraints and correct guidance, 3DGS could be easily trapped by these local optima. What's even worse is that the initialization of 3DGS relies entirely on SfM points, which are inherently sparse and noisy in regions with weak textures, and the lack of reliable seeds can lead to blind exploration of 3DGS in these regions, further exacerbating the likelihood of being trapped by the local optima. To improve the performance of 3DGS for indoor scene reconstruction, we need to, on the one hand, provide better reliable seeds and, on the other hand, add strong geometric constraints to help the optimization move towards the global optima." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.486, + 0.903 + ], + "angle": 0, + "content": "To obtain better initial seeds, we propose to extract more reliable geometric cues commonly found in the indoor scene and use them to help initialize the 3DGS. Specifically, we first extract 2D line segments and perform feature-based matching on these 2D lines to obtain 3D line cues. Matching only line segments can drastically reduce the computational complexity and the matching ambiguity, allowing us to get accurate and noise-free 3D line cues to help 3DGS understand the scene's construction. While SfM [26, 32, 34]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "points can provide good geometric priors in texture-rich regions, they usually behave as outliers in textureless or repeating-textured areas and can also trouble the reconstruction. Therefore, we filter the SfM points via Statistical Outlier Removal (SOR) and use the remaining points as reliable geometric cues in texture-rich regions. We further extend these cues to the weakly textured planar-like regions commonly existing in indoor scenes, such as walls, floors, ceilings, and furniture surfaces. We first extract the weakly textured regions from the 2D images and back-project them into 3D space using aligned monocular depths. We then analyze the spatial relationship between the extracted 3D lines and these back-projected planes, allowing us to extract reliable 3D plane-like cues. The above three types of geometric cues are fed to 3DGS for initialization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.573 + ], + "angle": 0, + "content": "Based on the distinctive characteristics of these geometric cues, we further divide the Gaussians into three types, namely line Gaussians, plane-like Gaussians, and SfM Gaussians, and design a geometric-cue-guided adaptive density control strategy to enhance the optimization of 3DGS marching towards global optima. As 3D lines are typically located at object edges with high spatial accuracy, we apply a tangent space densification method to these line Gaussians, encouraging them to expand to complete the object. For plane-like Gaussians, since the initial seeds from 3D plane-like region cues may introduce spatial offsets and redundancies, we disable the splitting and cloning operations for these Gaussians, focusing solely on optimizing Gaussian parameters and pruning processes. For SfM geometric cues, because SfM points are often noisier than 3D lines, we implement a densification strategy that prioritizes line Gaussian over SfM Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.575, + 0.871, + 0.589 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.591, + 0.904, + 0.635 + ], + "angle": 0, + "content": "- We propose a geometric cue extraction method for obtaining different types of reliable geometric priors from indoor scenes to guide the reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.636, + 0.905, + 0.695 + ], + "angle": 0, + "content": "- We classify Gaussian primitives according to their geometric types and design a geometric-cue-guided adaptive density control strategy to better utilize the unique properties of each type of geometric cue." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.696, + 0.905, + 0.741 + ], + "angle": 0, + "content": "- Our proposed IndoorGS demonstrates superior performance in both the rendering quality and the accuracy of extracted geometry across various indoor datasets." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.591, + 0.905, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.755, + 0.655, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "3DGS has been as a prominent 3D reconstruction technique in the past couple of years due to its training efficiency, rendering performance, and impressive results across various 3D tasks [24, 33, 37, 41]. The core of 3DGS lies in a carefully designed rasterization [53] method that efficiently leverages GPU computation by representing scenes as collections of Gaussian primitives. Despite the remarkable success of 3DGS and its subsequent modifica" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "845" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.086, + 0.904, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.328, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed IndoorGS. (a) The left part illustrates the extraction process of geometric cues, i.e., line cues, plane-like cues, and SfM cues, which are then used to initialize 3DGS. (b) The right part depicts the optimization process, where the 3DGS is further optimized using geometric-cue-guided adaptive density control and supervised by a series of loss functions." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.397, + 0.485, + 0.896 + ], + "angle": 0, + "content": "tions [23, 35, 46, 50] in novel view synthesis, it faces significant challenges in surface reconstruction, mainly due to the discrete and unstructured nature of its Gaussian primitives. Recent research has sought to enhance 3DGS-based surface reconstruction [5, 9, 10, 14, 47, 48], yielding improvements in geometrically rich environments with strong textures. However, these methods still struggle in weakly textured settings, such as indoor scenes, which has driven growing interest in 3DGS reconstruction for indoor environments. GaussianRoom [42] integrates neural signed distance functions (SDF) with 3DGS, incorporating monocular normal and edge priors to enhance indoor scene details. Kim et al.[16] proposed an approach that employs 3D Gaussians for object reconstruction while using meshes to capture the room layout. 360-GS [1] leverages four panoramic images to initialize 3D Gaussians based on room layout and depth maps derived from panoramas for comprehensive indoor reconstruction. FreeSplat [39] applies 2D backbones and a cost volume for multi-view aggregation, while MGSLAM [21] utilizes the Manhattan World hypothesis to improve geometric accuracy and completeness in real-time SLAM [2-4] reconstructions. Despite the contributions, these methods usually require specialized inputs or are inefficient, exhibiting limitations in texture-poor environments. Indoor scenes are rich in geometric cues like lines and planes, which are useful for 3DGS-based reconstructions. However, SfM often encounters issues in these environments due to large textureless surfaces and repetitive elements. This is where 3D line map reconstruction approaches [12, 13, 20] show natural advantages in overcoming weak and repetitive textures common in indoor settings. By fully leveraging these geometric cues, 3DGS is expected to better handle the unique challenges of indoor" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.399, + 0.658, + 0.412 + ], + "angle": 0, + "content": "scene reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.424, + 0.654, + 0.44 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.449, + 0.724, + 0.466 + ], + "angle": 0, + "content": "3.1. 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.471, + 0.908, + 0.517 + ], + "angle": 0, + "content": "3DGS [15] explicitly represents a 3D scene with a set of Gaussian primitives. Each of them is parameterized via a 3D covariance matrix \\(\\pmb{\\Sigma}\\) and mean \\(\\pmb{\\mu}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.522, + 0.907, + 0.552 + ], + "angle": 0, + "content": "\\[\nG (\\boldsymbol {p}) = \\exp \\left(- \\frac {1}{2} \\left(\\boldsymbol {p} - \\boldsymbol {\\mu}\\right) ^ {\\intercal} \\boldsymbol {\\Sigma} ^ {- 1} \\left(\\boldsymbol {p} - \\boldsymbol {\\mu}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.558, + 0.908, + 0.604 + ], + "angle": 0, + "content": "where \\(\\pmb{\\Sigma}\\) can be further factorized into \\(\\mathbf{RSS}^{\\intercal}\\mathbf{R}^{\\intercal}\\) with a scaling matrix \\(\\mathbf{S}\\) and a rotation matrix \\(\\mathbf{R}\\). \\(\\alpha_{i}\\) is the blending coefficient for a Gaussian with center \\(\\pmb{\\mu}_{i}\\) in screen space:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.61, + 0.907, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = o _ {i} \\cdot \\exp \\left(- \\frac {1}{2} (\\pmb {p} - \\pmb {\\mu} _ {i}) ^ {\\intercal} (\\pmb {\\Sigma} _ {i} ^ {\\prime}) ^ {- 1} (\\pmb {p} - \\pmb {\\mu} _ {i})\\right). (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.649, + 0.909, + 0.756 + ], + "angle": 0, + "content": "where \\(o_i\\) is the Gaussian opacity, the 3D Gaussian is transformed into the image coordinates with world-to-camera transform matrix \\(\\mathbf{W}\\) and projected onto pixel space with the Jacobian of the affine approximation for projection matrix \\(\\mathbf{J}\\), then the 2D covariance matrix \\(\\boldsymbol{\\Sigma}^{\\prime}\\) is computed as \\(\\boldsymbol{\\Sigma}^{\\prime} = \\mathbf{J}\\mathbf{W}\\boldsymbol{\\Sigma}\\mathbf{W}^{\\mathrm{T}}\\mathbf{J}^{\\mathrm{T}}\\). The rendering result is computed using point-based alpha-blending on each ray:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.762, + 0.907, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\hat {C} = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} T _ {i}, \\quad \\text {w h e r e} T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "where \\( T_{i} \\) is the accumulated transmittance at pixel location \\( \\pmb{p} \\), \\( \\mathrm{c}_i \\) is the view-dependent color defined by spherical harmonics (SH) coefficients, and \\( N \\) is the number of Gaussian primitives. A photometric loss function optimizes the Gaussian parameters to reduce the discrepancies between the rendered images and the actual observational data." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "846" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.364, + 0.108 + ], + "angle": 0, + "content": "3.2. Rasterizing Depth and Normal" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.16 + ], + "angle": 0, + "content": "Following RaDeGS [48], the center \\(\\mu_{i}\\) of a Gaussian \\(G_{i}\\) is initially projected into the camera space as \\(\\mu_{i}^{\\prime}\\). The depth of each pixel correspondent to the Gaussian is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.169, + 0.483, + 0.22 + ], + "angle": 0, + "content": "\\[\nd = z _ {i} ^ {\\prime} + \\boldsymbol {p} \\left( \\begin{array}{c} \\Delta x \\\\ \\Delta y \\end{array} \\right), \\quad \\boldsymbol {\\mu} _ {i} ^ {\\prime} = \\left[ \\begin{array}{c} x _ {i} ^ {\\prime} \\\\ y _ {i} ^ {\\prime} \\\\ z _ {i} ^ {\\prime} \\end{array} \\right] = \\boldsymbol {W} \\boldsymbol {\\mu} _ {i} + t, (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.485, + 0.32 + ], + "angle": 0, + "content": "where \\( z_{i}^{\\prime} \\) represents the depth of the Gaussian center, \\( \\Delta x = x_{i}^{\\prime} - x \\) and \\( \\Delta y = y_{i}^{\\prime} - y \\) denote the relative pixel positions. The vector \\( \\pmb{p} \\) is determined by the Gaussian parameters \\( [\\pmb{W},\\pmb{t}] \\in \\mathbb{R}^{3\\times 4} \\). Per-pixel z-depth estimates \\( \\hat{D} \\) are rendered using the discrete volume rendering approximation similar to color values:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.33, + 0.483, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\hat {D} = \\sum_ {i \\in N} d _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.43 + ], + "angle": 0, + "content": "where \\(d_{i}\\) is the \\(i^{\\mathrm{th}}\\) Gaussian's depth according to the Eq. (4). The normal direction of each projected Gaussian is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.441, + 0.483, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {n}} = - \\boldsymbol {J} ^ {\\intercal} \\left(\\frac {z _ {i} ^ {\\prime}}{z _ {i}} \\boldsymbol {p}, 1\\right) ^ {\\intercal} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.517 + ], + "angle": 0, + "content": "Normals are then alpha-blended, resulting in a single, unified per-pixel normal estimate:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.527, + 0.483, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {N}} = \\sum_ {i \\in N} \\hat {\\boldsymbol {n}} _ {i} \\alpha_ {i} T _ {i}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.571, + 0.265, + 0.589 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.703 + ], + "angle": 0, + "content": "Fig. 2 gives an overview of our methods. We first extract geometric cues, i.e., line, plane, and SfM point cues, from the input images (Section 4.1). Then, we divide Gaussians into three types according to the geometric cues and treat them differently in our geometric-cue-guided adaptive density control (Section 4.2). The training process is supervised by color and geometric loss (Section 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.713, + 0.338, + 0.728 + ], + "angle": 0, + "content": "4.1. Geometric Cues Extraction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.84 + ], + "angle": 0, + "content": "SfM points are often too sparse to guide the reconstruction of indoor scenes and sometimes contain geometric errors, which can cause lots of floaters in the subsequent 3DGS reconstruction process. Therefore, we need to provide the 3DGS with sufficient and accurate geometric cues. As depicted in Fig. 2, there are three types of geometric cues that we use to guide the reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Geometric Cues from 3D lines. In contrast to sparse points from SfM, which is structureless, 3D line segments can concisely encode the high-level layout, as they often delineate the main structural elements of the scene. We follow" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.259 + ], + "angle": 0, + "content": "the idea of LIMAP [20] to extract 3D lines, mainly consisting of 2D line extraction and line matching. Although rich 2D line segment features can be extracted from images using learning-based methods such as DeepLSD [29] or Sold2 [28], when encountering large continuous weakly textured or untextured areas, such as wall corner joint between white walls, they still have difficulties in obtaining distinctive 2D line segment features using only color information. However, we can easily extract distinctive line segment features from textureless regions on their normal maps, as shown in Fig. 3." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.27, + 0.71, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.27, + 0.896, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.376, + 0.71, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.377, + 0.895, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.478, + 0.633, + 0.489 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.477, + 0.825, + 0.488 + ], + "angle": 0, + "content": "Normal" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.499, + 0.907, + 0.569 + ], + "angle": 0, + "content": "Figure 3. The complementary effects of extracting structured lines from RGB images and normal maps. Top row: The line segments can be easily seen from the normal map, but extracting the corresponding line segments from the RGB image is difficult. Bottom row: The opposite case." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.704 + ], + "angle": 0, + "content": "However, at the beginning of the reconstruction, we do not know the scene's geometry and can not render the normal maps. Fortunately, nowadays, foundation models for monocular normal estimation can give us a good prior. Thus, we extracted 2D line segments from RGB images and monocular normal maps using the DeepLSD [29] method. Subsequently, the GlueStick [30] method was used to extract the line segments and match them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.796 + ], + "angle": 0, + "content": "Finally, we obtain the 3D line segments for the scene. To capture color information along these segments, we extract the corresponding 2D line colors from all visible views of each 3D line, then perform multiview color averaging to achieve a consistent color representation, and sample 3D points on 3D line segments in fixed steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Geometric Cues from SfM. The sparse points of SfM in indoor scenes often contain noise or outliers to repetitive structures, reflective surfaces, and limited textures. Although sparse 3D feature points are generally accurate in strong-texture regions, they are often misaligned in weak-texture regions, as 2D points are more challenging to match correctly across multiple views. Therefore, filtering SfM" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "847" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "points is crucial, as erroneous inputs can be amplified during adaptive density control (ADC) in 3DGS, causing floater artifacts. To address this, we apply a statistical outlier removal filter (SOR) [31], which calculates the average distance between each point and its K nearest neighbors, removing points with abnormally high average distances. The remaining SfM points serve as reliable geometric cues in texture-rich regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.482, + 0.318 + ], + "angle": 0, + "content": "Geometric Cues from Plane-like Regions. The plane-like regions in indoor scenes are usually weakly textured walls, ceilings, etc. For detecting these weakly textured regions in images, we use the Segment Anything Model (SAM) [17] for image segmentation, then determine texture strength by counting the density of SIFT [22] points in each segmented region." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.32, + 0.485, + 0.651 + ], + "angle": 0, + "content": "It is difficult to calculate the depth of these regions due to the lack of texture information. However, the positions of the structural line clues on their edges are accurate enough. Thus, we can locate these plane-like cues using the 3D line cues with the help of the monocular depth estimation model. Specifically, we first obtain the monocular depth estimation from Metric3D [44]. While the monocular depth estimation can reflect the relative positions of lines and planes, it remains inconsistent with absolute depths. Thus, we estimate a relative per-image depth scale by aligning monocular depth with line cues and SfM cues using linear regression, which is then used to adjust the monocular depths. Based on the adjusted monocular depths, the weakly textured region is back-projected into 3D space to form a point cloud. To ensure the reliability of the back-projected point cloud, we then calculate the distances between sampling points from 3D line cues and the back-projected point cloud, based on which we filter out points exceeding a preset threshold. The retained points form the plane-like cues, giving reliable priors for weakly textured regions in 3D space. The whole process is described as Algorithm 1 in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.663, + 0.422, + 0.68 + ], + "angle": 0, + "content": "4.2. Geometric-cue-guided ADC Strategies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Besides extracting points from geometric cues for initialization, we further let these cues guide the reconstruction process. To achieve this, we design Geometric-cue-guided Adaptive Densification Control (ADC) strategies for the three Gaussian types based on the positional accuracy and significance of their respective 3D spatial locations. In the following, we denote the Gaussians from SfM cues, line cues, and plane-like cues as \\( G_{M} \\), \\( G_{L} \\), and \\( G_{S} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.869 + ], + "angle": 0, + "content": "- Line Gaussian \\( G_{L} \\): They typically locate at object edges with high spatial accuracy. Thus, as discussed in detail below, we specifically design a tangent space densification method for them." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- SfM Gaussian \\(G_{M}\\): The ADC process for \\(G_{L}\\) takes precedence over \\(G_{M}\\) since \\(G_{M}\\) are often noisier than \\(G_{L}\\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "Specifically, we allow \\( G_{M} \\) to split and clone freely between 500-700 iterations to ensure it occupies strong texture regions correctly. Between iterations 700 and 3500, no ADC process is applied to \\( G_{M} \\); only the tangent space densification of \\( G_{L} \\) occurs during this period. After 3500 iterations, the ADC process for \\( G_{M} \\) resumes." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.243 + ], + "angle": 0, + "content": "- Plane-like Gaussian \\( G_{S} \\): Due to potential spatial offsets and redundancy of \\( G_{S} \\), we disable splitting and cloning operations for \\( G_{S} \\). Instead, only the parameters optimization and pruning processes are conducted for them." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.257, + 0.71, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.257, + 0.903, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.475 + ], + "angle": 0, + "content": "Figure 4. Densification in the tangential space. Left: Cloning. When the Gaussian \\( G_{i} \\) is cloned, an average gradient is added to its spatial location \\( \\mu_{i} \\). The projection of this gradient in the normal direction (the correction factor) is then subtracted, resulting in the new cloned green Gaussian \\( G_{i+1} \\). Right: Splitting. When the Gaussian \\( G_{i} \\) is split, a small random vector \\( v \\) is generated and added to its position. The component of \\( v \\) along the normal is set to 0, yielding the two newly split Gaussians \\( G_{i+1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.907, + 0.761 + ], + "angle": 0, + "content": "Densification in the tangential space. The vanilla 3D Gaussian Splatting (3DGS) is traditionally densified by randomly splitting or cloning Gaussian primitives within their Gaussian fields. While this method is flexible and easy to implement, it encounters difficulties in indoor environments as this unconstrained densification strategy can result in artifacts like floaters in weakly textured regions. Inspired by GeoGS [19], we propose a similar tangent space densification scheme. However, unlike GeoGS, whose initial point cloud is derived from PlanarSLAM [18] and whose normal vectors can be directly obtained from a spatial fit, we do not simply reduce the third scale of the Gaussian primitive to a very small value as we do not know the initial normal vectors. Instead, we introduce a soft constraint that flattens the Gaussian ellipsoid, akin to the approach used in Neusg [6], we directly minimize the minimum scale factor \\( S_{i} = \\text{diag}(s_{1}, s_{2}, s_{3}) \\) for each Gaussian that originates from 3D line points:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.768, + 0.905, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\left\\| \\min \\left(s _ {1}, s _ {2}, s _ {3}\\right) \\right\\| _ {1}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.79, + 0.905, + 0.866 + ], + "angle": 0, + "content": "In this case, Gaussian primitive becomes flat, forming dislike surfel, and the smallest scaling axis approximates the normal direction. Simultaneously, we align the Gaussians' normal with the monocular surface normal estimated by Metrci3D using the following normal constraints:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.871, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n} = \\frac {1}{| \\hat {N} |} \\sum \\left\\| \\hat {N} _ {i j} - N _ {\\text {m o n o}, i j} \\right\\| _ {1}, \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "848" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.152 + ], + "angle": 0, + "content": "where \\(\\hat{N}_{ij}\\) represents estimated normal values at pixel position \\((i,j)\\), and \\(N_{\\mathrm{mono},ij}\\) represents monocular normal values at pixel position \\((i,j)\\). We then define the geometric normal for a flat Gaussian primitive as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.167, + 0.483, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {n}} _ {f l a t} = \\boldsymbol {R} \\cdot \\operatorname {O n e H o t} (\\arg \\min (s _ {1}, s _ {2}, s _ {3})), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.197, + 0.484, + 0.408 + ], + "angle": 0, + "content": "where rotation matrix \\( \\mathbf{R} \\in SO(3) \\) which is obtained from its quaternion \\( \\mathbf{q} \\), OneHot(.) ∈ \\( \\mathbb{R}^3 \\) returns a unit vector with zeros everywhere except at the position where the scaling \\( s_i = (s_1, s_2, s_3) \\) is minimum. However, the orientation of the shortest axis can be ambiguous, as it may point either outward or inward relative to the surface. We leverage the relationship between the observation direction and the Gaussian normal to resolve this ambiguity. Specifically, the angle between the observation direction and the Gaussian normal should always be greater than 90 degrees. If this condition is not met, we simply invert the Gaussian normal vector by multiplying it by -1, ensuring the correct orientation. Then, we implement a tangent space densification strategy for these flat Gaussians, as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.483 + ], + "angle": 0, + "content": "When a cloning operation is performed, the position of the newly cloned Gaussian is determined by adding the average gradient to the position of the original Gaussian, along with a correction factor. The correction is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.487, + 0.483, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} ^ {i + 1} = \\boldsymbol {\\mu} ^ {i} + \\nabla \\boldsymbol {\\mu} ^ {i} - \\hat {\\mathbf {n}} _ {f l a t} \\cdot \\nabla \\boldsymbol {\\mu} ^ {i}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.514, + 0.484, + 0.56 + ], + "angle": 0, + "content": "where \\(\\pmb{\\mu}^i\\) is the position of the original Gaussian \\(G^i\\), \\(\\nabla \\pmb{\\mu}^i\\) is the accumulated average gradient of the original Gaussian \\(G_i\\), \\(\\hat{\\mathbf{n}}_{flat}\\) is the normal of the original Gaussian \\(G_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.484, + 0.681 + ], + "angle": 0, + "content": "When a splitting operation is performed, the position of the newly split Gaussian is determined by adding an offset to the original Gaussian's position. This offset is drawn from a normal distribution with a mean of 0 and the three scales of the Gaussian as the standard deviation. We set the offset along the axis corresponding to the most miniature scale to zero, ensuring no movement occurs along that direction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.484, + 0.759 + ], + "angle": 0, + "content": "The advantage of this approach is that the tangent space densification of \\( G_{L} \\) enables better splitting and cloning along surfaces, effectively suppressing the issue of \"floaters\" caused by excessive splitting and cloning of erroneous seeds in \\( G_{S} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.325, + 0.787 + ], + "angle": 0, + "content": "4.3. Loss Functions Definition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Despite adding geometric guidance to adaptive density control, Gaussians may still drift away from the correct locations during parameter optimization due to the weak constraints from color consistency loss at textureless regions, resulting in floaters or holes in the test views. Therefore, we add two geometric regularization losses to ensure the optimization proceeds in the right direction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "Monocular depth regularization loss. Monocular depth estimation faces challenges such as scale ambiguity, uneven scale distribution, and a lack of strict geometric consistency between depth maps. A simple linear transformation of monocular depth may destabilize 3D points and introduce distortions during alignment. To address this, the Pearson correlation coefficient (PCC) [8] can be used on small blocks of the depth map to extract robust geometric priors from monocular depth rather than imposing strict consistency. Specifically, we randomly sample \\(N\\) non-overlapping patches to compute the monocular depth regularization loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.283, + 0.907, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\frac {1}{N} \\sum_ {i} ^ {N} 1 - P C C \\left(\\hat {\\boldsymbol {P}} _ {i}, \\boldsymbol {P} _ {\\text {m o n o , i}}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.907, + 0.459 + ], + "angle": 0, + "content": "where \\(\\hat{P}_i\\in \\mathbb{R}^{S^2}\\) denotes the \\(i\\)th patch of \\(\\hat{D}\\), and \\(P_{\\mathrm{mono,i}}\\in \\mathbb{R}^{S^2}\\) denotes the \\(i\\)th patch of \\(D_{mono}\\), with the patch size \\(S\\) being a hyperparameter. PCC, similar to normalized cross-correlation, measures the similarity between the two depth patches. This loss function encourages local consistency between rendered and monocular depth maps, improving alignment over small regions while handling local variations, thus enhancing depth estimation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.906, + 0.521 + ], + "angle": 0, + "content": "Normal consistency loss. Following RaDeGS [48], we measure the consistency between the normal directions computed from the Gaussians and those derived from the depth map:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.53, + 0.906, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n c} = \\sum \\left(1 - \\hat {\\boldsymbol {N}} _ {i j} \\cdot \\vec {\\boldsymbol {N}} _ {i j}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.565, + 0.906, + 0.613 + ], + "angle": 0, + "content": "where \\(\\hat{N}_{ij}\\) is the rendered normal at pixel(i, j) from Eq. (7), \\(\\vec{N}_{ij}\\) is the normal estimated by the gradient of the depth map at pixel(i, j)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.613, + 0.906, + 0.643 + ], + "angle": 0, + "content": "Together with \\(\\mathcal{L}_{\\mathrm{n}}\\) and \\(\\mathcal{L}_{\\mathrm{s}}\\) used in the tangent space densification, our final loss function is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.654, + 0.905, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\mathrm {n}} + \\lambda_ {\\mathrm {s}} \\mathcal {L} _ {\\mathrm {s}} + \\lambda_ {\\mathrm {n c}} \\mathcal {L} _ {\\mathrm {n c}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.906, + 0.727 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{rgb}}\\) is the photometric loss proposed in 3DGS [15]. We set \\(\\lambda_{\\mathrm{d}} = 0.1\\), \\(\\lambda_{\\mathrm{n}} = 0.5\\), \\(\\lambda_{\\mathrm{s}} = 100.0\\) and \\(\\lambda_{\\mathrm{nc}} = 0.1\\) in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.646, + 0.757 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.87 + ], + "angle": 0, + "content": "In this section, we first demonstrate our IndoorGS's SOTA performance. We compare Gaussian-based methods, known for their high-quality rendering and fast training and rendering speed. We compare our method with 3DGS [15], GOF [47], RaDeGS [48] and PGSR [5]. And PGSR is the state-of-the-art surface reconstruction approach. We then analyze the contribution of each proposed component." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Dataset. We evaluate the performance of our approach on geometry reconstruction and rendering quality using 14" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "849" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.907, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.443, + 0.908, + 0.486 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison for geometry on Replica(the left two columns), Scannet++(the middle two columns), and Deep Blending(the right two columns) datasets. The experimental results demonstrate IndoorGS's robustness performance in less textured areas and ability to capture fine details as well." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.499, + 0.885, + 0.59 + ], + "angle": 0, + "content": "
ReplicaScannet++Deep BlendingAverage on all scenes
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓Time
3DGS39.520.9680.12129.460.9180.15229.410.9030.24332.790.9290.17210m
GOF40.980.9800.07429.820.9190.15329.210.9040.24633.330.9340.1582h
RaDeGS41.040.9800.07430.080.9210.14729.460.9070.24333.530.9360.15515m
PGSR40.020.9740.10129.990.9200.15429.220.8930.25533.080.9290.17025m
Ours41.110.9800.07330.450.9280.13130.060.9100.23533.880.9390.14635m
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.599, + 0.908, + 0.625 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison results of rendering quality for novel view synthesis on the Replica, Scannet++, and Deep Blending datasets. \"Red\" denote the best results, \"yellow\" denote the second best results. Test results on multiple datasets show that IndoorGS outperforms the other three methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.483, + 0.697 + ], + "angle": 0, + "content": "indoor scenes from publicly available datasets. This includes 8 scenes from Replica [36], 4 scenes from ScanNet++ [43] and 2 scenes from Deep Blending [11]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.483, + 0.806 + ], + "angle": 0, + "content": "Implementation details. We choose RaDeGS [48] as our baseline. Since our initialization points are denser than vanilla 3DGS, the densification stops earlier at 12,000 iterations. To extract the final mesh, we apply Truncated Signed Distance Function (TSDF) [27] surface reconstruction, then use Marching Cube [40] to extract mesh. All experiments are conducted on a single Nvidia RTX4080 GPU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Metrics. We chose three widely used image evaluation metrics for rendering evaluation: PSNR, SSIM, and LPIPS [49]. For mesh evaluation, we follow the evaluation protocol from monosdf [45] and Go-surf [38], report Chamfer- \\(L_{1}\\) distance \\((\\mathrm{C - }L_{1})\\) , F-score (F1) with a threshold of 5cm and Normal Consistency (NC)." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.647, + 0.912, + 0.743 + ], + "angle": 0, + "content": "
ReplicaScannet++Average on all scenes
C-L1↓F1↑NC↑C-L1↓F1↑NC↑C-L1↓F1↑NC↑
GOF0.0860.5760.7970.1050.4790.7020.0950.5280.749
RaDeGS0.0790.6000.8620.0960.4970.7610.0870.5490.811
PGSR0.0490.8180.8950.0790.6250.7780.0640.7220.837
Ours0.0360.8250.9400.0400.7760.8770.0380.8010.909
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.752, + 0.908, + 0.801 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison results of Chamfer distance, F-Score and Normal Consistency for reconstruction on the Replica and Scannet++ datasets. IndoorGS achieves the highest reconstruction accuracy across all metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.693, + 0.849 + ], + "angle": 0, + "content": "5.1. Mesh Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "IndoorGS demonstrates superior performance in indoor scene reconstruction, achieving the highest scores across all evaluated metrics—Chamfer Distance, F-Score, and Nor" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "850" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.905, + 0.278 + ], + "angle": 0, + "content": "
MethodGeometric CuesGeometric-cue-guided ADCLossMeshNovel View Synthesis
LineSfMPlaneLineSfMPlane\\(L_d\\)\\(L_n\\)\\(L_{nc}\\)C-L1↓F1↑NC↑PSNR↑SSIM↑LPIPS↓
w/o Line & Plane & SfM cues------✓✓✓0.0550.6810.85930.3080.9240.137
w/o Line & Plane cues-✓--✓-✓✓✓0.0510.7170.78930.5130.9270.135
w/o Plane cues✓✓-✓✓-✓✓✓0.0500.7280.86230.5830.9270.134
w/o All customized ADC✓✓✓---✓✓✓0.0520.7110.83930.4990.9260.135
w/o Line ADC✓✓✓-✓✓✓✓✓0.0510.7200.84830.5950.9260.135
w/o SfM ADC✓✓✓✓-✓✓✓✓0.0400.7710.87530.5730.9270.133
w/o Plane ADC✓✓✓✓✓-✓✓✓0.0460.7410.87130.3900.9250.137
w/o \\(L_d\\)✓✓✓✓✓✓-✓✓0.0440.7620.86430.5530.9260.134
w/o \\(L_n\\)✓✓✓✓✓✓✓-✓0.0510.7250.84830.5100.9260.137
w/o \\(L_{nc}\\)✓✓✓✓✓✓✓✓-0.0440.7440.80830.5050.9250.136
Full model✓✓✓✓✓✓✓✓✓0.0400.7760.87730.4530.9280.131
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.288, + 0.907, + 0.315 + ], + "angle": 0, + "content": "Table 3. Ablation study results. Geometric cues significantly improve reconstruction quality. We report mesh and novel view synthesis metrics on the Scannet++ dataset. Adding the SfM cues means using the filtered SfM points instead, and using \\( L_{n} \\) implies incorporating both \\( L_{n} \\) and \\( L_{s} \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.339, + 0.482, + 0.492 + ], + "angle": 0, + "content": "mal Consistency, on all of the Replica, Scannet++, and Deep Blending datasets as shown in Tab. 2 and Fig. 5. This consistent excellence highlights its effectiveness in accurately reconstructing complex indoor environments. Since the Deep Blending dataset does not include directly provided mesh ground truth (GT), we do not perform quantitative metrics calculations. In contrast to other methods, which often result in uneven surfaces in weakly textured regions and overly smooth surfaces along edges, IndoorGS achieves a more accurate and detailed reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.5, + 0.308, + 0.516 + ], + "angle": 0, + "content": "5.2. Rendering Comparison" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.522, + 0.483, + 0.84 + ], + "angle": 0, + "content": "To evaluate the rendering quality of our method, we compared the test-view rendering results of our approach with those from vanilla 3DGS, GOF, RaDeGS, and PGSR on Replica, ScanNet++, and Deep Blending datasets. As shown in Tab. 1, Our method outperforms all of the above Gaussian-based reconstruction methods in terms of novel view synthesis, achieving an average PSNR that is 0.35 dB higher than the second-best method across the three datasets. Notably, following vanilla 3DGS, our test set is obtained by sequential sampling from the original dataset. The test views are positioned close to the training views in indoor scenes. In this case, despite the model overfitting the training views and the geometry inaccurately estimated, the PSNR of the rendered images from the test views can remain relatively high, which undermines the effectiveness of the test set in evaluating the true capability of the model, as it fails to challenge the model with sufficiently diverse perspectives. We provide a detailed experimental analysis of this issue in the appendix. Additionally, our method takes only 10 minutes longer than PGSR despite including the 3D line map construction and planar cues extraction process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.255, + 0.864 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We conducted an ablation study on the ScanNet++ dataset to test the impact of each component we introduced." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.908, + 0.581 + ], + "angle": 0, + "content": "Specifically, we analyzed the effects of geometric cues, geometric-cue-guided ADC strategies, depth regularization, normal regularization, and normal consistency. The results are presented in Tab. 3. These results show that our full model performs best across all three mesh metrics and the final two NVS metrics (SSIM and LPIPS). Notably, as shown in the first three rows of the table, the step-by-step addition of geometric cues and the corresponding ADC strategy produces incremental performance improvement. Rows 4 to 7 further illustrate that introducing various ADC strategies after adding geometric cues results in further performance enhancement, and line-cue-guided ADC has the most significant impact among the three strategies. The final three rows show the positive effects of the three loss functions, with normal regularization providing the greatest enhancement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.601, + 0.634, + 0.617 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.795 + ], + "angle": 0, + "content": "In this paper, we propose IndoorGS, an indoor scene reconstruction method based on 3DGS. Our IndoorGS utilizes geometric cues such as filtered SfM points, line segments, and weakly textured planar surfaces, commonly found in indoor environments, as input data and takes a geometric-cues-guided ADC strategy for Gaussian scene optimization. We validate our method's rendering and reconstruction quality on the Replica, ScanNet++, and Deep Blending datasets. Experimental results demonstrate that our IndoorGS outperforms existing Gaussian-based techniques in geometric reconstruction accuracy and rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.814, + 0.666, + 0.831 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Project supported by the National Natural Science Foundation of China (Grant No. 62302174). The computation is completed in the HPC Platform of Huazhong University of Science and Technology." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.958 + ], + "angle": 0, + "content": "851" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Jiayang Bai, Letian Huang, Jie Guo, Wen Gong, Yuanqi Li, and Yanwen Guo. 360-gs: Layout-guided panoramic gaussian splatting for indoor roaming. arXiv preprint arXiv:2402.00763, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.323 + ], + "angle": 0, + "content": "[3] Danpeng Chen, Nan Wang, Runsen Xu, Weijian Xie, Hujun Bao, and Guofeng Zhang. Rnin-vio: Robust neural inertial navigation aided visual-inertial odometry in challenging scenes. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 275-283. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] Danpeng Chen, Shuai Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Hujun Bao, and Guofeng Zhang. Vip-slam: An efficient tightly-coupled rgb-d visual inertial planar slam. In 2022 International Conference on Robotics and Automation (ICRA), pages 5615-5621. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[5] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[6] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[7] Yutong Chen, Marko Mihajlovic, Xiyi Chen, Yiming Wang, Sergey Prokudin, and Siyu Tang. Splatformer: Point transformer for robust 3d gaussian splatting, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.606 + ], + "angle": 0, + "content": "[8] Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. Pearson correlation coefficient. Noise reduction in speech processing, pages 1-4, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[9] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[10] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[11] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[12] Manuel Hofer, Michael Maurer, and Horst Bischof. Line3d: Efficient 3d scene abstraction for the built environment. In Pattern Recognition: 37th German Conference, GCPR 2015, Aachen, Germany, October 7-10, 2015, Proceedings 37, pages 237-248. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Manuel Hofer, Michael Maurer, and Horst Bischof. Efficient 3d scene abstraction using line segments. Computer Vision and Image Understanding, 157:167-178, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[14] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.906, + 0.248 + ], + "angle": 0, + "content": "[16] Jiyeop Kim and Jongwoo Lim. Integrating meshes and 3d gaussians for indoor scene reconstruction with sam mask guidance. arXiv preprint arXiv:2407.16173, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[18] Yanyan Li, Raza Yunus, Nikolas Brasch, Nassir Navab, and Federico Tombari. Rgb-d slam with structural regularities. In 2021 IEEE international conference on Robotics and automation (ICRA), pages 11581-11587. IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[19] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering. arXiv preprint arXiv:2403.11324, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.908, + 0.489 + ], + "angle": 0, + "content": "[20] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.906, + 0.546 + ], + "angle": 0, + "content": "[21] Shuhong Liu, Heng Zhou, Liuzhuozheng Li, Yun Liu, Tianchen Deng, Yiming Zhou, and Mingrui Li. Structure gaussian slam with Manhattan world hypothesis. arXiv preprint arXiv:2405.20031, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.574 + ], + "angle": 0, + "content": "[22] G Lowe. Sift-the scale invariant feature transform. Int. J, 2 (91-110):2, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.906, + 0.646 + ], + "angle": 0, + "content": "[23] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[24] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.906, + 0.76 + ], + "angle": 0, + "content": "[25] Run Luo, Zikai Song, Lintao Ma, Jinlin Wei, Wei Yang, and Min Yang. Diffusiontrack: Diffusion model for multi-object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3991-3999, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[26] Pierre Moulon, Pascal Monasse, Romuald Perrot, and Renaud Marlet. Openmvg: Open multiple view geometry. In Reproducible Research in Pattern Recognition: First International Workshop, RRPR 2016, Cancún, Mexico, December 4, 2016, Revised Selected Papers 1, pages 60-74. Springer, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[27] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and track" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "852" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "ing. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[28] Rémi Pautrat, Juan-Ting Lin, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Sold2: Self-supervised occlusion-aware line description and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11368-11378, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[29] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17327-17336, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[30] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[31] Radu Bogdan Rusu and Steve Cousins. 3d is here: Point cloud library (pcl). In 2011 IEEE international conference on robotics and automation, pages 1-4. IEEE, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.483, + 0.425 + ], + "angle": 0, + "content": "[32] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[33] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint arXiv:2312.05133, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.538 + ], + "angle": 0, + "content": "[34] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International journal of computer vision, 80:189-210, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[35] Xiaowei Song, Jv Zheng, Shiran Yuan, Huan-ang Gao, Jingwei Zhao, Xiang He, Weihao Gu, and Hao Zhao. Saags: Scale-adaptive gaussian splatting for training-free antialIASing. arXiv preprint arXiv:2403.19615, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[36] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[37] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[38] Jingwen Wang, Tymoteusz Bleja, and Lourdes Agapito. Gosurf: Neural feature grid optimization for fast, high-fidelity rgb-d surface reconstruction. In 2022 International Conference on 3D Vision (3DV), pages 433-442. IEEE, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[39] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free-view synthesis of indoor scenes. arXiv preprint arXiv:2405.17958, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[40] LORENSEN WE. Marching cubes: A high resolution 3d surface construction algorithm. Computer graphics, 21(1): 7-12, 1987. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[41] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[42] Haodong Xiang, Xinghui Li, Xiansong Lai, Wanting Zhang, Zhichao Liao, Kai Cheng, and Xueping Liu. Gaussianroom: Improving 3d gaussian splatting with sdf guidance and monocular cues for indoor scene reconstruction. arXiv preprint arXiv:2405.19671, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[43] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[44] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9043-9053, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[45] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.417, + 0.905, + 0.485 + ], + "angle": 0, + "content": "[46] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.542 + ], + "angle": 0, + "content": "[47] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.544, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[48] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.905, + 0.726 + ], + "angle": 0, + "content": "[50] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.728, + 0.905, + 0.783 + ], + "angle": 0, + "content": "[51] Hang Zhou, Jiale Cai, Yuteng Ye, Yonghui Feng, Chenxing Gao, Junqing Yu, Zikai Song, and Wei Yang. Video anomaly detection with motion and appearance guided patch diffusion model. arXiv preprint arXiv:2412.09026, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.785, + 0.905, + 0.853 + ], + "angle": 0, + "content": "[52] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.855, + 0.905, + 0.897 + ], + "angle": 0, + "content": "[53] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "853" + } + ] +] \ No newline at end of file diff --git a/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/848b3be8-e7e7-4fb9-82b5-d4028969c01d_origin.pdf b/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/848b3be8-e7e7-4fb9-82b5-d4028969c01d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3ae75388f289ced1b2eb45870c289b23784d3eb1 --- /dev/null +++ b/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/848b3be8-e7e7-4fb9-82b5-d4028969c01d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f582009418e28c69b3ccbbc695358d9f09895693d076feac9dd722f3e45bded7 +size 5458268 diff --git a/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/full.md b/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..681912669f5030441e810d5be47b3eaf2475b520 --- /dev/null +++ b/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/full.md @@ -0,0 +1,345 @@ +# IndoorGS: Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction + +Cong Ruan Yuesong Wang* Tao Guan + +School of Computer Science & Technology, Huazhong University of Science and Technology, China {rc22, yuesongwang, qd.gt}@hust.edu.cn + +Bin Zhang Shenzhen Smart City Technology Development Group Co., Ltd., China + +zhangbin@smartcitysz.com + +Lili Ju +Department of Mathematics, University of South Carolina, USA + +ju@math.sc.edu + +![](images/86286184a81fa774ef4661008d3ba522052961a5bf96479b596ef63ebb4bad3b.jpg) + +![](images/a383e9358108d975821e845dc09ce6b7c15b1cb5e20e79316dc07a49954a2c72.jpg) +GOF + +![](images/afc5f02bb84e4364acfbc1b7e5ca1bf4c1edaa9334d1bd29d36a12ec3f8dd5de.jpg) + +![](images/f514a39c5ccc3b8f36aee53f46497a711199eb45ba54e6c3cd0aa41c287f74ec.jpg) +RaDeGS + +![](images/82d714b2771d7b9c57300e2b8531079a8e47abb6743c3c278306faedc403b6a8.jpg) + +![](images/6dd97cc05486b5981e971fff87df30dee8bf49442b5deb965f7db0ca011eb623.jpg) +PGSR +Figure 1. Our IndoorGS achieves state-of-the-art (SOTA) performance in indoor scene reconstruction. Compared to other methods that produce poor geometric results, our approach generates well-structural geometries while preserving all details. We can achieve superior rendering quality thanks to our precise geometry estimation. + +![](images/9c81823ce147160df633083428e050a13e4b4d44d151a356c65c6aa5e24376c7.jpg) + +![](images/f8e8c1c2b0debe1cd53f932432771077d304be39ce79c5f28ac3a4d7e1fcbd9c.jpg) +Ours + +# Abstract + +3D Gaussian Splatting (3DGS) has shown impressive performance in scene reconstruction, offering high rendering quality and rapid rendering speed with short training time. However, it often yields unsatisfactory results when applied to indoor scenes due to its poor ability to learn geometries without enough textural information. In this paper, we propose a new 3DGS-based method "IndoorGS", that leverages the commonly found yet important geometric + +cues in indoor scenes to improve the reconstruction quality. Specifically, we first extract 2D lines from input images and fuse them into 3D line cues via feature-based matching, which can provide a structural understanding of the target scene. We then apply the statistical outlier removal to refine Structure-from-Motion (SfM) points, ensuring robust cues in texture-rich areas. Based on these two types of cues, we further extract reliable 3D plane-like cues for textureless regions. Such geometric information will be utilized not only for initialization but also in the realization of a geometric-cue-guided adaptive density control (ADC) strategy. The proposed ADC approach is grounded in the principle of + +divide-and-conquer and optimizes the use of each type of geometric cues to enhance overall reconstruction performance. Extensive experiments on multiple indoor datasets show that our method can deliver much more accurate geometry and higher rendering quality for indoor scenes than existing 3DGS approaches. + +# 1. Introduction + +Indoor scene reconstruction is one of common tasks in computer vision, with broad applications in virtual reality, gaming, autonomous driving, robotics, and so on [25, 51]. Recently, 3D Gaussian Splitting (3DGS) [15] has emerged as a promising scene reconstruction method, which introduces a hybrid optimization method with Gaussian primitives similar to point clouds and combines differentiable backpropagation with adaptive density control. This method has demonstrated superior efficacy, exhibiting expedited training process, accelerated rendering speed, and high-quality view synthesis, particularly in highly textured outdoor environments. However, it often performs poorly when applied to indoor environments, which mainly attributes to the fact that 3DGS is not capable of reconstructing correct geometry when facing textureless regions, as there are too many potential combinations of geometry and appearance that can satisfy the training images. + +Although there has been a large amount of research works [5, 9, 10, 14, 47, 48] devoted to improving the geometry inference ability of 3DGS-based methods, their reconstruction quality is still unsatisfactory when dealing with indoor scenes, resulting in holes or floaters in the rendering images, as shown in Fig. 1. There are many local optima during the training process, and without strong constraints and correct guidance, 3DGS could be easily trapped by these local optima. What's even worse is that the initialization of 3DGS relies entirely on SfM points, which are inherently sparse and noisy in regions with weak textures, and the lack of reliable seeds can lead to blind exploration of 3DGS in these regions, further exacerbating the likelihood of being trapped by the local optima. To improve the performance of 3DGS for indoor scene reconstruction, we need to, on the one hand, provide better reliable seeds and, on the other hand, add strong geometric constraints to help the optimization move towards the global optima. + +To obtain better initial seeds, we propose to extract more reliable geometric cues commonly found in the indoor scene and use them to help initialize the 3DGS. Specifically, we first extract 2D line segments and perform feature-based matching on these 2D lines to obtain 3D line cues. Matching only line segments can drastically reduce the computational complexity and the matching ambiguity, allowing us to get accurate and noise-free 3D line cues to help 3DGS understand the scene's construction. While SfM [26, 32, 34] + +points can provide good geometric priors in texture-rich regions, they usually behave as outliers in textureless or repeating-textured areas and can also trouble the reconstruction. Therefore, we filter the SfM points via Statistical Outlier Removal (SOR) and use the remaining points as reliable geometric cues in texture-rich regions. We further extend these cues to the weakly textured planar-like regions commonly existing in indoor scenes, such as walls, floors, ceilings, and furniture surfaces. We first extract the weakly textured regions from the 2D images and back-project them into 3D space using aligned monocular depths. We then analyze the spatial relationship between the extracted 3D lines and these back-projected planes, allowing us to extract reliable 3D plane-like cues. The above three types of geometric cues are fed to 3DGS for initialization. + +Based on the distinctive characteristics of these geometric cues, we further divide the Gaussians into three types, namely line Gaussians, plane-like Gaussians, and SfM Gaussians, and design a geometric-cue-guided adaptive density control strategy to enhance the optimization of 3DGS marching towards global optima. As 3D lines are typically located at object edges with high spatial accuracy, we apply a tangent space densification method to these line Gaussians, encouraging them to expand to complete the object. For plane-like Gaussians, since the initial seeds from 3D plane-like region cues may introduce spatial offsets and redundancies, we disable the splitting and cloning operations for these Gaussians, focusing solely on optimizing Gaussian parameters and pruning processes. For SfM geometric cues, because SfM points are often noisier than 3D lines, we implement a densification strategy that prioritizes line Gaussian over SfM Gaussian. + +In summary, our main contributions are as follows: + +- We propose a geometric cue extraction method for obtaining different types of reliable geometric priors from indoor scenes to guide the reconstruction. +- We classify Gaussian primitives according to their geometric types and design a geometric-cue-guided adaptive density control strategy to better utilize the unique properties of each type of geometric cue. +- Our proposed IndoorGS demonstrates superior performance in both the rendering quality and the accuracy of extracted geometry across various indoor datasets. + +# 2. Related Work + +3DGS has been as a prominent 3D reconstruction technique in the past couple of years due to its training efficiency, rendering performance, and impressive results across various 3D tasks [24, 33, 37, 41]. The core of 3DGS lies in a carefully designed rasterization [53] method that efficiently leverages GPU computation by representing scenes as collections of Gaussian primitives. Despite the remarkable success of 3DGS and its subsequent modifica + +![](images/4eb2bc40b68ad2f362d4009f638cfb33739f4dd9bd934f9f926831cda4a6b27c.jpg) +Figure 2. Overview of the proposed IndoorGS. (a) The left part illustrates the extraction process of geometric cues, i.e., line cues, plane-like cues, and SfM cues, which are then used to initialize 3DGS. (b) The right part depicts the optimization process, where the 3DGS is further optimized using geometric-cue-guided adaptive density control and supervised by a series of loss functions. + +tions [23, 35, 46, 50] in novel view synthesis, it faces significant challenges in surface reconstruction, mainly due to the discrete and unstructured nature of its Gaussian primitives. Recent research has sought to enhance 3DGS-based surface reconstruction [5, 9, 10, 14, 47, 48], yielding improvements in geometrically rich environments with strong textures. However, these methods still struggle in weakly textured settings, such as indoor scenes, which has driven growing interest in 3DGS reconstruction for indoor environments. GaussianRoom [42] integrates neural signed distance functions (SDF) with 3DGS, incorporating monocular normal and edge priors to enhance indoor scene details. Kim et al.[16] proposed an approach that employs 3D Gaussians for object reconstruction while using meshes to capture the room layout. 360-GS [1] leverages four panoramic images to initialize 3D Gaussians based on room layout and depth maps derived from panoramas for comprehensive indoor reconstruction. FreeSplat [39] applies 2D backbones and a cost volume for multi-view aggregation, while MGSLAM [21] utilizes the Manhattan World hypothesis to improve geometric accuracy and completeness in real-time SLAM [2-4] reconstructions. Despite the contributions, these methods usually require specialized inputs or are inefficient, exhibiting limitations in texture-poor environments. Indoor scenes are rich in geometric cues like lines and planes, which are useful for 3DGS-based reconstructions. However, SfM often encounters issues in these environments due to large textureless surfaces and repetitive elements. This is where 3D line map reconstruction approaches [12, 13, 20] show natural advantages in overcoming weak and repetitive textures common in indoor settings. By fully leveraging these geometric cues, 3DGS is expected to better handle the unique challenges of indoor + +scene reconstruction. + +# 3. Preliminaries + +# 3.1. 3D Gaussian Splatting + +3DGS [15] explicitly represents a 3D scene with a set of Gaussian primitives. Each of them is parameterized via a 3D covariance matrix $\pmb{\Sigma}$ and mean $\pmb{\mu}$ : + +$$ +G (\boldsymbol {p}) = \exp \left(- \frac {1}{2} \left(\boldsymbol {p} - \boldsymbol {\mu}\right) ^ {\intercal} \boldsymbol {\Sigma} ^ {- 1} \left(\boldsymbol {p} - \boldsymbol {\mu}\right)\right), \tag {1} +$$ + +where $\pmb{\Sigma}$ can be further factorized into $\mathbf{RSS}^{\intercal}\mathbf{R}^{\intercal}$ with a scaling matrix $\mathbf{S}$ and a rotation matrix $\mathbf{R}$ . $\alpha_{i}$ is the blending coefficient for a Gaussian with center $\pmb{\mu}_{i}$ in screen space: + +$$ +\alpha_ {i} = o _ {i} \cdot \exp \left(- \frac {1}{2} (\pmb {p} - \pmb {\mu} _ {i}) ^ {\intercal} (\pmb {\Sigma} _ {i} ^ {\prime}) ^ {- 1} (\pmb {p} - \pmb {\mu} _ {i})\right). (2) +$$ + +where $o_i$ is the Gaussian opacity, the 3D Gaussian is transformed into the image coordinates with world-to-camera transform matrix $\mathbf{W}$ and projected onto pixel space with the Jacobian of the affine approximation for projection matrix $\mathbf{J}$ , then the 2D covariance matrix $\boldsymbol{\Sigma}^{\prime}$ is computed as $\boldsymbol{\Sigma}^{\prime} = \mathbf{J}\mathbf{W}\boldsymbol{\Sigma}\mathbf{W}^{\mathrm{T}}\mathbf{J}^{\mathrm{T}}$ . The rendering result is computed using point-based alpha-blending on each ray: + +$$ +\hat {C} = \sum_ {i \in N} c _ {i} \alpha_ {i} T _ {i}, \quad \text {w h e r e} T _ {i} = \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}) \tag {3} +$$ + +where $T_{i}$ is the accumulated transmittance at pixel location $\pmb{p}$ , $\mathrm{c}_i$ is the view-dependent color defined by spherical harmonics (SH) coefficients, and $N$ is the number of Gaussian primitives. A photometric loss function optimizes the Gaussian parameters to reduce the discrepancies between the rendered images and the actual observational data. + +# 3.2. Rasterizing Depth and Normal + +Following RaDeGS [48], the center $\mu_{i}$ of a Gaussian $G_{i}$ is initially projected into the camera space as $\mu_{i}^{\prime}$ . The depth of each pixel correspondent to the Gaussian is computed as: + +$$ +d = z _ {i} ^ {\prime} + \boldsymbol {p} \left( \begin{array}{c} \Delta x \\ \Delta y \end{array} \right), \quad \boldsymbol {\mu} _ {i} ^ {\prime} = \left[ \begin{array}{c} x _ {i} ^ {\prime} \\ y _ {i} ^ {\prime} \\ z _ {i} ^ {\prime} \end{array} \right] = \boldsymbol {W} \boldsymbol {\mu} _ {i} + t, (4) +$$ + +where $z_{i}^{\prime}$ represents the depth of the Gaussian center, $\Delta x = x_{i}^{\prime} - x$ and $\Delta y = y_{i}^{\prime} - y$ denote the relative pixel positions. The vector $\pmb{p}$ is determined by the Gaussian parameters $[\pmb{W},\pmb{t}] \in \mathbb{R}^{3\times 4}$ . Per-pixel z-depth estimates $\hat{D}$ are rendered using the discrete volume rendering approximation similar to color values: + +$$ +\hat {D} = \sum_ {i \in N} d _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {5} +$$ + +where $d_{i}$ is the $i^{\mathrm{th}}$ Gaussian's depth according to the Eq. (4). The normal direction of each projected Gaussian is computed as follows: + +$$ +\hat {\boldsymbol {n}} = - \boldsymbol {J} ^ {\intercal} \left(\frac {z _ {i} ^ {\prime}}{z _ {i}} \boldsymbol {p}, 1\right) ^ {\intercal} \tag {6} +$$ + +Normals are then alpha-blended, resulting in a single, unified per-pixel normal estimate: + +$$ +\hat {\boldsymbol {N}} = \sum_ {i \in N} \hat {\boldsymbol {n}} _ {i} \alpha_ {i} T _ {i}. \tag {7} +$$ + +# 4. Proposed Method + +Fig. 2 gives an overview of our methods. We first extract geometric cues, i.e., line, plane, and SfM point cues, from the input images (Section 4.1). Then, we divide Gaussians into three types according to the geometric cues and treat them differently in our geometric-cue-guided adaptive density control (Section 4.2). The training process is supervised by color and geometric loss (Section 4.3). + +# 4.1. Geometric Cues Extraction + +SfM points are often too sparse to guide the reconstruction of indoor scenes and sometimes contain geometric errors, which can cause lots of floaters in the subsequent 3DGS reconstruction process. Therefore, we need to provide the 3DGS with sufficient and accurate geometric cues. As depicted in Fig. 2, there are three types of geometric cues that we use to guide the reconstruction. + +Geometric Cues from 3D lines. In contrast to sparse points from SfM, which is structureless, 3D line segments can concisely encode the high-level layout, as they often delineate the main structural elements of the scene. We follow + +the idea of LIMAP [20] to extract 3D lines, mainly consisting of 2D line extraction and line matching. Although rich 2D line segment features can be extracted from images using learning-based methods such as DeepLSD [29] or Sold2 [28], when encountering large continuous weakly textured or untextured areas, such as wall corner joint between white walls, they still have difficulties in obtaining distinctive 2D line segment features using only color information. However, we can easily extract distinctive line segment features from textureless regions on their normal maps, as shown in Fig. 3. + +![](images/b2d5a0b6a063fea8388cbb78d2651ef015aab20ee61a7efc3e522f286f6fb59d.jpg) + +![](images/f227e2694a21010988a00558ba27d9ee217a96b0b7e73d06c4ea8a26fce85f06.jpg) + +![](images/21a41c1c872ff3abc3f23c0b2c51b29b9061e7b89a35800ea4f2906dab2be1bd.jpg) +RGB + +![](images/1c19b040043e13a978e48120f3ca13047609a5966a376950486752d425422482.jpg) +Normal +Figure 3. The complementary effects of extracting structured lines from RGB images and normal maps. Top row: The line segments can be easily seen from the normal map, but extracting the corresponding line segments from the RGB image is difficult. Bottom row: The opposite case. + +However, at the beginning of the reconstruction, we do not know the scene's geometry and can not render the normal maps. Fortunately, nowadays, foundation models for monocular normal estimation can give us a good prior. Thus, we extracted 2D line segments from RGB images and monocular normal maps using the DeepLSD [29] method. Subsequently, the GlueStick [30] method was used to extract the line segments and match them. + +Finally, we obtain the 3D line segments for the scene. To capture color information along these segments, we extract the corresponding 2D line colors from all visible views of each 3D line, then perform multiview color averaging to achieve a consistent color representation, and sample 3D points on 3D line segments in fixed steps. + +Geometric Cues from SfM. The sparse points of SfM in indoor scenes often contain noise or outliers to repetitive structures, reflective surfaces, and limited textures. Although sparse 3D feature points are generally accurate in strong-texture regions, they are often misaligned in weak-texture regions, as 2D points are more challenging to match correctly across multiple views. Therefore, filtering SfM + +points is crucial, as erroneous inputs can be amplified during adaptive density control (ADC) in 3DGS, causing floater artifacts. To address this, we apply a statistical outlier removal filter (SOR) [31], which calculates the average distance between each point and its K nearest neighbors, removing points with abnormally high average distances. The remaining SfM points serve as reliable geometric cues in texture-rich regions. + +Geometric Cues from Plane-like Regions. The plane-like regions in indoor scenes are usually weakly textured walls, ceilings, etc. For detecting these weakly textured regions in images, we use the Segment Anything Model (SAM) [17] for image segmentation, then determine texture strength by counting the density of SIFT [22] points in each segmented region. + +It is difficult to calculate the depth of these regions due to the lack of texture information. However, the positions of the structural line clues on their edges are accurate enough. Thus, we can locate these plane-like cues using the 3D line cues with the help of the monocular depth estimation model. Specifically, we first obtain the monocular depth estimation from Metric3D [44]. While the monocular depth estimation can reflect the relative positions of lines and planes, it remains inconsistent with absolute depths. Thus, we estimate a relative per-image depth scale by aligning monocular depth with line cues and SfM cues using linear regression, which is then used to adjust the monocular depths. Based on the adjusted monocular depths, the weakly textured region is back-projected into 3D space to form a point cloud. To ensure the reliability of the back-projected point cloud, we then calculate the distances between sampling points from 3D line cues and the back-projected point cloud, based on which we filter out points exceeding a preset threshold. The retained points form the plane-like cues, giving reliable priors for weakly textured regions in 3D space. The whole process is described as Algorithm 1 in the supplementary material. + +# 4.2. Geometric-cue-guided ADC Strategies + +Besides extracting points from geometric cues for initialization, we further let these cues guide the reconstruction process. To achieve this, we design Geometric-cue-guided Adaptive Densification Control (ADC) strategies for the three Gaussian types based on the positional accuracy and significance of their respective 3D spatial locations. In the following, we denote the Gaussians from SfM cues, line cues, and plane-like cues as $G_{M}$ , $G_{L}$ , and $G_{S}$ , respectively. + +- Line Gaussian $G_{L}$ : They typically locate at object edges with high spatial accuracy. Thus, as discussed in detail below, we specifically design a tangent space densification method for them. +- SfM Gaussian $G_{M}$ : The ADC process for $G_{L}$ takes precedence over $G_{M}$ since $G_{M}$ are often noisier than $G_{L}$ . + +Specifically, we allow $G_{M}$ to split and clone freely between 500-700 iterations to ensure it occupies strong texture regions correctly. Between iterations 700 and 3500, no ADC process is applied to $G_{M}$ ; only the tangent space densification of $G_{L}$ occurs during this period. After 3500 iterations, the ADC process for $G_{M}$ resumes. + +- Plane-like Gaussian $G_{S}$ : Due to potential spatial offsets and redundancy of $G_{S}$ , we disable splitting and cloning operations for $G_{S}$ . Instead, only the parameters optimization and pruning processes are conducted for them. + +![](images/8600296d216d3b058c1cfc1f4f5c9aa590730db32304e9983ec2bed6169184b0.jpg) +Figure 4. Densification in the tangential space. Left: Cloning. When the Gaussian $G_{i}$ is cloned, an average gradient is added to its spatial location $\mu_{i}$ . The projection of this gradient in the normal direction (the correction factor) is then subtracted, resulting in the new cloned green Gaussian $G_{i+1}$ . Right: Splitting. When the Gaussian $G_{i}$ is split, a small random vector $v$ is generated and added to its position. The component of $v$ along the normal is set to 0, yielding the two newly split Gaussians $G_{i+1}$ . + +![](images/d617d2d7634461cd87bc46f3de155e2962cf08eb6c7fe8a72453f7d3d0c00111.jpg) + +Densification in the tangential space. The vanilla 3D Gaussian Splatting (3DGS) is traditionally densified by randomly splitting or cloning Gaussian primitives within their Gaussian fields. While this method is flexible and easy to implement, it encounters difficulties in indoor environments as this unconstrained densification strategy can result in artifacts like floaters in weakly textured regions. Inspired by GeoGS [19], we propose a similar tangent space densification scheme. However, unlike GeoGS, whose initial point cloud is derived from PlanarSLAM [18] and whose normal vectors can be directly obtained from a spatial fit, we do not simply reduce the third scale of the Gaussian primitive to a very small value as we do not know the initial normal vectors. Instead, we introduce a soft constraint that flattens the Gaussian ellipsoid, akin to the approach used in Neusg [6], we directly minimize the minimum scale factor $S_{i} = \text{diag}(s_{1}, s_{2}, s_{3})$ for each Gaussian that originates from 3D line points: + +$$ +\mathcal {L} _ {s} = \left\| \min \left(s _ {1}, s _ {2}, s _ {3}\right) \right\| _ {1}. \tag {8} +$$ + +In this case, Gaussian primitive becomes flat, forming dislike surfel, and the smallest scaling axis approximates the normal direction. Simultaneously, we align the Gaussians' normal with the monocular surface normal estimated by Metrci3D using the following normal constraints: + +$$ +\mathcal {L} _ {n} = \frac {1}{| \hat {N} |} \sum \left\| \hat {N} _ {i j} - N _ {\text {m o n o}, i j} \right\| _ {1}, \tag {9} +$$ + +where $\hat{N}_{ij}$ represents estimated normal values at pixel position $(i,j)$ , and $N_{\mathrm{mono},ij}$ represents monocular normal values at pixel position $(i,j)$ . We then define the geometric normal for a flat Gaussian primitive as follows: + +$$ +\hat {\mathbf {n}} _ {f l a t} = \boldsymbol {R} \cdot \operatorname {O n e H o t} (\arg \min (s _ {1}, s _ {2}, s _ {3})), \tag {10} +$$ + +where rotation matrix $\mathbf{R} \in SO(3)$ which is obtained from its quaternion $\mathbf{q}$ , OneHot(.) ∈ $\mathbb{R}^3$ returns a unit vector with zeros everywhere except at the position where the scaling $s_i = (s_1, s_2, s_3)$ is minimum. However, the orientation of the shortest axis can be ambiguous, as it may point either outward or inward relative to the surface. We leverage the relationship between the observation direction and the Gaussian normal to resolve this ambiguity. Specifically, the angle between the observation direction and the Gaussian normal should always be greater than 90 degrees. If this condition is not met, we simply invert the Gaussian normal vector by multiplying it by -1, ensuring the correct orientation. Then, we implement a tangent space densification strategy for these flat Gaussians, as shown in Fig. 4. + +When a cloning operation is performed, the position of the newly cloned Gaussian is determined by adding the average gradient to the position of the original Gaussian, along with a correction factor. The correction is calculated as follows: + +$$ +\boldsymbol {\mu} ^ {i + 1} = \boldsymbol {\mu} ^ {i} + \nabla \boldsymbol {\mu} ^ {i} - \hat {\mathbf {n}} _ {f l a t} \cdot \nabla \boldsymbol {\mu} ^ {i}, \tag {11} +$$ + +where $\pmb{\mu}^i$ is the position of the original Gaussian $G^i$ , $\nabla \pmb{\mu}^i$ is the accumulated average gradient of the original Gaussian $G_i$ , $\hat{\mathbf{n}}_{flat}$ is the normal of the original Gaussian $G_i$ . + +When a splitting operation is performed, the position of the newly split Gaussian is determined by adding an offset to the original Gaussian's position. This offset is drawn from a normal distribution with a mean of 0 and the three scales of the Gaussian as the standard deviation. We set the offset along the axis corresponding to the most miniature scale to zero, ensuring no movement occurs along that direction. + +The advantage of this approach is that the tangent space densification of $G_{L}$ enables better splitting and cloning along surfaces, effectively suppressing the issue of "floaters" caused by excessive splitting and cloning of erroneous seeds in $G_{S}$ . + +# 4.3. Loss Functions Definition + +Despite adding geometric guidance to adaptive density control, Gaussians may still drift away from the correct locations during parameter optimization due to the weak constraints from color consistency loss at textureless regions, resulting in floaters or holes in the test views. Therefore, we add two geometric regularization losses to ensure the optimization proceeds in the right direction. + +Monocular depth regularization loss. Monocular depth estimation faces challenges such as scale ambiguity, uneven scale distribution, and a lack of strict geometric consistency between depth maps. A simple linear transformation of monocular depth may destabilize 3D points and introduce distortions during alignment. To address this, the Pearson correlation coefficient (PCC) [8] can be used on small blocks of the depth map to extract robust geometric priors from monocular depth rather than imposing strict consistency. Specifically, we randomly sample $N$ non-overlapping patches to compute the monocular depth regularization loss as follows: + +$$ +\mathcal {L} _ {d} = \frac {1}{N} \sum_ {i} ^ {N} 1 - P C C \left(\hat {\boldsymbol {P}} _ {i}, \boldsymbol {P} _ {\text {m o n o , i}}\right), \tag {12} +$$ + +where $\hat{P}_i\in \mathbb{R}^{S^2}$ denotes the $i$ th patch of $\hat{D}$ , and $P_{\mathrm{mono,i}}\in \mathbb{R}^{S^2}$ denotes the $i$ th patch of $D_{mono}$ , with the patch size $S$ being a hyperparameter. PCC, similar to normalized cross-correlation, measures the similarity between the two depth patches. This loss function encourages local consistency between rendered and monocular depth maps, improving alignment over small regions while handling local variations, thus enhancing depth estimation accuracy. + +Normal consistency loss. Following RaDeGS [48], we measure the consistency between the normal directions computed from the Gaussians and those derived from the depth map: + +$$ +\mathcal {L} _ {n c} = \sum \left(1 - \hat {\boldsymbol {N}} _ {i j} \cdot \vec {\boldsymbol {N}} _ {i j}\right), \tag {13} +$$ + +where $\hat{N}_{ij}$ is the rendered normal at pixel(i, j) from Eq. (7), $\vec{N}_{ij}$ is the normal estimated by the gradient of the depth map at pixel(i, j). + +Together with $\mathcal{L}_{\mathrm{n}}$ and $\mathcal{L}_{\mathrm{s}}$ used in the tangent space densification, our final loss function is defined as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {r g b}} + \lambda_ {\mathrm {d}} \mathcal {L} _ {\mathrm {d}} + \lambda_ {\mathrm {n}} \mathcal {L} _ {\mathrm {n}} + \lambda_ {\mathrm {s}} \mathcal {L} _ {\mathrm {s}} + \lambda_ {\mathrm {n c}} \mathcal {L} _ {\mathrm {n c}}, \tag {14} +$$ + +where $\mathcal{L}_{\mathrm{rgb}}$ is the photometric loss proposed in 3DGS [15]. We set $\lambda_{\mathrm{d}} = 0.1$ , $\lambda_{\mathrm{n}} = 0.5$ , $\lambda_{\mathrm{s}} = 100.0$ and $\lambda_{\mathrm{nc}} = 0.1$ in our experiments. + +# 5. Experiments + +In this section, we first demonstrate our IndoorGS's SOTA performance. We compare Gaussian-based methods, known for their high-quality rendering and fast training and rendering speed. We compare our method with 3DGS [15], GOF [47], RaDeGS [48] and PGSR [5]. And PGSR is the state-of-the-art surface reconstruction approach. We then analyze the contribution of each proposed component. + +Dataset. We evaluate the performance of our approach on geometry reconstruction and rendering quality using 14 + +![](images/b134ebe4a2a7e3aafd730e077a87fb78f811b87a76b78ed11547bffd359093d8.jpg) +Figure 5. Qualitative comparison for geometry on Replica(the left two columns), Scannet++(the middle two columns), and Deep Blending(the right two columns) datasets. The experimental results demonstrate IndoorGS's robustness performance in less textured areas and ability to capture fine details as well. + +
ReplicaScannet++Deep BlendingAverage on all scenes
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓Time
3DGS39.520.9680.12129.460.9180.15229.410.9030.24332.790.9290.17210m
GOF40.980.9800.07429.820.9190.15329.210.9040.24633.330.9340.1582h
RaDeGS41.040.9800.07430.080.9210.14729.460.9070.24333.530.9360.15515m
PGSR40.020.9740.10129.990.9200.15429.220.8930.25533.080.9290.17025m
Ours41.110.9800.07330.450.9280.13130.060.9100.23533.880.9390.14635m
+ +indoor scenes from publicly available datasets. This includes 8 scenes from Replica [36], 4 scenes from ScanNet++ [43] and 2 scenes from Deep Blending [11]. + +Implementation details. We choose RaDeGS [48] as our baseline. Since our initialization points are denser than vanilla 3DGS, the densification stops earlier at 12,000 iterations. To extract the final mesh, we apply Truncated Signed Distance Function (TSDF) [27] surface reconstruction, then use Marching Cube [40] to extract mesh. All experiments are conducted on a single Nvidia RTX4080 GPU. + +Metrics. We chose three widely used image evaluation metrics for rendering evaluation: PSNR, SSIM, and LPIPS [49]. For mesh evaluation, we follow the evaluation protocol from monosdf [45] and Go-surf [38], report Chamfer- $L_{1}$ distance $(\mathrm{C - }L_{1})$ , F-score (F1) with a threshold of 5cm and Normal Consistency (NC). + +Table 1. Quantitative comparison results of rendering quality for novel view synthesis on the Replica, Scannet++, and Deep Blending datasets. "Red" denote the best results, "yellow" denote the second best results. Test results on multiple datasets show that IndoorGS outperforms the other three methods. + +
ReplicaScannet++Average on all scenes
C-L1↓F1↑NC↑C-L1↓F1↑NC↑C-L1↓F1↑NC↑
GOF0.0860.5760.7970.1050.4790.7020.0950.5280.749
RaDeGS0.0790.6000.8620.0960.4970.7610.0870.5490.811
PGSR0.0490.8180.8950.0790.6250.7780.0640.7220.837
Ours0.0360.8250.9400.0400.7760.8770.0380.8010.909
+ +Table 2. Quantitative comparison results of Chamfer distance, F-Score and Normal Consistency for reconstruction on the Replica and Scannet++ datasets. IndoorGS achieves the highest reconstruction accuracy across all metrics. + +# 5.1. Mesh Comparison + +IndoorGS demonstrates superior performance in indoor scene reconstruction, achieving the highest scores across all evaluated metrics—Chamfer Distance, F-Score, and Nor + +
MethodGeometric CuesGeometric-cue-guided ADCLossMeshNovel View Synthesis
LineSfMPlaneLineSfMPlane\(L_d\)\(L_n\)\(L_{nc}\)C-L1↓F1↑NC↑PSNR↑SSIM↑LPIPS↓
w/o Line & Plane & SfM cues------✓✓✓0.0550.6810.85930.3080.9240.137
w/o Line & Plane cues-✓--✓-✓✓✓0.0510.7170.78930.5130.9270.135
w/o Plane cues✓✓-✓✓-✓✓✓0.0500.7280.86230.5830.9270.134
w/o All customized ADC✓✓✓---✓✓✓0.0520.7110.83930.4990.9260.135
w/o Line ADC✓✓✓-✓✓✓✓✓0.0510.7200.84830.5950.9260.135
w/o SfM ADC✓✓✓✓-✓✓✓✓0.0400.7710.87530.5730.9270.133
w/o Plane ADC✓✓✓✓✓-✓✓✓0.0460.7410.87130.3900.9250.137
w/o \(L_d\)✓✓✓✓✓✓-✓✓0.0440.7620.86430.5530.9260.134
w/o \(L_n\)✓✓✓✓✓✓✓-✓0.0510.7250.84830.5100.9260.137
w/o \(L_{nc}\)✓✓✓✓✓✓✓✓-0.0440.7440.80830.5050.9250.136
Full model✓✓✓✓✓✓✓✓✓0.0400.7760.87730.4530.9280.131
+ +Table 3. Ablation study results. Geometric cues significantly improve reconstruction quality. We report mesh and novel view synthesis metrics on the Scannet++ dataset. Adding the SfM cues means using the filtered SfM points instead, and using $L_{n}$ implies incorporating both $L_{n}$ and $L_{s}$ . + +mal Consistency, on all of the Replica, Scannet++, and Deep Blending datasets as shown in Tab. 2 and Fig. 5. This consistent excellence highlights its effectiveness in accurately reconstructing complex indoor environments. Since the Deep Blending dataset does not include directly provided mesh ground truth (GT), we do not perform quantitative metrics calculations. In contrast to other methods, which often result in uneven surfaces in weakly textured regions and overly smooth surfaces along edges, IndoorGS achieves a more accurate and detailed reconstruction. + +# 5.2. Rendering Comparison + +To evaluate the rendering quality of our method, we compared the test-view rendering results of our approach with those from vanilla 3DGS, GOF, RaDeGS, and PGSR on Replica, ScanNet++, and Deep Blending datasets. As shown in Tab. 1, Our method outperforms all of the above Gaussian-based reconstruction methods in terms of novel view synthesis, achieving an average PSNR that is 0.35 dB higher than the second-best method across the three datasets. Notably, following vanilla 3DGS, our test set is obtained by sequential sampling from the original dataset. The test views are positioned close to the training views in indoor scenes. In this case, despite the model overfitting the training views and the geometry inaccurately estimated, the PSNR of the rendered images from the test views can remain relatively high, which undermines the effectiveness of the test set in evaluating the true capability of the model, as it fails to challenge the model with sufficiently diverse perspectives. We provide a detailed experimental analysis of this issue in the appendix. Additionally, our method takes only 10 minutes longer than PGSR despite including the 3D line map construction and planar cues extraction process. + +# 5.3. Ablation Studies + +We conducted an ablation study on the ScanNet++ dataset to test the impact of each component we introduced. + +Specifically, we analyzed the effects of geometric cues, geometric-cue-guided ADC strategies, depth regularization, normal regularization, and normal consistency. The results are presented in Tab. 3. These results show that our full model performs best across all three mesh metrics and the final two NVS metrics (SSIM and LPIPS). Notably, as shown in the first three rows of the table, the step-by-step addition of geometric cues and the corresponding ADC strategy produces incremental performance improvement. Rows 4 to 7 further illustrate that introducing various ADC strategies after adding geometric cues results in further performance enhancement, and line-cue-guided ADC has the most significant impact among the three strategies. The final three rows show the positive effects of the three loss functions, with normal regularization providing the greatest enhancement. + +# 6. Conclusion + +In this paper, we propose IndoorGS, an indoor scene reconstruction method based on 3DGS. Our IndoorGS utilizes geometric cues such as filtered SfM points, line segments, and weakly textured planar surfaces, commonly found in indoor environments, as input data and takes a geometric-cues-guided ADC strategy for Gaussian scene optimization. We validate our method's rendering and reconstruction quality on the Replica, ScanNet++, and Deep Blending datasets. Experimental results demonstrate that our IndoorGS outperforms existing Gaussian-based techniques in geometric reconstruction accuracy and rendering quality. + +# Acknowledgment + +Project supported by the National Natural Science Foundation of China (Grant No. 62302174). The computation is completed in the HPC Platform of Huazhong University of Science and Technology. + +# References + +[1] Jiayang Bai, Letian Huang, Jie Guo, Wen Gong, Yuanqi Li, and Yanwen Guo. 360-gs: Layout-guided panoramic gaussian splatting for indoor roaming. arXiv preprint arXiv:2402.00763, 2024. 3 +[2] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 3 +[3] Danpeng Chen, Nan Wang, Runsen Xu, Weijian Xie, Hujun Bao, and Guofeng Zhang. Rnin-vio: Robust neural inertial navigation aided visual-inertial odometry in challenging scenes. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 275-283. IEEE, 2021. +[4] Danpeng Chen, Shuai Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Hujun Bao, and Guofeng Zhang. Vip-slam: An efficient tightly-coupled rgb-d visual inertial planar slam. In 2022 International Conference on Robotics and Automation (ICRA), pages 5615-5621. IEEE, 2022. 3 +[5] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 2, 3, 6 +[6] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 5 +[7] Yutong Chen, Marko Mihajlovic, Xiyi Chen, Yiming Wang, Sergey Prokudin, and Siyu Tang. Splatformer: Point transformer for robust 3d gaussian splatting, 2024. 2 +[8] Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. Pearson correlation coefficient. Noise reduction in speech processing, pages 1-4, 2009. 6 +[9] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3 +[10] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2, 3 +[11] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 7 +[12] Manuel Hofer, Michael Maurer, and Horst Bischof. Line3d: Efficient 3d scene abstraction for the built environment. In Pattern Recognition: 37th German Conference, GCPR 2015, Aachen, Germany, October 7-10, 2015, Proceedings 37, pages 237-248. Springer, 2015. 3 +[13] Manuel Hofer, Michael Maurer, and Horst Bischof. Efficient 3d scene abstraction using line segments. Computer Vision and Image Understanding, 157:167-178, 2017. 3 + +[14] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3 +[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 2, 3, 6 +[16] Jiyeop Kim and Jongwoo Lim. Integrating meshes and 3d gaussians for indoor scene reconstruction with sam mask guidance. arXiv preprint arXiv:2407.16173, 2024. 3 +[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 5 +[18] Yanyan Li, Raza Yunus, Nikolas Brasch, Nassir Navab, and Federico Tombari. Rgb-d slam with structural regularities. In 2021 IEEE international conference on Robotics and automation (ICRA), pages 11581-11587. IEEE, 2021. 5 +[19] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering. arXiv preprint arXiv:2403.11324, 2024. 5 +[20] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 3, 4 +[21] Shuhong Liu, Heng Zhou, Liuzhuozheng Li, Yun Liu, Tianchen Deng, Yiming Zhou, and Mingrui Li. Structure gaussian slam with Manhattan world hypothesis. arXiv preprint arXiv:2405.20031, 2024. 3 +[22] G Lowe. Sift-the scale invariant feature transform. Int. J, 2 (91-110):2, 2004. 5 +[23] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3 +[24] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2 +[25] Run Luo, Zikai Song, Lintao Ma, Jinlin Wei, Wei Yang, and Min Yang. Diffusiontrack: Diffusion model for multi-object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3991-3999, 2024. 2 +[26] Pierre Moulon, Pascal Monasse, Romuald Perrot, and Renaud Marlet. Openmvg: Open multiple view geometry. In Reproducible Research in Pattern Recognition: First International Workshop, RRPR 2016, Cancún, Mexico, December 4, 2016, Revised Selected Papers 1, pages 60-74. Springer, 2017. 2 +[27] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and track + +ing. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 7 +[28] Rémi Pautrat, Juan-Ting Lin, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Sold2: Self-supervised occlusion-aware line description and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11368-11378, 2021. 4 +[29] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17327-17336, 2023. 4 +[30] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 4 +[31] Radu Bogdan Rusu and Steve Cousins. 3d is here: Point cloud library (pcl). In 2011 IEEE international conference on robotics and automation, pages 1-4. IEEE, 2011. 5 +[32] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2 +[33] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint arXiv:2312.05133, 2023. 2 +[34] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International journal of computer vision, 80:189-210, 2008. 2 +[35] Xiaowei Song, Jv Zheng, Shiran Yuan, Huan-ang Gao, Jingwei Zhao, Xiang He, Weihao Gu, and Hao Zhao. Saags: Scale-adaptive gaussian splatting for training-free antialIASing. arXiv preprint arXiv:2403.19615, 2024. 3 +[36] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 7 +[37] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2 +[38] Jingwen Wang, Tymoteusz Bleja, and Lourdes Agapito. Gosurf: Neural feature grid optimization for fast, high-fidelity rgb-d surface reconstruction. In 2022 International Conference on 3D Vision (3DV), pages 433-442. IEEE, 2022. 7 +[39] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free-view synthesis of indoor scenes. arXiv preprint arXiv:2405.17958, 2024. 3 +[40] LORENSEN WE. Marching cubes: A high resolution 3d surface construction algorithm. Computer graphics, 21(1): 7-12, 1987. 7 +[41] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. + +4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2 +[42] Haodong Xiang, Xinghui Li, Xiansong Lai, Wanting Zhang, Zhichao Liao, Kai Cheng, and Xueping Liu. Gaussianroom: Improving 3d gaussian splatting with sdf guidance and monocular cues for indoor scene reconstruction. arXiv preprint arXiv:2405.19671, 2024. 3 +[43] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 7 +[44] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9043-9053, 2023. 5 +[45] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 7 +[46] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 3 +[47] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2, 3, 6 +[48] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2, 3, 4, 6, 7 +[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 +[50] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 3 +[51] Hang Zhou, Jiale Cai, Yuteng Ye, Yonghui Feng, Chenxing Gao, Junqing Yu, Zikai Song, and Wei Yang. Video anomaly detection with motion and appearance guided patch diffusion model. arXiv preprint arXiv:2412.09026, 2024. 2 +[52] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 2 +[53] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 2 \ No newline at end of file diff --git a/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/images.zip b/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9cd20708e252f620d99a1e4a63d65fde433a9b90 --- /dev/null +++ b/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df3589e79af8d275ac8085a80cd714db06aa0e468b085133ecb0d83aecd617c0 +size 602536 diff --git a/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/layout.json b/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a1e765918758ae64de2e2716ccb293139ef5b142 --- /dev/null +++ b/2025/IndoorGS_ Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction/layout.json @@ -0,0 +1,8977 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 125, + 103, + 486, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 103, + 486, + 138 + ], + "spans": [ + { + "bbox": [ + 125, + 103, + 486, + 138 + ], + "type": "text", + "content": "IndoorGS: Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 161, + 419, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 161, + 419, + 175 + ], + "spans": [ + { + "bbox": [ + 189, + 161, + 419, + 175 + ], + "type": "text", + "content": "Cong Ruan Yuesong Wang* Tao Guan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 175, + 548, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 548, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 548, + 203 + ], + "type": "text", + "content": "School of Computer Science & Technology, Huazhong University of Science and Technology, China {rc22, yuesongwang, qd.gt}@hust.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 209, + 477, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 209, + 477, + 238 + ], + "spans": [ + { + "bbox": [ + 133, + 209, + 477, + 238 + ], + "type": "text", + "content": "Bin Zhang Shenzhen Smart City Technology Development Group Co., Ltd., China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 239, + 239, + 372, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 239, + 372, + 251 + ], + "spans": [ + { + "bbox": [ + 239, + 239, + 372, + 251 + ], + "type": "text", + "content": "zhangbin@smartcitysz.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 149, + 257, + 460, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 257, + 460, + 285 + ], + "spans": [ + { + "bbox": [ + 149, + 257, + 460, + 285 + ], + "type": "text", + "content": "Lili Ju \nDepartment of Mathematics, University of South Carolina, USA" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 265, + 288, + 342, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 288, + 342, + 298 + ], + "spans": [ + { + "bbox": [ + 265, + 288, + 342, + 298 + ], + "type": "text", + "content": "ju@math.sc.edu" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 62, + 349, + 189, + 427 + ], + "blocks": [ + { + "bbox": [ + 62, + 349, + 189, + 427 + ], + "lines": [ + { + "bbox": [ + 62, + 349, + 189, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 349, + 189, + 427 + ], + "type": "image", + "image_path": "86286184a81fa774ef4661008d3ba522052961a5bf96479b596ef63ebb4bad3b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 62, + 430, + 188, + 505 + ], + "blocks": [ + { + "bbox": [ + 62, + 430, + 188, + 505 + ], + "lines": [ + { + "bbox": [ + 62, + 430, + 188, + 505 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 188, + 505 + ], + "type": "image", + "image_path": "a383e9358108d975821e845dc09ce6b7c15b1cb5e20e79316dc07a49954a2c72.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 506, + 145, + 516 + ], + "lines": [ + { + "bbox": [ + 119, + 506, + 145, + 516 + ], + "spans": [ + { + "bbox": [ + 119, + 506, + 145, + 516 + ], + "type": "text", + "content": "GOF" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 195, + 351, + 308, + 426 + ], + "blocks": [ + { + "bbox": [ + 195, + 351, + 308, + 426 + ], + "lines": [ + { + "bbox": [ + 195, + 351, + 308, + 426 + ], + "spans": [ + { + "bbox": [ + 195, + 351, + 308, + 426 + ], + "type": "image", + "image_path": "afc5f02bb84e4364acfbc1b7e5ca1bf4c1edaa9334d1bd29d36a12ec3f8dd5de.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 195, + 430, + 308, + 506 + ], + "blocks": [ + { + "bbox": [ + 195, + 430, + 308, + 506 + ], + "lines": [ + { + "bbox": [ + 195, + 430, + 308, + 506 + ], + "spans": [ + { + "bbox": [ + 195, + 430, + 308, + 506 + ], + "type": "image", + "image_path": "f514a39c5ccc3b8f36aee53f46497a711199eb45ba54e6c3cd0aa41c287f74ec.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 506, + 273, + 517 + ], + "lines": [ + { + "bbox": [ + 229, + 506, + 273, + 517 + ], + "spans": [ + { + "bbox": [ + 229, + 506, + 273, + 517 + ], + "type": "text", + "content": "RaDeGS" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 315, + 351, + 427, + 426 + ], + "blocks": [ + { + "bbox": [ + 315, + 351, + 427, + 426 + ], + "lines": [ + { + "bbox": [ + 315, + 351, + 427, + 426 + ], + "spans": [ + { + "bbox": [ + 315, + 351, + 427, + 426 + ], + "type": "image", + "image_path": "82d714b2771d7b9c57300e2b8531079a8e47abb6743c3c278306faedc403b6a8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 315, + 430, + 427, + 506 + ], + "blocks": [ + { + "bbox": [ + 315, + 430, + 427, + 506 + ], + "lines": [ + { + "bbox": [ + 315, + 430, + 427, + 506 + ], + "spans": [ + { + "bbox": [ + 315, + 430, + 427, + 506 + ], + "type": "image", + "image_path": "6dd97cc05486b5981e971fff87df30dee8bf49442b5deb965f7db0ca011eb623.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 506, + 388, + 517 + ], + "lines": [ + { + "bbox": [ + 355, + 506, + 388, + 517 + ], + "spans": [ + { + "bbox": [ + 355, + 506, + 388, + 517 + ], + "type": "text", + "content": "PGSR" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 524, + 555, + 554 + ], + "lines": [ + { + "bbox": [ + 55, + 524, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 555, + 554 + ], + "type": "text", + "content": "Figure 1. Our IndoorGS achieves state-of-the-art (SOTA) performance in indoor scene reconstruction. Compared to other methods that produce poor geometric results, our approach generates well-structural geometries while preserving all details. We can achieve superior rendering quality thanks to our precise geometry estimation." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 433, + 351, + 547, + 426 + ], + "blocks": [ + { + "bbox": [ + 433, + 351, + 547, + 426 + ], + "lines": [ + { + "bbox": [ + 433, + 351, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 433, + 351, + 547, + 426 + ], + "type": "image", + "image_path": "9c81823ce147160df633083428e050a13e4b4d44d151a356c65c6aa5e24376c7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 433, + 430, + 547, + 506 + ], + "blocks": [ + { + "bbox": [ + 433, + 430, + 547, + 506 + ], + "lines": [ + { + "bbox": [ + 433, + 430, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 433, + 430, + 547, + 506 + ], + "type": "image", + "image_path": "f8e8c1c2b0debe1cd53f932432771077d304be39ce79c5f28ac3a4d7e1fcbd9c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 506, + 504, + 516 + ], + "lines": [ + { + "bbox": [ + 477, + 506, + 504, + 516 + ], + "spans": [ + { + "bbox": [ + 477, + 506, + 504, + 516 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 151, + 569, + 200, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 569, + 200, + 582 + ], + "spans": [ + { + "bbox": [ + 151, + 569, + 200, + 582 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 54, + 595, + 295, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 595, + 295, + 692 + ], + "spans": [ + { + "bbox": [ + 54, + 595, + 295, + 692 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) has shown impressive performance in scene reconstruction, offering high rendering quality and rapid rendering speed with short training time. However, it often yields unsatisfactory results when applied to indoor scenes due to its poor ability to learn geometries without enough textural information. In this paper, we propose a new 3DGS-based method \"IndoorGS\", that leverages the commonly found yet important geometric" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 571, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 556, + 715 + ], + "type": "text", + "content": "cues in indoor scenes to improve the reconstruction quality. Specifically, we first extract 2D lines from input images and fuse them into 3D line cues via feature-based matching, which can provide a structural understanding of the target scene. We then apply the statistical outlier removal to refine Structure-from-Motion (SfM) points, ensuring robust cues in texture-rich areas. Based on these two types of cues, we further extract reliable 3D plane-like cues for textureless regions. Such geometric information will be utilized not only for initialization but also in the realization of a geometric-cue-guided adaptive density control (ADC) strategy. The proposed ADC approach is grounded in the principle of" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 704, + 285, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 704, + 285, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 704, + 285, + 714 + ], + "type": "text", + "content": "*Corresponding author. Contact him at yuesongwang@hust.edu.cn." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "844" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 298, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 298, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 298, + 145 + ], + "type": "text", + "content": "divide-and-conquer and optimizes the use of each type of geometric cues to enhance overall reconstruction performance. Extensive experiments on multiple indoor datasets show that our method can deliver much more accurate geometry and higher rendering quality for indoor scenes than existing 3DGS approaches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 136, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 136, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 136, + 179 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 187, + 297, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 297, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 297, + 390 + ], + "type": "text", + "content": "Indoor scene reconstruction is one of common tasks in computer vision, with broad applications in virtual reality, gaming, autonomous driving, robotics, and so on [25, 51]. Recently, 3D Gaussian Splitting (3DGS) [15] has emerged as a promising scene reconstruction method, which introduces a hybrid optimization method with Gaussian primitives similar to point clouds and combines differentiable backpropagation with adaptive density control. This method has demonstrated superior efficacy, exhibiting expedited training process, accelerated rendering speed, and high-quality view synthesis, particularly in highly textured outdoor environments. However, it often performs poorly when applied to indoor environments, which mainly attributes to the fact that 3DGS is not capable of reconstructing correct geometry when facing textureless regions, as there are too many potential combinations of geometry and appearance that can satisfy the training images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 391, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 605 + ], + "type": "text", + "content": "Although there has been a large amount of research works [5, 9, 10, 14, 47, 48] devoted to improving the geometry inference ability of 3DGS-based methods, their reconstruction quality is still unsatisfactory when dealing with indoor scenes, resulting in holes or floaters in the rendering images, as shown in Fig. 1. There are many local optima during the training process, and without strong constraints and correct guidance, 3DGS could be easily trapped by these local optima. What's even worse is that the initialization of 3DGS relies entirely on SfM points, which are inherently sparse and noisy in regions with weak textures, and the lack of reliable seeds can lead to blind exploration of 3DGS in these regions, further exacerbating the likelihood of being trapped by the local optima. To improve the performance of 3DGS for indoor scene reconstruction, we need to, on the one hand, provide better reliable seeds and, on the other hand, add strong geometric constraints to help the optimization move towards the global optima." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 606, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 297, + 715 + ], + "type": "text", + "content": "To obtain better initial seeds, we propose to extract more reliable geometric cues commonly found in the indoor scene and use them to help initialize the 3DGS. Specifically, we first extract 2D line segments and perform feature-based matching on these 2D lines to obtain 3D line cues. Matching only line segments can drastically reduce the computational complexity and the matching ambiguity, allowing us to get accurate and noise-free 3D line cues to help 3DGS understand the scene's construction. While SfM [26, 32, 34]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "points can provide good geometric priors in texture-rich regions, they usually behave as outliers in textureless or repeating-textured areas and can also trouble the reconstruction. Therefore, we filter the SfM points via Statistical Outlier Removal (SOR) and use the remaining points as reliable geometric cues in texture-rich regions. We further extend these cues to the weakly textured planar-like regions commonly existing in indoor scenes, such as walls, floors, ceilings, and furniture surfaces. We first extract the weakly textured regions from the 2D images and back-project them into 3D space using aligned monocular depths. We then analyze the spatial relationship between the extracted 3D lines and these back-projected planes, allowing us to extract reliable 3D plane-like cues. The above three types of geometric cues are fed to 3DGS for initialization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 251, + 553, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 453 + ], + "type": "text", + "content": "Based on the distinctive characteristics of these geometric cues, we further divide the Gaussians into three types, namely line Gaussians, plane-like Gaussians, and SfM Gaussians, and design a geometric-cue-guided adaptive density control strategy to enhance the optimization of 3DGS marching towards global optima. As 3D lines are typically located at object edges with high spatial accuracy, we apply a tangent space densification method to these line Gaussians, encouraging them to expand to complete the object. For plane-like Gaussians, since the initial seeds from 3D plane-like region cues may introduce spatial offsets and redundancies, we disable the splitting and cloning operations for these Gaussians, focusing solely on optimizing Gaussian parameters and pruning processes. For SfM geometric cues, because SfM points are often noisier than 3D lines, we implement a densification strategy that prioritizes line Gaussian over SfM Gaussian." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 455, + 533, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 455, + 533, + 466 + ], + "spans": [ + { + "bbox": [ + 324, + 455, + 533, + 466 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 468, + 553, + 586 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 468, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 553, + 502 + ], + "type": "text", + "content": "- We propose a geometric cue extraction method for obtaining different types of reliable geometric priors from indoor scenes to guide the reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 503, + 553, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 553, + 550 + ], + "type": "text", + "content": "- We classify Gaussian primitives according to their geometric types and design a geometric-cue-guided adaptive density control strategy to better utilize the unique properties of each type of geometric cue." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 551, + 553, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 553, + 586 + ], + "type": "text", + "content": "- Our proposed IndoorGS demonstrates superior performance in both the rendering quality and the accuracy of extracted geometry across various indoor datasets." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 597, + 400, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 400, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 400, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "3DGS has been as a prominent 3D reconstruction technique in the past couple of years due to its training efficiency, rendering performance, and impressive results across various 3D tasks [24, 33, 37, 41]. The core of 3DGS lies in a carefully designed rasterization [53] method that efficiently leverages GPU computation by representing scenes as collections of Gaussian primitives. Despite the remarkable success of 3DGS and its subsequent modifica" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "845" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 553, + 252 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 553, + 252 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 553, + 252 + ], + "type": "image", + "image_path": "4eb2bc40b68ad2f362d4009f638cfb33739f4dd9bd934f9f926831cda4a6b27c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 259, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 259, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 259, + 555, + 293 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed IndoorGS. (a) The left part illustrates the extraction process of geometric cues, i.e., line cues, plane-like cues, and SfM cues, which are then used to initialize 3DGS. (b) The right part depicts the optimization process, where the 3DGS is further optimized using geometric-cue-guided adaptive density control and supervised by a series of loss functions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 314, + 296, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 296, + 709 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 296, + 709 + ], + "type": "text", + "content": "tions [23, 35, 46, 50] in novel view synthesis, it faces significant challenges in surface reconstruction, mainly due to the discrete and unstructured nature of its Gaussian primitives. Recent research has sought to enhance 3DGS-based surface reconstruction [5, 9, 10, 14, 47, 48], yielding improvements in geometrically rich environments with strong textures. However, these methods still struggle in weakly textured settings, such as indoor scenes, which has driven growing interest in 3DGS reconstruction for indoor environments. GaussianRoom [42] integrates neural signed distance functions (SDF) with 3DGS, incorporating monocular normal and edge priors to enhance indoor scene details. Kim et al.[16] proposed an approach that employs 3D Gaussians for object reconstruction while using meshes to capture the room layout. 360-GS [1] leverages four panoramic images to initialize 3D Gaussians based on room layout and depth maps derived from panoramas for comprehensive indoor reconstruction. FreeSplat [39] applies 2D backbones and a cost volume for multi-view aggregation, while MGSLAM [21] utilizes the Manhattan World hypothesis to improve geometric accuracy and completeness in real-time SLAM [2-4] reconstructions. Despite the contributions, these methods usually require specialized inputs or are inefficient, exhibiting limitations in texture-poor environments. Indoor scenes are rich in geometric cues like lines and planes, which are useful for 3DGS-based reconstructions. However, SfM often encounters issues in these environments due to large textureless surfaces and repetitive elements. This is where 3D line map reconstruction approaches [12, 13, 20] show natural advantages in overcoming weak and repetitive textures common in indoor settings. By fully leveraging these geometric cues, 3DGS is expected to better handle the unique challenges of indoor" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 316, + 402, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 316, + 402, + 326 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 402, + 326 + ], + "type": "text", + "content": "scene reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 335, + 400, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 335, + 400, + 348 + ], + "spans": [ + { + "bbox": [ + 314, + 335, + 400, + 348 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 355, + 443, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 443, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 443, + 369 + ], + "type": "text", + "content": "3.1. 3D Gaussian Splatting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "type": "text", + "content": "3DGS [15] explicitly represents a 3D scene with a set of Gaussian primitives. Each of them is parameterized via a 3D covariance matrix " + }, + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "type": "text", + "content": " and mean " + }, + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 347, + 413, + 555, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 413, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 347, + 413, + 555, + 437 + ], + "type": "interline_equation", + "content": "G (\\boldsymbol {p}) = \\exp \\left(- \\frac {1}{2} \\left(\\boldsymbol {p} - \\boldsymbol {\\mu}\\right) ^ {\\intercal} \\boldsymbol {\\Sigma} ^ {- 1} \\left(\\boldsymbol {p} - \\boldsymbol {\\mu}\\right)\\right), \\tag {1}", + "image_path": "c2d84a24328647ecc99d389ea41efd2f97237b12c84b0bfd0730c19c514e6f62.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "text", + "content": " can be further factorized into " + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{RSS}^{\\intercal}\\mathbf{R}^{\\intercal}" + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "text", + "content": " with a scaling matrix " + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "text", + "content": " and a rotation matrix " + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "text", + "content": " is the blending coefficient for a Gaussian with center " + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{i}" + }, + { + "bbox": [ + 313, + 441, + 555, + 478 + ], + "type": "text", + "content": " in screen space:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 483, + 555, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 483, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 328, + 483, + 555, + 510 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = o _ {i} \\cdot \\exp \\left(- \\frac {1}{2} (\\pmb {p} - \\pmb {\\mu} _ {i}) ^ {\\intercal} (\\pmb {\\Sigma} _ {i} ^ {\\prime}) ^ {- 1} (\\pmb {p} - \\pmb {\\mu} _ {i})\\right). (2)", + "image_path": "f61655177cf0c97e37767225fa950fd5cb36181b6a1592e7310baa0aee788f1a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "text", + "content": " is the Gaussian opacity, the 3D Gaussian is transformed into the image coordinates with world-to-camera transform matrix " + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "text", + "content": " and projected onto pixel space with the Jacobian of the affine approximation for projection matrix " + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "text", + "content": ", then the 2D covariance matrix " + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}^{\\prime}" + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}^{\\prime} = \\mathbf{J}\\mathbf{W}\\boldsymbol{\\Sigma}\\mathbf{W}^{\\mathrm{T}}\\mathbf{J}^{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "text", + "content": ". The rendering result is computed using point-based alpha-blending on each ray:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 341, + 603, + 555, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 603, + 555, + 637 + ], + "spans": [ + { + "bbox": [ + 341, + 603, + 555, + 637 + ], + "type": "interline_equation", + "content": "\\hat {C} = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} T _ {i}, \\quad \\text {w h e r e} T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {3}", + "image_path": "1ab174ba666c01e3ab7c5352d59967f1f0e181903067ce9b6718e6a95b6e9bd1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " is the accumulated transmittance at pixel location " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{c}_i" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " is the view-dependent color defined by spherical harmonics (SH) coefficients, and " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " is the number of Gaussian primitives. A photometric loss function optimizes the Gaussian parameters to reduce the discrepancies between the rendered images and the actual observational data." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "846" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 222, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 222, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 222, + 85 + ], + "type": "text", + "content": "3.2. Rasterizing Depth and Normal" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 126 + ], + "type": "text", + "content": "Following RaDeGS [48], the center " + }, + { + "bbox": [ + 55, + 89, + 296, + 126 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 55, + 89, + 296, + 126 + ], + "type": "text", + "content": " of a Gaussian " + }, + { + "bbox": [ + 55, + 89, + 296, + 126 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 55, + 89, + 296, + 126 + ], + "type": "text", + "content": " is initially projected into the camera space as " + }, + { + "bbox": [ + 55, + 89, + 296, + 126 + ], + "type": "inline_equation", + "content": "\\mu_{i}^{\\prime}" + }, + { + "bbox": [ + 55, + 89, + 296, + 126 + ], + "type": "text", + "content": ". The depth of each pixel correspondent to the Gaussian is computed as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 133, + 295, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 133, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 66, + 133, + 295, + 174 + ], + "type": "interline_equation", + "content": "d = z _ {i} ^ {\\prime} + \\boldsymbol {p} \\left( \\begin{array}{c} \\Delta x \\\\ \\Delta y \\end{array} \\right), \\quad \\boldsymbol {\\mu} _ {i} ^ {\\prime} = \\left[ \\begin{array}{c} x _ {i} ^ {\\prime} \\\\ y _ {i} ^ {\\prime} \\\\ z _ {i} ^ {\\prime} \\end{array} \\right] = \\boldsymbol {W} \\boldsymbol {\\mu} _ {i} + t, (4)", + "image_path": "fe626f3db5e26080072fd42106e19dc22c3a70120265fc4c8a78f217e30aeaa3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "inline_equation", + "content": "z_{i}^{\\prime}" + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": " represents the depth of the Gaussian center, " + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\Delta x = x_{i}^{\\prime} - x" + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\Delta y = y_{i}^{\\prime} - y" + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": " denote the relative pixel positions. The vector " + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": " is determined by the Gaussian parameters " + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "inline_equation", + "content": "[\\pmb{W},\\pmb{t}] \\in \\mathbb{R}^{3\\times 4}" + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": ". Per-pixel z-depth estimates " + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": " are rendered using the discrete volume rendering approximation similar to color values:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 261, + 295, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 261, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 118, + 261, + 295, + 294 + ], + "type": "interline_equation", + "content": "\\hat {D} = \\sum_ {i \\in N} d _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {5}", + "image_path": "218845cb9b467376c8f9001d46703e987138e19299aa506a41cd5ab34fda0b23.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 304, + 296, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 304, + 296, + 340 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 55, + 304, + 296, + 340 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 304, + 296, + 340 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 304, + 296, + 340 + ], + "type": "text", + "content": " Gaussian's depth according to the Eq. (4). The normal direction of each projected Gaussian is computed as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 128, + 349, + 295, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 349, + 295, + 377 + ], + "spans": [ + { + "bbox": [ + 128, + 349, + 295, + 377 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {n}} = - \\boldsymbol {J} ^ {\\intercal} \\left(\\frac {z _ {i} ^ {\\prime}}{z _ {i}} \\boldsymbol {p}, 1\\right) ^ {\\intercal} \\tag {6}", + "image_path": "cfd7f4257e33c0d10a96658d6a2de1f89f56d5e36e33b8e5d075037d69dd6485.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "text", + "content": "Normals are then alpha-blended, resulting in a single, unified per-pixel normal estimate:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 417, + 295, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 417, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 137, + 417, + 295, + 443 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {N}} = \\sum_ {i \\in N} \\hat {\\boldsymbol {n}} _ {i} \\alpha_ {i} T _ {i}. \\tag {7}", + "image_path": "f3b611f16506f93314b2f4ba58f26db786c0e11c0160ac27d282a92c32b9fb08.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 452, + 162, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 162, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 162, + 466 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 472, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 556 + ], + "type": "text", + "content": "Fig. 2 gives an overview of our methods. We first extract geometric cues, i.e., line, plane, and SfM point cues, from the input images (Section 4.1). Then, we divide Gaussians into three types according to the geometric cues and treat them differently in our geometric-cue-guided adaptive density control (Section 4.2). The training process is supervised by color and geometric loss (Section 4.3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 564, + 206, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 206, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 206, + 576 + ], + "type": "text", + "content": "4.1. Geometric Cues Extraction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 582, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 665 + ], + "type": "text", + "content": "SfM points are often too sparse to guide the reconstruction of indoor scenes and sometimes contain geometric errors, which can cause lots of floaters in the subsequent 3DGS reconstruction process. Therefore, we need to provide the 3DGS with sufficient and accurate geometric cues. As depicted in Fig. 2, there are three types of geometric cues that we use to guide the reconstruction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "type": "text", + "content": "Geometric Cues from 3D lines. In contrast to sparse points from SfM, which is structureless, 3D line segments can concisely encode the high-level layout, as they often delineate the main structural elements of the scene. We follow" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": "the idea of LIMAP [20] to extract 3D lines, mainly consisting of 2D line extraction and line matching. Although rich 2D line segment features can be extracted from images using learning-based methods such as DeepLSD [29] or Sold2 [28], when encountering large continuous weakly textured or untextured areas, such as wall corner joint between white walls, they still have difficulties in obtaining distinctive 2D line segment features using only color information. However, we can easily extract distinctive line segment features from textureless regions on their normal maps, as shown in Fig. 3." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 321, + 213, + 434, + 293 + ], + "blocks": [ + { + "bbox": [ + 321, + 213, + 434, + 293 + ], + "lines": [ + { + "bbox": [ + 321, + 213, + 434, + 293 + ], + "spans": [ + { + "bbox": [ + 321, + 213, + 434, + 293 + ], + "type": "image", + "image_path": "b2d5a0b6a063fea8388cbb78d2651ef015aab20ee61a7efc3e522f286f6fb59d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 435, + 213, + 548, + 293 + ], + "blocks": [ + { + "bbox": [ + 435, + 213, + 548, + 293 + ], + "lines": [ + { + "bbox": [ + 435, + 213, + 548, + 293 + ], + "spans": [ + { + "bbox": [ + 435, + 213, + 548, + 293 + ], + "type": "image", + "image_path": "f227e2694a21010988a00558ba27d9ee217a96b0b7e73d06c4ea8a26fce85f06.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 321, + 297, + 434, + 375 + ], + "blocks": [ + { + "bbox": [ + 321, + 297, + 434, + 375 + ], + "lines": [ + { + "bbox": [ + 321, + 297, + 434, + 375 + ], + "spans": [ + { + "bbox": [ + 321, + 297, + 434, + 375 + ], + "type": "image", + "image_path": "21a41c1c872ff3abc3f23c0b2c51b29b9061e7b89a35800ea4f2906dab2be1bd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 378, + 387, + 387 + ], + "lines": [ + { + "bbox": [ + 368, + 378, + 387, + 387 + ], + "spans": [ + { + "bbox": [ + 368, + 378, + 387, + 387 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 435, + 298, + 547, + 376 + ], + "blocks": [ + { + "bbox": [ + 435, + 298, + 547, + 376 + ], + "lines": [ + { + "bbox": [ + 435, + 298, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 435, + 298, + 547, + 376 + ], + "type": "image", + "image_path": "1c19b040043e13a978e48120f3ca13047609a5966a376950486752d425422482.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 377, + 504, + 386 + ], + "lines": [ + { + "bbox": [ + 478, + 377, + 504, + 386 + ], + "spans": [ + { + "bbox": [ + 478, + 377, + 504, + 386 + ], + "type": "text", + "content": "Normal" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "type": "text", + "content": "Figure 3. The complementary effects of extracting structured lines from RGB images and normal maps. Top row: The line segments can be easily seen from the normal map, but extracting the corresponding line segments from the RGB image is difficult. Bottom row: The opposite case." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 463, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 557 + ], + "type": "text", + "content": "However, at the beginning of the reconstruction, we do not know the scene's geometry and can not render the normal maps. Fortunately, nowadays, foundation models for monocular normal estimation can give us a good prior. Thus, we extracted 2D line segments from RGB images and monocular normal maps using the DeepLSD [29] method. Subsequently, the GlueStick [30] method was used to extract the line segments and match them." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 558, + 554, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 630 + ], + "type": "text", + "content": "Finally, we obtain the 3D line segments for the scene. To capture color information along these segments, we extract the corresponding 2D line colors from all visible views of each 3D line, then perform multiview color averaging to achieve a consistent color representation, and sample 3D points on 3D line segments in fixed steps." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Geometric Cues from SfM. The sparse points of SfM in indoor scenes often contain noise or outliers to repetitive structures, reflective surfaces, and limited textures. Although sparse 3D feature points are generally accurate in strong-texture regions, they are often misaligned in weak-texture regions, as 2D points are more challenging to match correctly across multiple views. Therefore, filtering SfM" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "847" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "points is crucial, as erroneous inputs can be amplified during adaptive density control (ADC) in 3DGS, causing floater artifacts. To address this, we apply a statistical outlier removal filter (SOR) [31], which calculates the average distance between each point and its K nearest neighbors, removing points with abnormally high average distances. The remaining SfM points serve as reliable geometric cues in texture-rich regions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "text", + "content": "Geometric Cues from Plane-like Regions. The plane-like regions in indoor scenes are usually weakly textured walls, ceilings, etc. For detecting these weakly textured regions in images, we use the Segment Anything Model (SAM) [17] for image segmentation, then determine texture strength by counting the density of SIFT [22] points in each segmented region." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 253, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 515 + ], + "type": "text", + "content": "It is difficult to calculate the depth of these regions due to the lack of texture information. However, the positions of the structural line clues on their edges are accurate enough. Thus, we can locate these plane-like cues using the 3D line cues with the help of the monocular depth estimation model. Specifically, we first obtain the monocular depth estimation from Metric3D [44]. While the monocular depth estimation can reflect the relative positions of lines and planes, it remains inconsistent with absolute depths. Thus, we estimate a relative per-image depth scale by aligning monocular depth with line cues and SfM cues using linear regression, which is then used to adjust the monocular depths. Based on the adjusted monocular depths, the weakly textured region is back-projected into 3D space to form a point cloud. To ensure the reliability of the back-projected point cloud, we then calculate the distances between sampling points from 3D line cues and the back-projected point cloud, based on which we filter out points exceeding a preset threshold. The retained points form the plane-like cues, giving reliable priors for weakly textured regions in 3D space. The whole process is described as Algorithm 1 in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 525, + 258, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 258, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 258, + 538 + ], + "type": "text", + "content": "4.2. Geometric-cue-guided ADC Strategies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 543, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 639 + ], + "type": "text", + "content": "Besides extracting points from geometric cues for initialization, we further let these cues guide the reconstruction process. To achieve this, we design Geometric-cue-guided Adaptive Densification Control (ADC) strategies for the three Gaussian types based on the positional accuracy and significance of their respective 3D spatial locations. In the following, we denote the Gaussians from SfM cues, line cues, and plane-like cues as " + }, + { + "bbox": [ + 55, + 543, + 295, + 639 + ], + "type": "inline_equation", + "content": "G_{M}" + }, + { + "bbox": [ + 55, + 543, + 295, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 543, + 295, + 639 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 55, + 543, + 295, + 639 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 543, + 295, + 639 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 55, + 543, + 295, + 639 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 642, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 688 + ], + "type": "text", + "content": "- Line Gaussian " + }, + { + "bbox": [ + 55, + 642, + 295, + 688 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 55, + 642, + 295, + 688 + ], + "type": "text", + "content": ": They typically locate at object edges with high spatial accuracy. Thus, as discussed in detail below, we specifically design a tangent space densification method for them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- SfM Gaussian " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "G_{M}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": ": The ADC process for " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " takes precedence over " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "G_{M}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " since " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "G_{M}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " are often noisier than " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 553, + 144 + ], + "type": "text", + "content": "Specifically, we allow " + }, + { + "bbox": [ + 321, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "G_{M}" + }, + { + "bbox": [ + 321, + 72, + 553, + 144 + ], + "type": "text", + "content": " to split and clone freely between 500-700 iterations to ensure it occupies strong texture regions correctly. Between iterations 700 and 3500, no ADC process is applied to " + }, + { + "bbox": [ + 321, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "G_{M}" + }, + { + "bbox": [ + 321, + 72, + 553, + 144 + ], + "type": "text", + "content": "; only the tangent space densification of " + }, + { + "bbox": [ + 321, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 321, + 72, + 553, + 144 + ], + "type": "text", + "content": " occurs during this period. After 3500 iterations, the ADC process for " + }, + { + "bbox": [ + 321, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "G_{M}" + }, + { + "bbox": [ + 321, + 72, + 553, + 144 + ], + "type": "text", + "content": " resumes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "text", + "content": "- Plane-like Gaussian " + }, + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "text", + "content": ": Due to potential spatial offsets and redundancy of " + }, + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "text", + "content": ", we disable splitting and cloning operations for " + }, + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "text", + "content": ". Instead, only the parameters optimization and pruning processes are conducted for them." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 203, + 434, + 277 + ], + "blocks": [ + { + "bbox": [ + 318, + 203, + 434, + 277 + ], + "lines": [ + { + "bbox": [ + 318, + 203, + 434, + 277 + ], + "spans": [ + { + "bbox": [ + 318, + 203, + 434, + 277 + ], + "type": "image", + "image_path": "8600296d216d3b058c1cfc1f4f5c9aa590730db32304e9983ec2bed6169184b0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "text", + "content": "Figure 4. Densification in the tangential space. Left: Cloning. When the Gaussian " + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "text", + "content": " is cloned, an average gradient is added to its spatial location " + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "text", + "content": ". The projection of this gradient in the normal direction (the correction factor) is then subtracted, resulting in the new cloned green Gaussian " + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "inline_equation", + "content": "G_{i+1}" + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "text", + "content": ". Right: Splitting. When the Gaussian " + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "text", + "content": " is split, a small random vector " + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "text", + "content": " is generated and added to its position. The component of " + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "text", + "content": " along the normal is set to 0, yielding the two newly split Gaussians " + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "inline_equation", + "content": "G_{i+1}" + }, + { + "bbox": [ + 313, + 288, + 554, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 436, + 203, + 552, + 277 + ], + "blocks": [ + { + "bbox": [ + 436, + 203, + 552, + 277 + ], + "lines": [ + { + "bbox": [ + 436, + 203, + 552, + 277 + ], + "spans": [ + { + "bbox": [ + 436, + 203, + 552, + 277 + ], + "type": "image", + "image_path": "d617d2d7634461cd87bc46f3de155e2962cf08eb6c7fe8a72453f7d3d0c00111.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 387, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 602 + ], + "type": "text", + "content": "Densification in the tangential space. The vanilla 3D Gaussian Splatting (3DGS) is traditionally densified by randomly splitting or cloning Gaussian primitives within their Gaussian fields. While this method is flexible and easy to implement, it encounters difficulties in indoor environments as this unconstrained densification strategy can result in artifacts like floaters in weakly textured regions. Inspired by GeoGS [19], we propose a similar tangent space densification scheme. However, unlike GeoGS, whose initial point cloud is derived from PlanarSLAM [18] and whose normal vectors can be directly obtained from a spatial fit, we do not simply reduce the third scale of the Gaussian primitive to a very small value as we do not know the initial normal vectors. Instead, we introduce a soft constraint that flattens the Gaussian ellipsoid, akin to the approach used in Neusg [6], we directly minimize the minimum scale factor " + }, + { + "bbox": [ + 313, + 387, + 555, + 602 + ], + "type": "inline_equation", + "content": "S_{i} = \\text{diag}(s_{1}, s_{2}, s_{3})" + }, + { + "bbox": [ + 313, + 387, + 555, + 602 + ], + "type": "text", + "content": " for each Gaussian that originates from 3D line points:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 378, + 608, + 553, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 608, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 378, + 608, + 553, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\left\\| \\min \\left(s _ {1}, s _ {2}, s _ {3}\\right) \\right\\| _ {1}. \\tag {8}", + "image_path": "827fb92d687507ae7a4dd270f93f734e6589002ffccd3b21578be2f764e2d767.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 625, + 553, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 553, + 685 + ], + "type": "text", + "content": "In this case, Gaussian primitive becomes flat, forming dislike surfel, and the smallest scaling axis approximates the normal direction. Simultaneously, we align the Gaussians' normal with the monocular surface normal estimated by Metrci3D using the following normal constraints:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 359, + 689, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 689, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 359, + 689, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n} = \\frac {1}{| \\hat {N} |} \\sum \\left\\| \\hat {N} _ {i j} - N _ {\\text {m o n o}, i j} \\right\\| _ {1}, \\tag {9}", + "image_path": "6ff2fa2181ea9ce12d51d6ab76be3a3a5fe64611b48eb3eeddb97d2387a34677.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "848" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\hat{N}_{ij}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " represents estimated normal values at pixel position " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{mono},ij}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " represents monocular normal values at pixel position " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ". We then define the geometric normal for a flat Gaussian primitive as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 132, + 295, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 132, + 295, + 146 + ], + "spans": [ + { + "bbox": [ + 89, + 132, + 295, + 146 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {n}} _ {f l a t} = \\boldsymbol {R} \\cdot \\operatorname {O n e H o t} (\\arg \\min (s _ {1}, s _ {2}, s _ {3})), \\tag {10}", + "image_path": "4824189f761149074aba4cfe57b5d2ee7653fa011c7e45ec9a54dc60fe1c9568.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 296, + 323 + ], + "type": "text", + "content": "where rotation matrix " + }, + { + "bbox": [ + 55, + 156, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in SO(3)" + }, + { + "bbox": [ + 55, + 156, + 296, + 323 + ], + "type": "text", + "content": " which is obtained from its quaternion " + }, + { + "bbox": [ + 55, + 156, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 55, + 156, + 296, + 323 + ], + "type": "text", + "content": ", OneHot(.) ∈ " + }, + { + "bbox": [ + 55, + 156, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 156, + 296, + 323 + ], + "type": "text", + "content": " returns a unit vector with zeros everywhere except at the position where the scaling " + }, + { + "bbox": [ + 55, + 156, + 296, + 323 + ], + "type": "inline_equation", + "content": "s_i = (s_1, s_2, s_3)" + }, + { + "bbox": [ + 55, + 156, + 296, + 323 + ], + "type": "text", + "content": " is minimum. However, the orientation of the shortest axis can be ambiguous, as it may point either outward or inward relative to the surface. We leverage the relationship between the observation direction and the Gaussian normal to resolve this ambiguity. Specifically, the angle between the observation direction and the Gaussian normal should always be greater than 90 degrees. If this condition is not met, we simply invert the Gaussian normal vector by multiplying it by -1, ensuring the correct orientation. Then, we implement a tangent space densification strategy for these flat Gaussians, as shown in Fig. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 324, + 296, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 382 + ], + "type": "text", + "content": "When a cloning operation is performed, the position of the newly cloned Gaussian is determined by adding the average gradient to the position of the original Gaussian, along with a correction factor. The correction is calculated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 385, + 295, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 385, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 295, + 400 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} ^ {i + 1} = \\boldsymbol {\\mu} ^ {i} + \\nabla \\boldsymbol {\\mu} ^ {i} - \\hat {\\mathbf {n}} _ {f l a t} \\cdot \\nabla \\boldsymbol {\\mu} ^ {i}, \\tag {11}", + "image_path": "5cf15047b70d8dc5411c47b6ad1cd56b4a3141130c9cac51b982131ebf83d392.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^i" + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": " is the position of the original Gaussian " + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "inline_equation", + "content": "G^i" + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\nabla \\pmb{\\mu}^i" + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": " is the accumulated average gradient of the original Gaussian " + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "inline_equation", + "content": "G_i" + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}_{flat}" + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": " is the normal of the original Gaussian " + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "inline_equation", + "content": "G_i" + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 445, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 539 + ], + "type": "text", + "content": "When a splitting operation is performed, the position of the newly split Gaussian is determined by adding an offset to the original Gaussian's position. This offset is drawn from a normal distribution with a mean of 0 and the three scales of the Gaussian as the standard deviation. We set the offset along the axis corresponding to the most miniature scale to zero, ensuring no movement occurs along that direction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 541, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 296, + 601 + ], + "type": "text", + "content": "The advantage of this approach is that the tangent space densification of " + }, + { + "bbox": [ + 55, + 541, + 296, + 601 + ], + "type": "inline_equation", + "content": "G_{L}" + }, + { + "bbox": [ + 55, + 541, + 296, + 601 + ], + "type": "text", + "content": " enables better splitting and cloning along surfaces, effectively suppressing the issue of \"floaters\" caused by excessive splitting and cloning of erroneous seeds in " + }, + { + "bbox": [ + 55, + 541, + 296, + 601 + ], + "type": "inline_equation", + "content": "G_{S}" + }, + { + "bbox": [ + 55, + 541, + 296, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 611, + 198, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 198, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 198, + 623 + ], + "type": "text", + "content": "4.3. Loss Functions Definition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Despite adding geometric guidance to adaptive density control, Gaussians may still drift away from the correct locations during parameter optimization due to the weak constraints from color consistency loss at textureless regions, resulting in floaters or holes in the test views. Therefore, we add two geometric regularization losses to ensure the optimization proceeds in the right direction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "Monocular depth regularization loss. Monocular depth estimation faces challenges such as scale ambiguity, uneven scale distribution, and a lack of strict geometric consistency between depth maps. A simple linear transformation of monocular depth may destabilize 3D points and introduce distortions during alignment. To address this, the Pearson correlation coefficient (PCC) [8] can be used on small blocks of the depth map to extract robust geometric priors from monocular depth rather than imposing strict consistency. Specifically, we randomly sample " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " non-overlapping patches to compute the monocular depth regularization loss as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 358, + 224, + 555, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 224, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 358, + 224, + 555, + 257 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\frac {1}{N} \\sum_ {i} ^ {N} 1 - P C C \\left(\\hat {\\boldsymbol {P}} _ {i}, \\boldsymbol {P} _ {\\text {m o n o , i}}\\right), \\tag {12}", + "image_path": "79b0f07181d53fd33621d7adb86da7c5bebc5fc9ee2c5ef30a059a724f55b3c2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\hat{P}_i\\in \\mathbb{R}^{S^2}" + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "text", + "content": "th patch of " + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{mono,i}}\\in \\mathbb{R}^{S^2}" + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "text", + "content": "th patch of " + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "inline_equation", + "content": "D_{mono}" + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "text", + "content": ", with the patch size " + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "text", + "content": " being a hyperparameter. PCC, similar to normalized cross-correlation, measures the similarity between the two depth patches. This loss function encourages local consistency between rendered and monocular depth maps, improving alignment over small regions while handling local variations, thus enhancing depth estimation accuracy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 364, + 554, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 554, + 412 + ], + "type": "text", + "content": "Normal consistency loss. Following RaDeGS [48], we measure the consistency between the normal directions computed from the Gaussians and those derived from the depth map:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 376, + 419, + 554, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 419, + 554, + 439 + ], + "spans": [ + { + "bbox": [ + 376, + 419, + 554, + 439 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n c} = \\sum \\left(1 - \\hat {\\boldsymbol {N}} _ {i j} \\cdot \\vec {\\boldsymbol {N}} _ {i j}\\right), \\tag {13}", + "image_path": "1d7c53bfef43b5ec6c97dbb59e6af2cf1d525c79caf34eb47275093f651888ab.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 447, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 554, + 485 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 447, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\hat{N}_{ij}" + }, + { + "bbox": [ + 313, + 447, + 554, + 485 + ], + "type": "text", + "content": " is the rendered normal at pixel(i, j) from Eq. (7), " + }, + { + "bbox": [ + 313, + 447, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\vec{N}_{ij}" + }, + { + "bbox": [ + 313, + 447, + 554, + 485 + ], + "type": "text", + "content": " is the normal estimated by the gradient of the depth map at pixel(i, j)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 485, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 509 + ], + "type": "text", + "content": "Together with " + }, + { + "bbox": [ + 313, + 485, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{n}}" + }, + { + "bbox": [ + 313, + 485, + 554, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 485, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 485, + 554, + 509 + ], + "type": "text", + "content": " used in the tangent space densification, our final loss function is defined as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 336, + 517, + 553, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 517, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 336, + 517, + 553, + 532 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\mathrm {n}} + \\lambda_ {\\mathrm {s}} \\mathcal {L} _ {\\mathrm {s}} + \\lambda_ {\\mathrm {n c}} \\mathcal {L} _ {\\mathrm {n c}}, \\tag {14}", + "image_path": "17c58192b7612568c9743f829c4e2283f5264275c1cd7986f20bbad414804cd2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 539, + 554, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 554, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 539, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 313, + 539, + 554, + 575 + ], + "type": "text", + "content": " is the photometric loss proposed in 3DGS [15]. We set " + }, + { + "bbox": [ + 313, + 539, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{d}} = 0.1" + }, + { + "bbox": [ + 313, + 539, + 554, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 539, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{n}} = 0.5" + }, + { + "bbox": [ + 313, + 539, + 554, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 539, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{s}} = 100.0" + }, + { + "bbox": [ + 313, + 539, + 554, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 539, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{nc}} = 0.1" + }, + { + "bbox": [ + 313, + 539, + 554, + 575 + ], + "type": "text", + "content": " in our experiments." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 586, + 395, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 395, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 395, + 599 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": "In this section, we first demonstrate our IndoorGS's SOTA performance. We compare Gaussian-based methods, known for their high-quality rendering and fast training and rendering speed. We compare our method with 3DGS [15], GOF [47], RaDeGS [48] and PGSR [5]. And PGSR is the state-of-the-art surface reconstruction approach. We then analyze the contribution of each proposed component." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Dataset. We evaluate the performance of our approach on geometry reconstruction and rendering quality using 14" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "849" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 555, + 342 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 555, + 342 + ], + "type": "image", + "image_path": "b134ebe4a2a7e3aafd730e077a87fb78f811b87a76b78ed11547bffd359093d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 350, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 350, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 555, + 384 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison for geometry on Replica(the left two columns), Scannet++(the middle two columns), and Deep Blending(the right two columns) datasets. The experimental results demonstrate IndoorGS's robustness performance in less textured areas and ability to capture fine details as well." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 395, + 541, + 467 + ], + "blocks": [ + { + "bbox": [ + 70, + 395, + 541, + 467 + ], + "lines": [ + { + "bbox": [ + 70, + 395, + 541, + 467 + ], + "spans": [ + { + "bbox": [ + 70, + 395, + 541, + 467 + ], + "type": "table", + "html": "
ReplicaScannet++Deep BlendingAverage on all scenes
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓Time
3DGS39.520.9680.12129.460.9180.15229.410.9030.24332.790.9290.17210m
GOF40.980.9800.07429.820.9190.15329.210.9040.24633.330.9340.1582h
RaDeGS41.040.9800.07430.080.9210.14729.460.9070.24333.530.9360.15515m
PGSR40.020.9740.10129.990.9200.15429.220.8930.25533.080.9290.17025m
Ours41.110.9800.07330.450.9280.13130.060.9100.23533.880.9390.14635m
", + "image_path": "8901497c2772f66a5b991e6044bfb29c3de601ed0f321f823cd7d23d5f57660a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 514, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 552 + ], + "type": "text", + "content": "indoor scenes from publicly available datasets. This includes 8 scenes from Replica [36], 4 scenes from ScanNet++ [43] and 2 scenes from Deep Blending [11]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 554, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 638 + ], + "type": "text", + "content": "Implementation details. We choose RaDeGS [48] as our baseline. Since our initialization points are denser than vanilla 3DGS, the densification stops earlier at 12,000 iterations. To extract the final mesh, we apply Truncated Signed Distance Function (TSDF) [27] surface reconstruction, then use Marching Cube [40] to extract mesh. All experiments are conducted on a single Nvidia RTX4080 GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Metrics. We chose three widely used image evaluation metrics for rendering evaluation: PSNR, SSIM, and LPIPS [49]. For mesh evaluation, we follow the evaluation protocol from monosdf [45] and Go-surf [38], report Chamfer- " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " distance " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{C - }L_{1})" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " , F-score (F1) with a threshold of 5cm and Normal Consistency (NC)." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 512, + 558, + 588 + ], + "blocks": [ + { + "bbox": [ + 55, + 474, + 555, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 474, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 555, + 495 + ], + "type": "text", + "content": "Table 1. Quantitative comparison results of rendering quality for novel view synthesis on the Replica, Scannet++, and Deep Blending datasets. \"Red\" denote the best results, \"yellow\" denote the second best results. Test results on multiple datasets show that IndoorGS outperforms the other three methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 512, + 558, + 588 + ], + "lines": [ + { + "bbox": [ + 317, + 512, + 558, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 558, + 588 + ], + "type": "table", + "html": "
ReplicaScannet++Average on all scenes
C-L1↓F1↑NC↑C-L1↓F1↑NC↑C-L1↓F1↑NC↑
GOF0.0860.5760.7970.1050.4790.7020.0950.5280.749
RaDeGS0.0790.6000.8620.0960.4970.7610.0870.5490.811
PGSR0.0490.8180.8950.0790.6250.7780.0640.7220.837
Ours0.0360.8250.9400.0400.7760.8770.0380.8010.909
", + "image_path": "4fa0322421ceae312c4ff668a8f0f5675f9c9d154433e461220fef843885f275.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 595, + 555, + 634 + ], + "lines": [ + { + "bbox": [ + 313, + 595, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 595, + 555, + 634 + ], + "type": "text", + "content": "Table 2. Quantitative comparison results of Chamfer distance, F-Score and Normal Consistency for reconstruction on the Replica and Scannet++ datasets. IndoorGS achieves the highest reconstruction accuracy across all metrics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 658, + 424, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 424, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 424, + 672 + ], + "type": "text", + "content": "5.1. Mesh Comparison" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "IndoorGS demonstrates superior performance in indoor scene reconstruction, achieving the highest scores across all evaluated metrics—Chamfer Distance, F-Score, and Nor" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "850" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 220 + ], + "type": "table", + "html": "
MethodGeometric CuesGeometric-cue-guided ADCLossMeshNovel View Synthesis
LineSfMPlaneLineSfMPlane\\(L_d\\)\\(L_n\\)\\(L_{nc}\\)C-L1↓F1↑NC↑PSNR↑SSIM↑LPIPS↓
w/o Line & Plane & SfM cues------✓✓✓0.0550.6810.85930.3080.9240.137
w/o Line & Plane cues-✓--✓-✓✓✓0.0510.7170.78930.5130.9270.135
w/o Plane cues✓✓-✓✓-✓✓✓0.0500.7280.86230.5830.9270.134
w/o All customized ADC✓✓✓---✓✓✓0.0520.7110.83930.4990.9260.135
w/o Line ADC✓✓✓-✓✓✓✓✓0.0510.7200.84830.5950.9260.135
w/o SfM ADC✓✓✓✓-✓✓✓✓0.0400.7710.87530.5730.9270.133
w/o Plane ADC✓✓✓✓✓-✓✓✓0.0460.7410.87130.3900.9250.137
w/o \\(L_d\\)✓✓✓✓✓✓-✓✓0.0440.7620.86430.5530.9260.134
w/o \\(L_n\\)✓✓✓✓✓✓✓-✓0.0510.7250.84830.5100.9260.137
w/o \\(L_{nc}\\)✓✓✓✓✓✓✓✓-0.0440.7440.80830.5050.9250.136
Full model✓✓✓✓✓✓✓✓✓0.0400.7760.87730.4530.9280.131
", + "image_path": "f329bb9377e6770e9ec158f5f507355916c51ce680c7e1c107333e5ebb24191e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 555, + 249 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 555, + 249 + ], + "type": "text", + "content": "Table 3. Ablation study results. Geometric cues significantly improve reconstruction quality. We report mesh and novel view synthesis metrics on the Scannet++ dataset. Adding the SfM cues means using the filtered SfM points instead, and using " + }, + { + "bbox": [ + 55, + 228, + 555, + 249 + ], + "type": "inline_equation", + "content": "L_{n}" + }, + { + "bbox": [ + 55, + 228, + 555, + 249 + ], + "type": "text", + "content": " implies incorporating both " + }, + { + "bbox": [ + 55, + 228, + 555, + 249 + ], + "type": "inline_equation", + "content": "L_{n}" + }, + { + "bbox": [ + 55, + 228, + 555, + 249 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 555, + 249 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 55, + 228, + 555, + 249 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 268, + 294, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 268, + 294, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 294, + 389 + ], + "type": "text", + "content": "mal Consistency, on all of the Replica, Scannet++, and Deep Blending datasets as shown in Tab. 2 and Fig. 5. This consistent excellence highlights its effectiveness in accurately reconstructing complex indoor environments. Since the Deep Blending dataset does not include directly provided mesh ground truth (GT), we do not perform quantitative metrics calculations. In contrast to other methods, which often result in uneven surfaces in weakly textured regions and overly smooth surfaces along edges, IndoorGS achieves a more accurate and detailed reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 188, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 188, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 188, + 408 + ], + "type": "text", + "content": "5.2. Rendering Comparison" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 413, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 413, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 413, + 295, + 665 + ], + "type": "text", + "content": "To evaluate the rendering quality of our method, we compared the test-view rendering results of our approach with those from vanilla 3DGS, GOF, RaDeGS, and PGSR on Replica, ScanNet++, and Deep Blending datasets. As shown in Tab. 1, Our method outperforms all of the above Gaussian-based reconstruction methods in terms of novel view synthesis, achieving an average PSNR that is 0.35 dB higher than the second-best method across the three datasets. Notably, following vanilla 3DGS, our test set is obtained by sequential sampling from the original dataset. The test views are positioned close to the training views in indoor scenes. In this case, despite the model overfitting the training views and the geometry inaccurately estimated, the PSNR of the rendered images from the test views can remain relatively high, which undermines the effectiveness of the test set in evaluating the true capability of the model, as it fails to challenge the model with sufficiently diverse perspectives. We provide a detailed experimental analysis of this issue in the appendix. Additionally, our method takes only 10 minutes longer than PGSR despite including the 3D line map construction and planar cues extraction process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 156, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 156, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 156, + 684 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "We conducted an ablation study on the ScanNet++ dataset to test the impact of each component we introduced." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 268, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 460 + ], + "type": "text", + "content": "Specifically, we analyzed the effects of geometric cues, geometric-cue-guided ADC strategies, depth regularization, normal regularization, and normal consistency. The results are presented in Tab. 3. These results show that our full model performs best across all three mesh metrics and the final two NVS metrics (SSIM and LPIPS). Notably, as shown in the first three rows of the table, the step-by-step addition of geometric cues and the corresponding ADC strategy produces incremental performance improvement. Rows 4 to 7 further illustrate that introducing various ADC strategies after adding geometric cues results in further performance enhancement, and line-cue-guided ADC has the most significant impact among the three strategies. The final three rows show the positive effects of the three loss functions, with normal regularization providing the greatest enhancement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 497, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 629 + ], + "type": "text", + "content": "In this paper, we propose IndoorGS, an indoor scene reconstruction method based on 3DGS. Our IndoorGS utilizes geometric cues such as filtered SfM points, line segments, and weakly textured planar surfaces, commonly found in indoor environments, as input data and takes a geometric-cues-guided ADC strategy for Gaussian scene optimization. We validate our method's rendering and reconstruction quality on the Replica, ScanNet++, and Deep Blending datasets. Experimental results demonstrate that our IndoorGS outperforms existing Gaussian-based techniques in geometric reconstruction accuracy and rendering quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 644, + 407, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 407, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 407, + 658 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Project supported by the National Natural Science Foundation of China (Grant No. 62302174). The computation is completed in the HPC Platform of Huazhong University of Science and Technology." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 758 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 758 + ], + "type": "text", + "content": "851" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Jiayang Bai, Letian Huang, Jie Guo, Wen Gong, Yuanqi Li, and Yanwen Guo. 360-gs: Layout-guided panoramic gaussian splatting for indoor roaming. arXiv preprint arXiv:2402.00763, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 190 + ], + "type": "text", + "content": "[2] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 255 + ], + "type": "text", + "content": "[3] Danpeng Chen, Nan Wang, Runsen Xu, Weijian Xie, Hujun Bao, and Guofeng Zhang. Rnin-vio: Robust neural inertial navigation aided visual-inertial odometry in challenging scenes. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 275-283. IEEE, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "text", + "content": "[4] Danpeng Chen, Shuai Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Hujun Bao, and Guofeng Zhang. Vip-slam: An efficient tightly-coupled rgb-d visual inertial planar slam. In 2022 International Conference on Robotics and Automation (ICRA), pages 5615-5621. IEEE, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 313, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 295, + 367 + ], + "type": "text", + "content": "[5] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 401 + ], + "type": "text", + "content": "[6] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 402, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 295, + 435 + ], + "type": "text", + "content": "[7] Yutong Chen, Marko Mihajlovic, Xiyi Chen, Yiming Wang, Sergey Prokudin, and Siyu Tang. Splatformer: Point transformer for robust 3d gaussian splatting, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "text", + "content": "[8] Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. Pearson correlation coefficient. Noise reduction in speech processing, pages 1-4, 2009. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 524 + ], + "type": "text", + "content": "[9] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 579 + ], + "type": "text", + "content": "[10] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 580, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 624 + ], + "type": "text", + "content": "[11] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 680 + ], + "type": "text", + "content": "[12] Manuel Hofer, Michael Maurer, and Horst Bischof. Line3d: Efficient 3d scene abstraction for the built environment. In Pattern Recognition: 37th German Conference, GCPR 2015, Aachen, Germany, October 7-10, 2015, Proceedings 37, pages 237-248. Springer, 2015. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] Manuel Hofer, Michael Maurer, and Horst Bischof. Efficient 3d scene abstraction using line segments. Computer Vision and Image Understanding, 157:167-178, 2017. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[14] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 162 + ], + "type": "text", + "content": "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 554, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 196 + ], + "type": "text", + "content": "[16] Jiyeop Kim and Jongwoo Lim. Integrating meshes and 3d gaussians for indoor scene reconstruction with sam mask guidance. arXiv preprint arXiv:2407.16173, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 554, + 252 + ], + "type": "text", + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "text", + "content": "[18] Yanyan Li, Raza Yunus, Nikolas Brasch, Nassir Navab, and Federico Tombari. Rgb-d slam with structural regularities. In 2021 IEEE international conference on Robotics and automation (ICRA), pages 11581-11587. IEEE, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 298, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 554, + 342 + ], + "type": "text", + "content": "[19] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering. arXiv preprint arXiv:2403.11324, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 555, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 555, + 387 + ], + "type": "text", + "content": "[20] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "text", + "content": "[21] Shuhong Liu, Heng Zhou, Liuzhuozheng Li, Yun Liu, Tianchen Deng, Yiming Zhou, and Mingrui Li. Structure gaussian slam with Manhattan world hypothesis. arXiv preprint arXiv:2405.20031, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 554, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 454 + ], + "type": "text", + "content": "[22] G Lowe. Sift-the scale invariant feature transform. Int. J, 2 (91-110):2, 2004. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 456, + 554, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 554, + 511 + ], + "type": "text", + "content": "[23] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 513, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 554, + 555 + ], + "type": "text", + "content": "[24] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 557, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 554, + 601 + ], + "type": "text", + "content": "[25] Run Luo, Zikai Song, Lintao Ma, Jinlin Wei, Wei Yang, and Min Yang. Diffusiontrack: Diffusion model for multi-object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3991-3999, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 602, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 554, + 667 + ], + "type": "text", + "content": "[26] Pierre Moulon, Pascal Monasse, Romuald Perrot, and Renaud Marlet. Openmvg: Open multiple view geometry. In Reproducible Research in Pattern Recognition: First International Workshop, RRPR 2016, Cancún, Mexico, December 4, 2016, Revised Selected Papers 1, pages 60-74. Springer, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "type": "text", + "content": "[27] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and track" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "852" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "ing. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "type": "text", + "content": "[28] Rémi Pautrat, Juan-Ting Lin, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Sold2: Self-supervised occlusion-aware line description and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11368-11378, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "text", + "content": "[29] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17327-17336, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 295, + 259 + ], + "type": "text", + "content": "[30] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 293 + ], + "type": "text", + "content": "[31] Radu Bogdan Rusu and Steve Cousins. 3d is here: Point cloud library (pcl). In 2011 IEEE international conference on robotics and automation, pages 1-4. IEEE, 2011. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 295, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 295, + 336 + ], + "type": "text", + "content": "[32] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "text", + "content": "[33] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint arXiv:2312.05133, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 426 + ], + "type": "text", + "content": "[34] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International journal of computer vision, 80:189-210, 2008. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "type": "text", + "content": "[35] Xiaowei Song, Jv Zheng, Shiran Yuan, Huan-ang Gao, Jingwei Zhao, Xiang He, Weihao Gu, and Hao Zhao. Saags: Scale-adaptive gaussian splatting for training-free antialIASing. arXiv preprint arXiv:2403.19615, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 524 + ], + "type": "text", + "content": "[36] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 295, + 568 + ], + "type": "text", + "content": "[37] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 613 + ], + "type": "text", + "content": "[38] Jingwen Wang, Tymoteusz Bleja, and Lourdes Agapito. Gosurf: Neural feature grid optimization for fast, high-fidelity rgb-d surface reconstruction. In 2022 International Conference on 3D Vision (3DV), pages 433-442. IEEE, 2022. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 656 + ], + "type": "text", + "content": "[39] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free-view synthesis of indoor scenes. arXiv preprint arXiv:2405.17958, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 689 + ], + "type": "text", + "content": "[40] LORENSEN WE. Marching cubes: A high resolution 3d surface construction algorithm. Computer graphics, 21(1): 7-12, 1987. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "text", + "content": "[41] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 710 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "text", + "content": "4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "text", + "content": "[42] Haodong Xiang, Xinghui Li, Xiansong Lai, Wanting Zhang, Zhichao Liao, Kai Cheng, and Xueping Liu. Gaussianroom: Improving 3d gaussian splatting with sdf guidance and monocular cues for indoor scene reconstruction. arXiv preprint arXiv:2405.19671, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "text", + "content": "[43] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "text", + "content": "[44] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9043-9053, 2023. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "type": "text", + "content": "[45] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 330, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 330, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 330, + 553, + 384 + ], + "type": "text", + "content": "[46] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 386, + 553, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 429 + ], + "type": "text", + "content": "[47] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 430, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 474 + ], + "type": "text", + "content": "[48] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 475, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 529 + ], + "type": "text", + "content": "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 531, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 553, + 574 + ], + "type": "text", + "content": "[50] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 576, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 576, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 576, + 553, + 620 + ], + "type": "text", + "content": "[51] Hang Zhou, Jiale Cai, Yuteng Ye, Yonghui Feng, Chenxing Gao, Junqing Yu, Zikai Song, and Wei Yang. Video anomaly detection with motion and appearance guided patch diffusion model. arXiv preprint arXiv:2412.09026, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 621, + 553, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 621, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 316, + 621, + 553, + 675 + ], + "type": "text", + "content": "[52] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 677, + 553, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 553, + 710 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 553, + 710 + ], + "type": "text", + "content": "[53] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "853" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/6b151e98-3001-4e0f-b77a-ea1941036ad2_content_list.json b/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/6b151e98-3001-4e0f-b77a-ea1941036ad2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..700e34805b789d85547f518360f850c6388eee03 --- /dev/null +++ b/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/6b151e98-3001-4e0f-b77a-ea1941036ad2_content_list.json @@ -0,0 +1,1813 @@ +[ + { + "type": "text", + "text": "Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications", + "text_level": 1, + "bbox": [ + 94, + 128, + 903, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tong $\\mathsf{Bu}^{1,3\\dagger}$ , Maohua $\\mathsf{Li}^{2\\dagger}$ , Zhaofei $\\mathsf{Yu}^{1,3*}$", + "bbox": [ + 333, + 189, + 658, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\dagger$ Equal contribution, \\* Corresponding author", + "bbox": [ + 367, + 210, + 632, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "putong30@pku.edu.cn, maohuali@hhu.edu.cn, yuzf12@pku.edu.cn", + "bbox": [ + 233, + 229, + 756, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 292, + 326, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spiking Neural Networks (SNNs) have emerged as a promising substitute for Artificial Neural Networks (ANNs) due to their advantages of fast inference and low power consumption. However, the lack of efficient training algorithms has hindered their widespread adoption. Even efficient ANN-SNN conversion methods necessitate quantized training of ANNs to enhance the effectiveness of the conversion, incurring additional training costs. To address these challenges, we propose an efficient ANN-SNN conversion framework with only inference scale complexity. The conversion framework includes a local threshold balancing algorithm, which enables efficient calculation of the optimal thresholds and fine-grained adjustment of the threshold value by channel-wise scaling. We also introduce an effective delayed evaluation strategy to mitigate the influence of the spike propagation delays. We demonstrate the scalability of our framework in typical computer vision tasks: image classification, semantic segmentation, object detection, and video classification. Our algorithm outperforms existing methods, highlighting its practical applicability and efficiency. Moreover, we have evaluated the energy consumption of the converted SNNs, demonstrating their superior low-power advantage compared to conventional ANNs. This approach simplifies the deployment of SNNs by leveraging open-source pre-trained ANN models, enabling fast, low-power inference with negligible performance reduction. Code is available at https://github.com/putshua/Inference-scale-ANN-SNN.", + "bbox": [ + 88, + 325, + 485, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 763, + 220, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in large models have significantly reshaped the landscape of deep learning technology, industry, and the research community. These advanced models, characterized by their massive scale and unprecedented zero", + "bbox": [ + 89, + 789, + 485, + 851 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/61ab8c1b4dc0d063529ff2c2da68a7ba568537585be7cc2f9114ae143f00450a.jpg", + "image_caption": [ + "Figure 1. Compared to existing frameworks that require retraining a quantized ANN, the proposed framework is able to directly convert a pre-trained ANN to an SNN at inference-scale complexity, significantly reducing computational requirements, minimizing dependence on GPUs, and requiring only a small subset of the original dataset." + ], + "image_footnote": [], + "bbox": [ + 516, + 277, + 908, + 406 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shot generalizability in downstream tasks, greatly impact our daily lives. Nevertheless, the pursuit of larger models raises concerns about high energy consumption for model inference and training. The deployment of large models on resource-constrained devices has also become a challenge.", + "bbox": [ + 511, + 521, + 906, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spiking neural networks (SNNs), a well-established type of neural network that mimic the function of biological neurons and are considered the third generation of ANNs [44], offer a potential solution to these issues as an alternative to Artificial Neural Networks (ANNs). Unlike ANNs, spiking neurons encode information as discrete events (binary spikes or action potentials) and generate outputs based on both their inputs and internal state. This unique processing mechanism enables SNNs to effectively handle spatio-temporal information while leveraging sparse representations, leading to more efficient and biologically plausible information processing.", + "bbox": [ + 511, + 597, + 908, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the development of the neuromorphic computing hardware [10, 11, 17, 41, 42, 49, 50, 68, 69], SNNs can now be deployed on neuromorphic chips and further applied in power-limited scenarios [9, 56]. However, the lack of efficient training algorithms has hindered their widespread application. Recent learning methods have made significant progress in training deep convolutional SNNs [15, 26] and large spiking transformers using supervised learning algorithms [57, 67, 70]. While supervised training enables SNNs", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Institute for Artificial Intelligence, Peking University", + "bbox": [ + 112, + 862, + 403, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{2}$ School of Artificial Intelligence and Automation, Hohai University", + "bbox": [ + 112, + 875, + 477, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{3}$ School of Computer Science, Peking University", + "bbox": [ + 112, + 887, + 377, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24387", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to achieve performance comparable to ANNs with the same architecture and reduces inference time, the memory and time costs for training scale linearly with the number of inference time-steps, making it impractical for training large energy-efficient models. A more feasible approach is ANN-to-SNN conversion [6], which transforms pre-trained ANNs into SNNs with minimal computational overhead. However, converted SNNs typically require more inference time-steps to match the performance of their ANN counterparts, leading to increased latency and higher energy consumption. A potential solution is to re-train a modified ANN before conversion, which can significantly enhance SNN performance while reducing inference time-steps [4, 29]. Although retraining-based conversion methods generally improve performance across most tasks, the additional computational cost of re-training poses a notable overhead.", + "bbox": [ + 89, + 90, + 485, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose an efficient post-training ANN-SNN conversion algorithm that directly obtains high-performance SNNs from pre-trained ANNs with inference-scale complexity. Figure 1 illustrates the key difference between existing ANN-to-SNN conversion frameworks and our proposed conversion framework. Compared to training-required methods, our approach greatly reduces the need for GPU resources, requires fewer data samples, and shortens the overall conversion process. We employ a local-learning-rule based algorithm for rapid conversion from ANNs to SNNs, avoiding the requirements for computationally expensive backpropagation-based training or fine-tuning. Additionally, we introduce a delayed evaluation strategy to further enhance performance. This conversion framework simplifies the deployment of SNNs by leveraging open-source, pre-trained ANN models, allowing direct application of converted SNNs on neuromorphic hardware with only inference-level computational costs. Overall, the proposed conversion framework offers two potential advantages. First, the lightweight local learning algorithm reduces dependence on GPUs, enabling online mapping from ANN to SNN on neuromorphic chips. Second, this approach reduces the need for extensive retraining of models, making it highly suitable for the practical deployment of large models in resource-constrained environments. Our contribution can be summarized as follows:", + "bbox": [ + 91, + 338, + 486, + 729 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce an inference-scale complexity ANN-SNN conversion framework, enabling the conversion of pretrained ANN models into high-performance SNNs while maintaining computational costs at inference level or even lower.", + "- We propose an efficient ANN-SNN conversion framework, theoretically deriving an upper bound for the conversion error between the original ANN and the converted SNN. To minimize this error bound, we introduce a light-weight local learning method for threshold optimization. Additionally, we design a delayed evaluation strategy to mini" + ], + "bbox": [ + 89, + 734, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gate the influence of spike propagation delays.", + "bbox": [ + 526, + 90, + 834, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We experimentally demonstrate that our method is a lightweight, plug-and-play solution capable of integrating with various conversion methods for further performance improvement. Our framework successfully scales to three typical computer vision tasks, outperforming existing ANN-SNN conversion methods in classification tasks while also demonstrating its practical applicability in handling regression-based vision tasks.", + "bbox": [ + 513, + 106, + 906, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 239, + 663, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Researchers have endeavored to explore effective SNN training approaches for decades. Early studies mainly focused on the unsupervised Spike-Timing-Dependent Plasticity (STDP) algorithm [7, 25]. These studies aimed to utilize the STDP algorithm for shallow networks and feature extractors [13, 34, 46, 59, 62]. In parallel, supervised training approaches have been developed to achieve high-performance SNNs. Bohte et al. [1] were pioneers in applying the backpropagation algorithm to train SNNs, introducing a temporal-based backpropagation method that uses timestamps as intermediate variables during gradient computation. Wu et al. [64] treated SNNs as recurrently connected networks and utilized Back Propagation Through Time (BPTT) for training, termed spatial-temporal backpropagation (STBP). Unlike these methods, Cao et al. [6] proposed an ANN-SNN conversion method that achieves high-performance SNNs from pre-trained ANNs. Currently, research efforts are primarily focused on enhancing the efficiency of training algorithms for high-performance SNNs, leveraging the three methods mentioned above.", + "bbox": [ + 511, + 266, + 908, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on whether global backpropagation is required, SNN learning methods can be ascribed into post-training and training-involved approaches. Training-required methods, including direct training of SNNs [17, 18, 32, 48, 58, 64-66], re-training before ANN-SNN conversion [5, 12, 23, 27, 31, 38, 63] and fine-tuning of converted SNNs [2, 37, 51, 52], typically require the participation of the back-propagation algorithms, making them more computationally intensive and resource-demanding. Post-training learning methods mainly refer to conversion-based algorithms that directly convert the pre-trained ANN into SNN. Cao et al. [6] pioneered this approach by mapped the weights of a lightweight CNN network to an SNN, demonstrating significant energy efficiency improvements through hardware analysis. Diehl et al. [14] proposed a weight and threshold balancing method to mitigate performance degradation, while Hunsberger and Eliasmith [28] further extended conversion-based learning to LIF neurons. Rueckauer et al. [54] established a theoretical framework for conversion-based algorithms, leading to further optimizations such as robust normalization, an enhanced threshold balancing method designed to optimize the trade-off between inference latency and accuracy [55].", + "bbox": [ + 511, + 569, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24388", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Han and Roy [20] extended ANN-to-SNN conversion to temporal coding SNNs, improving accuracy and introducing a reset-by-subtraction mechanism instead of the traditional reset-to-zero, thereby preventing information loss [21]. Beyond classification tasks, conversion-based methods have been successfully applied to regression problems such as object detection, with Kim et al. [35] demonstrating the feasibility of spike-based object detection models. Further optimizations include adding an initial membrane potential for improved accuracy and reduced latency [4, 22], as well as employing a lightweight conversion pipeline to optimize thresholds and introduce a bias term at each inference step [36]. Although performance is limited, those lightweight training methods are able to acquire SNNs from pre-trained ANNs with negligible cost of multiple iterations of inference.", + "bbox": [ + 89, + 90, + 483, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 347, + 230, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Neuron Models", + "text_level": 1, + "bbox": [ + 89, + 372, + 246, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core components of conventional ANNs are point neurons, where their forward propagation can be represented as a combination of linear transformations and non-linear activations, that is", + "bbox": [ + 89, + 393, + 483, + 454 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} ^ {l} = A \\left(\\boldsymbol {w} ^ {l} \\boldsymbol {a} ^ {l - 1}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 465, + 483, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\boldsymbol{w}^l$ is the weights between layer $l - 1$ and layer $l$ , $\\boldsymbol{a}^l$ is the activation vector in layer $l$ , and $A(\\cdot)$ denotes the non-linear activation function.", + "bbox": [ + 89, + 496, + 483, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The neurons in SNNs are spiking neurons with temporal structures. Like most ANN-SNN conversion methods [3, 5, 6, 12, 14, 55], we employ the Integrate-and-Fire (IF) neuron [19, 30] in this paper. For computational convenience, we use the following discrete neuron function:", + "bbox": [ + 89, + 542, + 483, + 617 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} ^ {l} \\left(t ^ {-}\\right) = \\boldsymbol {v} ^ {l} (t - 1) + \\boldsymbol {w} ^ {l} \\boldsymbol {s} ^ {l - 1} (t) \\theta^ {l - 1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 628, + 483, + 646 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} ^ {l} (t) = H \\left(\\boldsymbol {v} ^ {l} \\left(t ^ {-}\\right) - \\theta^ {l}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 648, + 483, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} ^ {l} (t) = \\boldsymbol {v} ^ {l} \\left(t ^ {-}\\right) - \\theta^ {l} \\boldsymbol {s} ^ {l} (t). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 670, + 483, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $s^l(t)$ describes whether neurons in layer $l$ fire at timestep $t$ , with firing triggered when the membrane potential before firing $v^l(t^-)$ reaches the firing threshold $\\theta^l$ , as represented by the Heaviside step function $H(\\cdot)$ . To alleviate information loss, we adopt the \"reset-by-subtraction\" mechanism [21, 53] for updating the membrane potential $v^l(t)$ after spike emission. Specifically, instead of resetting to the resting potential, the membrane potential is reduced by $\\theta^l$ after emitting a spike.", + "bbox": [ + 89, + 700, + 485, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.ANN-SNN Conversion", + "text_level": 1, + "bbox": [ + 89, + 848, + 302, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pivotal idea of ANN-SNN conversion is to map the activation of analog neurons to the postsynaptic potential", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(or firing rate) of spiking neurons. Specifically, by defining the initial membrane potential as $\\mathbf{v}^l(0)$ and accumulating the neuron function (Equations 2 and 4) from 1 to T and dividing both sides by time-step T, we obtain:", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\boldsymbol {v} ^ {l} (T) - \\boldsymbol {v} ^ {l} (0)}{T} = \\frac {\\sum_ {t = 1} ^ {T} \\boldsymbol {w} ^ {l} \\boldsymbol {s} ^ {l - 1} (t) \\theta^ {l - 1} - \\sum_ {t = 1} ^ {T} \\boldsymbol {s} ^ {l} (t) \\theta^ {l}}{T}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 159, + 906, + 208 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This equation reveals a linear relationship between the average firing rate of neurons in adjacent layers by defining average postsynaptic potential as $r^l(t) = \\sum_{i=1}^{t} s^l(i) \\theta^l / t$ :", + "bbox": [ + 511, + 217, + 906, + 263 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {r} ^ {l} (T) = \\boldsymbol {w} ^ {l} \\boldsymbol {r} ^ {l - 1} (T) - \\frac {\\boldsymbol {v} ^ {l} (T) - \\boldsymbol {v} ^ {l} (0)}{T}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 272, + 906, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As $T \\to +\\infty$ , the conversion error can generally be assumed to approach zero. Therefore, using the equations presented, a trained ANN model can be converted into an SNN by replacing ReLU neurons with IF neurons, which forms the fundamental principle of ANN-SNN conversion. Since Equations 1 and 6 are not mathematically identical, some conversion error typically remains. To mitigate these errors, threshold balancing [14] and weight scaling algorithms [36, 55] play a crucial role. Both techniques aim to adjust synaptic weights or neuron thresholds according to the distribution range of neuron inputs, effectively reducing clipping errors. Previous studies have shown that weight scaling and threshold balancing are equivalent in effect and can achieve high-performance SNNs [4].", + "bbox": [ + 511, + 311, + 908, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 513, + 535, + 612, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our post-training ANN-SNN conversion framework. We begin by defining the conversion error and deriving its upper bound. After that, we propose a local threshold balancing method based on a local learning rule. Additionally, we introduce the channel-wise threshold balancing technique, the pre-neuron max pooling layer, and the delayed evaluation strategy. These techniques are seamlessly integrated into a comprehensive conversion framework to enhance both performance and efficiency.", + "bbox": [ + 511, + 560, + 908, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Conversion Error Bound", + "text_level": 1, + "bbox": [ + 511, + 705, + 741, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow the conversion error representation from [5] and define the conversion error $e^l$ in layer $l$ as the 2-norm of the difference between the postsynaptic potential of neurons in SNNs and the activation output of neurons in ANNs, that is", + "bbox": [ + 511, + 727, + 906, + 787 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ne ^ {l} = \\left\\| \\mathrm {S} \\left(\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}\\right) - \\mathrm {A} \\left(\\boldsymbol {z} ^ {l}\\right) \\right\\| _ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 795, + 906, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\hat{z}^l = \\pmb{w}^l\\pmb{r}^{l - 1}$ denotes the average input current from layer $l - 1$ in SNNs, and $\\pmb{z}^l = \\pmb{w}^l\\pmb{a}^{l - 1}$ represents the activation input from layer $l - 1$ in ANNs. Both the ANN and SNN models share the same weights, denoted as $\\pmb{w}^l$ . $\\mathrm{A}(z^l)$ represents the output of nonlinear ReLU activation function while", + "bbox": [ + 511, + 824, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24389", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathrm{S}(\\hat{z}^l;\\theta^l) = \\boldsymbol {r}^l (T)$ represents the average output synaptic potential at time-step $T$ with given average input $\\hat{z}^l$", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since our primary objective is to minimize the error of the final outputs, we define the conversion error between the ANN and SNN models as the layer-wise conversion error in the last layer $L$ , denoted as $e_{\\mathrm{model}} = e^{L}$ . A straightforward approach would be to directly use the conversion error between models as a loss function and optimize it, which can be viewed as a variant of knowledge distillation [66]. However, this method is computationally expensive in terms of both time and memory, resembling the challenges of supervised training of SNN. To address this, we scale the conversion error and derive an error upper bound, thereby simplifying the optimization process. We formalize this approach in the following theorem.", + "bbox": [ + 89, + 121, + 483, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 The layer-wise conversion error can be divided into intra-layer and inter-layer errors:", + "bbox": [ + 89, + 329, + 483, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne ^ {l} \\leqslant \\overbrace {\\| \\mathbf {S} (\\hat {\\boldsymbol {z}} ^ {l} ; \\theta^ {l}) - \\mathrm {A} (\\hat {\\boldsymbol {z}} ^ {l}) \\| _ {2}} ^ {\\text {i n t r a - l a y e r e r r o r}} + \\overbrace {\\| \\boldsymbol {w} ^ {l} \\| _ {2} e ^ {l - 1}} ^ {\\text {i n t e r - l a y e r e r r o r}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 369, + 483, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given that both ANN and SNN models receive the same input in the first layer, leading to $e^0 = 0$ , the upper bound for the conversion error between ANN and SNN models in an $L$ -layer fully-connected network is given by", + "bbox": [ + 88, + 417, + 483, + 478 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {\\text {m o d e l}} = e ^ {L} \\leqslant \\sum_ {l = 1} ^ {L} \\left(\\prod_ {k = l + 1} ^ {L} \\| \\boldsymbol {w} ^ {k} \\| _ {2}\\right) \\left\\| \\mathrm {S} \\left(\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}\\right) - \\mathrm {A} \\left(\\hat {\\boldsymbol {z}} ^ {l}\\right) \\right\\| _ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 487, + 483, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\| \\pmb{w}^k \\|_2$ is the matrix norm $(p = 2)$ or spectral norm of the weight matrix. The detailed proof is provided in the Appendix. Theorem 1 indicates that the conversion error is bounded by the weighted sum of errors across all layers. Based on this result, we introduce the local threshold balancing algorithm in the following section.", + "bbox": [ + 89, + 556, + 483, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Local Threshold Balancing Algorithm", + "text_level": 1, + "bbox": [ + 89, + 657, + 419, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The target of ANN-SNN conversion is to minimize the conversion error and achieve high-performance SNNs. An alternative approach is to optimize the error bound, defined as:", + "bbox": [ + 89, + 679, + 485, + 738 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname * {a r g m i n} _ {\\boldsymbol {\\theta}} \\mathbb {E} _ {x ^ {0} \\in \\mathcal {D}} \\\\ \\left[ \\sum_ {l = 1} ^ {L} \\left(\\prod_ {k = l + 1} ^ {L} \\| \\boldsymbol {w} ^ {k} \\| _ {2}\\right) \\left\\| \\mathrm {S} \\left(\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}\\right) - \\mathrm {A} \\left(\\hat {\\boldsymbol {z}} ^ {l}\\right) \\right\\| _ {2} \\right]. \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 748, + 483, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\theta$ represents the set of threshold values to be optimized across all layers, i.e., $\\pmb{\\theta} = \\{\\theta^1,\\theta^2,\\dots,\\theta^L\\}$ . $x^{0}$ denotes the input data sample drawn from the dataset $\\mathcal{D}$ . This approach leverages data-driven conversion, minimizing the expectation of the conversion error bound over the data distribution.", + "bbox": [ + 89, + 824, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To reduce computational costs during optimization, we introduce two key simplifications. Firstly, we use a greedy strategy to optimize the threshold $\\theta$ layer by layer, that is", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each layer $l$ , $\\arg \\min_{\\theta^l} \\mathbb{E}_{x^0 \\in \\mathcal{D}}$", + "bbox": [ + 527, + 145, + 746, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\left(\\prod_ {k = l + 1} ^ {L} \\| \\boldsymbol {w} ^ {k} \\| _ {2}\\right) \\left\\| \\mathbf {S} \\left(\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}\\right) - \\mathbf {A} \\left(\\hat {\\boldsymbol {z}} ^ {l}\\right) \\right\\| _ {2} \\right]. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 170, + 906, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Secondly, we simplify the function $\\mathrm{S}(\\cdot; \\theta)$ by neglecting its temporal dynamics and approximating it with the clipping function $\\mathrm{C}(\\cdot; \\theta)$ , as previous works have demonstrated that spiking neuron and clipping functions behave equivalently given sufficient number of time-steps [12]. The average input in the SNN can therefore be estimated as $\\hat{\\boldsymbol{z}}^0 = \\boldsymbol{z}^0$ ; $\\hat{\\boldsymbol{z}}^{l+1} = \\boldsymbol{w}^{l+1} \\mathrm{C}(\\hat{\\boldsymbol{z}}^l; \\theta^l)$ . Furthermore, We utilize the squared norm in the objective function for smoother optimization. The optimization problem then becomes:", + "bbox": [ + 511, + 220, + 906, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each layer $l$ , $\\arg \\min_{\\theta^l} \\mathbb{E}_{x^0 \\in \\mathcal{D}}$", + "bbox": [ + 527, + 366, + 746, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\left(\\prod_ {k = l + 1} ^ {L} \\| \\boldsymbol {w} ^ {k} \\| _ {2}\\right) \\| \\mathbf {C} (\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}) - \\mathbf {A} (\\hat {\\boldsymbol {z}} ^ {l}) \\| _ {2} ^ {2} \\right], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 391, + 906, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {w h e r e} \\mathrm {C} (x; \\theta) = \\min (\\max (0, x), \\theta). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 441, + 906, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, we can employ the stochastic gradient descent algorithm to optimize the threshold. Since $\\prod_{k = l + 1}^{L}\\| \\pmb{w}^{k}\\|_{2}$ can be considered as a constant when weights are fixed for each layer, this term can be absorbed into the learning rate. The final update rule for the local threshold balancing algorithm at each step is:", + "bbox": [ + 511, + 464, + 905, + 554 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\theta^ {l} = - \\sum_ {i = 0} ^ {N - 1} 2 \\left(\\hat {z} _ {i} ^ {l} - \\theta^ {l}\\right) H \\left(\\hat {z} _ {i} ^ {l} - \\theta^ {l}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 563, + 906, + 603 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {l} \\leftarrow \\theta^ {l} - \\eta \\Delta \\theta^ {l}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 606, + 906, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detailed derivation is provided in the Appendix. Here $\\Delta \\theta^l$ denotes the step size during optimization and $\\eta$ is the learning rate or update step size. $N$ is the number of elements in the input vector, and $\\hat{z}_i$ denotes each element in the vector $\\hat{z}^l$ . $H(\\cdot)$ is the Heaviside step function.", + "bbox": [ + 511, + 633, + 906, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, as shown in Figure 2, we jointly optimize the threshold for each layer by sampling a small set of image samples from the training dataset. Before starting the conversion process, we first replace all ReLU activation functions $\\mathrm{A}(\\cdot)$ with clipping functions $\\mathrm{C}(\\cdot)$ and initialize the threshold to zero. We then randomly sample data batches from the training dataset and perform model inference with these data samples. The thresholds are locally updated during the inference until convergence. Notably, the parameter $\\theta^l$ is updated locally, ensuring that the computational cost of each iteration is similar to the inference of the ANN model and the computational cost of the whole conversion process is at inference level.", + "bbox": [ + 511, + 710, + 908, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24390", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/58effa4ec822ff32ffdf4622e35ecaad2876a2fa3ab410222d5ecabd849e6eab.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed conversion framework. It requires only a small subset of data for conversion using local threshold balancing. During inference, the delayed evaluation technique is employed to enhance the accuracy of output estimation." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 472, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Pre-Neuron Max Pooling Layer", + "text_level": 1, + "bbox": [ + 89, + 287, + 369, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The conversion of the max pooling layer presents a significant challenge. As a pivotal feature in the process of down-sampling input representations, the max pooling layer is a fundamental component in convolutional neural networks and is widely used in most convolutional architectures. However, due to the binary nature of spiking neurons, max pooling layers cannot effectively perform downsampling in feature maps. Typically, in a single max-pooling operation, multiple elements may share the same value, preventing the extraction of the most salient features. Consequently, max pooling is often avoided as a downsampling layer in SNNs. In most previous ANN-SNN methods, the max pooling layer in ANN is often replaced by an average pooling layer, and the model is re-trained or fine-tuned before conversion.", + "bbox": [ + 88, + 310, + 485, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To convert architectures that contain max pooling layers, we propose a simple yet effective method. We replace all max-pooling layers with neuron layers, allowing the preneuron floating-point input to pass through the downsampling layer before being input into the neurons. We have the following theorem:", + "bbox": [ + 89, + 521, + 483, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2 The order of the max pooling layer and ReLU layer does not affect the ANN output results.", + "bbox": [ + 89, + 619, + 483, + 651 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {i} \\mathrm {R} (z) = \\mathrm {R} (\\max _ {i} z), \\text {w h e n} \\max (\\mathrm {z}) > 0. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 659, + 483, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\mathrm{R}(\\cdot)$ denotes the ReLU activation. The detailed proof is provided in the Appendix. Theorem 2 guarantees that such an operation does not affect the final performance. When deployed on neuromorphic chips or FPGAs, the max pooling operation can be efficiently implemented through a comparator, avoiding additional floating-point multiplications and minimizing power consumption.", + "bbox": [ + 89, + 688, + 485, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Channel-wise Threshold Balancing", + "text_level": 1, + "bbox": [ + 89, + 803, + 397, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During conversion, following Kim et al. [35], Li et al. [36], we employ a channel-wise local threshold balancing method applicable to both fully connected and convolutional layers.", + "bbox": [ + 89, + 824, + 483, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this approach, neurons in each channel share a single threshold for convolutional layers while each neuron has an", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "independent threshold for fully connected layers. Specifically, for convolutional layers, we channel-wisely determine the threshold value, while for fully connected layers, we optimize the threshold element-wise. The update function is given by:", + "bbox": [ + 511, + 90, + 906, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\theta_ {c} ^ {l} = - \\sum_ {i = 0} ^ {N - 1} 2 \\left(\\hat {z} _ {c, i} ^ {l} - \\theta_ {c} ^ {l}\\right) H \\left(\\hat {z} _ {c, i} ^ {l} - \\theta_ {c} ^ {l}\\right), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 176, + 906, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c$ denotes the channel index (or element index for fully connected layers) and $i$ denotes the element index within the current channel.", + "bbox": [ + 511, + 228, + 903, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After optimization, the thresholds for each channel can be individually absorbed into the corresponding channel of the convolutional kernel or the weight vectors of the fully connected layers. This channel-wise (or element-wise) operation ensures that the behavior of the IF neuron remains unaffected. This algorithm aims to more precisely determine the optimal threshold value while preserving the fundamental properties of spiking neurons.", + "bbox": [ + 511, + 273, + 906, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Delayed Evaluation Strategy", + "text_level": 1, + "bbox": [ + 511, + 402, + 769, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As introduced in [4], the first spike from the last layer often takes longer than expected to appear due to propagation delays between layers. This delay creates significant fluctuations in the firing rate during the initial time-steps, as no spikes are generated until the first one reaches each layer [29]. To address this issue, we propose an effective delayed step selection strategy building upon the delayed evaluation technique [29].", + "bbox": [ + 511, + 425, + 906, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 2, the basic idea of the delayed evaluation is to compute the average spike count in the last layer from time-step $t_0$ to $T$ , where $t_0$ denotes the delayed step, rather than averaging over all time-steps. Since the output is collected from the average spike count in the interval $[t_0, T]$ at the last layer, we can rewrite Equation 6 for the last layer as", + "bbox": [ + 511, + 546, + 908, + 651 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {r} ^ {L} \\left(t _ {0}, T\\right) = \\boldsymbol {w} ^ {L} \\boldsymbol {r} ^ {L - 1} \\left(t _ {0}, T\\right) - \\frac {\\boldsymbol {v} ^ {L} (T) - \\boldsymbol {v} ^ {L} \\left(t _ {0}\\right)}{T - t _ {0}}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 659, + 906, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\boldsymbol{r}^l(t_0, t) = \\sum_{i=t_0}^{t} \\boldsymbol{s}^l(i) \\theta^l / (t - t_0)$ denotes the average postsynaptic potential from time $t_0$ to $t$ , and $L$ represents the last layer index. When $T \\gg t_0$ , one can still guarantee that the basic equivalence of the average postsynaptic potential and ANN activations holds.", + "bbox": [ + 511, + 703, + 906, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our strategy, we do not start counting the average output spike until $\\widetilde{t_0}$ , where $\\widetilde{t_0}$ is the expected spike timing of the first spike in the last layer. We estimate $\\widetilde{t_0}$ based on the converted SNN model, denoted as", + "bbox": [ + 511, + 781, + 906, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {t _ {0}} = \\sum_ {l = 1} ^ {L} \\frac {\\theta^ {l} - v ^ {l} (0)}{\\max _ {i} \\left(\\overline {{\\mathbf {R} (\\hat {\\boldsymbol {z}} ^ {l})}}\\right)} \\approx \\sum_ {l = 1} ^ {L} \\frac {\\theta^ {l} - v ^ {l} (0)}{\\max _ {i} \\left(\\overline {{\\mathbf {R} (\\boldsymbol {z} ^ {l})}}\\right)}. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 849, + 906, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24391", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\max_i(\\cdot)$ selects the maximum element of a given vector. The notation $\\overline{(\\cdot)} = \\mathbb{E}_{x^0\\in \\mathcal{D}}[(\\cdot)]$ represents the expectation of the given input over dataset. Additionally, we use the ReLU function $\\mathbf{R}(\\cdot)$ to prevent negative estimated time. Before evaluation, we first calculate $\\widetilde{t_0}$ according to Equation 19 and consider it as an indicator for the delayed evaluation strategy. In practice, if the inference step $T$ is less than $\\widetilde{t_0} +4$ , we set $t_0 = T - 4$ and collect the average output over the time interval $[T - 4,T]$ . If $T$ exceeds $\\widetilde{t_0} +4$ , we use the interval $[\\widetilde{t_0},T]$ for output estimation. The value 4 is empirically chosen to ensure a sufficient delay while allowing adequate time for averaging spikes over time.", + "bbox": [ + 89, + 90, + 483, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This technique introduces minimal additional training cost and does not alter the model's inference process, making it seamlessly integrated into the proposed conversion framework without disrupting the overall pipeline.", + "bbox": [ + 89, + 272, + 483, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 345, + 222, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive experiments to demonstrate the effectiveness of our method and highlight its potential practical value. We first showcase the advantages of our approach by integrating it with existing conversion frameworks, comparing it with state-of-the-art post-training conversion algorithms, and conducting ablation experiments to validate the effectiveness of the whole conversion pipeline. Subsequently, we apply our algorithm to various visual tasks, including image classification, semantic segmentation, object detection, and video classification. The feasibility of our algorithm is highlighted on both classification and regression tasks, showcasing its generalizability across different datasets and its superiority over existing conversion algorithms.", + "bbox": [ + 88, + 371, + 483, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To underscore that our conversion method does not require additional training, we utilize open-source pre-trained ANN models for conversion, with most pre-trained sourced from TorchVision [45]. Additionally, we incorporate the membrane potential initialization algorithm [5], setting the membrane potential to half of the threshold before inference. The pseudocode for our conversion pipeline and detailed training settings are provided in the Appendix.", + "bbox": [ + 89, + 566, + 483, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Plug-and-Play Integration for Enhanced Performance", + "text_level": 1, + "bbox": [ + 89, + 696, + 482, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we aim to highlight that our proposed conversion method provides a lightweight, plug-and-play solution that can be seamlessly integrated with other conversion algorithms to enhance performance. We select two representative ANN-SNN conversion approaches from existing work: the post-training fine-tuning (calibration) conversion method LCP/ACP [36, 37], and the training-required conversion method QCFS [5]. We reproduce the conversion pipeline of both methods, incorporate a local threshold balancing method for threshold searching, and use our delayed evaluation strategy. The evaluation is conducted on the ImageNet", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7aa86646071b01c52c865c0ce523f18b998c358beddcc222bfed2c465150ccb8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodANNCFInference Step
3264128256512
LCP75.66×50.2163.6668.8972.1274.17
✓53.8067.1872.1874.0774.56
ACP75.66×64.5471.1273.4574.6175.35
✓65.9572.8774.2174.7375.26
QCFS73.30×64.4371.5173.2573.6773.76
✓71.9473.0673.5073.4973.54
", + "bbox": [ + 539, + 88, + 880, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Integration of the proposed conversion approach with existing framework", + "bbox": [ + 511, + 219, + 903, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/905acf9bacea199407f858162858455ebd244102c3cb873cfe60c8f74598e1f4.jpg", + "image_caption": [ + "(a) Ablation Study" + ], + "image_footnote": [], + "bbox": [ + 517, + 258, + 696, + 369 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/260936a879e100bc21238ee7869cc84fce9a1a415840a38493c06964c355a774.jpg", + "image_caption": [ + "(b) Effect of Iteration Steps", + "Figure 3. (a) Results on evaluating the impact of different techniques within the conversion framework. (b) Effect of iteration steps on SNN performance after conversion." + ], + "image_footnote": [], + "bbox": [ + 720, + 258, + 897, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dataset with the ResNet-34 architecture.", + "bbox": [ + 511, + 455, + 779, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results, as detailed in Table 1, showcase the performance of various configurations. The column labeled \"CF\" indicates whether the proposed method is used. Notably, the composite approaches, which combine our method with either LCP/ACP or QCFS, consistently outperform the individual conversion method across most timesteps. After incorporating our methods, both QCFS and LCP exhibit performance improvements exceeding $3\\%$ when $\\mathrm{T} = 32$ , while maintaining consistent performance as $T$ increases. This demonstrates the advantages of using our approach in conjunction with existing conversion techniques.", + "bbox": [ + 511, + 472, + 908, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Ablation Study for the Conversion Framework", + "text_level": 1, + "bbox": [ + 511, + 651, + 903, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of our proposed conversion framework by using the ResNet-34 architecture on the ImageNet dataset. An ablation study is conducted to evaluate the impact of different techniques within the framework. Specifically, as shown in Figure 3a, we compare the performance of four different SNNs obtained by combinations of different techniques. The baseline SNN is obtained using the Robust-Norm [54] algorithm (RN, red curve), which is a commonly used post-training ANN-SNN conversion method that sets the 99-th percentile activation value as the threshold. The other SNNs are obtained by using the layer-wise local threshold balancing algorithm (LTB, yellow curve), the channel-wise threshold balancing algorithm (CLTB, blue curve), and a combination of channel-wise threshold balancing and delayed evaluation (CLTB+DE, black curve).", + "bbox": [ + 511, + 674, + 910, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24392", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56d476b287f1f33b63b175734fd4e1b03b99d7d037d1c3c5ed10ae35c25b3c9d.jpg", + "image_caption": [ + "Figure 4. Effect of delayed steps" + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 472, + 193 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 3a presents the accuracy changes of the converted SNNs with respect to the increasing time-steps. The performance of the SNN obtained from the full conversion pipeline (black curve) consistently outperforms the other SNNs across all time-steps, demonstrating an excellent balance between fast inference and high performance of the proposed conversion framework. This is particularly evident when combined with the proposed delayed evaluation technique. Compared to the robust norm method (red curve), the peak performance of the SNNs obtained by the proposed local threshold balancing method (black, blue and yellow curves) more closely aligns with the original ANN accuracy. Furthermore, the use of the channel-wise threshold balancing (CLTB, blue curve) results in a noticeable performance improvement over the vanilla threshold balancing (LTB, yellow curve) algorithm.", + "bbox": [ + 89, + 239, + 485, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Effect of the Iteration Step", + "text_level": 1, + "bbox": [ + 89, + 476, + 331, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the influence of the iteration steps of the local threshold balancing algorithm. The iteration number of the local threshold balancing serves as the hyper-parameter of the method. We focus on the most straightforward approach by empirically evaluating the impact of different iteration steps. We converted five different SNNs from the pre-trained ImageNet ResNet-34 model, varying the iterations steps from 1000 to 5000. The batch size during local threshold balancing is consistently set to 100. Figure 3b shows the final accuracy of the different converted SNNs. As number of iterations increases, the peak performance of the converted SNNs at different inference steps approaches the accuracy of the original ANN. Furthermore, these SNNs demonstrate better performance at low time-steps when fewer iteration steps are used. On the ImageNet dataset, with only 1000 iterations of local threshold update, the performance gap between the SNN at 512 inference time-steps and the ANN is approximately $1\\%$ . With 5000 iterations, this difference is further reduced to around $0.2\\%$ . Notably, the accuracy surpasses $65\\%$ within just 55 time-steps, achieving a balance between inference time and performance.", + "bbox": [ + 89, + 498, + 483, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Effect of the Delayed Step", + "text_level": 1, + "bbox": [ + 89, + 824, + 326, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct experiments to demonstrate the effectiveness of the proposed delayed strategy on the ImageNet dataset using the ResNet-34 architecture. As shown in Figure 4, we evaluate the impact of the delayed step across four different", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5f417b2e54911da5c7f1ca199577b8754b8f43f2f38c6d2066efd33ad9f80a4a.jpg", + "table_caption": [], + "table_footnote": [ + "1 Standard ResNet-34 architecture without batch normalization.", + "2 Two more layers in the first conv-block compared to standard ResNet-34.", + "3 Standard ResNet-18/34 architecture." + ], + "table_body": "
ArchitectureMethodANNInference Step
3264128256≥512
Post-Training
ResNet-341RMP70.64---55.6560.08
ResNet-342LCP75.6655.1667.5672.4874.5375.44
MobileNetLCP73.400.2015.4755.9566.3572.19
MobileNetV2Ours71.8819.3949.9364.3168.8970.48
ResNet-183Ours69.7665.5568.5369.0369.2169.21
ResNet-343Ours73.3164.8871.1772.3072.4272.46
Training-involved
ResNet-343QCFS73.3064.4371.5173.2573.6773.76
ResNet-342ACP75.6664.6571.3073.9475.0075.45
ResNet-343Ours*73.3071.9473.0673.5073.4973.54
", + "bbox": [ + 524, + 90, + 897, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of the proposed method and previous works on the ImageNet dataset.", + "bbox": [ + 511, + 349, + 903, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "inference time-step settings: 32 steps (red curve), 64 steps (yellow curve), 128 steps (blue curve), and 256 steps (black curve). The grey vertical line indicates the estimated $\\widetilde{t_0}$ for the currently converted SNN. For each sub-figure, the x-axis represents the number of delayed steps, while the y-axis denotes the accuracy.", + "bbox": [ + 511, + 398, + 905, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although the delayed strategy is pre-determined before inference, the estimated $\\widetilde{t_0}$ can still serves as an indicator that provides a near-optimal strategy to select the delayed steps. When the total inference step is less than the estimated $\\widetilde{t_0}$ (red and yellow curves), the final performance gradually improves with the increase in the delayed steps, but drops significantly when the delayed step approaches the inference step. In the situation that the total inference step exceeds $\\widetilde{t_0}$ (blue and black curve), the performance first increases then fluctuates when delayed step is larger than $\\widetilde{t_0}$ , and ultimately experiences a sharp decline as the delayed steps get close to the inference step. The above results indicate that we can always reach near-optimal performance with the proposed delayed evaluation strategy.", + "bbox": [ + 511, + 489, + 906, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Test on Image Classification task", + "text_level": 1, + "bbox": [ + 511, + 712, + 802, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our conversion method on the classification task with the ImageNet dataset, employing different architectures including ResNet and MobileNet. To highlight the advantages of the proposed conversion algorithm, we conduct comparisons with previous post-training algorithms, showcasing its superior performance. When using the ResNet-34 architecture, the SNN converted by our algorithm exceeds $71\\%$ accuracy within 64 time steps. In contrast, the SNNs converted by the previous algorithm [20] and the calibrationrequired conversion algorithm [37] require longer time-steps to achieve comparable accuracy. We also implement the", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24393", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "proposed conversion framework on QCFS pre-trained models (denote as \"Our*\" in Table 2), and compare with the sota training-involved ANN-SNN techniques. The accuracy of the proposed method reaches $71.94\\%$ within 32 steps, demonstrating significant advantages compared to the fine-tuning required advanced calibration technique ACP [37] and training-involved vanilla QCFS [5]. The experiments on the ImageNet dataset demonstrate the superior performance of the proposed conversion algorithm and highlight its potential for large-scale applications.", + "bbox": [ + 89, + 90, + 482, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition, we demonstrate the energy-saving advantage of our method by theoretically estimating the energy efficiency of our converted SNN. We use the same energy estimation approach as Bu et al. [5]. With ResNet-34 architecture on ImageNet, SNN converted from our algorithm can achieve energy efficiency at 622FPS/W while maintaining $90\\%$ performance. In comparison, the energy efficiency of the original ANN is only 22FPS/W, which is 28 times lower than that of the converted SNN. For detailed energy consumption analysis, please refer to the Appendix.", + "bbox": [ + 89, + 243, + 483, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Test on Semantic Segmentation, Object Detection, and Video Classification Tasks", + "text_level": 1, + "bbox": [ + 89, + 407, + 482, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further extend the proposed ANN-SNN conversion method to semantic segmentation, object detection, and video classification tasks. For the semantic segmentation task, we evaluate our method on two commonly used datasets: Pascal VOC 2012 [16] augmented by SBD [24] and MS COCO 2017 [39]. We employ various semantic segmentation models, including FCN [43] and DeepLabV3 [8], provided by TorchVision [45] and other open-source repositories. As shown in Table 3, our method is compatible with tackling semantic segmentation tasks without any preparatory training. Specifically, we achieved a $53.50\\%$ mIoU with the FCN model using a ResNet-34 backbone in 64 timesteps on the Pascal VOC dataset and a $61.52\\%$ mIoU with the DeepLabV3 model using a ResNet-50 architecture in 128 timesteps on the more complex MS COCO 2017 dataset. Given that pixel-level classification tasks require precise model output, this excellent performance highlights the effectiveness and generalizability of the proposed method.", + "bbox": [ + 89, + 446, + 482, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For object detection tasks, we use the benchmark dataset MS COCO 2017 [39] with open-source models from TorchVision [45]. We employ the fully convolutional object detection method RetinaNet [40] and FCOS [60], and convert the pre-trained ANN models to SNNs using the proposed method. Table 3 shows detailed performance under different inference time-steps. Within 64 time-steps, the converted SNN can achieve $27.3\\%$ mAP with RetinaNet and $26.5\\%$ mAP with FCOS, which represents a significant improvement compared to previous explorations [35, 47]. The previous method required over 1000 time-steps to achieve comparable performance as ANN, which is detrimental to", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8aae9cddbf7cf2a218a85397141cb3194520dfd0f43667c49c6a5f24486bffa7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetArch.ANNInference Step
3264128256
Semantic Segmentation (mIoU%)
PascalVOCFCN-1847.3043.4947.0247.2647.23
FCN-3454.9946.6053.5054.7554.76
DeepLab-1855.7548.5953.9155.4055.60
DeepLab-3458.9541.2055.7458.3658.68
MS COCOFCN-5060.6728.6650.3857.0159.53
DeepLab-5063.0139.0256.6261.5262.61
Object detection (mAP%)
MS COCORetinaNet-5036.416.627.332.234.3
FCOS-5039.212.226.532.133.6
Video Classification (Acc%)
S-Kinetics-400MC3-1863.5255.2161.6062.4362.73
R3D-1862.5653.4660.3161.1261.15
", + "bbox": [ + 524, + 89, + 893, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Performance on semantic segmentation (mIoU%), object detection (mAP%), and video classification (Acc%) tasks.", + "bbox": [ + 511, + 321, + 903, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "real-time detection and energy efficiency.", + "bbox": [ + 511, + 369, + 787, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the video classification tasks, we use a small Kinetics-400 dataset [33], which is a subset of the original dataset. More specifically, we split the original validation set into two independent sets, one for our threshold balancing procedure, accounting for about $60\\%$ of the original validation set, and the other for the test. As a result, the new train set contains 12000 videos and the new test set contains 7881 videos. The converted MC3-18 [61] SNN model achieves $61.60\\%$ accuracy within 64 time-steps, showing great potential of our method in more complex computer vision tasks.", + "bbox": [ + 511, + 385, + 906, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 511, + 550, + 761, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces the concept of a lightweight posttraining ANN-SNN conversion algorithm with inference-scale computational cost, which can directly convert pretrained ANN models into SNNs without GPU-based training. The significant reduction in training overhead highlights the potential of the rapid deployment of large-scale SNNs in various scenarios and makes the instant on-chip conversion from pre-trained ANN to SNN possible. However, there is a tradeoff between performance and training cost, with the overall performance of the proposed post-training framework being slightly weaker compared to other training-involved SNN learning methods. Additionally, while this work focuses on the conversion of convolutional SNNs, future efforts will aim to extend such a conversion approach to transformer architectures and other complex layers.", + "bbox": [ + 511, + 574, + 908, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 815, + 692, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is funded by National Natural Science Foundation of China (62422601,62176003,62088102), Beijing Municipal Science and Technology Program (Z241100004224004) and Beijing Nova Program (20230484362,20240484703).", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24394", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sander M Bohte, Joost N Kok, and Johannes A La Poutre. Spikeprop: backpropagation for networks of spiking neurons. In European Symposium on Artificial Neural Networks, 2000. 2", + "[2] Velibor Bojkovic, Srinivas Anumasa, Giulia De Masi, Bin Gu, and Huan Xiong. Data driven threshold and potential initialization for spiking neural networks. In International Conference on Artificial Intelligence and Statistics, pages 4771-4779. PMLR, 2024. 2", + "[3] Romain Brette, Michelle Rudolph, Ted Carnevale, Michael Hines, David Beeman, James M Bower, Markus Diesmann, Abigail Morrison, Philip H Goodman, Frederick C Harris, et al. Simulation of networks of spiking neurons: a review of tools and strategies. Journal of Computational Neuroscience, 2007. 3", + "[4] Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Optimized potential initialization for low-latency spiking neural networks. In AAAI Conference on Artificial Intelligence, 2022. 2, 3, 5", + "[5] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks. In International Conference on Learning Representations, 2022. 2, 3, 6, 8", + "[6] Yongqiang Cao, Yang Chen, and Deepak Khosla. Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision, 2015. 2, 3", + "[7] Natalia Caporale and Yang Dan. Spike timing-dependent plasticity: a hebbian learning rule. Annual Review of Neuroscience, 31:25-46, 2008. 2", + "[8] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 8", + "[9] Yanqi Chen, Zhaofei Yu, Wei Fang, Zhengyu Ma, Tiejun Huang, and Yonghong Tian. State transition of dendritic spines improves learning of sparse spiking neural networks. In International Conference on Machine Learning, 2022. 1", + "[10] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 2018. 1", + "[11] Michael V DeBole, Brian Taba, Arnon Amir, Filipp Akopyan, Alexander Andreopoulos, William P Risk, Jeff Kusnitz, Carlos Ortega Otero, Tapan K Nayak, Rathinakumar Appuswamy, et al. TrueNorth: Accelerating from zero to 64 million neurons in 10 years. Computer, 2019. 1", + "[12] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. In International Conference on Learning Representations, 2021. 2, 3, 4", + "[13] Peter U Diehl and Matthew Cook. Unsupervised learning of digit recognition using spike-timing-dependent plasticity. Frontiers in Computational Neuroscience, 9:99, 2015. 2" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Peter U Diehl, Daniel Neil, Jonathan Binas, Matthew Cook, Shih-Chii Liu, and Michael Pfeiffer. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In International Joint Conference on Neural Networks, 2015. 2, 3", + "[15] Chaoteng Duan, Jianhao Ding, Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Temporal effective batch normalization in spiking neural networks. In Advances in Neural Information Processing Systems, 2022. 1", + "[16] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision, 2015. 8", + "[17] Biao Fang, Yuhao Zhang, Rui Yan, and Huajin Tang. Spike trains encoding optimization for spiking neural networks implementation in fpga. In International Conference on Advanced Computational Intelligence, 2020. 1, 2", + "[18] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In International Conference on Computer Vision, 2021. 2", + "[19] Wulfram Gerstner, Werner M Kistler, Richard Naud, and Liam Paninski. Neuronal dynamics: From single neurons to networks and models of cognition. Cambridge University Press, 2014. 3", + "[20] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, 2020. 3, 7", + "[21] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. RMP-SNN: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Computer Vision and Pattern Recognition, 2020. 3", + "[22] Zecheng Hao, Tong Bu, Jianhao Ding, Tiejun Huang, and Zhaofei Yu. Reducing ann-snn conversion error through residual membrane potential. arXiv preprint arXiv:2302.02091, 2023. 3", + "[23] Zecheng Hao, Jianhao Ding, Tong Bu, Tiejun Huang, and Zhaofei Yu. Bridging the gap between anns and snns by calibrating offset spikes. In International Conference on Learning Representations, 2023. 2", + "[24] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In International Conference on Computer Vision, 2011. 8", + "[25] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology press, 2005. 2", + "[26] JiaKui Hu, Man Yao, Xuerui Qiu, Yuhong Chou, Yuxuan Cai, Ning Qiao, Yonghong Tian, Xu Bo, and Guoqi Li. High-performance temporal reversible spiking neural networks with O(1) training memory and O(1) inference cost. In International Conference on Machine Learning, 2024. 1", + "[27] Zihan Huang, Xinyu Shi, Zecheng Hao, Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Towards high-performance spiking transformers from ann to snn conversion. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10688-10697, 2024. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24395", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Eric Hunsberger and Chris Eliasmith. Spiking deep networks with LIF neurons. arXiv preprint arXiv:1510.08829, 2015. 2", + "[29] Sungmin Hwang, Jeesoo Chang, Min-Hye Oh, Kyung Kyu Min, Taejin Jang, Kyungchul Park, Junsu Yu, Jong-Ho Lee, and Byung-Gook Park. Low-latency spiking neural networks using pre-charged membrane potential and delayed evaluation. Frontiers in Neuroscience, 2021. 2, 5", + "[30] Eugene M Izhikevich. Which model to use for cortical spiking neurons? IEEE Transactions on Neural Networks, 2004. 3", + "[31] Haiyan Jiang, Srinivas Anumasa, Giulia De Masi, Huan Xiong, and Bin Gu. A unified optimization framework of ann-snn conversion: towards optimal mapping from activation values to firing rates. In International Conference on Machine Learning, 2023. 2", + "[32] Haiyan Jiang, Vincent Zoonekynd, Giulia De Masi, Bin Gu, and Huan Xiong. TAB: Temporal accumulated batch normalization in spiking neural networks. In International Conference on Learning Representations, 2024. 2", + "[33] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 8", + "[34] Saeed Reza Kheradpisheh, Mohammad Ganjtabesh, Simon J Thorpe, and Timothee Masquelier. STDP-based spiking deep convolutional neural networks for object recognition. Neural Networks, 2018. 2", + "[35] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-YOLO: Spiking neural network for energy-efficient object detection. In AAAI Conference on Artificial Intelligence, 2020. 3, 5, 8", + "[36] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ANN: Towards efficient, accurate spiking neural networks calibration. In International Conference on Machine Learning, 2021. 3, 5, 6", + "[37] Yuhang Li, Shikuang Deng, Xin Dong, and Shi Gu. Error-aware conversion from ann to snn via post-training parameter calibration. International Journal of Computer Vision, 2024, 2, 6, 7, 8", + "[38] Yuhang Li, Tamar Geller, Youngeun Kim, and Priyadarshini Panda. SEENN: Towards temporal spiking early exit neural networks. Advances in Neural Information Processing Systems, 2024. 2", + "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision, 2014. 8", + "[40] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In International Conference on Computer Vision, 2017. 8", + "[41] Zibin Liu, Banglei Guan, Yang Shang, Shunkun Liang, Zhenbao Yu, and Qifeng Yu. Optical flow-guided 6dof object pose tracking with an event camera. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6501-6509, 2024. 1", + "[42] Zibin Liu, Banglei Guan, Yang Shang, Qifeng Yu, and Laurent Kneip. Line-based 6-dof object pose estimation and" + ], + "bbox": [ + 89, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tracking with an event camera. IEEE Transactions on Image Processing, 2024. 1", + "[43] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Computer Vision and Pattern Recognition, 2015. 8", + "[44] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural Networks, 1997. 1", + "[45] TorchVision maintainers and contributors. Torchvision: Pytorch's computer vision library. https://github.com/pytorch/vision, 2016.6,8", + "[46] Timothée Masquelier and Simon J Thorpe. Unsupervised learning of visual features through spike timing dependent plasticity. PLoS Computational Biology, 2007. 2", + "[47] Joaquín Royo Miquel, Silvia Tolu, Frederik ET Schöller, and Roberto Galeazzi. Retinanet object detector based on analog-to-spiking neural network conversion. In International Conference on Soft Computing & Machine Intelligence, 2021. 8", + "[48] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 2019. 2", + "[49] Nicolas Perez Nieves and Dan F. M. Goodman. Sparse spiking gradient descent. In Advances in Neural Information Processing Systems, 2021. 1", + "[50] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjic chip architecture. Nature, 2019. 1", + "[51] Nitin Rathi and Kaushik Roy. DIET-SNN: A low-latency spiking neural network with direct input encoding and leakage and threshold optimization. IEEE Transactions on Neural Networks and Learning Systems, 2021. 2", + "[52] Nitin Rathi, Gopalakrishnan Srinivasan, Priyadarshini Panda, and Kaushik Roy. Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation. In International Conference on Learning Representations, 2020. 2", + "[53] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, Michael Pfeiffer, and Shih-Chii Liu. Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in Neuroscience, 2017. 3", + "[54] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, Michael Pfeiffer, and Shih-Chii Liu. Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in Neuroscience, 2017. 2, 6", + "[55] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: VGG and residual architectures. Frontiers in Neuroscience, 2019. 2, 3", + "[56] Jiangrong Shen, Qi Xu, Jian K. Liu, Yueming Wang, Gang Pan, and Huajin Tang. ESL-SNNs: An evolutionary structure learning strategy for spiking neural networks. In AAAI Conference on Artificial Intelligence, 2023. 1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24396", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Xinyu Shi, Zecheng Hao, and Zhaofei Yu. Spikingresformer: Bridging resnet and vision transformer in spiking neural networks. In Computer Vision and Pattern Recognition, 2024. 1", + "[58] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in Neural Information Processing Systems, 2018. 2", + "[59] Sen Song, Kenneth D Miller, and Larry F Abbott. Competitive hebbian learning through spike-timing-dependent synaptic plasticity. Nature Neuroscience, 3(9):919-926, 2000. 2", + "[60] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: fully convolutional one-stage object detection. In International Conference on Computer Vision, 2019. 8", + "[61] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. CoRR, abs/1711.11248, 2017. 8", + "[62] John J Wade, Liam J McDaid, Jose A Santos, and Heather M Sayers. Swat: A spiking neural network training algorithm for classification problems. IEEE Transactions on Neural Networks, 2010. 2", + "[63] Xiaofeng Wu, Velibor Bojkovic, Bin Gu, Kun Suo, and Kai Zou. Ftbc: Forward temporal bias correction for optimizing ann-snn conversion. In European Conference on Computer Vision, pages 155-173. Springer, 2024. 2", + "[64] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in Neuroscience, 2018. 2", + "[65] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Lup ing Shi. Direct training for spiking neural networks: Faster, larger, better. In AAAI Conference on Artificial Intelligence, 2019.", + "[66] Qu Yang, Jibin Wu, Malu Zhang, Yansong Chua, Xinchao Wang, and Haizhou Li. Training spiking neural networks with local tandem learning. Advances in Neural Information Processing Systems, 2022. 2, 4", + "[67] Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, Xu Bo, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In International Conference on Learning Representations, 2024. 1", + "[68] Man Yao, Ole Richter, Guangshe Zhao, Ning Qiao, Yannan Xing, Dingheng Wang, Tianxiang Hu, Wei Fang, Tugba Demirci, Michele De Marchi, et al. Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip. Nature Communications, 2024. 1", + "[69] Friedemann Zenke and Emre O Neftci. Brain-inspired learning on neuromorphic substrates. Proceedings of the IEEE, 2021. 1", + "[70] Rui-Jie Zhu, Qihang Zhao, Guoqi Li, and Jason K Eshraghian. Spikegpt: Generative pre-trained language model with spiking neural networks. arXiv preprint arXiv:2302.13939, 2023.1" + ], + "bbox": [ + 91, + 92, + 485, + 866 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24397", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/6b151e98-3001-4e0f-b77a-ea1941036ad2_model.json b/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/6b151e98-3001-4e0f-b77a-ea1941036ad2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bcacaa6428b294fcd841464f366224908920ee77 --- /dev/null +++ b/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/6b151e98-3001-4e0f-b77a-ea1941036ad2_model.json @@ -0,0 +1,2598 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.13, + 0.904, + 0.177 + ], + "angle": 0, + "content": "Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.19, + 0.66, + 0.209 + ], + "angle": 0, + "content": "Tong \\(\\mathsf{Bu}^{1,3\\dagger}\\), Maohua \\(\\mathsf{Li}^{2\\dagger}\\), Zhaofei \\(\\mathsf{Yu}^{1,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.211, + 0.633, + 0.227 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contribution, \\* Corresponding author" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.23, + 0.758, + 0.244 + ], + "angle": 0, + "content": "putong30@pku.edu.cn, maohuali@hhu.edu.cn, yuzf12@pku.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.294, + 0.327, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.327, + 0.486, + 0.735 + ], + "angle": 0, + "content": "Spiking Neural Networks (SNNs) have emerged as a promising substitute for Artificial Neural Networks (ANNs) due to their advantages of fast inference and low power consumption. However, the lack of efficient training algorithms has hindered their widespread adoption. Even efficient ANN-SNN conversion methods necessitate quantized training of ANNs to enhance the effectiveness of the conversion, incurring additional training costs. To address these challenges, we propose an efficient ANN-SNN conversion framework with only inference scale complexity. The conversion framework includes a local threshold balancing algorithm, which enables efficient calculation of the optimal thresholds and fine-grained adjustment of the threshold value by channel-wise scaling. We also introduce an effective delayed evaluation strategy to mitigate the influence of the spike propagation delays. We demonstrate the scalability of our framework in typical computer vision tasks: image classification, semantic segmentation, object detection, and video classification. Our algorithm outperforms existing methods, highlighting its practical applicability and efficiency. Moreover, we have evaluated the energy consumption of the converted SNNs, demonstrating their superior low-power advantage compared to conventional ANNs. This approach simplifies the deployment of SNNs by leveraging open-source pre-trained ANN models, enabling fast, low-power inference with negligible performance reduction. Code is available at https://github.com/putshua/Inference-scale-ANN-SNN." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.765, + 0.222, + 0.78 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.486, + 0.852 + ], + "angle": 0, + "content": "Recent advancements in large models have significantly reshaped the landscape of deep learning technology, industry, and the research community. These advanced models, characterized by their massive scale and unprecedented zero" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.279, + 0.909, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.417, + 0.909, + 0.501 + ], + "angle": 0, + "content": "Figure 1. Compared to existing frameworks that require retraining a quantized ANN, the proposed framework is able to directly convert a pre-trained ANN to an SNN at inference-scale complexity, significantly reducing computational requirements, minimizing dependence on GPUs, and requiring only a small subset of the original dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.598 + ], + "angle": 0, + "content": "shot generalizability in downstream tasks, greatly impact our daily lives. Nevertheless, the pursuit of larger models raises concerns about high energy consumption for model inference and training. The deployment of large models on resource-constrained devices has also become a challenge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.909, + 0.765 + ], + "angle": 0, + "content": "Spiking neural networks (SNNs), a well-established type of neural network that mimic the function of biological neurons and are considered the third generation of ANNs [44], offer a potential solution to these issues as an alternative to Artificial Neural Networks (ANNs). Unlike ANNs, spiking neurons encode information as discrete events (binary spikes or action potentials) and generate outputs based on both their inputs and internal state. This unique processing mechanism enables SNNs to effectively handle spatio-temporal information while leveraging sparse representations, leading to more efficient and biologically plausible information processing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "With the development of the neuromorphic computing hardware [10, 11, 17, 41, 42, 49, 50, 68, 69], SNNs can now be deployed on neuromorphic chips and further applied in power-limited scenarios [9, 56]. However, the lack of efficient training algorithms has hindered their widespread application. Recent learning methods have made significant progress in training deep convolutional SNNs [15, 26] and large spiking transformers using supervised learning algorithms [57, 67, 70]. While supervised training enables SNNs" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.863, + 0.405, + 0.876 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute for Artificial Intelligence, Peking University" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.876, + 0.478, + 0.888 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Artificial Intelligence and Automation, Hohai University" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.379, + 0.9 + ], + "angle": 0, + "content": "\\(^{3}\\) School of Computer Science, Peking University" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.863, + 0.478, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24387" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.334 + ], + "angle": 0, + "content": "to achieve performance comparable to ANNs with the same architecture and reduces inference time, the memory and time costs for training scale linearly with the number of inference time-steps, making it impractical for training large energy-efficient models. A more feasible approach is ANN-to-SNN conversion [6], which transforms pre-trained ANNs into SNNs with minimal computational overhead. However, converted SNNs typically require more inference time-steps to match the performance of their ANN counterparts, leading to increased latency and higher energy consumption. A potential solution is to re-train a modified ANN before conversion, which can significantly enhance SNN performance while reducing inference time-steps [4, 29]. Although retraining-based conversion methods generally improve performance across most tasks, the additional computational cost of re-training poses a notable overhead." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.339, + 0.487, + 0.73 + ], + "angle": 0, + "content": "In this paper, we propose an efficient post-training ANN-SNN conversion algorithm that directly obtains high-performance SNNs from pre-trained ANNs with inference-scale complexity. Figure 1 illustrates the key difference between existing ANN-to-SNN conversion frameworks and our proposed conversion framework. Compared to training-required methods, our approach greatly reduces the need for GPU resources, requires fewer data samples, and shortens the overall conversion process. We employ a local-learning-rule based algorithm for rapid conversion from ANNs to SNNs, avoiding the requirements for computationally expensive backpropagation-based training or fine-tuning. Additionally, we introduce a delayed evaluation strategy to further enhance performance. This conversion framework simplifies the deployment of SNNs by leveraging open-source, pre-trained ANN models, allowing direct application of converted SNNs on neuromorphic hardware with only inference-level computational costs. Overall, the proposed conversion framework offers two potential advantages. First, the lightweight local learning algorithm reduces dependence on GPUs, enabling online mapping from ANN to SNN on neuromorphic chips. Second, this approach reduces the need for extensive retraining of models, making it highly suitable for the practical deployment of large models in resource-constrained environments. Our contribution can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.735, + 0.483, + 0.809 + ], + "angle": 0, + "content": "- We introduce an inference-scale complexity ANN-SNN conversion framework, enabling the conversion of pretrained ANN models into high-performance SNNs while maintaining computational costs at inference level or even lower." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.9 + ], + "angle": 0, + "content": "- We propose an efficient ANN-SNN conversion framework, theoretically deriving an upper bound for the conversion error between the original ANN and the converted SNN. To minimize this error bound, we introduce a light-weight local learning method for threshold optimization. Additionally, we design a delayed evaluation strategy to mini" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.735, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.835, + 0.107 + ], + "angle": 0, + "content": "gate the influence of spike propagation delays." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.908, + 0.228 + ], + "angle": 0, + "content": "- We experimentally demonstrate that our method is a lightweight, plug-and-play solution capable of integrating with various conversion methods for further performance improvement. Our framework successfully scales to three typical computer vision tasks, outperforming existing ANN-SNN conversion methods in classification tasks while also demonstrating its practical applicability in handling regression-based vision tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.241, + 0.665, + 0.258 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.267, + 0.909, + 0.568 + ], + "angle": 0, + "content": "Researchers have endeavored to explore effective SNN training approaches for decades. Early studies mainly focused on the unsupervised Spike-Timing-Dependent Plasticity (STDP) algorithm [7, 25]. These studies aimed to utilize the STDP algorithm for shallow networks and feature extractors [13, 34, 46, 59, 62]. In parallel, supervised training approaches have been developed to achieve high-performance SNNs. Bohte et al. [1] were pioneers in applying the backpropagation algorithm to train SNNs, introducing a temporal-based backpropagation method that uses timestamps as intermediate variables during gradient computation. Wu et al. [64] treated SNNs as recurrently connected networks and utilized Back Propagation Through Time (BPTT) for training, termed spatial-temporal backpropagation (STBP). Unlike these methods, Cao et al. [6] proposed an ANN-SNN conversion method that achieves high-performance SNNs from pre-trained ANNs. Currently, research efforts are primarily focused on enhancing the efficiency of training algorithms for high-performance SNNs, leveraging the three methods mentioned above." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.57, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Based on whether global backpropagation is required, SNN learning methods can be ascribed into post-training and training-involved approaches. Training-required methods, including direct training of SNNs [17, 18, 32, 48, 58, 64-66], re-training before ANN-SNN conversion [5, 12, 23, 27, 31, 38, 63] and fine-tuning of converted SNNs [2, 37, 51, 52], typically require the participation of the back-propagation algorithms, making them more computationally intensive and resource-demanding. Post-training learning methods mainly refer to conversion-based algorithms that directly convert the pre-trained ANN into SNN. Cao et al. [6] pioneered this approach by mapped the weights of a lightweight CNN network to an SNN, demonstrating significant energy efficiency improvements through hardware analysis. Diehl et al. [14] proposed a weight and threshold balancing method to mitigate performance degradation, while Hunsberger and Eliasmith [28] further extended conversion-based learning to LIF neurons. Rueckauer et al. [54] established a theoretical framework for conversion-based algorithms, leading to further optimizations such as robust normalization, an enhanced threshold balancing method designed to optimize the trade-off between inference latency and accuracy [55]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24388" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.332 + ], + "angle": 0, + "content": "Han and Roy [20] extended ANN-to-SNN conversion to temporal coding SNNs, improving accuracy and introducing a reset-by-subtraction mechanism instead of the traditional reset-to-zero, thereby preventing information loss [21]. Beyond classification tasks, conversion-based methods have been successfully applied to regression problems such as object detection, with Kim et al. [35] demonstrating the feasibility of spike-based object detection models. Further optimizations include adding an initial membrane potential for improved accuracy and reduced latency [4, 22], as well as employing a lightweight conversion pipeline to optimize thresholds and introduce a bias term at each inference step [36]. Although performance is limited, those lightweight training methods are able to acquire SNNs from pre-trained ANNs with negligible cost of multiple iterations of inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.348, + 0.232, + 0.363 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.373, + 0.248, + 0.387 + ], + "angle": 0, + "content": "3.1. Neuron Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.484, + 0.455 + ], + "angle": 0, + "content": "The core components of conventional ANNs are point neurons, where their forward propagation can be represented as a combination of linear transformations and non-linear activations, that is" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.466, + 0.484, + 0.486 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} ^ {l} = A \\left(\\boldsymbol {w} ^ {l} \\boldsymbol {a} ^ {l - 1}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.484, + 0.542 + ], + "angle": 0, + "content": "Here, \\( \\boldsymbol{w}^l \\) is the weights between layer \\( l - 1 \\) and layer \\( l \\), \\( \\boldsymbol{a}^l \\) is the activation vector in layer \\( l \\), and \\( A(\\cdot) \\) denotes the non-linear activation function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.543, + 0.485, + 0.618 + ], + "angle": 0, + "content": "The neurons in SNNs are spiking neurons with temporal structures. Like most ANN-SNN conversion methods [3, 5, 6, 12, 14, 55], we employ the Integrate-and-Fire (IF) neuron [19, 30] in this paper. For computational convenience, we use the following discrete neuron function:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.629, + 0.484, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} ^ {l} \\left(t ^ {-}\\right) = \\boldsymbol {v} ^ {l} (t - 1) + \\boldsymbol {w} ^ {l} \\boldsymbol {s} ^ {l - 1} (t) \\theta^ {l - 1}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.65, + 0.484, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} ^ {l} (t) = H \\left(\\boldsymbol {v} ^ {l} \\left(t ^ {-}\\right) - \\theta^ {l}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.671, + 0.484, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} ^ {l} (t) = \\boldsymbol {v} ^ {l} \\left(t ^ {-}\\right) - \\theta^ {l} \\boldsymbol {s} ^ {l} (t). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.486, + 0.839 + ], + "angle": 0, + "content": "Here \\( s^l(t) \\) describes whether neurons in layer \\( l \\) fire at timestep \\( t \\), with firing triggered when the membrane potential before firing \\( v^l(t^-) \\) reaches the firing threshold \\( \\theta^l \\), as represented by the Heaviside step function \\( H(\\cdot) \\). To alleviate information loss, we adopt the \"reset-by-subtraction\" mechanism [21, 53] for updating the membrane potential \\( v^l(t) \\) after spike emission. Specifically, instead of resetting to the resting potential, the membrane potential is reduced by \\( \\theta^l \\) after emitting a spike." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.303, + 0.863 + ], + "angle": 0, + "content": "3.2.ANN-SNN Conversion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The pivotal idea of ANN-SNN conversion is to map the activation of analog neurons to the postsynaptic potential" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "(or firing rate) of spiking neurons. Specifically, by defining the initial membrane potential as \\( \\mathbf{v}^l(0) \\) and accumulating the neuron function (Equations 2 and 4) from 1 to T and dividing both sides by time-step T, we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.16, + 0.907, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\frac {\\boldsymbol {v} ^ {l} (T) - \\boldsymbol {v} ^ {l} (0)}{T} = \\frac {\\sum_ {t = 1} ^ {T} \\boldsymbol {w} ^ {l} \\boldsymbol {s} ^ {l - 1} (t) \\theta^ {l - 1} - \\sum_ {t = 1} ^ {T} \\boldsymbol {s} ^ {l} (t) \\theta^ {l}}{T}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.218, + 0.908, + 0.265 + ], + "angle": 0, + "content": "This equation reveals a linear relationship between the average firing rate of neurons in adjacent layers by defining average postsynaptic potential as \\( r^l(t) = \\sum_{i=1}^{t} s^l(i) \\theta^l / t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.273, + 0.907, + 0.305 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {r} ^ {l} (T) = \\boldsymbol {w} ^ {l} \\boldsymbol {r} ^ {l - 1} (T) - \\frac {\\boldsymbol {v} ^ {l} (T) - \\boldsymbol {v} ^ {l} (0)}{T}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.909, + 0.525 + ], + "angle": 0, + "content": "As \\( T \\to +\\infty \\), the conversion error can generally be assumed to approach zero. Therefore, using the equations presented, a trained ANN model can be converted into an SNN by replacing ReLU neurons with IF neurons, which forms the fundamental principle of ANN-SNN conversion. Since Equations 1 and 6 are not mathematically identical, some conversion error typically remains. To mitigate these errors, threshold balancing [14] and weight scaling algorithms [36, 55] play a crucial role. Both techniques aim to adjust synaptic weights or neuron thresholds according to the distribution range of neuron inputs, effectively reducing clipping errors. Previous studies have shown that weight scaling and threshold balancing are equivalent in effect and can achieve high-performance SNNs [4]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.536, + 0.613, + 0.551 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.909, + 0.698 + ], + "angle": 0, + "content": "In this section, we present our post-training ANN-SNN conversion framework. We begin by defining the conversion error and deriving its upper bound. After that, we propose a local threshold balancing method based on a local learning rule. Additionally, we introduce the channel-wise threshold balancing technique, the pre-neuron max pooling layer, and the delayed evaluation strategy. These techniques are seamlessly integrated into a comprehensive conversion framework to enhance both performance and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.706, + 0.743, + 0.721 + ], + "angle": 0, + "content": "4.1. Conversion Error Bound" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.908, + 0.789 + ], + "angle": 0, + "content": "We follow the conversion error representation from [5] and define the conversion error \\( e^l \\) in layer \\( l \\) as the 2-norm of the difference between the postsynaptic potential of neurons in SNNs and the activation output of neurons in ANNs, that is" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.796, + 0.907, + 0.815 + ], + "angle": 0, + "content": "\\[\ne ^ {l} = \\left\\| \\mathrm {S} \\left(\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}\\right) - \\mathrm {A} \\left(\\boldsymbol {z} ^ {l}\\right) \\right\\| _ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Here, \\(\\hat{z}^l = \\pmb{w}^l\\pmb{r}^{l - 1}\\) denotes the average input current from layer \\(l - 1\\) in SNNs, and \\(\\pmb{z}^l = \\pmb{w}^l\\pmb{a}^{l - 1}\\) represents the activation input from layer \\(l - 1\\) in ANNs. Both the ANN and SNN models share the same weights, denoted as \\(\\pmb{w}^l\\). \\(\\mathrm{A}(z^l)\\) represents the output of nonlinear ReLU activation function while" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24389" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.483, + 0.121 + ], + "angle": 0, + "content": "\\(\\mathrm{S}(\\hat{z}^l;\\theta^l) = \\boldsymbol {r}^l (T)\\) represents the average output synaptic potential at time-step \\(T\\) with given average input \\(\\hat{z}^l\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.318 + ], + "angle": 0, + "content": "Since our primary objective is to minimize the error of the final outputs, we define the conversion error between the ANN and SNN models as the layer-wise conversion error in the last layer \\( L \\), denoted as \\( e_{\\mathrm{model}} = e^{L} \\). A straightforward approach would be to directly use the conversion error between models as a loss function and optimize it, which can be viewed as a variant of knowledge distillation [66]. However, this method is computationally expensive in terms of both time and memory, resembling the challenges of supervised training of SNN. To address this, we scale the conversion error and derive an error upper bound, thereby simplifying the optimization process. We formalize this approach in the following theorem." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.33, + 0.485, + 0.36 + ], + "angle": 0, + "content": "Theorem 1 The layer-wise conversion error can be divided into intra-layer and inter-layer errors:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.37, + 0.484, + 0.407 + ], + "angle": 0, + "content": "\\[\ne ^ {l} \\leqslant \\overbrace {\\| \\mathbf {S} (\\hat {\\boldsymbol {z}} ^ {l} ; \\theta^ {l}) - \\mathrm {A} (\\hat {\\boldsymbol {z}} ^ {l}) \\| _ {2}} ^ {\\text {i n t r a - l a y e r e r r o r}} + \\overbrace {\\| \\boldsymbol {w} ^ {l} \\| _ {2} e ^ {l - 1}} ^ {\\text {i n t e r - l a y e r e r r o r}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.418, + 0.484, + 0.479 + ], + "angle": 0, + "content": "Given that both ANN and SNN models receive the same input in the first layer, leading to \\( e^0 = 0 \\), the upper bound for the conversion error between ANN and SNN models in an \\( L \\)-layer fully-connected network is given by" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.488, + 0.484, + 0.545 + ], + "angle": 0, + "content": "\\[\ne _ {\\text {m o d e l}} = e ^ {L} \\leqslant \\sum_ {l = 1} ^ {L} \\left(\\prod_ {k = l + 1} ^ {L} \\| \\boldsymbol {w} ^ {k} \\| _ {2}\\right) \\left\\| \\mathrm {S} \\left(\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}\\right) - \\mathrm {A} \\left(\\hat {\\boldsymbol {z}} ^ {l}\\right) \\right\\| _ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.484, + 0.649 + ], + "angle": 0, + "content": "Here \\( \\| \\pmb{w}^k \\|_2 \\) is the matrix norm \\( (p = 2) \\) or spectral norm of the weight matrix. The detailed proof is provided in the Appendix. Theorem 1 indicates that the conversion error is bounded by the weighted sum of errors across all layers. Based on this result, we introduce the local threshold balancing algorithm in the following section." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.659, + 0.42, + 0.675 + ], + "angle": 0, + "content": "4.2. Local Threshold Balancing Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.486, + 0.739 + ], + "angle": 0, + "content": "The target of ANN-SNN conversion is to minimize the conversion error and achieve high-performance SNNs. An alternative approach is to optimize the error bound, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.749, + 0.484, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname * {a r g m i n} _ {\\boldsymbol {\\theta}} \\mathbb {E} _ {x ^ {0} \\in \\mathcal {D}} \\\\ \\left[ \\sum_ {l = 1} ^ {L} \\left(\\prod_ {k = l + 1} ^ {L} \\| \\boldsymbol {w} ^ {k} \\| _ {2}\\right) \\left\\| \\mathrm {S} \\left(\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}\\right) - \\mathrm {A} \\left(\\hat {\\boldsymbol {z}} ^ {l}\\right) \\right\\| _ {2} \\right]. \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Here \\(\\theta\\) represents the set of threshold values to be optimized across all layers, i.e., \\(\\pmb{\\theta} = \\{\\theta^1,\\theta^2,\\dots,\\theta^L\\}\\). \\(x^{0}\\) denotes the input data sample drawn from the dataset \\(\\mathcal{D}\\). This approach leverages data-driven conversion, minimizing the expectation of the conversion error bound over the data distribution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.908, + 0.138 + ], + "angle": 0, + "content": "To reduce computational costs during optimization, we introduce two key simplifications. Firstly, we use a greedy strategy to optimize the threshold \\(\\theta\\) layer by layer, that is" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.146, + 0.748, + 0.168 + ], + "angle": 0, + "content": "For each layer \\(l\\), \\(\\arg \\min_{\\theta^l} \\mathbb{E}_{x^0 \\in \\mathcal{D}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.171, + 0.907, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\left[ \\left(\\prod_ {k = l + 1} ^ {L} \\| \\boldsymbol {w} ^ {k} \\| _ {2}\\right) \\left\\| \\mathbf {S} \\left(\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}\\right) - \\mathbf {A} \\left(\\hat {\\boldsymbol {z}} ^ {l}\\right) \\right\\| _ {2} \\right]. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.222, + 0.907, + 0.358 + ], + "angle": 0, + "content": "Secondly, we simplify the function \\( \\mathrm{S}(\\cdot; \\theta) \\) by neglecting its temporal dynamics and approximating it with the clipping function \\( \\mathrm{C}(\\cdot; \\theta) \\), as previous works have demonstrated that spiking neuron and clipping functions behave equivalently given sufficient number of time-steps [12]. The average input in the SNN can therefore be estimated as \\( \\hat{\\boldsymbol{z}}^0 = \\boldsymbol{z}^0 \\); \\( \\hat{\\boldsymbol{z}}^{l+1} = \\boldsymbol{w}^{l+1} \\mathrm{C}(\\hat{\\boldsymbol{z}}^l; \\theta^l) \\). Furthermore, We utilize the squared norm in the objective function for smoother optimization. The optimization problem then becomes:" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.367, + 0.747, + 0.389 + ], + "angle": 0, + "content": "For each layer \\(l\\), \\(\\arg \\min_{\\theta^l} \\mathbb{E}_{x^0 \\in \\mathcal{D}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.392, + 0.907, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\left[ \\left(\\prod_ {k = l + 1} ^ {L} \\| \\boldsymbol {w} ^ {k} \\| _ {2}\\right) \\| \\mathbf {C} (\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}) - \\mathbf {A} (\\hat {\\boldsymbol {z}} ^ {l}) \\| _ {2} ^ {2} \\right], \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.443, + 0.907, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\operatorname {w h e r e} \\mathrm {C} (x; \\theta) = \\min (\\max (0, x), \\theta). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.906, + 0.555 + ], + "angle": 0, + "content": "Intuitively, we can employ the stochastic gradient descent algorithm to optimize the threshold. Since \\(\\prod_{k = l + 1}^{L}\\| \\pmb{w}^{k}\\|_{2}\\) can be considered as a constant when weights are fixed for each layer, this term can be absorbed into the learning rate. The final update rule for the local threshold balancing algorithm at each step is:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.564, + 0.907, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\Delta \\theta^ {l} = - \\sum_ {i = 0} ^ {N - 1} 2 \\left(\\hat {z} _ {i} ^ {l} - \\theta^ {l}\\right) H \\left(\\hat {z} _ {i} ^ {l} - \\theta^ {l}\\right), \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.607, + 0.907, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\theta^ {l} \\leftarrow \\theta^ {l} - \\eta \\Delta \\theta^ {l}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.635, + 0.908, + 0.711 + ], + "angle": 0, + "content": "Detailed derivation is provided in the Appendix. Here \\(\\Delta \\theta^l\\) denotes the step size during optimization and \\(\\eta\\) is the learning rate or update step size. \\(N\\) is the number of elements in the input vector, and \\(\\hat{z}_i\\) denotes each element in the vector \\(\\hat{z}^l\\). \\(H(\\cdot)\\) is the Heaviside step function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.909, + 0.906 + ], + "angle": 0, + "content": "In practice, as shown in Figure 2, we jointly optimize the threshold for each layer by sampling a small set of image samples from the training dataset. Before starting the conversion process, we first replace all ReLU activation functions \\(\\mathrm{A}(\\cdot)\\) with clipping functions \\(\\mathrm{C}(\\cdot)\\) and initialize the threshold to zero. We then randomly sample data batches from the training dataset and perform model inference with these data samples. The thresholds are locally updated during the inference until convergence. Notably, the parameter \\(\\theta^l\\) is updated locally, ensuring that the computational cost of each iteration is similar to the inference of the ANN model and the computational cost of the whole conversion process is at inference level." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24390" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.473, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.209, + 0.485, + 0.268 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed conversion framework. It requires only a small subset of data for conversion using local threshold balancing. During inference, the delayed evaluation technique is employed to enhance the accuracy of output estimation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.289, + 0.37, + 0.306 + ], + "angle": 0, + "content": "4.3. Pre-Neuron Max Pooling Layer" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.311, + 0.486, + 0.522 + ], + "angle": 0, + "content": "The conversion of the max pooling layer presents a significant challenge. As a pivotal feature in the process of down-sampling input representations, the max pooling layer is a fundamental component in convolutional neural networks and is widely used in most convolutional architectures. However, due to the binary nature of spiking neurons, max pooling layers cannot effectively perform downsampling in feature maps. Typically, in a single max-pooling operation, multiple elements may share the same value, preventing the extraction of the most salient features. Consequently, max pooling is often avoided as a downsampling layer in SNNs. In most previous ANN-SNN methods, the max pooling layer in ANN is often replaced by an average pooling layer, and the model is re-trained or fine-tuned before conversion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.485, + 0.613 + ], + "angle": 0, + "content": "To convert architectures that contain max pooling layers, we propose a simple yet effective method. We replace all max-pooling layers with neuron layers, allowing the preneuron floating-point input to pass through the downsampling layer before being input into the neurons. We have the following theorem:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.485, + 0.652 + ], + "angle": 0, + "content": "Theorem 2 The order of the max pooling layer and ReLU layer does not affect the ANN output results." + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.66, + 0.484, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\max _ {i} \\mathrm {R} (z) = \\mathrm {R} (\\max _ {i} z), \\text {w h e n} \\max (\\mathrm {z}) > 0. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.795 + ], + "angle": 0, + "content": "Here \\( \\mathrm{R}(\\cdot) \\) denotes the ReLU activation. The detailed proof is provided in the Appendix. Theorem 2 guarantees that such an operation does not affect the final performance. When deployed on neuromorphic chips or FPGAs, the max pooling operation can be efficiently implemented through a comparator, avoiding additional floating-point multiplications and minimizing power consumption." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.804, + 0.398, + 0.82 + ], + "angle": 0, + "content": "4.4. Channel-wise Threshold Balancing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.871 + ], + "angle": 0, + "content": "During conversion, following Kim et al. [35], Li et al. [36], we employ a channel-wise local threshold balancing method applicable to both fully connected and convolutional layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this approach, neurons in each channel share a single threshold for convolutional layers while each neuron has an" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.168 + ], + "angle": 0, + "content": "independent threshold for fully connected layers. Specifically, for convolutional layers, we channel-wisely determine the threshold value, while for fully connected layers, we optimize the threshold element-wise. The update function is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.178, + 0.907, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\Delta \\theta_ {c} ^ {l} = - \\sum_ {i = 0} ^ {N - 1} 2 \\left(\\hat {z} _ {c, i} ^ {l} - \\theta_ {c} ^ {l}\\right) H \\left(\\hat {z} _ {c, i} ^ {l} - \\theta_ {c} ^ {l}\\right), \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.905, + 0.273 + ], + "angle": 0, + "content": "where \\( c \\) denotes the channel index (or element index for fully connected layers) and \\( i \\) denotes the element index within the current channel." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.396 + ], + "angle": 0, + "content": "After optimization, the thresholds for each channel can be individually absorbed into the corresponding channel of the convolutional kernel or the weight vectors of the fully connected layers. This channel-wise (or element-wise) operation ensures that the behavior of the IF neuron remains unaffected. This algorithm aims to more precisely determine the optimal threshold value while preserving the fundamental properties of spiking neurons." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.404, + 0.771, + 0.421 + ], + "angle": 0, + "content": "4.5. Delayed Evaluation Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.908, + 0.547 + ], + "angle": 0, + "content": "As introduced in [4], the first spike from the last layer often takes longer than expected to appear due to propagation delays between layers. This delay creates significant fluctuations in the firing rate during the initial time-steps, as no spikes are generated until the first one reaches each layer [29]. To address this issue, we propose an effective delayed step selection strategy building upon the delayed evaluation technique [29]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.909, + 0.652 + ], + "angle": 0, + "content": "As shown in Figure 2, the basic idea of the delayed evaluation is to compute the average spike count in the last layer from time-step \\( t_0 \\) to \\( T \\), where \\( t_0 \\) denotes the delayed step, rather than averaging over all time-steps. Since the output is collected from the average spike count in the interval \\( [t_0, T] \\) at the last layer, we can rewrite Equation 6 for the last layer as" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.66, + 0.907, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {r} ^ {L} \\left(t _ {0}, T\\right) = \\boldsymbol {w} ^ {L} \\boldsymbol {r} ^ {L - 1} \\left(t _ {0}, T\\right) - \\frac {\\boldsymbol {v} ^ {L} (T) - \\boldsymbol {v} ^ {L} \\left(t _ {0}\\right)}{T - t _ {0}}, \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.908, + 0.781 + ], + "angle": 0, + "content": "where \\( \\boldsymbol{r}^l(t_0, t) = \\sum_{i=t_0}^{t} \\boldsymbol{s}^l(i) \\theta^l / (t - t_0) \\) denotes the average postsynaptic potential from time \\( t_0 \\) to \\( t \\), and \\( L \\) represents the last layer index. When \\( T \\gg t_0 \\), one can still guarantee that the basic equivalence of the average postsynaptic potential and ANN activations holds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.782, + 0.908, + 0.841 + ], + "angle": 0, + "content": "In our strategy, we do not start counting the average output spike until \\(\\widetilde{t_0}\\), where \\(\\widetilde{t_0}\\) is the expected spike timing of the first spike in the last layer. We estimate \\(\\widetilde{t_0}\\) based on the converted SNN model, denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.851, + 0.907, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\widetilde {t _ {0}} = \\sum_ {l = 1} ^ {L} \\frac {\\theta^ {l} - v ^ {l} (0)}{\\max _ {i} \\left(\\overline {{\\mathbf {R} (\\hat {\\boldsymbol {z}} ^ {l})}}\\right)} \\approx \\sum_ {l = 1} ^ {L} \\frac {\\theta^ {l} - v ^ {l} (0)}{\\max _ {i} \\left(\\overline {{\\mathbf {R} (\\boldsymbol {z} ^ {l})}}\\right)}. \\tag {19}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24391" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.272 + ], + "angle": 0, + "content": "Here \\(\\max_i(\\cdot)\\) selects the maximum element of a given vector. The notation \\(\\overline{(\\cdot)} = \\mathbb{E}_{x^0\\in \\mathcal{D}}[(\\cdot)]\\) represents the expectation of the given input over dataset. Additionally, we use the ReLU function \\(\\mathbf{R}(\\cdot)\\) to prevent negative estimated time. Before evaluation, we first calculate \\(\\widetilde{t_0}\\) according to Equation 19 and consider it as an indicator for the delayed evaluation strategy. In practice, if the inference step \\(T\\) is less than \\(\\widetilde{t_0} +4\\), we set \\(t_0 = T - 4\\) and collect the average output over the time interval \\([T - 4,T]\\). If \\(T\\) exceeds \\(\\widetilde{t_0} +4\\), we use the interval \\([\\widetilde{t_0},T]\\) for output estimation. The value 4 is empirically chosen to ensure a sufficient delay while allowing adequate time for averaging spikes over time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.335 + ], + "angle": 0, + "content": "This technique introduces minimal additional training cost and does not alter the model's inference process, making it seamlessly integrated into the proposed conversion framework without disrupting the overall pipeline." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.346, + 0.223, + 0.363 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.372, + 0.485, + 0.568 + ], + "angle": 0, + "content": "We conduct extensive experiments to demonstrate the effectiveness of our method and highlight its potential practical value. We first showcase the advantages of our approach by integrating it with existing conversion frameworks, comparing it with state-of-the-art post-training conversion algorithms, and conducting ablation experiments to validate the effectiveness of the whole conversion pipeline. Subsequently, we apply our algorithm to various visual tasks, including image classification, semantic segmentation, object detection, and video classification. The feasibility of our algorithm is highlighted on both classification and regression tasks, showcasing its generalizability across different datasets and its superiority over existing conversion algorithms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.691 + ], + "angle": 0, + "content": "To underscore that our conversion method does not require additional training, we utilize open-source pre-trained ANN models for conversion, with most pre-trained sourced from TorchVision [45]. Additionally, we incorporate the membrane potential initialization algorithm [5], setting the membrane potential to half of the threshold before inference. The pseudocode for our conversion pipeline and detailed training settings are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.698, + 0.483, + 0.729 + ], + "angle": 0, + "content": "5.1. Plug-and-Play Integration for Enhanced Performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we aim to highlight that our proposed conversion method provides a lightweight, plug-and-play solution that can be seamlessly integrated with other conversion algorithms to enhance performance. We select two representative ANN-SNN conversion approaches from existing work: the post-training fine-tuning (calibration) conversion method LCP/ACP [36, 37], and the training-required conversion method QCFS [5]. We reproduce the conversion pipeline of both methods, incorporate a local threshold balancing method for threshold searching, and use our delayed evaluation strategy. The evaluation is conducted on the ImageNet" + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.089, + 0.881, + 0.206 + ], + "angle": 0, + "content": "
MethodANNCFInference Step
3264128256512
LCP75.66×50.2163.6668.8972.1274.17
✓53.8067.1872.1874.0774.56
ACP75.66×64.5471.1273.4574.6175.35
✓65.9572.8774.2174.7375.26
QCFS73.30×64.4371.5173.2573.6773.76
✓71.9473.0673.5073.4973.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.22, + 0.905, + 0.248 + ], + "angle": 0, + "content": "Table 1. Integration of the proposed conversion approach with existing framework" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.259, + 0.697, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.374, + 0.657, + 0.387 + ], + "angle": 0, + "content": "(a) Ablation Study" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.259, + 0.898, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.374, + 0.882, + 0.387 + ], + "angle": 0, + "content": "(b) Effect of Iteration Steps" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.393, + 0.908, + 0.435 + ], + "angle": 0, + "content": "Figure 3. (a) Results on evaluating the impact of different techniques within the conversion framework. (b) Effect of iteration steps on SNN performance after conversion." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.457, + 0.78, + 0.471 + ], + "angle": 0, + "content": "dataset with the ResNet-34 architecture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.909, + 0.64 + ], + "angle": 0, + "content": "The results, as detailed in Table 1, showcase the performance of various configurations. The column labeled \"CF\" indicates whether the proposed method is used. Notably, the composite approaches, which combine our method with either LCP/ACP or QCFS, consistently outperform the individual conversion method across most timesteps. After incorporating our methods, both QCFS and LCP exhibit performance improvements exceeding \\(3\\%\\) when \\(\\mathrm{T} = 32\\), while maintaining consistent performance as \\(T\\) increases. This demonstrates the advantages of using our approach in conjunction with existing conversion techniques." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.905, + 0.667 + ], + "angle": 0, + "content": "5.2. Ablation Study for the Conversion Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.901 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of our proposed conversion framework by using the ResNet-34 architecture on the ImageNet dataset. An ablation study is conducted to evaluate the impact of different techniques within the framework. Specifically, as shown in Figure 3a, we compare the performance of four different SNNs obtained by combinations of different techniques. The baseline SNN is obtained using the Robust-Norm [54] algorithm (RN, red curve), which is a commonly used post-training ANN-SNN conversion method that sets the 99-th percentile activation value as the threshold. The other SNNs are obtained by using the layer-wise local threshold balancing algorithm (LTB, yellow curve), the channel-wise threshold balancing algorithm (CLTB, blue curve), and a combination of channel-wise threshold balancing and delayed evaluation (CLTB+DE, black curve)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24392" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.091, + 0.473, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.201, + 0.386, + 0.216 + ], + "angle": 0, + "content": "Figure 4. Effect of delayed steps" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.241, + 0.486, + 0.468 + ], + "angle": 0, + "content": "Figure 3a presents the accuracy changes of the converted SNNs with respect to the increasing time-steps. The performance of the SNN obtained from the full conversion pipeline (black curve) consistently outperforms the other SNNs across all time-steps, demonstrating an excellent balance between fast inference and high performance of the proposed conversion framework. This is particularly evident when combined with the proposed delayed evaluation technique. Compared to the robust norm method (red curve), the peak performance of the SNNs obtained by the proposed local threshold balancing method (black, blue and yellow curves) more closely aligns with the original ANN accuracy. Furthermore, the use of the channel-wise threshold balancing (CLTB, blue curve) results in a noticeable performance improvement over the vanilla threshold balancing (LTB, yellow curve) algorithm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.477, + 0.333, + 0.493 + ], + "angle": 0, + "content": "5.3. Effect of the Iteration Step" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.485, + 0.817 + ], + "angle": 0, + "content": "We evaluate the influence of the iteration steps of the local threshold balancing algorithm. The iteration number of the local threshold balancing serves as the hyper-parameter of the method. We focus on the most straightforward approach by empirically evaluating the impact of different iteration steps. We converted five different SNNs from the pre-trained ImageNet ResNet-34 model, varying the iterations steps from 1000 to 5000. The batch size during local threshold balancing is consistently set to 100. Figure 3b shows the final accuracy of the different converted SNNs. As number of iterations increases, the peak performance of the converted SNNs at different inference steps approaches the accuracy of the original ANN. Furthermore, these SNNs demonstrate better performance at low time-steps when fewer iteration steps are used. On the ImageNet dataset, with only 1000 iterations of local threshold update, the performance gap between the SNN at 512 inference time-steps and the ANN is approximately \\(1\\%\\). With 5000 iterations, this difference is further reduced to around \\(0.2\\%\\). Notably, the accuracy surpasses \\(65\\%\\) within just 55 time-steps, achieving a balance between inference time and performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.825, + 0.327, + 0.841 + ], + "angle": 0, + "content": "5.4. Effect of the Delayed Step" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We conduct experiments to demonstrate the effectiveness of the proposed delayed strategy on the ImageNet dataset using the ResNet-34 architecture. As shown in Figure 4, we evaluate the impact of the delayed step across four different" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.091, + 0.898, + 0.293 + ], + "angle": 0, + "content": "
ArchitectureMethodANNInference Step
3264128256≥512
Post-Training
ResNet-341RMP70.64---55.6560.08
ResNet-342LCP75.6655.1667.5672.4874.5375.44
MobileNetLCP73.400.2015.4755.9566.3572.19
MobileNetV2Ours71.8819.3949.9364.3168.8970.48
ResNet-183Ours69.7665.5568.5369.0369.2169.21
ResNet-343Ours73.3164.8871.1772.3072.4272.46
Training-involved
ResNet-343QCFS73.3064.4371.5173.2573.6773.76
ResNet-342ACP75.6664.6571.3073.9475.0075.45
ResNet-343Ours*73.3071.9473.0673.5073.4973.54
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.536, + 0.294, + 0.872, + 0.305 + ], + "angle": 0, + "content": "1 Standard ResNet-34 architecture without batch normalization." + }, + { + "type": "table_footnote", + "bbox": [ + 0.536, + 0.306, + 0.895, + 0.328 + ], + "angle": 0, + "content": "2 Two more layers in the first conv-block compared to standard ResNet-34." + }, + { + "type": "table_footnote", + "bbox": [ + 0.536, + 0.33, + 0.737, + 0.341 + ], + "angle": 0, + "content": "3 Standard ResNet-18/34 architecture." + }, + { + "type": "list", + "bbox": [ + 0.536, + 0.294, + 0.895, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.351, + 0.905, + 0.379 + ], + "angle": 0, + "content": "Table 2. Comparison of the proposed method and previous works on the ImageNet dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.906, + 0.49 + ], + "angle": 0, + "content": "inference time-step settings: 32 steps (red curve), 64 steps (yellow curve), 128 steps (blue curve), and 256 steps (black curve). The grey vertical line indicates the estimated \\(\\widetilde{t_0}\\) for the currently converted SNN. For each sub-figure, the x-axis represents the number of delayed steps, while the y-axis denotes the accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Although the delayed strategy is pre-determined before inference, the estimated \\(\\widetilde{t_0}\\) can still serves as an indicator that provides a near-optimal strategy to select the delayed steps. When the total inference step is less than the estimated \\(\\widetilde{t_0}\\) (red and yellow curves), the final performance gradually improves with the increase in the delayed steps, but drops significantly when the delayed step approaches the inference step. In the situation that the total inference step exceeds \\(\\widetilde{t_0}\\) (blue and black curve), the performance first increases then fluctuates when delayed step is larger than \\(\\widetilde{t_0}\\), and ultimately experiences a sharp decline as the delayed steps get close to the inference step. The above results indicate that we can always reach near-optimal performance with the proposed delayed evaluation strategy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.803, + 0.729 + ], + "angle": 0, + "content": "5.5. Test on Image Classification task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We evaluate our conversion method on the classification task with the ImageNet dataset, employing different architectures including ResNet and MobileNet. To highlight the advantages of the proposed conversion algorithm, we conduct comparisons with previous post-training algorithms, showcasing its superior performance. When using the ResNet-34 architecture, the SNN converted by our algorithm exceeds \\(71\\%\\) accuracy within 64 time steps. In contrast, the SNNs converted by the previous algorithm [20] and the calibrationrequired conversion algorithm [37] require longer time-steps to achieve comparable accuracy. We also implement the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24393" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.242 + ], + "angle": 0, + "content": "proposed conversion framework on QCFS pre-trained models (denote as \"Our*\" in Table 2), and compare with the sota training-involved ANN-SNN techniques. The accuracy of the proposed method reaches \\(71.94\\%\\) within 32 steps, demonstrating significant advantages compared to the fine-tuning required advanced calibration technique ACP [37] and training-involved vanilla QCFS [5]. The experiments on the ImageNet dataset demonstrate the superior performance of the proposed conversion algorithm and highlight its potential for large-scale applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.484, + 0.396 + ], + "angle": 0, + "content": "In addition, we demonstrate the energy-saving advantage of our method by theoretically estimating the energy efficiency of our converted SNN. We use the same energy estimation approach as Bu et al. [5]. With ResNet-34 architecture on ImageNet, SNN converted from our algorithm can achieve energy efficiency at 622FPS/W while maintaining \\(90\\%\\) performance. In comparison, the energy efficiency of the original ANN is only 22FPS/W, which is 28 times lower than that of the converted SNN. For detailed energy consumption analysis, please refer to the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.408, + 0.483, + 0.439 + ], + "angle": 0, + "content": "5.6. Test on Semantic Segmentation, Object Detection, and Video Classification Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.483, + 0.719 + ], + "angle": 0, + "content": "We further extend the proposed ANN-SNN conversion method to semantic segmentation, object detection, and video classification tasks. For the semantic segmentation task, we evaluate our method on two commonly used datasets: Pascal VOC 2012 [16] augmented by SBD [24] and MS COCO 2017 [39]. We employ various semantic segmentation models, including FCN [43] and DeepLabV3 [8], provided by TorchVision [45] and other open-source repositories. As shown in Table 3, our method is compatible with tackling semantic segmentation tasks without any preparatory training. Specifically, we achieved a \\(53.50\\%\\) mIoU with the FCN model using a ResNet-34 backbone in 64 timesteps on the Pascal VOC dataset and a \\(61.52\\%\\) mIoU with the DeepLabV3 model using a ResNet-50 architecture in 128 timesteps on the more complex MS COCO 2017 dataset. Given that pixel-level classification tasks require precise model output, this excellent performance highlights the effectiveness and generalizability of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For object detection tasks, we use the benchmark dataset MS COCO 2017 [39] with open-source models from TorchVision [45]. We employ the fully convolutional object detection method RetinaNet [40] and FCOS [60], and convert the pre-trained ANN models to SNNs using the proposed method. Table 3 shows detailed performance under different inference time-steps. Within 64 time-steps, the converted SNN can achieve \\(27.3\\%\\) mAP with RetinaNet and \\(26.5\\%\\) mAP with FCOS, which represents a significant improvement compared to previous explorations [35, 47]. The previous method required over 1000 time-steps to achieve comparable performance as ANN, which is detrimental to" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.09, + 0.895, + 0.31 + ], + "angle": 0, + "content": "
DatasetArch.ANNInference Step
3264128256
Semantic Segmentation (mIoU%)
PascalVOCFCN-1847.3043.4947.0247.2647.23
FCN-3454.9946.6053.5054.7554.76
DeepLab-1855.7548.5953.9155.4055.60
DeepLab-3458.9541.2055.7458.3658.68
MS COCOFCN-5060.6728.6650.3857.0159.53
DeepLab-5063.0139.0256.6261.5262.61
Object detection (mAP%)
MS COCORetinaNet-5036.416.627.332.234.3
FCOS-5039.212.226.532.133.6
Video Classification (Acc%)
S-Kinetics-400MC3-1863.5255.2161.6062.4362.73
R3D-1862.5653.4660.3161.1261.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.323, + 0.905, + 0.351 + ], + "angle": 0, + "content": "Table 3. Performance on semantic segmentation (mIoU%), object detection (mAP%), and video classification (Acc%) tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.371, + 0.788, + 0.386 + ], + "angle": 0, + "content": "real-time detection and energy efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.908, + 0.537 + ], + "angle": 0, + "content": "For the video classification tasks, we use a small Kinetics-400 dataset [33], which is a subset of the original dataset. More specifically, we split the original validation set into two independent sets, one for our threshold balancing procedure, accounting for about \\(60\\%\\) of the original validation set, and the other for the test. As a result, the new train set contains 12000 videos and the new test set contains 7881 videos. The converted MC3-18 [61] SNN model achieves \\(61.60\\%\\) accuracy within 64 time-steps, showing great potential of our method in more complex computer vision tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.551, + 0.763, + 0.566 + ], + "angle": 0, + "content": "6. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.909, + 0.803 + ], + "angle": 0, + "content": "This paper introduces the concept of a lightweight posttraining ANN-SNN conversion algorithm with inference-scale computational cost, which can directly convert pretrained ANN models into SNNs without GPU-based training. The significant reduction in training overhead highlights the potential of the rapid deployment of large-scale SNNs in various scenarios and makes the instant on-chip conversion from pre-trained ANN to SNN possible. However, there is a tradeoff between performance and training cost, with the overall performance of the proposed post-training framework being slightly weaker compared to other training-involved SNN learning methods. Additionally, while this work focuses on the conversion of convolutional SNNs, future efforts will aim to extend such a conversion approach to transformer architectures and other complex layers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.694, + 0.833 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.901 + ], + "angle": 0, + "content": "This work is funded by National Natural Science Foundation of China (62422601,62176003,62088102), Beijing Municipal Science and Technology Program (Z241100004224004) and Beijing Nova Program (20230484362,20240484703)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24394" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.169 + ], + "angle": 0, + "content": "[1] Sander M Bohte, Joost N Kok, and Johannes A La Poutre. Spikeprop: backpropagation for networks of spiking neurons. In European Symposium on Artificial Neural Networks, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Velibor Bojkovic, Srinivas Anumasa, Giulia De Masi, Bin Gu, and Huan Xiong. Data driven threshold and potential initialization for spiking neural networks. In International Conference on Artificial Intelligence and Statistics, pages 4771-4779. PMLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.484, + 0.323 + ], + "angle": 0, + "content": "[3] Romain Brette, Michelle Rudolph, Ted Carnevale, Michael Hines, David Beeman, James M Bower, Markus Diesmann, Abigail Morrison, Philip H Goodman, Frederick C Harris, et al. Simulation of networks of spiking neurons: a review of tools and strategies. Journal of Computational Neuroscience, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.484, + 0.38 + ], + "angle": 0, + "content": "[4] Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Optimized potential initialization for low-latency spiking neural networks. In AAAI Conference on Artificial Intelligence, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.484, + 0.45 + ], + "angle": 0, + "content": "[5] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks. In International Conference on Learning Representations, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.484, + 0.506 + ], + "angle": 0, + "content": "[6] Yongqiang Cao, Yang Chen, and Deepak Khosla. Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.484, + 0.549 + ], + "angle": 0, + "content": "[7] Natalia Caporale and Yang Dan. Spike timing-dependent plasticity: a hebbian learning rule. Annual Review of Neuroscience, 31:25-46, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.484, + 0.604 + ], + "angle": 0, + "content": "[8] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.662 + ], + "angle": 0, + "content": "[9] Yanqi Chen, Zhaofei Yu, Wei Fang, Zhengyu Ma, Tiejun Huang, and Yonghong Tian. State transition of dendritic spines improves learning of sparse spiking neural networks. In International Conference on Machine Learning, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[10] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[11] Michael V DeBole, Brian Taba, Arnon Amir, Filipp Akopyan, Alexander Andreopoulos, William P Risk, Jeff Kusnitz, Carlos Ortega Otero, Tapan K Nayak, Rathinakumar Appuswamy, et al. TrueNorth: Accelerating from zero to 64 million neurons in 10 years. Computer, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[12] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. In International Conference on Learning Representations, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Peter U Diehl and Matthew Cook. Unsupervised learning of digit recognition using spike-timing-dependent plasticity. Frontiers in Computational Neuroscience, 9:99, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[14] Peter U Diehl, Daniel Neil, Jonathan Binas, Matthew Cook, Shih-Chii Liu, and Michael Pfeiffer. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In International Joint Conference on Neural Networks, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[15] Chaoteng Duan, Jianhao Ding, Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Temporal effective batch normalization in spiking neural networks. In Advances in Neural Information Processing Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.908, + 0.275 + ], + "angle": 0, + "content": "[16] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.907, + 0.332 + ], + "angle": 0, + "content": "[17] Biao Fang, Yuhao Zhang, Rui Yan, and Huajin Tang. Spike trains encoding optimization for spiking neural networks implementation in fpga. In International Conference on Advanced Computational Intelligence, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.907, + 0.403 + ], + "angle": 0, + "content": "[18] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.906, + 0.459 + ], + "angle": 0, + "content": "[19] Wulfram Gerstner, Werner M Kistler, Richard Naud, and Liam Paninski. Neuronal dynamics: From single neurons to networks and models of cognition. Cambridge University Press, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[20] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.907, + 0.56 + ], + "angle": 0, + "content": "[21] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. RMP-SNN: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Computer Vision and Pattern Recognition, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.908, + 0.615 + ], + "angle": 0, + "content": "[22] Zecheng Hao, Tong Bu, Jianhao Ding, Tiejun Huang, and Zhaofei Yu. Reducing ann-snn conversion error through residual membrane potential. arXiv preprint arXiv:2302.02091, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[23] Zecheng Hao, Jianhao Ding, Tong Bu, Tiejun Huang, and Zhaofei Yu. Bridging the gap between anns and snns by calibrating offset spikes. In International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.907, + 0.73 + ], + "angle": 0, + "content": "[24] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In International Conference on Computer Vision, 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.907, + 0.76 + ], + "angle": 0, + "content": "[25] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology press, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[26] JiaKui Hu, Man Yao, Xuerui Qiu, Yuhong Chou, Yuxuan Cai, Ning Qiao, Yonghong Tian, Xu Bo, and Guoqi Li. High-performance temporal reversible spiking neural networks with O(1) training memory and O(1) inference cost. In International Conference on Machine Learning, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[27] Zihan Huang, Xinyu Shi, Zecheng Hao, Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Towards high-performance spiking transformers from ann to snn conversion. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10688-10697, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24395" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "[28] Eric Hunsberger and Chris Eliasmith. Spiking deep networks with LIF neurons. arXiv preprint arXiv:1510.08829, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.122, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[29] Sungmin Hwang, Jeesoo Chang, Min-Hye Oh, Kyung Kyu Min, Taejin Jang, Kyungchul Park, Junsu Yu, Jong-Ho Lee, and Byung-Gook Park. Low-latency spiking neural networks using pre-charged membrane potential and delayed evaluation. Frontiers in Neuroscience, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.192, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[30] Eugene M Izhikevich. Which model to use for cortical spiking neurons? IEEE Transactions on Neural Networks, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.222, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[31] Haiyan Jiang, Srinivas Anumasa, Giulia De Masi, Huan Xiong, and Bin Gu. A unified optimization framework of ann-snn conversion: towards optimal mapping from activation values to firing rates. In International Conference on Machine Learning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.292, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[32] Haiyan Jiang, Vincent Zoonekynd, Giulia De Masi, Bin Gu, and Huan Xiong. TAB: Temporal accumulated batch normalization in spiking neural networks. In International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.349, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[33] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.419, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[34] Saeed Reza Kheradpisheh, Mohammad Ganjtabesh, Simon J Thorpe, and Timothee Masquelier. STDP-based spiking deep convolutional neural networks for object recognition. Neural Networks, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.476, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[35] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-YOLO: Spiking neural network for energy-efficient object detection. In AAAI Conference on Artificial Intelligence, 2020. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.533, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[36] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ANN: Towards efficient, accurate spiking neural networks calibration. In International Conference on Machine Learning, 2021. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.59, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[37] Yuhang Li, Shikuang Deng, Xin Dong, and Shi Gu. Error-aware conversion from ann to snn via post-training parameter calibration. International Journal of Computer Vision, 2024, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.646, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[38] Yuhang Li, Tamar Geller, Youngeun Kim, and Priyadarshini Panda. SEENN: Towards temporal spiking early exit neural networks. Advances in Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[40] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In International Conference on Computer Vision, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[41] Zibin Liu, Banglei Guan, Yang Shang, Shunkun Liang, Zhenbao Yu, and Qifeng Yu. Optical flow-guided 6dof object pose tracking with an event camera. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6501-6509, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[42] Zibin Liu, Banglei Guan, Yang Shang, Qifeng Yu, and Laurent Kneip. Line-based 6-dof object pose estimation and" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tracking with an event camera. IEEE Transactions on Image Processing, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.164 + ], + "angle": 0, + "content": "[43] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Computer Vision and Pattern Recognition, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[44] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural Networks, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.907, + 0.248 + ], + "angle": 0, + "content": "[45] TorchVision maintainers and contributors. Torchvision: Pytorch's computer vision library. https://github.com/pytorch/vision, 2016.6,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.906, + 0.292 + ], + "angle": 0, + "content": "[46] Timothée Masquelier and Simon J Thorpe. Unsupervised learning of visual features through spike timing dependent plasticity. PLoS Computational Biology, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.294, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[47] Joaquín Royo Miquel, Silvia Tolu, Frederik ET Schöller, and Roberto Galeazzi. Retinanet object detector based on analog-to-spiking neural network conversion. In International Conference on Soft Computing & Machine Intelligence, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.907, + 0.42 + ], + "angle": 0, + "content": "[48] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.907, + 0.462 + ], + "angle": 0, + "content": "[49] Nicolas Perez Nieves and Dan F. M. Goodman. Sparse spiking gradient descent. In Advances in Neural Information Processing Systems, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.464, + 0.906, + 0.518 + ], + "angle": 0, + "content": "[50] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjic chip architecture. Nature, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.906, + 0.576 + ], + "angle": 0, + "content": "[51] Nitin Rathi and Kaushik Roy. DIET-SNN: A low-latency spiking neural network with direct input encoding and leakage and threshold optimization. IEEE Transactions on Neural Networks and Learning Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.578, + 0.907, + 0.645 + ], + "angle": 0, + "content": "[52] Nitin Rathi, Gopalakrishnan Srinivasan, Priyadarshini Panda, and Kaushik Roy. Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.907, + 0.716 + ], + "angle": 0, + "content": "[53] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, Michael Pfeiffer, and Shih-Chii Liu. Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in Neuroscience, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.907, + 0.786 + ], + "angle": 0, + "content": "[54] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, Michael Pfeiffer, and Shih-Chii Liu. Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in Neuroscience, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.843 + ], + "angle": 0, + "content": "[55] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: VGG and residual architectures. Frontiers in Neuroscience, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[56] Jiangrong Shen, Qi Xu, Jian K. Liu, Yueming Wang, Gang Pan, and Huajin Tang. ESL-SNNs: An evolutionary structure learning strategy for spiking neural networks. In AAAI Conference on Artificial Intelligence, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24396" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.146 + ], + "angle": 0, + "content": "[57] Xinyu Shi, Zecheng Hao, and Zhaofei Yu. Spikingresformer: Bridging resnet and vision transformer in spiking neural networks. In Computer Vision and Pattern Recognition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[58] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in Neural Information Processing Systems, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[59] Sen Song, Kenneth D Miller, and Larry F Abbott. Competitive hebbian learning through spike-timing-dependent synaptic plasticity. Nature Neuroscience, 3(9):919-926, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[60] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: fully convolutional one-stage object detection. In International Conference on Computer Vision, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[61] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. CoRR, abs/1711.11248, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.388 + ], + "angle": 0, + "content": "[62] John J Wade, Liam J McDaid, Jose A Santos, and Heather M Sayers. Swat: A spiking neural network training algorithm for classification problems. IEEE Transactions on Neural Networks, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[63] Xiaofeng Wu, Velibor Bojkovic, Bin Gu, Kun Suo, and Kai Zou. Ftbc: Forward temporal bias correction for optimizing ann-snn conversion. In European Conference on Computer Vision, pages 155-173. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[64] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in Neuroscience, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[65] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Lup ing Shi. Direct training for spiking neural networks: Faster, larger, better. In AAAI Conference on Artificial Intelligence, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[66] Qu Yang, Jibin Wu, Malu Zhang, Yansong Chua, Xinchao Wang, and Haizhou Li. Training spiking neural networks with local tandem learning. Advances in Neural Information Processing Systems, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.698 + ], + "angle": 0, + "content": "[67] Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, Xu Bo, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.483, + 0.77 + ], + "angle": 0, + "content": "[68] Man Yao, Ole Richter, Guangshe Zhao, Ning Qiao, Yannan Xing, Dingheng Wang, Tianxiang Hu, Wei Fang, Tugba Demirci, Michele De Marchi, et al. Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip. Nature Communications, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.483, + 0.812 + ], + "angle": 0, + "content": "[69] Friedemann Zenke and Emre O Neftci. Brain-inspired learning on neuromorphic substrates. Proceedings of the IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.814, + 0.483, + 0.867 + ], + "angle": 0, + "content": "[70] Rui-Jie Zhu, Qihang Zhao, Guoqi Li, and Jason K Eshraghian. Spikegpt: Generative pre-trained language model with spiking neural networks. arXiv preprint arXiv:2302.13939, 2023.1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "24397" + } + ] +] \ No newline at end of file diff --git a/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/6b151e98-3001-4e0f-b77a-ea1941036ad2_origin.pdf b/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/6b151e98-3001-4e0f-b77a-ea1941036ad2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e3e5a8c8f18139d7008ffc4ab52d6206afcf98cd --- /dev/null +++ b/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/6b151e98-3001-4e0f-b77a-ea1941036ad2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27447e0299dc9e191549d230391e05990be887e96bda5c8ff026bfd81e6de25f +size 2272751 diff --git a/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/full.md b/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d783f2f76e098953da7ee9c28994ea32276c4f11 --- /dev/null +++ b/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/full.md @@ -0,0 +1,382 @@ +# Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications + +Tong $\mathsf{Bu}^{1,3\dagger}$ , Maohua $\mathsf{Li}^{2\dagger}$ , Zhaofei $\mathsf{Yu}^{1,3*}$ + +$\dagger$ Equal contribution, \* Corresponding author + +putong30@pku.edu.cn, maohuali@hhu.edu.cn, yuzf12@pku.edu.cn + +# Abstract + +Spiking Neural Networks (SNNs) have emerged as a promising substitute for Artificial Neural Networks (ANNs) due to their advantages of fast inference and low power consumption. However, the lack of efficient training algorithms has hindered their widespread adoption. Even efficient ANN-SNN conversion methods necessitate quantized training of ANNs to enhance the effectiveness of the conversion, incurring additional training costs. To address these challenges, we propose an efficient ANN-SNN conversion framework with only inference scale complexity. The conversion framework includes a local threshold balancing algorithm, which enables efficient calculation of the optimal thresholds and fine-grained adjustment of the threshold value by channel-wise scaling. We also introduce an effective delayed evaluation strategy to mitigate the influence of the spike propagation delays. We demonstrate the scalability of our framework in typical computer vision tasks: image classification, semantic segmentation, object detection, and video classification. Our algorithm outperforms existing methods, highlighting its practical applicability and efficiency. Moreover, we have evaluated the energy consumption of the converted SNNs, demonstrating their superior low-power advantage compared to conventional ANNs. This approach simplifies the deployment of SNNs by leveraging open-source pre-trained ANN models, enabling fast, low-power inference with negligible performance reduction. Code is available at https://github.com/putshua/Inference-scale-ANN-SNN. + +# 1. Introduction + +Recent advancements in large models have significantly reshaped the landscape of deep learning technology, industry, and the research community. These advanced models, characterized by their massive scale and unprecedented zero + +![](images/61ab8c1b4dc0d063529ff2c2da68a7ba568537585be7cc2f9114ae143f00450a.jpg) +Figure 1. Compared to existing frameworks that require retraining a quantized ANN, the proposed framework is able to directly convert a pre-trained ANN to an SNN at inference-scale complexity, significantly reducing computational requirements, minimizing dependence on GPUs, and requiring only a small subset of the original dataset. + +shot generalizability in downstream tasks, greatly impact our daily lives. Nevertheless, the pursuit of larger models raises concerns about high energy consumption for model inference and training. The deployment of large models on resource-constrained devices has also become a challenge. + +Spiking neural networks (SNNs), a well-established type of neural network that mimic the function of biological neurons and are considered the third generation of ANNs [44], offer a potential solution to these issues as an alternative to Artificial Neural Networks (ANNs). Unlike ANNs, spiking neurons encode information as discrete events (binary spikes or action potentials) and generate outputs based on both their inputs and internal state. This unique processing mechanism enables SNNs to effectively handle spatio-temporal information while leveraging sparse representations, leading to more efficient and biologically plausible information processing. + +With the development of the neuromorphic computing hardware [10, 11, 17, 41, 42, 49, 50, 68, 69], SNNs can now be deployed on neuromorphic chips and further applied in power-limited scenarios [9, 56]. However, the lack of efficient training algorithms has hindered their widespread application. Recent learning methods have made significant progress in training deep convolutional SNNs [15, 26] and large spiking transformers using supervised learning algorithms [57, 67, 70]. While supervised training enables SNNs + +to achieve performance comparable to ANNs with the same architecture and reduces inference time, the memory and time costs for training scale linearly with the number of inference time-steps, making it impractical for training large energy-efficient models. A more feasible approach is ANN-to-SNN conversion [6], which transforms pre-trained ANNs into SNNs with minimal computational overhead. However, converted SNNs typically require more inference time-steps to match the performance of their ANN counterparts, leading to increased latency and higher energy consumption. A potential solution is to re-train a modified ANN before conversion, which can significantly enhance SNN performance while reducing inference time-steps [4, 29]. Although retraining-based conversion methods generally improve performance across most tasks, the additional computational cost of re-training poses a notable overhead. + +In this paper, we propose an efficient post-training ANN-SNN conversion algorithm that directly obtains high-performance SNNs from pre-trained ANNs with inference-scale complexity. Figure 1 illustrates the key difference between existing ANN-to-SNN conversion frameworks and our proposed conversion framework. Compared to training-required methods, our approach greatly reduces the need for GPU resources, requires fewer data samples, and shortens the overall conversion process. We employ a local-learning-rule based algorithm for rapid conversion from ANNs to SNNs, avoiding the requirements for computationally expensive backpropagation-based training or fine-tuning. Additionally, we introduce a delayed evaluation strategy to further enhance performance. This conversion framework simplifies the deployment of SNNs by leveraging open-source, pre-trained ANN models, allowing direct application of converted SNNs on neuromorphic hardware with only inference-level computational costs. Overall, the proposed conversion framework offers two potential advantages. First, the lightweight local learning algorithm reduces dependence on GPUs, enabling online mapping from ANN to SNN on neuromorphic chips. Second, this approach reduces the need for extensive retraining of models, making it highly suitable for the practical deployment of large models in resource-constrained environments. Our contribution can be summarized as follows: + +- We introduce an inference-scale complexity ANN-SNN conversion framework, enabling the conversion of pretrained ANN models into high-performance SNNs while maintaining computational costs at inference level or even lower. +- We propose an efficient ANN-SNN conversion framework, theoretically deriving an upper bound for the conversion error between the original ANN and the converted SNN. To minimize this error bound, we introduce a light-weight local learning method for threshold optimization. Additionally, we design a delayed evaluation strategy to mini + +gate the influence of spike propagation delays. + +- We experimentally demonstrate that our method is a lightweight, plug-and-play solution capable of integrating with various conversion methods for further performance improvement. Our framework successfully scales to three typical computer vision tasks, outperforming existing ANN-SNN conversion methods in classification tasks while also demonstrating its practical applicability in handling regression-based vision tasks. + +# 2. Related Works + +Researchers have endeavored to explore effective SNN training approaches for decades. Early studies mainly focused on the unsupervised Spike-Timing-Dependent Plasticity (STDP) algorithm [7, 25]. These studies aimed to utilize the STDP algorithm for shallow networks and feature extractors [13, 34, 46, 59, 62]. In parallel, supervised training approaches have been developed to achieve high-performance SNNs. Bohte et al. [1] were pioneers in applying the backpropagation algorithm to train SNNs, introducing a temporal-based backpropagation method that uses timestamps as intermediate variables during gradient computation. Wu et al. [64] treated SNNs as recurrently connected networks and utilized Back Propagation Through Time (BPTT) for training, termed spatial-temporal backpropagation (STBP). Unlike these methods, Cao et al. [6] proposed an ANN-SNN conversion method that achieves high-performance SNNs from pre-trained ANNs. Currently, research efforts are primarily focused on enhancing the efficiency of training algorithms for high-performance SNNs, leveraging the three methods mentioned above. + +Based on whether global backpropagation is required, SNN learning methods can be ascribed into post-training and training-involved approaches. Training-required methods, including direct training of SNNs [17, 18, 32, 48, 58, 64-66], re-training before ANN-SNN conversion [5, 12, 23, 27, 31, 38, 63] and fine-tuning of converted SNNs [2, 37, 51, 52], typically require the participation of the back-propagation algorithms, making them more computationally intensive and resource-demanding. Post-training learning methods mainly refer to conversion-based algorithms that directly convert the pre-trained ANN into SNN. Cao et al. [6] pioneered this approach by mapped the weights of a lightweight CNN network to an SNN, demonstrating significant energy efficiency improvements through hardware analysis. Diehl et al. [14] proposed a weight and threshold balancing method to mitigate performance degradation, while Hunsberger and Eliasmith [28] further extended conversion-based learning to LIF neurons. Rueckauer et al. [54] established a theoretical framework for conversion-based algorithms, leading to further optimizations such as robust normalization, an enhanced threshold balancing method designed to optimize the trade-off between inference latency and accuracy [55]. + +Han and Roy [20] extended ANN-to-SNN conversion to temporal coding SNNs, improving accuracy and introducing a reset-by-subtraction mechanism instead of the traditional reset-to-zero, thereby preventing information loss [21]. Beyond classification tasks, conversion-based methods have been successfully applied to regression problems such as object detection, with Kim et al. [35] demonstrating the feasibility of spike-based object detection models. Further optimizations include adding an initial membrane potential for improved accuracy and reduced latency [4, 22], as well as employing a lightweight conversion pipeline to optimize thresholds and introduce a bias term at each inference step [36]. Although performance is limited, those lightweight training methods are able to acquire SNNs from pre-trained ANNs with negligible cost of multiple iterations of inference. + +# 3. Preliminaries + +# 3.1. Neuron Models + +The core components of conventional ANNs are point neurons, where their forward propagation can be represented as a combination of linear transformations and non-linear activations, that is + +$$ +\boldsymbol {a} ^ {l} = A \left(\boldsymbol {w} ^ {l} \boldsymbol {a} ^ {l - 1}\right). \tag {1} +$$ + +Here, $\boldsymbol{w}^l$ is the weights between layer $l - 1$ and layer $l$ , $\boldsymbol{a}^l$ is the activation vector in layer $l$ , and $A(\cdot)$ denotes the non-linear activation function. + +The neurons in SNNs are spiking neurons with temporal structures. Like most ANN-SNN conversion methods [3, 5, 6, 12, 14, 55], we employ the Integrate-and-Fire (IF) neuron [19, 30] in this paper. For computational convenience, we use the following discrete neuron function: + +$$ +\boldsymbol {v} ^ {l} \left(t ^ {-}\right) = \boldsymbol {v} ^ {l} (t - 1) + \boldsymbol {w} ^ {l} \boldsymbol {s} ^ {l - 1} (t) \theta^ {l - 1}, \tag {2} +$$ + +$$ +\boldsymbol {s} ^ {l} (t) = H \left(\boldsymbol {v} ^ {l} \left(t ^ {-}\right) - \theta^ {l}\right), \tag {3} +$$ + +$$ +\boldsymbol {v} ^ {l} (t) = \boldsymbol {v} ^ {l} \left(t ^ {-}\right) - \theta^ {l} \boldsymbol {s} ^ {l} (t). \tag {4} +$$ + +Here $s^l(t)$ describes whether neurons in layer $l$ fire at timestep $t$ , with firing triggered when the membrane potential before firing $v^l(t^-)$ reaches the firing threshold $\theta^l$ , as represented by the Heaviside step function $H(\cdot)$ . To alleviate information loss, we adopt the "reset-by-subtraction" mechanism [21, 53] for updating the membrane potential $v^l(t)$ after spike emission. Specifically, instead of resetting to the resting potential, the membrane potential is reduced by $\theta^l$ after emitting a spike. + +# 3.2.ANN-SNN Conversion + +The pivotal idea of ANN-SNN conversion is to map the activation of analog neurons to the postsynaptic potential + +(or firing rate) of spiking neurons. Specifically, by defining the initial membrane potential as $\mathbf{v}^l(0)$ and accumulating the neuron function (Equations 2 and 4) from 1 to T and dividing both sides by time-step T, we obtain: + +$$ +\frac {\boldsymbol {v} ^ {l} (T) - \boldsymbol {v} ^ {l} (0)}{T} = \frac {\sum_ {t = 1} ^ {T} \boldsymbol {w} ^ {l} \boldsymbol {s} ^ {l - 1} (t) \theta^ {l - 1} - \sum_ {t = 1} ^ {T} \boldsymbol {s} ^ {l} (t) \theta^ {l}}{T}. \tag {5} +$$ + +This equation reveals a linear relationship between the average firing rate of neurons in adjacent layers by defining average postsynaptic potential as $r^l(t) = \sum_{i=1}^{t} s^l(i) \theta^l / t$ : + +$$ +\boldsymbol {r} ^ {l} (T) = \boldsymbol {w} ^ {l} \boldsymbol {r} ^ {l - 1} (T) - \frac {\boldsymbol {v} ^ {l} (T) - \boldsymbol {v} ^ {l} (0)}{T}. \tag {6} +$$ + +As $T \to +\infty$ , the conversion error can generally be assumed to approach zero. Therefore, using the equations presented, a trained ANN model can be converted into an SNN by replacing ReLU neurons with IF neurons, which forms the fundamental principle of ANN-SNN conversion. Since Equations 1 and 6 are not mathematically identical, some conversion error typically remains. To mitigate these errors, threshold balancing [14] and weight scaling algorithms [36, 55] play a crucial role. Both techniques aim to adjust synaptic weights or neuron thresholds according to the distribution range of neuron inputs, effectively reducing clipping errors. Previous studies have shown that weight scaling and threshold balancing are equivalent in effect and can achieve high-performance SNNs [4]. + +# 4. Methods + +In this section, we present our post-training ANN-SNN conversion framework. We begin by defining the conversion error and deriving its upper bound. After that, we propose a local threshold balancing method based on a local learning rule. Additionally, we introduce the channel-wise threshold balancing technique, the pre-neuron max pooling layer, and the delayed evaluation strategy. These techniques are seamlessly integrated into a comprehensive conversion framework to enhance both performance and efficiency. + +# 4.1. Conversion Error Bound + +We follow the conversion error representation from [5] and define the conversion error $e^l$ in layer $l$ as the 2-norm of the difference between the postsynaptic potential of neurons in SNNs and the activation output of neurons in ANNs, that is + +$$ +e ^ {l} = \left\| \mathrm {S} \left(\hat {\boldsymbol {z}} ^ {l}; \theta^ {l}\right) - \mathrm {A} \left(\boldsymbol {z} ^ {l}\right) \right\| _ {2}. \tag {7} +$$ + +Here, $\hat{z}^l = \pmb{w}^l\pmb{r}^{l - 1}$ denotes the average input current from layer $l - 1$ in SNNs, and $\pmb{z}^l = \pmb{w}^l\pmb{a}^{l - 1}$ represents the activation input from layer $l - 1$ in ANNs. Both the ANN and SNN models share the same weights, denoted as $\pmb{w}^l$ . $\mathrm{A}(z^l)$ represents the output of nonlinear ReLU activation function while + +$\mathrm{S}(\hat{z}^l;\theta^l) = \boldsymbol {r}^l (T)$ represents the average output synaptic potential at time-step $T$ with given average input $\hat{z}^l$ + +Since our primary objective is to minimize the error of the final outputs, we define the conversion error between the ANN and SNN models as the layer-wise conversion error in the last layer $L$ , denoted as $e_{\mathrm{model}} = e^{L}$ . A straightforward approach would be to directly use the conversion error between models as a loss function and optimize it, which can be viewed as a variant of knowledge distillation [66]. However, this method is computationally expensive in terms of both time and memory, resembling the challenges of supervised training of SNN. To address this, we scale the conversion error and derive an error upper bound, thereby simplifying the optimization process. We formalize this approach in the following theorem. + +Theorem 1 The layer-wise conversion error can be divided into intra-layer and inter-layer errors: + +$$ +e ^ {l} \leqslant \overbrace {\| \mathbf {S} (\hat {\boldsymbol {z}} ^ {l} ; \theta^ {l}) - \mathrm {A} (\hat {\boldsymbol {z}} ^ {l}) \| _ {2}} ^ {\text {i n t r a - l a y e r e r r o r}} + \overbrace {\| \boldsymbol {w} ^ {l} \| _ {2} e ^ {l - 1}} ^ {\text {i n t e r - l a y e r e r r o r}}. \tag {8} +$$ + +Given that both ANN and SNN models receive the same input in the first layer, leading to $e^0 = 0$ , the upper bound for the conversion error between ANN and SNN models in an $L$ -layer fully-connected network is given by + +$$ +e _ {\text {m o d e l}} = e ^ {L} \leqslant \sum_ {l = 1} ^ {L} \left(\prod_ {k = l + 1} ^ {L} \| \boldsymbol {w} ^ {k} \| _ {2}\right) \left\| \mathrm {S} \left(\hat {\boldsymbol {z}} ^ {l}; \theta^ {l}\right) - \mathrm {A} \left(\hat {\boldsymbol {z}} ^ {l}\right) \right\| _ {2} \tag {9} +$$ + +Here $\| \pmb{w}^k \|_2$ is the matrix norm $(p = 2)$ or spectral norm of the weight matrix. The detailed proof is provided in the Appendix. Theorem 1 indicates that the conversion error is bounded by the weighted sum of errors across all layers. Based on this result, we introduce the local threshold balancing algorithm in the following section. + +# 4.2. Local Threshold Balancing Algorithm + +The target of ANN-SNN conversion is to minimize the conversion error and achieve high-performance SNNs. An alternative approach is to optimize the error bound, defined as: + +$$ +\begin{array}{l} \operatorname * {a r g m i n} _ {\boldsymbol {\theta}} \mathbb {E} _ {x ^ {0} \in \mathcal {D}} \\ \left[ \sum_ {l = 1} ^ {L} \left(\prod_ {k = l + 1} ^ {L} \| \boldsymbol {w} ^ {k} \| _ {2}\right) \left\| \mathrm {S} \left(\hat {\boldsymbol {z}} ^ {l}; \theta^ {l}\right) - \mathrm {A} \left(\hat {\boldsymbol {z}} ^ {l}\right) \right\| _ {2} \right]. \tag {10} \\ \end{array} +$$ + +Here $\theta$ represents the set of threshold values to be optimized across all layers, i.e., $\pmb{\theta} = \{\theta^1,\theta^2,\dots,\theta^L\}$ . $x^{0}$ denotes the input data sample drawn from the dataset $\mathcal{D}$ . This approach leverages data-driven conversion, minimizing the expectation of the conversion error bound over the data distribution. + +To reduce computational costs during optimization, we introduce two key simplifications. Firstly, we use a greedy strategy to optimize the threshold $\theta$ layer by layer, that is + +For each layer $l$ , $\arg \min_{\theta^l} \mathbb{E}_{x^0 \in \mathcal{D}}$ + +$$ +\left[ \left(\prod_ {k = l + 1} ^ {L} \| \boldsymbol {w} ^ {k} \| _ {2}\right) \left\| \mathbf {S} \left(\hat {\boldsymbol {z}} ^ {l}; \theta^ {l}\right) - \mathbf {A} \left(\hat {\boldsymbol {z}} ^ {l}\right) \right\| _ {2} \right]. \tag {11} +$$ + +Secondly, we simplify the function $\mathrm{S}(\cdot; \theta)$ by neglecting its temporal dynamics and approximating it with the clipping function $\mathrm{C}(\cdot; \theta)$ , as previous works have demonstrated that spiking neuron and clipping functions behave equivalently given sufficient number of time-steps [12]. The average input in the SNN can therefore be estimated as $\hat{\boldsymbol{z}}^0 = \boldsymbol{z}^0$ ; $\hat{\boldsymbol{z}}^{l+1} = \boldsymbol{w}^{l+1} \mathrm{C}(\hat{\boldsymbol{z}}^l; \theta^l)$ . Furthermore, We utilize the squared norm in the objective function for smoother optimization. The optimization problem then becomes: + +For each layer $l$ , $\arg \min_{\theta^l} \mathbb{E}_{x^0 \in \mathcal{D}}$ + +$$ +\left[ \left(\prod_ {k = l + 1} ^ {L} \| \boldsymbol {w} ^ {k} \| _ {2}\right) \| \mathbf {C} (\hat {\boldsymbol {z}} ^ {l}; \theta^ {l}) - \mathbf {A} (\hat {\boldsymbol {z}} ^ {l}) \| _ {2} ^ {2} \right], \tag {12} +$$ + +$$ +\operatorname {w h e r e} \mathrm {C} (x; \theta) = \min (\max (0, x), \theta). \tag {13} +$$ + +Intuitively, we can employ the stochastic gradient descent algorithm to optimize the threshold. Since $\prod_{k = l + 1}^{L}\| \pmb{w}^{k}\|_{2}$ can be considered as a constant when weights are fixed for each layer, this term can be absorbed into the learning rate. The final update rule for the local threshold balancing algorithm at each step is: + +$$ +\Delta \theta^ {l} = - \sum_ {i = 0} ^ {N - 1} 2 \left(\hat {z} _ {i} ^ {l} - \theta^ {l}\right) H \left(\hat {z} _ {i} ^ {l} - \theta^ {l}\right), \tag {14} +$$ + +$$ +\theta^ {l} \leftarrow \theta^ {l} - \eta \Delta \theta^ {l}. \tag {15} +$$ + +Detailed derivation is provided in the Appendix. Here $\Delta \theta^l$ denotes the step size during optimization and $\eta$ is the learning rate or update step size. $N$ is the number of elements in the input vector, and $\hat{z}_i$ denotes each element in the vector $\hat{z}^l$ . $H(\cdot)$ is the Heaviside step function. + +In practice, as shown in Figure 2, we jointly optimize the threshold for each layer by sampling a small set of image samples from the training dataset. Before starting the conversion process, we first replace all ReLU activation functions $\mathrm{A}(\cdot)$ with clipping functions $\mathrm{C}(\cdot)$ and initialize the threshold to zero. We then randomly sample data batches from the training dataset and perform model inference with these data samples. The thresholds are locally updated during the inference until convergence. Notably, the parameter $\theta^l$ is updated locally, ensuring that the computational cost of each iteration is similar to the inference of the ANN model and the computational cost of the whole conversion process is at inference level. + +![](images/58effa4ec822ff32ffdf4622e35ecaad2876a2fa3ab410222d5ecabd849e6eab.jpg) +Figure 2. Illustration of the proposed conversion framework. It requires only a small subset of data for conversion using local threshold balancing. During inference, the delayed evaluation technique is employed to enhance the accuracy of output estimation. + +# 4.3. Pre-Neuron Max Pooling Layer + +The conversion of the max pooling layer presents a significant challenge. As a pivotal feature in the process of down-sampling input representations, the max pooling layer is a fundamental component in convolutional neural networks and is widely used in most convolutional architectures. However, due to the binary nature of spiking neurons, max pooling layers cannot effectively perform downsampling in feature maps. Typically, in a single max-pooling operation, multiple elements may share the same value, preventing the extraction of the most salient features. Consequently, max pooling is often avoided as a downsampling layer in SNNs. In most previous ANN-SNN methods, the max pooling layer in ANN is often replaced by an average pooling layer, and the model is re-trained or fine-tuned before conversion. + +To convert architectures that contain max pooling layers, we propose a simple yet effective method. We replace all max-pooling layers with neuron layers, allowing the preneuron floating-point input to pass through the downsampling layer before being input into the neurons. We have the following theorem: + +Theorem 2 The order of the max pooling layer and ReLU layer does not affect the ANN output results. + +$$ +\max _ {i} \mathrm {R} (z) = \mathrm {R} (\max _ {i} z), \text {w h e n} \max (\mathrm {z}) > 0. \tag {16} +$$ + +Here $\mathrm{R}(\cdot)$ denotes the ReLU activation. The detailed proof is provided in the Appendix. Theorem 2 guarantees that such an operation does not affect the final performance. When deployed on neuromorphic chips or FPGAs, the max pooling operation can be efficiently implemented through a comparator, avoiding additional floating-point multiplications and minimizing power consumption. + +# 4.4. Channel-wise Threshold Balancing + +During conversion, following Kim et al. [35], Li et al. [36], we employ a channel-wise local threshold balancing method applicable to both fully connected and convolutional layers. + +In this approach, neurons in each channel share a single threshold for convolutional layers while each neuron has an + +independent threshold for fully connected layers. Specifically, for convolutional layers, we channel-wisely determine the threshold value, while for fully connected layers, we optimize the threshold element-wise. The update function is given by: + +$$ +\Delta \theta_ {c} ^ {l} = - \sum_ {i = 0} ^ {N - 1} 2 \left(\hat {z} _ {c, i} ^ {l} - \theta_ {c} ^ {l}\right) H \left(\hat {z} _ {c, i} ^ {l} - \theta_ {c} ^ {l}\right), \tag {17} +$$ + +where $c$ denotes the channel index (or element index for fully connected layers) and $i$ denotes the element index within the current channel. + +After optimization, the thresholds for each channel can be individually absorbed into the corresponding channel of the convolutional kernel or the weight vectors of the fully connected layers. This channel-wise (or element-wise) operation ensures that the behavior of the IF neuron remains unaffected. This algorithm aims to more precisely determine the optimal threshold value while preserving the fundamental properties of spiking neurons. + +# 4.5. Delayed Evaluation Strategy + +As introduced in [4], the first spike from the last layer often takes longer than expected to appear due to propagation delays between layers. This delay creates significant fluctuations in the firing rate during the initial time-steps, as no spikes are generated until the first one reaches each layer [29]. To address this issue, we propose an effective delayed step selection strategy building upon the delayed evaluation technique [29]. + +As shown in Figure 2, the basic idea of the delayed evaluation is to compute the average spike count in the last layer from time-step $t_0$ to $T$ , where $t_0$ denotes the delayed step, rather than averaging over all time-steps. Since the output is collected from the average spike count in the interval $[t_0, T]$ at the last layer, we can rewrite Equation 6 for the last layer as + +$$ +\boldsymbol {r} ^ {L} \left(t _ {0}, T\right) = \boldsymbol {w} ^ {L} \boldsymbol {r} ^ {L - 1} \left(t _ {0}, T\right) - \frac {\boldsymbol {v} ^ {L} (T) - \boldsymbol {v} ^ {L} \left(t _ {0}\right)}{T - t _ {0}}, \tag {18} +$$ + +where $\boldsymbol{r}^l(t_0, t) = \sum_{i=t_0}^{t} \boldsymbol{s}^l(i) \theta^l / (t - t_0)$ denotes the average postsynaptic potential from time $t_0$ to $t$ , and $L$ represents the last layer index. When $T \gg t_0$ , one can still guarantee that the basic equivalence of the average postsynaptic potential and ANN activations holds. + +In our strategy, we do not start counting the average output spike until $\widetilde{t_0}$ , where $\widetilde{t_0}$ is the expected spike timing of the first spike in the last layer. We estimate $\widetilde{t_0}$ based on the converted SNN model, denoted as + +$$ +\widetilde {t _ {0}} = \sum_ {l = 1} ^ {L} \frac {\theta^ {l} - v ^ {l} (0)}{\max _ {i} \left(\overline {{\mathbf {R} (\hat {\boldsymbol {z}} ^ {l})}}\right)} \approx \sum_ {l = 1} ^ {L} \frac {\theta^ {l} - v ^ {l} (0)}{\max _ {i} \left(\overline {{\mathbf {R} (\boldsymbol {z} ^ {l})}}\right)}. \tag {19} +$$ + +Here $\max_i(\cdot)$ selects the maximum element of a given vector. The notation $\overline{(\cdot)} = \mathbb{E}_{x^0\in \mathcal{D}}[(\cdot)]$ represents the expectation of the given input over dataset. Additionally, we use the ReLU function $\mathbf{R}(\cdot)$ to prevent negative estimated time. Before evaluation, we first calculate $\widetilde{t_0}$ according to Equation 19 and consider it as an indicator for the delayed evaluation strategy. In practice, if the inference step $T$ is less than $\widetilde{t_0} +4$ , we set $t_0 = T - 4$ and collect the average output over the time interval $[T - 4,T]$ . If $T$ exceeds $\widetilde{t_0} +4$ , we use the interval $[\widetilde{t_0},T]$ for output estimation. The value 4 is empirically chosen to ensure a sufficient delay while allowing adequate time for averaging spikes over time. + +This technique introduces minimal additional training cost and does not alter the model's inference process, making it seamlessly integrated into the proposed conversion framework without disrupting the overall pipeline. + +# 5. Experiments + +We conduct extensive experiments to demonstrate the effectiveness of our method and highlight its potential practical value. We first showcase the advantages of our approach by integrating it with existing conversion frameworks, comparing it with state-of-the-art post-training conversion algorithms, and conducting ablation experiments to validate the effectiveness of the whole conversion pipeline. Subsequently, we apply our algorithm to various visual tasks, including image classification, semantic segmentation, object detection, and video classification. The feasibility of our algorithm is highlighted on both classification and regression tasks, showcasing its generalizability across different datasets and its superiority over existing conversion algorithms. + +To underscore that our conversion method does not require additional training, we utilize open-source pre-trained ANN models for conversion, with most pre-trained sourced from TorchVision [45]. Additionally, we incorporate the membrane potential initialization algorithm [5], setting the membrane potential to half of the threshold before inference. The pseudocode for our conversion pipeline and detailed training settings are provided in the Appendix. + +# 5.1. Plug-and-Play Integration for Enhanced Performance + +In this section, we aim to highlight that our proposed conversion method provides a lightweight, plug-and-play solution that can be seamlessly integrated with other conversion algorithms to enhance performance. We select two representative ANN-SNN conversion approaches from existing work: the post-training fine-tuning (calibration) conversion method LCP/ACP [36, 37], and the training-required conversion method QCFS [5]. We reproduce the conversion pipeline of both methods, incorporate a local threshold balancing method for threshold searching, and use our delayed evaluation strategy. The evaluation is conducted on the ImageNet + +
MethodANNCFInference Step
3264128256512
LCP75.66×50.2163.6668.8972.1274.17
✓53.8067.1872.1874.0774.56
ACP75.66×64.5471.1273.4574.6175.35
✓65.9572.8774.2174.7375.26
QCFS73.30×64.4371.5173.2573.6773.76
✓71.9473.0673.5073.4973.54
+ +Table 1. Integration of the proposed conversion approach with existing framework + +![](images/905acf9bacea199407f858162858455ebd244102c3cb873cfe60c8f74598e1f4.jpg) +(a) Ablation Study + +![](images/260936a879e100bc21238ee7869cc84fce9a1a415840a38493c06964c355a774.jpg) +(b) Effect of Iteration Steps +Figure 3. (a) Results on evaluating the impact of different techniques within the conversion framework. (b) Effect of iteration steps on SNN performance after conversion. + +dataset with the ResNet-34 architecture. + +The results, as detailed in Table 1, showcase the performance of various configurations. The column labeled "CF" indicates whether the proposed method is used. Notably, the composite approaches, which combine our method with either LCP/ACP or QCFS, consistently outperform the individual conversion method across most timesteps. After incorporating our methods, both QCFS and LCP exhibit performance improvements exceeding $3\%$ when $\mathrm{T} = 32$ , while maintaining consistent performance as $T$ increases. This demonstrates the advantages of using our approach in conjunction with existing conversion techniques. + +# 5.2. Ablation Study for the Conversion Framework + +We demonstrate the effectiveness of our proposed conversion framework by using the ResNet-34 architecture on the ImageNet dataset. An ablation study is conducted to evaluate the impact of different techniques within the framework. Specifically, as shown in Figure 3a, we compare the performance of four different SNNs obtained by combinations of different techniques. The baseline SNN is obtained using the Robust-Norm [54] algorithm (RN, red curve), which is a commonly used post-training ANN-SNN conversion method that sets the 99-th percentile activation value as the threshold. The other SNNs are obtained by using the layer-wise local threshold balancing algorithm (LTB, yellow curve), the channel-wise threshold balancing algorithm (CLTB, blue curve), and a combination of channel-wise threshold balancing and delayed evaluation (CLTB+DE, black curve). + +![](images/56d476b287f1f33b63b175734fd4e1b03b99d7d037d1c3c5ed10ae35c25b3c9d.jpg) +Figure 4. Effect of delayed steps + +Figure 3a presents the accuracy changes of the converted SNNs with respect to the increasing time-steps. The performance of the SNN obtained from the full conversion pipeline (black curve) consistently outperforms the other SNNs across all time-steps, demonstrating an excellent balance between fast inference and high performance of the proposed conversion framework. This is particularly evident when combined with the proposed delayed evaluation technique. Compared to the robust norm method (red curve), the peak performance of the SNNs obtained by the proposed local threshold balancing method (black, blue and yellow curves) more closely aligns with the original ANN accuracy. Furthermore, the use of the channel-wise threshold balancing (CLTB, blue curve) results in a noticeable performance improvement over the vanilla threshold balancing (LTB, yellow curve) algorithm. + +# 5.3. Effect of the Iteration Step + +We evaluate the influence of the iteration steps of the local threshold balancing algorithm. The iteration number of the local threshold balancing serves as the hyper-parameter of the method. We focus on the most straightforward approach by empirically evaluating the impact of different iteration steps. We converted five different SNNs from the pre-trained ImageNet ResNet-34 model, varying the iterations steps from 1000 to 5000. The batch size during local threshold balancing is consistently set to 100. Figure 3b shows the final accuracy of the different converted SNNs. As number of iterations increases, the peak performance of the converted SNNs at different inference steps approaches the accuracy of the original ANN. Furthermore, these SNNs demonstrate better performance at low time-steps when fewer iteration steps are used. On the ImageNet dataset, with only 1000 iterations of local threshold update, the performance gap between the SNN at 512 inference time-steps and the ANN is approximately $1\%$ . With 5000 iterations, this difference is further reduced to around $0.2\%$ . Notably, the accuracy surpasses $65\%$ within just 55 time-steps, achieving a balance between inference time and performance. + +# 5.4. Effect of the Delayed Step + +We conduct experiments to demonstrate the effectiveness of the proposed delayed strategy on the ImageNet dataset using the ResNet-34 architecture. As shown in Figure 4, we evaluate the impact of the delayed step across four different + +
ArchitectureMethodANNInference Step
3264128256≥512
Post-Training
ResNet-341RMP70.64---55.6560.08
ResNet-342LCP75.6655.1667.5672.4874.5375.44
MobileNetLCP73.400.2015.4755.9566.3572.19
MobileNetV2Ours71.8819.3949.9364.3168.8970.48
ResNet-183Ours69.7665.5568.5369.0369.2169.21
ResNet-343Ours73.3164.8871.1772.3072.4272.46
Training-involved
ResNet-343QCFS73.3064.4371.5173.2573.6773.76
ResNet-342ACP75.6664.6571.3073.9475.0075.45
ResNet-343Ours*73.3071.9473.0673.5073.4973.54
+ +1 Standard ResNet-34 architecture without batch normalization. +2 Two more layers in the first conv-block compared to standard ResNet-34. +3 Standard ResNet-18/34 architecture. + +Table 2. Comparison of the proposed method and previous works on the ImageNet dataset. + +inference time-step settings: 32 steps (red curve), 64 steps (yellow curve), 128 steps (blue curve), and 256 steps (black curve). The grey vertical line indicates the estimated $\widetilde{t_0}$ for the currently converted SNN. For each sub-figure, the x-axis represents the number of delayed steps, while the y-axis denotes the accuracy. + +Although the delayed strategy is pre-determined before inference, the estimated $\widetilde{t_0}$ can still serves as an indicator that provides a near-optimal strategy to select the delayed steps. When the total inference step is less than the estimated $\widetilde{t_0}$ (red and yellow curves), the final performance gradually improves with the increase in the delayed steps, but drops significantly when the delayed step approaches the inference step. In the situation that the total inference step exceeds $\widetilde{t_0}$ (blue and black curve), the performance first increases then fluctuates when delayed step is larger than $\widetilde{t_0}$ , and ultimately experiences a sharp decline as the delayed steps get close to the inference step. The above results indicate that we can always reach near-optimal performance with the proposed delayed evaluation strategy. + +# 5.5. Test on Image Classification task + +We evaluate our conversion method on the classification task with the ImageNet dataset, employing different architectures including ResNet and MobileNet. To highlight the advantages of the proposed conversion algorithm, we conduct comparisons with previous post-training algorithms, showcasing its superior performance. When using the ResNet-34 architecture, the SNN converted by our algorithm exceeds $71\%$ accuracy within 64 time steps. In contrast, the SNNs converted by the previous algorithm [20] and the calibrationrequired conversion algorithm [37] require longer time-steps to achieve comparable accuracy. We also implement the + +proposed conversion framework on QCFS pre-trained models (denote as "Our*" in Table 2), and compare with the sota training-involved ANN-SNN techniques. The accuracy of the proposed method reaches $71.94\%$ within 32 steps, demonstrating significant advantages compared to the fine-tuning required advanced calibration technique ACP [37] and training-involved vanilla QCFS [5]. The experiments on the ImageNet dataset demonstrate the superior performance of the proposed conversion algorithm and highlight its potential for large-scale applications. + +In addition, we demonstrate the energy-saving advantage of our method by theoretically estimating the energy efficiency of our converted SNN. We use the same energy estimation approach as Bu et al. [5]. With ResNet-34 architecture on ImageNet, SNN converted from our algorithm can achieve energy efficiency at 622FPS/W while maintaining $90\%$ performance. In comparison, the energy efficiency of the original ANN is only 22FPS/W, which is 28 times lower than that of the converted SNN. For detailed energy consumption analysis, please refer to the Appendix. + +# 5.6. Test on Semantic Segmentation, Object Detection, and Video Classification Tasks + +We further extend the proposed ANN-SNN conversion method to semantic segmentation, object detection, and video classification tasks. For the semantic segmentation task, we evaluate our method on two commonly used datasets: Pascal VOC 2012 [16] augmented by SBD [24] and MS COCO 2017 [39]. We employ various semantic segmentation models, including FCN [43] and DeepLabV3 [8], provided by TorchVision [45] and other open-source repositories. As shown in Table 3, our method is compatible with tackling semantic segmentation tasks without any preparatory training. Specifically, we achieved a $53.50\%$ mIoU with the FCN model using a ResNet-34 backbone in 64 timesteps on the Pascal VOC dataset and a $61.52\%$ mIoU with the DeepLabV3 model using a ResNet-50 architecture in 128 timesteps on the more complex MS COCO 2017 dataset. Given that pixel-level classification tasks require precise model output, this excellent performance highlights the effectiveness and generalizability of the proposed method. + +For object detection tasks, we use the benchmark dataset MS COCO 2017 [39] with open-source models from TorchVision [45]. We employ the fully convolutional object detection method RetinaNet [40] and FCOS [60], and convert the pre-trained ANN models to SNNs using the proposed method. Table 3 shows detailed performance under different inference time-steps. Within 64 time-steps, the converted SNN can achieve $27.3\%$ mAP with RetinaNet and $26.5\%$ mAP with FCOS, which represents a significant improvement compared to previous explorations [35, 47]. The previous method required over 1000 time-steps to achieve comparable performance as ANN, which is detrimental to + +
DatasetArch.ANNInference Step
3264128256
Semantic Segmentation (mIoU%)
PascalVOCFCN-1847.3043.4947.0247.2647.23
FCN-3454.9946.6053.5054.7554.76
DeepLab-1855.7548.5953.9155.4055.60
DeepLab-3458.9541.2055.7458.3658.68
MS COCOFCN-5060.6728.6650.3857.0159.53
DeepLab-5063.0139.0256.6261.5262.61
Object detection (mAP%)
MS COCORetinaNet-5036.416.627.332.234.3
FCOS-5039.212.226.532.133.6
Video Classification (Acc%)
S-Kinetics-400MC3-1863.5255.2161.6062.4362.73
R3D-1862.5653.4660.3161.1261.15
+ +Table 3. Performance on semantic segmentation (mIoU%), object detection (mAP%), and video classification (Acc%) tasks. + +real-time detection and energy efficiency. + +For the video classification tasks, we use a small Kinetics-400 dataset [33], which is a subset of the original dataset. More specifically, we split the original validation set into two independent sets, one for our threshold balancing procedure, accounting for about $60\%$ of the original validation set, and the other for the test. As a result, the new train set contains 12000 videos and the new test set contains 7881 videos. The converted MC3-18 [61] SNN model achieves $61.60\%$ accuracy within 64 time-steps, showing great potential of our method in more complex computer vision tasks. + +# 6. Conclusion and Limitation + +This paper introduces the concept of a lightweight posttraining ANN-SNN conversion algorithm with inference-scale computational cost, which can directly convert pretrained ANN models into SNNs without GPU-based training. The significant reduction in training overhead highlights the potential of the rapid deployment of large-scale SNNs in various scenarios and makes the instant on-chip conversion from pre-trained ANN to SNN possible. However, there is a tradeoff between performance and training cost, with the overall performance of the proposed post-training framework being slightly weaker compared to other training-involved SNN learning methods. Additionally, while this work focuses on the conversion of convolutional SNNs, future efforts will aim to extend such a conversion approach to transformer architectures and other complex layers. + +# 7. Acknowledgement + +This work is funded by National Natural Science Foundation of China (62422601,62176003,62088102), Beijing Municipal Science and Technology Program (Z241100004224004) and Beijing Nova Program (20230484362,20240484703). + +# References + +[1] Sander M Bohte, Joost N Kok, and Johannes A La Poutre. Spikeprop: backpropagation for networks of spiking neurons. In European Symposium on Artificial Neural Networks, 2000. 2 +[2] Velibor Bojkovic, Srinivas Anumasa, Giulia De Masi, Bin Gu, and Huan Xiong. Data driven threshold and potential initialization for spiking neural networks. In International Conference on Artificial Intelligence and Statistics, pages 4771-4779. PMLR, 2024. 2 +[3] Romain Brette, Michelle Rudolph, Ted Carnevale, Michael Hines, David Beeman, James M Bower, Markus Diesmann, Abigail Morrison, Philip H Goodman, Frederick C Harris, et al. Simulation of networks of spiking neurons: a review of tools and strategies. Journal of Computational Neuroscience, 2007. 3 +[4] Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Optimized potential initialization for low-latency spiking neural networks. In AAAI Conference on Artificial Intelligence, 2022. 2, 3, 5 +[5] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks. In International Conference on Learning Representations, 2022. 2, 3, 6, 8 +[6] Yongqiang Cao, Yang Chen, and Deepak Khosla. Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision, 2015. 2, 3 +[7] Natalia Caporale and Yang Dan. Spike timing-dependent plasticity: a hebbian learning rule. Annual Review of Neuroscience, 31:25-46, 2008. 2 +[8] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 8 +[9] Yanqi Chen, Zhaofei Yu, Wei Fang, Zhengyu Ma, Tiejun Huang, and Yonghong Tian. State transition of dendritic spines improves learning of sparse spiking neural networks. In International Conference on Machine Learning, 2022. 1 +[10] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 2018. 1 +[11] Michael V DeBole, Brian Taba, Arnon Amir, Filipp Akopyan, Alexander Andreopoulos, William P Risk, Jeff Kusnitz, Carlos Ortega Otero, Tapan K Nayak, Rathinakumar Appuswamy, et al. TrueNorth: Accelerating from zero to 64 million neurons in 10 years. Computer, 2019. 1 +[12] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. In International Conference on Learning Representations, 2021. 2, 3, 4 +[13] Peter U Diehl and Matthew Cook. Unsupervised learning of digit recognition using spike-timing-dependent plasticity. Frontiers in Computational Neuroscience, 9:99, 2015. 2 + +[14] Peter U Diehl, Daniel Neil, Jonathan Binas, Matthew Cook, Shih-Chii Liu, and Michael Pfeiffer. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In International Joint Conference on Neural Networks, 2015. 2, 3 +[15] Chaoteng Duan, Jianhao Ding, Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Temporal effective batch normalization in spiking neural networks. In Advances in Neural Information Processing Systems, 2022. 1 +[16] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision, 2015. 8 +[17] Biao Fang, Yuhao Zhang, Rui Yan, and Huajin Tang. Spike trains encoding optimization for spiking neural networks implementation in fpga. In International Conference on Advanced Computational Intelligence, 2020. 1, 2 +[18] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In International Conference on Computer Vision, 2021. 2 +[19] Wulfram Gerstner, Werner M Kistler, Richard Naud, and Liam Paninski. Neuronal dynamics: From single neurons to networks and models of cognition. Cambridge University Press, 2014. 3 +[20] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, 2020. 3, 7 +[21] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. RMP-SNN: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Computer Vision and Pattern Recognition, 2020. 3 +[22] Zecheng Hao, Tong Bu, Jianhao Ding, Tiejun Huang, and Zhaofei Yu. Reducing ann-snn conversion error through residual membrane potential. arXiv preprint arXiv:2302.02091, 2023. 3 +[23] Zecheng Hao, Jianhao Ding, Tong Bu, Tiejun Huang, and Zhaofei Yu. Bridging the gap between anns and snns by calibrating offset spikes. In International Conference on Learning Representations, 2023. 2 +[24] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In International Conference on Computer Vision, 2011. 8 +[25] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology press, 2005. 2 +[26] JiaKui Hu, Man Yao, Xuerui Qiu, Yuhong Chou, Yuxuan Cai, Ning Qiao, Yonghong Tian, Xu Bo, and Guoqi Li. High-performance temporal reversible spiking neural networks with O(1) training memory and O(1) inference cost. In International Conference on Machine Learning, 2024. 1 +[27] Zihan Huang, Xinyu Shi, Zecheng Hao, Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Towards high-performance spiking transformers from ann to snn conversion. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10688-10697, 2024. 2 + +[28] Eric Hunsberger and Chris Eliasmith. Spiking deep networks with LIF neurons. arXiv preprint arXiv:1510.08829, 2015. 2 +[29] Sungmin Hwang, Jeesoo Chang, Min-Hye Oh, Kyung Kyu Min, Taejin Jang, Kyungchul Park, Junsu Yu, Jong-Ho Lee, and Byung-Gook Park. Low-latency spiking neural networks using pre-charged membrane potential and delayed evaluation. Frontiers in Neuroscience, 2021. 2, 5 +[30] Eugene M Izhikevich. Which model to use for cortical spiking neurons? IEEE Transactions on Neural Networks, 2004. 3 +[31] Haiyan Jiang, Srinivas Anumasa, Giulia De Masi, Huan Xiong, and Bin Gu. A unified optimization framework of ann-snn conversion: towards optimal mapping from activation values to firing rates. In International Conference on Machine Learning, 2023. 2 +[32] Haiyan Jiang, Vincent Zoonekynd, Giulia De Masi, Bin Gu, and Huan Xiong. TAB: Temporal accumulated batch normalization in spiking neural networks. In International Conference on Learning Representations, 2024. 2 +[33] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 8 +[34] Saeed Reza Kheradpisheh, Mohammad Ganjtabesh, Simon J Thorpe, and Timothee Masquelier. STDP-based spiking deep convolutional neural networks for object recognition. Neural Networks, 2018. 2 +[35] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-YOLO: Spiking neural network for energy-efficient object detection. In AAAI Conference on Artificial Intelligence, 2020. 3, 5, 8 +[36] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ANN: Towards efficient, accurate spiking neural networks calibration. In International Conference on Machine Learning, 2021. 3, 5, 6 +[37] Yuhang Li, Shikuang Deng, Xin Dong, and Shi Gu. Error-aware conversion from ann to snn via post-training parameter calibration. International Journal of Computer Vision, 2024, 2, 6, 7, 8 +[38] Yuhang Li, Tamar Geller, Youngeun Kim, and Priyadarshini Panda. SEENN: Towards temporal spiking early exit neural networks. Advances in Neural Information Processing Systems, 2024. 2 +[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision, 2014. 8 +[40] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In International Conference on Computer Vision, 2017. 8 +[41] Zibin Liu, Banglei Guan, Yang Shang, Shunkun Liang, Zhenbao Yu, and Qifeng Yu. Optical flow-guided 6dof object pose tracking with an event camera. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6501-6509, 2024. 1 +[42] Zibin Liu, Banglei Guan, Yang Shang, Qifeng Yu, and Laurent Kneip. Line-based 6-dof object pose estimation and + +tracking with an event camera. IEEE Transactions on Image Processing, 2024. 1 +[43] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Computer Vision and Pattern Recognition, 2015. 8 +[44] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural Networks, 1997. 1 +[45] TorchVision maintainers and contributors. Torchvision: Pytorch's computer vision library. https://github.com/pytorch/vision, 2016.6,8 +[46] Timothée Masquelier and Simon J Thorpe. Unsupervised learning of visual features through spike timing dependent plasticity. PLoS Computational Biology, 2007. 2 +[47] Joaquín Royo Miquel, Silvia Tolu, Frederik ET Schöller, and Roberto Galeazzi. Retinanet object detector based on analog-to-spiking neural network conversion. In International Conference on Soft Computing & Machine Intelligence, 2021. 8 +[48] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 2019. 2 +[49] Nicolas Perez Nieves and Dan F. M. Goodman. Sparse spiking gradient descent. In Advances in Neural Information Processing Systems, 2021. 1 +[50] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjic chip architecture. Nature, 2019. 1 +[51] Nitin Rathi and Kaushik Roy. DIET-SNN: A low-latency spiking neural network with direct input encoding and leakage and threshold optimization. IEEE Transactions on Neural Networks and Learning Systems, 2021. 2 +[52] Nitin Rathi, Gopalakrishnan Srinivasan, Priyadarshini Panda, and Kaushik Roy. Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation. In International Conference on Learning Representations, 2020. 2 +[53] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, Michael Pfeiffer, and Shih-Chii Liu. Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in Neuroscience, 2017. 3 +[54] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, Michael Pfeiffer, and Shih-Chii Liu. Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in Neuroscience, 2017. 2, 6 +[55] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: VGG and residual architectures. Frontiers in Neuroscience, 2019. 2, 3 +[56] Jiangrong Shen, Qi Xu, Jian K. Liu, Yueming Wang, Gang Pan, and Huajin Tang. ESL-SNNs: An evolutionary structure learning strategy for spiking neural networks. In AAAI Conference on Artificial Intelligence, 2023. 1 + +[57] Xinyu Shi, Zecheng Hao, and Zhaofei Yu. Spikingresformer: Bridging resnet and vision transformer in spiking neural networks. In Computer Vision and Pattern Recognition, 2024. 1 +[58] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in Neural Information Processing Systems, 2018. 2 +[59] Sen Song, Kenneth D Miller, and Larry F Abbott. Competitive hebbian learning through spike-timing-dependent synaptic plasticity. Nature Neuroscience, 3(9):919-926, 2000. 2 +[60] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: fully convolutional one-stage object detection. In International Conference on Computer Vision, 2019. 8 +[61] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. CoRR, abs/1711.11248, 2017. 8 +[62] John J Wade, Liam J McDaid, Jose A Santos, and Heather M Sayers. Swat: A spiking neural network training algorithm for classification problems. IEEE Transactions on Neural Networks, 2010. 2 +[63] Xiaofeng Wu, Velibor Bojkovic, Bin Gu, Kun Suo, and Kai Zou. Ftbc: Forward temporal bias correction for optimizing ann-snn conversion. In European Conference on Computer Vision, pages 155-173. Springer, 2024. 2 +[64] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in Neuroscience, 2018. 2 +[65] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Lup ing Shi. Direct training for spiking neural networks: Faster, larger, better. In AAAI Conference on Artificial Intelligence, 2019. +[66] Qu Yang, Jibin Wu, Malu Zhang, Yansong Chua, Xinchao Wang, and Haizhou Li. Training spiking neural networks with local tandem learning. Advances in Neural Information Processing Systems, 2022. 2, 4 +[67] Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, Xu Bo, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In International Conference on Learning Representations, 2024. 1 +[68] Man Yao, Ole Richter, Guangshe Zhao, Ning Qiao, Yannan Xing, Dingheng Wang, Tianxiang Hu, Wei Fang, Tugba Demirci, Michele De Marchi, et al. Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip. Nature Communications, 2024. 1 +[69] Friedemann Zenke and Emre O Neftci. Brain-inspired learning on neuromorphic substrates. Proceedings of the IEEE, 2021. 1 +[70] Rui-Jie Zhu, Qihang Zhao, Guoqi Li, and Jason K Eshraghian. Spikegpt: Generative pre-trained language model with spiking neural networks. arXiv preprint arXiv:2302.13939, 2023.1 \ No newline at end of file diff --git a/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/images.zip b/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..64db617f7ebaf927bae44964e9397ca8a2a8e496 --- /dev/null +++ b/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96ed0878489eb8c98f463a995720ef19bf192c649cd3b3ac2480ca66814782ad +size 354575 diff --git a/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/layout.json b/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8b7bacbc1da749a1e7194030403c5a908efcf3b0 --- /dev/null +++ b/2025/Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications/layout.json @@ -0,0 +1,9897 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 553, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 553, + 140 + ], + "type": "text", + "content": "Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 204, + 150, + 403, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 150, + 403, + 165 + ], + "spans": [ + { + "bbox": [ + 204, + 150, + 403, + 165 + ], + "type": "text", + "content": "Tong " + }, + { + "bbox": [ + 204, + 150, + 403, + 165 + ], + "type": "inline_equation", + "content": "\\mathsf{Bu}^{1,3\\dagger}" + }, + { + "bbox": [ + 204, + 150, + 403, + 165 + ], + "type": "text", + "content": ", Maohua " + }, + { + "bbox": [ + 204, + 150, + 403, + 165 + ], + "type": "inline_equation", + "content": "\\mathsf{Li}^{2\\dagger}" + }, + { + "bbox": [ + 204, + 150, + 403, + 165 + ], + "type": "text", + "content": ", Zhaofei " + }, + { + "bbox": [ + 204, + 150, + 403, + 165 + ], + "type": "inline_equation", + "content": "\\mathsf{Yu}^{1,3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 167, + 387, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 167, + 387, + 179 + ], + "spans": [ + { + "bbox": [ + 225, + 167, + 387, + 179 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 225, + 167, + 387, + 179 + ], + "type": "text", + "content": " Equal contribution, \\* Corresponding author" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 182, + 463, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 182, + 463, + 193 + ], + "spans": [ + { + "bbox": [ + 143, + 182, + 463, + 193 + ], + "type": "text", + "content": "putong30@pku.edu.cn, maohuali@hhu.edu.cn, yuzf12@pku.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "spans": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 258, + 297, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 297, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 297, + 582 + ], + "type": "text", + "content": "Spiking Neural Networks (SNNs) have emerged as a promising substitute for Artificial Neural Networks (ANNs) due to their advantages of fast inference and low power consumption. However, the lack of efficient training algorithms has hindered their widespread adoption. Even efficient ANN-SNN conversion methods necessitate quantized training of ANNs to enhance the effectiveness of the conversion, incurring additional training costs. To address these challenges, we propose an efficient ANN-SNN conversion framework with only inference scale complexity. The conversion framework includes a local threshold balancing algorithm, which enables efficient calculation of the optimal thresholds and fine-grained adjustment of the threshold value by channel-wise scaling. We also introduce an effective delayed evaluation strategy to mitigate the influence of the spike propagation delays. We demonstrate the scalability of our framework in typical computer vision tasks: image classification, semantic segmentation, object detection, and video classification. Our algorithm outperforms existing methods, highlighting its practical applicability and efficiency. Moreover, we have evaluated the energy consumption of the converted SNNs, demonstrating their superior low-power advantage compared to conventional ANNs. This approach simplifies the deployment of SNNs by leveraging open-source pre-trained ANN models, enabling fast, low-power inference with negligible performance reduction. Code is available at https://github.com/putshua/Inference-scale-ANN-SNN." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 605, + 135, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 605, + 135, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 605, + 135, + 617 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 625, + 297, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 297, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 297, + 674 + ], + "type": "text", + "content": "Recent advancements in large models have significantly reshaped the landscape of deep learning technology, industry, and the research community. These advanced models, characterized by their massive scale and unprecedented zero" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 220, + 556, + 322 + ], + "blocks": [ + { + "bbox": [ + 316, + 220, + 556, + 322 + ], + "lines": [ + { + "bbox": [ + 316, + 220, + 556, + 322 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 556, + 322 + ], + "type": "image", + "image_path": "61ab8c1b4dc0d063529ff2c2da68a7ba568537585be7cc2f9114ae143f00450a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 330, + 556, + 396 + ], + "lines": [ + { + "bbox": [ + 313, + 330, + 556, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 556, + 396 + ], + "type": "text", + "content": "Figure 1. Compared to existing frameworks that require retraining a quantized ANN, the proposed framework is able to directly convert a pre-trained ANN to an SNN at inference-scale complexity, significantly reducing computational requirements, minimizing dependence on GPUs, and requiring only a small subset of the original dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 413, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 473 + ], + "type": "text", + "content": "shot generalizability in downstream tasks, greatly impact our daily lives. Nevertheless, the pursuit of larger models raises concerns about high energy consumption for model inference and training. The deployment of large models on resource-constrained devices has also become a challenge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 473, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 556, + 605 + ], + "type": "text", + "content": "Spiking neural networks (SNNs), a well-established type of neural network that mimic the function of biological neurons and are considered the third generation of ANNs [44], offer a potential solution to these issues as an alternative to Artificial Neural Networks (ANNs). Unlike ANNs, spiking neurons encode information as discrete events (binary spikes or action potentials) and generate outputs based on both their inputs and internal state. This unique processing mechanism enables SNNs to effectively handle spatio-temporal information while leveraging sparse representations, leading to more efficient and biologically plausible information processing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "With the development of the neuromorphic computing hardware [10, 11, 17, 41, 42, 49, 50, 68, 69], SNNs can now be deployed on neuromorphic chips and further applied in power-limited scenarios [9, 56]. However, the lack of efficient training algorithms has hindered their widespread application. Recent learning methods have made significant progress in training deep convolutional SNNs [15, 26] and large spiking transformers using supervised learning algorithms [57, 67, 70]. While supervised training enables SNNs" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 683, + 247, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 683, + 247, + 693 + ], + "spans": [ + { + "bbox": [ + 69, + 683, + 247, + 693 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 683, + 247, + 693 + ], + "type": "text", + "content": " Institute for Artificial Intelligence, Peking University" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 69, + 693, + 292, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 292, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 292, + 703 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 693, + 292, + 703 + ], + "type": "text", + "content": " School of Artificial Intelligence and Automation, Hohai University" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 69, + 703, + 231, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 231, + 712 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 231, + 712 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 703, + 231, + 712 + ], + "type": "text", + "content": " School of Computer Science, Peking University" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24387" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 264 + ], + "type": "text", + "content": "to achieve performance comparable to ANNs with the same architecture and reduces inference time, the memory and time costs for training scale linearly with the number of inference time-steps, making it impractical for training large energy-efficient models. A more feasible approach is ANN-to-SNN conversion [6], which transforms pre-trained ANNs into SNNs with minimal computational overhead. However, converted SNNs typically require more inference time-steps to match the performance of their ANN counterparts, leading to increased latency and higher energy consumption. A potential solution is to re-train a modified ANN before conversion, which can significantly enhance SNN performance while reducing inference time-steps [4, 29]. Although retraining-based conversion methods generally improve performance across most tasks, the additional computational cost of re-training poses a notable overhead." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 268, + 298, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 298, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 298, + 578 + ], + "type": "text", + "content": "In this paper, we propose an efficient post-training ANN-SNN conversion algorithm that directly obtains high-performance SNNs from pre-trained ANNs with inference-scale complexity. Figure 1 illustrates the key difference between existing ANN-to-SNN conversion frameworks and our proposed conversion framework. Compared to training-required methods, our approach greatly reduces the need for GPU resources, requires fewer data samples, and shortens the overall conversion process. We employ a local-learning-rule based algorithm for rapid conversion from ANNs to SNNs, avoiding the requirements for computationally expensive backpropagation-based training or fine-tuning. Additionally, we introduce a delayed evaluation strategy to further enhance performance. This conversion framework simplifies the deployment of SNNs by leveraging open-source, pre-trained ANN models, allowing direct application of converted SNNs on neuromorphic hardware with only inference-level computational costs. Overall, the proposed conversion framework offers two potential advantages. First, the lightweight local learning algorithm reduces dependence on GPUs, enabling online mapping from ANN to SNN on neuromorphic chips. Second, this approach reduces the need for extensive retraining of models, making it highly suitable for the practical deployment of large models in resource-constrained environments. Our contribution can be summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 582, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 582, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 640 + ], + "type": "text", + "content": "- We introduce an inference-scale complexity ANN-SNN conversion framework, enabling the conversion of pretrained ANN models into high-performance SNNs while maintaining computational costs at inference level or even lower." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 712 + ], + "type": "text", + "content": "- We propose an efficient ANN-SNN conversion framework, theoretically deriving an upper bound for the conversion error between the original ANN and the converted SNN. To minimize this error bound, we introduce a light-weight local learning method for threshold optimization. Additionally, we design a delayed evaluation strategy to mini" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 511, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 511, + 84 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 511, + 84 + ], + "type": "text", + "content": "gate the influence of spike propagation delays." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 84, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 555, + 180 + ], + "type": "text", + "content": "- We experimentally demonstrate that our method is a lightweight, plug-and-play solution capable of integrating with various conversion methods for further performance improvement. Our framework successfully scales to three typical computer vision tasks, outperforming existing ANN-SNN conversion methods in classification tasks while also demonstrating its practical applicability in handling regression-based vision tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 190, + 406, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 190, + 406, + 204 + ], + "spans": [ + { + "bbox": [ + 314, + 190, + 406, + 204 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 211, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 556, + 449 + ], + "type": "text", + "content": "Researchers have endeavored to explore effective SNN training approaches for decades. Early studies mainly focused on the unsupervised Spike-Timing-Dependent Plasticity (STDP) algorithm [7, 25]. These studies aimed to utilize the STDP algorithm for shallow networks and feature extractors [13, 34, 46, 59, 62]. In parallel, supervised training approaches have been developed to achieve high-performance SNNs. Bohte et al. [1] were pioneers in applying the backpropagation algorithm to train SNNs, introducing a temporal-based backpropagation method that uses timestamps as intermediate variables during gradient computation. Wu et al. [64] treated SNNs as recurrently connected networks and utilized Back Propagation Through Time (BPTT) for training, termed spatial-temporal backpropagation (STBP). Unlike these methods, Cao et al. [6] proposed an ANN-SNN conversion method that achieves high-performance SNNs from pre-trained ANNs. Currently, research efforts are primarily focused on enhancing the efficiency of training algorithms for high-performance SNNs, leveraging the three methods mentioned above." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "content": "Based on whether global backpropagation is required, SNN learning methods can be ascribed into post-training and training-involved approaches. Training-required methods, including direct training of SNNs [17, 18, 32, 48, 58, 64-66], re-training before ANN-SNN conversion [5, 12, 23, 27, 31, 38, 63] and fine-tuning of converted SNNs [2, 37, 51, 52], typically require the participation of the back-propagation algorithms, making them more computationally intensive and resource-demanding. Post-training learning methods mainly refer to conversion-based algorithms that directly convert the pre-trained ANN into SNN. Cao et al. [6] pioneered this approach by mapped the weights of a lightweight CNN network to an SNN, demonstrating significant energy efficiency improvements through hardware analysis. Diehl et al. [14] proposed a weight and threshold balancing method to mitigate performance degradation, while Hunsberger and Eliasmith [28] further extended conversion-based learning to LIF neurons. Rueckauer et al. [54] established a theoretical framework for conversion-based algorithms, leading to further optimizations such as robust normalization, an enhanced threshold balancing method designed to optimize the trade-off between inference latency and accuracy [55]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24388" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": "Han and Roy [20] extended ANN-to-SNN conversion to temporal coding SNNs, improving accuracy and introducing a reset-by-subtraction mechanism instead of the traditional reset-to-zero, thereby preventing information loss [21]. Beyond classification tasks, conversion-based methods have been successfully applied to regression problems such as object detection, with Kim et al. [35] demonstrating the feasibility of spike-based object detection models. Further optimizations include adding an initial membrane potential for improved accuracy and reduced latency [4, 22], as well as employing a lightweight conversion pipeline to optimize thresholds and introduce a bias term at each inference step [36]. Although performance is limited, those lightweight training methods are able to acquire SNNs from pre-trained ANNs with negligible cost of multiple iterations of inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 275, + 141, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 141, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 141, + 287 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 295, + 151, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 151, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 151, + 306 + ], + "type": "text", + "content": "3.1. Neuron Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 360 + ], + "type": "text", + "content": "The core components of conventional ANNs are point neurons, where their forward propagation can be represented as a combination of linear transformations and non-linear activations, that is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 369, + 296, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 369, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 135, + 369, + 296, + 384 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} ^ {l} = A \\left(\\boldsymbol {w} ^ {l} \\boldsymbol {a} ^ {l - 1}\\right). \\tag {1}", + "image_path": "81b3741f738b8fbfa1233f02fc7a608ee7b1fd390d720477149dc05c8301667a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}^l" + }, + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "text", + "content": " is the weights between layer " + }, + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "inline_equation", + "content": "l - 1" + }, + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "text", + "content": " and layer " + }, + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\boldsymbol{a}^l" + }, + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "text", + "content": " is the activation vector in layer " + }, + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "inline_equation", + "content": "A(\\cdot)" + }, + { + "bbox": [ + 55, + 393, + 296, + 429 + ], + "type": "text", + "content": " denotes the non-linear activation function." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 430, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 489 + ], + "type": "text", + "content": "The neurons in SNNs are spiking neurons with temporal structures. Like most ANN-SNN conversion methods [3, 5, 6, 12, 14, 55], we employ the Integrate-and-Fire (IF) neuron [19, 30] in this paper. For computational convenience, we use the following discrete neuron function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 97, + 498, + 296, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 498, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 97, + 498, + 296, + 512 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} ^ {l} \\left(t ^ {-}\\right) = \\boldsymbol {v} ^ {l} (t - 1) + \\boldsymbol {w} ^ {l} \\boldsymbol {s} ^ {l - 1} (t) \\theta^ {l - 1}, \\tag {2}", + "image_path": "3f0f2629463c842ce45ca447c0a600eddfc64a9cf5eba02475084fe651eedf1e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 514, + 296, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 514, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 296, + 529 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} ^ {l} (t) = H \\left(\\boldsymbol {v} ^ {l} \\left(t ^ {-}\\right) - \\theta^ {l}\\right), \\tag {3}", + "image_path": "2f987dabb7a3006c2c5087e6e31069588223e7523b678d89f751d34a260671bd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 531, + 296, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 531, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 296, + 546 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} ^ {l} (t) = \\boldsymbol {v} ^ {l} \\left(t ^ {-}\\right) - \\theta^ {l} \\boldsymbol {s} ^ {l} (t). \\tag {4}", + "image_path": "a47f8c4e10157b15278dcc01f6bdbdb8664d51f0d116c680633d47ae633bb086.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "inline_equation", + "content": "s^l(t)" + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "text", + "content": " describes whether neurons in layer " + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "text", + "content": " fire at timestep " + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "text", + "content": ", with firing triggered when the membrane potential before firing " + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "inline_equation", + "content": "v^l(t^-)" + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "text", + "content": " reaches the firing threshold " + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "inline_equation", + "content": "\\theta^l" + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "text", + "content": ", as represented by the Heaviside step function " + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "text", + "content": ". To alleviate information loss, we adopt the \"reset-by-subtraction\" mechanism [21, 53] for updating the membrane potential " + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "inline_equation", + "content": "v^l(t)" + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "text", + "content": " after spike emission. Specifically, instead of resetting to the resting potential, the membrane potential is reduced by " + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "inline_equation", + "content": "\\theta^l" + }, + { + "bbox": [ + 55, + 555, + 297, + 664 + ], + "type": "text", + "content": " after emitting a spike." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 672, + 185, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 185, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 185, + 683 + ], + "type": "text", + "content": "3.2.ANN-SNN Conversion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "The pivotal idea of ANN-SNN conversion is to map the activation of analog neurons to the postsynaptic potential" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "(or firing rate) of spiking neurons. Specifically, by defining the initial membrane potential as " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^l(0)" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " and accumulating the neuron function (Equations 2 and 4) from 1 to T and dividing both sides by time-step T, we obtain:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 126, + 555, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 126, + 555, + 165 + ], + "spans": [ + { + "bbox": [ + 318, + 126, + 555, + 165 + ], + "type": "interline_equation", + "content": "\\frac {\\boldsymbol {v} ^ {l} (T) - \\boldsymbol {v} ^ {l} (0)}{T} = \\frac {\\sum_ {t = 1} ^ {T} \\boldsymbol {w} ^ {l} \\boldsymbol {s} ^ {l - 1} (t) \\theta^ {l - 1} - \\sum_ {t = 1} ^ {T} \\boldsymbol {s} ^ {l} (t) \\theta^ {l}}{T}. \\tag {5}", + "image_path": "67a829a38d3c3fdd5e79fb5d63fff75c40c2a169c1d8f92bb3b3222391f5b6fe.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 172, + 555, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 209 + ], + "type": "text", + "content": "This equation reveals a linear relationship between the average firing rate of neurons in adjacent layers by defining average postsynaptic potential as " + }, + { + "bbox": [ + 313, + 172, + 555, + 209 + ], + "type": "inline_equation", + "content": "r^l(t) = \\sum_{i=1}^{t} s^l(i) \\theta^l / t" + }, + { + "bbox": [ + 313, + 172, + 555, + 209 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 353, + 216, + 555, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 216, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 353, + 216, + 555, + 241 + ], + "type": "interline_equation", + "content": "\\boldsymbol {r} ^ {l} (T) = \\boldsymbol {w} ^ {l} \\boldsymbol {r} ^ {l - 1} (T) - \\frac {\\boldsymbol {v} ^ {l} (T) - \\boldsymbol {v} ^ {l} (0)}{T}. \\tag {6}", + "image_path": "c7b5633ff56748489574c6ab25abb68170ee8c8d099df8b47c04eafbde71b315.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 247, + 556, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 556, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 556, + 415 + ], + "type": "text", + "content": "As " + }, + { + "bbox": [ + 313, + 247, + 556, + 415 + ], + "type": "inline_equation", + "content": "T \\to +\\infty" + }, + { + "bbox": [ + 313, + 247, + 556, + 415 + ], + "type": "text", + "content": ", the conversion error can generally be assumed to approach zero. Therefore, using the equations presented, a trained ANN model can be converted into an SNN by replacing ReLU neurons with IF neurons, which forms the fundamental principle of ANN-SNN conversion. Since Equations 1 and 6 are not mathematically identical, some conversion error typically remains. To mitigate these errors, threshold balancing [14] and weight scaling algorithms [36, 55] play a crucial role. Both techniques aim to adjust synaptic weights or neuron thresholds according to the distribution range of neuron inputs, effectively reducing clipping errors. Previous studies have shown that weight scaling and threshold balancing are equivalent in effect and can achieve high-performance SNNs [4]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 424, + 375, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 424, + 375, + 436 + ], + "spans": [ + { + "bbox": [ + 314, + 424, + 375, + 436 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 444, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 556, + 552 + ], + "type": "text", + "content": "In this section, we present our post-training ANN-SNN conversion framework. We begin by defining the conversion error and deriving its upper bound. After that, we propose a local threshold balancing method based on a local learning rule. Additionally, we introduce the channel-wise threshold balancing technique, the pre-neuron max pooling layer, and the delayed evaluation strategy. These techniques are seamlessly integrated into a comprehensive conversion framework to enhance both performance and efficiency." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 559, + 454, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 454, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 454, + 571 + ], + "type": "text", + "content": "4.1. Conversion Error Bound" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 576, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 555, + 624 + ], + "type": "text", + "content": "We follow the conversion error representation from [5] and define the conversion error " + }, + { + "bbox": [ + 313, + 576, + 555, + 624 + ], + "type": "inline_equation", + "content": "e^l" + }, + { + "bbox": [ + 313, + 576, + 555, + 624 + ], + "type": "text", + "content": " in layer " + }, + { + "bbox": [ + 313, + 576, + 555, + 624 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 576, + 555, + 624 + ], + "type": "text", + "content": " as the 2-norm of the difference between the postsynaptic potential of neurons in SNNs and the activation output of neurons in ANNs, that is" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 377, + 630, + 555, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 630, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 377, + 630, + 555, + 645 + ], + "type": "interline_equation", + "content": "e ^ {l} = \\left\\| \\mathrm {S} \\left(\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}\\right) - \\mathrm {A} \\left(\\boldsymbol {z} ^ {l}\\right) \\right\\| _ {2}. \\tag {7}", + "image_path": "2a0b7f77be9f90a5b8479538b93e9d34f5566fcfb3584968bf34062194d815a6.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\hat{z}^l = \\pmb{w}^l\\pmb{r}^{l - 1}" + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "content": " denotes the average input current from layer " + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "inline_equation", + "content": "l - 1" + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "content": " in SNNs, and " + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{z}^l = \\pmb{w}^l\\pmb{a}^{l - 1}" + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "content": " represents the activation input from layer " + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "inline_equation", + "content": "l - 1" + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "content": " in ANNs. Both the ANN and SNN models share the same weights, denoted as " + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{w}^l" + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{A}(z^l)" + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "content": " represents the output of nonlinear ReLU activation function while" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24389" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\mathrm{S}(\\hat{z}^l;\\theta^l) = \\boldsymbol {r}^l (T)" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": " represents the average output synaptic potential at time-step " + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": " with given average input " + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\hat{z}^l" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "text", + "content": "Since our primary objective is to minimize the error of the final outputs, we define the conversion error between the ANN and SNN models as the layer-wise conversion error in the last layer " + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{model}} = e^{L}" + }, + { + "bbox": [ + 55, + 96, + 296, + 251 + ], + "type": "text", + "content": ". A straightforward approach would be to directly use the conversion error between models as a loss function and optimize it, which can be viewed as a variant of knowledge distillation [66]. However, this method is computationally expensive in terms of both time and memory, resembling the challenges of supervised training of SNN. To address this, we scale the conversion error and derive an error upper bound, thereby simplifying the optimization process. We formalize this approach in the following theorem." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 261, + 296, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 296, + 285 + ], + "type": "text", + "content": "Theorem 1 The layer-wise conversion error can be divided into intra-layer and inter-layer errors:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 293, + 296, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 293, + 296, + 322 + ], + "spans": [ + { + "bbox": [ + 90, + 293, + 296, + 322 + ], + "type": "interline_equation", + "content": "e ^ {l} \\leqslant \\overbrace {\\| \\mathbf {S} (\\hat {\\boldsymbol {z}} ^ {l} ; \\theta^ {l}) - \\mathrm {A} (\\hat {\\boldsymbol {z}} ^ {l}) \\| _ {2}} ^ {\\text {i n t r a - l a y e r e r r o r}} + \\overbrace {\\| \\boldsymbol {w} ^ {l} \\| _ {2} e ^ {l - 1}} ^ {\\text {i n t e r - l a y e r e r r o r}}. \\tag {8}", + "image_path": "46efd8bc62c8d54c77559fc7c3dd242b71a74f31843261e13dd262a675d0576d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 331, + 296, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 331, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 296, + 379 + ], + "type": "text", + "content": "Given that both ANN and SNN models receive the same input in the first layer, leading to " + }, + { + "bbox": [ + 54, + 331, + 296, + 379 + ], + "type": "inline_equation", + "content": "e^0 = 0" + }, + { + "bbox": [ + 54, + 331, + 296, + 379 + ], + "type": "text", + "content": ", the upper bound for the conversion error between ANN and SNN models in an " + }, + { + "bbox": [ + 54, + 331, + 296, + 379 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 54, + 331, + 296, + 379 + ], + "type": "text", + "content": "-layer fully-connected network is given by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 386, + 296, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 386, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 58, + 386, + 296, + 431 + ], + "type": "interline_equation", + "content": "e _ {\\text {m o d e l}} = e ^ {L} \\leqslant \\sum_ {l = 1} ^ {L} \\left(\\prod_ {k = l + 1} ^ {L} \\| \\boldsymbol {w} ^ {k} \\| _ {2}\\right) \\left\\| \\mathrm {S} \\left(\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}\\right) - \\mathrm {A} \\left(\\hat {\\boldsymbol {z}} ^ {l}\\right) \\right\\| _ {2} \\tag {9}", + "image_path": "9dcda0e59f356bd82fca118bbfb7bc204d6aec7c3d3c89ebba64c288018ec6bb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 441, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 514 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 441, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\| \\pmb{w}^k \\|_2" + }, + { + "bbox": [ + 55, + 441, + 296, + 514 + ], + "type": "text", + "content": " is the matrix norm " + }, + { + "bbox": [ + 55, + 441, + 296, + 514 + ], + "type": "inline_equation", + "content": "(p = 2)" + }, + { + "bbox": [ + 55, + 441, + 296, + 514 + ], + "type": "text", + "content": " or spectral norm of the weight matrix. The detailed proof is provided in the Appendix. Theorem 1 indicates that the conversion error is bounded by the weighted sum of errors across all layers. Based on this result, we introduce the local threshold balancing algorithm in the following section." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 521, + 257, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 257, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 257, + 534 + ], + "type": "text", + "content": "4.2. Local Threshold Balancing Algorithm" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 538, + 297, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 297, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 297, + 585 + ], + "type": "text", + "content": "The target of ANN-SNN conversion is to minimize the conversion error and achieve high-performance SNNs. An alternative approach is to optimize the error bound, defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 593, + 296, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 593, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 64, + 593, + 296, + 646 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname * {a r g m i n} _ {\\boldsymbol {\\theta}} \\mathbb {E} _ {x ^ {0} \\in \\mathcal {D}} \\\\ \\left[ \\sum_ {l = 1} ^ {L} \\left(\\prod_ {k = l + 1} ^ {L} \\| \\boldsymbol {w} ^ {k} \\| _ {2}\\right) \\left\\| \\mathrm {S} \\left(\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}\\right) - \\mathrm {A} \\left(\\hat {\\boldsymbol {z}} ^ {l}\\right) \\right\\| _ {2} \\right]. \\tag {10} \\\\ \\end{array}", + "image_path": "337c411c70240250636f20dda3735b6a8f882a25effc2f063bc7e95d5b087c4d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "content": " represents the set of threshold values to be optimized across all layers, i.e., " + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} = \\{\\theta^1,\\theta^2,\\dots,\\theta^L\\}" + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "inline_equation", + "content": "x^{0}" + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "content": " denotes the input data sample drawn from the dataset " + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "content": ". This approach leverages data-driven conversion, minimizing the expectation of the conversion error bound over the data distribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "To reduce computational costs during optimization, we introduce two key simplifications. Firstly, we use a greedy strategy to optimize the threshold " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": " layer by layer, that is" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 115, + 457, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 115, + 457, + 133 + ], + "spans": [ + { + "bbox": [ + 323, + 115, + 457, + 133 + ], + "type": "text", + "content": "For each layer " + }, + { + "bbox": [ + 323, + 115, + 457, + 133 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 323, + 115, + 457, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 323, + 115, + 457, + 133 + ], + "type": "inline_equation", + "content": "\\arg \\min_{\\theta^l} \\mathbb{E}_{x^0 \\in \\mathcal{D}}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 356, + 135, + 555, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 135, + 555, + 169 + ], + "spans": [ + { + "bbox": [ + 356, + 135, + 555, + 169 + ], + "type": "interline_equation", + "content": "\\left[ \\left(\\prod_ {k = l + 1} ^ {L} \\| \\boldsymbol {w} ^ {k} \\| _ {2}\\right) \\left\\| \\mathbf {S} \\left(\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}\\right) - \\mathbf {A} \\left(\\hat {\\boldsymbol {z}} ^ {l}\\right) \\right\\| _ {2} \\right]. \\tag {11}", + "image_path": "b0e4e65b87202ccb4b8f3fac3219265a2708bf76f9c7d7760fcb63c652fd746d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 175, + 555, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 283 + ], + "type": "text", + "content": "Secondly, we simplify the function " + }, + { + "bbox": [ + 313, + 175, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathrm{S}(\\cdot; \\theta)" + }, + { + "bbox": [ + 313, + 175, + 555, + 283 + ], + "type": "text", + "content": " by neglecting its temporal dynamics and approximating it with the clipping function " + }, + { + "bbox": [ + 313, + 175, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathrm{C}(\\cdot; \\theta)" + }, + { + "bbox": [ + 313, + 175, + 555, + 283 + ], + "type": "text", + "content": ", as previous works have demonstrated that spiking neuron and clipping functions behave equivalently given sufficient number of time-steps [12]. The average input in the SNN can therefore be estimated as " + }, + { + "bbox": [ + 313, + 175, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{z}}^0 = \\boldsymbol{z}^0" + }, + { + "bbox": [ + 313, + 175, + 555, + 283 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 175, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{z}}^{l+1} = \\boldsymbol{w}^{l+1} \\mathrm{C}(\\hat{\\boldsymbol{z}}^l; \\theta^l)" + }, + { + "bbox": [ + 313, + 175, + 555, + 283 + ], + "type": "text", + "content": ". Furthermore, We utilize the squared norm in the objective function for smoother optimization. The optimization problem then becomes:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 323, + 290, + 457, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 290, + 457, + 308 + ], + "spans": [ + { + "bbox": [ + 323, + 290, + 457, + 308 + ], + "type": "text", + "content": "For each layer " + }, + { + "bbox": [ + 323, + 290, + 457, + 308 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 323, + 290, + 457, + 308 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 323, + 290, + 457, + 308 + ], + "type": "inline_equation", + "content": "\\arg \\min_{\\theta^l} \\mathbb{E}_{x^0 \\in \\mathcal{D}}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 356, + 310, + 555, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 310, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 356, + 310, + 555, + 344 + ], + "type": "interline_equation", + "content": "\\left[ \\left(\\prod_ {k = l + 1} ^ {L} \\| \\boldsymbol {w} ^ {k} \\| _ {2}\\right) \\| \\mathbf {C} (\\hat {\\boldsymbol {z}} ^ {l}; \\theta^ {l}) - \\mathbf {A} (\\hat {\\boldsymbol {z}} ^ {l}) \\| _ {2} ^ {2} \\right], \\tag {12}", + "image_path": "5f83dfe417879f2ecd7747ddf7424073e80dbd91e9696fdeb6ce2e2facfd3bfa.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 350, + 555, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 350, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 359, + 350, + 555, + 364 + ], + "type": "interline_equation", + "content": "\\operatorname {w h e r e} \\mathrm {C} (x; \\theta) = \\min (\\max (0, x), \\theta). \\tag {13}", + "image_path": "a466dd775a9be426da0d9fe734f189eb97369ee8820d518a2b6d7a7abbda5117.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 368, + 554, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 554, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 554, + 439 + ], + "type": "text", + "content": "Intuitively, we can employ the stochastic gradient descent algorithm to optimize the threshold. Since " + }, + { + "bbox": [ + 313, + 368, + 554, + 439 + ], + "type": "inline_equation", + "content": "\\prod_{k = l + 1}^{L}\\| \\pmb{w}^{k}\\|_{2}" + }, + { + "bbox": [ + 313, + 368, + 554, + 439 + ], + "type": "text", + "content": " can be considered as a constant when weights are fixed for each layer, this term can be absorbed into the learning rate. The final update rule for the local threshold balancing algorithm at each step is:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 359, + 446, + 555, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 446, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 359, + 446, + 555, + 478 + ], + "type": "interline_equation", + "content": "\\Delta \\theta^ {l} = - \\sum_ {i = 0} ^ {N - 1} 2 \\left(\\hat {z} _ {i} ^ {l} - \\theta^ {l}\\right) H \\left(\\hat {z} _ {i} ^ {l} - \\theta^ {l}\\right), \\tag {14}", + "image_path": "d96ca597bd5ca7fde1d5bc4ad79ae0acc2eb52a7591bfc40290de6554985346a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 367, + 480, + 555, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 480, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 367, + 480, + 555, + 495 + ], + "type": "interline_equation", + "content": "\\theta^ {l} \\leftarrow \\theta^ {l} - \\eta \\Delta \\theta^ {l}. \\tag {15}", + "image_path": "c6fd5e8a6928e946dd857bc876b209d8201e091001b660ec4d7e347433d6de61.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "text", + "content": "Detailed derivation is provided in the Appendix. Here " + }, + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "inline_equation", + "content": "\\Delta \\theta^l" + }, + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "text", + "content": " denotes the step size during optimization and " + }, + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "text", + "content": " is the learning rate or update step size. " + }, + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "text", + "content": " is the number of elements in the input vector, and " + }, + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "inline_equation", + "content": "\\hat{z}_i" + }, + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "text", + "content": " denotes each element in the vector " + }, + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "inline_equation", + "content": "\\hat{z}^l" + }, + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 313, + 502, + 555, + 563 + ], + "type": "text", + "content": " is the Heaviside step function." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 563, + 556, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 556, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 556, + 717 + ], + "type": "text", + "content": "In practice, as shown in Figure 2, we jointly optimize the threshold for each layer by sampling a small set of image samples from the training dataset. Before starting the conversion process, we first replace all ReLU activation functions " + }, + { + "bbox": [ + 313, + 563, + 556, + 717 + ], + "type": "inline_equation", + "content": "\\mathrm{A}(\\cdot)" + }, + { + "bbox": [ + 313, + 563, + 556, + 717 + ], + "type": "text", + "content": " with clipping functions " + }, + { + "bbox": [ + 313, + 563, + 556, + 717 + ], + "type": "inline_equation", + "content": "\\mathrm{C}(\\cdot)" + }, + { + "bbox": [ + 313, + 563, + 556, + 717 + ], + "type": "text", + "content": " and initialize the threshold to zero. We then randomly sample data batches from the training dataset and perform model inference with these data samples. The thresholds are locally updated during the inference until convergence. Notably, the parameter " + }, + { + "bbox": [ + 313, + 563, + 556, + 717 + ], + "type": "inline_equation", + "content": "\\theta^l" + }, + { + "bbox": [ + 313, + 563, + 556, + 717 + ], + "type": "text", + "content": " is updated locally, ensuring that the computational cost of each iteration is similar to the inference of the ANN model and the computational cost of the whole conversion process is at inference level." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24390" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 289, + 157 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 289, + 157 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 289, + 157 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 289, + 157 + ], + "type": "image", + "image_path": "58effa4ec822ff32ffdf4622e35ecaad2876a2fa3ab410222d5ecabd849e6eab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 165, + 296, + 212 + ], + "lines": [ + { + "bbox": [ + 55, + 165, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 296, + 212 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed conversion framework. It requires only a small subset of data for conversion using local threshold balancing. During inference, the delayed evaluation technique is employed to enhance the accuracy of output estimation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 226, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 226, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 226, + 242 + ], + "type": "text", + "content": "4.3. Pre-Neuron Max Pooling Layer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 246, + 297, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 246, + 297, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 246, + 297, + 413 + ], + "type": "text", + "content": "The conversion of the max pooling layer presents a significant challenge. As a pivotal feature in the process of down-sampling input representations, the max pooling layer is a fundamental component in convolutional neural networks and is widely used in most convolutional architectures. However, due to the binary nature of spiking neurons, max pooling layers cannot effectively perform downsampling in feature maps. Typically, in a single max-pooling operation, multiple elements may share the same value, preventing the extraction of the most salient features. Consequently, max pooling is often avoided as a downsampling layer in SNNs. In most previous ANN-SNN methods, the max pooling layer in ANN is often replaced by an average pooling layer, and the model is re-trained or fine-tuned before conversion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 485 + ], + "type": "text", + "content": "To convert architectures that contain max pooling layers, we propose a simple yet effective method. We replace all max-pooling layers with neuron layers, allowing the preneuron floating-point input to pass through the downsampling layer before being input into the neurons. We have the following theorem:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 491, + 296, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 516 + ], + "type": "text", + "content": "Theorem 2 The order of the max pooling layer and ReLU layer does not affect the ANN output results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 522, + 296, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 522, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 75, + 522, + 296, + 540 + ], + "type": "interline_equation", + "content": "\\max _ {i} \\mathrm {R} (z) = \\mathrm {R} (\\max _ {i} z), \\text {w h e n} \\max (\\mathrm {z}) > 0. \\tag {16}", + "image_path": "8c1fe6ceffd8d76fb3d5147740d0d44501e70aa6c4bd3eb7cc79d85b83669523.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 297, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 629 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 545, + 297, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{R}(\\cdot)" + }, + { + "bbox": [ + 55, + 545, + 297, + 629 + ], + "type": "text", + "content": " denotes the ReLU activation. The detailed proof is provided in the Appendix. Theorem 2 guarantees that such an operation does not affect the final performance. When deployed on neuromorphic chips or FPGAs, the max pooling operation can be efficiently implemented through a comparator, avoiding additional floating-point multiplications and minimizing power consumption." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 636, + 243, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 243, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 243, + 649 + ], + "type": "text", + "content": "4.4. Channel-wise Threshold Balancing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "text", + "content": "During conversion, following Kim et al. [35], Li et al. [36], we employ a channel-wise local threshold balancing method applicable to both fully connected and convolutional layers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "In this approach, neurons in each channel share a single threshold for convolutional layers while each neuron has an" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": "independent threshold for fully connected layers. Specifically, for convolutional layers, we channel-wisely determine the threshold value, while for fully connected layers, we optimize the threshold element-wise. The update function is given by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 352, + 140, + 555, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 140, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 352, + 140, + 555, + 174 + ], + "type": "interline_equation", + "content": "\\Delta \\theta_ {c} ^ {l} = - \\sum_ {i = 0} ^ {N - 1} 2 \\left(\\hat {z} _ {c, i} ^ {l} - \\theta_ {c} ^ {l}\\right) H \\left(\\hat {z} _ {c, i} ^ {l} - \\theta_ {c} ^ {l}\\right), \\tag {17}", + "image_path": "7c3beb452f04b2f8ca1b29336cbe7e23ac2bc5b091340d4ba8667f7897ec8b80.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 181, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 216 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 181, + 553, + 216 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 181, + 553, + 216 + ], + "type": "text", + "content": " denotes the channel index (or element index for fully connected layers) and " + }, + { + "bbox": [ + 313, + 181, + 553, + 216 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 181, + 553, + 216 + ], + "type": "text", + "content": " denotes the element index within the current channel." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 217, + 555, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 313 + ], + "type": "text", + "content": "After optimization, the thresholds for each channel can be individually absorbed into the corresponding channel of the convolutional kernel or the weight vectors of the fully connected layers. This channel-wise (or element-wise) operation ensures that the behavior of the IF neuron remains unaffected. This algorithm aims to more precisely determine the optimal threshold value while preserving the fundamental properties of spiking neurons." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 319, + 471, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 471, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 471, + 333 + ], + "type": "text", + "content": "4.5. Delayed Evaluation Strategy" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 337, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 433 + ], + "type": "text", + "content": "As introduced in [4], the first spike from the last layer often takes longer than expected to appear due to propagation delays between layers. This delay creates significant fluctuations in the firing rate during the initial time-steps, as no spikes are generated until the first one reaches each layer [29]. To address this issue, we propose an effective delayed step selection strategy building upon the delayed evaluation technique [29]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "type": "text", + "content": "As shown in Figure 2, the basic idea of the delayed evaluation is to compute the average spike count in the last layer from time-step " + }, + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "type": "text", + "content": " denotes the delayed step, rather than averaging over all time-steps. Since the output is collected from the average spike count in the interval " + }, + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "type": "inline_equation", + "content": "[t_0, T]" + }, + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "type": "text", + "content": " at the last layer, we can rewrite Equation 6 for the last layer as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 324, + 522, + 555, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 522, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 324, + 522, + 555, + 550 + ], + "type": "interline_equation", + "content": "\\boldsymbol {r} ^ {L} \\left(t _ {0}, T\\right) = \\boldsymbol {w} ^ {L} \\boldsymbol {r} ^ {L - 1} \\left(t _ {0}, T\\right) - \\frac {\\boldsymbol {v} ^ {L} (T) - \\boldsymbol {v} ^ {L} \\left(t _ {0}\\right)}{T - t _ {0}}, \\tag {18}", + "image_path": "0862157f34123e0bb41212b69846f69e1e6e2b1affc64390b98fb7c18fac03e0.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 557, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 618 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 557, + 555, + 618 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}^l(t_0, t) = \\sum_{i=t_0}^{t} \\boldsymbol{s}^l(i) \\theta^l / (t - t_0)" + }, + { + "bbox": [ + 313, + 557, + 555, + 618 + ], + "type": "text", + "content": " denotes the average postsynaptic potential from time " + }, + { + "bbox": [ + 313, + 557, + 555, + 618 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 313, + 557, + 555, + 618 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 557, + 555, + 618 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 557, + 555, + 618 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 557, + 555, + 618 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 557, + 555, + 618 + ], + "type": "text", + "content": " represents the last layer index. When " + }, + { + "bbox": [ + 313, + 557, + 555, + 618 + ], + "type": "inline_equation", + "content": "T \\gg t_0" + }, + { + "bbox": [ + 313, + 557, + 555, + 618 + ], + "type": "text", + "content": ", one can still guarantee that the basic equivalence of the average postsynaptic potential and ANN activations holds." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "text", + "content": "In our strategy, we do not start counting the average output spike until " + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\widetilde{t_0}" + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\widetilde{t_0}" + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "text", + "content": " is the expected spike timing of the first spike in the last layer. We estimate " + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\widetilde{t_0}" + }, + { + "bbox": [ + 313, + 619, + 555, + 666 + ], + "type": "text", + "content": " based on the converted SNN model, denoted as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 328, + 673, + 555, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 673, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 328, + 673, + 555, + 712 + ], + "type": "interline_equation", + "content": "\\widetilde {t _ {0}} = \\sum_ {l = 1} ^ {L} \\frac {\\theta^ {l} - v ^ {l} (0)}{\\max _ {i} \\left(\\overline {{\\mathbf {R} (\\hat {\\boldsymbol {z}} ^ {l})}}\\right)} \\approx \\sum_ {l = 1} ^ {L} \\frac {\\theta^ {l} - v ^ {l} (0)}{\\max _ {i} \\left(\\overline {{\\mathbf {R} (\\boldsymbol {z} ^ {l})}}\\right)}. \\tag {19}", + "image_path": "9cc5051427e0257529cbf56f03b94d997690937798d422b971a85b2b23321e90.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24391" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\max_i(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " selects the maximum element of a given vector. The notation " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\overline{(\\cdot)} = \\mathbb{E}_{x^0\\in \\mathcal{D}}[(\\cdot)]" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " represents the expectation of the given input over dataset. Additionally, we use the ReLU function " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " to prevent negative estimated time. Before evaluation, we first calculate " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\widetilde{t_0}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " according to Equation 19 and consider it as an indicator for the delayed evaluation strategy. In practice, if the inference step " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " is less than " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\widetilde{t_0} +4" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "t_0 = T - 4" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " and collect the average output over the time interval " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "[T - 4,T]" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " exceeds " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\widetilde{t_0} +4" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", we use the interval " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "[\\widetilde{t_0},T]" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " for output estimation. The value 4 is empirically chosen to ensure a sufficient delay while allowing adequate time for averaging spikes over time." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 296, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 265 + ], + "type": "text", + "content": "This technique introduces minimal additional training cost and does not alter the model's inference process, making it seamlessly integrated into the proposed conversion framework without disrupting the overall pipeline." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 274, + 136, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 136, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 136, + 287 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 294, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 294, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 296, + 449 + ], + "type": "text", + "content": "We conduct extensive experiments to demonstrate the effectiveness of our method and highlight its potential practical value. We first showcase the advantages of our approach by integrating it with existing conversion frameworks, comparing it with state-of-the-art post-training conversion algorithms, and conducting ablation experiments to validate the effectiveness of the whole conversion pipeline. Subsequently, we apply our algorithm to various visual tasks, including image classification, semantic segmentation, object detection, and video classification. The feasibility of our algorithm is highlighted on both classification and regression tasks, showcasing its generalizability across different datasets and its superiority over existing conversion algorithms." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 547 + ], + "type": "text", + "content": "To underscore that our conversion method does not require additional training, we utilize open-source pre-trained ANN models for conversion, with most pre-trained sourced from TorchVision [45]. Additionally, we incorporate the membrane potential initialization algorithm [5], setting the membrane potential to half of the threshold before inference. The pseudocode for our conversion pipeline and detailed training settings are provided in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 295, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 577 + ], + "type": "text", + "content": "5.1. Plug-and-Play Integration for Enhanced Performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "In this section, we aim to highlight that our proposed conversion method provides a lightweight, plug-and-play solution that can be seamlessly integrated with other conversion algorithms to enhance performance. We select two representative ANN-SNN conversion approaches from existing work: the post-training fine-tuning (calibration) conversion method LCP/ACP [36, 37], and the training-required conversion method QCFS [5]. We reproduce the conversion pipeline of both methods, incorporate a local threshold balancing method for threshold searching, and use our delayed evaluation strategy. The evaluation is conducted on the ImageNet" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 330, + 70, + 539, + 163 + ], + "blocks": [ + { + "bbox": [ + 330, + 70, + 539, + 163 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 539, + 163 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 539, + 163 + ], + "type": "table", + "html": "
MethodANNCFInference Step
3264128256512
LCP75.66×50.2163.6668.8972.1274.17
✓53.8067.1872.1874.0774.56
ACP75.66×64.5471.1273.4574.6175.35
✓65.9572.8774.2174.7375.26
QCFS73.30×64.4371.5173.2573.6773.76
✓71.9473.0673.5073.4973.54
", + "image_path": "7aa86646071b01c52c865c0ce523f18b998c358beddcc222bfed2c465150ccb8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 174, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 313, + 174, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 553, + 196 + ], + "type": "text", + "content": "Table 1. Integration of the proposed conversion approach with existing framework" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 205, + 426, + 293 + ], + "blocks": [ + { + "bbox": [ + 317, + 205, + 426, + 293 + ], + "lines": [ + { + "bbox": [ + 317, + 205, + 426, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 205, + 426, + 293 + ], + "type": "image", + "image_path": "905acf9bacea199407f858162858455ebd244102c3cb873cfe60c8f74598e1f4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 296, + 402, + 306 + ], + "lines": [ + { + "bbox": [ + 341, + 296, + 402, + 306 + ], + "spans": [ + { + "bbox": [ + 341, + 296, + 402, + 306 + ], + "type": "text", + "content": "(a) Ablation Study" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 441, + 205, + 549, + 293 + ], + "blocks": [ + { + "bbox": [ + 441, + 205, + 549, + 293 + ], + "lines": [ + { + "bbox": [ + 441, + 205, + 549, + 293 + ], + "spans": [ + { + "bbox": [ + 441, + 205, + 549, + 293 + ], + "type": "image", + "image_path": "260936a879e100bc21238ee7869cc84fce9a1a415840a38493c06964c355a774.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 296, + 539, + 306 + ], + "lines": [ + { + "bbox": [ + 451, + 296, + 539, + 306 + ], + "spans": [ + { + "bbox": [ + 451, + 296, + 539, + 306 + ], + "type": "text", + "content": "(b) Effect of Iteration Steps" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 311, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 344 + ], + "type": "text", + "content": "Figure 3. (a) Results on evaluating the impact of different techniques within the conversion framework. (b) Effect of iteration steps on SNN performance after conversion." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 361, + 477, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 477, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 477, + 373 + ], + "type": "text", + "content": "dataset with the ResNet-34 architecture." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 374, + 556, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 556, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 556, + 506 + ], + "type": "text", + "content": "The results, as detailed in Table 1, showcase the performance of various configurations. The column labeled \"CF\" indicates whether the proposed method is used. Notably, the composite approaches, which combine our method with either LCP/ACP or QCFS, consistently outperform the individual conversion method across most timesteps. After incorporating our methods, both QCFS and LCP exhibit performance improvements exceeding " + }, + { + "bbox": [ + 313, + 374, + 556, + 506 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 313, + 374, + 556, + 506 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 374, + 556, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 32" + }, + { + "bbox": [ + 313, + 374, + 556, + 506 + ], + "type": "text", + "content": ", while maintaining consistent performance as " + }, + { + "bbox": [ + 313, + 374, + 556, + 506 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 374, + 556, + 506 + ], + "type": "text", + "content": " increases. This demonstrates the advantages of using our approach in conjunction with existing conversion techniques." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 516, + 553, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 553, + 528 + ], + "type": "text", + "content": "5.2. Ablation Study for the Conversion Framework" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "content": "We demonstrate the effectiveness of our proposed conversion framework by using the ResNet-34 architecture on the ImageNet dataset. An ablation study is conducted to evaluate the impact of different techniques within the framework. Specifically, as shown in Figure 3a, we compare the performance of four different SNNs obtained by combinations of different techniques. The baseline SNN is obtained using the Robust-Norm [54] algorithm (RN, red curve), which is a commonly used post-training ANN-SNN conversion method that sets the 99-th percentile activation value as the threshold. The other SNNs are obtained by using the layer-wise local threshold balancing algorithm (LTB, yellow curve), the channel-wise threshold balancing algorithm (CLTB, blue curve), and a combination of channel-wise threshold balancing and delayed evaluation (CLTB+DE, black curve)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24392" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 289, + 153 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 289, + 153 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 289, + 153 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 289, + 153 + ], + "type": "image", + "image_path": "56d476b287f1f33b63b175734fd4e1b03b99d7d037d1c3c5ed10ae35c25b3c9d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 159, + 236, + 171 + ], + "lines": [ + { + "bbox": [ + 115, + 159, + 236, + 171 + ], + "spans": [ + { + "bbox": [ + 115, + 159, + 236, + 171 + ], + "type": "text", + "content": "Figure 4. Effect of delayed steps" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 297, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 297, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 297, + 370 + ], + "type": "text", + "content": "Figure 3a presents the accuracy changes of the converted SNNs with respect to the increasing time-steps. The performance of the SNN obtained from the full conversion pipeline (black curve) consistently outperforms the other SNNs across all time-steps, demonstrating an excellent balance between fast inference and high performance of the proposed conversion framework. This is particularly evident when combined with the proposed delayed evaluation technique. Compared to the robust norm method (red curve), the peak performance of the SNNs obtained by the proposed local threshold balancing method (black, blue and yellow curves) more closely aligns with the original ANN accuracy. Furthermore, the use of the channel-wise threshold balancing (CLTB, blue curve) results in a noticeable performance improvement over the vanilla threshold balancing (LTB, yellow curve) algorithm." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 377, + 203, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 203, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 203, + 390 + ], + "type": "text", + "content": "5.3. Effect of the Iteration Step" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 395, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 647 + ], + "type": "text", + "content": "We evaluate the influence of the iteration steps of the local threshold balancing algorithm. The iteration number of the local threshold balancing serves as the hyper-parameter of the method. We focus on the most straightforward approach by empirically evaluating the impact of different iteration steps. We converted five different SNNs from the pre-trained ImageNet ResNet-34 model, varying the iterations steps from 1000 to 5000. The batch size during local threshold balancing is consistently set to 100. Figure 3b shows the final accuracy of the different converted SNNs. As number of iterations increases, the peak performance of the converted SNNs at different inference steps approaches the accuracy of the original ANN. Furthermore, these SNNs demonstrate better performance at low time-steps when fewer iteration steps are used. On the ImageNet dataset, with only 1000 iterations of local threshold update, the performance gap between the SNN at 512 inference time-steps and the ANN is approximately " + }, + { + "bbox": [ + 55, + 395, + 296, + 647 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 395, + 296, + 647 + ], + "type": "text", + "content": ". With 5000 iterations, this difference is further reduced to around " + }, + { + "bbox": [ + 55, + 395, + 296, + 647 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 55, + 395, + 296, + 647 + ], + "type": "text", + "content": ". Notably, the accuracy surpasses " + }, + { + "bbox": [ + 55, + 395, + 296, + 647 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 55, + 395, + 296, + 647 + ], + "type": "text", + "content": " within just 55 time-steps, achieving a balance between inference time and performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 200, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 200, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 200, + 666 + ], + "type": "text", + "content": "5.4. Effect of the Delayed Step" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "We conduct experiments to demonstrate the effectiveness of the proposed delayed strategy on the ImageNet dataset using the ResNet-34 architecture. As shown in Figure 4, we evaluate the impact of the delayed step across four different" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 321, + 72, + 549, + 232 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 549, + 232 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 549, + 232 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 549, + 232 + ], + "type": "table", + "html": "
ArchitectureMethodANNInference Step
3264128256≥512
Post-Training
ResNet-341RMP70.64---55.6560.08
ResNet-342LCP75.6655.1667.5672.4874.5375.44
MobileNetLCP73.400.2015.4755.9566.3572.19
MobileNetV2Ours71.8819.3949.9364.3168.8970.48
ResNet-183Ours69.7665.5568.5369.0369.2169.21
ResNet-343Ours73.3164.8871.1772.3072.4272.46
Training-involved
ResNet-343QCFS73.3064.4371.5173.2573.6773.76
ResNet-342ACP75.6664.6571.3073.9475.0075.45
ResNet-343Ours*73.3071.9473.0673.5073.4973.54
", + "image_path": "5f417b2e54911da5c7f1ca199577b8754b8f43f2f38c6d2066efd33ad9f80a4a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 328, + 232, + 533, + 241 + ], + "lines": [ + { + "bbox": [ + 328, + 232, + 533, + 241 + ], + "spans": [ + { + "bbox": [ + 328, + 232, + 533, + 241 + ], + "type": "text", + "content": "1 Standard ResNet-34 architecture without batch normalization." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 328, + 242, + 547, + 259 + ], + "lines": [ + { + "bbox": [ + 328, + 242, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 328, + 242, + 547, + 259 + ], + "type": "text", + "content": "2 Two more layers in the first conv-block compared to standard ResNet-34." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 328, + 261, + 451, + 270 + ], + "lines": [ + { + "bbox": [ + 328, + 261, + 451, + 270 + ], + "spans": [ + { + "bbox": [ + 328, + 261, + 451, + 270 + ], + "type": "text", + "content": "3 Standard ResNet-18/34 architecture." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 277, + 553, + 300 + ], + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 300 + ], + "type": "text", + "content": "Table 2. Comparison of the proposed method and previous works on the ImageNet dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 316, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 388 + ], + "type": "text", + "content": "inference time-step settings: 32 steps (red curve), 64 steps (yellow curve), 128 steps (blue curve), and 256 steps (black curve). The grey vertical line indicates the estimated " + }, + { + "bbox": [ + 313, + 316, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\widetilde{t_0}" + }, + { + "bbox": [ + 313, + 316, + 554, + 388 + ], + "type": "text", + "content": " for the currently converted SNN. For each sub-figure, the x-axis represents the number of delayed steps, while the y-axis denotes the accuracy." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "text", + "content": "Although the delayed strategy is pre-determined before inference, the estimated " + }, + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "inline_equation", + "content": "\\widetilde{t_0}" + }, + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "text", + "content": " can still serves as an indicator that provides a near-optimal strategy to select the delayed steps. When the total inference step is less than the estimated " + }, + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "inline_equation", + "content": "\\widetilde{t_0}" + }, + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "text", + "content": " (red and yellow curves), the final performance gradually improves with the increase in the delayed steps, but drops significantly when the delayed step approaches the inference step. In the situation that the total inference step exceeds " + }, + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "inline_equation", + "content": "\\widetilde{t_0}" + }, + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "text", + "content": " (blue and black curve), the performance first increases then fluctuates when delayed step is larger than " + }, + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "inline_equation", + "content": "\\widetilde{t_0}" + }, + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "text", + "content": ", and ultimately experiences a sharp decline as the delayed steps get close to the inference step. The above results indicate that we can always reach near-optimal performance with the proposed delayed evaluation strategy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 564, + 491, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 491, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 491, + 577 + ], + "type": "text", + "content": "5.5. Test on Image Classification task" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "We evaluate our conversion method on the classification task with the ImageNet dataset, employing different architectures including ResNet and MobileNet. To highlight the advantages of the proposed conversion algorithm, we conduct comparisons with previous post-training algorithms, showcasing its superior performance. When using the ResNet-34 architecture, the SNN converted by our algorithm exceeds " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "71\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " accuracy within 64 time steps. In contrast, the SNNs converted by the previous algorithm [20] and the calibrationrequired conversion algorithm [37] require longer time-steps to achieve comparable accuracy. We also implement the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24393" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 191 + ], + "type": "text", + "content": "proposed conversion framework on QCFS pre-trained models (denote as \"Our*\" in Table 2), and compare with the sota training-involved ANN-SNN techniques. The accuracy of the proposed method reaches " + }, + { + "bbox": [ + 55, + 72, + 295, + 191 + ], + "type": "inline_equation", + "content": "71.94\\%" + }, + { + "bbox": [ + 55, + 72, + 295, + 191 + ], + "type": "text", + "content": " within 32 steps, demonstrating significant advantages compared to the fine-tuning required advanced calibration technique ACP [37] and training-involved vanilla QCFS [5]. The experiments on the ImageNet dataset demonstrate the superior performance of the proposed conversion algorithm and highlight its potential for large-scale applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 193, + 296, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 313 + ], + "type": "text", + "content": "In addition, we demonstrate the energy-saving advantage of our method by theoretically estimating the energy efficiency of our converted SNN. We use the same energy estimation approach as Bu et al. [5]. With ResNet-34 architecture on ImageNet, SNN converted from our algorithm can achieve energy efficiency at 622FPS/W while maintaining " + }, + { + "bbox": [ + 55, + 193, + 296, + 313 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 193, + 296, + 313 + ], + "type": "text", + "content": " performance. In comparison, the energy efficiency of the original ANN is only 22FPS/W, which is 28 times lower than that of the converted SNN. For detailed energy consumption analysis, please refer to the Appendix." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 323, + 295, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 295, + 347 + ], + "type": "text", + "content": "5.6. Test on Semantic Segmentation, Object Detection, and Video Classification Tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 569 + ], + "type": "text", + "content": "We further extend the proposed ANN-SNN conversion method to semantic segmentation, object detection, and video classification tasks. For the semantic segmentation task, we evaluate our method on two commonly used datasets: Pascal VOC 2012 [16] augmented by SBD [24] and MS COCO 2017 [39]. We employ various semantic segmentation models, including FCN [43] and DeepLabV3 [8], provided by TorchVision [45] and other open-source repositories. As shown in Table 3, our method is compatible with tackling semantic segmentation tasks without any preparatory training. Specifically, we achieved a " + }, + { + "bbox": [ + 55, + 354, + 295, + 569 + ], + "type": "inline_equation", + "content": "53.50\\%" + }, + { + "bbox": [ + 55, + 354, + 295, + 569 + ], + "type": "text", + "content": " mIoU with the FCN model using a ResNet-34 backbone in 64 timesteps on the Pascal VOC dataset and a " + }, + { + "bbox": [ + 55, + 354, + 295, + 569 + ], + "type": "inline_equation", + "content": "61.52\\%" + }, + { + "bbox": [ + 55, + 354, + 295, + 569 + ], + "type": "text", + "content": " mIoU with the DeepLabV3 model using a ResNet-50 architecture in 128 timesteps on the more complex MS COCO 2017 dataset. Given that pixel-level classification tasks require precise model output, this excellent performance highlights the effectiveness and generalizability of the proposed method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "For object detection tasks, we use the benchmark dataset MS COCO 2017 [39] with open-source models from TorchVision [45]. We employ the fully convolutional object detection method RetinaNet [40] and FCOS [60], and convert the pre-trained ANN models to SNNs using the proposed method. Table 3 shows detailed performance under different inference time-steps. Within 64 time-steps, the converted SNN can achieve " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "27.3\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " mAP with RetinaNet and " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "26.5\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " mAP with FCOS, which represents a significant improvement compared to previous explorations [35, 47]. The previous method required over 1000 time-steps to achieve comparable performance as ANN, which is detrimental to" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 321, + 71, + 547, + 245 + ], + "blocks": [ + { + "bbox": [ + 321, + 71, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 321, + 71, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 547, + 245 + ], + "type": "table", + "html": "
DatasetArch.ANNInference Step
3264128256
Semantic Segmentation (mIoU%)
PascalVOCFCN-1847.3043.4947.0247.2647.23
FCN-3454.9946.6053.5054.7554.76
DeepLab-1855.7548.5953.9155.4055.60
DeepLab-3458.9541.2055.7458.3658.68
MS COCOFCN-5060.6728.6650.3857.0159.53
DeepLab-5063.0139.0256.6261.5262.61
Object detection (mAP%)
MS COCORetinaNet-5036.416.627.332.234.3
FCOS-5039.212.226.532.133.6
Video Classification (Acc%)
S-Kinetics-400MC3-1863.5255.2161.6062.4362.73
R3D-1862.5653.4660.3161.1261.15
", + "image_path": "8aae9cddbf7cf2a218a85397141cb3194520dfd0f43667c49c6a5f24486bffa7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 255, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 277 + ], + "type": "text", + "content": "Table 3. Performance on semantic segmentation (mIoU%), object detection (mAP%), and video classification (Acc%) tasks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 293, + 482, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 482, + 305 + ], + "type": "text", + "content": "real-time detection and energy efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 305, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 555, + 425 + ], + "type": "text", + "content": "For the video classification tasks, we use a small Kinetics-400 dataset [33], which is a subset of the original dataset. More specifically, we split the original validation set into two independent sets, one for our threshold balancing procedure, accounting for about " + }, + { + "bbox": [ + 313, + 305, + 555, + 425 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 305, + 555, + 425 + ], + "type": "text", + "content": " of the original validation set, and the other for the test. As a result, the new train set contains 12000 videos and the new test set contains 7881 videos. The converted MC3-18 [61] SNN model achieves " + }, + { + "bbox": [ + 313, + 305, + 555, + 425 + ], + "type": "inline_equation", + "content": "61.60\\%" + }, + { + "bbox": [ + 313, + 305, + 555, + 425 + ], + "type": "text", + "content": " accuracy within 64 time-steps, showing great potential of our method in more complex computer vision tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 436, + 466, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 466, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 466, + 448 + ], + "type": "text", + "content": "6. Conclusion and Limitation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 455, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 556, + 635 + ], + "type": "text", + "content": "This paper introduces the concept of a lightweight posttraining ANN-SNN conversion algorithm with inference-scale computational cost, which can directly convert pretrained ANN models into SNNs without GPU-based training. The significant reduction in training overhead highlights the potential of the rapid deployment of large-scale SNNs in various scenarios and makes the instant on-chip conversion from pre-trained ANN to SNN possible. However, there is a tradeoff between performance and training cost, with the overall performance of the proposed post-training framework being slightly weaker compared to other training-involved SNN learning methods. Additionally, while this work focuses on the conversion of convolutional SNNs, future efforts will aim to extend such a conversion approach to transformer architectures and other complex layers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 646, + 424, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 424, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 424, + 659 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "This work is funded by National Natural Science Foundation of China (62422601,62176003,62088102), Beijing Municipal Science and Technology Program (Z241100004224004) and Beijing Nova Program (20230484362,20240484703)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24394" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 133 + ], + "type": "text", + "content": "[1] Sander M Bohte, Joost N Kok, and Johannes A La Poutre. Spikeprop: backpropagation for networks of spiking neurons. In European Symposium on Artificial Neural Networks, 2000. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] Velibor Bojkovic, Srinivas Anumasa, Giulia De Masi, Bin Gu, and Huan Xiong. Data driven threshold and potential initialization for spiking neural networks. In International Conference on Artificial Intelligence and Statistics, pages 4771-4779. PMLR, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 296, + 255 + ], + "type": "text", + "content": "[3] Romain Brette, Michelle Rudolph, Ted Carnevale, Michael Hines, David Beeman, James M Bower, Markus Diesmann, Abigail Morrison, Philip H Goodman, Frederick C Harris, et al. Simulation of networks of spiking neurons: a review of tools and strategies. Journal of Computational Neuroscience, 2007. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "type": "text", + "content": "[4] Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Optimized potential initialization for low-latency spiking neural networks. In AAAI Conference on Artificial Intelligence, 2022. 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 296, + 356 + ], + "type": "text", + "content": "[5] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks. In International Conference on Learning Representations, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 296, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 296, + 400 + ], + "type": "text", + "content": "[6] Yongqiang Cao, Yang Chen, and Deepak Khosla. Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision, 2015. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 402, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 296, + 434 + ], + "type": "text", + "content": "[7] Natalia Caporale and Yang Dan. Spike timing-dependent plasticity: a hebbian learning rule. Annual Review of Neuroscience, 31:25-46, 2008. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 296, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 296, + 478 + ], + "type": "text", + "content": "[8] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "type": "text", + "content": "[9] Yanqi Chen, Zhaofei Yu, Wei Fang, Zhengyu Ma, Tiejun Huang, and Yonghong Tian. State transition of dendritic spines improves learning of sparse spiking neural networks. In International Conference on Machine Learning, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 579 + ], + "type": "text", + "content": "[10] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 580, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 296, + 635 + ], + "type": "text", + "content": "[11] Michael V DeBole, Brian Taba, Arnon Amir, Filipp Akopyan, Alexander Andreopoulos, William P Risk, Jeff Kusnitz, Carlos Ortega Otero, Tapan K Nayak, Rathinakumar Appuswamy, et al. TrueNorth: Accelerating from zero to 64 million neurons in 10 years. Computer, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 678 + ], + "type": "text", + "content": "[12] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. In International Conference on Learning Representations, 2021. 2, 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[13] Peter U Diehl and Matthew Cook. Unsupervised learning of digit recognition using spike-timing-dependent plasticity. Frontiers in Computational Neuroscience, 9:99, 2015. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[14] Peter U Diehl, Daniel Neil, Jonathan Binas, Matthew Cook, Shih-Chii Liu, and Michael Pfeiffer. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In International Joint Conference on Neural Networks, 2015. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 173 + ], + "type": "text", + "content": "[15] Chaoteng Duan, Jianhao Ding, Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Temporal effective batch normalization in spiking neural networks. In Advances in Neural Information Processing Systems, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 175, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 555, + 217 + ], + "type": "text", + "content": "[16] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision, 2015. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 555, + 262 + ], + "type": "text", + "content": "[17] Biao Fang, Yuhao Zhang, Rui Yan, and Huajin Tang. Spike trains encoding optimization for spiking neural networks implementation in fpga. In International Conference on Advanced Computational Intelligence, 2020. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 264, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 555, + 319 + ], + "type": "text", + "content": "[18] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 320, + 554, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 554, + 363 + ], + "type": "text", + "content": "[19] Wulfram Gerstner, Werner M Kistler, Richard Naud, and Liam Paninski. Neuronal dynamics: From single neurons to networks and models of cognition. Cambridge University Press, 2014. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "type": "text", + "content": "[20] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, 2020. 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 555, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 555, + 443 + ], + "type": "text", + "content": "[21] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. RMP-SNN: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Computer Vision and Pattern Recognition, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 444, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 555, + 487 + ], + "type": "text", + "content": "[22] Zecheng Hao, Tong Bu, Jianhao Ding, Tiejun Huang, and Zhaofei Yu. Reducing ann-snn conversion error through residual membrane potential. arXiv preprint arXiv:2302.02091, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 489, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 555, + 533 + ], + "type": "text", + "content": "[23] Zecheng Hao, Jianhao Ding, Tong Bu, Tiejun Huang, and Zhaofei Yu. Bridging the gap between anns and snns by calibrating offset spikes. In International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 535, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 555, + 578 + ], + "type": "text", + "content": "[24] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In International Conference on Computer Vision, 2011. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 579, + 555, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 555, + 601 + ], + "type": "text", + "content": "[25] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology press, 2005. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 602, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 555, + 657 + ], + "type": "text", + "content": "[26] JiaKui Hu, Man Yao, Xuerui Qiu, Yuhong Chou, Yuxuan Cai, Ning Qiao, Yonghong Tian, Xu Bo, and Guoqi Li. High-performance temporal reversible spiking neural networks with O(1) training memory and O(1) inference cost. In International Conference on Machine Learning, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "text", + "content": "[27] Zihan Huang, Xinyu Shi, Zecheng Hao, Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Towards high-performance spiking transformers from ann to snn conversion. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10688-10697, 2024. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24395" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 95 + ], + "type": "text", + "content": "[28] Eric Hunsberger and Chris Eliasmith. Spiking deep networks with LIF neurons. arXiv preprint arXiv:1510.08829, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 150 + ], + "type": "text", + "content": "[29] Sungmin Hwang, Jeesoo Chang, Min-Hye Oh, Kyung Kyu Min, Taejin Jang, Kyungchul Park, Junsu Yu, Jong-Ho Lee, and Byung-Gook Park. Low-latency spiking neural networks using pre-charged membrane potential and delayed evaluation. Frontiers in Neuroscience, 2021. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 152, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 294, + 174 + ], + "type": "text", + "content": "[30] Eugene M Izhikevich. Which model to use for cortical spiking neurons? IEEE Transactions on Neural Networks, 2004. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 175, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 296, + 229 + ], + "type": "text", + "content": "[31] Haiyan Jiang, Srinivas Anumasa, Giulia De Masi, Huan Xiong, and Bin Gu. A unified optimization framework of ann-snn conversion: towards optimal mapping from activation values to firing rates. In International Conference on Machine Learning, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 231, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 231, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 231, + 296, + 274 + ], + "type": "text", + "content": "[32] Haiyan Jiang, Vincent Zoonekynd, Giulia De Masi, Bin Gu, and Huan Xiong. TAB: Temporal accumulated batch normalization in spiking neural networks. In International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 276, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 276, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 57, + 276, + 296, + 330 + ], + "type": "text", + "content": "[33] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 331, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 331, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 57, + 331, + 294, + 374 + ], + "type": "text", + "content": "[34] Saeed Reza Kheradpisheh, Mohammad Ganjtabesh, Simon J Thorpe, and Timothee Masquelier. STDP-based spiking deep convolutional neural networks for object recognition. Neural Networks, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 376, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 376, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 57, + 376, + 296, + 421 + ], + "type": "text", + "content": "[35] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-YOLO: Spiking neural network for energy-efficient object detection. In AAAI Conference on Artificial Intelligence, 2020. 3, 5, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 422, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 422, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 57, + 422, + 296, + 465 + ], + "type": "text", + "content": "[36] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ANN: Towards efficient, accurate spiking neural networks calibration. In International Conference on Machine Learning, 2021. 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 467, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 467, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 57, + 467, + 296, + 510 + ], + "type": "text", + "content": "[37] Yuhang Li, Shikuang Deng, Xin Dong, and Shi Gu. Error-aware conversion from ann to snn via post-training parameter calibration. International Journal of Computer Vision, 2024, 2, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 511, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 511, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 511, + 296, + 555 + ], + "type": "text", + "content": "[38] Yuhang Li, Tamar Geller, Youngeun Kim, and Priyadarshini Panda. SEENN: Towards temporal spiking early exit neural networks. Advances in Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 557, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 296, + 600 + ], + "type": "text", + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision, 2014. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 296, + 634 + ], + "type": "text", + "content": "[40] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In International Conference on Computer Vision, 2017. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "text", + "content": "[41] Zibin Liu, Banglei Guan, Yang Shang, Shunkun Liang, Zhenbao Yu, and Qifeng Yu. Optical flow-guided 6dof object pose tracking with an event camera. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6501-6509, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[42] Zibin Liu, Banglei Guan, Yang Shang, Qifeng Yu, and Laurent Kneip. Line-based 6-dof object pose estimation and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tracking with an event camera. IEEE Transactions on Image Processing, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "type": "text", + "content": "[43] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Computer Vision and Pattern Recognition, 2015. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 555, + 162 + ], + "type": "text", + "content": "[44] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural Networks, 1997. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 164, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 555, + 196 + ], + "type": "text", + "content": "[45] TorchVision maintainers and contributors. Torchvision: Pytorch's computer vision library. https://github.com/pytorch/vision, 2016.6,8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 198, + 554, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 554, + 231 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 554, + 231 + ], + "type": "text", + "content": "[46] Timothée Masquelier and Simon J Thorpe. Unsupervised learning of visual features through spike timing dependent plasticity. PLoS Computational Biology, 2007. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 232, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 555, + 285 + ], + "type": "text", + "content": "[47] Joaquín Royo Miquel, Silvia Tolu, Frederik ET Schöller, and Roberto Galeazzi. Retinanet object detector based on analog-to-spiking neural network conversion. In International Conference on Soft Computing & Machine Intelligence, 2021. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 288, + 555, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 555, + 332 + ], + "type": "text", + "content": "[48] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 333, + 555, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 555, + 365 + ], + "type": "text", + "content": "[49] Nicolas Perez Nieves and Dan F. M. Goodman. Sparse spiking gradient descent. In Advances in Neural Information Processing Systems, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 367, + 554, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 367, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 367, + 554, + 410 + ], + "type": "text", + "content": "[50] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjic chip architecture. Nature, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 412, + 554, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 554, + 456 + ], + "type": "text", + "content": "[51] Nitin Rathi and Kaushik Roy. DIET-SNN: A low-latency spiking neural network with direct input encoding and leakage and threshold optimization. IEEE Transactions on Neural Networks and Learning Systems, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 457, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 457, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 457, + 555, + 510 + ], + "type": "text", + "content": "[52] Nitin Rathi, Gopalakrishnan Srinivasan, Priyadarshini Panda, and Kaushik Roy. Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 513, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 555, + 567 + ], + "type": "text", + "content": "[53] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, Michael Pfeiffer, and Shih-Chii Liu. Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in Neuroscience, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 568, + 555, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 555, + 622 + ], + "type": "text", + "content": "[54] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, Michael Pfeiffer, and Shih-Chii Liu. Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in Neuroscience, 2017. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 667 + ], + "type": "text", + "content": "[55] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: VGG and residual architectures. Frontiers in Neuroscience, 2019. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[56] Jiangrong Shen, Qi Xu, Jian K. Liu, Yueming Wang, Gang Pan, and Huajin Tang. ESL-SNNs: An evolutionary structure learning strategy for spiking neural networks. In AAAI Conference on Artificial Intelligence, 2023. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24396" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 686 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "type": "text", + "content": "[57] Xinyu Shi, Zecheng Hao, and Zhaofei Yu. Spikingresformer: Bridging resnet and vision transformer in spiking neural networks. In Computer Vision and Pattern Recognition, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "text", + "content": "[58] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in Neural Information Processing Systems, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[59] Sen Song, Kenneth D Miller, and Larry F Abbott. Competitive hebbian learning through spike-timing-dependent synaptic plasticity. Nature Neuroscience, 3(9):919-926, 2000. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 217 + ], + "type": "text", + "content": "[60] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: fully convolutional one-stage object detection. In International Conference on Computer Vision, 2019. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 261 + ], + "type": "text", + "content": "[61] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. CoRR, abs/1711.11248, 2017. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 307 + ], + "type": "text", + "content": "[62] John J Wade, Liam J McDaid, Jose A Santos, and Heather M Sayers. Swat: A spiking neural network training algorithm for classification problems. IEEE Transactions on Neural Networks, 2010. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "text", + "content": "[63] Xiaofeng Wu, Velibor Bojkovic, Bin Gu, Kun Suo, and Kai Zou. Ftbc: Forward temporal bias correction for optimizing ann-snn conversion. In European Conference on Computer Vision, pages 155-173. Springer, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 396 + ], + "type": "text", + "content": "[64] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in Neuroscience, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 441 + ], + "type": "text", + "content": "[65] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Lup ing Shi. Direct training for spiking neural networks: Faster, larger, better. In AAAI Conference on Artificial Intelligence, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 444, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 486 + ], + "type": "text", + "content": "[66] Qu Yang, Jibin Wu, Malu Zhang, Yansong Chua, Xinchao Wang, and Haizhou Li. Training spiking neural networks with local tandem learning. Advances in Neural Information Processing Systems, 2022. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 489, + 295, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 552 + ], + "type": "text", + "content": "[67] Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, Xu Bo, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 555, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 295, + 609 + ], + "type": "text", + "content": "[68] Man Yao, Ole Richter, Guangshe Zhao, Ning Qiao, Yannan Xing, Dingheng Wang, Tianxiang Hu, Wei Fang, Tugba Demirci, Michele De Marchi, et al. Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip. Nature Communications, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 612, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 295, + 643 + ], + "type": "text", + "content": "[69] Friedemann Zenke and Emre O Neftci. Brain-inspired learning on neuromorphic substrates. Proceedings of the IEEE, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 644, + 295, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 644, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 56, + 644, + 295, + 686 + ], + "type": "text", + "content": "[70] Rui-Jie Zhu, Qihang Zhao, Guoqi Li, and Jason K Eshraghian. Spikegpt: Generative pre-trained language model with spiking neural networks. arXiv preprint arXiv:2302.13939, 2023.1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24397" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/bab57aa6-5f36-4253-ae3a-1e5e1f4632d7_content_list.json b/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/bab57aa6-5f36-4253-ae3a-1e5e1f4632d7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..be27d29fcdb5f922dc36e7427a1fdc5da9251fad --- /dev/null +++ b/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/bab57aa6-5f36-4253-ae3a-1e5e1f4632d7_content_list.json @@ -0,0 +1,1579 @@ +[ + { + "type": "text", + "text": "Infighting in the Dark: Multi-Label Backdoor Attack in Federated Learning", + "text_level": 1, + "bbox": [ + 114, + 130, + 883, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ye Li $^{1}$ , Yanchao Zhao $^{1*}$ , Chengcheng Zhu $^{2}$ , Jiale Zhang $^{2*}$ $^{1}$ Nanjing University of Aeronautics and Astronautics $^{2}$ Yangzhou University \n{miles.li, yczhao}@nuaa.edu.cn \nchengchengzhu2022@126.com, jialezhang@yzu.edu.cn", + "bbox": [ + 184, + 179, + 810, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning (FL), a privacy-preserving decentralized machine learning framework, has been shown to be vulnerable to backdoor attacks. Current research primarily focuses on the Single-Label Backdoor Attack (SBA), wherein adversaries share a consistent target. However, a critical fact is overlooked: adversaries may be non-cooperative, have distinct targets, and operate independently, which exhibits a more practical scenario called Multi-Label Backdoor Attack (MBA). Unfortunately, prior works are ineffective in the MBA scenario since non-cooperative attackers exclude each other. In this work, we conduct an in-depth investigation to uncover the inherent constraints of the exclusion: similar backdoor mappings are constructed for different targets, resulting in conflicts among backdoor functions. To address this limitation, we propose Mirage, the first non-cooperative MBA strategy in FL that allows attackers to inject effective and persistent backdoors into the global model without collusion by constructing in-distribution (ID) backdoor mapping. Specifically, we introduce an adversarial adaptation method to bridge the backdoor features and the target distribution in an ID manner. Additionally, we further leverage a constrained optimization method to ensure the ID mapping survives in the global training dynamics. Extensive evaluations demonstrate that Mirage outperforms various state-of-the-art attacks and bypasses existing defenses, achieving an average ASR greater than $97\\%$ and maintaining over $90\\%$ after 900 rounds. This work aims to alert researchers to this potential threat and inspire the design of effective defense mechanisms. Code has been made open-source.", + "bbox": [ + 89, + 319, + 483, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 803, + 220, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed the flourishing of federated learning, an advanced privacy-preserving distributed learning paradigm [24, 26]. FL allows participants to collab", + "bbox": [ + 89, + 829, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0ddbdf872f33853402ca79f8266fb5f8851611c6766853843c188eca1dfb8883.jpg", + "image_caption": [ + "Figure 1. \"Blue\" and \"Red\" construct similar OOD mappings (see t-SNE), resulting in a conflict for the neuron weights. Only the dominant attacker can induce the model to output high feature attributions for its backdoor samples (see FA of $k$ -th round) and successfully perform the attack. Similar in $t$ -th round." + ], + "image_footnote": [], + "bbox": [ + 524, + 284, + 717, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dc5b1c26396d5cda41147e7343923cf1fa2364fec9b37fd99cf135cb8c3d2f5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 285, + 895, + 433 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "oratively build a neural network model without leaking their data. Following its paradigm, a wide range of FL applications have emerged in various safety and privacy-critical application domains, including medical image analysis [10, 15], face recognition [30], and others [1, 11].", + "bbox": [ + 511, + 515, + 906, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite enhanced privacy, the distributed nature of FL makes them more vulnerable to various adversarial attacks, with backdoor attacks posing the greatest threat. For backdoor attacks in FL, an attacker injects a backdoor into the local model, which is then submitted to the FL system, causing the global model to inherit the backdoor after aggregation. Consequently, the global model functions correctly on clean data but misclassifies inputs stamped with the attacker's chosen backdoor trigger as the target class. Numerous studies have been done to improve the persistence and effectiveness of backdoors [2, 6, 43, 48], which pose significant threats to various applications, including computer vision [40], IoT networks [28], and healthcare systems [16].", + "bbox": [ + 511, + 595, + 908, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in backdoor attacks in FL, most, if not all, assume that attackers in training are colluding and sharing the same target class, which can be denoted as Single-Label Backdoor Attacks (SBA). However, situations become more complex in real-world FL applications, especially in large-scale FL scenarios [18, 22], where attackers engage in training individually and compromise", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 109, + 886, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25770", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the global model for their own purposes, i.e., Multi-Label Backdoor Attack (MBA). Intuitively, non-cooperative adversaries separately attack the global model and do not interfere with each other. Unfortunately, our empirical results indicate that directly applying existing attack methods to multiple adversaries leads to disorganized obstacles, making backdoors not as effective as they are in SBA scenarios. As illustrated in Fig. 1, two backdoors cannot be effective simultaneously because attackers construct similar Out-of-Distribution (OOD) backdoor mappings, indicating that their backdoor samples distribute together but outside the distributions of the target classes. This results in attackers excluding one another, with only the dominant attacker capable of embedding a backdoor in the global model, consequently causing the global model to output high feature attributions for its backdoor samples, thereby achieving a high attack success rate (ASR). It further raises a new problem: Can MBA successfully launch without collusion?", + "bbox": [ + 89, + 90, + 480, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned issue, we first conduct an in-depth investigation to uncover the inherent constraints of the exclusions of SBAs: they create similar backdoor mappings for distinct targets, resulting in conflicts among backdoor functions. Motivated by this observation, we propose Mirage, the first non-cooperative MBA strategy in FL, wherein multiple adversaries individually undermine the global model without collusion. At a high level, instead of coordinating the various OOD mappings, Mirage aims to ensure that the backdoor samples are processed as clean samples by constructing an effective and persistent ID mapping between backdoor features and target distributions. To achieve this goal, we introduce a trigger optimization method based on adversarial adaptation, which adaptively optimizes the trigger to fool a carefully designed OOD sample detector, thereby constructing the ID mapping. Furthermore, considering that the distributions change with global training, we further introduce a constrained optimization method to enhance the persistence of ID mapping by tightening the distribution of backdoor features, allowing the backdoor mapping to live within the global training dynamics. In summary, our contributions are as follows:", + "bbox": [ + 91, + 367, + 482, + 699 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Uncovering the inherent constraints of extending existing SBAs to MBAs. To the best of our knowledge, we are the first to investigate and elucidate the inherent constraints of extending existing SBAs to MBAs. We underscore the practical threat posed by such attacks in large-scale FL applications by proposing a multi-label backdoor attack method based on the in-distribution backdoor.", + "- Effective ID Mapping Construction. We propose a novel backdoor optimization method based on adversarial adaptation, which enables attackers to build the bridge between their backdoor and the distribution of their target class, allowing multiple adversaries to simultaneously launch backdoor attacks without interfering with others." + ], + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Persistent ID Mapping Enhancement. We additionally introduce a constrained optimization method to enhance the ID mapping by minimizing the similarity of features between backdoorsed samples and benign samples. This tightens the ID backdoor mapping and ensures its persistence within the global training dynamics.", + "- Evaluations and Discussions. Extensive experiments demonstrate that Mirage outperforms SOTA attacks across various settings. Furthermore, we discuss the vulnerability of existing defense methods to MBAs and present potential countermeasures against Mirage, highlighting the urgent demand for tailored defenses." + ], + "bbox": [ + 513, + 90, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 287, + 653, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Backdoor Attacks in Federated Learning", + "text_level": 1, + "bbox": [ + 513, + 313, + 864, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backdoor attacks aim to compromise the victim model to produce the attacker-desired outputs if and only if the inputs are assigned to a specific trigger pattern. In light of this, a few researchers also start to notice the threat of backdoor attacks on FL systems. Depending on whether the backdoor trigger is optimized, current backdoor attacks fall under the categories of either static-trigger attacks [2, 6, 20, 44, 48] or trigger-optimization attacks [7, 25, 43, 47].", + "bbox": [ + 511, + 335, + 903, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The former pre-selected a fixed trigger pattern to poison the local training set and combine them with manually manipulating methods to improve the backdoor effectiveness. For example, [2, 3] aim to replace the global model by uploading a scaled-up poisoned model. In advance, [7, 34, 48] suggest leveraging the unimportant redundant neurons to prevent the backdoor from being erased shortly. Particularly, Dai et al. [6] introduce contrastive learning from the feature perspective of the backdoor to enhance its durability. The latter optimizes trigger patterns according to different tasks and generally achieves better effectiveness and utility. 3DFed [19] and A3FL [43] introduced the backdoor unlearning in an adversarial way to obtain a robust trigger to make the attacks more covert and persistent. DBA [41] and NBA [27] are the most similar work with us. However, DBA focuses on attacking the same target class, which is also an SBA situation. NBA investigated the noncooperative multi-label attack in FL but neither revealed the inner reason for the inapplicability of previous works nor proposed an effective method.", + "bbox": [ + 511, + 458, + 903, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Backdoor Defenses in FL", + "text_level": 1, + "bbox": [ + 513, + 771, + 743, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, researchers have noticed the threat of backdoor attacks in FL systems and actively explored various backdoor defense techniques, which can be categorized into either detection [8, 21, 39] or mitigation methods [4, 29, 45, 46, 49].", + "bbox": [ + 511, + 794, + 903, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The detection-based methods aim to utilize anomaly detection techniques to determine whether a model is backdoored. Foolsgold [8] assigns lower aggregation weights", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25771", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to updates with high pairwise cosine similarities, thereby mitigating the impact of backdoor updates. RFLBAT [39] discriminates the malicious model according to the difference between poisoned updates and clean updates in low-dimensional projection space. In advance, Li et al. [21] proposed BackdoorIndicator (referred to as \"Indicator\" hereafter), which detects potentially poisoned models based on the OOD properties of backdoor samples. Mitigation-based methods explored how to purify the backdoor model by eliminating backdoor triggers. MultiKrum [4] is a byzantine-robust aggregation protocol that aggregates the global model of having the smallest $\\mathcal{L}_2$ distance. FLAME [29] adapts the weak DP method by noise boundary proof and a dynamic clipping bound, which is shown to alleviate the backdoor attack while still retaining a high main task accuracy.", + "bbox": [ + 89, + 90, + 483, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Attack intuitions", + "text_level": 1, + "bbox": [ + 89, + 349, + 256, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we investigate the inherent constraints that affect the effectiveness of backdoors when extending previous works to MBA scenarios and further provide the attack intuitions based on this investigation.", + "bbox": [ + 89, + 376, + 483, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Inherent constraints of SBA methods", + "text_level": 1, + "bbox": [ + 89, + 449, + 410, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recall that, from the neuron activation perspective, the backdoor function in a model is highly related to the specific activation paths across neurons. This results in the model being sensitive to triggered inputs and producing anomalously high feature attribution values [9, 34, 36]. Correspondingly, such high values result in backdoor samples distributed out of the target clean distribution in feature space, i.e. the backdoor function constructs the OOD backdoor mapping [21]. In MBA scenarios, attackers may adopt similar strategies to construct the activation paths, such as using redundant neurons, which leads to the construction of similar OOD mappings and competition for neuron weights. Specifically, different attackers have varying requirements for the output values of specific neurons within the pathway. However, the dominant attacker controls most neurons, ensuring they produce the desired values, which results in others failing to achieve their backdoor function. As shown in Fig. 2a, two attackers aim to construct persistent and effective OOD backdoor mappings by targeting redundant neurons. Such a strategy causes them to leverage overlapped activation paths to achieve different functions, which further results in infighting among adversaries. This drives us to find a solution that allows attackers to select different activation paths non-cooperatively.", + "bbox": [ + 89, + 472, + 483, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Attack intuitions and challenges", + "text_level": 1, + "bbox": [ + 89, + 847, + 372, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To alleviate the exclusion, it is critical to design a mapping strategy that ensures different attackers can independent", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0d649c628cdfb509da23a4774260195bc893edb30981bbd493a5bb047cefb70a.jpg", + "image_caption": [ + "(a) SBA's mapping." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 647, + 190 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6864a2cca13c4d70506a193ba3d89a2ee1b8688d4edf7548f8bcca559aa54d2f.jpg", + "image_caption": [ + "(b) Separate mapping." + ], + "image_footnote": [], + "bbox": [ + 647, + 88, + 779, + 189 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e83bbc755a5c97ac8a5cef83bf5780f6856973cd1c45dfe407503ab5785da6ad.jpg", + "image_caption": [ + "(c) ID mapping.", + "Figure 2. Illustrations of different mapping strategy." + ], + "image_footnote": [], + "bbox": [ + 781, + 88, + 911, + 189 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mappings with no intersections. Achieving this is challenging because attackers share no information with others and are unaware of each other's existence. Surprisingly, our empirical investigation reveals that A3FL can fulfill the aforementioned requirement by combining machine unlearning [37, 42], which indirectly stimulates the eliminations performed by other attackers. As shown in Fig 2b, the unlearning method adversarially optimizes the trigger to construct a distinct OOD backdoor mapping that differs from other attackers. Although this method mitigates the exclusions, it enhances the OOD characteristics of the backdoor functions, consequently introducing more challenges in bypassing the OOD-based defense methods, such as Indicator [21]. The OOD-based defense method leads us to a counterintuitive thought: Can a backdoor be triggered via a clean activation path? In other words, construct the ID mapping for the backdoor features and target distributions (present in Fig 2c). Therefore, once the adversaries construct the ID backdoor mapping to the target class, they naturally alleviate potential exclusions, even without collusion, since additional measures are no longer needed to circumvent the backdoor activation paths of other attackers. However, constructing the ID backdoor mapping remains challenging due to the following issues:", + "bbox": [ + 511, + 220, + 906, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenge 1: How to construct the ID mapping?", + "text_level": 1, + "bbox": [ + 532, + 582, + 862, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While backdoor attacks have emerged as a prominent topic in FL, to the best of our knowledge, no prior studies investigated the ID backdoor within FL. Intuitively, the backdoor sample should be processed as a clean sample following the clean activation path, meaning it should share similar feature outputs. One promising way is leveraging feature alignment. However, this may disrupt the relationships among clean classes, making it challenging for attackers to maintain the ID characteristics. Therefore, there is a pressing need to devise a novel method that ensures the effective construction of ID mapping.", + "bbox": [ + 511, + 598, + 906, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenge 2: How to maintain the persistence of ID mapping?", + "text_level": 1, + "bbox": [ + 511, + 763, + 905, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One key challenge for backdoor attacks in FL is that the continued dynamic clean updates may break the well-constructed backdoor mappings. Previous research focuses on leveraging neurons that are updated less frequently to build the neural activation path, which cannot be applied in ID mapping since our objective is to construct the bridge between the trigger pattern and the clean target activation path.", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25772", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, these maintenance requirements for ID mapping further exacerbate the challenges.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 89, + 133, + 225, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To formulate the attack scenario, we first introduce the threat model. Driven by attack intuitions, we propose two methods that enable attackers to construct an efficient and persistent ID mapping without cooperation.", + "bbox": [ + 89, + 159, + 482, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Threat model", + "text_level": 1, + "bbox": [ + 89, + 228, + 230, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adversaries' Capabilities and knowledge. We consider a FL system running image classification tasks and assume that adversaries are FL participants, which enables attackers to control their own local datasets and training processes. Note that adversaries do not have access to any information regarding benign clients or other adversaries, including but not limited to datasets, models, and target classes.", + "bbox": [ + 89, + 250, + 482, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adversaries' goal. Adversaries aim to plant a backdoor into the global model that makes it misclassify backdoor data to a specific class while leaving other tasks unaffected.", + "bbox": [ + 89, + 356, + 483, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Overview of Mirage", + "text_level": 1, + "bbox": [ + 89, + 410, + 279, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the attack intuition, our objective is to construct an ID mapping to bridge the backdoor features and distribution of the target class, thereby freeing attackers from the exclusion of using similar backdoor activation paths. This presents two significant challenges: how to construct a qualified ID mapping and how to maintain it. To address these challenges, we propose Mirage, the first multilabel backdoor attack in FL, which enables attackers to noncooperatively compromise the global model by constructing an ID mapping between the backdoor features and target distributions. As illustrated in Fig. 3, Mirage introduces two components to construct an effective and persistent ID backdoor mapping. On the one hand, we propose a backdoor optimization method based on adversarial adaptation, which constructs the ID mapping by maximizing the misclassification probability of a well-designed backdoor sample detector. On the other hand, we deploy a constrained optimization strategy to tighten the backdoor distribution, ensuring its survival in the global training dynamics. Next, we provide details of each component.", + "bbox": [ + 89, + 431, + 483, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Effective ID Mapping Construction", + "text_level": 1, + "bbox": [ + 89, + 742, + 398, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the first challenge, we aim to adversarially optimize the trigger pattern to make it able to construct the ID mapping. In order to optimize the trigger, we employ an OOD sample detector to determine whether a backdoored sample is an OOD sample and further optimize the trigger pattern to fool the detector. This introduces another problem: How can an attacker obtain an accurate detector? Previous research suggests training a new model as the detector [47]. However, retraining the detector with each trigger", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "update incurs significant computational overhead, and the local samples limit its effectiveness. These shortcomings prompt us to seek another solution to obtain a detector that is both effective and lightweight for updates. Recall that the FL training process allows clients to obtain the latest global model, which contains a well-trained feature extractor. Consequently, we consider leveraging the feature extractor of the global model as the primary component and training a binary classifier to distinguish the OOD samples. Such an approach is a common practice in transfer learning, consumes only a few computing resources, and requires little data.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to obtain an accurate detector, we first prepare the dataset for training. The local dataset for each attacker can be divided into $D_{cln} = \\{(x,y) \\in D_{att} | + \\}$ and $D_{bd} = \\{(x \\oplus \\delta, y) \\in D_{att} | - \\}$ , where $+$ indicates the ground truth label of a sample is $c$ , and $-$ is the opposite and $\\delta$ represents the trigger pattern. Then, the detector dataset $D_{detector}$ ( $D_d$ for short) is composed as follows:", + "bbox": [ + 511, + 272, + 906, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {d} = \\{(x, -) \\mid x \\in D _ {b d} \\} \\cup \\{(x, +) \\mid x \\in D _ {c l n} \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 390, + 906, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To make the ID mapping construction more effective and flexible, we adopt an adversarial training-like method to optimize the trigger pattern. Specifically, $\\theta_{detector}$ is trained on the $D_{detector}$ to differentiate between clean and backdoored samples by minimizing the binary cross-entropy (BCE) loss.", + "bbox": [ + 511, + 420, + 905, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concurrently, the trigger is optimized to fool the detector by maximizing the binary classification loss. Such a min-max game equips the trigger with the ability to activate the clean activation path of the target class in the global model, naturally avoiding conflicts with other attackers. The min-max process is mathematically described as follows:", + "bbox": [ + 511, + 511, + 905, + 601 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\delta = \\operatorname {a r g m i n} \\mathbb {E} _ {(x, y) \\sim D _ {\\text {d e t e c t o r}}} [ \\mathcal {L} (x, +, \\theta) | y = - ] \\\\ s. t. \\theta = \\underset {\\theta} {\\overset {\\delta} {\\operatorname {a r g m i n}}} \\mathbb {E} _ {(x, y) \\sim D _ {\\text {d e t e c t o r}}} [ \\mathcal {L} (x, y, \\theta) ] \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 613, + 903, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta$ is the detector's parameters, with its feature extractor frozen and parameterized by the weights of the global model, and $\\mathcal{L}$ is the BCE loss. For better implementation, we replace the maximization problem with minimizing the BCE loss of misclassification for the backdoored samples.", + "bbox": [ + 511, + 671, + 905, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.4. Persistent ID Mapping Enhancement", + "text_level": 1, + "bbox": [ + 511, + 757, + 834, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although using an OOD detector and an adversarial strategy restricts the trigger pattern $\\delta$ to reveal the ID feature, our empirical results indicate that the backdoored samples still do not meet our expectations. This is because the adversarial strategy does not constrain the tightness of backdoor mapping, which causes the margin-distributed backdoored samples not in-distribution anymore with the dynamic changes of training and failing to activate the target", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25773", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2e7879dd771ac7d4de94e80976837e864c9225773ab16729370982d3425f267f.jpg", + "image_caption": [ + "Figure 3. The workflow of Mirage. Step 1: Train an OOD sample detector $\\theta_{\\text{Detector}}$ . Step 2: Construct ID mapping by maximizing misclassification probabilities of backdoor samples on the detector. Step 3: Tighten backdoor distribution by minimizing the feature similarities between backdoor samples and benign samples. Step 4: Optimize the trigger by minimizing $L_{\\text{Enhance}}$ and $L_{\\text{detector}}$ ." + ], + "image_footnote": [], + "bbox": [ + 130, + 87, + 867, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "activation path. One promising solution is to leverage distribution estimation methods, like KDE [33], to obtain the actual distribution of the target class and use data augmentation to generate samples to improve the effectiveness of the detector. Although it has been proven effective, it requires countless computing resources. Therefore, we introduce a constrained optimization method to enhance the ID mapping by tightening the distribution of backdoored samples. Since the adversarial strategy successfully constructs the ID mapping, we aim to \"push\" the backdoor distribution to make it tighter. Specifically, we minimize the similarities between clean samples $x$ and their corresponding backdoored samples $x \\oplus \\delta$ , allowing the backdoored sample to deviate as much as possible from the original distribution. Additionally, we establish the direction by minimizing the loss of the backdoor samples on the global model. Formally, it can be formulated as the following optimization problem:", + "bbox": [ + 91, + 315, + 483, + 573 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta = \\operatorname {a r g m i n} \\mathrm {C S} \\left[ \\left(\\theta_ {f} (x), \\theta_ {f} (\\hat {x}) \\right] + \\mathrm {C E} (\\hat {x}, \\hat {y}, \\theta) \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 580, + 482, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_{f}$ represents the feature extractor of the global model $\\theta$ . $\\hat{x} = x \\oplus \\delta, x \\in D_{att}$ , $\\oplus$ is the backdoored operator. $\\mathrm{CS}(\\cdot)$ and $\\mathrm{CE}(\\cdot)$ represent the Cosine-similarity and the Cross-entropy loss, separately. Therefore, the backdoor distribution can be tightened into the target distribution and live within the global training dynamics.", + "bbox": [ + 89, + 606, + 483, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We summarized Mirage in Algorithm 1, and the detailed description is provided in Appendix B.", + "bbox": [ + 89, + 696, + 483, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 739, + 223, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we provide extensive experimental results to demonstrate the effectiveness of Mirage by comparing its performance with several SOTA backdoor attack methods from multiple angles. Our experiments run on a FL system performing image classification tasks, using an NVIDIA GeForce RTX 4090 GPU equipped with 24GB of memory. The detailed setups are listed as follows:", + "bbox": [ + 89, + 763, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset: We conduct experiments on three widely used public real-world datasets, CIFAR-10, CIFAR-100 [17] and", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Mirage at $i$ -th attacker." + ], + "code_body": "Data: $\\theta_{t},\\theta_{detector},\\theta_{fe},\\mathcal{D}_{i},\\delta ,\\eta ,c,E.$ \n1 $\\theta_t^i = \\theta_t$ \n2 /\\* Trigger adversarial adaption\\*/ \n3 for epoch $e = 1,\\ldots ,E$ do \n4 Generate $D_{detector}$ by Eq 1; \n5 $\\theta_{detector} =$ argmin $\\mathbb{E}_{(x,y)\\sim D_{detector}}[\\mathcal{L}(x,y,\\theta_{detector})]$ . \n6 for batch $\\mathcal{B}\\in D_i$ do \n7 $\\begin{array}{r}\\mathcal{L}_{detector} = \\\\ \\frac{1}{|\\mathcal{B}|}\\sum_{(x,y)\\in \\mathcal{B}}\\mathcal{L}_{BCE}(x\\oplus \\delta , + ,\\theta_{detector});\\\\ \\mathcal{L}_{Enhance} = \\frac{1}{|\\mathcal{B}|}\\sum_{(x,y)\\in \\mathcal{B}}\\mathcal{L}_{CE}(x\\oplus \\delta ,\\hat{y},\\theta_t^i)\\\\ +\\mathcal{L}_{CS}((\\theta_{fe}(x),\\theta_{fe}(x\\oplus \\delta));\\\\ \\delta \\gets \\delta -\\eta \\cdot \\nabla_{\\delta}(\\mathcal{L}_{detector} + \\mathcal{L}_{Enhance});\\\\ end \\end{array}$ \n10 end \n13 /\\* Local training\\*/ \n14 Get the poisoned local dataset $\\mathcal{D}_i^{poison}$ with $\\delta$ \n15 Local optimize $\\theta_t^\\prime$ on $\\mathcal{D}_i^{poison}$ . \n16 $\\Delta_i^{t + 1} = \\theta_t - \\theta_t'$ . \n17 Upload $\\Delta_i^{t + 1}$ to the server;", + "bbox": [ + 513, + 335, + 903, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "GTSRB [13]. Details are provided in Appendix C.", + "bbox": [ + 511, + 702, + 846, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Federated learning setup: By default, we set the number of FL clients $N = 100$ . The server randomly selects $M = 10$ clients at each communication round to contribute to the global model. The default global model architecture is ResNet-18 [12], which is a classic and effective model for image classification. Based on previous research, we randomly split the dataset among clients non-IID using Dirichlet sampling [14] with a hyperparameter $\\alpha = 1$ . For each selected benign client, they train the local model for two local epochs using an SGD optimizer with a learning rate of 0.01. The batch size is 64 for CIFAR-10 and CIFAR-100 and 32 for GTSRB. The entire training process continues", + "bbox": [ + 511, + 719, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25774", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for 2100 communication rounds.", + "bbox": [ + 89, + 90, + 308, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attack and defense setup: We compare the effectiveness of Mirage with SOTA methods, including A3FL [43], Chameleon [6], Neurotoxin [48], PGD [38], and Vanilla [2]. We further extend the aforementioned methods to MBA scenario, wherein attackers independently execute their attacks. Unless otherwise specified, we set $N = 3$ , with the target classes being 0, 1, and 2. In line with A3FL, attackers of Mirage optimize their triggers by PGD and poison $12.5\\%$ of $D_{att}$ . Blend [5] is utilized as a backdoor trigger across all attacks. The attack window starts from the 2000-th round and continues for 100 rounds, so attackers are forced to wait for selection and cannot launch a continuous attack.", + "bbox": [ + 89, + 106, + 483, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also evaluate those attacks against six SOTA defenses: Multi-Krum [4], DeepSight [31], Foolsgold [8], RFLBAT [39], FLAME [29] and BakcdoorIndicator [21]. For those attacks and defense methods, we follow the original implementations. Note that we selected the random noise dataset as the source of the indicator dataset.", + "bbox": [ + 89, + 287, + 483, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation metrics", + "text_level": 1, + "bbox": [ + 89, + 390, + 272, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous works, we comprehensively evaluate Mirage using accuracy (Acc), attack success rate (ASR), and lifespan. Specifically, Acc reflects the attack stealthiness, which indicates the performance of the backdoor model on the main tasks. ASR represents the ratio of backdoors samples misclassified as the labels specified by the attackers. Note that we report the mean ASR of three attackers on the global model at the end of FL. To evaluate the persistence of our attack, we introduce lifespan, denoted as the period that starts at the end of the attack window and ends when the ASR decreases to below a chosen threshold.", + "bbox": [ + 89, + 412, + 483, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Attack performance", + "text_level": 1, + "bbox": [ + 89, + 590, + 282, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we evaluate the attack performance of Mirage across various datasets and compare it to different SOTA attack methods. The results are summarized in Table 1, and the best results are highlighted in Bold. The experiment shows that Mirage achieves remarkable results across three tasks and various SOTA defense methods compared to the baseline methods. In terms of accuracy, Mirage achieves the best accuracy in over half of the tasks and obtains suboptimal results for the remaining tasks, losing to the best ones by an average of only $0.498\\%$ . Regarding effectiveness, Mirage achieves the best ASRs across all the tasks, with the gap among attackers being acceptable (the maximum gap is $6.99\\%$ , and the average gap is $2.19\\%$ ). Those results indicate that the design of constructing the ID mapping between backdoor features and target class allows attackers to independently attack the FL model without concerns about the potential exclusion, resulting in a global model that excels in both main and multiple backdoor tasks.", + "bbox": [ + 89, + 613, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, A3FL is a relatively competitive method", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0eb1145d418053a35ff027e368df1827a1b78f41c973be07acd4a0209dc9f924.jpg", + "image_caption": [ + "Figure 4. Persistent evaluations." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 901, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that leverages unlearning to simulate the potential exclusion and optimize the trigger to construct robust activation paths independent of other attackers. However, as described in Sec. 3.2, such an approach may enhance its OOD mapping concerning the benign distribution and increase the risk of being detected by OOD-based detection methods, such as Indicator [21]. The results of A3FL with the defense of Indicator support our analysis. Besides, the Indicator does not perform as expected, we will discuss it in Section 6.3.", + "bbox": [ + 511, + 227, + 906, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Lifespan", + "text_level": 1, + "bbox": [ + 513, + 373, + 617, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Secondly, we evaluate the durability of attacks by comparing their lifespans. We extend the FL training process from 2100 to 3000 rounds while keeping the attack window to the first 100 rounds. Fig. 4 illustrates the attack success rate without any defense method. Mirage achieves similar results with A3FL, where the ASRs exceed $90\\%$ at the end, indicating a lifespan of over 900 rounds. In contrast, the ASRs of all other baseline attacks are constrained by potential exclusion between attackers, as described before, making them ineffective and falling rapidly. These experimental results suggest that Mirage can achieve comparable durability to A3FL, which focuses on enhancing the persistence of backdoors in FL.", + "bbox": [ + 511, + 396, + 906, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. The impact of different models", + "text_level": 1, + "bbox": [ + 511, + 603, + 785, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate Mirage across different model architectures. Following previous research, we select widely used image classification model architectures: ResNet18 and ResNet34 [12], VGG11 and VGG19 [35], and MobileNet-V2 [32]. We set the parameters to default and only change the model architectures. As shown in Tab. 2, the attack performance of Mirage remains high in ASRs. The decrease in MobileNet-V2 is mainly attributed to the low utility of its model performance, which significantly reduces the detector's utility, resulting in sub-optimal ID mapping construction for certain attackers. However, the average ASRs remain acceptable. In conclusion, based on the results presented above, we can conclude that Mirage is not sensitive to specific model structures.", + "bbox": [ + 511, + 626, + 906, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.5. The impact of adversary number", + "text_level": 1, + "bbox": [ + 511, + 848, + 803, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also study the impact of the number of adversaries, which can be considered as the number of different targets", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25775", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/36da78277987ff9989e904e32fd36a9623b89f4e602b4bf3153d45994be1f6d5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetDefenseVanillaPGDNeurotoxinChameleonA3FLMirage
Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)
CIFAR-10nodefense91.7631.88*87.6891.5231.25*92.7892.0352.99*90.2992.3539.65*79.9891.7399.52*0.2392.1699.54*0.27
deepsight91.4431.71*88.2990.8131.083*92.6392.3139.24*83.5092.3236.66*79.6692.1796.56*7.7491.2897.39*3.82
foolsgold88.7932.50*89.5090.7131.67*93.9691.0480.91*21.8390.6645.4667.2589.0095.30*12.0291.3297.35*2.51
Indicator85.0828.7685.3285.8320.6059.5085.1329.9860.32*86.138.215.1385.7570.33*88.4591.1093.46*3.90
multikrum92.361.594.5692.391.183.2592.641.895.1892.461.905.2592.1878.15*64.4192.1092.30*3.12
rfbat86.8330.5588.3488.2833.27*99.7987.2233.27*73.1286.9737.1586.4285.4687.83*10.2185.7497.29*2.11
Flame92.561.835.2692.513.399.9092.0240.97*94.0492.281.383.7192.2579.40*60.2392.3695.9*2.42
CIFAR-100nodefense69.4032.62*94.3170.5332.28*93.8971.0732.62*97.0571.6624.0727.5071.2698.67*2.4771.6599.05*0.77
deepsight70.0232.86*89.0270.5432.23*96.3270.5333.14*93.2871.3024.0626.1771.6892.05*18.4471.5698.67*2.43
foolsgold70.0233.3089.8170.1532.65*97.3471.0732.63*97.0571.6024.2027.1171.7898.71*1.6471.9499.18*1.00
Indicator67.736.8020.1263.855.8817.2860.9811.8932.3461.356.0713.4365.3537.13*92.7268.2299.8*0.10
multikrum71.810.380.7771.140.310.5971.630.360.6671.850.210.3871.7691.28*18.5571.8693.57*5.14
rfbat62.360.300.4163.100.600.1262.150.561.0061.4918.0127.3161.3594.34*16.5663.1995.68*6.99
Flame71.750.210.2872.090.330.6272.020.270.4572.100.300.5271.2291.48*19.8371.8294.22*3.16
GTSRBnodefense96.1033.32*98.2496.7333.10*98.8095.9338.39*95.7996.0845.62*87.0696.6599.63*0.4996.9799.73*0.48
deepsight95.9533.8*97.4495.9733.36*98.3596.1233.06*99.1796.1450.1476.7195.9594.78*12.6496.2096.76*2.53
foolsgold96.0833.47*97.9296.9433.06*98.7595.8837.57*96.3696.0451.9178.0696.6297.68*6.9596.9599.88*0.27
Indicator91.5750.9291.5389.4527.2778.1489.0817.0950.9792.912.765.3885.5185.05*25.2695.1299.73*0.70
multikrum96.244.1412.2096.972.747.7296.124.8414.3296.134.7213.9796.6298.18*4.1696.9298.99*1.73
rfbat94.4056.61*90.6289.4433.26*99.7294.5333.88*97.8094.7554.39*83.0495.2397.86*6.4294.2199.97*0.08
Flame95.886.1918.4596.972.747.7295.925.1515.3295.925.9517.7196.5299.31*1.5296.9299.32*1.05
", + "bbox": [ + 91, + 88, + 906, + 295 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c830df29a3bb5c13eb36b95aea2e3467153fb083ac2b4eb3970241b03bcfbaec.jpg", + "table_caption": [ + "Table 1. Performance of Mirage compared with baseline attack methods against the SOTA defense methods. The Acc higher represents better, and so does the ASR. \"GAP\" represents the difference between the highest and the lowest ASR, and attack methods should maintain their GAP close to zero. An \" * \" after ASR indicates that at least one of the three attackers achieved over ${90}\\%$ ASR." + ], + "table_footnote": [], + "table_body": "
ResNet18ResNet34VGG11VGG19MobileNet-V2
Acc92.1693.8991.1492.0785.38
ASR98.899.6399.2199.8297.65
", + "bbox": [ + 122, + 359, + 450, + 404 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a1e3ac5d3b81e792a0122d917ccb5fb51c910e265060b0e1e8b05f46a900d1df.jpg", + "table_caption": [ + "Table 2. Results on different model architectures." + ], + "table_footnote": [], + "table_body": "
Dataset0.515101000
CIFAR100Acc71.4671.6571.6871.5071.81
ASR99.5999.0599.1699.2699.21
GTSRBAcc96.9696.9796.6296.5796.70
ASR99.5499.7399.4798.2698.86
CIFAR10Acc91.6192.1692.4392.5192.38
ASR98.9598.8099.0399.3999.15
", + "bbox": [ + 143, + 431, + 431, + 523 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/70d471edf15185594ff36e1fafad868330dec85c28c5308dfd35a326f227137a.jpg", + "table_caption": [ + "Table 3. Results on different non-IID settings." + ], + "table_footnote": [], + "table_body": "
Attacker Number12345
Acc92.3792.5392.1692.1192.12
ASR99.0699.1598.8098.37598.484
GAP0.000.490.381.421.99
", + "bbox": [ + 143, + 551, + 431, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Mirage with different attacker number.", + "bbox": [ + 143, + 614, + 426, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "since we assign different targets to the attackers. Table 4 presents the experimental results on CIFAR-10 with varying attack numbers ranging from 1 to 5, evaluating the performance of the global model on the main tasks and different backdoor tasks. Note that other settings follow the default configuration. We observe that the increase in attack numbers has a negligible impact on Mirage, with the accuracy of main tasks maintained above $92\\%$ and the average ASR exceeding $98\\%$ . The experimental results indicate that backdoor injection without collusion can be effectively implemented by constructing ID mapping. We also conducted experiments on different datasets (see Appendix D), and the results support the above conclusions.", + "bbox": [ + 89, + 641, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.6. Performance under varying non-IID degrees", + "text_level": 1, + "bbox": [ + 89, + 848, + 467, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 3 exhibits the attack performance of Mirage under varying non-IID degrees. We set different sampling param", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "eters $\\alpha$ for Mirage, with larger values indicating a distribution that is closer to IID. The results indicate that Mirage can achieve stable performance across different non-IID degrees. As described in Sec. 4.4, the adversarial strategy may optimize the backdoor distribution to the margin of the target distribution. To address this issue, we introduce a constrained optimization method, tightening the backdoor distribution into the objective distribution and ensuring successful attacks across different distributions.", + "bbox": [ + 511, + 363, + 906, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.7. Illustration of backdoor distribution", + "text_level": 1, + "bbox": [ + 511, + 510, + 828, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The preceding experiments leverage t-SNE to illustrate the backdoor distribution for Mirage and baseline methods. The results are in Fig. 5. The backdoor distribution from Fig. 5(a) to Fig. 5(d) further corroborates our presentation in 1, that different attackers may share similar neuron activation paths, resulting in their potential mitigation since their distribution is close to each other. On the other aspect, A3FL paves the activation path by adversariably unlearning the backdoor, which enhances their OOD mapping, producing clear boundaries between classes and tight clustering within each class. Although effective, this poses more threats to A3FL since the OOD-sample-based detection method can detect their backdoor models effectively. Looking back to Mirage, which is pictured in Fig. 5(f), backdoor samples present in-distribution with the target clean distribution. The results further indicate that the proposed Mirage can construct the ID mapping between backdoor features and clean distributions of the target class.", + "bbox": [ + 511, + 532, + 906, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.8. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 816, + 663, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sections 4.3 and 4.4 propose two hybrid techniques to construct and enhance the ID mapping. Here, we conduct an ablation study to show the effects of different techniques for our method. The experiment adheres to the settings", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25776", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c80a62ca8377a5018d653c1b71f1e7795ac3f9f31dcd4f8725bc3b91553e1e77.jpg", + "image_caption": [ + "Figure 5. Illustration of backdoor distribution." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 893, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9acf7e4ede2ad0a10fd9d0f4ece7275c1dcdad459d82955fca221a65239fcf45.jpg", + "image_caption": [ + "Figure 6. Ablation study." + ], + "image_footnote": [], + "bbox": [ + 138, + 202, + 428, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "outlined in Section 5.3. Fig. 6 presents the average ASR for Mirage and Mirage without enhancement. We can see that introducing the ID mapping enhancement significantly improves Mirage's performance. As described in Section 4.4, although the adversarial strategy successfully constructs the ID mapping, the backdoors may not fully activate the clean path, making the backdoor distribution map to the target distribution's margin, rendering it ineffective and non-persistent. Therefore, we introduce the constrained optimization method to tighten the backdoor distribution, ensuring backdoor samples will not be misidentified as the target distribution changes.", + "bbox": [ + 89, + 301, + 483, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 89, + 500, + 205, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1. Does the pixel trigger still effective?", + "text_level": 1, + "bbox": [ + 89, + 523, + 400, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the previous experiments, we adopted the blend trigger as the trigger pattern. Here, we conduct experiments with pixel triggers to evaluate their effectiveness. The results (presented in Appendix E) show that the pixel trigger remains effective. According to the t-SNE analysis of the feature attribution, the backdoor samples cannot aggregate together like the blend trigger. The main reason is that the pixel trigger can only activate a subset of neurons along the activation path, while other features typically activate the paths belonging to their original classes. However, it still constructs the ID mapping to the target distribution and successfully embeds the backdoors for different attackers.", + "bbox": [ + 89, + 547, + 483, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. One-shot Mirage", + "text_level": 1, + "bbox": [ + 89, + 741, + 256, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mirage can perform a one-shot attack by increasing adversarial training epochs, allowing the trigger to directly simulate features of the target class, like adversarial samples. However, it does not construct persistent and effective ID backdoor mapping, resulting in low durability and side effects on the main tasks. Therefore, we enable the model to learn from the backdoor samples and construct effective and persistent ID mapping by injecting more rounds and fewer batches to achieve high-confidence classification.", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Potential defense methods", + "text_level": 1, + "bbox": [ + 513, + 203, + 750, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "According to Section 5.2, we can see that although Indicator [2], the most powerful detection method for now, can detect at least one of the three A3FL attackers, the rest can still successfully inject the backdoor. This is because it constructs limited OOD mappings, succeeding only under the assumption that the attacker constructs similar OOD mappings. Therefore, it cannot be as effective against dynamic attacks (A3FL) as it is against static attacks (Chameleon, etc.). Furthermore, for A3FL, such misdetections are sufficient for the attacker to inject a persistent backdoor into the global model. Thus, following the works with foresight [23, 27], we urge researchers to focus on MBA and expand their efforts to defend against it.", + "bbox": [ + 511, + 224, + 906, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed Mirage does not follow the traditional backdoor activation scheme, making it challenging for previous work to resist. This motivates us to seek solutions to address this threat to real-world FL applications. The advanced explainable method may effectively detect our attack, however, detecting all backdoored models remains challenging. Input detection may also identify our attack, as clean samples differ from backdoored samples.", + "bbox": [ + 511, + 421, + 905, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 556, + 633, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we uncover the inherent constraints of SBAs when extending them to MBA scenarios: Non-cooperative attackers will exclude others as they construct similar backdoor mappings. Based on this observation, we introduce Mirage, the first non-cooperative MBA strategy in FL. Mirage ensures the backdoor functions of different attackers by constructing effective and persistent ID backdoor mappings. Specifically, an adversarial trigger optimization method is introduced to construct the ID mapping between backdoor features and target distributions. Subsequently, Mirage leverages a constrained optimization method to tighten the ID mapping for persistence. Our comprehensive experiments demonstrate that Mirage outperforms existing backdoor attacks across different settings. For future work, investigating defense methods will be our primary focus.", + "bbox": [ + 511, + 580, + 906, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 513, + 808, + 647, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant(No. 62172215) and in part by the A3 Foresight Program of NSFC (Grant No. 62061146002)", + "bbox": [ + 511, + 823, + 905, + 882 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25777", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Apple. Private federated learning (neurips 2019 expo talk abstract). https://nips.cc/ExpoConferences/2019/schedule?talkid=40. accessed: 2020-05-22. 1", + "[2] Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov. How to backdoor federated learning. In International conference on artificial intelligence and statistics, pages 2938-2948. PMLR, 2020. 1, 2, 6, 8", + "[3] Arjun Nitin Bhagoji, Supriyo Chakraborty, Prateek Mittal, and Seraphin Calo. Analyzing federated learning through an adversarial lens. In International conference on machine learning, pages 634-643. PMLR, 2019. 2", + "[4] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 3, 6", + "[5] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 6, 1", + "[6] Yanbo Dai and Songze Li. Chameleon: Adapting to peer images for planting durable backdoors in federated learning. In International Conference on Machine Learning, pages 6712-6725. PMLR, 2023. 1, 2, 6", + "[7] Pei Fang and Jinghui Chen. On the vulnerability of backdoor defenses for federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11800-11808, 2023. 2", + "[8] Clement Fung, Chris J. M. Yoon, and Ivan Beschastnikh. The limitations of federated learning in sybil settings. In 23rd International Symposium on Research in Attacks, Intrusions and Defenses, RAID 2020, San Sebastian, Spain, October 14-15, 2020, pages 301-316. USENIX Association, 2020. 2, 6", + "[9] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. BadNets: Evaluating Backdooring Attacks on Deep Neural Networks. IEEE Access, 7:47230-47244, 2019. 3", + "[10] Hao Guan, Pew-Thian Yap, Andrea Bozoki, and Mingxia Liu. Federated learning for medical image analysis: A survey. Pattern Recognition, page 110424, 2024. 1", + "[11] Andrew Hard, Kanishka Rao, Rajiv Mathews, Swaroop Ramaswamy, Françoise Beaufays, Sean Augenstein, Hubert Eichner, Chloe Kiddon, and Daniel Ramage. Federated Learning for Mobile Keyboard Prediction, 2018. 1", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 5, 6", + "[13] Sebastian Houben, Johannes Stallkamp, Jan Salmen, Marc Schlipsing, and Christian Igel. Detection of traffic signs in real-world images: The german traffic sign detection benchmark. In The 2013 international joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2013. 5" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. 5", + "[15] Meirui Jiang, Zirui Wang, and Qi Dou. Harmofl: Harmonizing local and global drifts in federated learning on heterogeneous medical images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1087-1095, 2022. 1", + "[16] Ruinan Jin and Xiaoxiao Li. Backdoor attack and defense in federated generative adversarial network-based medical image synthesis. Medical Image Analysis, 90:102965, 2023. 1", + "[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009. 5, 1", + "[18] Fan Lai, Yinwei Dai, Sanjay Singapuram, Jiachen Liu, Xi-angfeng Zhu, Harsha Madhyastha, and Mosharaf Chowdhury. Fedscale: Benchmarking model and system performance of federated learning at scale. In International conference on machine learning, pages 11814–11827. PMLR, 2022. 1", + "[19] Haoyang Li, Qingqing Ye, Haibo Hu, Jin Li, Leixia Wang, Chengfang Fang, and Jie Shi. 3dfed: Adaptive and extensible framework for covert backdoor attack in federated learning. In 44th IEEE Symposium on Security and Privacy, SP 2023, San Francisco, CA, USA, May 21-25, 2023, pages 1893-1907. IEEE, 2023. 2", + "[20] Minghui Li, Wei Wan, Yuxuan Ning, Shengshan Hu, Lulu Xue, Leo Yu Zhang, and Yichen Wang. Darkfed: A data-free backdoor attack in federated learning. arXiv preprint arXiv:2405.03299, 2024. 2", + "[21] Songze Li and Yanbo Dai. BackdoorIndicator: Leveraging OOD data for proactive backdoor detection in federated learning. In 33rd USENIX Security Symposium (USENIX Security 24), pages 4193-4210, Philadelphia, PA, 2024. USENIX Association. 2, 3, 6", + "[22] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1", + "[23] Yige Li, Jiabo He, Hanxun Huang, Jun Sun, and Xingjun Ma. Shortcuts everywhere and nowhere: Exploring multi-trigger backdoor attacks, 2024. 8", + "[24] Ye Li, Jiale Zhang, Yanchao Zhao, Bing Chen, and Shui Yu. Fairness-aware federated learning framework on heterogeneous data distributions. In ICC 2024-IEEE International Conference on Communications, pages 728-733. IEEE, 2024. 1", + "[25] Xiaoting Lyu, Yufei Han, Wei Wang, Jingkai Liu, Bin Wang, Jiqiang Liu, and Xiangliang Zhang. Poisoning with cerberus: Stealthy and colluded backdoor attack against federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9020-9028, 2023. 2", + "[26] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25778", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 1", + "[27] Tuan Nguyen, Dung Thuy Nguyen, Khoa D Doan, and Kok-Seng Wong. Non-cooperative backdoor attacks in federated learning: A new threat landscape. arXiv preprint arXiv:2407.07917, 2024. 2, 8, 1", + "[28] Thien Duc Nguyen, Phillip Rieger, Markus Miettinen, and Ahmad-Reza Sadeghi. Poisoning attacks on federated learning-based IoT intrusion detection system. In Proc. Workshop Decentralized IoT Syst. Secur.(DISS), 2020. 1", + "[29] Thien Duc Nguyen, Phillip Rieger, Huili Chen, Hossein Yalame, Helen Möllering, Hossein Fereidooni, Samuel Marchal, Markus Miettinen, Azalia Mirhoseini, Shaza Zeitouni, Farinaz Koushanfar, Ahmad-Reza Sadeghi, and Thomas Schneider. FLAME: taming backdoors in federated learning. In 31st USENIX Security Symposium, USENIX Security 2022, Boston, MA, USA, August 10-12, 2022, pages 1415-1432. USENIX Association, 2022. 2, 3, 6", + "[30] Yifan Niu and Weihong Deng. Federated learning for face recognition with gradient correction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1999–2007, 2022. 1", + "[31] Phillip Rieger, Thien Duc Nguyen, Markus Miettinen, and Ahmad-Reza Sadeghi. Deepsight: Mitigating backdoor attacks in federated learning through deep model inspection. In 29th Annual Network and Distributed System Security Symposium, NDSS 2022, San Diego, California, USA, April 24-28, 2022. The Internet Society, 2022. 6", + "[32] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE conference on computer vision and pattern recognition*, pages 4510–4520, 2018. 6", + "[33] Simon J Sheather and Michael C Jones. A reliable data-based bandwidth selection method for kernel density estimation. Journal of the Royal Statistical Society: Series B (Methodological), 53(3):683-690, 1991. 5", + "[34] Chenghui Shi, Shouling Ji, Xudong Pan, Xuhong Zhang, Mi Zhang, Min Yang, Jun Zhou, Jianwei Yin, and Ting Wang. Towards practical backdoor attacks on federated learning systems. IEEE Transactions on Dependable and Secure Computing, 2024. 2, 3", + "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6", + "[36] Guanhong Tao, Zhenting Wang, Shiwei Feng, Guangyu Shen, Shiqing Ma, and Xiangyu Zhang. DRUPE: Distribution Preserving Backdoor Attack in Self-supervised Learning. In 2024 IEEE Symposium on Security and Privacy (SP), pages 29-29. IEEE Computer Society, 2023. 3", + "[37] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE symposium on security and privacy (SP), pages 707-723. IEEE, 2019. 3" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris S. Papailiopoulos. Attack of the tails: Yes, you really can backdoor federated learning. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 6", + "[39] Yongkang Wang, Dihua Zhai, Yufeng Zhan, and Yuanqing Xia. RFLBAT: A robust federated learning algorithm against backdoor attack. CoRR, abs/2201.03772, 2022. 2, 3, 6", + "[40] Emily Wenger, Josephine Passananti, Arjun Nitin Bhagoji, Yuanshun Yao, Haitao Zheng, and Ben Y Zhao. Backdoor attacks against deep learning systems in the physical world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6206-6215, 2021. 1", + "[41] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. DBA: distributed backdoor attacks against federated learning. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 2", + "[42] Yi Zeng, Si Chen, Won Park, Z Morley Mao, Ming Jin, and Ruoxi Jia. Adversarial unlearning of backdoors via implicit hypergradient. arXiv preprint arXiv:2110.03735, 2021. 3", + "[43] Hangfan Zhang, Jinyuan Jia, Jinghui Chen, Lu Lin, and Dinghao Wu. A3FL: adversarially adaptive backdoor attacks to federated learning. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 61213-61233, 2023. 1, 2, 6", + "[44] Jiale Zhang, Chengcheng Zhu, Xiaobing Sun Di Wu, Jianming Yong, and Guodong Long. Badfss: backdoor attacks on federated self-supervised learning. In Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI), 2024. 2", + "[45] Jiale Zhang, Chengcheng Zhu, Chunpeng Ge, Chuan Ma, Yanchao Zhao, Xiaobing Sun, and Bing Chen. BadCleaner: Defending Backdoor Attacks in Federated Learning via Attention-Based Multi-Teacher Distillation. IEEE Transactions on Dependable and Secure Computing, 21(5):4559–4573, 2024. 2", + "[46] Jiale Zhang, Chengcheng Zhu, Xiaobing Sun, Chunpeng Ge, Bing Chen, Willy Susilo, and Shui Yu. FLPurifier: Backdoor Defense in Federated Learning via Decoupled Contrastive Training. IEEE Transactions on Information Forensics and Security, 19:4752-4766, 2024. 2", + "[47] Quan Zhang, Yifeng Ding, Yongqiang Tian, Jianmin Guo, Min Yuan, and Yu Jiang. Advdoor: adversarial backdoor attack of deep learning system. In Proceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis, pages 127-138, 2021. 2, 4", + "[48] Zhengming Zhang, Ashwinee Panda, Linyue Song, Yaoqing Yang, Michael Mahoney, Prateek Mittal, Ramchandran Kannan, and Joseph Gonzalez. Neurotoxin: Durable backdoors in federated learning. In International Conference on Machine Learning, pages 26429-26446. PMLR, 2022. 1, 2, 6", + "[49] Chengcheng Zhu, Jiale Zhang, Xiaobing Sun, Bing Chen, and Weizhi Meng. Adfl: Defending backdoor attacks in federated learning via adversarial distillation. Computers & Security, 132:103366, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25779", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/bab57aa6-5f36-4253-ae3a-1e5e1f4632d7_model.json b/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/bab57aa6-5f36-4253-ae3a-1e5e1f4632d7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ebd81d6d281d478ef598f621a1f598d05f50a5d5 --- /dev/null +++ b/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/bab57aa6-5f36-4253-ae3a-1e5e1f4632d7_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.131, + 0.885, + 0.155 + ], + "angle": 0, + "content": "Infighting in the Dark: Multi-Label Backdoor Attack in Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.18, + 0.812, + 0.253 + ], + "angle": 0, + "content": "Ye Li\\(^{1}\\), Yanchao Zhao\\(^{1*}\\), Chengcheng Zhu\\(^{2}\\), Jiale Zhang\\(^{2*}\\) \n\\(^{1}\\) Nanjing University of Aeronautics and Astronautics \\(^{2}\\) Yangzhou University \n{miles.li, yczhao}@nuaa.edu.cn \nchengchengzhu2022@126.com, jialezhang@yzu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.485, + 0.775 + ], + "angle": 0, + "content": "Federated Learning (FL), a privacy-preserving decentralized machine learning framework, has been shown to be vulnerable to backdoor attacks. Current research primarily focuses on the Single-Label Backdoor Attack (SBA), wherein adversaries share a consistent target. However, a critical fact is overlooked: adversaries may be non-cooperative, have distinct targets, and operate independently, which exhibits a more practical scenario called Multi-Label Backdoor Attack (MBA). Unfortunately, prior works are ineffective in the MBA scenario since non-cooperative attackers exclude each other. In this work, we conduct an in-depth investigation to uncover the inherent constraints of the exclusion: similar backdoor mappings are constructed for different targets, resulting in conflicts among backdoor functions. To address this limitation, we propose Mirage, the first non-cooperative MBA strategy in FL that allows attackers to inject effective and persistent backdoors into the global model without collusion by constructing in-distribution (ID) backdoor mapping. Specifically, we introduce an adversarial adaptation method to bridge the backdoor features and the target distribution in an ID manner. Additionally, we further leverage a constrained optimization method to ensure the ID mapping survives in the global training dynamics. Extensive evaluations demonstrate that Mirage outperforms various state-of-the-art attacks and bypasses existing defenses, achieving an average ASR greater than \\(97\\%\\) and maintaining over \\(90\\%\\) after 900 rounds. This work aims to alert researchers to this potential threat and inspire the design of effective defense mechanisms. Code has been made open-source." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.804, + 0.222, + 0.82 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.83, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Recent years have witnessed the flourishing of federated learning, an advanced privacy-preserving distributed learning paradigm [24, 26]. FL allows participants to collab" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.285, + 0.718, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.286, + 0.897, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.509 + ], + "angle": 0, + "content": "Figure 1. \"Blue\" and \"Red\" construct similar OOD mappings (see t-SNE), resulting in a conflict for the neuron weights. Only the dominant attacker can induce the model to output high feature attributions for its backdoor samples (see FA of \\( k \\)-th round) and successfully perform the attack. Similar in \\( t \\)-th round." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.594 + ], + "angle": 0, + "content": "oratively build a neural network model without leaking their data. Following its paradigm, a wide range of FL applications have emerged in various safety and privacy-critical application domains, including medical image analysis [10, 15], face recognition [30], and others [1, 11]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Despite enhanced privacy, the distributed nature of FL makes them more vulnerable to various adversarial attacks, with backdoor attacks posing the greatest threat. For backdoor attacks in FL, an attacker injects a backdoor into the local model, which is then submitted to the FL system, causing the global model to inherit the backdoor after aggregation. Consequently, the global model functions correctly on clean data but misclassifies inputs stamped with the attacker's chosen backdoor trigger as the target class. Numerous studies have been done to improve the persistence and effectiveness of backdoors [2, 6, 43, 48], which pose significant threats to various applications, including computer vision [40], IoT networks [28], and healthcare systems [16]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recent advancements in backdoor attacks in FL, most, if not all, assume that attackers in training are colluding and sharing the same target class, which can be denoted as Single-Label Backdoor Attacks (SBA). However, situations become more complex in real-world FL applications, especially in large-scale FL scenarios [18, 22], where attackers engage in training individually and compromise" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.242, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25770" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.363 + ], + "angle": 0, + "content": "the global model for their own purposes, i.e., Multi-Label Backdoor Attack (MBA). Intuitively, non-cooperative adversaries separately attack the global model and do not interfere with each other. Unfortunately, our empirical results indicate that directly applying existing attack methods to multiple adversaries leads to disorganized obstacles, making backdoors not as effective as they are in SBA scenarios. As illustrated in Fig. 1, two backdoors cannot be effective simultaneously because attackers construct similar Out-of-Distribution (OOD) backdoor mappings, indicating that their backdoor samples distribute together but outside the distributions of the target classes. This results in attackers excluding one another, with only the dominant attacker capable of embedding a backdoor in the global model, consequently causing the global model to output high feature attributions for its backdoor samples, thereby achieving a high attack success rate (ASR). It further raises a new problem: Can MBA successfully launch without collusion?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.368, + 0.483, + 0.7 + ], + "angle": 0, + "content": "To address the aforementioned issue, we first conduct an in-depth investigation to uncover the inherent constraints of the exclusions of SBAs: they create similar backdoor mappings for distinct targets, resulting in conflicts among backdoor functions. Motivated by this observation, we propose Mirage, the first non-cooperative MBA strategy in FL, wherein multiple adversaries individually undermine the global model without collusion. At a high level, instead of coordinating the various OOD mappings, Mirage aims to ensure that the backdoor samples are processed as clean samples by constructing an effective and persistent ID mapping between backdoor features and target distributions. To achieve this goal, we introduce a trigger optimization method based on adversarial adaptation, which adaptively optimizes the trigger to fool a carefully designed OOD sample detector, thereby constructing the ID mapping. Furthermore, considering that the distributions change with global training, we further introduce a constrained optimization method to enhance the persistence of ID mapping by tightening the distribution of backdoor features, allowing the backdoor mapping to live within the global training dynamics. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.81 + ], + "angle": 0, + "content": "- Uncovering the inherent constraints of extending existing SBAs to MBAs. To the best of our knowledge, we are the first to investigate and elucidate the inherent constraints of extending existing SBAs to MBAs. We underscore the practical threat posed by such attacks in large-scale FL applications by proposing a multi-label backdoor attack method based on the in-distribution backdoor." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.482, + 0.901 + ], + "angle": 0, + "content": "- Effective ID Mapping Construction. We propose a novel backdoor optimization method based on adversarial adaptation, which enables attackers to build the bridge between their backdoor and the distribution of their target class, allowing multiple adversaries to simultaneously launch backdoor attacks without interfering with others." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- Persistent ID Mapping Enhancement. We additionally introduce a constrained optimization method to enhance the ID mapping by minimizing the similarity of features between backdoorsed samples and benign samples. This tightens the ID backdoor mapping and ensures its persistence within the global training dynamics." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.273 + ], + "angle": 0, + "content": "- Evaluations and Discussions. Extensive experiments demonstrate that Mirage outperforms SOTA attacks across various settings. Furthermore, we discuss the vulnerability of existing defense methods to MBAs and present potential countermeasures against Mirage, highlighting the urgent demand for tailored defenses." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.289, + 0.655, + 0.305 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.314, + 0.866, + 0.331 + ], + "angle": 0, + "content": "2.1. Backdoor Attacks in Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.337, + 0.905, + 0.457 + ], + "angle": 0, + "content": "Backdoor attacks aim to compromise the victim model to produce the attacker-desired outputs if and only if the inputs are assigned to a specific trigger pattern. In light of this, a few researchers also start to notice the threat of backdoor attacks on FL systems. Depending on whether the backdoor trigger is optimized, current backdoor attacks fall under the categories of either static-trigger attacks [2, 6, 20, 44, 48] or trigger-optimization attacks [7, 25, 43, 47]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.459, + 0.905, + 0.76 + ], + "angle": 0, + "content": "The former pre-selected a fixed trigger pattern to poison the local training set and combine them with manually manipulating methods to improve the backdoor effectiveness. For example, [2, 3] aim to replace the global model by uploading a scaled-up poisoned model. In advance, [7, 34, 48] suggest leveraging the unimportant redundant neurons to prevent the backdoor from being erased shortly. Particularly, Dai et al. [6] introduce contrastive learning from the feature perspective of the backdoor to enhance its durability. The latter optimizes trigger patterns according to different tasks and generally achieves better effectiveness and utility. 3DFed [19] and A3FL [43] introduced the backdoor unlearning in an adversarial way to obtain a robust trigger to make the attacks more covert and persistent. DBA [41] and NBA [27] are the most similar work with us. However, DBA focuses on attacking the same target class, which is also an SBA situation. NBA investigated the noncooperative multi-label attack in FL but neither revealed the inner reason for the inapplicability of previous works nor proposed an effective method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.772, + 0.744, + 0.787 + ], + "angle": 0, + "content": "2.2. Backdoor Defenses in FL" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.795, + 0.905, + 0.855 + ], + "angle": 0, + "content": "Recently, researchers have noticed the threat of backdoor attacks in FL systems and actively explored various backdoor defense techniques, which can be categorized into either detection [8, 21, 39] or mitigation methods [4, 29, 45, 46, 49]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The detection-based methods aim to utilize anomaly detection techniques to determine whether a model is backdoored. Foolsgold [8] assigns lower aggregation weights" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25771" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.334 + ], + "angle": 0, + "content": "to updates with high pairwise cosine similarities, thereby mitigating the impact of backdoor updates. RFLBAT [39] discriminates the malicious model according to the difference between poisoned updates and clean updates in low-dimensional projection space. In advance, Li et al. [21] proposed BackdoorIndicator (referred to as \"Indicator\" hereafter), which detects potentially poisoned models based on the OOD properties of backdoor samples. Mitigation-based methods explored how to purify the backdoor model by eliminating backdoor triggers. MultiKrum [4] is a byzantine-robust aggregation protocol that aggregates the global model of having the smallest \\(\\mathcal{L}_2\\) distance. FLAME [29] adapts the weak DP method by noise boundary proof and a dynamic clipping bound, which is shown to alleviate the backdoor attack while still retaining a high main task accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.35, + 0.258, + 0.366 + ], + "angle": 0, + "content": "3. Attack intuitions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.484, + 0.438 + ], + "angle": 0, + "content": "In this section, we investigate the inherent constraints that affect the effectiveness of backdoors when extending previous works to MBA scenarios and further provide the attack intuitions based on this investigation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.45, + 0.411, + 0.465 + ], + "angle": 0, + "content": "3.1. Inherent constraints of SBA methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.485, + 0.837 + ], + "angle": 0, + "content": "Recall that, from the neuron activation perspective, the backdoor function in a model is highly related to the specific activation paths across neurons. This results in the model being sensitive to triggered inputs and producing anomalously high feature attribution values [9, 34, 36]. Correspondingly, such high values result in backdoor samples distributed out of the target clean distribution in feature space, i.e. the backdoor function constructs the OOD backdoor mapping [21]. In MBA scenarios, attackers may adopt similar strategies to construct the activation paths, such as using redundant neurons, which leads to the construction of similar OOD mappings and competition for neuron weights. Specifically, different attackers have varying requirements for the output values of specific neurons within the pathway. However, the dominant attacker controls most neurons, ensuring they produce the desired values, which results in others failing to achieve their backdoor function. As shown in Fig. 2a, two attackers aim to construct persistent and effective OOD backdoor mappings by targeting redundant neurons. Such a strategy causes them to leverage overlapped activation paths to achieve different functions, which further results in infighting among adversaries. This drives us to find a solution that allows attackers to select different activation paths non-cooperatively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.373, + 0.864 + ], + "angle": 0, + "content": "3.2. Attack intuitions and challenges" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To alleviate the exclusion, it is critical to design a mapping strategy that ensures different attackers can independent" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.648, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.193, + 0.628, + 0.205 + ], + "angle": 0, + "content": "(a) SBA's mapping." + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.089, + 0.781, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.193, + 0.767, + 0.205 + ], + "angle": 0, + "content": "(b) Separate mapping." + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.089, + 0.913, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.193, + 0.883, + 0.205 + ], + "angle": 0, + "content": "(c) ID mapping." + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.207, + 0.864, + 0.22 + ], + "angle": 0, + "content": "Figure 2. Illustrations of different mapping strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.221, + 0.907, + 0.582 + ], + "angle": 0, + "content": "mappings with no intersections. Achieving this is challenging because attackers share no information with others and are unaware of each other's existence. Surprisingly, our empirical investigation reveals that A3FL can fulfill the aforementioned requirement by combining machine unlearning [37, 42], which indirectly stimulates the eliminations performed by other attackers. As shown in Fig 2b, the unlearning method adversarially optimizes the trigger to construct a distinct OOD backdoor mapping that differs from other attackers. Although this method mitigates the exclusions, it enhances the OOD characteristics of the backdoor functions, consequently introducing more challenges in bypassing the OOD-based defense methods, such as Indicator [21]. The OOD-based defense method leads us to a counterintuitive thought: Can a backdoor be triggered via a clean activation path? In other words, construct the ID mapping for the backdoor features and target distributions (present in Fig 2c). Therefore, once the adversaries construct the ID backdoor mapping to the target class, they naturally alleviate potential exclusions, even without collusion, since additional measures are no longer needed to circumvent the backdoor activation paths of other attackers. However, constructing the ID backdoor mapping remains challenging due to the following issues:" + }, + { + "type": "title", + "bbox": [ + 0.534, + 0.583, + 0.863, + 0.598 + ], + "angle": 0, + "content": "Challenge 1: How to construct the ID mapping?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.764 + ], + "angle": 0, + "content": "While backdoor attacks have emerged as a prominent topic in FL, to the best of our knowledge, no prior studies investigated the ID backdoor within FL. Intuitively, the backdoor sample should be processed as a clean sample following the clean activation path, meaning it should share similar feature outputs. One promising way is leveraging feature alignment. However, this may disrupt the relationships among clean classes, making it challenging for attackers to maintain the ID characteristics. Therefore, there is a pressing need to devise a novel method that ensures the effective construction of ID mapping." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.765, + 0.906, + 0.796 + ], + "angle": 0, + "content": "Challenge 2: How to maintain the persistence of ID mapping?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "One key challenge for backdoor attacks in FL is that the continued dynamic clean updates may break the well-constructed backdoor mappings. Previous research focuses on leveraging neurons that are updated less frequently to build the neural activation path, which cannot be applied in ID mapping since our objective is to construct the bridge between the trigger pattern and the clean target activation path." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25772" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "Therefore, these maintenance requirements for ID mapping further exacerbate the challenges." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.135, + 0.226, + 0.153 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.16, + 0.483, + 0.221 + ], + "angle": 0, + "content": "To formulate the attack scenario, we first introduce the threat model. Driven by attack intuitions, we propose two methods that enable attackers to construct an efficient and persistent ID mapping without cooperation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.229, + 0.232, + 0.243 + ], + "angle": 0, + "content": "4.1. Threat model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.251, + 0.483, + 0.356 + ], + "angle": 0, + "content": "Adversaries' Capabilities and knowledge. We consider a FL system running image classification tasks and assume that adversaries are FL participants, which enables attackers to control their own local datasets and training processes. Note that adversaries do not have access to any information regarding benign clients or other adversaries, including but not limited to datasets, models, and target classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.402 + ], + "angle": 0, + "content": "Adversaries' goal. Adversaries aim to plant a backdoor into the global model that makes it misclassify backdoor data to a specific class while leaving other tasks unaffected." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.411, + 0.281, + 0.426 + ], + "angle": 0, + "content": "4.2. Overview of Mirage" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.735 + ], + "angle": 0, + "content": "Following the attack intuition, our objective is to construct an ID mapping to bridge the backdoor features and distribution of the target class, thereby freeing attackers from the exclusion of using similar backdoor activation paths. This presents two significant challenges: how to construct a qualified ID mapping and how to maintain it. To address these challenges, we propose Mirage, the first multilabel backdoor attack in FL, which enables attackers to noncooperatively compromise the global model by constructing an ID mapping between the backdoor features and target distributions. As illustrated in Fig. 3, Mirage introduces two components to construct an effective and persistent ID backdoor mapping. On the one hand, we propose a backdoor optimization method based on adversarial adaptation, which constructs the ID mapping by maximizing the misclassification probability of a well-designed backdoor sample detector. On the other hand, we deploy a constrained optimization strategy to tighten the backdoor distribution, ensuring its survival in the global training dynamics. Next, we provide details of each component." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.399, + 0.759 + ], + "angle": 0, + "content": "4.3. Effective ID Mapping Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address the first challenge, we aim to adversarially optimize the trigger pattern to make it able to construct the ID mapping. In order to optimize the trigger, we employ an OOD sample detector to determine whether a backdoored sample is an OOD sample and further optimize the trigger pattern to fool the detector. This introduces another problem: How can an attacker obtain an accurate detector? Previous research suggests training a new model as the detector [47]. However, retraining the detector with each trigger" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "update incurs significant computational overhead, and the local samples limit its effectiveness. These shortcomings prompt us to seek another solution to obtain a detector that is both effective and lightweight for updates. Recall that the FL training process allows clients to obtain the latest global model, which contains a well-trained feature extractor. Consequently, we consider leveraging the feature extractor of the global model as the primary component and training a binary classifier to distinguish the OOD samples. Such an approach is a common practice in transfer learning, consumes only a few computing resources, and requires little data." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.907, + 0.379 + ], + "angle": 0, + "content": "In order to obtain an accurate detector, we first prepare the dataset for training. The local dataset for each attacker can be divided into \\( D_{cln} = \\{(x,y) \\in D_{att} | + \\} \\) and \\( D_{bd} = \\{(x \\oplus \\delta, y) \\in D_{att} | - \\} \\), where \\( + \\) indicates the ground truth label of a sample is \\( c \\), and \\( - \\) is the opposite and \\( \\delta \\) represents the trigger pattern. Then, the detector dataset \\( D_{detector} \\) (\\( D_d \\) for short) is composed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.391, + 0.907, + 0.409 + ], + "angle": 0, + "content": "\\[\nD _ {d} = \\{(x, -) \\mid x \\in D _ {b d} \\} \\cup \\{(x, +) \\mid x \\in D _ {c l n} \\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.906, + 0.51 + ], + "angle": 0, + "content": "To make the ID mapping construction more effective and flexible, we adopt an adversarial training-like method to optimize the trigger pattern. Specifically, \\(\\theta_{detector}\\) is trained on the \\(D_{detector}\\) to differentiate between clean and backdoored samples by minimizing the binary cross-entropy (BCE) loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.906, + 0.602 + ], + "angle": 0, + "content": "Concurrently, the trigger is optimized to fool the detector by maximizing the binary classification loss. Such a min-max game equips the trigger with the ability to activate the clean activation path of the target class in the global model, naturally avoiding conflicts with other attackers. The min-max process is mathematically described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.614, + 0.905, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\delta = \\operatorname {a r g m i n} \\mathbb {E} _ {(x, y) \\sim D _ {\\text {d e t e c t o r}}} [ \\mathcal {L} (x, +, \\theta) | y = - ] \\\\ s. t. \\theta = \\underset {\\theta} {\\overset {\\delta} {\\operatorname {a r g m i n}}} \\mathbb {E} _ {(x, y) \\sim D _ {\\text {d e t e c t o r}}} [ \\mathcal {L} (x, y, \\theta) ] \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.748 + ], + "angle": 0, + "content": "where \\(\\theta\\) is the detector's parameters, with its feature extractor frozen and parameterized by the weights of the global model, and \\(\\mathcal{L}\\) is the BCE loss. For better implementation, we replace the maximization problem with minimizing the BCE loss of misclassification for the backdoored samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.835, + 0.774 + ], + "angle": 0, + "content": "4.4. Persistent ID Mapping Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Although using an OOD detector and an adversarial strategy restricts the trigger pattern \\(\\delta\\) to reveal the ID feature, our empirical results indicate that the backdoored samples still do not meet our expectations. This is because the adversarial strategy does not constrain the tightness of backdoor mapping, which causes the margin-distributed backdoored samples not in-distribution anymore with the dynamic changes of training and failing to activate the target" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25773" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.088, + 0.868, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.261, + 0.908, + 0.306 + ], + "angle": 0, + "content": "Figure 3. The workflow of Mirage. Step 1: Train an OOD sample detector \\(\\theta_{\\text{Detector}}\\). Step 2: Construct ID mapping by maximizing misclassification probabilities of backdoor samples on the detector. Step 3: Tighten backdoor distribution by minimizing the feature similarities between backdoor samples and benign samples. Step 4: Optimize the trigger by minimizing \\(L_{\\text{Enhance}}\\) and \\(L_{\\text{detector}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.316, + 0.485, + 0.574 + ], + "angle": 0, + "content": "activation path. One promising solution is to leverage distribution estimation methods, like KDE [33], to obtain the actual distribution of the target class and use data augmentation to generate samples to improve the effectiveness of the detector. Although it has been proven effective, it requires countless computing resources. Therefore, we introduce a constrained optimization method to enhance the ID mapping by tightening the distribution of backdoored samples. Since the adversarial strategy successfully constructs the ID mapping, we aim to \"push\" the backdoor distribution to make it tighter. Specifically, we minimize the similarities between clean samples \\( x \\) and their corresponding backdoored samples \\( x \\oplus \\delta \\), allowing the backdoored sample to deviate as much as possible from the original distribution. Additionally, we establish the direction by minimizing the loss of the backdoor samples on the global model. Formally, it can be formulated as the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.581, + 0.483, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\delta = \\operatorname {a r g m i n} \\mathrm {C S} \\left[ \\left(\\theta_ {f} (x), \\theta_ {f} (\\hat {x}) \\right] + \\mathrm {C E} (\\hat {x}, \\hat {y}, \\theta) \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.698 + ], + "angle": 0, + "content": "where \\(\\theta_{f}\\) represents the feature extractor of the global model \\(\\theta\\). \\(\\hat{x} = x \\oplus \\delta, x \\in D_{att}\\), \\(\\oplus\\) is the backdoored operator. \\(\\mathrm{CS}(\\cdot)\\) and \\(\\mathrm{CE}(\\cdot)\\) represent the Cosine-similarity and the Cross-entropy loss, separately. Therefore, the backdoor distribution can be tightened into the target distribution and live within the global training dynamics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.484, + 0.729 + ], + "angle": 0, + "content": "We summarized Mirage in Algorithm 1, and the detailed description is provided in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.224, + 0.758 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.87 + ], + "angle": 0, + "content": "In this section, we provide extensive experimental results to demonstrate the effectiveness of Mirage by comparing its performance with several SOTA backdoor attack methods from multiple angles. Our experiments run on a FL system performing image classification tasks, using an NVIDIA GeForce RTX 4090 GPU equipped with 24GB of memory. The detailed setups are listed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Dataset: We conduct experiments on three widely used public real-world datasets, CIFAR-10, CIFAR-100 [17] and" + }, + { + "type": "code_caption", + "bbox": [ + 0.526, + 0.32, + 0.772, + 0.335 + ], + "angle": 0, + "content": "Algorithm 1: Mirage at \\(i\\)-th attacker." + }, + { + "type": "algorithm", + "bbox": [ + 0.514, + 0.337, + 0.905, + 0.671 + ], + "angle": 0, + "content": "Data: \\(\\theta_{t},\\theta_{detector},\\theta_{fe},\\mathcal{D}_{i},\\delta ,\\eta ,c,E.\\) \n1 \\(\\theta_t^i = \\theta_t\\) \n2 /\\* Trigger adversarial adaption\\*/ \n3 for epoch \\(e = 1,\\ldots ,E\\) do \n4 Generate \\(D_{detector}\\) by Eq 1; \n5 \\(\\theta_{detector} =\\) argmin \\(\\mathbb{E}_{(x,y)\\sim D_{detector}}[\\mathcal{L}(x,y,\\theta_{detector})]\\) . \n6 for batch \\(\\mathcal{B}\\in D_i\\) do \n7 \\(\\begin{array}{r}\\mathcal{L}_{detector} = \\\\ \\frac{1}{|\\mathcal{B}|}\\sum_{(x,y)\\in \\mathcal{B}}\\mathcal{L}_{BCE}(x\\oplus \\delta , + ,\\theta_{detector});\\\\ \\mathcal{L}_{Enhance} = \\frac{1}{|\\mathcal{B}|}\\sum_{(x,y)\\in \\mathcal{B}}\\mathcal{L}_{CE}(x\\oplus \\delta ,\\hat{y},\\theta_t^i)\\\\ +\\mathcal{L}_{CS}((\\theta_{fe}(x),\\theta_{fe}(x\\oplus \\delta));\\\\ \\delta \\gets \\delta -\\eta \\cdot \\nabla_{\\delta}(\\mathcal{L}_{detector} + \\mathcal{L}_{Enhance});\\\\ end \\end{array}\\) \n10 end \n13 /\\* Local training\\*/ \n14 Get the poisoned local dataset \\(\\mathcal{D}_i^{poison}\\) with \\(\\delta\\) \n15 Local optimize \\(\\theta_t^\\prime\\) on \\(\\mathcal{D}_i^{poison}\\) . \n16 \\(\\Delta_i^{t + 1} = \\theta_t - \\theta_t'\\) . \n17 Upload \\(\\Delta_i^{t + 1}\\) to the server;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.703, + 0.848, + 0.719 + ], + "angle": 0, + "content": "GTSRB [13]. Details are provided in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Federated learning setup: By default, we set the number of FL clients \\( N = 100 \\). The server randomly selects \\( M = 10 \\) clients at each communication round to contribute to the global model. The default global model architecture is ResNet-18 [12], which is a classic and effective model for image classification. Based on previous research, we randomly split the dataset among clients non-IID using Dirichlet sampling [14] with a hyperparameter \\( \\alpha = 1 \\). For each selected benign client, they train the local model for two local epochs using an SGD optimizer with a learning rate of 0.01. The batch size is 64 for CIFAR-10 and CIFAR-100 and 32 for GTSRB. The entire training process continues" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25774" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.309, + 0.105 + ], + "angle": 0, + "content": "for 2100 communication rounds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.484, + 0.288 + ], + "angle": 0, + "content": "Attack and defense setup: We compare the effectiveness of Mirage with SOTA methods, including A3FL [43], Chameleon [6], Neurotoxin [48], PGD [38], and Vanilla [2]. We further extend the aforementioned methods to MBA scenario, wherein attackers independently execute their attacks. Unless otherwise specified, we set \\( N = 3 \\), with the target classes being 0, 1, and 2. In line with A3FL, attackers of Mirage optimize their triggers by PGD and poison \\( 12.5\\% \\) of \\( D_{att} \\). Blend [5] is utilized as a backdoor trigger across all attacks. The attack window starts from the 2000-th round and continues for 100 rounds, so attackers are forced to wait for selection and cannot launch a continuous attack." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.38 + ], + "angle": 0, + "content": "We also evaluate those attacks against six SOTA defenses: Multi-Krum [4], DeepSight [31], Foolsgold [8], RFLBAT [39], FLAME [29] and BakcdoorIndicator [21]. For those attacks and defense methods, we follow the original implementations. Note that we selected the random noise dataset as the source of the indicator dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.391, + 0.273, + 0.405 + ], + "angle": 0, + "content": "5.1. Evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Following previous works, we comprehensively evaluate Mirage using accuracy (Acc), attack success rate (ASR), and lifespan. Specifically, Acc reflects the attack stealthiness, which indicates the performance of the backdoor model on the main tasks. ASR represents the ratio of backdoors samples misclassified as the labels specified by the attackers. Note that we report the mean ASR of three attackers on the global model at the end of FL. To evaluate the persistence of our attack, we introduce lifespan, denoted as the period that starts at the end of the attack window and ends when the ASR decreases to below a chosen threshold." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.591, + 0.283, + 0.607 + ], + "angle": 0, + "content": "5.2. Attack performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.885 + ], + "angle": 0, + "content": "First, we evaluate the attack performance of Mirage across various datasets and compare it to different SOTA attack methods. The results are summarized in Table 1, and the best results are highlighted in Bold. The experiment shows that Mirage achieves remarkable results across three tasks and various SOTA defense methods compared to the baseline methods. In terms of accuracy, Mirage achieves the best accuracy in over half of the tasks and obtains suboptimal results for the remaining tasks, losing to the best ones by an average of only \\(0.498\\%\\). Regarding effectiveness, Mirage achieves the best ASRs across all the tasks, with the gap among attackers being acceptable (the maximum gap is \\(6.99\\%\\), and the average gap is \\(2.19\\%\\)). Those results indicate that the design of constructing the ID mapping between backdoor features and target class allows attackers to independently attack the FL model without concerns about the potential exclusion, resulting in a global model that excels in both main and multiple backdoor tasks." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Furthermore, A3FL is a relatively competitive method" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.903, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.215, + 0.806, + 0.228 + ], + "angle": 0, + "content": "Figure 4. Persistent evaluations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.907, + 0.365 + ], + "angle": 0, + "content": "that leverages unlearning to simulate the potential exclusion and optimize the trigger to construct robust activation paths independent of other attackers. However, as described in Sec. 3.2, such an approach may enhance its OOD mapping concerning the benign distribution and increase the risk of being detected by OOD-based detection methods, such as Indicator [21]. The results of A3FL with the defense of Indicator support our analysis. Besides, the Indicator does not perform as expected, we will discuss it in Section 6.3." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.374, + 0.618, + 0.39 + ], + "angle": 0, + "content": "5.3. Lifespan" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.907, + 0.593 + ], + "angle": 0, + "content": "Secondly, we evaluate the durability of attacks by comparing their lifespans. We extend the FL training process from 2100 to 3000 rounds while keeping the attack window to the first 100 rounds. Fig. 4 illustrates the attack success rate without any defense method. Mirage achieves similar results with A3FL, where the ASRs exceed \\(90\\%\\) at the end, indicating a lifespan of over 900 rounds. In contrast, the ASRs of all other baseline attacks are constrained by potential exclusion between attackers, as described before, making them ineffective and falling rapidly. These experimental results suggest that Mirage can achieve comparable durability to A3FL, which focuses on enhancing the persistence of backdoors in FL." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.787, + 0.62 + ], + "angle": 0, + "content": "5.4. The impact of different models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.839 + ], + "angle": 0, + "content": "In this section, we evaluate Mirage across different model architectures. Following previous research, we select widely used image classification model architectures: ResNet18 and ResNet34 [12], VGG11 and VGG19 [35], and MobileNet-V2 [32]. We set the parameters to default and only change the model architectures. As shown in Tab. 2, the attack performance of Mirage remains high in ASRs. The decrease in MobileNet-V2 is mainly attributed to the low utility of its model performance, which significantly reduces the detector's utility, resulting in sub-optimal ID mapping construction for certain attackers. However, the average ASRs remain acceptable. In conclusion, based on the results presented above, we can conclude that Mirage is not sensitive to specific model structures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.805, + 0.865 + ], + "angle": 0, + "content": "5.5. The impact of adversary number" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We also study the impact of the number of adversaries, which can be considered as the number of different targets" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "25775" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.296 + ], + "angle": 0, + "content": "
DatasetDefenseVanillaPGDNeurotoxinChameleonA3FLMirage
Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)
CIFAR-10nodefense91.7631.88*87.6891.5231.25*92.7892.0352.99*90.2992.3539.65*79.9891.7399.52*0.2392.1699.54*0.27
deepsight91.4431.71*88.2990.8131.083*92.6392.3139.24*83.5092.3236.66*79.6692.1796.56*7.7491.2897.39*3.82
foolsgold88.7932.50*89.5090.7131.67*93.9691.0480.91*21.8390.6645.4667.2589.0095.30*12.0291.3297.35*2.51
Indicator85.0828.7685.3285.8320.6059.5085.1329.9860.32*86.138.215.1385.7570.33*88.4591.1093.46*3.90
multikrum92.361.594.5692.391.183.2592.641.895.1892.461.905.2592.1878.15*64.4192.1092.30*3.12
rfbat86.8330.5588.3488.2833.27*99.7987.2233.27*73.1286.9737.1586.4285.4687.83*10.2185.7497.29*2.11
Flame92.561.835.2692.513.399.9092.0240.97*94.0492.281.383.7192.2579.40*60.2392.3695.9*2.42
CIFAR-100nodefense69.4032.62*94.3170.5332.28*93.8971.0732.62*97.0571.6624.0727.5071.2698.67*2.4771.6599.05*0.77
deepsight70.0232.86*89.0270.5432.23*96.3270.5333.14*93.2871.3024.0626.1771.6892.05*18.4471.5698.67*2.43
foolsgold70.0233.3089.8170.1532.65*97.3471.0732.63*97.0571.6024.2027.1171.7898.71*1.6471.9499.18*1.00
Indicator67.736.8020.1263.855.8817.2860.9811.8932.3461.356.0713.4365.3537.13*92.7268.2299.8*0.10
multikrum71.810.380.7771.140.310.5971.630.360.6671.850.210.3871.7691.28*18.5571.8693.57*5.14
rfbat62.360.300.4163.100.600.1262.150.561.0061.4918.0127.3161.3594.34*16.5663.1995.68*6.99
Flame71.750.210.2872.090.330.6272.020.270.4572.100.300.5271.2291.48*19.8371.8294.22*3.16
GTSRBnodefense96.1033.32*98.2496.7333.10*98.8095.9338.39*95.7996.0845.62*87.0696.6599.63*0.4996.9799.73*0.48
deepsight95.9533.8*97.4495.9733.36*98.3596.1233.06*99.1796.1450.1476.7195.9594.78*12.6496.2096.76*2.53
foolsgold96.0833.47*97.9296.9433.06*98.7595.8837.57*96.3696.0451.9178.0696.6297.68*6.9596.9599.88*0.27
Indicator91.5750.9291.5389.4527.2778.1489.0817.0950.9792.912.765.3885.5185.05*25.2695.1299.73*0.70
multikrum96.244.1412.2096.972.747.7296.124.8414.3296.134.7213.9796.6298.18*4.1696.9298.99*1.73
rfbat94.4056.61*90.6289.4433.26*99.7294.5333.88*97.8094.7554.39*83.0495.2397.86*6.4294.2199.97*0.08
Flame95.886.1918.4596.972.747.7295.925.1515.3295.925.9517.7196.5299.31*1.5296.9299.32*1.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.306, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Table 1. Performance of Mirage compared with baseline attack methods against the SOTA defense methods. The Acc higher represents better, and so does the ASR. \"GAP\" represents the difference between the highest and the lowest ASR, and attack methods should maintain their GAP close to zero. An \" * \" after ASR indicates that at least one of the three attackers achieved over \\( {90}\\% \\) ASR." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.36, + 0.452, + 0.405 + ], + "angle": 0, + "content": "
ResNet18ResNet34VGG11VGG19MobileNet-V2
Acc92.1693.8991.1492.0785.38
ASR98.899.6399.2199.8297.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.141, + 0.415, + 0.433, + 0.428 + ], + "angle": 0, + "content": "Table 2. Results on different model architectures." + }, + { + "type": "table", + "bbox": [ + 0.144, + 0.433, + 0.432, + 0.524 + ], + "angle": 0, + "content": "
Dataset0.515101000
CIFAR100Acc71.4671.6571.6871.5071.81
ASR99.5999.0599.1699.2699.21
GTSRBAcc96.9696.9796.6296.5796.70
ASR99.5499.7399.4798.2698.86
CIFAR10Acc91.6192.1692.4392.5192.38
ASR98.9598.8099.0399.3999.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.15, + 0.535, + 0.424, + 0.549 + ], + "angle": 0, + "content": "Table 3. Results on different non-IID settings." + }, + { + "type": "table", + "bbox": [ + 0.145, + 0.553, + 0.432, + 0.605 + ], + "angle": 0, + "content": "
Attacker Number12345
Acc92.3792.5392.1692.1192.12
ASR99.0699.1598.8098.37598.484
GAP0.000.490.381.421.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.615, + 0.428, + 0.629 + ], + "angle": 0, + "content": "Table 4. Mirage with different attacker number." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.839 + ], + "angle": 0, + "content": "since we assign different targets to the attackers. Table 4 presents the experimental results on CIFAR-10 with varying attack numbers ranging from 1 to 5, evaluating the performance of the global model on the main tasks and different backdoor tasks. Note that other settings follow the default configuration. We observe that the increase in attack numbers has a negligible impact on Mirage, with the accuracy of main tasks maintained above \\(92\\%\\) and the average ASR exceeding \\(98\\%\\). The experimental results indicate that backdoor injection without collusion can be effectively implemented by constructing ID mapping. We also conducted experiments on different datasets (see Appendix D), and the results support the above conclusions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.468, + 0.866 + ], + "angle": 0, + "content": "5.6. Performance under varying non-IID degrees" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Tab. 3 exhibits the attack performance of Mirage under varying non-IID degrees. We set different sampling param" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.501 + ], + "angle": 0, + "content": "eters \\(\\alpha\\) for Mirage, with larger values indicating a distribution that is closer to IID. The results indicate that Mirage can achieve stable performance across different non-IID degrees. As described in Sec. 4.4, the adversarial strategy may optimize the backdoor distribution to the margin of the target distribution. To address this issue, we introduce a constrained optimization method, tightening the backdoor distribution into the objective distribution and ensuring successful attacks across different distributions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.511, + 0.829, + 0.527 + ], + "angle": 0, + "content": "5.7. Illustration of backdoor distribution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.907, + 0.807 + ], + "angle": 0, + "content": "The preceding experiments leverage t-SNE to illustrate the backdoor distribution for Mirage and baseline methods. The results are in Fig. 5. The backdoor distribution from Fig. 5(a) to Fig. 5(d) further corroborates our presentation in 1, that different attackers may share similar neuron activation paths, resulting in their potential mitigation since their distribution is close to each other. On the other aspect, A3FL paves the activation path by adversariably unlearning the backdoor, which enhances their OOD mapping, producing clear boundaries between classes and tight clustering within each class. Although effective, this poses more threats to A3FL since the OOD-sample-based detection method can detect their backdoor models effectively. Looking back to Mirage, which is pictured in Fig. 5(f), backdoor samples present in-distribution with the target clean distribution. The results further indicate that the proposed Mirage can construct the ID mapping between backdoor features and clean distributions of the target class." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.665, + 0.834 + ], + "angle": 0, + "content": "5.8. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Sections 4.3 and 4.4 propose two hybrid techniques to construct and enhance the ID mapping. Here, we conduct an ablation study to show the effects of different techniques for our method. The experiment adheres to the settings" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25776" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.089, + 0.895, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.19, + 0.635, + 0.202 + ], + "angle": 0, + "content": "Figure 5. Illustration of backdoor distribution." + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.203, + 0.429, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.289, + 0.362, + 0.302 + ], + "angle": 0, + "content": "Figure 6. Ablation study." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.485 + ], + "angle": 0, + "content": "outlined in Section 5.3. Fig. 6 presents the average ASR for Mirage and Mirage without enhancement. We can see that introducing the ID mapping enhancement significantly improves Mirage's performance. As described in Section 4.4, although the adversarial strategy successfully constructs the ID mapping, the backdoors may not fully activate the clean path, making the backdoor distribution map to the target distribution's margin, rendering it ineffective and non-persistent. Therefore, we introduce the constrained optimization method to tighten the backdoor distribution, ensuring backdoor samples will not be misidentified as the target distribution changes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.5, + 0.206, + 0.515 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.525, + 0.401, + 0.542 + ], + "angle": 0, + "content": "6.1. Does the pixel trigger still effective?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.484, + 0.73 + ], + "angle": 0, + "content": "In the previous experiments, we adopted the blend trigger as the trigger pattern. Here, we conduct experiments with pixel triggers to evaluate their effectiveness. The results (presented in Appendix E) show that the pixel trigger remains effective. According to the t-SNE analysis of the feature attribution, the backdoor samples cannot aggregate together like the blend trigger. The main reason is that the pixel trigger can only activate a subset of neurons along the activation path, while other features typically activate the paths belonging to their original classes. However, it still constructs the ID mapping to the target distribution and successfully embeds the backdoors for different attackers." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.258, + 0.759 + ], + "angle": 0, + "content": "6.2. One-shot Mirage" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Mirage can perform a one-shot attack by increasing adversarial training epochs, allowing the trigger to directly simulate features of the target class, like adversarial samples. However, it does not construct persistent and effective ID backdoor mapping, resulting in low durability and side effects on the main tasks. Therefore, we enable the model to learn from the backdoor samples and construct effective and persistent ID mapping by injecting more rounds and fewer batches to achieve high-confidence classification." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.204, + 0.75, + 0.218 + ], + "angle": 0, + "content": "6.3. Potential defense methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.226, + 0.907, + 0.421 + ], + "angle": 0, + "content": "According to Section 5.2, we can see that although Indicator [2], the most powerful detection method for now, can detect at least one of the three A3FL attackers, the rest can still successfully inject the backdoor. This is because it constructs limited OOD mappings, succeeding only under the assumption that the attacker constructs similar OOD mappings. Therefore, it cannot be as effective against dynamic attacks (A3FL) as it is against static attacks (Chameleon, etc.). Furthermore, for A3FL, such misdetections are sufficient for the attacker to inject a persistent backdoor into the global model. Thus, following the works with foresight [23, 27], we urge researchers to focus on MBA and expand their efforts to defend against it." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.906, + 0.543 + ], + "angle": 0, + "content": "Our proposed Mirage does not follow the traditional backdoor activation scheme, making it challenging for previous work to resist. This motivates us to seek solutions to address this threat to real-world FL applications. The advanced explainable method may effectively detect our attack, however, detecting all backdoored models remains challenging. Input detection may also identify our attack, as clean samples differ from backdoored samples." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.557, + 0.634, + 0.572 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.808 + ], + "angle": 0, + "content": "In this paper, we uncover the inherent constraints of SBAs when extending them to MBA scenarios: Non-cooperative attackers will exclude others as they construct similar backdoor mappings. Based on this observation, we introduce Mirage, the first non-cooperative MBA strategy in FL. Mirage ensures the backdoor functions of different attackers by constructing effective and persistent ID backdoor mappings. Specifically, an adversarial trigger optimization method is introduced to construct the ID mapping between backdoor features and target distributions. Subsequently, Mirage leverages a constrained optimization method to tighten the ID mapping for persistence. Our comprehensive experiments demonstrate that Mirage outperforms existing backdoor attacks across different settings. For future work, investigating defense methods will be our primary focus." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.809, + 0.648, + 0.823 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.824, + 0.906, + 0.883 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant(No. 62172215) and in part by the A3 Foresight Program of NSFC (Grant No. 62061146002)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25777" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.157 + ], + "angle": 0, + "content": "[1] Apple. Private federated learning (neurips 2019 expo talk abstract). https://nips.cc/ExpoConferences/2019/schedule?talkid=40. accessed: 2020-05-22. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov. How to backdoor federated learning. In International conference on artificial intelligence and statistics, pages 2938-2948. PMLR, 2020. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Arjun Nitin Bhagoji, Supriyo Chakraborty, Prateek Mittal, and Seraphin Calo. Analyzing federated learning through an adversarial lens. In International conference on machine learning, pages 634-643. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.483, + 0.356 + ], + "angle": 0, + "content": "[4] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[5] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[6] Yanbo Dai and Songze Li. Chameleon: Adapting to peer images for planting durable backdoors in federated learning. In International Conference on Machine Learning, pages 6712-6725. PMLR, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.483, + 0.528 + ], + "angle": 0, + "content": "[7] Pei Fang and Jinghui Chen. On the vulnerability of backdoor defenses for federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11800-11808, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[8] Clement Fung, Chris J. M. Yoon, and Ivan Beschastnikh. The limitations of federated learning in sybil settings. In 23rd International Symposium on Research in Attacks, Intrusions and Defenses, RAID 2020, San Sebastian, Spain, October 14-15, 2020, pages 301-316. USENIX Association, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[9] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. BadNets: Evaluating Backdooring Attacks on Deep Neural Networks. IEEE Access, 7:47230-47244, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[10] Hao Guan, Pew-Thian Yap, Andrea Bozoki, and Mingxia Liu. Federated learning for medical image analysis: A survey. Pattern Recognition, page 110424, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[11] Andrew Hard, Kanishka Rao, Rajiv Mathews, Swaroop Ramaswamy, Françoise Beaufays, Sean Augenstein, Hubert Eichner, Chloe Kiddon, and Daniel Ramage. Federated Learning for Mobile Keyboard Prediction, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Sebastian Houben, Johannes Stallkamp, Jan Salmen, Marc Schlipsing, and Christian Igel. Detection of traffic signs in real-world images: The german traffic sign detection benchmark. In The 2013 international joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2013. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[14] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[15] Meirui Jiang, Zirui Wang, and Qi Dou. Harmofl: Harmonizing local and global drifts in federated learning on heterogeneous medical images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1087-1095, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[16] Ruinan Jin and Xiaoxiao Li. Backdoor attack and defense in federated generative adversarial network-based medical image synthesis. Medical Image Analysis, 90:102965, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[18] Fan Lai, Yinwei Dai, Sanjay Singapuram, Jiachen Liu, Xi-angfeng Zhu, Harsha Madhyastha, and Mosharaf Chowdhury. Fedscale: Benchmarking model and system performance of federated learning at scale. In International conference on machine learning, pages 11814–11827. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[19] Haoyang Li, Qingqing Ye, Haibo Hu, Jin Li, Leixia Wang, Chengfang Fang, and Jie Shi. 3dfed: Adaptive and extensible framework for covert backdoor attack in federated learning. In 44th IEEE Symposium on Security and Privacy, SP 2023, San Francisco, CA, USA, May 21-25, 2023, pages 1893-1907. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[20] Minghui Li, Wei Wan, Yuxuan Ning, Shengshan Hu, Lulu Xue, Leo Yu Zhang, and Yichen Wang. Darkfed: A data-free backdoor attack in federated learning. arXiv preprint arXiv:2405.03299, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[21] Songze Li and Yanbo Dai. BackdoorIndicator: Leveraging OOD data for proactive backdoor detection in federated learning. In 33rd USENIX Security Symposium (USENIX Security 24), pages 4193-4210, Philadelphia, PA, 2024. USENIX Association. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[22] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[23] Yige Li, Jiabo He, Hanxun Huang, Jun Sun, and Xingjun Ma. Shortcuts everywhere and nowhere: Exploring multi-trigger backdoor attacks, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[24] Ye Li, Jiale Zhang, Yanchao Zhao, Bing Chen, and Shui Yu. Fairness-aware federated learning framework on heterogeneous data distributions. In ICC 2024-IEEE International Conference on Communications, pages 728-733. IEEE, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[25] Xiaoting Lyu, Yufei Han, Wei Wang, Jingkai Liu, Bin Wang, Jiqiang Liu, and Xiangliang Zhang. Poisoning with cerberus: Stealthy and colluded backdoor attack against federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9020-9028, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25778" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "on Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[27] Tuan Nguyen, Dung Thuy Nguyen, Khoa D Doan, and Kok-Seng Wong. Non-cooperative backdoor attacks in federated learning: A new threat landscape. arXiv preprint arXiv:2407.07917, 2024. 2, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[28] Thien Duc Nguyen, Phillip Rieger, Markus Miettinen, and Ahmad-Reza Sadeghi. Poisoning attacks on federated learning-based IoT intrusion detection system. In Proc. Workshop Decentralized IoT Syst. Secur.(DISS), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[29] Thien Duc Nguyen, Phillip Rieger, Huili Chen, Hossein Yalame, Helen Möllering, Hossein Fereidooni, Samuel Marchal, Markus Miettinen, Azalia Mirhoseini, Shaza Zeitouni, Farinaz Koushanfar, Ahmad-Reza Sadeghi, and Thomas Schneider. FLAME: taming backdoors in federated learning. In 31st USENIX Security Symposium, USENIX Security 2022, Boston, MA, USA, August 10-12, 2022, pages 1415-1432. USENIX Association, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[30] Yifan Niu and Weihong Deng. Federated learning for face recognition with gradient correction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1999–2007, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[31] Phillip Rieger, Thien Duc Nguyen, Markus Miettinen, and Ahmad-Reza Sadeghi. Deepsight: Mitigating backdoor attacks in federated learning through deep model inspection. In 29th Annual Network and Distributed System Security Symposium, NDSS 2022, San Diego, California, USA, April 24-28, 2022. The Internet Society, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[32] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE conference on computer vision and pattern recognition*, pages 4510–4520, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[33] Simon J Sheather and Michael C Jones. A reliable data-based bandwidth selection method for kernel density estimation. Journal of the Royal Statistical Society: Series B (Methodological), 53(3):683-690, 1991. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[34] Chenghui Shi, Shouling Ji, Xudong Pan, Xuhong Zhang, Mi Zhang, Min Yang, Jun Zhou, Jianwei Yin, and Ting Wang. Towards practical backdoor attacks on federated learning systems. IEEE Transactions on Dependable and Secure Computing, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[36] Guanhong Tao, Zhenting Wang, Shiwei Feng, Guangyu Shen, Shiqing Ma, and Xiangyu Zhang. DRUPE: Distribution Preserving Backdoor Attack in Self-supervised Learning. In 2024 IEEE Symposium on Security and Privacy (SP), pages 29-29. IEEE Computer Society, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE symposium on security and privacy (SP), pages 707-723. IEEE, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[38] Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris S. Papailiopoulos. Attack of the tails: Yes, you really can backdoor federated learning. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[39] Yongkang Wang, Dihua Zhai, Yufeng Zhan, and Yuanqing Xia. RFLBAT: A robust federated learning algorithm against backdoor attack. CoRR, abs/2201.03772, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[40] Emily Wenger, Josephine Passananti, Arjun Nitin Bhagoji, Yuanshun Yao, Haitao Zheng, and Ben Y Zhao. Backdoor attacks against deep learning systems in the physical world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6206-6215, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[41] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. DBA: distributed backdoor attacks against federated learning. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[42] Yi Zeng, Si Chen, Won Park, Z Morley Mao, Ming Jin, and Ruoxi Jia. Adversarial unlearning of backdoors via implicit hypergradient. arXiv preprint arXiv:2110.03735, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[43] Hangfan Zhang, Jinyuan Jia, Jinghui Chen, Lu Lin, and Dinghao Wu. A3FL: adversarially adaptive backdoor attacks to federated learning. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 61213-61233, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[44] Jiale Zhang, Chengcheng Zhu, Xiaobing Sun Di Wu, Jianming Yong, and Guodong Long. Badfss: backdoor attacks on federated self-supervised learning. In Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[45] Jiale Zhang, Chengcheng Zhu, Chunpeng Ge, Chuan Ma, Yanchao Zhao, Xiaobing Sun, and Bing Chen. BadCleaner: Defending Backdoor Attacks in Federated Learning via Attention-Based Multi-Teacher Distillation. IEEE Transactions on Dependable and Secure Computing, 21(5):4559–4573, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[46] Jiale Zhang, Chengcheng Zhu, Xiaobing Sun, Chunpeng Ge, Bing Chen, Willy Susilo, and Shui Yu. FLPurifier: Backdoor Defense in Federated Learning via Decoupled Contrastive Training. IEEE Transactions on Information Forensics and Security, 19:4752-4766, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[47] Quan Zhang, Yifeng Ding, Yongqiang Tian, Jianmin Guo, Min Yuan, and Yu Jiang. Advdoor: adversarial backdoor attack of deep learning system. In Proceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis, pages 127-138, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[48] Zhengming Zhang, Ashwinee Panda, Linyue Song, Yaoqing Yang, Michael Mahoney, Prateek Mittal, Ramchandran Kannan, and Joseph Gonzalez. Neurotoxin: Durable backdoors in federated learning. In International Conference on Machine Learning, pages 26429-26446. PMLR, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Chengcheng Zhu, Jiale Zhang, Xiaobing Sun, Bing Chen, and Weizhi Meng. Adfl: Defending backdoor attacks in federated learning via adversarial distillation. Computers & Security, 132:103366, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25779" + } + ] +] \ No newline at end of file diff --git a/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/bab57aa6-5f36-4253-ae3a-1e5e1f4632d7_origin.pdf b/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/bab57aa6-5f36-4253-ae3a-1e5e1f4632d7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e8f1aac7d5b7983c1937c28cfd9b1fe8a32be969 --- /dev/null +++ b/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/bab57aa6-5f36-4253-ae3a-1e5e1f4632d7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f91b4f4d7dc33da48648e628ca3e9a867546a1cabda2de17866c5d1f0f2c3902 +size 3453568 diff --git a/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/full.md b/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3ac6a88f95c6d10b00459595c9fe0d2a4b24ba1e --- /dev/null +++ b/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/full.md @@ -0,0 +1,316 @@ +# Infighting in the Dark: Multi-Label Backdoor Attack in Federated Learning + +Ye Li $^{1}$ , Yanchao Zhao $^{1*}$ , Chengcheng Zhu $^{2}$ , Jiale Zhang $^{2*}$ $^{1}$ Nanjing University of Aeronautics and Astronautics $^{2}$ Yangzhou University +{miles.li, yczhao}@nuaa.edu.cn +chengchengzhu2022@126.com, jialezhang@yzu.edu.cn + +# Abstract + +Federated Learning (FL), a privacy-preserving decentralized machine learning framework, has been shown to be vulnerable to backdoor attacks. Current research primarily focuses on the Single-Label Backdoor Attack (SBA), wherein adversaries share a consistent target. However, a critical fact is overlooked: adversaries may be non-cooperative, have distinct targets, and operate independently, which exhibits a more practical scenario called Multi-Label Backdoor Attack (MBA). Unfortunately, prior works are ineffective in the MBA scenario since non-cooperative attackers exclude each other. In this work, we conduct an in-depth investigation to uncover the inherent constraints of the exclusion: similar backdoor mappings are constructed for different targets, resulting in conflicts among backdoor functions. To address this limitation, we propose Mirage, the first non-cooperative MBA strategy in FL that allows attackers to inject effective and persistent backdoors into the global model without collusion by constructing in-distribution (ID) backdoor mapping. Specifically, we introduce an adversarial adaptation method to bridge the backdoor features and the target distribution in an ID manner. Additionally, we further leverage a constrained optimization method to ensure the ID mapping survives in the global training dynamics. Extensive evaluations demonstrate that Mirage outperforms various state-of-the-art attacks and bypasses existing defenses, achieving an average ASR greater than $97\%$ and maintaining over $90\%$ after 900 rounds. This work aims to alert researchers to this potential threat and inspire the design of effective defense mechanisms. Code has been made open-source. + +# 1. Introduction + +Recent years have witnessed the flourishing of federated learning, an advanced privacy-preserving distributed learning paradigm [24, 26]. FL allows participants to collab + +![](images/0ddbdf872f33853402ca79f8266fb5f8851611c6766853843c188eca1dfb8883.jpg) +Figure 1. "Blue" and "Red" construct similar OOD mappings (see t-SNE), resulting in a conflict for the neuron weights. Only the dominant attacker can induce the model to output high feature attributions for its backdoor samples (see FA of $k$ -th round) and successfully perform the attack. Similar in $t$ -th round. + +![](images/dc5b1c26396d5cda41147e7343923cf1fa2364fec9b37fd99cf135cb8c3d2f5f.jpg) + +oratively build a neural network model without leaking their data. Following its paradigm, a wide range of FL applications have emerged in various safety and privacy-critical application domains, including medical image analysis [10, 15], face recognition [30], and others [1, 11]. + +Despite enhanced privacy, the distributed nature of FL makes them more vulnerable to various adversarial attacks, with backdoor attacks posing the greatest threat. For backdoor attacks in FL, an attacker injects a backdoor into the local model, which is then submitted to the FL system, causing the global model to inherit the backdoor after aggregation. Consequently, the global model functions correctly on clean data but misclassifies inputs stamped with the attacker's chosen backdoor trigger as the target class. Numerous studies have been done to improve the persistence and effectiveness of backdoors [2, 6, 43, 48], which pose significant threats to various applications, including computer vision [40], IoT networks [28], and healthcare systems [16]. + +Recent advancements in backdoor attacks in FL, most, if not all, assume that attackers in training are colluding and sharing the same target class, which can be denoted as Single-Label Backdoor Attacks (SBA). However, situations become more complex in real-world FL applications, especially in large-scale FL scenarios [18, 22], where attackers engage in training individually and compromise + +the global model for their own purposes, i.e., Multi-Label Backdoor Attack (MBA). Intuitively, non-cooperative adversaries separately attack the global model and do not interfere with each other. Unfortunately, our empirical results indicate that directly applying existing attack methods to multiple adversaries leads to disorganized obstacles, making backdoors not as effective as they are in SBA scenarios. As illustrated in Fig. 1, two backdoors cannot be effective simultaneously because attackers construct similar Out-of-Distribution (OOD) backdoor mappings, indicating that their backdoor samples distribute together but outside the distributions of the target classes. This results in attackers excluding one another, with only the dominant attacker capable of embedding a backdoor in the global model, consequently causing the global model to output high feature attributions for its backdoor samples, thereby achieving a high attack success rate (ASR). It further raises a new problem: Can MBA successfully launch without collusion? + +To address the aforementioned issue, we first conduct an in-depth investigation to uncover the inherent constraints of the exclusions of SBAs: they create similar backdoor mappings for distinct targets, resulting in conflicts among backdoor functions. Motivated by this observation, we propose Mirage, the first non-cooperative MBA strategy in FL, wherein multiple adversaries individually undermine the global model without collusion. At a high level, instead of coordinating the various OOD mappings, Mirage aims to ensure that the backdoor samples are processed as clean samples by constructing an effective and persistent ID mapping between backdoor features and target distributions. To achieve this goal, we introduce a trigger optimization method based on adversarial adaptation, which adaptively optimizes the trigger to fool a carefully designed OOD sample detector, thereby constructing the ID mapping. Furthermore, considering that the distributions change with global training, we further introduce a constrained optimization method to enhance the persistence of ID mapping by tightening the distribution of backdoor features, allowing the backdoor mapping to live within the global training dynamics. In summary, our contributions are as follows: + +- Uncovering the inherent constraints of extending existing SBAs to MBAs. To the best of our knowledge, we are the first to investigate and elucidate the inherent constraints of extending existing SBAs to MBAs. We underscore the practical threat posed by such attacks in large-scale FL applications by proposing a multi-label backdoor attack method based on the in-distribution backdoor. +- Effective ID Mapping Construction. We propose a novel backdoor optimization method based on adversarial adaptation, which enables attackers to build the bridge between their backdoor and the distribution of their target class, allowing multiple adversaries to simultaneously launch backdoor attacks without interfering with others. + +- Persistent ID Mapping Enhancement. We additionally introduce a constrained optimization method to enhance the ID mapping by minimizing the similarity of features between backdoorsed samples and benign samples. This tightens the ID backdoor mapping and ensures its persistence within the global training dynamics. +- Evaluations and Discussions. Extensive experiments demonstrate that Mirage outperforms SOTA attacks across various settings. Furthermore, we discuss the vulnerability of existing defense methods to MBAs and present potential countermeasures against Mirage, highlighting the urgent demand for tailored defenses. + +# 2. Related Work + +# 2.1. Backdoor Attacks in Federated Learning + +Backdoor attacks aim to compromise the victim model to produce the attacker-desired outputs if and only if the inputs are assigned to a specific trigger pattern. In light of this, a few researchers also start to notice the threat of backdoor attacks on FL systems. Depending on whether the backdoor trigger is optimized, current backdoor attacks fall under the categories of either static-trigger attacks [2, 6, 20, 44, 48] or trigger-optimization attacks [7, 25, 43, 47]. + +The former pre-selected a fixed trigger pattern to poison the local training set and combine them with manually manipulating methods to improve the backdoor effectiveness. For example, [2, 3] aim to replace the global model by uploading a scaled-up poisoned model. In advance, [7, 34, 48] suggest leveraging the unimportant redundant neurons to prevent the backdoor from being erased shortly. Particularly, Dai et al. [6] introduce contrastive learning from the feature perspective of the backdoor to enhance its durability. The latter optimizes trigger patterns according to different tasks and generally achieves better effectiveness and utility. 3DFed [19] and A3FL [43] introduced the backdoor unlearning in an adversarial way to obtain a robust trigger to make the attacks more covert and persistent. DBA [41] and NBA [27] are the most similar work with us. However, DBA focuses on attacking the same target class, which is also an SBA situation. NBA investigated the noncooperative multi-label attack in FL but neither revealed the inner reason for the inapplicability of previous works nor proposed an effective method. + +# 2.2. Backdoor Defenses in FL + +Recently, researchers have noticed the threat of backdoor attacks in FL systems and actively explored various backdoor defense techniques, which can be categorized into either detection [8, 21, 39] or mitigation methods [4, 29, 45, 46, 49]. + +The detection-based methods aim to utilize anomaly detection techniques to determine whether a model is backdoored. Foolsgold [8] assigns lower aggregation weights + +to updates with high pairwise cosine similarities, thereby mitigating the impact of backdoor updates. RFLBAT [39] discriminates the malicious model according to the difference between poisoned updates and clean updates in low-dimensional projection space. In advance, Li et al. [21] proposed BackdoorIndicator (referred to as "Indicator" hereafter), which detects potentially poisoned models based on the OOD properties of backdoor samples. Mitigation-based methods explored how to purify the backdoor model by eliminating backdoor triggers. MultiKrum [4] is a byzantine-robust aggregation protocol that aggregates the global model of having the smallest $\mathcal{L}_2$ distance. FLAME [29] adapts the weak DP method by noise boundary proof and a dynamic clipping bound, which is shown to alleviate the backdoor attack while still retaining a high main task accuracy. + +# 3. Attack intuitions + +In this section, we investigate the inherent constraints that affect the effectiveness of backdoors when extending previous works to MBA scenarios and further provide the attack intuitions based on this investigation. + +# 3.1. Inherent constraints of SBA methods + +Recall that, from the neuron activation perspective, the backdoor function in a model is highly related to the specific activation paths across neurons. This results in the model being sensitive to triggered inputs and producing anomalously high feature attribution values [9, 34, 36]. Correspondingly, such high values result in backdoor samples distributed out of the target clean distribution in feature space, i.e. the backdoor function constructs the OOD backdoor mapping [21]. In MBA scenarios, attackers may adopt similar strategies to construct the activation paths, such as using redundant neurons, which leads to the construction of similar OOD mappings and competition for neuron weights. Specifically, different attackers have varying requirements for the output values of specific neurons within the pathway. However, the dominant attacker controls most neurons, ensuring they produce the desired values, which results in others failing to achieve their backdoor function. As shown in Fig. 2a, two attackers aim to construct persistent and effective OOD backdoor mappings by targeting redundant neurons. Such a strategy causes them to leverage overlapped activation paths to achieve different functions, which further results in infighting among adversaries. This drives us to find a solution that allows attackers to select different activation paths non-cooperatively. + +# 3.2. Attack intuitions and challenges + +To alleviate the exclusion, it is critical to design a mapping strategy that ensures different attackers can independent + +![](images/0d649c628cdfb509da23a4774260195bc893edb30981bbd493a5bb047cefb70a.jpg) +(a) SBA's mapping. + +![](images/6864a2cca13c4d70506a193ba3d89a2ee1b8688d4edf7548f8bcca559aa54d2f.jpg) +(b) Separate mapping. + +![](images/e83bbc755a5c97ac8a5cef83bf5780f6856973cd1c45dfe407503ab5785da6ad.jpg) +(c) ID mapping. +Figure 2. Illustrations of different mapping strategy. + +mappings with no intersections. Achieving this is challenging because attackers share no information with others and are unaware of each other's existence. Surprisingly, our empirical investigation reveals that A3FL can fulfill the aforementioned requirement by combining machine unlearning [37, 42], which indirectly stimulates the eliminations performed by other attackers. As shown in Fig 2b, the unlearning method adversarially optimizes the trigger to construct a distinct OOD backdoor mapping that differs from other attackers. Although this method mitigates the exclusions, it enhances the OOD characteristics of the backdoor functions, consequently introducing more challenges in bypassing the OOD-based defense methods, such as Indicator [21]. The OOD-based defense method leads us to a counterintuitive thought: Can a backdoor be triggered via a clean activation path? In other words, construct the ID mapping for the backdoor features and target distributions (present in Fig 2c). Therefore, once the adversaries construct the ID backdoor mapping to the target class, they naturally alleviate potential exclusions, even without collusion, since additional measures are no longer needed to circumvent the backdoor activation paths of other attackers. However, constructing the ID backdoor mapping remains challenging due to the following issues: + +# Challenge 1: How to construct the ID mapping? + +While backdoor attacks have emerged as a prominent topic in FL, to the best of our knowledge, no prior studies investigated the ID backdoor within FL. Intuitively, the backdoor sample should be processed as a clean sample following the clean activation path, meaning it should share similar feature outputs. One promising way is leveraging feature alignment. However, this may disrupt the relationships among clean classes, making it challenging for attackers to maintain the ID characteristics. Therefore, there is a pressing need to devise a novel method that ensures the effective construction of ID mapping. + +# Challenge 2: How to maintain the persistence of ID mapping? + +One key challenge for backdoor attacks in FL is that the continued dynamic clean updates may break the well-constructed backdoor mappings. Previous research focuses on leveraging neurons that are updated less frequently to build the neural activation path, which cannot be applied in ID mapping since our objective is to construct the bridge between the trigger pattern and the clean target activation path. + +Therefore, these maintenance requirements for ID mapping further exacerbate the challenges. + +# 4. Methodology + +To formulate the attack scenario, we first introduce the threat model. Driven by attack intuitions, we propose two methods that enable attackers to construct an efficient and persistent ID mapping without cooperation. + +# 4.1. Threat model + +Adversaries' Capabilities and knowledge. We consider a FL system running image classification tasks and assume that adversaries are FL participants, which enables attackers to control their own local datasets and training processes. Note that adversaries do not have access to any information regarding benign clients or other adversaries, including but not limited to datasets, models, and target classes. + +Adversaries' goal. Adversaries aim to plant a backdoor into the global model that makes it misclassify backdoor data to a specific class while leaving other tasks unaffected. + +# 4.2. Overview of Mirage + +Following the attack intuition, our objective is to construct an ID mapping to bridge the backdoor features and distribution of the target class, thereby freeing attackers from the exclusion of using similar backdoor activation paths. This presents two significant challenges: how to construct a qualified ID mapping and how to maintain it. To address these challenges, we propose Mirage, the first multilabel backdoor attack in FL, which enables attackers to noncooperatively compromise the global model by constructing an ID mapping between the backdoor features and target distributions. As illustrated in Fig. 3, Mirage introduces two components to construct an effective and persistent ID backdoor mapping. On the one hand, we propose a backdoor optimization method based on adversarial adaptation, which constructs the ID mapping by maximizing the misclassification probability of a well-designed backdoor sample detector. On the other hand, we deploy a constrained optimization strategy to tighten the backdoor distribution, ensuring its survival in the global training dynamics. Next, we provide details of each component. + +# 4.3. Effective ID Mapping Construction + +To address the first challenge, we aim to adversarially optimize the trigger pattern to make it able to construct the ID mapping. In order to optimize the trigger, we employ an OOD sample detector to determine whether a backdoored sample is an OOD sample and further optimize the trigger pattern to fool the detector. This introduces another problem: How can an attacker obtain an accurate detector? Previous research suggests training a new model as the detector [47]. However, retraining the detector with each trigger + +update incurs significant computational overhead, and the local samples limit its effectiveness. These shortcomings prompt us to seek another solution to obtain a detector that is both effective and lightweight for updates. Recall that the FL training process allows clients to obtain the latest global model, which contains a well-trained feature extractor. Consequently, we consider leveraging the feature extractor of the global model as the primary component and training a binary classifier to distinguish the OOD samples. Such an approach is a common practice in transfer learning, consumes only a few computing resources, and requires little data. + +In order to obtain an accurate detector, we first prepare the dataset for training. The local dataset for each attacker can be divided into $D_{cln} = \{(x,y) \in D_{att} | + \}$ and $D_{bd} = \{(x \oplus \delta, y) \in D_{att} | - \}$ , where $+$ indicates the ground truth label of a sample is $c$ , and $-$ is the opposite and $\delta$ represents the trigger pattern. Then, the detector dataset $D_{detector}$ ( $D_d$ for short) is composed as follows: + +$$ +D _ {d} = \{(x, -) \mid x \in D _ {b d} \} \cup \{(x, +) \mid x \in D _ {c l n} \}, \tag {1} +$$ + +To make the ID mapping construction more effective and flexible, we adopt an adversarial training-like method to optimize the trigger pattern. Specifically, $\theta_{detector}$ is trained on the $D_{detector}$ to differentiate between clean and backdoored samples by minimizing the binary cross-entropy (BCE) loss. + +Concurrently, the trigger is optimized to fool the detector by maximizing the binary classification loss. Such a min-max game equips the trigger with the ability to activate the clean activation path of the target class in the global model, naturally avoiding conflicts with other attackers. The min-max process is mathematically described as follows: + +$$ +\begin{array}{l} \delta = \operatorname {a r g m i n} \mathbb {E} _ {(x, y) \sim D _ {\text {d e t e c t o r}}} [ \mathcal {L} (x, +, \theta) | y = - ] \\ s. t. \theta = \underset {\theta} {\overset {\delta} {\operatorname {a r g m i n}}} \mathbb {E} _ {(x, y) \sim D _ {\text {d e t e c t o r}}} [ \mathcal {L} (x, y, \theta) ] \tag {2} \\ \end{array} +$$ + +where $\theta$ is the detector's parameters, with its feature extractor frozen and parameterized by the weights of the global model, and $\mathcal{L}$ is the BCE loss. For better implementation, we replace the maximization problem with minimizing the BCE loss of misclassification for the backdoored samples. + +# 4.4. Persistent ID Mapping Enhancement + +Although using an OOD detector and an adversarial strategy restricts the trigger pattern $\delta$ to reveal the ID feature, our empirical results indicate that the backdoored samples still do not meet our expectations. This is because the adversarial strategy does not constrain the tightness of backdoor mapping, which causes the margin-distributed backdoored samples not in-distribution anymore with the dynamic changes of training and failing to activate the target + +![](images/2e7879dd771ac7d4de94e80976837e864c9225773ab16729370982d3425f267f.jpg) +Figure 3. The workflow of Mirage. Step 1: Train an OOD sample detector $\theta_{\text{Detector}}$ . Step 2: Construct ID mapping by maximizing misclassification probabilities of backdoor samples on the detector. Step 3: Tighten backdoor distribution by minimizing the feature similarities between backdoor samples and benign samples. Step 4: Optimize the trigger by minimizing $L_{\text{Enhance}}$ and $L_{\text{detector}}$ . + +activation path. One promising solution is to leverage distribution estimation methods, like KDE [33], to obtain the actual distribution of the target class and use data augmentation to generate samples to improve the effectiveness of the detector. Although it has been proven effective, it requires countless computing resources. Therefore, we introduce a constrained optimization method to enhance the ID mapping by tightening the distribution of backdoored samples. Since the adversarial strategy successfully constructs the ID mapping, we aim to "push" the backdoor distribution to make it tighter. Specifically, we minimize the similarities between clean samples $x$ and their corresponding backdoored samples $x \oplus \delta$ , allowing the backdoored sample to deviate as much as possible from the original distribution. Additionally, we establish the direction by minimizing the loss of the backdoor samples on the global model. Formally, it can be formulated as the following optimization problem: + +$$ +\delta = \operatorname {a r g m i n} \mathrm {C S} \left[ \left(\theta_ {f} (x), \theta_ {f} (\hat {x}) \right] + \mathrm {C E} (\hat {x}, \hat {y}, \theta) \right. \tag {3} +$$ + +where $\theta_{f}$ represents the feature extractor of the global model $\theta$ . $\hat{x} = x \oplus \delta, x \in D_{att}$ , $\oplus$ is the backdoored operator. $\mathrm{CS}(\cdot)$ and $\mathrm{CE}(\cdot)$ represent the Cosine-similarity and the Cross-entropy loss, separately. Therefore, the backdoor distribution can be tightened into the target distribution and live within the global training dynamics. + +We summarized Mirage in Algorithm 1, and the detailed description is provided in Appendix B. + +# 5. Experiments + +In this section, we provide extensive experimental results to demonstrate the effectiveness of Mirage by comparing its performance with several SOTA backdoor attack methods from multiple angles. Our experiments run on a FL system performing image classification tasks, using an NVIDIA GeForce RTX 4090 GPU equipped with 24GB of memory. The detailed setups are listed as follows: + +Dataset: We conduct experiments on three widely used public real-world datasets, CIFAR-10, CIFAR-100 [17] and + +Algorithm 1: Mirage at $i$ -th attacker. +Data: $\theta_{t},\theta_{detector},\theta_{fe},\mathcal{D}_{i},\delta ,\eta ,c,E.$ +1 $\theta_t^i = \theta_t$ +2 /\* Trigger adversarial adaption\*/ +3 for epoch $e = 1,\ldots ,E$ do +4 Generate $D_{detector}$ by Eq 1; +5 $\theta_{detector} =$ argmin $\mathbb{E}_{(x,y)\sim D_{detector}}[\mathcal{L}(x,y,\theta_{detector})]$ . +6 for batch $\mathcal{B}\in D_i$ do +7 $\begin{array}{r}\mathcal{L}_{detector} = \\ \frac{1}{|\mathcal{B}|}\sum_{(x,y)\in \mathcal{B}}\mathcal{L}_{BCE}(x\oplus \delta , + ,\theta_{detector});\\ \mathcal{L}_{Enhance} = \frac{1}{|\mathcal{B}|}\sum_{(x,y)\in \mathcal{B}}\mathcal{L}_{CE}(x\oplus \delta ,\hat{y},\theta_t^i)\\ +\mathcal{L}_{CS}((\theta_{fe}(x),\theta_{fe}(x\oplus \delta));\\ \delta \gets \delta -\eta \cdot \nabla_{\delta}(\mathcal{L}_{detector} + \mathcal{L}_{Enhance});\\ end \end{array}$ +10 end +13 /\* Local training\*/ +14 Get the poisoned local dataset $\mathcal{D}_i^{poison}$ with $\delta$ +15 Local optimize $\theta_t^\prime$ on $\mathcal{D}_i^{poison}$ . +16 $\Delta_i^{t + 1} = \theta_t - \theta_t'$ . +17 Upload $\Delta_i^{t + 1}$ to the server; + +GTSRB [13]. Details are provided in Appendix C. + +Federated learning setup: By default, we set the number of FL clients $N = 100$ . The server randomly selects $M = 10$ clients at each communication round to contribute to the global model. The default global model architecture is ResNet-18 [12], which is a classic and effective model for image classification. Based on previous research, we randomly split the dataset among clients non-IID using Dirichlet sampling [14] with a hyperparameter $\alpha = 1$ . For each selected benign client, they train the local model for two local epochs using an SGD optimizer with a learning rate of 0.01. The batch size is 64 for CIFAR-10 and CIFAR-100 and 32 for GTSRB. The entire training process continues + +for 2100 communication rounds. + +Attack and defense setup: We compare the effectiveness of Mirage with SOTA methods, including A3FL [43], Chameleon [6], Neurotoxin [48], PGD [38], and Vanilla [2]. We further extend the aforementioned methods to MBA scenario, wherein attackers independently execute their attacks. Unless otherwise specified, we set $N = 3$ , with the target classes being 0, 1, and 2. In line with A3FL, attackers of Mirage optimize their triggers by PGD and poison $12.5\%$ of $D_{att}$ . Blend [5] is utilized as a backdoor trigger across all attacks. The attack window starts from the 2000-th round and continues for 100 rounds, so attackers are forced to wait for selection and cannot launch a continuous attack. + +We also evaluate those attacks against six SOTA defenses: Multi-Krum [4], DeepSight [31], Foolsgold [8], RFLBAT [39], FLAME [29] and BakcdoorIndicator [21]. For those attacks and defense methods, we follow the original implementations. Note that we selected the random noise dataset as the source of the indicator dataset. + +# 5.1. Evaluation metrics + +Following previous works, we comprehensively evaluate Mirage using accuracy (Acc), attack success rate (ASR), and lifespan. Specifically, Acc reflects the attack stealthiness, which indicates the performance of the backdoor model on the main tasks. ASR represents the ratio of backdoors samples misclassified as the labels specified by the attackers. Note that we report the mean ASR of three attackers on the global model at the end of FL. To evaluate the persistence of our attack, we introduce lifespan, denoted as the period that starts at the end of the attack window and ends when the ASR decreases to below a chosen threshold. + +# 5.2. Attack performance + +First, we evaluate the attack performance of Mirage across various datasets and compare it to different SOTA attack methods. The results are summarized in Table 1, and the best results are highlighted in Bold. The experiment shows that Mirage achieves remarkable results across three tasks and various SOTA defense methods compared to the baseline methods. In terms of accuracy, Mirage achieves the best accuracy in over half of the tasks and obtains suboptimal results for the remaining tasks, losing to the best ones by an average of only $0.498\%$ . Regarding effectiveness, Mirage achieves the best ASRs across all the tasks, with the gap among attackers being acceptable (the maximum gap is $6.99\%$ , and the average gap is $2.19\%$ ). Those results indicate that the design of constructing the ID mapping between backdoor features and target class allows attackers to independently attack the FL model without concerns about the potential exclusion, resulting in a global model that excels in both main and multiple backdoor tasks. + +Furthermore, A3FL is a relatively competitive method + +![](images/0eb1145d418053a35ff027e368df1827a1b78f41c973be07acd4a0209dc9f924.jpg) +Figure 4. Persistent evaluations. + +that leverages unlearning to simulate the potential exclusion and optimize the trigger to construct robust activation paths independent of other attackers. However, as described in Sec. 3.2, such an approach may enhance its OOD mapping concerning the benign distribution and increase the risk of being detected by OOD-based detection methods, such as Indicator [21]. The results of A3FL with the defense of Indicator support our analysis. Besides, the Indicator does not perform as expected, we will discuss it in Section 6.3. + +# 5.3. Lifespan + +Secondly, we evaluate the durability of attacks by comparing their lifespans. We extend the FL training process from 2100 to 3000 rounds while keeping the attack window to the first 100 rounds. Fig. 4 illustrates the attack success rate without any defense method. Mirage achieves similar results with A3FL, where the ASRs exceed $90\%$ at the end, indicating a lifespan of over 900 rounds. In contrast, the ASRs of all other baseline attacks are constrained by potential exclusion between attackers, as described before, making them ineffective and falling rapidly. These experimental results suggest that Mirage can achieve comparable durability to A3FL, which focuses on enhancing the persistence of backdoors in FL. + +# 5.4. The impact of different models + +In this section, we evaluate Mirage across different model architectures. Following previous research, we select widely used image classification model architectures: ResNet18 and ResNet34 [12], VGG11 and VGG19 [35], and MobileNet-V2 [32]. We set the parameters to default and only change the model architectures. As shown in Tab. 2, the attack performance of Mirage remains high in ASRs. The decrease in MobileNet-V2 is mainly attributed to the low utility of its model performance, which significantly reduces the detector's utility, resulting in sub-optimal ID mapping construction for certain attackers. However, the average ASRs remain acceptable. In conclusion, based on the results presented above, we can conclude that Mirage is not sensitive to specific model structures. + +# 5.5. The impact of adversary number + +We also study the impact of the number of adversaries, which can be considered as the number of different targets + +
DatasetDefenseVanillaPGDNeurotoxinChameleonA3FLMirage
Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)
CIFAR-10nodefense91.7631.88*87.6891.5231.25*92.7892.0352.99*90.2992.3539.65*79.9891.7399.52*0.2392.1699.54*0.27
deepsight91.4431.71*88.2990.8131.083*92.6392.3139.24*83.5092.3236.66*79.6692.1796.56*7.7491.2897.39*3.82
foolsgold88.7932.50*89.5090.7131.67*93.9691.0480.91*21.8390.6645.4667.2589.0095.30*12.0291.3297.35*2.51
Indicator85.0828.7685.3285.8320.6059.5085.1329.9860.32*86.138.215.1385.7570.33*88.4591.1093.46*3.90
multikrum92.361.594.5692.391.183.2592.641.895.1892.461.905.2592.1878.15*64.4192.1092.30*3.12
rfbat86.8330.5588.3488.2833.27*99.7987.2233.27*73.1286.9737.1586.4285.4687.83*10.2185.7497.29*2.11
Flame92.561.835.2692.513.399.9092.0240.97*94.0492.281.383.7192.2579.40*60.2392.3695.9*2.42
CIFAR-100nodefense69.4032.62*94.3170.5332.28*93.8971.0732.62*97.0571.6624.0727.5071.2698.67*2.4771.6599.05*0.77
deepsight70.0232.86*89.0270.5432.23*96.3270.5333.14*93.2871.3024.0626.1771.6892.05*18.4471.5698.67*2.43
foolsgold70.0233.3089.8170.1532.65*97.3471.0732.63*97.0571.6024.2027.1171.7898.71*1.6471.9499.18*1.00
Indicator67.736.8020.1263.855.8817.2860.9811.8932.3461.356.0713.4365.3537.13*92.7268.2299.8*0.10
multikrum71.810.380.7771.140.310.5971.630.360.6671.850.210.3871.7691.28*18.5571.8693.57*5.14
rfbat62.360.300.4163.100.600.1262.150.561.0061.4918.0127.3161.3594.34*16.5663.1995.68*6.99
Flame71.750.210.2872.090.330.6272.020.270.4572.100.300.5271.2291.48*19.8371.8294.22*3.16
GTSRBnodefense96.1033.32*98.2496.7333.10*98.8095.9338.39*95.7996.0845.62*87.0696.6599.63*0.4996.9799.73*0.48
deepsight95.9533.8*97.4495.9733.36*98.3596.1233.06*99.1796.1450.1476.7195.9594.78*12.6496.2096.76*2.53
foolsgold96.0833.47*97.9296.9433.06*98.7595.8837.57*96.3696.0451.9178.0696.6297.68*6.9596.9599.88*0.27
Indicator91.5750.9291.5389.4527.2778.1489.0817.0950.9792.912.765.3885.5185.05*25.2695.1299.73*0.70
multikrum96.244.1412.2096.972.747.7296.124.8414.3296.134.7213.9796.6298.18*4.1696.9298.99*1.73
rfbat94.4056.61*90.6289.4433.26*99.7294.5333.88*97.8094.7554.39*83.0495.2397.86*6.4294.2199.97*0.08
Flame95.886.1918.4596.972.747.7295.925.1515.3295.925.9517.7196.5299.31*1.5296.9299.32*1.05
+ +Table 1. Performance of Mirage compared with baseline attack methods against the SOTA defense methods. The Acc higher represents better, and so does the ASR. "GAP" represents the difference between the highest and the lowest ASR, and attack methods should maintain their GAP close to zero. An " * " after ASR indicates that at least one of the three attackers achieved over ${90}\%$ ASR. + +
ResNet18ResNet34VGG11VGG19MobileNet-V2
Acc92.1693.8991.1492.0785.38
ASR98.899.6399.2199.8297.65
+ +Table 2. Results on different model architectures. + +
Dataset0.515101000
CIFAR100Acc71.4671.6571.6871.5071.81
ASR99.5999.0599.1699.2699.21
GTSRBAcc96.9696.9796.6296.5796.70
ASR99.5499.7399.4798.2698.86
CIFAR10Acc91.6192.1692.4392.5192.38
ASR98.9598.8099.0399.3999.15
+ +Table 3. Results on different non-IID settings. + +
Attacker Number12345
Acc92.3792.5392.1692.1192.12
ASR99.0699.1598.8098.37598.484
GAP0.000.490.381.421.99
+ +Table 4. Mirage with different attacker number. + +since we assign different targets to the attackers. Table 4 presents the experimental results on CIFAR-10 with varying attack numbers ranging from 1 to 5, evaluating the performance of the global model on the main tasks and different backdoor tasks. Note that other settings follow the default configuration. We observe that the increase in attack numbers has a negligible impact on Mirage, with the accuracy of main tasks maintained above $92\%$ and the average ASR exceeding $98\%$ . The experimental results indicate that backdoor injection without collusion can be effectively implemented by constructing ID mapping. We also conducted experiments on different datasets (see Appendix D), and the results support the above conclusions. + +# 5.6. Performance under varying non-IID degrees + +Tab. 3 exhibits the attack performance of Mirage under varying non-IID degrees. We set different sampling param + +eters $\alpha$ for Mirage, with larger values indicating a distribution that is closer to IID. The results indicate that Mirage can achieve stable performance across different non-IID degrees. As described in Sec. 4.4, the adversarial strategy may optimize the backdoor distribution to the margin of the target distribution. To address this issue, we introduce a constrained optimization method, tightening the backdoor distribution into the objective distribution and ensuring successful attacks across different distributions. + +# 5.7. Illustration of backdoor distribution + +The preceding experiments leverage t-SNE to illustrate the backdoor distribution for Mirage and baseline methods. The results are in Fig. 5. The backdoor distribution from Fig. 5(a) to Fig. 5(d) further corroborates our presentation in 1, that different attackers may share similar neuron activation paths, resulting in their potential mitigation since their distribution is close to each other. On the other aspect, A3FL paves the activation path by adversariably unlearning the backdoor, which enhances their OOD mapping, producing clear boundaries between classes and tight clustering within each class. Although effective, this poses more threats to A3FL since the OOD-sample-based detection method can detect their backdoor models effectively. Looking back to Mirage, which is pictured in Fig. 5(f), backdoor samples present in-distribution with the target clean distribution. The results further indicate that the proposed Mirage can construct the ID mapping between backdoor features and clean distributions of the target class. + +# 5.8. Ablation study + +Sections 4.3 and 4.4 propose two hybrid techniques to construct and enhance the ID mapping. Here, we conduct an ablation study to show the effects of different techniques for our method. The experiment adheres to the settings + +![](images/c80a62ca8377a5018d653c1b71f1e7795ac3f9f31dcd4f8725bc3b91553e1e77.jpg) +Figure 5. Illustration of backdoor distribution. + +![](images/9acf7e4ede2ad0a10fd9d0f4ece7275c1dcdad459d82955fca221a65239fcf45.jpg) +Figure 6. Ablation study. + +outlined in Section 5.3. Fig. 6 presents the average ASR for Mirage and Mirage without enhancement. We can see that introducing the ID mapping enhancement significantly improves Mirage's performance. As described in Section 4.4, although the adversarial strategy successfully constructs the ID mapping, the backdoors may not fully activate the clean path, making the backdoor distribution map to the target distribution's margin, rendering it ineffective and non-persistent. Therefore, we introduce the constrained optimization method to tighten the backdoor distribution, ensuring backdoor samples will not be misidentified as the target distribution changes. + +# 6. Discussion + +# 6.1. Does the pixel trigger still effective? + +In the previous experiments, we adopted the blend trigger as the trigger pattern. Here, we conduct experiments with pixel triggers to evaluate their effectiveness. The results (presented in Appendix E) show that the pixel trigger remains effective. According to the t-SNE analysis of the feature attribution, the backdoor samples cannot aggregate together like the blend trigger. The main reason is that the pixel trigger can only activate a subset of neurons along the activation path, while other features typically activate the paths belonging to their original classes. However, it still constructs the ID mapping to the target distribution and successfully embeds the backdoors for different attackers. + +# 6.2. One-shot Mirage + +Mirage can perform a one-shot attack by increasing adversarial training epochs, allowing the trigger to directly simulate features of the target class, like adversarial samples. However, it does not construct persistent and effective ID backdoor mapping, resulting in low durability and side effects on the main tasks. Therefore, we enable the model to learn from the backdoor samples and construct effective and persistent ID mapping by injecting more rounds and fewer batches to achieve high-confidence classification. + +# 6.3. Potential defense methods + +According to Section 5.2, we can see that although Indicator [2], the most powerful detection method for now, can detect at least one of the three A3FL attackers, the rest can still successfully inject the backdoor. This is because it constructs limited OOD mappings, succeeding only under the assumption that the attacker constructs similar OOD mappings. Therefore, it cannot be as effective against dynamic attacks (A3FL) as it is against static attacks (Chameleon, etc.). Furthermore, for A3FL, such misdetections are sufficient for the attacker to inject a persistent backdoor into the global model. Thus, following the works with foresight [23, 27], we urge researchers to focus on MBA and expand their efforts to defend against it. + +Our proposed Mirage does not follow the traditional backdoor activation scheme, making it challenging for previous work to resist. This motivates us to seek solutions to address this threat to real-world FL applications. The advanced explainable method may effectively detect our attack, however, detecting all backdoored models remains challenging. Input detection may also identify our attack, as clean samples differ from backdoored samples. + +# 7. Conclusion + +In this paper, we uncover the inherent constraints of SBAs when extending them to MBA scenarios: Non-cooperative attackers will exclude others as they construct similar backdoor mappings. Based on this observation, we introduce Mirage, the first non-cooperative MBA strategy in FL. Mirage ensures the backdoor functions of different attackers by constructing effective and persistent ID backdoor mappings. Specifically, an adversarial trigger optimization method is introduced to construct the ID mapping between backdoor features and target distributions. Subsequently, Mirage leverages a constrained optimization method to tighten the ID mapping for persistence. Our comprehensive experiments demonstrate that Mirage outperforms existing backdoor attacks across different settings. For future work, investigating defense methods will be our primary focus. + +# Acknowledgement + +This work was supported in part by the National Natural Science Foundation of China under Grant(No. 62172215) and in part by the A3 Foresight Program of NSFC (Grant No. 62061146002) + +# References + +[1] Apple. Private federated learning (neurips 2019 expo talk abstract). https://nips.cc/ExpoConferences/2019/schedule?talkid=40. accessed: 2020-05-22. 1 +[2] Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov. How to backdoor federated learning. In International conference on artificial intelligence and statistics, pages 2938-2948. PMLR, 2020. 1, 2, 6, 8 +[3] Arjun Nitin Bhagoji, Supriyo Chakraborty, Prateek Mittal, and Seraphin Calo. Analyzing federated learning through an adversarial lens. In International conference on machine learning, pages 634-643. PMLR, 2019. 2 +[4] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 3, 6 +[5] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 6, 1 +[6] Yanbo Dai and Songze Li. Chameleon: Adapting to peer images for planting durable backdoors in federated learning. In International Conference on Machine Learning, pages 6712-6725. PMLR, 2023. 1, 2, 6 +[7] Pei Fang and Jinghui Chen. On the vulnerability of backdoor defenses for federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11800-11808, 2023. 2 +[8] Clement Fung, Chris J. M. Yoon, and Ivan Beschastnikh. The limitations of federated learning in sybil settings. In 23rd International Symposium on Research in Attacks, Intrusions and Defenses, RAID 2020, San Sebastian, Spain, October 14-15, 2020, pages 301-316. USENIX Association, 2020. 2, 6 +[9] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. BadNets: Evaluating Backdooring Attacks on Deep Neural Networks. IEEE Access, 7:47230-47244, 2019. 3 +[10] Hao Guan, Pew-Thian Yap, Andrea Bozoki, and Mingxia Liu. Federated learning for medical image analysis: A survey. Pattern Recognition, page 110424, 2024. 1 +[11] Andrew Hard, Kanishka Rao, Rajiv Mathews, Swaroop Ramaswamy, Françoise Beaufays, Sean Augenstein, Hubert Eichner, Chloe Kiddon, and Daniel Ramage. Federated Learning for Mobile Keyboard Prediction, 2018. 1 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 5, 6 +[13] Sebastian Houben, Johannes Stallkamp, Jan Salmen, Marc Schlipsing, and Christian Igel. Detection of traffic signs in real-world images: The german traffic sign detection benchmark. In The 2013 international joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2013. 5 + +[14] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. 5 +[15] Meirui Jiang, Zirui Wang, and Qi Dou. Harmofl: Harmonizing local and global drifts in federated learning on heterogeneous medical images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1087-1095, 2022. 1 +[16] Ruinan Jin and Xiaoxiao Li. Backdoor attack and defense in federated generative adversarial network-based medical image synthesis. Medical Image Analysis, 90:102965, 2023. 1 +[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009. 5, 1 +[18] Fan Lai, Yinwei Dai, Sanjay Singapuram, Jiachen Liu, Xi-angfeng Zhu, Harsha Madhyastha, and Mosharaf Chowdhury. Fedscale: Benchmarking model and system performance of federated learning at scale. In International conference on machine learning, pages 11814–11827. PMLR, 2022. 1 +[19] Haoyang Li, Qingqing Ye, Haibo Hu, Jin Li, Leixia Wang, Chengfang Fang, and Jie Shi. 3dfed: Adaptive and extensible framework for covert backdoor attack in federated learning. In 44th IEEE Symposium on Security and Privacy, SP 2023, San Francisco, CA, USA, May 21-25, 2023, pages 1893-1907. IEEE, 2023. 2 +[20] Minghui Li, Wei Wan, Yuxuan Ning, Shengshan Hu, Lulu Xue, Leo Yu Zhang, and Yichen Wang. Darkfed: A data-free backdoor attack in federated learning. arXiv preprint arXiv:2405.03299, 2024. 2 +[21] Songze Li and Yanbo Dai. BackdoorIndicator: Leveraging OOD data for proactive backdoor detection in federated learning. In 33rd USENIX Security Symposium (USENIX Security 24), pages 4193-4210, Philadelphia, PA, 2024. USENIX Association. 2, 3, 6 +[22] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1 +[23] Yige Li, Jiabo He, Hanxun Huang, Jun Sun, and Xingjun Ma. Shortcuts everywhere and nowhere: Exploring multi-trigger backdoor attacks, 2024. 8 +[24] Ye Li, Jiale Zhang, Yanchao Zhao, Bing Chen, and Shui Yu. Fairness-aware federated learning framework on heterogeneous data distributions. In ICC 2024-IEEE International Conference on Communications, pages 728-733. IEEE, 2024. 1 +[25] Xiaoting Lyu, Yufei Han, Wei Wang, Jingkai Liu, Bin Wang, Jiqiang Liu, and Xiangliang Zhang. Poisoning with cerberus: Stealthy and colluded backdoor attack against federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9020-9028, 2023. 2 +[26] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference + +on Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 1 +[27] Tuan Nguyen, Dung Thuy Nguyen, Khoa D Doan, and Kok-Seng Wong. Non-cooperative backdoor attacks in federated learning: A new threat landscape. arXiv preprint arXiv:2407.07917, 2024. 2, 8, 1 +[28] Thien Duc Nguyen, Phillip Rieger, Markus Miettinen, and Ahmad-Reza Sadeghi. Poisoning attacks on federated learning-based IoT intrusion detection system. In Proc. Workshop Decentralized IoT Syst. Secur.(DISS), 2020. 1 +[29] Thien Duc Nguyen, Phillip Rieger, Huili Chen, Hossein Yalame, Helen Möllering, Hossein Fereidooni, Samuel Marchal, Markus Miettinen, Azalia Mirhoseini, Shaza Zeitouni, Farinaz Koushanfar, Ahmad-Reza Sadeghi, and Thomas Schneider. FLAME: taming backdoors in federated learning. In 31st USENIX Security Symposium, USENIX Security 2022, Boston, MA, USA, August 10-12, 2022, pages 1415-1432. USENIX Association, 2022. 2, 3, 6 +[30] Yifan Niu and Weihong Deng. Federated learning for face recognition with gradient correction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1999–2007, 2022. 1 +[31] Phillip Rieger, Thien Duc Nguyen, Markus Miettinen, and Ahmad-Reza Sadeghi. Deepsight: Mitigating backdoor attacks in federated learning through deep model inspection. In 29th Annual Network and Distributed System Security Symposium, NDSS 2022, San Diego, California, USA, April 24-28, 2022. The Internet Society, 2022. 6 +[32] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE conference on computer vision and pattern recognition*, pages 4510–4520, 2018. 6 +[33] Simon J Sheather and Michael C Jones. A reliable data-based bandwidth selection method for kernel density estimation. Journal of the Royal Statistical Society: Series B (Methodological), 53(3):683-690, 1991. 5 +[34] Chenghui Shi, Shouling Ji, Xudong Pan, Xuhong Zhang, Mi Zhang, Min Yang, Jun Zhou, Jianwei Yin, and Ting Wang. Towards practical backdoor attacks on federated learning systems. IEEE Transactions on Dependable and Secure Computing, 2024. 2, 3 +[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6 +[36] Guanhong Tao, Zhenting Wang, Shiwei Feng, Guangyu Shen, Shiqing Ma, and Xiangyu Zhang. DRUPE: Distribution Preserving Backdoor Attack in Self-supervised Learning. In 2024 IEEE Symposium on Security and Privacy (SP), pages 29-29. IEEE Computer Society, 2023. 3 +[37] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE symposium on security and privacy (SP), pages 707-723. IEEE, 2019. 3 + +[38] Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris S. Papailiopoulos. Attack of the tails: Yes, you really can backdoor federated learning. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 6 +[39] Yongkang Wang, Dihua Zhai, Yufeng Zhan, and Yuanqing Xia. RFLBAT: A robust federated learning algorithm against backdoor attack. CoRR, abs/2201.03772, 2022. 2, 3, 6 +[40] Emily Wenger, Josephine Passananti, Arjun Nitin Bhagoji, Yuanshun Yao, Haitao Zheng, and Ben Y Zhao. Backdoor attacks against deep learning systems in the physical world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6206-6215, 2021. 1 +[41] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. DBA: distributed backdoor attacks against federated learning. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 2 +[42] Yi Zeng, Si Chen, Won Park, Z Morley Mao, Ming Jin, and Ruoxi Jia. Adversarial unlearning of backdoors via implicit hypergradient. arXiv preprint arXiv:2110.03735, 2021. 3 +[43] Hangfan Zhang, Jinyuan Jia, Jinghui Chen, Lu Lin, and Dinghao Wu. A3FL: adversarially adaptive backdoor attacks to federated learning. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 61213-61233, 2023. 1, 2, 6 +[44] Jiale Zhang, Chengcheng Zhu, Xiaobing Sun Di Wu, Jianming Yong, and Guodong Long. Badfss: backdoor attacks on federated self-supervised learning. In Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI), 2024. 2 +[45] Jiale Zhang, Chengcheng Zhu, Chunpeng Ge, Chuan Ma, Yanchao Zhao, Xiaobing Sun, and Bing Chen. BadCleaner: Defending Backdoor Attacks in Federated Learning via Attention-Based Multi-Teacher Distillation. IEEE Transactions on Dependable and Secure Computing, 21(5):4559–4573, 2024. 2 +[46] Jiale Zhang, Chengcheng Zhu, Xiaobing Sun, Chunpeng Ge, Bing Chen, Willy Susilo, and Shui Yu. FLPurifier: Backdoor Defense in Federated Learning via Decoupled Contrastive Training. IEEE Transactions on Information Forensics and Security, 19:4752-4766, 2024. 2 +[47] Quan Zhang, Yifeng Ding, Yongqiang Tian, Jianmin Guo, Min Yuan, and Yu Jiang. Advdoor: adversarial backdoor attack of deep learning system. In Proceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis, pages 127-138, 2021. 2, 4 +[48] Zhengming Zhang, Ashwinee Panda, Linyue Song, Yaoqing Yang, Michael Mahoney, Prateek Mittal, Ramchandran Kannan, and Joseph Gonzalez. Neurotoxin: Durable backdoors in federated learning. In International Conference on Machine Learning, pages 26429-26446. PMLR, 2022. 1, 2, 6 +[49] Chengcheng Zhu, Jiale Zhang, Xiaobing Sun, Bing Chen, and Weizhi Meng. Adfl: Defending backdoor attacks in federated learning via adversarial distillation. Computers & Security, 132:103366, 2023. 2 \ No newline at end of file diff --git a/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/images.zip b/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2de84e9721fcc8416f1482fbbb127b95d7659848 --- /dev/null +++ b/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc58d472f39a52421b9f83ebcbcebec11a8ea2125c30dd1d11aa75e97c2e991e +size 449963 diff --git a/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/layout.json b/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..afd8537f9cb6fa3a74b505151fb287001c74f8b5 --- /dev/null +++ b/2025/Infighting in the Dark_ Multi-Label Backdoor Attack in Federated Learning/layout.json @@ -0,0 +1,7916 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 541, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 541, + 122 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 541, + 122 + ], + "type": "text", + "content": "Infighting in the Dark: Multi-Label Backdoor Attack in Federated Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "spans": [ + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "type": "text", + "content": "Ye Li" + }, + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "type": "text", + "content": ", Yanchao Zhao" + }, + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "type": "text", + "content": ", Chengcheng Zhu" + }, + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "type": "text", + "content": ", Jiale Zhang" + }, + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "type": "text", + "content": " Nanjing University of Aeronautics and Astronautics " + }, + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 496, + 200 + ], + "type": "text", + "content": " Yangzhou University \n{miles.li, yczhao}@nuaa.edu.cn \nchengchengzhu2022@126.com, jialezhang@yzu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 253, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 613 + ], + "type": "text", + "content": "Federated Learning (FL), a privacy-preserving decentralized machine learning framework, has been shown to be vulnerable to backdoor attacks. Current research primarily focuses on the Single-Label Backdoor Attack (SBA), wherein adversaries share a consistent target. However, a critical fact is overlooked: adversaries may be non-cooperative, have distinct targets, and operate independently, which exhibits a more practical scenario called Multi-Label Backdoor Attack (MBA). Unfortunately, prior works are ineffective in the MBA scenario since non-cooperative attackers exclude each other. In this work, we conduct an in-depth investigation to uncover the inherent constraints of the exclusion: similar backdoor mappings are constructed for different targets, resulting in conflicts among backdoor functions. To address this limitation, we propose Mirage, the first non-cooperative MBA strategy in FL that allows attackers to inject effective and persistent backdoors into the global model without collusion by constructing in-distribution (ID) backdoor mapping. Specifically, we introduce an adversarial adaptation method to bridge the backdoor features and the target distribution in an ID manner. Additionally, we further leverage a constrained optimization method to ensure the ID mapping survives in the global training dynamics. Extensive evaluations demonstrate that Mirage outperforms various state-of-the-art attacks and bypasses existing defenses, achieving an average ASR greater than " + }, + { + "bbox": [ + 55, + 253, + 296, + 613 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 55, + 253, + 296, + 613 + ], + "type": "text", + "content": " and maintaining over " + }, + { + "bbox": [ + 55, + 253, + 296, + 613 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 253, + 296, + 613 + ], + "type": "text", + "content": " after 900 rounds. This work aims to alert researchers to this potential threat and inspire the design of effective defense mechanisms. Code has been made open-source." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 636, + 135, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 135, + 649 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 135, + 649 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 657, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 296, + 694 + ], + "type": "text", + "content": "Recent years have witnessed the flourishing of federated learning, an advanced privacy-preserving distributed learning paradigm [24, 26]. FL allows participants to collab" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 225, + 439, + 344 + ], + "blocks": [ + { + "bbox": [ + 321, + 225, + 439, + 344 + ], + "lines": [ + { + "bbox": [ + 321, + 225, + 439, + 344 + ], + "spans": [ + { + "bbox": [ + 321, + 225, + 439, + 344 + ], + "type": "image", + "image_path": "0ddbdf872f33853402ca79f8266fb5f8851611c6766853843c188eca1dfb8883.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. \"Blue\" and \"Red\" construct similar OOD mappings (see t-SNE), resulting in a conflict for the neuron weights. Only the dominant attacker can induce the model to output high feature attributions for its backdoor samples (see FA of " + }, + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "type": "text", + "content": "-th round) and successfully perform the attack. Similar in " + }, + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "type": "text", + "content": "-th round." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 440, + 226, + 548, + 343 + ], + "blocks": [ + { + "bbox": [ + 440, + 226, + 548, + 343 + ], + "lines": [ + { + "bbox": [ + 440, + 226, + 548, + 343 + ], + "spans": [ + { + "bbox": [ + 440, + 226, + 548, + 343 + ], + "type": "image", + "image_path": "dc5b1c26396d5cda41147e7343923cf1fa2364fec9b37fd99cf135cb8c3d2f5f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 408, + 555, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 470 + ], + "type": "text", + "content": "oratively build a neural network model without leaking their data. Following its paradigm, a wide range of FL applications have emerged in various safety and privacy-critical application domains, including medical image analysis [10, 15], face recognition [30], and others [1, 11]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 472, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 628 + ], + "type": "text", + "content": "Despite enhanced privacy, the distributed nature of FL makes them more vulnerable to various adversarial attacks, with backdoor attacks posing the greatest threat. For backdoor attacks in FL, an attacker injects a backdoor into the local model, which is then submitted to the FL system, causing the global model to inherit the backdoor after aggregation. Consequently, the global model functions correctly on clean data but misclassifies inputs stamped with the attacker's chosen backdoor trigger as the target class. Numerous studies have been done to improve the persistence and effectiveness of backdoors [2, 6, 43, 48], which pose significant threats to various applications, including computer vision [40], IoT networks [28], and healthcare systems [16]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Recent advancements in backdoor attacks in FL, most, if not all, assume that attackers in training are colluding and sharing the same target class, which can be denoted as Single-Label Backdoor Attacks (SBA). However, situations become more complex in real-world FL applications, especially in large-scale FL scenarios [18, 22], where attackers engage in training individually and compromise" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 148, + 712 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25770" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "type": "text", + "content": "the global model for their own purposes, i.e., Multi-Label Backdoor Attack (MBA). Intuitively, non-cooperative adversaries separately attack the global model and do not interfere with each other. Unfortunately, our empirical results indicate that directly applying existing attack methods to multiple adversaries leads to disorganized obstacles, making backdoors not as effective as they are in SBA scenarios. As illustrated in Fig. 1, two backdoors cannot be effective simultaneously because attackers construct similar Out-of-Distribution (OOD) backdoor mappings, indicating that their backdoor samples distribute together but outside the distributions of the target classes. This results in attackers excluding one another, with only the dominant attacker capable of embedding a backdoor in the global model, consequently causing the global model to output high feature attributions for its backdoor samples, thereby achieving a high attack success rate (ASR). It further raises a new problem: Can MBA successfully launch without collusion?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 291, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 295, + 554 + ], + "type": "text", + "content": "To address the aforementioned issue, we first conduct an in-depth investigation to uncover the inherent constraints of the exclusions of SBAs: they create similar backdoor mappings for distinct targets, resulting in conflicts among backdoor functions. Motivated by this observation, we propose Mirage, the first non-cooperative MBA strategy in FL, wherein multiple adversaries individually undermine the global model without collusion. At a high level, instead of coordinating the various OOD mappings, Mirage aims to ensure that the backdoor samples are processed as clean samples by constructing an effective and persistent ID mapping between backdoor features and target distributions. To achieve this goal, we introduce a trigger optimization method based on adversarial adaptation, which adaptively optimizes the trigger to fool a carefully designed OOD sample detector, thereby constructing the ID mapping. Furthermore, considering that the distributions change with global training, we further introduce a constrained optimization method to enhance the persistence of ID mapping by tightening the distribution of backdoor features, allowing the backdoor mapping to live within the global training dynamics. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "text", + "content": "- Uncovering the inherent constraints of extending existing SBAs to MBAs. To the best of our knowledge, we are the first to investigate and elucidate the inherent constraints of extending existing SBAs to MBAs. We underscore the practical threat posed by such attacks in large-scale FL applications by proposing a multi-label backdoor attack method based on the in-distribution backdoor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "content": "- Effective ID Mapping Construction. We propose a novel backdoor optimization method based on adversarial adaptation, which enables attackers to build the bridge between their backdoor and the distribution of their target class, allowing multiple adversaries to simultaneously launch backdoor attacks without interfering with others." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 553, + 216 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 144 + ], + "type": "text", + "content": "- Persistent ID Mapping Enhancement. We additionally introduce a constrained optimization method to enhance the ID mapping by minimizing the similarity of features between backdoorsed samples and benign samples. This tightens the ID backdoor mapping and ensures its persistence within the global training dynamics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 144, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 216 + ], + "type": "text", + "content": "- Evaluations and Discussions. Extensive experiments demonstrate that Mirage outperforms SOTA attacks across various settings. Furthermore, we discuss the vulnerability of existing defense methods to MBAs and present potential countermeasures against Mirage, highlighting the urgent demand for tailored defenses." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 228, + 400, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 400, + 241 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 400, + 241 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 248, + 529, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 248, + 529, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 248, + 529, + 262 + ], + "type": "text", + "content": "2.1. Backdoor Attacks in Federated Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 266, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 361 + ], + "type": "text", + "content": "Backdoor attacks aim to compromise the victim model to produce the attacker-desired outputs if and only if the inputs are assigned to a specific trigger pattern. In light of this, a few researchers also start to notice the threat of backdoor attacks on FL systems. Depending on whether the backdoor trigger is optimized, current backdoor attacks fall under the categories of either static-trigger attacks [2, 6, 20, 44, 48] or trigger-optimization attacks [7, 25, 43, 47]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 363, + 553, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 553, + 601 + ], + "type": "text", + "content": "The former pre-selected a fixed trigger pattern to poison the local training set and combine them with manually manipulating methods to improve the backdoor effectiveness. For example, [2, 3] aim to replace the global model by uploading a scaled-up poisoned model. In advance, [7, 34, 48] suggest leveraging the unimportant redundant neurons to prevent the backdoor from being erased shortly. Particularly, Dai et al. [6] introduce contrastive learning from the feature perspective of the backdoor to enhance its durability. The latter optimizes trigger patterns according to different tasks and generally achieves better effectiveness and utility. 3DFed [19] and A3FL [43] introduced the backdoor unlearning in an adversarial way to obtain a robust trigger to make the attacks more covert and persistent. DBA [41] and NBA [27] are the most similar work with us. However, DBA focuses on attacking the same target class, which is also an SBA situation. NBA investigated the noncooperative multi-label attack in FL but neither revealed the inner reason for the inapplicability of previous works nor proposed an effective method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 611, + 455, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 611, + 455, + 623 + ], + "spans": [ + { + "bbox": [ + 314, + 611, + 455, + 623 + ], + "type": "text", + "content": "2.2. Backdoor Defenses in FL" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 677 + ], + "type": "text", + "content": "Recently, researchers have noticed the threat of backdoor attacks in FL systems and actively explored various backdoor defense techniques, which can be categorized into either detection [8, 21, 39] or mitigation methods [4, 29, 45, 46, 49]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "The detection-based methods aim to utilize anomaly detection techniques to determine whether a model is backdoored. Foolsgold [8] assigns lower aggregation weights" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25771" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": "to updates with high pairwise cosine similarities, thereby mitigating the impact of backdoor updates. RFLBAT [39] discriminates the malicious model according to the difference between poisoned updates and clean updates in low-dimensional projection space. In advance, Li et al. [21] proposed BackdoorIndicator (referred to as \"Indicator\" hereafter), which detects potentially poisoned models based on the OOD properties of backdoor samples. Mitigation-based methods explored how to purify the backdoor model by eliminating backdoor triggers. MultiKrum [4] is a byzantine-robust aggregation protocol that aggregates the global model of having the smallest " + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": " distance. FLAME [29] adapts the weak DP method by noise boundary proof and a dynamic clipping bound, which is shown to alleviate the backdoor attack while still retaining a high main task accuracy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 157, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 157, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 157, + 289 + ], + "type": "text", + "content": "3. Attack intuitions" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 298, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 346 + ], + "type": "text", + "content": "In this section, we investigate the inherent constraints that affect the effectiveness of backdoors when extending previous works to MBA scenarios and further provide the attack intuitions based on this investigation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 356, + 251, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 251, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 251, + 368 + ], + "type": "text", + "content": "3.1. Inherent constraints of SBA methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 374, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 662 + ], + "type": "text", + "content": "Recall that, from the neuron activation perspective, the backdoor function in a model is highly related to the specific activation paths across neurons. This results in the model being sensitive to triggered inputs and producing anomalously high feature attribution values [9, 34, 36]. Correspondingly, such high values result in backdoor samples distributed out of the target clean distribution in feature space, i.e. the backdoor function constructs the OOD backdoor mapping [21]. In MBA scenarios, attackers may adopt similar strategies to construct the activation paths, such as using redundant neurons, which leads to the construction of similar OOD mappings and competition for neuron weights. Specifically, different attackers have varying requirements for the output values of specific neurons within the pathway. However, the dominant attacker controls most neurons, ensuring they produce the desired values, which results in others failing to achieve their backdoor function. As shown in Fig. 2a, two attackers aim to construct persistent and effective OOD backdoor mappings by targeting redundant neurons. Such a strategy causes them to leverage overlapped activation paths to achieve different functions, which further results in infighting among adversaries. This drives us to find a solution that allows attackers to select different activation paths non-cooperatively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 671, + 228, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 228, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 228, + 684 + ], + "type": "text", + "content": "3.2. Attack intuitions and challenges" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "To alleviate the exclusion, it is critical to design a mapping strategy that ensures different attackers can independent" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 396, + 151 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 396, + 151 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 396, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 396, + 151 + ], + "type": "image", + "image_path": "0d649c628cdfb509da23a4774260195bc893edb30981bbd493a5bb047cefb70a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 152, + 384, + 162 + ], + "lines": [ + { + "bbox": [ + 320, + 152, + 384, + 162 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 384, + 162 + ], + "type": "text", + "content": "(a) SBA's mapping." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 396, + 70, + 477, + 150 + ], + "blocks": [ + { + "bbox": [ + 396, + 70, + 477, + 150 + ], + "lines": [ + { + "bbox": [ + 396, + 70, + 477, + 150 + ], + "spans": [ + { + "bbox": [ + 396, + 70, + 477, + 150 + ], + "type": "image", + "image_path": "6864a2cca13c4d70506a193ba3d89a2ee1b8688d4edf7548f8bcca559aa54d2f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 152, + 469, + 162 + ], + "lines": [ + { + "bbox": [ + 397, + 152, + 469, + 162 + ], + "spans": [ + { + "bbox": [ + 397, + 152, + 469, + 162 + ], + "type": "text", + "content": "(b) Separate mapping." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 478, + 70, + 558, + 150 + ], + "blocks": [ + { + "bbox": [ + 478, + 70, + 558, + 150 + ], + "lines": [ + { + "bbox": [ + 478, + 70, + 558, + 150 + ], + "spans": [ + { + "bbox": [ + 478, + 70, + 558, + 150 + ], + "type": "image", + "image_path": "e83bbc755a5c97ac8a5cef83bf5780f6856973cd1c45dfe407503ab5785da6ad.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 152, + 540, + 162 + ], + "lines": [ + { + "bbox": [ + 488, + 152, + 540, + 162 + ], + "spans": [ + { + "bbox": [ + 488, + 152, + 540, + 162 + ], + "type": "text", + "content": "(c) ID mapping." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 338, + 163, + 528, + 174 + ], + "lines": [ + { + "bbox": [ + 338, + 163, + 528, + 174 + ], + "spans": [ + { + "bbox": [ + 338, + 163, + 528, + 174 + ], + "type": "text", + "content": "Figure 2. Illustrations of different mapping strategy." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 175, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 460 + ], + "type": "text", + "content": "mappings with no intersections. Achieving this is challenging because attackers share no information with others and are unaware of each other's existence. Surprisingly, our empirical investigation reveals that A3FL can fulfill the aforementioned requirement by combining machine unlearning [37, 42], which indirectly stimulates the eliminations performed by other attackers. As shown in Fig 2b, the unlearning method adversarially optimizes the trigger to construct a distinct OOD backdoor mapping that differs from other attackers. Although this method mitigates the exclusions, it enhances the OOD characteristics of the backdoor functions, consequently introducing more challenges in bypassing the OOD-based defense methods, such as Indicator [21]. The OOD-based defense method leads us to a counterintuitive thought: Can a backdoor be triggered via a clean activation path? In other words, construct the ID mapping for the backdoor features and target distributions (present in Fig 2c). Therefore, once the adversaries construct the ID backdoor mapping to the target class, they naturally alleviate potential exclusions, even without collusion, since additional measures are no longer needed to circumvent the backdoor activation paths of other attackers. However, constructing the ID backdoor mapping remains challenging due to the following issues:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 461, + 528, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 461, + 528, + 473 + ], + "spans": [ + { + "bbox": [ + 326, + 461, + 528, + 473 + ], + "type": "text", + "content": "Challenge 1: How to construct the ID mapping?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 474, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 605 + ], + "type": "text", + "content": "While backdoor attacks have emerged as a prominent topic in FL, to the best of our knowledge, no prior studies investigated the ID backdoor within FL. Intuitively, the backdoor sample should be processed as a clean sample following the clean activation path, meaning it should share similar feature outputs. One promising way is leveraging feature alignment. However, this may disrupt the relationships among clean classes, making it challenging for attackers to maintain the ID characteristics. Therefore, there is a pressing need to devise a novel method that ensures the effective construction of ID mapping." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 554, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 630 + ], + "type": "text", + "content": "Challenge 2: How to maintain the persistence of ID mapping?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "One key challenge for backdoor attacks in FL is that the continued dynamic clean updates may break the well-constructed backdoor mappings. Previous research focuses on leveraging neurons that are updated less frequently to build the neural activation path, which cannot be applied in ID mapping since our objective is to construct the bridge between the trigger pattern and the clean target activation path." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25772" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "Therefore, these maintenance requirements for ID mapping further exacerbate the challenges." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 138, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 138, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 138, + 121 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 126, + 295, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 295, + 175 + ], + "type": "text", + "content": "To formulate the attack scenario, we first introduce the threat model. Driven by attack intuitions, we propose two methods that enable attackers to construct an efficient and persistent ID mapping without cooperation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 181, + 141, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 181, + 141, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 141, + 192 + ], + "type": "text", + "content": "4.1. Threat model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 198, + 295, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 281 + ], + "type": "text", + "content": "Adversaries' Capabilities and knowledge. We consider a FL system running image classification tasks and assume that adversaries are FL participants, which enables attackers to control their own local datasets and training processes. Note that adversaries do not have access to any information regarding benign clients or other adversaries, including but not limited to datasets, models, and target classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 282, + 296, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 318 + ], + "type": "text", + "content": "Adversaries' goal. Adversaries aim to plant a backdoor into the global model that makes it misclassify backdoor data to a specific class while leaving other tasks unaffected." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 325, + 171, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 171, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 171, + 337 + ], + "type": "text", + "content": "4.2. Overview of Mirage" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 342, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 582 + ], + "type": "text", + "content": "Following the attack intuition, our objective is to construct an ID mapping to bridge the backdoor features and distribution of the target class, thereby freeing attackers from the exclusion of using similar backdoor activation paths. This presents two significant challenges: how to construct a qualified ID mapping and how to maintain it. To address these challenges, we propose Mirage, the first multilabel backdoor attack in FL, which enables attackers to noncooperatively compromise the global model by constructing an ID mapping between the backdoor features and target distributions. As illustrated in Fig. 3, Mirage introduces two components to construct an effective and persistent ID backdoor mapping. On the one hand, we propose a backdoor optimization method based on adversarial adaptation, which constructs the ID mapping by maximizing the misclassification probability of a well-designed backdoor sample detector. On the other hand, we deploy a constrained optimization strategy to tighten the backdoor distribution, ensuring its survival in the global training dynamics. Next, we provide details of each component." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 588, + 244, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 244, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 244, + 601 + ], + "type": "text", + "content": "4.3. Effective ID Mapping Construction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "To address the first challenge, we aim to adversarially optimize the trigger pattern to make it able to construct the ID mapping. In order to optimize the trigger, we employ an OOD sample detector to determine whether a backdoored sample is an OOD sample and further optimize the trigger pattern to fool the detector. This introduces another problem: How can an attacker obtain an accurate detector? Previous research suggests training a new model as the detector [47]. However, retraining the detector with each trigger" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "update incurs significant computational overhead, and the local samples limit its effectiveness. These shortcomings prompt us to seek another solution to obtain a detector that is both effective and lightweight for updates. Recall that the FL training process allows clients to obtain the latest global model, which contains a well-trained feature extractor. Consequently, we consider leveraging the feature extractor of the global model as the primary component and training a binary classifier to distinguish the OOD samples. Such an approach is a common practice in transfer learning, consumes only a few computing resources, and requires little data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "text", + "content": "In order to obtain an accurate detector, we first prepare the dataset for training. The local dataset for each attacker can be divided into " + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "inline_equation", + "content": "D_{cln} = \\{(x,y) \\in D_{att} | + \\}" + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "inline_equation", + "content": "D_{bd} = \\{(x \\oplus \\delta, y) \\in D_{att} | - \\}" + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "text", + "content": " indicates the ground truth label of a sample is " + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "inline_equation", + "content": "-" + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "text", + "content": " is the opposite and " + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "text", + "content": " represents the trigger pattern. Then, the detector dataset " + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "inline_equation", + "content": "D_{detector}" + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "inline_equation", + "content": "D_d" + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "text", + "content": " for short) is composed as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 309, + 555, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 309, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 326, + 309, + 555, + 323 + ], + "type": "interline_equation", + "content": "D _ {d} = \\{(x, -) \\mid x \\in D _ {b d} \\} \\cup \\{(x, +) \\mid x \\in D _ {c l n} \\}, \\tag {1}", + "image_path": "45b11873ff0000774a73b12c1e6654fd0876baaa921c25b6d240328fd67666d9.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 333, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 554, + 403 + ], + "type": "text", + "content": "To make the ID mapping construction more effective and flexible, we adopt an adversarial training-like method to optimize the trigger pattern. Specifically, " + }, + { + "bbox": [ + 313, + 333, + 554, + 403 + ], + "type": "inline_equation", + "content": "\\theta_{detector}" + }, + { + "bbox": [ + 313, + 333, + 554, + 403 + ], + "type": "text", + "content": " is trained on the " + }, + { + "bbox": [ + 313, + 333, + 554, + 403 + ], + "type": "inline_equation", + "content": "D_{detector}" + }, + { + "bbox": [ + 313, + 333, + 554, + 403 + ], + "type": "text", + "content": " to differentiate between clean and backdoored samples by minimizing the binary cross-entropy (BCE) loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 405, + 554, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 554, + 476 + ], + "type": "text", + "content": "Concurrently, the trigger is optimized to fool the detector by maximizing the binary classification loss. Such a min-max game equips the trigger with the ability to activate the clean activation path of the target class in the global model, naturally avoiding conflicts with other attackers. The min-max process is mathematically described as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 329, + 486, + 553, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 486, + 553, + 522 + ], + "spans": [ + { + "bbox": [ + 329, + 486, + 553, + 522 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\delta = \\operatorname {a r g m i n} \\mathbb {E} _ {(x, y) \\sim D _ {\\text {d e t e c t o r}}} [ \\mathcal {L} (x, +, \\theta) | y = - ] \\\\ s. t. \\theta = \\underset {\\theta} {\\overset {\\delta} {\\operatorname {a r g m i n}}} \\mathbb {E} _ {(x, y) \\sim D _ {\\text {d e t e c t o r}}} [ \\mathcal {L} (x, y, \\theta) ] \\tag {2} \\\\ \\end{array}", + "image_path": "d755691a064f40628f03fa0bec22095c80498293a4618476083f3bc050e918e3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 532, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 532, + 554, + 592 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 532, + 554, + 592 + ], + "type": "text", + "content": " is the detector's parameters, with its feature extractor frozen and parameterized by the weights of the global model, and " + }, + { + "bbox": [ + 313, + 532, + 554, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 532, + 554, + 592 + ], + "type": "text", + "content": " is the BCE loss. For better implementation, we replace the maximization problem with minimizing the BCE loss of misclassification for the backdoored samples." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 600, + 511, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 511, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 511, + 613 + ], + "type": "text", + "content": "4.4. Persistent ID Mapping Enhancement" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Although using an OOD detector and an adversarial strategy restricts the trigger pattern " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " to reveal the ID feature, our empirical results indicate that the backdoored samples still do not meet our expectations. This is because the adversarial strategy does not constrain the tightness of backdoor mapping, which causes the margin-distributed backdoored samples not in-distribution anymore with the dynamic changes of training and failing to activate the target" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25773" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 69, + 531, + 203 + ], + "blocks": [ + { + "bbox": [ + 80, + 69, + 531, + 203 + ], + "lines": [ + { + "bbox": [ + 80, + 69, + 531, + 203 + ], + "spans": [ + { + "bbox": [ + 80, + 69, + 531, + 203 + ], + "type": "image", + "image_path": "2e7879dd771ac7d4de94e80976837e864c9225773ab16729370982d3425f267f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 206, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 54, + 206, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 54, + 206, + 555, + 242 + ], + "type": "text", + "content": "Figure 3. The workflow of Mirage. Step 1: Train an OOD sample detector " + }, + { + "bbox": [ + 54, + 206, + 555, + 242 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{Detector}}" + }, + { + "bbox": [ + 54, + 206, + 555, + 242 + ], + "type": "text", + "content": ". Step 2: Construct ID mapping by maximizing misclassification probabilities of backdoor samples on the detector. Step 3: Tighten backdoor distribution by minimizing the feature similarities between backdoor samples and benign samples. Step 4: Optimize the trigger by minimizing " + }, + { + "bbox": [ + 54, + 206, + 555, + 242 + ], + "type": "inline_equation", + "content": "L_{\\text{Enhance}}" + }, + { + "bbox": [ + 54, + 206, + 555, + 242 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 206, + 555, + 242 + ], + "type": "inline_equation", + "content": "L_{\\text{detector}}" + }, + { + "bbox": [ + 54, + 206, + 555, + 242 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 250, + 296, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 296, + 454 + ], + "type": "text", + "content": "activation path. One promising solution is to leverage distribution estimation methods, like KDE [33], to obtain the actual distribution of the target class and use data augmentation to generate samples to improve the effectiveness of the detector. Although it has been proven effective, it requires countless computing resources. Therefore, we introduce a constrained optimization method to enhance the ID mapping by tightening the distribution of backdoored samples. Since the adversarial strategy successfully constructs the ID mapping, we aim to \"push\" the backdoor distribution to make it tighter. Specifically, we minimize the similarities between clean samples " + }, + { + "bbox": [ + 56, + 250, + 296, + 454 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 56, + 250, + 296, + 454 + ], + "type": "text", + "content": " and their corresponding backdoored samples " + }, + { + "bbox": [ + 56, + 250, + 296, + 454 + ], + "type": "inline_equation", + "content": "x \\oplus \\delta" + }, + { + "bbox": [ + 56, + 250, + 296, + 454 + ], + "type": "text", + "content": ", allowing the backdoored sample to deviate as much as possible from the original distribution. Additionally, we establish the direction by minimizing the loss of the backdoor samples on the global model. Formally, it can be formulated as the following optimization problem:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 460, + 295, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 460, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 75, + 460, + 295, + 475 + ], + "type": "interline_equation", + "content": "\\delta = \\operatorname {a r g m i n} \\mathrm {C S} \\left[ \\left(\\theta_ {f} (x), \\theta_ {f} (\\hat {x}) \\right] + \\mathrm {C E} (\\hat {x}, \\hat {y}, \\theta) \\right. \\tag {3}", + "image_path": "eecfaab344ea59838d73ac7e4c2b36c7365765112464c3099db6807f5f9eb750.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\theta_{f}" + }, + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "text", + "content": " represents the feature extractor of the global model " + }, + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\hat{x} = x \\oplus \\delta, x \\in D_{att}" + }, + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "text", + "content": " is the backdoored operator. " + }, + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{CS}(\\cdot)" + }, + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}(\\cdot)" + }, + { + "bbox": [ + 55, + 480, + 296, + 552 + ], + "type": "text", + "content": " represent the Cosine-similarity and the Cross-entropy loss, separately. Therefore, the backdoor distribution can be tightened into the target distribution and live within the global training dynamics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 296, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 577 + ], + "type": "text", + "content": "We summarized Mirage in Algorithm 1, and the detailed description is provided in Appendix B." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 586, + 137, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 137, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 137, + 600 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "In this section, we provide extensive experimental results to demonstrate the effectiveness of Mirage by comparing its performance with several SOTA backdoor attack methods from multiple angles. Our experiments run on a FL system performing image classification tasks, using an NVIDIA GeForce RTX 4090 GPU equipped with 24GB of memory. The detailed setups are listed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Dataset: We conduct experiments on three widely used public real-world datasets, CIFAR-10, CIFAR-100 [17] and" + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 314, + 266, + 553, + 531 + ], + "blocks": [ + { + "bbox": [ + 321, + 253, + 472, + 265 + ], + "lines": [ + { + "bbox": [ + 321, + 253, + 472, + 265 + ], + "spans": [ + { + "bbox": [ + 321, + 253, + 472, + 265 + ], + "type": "text", + "content": "Algorithm 1: Mirage at " + }, + { + "bbox": [ + 321, + 253, + 472, + 265 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 321, + 253, + 472, + 265 + ], + "type": "text", + "content": "-th attacker." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "lines": [ + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": "Data: " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "\\theta_{t},\\theta_{detector},\\theta_{fe},\\mathcal{D}_{i},\\delta ,\\eta ,c,E." + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " \n1 " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "\\theta_t^i = \\theta_t" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " \n2 /\\* Trigger adversarial adaption\\*/ \n3 for epoch " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "e = 1,\\ldots ,E" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " do \n4 Generate " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "D_{detector}" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " by Eq 1; \n5 " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "\\theta_{detector} =" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " argmin " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{(x,y)\\sim D_{detector}}[\\mathcal{L}(x,y,\\theta_{detector})]" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " . \n6 for batch " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{B}\\in D_i" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " do \n7 " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "\\begin{array}{r}\\mathcal{L}_{detector} = \\\\ \\frac{1}{|\\mathcal{B}|}\\sum_{(x,y)\\in \\mathcal{B}}\\mathcal{L}_{BCE}(x\\oplus \\delta , + ,\\theta_{detector});\\\\ \\mathcal{L}_{Enhance} = \\frac{1}{|\\mathcal{B}|}\\sum_{(x,y)\\in \\mathcal{B}}\\mathcal{L}_{CE}(x\\oplus \\delta ,\\hat{y},\\theta_t^i)\\\\ +\\mathcal{L}_{CS}((\\theta_{fe}(x),\\theta_{fe}(x\\oplus \\delta));\\\\ \\delta \\gets \\delta -\\eta \\cdot \\nabla_{\\delta}(\\mathcal{L}_{detector} + \\mathcal{L}_{Enhance});\\\\ end \\end{array}" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " \n10 end \n13 /\\* Local training\\*/ \n14 Get the poisoned local dataset " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i^{poison}" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " \n15 Local optimize " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "\\theta_t^\\prime" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i^{poison}" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " . \n16 " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "\\Delta_i^{t + 1} = \\theta_t - \\theta_t'" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " . \n17 Upload " + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "inline_equation", + "content": "\\Delta_i^{t + 1}" + }, + { + "bbox": [ + 314, + 266, + 553, + 531 + ], + "type": "text", + "content": " to the server;" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + } + ], + "index": 10, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 556, + 518, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 518, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 518, + 569 + ], + "type": "text", + "content": "GTSRB [13]. Details are provided in Appendix C." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": "Federated learning setup: By default, we set the number of FL clients " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "N = 100" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": ". The server randomly selects " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "M = 10" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " clients at each communication round to contribute to the global model. The default global model architecture is ResNet-18 [12], which is a classic and effective model for image classification. Based on previous research, we randomly split the dataset among clients non-IID using Dirichlet sampling [14] with a hyperparameter " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": ". For each selected benign client, they train the local model for two local epochs using an SGD optimizer with a learning rate of 0.01. The batch size is 64 for CIFAR-10 and CIFAR-100 and 32 for GTSRB. The entire training process continues" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25774" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 189, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 189, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 189, + 83 + ], + "type": "text", + "content": "for 2100 communication rounds." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 296, + 228 + ], + "type": "text", + "content": "Attack and defense setup: We compare the effectiveness of Mirage with SOTA methods, including A3FL [43], Chameleon [6], Neurotoxin [48], PGD [38], and Vanilla [2]. We further extend the aforementioned methods to MBA scenario, wherein attackers independently execute their attacks. Unless otherwise specified, we set " + }, + { + "bbox": [ + 55, + 84, + 296, + 228 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 55, + 84, + 296, + 228 + ], + "type": "text", + "content": ", with the target classes being 0, 1, and 2. In line with A3FL, attackers of Mirage optimize their triggers by PGD and poison " + }, + { + "bbox": [ + 55, + 84, + 296, + 228 + ], + "type": "inline_equation", + "content": "12.5\\%" + }, + { + "bbox": [ + 55, + 84, + 296, + 228 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 84, + 296, + 228 + ], + "type": "inline_equation", + "content": "D_{att}" + }, + { + "bbox": [ + 55, + 84, + 296, + 228 + ], + "type": "text", + "content": ". Blend [5] is utilized as a backdoor trigger across all attacks. The attack window starts from the 2000-th round and continues for 100 rounds, so attackers are forced to wait for selection and cannot launch a continuous attack." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "content": "We also evaluate those attacks against six SOTA defenses: Multi-Krum [4], DeepSight [31], Foolsgold [8], RFLBAT [39], FLAME [29] and BakcdoorIndicator [21]. For those attacks and defense methods, we follow the original implementations. Note that we selected the random noise dataset as the source of the indicator dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 309, + 167, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 167, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 167, + 320 + ], + "type": "text", + "content": "5.1. Evaluation metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "type": "text", + "content": "Following previous works, we comprehensively evaluate Mirage using accuracy (Acc), attack success rate (ASR), and lifespan. Specifically, Acc reflects the attack stealthiness, which indicates the performance of the backdoor model on the main tasks. ASR represents the ratio of backdoors samples misclassified as the labels specified by the attackers. Note that we report the mean ASR of three attackers on the global model at the end of FL. To evaluate the persistence of our attack, we introduce lifespan, denoted as the period that starts at the end of the attack window and ends when the ASR decreases to below a chosen threshold." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 173, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 173, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 173, + 480 + ], + "type": "text", + "content": "5.2. Attack performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 700 + ], + "type": "text", + "content": "First, we evaluate the attack performance of Mirage across various datasets and compare it to different SOTA attack methods. The results are summarized in Table 1, and the best results are highlighted in Bold. The experiment shows that Mirage achieves remarkable results across three tasks and various SOTA defense methods compared to the baseline methods. In terms of accuracy, Mirage achieves the best accuracy in over half of the tasks and obtains suboptimal results for the remaining tasks, losing to the best ones by an average of only " + }, + { + "bbox": [ + 55, + 486, + 295, + 700 + ], + "type": "inline_equation", + "content": "0.498\\%" + }, + { + "bbox": [ + 55, + 486, + 295, + 700 + ], + "type": "text", + "content": ". Regarding effectiveness, Mirage achieves the best ASRs across all the tasks, with the gap among attackers being acceptable (the maximum gap is " + }, + { + "bbox": [ + 55, + 486, + 295, + 700 + ], + "type": "inline_equation", + "content": "6.99\\%" + }, + { + "bbox": [ + 55, + 486, + 295, + 700 + ], + "type": "text", + "content": ", and the average gap is " + }, + { + "bbox": [ + 55, + 486, + 295, + 700 + ], + "type": "inline_equation", + "content": "2.19\\%" + }, + { + "bbox": [ + 55, + 486, + 295, + 700 + ], + "type": "text", + "content": "). Those results indicate that the design of constructing the ID mapping between backdoor features and target class allows attackers to independently attack the FL model without concerns about the potential exclusion, resulting in a global model that excels in both main and multiple backdoor tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Furthermore, A3FL is a relatively competitive method" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 552, + 163 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 552, + 163 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 552, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 552, + 163 + ], + "type": "image", + "image_path": "0eb1145d418053a35ff027e368df1827a1b78f41c973be07acd4a0209dc9f924.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 170, + 493, + 180 + ], + "lines": [ + { + "bbox": [ + 374, + 170, + 493, + 180 + ], + "spans": [ + { + "bbox": [ + 374, + 170, + 493, + 180 + ], + "type": "text", + "content": "Figure 4. Persistent evaluations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 180, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 289 + ], + "type": "text", + "content": "that leverages unlearning to simulate the potential exclusion and optimize the trigger to construct robust activation paths independent of other attackers. However, as described in Sec. 3.2, such an approach may enhance its OOD mapping concerning the benign distribution and increase the risk of being detected by OOD-based detection methods, such as Indicator [21]. The results of A3FL with the defense of Indicator support our analysis. Besides, the Indicator does not perform as expected, we will discuss it in Section 6.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 296, + 378, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 296, + 378, + 308 + ], + "spans": [ + { + "bbox": [ + 314, + 296, + 378, + 308 + ], + "type": "text", + "content": "5.3. Lifespan" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 314, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 555, + 469 + ], + "type": "text", + "content": "Secondly, we evaluate the durability of attacks by comparing their lifespans. We extend the FL training process from 2100 to 3000 rounds while keeping the attack window to the first 100 rounds. Fig. 4 illustrates the attack success rate without any defense method. Mirage achieves similar results with A3FL, where the ASRs exceed " + }, + { + "bbox": [ + 313, + 314, + 555, + 469 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 314, + 555, + 469 + ], + "type": "text", + "content": " at the end, indicating a lifespan of over 900 rounds. In contrast, the ASRs of all other baseline attacks are constrained by potential exclusion between attackers, as described before, making them ineffective and falling rapidly. These experimental results suggest that Mirage can achieve comparable durability to A3FL, which focuses on enhancing the persistence of backdoors in FL." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 478, + 481, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 481, + 491 + ], + "type": "text", + "content": "5.4. The impact of different models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 496, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 664 + ], + "type": "text", + "content": "In this section, we evaluate Mirage across different model architectures. Following previous research, we select widely used image classification model architectures: ResNet18 and ResNet34 [12], VGG11 and VGG19 [35], and MobileNet-V2 [32]. We set the parameters to default and only change the model architectures. As shown in Tab. 2, the attack performance of Mirage remains high in ASRs. The decrease in MobileNet-V2 is mainly attributed to the low utility of its model performance, which significantly reduces the detector's utility, resulting in sub-optimal ID mapping construction for certain attackers. However, the average ASRs remain acceptable. In conclusion, based on the results presented above, we can conclude that Mirage is not sensitive to specific model structures." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 672, + 492, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 492, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 492, + 685 + ], + "type": "text", + "content": "5.5. The impact of adversary number" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "We also study the impact of the number of adversaries, which can be considered as the number of different targets" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25775" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 234 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 234 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 234 + ], + "type": "table", + "html": "
DatasetDefenseVanillaPGDNeurotoxinChameleonA3FLMirage
Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)Acc (↑)ASR (↑)GAP (↓)
CIFAR-10nodefense91.7631.88*87.6891.5231.25*92.7892.0352.99*90.2992.3539.65*79.9891.7399.52*0.2392.1699.54*0.27
deepsight91.4431.71*88.2990.8131.083*92.6392.3139.24*83.5092.3236.66*79.6692.1796.56*7.7491.2897.39*3.82
foolsgold88.7932.50*89.5090.7131.67*93.9691.0480.91*21.8390.6645.4667.2589.0095.30*12.0291.3297.35*2.51
Indicator85.0828.7685.3285.8320.6059.5085.1329.9860.32*86.138.215.1385.7570.33*88.4591.1093.46*3.90
multikrum92.361.594.5692.391.183.2592.641.895.1892.461.905.2592.1878.15*64.4192.1092.30*3.12
rfbat86.8330.5588.3488.2833.27*99.7987.2233.27*73.1286.9737.1586.4285.4687.83*10.2185.7497.29*2.11
Flame92.561.835.2692.513.399.9092.0240.97*94.0492.281.383.7192.2579.40*60.2392.3695.9*2.42
CIFAR-100nodefense69.4032.62*94.3170.5332.28*93.8971.0732.62*97.0571.6624.0727.5071.2698.67*2.4771.6599.05*0.77
deepsight70.0232.86*89.0270.5432.23*96.3270.5333.14*93.2871.3024.0626.1771.6892.05*18.4471.5698.67*2.43
foolsgold70.0233.3089.8170.1532.65*97.3471.0732.63*97.0571.6024.2027.1171.7898.71*1.6471.9499.18*1.00
Indicator67.736.8020.1263.855.8817.2860.9811.8932.3461.356.0713.4365.3537.13*92.7268.2299.8*0.10
multikrum71.810.380.7771.140.310.5971.630.360.6671.850.210.3871.7691.28*18.5571.8693.57*5.14
rfbat62.360.300.4163.100.600.1262.150.561.0061.4918.0127.3161.3594.34*16.5663.1995.68*6.99
Flame71.750.210.2872.090.330.6272.020.270.4572.100.300.5271.2291.48*19.8371.8294.22*3.16
GTSRBnodefense96.1033.32*98.2496.7333.10*98.8095.9338.39*95.7996.0845.62*87.0696.6599.63*0.4996.9799.73*0.48
deepsight95.9533.8*97.4495.9733.36*98.3596.1233.06*99.1796.1450.1476.7195.9594.78*12.6496.2096.76*2.53
foolsgold96.0833.47*97.9296.9433.06*98.7595.8837.57*96.3696.0451.9178.0696.6297.68*6.9596.9599.88*0.27
Indicator91.5750.9291.5389.4527.2778.1489.0817.0950.9792.912.765.3885.5185.05*25.2695.1299.73*0.70
multikrum96.244.1412.2096.972.747.7296.124.8414.3296.134.7213.9796.6298.18*4.1696.9298.99*1.73
rfbat94.4056.61*90.6289.4433.26*99.7294.5333.88*97.8094.7554.39*83.0495.2397.86*6.4294.2199.97*0.08
Flame95.886.1918.4596.972.747.7295.925.1515.3295.925.9517.7196.5299.31*1.5296.9299.32*1.05
", + "image_path": "36da78277987ff9989e904e32fd36a9623b89f4e602b4bf3153d45994be1f6d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 75, + 285, + 276, + 320 + ], + "blocks": [ + { + "bbox": [ + 55, + 242, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 242, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 555, + 275 + ], + "type": "text", + "content": "Table 1. Performance of Mirage compared with baseline attack methods against the SOTA defense methods. The Acc higher represents better, and so does the ASR. \"GAP\" represents the difference between the highest and the lowest ASR, and attack methods should maintain their GAP close to zero. An \" * \" after ASR indicates that at least one of the three attackers achieved over " + }, + { + "bbox": [ + 55, + 242, + 555, + 275 + ], + "type": "inline_equation", + "content": "{90}\\%" + }, + { + "bbox": [ + 55, + 242, + 555, + 275 + ], + "type": "text", + "content": " ASR." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 285, + 276, + 320 + ], + "lines": [ + { + "bbox": [ + 75, + 285, + 276, + 320 + ], + "spans": [ + { + "bbox": [ + 75, + 285, + 276, + 320 + ], + "type": "table", + "html": "
ResNet18ResNet34VGG11VGG19MobileNet-V2
Acc92.1693.8991.1492.0785.38
ASR98.899.6399.2199.8297.65
", + "image_path": "c830df29a3bb5c13eb36b95aea2e3467153fb083ac2b4eb3970241b03bcfbaec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 88, + 342, + 264, + 415 + ], + "blocks": [ + { + "bbox": [ + 86, + 328, + 264, + 338 + ], + "lines": [ + { + "bbox": [ + 86, + 328, + 264, + 338 + ], + "spans": [ + { + "bbox": [ + 86, + 328, + 264, + 338 + ], + "type": "text", + "content": "Table 2. Results on different model architectures." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 342, + 264, + 415 + ], + "lines": [ + { + "bbox": [ + 88, + 342, + 264, + 415 + ], + "spans": [ + { + "bbox": [ + 88, + 342, + 264, + 415 + ], + "type": "table", + "html": "
Dataset0.515101000
CIFAR100Acc71.4671.6571.6871.5071.81
ASR99.5999.0599.1699.2699.21
GTSRBAcc96.9696.9796.6296.5796.70
ASR99.5499.7399.4798.2698.86
CIFAR10Acc91.6192.1692.4392.5192.38
ASR98.9598.8099.0399.3999.15
", + "image_path": "a1e3ac5d3b81e792a0122d917ccb5fb51c910e265060b0e1e8b05f46a900d1df.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 88, + 437, + 264, + 479 + ], + "blocks": [ + { + "bbox": [ + 91, + 423, + 259, + 434 + ], + "lines": [ + { + "bbox": [ + 91, + 423, + 259, + 434 + ], + "spans": [ + { + "bbox": [ + 91, + 423, + 259, + 434 + ], + "type": "text", + "content": "Table 3. Results on different non-IID settings." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 437, + 264, + 479 + ], + "lines": [ + { + "bbox": [ + 88, + 437, + 264, + 479 + ], + "spans": [ + { + "bbox": [ + 88, + 437, + 264, + 479 + ], + "type": "table", + "html": "
Attacker Number12345
Acc92.3792.5392.1692.1192.12
ASR99.0699.1598.8098.37598.484
GAP0.000.490.381.421.99
", + "image_path": "70d471edf15185594ff36e1fafad868330dec85c28c5308dfd35a326f227137a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 487, + 261, + 498 + ], + "lines": [ + { + "bbox": [ + 88, + 487, + 261, + 498 + ], + "spans": [ + { + "bbox": [ + 88, + 487, + 261, + 498 + ], + "type": "text", + "content": "Table 4. Mirage with different attacker number." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 508, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 664 + ], + "type": "text", + "content": "since we assign different targets to the attackers. Table 4 presents the experimental results on CIFAR-10 with varying attack numbers ranging from 1 to 5, evaluating the performance of the global model on the main tasks and different backdoor tasks. Note that other settings follow the default configuration. We observe that the increase in attack numbers has a negligible impact on Mirage, with the accuracy of main tasks maintained above " + }, + { + "bbox": [ + 55, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 55, + 508, + 295, + 664 + ], + "type": "text", + "content": " and the average ASR exceeding " + }, + { + "bbox": [ + 55, + 508, + 295, + 664 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 55, + 508, + 295, + 664 + ], + "type": "text", + "content": ". The experimental results indicate that backdoor injection without collusion can be effectively implemented by constructing ID mapping. We also conducted experiments on different datasets (see Appendix D), and the results support the above conclusions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 672, + 286, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 286, + 685 + ], + "type": "text", + "content": "5.6. Performance under varying non-IID degrees" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Tab. 3 exhibits the attack performance of Mirage under varying non-IID degrees. We set different sampling param" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "content": "eters " + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "content": " for Mirage, with larger values indicating a distribution that is closer to IID. The results indicate that Mirage can achieve stable performance across different non-IID degrees. As described in Sec. 4.4, the adversarial strategy may optimize the backdoor distribution to the margin of the target distribution. To address this issue, we introduce a constrained optimization method, tightening the backdoor distribution into the objective distribution and ensuring successful attacks across different distributions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 404, + 507, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 507, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 507, + 417 + ], + "type": "text", + "content": "5.7. Illustration of backdoor distribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 422, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 639 + ], + "type": "text", + "content": "The preceding experiments leverage t-SNE to illustrate the backdoor distribution for Mirage and baseline methods. The results are in Fig. 5. The backdoor distribution from Fig. 5(a) to Fig. 5(d) further corroborates our presentation in 1, that different attackers may share similar neuron activation paths, resulting in their potential mitigation since their distribution is close to each other. On the other aspect, A3FL paves the activation path by adversariably unlearning the backdoor, which enhances their OOD mapping, producing clear boundaries between classes and tight clustering within each class. Although effective, this poses more threats to A3FL since the OOD-sample-based detection method can detect their backdoor models effectively. Looking back to Mirage, which is pictured in Fig. 5(f), backdoor samples present in-distribution with the target clean distribution. The results further indicate that the proposed Mirage can construct the ID mapping between backdoor features and clean distributions of the target class." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 647, + 406, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 406, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 406, + 660 + ], + "type": "text", + "content": "5.8. Ablation study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "Sections 4.3 and 4.4 propose two hybrid techniques to construct and enhance the ID mapping. Here, we conduct an ablation study to show the effects of different techniques for our method. The experiment adheres to the settings" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25776" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 547, + 150 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 547, + 150 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 547, + 150 + ], + "type": "image", + "image_path": "c80a62ca8377a5018d653c1b71f1e7795ac3f9f31dcd4f8725bc3b91553e1e77.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 150, + 388, + 159 + ], + "lines": [ + { + "bbox": [ + 220, + 150, + 388, + 159 + ], + "spans": [ + { + "bbox": [ + 220, + 150, + 388, + 159 + ], + "type": "text", + "content": "Figure 5. Illustration of backdoor distribution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 85, + 160, + 262, + 228 + ], + "blocks": [ + { + "bbox": [ + 85, + 160, + 262, + 228 + ], + "lines": [ + { + "bbox": [ + 85, + 160, + 262, + 228 + ], + "spans": [ + { + "bbox": [ + 85, + 160, + 262, + 228 + ], + "type": "image", + "image_path": "9acf7e4ede2ad0a10fd9d0f4ece7275c1dcdad459d82955fca221a65239fcf45.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 228, + 221, + 239 + ], + "lines": [ + { + "bbox": [ + 129, + 228, + 221, + 239 + ], + "spans": [ + { + "bbox": [ + 129, + 228, + 221, + 239 + ], + "type": "text", + "content": "Figure 6. Ablation study." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 239, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 384 + ], + "type": "text", + "content": "outlined in Section 5.3. Fig. 6 presents the average ASR for Mirage and Mirage without enhancement. We can see that introducing the ID mapping enhancement significantly improves Mirage's performance. As described in Section 4.4, although the adversarial strategy successfully constructs the ID mapping, the backdoors may not fully activate the clean path, making the backdoor distribution map to the target distribution's margin, rendering it ineffective and non-persistent. Therefore, we introduce the constrained optimization method to tighten the backdoor distribution, ensuring backdoor samples will not be misidentified as the target distribution changes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 126, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 126, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 126, + 407 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 245, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 245, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 245, + 429 + ], + "type": "text", + "content": "6.1. Does the pixel trigger still effective?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 434, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 578 + ], + "type": "text", + "content": "In the previous experiments, we adopted the blend trigger as the trigger pattern. Here, we conduct experiments with pixel triggers to evaluate their effectiveness. The results (presented in Appendix E) show that the pixel trigger remains effective. According to the t-SNE analysis of the feature attribution, the backdoor samples cannot aggregate together like the blend trigger. The main reason is that the pixel trigger can only activate a subset of neurons along the activation path, while other features typically activate the paths belonging to their original classes. However, it still constructs the ID mapping to the target distribution and successfully embeds the backdoors for different attackers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 587, + 157, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 157, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 157, + 601 + ], + "type": "text", + "content": "6.2. One-shot Mirage" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Mirage can perform a one-shot attack by increasing adversarial training epochs, allowing the trigger to directly simulate features of the target class, like adversarial samples. However, it does not construct persistent and effective ID backdoor mapping, resulting in low durability and side effects on the main tasks. Therefore, we enable the model to learn from the backdoor samples and construct effective and persistent ID mapping by injecting more rounds and fewer batches to achieve high-confidence classification." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 161, + 459, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 161, + 459, + 172 + ], + "spans": [ + { + "bbox": [ + 314, + 161, + 459, + 172 + ], + "type": "text", + "content": "6.3. Potential defense methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 178, + 555, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 333 + ], + "type": "text", + "content": "According to Section 5.2, we can see that although Indicator [2], the most powerful detection method for now, can detect at least one of the three A3FL attackers, the rest can still successfully inject the backdoor. This is because it constructs limited OOD mappings, succeeding only under the assumption that the attacker constructs similar OOD mappings. Therefore, it cannot be as effective against dynamic attacks (A3FL) as it is against static attacks (Chameleon, etc.). Furthermore, for A3FL, such misdetections are sufficient for the attacker to inject a persistent backdoor into the global model. Thus, following the works with foresight [23, 27], we urge researchers to focus on MBA and expand their efforts to defend against it." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 334, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 554, + 430 + ], + "type": "text", + "content": "Our proposed Mirage does not follow the traditional backdoor activation scheme, making it challenging for previous work to resist. This motivates us to seek solutions to address this threat to real-world FL applications. The advanced explainable method may effectively detect our attack, however, detecting all backdoored models remains challenging. Input detection may also identify our attack, as clean samples differ from backdoored samples." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 441, + 388, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 388, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 388, + 453 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 460, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 639 + ], + "type": "text", + "content": "In this paper, we uncover the inherent constraints of SBAs when extending them to MBA scenarios: Non-cooperative attackers will exclude others as they construct similar backdoor mappings. Based on this observation, we introduce Mirage, the first non-cooperative MBA strategy in FL. Mirage ensures the backdoor functions of different attackers by constructing effective and persistent ID backdoor mappings. Specifically, an adversarial trigger optimization method is introduced to construct the ID mapping between backdoor features and target distributions. Subsequently, Mirage leverages a constrained optimization method to tighten the ID mapping for persistence. Our comprehensive experiments demonstrate that Mirage outperforms existing backdoor attacks across different settings. For future work, investigating defense methods will be our primary focus." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 640, + 396, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 640, + 396, + 651 + ], + "spans": [ + { + "bbox": [ + 314, + 640, + 396, + 651 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 652, + 554, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 554, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 554, + 699 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant(No. 62172215) and in part by the A3 Foresight Program of NSFC (Grant No. 62061146002)" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25777" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "text", + "content": "[1] Apple. Private federated learning (neurips 2019 expo talk abstract). https://nips.cc/ExpoConferences/2019/schedule?talkid=40. accessed: 2020-05-22. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "type": "text", + "content": "[2] Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov. How to backdoor federated learning. In International conference on artificial intelligence and statistics, pages 2938-2948. PMLR, 2020. 1, 2, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "text", + "content": "[3] Arjun Nitin Bhagoji, Supriyo Chakraborty, Prateek Mittal, and Seraphin Calo. Analyzing federated learning through an adversarial lens. In International conference on machine learning, pages 634-643. PMLR, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 295, + 281 + ], + "type": "text", + "content": "[4] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "type": "text", + "content": "[5] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 6, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 295, + 372 + ], + "type": "text", + "content": "[6] Yanbo Dai and Songze Li. Chameleon: Adapting to peer images for planting durable backdoors in federated learning. In International Conference on Machine Learning, pages 6712-6725. PMLR, 2023. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 374, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 295, + 418 + ], + "type": "text", + "content": "[7] Pei Fang and Jinghui Chen. On the vulnerability of backdoor defenses for federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11800-11808, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 420, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 485 + ], + "type": "text", + "content": "[8] Clement Fung, Chris J. M. Yoon, and Ivan Beschastnikh. The limitations of federated learning in sybil settings. In 23rd International Symposium on Research in Attacks, Intrusions and Defenses, RAID 2020, San Sebastian, Spain, October 14-15, 2020, pages 301-316. USENIX Association, 2020. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 295, + 520 + ], + "type": "text", + "content": "[9] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. BadNets: Evaluating Backdooring Attacks on Deep Neural Networks. IEEE Access, 7:47230-47244, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 522, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 295, + 555 + ], + "type": "text", + "content": "[10] Hao Guan, Pew-Thian Yap, Andrea Bozoki, and Mingxia Liu. Federated learning for medical image analysis: A survey. Pattern Recognition, page 110424, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 556, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 295, + 601 + ], + "type": "text", + "content": "[11] Andrew Hard, Kanishka Rao, Rajiv Mathews, Swaroop Ramaswamy, Françoise Beaufays, Sean Augenstein, Hubert Eichner, Chloe Kiddon, and Daniel Ramage. Federated Learning for Mobile Keyboard Prediction, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 657 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[13] Sebastian Houben, Johannes Stallkamp, Jan Salmen, Marc Schlipsing, and Christian Igel. Detection of traffic signs in real-world images: The german traffic sign detection benchmark. In The 2013 international joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2013. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[14] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "text", + "content": "[15] Meirui Jiang, Zirui Wang, and Qi Dou. Harmofl: Harmonizing local and global drifts in federated learning on heterogeneous medical images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1087-1095, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "text", + "content": "[16] Ruinan Jin and Xiaoxiao Li. Backdoor attack and defense in federated generative adversarial network-based medical image synthesis. Medical Image Analysis, 90:102965, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 240 + ], + "type": "text", + "content": "[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009. 5, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 243, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 307 + ], + "type": "text", + "content": "[18] Fan Lai, Yinwei Dai, Sanjay Singapuram, Jiachen Liu, Xi-angfeng Zhu, Harsha Madhyastha, and Mosharaf Chowdhury. Fedscale: Benchmarking model and system performance of federated learning at scale. In International conference on machine learning, pages 11814–11827. PMLR, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 310, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 374 + ], + "type": "text", + "content": "[19] Haoyang Li, Qingqing Ye, Haibo Hu, Jin Li, Leixia Wang, Chengfang Fang, and Jie Shi. 3dfed: Adaptive and extensible framework for covert backdoor attack in federated learning. In 44th IEEE Symposium on Security and Privacy, SP 2023, San Francisco, CA, USA, May 21-25, 2023, pages 1893-1907. IEEE, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "text", + "content": "[20] Minghui Li, Wei Wan, Yuxuan Ning, Shengshan Hu, Lulu Xue, Leo Yu Zhang, and Yichen Wang. Darkfed: A data-free backdoor attack in federated learning. arXiv preprint arXiv:2405.03299, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "text", + "content": "[21] Songze Li and Yanbo Dai. BackdoorIndicator: Leveraging OOD data for proactive backdoor detection in federated learning. In 33rd USENIX Security Symposium (USENIX Security 24), pages 4193-4210, Philadelphia, PA, 2024. USENIX Association. 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "text", + "content": "[22] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[23] Yige Li, Jiabo He, Hanxun Huang, Jun Sun, and Xingjun Ma. Shortcuts everywhere and nowhere: Exploring multi-trigger backdoor attacks, 2024. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 557, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 610 + ], + "type": "text", + "content": "[24] Ye Li, Jiale Zhang, Yanchao Zhao, Bing Chen, and Shui Yu. Fairness-aware federated learning framework on heterogeneous data distributions. In ICC 2024-IEEE International Conference on Communications, pages 728-733. IEEE, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 613, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 668 + ], + "type": "text", + "content": "[25] Xiaoting Lyu, Yufei Han, Wei Wang, Jingkai Liu, Bin Wang, Jiqiang Liu, and Xiangliang Zhang. Poisoning with cerberus: Stealthy and colluded backdoor attack against federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9020-9028, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[26] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25778" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "on Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[27] Tuan Nguyen, Dung Thuy Nguyen, Khoa D Doan, and Kok-Seng Wong. Non-cooperative backdoor attacks in federated learning: A new threat landscape. arXiv preprint arXiv:2407.07917, 2024. 2, 8, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 294, + 185 + ], + "type": "text", + "content": "[28] Thien Duc Nguyen, Phillip Rieger, Markus Miettinen, and Ahmad-Reza Sadeghi. Poisoning attacks on federated learning-based IoT intrusion detection system. In Proc. Workshop Decentralized IoT Syst. Secur.(DISS), 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 274 + ], + "type": "text", + "content": "[29] Thien Duc Nguyen, Phillip Rieger, Huili Chen, Hossein Yalame, Helen Möllering, Hossein Fereidooni, Samuel Marchal, Markus Miettinen, Azalia Mirhoseini, Shaza Zeitouni, Farinaz Koushanfar, Ahmad-Reza Sadeghi, and Thomas Schneider. FLAME: taming backdoors in federated learning. In 31st USENIX Security Symposium, USENIX Security 2022, Boston, MA, USA, August 10-12, 2022, pages 1415-1432. USENIX Association, 2022. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "text", + "content": "[30] Yifan Niu and Weihong Deng. Federated learning for face recognition with gradient correction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1999–2007, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 387 + ], + "type": "text", + "content": "[31] Phillip Rieger, Thien Duc Nguyen, Markus Miettinen, and Ahmad-Reza Sadeghi. Deepsight: Mitigating backdoor attacks in federated learning through deep model inspection. In 29th Annual Network and Distributed System Security Symposium, NDSS 2022, San Diego, California, USA, April 24-28, 2022. The Internet Society, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "text", + "content": "[32] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE conference on computer vision and pattern recognition*, pages 4510–4520, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "type": "text", + "content": "[33] Simon J Sheather and Michael C Jones. A reliable data-based bandwidth selection method for kernel density estimation. Journal of the Royal Statistical Society: Series B (Methodological), 53(3):683-690, 1991. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "text", + "content": "[34] Chenghui Shi, Shouling Ji, Xudong Pan, Xuhong Zhang, Mi Zhang, Min Yang, Jun Zhou, Jianwei Yin, and Ting Wang. Towards practical backdoor attacks on federated learning systems. IEEE Transactions on Dependable and Secure Computing, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "type": "text", + "content": "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "text", + "content": "[36] Guanhong Tao, Zhenting Wang, Shiwei Feng, Guangyu Shen, Shiqing Ma, and Xiangyu Zhang. DRUPE: Distribution Preserving Backdoor Attack in Self-supervised Learning. In 2024 IEEE Symposium on Security and Privacy (SP), pages 29-29. IEEE Computer Society, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[37] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE symposium on security and privacy (SP), pages 707-723. IEEE, 2019. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 149 + ], + "type": "text", + "content": "[38] Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris S. Papailiopoulos. Attack of the tails: Yes, you really can backdoor federated learning. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 150, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 183 + ], + "type": "text", + "content": "[39] Yongkang Wang, Dihua Zhai, Yufeng Zhan, and Yuanqing Xia. RFLBAT: A robust federated learning algorithm against backdoor attack. CoRR, abs/2201.03772, 2022. 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "type": "text", + "content": "[40] Emily Wenger, Josephine Passananti, Arjun Nitin Bhagoji, Yuanshun Yao, Haitao Zheng, and Ben Y Zhao. Backdoor attacks against deep learning systems in the physical world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6206-6215, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[41] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. DBA: distributed backdoor attacks against federated learning. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "type": "text", + "content": "[42] Yi Zeng, Si Chen, Won Park, Z Morley Mao, Ming Jin, and Ruoxi Jia. Adversarial unlearning of backdoors via implicit hypergradient. arXiv preprint arXiv:2110.03735, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[43] Hangfan Zhang, Jinyuan Jia, Jinghui Chen, Lu Lin, and Dinghao Wu. A3FL: adversarially adaptive backdoor attacks to federated learning. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 61213-61233, 2023. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "text", + "content": "[44] Jiale Zhang, Chengcheng Zhu, Xiaobing Sun Di Wu, Jianming Yong, and Guodong Long. Badfss: backdoor attacks on federated self-supervised learning. In Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI), 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 438, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 502 + ], + "type": "text", + "content": "[45] Jiale Zhang, Chengcheng Zhu, Chunpeng Ge, Chuan Ma, Yanchao Zhao, Xiaobing Sun, and Bing Chen. BadCleaner: Defending Backdoor Attacks in Federated Learning via Attention-Based Multi-Teacher Distillation. IEEE Transactions on Dependable and Secure Computing, 21(5):4559–4573, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 504, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 558 + ], + "type": "text", + "content": "[46] Jiale Zhang, Chengcheng Zhu, Xiaobing Sun, Chunpeng Ge, Bing Chen, Willy Susilo, and Shui Yu. FLPurifier: Backdoor Defense in Federated Learning via Decoupled Contrastive Training. IEEE Transactions on Information Forensics and Security, 19:4752-4766, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "text", + "content": "[47] Quan Zhang, Yifeng Ding, Yongqiang Tian, Jianmin Guo, Min Yuan, and Yu Jiang. Advdoor: adversarial backdoor attack of deep learning system. In Proceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis, pages 127-138, 2021. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 615, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 669 + ], + "type": "text", + "content": "[48] Zhengming Zhang, Ashwinee Panda, Linyue Song, Yaoqing Yang, Michael Mahoney, Prateek Mittal, Ramchandran Kannan, and Joseph Gonzalez. Neurotoxin: Durable backdoors in federated learning. In International Conference on Machine Learning, pages 26429-26446. PMLR, 2022. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[49] Chengcheng Zhu, Jiale Zhang, Xiaobing Sun, Bing Chen, and Weizhi Meng. Adfl: Defending backdoor attacks in federated learning via adversarial distillation. Computers & Security, 132:103366, 2023. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25779" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/d1db9220-1c65-401b-9202-6e181bfed6b4_content_list.json b/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/d1db9220-1c65-401b-9202-6e181bfed6b4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6d38ac3feb1465c981547b9bb4c832c624f358e6 --- /dev/null +++ b/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/d1db9220-1c65-401b-9202-6e181bfed6b4_content_list.json @@ -0,0 +1,1868 @@ +[ + { + "type": "text", + "text": "Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis", + "text_level": 1, + "bbox": [ + 233, + 130, + 767, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jian Han*, Jinlai Liu*, Yi Jiang*, Bin Yan, Yuqi Zhang, Zehuan Yuan†, Bingyue Peng, Xiaobing Liu ByteDance", + "bbox": [ + 272, + 203, + 725, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hanjian.thul23, liujinlai.licio, jiangyi.enjoy, yanbin/master}@bytedance.com {zhangyuqi.hi, yuanzehuan, bingyue.peng, will.liu}@bytedance.com", + "bbox": [ + 171, + 266, + 826, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Code: https://github.com/FoundationVision/Infinity Online demo: https://opensource.bytedance.com/gmpt/t2i/invite", + "bbox": [ + 240, + 306, + 754, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/29cf08ae7f20768842fe4f79bf08cf43ea65d94f738a849f470d7531dd6cfd79.jpg", + "image_caption": [ + "Figure 1. High-resolution image synthesis results from our Infinity-2B model, showcasing its capabilities in precise prompt following, spatial reasoning, text rendering, and aesthetics across different styles and aspect ratios." + ], + "image_footnote": [], + "bbox": [ + 93, + 349, + 254, + 700 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a532c938a6514fd71575c693da922625a94bc97714c20574c862c7a725d7e4db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 349, + 385, + 595 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/936ccb57f533d7860fce82f267505220b8a64aaa08ad88a046af339bb702ac82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 349, + 589, + 595 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/99ad381f612a32ea4b9bf615affc70f31021ee0c901df18d1ecc1de90f57a986.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 349, + 736, + 488 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3c8e34c488ae9240536902d9e42fc3f64bb1e35efc4701f16a21324ec08da1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 349, + 903, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 739, + 326, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Infinity, a Bitwise Visual AutoRegressive Modeling capable of generating high-resolution, photorealistic images following language instruction. Infinity refactors visual autoregressive model under a bitwise token prediction framework with an infinite-vocabulary classifier and bitwise self-correction mechanism. By theoretically expanding the tokenizer vocabulary size to infinity in Transformer, our", + "bbox": [ + 88, + 771, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "method significantly unleashes powerful scaling capabilities to infinity compared to vanilla VAR. Extensive experiments indicate Infinity outperforms AutoRegressive Text-to-Image models by large margins, matches or surpasses leading diffusion models. Without extra optimization, Infinity generates a $1024 \\times 1024$ image in 0.8s, $2.6 \\times$ faster than SD3-Medium, making it the fastest Text-to-Image model. All the code and models are available to promote further exploration of Infinity for visual generation.", + "bbox": [ + 511, + 741, + 906, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † Corresponding author.", + "bbox": [ + 107, + 886, + 346, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15733", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual generation[20, 27, 45, 51, 55] has recently seen rapid advancements, enabling high-quality, high-resolution images and video synthesis[7, 22]. Text-to-image generation[6, 22, 46, 48, 49, 53] is one of the most challenging tasks due to its need for complex language adherence and intricate scene creation. Currently, visual generation is primarily divided into two main approaches: Diffusion models and AutoRegressive models.", + "bbox": [ + 89, + 114, + 483, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models[20, 22, 27, 45, 46, 55], trained to invert the forward paths of data toward random noise, effectively generate images through a continuous denoising process. AutoRegressive models[14, 21, 21, 58, 64, 73], on the other hand, harness the scalability and generalizability of language models[2, 3, 16, 28, 60, 63, 65, 66, 72] by employing a visual tokenizer[50, 67, 78] to convert images into discrete tokens and causally optimize these tokens, allowing image generation through next-token prediction or next-scale prediction. AutoRegressive models encounter significant challenges in high-resolution text-to-image synthesis[58, 70, 79]. They exhibit inferior reconstruction quality when utilizing discrete tokens as opposed to continuous tokens. Additionally, their generated visual contents are less detailed than those by diffusion models. Inefficiency and latency in visual generation, resulting from the raster scan method of next-token prediction, further exacerbate these issues.", + "bbox": [ + 89, + 234, + 483, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Visual AutoRegressive Modeling (VAR)[64] redefined autoregressive learning on images as coarse-to-fine \"next-scale prediction\". It demonstrates superior generalization and scaling capabilities compared to diffusion transformers while requiring fewer steps. VAR leverages the powerful scaling properties of LLMs [25, 31] and can simultaneously refine previous scale steps, benefiting from the strengths of diffusion models as well. However, the index-wise discrete tokenizer[21, 38, 42, 64, 67, 69, 83] employed in AutoRegressive or Visual AutoRegressive models faces significant quantization errors with a limited vocabulary size resulting in difficulties in reconstructing fine-grained details, especially in high-resolution images. In the generation stage, index-wise tokens suffer from fuzzy supervision leading to visual detail loss and local distortions. Moreover, train-test discrepancies from teacher-forcing training, inherent to LLMs, amplify cumulative errors in visual details. These challenges make index-wise tokens a significant bottleneck for AutoRegressive models.", + "bbox": [ + 89, + 507, + 483, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a novel bitwise modeling framework that substitutes index-wise tokens with bitwise tokens throughout the process. Our bitwise modeling framework consists of three primary modules: bitwise visual tokenizer, bitwise infinite-vocabulary classifier, and bitwise self-correction. Inspired by the success and widespread adoption of binary vector quantization[80, 85], we scaled up the tokenizer vo", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0528a35add7e85d91bb366c2a15151c954a45def3e393ee5e18c0d89264b490e.jpg", + "image_caption": [ + "Figure 2. Left: Conventional classifier predicts $2^{d}$ indices. Right: Infinite-Vocabulary Classifier predicts $d$ bits instead, being much more efficient than index-wise classifier when $d$ is large." + ], + "image_footnote": [], + "bbox": [ + 522, + 90, + 903, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cabulary to $2^{64}$ , significantly surpassing all previous AutoRegressive model vocabularies[58, 81]. This expansion allows for reconstruction quality that far exceeds previous discrete tokenizers, achieving results comparable to continuous VAE[51], with scores improving from 0.87 to 0.33 on ImageNet-256 benchmark[19]. In Fig.2, we transform the conventional token prediction from a large integer into binary bits in a bitwise infinite-vocabulary classifier to address optimization and computation challenges, enabling the learning of massive vocabularies in Visual AutoRegressive models. Additionally, we incorporated bitwise self-correction during training by randomly flipping some bits to simulate prediction mistakes and re-quantizing the residual features, thus endowing the system with self-correcting capabilities. Our method, Infinity: Bitwise Visual AutoRegressive Modeling, maintains the scaling and speed advantages of VAR while achieving detailed reconstruction and generation quality comparable to that of continuous Diffusion models.", + "bbox": [ + 511, + 320, + 906, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Infinity outperforms all previous AutoRegressive text-to-image models by large margins and matches or surpasses state-of-the-art diffusion models including SDXL[46], PixArt-Sigma[13],DALL-E3[6] and Stable-Diffusion 3[22] in both quantitative evaluations and human preference scores. The core contributions of our work are as follows:", + "bbox": [ + 511, + 609, + 905, + 700 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose Infinity, an autoregressive model utilizing Bitwise Modeling, which significantly enhances scalability and visual detail representation capabilities of discrete generative models. This framework opens up new possibilities of 'infinity' within the discrete generation community.", + "2. Infinity demonstrates the potential of scaling tokenizers and transformers by achieving near-continuous tokenizer performance with its image tokenizer, surpassing diffusion models in high-quality text-to-image generation.", + "3. Infinity enables an autoregressive text-to-image model to achieve exceptional prompt adherence and superior" + ], + "bbox": [ + 529, + 704, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15734", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image generation quality, while also delivering unprecedented inference speed.", + "bbox": [ + 125, + 90, + 482, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 135, + 233, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. AutoRegressive Models", + "text_level": 1, + "bbox": [ + 89, + 160, + 305, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AutoRegressive models, leveraging the powerful scaling capabilities of LLMs[8, 16, 47, 65, 66], use discrete image tokenizers[21, 50, 67] in conjunction with transformers to generate images based on next-token prediction. VQ-based methods [21, 35, 50, 58, 67] employ vector quantization to convert image patches into index-wise tokens and use a decoder-only transformer to predict the next token index. However, these methods are limited by the lack of scaled-up transformers and the quantization error inherent in VQ-VAE[67], preventing them from achieving performance on par with diffusion models. Parti [79], Emu3 [70], chameleon[62], Liquid[73], and VideoPoet[32] scaled up autoregressive models in text-to-image or video synthesis. Inspired by the spatial structure of visual information, Visual AutoRegressive modeling(VAR)[64] redefines the autoregressive modeling on images as next-scale prediction framework, significantly improving generation quality and sampling speed. HART[61] adopted hybrid tokenizers based on VAR. MAR[37], Fluid[23] proposed random-order models and employed continuous tokenizer rather than discrete tokenizer.", + "bbox": [ + 89, + 183, + 483, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Diffusion Models.", + "text_level": 1, + "bbox": [ + 89, + 510, + 263, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models have recently experienced rapid advancements in several directions. Denoising learning mechanisms [27, 44] and improvements in sampling efficiency [4, 40, 41, 55, 56] have been continuously optimized to generate high-quality images. Bit Diffusion [15] represents data with analog bits and then trains a continuous diffusion model. Latent diffusion models [51] were the first to propose modeling in the latent space rather than the pixel space. Recently, latent diffusion models [18, 22] have scaled up VAEs to enhance representations in the latent space. DiT [45] and U-ViT [5] employ more scalable transformers to model diffusion, achieving superior results. Consequently, mainstream text-to-image diffusion models [6, 12, 22] have adopted the DiT architecture due to its effectiveness.", + "bbox": [ + 89, + 534, + 483, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Scaling Law", + "text_level": 1, + "bbox": [ + 89, + 770, + 220, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scaling laws in autoregressive language models reveal a power-law relationship between model size, dataset size, and compute with test set cross-entropy loss [1, 25, 31]. These laws help predict larger model performance, leading to efficient resource allocation and ongoing improvements without saturation [8, 28, 65, 66, 72, 84]. This has inspired research into scaling in visual generation [7, 22, 59, 64, 82].", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Infinity Setup", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Visual AutoRegressive Modeling", + "text_level": 1, + "bbox": [ + 513, + 114, + 797, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Infinity incorporates a visual tokenizer and a transformer for image synthesis. During the training stage, a sample consists of a text prompt $t$ and a ground truth image $im$ . The proposed visual tokenizer first encodes the image $im$ into a feature map $f \\in \\mathbb{R}^{h \\times w \\times C}$ with stride $s$ and then quantizes the feature map $f$ into $K$ multi-scale residual maps $(r_1, r_2, \\dots, r_K)$ . The resolution of $r_k$ is $h_k \\times w_k$ and it grows larger gradually from $k = 1 \\rightarrow K$ . Based on this sequence of residuals, we can gradually approximate the continuous feature $f$ as in Eq.1", + "bbox": [ + 511, + 137, + 906, + 290 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {k} = \\sum_ {i = 1} ^ {k} u p \\left(r _ {i}, (h, w)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 300, + 905, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "up means bilinear upsampling and $f_{k}$ is the cumulative sum of the upsampled $r_{\\leq k}$ .", + "bbox": [ + 511, + 352, + 906, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Subsequently, transformer learns to predict residuals $r$ of the next scale conditioned on previous predictions and the text input in an autoregressive manner. Formally, the autoregressive likelihood can be formulated as:", + "bbox": [ + 511, + 383, + 905, + 444 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(r _ {1}, \\dots , r _ {K}\\right) = \\prod_ {k = 1} ^ {K} p \\left(r _ {k} \\mid r _ {1}, \\dots , r _ {k - 1}, \\psi (t)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 455, + 906, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\psi(t)$ is the text embeddings from Flan-T5 [17] model. $(r_1, \\dots, r_{k-1}, \\psi(t))$ serves as the prefixed context When predicting $r_k$ . Besides, the text embeddings $\\psi(t)$ further guide the prediction through a cross-attention mechanism. In particular, as shown in Fig. 3, the text embeddings $\\psi(t) \\in \\mathbb{R}^{L \\times C}$ is projected into a $\\langle \\mathrm{SOS} \\rangle \\in \\mathbb{R}^{1 \\times 1 \\times h}$ as the input of the first scale, where $h$ is the hidden dimension of transformer. The transformer is required to predict $r_1$ based on $\\langle \\mathrm{SOS} \\rangle$ in the first scale. In the latter $k$ -th scale, to match the spatial size of the input and the output label $R_k$ , we take the downsampled feature $\\widetilde{f}_{k-1}$ from the last scale $k-1$ as the input to predict $R_k$ in parallel.", + "bbox": [ + 511, + 508, + 908, + 691 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {f} _ {k - 1} = \\operatorname {d o w n} \\left(f _ {k - 1}, \\left(h _ {k}, w _ {k}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 702, + 906, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $down$ is bilinear downsampling and the spatial size of both $\\widetilde{f}_{k-1}$ and $r_k$ is $(h_k, w_k)$ . In previous index-wise [64] representations, the shape of $r_k$ is $(h_k, w_k, V_d)$ . $V_d$ is the vocabulary size of the visual tokenizer. For binary quantization [80, 85] with code embedding dimension $d$ , $V_d = 2^d$ .", + "bbox": [ + 511, + 732, + 906, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The transformer consists of a stack of repeated blocks, where each block includes RoPE2D [26], Self-Attention, Cross Attention, and FFN layers. The text embeddings $\\psi(t)$ provide effective guidance for image synthesis in each cross-attention layer. During the training stage, we exploit a block-wise causal attention mask to ensure that the", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15735", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d2b4972cb2f63c6654518ef6ea4f2df2d10093ee717cf02fbba77d7f224d51b3.jpg", + "image_caption": [ + "Figure 3. Framework of Infinity. Infinity introduces bitwise modeling. When predicting $r_k$ , the sequence $(r_1, r_2, \\dots, r_{k-1})$ serves as the prefixed context and the text embeddings $\\psi(t)$ guides the prediction through a cross attention mechanism." + ], + "image_footnote": [], + "bbox": [ + 166, + 92, + 529, + 393 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/925163793b1f89868a0e859d1d3180ee3d889ea4f0402927d32ecfcd68510870.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 89, + 823, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "transformer can only attend to its prefixed context, i.e., $(\\langle \\mathrm{SOS}\\rangle ,\\widetilde{f}_1,\\dots ,\\widetilde{f}_{k - 1})$ , when predicting $f_{k}$ . During the inference stage, we perform KV-Caching to speed up inference and there's no need for masking.", + "bbox": [ + 89, + 436, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Infinite Tokenizer", + "text_level": 1, + "bbox": [ + 89, + 510, + 264, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Increasing the vocabulary size has significant potential for improving reconstruction and generation quality. However, directly enlarging the vocabulary in existing tokenizers[58, 64, 71, 81] leads to a substantial increase in memory consumption and computational burden. To address these challenges and fully exploit the potential of discrete tokenizers, this paper proposes a new bitwise multi-scale residual quantizer, which significantly reduces memory usage, enabling the training of extremely large vocabulary, e.g. $2^{64}$ .", + "bbox": [ + 89, + 532, + 483, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bitwise Multi-scale Residual Quantizer. We replace the original vector quantizer of VAR [64] with a dimension-independent bitwise quantizer. In this paper, we consider two candidates, LFQ [81] and BSQ[85]. Given $K$ scales in the multi-scale quantizer, on the $k$ -th scale, the input continuous residual vector $z_{k} \\in \\mathbb{R}^{d}$ are quantized into binary output $q_{k}$ as shown below.", + "bbox": [ + 89, + 670, + 483, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {k} = \\left\\{ \\begin{array}{l l} \\operatorname {s i g n} \\left(z _ {k}\\right) & \\text {l o o k u p - f r e e q u a n t i z e r} \\\\ \\frac {1}{\\sqrt {d}} \\operatorname {s i g n} \\left(\\frac {z _ {k}}{\\left| z _ {k} \\right|}\\right) & \\text {b i n a r y s p h e r i c a l q u a n t i z e r} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 787, + 483, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To encourage codebook utilization, an entropy penalty $\\mathcal{L}_{entropy} = \\mathbb{E}[H(q(z))] - H[\\mathbb{E}(q(z))]$ [30] is adopted. To obtain the distribution of $q(z)$ , we need to compute the similarities between the input $z$ and the whole codebook when", + "bbox": [ + 89, + 840, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "using LFQ. However, this leads to unaffordable space and time complexity of $O(2^d)$ . When the codebook dimension $d$ becomes large, e.g. 20, an out-of-memory (OOM) issue is faced as shown in Tab. 3. By contrast, since both input and output in BSQ are unit vectors, BSQ[85] proposes an approximation formula for the entropy penalty, reducing the computational complexity to $O(d)$ . As shown in Tab 3, there is no obvious increase in memory consumption for BSQ even when codebook size is $2^{64}$ . Unless otherwise stated, we adopt BSQ by default.", + "bbox": [ + 511, + 436, + 906, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Infinite-Vocabulary Classifier", + "text_level": 1, + "bbox": [ + 511, + 601, + 777, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The visual tokenizer obtains discrete labels by quantizing residual features. Consequently, the transformer predicts next-scale residual features' labels $\\mathbf{y}_k \\in [0, V_d)^{h_k \\times w_k}$ and optimizes the target through the cross-entropy loss. Previous works [64, 80] directly predict these index labels using a classifier of $V_d$ classes. However, it suffers from two drawbacks, huge computational costs and fuzzy supervision.", + "bbox": [ + 511, + 626, + 905, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Section 3.2, we exploit a bitwise VQ-VAE as the visual tokenizer, where the vocabulary size $V_{d}$ is extremely large. For example, if $V_{d} = 2^{32}$ and $h = 2048$ , a conventional classifier would require a weight matrix $W \\in \\mathbb{R}^{h \\times V_d}$ of 8.8 trillion parameters, which exceeds the limits of current computational resources.", + "bbox": [ + 511, + 733, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, VQ-VAE exploits the sign function during quantization as in Eq.4. After that, the positive elements are multiplied with the corresponding base and summed to get the index label $\\mathbf{y}_k(m,n)$ as in Eq.5, where $m\\in [0,h_k)$ and $n\\in [0,w_k)$ .", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15736", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {k} (m, n) = \\sum_ {p = 0} ^ {d - 1} \\mathbb {1} _ {r _ {k} (m, n, p) > 0} \\cdot 2 ^ {p} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 107, + 482, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Owing to the merits of the quantization method, slight perturbations to those near-zero features cause a significant change in the label. As a result, the conventional index-wise classifier [10, 64, 81] is difficult to optimize.", + "bbox": [ + 89, + 159, + 482, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address these problems in computation and optimization, we propose an Infinite-Vocabulary Classifier (IVC). In particular, instead of using a conventional classifier with $V_{d}$ classes, we use $d$ binary classifiers in parallel to predict if the next-scale residual $r_k(m,n,p)$ is positive or negative, where $d = \\log_2(V_d)$ . The proposed Infinite-Vocabulary Classifier is much more efficient in memory and parameters compared to the conventional classifier. When $V_{d} = 2^{16}$ and $h = 2048$ , it saves $99.95\\%$ parameters and GPU memory. Besides, when there exist near-zero values that confuse the model in some dimensions, the supervision in other dimensions is still clear. Therefore, compared with conventional index-wise classifiers, the proposed Infinite-Vocabulary Classifier is easier to optimize.", + "bbox": [ + 89, + 220, + 483, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Bitwise Self-Correction", + "text_level": 1, + "bbox": [ + 89, + 446, + 305, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Weakness of teacher-forcing training. VAR [64] inherits the teacher-forcing training from LLMs. However, next-scale prediction in vision is quite different from next-token prediction in language. Specifically, we cannot decode the complete image until residuals $r_k$ from all scales are obtained. We find that the teacher-forcing training brings about severe train-test discrepancies for visual generation. In particular, the teacher-forcing training makes the transformer only refine features in each scale without the ability to recognize and correct mistakes. Mistakes made in former scales will be propagated and amplified in latter scales, finally messing up generated images (left images in Fig.10).", + "bbox": [ + 89, + 470, + 482, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we propose the bitwise self-correction to address this issue. In particular, we obtain $r_k^{flip}$ via randomly flipping the bits in $r_k$ with a probability from $0\\%$ to $30\\%$ , imitating the errors in the prediction of the $k$ -th scale.", + "bbox": [ + 89, + 654, + 482, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here comes the key component of bitwise self-correction. $r_k^{flip}$ contains errors while $r_k$ doesn't. After replacing $r_k$ with $r_k^{flip}$ as predictions on the $k$ -th scale, we need to recompute the transformer input $\\widetilde{f}_k$ . Besides, requantization is performed to get a new target $r_{k+1}$ as in Eq.6. Refer to Fig.3 (right) for a simplified illustration.", + "bbox": [ + 89, + 715, + 482, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} f _ {k} ^ {\\text {f l i p}} = \\sum_ {i = 1} ^ {k} u p \\left(r _ {i} ^ {\\text {f l i p}}, (h, w)\\right) \\\\ \\widetilde {f} _ {k} = \\operatorname {d o w n} \\left(f _ {k} ^ {\\text {f l i p}}, \\left(h _ {k + 1}, w _ {k + 1}\\right)\\right) \\\\ r _ {k + 1} = \\operatorname {q u a n t} \\left(\\operatorname {d o w n} \\left(f - f _ {k} ^ {\\text {f l i p}}, \\left(h _ {k + 1}, w _ {k + 1}\\right)\\right)\\right) \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 818, + 482, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each scale undergoes the same process of random-flipping and re-quantization. Here the transformer takes partially randomly flipped features as inputs, taking the prediction errors into consideration. The re-quantized bit labels enable the transformer to auto-correct errors made in former predictions. In such a way, we address the train-test discrepancy caused by teacher-forcing training and empower Infinity to have the self-correction ability.", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Dynamic Resolution Generation", + "text_level": 1, + "bbox": [ + 511, + 222, + 797, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Infinity can generate photo-realistic images with various aspect ratios, which is significantly different from VAR [64] that can only generate square images. The main obstacles of generating various aspect ratio images lie in two folds. The first is to define the height $h_k$ and width $w_k$ of $r_k$ based on varying aspect ratios. In the supplementary material, we pre-define a list of scales, also called scale schedule, as $\\{(h_1^r, w_1^r), \\dots, (h_K^r, w_K^r)\\}$ for each aspect ratio. We ensure that the aspect ratio of each tuple $(h_k^r, w_k^r)$ is approximately equal to $r$ , especially in the latter prediction scales. Additionally, for different aspect ratios at the same scale $k$ , we keep the area of $h_k^r \\times w_k^r$ to be roughly equal, ensuring that the training sequence lengths are roughly the same.", + "bbox": [ + 511, + 244, + 906, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Secondly, we need to carefully design a resolution-aware positional encoding method to handle features of various scales and aspect ratios. This issue poses a significant challenge, as the existing solutions [26, 43, 57, 64, 68] exhibit substantial limitations under such conditions. In this paper, we apply RoPE2d [26] on features of each scale to preserve the intrinsic 2D structure of images. Additionally, we exploit learnable scale embeddings to avoid confusion between features of different scales. Compared to learnable APE element-wise applied on features, learnable embeddings applied on scales bring fewer parameters, can adapt to varying sequence lengths, and are easier to optimize.", + "bbox": [ + 511, + 441, + 908, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 637, + 638, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 511, + 662, + 609, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Curation. We curated a large-scale dataset from open-source academic data and high-quality internally collected data. The pre-training dataset is constructed by collecting and cleaning open-source academic datasets such as LAION [54], COYO [9], OpenImages [33]. We exploit an OCR model and a watermark detection model to filter undesired images with too many texts or watermarks. Additionally, we employ Aesthetic-V2 to filter out images with low aesthetic scores.", + "bbox": [ + 511, + 685, + 906, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Training", + "text_level": 1, + "bbox": [ + 511, + 830, + 617, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Progressive Training. Infinity redefines text-to-image as a coarse-to-fine, next-scale prediction task. In line with its architecture, we propose to train Infinity in a progressive", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15737", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "strategy. Specifically, we first train Infinity of 2B parameters on the pre-training dataset of over 100 million images with 256 resolution for 150k iterations using a batch size of 4096 and a learning rate of 6e-5. Then we switch to 512 resolution and train 110k iterations using the same hyperparameters. Next, we fine-tune Infinity at 1024 resolution with a smaller, high-quality dataset of 60 million images. In this stage, we train Infinity for 60k iterations using a batch size of 2048 and a learning rate of 2e-5. All training stages use images of varying aspect ratios.", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 89, + 253, + 326, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1 Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 277, + 277, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall Results. Fig.1 presents generated images from our Infinity-2B model, showcasing Infinity's strong capabilities in generating high-fidelity images from various categories following user prompts. Qualitative comparison results among Infinity and other top-tier models can be found in the supplementary materials.", + "bbox": [ + 89, + 301, + 483, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt-Following. Fig.11 in the appendix presents three examples demonstrating the superior prompt-following ability of Infinity. As highlighted in red, Infinity consistently adheres to user prompts, whether they are short or extremely long. We attribute these improvements to scaling autoregressive modeling and visual tokenizer's vocabulary.", + "bbox": [ + 89, + 393, + 483, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text Rendering. As illustrated in Fig.12 in the appendix, Infinity can render text according to user prompts across diverse categories. Despite diverse backgrounds and subjects, Infinity accurately renders corresponding texts according to user requirements, such as fonts, styles, colors, and more.", + "bbox": [ + 89, + 484, + 483, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2 Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 571, + 287, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark. As in Tab 1, on GenEval[24], our model with a re-writer achieves the best overall score of 0.73. Besides, Infinity also reaches the highest position reasoning score of 0.49. On DPG [29]. Our model reaches an overall score of 83.5, surpassing SDXL [46], Playground v2.5 [36], and DALLE 3 [6]. What's more, Infinity achieves the best relation score of 90.76 among all open-source T2I models, demonstrating its stronger ability to generate spatially consistent images based on user prompts.", + "bbox": [ + 89, + 597, + 483, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Human Preference Evaluation. We conduct human preference evaluation in both human studies and benchmarks. As in Fig.5, the generation results of Infinity are more frequently selected by humans in terms of overall quality, prompt following, and visual aesthetics in contrast to other open-sourced T2I models. Please refer to the supplementary material for more details. Tab.2 lists the results of three human preference benchmarks, i.e., ImageReward [77], HPSv2 [74], and VQAScore[39]. Infinity reaches the highest ImageReward and HPSv2.1, indicating our method could generate images that are more appealing to humans.", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bb33b2d224153267eb0423b8ed5f0d4d6760e7113812aaaeb6feb138adffaa93.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods# ParamsGenEval↑DPG↑
Two Obj.PositionColor Attri.OverallGlobalRelationOverall
Diffusion Models
LDM [52]1.4B0.290.020.050.37---
SDv1.5 [52]0.9B0.380.040.060.4374.6373.4963.2
PixArt-alpha [11]0.6B0.500.080.070.4874.9782.5771.1
SDv2.1 [52]0.9B0.510.070.170.5077.6780.7268.1
DALL-E 2 [49]6.5B0.660.100.190.52---
DALL-E 3 [6]----\\( 0.67^† \\)90.9790.5883.5
SDXL [46]2.6B0.740.150.230.5583.2786.7674.7
PixArt-Sigma [13]0.6B0.620.140.270.5586.8986.5980.5
SD3 (d=24) [22]2B0.740.340.360.62--84.1
SD3 (d=38) [22]8B0.890.340.470.71---
SANA-1.0 [75]1.6B---0.66--84.8
FLUX-dev [34]12B---0.67--84.0
FLUX-schnell [34]12B---0.71--84.8
AutoRegressive Models
LlamaGen [58]0.8B0.340.070.040.3265.2
Chameleon [62]7B---0.39---
HART [61]0.7B---0.56--80.9
Show-o [76]1.3B0.800.310.500.68--67.5
Emu3 [70]8.5B\\( 0.81^† \\)\\( 0.49^† \\)\\( 0.45^† \\)\\( 0.66^† \\)--81.6
Infinity2B\\( 0.85^† \\)\\( 0.49^† \\)\\( 0.57^† \\)\\( 0.73^† \\)93.1190.7683.5
", + "bbox": [ + 517, + 89, + 903, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Evaluation on the GenEval [24] and DPG [29] benchmark. A blue background indicates methods with larger model sizes than ours. $\\dagger$ result is with prompt rewriting.", + "bbox": [ + 511, + 297, + 906, + 339 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5dcdd0ea2045b3c7e7cb134d85da92a07dadd4d32390d37f4cb0be2509ae44c2.jpg", + "image_caption": [ + "Figure 4. Generation Latency (s) on $1024 \\times 1024$ . Infinity shows a significant advantage in generation latency compared to diffusion models." + ], + "image_footnote": [], + "bbox": [ + 522, + 354, + 888, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e9d1a01261deb23886cb2b5dcb1fc85737f30904106a3e733214f0c02d0c0702.jpg", + "image_caption": [ + "Figure 5. Human Preference Evaluation. We ask users to select the better one in a side-by-side comparison in terms of Overall Quality, Prompt Following, and Visual Aesthetics. Infinity is more preferred by humans compared to other methods." + ], + "image_footnote": [], + "bbox": [ + 557, + 513, + 846, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference Latency. As shown in Fig.4, Infinity demonstrates a significant advantage in generation speed compared to diffusion models across different model sizes. Without extra optimization, Infinity generates a $1024 \\times 1024$ image in 0.8s, $2.6 \\times$ faster than SD3-Medium. The speed advantage becomes more substantial as the model size increases. Infinity achieves $7 \\times$ faster inference latency compared to SD3.5 (8B) [22] and Flux-Dev (12B). Besides, Infinity still achieves $3.7 \\times$ speedup than Flux-Dev even with more parameters (20B v.s. 12B).", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15738", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4e65dae9ca756f853f51306daec807c7d738f333d8ecca51701b7e2e1480c625.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods# ParamsImageReward↑HPSv2.1↑VQAScore↑
RankScoreRankScoreRankTime
SD-XL [46]2.6B40.600430.0630.832
SD3-Medium [22]2B30.871330.9110.906
PixArt Sigma [13]0.6B20.872231.47--
Infinity2B10.953132.0520.894
", + "bbox": [ + 94, + 88, + 478, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Scaling Visual Tokenizer's Vocabulary", + "text_level": 1, + "bbox": [ + 89, + 237, + 419, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling up the Vocabulary Benefits Reconstruction. Restricted by the vocabulary size, discrete VQ-VAEs have always lagged behind continuous ones, hindering the performance of AR-based T2I models. In this work, we successfully train a discrete VQ-VAE matching its continuous counterparts by scaling up the vocabulary size. As in Tab. 4, we observe consistent rFID improvements as scaling up the vocabulary size from $2^{16}$ to $2^{64}$ . It's noteworthy that our discrete tokenizer achieves a rFID of 0.61 on ImageNet-256 when $V_{d} = 2^{32}$ , outperforming SD [52]'s contiguous VAE.", + "bbox": [ + 89, + 260, + 482, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Infinite Vocabulary Classifier Benefits Generation. We compare predicting bit labels with IVC to predicting index labels using a conventional classifier under the vocabulary size of $2^{16}$ since a larger vocabulary causes OOM for the conventional classifier. We utilize the reconstruction loss on $\\mathbf{R}_k$ , the Fréchet Inception Distance (FID) on the validation dataset and ImageReward for comprehensive evaluation. As shown in Tab.5, IVC achieves lower reconstruction loss and FID, suggesting IVC has better fitting capabilities. Beyond the quantitative results, training Infinity with IVC yields images with richer details as in Fig.6, which is consistent with a higher ImageReward.", + "bbox": [ + 89, + 412, + 482, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling Up the Vocabulary Benefits Generation. We then scale up the vocabulary size to $2^{32}$ during training the T2I model, which exceeds the range of the Int32 data type and can be considered infinitely large. In Fig.7, we illustrate the effect of scaling up the vocabulary from $2^{16}$ to $2^{32}$ for image generation. For small models (125M and 361M), the vocabulary size of $2^{16}$ converges faster and achieves better results. However, as we scaled up the transformer to 2.2B, the vocabulary size of $2^{32}$ beats $2^{16}$ after 40K iterations. Therefore, it's worthwhile to scale up the vocabulary along with scaling up the transformer. As illustrated in Tab.1,2, with infinite vocabulary and IVC, Infinity achieves superior performance among various benchmarks, redefining performance limits in autoregressive visual synthesis.", + "bbox": [ + 89, + 594, + 482, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Scaling Visual AutoRegressive Modeling", + "text_level": 1, + "bbox": [ + 89, + 816, + 434, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig.8, we depict the validation loss against the total training iterations and computational FLOPs for various model sizes of Infinity. We consistently notice a reduction in validation loss with an increase in training steps and computa", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/14871b2604d627dedf110fcfe5a984ff12d600889058693dd56b8ea2da8b1b90.jpg", + "table_caption": [ + "Table 2. Human Preference Metrics. Infinity reaches the highest ImageReward and HPSv2.1 score, also with comparable VQAS-score to SD3." + ], + "table_footnote": [], + "table_body": "
Quantizerd=16d=18d=20d=32d=64
LFQ37.653.7OOMOOMOOM
BSQ32.432.432.432.432.4
", + "bbox": [ + 545, + 88, + 870, + 137 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b9294d09a1650a05688fe481dbd190a5c7fc4ad97366b2c34c3282c55a9700eb.jpg", + "table_caption": [ + "Table 3. Comparison of memory consumption (GB) between different quantizers during training. As codebook dimension $d$ increases, MSR-BSQ shows significant advantages over MSR-LFQ, enabling nearly infinite vocabulary size of $2^{64}$ ." + ], + "table_footnote": [], + "table_body": "
VAE (stride=16)TYPEIN-256 rFID↓IN-512 rFID↓
Vd=216Discrete1.220.31
Vd=224Discrete0.750.30
Vd=232Discrete0.610.23
Vd=264Discrete0.330.15
SD VAE [52]Continual0.87N/A
", + "bbox": [ + 544, + 202, + 870, + 297 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4875f38f5087508b52926a82b6ec6b104b92df93da7b06662751e16ab13b26b8.jpg", + "table_caption": [ + "Table 4. By scaling up visual tokenizer's vocabulary, discrete tokenizer surpasses continuous VAE of SD [51] on ImageNet-rFID." + ], + "table_footnote": [], + "table_body": "
Classifier# ParamsvRAMRecons. Loss↓FID↓ImageReward↑
Convention124M2GB0.1845.950.79
IVC0.65M10MB0.1805.320.91
", + "bbox": [ + 517, + 332, + 901, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. In contrast to the conventional classifier, IVC saves $99.95\\%$ parameters (vRAM) and reaches better performance.", + "bbox": [ + 511, + 380, + 903, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c2668e9c39b7ed139637e89a375b29aa78f2c0f7ea3eeba4912ba9f690c84be.jpg", + "image_caption": [ + "Figure 6. Impact of Infinite-Vocabulary Classifier. Predicting bitwise labels with the Infinite-Vocabulary Classifier (Right) generates images with richer details compared to predicting index-wise labels using a conventional classifier (Left)." + ], + "image_footnote": [], + "bbox": [ + 521, + 412, + 900, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tional FLOPs. Nevertheless, the advantages gained from training smaller models for extended periods lag behind those obtained from training larger models for shorter durations. This trend aligns with findings in language models, emphasizing the promising outlook for increasing model sizes with appropriate training.", + "bbox": [ + 511, + 627, + 905, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig.8, we plot GenEval, ImageReward, and HPSv2 scores against validation loss for different model sizes ranging from 125M to 4.7B. We observe a strong correlation between validation loss and evaluation metrics. To further quantify their correlation, we calculate the Pearson correlation coefficients through linear regression. The correlation coefficients for GenEval, ImageReward, and HPSv2 are -0.983, -0.981, and -0.979, respectively. These results demonstrate a nearly linear correlation between validation loss and the evaluation metrics when scaling up model sizes from 125M to 4.7B. This promising phenomenon encourages us to scale up Infinity to achieve better performance.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15739", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3657c0a75737e60d587cc98522964eb09c2a7e9ed8652ffeea7ed65a9f31878c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 90, + 225, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7236248afff6ec234490ceab5a9e6ea49dbb2683706eb51d3da7a561b4a0d382.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 90, + 351, + 171 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5ed7059f5d1331ad41e4eace9131cbe2dd3e25befb2d8606e4fdd35d93f27c45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 90, + 478, + 171 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c12ed33fc6b5c76d1cc5a4ef9ac3717a4e02a00caa0b04c10cc9f5d02f7e5dd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 172, + 223, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cbef24ae022b36e2f60c9bf522df068337a9a8d30c20ef1f83bf49070c0e348c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 172, + 349, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d87333dfb3b6400225739574a4f6d3fa2aa3743eb42f8b27bfc255544d652407.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 172, + 478, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f8cefe67af63d230132a9b972eb2b46ef9b15a61752a255d4756ac9e3a7b8f4.jpg", + "image_caption": [ + "Figure 7. Effects of Scaling Up the Vocabulary. We analyze the impact of scaling the vocabulary size under consistent training hyperparameters. Vocabulary size $V_{d} = 2^{16}$ converges faster and achieves better results for small models (125M and 361M parameters). As we scale up the model size to 2.2B, Infinity with a vocabulary size $V_{d} = 2^{32}$ beats $V_{d} = 2^{16}$ . Experiment with 5M high-quality data under $256 \\times 256$ resolution." + ], + "image_footnote": [], + "bbox": [ + 101, + 255, + 223, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d372e3b5e9ea100b49fd408da60d9cc641722859e5bcfdee42002d37f1134a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 255, + 351, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d42ecf68d027aaef132a35fa5ee0537d243001455acc1b7f25a0c8bf60f0f79e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 255, + 478, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f0695e4cbdaa9b4cb848ef48843156eeb498849b8b1f494965c0d173bbb9bca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 440, + 222, + 529 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f21d9d2d0e161034b4d7db26ce485706acc3bf9dede5ed150b70a85aca8ed623.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 440, + 349, + 529 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c73df1284253e488e649031b73aae44c0a8121821655c6a78b7266a0267e4666.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 440, + 477, + 529 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/403493a93b7372ceffcf27358e714c138cc4fc5616467ca7ecae05f9c4445e23.jpg", + "image_caption": [ + "Figure 8. Scaling Visual AutoRegressive Modeling effects. We analyze the impact of scaling model size under consistent training hyperparameters (Experiment with 10M pre-training data under 256 resolution). Validation loss smoothly decreases as a function of the model size and training iters. Besides, validation loss is a strong predictor of overall performance. There is a strong correlation between validation loss and holistic image evaluation metrics." + ], + "image_footnote": [], + "bbox": [ + 96, + 531, + 220, + 609 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/754992fc02106e037ab433342be1b7d57a48f5c71ac034e22e5c90a9c2013047.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 531, + 349, + 609 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e0606cbb376018f3b27cf454bf6e228c19258124208168dcc13c080efd011cac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 531, + 477, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of scaling effect. To delve deeper into the scaling effect of Infinity, we compare a set of generated $256 \\times 256$ images of three model sizes (125M, 940M, 4.7B) across three distinct training schedules (10K, 40K, 90K iterations) as illustrated in Fig.9. The semantics and visual quality of generated images improve steadily when scaling up model size and training compute, which is consistent with the scaling behaviors of Infinity.", + "bbox": [ + 89, + 715, + 483, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Bitwise Self-Correction", + "text_level": 1, + "bbox": [ + 89, + 845, + 305, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab.6 and Fig.10, we list the evaluation metrics and present images generated by models trained using teacher-", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd4ed85f51c920bb7c7daaeb3694c863b3ab92b55b930420581f6f8fbf02b5f9.jpg", + "image_caption": [ + "Figure 9. Semantics and visual quality improve consistently with scaling up model size and training compute." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51edd01475b6da891194f4a3a1fce98d77ef4eaedf934b440d38a2cd75aa3cbc.jpg", + "image_caption": [ + "Figure 10. Impact of Self-Correction. Teacher-forcing training introduces great train-test discrepancy which degrades performance during inference (left). Bitwise Self-Correction autocorrects mistakes and thus generates better results (right)." + ], + "image_footnote": [], + "bbox": [ + 532, + 320, + 882, + 412 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dff2d468a23427772b8f73c13a341560887475dac1c2792f4a1fce31ec20a198.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID↓ImageReward↑HPSv2.1↑
Baseline9.760.5229.53
Baseline + Random Flip9.690.5229.20
Baseline + Bitwise Self-Correction3.470.7630.71
", + "bbox": [ + 517, + 470, + 901, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Bitwise Self-Correction makes significant improvements. Experiment with 5M high quality data and $512 \\times 512$ resolution. FID is measured on the validation set with 40K images.", + "bbox": [ + 511, + 531, + 903, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "forcing and bitwise self-correction methods. Substantial advantages are observed after applying bitwise self-correction, primarily driven by the self-correction mechanism rather than applying flipping. Simply random flipping $\\mathbf{R}_k$ doesn't bring improvements. Self-correction considers errors and performs re-quantization to correct them. We emphasize that self-correction is essential for AR-based T2I models since it empowers models to correct errors automatically, significantly mitigating the train-test discrepancy.", + "bbox": [ + 511, + 575, + 906, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 724, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce Infinity, a bitwise visual autoregressive model to perform Text-to-Image generation. Infinity is a pioneering framework for bitwise token modeling with the IVC and self-correction innovation. Extensive results demonstrate Infinity significantly raised the upper limit for Autoregressive Text-To-Image models, matching or surpassing leading diffusion models. We believe Infinity will significantly promote the development of autoregressive visual modeling and inspire the community for faster and more realistic generation models.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15740", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 89, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Many colleagues from ByteDance supported this work. We are grateful to Guanyang Deng for his efforts in data processing. We also thank Chongxi Wang and Taekmin Kim for their contributions to model deployment. Special thanks to Xiaoxiao Qin for her work in human preference evaluation. Additionally, we are thankful to Hui Wu, Fu Li, Xing Wang, Hongxiang Hao, and Chuan Li for their contributions to infrastructure.", + "bbox": [ + 89, + 114, + 485, + 238 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 251, + 187, + 266 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3", + "[2] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2", + "[3] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2", + "[4] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503, 2022. 3", + "[5] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22669-22679, 2023. 3", + "[6] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 2, 3, 6", + "[7] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. OpenAI, 2024. 2, 3", + "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3", + "[9] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.5", + "[10] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer" + ], + "bbox": [ + 93, + 276, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition, pages 11315-11325, 2022. 5", + "[11] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 6", + "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 3", + "[13] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2, 6, 7", + "[14] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2", + "[15] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202, 2022.3", + "[16] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 2, 3", + "[17] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 3", + "[18] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 3", + "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2", + "[20] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[21] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 3", + "[22] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15741", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Lijie Fan, Tianhong Li, Siyang Qin, Yuanshen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens, 2024. 3", + "[24] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36, 2024. 6", + "[25] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 2, 3", + "[26] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. In European Conference on Computer Vision, pages 289-305. Springer, 2025. 3, 5", + "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3", + "[28] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022. 2, 3", + "[29] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 6", + "[30] Aren Jansen, Daniel PW Ellis, Shawn Hershey, R Channing Moore, Manoj Plakal, Ashok C Popat, and Rif A Saurous. Coincidence, categorization, and consolidation: Learning to recognize sounds with minimal supervision. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 121-125. IEEE, 2020. 4", + "[31] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2, 3", + "[32] Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jose Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, Krishna Somandepalli, Hassan Akbari, Yair Alon, Yong Cheng, Josh Dillon, Agrim Gupta, Meera Hahn, Anja Hauth, David Hendon, Alonso Martinez, David Minnen, Mikhail Sirotenko, Kihyuk Sohn, Xuan Yang, Hartwig Adam, Ming-Hsuan Yang, Irfan Essa, Huisheng Wang, David A. Ross, Bryan Seybold, and Lu Jiang. Videopoet: A large language model for zero-shot video generation, 2024. 3", + "[33] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 5" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Black Forest Labs. Flux. https://blackforestlabs.ai/announcing-black-forest-labs/,2024.6", + "[35] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523-11532, 2022. 3", + "[36] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 6", + "[37] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. Advances in Neural Information Processing Systems, 37:56424-56445, 2024. 3", + "[38] Xiang Li, Hao Chen, Kai Qiu, Jason Kuen, Jiumiang Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. arXiv preprint arXiv:2410.01756, 2024. 2", + "[39] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2024. 6", + "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3", + "[41] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 3", + "[42] Chuofan Ma, Yi Jiang, Junfeng Wu, Jihan Yang, Xin Yu, Zehuan Yuan, Bingyue Peng, and Xiaojuan Qi. Unitok: A unified tokenizer for visual generation and understanding. arXiv preprint arXiv:2502.20321, 2025. 2", + "[43] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 5", + "[44] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 3", + "[45] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 3", + "[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 6, 7", + "[47] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15742", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2", + "[49] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2, 6", + "[50] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2, 3", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 7", + "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 6, 7", + "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2", + "[54] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5", + "[55] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3", + "[56] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3", + "[57] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5", + "[58] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2, 3, 4, 6", + "[59] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 3", + "[60] Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, et al. Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137, 2021. 2", + "[61] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 3, 6", + "[62] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024.3, 6", + "[63] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2", + "[64] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2404.02905, 2024. 2, 3, 4, 5", + "[65] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 3", + "[66] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3", + "[67] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3", + "[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5", + "[69] Junke Wang, Yi Jiang, Zehuan Yuan, Bingyue Peng, Zuxuan Wu, and Yu-Gang Jiang. Omnitokensizer: A joint image-video tokenizer for visual generation. Advances in Neural Information Processing Systems, 37:28281-28295, 2024. 2", + "[70] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 6", + "[71] Mark Weber, Lijun Yu, Qihang Yu, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. Maskbit: Embedding-free image generation via bit tokens. arXiv preprint arXiv:2409.16211, 2024. 4", + "[72] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2, 3", + "[73] Junfeng Wu, Yi Jiang, Chuofan Ma, Yuliang Liu, Hengshuang Zhao, Zehuan Yuan, Song Bai, and Xiang Bai. Liquid: Language models are scalable and unified multi-modal generators. arXiv preprint arXiv:2412.04332, 2024. 2, 3", + "[74] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15743", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6", + "[75] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 6", + "[76] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 6", + "[77] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 6", + "[78] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2", + "[79] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2, 3", + "[80] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 2, 3, 4", + "[81] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 2, 4, 5", + "[82] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 3", + "[83] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. arXiv preprint arXiv:2406.07550, 2024. 2", + "[84] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.3", + "[85] Yue Zhao, Yuanjun Xiong, and Philipp Krahenbuhl. Image and video tokenization with binary spherical quantization. arXiv preprint arXiv:2406.07548, 2024. 2, 3, 4" + ], + "bbox": [ + 91, + 90, + 483, + 854 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "15744", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/d1db9220-1c65-401b-9202-6e181bfed6b4_model.json b/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/d1db9220-1c65-401b-9202-6e181bfed6b4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cb0dad605578aed20640b2f251d975fec8fb58c8 --- /dev/null +++ b/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/d1db9220-1c65-401b-9202-6e181bfed6b4_model.json @@ -0,0 +1,2798 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.768, + 0.177 + ], + "angle": 0, + "content": "Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.204, + 0.726, + 0.258 + ], + "angle": 0, + "content": "Jian Han*, Jinlai Liu*, Yi Jiang*, Bin Yan, Yuqi Zhang, Zehuan Yuan†, Bingyue Peng, Xiaobing Liu ByteDance" + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.267, + 0.828, + 0.301 + ], + "angle": 0, + "content": "{hanjian.thul23, liujinlai.licio, jiangyi.enjoy, yanbin/master}@bytedance.com {zhangyuqi.hi, yuanzehuan, bingyue.peng, will.liu}@bytedance.com" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.307, + 0.756, + 0.343 + ], + "angle": 0, + "content": "Code: https://github.com/FoundationVision/Infinity Online demo: https://opensource.bytedance.com/gmpt/t2i/invite" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.351, + 0.255, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.351, + 0.387, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.351, + 0.591, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.351, + 0.737, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.351, + 0.904, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.703, + 0.908, + 0.731 + ], + "angle": 0, + "content": "Figure 1. High-resolution image synthesis results from our Infinity-2B model, showcasing its capabilities in precise prompt following, spatial reasoning, text rendering, and aesthetics across different styles and aspect ratios." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.74, + 0.327, + 0.756 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.772, + 0.483, + 0.879 + ], + "angle": 0, + "content": "We present Infinity, a Bitwise Visual AutoRegressive Modeling capable of generating high-resolution, photorealistic images following language instruction. Infinity refactors visual autoregressive model under a bitwise token prediction framework with an infinite-vocabulary classifier and bitwise self-correction mechanism. By theoretically expanding the tokenizer vocabulary size to infinity in Transformer, our" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.742, + 0.907, + 0.879 + ], + "angle": 0, + "content": "method significantly unleashes powerful scaling capabilities to infinity compared to vanilla VAR. Extensive experiments indicate Infinity outperforms AutoRegressive Text-to-Image models by large margins, matches or surpasses leading diffusion models. Without extra optimization, Infinity generates a \\(1024 \\times 1024\\) image in 0.8s, \\(2.6 \\times\\) faster than SD3-Medium, making it the fastest Text-to-Image model. All the code and models are available to promote further exploration of Infinity for visual generation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.348, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15733" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.236 + ], + "angle": 0, + "content": "Visual generation[20, 27, 45, 51, 55] has recently seen rapid advancements, enabling high-quality, high-resolution images and video synthesis[7, 22]. Text-to-image generation[6, 22, 46, 48, 49, 53] is one of the most challenging tasks due to its need for complex language adherence and intricate scene creation. Currently, visual generation is primarily divided into two main approaches: Diffusion models and AutoRegressive models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.236, + 0.484, + 0.508 + ], + "angle": 0, + "content": "Diffusion models[20, 22, 27, 45, 46, 55], trained to invert the forward paths of data toward random noise, effectively generate images through a continuous denoising process. AutoRegressive models[14, 21, 21, 58, 64, 73], on the other hand, harness the scalability and generalizability of language models[2, 3, 16, 28, 60, 63, 65, 66, 72] by employing a visual tokenizer[50, 67, 78] to convert images into discrete tokens and causally optimize these tokens, allowing image generation through next-token prediction or next-scale prediction. AutoRegressive models encounter significant challenges in high-resolution text-to-image synthesis[58, 70, 79]. They exhibit inferior reconstruction quality when utilizing discrete tokens as opposed to continuous tokens. Additionally, their generated visual contents are less detailed than those by diffusion models. Inefficiency and latency in visual generation, resulting from the raster scan method of next-token prediction, further exacerbate these issues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.795 + ], + "angle": 0, + "content": "Recently, Visual AutoRegressive Modeling (VAR)[64] redefined autoregressive learning on images as coarse-to-fine \"next-scale prediction\". It demonstrates superior generalization and scaling capabilities compared to diffusion transformers while requiring fewer steps. VAR leverages the powerful scaling properties of LLMs [25, 31] and can simultaneously refine previous scale steps, benefiting from the strengths of diffusion models as well. However, the index-wise discrete tokenizer[21, 38, 42, 64, 67, 69, 83] employed in AutoRegressive or Visual AutoRegressive models faces significant quantization errors with a limited vocabulary size resulting in difficulties in reconstructing fine-grained details, especially in high-resolution images. In the generation stage, index-wise tokens suffer from fuzzy supervision leading to visual detail loss and local distortions. Moreover, train-test discrepancies from teacher-forcing training, inherent to LLMs, amplify cumulative errors in visual details. These challenges make index-wise tokens a significant bottleneck for AutoRegressive models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We propose a novel bitwise modeling framework that substitutes index-wise tokens with bitwise tokens throughout the process. Our bitwise modeling framework consists of three primary modules: bitwise visual tokenizer, bitwise infinite-vocabulary classifier, and bitwise self-correction. Inspired by the success and widespread adoption of binary vector quantization[80, 85], we scaled up the tokenizer vo" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.091, + 0.905, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.265, + 0.907, + 0.308 + ], + "angle": 0, + "content": "Figure 2. Left: Conventional classifier predicts \\(2^{d}\\) indices. Right: Infinite-Vocabulary Classifier predicts \\(d\\) bits instead, being much more efficient than index-wise classifier when \\(d\\) is large." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.907, + 0.609 + ], + "angle": 0, + "content": "cabulary to \\(2^{64}\\), significantly surpassing all previous AutoRegressive model vocabularies[58, 81]. This expansion allows for reconstruction quality that far exceeds previous discrete tokenizers, achieving results comparable to continuous VAE[51], with scores improving from 0.87 to 0.33 on ImageNet-256 benchmark[19]. In Fig.2, we transform the conventional token prediction from a large integer into binary bits in a bitwise infinite-vocabulary classifier to address optimization and computation challenges, enabling the learning of massive vocabularies in Visual AutoRegressive models. Additionally, we incorporated bitwise self-correction during training by randomly flipping some bits to simulate prediction mistakes and re-quantizing the residual features, thus endowing the system with self-correcting capabilities. Our method, Infinity: Bitwise Visual AutoRegressive Modeling, maintains the scaling and speed advantages of VAR while achieving detailed reconstruction and generation quality comparable to that of continuous Diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.611, + 0.906, + 0.702 + ], + "angle": 0, + "content": "Infinity outperforms all previous AutoRegressive text-to-image models by large margins and matches or surpasses state-of-the-art diffusion models including SDXL[46], PixArt-Sigma[13],DALL-E3[6] and Stable-Diffusion 3[22] in both quantitative evaluations and human preference scores. The core contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.705, + 0.907, + 0.796 + ], + "angle": 0, + "content": "1. We propose Infinity, an autoregressive model utilizing Bitwise Modeling, which significantly enhances scalability and visual detail representation capabilities of discrete generative models. This framework opens up new possibilities of 'infinity' within the discrete generation community." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.796, + 0.906, + 0.871 + ], + "angle": 0, + "content": "2. Infinity demonstrates the potential of scaling tokenizers and transformers by achieving near-continuous tokenizer performance with its image tokenizer, surpassing diffusion models in high-quality text-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "3. Infinity enables an autoregressive text-to-image model to achieve exceptional prompt adherence and superior" + }, + { + "type": "list", + "bbox": [ + 0.531, + 0.705, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15734" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.127, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "image generation quality, while also delivering unprecedented inference speed." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.136, + 0.234, + 0.151 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.161, + 0.307, + 0.178 + ], + "angle": 0, + "content": "2.1. AutoRegressive Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.485, + 0.502 + ], + "angle": 0, + "content": "AutoRegressive models, leveraging the powerful scaling capabilities of LLMs[8, 16, 47, 65, 66], use discrete image tokenizers[21, 50, 67] in conjunction with transformers to generate images based on next-token prediction. VQ-based methods [21, 35, 50, 58, 67] employ vector quantization to convert image patches into index-wise tokens and use a decoder-only transformer to predict the next token index. However, these methods are limited by the lack of scaled-up transformers and the quantization error inherent in VQ-VAE[67], preventing them from achieving performance on par with diffusion models. Parti [79], Emu3 [70], chameleon[62], Liquid[73], and VideoPoet[32] scaled up autoregressive models in text-to-image or video synthesis. Inspired by the spatial structure of visual information, Visual AutoRegressive modeling(VAR)[64] redefines the autoregressive modeling on images as next-scale prediction framework, significantly improving generation quality and sampling speed. HART[61] adopted hybrid tokenizers based on VAR. MAR[37], Fluid[23] proposed random-order models and employed continuous tokenizer rather than discrete tokenizer." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.511, + 0.264, + 0.526 + ], + "angle": 0, + "content": "2.2. Diffusion Models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.761 + ], + "angle": 0, + "content": "Diffusion models have recently experienced rapid advancements in several directions. Denoising learning mechanisms [27, 44] and improvements in sampling efficiency [4, 40, 41, 55, 56] have been continuously optimized to generate high-quality images. Bit Diffusion [15] represents data with analog bits and then trains a continuous diffusion model. Latent diffusion models [51] were the first to propose modeling in the latent space rather than the pixel space. Recently, latent diffusion models [18, 22] have scaled up VAEs to enhance representations in the latent space. DiT [45] and U-ViT [5] employ more scalable transformers to model diffusion, achieving superior results. Consequently, mainstream text-to-image diffusion models [6, 12, 22] have adopted the DiT architecture due to its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.222, + 0.787 + ], + "angle": 0, + "content": "2.3. Scaling Law" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Scaling laws in autoregressive language models reveal a power-law relationship between model size, dataset size, and compute with test set cross-entropy loss [1, 25, 31]. These laws help predict larger model performance, leading to efficient resource allocation and ongoing improvements without saturation [8, 28, 65, 66, 72, 84]. This has inspired research into scaling in visual generation [7, 22, 59, 64, 82]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.108 + ], + "angle": 0, + "content": "3. Infinity Setup" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.799, + 0.132 + ], + "angle": 0, + "content": "3.1. Visual AutoRegressive Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.291 + ], + "angle": 0, + "content": "Infinity incorporates a visual tokenizer and a transformer for image synthesis. During the training stage, a sample consists of a text prompt \\( t \\) and a ground truth image \\( im \\). The proposed visual tokenizer first encodes the image \\( im \\) into a feature map \\( f \\in \\mathbb{R}^{h \\times w \\times C} \\) with stride \\( s \\) and then quantizes the feature map \\( f \\) into \\( K \\) multi-scale residual maps \\( (r_1, r_2, \\dots, r_K) \\). The resolution of \\( r_k \\) is \\( h_k \\times w_k \\) and it grows larger gradually from \\( k = 1 \\rightarrow K \\). Based on this sequence of residuals, we can gradually approximate the continuous feature \\( f \\) as in Eq.1" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.301, + 0.906, + 0.342 + ], + "angle": 0, + "content": "\\[\nf _ {k} = \\sum_ {i = 1} ^ {k} u p \\left(r _ {i}, (h, w)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.353, + 0.907, + 0.384 + ], + "angle": 0, + "content": "up means bilinear upsampling and \\( f_{k} \\) is the cumulative sum of the upsampled \\( r_{\\leq k} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.906, + 0.445 + ], + "angle": 0, + "content": "Subsequently, transformer learns to predict residuals \\( r \\) of the next scale conditioned on previous predictions and the text input in an autoregressive manner. Formally, the autoregressive likelihood can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.456, + 0.907, + 0.497 + ], + "angle": 0, + "content": "\\[\np \\left(r _ {1}, \\dots , r _ {K}\\right) = \\prod_ {k = 1} ^ {K} p \\left(r _ {k} \\mid r _ {1}, \\dots , r _ {k - 1}, \\psi (t)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.909, + 0.693 + ], + "angle": 0, + "content": "where \\(\\psi(t)\\) is the text embeddings from Flan-T5 [17] model. \\((r_1, \\dots, r_{k-1}, \\psi(t))\\) serves as the prefixed context When predicting \\(r_k\\). Besides, the text embeddings \\(\\psi(t)\\) further guide the prediction through a cross-attention mechanism. In particular, as shown in Fig. 3, the text embeddings \\(\\psi(t) \\in \\mathbb{R}^{L \\times C}\\) is projected into a \\(\\langle \\mathrm{SOS} \\rangle \\in \\mathbb{R}^{1 \\times 1 \\times h}\\) as the input of the first scale, where \\(h\\) is the hidden dimension of transformer. The transformer is required to predict \\(r_1\\) based on \\(\\langle \\mathrm{SOS} \\rangle\\) in the first scale. In the latter \\(k\\)-th scale, to match the spatial size of the input and the output label \\(R_k\\), we take the downsampled feature \\(\\widetilde{f}_{k-1}\\) from the last scale \\(k-1\\) as the input to predict \\(R_k\\) in parallel." + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.703, + 0.907, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\widetilde {f} _ {k - 1} = \\operatorname {d o w n} \\left(f _ {k - 1}, \\left(h _ {k}, w _ {k}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.907, + 0.81 + ], + "angle": 0, + "content": "where \\( down \\) is bilinear downsampling and the spatial size of both \\( \\widetilde{f}_{k-1} \\) and \\( r_k \\) is \\( (h_k, w_k) \\). In previous index-wise [64] representations, the shape of \\( r_k \\) is \\( (h_k, w_k, V_d) \\). \\( V_d \\) is the vocabulary size of the visual tokenizer. For binary quantization [80, 85] with code embedding dimension \\( d \\), \\( V_d = 2^d \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The transformer consists of a stack of repeated blocks, where each block includes RoPE2D [26], Self-Attention, Cross Attention, and FFN layers. The text embeddings \\(\\psi(t)\\) provide effective guidance for image synthesis in each cross-attention layer. During the training stage, we exploit a block-wise causal attention mask to ensure that the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15735" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.168, + 0.093, + 0.531, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.09, + 0.825, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.401, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Figure 3. Framework of Infinity. Infinity introduces bitwise modeling. When predicting \\( r_k \\), the sequence \\( (r_1, r_2, \\dots, r_{k-1}) \\) serves as the prefixed context and the text embeddings \\( \\psi(t) \\) guides the prediction through a cross attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.501 + ], + "angle": 0, + "content": "transformer can only attend to its prefixed context, i.e., \\((\\langle \\mathrm{SOS}\\rangle ,\\widetilde{f}_1,\\dots ,\\widetilde{f}_{k - 1})\\) , when predicting \\(f_{k}\\) . During the inference stage, we perform KV-Caching to speed up inference and there's no need for masking." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.511, + 0.266, + 0.525 + ], + "angle": 0, + "content": "3.2. Infinite Tokenizer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.67 + ], + "angle": 0, + "content": "Increasing the vocabulary size has significant potential for improving reconstruction and generation quality. However, directly enlarging the vocabulary in existing tokenizers[58, 64, 71, 81] leads to a substantial increase in memory consumption and computational burden. To address these challenges and fully exploit the potential of discrete tokenizers, this paper proposes a new bitwise multi-scale residual quantizer, which significantly reduces memory usage, enabling the training of extremely large vocabulary, e.g. \\(2^{64}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Bitwise Multi-scale Residual Quantizer. We replace the original vector quantizer of VAR [64] with a dimension-independent bitwise quantizer. In this paper, we consider two candidates, LFQ [81] and BSQ[85]. Given \\(K\\) scales in the multi-scale quantizer, on the \\(k\\)-th scale, the input continuous residual vector \\(z_{k} \\in \\mathbb{R}^{d}\\) are quantized into binary output \\(q_{k}\\) as shown below." + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.789, + 0.484, + 0.831 + ], + "angle": 0, + "content": "\\[\nq _ {k} = \\left\\{ \\begin{array}{l l} \\operatorname {s i g n} \\left(z _ {k}\\right) & \\text {l o o k u p - f r e e q u a n t i z e r} \\\\ \\frac {1}{\\sqrt {d}} \\operatorname {s i g n} \\left(\\frac {z _ {k}}{\\left| z _ {k} \\right|}\\right) & \\text {b i n a r y s p h e r i c a l q u a n t i z e r} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.485, + 0.903 + ], + "angle": 0, + "content": "To encourage codebook utilization, an entropy penalty \\(\\mathcal{L}_{entropy} = \\mathbb{E}[H(q(z))] - H[\\mathbb{E}(q(z))]\\) [30] is adopted. To obtain the distribution of \\(q(z)\\), we need to compute the similarities between the input \\(z\\) and the whole codebook when" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.59 + ], + "angle": 0, + "content": "using LFQ. However, this leads to unaffordable space and time complexity of \\( O(2^d) \\). When the codebook dimension \\( d \\) becomes large, e.g. 20, an out-of-memory (OOM) issue is faced as shown in Tab. 3. By contrast, since both input and output in BSQ are unit vectors, BSQ[85] proposes an approximation formula for the entropy penalty, reducing the computational complexity to \\( O(d) \\). As shown in Tab 3, there is no obvious increase in memory consumption for BSQ even when codebook size is \\( 2^{64} \\). Unless otherwise stated, we adopt BSQ by default." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.779, + 0.619 + ], + "angle": 0, + "content": "3.3. Infinite-Vocabulary Classifier" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.906, + 0.734 + ], + "angle": 0, + "content": "The visual tokenizer obtains discrete labels by quantizing residual features. Consequently, the transformer predicts next-scale residual features' labels \\(\\mathbf{y}_k \\in [0, V_d)^{h_k \\times w_k}\\) and optimizes the target through the cross-entropy loss. Previous works [64, 80] directly predict these index labels using a classifier of \\(V_d\\) classes. However, it suffers from two drawbacks, huge computational costs and fuzzy supervision." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.734, + 0.907, + 0.825 + ], + "angle": 0, + "content": "As illustrated in Section 3.2, we exploit a bitwise VQ-VAE as the visual tokenizer, where the vocabulary size \\(V_{d}\\) is extremely large. For example, if \\(V_{d} = 2^{32}\\) and \\(h = 2048\\), a conventional classifier would require a weight matrix \\(W \\in \\mathbb{R}^{h \\times V_d}\\) of 8.8 trillion parameters, which exceeds the limits of current computational resources." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Moreover, VQ-VAE exploits the sign function during quantization as in Eq.4. After that, the positive elements are multiplied with the corresponding base and summed to get the index label \\(\\mathbf{y}_k(m,n)\\) as in Eq.5, where \\(m\\in [0,h_k)\\) and \\(n\\in [0,w_k)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15736" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.172, + 0.108, + 0.483, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {k} (m, n) = \\sum_ {p = 0} ^ {d - 1} \\mathbb {1} _ {r _ {k} (m, n, p) > 0} \\cdot 2 ^ {p} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.16, + 0.483, + 0.221 + ], + "angle": 0, + "content": "Owing to the merits of the quantization method, slight perturbations to those near-zero features cause a significant change in the label. As a result, the conventional index-wise classifier [10, 64, 81] is difficult to optimize." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.484, + 0.435 + ], + "angle": 0, + "content": "To address these problems in computation and optimization, we propose an Infinite-Vocabulary Classifier (IVC). In particular, instead of using a conventional classifier with \\( V_{d} \\) classes, we use \\( d \\) binary classifiers in parallel to predict if the next-scale residual \\( r_k(m,n,p) \\) is positive or negative, where \\( d = \\log_2(V_d) \\). The proposed Infinite-Vocabulary Classifier is much more efficient in memory and parameters compared to the conventional classifier. When \\( V_{d} = 2^{16} \\) and \\( h = 2048 \\), it saves \\( 99.95\\% \\) parameters and GPU memory. Besides, when there exist near-zero values that confuse the model in some dimensions, the supervision in other dimensions is still clear. Therefore, compared with conventional index-wise classifiers, the proposed Infinite-Vocabulary Classifier is easier to optimize." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.448, + 0.307, + 0.463 + ], + "angle": 0, + "content": "3.4. Bitwise Self-Correction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.483, + 0.653 + ], + "angle": 0, + "content": "Weakness of teacher-forcing training. VAR [64] inherits the teacher-forcing training from LLMs. However, next-scale prediction in vision is quite different from next-token prediction in language. Specifically, we cannot decode the complete image until residuals \\( r_k \\) from all scales are obtained. We find that the teacher-forcing training brings about severe train-test discrepancies for visual generation. In particular, the teacher-forcing training makes the transformer only refine features in each scale without the ability to recognize and correct mistakes. Mistakes made in former scales will be propagated and amplified in latter scales, finally messing up generated images (left images in Fig.10)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.716 + ], + "angle": 0, + "content": "In this work, we propose the bitwise self-correction to address this issue. In particular, we obtain \\( r_k^{flip} \\) via randomly flipping the bits in \\( r_k \\) with a probability from \\( 0\\% \\) to \\( 30\\% \\), imitating the errors in the prediction of the \\( k \\)-th scale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.813 + ], + "angle": 0, + "content": "Here comes the key component of bitwise self-correction. \\( r_k^{flip} \\) contains errors while \\( r_k \\) doesn't. After replacing \\( r_k \\) with \\( r_k^{flip} \\) as predictions on the \\( k \\)-th scale, we need to recompute the transformer input \\( \\widetilde{f}_k \\). Besides, requantization is performed to get a new target \\( r_{k+1} \\) as in Eq.6. Refer to Fig.3 (right) for a simplified illustration." + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.819, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} f _ {k} ^ {\\text {f l i p}} = \\sum_ {i = 1} ^ {k} u p \\left(r _ {i} ^ {\\text {f l i p}}, (h, w)\\right) \\\\ \\widetilde {f} _ {k} = \\operatorname {d o w n} \\left(f _ {k} ^ {\\text {f l i p}}, \\left(h _ {k + 1}, w _ {k + 1}\\right)\\right) \\\\ r _ {k + 1} = \\operatorname {q u a n t} \\left(\\operatorname {d o w n} \\left(f - f _ {k} ^ {\\text {f l i p}}, \\left(h _ {k + 1}, w _ {k + 1}\\right)\\right)\\right) \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "Each scale undergoes the same process of random-flipping and re-quantization. Here the transformer takes partially randomly flipped features as inputs, taking the prediction errors into consideration. The re-quantized bit labels enable the transformer to auto-correct errors made in former predictions. In such a way, we address the train-test discrepancy caused by teacher-forcing training and empower Infinity to have the self-correction ability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.223, + 0.798, + 0.238 + ], + "angle": 0, + "content": "3.5. Dynamic Resolution Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Infinity can generate photo-realistic images with various aspect ratios, which is significantly different from VAR [64] that can only generate square images. The main obstacles of generating various aspect ratio images lie in two folds. The first is to define the height \\( h_k \\) and width \\( w_k \\) of \\( r_k \\) based on varying aspect ratios. In the supplementary material, we pre-define a list of scales, also called scale schedule, as \\( \\{(h_1^r, w_1^r), \\dots, (h_K^r, w_K^r)\\} \\) for each aspect ratio. We ensure that the aspect ratio of each tuple \\( (h_k^r, w_k^r) \\) is approximately equal to \\( r \\), especially in the latter prediction scales. Additionally, for different aspect ratios at the same scale \\( k \\), we keep the area of \\( h_k^r \\times w_k^r \\) to be roughly equal, ensuring that the training sequence lengths are roughly the same." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.909, + 0.625 + ], + "angle": 0, + "content": "Secondly, we need to carefully design a resolution-aware positional encoding method to handle features of various scales and aspect ratios. This issue poses a significant challenge, as the existing solutions [26, 43, 57, 64, 68] exhibit substantial limitations under such conditions. In this paper, we apply RoPE2d [26] on features of each scale to preserve the intrinsic 2D structure of images. Additionally, we exploit learnable scale embeddings to avoid confusion between features of different scales. Compared to learnable APE element-wise applied on features, learnable embeddings applied on scales bring fewer parameters, can adapt to varying sequence lengths, and are easier to optimize." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.638, + 0.64, + 0.656 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.663, + 0.61, + 0.678 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.907, + 0.822 + ], + "angle": 0, + "content": "Data Curation. We curated a large-scale dataset from open-source academic data and high-quality internally collected data. The pre-training dataset is constructed by collecting and cleaning open-source academic datasets such as LAION [54], COYO [9], OpenImages [33]. We exploit an OCR model and a watermark detection model to filter undesired images with too many texts or watermarks. Additionally, we employ Aesthetic-V2 to filter out images with low aesthetic scores." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.619, + 0.848 + ], + "angle": 0, + "content": "4.2. Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Progressive Training. Infinity redefines text-to-image as a coarse-to-fine, next-scale prediction task. In line with its architecture, we propose to train Infinity in a progressive" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15737" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.244 + ], + "angle": 0, + "content": "strategy. Specifically, we first train Infinity of 2B parameters on the pre-training dataset of over 100 million images with 256 resolution for 150k iterations using a batch size of 4096 and a learning rate of 6e-5. Then we switch to 512 resolution and train 110k iterations using the same hyperparameters. Next, we fine-tune Infinity at 1024 resolution with a smaller, high-quality dataset of 60 million images. In this stage, we train Infinity for 60k iterations using a batch size of 2048 and a learning rate of 2e-5. All training stages use images of varying aspect ratios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.255, + 0.328, + 0.27 + ], + "angle": 0, + "content": "4.3. Text-to-Image Generation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.278, + 0.279, + 0.294 + ], + "angle": 0, + "content": "4.3.1 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.394 + ], + "angle": 0, + "content": "Overall Results. Fig.1 presents generated images from our Infinity-2B model, showcasing Infinity's strong capabilities in generating high-fidelity images from various categories following user prompts. Qualitative comparison results among Infinity and other top-tier models can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.484, + 0.485 + ], + "angle": 0, + "content": "Prompt-Following. Fig.11 in the appendix presents three examples demonstrating the superior prompt-following ability of Infinity. As highlighted in red, Infinity consistently adheres to user prompts, whether they are short or extremely long. We attribute these improvements to scaling autoregressive modeling and visual tokenizer's vocabulary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.562 + ], + "angle": 0, + "content": "Text Rendering. As illustrated in Fig.12 in the appendix, Infinity can render text according to user prompts across diverse categories. Despite diverse backgrounds and subjects, Infinity accurately renders corresponding texts according to user requirements, such as fonts, styles, colors, and more." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.573, + 0.289, + 0.589 + ], + "angle": 0, + "content": "4.3.2 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Benchmark. As in Tab 1, on GenEval[24], our model with a re-writer achieves the best overall score of 0.73. Besides, Infinity also reaches the highest position reasoning score of 0.49. On DPG [29]. Our model reaches an overall score of 83.5, surpassing SDXL [46], Playground v2.5 [36], and DALLE 3 [6]. What's more, Infinity achieves the best relation score of 90.76 among all open-source T2I models, demonstrating its stronger ability to generate spatially consistent images based on user prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Human Preference Evaluation. We conduct human preference evaluation in both human studies and benchmarks. As in Fig.5, the generation results of Infinity are more frequently selected by humans in terms of overall quality, prompt following, and visual aesthetics in contrast to other open-sourced T2I models. Please refer to the supplementary material for more details. Tab.2 lists the results of three human preference benchmarks, i.e., ImageReward [77], HPSv2 [74], and VQAScore[39]. Infinity reaches the highest ImageReward and HPSv2.1, indicating our method could generate images that are more appealing to humans." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.287 + ], + "angle": 0, + "content": "
Methods# ParamsGenEval↑DPG↑
Two Obj.PositionColor Attri.OverallGlobalRelationOverall
Diffusion Models
LDM [52]1.4B0.290.020.050.37---
SDv1.5 [52]0.9B0.380.040.060.4374.6373.4963.2
PixArt-alpha [11]0.6B0.500.080.070.4874.9782.5771.1
SDv2.1 [52]0.9B0.510.070.170.5077.6780.7268.1
DALL-E 2 [49]6.5B0.660.100.190.52---
DALL-E 3 [6]----\\( 0.67^† \\)90.9790.5883.5
SDXL [46]2.6B0.740.150.230.5583.2786.7674.7
PixArt-Sigma [13]0.6B0.620.140.270.5586.8986.5980.5
SD3 (d=24) [22]2B0.740.340.360.62--84.1
SD3 (d=38) [22]8B0.890.340.470.71---
SANA-1.0 [75]1.6B---0.66--84.8
FLUX-dev [34]12B---0.67--84.0
FLUX-schnell [34]12B---0.71--84.8
AutoRegressive Models
LlamaGen [58]0.8B0.340.070.040.3265.2
Chameleon [62]7B---0.39---
HART [61]0.7B---0.56--80.9
Show-o [76]1.3B0.800.310.500.68--67.5
Emu3 [70]8.5B\\( 0.81^† \\)\\( 0.49^† \\)\\( 0.45^† \\)\\( 0.66^† \\)--81.6
Infinity2B\\( 0.85^† \\)\\( 0.49^† \\)\\( 0.57^† \\)\\( 0.73^† \\)93.1190.7683.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.298, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Table 1. Evaluation on the GenEval [24] and DPG [29] benchmark. A blue background indicates methods with larger model sizes than ours. \\(\\dagger\\) result is with prompt rewriting." + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.356, + 0.89, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.466, + 0.907, + 0.506 + ], + "angle": 0, + "content": "Figure 4. Generation Latency (s) on \\(1024 \\times 1024\\). Infinity shows a significant advantage in generation latency compared to diffusion models." + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.515, + 0.847, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.675, + 0.907, + 0.731 + ], + "angle": 0, + "content": "Figure 5. Human Preference Evaluation. We ask users to select the better one in a side-by-side comparison in terms of Overall Quality, Prompt Following, and Visual Aesthetics. Infinity is more preferred by humans compared to other methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Inference Latency. As shown in Fig.4, Infinity demonstrates a significant advantage in generation speed compared to diffusion models across different model sizes. Without extra optimization, Infinity generates a \\(1024 \\times 1024\\) image in 0.8s, \\(2.6 \\times\\) faster than SD3-Medium. The speed advantage becomes more substantial as the model size increases. Infinity achieves \\(7 \\times\\) faster inference latency compared to SD3.5 (8B) [22] and Flux-Dev (12B). Besides, Infinity still achieves \\(3.7 \\times\\) speedup than Flux-Dev even with more parameters (20B v.s. 12B)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15738" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.479, + 0.187 + ], + "angle": 0, + "content": "
Methods# ParamsImageReward↑HPSv2.1↑VQAScore↑
RankScoreRankScoreRankTime
SD-XL [46]2.6B40.600430.0630.832
SD3-Medium [22]2B30.871330.9110.906
PixArt Sigma [13]0.6B20.872231.47--
Infinity2B10.953132.0520.894
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.189, + 0.483, + 0.232 + ], + "angle": 0, + "content": "Table 2. Human Preference Metrics. Infinity reaches the highest ImageReward and HPSv2.1 score, also with comparable VQAS-score to SD3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.238, + 0.421, + 0.254 + ], + "angle": 0, + "content": "4.4. Scaling Visual Tokenizer's Vocabulary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.413 + ], + "angle": 0, + "content": "Scaling up the Vocabulary Benefits Reconstruction. Restricted by the vocabulary size, discrete VQ-VAEs have always lagged behind continuous ones, hindering the performance of AR-based T2I models. In this work, we successfully train a discrete VQ-VAE matching its continuous counterparts by scaling up the vocabulary size. As in Tab. 4, we observe consistent rFID improvements as scaling up the vocabulary size from \\(2^{16}\\) to \\(2^{64}\\). It's noteworthy that our discrete tokenizer achieves a rFID of 0.61 on ImageNet-256 when \\(V_{d} = 2^{32}\\), outperforming SD [52]'s contiguous VAE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.483, + 0.594 + ], + "angle": 0, + "content": "Infinite Vocabulary Classifier Benefits Generation. We compare predicting bit labels with IVC to predicting index labels using a conventional classifier under the vocabulary size of \\(2^{16}\\) since a larger vocabulary causes OOM for the conventional classifier. We utilize the reconstruction loss on \\(\\mathbf{R}_k\\), the Fréchet Inception Distance (FID) on the validation dataset and ImageReward for comprehensive evaluation. As shown in Tab.5, IVC achieves lower reconstruction loss and FID, suggesting IVC has better fitting capabilities. Beyond the quantitative results, training Infinity with IVC yields images with richer details as in Fig.6, which is consistent with a higher ImageReward." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Scaling Up the Vocabulary Benefits Generation. We then scale up the vocabulary size to \\(2^{32}\\) during training the T2I model, which exceeds the range of the Int32 data type and can be considered infinitely large. In Fig.7, we illustrate the effect of scaling up the vocabulary from \\(2^{16}\\) to \\(2^{32}\\) for image generation. For small models (125M and 361M), the vocabulary size of \\(2^{16}\\) converges faster and achieves better results. However, as we scaled up the transformer to 2.2B, the vocabulary size of \\(2^{32}\\) beats \\(2^{16}\\) after 40K iterations. Therefore, it's worthwhile to scale up the vocabulary along with scaling up the transformer. As illustrated in Tab.1,2, with infinite vocabulary and IVC, Infinity achieves superior performance among various benchmarks, redefining performance limits in autoregressive visual synthesis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.436, + 0.833 + ], + "angle": 0, + "content": "4.5. Scaling Visual AutoRegressive Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In Fig.8, we depict the validation loss against the total training iterations and computational FLOPs for various model sizes of Infinity. We consistently notice a reduction in validation loss with an increase in training steps and computa" + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.089, + 0.872, + 0.138 + ], + "angle": 0, + "content": "
Quantizerd=16d=18d=20d=32d=64
LFQ37.653.7OOMOOMOOM
BSQ32.432.432.432.432.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.141, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Table 3. Comparison of memory consumption (GB) between different quantizers during training. As codebook dimension \\(d\\) increases, MSR-BSQ shows significant advantages over MSR-LFQ, enabling nearly infinite vocabulary size of \\(2^{64}\\)." + }, + { + "type": "table", + "bbox": [ + 0.545, + 0.203, + 0.872, + 0.299 + ], + "angle": 0, + "content": "
VAE (stride=16)TYPEIN-256 rFID↓IN-512 rFID↓
Vd=216Discrete1.220.31
Vd=224Discrete0.750.30
Vd=232Discrete0.610.23
Vd=264Discrete0.330.15
SD VAE [52]Continual0.87N/A
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.3, + 0.905, + 0.328 + ], + "angle": 0, + "content": "Table 4. By scaling up visual tokenizer's vocabulary, discrete tokenizer surpasses continuous VAE of SD [51] on ImageNet-rFID." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.333, + 0.903, + 0.378 + ], + "angle": 0, + "content": "
Classifier# ParamsvRAMRecons. Loss↓FID↓ImageReward↑
Convention124M2GB0.1845.950.79
IVC0.65M10MB0.1805.320.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.381, + 0.905, + 0.409 + ], + "angle": 0, + "content": "Table 5. In contrast to the conventional classifier, IVC saves \\(99.95\\%\\) parameters (vRAM) and reaches better performance." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.414, + 0.901, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.565, + 0.905, + 0.62 + ], + "angle": 0, + "content": "Figure 6. Impact of Infinite-Vocabulary Classifier. Predicting bitwise labels with the Infinite-Vocabulary Classifier (Right) generates images with richer details compared to predicting index-wise labels using a conventional classifier (Left)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.718 + ], + "angle": 0, + "content": "tional FLOPs. Nevertheless, the advantages gained from training smaller models for extended periods lag behind those obtained from training larger models for shorter durations. This trend aligns with findings in language models, emphasizing the promising outlook for increasing model sizes with appropriate training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In Fig.8, we plot GenEval, ImageReward, and HPSv2 scores against validation loss for different model sizes ranging from 125M to 4.7B. We observe a strong correlation between validation loss and evaluation metrics. To further quantify their correlation, we calculate the Pearson correlation coefficients through linear regression. The correlation coefficients for GenEval, ImageReward, and HPSv2 are -0.983, -0.981, and -0.979, respectively. These results demonstrate a nearly linear correlation between validation loss and the evaluation metrics when scaling up model sizes from 125M to 4.7B. This promising phenomenon encourages us to scale up Infinity to achieve better performance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15739" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.091, + 0.226, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.092, + 0.352, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.092, + 0.479, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.174, + 0.225, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.174, + 0.351, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.174, + 0.479, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.256, + 0.225, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.256, + 0.352, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.256, + 0.479, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.338, + 0.484, + 0.435 + ], + "angle": 0, + "content": "Figure 7. Effects of Scaling Up the Vocabulary. We analyze the impact of scaling the vocabulary size under consistent training hyperparameters. Vocabulary size \\( V_{d} = 2^{16} \\) converges faster and achieves better results for small models (125M and 361M parameters). As we scale up the model size to 2.2B, Infinity with a vocabulary size \\( V_{d} = 2^{32} \\) beats \\( V_{d} = 2^{16} \\). Experiment with 5M high-quality data under \\( 256 \\times 256 \\) resolution." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.441, + 0.223, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.441, + 0.351, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.441, + 0.478, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.532, + 0.222, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.532, + 0.351, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.532, + 0.478, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.71 + ], + "angle": 0, + "content": "Figure 8. Scaling Visual AutoRegressive Modeling effects. We analyze the impact of scaling model size under consistent training hyperparameters (Experiment with 10M pre-training data under 256 resolution). Validation loss smoothly decreases as a function of the model size and training iters. Besides, validation loss is a strong predictor of overall performance. There is a strong correlation between validation loss and holistic image evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.716, + 0.484, + 0.837 + ], + "angle": 0, + "content": "Visualization of scaling effect. To delve deeper into the scaling effect of Infinity, we compare a set of generated \\(256 \\times 256\\) images of three model sizes (125M, 940M, 4.7B) across three distinct training schedules (10K, 40K, 90K iterations) as illustrated in Fig.9. The semantics and visual quality of generated images improve steadily when scaling up model size and training compute, which is consistent with the scaling behaviors of Infinity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.307, + 0.862 + ], + "angle": 0, + "content": "4.6. Bitwise Self-Correction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In Tab.6 and Fig.10, we list the evaluation metrics and present images generated by models trained using teacher-" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.293, + 0.905, + 0.32 + ], + "angle": 0, + "content": "Figure 9. Semantics and visual quality improve consistently with scaling up model size and training compute." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.321, + 0.883, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.414, + 0.905, + 0.469 + ], + "angle": 0, + "content": "Figure 10. Impact of Self-Correction. Teacher-forcing training introduces great train-test discrepancy which degrades performance during inference (left). Bitwise Self-Correction autocorrects mistakes and thus generates better results (right)." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.472, + 0.903, + 0.529 + ], + "angle": 0, + "content": "
MethodFID↓ImageReward↑HPSv2.1↑
Baseline9.760.5229.53
Baseline + Random Flip9.690.5229.20
Baseline + Bitwise Self-Correction3.470.7630.71
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.532, + 0.905, + 0.574 + ], + "angle": 0, + "content": "Table 6. Bitwise Self-Correction makes significant improvements. Experiment with 5M high quality data and \\(512 \\times 512\\) resolution. FID is measured on the validation set with 40K images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.907, + 0.712 + ], + "angle": 0, + "content": "forcing and bitwise self-correction methods. Substantial advantages are observed after applying bitwise self-correction, primarily driven by the self-correction mechanism rather than applying flipping. Simply random flipping \\(\\mathbf{R}_k\\) doesn't bring improvements. Self-correction considers errors and performs re-quantization to correct them. We emphasize that self-correction is essential for AR-based T2I models since it empowers models to correct errors automatically, significantly mitigating the train-test discrepancy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.634, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce Infinity, a bitwise visual autoregressive model to perform Text-to-Image generation. Infinity is a pioneering framework for bitwise token modeling with the IVC and self-correction innovation. Extensive results demonstrate Infinity significantly raised the upper limit for Autoregressive Text-To-Image models, matching or surpassing leading diffusion models. We believe Infinity will significantly promote the development of autoregressive visual modeling and inspire the community for faster and more realistic generation models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15740" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.239 + ], + "angle": 0, + "content": "Many colleagues from ByteDance supported this work. We are grateful to Guanyang Deng for his efforts in data processing. We also thank Chongxi Wang and Taekmin Kim for their contributions to model deployment. Special thanks to Xiaoxiao Qin for her work in human preference evaluation. Additionally, we are thankful to Hui Wu, Fu Li, Xing Wang, Hongxiang Hao, and Chuan Li for their contributions to infrastructure." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.252, + 0.188, + 0.267 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.277, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.347, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[2] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.404, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.461, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[4] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[5] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22669-22679, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[6] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[7] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. OpenAI, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.73, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[9] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[10] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.277, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Vision and Pattern Recognition, pages 11315-11325, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[11] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[13] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[14] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[15] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[16] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[17] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[18] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.717 + ], + "angle": 0, + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[20] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[21] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15741" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[23] Lijie Fan, Tianhong Li, Siyang Qin, Yuanshen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[24] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[25] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.329 + ], + "angle": 0, + "content": "[26] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. In European Conference on Computer Vision, pages 289-305. Springer, 2025. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.373, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[28] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[29] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[30] Aren Jansen, Daniel PW Ellis, Shawn Hershey, R Channing Moore, Manoj Plakal, Ashok C Popat, and Rif A Saurous. Coincidence, categorization, and consolidation: Learning to recognize sounds with minimal supervision. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 121-125. IEEE, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[31] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[32] Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jose Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, Krishna Somandepalli, Hassan Akbari, Yair Alon, Yong Cheng, Josh Dillon, Agrim Gupta, Meera Hahn, Anja Hauth, David Hendon, Alonso Martinez, David Minnen, Mikhail Sirotenko, Kihyuk Sohn, Xuan Yang, Hartwig Adam, Ming-Hsuan Yang, Irfan Essa, Huisheng Wang, David A. Ross, Bryan Seybold, and Lu Jiang. Videopoet: A large language model for zero-shot video generation, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[33] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[34] Black Forest Labs. Flux. https://blackforestlabs.ai/announcing-black-forest-labs/,2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[35] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523-11532, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[36] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[37] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. Advances in Neural Information Processing Systems, 37:56424-56445, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[38] Xiang Li, Hao Chen, Kai Qiu, Jason Kuen, Jiumiang Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. arXiv preprint arXiv:2410.01756, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[39] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.508 + ], + "angle": 0, + "content": "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[41] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[42] Chuofan Ma, Yi Jiang, Junfeng Wu, Jihan Yang, Xin Yu, Zehuan Yuan, Bingyue Peng, and Xiaojuan Qi. Unitok: A unified tokenizer for visual generation and understanding. arXiv preprint arXiv:2502.20321, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[43] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[44] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[45] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[47] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15742" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[48] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.161, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[49] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[50] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[54] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[55] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[56] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[57] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[58] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[59] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[60] Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, et al. Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[61] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[62] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024.3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[63] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[64] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2404.02905, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[65] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[66] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[67] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[69] Junke Wang, Yi Jiang, Zehuan Yuan, Bingyue Peng, Zuxuan Wu, and Yu-Gang Jiang. Omnitokensizer: A joint image-video tokenizer for visual generation. Advances in Neural Information Processing Systems, 37:28281-28295, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[70] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[71] Mark Weber, Lijun Yu, Qihang Yu, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. Maskbit: Embedding-free image generation via bit tokens. arXiv preprint arXiv:2409.16211, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[72] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[73] Junfeng Wu, Yi Jiang, Chuofan Ma, Yuliang Liu, Hengshuang Zhao, Zehuan Yuan, Song Bai, and Xiang Bai. Liquid: Language models are scalable and unified multi-modal generators. arXiv preprint arXiv:2412.04332, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[74] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15743" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[75] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[76] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[77] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[78] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.484, + 0.473 + ], + "angle": 0, + "content": "[79] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.484, + 0.543 + ], + "angle": 0, + "content": "[80] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[81] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.484, + 0.684 + ], + "angle": 0, + "content": "[82] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.484, + 0.74 + ], + "angle": 0, + "content": "[83] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. arXiv preprint arXiv:2406.07550, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.743, + 0.484, + 0.809 + ], + "angle": 0, + "content": "[84] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.812, + 0.484, + 0.855 + ], + "angle": 0, + "content": "[85] Yue Zhao, Yuanjun Xiong, and Philipp Krahenbuhl. Image and video tokenization with binary spherical quantization. arXiv preprint arXiv:2406.07548, 2024. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.855 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15744" + } + ] +] \ No newline at end of file diff --git a/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/d1db9220-1c65-401b-9202-6e181bfed6b4_origin.pdf b/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/d1db9220-1c65-401b-9202-6e181bfed6b4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7ba5c3e480ba5f721a5cb91e444106e3691c9e53 --- /dev/null +++ b/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/d1db9220-1c65-401b-9202-6e181bfed6b4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15e7a2b71a5eea716762dfc53d7729677dbfd77f00836501010baa8bc48cff71 +size 6723478 diff --git a/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/full.md b/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..25466d932ad78aead8c92112271aae279f4a9a3c --- /dev/null +++ b/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/full.md @@ -0,0 +1,376 @@ +# Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis + +Jian Han*, Jinlai Liu*, Yi Jiang*, Bin Yan, Yuqi Zhang, Zehuan Yuan†, Bingyue Peng, Xiaobing Liu ByteDance + +{hanjian.thul23, liujinlai.licio, jiangyi.enjoy, yanbin/master}@bytedance.com {zhangyuqi.hi, yuanzehuan, bingyue.peng, will.liu}@bytedance.com + +Code: https://github.com/FoundationVision/Infinity Online demo: https://opensource.bytedance.com/gmpt/t2i/invite + +![](images/29cf08ae7f20768842fe4f79bf08cf43ea65d94f738a849f470d7531dd6cfd79.jpg) +Figure 1. High-resolution image synthesis results from our Infinity-2B model, showcasing its capabilities in precise prompt following, spatial reasoning, text rendering, and aesthetics across different styles and aspect ratios. + +![](images/a532c938a6514fd71575c693da922625a94bc97714c20574c862c7a725d7e4db.jpg) + +![](images/936ccb57f533d7860fce82f267505220b8a64aaa08ad88a046af339bb702ac82.jpg) + +![](images/99ad381f612a32ea4b9bf615affc70f31021ee0c901df18d1ecc1de90f57a986.jpg) + +![](images/b3c8e34c488ae9240536902d9e42fc3f64bb1e35efc4701f16a21324ec08da1f.jpg) + +# Abstract + +We present Infinity, a Bitwise Visual AutoRegressive Modeling capable of generating high-resolution, photorealistic images following language instruction. Infinity refactors visual autoregressive model under a bitwise token prediction framework with an infinite-vocabulary classifier and bitwise self-correction mechanism. By theoretically expanding the tokenizer vocabulary size to infinity in Transformer, our + +method significantly unleashes powerful scaling capabilities to infinity compared to vanilla VAR. Extensive experiments indicate Infinity outperforms AutoRegressive Text-to-Image models by large margins, matches or surpasses leading diffusion models. Without extra optimization, Infinity generates a $1024 \times 1024$ image in 0.8s, $2.6 \times$ faster than SD3-Medium, making it the fastest Text-to-Image model. All the code and models are available to promote further exploration of Infinity for visual generation. + +# 1. Introduction + +Visual generation[20, 27, 45, 51, 55] has recently seen rapid advancements, enabling high-quality, high-resolution images and video synthesis[7, 22]. Text-to-image generation[6, 22, 46, 48, 49, 53] is one of the most challenging tasks due to its need for complex language adherence and intricate scene creation. Currently, visual generation is primarily divided into two main approaches: Diffusion models and AutoRegressive models. + +Diffusion models[20, 22, 27, 45, 46, 55], trained to invert the forward paths of data toward random noise, effectively generate images through a continuous denoising process. AutoRegressive models[14, 21, 21, 58, 64, 73], on the other hand, harness the scalability and generalizability of language models[2, 3, 16, 28, 60, 63, 65, 66, 72] by employing a visual tokenizer[50, 67, 78] to convert images into discrete tokens and causally optimize these tokens, allowing image generation through next-token prediction or next-scale prediction. AutoRegressive models encounter significant challenges in high-resolution text-to-image synthesis[58, 70, 79]. They exhibit inferior reconstruction quality when utilizing discrete tokens as opposed to continuous tokens. Additionally, their generated visual contents are less detailed than those by diffusion models. Inefficiency and latency in visual generation, resulting from the raster scan method of next-token prediction, further exacerbate these issues. + +Recently, Visual AutoRegressive Modeling (VAR)[64] redefined autoregressive learning on images as coarse-to-fine "next-scale prediction". It demonstrates superior generalization and scaling capabilities compared to diffusion transformers while requiring fewer steps. VAR leverages the powerful scaling properties of LLMs [25, 31] and can simultaneously refine previous scale steps, benefiting from the strengths of diffusion models as well. However, the index-wise discrete tokenizer[21, 38, 42, 64, 67, 69, 83] employed in AutoRegressive or Visual AutoRegressive models faces significant quantization errors with a limited vocabulary size resulting in difficulties in reconstructing fine-grained details, especially in high-resolution images. In the generation stage, index-wise tokens suffer from fuzzy supervision leading to visual detail loss and local distortions. Moreover, train-test discrepancies from teacher-forcing training, inherent to LLMs, amplify cumulative errors in visual details. These challenges make index-wise tokens a significant bottleneck for AutoRegressive models. + +We propose a novel bitwise modeling framework that substitutes index-wise tokens with bitwise tokens throughout the process. Our bitwise modeling framework consists of three primary modules: bitwise visual tokenizer, bitwise infinite-vocabulary classifier, and bitwise self-correction. Inspired by the success and widespread adoption of binary vector quantization[80, 85], we scaled up the tokenizer vo + +![](images/0528a35add7e85d91bb366c2a15151c954a45def3e393ee5e18c0d89264b490e.jpg) +Figure 2. Left: Conventional classifier predicts $2^{d}$ indices. Right: Infinite-Vocabulary Classifier predicts $d$ bits instead, being much more efficient than index-wise classifier when $d$ is large. + +cabulary to $2^{64}$ , significantly surpassing all previous AutoRegressive model vocabularies[58, 81]. This expansion allows for reconstruction quality that far exceeds previous discrete tokenizers, achieving results comparable to continuous VAE[51], with scores improving from 0.87 to 0.33 on ImageNet-256 benchmark[19]. In Fig.2, we transform the conventional token prediction from a large integer into binary bits in a bitwise infinite-vocabulary classifier to address optimization and computation challenges, enabling the learning of massive vocabularies in Visual AutoRegressive models. Additionally, we incorporated bitwise self-correction during training by randomly flipping some bits to simulate prediction mistakes and re-quantizing the residual features, thus endowing the system with self-correcting capabilities. Our method, Infinity: Bitwise Visual AutoRegressive Modeling, maintains the scaling and speed advantages of VAR while achieving detailed reconstruction and generation quality comparable to that of continuous Diffusion models. + +Infinity outperforms all previous AutoRegressive text-to-image models by large margins and matches or surpasses state-of-the-art diffusion models including SDXL[46], PixArt-Sigma[13],DALL-E3[6] and Stable-Diffusion 3[22] in both quantitative evaluations and human preference scores. The core contributions of our work are as follows: + +1. We propose Infinity, an autoregressive model utilizing Bitwise Modeling, which significantly enhances scalability and visual detail representation capabilities of discrete generative models. This framework opens up new possibilities of 'infinity' within the discrete generation community. +2. Infinity demonstrates the potential of scaling tokenizers and transformers by achieving near-continuous tokenizer performance with its image tokenizer, surpassing diffusion models in high-quality text-to-image generation. +3. Infinity enables an autoregressive text-to-image model to achieve exceptional prompt adherence and superior + +image generation quality, while also delivering unprecedented inference speed. + +# 2. Related Work + +# 2.1. AutoRegressive Models + +AutoRegressive models, leveraging the powerful scaling capabilities of LLMs[8, 16, 47, 65, 66], use discrete image tokenizers[21, 50, 67] in conjunction with transformers to generate images based on next-token prediction. VQ-based methods [21, 35, 50, 58, 67] employ vector quantization to convert image patches into index-wise tokens and use a decoder-only transformer to predict the next token index. However, these methods are limited by the lack of scaled-up transformers and the quantization error inherent in VQ-VAE[67], preventing them from achieving performance on par with diffusion models. Parti [79], Emu3 [70], chameleon[62], Liquid[73], and VideoPoet[32] scaled up autoregressive models in text-to-image or video synthesis. Inspired by the spatial structure of visual information, Visual AutoRegressive modeling(VAR)[64] redefines the autoregressive modeling on images as next-scale prediction framework, significantly improving generation quality and sampling speed. HART[61] adopted hybrid tokenizers based on VAR. MAR[37], Fluid[23] proposed random-order models and employed continuous tokenizer rather than discrete tokenizer. + +# 2.2. Diffusion Models. + +Diffusion models have recently experienced rapid advancements in several directions. Denoising learning mechanisms [27, 44] and improvements in sampling efficiency [4, 40, 41, 55, 56] have been continuously optimized to generate high-quality images. Bit Diffusion [15] represents data with analog bits and then trains a continuous diffusion model. Latent diffusion models [51] were the first to propose modeling in the latent space rather than the pixel space. Recently, latent diffusion models [18, 22] have scaled up VAEs to enhance representations in the latent space. DiT [45] and U-ViT [5] employ more scalable transformers to model diffusion, achieving superior results. Consequently, mainstream text-to-image diffusion models [6, 12, 22] have adopted the DiT architecture due to its effectiveness. + +# 2.3. Scaling Law + +Scaling laws in autoregressive language models reveal a power-law relationship between model size, dataset size, and compute with test set cross-entropy loss [1, 25, 31]. These laws help predict larger model performance, leading to efficient resource allocation and ongoing improvements without saturation [8, 28, 65, 66, 72, 84]. This has inspired research into scaling in visual generation [7, 22, 59, 64, 82]. + +# 3. Infinity Setup + +# 3.1. Visual AutoRegressive Modeling + +Infinity incorporates a visual tokenizer and a transformer for image synthesis. During the training stage, a sample consists of a text prompt $t$ and a ground truth image $im$ . The proposed visual tokenizer first encodes the image $im$ into a feature map $f \in \mathbb{R}^{h \times w \times C}$ with stride $s$ and then quantizes the feature map $f$ into $K$ multi-scale residual maps $(r_1, r_2, \dots, r_K)$ . The resolution of $r_k$ is $h_k \times w_k$ and it grows larger gradually from $k = 1 \rightarrow K$ . Based on this sequence of residuals, we can gradually approximate the continuous feature $f$ as in Eq.1 + +$$ +f _ {k} = \sum_ {i = 1} ^ {k} u p \left(r _ {i}, (h, w)\right) \tag {1} +$$ + +up means bilinear upsampling and $f_{k}$ is the cumulative sum of the upsampled $r_{\leq k}$ . + +Subsequently, transformer learns to predict residuals $r$ of the next scale conditioned on previous predictions and the text input in an autoregressive manner. Formally, the autoregressive likelihood can be formulated as: + +$$ +p \left(r _ {1}, \dots , r _ {K}\right) = \prod_ {k = 1} ^ {K} p \left(r _ {k} \mid r _ {1}, \dots , r _ {k - 1}, \psi (t)\right), \tag {2} +$$ + +where $\psi(t)$ is the text embeddings from Flan-T5 [17] model. $(r_1, \dots, r_{k-1}, \psi(t))$ serves as the prefixed context When predicting $r_k$ . Besides, the text embeddings $\psi(t)$ further guide the prediction through a cross-attention mechanism. In particular, as shown in Fig. 3, the text embeddings $\psi(t) \in \mathbb{R}^{L \times C}$ is projected into a $\langle \mathrm{SOS} \rangle \in \mathbb{R}^{1 \times 1 \times h}$ as the input of the first scale, where $h$ is the hidden dimension of transformer. The transformer is required to predict $r_1$ based on $\langle \mathrm{SOS} \rangle$ in the first scale. In the latter $k$ -th scale, to match the spatial size of the input and the output label $R_k$ , we take the downsampled feature $\widetilde{f}_{k-1}$ from the last scale $k-1$ as the input to predict $R_k$ in parallel. + +$$ +\widetilde {f} _ {k - 1} = \operatorname {d o w n} \left(f _ {k - 1}, \left(h _ {k}, w _ {k}\right)\right), \tag {3} +$$ + +where $down$ is bilinear downsampling and the spatial size of both $\widetilde{f}_{k-1}$ and $r_k$ is $(h_k, w_k)$ . In previous index-wise [64] representations, the shape of $r_k$ is $(h_k, w_k, V_d)$ . $V_d$ is the vocabulary size of the visual tokenizer. For binary quantization [80, 85] with code embedding dimension $d$ , $V_d = 2^d$ . + +The transformer consists of a stack of repeated blocks, where each block includes RoPE2D [26], Self-Attention, Cross Attention, and FFN layers. The text embeddings $\psi(t)$ provide effective guidance for image synthesis in each cross-attention layer. During the training stage, we exploit a block-wise causal attention mask to ensure that the + +![](images/d2b4972cb2f63c6654518ef6ea4f2df2d10093ee717cf02fbba77d7f224d51b3.jpg) +Figure 3. Framework of Infinity. Infinity introduces bitwise modeling. When predicting $r_k$ , the sequence $(r_1, r_2, \dots, r_{k-1})$ serves as the prefixed context and the text embeddings $\psi(t)$ guides the prediction through a cross attention mechanism. + +![](images/925163793b1f89868a0e859d1d3180ee3d889ea4f0402927d32ecfcd68510870.jpg) + +transformer can only attend to its prefixed context, i.e., $(\langle \mathrm{SOS}\rangle ,\widetilde{f}_1,\dots ,\widetilde{f}_{k - 1})$ , when predicting $f_{k}$ . During the inference stage, we perform KV-Caching to speed up inference and there's no need for masking. + +# 3.2. Infinite Tokenizer + +Increasing the vocabulary size has significant potential for improving reconstruction and generation quality. However, directly enlarging the vocabulary in existing tokenizers[58, 64, 71, 81] leads to a substantial increase in memory consumption and computational burden. To address these challenges and fully exploit the potential of discrete tokenizers, this paper proposes a new bitwise multi-scale residual quantizer, which significantly reduces memory usage, enabling the training of extremely large vocabulary, e.g. $2^{64}$ . + +Bitwise Multi-scale Residual Quantizer. We replace the original vector quantizer of VAR [64] with a dimension-independent bitwise quantizer. In this paper, we consider two candidates, LFQ [81] and BSQ[85]. Given $K$ scales in the multi-scale quantizer, on the $k$ -th scale, the input continuous residual vector $z_{k} \in \mathbb{R}^{d}$ are quantized into binary output $q_{k}$ as shown below. + +$$ +q _ {k} = \left\{ \begin{array}{l l} \operatorname {s i g n} \left(z _ {k}\right) & \text {l o o k u p - f r e e q u a n t i z e r} \\ \frac {1}{\sqrt {d}} \operatorname {s i g n} \left(\frac {z _ {k}}{\left| z _ {k} \right|}\right) & \text {b i n a r y s p h e r i c a l q u a n t i z e r} \end{array} \right. \tag {4} +$$ + +To encourage codebook utilization, an entropy penalty $\mathcal{L}_{entropy} = \mathbb{E}[H(q(z))] - H[\mathbb{E}(q(z))]$ [30] is adopted. To obtain the distribution of $q(z)$ , we need to compute the similarities between the input $z$ and the whole codebook when + +using LFQ. However, this leads to unaffordable space and time complexity of $O(2^d)$ . When the codebook dimension $d$ becomes large, e.g. 20, an out-of-memory (OOM) issue is faced as shown in Tab. 3. By contrast, since both input and output in BSQ are unit vectors, BSQ[85] proposes an approximation formula for the entropy penalty, reducing the computational complexity to $O(d)$ . As shown in Tab 3, there is no obvious increase in memory consumption for BSQ even when codebook size is $2^{64}$ . Unless otherwise stated, we adopt BSQ by default. + +# 3.3. Infinite-Vocabulary Classifier + +The visual tokenizer obtains discrete labels by quantizing residual features. Consequently, the transformer predicts next-scale residual features' labels $\mathbf{y}_k \in [0, V_d)^{h_k \times w_k}$ and optimizes the target through the cross-entropy loss. Previous works [64, 80] directly predict these index labels using a classifier of $V_d$ classes. However, it suffers from two drawbacks, huge computational costs and fuzzy supervision. + +As illustrated in Section 3.2, we exploit a bitwise VQ-VAE as the visual tokenizer, where the vocabulary size $V_{d}$ is extremely large. For example, if $V_{d} = 2^{32}$ and $h = 2048$ , a conventional classifier would require a weight matrix $W \in \mathbb{R}^{h \times V_d}$ of 8.8 trillion parameters, which exceeds the limits of current computational resources. + +Moreover, VQ-VAE exploits the sign function during quantization as in Eq.4. After that, the positive elements are multiplied with the corresponding base and summed to get the index label $\mathbf{y}_k(m,n)$ as in Eq.5, where $m\in [0,h_k)$ and $n\in [0,w_k)$ . + +$$ +\mathbf {y} _ {k} (m, n) = \sum_ {p = 0} ^ {d - 1} \mathbb {1} _ {r _ {k} (m, n, p) > 0} \cdot 2 ^ {p} \tag {5} +$$ + +Owing to the merits of the quantization method, slight perturbations to those near-zero features cause a significant change in the label. As a result, the conventional index-wise classifier [10, 64, 81] is difficult to optimize. + +To address these problems in computation and optimization, we propose an Infinite-Vocabulary Classifier (IVC). In particular, instead of using a conventional classifier with $V_{d}$ classes, we use $d$ binary classifiers in parallel to predict if the next-scale residual $r_k(m,n,p)$ is positive or negative, where $d = \log_2(V_d)$ . The proposed Infinite-Vocabulary Classifier is much more efficient in memory and parameters compared to the conventional classifier. When $V_{d} = 2^{16}$ and $h = 2048$ , it saves $99.95\%$ parameters and GPU memory. Besides, when there exist near-zero values that confuse the model in some dimensions, the supervision in other dimensions is still clear. Therefore, compared with conventional index-wise classifiers, the proposed Infinite-Vocabulary Classifier is easier to optimize. + +# 3.4. Bitwise Self-Correction + +Weakness of teacher-forcing training. VAR [64] inherits the teacher-forcing training from LLMs. However, next-scale prediction in vision is quite different from next-token prediction in language. Specifically, we cannot decode the complete image until residuals $r_k$ from all scales are obtained. We find that the teacher-forcing training brings about severe train-test discrepancies for visual generation. In particular, the teacher-forcing training makes the transformer only refine features in each scale without the ability to recognize and correct mistakes. Mistakes made in former scales will be propagated and amplified in latter scales, finally messing up generated images (left images in Fig.10). + +In this work, we propose the bitwise self-correction to address this issue. In particular, we obtain $r_k^{flip}$ via randomly flipping the bits in $r_k$ with a probability from $0\%$ to $30\%$ , imitating the errors in the prediction of the $k$ -th scale. + +Here comes the key component of bitwise self-correction. $r_k^{flip}$ contains errors while $r_k$ doesn't. After replacing $r_k$ with $r_k^{flip}$ as predictions on the $k$ -th scale, we need to recompute the transformer input $\widetilde{f}_k$ . Besides, requantization is performed to get a new target $r_{k+1}$ as in Eq.6. Refer to Fig.3 (right) for a simplified illustration. + +$$ +\left\{ \begin{array}{l} f _ {k} ^ {\text {f l i p}} = \sum_ {i = 1} ^ {k} u p \left(r _ {i} ^ {\text {f l i p}}, (h, w)\right) \\ \widetilde {f} _ {k} = \operatorname {d o w n} \left(f _ {k} ^ {\text {f l i p}}, \left(h _ {k + 1}, w _ {k + 1}\right)\right) \\ r _ {k + 1} = \operatorname {q u a n t} \left(\operatorname {d o w n} \left(f - f _ {k} ^ {\text {f l i p}}, \left(h _ {k + 1}, w _ {k + 1}\right)\right)\right) \end{array} \right. \tag {6} +$$ + +Each scale undergoes the same process of random-flipping and re-quantization. Here the transformer takes partially randomly flipped features as inputs, taking the prediction errors into consideration. The re-quantized bit labels enable the transformer to auto-correct errors made in former predictions. In such a way, we address the train-test discrepancy caused by teacher-forcing training and empower Infinity to have the self-correction ability. + +# 3.5. Dynamic Resolution Generation + +Infinity can generate photo-realistic images with various aspect ratios, which is significantly different from VAR [64] that can only generate square images. The main obstacles of generating various aspect ratio images lie in two folds. The first is to define the height $h_k$ and width $w_k$ of $r_k$ based on varying aspect ratios. In the supplementary material, we pre-define a list of scales, also called scale schedule, as $\{(h_1^r, w_1^r), \dots, (h_K^r, w_K^r)\}$ for each aspect ratio. We ensure that the aspect ratio of each tuple $(h_k^r, w_k^r)$ is approximately equal to $r$ , especially in the latter prediction scales. Additionally, for different aspect ratios at the same scale $k$ , we keep the area of $h_k^r \times w_k^r$ to be roughly equal, ensuring that the training sequence lengths are roughly the same. + +Secondly, we need to carefully design a resolution-aware positional encoding method to handle features of various scales and aspect ratios. This issue poses a significant challenge, as the existing solutions [26, 43, 57, 64, 68] exhibit substantial limitations under such conditions. In this paper, we apply RoPE2d [26] on features of each scale to preserve the intrinsic 2D structure of images. Additionally, we exploit learnable scale embeddings to avoid confusion between features of different scales. Compared to learnable APE element-wise applied on features, learnable embeddings applied on scales bring fewer parameters, can adapt to varying sequence lengths, and are easier to optimize. + +# 4. Experiment + +# 4.1. Dataset + +Data Curation. We curated a large-scale dataset from open-source academic data and high-quality internally collected data. The pre-training dataset is constructed by collecting and cleaning open-source academic datasets such as LAION [54], COYO [9], OpenImages [33]. We exploit an OCR model and a watermark detection model to filter undesired images with too many texts or watermarks. Additionally, we employ Aesthetic-V2 to filter out images with low aesthetic scores. + +# 4.2. Training + +Progressive Training. Infinity redefines text-to-image as a coarse-to-fine, next-scale prediction task. In line with its architecture, we propose to train Infinity in a progressive + +strategy. Specifically, we first train Infinity of 2B parameters on the pre-training dataset of over 100 million images with 256 resolution for 150k iterations using a batch size of 4096 and a learning rate of 6e-5. Then we switch to 512 resolution and train 110k iterations using the same hyperparameters. Next, we fine-tune Infinity at 1024 resolution with a smaller, high-quality dataset of 60 million images. In this stage, we train Infinity for 60k iterations using a batch size of 2048 and a learning rate of 2e-5. All training stages use images of varying aspect ratios. + +# 4.3. Text-to-Image Generation + +# 4.3.1 Qualitative Results + +Overall Results. Fig.1 presents generated images from our Infinity-2B model, showcasing Infinity's strong capabilities in generating high-fidelity images from various categories following user prompts. Qualitative comparison results among Infinity and other top-tier models can be found in the supplementary materials. + +Prompt-Following. Fig.11 in the appendix presents three examples demonstrating the superior prompt-following ability of Infinity. As highlighted in red, Infinity consistently adheres to user prompts, whether they are short or extremely long. We attribute these improvements to scaling autoregressive modeling and visual tokenizer's vocabulary. + +Text Rendering. As illustrated in Fig.12 in the appendix, Infinity can render text according to user prompts across diverse categories. Despite diverse backgrounds and subjects, Infinity accurately renders corresponding texts according to user requirements, such as fonts, styles, colors, and more. + +# 4.3.2 Quantitative Results + +Benchmark. As in Tab 1, on GenEval[24], our model with a re-writer achieves the best overall score of 0.73. Besides, Infinity also reaches the highest position reasoning score of 0.49. On DPG [29]. Our model reaches an overall score of 83.5, surpassing SDXL [46], Playground v2.5 [36], and DALLE 3 [6]. What's more, Infinity achieves the best relation score of 90.76 among all open-source T2I models, demonstrating its stronger ability to generate spatially consistent images based on user prompts. + +Human Preference Evaluation. We conduct human preference evaluation in both human studies and benchmarks. As in Fig.5, the generation results of Infinity are more frequently selected by humans in terms of overall quality, prompt following, and visual aesthetics in contrast to other open-sourced T2I models. Please refer to the supplementary material for more details. Tab.2 lists the results of three human preference benchmarks, i.e., ImageReward [77], HPSv2 [74], and VQAScore[39]. Infinity reaches the highest ImageReward and HPSv2.1, indicating our method could generate images that are more appealing to humans. + +
Methods# ParamsGenEval↑DPG↑
Two Obj.PositionColor Attri.OverallGlobalRelationOverall
Diffusion Models
LDM [52]1.4B0.290.020.050.37---
SDv1.5 [52]0.9B0.380.040.060.4374.6373.4963.2
PixArt-alpha [11]0.6B0.500.080.070.4874.9782.5771.1
SDv2.1 [52]0.9B0.510.070.170.5077.6780.7268.1
DALL-E 2 [49]6.5B0.660.100.190.52---
DALL-E 3 [6]----\( 0.67^† \)90.9790.5883.5
SDXL [46]2.6B0.740.150.230.5583.2786.7674.7
PixArt-Sigma [13]0.6B0.620.140.270.5586.8986.5980.5
SD3 (d=24) [22]2B0.740.340.360.62--84.1
SD3 (d=38) [22]8B0.890.340.470.71---
SANA-1.0 [75]1.6B---0.66--84.8
FLUX-dev [34]12B---0.67--84.0
FLUX-schnell [34]12B---0.71--84.8
AutoRegressive Models
LlamaGen [58]0.8B0.340.070.040.3265.2
Chameleon [62]7B---0.39---
HART [61]0.7B---0.56--80.9
Show-o [76]1.3B0.800.310.500.68--67.5
Emu3 [70]8.5B\( 0.81^† \)\( 0.49^† \)\( 0.45^† \)\( 0.66^† \)--81.6
Infinity2B\( 0.85^† \)\( 0.49^† \)\( 0.57^† \)\( 0.73^† \)93.1190.7683.5
+ +Table 1. Evaluation on the GenEval [24] and DPG [29] benchmark. A blue background indicates methods with larger model sizes than ours. $\dagger$ result is with prompt rewriting. + +![](images/5dcdd0ea2045b3c7e7cb134d85da92a07dadd4d32390d37f4cb0be2509ae44c2.jpg) +Figure 4. Generation Latency (s) on $1024 \times 1024$ . Infinity shows a significant advantage in generation latency compared to diffusion models. + +![](images/e9d1a01261deb23886cb2b5dcb1fc85737f30904106a3e733214f0c02d0c0702.jpg) +Figure 5. Human Preference Evaluation. We ask users to select the better one in a side-by-side comparison in terms of Overall Quality, Prompt Following, and Visual Aesthetics. Infinity is more preferred by humans compared to other methods. + +Inference Latency. As shown in Fig.4, Infinity demonstrates a significant advantage in generation speed compared to diffusion models across different model sizes. Without extra optimization, Infinity generates a $1024 \times 1024$ image in 0.8s, $2.6 \times$ faster than SD3-Medium. The speed advantage becomes more substantial as the model size increases. Infinity achieves $7 \times$ faster inference latency compared to SD3.5 (8B) [22] and Flux-Dev (12B). Besides, Infinity still achieves $3.7 \times$ speedup than Flux-Dev even with more parameters (20B v.s. 12B). + +
Methods# ParamsImageReward↑HPSv2.1↑VQAScore↑
RankScoreRankScoreRankTime
SD-XL [46]2.6B40.600430.0630.832
SD3-Medium [22]2B30.871330.9110.906
PixArt Sigma [13]0.6B20.872231.47--
Infinity2B10.953132.0520.894
+ +# 4.4. Scaling Visual Tokenizer's Vocabulary + +Scaling up the Vocabulary Benefits Reconstruction. Restricted by the vocabulary size, discrete VQ-VAEs have always lagged behind continuous ones, hindering the performance of AR-based T2I models. In this work, we successfully train a discrete VQ-VAE matching its continuous counterparts by scaling up the vocabulary size. As in Tab. 4, we observe consistent rFID improvements as scaling up the vocabulary size from $2^{16}$ to $2^{64}$ . It's noteworthy that our discrete tokenizer achieves a rFID of 0.61 on ImageNet-256 when $V_{d} = 2^{32}$ , outperforming SD [52]'s contiguous VAE. + +Infinite Vocabulary Classifier Benefits Generation. We compare predicting bit labels with IVC to predicting index labels using a conventional classifier under the vocabulary size of $2^{16}$ since a larger vocabulary causes OOM for the conventional classifier. We utilize the reconstruction loss on $\mathbf{R}_k$ , the Fréchet Inception Distance (FID) on the validation dataset and ImageReward for comprehensive evaluation. As shown in Tab.5, IVC achieves lower reconstruction loss and FID, suggesting IVC has better fitting capabilities. Beyond the quantitative results, training Infinity with IVC yields images with richer details as in Fig.6, which is consistent with a higher ImageReward. + +Scaling Up the Vocabulary Benefits Generation. We then scale up the vocabulary size to $2^{32}$ during training the T2I model, which exceeds the range of the Int32 data type and can be considered infinitely large. In Fig.7, we illustrate the effect of scaling up the vocabulary from $2^{16}$ to $2^{32}$ for image generation. For small models (125M and 361M), the vocabulary size of $2^{16}$ converges faster and achieves better results. However, as we scaled up the transformer to 2.2B, the vocabulary size of $2^{32}$ beats $2^{16}$ after 40K iterations. Therefore, it's worthwhile to scale up the vocabulary along with scaling up the transformer. As illustrated in Tab.1,2, with infinite vocabulary and IVC, Infinity achieves superior performance among various benchmarks, redefining performance limits in autoregressive visual synthesis. + +# 4.5. Scaling Visual AutoRegressive Modeling + +In Fig.8, we depict the validation loss against the total training iterations and computational FLOPs for various model sizes of Infinity. We consistently notice a reduction in validation loss with an increase in training steps and computa + +Table 2. Human Preference Metrics. Infinity reaches the highest ImageReward and HPSv2.1 score, also with comparable VQAS-score to SD3. + +
Quantizerd=16d=18d=20d=32d=64
LFQ37.653.7OOMOOMOOM
BSQ32.432.432.432.432.4
+ +Table 3. Comparison of memory consumption (GB) between different quantizers during training. As codebook dimension $d$ increases, MSR-BSQ shows significant advantages over MSR-LFQ, enabling nearly infinite vocabulary size of $2^{64}$ . + +
VAE (stride=16)TYPEIN-256 rFID↓IN-512 rFID↓
Vd=216Discrete1.220.31
Vd=224Discrete0.750.30
Vd=232Discrete0.610.23
Vd=264Discrete0.330.15
SD VAE [52]Continual0.87N/A
+ +Table 4. By scaling up visual tokenizer's vocabulary, discrete tokenizer surpasses continuous VAE of SD [51] on ImageNet-rFID. + +
Classifier# ParamsvRAMRecons. Loss↓FID↓ImageReward↑
Convention124M2GB0.1845.950.79
IVC0.65M10MB0.1805.320.91
+ +Table 5. In contrast to the conventional classifier, IVC saves $99.95\%$ parameters (vRAM) and reaches better performance. + +![](images/3c2668e9c39b7ed139637e89a375b29aa78f2c0f7ea3eeba4912ba9f690c84be.jpg) +Figure 6. Impact of Infinite-Vocabulary Classifier. Predicting bitwise labels with the Infinite-Vocabulary Classifier (Right) generates images with richer details compared to predicting index-wise labels using a conventional classifier (Left). + +tional FLOPs. Nevertheless, the advantages gained from training smaller models for extended periods lag behind those obtained from training larger models for shorter durations. This trend aligns with findings in language models, emphasizing the promising outlook for increasing model sizes with appropriate training. + +In Fig.8, we plot GenEval, ImageReward, and HPSv2 scores against validation loss for different model sizes ranging from 125M to 4.7B. We observe a strong correlation between validation loss and evaluation metrics. To further quantify their correlation, we calculate the Pearson correlation coefficients through linear regression. The correlation coefficients for GenEval, ImageReward, and HPSv2 are -0.983, -0.981, and -0.979, respectively. These results demonstrate a nearly linear correlation between validation loss and the evaluation metrics when scaling up model sizes from 125M to 4.7B. This promising phenomenon encourages us to scale up Infinity to achieve better performance. + +![](images/3657c0a75737e60d587cc98522964eb09c2a7e9ed8652ffeea7ed65a9f31878c.jpg) + +![](images/7236248afff6ec234490ceab5a9e6ea49dbb2683706eb51d3da7a561b4a0d382.jpg) + +![](images/5ed7059f5d1331ad41e4eace9131cbe2dd3e25befb2d8606e4fdd35d93f27c45.jpg) + +![](images/c12ed33fc6b5c76d1cc5a4ef9ac3717a4e02a00caa0b04c10cc9f5d02f7e5dd5.jpg) + +![](images/cbef24ae022b36e2f60c9bf522df068337a9a8d30c20ef1f83bf49070c0e348c.jpg) + +![](images/d87333dfb3b6400225739574a4f6d3fa2aa3743eb42f8b27bfc255544d652407.jpg) + +![](images/3f8cefe67af63d230132a9b972eb2b46ef9b15a61752a255d4756ac9e3a7b8f4.jpg) +Figure 7. Effects of Scaling Up the Vocabulary. We analyze the impact of scaling the vocabulary size under consistent training hyperparameters. Vocabulary size $V_{d} = 2^{16}$ converges faster and achieves better results for small models (125M and 361M parameters). As we scale up the model size to 2.2B, Infinity with a vocabulary size $V_{d} = 2^{32}$ beats $V_{d} = 2^{16}$ . Experiment with 5M high-quality data under $256 \times 256$ resolution. + +![](images/5d372e3b5e9ea100b49fd408da60d9cc641722859e5bcfdee42002d37f1134a8.jpg) + +![](images/d42ecf68d027aaef132a35fa5ee0537d243001455acc1b7f25a0c8bf60f0f79e.jpg) + +![](images/2f0695e4cbdaa9b4cb848ef48843156eeb498849b8b1f494965c0d173bbb9bca.jpg) + +![](images/f21d9d2d0e161034b4d7db26ce485706acc3bf9dede5ed150b70a85aca8ed623.jpg) + +![](images/c73df1284253e488e649031b73aae44c0a8121821655c6a78b7266a0267e4666.jpg) + +![](images/403493a93b7372ceffcf27358e714c138cc4fc5616467ca7ecae05f9c4445e23.jpg) +Figure 8. Scaling Visual AutoRegressive Modeling effects. We analyze the impact of scaling model size under consistent training hyperparameters (Experiment with 10M pre-training data under 256 resolution). Validation loss smoothly decreases as a function of the model size and training iters. Besides, validation loss is a strong predictor of overall performance. There is a strong correlation between validation loss and holistic image evaluation metrics. + +![](images/754992fc02106e037ab433342be1b7d57a48f5c71ac034e22e5c90a9c2013047.jpg) + +![](images/e0606cbb376018f3b27cf454bf6e228c19258124208168dcc13c080efd011cac.jpg) + +Visualization of scaling effect. To delve deeper into the scaling effect of Infinity, we compare a set of generated $256 \times 256$ images of three model sizes (125M, 940M, 4.7B) across three distinct training schedules (10K, 40K, 90K iterations) as illustrated in Fig.9. The semantics and visual quality of generated images improve steadily when scaling up model size and training compute, which is consistent with the scaling behaviors of Infinity. + +# 4.6. Bitwise Self-Correction + +In Tab.6 and Fig.10, we list the evaluation metrics and present images generated by models trained using teacher- + +![](images/dd4ed85f51c920bb7c7daaeb3694c863b3ab92b55b930420581f6f8fbf02b5f9.jpg) +Figure 9. Semantics and visual quality improve consistently with scaling up model size and training compute. + +![](images/51edd01475b6da891194f4a3a1fce98d77ef4eaedf934b440d38a2cd75aa3cbc.jpg) +Figure 10. Impact of Self-Correction. Teacher-forcing training introduces great train-test discrepancy which degrades performance during inference (left). Bitwise Self-Correction autocorrects mistakes and thus generates better results (right). + +
MethodFID↓ImageReward↑HPSv2.1↑
Baseline9.760.5229.53
Baseline + Random Flip9.690.5229.20
Baseline + Bitwise Self-Correction3.470.7630.71
+ +Table 6. Bitwise Self-Correction makes significant improvements. Experiment with 5M high quality data and $512 \times 512$ resolution. FID is measured on the validation set with 40K images. + +forcing and bitwise self-correction methods. Substantial advantages are observed after applying bitwise self-correction, primarily driven by the self-correction mechanism rather than applying flipping. Simply random flipping $\mathbf{R}_k$ doesn't bring improvements. Self-correction considers errors and performs re-quantization to correct them. We emphasize that self-correction is essential for AR-based T2I models since it empowers models to correct errors automatically, significantly mitigating the train-test discrepancy. + +# 5. Conclusion + +We introduce Infinity, a bitwise visual autoregressive model to perform Text-to-Image generation. Infinity is a pioneering framework for bitwise token modeling with the IVC and self-correction innovation. Extensive results demonstrate Infinity significantly raised the upper limit for Autoregressive Text-To-Image models, matching or surpassing leading diffusion models. We believe Infinity will significantly promote the development of autoregressive visual modeling and inspire the community for faster and more realistic generation models. + +# 6. Acknowledgment + +Many colleagues from ByteDance supported this work. We are grateful to Guanyang Deng for his efforts in data processing. We also thank Chongxi Wang and Taekmin Kim for their contributions to model deployment. Special thanks to Xiaoxiao Qin for her work in human preference evaluation. Additionally, we are thankful to Hui Wu, Fu Li, Xing Wang, Hongxiang Hao, and Chuan Li for their contributions to infrastructure. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[2] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2 +[3] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2 +[4] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503, 2022. 3 +[5] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22669-22679, 2023. 3 +[6] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 2, 3, 6 +[7] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. OpenAI, 2024. 2, 3 +[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3 +[9] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.5 +[10] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer + +Vision and Pattern Recognition, pages 11315-11325, 2022. 5 +[11] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 6 +[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 3 +[13] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2, 6, 7 +[14] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2 +[15] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202, 2022.3 +[16] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 2, 3 +[17] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 3 +[18] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 3 +[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2 +[20] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[21] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 3 +[22] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 6, 7 + +[23] Lijie Fan, Tianhong Li, Siyang Qin, Yuanshen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens, 2024. 3 +[24] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36, 2024. 6 +[25] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 2, 3 +[26] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. In European Conference on Computer Vision, pages 289-305. Springer, 2025. 3, 5 +[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3 +[28] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022. 2, 3 +[29] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 6 +[30] Aren Jansen, Daniel PW Ellis, Shawn Hershey, R Channing Moore, Manoj Plakal, Ashok C Popat, and Rif A Saurous. Coincidence, categorization, and consolidation: Learning to recognize sounds with minimal supervision. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 121-125. IEEE, 2020. 4 +[31] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2, 3 +[32] Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jose Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, Krishna Somandepalli, Hassan Akbari, Yair Alon, Yong Cheng, Josh Dillon, Agrim Gupta, Meera Hahn, Anja Hauth, David Hendon, Alonso Martinez, David Minnen, Mikhail Sirotenko, Kihyuk Sohn, Xuan Yang, Hartwig Adam, Ming-Hsuan Yang, Irfan Essa, Huisheng Wang, David A. Ross, Bryan Seybold, and Lu Jiang. Videopoet: A large language model for zero-shot video generation, 2024. 3 +[33] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 5 + +[34] Black Forest Labs. Flux. https://blackforestlabs.ai/announcing-black-forest-labs/,2024.6 +[35] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523-11532, 2022. 3 +[36] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 6 +[37] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. Advances in Neural Information Processing Systems, 37:56424-56445, 2024. 3 +[38] Xiang Li, Hao Chen, Kai Qiu, Jason Kuen, Jiumiang Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. arXiv preprint arXiv:2410.01756, 2024. 2 +[39] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2024. 6 +[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3 +[41] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 3 +[42] Chuofan Ma, Yi Jiang, Junfeng Wu, Jihan Yang, Xin Yu, Zehuan Yuan, Bingyue Peng, and Xiaojuan Qi. Unitok: A unified tokenizer for visual generation and understanding. arXiv preprint arXiv:2502.20321, 2025. 2 +[43] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 5 +[44] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 3 +[45] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 3 +[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 6, 7 +[47] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3 + +[48] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2 +[49] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2, 6 +[50] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2, 3 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 7 +[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 6, 7 +[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2 +[54] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5 +[55] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3 +[56] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3 +[57] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5 +[58] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2, 3, 4, 6 +[59] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 3 +[60] Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, et al. Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137, 2021. 2 +[61] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and + +Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 3, 6 +[62] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024.3, 6 +[63] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2 +[64] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2404.02905, 2024. 2, 3, 4, 5 +[65] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 3 +[66] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3 +[67] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3 +[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5 +[69] Junke Wang, Yi Jiang, Zehuan Yuan, Bingyue Peng, Zuxuan Wu, and Yu-Gang Jiang. Omnitokensizer: A joint image-video tokenizer for visual generation. Advances in Neural Information Processing Systems, 37:28281-28295, 2024. 2 +[70] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 6 +[71] Mark Weber, Lijun Yu, Qihang Yu, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. Maskbit: Embedding-free image generation via bit tokens. arXiv preprint arXiv:2409.16211, 2024. 4 +[72] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2, 3 +[73] Junfeng Wu, Yi Jiang, Chuofan Ma, Yuliang Liu, Hengshuang Zhao, Zehuan Yuan, Song Bai, and Xiang Bai. Liquid: Language models are scalable and unified multi-modal generators. arXiv preprint arXiv:2412.04332, 2024. 2, 3 +[74] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score + +v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6 +[75] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 6 +[76] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 6 +[77] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 6 +[78] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2 +[79] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2, 3 +[80] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 2, 3, 4 +[81] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 2, 4, 5 +[82] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 3 +[83] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. arXiv preprint arXiv:2406.07550, 2024. 2 +[84] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.3 +[85] Yue Zhao, Yuanjun Xiong, and Philipp Krahenbuhl. Image and video tokenization with binary spherical quantization. arXiv preprint arXiv:2406.07548, 2024. 2, 3, 4 \ No newline at end of file diff --git a/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/images.zip b/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1e216ff3ba6ccad383580ddf34fc4bd2716fcccd --- /dev/null +++ b/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5239dbeb32668797e3bf9f22047d566ecba3414fe524b0f694c0bfb07f86df3 +size 685506 diff --git a/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/layout.json b/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8901ec1d4d46c26969376ff8183e1487ebb581ff --- /dev/null +++ b/2025/Infinity_ Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis/layout.json @@ -0,0 +1,11242 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 470, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 470, + 140 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 470, + 140 + ], + "type": "text", + "content": "Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 167, + 161, + 444, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 161, + 444, + 204 + ], + "spans": [ + { + "bbox": [ + 167, + 161, + 444, + 204 + ], + "type": "text", + "content": "Jian Han*, Jinlai Liu*, Yi Jiang*, Bin Yan, Yuqi Zhang, Zehuan Yuan†, Bingyue Peng, Xiaobing Liu ByteDance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 211, + 506, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 211, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 506, + 238 + ], + "type": "text", + "content": "{hanjian.thul23, liujinlai.licio, jiangyi.enjoy, yanbin/master}@bytedance.com {zhangyuqi.hi, yuanzehuan, bingyue.peng, will.liu}@bytedance.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 243, + 462, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 243, + 462, + 271 + ], + "spans": [ + { + "bbox": [ + 147, + 243, + 462, + 271 + ], + "type": "text", + "content": "Code: https://github.com/FoundationVision/Infinity Online demo: https://opensource.bytedance.com/gmpt/t2i/invite" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 277, + 156, + 555 + ], + "blocks": [ + { + "bbox": [ + 57, + 277, + 156, + 555 + ], + "lines": [ + { + "bbox": [ + 57, + 277, + 156, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 277, + 156, + 555 + ], + "type": "image", + "image_path": "29cf08ae7f20768842fe4f79bf08cf43ea65d94f738a849f470d7531dd6cfd79.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 556, + 555, + 578 + ], + "lines": [ + { + "bbox": [ + 55, + 556, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 555, + 578 + ], + "type": "text", + "content": "Figure 1. High-resolution image synthesis results from our Infinity-2B model, showcasing its capabilities in precise prompt following, spatial reasoning, text rendering, and aesthetics across different styles and aspect ratios." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 155, + 277, + 236, + 472 + ], + "blocks": [ + { + "bbox": [ + 155, + 277, + 236, + 472 + ], + "lines": [ + { + "bbox": [ + 155, + 277, + 236, + 472 + ], + "spans": [ + { + "bbox": [ + 155, + 277, + 236, + 472 + ], + "type": "image", + "image_path": "a532c938a6514fd71575c693da922625a94bc97714c20574c862c7a725d7e4db.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 236, + 277, + 361, + 472 + ], + "blocks": [ + { + "bbox": [ + 236, + 277, + 361, + 472 + ], + "lines": [ + { + "bbox": [ + 236, + 277, + 361, + 472 + ], + "spans": [ + { + "bbox": [ + 236, + 277, + 361, + 472 + ], + "type": "image", + "image_path": "936ccb57f533d7860fce82f267505220b8a64aaa08ad88a046af339bb702ac82.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 361, + 277, + 451, + 387 + ], + "blocks": [ + { + "bbox": [ + 361, + 277, + 451, + 387 + ], + "lines": [ + { + "bbox": [ + 361, + 277, + 451, + 387 + ], + "spans": [ + { + "bbox": [ + 361, + 277, + 451, + 387 + ], + "type": "image", + "image_path": "99ad381f612a32ea4b9bf615affc70f31021ee0c901df18d1ecc1de90f57a986.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 451, + 277, + 553, + 400 + ], + "blocks": [ + { + "bbox": [ + 451, + 277, + 553, + 400 + ], + "lines": [ + { + "bbox": [ + 451, + 277, + 553, + 400 + ], + "spans": [ + { + "bbox": [ + 451, + 277, + 553, + 400 + ], + "type": "image", + "image_path": "b3c8e34c488ae9240536902d9e42fc3f64bb1e35efc4701f16a21324ec08da1f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 152, + 586, + 200, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 586, + 200, + 598 + ], + "spans": [ + { + "bbox": [ + 152, + 586, + 200, + 598 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 611, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 611, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 611, + 295, + 696 + ], + "type": "text", + "content": "We present Infinity, a Bitwise Visual AutoRegressive Modeling capable of generating high-resolution, photorealistic images following language instruction. Infinity refactors visual autoregressive model under a bitwise token prediction framework with an infinite-vocabulary classifier and bitwise self-correction mechanism. By theoretically expanding the tokenizer vocabulary size to infinity in Transformer, our" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 587, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 555, + 696 + ], + "type": "text", + "content": "method significantly unleashes powerful scaling capabilities to infinity compared to vanilla VAR. Extensive experiments indicate Infinity outperforms AutoRegressive Text-to-Image models by large margins, matches or surpasses leading diffusion models. Without extra optimization, Infinity generates a " + }, + { + "bbox": [ + 313, + 587, + 555, + 696 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 587, + 555, + 696 + ], + "type": "text", + "content": " image in 0.8s, " + }, + { + "bbox": [ + 313, + 587, + 555, + 696 + ], + "type": "inline_equation", + "content": "2.6 \\times" + }, + { + "bbox": [ + 313, + 587, + 555, + 696 + ], + "type": "text", + "content": " faster than SD3-Medium, making it the fastest Text-to-Image model. All the code and models are available to promote further exploration of Infinity for visual generation." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "type": "text", + "content": "*Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15733" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 186 + ], + "type": "text", + "content": "Visual generation[20, 27, 45, 51, 55] has recently seen rapid advancements, enabling high-quality, high-resolution images and video synthesis[7, 22]. Text-to-image generation[6, 22, 46, 48, 49, 53] is one of the most challenging tasks due to its need for complex language adherence and intricate scene creation. Currently, visual generation is primarily divided into two main approaches: Diffusion models and AutoRegressive models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 186, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 402 + ], + "type": "text", + "content": "Diffusion models[20, 22, 27, 45, 46, 55], trained to invert the forward paths of data toward random noise, effectively generate images through a continuous denoising process. AutoRegressive models[14, 21, 21, 58, 64, 73], on the other hand, harness the scalability and generalizability of language models[2, 3, 16, 28, 60, 63, 65, 66, 72] by employing a visual tokenizer[50, 67, 78] to convert images into discrete tokens and causally optimize these tokens, allowing image generation through next-token prediction or next-scale prediction. AutoRegressive models encounter significant challenges in high-resolution text-to-image synthesis[58, 70, 79]. They exhibit inferior reconstruction quality when utilizing discrete tokens as opposed to continuous tokens. Additionally, their generated visual contents are less detailed than those by diffusion models. Inefficiency and latency in visual generation, resulting from the raster scan method of next-token prediction, further exacerbate these issues." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 402, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 629 + ], + "type": "text", + "content": "Recently, Visual AutoRegressive Modeling (VAR)[64] redefined autoregressive learning on images as coarse-to-fine \"next-scale prediction\". It demonstrates superior generalization and scaling capabilities compared to diffusion transformers while requiring fewer steps. VAR leverages the powerful scaling properties of LLMs [25, 31] and can simultaneously refine previous scale steps, benefiting from the strengths of diffusion models as well. However, the index-wise discrete tokenizer[21, 38, 42, 64, 67, 69, 83] employed in AutoRegressive or Visual AutoRegressive models faces significant quantization errors with a limited vocabulary size resulting in difficulties in reconstructing fine-grained details, especially in high-resolution images. In the generation stage, index-wise tokens suffer from fuzzy supervision leading to visual detail loss and local distortions. Moreover, train-test discrepancies from teacher-forcing training, inherent to LLMs, amplify cumulative errors in visual details. These challenges make index-wise tokens a significant bottleneck for AutoRegressive models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "We propose a novel bitwise modeling framework that substitutes index-wise tokens with bitwise tokens throughout the process. Our bitwise modeling framework consists of three primary modules: bitwise visual tokenizer, bitwise infinite-vocabulary classifier, and bitwise self-correction. Inspired by the success and widespread adoption of binary vector quantization[80, 85], we scaled up the tokenizer vo" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 553, + 203 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 553, + 203 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 553, + 203 + ], + "type": "image", + "image_path": "0528a35add7e85d91bb366c2a15151c954a45def3e393ee5e18c0d89264b490e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "type": "text", + "content": "Figure 2. Left: Conventional classifier predicts " + }, + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "type": "inline_equation", + "content": "2^{d}" + }, + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "type": "text", + "content": " indices. Right: Infinite-Vocabulary Classifier predicts " + }, + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "type": "text", + "content": " bits instead, being much more efficient than index-wise classifier when " + }, + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "type": "text", + "content": " is large." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 254, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 482 + ], + "type": "text", + "content": "cabulary to " + }, + { + "bbox": [ + 313, + 254, + 555, + 482 + ], + "type": "inline_equation", + "content": "2^{64}" + }, + { + "bbox": [ + 313, + 254, + 555, + 482 + ], + "type": "text", + "content": ", significantly surpassing all previous AutoRegressive model vocabularies[58, 81]. This expansion allows for reconstruction quality that far exceeds previous discrete tokenizers, achieving results comparable to continuous VAE[51], with scores improving from 0.87 to 0.33 on ImageNet-256 benchmark[19]. In Fig.2, we transform the conventional token prediction from a large integer into binary bits in a bitwise infinite-vocabulary classifier to address optimization and computation challenges, enabling the learning of massive vocabularies in Visual AutoRegressive models. Additionally, we incorporated bitwise self-correction during training by randomly flipping some bits to simulate prediction mistakes and re-quantizing the residual features, thus endowing the system with self-correcting capabilities. Our method, Infinity: Bitwise Visual AutoRegressive Modeling, maintains the scaling and speed advantages of VAR while achieving detailed reconstruction and generation quality comparable to that of continuous Diffusion models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 483, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 555 + ], + "type": "text", + "content": "Infinity outperforms all previous AutoRegressive text-to-image models by large margins and matches or surpasses state-of-the-art diffusion models including SDXL[46], PixArt-Sigma[13],DALL-E3[6] and Stable-Diffusion 3[22] in both quantitative evaluations and human preference scores. The core contributions of our work are as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 324, + 558, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 325, + 558, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 558, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 325, + 558, + 555, + 630 + ], + "type": "text", + "content": "1. We propose Infinity, an autoregressive model utilizing Bitwise Modeling, which significantly enhances scalability and visual detail representation capabilities of discrete generative models. This framework opens up new possibilities of 'infinity' within the discrete generation community." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 630, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 630, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 324, + 630, + 554, + 689 + ], + "type": "text", + "content": "2. Infinity demonstrates the potential of scaling tokenizers and transformers by achieving near-continuous tokenizer performance with its image tokenizer, surpassing diffusion models in high-quality text-to-image generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 324, + 689, + 554, + 714 + ], + "type": "text", + "content": "3. Infinity enables an autoregressive text-to-image model to achieve exceptional prompt adherence and superior" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15734" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 295, + 96 + ], + "type": "text", + "content": "image generation quality, while also delivering unprecedented inference speed." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 143, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 143, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 143, + 119 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 127, + 187, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 127, + 187, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 127, + 187, + 140 + ], + "type": "text", + "content": "2.1. AutoRegressive Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 145, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 296, + 397 + ], + "type": "text", + "content": "AutoRegressive models, leveraging the powerful scaling capabilities of LLMs[8, 16, 47, 65, 66], use discrete image tokenizers[21, 50, 67] in conjunction with transformers to generate images based on next-token prediction. VQ-based methods [21, 35, 50, 58, 67] employ vector quantization to convert image patches into index-wise tokens and use a decoder-only transformer to predict the next token index. However, these methods are limited by the lack of scaled-up transformers and the quantization error inherent in VQ-VAE[67], preventing them from achieving performance on par with diffusion models. Parti [79], Emu3 [70], chameleon[62], Liquid[73], and VideoPoet[32] scaled up autoregressive models in text-to-image or video synthesis. Inspired by the spatial structure of visual information, Visual AutoRegressive modeling(VAR)[64] redefines the autoregressive modeling on images as next-scale prediction framework, significantly improving generation quality and sampling speed. HART[61] adopted hybrid tokenizers based on VAR. MAR[37], Fluid[23] proposed random-order models and employed continuous tokenizer rather than discrete tokenizer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 404, + 161, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 161, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 161, + 416 + ], + "type": "text", + "content": "2.2. Diffusion Models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 602 + ], + "type": "text", + "content": "Diffusion models have recently experienced rapid advancements in several directions. Denoising learning mechanisms [27, 44] and improvements in sampling efficiency [4, 40, 41, 55, 56] have been continuously optimized to generate high-quality images. Bit Diffusion [15] represents data with analog bits and then trains a continuous diffusion model. Latent diffusion models [51] were the first to propose modeling in the latent space rather than the pixel space. Recently, latent diffusion models [18, 22] have scaled up VAEs to enhance representations in the latent space. DiT [45] and U-ViT [5] employ more scalable transformers to model diffusion, achieving superior results. Consequently, mainstream text-to-image diffusion models [6, 12, 22] have adopted the DiT architecture due to its effectiveness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 610, + 135, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 135, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 135, + 623 + ], + "type": "text", + "content": "2.3. Scaling Law" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Scaling laws in autoregressive language models reveal a power-law relationship between model size, dataset size, and compute with test set cross-entropy loss [1, 25, 31]. These laws help predict larger model performance, leading to efficient resource allocation and ongoing improvements without saturation [8, 28, 65, 66, 72, 84]. This has inspired research into scaling in visual generation [7, 22, 59, 64, 82]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 71, + 400, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 85 + ], + "type": "text", + "content": "3. Infinity Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 91, + 488, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 488, + 104 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 488, + 104 + ], + "type": "text", + "content": "3.1. Visual AutoRegressive Modeling" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "content": "Infinity incorporates a visual tokenizer and a transformer for image synthesis. During the training stage, a sample consists of a text prompt " + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "content": " and a ground truth image " + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "inline_equation", + "content": "im" + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "content": ". The proposed visual tokenizer first encodes the image " + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "inline_equation", + "content": "im" + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "content": " into a feature map " + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "inline_equation", + "content": "f \\in \\mathbb{R}^{h \\times w \\times C}" + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "content": " with stride " + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "content": " and then quantizes the feature map " + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "content": " multi-scale residual maps " + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "inline_equation", + "content": "(r_1, r_2, \\dots, r_K)" + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "content": ". The resolution of " + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "inline_equation", + "content": "h_k \\times w_k" + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "content": " and it grows larger gradually from " + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "inline_equation", + "content": "k = 1 \\rightarrow K" + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "content": ". Based on this sequence of residuals, we can gradually approximate the continuous feature " + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 109, + 555, + 230 + ], + "type": "text", + "content": " as in Eq.1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 385, + 238, + 554, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 238, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 385, + 238, + 554, + 270 + ], + "type": "interline_equation", + "content": "f _ {k} = \\sum_ {i = 1} ^ {k} u p \\left(r _ {i}, (h, w)\\right) \\tag {1}", + "image_path": "2d946a60f4de5807c8078e58bb711f2562dfe555f4bc45a430751dcda0d241f8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 279, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 304 + ], + "type": "text", + "content": "up means bilinear upsampling and " + }, + { + "bbox": [ + 313, + 279, + 555, + 304 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 313, + 279, + 555, + 304 + ], + "type": "text", + "content": " is the cumulative sum of the upsampled " + }, + { + "bbox": [ + 313, + 279, + 555, + 304 + ], + "type": "inline_equation", + "content": "r_{\\leq k}" + }, + { + "bbox": [ + 313, + 279, + 555, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 304, + 554, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 352 + ], + "type": "text", + "content": "Subsequently, transformer learns to predict residuals " + }, + { + "bbox": [ + 313, + 304, + 554, + 352 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 304, + 554, + 352 + ], + "type": "text", + "content": " of the next scale conditioned on previous predictions and the text input in an autoregressive manner. Formally, the autoregressive likelihood can be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 361, + 555, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 361, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 340, + 361, + 555, + 393 + ], + "type": "interline_equation", + "content": "p \\left(r _ {1}, \\dots , r _ {K}\\right) = \\prod_ {k = 1} ^ {K} p \\left(r _ {k} \\mid r _ {1}, \\dots , r _ {k - 1}, \\psi (t)\\right), \\tag {2}", + "image_path": "8d98865408bbe4bb6f462e4e8b97a4fa9280cdb5cd1f720b3f3ab582e7a1f4c3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\psi(t)" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": " is the text embeddings from Flan-T5 [17] model. " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "(r_1, \\dots, r_{k-1}, \\psi(t))" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": " serves as the prefixed context When predicting " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": ". Besides, the text embeddings " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\psi(t)" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": " further guide the prediction through a cross-attention mechanism. In particular, as shown in Fig. 3, the text embeddings " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\psi(t) \\in \\mathbb{R}^{L \\times C}" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": " is projected into a " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{SOS} \\rangle \\in \\mathbb{R}^{1 \\times 1 \\times h}" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": " as the input of the first scale, where " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": " is the hidden dimension of transformer. The transformer is required to predict " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{SOS} \\rangle" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": " in the first scale. In the latter " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": "-th scale, to match the spatial size of the input and the output label " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "R_k" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": ", we take the downsampled feature " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_{k-1}" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": " from the last scale " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "k-1" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": " as the input to predict " + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "inline_equation", + "content": "R_k" + }, + { + "bbox": [ + 313, + 403, + 556, + 548 + ], + "type": "text", + "content": " in parallel." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 370, + 556, + 555, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 556, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 370, + 556, + 555, + 572 + ], + "type": "interline_equation", + "content": "\\widetilde {f} _ {k - 1} = \\operatorname {d o w n} \\left(f _ {k - 1}, \\left(h _ {k}, w _ {k}\\right)\\right), \\tag {3}", + "image_path": "8387f671d5420944276bc6fcb4ab1cf8dba18a1e486a3e6ef53ad95aa5bab41e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "inline_equation", + "content": "down" + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "text", + "content": " is bilinear downsampling and the spatial size of both " + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_{k-1}" + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "inline_equation", + "content": "(h_k, w_k)" + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "text", + "content": ". In previous index-wise [64] representations, the shape of " + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "inline_equation", + "content": "(h_k, w_k, V_d)" + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "inline_equation", + "content": "V_d" + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "text", + "content": " is the vocabulary size of the visual tokenizer. For binary quantization [80, 85] with code embedding dimension " + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "inline_equation", + "content": "V_d = 2^d" + }, + { + "bbox": [ + 313, + 580, + 555, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "The transformer consists of a stack of repeated blocks, where each block includes RoPE2D [26], Self-Attention, Cross Attention, and FFN layers. The text embeddings " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\psi(t)" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " provide effective guidance for image synthesis in each cross-attention layer. During the training stage, we exploit a block-wise causal attention mask to ensure that the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15735" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 73, + 324, + 312 + ], + "blocks": [ + { + "bbox": [ + 102, + 73, + 324, + 312 + ], + "lines": [ + { + "bbox": [ + 102, + 73, + 324, + 312 + ], + "spans": [ + { + "bbox": [ + 102, + 73, + 324, + 312 + ], + "type": "image", + "image_path": "d2b4972cb2f63c6654518ef6ea4f2df2d10093ee717cf02fbba77d7f224d51b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "type": "text", + "content": "Figure 3. Framework of Infinity. Infinity introduces bitwise modeling. When predicting " + }, + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "type": "text", + "content": ", the sequence " + }, + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "type": "inline_equation", + "content": "(r_1, r_2, \\dots, r_{k-1})" + }, + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "type": "text", + "content": " serves as the prefixed context and the text embeddings " + }, + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\psi(t)" + }, + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "type": "text", + "content": " guides the prediction through a cross attention mechanism." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 360, + 71, + 504, + 312 + ], + "blocks": [ + { + "bbox": [ + 360, + 71, + 504, + 312 + ], + "lines": [ + { + "bbox": [ + 360, + 71, + 504, + 312 + ], + "spans": [ + { + "bbox": [ + 360, + 71, + 504, + 312 + ], + "type": "image", + "image_path": "925163793b1f89868a0e859d1d3180ee3d889ea4f0402927d32ecfcd68510870.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 346, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 396 + ], + "type": "text", + "content": "transformer can only attend to its prefixed context, i.e., " + }, + { + "bbox": [ + 55, + 346, + 296, + 396 + ], + "type": "inline_equation", + "content": "(\\langle \\mathrm{SOS}\\rangle ,\\widetilde{f}_1,\\dots ,\\widetilde{f}_{k - 1})" + }, + { + "bbox": [ + 55, + 346, + 296, + 396 + ], + "type": "text", + "content": " , when predicting " + }, + { + "bbox": [ + 55, + 346, + 296, + 396 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 55, + 346, + 296, + 396 + ], + "type": "text", + "content": " . During the inference stage, we perform KV-Caching to speed up inference and there's no need for masking." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 404, + 162, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 162, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 162, + 415 + ], + "type": "text", + "content": "3.2. Infinite Tokenizer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 422, + 296, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 530 + ], + "type": "text", + "content": "Increasing the vocabulary size has significant potential for improving reconstruction and generation quality. However, directly enlarging the vocabulary in existing tokenizers[58, 64, 71, 81] leads to a substantial increase in memory consumption and computational burden. To address these challenges and fully exploit the potential of discrete tokenizers, this paper proposes a new bitwise multi-scale residual quantizer, which significantly reduces memory usage, enabling the training of extremely large vocabulary, e.g. " + }, + { + "bbox": [ + 55, + 422, + 296, + 530 + ], + "type": "inline_equation", + "content": "2^{64}" + }, + { + "bbox": [ + 55, + 422, + 296, + 530 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 531, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 296, + 616 + ], + "type": "text", + "content": "Bitwise Multi-scale Residual Quantizer. We replace the original vector quantizer of VAR [64] with a dimension-independent bitwise quantizer. In this paper, we consider two candidates, LFQ [81] and BSQ[85]. Given " + }, + { + "bbox": [ + 55, + 531, + 296, + 616 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 531, + 296, + 616 + ], + "type": "text", + "content": " scales in the multi-scale quantizer, on the " + }, + { + "bbox": [ + 55, + 531, + 296, + 616 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 531, + 296, + 616 + ], + "type": "text", + "content": "-th scale, the input continuous residual vector " + }, + { + "bbox": [ + 55, + 531, + 296, + 616 + ], + "type": "inline_equation", + "content": "z_{k} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 531, + 296, + 616 + ], + "type": "text", + "content": " are quantized into binary output " + }, + { + "bbox": [ + 55, + 531, + 296, + 616 + ], + "type": "inline_equation", + "content": "q_{k}" + }, + { + "bbox": [ + 55, + 531, + 296, + 616 + ], + "type": "text", + "content": " as shown below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 624, + 296, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 624, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 67, + 624, + 296, + 658 + ], + "type": "interline_equation", + "content": "q _ {k} = \\left\\{ \\begin{array}{l l} \\operatorname {s i g n} \\left(z _ {k}\\right) & \\text {l o o k u p - f r e e q u a n t i z e r} \\\\ \\frac {1}{\\sqrt {d}} \\operatorname {s i g n} \\left(\\frac {z _ {k}}{\\left| z _ {k} \\right|}\\right) & \\text {b i n a r y s p h e r i c a l q u a n t i z e r} \\end{array} \\right. \\tag {4}", + "image_path": "25eea374b1e0579489361a9c485c831ed5601e0c087a7e60f26da2cdc0db3a5f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": "To encourage codebook utilization, an entropy penalty " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{entropy} = \\mathbb{E}[H(q(z))] - H[\\mathbb{E}(q(z))]" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " [30] is adopted. To obtain the distribution of " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "q(z)" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": ", we need to compute the similarities between the input " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " and the whole codebook when" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "text", + "content": "using LFQ. However, this leads to unaffordable space and time complexity of " + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "inline_equation", + "content": "O(2^d)" + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "text", + "content": ". When the codebook dimension " + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "text", + "content": " becomes large, e.g. 20, an out-of-memory (OOM) issue is faced as shown in Tab. 3. By contrast, since both input and output in BSQ are unit vectors, BSQ[85] proposes an approximation formula for the entropy penalty, reducing the computational complexity to " + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "inline_equation", + "content": "O(d)" + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "text", + "content": ". As shown in Tab 3, there is no obvious increase in memory consumption for BSQ even when codebook size is " + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "inline_equation", + "content": "2^{64}" + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "text", + "content": ". Unless otherwise stated, we adopt BSQ by default." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 476, + 476, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 476, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 476, + 490 + ], + "type": "text", + "content": "3.3. Infinite-Vocabulary Classifier" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 496, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 554, + 581 + ], + "type": "text", + "content": "The visual tokenizer obtains discrete labels by quantizing residual features. Consequently, the transformer predicts next-scale residual features' labels " + }, + { + "bbox": [ + 313, + 496, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_k \\in [0, V_d)^{h_k \\times w_k}" + }, + { + "bbox": [ + 313, + 496, + 554, + 581 + ], + "type": "text", + "content": " and optimizes the target through the cross-entropy loss. Previous works [64, 80] directly predict these index labels using a classifier of " + }, + { + "bbox": [ + 313, + 496, + 554, + 581 + ], + "type": "inline_equation", + "content": "V_d" + }, + { + "bbox": [ + 313, + 496, + 554, + 581 + ], + "type": "text", + "content": " classes. However, it suffers from two drawbacks, huge computational costs and fuzzy supervision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "text", + "content": "As illustrated in Section 3.2, we exploit a bitwise VQ-VAE as the visual tokenizer, where the vocabulary size " + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "inline_equation", + "content": "V_{d}" + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "text", + "content": " is extremely large. For example, if " + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "inline_equation", + "content": "V_{d} = 2^{32}" + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "inline_equation", + "content": "h = 2048" + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "text", + "content": ", a conventional classifier would require a weight matrix " + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{h \\times V_d}" + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "text", + "content": " of 8.8 trillion parameters, which exceeds the limits of current computational resources." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "Moreover, VQ-VAE exploits the sign function during quantization as in Eq.4. After that, the positive elements are multiplied with the corresponding base and summed to get the index label " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_k(m,n)" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " as in Eq.5, where " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "m\\in [0,h_k)" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "n\\in [0,w_k)" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15736" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 105, + 85, + 295, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 85, + 295, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 85, + 295, + 119 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {k} (m, n) = \\sum_ {p = 0} ^ {d - 1} \\mathbb {1} _ {r _ {k} (m, n, p) > 0} \\cdot 2 ^ {p} \\tag {5}", + "image_path": "0c7f2055b2ac1bc6c519cf33a85f38d3beea1cbd26a095a4dfe12cf3a848ba6b.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 126, + 295, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 295, + 175 + ], + "type": "text", + "content": "Owing to the merits of the quantization method, slight perturbations to those near-zero features cause a significant change in the label. As a result, the conventional index-wise classifier [10, 64, 81] is difficult to optimize." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "text", + "content": "To address these problems in computation and optimization, we propose an Infinite-Vocabulary Classifier (IVC). In particular, instead of using a conventional classifier with " + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "inline_equation", + "content": "V_{d}" + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "text", + "content": " classes, we use " + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "text", + "content": " binary classifiers in parallel to predict if the next-scale residual " + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "inline_equation", + "content": "r_k(m,n,p)" + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "text", + "content": " is positive or negative, where " + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "inline_equation", + "content": "d = \\log_2(V_d)" + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "text", + "content": ". The proposed Infinite-Vocabulary Classifier is much more efficient in memory and parameters compared to the conventional classifier. When " + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "inline_equation", + "content": "V_{d} = 2^{16}" + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "inline_equation", + "content": "h = 2048" + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "text", + "content": ", it saves " + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "inline_equation", + "content": "99.95\\%" + }, + { + "bbox": [ + 55, + 175, + 296, + 344 + ], + "type": "text", + "content": " parameters and GPU memory. Besides, when there exist near-zero values that confuse the model in some dimensions, the supervision in other dimensions is still clear. Therefore, compared with conventional index-wise classifiers, the proposed Infinite-Vocabulary Classifier is easier to optimize." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 187, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 187, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 187, + 366 + ], + "type": "text", + "content": "3.4. Bitwise Self-Correction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 517 + ], + "type": "text", + "content": "Weakness of teacher-forcing training. VAR [64] inherits the teacher-forcing training from LLMs. However, next-scale prediction in vision is quite different from next-token prediction in language. Specifically, we cannot decode the complete image until residuals " + }, + { + "bbox": [ + 55, + 373, + 295, + 517 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 55, + 373, + 295, + 517 + ], + "type": "text", + "content": " from all scales are obtained. We find that the teacher-forcing training brings about severe train-test discrepancies for visual generation. In particular, the teacher-forcing training makes the transformer only refine features in each scale without the ability to recognize and correct mistakes. Mistakes made in former scales will be propagated and amplified in latter scales, finally messing up generated images (left images in Fig.10)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "text", + "content": "In this work, we propose the bitwise self-correction to address this issue. In particular, we obtain " + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "inline_equation", + "content": "r_k^{flip}" + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "text", + "content": " via randomly flipping the bits in " + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "text", + "content": " with a probability from " + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "text", + "content": ", imitating the errors in the prediction of the " + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "text", + "content": "-th scale." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "text", + "content": "Here comes the key component of bitwise self-correction. " + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "inline_equation", + "content": "r_k^{flip}" + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "text", + "content": " contains errors while " + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "text", + "content": " doesn't. After replacing " + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "inline_equation", + "content": "r_k^{flip}" + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "text", + "content": " as predictions on the " + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "text", + "content": "-th scale, we need to recompute the transformer input " + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_k" + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "text", + "content": ". Besides, requantization is performed to get a new target " + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "inline_equation", + "content": "r_{k+1}" + }, + { + "bbox": [ + 55, + 567, + 295, + 643 + ], + "type": "text", + "content": " as in Eq.6. Refer to Fig.3 (right) for a simplified illustration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 648, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 648, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 68, + 648, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} f _ {k} ^ {\\text {f l i p}} = \\sum_ {i = 1} ^ {k} u p \\left(r _ {i} ^ {\\text {f l i p}}, (h, w)\\right) \\\\ \\widetilde {f} _ {k} = \\operatorname {d o w n} \\left(f _ {k} ^ {\\text {f l i p}}, \\left(h _ {k + 1}, w _ {k + 1}\\right)\\right) \\\\ r _ {k + 1} = \\operatorname {q u a n t} \\left(\\operatorname {d o w n} \\left(f - f _ {k} ^ {\\text {f l i p}}, \\left(h _ {k + 1}, w _ {k + 1}\\right)\\right)\\right) \\end{array} \\right. \\tag {6}", + "image_path": "ff59878281ec7d8e8db3969c7fe8ea8a4bbb0329d7d05e8b0f0d08c8e2f69469.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "Each scale undergoes the same process of random-flipping and re-quantization. Here the transformer takes partially randomly flipped features as inputs, taking the prediction errors into consideration. The re-quantized bit labels enable the transformer to auto-correct errors made in former predictions. In such a way, we address the train-test discrepancy caused by teacher-forcing training and empower Infinity to have the self-correction ability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 176, + 488, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 488, + 188 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 488, + 188 + ], + "type": "text", + "content": "3.5. Dynamic Resolution Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": "Infinity can generate photo-realistic images with various aspect ratios, which is significantly different from VAR [64] that can only generate square images. The main obstacles of generating various aspect ratio images lie in two folds. The first is to define the height " + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "inline_equation", + "content": "h_k" + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "inline_equation", + "content": "w_k" + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": " based on varying aspect ratios. In the supplementary material, we pre-define a list of scales, also called scale schedule, as " + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\{(h_1^r, w_1^r), \\dots, (h_K^r, w_K^r)\\}" + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": " for each aspect ratio. We ensure that the aspect ratio of each tuple " + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "inline_equation", + "content": "(h_k^r, w_k^r)" + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": " is approximately equal to " + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": ", especially in the latter prediction scales. Additionally, for different aspect ratios at the same scale " + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": ", we keep the area of " + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "inline_equation", + "content": "h_k^r \\times w_k^r" + }, + { + "bbox": [ + 313, + 194, + 555, + 350 + ], + "type": "text", + "content": " to be roughly equal, ensuring that the training sequence lengths are roughly the same." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 350, + 556, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 495 + ], + "type": "text", + "content": "Secondly, we need to carefully design a resolution-aware positional encoding method to handle features of various scales and aspect ratios. This issue poses a significant challenge, as the existing solutions [26, 43, 57, 64, 68] exhibit substantial limitations under such conditions. In this paper, we apply RoPE2d [26] on features of each scale to preserve the intrinsic 2D structure of images. Additionally, we exploit learnable scale embeddings to avoid confusion between features of different scales. Compared to learnable APE element-wise applied on features, learnable embeddings applied on scales bring fewer parameters, can adapt to varying sequence lengths, and are easier to optimize." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 505, + 391, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 391, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 391, + 519 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 525, + 373, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 373, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 373, + 536 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 543, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 555, + 651 + ], + "type": "text", + "content": "Data Curation. We curated a large-scale dataset from open-source academic data and high-quality internally collected data. The pre-training dataset is constructed by collecting and cleaning open-source academic datasets such as LAION [54], COYO [9], OpenImages [33]. We exploit an OCR model and a watermark detection model to filter undesired images with too many texts or watermarks. Additionally, we employ Aesthetic-V2 to filter out images with low aesthetic scores." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 658, + 378, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 378, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 378, + 671 + ], + "type": "text", + "content": "4.2. Training" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "Progressive Training. Infinity redefines text-to-image as a coarse-to-fine, next-scale prediction task. In line with its architecture, we propose to train Infinity in a progressive" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15737" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "strategy. Specifically, we first train Infinity of 2B parameters on the pre-training dataset of over 100 million images with 256 resolution for 150k iterations using a batch size of 4096 and a learning rate of 6e-5. Then we switch to 512 resolution and train 110k iterations using the same hyperparameters. Next, we fine-tune Infinity at 1024 resolution with a smaller, high-quality dataset of 60 million images. In this stage, we train Infinity for 60k iterations using a batch size of 2048 and a learning rate of 2e-5. All training stages use images of varying aspect ratios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 200, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 200, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 200, + 213 + ], + "type": "text", + "content": "4.3. Text-to-Image Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 220, + 170, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 170, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 170, + 232 + ], + "type": "text", + "content": "4.3.1 Qualitative Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 239, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 312 + ], + "type": "text", + "content": "Overall Results. Fig.1 presents generated images from our Infinity-2B model, showcasing Infinity's strong capabilities in generating high-fidelity images from various categories following user prompts. Qualitative comparison results among Infinity and other top-tier models can be found in the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 312, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 384 + ], + "type": "text", + "content": "Prompt-Following. Fig.11 in the appendix presents three examples demonstrating the superior prompt-following ability of Infinity. As highlighted in red, Infinity consistently adheres to user prompts, whether they are short or extremely long. We attribute these improvements to scaling autoregressive modeling and visual tokenizer's vocabulary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 384, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 445 + ], + "type": "text", + "content": "Text Rendering. As illustrated in Fig.12 in the appendix, Infinity can render text according to user prompts across diverse categories. Despite diverse backgrounds and subjects, Infinity accurately renders corresponding texts according to user requirements, such as fonts, styles, colors, and more." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 453, + 176, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 176, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 176, + 466 + ], + "type": "text", + "content": "4.3.2 Quantitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 473, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 581 + ], + "type": "text", + "content": "Benchmark. As in Tab 1, on GenEval[24], our model with a re-writer achieves the best overall score of 0.73. Besides, Infinity also reaches the highest position reasoning score of 0.49. On DPG [29]. Our model reaches an overall score of 83.5, surpassing SDXL [46], Playground v2.5 [36], and DALLE 3 [6]. What's more, Infinity achieves the best relation score of 90.76 among all open-source T2I models, demonstrating its stronger ability to generate spatially consistent images based on user prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Human Preference Evaluation. We conduct human preference evaluation in both human studies and benchmarks. As in Fig.5, the generation results of Infinity are more frequently selected by humans in terms of overall quality, prompt following, and visual aesthetics in contrast to other open-sourced T2I models. Please refer to the supplementary material for more details. Tab.2 lists the results of three human preference benchmarks, i.e., ImageReward [77], HPSv2 [74], and VQAScore[39]. Infinity reaches the highest ImageReward and HPSv2.1, indicating our method could generate images that are more appealing to humans." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 227 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 227 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 227 + ], + "type": "table", + "html": "
Methods# ParamsGenEval↑DPG↑
Two Obj.PositionColor Attri.OverallGlobalRelationOverall
Diffusion Models
LDM [52]1.4B0.290.020.050.37---
SDv1.5 [52]0.9B0.380.040.060.4374.6373.4963.2
PixArt-alpha [11]0.6B0.500.080.070.4874.9782.5771.1
SDv2.1 [52]0.9B0.510.070.170.5077.6780.7268.1
DALL-E 2 [49]6.5B0.660.100.190.52---
DALL-E 3 [6]----\\( 0.67^† \\)90.9790.5883.5
SDXL [46]2.6B0.740.150.230.5583.2786.7674.7
PixArt-Sigma [13]0.6B0.620.140.270.5586.8986.5980.5
SD3 (d=24) [22]2B0.740.340.360.62--84.1
SD3 (d=38) [22]8B0.890.340.470.71---
SANA-1.0 [75]1.6B---0.66--84.8
FLUX-dev [34]12B---0.67--84.0
FLUX-schnell [34]12B---0.71--84.8
AutoRegressive Models
LlamaGen [58]0.8B0.340.070.040.3265.2
Chameleon [62]7B---0.39---
HART [61]0.7B---0.56--80.9
Show-o [76]1.3B0.800.310.500.68--67.5
Emu3 [70]8.5B\\( 0.81^† \\)\\( 0.49^† \\)\\( 0.45^† \\)\\( 0.66^† \\)--81.6
Infinity2B\\( 0.85^† \\)\\( 0.49^† \\)\\( 0.57^† \\)\\( 0.73^† \\)93.1190.7683.5
", + "image_path": "bb33b2d224153267eb0423b8ed5f0d4d6760e7113812aaaeb6feb138adffaa93.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 236, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 269 + ], + "type": "text", + "content": "Table 1. Evaluation on the GenEval [24] and DPG [29] benchmark. A blue background indicates methods with larger model sizes than ours. " + }, + { + "bbox": [ + 313, + 236, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 236, + 555, + 269 + ], + "type": "text", + "content": " result is with prompt rewriting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 320, + 281, + 544, + 365 + ], + "blocks": [ + { + "bbox": [ + 320, + 281, + 544, + 365 + ], + "lines": [ + { + "bbox": [ + 320, + 281, + 544, + 365 + ], + "spans": [ + { + "bbox": [ + 320, + 281, + 544, + 365 + ], + "type": "image", + "image_path": "5dcdd0ea2045b3c7e7cb134d85da92a07dadd4d32390d37f4cb0be2509ae44c2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 369, + 555, + 400 + ], + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 400 + ], + "type": "text", + "content": "Figure 4. Generation Latency (s) on " + }, + { + "bbox": [ + 313, + 369, + 555, + 400 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 369, + 555, + 400 + ], + "type": "text", + "content": ". Infinity shows a significant advantage in generation latency compared to diffusion models." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 341, + 407, + 518, + 533 + ], + "blocks": [ + { + "bbox": [ + 341, + 407, + 518, + 533 + ], + "lines": [ + { + "bbox": [ + 341, + 407, + 518, + 533 + ], + "spans": [ + { + "bbox": [ + 341, + 407, + 518, + 533 + ], + "type": "image", + "image_path": "e9d1a01261deb23886cb2b5dcb1fc85737f30904106a3e733214f0c02d0c0702.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 534, + 555, + 578 + ], + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 578 + ], + "type": "text", + "content": "Figure 5. Human Preference Evaluation. We ask users to select the better one in a side-by-side comparison in terms of Overall Quality, Prompt Following, and Visual Aesthetics. Infinity is more preferred by humans compared to other methods." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Inference Latency. As shown in Fig.4, Infinity demonstrates a significant advantage in generation speed compared to diffusion models across different model sizes. Without extra optimization, Infinity generates a " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " image in 0.8s, " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "2.6 \\times" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " faster than SD3-Medium. The speed advantage becomes more substantial as the model size increases. Infinity achieves " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "7 \\times" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " faster inference latency compared to SD3.5 (8B) [22] and Flux-Dev (12B). Besides, Infinity still achieves " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "3.7 \\times" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " speedup than Flux-Dev even with more parameters (20B v.s. 12B)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15738" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 148 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 148 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 148 + ], + "type": "table", + "html": "
Methods# ParamsImageReward↑HPSv2.1↑VQAScore↑
RankScoreRankScoreRankTime
SD-XL [46]2.6B40.600430.0630.832
SD3-Medium [22]2B30.871330.9110.906
PixArt Sigma [13]0.6B20.872231.47--
Infinity2B10.953132.0520.894
", + "image_path": "4e65dae9ca756f853f51306daec807c7d738f333d8ecca51701b7e2e1480c625.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 188, + 257, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 257, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 257, + 201 + ], + "type": "text", + "content": "4.4. Scaling Visual Tokenizer's Vocabulary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 206, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 327 + ], + "type": "text", + "content": "Scaling up the Vocabulary Benefits Reconstruction. Restricted by the vocabulary size, discrete VQ-VAEs have always lagged behind continuous ones, hindering the performance of AR-based T2I models. In this work, we successfully train a discrete VQ-VAE matching its continuous counterparts by scaling up the vocabulary size. As in Tab. 4, we observe consistent rFID improvements as scaling up the vocabulary size from " + }, + { + "bbox": [ + 55, + 206, + 295, + 327 + ], + "type": "inline_equation", + "content": "2^{16}" + }, + { + "bbox": [ + 55, + 206, + 295, + 327 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 206, + 295, + 327 + ], + "type": "inline_equation", + "content": "2^{64}" + }, + { + "bbox": [ + 55, + 206, + 295, + 327 + ], + "type": "text", + "content": ". It's noteworthy that our discrete tokenizer achieves a rFID of 0.61 on ImageNet-256 when " + }, + { + "bbox": [ + 55, + 206, + 295, + 327 + ], + "type": "inline_equation", + "content": "V_{d} = 2^{32}" + }, + { + "bbox": [ + 55, + 206, + 295, + 327 + ], + "type": "text", + "content": ", outperforming SD [52]'s contiguous VAE." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 327, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 470 + ], + "type": "text", + "content": "Infinite Vocabulary Classifier Benefits Generation. We compare predicting bit labels with IVC to predicting index labels using a conventional classifier under the vocabulary size of " + }, + { + "bbox": [ + 55, + 327, + 295, + 470 + ], + "type": "inline_equation", + "content": "2^{16}" + }, + { + "bbox": [ + 55, + 327, + 295, + 470 + ], + "type": "text", + "content": " since a larger vocabulary causes OOM for the conventional classifier. We utilize the reconstruction loss on " + }, + { + "bbox": [ + 55, + 327, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_k" + }, + { + "bbox": [ + 55, + 327, + 295, + 470 + ], + "type": "text", + "content": ", the Fréchet Inception Distance (FID) on the validation dataset and ImageReward for comprehensive evaluation. As shown in Tab.5, IVC achieves lower reconstruction loss and FID, suggesting IVC has better fitting capabilities. Beyond the quantitative results, training Infinity with IVC yields images with richer details as in Fig.6, which is consistent with a higher ImageReward." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "content": "Scaling Up the Vocabulary Benefits Generation. We then scale up the vocabulary size to " + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "inline_equation", + "content": "2^{32}" + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "content": " during training the T2I model, which exceeds the range of the Int32 data type and can be considered infinitely large. In Fig.7, we illustrate the effect of scaling up the vocabulary from " + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "inline_equation", + "content": "2^{16}" + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "inline_equation", + "content": "2^{32}" + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "content": " for image generation. For small models (125M and 361M), the vocabulary size of " + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "inline_equation", + "content": "2^{16}" + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "content": " converges faster and achieves better results. However, as we scaled up the transformer to 2.2B, the vocabulary size of " + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "inline_equation", + "content": "2^{32}" + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "content": " beats " + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "inline_equation", + "content": "2^{16}" + }, + { + "bbox": [ + 55, + 471, + 295, + 639 + ], + "type": "text", + "content": " after 40K iterations. Therefore, it's worthwhile to scale up the vocabulary along with scaling up the transformer. As illustrated in Tab.1,2, with infinite vocabulary and IVC, Infinity achieves superior performance among various benchmarks, redefining performance limits in autoregressive visual synthesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 647, + 266, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 266, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 266, + 659 + ], + "type": "text", + "content": "4.5. Scaling Visual AutoRegressive Modeling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "In Fig.8, we depict the validation loss against the total training iterations and computational FLOPs for various model sizes of Infinity. We consistently notice a reduction in validation loss with an increase in training steps and computa" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 334, + 70, + 533, + 109 + ], + "blocks": [ + { + "bbox": [ + 55, + 149, + 295, + 183 + ], + "lines": [ + { + "bbox": [ + 55, + 149, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 295, + 183 + ], + "type": "text", + "content": "Table 2. Human Preference Metrics. Infinity reaches the highest ImageReward and HPSv2.1 score, also with comparable VQAS-score to SD3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 70, + 533, + 109 + ], + "lines": [ + { + "bbox": [ + 334, + 70, + 533, + 109 + ], + "spans": [ + { + "bbox": [ + 334, + 70, + 533, + 109 + ], + "type": "table", + "html": "
Quantizerd=16d=18d=20d=32d=64
LFQ37.653.7OOMOOMOOM
BSQ32.432.432.432.432.4
", + "image_path": "14871b2604d627dedf110fcfe5a984ff12d600889058693dd56b8ea2da8b1b90.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 333, + 160, + 533, + 236 + ], + "blocks": [ + { + "bbox": [ + 313, + 111, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 313, + 111, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 553, + 156 + ], + "type": "text", + "content": "Table 3. Comparison of memory consumption (GB) between different quantizers during training. As codebook dimension " + }, + { + "bbox": [ + 313, + 111, + 553, + 156 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 111, + 553, + 156 + ], + "type": "text", + "content": " increases, MSR-BSQ shows significant advantages over MSR-LFQ, enabling nearly infinite vocabulary size of " + }, + { + "bbox": [ + 313, + 111, + 553, + 156 + ], + "type": "inline_equation", + "content": "2^{64}" + }, + { + "bbox": [ + 313, + 111, + 553, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 333, + 160, + 533, + 236 + ], + "lines": [ + { + "bbox": [ + 333, + 160, + 533, + 236 + ], + "spans": [ + { + "bbox": [ + 333, + 160, + 533, + 236 + ], + "type": "table", + "html": "
VAE (stride=16)TYPEIN-256 rFID↓IN-512 rFID↓
Vd=216Discrete1.220.31
Vd=224Discrete0.750.30
Vd=232Discrete0.610.23
Vd=264Discrete0.330.15
SD VAE [52]Continual0.87N/A
", + "image_path": "b9294d09a1650a05688fe481dbd190a5c7fc4ad97366b2c34c3282c55a9700eb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 263, + 552, + 299 + ], + "blocks": [ + { + "bbox": [ + 313, + 237, + 553, + 259 + ], + "lines": [ + { + "bbox": [ + 313, + 237, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 553, + 259 + ], + "type": "text", + "content": "Table 4. By scaling up visual tokenizer's vocabulary, discrete tokenizer surpasses continuous VAE of SD [51] on ImageNet-rFID." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 263, + 552, + 299 + ], + "lines": [ + { + "bbox": [ + 317, + 263, + 552, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 263, + 552, + 299 + ], + "type": "table", + "html": "
Classifier# ParamsvRAMRecons. Loss↓FID↓ImageReward↑
Convention124M2GB0.1845.950.79
IVC0.65M10MB0.1805.320.91
", + "image_path": "4875f38f5087508b52926a82b6ec6b104b92df93da7b06662751e16ab13b26b8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 301, + 553, + 323 + ], + "lines": [ + { + "bbox": [ + 313, + 301, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 553, + 323 + ], + "type": "text", + "content": "Table 5. In contrast to the conventional classifier, IVC saves " + }, + { + "bbox": [ + 313, + 301, + 553, + 323 + ], + "type": "inline_equation", + "content": "99.95\\%" + }, + { + "bbox": [ + 313, + 301, + 553, + 323 + ], + "type": "text", + "content": " parameters (vRAM) and reaches better performance." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 319, + 327, + 551, + 445 + ], + "blocks": [ + { + "bbox": [ + 319, + 327, + 551, + 445 + ], + "lines": [ + { + "bbox": [ + 319, + 327, + 551, + 445 + ], + "spans": [ + { + "bbox": [ + 319, + 327, + 551, + 445 + ], + "type": "image", + "image_path": "3c2668e9c39b7ed139637e89a375b29aa78f2c0f7ea3eeba4912ba9f690c84be.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 447, + 553, + 491 + ], + "lines": [ + { + "bbox": [ + 313, + 447, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 553, + 491 + ], + "type": "text", + "content": "Figure 6. Impact of Infinite-Vocabulary Classifier. Predicting bitwise labels with the Infinite-Vocabulary Classifier (Right) generates images with richer details compared to predicting index-wise labels using a conventional classifier (Left)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 497, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 568 + ], + "type": "text", + "content": "tional FLOPs. Nevertheless, the advantages gained from training smaller models for extended periods lag behind those obtained from training larger models for shorter durations. This trend aligns with findings in language models, emphasizing the promising outlook for increasing model sizes with appropriate training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In Fig.8, we plot GenEval, ImageReward, and HPSv2 scores against validation loss for different model sizes ranging from 125M to 4.7B. We observe a strong correlation between validation loss and evaluation metrics. To further quantify their correlation, we calculate the Pearson correlation coefficients through linear regression. The correlation coefficients for GenEval, ImageReward, and HPSv2 are -0.983, -0.981, and -0.979, respectively. These results demonstrate a nearly linear correlation between validation loss and the evaluation metrics when scaling up model sizes from 125M to 4.7B. This promising phenomenon encourages us to scale up Infinity to achieve better performance." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15739" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 138, + 137 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 138, + 137 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 138, + 137 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 138, + 137 + ], + "type": "image", + "image_path": "3657c0a75737e60d587cc98522964eb09c2a7e9ed8652ffeea7ed65a9f31878c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 140, + 72, + 215, + 136 + ], + "blocks": [ + { + "bbox": [ + 140, + 72, + 215, + 136 + ], + "lines": [ + { + "bbox": [ + 140, + 72, + 215, + 136 + ], + "spans": [ + { + "bbox": [ + 140, + 72, + 215, + 136 + ], + "type": "image", + "image_path": "7236248afff6ec234490ceab5a9e6ea49dbb2683706eb51d3da7a561b4a0d382.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 216, + 72, + 293, + 136 + ], + "blocks": [ + { + "bbox": [ + 216, + 72, + 293, + 136 + ], + "lines": [ + { + "bbox": [ + 216, + 72, + 293, + 136 + ], + "spans": [ + { + "bbox": [ + 216, + 72, + 293, + 136 + ], + "type": "image", + "image_path": "5ed7059f5d1331ad41e4eace9131cbe2dd3e25befb2d8606e4fdd35d93f27c45.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 137, + 137, + 201 + ], + "blocks": [ + { + "bbox": [ + 62, + 137, + 137, + 201 + ], + "lines": [ + { + "bbox": [ + 62, + 137, + 137, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 137, + 137, + 201 + ], + "type": "image", + "image_path": "c12ed33fc6b5c76d1cc5a4ef9ac3717a4e02a00caa0b04c10cc9f5d02f7e5dd5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 137, + 214, + 201 + ], + "blocks": [ + { + "bbox": [ + 140, + 137, + 214, + 201 + ], + "lines": [ + { + "bbox": [ + 140, + 137, + 214, + 201 + ], + "spans": [ + { + "bbox": [ + 140, + 137, + 214, + 201 + ], + "type": "image", + "image_path": "cbef24ae022b36e2f60c9bf522df068337a9a8d30c20ef1f83bf49070c0e348c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 216, + 137, + 293, + 201 + ], + "blocks": [ + { + "bbox": [ + 216, + 137, + 293, + 201 + ], + "lines": [ + { + "bbox": [ + 216, + 137, + 293, + 201 + ], + "spans": [ + { + "bbox": [ + 216, + 137, + 293, + 201 + ], + "type": "image", + "image_path": "d87333dfb3b6400225739574a4f6d3fa2aa3743eb42f8b27bfc255544d652407.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 62, + 202, + 137, + 266 + ], + "blocks": [ + { + "bbox": [ + 62, + 202, + 137, + 266 + ], + "lines": [ + { + "bbox": [ + 62, + 202, + 137, + 266 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 137, + 266 + ], + "type": "image", + "image_path": "3f8cefe67af63d230132a9b972eb2b46ef9b15a61752a255d4756ac9e3a7b8f4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 267, + 296, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 267, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 296, + 344 + ], + "type": "text", + "content": "Figure 7. Effects of Scaling Up the Vocabulary. We analyze the impact of scaling the vocabulary size under consistent training hyperparameters. Vocabulary size " + }, + { + "bbox": [ + 55, + 267, + 296, + 344 + ], + "type": "inline_equation", + "content": "V_{d} = 2^{16}" + }, + { + "bbox": [ + 55, + 267, + 296, + 344 + ], + "type": "text", + "content": " converges faster and achieves better results for small models (125M and 361M parameters). As we scale up the model size to 2.2B, Infinity with a vocabulary size " + }, + { + "bbox": [ + 55, + 267, + 296, + 344 + ], + "type": "inline_equation", + "content": "V_{d} = 2^{32}" + }, + { + "bbox": [ + 55, + 267, + 296, + 344 + ], + "type": "text", + "content": " beats " + }, + { + "bbox": [ + 55, + 267, + 296, + 344 + ], + "type": "inline_equation", + "content": "V_{d} = 2^{16}" + }, + { + "bbox": [ + 55, + 267, + 296, + 344 + ], + "type": "text", + "content": ". Experiment with 5M high-quality data under " + }, + { + "bbox": [ + 55, + 267, + 296, + 344 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 267, + 296, + 344 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 140, + 202, + 215, + 266 + ], + "blocks": [ + { + "bbox": [ + 140, + 202, + 215, + 266 + ], + "lines": [ + { + "bbox": [ + 140, + 202, + 215, + 266 + ], + "spans": [ + { + "bbox": [ + 140, + 202, + 215, + 266 + ], + "type": "image", + "image_path": "5d372e3b5e9ea100b49fd408da60d9cc641722859e5bcfdee42002d37f1134a8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 216, + 202, + 293, + 266 + ], + "blocks": [ + { + "bbox": [ + 216, + 202, + 293, + 266 + ], + "lines": [ + { + "bbox": [ + 216, + 202, + 293, + 266 + ], + "spans": [ + { + "bbox": [ + 216, + 202, + 293, + 266 + ], + "type": "image", + "image_path": "d42ecf68d027aaef132a35fa5ee0537d243001455acc1b7f25a0c8bf60f0f79e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 349, + 136, + 419 + ], + "blocks": [ + { + "bbox": [ + 59, + 349, + 136, + 419 + ], + "lines": [ + { + "bbox": [ + 59, + 349, + 136, + 419 + ], + "spans": [ + { + "bbox": [ + 59, + 349, + 136, + 419 + ], + "type": "image", + "image_path": "2f0695e4cbdaa9b4cb848ef48843156eeb498849b8b1f494965c0d173bbb9bca.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 139, + 349, + 214, + 419 + ], + "blocks": [ + { + "bbox": [ + 139, + 349, + 214, + 419 + ], + "lines": [ + { + "bbox": [ + 139, + 349, + 214, + 419 + ], + "spans": [ + { + "bbox": [ + 139, + 349, + 214, + 419 + ], + "type": "image", + "image_path": "f21d9d2d0e161034b4d7db26ce485706acc3bf9dede5ed150b70a85aca8ed623.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 216, + 349, + 292, + 419 + ], + "blocks": [ + { + "bbox": [ + 216, + 349, + 292, + 419 + ], + "lines": [ + { + "bbox": [ + 216, + 349, + 292, + 419 + ], + "spans": [ + { + "bbox": [ + 216, + 349, + 292, + 419 + ], + "type": "image", + "image_path": "c73df1284253e488e649031b73aae44c0a8121821655c6a78b7266a0267e4666.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 59, + 421, + 135, + 483 + ], + "blocks": [ + { + "bbox": [ + 59, + 421, + 135, + 483 + ], + "lines": [ + { + "bbox": [ + 59, + 421, + 135, + 483 + ], + "spans": [ + { + "bbox": [ + 59, + 421, + 135, + 483 + ], + "type": "image", + "image_path": "403493a93b7372ceffcf27358e714c138cc4fc5616467ca7ecae05f9c4445e23.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 485, + 296, + 562 + ], + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 562 + ], + "type": "text", + "content": "Figure 8. Scaling Visual AutoRegressive Modeling effects. We analyze the impact of scaling model size under consistent training hyperparameters (Experiment with 10M pre-training data under 256 resolution). Validation loss smoothly decreases as a function of the model size and training iters. Besides, validation loss is a strong predictor of overall performance. There is a strong correlation between validation loss and holistic image evaluation metrics." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 138, + 421, + 214, + 483 + ], + "blocks": [ + { + "bbox": [ + 138, + 421, + 214, + 483 + ], + "lines": [ + { + "bbox": [ + 138, + 421, + 214, + 483 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 214, + 483 + ], + "type": "image", + "image_path": "754992fc02106e037ab433342be1b7d57a48f5c71ac034e22e5c90a9c2013047.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 217, + 421, + 292, + 483 + ], + "blocks": [ + { + "bbox": [ + 217, + 421, + 292, + 483 + ], + "lines": [ + { + "bbox": [ + 217, + 421, + 292, + 483 + ], + "spans": [ + { + "bbox": [ + 217, + 421, + 292, + 483 + ], + "type": "image", + "image_path": "e0606cbb376018f3b27cf454bf6e228c19258124208168dcc13c080efd011cac.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 567, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 662 + ], + "type": "text", + "content": "Visualization of scaling effect. To delve deeper into the scaling effect of Infinity, we compare a set of generated " + }, + { + "bbox": [ + 55, + 567, + 296, + 662 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 567, + 296, + 662 + ], + "type": "text", + "content": " images of three model sizes (125M, 940M, 4.7B) across three distinct training schedules (10K, 40K, 90K iterations) as illustrated in Fig.9. The semantics and visual quality of generated images improve steadily when scaling up model size and training compute, which is consistent with the scaling behaviors of Infinity." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 670, + 187, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 187, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 187, + 682 + ], + "type": "text", + "content": "4.6. Bitwise Self-Correction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "In Tab.6 and Fig.10, we list the evaluation metrics and present images generated by models trained using teacher-" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 232 + ], + "type": "image", + "image_path": "dd4ed85f51c920bb7c7daaeb3694c863b3ab92b55b930420581f6f8fbf02b5f9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 232, + 553, + 253 + ], + "lines": [ + { + "bbox": [ + 314, + 232, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 314, + 232, + 553, + 253 + ], + "type": "text", + "content": "Figure 9. Semantics and visual quality improve consistently with scaling up model size and training compute." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 326, + 254, + 540, + 327 + ], + "blocks": [ + { + "bbox": [ + 326, + 254, + 540, + 327 + ], + "lines": [ + { + "bbox": [ + 326, + 254, + 540, + 327 + ], + "spans": [ + { + "bbox": [ + 326, + 254, + 540, + 327 + ], + "type": "image", + "image_path": "51edd01475b6da891194f4a3a1fce98d77ef4eaedf934b440d38a2cd75aa3cbc.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 327, + 553, + 371 + ], + "lines": [ + { + "bbox": [ + 313, + 327, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 553, + 371 + ], + "type": "text", + "content": "Figure 10. Impact of Self-Correction. Teacher-forcing training introduces great train-test discrepancy which degrades performance during inference (left). Bitwise Self-Correction autocorrects mistakes and thus generates better results (right)." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 317, + 373, + 552, + 418 + ], + "blocks": [ + { + "bbox": [ + 317, + 373, + 552, + 418 + ], + "lines": [ + { + "bbox": [ + 317, + 373, + 552, + 418 + ], + "spans": [ + { + "bbox": [ + 317, + 373, + 552, + 418 + ], + "type": "table", + "html": "
MethodFID↓ImageReward↑HPSv2.1↑
Baseline9.760.5229.53
Baseline + Random Flip9.690.5229.20
Baseline + Bitwise Self-Correction3.470.7630.71
", + "image_path": "dff2d468a23427772b8f73c13a341560887475dac1c2792f4a1fce31ec20a198.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 421, + 553, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 553, + 454 + ], + "type": "text", + "content": "Table 6. Bitwise Self-Correction makes significant improvements. Experiment with 5M high quality data and " + }, + { + "bbox": [ + 313, + 421, + 553, + 454 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 421, + 553, + 454 + ], + "type": "text", + "content": " resolution. FID is measured on the validation set with 40K images." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 456, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 563 + ], + "type": "text", + "content": "forcing and bitwise self-correction methods. Substantial advantages are observed after applying bitwise self-correction, primarily driven by the self-correction mechanism rather than applying flipping. Simply random flipping " + }, + { + "bbox": [ + 313, + 456, + 555, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_k" + }, + { + "bbox": [ + 313, + 456, + 555, + 563 + ], + "type": "text", + "content": " doesn't bring improvements. Self-correction considers errors and performs re-quantization to correct them. We emphasize that self-correction is essential for AR-based T2I models since it empowers models to correct errors automatically, significantly mitigating the train-test discrepancy." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "We introduce Infinity, a bitwise visual autoregressive model to perform Text-to-Image generation. Infinity is a pioneering framework for bitwise token modeling with the IVC and self-correction innovation. Extensive results demonstrate Infinity significantly raised the upper limit for Autoregressive Text-To-Image models, matching or surpassing leading diffusion models. We believe Infinity will significantly promote the development of autoregressive visual modeling and inspire the community for faster and more realistic generation models." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15740" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 189 + ], + "type": "text", + "content": "Many colleagues from ByteDance supported this work. We are grateful to Guanyang Deng for his efforts in data processing. We also thank Chongxi Wang and Taekmin Kim for their contributions to model deployment. Special thanks to Xiaoxiao Qin for her work in human preference evaluation. Additionally, we are thankful to Hui Wu, Fu Li, Xing Wang, Hongxiang Hao, and Chuan Li for their contributions to infrastructure." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 219, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 274, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 319 + ], + "type": "text", + "content": "[2] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 319, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 319, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 61, + 319, + 296, + 364 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 365, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 365, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 61, + 365, + 296, + 408 + ], + "type": "text", + "content": "[4] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 410, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 296, + 464 + ], + "type": "text", + "content": "[5] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22669-22679, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 466, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 296, + 521 + ], + "type": "text", + "content": "[6] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 522, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 578 + ], + "type": "text", + "content": "[7] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. OpenAI, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 578, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 296, + 634 + ], + "type": "text", + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "type": "text", + "content": "[9] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "text", + "content": "[10] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "Vision and Pattern Recognition, pages 11315-11325, 2022. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "text", + "content": "[11] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "text", + "content": "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[13] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[14] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[15] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202, 2022.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 408 + ], + "type": "text", + "content": "[16] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "text", + "content": "[17] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 466, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 521 + ], + "type": "text", + "content": "[18] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "type": "text", + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "text", + "content": "[20] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 646 + ], + "type": "text", + "content": "[21] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "text", + "content": "[22] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15741" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[23] Lijie Fan, Tianhong Li, Siyang Qin, Yuanshen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[24] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "text", + "content": "[25] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 260 + ], + "type": "text", + "content": "[26] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. In European Conference on Computer Vision, pages 289-305. Springer, 2025. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 294 + ], + "type": "text", + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 295, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 295, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 295, + 294, + 349 + ], + "type": "text", + "content": "[28] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[29] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 470 + ], + "type": "text", + "content": "[30] Aren Jansen, Daniel PW Ellis, Shawn Hershey, R Channing Moore, Manoj Plakal, Ashok C Popat, and Rif A Saurous. Coincidence, categorization, and consolidation: Learning to recognize sounds with minimal supervision. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 121-125. IEEE, 2020. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[31] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 635 + ], + "type": "text", + "content": "[32] Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jose Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, Krishna Somandepalli, Hassan Akbari, Yair Alon, Yong Cheng, Josh Dillon, Agrim Gupta, Meera Hahn, Anja Hauth, David Hendon, Alonso Martinez, David Minnen, Mikhail Sirotenko, Kihyuk Sohn, Xuan Yang, Hartwig Adam, Ming-Hsuan Yang, Irfan Essa, Huisheng Wang, David A. Ross, Bryan Seybold, and Lu Jiang. Videopoet: A large language model for zero-shot video generation, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 636, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 712 + ], + "type": "text", + "content": "[33] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 5" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[34] Black Forest Labs. Flux. https://blackforestlabs.ai/announcing-black-forest-labs/,2024.6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "text", + "content": "[35] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523-11532, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[36] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "text", + "content": "[37] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. Advances in Neural Information Processing Systems, 37:56424-56445, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "text", + "content": "[38] Xiang Li, Hao Chen, Kai Qiu, Jason Kuen, Jiumiang Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. arXiv preprint arXiv:2410.01756, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "text", + "content": "[39] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 350, + 553, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 402 + ], + "type": "text", + "content": "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "text", + "content": "[41] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "text", + "content": "[42] Chuofan Ma, Yi Jiang, Junfeng Wu, Jihan Yang, Xin Yu, Zehuan Yuan, Bingyue Peng, and Xiaojuan Qi. Unitok: A unified tokenizer for visual generation and understanding. arXiv preprint arXiv:2502.20321, 2025. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "text", + "content": "[43] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "type": "text", + "content": "[44] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[45] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 625, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 679 + ], + "type": "text", + "content": "[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 714 + ], + "type": "text", + "content": "[47] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15742" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[48] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 127, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 127, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 127, + 294, + 171 + ], + "type": "text", + "content": "[49] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "type": "text", + "content": "[50] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 259 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "text", + "content": "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "type": "text", + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "type": "text", + "content": "[54] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "text", + "content": "[55] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "text", + "content": "[56] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "text", + "content": "[57] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "text", + "content": "[58] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[59] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "text", + "content": "[60] Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, et al. Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[61] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "text", + "content": "Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[62] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024.3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[63] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "text", + "content": "[64] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2404.02905, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "text", + "content": "[65] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "type": "text", + "content": "[66] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[67] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "text", + "content": "[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "type": "text", + "content": "[69] Junke Wang, Yi Jiang, Zehuan Yuan, Bingyue Peng, Zuxuan Wu, and Yu-Gang Jiang. Omnitokensizer: A joint image-video tokenizer for visual generation. Advances in Neural Information Processing Systems, 37:28281-28295, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 478, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 531 + ], + "type": "text", + "content": "[70] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "text", + "content": "[71] Mark Weber, Lijun Yu, Qihang Yu, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. Maskbit: Embedding-free image generation via bit tokens. arXiv preprint arXiv:2409.16211, 2024. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "type": "text", + "content": "[72] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "text", + "content": "[73] Junfeng Wu, Yi Jiang, Chuofan Ma, Yuliang Liu, Hengshuang Zhao, Zehuan Yuan, Song Bai, and Xiang Bai. Liquid: Language models are scalable and unified multi-modal generators. arXiv preprint arXiv:2412.04332, 2024. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[74] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15743" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 677 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "text", + "content": "v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[75] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "text", + "content": "[76] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 274 + ], + "type": "text", + "content": "[77] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 296, + 319 + ], + "type": "text", + "content": "[78] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 374 + ], + "type": "text", + "content": "[79] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 430 + ], + "type": "text", + "content": "[80] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 296, + 485 + ], + "type": "text", + "content": "[81] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 2, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 296, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 296, + 541 + ], + "type": "text", + "content": "[82] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 543, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 296, + 586 + ], + "type": "text", + "content": "[83] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. arXiv preprint arXiv:2406.07550, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 588, + 296, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 588, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 56, + 588, + 296, + 640 + ], + "type": "text", + "content": "[84] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 643, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 643, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 643, + 296, + 677 + ], + "type": "text", + "content": "[85] Yue Zhao, Yuanjun Xiong, and Philipp Krahenbuhl. Image and video tokenization with binary spherical quantization. arXiv preprint arXiv:2406.07548, 2024. 2, 3, 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15744" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/1014c4e5-3a69-4fa3-98b3-a879e2f45651_content_list.json b/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/1014c4e5-3a69-4fa3-98b3-a879e2f45651_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1e122aac8f3e5c45281020b601d41684f48fccc1 --- /dev/null +++ b/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/1014c4e5-3a69-4fa3-98b3-a879e2f45651_content_list.json @@ -0,0 +1,2335 @@ +[ + { + "type": "text", + "text": "InsTaG: Learning Personalized 3D Talking Head from Few-Second Video", + "text_level": 1, + "bbox": [ + 127, + 130, + 869, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiahe Li $^{1}$ Jiawei Zhang $^{1}$ Xiao Bai $^{1*}$ Jin Zheng $^{1*}$ Jun Zhou $^{2}$ Lin Gu $^{3,4}$", + "bbox": [ + 174, + 179, + 818, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, State Key Laboratory of Complex & Critical Software Environment, Jiangxi Research Institute, Beihang University", + "bbox": [ + 101, + 198, + 893, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Information and Communication Technology, Griffith University", + "bbox": [ + 197, + 233, + 799, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ RIKEN AIP $^{4}$ The University of Tokyo", + "bbox": [ + 328, + 251, + 661, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9f2aa9bfb22be44ffd9dd7f4038aaa8dadc1a26e408367a36f356cc4c37a4632.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 287, + 205, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/31bd0f72d9a39afcaaf2ac6a13996ced8b6147ef6edbb82c5161baa26ec608da.jpg", + "image_caption": [ + "4 5-Second Video" + ], + "image_footnote": [], + "bbox": [ + 94, + 369, + 197, + 445 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db056d0232130caf05033539929122d44cca178ce16123337e72b896e9af1390.jpg", + "image_caption": [ + "Instant Adaptation" + ], + "image_footnote": [], + "bbox": [ + 215, + 287, + 413, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/710fe4799ae419a505052e72412b92eb1e166b2ab4bb121d60d5fdafd2b9918f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 450, + 259, + 463 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1fd73e80e5b2399c41bae8227a9e9fd8486aae97f580ba66d2d856aab3935666.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 287, + 514, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b2b14592a19a3d30f800c922b2f4c43c2bbdf9aa3c091d2dbe9a1f5eaa348f77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 287, + 607, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/faa09c48ccaf4997f752807f6f5dfcd9897dd5bec9184c77b6908dd01882456b.jpg", + "image_caption": [ + "#" + ], + "image_footnote": [], + "bbox": [ + 522, + 361, + 532, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14df0e95c8c120e7e913efb2a329be7c0655ee930e4a479e3fc84cc8ddfa6be6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 361, + 545, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3b7295e2f9ed1ccdd8a11ee0e8e8a4dd57d0454327e0c86c5ec22f2edb767315.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 549, + 361, + 558, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/979e1435fd3b909ace6fed21011c215a43e6ab1e51988dc90eddb9b56673ffdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 361, + 573, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba94422f9262a5181b40113feb9309c8e903a71447396d61fd8ecd35a92bd556.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 287, + 697, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a07d2db5b7ab3913bf8f22f69c41d5028ce01656a8509ae73eebb5b59e03f213.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 361, + 586, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e8814e81e349ac0b8b7b9897e941e2a83c87b20ec6b0cbc98d3da4fb5e1ca1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 361, + 599, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d19bf764db921942ac2e1693fd8aed4bb8a98c902b597d05a4f6e53cead7f57e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 361, + 612, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b53bc7e94c1cde46dab2794c46bdec028fd75e65762c4ee86419ab45213e821.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 361, + 627, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa8042717dc08e538db267d8e2d514512f7b1841f3fc42f49af7df2decf7d1c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 287, + 700, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4a115dc5eda4726d769dc52b70af36625c2dc2230503487514c1f9e14493fceb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 361, + 638, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1127508aa857436268f37440fdb11efb9e40ef223deb934009ca9fd95f83baaa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 361, + 653, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c9e4e8ef72f7b97dd5a2369f1ff876d98555194139a903fa3a7c2778258d7746.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 361, + 666, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e476341d19c3f90f3c5a9dce5d26b2448e3411bb11f535e3820f0c9f57e42777.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 287, + 797, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d39441fc7c6eacead48266ced4a3179b907026d24bfa10edd43bb0f55ba8e62d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 361, + 681, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1203f747297cbd8365b3806e7c9ffc6db20ce07e116d8b235409c1fb6433855f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 361, + 694, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/60892eced30bd09f1b8bfc6573686450a31de1f729d1674247691e057630c227.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 361, + 710, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5b4a430ad266f1c47c9334eeda81aa3f5a0dabe4dd92930ddb373473e9596b96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 361, + 723, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 728, + 287, + 797, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/46b31cc39b99e9e8266c9360345e4d48970b664390db3d16f9f9c416f709536c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 361, + 741, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d831fb9e16dfc3f7e0d302366dfbf5b69997f0aa1d149c047b320c684f47d5be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 361, + 759, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1532240c6186fa412b466c21031a91df84f3f7cb45468dccca1c8285012e2535.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 361, + 774, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b68e9b08100617e52c6bd73ceb7592a5100c9c082a370c23ac5e847f723d86d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 361, + 790, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/16687558939e41f2b833193f4931721e13487586c766626ad451d12dece8d7e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 287, + 901, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/898d194d65f7295f6612539a1041ae70d32a119535f5f54f480051a77fed5630.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 361, + 808, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/189b59b52811878ed832a360ec76b213c84d09df68652dad4612dfb385d59d2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 812, + 361, + 823, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f26a35b278fded0c1c6fc99521b32d643164a0bb090e73cfb8517b1e02ede095.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 361, + 838, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0fdfcfe25cc7037066b71df990b0c5665160b494a5002815df4fc93aaa03fe31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 841, + 361, + 852, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd3cf62c03c65ec8033f06af8339a6bc0280d88711159e63042cebc5b9c70784.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 859, + 361, + 867, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0499b94092db0dd2149c7623c669e599f128559cf21357b8bce7aa78beb590ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 872, + 361, + 883, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/302e18d82be343df5ca81db567fb0a7587a3ef226537040b5fe7afaca51e56e4.jpg", + "image_caption": [ + "GeneFace" + ], + "image_footnote": [], + "bbox": [ + 424, + 376, + 514, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f0b93c63e6361411c769c5e974c10b166d1826cd373ef0c359b0969a6907e051.jpg", + "image_caption": [ + "TalkingGaussian" + ], + "image_footnote": [], + "bbox": [ + 516, + 376, + 607, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c90f1ef5c16190be98058b94f21a1e7d3b5bc1fc7ceb0894726b283900282161.jpg", + "image_caption": [ + "MimicTalk" + ], + "image_footnote": [], + "bbox": [ + 607, + 376, + 700, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a2f7b8d2297913a4c533a423459b431f2a483c516450a7278c4de362dde8a15.jpg", + "image_caption": [ + "In" + ], + "image_footnote": [], + "bbox": [ + 705, + 376, + 800, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db37008c48d8d892b18be525425933e48ec5827c25dd08a96bdb0cc057b413ad.jpg", + "image_caption": [ + "Figure 1. With only 5-second video data, InsTaG outperforms the state-of-the-arts [28, 52, 53] by delivering high-quality personalized lip synchronization and realistic rendering with the fastest adaptation, meanwhile attaining low memory overhead and real-time inference." + ], + "image_footnote": [], + "bbox": [ + 715, + 452, + 725, + 464 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3410ea1cd96cd84e0e07147237f20c9a455ab8f3bda0ef45a895ca84c257ec6c.jpg", + "image_caption": [ + "G" + ], + "image_footnote": [], + "bbox": [ + 753, + 453, + 764, + 463 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/33690756a5d4706e6120b3ded57eae826e1e6b7789f24e2c8bcd137c23f78100.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 808, + 376, + 903, + 449 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 521, + 325, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite exhibiting impressive performance in synthesizing lifelike personalized 3D talking heads, prevailing methods based on radiance fields suffer from high demands for training data and time for each new identity. This paper introduces InsTaG, a 3D talking head synthesis framework that allows a fast learning of realistic personalized 3D talking head from few training data. Built upon a lightweight 3DGS person-specific synthesizer with universal motion priors, InsTaG achieves high-quality and fast adaptation while preserving high-level personalization and efficiency. As preparation, we first propose an Identity-Free Pre-training strategy that enables the pre-training of the person-specific model and encourages the collection of universal motion priors from long-video data corpus. To fully exploit the universal motion priors to learn an unseen new identity, we then present a Motion-Aligned Adaptation strategy to adaptively align the target head to the pre-trained field, and constrain a robust dynamic head structure under few training data. Experiments demonstrate our outstanding performance and efficiency under various data scenarios to render high-quality personalized talking heads. Project page:", + "bbox": [ + 88, + 565, + 483, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://fictionarry.github.io/InsTaG/.", + "bbox": [ + 89, + 883, + 465, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 520, + 643, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Audio-driven talking head synthesis has become an important technique for various digital applications like video production, virtual reality, and human-computer interaction. The recent advancements in Neural Radiance Fields (NeRF) [35] and 3D Gaussian Splatting (3DGS) [22] have led to significant progress in 3D representations. Various methods [16, 27, 28, 31, 40, 52] now utilize radiance fields to create high-fidelity 3D talking head videos, capturing personalized talking styles with impressive photorealistic image quality.", + "bbox": [ + 511, + 545, + 906, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most of these methods adopt person-specific training. Given a video clip with the target talking person, a whole person-specific model is trained to memorize the target talking head. Despite their tremendous improvements in high-fidelity rendering, personalized dynamics, and fast running speed, as a trade-off, the model holds no generalizability for an unseen identity. Thus, they require a large amount of high-quality video frames with a long training time to adapt to each new identity. Some image-referred methods [30, 40, 53, 54] can quickly generalize to different identities, nevertheless, the quality of personalization and efficiency drop. A question is then raised: How can we use the least training data to rapidly learn a faithful 3D talking", + "bbox": [ + 511, + 681, + 908, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors: Xiao Bai and Jin Zheng.", + "bbox": [ + 537, + 886, + 803, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10690", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "head, while preserving a high level of personalization?", + "bbox": [ + 89, + 90, + 454, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Observing previous radiance-field-based methods, we notice that: 1) While the training process includes a long time to learn a basic audio-driven motion that with only little personalized style, the static 3D head can be well reconstructed in just a few early iterations; 2) Though the results may exhibit jittering and poor quality when trained with limited data, a portion of the lip motions remains recognizable with distinct personalized style. These suggest that even a short video clip can provide rich personalization cues to enable the learning of a specific talking head with little time cost, as long as common audio-motion knowledge is obtained. Motivated by this, by preparing a pre-trained identity-free 3D motion field as prior and aligning it to the new identity to facilitate the adaptation, we propose Instant Talking Head Synthesis with Gaussian Splatting (InsTaG), a framework that allows fast learning of high-fidelity personalized 3D talking head from a video even as short as only a few seconds while attaining high efficiency.", + "bbox": [ + 89, + 109, + 483, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from previous few-shot methods [30, 53, 54] that imitate a new identity from inputs, InsTaG memorizes the whole talking head in a lightweight person-specific model to obtain high-level personalization and efficiency. Upon such a person-specific paradigm, we fully decouple the learning of universal motion and personalization into pre-training and adaptation, preparing a pre-trained 3D motion field for future person-specific learning of new identity. However, a challenge with this design lies in pre-training with multi-person data, where the person-specific model can not fit varying identities and estimate their varied personalized motions, leading to an ineffective optimization. To tackle this problem, we present an Identity-Free Pretraining strategy. By keeping a series of temporary personal fields to store the identity information and filter out the personalized motions, we extract the common knowledge of talking motion from long-video corpus and produce a universal motion field that contains identity-free motion prior. To further facilitate this motion detachment, a negative contrast loss is introduced to encourage the diversity of learned personalized motion between each sample, maximizing the retaining of universal priors for later adaptation.", + "bbox": [ + 89, + 383, + 483, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fully exploit the universal motion priors to enable fast and high-quality new-identity adaptation, we present a Motion-Aligned Adaptation strategy to align the unseen talking head to the pre-trained motion field. First, a motion aligner is developed to learn a condition-independent primitive-wise coordinate offset and a motion scalar. It helps the intra-alignment between the personalized and universal motions in the field, maximizing the retention of learned knowledge during few-shot fine-tuning. Moreover, to enhance the generalizability by face-mouth decomposition [58] while preserving a robust dynamic face structure, we introduce a face-mouth hook technique to inter-align the", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "inside mouth motion with the face, which efficiently improves the fidelity as well as motion quality. To further improve the robustness of view direction, the new identity's head structure is aligned with estimated geometry priors for regularization. By incorporating the alignments in the adaptation, the person-specific model finally synthesizes high-fidelity personalized 3D talking heads at a fast speed.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Integrating the two proposed strategies, InsTaG can learn a personalized 3D talking head from even a short 5-second video, achieving state-of-the-art visual quality and personalization compared to existing methods, with high lip-synchronization and real-time inference. Experiments demonstrate the outperforming efficiency and generalizability of InsTaG for various identities, genders and languages.", + "bbox": [ + 511, + 198, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 531, + 303, + 872, + 316 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- An Identity-Free Pre-training strategy that enables the pretraining of the person-specific model by filtering out conflict identity and personalized motions, meanwhile maximizing the capturing of universal motion priors.", + "- A Motion-Aligned Adaptation strategy, which inter- and intra-aligns the talking head with the motion field for efficient new-identity adaptation, attaining robust, realistic reconstruction and personalized lip-synchronization.", + "- Extensive experiments show that our InsTaG learns realistic personalized talking heads from few data, while attaining high lip-synchronization and efficiency, outperforming state-of-the-art methods under various scenarios." + ], + "bbox": [ + 513, + 321, + 903, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 515, + 653, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Talking Head Synthesis. Talking head synthesis aims to use arbitrary audio to reenact a talking person to generate audio-visual synchronized videos. Early works [5, 11, 21, 38, 45, 47, 62] are mainly built on generative models, creating talking heads by manipulating the given 2D images. Later, to solve the temporal inconsistency when the head moves, 3D-based methods [33, 44, 55, 60] utilize explicit 3D face structure and successfully improve the naturalness.", + "bbox": [ + 511, + 537, + 906, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, radiance fields like NeRF [35] and 3D Gaussian Splitting (3DGS) [22] have been introduced as the representation to allow 3D talking head synthesize. Inheriting the radiance field optimization, most methods [7, 16, 25, 28, 31, 37, 42] train a person-specific model on minutes of high-quality video, achieving success in reconstructing photorealistic rendering and personalized talking style. However, the strict quality requirements of the training video data and the long training time for every adaptation to a new identity have hugely limited their application. Although some works [30, 40, 52-54, 56] attempt to solve the problem by a one-shot generator [30, 53, 54] or take pre-trained motions from external modules to loosely join the adaptation [40, 52] to achieve one- or few-shot, they often compromise image quality and loses the personalization as the trade-off. A concurrent work MimicTalk [53] injects LoRA [20] to", + "bbox": [ + 511, + 659, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10691", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0113c62eedf6e93077e4e9fb7d58a434dbdfabbe6183c8414f97c2b72a48b895.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 89, + 898, + 233 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9df1aad750b0176fcfce7ac78676c69e2f4b857b7a8331d4eb549f54f6a40e0a.jpg", + "image_caption": [ + "Figure 2. Overview of InsTaG. For preparation, InsTaG collects the common knowledge of talking motion from a long-video corpus by Identity-Free Pre-training, storing it as a motion field. Given a short video with a new identity, the Motion-Aligned Adaptation strategy builds a robust and fast person-specific synthesizer with the pre-trained motion field to learn a high-quality personalized 3D talking head." + ], + "image_footnote": [], + "bbox": [ + 94, + 236, + 898, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "improve fine-tuning efficiency, but its inference is slow due to the large backbone. Instead, by enabling the pretraining of person-specific models and using pre-trained motion field to directly drive the talking head, our method obtains a more compact and consistent architecture to fully exploit the motion priors in an end-to-end way, facilitating modeling precise lip-synchronization with high efficiency.", + "bbox": [ + 88, + 444, + 482, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-shot 3D Head Reconstruction. Fitting high-fidelity 3D human heads from RGB images has been a hot topic. Early model-based methods [14, 43, 50, 57] focus on fitting a 3D head with 3DMMs [1, 29, 36] but are short in detail. Combined with neural fields, some model-free methods [2, 34, 39] learn static heads from a few images with structure priors. Although many works [3, 12, 15, 19, 41, 49, 51] attempt to reconstruct dynamic heads from few videos, their parametric controls are difficult to predict from audio and limit the mouth motion granularity, making it hard for them to express the highly personalized 3D talking head with delicate audio-driven mouth motions achieved by InsTaG.", + "bbox": [ + 88, + 551, + 482, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some generative methods [13, 17, 30, 48, 54, 59] focus on reconstructing talking heads with 3D structure, however, they can hardly infer personalized talking style from only one image, leading to weak personalization and fidelity. Learning 3D talking heads with few-shot NeRF, recent works [24, 40, 53] introduce 2D-to-3D modules with pre-training but are often slow due to their overhead. Instead, with an identity-free motion prior and person-specific adaptation instead of expensive 2D-to-3D priors, our InS-TaG can rapidly learn a personalized audio-driven 3D talking head from few data while attaining fast inference.", + "bbox": [ + 88, + 734, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. InsTaG", + "text_level": 1, + "bbox": [ + 511, + 443, + 601, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, InsTaG synthesizes personalized 3D talking head based on Gaussian Splatting [22] (Sec. 3.1), consisting of an Identity-Free Pre-training (Sec. 3.2) strategy to fetch universal motion priors and a Motion-Aligned Adaptation strategy (Sec. 3.3) to learn a new identity. We describe the detailed designs in the following sections.", + "bbox": [ + 511, + 465, + 906, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3DGS Talking Head Synthesizer", + "text_level": 1, + "bbox": [ + 511, + 561, + 797, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework aims to learn personalized 3D talking head with fast inference and instant adaptation, therefore, a lightweight model with high representation ability is needed to be as the 3D synthesizer. Generally, existing methods can be categorized into image-based [30, 53, 54] and person-specific [16, 25, 28, 42]. The former often possesses generalizability but is relatively slow and weak in personalization, while the latter matches our requirements.", + "bbox": [ + 511, + 582, + 905, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminaries. A person-specific synthesizer based on radiance fields can be formulated as a mapping $\\mathcal{F}$ that from given conditions to a synthesized head image $I$ :", + "bbox": [ + 511, + 704, + 905, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F}: (\\boldsymbol {a}, \\boldsymbol {e}, [ R, t ]; \\theta) \\rightarrow I \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 756, + 906, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the conditions include audio $\\mathbf{a}$ , upper-face expression control $\\mathbf{e}$ and camera pose $[R, t]$ that represents the head pose, and the parameters of the radiance fields $\\theta$ .", + "bbox": [ + 511, + 779, + 905, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Under such an architecture, we require the motion learning to be entirely irrelevant to appearance for pre-training. Hence, we leverage the recent 3DGS-based synthesizer with a face-mouth decomposition [28], including a pair of structure fields and motion fields separately for a face branch and", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10692", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "inside-mouth branch to synthesize the talking head image $I$ . Structure Field. The structure field stores the static Gaussian primitives with parameters $\\theta = \\{\\mu, s, q, \\alpha, f\\}$ . Both the face branch and inside-mouth branch obtain a private structure field for their own target structure.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion Field. The motion field predicts a point-wise deformation $\\delta = \\{\\Delta \\mu, \\Delta s, \\Delta q\\}$ with a tri-plane hash encoder $\\mathcal{H}$ [25] with a region attention (RA) module to store spatial relation, and an MLP decoder to predict deformation. For each query primitive with center $\\mu$ , the neural field $\\mathcal{D}$ predicts its deformation $\\delta$ from the given condition feature set $\\mathbf{C}$ including audio $a$ and upper-face expression signal $e$ :", + "bbox": [ + 89, + 167, + 482, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta = \\mathcal {D} (\\mu , \\mathbf {C}) = \\operatorname {M L P} (\\mathcal {H} (\\mu) \\oplus \\mathbf {C}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 279, + 482, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes concatenation.", + "bbox": [ + 89, + 303, + 303, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rendering. During rendering, the motion field deforms the Gaussians $\\theta$ depending on the inputs $a$ and $e$ , and then a 2D image is rendered. This process happens in both the face and inside-mouth branches, and the output results are combined to get the final talking head image.", + "bbox": [ + 89, + 316, + 482, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following, unless necessary, we do not show the face-mouth decomposition for easier understanding.", + "bbox": [ + 89, + 393, + 482, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Identity-Free Pre-training", + "text_level": 1, + "bbox": [ + 89, + 428, + 328, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although with personalization knowledge, a short video clip is insufficient to train a talking head with personalized audio-motion mapping from scratch. To this end, we extract the common motion knowledge from long videos by pre-training as compensation for the new identity.", + "bbox": [ + 89, + 446, + 482, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to existing generative pre-training solutions [30, 40, 53] that allow multiple-identity inputs to train the same model, an identify-conflict problem lies in our method. Firstly, different identities can not coexist in one person-specific model. Secondly, since the model can not separate different people, the conflicting personalized motion hinders the model convergence and pollutes the pretrained module by overfitting the noises.", + "bbox": [ + 89, + 523, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Strategy. To tackle this problem, we design storing the identity-dependent knowledge in a Personalized Field distributed to each training video, therefore detaching the identity influence of the learning of universal motion to pre-train a Universal Motion Field. Specifically, each Personalized Field consists of a personalized structure field with parameter $\\theta_{P}$ and a smaller motion field $\\mathcal{D}_P$ to retain the identity appearance and personalized motion of its owner video.", + "bbox": [ + 89, + 643, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Universal Motion Field (UMF). Serving as a pre-trained module, UMF aims to predict a commonly correct facial motion that suits most identities. Queried by an arbitrary Gaussian, the field $\\mathcal{U}$ predicts the universal deformation $\\delta_U$ without personalization according to Eq. (2):", + "bbox": [ + 89, + 763, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {U} = \\mathcal {U} (\\mu , \\mathbf {C}) = \\operatorname {M L P} \\left(\\mathcal {H} _ {U} (\\mu) \\oplus \\mathbf {C}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 847, + 482, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Personalized Fields. During training, Personalized Fields cooperate with one sharing Universal Motion Field to syn", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "thesize personalized talking heads, enabling the supervision from videos with different identities. Practically, given an audio feature $\\mathbf{a}$ and expression signal $\\mathbf{e}$ packaged by $\\mathbf{C}$ , the UMF $\\mathcal{U}$ would predict an identity-free deformation queried by $\\mu_P^i \\in \\theta_P^i$ , serving as a basic deformation for the $i$ -th person. Meanwhile, the $i$ -th personalized field predicts a residual motion $\\delta_P^i$ as the personalized adjustment:", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {P} ^ {i} = \\mathcal {D} _ {P} ^ {i} \\left(\\mu_ {P} ^ {i}, \\mathbf {C}\\right) = \\operatorname {M L P} \\left(\\mathcal {H} _ {P} ^ {i} \\left(\\mu_ {P} ^ {i}\\right) \\oplus \\mathbf {C}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 200, + 903, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After that, the two deformations are combined to deform the personalized structure fields for rendering:", + "bbox": [ + 511, + 223, + 903, + 253 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R}: ([ R, t ], \\widetilde {\\theta} _ {P} ^ {i}) \\rightarrow I, \\quad \\widetilde {\\theta} _ {P} ^ {i} = \\theta_ {P} ^ {i} + \\delta_ {U} + \\delta_ {P} ^ {i}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 258, + 903, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{R}$ refers to the 3DGS rendering process.", + "bbox": [ + 511, + 285, + 825, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since then, videos with different talking identities can be used for pre-training with no identity conflicts, and the personalized motions are isolated from each other by staying at the personalized fields, detached from the universal ones.", + "bbox": [ + 511, + 311, + 903, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Negative Contrast Loss. Despite the detachment of personalized and universal motions enabled, a new problem emerges: how to decide whether a motion should be personalized or universal? Without supervision, a native separation from gradient would be highly biased because of uncertainties like initialization and data disparity, making some personalized motions inappropriately stored into the universal field but too much universal knowledge omitted.", + "bbox": [ + 511, + 383, + 905, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For this problem, we introduce a Negative Contrast Loss with two principles: 1) the ideal personalized motions are expected to be the minimum additional motions that enable smooth model convergence, thus the fewer the better; 2) subject to the first point, the personalized motions of one identity should be different from that of the others. Therefore, we design the loss by a truncated dot product:", + "bbox": [ + 511, + 515, + 905, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C} \\left(\\Delta \\mu_ {P} ^ {i}, \\Delta \\mu_ {P} ^ {j}\\right) = \\mathbb {I} _ {\\text {t r u n c}} \\left(\\Delta \\mu_ {P} ^ {i} \\cdot \\Delta \\mu_ {P} ^ {j}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 626, + 903, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(\\Delta \\mu_P^i,\\Delta \\mu_P^j)$ is a pair of personalized $\\mu$ deformation from $i$ -th and $j$ -th personalized fields, queried by the same $\\mu_{P}$ and $\\mathbf{C}$ in Eq. (4), and $\\mathbb{I}_{trunc}$ is a truncation function:", + "bbox": [ + 511, + 651, + 903, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {I} _ {\\text {t r u n c}} (x) = \\left\\{ \\begin{array}{l l} x, & x > 0 \\\\ 0, & x \\leq 0 \\end{array} \\right.. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 702, + 903, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This negative-sample-based contrastive loss works from two aspects: 1) when the two participants are close in the 3D space, the loss would encourage them to be either different in direction or less in effect; 2) if the participants are already significantly different, or if one of the participants is close to 0, the loss stops pushing them too far.", + "bbox": [ + 511, + 739, + 903, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By contrasting the negative personalized $\\Delta \\mu$ pairs, the proposed loss can encourage learned personalized motion to be diverse between the different personalized fields, therefore maximizing the filtering to get universal motions.", + "bbox": [ + 511, + 840, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10693", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Motion-Aligned Adaptation", + "text_level": 1, + "bbox": [ + 89, + 90, + 341, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the pre-trained UMF from Sec. 3.2, we introduce a Motion-Aligned Adaptation strategy to allow fast and high-quality learning of new-identity talking heads from a short video. Different from existing works [30, 40, 53] that take motions to guide image generation, we directly drive Gaussians by deformation in the final rendering for more realistic visual quality, bringing a challenge in precisely aligning the pre-trained motion to the new identity's head.", + "bbox": [ + 89, + 113, + 483, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion Aligner. Due to the diversity of human faces, the pre-trained UMF can not natively match all unseen identities suitably. Except for the personalization that we aim to learn, this motion misalignment exists in two aspects: 1) the target facial structure may be biased with the implicit one from pre-training in UMF; 2) the motion of the target identity may have a scale gap from the universal motion.", + "bbox": [ + 89, + 238, + 483, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, we propose to wrap a Motion Aligner on the motion field to learn the adjustment. Specifically, a multiresolution positional encoder $\\mathcal{H}_A$ is first added to store the spatial information. Then, before querying the motion field by $\\mu$ from the structure field, a coordinate offset $\\Delta \\mu_A$ is predicted as bias adjustment:", + "bbox": [ + 89, + 348, + 483, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mu_ {A} = \\operatorname {M L P} \\left(\\mathcal {H} _ {A} (\\mu)\\right) \\in \\mathbb {R} ^ {3}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 443, + 480, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After getting the universal deformation $\\delta_U$ , the $\\Delta \\mu$ part is multiplied by a scaling factor $\\tau_A \\in \\mathbb{R}^3$ predicted in the same way as Eq. (8) for scale alignment, and finally output the adjusted deformation $\\delta_U' = \\{\\Delta \\mu', \\Delta s, \\Delta q\\}$ , where", + "bbox": [ + 89, + 467, + 483, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mu^ {\\prime} = \\Delta \\mu \\times \\tau_ {A}, \\quad \\Delta \\mu \\in \\mathcal {U} (\\mu + \\Delta \\mu_ {A}, \\mathbf {C}). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 536, + 480, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\delta_U^\\prime$ is used as the final deformation for rendering.", + "bbox": [ + 89, + 560, + 450, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Face-Mouth Hook. As a deformation-based framework, the asynchronous inside and out mouth motions often deteriorate the modeling [28]. As in Sec. 3.1, we adopt a face-mouth decomposition [28] to solve this problem. It separates the modeling of the face (including lips) and inside of the mouth into two branches. However, when data is reduced, these two parts may fail to learn to move in harmony, causing misalignment with low visual quality. Without an existing mechanism to ensure the inter-alignment, we suggest setting a hook in the motion field to allow the mouth to actively align to the face, as shown in Figure 3.", + "bbox": [ + 89, + 579, + 483, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a predicted deformation $\\delta_F$ from the face branch that represents face motion, we attempt to fetch the motions on lips from it to guide the inside-mouth deformation $\\delta_M$ . Considering the lip motion tends to cause the largest deformation amount of the whole face, we take the maximal and minimal $\\Delta \\mu^F \\in \\delta_F$ (denoted as $\\Delta \\mu_{max/min}$ ) as the motion cues, and calculate a distance $\\mu_{dist} = \\Delta \\mu_{max} - \\Delta \\mu_{min}$ for a direct scale cue. Then, this set of features $\\varphi = \\{\\Delta \\mu_{max}, \\Delta \\mu_{min}, \\mu_{dist}\\}$ are input in the MLP decoder in the inside-mouth branch motion field to predict a deforma", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b3f88ffa2ca7c54db507cbd5d6e6a90a492b05bea8da256ca2cd72c2db0630a5.jpg", + "image_caption": [ + "Figure 3. Illustration of Face-Mouth Hook. We hook the motion of the mouth to the generated face motion, allowing an alignment across two branches to enhance robustness under few training data." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 903, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion $\\Delta \\mu$ . The mouth motion field can be formulated as:", + "bbox": [ + 511, + 266, + 879, + 281 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mu^ {M} = \\mathcal {U} ^ {\\mathcal {M}} (\\mu , \\{\\boldsymbol {a}, \\varphi \\}). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 286, + 903, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Representing a structure difference, $\\mathcal{U}^{\\mathcal{M}}$ does not take $e$ as input and does not predict deformations on $r$ and $s$ .", + "bbox": [ + 511, + 311, + 903, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, considering the opening degree of the inside mouth should align with that of the lips, we hook the motion scaling prediction to the facial motion scale cue $\\mu_{dist}$ . To achieve this, a scaling factor $\\tau_{M}$ is predicted for each $\\mu$ by an MLP to get the final primitive-wise deformation:", + "bbox": [ + 511, + 342, + 905, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {\\text {h o o k}} ^ {M} = \\Delta \\mu^ {M} \\times \\tau_ {M}, \\quad \\tau_ {M} = \\operatorname {M L P} \\left(\\mathcal {H} _ {U} ^ {M} (\\mu) \\oplus \\mu_ {\\text {d i s t}}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 425, + 903, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{H}_U^M$ is the $\\mathcal{H}$ of the UMF in the inside-mouth branch.", + "bbox": [ + 511, + 452, + 903, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This hook technique establishes a connection between the motion fields to guide the mouth motion to align to the face, while the former's scale is also explicitly connected to the latter, bringing robust performance on the face-mouth alignment even when with only few training data.", + "bbox": [ + 511, + 469, + 905, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Geometry Prior Regularizer. To further reduce the geometry ambiguity caused by the lack of view cover, we leverage the power of advanced human geometry estimator [23] to provide extra geometry cues for regularization. To solve the geometry degradation [26] and instability of 3DGS at unseen views [6], we add two regularizers on the depth and surface normal. Following [6] we first calculate the 2D depth $D$ and normal $N_{s}$ when synthesizing image $I$ . Then, the estimated monocular depth map $\\check{D}$ and normal $\\check{N}$ from the ground-truth image are used for regularization:", + "bbox": [ + 511, + 544, + 906, + 695 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G e o} = \\lambda_ {D} L _ {D} (D, \\check {D}) + \\lambda_ {N} \\sum_ {i = 0} ^ {m} \\sum_ {j = 0} ^ {n} \\left(1 - N _ {i, j} \\cdot \\check {N} _ {i, j}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 699, + 903, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{D}$ is a scale-invariant depth loss, $N_{i,j}$ denotes the normal at pixel $(i,j)$ , and $(m,n)$ is the shape of image $I$ .", + "bbox": [ + 511, + 744, + 903, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Details", + "text_level": 1, + "bbox": [ + 511, + 786, + 674, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Photometric Loss. The training processes are mainly supervised by the image photometric loss $\\mathcal{L}_I$ between the synthesized image $I$ and its ground-truth $I_{GT}$ . Following 3DGS [22], $\\mathcal{L}_I$ consists of an L1 and a D-SSIM [46] terms. Pre-training. During pre-training, we collect $k$ videos as training data to simultaneously train the UMF. For each syn", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10694", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/451012626d85b042f3ed1dc6bd2f5a5d4477c14c83333f009b17feb2fcf699f9.jpg", + "image_caption": [ + "Figure 4. Qualitative Comparison on Synchronization. Our method performs best in both lip-synchronization and visual quality. \"Real3DP\" and \"TG\" denote [54] and [28]. Better visualized with zoom-in. We recommend watching the supplementary video." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 480, + 343 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e57164751141bdc6e749fb9db598847c9e386a38adaa087c3073e5abeac944d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 88, + 895, + 343 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9a217e33877e8d54cc32d1ac34bbb57c08e98a6e83e10df91b1be5be8b476e87.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSettingRendering QualityMotion QualityEfficiency
PSNR (A/F) ↑ N/ALPIPS (A/F) ↓ 0 / 0SSIM ↑ 1.000LMD ↓ 0AUE-(L/U) ↓ 0 / 0Sync-C ↑ 8.897Training ↓ -FPS ↑ -
Ground TruthReal-time
EAT [13]One-shot- / 22.44- / 0.0600.712-1.309 / 1.0117.577-11.8X
Real3DPortrait [54]23.58* / 25.370.104* / 0.0540.8213.4071.148 / 1.0987.549-8.4X
RAD-NeRF [42]Trained28.36 / 25.510.048 / 0.0350.8353.5551.295 / 0.7322.2495 hours28.6✓
ER-NeRF [25]28.23 / 25.630.040 / 0.0310.8443.5411.327 / 0.4513.0742 hours30.8✓
GaussianTalk [7]From Scratch28.18 / 25.610.043 / 0.0320.8363.6471.379 / 0.4151.97067 min64.5✓
TalkingGaussian [28]28.32 / 26.010.041 / 0.0280.8433.5881.158 / 0.3163.55631 min114.2✓
DFRF [40]Adaptation & Few-shot28.48 / 26.260.066 / 0.0340.8593.4361.247 / 0.6833.9579 hours0.03X
GeneFace [52]27.23 / 25.080.052 / 0.0360.8233.5101.310 / 0.6843.6517 hours18.1X
MimicTalk [53]24.69* / 26.270.075* / 0.0310.8533.4890.958 / 0.7756.92616 min8.2X
InsTaG (Ours)28.86 / 26.320.039 / 0.0260.8613.1670.926 / 0.3135.31813 min82.5✓
", + "bbox": [ + 96, + 385, + 906, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative Comparison in self-reconstruction setting with 5s training data. InsTaG achieves the best image quality and personalized dynamics, with fast adaption and real-time inference. We mark the best and second-best results. (*: with synthesized torso.)", + "bbox": [ + 89, + 569, + 906, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "thesized image $I^i$ from $i$ -th video, we first calculate its photometric loss, and then calculate the Negative Contrast Loss to each other. The pre-training loss for the $i$ -th video is:", + "bbox": [ + 89, + 613, + 482, + 657 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p r e} ^ {i} = \\mathcal {L} _ {I} \\left(I ^ {i}, I _ {G T} ^ {i}\\right) + \\lambda_ {C} \\sum_ {j = 1, j \\neq i} ^ {k} \\mathcal {L} _ {C} \\left(\\Delta \\mu_ {P} ^ {i} \\cdot \\Delta \\mu_ {P} ^ {j}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 659, + 482, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adaptation. For adaptation, we initialize a person-specific model with the pre-trained UMF. A warm-up stage first runs without a Motion Aligner and geometry loss for a start. After that, the model is trained with full loss:", + "bbox": [ + 89, + 702, + 482, + 762 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d a} = \\mathcal {L} _ {I} (I, I _ {G T}) + \\mathcal {L} _ {G e o} (D, N). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 768, + 482, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 787, + 223, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. In the experiments, we collect a long-video corpus with 5 long speaking videos from [25, 52] for pre-training, and follow prevailing settings [16, 25, 28, 52] to take 4 videos from publicly-released sets [40, 52] for the test, with no identity overlap of the two parts. The videos have an average length of about 800-8000 frames in 25 FPS with a", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "center portrait, all are cropped and resized to $512 \\times 512$ .", + "bbox": [ + 511, + 613, + 885, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our method is implemented on PyTorch. The pre-training stage lasts 150,000 iterations and adaptation lasts 10,000. An AdamW [32] optimizer is used with learning rates of 5e-3 and 5e-4 separately for the grid and neural networks to train UMF. $\\lambda_{D}$ , $\\lambda_{N}$ and $\\lambda_{C}$ are set to 1e-2, 1e-3 and 1. All experiments are performed on RTX 3080 Ti GPUs. The overall pre-training takes about 5 hours. The head pose is estimated through BFM [36]. A DeepSpeech model [18] is used to extract basic audio features following previous works [16, 25, 28].", + "bbox": [ + 511, + 628, + 905, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison Baselines. To compare with the state-of-the-art methods, we take the baselines from the relevant person-specific methods DFRF [40], RAD-NeRF [42], GeneFace [52], ER-NeRF [25], TalkingGaussian [58], GaussianTalker [7] and a concurrent method MimicTalk [53]. We also involve the one-shot EAT [13] and Real3DPortrait [54]. All the methods are reproduced by their official codebase. If officially supported, we use DeepSpeech as audio extractor.", + "bbox": [ + 511, + 780, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10695", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/14493a5e2c805b7245f39c0d37bfb7f63beb9772db764f24e3bc7f7d048b6601.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparison on Reconstruction Quality. Our method performs the best in rendering photorealistic talking heads with fine details. \"Real3DP\" and \"TG\" denote Real3DPortrait [54] and TalkingGaussian [28]. Better visualized with zoom-in." + ], + "image_footnote": [], + "bbox": [ + 101, + 85, + 901, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4e0fcca028354b5e541eefcfe785bb2cd6457208d54ef9edb8fda253586571d6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodType5s10s20s
PSNR ↑AUE-L ↓Sync ↑PSNR ↑AUE-L ↓Sync ↑PSNR ↑AUE-L ↓Sync ↑
ER-NeRF [25]From Scratch28.2351.3273.07428.7281.0783.24029.2870.9414.638
TalkingGaussian [28]28.3211.1583.55629.1301.0694.66329.5361.0734.776
DFRF [40]Adaptation28.4811.2473.95728.8811.0724.21029.2531.1374.333
GeneFace [52]28.2271.3103.65127.9971.0364.86928.8251.0334.772
InsTaG (Ours)28.8630.9265.31829.4210.8815.63630.3520.7105.764
", + "bbox": [ + 96, + 332, + 903, + 429 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f449b687bed18ed206ff2f2db282ddf0080ca83cf958b7b980321cd805c1d88d.jpg", + "table_caption": [ + "Table 2. Quantitative Comparison in self-reconstruction setting to person-specific baselines with different data amounts. Our method performs best in various data scenarios, exhibiting robustness. The best and second-best results are in bold and underline." + ], + "table_footnote": [], + "table_body": "
MethodsAudio English, MaleAudio German, Female
Sync-E ↓ 10.443Sync-C ↑ 4.342Sync-E ↓ 9.942Sync-C ↑ 4.949
Full Baseline [28]
ER-NeRF [25]11.7113.04311.5022.853
TalkingGaussian [28]10.6754.31710.8983.708
DFRF [40]10.5154.72610.4824.329
GeneFace [52]9.9954.73411.0294.162
InsTaG (Ours)9.8864.8289.7334.990
", + "bbox": [ + 96, + 474, + 480, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison in cross-domain setting with 20s training data. We take TalkingGaussian [28] with full data for reference.", + "bbox": [ + 89, + 595, + 482, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 633, + 212, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison Settings. The quantitative comparison contains two settings: 1) Self-reconstruction setting. We split each of the 4 adaptation videos into few-second training sets and a test set of at least 12s, and use the unseen test data as input and ground truth for evaluation. 2) Cross-domain setting. We use an unseen English male audio from NVP [44] and a German female one from [16] to test the cross-domain generalizability. Considering some baselines with failure cases, we use 20s from a certain video to train per method to represent its best performance for testing.", + "bbox": [ + 88, + 657, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics and Measurements. We employ PSNR, LPIPS [61], and SSIM [46] to evaluate image quality. Especially, we report PSNR and LPIPS separately on an all image (\"A\") and only the aligned face (\"F\"). SSIM is only measured on the face. For lip synchronization, we use LMD [4] and SyncNet [8, 9] confidence (Sync-C) and error (Sync-E). Be", + "bbox": [ + 88, + 809, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sides, action unit [10] error (AUE) is used on upper-face (AUE-U) and lower-face (AUE-L) following [28], in which AUE-L can reflect the similarity of the personalized talking actions from the target. We also statistic the training time (fine-tuning time for few-shot or adaptation-based methods) and inference FPS to evaluate efficiency.", + "bbox": [ + 511, + 477, + 906, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Evaluation. 1) Table 1 reports the results of self-reconstruction setting with 5s training data. Our method achieves the best image quality, demonstrating our ability to faithfully generate a realistic talking head. Inheriting the traditions of large corpus pre-training, the one-shot methods [13, 54] and MimicTalk [53] achieve very high Sync scores but are relatively slow. Also getting a high Sync, InsTaG achieves the best in LMD and AUE-L, showing our advantage in preserving personalization. While all baselines can not simultaneously achieve high synchronization and real-time inference, our method produces the best overall performance including the fastest adaptation and high speed. 2) When data increases, the results in Table 2 show that InsTaG still outperforms other person-specific models in all aspects, demonstrating our wider application value and robustness to various scenarios. 3) In the results from Table 3 in cross-domain setting, where the methods are tested on two difficult cross-domain scenarios, the advantages of pre-training are further amplified. GeneFace [52] and DFRF [40] surpass all models trained from scratch. At the same time, InsTaG exceeds the baseline [28] trained on about 1-minute full training data, demonstrating the out", + "bbox": [ + 511, + 568, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10696", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4313efe0e902472506b0984c5df0320e7ead22d496cde95134172915877cf6d4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsER-NeRF [25]TalkingGaussian [28]GaussianTalker [7]DFRF [40]GeneFace [52]MimicTalk [53]InsTaG (Ours)
Lip-sync Accuracy2.001.901.503.302.704.804.50
Image Quality3.403.704.002.903.103.704.80
Identity Preserving2.401.501.702.903.004.204.60
Video Realness1.401.501.402.801.703.203.90
", + "bbox": [ + 98, + 88, + 903, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e526eaca366b3a275bf2dbb1e7827758ffee23a61bbb9ef254f2ab22ff2d3e47.jpg", + "table_caption": [ + "Table 4. User Study. The rating is in the range of 1-5, higher denotes better. The best and second-best results are in bold and underline." + ], + "table_footnote": [], + "table_body": "
Pre-training PFieldNCLossAdaptation AlignerHookPSNR ↑LPIPS ↓LMD ↓Sync-C ↑
28.660.0443.7342.019
✓28.760.0403.2394.816
✓✓28.590.0413.2235.124
✓✓✓28.690.0403.2745.233
✓✓✓28.660.0403.2575.086
✓✓✓28.790.0403.2264.994
✓✓✓✓28.860.0393.1675.318
", + "bbox": [ + 98, + 199, + 480, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "standing robustness and generalizability of our method.", + "bbox": [ + 89, + 369, + 457, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Evaluation. To qualitatively evaluate the results, we show the generated sequences and rendering details from self-reconstruction setting. In Figure 4, it can be observed that the models trained from scratch (ER-NeRF, TG) [27, 28] behave poorly in synchronization (yellow box), while the adaptation-based methods (DFRF, GeneFace, MimicTalk) [40, 52, 53] perform better but generate blurry results (orange circle). Referring to Figure 5, while all methods synthesize either blurs or heavily distorted details (red arrow), our method can reconstruct photorealistic appearance and delicate details. By benefitting from the advantages of both pre-training and person-specific training, InsTaG achieves the best in both synchronization and image quality. We strongly recommend watching our supplementary video for an intuitive comparison.", + "bbox": [ + 88, + 385, + 482, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study. To better judge the quality in real human-participated scenarios, we conducted a user study with 28 generated talking head videos from 7 methods with 5s training data. We invited 10 attendees to rate each anonymous method from four aspects: (1) Lip-sync Accuracy; (2) Video Realness; (3) Identity Preserving and (4) Video Realness. The results in Table 4 show that our InsTaG achieves the best in three aspects and the second-best in lip-sync, which is only after the concurrent MimicTalk [53], demonstrating our outstanding visual quality via human judgment.", + "bbox": [ + 89, + 611, + 482, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 771, + 243, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct the ablation study to prove the effectiveness of our contributions. The results are reported in Table 5.", + "bbox": [ + 89, + 794, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Identity-Free Pre-training. In the first row, we show the results pre-trained without Personalized Fields (PField), which produces bad synchronization due to the identity-conflict problem. After applied PField, the pre-training works with obvious effect. Combined with PField, Nega", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f5f41c69e4fb401cf13f2f3d7c5df5ce6dff4aa9b93deec51e762cf0102963f1.jpg", + "table_caption": [ + "Table 5. Ablation Study. We verify the components under 5s self-reconstruction setting, showing the effect of our contributions." + ], + "table_footnote": [], + "table_body": "
BackboneHookPSNR ↑LPIPS ↓LMD ↓Sync-C ↑
Single Branch-28.590.0493.2334.675
Face-Mouth Decom.×28.690.0403.2745.233
Face-Mouth Decomp.✓28.860.0393.1675.318
", + "bbox": [ + 519, + 199, + 903, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation Study on Mouth Motion Representation. Introducing FMD [28] can effectively improve lip-synchronization, while our FM hook technique further boosts image quality.", + "bbox": [ + 511, + 257, + 906, + 299 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66725b671ecca2c21363a900d66a6c5a43d2f32026d9990a819f350a1becf75b.jpg", + "image_caption": [ + "Figure 6. Visualization Results of Face-Mouth Hook. FMD improves performance but also causes misalignment when with inadequate data, which can be well solved by our Face-Mouth Hook." + ], + "image_footnote": [], + "bbox": [ + 524, + 311, + 704, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb8efee902a7fcfda286531204fa94ef8221898c476e788be6ad4913b94c9d7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 311, + 895, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tive Contrast Loss (NCLoss) shows its impact on lip-sync improvement, offering higher Sync-C. This demonstrates the necessity of PFiled in enabling the pre-training and the effect of NCLoss on encouraging knowledge collection.", + "bbox": [ + 511, + 446, + 903, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motion-Aligned Adaptation. We apply our Motion Aligner (Aligner) and Face-Mouth Hook (FM Hook) to illustrate their effect on adaptation. Although the performance is improved by applying our pre-training strategy, the Aligner can bring extra enhancement by providing adaptation adjustments. Furthermore, the Hook works in both visual quality and dynamics by correcting inside-mouth misalignment. Combining these two adaptation techniques, the performance achieves the best among all settings.", + "bbox": [ + 511, + 506, + 905, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mouth Motion Representation. To further explain the necessity of Face-Mouth Decomposition (FMD) [28] and our FM Hook, we take studies by using one single branch to represent the two parts. Results in Table 6 show the effect of FMD. However, simply introducing it would cause significant misalignment that lowers the visual quality, as in Figure 6. By applying our FM Hook, this drawback can be well solved to benefit FMD with robust performance, generating synchronized talking heads with high visual quality.", + "bbox": [ + 511, + 642, + 906, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Ethical Consideration", + "text_level": 1, + "bbox": [ + 511, + 787, + 720, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We hope our method can promote the healthy development of digital industries. However, it is important to note that it could be misused for malicious purposes, potentially causing negative impacts. As part of our responsibility, we will assist in developing deepfake detection tools. We recommend the responsible use of this and all similar techniques.", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10697", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 191, + 187, + 208 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] James Booth, Epameinondas Antonakos, Stylianos Ploumpis, George Trigeorgis, Yannis Panagakis, and Stefanos Zafeiriou. 3d face morphable models\" in-the-wild\". In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 48-57, 2017. 3", + "[2] Marcel C Bühler, Kripasindhu Sarkar, Tanmay Shah, Gengyan Li, Daoye Wang, Leonhard Helminger, Sergio Orts-Escalano, Dmitry Lagun, Otmar Hilliges, Thabo Beeler, et al. Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3402-3413, 2023. 3", + "[3] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, et al. Authentic volumetric avatars from a phone scan. ACM Transactions on Graphics (TOG), 41(4):1-19, 2022. 3", + "[4] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Computer Vision-ECCV 2018: 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part VII 15, pages 538-553. Springer, 2018. 7", + "[5] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019. 2", + "[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 5", + "[7] Kyusun Cho, JoungBin Lee, Heeji Yoon, Yeobin Hong, Jaehoon Ko, Sangjun Ahn, and Seungryong Kim. Gaussian talker: Real-time talking head synthesis with 3d gaussian splatting. In ACM Multimedia 2024, 2024. 2, 6, 8", + "[8] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 87-103. Springer, 2017. 7", + "[9] Joon Son Chung and Andrew Zisserman. Out of time: Automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7", + "[10] Paul Ekman and Wallace V. Friesen. Facial Action Coding System: Manual. Palo Alto: Consulting Psychologists Press, 1978. 7" + ], + "bbox": [ + 93, + 215, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In this work, we are supported by the National Natural Science Foundation of China 62276016, 62372029. Lin Gu is supported by JST Moonshot R&D Grant Number JP-MJMS2011 Japan.", + "[11] Tony Ezzat, Gadi Geiger, and Tomaso Poggio. Trainable videorealistic speech animation. ACM Transactions on Graphics (TOG), 21(3):388-398, 2002. 2", + "[12] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3", + "[13] Yuan Gan, Zongxin Yang, Xihang Yue, Lingyun Sun, and Yi Yang. Efficient emotional adaptation for audio-driven talking-head generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22634-22645, 2023. 3, 6, 7", + "[14] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1155-1164, 2019. 3", + "[15] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rinz, Lourdes Agapito, and Matthias Nießner. Mononphm: Dynamic head reconstruction from monocular videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10747-10758, 2024. 3", + "[16] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 1, 2, 3, 6, 7", + "[17] Siddharth Gururani, Arun Mallya, Ting-Chun Wang, Rafael Valle, and Ming-Yu Liu. Space: Speech-driven portrait animation with controllable expression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20914-20923, 2023. 3", + "[18] Awni Hannun, Carl Case, Jared Casper, Bryan Catanzaro, Greg Diamos, Erich Elsen, Ryan Prenger, Sanjeev Satheesh, Shubho Sengupta, Adam Coates, et al. Deep speech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:1412.5567, 2014. 6", + "[19] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 3", + "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 2", + "[21] Amir Jamaludin, Joon Son Chung, and Andrew Zisserman. You said that?: Synthesising talking faces from audio. International Journal of Computer Vision, 127:1767-1779, 2019. 2", + "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5", + "[23] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10698", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 5", + "[24] Dongze Li, Kang Zhao, Wei Wang, Bo Peng, Yingya Zhang, Jing Dong, and Tieniu Tan. Ae-nerf: Audio enhanced neural radiance field for few shot talking head synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3037–3045, 2024. 3", + "[25] Jiahe Li, Jiawei Zhang, Xiao Bai, Jun Zhou, and Lin Gu. Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7568-7578, 2023. 2, 3, 4, 6, 7, 8", + "[26] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 5", + "[27] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Jun Zhou, and Lin Gu. Er-nerf++: Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. Information Fusion, 110:102456, 2024. 1, 8", + "[28] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Talkinggaussian: Structure-persistent 3d talking head synthesis via gaussian splatting. In European Conference on Computer Vision, pages 127-145. Springer, 2025. 1, 2, 3, 5, 6, 7, 8", + "[29] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 3", + "[30] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhigang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17969-17978, 2023. 1, 2, 3, 4, 5", + "[31] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 106-125. Springer, 2022. 1, 2", + "[32] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6", + "[33] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2", + "[34] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, pages 179-197. Springer, 2022. 3", + "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421. Springer, 2020. 1, 2", + "[36] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. Ieee, 2009. 3, 6", + "[37] Ziqiao Peng, Wentao Hu, Yue Shi, Xiangyu Zhu, Xiaomei Zhang, Hao Zhao, Jun He, Hongyan Liu, and Zhaoxin Fan. Synctalk: The devil is in the synchronization for talking head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 666-676, 2024. 2", + "[38] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2", + "[39] Eduard Ramon, Gil Triginer, Janna Escur, Albert Pumarola, Jaime Garcia, Xavier Giro-i Nieto, and Francesc Moreno-Noguer. H3d-net: Few-shot high-fidelity 3d head reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5620-5629, 2021. 3", + "[40] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII, pages 666-682. Springer, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[41] Luchuan Song, Pinxin Liu, Lele Chen, Guojun Yin, and Chenliang Xu. Tri²-plane: Thinking head avatar via feature pyramid. In European Conference on Computer Vision, pages 1–20. Springer, 2025. 3", + "[42] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 2, 3, 6", + "[43] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 3", + "[44] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 716-731. Springer, 2020. 2, 7", + "[45] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI, pages 700–717. Springer, 2020. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10699", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5, 7", + "[47] Olivia Wiles, A Sophia Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In Computer Vision-ECCV 2018: 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XIII 15, pages 690-706. Springer, 2018. 2", + "[48] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024.3", + "[49] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 3", + "[50] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 3", + "[51] Xihe Yang, Xingyu Chen, Daiheng Gao, Shaohui Wang, Xiaoguang Han, and Baoyuan Wang. Have-fun: Human avatar reconstruction from few-shot unconstrained images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 742-752, 2024. 3", + "[52] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, Jinzheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 6, 7, 8", + "[53] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiangwei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, Chen Zhang, Xiang Yin, Zejun Ma, and Zhou Zhao. Mimictalk: Mimicking a personalized and expressive 3d talking face in few minutes. Advances in neural information processing systems, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[54] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiawei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6, 7", + "[55] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137, 2020. 2", + "[56] Hongyun Yu, Zhan Qu, Qihang Yu, Jianchuan Chen, Zhonghua Jiang, Zhiwen Chen, Shengyu Zhang, Jimin Xu, Fei Wu, Chengfei Lv, et al. Gaussian talker: Speaker-specific talking head synthesis via 3d gaussian splattering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3548-3557, 2024. 2" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Yu Yu, Kenneth Alberto Funes Mora, and Jean-Marc Odobez. Headfusion: $360^{\\circ}$ head pose tracking combining 3d morphable model and 3d reconstruction. IEEE transactions on pattern analysis and machine intelligence, 40(11): 2653-2667, 2018. 3", + "[58] Zhentao Yu, Zixin Yin, Deyu Zhou, Duomin Wang, Finn Wong, and Baoyuan Wang. Talking head generation with probabilistic audio-to-visual diffusion priors. arXiv preprint arXiv:2212.04248, 2022. 2, 6", + "[59] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9459-9468, 2019. 3", + "[60] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3867-3876, 2021. 2", + "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7", + "[62] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2" + ], + "bbox": [ + 516, + 90, + 903, + 512 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10700", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/1014c4e5-3a69-4fa3-98b3-a879e2f45651_model.json b/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/1014c4e5-3a69-4fa3-98b3-a879e2f45651_model.json new file mode 100644 index 0000000000000000000000000000000000000000..33e35ceed25358f4918c13bb4bc8d882daa3d3b1 --- /dev/null +++ b/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/1014c4e5-3a69-4fa3-98b3-a879e2f45651_model.json @@ -0,0 +1,3038 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.131, + 0.87, + 0.152 + ], + "angle": 0, + "content": "InsTaG: Learning Personalized 3D Talking Head from Few-Second Video" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.18, + 0.82, + 0.199 + ], + "angle": 0, + "content": "Jiahe Li\\(^{1}\\) Jiawei Zhang\\(^{1}\\) Xiao Bai\\(^{1*}\\) Jin Zheng\\(^{1*}\\) Jun Zhou\\(^{2}\\) Lin Gu\\(^{3,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.199, + 0.895, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, State Key Laboratory of Complex & Critical Software Environment, Jiangxi Research Institute, Beihang University" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.234, + 0.8, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Information and Communication Technology, Griffith University" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.252, + 0.662, + 0.27 + ], + "angle": 0, + "content": "\\(^{3}\\)RIKEN AIP \\(^{4}\\)The University of Tokyo" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.288, + 0.206, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.371, + 0.198, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.45, + 0.199, + 0.464 + ], + "angle": 0, + "content": "4 5-Second Video" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.289, + 0.415, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.452, + 0.26, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.453, + 0.371, + 0.464 + ], + "angle": 0, + "content": "Instant Adaptation" + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.288, + 0.516, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.362, + 0.535, + 0.37 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.288, + 0.608, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.362, + 0.534, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.362, + 0.546, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.362, + 0.56, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.362, + 0.574, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.288, + 0.699, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.362, + 0.587, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.362, + 0.601, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.362, + 0.614, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.362, + 0.628, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.288, + 0.701, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.628, + 0.362, + 0.639, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.362, + 0.654, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.362, + 0.668, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.288, + 0.799, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.362, + 0.682, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.362, + 0.696, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.362, + 0.711, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.362, + 0.725, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.288, + 0.798, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.362, + 0.742, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.362, + 0.761, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.362, + 0.776, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.362, + 0.792, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.288, + 0.903, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.362, + 0.809, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.362, + 0.824, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.362, + 0.839, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.843, + 0.362, + 0.854, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.86, + 0.362, + 0.869, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.874, + 0.362, + 0.884, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.377, + 0.516, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.453, + 0.498, + 0.464 + ], + "angle": 0, + "content": "GeneFace" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.377, + 0.608, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.454, + 0.605, + 0.465 + ], + "angle": 0, + "content": "TalkingGaussian" + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.377, + 0.702, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.454, + 0.683, + 0.464 + ], + "angle": 0, + "content": "MimicTalk" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.377, + 0.801, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.453, + 0.726, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.454, + 0.749, + 0.464 + ], + "angle": 0, + "content": "In" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.454, + 0.765, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.454, + 0.777, + 0.464 + ], + "angle": 0, + "content": "G" + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.377, + 0.904, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.818, + 0.454, + 0.893, + 0.464 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.48, + 0.907, + 0.509 + ], + "angle": 0, + "content": "Figure 1. With only 5-second video data, InsTaG outperforms the state-of-the-arts [28, 52, 53] by delivering high-quality personalized lip synchronization and realistic rendering with the fastest adaptation, meanwhile attaining low memory overhead and real-time inference." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.522, + 0.326, + 0.536 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.566, + 0.485, + 0.883 + ], + "angle": 0, + "content": "Despite exhibiting impressive performance in synthesizing lifelike personalized 3D talking heads, prevailing methods based on radiance fields suffer from high demands for training data and time for each new identity. This paper introduces InsTaG, a 3D talking head synthesis framework that allows a fast learning of realistic personalized 3D talking head from few training data. Built upon a lightweight 3DGS person-specific synthesizer with universal motion priors, InsTaG achieves high-quality and fast adaptation while preserving high-level personalization and efficiency. As preparation, we first propose an Identity-Free Pre-training strategy that enables the pre-training of the person-specific model and encourages the collection of universal motion priors from long-video data corpus. To fully exploit the universal motion priors to learn an unseen new identity, we then present a Motion-Aligned Adaptation strategy to adaptively align the target head to the pre-trained field, and constrain a robust dynamic head structure under few training data. Experiments demonstrate our outstanding performance and efficiency under various data scenarios to render high-quality personalized talking heads. Project page:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.884, + 0.466, + 0.898 + ], + "angle": 0, + "content": "https://fictionarry.github.io/InsTaG/." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.521, + 0.645, + 0.536 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.907, + 0.682 + ], + "angle": 0, + "content": "Audio-driven talking head synthesis has become an important technique for various digital applications like video production, virtual reality, and human-computer interaction. The recent advancements in Neural Radiance Fields (NeRF) [35] and 3D Gaussian Splatting (3DGS) [22] have led to significant progress in 3D representations. Various methods [16, 27, 28, 31, 40, 52] now utilize radiance fields to create high-fidelity 3D talking head videos, capturing personalized talking styles with impressive photorealistic image quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.909, + 0.879 + ], + "angle": 0, + "content": "Most of these methods adopt person-specific training. Given a video clip with the target talking person, a whole person-specific model is trained to memorize the target talking head. Despite their tremendous improvements in high-fidelity rendering, personalized dynamics, and fast running speed, as a trade-off, the model holds no generalizability for an unseen identity. Thus, they require a large amount of high-quality video frames with a long training time to adapt to each new identity. Some image-referred methods [30, 40, 53, 54] can quickly generalize to different identities, nevertheless, the quality of personalization and efficiency drop. A question is then raised: How can we use the least training data to rapidly learn a faithful 3D talking" + }, + { + "type": "page_footnote", + "bbox": [ + 0.538, + 0.887, + 0.804, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors: Xiao Bai and Jin Zheng." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10690" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.455, + 0.108 + ], + "angle": 0, + "content": "head, while preserving a high level of personalization?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.484, + 0.382 + ], + "angle": 0, + "content": "Observing previous radiance-field-based methods, we notice that: 1) While the training process includes a long time to learn a basic audio-driven motion that with only little personalized style, the static 3D head can be well reconstructed in just a few early iterations; 2) Though the results may exhibit jittering and poor quality when trained with limited data, a portion of the lip motions remains recognizable with distinct personalized style. These suggest that even a short video clip can provide rich personalization cues to enable the learning of a specific talking head with little time cost, as long as common audio-motion knowledge is obtained. Motivated by this, by preparing a pre-trained identity-free 3D motion field as prior and aligning it to the new identity to facilitate the adaptation, we propose Instant Talking Head Synthesis with Gaussian Splatting (InsTaG), a framework that allows fast learning of high-fidelity personalized 3D talking head from a video even as short as only a few seconds while attaining high efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Different from previous few-shot methods [30, 53, 54] that imitate a new identity from inputs, InsTaG memorizes the whole talking head in a lightweight person-specific model to obtain high-level personalization and efficiency. Upon such a person-specific paradigm, we fully decouple the learning of universal motion and personalization into pre-training and adaptation, preparing a pre-trained 3D motion field for future person-specific learning of new identity. However, a challenge with this design lies in pre-training with multi-person data, where the person-specific model can not fit varying identities and estimate their varied personalized motions, leading to an ineffective optimization. To tackle this problem, we present an Identity-Free Pretraining strategy. By keeping a series of temporary personal fields to store the identity information and filter out the personalized motions, we extract the common knowledge of talking motion from long-video corpus and produce a universal motion field that contains identity-free motion prior. To further facilitate this motion detachment, a negative contrast loss is introduced to encourage the diversity of learned personalized motion between each sample, maximizing the retaining of universal priors for later adaptation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To fully exploit the universal motion priors to enable fast and high-quality new-identity adaptation, we present a Motion-Aligned Adaptation strategy to align the unseen talking head to the pre-trained motion field. First, a motion aligner is developed to learn a condition-independent primitive-wise coordinate offset and a motion scalar. It helps the intra-alignment between the personalized and universal motions in the field, maximizing the retention of learned knowledge during few-shot fine-tuning. Moreover, to enhance the generalizability by face-mouth decomposition [58] while preserving a robust dynamic face structure, we introduce a face-mouth hook technique to inter-align the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "inside mouth motion with the face, which efficiently improves the fidelity as well as motion quality. To further improve the robustness of view direction, the new identity's head structure is aligned with estimated geometry priors for regularization. By incorporating the alignments in the adaptation, the person-specific model finally synthesizes high-fidelity personalized 3D talking heads at a fast speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.304 + ], + "angle": 0, + "content": "Integrating the two proposed strategies, InsTaG can learn a personalized 3D talking head from even a short 5-second video, achieving state-of-the-art visual quality and personalization compared to existing methods, with high lip-synchronization and real-time inference. Experiments demonstrate the outperforming efficiency and generalizability of InsTaG for various identities, genders and languages." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.304, + 0.873, + 0.318 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.322, + 0.905, + 0.382 + ], + "angle": 0, + "content": "- An Identity-Free Pre-training strategy that enables the pretraining of the person-specific model by filtering out conflict identity and personalized motions, meanwhile maximizing the capturing of universal motion priors." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.382, + 0.905, + 0.442 + ], + "angle": 0, + "content": "- A Motion-Aligned Adaptation strategy, which inter- and intra-aligns the talking head with the motion field for efficient new-identity adaptation, attaining robust, realistic reconstruction and personalized lip-synchronization." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.443, + 0.905, + 0.502 + ], + "angle": 0, + "content": "- Extensive experiments show that our InsTaG learns realistic personalized talking heads from few data, while attaining high lip-synchronization and efficiency, outperforming state-of-the-art methods under various scenarios." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.322, + 0.905, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.516, + 0.655, + 0.532 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.659 + ], + "angle": 0, + "content": "3D Talking Head Synthesis. Talking head synthesis aims to use arbitrary audio to reenact a talking person to generate audio-visual synchronized videos. Early works [5, 11, 21, 38, 45, 47, 62] are mainly built on generative models, creating talking heads by manipulating the given 2D images. Later, to solve the temporal inconsistency when the head moves, 3D-based methods [33, 44, 55, 60] utilize explicit 3D face structure and successfully improve the naturalness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recently, radiance fields like NeRF [35] and 3D Gaussian Splitting (3DGS) [22] have been introduced as the representation to allow 3D talking head synthesize. Inheriting the radiance field optimization, most methods [7, 16, 25, 28, 31, 37, 42] train a person-specific model on minutes of high-quality video, achieving success in reconstructing photorealistic rendering and personalized talking style. However, the strict quality requirements of the training video data and the long training time for every adaptation to a new identity have hugely limited their application. Although some works [30, 40, 52-54, 56] attempt to solve the problem by a one-shot generator [30, 53, 54] or take pre-trained motions from external modules to loosely join the adaptation [40, 52] to achieve one- or few-shot, they often compromise image quality and loses the personalization as the trade-off. A concurrent work MimicTalk [53] injects LoRA [20] to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10691" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.9, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.237, + 0.9, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.389, + 0.907, + 0.432 + ], + "angle": 0, + "content": "Figure 2. Overview of InsTaG. For preparation, InsTaG collects the common knowledge of talking motion from a long-video corpus by Identity-Free Pre-training, storing it as a motion field. Given a short video with a new identity, the Motion-Aligned Adaptation strategy builds a robust and fast person-specific synthesizer with the pre-trained motion field to learn a high-quality personalized 3D talking head." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.445, + 0.483, + 0.55 + ], + "angle": 0, + "content": "improve fine-tuning efficiency, but its inference is slow due to the large backbone. Instead, by enabling the pretraining of person-specific models and using pre-trained motion field to directly drive the talking head, our method obtains a more compact and consistent architecture to fully exploit the motion priors in an end-to-end way, facilitating modeling precise lip-synchronization with high efficiency." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.552, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Few-shot 3D Head Reconstruction. Fitting high-fidelity 3D human heads from RGB images has been a hot topic. Early model-based methods [14, 43, 50, 57] focus on fitting a 3D head with 3DMMs [1, 29, 36] but are short in detail. Combined with neural fields, some model-free methods [2, 34, 39] learn static heads from a few images with structure priors. Although many works [3, 12, 15, 19, 41, 49, 51] attempt to reconstruct dynamic heads from few videos, their parametric controls are difficult to predict from audio and limit the mouth motion granularity, making it hard for them to express the highly personalized 3D talking head with delicate audio-driven mouth motions achieved by InsTaG." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.736, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Some generative methods [13, 17, 30, 48, 54, 59] focus on reconstructing talking heads with 3D structure, however, they can hardly infer personalized talking style from only one image, leading to weak personalization and fidelity. Learning 3D talking heads with few-shot NeRF, recent works [24, 40, 53] introduce 2D-to-3D modules with pre-training but are often slow due to their overhead. Instead, with an identity-free motion prior and person-specific adaptation instead of expensive 2D-to-3D priors, our InS-TaG can rapidly learn a personalized audio-driven 3D talking head from few data while attaining fast inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.444, + 0.602, + 0.459 + ], + "angle": 0, + "content": "3. InsTaG" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.907, + 0.557 + ], + "angle": 0, + "content": "As shown in Figure 2, InsTaG synthesizes personalized 3D talking head based on Gaussian Splatting [22] (Sec. 3.1), consisting of an Identity-Free Pre-training (Sec. 3.2) strategy to fetch universal motion priors and a Motion-Aligned Adaptation strategy (Sec. 3.3) to learn a new identity. We describe the detailed designs in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.563, + 0.799, + 0.58 + ], + "angle": 0, + "content": "3.1. 3DGS Talking Head Synthesizer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Our framework aims to learn personalized 3D talking head with fast inference and instant adaptation, therefore, a lightweight model with high representation ability is needed to be as the 3D synthesizer. Generally, existing methods can be categorized into image-based [30, 53, 54] and person-specific [16, 25, 28, 42]. The former often possesses generalizability but is relatively slow and weak in personalization, while the latter matches our requirements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.75 + ], + "angle": 0, + "content": "Preliminaries. A person-specific synthesizer based on radiance fields can be formulated as a mapping \\(\\mathcal{F}\\) that from given conditions to a synthesized head image \\(I\\):" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.757, + 0.907, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F}: (\\boldsymbol {a}, \\boldsymbol {e}, [ R, t ]; \\theta) \\rightarrow I \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.825 + ], + "angle": 0, + "content": "where the conditions include audio \\( \\mathbf{a} \\), upper-face expression control \\( \\mathbf{e} \\) and camera pose \\( [R, t] \\) that represents the head pose, and the parameters of the radiance fields \\( \\theta \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Under such an architecture, we require the motion learning to be entirely irrelevant to appearance for pre-training. Hence, we leverage the recent 3DGS-based synthesizer with a face-mouth decomposition [28], including a pair of structure fields and motion fields separately for a face branch and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10692" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "inside-mouth branch to synthesize the talking head image \\(I\\). Structure Field. The structure field stores the static Gaussian primitives with parameters \\(\\theta = \\{\\mu, s, q, \\alpha, f\\}\\). Both the face branch and inside-mouth branch obtain a private structure field for their own target structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.273 + ], + "angle": 0, + "content": "Motion Field. The motion field predicts a point-wise deformation \\(\\delta = \\{\\Delta \\mu, \\Delta s, \\Delta q\\}\\) with a tri-plane hash encoder \\(\\mathcal{H}\\) [25] with a region attention (RA) module to store spatial relation, and an MLP decoder to predict deformation. For each query primitive with center \\(\\mu\\), the neural field \\(\\mathcal{D}\\) predicts its deformation \\(\\delta\\) from the given condition feature set \\(\\mathbf{C}\\) including audio \\(a\\) and upper-face expression signal \\(e\\):" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.28, + 0.483, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\delta = \\mathcal {D} (\\mu , \\mathbf {C}) = \\operatorname {M L P} (\\mathcal {H} (\\mu) \\oplus \\mathbf {C}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.305, + 0.317 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes concatenation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.483, + 0.394 + ], + "angle": 0, + "content": "Rendering. During rendering, the motion field deforms the Gaussians \\(\\theta\\) depending on the inputs \\(a\\) and \\(e\\), and then a 2D image is rendered. This process happens in both the face and inside-mouth branches, and the output results are combined to get the final talking head image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.483, + 0.424 + ], + "angle": 0, + "content": "In the following, unless necessary, we do not show the face-mouth decomposition for easier understanding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.429, + 0.329, + 0.445 + ], + "angle": 0, + "content": "3.2. Identity-Free Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.483, + 0.524 + ], + "angle": 0, + "content": "Although with personalization knowledge, a short video clip is insufficient to train a talking head with personalized audio-motion mapping from scratch. To this end, we extract the common motion knowledge from long videos by pre-training as compensation for the new identity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.644 + ], + "angle": 0, + "content": "In contrast to existing generative pre-training solutions [30, 40, 53] that allow multiple-identity inputs to train the same model, an identify-conflict problem lies in our method. Firstly, different identities can not coexist in one person-specific model. Secondly, since the model can not separate different people, the conflicting personalized motion hinders the model convergence and pollutes the pretrained module by overfitting the noises." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Strategy. To tackle this problem, we design storing the identity-dependent knowledge in a Personalized Field distributed to each training video, therefore detaching the identity influence of the learning of universal motion to pre-train a Universal Motion Field. Specifically, each Personalized Field consists of a personalized structure field with parameter \\(\\theta_{P}\\) and a smaller motion field \\(\\mathcal{D}_P\\) to retain the identity appearance and personalized motion of its owner video." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Universal Motion Field (UMF). Serving as a pre-trained module, UMF aims to predict a commonly correct facial motion that suits most identities. Queried by an arbitrary Gaussian, the field \\(\\mathcal{U}\\) predicts the universal deformation \\(\\delta_U\\) without personalization according to Eq. (2):" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.848, + 0.483, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\delta_ {U} = \\mathcal {U} (\\mu , \\mathbf {C}) = \\operatorname {M L P} \\left(\\mathcal {H} _ {U} (\\mu) \\oplus \\mathbf {C}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Personalized Fields. During training, Personalized Fields cooperate with one sharing Universal Motion Field to syn" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "thesize personalized talking heads, enabling the supervision from videos with different identities. Practically, given an audio feature \\( \\mathbf{a} \\) and expression signal \\( \\mathbf{e} \\) packaged by \\( \\mathbf{C} \\), the UMF \\( \\mathcal{U} \\) would predict an identity-free deformation queried by \\( \\mu_P^i \\in \\theta_P^i \\), serving as a basic deformation for the \\( i \\)-th person. Meanwhile, the \\( i \\)-th personalized field predicts a residual motion \\( \\delta_P^i \\) as the personalized adjustment:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.201, + 0.905, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\delta_ {P} ^ {i} = \\mathcal {D} _ {P} ^ {i} \\left(\\mu_ {P} ^ {i}, \\mathbf {C}\\right) = \\operatorname {M L P} \\left(\\mathcal {H} _ {P} ^ {i} \\left(\\mu_ {P} ^ {i}\\right) \\oplus \\mathbf {C}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.224, + 0.905, + 0.254 + ], + "angle": 0, + "content": "After that, the two deformations are combined to deform the personalized structure fields for rendering:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.26, + 0.905, + 0.28 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R}: ([ R, t ], \\widetilde {\\theta} _ {P} ^ {i}) \\rightarrow I, \\quad \\widetilde {\\theta} _ {P} ^ {i} = \\theta_ {P} ^ {i} + \\delta_ {U} + \\delta_ {P} ^ {i}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.286, + 0.826, + 0.301 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}\\) refers to the 3DGS rendering process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.905, + 0.372 + ], + "angle": 0, + "content": "Since then, videos with different talking identities can be used for pre-training with no identity conflicts, and the personalized motions are isolated from each other by staying at the personalized fields, detached from the universal ones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.906, + 0.504 + ], + "angle": 0, + "content": "Negative Contrast Loss. Despite the detachment of personalized and universal motions enabled, a new problem emerges: how to decide whether a motion should be personalized or universal? Without supervision, a native separation from gradient would be highly biased because of uncertainties like initialization and data disparity, making some personalized motions inappropriately stored into the universal field but too much universal knowledge omitted." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.621 + ], + "angle": 0, + "content": "For this problem, we introduce a Negative Contrast Loss with two principles: 1) the ideal personalized motions are expected to be the minimum additional motions that enable smooth model convergence, thus the fewer the better; 2) subject to the first point, the personalized motions of one identity should be different from that of the others. Therefore, we design the loss by a truncated dot product:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.627, + 0.905, + 0.646 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C} \\left(\\Delta \\mu_ {P} ^ {i}, \\Delta \\mu_ {P} ^ {j}\\right) = \\mathbb {I} _ {\\text {t r u n c}} \\left(\\Delta \\mu_ {P} ^ {i} \\cdot \\Delta \\mu_ {P} ^ {j}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.905, + 0.699 + ], + "angle": 0, + "content": "where \\((\\Delta \\mu_P^i,\\Delta \\mu_P^j)\\) is a pair of personalized \\(\\mu\\) deformation from \\(i\\) -th and \\(j\\) -th personalized fields, queried by the same \\(\\mu_{P}\\) and \\(\\mathbf{C}\\) in Eq. (4), and \\(\\mathbb{I}_{trunc}\\) is a truncation function:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.703, + 0.905, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\mathbb {I} _ {\\text {t r u n c}} (x) = \\left\\{ \\begin{array}{l l} x, & x > 0 \\\\ 0, & x \\leq 0 \\end{array} \\right.. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.74, + 0.905, + 0.83 + ], + "angle": 0, + "content": "This negative-sample-based contrastive loss works from two aspects: 1) when the two participants are close in the 3D space, the loss would encourage them to be either different in direction or less in effect; 2) if the participants are already significantly different, or if one of the participants is close to 0, the loss stops pushing them too far." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.842, + 0.905, + 0.902 + ], + "angle": 0, + "content": "By contrasting the negative personalized \\(\\Delta \\mu\\) pairs, the proposed loss can encourage learned personalized motion to be diverse between the different personalized fields, therefore maximizing the filtering to get universal motions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10693" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.342, + 0.108 + ], + "angle": 0, + "content": "3.3. Motion-Aligned Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.236 + ], + "angle": 0, + "content": "Based on the pre-trained UMF from Sec. 3.2, we introduce a Motion-Aligned Adaptation strategy to allow fast and high-quality learning of new-identity talking heads from a short video. Different from existing works [30, 40, 53] that take motions to guide image generation, we directly drive Gaussians by deformation in the final rendering for more realistic visual quality, bringing a challenge in precisely aligning the pre-trained motion to the new identity's head." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.239, + 0.484, + 0.345 + ], + "angle": 0, + "content": "Motion Aligner. Due to the diversity of human faces, the pre-trained UMF can not natively match all unseen identities suitably. Except for the personalization that we aim to learn, this motion misalignment exists in two aspects: 1) the target facial structure may be biased with the implicit one from pre-training in UMF; 2) the motion of the target identity may have a scale gap from the universal motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.484, + 0.44 + ], + "angle": 0, + "content": "To this end, we propose to wrap a Motion Aligner on the motion field to learn the adjustment. Specifically, a multiresolution positional encoder \\(\\mathcal{H}_A\\) is first added to store the spatial information. Then, before querying the motion field by \\(\\mu\\) from the structure field, a coordinate offset \\(\\Delta \\mu_A\\) is predicted as bias adjustment:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.444, + 0.482, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mu_ {A} = \\operatorname {M L P} \\left(\\mathcal {H} _ {A} (\\mu)\\right) \\in \\mathbb {R} ^ {3}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.484, + 0.53 + ], + "angle": 0, + "content": "After getting the universal deformation \\(\\delta_U\\), the \\(\\Delta \\mu\\) part is multiplied by a scaling factor \\(\\tau_A \\in \\mathbb{R}^3\\) predicted in the same way as Eq. (8) for scale alignment, and finally output the adjusted deformation \\(\\delta_U' = \\{\\Delta \\mu', \\Delta s, \\Delta q\\}\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.537, + 0.482, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mu^ {\\prime} = \\Delta \\mu \\times \\tau_ {A}, \\quad \\Delta \\mu \\in \\mathcal {U} (\\mu + \\Delta \\mu_ {A}, \\mathbf {C}). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.452, + 0.577 + ], + "angle": 0, + "content": "Here \\(\\delta_U^\\prime\\) is used as the final deformation for rendering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.746 + ], + "angle": 0, + "content": "Face-Mouth Hook. As a deformation-based framework, the asynchronous inside and out mouth motions often deteriorate the modeling [28]. As in Sec. 3.1, we adopt a face-mouth decomposition [28] to solve this problem. It separates the modeling of the face (including lips) and inside of the mouth into two branches. However, when data is reduced, these two parts may fail to learn to move in harmony, causing misalignment with low visual quality. Without an existing mechanism to ensure the inter-alignment, we suggest setting a hook in the motion field to allow the mouth to actively align to the face, as shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Given a predicted deformation \\(\\delta_F\\) from the face branch that represents face motion, we attempt to fetch the motions on lips from it to guide the inside-mouth deformation \\(\\delta_M\\). Considering the lip motion tends to cause the largest deformation amount of the whole face, we take the maximal and minimal \\(\\Delta \\mu^F \\in \\delta_F\\) (denoted as \\(\\Delta \\mu_{max/min}\\)) as the motion cues, and calculate a distance \\(\\mu_{dist} = \\Delta \\mu_{max} - \\Delta \\mu_{min}\\) for a direct scale cue. Then, this set of features \\(\\varphi = \\{\\Delta \\mu_{max}, \\Delta \\mu_{min}, \\mu_{dist}\\}\\) are input in the MLP decoder in the inside-mouth branch motion field to predict a deforma" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.212, + 0.907, + 0.254 + ], + "angle": 0, + "content": "Figure 3. Illustration of Face-Mouth Hook. We hook the motion of the mouth to the generated face motion, allowing an alignment across two branches to enhance robustness under few training data." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.267, + 0.88, + 0.282 + ], + "angle": 0, + "content": "tion \\(\\Delta \\mu\\). The mouth motion field can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.287, + 0.905, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mu^ {M} = \\mathcal {U} ^ {\\mathcal {M}} (\\mu , \\{\\boldsymbol {a}, \\varphi \\}). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.312, + 0.905, + 0.342 + ], + "angle": 0, + "content": "Representing a structure difference, \\(\\mathcal{U}^{\\mathcal{M}}\\) does not take \\(e\\) as input and does not predict deformations on \\(r\\) and \\(s\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.343, + 0.906, + 0.419 + ], + "angle": 0, + "content": "Moreover, considering the opening degree of the inside mouth should align with that of the lips, we hook the motion scaling prediction to the facial motion scale cue \\(\\mu_{dist}\\). To achieve this, a scaling factor \\(\\tau_{M}\\) is predicted for each \\(\\mu\\) by an MLP to get the final primitive-wise deformation:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.426, + 0.905, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\mu_ {\\text {h o o k}} ^ {M} = \\Delta \\mu^ {M} \\times \\tau_ {M}, \\quad \\tau_ {M} = \\operatorname {M L P} \\left(\\mathcal {H} _ {U} ^ {M} (\\mu) \\oplus \\mu_ {\\text {d i s t}}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.453, + 0.905, + 0.469 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}_U^M\\) is the \\(\\mathcal{H}\\) of the UMF in the inside-mouth branch." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.544 + ], + "angle": 0, + "content": "This hook technique establishes a connection between the motion fields to guide the mouth motion to align to the face, while the former's scale is also explicitly connected to the latter, bringing robust performance on the face-mouth alignment even when with only few training data." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.545, + 0.907, + 0.696 + ], + "angle": 0, + "content": "Geometry Prior Regularizer. To further reduce the geometry ambiguity caused by the lack of view cover, we leverage the power of advanced human geometry estimator [23] to provide extra geometry cues for regularization. To solve the geometry degradation [26] and instability of 3DGS at unseen views [6], we add two regularizers on the depth and surface normal. Following [6] we first calculate the 2D depth \\( D \\) and normal \\( N_{s} \\) when synthesizing image \\( I \\). Then, the estimated monocular depth map \\( \\check{D} \\) and normal \\( \\check{N} \\) from the ground-truth image are used for regularization:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.7, + 0.905, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G e o} = \\lambda_ {D} L _ {D} (D, \\check {D}) + \\lambda_ {N} \\sum_ {i = 0} ^ {m} \\sum_ {j = 0} ^ {n} \\left(1 - N _ {i, j} \\cdot \\check {N} _ {i, j}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.746, + 0.905, + 0.777 + ], + "angle": 0, + "content": "where \\(L_{D}\\) is a scale-invariant depth loss, \\(N_{i,j}\\) denotes the normal at pixel \\((i,j)\\), and \\((m,n)\\) is the shape of image \\(I\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.676, + 0.803 + ], + "angle": 0, + "content": "3.4. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Photometric Loss. The training processes are mainly supervised by the image photometric loss \\(\\mathcal{L}_I\\) between the synthesized image \\(I\\) and its ground-truth \\(I_{GT}\\). Following 3DGS [22], \\(\\mathcal{L}_I\\) consists of an L1 and a D-SSIM [46] terms. Pre-training. During pre-training, we collect \\(k\\) videos as training data to simultaneously train the UMF. For each syn" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10694" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.482, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.089, + 0.897, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.348, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Figure 4. Qualitative Comparison on Synchronization. Our method performs best in both lip-synchronization and visual quality. \"Real3DP\" and \"TG\" denote [54] and [28]. Better visualized with zoom-in. We recommend watching the supplementary video." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.386, + 0.907, + 0.564 + ], + "angle": 0, + "content": "
MethodsSettingRendering QualityMotion QualityEfficiency
PSNR (A/F) ↑ N/ALPIPS (A/F) ↓ 0 / 0SSIM ↑ 1.000LMD ↓ 0AUE-(L/U) ↓ 0 / 0Sync-C ↑ 8.897Training ↓ -FPS ↑ -
Ground TruthReal-time
EAT [13]One-shot- / 22.44- / 0.0600.712-1.309 / 1.0117.577-11.8X
Real3DPortrait [54]23.58* / 25.370.104* / 0.0540.8213.4071.148 / 1.0987.549-8.4X
RAD-NeRF [42]Trained28.36 / 25.510.048 / 0.0350.8353.5551.295 / 0.7322.2495 hours28.6✓
ER-NeRF [25]28.23 / 25.630.040 / 0.0310.8443.5411.327 / 0.4513.0742 hours30.8✓
GaussianTalk [7]From Scratch28.18 / 25.610.043 / 0.0320.8363.6471.379 / 0.4151.97067 min64.5✓
TalkingGaussian [28]28.32 / 26.010.041 / 0.0280.8433.5881.158 / 0.3163.55631 min114.2✓
DFRF [40]Adaptation & Few-shot28.48 / 26.260.066 / 0.0340.8593.4361.247 / 0.6833.9579 hours0.03X
GeneFace [52]27.23 / 25.080.052 / 0.0360.8233.5101.310 / 0.6843.6517 hours18.1X
MimicTalk [53]24.69* / 26.270.075* / 0.0310.8533.4890.958 / 0.7756.92616 min8.2X
InsTaG (Ours)28.86 / 26.320.039 / 0.0260.8613.1670.926 / 0.3135.31813 min82.5✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.57, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparison in self-reconstruction setting with 5s training data. InsTaG achieves the best image quality and personalized dynamics, with fast adaption and real-time inference. We mark the best and second-best results. (*: with synthesized torso.)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.659 + ], + "angle": 0, + "content": "thesized image \\(I^i\\) from \\(i\\)-th video, we first calculate its photometric loss, and then calculate the Negative Contrast Loss to each other. The pre-training loss for the \\(i\\)-th video is:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.66, + 0.483, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p r e} ^ {i} = \\mathcal {L} _ {I} \\left(I ^ {i}, I _ {G T} ^ {i}\\right) + \\lambda_ {C} \\sum_ {j = 1, j \\neq i} ^ {k} \\mathcal {L} _ {C} \\left(\\Delta \\mu_ {P} ^ {i} \\cdot \\Delta \\mu_ {P} ^ {j}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Adaptation. For adaptation, we initialize a person-specific model with the pre-trained UMF. A warm-up stage first runs without a Motion Aligner and geometry loss for a start. After that, the model is trained with full loss:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.769, + 0.483, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d a} = \\mathcal {L} _ {I} (I, I _ {G T}) + \\mathcal {L} _ {G e o} (D, N). \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.224, + 0.806 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Dataset. In the experiments, we collect a long-video corpus with 5 long speaking videos from [25, 52] for pre-training, and follow prevailing settings [16, 25, 28, 52] to take 4 videos from publicly-released sets [40, 52] for the test, with no identity overlap of the two parts. The videos have an average length of about 800-8000 frames in 25 FPS with a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.887, + 0.629 + ], + "angle": 0, + "content": "center portrait, all are cropped and resized to \\(512 \\times 512\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Implementation Details. Our method is implemented on PyTorch. The pre-training stage lasts 150,000 iterations and adaptation lasts 10,000. An AdamW [32] optimizer is used with learning rates of 5e-3 and 5e-4 separately for the grid and neural networks to train UMF. \\(\\lambda_{D}\\), \\(\\lambda_{N}\\) and \\(\\lambda_{C}\\) are set to 1e-2, 1e-3 and 1. All experiments are performed on RTX 3080 Ti GPUs. The overall pre-training takes about 5 hours. The head pose is estimated through BFM [36]. A DeepSpeech model [18] is used to extract basic audio features following previous works [16, 25, 28]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Comparison Baselines. To compare with the state-of-the-art methods, we take the baselines from the relevant person-specific methods DFRF [40], RAD-NeRF [42], GeneFace [52], ER-NeRF [25], TalkingGaussian [58], GaussianTalker [7] and a concurrent method MimicTalk [53]. We also involve the one-shot EAT [13] and Real3DPortrait [54]. All the methods are reproduced by their official codebase. If officially supported, we use DeepSpeech as audio extractor." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10695" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.087, + 0.903, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.298, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparison on Reconstruction Quality. Our method performs the best in rendering photorealistic talking heads with fine details. \"Real3DP\" and \"TG\" denote Real3DPortrait [54] and TalkingGaussian [28]. Better visualized with zoom-in." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.333, + 0.905, + 0.43 + ], + "angle": 0, + "content": "
MethodType5s10s20s
PSNR ↑AUE-L ↓Sync ↑PSNR ↑AUE-L ↓Sync ↑PSNR ↑AUE-L ↓Sync ↑
ER-NeRF [25]From Scratch28.2351.3273.07428.7281.0783.24029.2870.9414.638
TalkingGaussian [28]28.3211.1583.55629.1301.0694.66329.5361.0734.776
DFRF [40]Adaptation28.4811.2473.95728.8811.0724.21029.2531.1374.333
GeneFace [52]28.2271.3103.65127.9971.0364.86928.8251.0334.772
InsTaG (Ours)28.8630.9265.31829.4210.8815.63630.3520.7105.764
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.436, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Table 2. Quantitative Comparison in self-reconstruction setting to person-specific baselines with different data amounts. Our method performs best in various data scenarios, exhibiting robustness. The best and second-best results are in bold and underline." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.475, + 0.481, + 0.591 + ], + "angle": 0, + "content": "
MethodsAudio English, MaleAudio German, Female
Sync-E ↓ 10.443Sync-C ↑ 4.342Sync-E ↓ 9.942Sync-C ↑ 4.949
Full Baseline [28]
ER-NeRF [25]11.7113.04311.5022.853
TalkingGaussian [28]10.6754.31710.8983.708
DFRF [40]10.5154.72610.4824.329
GeneFace [52]9.9954.73411.0294.162
InsTaG (Ours)9.8864.8289.7334.990
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.625 + ], + "angle": 0, + "content": "Table 3. Comparison in cross-domain setting with 20s training data. We take TalkingGaussian [28] with full data for reference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.214, + 0.649 + ], + "angle": 0, + "content": "4.1. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Comparison Settings. The quantitative comparison contains two settings: 1) Self-reconstruction setting. We split each of the 4 adaptation videos into few-second training sets and a test set of at least 12s, and use the unseen test data as input and ground truth for evaluation. 2) Cross-domain setting. We use an unseen English male audio from NVP [44] and a German female one from [16] to test the cross-domain generalizability. Considering some baselines with failure cases, we use 20s from a certain video to train per method to represent its best performance for testing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Metrics and Measurements. We employ PSNR, LPIPS [61], and SSIM [46] to evaluate image quality. Especially, we report PSNR and LPIPS separately on an all image (\"A\") and only the aligned face (\"F\"). SSIM is only measured on the face. For lip synchronization, we use LMD [4] and SyncNet [8, 9] confidence (Sync-C) and error (Sync-E). Be" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.569 + ], + "angle": 0, + "content": "sides, action unit [10] error (AUE) is used on upper-face (AUE-U) and lower-face (AUE-L) following [28], in which AUE-L can reflect the similarity of the personalized talking actions from the target. We also statistic the training time (fine-tuning time for few-shot or adaptation-based methods) and inference FPS to evaluate efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Quantitative Evaluation. 1) Table 1 reports the results of self-reconstruction setting with 5s training data. Our method achieves the best image quality, demonstrating our ability to faithfully generate a realistic talking head. Inheriting the traditions of large corpus pre-training, the one-shot methods [13, 54] and MimicTalk [53] achieve very high Sync scores but are relatively slow. Also getting a high Sync, InsTaG achieves the best in LMD and AUE-L, showing our advantage in preserving personalization. While all baselines can not simultaneously achieve high synchronization and real-time inference, our method produces the best overall performance including the fastest adaptation and high speed. 2) When data increases, the results in Table 2 show that InsTaG still outperforms other person-specific models in all aspects, demonstrating our wider application value and robustness to various scenarios. 3) In the results from Table 3 in cross-domain setting, where the methods are tested on two difficult cross-domain scenarios, the advantages of pre-training are further amplified. GeneFace [52] and DFRF [40] surpass all models trained from scratch. At the same time, InsTaG exceeds the baseline [28] trained on about 1-minute full training data, demonstrating the out" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10696" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.904, + 0.168 + ], + "angle": 0, + "content": "
MethodsER-NeRF [25]TalkingGaussian [28]GaussianTalker [7]DFRF [40]GeneFace [52]MimicTalk [53]InsTaG (Ours)
Lip-sync Accuracy2.001.901.503.302.704.804.50
Image Quality3.403.704.002.903.103.704.80
Identity Preserving2.401.501.702.903.004.204.60
Video Realness1.401.501.402.801.703.203.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.171, + 0.899, + 0.186 + ], + "angle": 0, + "content": "Table 4. User Study. The rating is in the range of 1-5, higher denotes better. The best and second-best results are in bold and underline." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.2, + 0.481, + 0.329 + ], + "angle": 0, + "content": "
Pre-training PFieldNCLossAdaptation AlignerHookPSNR ↑LPIPS ↓LMD ↓Sync-C ↑
28.660.0443.7342.019
✓28.760.0403.2394.816
✓✓28.590.0413.2235.124
✓✓✓28.690.0403.2745.233
✓✓✓28.660.0403.2575.086
✓✓✓28.790.0403.2264.994
✓✓✓✓28.860.0393.1675.318
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.334, + 0.482, + 0.363 + ], + "angle": 0, + "content": "Table 5. Ablation Study. We verify the components under 5s self-reconstruction setting, showing the effect of our contributions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.458, + 0.385 + ], + "angle": 0, + "content": "standing robustness and generalizability of our method." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.386, + 0.483, + 0.611 + ], + "angle": 0, + "content": "Qualitative Evaluation. To qualitatively evaluate the results, we show the generated sequences and rendering details from self-reconstruction setting. In Figure 4, it can be observed that the models trained from scratch (ER-NeRF, TG) [27, 28] behave poorly in synchronization (yellow box), while the adaptation-based methods (DFRF, GeneFace, MimicTalk) [40, 52, 53] perform better but generate blurry results (orange circle). Referring to Figure 5, while all methods synthesize either blurs or heavily distorted details (red arrow), our method can reconstruct photorealistic appearance and delicate details. By benefitting from the advantages of both pre-training and person-specific training, InsTaG achieves the best in both synchronization and image quality. We strongly recommend watching our supplementary video for an intuitive comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.765 + ], + "angle": 0, + "content": "User Study. To better judge the quality in real human-participated scenarios, we conducted a user study with 28 generated talking head videos from 7 methods with 5s training data. We invited 10 attendees to rate each anonymous method from four aspects: (1) Lip-sync Accuracy; (2) Video Realness; (3) Identity Preserving and (4) Video Realness. The results in Table 4 show that our InsTaG achieves the best in three aspects and the second-best in lip-sync, which is only after the concurrent MimicTalk [53], demonstrating our outstanding visual quality via human judgment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.245, + 0.789 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.825 + ], + "angle": 0, + "content": "We conduct the ablation study to prove the effectiveness of our contributions. The results are reported in Table 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Identity-Free Pre-training. In the first row, we show the results pre-trained without Personalized Fields (PField), which produces bad synchronization due to the identity-conflict problem. After applied PField, the pre-training works with obvious effect. Combined with PField, Nega" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.2, + 0.905, + 0.253 + ], + "angle": 0, + "content": "
BackboneHookPSNR ↑LPIPS ↓LMD ↓Sync-C ↑
Single Branch-28.590.0493.2334.675
Face-Mouth Decom.×28.690.0403.2745.233
Face-Mouth Decomp.✓28.860.0393.1675.318
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.258, + 0.907, + 0.3 + ], + "angle": 0, + "content": "Table 6. Ablation Study on Mouth Motion Representation. Introducing FMD [28] can effectively improve lip-synchronization, while our FM hook technique further boosts image quality." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.313, + 0.705, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.313, + 0.896, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.393, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 6. Visualization Results of Face-Mouth Hook. FMD improves performance but also causes misalignment when with inadequate data, which can be well solved by our Face-Mouth Hook." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.905, + 0.506 + ], + "angle": 0, + "content": "tive Contrast Loss (NCLoss) shows its impact on lip-sync improvement, offering higher Sync-C. This demonstrates the necessity of PFiled in enabling the pre-training and the effect of NCLoss on encouraging knowledge collection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.643 + ], + "angle": 0, + "content": "Motion-Aligned Adaptation. We apply our Motion Aligner (Aligner) and Face-Mouth Hook (FM Hook) to illustrate their effect on adaptation. Although the performance is improved by applying our pre-training strategy, the Aligner can bring extra enhancement by providing adaptation adjustments. Furthermore, the Hook works in both visual quality and dynamics by correcting inside-mouth misalignment. Combining these two adaptation techniques, the performance achieves the best among all settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Mouth Motion Representation. To further explain the necessity of Face-Mouth Decomposition (FMD) [28] and our FM Hook, we take studies by using one single branch to represent the two parts. Results in Table 6 show the effect of FMD. However, simply introducing it would cause significant misalignment that lowers the visual quality, as in Figure 6. By applying our FM Hook, this drawback can be well solved to benefit FMD with robust performance, generating synchronized talking heads with high visual quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.722, + 0.804 + ], + "angle": 0, + "content": "5. Ethical Consideration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We hope our method can promote the healthy development of digital industries. However, it is important to note that it could be misused for malicious purposes, potentially causing negative impacts. As part of our responsibility, we will assist in developing deepfake detection tools. We recommend the responsible use of this and all similar techniques." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10697" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.178 + ], + "angle": 0, + "content": "In this work, we are supported by the National Natural Science Foundation of China 62276016, 62372029. Lin Gu is supported by JST Moonshot R&D Grant Number JP-MJMS2011 Japan." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.193, + 0.188, + 0.209 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.217, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[1] James Booth, Epameinondas Antonakos, Stylianos Ploumpis, George Trigeorgis, Yannis Panagakis, and Stefanos Zafeiriou. 3d face morphable models\" in-the-wild\". In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 48-57, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.289, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[2] Marcel C Bühler, Kripasindhu Sarkar, Tanmay Shah, Gengyan Li, Daoye Wang, Leonhard Helminger, Sergio Orts-Escalano, Dmitry Lagun, Otmar Hilliges, Thabo Beeler, et al. Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3402-3413, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.389, + 0.483, + 0.457 + ], + "angle": 0, + "content": "[3] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, et al. Authentic volumetric avatars from a phone scan. ACM Transactions on Graphics (TOG), 41(4):1-19, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.484, + 0.529 + ], + "angle": 0, + "content": "[4] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Computer Vision-ECCV 2018: 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part VII 15, pages 538-553. Springer, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.484, + 0.6 + ], + "angle": 0, + "content": "[5] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.484, + 0.658 + ], + "angle": 0, + "content": "[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[7] Kyusun Cho, JoungBin Lee, Heeji Yoon, Yeobin Hong, Jaehoon Ko, Sangjun Ahn, and Seungryong Kim. Gaussian talker: Real-time talking head synthesis with 3d gaussian splatting. In ACM Multimedia 2024, 2024. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[8] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 87-103. Springer, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[9] Joon Son Chung and Andrew Zisserman. Out of time: Automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[10] Paul Ekman and Wallace V. Friesen. Facial Action Coding System: Manual. Palo Alto: Consulting Psychologists Press, 1978. 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.217, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[11] Tony Ezzat, Gadi Geiger, and Tomaso Poggio. Trainable videorealistic speech animation. ACM Transactions on Graphics (TOG), 21(3):388-398, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[12] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[13] Yuan Gan, Zongxin Yang, Xihang Yue, Lingyun Sun, and Yi Yang. Efficient emotional adaptation for audio-driven talking-head generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22634-22645, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.907, + 0.343 + ], + "angle": 0, + "content": "[14] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1155-1164, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[15] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rinz, Lourdes Agapito, and Matthias Nießner. Mononphm: Dynamic head reconstruction from monocular videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10747-10758, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.907, + 0.496 + ], + "angle": 0, + "content": "[16] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[17] Siddharth Gururani, Arun Mallya, Ting-Chun Wang, Rafael Valle, and Ming-Yu Liu. Space: Speech-driven portrait animation with controllable expression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20914-20923, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.567, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[18] Awni Hannun, Carl Case, Jared Casper, Bryan Catanzaro, Greg Diamos, Erich Elsen, Ryan Prenger, Sanjeev Satheesh, Shubho Sengupta, Adam Coates, et al. Deep speech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:1412.5567, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[19] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.705, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.904, + 0.815 + ], + "angle": 0, + "content": "[21] Amir Jamaludin, Joon Son Chung, and Andrew Zisserman. You said that?: Synthesising talking faces from audio. International Journal of Computer Vision, 127:1767-1779, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10698" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[24] Dongze Li, Kang Zhao, Wei Wang, Bo Peng, Yingya Zhang, Jing Dong, and Tieniu Tan. Ae-nerf: Audio enhanced neural radiance field for few shot talking head synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3037–3045, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[25] Jiahe Li, Jiawei Zhang, Xiao Bai, Jun Zhou, and Lin Gu. Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7568-7578, 2023. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[26] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[27] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Jun Zhou, and Lin Gu. Er-nerf++: Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. Information Fusion, 110:102456, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[28] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Talkinggaussian: Structure-persistent 3d talking head synthesis via gaussian splatting. In European Conference on Computer Vision, pages 127-145. Springer, 2025. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[29] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[30] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhigang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17969-17978, 2023. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[31] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 106-125. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[32] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[33] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[34] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, pages 179-197. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[36] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. Ieee, 2009. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[37] Ziqiao Peng, Wentao Hu, Yue Shi, Xiangyu Zhu, Xiaomei Zhang, Hao Zhao, Jun He, Hongyan Liu, and Zhaoxin Fan. Synctalk: The devil is in the synchronization for talking head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 666-676, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[38] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[39] Eduard Ramon, Gil Triginer, Janna Escur, Albert Pumarola, Jaime Garcia, Xavier Giro-i Nieto, and Francesc Moreno-Noguer. H3d-net: Few-shot high-fidelity 3d head reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5620-5629, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[40] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII, pages 666-682. Springer, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.576 + ], + "angle": 0, + "content": "[41] Luchuan Song, Pinxin Liu, Lele Chen, Guojun Yin, and Chenliang Xu. Tri²-plane: Thinking head avatar via feature pyramid. In European Conference on Computer Vision, pages 1–20. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.905, + 0.647 + ], + "angle": 0, + "content": "[42] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[43] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[44] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 716-731. Springer, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[45] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI, pages 700–717. Springer, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10699" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[47] Olivia Wiles, A Sophia Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In Computer Vision-ECCV 2018: 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XIII 15, pages 690-706. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[48] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.306, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[49] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[50] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[51] Xihe Yang, Xingyu Chen, Daiheng Gao, Shaohui Wang, Xiaoguang Han, and Baoyuan Wang. Have-fun: Human avatar reconstruction from few-shot unconstrained images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 742-752, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[52] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, Jinzheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[53] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiangwei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, Chen Zhang, Xiang Yin, Zejun Ma, and Zhou Zhao. Mimictalk: Mimicking a personalized and expressive 3d talking face in few minutes. Advances in neural information processing systems, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[54] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiawei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[55] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[56] Hongyun Yu, Zhan Qu, Qihang Yu, Jianchuan Chen, Zhonghua Jiang, Zhiwen Chen, Shengyu Zhang, Jimin Xu, Fei Wu, Chengfei Lv, et al. Gaussian talker: Speaker-specific talking head synthesis via 3d gaussian splattering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3548-3557, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[57] Yu Yu, Kenneth Alberto Funes Mora, and Jean-Marc Odobez. Headfusion: \\(360^{\\circ}\\) head pose tracking combining 3d morphable model and 3d reconstruction. IEEE transactions on pattern analysis and machine intelligence, 40(11): 2653-2667, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[58] Zhentao Yu, Zixin Yin, Deyu Zhou, Duomin Wang, Finn Wong, and Baoyuan Wang. Talking head generation with probabilistic audio-to-visual diffusion priors. arXiv preprint arXiv:2212.04248, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[59] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9459-9468, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[60] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3867-3876, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.513 + ], + "angle": 0, + "content": "[62] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10700" + } + ] +] \ No newline at end of file diff --git a/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/1014c4e5-3a69-4fa3-98b3-a879e2f45651_origin.pdf b/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/1014c4e5-3a69-4fa3-98b3-a879e2f45651_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e0910e06f74f65a268159d131e68b87dd1dfa7af --- /dev/null +++ b/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/1014c4e5-3a69-4fa3-98b3-a879e2f45651_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c8d52096bdffede0a3769a4a3abc5c2dda88cfac4ba451f98ecc7e0fea9c71e +size 5250268 diff --git a/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/full.md b/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b7efce98575d63edf7808c9feaceaaf9767adb05 --- /dev/null +++ b/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/full.md @@ -0,0 +1,455 @@ +# InsTaG: Learning Personalized 3D Talking Head from Few-Second Video + +Jiahe Li $^{1}$ Jiawei Zhang $^{1}$ Xiao Bai $^{1*}$ Jin Zheng $^{1*}$ Jun Zhou $^{2}$ Lin Gu $^{3,4}$ + +$^{1}$ School of Computer Science and Engineering, State Key Laboratory of Complex & Critical Software Environment, Jiangxi Research Institute, Beihang University + +$^{2}$ School of Information and Communication Technology, Griffith University + +$^{3}$ RIKEN AIP $^{4}$ The University of Tokyo + +![](images/9f2aa9bfb22be44ffd9dd7f4038aaa8dadc1a26e408367a36f356cc4c37a4632.jpg) + +![](images/31bd0f72d9a39afcaaf2ac6a13996ced8b6147ef6edbb82c5161baa26ec608da.jpg) +4 5-Second Video + +![](images/db056d0232130caf05033539929122d44cca178ce16123337e72b896e9af1390.jpg) +Instant Adaptation + +![](images/710fe4799ae419a505052e72412b92eb1e166b2ab4bb121d60d5fdafd2b9918f.jpg) + +![](images/1fd73e80e5b2399c41bae8227a9e9fd8486aae97f580ba66d2d856aab3935666.jpg) + +![](images/b2b14592a19a3d30f800c922b2f4c43c2bbdf9aa3c091d2dbe9a1f5eaa348f77.jpg) + +![](images/faa09c48ccaf4997f752807f6f5dfcd9897dd5bec9184c77b6908dd01882456b.jpg) +# + +![](images/14df0e95c8c120e7e913efb2a329be7c0655ee930e4a479e3fc84cc8ddfa6be6.jpg) + +![](images/3b7295e2f9ed1ccdd8a11ee0e8e8a4dd57d0454327e0c86c5ec22f2edb767315.jpg) + +![](images/979e1435fd3b909ace6fed21011c215a43e6ab1e51988dc90eddb9b56673ffdf.jpg) + +![](images/ba94422f9262a5181b40113feb9309c8e903a71447396d61fd8ecd35a92bd556.jpg) + +![](images/a07d2db5b7ab3913bf8f22f69c41d5028ce01656a8509ae73eebb5b59e03f213.jpg) + +![](images/8e8814e81e349ac0b8b7b9897e941e2a83c87b20ec6b0cbc98d3da4fb5e1ca1a.jpg) + +![](images/d19bf764db921942ac2e1693fd8aed4bb8a98c902b597d05a4f6e53cead7f57e.jpg) + +![](images/6b53bc7e94c1cde46dab2794c46bdec028fd75e65762c4ee86419ab45213e821.jpg) + +![](images/fa8042717dc08e538db267d8e2d514512f7b1841f3fc42f49af7df2decf7d1c6.jpg) + +![](images/4a115dc5eda4726d769dc52b70af36625c2dc2230503487514c1f9e14493fceb.jpg) + +![](images/1127508aa857436268f37440fdb11efb9e40ef223deb934009ca9fd95f83baaa.jpg) + +![](images/c9e4e8ef72f7b97dd5a2369f1ff876d98555194139a903fa3a7c2778258d7746.jpg) + +![](images/e476341d19c3f90f3c5a9dce5d26b2448e3411bb11f535e3820f0c9f57e42777.jpg) + +![](images/d39441fc7c6eacead48266ced4a3179b907026d24bfa10edd43bb0f55ba8e62d.jpg) + +![](images/1203f747297cbd8365b3806e7c9ffc6db20ce07e116d8b235409c1fb6433855f.jpg) + +![](images/60892eced30bd09f1b8bfc6573686450a31de1f729d1674247691e057630c227.jpg) + +![](images/5b4a430ad266f1c47c9334eeda81aa3f5a0dabe4dd92930ddb373473e9596b96.jpg) + +![](images/46b31cc39b99e9e8266c9360345e4d48970b664390db3d16f9f9c416f709536c.jpg) + +![](images/d831fb9e16dfc3f7e0d302366dfbf5b69997f0aa1d149c047b320c684f47d5be.jpg) + +![](images/1532240c6186fa412b466c21031a91df84f3f7cb45468dccca1c8285012e2535.jpg) + +![](images/8b68e9b08100617e52c6bd73ceb7592a5100c9c082a370c23ac5e847f723d86d.jpg) + +![](images/16687558939e41f2b833193f4931721e13487586c766626ad451d12dece8d7e3.jpg) + +![](images/898d194d65f7295f6612539a1041ae70d32a119535f5f54f480051a77fed5630.jpg) + +![](images/189b59b52811878ed832a360ec76b213c84d09df68652dad4612dfb385d59d2a.jpg) + +![](images/f26a35b278fded0c1c6fc99521b32d643164a0bb090e73cfb8517b1e02ede095.jpg) + +![](images/0fdfcfe25cc7037066b71df990b0c5665160b494a5002815df4fc93aaa03fe31.jpg) + +![](images/bd3cf62c03c65ec8033f06af8339a6bc0280d88711159e63042cebc5b9c70784.jpg) + +![](images/0499b94092db0dd2149c7623c669e599f128559cf21357b8bce7aa78beb590ec.jpg) + +![](images/302e18d82be343df5ca81db567fb0a7587a3ef226537040b5fe7afaca51e56e4.jpg) +GeneFace + +![](images/f0b93c63e6361411c769c5e974c10b166d1826cd373ef0c359b0969a6907e051.jpg) +TalkingGaussian + +![](images/c90f1ef5c16190be98058b94f21a1e7d3b5bc1fc7ceb0894726b283900282161.jpg) +MimicTalk + +![](images/7a2f7b8d2297913a4c533a423459b431f2a483c516450a7278c4de362dde8a15.jpg) +In + +![](images/db37008c48d8d892b18be525425933e48ec5827c25dd08a96bdb0cc057b413ad.jpg) +Figure 1. With only 5-second video data, InsTaG outperforms the state-of-the-arts [28, 52, 53] by delivering high-quality personalized lip synchronization and realistic rendering with the fastest adaptation, meanwhile attaining low memory overhead and real-time inference. + +![](images/3410ea1cd96cd84e0e07147237f20c9a455ab8f3bda0ef45a895ca84c257ec6c.jpg) +G + +![](images/33690756a5d4706e6120b3ded57eae826e1e6b7789f24e2c8bcd137c23f78100.jpg) +Ground Truth + +# Abstract + +Despite exhibiting impressive performance in synthesizing lifelike personalized 3D talking heads, prevailing methods based on radiance fields suffer from high demands for training data and time for each new identity. This paper introduces InsTaG, a 3D talking head synthesis framework that allows a fast learning of realistic personalized 3D talking head from few training data. Built upon a lightweight 3DGS person-specific synthesizer with universal motion priors, InsTaG achieves high-quality and fast adaptation while preserving high-level personalization and efficiency. As preparation, we first propose an Identity-Free Pre-training strategy that enables the pre-training of the person-specific model and encourages the collection of universal motion priors from long-video data corpus. To fully exploit the universal motion priors to learn an unseen new identity, we then present a Motion-Aligned Adaptation strategy to adaptively align the target head to the pre-trained field, and constrain a robust dynamic head structure under few training data. Experiments demonstrate our outstanding performance and efficiency under various data scenarios to render high-quality personalized talking heads. Project page: + +https://fictionarry.github.io/InsTaG/. + +# 1. Introduction + +Audio-driven talking head synthesis has become an important technique for various digital applications like video production, virtual reality, and human-computer interaction. The recent advancements in Neural Radiance Fields (NeRF) [35] and 3D Gaussian Splatting (3DGS) [22] have led to significant progress in 3D representations. Various methods [16, 27, 28, 31, 40, 52] now utilize radiance fields to create high-fidelity 3D talking head videos, capturing personalized talking styles with impressive photorealistic image quality. + +Most of these methods adopt person-specific training. Given a video clip with the target talking person, a whole person-specific model is trained to memorize the target talking head. Despite their tremendous improvements in high-fidelity rendering, personalized dynamics, and fast running speed, as a trade-off, the model holds no generalizability for an unseen identity. Thus, they require a large amount of high-quality video frames with a long training time to adapt to each new identity. Some image-referred methods [30, 40, 53, 54] can quickly generalize to different identities, nevertheless, the quality of personalization and efficiency drop. A question is then raised: How can we use the least training data to rapidly learn a faithful 3D talking + +head, while preserving a high level of personalization? + +Observing previous radiance-field-based methods, we notice that: 1) While the training process includes a long time to learn a basic audio-driven motion that with only little personalized style, the static 3D head can be well reconstructed in just a few early iterations; 2) Though the results may exhibit jittering and poor quality when trained with limited data, a portion of the lip motions remains recognizable with distinct personalized style. These suggest that even a short video clip can provide rich personalization cues to enable the learning of a specific talking head with little time cost, as long as common audio-motion knowledge is obtained. Motivated by this, by preparing a pre-trained identity-free 3D motion field as prior and aligning it to the new identity to facilitate the adaptation, we propose Instant Talking Head Synthesis with Gaussian Splatting (InsTaG), a framework that allows fast learning of high-fidelity personalized 3D talking head from a video even as short as only a few seconds while attaining high efficiency. + +Different from previous few-shot methods [30, 53, 54] that imitate a new identity from inputs, InsTaG memorizes the whole talking head in a lightweight person-specific model to obtain high-level personalization and efficiency. Upon such a person-specific paradigm, we fully decouple the learning of universal motion and personalization into pre-training and adaptation, preparing a pre-trained 3D motion field for future person-specific learning of new identity. However, a challenge with this design lies in pre-training with multi-person data, where the person-specific model can not fit varying identities and estimate their varied personalized motions, leading to an ineffective optimization. To tackle this problem, we present an Identity-Free Pretraining strategy. By keeping a series of temporary personal fields to store the identity information and filter out the personalized motions, we extract the common knowledge of talking motion from long-video corpus and produce a universal motion field that contains identity-free motion prior. To further facilitate this motion detachment, a negative contrast loss is introduced to encourage the diversity of learned personalized motion between each sample, maximizing the retaining of universal priors for later adaptation. + +To fully exploit the universal motion priors to enable fast and high-quality new-identity adaptation, we present a Motion-Aligned Adaptation strategy to align the unseen talking head to the pre-trained motion field. First, a motion aligner is developed to learn a condition-independent primitive-wise coordinate offset and a motion scalar. It helps the intra-alignment between the personalized and universal motions in the field, maximizing the retention of learned knowledge during few-shot fine-tuning. Moreover, to enhance the generalizability by face-mouth decomposition [58] while preserving a robust dynamic face structure, we introduce a face-mouth hook technique to inter-align the + +inside mouth motion with the face, which efficiently improves the fidelity as well as motion quality. To further improve the robustness of view direction, the new identity's head structure is aligned with estimated geometry priors for regularization. By incorporating the alignments in the adaptation, the person-specific model finally synthesizes high-fidelity personalized 3D talking heads at a fast speed. + +Integrating the two proposed strategies, InsTaG can learn a personalized 3D talking head from even a short 5-second video, achieving state-of-the-art visual quality and personalization compared to existing methods, with high lip-synchronization and real-time inference. Experiments demonstrate the outperforming efficiency and generalizability of InsTaG for various identities, genders and languages. + +Our main contributions are summarized as follows: + +- An Identity-Free Pre-training strategy that enables the pretraining of the person-specific model by filtering out conflict identity and personalized motions, meanwhile maximizing the capturing of universal motion priors. +- A Motion-Aligned Adaptation strategy, which inter- and intra-aligns the talking head with the motion field for efficient new-identity adaptation, attaining robust, realistic reconstruction and personalized lip-synchronization. +- Extensive experiments show that our InsTaG learns realistic personalized talking heads from few data, while attaining high lip-synchronization and efficiency, outperforming state-of-the-art methods under various scenarios. + +# 2. Related Work + +3D Talking Head Synthesis. Talking head synthesis aims to use arbitrary audio to reenact a talking person to generate audio-visual synchronized videos. Early works [5, 11, 21, 38, 45, 47, 62] are mainly built on generative models, creating talking heads by manipulating the given 2D images. Later, to solve the temporal inconsistency when the head moves, 3D-based methods [33, 44, 55, 60] utilize explicit 3D face structure and successfully improve the naturalness. + +Recently, radiance fields like NeRF [35] and 3D Gaussian Splitting (3DGS) [22] have been introduced as the representation to allow 3D talking head synthesize. Inheriting the radiance field optimization, most methods [7, 16, 25, 28, 31, 37, 42] train a person-specific model on minutes of high-quality video, achieving success in reconstructing photorealistic rendering and personalized talking style. However, the strict quality requirements of the training video data and the long training time for every adaptation to a new identity have hugely limited their application. Although some works [30, 40, 52-54, 56] attempt to solve the problem by a one-shot generator [30, 53, 54] or take pre-trained motions from external modules to loosely join the adaptation [40, 52] to achieve one- or few-shot, they often compromise image quality and loses the personalization as the trade-off. A concurrent work MimicTalk [53] injects LoRA [20] to + +![](images/0113c62eedf6e93077e4e9fb7d58a434dbdfabbe6183c8414f97c2b72a48b895.jpg) + +![](images/9df1aad750b0176fcfce7ac78676c69e2f4b857b7a8331d4eb549f54f6a40e0a.jpg) +Figure 2. Overview of InsTaG. For preparation, InsTaG collects the common knowledge of talking motion from a long-video corpus by Identity-Free Pre-training, storing it as a motion field. Given a short video with a new identity, the Motion-Aligned Adaptation strategy builds a robust and fast person-specific synthesizer with the pre-trained motion field to learn a high-quality personalized 3D talking head. + +improve fine-tuning efficiency, but its inference is slow due to the large backbone. Instead, by enabling the pretraining of person-specific models and using pre-trained motion field to directly drive the talking head, our method obtains a more compact and consistent architecture to fully exploit the motion priors in an end-to-end way, facilitating modeling precise lip-synchronization with high efficiency. + +Few-shot 3D Head Reconstruction. Fitting high-fidelity 3D human heads from RGB images has been a hot topic. Early model-based methods [14, 43, 50, 57] focus on fitting a 3D head with 3DMMs [1, 29, 36] but are short in detail. Combined with neural fields, some model-free methods [2, 34, 39] learn static heads from a few images with structure priors. Although many works [3, 12, 15, 19, 41, 49, 51] attempt to reconstruct dynamic heads from few videos, their parametric controls are difficult to predict from audio and limit the mouth motion granularity, making it hard for them to express the highly personalized 3D talking head with delicate audio-driven mouth motions achieved by InsTaG. + +Some generative methods [13, 17, 30, 48, 54, 59] focus on reconstructing talking heads with 3D structure, however, they can hardly infer personalized talking style from only one image, leading to weak personalization and fidelity. Learning 3D talking heads with few-shot NeRF, recent works [24, 40, 53] introduce 2D-to-3D modules with pre-training but are often slow due to their overhead. Instead, with an identity-free motion prior and person-specific adaptation instead of expensive 2D-to-3D priors, our InS-TaG can rapidly learn a personalized audio-driven 3D talking head from few data while attaining fast inference. + +# 3. InsTaG + +As shown in Figure 2, InsTaG synthesizes personalized 3D talking head based on Gaussian Splatting [22] (Sec. 3.1), consisting of an Identity-Free Pre-training (Sec. 3.2) strategy to fetch universal motion priors and a Motion-Aligned Adaptation strategy (Sec. 3.3) to learn a new identity. We describe the detailed designs in the following sections. + +# 3.1. 3DGS Talking Head Synthesizer + +Our framework aims to learn personalized 3D talking head with fast inference and instant adaptation, therefore, a lightweight model with high representation ability is needed to be as the 3D synthesizer. Generally, existing methods can be categorized into image-based [30, 53, 54] and person-specific [16, 25, 28, 42]. The former often possesses generalizability but is relatively slow and weak in personalization, while the latter matches our requirements. + +Preliminaries. A person-specific synthesizer based on radiance fields can be formulated as a mapping $\mathcal{F}$ that from given conditions to a synthesized head image $I$ : + +$$ +\mathcal {F}: (\boldsymbol {a}, \boldsymbol {e}, [ R, t ]; \theta) \rightarrow I \tag {1} +$$ + +where the conditions include audio $\mathbf{a}$ , upper-face expression control $\mathbf{e}$ and camera pose $[R, t]$ that represents the head pose, and the parameters of the radiance fields $\theta$ . + +Under such an architecture, we require the motion learning to be entirely irrelevant to appearance for pre-training. Hence, we leverage the recent 3DGS-based synthesizer with a face-mouth decomposition [28], including a pair of structure fields and motion fields separately for a face branch and + +inside-mouth branch to synthesize the talking head image $I$ . Structure Field. The structure field stores the static Gaussian primitives with parameters $\theta = \{\mu, s, q, \alpha, f\}$ . Both the face branch and inside-mouth branch obtain a private structure field for their own target structure. + +Motion Field. The motion field predicts a point-wise deformation $\delta = \{\Delta \mu, \Delta s, \Delta q\}$ with a tri-plane hash encoder $\mathcal{H}$ [25] with a region attention (RA) module to store spatial relation, and an MLP decoder to predict deformation. For each query primitive with center $\mu$ , the neural field $\mathcal{D}$ predicts its deformation $\delta$ from the given condition feature set $\mathbf{C}$ including audio $a$ and upper-face expression signal $e$ : + +$$ +\delta = \mathcal {D} (\mu , \mathbf {C}) = \operatorname {M L P} (\mathcal {H} (\mu) \oplus \mathbf {C}), \tag {2} +$$ + +where $\oplus$ denotes concatenation. + +Rendering. During rendering, the motion field deforms the Gaussians $\theta$ depending on the inputs $a$ and $e$ , and then a 2D image is rendered. This process happens in both the face and inside-mouth branches, and the output results are combined to get the final talking head image. + +In the following, unless necessary, we do not show the face-mouth decomposition for easier understanding. + +# 3.2. Identity-Free Pre-training + +Although with personalization knowledge, a short video clip is insufficient to train a talking head with personalized audio-motion mapping from scratch. To this end, we extract the common motion knowledge from long videos by pre-training as compensation for the new identity. + +In contrast to existing generative pre-training solutions [30, 40, 53] that allow multiple-identity inputs to train the same model, an identify-conflict problem lies in our method. Firstly, different identities can not coexist in one person-specific model. Secondly, since the model can not separate different people, the conflicting personalized motion hinders the model convergence and pollutes the pretrained module by overfitting the noises. + +Strategy. To tackle this problem, we design storing the identity-dependent knowledge in a Personalized Field distributed to each training video, therefore detaching the identity influence of the learning of universal motion to pre-train a Universal Motion Field. Specifically, each Personalized Field consists of a personalized structure field with parameter $\theta_{P}$ and a smaller motion field $\mathcal{D}_P$ to retain the identity appearance and personalized motion of its owner video. + +Universal Motion Field (UMF). Serving as a pre-trained module, UMF aims to predict a commonly correct facial motion that suits most identities. Queried by an arbitrary Gaussian, the field $\mathcal{U}$ predicts the universal deformation $\delta_U$ without personalization according to Eq. (2): + +$$ +\delta_ {U} = \mathcal {U} (\mu , \mathbf {C}) = \operatorname {M L P} \left(\mathcal {H} _ {U} (\mu) \oplus \mathbf {C}\right). \tag {3} +$$ + +Personalized Fields. During training, Personalized Fields cooperate with one sharing Universal Motion Field to syn + +thesize personalized talking heads, enabling the supervision from videos with different identities. Practically, given an audio feature $\mathbf{a}$ and expression signal $\mathbf{e}$ packaged by $\mathbf{C}$ , the UMF $\mathcal{U}$ would predict an identity-free deformation queried by $\mu_P^i \in \theta_P^i$ , serving as a basic deformation for the $i$ -th person. Meanwhile, the $i$ -th personalized field predicts a residual motion $\delta_P^i$ as the personalized adjustment: + +$$ +\delta_ {P} ^ {i} = \mathcal {D} _ {P} ^ {i} \left(\mu_ {P} ^ {i}, \mathbf {C}\right) = \operatorname {M L P} \left(\mathcal {H} _ {P} ^ {i} \left(\mu_ {P} ^ {i}\right) \oplus \mathbf {C}\right). \tag {4} +$$ + +After that, the two deformations are combined to deform the personalized structure fields for rendering: + +$$ +\mathcal {R}: ([ R, t ], \widetilde {\theta} _ {P} ^ {i}) \rightarrow I, \quad \widetilde {\theta} _ {P} ^ {i} = \theta_ {P} ^ {i} + \delta_ {U} + \delta_ {P} ^ {i}, \tag {5} +$$ + +where $\mathcal{R}$ refers to the 3DGS rendering process. + +Since then, videos with different talking identities can be used for pre-training with no identity conflicts, and the personalized motions are isolated from each other by staying at the personalized fields, detached from the universal ones. + +Negative Contrast Loss. Despite the detachment of personalized and universal motions enabled, a new problem emerges: how to decide whether a motion should be personalized or universal? Without supervision, a native separation from gradient would be highly biased because of uncertainties like initialization and data disparity, making some personalized motions inappropriately stored into the universal field but too much universal knowledge omitted. + +For this problem, we introduce a Negative Contrast Loss with two principles: 1) the ideal personalized motions are expected to be the minimum additional motions that enable smooth model convergence, thus the fewer the better; 2) subject to the first point, the personalized motions of one identity should be different from that of the others. Therefore, we design the loss by a truncated dot product: + +$$ +\mathcal {L} _ {C} \left(\Delta \mu_ {P} ^ {i}, \Delta \mu_ {P} ^ {j}\right) = \mathbb {I} _ {\text {t r u n c}} \left(\Delta \mu_ {P} ^ {i} \cdot \Delta \mu_ {P} ^ {j}\right), \tag {6} +$$ + +where $(\Delta \mu_P^i,\Delta \mu_P^j)$ is a pair of personalized $\mu$ deformation from $i$ -th and $j$ -th personalized fields, queried by the same $\mu_{P}$ and $\mathbf{C}$ in Eq. (4), and $\mathbb{I}_{trunc}$ is a truncation function: + +$$ +\mathbb {I} _ {\text {t r u n c}} (x) = \left\{ \begin{array}{l l} x, & x > 0 \\ 0, & x \leq 0 \end{array} \right.. \tag {7} +$$ + +This negative-sample-based contrastive loss works from two aspects: 1) when the two participants are close in the 3D space, the loss would encourage them to be either different in direction or less in effect; 2) if the participants are already significantly different, or if one of the participants is close to 0, the loss stops pushing them too far. + +By contrasting the negative personalized $\Delta \mu$ pairs, the proposed loss can encourage learned personalized motion to be diverse between the different personalized fields, therefore maximizing the filtering to get universal motions. + +# 3.3. Motion-Aligned Adaptation + +Based on the pre-trained UMF from Sec. 3.2, we introduce a Motion-Aligned Adaptation strategy to allow fast and high-quality learning of new-identity talking heads from a short video. Different from existing works [30, 40, 53] that take motions to guide image generation, we directly drive Gaussians by deformation in the final rendering for more realistic visual quality, bringing a challenge in precisely aligning the pre-trained motion to the new identity's head. + +Motion Aligner. Due to the diversity of human faces, the pre-trained UMF can not natively match all unseen identities suitably. Except for the personalization that we aim to learn, this motion misalignment exists in two aspects: 1) the target facial structure may be biased with the implicit one from pre-training in UMF; 2) the motion of the target identity may have a scale gap from the universal motion. + +To this end, we propose to wrap a Motion Aligner on the motion field to learn the adjustment. Specifically, a multiresolution positional encoder $\mathcal{H}_A$ is first added to store the spatial information. Then, before querying the motion field by $\mu$ from the structure field, a coordinate offset $\Delta \mu_A$ is predicted as bias adjustment: + +$$ +\Delta \mu_ {A} = \operatorname {M L P} \left(\mathcal {H} _ {A} (\mu)\right) \in \mathbb {R} ^ {3}. \tag {8} +$$ + +After getting the universal deformation $\delta_U$ , the $\Delta \mu$ part is multiplied by a scaling factor $\tau_A \in \mathbb{R}^3$ predicted in the same way as Eq. (8) for scale alignment, and finally output the adjusted deformation $\delta_U' = \{\Delta \mu', \Delta s, \Delta q\}$ , where + +$$ +\Delta \mu^ {\prime} = \Delta \mu \times \tau_ {A}, \quad \Delta \mu \in \mathcal {U} (\mu + \Delta \mu_ {A}, \mathbf {C}). \tag {9} +$$ + +Here $\delta_U^\prime$ is used as the final deformation for rendering. + +Face-Mouth Hook. As a deformation-based framework, the asynchronous inside and out mouth motions often deteriorate the modeling [28]. As in Sec. 3.1, we adopt a face-mouth decomposition [28] to solve this problem. It separates the modeling of the face (including lips) and inside of the mouth into two branches. However, when data is reduced, these two parts may fail to learn to move in harmony, causing misalignment with low visual quality. Without an existing mechanism to ensure the inter-alignment, we suggest setting a hook in the motion field to allow the mouth to actively align to the face, as shown in Figure 3. + +Given a predicted deformation $\delta_F$ from the face branch that represents face motion, we attempt to fetch the motions on lips from it to guide the inside-mouth deformation $\delta_M$ . Considering the lip motion tends to cause the largest deformation amount of the whole face, we take the maximal and minimal $\Delta \mu^F \in \delta_F$ (denoted as $\Delta \mu_{max/min}$ ) as the motion cues, and calculate a distance $\mu_{dist} = \Delta \mu_{max} - \Delta \mu_{min}$ for a direct scale cue. Then, this set of features $\varphi = \{\Delta \mu_{max}, \Delta \mu_{min}, \mu_{dist}\}$ are input in the MLP decoder in the inside-mouth branch motion field to predict a deforma + +![](images/b3f88ffa2ca7c54db507cbd5d6e6a90a492b05bea8da256ca2cd72c2db0630a5.jpg) +Figure 3. Illustration of Face-Mouth Hook. We hook the motion of the mouth to the generated face motion, allowing an alignment across two branches to enhance robustness under few training data. + +tion $\Delta \mu$ . The mouth motion field can be formulated as: + +$$ +\Delta \mu^ {M} = \mathcal {U} ^ {\mathcal {M}} (\mu , \{\boldsymbol {a}, \varphi \}). \tag {10} +$$ + +Representing a structure difference, $\mathcal{U}^{\mathcal{M}}$ does not take $e$ as input and does not predict deformations on $r$ and $s$ . + +Moreover, considering the opening degree of the inside mouth should align with that of the lips, we hook the motion scaling prediction to the facial motion scale cue $\mu_{dist}$ . To achieve this, a scaling factor $\tau_{M}$ is predicted for each $\mu$ by an MLP to get the final primitive-wise deformation: + +$$ +\mu_ {\text {h o o k}} ^ {M} = \Delta \mu^ {M} \times \tau_ {M}, \quad \tau_ {M} = \operatorname {M L P} \left(\mathcal {H} _ {U} ^ {M} (\mu) \oplus \mu_ {\text {d i s t}}\right), \tag {11} +$$ + +where $\mathcal{H}_U^M$ is the $\mathcal{H}$ of the UMF in the inside-mouth branch. + +This hook technique establishes a connection between the motion fields to guide the mouth motion to align to the face, while the former's scale is also explicitly connected to the latter, bringing robust performance on the face-mouth alignment even when with only few training data. + +Geometry Prior Regularizer. To further reduce the geometry ambiguity caused by the lack of view cover, we leverage the power of advanced human geometry estimator [23] to provide extra geometry cues for regularization. To solve the geometry degradation [26] and instability of 3DGS at unseen views [6], we add two regularizers on the depth and surface normal. Following [6] we first calculate the 2D depth $D$ and normal $N_{s}$ when synthesizing image $I$ . Then, the estimated monocular depth map $\check{D}$ and normal $\check{N}$ from the ground-truth image are used for regularization: + +$$ +\mathcal {L} _ {G e o} = \lambda_ {D} L _ {D} (D, \check {D}) + \lambda_ {N} \sum_ {i = 0} ^ {m} \sum_ {j = 0} ^ {n} \left(1 - N _ {i, j} \cdot \check {N} _ {i, j}\right), \tag {12} +$$ + +where $L_{D}$ is a scale-invariant depth loss, $N_{i,j}$ denotes the normal at pixel $(i,j)$ , and $(m,n)$ is the shape of image $I$ . + +# 3.4. Training Details + +Photometric Loss. The training processes are mainly supervised by the image photometric loss $\mathcal{L}_I$ between the synthesized image $I$ and its ground-truth $I_{GT}$ . Following 3DGS [22], $\mathcal{L}_I$ consists of an L1 and a D-SSIM [46] terms. Pre-training. During pre-training, we collect $k$ videos as training data to simultaneously train the UMF. For each syn + +![](images/451012626d85b042f3ed1dc6bd2f5a5d4477c14c83333f009b17feb2fcf699f9.jpg) +Figure 4. Qualitative Comparison on Synchronization. Our method performs best in both lip-synchronization and visual quality. "Real3DP" and "TG" denote [54] and [28]. Better visualized with zoom-in. We recommend watching the supplementary video. + +![](images/e57164751141bdc6e749fb9db598847c9e386a38adaa087c3073e5abeac944d3.jpg) + +
MethodsSettingRendering QualityMotion QualityEfficiency
PSNR (A/F) ↑ N/ALPIPS (A/F) ↓ 0 / 0SSIM ↑ 1.000LMD ↓ 0AUE-(L/U) ↓ 0 / 0Sync-C ↑ 8.897Training ↓ -FPS ↑ -
Ground TruthReal-time
EAT [13]One-shot- / 22.44- / 0.0600.712-1.309 / 1.0117.577-11.8X
Real3DPortrait [54]23.58* / 25.370.104* / 0.0540.8213.4071.148 / 1.0987.549-8.4X
RAD-NeRF [42]Trained28.36 / 25.510.048 / 0.0350.8353.5551.295 / 0.7322.2495 hours28.6✓
ER-NeRF [25]28.23 / 25.630.040 / 0.0310.8443.5411.327 / 0.4513.0742 hours30.8✓
GaussianTalk [7]From Scratch28.18 / 25.610.043 / 0.0320.8363.6471.379 / 0.4151.97067 min64.5✓
TalkingGaussian [28]28.32 / 26.010.041 / 0.0280.8433.5881.158 / 0.3163.55631 min114.2✓
DFRF [40]Adaptation & Few-shot28.48 / 26.260.066 / 0.0340.8593.4361.247 / 0.6833.9579 hours0.03X
GeneFace [52]27.23 / 25.080.052 / 0.0360.8233.5101.310 / 0.6843.6517 hours18.1X
MimicTalk [53]24.69* / 26.270.075* / 0.0310.8533.4890.958 / 0.7756.92616 min8.2X
InsTaG (Ours)28.86 / 26.320.039 / 0.0260.8613.1670.926 / 0.3135.31813 min82.5✓
+ +Table 1. Quantitative Comparison in self-reconstruction setting with 5s training data. InsTaG achieves the best image quality and personalized dynamics, with fast adaption and real-time inference. We mark the best and second-best results. (*: with synthesized torso.) + +thesized image $I^i$ from $i$ -th video, we first calculate its photometric loss, and then calculate the Negative Contrast Loss to each other. The pre-training loss for the $i$ -th video is: + +$$ +\mathcal {L} _ {p r e} ^ {i} = \mathcal {L} _ {I} \left(I ^ {i}, I _ {G T} ^ {i}\right) + \lambda_ {C} \sum_ {j = 1, j \neq i} ^ {k} \mathcal {L} _ {C} \left(\Delta \mu_ {P} ^ {i} \cdot \Delta \mu_ {P} ^ {j}\right). \tag {13} +$$ + +Adaptation. For adaptation, we initialize a person-specific model with the pre-trained UMF. A warm-up stage first runs without a Motion Aligner and geometry loss for a start. After that, the model is trained with full loss: + +$$ +\mathcal {L} _ {a d a} = \mathcal {L} _ {I} (I, I _ {G T}) + \mathcal {L} _ {G e o} (D, N). \tag {14} +$$ + +# 4. Experiments + +Dataset. In the experiments, we collect a long-video corpus with 5 long speaking videos from [25, 52] for pre-training, and follow prevailing settings [16, 25, 28, 52] to take 4 videos from publicly-released sets [40, 52] for the test, with no identity overlap of the two parts. The videos have an average length of about 800-8000 frames in 25 FPS with a + +center portrait, all are cropped and resized to $512 \times 512$ . + +Implementation Details. Our method is implemented on PyTorch. The pre-training stage lasts 150,000 iterations and adaptation lasts 10,000. An AdamW [32] optimizer is used with learning rates of 5e-3 and 5e-4 separately for the grid and neural networks to train UMF. $\lambda_{D}$ , $\lambda_{N}$ and $\lambda_{C}$ are set to 1e-2, 1e-3 and 1. All experiments are performed on RTX 3080 Ti GPUs. The overall pre-training takes about 5 hours. The head pose is estimated through BFM [36]. A DeepSpeech model [18] is used to extract basic audio features following previous works [16, 25, 28]. + +Comparison Baselines. To compare with the state-of-the-art methods, we take the baselines from the relevant person-specific methods DFRF [40], RAD-NeRF [42], GeneFace [52], ER-NeRF [25], TalkingGaussian [58], GaussianTalker [7] and a concurrent method MimicTalk [53]. We also involve the one-shot EAT [13] and Real3DPortrait [54]. All the methods are reproduced by their official codebase. If officially supported, we use DeepSpeech as audio extractor. + +![](images/14493a5e2c805b7245f39c0d37bfb7f63beb9772db764f24e3bc7f7d048b6601.jpg) +Figure 5. Qualitative Comparison on Reconstruction Quality. Our method performs the best in rendering photorealistic talking heads with fine details. "Real3DP" and "TG" denote Real3DPortrait [54] and TalkingGaussian [28]. Better visualized with zoom-in. + +
MethodType5s10s20s
PSNR ↑AUE-L ↓Sync ↑PSNR ↑AUE-L ↓Sync ↑PSNR ↑AUE-L ↓Sync ↑
ER-NeRF [25]From Scratch28.2351.3273.07428.7281.0783.24029.2870.9414.638
TalkingGaussian [28]28.3211.1583.55629.1301.0694.66329.5361.0734.776
DFRF [40]Adaptation28.4811.2473.95728.8811.0724.21029.2531.1374.333
GeneFace [52]28.2271.3103.65127.9971.0364.86928.8251.0334.772
InsTaG (Ours)28.8630.9265.31829.4210.8815.63630.3520.7105.764
+ +Table 2. Quantitative Comparison in self-reconstruction setting to person-specific baselines with different data amounts. Our method performs best in various data scenarios, exhibiting robustness. The best and second-best results are in bold and underline. + +
MethodsAudio English, MaleAudio German, Female
Sync-E ↓ 10.443Sync-C ↑ 4.342Sync-E ↓ 9.942Sync-C ↑ 4.949
Full Baseline [28]
ER-NeRF [25]11.7113.04311.5022.853
TalkingGaussian [28]10.6754.31710.8983.708
DFRF [40]10.5154.72610.4824.329
GeneFace [52]9.9954.73411.0294.162
InsTaG (Ours)9.8864.8289.7334.990
+ +Table 3. Comparison in cross-domain setting with 20s training data. We take TalkingGaussian [28] with full data for reference. + +# 4.1. Evaluation + +Comparison Settings. The quantitative comparison contains two settings: 1) Self-reconstruction setting. We split each of the 4 adaptation videos into few-second training sets and a test set of at least 12s, and use the unseen test data as input and ground truth for evaluation. 2) Cross-domain setting. We use an unseen English male audio from NVP [44] and a German female one from [16] to test the cross-domain generalizability. Considering some baselines with failure cases, we use 20s from a certain video to train per method to represent its best performance for testing. + +Metrics and Measurements. We employ PSNR, LPIPS [61], and SSIM [46] to evaluate image quality. Especially, we report PSNR and LPIPS separately on an all image ("A") and only the aligned face ("F"). SSIM is only measured on the face. For lip synchronization, we use LMD [4] and SyncNet [8, 9] confidence (Sync-C) and error (Sync-E). Be + +sides, action unit [10] error (AUE) is used on upper-face (AUE-U) and lower-face (AUE-L) following [28], in which AUE-L can reflect the similarity of the personalized talking actions from the target. We also statistic the training time (fine-tuning time for few-shot or adaptation-based methods) and inference FPS to evaluate efficiency. + +Quantitative Evaluation. 1) Table 1 reports the results of self-reconstruction setting with 5s training data. Our method achieves the best image quality, demonstrating our ability to faithfully generate a realistic talking head. Inheriting the traditions of large corpus pre-training, the one-shot methods [13, 54] and MimicTalk [53] achieve very high Sync scores but are relatively slow. Also getting a high Sync, InsTaG achieves the best in LMD and AUE-L, showing our advantage in preserving personalization. While all baselines can not simultaneously achieve high synchronization and real-time inference, our method produces the best overall performance including the fastest adaptation and high speed. 2) When data increases, the results in Table 2 show that InsTaG still outperforms other person-specific models in all aspects, demonstrating our wider application value and robustness to various scenarios. 3) In the results from Table 3 in cross-domain setting, where the methods are tested on two difficult cross-domain scenarios, the advantages of pre-training are further amplified. GeneFace [52] and DFRF [40] surpass all models trained from scratch. At the same time, InsTaG exceeds the baseline [28] trained on about 1-minute full training data, demonstrating the out + +
MethodsER-NeRF [25]TalkingGaussian [28]GaussianTalker [7]DFRF [40]GeneFace [52]MimicTalk [53]InsTaG (Ours)
Lip-sync Accuracy2.001.901.503.302.704.804.50
Image Quality3.403.704.002.903.103.704.80
Identity Preserving2.401.501.702.903.004.204.60
Video Realness1.401.501.402.801.703.203.90
+ +Table 4. User Study. The rating is in the range of 1-5, higher denotes better. The best and second-best results are in bold and underline. + +
Pre-training PFieldNCLossAdaptation AlignerHookPSNR ↑LPIPS ↓LMD ↓Sync-C ↑
28.660.0443.7342.019
✓28.760.0403.2394.816
✓✓28.590.0413.2235.124
✓✓✓28.690.0403.2745.233
✓✓✓28.660.0403.2575.086
✓✓✓28.790.0403.2264.994
✓✓✓✓28.860.0393.1675.318
+ +standing robustness and generalizability of our method. + +Qualitative Evaluation. To qualitatively evaluate the results, we show the generated sequences and rendering details from self-reconstruction setting. In Figure 4, it can be observed that the models trained from scratch (ER-NeRF, TG) [27, 28] behave poorly in synchronization (yellow box), while the adaptation-based methods (DFRF, GeneFace, MimicTalk) [40, 52, 53] perform better but generate blurry results (orange circle). Referring to Figure 5, while all methods synthesize either blurs or heavily distorted details (red arrow), our method can reconstruct photorealistic appearance and delicate details. By benefitting from the advantages of both pre-training and person-specific training, InsTaG achieves the best in both synchronization and image quality. We strongly recommend watching our supplementary video for an intuitive comparison. + +User Study. To better judge the quality in real human-participated scenarios, we conducted a user study with 28 generated talking head videos from 7 methods with 5s training data. We invited 10 attendees to rate each anonymous method from four aspects: (1) Lip-sync Accuracy; (2) Video Realness; (3) Identity Preserving and (4) Video Realness. The results in Table 4 show that our InsTaG achieves the best in three aspects and the second-best in lip-sync, which is only after the concurrent MimicTalk [53], demonstrating our outstanding visual quality via human judgment. + +# 4.2. Ablation Study + +We conduct the ablation study to prove the effectiveness of our contributions. The results are reported in Table 5. + +Identity-Free Pre-training. In the first row, we show the results pre-trained without Personalized Fields (PField), which produces bad synchronization due to the identity-conflict problem. After applied PField, the pre-training works with obvious effect. Combined with PField, Nega + +Table 5. Ablation Study. We verify the components under 5s self-reconstruction setting, showing the effect of our contributions. + +
BackboneHookPSNR ↑LPIPS ↓LMD ↓Sync-C ↑
Single Branch-28.590.0493.2334.675
Face-Mouth Decom.×28.690.0403.2745.233
Face-Mouth Decomp.✓28.860.0393.1675.318
+ +Table 6. Ablation Study on Mouth Motion Representation. Introducing FMD [28] can effectively improve lip-synchronization, while our FM hook technique further boosts image quality. + +![](images/66725b671ecca2c21363a900d66a6c5a43d2f32026d9990a819f350a1becf75b.jpg) +Figure 6. Visualization Results of Face-Mouth Hook. FMD improves performance but also causes misalignment when with inadequate data, which can be well solved by our Face-Mouth Hook. + +![](images/eb8efee902a7fcfda286531204fa94ef8221898c476e788be6ad4913b94c9d7d.jpg) + +tive Contrast Loss (NCLoss) shows its impact on lip-sync improvement, offering higher Sync-C. This demonstrates the necessity of PFiled in enabling the pre-training and the effect of NCLoss on encouraging knowledge collection. + +Motion-Aligned Adaptation. We apply our Motion Aligner (Aligner) and Face-Mouth Hook (FM Hook) to illustrate their effect on adaptation. Although the performance is improved by applying our pre-training strategy, the Aligner can bring extra enhancement by providing adaptation adjustments. Furthermore, the Hook works in both visual quality and dynamics by correcting inside-mouth misalignment. Combining these two adaptation techniques, the performance achieves the best among all settings. + +Mouth Motion Representation. To further explain the necessity of Face-Mouth Decomposition (FMD) [28] and our FM Hook, we take studies by using one single branch to represent the two parts. Results in Table 6 show the effect of FMD. However, simply introducing it would cause significant misalignment that lowers the visual quality, as in Figure 6. By applying our FM Hook, this drawback can be well solved to benefit FMD with robust performance, generating synchronized talking heads with high visual quality. + +# 5. Ethical Consideration + +We hope our method can promote the healthy development of digital industries. However, it is important to note that it could be misused for malicious purposes, potentially causing negative impacts. As part of our responsibility, we will assist in developing deepfake detection tools. We recommend the responsible use of this and all similar techniques. + +# Acknowledgements + +# References + +[1] James Booth, Epameinondas Antonakos, Stylianos Ploumpis, George Trigeorgis, Yannis Panagakis, and Stefanos Zafeiriou. 3d face morphable models" in-the-wild". In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 48-57, 2017. 3 +[2] Marcel C Bühler, Kripasindhu Sarkar, Tanmay Shah, Gengyan Li, Daoye Wang, Leonhard Helminger, Sergio Orts-Escalano, Dmitry Lagun, Otmar Hilliges, Thabo Beeler, et al. Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3402-3413, 2023. 3 +[3] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, et al. Authentic volumetric avatars from a phone scan. ACM Transactions on Graphics (TOG), 41(4):1-19, 2022. 3 +[4] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Computer Vision-ECCV 2018: 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part VII 15, pages 538-553. Springer, 2018. 7 +[5] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019. 2 +[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 5 +[7] Kyusun Cho, JoungBin Lee, Heeji Yoon, Yeobin Hong, Jaehoon Ko, Sangjun Ahn, and Seungryong Kim. Gaussian talker: Real-time talking head synthesis with 3d gaussian splatting. In ACM Multimedia 2024, 2024. 2, 6, 8 +[8] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 87-103. Springer, 2017. 7 +[9] Joon Son Chung and Andrew Zisserman. Out of time: Automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7 +[10] Paul Ekman and Wallace V. Friesen. Facial Action Coding System: Manual. Palo Alto: Consulting Psychologists Press, 1978. 7 + +In this work, we are supported by the National Natural Science Foundation of China 62276016, 62372029. Lin Gu is supported by JST Moonshot R&D Grant Number JP-MJMS2011 Japan. +[11] Tony Ezzat, Gadi Geiger, and Tomaso Poggio. Trainable videorealistic speech animation. ACM Transactions on Graphics (TOG), 21(3):388-398, 2002. 2 +[12] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3 +[13] Yuan Gan, Zongxin Yang, Xihang Yue, Lingyun Sun, and Yi Yang. Efficient emotional adaptation for audio-driven talking-head generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22634-22645, 2023. 3, 6, 7 +[14] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1155-1164, 2019. 3 +[15] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rinz, Lourdes Agapito, and Matthias Nießner. Mononphm: Dynamic head reconstruction from monocular videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10747-10758, 2024. 3 +[16] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 1, 2, 3, 6, 7 +[17] Siddharth Gururani, Arun Mallya, Ting-Chun Wang, Rafael Valle, and Ming-Yu Liu. Space: Speech-driven portrait animation with controllable expression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20914-20923, 2023. 3 +[18] Awni Hannun, Carl Case, Jared Casper, Bryan Catanzaro, Greg Diamos, Erich Elsen, Ryan Prenger, Sanjeev Satheesh, Shubho Sengupta, Adam Coates, et al. Deep speech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:1412.5567, 2014. 6 +[19] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 3 +[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 2 +[21] Amir Jamaludin, Joon Son Chung, and Andrew Zisserman. You said that?: Synthesising talking faces from audio. International Journal of Computer Vision, 127:1767-1779, 2019. 2 +[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5 +[23] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and + +Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 5 +[24] Dongze Li, Kang Zhao, Wei Wang, Bo Peng, Yingya Zhang, Jing Dong, and Tieniu Tan. Ae-nerf: Audio enhanced neural radiance field for few shot talking head synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3037–3045, 2024. 3 +[25] Jiahe Li, Jiawei Zhang, Xiao Bai, Jun Zhou, and Lin Gu. Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7568-7578, 2023. 2, 3, 4, 6, 7, 8 +[26] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 5 +[27] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Jun Zhou, and Lin Gu. Er-nerf++: Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. Information Fusion, 110:102456, 2024. 1, 8 +[28] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Talkinggaussian: Structure-persistent 3d talking head synthesis via gaussian splatting. In European Conference on Computer Vision, pages 127-145. Springer, 2025. 1, 2, 3, 5, 6, 7, 8 +[29] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 3 +[30] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhigang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17969-17978, 2023. 1, 2, 3, 4, 5 +[31] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 106-125. Springer, 2022. 1, 2 +[32] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6 +[33] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2 +[34] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, pages 179-197. Springer, 2022. 3 +[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: + +Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421. Springer, 2020. 1, 2 +[36] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. Ieee, 2009. 3, 6 +[37] Ziqiao Peng, Wentao Hu, Yue Shi, Xiangyu Zhu, Xiaomei Zhang, Hao Zhao, Jun He, Hongyan Liu, and Zhaoxin Fan. Synctalk: The devil is in the synchronization for talking head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 666-676, 2024. 2 +[38] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2 +[39] Eduard Ramon, Gil Triginer, Janna Escur, Albert Pumarola, Jaime Garcia, Xavier Giro-i Nieto, and Francesc Moreno-Noguer. H3d-net: Few-shot high-fidelity 3d head reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5620-5629, 2021. 3 +[40] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII, pages 666-682. Springer, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[41] Luchuan Song, Pinxin Liu, Lele Chen, Guojun Yin, and Chenliang Xu. Tri²-plane: Thinking head avatar via feature pyramid. In European Conference on Computer Vision, pages 1–20. Springer, 2025. 3 +[42] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 2, 3, 6 +[43] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 3 +[44] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 716-731. Springer, 2020. 2, 7 +[45] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI, pages 700–717. Springer, 2020. 2 + +[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5, 7 +[47] Olivia Wiles, A Sophia Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In Computer Vision-ECCV 2018: 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XIII 15, pages 690-706. Springer, 2018. 2 +[48] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024.3 +[49] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 3 +[50] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 3 +[51] Xihe Yang, Xingyu Chen, Daiheng Gao, Shaohui Wang, Xiaoguang Han, and Baoyuan Wang. Have-fun: Human avatar reconstruction from few-shot unconstrained images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 742-752, 2024. 3 +[52] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, Jinzheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 6, 7, 8 +[53] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiangwei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, Chen Zhang, Xiang Yin, Zejun Ma, and Zhou Zhao. Mimictalk: Mimicking a personalized and expressive 3d talking face in few minutes. Advances in neural information processing systems, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[54] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiawei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6, 7 +[55] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137, 2020. 2 +[56] Hongyun Yu, Zhan Qu, Qihang Yu, Jianchuan Chen, Zhonghua Jiang, Zhiwen Chen, Shengyu Zhang, Jimin Xu, Fei Wu, Chengfei Lv, et al. Gaussian talker: Speaker-specific talking head synthesis via 3d gaussian splattering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3548-3557, 2024. 2 + +[57] Yu Yu, Kenneth Alberto Funes Mora, and Jean-Marc Odobez. Headfusion: $360^{\circ}$ head pose tracking combining 3d morphable model and 3d reconstruction. IEEE transactions on pattern analysis and machine intelligence, 40(11): 2653-2667, 2018. 3 +[58] Zhentao Yu, Zixin Yin, Deyu Zhou, Duomin Wang, Finn Wong, and Baoyuan Wang. Talking head generation with probabilistic audio-to-visual diffusion priors. arXiv preprint arXiv:2212.04248, 2022. 2, 6 +[59] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9459-9468, 2019. 3 +[60] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3867-3876, 2021. 2 +[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 +[62] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2 \ No newline at end of file diff --git a/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/images.zip b/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b4e876d596ef14e877e8608e2c91957759f81aa1 --- /dev/null +++ b/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f9a8613147591535cf5c1294c6a85e1a1902d4d1787c730461920b815552b32 +size 852537 diff --git a/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/layout.json b/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..27d78350ae22e10af2259ce90edc430d2068a7f2 --- /dev/null +++ b/2025/InsTaG_ Learning Personalized 3D Talking Head from Few-Second Video/layout.json @@ -0,0 +1,11648 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 532, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 532, + 120 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 532, + 120 + ], + "type": "text", + "content": "InsTaG: Learning Personalized 3D Talking Head from Few-Second Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "spans": [ + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "type": "text", + "content": "Jiahe Li" + }, + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "type": "text", + "content": " Jiawei Zhang" + }, + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "type": "text", + "content": " Xiao Bai" + }, + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "type": "text", + "content": " Jin Zheng" + }, + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "type": "text", + "content": " Jun Zhou" + }, + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "type": "text", + "content": " Lin Gu" + }, + { + "bbox": [ + 107, + 142, + 501, + 157 + ], + "type": "inline_equation", + "content": "^{3,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 157, + 547, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 157, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 62, + 157, + 547, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 157, + 547, + 185 + ], + "type": "text", + "content": "School of Computer Science and Engineering, State Key Laboratory of Complex & Critical Software Environment, Jiangxi Research Institute, Beihang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 185, + 489, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 185, + 489, + 199 + ], + "spans": [ + { + "bbox": [ + 121, + 185, + 489, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 185, + 489, + 199 + ], + "type": "text", + "content": "School of Information and Communication Technology, Griffith University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 201, + 199, + 405, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 199, + 405, + 213 + ], + "spans": [ + { + "bbox": [ + 201, + 199, + 405, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 201, + 199, + 405, + 213 + ], + "type": "text", + "content": "RIKEN AIP " + }, + { + "bbox": [ + 201, + 199, + 405, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 201, + 199, + 405, + 213 + ], + "type": "text", + "content": "The University of Tokyo" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 228, + 126, + 289 + ], + "blocks": [ + { + "bbox": [ + 58, + 228, + 126, + 289 + ], + "lines": [ + { + "bbox": [ + 58, + 228, + 126, + 289 + ], + "spans": [ + { + "bbox": [ + 58, + 228, + 126, + 289 + ], + "type": "image", + "image_path": "9f2aa9bfb22be44ffd9dd7f4038aaa8dadc1a26e408367a36f356cc4c37a4632.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 58, + 293, + 121, + 353 + ], + "blocks": [ + { + "bbox": [ + 58, + 293, + 121, + 353 + ], + "lines": [ + { + "bbox": [ + 58, + 293, + 121, + 353 + ], + "spans": [ + { + "bbox": [ + 58, + 293, + 121, + 353 + ], + "type": "image", + "image_path": "31bd0f72d9a39afcaaf2ac6a13996ced8b6147ef6edbb82c5161baa26ec608da.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 356, + 121, + 367 + ], + "lines": [ + { + "bbox": [ + 57, + 356, + 121, + 367 + ], + "spans": [ + { + "bbox": [ + 57, + 356, + 121, + 367 + ], + "type": "text", + "content": "4 5-Second Video" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 132, + 228, + 253, + 356 + ], + "blocks": [ + { + "bbox": [ + 132, + 228, + 253, + 356 + ], + "lines": [ + { + "bbox": [ + 132, + 228, + 253, + 356 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 253, + 356 + ], + "type": "image", + "image_path": "db056d0232130caf05033539929122d44cca178ce16123337e72b896e9af1390.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 358, + 227, + 367 + ], + "lines": [ + { + "bbox": [ + 162, + 358, + 227, + 367 + ], + "spans": [ + { + "bbox": [ + 162, + 358, + 227, + 367 + ], + "type": "text", + "content": "Instant Adaptation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 147, + 357, + 159, + 367 + ], + "blocks": [ + { + "bbox": [ + 147, + 357, + 159, + 367 + ], + "lines": [ + { + "bbox": [ + 147, + 357, + 159, + 367 + ], + "spans": [ + { + "bbox": [ + 147, + 357, + 159, + 367 + ], + "type": "image", + "image_path": "710fe4799ae419a505052e72412b92eb1e166b2ab4bb121d60d5fdafd2b9918f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 259, + 228, + 315, + 285 + ], + "blocks": [ + { + "bbox": [ + 259, + 228, + 315, + 285 + ], + "lines": [ + { + "bbox": [ + 259, + 228, + 315, + 285 + ], + "spans": [ + { + "bbox": [ + 259, + 228, + 315, + 285 + ], + "type": "image", + "image_path": "1fd73e80e5b2399c41bae8227a9e9fd8486aae97f580ba66d2d856aab3935666.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 228, + 372, + 285 + ], + "blocks": [ + { + "bbox": [ + 316, + 228, + 372, + 285 + ], + "lines": [ + { + "bbox": [ + 316, + 228, + 372, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 372, + 285 + ], + "type": "image", + "image_path": "b2b14592a19a3d30f800c922b2f4c43c2bbdf9aa3c091d2dbe9a1f5eaa348f77.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 320, + 286, + 326, + 293 + ], + "blocks": [ + { + "bbox": [ + 271, + 286, + 327, + 293 + ], + "lines": [ + { + "bbox": [ + 271, + 286, + 327, + 293 + ], + "spans": [ + { + "bbox": [ + 271, + 286, + 327, + 293 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 286, + 326, + 293 + ], + "lines": [ + { + "bbox": [ + 320, + 286, + 326, + 293 + ], + "spans": [ + { + "bbox": [ + 320, + 286, + 326, + 293 + ], + "type": "image", + "image_path": "faa09c48ccaf4997f752807f6f5dfcd9897dd5bec9184c77b6908dd01882456b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 328, + 286, + 334, + 293 + ], + "blocks": [ + { + "bbox": [ + 328, + 286, + 334, + 293 + ], + "lines": [ + { + "bbox": [ + 328, + 286, + 334, + 293 + ], + "spans": [ + { + "bbox": [ + 328, + 286, + 334, + 293 + ], + "type": "image", + "image_path": "14df0e95c8c120e7e913efb2a329be7c0655ee930e4a479e3fc84cc8ddfa6be6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 336, + 286, + 342, + 293 + ], + "blocks": [ + { + "bbox": [ + 336, + 286, + 342, + 293 + ], + "lines": [ + { + "bbox": [ + 336, + 286, + 342, + 293 + ], + "spans": [ + { + "bbox": [ + 336, + 286, + 342, + 293 + ], + "type": "image", + "image_path": "3b7295e2f9ed1ccdd8a11ee0e8e8a4dd57d0454327e0c86c5ec22f2edb767315.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 344, + 286, + 351, + 293 + ], + "blocks": [ + { + "bbox": [ + 344, + 286, + 351, + 293 + ], + "lines": [ + { + "bbox": [ + 344, + 286, + 351, + 293 + ], + "spans": [ + { + "bbox": [ + 344, + 286, + 351, + 293 + ], + "type": "image", + "image_path": "979e1435fd3b909ace6fed21011c215a43e6ab1e51988dc90eddb9b56673ffdf.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 353, + 228, + 427, + 285 + ], + "blocks": [ + { + "bbox": [ + 353, + 228, + 427, + 285 + ], + "lines": [ + { + "bbox": [ + 353, + 228, + 427, + 285 + ], + "spans": [ + { + "bbox": [ + 353, + 228, + 427, + 285 + ], + "type": "image", + "image_path": "ba94422f9262a5181b40113feb9309c8e903a71447396d61fd8ecd35a92bd556.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 353, + 286, + 359, + 293 + ], + "blocks": [ + { + "bbox": [ + 353, + 286, + 359, + 293 + ], + "lines": [ + { + "bbox": [ + 353, + 286, + 359, + 293 + ], + "spans": [ + { + "bbox": [ + 353, + 286, + 359, + 293 + ], + "type": "image", + "image_path": "a07d2db5b7ab3913bf8f22f69c41d5028ce01656a8509ae73eebb5b59e03f213.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 361, + 286, + 367, + 293 + ], + "blocks": [ + { + "bbox": [ + 361, + 286, + 367, + 293 + ], + "lines": [ + { + "bbox": [ + 361, + 286, + 367, + 293 + ], + "spans": [ + { + "bbox": [ + 361, + 286, + 367, + 293 + ], + "type": "image", + "image_path": "8e8814e81e349ac0b8b7b9897e941e2a83c87b20ec6b0cbc98d3da4fb5e1ca1a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 369, + 286, + 375, + 293 + ], + "blocks": [ + { + "bbox": [ + 369, + 286, + 375, + 293 + ], + "lines": [ + { + "bbox": [ + 369, + 286, + 375, + 293 + ], + "spans": [ + { + "bbox": [ + 369, + 286, + 375, + 293 + ], + "type": "image", + "image_path": "d19bf764db921942ac2e1693fd8aed4bb8a98c902b597d05a4f6e53cead7f57e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 377, + 286, + 384, + 293 + ], + "blocks": [ + { + "bbox": [ + 377, + 286, + 384, + 293 + ], + "lines": [ + { + "bbox": [ + 377, + 286, + 384, + 293 + ], + "spans": [ + { + "bbox": [ + 377, + 286, + 384, + 293 + ], + "type": "image", + "image_path": "6b53bc7e94c1cde46dab2794c46bdec028fd75e65762c4ee86419ab45213e821.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 386, + 228, + 429, + 285 + ], + "blocks": [ + { + "bbox": [ + 386, + 228, + 429, + 285 + ], + "lines": [ + { + "bbox": [ + 386, + 228, + 429, + 285 + ], + "spans": [ + { + "bbox": [ + 386, + 228, + 429, + 285 + ], + "type": "image", + "image_path": "fa8042717dc08e538db267d8e2d514512f7b1841f3fc42f49af7df2decf7d1c6.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 384, + 286, + 391, + 293 + ], + "blocks": [ + { + "bbox": [ + 384, + 286, + 391, + 293 + ], + "lines": [ + { + "bbox": [ + 384, + 286, + 391, + 293 + ], + "spans": [ + { + "bbox": [ + 384, + 286, + 391, + 293 + ], + "type": "image", + "image_path": "4a115dc5eda4726d769dc52b70af36625c2dc2230503487514c1f9e14493fceb.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 393, + 286, + 400, + 293 + ], + "blocks": [ + { + "bbox": [ + 393, + 286, + 400, + 293 + ], + "lines": [ + { + "bbox": [ + 393, + 286, + 400, + 293 + ], + "spans": [ + { + "bbox": [ + 393, + 286, + 400, + 293 + ], + "type": "image", + "image_path": "1127508aa857436268f37440fdb11efb9e40ef223deb934009ca9fd95f83baaa.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 402, + 286, + 408, + 293 + ], + "blocks": [ + { + "bbox": [ + 402, + 286, + 408, + 293 + ], + "lines": [ + { + "bbox": [ + 402, + 286, + 408, + 293 + ], + "spans": [ + { + "bbox": [ + 402, + 286, + 408, + 293 + ], + "type": "image", + "image_path": "c9e4e8ef72f7b97dd5a2369f1ff876d98555194139a903fa3a7c2778258d7746.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 410, + 228, + 488, + 285 + ], + "blocks": [ + { + "bbox": [ + 410, + 228, + 488, + 285 + ], + "lines": [ + { + "bbox": [ + 410, + 228, + 488, + 285 + ], + "spans": [ + { + "bbox": [ + 410, + 228, + 488, + 285 + ], + "type": "image", + "image_path": "e476341d19c3f90f3c5a9dce5d26b2448e3411bb11f535e3820f0c9f57e42777.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 410, + 286, + 417, + 293 + ], + "blocks": [ + { + "bbox": [ + 410, + 286, + 417, + 293 + ], + "lines": [ + { + "bbox": [ + 410, + 286, + 417, + 293 + ], + "spans": [ + { + "bbox": [ + 410, + 286, + 417, + 293 + ], + "type": "image", + "image_path": "d39441fc7c6eacead48266ced4a3179b907026d24bfa10edd43bb0f55ba8e62d.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 419, + 286, + 425, + 293 + ], + "blocks": [ + { + "bbox": [ + 419, + 286, + 425, + 293 + ], + "lines": [ + { + "bbox": [ + 419, + 286, + 425, + 293 + ], + "spans": [ + { + "bbox": [ + 419, + 286, + 425, + 293 + ], + "type": "image", + "image_path": "1203f747297cbd8365b3806e7c9ffc6db20ce07e116d8b235409c1fb6433855f.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 428, + 286, + 435, + 293 + ], + "blocks": [ + { + "bbox": [ + 428, + 286, + 435, + 293 + ], + "lines": [ + { + "bbox": [ + 428, + 286, + 435, + 293 + ], + "spans": [ + { + "bbox": [ + 428, + 286, + 435, + 293 + ], + "type": "image", + "image_path": "60892eced30bd09f1b8bfc6573686450a31de1f729d1674247691e057630c227.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 436, + 286, + 443, + 293 + ], + "blocks": [ + { + "bbox": [ + 436, + 286, + 443, + 293 + ], + "lines": [ + { + "bbox": [ + 436, + 286, + 443, + 293 + ], + "spans": [ + { + "bbox": [ + 436, + 286, + 443, + 293 + ], + "type": "image", + "image_path": "5b4a430ad266f1c47c9334eeda81aa3f5a0dabe4dd92930ddb373473e9596b96.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 446, + 228, + 488, + 285 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 228, + 488, + 285 + ], + "spans": [ + { + "bbox": [ + 446, + 228, + 488, + 285 + ], + "type": "image", + "image_path": "3c3432b297808b9495d03ed82df12454a1d565db357004c3d43803ebe971b613.jpg" + } + ] + } + ], + "index": 34, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 446, + 286, + 454, + 293 + ], + "blocks": [ + { + "bbox": [ + 446, + 286, + 454, + 293 + ], + "lines": [ + { + "bbox": [ + 446, + 286, + 454, + 293 + ], + "spans": [ + { + "bbox": [ + 446, + 286, + 454, + 293 + ], + "type": "image", + "image_path": "46b31cc39b99e9e8266c9360345e4d48970b664390db3d16f9f9c416f709536c.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 459, + 286, + 465, + 293 + ], + "blocks": [ + { + "bbox": [ + 459, + 286, + 465, + 293 + ], + "lines": [ + { + "bbox": [ + 459, + 286, + 465, + 293 + ], + "spans": [ + { + "bbox": [ + 459, + 286, + 465, + 293 + ], + "type": "image", + "image_path": "d831fb9e16dfc3f7e0d302366dfbf5b69997f0aa1d149c047b320c684f47d5be.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 468, + 286, + 474, + 293 + ], + "blocks": [ + { + "bbox": [ + 468, + 286, + 474, + 293 + ], + "lines": [ + { + "bbox": [ + 468, + 286, + 474, + 293 + ], + "spans": [ + { + "bbox": [ + 468, + 286, + 474, + 293 + ], + "type": "image", + "image_path": "1532240c6186fa412b466c21031a91df84f3f7cb45468dccca1c8285012e2535.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 477, + 286, + 484, + 293 + ], + "blocks": [ + { + "bbox": [ + 477, + 286, + 484, + 293 + ], + "lines": [ + { + "bbox": [ + 477, + 286, + 484, + 293 + ], + "spans": [ + { + "bbox": [ + 477, + 286, + 484, + 293 + ], + "type": "image", + "image_path": "8b68e9b08100617e52c6bd73ceb7592a5100c9c082a370c23ac5e847f723d86d.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 487, + 228, + 552, + 285 + ], + "blocks": [ + { + "bbox": [ + 487, + 228, + 552, + 285 + ], + "lines": [ + { + "bbox": [ + 487, + 228, + 552, + 285 + ], + "spans": [ + { + "bbox": [ + 487, + 228, + 552, + 285 + ], + "type": "image", + "image_path": "16687558939e41f2b833193f4931721e13487586c766626ad451d12dece8d7e3.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 488, + 286, + 495, + 293 + ], + "blocks": [ + { + "bbox": [ + 488, + 286, + 495, + 293 + ], + "lines": [ + { + "bbox": [ + 488, + 286, + 495, + 293 + ], + "spans": [ + { + "bbox": [ + 488, + 286, + 495, + 293 + ], + "type": "image", + "image_path": "898d194d65f7295f6612539a1041ae70d32a119535f5f54f480051a77fed5630.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 497, + 286, + 504, + 293 + ], + "blocks": [ + { + "bbox": [ + 497, + 286, + 504, + 293 + ], + "lines": [ + { + "bbox": [ + 497, + 286, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 497, + 286, + 504, + 293 + ], + "type": "image", + "image_path": "189b59b52811878ed832a360ec76b213c84d09df68652dad4612dfb385d59d2a.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 507, + 286, + 513, + 293 + ], + "blocks": [ + { + "bbox": [ + 507, + 286, + 513, + 293 + ], + "lines": [ + { + "bbox": [ + 507, + 286, + 513, + 293 + ], + "spans": [ + { + "bbox": [ + 507, + 286, + 513, + 293 + ], + "type": "image", + "image_path": "f26a35b278fded0c1c6fc99521b32d643164a0bb090e73cfb8517b1e02ede095.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 515, + 286, + 522, + 293 + ], + "blocks": [ + { + "bbox": [ + 515, + 286, + 522, + 293 + ], + "lines": [ + { + "bbox": [ + 515, + 286, + 522, + 293 + ], + "spans": [ + { + "bbox": [ + 515, + 286, + 522, + 293 + ], + "type": "image", + "image_path": "0fdfcfe25cc7037066b71df990b0c5665160b494a5002815df4fc93aaa03fe31.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 526, + 286, + 531, + 293 + ], + "blocks": [ + { + "bbox": [ + 526, + 286, + 531, + 293 + ], + "lines": [ + { + "bbox": [ + 526, + 286, + 531, + 293 + ], + "spans": [ + { + "bbox": [ + 526, + 286, + 531, + 293 + ], + "type": "image", + "image_path": "bd3cf62c03c65ec8033f06af8339a6bc0280d88711159e63042cebc5b9c70784.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 534, + 286, + 541, + 293 + ], + "blocks": [ + { + "bbox": [ + 534, + 286, + 541, + 293 + ], + "lines": [ + { + "bbox": [ + 534, + 286, + 541, + 293 + ], + "spans": [ + { + "bbox": [ + 534, + 286, + 541, + 293 + ], + "type": "image", + "image_path": "0499b94092db0dd2149c7623c669e599f128559cf21357b8bce7aa78beb590ec.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 260, + 298, + 315, + 356 + ], + "blocks": [ + { + "bbox": [ + 260, + 298, + 315, + 356 + ], + "lines": [ + { + "bbox": [ + 260, + 298, + 315, + 356 + ], + "spans": [ + { + "bbox": [ + 260, + 298, + 315, + 356 + ], + "type": "image", + "image_path": "302e18d82be343df5ca81db567fb0a7587a3ef226537040b5fe7afaca51e56e4.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 358, + 304, + 367 + ], + "lines": [ + { + "bbox": [ + 271, + 358, + 304, + 367 + ], + "spans": [ + { + "bbox": [ + 271, + 358, + 304, + 367 + ], + "type": "text", + "content": "GeneFace" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 316, + 298, + 372, + 356 + ], + "blocks": [ + { + "bbox": [ + 316, + 298, + 372, + 356 + ], + "lines": [ + { + "bbox": [ + 316, + 298, + 372, + 356 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 372, + 356 + ], + "type": "image", + "image_path": "f0b93c63e6361411c769c5e974c10b166d1826cd373ef0c359b0969a6907e051.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 359, + 370, + 368 + ], + "lines": [ + { + "bbox": [ + 320, + 359, + 370, + 368 + ], + "spans": [ + { + "bbox": [ + 320, + 359, + 370, + 368 + ], + "type": "text", + "content": "TalkingGaussian" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 372, + 298, + 429, + 356 + ], + "blocks": [ + { + "bbox": [ + 372, + 298, + 429, + 356 + ], + "lines": [ + { + "bbox": [ + 372, + 298, + 429, + 356 + ], + "spans": [ + { + "bbox": [ + 372, + 298, + 429, + 356 + ], + "type": "image", + "image_path": "c90f1ef5c16190be98058b94f21a1e7d3b5bc1fc7ceb0894726b283900282161.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 359, + 417, + 367 + ], + "lines": [ + { + "bbox": [ + 383, + 359, + 417, + 367 + ], + "spans": [ + { + "bbox": [ + 383, + 359, + 417, + 367 + ], + "type": "text", + "content": "MimicTalk" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 432, + 298, + 490, + 356 + ], + "blocks": [ + { + "bbox": [ + 432, + 298, + 490, + 356 + ], + "lines": [ + { + "bbox": [ + 432, + 298, + 490, + 356 + ], + "spans": [ + { + "bbox": [ + 432, + 298, + 490, + 356 + ], + "type": "image", + "image_path": "7a2f7b8d2297913a4c533a423459b431f2a483c516450a7278c4de362dde8a15.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 448, + 359, + 458, + 367 + ], + "lines": [ + { + "bbox": [ + 448, + 359, + 458, + 367 + ], + "spans": [ + { + "bbox": [ + 448, + 359, + 458, + 367 + ], + "type": "text", + "content": "In" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 438, + 358, + 444, + 368 + ], + "blocks": [ + { + "bbox": [ + 438, + 358, + 444, + 368 + ], + "lines": [ + { + "bbox": [ + 438, + 358, + 444, + 368 + ], + "spans": [ + { + "bbox": [ + 438, + 358, + 444, + 368 + ], + "type": "image", + "image_path": "db37008c48d8d892b18be525425933e48ec5827c25dd08a96bdb0cc057b413ad.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 380, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. With only 5-second video data, InsTaG outperforms the state-of-the-arts [28, 52, 53] by delivering high-quality personalized lip synchronization and realistic rendering with the fastest adaptation, meanwhile attaining low memory overhead and real-time inference." + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 461, + 359, + 468, + 367 + ], + "blocks": [ + { + "bbox": [ + 461, + 359, + 468, + 367 + ], + "lines": [ + { + "bbox": [ + 461, + 359, + 468, + 367 + ], + "spans": [ + { + "bbox": [ + 461, + 359, + 468, + 367 + ], + "type": "image", + "image_path": "3410ea1cd96cd84e0e07147237f20c9a455ab8f3bda0ef45a895ca84c257ec6c.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 359, + 475, + 367 + ], + "lines": [ + { + "bbox": [ + 470, + 359, + 475, + 367 + ], + "spans": [ + { + "bbox": [ + 470, + 359, + 475, + 367 + ], + "type": "text", + "content": "G" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 495, + 298, + 553, + 356 + ], + "blocks": [ + { + "bbox": [ + 495, + 298, + 553, + 356 + ], + "lines": [ + { + "bbox": [ + 495, + 298, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 495, + 298, + 553, + 356 + ], + "type": "image", + "image_path": "33690756a5d4706e6120b3ded57eae826e1e6b7789f24e2c8bcd137c23f78100.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 359, + 546, + 367 + ], + "lines": [ + { + "bbox": [ + 500, + 359, + 546, + 367 + ], + "spans": [ + { + "bbox": [ + 500, + 359, + 546, + 367 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 57 + }, + { + "bbox": [ + 152, + 413, + 199, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 413, + 199, + 424 + ], + "spans": [ + { + "bbox": [ + 152, + 413, + 199, + 424 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 54, + 448, + 296, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 448, + 296, + 699 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 296, + 699 + ], + "type": "text", + "content": "Despite exhibiting impressive performance in synthesizing lifelike personalized 3D talking heads, prevailing methods based on radiance fields suffer from high demands for training data and time for each new identity. This paper introduces InsTaG, a 3D talking head synthesis framework that allows a fast learning of realistic personalized 3D talking head from few training data. Built upon a lightweight 3DGS person-specific synthesizer with universal motion priors, InsTaG achieves high-quality and fast adaptation while preserving high-level personalization and efficiency. As preparation, we first propose an Identity-Free Pre-training strategy that enables the pre-training of the person-specific model and encourages the collection of universal motion priors from long-video data corpus. To fully exploit the universal motion priors to learn an unseen new identity, we then present a Motion-Aligned Adaptation strategy to adaptively align the target head to the pre-trained field, and constrain a robust dynamic head structure under few training data. Experiments demonstrate our outstanding performance and efficiency under various data scenarios to render high-quality personalized talking heads. Project page:" + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 55, + 700, + 285, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 700, + 285, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 700, + 285, + 711 + ], + "type": "text", + "content": "https://fictionarry.github.io/InsTaG/." + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 315, + 412, + 394, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 412, + 394, + 424 + ], + "spans": [ + { + "bbox": [ + 315, + 412, + 394, + 424 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 313, + 432, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 540 + ], + "type": "text", + "content": "Audio-driven talking head synthesis has become an important technique for various digital applications like video production, virtual reality, and human-computer interaction. The recent advancements in Neural Radiance Fields (NeRF) [35] and 3D Gaussian Splatting (3DGS) [22] have led to significant progress in 3D representations. Various methods [16, 27, 28, 31, 40, 52] now utilize radiance fields to create high-fidelity 3D talking head videos, capturing personalized talking styles with impressive photorealistic image quality." + } + ] + } + ], + "index": 64 + }, + { + "bbox": [ + 313, + 540, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 556, + 696 + ], + "type": "text", + "content": "Most of these methods adopt person-specific training. Given a video clip with the target talking person, a whole person-specific model is trained to memorize the target talking head. Despite their tremendous improvements in high-fidelity rendering, personalized dynamics, and fast running speed, as a trade-off, the model holds no generalizability for an unseen identity. Thus, they require a large amount of high-quality video frames with a long training time to adapt to each new identity. Some image-referred methods [30, 40, 53, 54] can quickly generalize to different identities, nevertheless, the quality of personalization and efficiency drop. A question is then raised: How can we use the least training data to rapidly learn a faithful 3D talking" + } + ] + } + ], + "index": 65 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 329, + 702, + 492, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 702, + 492, + 713 + ], + "spans": [ + { + "bbox": [ + 329, + 702, + 492, + 713 + ], + "type": "text", + "content": "*Corresponding authors: Xiao Bai and Jin Zheng." + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10690" + } + ] + } + ], + "index": 67 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 278, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 278, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 278, + 85 + ], + "type": "text", + "content": "head, while preserving a high level of personalization?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 302 + ], + "type": "text", + "content": "Observing previous radiance-field-based methods, we notice that: 1) While the training process includes a long time to learn a basic audio-driven motion that with only little personalized style, the static 3D head can be well reconstructed in just a few early iterations; 2) Though the results may exhibit jittering and poor quality when trained with limited data, a portion of the lip motions remains recognizable with distinct personalized style. These suggest that even a short video clip can provide rich personalization cues to enable the learning of a specific talking head with little time cost, as long as common audio-motion knowledge is obtained. Motivated by this, by preparing a pre-trained identity-free 3D motion field as prior and aligning it to the new identity to facilitate the adaptation, we propose Instant Talking Head Synthesis with Gaussian Splatting (InsTaG), a framework that allows fast learning of high-fidelity personalized 3D talking head from a video even as short as only a few seconds while attaining high efficiency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 304, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 568 + ], + "type": "text", + "content": "Different from previous few-shot methods [30, 53, 54] that imitate a new identity from inputs, InsTaG memorizes the whole talking head in a lightweight person-specific model to obtain high-level personalization and efficiency. Upon such a person-specific paradigm, we fully decouple the learning of universal motion and personalization into pre-training and adaptation, preparing a pre-trained 3D motion field for future person-specific learning of new identity. However, a challenge with this design lies in pre-training with multi-person data, where the person-specific model can not fit varying identities and estimate their varied personalized motions, leading to an ineffective optimization. To tackle this problem, we present an Identity-Free Pretraining strategy. By keeping a series of temporary personal fields to store the identity information and filter out the personalized motions, we extract the common knowledge of talking motion from long-video corpus and produce a universal motion field that contains identity-free motion prior. To further facilitate this motion detachment, a negative contrast loss is introduced to encourage the diversity of learned personalized motion between each sample, maximizing the retaining of universal priors for later adaptation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "To fully exploit the universal motion priors to enable fast and high-quality new-identity adaptation, we present a Motion-Aligned Adaptation strategy to align the unseen talking head to the pre-trained motion field. First, a motion aligner is developed to learn a condition-independent primitive-wise coordinate offset and a motion scalar. It helps the intra-alignment between the personalized and universal motions in the field, maximizing the retention of learned knowledge during few-shot fine-tuning. Moreover, to enhance the generalizability by face-mouth decomposition [58] while preserving a robust dynamic face structure, we introduce a face-mouth hook technique to inter-align the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "inside mouth motion with the face, which efficiently improves the fidelity as well as motion quality. To further improve the robustness of view direction, the new identity's head structure is aligned with estimated geometry priors for regularization. By incorporating the alignments in the adaptation, the person-specific model finally synthesizes high-fidelity personalized 3D talking heads at a fast speed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 157, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 240 + ], + "type": "text", + "content": "Integrating the two proposed strategies, InsTaG can learn a personalized 3D talking head from even a short 5-second video, achieving state-of-the-art visual quality and personalization compared to existing methods, with high lip-synchronization and real-time inference. Experiments demonstrate the outperforming efficiency and generalizability of InsTaG for various identities, genders and languages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 325, + 240, + 534, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 240, + 534, + 251 + ], + "spans": [ + { + "bbox": [ + 325, + 240, + 534, + 251 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 255, + 553, + 397 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 255, + 553, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 255, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 314, + 255, + 553, + 302 + ], + "type": "text", + "content": "- An Identity-Free Pre-training strategy that enables the pretraining of the person-specific model by filtering out conflict identity and personalized motions, meanwhile maximizing the capturing of universal motion priors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 302, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 302, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 314, + 302, + 553, + 350 + ], + "type": "text", + "content": "- A Motion-Aligned Adaptation strategy, which inter- and intra-aligns the talking head with the motion field for efficient new-identity adaptation, attaining robust, realistic reconstruction and personalized lip-synchronization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 350, + 553, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 350, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 314, + 350, + 553, + 397 + ], + "type": "text", + "content": "- Extensive experiments show that our InsTaG learns realistic personalized talking heads from few data, while attaining high lip-synchronization and efficiency, outperforming state-of-the-art methods under various scenarios." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 408, + 400, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 408, + 400, + 421 + ], + "spans": [ + { + "bbox": [ + 314, + 408, + 400, + 421 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": "3D Talking Head Synthesis. Talking head synthesis aims to use arbitrary audio to reenact a talking person to generate audio-visual synchronized videos. Early works [5, 11, 21, 38, 45, 47, 62] are mainly built on generative models, creating talking heads by manipulating the given 2D images. Later, to solve the temporal inconsistency when the head moves, 3D-based methods [33, 44, 55, 60] utilize explicit 3D face structure and successfully improve the naturalness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": "Recently, radiance fields like NeRF [35] and 3D Gaussian Splitting (3DGS) [22] have been introduced as the representation to allow 3D talking head synthesize. Inheriting the radiance field optimization, most methods [7, 16, 25, 28, 31, 37, 42] train a person-specific model on minutes of high-quality video, achieving success in reconstructing photorealistic rendering and personalized talking style. However, the strict quality requirements of the training video data and the long training time for every adaptation to a new identity have hugely limited their application. Although some works [30, 40, 52-54, 56] attempt to solve the problem by a one-shot generator [30, 53, 54] or take pre-trained motions from external modules to loosely join the adaptation [40, 52] to achieve one- or few-shot, they often compromise image quality and loses the personalization as the trade-off. A concurrent work MimicTalk [53] injects LoRA [20] to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10691" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 550, + 185 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 550, + 185 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 550, + 185 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 550, + 185 + ], + "type": "image", + "image_path": "0113c62eedf6e93077e4e9fb7d58a434dbdfabbe6183c8414f97c2b72a48b895.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 187, + 550, + 299 + ], + "blocks": [ + { + "bbox": [ + 58, + 187, + 550, + 299 + ], + "lines": [ + { + "bbox": [ + 58, + 187, + 550, + 299 + ], + "spans": [ + { + "bbox": [ + 58, + 187, + 550, + 299 + ], + "type": "image", + "image_path": "9df1aad750b0176fcfce7ac78676c69e2f4b857b7a8331d4eb549f54f6a40e0a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 308, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 342 + ], + "type": "text", + "content": "Figure 2. Overview of InsTaG. For preparation, InsTaG collects the common knowledge of talking motion from a long-video corpus by Identity-Free Pre-training, storing it as a motion field. Given a short video with a new identity, the Motion-Aligned Adaptation strategy builds a robust and fast person-specific synthesizer with the pre-trained motion field to learn a high-quality personalized 3D talking head." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 352, + 295, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 352, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 54, + 352, + 295, + 435 + ], + "type": "text", + "content": "improve fine-tuning efficiency, but its inference is slow due to the large backbone. Instead, by enabling the pretraining of person-specific models and using pre-trained motion field to directly drive the talking head, our method obtains a more compact and consistent architecture to fully exploit the motion priors in an end-to-end way, facilitating modeling precise lip-synchronization with high efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "content": "Few-shot 3D Head Reconstruction. Fitting high-fidelity 3D human heads from RGB images has been a hot topic. Early model-based methods [14, 43, 50, 57] focus on fitting a 3D head with 3DMMs [1, 29, 36] but are short in detail. Combined with neural fields, some model-free methods [2, 34, 39] learn static heads from a few images with structure priors. Although many works [3, 12, 15, 19, 41, 49, 51] attempt to reconstruct dynamic heads from few videos, their parametric controls are difficult to predict from audio and limit the mouth motion granularity, making it hard for them to express the highly personalized 3D talking head with delicate audio-driven mouth motions achieved by InsTaG." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "text", + "content": "Some generative methods [13, 17, 30, 48, 54, 59] focus on reconstructing talking heads with 3D structure, however, they can hardly infer personalized talking style from only one image, leading to weak personalization and fidelity. Learning 3D talking heads with few-shot NeRF, recent works [24, 40, 53] introduce 2D-to-3D modules with pre-training but are often slow due to their overhead. Instead, with an identity-free motion prior and person-specific adaptation instead of expensive 2D-to-3D priors, our InS-TaG can rapidly learn a personalized audio-driven 3D talking head from few data while attaining fast inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 351, + 368, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 368, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 368, + 363 + ], + "type": "text", + "content": "3. InsTaG" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 369, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 441 + ], + "type": "text", + "content": "As shown in Figure 2, InsTaG synthesizes personalized 3D talking head based on Gaussian Splatting [22] (Sec. 3.1), consisting of an Identity-Free Pre-training (Sec. 3.2) strategy to fetch universal motion priors and a Motion-Aligned Adaptation strategy (Sec. 3.3) to learn a new identity. We describe the detailed designs in the following sections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 445, + 488, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 488, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 488, + 459 + ], + "type": "text", + "content": "3.1. 3DGS Talking Head Synthesizer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 461, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 557 + ], + "type": "text", + "content": "Our framework aims to learn personalized 3D talking head with fast inference and instant adaptation, therefore, a lightweight model with high representation ability is needed to be as the 3D synthesizer. Generally, existing methods can be categorized into image-based [30, 53, 54] and person-specific [16, 25, 28, 42]. The former often possesses generalizability but is relatively slow and weak in personalization, while the latter matches our requirements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 594 + ], + "type": "text", + "content": "Preliminaries. A person-specific synthesizer based on radiance fields can be formulated as a mapping " + }, + { + "bbox": [ + 313, + 558, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 558, + 554, + 594 + ], + "type": "text", + "content": " that from given conditions to a synthesized head image " + }, + { + "bbox": [ + 313, + 558, + 554, + 594 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 558, + 554, + 594 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 383, + 599, + 555, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 599, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 383, + 599, + 555, + 612 + ], + "type": "interline_equation", + "content": "\\mathcal {F}: (\\boldsymbol {a}, \\boldsymbol {e}, [ R, t ]; \\theta) \\rightarrow I \\tag {1}", + "image_path": "23433cdf56d06631092e70d5ae3545e7617d974962387cf01216f63db6984245.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "type": "text", + "content": "where the conditions include audio " + }, + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "type": "text", + "content": ", upper-face expression control " + }, + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "type": "text", + "content": " and camera pose " + }, + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "type": "inline_equation", + "content": "[R, t]" + }, + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "type": "text", + "content": " that represents the head pose, and the parameters of the radiance fields " + }, + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Under such an architecture, we require the motion learning to be entirely irrelevant to appearance for pre-training. Hence, we leverage the recent 3DGS-based synthesizer with a face-mouth decomposition [28], including a pair of structure fields and motion fields separately for a face branch and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10692" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "inside-mouth branch to synthesize the talking head image " + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": ". Structure Field. The structure field stores the static Gaussian primitives with parameters " + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "inline_equation", + "content": "\\theta = \\{\\mu, s, q, \\alpha, f\\}" + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": ". Both the face branch and inside-mouth branch obtain a private structure field for their own target structure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "text", + "content": "Motion Field. The motion field predicts a point-wise deformation " + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\delta = \\{\\Delta \\mu, \\Delta s, \\Delta q\\}" + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "text", + "content": " with a tri-plane hash encoder " + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "text", + "content": " [25] with a region attention (RA) module to store spatial relation, and an MLP decoder to predict deformation. For each query primitive with center " + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "text", + "content": ", the neural field " + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "text", + "content": " predicts its deformation " + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "text", + "content": " from the given condition feature set " + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "text", + "content": " including audio " + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "text", + "content": " and upper-face expression signal " + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 133, + 295, + 216 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 103, + 221, + 295, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 221, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 103, + 221, + 295, + 235 + ], + "type": "interline_equation", + "content": "\\delta = \\mathcal {D} (\\mu , \\mathbf {C}) = \\operatorname {M L P} (\\mathcal {H} (\\mu) \\oplus \\mathbf {C}), \\tag {2}", + "image_path": "5c10181dc4cd970bdd9fa6aec09ee11f0013c6593ed758e6e55a0b76f96270a5.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 240, + 186, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 186, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 186, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 240, + 186, + 251 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 55, + 240, + 186, + 251 + ], + "type": "text", + "content": " denotes concatenation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "text", + "content": "Rendering. During rendering, the motion field deforms the Gaussians " + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "text", + "content": " depending on the inputs " + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "text", + "content": ", and then a 2D image is rendered. This process happens in both the face and inside-mouth branches, and the output results are combined to get the final talking head image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 312, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 335 + ], + "type": "text", + "content": "In the following, unless necessary, we do not show the face-mouth decomposition for easier understanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 339, + 201, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 201, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 201, + 352 + ], + "type": "text", + "content": "3.2. Identity-Free Pre-training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 354, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 415 + ], + "type": "text", + "content": "Although with personalization knowledge, a short video clip is insufficient to train a talking head with personalized audio-motion mapping from scratch. To this end, we extract the common motion knowledge from long videos by pre-training as compensation for the new identity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 415, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 510 + ], + "type": "text", + "content": "In contrast to existing generative pre-training solutions [30, 40, 53] that allow multiple-identity inputs to train the same model, an identify-conflict problem lies in our method. Firstly, different identities can not coexist in one person-specific model. Secondly, since the model can not separate different people, the conflicting personalized motion hinders the model convergence and pollutes the pretrained module by overfitting the noises." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "content": "Strategy. To tackle this problem, we design storing the identity-dependent knowledge in a Personalized Field distributed to each training video, therefore detaching the identity influence of the learning of universal motion to pre-train a Universal Motion Field. Specifically, each Personalized Field consists of a personalized structure field with parameter " + }, + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\theta_{P}" + }, + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "content": " and a smaller motion field " + }, + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_P" + }, + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "content": " to retain the identity appearance and personalized motion of its owner video." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": "Universal Motion Field (UMF). Serving as a pre-trained module, UMF aims to predict a commonly correct facial motion that suits most identities. Queried by an arbitrary Gaussian, the field " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": " predicts the universal deformation " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\delta_U" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": " without personalization according to Eq. (2):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 97, + 671, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 671, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 97, + 671, + 295, + 685 + ], + "type": "interline_equation", + "content": "\\delta_ {U} = \\mathcal {U} (\\mu , \\mathbf {C}) = \\operatorname {M L P} \\left(\\mathcal {H} _ {U} (\\mu) \\oplus \\mathbf {C}\\right). \\tag {3}", + "image_path": "fcacd36d7e0285f0a81d09c14473a4d10970e1ec7a5987d47e77a7615d456177.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Personalized Fields. During training, Personalized Fields cooperate with one sharing Universal Motion Field to syn" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "thesize personalized talking heads, enabling the supervision from videos with different identities. Practically, given an audio feature " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " and expression signal " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " packaged by " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", the UMF " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " would predict an identity-free deformation queried by " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mu_P^i \\in \\theta_P^i" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", serving as a basic deformation for the " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "-th person. Meanwhile, the " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "-th personalized field predicts a residual motion " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\delta_P^i" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " as the personalized adjustment:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 345, + 159, + 553, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 159, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 345, + 159, + 553, + 172 + ], + "type": "interline_equation", + "content": "\\delta_ {P} ^ {i} = \\mathcal {D} _ {P} ^ {i} \\left(\\mu_ {P} ^ {i}, \\mathbf {C}\\right) = \\operatorname {M L P} \\left(\\mathcal {H} _ {P} ^ {i} \\left(\\mu_ {P} ^ {i}\\right) \\oplus \\mathbf {C}\\right). \\tag {4}", + "image_path": "bd3f8223f91658254c5e8609f3e3daeee0ad6584e474df65ee6f2008b2a93c27.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "text", + "content": "After that, the two deformations are combined to deform the personalized structure fields for rendering:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 341, + 205, + 553, + 221 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 205, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 341, + 205, + 553, + 221 + ], + "type": "interline_equation", + "content": "\\mathcal {R}: ([ R, t ], \\widetilde {\\theta} _ {P} ^ {i}) \\rightarrow I, \\quad \\widetilde {\\theta} _ {P} ^ {i} = \\theta_ {P} ^ {i} + \\delta_ {U} + \\delta_ {P} ^ {i}, \\tag {5}", + "image_path": "37f3d9230f0eb5512848666d4b8f9666eec60f3a49778a42cd28a2336fcd4278.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 226, + 505, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 505, + 238 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 226, + 505, + 238 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 226, + 505, + 238 + ], + "type": "text", + "content": " refers to the 3DGS rendering process." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 247, + 553, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 294 + ], + "type": "text", + "content": "Since then, videos with different talking identities can be used for pre-training with no identity conflicts, and the personalized motions are isolated from each other by staying at the personalized fields, detached from the universal ones." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": "Negative Contrast Loss. Despite the detachment of personalized and universal motions enabled, a new problem emerges: how to decide whether a motion should be personalized or universal? Without supervision, a native separation from gradient would be highly biased because of uncertainties like initialization and data disparity, making some personalized motions inappropriately stored into the universal field but too much universal knowledge omitted." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 408, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 491 + ], + "type": "text", + "content": "For this problem, we introduce a Negative Contrast Loss with two principles: 1) the ideal personalized motions are expected to be the minimum additional motions that enable smooth model convergence, thus the fewer the better; 2) subject to the first point, the personalized motions of one identity should be different from that of the others. Therefore, we design the loss by a truncated dot product:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 350, + 496, + 553, + 511 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 496, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 350, + 496, + 553, + 511 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C} \\left(\\Delta \\mu_ {P} ^ {i}, \\Delta \\mu_ {P} ^ {j}\\right) = \\mathbb {I} _ {\\text {t r u n c}} \\left(\\Delta \\mu_ {P} ^ {i} \\cdot \\Delta \\mu_ {P} ^ {j}\\right), \\tag {6}", + "image_path": "4004a4482a095d06c77b8fb973107f7030aab03105fb1d7490641bd1dd3c840c.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "inline_equation", + "content": "(\\Delta \\mu_P^i,\\Delta \\mu_P^j)" + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "text", + "content": " is a pair of personalized " + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "text", + "content": " deformation from " + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "text", + "content": " -th and " + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "text", + "content": " -th personalized fields, queried by the same " + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "inline_equation", + "content": "\\mu_{P}" + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "text", + "content": " in Eq. (4), and " + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_{trunc}" + }, + { + "bbox": [ + 313, + 516, + 553, + 553 + ], + "type": "text", + "content": " is a truncation function:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 373, + 556, + 553, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 556, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 373, + 556, + 553, + 582 + ], + "type": "interline_equation", + "content": "\\mathbb {I} _ {\\text {t r u n c}} (x) = \\left\\{ \\begin{array}{l l} x, & x > 0 \\\\ 0, & x \\leq 0 \\end{array} \\right.. \\tag {7}", + "image_path": "9bb4aa0d81b96e47d85819ccedf01483c3c3214c60af94a69f89edec522f71e7.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 586, + 553, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 553, + 657 + ], + "type": "text", + "content": "This negative-sample-based contrastive loss works from two aspects: 1) when the two participants are close in the 3D space, the loss would encourage them to be either different in direction or less in effect; 2) if the participants are already significantly different, or if one of the participants is close to 0, the loss stops pushing them too far." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "text", + "content": "By contrasting the negative personalized " + }, + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\Delta \\mu" + }, + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "text", + "content": " pairs, the proposed loss can encourage learned personalized motion to be diverse between the different personalized fields, therefore maximizing the filtering to get universal motions." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10693" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 209, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 209, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 209, + 85 + ], + "type": "text", + "content": "3.3. Motion-Aligned Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 186 + ], + "type": "text", + "content": "Based on the pre-trained UMF from Sec. 3.2, we introduce a Motion-Aligned Adaptation strategy to allow fast and high-quality learning of new-identity talking heads from a short video. Different from existing works [30, 40, 53] that take motions to guide image generation, we directly drive Gaussians by deformation in the final rendering for more realistic visual quality, bringing a challenge in precisely aligning the pre-trained motion to the new identity's head." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 189, + 296, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 296, + 273 + ], + "type": "text", + "content": "Motion Aligner. Due to the diversity of human faces, the pre-trained UMF can not natively match all unseen identities suitably. Except for the personalization that we aim to learn, this motion misalignment exists in two aspects: 1) the target facial structure may be biased with the implicit one from pre-training in UMF; 2) the motion of the target identity may have a scale gap from the universal motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "content": "To this end, we propose to wrap a Motion Aligner on the motion field to learn the adjustment. Specifically, a multiresolution positional encoder " + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_A" + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "content": " is first added to store the spatial information. Then, before querying the motion field by " + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "content": " from the structure field, a coordinate offset " + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\Delta \\mu_A" + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "content": " is predicted as bias adjustment:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 351, + 294, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 351, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 115, + 351, + 294, + 365 + ], + "type": "interline_equation", + "content": "\\Delta \\mu_ {A} = \\operatorname {M L P} \\left(\\mathcal {H} _ {A} (\\mu)\\right) \\in \\mathbb {R} ^ {3}. \\tag {8}", + "image_path": "965334fb7d86ea4996c529b7bd33b78bf9b0c85d4486b9b34dd1cf1b119dd87a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "text", + "content": "After getting the universal deformation " + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\delta_U" + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\Delta \\mu" + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "text", + "content": " part is multiplied by a scaling factor " + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\tau_A \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "text", + "content": " predicted in the same way as Eq. (8) for scale alignment, and finally output the adjusted deformation " + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\delta_U' = \\{\\Delta \\mu', \\Delta s, \\Delta q\\}" + }, + { + "bbox": [ + 55, + 370, + 296, + 419 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 425, + 294, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 425, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 85, + 425, + 294, + 437 + ], + "type": "interline_equation", + "content": "\\Delta \\mu^ {\\prime} = \\Delta \\mu \\times \\tau_ {A}, \\quad \\Delta \\mu \\in \\mathcal {U} (\\mu + \\Delta \\mu_ {A}, \\mathbf {C}). \\tag {9}", + "image_path": "5847e7ed4b0aed131ce3c915acbf605c865a1f716aa7cec019f089a3287782ff.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 444, + 276, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 276, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 276, + 456 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 444, + 276, + 456 + ], + "type": "inline_equation", + "content": "\\delta_U^\\prime" + }, + { + "bbox": [ + 55, + 444, + 276, + 456 + ], + "type": "text", + "content": " is used as the final deformation for rendering." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 459, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 590 + ], + "type": "text", + "content": "Face-Mouth Hook. As a deformation-based framework, the asynchronous inside and out mouth motions often deteriorate the modeling [28]. As in Sec. 3.1, we adopt a face-mouth decomposition [28] to solve this problem. It separates the modeling of the face (including lips) and inside of the mouth into two branches. However, when data is reduced, these two parts may fail to learn to move in harmony, causing misalignment with low visual quality. Without an existing mechanism to ensure the inter-alignment, we suggest setting a hook in the motion field to allow the mouth to actively align to the face, as shown in Figure 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Given a predicted deformation " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\delta_F" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " from the face branch that represents face motion, we attempt to fetch the motions on lips from it to guide the inside-mouth deformation " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\delta_M" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ". Considering the lip motion tends to cause the largest deformation amount of the whole face, we take the maximal and minimal " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mu^F \\in \\delta_F" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " (denoted as " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mu_{max/min}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ") as the motion cues, and calculate a distance " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mu_{dist} = \\Delta \\mu_{max} - \\Delta \\mu_{min}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " for a direct scale cue. Then, this set of features " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\varphi = \\{\\Delta \\mu_{max}, \\Delta \\mu_{min}, \\mu_{dist}\\}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " are input in the MLP decoder in the inside-mouth branch motion field to predict a deforma" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 553, + 163 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 163 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 163 + ], + "type": "image", + "image_path": "b3f88ffa2ca7c54db507cbd5d6e6a90a492b05bea8da256ca2cd72c2db0630a5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 167, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 201 + ], + "type": "text", + "content": "Figure 3. Illustration of Face-Mouth Hook. We hook the motion of the mouth to the generated face motion, allowing an alignment across two branches to enhance robustness under few training data." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 211, + 538, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 538, + 223 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 538, + 223 + ], + "type": "text", + "content": "tion " + }, + { + "bbox": [ + 313, + 211, + 538, + 223 + ], + "type": "inline_equation", + "content": "\\Delta \\mu" + }, + { + "bbox": [ + 313, + 211, + 538, + 223 + ], + "type": "text", + "content": ". The mouth motion field can be formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 381, + 227, + 553, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 227, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 381, + 227, + 553, + 242 + ], + "type": "interline_equation", + "content": "\\Delta \\mu^ {M} = \\mathcal {U} ^ {\\mathcal {M}} (\\mu , \\{\\boldsymbol {a}, \\varphi \\}). \\tag {10}", + "image_path": "543931570b6b3fedb79d3421796373858cb081d439a61bbaf25fb7963bd3f69c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "text", + "content": "Representing a structure difference, " + }, + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^{\\mathcal{M}}" + }, + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "text", + "content": " does not take " + }, + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "text", + "content": " as input and does not predict deformations on " + }, + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "text", + "content": "Moreover, considering the opening degree of the inside mouth should align with that of the lips, we hook the motion scaling prediction to the facial motion scale cue " + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "inline_equation", + "content": "\\mu_{dist}" + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "text", + "content": ". To achieve this, a scaling factor " + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "inline_equation", + "content": "\\tau_{M}" + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "text", + "content": " is predicted for each " + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "text", + "content": " by an MLP to get the final primitive-wise deformation:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 337, + 553, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 337, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 318, + 337, + 553, + 353 + ], + "type": "interline_equation", + "content": "\\mu_ {\\text {h o o k}} ^ {M} = \\Delta \\mu^ {M} \\times \\tau_ {M}, \\quad \\tau_ {M} = \\operatorname {M L P} \\left(\\mathcal {H} _ {U} ^ {M} (\\mu) \\oplus \\mu_ {\\text {d i s t}}\\right), \\tag {11}", + "image_path": "af963843daecfa4594de761045a2c1623406d46e0ac7147a9a64b75b52b4a179.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 358, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 553, + 371 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 358, + 553, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_U^M" + }, + { + "bbox": [ + 313, + 358, + 553, + 371 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 358, + 553, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 358, + 553, + 371 + ], + "type": "text", + "content": " of the UMF in the inside-mouth branch." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 372, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 430 + ], + "type": "text", + "content": "This hook technique establishes a connection between the motion fields to guide the mouth motion to align to the face, while the former's scale is also explicitly connected to the latter, bringing robust performance on the face-mouth alignment even when with only few training data." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "text", + "content": "Geometry Prior Regularizer. To further reduce the geometry ambiguity caused by the lack of view cover, we leverage the power of advanced human geometry estimator [23] to provide extra geometry cues for regularization. To solve the geometry degradation [26] and instability of 3DGS at unseen views [6], we add two regularizers on the depth and surface normal. Following [6] we first calculate the 2D depth " + }, + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "text", + "content": " and normal " + }, + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "text", + "content": " when synthesizing image " + }, + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "text", + "content": ". Then, the estimated monocular depth map " + }, + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "inline_equation", + "content": "\\check{D}" + }, + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "text", + "content": " and normal " + }, + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "inline_equation", + "content": "\\check{N}" + }, + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "text", + "content": " from the ground-truth image are used for regularization:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 554, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 554, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 318, + 554, + 553, + 586 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G e o} = \\lambda_ {D} L _ {D} (D, \\check {D}) + \\lambda_ {N} \\sum_ {i = 0} ^ {m} \\sum_ {j = 0} ^ {n} \\left(1 - N _ {i, j} \\cdot \\check {N} _ {i, j}\\right), \\tag {12}", + "image_path": "1e1b11bfa2de0cd0117d033ec974c180ac2936faf6fae511123658b2b8769c32.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 590, + 553, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 553, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 553, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 590, + 553, + 615 + ], + "type": "inline_equation", + "content": "L_{D}" + }, + { + "bbox": [ + 313, + 590, + 553, + 615 + ], + "type": "text", + "content": " is a scale-invariant depth loss, " + }, + { + "bbox": [ + 313, + 590, + 553, + 615 + ], + "type": "inline_equation", + "content": "N_{i,j}" + }, + { + "bbox": [ + 313, + 590, + 553, + 615 + ], + "type": "text", + "content": " denotes the normal at pixel " + }, + { + "bbox": [ + 313, + 590, + 553, + 615 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 590, + 553, + 615 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 590, + 553, + 615 + ], + "type": "inline_equation", + "content": "(m,n)" + }, + { + "bbox": [ + 313, + 590, + 553, + 615 + ], + "type": "text", + "content": " is the shape of image " + }, + { + "bbox": [ + 313, + 590, + 553, + 615 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 590, + 553, + 615 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 623, + 413, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 413, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 413, + 635 + ], + "type": "text", + "content": "3.4. Training Details" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "Photometric Loss. The training processes are mainly supervised by the image photometric loss " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_I" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": " between the synthesized image " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": " and its ground-truth " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "I_{GT}" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": ". Following 3DGS [22], " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_I" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": " consists of an L1 and a D-SSIM [46] terms. Pre-training. During pre-training, we collect " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": " videos as training data to simultaneously train the UMF. For each syn" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10694" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 294, + 272 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 294, + 272 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 294, + 272 + ], + "type": "image", + "image_path": "451012626d85b042f3ed1dc6bd2f5a5d4477c14c83333f009b17feb2fcf699f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 275, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 275, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 555, + 299 + ], + "type": "text", + "content": "Figure 4. Qualitative Comparison on Synchronization. Our method performs best in both lip-synchronization and visual quality. \"Real3DP\" and \"TG\" denote [54] and [28]. Better visualized with zoom-in. We recommend watching the supplementary video." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 300, + 70, + 548, + 272 + ], + "blocks": [ + { + "bbox": [ + 300, + 70, + 548, + 272 + ], + "lines": [ + { + "bbox": [ + 300, + 70, + 548, + 272 + ], + "spans": [ + { + "bbox": [ + 300, + 70, + 548, + 272 + ], + "type": "image", + "image_path": "e57164751141bdc6e749fb9db598847c9e386a38adaa087c3073e5abeac944d3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 305, + 555, + 446 + ], + "blocks": [ + { + "bbox": [ + 59, + 305, + 555, + 446 + ], + "lines": [ + { + "bbox": [ + 59, + 305, + 555, + 446 + ], + "spans": [ + { + "bbox": [ + 59, + 305, + 555, + 446 + ], + "type": "table", + "html": "
MethodsSettingRendering QualityMotion QualityEfficiency
PSNR (A/F) ↑ N/ALPIPS (A/F) ↓ 0 / 0SSIM ↑ 1.000LMD ↓ 0AUE-(L/U) ↓ 0 / 0Sync-C ↑ 8.897Training ↓ -FPS ↑ -
Ground TruthReal-time
EAT [13]One-shot- / 22.44- / 0.0600.712-1.309 / 1.0117.577-11.8X
Real3DPortrait [54]23.58* / 25.370.104* / 0.0540.8213.4071.148 / 1.0987.549-8.4X
RAD-NeRF [42]Trained28.36 / 25.510.048 / 0.0350.8353.5551.295 / 0.7322.2495 hours28.6✓
ER-NeRF [25]28.23 / 25.630.040 / 0.0310.8443.5411.327 / 0.4513.0742 hours30.8✓
GaussianTalk [7]From Scratch28.18 / 25.610.043 / 0.0320.8363.6471.379 / 0.4151.97067 min64.5✓
TalkingGaussian [28]28.32 / 26.010.041 / 0.0280.8433.5881.158 / 0.3163.55631 min114.2✓
DFRF [40]Adaptation & Few-shot28.48 / 26.260.066 / 0.0340.8593.4361.247 / 0.6833.9579 hours0.03X
GeneFace [52]27.23 / 25.080.052 / 0.0360.8233.5101.310 / 0.6843.6517 hours18.1X
MimicTalk [53]24.69* / 26.270.075* / 0.0310.8533.4890.958 / 0.7756.92616 min8.2X
InsTaG (Ours)28.86 / 26.320.039 / 0.0260.8613.1670.926 / 0.3135.31813 min82.5✓
", + "image_path": "9a217e33877e8d54cc32d1ac34bbb57c08e98a6e83e10df91b1be5be8b476e87.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 451, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 55, + 451, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 555, + 473 + ], + "type": "text", + "content": "Table 1. Quantitative Comparison in self-reconstruction setting with 5s training data. InsTaG achieves the best image quality and personalized dynamics, with fast adaption and real-time inference. We mark the best and second-best results. (*: with synthesized torso.)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "content": "thesized image " + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "inline_equation", + "content": "I^i" + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "content": "-th video, we first calculate its photometric loss, and then calculate the Negative Contrast Loss to each other. The pre-training loss for the " + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "content": "-th video is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 60, + 522, + 295, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 522, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 60, + 522, + 295, + 555 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p r e} ^ {i} = \\mathcal {L} _ {I} \\left(I ^ {i}, I _ {G T} ^ {i}\\right) + \\lambda_ {C} \\sum_ {j = 1, j \\neq i} ^ {k} \\mathcal {L} _ {C} \\left(\\Delta \\mu_ {P} ^ {i} \\cdot \\Delta \\mu_ {P} ^ {j}\\right). \\tag {13}", + "image_path": "a25f9ed10144c5dcbbf623a679f9af75787d2cc4272ad707534217e018258fce.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 556, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 604 + ], + "type": "text", + "content": "Adaptation. For adaptation, we initialize a person-specific model with the pre-trained UMF. A warm-up stage first runs without a Motion Aligner and geometry loss for a start. After that, the model is trained with full loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 101, + 609, + 295, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 609, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 101, + 609, + 295, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d a} = \\mathcal {L} _ {I} (I, I _ {G T}) + \\mathcal {L} _ {G e o} (D, N). \\tag {14}", + "image_path": "fa0400b6c38adb95801e81efcf6ff60503ae9842d2e3a250555bc2251fbb1bef.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 624, + 137, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 137, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 137, + 638 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "Dataset. In the experiments, we collect a long-video corpus with 5 long speaking videos from [25, 52] for pre-training, and follow prevailing settings [16, 25, 28, 52] to take 4 videos from publicly-released sets [40, 52] for the test, with no identity overlap of the two parts. The videos have an average length of about 800-8000 frames in 25 FPS with a" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 542, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 542, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 542, + 498 + ], + "type": "text", + "content": "center portrait, all are cropped and resized to " + }, + { + "bbox": [ + 313, + 486, + 542, + 498 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 486, + 542, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "type": "text", + "content": "Implementation Details. Our method is implemented on PyTorch. The pre-training stage lasts 150,000 iterations and adaptation lasts 10,000. An AdamW [32] optimizer is used with learning rates of 5e-3 and 5e-4 separately for the grid and neural networks to train UMF. " + }, + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{D}" + }, + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{N}" + }, + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{C}" + }, + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "type": "text", + "content": " are set to 1e-2, 1e-3 and 1. All experiments are performed on RTX 3080 Ti GPUs. The overall pre-training takes about 5 hours. The head pose is estimated through BFM [36]. A DeepSpeech model [18] is used to extract basic audio features following previous works [16, 25, 28]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": "Comparison Baselines. To compare with the state-of-the-art methods, we take the baselines from the relevant person-specific methods DFRF [40], RAD-NeRF [42], GeneFace [52], ER-NeRF [25], TalkingGaussian [58], GaussianTalker [7] and a concurrent method MimicTalk [53]. We also involve the one-shot EAT [13] and Real3DPortrait [54]. All the methods are reproduced by their official codebase. If officially supported, we use DeepSpeech as audio extractor." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10695" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 68, + 552, + 232 + ], + "blocks": [ + { + "bbox": [ + 62, + 68, + 552, + 232 + ], + "lines": [ + { + "bbox": [ + 62, + 68, + 552, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 68, + 552, + 232 + ], + "type": "image", + "image_path": "14493a5e2c805b7245f39c0d37bfb7f63beb9772db764f24e3bc7f7d048b6601.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparison on Reconstruction Quality. Our method performs the best in rendering photorealistic talking heads with fine details. \"Real3DP\" and \"TG\" denote Real3DPortrait [54] and TalkingGaussian [28]. Better visualized with zoom-in." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 263, + 553, + 340 + ], + "blocks": [ + { + "bbox": [ + 59, + 263, + 553, + 340 + ], + "lines": [ + { + "bbox": [ + 59, + 263, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 59, + 263, + 553, + 340 + ], + "type": "table", + "html": "
MethodType5s10s20s
PSNR ↑AUE-L ↓Sync ↑PSNR ↑AUE-L ↓Sync ↑PSNR ↑AUE-L ↓Sync ↑
ER-NeRF [25]From Scratch28.2351.3273.07428.7281.0783.24029.2870.9414.638
TalkingGaussian [28]28.3211.1583.55629.1301.0694.66329.5361.0734.776
DFRF [40]Adaptation28.4811.2473.95728.8811.0724.21029.2531.1374.333
GeneFace [52]28.2271.3103.65127.9971.0364.86928.8251.0334.772
InsTaG (Ours)28.8630.9265.31829.4210.8815.63630.3520.7105.764
", + "image_path": "4e0fcca028354b5e541eefcfe785bb2cd6457208d54ef9edb8fda253586571d6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 376, + 294, + 468 + ], + "blocks": [ + { + "bbox": [ + 55, + 345, + 555, + 368 + ], + "lines": [ + { + "bbox": [ + 55, + 345, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 555, + 368 + ], + "type": "text", + "content": "Table 2. Quantitative Comparison in self-reconstruction setting to person-specific baselines with different data amounts. Our method performs best in various data scenarios, exhibiting robustness. The best and second-best results are in bold and underline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 376, + 294, + 468 + ], + "lines": [ + { + "bbox": [ + 59, + 376, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 59, + 376, + 294, + 468 + ], + "type": "table", + "html": "
MethodsAudio English, MaleAudio German, Female
Sync-E ↓ 10.443Sync-C ↑ 4.342Sync-E ↓ 9.942Sync-C ↑ 4.949
Full Baseline [28]
ER-NeRF [25]11.7113.04311.5022.853
TalkingGaussian [28]10.6754.31710.8983.708
DFRF [40]10.5154.72610.4824.329
GeneFace [52]9.9954.73411.0294.162
InsTaG (Ours)9.8864.8289.7334.990
", + "image_path": "f449b687bed18ed206ff2f2db282ddf0080ca83cf958b7b980321cd805c1d88d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 295, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 495 + ], + "type": "text", + "content": "Table 3. Comparison in cross-domain setting with 20s training data. We take TalkingGaussian [28] with full data for reference." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 502, + 130, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 130, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 130, + 514 + ], + "type": "text", + "content": "4.1. Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "text", + "content": "Comparison Settings. The quantitative comparison contains two settings: 1) Self-reconstruction setting. We split each of the 4 adaptation videos into few-second training sets and a test set of at least 12s, and use the unseen test data as input and ground truth for evaluation. 2) Cross-domain setting. We use an unseen English male audio from NVP [44] and a German female one from [16] to test the cross-domain generalizability. Considering some baselines with failure cases, we use 20s from a certain video to train per method to represent its best performance for testing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": "Metrics and Measurements. We employ PSNR, LPIPS [61], and SSIM [46] to evaluate image quality. Especially, we report PSNR and LPIPS separately on an all image (\"A\") and only the aligned face (\"F\"). SSIM is only measured on the face. For lip synchronization, we use LMD [4] and SyncNet [8, 9] confidence (Sync-C) and error (Sync-E). Be" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 378, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 450 + ], + "type": "text", + "content": "sides, action unit [10] error (AUE) is used on upper-face (AUE-U) and lower-face (AUE-L) following [28], in which AUE-L can reflect the similarity of the personalized talking actions from the target. We also statistic the training time (fine-tuning time for few-shot or adaptation-based methods) and inference FPS to evaluate efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": "Quantitative Evaluation. 1) Table 1 reports the results of self-reconstruction setting with 5s training data. Our method achieves the best image quality, demonstrating our ability to faithfully generate a realistic talking head. Inheriting the traditions of large corpus pre-training, the one-shot methods [13, 54] and MimicTalk [53] achieve very high Sync scores but are relatively slow. Also getting a high Sync, InsTaG achieves the best in LMD and AUE-L, showing our advantage in preserving personalization. While all baselines can not simultaneously achieve high synchronization and real-time inference, our method produces the best overall performance including the fastest adaptation and high speed. 2) When data increases, the results in Table 2 show that InsTaG still outperforms other person-specific models in all aspects, demonstrating our wider application value and robustness to various scenarios. 3) In the results from Table 3 in cross-domain setting, where the methods are tested on two difficult cross-domain scenarios, the advantages of pre-training are further amplified. GeneFace [52] and DFRF [40] surpass all models trained from scratch. At the same time, InsTaG exceeds the baseline [28] trained on about 1-minute full training data, demonstrating the out" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10696" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 553, + 133 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 553, + 133 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 553, + 133 + ], + "type": "table", + "html": "
MethodsER-NeRF [25]TalkingGaussian [28]GaussianTalker [7]DFRF [40]GeneFace [52]MimicTalk [53]InsTaG (Ours)
Lip-sync Accuracy2.001.901.503.302.704.804.50
Image Quality3.403.704.002.903.103.704.80
Identity Preserving2.401.501.702.903.004.204.60
Video Realness1.401.501.402.801.703.203.90
", + "image_path": "4313efe0e902472506b0984c5df0320e7ead22d496cde95134172915877cf6d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 158, + 294, + 260 + ], + "blocks": [ + { + "bbox": [ + 59, + 135, + 550, + 147 + ], + "lines": [ + { + "bbox": [ + 59, + 135, + 550, + 147 + ], + "spans": [ + { + "bbox": [ + 59, + 135, + 550, + 147 + ], + "type": "text", + "content": "Table 4. User Study. The rating is in the range of 1-5, higher denotes better. The best and second-best results are in bold and underline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 158, + 294, + 260 + ], + "lines": [ + { + "bbox": [ + 60, + 158, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 60, + 158, + 294, + 260 + ], + "type": "table", + "html": "
Pre-training PFieldNCLossAdaptation AlignerHookPSNR ↑LPIPS ↓LMD ↓Sync-C ↑
28.660.0443.7342.019
✓28.760.0403.2394.816
✓✓28.590.0413.2235.124
✓✓✓28.690.0403.2745.233
✓✓✓28.660.0403.2575.086
✓✓✓28.790.0403.2264.994
✓✓✓✓28.860.0393.1675.318
", + "image_path": "e526eaca366b3a275bf2dbb1e7827758ffee23a61bbb9ef254f2ab22ff2d3e47.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 293, + 280, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 280, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 280, + 304 + ], + "type": "text", + "content": "standing robustness and generalizability of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 305, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 305, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 295, + 483 + ], + "type": "text", + "content": "Qualitative Evaluation. To qualitatively evaluate the results, we show the generated sequences and rendering details from self-reconstruction setting. In Figure 4, it can be observed that the models trained from scratch (ER-NeRF, TG) [27, 28] behave poorly in synchronization (yellow box), while the adaptation-based methods (DFRF, GeneFace, MimicTalk) [40, 52, 53] perform better but generate blurry results (orange circle). Referring to Figure 5, while all methods synthesize either blurs or heavily distorted details (red arrow), our method can reconstruct photorealistic appearance and delicate details. By benefitting from the advantages of both pre-training and person-specific training, InsTaG achieves the best in both synchronization and image quality. We strongly recommend watching our supplementary video for an intuitive comparison." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 484, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 605 + ], + "type": "text", + "content": "User Study. To better judge the quality in real human-participated scenarios, we conducted a user study with 28 generated talking head videos from 7 methods with 5s training data. We invited 10 attendees to rate each anonymous method from four aspects: (1) Lip-sync Accuracy; (2) Video Realness; (3) Identity Preserving and (4) Video Realness. The results in Table 4 show that our InsTaG achieves the best in three aspects and the second-best in lip-sync, which is only after the concurrent MimicTalk [53], demonstrating our outstanding visual quality via human judgment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 611, + 149, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 149, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 149, + 624 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "text", + "content": "We conduct the ablation study to prove the effectiveness of our contributions. The results are reported in Table 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Identity-Free Pre-training. In the first row, we show the results pre-trained without Personalized Fields (PField), which produces bad synchronization due to the identity-conflict problem. After applied PField, the pre-training works with obvious effect. Combined with PField, Nega" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 318, + 158, + 553, + 200 + ], + "blocks": [ + { + "bbox": [ + 55, + 264, + 294, + 287 + ], + "lines": [ + { + "bbox": [ + 55, + 264, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 294, + 287 + ], + "type": "text", + "content": "Table 5. Ablation Study. We verify the components under 5s self-reconstruction setting, showing the effect of our contributions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 158, + 553, + 200 + ], + "lines": [ + { + "bbox": [ + 318, + 158, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 318, + 158, + 553, + 200 + ], + "type": "table", + "html": "
BackboneHookPSNR ↑LPIPS ↓LMD ↓Sync-C ↑
Single Branch-28.590.0493.2334.675
Face-Mouth Decom.×28.690.0403.2745.233
Face-Mouth Decomp.✓28.860.0393.1675.318
", + "image_path": "f5f41c69e4fb401cf13f2f3d7c5df5ce6dff4aa9b93deec51e762cf0102963f1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 204, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 237 + ], + "type": "text", + "content": "Table 6. Ablation Study on Mouth Motion Representation. Introducing FMD [28] can effectively improve lip-synchronization, while our FM hook technique further boosts image quality." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 321, + 247, + 431, + 308 + ], + "blocks": [ + { + "bbox": [ + 321, + 247, + 431, + 308 + ], + "lines": [ + { + "bbox": [ + 321, + 247, + 431, + 308 + ], + "spans": [ + { + "bbox": [ + 321, + 247, + 431, + 308 + ], + "type": "image", + "image_path": "66725b671ecca2c21363a900d66a6c5a43d2f32026d9990a819f350a1becf75b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 311, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 344 + ], + "type": "text", + "content": "Figure 6. Visualization Results of Face-Mouth Hook. FMD improves performance but also causes misalignment when with inadequate data, which can be well solved by our Face-Mouth Hook." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 439, + 247, + 548, + 308 + ], + "blocks": [ + { + "bbox": [ + 439, + 247, + 548, + 308 + ], + "lines": [ + { + "bbox": [ + 439, + 247, + 548, + 308 + ], + "spans": [ + { + "bbox": [ + 439, + 247, + 548, + 308 + ], + "type": "image", + "image_path": "eb8efee902a7fcfda286531204fa94ef8221898c476e788be6ad4913b94c9d7d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 354, + 553, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 400 + ], + "type": "text", + "content": "tive Contrast Loss (NCLoss) shows its impact on lip-sync improvement, offering higher Sync-C. This demonstrates the necessity of PFiled in enabling the pre-training and the effect of NCLoss on encouraging knowledge collection." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 401, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 509 + ], + "type": "text", + "content": "Motion-Aligned Adaptation. We apply our Motion Aligner (Aligner) and Face-Mouth Hook (FM Hook) to illustrate their effect on adaptation. Although the performance is improved by applying our pre-training strategy, the Aligner can bring extra enhancement by providing adaptation adjustments. Furthermore, the Hook works in both visual quality and dynamics by correcting inside-mouth misalignment. Combining these two adaptation techniques, the performance achieves the best among all settings." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": "Mouth Motion Representation. To further explain the necessity of Face-Mouth Decomposition (FMD) [28] and our FM Hook, we take studies by using one single branch to represent the two parts. Results in Table 6 show the effect of FMD. However, simply introducing it would cause significant misalignment that lowers the visual quality, as in Figure 6. By applying our FM Hook, this drawback can be well solved to benefit FMD with robust performance, generating synchronized talking heads with high visual quality." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 624, + 441, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 441, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 441, + 636 + ], + "type": "text", + "content": "5. Ethical Consideration" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "We hope our method can promote the healthy development of digital industries. However, it is important to note that it could be misused for malicious purposes, potentially causing negative impacts. As part of our responsibility, we will assist in developing deepfake detection tools. We recommend the responsible use of this and all similar techniques." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10697" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 171, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 228 + ], + "type": "text", + "content": "[1] James Booth, Epameinondas Antonakos, Stylianos Ploumpis, George Trigeorgis, Yannis Panagakis, and Stefanos Zafeiriou. 3d face morphable models\" in-the-wild\". In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 48-57, 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 228, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 306 + ], + "type": "text", + "content": "[2] Marcel C Bühler, Kripasindhu Sarkar, Tanmay Shah, Gengyan Li, Daoye Wang, Leonhard Helminger, Sergio Orts-Escalano, Dmitry Lagun, Otmar Hilliges, Thabo Beeler, et al. Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3402-3413, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 308, + 295, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 308, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 61, + 308, + 295, + 361 + ], + "type": "text", + "content": "[3] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, et al. Authentic volumetric avatars from a phone scan. ACM Transactions on Graphics (TOG), 41(4):1-19, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 363, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 296, + 418 + ], + "type": "text", + "content": "[4] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Computer Vision-ECCV 2018: 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part VII 15, pages 538-553. Springer, 2018. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 296, + 475 + ], + "type": "text", + "content": "[5] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 476, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 296, + 521 + ], + "type": "text", + "content": "[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "text", + "content": "[7] Kyusun Cho, JoungBin Lee, Heeji Yoon, Yeobin Hong, Jaehoon Ko, Sangjun Ahn, and Seungryong Kim. Gaussian talker: Real-time talking head synthesis with 3d gaussian splatting. In ACM Multimedia 2024, 2024. 2, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 567, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 622 + ], + "type": "text", + "content": "[8] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 87-103. Springer, 2017. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "text", + "content": "[9] Joon Son Chung and Andrew Zisserman. Out of time: Automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "text", + "content": "[10] Paul Ekman and Wallace V. Friesen. Facial Action Coding System: Manual. Palo Alto: Consulting Psychologists Press, 1978. 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "In this work, we are supported by the National Natural Science Foundation of China 62276016, 62372029. Lin Gu is supported by JST Moonshot R&D Grant Number JP-MJMS2011 Japan." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[11] Tony Ezzat, Gadi Geiger, and Tomaso Poggio. Trainable videorealistic speech animation. ACM Transactions on Graphics (TOG), 21(3):388-398, 2002. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "type": "text", + "content": "[12] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 162, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 215 + ], + "type": "text", + "content": "[13] Yuan Gan, Zongxin Yang, Xihang Yue, Lingyun Sun, and Yi Yang. Efficient emotional adaptation for audio-driven talking-head generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22634-22645, 2023. 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 271 + ], + "type": "text", + "content": "[14] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1155-1164, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 272, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 337 + ], + "type": "text", + "content": "[15] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rinz, Lourdes Agapito, and Matthias Nießner. Mononphm: Dynamic head reconstruction from monocular videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10747-10758, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 338, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 555, + 392 + ], + "type": "text", + "content": "[16] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 394, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 448 + ], + "type": "text", + "content": "[17] Siddharth Gururani, Arun Mallya, Ting-Chun Wang, Rafael Valle, and Ming-Yu Liu. Space: Speech-driven portrait animation with controllable expression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20914-20923, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "type": "text", + "content": "[18] Awni Hannun, Carl Case, Jared Casper, Bryan Catanzaro, Greg Diamos, Erich Elsen, Ryan Prenger, Sanjeev Satheesh, Shubho Sengupta, Adam Coates, et al. Deep speech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:1412.5567, 2014. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "type": "text", + "content": "[19] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 558, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 553, + 603 + ], + "type": "text", + "content": "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 604, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 645 + ], + "type": "text", + "content": "[21] Amir Jamaludin, Joon Son Chung, and Andrew Zisserman. You said that?: Synthesising talking faces from audio. International Journal of Computer Vision, 127:1767-1779, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[23] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10698" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "type": "text", + "content": "[24] Dongze Li, Kang Zhao, Wei Wang, Bo Peng, Yingya Zhang, Jing Dong, and Tieniu Tan. Ae-nerf: Audio enhanced neural radiance field for few shot talking head synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3037–3045, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 295, + 216 + ], + "type": "text", + "content": "[25] Jiahe Li, Jiawei Zhang, Xiao Bai, Jun Zhou, and Lin Gu. Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7568-7578, 2023. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "type": "text", + "content": "[26] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "type": "text", + "content": "[27] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Jun Zhou, and Lin Gu. Er-nerf++: Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. Information Fusion, 110:102456, 2024. 1, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "text", + "content": "[28] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Talkinggaussian: Structure-persistent 3d talking head synthesis via gaussian splatting. In European Conference on Computer Vision, pages 127-145. Springer, 2025. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "type": "text", + "content": "[29] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 502 + ], + "type": "text", + "content": "[30] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhigang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17969-17978, 2023. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 503, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 569 + ], + "type": "text", + "content": "[31] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 106-125. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "text", + "content": "[32] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "type": "text", + "content": "[33] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "type": "text", + "content": "[34] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, pages 179-197. Springer, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 163 + ], + "type": "text", + "content": "[36] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. Ieee, 2009. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 164, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 229 + ], + "type": "text", + "content": "[37] Ziqiao Peng, Wentao Hu, Yue Shi, Xiangyu Zhu, Xiaomei Zhang, Hao Zhao, Jun He, Hongyan Liu, and Zhaoxin Fan. Synctalk: The devil is in the synchronization for talking head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 666-676, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 285 + ], + "type": "text", + "content": "[38] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 286, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 342 + ], + "type": "text", + "content": "[39] Eduard Ramon, Gil Triginer, Janna Escur, Albert Pumarola, Jaime Garcia, Xavier Giro-i Nieto, and Francesc Moreno-Noguer. H3d-net: Few-shot high-fidelity 3d head reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5620-5629, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 343, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 410 + ], + "type": "text", + "content": "[40] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII, pages 666-682. Springer, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 553, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 456 + ], + "type": "text", + "content": "[41] Luchuan Song, Pinxin Liu, Lele Chen, Guojun Yin, and Chenliang Xu. Tri²-plane: Thinking head avatar via feature pyramid. In European Conference on Computer Vision, pages 1–20. Springer, 2025. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 456, + 553, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 512 + ], + "type": "text", + "content": "[42] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 513, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 553, + 568 + ], + "type": "text", + "content": "[43] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 569, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 635 + ], + "type": "text", + "content": "[44] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 716-731. Springer, 2020. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 635, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 712 + ], + "type": "text", + "content": "[45] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI, pages 700–717. Springer, 2020. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10699" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 185 + ], + "type": "text", + "content": "[47] Olivia Wiles, A Sophia Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In Computer Vision-ECCV 2018: 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XIII 15, pages 690-706. Springer, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "text", + "content": "[48] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 242, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 294, + 297 + ], + "type": "text", + "content": "[49] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 298, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 363 + ], + "type": "text", + "content": "[50] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 421 + ], + "type": "text", + "content": "[51] Xihe Yang, Xingyu Chen, Daiheng Gao, Shaohui Wang, Xiaoguang Han, and Baoyuan Wang. Have-fun: Human avatar reconstruction from few-shot unconstrained images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 742-752, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "text", + "content": "[52] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, Jinzheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 478, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 544 + ], + "type": "text", + "content": "[53] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiangwei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, Chen Zhang, Xiang Yin, Zejun Ma, and Zhou Zhao. Mimictalk: Mimicking a personalized and expressive 3d talking face in few minutes. Advances in neural information processing systems, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 545, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 601 + ], + "type": "text", + "content": "[54] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiawei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "type": "text", + "content": "[55] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "text", + "content": "[56] Hongyun Yu, Zhan Qu, Qihang Yu, Jianchuan Chen, Zhonghua Jiang, Zhiwen Chen, Shengyu Zhang, Jimin Xu, Fei Wu, Chengfei Lv, et al. Gaussian talker: Speaker-specific talking head synthesis via 3d gaussian splattering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3548-3557, 2024. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 406 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[57] Yu Yu, Kenneth Alberto Funes Mora, and Jean-Marc Odobez. Headfusion: " + }, + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": " head pose tracking combining 3d morphable model and 3d reconstruction. IEEE transactions on pattern analysis and machine intelligence, 40(11): 2653-2667, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[58] Zhentao Yu, Zixin Yin, Deyu Zhou, Duomin Wang, Finn Wong, and Baoyuan Wang. Talking head generation with probabilistic audio-to-visual diffusion priors. arXiv preprint arXiv:2212.04248, 2022. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "text", + "content": "[59] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9459-9468, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "type": "text", + "content": "[60] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3867-3876, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "type": "text", + "content": "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 342, + 553, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 406 + ], + "type": "text", + "content": "[62] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10700" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/cb8122c4-a3cd-447b-9677-ef086abe8658_content_list.json b/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/cb8122c4-a3cd-447b-9677-ef086abe8658_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..51997b1f9a60e80726a5482f711dd880948bdb4d --- /dev/null +++ b/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/cb8122c4-a3cd-447b-9677-ef086abe8658_content_list.json @@ -0,0 +1,1754 @@ +[ + { + "type": "text", + "text": "Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 215, + 128, + 782, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhao Dong $^{1*}$ Zuyan Liu $^{2,3*}$ Hai-Long Sun $^{3,4}$ Jingkang Yang $^{1}$ Winston Hu $^{2}$ Yongming Rao $^{2\\boxtimes}$ Ziwei Liu $^{1\\boxtimes}$ $^{1}$ S-Lab, Nanyang Technological University $^{2}$ Tencent \n $^{3}$ Tsinghua University $^{4}$ Nanjing University \nhttps://github.com/dongyh20/Insight-V", + "bbox": [ + 236, + 202, + 759, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Language Models (LLMs) demonstrate enhanced capabilities and reliability by reasoning more, evolving from Chain-of-Thought prompting to product-level solutions like OpenAI o1. Despite various efforts to improve LLM reasoning, high-quality long-chain reasoning data and optimized training pipelines still remain inadequately explored in vision-language tasks. In this paper, we present Insight-V, an early effort to 1) scalably produce long and robust reasoning data for complex multi-modal tasks, and 2) an effective training pipeline to enhance the reasoning capabilities of multi-modal large language models (MLLMs). Specifically, to create long and structured reasoning data without human labor, we design a two-step pipeline with a progressive strategy to generate sufficiently long and diverse reasoning paths and a multi-granularity assessment method to ensure data quality. We observe that directly supervising MLLMs with such long and complex reasoning data will not yield ideal reasoning ability. To tackle this problem, we design a multi-agent system consisting of a reasoning agent dedicated to performing long-chain reasoning and a summary agent trained to judge and summarize reasoning results. We further incorporate an iterative DPO algorithm to enhance the reasoning agent's generation stability and quality. Based on the popular LLaVA-NeXT model and our stronger base MLLM, we demonstrate significant performance gains across challenging multi-modal benchmarks requiring visual reasoning. Benefiting from our multi-agent system, Insight-V can also easily maintain or improve performance on perception-focused multi-modal tasks.", + "bbox": [ + 88, + 358, + 485, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 823, + 223, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The development of artificial general intelligence requires models that can seamlessly understand and respond to multi-", + "bbox": [ + 89, + 848, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/909c928c7d743da8c8fe481207aa002a384c19e57aa64d5f63df3e8d9ce7c3a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 327, + 919, + 551 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/359944c99fc5a9953c0ada46ead258b10e388777a65b063fc90ca5fe55d907a2.jpg", + "image_caption": [ + "Figure 1. Illustration and Performance of Insight-V. Insight-V consists of two agents, one dedicated to reasoning and the other to summarization, driving significant improvements in performance across various visual reasoning benchmarks." + ], + "image_footnote": [], + "bbox": [ + 519, + 553, + 921, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "modal data. Recent advancements in Large Language Models (LLMs) [9, 36, 37, 41] and Multi-modal LLMs (MLLMs) [6, 21-23, 29, 38, 49] have significantly facilitated this progress across various fields, ranging from common question-answering [6, 17, 22, 27, 38] to autonomous driving [32, 42] and robotics [8, 47]. Despite the substantial progress made in enhancing the performance of MLLMs on a wide range of tasks, enabling MLLMs to perform human-level reasoning remains a key challenge. This area remains underexplored and has yet to fully realize its potential.", + "bbox": [ + 509, + 718, + 908, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing efforts [44, 48] to enhance the reasoning capabilities of LLMs through long-chain reasoning have", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors contributed equally to this research. Corresponding authors.", + "bbox": [ + 107, + 886, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9062", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "demonstrated considerable progress, largely benefiting from the availability of structured, high-quality data and well-established training pipelines. In contrast, teaching MLLMs to perform long-chain visual reasoning remains a significant challenge, primarily due to the lack of large-scale, high-quality datasets and efficient and effective training strategies. Compared to text-only data, visual reasoning data is not only more expensive to collect but also requires significant human labor for detailed annotation and validation, due to the absence of an effective data generation pipeline. Moreover, while previous work [58] has demonstrated that directly applying chain-of-thought [44] reasoning can improve the capabilities of MLLMs, other research [56, 57] suggests that current training approaches have limited effectiveness in enhancing CoT reasoning. This highlights the inability of current MLLMs to leverage visual cues for precise step-by-step problem-solving, emphasizing the need for an effective training procedure that enables MLLMs to reason in detail while maintaining clear visual perception.", + "bbox": [ + 93, + 90, + 482, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose Insight-V, which incorporates two innovative designs to enhance reasoning capabilities. First, we introduce a data generation pipeline consisting of two key steps: a progressive strategy to generate structured, long-chain reasoning data with diverse reasoning paths, and a multi-granularity assessment system to evaluate and score these paths at different levels. Through automatic generation, assessment, and ranking strategies, the pipeline effectively operates without the need for human labor and makes the reasoning dataset more scalable for enhancing reasoning capabilities. To further improve MLLM reasoning beyond data scaling, we design a multi-agent system, as illustrated in Figure 1, that decomposes the problem-solving process into two distinct steps: reasoning and summarization. The reasoning agent generates a detailed reasoning process for the input query, while the summarization agent identifies key information within the reasoning process and selectively answers the question. To refine the quality of the reasoning, we employ an iterative DPO approach to enhance reasoning capabilities. The two agents collaborate to further improve the reasoning quality. Our findings demonstrate that this system significantly enhances the performance of various MLLMs across a broad range of visual reasoning benchmarks.", + "bbox": [ + 93, + 383, + 482, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate Insight-V by integrating our system into the widely used LLaVA-NeXT [22] model. To further highlight the potential of our design in advancing state-of-the-art models, we construct a robust base MLLM and apply our system to demonstrate its ability to enhance SOTA models across various visual reasoning benchmarks. Notably, Insight-V improves LLaVA-NeXT in an average performance of $7.0\\%$ across seven challenging visual reasoning benchmarks, while a $2.9\\%$ improvement is observed with the strong base MLLM, underscoring the effectiveness and generalizability", + "bbox": [ + 93, + 750, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of Insight-V.", + "bbox": [ + 517, + 92, + 599, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, Insight-V offers 1) a scalable data generation pipeline for long-chain, high-quality reasoning data, 2) a multi-agent system that decomposes visual reasoning tasks into reasoning and summarization, and 3) a two-stage training pipeline to enhance visual reasoning capabilities. Together, these contributions address key challenges in visual reasoning, providing a solid foundation for future research in MLLM reasoning.", + "bbox": [ + 517, + 108, + 905, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 517, + 243, + 651, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision-Language Reasoning", + "text_level": 1, + "bbox": [ + 517, + 268, + 763, + 285 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in MLLMs [1, 17, 20-23, 27, 29, 38] have equipped these models with robust capabilities across diverse domains, including visual understanding [20, 38], mathematics [19], college-level questions [6], and scientific inquiries. In visual understanding, most research [18, 22, 28, 43, 46] emphasizes fine-grained detail analysis and localization, training models to perform visual reasoning with tailored datasets to enhance interpretive capabilities. For mathematics and expert-level reasoning, existing methods [11, 56, 57] predominantly derive from Chain-of-Thought [44] approaches, training MLLMs to generate step-by-step reasoning across various subjects. However, these approaches often focus primarily on improving dataset quality through Chain-of-Thought, overlooking the importance of structured reasoning paths and extended reasoning chains in advancing model reasoning capabilities. Additionally, significant challenges arise when relying on a single model to manage the entire reasoning process for complex tasks, underscoring the need for a multi-agent system to decompose and enhance this process. In this work, we tackle these challenges by introducing a scalable reasoning data generation pipeline and implementing a multi-agent system for reasoning and summarization decomposition, enhancing the overall reasoning capabilities of existing MLLMs.", + "bbox": [ + 516, + 292, + 905, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Vision-Language Alignment", + "text_level": 1, + "bbox": [ + 517, + 666, + 763, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To align the model more closely with human preferences, several alignment techniques are employed for MLLMs. A widely used approach is Reinforcement Learning from Human Feedback [2] (RLHF), which iteratively refines the model's responses based on human feedback, enhancing both response quality and interpretability. To further improve MLLM capabilities, Direct Preference Optimization [39] (DPO) is introduced to simplify the alignment process. By directly training on human preference data, DPO optimizes the model's outputs to better match human-selected responses. However, traditional DPO is primarily focused on offline scenarios, and as the model evolves, the effectiveness of this approach may significantly diminish. To address this, Iterative DPO [5] has been proposed, which optimizes pref", + "bbox": [ + 516, + 689, + 905, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9063", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "erence pairs through DPO at each iteration. It then generates new preference pairs for the next iteration using the updated model and evaluates them with a reward model. In this paper, we use iterative DPO to achieve stronger alignment and enhance the model's reasoning capabilities.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 195, + 225, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we provide a comprehensive description of the proposed Insight-V system, detailing its architecture and key contributions. Section 3.1 presents an overview of Insight-V, highlighting the core concepts of our approach. The design of Insight-V is structured around three primary components: 1) a carefully constructed pipeline for structured, scalable reasoning data generation, as described in Section 3.2; 2) a multi-agent MLLM system that supports complex visual reasoning, as detailed in Sections 3.3; and 3) a streamlined yet effective training pipeline to enhance overall performance, as outlined in Section 3.4. Together, these components form a cohesive system that effectively tackles the challenges of performing detailed, long-chain reasoning while preserving visual perception capabilities.", + "bbox": [ + 89, + 224, + 483, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 460, + 200, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Advancing reasoning capabilities of LLMs has been a focal point of extensive research. Despite these efforts, the reasoning potential within multi-modal LLMs has barely been explored. Most approaches aim to strengthen reasoning at the inference stage, assuming the model has already acquired robust reasoning skills. Other approaches optimize model parameters using chain-of-thought data, enabling models to mimic human reasoning processes. However, these methods present significant challenges for current general-purpose MLLMs, as they require the model to develop reasoning skills while retaining prior capabilities, which often results in only modest performance gains. Additionally, the lack of structured, high-quality training data impedes training models with advanced reasoning capabilities.", + "bbox": [ + 89, + 488, + 483, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To fully leverage the reasoning capabilities of MLLMs, we propose Insight-V, a novel system comprising two MLLMs dedicated to reasoning and summarization, respectively. The reasoning model is tasked with generating a detailed reasoning process to assist in problem-solving, while the summary model evaluates this reasoning as supplementary information to assess its relevance and utility for answering the question. We also construct a structured, high-quality dataset to train both agents. We posit that this multi-agent system can enhance the reasoning strengths of MLLMs by decomposing the problem-solving process into distinct reasoning and summarization phases, thereby driving substantial performance improvements.", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/952a18fd2a58e66250ef7efc80fbbba15454c2b5615fa32def0110ed62c8a683.jpg", + "image_caption": [ + "Figure 2. Data Generation Pipeline of Insight-V. The reasoning processes are generated progressively through a reasoning generator, and then fed into a multi-granularity assessment system to ensure high-quality reasoning." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 919, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Construction of Structured Reasoning Data", + "text_level": 1, + "bbox": [ + 511, + 349, + 883, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous studies [56, 58] have explored the integration of reasoning capabilities into MLLMs. However, training MLLMs to develop robust reasoning skills remains a considerable challenge, particularly due to data limitations. To address this, we introduce our data generation pipeline in this section, designed to produce high-quality, long-chain reasoning data using a progressive generation process and multi-granularity assessment. As shown in Figure 2, this scalable approach enables us to generate high-quality data to enhance the model's reasoning capabilities effectively.", + "bbox": [ + 511, + 371, + 908, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Progressive Long-Chain Reasoning Data Generation.", + "text_level": 1, + "bbox": [ + 511, + 541, + 897, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each input query, we first employ a reasoning generator to produce a structured reasoning process in JSON format to address the problem. At each step, the reasoning generator provides a brief summary of the current step, a detailed reasoning response, and an action for the following step. If the action is continue, the model proceeds with an additional reasoning step in the next iteration; if the action is summary, the model generates a final summary and answer based on the complete reasoning process in the subsequent iteration. Specifically, For a multi-modal model $M$ , input image $I$ , and question $Q$ , the data generation process for each step is represented as follows:", + "bbox": [ + 511, + 556, + 908, + 737 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {t} = M (I, Q, \\left[ R _ {1} \\dots R _ {t - 1} \\right], A),\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 748, + 820, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {a n s} = M (I, Q, \\left[ R _ {1} \\dots R _ {n} \\right]),\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 768, + 803, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $R_{t}$ and $R_{ans}$ denote the response at the $t$ -th step and the final answer, respectively, $R_{i}$ represents the reasoning generated by the model at the $i$ -th step, $n$ represents the total reasoning steps, and $A$ is the action determined in the previous step. By repeating this process $N$ times, we can iteratively sample $N$ structured responses for each query. The generation parameters are adjusted to encourage the", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9064", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "model to produce outputs with various information and steps, allowing us to identify the most effective reasoning chain for each question.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Granularity Assessment. After obtaining the structured responses, we utilize an assessment pipeline to ensure data quality. Specifically, we first apply a strong LLM, such as Qwen2 [37], for direct answer filtering. As shown in Figure 2, the model is provided with the generated final answer and the ground truth answer and is tasked with determining whether the generated answer is correct, serving as an approximate indicator of the validity of the associated reasoning chain. Once responses with incorrect answers are filtered out, the remaining reasoning processes are passed to a reasoning path scoring agent. Here, an advanced multimodal model, such as Qwen2-VL [38], is supplied with the image, question, reasoning path, and ground truth answer, and is prompted to evaluate the reasoning path. The scoring agent assesses each response based on the step-by-step accuracy of the reasoning path and the level of detail in the reasoning. To ensure consistency in scoring across different data samples, we aggregate all responses for each question and process them in a single pass. The model then generates scores for each response, ranging from 1 to 100.", + "bbox": [ + 91, + 156, + 483, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through the above two steps, we construct a structured, high-quality dataset that provides detailed reasoning for each question, effectively supporting the training of our models.", + "bbox": [ + 89, + 459, + 483, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Model Design", + "text_level": 1, + "bbox": [ + 89, + 515, + 232, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After constructing the dataset, we develop a multi-agent framework to enhance overall reasoning capabilities through collaborative agent interaction. Specifically, we first train a reasoning agent to generate a detailed reasoning process for each problem. Then, a summary agent is employed to answer the question, selectively utilizing the reasoning process based on its assessment. Together, these two agents collaborate to improve reasoning performance effectively.", + "bbox": [ + 89, + 537, + 483, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reasoning Agent. Previous approaches typically combine reasoning and question-answering within a single process, which poses challenges for MLLMs. Generating a long-chain reasoning process can introduce errors, and directly answering questions based on flawed reasoning often leads to poorer results. To address this, we propose a specialized reasoning agent designed to generate a detailed, step-by-step reasoning process in response to an input query. We construct the reasoning dataset by selecting the highest-scoring reasoning path for each question. After training on this dataset, the model transforms into a reasoning agent with enhanced reasoning capabilities, enabling it to generate more detailed, structured reasoning processes.", + "bbox": [ + 89, + 666, + 483, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Summary Agent. Summarization plays a critical role in enabling models to accurately answer questions. After", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fb79206494bf4a64124f608becbcf2d97ebcac1cc39902ff3abf9ae690cb7102.jpg", + "image_caption": [ + "Figure 3. Overview of Insight-V Model Design. We derive a multi-agent system from a single model. By decomposing the task into reasoning and summarization, the two agents collaborate to enhance the overall reasoning capability." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 919, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generating multi-step reasoning, summarization provides a cohesive understanding of the reasoning process, ultimately guiding the model to the final answer. However, since the response generated by the reasoning agent may contain errors, we develop a summarization model robust to inaccuracies in the reasoning path, selectively incorporating or disregarding elements as needed. This approach maximizes the reasoning model's effectiveness while minimizing the risk of introducing misleading information.", + "bbox": [ + 511, + 353, + 908, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance the robustness of the summary agent, we carefully curate its training dataset. We utilize the collected dataset, which is comprised of two types: data with optimal reasoning processes and data with flawed reasoning processes for the summarization task. This method prevents the model from simply copying reasoning outcomes and encourages critical evaluation of reasoning quality. To further promote critical analysis by the summary agent, we select flawed reasoning samples based on their performance scores. Specifically, we draw flawed responses from varying score ranges to create a dataset with different levels of error, prompting the model to assess reasoning processes at various granularities. To better align the summary model with the reasoning agent, we also incorporate question-reasoning pairs generated by the reasoning agent to enhance collaboration between the two agents. Additionally, to preserve the original multi-modal capabilities, we supplement the dataset with standard question-answering data to sustain the summary agent's performance in direct question-answering.", + "bbox": [ + 511, + 489, + 908, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Training Pipeline", + "text_level": 1, + "bbox": [ + 511, + 787, + 684, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The training pipeline for Insight-V is designed to be straightforward and efficient, utilizing a two-stage strategy. For both the reasoning agent and the summary agent, we begin with a well-trained MLLM. In the first stage, we apply supervised fine-tuning to train the agents to fulfill their designated roles using corresponding datasets. In the second stage, we", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9065", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "implement direct preference optimization, following prior research [45, 57]. This optimization is applied to the reasoning model, aligning it with human reasoning processes in a simple yet effective manner. These two stages enable the development of a robust system with enhanced visual reasoning capabilities.", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1. Supervised Fine-tuning for Multi-agent System", + "text_level": 1, + "bbox": [ + 89, + 186, + 464, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To perform supervised fine-tuning and obtain the two agents, we first train a base multi-modal model, following established methodologies. This model can address general visual question-answering tasks and gain foundational vision-language skills. We compile a high-quality image-text dataset focused on knowledge learning to train the base model. This data is sourced from various open-source academic datasets, including LLaVA-NeXT [22], Cauldron [15], and Cambrian-1 [43]. Once the base model is trained, we further fine-tune the two agents, initializing them from the base model. For the reasoning agent, we utilize the curated reasoning dataset to develop step-by-step reasoning capabilities. For the summary agent, we formulate a dataset as outlined in Section 3.3 and sample about one million general image-text pairs from the dataset used for the base model, preserving its original visual perception abilities.", + "bbox": [ + 89, + 205, + 485, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2. Enhanced Reasoning with RL", + "text_level": 1, + "bbox": [ + 89, + 452, + 344, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Preference learning has gained increasing focus in the field of large language models. The primary aim is to fine-tune model outputs to align better with human (or expert) preferences, creating outputs more suited to real-world applications. Let's assume a preference dataset defined as $\\mathcal{D} = \\{(x^{(i)},y_w^{(i)},y_l^{(i)})\\}_{i = 1,\\dots ,|\\mathcal{D}|}$ , where each $x^{(i)}$ is prompt, and $y_{w}^{(i)}$ and $y_{l}^{(i)}$ represent the preferred and less preferred responses, respectively. We denote $y_{w}\\succ y_{l}\\mid x$ to signify that $y_{w}$ is preferred over $y_{l}$ for prompt $x$ .", + "bbox": [ + 89, + 470, + 483, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the true distribution of human preferences cannot be directly observed, we approximate it with a latent reward model $r^{*}(x,y)$ , assuming that higher rewards correspond to stronger preferences. Following the approach by Rafailov et al. [39], we can model the human preference distribution $p^{*}$ using the Bradley-Terry (BT) model [3]:", + "bbox": [ + 89, + 613, + 483, + 704 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p ^ {*} (y _ {1} \\succ y _ {2} \\mid x) = \\frac {\\exp (r ^ {*} (x , y _ {1}))}{\\exp (r ^ {*} (x , y _ {1})) + \\exp (r ^ {*} (x , y _ {2}))} \\\\ = \\sigma \\left(r ^ {*} (x, y _ {1}) - r ^ {*} (x, y _ {2})\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 709, + 460, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ denotes the logistic function.", + "bbox": [ + 89, + 766, + 339, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To estimate the parameters of the reward model, we can apply maximum likelihood estimation by minimizing the negative log-likelihood:", + "bbox": [ + 89, + 781, + 483, + 825 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R} \\left(r _ {\\phi}, \\mathcal {D}\\right) = - \\mathbb {E} _ {\\left(x, y _ {w}, y _ {l}\\right) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(r _ {\\phi} \\left(x, y _ {w}\\right) - r _ {\\phi} \\left(x, y _ {l}\\right)\\right) \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 832, + 482, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r_{\\phi}$ is a parameterized reward model. This approach allows us to approximate the preference distribution and fine-tune the model to capture human-like preferences effectively.", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The traditional DPO algorithm operates in an offline setting. During DPO training, as model parameters continuously change, the preference dataset generated offline can gradually diverge from the model's current distribution, which weakens the effectiveness of the DPO algorithm. To address this issue, we employ an iterative DPO algorithm. By conducting multiple rounds of DPO training and sampling, this approach enables the model to better approximate an online setting during training, thus further enhancing its performance. Specifically, our approach involves training a sequence of models $M_{1},\\ldots ,M_{T}$ , where each subsequent model $M_{t + 1}$ utilizes preference data $\\mathcal{D}_t$ generated by the $t$ -th model. We apply this complete training process to the fine-tuned reasoning agent, enabling the model to better align with human preferences and produce structured, detailed reasoning steps for complex questions, which supports the summary agent more effectively.", + "bbox": [ + 511, + 90, + 906, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 361, + 645, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct extensive experiments across multiple vision-language benchmarks to validate the effectiveness of our method. In this section, we first introduce the implementation details of Insight-V in Section 4.1. Then we present a comparison with state-of-the-art MLLMs, outlining the primary results of our method on visual reasoning tasks as well as additional results on general image understanding in Section 4.2. Moreover, we offer further analytical experiments and essential ablation studies on design choices in Section 4.3, along with qualitative results shown in Section 4.4 for more insights.", + "bbox": [ + 511, + 386, + 908, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 561, + 730, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We integrate the Insight-V system with various MLLMs to demonstrate the broad applicability of our approach. Our initial implementation with Insight-V on LLaVA-NeXT-LLaMA3 [22] illustrates the method's effectiveness. To further validate its generalizability and establish a solid baseline against state-of-the-art MLLMs, we additionally train a base multi-modal model using the Qwen-2.5-7B [41] LLM. During pretraining, we utilize the 558K captioning dataset from LLaVA-1.5 [21], unfreezing the connector parameters. This is followed by supervised fine-tuning with a curated dataset of approximately 4 million images, using a learning rate of 2e-5 as guided by prior researches [21, 22]. This two-stage training process equips the baseline model with essential visual perception abilities, achieving competitive results on vision-language benchmarks.", + "bbox": [ + 511, + 583, + 908, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then initialize two agents from the baseline model, performing targeted fine-tuning to obtain the final agents. For the reasoning agent, we compile a dataset of 200K images and train the model over 2 epochs with a learning rate of 5e-6. For the summary agent, we use a dataset of 1.2 million images, applying a learning rate of 1e-5 and train", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9066", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8e7a4db7ddf76a1fa5d9a04b1cf11bf5f7830f5875941835485d58bc2c3ade2d.jpg", + "table_caption": [ + "Table 1. Results on Visual Reasoning Tasks. We conduct evaluation experiments across 7 benchmarks, covering both general reasoning and task-specific reasoning assessments. Insight-V exhibits notable effectiveness and generalizability when applied to LLaVA-NeXT and our baseline model, surpassing other state-of-the-art MLLMs by a large margin." + ], + "table_footnote": [], + "table_body": "
ModelSizeMMMUMMMU-ProMMBenchMMEChartQAMMStarMathVistaAverage
DeepSeek-VL [29]7B35.4-73.5-/-59.137.136.1-
VILA-1.5 [20]8B38.6-75.31634.9/--39.7--
Cambrian-1 [43]8B42.7-75.91547.1/-73.3-49.0-
InternLM-XComposer2 [7]7B41.1-77.62220.471.856.259.5-
POINTS [26]7B51.4-78.02184.1-60.963.0-
IXC-2.5 [55]7B42.9-79.42233.182.259.963.7-
Bunny-LLaMA3 [12]8B43.4-77.21588.9/321.1--34.4-
MM-1.5 [54]7B41.8--1514.9/346.478.6-47.6-
MiniCPM-LLaMA3-V 2.5 [49]8B45.819.677.22024.6-51.854.3-
MiniCPM-V-2.6 [50]7B49.827.278.02268.7-57.560.6-
Qwen2-VL [38]7B53.7-81.0-83.060.761.4-
Idefics3-LLaMA3 [14]8B46.622.977.51937.474.855.958.448.1
Ovis1.5-LLaMA3 [31]8B48.323.676.61948.576.457.363.049.4
LLaVA-NeXT-LLaMA3 [22]8B36.913.272.31611.1/346.069.443.145.940.2
+ Multi-Agent8B40.817.877.61603.7/469.374.652.647.444.5
+ Iterative DPO (Insight-V-LLaVA)8B42.021.081.71583.9/485.477.457.449.847.2 (+7.0)
Our Base Model7B47.122.681.31573.7/482.575.757.056.948.7
+ Multi-Agent7B49.723.882.21662.2/629.381.258.658.750.7
+ Iterative DPO (Insight-V)7B50.224.982.31685.1/627.081.561.559.951.6 (+2.9)
", + "bbox": [ + 91, + 142, + 906, + 382 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/df280f877b6050045bac5914d8c379a17cf479b1e3a2ef99896b466af22f9726.jpg", + "table_caption": [ + "Table 2. Results on other multimodal benchmarks. Insight-V enhances reasoning capabilities without compromising general visual perception and even achieves improvements on benchmarks requiring perception ability more." + ], + "table_footnote": [], + "table_body": "
ModelTextVQADocVQAOCRbenchAI2D
LLaVA-NeXT-LLaMA365.278.255371.5
+ Multi-Agent68.981.863175.7
+ Iterative DPO (Insight-V-LLaVA)70.582.966377.3
Our Base Model75.490.271379.7
+ Multi-Agent77.091.473880.1
+ Iterative DPO (Insight-V)76.891.573579.8
", + "bbox": [ + 91, + 460, + 483, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing for 1 epoch. Additionally, we apply Direct Preference Optimization (DPO) to the reasoning agent, using approximately 15K preference data and training for 1 epoch at a learning rate of 5e-7. This DPO process is iteratively conducted across 3 rounds by using the model from the previous stage to generate preference data, thus enhancing the agent's reasoning capabilities. The lmms-eval [16] is utilized for fast evaluation. Further training details are provided in the appendix.", + "bbox": [ + 88, + 563, + 482, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results on Visual Reasoning", + "text_level": 1, + "bbox": [ + 89, + 710, + 390, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup. We present comprehensive experimental results on various visual reasoning benchmarks, which require the model to handle complex and challenging questions, demonstrating strong reasoning capabilities. We select several well-known and representative benchmarks, covering comprehensive evaluation, chart understanding, and mathematical problem-solving. MMMU [52] and MMMU-Pro [53] evaluate the model's expert-level perception and reasoning abilities across various topics. MMBench [24] is a widely used benchmark for comprehensive MLLM evaluation on images. MME [10] includes 14 challenging subtasks for assessing", + "bbox": [ + 88, + 734, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/32ece75a2acbb8abd5474ad89b6d01f2c9d93f626938acada62cd447252ee845.jpg", + "table_caption": [ + "Table 3. Ablations on the Insight-V Design Choice. The multiagent design outperforms other configurations, highlighting the critical role of reasoning and summarization decomposition." + ], + "table_footnote": [], + "table_body": "
ModelMMMUChartQAMathVistaMMStarAvg
Baseline47.175.756.957.059.2
Vanilla - Direct SFT47.079.257.658.460.6
Multi-Turn Supervised48.179.657.958.261.0
Summary Agent Only47.576.357.357.959.8
Multi-Agent49.781.258.758.662.1
", + "bbox": [ + 516, + 446, + 906, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "visual perception and cognition. ChartQA [33] focuses on logical reasoning with charts, while MathVista [30] assesses problem-solving skills in math, using GPT-4-0613 [35] as the evaluator, following standard practice. MMStar [4] tests a range of tasks with varying difficulty levels.", + "bbox": [ + 511, + 541, + 906, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Main Results. The experimental results in Table 1 confirm the effectiveness and generalizability of Insight-V on visual reasoning tasks. Applied to LLaVA-NeXT and our baseline model, the Insight-V system substantially boosts performance across challenging visual reasoning benchmarks, achieving competitive outcomes against state-of-the-art methods. We report an average improvement across all benchmarks: Insight-V yields an average increase of $4.9\\%$ and $3.2\\%$ when applied to LLaVA-NeXT and the base model, underscoring the system's effectiveness. Following Direct Preference Optimization, an additional improvement of $2.6\\%$ and $1.0\\%$ was observed, illustrating the value of reinforcement learning algorithms in enhancing reasoning capabilities. Specifically, on general reasoning benchmarks, MMMU, MMMU-Pro, and MMBench, we noted gains of $3.1\\%$ , $2.3\\%$ , and $1.0\\%$ , respectively. The highest improvement was on MME, with a $9.1\\%$ increase, highlighting Insight-V's strength in tasks requiring both perception and", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9067", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a244efeef3264345029049f015ebfe359b3db79d1890b57e3ed256e25ecda321.jpg", + "image_caption": [ + "Figure 4. Ablations on the amount of training data. The reasoning agent benefits from data scaling, providing more valuable insights for the summary agent." + ], + "image_footnote": [], + "bbox": [ + 109, + 99, + 444, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "reasoning. On ChartQA, which entails complex chart-based reasoning, we achieved a $5.8\\%$ increase. Performance gains on MMStar and MathVista, with improvements of $4.5\\%$ and $3.0\\%$ , further validate our approach's versatility in tackling math problems that demand sophisticated reasoning and numerical calculation skills. This strong performance across all benchmarks underscores the efficacy of Insight-V and its potential to advance visual reasoning capabilities.", + "bbox": [ + 89, + 320, + 483, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Other Multi-modal Benchmarks. To verify that Insight-V enhances reasoning capabilities without compromising basic visual perception skills, we conduct experiments on benchmarks requiring only fundamental image understanding. We selected TextVQ [40], DocVQA [34], OCRBench [25], and AI2D [13] for this purpose, as these benchmarks focus on basic visual interpretation rather than complex reasoning.", + "bbox": [ + 89, + 450, + 483, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 2, integrating Insight-V enables models to maintain strong performance across general visual perception benchmarks. Additionally, applying Insight-V yields performance improvements on both LLaVA-NeXT and our base model, as it enhances general visual understanding by guiding the model's attention to relevant regions, thereby improving visual perception accuracy. Following DPO, the model demonstrates advanced reasoning capabilities while achieving competitive results on standard image understanding benchmarks compared to baseline models, underscoring Insight-V's robustness. Consistent results of both models further confirm Insight-V's generalizability.", + "bbox": [ + 89, + 571, + 483, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Further Analysis", + "text_level": 1, + "bbox": [ + 89, + 763, + 256, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present comprehensive experiments to validate the design choices of Insight-V, emphasizing our approach's key contributions. Additionally, we include a case study to further demonstrate the qualitative effectiveness of Insight-V.", + "bbox": [ + 89, + 786, + 483, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Multi-agent System. To evaluate the effectiveness of the multi-agent system, we comprehensively", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7b1304471774191c54149c5786110c6360ec529e7c0d7f62cb97d00f23fd80e5.jpg", + "table_caption": [ + "Table 4. Ablations on the DPO training strategy. Iterative DPO progressively enhances the model's reasoning capabilities, leading to improved performance." + ], + "table_footnote": [], + "table_body": "
ModelMMMUChartQAMathVistaMMStarAvg
Insight-V (Multi-Agent)49.781.258.758.662.1
+ RLAIF49.581.459.159.262.3
+ DPO50.880.859.359.962.7
+ Iterative DPO50.281.559.961.563.3
", + "bbox": [ + 517, + 142, + 906, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compare Insight-V with alternative design choices. We begin by reporting the performance of the summary agent without a reasoning process to underscore the importance of the reasoning agent. For a comprehensive comparison, we also restructure the collected data using a Chain-of-Thought (CoT) template and train a model to perform sequential reasoning and subsequently answer questions. Additionally, we enable a model that can handle multi-turn conversations, facilitating reasoning and summarization in a multi-turn format. We use MMMU, ChartQA, MathVista, and MMStar as representative benchmarks to assess the performance across different methods.", + "bbox": [ + 511, + 222, + 906, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 3, the results indicate that the multi-agent system plays a critical role in enhancing the system's visual understanding capabilities. Using only the summary agent without a reasoning process results in limited improvements in reasoning tasks, as the model lacks the necessary reasoning framework for optimal performance. Training the model to perform chain-of-thoughts reasoning, which is denoted as Insight-V (Vanilla-Direct SFT), yields modest gains, as this approach does not sufficiently emphasize critical judgments within the reasoning process, merging reasoning and summarization into a single task. Furthermore, models trained with multi-turn conversations still produce sub-optimal results, underscoring the importance of a multi-agent system that separates reasoning and summarization.", + "bbox": [ + 511, + 404, + 908, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data Scaling Law of Reasoning Agent. To assess the effectiveness of the reasoning agent, we perform ablation experiments on the data volume for training. As shown in Figure 4, we compare reasoning agents trained on varying data sizes from 50K to 200K. For a fair and comprehensive comparison, we employ the same summary agent across evaluations and report results on six benchmarks. The findings clearly indicate that the reasoning agent benefits from increased data. With limited data, the reasoning agent struggles to generalize and fails to provide useful input to the summary model, resulting in performance even worse than baseline models. In contrast, training in larger datasets improves the capabilities of the reasoning agent, allowing it to perform step-by-step reasoning and provide valuable insights that help the summary agent solve tasks effectively.", + "bbox": [ + 511, + 622, + 908, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of RL Algorithms. We assess DPO algorithms to identify optimal strategies for enhancing the reasoning process. To examine the effects of dataset composition on DPO", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9068", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c7a25ec1fd61936835c8b31aeb85628a7776f1ee1f0180028a4b27df3229af3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 114, + 215, + 175 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Referring to Table 3-1, which is INCORRECT. If pots sell for $20 each then:", + "bbox": [ + 225, + 104, + 444, + 121 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(A) The marginal product of the third worker is five pots.", + "(B) the marginal revenue product of labor equals the marginal product of labor multiplied by the additional revenue.", + " (C) the marginal revenue product of labor of the second worker is $260.", + "(D) the marginal revenue from selling the eighteenth pot is $20." + ], + "bbox": [ + 225, + 125, + 455, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Chain-of-Thought", + "text_level": 1, + "bbox": [ + 741, + 94, + 857, + 107 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A. ... It is incorrect because the marginal product of the third worker is 5 pots (18 - 13 = 5). However, the question incorrectly states that ... of the third worker is five pots, which should be ... of the second worker.", + "B. ... This statement is correct. It is calculated by multiplying the marginal product of labor by the price per unit. In this case, ...", + " C......This statement is incorrect. The marginal product of the second worker is 5 pots, and the price per pot is $20, so ... is 5 * $20 = $100.", + "D. ... This statement is correct. The ... is equal to the price per unit, which in this case is $20.", + "So the answer is A." + ], + "bbox": [ + 475, + 111, + 864, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Insight-V (Vanilla)", + "text_level": 1, + "bbox": [ + 130, + 196, + 250, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Insight-V (Multi-Agent)", + "text_level": 1, + "bbox": [ + 720, + 196, + 870, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DIRECT SFT", + "text_level": 1, + "bbox": [ + 137, + 220, + 207, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfcce8980bd3b9fc188f00aa812a4d81cfb9c0eb96fdfb368a54bfe6f451fb24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 237, + 158, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Identifying Key Information: The image presents a table showing ... The question asks about ...", + "bbox": [ + 171, + 236, + 400, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/778b689d6fe9a1c351ac842731aded0a7e487ba15f80770e6cfa87242cd79189.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 267, + 158, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Calculating Marginal Product of Labor: To find the marginal product of labor, ... For example, ... subtracting the total pots produced with two workers from that with three workers.", + "bbox": [ + 171, + 257, + 400, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ee4fd4c6ffcaf143575a303ab2ee389b42bed6d4f3abaec1678d21f5a47019c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 300, + 158, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Determining Marginal Revenue Product: Next, we need to determine ... We multiply the marginal product of each worker by \\(20 to \\dots", + "bbox": [ + 171, + 297, + 400, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/307d65e41ad124e925b0ede0019b72ec5c574315f66f158b4e0bd54cf40d64ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 340, + 158, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparing Options: Now that ... Option (A) states ..., which does not matches... . Option (B) suggests that ..., but this does not match ... . Option (C) indicates that ..., which also does not match ... . Option (D) claims that ..., which aligns with...", + "bbox": [ + 169, + 328, + 398, + 373 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f970d9032c04744ec79cedb7ee0cbfe3121a27642079c5318b859502a6af015.jpg", + "image_caption": [ + "Figure 5. Qualitative Results of Insight-V. We present qualitative comparisons of Insight-V with Chain-of-Thought and Insight-V with direct SFT (Vanilla). For the Insight-V system, the reasoning agent delivers a more coherent and structured reasoning process, guiding the summary agent to the correct answer, whereas other methods struggle with complex reasoning and fail to solve such challenging problems." + ], + "image_footnote": [], + "bbox": [ + 135, + 380, + 158, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FINAL ANSWER: (D) the marginal revenue from selling the eighteenth pot is $20.", + "bbox": [ + 171, + 378, + 395, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REASONING AGENT", + "text_level": 1, + "bbox": [ + 444, + 219, + 560, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29d50873d1110d083270fec76951095889a5e45aff2cdcc6e0dbed0bb578fcb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 237, + 465, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Identifying Key Information: To begin solving this problem, we need to ... This involves understanding the table provided, which ...", + "bbox": [ + 477, + 236, + 862, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/afbc7a818f7a055823802f6eb3d8d40929037def6e8959b6eacdfe9420e9de78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 258, + 465, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analyzing Table Data: Next, we should analyze ... in Table 3-1 to ... . Specifically, we need to calculate... to determine which option is incorrect.", + "bbox": [ + 475, + 258, + 859, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/767b7adf363d8251449e2d718ed93946cd5cf7001697b766315abfa4d77c2e7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 281, + 465, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparing Marginal Products: Option (A) suggests .... Based on Table 3-1, we find ... (from 13 to 18), matching option (A).", + "bbox": [ + 475, + 280, + 859, + 299 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b682c5e385bd75a15c1d2b276cb37078c4522e84fad83bd37ef1a89a990cc93d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 308, + 465, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Calculating Marginal Revenue Product of Labor: Option (B) states.... It is consistent with option (B), as it takes into account both the marginal product and the additional revenue per unit of output.", + "bbox": [ + 475, + 301, + 859, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b0220de4bbd6d3766239bb19dff80782ce63b64e3fb19ed27b1a3f836751f7d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 335, + 465, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analyzing Option (C): Option (C) claims ... For the second worker, ... We can calculate that ... of the second worker is 7 multiplies $20, which is$ 140, contradicting option (C).", + "bbox": [ + 475, + 332, + 862, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/553a02bf49b2decbc863c0c383795908b9070b1c39f5494fc5d8e1b880fba4ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 356, + 465, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Calculating Marginal Revenue: Option (D) indicates ... It can be ... that the pots sell $20 each. This means that ... is equal to $20, matching option (D).", + "bbox": [ + 475, + 354, + 864, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SUMMARY AGENT", + "text_level": 1, + "bbox": [ + 444, + 376, + 549, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c59585149320464353ff81377b9ff8fd6f9e297dd040872780c31df26c96a79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 388, + 465, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on the reasoning process above, The incorrect option is option (C).", + "bbox": [ + 475, + 388, + 779, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "training, we compare our curated dataset against the widely-used RLAIF-V [51] dataset, which contains 80K DPO data pairs for alignment. We also investigate the potential benefits of iterative DPO in advancing the model's reasoning capabilities. For an unbiased comparison, we subsample the RLAIF-V dataset to approximately 15K data, aligning it with the size of our dataset. The results, summarized in Table 4, show an average improvement of approximately $0.2\\%$ on the evaluated benchmarks when trained with the RLAIF-V dataset, whereas our curated dataset achieves greater gains of $0.6\\%$ . This suggests that a DPO dataset based on model-generated rationales, rather than externally sourced data, more effectively boosts reasoning accuracy. Moreover, conducting two additional rounds of DPO training using the same methodology yields further performance gains of $0.6\\%$ , underscoring that iterative DPO training progressively refines the model's ability to generate precise and high-quality reasoning processes compared to a single training pass.", + "bbox": [ + 91, + 470, + 483, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 756, + 272, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present qualitative comparisons in Figure 5 to illustrate the improvements introduced by integrating the Insight-V system. Specifically, we provide an example that highlights advancements in the reasoning process. In this example, the model is challenged with a multiple-choice question based on a table. We compare our approach with a direct Chain-of-Thought application, and the model undergoes supervised fine-tuning without incorporating a multi-agent sys", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tem, which is denoted as Insight-V (Vanilla). It is evident that applying Chain-of-Thought directly results in suboptimal reasoning and leads to incorrect answers. Although fine-tuning a single model yields somewhat better reasoning, the model begins to falter as the reasoning chain lengthens, ultimately arriving at incorrect answers. This outcome arises because the model is required to generate both reasoning steps and the final answer simultaneously, limiting its judgment capabilities and weakening its robustness in handling flawed reasoning chains. In contrast, employing Insight-V enables more logical step decomposition and a structured reasoning chain. The model can analyze each option step-by-step and perform detailed calculations, which the other two methods cannot achieve. The fine-tuned summary model is able to evaluate the reasoning process and determine whether to derive the final answer based on it, significantly enhancing system robustness and ensuring correct answers.", + "bbox": [ + 511, + 469, + 906, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 739, + 632, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce Insight-V, a novel system that combines a scalable data generation pipeline for long-chain reasoning data and an effective multi-agent training pipeline to enhance the reasoning capabilities of MLLMs. By developing this system, we provide a scalable approach to model training aimed at improving reasoning performance. Our extensive evaluation across various benchmarks demonstrates the effectiveness of our approach, paving the way for equipping MLLMs with enhanced reasoning capabilities.", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9069", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. 2023. 2", + "[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 2", + "[3] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 5", + "[4] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 6", + "[5] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 2", + "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1, 2", + "[7] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 6", + "[8] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1", + "[9] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1", + "[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 6", + "[11] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023. 2", + "[12] Muyang He, Yexin Liu, Boya Wu, Jianhao Yuan, Yueze Wang, Tiejun Huang, and Bo Zhao. Efficient multimodal learning from data-centric perspective. arXiv preprint arXiv:2402.11530, 2024. 6" + ], + "bbox": [ + 93, + 114, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In ECCV, pages 235-251. Springer, 2016. 7", + "[14] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. arXiv preprint arXiv:2408.12637, 2024. 6", + "[15] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 5", + "[16] Bo Li. Xinrun du, yuhao dong, haotian liu, yuanhan zhang, ge zhang, chunyuan li, and ziwei liu. Lmms-eval: Accelerating the development of large multimoal models, 2024. 6", + "[17] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2", + "[18] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 2", + "[19] Zhenwen Liang, Tianyu Yang, Jipeng Zhang, and Xiangliang Zhang. Unimath: A foundational and multimodal mathematical reasoner. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7126-7133, 2023. 2", + "[20] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2, 6", + "[21] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 5", + "[22] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.1,2,5,6", + "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1, 2", + "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 6", + "[25] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 7", + "[26] Yuan Liu, Zhongyin Zhao, Ziyuan Zhuang, Le Tian, Xiao Zhou, and Jie Zhou. Points: Improving your vision-language model with affordable strategies. arXiv preprint arXiv:2409.04828, 2024. 6", + "[27] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9070", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Zuyan Liu, Yuhao Dong, Yongming Rao, Jie Zhou, and Jiwen Lu. Chain-of-spot: Interactive reasoning improves large vision-language models. arXiv preprint arXiv:2403.12966, 2024. 2", + "[29] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Yaofeng Sun, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 1, 2, 6", + "[30] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 6", + "[31] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv:2405.20797, 2024. 6", + "[32] Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. Dolphins: Multimodal language model for driving. arXiv preprint arXiv:2312.00438, 2023. 1", + "[33] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 6", + "[34] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 7", + "[35] OpenAI. Gpt-4 technical report. ArXiv:abs/2303.08774, 2023.6", + "[36] OpenAI. Hello gpt-4o — openai. OpenAI Blog, 2024. 1", + "[37] QwenTeam. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1, 4", + "[38] QwenTeam. Qwen2-vl: To see the world more clearly. Wwen Blog, 2024. 1, 2, 4, 6", + "[39] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2, 5", + "[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 7", + "[41] Qwen Team. Qwen2.5: A party of foundation models, 2024. 1, 5", + "[42] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1", + "[43] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 2, 5, 6" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2", + "[45] Tianyi Xiong, Bo Li, Dong Guo, Huizhuo Yuan, Quanquan Gu, and Chunyuan Li. Llava-onevision-chat: Improving chat with preference learning, 2024. 5", + "[46] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 2", + "[47] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, et al. Octopus: Embodied vision-language programmer from environmental feedback. arXiv preprint arXiv:2310.08588, 2023. 1", + "[48] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[49] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint 2408.01800, 2024. 1, 6", + "[50] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 6", + "[51] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 8", + "[52] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 6", + "[53] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Ming Yin, Botao Yu, Ge Zhang, et al. Mmmu-pro: A more robust multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024.6", + "[54] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Dufter, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, et al. Mm1. 5: Methods, analysis & insights from multimodal llm fine-tuning. arXiv preprint arXiv:2409.20566, 2024. 6", + "[55] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang," + ], + "bbox": [ + 516, + 90, + 908, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9071", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 6", + "[56] Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Yichi Zhang, Ziyu Guo, Chengzhuo Tong, Jiaming Liu, Aojun Zhou, Bin Wei, Shanghang Zhang, et al. Mavis: Mathematical visual instruction tuning. arXiv preprint arXiv:2407.08739, 2024. 2, 3", + "[57] Ruohong Zhang, Bowen Zhang, Yanghao Li, Haotian Zhang, Zhiqing Sun, Zhe Gan, Yinfei Yang, Ruoming Pang, and Yiming Yang. Improve vision language model chain-of-thought reasoning. arXiv preprint arXiv:2410.16198, 2024. 2, 5", + "[58] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2, 3" + ], + "bbox": [ + 91, + 90, + 483, + 371 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9072", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/cb8122c4-a3cd-447b-9677-ef086abe8658_model.json b/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/cb8122c4-a3cd-447b-9677-ef086abe8658_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cc78dda8a847cc555fb7eecfbd70b988a9be2a72 --- /dev/null +++ b/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/cb8122c4-a3cd-447b-9677-ef086abe8658_model.json @@ -0,0 +1,2444 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.13, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.203, + 0.761, + 0.294 + ], + "angle": 0, + "content": "Yuhao Dong\\(^{1*}\\) Zuyan Liu\\(^{2,3*}\\) Hai-Long Sun\\(^{3,4}\\) Jingkang Yang\\(^{1}\\) Winston Hu\\(^{2}\\) Yongming Rao\\(^{2\\boxtimes}\\) Ziwei Liu\\(^{1\\boxtimes}\\) \n\\(^{1}\\) S-Lab, Nanyang Technological University \\(^{2}\\) Tencent \n\\(^{3}\\) Tsinghua University \\(^{4}\\) Nanjing University \nhttps://github.com/dongyh20/Insight-V" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.486, + 0.798 + ], + "angle": 0, + "content": "Large Language Models (LLMs) demonstrate enhanced capabilities and reliability by reasoning more, evolving from Chain-of-Thought prompting to product-level solutions like OpenAI o1. Despite various efforts to improve LLM reasoning, high-quality long-chain reasoning data and optimized training pipelines still remain inadequately explored in vision-language tasks. In this paper, we present Insight-V, an early effort to 1) scalably produce long and robust reasoning data for complex multi-modal tasks, and 2) an effective training pipeline to enhance the reasoning capabilities of multi-modal large language models (MLLMs). Specifically, to create long and structured reasoning data without human labor, we design a two-step pipeline with a progressive strategy to generate sufficiently long and diverse reasoning paths and a multi-granularity assessment method to ensure data quality. We observe that directly supervising MLLMs with such long and complex reasoning data will not yield ideal reasoning ability. To tackle this problem, we design a multi-agent system consisting of a reasoning agent dedicated to performing long-chain reasoning and a summary agent trained to judge and summarize reasoning results. We further incorporate an iterative DPO algorithm to enhance the reasoning agent's generation stability and quality. Based on the popular LLaVA-NeXT model and our stronger base MLLM, we demonstrate significant performance gains across challenging multi-modal benchmarks requiring visual reasoning. Benefiting from our multi-agent system, Insight-V can also easily maintain or improve performance on perception-focused multi-modal tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.224, + 0.84 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.485, + 0.88 + ], + "angle": 0, + "content": "The development of artificial general intelligence requires models that can seamlessly understand and respond to multi-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*Authors contributed equally to this research. Corresponding authors." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.328, + 0.921, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.554, + 0.922, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.646, + 0.908, + 0.703 + ], + "angle": 0, + "content": "Figure 1. Illustration and Performance of Insight-V. Insight-V consists of two agents, one dedicated to reasoning and the other to summarization, driving significant improvements in performance across various visual reasoning benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.719, + 0.909, + 0.87 + ], + "angle": 0, + "content": "modal data. Recent advancements in Large Language Models (LLMs) [9, 36, 37, 41] and Multi-modal LLMs (MLLMs) [6, 21-23, 29, 38, 49] have significantly facilitated this progress across various fields, ranging from common question-answering [6, 17, 22, 27, 38] to autonomous driving [32, 42] and robotics [8, 47]. Despite the substantial progress made in enhancing the performance of MLLMs on a wide range of tasks, enabling MLLMs to perform human-level reasoning remains a key challenge. This area remains underexplored and has yet to fully realize its potential." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Existing efforts [44, 48] to enhance the reasoning capabilities of LLMs through long-chain reasoning have" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9062" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.483, + 0.378 + ], + "angle": 0, + "content": "demonstrated considerable progress, largely benefiting from the availability of structured, high-quality data and well-established training pipelines. In contrast, teaching MLLMs to perform long-chain visual reasoning remains a significant challenge, primarily due to the lack of large-scale, high-quality datasets and efficient and effective training strategies. Compared to text-only data, visual reasoning data is not only more expensive to collect but also requires significant human labor for detailed annotation and validation, due to the absence of an effective data generation pipeline. Moreover, while previous work [58] has demonstrated that directly applying chain-of-thought [44] reasoning can improve the capabilities of MLLMs, other research [56, 57] suggests that current training approaches have limited effectiveness in enhancing CoT reasoning. This highlights the inability of current MLLMs to leverage visual cues for precise step-by-step problem-solving, emphasizing the need for an effective training procedure that enables MLLMs to reason in detail while maintaining clear visual perception." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.384, + 0.483, + 0.744 + ], + "angle": 0, + "content": "To address these challenges, we propose Insight-V, which incorporates two innovative designs to enhance reasoning capabilities. First, we introduce a data generation pipeline consisting of two key steps: a progressive strategy to generate structured, long-chain reasoning data with diverse reasoning paths, and a multi-granularity assessment system to evaluate and score these paths at different levels. Through automatic generation, assessment, and ranking strategies, the pipeline effectively operates without the need for human labor and makes the reasoning dataset more scalable for enhancing reasoning capabilities. To further improve MLLM reasoning beyond data scaling, we design a multi-agent system, as illustrated in Figure 1, that decomposes the problem-solving process into two distinct steps: reasoning and summarization. The reasoning agent generates a detailed reasoning process for the input query, while the summarization agent identifies key information within the reasoning process and selectively answers the question. To refine the quality of the reasoning, we employ an iterative DPO approach to enhance reasoning capabilities. The two agents collaborate to further improve the reasoning quality. Our findings demonstrate that this system significantly enhances the performance of various MLLMs across a broad range of visual reasoning benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.75, + 0.483, + 0.9 + ], + "angle": 0, + "content": "We evaluate Insight-V by integrating our system into the widely used LLaVA-NeXT [22] model. To further highlight the potential of our design in advancing state-of-the-art models, we construct a robust base MLLM and apply our system to demonstrate its ability to enhance SOTA models across various visual reasoning benchmarks. Notably, Insight-V improves LLaVA-NeXT in an average performance of \\(7.0\\%\\) across seven challenging visual reasoning benchmarks, while a \\(2.9\\%\\) improvement is observed with the strong base MLLM, underscoring the effectiveness and generalizability" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.093, + 0.6, + 0.106 + ], + "angle": 0, + "content": "of Insight-V." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.109, + 0.906, + 0.228 + ], + "angle": 0, + "content": "In summary, Insight-V offers 1) a scalable data generation pipeline for long-chain, high-quality reasoning data, 2) a multi-agent system that decomposes visual reasoning tasks into reasoning and summarization, and 3) a two-stage training pipeline to enhance visual reasoning capabilities. Together, these contributions address key challenges in visual reasoning, providing a solid foundation for future research in MLLM reasoning." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.244, + 0.653, + 0.26 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.269, + 0.764, + 0.286 + ], + "angle": 0, + "content": "2.1. Vision-Language Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.293, + 0.906, + 0.655 + ], + "angle": 0, + "content": "Recent advancements in MLLMs [1, 17, 20-23, 27, 29, 38] have equipped these models with robust capabilities across diverse domains, including visual understanding [20, 38], mathematics [19], college-level questions [6], and scientific inquiries. In visual understanding, most research [18, 22, 28, 43, 46] emphasizes fine-grained detail analysis and localization, training models to perform visual reasoning with tailored datasets to enhance interpretive capabilities. For mathematics and expert-level reasoning, existing methods [11, 56, 57] predominantly derive from Chain-of-Thought [44] approaches, training MLLMs to generate step-by-step reasoning across various subjects. However, these approaches often focus primarily on improving dataset quality through Chain-of-Thought, overlooking the importance of structured reasoning paths and extended reasoning chains in advancing model reasoning capabilities. Additionally, significant challenges arise when relying on a single model to manage the entire reasoning process for complex tasks, underscoring the need for a multi-agent system to decompose and enhance this process. In this work, we tackle these challenges by introducing a scalable reasoning data generation pipeline and implementing a multi-agent system for reasoning and summarization decomposition, enhancing the overall reasoning capabilities of existing MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.667, + 0.764, + 0.683 + ], + "angle": 0, + "content": "2.2. Vision-Language Alignment" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.69, + 0.906, + 0.9 + ], + "angle": 0, + "content": "To align the model more closely with human preferences, several alignment techniques are employed for MLLMs. A widely used approach is Reinforcement Learning from Human Feedback [2] (RLHF), which iteratively refines the model's responses based on human feedback, enhancing both response quality and interpretability. To further improve MLLM capabilities, Direct Preference Optimization [39] (DPO) is introduced to simplify the alignment process. By directly training on human preference data, DPO optimizes the model's outputs to better match human-selected responses. However, traditional DPO is primarily focused on offline scenarios, and as the model evolves, the effectiveness of this approach may significantly diminish. To address this, Iterative DPO [5] has been proposed, which optimizes pref" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9063" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "erence pairs through DPO at each iteration. It then generates new preference pairs for the next iteration using the updated model and evaluates them with a reward model. In this paper, we use iterative DPO to achieve stronger alignment and enhance the model's reasoning capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.196, + 0.226, + 0.213 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.226, + 0.485, + 0.439 + ], + "angle": 0, + "content": "In this section, we provide a comprehensive description of the proposed Insight-V system, detailing its architecture and key contributions. Section 3.1 presents an overview of Insight-V, highlighting the core concepts of our approach. The design of Insight-V is structured around three primary components: 1) a carefully constructed pipeline for structured, scalable reasoning data generation, as described in Section 3.2; 2) a multi-agent MLLM system that supports complex visual reasoning, as detailed in Sections 3.3; and 3) a streamlined yet effective training pipeline to enhance overall performance, as outlined in Section 3.4. Together, these components form a cohesive system that effectively tackles the challenges of performing detailed, long-chain reasoning while preserving visual perception capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.461, + 0.201, + 0.476 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.7 + ], + "angle": 0, + "content": "Advancing reasoning capabilities of LLMs has been a focal point of extensive research. Despite these efforts, the reasoning potential within multi-modal LLMs has barely been explored. Most approaches aim to strengthen reasoning at the inference stage, assuming the model has already acquired robust reasoning skills. Other approaches optimize model parameters using chain-of-thought data, enabling models to mimic human reasoning processes. However, these methods present significant challenges for current general-purpose MLLMs, as they require the model to develop reasoning skills while retaining prior capabilities, which often results in only modest performance gains. Additionally, the lack of structured, high-quality training data impedes training models with advanced reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To fully leverage the reasoning capabilities of MLLMs, we propose Insight-V, a novel system comprising two MLLMs dedicated to reasoning and summarization, respectively. The reasoning model is tasked with generating a detailed reasoning process to assist in problem-solving, while the summary model evaluates this reasoning as supplementary information to assess its relevance and utility for answering the question. We also construct a structured, high-quality dataset to train both agents. We posit that this multi-agent system can enhance the reasoning strengths of MLLMs by decomposing the problem-solving process into distinct reasoning and summarization phases, thereby driving substantial performance improvements." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.92, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.282, + 0.908, + 0.338 + ], + "angle": 0, + "content": "Figure 2. Data Generation Pipeline of Insight-V. The reasoning processes are generated progressively through a reasoning generator, and then fed into a multi-granularity assessment system to ensure high-quality reasoning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.351, + 0.884, + 0.367 + ], + "angle": 0, + "content": "3.2. Construction of Structured Reasoning Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.909, + 0.525 + ], + "angle": 0, + "content": "Previous studies [56, 58] have explored the integration of reasoning capabilities into MLLMs. However, training MLLMs to develop robust reasoning skills remains a considerable challenge, particularly due to data limitations. To address this, we introduce our data generation pipeline in this section, designed to produce high-quality, long-chain reasoning data using a progressive generation process and multi-granularity assessment. As shown in Figure 2, this scalable approach enables us to generate high-quality data to enhance the model's reasoning capabilities effectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.898, + 0.557 + ], + "angle": 0, + "content": "Progressive Long-Chain Reasoning Data Generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.909, + 0.738 + ], + "angle": 0, + "content": "For each input query, we first employ a reasoning generator to produce a structured reasoning process in JSON format to address the problem. At each step, the reasoning generator provides a brief summary of the current step, a detailed reasoning response, and an action for the following step. If the action is continue, the model proceeds with an additional reasoning step in the next iteration; if the action is summary, the model generates a final summary and answer based on the complete reasoning process in the subsequent iteration. Specifically, For a multi-modal model \\( M \\), input image \\( I \\), and question \\( Q \\), the data generation process for each step is represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.749, + 0.821, + 0.765 + ], + "angle": 0, + "content": "\\[\nR _ {t} = M (I, Q, \\left[ R _ {1} \\dots R _ {t - 1} \\right], A),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.769, + 0.805, + 0.785 + ], + "angle": 0, + "content": "\\[\nR _ {a n s} = M (I, Q, \\left[ R _ {1} \\dots R _ {n} \\right]),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "where \\( R_{t} \\) and \\( R_{ans} \\) denote the response at the \\( t \\)-th step and the final answer, respectively, \\( R_{i} \\) represents the reasoning generated by the model at the \\( i \\)-th step, \\( n \\) represents the total reasoning steps, and \\( A \\) is the action determined in the previous step. By repeating this process \\( N \\) times, we can iteratively sample \\( N \\) structured responses for each query. The generation parameters are adjusted to encourage the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9064" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "model to produce outputs with various information and steps, allowing us to identify the most effective reasoning chain for each question." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.157, + 0.485, + 0.459 + ], + "angle": 0, + "content": "Multi-Granularity Assessment. After obtaining the structured responses, we utilize an assessment pipeline to ensure data quality. Specifically, we first apply a strong LLM, such as Qwen2 [37], for direct answer filtering. As shown in Figure 2, the model is provided with the generated final answer and the ground truth answer and is tasked with determining whether the generated answer is correct, serving as an approximate indicator of the validity of the associated reasoning chain. Once responses with incorrect answers are filtered out, the remaining reasoning processes are passed to a reasoning path scoring agent. Here, an advanced multimodal model, such as Qwen2-VL [38], is supplied with the image, question, reasoning path, and ground truth answer, and is prompted to evaluate the reasoning path. The scoring agent assesses each response based on the step-by-step accuracy of the reasoning path and the level of detail in the reasoning. To ensure consistency in scoring across different data samples, we aggregate all responses for each question and process them in a single pass. The model then generates scores for each response, ranging from 1 to 100." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.484, + 0.506 + ], + "angle": 0, + "content": "Through the above two steps, we construct a structured, high-quality dataset that provides detailed reasoning for each question, effectively supporting the training of our models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.516, + 0.233, + 0.532 + ], + "angle": 0, + "content": "3.3. Model Design" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.659 + ], + "angle": 0, + "content": "After constructing the dataset, we develop a multi-agent framework to enhance overall reasoning capabilities through collaborative agent interaction. Specifically, we first train a reasoning agent to generate a detailed reasoning process for each problem. Then, a summary agent is employed to answer the question, selectively utilizing the reasoning process based on its assessment. Together, these two agents collaborate to improve reasoning performance effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Reasoning Agent. Previous approaches typically combine reasoning and question-answering within a single process, which poses challenges for MLLMs. Generating a long-chain reasoning process can introduce errors, and directly answering questions based on flawed reasoning often leads to poorer results. To address this, we propose a specialized reasoning agent designed to generate a detailed, step-by-step reasoning process in response to an input query. We construct the reasoning dataset by selecting the highest-scoring reasoning path for each question. After training on this dataset, the model transforms into a reasoning agent with enhanced reasoning capabilities, enabling it to generate more detailed, structured reasoning processes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Summary Agent. Summarization plays a critical role in enabling models to accurately answer questions. After" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.92, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.283, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Figure 3. Overview of Insight-V Model Design. We derive a multi-agent system from a single model. By decomposing the task into reasoning and summarization, the two agents collaborate to enhance the overall reasoning capability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.909, + 0.49 + ], + "angle": 0, + "content": "generating multi-step reasoning, summarization provides a cohesive understanding of the reasoning process, ultimately guiding the model to the final answer. However, since the response generated by the reasoning agent may contain errors, we develop a summarization model robust to inaccuracies in the reasoning path, selectively incorporating or disregarding elements as needed. This approach maximizes the reasoning model's effectiveness while minimizing the risk of introducing misleading information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.909, + 0.779 + ], + "angle": 0, + "content": "To enhance the robustness of the summary agent, we carefully curate its training dataset. We utilize the collected dataset, which is comprised of two types: data with optimal reasoning processes and data with flawed reasoning processes for the summarization task. This method prevents the model from simply copying reasoning outcomes and encourages critical evaluation of reasoning quality. To further promote critical analysis by the summary agent, we select flawed reasoning samples based on their performance scores. Specifically, we draw flawed responses from varying score ranges to create a dataset with different levels of error, prompting the model to assess reasoning processes at various granularities. To better align the summary model with the reasoning agent, we also incorporate question-reasoning pairs generated by the reasoning agent to enhance collaboration between the two agents. Additionally, to preserve the original multi-modal capabilities, we supplement the dataset with standard question-answering data to sustain the summary agent's performance in direct question-answering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.685, + 0.804 + ], + "angle": 0, + "content": "3.4. Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The training pipeline for Insight-V is designed to be straightforward and efficient, utilizing a two-stage strategy. For both the reasoning agent and the summary agent, we begin with a well-trained MLLM. In the first stage, we apply supervised fine-tuning to train the agents to fulfill their designated roles using corresponding datasets. In the second stage, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9065" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "implement direct preference optimization, following prior research [45, 57]. This optimization is applied to the reasoning model, aligning it with human reasoning processes in a simple yet effective manner. These two stages enable the development of a robust system with enhanced visual reasoning capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.187, + 0.465, + 0.203 + ], + "angle": 0, + "content": "3.4.1. Supervised Fine-tuning for Multi-agent System" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.207, + 0.486, + 0.449 + ], + "angle": 0, + "content": "To perform supervised fine-tuning and obtain the two agents, we first train a base multi-modal model, following established methodologies. This model can address general visual question-answering tasks and gain foundational vision-language skills. We compile a high-quality image-text dataset focused on knowledge learning to train the base model. This data is sourced from various open-source academic datasets, including LLaVA-NeXT [22], Cauldron [15], and Cambrian-1 [43]. Once the base model is trained, we further fine-tune the two agents, initializing them from the base model. For the reasoning agent, we utilize the curated reasoning dataset to develop step-by-step reasoning capabilities. For the summary agent, we formulate a dataset as outlined in Section 3.3 and sample about one million general image-text pairs from the dataset used for the base model, preserving its original visual perception abilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.345, + 0.468 + ], + "angle": 0, + "content": "3.4.2. Enhanced Reasoning with RL" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Preference learning has gained increasing focus in the field of large language models. The primary aim is to fine-tune model outputs to align better with human (or expert) preferences, creating outputs more suited to real-world applications. Let's assume a preference dataset defined as \\(\\mathcal{D} = \\{(x^{(i)},y_w^{(i)},y_l^{(i)})\\}_{i = 1,\\dots ,|\\mathcal{D}|}\\), where each \\(x^{(i)}\\) is prompt, and \\(y_{w}^{(i)}\\) and \\(y_{l}^{(i)}\\) represent the preferred and less preferred responses, respectively. We denote \\(y_{w}\\succ y_{l}\\mid x\\) to signify that \\(y_{w}\\) is preferred over \\(y_{l}\\) for prompt \\(x\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Since the true distribution of human preferences cannot be directly observed, we approximate it with a latent reward model \\( r^{*}(x,y) \\), assuming that higher rewards correspond to stronger preferences. Following the approach by Rafailov et al. [39], we can model the human preference distribution \\( p^{*} \\) using the Bradley-Terry (BT) model [3]:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.71, + 0.462, + 0.762 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p ^ {*} (y _ {1} \\succ y _ {2} \\mid x) = \\frac {\\exp (r ^ {*} (x , y _ {1}))}{\\exp (r ^ {*} (x , y _ {1})) + \\exp (r ^ {*} (x , y _ {2}))} \\\\ = \\sigma \\left(r ^ {*} (x, y _ {1}) - r ^ {*} (x, y _ {2})\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.767, + 0.341, + 0.781 + ], + "angle": 0, + "content": "where \\(\\sigma\\) denotes the logistic function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.782, + 0.484, + 0.827 + ], + "angle": 0, + "content": "To estimate the parameters of the reward model, we can apply maximum likelihood estimation by minimizing the negative log-likelihood:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.833, + 0.483, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R} \\left(r _ {\\phi}, \\mathcal {D}\\right) = - \\mathbb {E} _ {\\left(x, y _ {w}, y _ {l}\\right) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(r _ {\\phi} \\left(x, y _ {w}\\right) - r _ {\\phi} \\left(x, y _ {l}\\right)\\right) \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "where \\(r_{\\phi}\\) is a parameterized reward model. This approach allows us to approximate the preference distribution and fine-tune the model to capture human-like preferences effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.35 + ], + "angle": 0, + "content": "The traditional DPO algorithm operates in an offline setting. During DPO training, as model parameters continuously change, the preference dataset generated offline can gradually diverge from the model's current distribution, which weakens the effectiveness of the DPO algorithm. To address this issue, we employ an iterative DPO algorithm. By conducting multiple rounds of DPO training and sampling, this approach enables the model to better approximate an online setting during training, thus further enhancing its performance. Specifically, our approach involves training a sequence of models \\(M_{1},\\ldots ,M_{T}\\), where each subsequent model \\(M_{t + 1}\\) utilizes preference data \\(\\mathcal{D}_t\\) generated by the \\(t\\)-th model. We apply this complete training process to the fine-tuned reasoning agent, enabling the model to better align with human preferences and produce structured, detailed reasoning steps for complex questions, which supports the summary agent more effectively." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.646, + 0.379 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.909, + 0.553 + ], + "angle": 0, + "content": "We conduct extensive experiments across multiple vision-language benchmarks to validate the effectiveness of our method. In this section, we first introduce the implementation details of Insight-V in Section 4.1. Then we present a comparison with state-of-the-art MLLMs, outlining the primary results of our method on visual reasoning tasks as well as additional results on general image understanding in Section 4.2. Moreover, we offer further analytical experiments and essential ablation studies on design choices in Section 4.3, along with qualitative results shown in Section 4.4 for more insights." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.731, + 0.578 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.81 + ], + "angle": 0, + "content": "We integrate the Insight-V system with various MLLMs to demonstrate the broad applicability of our approach. Our initial implementation with Insight-V on LLaVA-NeXT-LLaMA3 [22] illustrates the method's effectiveness. To further validate its generalizability and establish a solid baseline against state-of-the-art MLLMs, we additionally train a base multi-modal model using the Qwen-2.5-7B [41] LLM. During pretraining, we utilize the 558K captioning dataset from LLaVA-1.5 [21], unfreezing the connector parameters. This is followed by supervised fine-tuning with a curated dataset of approximately 4 million images, using a learning rate of 2e-5 as guided by prior researches [21, 22]. This two-stage training process equips the baseline model with essential visual perception abilities, achieving competitive results on vision-language benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We then initialize two agents from the baseline model, performing targeted fine-tuning to obtain the final agents. For the reasoning agent, we compile a dataset of 200K images and train the model over 2 epochs with a learning rate of 5e-6. For the summary agent, we use a dataset of 1.2 million images, applying a learning rate of 1e-5 and train" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9066" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. Results on Visual Reasoning Tasks. We conduct evaluation experiments across 7 benchmarks, covering both general reasoning and task-specific reasoning assessments. Insight-V exhibits notable effectiveness and generalizability when applied to LLaVA-NeXT and our baseline model, surpassing other state-of-the-art MLLMs by a large margin." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.143, + 0.907, + 0.383 + ], + "angle": 0, + "content": "
ModelSizeMMMUMMMU-ProMMBenchMMEChartQAMMStarMathVistaAverage
DeepSeek-VL [29]7B35.4-73.5-/-59.137.136.1-
VILA-1.5 [20]8B38.6-75.31634.9/--39.7--
Cambrian-1 [43]8B42.7-75.91547.1/-73.3-49.0-
InternLM-XComposer2 [7]7B41.1-77.62220.471.856.259.5-
POINTS [26]7B51.4-78.02184.1-60.963.0-
IXC-2.5 [55]7B42.9-79.42233.182.259.963.7-
Bunny-LLaMA3 [12]8B43.4-77.21588.9/321.1--34.4-
MM-1.5 [54]7B41.8--1514.9/346.478.6-47.6-
MiniCPM-LLaMA3-V 2.5 [49]8B45.819.677.22024.6-51.854.3-
MiniCPM-V-2.6 [50]7B49.827.278.02268.7-57.560.6-
Qwen2-VL [38]7B53.7-81.0-83.060.761.4-
Idefics3-LLaMA3 [14]8B46.622.977.51937.474.855.958.448.1
Ovis1.5-LLaMA3 [31]8B48.323.676.61948.576.457.363.049.4
LLaVA-NeXT-LLaMA3 [22]8B36.913.272.31611.1/346.069.443.145.940.2
+ Multi-Agent8B40.817.877.61603.7/469.374.652.647.444.5
+ Iterative DPO (Insight-V-LLaVA)8B42.021.081.71583.9/485.477.457.449.847.2 (+7.0)
Our Base Model7B47.122.681.31573.7/482.575.757.056.948.7
+ Multi-Agent7B49.723.882.21662.2/629.381.258.658.750.7
+ Iterative DPO (Insight-V)7B50.224.982.31685.1/627.081.561.559.951.6 (+2.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.395, + 0.486, + 0.452 + ], + "angle": 0, + "content": "Table 2. Results on other multimodal benchmarks. Insight-V enhances reasoning capabilities without compromising general visual perception and even achieves improvements on benchmarks requiring perception ability more." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.462, + 0.485, + 0.55 + ], + "angle": 0, + "content": "
ModelTextVQADocVQAOCRbenchAI2D
LLaVA-NeXT-LLaMA365.278.255371.5
+ Multi-Agent68.981.863175.7
+ Iterative DPO (Insight-V-LLaVA)70.582.966377.3
Our Base Model75.490.271379.7
+ Multi-Agent77.091.473880.1
+ Iterative DPO (Insight-V)76.891.573579.8
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.564, + 0.483, + 0.701 + ], + "angle": 0, + "content": "ing for 1 epoch. Additionally, we apply Direct Preference Optimization (DPO) to the reasoning agent, using approximately 15K preference data and training for 1 epoch at a learning rate of 5e-7. This DPO process is iteratively conducted across 3 rounds by using the model from the previous stage to generate preference data, thus enhancing the agent's reasoning capabilities. The lmms-eval [16] is utilized for fast evaluation. Further training details are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.712, + 0.391, + 0.729 + ], + "angle": 0, + "content": "4.2. Main Results on Visual Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Setup. We present comprehensive experimental results on various visual reasoning benchmarks, which require the model to handle complex and challenging questions, demonstrating strong reasoning capabilities. We select several well-known and representative benchmarks, covering comprehensive evaluation, chart understanding, and mathematical problem-solving. MMMU [52] and MMMU-Pro [53] evaluate the model's expert-level perception and reasoning abilities across various topics. MMBench [24] is a widely used benchmark for comprehensive MLLM evaluation on images. MME [10] includes 14 challenging subtasks for assessing" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.395, + 0.907, + 0.438 + ], + "angle": 0, + "content": "Table 3. Ablations on the Insight-V Design Choice. The multiagent design outperforms other configurations, highlighting the critical role of reasoning and summarization decomposition." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.448, + 0.907, + 0.527 + ], + "angle": 0, + "content": "
ModelMMMUChartQAMathVistaMMStarAvg
Baseline47.175.756.957.059.2
Vanilla - Direct SFT47.079.257.658.460.6
Multi-Turn Supervised48.179.657.958.261.0
Summary Agent Only47.576.357.357.959.8
Multi-Agent49.781.258.758.662.1
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.907, + 0.619 + ], + "angle": 0, + "content": "visual perception and cognition. ChartQA [33] focuses on logical reasoning with charts, while MathVista [30] assesses problem-solving skills in math, using GPT-4-0613 [35] as the evaluator, following standard practice. MMStar [4] tests a range of tasks with varying difficulty levels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Main Results. The experimental results in Table 1 confirm the effectiveness and generalizability of Insight-V on visual reasoning tasks. Applied to LLaVA-NeXT and our baseline model, the Insight-V system substantially boosts performance across challenging visual reasoning benchmarks, achieving competitive outcomes against state-of-the-art methods. We report an average improvement across all benchmarks: Insight-V yields an average increase of \\(4.9\\%\\) and \\(3.2\\%\\) when applied to LLaVA-NeXT and the base model, underscoring the system's effectiveness. Following Direct Preference Optimization, an additional improvement of \\(2.6\\%\\) and \\(1.0\\%\\) was observed, illustrating the value of reinforcement learning algorithms in enhancing reasoning capabilities. Specifically, on general reasoning benchmarks, MMMU, MMMU-Pro, and MMBench, we noted gains of \\(3.1\\%\\), \\(2.3\\%\\), and \\(1.0\\%\\), respectively. The highest improvement was on MME, with a \\(9.1\\%\\) increase, highlighting Insight-V's strength in tasks requiring both perception and" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9067" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.1, + 0.445, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.263, + 0.486, + 0.307 + ], + "angle": 0, + "content": "Figure 4. Ablations on the amount of training data. The reasoning agent benefits from data scaling, providing more valuable insights for the summary agent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.442 + ], + "angle": 0, + "content": "reasoning. On ChartQA, which entails complex chart-based reasoning, we achieved a \\(5.8\\%\\) increase. Performance gains on MMStar and MathVista, with improvements of \\(4.5\\%\\) and \\(3.0\\%\\), further validate our approach's versatility in tackling math problems that demand sophisticated reasoning and numerical calculation skills. This strong performance across all benchmarks underscores the efficacy of Insight-V and its potential to advance visual reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.571 + ], + "angle": 0, + "content": "Results on Other Multi-modal Benchmarks. To verify that Insight-V enhances reasoning capabilities without compromising basic visual perception skills, we conduct experiments on benchmarks requiring only fundamental image understanding. We selected TextVQ [40], DocVQA [34], OCRBench [25], and AI2D [13] for this purpose, as these benchmarks focus on basic visual interpretation rather than complex reasoning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.484, + 0.753 + ], + "angle": 0, + "content": "As shown in Table 2, integrating Insight-V enables models to maintain strong performance across general visual perception benchmarks. Additionally, applying Insight-V yields performance improvements on both LLaVA-NeXT and our base model, as it enhances general visual understanding by guiding the model's attention to relevant regions, thereby improving visual perception accuracy. Following DPO, the model demonstrates advanced reasoning capabilities while achieving competitive results on standard image understanding benchmarks compared to baseline models, underscoring Insight-V's robustness. Consistent results of both models further confirm Insight-V's generalizability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.764, + 0.258, + 0.78 + ], + "angle": 0, + "content": "4.3. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.484, + 0.862 + ], + "angle": 0, + "content": "In this section, we present comprehensive experiments to validate the design choices of Insight-V, emphasizing our approach's key contributions. Additionally, we include a case study to further demonstrate the qualitative effectiveness of Insight-V." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of Multi-agent System. To evaluate the effectiveness of the multi-agent system, we comprehensively" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 4. Ablations on the DPO training strategy. Iterative DPO progressively enhances the model's reasoning capabilities, leading to improved performance." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.143, + 0.907, + 0.21 + ], + "angle": 0, + "content": "
ModelMMMUChartQAMathVistaMMStarAvg
Insight-V (Multi-Agent)49.781.258.758.662.1
+ RLAIF49.581.459.159.262.3
+ DPO50.880.859.359.962.7
+ Iterative DPO50.281.559.961.563.3
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.223, + 0.907, + 0.403 + ], + "angle": 0, + "content": "compare Insight-V with alternative design choices. We begin by reporting the performance of the summary agent without a reasoning process to underscore the importance of the reasoning agent. For a comprehensive comparison, we also restructure the collected data using a Chain-of-Thought (CoT) template and train a model to perform sequential reasoning and subsequently answer questions. Additionally, we enable a model that can handle multi-turn conversations, facilitating reasoning and summarization in a multi-turn format. We use MMMU, ChartQA, MathVista, and MMStar as representative benchmarks to assess the performance across different methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.909, + 0.617 + ], + "angle": 0, + "content": "As shown in Table 3, the results indicate that the multi-agent system plays a critical role in enhancing the system's visual understanding capabilities. Using only the summary agent without a reasoning process results in limited improvements in reasoning tasks, as the model lacks the necessary reasoning framework for optimal performance. Training the model to perform chain-of-thoughts reasoning, which is denoted as Insight-V (Vanilla-Direct SFT), yields modest gains, as this approach does not sufficiently emphasize critical judgments within the reasoning process, merging reasoning and summarization into a single task. Furthermore, models trained with multi-turn conversations still produce sub-optimal results, underscoring the importance of a multi-agent system that separates reasoning and summarization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.909, + 0.849 + ], + "angle": 0, + "content": "Data Scaling Law of Reasoning Agent. To assess the effectiveness of the reasoning agent, we perform ablation experiments on the data volume for training. As shown in Figure 4, we compare reasoning agents trained on varying data sizes from 50K to 200K. For a fair and comprehensive comparison, we employ the same summary agent across evaluations and report results on six benchmarks. The findings clearly indicate that the reasoning agent benefits from increased data. With limited data, the reasoning agent struggles to generalize and fails to provide useful input to the summary model, resulting in performance even worse than baseline models. In contrast, training in larger datasets improves the capabilities of the reasoning agent, allowing it to perform step-by-step reasoning and provide valuable insights that help the summary agent solve tasks effectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Effects of RL Algorithms. We assess DPO algorithms to identify optimal strategies for enhancing the reasoning process. To examine the effects of dataset composition on DPO" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9068" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.116, + 0.217, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.105, + 0.445, + 0.122 + ], + "angle": 0, + "content": "Referring to Table 3-1, which is INCORRECT. If pots sell for $20 each then:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.125, + 0.444, + 0.134 + ], + "angle": 0, + "content": "(A) The marginal product of the third worker is five pots." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.134, + 0.456, + 0.157 + ], + "angle": 0, + "content": "(B) the marginal revenue product of labor equals the marginal product of labor multiplied by the additional revenue." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.157, + 0.456, + 0.173 + ], + "angle": 0, + "content": " (C) the marginal revenue product of labor of the second worker is $260." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.173, + 0.456, + 0.19 + ], + "angle": 0, + "content": "(D) the marginal revenue from selling the eighteenth pot is $20." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.125, + 0.456, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.743, + 0.095, + 0.859, + 0.108 + ], + "angle": 0, + "content": "Chain-of-Thought" + }, + { + "type": "text", + "bbox": [ + 0.477, + 0.112, + 0.854, + 0.139 + ], + "angle": 0, + "content": "A. ... It is incorrect because the marginal product of the third worker is 5 pots (18 - 13 = 5). However, the question incorrectly states that ... of the third worker is five pots, which should be ... of the second worker." + }, + { + "type": "text", + "bbox": [ + 0.478, + 0.139, + 0.866, + 0.156 + ], + "angle": 0, + "content": "B. ... This statement is correct. It is calculated by multiplying the marginal product of labor by the price per unit. In this case, ..." + }, + { + "type": "text", + "bbox": [ + 0.478, + 0.156, + 0.858, + 0.173 + ], + "angle": 0, + "content": " C......This statement is incorrect. The marginal product of the second worker is 5 pots, and the price per pot is $20, so ... is 5 * $20 = $100." + }, + { + "type": "text", + "bbox": [ + 0.478, + 0.173, + 0.844, + 0.181 + ], + "angle": 0, + "content": "D. ... This statement is correct. The ... is equal to the price per unit, which in this case is $20." + }, + { + "type": "text", + "bbox": [ + 0.478, + 0.181, + 0.56, + 0.19 + ], + "angle": 0, + "content": "So the answer is A." + }, + { + "type": "list", + "bbox": [ + 0.477, + 0.112, + 0.866, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.131, + 0.197, + 0.251, + 0.211 + ], + "angle": 0, + "content": "Insight-V (Vanilla)" + }, + { + "type": "title", + "bbox": [ + 0.722, + 0.197, + 0.872, + 0.21 + ], + "angle": 0, + "content": "Insight-V (Multi-Agent)" + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.221, + 0.208, + 0.232 + ], + "angle": 0, + "content": "DIRECT SFT" + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.238, + 0.16, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.237, + 0.401, + 0.255 + ], + "angle": 0, + "content": "Identifying Key Information: The image presents a table showing ... The question asks about ..." + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.268, + 0.159, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.258, + 0.401, + 0.293 + ], + "angle": 0, + "content": "Calculating Marginal Product of Labor: To find the marginal product of labor, ... For example, ... subtracting the total pots produced with two workers from that with three workers." + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.301, + 0.16, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.298, + 0.401, + 0.324 + ], + "angle": 0, + "content": "Determining Marginal Revenue Product: Next, we need to determine ... We multiply the marginal product of each worker by \\(20 to \\dots" + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.341, + 0.159, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.329, + 0.4, + 0.374 + ], + "angle": 0, + "content": "Comparing Options: Now that ... Option (A) states ..., which does not matches... . Option (B) suggests that ..., but this does not match ... . Option (C) indicates that ..., which also does not match ... . Option (D) claims that ..., which aligns with..." + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.381, + 0.159, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.38, + 0.397, + 0.399 + ], + "angle": 0, + "content": "FINAL ANSWER: (D) the marginal revenue from selling the eighteenth pot is $20." + }, + { + "type": "title", + "bbox": [ + 0.446, + 0.22, + 0.561, + 0.231 + ], + "angle": 0, + "content": "REASONING AGENT" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.238, + 0.466, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.478, + 0.237, + 0.863, + 0.255 + ], + "angle": 0, + "content": "Identifying Key Information: To begin solving this problem, we need to ... This involves understanding the table provided, which ..." + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.26, + 0.466, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.477, + 0.26, + 0.86, + 0.278 + ], + "angle": 0, + "content": "Analyzing Table Data: Next, we should analyze ... in Table 3-1 to ... . Specifically, we need to calculate... to determine which option is incorrect." + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.282, + 0.466, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.477, + 0.281, + 0.86, + 0.3 + ], + "angle": 0, + "content": "Comparing Marginal Products: Option (A) suggests .... Based on Table 3-1, we find ... (from 13 to 18), matching option (A)." + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.309, + 0.466, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.477, + 0.303, + 0.86, + 0.33 + ], + "angle": 0, + "content": "Calculating Marginal Revenue Product of Labor: Option (B) states.... It is consistent with option (B), as it takes into account both the marginal product and the additional revenue per unit of output." + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.336, + 0.466, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.477, + 0.333, + 0.863, + 0.351 + ], + "angle": 0, + "content": "Analyzing Option (C): Option (C) claims ... For the second worker, ... We can calculate that ... of the second worker is 7 multiplies \\(20, which is \\)140, contradicting option (C)." + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.357, + 0.466, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.477, + 0.355, + 0.866, + 0.373 + ], + "angle": 0, + "content": "Calculating Marginal Revenue: Option (D) indicates ... It can be ... that the pots sell $20 each. This means that ... is equal to $20, matching option (D)." + }, + { + "type": "title", + "bbox": [ + 0.446, + 0.377, + 0.55, + 0.387 + ], + "angle": 0, + "content": "SUMMARY AGENT" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.39, + 0.466, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.477, + 0.39, + 0.78, + 0.4 + ], + "angle": 0, + "content": "Based on the reasoning process above, The incorrect option is option (C)." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.415, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Figure 5. Qualitative Results of Insight-V. We present qualitative comparisons of Insight-V with Chain-of-Thought and Insight-V with direct SFT (Vanilla). For the Insight-V system, the reasoning agent delivers a more coherent and structured reasoning process, guiding the summary agent to the correct answer, whereas other methods struggle with complex reasoning and fail to solve such challenging problems." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.471, + 0.485, + 0.743 + ], + "angle": 0, + "content": "training, we compare our curated dataset against the widely-used RLAIF-V [51] dataset, which contains 80K DPO data pairs for alignment. We also investigate the potential benefits of iterative DPO in advancing the model's reasoning capabilities. For an unbiased comparison, we subsample the RLAIF-V dataset to approximately 15K data, aligning it with the size of our dataset. The results, summarized in Table 4, show an average improvement of approximately \\(0.2\\%\\) on the evaluated benchmarks when trained with the RLAIF-V dataset, whereas our curated dataset achieves greater gains of \\(0.6\\%\\). This suggests that a DPO dataset based on model-generated rationales, rather than externally sourced data, more effectively boosts reasoning accuracy. Moreover, conducting two additional rounds of DPO training using the same methodology yields further performance gains of \\(0.6\\%\\), underscoring that iterative DPO training progressively refines the model's ability to generate precise and high-quality reasoning processes compared to a single training pass." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.274, + 0.772 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We present qualitative comparisons in Figure 5 to illustrate the improvements introduced by integrating the Insight-V system. Specifically, we provide an example that highlights advancements in the reasoning process. In this example, the model is challenged with a multiple-choice question based on a table. We compare our approach with a direct Chain-of-Thought application, and the model undergoes supervised fine-tuning without incorporating a multi-agent sys" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.908, + 0.728 + ], + "angle": 0, + "content": "tem, which is denoted as Insight-V (Vanilla). It is evident that applying Chain-of-Thought directly results in suboptimal reasoning and leads to incorrect answers. Although fine-tuning a single model yields somewhat better reasoning, the model begins to falter as the reasoning chain lengthens, ultimately arriving at incorrect answers. This outcome arises because the model is required to generate both reasoning steps and the final answer simultaneously, limiting its judgment capabilities and weakening its robustness in handling flawed reasoning chains. In contrast, employing Insight-V enables more logical step decomposition and a structured reasoning chain. The model can analyze each option step-by-step and perform detailed calculations, which the other two methods cannot achieve. The fine-tuned summary model is able to evaluate the reasoning process and determine whether to derive the final answer based on it, significantly enhancing system robustness and ensuring correct answers." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.74, + 0.633, + 0.755 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this paper, we introduce Insight-V, a novel system that combines a scalable data generation pipeline for long-chain reasoning data and an effective multi-agent training pipeline to enhance the reasoning capabilities of MLLMs. By developing this system, we provide a scalable approach to model training aimed at improving reasoning performance. Our extensive evaluation across various benchmarks demonstrates the effectiveness of our approach, paving the way for equipping MLLMs with enhanced reasoning capabilities." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9069" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.171 + ], + "angle": 0, + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.484, + 0.283 + ], + "angle": 0, + "content": "[3] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[4] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.484, + 0.409 + ], + "angle": 0, + "content": "[5] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.484, + 0.48 + ], + "angle": 0, + "content": "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.563 + ], + "angle": 0, + "content": "[7] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[8] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[9] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[11] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[12] Muyang He, Yexin Liu, Boya Wu, Jianhao Yuan, Yueze Wang, Tiejun Huang, and Bo Zhao. Efficient multimodal learning from data-centric perspective. arXiv preprint arXiv:2402.11530, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[13] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In ECCV, pages 235-251. Springer, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[14] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. arXiv preprint arXiv:2408.12637, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[15] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[16] Bo Li. Xinrun du, yuhao dong, haotian liu, yuanhan zhang, ge zhang, chunyuan li, and ziwei liu. Lmms-eval: Accelerating the development of large multimoal models, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.907, + 0.347 + ], + "angle": 0, + "content": "[17] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.907, + 0.43 + ], + "angle": 0, + "content": "[18] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.907, + 0.501 + ], + "angle": 0, + "content": "[19] Zhenwen Liang, Tianyu Yang, Jipeng Zhang, and Xiangliang Zhang. Unimath: A foundational and multimodal mathematical reasoner. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7126-7133, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.907, + 0.558 + ], + "angle": 0, + "content": "[20] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.907, + 0.602 + ], + "angle": 0, + "content": "[21] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.907, + 0.644 + ], + "angle": 0, + "content": "[22] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.1,2,5,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[25] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.907, + 0.843 + ], + "angle": 0, + "content": "[26] Yuan Liu, Zhongyin Zhao, Ziyuan Zhuang, Le Tian, Xiao Zhou, and Jie Zhou. Points: Improving your vision-language model with affordable strategies. arXiv preprint arXiv:2409.04828, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[27] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9070" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.146 + ], + "angle": 0, + "content": "[28] Zuyan Liu, Yuhao Dong, Yongming Rao, Jie Zhou, and Jiwen Lu. Chain-of-spot: Interactive reasoning improves large vision-language models. arXiv preprint arXiv:2403.12966, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.485, + 0.217 + ], + "angle": 0, + "content": "[29] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Yaofeng Sun, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[30] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[31] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv:2405.20797, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[32] Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. Dolphins: Multimodal language model for driving. arXiv preprint arXiv:2312.00438, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[33] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[34] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.496, + 0.484, + 0.523 + ], + "angle": 0, + "content": "[35] OpenAI. Gpt-4 technical report. ArXiv:abs/2303.08774, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.524, + 0.461, + 0.538 + ], + "angle": 0, + "content": "[36] OpenAI. Hello gpt-4o — openai. OpenAI Blog, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.484, + 0.566 + ], + "angle": 0, + "content": "[37] QwenTeam. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[38] QwenTeam. Qwen2-vl: To see the world more clearly. Wwen Blog, 2024. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.663 + ], + "angle": 0, + "content": "[39] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[41] Qwen Team. Qwen2.5: A party of foundation models, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[42] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[43] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.91, + 0.161 + ], + "angle": 0, + "content": "[44] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.908, + 0.206 + ], + "angle": 0, + "content": "[45] Tianyi Xiong, Bo Li, Dong Guo, Huizhuo Yuan, Quanquan Gu, and Chunyuan Li. Llava-onevision-chat: Improving chat with preference learning, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[46] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.908, + 0.346 + ], + "angle": 0, + "content": "[47] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, et al. Octopus: Embodied vision-language programmer from environmental feedback. arXiv preprint arXiv:2310.08588, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.909, + 0.417 + ], + "angle": 0, + "content": "[48] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.908, + 0.517 + ], + "angle": 0, + "content": "[49] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint 2408.01800, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.908, + 0.573 + ], + "angle": 0, + "content": "[50] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.908, + 0.644 + ], + "angle": 0, + "content": "[51] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.909, + 0.73 + ], + "angle": 0, + "content": "[52] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.909, + 0.8 + ], + "angle": 0, + "content": "[53] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Ming Yin, Botao Yu, Ge Zhang, et al. Mmmu-pro: A more robust multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.801, + 0.909, + 0.87 + ], + "angle": 0, + "content": "[54] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Dufter, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, et al. Mm1. 5: Methods, analysis & insights from multimodal llm fine-tuning. arXiv preprint arXiv:2409.20566, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.909, + 0.902 + ], + "angle": 0, + "content": "[55] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.91, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9071" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.189 + ], + "angle": 0, + "content": "Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[56] Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Yichi Zhang, Ziyu Guo, Chengzhuo Tong, Jiaming Liu, Aojun Zhou, Bin Wei, Shanghang Zhang, et al. Mavis: Mathematical visual instruction tuning. arXiv preprint arXiv:2407.08739, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[57] Ruohong Zhang, Bowen Zhang, Yanghao Li, Haotian Zhang, Zhiqing Sun, Zhe Gan, Yinfei Yang, Ruoming Pang, and Yiming Yang. Improve vision language model chain-of-thought reasoning. arXiv preprint arXiv:2410.16198, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.318, + 0.484, + 0.372 + ], + "angle": 0, + "content": "[58] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9072" + } + ] +] \ No newline at end of file diff --git a/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/cb8122c4-a3cd-447b-9677-ef086abe8658_origin.pdf b/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/cb8122c4-a3cd-447b-9677-ef086abe8658_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e31f37b52094a252017e72ccef34889d93d1a130 --- /dev/null +++ b/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/cb8122c4-a3cd-447b-9677-ef086abe8658_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1f15da8de55687283e1ad0f46c2e515ee03a506ffdd3d89ebfc8315fc3251a5 +size 1192973 diff --git a/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/full.md b/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e2c9d01a7eae065955d85c35edd6eebb79d6b411 --- /dev/null +++ b/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/full.md @@ -0,0 +1,341 @@ +# Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models + +Yuhao Dong $^{1*}$ Zuyan Liu $^{2,3*}$ Hai-Long Sun $^{3,4}$ Jingkang Yang $^{1}$ Winston Hu $^{2}$ Yongming Rao $^{2\boxtimes}$ Ziwei Liu $^{1\boxtimes}$ $^{1}$ S-Lab, Nanyang Technological University $^{2}$ Tencent + $^{3}$ Tsinghua University $^{4}$ Nanjing University +https://github.com/dongyh20/Insight-V + +# Abstract + +Large Language Models (LLMs) demonstrate enhanced capabilities and reliability by reasoning more, evolving from Chain-of-Thought prompting to product-level solutions like OpenAI o1. Despite various efforts to improve LLM reasoning, high-quality long-chain reasoning data and optimized training pipelines still remain inadequately explored in vision-language tasks. In this paper, we present Insight-V, an early effort to 1) scalably produce long and robust reasoning data for complex multi-modal tasks, and 2) an effective training pipeline to enhance the reasoning capabilities of multi-modal large language models (MLLMs). Specifically, to create long and structured reasoning data without human labor, we design a two-step pipeline with a progressive strategy to generate sufficiently long and diverse reasoning paths and a multi-granularity assessment method to ensure data quality. We observe that directly supervising MLLMs with such long and complex reasoning data will not yield ideal reasoning ability. To tackle this problem, we design a multi-agent system consisting of a reasoning agent dedicated to performing long-chain reasoning and a summary agent trained to judge and summarize reasoning results. We further incorporate an iterative DPO algorithm to enhance the reasoning agent's generation stability and quality. Based on the popular LLaVA-NeXT model and our stronger base MLLM, we demonstrate significant performance gains across challenging multi-modal benchmarks requiring visual reasoning. Benefiting from our multi-agent system, Insight-V can also easily maintain or improve performance on perception-focused multi-modal tasks. + +# 1. Introduction + +The development of artificial general intelligence requires models that can seamlessly understand and respond to multi- + +![](images/909c928c7d743da8c8fe481207aa002a384c19e57aa64d5f63df3e8d9ce7c3a8.jpg) + +![](images/359944c99fc5a9953c0ada46ead258b10e388777a65b063fc90ca5fe55d907a2.jpg) +Figure 1. Illustration and Performance of Insight-V. Insight-V consists of two agents, one dedicated to reasoning and the other to summarization, driving significant improvements in performance across various visual reasoning benchmarks. + +modal data. Recent advancements in Large Language Models (LLMs) [9, 36, 37, 41] and Multi-modal LLMs (MLLMs) [6, 21-23, 29, 38, 49] have significantly facilitated this progress across various fields, ranging from common question-answering [6, 17, 22, 27, 38] to autonomous driving [32, 42] and robotics [8, 47]. Despite the substantial progress made in enhancing the performance of MLLMs on a wide range of tasks, enabling MLLMs to perform human-level reasoning remains a key challenge. This area remains underexplored and has yet to fully realize its potential. + +Existing efforts [44, 48] to enhance the reasoning capabilities of LLMs through long-chain reasoning have + +demonstrated considerable progress, largely benefiting from the availability of structured, high-quality data and well-established training pipelines. In contrast, teaching MLLMs to perform long-chain visual reasoning remains a significant challenge, primarily due to the lack of large-scale, high-quality datasets and efficient and effective training strategies. Compared to text-only data, visual reasoning data is not only more expensive to collect but also requires significant human labor for detailed annotation and validation, due to the absence of an effective data generation pipeline. Moreover, while previous work [58] has demonstrated that directly applying chain-of-thought [44] reasoning can improve the capabilities of MLLMs, other research [56, 57] suggests that current training approaches have limited effectiveness in enhancing CoT reasoning. This highlights the inability of current MLLMs to leverage visual cues for precise step-by-step problem-solving, emphasizing the need for an effective training procedure that enables MLLMs to reason in detail while maintaining clear visual perception. + +To address these challenges, we propose Insight-V, which incorporates two innovative designs to enhance reasoning capabilities. First, we introduce a data generation pipeline consisting of two key steps: a progressive strategy to generate structured, long-chain reasoning data with diverse reasoning paths, and a multi-granularity assessment system to evaluate and score these paths at different levels. Through automatic generation, assessment, and ranking strategies, the pipeline effectively operates without the need for human labor and makes the reasoning dataset more scalable for enhancing reasoning capabilities. To further improve MLLM reasoning beyond data scaling, we design a multi-agent system, as illustrated in Figure 1, that decomposes the problem-solving process into two distinct steps: reasoning and summarization. The reasoning agent generates a detailed reasoning process for the input query, while the summarization agent identifies key information within the reasoning process and selectively answers the question. To refine the quality of the reasoning, we employ an iterative DPO approach to enhance reasoning capabilities. The two agents collaborate to further improve the reasoning quality. Our findings demonstrate that this system significantly enhances the performance of various MLLMs across a broad range of visual reasoning benchmarks. + +We evaluate Insight-V by integrating our system into the widely used LLaVA-NeXT [22] model. To further highlight the potential of our design in advancing state-of-the-art models, we construct a robust base MLLM and apply our system to demonstrate its ability to enhance SOTA models across various visual reasoning benchmarks. Notably, Insight-V improves LLaVA-NeXT in an average performance of $7.0\%$ across seven challenging visual reasoning benchmarks, while a $2.9\%$ improvement is observed with the strong base MLLM, underscoring the effectiveness and generalizability + +of Insight-V. + +In summary, Insight-V offers 1) a scalable data generation pipeline for long-chain, high-quality reasoning data, 2) a multi-agent system that decomposes visual reasoning tasks into reasoning and summarization, and 3) a two-stage training pipeline to enhance visual reasoning capabilities. Together, these contributions address key challenges in visual reasoning, providing a solid foundation for future research in MLLM reasoning. + +# 2. Related Work + +# 2.1. Vision-Language Reasoning + +Recent advancements in MLLMs [1, 17, 20-23, 27, 29, 38] have equipped these models with robust capabilities across diverse domains, including visual understanding [20, 38], mathematics [19], college-level questions [6], and scientific inquiries. In visual understanding, most research [18, 22, 28, 43, 46] emphasizes fine-grained detail analysis and localization, training models to perform visual reasoning with tailored datasets to enhance interpretive capabilities. For mathematics and expert-level reasoning, existing methods [11, 56, 57] predominantly derive from Chain-of-Thought [44] approaches, training MLLMs to generate step-by-step reasoning across various subjects. However, these approaches often focus primarily on improving dataset quality through Chain-of-Thought, overlooking the importance of structured reasoning paths and extended reasoning chains in advancing model reasoning capabilities. Additionally, significant challenges arise when relying on a single model to manage the entire reasoning process for complex tasks, underscoring the need for a multi-agent system to decompose and enhance this process. In this work, we tackle these challenges by introducing a scalable reasoning data generation pipeline and implementing a multi-agent system for reasoning and summarization decomposition, enhancing the overall reasoning capabilities of existing MLLMs. + +# 2.2. Vision-Language Alignment + +To align the model more closely with human preferences, several alignment techniques are employed for MLLMs. A widely used approach is Reinforcement Learning from Human Feedback [2] (RLHF), which iteratively refines the model's responses based on human feedback, enhancing both response quality and interpretability. To further improve MLLM capabilities, Direct Preference Optimization [39] (DPO) is introduced to simplify the alignment process. By directly training on human preference data, DPO optimizes the model's outputs to better match human-selected responses. However, traditional DPO is primarily focused on offline scenarios, and as the model evolves, the effectiveness of this approach may significantly diminish. To address this, Iterative DPO [5] has been proposed, which optimizes pref + +erence pairs through DPO at each iteration. It then generates new preference pairs for the next iteration using the updated model and evaluates them with a reward model. In this paper, we use iterative DPO to achieve stronger alignment and enhance the model's reasoning capabilities. + +# 3. Methodology + +In this section, we provide a comprehensive description of the proposed Insight-V system, detailing its architecture and key contributions. Section 3.1 presents an overview of Insight-V, highlighting the core concepts of our approach. The design of Insight-V is structured around three primary components: 1) a carefully constructed pipeline for structured, scalable reasoning data generation, as described in Section 3.2; 2) a multi-agent MLLM system that supports complex visual reasoning, as detailed in Sections 3.3; and 3) a streamlined yet effective training pipeline to enhance overall performance, as outlined in Section 3.4. Together, these components form a cohesive system that effectively tackles the challenges of performing detailed, long-chain reasoning while preserving visual perception capabilities. + +# 3.1. Overview + +Advancing reasoning capabilities of LLMs has been a focal point of extensive research. Despite these efforts, the reasoning potential within multi-modal LLMs has barely been explored. Most approaches aim to strengthen reasoning at the inference stage, assuming the model has already acquired robust reasoning skills. Other approaches optimize model parameters using chain-of-thought data, enabling models to mimic human reasoning processes. However, these methods present significant challenges for current general-purpose MLLMs, as they require the model to develop reasoning skills while retaining prior capabilities, which often results in only modest performance gains. Additionally, the lack of structured, high-quality training data impedes training models with advanced reasoning capabilities. + +To fully leverage the reasoning capabilities of MLLMs, we propose Insight-V, a novel system comprising two MLLMs dedicated to reasoning and summarization, respectively. The reasoning model is tasked with generating a detailed reasoning process to assist in problem-solving, while the summary model evaluates this reasoning as supplementary information to assess its relevance and utility for answering the question. We also construct a structured, high-quality dataset to train both agents. We posit that this multi-agent system can enhance the reasoning strengths of MLLMs by decomposing the problem-solving process into distinct reasoning and summarization phases, thereby driving substantial performance improvements. + +![](images/952a18fd2a58e66250ef7efc80fbbba15454c2b5615fa32def0110ed62c8a683.jpg) +Figure 2. Data Generation Pipeline of Insight-V. The reasoning processes are generated progressively through a reasoning generator, and then fed into a multi-granularity assessment system to ensure high-quality reasoning. + +# 3.2. Construction of Structured Reasoning Data + +Previous studies [56, 58] have explored the integration of reasoning capabilities into MLLMs. However, training MLLMs to develop robust reasoning skills remains a considerable challenge, particularly due to data limitations. To address this, we introduce our data generation pipeline in this section, designed to produce high-quality, long-chain reasoning data using a progressive generation process and multi-granularity assessment. As shown in Figure 2, this scalable approach enables us to generate high-quality data to enhance the model's reasoning capabilities effectively. + +# Progressive Long-Chain Reasoning Data Generation. + +For each input query, we first employ a reasoning generator to produce a structured reasoning process in JSON format to address the problem. At each step, the reasoning generator provides a brief summary of the current step, a detailed reasoning response, and an action for the following step. If the action is continue, the model proceeds with an additional reasoning step in the next iteration; if the action is summary, the model generates a final summary and answer based on the complete reasoning process in the subsequent iteration. Specifically, For a multi-modal model $M$ , input image $I$ , and question $Q$ , the data generation process for each step is represented as follows: + +$$ +R _ {t} = M (I, Q, \left[ R _ {1} \dots R _ {t - 1} \right], A), +$$ + +$$ +R _ {a n s} = M (I, Q, \left[ R _ {1} \dots R _ {n} \right]), +$$ + +where $R_{t}$ and $R_{ans}$ denote the response at the $t$ -th step and the final answer, respectively, $R_{i}$ represents the reasoning generated by the model at the $i$ -th step, $n$ represents the total reasoning steps, and $A$ is the action determined in the previous step. By repeating this process $N$ times, we can iteratively sample $N$ structured responses for each query. The generation parameters are adjusted to encourage the + +model to produce outputs with various information and steps, allowing us to identify the most effective reasoning chain for each question. + +Multi-Granularity Assessment. After obtaining the structured responses, we utilize an assessment pipeline to ensure data quality. Specifically, we first apply a strong LLM, such as Qwen2 [37], for direct answer filtering. As shown in Figure 2, the model is provided with the generated final answer and the ground truth answer and is tasked with determining whether the generated answer is correct, serving as an approximate indicator of the validity of the associated reasoning chain. Once responses with incorrect answers are filtered out, the remaining reasoning processes are passed to a reasoning path scoring agent. Here, an advanced multimodal model, such as Qwen2-VL [38], is supplied with the image, question, reasoning path, and ground truth answer, and is prompted to evaluate the reasoning path. The scoring agent assesses each response based on the step-by-step accuracy of the reasoning path and the level of detail in the reasoning. To ensure consistency in scoring across different data samples, we aggregate all responses for each question and process them in a single pass. The model then generates scores for each response, ranging from 1 to 100. + +Through the above two steps, we construct a structured, high-quality dataset that provides detailed reasoning for each question, effectively supporting the training of our models. + +# 3.3. Model Design + +After constructing the dataset, we develop a multi-agent framework to enhance overall reasoning capabilities through collaborative agent interaction. Specifically, we first train a reasoning agent to generate a detailed reasoning process for each problem. Then, a summary agent is employed to answer the question, selectively utilizing the reasoning process based on its assessment. Together, these two agents collaborate to improve reasoning performance effectively. + +Reasoning Agent. Previous approaches typically combine reasoning and question-answering within a single process, which poses challenges for MLLMs. Generating a long-chain reasoning process can introduce errors, and directly answering questions based on flawed reasoning often leads to poorer results. To address this, we propose a specialized reasoning agent designed to generate a detailed, step-by-step reasoning process in response to an input query. We construct the reasoning dataset by selecting the highest-scoring reasoning path for each question. After training on this dataset, the model transforms into a reasoning agent with enhanced reasoning capabilities, enabling it to generate more detailed, structured reasoning processes. + +Summary Agent. Summarization plays a critical role in enabling models to accurately answer questions. After + +![](images/fb79206494bf4a64124f608becbcf2d97ebcac1cc39902ff3abf9ae690cb7102.jpg) +Figure 3. Overview of Insight-V Model Design. We derive a multi-agent system from a single model. By decomposing the task into reasoning and summarization, the two agents collaborate to enhance the overall reasoning capability. + +generating multi-step reasoning, summarization provides a cohesive understanding of the reasoning process, ultimately guiding the model to the final answer. However, since the response generated by the reasoning agent may contain errors, we develop a summarization model robust to inaccuracies in the reasoning path, selectively incorporating or disregarding elements as needed. This approach maximizes the reasoning model's effectiveness while minimizing the risk of introducing misleading information. + +To enhance the robustness of the summary agent, we carefully curate its training dataset. We utilize the collected dataset, which is comprised of two types: data with optimal reasoning processes and data with flawed reasoning processes for the summarization task. This method prevents the model from simply copying reasoning outcomes and encourages critical evaluation of reasoning quality. To further promote critical analysis by the summary agent, we select flawed reasoning samples based on their performance scores. Specifically, we draw flawed responses from varying score ranges to create a dataset with different levels of error, prompting the model to assess reasoning processes at various granularities. To better align the summary model with the reasoning agent, we also incorporate question-reasoning pairs generated by the reasoning agent to enhance collaboration between the two agents. Additionally, to preserve the original multi-modal capabilities, we supplement the dataset with standard question-answering data to sustain the summary agent's performance in direct question-answering. + +# 3.4. Training Pipeline + +The training pipeline for Insight-V is designed to be straightforward and efficient, utilizing a two-stage strategy. For both the reasoning agent and the summary agent, we begin with a well-trained MLLM. In the first stage, we apply supervised fine-tuning to train the agents to fulfill their designated roles using corresponding datasets. In the second stage, we + +implement direct preference optimization, following prior research [45, 57]. This optimization is applied to the reasoning model, aligning it with human reasoning processes in a simple yet effective manner. These two stages enable the development of a robust system with enhanced visual reasoning capabilities. + +# 3.4.1. Supervised Fine-tuning for Multi-agent System + +To perform supervised fine-tuning and obtain the two agents, we first train a base multi-modal model, following established methodologies. This model can address general visual question-answering tasks and gain foundational vision-language skills. We compile a high-quality image-text dataset focused on knowledge learning to train the base model. This data is sourced from various open-source academic datasets, including LLaVA-NeXT [22], Cauldron [15], and Cambrian-1 [43]. Once the base model is trained, we further fine-tune the two agents, initializing them from the base model. For the reasoning agent, we utilize the curated reasoning dataset to develop step-by-step reasoning capabilities. For the summary agent, we formulate a dataset as outlined in Section 3.3 and sample about one million general image-text pairs from the dataset used for the base model, preserving its original visual perception abilities. + +# 3.4.2. Enhanced Reasoning with RL + +Preference learning has gained increasing focus in the field of large language models. The primary aim is to fine-tune model outputs to align better with human (or expert) preferences, creating outputs more suited to real-world applications. Let's assume a preference dataset defined as $\mathcal{D} = \{(x^{(i)},y_w^{(i)},y_l^{(i)})\}_{i = 1,\dots ,|\mathcal{D}|}$ , where each $x^{(i)}$ is prompt, and $y_{w}^{(i)}$ and $y_{l}^{(i)}$ represent the preferred and less preferred responses, respectively. We denote $y_{w}\succ y_{l}\mid x$ to signify that $y_{w}$ is preferred over $y_{l}$ for prompt $x$ . + +Since the true distribution of human preferences cannot be directly observed, we approximate it with a latent reward model $r^{*}(x,y)$ , assuming that higher rewards correspond to stronger preferences. Following the approach by Rafailov et al. [39], we can model the human preference distribution $p^{*}$ using the Bradley-Terry (BT) model [3]: + +$$ +\begin{array}{l} p ^ {*} (y _ {1} \succ y _ {2} \mid x) = \frac {\exp (r ^ {*} (x , y _ {1}))}{\exp (r ^ {*} (x , y _ {1})) + \exp (r ^ {*} (x , y _ {2}))} \\ = \sigma \left(r ^ {*} (x, y _ {1}) - r ^ {*} (x, y _ {2})\right), \\ \end{array} +$$ + +where $\sigma$ denotes the logistic function. + +To estimate the parameters of the reward model, we can apply maximum likelihood estimation by minimizing the negative log-likelihood: + +$$ +\mathcal {L} _ {R} \left(r _ {\phi}, \mathcal {D}\right) = - \mathbb {E} _ {\left(x, y _ {w}, y _ {l}\right) \sim \mathcal {D}} \left[ \log \sigma \left(r _ {\phi} \left(x, y _ {w}\right) - r _ {\phi} \left(x, y _ {l}\right)\right) \right], +$$ + +where $r_{\phi}$ is a parameterized reward model. This approach allows us to approximate the preference distribution and fine-tune the model to capture human-like preferences effectively. + +The traditional DPO algorithm operates in an offline setting. During DPO training, as model parameters continuously change, the preference dataset generated offline can gradually diverge from the model's current distribution, which weakens the effectiveness of the DPO algorithm. To address this issue, we employ an iterative DPO algorithm. By conducting multiple rounds of DPO training and sampling, this approach enables the model to better approximate an online setting during training, thus further enhancing its performance. Specifically, our approach involves training a sequence of models $M_{1},\ldots ,M_{T}$ , where each subsequent model $M_{t + 1}$ utilizes preference data $\mathcal{D}_t$ generated by the $t$ -th model. We apply this complete training process to the fine-tuned reasoning agent, enabling the model to better align with human preferences and produce structured, detailed reasoning steps for complex questions, which supports the summary agent more effectively. + +# 4. Experiments + +We conduct extensive experiments across multiple vision-language benchmarks to validate the effectiveness of our method. In this section, we first introduce the implementation details of Insight-V in Section 4.1. Then we present a comparison with state-of-the-art MLLMs, outlining the primary results of our method on visual reasoning tasks as well as additional results on general image understanding in Section 4.2. Moreover, we offer further analytical experiments and essential ablation studies on design choices in Section 4.3, along with qualitative results shown in Section 4.4 for more insights. + +# 4.1. Implementation Details + +We integrate the Insight-V system with various MLLMs to demonstrate the broad applicability of our approach. Our initial implementation with Insight-V on LLaVA-NeXT-LLaMA3 [22] illustrates the method's effectiveness. To further validate its generalizability and establish a solid baseline against state-of-the-art MLLMs, we additionally train a base multi-modal model using the Qwen-2.5-7B [41] LLM. During pretraining, we utilize the 558K captioning dataset from LLaVA-1.5 [21], unfreezing the connector parameters. This is followed by supervised fine-tuning with a curated dataset of approximately 4 million images, using a learning rate of 2e-5 as guided by prior researches [21, 22]. This two-stage training process equips the baseline model with essential visual perception abilities, achieving competitive results on vision-language benchmarks. + +We then initialize two agents from the baseline model, performing targeted fine-tuning to obtain the final agents. For the reasoning agent, we compile a dataset of 200K images and train the model over 2 epochs with a learning rate of 5e-6. For the summary agent, we use a dataset of 1.2 million images, applying a learning rate of 1e-5 and train + +Table 1. Results on Visual Reasoning Tasks. We conduct evaluation experiments across 7 benchmarks, covering both general reasoning and task-specific reasoning assessments. Insight-V exhibits notable effectiveness and generalizability when applied to LLaVA-NeXT and our baseline model, surpassing other state-of-the-art MLLMs by a large margin. + +
ModelSizeMMMUMMMU-ProMMBenchMMEChartQAMMStarMathVistaAverage
DeepSeek-VL [29]7B35.4-73.5-/-59.137.136.1-
VILA-1.5 [20]8B38.6-75.31634.9/--39.7--
Cambrian-1 [43]8B42.7-75.91547.1/-73.3-49.0-
InternLM-XComposer2 [7]7B41.1-77.62220.471.856.259.5-
POINTS [26]7B51.4-78.02184.1-60.963.0-
IXC-2.5 [55]7B42.9-79.42233.182.259.963.7-
Bunny-LLaMA3 [12]8B43.4-77.21588.9/321.1--34.4-
MM-1.5 [54]7B41.8--1514.9/346.478.6-47.6-
MiniCPM-LLaMA3-V 2.5 [49]8B45.819.677.22024.6-51.854.3-
MiniCPM-V-2.6 [50]7B49.827.278.02268.7-57.560.6-
Qwen2-VL [38]7B53.7-81.0-83.060.761.4-
Idefics3-LLaMA3 [14]8B46.622.977.51937.474.855.958.448.1
Ovis1.5-LLaMA3 [31]8B48.323.676.61948.576.457.363.049.4
LLaVA-NeXT-LLaMA3 [22]8B36.913.272.31611.1/346.069.443.145.940.2
+ Multi-Agent8B40.817.877.61603.7/469.374.652.647.444.5
+ Iterative DPO (Insight-V-LLaVA)8B42.021.081.71583.9/485.477.457.449.847.2 (+7.0)
Our Base Model7B47.122.681.31573.7/482.575.757.056.948.7
+ Multi-Agent7B49.723.882.21662.2/629.381.258.658.750.7
+ Iterative DPO (Insight-V)7B50.224.982.31685.1/627.081.561.559.951.6 (+2.9)
+ +Table 2. Results on other multimodal benchmarks. Insight-V enhances reasoning capabilities without compromising general visual perception and even achieves improvements on benchmarks requiring perception ability more. + +
ModelTextVQADocVQAOCRbenchAI2D
LLaVA-NeXT-LLaMA365.278.255371.5
+ Multi-Agent68.981.863175.7
+ Iterative DPO (Insight-V-LLaVA)70.582.966377.3
Our Base Model75.490.271379.7
+ Multi-Agent77.091.473880.1
+ Iterative DPO (Insight-V)76.891.573579.8
+ +ing for 1 epoch. Additionally, we apply Direct Preference Optimization (DPO) to the reasoning agent, using approximately 15K preference data and training for 1 epoch at a learning rate of 5e-7. This DPO process is iteratively conducted across 3 rounds by using the model from the previous stage to generate preference data, thus enhancing the agent's reasoning capabilities. The lmms-eval [16] is utilized for fast evaluation. Further training details are provided in the appendix. + +# 4.2. Main Results on Visual Reasoning + +Setup. We present comprehensive experimental results on various visual reasoning benchmarks, which require the model to handle complex and challenging questions, demonstrating strong reasoning capabilities. We select several well-known and representative benchmarks, covering comprehensive evaluation, chart understanding, and mathematical problem-solving. MMMU [52] and MMMU-Pro [53] evaluate the model's expert-level perception and reasoning abilities across various topics. MMBench [24] is a widely used benchmark for comprehensive MLLM evaluation on images. MME [10] includes 14 challenging subtasks for assessing + +Table 3. Ablations on the Insight-V Design Choice. The multiagent design outperforms other configurations, highlighting the critical role of reasoning and summarization decomposition. + +
ModelMMMUChartQAMathVistaMMStarAvg
Baseline47.175.756.957.059.2
Vanilla - Direct SFT47.079.257.658.460.6
Multi-Turn Supervised48.179.657.958.261.0
Summary Agent Only47.576.357.357.959.8
Multi-Agent49.781.258.758.662.1
+ +visual perception and cognition. ChartQA [33] focuses on logical reasoning with charts, while MathVista [30] assesses problem-solving skills in math, using GPT-4-0613 [35] as the evaluator, following standard practice. MMStar [4] tests a range of tasks with varying difficulty levels. + +Main Results. The experimental results in Table 1 confirm the effectiveness and generalizability of Insight-V on visual reasoning tasks. Applied to LLaVA-NeXT and our baseline model, the Insight-V system substantially boosts performance across challenging visual reasoning benchmarks, achieving competitive outcomes against state-of-the-art methods. We report an average improvement across all benchmarks: Insight-V yields an average increase of $4.9\%$ and $3.2\%$ when applied to LLaVA-NeXT and the base model, underscoring the system's effectiveness. Following Direct Preference Optimization, an additional improvement of $2.6\%$ and $1.0\%$ was observed, illustrating the value of reinforcement learning algorithms in enhancing reasoning capabilities. Specifically, on general reasoning benchmarks, MMMU, MMMU-Pro, and MMBench, we noted gains of $3.1\%$ , $2.3\%$ , and $1.0\%$ , respectively. The highest improvement was on MME, with a $9.1\%$ increase, highlighting Insight-V's strength in tasks requiring both perception and + +![](images/a244efeef3264345029049f015ebfe359b3db79d1890b57e3ed256e25ecda321.jpg) +Figure 4. Ablations on the amount of training data. The reasoning agent benefits from data scaling, providing more valuable insights for the summary agent. + +reasoning. On ChartQA, which entails complex chart-based reasoning, we achieved a $5.8\%$ increase. Performance gains on MMStar and MathVista, with improvements of $4.5\%$ and $3.0\%$ , further validate our approach's versatility in tackling math problems that demand sophisticated reasoning and numerical calculation skills. This strong performance across all benchmarks underscores the efficacy of Insight-V and its potential to advance visual reasoning capabilities. + +Results on Other Multi-modal Benchmarks. To verify that Insight-V enhances reasoning capabilities without compromising basic visual perception skills, we conduct experiments on benchmarks requiring only fundamental image understanding. We selected TextVQ [40], DocVQA [34], OCRBench [25], and AI2D [13] for this purpose, as these benchmarks focus on basic visual interpretation rather than complex reasoning. + +As shown in Table 2, integrating Insight-V enables models to maintain strong performance across general visual perception benchmarks. Additionally, applying Insight-V yields performance improvements on both LLaVA-NeXT and our base model, as it enhances general visual understanding by guiding the model's attention to relevant regions, thereby improving visual perception accuracy. Following DPO, the model demonstrates advanced reasoning capabilities while achieving competitive results on standard image understanding benchmarks compared to baseline models, underscoring Insight-V's robustness. Consistent results of both models further confirm Insight-V's generalizability. + +# 4.3. Further Analysis + +In this section, we present comprehensive experiments to validate the design choices of Insight-V, emphasizing our approach's key contributions. Additionally, we include a case study to further demonstrate the qualitative effectiveness of Insight-V. + +Effectiveness of Multi-agent System. To evaluate the effectiveness of the multi-agent system, we comprehensively + +Table 4. Ablations on the DPO training strategy. Iterative DPO progressively enhances the model's reasoning capabilities, leading to improved performance. + +
ModelMMMUChartQAMathVistaMMStarAvg
Insight-V (Multi-Agent)49.781.258.758.662.1
+ RLAIF49.581.459.159.262.3
+ DPO50.880.859.359.962.7
+ Iterative DPO50.281.559.961.563.3
+ +compare Insight-V with alternative design choices. We begin by reporting the performance of the summary agent without a reasoning process to underscore the importance of the reasoning agent. For a comprehensive comparison, we also restructure the collected data using a Chain-of-Thought (CoT) template and train a model to perform sequential reasoning and subsequently answer questions. Additionally, we enable a model that can handle multi-turn conversations, facilitating reasoning and summarization in a multi-turn format. We use MMMU, ChartQA, MathVista, and MMStar as representative benchmarks to assess the performance across different methods. + +As shown in Table 3, the results indicate that the multi-agent system plays a critical role in enhancing the system's visual understanding capabilities. Using only the summary agent without a reasoning process results in limited improvements in reasoning tasks, as the model lacks the necessary reasoning framework for optimal performance. Training the model to perform chain-of-thoughts reasoning, which is denoted as Insight-V (Vanilla-Direct SFT), yields modest gains, as this approach does not sufficiently emphasize critical judgments within the reasoning process, merging reasoning and summarization into a single task. Furthermore, models trained with multi-turn conversations still produce sub-optimal results, underscoring the importance of a multi-agent system that separates reasoning and summarization. + +Data Scaling Law of Reasoning Agent. To assess the effectiveness of the reasoning agent, we perform ablation experiments on the data volume for training. As shown in Figure 4, we compare reasoning agents trained on varying data sizes from 50K to 200K. For a fair and comprehensive comparison, we employ the same summary agent across evaluations and report results on six benchmarks. The findings clearly indicate that the reasoning agent benefits from increased data. With limited data, the reasoning agent struggles to generalize and fails to provide useful input to the summary model, resulting in performance even worse than baseline models. In contrast, training in larger datasets improves the capabilities of the reasoning agent, allowing it to perform step-by-step reasoning and provide valuable insights that help the summary agent solve tasks effectively. + +Effects of RL Algorithms. We assess DPO algorithms to identify optimal strategies for enhancing the reasoning process. To examine the effects of dataset composition on DPO + +![](images/3c7a25ec1fd61936835c8b31aeb85628a7776f1ee1f0180028a4b27df3229af3.jpg) + +Referring to Table 3-1, which is INCORRECT. If pots sell for $20 each then: + +(A) The marginal product of the third worker is five pots. +(B) the marginal revenue product of labor equals the marginal product of labor multiplied by the additional revenue. + (C) the marginal revenue product of labor of the second worker is $260. +(D) the marginal revenue from selling the eighteenth pot is $20. + +# Chain-of-Thought + +A. ... It is incorrect because the marginal product of the third worker is 5 pots (18 - 13 = 5). However, the question incorrectly states that ... of the third worker is five pots, which should be ... of the second worker. +B. ... This statement is correct. It is calculated by multiplying the marginal product of labor by the price per unit. In this case, ... + C......This statement is incorrect. The marginal product of the second worker is 5 pots, and the price per pot is $20, so ... is 5 * $20 = $100. +D. ... This statement is correct. The ... is equal to the price per unit, which in this case is $20. +So the answer is A. + +# Insight-V (Vanilla) + +# Insight-V (Multi-Agent) + +# DIRECT SFT + +![](images/cfcce8980bd3b9fc188f00aa812a4d81cfb9c0eb96fdfb368a54bfe6f451fb24.jpg) + +Identifying Key Information: The image presents a table showing ... The question asks about ... + +![](images/778b689d6fe9a1c351ac842731aded0a7e487ba15f80770e6cfa87242cd79189.jpg) + +Calculating Marginal Product of Labor: To find the marginal product of labor, ... For example, ... subtracting the total pots produced with two workers from that with three workers. + +![](images/ee4fd4c6ffcaf143575a303ab2ee389b42bed6d4f3abaec1678d21f5a47019c3.jpg) + +Determining Marginal Revenue Product: Next, we need to determine ... We multiply the marginal product of each worker by \(20 to \dots + +![](images/307d65e41ad124e925b0ede0019b72ec5c574315f66f158b4e0bd54cf40d64ba.jpg) + +Comparing Options: Now that ... Option (A) states ..., which does not matches... . Option (B) suggests that ..., but this does not match ... . Option (C) indicates that ..., which also does not match ... . Option (D) claims that ..., which aligns with... + +![](images/1f970d9032c04744ec79cedb7ee0cbfe3121a27642079c5318b859502a6af015.jpg) +Figure 5. Qualitative Results of Insight-V. We present qualitative comparisons of Insight-V with Chain-of-Thought and Insight-V with direct SFT (Vanilla). For the Insight-V system, the reasoning agent delivers a more coherent and structured reasoning process, guiding the summary agent to the correct answer, whereas other methods struggle with complex reasoning and fail to solve such challenging problems. + +FINAL ANSWER: (D) the marginal revenue from selling the eighteenth pot is $20. + +# REASONING AGENT + +![](images/29d50873d1110d083270fec76951095889a5e45aff2cdcc6e0dbed0bb578fcb4.jpg) + +Identifying Key Information: To begin solving this problem, we need to ... This involves understanding the table provided, which ... + +![](images/afbc7a818f7a055823802f6eb3d8d40929037def6e8959b6eacdfe9420e9de78.jpg) + +Analyzing Table Data: Next, we should analyze ... in Table 3-1 to ... . Specifically, we need to calculate... to determine which option is incorrect. + +![](images/767b7adf363d8251449e2d718ed93946cd5cf7001697b766315abfa4d77c2e7f.jpg) + +Comparing Marginal Products: Option (A) suggests .... Based on Table 3-1, we find ... (from 13 to 18), matching option (A). + +![](images/b682c5e385bd75a15c1d2b276cb37078c4522e84fad83bd37ef1a89a990cc93d.jpg) + +Calculating Marginal Revenue Product of Labor: Option (B) states.... It is consistent with option (B), as it takes into account both the marginal product and the additional revenue per unit of output. + +![](images/b0220de4bbd6d3766239bb19dff80782ce63b64e3fb19ed27b1a3f836751f7d0.jpg) + +Analyzing Option (C): Option (C) claims ... For the second worker, ... We can calculate that ... of the second worker is 7 multiplies $20, which is$ 140, contradicting option (C). + +![](images/553a02bf49b2decbc863c0c383795908b9070b1c39f5494fc5d8e1b880fba4ed.jpg) + +Calculating Marginal Revenue: Option (D) indicates ... It can be ... that the pots sell $20 each. This means that ... is equal to $20, matching option (D). + +# SUMMARY AGENT + +![](images/1c59585149320464353ff81377b9ff8fd6f9e297dd040872780c31df26c96a79.jpg) + +Based on the reasoning process above, The incorrect option is option (C). + +training, we compare our curated dataset against the widely-used RLAIF-V [51] dataset, which contains 80K DPO data pairs for alignment. We also investigate the potential benefits of iterative DPO in advancing the model's reasoning capabilities. For an unbiased comparison, we subsample the RLAIF-V dataset to approximately 15K data, aligning it with the size of our dataset. The results, summarized in Table 4, show an average improvement of approximately $0.2\%$ on the evaluated benchmarks when trained with the RLAIF-V dataset, whereas our curated dataset achieves greater gains of $0.6\%$ . This suggests that a DPO dataset based on model-generated rationales, rather than externally sourced data, more effectively boosts reasoning accuracy. Moreover, conducting two additional rounds of DPO training using the same methodology yields further performance gains of $0.6\%$ , underscoring that iterative DPO training progressively refines the model's ability to generate precise and high-quality reasoning processes compared to a single training pass. + +# 4.4. Qualitative Results + +We present qualitative comparisons in Figure 5 to illustrate the improvements introduced by integrating the Insight-V system. Specifically, we provide an example that highlights advancements in the reasoning process. In this example, the model is challenged with a multiple-choice question based on a table. We compare our approach with a direct Chain-of-Thought application, and the model undergoes supervised fine-tuning without incorporating a multi-agent sys + +tem, which is denoted as Insight-V (Vanilla). It is evident that applying Chain-of-Thought directly results in suboptimal reasoning and leads to incorrect answers. Although fine-tuning a single model yields somewhat better reasoning, the model begins to falter as the reasoning chain lengthens, ultimately arriving at incorrect answers. This outcome arises because the model is required to generate both reasoning steps and the final answer simultaneously, limiting its judgment capabilities and weakening its robustness in handling flawed reasoning chains. In contrast, employing Insight-V enables more logical step decomposition and a structured reasoning chain. The model can analyze each option step-by-step and perform detailed calculations, which the other two methods cannot achieve. The fine-tuned summary model is able to evaluate the reasoning process and determine whether to derive the final answer based on it, significantly enhancing system robustness and ensuring correct answers. + +# 5. Conclusion + +In this paper, we introduce Insight-V, a novel system that combines a scalable data generation pipeline for long-chain reasoning data and an effective multi-agent training pipeline to enhance the reasoning capabilities of MLLMs. By developing this system, we provide a scalable approach to model training aimed at improving reasoning performance. Our extensive evaluation across various benchmarks demonstrates the effectiveness of our approach, paving the way for equipping MLLMs with enhanced reasoning capabilities. + +# References + +[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. 2023. 2 +[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 2 +[3] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 5 +[4] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 6 +[5] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 2 +[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1, 2 +[7] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 6 +[8] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1 +[9] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1 +[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 6 +[11] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023. 2 +[12] Muyang He, Yexin Liu, Boya Wu, Jianhao Yuan, Yueze Wang, Tiejun Huang, and Bo Zhao. Efficient multimodal learning from data-centric perspective. arXiv preprint arXiv:2402.11530, 2024. 6 + +[13] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In ECCV, pages 235-251. Springer, 2016. 7 +[14] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. arXiv preprint arXiv:2408.12637, 2024. 6 +[15] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 5 +[16] Bo Li. Xinrun du, yuhao dong, haotian liu, yuanhan zhang, ge zhang, chunyuan li, and ziwei liu. Lmms-eval: Accelerating the development of large multimoal models, 2024. 6 +[17] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2 +[18] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 2 +[19] Zhenwen Liang, Tianyu Yang, Jipeng Zhang, and Xiangliang Zhang. Unimath: A foundational and multimodal mathematical reasoner. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7126-7133, 2023. 2 +[20] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2, 6 +[21] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 5 +[22] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.1,2,5,6 +[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1, 2 +[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 6 +[25] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 7 +[26] Yuan Liu, Zhongyin Zhao, Ziyuan Zhuang, Le Tian, Xiao Zhou, and Jie Zhou. Points: Improving your vision-language model with affordable strategies. arXiv preprint arXiv:2409.04828, 2024. 6 +[27] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 1, 2 + +[28] Zuyan Liu, Yuhao Dong, Yongming Rao, Jie Zhou, and Jiwen Lu. Chain-of-spot: Interactive reasoning improves large vision-language models. arXiv preprint arXiv:2403.12966, 2024. 2 +[29] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Yaofeng Sun, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 1, 2, 6 +[30] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 6 +[31] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv:2405.20797, 2024. 6 +[32] Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. Dolphins: Multimodal language model for driving. arXiv preprint arXiv:2312.00438, 2023. 1 +[33] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 6 +[34] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 7 +[35] OpenAI. Gpt-4 technical report. ArXiv:abs/2303.08774, 2023.6 +[36] OpenAI. Hello gpt-4o — openai. OpenAI Blog, 2024. 1 +[37] QwenTeam. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1, 4 +[38] QwenTeam. Qwen2-vl: To see the world more clearly. Wwen Blog, 2024. 1, 2, 4, 6 +[39] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2, 5 +[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 7 +[41] Qwen Team. Qwen2.5: A party of foundation models, 2024. 1, 5 +[42] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1 +[43] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 2, 5, 6 + +[44] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2 +[45] Tianyi Xiong, Bo Li, Dong Guo, Huizhuo Yuan, Quanquan Gu, and Chunyuan Li. Llava-onevision-chat: Improving chat with preference learning, 2024. 5 +[46] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 2 +[47] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, et al. Octopus: Embodied vision-language programmer from environmental feedback. arXiv preprint arXiv:2310.08588, 2023. 1 +[48] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 1 +[49] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint 2408.01800, 2024. 1, 6 +[50] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 6 +[51] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 8 +[52] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 6 +[53] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Ming Yin, Botao Yu, Ge Zhang, et al. Mmmu-pro: A more robust multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024.6 +[54] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Dufter, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, et al. Mm1. 5: Methods, analysis & insights from multimodal llm fine-tuning. arXiv preprint arXiv:2409.20566, 2024. 6 +[55] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, + +Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 6 +[56] Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Yichi Zhang, Ziyu Guo, Chengzhuo Tong, Jiaming Liu, Aojun Zhou, Bin Wei, Shanghang Zhang, et al. Mavis: Mathematical visual instruction tuning. arXiv preprint arXiv:2407.08739, 2024. 2, 3 +[57] Ruohong Zhang, Bowen Zhang, Yanghao Li, Haotian Zhang, Zhiqing Sun, Zhe Gan, Yinfei Yang, Ruoming Pang, and Yiming Yang. Improve vision language model chain-of-thought reasoning. arXiv preprint arXiv:2410.16198, 2024. 2, 5 +[58] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2, 3 \ No newline at end of file diff --git a/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/images.zip b/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..22bb4bd924df0ab9a578a15f7d2454f8c8cc11a5 --- /dev/null +++ b/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c10fc9027c45c4ebecb1e5d0e10f11f6444220b4bafc32ab8e2aa8d5d4959d4 +size 384552 diff --git a/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/layout.json b/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..50ab5cd052597e6cc3448e95ceacd81d72b76518 --- /dev/null +++ b/2025/Insight-V_ Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models/layout.json @@ -0,0 +1,8636 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 132, + 102, + 479, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 102, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 102, + 479, + 140 + ], + "type": "text", + "content": "Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "spans": [ + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "text", + "content": "Yuhao Dong" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "text", + "content": " Zuyan Liu" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "text", + "content": " Hai-Long Sun" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "text", + "content": " Jingkang Yang" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "text", + "content": " Winston Hu" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "text", + "content": " Yongming Rao" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "inline_equation", + "content": "^{2\\boxtimes}" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "text", + "content": " Ziwei Liu" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "text", + "content": " S-Lab, Nanyang Technological University " + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "text", + "content": " Tencent \n" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "text", + "content": " Tsinghua University " + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 145, + 160, + 465, + 232 + ], + "type": "text", + "content": " Nanjing University \nhttps://github.com/dongyh20/Insight-V" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 284, + 297, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 297, + 632 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 297, + 632 + ], + "type": "text", + "content": "Large Language Models (LLMs) demonstrate enhanced capabilities and reliability by reasoning more, evolving from Chain-of-Thought prompting to product-level solutions like OpenAI o1. Despite various efforts to improve LLM reasoning, high-quality long-chain reasoning data and optimized training pipelines still remain inadequately explored in vision-language tasks. In this paper, we present Insight-V, an early effort to 1) scalably produce long and robust reasoning data for complex multi-modal tasks, and 2) an effective training pipeline to enhance the reasoning capabilities of multi-modal large language models (MLLMs). Specifically, to create long and structured reasoning data without human labor, we design a two-step pipeline with a progressive strategy to generate sufficiently long and diverse reasoning paths and a multi-granularity assessment method to ensure data quality. We observe that directly supervising MLLMs with such long and complex reasoning data will not yield ideal reasoning ability. To tackle this problem, we design a multi-agent system consisting of a reasoning agent dedicated to performing long-chain reasoning and a summary agent trained to judge and summarize reasoning results. We further incorporate an iterative DPO algorithm to enhance the reasoning agent's generation stability and quality. Based on the popular LLaVA-NeXT model and our stronger base MLLM, we demonstrate significant performance gains across challenging multi-modal benchmarks requiring visual reasoning. Benefiting from our multi-agent system, Insight-V can also easily maintain or improve performance on perception-focused multi-modal tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 652, + 137, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 137, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 137, + 665 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 672, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 296, + 696 + ], + "type": "text", + "content": "The development of artificial general intelligence requires models that can seamlessly understand and respond to multi-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 259, + 563, + 437 + ], + "blocks": [ + { + "bbox": [ + 318, + 259, + 563, + 437 + ], + "lines": [ + { + "bbox": [ + 318, + 259, + 563, + 437 + ], + "spans": [ + { + "bbox": [ + 318, + 259, + 563, + 437 + ], + "type": "image", + "image_path": "909c928c7d743da8c8fe481207aa002a384c19e57aa64d5f63df3e8d9ce7c3a8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 438, + 564, + 500 + ], + "blocks": [ + { + "bbox": [ + 318, + 438, + 564, + 500 + ], + "lines": [ + { + "bbox": [ + 318, + 438, + 564, + 500 + ], + "spans": [ + { + "bbox": [ + 318, + 438, + 564, + 500 + ], + "type": "image", + "image_path": "359944c99fc5a9953c0ada46ead258b10e388777a65b063fc90ca5fe55d907a2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 511, + 555, + 556 + ], + "lines": [ + { + "bbox": [ + 313, + 511, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 555, + 556 + ], + "type": "text", + "content": "Figure 1. Illustration and Performance of Insight-V. Insight-V consists of two agents, one dedicated to reasoning and the other to summarization, driving significant improvements in performance across various visual reasoning benchmarks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 569, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 569, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 569, + 556, + 689 + ], + "type": "text", + "content": "modal data. Recent advancements in Large Language Models (LLMs) [9, 36, 37, 41] and Multi-modal LLMs (MLLMs) [6, 21-23, 29, 38, 49] have significantly facilitated this progress across various fields, ranging from common question-answering [6, 17, 22, 27, 38] to autonomous driving [32, 42] and robotics [8, 47]. Despite the substantial progress made in enhancing the performance of MLLMs on a wide range of tasks, enabling MLLMs to perform human-level reasoning remains a key challenge. This area remains underexplored and has yet to fully realize its potential." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Existing efforts [44, 48] to enhance the reasoning capabilities of LLMs through long-chain reasoning have" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "type": "text", + "content": "*Authors contributed equally to this research. Corresponding authors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9062" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 295, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 295, + 299 + ], + "type": "text", + "content": "demonstrated considerable progress, largely benefiting from the availability of structured, high-quality data and well-established training pipelines. In contrast, teaching MLLMs to perform long-chain visual reasoning remains a significant challenge, primarily due to the lack of large-scale, high-quality datasets and efficient and effective training strategies. Compared to text-only data, visual reasoning data is not only more expensive to collect but also requires significant human labor for detailed annotation and validation, due to the absence of an effective data generation pipeline. Moreover, while previous work [58] has demonstrated that directly applying chain-of-thought [44] reasoning can improve the capabilities of MLLMs, other research [56, 57] suggests that current training approaches have limited effectiveness in enhancing CoT reasoning. This highlights the inability of current MLLMs to leverage visual cues for precise step-by-step problem-solving, emphasizing the need for an effective training procedure that enables MLLMs to reason in detail while maintaining clear visual perception." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 304, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 304, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 304, + 295, + 589 + ], + "type": "text", + "content": "To address these challenges, we propose Insight-V, which incorporates two innovative designs to enhance reasoning capabilities. First, we introduce a data generation pipeline consisting of two key steps: a progressive strategy to generate structured, long-chain reasoning data with diverse reasoning paths, and a multi-granularity assessment system to evaluate and score these paths at different levels. Through automatic generation, assessment, and ranking strategies, the pipeline effectively operates without the need for human labor and makes the reasoning dataset more scalable for enhancing reasoning capabilities. To further improve MLLM reasoning beyond data scaling, we design a multi-agent system, as illustrated in Figure 1, that decomposes the problem-solving process into two distinct steps: reasoning and summarization. The reasoning agent generates a detailed reasoning process for the input query, while the summarization agent identifies key information within the reasoning process and selectively answers the question. To refine the quality of the reasoning, we employ an iterative DPO approach to enhance reasoning capabilities. The two agents collaborate to further improve the reasoning quality. Our findings demonstrate that this system significantly enhances the performance of various MLLMs across a broad range of visual reasoning benchmarks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 594, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 594, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 594, + 295, + 712 + ], + "type": "text", + "content": "We evaluate Insight-V by integrating our system into the widely used LLaVA-NeXT [22] model. To further highlight the potential of our design in advancing state-of-the-art models, we construct a robust base MLLM and apply our system to demonstrate its ability to enhance SOTA models across various visual reasoning benchmarks. Notably, Insight-V improves LLaVA-NeXT in an average performance of " + }, + { + "bbox": [ + 57, + 594, + 295, + 712 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 57, + 594, + 295, + 712 + ], + "type": "text", + "content": " across seven challenging visual reasoning benchmarks, while a " + }, + { + "bbox": [ + 57, + 594, + 295, + 712 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 57, + 594, + 295, + 712 + ], + "type": "text", + "content": " improvement is observed with the strong base MLLM, underscoring the effectiveness and generalizability" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 317, + 73, + 367, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 73, + 367, + 83 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 367, + 83 + ], + "type": "text", + "content": "of Insight-V." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 86, + 554, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 86, + 554, + 180 + ], + "spans": [ + { + "bbox": [ + 317, + 86, + 554, + 180 + ], + "type": "text", + "content": "In summary, Insight-V offers 1) a scalable data generation pipeline for long-chain, high-quality reasoning data, 2) a multi-agent system that decomposes visual reasoning tasks into reasoning and summarization, and 3) a two-stage training pipeline to enhance visual reasoning capabilities. Together, these contributions address key challenges in visual reasoning, providing a solid foundation for future research in MLLM reasoning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 193, + 399, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 193, + 399, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 193, + 399, + 205 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 213, + 467, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 213, + 467, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 467, + 226 + ], + "type": "text", + "content": "2.1. Vision-Language Reasoning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 232, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 554, + 518 + ], + "type": "text", + "content": "Recent advancements in MLLMs [1, 17, 20-23, 27, 29, 38] have equipped these models with robust capabilities across diverse domains, including visual understanding [20, 38], mathematics [19], college-level questions [6], and scientific inquiries. In visual understanding, most research [18, 22, 28, 43, 46] emphasizes fine-grained detail analysis and localization, training models to perform visual reasoning with tailored datasets to enhance interpretive capabilities. For mathematics and expert-level reasoning, existing methods [11, 56, 57] predominantly derive from Chain-of-Thought [44] approaches, training MLLMs to generate step-by-step reasoning across various subjects. However, these approaches often focus primarily on improving dataset quality through Chain-of-Thought, overlooking the importance of structured reasoning paths and extended reasoning chains in advancing model reasoning capabilities. Additionally, significant challenges arise when relying on a single model to manage the entire reasoning process for complex tasks, underscoring the need for a multi-agent system to decompose and enhance this process. In this work, we tackle these challenges by introducing a scalable reasoning data generation pipeline and implementing a multi-agent system for reasoning and summarization decomposition, enhancing the overall reasoning capabilities of existing MLLMs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 528, + 467, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 528, + 467, + 540 + ], + "spans": [ + { + "bbox": [ + 317, + 528, + 467, + 540 + ], + "type": "text", + "content": "2.2. Vision-Language Alignment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 546, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 554, + 712 + ], + "type": "text", + "content": "To align the model more closely with human preferences, several alignment techniques are employed for MLLMs. A widely used approach is Reinforcement Learning from Human Feedback [2] (RLHF), which iteratively refines the model's responses based on human feedback, enhancing both response quality and interpretability. To further improve MLLM capabilities, Direct Preference Optimization [39] (DPO) is introduced to simplify the alignment process. By directly training on human preference data, DPO optimizes the model's outputs to better match human-selected responses. However, traditional DPO is primarily focused on offline scenarios, and as the model evolves, the effectiveness of this approach may significantly diminish. To address this, Iterative DPO [5] has been proposed, which optimizes pref" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "9063" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "erence pairs through DPO at each iteration. It then generates new preference pairs for the next iteration using the updated model and evaluates them with a reward model. In this paper, we use iterative DPO to achieve stronger alignment and enhance the model's reasoning capabilities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 155, + 138, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 138, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 138, + 168 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 178, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 296, + 347 + ], + "type": "text", + "content": "In this section, we provide a comprehensive description of the proposed Insight-V system, detailing its architecture and key contributions. Section 3.1 presents an overview of Insight-V, highlighting the core concepts of our approach. The design of Insight-V is structured around three primary components: 1) a carefully constructed pipeline for structured, scalable reasoning data generation, as described in Section 3.2; 2) a multi-agent MLLM system that supports complex visual reasoning, as detailed in Sections 3.3; and 3) a streamlined yet effective training pipeline to enhance overall performance, as outlined in Section 3.4. Together, these components form a cohesive system that effectively tackles the challenges of performing detailed, long-chain reasoning while preserving visual perception capabilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 123, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 123, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 123, + 376 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 387, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 554 + ], + "type": "text", + "content": "Advancing reasoning capabilities of LLMs has been a focal point of extensive research. Despite these efforts, the reasoning potential within multi-modal LLMs has barely been explored. Most approaches aim to strengthen reasoning at the inference stage, assuming the model has already acquired robust reasoning skills. Other approaches optimize model parameters using chain-of-thought data, enabling models to mimic human reasoning processes. However, these methods present significant challenges for current general-purpose MLLMs, as they require the model to develop reasoning skills while retaining prior capabilities, which often results in only modest performance gains. Additionally, the lack of structured, high-quality training data impedes training models with advanced reasoning capabilities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "To fully leverage the reasoning capabilities of MLLMs, we propose Insight-V, a novel system comprising two MLLMs dedicated to reasoning and summarization, respectively. The reasoning model is tasked with generating a detailed reasoning process to assist in problem-solving, while the summary model evaluates this reasoning as supplementary information to assess its relevance and utility for answering the question. We also construct a structured, high-quality dataset to train both agents. We posit that this multi-agent system can enhance the reasoning strengths of MLLMs by decomposing the problem-solving process into distinct reasoning and summarization phases, thereby driving substantial performance improvements." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 563, + 213 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 563, + 213 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 563, + 213 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 563, + 213 + ], + "type": "image", + "image_path": "952a18fd2a58e66250ef7efc80fbbba15454c2b5615fa32def0110ed62c8a683.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 223, + 555, + 267 + ], + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 267 + ], + "type": "text", + "content": "Figure 2. Data Generation Pipeline of Insight-V. The reasoning processes are generated progressively through a reasoning generator, and then fed into a multi-granularity assessment system to ensure high-quality reasoning." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 277, + 541, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 541, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 541, + 290 + ], + "type": "text", + "content": "3.2. Construction of Structured Reasoning Data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 294, + 556, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 556, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 556, + 415 + ], + "type": "text", + "content": "Previous studies [56, 58] have explored the integration of reasoning capabilities into MLLMs. However, training MLLMs to develop robust reasoning skills remains a considerable challenge, particularly due to data limitations. To address this, we introduce our data generation pipeline in this section, designed to produce high-quality, long-chain reasoning data using a progressive generation process and multi-granularity assessment. As shown in Figure 2, this scalable approach enables us to generate high-quality data to enhance the model's reasoning capabilities effectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 429, + 549, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 549, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 549, + 441 + ], + "type": "text", + "content": "Progressive Long-Chain Reasoning Data Generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 441, + 556, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 556, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 556, + 584 + ], + "type": "text", + "content": "For each input query, we first employ a reasoning generator to produce a structured reasoning process in JSON format to address the problem. At each step, the reasoning generator provides a brief summary of the current step, a detailed reasoning response, and an action for the following step. If the action is continue, the model proceeds with an additional reasoning step in the next iteration; if the action is summary, the model generates a final summary and answer based on the complete reasoning process in the subsequent iteration. Specifically, For a multi-modal model " + }, + { + "bbox": [ + 313, + 441, + 556, + 584 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 441, + 556, + 584 + ], + "type": "text", + "content": ", input image " + }, + { + "bbox": [ + 313, + 441, + 556, + 584 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 441, + 556, + 584 + ], + "type": "text", + "content": ", and question " + }, + { + "bbox": [ + 313, + 441, + 556, + 584 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 441, + 556, + 584 + ], + "type": "text", + "content": ", the data generation process for each step is represented as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 365, + 593, + 502, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 593, + 502, + 605 + ], + "spans": [ + { + "bbox": [ + 365, + 593, + 502, + 605 + ], + "type": "interline_equation", + "content": "R _ {t} = M (I, Q, \\left[ R _ {1} \\dots R _ {t - 1} \\right], A),", + "image_path": "5812b4d094b279dcf8a31f9b8624ad5a70e17fdf56359e15bb387f757d5b1a45.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 367, + 609, + 492, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 609, + 492, + 621 + ], + "spans": [ + { + "bbox": [ + 367, + 609, + 492, + 621 + ], + "type": "interline_equation", + "content": "R _ {a n s} = M (I, Q, \\left[ R _ {1} \\dots R _ {n} \\right]),", + "image_path": "c958cb4d9c8f513806cef4e224dcfad8cfc06350f8feb5a57e77059215b96951.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "R_{t}" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "R_{ans}" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": " denote the response at the " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "-th step and the final answer, respectively, " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": " represents the reasoning generated by the model at the " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "-th step, " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": " represents the total reasoning steps, and " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": " is the action determined in the previous step. By repeating this process " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": " times, we can iteratively sample " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": " structured responses for each query. The generation parameters are adjusted to encourage the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9064" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "model to produce outputs with various information and steps, allowing us to identify the most effective reasoning chain for each question." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 124, + 296, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 124, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 124, + 296, + 363 + ], + "type": "text", + "content": "Multi-Granularity Assessment. After obtaining the structured responses, we utilize an assessment pipeline to ensure data quality. Specifically, we first apply a strong LLM, such as Qwen2 [37], for direct answer filtering. As shown in Figure 2, the model is provided with the generated final answer and the ground truth answer and is tasked with determining whether the generated answer is correct, serving as an approximate indicator of the validity of the associated reasoning chain. Once responses with incorrect answers are filtered out, the remaining reasoning processes are passed to a reasoning path scoring agent. Here, an advanced multimodal model, such as Qwen2-VL [38], is supplied with the image, question, reasoning path, and ground truth answer, and is prompted to evaluate the reasoning path. The scoring agent assesses each response based on the step-by-step accuracy of the reasoning path and the level of detail in the reasoning. To ensure consistency in scoring across different data samples, we aggregate all responses for each question and process them in a single pass. The model then generates scores for each response, ranging from 1 to 100." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 364, + 296, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 296, + 400 + ], + "type": "text", + "content": "Through the above two steps, we construct a structured, high-quality dataset that provides detailed reasoning for each question, effectively supporting the training of our models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 408, + 142, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 142, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 142, + 421 + ], + "type": "text", + "content": "3.3. Model Design" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "text", + "content": "After constructing the dataset, we develop a multi-agent framework to enhance overall reasoning capabilities through collaborative agent interaction. Specifically, we first train a reasoning agent to generate a detailed reasoning process for each problem. Then, a summary agent is employed to answer the question, selectively utilizing the reasoning process based on its assessment. Together, these two agents collaborate to improve reasoning performance effectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 528, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 296, + 684 + ], + "type": "text", + "content": "Reasoning Agent. Previous approaches typically combine reasoning and question-answering within a single process, which poses challenges for MLLMs. Generating a long-chain reasoning process can introduce errors, and directly answering questions based on flawed reasoning often leads to poorer results. To address this, we propose a specialized reasoning agent designed to generate a detailed, step-by-step reasoning process in response to an input query. We construct the reasoning dataset by selecting the highest-scoring reasoning path for each question. After training on this dataset, the model transforms into a reasoning agent with enhanced reasoning capabilities, enabling it to generate more detailed, structured reasoning processes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Summary Agent. Summarization plays a critical role in enabling models to accurately answer questions. After" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 563, + 215 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 563, + 215 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 563, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 563, + 215 + ], + "type": "image", + "image_path": "fb79206494bf4a64124f608becbcf2d97ebcac1cc39902ff3abf9ae690cb7102.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "type": "text", + "content": "Figure 3. Overview of Insight-V Model Design. We derive a multi-agent system from a single model. By decomposing the task into reasoning and summarization, the two agents collaborate to enhance the overall reasoning capability." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 280, + 556, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 556, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 556, + 388 + ], + "type": "text", + "content": "generating multi-step reasoning, summarization provides a cohesive understanding of the reasoning process, ultimately guiding the model to the final answer. However, since the response generated by the reasoning agent may contain errors, we develop a summarization model robust to inaccuracies in the reasoning path, selectively incorporating or disregarding elements as needed. This approach maximizes the reasoning model's effectiveness while minimizing the risk of introducing misleading information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 388, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 616 + ], + "type": "text", + "content": "To enhance the robustness of the summary agent, we carefully curate its training dataset. We utilize the collected dataset, which is comprised of two types: data with optimal reasoning processes and data with flawed reasoning processes for the summarization task. This method prevents the model from simply copying reasoning outcomes and encourages critical evaluation of reasoning quality. To further promote critical analysis by the summary agent, we select flawed reasoning samples based on their performance scores. Specifically, we draw flawed responses from varying score ranges to create a dataset with different levels of error, prompting the model to assess reasoning processes at various granularities. To better align the summary model with the reasoning agent, we also incorporate question-reasoning pairs generated by the reasoning agent to enhance collaboration between the two agents. Additionally, to preserve the original multi-modal capabilities, we supplement the dataset with standard question-answering data to sustain the summary agent's performance in direct question-answering." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 624, + 419, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 419, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 419, + 636 + ], + "type": "text", + "content": "3.4. Training Pipeline" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "The training pipeline for Insight-V is designed to be straightforward and efficient, utilizing a two-stage strategy. For both the reasoning agent and the summary agent, we begin with a well-trained MLLM. In the first stage, we apply supervised fine-tuning to train the agents to fulfill their designated roles using corresponding datasets. In the second stage, we" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9065" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "implement direct preference optimization, following prior research [45, 57]. This optimization is applied to the reasoning model, aligning it with human reasoning processes in a simple yet effective manner. These two stages enable the development of a robust system with enhanced visual reasoning capabilities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 148, + 284, + 160 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 284, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 284, + 160 + ], + "type": "text", + "content": "3.4.1. Supervised Fine-tuning for Multi-agent System" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 297, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 297, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 297, + 355 + ], + "type": "text", + "content": "To perform supervised fine-tuning and obtain the two agents, we first train a base multi-modal model, following established methodologies. This model can address general visual question-answering tasks and gain foundational vision-language skills. We compile a high-quality image-text dataset focused on knowledge learning to train the base model. This data is sourced from various open-source academic datasets, including LLaVA-NeXT [22], Cauldron [15], and Cambrian-1 [43]. Once the base model is trained, we further fine-tune the two agents, initializing them from the base model. For the reasoning agent, we utilize the curated reasoning dataset to develop step-by-step reasoning capabilities. For the summary agent, we formulate a dataset as outlined in Section 3.3 and sample about one million general image-text pairs from the dataset used for the base model, preserving its original visual perception abilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 211, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 211, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 211, + 370 + ], + "type": "text", + "content": "3.4.2. Enhanced Reasoning with RL" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "text", + "content": "Preference learning has gained increasing focus in the field of large language models. The primary aim is to fine-tune model outputs to align better with human (or expert) preferences, creating outputs more suited to real-world applications. Let's assume a preference dataset defined as " + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x^{(i)},y_w^{(i)},y_l^{(i)})\\}_{i = 1,\\dots ,|\\mathcal{D}|}" + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "inline_equation", + "content": "x^{(i)}" + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "text", + "content": " is prompt, and " + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "inline_equation", + "content": "y_{w}^{(i)}" + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "inline_equation", + "content": "y_{l}^{(i)}" + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "text", + "content": " represent the preferred and less preferred responses, respectively. We denote " + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "inline_equation", + "content": "y_{w}\\succ y_{l}\\mid x" + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "text", + "content": " to signify that " + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "inline_equation", + "content": "y_{w}" + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "text", + "content": " is preferred over " + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "inline_equation", + "content": "y_{l}" + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "text", + "content": " for prompt " + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 373, + 296, + 486 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "text", + "content": "Since the true distribution of human preferences cannot be directly observed, we approximate it with a latent reward model " + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "inline_equation", + "content": "r^{*}(x,y)" + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "text", + "content": ", assuming that higher rewards correspond to stronger preferences. Following the approach by Rafailov et al. [39], we can model the human preference distribution " + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "inline_equation", + "content": "p^{*}" + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "text", + "content": " using the Bradley-Terry (BT) model [3]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 562, + 282, + 603 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 562, + 282, + 603 + ], + "spans": [ + { + "bbox": [ + 69, + 562, + 282, + 603 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p ^ {*} (y _ {1} \\succ y _ {2} \\mid x) = \\frac {\\exp (r ^ {*} (x , y _ {1}))}{\\exp (r ^ {*} (x , y _ {1})) + \\exp (r ^ {*} (x , y _ {2}))} \\\\ = \\sigma \\left(r ^ {*} (x, y _ {1}) - r ^ {*} (x, y _ {2})\\right), \\\\ \\end{array}", + "image_path": "efd822352de70f7bd4326567acb4d0455e78e96fc73f63092e0962377be07abb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 607, + 208, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 607, + 208, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 607, + 208, + 618 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 607, + 208, + 618 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 607, + 208, + 618 + ], + "type": "text", + "content": " denotes the logistic function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 619, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 296, + 654 + ], + "type": "text", + "content": "To estimate the parameters of the reward model, we can apply maximum likelihood estimation by minimizing the negative log-likelihood:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 659, + 295, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 295, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R} \\left(r _ {\\phi}, \\mathcal {D}\\right) = - \\mathbb {E} _ {\\left(x, y _ {w}, y _ {l}\\right) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(r _ {\\phi} \\left(x, y _ {w}\\right) - r _ {\\phi} \\left(x, y _ {l}\\right)\\right) \\right],", + "image_path": "4e31d9ab02edebabdb3d180e46b590ed234f141ae8669cf3bcafe919ed80ae36.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "r_{\\phi}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " is a parameterized reward model. This approach allows us to approximate the preference distribution and fine-tune the model to capture human-like preferences effectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "content": "The traditional DPO algorithm operates in an offline setting. During DPO training, as model parameters continuously change, the preference dataset generated offline can gradually diverge from the model's current distribution, which weakens the effectiveness of the DPO algorithm. To address this issue, we employ an iterative DPO algorithm. By conducting multiple rounds of DPO training and sampling, this approach enables the model to better approximate an online setting during training, thus further enhancing its performance. Specifically, our approach involves training a sequence of models " + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "inline_equation", + "content": "M_{1},\\ldots ,M_{T}" + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "content": ", where each subsequent model " + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "inline_equation", + "content": "M_{t + 1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "content": " utilizes preference data " + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "content": " generated by the " + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "content": "-th model. We apply this complete training process to the fine-tuned reasoning agent, enabling the model to better align with human preferences and produce structured, detailed reasoning steps for complex questions, which supports the summary agent more effectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 286, + 395, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 395, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 395, + 300 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "text", + "content": "We conduct extensive experiments across multiple vision-language benchmarks to validate the effectiveness of our method. In this section, we first introduce the implementation details of Insight-V in Section 4.1. Then we present a comparison with state-of-the-art MLLMs, outlining the primary results of our method on visual reasoning tasks as well as additional results on general image understanding in Section 4.2. Moreover, we offer further analytical experiments and essential ablation studies on design choices in Section 4.3, along with qualitative results shown in Section 4.4 for more insights." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 445, + 447, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 447, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 447, + 457 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 462, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 641 + ], + "type": "text", + "content": "We integrate the Insight-V system with various MLLMs to demonstrate the broad applicability of our approach. Our initial implementation with Insight-V on LLaVA-NeXT-LLaMA3 [22] illustrates the method's effectiveness. To further validate its generalizability and establish a solid baseline against state-of-the-art MLLMs, we additionally train a base multi-modal model using the Qwen-2.5-7B [41] LLM. During pretraining, we utilize the 558K captioning dataset from LLaVA-1.5 [21], unfreezing the connector parameters. This is followed by supervised fine-tuning with a curated dataset of approximately 4 million images, using a learning rate of 2e-5 as guided by prior researches [21, 22]. This two-stage training process equips the baseline model with essential visual perception abilities, achieving competitive results on vision-language benchmarks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "We then initialize two agents from the baseline model, performing targeted fine-tuning to obtain the final agents. For the reasoning agent, we compile a dataset of 200K images and train the model over 2 epochs with a learning rate of 5e-6. For the summary agent, we use a dataset of 1.2 million images, applying a learning rate of 1e-5 and train" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9066" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 113, + 555, + 303 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Results on Visual Reasoning Tasks. We conduct evaluation experiments across 7 benchmarks, covering both general reasoning and task-specific reasoning assessments. Insight-V exhibits notable effectiveness and generalizability when applied to LLaVA-NeXT and our baseline model, surpassing other state-of-the-art MLLMs by a large margin." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 113, + 555, + 303 + ], + "lines": [ + { + "bbox": [ + 56, + 113, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 113, + 555, + 303 + ], + "type": "table", + "html": "
ModelSizeMMMUMMMU-ProMMBenchMMEChartQAMMStarMathVistaAverage
DeepSeek-VL [29]7B35.4-73.5-/-59.137.136.1-
VILA-1.5 [20]8B38.6-75.31634.9/--39.7--
Cambrian-1 [43]8B42.7-75.91547.1/-73.3-49.0-
InternLM-XComposer2 [7]7B41.1-77.62220.471.856.259.5-
POINTS [26]7B51.4-78.02184.1-60.963.0-
IXC-2.5 [55]7B42.9-79.42233.182.259.963.7-
Bunny-LLaMA3 [12]8B43.4-77.21588.9/321.1--34.4-
MM-1.5 [54]7B41.8--1514.9/346.478.6-47.6-
MiniCPM-LLaMA3-V 2.5 [49]8B45.819.677.22024.6-51.854.3-
MiniCPM-V-2.6 [50]7B49.827.278.02268.7-57.560.6-
Qwen2-VL [38]7B53.7-81.0-83.060.761.4-
Idefics3-LLaMA3 [14]8B46.622.977.51937.474.855.958.448.1
Ovis1.5-LLaMA3 [31]8B48.323.676.61948.576.457.363.049.4
LLaVA-NeXT-LLaMA3 [22]8B36.913.272.31611.1/346.069.443.145.940.2
+ Multi-Agent8B40.817.877.61603.7/469.374.652.647.444.5
+ Iterative DPO (Insight-V-LLaVA)8B42.021.081.71583.9/485.477.457.449.847.2 (+7.0)
Our Base Model7B47.122.681.31573.7/482.575.757.056.948.7
+ Multi-Agent7B49.723.882.21662.2/629.381.258.658.750.7
+ Iterative DPO (Insight-V)7B50.224.982.31685.1/627.081.561.559.951.6 (+2.9)
", + "image_path": "8e7a4db7ddf76a1fa5d9a04b1cf11bf5f7830f5875941835485d58bc2c3ade2d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 365, + 296, + 435 + ], + "blocks": [ + { + "bbox": [ + 55, + 312, + 297, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 312, + 297, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 297, + 357 + ], + "type": "text", + "content": "Table 2. Results on other multimodal benchmarks. Insight-V enhances reasoning capabilities without compromising general visual perception and even achieves improvements on benchmarks requiring perception ability more." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 365, + 296, + 435 + ], + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 435 + ], + "type": "table", + "html": "
ModelTextVQADocVQAOCRbenchAI2D
LLaVA-NeXT-LLaMA365.278.255371.5
+ Multi-Agent68.981.863175.7
+ Iterative DPO (Insight-V-LLaVA)70.582.966377.3
Our Base Model75.490.271379.7
+ Multi-Agent77.091.473880.1
+ Iterative DPO (Insight-V)76.891.573579.8
", + "image_path": "df280f877b6050045bac5914d8c379a17cf479b1e3a2ef99896b466af22f9726.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 446, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 446, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 446, + 295, + 555 + ], + "type": "text", + "content": "ing for 1 epoch. Additionally, we apply Direct Preference Optimization (DPO) to the reasoning agent, using approximately 15K preference data and training for 1 epoch at a learning rate of 5e-7. This DPO process is iteratively conducted across 3 rounds by using the model from the previous stage to generate preference data, thus enhancing the agent's reasoning capabilities. The lmms-eval [16] is utilized for fast evaluation. Further training details are provided in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 563, + 239, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 239, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 239, + 577 + ], + "type": "text", + "content": "4.2. Main Results on Visual Reasoning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 297, + 715 + ], + "type": "text", + "content": "Setup. We present comprehensive experimental results on various visual reasoning benchmarks, which require the model to handle complex and challenging questions, demonstrating strong reasoning capabilities. We select several well-known and representative benchmarks, covering comprehensive evaluation, chart understanding, and mathematical problem-solving. MMMU [52] and MMMU-Pro [53] evaluate the model's expert-level perception and reasoning abilities across various topics. MMBench [24] is a widely used benchmark for comprehensive MLLM evaluation on images. MME [10] includes 14 challenging subtasks for assessing" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 354, + 555, + 417 + ], + "blocks": [ + { + "bbox": [ + 313, + 312, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 346 + ], + "type": "text", + "content": "Table 3. Ablations on the Insight-V Design Choice. The multiagent design outperforms other configurations, highlighting the critical role of reasoning and summarization decomposition." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 354, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 417 + ], + "type": "table", + "html": "
ModelMMMUChartQAMathVistaMMStarAvg
Baseline47.175.756.957.059.2
Vanilla - Direct SFT47.079.257.658.460.6
Multi-Turn Supervised48.179.657.958.261.0
Summary Agent Only47.576.357.357.959.8
Multi-Agent49.781.258.758.662.1
", + "image_path": "32ece75a2acbb8abd5474ad89b6d01f2c9d93f626938acada62cd447252ee845.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 429, + 555, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 490 + ], + "type": "text", + "content": "visual perception and cognition. ChartQA [33] focuses on logical reasoning with charts, while MathVista [30] assesses problem-solving skills in math, using GPT-4-0613 [35] as the evaluator, following standard practice. MMStar [4] tests a range of tasks with varying difficulty levels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "Main Results. The experimental results in Table 1 confirm the effectiveness and generalizability of Insight-V on visual reasoning tasks. Applied to LLaVA-NeXT and our baseline model, the Insight-V system substantially boosts performance across challenging visual reasoning benchmarks, achieving competitive outcomes against state-of-the-art methods. We report an average improvement across all benchmarks: Insight-V yields an average increase of " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": " when applied to LLaVA-NeXT and the base model, underscoring the system's effectiveness. Following Direct Preference Optimization, an additional improvement of " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": " was observed, illustrating the value of reinforcement learning algorithms in enhancing reasoning capabilities. Specifically, on general reasoning benchmarks, MMMU, MMMU-Pro, and MMBench, we noted gains of " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": ", respectively. The highest improvement was on MME, with a " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "9.1\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": " increase, highlighting Insight-V's strength in tasks requiring both perception and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9067" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 79, + 272, + 200 + ], + "blocks": [ + { + "bbox": [ + 67, + 79, + 272, + 200 + ], + "lines": [ + { + "bbox": [ + 67, + 79, + 272, + 200 + ], + "spans": [ + { + "bbox": [ + 67, + 79, + 272, + 200 + ], + "type": "image", + "image_path": "a244efeef3264345029049f015ebfe359b3db79d1890b57e3ed256e25ecda321.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 208, + 297, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 297, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 297, + 243 + ], + "type": "text", + "content": "Figure 4. Ablations on the amount of training data. The reasoning agent benefits from data scaling, providing more valuable insights for the summary agent." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "text", + "content": "reasoning. On ChartQA, which entails complex chart-based reasoning, we achieved a " + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "text", + "content": " increase. Performance gains on MMStar and MathVista, with improvements of " + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "text", + "content": ", further validate our approach's versatility in tackling math problems that demand sophisticated reasoning and numerical calculation skills. This strong performance across all benchmarks underscores the efficacy of Insight-V and its potential to advance visual reasoning capabilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 452 + ], + "type": "text", + "content": "Results on Other Multi-modal Benchmarks. To verify that Insight-V enhances reasoning capabilities without compromising basic visual perception skills, we conduct experiments on benchmarks requiring only fundamental image understanding. We selected TextVQ [40], DocVQA [34], OCRBench [25], and AI2D [13] for this purpose, as these benchmarks focus on basic visual interpretation rather than complex reasoning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 296, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 596 + ], + "type": "text", + "content": "As shown in Table 2, integrating Insight-V enables models to maintain strong performance across general visual perception benchmarks. Additionally, applying Insight-V yields performance improvements on both LLaVA-NeXT and our base model, as it enhances general visual understanding by guiding the model's attention to relevant regions, thereby improving visual perception accuracy. Following DPO, the model demonstrates advanced reasoning capabilities while achieving competitive results on standard image understanding benchmarks compared to baseline models, underscoring Insight-V's robustness. Consistent results of both models further confirm Insight-V's generalizability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 157, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 157, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 157, + 617 + ], + "type": "text", + "content": "4.3. Further Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 623, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 296, + 682 + ], + "type": "text", + "content": "In this section, we present comprehensive experiments to validate the design choices of Insight-V, emphasizing our approach's key contributions. Additionally, we include a case study to further demonstrate the qualitative effectiveness of Insight-V." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Effectiveness of Multi-agent System. To evaluate the effectiveness of the multi-agent system, we comprehensively" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 113, + 555, + 166 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "type": "text", + "content": "Table 4. Ablations on the DPO training strategy. Iterative DPO progressively enhances the model's reasoning capabilities, leading to improved performance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 113, + 555, + 166 + ], + "lines": [ + { + "bbox": [ + 317, + 113, + 555, + 166 + ], + "spans": [ + { + "bbox": [ + 317, + 113, + 555, + 166 + ], + "type": "table", + "html": "
ModelMMMUChartQAMathVistaMMStarAvg
Insight-V (Multi-Agent)49.781.258.758.662.1
+ RLAIF49.581.459.159.262.3
+ DPO50.880.859.359.962.7
+ Iterative DPO50.281.559.961.563.3
", + "image_path": "7b1304471774191c54149c5786110c6360ec529e7c0d7f62cb97d00f23fd80e5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 176, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 555, + 319 + ], + "type": "text", + "content": "compare Insight-V with alternative design choices. We begin by reporting the performance of the summary agent without a reasoning process to underscore the importance of the reasoning agent. For a comprehensive comparison, we also restructure the collected data using a Chain-of-Thought (CoT) template and train a model to perform sequential reasoning and subsequently answer questions. Additionally, we enable a model that can handle multi-turn conversations, facilitating reasoning and summarization in a multi-turn format. We use MMMU, ChartQA, MathVista, and MMStar as representative benchmarks to assess the performance across different methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 320, + 556, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 556, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 556, + 488 + ], + "type": "text", + "content": "As shown in Table 3, the results indicate that the multi-agent system plays a critical role in enhancing the system's visual understanding capabilities. Using only the summary agent without a reasoning process results in limited improvements in reasoning tasks, as the model lacks the necessary reasoning framework for optimal performance. Training the model to perform chain-of-thoughts reasoning, which is denoted as Insight-V (Vanilla-Direct SFT), yields modest gains, as this approach does not sufficiently emphasize critical judgments within the reasoning process, merging reasoning and summarization into a single task. Furthermore, models trained with multi-turn conversations still produce sub-optimal results, underscoring the importance of a multi-agent system that separates reasoning and summarization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 493, + 556, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 556, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 556, + 672 + ], + "type": "text", + "content": "Data Scaling Law of Reasoning Agent. To assess the effectiveness of the reasoning agent, we perform ablation experiments on the data volume for training. As shown in Figure 4, we compare reasoning agents trained on varying data sizes from 50K to 200K. For a fair and comprehensive comparison, we employ the same summary agent across evaluations and report results on six benchmarks. The findings clearly indicate that the reasoning agent benefits from increased data. With limited data, the reasoning agent struggles to generalize and fails to provide useful input to the summary model, resulting in performance even worse than baseline models. In contrast, training in larger datasets improves the capabilities of the reasoning agent, allowing it to perform step-by-step reasoning and provide valuable insights that help the summary agent solve tasks effectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Effects of RL Algorithms. We assess DPO algorithms to identify optimal strategies for enhancing the reasoning process. To examine the effects of dataset composition on DPO" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9068" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 91, + 132, + 139 + ], + "blocks": [ + { + "bbox": [ + 67, + 91, + 132, + 139 + ], + "lines": [ + { + "bbox": [ + 67, + 91, + 132, + 139 + ], + "spans": [ + { + "bbox": [ + 67, + 91, + 132, + 139 + ], + "type": "image", + "image_path": "3c7a25ec1fd61936835c8b31aeb85628a7776f1ee1f0180028a4b27df3229af3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 83, + 272, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 83, + 272, + 96 + ], + "spans": [ + { + "bbox": [ + 138, + 83, + 272, + 96 + ], + "type": "text", + "content": "Referring to Table 3-1, which is INCORRECT. If pots sell for $20 each then:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 99, + 279, + 150 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 99, + 271, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 99, + 271, + 106 + ], + "spans": [ + { + "bbox": [ + 138, + 99, + 271, + 106 + ], + "type": "text", + "content": "(A) The marginal product of the third worker is five pots." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 106, + 279, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 106, + 279, + 124 + ], + "spans": [ + { + "bbox": [ + 138, + 106, + 279, + 124 + ], + "type": "text", + "content": "(B) the marginal revenue product of labor equals the marginal product of labor multiplied by the additional revenue." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 124, + 279, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 124, + 279, + 137 + ], + "spans": [ + { + "bbox": [ + 138, + 124, + 279, + 137 + ], + "type": "text", + "content": " (C) the marginal revenue product of labor of the second worker is $260." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 137, + 279, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 279, + 150 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 279, + 150 + ], + "type": "text", + "content": "(D) the marginal revenue from selling the eighteenth pot is $20." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 454, + 75, + 525, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 454, + 75, + 525, + 85 + ], + "spans": [ + { + "bbox": [ + 454, + 75, + 525, + 85 + ], + "type": "text", + "content": "Chain-of-Thought" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 291, + 88, + 529, + 150 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 291, + 88, + 522, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 88, + 522, + 110 + ], + "spans": [ + { + "bbox": [ + 291, + 88, + 522, + 110 + ], + "type": "text", + "content": "A. ... It is incorrect because the marginal product of the third worker is 5 pots (18 - 13 = 5). However, the question incorrectly states that ... of the third worker is five pots, which should be ... of the second worker." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 292, + 110, + 529, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 110, + 529, + 123 + ], + "spans": [ + { + "bbox": [ + 292, + 110, + 529, + 123 + ], + "type": "text", + "content": "B. ... This statement is correct. It is calculated by multiplying the marginal product of labor by the price per unit. In this case, ..." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 292, + 123, + 525, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 123, + 525, + 137 + ], + "spans": [ + { + "bbox": [ + 292, + 123, + 525, + 137 + ], + "type": "text", + "content": " C......This statement is incorrect. The marginal product of the second worker is 5 pots, and the price per pot is $20, so ... is 5 * $20 = $100." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 292, + 137, + 516, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 137, + 516, + 143 + ], + "spans": [ + { + "bbox": [ + 292, + 137, + 516, + 143 + ], + "type": "text", + "content": "D. ... This statement is correct. The ... is equal to the price per unit, which in this case is $20." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 292, + 143, + 342, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 143, + 342, + 150 + ], + "spans": [ + { + "bbox": [ + 292, + 143, + 342, + 150 + ], + "type": "text", + "content": "So the answer is A." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 80, + 156, + 153, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 156, + 153, + 167 + ], + "spans": [ + { + "bbox": [ + 80, + 156, + 153, + 167 + ], + "type": "text", + "content": "Insight-V (Vanilla)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 441, + 156, + 533, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 441, + 156, + 533, + 166 + ], + "spans": [ + { + "bbox": [ + 441, + 156, + 533, + 166 + ], + "type": "text", + "content": "Insight-V (Multi-Agent)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 84, + 175, + 127, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 175, + 127, + 183 + ], + "spans": [ + { + "bbox": [ + 84, + 175, + 127, + 183 + ], + "type": "text", + "content": "DIRECT SFT" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 83, + 188, + 97, + 201 + ], + "blocks": [ + { + "bbox": [ + 83, + 188, + 97, + 201 + ], + "lines": [ + { + "bbox": [ + 83, + 188, + 97, + 201 + ], + "spans": [ + { + "bbox": [ + 83, + 188, + 97, + 201 + ], + "type": "image", + "image_path": "cfcce8980bd3b9fc188f00aa812a4d81cfb9c0eb96fdfb368a54bfe6f451fb24.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 187, + 245, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 187, + 245, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 245, + 201 + ], + "type": "text", + "content": "Identifying Key Information: The image presents a table showing ... The question asks about ..." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 83, + 212, + 97, + 224 + ], + "blocks": [ + { + "bbox": [ + 83, + 212, + 97, + 224 + ], + "lines": [ + { + "bbox": [ + 83, + 212, + 97, + 224 + ], + "spans": [ + { + "bbox": [ + 83, + 212, + 97, + 224 + ], + "type": "image", + "image_path": "778b689d6fe9a1c351ac842731aded0a7e487ba15f80770e6cfa87242cd79189.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 204, + 245, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 204, + 245, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 245, + 232 + ], + "type": "text", + "content": "Calculating Marginal Product of Labor: To find the marginal product of labor, ... For example, ... subtracting the total pots produced with two workers from that with three workers." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 83, + 238, + 97, + 251 + ], + "blocks": [ + { + "bbox": [ + 83, + 238, + 97, + 251 + ], + "lines": [ + { + "bbox": [ + 83, + 238, + 97, + 251 + ], + "spans": [ + { + "bbox": [ + 83, + 238, + 97, + 251 + ], + "type": "image", + "image_path": "ee4fd4c6ffcaf143575a303ab2ee389b42bed6d4f3abaec1678d21f5a47019c3.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 236, + 245, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 236, + 245, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 245, + 256 + ], + "type": "text", + "content": "Determining Marginal Revenue Product: Next, we need to determine ... We multiply the marginal product of each worker by \\(20 to \\dots" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 83, + 270, + 97, + 282 + ], + "blocks": [ + { + "bbox": [ + 83, + 270, + 97, + 282 + ], + "lines": [ + { + "bbox": [ + 83, + 270, + 97, + 282 + ], + "spans": [ + { + "bbox": [ + 83, + 270, + 97, + 282 + ], + "type": "image", + "image_path": "307d65e41ad124e925b0ede0019b72ec5c574315f66f158b4e0bd54cf40d64ba.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 260, + 244, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 260, + 244, + 296 + ], + "spans": [ + { + "bbox": [ + 104, + 260, + 244, + 296 + ], + "type": "text", + "content": "Comparing Options: Now that ... Option (A) states ..., which does not matches... . Option (B) suggests that ..., but this does not match ... . Option (C) indicates that ..., which also does not match ... . Option (D) claims that ..., which aligns with..." + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 83, + 301, + 97, + 315 + ], + "blocks": [ + { + "bbox": [ + 83, + 301, + 97, + 315 + ], + "lines": [ + { + "bbox": [ + 83, + 301, + 97, + 315 + ], + "spans": [ + { + "bbox": [ + 83, + 301, + 97, + 315 + ], + "type": "image", + "image_path": "1f970d9032c04744ec79cedb7ee0cbfe3121a27642079c5318b859502a6af015.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 328, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 328, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 555, + 361 + ], + "type": "text", + "content": "Figure 5. Qualitative Results of Insight-V. We present qualitative comparisons of Insight-V with Chain-of-Thought and Insight-V with direct SFT (Vanilla). For the Insight-V system, the reasoning agent delivers a more coherent and structured reasoning process, guiding the summary agent to the correct answer, whereas other methods struggle with complex reasoning and fail to solve such challenging problems." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 300, + 242, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 300, + 242, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 242, + 316 + ], + "type": "text", + "content": "FINAL ANSWER: (D) the marginal revenue from selling the eighteenth pot is $20." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 272, + 174, + 343, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 174, + 343, + 182 + ], + "spans": [ + { + "bbox": [ + 272, + 174, + 343, + 182 + ], + "type": "text", + "content": "REASONING AGENT" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 272, + 188, + 285, + 201 + ], + "blocks": [ + { + "bbox": [ + 272, + 188, + 285, + 201 + ], + "lines": [ + { + "bbox": [ + 272, + 188, + 285, + 201 + ], + "spans": [ + { + "bbox": [ + 272, + 188, + 285, + 201 + ], + "type": "image", + "image_path": "29d50873d1110d083270fec76951095889a5e45aff2cdcc6e0dbed0bb578fcb4.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 292, + 187, + 528, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 187, + 528, + 201 + ], + "spans": [ + { + "bbox": [ + 292, + 187, + 528, + 201 + ], + "type": "text", + "content": "Identifying Key Information: To begin solving this problem, we need to ... This involves understanding the table provided, which ..." + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 272, + 205, + 285, + 217 + ], + "blocks": [ + { + "bbox": [ + 272, + 205, + 285, + 217 + ], + "lines": [ + { + "bbox": [ + 272, + 205, + 285, + 217 + ], + "spans": [ + { + "bbox": [ + 272, + 205, + 285, + 217 + ], + "type": "image", + "image_path": "afbc7a818f7a055823802f6eb3d8d40929037def6e8959b6eacdfe9420e9de78.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 291, + 205, + 526, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 205, + 526, + 220 + ], + "spans": [ + { + "bbox": [ + 291, + 205, + 526, + 220 + ], + "type": "text", + "content": "Analyzing Table Data: Next, we should analyze ... in Table 3-1 to ... . Specifically, we need to calculate... to determine which option is incorrect." + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 272, + 223, + 285, + 236 + ], + "blocks": [ + { + "bbox": [ + 272, + 223, + 285, + 236 + ], + "lines": [ + { + "bbox": [ + 272, + 223, + 285, + 236 + ], + "spans": [ + { + "bbox": [ + 272, + 223, + 285, + 236 + ], + "type": "image", + "image_path": "767b7adf363d8251449e2d718ed93946cd5cf7001697b766315abfa4d77c2e7f.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 291, + 222, + 526, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 222, + 526, + 237 + ], + "spans": [ + { + "bbox": [ + 291, + 222, + 526, + 237 + ], + "type": "text", + "content": "Comparing Marginal Products: Option (A) suggests .... Based on Table 3-1, we find ... (from 13 to 18), matching option (A)." + } + ] + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 272, + 244, + 285, + 257 + ], + "blocks": [ + { + "bbox": [ + 272, + 244, + 285, + 257 + ], + "lines": [ + { + "bbox": [ + 272, + 244, + 285, + 257 + ], + "spans": [ + { + "bbox": [ + 272, + 244, + 285, + 257 + ], + "type": "image", + "image_path": "b682c5e385bd75a15c1d2b276cb37078c4522e84fad83bd37ef1a89a990cc93d.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 291, + 239, + 526, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 239, + 526, + 261 + ], + "spans": [ + { + "bbox": [ + 291, + 239, + 526, + 261 + ], + "type": "text", + "content": "Calculating Marginal Revenue Product of Labor: Option (B) states.... It is consistent with option (B), as it takes into account both the marginal product and the additional revenue per unit of output." + } + ] + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 272, + 266, + 285, + 277 + ], + "blocks": [ + { + "bbox": [ + 272, + 266, + 285, + 277 + ], + "lines": [ + { + "bbox": [ + 272, + 266, + 285, + 277 + ], + "spans": [ + { + "bbox": [ + 272, + 266, + 285, + 277 + ], + "type": "image", + "image_path": "b0220de4bbd6d3766239bb19dff80782ce63b64e3fb19ed27b1a3f836751f7d0.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 291, + 263, + 528, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 263, + 528, + 277 + ], + "spans": [ + { + "bbox": [ + 291, + 263, + 528, + 277 + ], + "type": "text", + "content": "Analyzing Option (C): Option (C) claims ... For the second worker, ... We can calculate that ... of the second worker is 7 multiplies " + }, + { + "bbox": [ + 291, + 263, + 528, + 277 + ], + "type": "inline_equation", + "content": "20, which is" + }, + { + "bbox": [ + 291, + 263, + 528, + 277 + ], + "type": "text", + "content": "140, contradicting option (C)." + } + ] + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 272, + 282, + 285, + 294 + ], + "blocks": [ + { + "bbox": [ + 272, + 282, + 285, + 294 + ], + "lines": [ + { + "bbox": [ + 272, + 282, + 285, + 294 + ], + "spans": [ + { + "bbox": [ + 272, + 282, + 285, + 294 + ], + "type": "image", + "image_path": "553a02bf49b2decbc863c0c383795908b9070b1c39f5494fc5d8e1b880fba4ed.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "bbox": [ + 291, + 281, + 529, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 281, + 529, + 295 + ], + "spans": [ + { + "bbox": [ + 291, + 281, + 529, + 295 + ], + "type": "text", + "content": "Calculating Marginal Revenue: Option (D) indicates ... It can be ... that the pots sell $20 each. This means that ... is equal to $20, matching option (D)." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 272, + 298, + 336, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 298, + 336, + 306 + ], + "spans": [ + { + "bbox": [ + 272, + 298, + 336, + 306 + ], + "type": "text", + "content": "SUMMARY AGENT" + } + ] + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 272, + 308, + 285, + 319 + ], + "blocks": [ + { + "bbox": [ + 272, + 308, + 285, + 319 + ], + "lines": [ + { + "bbox": [ + 272, + 308, + 285, + 319 + ], + "spans": [ + { + "bbox": [ + 272, + 308, + 285, + 319 + ], + "type": "image", + "image_path": "1c59585149320464353ff81377b9ff8fd6f9e297dd040872780c31df26c96a79.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "bbox": [ + 291, + 308, + 477, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 308, + 477, + 316 + ], + "spans": [ + { + "bbox": [ + 291, + 308, + 477, + 316 + ], + "type": "text", + "content": "Based on the reasoning process above, The incorrect option is option (C)." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 56, + 373, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 373, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 296, + 588 + ], + "type": "text", + "content": "training, we compare our curated dataset against the widely-used RLAIF-V [51] dataset, which contains 80K DPO data pairs for alignment. We also investigate the potential benefits of iterative DPO in advancing the model's reasoning capabilities. For an unbiased comparison, we subsample the RLAIF-V dataset to approximately 15K data, aligning it with the size of our dataset. The results, summarized in Table 4, show an average improvement of approximately " + }, + { + "bbox": [ + 56, + 373, + 296, + 588 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 56, + 373, + 296, + 588 + ], + "type": "text", + "content": " on the evaluated benchmarks when trained with the RLAIF-V dataset, whereas our curated dataset achieves greater gains of " + }, + { + "bbox": [ + 56, + 373, + 296, + 588 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 56, + 373, + 296, + 588 + ], + "type": "text", + "content": ". This suggests that a DPO dataset based on model-generated rationales, rather than externally sourced data, more effectively boosts reasoning accuracy. Moreover, conducting two additional rounds of DPO training using the same methodology yields further performance gains of " + }, + { + "bbox": [ + 56, + 373, + 296, + 588 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 56, + 373, + 296, + 588 + ], + "type": "text", + "content": ", underscoring that iterative DPO training progressively refines the model's ability to generate precise and high-quality reasoning processes compared to a single training pass." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 55, + 599, + 167, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 167, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 167, + 611 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "We present qualitative comparisons in Figure 5 to illustrate the improvements introduced by integrating the Insight-V system. Specifically, we provide an example that highlights advancements in the reasoning process. In this example, the model is challenged with a multiple-choice question based on a table. We compare our approach with a direct Chain-of-Thought application, and the model undergoes supervised fine-tuning without incorporating a multi-agent sys" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 372, + 555, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 576 + ], + "type": "text", + "content": "tem, which is denoted as Insight-V (Vanilla). It is evident that applying Chain-of-Thought directly results in suboptimal reasoning and leads to incorrect answers. Although fine-tuning a single model yields somewhat better reasoning, the model begins to falter as the reasoning chain lengthens, ultimately arriving at incorrect answers. This outcome arises because the model is required to generate both reasoning steps and the final answer simultaneously, limiting its judgment capabilities and weakening its robustness in handling flawed reasoning chains. In contrast, employing Insight-V enables more logical step decomposition and a structured reasoning chain. The model can analyze each option step-by-step and perform detailed calculations, which the other two methods cannot achieve. The fine-tuned summary model is able to evaluate the reasoning process and determine whether to derive the final answer based on it, significantly enhancing system robustness and ensuring correct answers." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 314, + 586, + 387, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 387, + 597 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 387, + 597 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "In this paper, we introduce Insight-V, a novel system that combines a scalable data generation pipeline for long-chain reasoning data and an effective multi-agent training pipeline to enhance the reasoning capabilities of MLLMs. By developing this system, we provide a scalable approach to model training aimed at improving reasoning performance. Our extensive evaluation across various benchmarks demonstrates the effectiveness of our approach, paving the way for equipping MLLMs with enhanced reasoning capabilities." + } + ] + } + ], + "index": 49 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9069" + } + ] + } + ], + "index": 50 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 135 + ], + "type": "text", + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 296, + 224 + ], + "type": "text", + "content": "[3] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 225, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 296, + 278 + ], + "type": "text", + "content": "[4] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 296, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 296, + 323 + ], + "type": "text", + "content": "[5] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 325, + 296, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 296, + 380 + ], + "type": "text", + "content": "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 445 + ], + "type": "text", + "content": "[7] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 501 + ], + "type": "text", + "content": "[8] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 555 + ], + "type": "text", + "content": "[9] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 296, + 612 + ], + "type": "text", + "content": "[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 614, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 296, + 667 + ], + "type": "text", + "content": "[11] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[12] Muyang He, Yexin Liu, Boya Wu, Jianhao Yuan, Yueze Wang, Tiejun Huang, and Bo Zhao. Efficient multimodal learning from data-centric perspective. arXiv preprint arXiv:2402.11530, 2024. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[13] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In ECCV, pages 235-251. Springer, 2016. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "text", + "content": "[14] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. arXiv preprint arXiv:2408.12637, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 163, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 555, + 195 + ], + "type": "text", + "content": "[15] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 554, + 229 + ], + "type": "text", + "content": "[16] Bo Li. Xinrun du, yuhao dong, haotian liu, yuanhan zhang, ge zhang, chunyuan li, and ziwei liu. Lmms-eval: Accelerating the development of large multimoal models, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 232, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 555, + 274 + ], + "type": "text", + "content": "[17] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 277, + 555, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 555, + 340 + ], + "type": "text", + "content": "[18] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 343, + 555, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 555, + 396 + ], + "type": "text", + "content": "[19] Zhenwen Liang, Tianyu Yang, Jipeng Zhang, and Xiangliang Zhang. Unimath: A foundational and multimodal mathematical reasoner. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7126-7133, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 399, + 555, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 555, + 441 + ], + "type": "text", + "content": "[20] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 444, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 555, + 476 + ], + "type": "text", + "content": "[21] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 478, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 555, + 510 + ], + "type": "text", + "content": "[22] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.1,2,5,6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 512, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 555, + 533 + ], + "type": "text", + "content": "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 535, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 555, + 578 + ], + "type": "text", + "content": "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 580, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 554, + 623 + ], + "type": "text", + "content": "[25] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 625, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 555, + 667 + ], + "type": "text", + "content": "[26] Yuan Liu, Zhongyin Zhao, Ziyuan Zhuang, Le Tian, Xiao Zhou, and Jie Zhou. Points: Improving your vision-language model with affordable strategies. arXiv preprint arXiv:2409.04828, 2024. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[27] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 1, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9070" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 115 + ], + "type": "text", + "content": "[28] Zuyan Liu, Yuhao Dong, Yongming Rao, Jie Zhou, and Jiwen Lu. Chain-of-spot: Interactive reasoning improves large vision-language models. arXiv preprint arXiv:2403.12966, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 296, + 171 + ], + "type": "text", + "content": "[29] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Yaofeng Sun, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 227 + ], + "type": "text", + "content": "[30] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 270 + ], + "type": "text", + "content": "[31] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv:2405.20797, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 271, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 296, + 304 + ], + "type": "text", + "content": "[32] Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. Dolphins: Multimodal language model for driving. arXiv preprint arXiv:2312.00438, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 296, + 348 + ], + "type": "text", + "content": "[33] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 349, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 296, + 392 + ], + "type": "text", + "content": "[34] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 392, + 296, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 392, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 296, + 414 + ], + "type": "text", + "content": "[35] OpenAI. Gpt-4 technical report. ArXiv:abs/2303.08774, 2023.6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 282, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 282, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 282, + 426 + ], + "type": "text", + "content": "[36] OpenAI. Hello gpt-4o — openai. OpenAI Blog, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 448 + ], + "type": "text", + "content": "[37] QwenTeam. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 449, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 470 + ], + "type": "text", + "content": "[38] QwenTeam. Qwen2-vl: To see the world more clearly. Wwen Blog, 2024. 1, 2, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 471, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 525 + ], + "type": "text", + "content": "[39] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 525, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 296, + 581 + ], + "type": "text", + "content": "[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 582, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 602 + ], + "type": "text", + "content": "[41] Qwen Team. Qwen2.5: A party of foundation models, 2024. 1, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 603, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 296, + 657 + ], + "type": "text", + "content": "[42] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "text", + "content": "[43] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 2, 5, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 556, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 556, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 556, + 127 + ], + "type": "text", + "content": "[44] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 129, + 555, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 163 + ], + "type": "text", + "content": "[45] Tianyi Xiong, Bo Li, Dong Guo, Huizhuo Yuan, Quanquan Gu, and Chunyuan Li. Llava-onevision-chat: Improving chat with preference learning, 2024. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "text", + "content": "[46] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 219, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 555, + 274 + ], + "type": "text", + "content": "[47] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, et al. Octopus: Embodied vision-language programmer from environmental feedback. arXiv preprint arXiv:2310.08588, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 275, + 556, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 556, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 556, + 330 + ], + "type": "text", + "content": "[48] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 555, + 409 + ], + "type": "text", + "content": "[49] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint 2408.01800, 2024. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 410, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 555, + 453 + ], + "type": "text", + "content": "[50] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 455, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 555, + 510 + ], + "type": "text", + "content": "[51] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 510, + 556, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 556, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 556, + 578 + ], + "type": "text", + "content": "[52] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 578, + 556, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 556, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 556, + 633 + ], + "type": "text", + "content": "[53] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Ming Yin, Botao Yu, Ge Zhang, et al. Mmmu-pro: A more robust multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024.6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 634, + 556, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 634, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 634, + 556, + 689 + ], + "type": "text", + "content": "[54] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Dufter, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, et al. Mm1. 5: Methods, analysis & insights from multimodal llm fine-tuning. arXiv preprint arXiv:2409.20566, 2024. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 690, + 556, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 556, + 714 + ], + "type": "text", + "content": "[55] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9071" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 294 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 149 + ], + "type": "text", + "content": "Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 151, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 296, + 205 + ], + "type": "text", + "content": "[56] Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Yichi Zhang, Ziyu Guo, Chengzhuo Tong, Jiaming Liu, Aojun Zhou, Bin Wei, Shanghang Zhang, et al. Mavis: Mathematical visual instruction tuning. arXiv preprint arXiv:2407.08739, 2024. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 206, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 296, + 251 + ], + "type": "text", + "content": "[57] Ruohong Zhang, Bowen Zhang, Yanghao Li, Haotian Zhang, Zhiqing Sun, Zhe Gan, Yinfei Yang, Ruoming Pang, and Yiming Yang. Improve vision language model chain-of-thought reasoning. arXiv preprint arXiv:2410.16198, 2024. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 296, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 296, + 294 + ], + "type": "text", + "content": "[58] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2, 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9072" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/fb96c75a-1b6e-4864-bf83-3b9ecc089e1a_content_list.json b/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/fb96c75a-1b6e-4864-bf83-3b9ecc089e1a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..684b8b46731b211a8d8dcc9ff61e4c7f50842251 --- /dev/null +++ b/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/fb96c75a-1b6e-4864-bf83-3b9ecc089e1a_content_list.json @@ -0,0 +1,2142 @@ +[ + { + "type": "text", + "text": "InsightEdit: Towards Better Instruction Following for Image Editing", + "text_level": 1, + "bbox": [ + 151, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingjing Xu $^{1,2*}$ , Jie Kong $^{1*†}$ , Jiazhi Wang $^{1}$ , Xiao Pan $^{2}$ , Bo Lin $^{2‡}$ , Qiang Liu $^{1}$", + "bbox": [ + 199, + 179, + 795, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "101.AI, Zhejiang University", + "bbox": [ + 380, + 215, + 614, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{poppyxu, xiaopan, rainbowlin}@zju.edu.cn, {kongjie, wangjiazhi, liuqiang}@01.ai", + "bbox": [ + 143, + 236, + 844, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 328, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we focus on the task of instruction-based image editing. Previous works like InstructPix2Pix, Instruct-Diffusion, and SmartEdit have explored end-to-end editing. However, two limitations still remain: First, existing datasets suffer from low resolution, poor background consistency, and overly simplistic instructions. Second, current approaches mainly condition on the text while the rich image information is underexplored, therefore inferior in complex instruction following and maintaining background consistency. Targeting these issues, we first curated the AdvancedEdit dataset using a novel data construction pipeline, formulating a large-scale dataset with high visual quality, complex instructions, and good background consistency. Then, to further inject the rich image information, we introduce a two-stream bridging mechanism utilizing both the textual and visual features reasoned by the powerful Multimodal Large Language Models (MLLM) to guide the image editing process more precisely. Extensive results demonstrate that our approach, InsightEdit, achieves state-of-the-art performance, excelling in complex instruction following and maintaining high background consistency with the original image. The project page is https://poppyxu.github.io/InsightEdit_web/.", + "bbox": [ + 89, + 316, + 485, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 691, + 220, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image editing involves altering an image's appearance, structure, or content, encompassing a range of changes from subtle adjustments to major transformations [13]. It has witnessed significant advancements in image editing tasks alongside the development of diffusion models [2, 4, 9, 16, 31], leading to impressive results.", + "bbox": [ + 89, + 718, + 482, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing works like InstructPix2Pix [2], MagicBrush [39], and InstructDiffusion [9] have investigated end-to-end image editing. Some recent efforts [8, 14]", + "bbox": [ + 89, + 809, + 483, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "explore using MLLM to comprehend instructions, thereby enhancing the ability to cope with complex editing tasks. However, two challenges still remain: (i) Lack of high-quality datasets. As illustrated in Table 1, current datasets suffer from low resolution, poor visual quality and background consistency, and overly simplistic, template-based instructions. Specifically, most datasets rely on Prompt2Prompt [10] method, which struggles to achieve precise control over the generated images and fails to maintain consistency in unedited regions. These limitations hinder the image editing model's capability in both complex instruction following and high-fidelity target image generation. (ii) Lack of rich image condition. Current methods primarily use the CLIP text encoder to provide conditions, yet it often exhibits limited ability to understand the instructions. Some approaches further leverage the more advanced MLLM to better comprehend instructions. However, they still mainly focus on understanding instructions at the textual level, while neglecting to capture the rich visual semantics of the image, therefore showing weak performance in complex instruction following and maintaining background consistency.", + "bbox": [ + 511, + 287, + 906, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the dataset issue, we propose an automated data construction pipeline to generate editing pairs with complex instructions and strong background consistency. Leveraging the perceptive capabilities of MLLM, our approach extracts detailed object information from high-resolution images and utilizes an advanced mask-based editing model to produce realistic and controllable edits. Using this method, we introduce AdvancedEdit Dataset, a comprehensive collection comprising over 2,500,000 editing pairs with high visual quality, complex instructions, and good background consistency.", + "bbox": [ + 511, + 622, + 908, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the lack of rich image condition, we employ a two-stream bridging mechanism to integrate both high-level textual and rich visual information into the denoising process of the diffusion model. This interaction effectively extracts the conditional information, enabling precise and controllable image editing.", + "bbox": [ + 511, + 791, + 908, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our method achieves state-of-the-art performance on", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 107, + 862, + 222, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Primary Corresponding author.", + "bbox": [ + 109, + 875, + 281, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Co-Corresponding author.", + "bbox": [ + 109, + 887, + 254, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2694", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c7dc7f87eb72e6d6704d242315e121e5d7054de52ef619824a36ab46178c060c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 88, + 354, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6bb8b9091b919a393682da3bc13ceb1d64476149ae5b890faf7660723abda5f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 251, + 354, + 381 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f4e0946b6093aa9caf613b4753e7328ec9a6f513da4daf7dbaf6117b4566b101.jpg", + "image_caption": [ + "\"Locate the green items suspended from above and substitute them with pendant lighting.\"" + ], + "image_footnote": [], + "bbox": [ + 367, + 88, + 630, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/981f049a89af991e806e9c391b21c230d13681ceabf01aecbb72b865e68bfa83.jpg", + "image_caption": [ + "\"Find the white container and change it to a wooden one.\"" + ], + "image_footnote": [], + "bbox": [ + 367, + 251, + 630, + 353 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c127acf9a33ba1e2342e9cb929b6b86d8437c5997837e3bc203c33efc6e71c10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 88, + 905, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d70b1edb69ce5b9248e612f19bf6c5de34f27cef5a200a18fecbcf5e3fc548c9.jpg", + "image_caption": [ + "\"Please locate the tallest building and add a clock tower in front of it.\"", + "Figure 1. We propose InsightEdit, an end-to-end instruction-based image editing model, trained on high-quality data and designed to fully harness the capabilities of Multimodal Large Language Models (MLLM), achieving high-quality edits with strong instruction-following and background consistency." + ], + "image_footnote": [], + "bbox": [ + 640, + 251, + 905, + 353 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/ca2eb1ceacedfc5f6adae12775adad1d8bf82533e8a7f0767fa39c0f766de060.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetEditing PairsResolutionGood Background ConsistencyComplex InstructionsHigh Visual QualityAutomated Pipeline
InstructPix2Pix [2]313,0105122XXX✓
MagicBrush [39]10,38810242✓X✓X
HIVE [42]1,100,000+5122XXX✓
FaithfulEdit [6]52,20810242XX✓✓
UltraEdit [43]4,108,2625122✓✓X✓
HQEdit [15]187,3509002XX✓✓
EditWorld [35]10,000+5122X✓✓✓
AdvancedEdit (Ours)2,536,67410242✓✓✓✓
", + "bbox": [ + 161, + 455, + 834, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison between previous datasets and ours. We provide more detailed comparisons in the appendix.", + "bbox": [ + 151, + 607, + 841, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "both the Reason-Edit [14] and AdvancedEdit-Eval, demonstrating its strong capability in complex instruction following and maintaining background consistency.", + "bbox": [ + 89, + 648, + 482, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 109, + 699, + 410, + 714 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose an automated data construction pipeline for image editing, specifically engineered to facilitate the training of models for complex instruction-based image editing tasks.", + "2. We produce the AdvancedEdit Dataset, a large-scale, high-quality, and fine-grained image editing dataset with complex instructions and good background consistency. Additionally, we introduce the AdvancedEdit-Eval dataset to assess the model's ability to handle complex instructions.", + "3. We present InsightEdit, which utilizes a two-stream bridging mechanism to guide the image editing pro" + ], + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cess with both textual and image features reasoned by MLLM. Extensive experiments are conducted to prove the model's ability to follow complex instructions and maintain high background consistency.", + "bbox": [ + 532, + 648, + 906, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 727, + 653, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Editing Methods", + "text_level": 1, + "bbox": [ + 511, + 753, + 730, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in image editing can be broadly categorized into mask-based and mask-free approaches. Mask-based image editing [1, 31-33] requires the original image, instructions, and mask image indicating the regions to be edited as input. The mask image typically needs to be manually specified by the user. Blended Latent Diffusion [1] extends Latent Diffusion Model [28] by integrating them into a localized image editing framework, enabling the tar", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2695", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "geted redrawing of specific regions during the denoising process. BrushNet [16] is inspired by the architecture of ControlNet [40]. It employs a dual UNet [29] structure, where an additional branch is used to enhance feature extraction from the masked regions. Power-Paint [44] focuses on the distinct characteristics of various image editing tasks by learning a token to classify different editing tasks. Mask-based approaches enable localized redrawing and fine-grained control, resulting in a stable and high-visual-quality output. However, they require additional mask information and are sensitive to mask shape and the specific editing task.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mask-free methods [2, 9, 14, 19] present a promising alternative by reducing the dependency on explicit masks and allowing for more flexible and intuitive image editing. InstructPix2Pix [2] utilizes GPT-3 [25] and a Prompt2Prompt [10] methodology to construct an image editing dataset. InstructDiffusion [9] builds upon the network design of InstructPix2Pix, aiming to unify various vision tasks through joint training. MGIE [19] enhances instruction-based image editing by learning to generate expressive instructions with MLLM. SmartEdit [14] utilizes MLLM to generate text embeddings, leveraging the MLLM's advanced reasoning and comprehension capabilities.", + "bbox": [ + 89, + 273, + 480, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although instruction-based image editing methods do not require explicit mask guidance, making them more intuitive and convenient, they often yield lower image quality than mask-based approaches. They also face challenges in following complex instructions and maintaining background consistency.", + "bbox": [ + 89, + 470, + 480, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Image Editing Datasets", + "text_level": 1, + "bbox": [ + 89, + 573, + 305, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several works have recently introduced image editing datasets. InstructPix2Pix [2] introduces a large-scale image editing dataset created through fine-tuned GPT-3 and a Prompt-to-Prompt methodology integrated with Stable Diffusion. MagicBrush [39] further provides a limited-scale, manually annotated dataset for instruction-guided real image editing. Subsequent work, HIVE [42] introduces more training triplets and human ranking results to provide stronger supervisory signals for improved model training. FaithfulEdits [6] employed inpainting techniques followed by a filtering process using Visual Question Answering (VQA) models. UltraEdit [43] explores region-based data generation, employing mask generation to facilitate object redrawing in specified areas. Some works explore the diversity of instructions, such as EditWorld [35], which defines and categorizes instructions grounded in various world scenarios.", + "bbox": [ + 89, + 597, + 482, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Table 1, despite the notable contributions of these datasets, current image editing datasets still face several challenges, including the inability to scale the data con", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "struction process, low dataset resolution, poor background consistency, and a lack of fine-grained instructions for complex understanding and reasoning tasks.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset Construction", + "text_level": 1, + "bbox": [ + 513, + 152, + 718, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Automated Pipeline", + "text_level": 1, + "bbox": [ + 513, + 176, + 700, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose an automated data construction pipeline focused on generating high-fidelity, fine-grained image-editing pairs with detailed instructions that demonstrate advanced reasoning and understanding. We categorize the image editing tasks into three types: removal, addition, and replacement. Figure 2 presents our data preparation workflow.", + "bbox": [ + 511, + 200, + 903, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step 1: Caption & Object Extraction. We leverage the advanced comprehension capabilities of MLLM to generate a global caption that effectively conveys the image's content. Using this caption, we utilize LLM to extract a JSON list of objects, identifying those with physical significance. Each object is defined by a simple caption (e.g., blue T-shirt) and a detailed description (e.g., A shirt worn by the man, typically made of cotton, featuring a blue color).", + "bbox": [ + 511, + 308, + 903, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step 2: Mask Generation. For mask generation, we apply GroundedSAM [27] to extract local masks for each object, filtering out low-confidence masks based on a predefined threshold, thus obtaining accurate object and mask pairs for further processing.", + "bbox": [ + 511, + 430, + 903, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step 3: Editing Pair Construction. As mentioned in Section 2.2, mask-based image editing models exhibit superior image generation capabilities, offering better control over the generation process for specific tasks. In the process of constructing the edited images, we employed the state-of-the-art mask-based methods [16, 44] to generate the target image, enabling the creation of more fine-grained and controllable image editing pairs. We categorized the image editing generation tasks into three types: removal, addition, and replacement.", + "bbox": [ + 511, + 506, + 903, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For removal, we apply a mask-based approach, using both the original image and the object mask to generate the post-removal image, simultaneously generating and storing the template instruction: \"remove the [object].\" For addition, we swap the source and target images. The template instruction is: \"add the [object].\"", + "bbox": [ + 511, + 657, + 903, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the replacement task, we leverage MLLM's creative capabilities to propose an alternative object for the scene, often resulting in plausible and innovative outcomes. Simultaneously, we generate and store the template instruction: \"replace the [source object] with the [target object].\".", + "bbox": [ + 511, + 750, + 903, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step 4: Instruction Recaptioning. To diversify the complexity of the instructions, we recaption the instructions into both simple and advanced versions. The simple instructions are generated through synonym replacement and task template modifications. For instance, the instruction \"replace", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2696", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8c589f15fece6d7542f2e747c02b17c50ce8144eb60512fabdfef654b9d079b5.jpg", + "image_caption": [ + "Figure 2. The overall data construction pipeline. (1) Captioning & Object Extraction: Utilizing VLM to generate a global caption from the source image, and further get an object JSON list contains both simple caption and detailed caption. (2) Mask Generation: Utilizing GroundedSAM to obtain the corresponding mask of each object. (3) Editing Pair Construction: Utilizing mask-based image editing model to construct target image and templated instruction. (4) Instruction Recaptioning: Utilizing VLM to rewrite instruction to gain diverse instructions. (5) Quality Evaluation: Filtering the datasets using VIEScore." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 361, + 323 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0c283059f42ef96137bf57d8da4975ef7a356feb0d6ed2142a8e1c89f9609e29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 88, + 617, + 321 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fdc20728baf73486c2f36cdfc3f2efbe925edef9bf84d97e42a78db9772ed462.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 88, + 903, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the [acoustic guitar] with [white guitar]\" is rewritten as \"Swap the [acoustic guitar] for the [white guitar].\"", + "bbox": [ + 89, + 419, + 480, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast, the advanced instructions are designed to increase complexity in two key ways. First, we replace simple object descriptions with detailed ones prepared in Step 1 and apply different task templates. Second, we recaption instructions to gain reasoning capabilities, such as \"What is the man holding in his hands? Please change the color of this object to white.\" This approach challenges the editing model to generate images with more intricate details and greater precision.", + "bbox": [ + 89, + 450, + 483, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 5: Quality Evaluation. To construct a high-quality dataset, we assess and filter the image-editing pairs using VIEScore [20], which aligns evaluations with human preferences via MLLM. The evaluation framework consists of two components: semantic consistency, which measures instruction following and evaluates whether the image has been over-edited, and perceptual quality, which evaluates image fidelity, including the presence of artifacts or blurriness. The rating example of VIEScore can be seen in Figure 2.", + "bbox": [ + 89, + 585, + 483, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. AdvancedEdit Dataset", + "text_level": 1, + "bbox": [ + 89, + 747, + 302, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We choose Pexels1, a high-quality real-world photographic image dataset, as our source data. Pexels is a popular platform offering a large collection of high-resolution, royalty-free images contributed by photographers around the world. The images in the Pexels dataset generally exhibit an average resolution of approximately 2K or higher. This dataset contains over 1 million photographic images and encom-", + "bbox": [ + 89, + 770, + 483, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "passes a broad range of subjects, covering various world scenes. Given that each image contains multiple objects and is associated with various instruction tasks, our dataset comprises a total of 2,536,674 editing pairs. A more detailed analysis of the proposed dataset is provided in the appendix.", + "bbox": [ + 511, + 419, + 906, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define the image editing pairs with simple instructions as SimpleEdit, and image editing pairs with complex instructions as AdvancedEdit. We will discuss the influence of instructions with varying complexity in Section 5.3. Additionally, we introduce the AdvancedEdit-Eval, comprising 300 curated image pairs, covering a range of tasks including removal, addition, and replacement. The evaluation dataset includes a variety of intricate and nuanced image editing scenarios, necessitating that the model possesses a certain level of understanding and reasoning capabilities.", + "bbox": [ + 511, + 496, + 905, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 661, + 604, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. The overall architecture of InsightEdit is depicted in Figure 3. It mainly consists of a comprehension module, a bridging module, and a generation module. Specifically, the comprehension module leverages MLLM to comprehend the image editing task; the bridging module integrates both text and image features into the denoising process of the diffusion model; and the generation module receives editing guidance via the diffusion model to generate the target image.", + "bbox": [ + 511, + 686, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Comprehension Module", + "text_level": 1, + "bbox": [ + 511, + 834, + 735, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal large language model receives the original image and the editing instruction to comprehend the image editing task. In the comprehension module, we use LLaVA", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1https://www.pexels.com/", + "bbox": [ + 109, + 886, + 250, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2697", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/053b18cac162e469609c292dc3ff0d9a7fc2f48c4136078261aca143b64928c4.jpg", + "image_caption": [ + "Figure 3. The overall architecture of InsightEdit. It mainly consists of three parts: (1) Comprehension Module: A comprehension module that leverages MLLM to perceive and comprehend the image editing task; (2) Bridging Module: A bridging module that better interacts and extracts both the textual and image features; (3) Generation Module: A generation module that receives editing guidance via diffusion model to generate the target image." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 591, + 314 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/91b662947536a85858a387af89f98f6e94c430739aed7cc78ca62dd3b13b0e8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 89, + 903, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7B [24] as our vision-language foundation model. The feature of source image $\\mathbf{I}_{src}$ is extracted by a vision encoder $E_{\\phi}(\\cdot)$ and the text instruction is tokenized as text embeddings $c$ . Aligned by a fully connect layer, the original image features are sent into the LLaVA decoder [30] along with text embeddings. The mentioned process is represented as:", + "bbox": [ + 89, + 400, + 483, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nv = \\operatorname {F C} \\left(E _ {\\phi} \\left(\\mathbf {I} _ {s r c}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 505, + 480, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh = \\mathrm {L L a V A} _ {\\omega} (v, c).\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 523, + 354, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by GILL [19], we extend the vocabulary of the LLM by introducing $r$ special [MM] tokens, where \"MM\" indicates multi-modality comprehension of text and image information via MLLM. These tokens are appended to the end of the instruction $c$ , and the model is trained by minimizing the negative log-likelihood of predicting [MM] tokens, conditioned on the previously generated tokens. The loss function can be represented below:", + "bbox": [ + 89, + 553, + 483, + 675 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {L L M}} (c) = - \\sum_ {i = 1} ^ {r} \\log p _ {\\left\\{\\omega \\cup \\theta \\right\\}} ([ \\mathbf {M M} _ {i} ] \\mid v, c, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 688, + 482, + 727 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ \\mathbf {M M} _ {1} ], \\dots , [ \\mathbf {M M} _ {i - 1} ]).\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 729, + 418, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The majority of the LLM parameters are kept frozen, while the LoRA [12] is employed to facilitate efficient training.", + "bbox": [ + 89, + 760, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Bridging Module", + "text_level": 1, + "bbox": [ + 89, + 804, + 261, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The textual features provide high-level information while the image features provide more detailed condition. Therefore, in the bridging module, to comprehensively integrate both textual and image features as the conditions, we design a dual-stream condition alignment methodology.", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Textual Branch via Q-Former & BIM. For the textual branch, similar to SmartEdit [14], we start from the text-aligned Q-Former [5, 22] which has primarily aligned with the original clip text embedding space. Then, given the learnable tokens $q$ as inputs, we propose to further extract the reasoned textual information from the [MM] token hidden states $h$ via cross-attention:", + "bbox": [ + 511, + 400, + 906, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq ^ {\\prime} = Q _ {\\beta} (q, h), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 656, + 517, + 903, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q^{\\prime}$ represents the Q-former outputs that mainly encode the text instruction information.", + "bbox": [ + 511, + 544, + 903, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, following SmartEdit [14], we employ the BIM module that enables bidirectional information exchange between the input image and the MLLM output:", + "bbox": [ + 511, + 575, + 905, + 621 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {t x t}, v _ {t x t} = \\operatorname {B I M} \\left(E _ {\\phi} \\left(\\mathbf {I} _ {s r c}\\right), q ^ {\\prime}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 631, + 903, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{txt}$ and $v_{txt}$ are the interacted features obtained from the BIM module. $f_{txt}$ is used as the final textual condition to UNet and $v_{txt}$ is the textual-aware vision features which is later added with U-Net inputs.", + "bbox": [ + 511, + 659, + 905, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Branch via IAA. For the image branch, we propose an Image Alignment Adapter (IAA) module to activate the image information reasoned by MLLM, which consists of a mapper, a linear layer, and a layer normalization.", + "bbox": [ + 511, + 719, + 905, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In detail, the mapper is a Multi-Layer Perceptron (MLP) that transforms the [MM] token hidden states $h \\in \\mathbb{R}^{r \\times 4096}$ into a single embedding $\\operatorname{Mapper}(h) \\in \\mathbb{R}^{1 \\times 768}$ , which is then aligned with the target image features extracted by a CLIP vision encoder. Intuitively, the target image contains direct information about the editing goal, e.g., the expected background, therefore can provide explicit supervision. After that, the linear is further employed to expand", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2698", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/270da8f0266e1083e661f5dc11abb3abb32d9176503fd9adbd36234ef0e66f28.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on AdvancedEdit-Eval." + ], + "table_footnote": [], + "table_body": "
MethodsVIEScore↑CLIPScore↑PSNR↑SSIM↑LPIPS↓
InstructPix2Pix0.34219.52820.1920.6940.182
MagicBrush0.35219.75122.6360.7430.132
InstructDiffusion0.31819.39018.0250.6240.223
MGIE0.36119.04720.0740.6910.199
SmartEdit-7B0.68220.11420.1150.6510.131
InsightEdit0.73820.39521.2670.6750.112
InsightEdit with AdvancedEdit0.83121.00222.8710.7160.071
", + "bbox": [ + 212, + 102, + 784, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the $\\operatorname{Mapper}(h) \\in \\mathbb{R}^{1 \\times 768}$ back to token sequences $\\mathbb{R}^{N \\times 768}$ ( $N$ is set as 4 by default) followed by a layer normalization. Formally, the mentioned process is represented as:", + "bbox": [ + 89, + 234, + 480, + 282 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {I A A}} = \\left\\| \\operatorname {C L I P} \\left(\\mathbf {I} _ {\\text {t a r}}\\right) - \\operatorname {M a p p e r} (h) \\right\\| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 292, + 480, + 319 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i m g} = \\operatorname {I A A} (h),\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 316, + 272, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{IAA}}$ is the supervision from the target image, and $f_{img}$ represents the output of the IAA, which will be sent to further cross-attention operation in U-Net.", + "bbox": [ + 89, + 343, + 482, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to textual features, the image features contain more detailed and comprehensive information, which can guide image editing more explicitly.", + "bbox": [ + 89, + 388, + 483, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Generation Module", + "text_level": 1, + "bbox": [ + 89, + 445, + 277, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the process of target image generation, inspired by IP-adapter [37], we apply a decoupled cross-attention mechanism to integrate both textual and image features. An additional cross-attention layer is employed on top of the cross-attention layer in the original UNet block to interact with the image features. As the text cross-attention and image cross-attention are detached, we can also adjust the weight of the image condition in the inference stage:", + "bbox": [ + 89, + 464, + 483, + 585 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {Z} = \\text {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K} _ {\\mathrm {t x t}}, \\mathbf {V} _ {\\mathrm {t x t}}) \\tag {6} \\\\ + \\lambda \\cdot \\text {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K} _ {\\mathrm {i m g}}, \\mathbf {V} _ {\\mathrm {i m g}}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 598, + 480, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{K}_{\\mathrm{img}}$ and $\\mathbf{V}_{\\mathrm{img}}$ are projected by $f_{img}$ , $\\mathbf{K}_{\\mathrm{txt}}$ and $\\mathbf{V}_{\\mathrm{txt}}$ are projected by $f_{txt}$ .", + "bbox": [ + 89, + 645, + 482, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the diffusion model, we concatenate the encoded image latent $\\mathcal{E}(\\mathbf{I}_{src})$ with the noisy latent $z_{t}$ . The process can be formulated as:", + "bbox": [ + 89, + 676, + 483, + 720 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i f f}} = \\mathbb {E} _ {\\mathcal {E} (y), \\mathcal {E} (x), c _ {T}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\\\ \\left[ \\| \\epsilon - \\epsilon_ {\\delta} (t, \\operatorname {c o n c a t} [ z _ {t}, \\mathcal {E} (x) ] + v _ {t}, f _ {t x t}, f _ {i m g}) \\| _ {2} ^ {2} \\right]. \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 746, + 482, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 89, + 825, + 176, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 89, + 849, + 282, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. To train InsightEdit, we use CC12M [7], InstructPix2Pix [2], MagicBrush [39],", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "COCO [23], RefCOCO, GRefCOCO [17, 38], CO-COStuff [3], LISA [21, 36], ReasonEdit [14] and our proposed AdvancedEdit. Due to the limited resources, we only used 202, 822 editing pairs from AdvancedEdit for the experiments.", + "bbox": [ + 511, + 236, + 903, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "InsightEdit is trained in three stages, as detailed in the Appendix, and the training is conducted on a single machine with 8 H100 GPUs. We adopt the AdamW optimizer [18] as the optimizer during three stages. In the first training stage, the learning rate is $2e^{-4}$ and weight decay is 0. In the second stage and the third stage, the values of learning rate, weight decay, and warm-up ratio were set to $1e^{-5}$ , 0, and 0.001, respectively. In the data construction pipeline, the MLLM we utilize is GPT-4o $^2$ , while the LLM applied is Qwen2 [34].", + "bbox": [ + 511, + 313, + 905, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. For evaluating the background consistency, we use PSNR, SSIM, and LPIPS [11, 41]. For the edited area, we calculate the CLIPScore [26] by comparing the edited object with its ground truth (GT) label. To more accurately evaluate editing effects and align with human preferences, we employ VIEScore [20], which utilizes MLLM to evaluate editing performance from two perspectives: instruction-following and background consistency. We emphasize that VIEScore might better align with human preferences, making it a more convincing measure [20].", + "bbox": [ + 511, + 465, + 905, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison with State-of-the-art", + "text_level": 1, + "bbox": [ + 511, + 627, + 805, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on AdvancedEdit-Eval. We compare our method with existing state-of-the-art instruction-based image methods, including instructPix2Pix, MagicBrush, InstructDiffusion, MGIE, and SmartEdit. First, we conduct comparative experiments on the AdvancedEdit-Eval as introduced in Section 3.2. The quantitative results presented in Table 2 show that InsightEdit shows good performance across all five metrics. The experimental results in PSNR and LPIPS scores prove that InsightEdit excels in preserving non-edited regions, while the enhanced CLIPScore validates our method's superior ability to follow instructions with greater accuracy. Specifically, in terms of VIEScore, both InsightEdit itself and InsightEdit trained on the AdvancedEdit dataset achieve state-of-the-art results, demonstrating that the model's generated outputs are preferred by", + "bbox": [ + 511, + 648, + 906, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://openai.com/", + "bbox": [ + 529, + 886, + 643, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2699", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/19b3a21691411425ff727ad8b174345ea74fdc773294a4c1757df30d51797c90.jpg", + "table_caption": [ + "Table 3. Quantitative comparison on Reason-Edit. All the methods we compared have been fine-tuned using the same training data as that used by SmartEdit." + ], + "table_footnote": [], + "table_body": "
MethodsUnderstanding ScenariosReasoning Scenarios
VIEScore↑CLIPScore↑PSNR↑SSIM↑LPIPS↓VIEScore↑CLIPScore↑PSNR↑SSIM↑LPIPS↓
InstructPix2Pix0.49322.76221.5760.7210.0890.54819.41324.2340.7070.083
MagicBrush0.42222.62018.1200.6800.1430.53019.75522.1010.6940.113
InstructDiffusion0.45723.08023.2580.7430.0670.67019.52321.4530.6660.117
MGIE0.37321.94716.0940.6950.2000.54018.03722.9770.7430.144
SmartEdit-7B0.86623.61122.0490.7310.0870.83520.95025.2580.7420.055
InsightEdit0.90123.73423.5880.7490.0670.89320.86725.7120.7470.049
InsightEdit with AdvancedEdit0.93424.42124.5030.7600.0540.94721.14126.0900.7500.047
", + "bbox": [ + 91, + 116, + 906, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8996cd85997d748b8cf9b7dd045aa3059304a60676708cea9bf74293e4ad052b.jpg", + "image_caption": [ + "Source Image" + ], + "image_footnote": [], + "bbox": [ + 133, + 251, + 235, + 329 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/715447c62efa55a4b9e9a38ff7c2823d9815558b62fd2a208eaa17029f62b351.jpg", + "image_caption": [ + "InstructPixPix", + "\"Identify the the cinnamon sticks next to the bowl and remove them from the wooden surface.\"" + ], + "image_footnote": [], + "bbox": [ + 259, + 251, + 361, + 328 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c40d5175cea8353a97b1f27429a65c944c7c27dc0504da14995139e90d1b527.jpg", + "image_caption": [ + "MagicBrush InstructDiffusion" + ], + "image_footnote": [], + "bbox": [ + 361, + 251, + 460, + 328 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bab0c5aa878c33c95b4c3cfb3c07633c9c9983310819c72b594df6f8ae538bf3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 251, + 560, + 328 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b9cbe9bf00783a3f256293406a22974eee2985b44860717c5173ace1a08c1ad.jpg", + "image_caption": [ + "n MGIE" + ], + "image_footnote": [], + "bbox": [ + 563, + 251, + 661, + 328 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46e689f02cfe57e29e6fa82763f86f797869da10dbe17e9f7262bcb39bd4ce8e.jpg", + "image_caption": [ + "SmartEdit" + ], + "image_footnote": [], + "bbox": [ + 663, + 251, + 761, + 328 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/600ae3f3c78d2af20480a9f8b79e7addab8c285b6d505c15deb808b980513db2.jpg", + "image_caption": [ + "InsightEdit" + ], + "image_footnote": [], + "bbox": [ + 764, + 251, + 864, + 328 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d97cef3c3a776c4b41ce8efc2024eb70255d1e476859390e91022aaa4f0e755.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 347, + 235, + 425 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14075f120499dbb64bf154579a77b1d342977872642af1999d9766004ec1fd27.jpg", + "image_caption": [ + "\"Identify the colorful flowers in the scene and modify them to a vibrant yellow hue.\"" + ], + "image_footnote": [], + "bbox": [ + 259, + 347, + 359, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f06ec95ce02751ea04e754fb529972f177d8da2af06598b80c703f4bfff33383.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 347, + 460, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0761d918299cba6a7d4c149a0a743e53b99d579c4c1c246078ac1afb72a06612.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 347, + 560, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/215ed51fa28efdd2c9526463f74df1ed31b61c46c47edd328bda4a4348bc5553.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 347, + 661, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61a3a414cbdd1de3b55f9e8ec48a64f838b30252a96612361a5ad54fd3ddec74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 347, + 761, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/265427873a64724e3d8242cd07b04c25083a1213766927cc7663ea4ff55f3a33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 347, + 864, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9e8b40a317fc2dbddd4f80b1bbe3eef969afd4e6538769464eedede098aaa1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 441, + 235, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f3b8e4f4a6614cc07396e92208efcc2bc1fd6edf06dd7b2ccc190d9db3bf2e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 441, + 359, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be2134f83ee258b5da9f97ac71dc14995fa8c62a83c1cb403dd0c703e212c632.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 441, + 460, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/095d63cead02e28cc4726abac328694283939853f1cc3978743141f168b3e49c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 441, + 560, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e0688527964835f4ed51d8a5035219f08e5388b1aff6ce0dd0698d0db017b7da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 441, + 661, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/135a8847e1a85bcb9de5ecf311fc26d19e78c69effca41b67e2c6e4af7226054.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 441, + 761, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/efb725dfe94c902325446d147b1d1036c0137780664586e1e8bf3611c38facf2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 441, + 864, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe75ee3ce9a3ece48c1d3ad701613a65e9b613d4b6c905d5432f90a9a0777399.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 537, + 235, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/32da01294b3d084a381d2de8f71c631616b376ab2f100a02935037693c965905.jpg", + "image_caption": [ + "\"Add an ornate, large crystal chandelier hanging from the center of the ceiling.\"", + "\"There is a person sitting in the center of the image, remove it.\"" + ], + "image_footnote": [], + "bbox": [ + 259, + 537, + 359, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c7b6ae1f9b5417e6229f74a460a1f1e953192f2280e3468f54c320a2e1f3fbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 537, + 460, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f60cdce23972f672d834616bbc61335910a5e034dcd3d81b866be4ef563db57a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 537, + 560, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5292e0289c63c2d2c92461b9ebcf5a7ad1c94becbb4dbc1cfb5edb28bbdadbd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 537, + 661, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf1adffa360ca40c42e0ade0085d54d50e9885b34c43b79ecb1f425218fb1148.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 537, + 761, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36f2bc973f8a8d3bcd909e31ebdd8887dcf496f0ef9a3f86ce053cf24b1068be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 537, + 864, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8725f5438b4b4f9fc8b71cdf3b350b9d447b69beba70881cca3da54d569e3523.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison on AdvancedEdit. InsightEdit shows superior instruction following and background consistency capability." + ], + "image_footnote": [], + "bbox": [ + 133, + 633, + 235, + 712 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e9edd03326b5583b2b259a5250db061dbd747c80d5b1eb0002ab92e5a5ab046.jpg", + "image_caption": [ + "\"Please observe if there's a house in the image and change its appearance to a cozy cabin with a warm, inviting glow from within.\"" + ], + "image_footnote": [], + "bbox": [ + 259, + 633, + 359, + 712 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1eeb5fd9d19789ef53f81ec70fe8efa93ed53cb048b190b863708e84866eb423.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 633, + 460, + 712 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c8247aac02ad0619c71ebe6795670748be763529ca4d834171a04cae35faccb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 633, + 560, + 712 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57f914d0468f0e00ff9340ea949384ac5fcf54854ad4b5656e03d378cb38be5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 633, + 661, + 712 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6362222ae5c33dc04ac8a696f3995e2f13bf28b25fc109ca2920e10cb4bd42d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 633, + 761, + 712 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c8f9d7fc05eb30d1c630039b9379e75ca1196d324b71d04865af4c859719a6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 633, + 864, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "humans over other methods, whether regarding background preservation or the ability to follow complex instructions.", + "bbox": [ + 89, + 772, + 482, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 illustrates the qualitative results. As shown in the last column, our method trained on the AdvancedEdit dataset demonstrates a clear advantage over competing approaches. As can be seen, other methods struggle to maintain background consistency with the original image, often introducing artifacts or unintended alterations, while In", + "bbox": [ + 89, + 810, + 483, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sightEdit provides specific editing on the required regions. In the meanwhile, InsightEdit performs strong instruction following capabilities, having a great understanding of spatial, color, number, and detailed objects.", + "bbox": [ + 511, + 772, + 906, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Reason-Edit. Additionally, we conducted comparative experiments on the Reason-Edit[14], which is specifically designed to assess editing abilities with an emphasis on understanding and reasoning across various sce", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2700", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9c9fc26848dc01cb8483263136a00e6fc369ee95a6c44d277d83f4c688fa4f78.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IAAPSNR↑SSIM↑LPIPS↓CLIPScore↑VIEScore↑
22.3480.6920.09520.6527.307
✓22.8710.7160.07121.0027.545
", + "bbox": [ + 93, + 88, + 482, + 121 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d90bdf94379e4ddbae2a55307eb0d604bd7bb68637d751bb1b7cd42b79e71b81.jpg", + "image_caption": [ + "\"Could you swap the white flowers with a diverse collection of colorful wildflowers, creating a vivid and colorful foreground?\"" + ], + "image_footnote": [], + "bbox": [ + 109, + 152, + 439, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e93fbd28f48acfa79239cac18a2c0a58b82993a781137cd93f1d8cf6ddb738d.jpg", + "image_caption": [ + "Transform the worn wooden table into a sleek, modern marble countertop while maintaining the arrangement of the objects." + ], + "image_footnote": [], + "bbox": [ + 112, + 273, + 439, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "narios (i.e. color, position, mirror, multiple-objects, reason, and size). The results in Table 3 show that InsightEdit outperforms other metrics, except for a slight CLIP score difference in reasoning tasks, where it slightly trails SmartEdit. Furthermore, when trained on our AdvancedEdit dataset, InsightEdit demonstrates improvements across all metrics, highlighting the model's ability to leverage the dataset to enhance both comprehension and reasoning capabilities, while also emphasizing the intrinsic value of the data used.", + "bbox": [ + 89, + 416, + 483, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 561, + 254, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Studies on IAA. To validate the effectiveness of IAA, we conduct comparative experiments on InsightEdit with the AdvancedEdit dataset. Table 4 shows the quantitative results, it can be seen that, with the IAA module, InsightEdit leads to notable improvements across various evaluation metrics. The qualitative results can be seen in Figure 5. In the first example, without the IAA module, the house is erroneously transformed into wildflowers, resulting in poor background consistency due to the absence of the red house. In contrast, with the IAA module, the red house is preserved accurately, maintaining consistency with the source image. In the second example, the marble table is edited poorly without the IAA module, while the addition of the IAA module leads to a more natural edit that aligns well with the instruction.", + "bbox": [ + 89, + 583, + 483, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Studies on Instruction Complexity. We investigate the impact of instruction complexity and granularity on model performance by comparing two datasets introduced in Section 3.1: SimpleEdit Dataset and AdvancedEdit Dataset. Table 5 demonstrates that incorporating the SimpleEdit Dataset alone enhances the model's per", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ce84b54749afa34d997c90bb1de38a84560940d55640544553dea27ba9662ece.jpg", + "table_caption": [ + "Table 4. Ablation of IAA. Our IAA module performs best." + ], + "table_footnote": [], + "table_body": "
SimpleEdit AdvancedEditPSNR↑SSIM↑LPIPS↓CLIPScore↑VIEScore↑
21.2670.6750.11220.3956.974
✓22.2600.7080.08020.5577.213
✓22.8710.7160.07121.0027.545
", + "bbox": [ + 514, + 88, + 906, + 131 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation of AdvancedEdit. Our method with AdvancedEdit performs best.", + "bbox": [ + 513, + 133, + 903, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/921b9dcd1015ad834c3d0fdbb40d162a1758aeaec682dc409a59363cbcec7d9a.jpg", + "image_caption": [ + "Figure 5. Demonstration of the effectiveness of the IAA module.", + "\"Can you identify the specific bright light source from the small window on the staircase wall, remove it smoothly and blend the area to match the surrounding tones?\"", + "Figure 6. Demonstration of the effectiveness of AdvancedEdit." + ], + "image_footnote": [], + "bbox": [ + 517, + 175, + 900, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formance, primarily due to the high-quality outputs generated by our mask-based image editing approach. However, the inclusion of the AdvancedEdit Dataset leads to substantial improvements, particularly in CLIPScore and VI-EScore, highlighting the importance of increased instruction detail and the value of integrating complex instructions. The qualitative comparison results, as illustrated in Figure 6, reveal that without any data augmentation, the method struggles to understand the instructions and produce effective edits in both two examples. When the SimpleEdit dataset is introduced, there is some improvement, though the edit quality remains suboptimal. However, with the use of the AdvancedEdit dataset, the method demonstrates the ability to accurately interpret editing instructions, resulting in vivid and precise edits. For instance, in the first example, the method produces finely detailed orange and black butterfly wings, while in the second example, it accurately identifies and removes a light source, maintaining harmony in the surrounding content. These results further show that advanced instruction can handle more complex image editing scenarios and generate more vivid editing results.", + "bbox": [ + 511, + 412, + 906, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 739, + 761, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, we present InsightEdit, an end-to-end instruction-based image editing method, utilizing a two-stream bridging mechanism to integrate both the textual and visual features reasoned by the powerful MLLM into the diffusion model to guide the image editing process more precisely. We design an automatic data construction pipeline to generate a large-scale, high-quality, and high-fidelity image editing dataset, considering both image and instruction perspectives.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2701", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023.", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023.", + "[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018.", + "[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023.", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020.", + "[6] Tuhin Chakrabarty, Kanishk Singh, Arkadiy Saakyan, and Smaranda Muresan. Learning to follow object-centric image editing instructions faithfully. arXiv preprint arXiv:2310.19145, 2023.", + "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021.", + "[8] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023.", + "[9] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024.", + "[10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022.", + "[11] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010.", + "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021.", + "[13] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525, 2024.", + "[14] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024.", + "[15] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024.", + "[16] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024.", + "[17] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014.", + "[18] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.", + "[19] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. Advances in Neural Information Processing Systems, 36, 2024.", + "[20] Max Ku, Dongfu Jiang, Cong Wei, Xiang Yue, and Wenhu Chen. Viescore: Towards explainable metrics for conditional image synthesis evaluation. arXiv preprint arXiv:2312.14867, 2023.", + "[21] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023.", + "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023.", + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014.", + "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024.", + "[25] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020.", + "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021.", + "[27] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2702", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024.", + "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022.", + "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015.", + "[30] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.", + "[31] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18359-18369, 2023.", + "[32] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023.", + "[33] Shaoan Xie, Yang Zhao, Zhisheng Xiao, Kelvin CK Chan, Yandong Li, Yanwu Xu, Kun Zhang, and Tingbo Hou. Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. arXiv preprint arXiv:2312.03771, 2023.", + "[34] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024.", + "[35] Ling Yang, Bohan Zeng, Jiaming Liu, Hong Li, Minghao Xu, Wentao Zhang, and Shuicheng Yan. Editworld: Simulating world dynamics for instruction-following image editing. arXiv preprint arXiv:2405.14785, 2024.", + "[36] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. arXiv preprint arXiv:2312.17240, 2023.", + "[37] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023.", + "[38] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016." + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024.", + "[40] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023.", + "[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018.", + "[42] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9026-9036, 2024.", + "[43] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024.", + "[44] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. arXiv preprint arXiv:2312.03594, 2023." + ], + "bbox": [ + 516, + 90, + 903, + 444 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2703", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/fb96c75a-1b6e-4864-bf83-3b9ecc089e1a_model.json b/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/fb96c75a-1b6e-4864-bf83-3b9ecc089e1a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4d9b0451042d50c5bd156e1f922e38deec6e9a85 --- /dev/null +++ b/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/fb96c75a-1b6e-4864-bf83-3b9ecc089e1a_model.json @@ -0,0 +1,2750 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.131, + 0.848, + 0.154 + ], + "angle": 0, + "content": "InsightEdit: Towards Better Instruction Following for Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.18, + 0.797, + 0.201 + ], + "angle": 0, + "content": "Yingjing Xu\\(^{1,2*}\\), Jie Kong\\(^{1*†}\\), Jiazhi Wang\\(^{1}\\), Xiao Pan\\(^{2}\\), Bo Lin\\(^{2‡}\\), Qiang Liu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.216, + 0.616, + 0.234 + ], + "angle": 0, + "content": "101.AI, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.237, + 0.846, + 0.253 + ], + "angle": 0, + "content": "{poppyxu, xiaopan, rainbowlin}@zju.edu.cn, {kongjie, wangjiazhi, liuqiang}@01.ai" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.329, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.486, + 0.667 + ], + "angle": 0, + "content": "In this paper, we focus on the task of instruction-based image editing. Previous works like InstructPix2Pix, Instruct-Diffusion, and SmartEdit have explored end-to-end editing. However, two limitations still remain: First, existing datasets suffer from low resolution, poor background consistency, and overly simplistic instructions. Second, current approaches mainly condition on the text while the rich image information is underexplored, therefore inferior in complex instruction following and maintaining background consistency. Targeting these issues, we first curated the AdvancedEdit dataset using a novel data construction pipeline, formulating a large-scale dataset with high visual quality, complex instructions, and good background consistency. Then, to further inject the rich image information, we introduce a two-stream bridging mechanism utilizing both the textual and visual features reasoned by the powerful Multimodal Large Language Models (MLLM) to guide the image editing process more precisely. Extensive results demonstrate that our approach, InsightEdit, achieves state-of-the-art performance, excelling in complex instruction following and maintaining high background consistency with the original image. The project page is https://poppyxu.github.io/InsightEdit_web/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.693, + 0.222, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Image editing involves altering an image's appearance, structure, or content, encompassing a range of changes from subtle adjustments to major transformations [13]. It has witnessed significant advancements in image editing tasks alongside the development of diffusion models [2, 4, 9, 16, 31], leading to impressive results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Existing works like InstructPix2Pix [2], MagicBrush [39], and InstructDiffusion [9] have investigated end-to-end image editing. Some recent efforts [8, 14]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.908, + 0.621 + ], + "angle": 0, + "content": "explore using MLLM to comprehend instructions, thereby enhancing the ability to cope with complex editing tasks. However, two challenges still remain: (i) Lack of high-quality datasets. As illustrated in Table 1, current datasets suffer from low resolution, poor visual quality and background consistency, and overly simplistic, template-based instructions. Specifically, most datasets rely on Prompt2Prompt [10] method, which struggles to achieve precise control over the generated images and fails to maintain consistency in unedited regions. These limitations hinder the image editing model's capability in both complex instruction following and high-fidelity target image generation. (ii) Lack of rich image condition. Current methods primarily use the CLIP text encoder to provide conditions, yet it often exhibits limited ability to understand the instructions. Some approaches further leverage the more advanced MLLM to better comprehend instructions. However, they still mainly focus on understanding instructions at the textual level, while neglecting to capture the rich visual semantics of the image, therefore showing weak performance in complex instruction following and maintaining background consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.909, + 0.789 + ], + "angle": 0, + "content": "To address the dataset issue, we propose an automated data construction pipeline to generate editing pairs with complex instructions and strong background consistency. Leveraging the perceptive capabilities of MLLM, our approach extracts detailed object information from high-resolution images and utilizes an advanced mask-based editing model to produce realistic and controllable edits. Using this method, we introduce AdvancedEdit Dataset, a comprehensive collection comprising over 2,500,000 editing pairs with high visual quality, complex instructions, and good background consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.909, + 0.883 + ], + "angle": 0, + "content": "To address the lack of rich image condition, we employ a two-stream bridging mechanism to integrate both high-level textual and rich visual information into the denoising process of the diffusion model. This interaction effectively extracts the conditional information, enabling precise and controllable image editing." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Our method achieves state-of-the-art performance on" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.863, + 0.223, + 0.876 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.876, + 0.282, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Primary Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.255, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Co-Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.863, + 0.282, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2694" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.355, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.252, + 0.356, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.089, + 0.631, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.213, + 0.612, + 0.25 + ], + "angle": 0, + "content": "\"Locate the green items suspended from above and substitute them with pendant lighting.\"" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.252, + 0.631, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.357, + 0.63, + 0.382 + ], + "angle": 0, + "content": "\"Find the white container and change it to a wooden one.\"" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.089, + 0.906, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.252, + 0.906, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.357, + 0.904, + 0.382 + ], + "angle": 0, + "content": "\"Please locate the tallest building and add a clock tower in front of it.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.401, + 0.908, + 0.443 + ], + "angle": 0, + "content": "Figure 1. We propose InsightEdit, an end-to-end instruction-based image editing model, trained on high-quality data and designed to fully harness the capabilities of Multimodal Large Language Models (MLLM), achieving high-quality edits with strong instruction-following and background consistency." + }, + { + "type": "table", + "bbox": [ + 0.163, + 0.456, + 0.836, + 0.607 + ], + "angle": 0, + "content": "
DatasetEditing PairsResolutionGood Background ConsistencyComplex InstructionsHigh Visual QualityAutomated Pipeline
InstructPix2Pix [2]313,0105122XXX✓
MagicBrush [39]10,38810242✓X✓X
HIVE [42]1,100,000+5122XXX✓
FaithfulEdit [6]52,20810242XX✓✓
UltraEdit [43]4,108,2625122✓✓X✓
HQEdit [15]187,3509002XX✓✓
EditWorld [35]10,000+5122X✓✓✓
AdvancedEdit (Ours)2,536,67410242✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.152, + 0.608, + 0.843, + 0.623 + ], + "angle": 0, + "content": "Table 1. Comparison between previous datasets and ours. We provide more detailed comparisons in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.483, + 0.695 + ], + "angle": 0, + "content": "both the Reason-Edit [14] and AdvancedEdit-Eval, demonstrating its strong capability in complex instruction following and maintaining background consistency." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.7, + 0.411, + 0.715 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.72, + 0.483, + 0.78 + ], + "angle": 0, + "content": "1. We propose an automated data construction pipeline for image editing, specifically engineered to facilitate the training of models for complex instruction-based image editing tasks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.78, + 0.483, + 0.87 + ], + "angle": 0, + "content": "2. We produce the AdvancedEdit Dataset, a large-scale, high-quality, and fine-grained image editing dataset with complex instructions and good background consistency. Additionally, we introduce the AdvancedEdit-Eval dataset to assess the model's ability to handle complex instructions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "3. We present InsightEdit, which utilizes a two-stream bridging mechanism to guide the image editing pro" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.649, + 0.907, + 0.71 + ], + "angle": 0, + "content": "cess with both textual and image features reasoned by MLLM. Extensive experiments are conducted to prove the model's ability to follow complex instructions and maintain high background consistency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.728, + 0.655, + 0.744 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.731, + 0.772 + ], + "angle": 0, + "content": "2.1. Image Editing Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Recent advancements in image editing can be broadly categorized into mask-based and mask-free approaches. Mask-based image editing [1, 31-33] requires the original image, instructions, and mask image indicating the regions to be edited as input. The mask image typically needs to be manually specified by the user. Blended Latent Diffusion [1] extends Latent Diffusion Model [28] by integrating them into a localized image editing framework, enabling the tar" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2695" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "geted redrawing of specific regions during the denoising process. BrushNet [16] is inspired by the architecture of ControlNet [40]. It employs a dual UNet [29] structure, where an additional branch is used to enhance feature extraction from the masked regions. Power-Paint [44] focuses on the distinct characteristics of various image editing tasks by learning a token to classify different editing tasks. Mask-based approaches enable localized redrawing and fine-grained control, resulting in a stable and high-visual-quality output. However, they require additional mask information and are sensitive to mask shape and the specific editing task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.482, + 0.468 + ], + "angle": 0, + "content": "Mask-free methods [2, 9, 14, 19] present a promising alternative by reducing the dependency on explicit masks and allowing for more flexible and intuitive image editing. InstructPix2Pix [2] utilizes GPT-3 [25] and a Prompt2Prompt [10] methodology to construct an image editing dataset. InstructDiffusion [9] builds upon the network design of InstructPix2Pix, aiming to unify various vision tasks through joint training. MGIE [19] enhances instruction-based image editing by learning to generate expressive instructions with MLLM. SmartEdit [14] utilizes MLLM to generate text embeddings, leveraging the MLLM's advanced reasoning and comprehension capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.482, + 0.561 + ], + "angle": 0, + "content": "Although instruction-based image editing methods do not require explicit mask guidance, making them more intuitive and convenient, they often yield lower image quality than mask-based approaches. They also face challenges in following complex instructions and maintaining background consistency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.307, + 0.59 + ], + "angle": 0, + "content": "2.2. Image Editing Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Several works have recently introduced image editing datasets. InstructPix2Pix [2] introduces a large-scale image editing dataset created through fine-tuned GPT-3 and a Prompt-to-Prompt methodology integrated with Stable Diffusion. MagicBrush [39] further provides a limited-scale, manually annotated dataset for instruction-guided real image editing. Subsequent work, HIVE [42] introduces more training triplets and human ranking results to provide stronger supervisory signals for improved model training. FaithfulEdits [6] employed inpainting techniques followed by a filtering process using Visual Question Answering (VQA) models. UltraEdit [43] explores region-based data generation, employing mask generation to facilitate object redrawing in specified areas. Some works explore the diversity of instructions, such as EditWorld [35], which defines and categorizes instructions grounded in various world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, despite the notable contributions of these datasets, current image editing datasets still face several challenges, including the inability to scale the data con" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "struction process, low dataset resolution, poor background consistency, and a lack of fine-grained instructions for complex understanding and reasoning tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.153, + 0.719, + 0.168 + ], + "angle": 0, + "content": "3. Dataset Construction" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.178, + 0.702, + 0.194 + ], + "angle": 0, + "content": "3.1. Automated Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.202, + 0.905, + 0.306 + ], + "angle": 0, + "content": "We propose an automated data construction pipeline focused on generating high-fidelity, fine-grained image-editing pairs with detailed instructions that demonstrate advanced reasoning and understanding. We categorize the image editing tasks into three types: removal, addition, and replacement. Figure 2 presents our data preparation workflow." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.309, + 0.905, + 0.429 + ], + "angle": 0, + "content": "Step 1: Caption & Object Extraction. We leverage the advanced comprehension capabilities of MLLM to generate a global caption that effectively conveys the image's content. Using this caption, we utilize LLM to extract a JSON list of objects, identifying those with physical significance. Each object is defined by a simple caption (e.g., blue T-shirt) and a detailed description (e.g., A shirt worn by the man, typically made of cotton, featuring a blue color)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.431, + 0.905, + 0.505 + ], + "angle": 0, + "content": "Step 2: Mask Generation. For mask generation, we apply GroundedSAM [27] to extract local masks for each object, filtering out low-confidence masks based on a predefined threshold, thus obtaining accurate object and mask pairs for further processing." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.507, + 0.905, + 0.657 + ], + "angle": 0, + "content": "Step 3: Editing Pair Construction. As mentioned in Section 2.2, mask-based image editing models exhibit superior image generation capabilities, offering better control over the generation process for specific tasks. In the process of constructing the edited images, we employed the state-of-the-art mask-based methods [16, 44] to generate the target image, enabling the creation of more fine-grained and controllable image editing pairs. We categorized the image editing generation tasks into three types: removal, addition, and replacement." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.658, + 0.905, + 0.748 + ], + "angle": 0, + "content": "For removal, we apply a mask-based approach, using both the original image and the object mask to generate the post-removal image, simultaneously generating and storing the template instruction: \"remove the [object].\" For addition, we swap the source and target images. The template instruction is: \"add the [object].\"" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.75, + 0.905, + 0.825 + ], + "angle": 0, + "content": "For the replacement task, we leverage MLLM's creative capabilities to propose an alternative object for the scene, often resulting in plausible and innovative outcomes. Simultaneously, we generate and store the template instruction: \"replace the [source object] with the [target object].\"." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Step 4: Instruction Recaptioning. To diversify the complexity of the instructions, we recaption the instructions into both simple and advanced versions. The simple instructions are generated through synonym replacement and task template modifications. For instance, the instruction \"replace" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2696" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.362, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.089, + 0.619, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.089, + 0.904, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.324, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Figure 2. The overall data construction pipeline. (1) Captioning & Object Extraction: Utilizing VLM to generate a global caption from the source image, and further get an object JSON list contains both simple caption and detailed caption. (2) Mask Generation: Utilizing GroundedSAM to obtain the corresponding mask of each object. (3) Editing Pair Construction: Utilizing mask-based image editing model to construct target image and templated instruction. (4) Instruction Recaptioning: Utilizing VLM to rewrite instruction to gain diverse instructions. (5) Quality Evaluation: Filtering the datasets using VIEScore." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.42, + 0.482, + 0.451 + ], + "angle": 0, + "content": "the [acoustic guitar] with [white guitar]\" is rewritten as \"Swap the [acoustic guitar] for the [white guitar].\"" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.586 + ], + "angle": 0, + "content": "In contrast, the advanced instructions are designed to increase complexity in two key ways. First, we replace simple object descriptions with detailed ones prepared in Step 1 and apply different task templates. Second, we recaption instructions to gain reasoning capabilities, such as \"What is the man holding in his hands? Please change the color of this object to white.\" This approach challenges the editing model to generate images with more intricate details and greater precision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.484, + 0.736 + ], + "angle": 0, + "content": "Step 5: Quality Evaluation. To construct a high-quality dataset, we assess and filter the image-editing pairs using VIEScore [20], which aligns evaluations with human preferences via MLLM. The evaluation framework consists of two components: semantic consistency, which measures instruction following and evaluates whether the image has been over-edited, and perceptual quality, which evaluates image fidelity, including the presence of artifacts or blurriness. The rating example of VIEScore can be seen in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.748, + 0.303, + 0.763 + ], + "angle": 0, + "content": "3.2. AdvancedEdit Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.484, + 0.877 + ], + "angle": 0, + "content": "We choose Pexels1, a high-quality real-world photographic image dataset, as our source data. Pexels is a popular platform offering a large collection of high-resolution, royalty-free images contributed by photographers around the world. The images in the Pexels dataset generally exhibit an average resolution of approximately 2K or higher. This dataset contains over 1 million photographic images and encom-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.907, + 0.496 + ], + "angle": 0, + "content": "passes a broad range of subjects, covering various world scenes. Given that each image contains multiple objects and is associated with various instruction tasks, our dataset comprises a total of 2,536,674 editing pairs. A more detailed analysis of the proposed dataset is provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.648 + ], + "angle": 0, + "content": "We define the image editing pairs with simple instructions as SimpleEdit, and image editing pairs with complex instructions as AdvancedEdit. We will discuss the influence of instructions with varying complexity in Section 5.3. Additionally, we introduce the AdvancedEdit-Eval, comprising 300 curated image pairs, covering a range of tasks including removal, addition, and replacement. The evaluation dataset includes a variety of intricate and nuanced image editing scenarios, necessitating that the model possesses a certain level of understanding and reasoning capabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.662, + 0.605, + 0.677 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Overview. The overall architecture of InsightEdit is depicted in Figure 3. It mainly consists of a comprehension module, a bridging module, and a generation module. Specifically, the comprehension module leverages MLLM to comprehend the image editing task; the bridging module integrates both text and image features into the denoising process of the diffusion model; and the generation module receives editing guidance via the diffusion model to generate the target image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.835, + 0.736, + 0.851 + ], + "angle": 0, + "content": "4.1. Comprehension Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Multimodal large language model receives the original image and the editing instruction to comprehend the image editing task. In the comprehension module, we use LLaVA" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.251, + 0.901 + ], + "angle": 0, + "content": "1https://www.pexels.com/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2697" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.09, + 0.592, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.09, + 0.905, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.318, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Figure 3. The overall architecture of InsightEdit. It mainly consists of three parts: (1) Comprehension Module: A comprehension module that leverages MLLM to perceive and comprehend the image editing task; (2) Bridging Module: A bridging module that better interacts and extracts both the textual and image features; (3) Generation Module: A generation module that receives editing guidance via diffusion model to generate the target image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.493 + ], + "angle": 0, + "content": "7B [24] as our vision-language foundation model. The feature of source image \\(\\mathbf{I}_{src}\\) is extracted by a vision encoder \\(E_{\\phi}(\\cdot)\\) and the text instruction is tokenized as text embeddings \\(c\\). Aligned by a fully connect layer, the original image features are sent into the LLaVA decoder [30] along with text embeddings. The mentioned process is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.506, + 0.482, + 0.529 + ], + "angle": 0, + "content": "\\[\nv = \\operatorname {F C} \\left(E _ {\\phi} \\left(\\mathbf {I} _ {s r c}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.524, + 0.355, + 0.541 + ], + "angle": 0, + "content": "\\[\nh = \\mathrm {L L a V A} _ {\\omega} (v, c).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.676 + ], + "angle": 0, + "content": "Inspired by GILL [19], we extend the vocabulary of the LLM by introducing \\( r \\) special [MM] tokens, where \"MM\" indicates multi-modality comprehension of text and image information via MLLM. These tokens are appended to the end of the instruction \\( c \\), and the model is trained by minimizing the negative log-likelihood of predicting [MM] tokens, conditioned on the previously generated tokens. The loss function can be represented below:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.689, + 0.483, + 0.728 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {L L M}} (c) = - \\sum_ {i = 1} ^ {r} \\log p _ {\\left\\{\\omega \\cup \\theta \\right\\}} ([ \\mathbf {M M} _ {i} ] \\mid v, c, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.262, + 0.731, + 0.419, + 0.749 + ], + "angle": 0, + "content": "\\[\n[ \\mathbf {M M} _ {1} ], \\dots , [ \\mathbf {M M} _ {i - 1} ]).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.483, + 0.793 + ], + "angle": 0, + "content": "The majority of the LLM parameters are kept frozen, while the LoRA [12] is employed to facilitate efficient training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.805, + 0.262, + 0.822 + ], + "angle": 0, + "content": "4.2. Bridging Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The textual features provide high-level information while the image features provide more detailed condition. Therefore, in the bridging module, to comprehensively integrate both textual and image features as the conditions, we design a dual-stream condition alignment methodology." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.907, + 0.507 + ], + "angle": 0, + "content": "Textual Branch via Q-Former & BIM. For the textual branch, similar to SmartEdit [14], we start from the text-aligned Q-Former [5, 22] which has primarily aligned with the original clip text embedding space. Then, given the learnable tokens \\( q \\) as inputs, we propose to further extract the reasoned textual information from the [MM] token hidden states \\( h \\) via cross-attention:" + }, + { + "type": "equation", + "bbox": [ + 0.657, + 0.518, + 0.905, + 0.535 + ], + "angle": 0, + "content": "\\[\nq ^ {\\prime} = Q _ {\\beta} (q, h), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.905, + 0.574 + ], + "angle": 0, + "content": "where \\(q^{\\prime}\\) represents the Q-former outputs that mainly encode the text instruction information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.906, + 0.622 + ], + "angle": 0, + "content": "Additionally, following SmartEdit [14], we employ the BIM module that enables bidirectional information exchange between the input image and the MLLM output:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.632, + 0.905, + 0.65 + ], + "angle": 0, + "content": "\\[\nf _ {t x t}, v _ {t x t} = \\operatorname {B I M} \\left(E _ {\\phi} \\left(\\mathbf {I} _ {s r c}\\right), q ^ {\\prime}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.72 + ], + "angle": 0, + "content": "where \\( f_{txt} \\) and \\( v_{txt} \\) are the interacted features obtained from the BIM module. \\( f_{txt} \\) is used as the final textual condition to UNet and \\( v_{txt} \\) is the textual-aware vision features which is later added with U-Net inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Image Branch via IAA. For the image branch, we propose an Image Alignment Adapter (IAA) module to activate the image information reasoned by MLLM, which consists of a mapper, a linear layer, and a layer normalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In detail, the mapper is a Multi-Layer Perceptron (MLP) that transforms the [MM] token hidden states \\( h \\in \\mathbb{R}^{r \\times 4096} \\) into a single embedding \\( \\operatorname{Mapper}(h) \\in \\mathbb{R}^{1 \\times 768} \\), which is then aligned with the target image features extracted by a CLIP vision encoder. Intuitively, the target image contains direct information about the editing goal, e.g., the expected background, therefore can provide explicit supervision. After that, the linear is further employed to expand" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2698" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.318, + 0.089, + 0.679, + 0.103 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on AdvancedEdit-Eval." + }, + { + "type": "table", + "bbox": [ + 0.214, + 0.103, + 0.785, + 0.213 + ], + "angle": 0, + "content": "
MethodsVIEScore↑CLIPScore↑PSNR↑SSIM↑LPIPS↓
InstructPix2Pix0.34219.52820.1920.6940.182
MagicBrush0.35219.75122.6360.7430.132
InstructDiffusion0.31819.39018.0250.6240.223
MGIE0.36119.04720.0740.6910.199
SmartEdit-7B0.68220.11420.1150.6510.131
InsightEdit0.73820.39521.2670.6750.112
InsightEdit with AdvancedEdit0.83121.00222.8710.7160.071
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.236, + 0.482, + 0.283 + ], + "angle": 0, + "content": "the \\(\\operatorname{Mapper}(h) \\in \\mathbb{R}^{1 \\times 768}\\) back to token sequences \\(\\mathbb{R}^{N \\times 768}\\) (\\(N\\) is set as 4 by default) followed by a layer normalization. Formally, the mentioned process is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.294, + 0.482, + 0.32 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {I A A}} = \\left\\| \\operatorname {C L I P} \\left(\\mathbf {I} _ {\\text {t a r}}\\right) - \\operatorname {M a p p e r} (h) \\right\\| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.317, + 0.273, + 0.332 + ], + "angle": 0, + "content": "\\[\nf _ {i m g} = \\operatorname {I A A} (h),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.344, + 0.483, + 0.388 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{IAA}}\\) is the supervision from the target image, and \\(f_{img}\\) represents the output of the IAA, which will be sent to further cross-attention operation in U-Net." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.39, + 0.484, + 0.435 + ], + "angle": 0, + "content": "Compared to textual features, the image features contain more detailed and comprehensive information, which can guide image editing more explicitly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.446, + 0.279, + 0.459 + ], + "angle": 0, + "content": "4.3. Generation Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.465, + 0.484, + 0.587 + ], + "angle": 0, + "content": "In the process of target image generation, inspired by IP-adapter [37], we apply a decoupled cross-attention mechanism to integrate both textual and image features. An additional cross-attention layer is employed on top of the cross-attention layer in the original UNet block to interact with the image features. As the text cross-attention and image cross-attention are detached, we can also adjust the weight of the image condition in the inference stage:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.599, + 0.482, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {Z} = \\text {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K} _ {\\mathrm {t x t}}, \\mathbf {V} _ {\\mathrm {t x t}}) \\tag {6} \\\\ + \\lambda \\cdot \\text {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K} _ {\\mathrm {i m g}}, \\mathbf {V} _ {\\mathrm {i m g}}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.483, + 0.676 + ], + "angle": 0, + "content": "where \\(\\mathbf{K}_{\\mathrm{img}}\\) and \\(\\mathbf{V}_{\\mathrm{img}}\\) are projected by \\(f_{img}\\), \\(\\mathbf{K}_{\\mathrm{txt}}\\) and \\(\\mathbf{V}_{\\mathrm{txt}}\\) are projected by \\(f_{txt}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.677, + 0.484, + 0.721 + ], + "angle": 0, + "content": "For the diffusion model, we concatenate the encoded image latent \\(\\mathcal{E}(\\mathbf{I}_{src})\\) with the noisy latent \\(z_{t}\\). The process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i f f}} = \\mathbb {E} _ {\\mathcal {E} (y), \\mathcal {E} (x), c _ {T}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\\\ \\left[ \\| \\epsilon - \\epsilon_ {\\delta} (t, \\operatorname {c o n c a t} [ z _ {t}, \\mathcal {E} (x) ] + v _ {t}, f _ {t x t}, f _ {i m g}) \\| _ {2} ^ {2} \\right]. \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.826, + 0.178, + 0.84 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.851, + 0.284, + 0.867 + ], + "angle": 0, + "content": "5.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. To train InsightEdit, we use CC12M [7], InstructPix2Pix [2], MagicBrush [39]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.237, + 0.905, + 0.313 + ], + "angle": 0, + "content": "COCO [23], RefCOCO, GRefCOCO [17, 38], CO-COStuff [3], LISA [21, 36], ReasonEdit [14] and our proposed AdvancedEdit. Due to the limited resources, we only used 202, 822 editing pairs from AdvancedEdit for the experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.906, + 0.465 + ], + "angle": 0, + "content": "InsightEdit is trained in three stages, as detailed in the Appendix, and the training is conducted on a single machine with 8 H100 GPUs. We adopt the AdamW optimizer [18] as the optimizer during three stages. In the first training stage, the learning rate is \\(2e^{-4}\\) and weight decay is 0. In the second stage and the third stage, the values of learning rate, weight decay, and warm-up ratio were set to \\(1e^{-5}\\), 0, and 0.001, respectively. In the data construction pipeline, the MLLM we utilize is GPT-4o\\(^2\\), while the LLM applied is Qwen2 [34]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.906, + 0.618 + ], + "angle": 0, + "content": "Metrics. For evaluating the background consistency, we use PSNR, SSIM, and LPIPS [11, 41]. For the edited area, we calculate the CLIPScore [26] by comparing the edited object with its ground truth (GT) label. To more accurately evaluate editing effects and align with human preferences, we employ VIEScore [20], which utilizes MLLM to evaluate editing performance from two perspectives: instruction-following and background consistency. We emphasize that VIEScore might better align with human preferences, making it a more convincing measure [20]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.628, + 0.807, + 0.644 + ], + "angle": 0, + "content": "5.2. Comparison with State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.907, + 0.875 + ], + "angle": 0, + "content": "Results on AdvancedEdit-Eval. We compare our method with existing state-of-the-art instruction-based image methods, including instructPix2Pix, MagicBrush, InstructDiffusion, MGIE, and SmartEdit. First, we conduct comparative experiments on the AdvancedEdit-Eval as introduced in Section 3.2. The quantitative results presented in Table 2 show that InsightEdit shows good performance across all five metrics. The experimental results in PSNR and LPIPS scores prove that InsightEdit excels in preserving non-edited regions, while the enhanced CLIPScore validates our method's superior ability to follow instructions with greater accuracy. Specifically, in terms of VIEScore, both InsightEdit itself and InsightEdit trained on the AdvancedEdit dataset achieve state-of-the-art results, demonstrating that the model's generated outputs are preferred by" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.644, + 0.9 + ], + "angle": 0, + "content": "\\(^{2}\\)https://openai.com/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2699" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.117 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison on Reason-Edit. All the methods we compared have been fine-tuned using the same training data as that used by SmartEdit." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.117, + 0.907, + 0.226 + ], + "angle": 0, + "content": "
MethodsUnderstanding ScenariosReasoning Scenarios
VIEScore↑CLIPScore↑PSNR↑SSIM↑LPIPS↓VIEScore↑CLIPScore↑PSNR↑SSIM↑LPIPS↓
InstructPix2Pix0.49322.76221.5760.7210.0890.54819.41324.2340.7070.083
MagicBrush0.42222.62018.1200.6800.1430.53019.75522.1010.6940.113
InstructDiffusion0.45723.08023.2580.7430.0670.67019.52321.4530.6660.117
MGIE0.37321.94716.0940.6950.2000.54018.03722.9770.7430.144
SmartEdit-7B0.86623.61122.0490.7310.0870.83520.95025.2580.7420.055
InsightEdit0.90123.73423.5880.7490.0670.89320.86725.7120.7470.049
InsightEdit with AdvancedEdit0.93424.42124.5030.7600.0540.94721.14126.0900.7500.047
" + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.239, + 0.23, + 0.252 + ], + "angle": 0, + "content": "Source Image" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.252, + 0.236, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.239, + 0.361, + 0.251 + ], + "angle": 0, + "content": "InstructPixPix" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.252, + 0.362, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.239, + 0.573, + 0.252 + ], + "angle": 0, + "content": "MagicBrush InstructDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.252, + 0.461, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.463, + 0.252, + 0.562, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.239, + 0.635, + 0.251 + ], + "angle": 0, + "content": "n MGIE" + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.252, + 0.663, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.239, + 0.75, + 0.251 + ], + "angle": 0, + "content": "SmartEdit" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.252, + 0.763, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.239, + 0.856, + 0.252 + ], + "angle": 0, + "content": "InsightEdit" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.252, + 0.865, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.33, + 0.865, + 0.343 + ], + "angle": 0, + "content": "\"Identify the the cinnamon sticks next to the bowl and remove them from the wooden surface.\"" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.348, + 0.236, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.348, + 0.36, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.348, + 0.461, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.463, + 0.348, + 0.562, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.348, + 0.663, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.348, + 0.763, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.348, + 0.865, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.427, + 0.818, + 0.44 + ], + "angle": 0, + "content": "\"Identify the colorful flowers in the scene and modify them to a vibrant yellow hue.\"" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.443, + 0.236, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.443, + 0.361, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.443, + 0.461, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.463, + 0.443, + 0.562, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.443, + 0.663, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.443, + 0.763, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.443, + 0.865, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.523, + 0.803, + 0.537 + ], + "angle": 0, + "content": "\"Add an ornate, large crystal chandelier hanging from the center of the ceiling.\"" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.539, + 0.236, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.539, + 0.361, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.539, + 0.461, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.463, + 0.539, + 0.562, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.539, + 0.663, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.539, + 0.763, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.539, + 0.865, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.619, + 0.746, + 0.632 + ], + "angle": 0, + "content": "\"There is a person sitting in the center of the image, remove it.\"" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.635, + 0.236, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.635, + 0.361, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.635, + 0.461, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.463, + 0.635, + 0.562, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.635, + 0.663, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.635, + 0.763, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.635, + 0.865, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.715, + 0.862, + 0.741 + ], + "angle": 0, + "content": "\"Please observe if there's a house in the image and change its appearance to a cozy cabin with a warm, inviting glow from within.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.741, + 0.908, + 0.768 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on AdvancedEdit. InsightEdit shows superior instruction following and background consistency capability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.483, + 0.804 + ], + "angle": 0, + "content": "humans over other methods, whether regarding background preservation or the ability to follow complex instructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Figure 4 illustrates the qualitative results. As shown in the last column, our method trained on the AdvancedEdit dataset demonstrates a clear advantage over competing approaches. As can be seen, other methods struggle to maintain background consistency with the original image, often introducing artifacts or unintended alterations, while In" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.773, + 0.907, + 0.834 + ], + "angle": 0, + "content": "sightEdit provides specific editing on the required regions. In the meanwhile, InsightEdit performs strong instruction following capabilities, having a great understanding of spatial, color, number, and detailed objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Results on Reason-Edit. Additionally, we conducted comparative experiments on the Reason-Edit[14], which is specifically designed to assess editing abilities with an emphasis on understanding and reasoning across various sce" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2700" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.483, + 0.122 + ], + "angle": 0, + "content": "
IAAPSNR↑SSIM↑LPIPS↓CLIPScore↑VIEScore↑
22.3480.6920.09520.6527.307
✓22.8710.7160.07121.0027.545
" + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.124, + 0.462, + 0.136 + ], + "angle": 0, + "content": "Table 4. Ablation of IAA. Our IAA module performs best." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.153, + 0.44, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.247, + 0.425, + 0.267 + ], + "angle": 0, + "content": "\"Could you swap the white flowers with a diverse collection of colorful wildflowers, creating a vivid and colorful foreground?\"" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.274, + 0.44, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.357, + 0.428, + 0.375 + ], + "angle": 0, + "content": "Transform the worn wooden table into a sleek, modern marble countertop while maintaining the arrangement of the objects." + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.376, + 0.479, + 0.388 + ], + "angle": 0, + "content": "Figure 5. Demonstration of the effectiveness of the IAA module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.552 + ], + "angle": 0, + "content": "narios (i.e. color, position, mirror, multiple-objects, reason, and size). The results in Table 3 show that InsightEdit outperforms other metrics, except for a slight CLIP score difference in reasoning tasks, where it slightly trails SmartEdit. Furthermore, when trained on our AdvancedEdit dataset, InsightEdit demonstrates improvements across all metrics, highlighting the model's ability to leverage the dataset to enhance both comprehension and reasoning capabilities, while also emphasizing the intrinsic value of the data used." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.563, + 0.255, + 0.577 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.809 + ], + "angle": 0, + "content": "Ablation Studies on IAA. To validate the effectiveness of IAA, we conduct comparative experiments on InsightEdit with the AdvancedEdit dataset. Table 4 shows the quantitative results, it can be seen that, with the IAA module, InsightEdit leads to notable improvements across various evaluation metrics. The qualitative results can be seen in Figure 5. In the first example, without the IAA module, the house is erroneously transformed into wildflowers, resulting in poor background consistency due to the absence of the red house. In contrast, with the IAA module, the red house is preserved accurately, maintaining consistency with the source image. In the second example, the marble table is edited poorly without the IAA module, while the addition of the IAA module leads to a more natural edit that aligns well with the instruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Ablation Studies on Instruction Complexity. We investigate the impact of instruction complexity and granularity on model performance by comparing two datasets introduced in Section 3.1: SimpleEdit Dataset and AdvancedEdit Dataset. Table 5 demonstrates that incorporating the SimpleEdit Dataset alone enhances the model's per" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "
SimpleEdit AdvancedEditPSNR↑SSIM↑LPIPS↓CLIPScore↑VIEScore↑
21.2670.6750.11220.3956.974
✓22.2600.7080.08020.5577.213
✓22.8710.7160.07121.0027.545
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.134, + 0.905, + 0.16 + ], + "angle": 0, + "content": "Table 5. Ablation of AdvancedEdit. Our method with AdvancedEdit performs best." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.176, + 0.901, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.356, + 0.886, + 0.373 + ], + "angle": 0, + "content": "\"Can you identify the specific bright light source from the small window on the staircase wall, remove it smoothly and blend the area to match the surrounding tones?\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.373, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Figure 6. Demonstration of the effectiveness of AdvancedEdit." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.73 + ], + "angle": 0, + "content": "formance, primarily due to the high-quality outputs generated by our mask-based image editing approach. However, the inclusion of the AdvancedEdit Dataset leads to substantial improvements, particularly in CLIPScore and VI-EScore, highlighting the importance of increased instruction detail and the value of integrating complex instructions. The qualitative comparison results, as illustrated in Figure 6, reveal that without any data augmentation, the method struggles to understand the instructions and produce effective edits in both two examples. When the SimpleEdit dataset is introduced, there is some improvement, though the edit quality remains suboptimal. However, with the use of the AdvancedEdit dataset, the method demonstrates the ability to accurately interpret editing instructions, resulting in vivid and precise edits. For instance, in the first example, the method produces finely detailed orange and black butterfly wings, while in the second example, it accurately identifies and removes a light source, maintaining harmony in the surrounding content. These results further show that advanced instruction can handle more complex image editing scenarios and generate more vivid editing results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.762, + 0.755 + ], + "angle": 0, + "content": "6. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In conclusion, we present InsightEdit, an end-to-end instruction-based image editing method, utilizing a two-stream bridging mechanism to integrate both the textual and visual features reasoned by the powerful MLLM into the diffusion model to guide the image editing process more precisely. We design an automatic data construction pipeline to generate a large-scale, high-quality, and high-fidelity image editing dataset, considering both image and instruction perspectives." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2701" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.483, + 0.156 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.158, + 0.483, + 0.213 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.339 + ], + "angle": 0, + "content": "[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[6] Tuhin Chakrabarty, Kanishk Singh, Arkadiy Saakyan, and Smaranda Muresan. Learning to follow object-centric image editing instructions faithfully. arXiv preprint arXiv:2310.19145, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.578 + ], + "angle": 0, + "content": "[8] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[9] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[11] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[13] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[15] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[16] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[17] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.904, + 0.375 + ], + "angle": 0, + "content": "[18] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[19] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. Advances in Neural Information Processing Systems, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[20] Max Ku, Dongfu Jiang, Cong Wei, Xiang Yue, and Wenhu Chen. Viescore: Towards explainable metrics for conditional image synthesis evaluation. arXiv preprint arXiv:2312.14867, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[21] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[25] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2702" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.301 + ], + "angle": 0, + "content": "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[30] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[31] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18359-18369, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[32] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[33] Shaoan Xie, Yang Zhao, Zhisheng Xiao, Kelvin CK Chan, Yandong Li, Yanwu Xu, Kun Zhang, and Tingbo Hou. Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. arXiv preprint arXiv:2312.03771, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[34] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[35] Ling Yang, Bohan Zeng, Jiaming Liu, Hong Li, Minghao Xu, Wentao Zhang, and Shuicheng Yan. Editworld: Simulating world dynamics for instruction-following image editing. arXiv preprint arXiv:2405.14785, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[36] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. arXiv preprint arXiv:2312.17240, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[37] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[39] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[40] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[42] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9026-9036, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[43] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[44] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. arXiv preprint arXiv:2312.03594, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2703" + } + ] +] \ No newline at end of file diff --git a/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/fb96c75a-1b6e-4864-bf83-3b9ecc089e1a_origin.pdf b/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/fb96c75a-1b6e-4864-bf83-3b9ecc089e1a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bbe5b00020512bec57ed733b5d9aab60ca120e48 --- /dev/null +++ b/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/fb96c75a-1b6e-4864-bf83-3b9ecc089e1a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3f75a89d3204ef3e660327bf57aa4c96737a7e41805907e24522cc963e0a4b8 +size 7798707 diff --git a/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/full.md b/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..99b6bbf8108012aa448a0516cec531ce2f08d61c --- /dev/null +++ b/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/full.md @@ -0,0 +1,402 @@ +# InsightEdit: Towards Better Instruction Following for Image Editing + +Yingjing Xu $^{1,2*}$ , Jie Kong $^{1*†}$ , Jiazhi Wang $^{1}$ , Xiao Pan $^{2}$ , Bo Lin $^{2‡}$ , Qiang Liu $^{1}$ + +101.AI, Zhejiang University + +{poppyxu, xiaopan, rainbowlin}@zju.edu.cn, {kongjie, wangjiazhi, liuqiang}@01.ai + +# Abstract + +In this paper, we focus on the task of instruction-based image editing. Previous works like InstructPix2Pix, Instruct-Diffusion, and SmartEdit have explored end-to-end editing. However, two limitations still remain: First, existing datasets suffer from low resolution, poor background consistency, and overly simplistic instructions. Second, current approaches mainly condition on the text while the rich image information is underexplored, therefore inferior in complex instruction following and maintaining background consistency. Targeting these issues, we first curated the AdvancedEdit dataset using a novel data construction pipeline, formulating a large-scale dataset with high visual quality, complex instructions, and good background consistency. Then, to further inject the rich image information, we introduce a two-stream bridging mechanism utilizing both the textual and visual features reasoned by the powerful Multimodal Large Language Models (MLLM) to guide the image editing process more precisely. Extensive results demonstrate that our approach, InsightEdit, achieves state-of-the-art performance, excelling in complex instruction following and maintaining high background consistency with the original image. The project page is https://poppyxu.github.io/InsightEdit_web/. + +# 1. Introduction + +Image editing involves altering an image's appearance, structure, or content, encompassing a range of changes from subtle adjustments to major transformations [13]. It has witnessed significant advancements in image editing tasks alongside the development of diffusion models [2, 4, 9, 16, 31], leading to impressive results. + +Existing works like InstructPix2Pix [2], MagicBrush [39], and InstructDiffusion [9] have investigated end-to-end image editing. Some recent efforts [8, 14] + +explore using MLLM to comprehend instructions, thereby enhancing the ability to cope with complex editing tasks. However, two challenges still remain: (i) Lack of high-quality datasets. As illustrated in Table 1, current datasets suffer from low resolution, poor visual quality and background consistency, and overly simplistic, template-based instructions. Specifically, most datasets rely on Prompt2Prompt [10] method, which struggles to achieve precise control over the generated images and fails to maintain consistency in unedited regions. These limitations hinder the image editing model's capability in both complex instruction following and high-fidelity target image generation. (ii) Lack of rich image condition. Current methods primarily use the CLIP text encoder to provide conditions, yet it often exhibits limited ability to understand the instructions. Some approaches further leverage the more advanced MLLM to better comprehend instructions. However, they still mainly focus on understanding instructions at the textual level, while neglecting to capture the rich visual semantics of the image, therefore showing weak performance in complex instruction following and maintaining background consistency. + +To address the dataset issue, we propose an automated data construction pipeline to generate editing pairs with complex instructions and strong background consistency. Leveraging the perceptive capabilities of MLLM, our approach extracts detailed object information from high-resolution images and utilizes an advanced mask-based editing model to produce realistic and controllable edits. Using this method, we introduce AdvancedEdit Dataset, a comprehensive collection comprising over 2,500,000 editing pairs with high visual quality, complex instructions, and good background consistency. + +To address the lack of rich image condition, we employ a two-stream bridging mechanism to integrate both high-level textual and rich visual information into the denoising process of the diffusion model. This interaction effectively extracts the conditional information, enabling precise and controllable image editing. + +Our method achieves state-of-the-art performance on + +![](images/c7dc7f87eb72e6d6704d242315e121e5d7054de52ef619824a36ab46178c060c.jpg) + +![](images/6bb8b9091b919a393682da3bc13ceb1d64476149ae5b890faf7660723abda5f2.jpg) + +![](images/f4e0946b6093aa9caf613b4753e7328ec9a6f513da4daf7dbaf6117b4566b101.jpg) +"Locate the green items suspended from above and substitute them with pendant lighting." + +![](images/981f049a89af991e806e9c391b21c230d13681ceabf01aecbb72b865e68bfa83.jpg) +"Find the white container and change it to a wooden one." + +![](images/c127acf9a33ba1e2342e9cb929b6b86d8437c5997837e3bc203c33efc6e71c10.jpg) + +![](images/d70b1edb69ce5b9248e612f19bf6c5de34f27cef5a200a18fecbcf5e3fc548c9.jpg) +"Please locate the tallest building and add a clock tower in front of it." +Figure 1. We propose InsightEdit, an end-to-end instruction-based image editing model, trained on high-quality data and designed to fully harness the capabilities of Multimodal Large Language Models (MLLM), achieving high-quality edits with strong instruction-following and background consistency. + +
DatasetEditing PairsResolutionGood Background ConsistencyComplex InstructionsHigh Visual QualityAutomated Pipeline
InstructPix2Pix [2]313,0105122XXX✓
MagicBrush [39]10,38810242✓X✓X
HIVE [42]1,100,000+5122XXX✓
FaithfulEdit [6]52,20810242XX✓✓
UltraEdit [43]4,108,2625122✓✓X✓
HQEdit [15]187,3509002XX✓✓
EditWorld [35]10,000+5122X✓✓✓
AdvancedEdit (Ours)2,536,67410242✓✓✓✓
+ +Table 1. Comparison between previous datasets and ours. We provide more detailed comparisons in the appendix. + +both the Reason-Edit [14] and AdvancedEdit-Eval, demonstrating its strong capability in complex instruction following and maintaining background consistency. + +In summary, our contributions are as follows: + +1. We propose an automated data construction pipeline for image editing, specifically engineered to facilitate the training of models for complex instruction-based image editing tasks. +2. We produce the AdvancedEdit Dataset, a large-scale, high-quality, and fine-grained image editing dataset with complex instructions and good background consistency. Additionally, we introduce the AdvancedEdit-Eval dataset to assess the model's ability to handle complex instructions. +3. We present InsightEdit, which utilizes a two-stream bridging mechanism to guide the image editing pro + +cess with both textual and image features reasoned by MLLM. Extensive experiments are conducted to prove the model's ability to follow complex instructions and maintain high background consistency. + +# 2. Related Work + +# 2.1. Image Editing Methods + +Recent advancements in image editing can be broadly categorized into mask-based and mask-free approaches. Mask-based image editing [1, 31-33] requires the original image, instructions, and mask image indicating the regions to be edited as input. The mask image typically needs to be manually specified by the user. Blended Latent Diffusion [1] extends Latent Diffusion Model [28] by integrating them into a localized image editing framework, enabling the tar + +geted redrawing of specific regions during the denoising process. BrushNet [16] is inspired by the architecture of ControlNet [40]. It employs a dual UNet [29] structure, where an additional branch is used to enhance feature extraction from the masked regions. Power-Paint [44] focuses on the distinct characteristics of various image editing tasks by learning a token to classify different editing tasks. Mask-based approaches enable localized redrawing and fine-grained control, resulting in a stable and high-visual-quality output. However, they require additional mask information and are sensitive to mask shape and the specific editing task. + +Mask-free methods [2, 9, 14, 19] present a promising alternative by reducing the dependency on explicit masks and allowing for more flexible and intuitive image editing. InstructPix2Pix [2] utilizes GPT-3 [25] and a Prompt2Prompt [10] methodology to construct an image editing dataset. InstructDiffusion [9] builds upon the network design of InstructPix2Pix, aiming to unify various vision tasks through joint training. MGIE [19] enhances instruction-based image editing by learning to generate expressive instructions with MLLM. SmartEdit [14] utilizes MLLM to generate text embeddings, leveraging the MLLM's advanced reasoning and comprehension capabilities. + +Although instruction-based image editing methods do not require explicit mask guidance, making them more intuitive and convenient, they often yield lower image quality than mask-based approaches. They also face challenges in following complex instructions and maintaining background consistency. + +# 2.2. Image Editing Datasets + +Several works have recently introduced image editing datasets. InstructPix2Pix [2] introduces a large-scale image editing dataset created through fine-tuned GPT-3 and a Prompt-to-Prompt methodology integrated with Stable Diffusion. MagicBrush [39] further provides a limited-scale, manually annotated dataset for instruction-guided real image editing. Subsequent work, HIVE [42] introduces more training triplets and human ranking results to provide stronger supervisory signals for improved model training. FaithfulEdits [6] employed inpainting techniques followed by a filtering process using Visual Question Answering (VQA) models. UltraEdit [43] explores region-based data generation, employing mask generation to facilitate object redrawing in specified areas. Some works explore the diversity of instructions, such as EditWorld [35], which defines and categorizes instructions grounded in various world scenarios. + +As shown in Table 1, despite the notable contributions of these datasets, current image editing datasets still face several challenges, including the inability to scale the data con + +struction process, low dataset resolution, poor background consistency, and a lack of fine-grained instructions for complex understanding and reasoning tasks. + +# 3. Dataset Construction + +# 3.1. Automated Pipeline + +We propose an automated data construction pipeline focused on generating high-fidelity, fine-grained image-editing pairs with detailed instructions that demonstrate advanced reasoning and understanding. We categorize the image editing tasks into three types: removal, addition, and replacement. Figure 2 presents our data preparation workflow. + +Step 1: Caption & Object Extraction. We leverage the advanced comprehension capabilities of MLLM to generate a global caption that effectively conveys the image's content. Using this caption, we utilize LLM to extract a JSON list of objects, identifying those with physical significance. Each object is defined by a simple caption (e.g., blue T-shirt) and a detailed description (e.g., A shirt worn by the man, typically made of cotton, featuring a blue color). + +Step 2: Mask Generation. For mask generation, we apply GroundedSAM [27] to extract local masks for each object, filtering out low-confidence masks based on a predefined threshold, thus obtaining accurate object and mask pairs for further processing. + +Step 3: Editing Pair Construction. As mentioned in Section 2.2, mask-based image editing models exhibit superior image generation capabilities, offering better control over the generation process for specific tasks. In the process of constructing the edited images, we employed the state-of-the-art mask-based methods [16, 44] to generate the target image, enabling the creation of more fine-grained and controllable image editing pairs. We categorized the image editing generation tasks into three types: removal, addition, and replacement. + +For removal, we apply a mask-based approach, using both the original image and the object mask to generate the post-removal image, simultaneously generating and storing the template instruction: "remove the [object]." For addition, we swap the source and target images. The template instruction is: "add the [object]." + +For the replacement task, we leverage MLLM's creative capabilities to propose an alternative object for the scene, often resulting in plausible and innovative outcomes. Simultaneously, we generate and store the template instruction: "replace the [source object] with the [target object].". + +Step 4: Instruction Recaptioning. To diversify the complexity of the instructions, we recaption the instructions into both simple and advanced versions. The simple instructions are generated through synonym replacement and task template modifications. For instance, the instruction "replace + +![](images/8c589f15fece6d7542f2e747c02b17c50ce8144eb60512fabdfef654b9d079b5.jpg) +Figure 2. The overall data construction pipeline. (1) Captioning & Object Extraction: Utilizing VLM to generate a global caption from the source image, and further get an object JSON list contains both simple caption and detailed caption. (2) Mask Generation: Utilizing GroundedSAM to obtain the corresponding mask of each object. (3) Editing Pair Construction: Utilizing mask-based image editing model to construct target image and templated instruction. (4) Instruction Recaptioning: Utilizing VLM to rewrite instruction to gain diverse instructions. (5) Quality Evaluation: Filtering the datasets using VIEScore. + +![](images/0c283059f42ef96137bf57d8da4975ef7a356feb0d6ed2142a8e1c89f9609e29.jpg) + +![](images/fdc20728baf73486c2f36cdfc3f2efbe925edef9bf84d97e42a78db9772ed462.jpg) + +the [acoustic guitar] with [white guitar]" is rewritten as "Swap the [acoustic guitar] for the [white guitar]." + +In contrast, the advanced instructions are designed to increase complexity in two key ways. First, we replace simple object descriptions with detailed ones prepared in Step 1 and apply different task templates. Second, we recaption instructions to gain reasoning capabilities, such as "What is the man holding in his hands? Please change the color of this object to white." This approach challenges the editing model to generate images with more intricate details and greater precision. + +Step 5: Quality Evaluation. To construct a high-quality dataset, we assess and filter the image-editing pairs using VIEScore [20], which aligns evaluations with human preferences via MLLM. The evaluation framework consists of two components: semantic consistency, which measures instruction following and evaluates whether the image has been over-edited, and perceptual quality, which evaluates image fidelity, including the presence of artifacts or blurriness. The rating example of VIEScore can be seen in Figure 2. + +# 3.2. AdvancedEdit Dataset + +We choose Pexels1, a high-quality real-world photographic image dataset, as our source data. Pexels is a popular platform offering a large collection of high-resolution, royalty-free images contributed by photographers around the world. The images in the Pexels dataset generally exhibit an average resolution of approximately 2K or higher. This dataset contains over 1 million photographic images and encom- + +passes a broad range of subjects, covering various world scenes. Given that each image contains multiple objects and is associated with various instruction tasks, our dataset comprises a total of 2,536,674 editing pairs. A more detailed analysis of the proposed dataset is provided in the appendix. + +We define the image editing pairs with simple instructions as SimpleEdit, and image editing pairs with complex instructions as AdvancedEdit. We will discuss the influence of instructions with varying complexity in Section 5.3. Additionally, we introduce the AdvancedEdit-Eval, comprising 300 curated image pairs, covering a range of tasks including removal, addition, and replacement. The evaluation dataset includes a variety of intricate and nuanced image editing scenarios, necessitating that the model possesses a certain level of understanding and reasoning capabilities. + +# 4. Method + +Overview. The overall architecture of InsightEdit is depicted in Figure 3. It mainly consists of a comprehension module, a bridging module, and a generation module. Specifically, the comprehension module leverages MLLM to comprehend the image editing task; the bridging module integrates both text and image features into the denoising process of the diffusion model; and the generation module receives editing guidance via the diffusion model to generate the target image. + +# 4.1. Comprehension Module + +Multimodal large language model receives the original image and the editing instruction to comprehend the image editing task. In the comprehension module, we use LLaVA + +![](images/053b18cac162e469609c292dc3ff0d9a7fc2f48c4136078261aca143b64928c4.jpg) +Figure 3. The overall architecture of InsightEdit. It mainly consists of three parts: (1) Comprehension Module: A comprehension module that leverages MLLM to perceive and comprehend the image editing task; (2) Bridging Module: A bridging module that better interacts and extracts both the textual and image features; (3) Generation Module: A generation module that receives editing guidance via diffusion model to generate the target image. + +![](images/91b662947536a85858a387af89f98f6e94c430739aed7cc78ca62dd3b13b0e8d.jpg) + +7B [24] as our vision-language foundation model. The feature of source image $\mathbf{I}_{src}$ is extracted by a vision encoder $E_{\phi}(\cdot)$ and the text instruction is tokenized as text embeddings $c$ . Aligned by a fully connect layer, the original image features are sent into the LLaVA decoder [30] along with text embeddings. The mentioned process is represented as: + +$$ +v = \operatorname {F C} \left(E _ {\phi} \left(\mathbf {I} _ {s r c}\right)\right), \tag {1} +$$ + +$$ +h = \mathrm {L L a V A} _ {\omega} (v, c). +$$ + +Inspired by GILL [19], we extend the vocabulary of the LLM by introducing $r$ special [MM] tokens, where "MM" indicates multi-modality comprehension of text and image information via MLLM. These tokens are appended to the end of the instruction $c$ , and the model is trained by minimizing the negative log-likelihood of predicting [MM] tokens, conditioned on the previously generated tokens. The loss function can be represented below: + +$$ +L _ {\mathrm {L L M}} (c) = - \sum_ {i = 1} ^ {r} \log p _ {\left\{\omega \cup \theta \right\}} ([ \mathbf {M M} _ {i} ] \mid v, c, \tag {2} +$$ + +$$ +[ \mathbf {M M} _ {1} ], \dots , [ \mathbf {M M} _ {i - 1} ]). +$$ + +The majority of the LLM parameters are kept frozen, while the LoRA [12] is employed to facilitate efficient training. + +# 4.2. Bridging Module + +The textual features provide high-level information while the image features provide more detailed condition. Therefore, in the bridging module, to comprehensively integrate both textual and image features as the conditions, we design a dual-stream condition alignment methodology. + +Textual Branch via Q-Former & BIM. For the textual branch, similar to SmartEdit [14], we start from the text-aligned Q-Former [5, 22] which has primarily aligned with the original clip text embedding space. Then, given the learnable tokens $q$ as inputs, we propose to further extract the reasoned textual information from the [MM] token hidden states $h$ via cross-attention: + +$$ +q ^ {\prime} = Q _ {\beta} (q, h), \tag {3} +$$ + +where $q^{\prime}$ represents the Q-former outputs that mainly encode the text instruction information. + +Additionally, following SmartEdit [14], we employ the BIM module that enables bidirectional information exchange between the input image and the MLLM output: + +$$ +f _ {t x t}, v _ {t x t} = \operatorname {B I M} \left(E _ {\phi} \left(\mathbf {I} _ {s r c}\right), q ^ {\prime}\right), \tag {4} +$$ + +where $f_{txt}$ and $v_{txt}$ are the interacted features obtained from the BIM module. $f_{txt}$ is used as the final textual condition to UNet and $v_{txt}$ is the textual-aware vision features which is later added with U-Net inputs. + +Image Branch via IAA. For the image branch, we propose an Image Alignment Adapter (IAA) module to activate the image information reasoned by MLLM, which consists of a mapper, a linear layer, and a layer normalization. + +In detail, the mapper is a Multi-Layer Perceptron (MLP) that transforms the [MM] token hidden states $h \in \mathbb{R}^{r \times 4096}$ into a single embedding $\operatorname{Mapper}(h) \in \mathbb{R}^{1 \times 768}$ , which is then aligned with the target image features extracted by a CLIP vision encoder. Intuitively, the target image contains direct information about the editing goal, e.g., the expected background, therefore can provide explicit supervision. After that, the linear is further employed to expand + +Table 2. Quantitative comparison on AdvancedEdit-Eval. + +
MethodsVIEScore↑CLIPScore↑PSNR↑SSIM↑LPIPS↓
InstructPix2Pix0.34219.52820.1920.6940.182
MagicBrush0.35219.75122.6360.7430.132
InstructDiffusion0.31819.39018.0250.6240.223
MGIE0.36119.04720.0740.6910.199
SmartEdit-7B0.68220.11420.1150.6510.131
InsightEdit0.73820.39521.2670.6750.112
InsightEdit with AdvancedEdit0.83121.00222.8710.7160.071
+ +the $\operatorname{Mapper}(h) \in \mathbb{R}^{1 \times 768}$ back to token sequences $\mathbb{R}^{N \times 768}$ ( $N$ is set as 4 by default) followed by a layer normalization. Formally, the mentioned process is represented as: + +$$ +\mathcal {L} _ {\mathrm {I A A}} = \left\| \operatorname {C L I P} \left(\mathbf {I} _ {\text {t a r}}\right) - \operatorname {M a p p e r} (h) \right\| _ {2} ^ {2}, \tag {5} +$$ + +$$ +f _ {i m g} = \operatorname {I A A} (h), +$$ + +where $\mathcal{L}_{\mathrm{IAA}}$ is the supervision from the target image, and $f_{img}$ represents the output of the IAA, which will be sent to further cross-attention operation in U-Net. + +Compared to textual features, the image features contain more detailed and comprehensive information, which can guide image editing more explicitly. + +# 4.3. Generation Module + +In the process of target image generation, inspired by IP-adapter [37], we apply a decoupled cross-attention mechanism to integrate both textual and image features. An additional cross-attention layer is employed on top of the cross-attention layer in the original UNet block to interact with the image features. As the text cross-attention and image cross-attention are detached, we can also adjust the weight of the image condition in the inference stage: + +$$ +\begin{array}{l} \mathbf {Z} = \text {A t t e n t i o n} (\mathbf {Q}, \mathbf {K} _ {\mathrm {t x t}}, \mathbf {V} _ {\mathrm {t x t}}) \tag {6} \\ + \lambda \cdot \text {A t t e n t i o n} (\mathbf {Q}, \mathbf {K} _ {\mathrm {i m g}}, \mathbf {V} _ {\mathrm {i m g}}), \\ \end{array} +$$ + +where $\mathbf{K}_{\mathrm{img}}$ and $\mathbf{V}_{\mathrm{img}}$ are projected by $f_{img}$ , $\mathbf{K}_{\mathrm{txt}}$ and $\mathbf{V}_{\mathrm{txt}}$ are projected by $f_{txt}$ . + +For the diffusion model, we concatenate the encoded image latent $\mathcal{E}(\mathbf{I}_{src})$ with the noisy latent $z_{t}$ . The process can be formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {d i f f}} = \mathbb {E} _ {\mathcal {E} (y), \mathcal {E} (x), c _ {T}, \epsilon \sim \mathcal {N} (0, 1), t} \\ \left[ \| \epsilon - \epsilon_ {\delta} (t, \operatorname {c o n c a t} [ z _ {t}, \mathcal {E} (x) ] + v _ {t}, f _ {t x t}, f _ {i m g}) \| _ {2} ^ {2} \right]. \tag {7} \\ \end{array} +$$ + +# 5. Results + +# 5.1. Experiment Settings + +Implementation Details. To train InsightEdit, we use CC12M [7], InstructPix2Pix [2], MagicBrush [39], + +COCO [23], RefCOCO, GRefCOCO [17, 38], CO-COStuff [3], LISA [21, 36], ReasonEdit [14] and our proposed AdvancedEdit. Due to the limited resources, we only used 202, 822 editing pairs from AdvancedEdit for the experiments. + +InsightEdit is trained in three stages, as detailed in the Appendix, and the training is conducted on a single machine with 8 H100 GPUs. We adopt the AdamW optimizer [18] as the optimizer during three stages. In the first training stage, the learning rate is $2e^{-4}$ and weight decay is 0. In the second stage and the third stage, the values of learning rate, weight decay, and warm-up ratio were set to $1e^{-5}$ , 0, and 0.001, respectively. In the data construction pipeline, the MLLM we utilize is GPT-4o $^2$ , while the LLM applied is Qwen2 [34]. + +Metrics. For evaluating the background consistency, we use PSNR, SSIM, and LPIPS [11, 41]. For the edited area, we calculate the CLIPScore [26] by comparing the edited object with its ground truth (GT) label. To more accurately evaluate editing effects and align with human preferences, we employ VIEScore [20], which utilizes MLLM to evaluate editing performance from two perspectives: instruction-following and background consistency. We emphasize that VIEScore might better align with human preferences, making it a more convincing measure [20]. + +# 5.2. Comparison with State-of-the-art + +Results on AdvancedEdit-Eval. We compare our method with existing state-of-the-art instruction-based image methods, including instructPix2Pix, MagicBrush, InstructDiffusion, MGIE, and SmartEdit. First, we conduct comparative experiments on the AdvancedEdit-Eval as introduced in Section 3.2. The quantitative results presented in Table 2 show that InsightEdit shows good performance across all five metrics. The experimental results in PSNR and LPIPS scores prove that InsightEdit excels in preserving non-edited regions, while the enhanced CLIPScore validates our method's superior ability to follow instructions with greater accuracy. Specifically, in terms of VIEScore, both InsightEdit itself and InsightEdit trained on the AdvancedEdit dataset achieve state-of-the-art results, demonstrating that the model's generated outputs are preferred by + +Table 3. Quantitative comparison on Reason-Edit. All the methods we compared have been fine-tuned using the same training data as that used by SmartEdit. + +
MethodsUnderstanding ScenariosReasoning Scenarios
VIEScore↑CLIPScore↑PSNR↑SSIM↑LPIPS↓VIEScore↑CLIPScore↑PSNR↑SSIM↑LPIPS↓
InstructPix2Pix0.49322.76221.5760.7210.0890.54819.41324.2340.7070.083
MagicBrush0.42222.62018.1200.6800.1430.53019.75522.1010.6940.113
InstructDiffusion0.45723.08023.2580.7430.0670.67019.52321.4530.6660.117
MGIE0.37321.94716.0940.6950.2000.54018.03722.9770.7430.144
SmartEdit-7B0.86623.61122.0490.7310.0870.83520.95025.2580.7420.055
InsightEdit0.90123.73423.5880.7490.0670.89320.86725.7120.7470.049
InsightEdit with AdvancedEdit0.93424.42124.5030.7600.0540.94721.14126.0900.7500.047
+ +![](images/8996cd85997d748b8cf9b7dd045aa3059304a60676708cea9bf74293e4ad052b.jpg) +Source Image + +![](images/715447c62efa55a4b9e9a38ff7c2823d9815558b62fd2a208eaa17029f62b351.jpg) +InstructPixPix +"Identify the the cinnamon sticks next to the bowl and remove them from the wooden surface." + +![](images/3c40d5175cea8353a97b1f27429a65c944c7c27dc0504da14995139e90d1b527.jpg) +MagicBrush InstructDiffusion + +![](images/bab0c5aa878c33c95b4c3cfb3c07633c9c9983310819c72b594df6f8ae538bf3.jpg) + +![](images/0b9cbe9bf00783a3f256293406a22974eee2985b44860717c5173ace1a08c1ad.jpg) +n MGIE + +![](images/46e689f02cfe57e29e6fa82763f86f797869da10dbe17e9f7262bcb39bd4ce8e.jpg) +SmartEdit + +![](images/600ae3f3c78d2af20480a9f8b79e7addab8c285b6d505c15deb808b980513db2.jpg) +InsightEdit + +![](images/1d97cef3c3a776c4b41ce8efc2024eb70255d1e476859390e91022aaa4f0e755.jpg) + +![](images/14075f120499dbb64bf154579a77b1d342977872642af1999d9766004ec1fd27.jpg) +"Identify the colorful flowers in the scene and modify them to a vibrant yellow hue." + +![](images/f06ec95ce02751ea04e754fb529972f177d8da2af06598b80c703f4bfff33383.jpg) + +![](images/0761d918299cba6a7d4c149a0a743e53b99d579c4c1c246078ac1afb72a06612.jpg) + +![](images/215ed51fa28efdd2c9526463f74df1ed31b61c46c47edd328bda4a4348bc5553.jpg) + +![](images/61a3a414cbdd1de3b55f9e8ec48a64f838b30252a96612361a5ad54fd3ddec74.jpg) + +![](images/265427873a64724e3d8242cd07b04c25083a1213766927cc7663ea4ff55f3a33.jpg) + +![](images/f9e8b40a317fc2dbddd4f80b1bbe3eef969afd4e6538769464eedede098aaa1e.jpg) + +![](images/6f3b8e4f4a6614cc07396e92208efcc2bc1fd6edf06dd7b2ccc190d9db3bf2e0.jpg) + +![](images/be2134f83ee258b5da9f97ac71dc14995fa8c62a83c1cb403dd0c703e212c632.jpg) + +![](images/095d63cead02e28cc4726abac328694283939853f1cc3978743141f168b3e49c.jpg) + +![](images/e0688527964835f4ed51d8a5035219f08e5388b1aff6ce0dd0698d0db017b7da.jpg) + +![](images/135a8847e1a85bcb9de5ecf311fc26d19e78c69effca41b67e2c6e4af7226054.jpg) + +![](images/efb725dfe94c902325446d147b1d1036c0137780664586e1e8bf3611c38facf2.jpg) + +![](images/fe75ee3ce9a3ece48c1d3ad701613a65e9b613d4b6c905d5432f90a9a0777399.jpg) + +![](images/32da01294b3d084a381d2de8f71c631616b376ab2f100a02935037693c965905.jpg) +"Add an ornate, large crystal chandelier hanging from the center of the ceiling." +"There is a person sitting in the center of the image, remove it." + +![](images/6c7b6ae1f9b5417e6229f74a460a1f1e953192f2280e3468f54c320a2e1f3fbc.jpg) + +![](images/f60cdce23972f672d834616bbc61335910a5e034dcd3d81b866be4ef563db57a.jpg) + +![](images/5292e0289c63c2d2c92461b9ebcf5a7ad1c94becbb4dbc1cfb5edb28bbdadbd1.jpg) + +![](images/cf1adffa360ca40c42e0ade0085d54d50e9885b34c43b79ecb1f425218fb1148.jpg) + +![](images/36f2bc973f8a8d3bcd909e31ebdd8887dcf496f0ef9a3f86ce053cf24b1068be.jpg) + +![](images/8725f5438b4b4f9fc8b71cdf3b350b9d447b69beba70881cca3da54d569e3523.jpg) +Figure 4. Qualitative comparison on AdvancedEdit. InsightEdit shows superior instruction following and background consistency capability. + +![](images/8e9edd03326b5583b2b259a5250db061dbd747c80d5b1eb0002ab92e5a5ab046.jpg) +"Please observe if there's a house in the image and change its appearance to a cozy cabin with a warm, inviting glow from within." + +![](images/1eeb5fd9d19789ef53f81ec70fe8efa93ed53cb048b190b863708e84866eb423.jpg) + +![](images/5c8247aac02ad0619c71ebe6795670748be763529ca4d834171a04cae35faccb.jpg) + +![](images/57f914d0468f0e00ff9340ea949384ac5fcf54854ad4b5656e03d378cb38be5d.jpg) + +![](images/a6362222ae5c33dc04ac8a696f3995e2f13bf28b25fc109ca2920e10cb4bd42d.jpg) + +![](images/1c8f9d7fc05eb30d1c630039b9379e75ca1196d324b71d04865af4c859719a6d.jpg) + +humans over other methods, whether regarding background preservation or the ability to follow complex instructions. + +Figure 4 illustrates the qualitative results. As shown in the last column, our method trained on the AdvancedEdit dataset demonstrates a clear advantage over competing approaches. As can be seen, other methods struggle to maintain background consistency with the original image, often introducing artifacts or unintended alterations, while In + +sightEdit provides specific editing on the required regions. In the meanwhile, InsightEdit performs strong instruction following capabilities, having a great understanding of spatial, color, number, and detailed objects. + +Results on Reason-Edit. Additionally, we conducted comparative experiments on the Reason-Edit[14], which is specifically designed to assess editing abilities with an emphasis on understanding and reasoning across various sce + +
IAAPSNR↑SSIM↑LPIPS↓CLIPScore↑VIEScore↑
22.3480.6920.09520.6527.307
✓22.8710.7160.07121.0027.545
+ +![](images/d90bdf94379e4ddbae2a55307eb0d604bd7bb68637d751bb1b7cd42b79e71b81.jpg) +"Could you swap the white flowers with a diverse collection of colorful wildflowers, creating a vivid and colorful foreground?" + +![](images/7e93fbd28f48acfa79239cac18a2c0a58b82993a781137cd93f1d8cf6ddb738d.jpg) +Transform the worn wooden table into a sleek, modern marble countertop while maintaining the arrangement of the objects. + +narios (i.e. color, position, mirror, multiple-objects, reason, and size). The results in Table 3 show that InsightEdit outperforms other metrics, except for a slight CLIP score difference in reasoning tasks, where it slightly trails SmartEdit. Furthermore, when trained on our AdvancedEdit dataset, InsightEdit demonstrates improvements across all metrics, highlighting the model's ability to leverage the dataset to enhance both comprehension and reasoning capabilities, while also emphasizing the intrinsic value of the data used. + +# 5.3. Ablation Studies + +Ablation Studies on IAA. To validate the effectiveness of IAA, we conduct comparative experiments on InsightEdit with the AdvancedEdit dataset. Table 4 shows the quantitative results, it can be seen that, with the IAA module, InsightEdit leads to notable improvements across various evaluation metrics. The qualitative results can be seen in Figure 5. In the first example, without the IAA module, the house is erroneously transformed into wildflowers, resulting in poor background consistency due to the absence of the red house. In contrast, with the IAA module, the red house is preserved accurately, maintaining consistency with the source image. In the second example, the marble table is edited poorly without the IAA module, while the addition of the IAA module leads to a more natural edit that aligns well with the instruction. + +Ablation Studies on Instruction Complexity. We investigate the impact of instruction complexity and granularity on model performance by comparing two datasets introduced in Section 3.1: SimpleEdit Dataset and AdvancedEdit Dataset. Table 5 demonstrates that incorporating the SimpleEdit Dataset alone enhances the model's per + +Table 4. Ablation of IAA. Our IAA module performs best. + +
SimpleEdit AdvancedEditPSNR↑SSIM↑LPIPS↓CLIPScore↑VIEScore↑
21.2670.6750.11220.3956.974
✓22.2600.7080.08020.5577.213
✓22.8710.7160.07121.0027.545
+ +Table 5. Ablation of AdvancedEdit. Our method with AdvancedEdit performs best. + +![](images/921b9dcd1015ad834c3d0fdbb40d162a1758aeaec682dc409a59363cbcec7d9a.jpg) +Figure 5. Demonstration of the effectiveness of the IAA module. +"Can you identify the specific bright light source from the small window on the staircase wall, remove it smoothly and blend the area to match the surrounding tones?" +Figure 6. Demonstration of the effectiveness of AdvancedEdit. + +formance, primarily due to the high-quality outputs generated by our mask-based image editing approach. However, the inclusion of the AdvancedEdit Dataset leads to substantial improvements, particularly in CLIPScore and VI-EScore, highlighting the importance of increased instruction detail and the value of integrating complex instructions. The qualitative comparison results, as illustrated in Figure 6, reveal that without any data augmentation, the method struggles to understand the instructions and produce effective edits in both two examples. When the SimpleEdit dataset is introduced, there is some improvement, though the edit quality remains suboptimal. However, with the use of the AdvancedEdit dataset, the method demonstrates the ability to accurately interpret editing instructions, resulting in vivid and precise edits. For instance, in the first example, the method produces finely detailed orange and black butterfly wings, while in the second example, it accurately identifies and removes a light source, maintaining harmony in the surrounding content. These results further show that advanced instruction can handle more complex image editing scenarios and generate more vivid editing results. + +# 6. Discussion and Conclusion + +In conclusion, we present InsightEdit, an end-to-end instruction-based image editing method, utilizing a two-stream bridging mechanism to integrate both the textual and visual features reasoned by the powerful MLLM into the diffusion model to guide the image editing process more precisely. We design an automatic data construction pipeline to generate a large-scale, high-quality, and high-fidelity image editing dataset, considering both image and instruction perspectives. + +# References + +[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. +[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. +[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. +[6] Tuhin Chakrabarty, Kanishk Singh, Arkadiy Saakyan, and Smaranda Muresan. Learning to follow object-centric image editing instructions faithfully. arXiv preprint arXiv:2310.19145, 2023. +[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. +[8] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. +[9] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024. +[10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. +[11] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. +[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. +[13] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525, 2024. +[14] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui + +Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024. +[15] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. +[16] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. +[17] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. +[18] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. +[19] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. Advances in Neural Information Processing Systems, 36, 2024. +[20] Max Ku, Dongfu Jiang, Cong Wei, Xiang Yue, and Wenhu Chen. Viescore: Towards explainable metrics for conditional image synthesis evaluation. arXiv preprint arXiv:2312.14867, 2023. +[21] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. +[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. +[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. +[25] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020. +[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. +[27] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, + +Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. +[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. +[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. +[30] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. +[31] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18359-18369, 2023. +[32] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. +[33] Shaoan Xie, Yang Zhao, Zhisheng Xiao, Kelvin CK Chan, Yandong Li, Yanwu Xu, Kun Zhang, and Tingbo Hou. Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. arXiv preprint arXiv:2312.03771, 2023. +[34] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. +[35] Ling Yang, Bohan Zeng, Jiaming Liu, Hong Li, Minghao Xu, Wentao Zhang, and Shuicheng Yan. Editworld: Simulating world dynamics for instruction-following image editing. arXiv preprint arXiv:2405.14785, 2024. +[36] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. arXiv preprint arXiv:2312.17240, 2023. +[37] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. +[38] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. + +[39] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. +[40] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. +[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. +[42] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9026-9036, 2024. +[43] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. +[44] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. arXiv preprint arXiv:2312.03594, 2023. \ No newline at end of file diff --git a/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/images.zip b/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8eff9416e95bb48a1d0eb27f6917d0266284335a --- /dev/null +++ b/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e35836adfb85cfc850fa69272557eb3beefe8dec766840908c184770a686f951 +size 960835 diff --git a/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/layout.json b/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cdba17ee5d32f7d9625ef7ae0e716126673144c5 --- /dev/null +++ b/2025/InsightEdit_ Towards Better Instruction Following for Image Editing/layout.json @@ -0,0 +1,9575 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 518, + 121 + ], + "type": "text", + "content": "InsightEdit: Towards Better Instruction Following for Image Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "spans": [ + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "type": "text", + "content": "Yingjing Xu" + }, + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "type": "text", + "content": ", Jie Kong" + }, + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "type": "inline_equation", + "content": "^{1*†}" + }, + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "type": "text", + "content": ", Jiazhi Wang" + }, + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "type": "text", + "content": ", Xiao Pan" + }, + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "type": "text", + "content": ", Bo Lin" + }, + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "type": "inline_equation", + "content": "^{2‡}" + }, + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "type": "text", + "content": ", Qiang Liu" + }, + { + "bbox": [ + 122, + 142, + 487, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 233, + 171, + 376, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 171, + 376, + 185 + ], + "spans": [ + { + "bbox": [ + 233, + 171, + 376, + 185 + ], + "type": "text", + "content": "101.AI, Zhejiang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 187, + 517, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 187, + 517, + 200 + ], + "spans": [ + { + "bbox": [ + 88, + 187, + 517, + 200 + ], + "type": "text", + "content": "{poppyxu, xiaopan, rainbowlin}@zju.edu.cn, {kongjie, wangjiazhi, liuqiang}@01.ai" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 201, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 201, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 201, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 251, + 297, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 297, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 297, + 528 + ], + "type": "text", + "content": "In this paper, we focus on the task of instruction-based image editing. Previous works like InstructPix2Pix, Instruct-Diffusion, and SmartEdit have explored end-to-end editing. However, two limitations still remain: First, existing datasets suffer from low resolution, poor background consistency, and overly simplistic instructions. Second, current approaches mainly condition on the text while the rich image information is underexplored, therefore inferior in complex instruction following and maintaining background consistency. Targeting these issues, we first curated the AdvancedEdit dataset using a novel data construction pipeline, formulating a large-scale dataset with high visual quality, complex instructions, and good background consistency. Then, to further inject the rich image information, we introduce a two-stream bridging mechanism utilizing both the textual and visual features reasoned by the powerful Multimodal Large Language Models (MLLM) to guide the image editing process more precisely. Extensive results demonstrate that our approach, InsightEdit, achieves state-of-the-art performance, excelling in complex instruction following and maintaining high background consistency with the original image. The project page is https://poppyxu.github.io/InsightEdit_web/." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 548, + 135, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 135, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 135, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "text", + "content": "Image editing involves altering an image's appearance, structure, or content, encompassing a range of changes from subtle adjustments to major transformations [13]. It has witnessed significant advancements in image editing tasks alongside the development of diffusion models [2, 4, 9, 16, 31], leading to impressive results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 677 + ], + "type": "text", + "content": "Existing works like InstructPix2Pix [2], MagicBrush [39], and InstructDiffusion [9] have investigated end-to-end image editing. Some recent efforts [8, 14]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 228, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 491 + ], + "type": "text", + "content": "explore using MLLM to comprehend instructions, thereby enhancing the ability to cope with complex editing tasks. However, two challenges still remain: (i) Lack of high-quality datasets. As illustrated in Table 1, current datasets suffer from low resolution, poor visual quality and background consistency, and overly simplistic, template-based instructions. Specifically, most datasets rely on Prompt2Prompt [10] method, which struggles to achieve precise control over the generated images and fails to maintain consistency in unedited regions. These limitations hinder the image editing model's capability in both complex instruction following and high-fidelity target image generation. (ii) Lack of rich image condition. Current methods primarily use the CLIP text encoder to provide conditions, yet it often exhibits limited ability to understand the instructions. Some approaches further leverage the more advanced MLLM to better comprehend instructions. However, they still mainly focus on understanding instructions at the textual level, while neglecting to capture the rich visual semantics of the image, therefore showing weak performance in complex instruction following and maintaining background consistency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 493, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 556, + 624 + ], + "type": "text", + "content": "To address the dataset issue, we propose an automated data construction pipeline to generate editing pairs with complex instructions and strong background consistency. Leveraging the perceptive capabilities of MLLM, our approach extracts detailed object information from high-resolution images and utilizes an advanced mask-based editing model to produce realistic and controllable edits. Using this method, we introduce AdvancedEdit Dataset, a comprehensive collection comprising over 2,500,000 editing pairs with high visual quality, complex instructions, and good background consistency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 627, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 556, + 699 + ], + "type": "text", + "content": "To address the lack of rich image condition, we employ a two-stream bridging mechanism to integrate both high-level textual and rich visual information into the denoising process of the diffusion model. This interaction effectively extracts the conditional information, enabling precise and controllable image editing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "Our method achieves state-of-the-art performance on" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 683, + 136, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 136, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 136, + 693 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 693, + 172, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 172, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 172, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 67, + 693, + 172, + 703 + ], + "type": "text", + "content": " Primary Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 703, + 156, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 156, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 156, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 67, + 703, + 156, + 712 + ], + "type": "text", + "content": " Co-Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2694" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 217, + 190 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 217, + 190 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 217, + 190 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 217, + 190 + ], + "type": "image", + "image_path": "c7dc7f87eb72e6d6704d242315e121e5d7054de52ef619824a36ab46178c060c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 199, + 217, + 302 + ], + "blocks": [ + { + "bbox": [ + 57, + 199, + 217, + 302 + ], + "lines": [ + { + "bbox": [ + 57, + 199, + 217, + 302 + ], + "spans": [ + { + "bbox": [ + 57, + 199, + 217, + 302 + ], + "type": "image", + "image_path": "6bb8b9091b919a393682da3bc13ceb1d64476149ae5b890faf7660723abda5f2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 225, + 70, + 386, + 166 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 386, + 166 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 386, + 166 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 386, + 166 + ], + "type": "image", + "image_path": "f4e0946b6093aa9caf613b4753e7328ec9a6f513da4daf7dbaf6117b4566b101.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 168, + 374, + 198 + ], + "lines": [ + { + "bbox": [ + 224, + 168, + 374, + 198 + ], + "spans": [ + { + "bbox": [ + 224, + 168, + 374, + 198 + ], + "type": "text", + "content": "\"Locate the green items suspended from above and substitute them with pendant lighting.\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 199, + 386, + 280 + ], + "blocks": [ + { + "bbox": [ + 225, + 199, + 386, + 280 + ], + "lines": [ + { + "bbox": [ + 225, + 199, + 386, + 280 + ], + "spans": [ + { + "bbox": [ + 225, + 199, + 386, + 280 + ], + "type": "image", + "image_path": "981f049a89af991e806e9c391b21c230d13681ceabf01aecbb72b865e68bfa83.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 282, + 385, + 302 + ], + "lines": [ + { + "bbox": [ + 224, + 282, + 385, + 302 + ], + "spans": [ + { + "bbox": [ + 224, + 282, + 385, + 302 + ], + "type": "text", + "content": "\"Find the white container and change it to a wooden one.\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 392, + 70, + 554, + 190 + ], + "blocks": [ + { + "bbox": [ + 392, + 70, + 554, + 190 + ], + "lines": [ + { + "bbox": [ + 392, + 70, + 554, + 190 + ], + "spans": [ + { + "bbox": [ + 392, + 70, + 554, + 190 + ], + "type": "image", + "image_path": "c127acf9a33ba1e2342e9cb929b6b86d8437c5997837e3bc203c33efc6e71c10.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 392, + 199, + 554, + 280 + ], + "blocks": [ + { + "bbox": [ + 392, + 199, + 554, + 280 + ], + "lines": [ + { + "bbox": [ + 392, + 199, + 554, + 280 + ], + "spans": [ + { + "bbox": [ + 392, + 199, + 554, + 280 + ], + "type": "image", + "image_path": "d70b1edb69ce5b9248e612f19bf6c5de34f27cef5a200a18fecbcf5e3fc548c9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 282, + 553, + 302 + ], + "lines": [ + { + "bbox": [ + 392, + 282, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 392, + 282, + 553, + 302 + ], + "type": "text", + "content": "\"Please locate the tallest building and add a clock tower in front of it.\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 317, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 317, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 555, + 350 + ], + "type": "text", + "content": "Figure 1. We propose InsightEdit, an end-to-end instruction-based image editing model, trained on high-quality data and designed to fully harness the capabilities of Multimodal Large Language Models (MLLM), achieving high-quality edits with strong instruction-following and background consistency." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 99, + 361, + 511, + 480 + ], + "blocks": [ + { + "bbox": [ + 99, + 361, + 511, + 480 + ], + "lines": [ + { + "bbox": [ + 99, + 361, + 511, + 480 + ], + "spans": [ + { + "bbox": [ + 99, + 361, + 511, + 480 + ], + "type": "table", + "html": "
DatasetEditing PairsResolutionGood Background ConsistencyComplex InstructionsHigh Visual QualityAutomated Pipeline
InstructPix2Pix [2]313,0105122XXX✓
MagicBrush [39]10,38810242✓X✓X
HIVE [42]1,100,000+5122XXX✓
FaithfulEdit [6]52,20810242XX✓✓
UltraEdit [43]4,108,2625122✓✓X✓
HQEdit [15]187,3509002XX✓✓
EditWorld [35]10,000+5122X✓✓✓
AdvancedEdit (Ours)2,536,67410242✓✓✓✓
", + "image_path": "ca2eb1ceacedfc5f6adae12775adad1d8bf82533e8a7f0767fa39c0f766de060.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 93, + 481, + 515, + 493 + ], + "lines": [ + { + "bbox": [ + 93, + 481, + 515, + 493 + ], + "spans": [ + { + "bbox": [ + 93, + 481, + 515, + 493 + ], + "type": "text", + "content": "Table 1. Comparison between previous datasets and ours. We provide more detailed comparisons in the appendix." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 514, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 550 + ], + "type": "text", + "content": "both the Reason-Edit [14] and AdvancedEdit-Eval, demonstrating its strong capability in complex instruction following and maintaining background consistency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 554, + 251, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 554, + 251, + 566 + ], + "spans": [ + { + "bbox": [ + 67, + 554, + 251, + 566 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 570, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 617 + ], + "type": "text", + "content": "1. We propose an automated data construction pipeline for image editing, specifically engineered to facilitate the training of models for complex instruction-based image editing tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "content": "2. We produce the AdvancedEdit Dataset, a large-scale, high-quality, and fine-grained image editing dataset with complex instructions and good background consistency. Additionally, we introduce the AdvancedEdit-Eval dataset to assess the model's ability to handle complex instructions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "3. We present InsightEdit, which utilizes a two-stream bridging mechanism to guide the image editing pro" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 514, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 514, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 326, + 514, + 555, + 562 + ], + "type": "text", + "content": "cess with both textual and image features reasoned by MLLM. Extensive experiments are conducted to prove the model's ability to follow complex instructions and maintain high background consistency." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 576, + 400, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 576, + 400, + 589 + ], + "spans": [ + { + "bbox": [ + 314, + 576, + 400, + 589 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 597, + 447, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 447, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 447, + 611 + ], + "type": "text", + "content": "2.1. Image Editing Methods" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Recent advancements in image editing can be broadly categorized into mask-based and mask-free approaches. Mask-based image editing [1, 31-33] requires the original image, instructions, and mask image indicating the regions to be edited as input. The mask image typically needs to be manually specified by the user. Blended Latent Diffusion [1] extends Latent Diffusion Model [28] by integrating them into a localized image editing framework, enabling the tar" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2695" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "geted redrawing of specific regions during the denoising process. BrushNet [16] is inspired by the architecture of ControlNet [40]. It employs a dual UNet [29] structure, where an additional branch is used to enhance feature extraction from the masked regions. Power-Paint [44] focuses on the distinct characteristics of various image editing tasks by learning a token to classify different editing tasks. Mask-based approaches enable localized redrawing and fine-grained control, resulting in a stable and high-visual-quality output. However, they require additional mask information and are sensitive to mask shape and the specific editing task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 294, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 294, + 370 + ], + "type": "text", + "content": "Mask-free methods [2, 9, 14, 19] present a promising alternative by reducing the dependency on explicit masks and allowing for more flexible and intuitive image editing. InstructPix2Pix [2] utilizes GPT-3 [25] and a Prompt2Prompt [10] methodology to construct an image editing dataset. InstructDiffusion [9] builds upon the network design of InstructPix2Pix, aiming to unify various vision tasks through joint training. MGIE [19] enhances instruction-based image editing by learning to generate expressive instructions with MLLM. SmartEdit [14] utilizes MLLM to generate text embeddings, leveraging the MLLM's advanced reasoning and comprehension capabilities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 373, + 294, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 294, + 444 + ], + "type": "text", + "content": "Although instruction-based image editing methods do not require explicit mask guidance, making them more intuitive and convenient, they often yield lower image quality than mask-based approaches. They also face challenges in following complex instructions and maintaining background consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 454, + 187, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 187, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 187, + 467 + ], + "type": "text", + "content": "2.2. Image Editing Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 674 + ], + "type": "text", + "content": "Several works have recently introduced image editing datasets. InstructPix2Pix [2] introduces a large-scale image editing dataset created through fine-tuned GPT-3 and a Prompt-to-Prompt methodology integrated with Stable Diffusion. MagicBrush [39] further provides a limited-scale, manually annotated dataset for instruction-guided real image editing. Subsequent work, HIVE [42] introduces more training triplets and human ranking results to provide stronger supervisory signals for improved model training. FaithfulEdits [6] employed inpainting techniques followed by a filtering process using Visual Question Answering (VQA) models. UltraEdit [43] explores region-based data generation, employing mask generation to facilitate object redrawing in specified areas. Some works explore the diversity of instructions, such as EditWorld [35], which defines and categorizes instructions grounded in various world scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "As shown in Table 1, despite the notable contributions of these datasets, current image editing datasets still face several challenges, including the inability to scale the data con" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "struction process, low dataset resolution, poor background consistency, and a lack of fine-grained instructions for complex understanding and reasoning tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 121, + 440, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 440, + 133 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 440, + 133 + ], + "type": "text", + "content": "3. Dataset Construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 140, + 429, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 140, + 429, + 153 + ], + "spans": [ + { + "bbox": [ + 314, + 140, + 429, + 153 + ], + "type": "text", + "content": "3.1. Automated Pipeline" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 159, + 553, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 553, + 242 + ], + "type": "text", + "content": "We propose an automated data construction pipeline focused on generating high-fidelity, fine-grained image-editing pairs with detailed instructions that demonstrate advanced reasoning and understanding. We categorize the image editing tasks into three types: removal, addition, and replacement. Figure 2 presents our data preparation workflow." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 244, + 553, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 553, + 339 + ], + "type": "text", + "content": "Step 1: Caption & Object Extraction. We leverage the advanced comprehension capabilities of MLLM to generate a global caption that effectively conveys the image's content. Using this caption, we utilize LLM to extract a JSON list of objects, identifying those with physical significance. Each object is defined by a simple caption (e.g., blue T-shirt) and a detailed description (e.g., A shirt worn by the man, typically made of cotton, featuring a blue color)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 341, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 553, + 399 + ], + "type": "text", + "content": "Step 2: Mask Generation. For mask generation, we apply GroundedSAM [27] to extract local masks for each object, filtering out low-confidence masks based on a predefined threshold, thus obtaining accurate object and mask pairs for further processing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 401, + 553, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 553, + 520 + ], + "type": "text", + "content": "Step 3: Editing Pair Construction. As mentioned in Section 2.2, mask-based image editing models exhibit superior image generation capabilities, offering better control over the generation process for specific tasks. In the process of constructing the edited images, we employed the state-of-the-art mask-based methods [16, 44] to generate the target image, enabling the creation of more fine-grained and controllable image editing pairs. We categorized the image editing generation tasks into three types: removal, addition, and replacement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 553, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 592 + ], + "type": "text", + "content": "For removal, we apply a mask-based approach, using both the original image and the object mask to generate the post-removal image, simultaneously generating and storing the template instruction: \"remove the [object].\" For addition, we swap the source and target images. The template instruction is: \"add the [object].\"" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 653 + ], + "type": "text", + "content": "For the replacement task, we leverage MLLM's creative capabilities to propose an alternative object for the scene, often resulting in plausible and innovative outcomes. Simultaneously, we generate and store the template instruction: \"replace the [source object] with the [target object].\"." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Step 4: Instruction Recaptioning. To diversify the complexity of the instructions, we recaption the instructions into both simple and advanced versions. The simple instructions are generated through synonym replacement and task template modifications. For instance, the instruction \"replace" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2696" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 221, + 256 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 221, + 256 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 221, + 256 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 221, + 256 + ], + "type": "image", + "image_path": "8c589f15fece6d7542f2e747c02b17c50ce8144eb60512fabdfef654b9d079b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 256, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 256, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 555, + 312 + ], + "type": "text", + "content": "Figure 2. The overall data construction pipeline. (1) Captioning & Object Extraction: Utilizing VLM to generate a global caption from the source image, and further get an object JSON list contains both simple caption and detailed caption. (2) Mask Generation: Utilizing GroundedSAM to obtain the corresponding mask of each object. (3) Editing Pair Construction: Utilizing mask-based image editing model to construct target image and templated instruction. (4) Instruction Recaptioning: Utilizing VLM to rewrite instruction to gain diverse instructions. (5) Quality Evaluation: Filtering the datasets using VIEScore." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 70, + 378, + 255 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 378, + 255 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 378, + 255 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 378, + 255 + ], + "type": "image", + "image_path": "0c283059f42ef96137bf57d8da4975ef7a356feb0d6ed2142a8e1c89f9609e29.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 378, + 70, + 553, + 255 + ], + "blocks": [ + { + "bbox": [ + 378, + 70, + 553, + 255 + ], + "lines": [ + { + "bbox": [ + 378, + 70, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 378, + 70, + 553, + 255 + ], + "type": "image", + "image_path": "fdc20728baf73486c2f36cdfc3f2efbe925edef9bf84d97e42a78db9772ed462.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 332, + 294, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 294, + 357 + ], + "type": "text", + "content": "the [acoustic guitar] with [white guitar]\" is rewritten as \"Swap the [acoustic guitar] for the [white guitar].\"" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 357, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 464 + ], + "type": "text", + "content": "In contrast, the advanced instructions are designed to increase complexity in two key ways. First, we replace simple object descriptions with detailed ones prepared in Step 1 and apply different task templates. Second, we recaption instructions to gain reasoning capabilities, such as \"What is the man holding in his hands? Please change the color of this object to white.\" This approach challenges the editing model to generate images with more intricate details and greater precision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 464, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 582 + ], + "type": "text", + "content": "Step 5: Quality Evaluation. To construct a high-quality dataset, we assess and filter the image-editing pairs using VIEScore [20], which aligns evaluations with human preferences via MLLM. The evaluation framework consists of two components: semantic consistency, which measures instruction following and evaluates whether the image has been over-edited, and perceptual quality, which evaluates image fidelity, including the presence of artifacts or blurriness. The rating example of VIEScore can be seen in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 592, + 185, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 185, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 185, + 604 + ], + "type": "text", + "content": "3.2. AdvancedEdit Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 610, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 296, + 694 + ], + "type": "text", + "content": "We choose Pexels1, a high-quality real-world photographic image dataset, as our source data. Pexels is a popular platform offering a large collection of high-resolution, royalty-free images contributed by photographers around the world. The images in the Pexels dataset generally exhibit an average resolution of approximately 2K or higher. This dataset contains over 1 million photographic images and encom-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 332, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 392 + ], + "type": "text", + "content": "passes a broad range of subjects, covering various world scenes. Given that each image contains multiple objects and is associated with various instruction tasks, our dataset comprises a total of 2,536,674 editing pairs. A more detailed analysis of the proposed dataset is provided in the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "text", + "content": "We define the image editing pairs with simple instructions as SimpleEdit, and image editing pairs with complex instructions as AdvancedEdit. We will discuss the influence of instructions with varying complexity in Section 5.3. Additionally, we introduce the AdvancedEdit-Eval, comprising 300 curated image pairs, covering a range of tasks including removal, addition, and replacement. The evaluation dataset includes a variety of intricate and nuanced image editing scenarios, necessitating that the model possesses a certain level of understanding and reasoning capabilities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 524, + 370, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 370, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 370, + 536 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 544, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 653 + ], + "type": "text", + "content": "Overview. The overall architecture of InsightEdit is depicted in Figure 3. It mainly consists of a comprehension module, a bridging module, and a generation module. Specifically, the comprehension module leverages MLLM to comprehend the image editing task; the bridging module integrates both text and image features into the denoising process of the diffusion model; and the generation module receives editing guidance via the diffusion model to generate the target image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 661, + 450, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 661, + 450, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 661, + 450, + 673 + ], + "type": "text", + "content": "4.1. Comprehension Module" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Multimodal large language model receives the original image and the editing instruction to comprehend the image editing task. In the comprehension module, we use LLaVA" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 702, + 153, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 153, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 153, + 713 + ], + "type": "text", + "content": "1https://www.pexels.com/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2697" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 362, + 249 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 362, + 249 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 362, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 362, + 249 + ], + "type": "image", + "image_path": "053b18cac162e469609c292dc3ff0d9a7fc2f48c4136078261aca143b64928c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 298 + ], + "type": "text", + "content": "Figure 3. The overall architecture of InsightEdit. It mainly consists of three parts: (1) Comprehension Module: A comprehension module that leverages MLLM to perceive and comprehend the image editing task; (2) Bridging Module: A bridging module that better interacts and extracts both the textual and image features; (3) Generation Module: A generation module that receives editing guidance via diffusion model to generate the target image." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 362, + 71, + 553, + 250 + ], + "blocks": [ + { + "bbox": [ + 362, + 71, + 553, + 250 + ], + "lines": [ + { + "bbox": [ + 362, + 71, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 362, + 71, + 553, + 250 + ], + "type": "image", + "image_path": "91b662947536a85858a387af89f98f6e94c430739aed7cc78ca62dd3b13b0e8d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 317, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 390 + ], + "type": "text", + "content": "7B [24] as our vision-language foundation model. The feature of source image " + }, + { + "bbox": [ + 55, + 317, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{src}" + }, + { + "bbox": [ + 55, + 317, + 296, + 390 + ], + "type": "text", + "content": " is extracted by a vision encoder " + }, + { + "bbox": [ + 55, + 317, + 296, + 390 + ], + "type": "inline_equation", + "content": "E_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 55, + 317, + 296, + 390 + ], + "type": "text", + "content": " and the text instruction is tokenized as text embeddings " + }, + { + "bbox": [ + 55, + 317, + 296, + 390 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 317, + 296, + 390 + ], + "type": "text", + "content": ". Aligned by a fully connect layer, the original image features are sent into the LLaVA decoder [30] along with text embeddings. The mentioned process is represented as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 400, + 294, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 400, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 400, + 294, + 418 + ], + "type": "interline_equation", + "content": "v = \\operatorname {F C} \\left(E _ {\\phi} \\left(\\mathbf {I} _ {s r c}\\right)\\right), \\tag {1}", + "image_path": "499b30f2c62a7a4287417f7782620e5f4b1f0ae74e5bdb5840460407fbe1efc1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 415, + 217, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 415, + 217, + 428 + ], + "spans": [ + { + "bbox": [ + 134, + 415, + 217, + 428 + ], + "type": "interline_equation", + "content": "h = \\mathrm {L L a V A} _ {\\omega} (v, c).", + "image_path": "aa5ff989152954bc1f11552b5edfe5a321d95d87e2a5ae5a044b98ae19cbee58.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "text", + "content": "Inspired by GILL [19], we extend the vocabulary of the LLM by introducing " + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "text", + "content": " special [MM] tokens, where \"MM\" indicates multi-modality comprehension of text and image information via MLLM. These tokens are appended to the end of the instruction " + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "text", + "content": ", and the model is trained by minimizing the negative log-likelihood of predicting [MM] tokens, conditioned on the previously generated tokens. The loss function can be represented below:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 545, + 295, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 545, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 88, + 545, + 295, + 576 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {L L M}} (c) = - \\sum_ {i = 1} ^ {r} \\log p _ {\\left\\{\\omega \\cup \\theta \\right\\}} ([ \\mathbf {M M} _ {i} ] \\mid v, c, \\tag {2}", + "image_path": "d00896caad58305ab3fdbe077cfee975ed13ace8d33d75c6325a2b3975aa9f5c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 578, + 256, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 578, + 256, + 593 + ], + "spans": [ + { + "bbox": [ + 160, + 578, + 256, + 593 + ], + "type": "interline_equation", + "content": "[ \\mathbf {M M} _ {1} ], \\dots , [ \\mathbf {M M} _ {i - 1} ]).", + "image_path": "6d2cb045c10d30d21c3ee580449528d04ef55f7b0fb0291d0a08da1e85f8649e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 602, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 295, + 628 + ], + "type": "text", + "content": "The majority of the LLM parameters are kept frozen, while the LoRA [12] is employed to facilitate efficient training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 637, + 160, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 160, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 160, + 651 + ], + "type": "text", + "content": "4.2. Bridging Module" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "The textual features provide high-level information while the image features provide more detailed condition. Therefore, in the bridging module, to comprehensively integrate both textual and image features as the conditions, we design a dual-stream condition alignment methodology." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 317, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 401 + ], + "type": "text", + "content": "Textual Branch via Q-Former & BIM. For the textual branch, similar to SmartEdit [14], we start from the text-aligned Q-Former [5, 22] which has primarily aligned with the original clip text embedding space. Then, given the learnable tokens " + }, + { + "bbox": [ + 313, + 317, + 555, + 401 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 317, + 555, + 401 + ], + "type": "text", + "content": " as inputs, we propose to further extract the reasoned textual information from the [MM] token hidden states " + }, + { + "bbox": [ + 313, + 317, + 555, + 401 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 317, + 555, + 401 + ], + "type": "text", + "content": " via cross-attention:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 402, + 410, + 553, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 410, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 402, + 410, + 553, + 423 + ], + "type": "interline_equation", + "content": "q ^ {\\prime} = Q _ {\\beta} (q, h), \\tag {3}", + "image_path": "a5caef5cf70ece44efca3e6dad73956cf633f6d9c1731e5ccda1745462ff5d30.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 431, + 553, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 553, + 454 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 431, + 553, + 454 + ], + "type": "inline_equation", + "content": "q^{\\prime}" + }, + { + "bbox": [ + 313, + 431, + 553, + 454 + ], + "type": "text", + "content": " represents the Q-former outputs that mainly encode the text instruction information." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 456, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 492 + ], + "type": "text", + "content": "Additionally, following SmartEdit [14], we employ the BIM module that enables bidirectional information exchange between the input image and the MLLM output:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 367, + 500, + 553, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 500, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 367, + 500, + 553, + 514 + ], + "type": "interline_equation", + "content": "f _ {t x t}, v _ {t x t} = \\operatorname {B I M} \\left(E _ {\\phi} \\left(\\mathbf {I} _ {s r c}\\right), q ^ {\\prime}\\right), \\tag {4}", + "image_path": "abd7de3f085593830791afea689d3234f403bc8dc82b2d6099b5f8b886c64200.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "inline_equation", + "content": "f_{txt}" + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "inline_equation", + "content": "v_{txt}" + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "text", + "content": " are the interacted features obtained from the BIM module. " + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "inline_equation", + "content": "f_{txt}" + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "text", + "content": " is used as the final textual condition to UNet and " + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "inline_equation", + "content": "v_{txt}" + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "text", + "content": " is the textual-aware vision features which is later added with U-Net inputs." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 570, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 617 + ], + "type": "text", + "content": "Image Branch via IAA. For the image branch, we propose an Image Alignment Adapter (IAA) module to activate the image information reasoned by MLLM, which consists of a mapper, a linear layer, and a layer normalization." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "In detail, the mapper is a Multi-Layer Perceptron (MLP) that transforms the [MM] token hidden states " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "h \\in \\mathbb{R}^{r \\times 4096}" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": " into a single embedding " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\operatorname{Mapper}(h) \\in \\mathbb{R}^{1 \\times 768}" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": ", which is then aligned with the target image features extracted by a CLIP vision encoder. Intuitively, the target image contains direct information about the editing goal, e.g., the expected background, therefore can provide explicit supervision. After that, the linear is further employed to expand" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2698" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 130, + 81, + 480, + 168 + ], + "blocks": [ + { + "bbox": [ + 194, + 70, + 415, + 81 + ], + "lines": [ + { + "bbox": [ + 194, + 70, + 415, + 81 + ], + "spans": [ + { + "bbox": [ + 194, + 70, + 415, + 81 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on AdvancedEdit-Eval." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 130, + 81, + 480, + 168 + ], + "lines": [ + { + "bbox": [ + 130, + 81, + 480, + 168 + ], + "spans": [ + { + "bbox": [ + 130, + 81, + 480, + 168 + ], + "type": "table", + "html": "
MethodsVIEScore↑CLIPScore↑PSNR↑SSIM↑LPIPS↓
InstructPix2Pix0.34219.52820.1920.6940.182
MagicBrush0.35219.75122.6360.7430.132
InstructDiffusion0.31819.39018.0250.6240.223
MGIE0.36119.04720.0740.6910.199
SmartEdit-7B0.68220.11420.1150.6510.131
InsightEdit0.73820.39521.2670.6750.112
InsightEdit with AdvancedEdit0.83121.00222.8710.7160.071
", + "image_path": "270da8f0266e1083e661f5dc11abb3abb32d9176503fd9adbd36234ef0e66f28.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 186, + 294, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 294, + 224 + ], + "type": "text", + "content": "the " + }, + { + "bbox": [ + 55, + 186, + 294, + 224 + ], + "type": "inline_equation", + "content": "\\operatorname{Mapper}(h) \\in \\mathbb{R}^{1 \\times 768}" + }, + { + "bbox": [ + 55, + 186, + 294, + 224 + ], + "type": "text", + "content": " back to token sequences " + }, + { + "bbox": [ + 55, + 186, + 294, + 224 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N \\times 768}" + }, + { + "bbox": [ + 55, + 186, + 294, + 224 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 186, + 294, + 224 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 186, + 294, + 224 + ], + "type": "text", + "content": " is set as 4 by default) followed by a layer normalization. Formally, the mentioned process is represented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 232, + 294, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 232, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 95, + 232, + 294, + 253 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {I A A}} = \\left\\| \\operatorname {C L I P} \\left(\\mathbf {I} _ {\\text {t a r}}\\right) - \\operatorname {M a p p e r} (h) \\right\\| _ {2} ^ {2}, \\tag {5}", + "image_path": "a6f4a654b6035368f981d63ebdad25b286c9b8ae1b4630c086762e3a9dc01328.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 251, + 167, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 251, + 167, + 262 + ], + "spans": [ + { + "bbox": [ + 97, + 251, + 167, + 262 + ], + "type": "interline_equation", + "content": "f _ {i m g} = \\operatorname {I A A} (h),", + "image_path": "da7026364fcf2f733073f5252daa3a76e4bd7aaae1fe9e94c1f38aff01116c85.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 272, + 295, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 295, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 272, + 295, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{IAA}}" + }, + { + "bbox": [ + 55, + 272, + 295, + 307 + ], + "type": "text", + "content": " is the supervision from the target image, and " + }, + { + "bbox": [ + 55, + 272, + 295, + 307 + ], + "type": "inline_equation", + "content": "f_{img}" + }, + { + "bbox": [ + 55, + 272, + 295, + 307 + ], + "type": "text", + "content": " represents the output of the IAA, which will be sent to further cross-attention operation in U-Net." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 308, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 344 + ], + "type": "text", + "content": "Compared to textual features, the image features contain more detailed and comprehensive information, which can guide image editing more explicitly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 353, + 170, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 170, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 170, + 363 + ], + "type": "text", + "content": "4.3. Generation Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 368, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 296, + 464 + ], + "type": "text", + "content": "In the process of target image generation, inspired by IP-adapter [37], we apply a decoupled cross-attention mechanism to integrate both textual and image features. An additional cross-attention layer is employed on top of the cross-attention layer in the original UNet block to interact with the image features. As the text cross-attention and image cross-attention are detached, we can also adjust the weight of the image condition in the inference stage:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 93, + 474, + 294, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 474, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 93, + 474, + 294, + 502 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {Z} = \\text {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K} _ {\\mathrm {t x t}}, \\mathbf {V} _ {\\mathrm {t x t}}) \\tag {6} \\\\ + \\lambda \\cdot \\text {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K} _ {\\mathrm {i m g}}, \\mathbf {V} _ {\\mathrm {i m g}}), \\\\ \\end{array}", + "image_path": "55e9ec70e48273c23791df8d2e055559ef0b3406645b8429fe4d1e7b3a005f1d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "text", + "content": " are projected by " + }, + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "inline_equation", + "content": "f_{img}" + }, + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "text", + "content": " are projected by " + }, + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "inline_equation", + "content": "f_{txt}" + }, + { + "bbox": [ + 55, + 511, + 295, + 535 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 536, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 296, + 571 + ], + "type": "text", + "content": "For the diffusion model, we concatenate the encoded image latent " + }, + { + "bbox": [ + 55, + 536, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\mathbf{I}_{src})" + }, + { + "bbox": [ + 55, + 536, + 296, + 571 + ], + "type": "text", + "content": " with the noisy latent " + }, + { + "bbox": [ + 55, + 536, + 296, + 571 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 536, + 296, + 571 + ], + "type": "text", + "content": ". The process can be formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 295, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 633 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {d i f f}} = \\mathbb {E} _ {\\mathcal {E} (y), \\mathcal {E} (x), c _ {T}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\\\ \\left[ \\| \\epsilon - \\epsilon_ {\\delta} (t, \\operatorname {c o n c a t} [ z _ {t}, \\mathcal {E} (x) ] + v _ {t}, f _ {t x t}, f _ {i m g}) \\| _ {2} ^ {2} \\right]. \\tag {7} \\\\ \\end{array}", + "image_path": "45eb30bd8b6e0297e0a21b632fa96643646beda5a26d4425d12a8a97a5d1a2da.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 654, + 108, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 108, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 108, + 665 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 673, + 173, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 173, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 173, + 686 + ], + "type": "text", + "content": "5.1. Experiment Settings" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Implementation Details. To train InsightEdit, we use CC12M [7], InstructPix2Pix [2], MagicBrush [39]," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 187, + 553, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 553, + 247 + ], + "type": "text", + "content": "COCO [23], RefCOCO, GRefCOCO [17, 38], CO-COStuff [3], LISA [21, 36], ReasonEdit [14] and our proposed AdvancedEdit. Due to the limited resources, we only used 202, 822 editing pairs from AdvancedEdit for the experiments." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 248, + 554, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 554, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 554, + 368 + ], + "type": "text", + "content": "InsightEdit is trained in three stages, as detailed in the Appendix, and the training is conducted on a single machine with 8 H100 GPUs. We adopt the AdamW optimizer [18] as the optimizer during three stages. In the first training stage, the learning rate is " + }, + { + "bbox": [ + 313, + 248, + 554, + 368 + ], + "type": "inline_equation", + "content": "2e^{-4}" + }, + { + "bbox": [ + 313, + 248, + 554, + 368 + ], + "type": "text", + "content": " and weight decay is 0. In the second stage and the third stage, the values of learning rate, weight decay, and warm-up ratio were set to " + }, + { + "bbox": [ + 313, + 248, + 554, + 368 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 313, + 248, + 554, + 368 + ], + "type": "text", + "content": ", 0, and 0.001, respectively. In the data construction pipeline, the MLLM we utilize is GPT-4o" + }, + { + "bbox": [ + 313, + 248, + 554, + 368 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 248, + 554, + 368 + ], + "type": "text", + "content": ", while the LLM applied is Qwen2 [34]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 369, + 554, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 489 + ], + "type": "text", + "content": "Metrics. For evaluating the background consistency, we use PSNR, SSIM, and LPIPS [11, 41]. For the edited area, we calculate the CLIPScore [26] by comparing the edited object with its ground truth (GT) label. To more accurately evaluate editing effects and align with human preferences, we employ VIEScore [20], which utilizes MLLM to evaluate editing performance from two perspectives: instruction-following and background consistency. We emphasize that VIEScore might better align with human preferences, making it a more convincing measure [20]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 497, + 493, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 493, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 493, + 510 + ], + "type": "text", + "content": "5.2. Comparison with State-of-the-art" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 514, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 693 + ], + "type": "text", + "content": "Results on AdvancedEdit-Eval. We compare our method with existing state-of-the-art instruction-based image methods, including instructPix2Pix, MagicBrush, InstructDiffusion, MGIE, and SmartEdit. First, we conduct comparative experiments on the AdvancedEdit-Eval as introduced in Section 3.2. The quantitative results presented in Table 2 show that InsightEdit shows good performance across all five metrics. The experimental results in PSNR and LPIPS scores prove that InsightEdit excels in preserving non-edited regions, while the enhanced CLIPScore validates our method's superior ability to follow instructions with greater accuracy. Specifically, in terms of VIEScore, both InsightEdit itself and InsightEdit trained on the AdvancedEdit dataset achieve state-of-the-art results, demonstrating that the model's generated outputs are preferred by" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 394, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 394, + 712 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 394, + 712 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 324, + 702, + 394, + 712 + ], + "type": "text", + "content": "https://openai.com/" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2699" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 92, + 555, + 178 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "type": "text", + "content": "Table 3. Quantitative comparison on Reason-Edit. All the methods we compared have been fine-tuned using the same training data as that used by SmartEdit." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 92, + 555, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 92, + 555, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 92, + 555, + 178 + ], + "type": "table", + "html": "
MethodsUnderstanding ScenariosReasoning Scenarios
VIEScore↑CLIPScore↑PSNR↑SSIM↑LPIPS↓VIEScore↑CLIPScore↑PSNR↑SSIM↑LPIPS↓
InstructPix2Pix0.49322.76221.5760.7210.0890.54819.41324.2340.7070.083
MagicBrush0.42222.62018.1200.6800.1430.53019.75522.1010.6940.113
InstructDiffusion0.45723.08023.2580.7430.0670.67019.52321.4530.6660.117
MGIE0.37321.94716.0940.6950.2000.54018.03722.9770.7430.144
SmartEdit-7B0.86623.61122.0490.7310.0870.83520.95025.2580.7420.055
InsightEdit0.90123.73423.5880.7490.0670.89320.86725.7120.7470.049
InsightEdit with AdvancedEdit0.93424.42124.5030.7600.0540.94721.14126.0900.7500.047
", + "image_path": "19b3a21691411425ff727ad8b174345ea74fdc773294a4c1757df30d51797c90.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 82, + 199, + 144, + 261 + ], + "blocks": [ + { + "bbox": [ + 82, + 189, + 140, + 199 + ], + "lines": [ + { + "bbox": [ + 82, + 189, + 140, + 199 + ], + "spans": [ + { + "bbox": [ + 82, + 189, + 140, + 199 + ], + "type": "text", + "content": "Source Image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 199, + 144, + 261 + ], + "lines": [ + { + "bbox": [ + 82, + 199, + 144, + 261 + ], + "spans": [ + { + "bbox": [ + 82, + 199, + 144, + 261 + ], + "type": "image", + "image_path": "8996cd85997d748b8cf9b7dd045aa3059304a60676708cea9bf74293e4ad052b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 159, + 199, + 221, + 260 + ], + "blocks": [ + { + "bbox": [ + 159, + 189, + 220, + 198 + ], + "lines": [ + { + "bbox": [ + 159, + 189, + 220, + 198 + ], + "spans": [ + { + "bbox": [ + 159, + 189, + 220, + 198 + ], + "type": "text", + "content": "InstructPixPix" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 199, + 221, + 260 + ], + "lines": [ + { + "bbox": [ + 159, + 199, + 221, + 260 + ], + "spans": [ + { + "bbox": [ + 159, + 199, + 221, + 260 + ], + "type": "image", + "image_path": "715447c62efa55a4b9e9a38ff7c2823d9815558b62fd2a208eaa17029f62b351.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 261, + 529, + 271 + ], + "lines": [ + { + "bbox": [ + 159, + 261, + 529, + 271 + ], + "spans": [ + { + "bbox": [ + 159, + 261, + 529, + 271 + ], + "type": "text", + "content": "\"Identify the the cinnamon sticks next to the bowl and remove them from the wooden surface.\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 221, + 199, + 282, + 260 + ], + "blocks": [ + { + "bbox": [ + 225, + 189, + 350, + 199 + ], + "lines": [ + { + "bbox": [ + 225, + 189, + 350, + 199 + ], + "spans": [ + { + "bbox": [ + 225, + 189, + 350, + 199 + ], + "type": "text", + "content": "MagicBrush InstructDiffusion" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 221, + 199, + 282, + 260 + ], + "lines": [ + { + "bbox": [ + 221, + 199, + 282, + 260 + ], + "spans": [ + { + "bbox": [ + 221, + 199, + 282, + 260 + ], + "type": "image", + "image_path": "3c40d5175cea8353a97b1f27429a65c944c7c27dc0504da14995139e90d1b527.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 283, + 199, + 343, + 260 + ], + "blocks": [ + { + "bbox": [ + 283, + 199, + 343, + 260 + ], + "lines": [ + { + "bbox": [ + 283, + 199, + 343, + 260 + ], + "spans": [ + { + "bbox": [ + 283, + 199, + 343, + 260 + ], + "type": "image", + "image_path": "bab0c5aa878c33c95b4c3cfb3c07633c9c9983310819c72b594df6f8ae538bf3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 345, + 199, + 405, + 260 + ], + "blocks": [ + { + "bbox": [ + 345, + 189, + 388, + 198 + ], + "lines": [ + { + "bbox": [ + 345, + 189, + 388, + 198 + ], + "spans": [ + { + "bbox": [ + 345, + 189, + 388, + 198 + ], + "type": "text", + "content": "n MGIE" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 345, + 199, + 405, + 260 + ], + "lines": [ + { + "bbox": [ + 345, + 199, + 405, + 260 + ], + "spans": [ + { + "bbox": [ + 345, + 199, + 405, + 260 + ], + "type": "image", + "image_path": "0b9cbe9bf00783a3f256293406a22974eee2985b44860717c5173ace1a08c1ad.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 406, + 199, + 466, + 260 + ], + "blocks": [ + { + "bbox": [ + 412, + 189, + 459, + 198 + ], + "lines": [ + { + "bbox": [ + 412, + 189, + 459, + 198 + ], + "spans": [ + { + "bbox": [ + 412, + 189, + 459, + 198 + ], + "type": "text", + "content": "SmartEdit" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 199, + 466, + 260 + ], + "lines": [ + { + "bbox": [ + 406, + 199, + 466, + 260 + ], + "spans": [ + { + "bbox": [ + 406, + 199, + 466, + 260 + ], + "type": "image", + "image_path": "46e689f02cfe57e29e6fa82763f86f797869da10dbe17e9f7262bcb39bd4ce8e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 468, + 199, + 529, + 260 + ], + "blocks": [ + { + "bbox": [ + 476, + 189, + 523, + 199 + ], + "lines": [ + { + "bbox": [ + 476, + 189, + 523, + 199 + ], + "spans": [ + { + "bbox": [ + 476, + 189, + 523, + 199 + ], + "type": "text", + "content": "InsightEdit" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 468, + 199, + 529, + 260 + ], + "lines": [ + { + "bbox": [ + 468, + 199, + 529, + 260 + ], + "spans": [ + { + "bbox": [ + 468, + 199, + 529, + 260 + ], + "type": "image", + "image_path": "600ae3f3c78d2af20480a9f8b79e7addab8c285b6d505c15deb808b980513db2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 82, + 275, + 144, + 337 + ], + "blocks": [ + { + "bbox": [ + 82, + 275, + 144, + 337 + ], + "lines": [ + { + "bbox": [ + 82, + 275, + 144, + 337 + ], + "spans": [ + { + "bbox": [ + 82, + 275, + 144, + 337 + ], + "type": "image", + "image_path": "1d97cef3c3a776c4b41ce8efc2024eb70255d1e476859390e91022aaa4f0e755.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 159, + 275, + 220, + 336 + ], + "blocks": [ + { + "bbox": [ + 159, + 275, + 220, + 336 + ], + "lines": [ + { + "bbox": [ + 159, + 275, + 220, + 336 + ], + "spans": [ + { + "bbox": [ + 159, + 275, + 220, + 336 + ], + "type": "image", + "image_path": "14075f120499dbb64bf154579a77b1d342977872642af1999d9766004ec1fd27.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 338, + 500, + 348 + ], + "lines": [ + { + "bbox": [ + 174, + 338, + 500, + 348 + ], + "spans": [ + { + "bbox": [ + 174, + 338, + 500, + 348 + ], + "type": "text", + "content": "\"Identify the colorful flowers in the scene and modify them to a vibrant yellow hue.\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 221, + 275, + 282, + 336 + ], + "blocks": [ + { + "bbox": [ + 221, + 275, + 282, + 336 + ], + "lines": [ + { + "bbox": [ + 221, + 275, + 282, + 336 + ], + "spans": [ + { + "bbox": [ + 221, + 275, + 282, + 336 + ], + "type": "image", + "image_path": "f06ec95ce02751ea04e754fb529972f177d8da2af06598b80c703f4bfff33383.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 283, + 275, + 343, + 336 + ], + "blocks": [ + { + "bbox": [ + 283, + 275, + 343, + 336 + ], + "lines": [ + { + "bbox": [ + 283, + 275, + 343, + 336 + ], + "spans": [ + { + "bbox": [ + 283, + 275, + 343, + 336 + ], + "type": "image", + "image_path": "0761d918299cba6a7d4c149a0a743e53b99d579c4c1c246078ac1afb72a06612.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 345, + 275, + 405, + 336 + ], + "blocks": [ + { + "bbox": [ + 345, + 275, + 405, + 336 + ], + "lines": [ + { + "bbox": [ + 345, + 275, + 405, + 336 + ], + "spans": [ + { + "bbox": [ + 345, + 275, + 405, + 336 + ], + "type": "image", + "image_path": "215ed51fa28efdd2c9526463f74df1ed31b61c46c47edd328bda4a4348bc5553.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 406, + 275, + 466, + 336 + ], + "blocks": [ + { + "bbox": [ + 406, + 275, + 466, + 336 + ], + "lines": [ + { + "bbox": [ + 406, + 275, + 466, + 336 + ], + "spans": [ + { + "bbox": [ + 406, + 275, + 466, + 336 + ], + "type": "image", + "image_path": "61a3a414cbdd1de3b55f9e8ec48a64f838b30252a96612361a5ad54fd3ddec74.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 468, + 275, + 529, + 336 + ], + "blocks": [ + { + "bbox": [ + 468, + 275, + 529, + 336 + ], + "lines": [ + { + "bbox": [ + 468, + 275, + 529, + 336 + ], + "spans": [ + { + "bbox": [ + 468, + 275, + 529, + 336 + ], + "type": "image", + "image_path": "265427873a64724e3d8242cd07b04c25083a1213766927cc7663ea4ff55f3a33.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 82, + 350, + 144, + 412 + ], + "blocks": [ + { + "bbox": [ + 82, + 350, + 144, + 412 + ], + "lines": [ + { + "bbox": [ + 82, + 350, + 144, + 412 + ], + "spans": [ + { + "bbox": [ + 82, + 350, + 144, + 412 + ], + "type": "image", + "image_path": "f9e8b40a317fc2dbddd4f80b1bbe3eef969afd4e6538769464eedede098aaa1e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 159, + 350, + 220, + 412 + ], + "blocks": [ + { + "bbox": [ + 159, + 350, + 220, + 412 + ], + "lines": [ + { + "bbox": [ + 159, + 350, + 220, + 412 + ], + "spans": [ + { + "bbox": [ + 159, + 350, + 220, + 412 + ], + "type": "image", + "image_path": "6f3b8e4f4a6614cc07396e92208efcc2bc1fd6edf06dd7b2ccc190d9db3bf2e0.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 221, + 350, + 282, + 412 + ], + "blocks": [ + { + "bbox": [ + 221, + 350, + 282, + 412 + ], + "lines": [ + { + "bbox": [ + 221, + 350, + 282, + 412 + ], + "spans": [ + { + "bbox": [ + 221, + 350, + 282, + 412 + ], + "type": "image", + "image_path": "be2134f83ee258b5da9f97ac71dc14995fa8c62a83c1cb403dd0c703e212c632.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 283, + 350, + 343, + 412 + ], + "blocks": [ + { + "bbox": [ + 283, + 350, + 343, + 412 + ], + "lines": [ + { + "bbox": [ + 283, + 350, + 343, + 412 + ], + "spans": [ + { + "bbox": [ + 283, + 350, + 343, + 412 + ], + "type": "image", + "image_path": "095d63cead02e28cc4726abac328694283939853f1cc3978743141f168b3e49c.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 345, + 350, + 405, + 412 + ], + "blocks": [ + { + "bbox": [ + 345, + 350, + 405, + 412 + ], + "lines": [ + { + "bbox": [ + 345, + 350, + 405, + 412 + ], + "spans": [ + { + "bbox": [ + 345, + 350, + 405, + 412 + ], + "type": "image", + "image_path": "e0688527964835f4ed51d8a5035219f08e5388b1aff6ce0dd0698d0db017b7da.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 406, + 350, + 466, + 412 + ], + "blocks": [ + { + "bbox": [ + 406, + 350, + 466, + 412 + ], + "lines": [ + { + "bbox": [ + 406, + 350, + 466, + 412 + ], + "spans": [ + { + "bbox": [ + 406, + 350, + 466, + 412 + ], + "type": "image", + "image_path": "135a8847e1a85bcb9de5ecf311fc26d19e78c69effca41b67e2c6e4af7226054.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 468, + 350, + 529, + 412 + ], + "blocks": [ + { + "bbox": [ + 468, + 350, + 529, + 412 + ], + "lines": [ + { + "bbox": [ + 468, + 350, + 529, + 412 + ], + "spans": [ + { + "bbox": [ + 468, + 350, + 529, + 412 + ], + "type": "image", + "image_path": "efb725dfe94c902325446d147b1d1036c0137780664586e1e8bf3611c38facf2.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 82, + 426, + 144, + 488 + ], + "blocks": [ + { + "bbox": [ + 82, + 426, + 144, + 488 + ], + "lines": [ + { + "bbox": [ + 82, + 426, + 144, + 488 + ], + "spans": [ + { + "bbox": [ + 82, + 426, + 144, + 488 + ], + "type": "image", + "image_path": "fe75ee3ce9a3ece48c1d3ad701613a65e9b613d4b6c905d5432f90a9a0777399.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 159, + 426, + 220, + 488 + ], + "blocks": [ + { + "bbox": [ + 182, + 414, + 491, + 425 + ], + "lines": [ + { + "bbox": [ + 182, + 414, + 491, + 425 + ], + "spans": [ + { + "bbox": [ + 182, + 414, + 491, + 425 + ], + "type": "text", + "content": "\"Add an ornate, large crystal chandelier hanging from the center of the ceiling.\"" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 426, + 220, + 488 + ], + "lines": [ + { + "bbox": [ + 159, + 426, + 220, + 488 + ], + "spans": [ + { + "bbox": [ + 159, + 426, + 220, + 488 + ], + "type": "image", + "image_path": "32da01294b3d084a381d2de8f71c631616b376ab2f100a02935037693c965905.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 490, + 456, + 500 + ], + "lines": [ + { + "bbox": [ + 208, + 490, + 456, + 500 + ], + "spans": [ + { + "bbox": [ + 208, + 490, + 456, + 500 + ], + "type": "text", + "content": "\"There is a person sitting in the center of the image, remove it.\"" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 221, + 426, + 282, + 488 + ], + "blocks": [ + { + "bbox": [ + 221, + 426, + 282, + 488 + ], + "lines": [ + { + "bbox": [ + 221, + 426, + 282, + 488 + ], + "spans": [ + { + "bbox": [ + 221, + 426, + 282, + 488 + ], + "type": "image", + "image_path": "6c7b6ae1f9b5417e6229f74a460a1f1e953192f2280e3468f54c320a2e1f3fbc.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 283, + 426, + 343, + 488 + ], + "blocks": [ + { + "bbox": [ + 283, + 426, + 343, + 488 + ], + "lines": [ + { + "bbox": [ + 283, + 426, + 343, + 488 + ], + "spans": [ + { + "bbox": [ + 283, + 426, + 343, + 488 + ], + "type": "image", + "image_path": "f60cdce23972f672d834616bbc61335910a5e034dcd3d81b866be4ef563db57a.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 345, + 426, + 405, + 488 + ], + "blocks": [ + { + "bbox": [ + 345, + 426, + 405, + 488 + ], + "lines": [ + { + "bbox": [ + 345, + 426, + 405, + 488 + ], + "spans": [ + { + "bbox": [ + 345, + 426, + 405, + 488 + ], + "type": "image", + "image_path": "5292e0289c63c2d2c92461b9ebcf5a7ad1c94becbb4dbc1cfb5edb28bbdadbd1.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 406, + 426, + 466, + 488 + ], + "blocks": [ + { + "bbox": [ + 406, + 426, + 466, + 488 + ], + "lines": [ + { + "bbox": [ + 406, + 426, + 466, + 488 + ], + "spans": [ + { + "bbox": [ + 406, + 426, + 466, + 488 + ], + "type": "image", + "image_path": "cf1adffa360ca40c42e0ade0085d54d50e9885b34c43b79ecb1f425218fb1148.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 468, + 426, + 529, + 488 + ], + "blocks": [ + { + "bbox": [ + 468, + 426, + 529, + 488 + ], + "lines": [ + { + "bbox": [ + 468, + 426, + 529, + 488 + ], + "spans": [ + { + "bbox": [ + 468, + 426, + 529, + 488 + ], + "type": "image", + "image_path": "36f2bc973f8a8d3bcd909e31ebdd8887dcf496f0ef9a3f86ce053cf24b1068be.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 82, + 502, + 144, + 564 + ], + "blocks": [ + { + "bbox": [ + 82, + 502, + 144, + 564 + ], + "lines": [ + { + "bbox": [ + 82, + 502, + 144, + 564 + ], + "spans": [ + { + "bbox": [ + 82, + 502, + 144, + 564 + ], + "type": "image", + "image_path": "8725f5438b4b4f9fc8b71cdf3b350b9d447b69beba70881cca3da54d569e3523.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 586, + 555, + 608 + ], + "lines": [ + { + "bbox": [ + 55, + 586, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 555, + 608 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on AdvancedEdit. InsightEdit shows superior instruction following and background consistency capability." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 159, + 502, + 220, + 564 + ], + "blocks": [ + { + "bbox": [ + 159, + 502, + 220, + 564 + ], + "lines": [ + { + "bbox": [ + 159, + 502, + 220, + 564 + ], + "spans": [ + { + "bbox": [ + 159, + 502, + 220, + 564 + ], + "type": "image", + "image_path": "8e9edd03326b5583b2b259a5250db061dbd747c80d5b1eb0002ab92e5a5ab046.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 566, + 527, + 586 + ], + "lines": [ + { + "bbox": [ + 160, + 566, + 527, + 586 + ], + "spans": [ + { + "bbox": [ + 160, + 566, + 527, + 586 + ], + "type": "text", + "content": "\"Please observe if there's a house in the image and change its appearance to a cozy cabin with a warm, inviting glow from within.\"" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 221, + 502, + 282, + 564 + ], + "blocks": [ + { + "bbox": [ + 221, + 502, + 282, + 564 + ], + "lines": [ + { + "bbox": [ + 221, + 502, + 282, + 564 + ], + "spans": [ + { + "bbox": [ + 221, + 502, + 282, + 564 + ], + "type": "image", + "image_path": "1eeb5fd9d19789ef53f81ec70fe8efa93ed53cb048b190b863708e84866eb423.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 283, + 502, + 343, + 564 + ], + "blocks": [ + { + "bbox": [ + 283, + 502, + 343, + 564 + ], + "lines": [ + { + "bbox": [ + 283, + 502, + 343, + 564 + ], + "spans": [ + { + "bbox": [ + 283, + 502, + 343, + 564 + ], + "type": "image", + "image_path": "5c8247aac02ad0619c71ebe6795670748be763529ca4d834171a04cae35faccb.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 345, + 502, + 405, + 564 + ], + "blocks": [ + { + "bbox": [ + 345, + 502, + 405, + 564 + ], + "lines": [ + { + "bbox": [ + 345, + 502, + 405, + 564 + ], + "spans": [ + { + "bbox": [ + 345, + 502, + 405, + 564 + ], + "type": "image", + "image_path": "57f914d0468f0e00ff9340ea949384ac5fcf54854ad4b5656e03d378cb38be5d.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 406, + 502, + 466, + 564 + ], + "blocks": [ + { + "bbox": [ + 406, + 502, + 466, + 564 + ], + "lines": [ + { + "bbox": [ + 406, + 502, + 466, + 564 + ], + "spans": [ + { + "bbox": [ + 406, + 502, + 466, + 564 + ], + "type": "image", + "image_path": "a6362222ae5c33dc04ac8a696f3995e2f13bf28b25fc109ca2920e10cb4bd42d.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 468, + 502, + 529, + 564 + ], + "blocks": [ + { + "bbox": [ + 468, + 502, + 529, + 564 + ], + "lines": [ + { + "bbox": [ + 468, + 502, + 529, + 564 + ], + "spans": [ + { + "bbox": [ + 468, + 502, + 529, + 564 + ], + "type": "image", + "image_path": "1c8f9d7fc05eb30d1c630039b9379e75ca1196d324b71d04865af4c859719a6d.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "type": "text", + "content": "humans over other methods, whether regarding background preservation or the ability to follow complex instructions." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "type": "text", + "content": "Figure 4 illustrates the qualitative results. As shown in the last column, our method trained on the AdvancedEdit dataset demonstrates a clear advantage over competing approaches. As can be seen, other methods struggle to maintain background consistency with the original image, often introducing artifacts or unintended alterations, while In" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 612, + 555, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 555, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 555, + 660 + ], + "type": "text", + "content": "sightEdit provides specific editing on the required regions. In the meanwhile, InsightEdit performs strong instruction following capabilities, having a great understanding of spatial, color, number, and detailed objects." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Results on Reason-Edit. Additionally, we conducted comparative experiments on the Reason-Edit[14], which is specifically designed to assess editing abilities with an emphasis on understanding and reasoning across various sce" + } + ] + } + ], + "index": 52 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2700" + } + ] + } + ], + "index": 53 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 295, + 96 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 295, + 96 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 295, + 96 + ], + "type": "table", + "html": "
IAAPSNR↑SSIM↑LPIPS↓CLIPScore↑VIEScore↑
22.3480.6920.09520.6527.307
✓22.8710.7160.07121.0027.545
", + "image_path": "9c9fc26848dc01cb8483263136a00e6fc369ee95a6c44d277d83f4c688fa4f78.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 121, + 269, + 194 + ], + "blocks": [ + { + "bbox": [ + 67, + 121, + 269, + 194 + ], + "lines": [ + { + "bbox": [ + 67, + 121, + 269, + 194 + ], + "spans": [ + { + "bbox": [ + 67, + 121, + 269, + 194 + ], + "type": "image", + "image_path": "d90bdf94379e4ddbae2a55307eb0d604bd7bb68637d751bb1b7cd42b79e71b81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 195, + 260, + 211 + ], + "lines": [ + { + "bbox": [ + 75, + 195, + 260, + 211 + ], + "spans": [ + { + "bbox": [ + 75, + 195, + 260, + 211 + ], + "type": "text", + "content": "\"Could you swap the white flowers with a diverse collection of colorful wildflowers, creating a vivid and colorful foreground?\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 69, + 217, + 269, + 281 + ], + "blocks": [ + { + "bbox": [ + 69, + 217, + 269, + 281 + ], + "lines": [ + { + "bbox": [ + 69, + 217, + 269, + 281 + ], + "spans": [ + { + "bbox": [ + 69, + 217, + 269, + 281 + ], + "type": "image", + "image_path": "7e93fbd28f48acfa79239cac18a2c0a58b82993a781137cd93f1d8cf6ddb738d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 282, + 261, + 297 + ], + "lines": [ + { + "bbox": [ + 76, + 282, + 261, + 297 + ], + "spans": [ + { + "bbox": [ + 76, + 282, + 261, + 297 + ], + "type": "text", + "content": "Transform the worn wooden table into a sleek, modern marble countertop while maintaining the arrangement of the objects." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 330, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 437 + ], + "type": "text", + "content": "narios (i.e. color, position, mirror, multiple-objects, reason, and size). The results in Table 3 show that InsightEdit outperforms other metrics, except for a slight CLIP score difference in reasoning tasks, where it slightly trails SmartEdit. Furthermore, when trained on our AdvancedEdit dataset, InsightEdit demonstrates improvements across all metrics, highlighting the model's ability to leverage the dataset to enhance both comprehension and reasoning capabilities, while also emphasizing the intrinsic value of the data used." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 445, + 156, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 156, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 156, + 456 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 462, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 640 + ], + "type": "text", + "content": "Ablation Studies on IAA. To validate the effectiveness of IAA, we conduct comparative experiments on InsightEdit with the AdvancedEdit dataset. Table 4 shows the quantitative results, it can be seen that, with the IAA module, InsightEdit leads to notable improvements across various evaluation metrics. The qualitative results can be seen in Figure 5. In the first example, without the IAA module, the house is erroneously transformed into wildflowers, resulting in poor background consistency due to the absence of the red house. In contrast, with the IAA module, the red house is preserved accurately, maintaining consistency with the source image. In the second example, the marble table is edited poorly without the IAA module, while the addition of the IAA module leads to a more natural edit that aligns well with the instruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Ablation Studies on Instruction Complexity. We investigate the impact of instruction complexity and granularity on model performance by comparing two datasets introduced in Section 3.1: SimpleEdit Dataset and AdvancedEdit Dataset. Table 5 demonstrates that incorporating the SimpleEdit Dataset alone enhances the model's per" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 555, + 104 + ], + "blocks": [ + { + "bbox": [ + 67, + 98, + 282, + 107 + ], + "lines": [ + { + "bbox": [ + 67, + 98, + 282, + 107 + ], + "spans": [ + { + "bbox": [ + 67, + 98, + 282, + 107 + ], + "type": "text", + "content": "Table 4. Ablation of IAA. Our IAA module performs best." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 555, + 104 + ], + "type": "table", + "html": "
SimpleEdit AdvancedEditPSNR↑SSIM↑LPIPS↓CLIPScore↑VIEScore↑
21.2670.6750.11220.3956.974
✓22.2600.7080.08020.5577.213
✓22.8710.7160.07121.0027.545
", + "image_path": "ce84b54749afa34d997c90bb1de38a84560940d55640544553dea27ba9662ece.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 106, + 553, + 126 + ], + "lines": [ + { + "bbox": [ + 314, + 106, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 314, + 106, + 553, + 126 + ], + "type": "text", + "content": "Table 5. Ablation of AdvancedEdit. Our method with AdvancedEdit performs best." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 139, + 551, + 281 + ], + "blocks": [ + { + "bbox": [ + 58, + 297, + 293, + 307 + ], + "lines": [ + { + "bbox": [ + 58, + 297, + 293, + 307 + ], + "spans": [ + { + "bbox": [ + 58, + 297, + 293, + 307 + ], + "type": "text", + "content": "Figure 5. Demonstration of the effectiveness of the IAA module." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 139, + 551, + 281 + ], + "lines": [ + { + "bbox": [ + 317, + 139, + 551, + 281 + ], + "spans": [ + { + "bbox": [ + 317, + 139, + 551, + 281 + ], + "type": "image", + "image_path": "921b9dcd1015ad834c3d0fdbb40d162a1758aeaec682dc409a59363cbcec7d9a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 281, + 542, + 295 + ], + "lines": [ + { + "bbox": [ + 318, + 281, + 542, + 295 + ], + "spans": [ + { + "bbox": [ + 318, + 281, + 542, + 295 + ], + "type": "text", + "content": "\"Can you identify the specific bright light source from the small window on the staircase wall, remove it smoothly and blend the area to match the surrounding tones?\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 295, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 320, + 295, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 320, + 295, + 547, + 304 + ], + "type": "text", + "content": "Figure 6. Demonstration of the effectiveness of AdvancedEdit." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 327, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 578 + ], + "type": "text", + "content": "formance, primarily due to the high-quality outputs generated by our mask-based image editing approach. However, the inclusion of the AdvancedEdit Dataset leads to substantial improvements, particularly in CLIPScore and VI-EScore, highlighting the importance of increased instruction detail and the value of integrating complex instructions. The qualitative comparison results, as illustrated in Figure 6, reveal that without any data augmentation, the method struggles to understand the instructions and produce effective edits in both two examples. When the SimpleEdit dataset is introduced, there is some improvement, though the edit quality remains suboptimal. However, with the use of the AdvancedEdit dataset, the method demonstrates the ability to accurately interpret editing instructions, resulting in vivid and precise edits. For instance, in the first example, the method produces finely detailed orange and black butterfly wings, while in the second example, it accurately identifies and removes a light source, maintaining harmony in the surrounding content. These results further show that advanced instruction can handle more complex image editing scenarios and generate more vivid editing results." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 586, + 466, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 466, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 466, + 597 + ], + "type": "text", + "content": "6. Discussion and Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "In conclusion, we present InsightEdit, an end-to-end instruction-based image editing method, utilizing a two-stream bridging mechanism to integrate both the textual and visual features reasoned by the powerful MLLM into the diffusion model to guide the image editing process more precisely. We design an automatic data construction pipeline to generate a large-scale, high-quality, and high-fidelity image editing dataset, considering both image and instruction perspectives." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "2701" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "text", + "content": "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 296, + 213 + ], + "type": "text", + "content": "[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 295, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 268 + ], + "type": "text", + "content": "[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 357 + ], + "type": "text", + "content": "[6] Tuhin Chakrabarty, Kanishk Singh, Arkadiy Saakyan, and Smaranda Muresan. Learning to follow object-centric image editing instructions faithfully. arXiv preprint arXiv:2310.19145, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 413 + ], + "type": "text", + "content": "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 457 + ], + "type": "text", + "content": "[8] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 459, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 295, + 523 + ], + "type": "text", + "content": "[9] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "type": "text", + "content": "[10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "type": "text", + "content": "[11] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 603, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 646 + ], + "type": "text", + "content": "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "text", + "content": "[13] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[14] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 128 + ], + "type": "text", + "content": "Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[15] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "text", + "content": "[16] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "text", + "content": "[17] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 276, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 297 + ], + "type": "text", + "content": "[18] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "type": "text", + "content": "[19] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. Advances in Neural Information Processing Systems, 36, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 333, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 376 + ], + "type": "text", + "content": "[20] Max Ku, Dongfu Jiang, Cong Wei, Xiang Yue, and Wenhu Chen. Viescore: Towards explainable metrics for conditional image synthesis evaluation. arXiv preprint arXiv:2312.14867, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "type": "text", + "content": "[21] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "type": "text", + "content": "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 478, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 544 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "text", + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 580, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 623 + ], + "type": "text", + "content": "[25] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "text", + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[27] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2702" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 115 + ], + "type": "text", + "content": "Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "text", + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 238 + ], + "type": "text", + "content": "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "text", + "content": "[30] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 369 + ], + "type": "text", + "content": "[31] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18359-18369, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "type": "text", + "content": "[32] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 426, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 479 + ], + "type": "text", + "content": "[33] Shaoan Xie, Yang Zhao, Zhisheng Xiao, Kelvin CK Chan, Yandong Li, Yanwu Xu, Kun Zhang, and Tingbo Hou. Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. arXiv preprint arXiv:2312.03771, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 481, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 524 + ], + "type": "text", + "content": "[34] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 526, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 569 + ], + "type": "text", + "content": "[35] Ling Yang, Bohan Zeng, Jiaming Liu, Hong Li, Minghao Xu, Wentao Zhang, and Shuicheng Yan. Editworld: Simulating world dynamics for instruction-following image editing. arXiv preprint arXiv:2405.14785, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "text", + "content": "[36] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. arXiv preprint arXiv:2312.17240, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "type": "text", + "content": "[37] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[38] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 352 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[39] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[40] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 262 + ], + "type": "text", + "content": "[42] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9026-9036, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "text", + "content": "[43] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "text", + "content": "[44] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. arXiv preprint arXiv:2312.03594, 2023." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2703" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Insightful Instance Features for 3D Instance Segmentation/275c7e3b-7085-4434-b01d-e129139897fa_content_list.json b/2025/Insightful Instance Features for 3D Instance Segmentation/275c7e3b-7085-4434-b01d-e129139897fa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9ce952750135f1148f3e09e4ff831efaf7b32522 --- /dev/null +++ b/2025/Insightful Instance Features for 3D Instance Segmentation/275c7e3b-7085-4434-b01d-e129139897fa_content_list.json @@ -0,0 +1,1459 @@ +[ + { + "type": "text", + "text": "Insightful Instance Features for 3D Instance Segmentation", + "text_level": 1, + "bbox": [ + 204, + 130, + 794, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wonseok Roh $^{1}$ Hwanhee Jung $^{1}$ Giljoo Nam $^{2}$ Dong In Lee $^{1}$ \nHyeongcheol Park $^{1}$ Sang Ho Yoon $^{3}$ Jungseock Joo $^{4}$ Sangpil Kim $^{1*}$ $^{1}$ Korea University $^{2}$ Meta $^{3}$ KAIST $^{4}$ UCLA", + "bbox": [ + 207, + 179, + 795, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 328, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent 3D Instance Segmentation methods typically encode hundreds of instance-wise candidates with instancespecific information in various ways and refine them into final masks. However, they have yet to fully explore the benefit of these candidates. They overlook the valuable cues encoded in multiple candidates that represent different parts of the same instance, resulting in fragments. Also, they often fail to capture the precise spatial range of 3D instances, primarily due to inherent noises from sparse and unordered point clouds. In this work, to address these challenges, we propose IKNE, a novel instance-wise knowledge enhancement approach. We first introduce an Instance-wise Knowledge Aggregation (IKA) to associate scattered single instance details by optimizing correlations among candidates representing the same instance. Moreover, we present an Instance-wise Structural Guidance (ISG) to enhance the spatial understanding of candidates using structural cues from ambiguity-reduced features. Here, we utilize a simple yet effective truncated singular value decomposition algorithm to minimize inherent noises of 3D features. In our extensive experiments on large-scale datasets, ScanNetV2, ScanNet200, S3DIS, and STPLS3D, IKNE outperforms existing works. We validate the effectiveness of our modules in both kernel-based and transformer-based architectures.", + "bbox": [ + 88, + 304, + 485, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 686, + 223, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Instance Segmentation (3DIS) is a fundamental 3D computer vision task that significantly contributes to real-world applications. Given 3D point cloud scenes, the 3DIS task identifies respective instances and assigns semantic class labels with the goal of comprehensively understanding entire spatial environments. In real-life 3D scenarios, substantial occlusion and truncation commonly arise, especially when objects are overlapped or obscured by others. To tackle these challenges, early approaches mainly focused on accurately generating region proposals (top-down) [21, 50, 52] or effectively grouping points with clus", + "bbox": [ + 89, + 710, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tering algorithms (bottom-up) [6, 24, 45]. However, they presuppose that the intermediate processes, such as bounding box detection [18] or heuristic voting mechanism [34], generate near perfect results, which is often not the case.", + "bbox": [ + 511, + 273, + 906, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, kernel-based [19, 20, 32, 48] and transformer-based [29, 31, 39, 43] 3DIS approaches have been proposed, aiming to overcome the limitations of traditional frameworks. Kernel-based methods leverage instance-aware kernels for dynamic convolution, which decodes instance masks. They represent instances as kernels, replacing clustering algorithms with point sampling processes. On the other hand, transformer-based methods train instance queries to identify features about individual objects using attention mechanisms. These queries enable the direct prediction of per-point categories and instance labels. In both architectures, models utilize hundreds of instance-wise candidate features, to estimate the final instance mask. For example, ISBNet [32] (kernel-based) produces 256 instance kernels, and MAFT [29] (transformer-based) leverages 400 instance queries. However, despite the significant improvements based on their architectural advantages, they have yet to fully optimize the handling of these numerous candidates and still face several challenges, as illustrated in Fig. 1. To tackle these challenges, we introduce two main modules, IKA and ISG, each thoughtfully designed to enhance the informativeness of instance candidate features.", + "bbox": [ + 511, + 338, + 908, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "First, we empirically observe that multiple candidates from previous works often encode cues from the same single instance, as shown in Fig. 1 (a). These candidates are then typically decoded into considerable fragment masks representing different components of each instance, such as the backrest or armrest of the sofa. In this work, to estimate the complete coverage of each instance, we introduce an Instance-wise Knowledge Aggregation (IKA) network, which integrates the scattered intelligence from multiple fragments beyond relying solely on a few fragments based on confidence scores. Specifically, we optimize the correlations among instance-wise candidate features, encouraging the network to be aware of the distributed instance-specific knowledge. We initially identify candidates representing the same instance based on the similarity of their features, as", + "bbox": [ + 511, + 674, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author (spk7@korea.ac.kr)", + "bbox": [ + 107, + 886, + 339, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14057", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/30a9a2e2005cbac10a94477050abc7efac50968382e685ce4335538ed8967ea2.jpg", + "image_caption": [ + "(a) Per-point masks of instance-wise candidates" + ], + "image_footnote": [], + "bbox": [ + 94, + 74, + 500, + 246 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4f5a6da64acca8cbfad76fc499301524148b1bec0e9c30e97c2c73d325ad47a4.jpg", + "image_caption": [ + "(b) Instance misinterpretation cases", + "Figure 1. Examples of two challenging cases of 3DIS. (a) Per-point masks of instance-wise candidates: highly correlated instance-wise candidates usually represent incomplete fragments of the same single instance, such as the backrest or armrest of a sofa. And (b) Instance misinterpretation cases: existing studies often confuse instances from backgrounds (Scene1) or misunderstand the spatial range and appearance of instances (Scene2). Note that we also provide additional examples for both cases in the supplementary material (Fig. 1)." + ], + "image_footnote": [], + "bbox": [ + 504, + 74, + 903, + 246 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "highly correlated candidates usually capture the information from the same instance. Then, inspired by correlation regularization mechanisms [2, 23, 53], we softly guide identified candidates to be closer to each other in the latent space, aiming to enhance solidarity among them. As a result, IKA facilitates the network to learn relationships between candidates and aggregate their comprehensive details, maximizing the advantages of using hundreds of candidates.", + "bbox": [ + 88, + 340, + 480, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to IKA, we explore additional avenues to enrich the knowledge of instance-wise candidates. Unlike the human visual cognitive system, which perceives instances with sufficient prior knowledge, the network understands instances depending on their external properties from point features. However, point features are frequently unstable, primarily due to ambiguous inherent noises [13, 36, 47, 49] caused by sparse and incomplete point clouds. These noises are likely to confuse the models, making it hard to precisely interpret the specific spatial range and unique appearance of instances, as shown in Fig. 1 (b). To mitigate the negative impact of such noises, we propose an Instance-wise Structural Guidance (ISG) network specifically designed to improve candidates' structural understanding of 3D instances from point features. Here, we utilize the truncated singular value decomposition (SVD) [14, 42] algorithm to strengthen essential cues (e.g., shape) of instance features while explicitly reducing fuzzy noises. Then, we effectively transfer fundamental clues for object cognition from ambiguity-minimized features to corresponding original features based on the cross-correlations between them. Ultimately, our instance features are now highly informative across complex 3D instances with aggregated and structural knowledge from IKA and ISG. Our novel modules are applicable to frameworks employing considerable instance candidates, regardless of their underlying model structures, including kernel-based and transformer-based architectures.", + "bbox": [ + 93, + 464, + 482, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given landmark datasets for 3DIS, ScanNetV2 [7], ScanNet200 [38], S3DIS [1], and STPLS3D [5], we vali", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "date the effectiveness of our novel framework. Our method outperforms the existing state-of-the-art methods. To summarize, our main contributions are listed as follows:", + "bbox": [ + 511, + 340, + 903, + 386 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To extend the benefits of producing multiple instance-wise candidates, we introduce the Instance-wise Knowledge Aggregation (IKA) network, which associates scattered instance-specific information of the same single instance by optimizing correlations among them.", + "- We design the Instance-wise Structural Guidance (ISG) network to improve the structural knowledge of candidates. Specifically, we use a simple yet strong truncated singular value decomposition process to emphasize essential cues from point features while filtering out noises.", + "- We analyze the effectiveness of our framework, IKNE, on four challenging datasets. Comprehensive experiments in various 3D scenarios demonstrate that our novel method achieves new State-of-the-Art performance on 3DIS." + ], + "bbox": [ + 513, + 388, + 903, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 614, + 653, + 631 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Instance Segmentation (3DIS) aims to distinguish individual instances within 3D scenes and assign corresponding semantic categories. Typically, 3DIS approaches can be categorized into four types: proposal-based [21, 50, 52], grouping-based [6, 11, 12, 24, 45], kernel-based [19, 20, 32, 48], and transformer-based [29, 31, 37, 39, 43]. Based on the impressive achievements of detection methods [18, 46], proposal-based approaches first detect instances and employ proposed regions as hard references to predict masks. However, these strategies, such as 3D-SIS [21] or 3D-BoNet [50], heavily rely on the quality of the outputs from the detection process. On the other hand, grouping-based methods catch 3D instances through a bottom-up pipeline, aggregating closely related points into instances using predicted semantic categories and center offsets. Yet, they still depend on intermediate manually tuned processing, such as point grouping [24] or heuristic voting mechanisms [34].", + "bbox": [ + 509, + 643, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14058", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/919a97be83018fe598f7553d2c6c20071c8295fe45b50e6b4b650bc5b690841e.jpg", + "image_caption": [ + "Figure 2. An overview of our framework IKNE. Built upon the modern kernel-based structure, our model consists of four main modules: (1) Sparse Convolutional 3D Backbone; (2) Instance-wise Knowledge Aggregation (IKA, Sec. 3.2), which softly associates scattered cues from highly correlated instance features of $F_{inst}$ ; (3) Instance-wise Structural Guidance (ISG, Sec. 3.3), which smartly instructs $F_{inst}$ with essential clues from $\\tilde{F}_{inst}$ , using an effective SVD algorithm; and (4) Dynamic Convolution, which produces $\\hat{M}_f$ from kernels $K$ ." + ], + "image_footnote": [], + "bbox": [ + 96, + 74, + 906, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these limitations, kernel-based approaches have been introduced. For example, DyCo3D [19] employs a clustering algorithm to generate kernels for dynamic convolutions to predict instance masks, while ISBNet [32] presents a cluster-free method using instance-wise kernels. Recently, Spherical Mask [41], based on ISBNet, has overcome the low-quality results of coarse-to-fine strategies by using spherical representation to mitigate the propagation of false negatives and overestimation of instance size. Furthermore, transformer-based methods, Mask3D [39], MAFT [29], and EASE [37] utilize instance-wise queries to encode individual instance information through attention mechanisms. In this work, we focus on improving the quality of numerous instance-wise candidates introduced from recent studies with two novel approaches, IKA and ISG.", + "bbox": [ + 88, + 359, + 480, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Singular Value Decomposition (SVD) is a fundamental algorithm in linear algebra, widely used for matrix factorization [14, 27, 42], dimensionality reduction [22, 51], and 2D image processing [9, 16, 26, 35, 40]. SVD decomposes a matrix $A \\in \\mathbb{R}^{m \\times n}$ into the dot product of three matrices, $U \\in \\mathbb{R}^{m \\times m}$ , $\\Sigma \\in \\mathbb{R}^{m \\times n}$ , and $V \\in \\mathbb{R}^{n \\times n}$ , each containing meaningful vectors and scalar values, as follows:", + "bbox": [ + 89, + 589, + 482, + 695 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA = U \\cdot \\Sigma \\cdot V ^ {T} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 699, + 480, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $U$ and $V$ are the matrices with orthonormal columns, satisfying $U^T U = V^T V = I$ , and $\\Sigma$ is a diagonal matrix with nonnegative singular values $\\sigma$ in descending order as $\\sigma_1 \\geqslant \\sigma_2 \\geqslant \\dots \\geqslant \\sigma_r$ . Based on SVD, truncated SVD [17] is an insightful dimension-reduction technique that minimizes nonessential noises while highlighting essential cues from the original matrix $A$ . It produces a low-rank approximation $\\tilde{A} \\in \\mathbb{R}^{m \\times n}$ by retaining the most important top $k$ singular values and their corresponding vectors from $U$ , $\\Sigma$ , and $V$ :", + "bbox": [ + 89, + 719, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {A} = U _ {k} \\cdot \\Sigma_ {k} \\cdot V _ {k} ^ {T} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 858, + 480, + 877 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $U_{k}\\in \\mathbb{R}^{m\\times k}$ $\\Sigma_{k}\\in \\mathbb{R}^{k\\times k}$ , and $V_{k}\\in \\mathbb{R}^{n\\times k}$ are the", + "bbox": [ + 89, + 883, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "truncated matrices with the top $k$ singular value vectors.", + "bbox": [ + 511, + 359, + 883, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the field of 2D image processing, SVD has been extensively utilized for low-rank approximation. [16, 35] employ SVD to factorize and estimate low-rank patches, effectively reducing noise and enhancing image quality. Similarly, [4, 25] operate SVD to detect tampering by extracting and analyzing low-dimensional representations of images for image forensics. Also, SVD showcases robust performance in image compression [3, 33] and recovery [40]. Further, [9] leverages SVD within CNNs to learn subspace information for hyperspectral image reconstruction. However, despite its potential benefits, the introduction of SVD has received limited attention in 3D vision, especially for point clouds. Based on these observations, we apply SVD to address point features that contain inherent noises due to the unordered properties of point clouds, facilitating the model to become more robust against complex ambiguities.", + "bbox": [ + 511, + 376, + 906, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 630, + 604, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce a novel 3D Instance Segmentation framework IKNE. We provide an overview of the whole pipeline (Sec. 3.1 and Fig. 2), followed by detailed technical discussions: Instance-wise Knowledge Aggregation (Sec. 3.2) and Instance-wise Structural Guidance (Sec. 3.3).", + "bbox": [ + 511, + 655, + 905, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 741, + 624, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent 3DIS methods [19, 20, 32, 37, 39] commonly follow a similar paradigm; they spread hundreds of instance candidates across 3D scenes to capture instance-specific information in various ways and refine them into final instance masks. In this work, we encourage these numerous instance candidates to be highly informative, improving their understanding of diverse real-world instances. We describe our instance-wise knowledge enhancement approaches based on the modern kernel-based architecture [32]. As shown in", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14059", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2, our model consists of four main modules: (1) Sparse Convolutional 3D Backbone; (2) Instance-wise Knowledge Aggregation, which relates the scattered information of individual instances; (3) Instance-wise Structural Guidance, which effectively instructs candidates with spatial cues from augmented instance features; and (4) Dynamic Convolution network. Note that we also describe our transformer-based architecture in the supplementary material (Sec.3).", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Kernel-based 3DIS Architecture. First, sparse convolutional U-Net backbone [15] takes a colored point cloud $P \\in \\mathbb{R}^{N_p \\times 6}$ as input and voxelizes $P$ into voxels to extract point-wise feature maps $F_{p} \\in \\mathbb{R}^{N_{p} \\times D}$ . Given the point features, the point aggregator samples a set of instance candidate features, referring to point positions based on the Farthest Point Sampling (FPS) [10]. Specifically, we employ two-stage point aggregator blocks to produce optimal instance features across 3D scenes. In the first stage, $N_{k}^{\\prime}$ number of candidates $F_{inst}^{\\prime} \\in \\mathbb{R}^{N_{k}^{\\prime} \\times D}$ are sampled from full-resolution feature $F_{p}$ , while in the second stage, $N_{k}$ ( $< N_{k}^{\\prime}$ ) instance features $F_{inst} \\in \\mathbb{R}^{N_{k} \\times D}$ are sampled based on the $F_{inst}^{\\prime}$ . We then classify the category $C_{i}$ for $i = \\{1,2,\\dots,N_{c}\\}$ of each instance from $F_{inst}$ via linear classification head $f_{\\mathrm{cls}}$ using cross-entropy loss as follows:", + "bbox": [ + 89, + 217, + 483, + 446 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = - \\mathbb {E} _ {c, w _ {c} \\sim \\mathbb {D}} \\left[ \\sum_ {r \\in N _ {c}} w _ {c} [ r ] \\log f _ {\\mathrm {c l s}} \\left(F _ {\\text {i n s t}}\\right) [ r ] \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 455, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{c}$ indicates one-hot encoded category labels and $\\mathbb{D}$ represents (input) data distribution. Also, we predict instance kernels $K\\in \\mathbb{R}^{N_k\\times D}$ based on $F_{inst}$ through MLPs. Furthermore, each point-wise prediction head processes $F_{p}$ to generate mask features $F_{mask}\\in \\mathbb{R}^{N_p\\times D}$ and axis-aligned bounding boxes $F_{box}\\in \\mathbb{R}^{N_p\\times 6}$ as auxiliary information for mask prediction. Finally, the dynamic convolution network produces the final instance mask $\\hat{M}_f$ using instance kernels $K$ and $F_{mask}$ . To optimize $\\hat{M}_f$ , we formulate the following loss $\\mathcal{L}_{mask}$ as a sum of the two losses:", + "bbox": [ + 89, + 507, + 483, + 661 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {\\mathrm {B C E}} \\mathcal {L} _ {\\mathrm {B C E}} \\left(M _ {f}, \\hat {M} _ {f}\\right) + \\lambda_ {\\mathrm {d i c e}} \\mathcal {L} _ {\\mathrm {d i c e}} \\left(M _ {f}, \\hat {M} _ {f}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 672, + 482, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_{f}$ is the ground-truth mask, and $\\mathcal{L}_{\\mathrm{BCE}}$ and $\\mathcal{L}_{\\mathrm{dice}}$ denote binary cross-entropy and Dice losses [8], respectively.", + "bbox": [ + 89, + 700, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Instance-wise Knowledge Aggregation (IKA)", + "text_level": 1, + "bbox": [ + 89, + 741, + 470, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent studies [19, 20, 29, 31, 32, 39, 48] in 3DIS deal with instance-wise features $F_{inst}$ as essential components, aiming to encode rich semantic and geometric information about instances. They spread hundreds of candidates over expansive 3D scenes to cover as many instances as possible; this number $N_{k}$ far exceeds the total number $N_{i}$ of instances present in any single scene (see Tab.4 in Supp.). Although these candidates provide practical information for mask prediction, they often represent separate parts of each instance,", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/300b9787ccc7195a5588487da0f6cc05465952974eb66aa20a386dc31e0d56e9.jpg", + "image_caption": [ + "Figure 3. In IKA, we specify highly correlated instance candidates $(I_{ka} > \\tau)$ representing identical instances based on the correlation matrix $I_{ka}$ . Then, we softly aggregate scattered clues of related candidates, optimizing $I_{ka}$ with pseudo-binary label $I_{ka}^{*}$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "resulting in fragmented instance masks. With these insights, we focus on associating scattered knowledge to fully leverage the advantages of multiple instance candidates.", + "bbox": [ + 511, + 344, + 903, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To handle multiple candidates, we first identify those representing the same instance based on their pairwise affinity. Specifically, we compute a correlation matrix $I_{ka} \\in \\mathbb{R}^{N_k \\times N_k}$ using the dot product of $F_{inst}$ and its transpose $F_{inst}^T$ , then normalize $I_{ka}$ to a range between 0 and 1. Here, we note that candidates representing the same instance yield relatively high correlation values. Thus, based on $I_{ka}$ , we consider pairwise candidates with correlation values surpassing a predefined threshold $\\tau$ as fragments from the same instance. Then, as shown in Fig. 3, we softly guide these candidates by optimizing their correlations in $I_{ka}$ , motivated by the success of recent self-supervised learning strategies [2, 53]. We extend this approach to further enrich coherence among highly correlated candidate features while also encouraging the network to be aware of scattered cues for each instance. In particular, we dynamically construct a pseudo-binary correlation label matrix $I_{ka}^* \\in \\{0,1\\}^{N_k \\times N_k}$ , where elements of highly correlated pairs $(\\tau)$ in $I_{ka}$ are 1, while the rest $(\\leqslant \\tau)$ are 0 for all candidate pairs. Given $I_{ka}$ and $I_{ka}^*$ , we formulate the element-wise correlation regularization loss $\\mathcal{L}_{KA}$ as follows:", + "bbox": [ + 511, + 390, + 906, + 708 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {k a} = F _ {i n s t} \\cdot F _ {i n s t} ^ {T}, I _ {k a} ^ {*} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} I _ {k a} (i, j) > \\tau \\\\ 0 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 709, + 906, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {K A} = \\sum_ {I _ {k a} > \\tau} \\| I _ {k a} ^ {*} (i, j) - I _ {k a} (i, j) \\| _ {2} \\\\ + w \\cdot \\sum_ {I _ {k a} \\leqslant \\tau} \\| I _ {k a} ^ {*} (i, j) - I _ {k a} (i, j) \\| _ {2} \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 760, + 905, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\| _2$ denotes the $L_{2}$ norm, and $w$ is the weight value used to balance between highly correlated pairs and others. In Eq. 6, the first term facilitates candidates representing the identical instance closer together in the latent space,", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14060", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "whereas the second term decreases correlations between unrelated candidates, thereby reducing confusion from irrelevant knowledge. In conclusion, IKA establishes practical connections that strengthen solidarity across multiple candidates encoding scattered cues by regularizing loss $\\mathcal{L}_{KA}$ .", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Instance-wise Structural Guidance (ISG)", + "text_level": 1, + "bbox": [ + 89, + 174, + 441, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In real-world scenarios, 3D instances are usually placed near each other or overlapping, arranged in inconsistent patterns without following conventional standards. Also, these instances are represented as sparse point clouds, which often induce fuzzy inherent noises in point-wise features. Hence, the instance-wise features $F_{inst}$ from these settings are unclear, causing models to misinterpret the spatial extent of each instance. In this work, we explicitly reduce such ambiguity of $F_{inst}$ by leveraging intelligent structural guidance from clarity-enhanced instance-wise feature $\\tilde{F}_{inst}$ .", + "bbox": [ + 89, + 198, + 483, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For accurate segmentation, it is crucial to capture the clear morphological shape of complex instances. From this intuition, we aim to encourage the model to better learn the structural appearance of instances. Thus, we implement effective structural guidance based on a truncated singular value decomposition (SVD) [14, 42] algorithm primarily used for noise reduction. As illustrated in Fig. 4, given $F_{inst}^{\\prime} \\in \\mathbb{R}^{N_k^{\\prime} \\times D}$ from the first point aggregator, we initially decompose $F_{inst}^{\\prime}$ into three matrices based on SVD: (1) $U \\in \\mathbb{R}^{N_k^{\\prime} \\times N_k^{\\prime}}$ , (2) $\\Sigma \\in \\mathbb{R}^{N_k^{\\prime} \\times D}$ , and (3) $V \\in \\mathbb{R}^{D \\times D}$ , as:", + "bbox": [ + 89, + 349, + 483, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i n s t} ^ {\\prime} = U \\cdot \\Sigma \\cdot V ^ {T} = \\sum_ {i} ^ {N _ {k} ^ {\\prime}} u _ {i} \\cdot \\sigma_ {i} \\cdot v _ {i} ^ {T} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 508, + 482, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $U = (u_{1},\\ldots ,u_{n})$ and $V = (v_{1},\\dots ,v_{n})$ are the matrices with orthonormal columns, and $\\Sigma = diag(\\sigma_1,\\dots,\\sigma_n)$ is a diagonal matrix containing singular values, arranged in the descending order. Typically, vectors with higher singular values in $\\Sigma$ include meaningful information for object cognition, while the lower-rank vectors are regarded as less important, potentially causing ambiguity. Then, we preserve only the top $t$ singular values from $\\Sigma$ , along with their corresponding singular vectors from $U$ and $V$ , facilitating an ambiguity-reduced yet information-rich representation of the original features $F_{inst}^{\\prime}$ as follows:", + "bbox": [ + 89, + 555, + 483, + 723 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {F} _ {i n s t} ^ {\\prime} \\in \\mathbb {R} ^ {N _ {k} ^ {\\prime} \\times D} = \\tilde {U} \\cdot \\tilde {\\Sigma} \\cdot \\tilde {V} ^ {T} = \\sum_ {i} ^ {N _ {k} ^ {\\prime}} \\tilde {u} _ {i} \\cdot \\tilde {\\sigma} _ {i} \\cdot \\tilde {v} _ {i} ^ {T} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 729, + 482, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{U} \\in \\mathbb{R}^{N_k' \\times t}$ , $\\tilde{\\Sigma} \\in \\mathbb{R}^{t \\times t}$ , and $\\tilde{V} \\in \\mathbb{R}^{D \\times t}$ are truncated matrices using only $t$ columns from $U$ , $\\Sigma$ , and $V$ , then $\\tilde{F}_{inst}'$ is reconstructed from them. Based on both $F_{inst}'$ and $\\tilde{F}_{inst}'$ , we sample instance-wise features $F_{inst}$ and $F_{inst}$ , respectively, employing the shared second stage point aggregator. Given $F_{inst}$ and $\\tilde{F}_{inst}$ , we calculate the cross-correlation matrix $I_{sg} \\in \\mathbb{R}^{N_k \\times N_k}$ and also generate a pseudo-binary correlation label $I_{sg}^* \\in \\{0, 1\\}^{N_k \\times N_k}$ , as in the IKA network.", + "bbox": [ + 89, + 777, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/888d7c58af97bbd2c756c3920ad5fb856cd7a5755b7c55dc218a84f94a591b1c.jpg", + "image_caption": [ + "Original Instance Feats.", + "Figure 4. In ISG, we decompose $F_{inst}^{\\prime}$ into three matrices $(U, \\Sigma, V)$ using the SVD algorithm and keep top $t$ singular vectors to filter out less important noises. Then, we effectively enhance $F_{inst}$ with highlighted essential cues from $\\tilde{F}_{inst}$ by optimizing $\\mathcal{L}_{SG}$ ." + ], + "image_footnote": [], + "bbox": [ + 516, + 94, + 906, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then effectively transfer structural clues from $\\tilde{F}_{inst}$ , utilizing the correlation regularization loss $\\mathcal{L}_{SG}$ as follows:", + "bbox": [ + 511, + 333, + 905, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {s g} = F _ {i n s t} \\cdot \\tilde {F} _ {i n s t} ^ {T}, I _ {s g} ^ {*} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} I _ {s g} (i, j) > \\tau \\\\ 0 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 367, + 906, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {S G} = \\sum_ {I _ {s g} > \\tau} \\left| \\left| I _ {s g} ^ {*} (i, j) - I _ {s g} (i, j) \\right| \\right| _ {2} \\\\ + \\tilde {w} \\cdot \\sum_ {I _ {s g} \\leqslant \\tau} \\left\\| I _ {s g} ^ {*} (i, j) - I _ {s g} (i, j) \\right\\| _ {2} \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 412, + 905, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{w}$ represents the balancing hyper-parameter, and the role of each term in Eq. 10 follows Eq. 6. By optimizing $\\mathcal{L}_{SG}$ , ISG conditionally instructs instance features $F_{inst}$ to incorporate spatial configuration details from clarity-enhanced features $\\tilde{F}_{inst}$ , improving discriminative power. Loss Function. Finally, our full network minimizes $\\mathcal{L}_{total}$ :", + "bbox": [ + 511, + 489, + 905, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} \\\\ + \\lambda_ {K A} \\mathcal {L} _ {K A} + \\lambda_ {S G} \\mathcal {L} _ {S G} \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 593, + 903, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where each $\\lambda$ is a hyper-parameter decided via extensive grid searches to balance the impact of respective loss terms.", + "bbox": [ + 511, + 638, + 905, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 681, + 645, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 705, + 702, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We train and evaluate the overall performance using four landmark benchmarks for 3D Instance Segmentation (3DIS): ScanNetV2 [7], ScanNet200 [38], S3DIS [1], and STPLS3D [5]. These four datasets provide 3D point cloud data collected in diverse real-world environments. Note that detailed descriptions of each dataset and all implementation details are provided in Supplementary Material.", + "bbox": [ + 511, + 731, + 906, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate the 3DIS performance using the Average Precision (AP), a conventional metric in computer vision tasks. We report the mean average precision (mAP) across IoU (Intersection of Union) thresholds", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14061", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cf96c15cb1c6a51e8bd48c1b1d0685f0d2ab2c75606cfd6f0de9914985fbd37f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodScanNetV2 ValScanNetV2 Test
mAP\\( mAP_{50} \\)\\( mAP_{25} \\)mAP\\( mAP_{50} \\)\\( mAP_{25} \\)
3D-SIS [21]-18.735.716.138.255.8
MTML [28]20.340.255.428.254.973.1
PointGroup [24]34.856.771.340.763.677.8
DyCo3D [19]35.457.672.939.564.176.1
3D-MPA [12]35.559.172.435.561.173.7
MaskGroup [54]42.063.363.343.466.479.2
SoftGroup [44]46.067.678.950.476.186.5
SSTNet [30]49.464.374.050.669.878.9
Mask3D [39]55.273.782.956.678.082.9
SPFormer [43]56.373.982.954.977.085.1
Q-Former [31]56.574.283.358.378.787.4
ISBNet [32]56.873.381.355.976.387.0
MAFT [29]58.475.984.559.678.686.0
EASE [37]60.277.284.259.878.986.7
SphMask [41]62.379.988.261.681.287.5
IKNE (Ours)62.981.888.761.781.288.0
", + "bbox": [ + 91, + 78, + 486, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "incremented by $5\\%$ , ranging from $50\\%$ to $95\\%$ . Also, we assess $\\mathrm{mAP}_{50}$ and $\\mathrm{mAP}_{25}$ , which measure accuracy based on IoU thresholds of $50\\%$ and $25\\%$ . For the S3DIS [1], we further provide mean precision (mPrec) and mean recall (mRec) at a $50\\%$ IoU threshold, following previous works.", + "bbox": [ + 89, + 436, + 483, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 89, + 527, + 483, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we quantitatively compare our proposed framework IKNE with previous methods. Despite significant advances with high scores, these methods, regardless of architecture, still waste considerable instance candidates that may contain valuable information. To maximize the benefits of using numerous candidates, we present novel instance-wise knowledge aggregation (IKA) and guidance (ISG) networks. In Tab. 1, we first evaluate mean average precision (mAP) with different IoU thresholds across 18 classes on the ScanNetV2 [7] dataset. IKNE generally outperforms prior works, achieving new state-of-the-art scores in terms of mAP, $\\mathrm{mAP}_{50}$ , and $\\mathrm{mAP}_{25}$ on both validation and hidden test sets. While our method slightly exceeds the latest coarse-to-fine approach, Spherical Mask [41], ours transcends our primary baselines, which rely on hundreds of instance candidates, such as ISBNet [32] (kernel-based) and MAFT [29] (transformer), by large margins across all metrics. In Tab. 2, we assess 3DIS performance on Area 5 and the 6-fold cross-validation set of the S3DIS [1] dataset. For Area 5 evaluation, we train with Areas 1 to 6, excluding 5, and validate with Area 5; for 6-fold cross-validation, we average the validation scores across all 6 areas. IKNE demonstrates impressive performance in both evaluations, reaching", + "bbox": [ + 88, + 553, + 486, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dfcde9732cfbaea6622f2c9f5792988cd665fb3117a36a0a9626b7077e0b9e4d.jpg", + "table_caption": [ + "Table 1. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the ScanNetV2 [7] dataset." + ], + "table_footnote": [], + "table_body": "
MethodS3DIS Area 5S3DIS 6-fold CV
AP\\( AP_{50} \\)\\( Prec_{50} \\)\\( Rec_{50} \\)AP\\( AP_{50} \\)\\( Prec_{50} \\)\\( Rec_{50} \\)
3D-Bonet [50]--57.540.2--65.647.6
PointGroup [24]-57.861.962.1-64.069.669.2
3D-MPA [12]--63.158.0--66.764.1
DyCo3D [19]--64.364.2----
MaskGroup [54]-65.062.964.7-69.966.669.6
SSTNet [30]42.759.365.564.254.167.873.573.4
SoftGroup [44]51.666.173.666.654.468.975.369.8
SPFormer [43]-66.872.867.1-69.274.071.1
ISBNet [32]56.367.570.572.060.870.577.577.1
DKNet [48]--70.865.3--75.371.1
Mask3D [39]56.668.468.766.364.575.572.874.5
Q-Former [31]57.769.970.572.262.073.372.773.4
MAFT [29]-69.1------
EASE [37]59.071.669.468.765.376.473.674.6
SphMask [41]60.572.371.376.364.072.378.177.7
IKNE (Ours)61.173.072.077.065.876.979.478.9
", + "bbox": [ + 514, + 78, + 911, + 390 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/274f2619bd16d4d5826dfb08c93dce4a990f093ef21b90da5bf1dfce20c3f535.jpg", + "table_caption": [ + "Table 2. Comparison of 3DIS performance with state-of-the-art approaches on the S3DIS [1] Area 5 and 6-fold cross validation." + ], + "table_footnote": [], + "table_body": "
MethodmAP\\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{25} \\)
SPFormer [43]25.233.839.6
Mask3D [39]27.437.042.3
Q-Former [31]28.137.143.4
MAFT [29]29.238.243.3
IKNE (Ours)29.938.944.9
", + "bbox": [ + 517, + 435, + 898, + 542 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/455f2596ccc90686327de63480b344fc07ff919772433b61e2e9ade19d0b2ded.jpg", + "table_caption": [ + "Table 3. Comparison of 3DIS performance with state-of-the-art approaches on the ScanNet200 [38] validation (with 200 classes)." + ], + "table_footnote": [], + "table_body": "
MethodmAP\\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{25} \\)
PointGroup [24]23.338.548.6
HAIS [6]35.146.752.8
SoftGroup [44]47.363.171.4
Mask3D [39]63.479.285.6
IKNE (Ours)64.981.287.6
", + "bbox": [ + 514, + 584, + 916, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the STPLS3D [5] test dataset.", + "bbox": [ + 511, + 696, + 906, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing $61.1 / 73.0$ (Area 5) and $65.8 / 76.9$ (6-fold) in $\\mathrm{mAP} / \\mathrm{mAP}_{50}$ . Also, in Tab. 3, ours achieves the highest scores on the ScanNet200 [38] dataset containing fine-grained 200 categories. Notably, as shown in Tab. 4, IKNE surpasses existing methods on the STPLS3D [5] dataset, including outdoor 3D scenes, with improvements of up to $+1.5 / 2.0 / 2.0$ in $\\mathrm{mAP} / \\mathrm{mAP}_{50} / \\mathrm{mAP}_{25}$ . These results highlight that we successfully handle instance candidates, improving their informativeness for understanding real-world instances from both indoor and outdoor environments. Note that “-” in each table indicates unreported scores from published papers.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14062", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/30486668d0d33bed999d583ba27cb4e630bcb91612e0b75f200192fa7035cd0e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method (Kernel-based)ScanNet ValS3DIS Area 5
Baseline (K)IKAISGmAP / mAP50mAP / mAP50
✓--56.8 / 73.356.3 / 67.5
✓✓-62.7 / 80.860.4 / 72.6
✓-✓62.8 / 81.160.6 / 71.9
✓✓✓62.9 / 81.861.1 / 73.0
", + "bbox": [ + 94, + 78, + 482, + 170 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c20976d700703fec1449f3d521e5bdfd72eb7cd4f12f1961559ce8a55de311ad.jpg", + "table_caption": [ + "Table 5. Ablation study to see the effect of two main modules (IKA and ISG) based on the kernel-based pipeline (K)." + ], + "table_footnote": [], + "table_body": "
Method (Transformer)ScanNet ValS3DIS Area 5
Baseline (T)IKAISGmAP / mAP50mAP / mAP50
✓--58.4 / 75.956.6 / 68.4
✓✓-59.1 / 76.258.8 / 70.7
✓-✓59.4 / 77.258.4 / 70.1
✓✓✓60.8 / 77.959.0 / 71.6
", + "bbox": [ + 93, + 208, + 483, + 301 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7ddc6c0d7561408e9c1c2e835a22363ab73347923fb2f706f4682f10f5cc656f.jpg", + "table_caption": [ + "Table 6. Ablation study to see the effect of two main modules (IKA and ISG) based on the transformer-based pipeline (T)." + ], + "table_footnote": [], + "table_body": "
BaselineMethodKernel-basedTransformer
IKAISGAverage IoUAverage IoU
✓--0.86270.8486
✓✓-0.88280.8728
✓-✓0.88250.8730
✓✓✓0.89040.8803
", + "bbox": [ + 94, + 339, + 483, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 481, + 254, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of IKA and ISG Networks. We evaluate variants of our method w/ and w/o IKA and ISG modules based on kernel (K) and transformer-based (T) architectures. As shown in Tab. 5 and 6, the addition of each module surpasses the score of baseline models [29, 32, 39] across all experiments, regardless of model structure. Specifically, IKA boosts performance, with gains of up to $+5.9 / 7.5$ for ScanNetV2 [7] and $+4.1 / 5.1$ for S3DIS [1] in $\\mathrm{mAP} / \\mathrm{mAP}_{50}$ , underscoring the importance of integrating scattered clues from multiple candidates. Also, ISG contributes to the spatial understanding of instance candidates via intelligent structural guidance from clarity-enhanced features, improving up to $+6.0 / 7.8$ and $+4.3 / 4.4$ . Eventually, utilizing all modules together gains further advances, resulting in the best performance with notable progress on both structures. These results demonstrate each module plays a meaningful role in improving the discriminative capability of instance features.", + "bbox": [ + 88, + 507, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, to assess how each module influences the accurate determination of instance mask coverage, we use the IoU metric, which is closely associated with segmentation performance. IoU calculates the ratio of the intersection to the union of predicted and ground truth masks. In Tab. 7, we compare the average IoU of our predicted instance masks and those of the baselines for both architectures. As shown in Fig. 1 (a) and Fig.1 (a) (in Supp.), for previous works, a single instance is often separated into sev-", + "bbox": [ + 88, + 765, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/79e068e56c904992c6d4d42c005aa1331e2eb4bafd421a2efb0292a3086faa22.jpg", + "table_caption": [ + "Table 7. Ablation study to see the effect of two main modules across both pipelines, using IoU scores on the ScanNetV2 [7] val." + ], + "table_footnote": [], + "table_body": "
ArchitectureThreshold τScanNet ValS3DIS Area 5
mAP / mAP50mAP / mAP50
Baseline (K)-56.8 / 73.356.3 / 67.5
Kernel-based0.653.6 / 73.455.3 / 65.5
Kernel-based0.757.2 / 76.459.1 / 69.2
Kernel-based0.861.1 / 79.960.7 / 72.3
Kernel-based0.962.9 / 81.861.1 / 73.0
Baseline (T)-58.4 / 75.956.6 / 68.4
Transformer0.657.9 / 75.856.3 / 67.5
Transformer0.759.2 / 76.257.4 / 68.6
Transformer0.859.9 / 77.158.6 / 70.4
Transformer0.960.8 / 77.959.0 / 71.6
", + "bbox": [ + 517, + 78, + 905, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Ablation study to investigate the threshold $\\tau$ for determining instance-wise candidates representing the same instance.", + "bbox": [ + 511, + 257, + 903, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a966361c95c2551373d737051f9b0fec27598f3f8fad077fe92df7ddef9edc1.jpg", + "image_caption": [ + "Figure 5. Comparison to explore the optimal top $t$ for reducing ambiguity while keeping important cues of instance features." + ], + "image_footnote": [], + "bbox": [ + 517, + 294, + 710, + 482 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/868289e5a6cf919f05113b6dcd340f90a55d7574a560f2e6ddffad8e790aa8d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 295, + 903, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "eral fragments. These fragments potentially cause low IoU scores. The IoU improvements with IKA across both architectures confirm its effectiveness in reducing the negative impact of multiple fragments and enhancing the accuracy of mask coverage. Also, the IoU tends to be low, particularly when instance masks include the surrounding backgrounds or when adjacent instances are not clearly distinguished, as in Fig. 1 (b) and Fig.1 (b) (in Supp.), leading to overestimated or underestimated instance masks. In addition to IKA, ISG enhances the IoU scores, verifying our guidance is sufficiently practical. We further validate the significance of ours with additional experiments in Supp. (see Sec.4).", + "bbox": [ + 511, + 532, + 906, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Thresholds $\\tau$ for Instance-wise Identification. We specify candidates representing the same instance using predefined threshold $\\tau$ from the correlation matrices $I_{ka}$ and $I_{sg}$ . In Tab. 8, based on kernel (K) and transformer (T) architectures [29, 32, 39], we explore the impact of $\\tau$ for useful knowledge interaction. When $\\tau$ is lower than 0.7, the scores degrade due to the inclusion of inaccurate information from different instance representative features. Yet, with more precise identification using the higher $\\tau$ ( $\\geqslant 0.8$ ), our knowledge interaction yields considerable improvements. This result emphasizes that clues from accurately identified candidates are precious for enhancing instance interpretation. We", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14063", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/368dcd2e9bd0908366bd5301ac51392c02a22e2858876da82abdf0f6f7933ce5.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of instance (Inst.) and semantic (Sem.) masks between the baselines (ISBNet (K) [32], MAFT (T) [29]) and ours with Ground Truth masks on the ScanNetV2. The critical differences are highlighted using yellow boxes. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 93, + 65, + 903, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "also examine $\\tau$ for the STPLS3D [5] in Supp. (Sec.4.3).", + "bbox": [ + 89, + 314, + 460, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of the Top t values. In this work, we truncate the top $t$ vectors of $(U,\\Sigma ,V)$ from $F_{inst}^{\\prime}$ and reconstruct $\\tilde{F}_{inst}^{\\prime}$ using $(\\tilde{U},\\tilde{\\Sigma},\\tilde{V})$ based on a truncated SVD algorithm [17]. Here, the $t$ value impacts the balance between data compression and information preservation. Therefore, it is crucial to find the optimal value of $t$ for proper balance. To this end, we conduct experiments using various $t$ for both kernel (blue) and transformer (orange) architectures in Fig. 5. We represent $t$ as a percentage, so the x-axis indicates the proportion of vectors retained during truncation across three matrices. We observe that the trends of all experiments are similar. In settings with a lower $t$ percentage (0.1), the scores consistently decrease due to large information loss. However, with a higher $t$ (0.7), the effect of noise reduction is insufficient; thereby, the performance is comparable to baselines [29, 32]. We ultimately set $t$ as 0.5, resulting in the highest accuracy while allowing us to preserve meaningful information and effectively filter out inherent noises.", + "bbox": [ + 88, + 330, + 483, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 614, + 272, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Comparison. In Fig. 6, we qualitatively validate the effectiveness of our novel framework IKNE. We visualize the predicted semantic (Sem.) and instance (Inst.) masks of the kernel (K) [32] and transformer (T) [29] baselines with ours on the ScanNetV2 [7]. We also provide the corresponding ground truth for solid comparisons. In Scene 1, ours accurately segments the cabinet as a single instance, whereas the baselines separate it into multiple fragments. In Scene 2, ours correctly distinguishes neighboring instances (four chairs and a table), while baselines erroneously recognize two chairs as one (K) or a table and a chair as one (T). Also, in Scene 3, where objects are disorderly adjacent, ours clearly captures the spatial range of the desk and chairs compared to the baselines. These qualitative results verify that our instance features contain highly informative cues.", + "bbox": [ + 89, + 640, + 482, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Structural Guidance from $\\tilde{\\mathbf{F}}_{\\mathrm{inst}}$ . In Fig. 7, we qualitatively confirm the significance of structural cues from the", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc58b0ff2a13755be704cf1499da2153084d4d029327480edc6d57100f8932af.jpg", + "image_caption": [ + "Figure 7. Instance feature visualizations of the original features $F_{inst}$ (blue) and ambiguity-reduced features $\\bar{F}_{inst}$ (red)." + ], + "image_footnote": [], + "bbox": [ + 516, + 311, + 900, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "noise-reduced features $\\tilde{F}_{inst}$ in smartly guiding the original features $F_{inst}$ . As shown in Scene 1 and 2, instance features (red) from $\\tilde{F}_{inst}$ accurately capture the spatial range of instances (chair and desk), whereas instance features (blue) from $F_{inst}$ struggle to cover the complete coverage. Also, in Scene 3, where the distinction is unclear (blue) from surroundings (curtain and wall), the candidate (red) from $\\tilde{F}_{inst}$ clearly segments such instances. These findings highlight guidance from clarity-enhanced features is worthwhile.", + "bbox": [ + 511, + 547, + 906, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 693, + 633, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel instance-wise knowledge enhancement approach, IKNE, for 3D instance segmentation. We focus on optimizing the efficiency of hundreds of instance candidates by effectively handling those representing the same instance. To address these candidates, we propose an IKA to integrate spread clues based on correlation regularization. Also, we present an ISG, which enhances instance features with fundamental cues for object understanding from ambiguity-reduced instance features using the truncated SVD. Our extensive experiments demonstrate the effectiveness of our methods, achieving new state-of-the-art results on both kernel and transformer architectures.", + "bbox": [ + 511, + 719, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14064", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2024 (RS-2023-00227409, Development of sketch-based semantic 3D modeling technology for creating user-centric Metaverse content spaces for indoor spaces, $90\\%$ ; RS-2024-00345025, International Collaborative Research and Global Talent Development for the Development of Copyright Management and Protection Technologies for Generative AI, $5\\%$ ), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2019-II190079, Artificial Intelligence Graduate School Program (Korea University), $1\\%$ ; IITP-2025-RS-2024-00436857, $2\\%$ ), the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2025-00521602, $2\\%$ ), and Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT (MSIT, Korea) & Gwangju Metropolitan City.", + "bbox": [ + 89, + 113, + 485, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 420, + 187, + 436 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 5, 6, 7", + "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. 2, 4", + "[3] Ori Bryt and Michael Elad. Compression of facial images using the k-svd algorithm. Journal of Visual Communication and Image Representation, 19(4):270-282, 2008. 3", + "[4] Chin-Chen Chang, Piyu Tsai, and Chia-Chen Lin. Svd-based digital image watermarking scheme. Pattern Recognition Letters, 26(10):1577-1586, 2005. 3", + "[5] Meida Chen, Qingyong Hu, Zifan Yu, Hugues Thomas, Andrew Feng, Yu Hou, Kyle McCullough, Fengbo Ren, and Lucio Soibelman. Stpls3d: A large-scale synthetic and real aerial photogrammetry 3d point cloud dataset. arXiv preprint arXiv:2203.09065, 2022. 2, 5, 6, 8", + "[6] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 1, 2, 6", + "[7] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5, 6, 7, 8", + "[8] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In" + ], + "bbox": [ + 99, + 445, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the European conference on computer vision (ECCV), pages 562-578, 2018. 4", + "[9] Renwei Dian, Shutao Li, and Xudong Kang. Regularizing hyperspectral and multispectral image fusion by cnn denoiser. IEEE transactions on neural networks and learning systems, 32(3):1124-1135, 2020. 3", + "[10] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. IEEE transactions on image processing, 6(9):1305-1315, 1997. 4", + "[11] Cathrin Elich, Francis Engelmann, Theodora Kontogianni, and Bastian Leibe. 3d bird's-eye-view instance segmentation. In German Conference on Pattern Recognition, pages 48-61. Springer, 2019. 2", + "[12] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2, 6", + "[13] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3722-3731, 2021. 2", + "[14] Gene H Golub and Christian Reinsch. Singular value decomposition and least squares solutions. In Handbook for Automatic Computation: Volume II: Linear Algebra, pages 134-151. Springer, 1971. 2, 3, 5", + "[15] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 4", + "[16] Qiang Guo, Caiming Zhang, Yunfeng Zhang, and Hui Liu. An efficient svd-based method for image denoising. IEEE transactions on Circuits and Systems for Video Technology, 26(5):868-880, 2015. 3", + "[17] Per Christian Hansen. The truncated svd as a method for regularization. BIT Numerical Mathematics, 27:534-553, 1987. 3, 8", + "[18] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2", + "[19] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3d: Robust instance segmentation of 3d point clouds through dynamic convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 354-363, 2021. 1, 2, 3, 4, 6", + "[20] Tong He, Wei Yin, Chunhua Shen, and Anton Van den Hengel. Pointinst3d: Segmenting 3d instances by points. In European Conference on Computer Vision, pages 286-302. Springer, 2022. 1, 2, 3, 4", + "[21] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 1, 2, 6" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14065", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Peg Howland, Moongu Jeon, and Haesun Park. Structure preserving dimension reduction for clustered text data based on the generalized singular value decomposition. SIAM Journal on Matrix Analysis and Applications, 25(1):165-179, 2003. 3", + "[23] Sujin Jang, Dae Ung Jo, Sung Ju Hwang, Dongwook Lee, and Daehyun Ji. Stxd: Structural and temporal cross-modal distillation for multi-view 3d object detection. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[24] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 1, 2, 6", + "[25] XiaoBing Kang and ShengMin Wei. Identifying tampered regions using singular value decomposition in digital image forensics. In 2008 International conference on computer science and software engineering, pages 926-930. IEEE, 2008. 3", + "[26] Sungjune Kim, Hyung-gun Chi, Hyerin Lim, Karthik Ramani, Jinkyu Kim, and Sangpil Kim. Higher-order relational reasoning for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15251-15260, 2024. 3", + "[27] Virginia Klema and Alan Laub. The singular value decomposition: Its computation and some applications. IEEE Transactions on automatic control, 25(2):164-176, 1980. 3", + "[28] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9256-9266, 2019. 6", + "[29] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-attention-free transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3693–3703, 2023. 1, 2, 3, 4, 6, 7, 8", + "[30] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 6", + "[31] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18516-18526, 2023. 1, 2, 4, 6", + "[32] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. Isbnet: a 3d point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13550-13559, 2023. 1, 2, 3, 4, 6, 7, 8", + "[33] HS Prasantha, HL Shashidhara, and KN Balasubramanya Murthy. Image compression using svd. In International conference on computational intelligence and multimedia applications (ICCIMA 2007), pages 143-145. IEEE, 2007. 3", + "[34] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1, 2", + "[35] Ajit Rajwade, Anand Rangarajan, and Arunava Banerjee. Image denoising using the higher order singular value decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(4):849-862, 2012. 3", + "[36] Zhongzheng Ren, Ishan Misra, Alexander G Schwing, and Rohit Girdhar. 3d spatial recognition without spatially labeled 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13213, 2021. 2", + "[37] Wonseok Roh, Hwanhee Jung, Giljoo Nam, Jinseop Yeom, Hyunjie Park, Sang Ho Yoon, and Sanggil Kim. Edge-aware 3d instance segmentation network with intelligent semantic prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20644–20653, 2024. 2, 3, 6", + "[38] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2, 5, 6", + "[39] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d for 3d semantic instance segmentation. arXiv preprint arXiv:2210.03105, 2022. 1, 2, 3, 4, 6, 7", + "[40] Qiquan Shi, Yiu-Ming Cheung, and Jian Lou. Robust tensor svd and recovery with rank estimation. IEEE Transactions on Cybernetics, 52(10):10667-10682, 2021. 3", + "[41] Sangyun Shin, Kaichen Zhou, Madhu Vankadari, Andrew Markham, and Niki Trigoni. Spherical mask: Coarse-to-fine 3d point cloud instance segmentation with spherical representation. arXiv preprint arXiv:2312.11269, 2023. 3, 6", + "[42] Gilbert W Stewart. On the early history of the singular value decomposition. SIAM review, 35(4):551-566, 1993. 2, 3, 5", + "[43] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 1, 2, 6", + "[44] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 6", + "[45] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2569-2578, 2018. 1, 2", + "[46] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2", + "[47] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scenegraphfusion: Incremental 3d scene graph prediction from rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7515-7525, 2021. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14066", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Yizheng Wu, Min Shi, Shuaiyuan Du, Hao Lu, Zhiguo Cao, and Weicai Zhong. 3d instances as 1d kernels. In European Conference on Computer Vision, pages 235-252. Springer, 2022. 1, 2, 4, 6", + "[49] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Dening Lu, Mingqiang Wei, and Jun Wang. Venet: Voting enhancement network for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3712-3721, 2021. 2", + "[50] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds. Advances in neural information processing systems, 32, 2019. 1, 2, 6", + "[51] Dan Yang, Zongming Ma, and Andreas Buja. A sparse singular value decomposition method for high-dimensional data. Journal of Computational and Graphical Statistics, 23 (4):923-942, 2014. 3", + "[52] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3947-3956, 2019. 1, 2", + "[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pages 12310-12320. PMLR, 2021. 2, 4", + "[54] Min Zhong, Xinghao Chen, Xiaokang Chen, Gang Zeng, and Yunhe Wang. Maskgroup: Hierarchical point grouping and masking for 3d instance segmentation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 6" + ], + "bbox": [ + 91, + 92, + 482, + 541 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14067", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Insightful Instance Features for 3D Instance Segmentation/275c7e3b-7085-4434-b01d-e129139897fa_model.json b/2025/Insightful Instance Features for 3D Instance Segmentation/275c7e3b-7085-4434-b01d-e129139897fa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..921faa7a111009e3d7b8b18e01e73a6b773df69e --- /dev/null +++ b/2025/Insightful Instance Features for 3D Instance Segmentation/275c7e3b-7085-4434-b01d-e129139897fa_model.json @@ -0,0 +1,2136 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.131, + 0.795, + 0.154 + ], + "angle": 0, + "content": "Insightful Instance Features for 3D Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.18, + 0.796, + 0.245 + ], + "angle": 0, + "content": "Wonseok Roh\\(^{1}\\) Hwanhee Jung\\(^{1}\\) Giljoo Nam\\(^{2}\\) Dong In Lee\\(^{1}\\) \nHyeongcheol Park\\(^{1}\\) Sang Ho Yoon\\(^{3}\\) Jungseock Joo\\(^{4}\\) Sangpil Kim\\(^{1*}\\) \n\\(^{1}\\)Korea University \\(^{2}\\)Meta \\(^{3}\\)KAIST \\(^{4}\\)UCLA" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.329, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.305, + 0.486, + 0.669 + ], + "angle": 0, + "content": "Recent 3D Instance Segmentation methods typically encode hundreds of instance-wise candidates with instancespecific information in various ways and refine them into final masks. However, they have yet to fully explore the benefit of these candidates. They overlook the valuable cues encoded in multiple candidates that represent different parts of the same instance, resulting in fragments. Also, they often fail to capture the precise spatial range of 3D instances, primarily due to inherent noises from sparse and unordered point clouds. In this work, to address these challenges, we propose IKNE, a novel instance-wise knowledge enhancement approach. We first introduce an Instance-wise Knowledge Aggregation (IKA) to associate scattered single instance details by optimizing correlations among candidates representing the same instance. Moreover, we present an Instance-wise Structural Guidance (ISG) to enhance the spatial understanding of candidates using structural cues from ambiguity-reduced features. Here, we utilize a simple yet effective truncated singular value decomposition algorithm to minimize inherent noises of 3D features. In our extensive experiments on large-scale datasets, ScanNetV2, ScanNet200, S3DIS, and STPLS3D, IKNE outperforms existing works. We validate the effectiveness of our modules in both kernel-based and transformer-based architectures." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.687, + 0.224, + 0.702 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.485, + 0.879 + ], + "angle": 0, + "content": "3D Instance Segmentation (3DIS) is a fundamental 3D computer vision task that significantly contributes to real-world applications. Given 3D point cloud scenes, the 3DIS task identifies respective instances and assigns semantic class labels with the goal of comprehensively understanding entire spatial environments. In real-life 3D scenarios, substantial occlusion and truncation commonly arise, especially when objects are overlapped or obscured by others. To tackle these challenges, early approaches mainly focused on accurately generating region proposals (top-down) [21, 50, 52] or effectively grouping points with clus" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.336 + ], + "angle": 0, + "content": "tering algorithms (bottom-up) [6, 24, 45]. However, they presuppose that the intermediate processes, such as bounding box detection [18] or heuristic voting mechanism [34], generate near perfect results, which is often not the case." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.909, + 0.672 + ], + "angle": 0, + "content": "Recently, kernel-based [19, 20, 32, 48] and transformer-based [29, 31, 39, 43] 3DIS approaches have been proposed, aiming to overcome the limitations of traditional frameworks. Kernel-based methods leverage instance-aware kernels for dynamic convolution, which decodes instance masks. They represent instances as kernels, replacing clustering algorithms with point sampling processes. On the other hand, transformer-based methods train instance queries to identify features about individual objects using attention mechanisms. These queries enable the direct prediction of per-point categories and instance labels. In both architectures, models utilize hundreds of instance-wise candidate features, to estimate the final instance mask. For example, ISBNet [32] (kernel-based) produces 256 instance kernels, and MAFT [29] (transformer-based) leverages 400 instance queries. However, despite the significant improvements based on their architectural advantages, they have yet to fully optimize the handling of these numerous candidates and still face several challenges, as illustrated in Fig. 1. To tackle these challenges, we introduce two main modules, IKA and ISG, each thoughtfully designed to enhance the informativeness of instance candidate features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.903 + ], + "angle": 0, + "content": "First, we empirically observe that multiple candidates from previous works often encode cues from the same single instance, as shown in Fig. 1 (a). These candidates are then typically decoded into considerable fragment masks representing different components of each instance, such as the backrest or armrest of the sofa. In this work, to estimate the complete coverage of each instance, we introduce an Instance-wise Knowledge Aggregation (IKA) network, which integrates the scattered intelligence from multiple fragments beyond relying solely on a few fragments based on confidence scores. Specifically, we optimize the correlations among instance-wise candidate features, encouraging the network to be aware of the distributed instance-specific knowledge. We initially identify candidates representing the same instance based on the similarity of their features, as" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.341, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author (spk7@korea.ac.kr)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14057" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.075, + 0.5, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.248, + 0.434, + 0.261 + ], + "angle": 0, + "content": "(a) Per-point masks of instance-wise candidates" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.075, + 0.905, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.248, + 0.802, + 0.262 + ], + "angle": 0, + "content": "(b) Instance misinterpretation cases" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.272, + 0.907, + 0.33 + ], + "angle": 0, + "content": "Figure 1. Examples of two challenging cases of 3DIS. (a) Per-point masks of instance-wise candidates: highly correlated instance-wise candidates usually represent incomplete fragments of the same single instance, such as the backrest or armrest of a sofa. And (b) Instance misinterpretation cases: existing studies often confuse instances from backgrounds (Scene1) or misunderstand the spatial range and appearance of instances (Scene2). Note that we also provide additional examples for both cases in the supplementary material (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.482, + 0.463 + ], + "angle": 0, + "content": "highly correlated candidates usually capture the information from the same instance. Then, inspired by correlation regularization mechanisms [2, 23, 53], we softly guide identified candidates to be closer to each other in the latent space, aiming to enhance solidarity among them. As a result, IKA facilitates the network to learn relationships between candidates and aggregate their comprehensive details, maximizing the advantages of using hundreds of candidates." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.465, + 0.483, + 0.869 + ], + "angle": 0, + "content": "In addition to IKA, we explore additional avenues to enrich the knowledge of instance-wise candidates. Unlike the human visual cognitive system, which perceives instances with sufficient prior knowledge, the network understands instances depending on their external properties from point features. However, point features are frequently unstable, primarily due to ambiguous inherent noises [13, 36, 47, 49] caused by sparse and incomplete point clouds. These noises are likely to confuse the models, making it hard to precisely interpret the specific spatial range and unique appearance of instances, as shown in Fig. 1 (b). To mitigate the negative impact of such noises, we propose an Instance-wise Structural Guidance (ISG) network specifically designed to improve candidates' structural understanding of 3D instances from point features. Here, we utilize the truncated singular value decomposition (SVD) [14, 42] algorithm to strengthen essential cues (e.g., shape) of instance features while explicitly reducing fuzzy noises. Then, we effectively transfer fundamental clues for object cognition from ambiguity-minimized features to corresponding original features based on the cross-correlations between them. Ultimately, our instance features are now highly informative across complex 3D instances with aggregated and structural knowledge from IKA and ISG. Our novel modules are applicable to frameworks employing considerable instance candidates, regardless of their underlying model structures, including kernel-based and transformer-based architectures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Given landmark datasets for 3DIS, ScanNetV2 [7], ScanNet200 [38], S3DIS [1], and STPLS3D [5], we vali" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.905, + 0.387 + ], + "angle": 0, + "content": "date the effectiveness of our novel framework. Our method outperforms the existing state-of-the-art methods. To summarize, our main contributions are listed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.39, + 0.905, + 0.466 + ], + "angle": 0, + "content": "- To extend the benefits of producing multiple instance-wise candidates, we introduce the Instance-wise Knowledge Aggregation (IKA) network, which associates scattered instance-specific information of the same single instance by optimizing correlations among them." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.466, + 0.905, + 0.542 + ], + "angle": 0, + "content": "- We design the Instance-wise Structural Guidance (ISG) network to improve the structural knowledge of candidates. Specifically, we use a simple yet strong truncated singular value decomposition process to emphasize essential cues from point features while filtering out noises." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.542, + 0.905, + 0.601 + ], + "angle": 0, + "content": "- We analyze the effectiveness of our framework, IKNE, on four challenging datasets. Comprehensive experiments in various 3D scenarios demonstrate that our novel method achieves new State-of-the-Art performance on 3DIS." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.39, + 0.905, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.616, + 0.655, + 0.632 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.903 + ], + "angle": 0, + "content": "3D Instance Segmentation (3DIS) aims to distinguish individual instances within 3D scenes and assign corresponding semantic categories. Typically, 3DIS approaches can be categorized into four types: proposal-based [21, 50, 52], grouping-based [6, 11, 12, 24, 45], kernel-based [19, 20, 32, 48], and transformer-based [29, 31, 37, 39, 43]. Based on the impressive achievements of detection methods [18, 46], proposal-based approaches first detect instances and employ proposed regions as hard references to predict masks. However, these strategies, such as 3D-SIS [21] or 3D-BoNet [50], heavily rely on the quality of the outputs from the detection process. On the other hand, grouping-based methods catch 3D instances through a bottom-up pipeline, aggregating closely related points into instances using predicted semantic categories and center offsets. Yet, they still depend on intermediate manually tuned processing, such as point grouping [24] or heuristic voting mechanisms [34]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14058" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.075, + 0.907, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.294, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Figure 2. An overview of our framework IKNE. Built upon the modern kernel-based structure, our model consists of four main modules: (1) Sparse Convolutional 3D Backbone; (2) Instance-wise Knowledge Aggregation (IKA, Sec. 3.2), which softly associates scattered cues from highly correlated instance features of \\( F_{inst} \\); (3) Instance-wise Structural Guidance (ISG, Sec. 3.3), which smartly instructs \\( F_{inst} \\) with essential clues from \\( \\tilde{F}_{inst} \\), using an effective SVD algorithm; and (4) Dynamic Convolution, which produces \\( \\hat{M}_f \\) from kernels \\( K \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.361, + 0.482, + 0.588 + ], + "angle": 0, + "content": "To address these limitations, kernel-based approaches have been introduced. For example, DyCo3D [19] employs a clustering algorithm to generate kernels for dynamic convolutions to predict instance masks, while ISBNet [32] presents a cluster-free method using instance-wise kernels. Recently, Spherical Mask [41], based on ISBNet, has overcome the low-quality results of coarse-to-fine strategies by using spherical representation to mitigate the propagation of false negatives and overestimation of instance size. Furthermore, transformer-based methods, Mask3D [39], MAFT [29], and EASE [37] utilize instance-wise queries to encode individual instance information through attention mechanisms. In this work, we focus on improving the quality of numerous instance-wise candidates introduced from recent studies with two novel approaches, IKA and ISG." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.483, + 0.696 + ], + "angle": 0, + "content": "Singular Value Decomposition (SVD) is a fundamental algorithm in linear algebra, widely used for matrix factorization [14, 27, 42], dimensionality reduction [22, 51], and 2D image processing [9, 16, 26, 35, 40]. SVD decomposes a matrix \\(A \\in \\mathbb{R}^{m \\times n}\\) into the dot product of three matrices, \\(U \\in \\mathbb{R}^{m \\times m}\\), \\(\\Sigma \\in \\mathbb{R}^{m \\times n}\\), and \\(V \\in \\mathbb{R}^{n \\times n}\\), each containing meaningful vectors and scalar values, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.7, + 0.482, + 0.717 + ], + "angle": 0, + "content": "\\[\nA = U \\cdot \\Sigma \\cdot V ^ {T} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.856 + ], + "angle": 0, + "content": "where \\(U\\) and \\(V\\) are the matrices with orthonormal columns, satisfying \\(U^T U = V^T V = I\\), and \\(\\Sigma\\) is a diagonal matrix with nonnegative singular values \\(\\sigma\\) in descending order as \\(\\sigma_1 \\geqslant \\sigma_2 \\geqslant \\dots \\geqslant \\sigma_r\\). Based on SVD, truncated SVD [17] is an insightful dimension-reduction technique that minimizes nonessential noises while highlighting essential cues from the original matrix \\(A\\). It produces a low-rank approximation \\(\\tilde{A} \\in \\mathbb{R}^{m \\times n}\\) by retaining the most important top \\(k\\) singular values and their corresponding vectors from \\(U\\), \\(\\Sigma\\), and \\(V\\):" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.859, + 0.482, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\tilde {A} = U _ {k} \\cdot \\Sigma_ {k} \\cdot V _ {k} ^ {T} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.885, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\(U_{k}\\in \\mathbb{R}^{m\\times k}\\) \\(\\Sigma_{k}\\in \\mathbb{R}^{k\\times k}\\) , and \\(V_{k}\\in \\mathbb{R}^{n\\times k}\\) are the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.361, + 0.884, + 0.376 + ], + "angle": 0, + "content": "truncated matrices with the top \\(k\\) singular value vectors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.907, + 0.619 + ], + "angle": 0, + "content": "In the field of 2D image processing, SVD has been extensively utilized for low-rank approximation. [16, 35] employ SVD to factorize and estimate low-rank patches, effectively reducing noise and enhancing image quality. Similarly, [4, 25] operate SVD to detect tampering by extracting and analyzing low-dimensional representations of images for image forensics. Also, SVD showcases robust performance in image compression [3, 33] and recovery [40]. Further, [9] leverages SVD within CNNs to learn subspace information for hyperspectral image reconstruction. However, despite its potential benefits, the introduction of SVD has received limited attention in 3D vision, especially for point clouds. Based on these observations, we apply SVD to address point features that contain inherent noises due to the unordered properties of point clouds, facilitating the model to become more robust against complex ambiguities." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.631, + 0.605, + 0.645 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.906, + 0.732 + ], + "angle": 0, + "content": "In this section, we introduce a novel 3D Instance Segmentation framework IKNE. We provide an overview of the whole pipeline (Sec. 3.1 and Fig. 2), followed by detailed technical discussions: Instance-wise Knowledge Aggregation (Sec. 3.2) and Instance-wise Structural Guidance (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.625, + 0.756 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Recent 3DIS methods [19, 20, 32, 37, 39] commonly follow a similar paradigm; they spread hundreds of instance candidates across 3D scenes to capture instance-specific information in various ways and refine them into final instance masks. In this work, we encourage these numerous instance candidates to be highly informative, improving their understanding of diverse real-world instances. We describe our instance-wise knowledge enhancement approaches based on the modern kernel-based architecture [32]. As shown in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14059" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "Fig. 2, our model consists of four main modules: (1) Sparse Convolutional 3D Backbone; (2) Instance-wise Knowledge Aggregation, which relates the scattered information of individual instances; (3) Instance-wise Structural Guidance, which effectively instructs candidates with spatial cues from augmented instance features; and (4) Dynamic Convolution network. Note that we also describe our transformer-based architecture in the supplementary material (Sec.3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.218, + 0.484, + 0.448 + ], + "angle": 0, + "content": "Kernel-based 3DIS Architecture. First, sparse convolutional U-Net backbone [15] takes a colored point cloud \\( P \\in \\mathbb{R}^{N_p \\times 6} \\) as input and voxelizes \\( P \\) into voxels to extract point-wise feature maps \\( F_{p} \\in \\mathbb{R}^{N_{p} \\times D} \\). Given the point features, the point aggregator samples a set of instance candidate features, referring to point positions based on the Farthest Point Sampling (FPS) [10]. Specifically, we employ two-stage point aggregator blocks to produce optimal instance features across 3D scenes. In the first stage, \\( N_{k}^{\\prime} \\) number of candidates \\( F_{inst}^{\\prime} \\in \\mathbb{R}^{N_{k}^{\\prime} \\times D} \\) are sampled from full-resolution feature \\( F_{p} \\), while in the second stage, \\( N_{k} \\) (\\( < N_{k}^{\\prime} \\)) instance features \\( F_{inst} \\in \\mathbb{R}^{N_{k} \\times D} \\) are sampled based on the \\( F_{inst}^{\\prime} \\). We then classify the category \\( C_{i} \\) for \\( i = \\{1,2,\\dots,N_{c}\\} \\) of each instance from \\( F_{inst} \\) via linear classification head \\( f_{\\mathrm{cls}} \\) using cross-entropy loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.456, + 0.484, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = - \\mathbb {E} _ {c, w _ {c} \\sim \\mathbb {D}} \\left[ \\sum_ {r \\in N _ {c}} w _ {c} [ r ] \\log f _ {\\mathrm {c l s}} \\left(F _ {\\text {i n s t}}\\right) [ r ] \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.662 + ], + "angle": 0, + "content": "where \\(w_{c}\\) indicates one-hot encoded category labels and \\(\\mathbb{D}\\) represents (input) data distribution. Also, we predict instance kernels \\(K\\in \\mathbb{R}^{N_k\\times D}\\) based on \\(F_{inst}\\) through MLPs. Furthermore, each point-wise prediction head processes \\(F_{p}\\) to generate mask features \\(F_{mask}\\in \\mathbb{R}^{N_p\\times D}\\) and axis-aligned bounding boxes \\(F_{box}\\in \\mathbb{R}^{N_p\\times 6}\\) as auxiliary information for mask prediction. Finally, the dynamic convolution network produces the final instance mask \\(\\hat{M}_f\\) using instance kernels \\(K\\) and \\(F_{mask}\\). To optimize \\(\\hat{M}_f\\), we formulate the following loss \\(\\mathcal{L}_{mask}\\) as a sum of the two losses:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.674, + 0.483, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {\\mathrm {B C E}} \\mathcal {L} _ {\\mathrm {B C E}} \\left(M _ {f}, \\hat {M} _ {f}\\right) + \\lambda_ {\\mathrm {d i c e}} \\mathcal {L} _ {\\mathrm {d i c e}} \\left(M _ {f}, \\hat {M} _ {f}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.734 + ], + "angle": 0, + "content": "where \\(M_{f}\\) is the ground-truth mask, and \\(\\mathcal{L}_{\\mathrm{BCE}}\\) and \\(\\mathcal{L}_{\\mathrm{dice}}\\) denote binary cross-entropy and Dice losses [8], respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.472, + 0.759 + ], + "angle": 0, + "content": "3.2. Instance-wise Knowledge Aggregation (IKA)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Recent studies [19, 20, 29, 31, 32, 39, 48] in 3DIS deal with instance-wise features \\( F_{inst} \\) as essential components, aiming to encode rich semantic and geometric information about instances. They spread hundreds of candidates over expansive 3D scenes to cover as many instances as possible; this number \\( N_{k} \\) far exceeds the total number \\( N_{i} \\) of instances present in any single scene (see Tab.4 in Supp.). Although these candidates provide practical information for mask prediction, they often represent separate parts of each instance," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.278, + 0.907, + 0.335 + ], + "angle": 0, + "content": "Figure 3. In IKA, we specify highly correlated instance candidates \\((I_{ka} > \\tau)\\) representing identical instances based on the correlation matrix \\(I_{ka}\\). Then, we softly aggregate scattered clues of related candidates, optimizing \\(I_{ka}\\) with pseudo-binary label \\(I_{ka}^{*}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.905, + 0.391 + ], + "angle": 0, + "content": "resulting in fragmented instance masks. With these insights, we focus on associating scattered knowledge to fully leverage the advantages of multiple instance candidates." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.391, + 0.907, + 0.709 + ], + "angle": 0, + "content": "To handle multiple candidates, we first identify those representing the same instance based on their pairwise affinity. Specifically, we compute a correlation matrix \\( I_{ka} \\in \\mathbb{R}^{N_k \\times N_k} \\) using the dot product of \\( F_{inst} \\) and its transpose \\( F_{inst}^T \\), then normalize \\( I_{ka} \\) to a range between 0 and 1. Here, we note that candidates representing the same instance yield relatively high correlation values. Thus, based on \\( I_{ka} \\), we consider pairwise candidates with correlation values surpassing a predefined threshold \\( \\tau \\) as fragments from the same instance. Then, as shown in Fig. 3, we softly guide these candidates by optimizing their correlations in \\( I_{ka} \\), motivated by the success of recent self-supervised learning strategies [2, 53]. We extend this approach to further enrich coherence among highly correlated candidate features while also encouraging the network to be aware of scattered cues for each instance. In particular, we dynamically construct a pseudo-binary correlation label matrix \\( I_{ka}^* \\in \\{0,1\\}^{N_k \\times N_k} \\), where elements of highly correlated pairs \\( (\\tau) \\) in \\( I_{ka} \\) are 1, while the rest \\( (\\leqslant \\tau) \\) are 0 for all candidate pairs. Given \\( I_{ka} \\) and \\( I_{ka}^* \\), we formulate the element-wise correlation regularization loss \\( \\mathcal{L}_{KA} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.71, + 0.907, + 0.751 + ], + "angle": 0, + "content": "\\[\nI _ {k a} = F _ {i n s t} \\cdot F _ {i n s t} ^ {T}, I _ {k a} ^ {*} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} I _ {k a} (i, j) > \\tau \\\\ 0 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.761, + 0.906, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {K A} = \\sum_ {I _ {k a} > \\tau} \\| I _ {k a} ^ {*} (i, j) - I _ {k a} (i, j) \\| _ {2} \\\\ + w \\cdot \\sum_ {I _ {k a} \\leqslant \\tau} \\| I _ {k a} ^ {*} (i, j) - I _ {k a} (i, j) \\| _ {2} \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\| _2\\) denotes the \\(L_{2}\\) norm, and \\(w\\) is the weight value used to balance between highly correlated pairs and others. In Eq. 6, the first term facilitates candidates representing the identical instance closer together in the latent space," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14060" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.169 + ], + "angle": 0, + "content": "whereas the second term decreases correlations between unrelated candidates, thereby reducing confusion from irrelevant knowledge. In conclusion, IKA establishes practical connections that strengthen solidarity across multiple candidates encoding scattered cues by regularizing loss \\(\\mathcal{L}_{KA}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.175, + 0.442, + 0.192 + ], + "angle": 0, + "content": "3.3. Instance-wise Structural Guidance (ISG)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.484, + 0.35 + ], + "angle": 0, + "content": "In real-world scenarios, 3D instances are usually placed near each other or overlapping, arranged in inconsistent patterns without following conventional standards. Also, these instances are represented as sparse point clouds, which often induce fuzzy inherent noises in point-wise features. Hence, the instance-wise features \\( F_{inst} \\) from these settings are unclear, causing models to misinterpret the spatial extent of each instance. In this work, we explicitly reduce such ambiguity of \\( F_{inst} \\) by leveraging intelligent structural guidance from clarity-enhanced instance-wise feature \\( \\tilde{F}_{inst} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.484, + 0.504 + ], + "angle": 0, + "content": "For accurate segmentation, it is crucial to capture the clear morphological shape of complex instances. From this intuition, we aim to encourage the model to better learn the structural appearance of instances. Thus, we implement effective structural guidance based on a truncated singular value decomposition (SVD) [14, 42] algorithm primarily used for noise reduction. As illustrated in Fig. 4, given \\( F_{inst}^{\\prime} \\in \\mathbb{R}^{N_k^{\\prime} \\times D} \\) from the first point aggregator, we initially decompose \\( F_{inst}^{\\prime} \\) into three matrices based on SVD: (1) \\( U \\in \\mathbb{R}^{N_k^{\\prime} \\times N_k^{\\prime}} \\), (2) \\( \\Sigma \\in \\mathbb{R}^{N_k^{\\prime} \\times D} \\), and (3) \\( V \\in \\mathbb{R}^{D \\times D} \\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.509, + 0.483, + 0.55 + ], + "angle": 0, + "content": "\\[\nF _ {i n s t} ^ {\\prime} = U \\cdot \\Sigma \\cdot V ^ {T} = \\sum_ {i} ^ {N _ {k} ^ {\\prime}} u _ {i} \\cdot \\sigma_ {i} \\cdot v _ {i} ^ {T} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.484, + 0.724 + ], + "angle": 0, + "content": "where \\(U = (u_{1},\\ldots ,u_{n})\\) and \\(V = (v_{1},\\dots ,v_{n})\\) are the matrices with orthonormal columns, and \\(\\Sigma = diag(\\sigma_1,\\dots,\\sigma_n)\\) is a diagonal matrix containing singular values, arranged in the descending order. Typically, vectors with higher singular values in \\(\\Sigma\\) include meaningful information for object cognition, while the lower-rank vectors are regarded as less important, potentially causing ambiguity. Then, we preserve only the top \\(t\\) singular values from \\(\\Sigma\\), along with their corresponding singular vectors from \\(U\\) and \\(V\\), facilitating an ambiguity-reduced yet information-rich representation of the original features \\(F_{inst}^{\\prime}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.73, + 0.483, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\tilde {F} _ {i n s t} ^ {\\prime} \\in \\mathbb {R} ^ {N _ {k} ^ {\\prime} \\times D} = \\tilde {U} \\cdot \\tilde {\\Sigma} \\cdot \\tilde {V} ^ {T} = \\sum_ {i} ^ {N _ {k} ^ {\\prime}} \\tilde {u} _ {i} \\cdot \\tilde {\\sigma} _ {i} \\cdot \\tilde {v} _ {i} ^ {T} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.484, + 0.904 + ], + "angle": 0, + "content": "where \\(\\tilde{U} \\in \\mathbb{R}^{N_k' \\times t}\\), \\(\\tilde{\\Sigma} \\in \\mathbb{R}^{t \\times t}\\), and \\(\\tilde{V} \\in \\mathbb{R}^{D \\times t}\\) are truncated matrices using only \\(t\\) columns from \\(U\\), \\(\\Sigma\\), and \\(V\\), then \\(\\tilde{F}_{inst}'\\) is reconstructed from them. Based on both \\(F_{inst}'\\) and \\(\\tilde{F}_{inst}'\\), we sample instance-wise features \\(F_{inst}\\) and \\(F_{inst}\\), respectively, employing the shared second stage point aggregator. Given \\(F_{inst}\\) and \\(\\tilde{F}_{inst}\\), we calculate the cross-correlation matrix \\(I_{sg} \\in \\mathbb{R}^{N_k \\times N_k}\\) and also generate a pseudo-binary correlation label \\(I_{sg}^* \\in \\{0, 1\\}^{N_k \\times N_k}\\), as in the IKA network." + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.084, + 0.646, + 0.095 + ], + "angle": 0, + "content": "Original Instance Feats." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.095, + 0.907, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.271, + 0.907, + 0.328 + ], + "angle": 0, + "content": "Figure 4. In ISG, we decompose \\( F_{inst}^{\\prime} \\) into three matrices \\( (U, \\Sigma, V) \\) using the SVD algorithm and keep top \\( t \\) singular vectors to filter out less important noises. Then, we effectively enhance \\( F_{inst} \\) with highlighted essential cues from \\( \\tilde{F}_{inst} \\) by optimizing \\( \\mathcal{L}_{SG} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.906, + 0.365 + ], + "angle": 0, + "content": "We then effectively transfer structural clues from \\( \\tilde{F}_{inst} \\), utilizing the correlation regularization loss \\( \\mathcal{L}_{SG} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.368, + 0.907, + 0.408 + ], + "angle": 0, + "content": "\\[\nI _ {s g} = F _ {i n s t} \\cdot \\tilde {F} _ {i n s t} ^ {T}, I _ {s g} ^ {*} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} I _ {s g} (i, j) > \\tau \\\\ 0 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.414, + 0.906, + 0.483 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {S G} = \\sum_ {I _ {s g} > \\tau} \\left| \\left| I _ {s g} ^ {*} (i, j) - I _ {s g} (i, j) \\right| \\right| _ {2} \\\\ + \\tilde {w} \\cdot \\sum_ {I _ {s g} \\leqslant \\tau} \\left\\| I _ {s g} ^ {*} (i, j) - I _ {s g} (i, j) \\right\\| _ {2} \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.491, + 0.906, + 0.582 + ], + "angle": 0, + "content": "where \\(\\tilde{w}\\) represents the balancing hyper-parameter, and the role of each term in Eq. 10 follows Eq. 6. By optimizing \\(\\mathcal{L}_{SG}\\), ISG conditionally instructs instance features \\(F_{inst}\\) to incorporate spatial configuration details from clarity-enhanced features \\(\\tilde{F}_{inst}\\), improving discriminative power. Loss Function. Finally, our full network minimizes \\(\\mathcal{L}_{total}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.594, + 0.905, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} \\\\ + \\lambda_ {K A} \\mathcal {L} _ {K A} + \\lambda_ {S G} \\mathcal {L} _ {S G} \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.639, + 0.906, + 0.67 + ], + "angle": 0, + "content": "where each \\(\\lambda\\) is a hyper-parameter decided via extensive grid searches to balance the impact of respective loss terms." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.646, + 0.699 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.706, + 0.704, + 0.723 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Datasets. We train and evaluate the overall performance using four landmark benchmarks for 3D Instance Segmentation (3DIS): ScanNetV2 [7], ScanNet200 [38], S3DIS [1], and STPLS3D [5]. These four datasets provide 3D point cloud data collected in diverse real-world environments. Note that detailed descriptions of each dataset and all implementation details are provided in Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate the 3DIS performance using the Average Precision (AP), a conventional metric in computer vision tasks. We report the mean average precision (mAP) across IoU (Intersection of Union) thresholds" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14061" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.079, + 0.487, + 0.391 + ], + "angle": 0, + "content": "
MethodScanNetV2 ValScanNetV2 Test
mAP\\( mAP_{50} \\)\\( mAP_{25} \\)mAP\\( mAP_{50} \\)\\( mAP_{25} \\)
3D-SIS [21]-18.735.716.138.255.8
MTML [28]20.340.255.428.254.973.1
PointGroup [24]34.856.771.340.763.677.8
DyCo3D [19]35.457.672.939.564.176.1
3D-MPA [12]35.559.172.435.561.173.7
MaskGroup [54]42.063.363.343.466.479.2
SoftGroup [44]46.067.678.950.476.186.5
SSTNet [30]49.464.374.050.669.878.9
Mask3D [39]55.273.782.956.678.082.9
SPFormer [43]56.373.982.954.977.085.1
Q-Former [31]56.574.283.358.378.787.4
ISBNet [32]56.873.381.355.976.387.0
MAFT [29]58.475.984.559.678.686.0
EASE [37]60.277.284.259.878.986.7
SphMask [41]62.379.988.261.681.287.5
IKNE (Ours)62.981.888.761.781.288.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.397, + 0.484, + 0.426 + ], + "angle": 0, + "content": "Table 1. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the ScanNetV2 [7] dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.514 + ], + "angle": 0, + "content": "incremented by \\(5\\%\\), ranging from \\(50\\%\\) to \\(95\\%\\). Also, we assess \\(\\mathrm{mAP}_{50}\\) and \\(\\mathrm{mAP}_{25}\\), which measure accuracy based on IoU thresholds of \\(50\\%\\) and \\(25\\%\\). For the S3DIS [1], we further provide mean precision (mPrec) and mean recall (mRec) at a \\(50\\%\\) IoU threshold, following previous works." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.546 + ], + "angle": 0, + "content": "4.2. Performance Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.554, + 0.487, + 0.903 + ], + "angle": 0, + "content": "In this section, we quantitatively compare our proposed framework IKNE with previous methods. Despite significant advances with high scores, these methods, regardless of architecture, still waste considerable instance candidates that may contain valuable information. To maximize the benefits of using numerous candidates, we present novel instance-wise knowledge aggregation (IKA) and guidance (ISG) networks. In Tab. 1, we first evaluate mean average precision (mAP) with different IoU thresholds across 18 classes on the ScanNetV2 [7] dataset. IKNE generally outperforms prior works, achieving new state-of-the-art scores in terms of mAP, \\(\\mathrm{mAP}_{50}\\), and \\(\\mathrm{mAP}_{25}\\) on both validation and hidden test sets. While our method slightly exceeds the latest coarse-to-fine approach, Spherical Mask [41], ours transcends our primary baselines, which rely on hundreds of instance candidates, such as ISBNet [32] (kernel-based) and MAFT [29] (transformer), by large margins across all metrics. In Tab. 2, we assess 3DIS performance on Area 5 and the 6-fold cross-validation set of the S3DIS [1] dataset. For Area 5 evaluation, we train with Areas 1 to 6, excluding 5, and validate with Area 5; for 6-fold cross-validation, we average the validation scores across all 6 areas. IKNE demonstrates impressive performance in both evaluations, reaching" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.079, + 0.913, + 0.391 + ], + "angle": 0, + "content": "
MethodS3DIS Area 5S3DIS 6-fold CV
AP\\( AP_{50} \\)\\( Prec_{50} \\)\\( Rec_{50} \\)AP\\( AP_{50} \\)\\( Prec_{50} \\)\\( Rec_{50} \\)
3D-Bonet [50]--57.540.2--65.647.6
PointGroup [24]-57.861.962.1-64.069.669.2
3D-MPA [12]--63.158.0--66.764.1
DyCo3D [19]--64.364.2----
MaskGroup [54]-65.062.964.7-69.966.669.6
SSTNet [30]42.759.365.564.254.167.873.573.4
SoftGroup [44]51.666.173.666.654.468.975.369.8
SPFormer [43]-66.872.867.1-69.274.071.1
ISBNet [32]56.367.570.572.060.870.577.577.1
DKNet [48]--70.865.3--75.371.1
Mask3D [39]56.668.468.766.364.575.572.874.5
Q-Former [31]57.769.970.572.262.073.372.773.4
MAFT [29]-69.1------
EASE [37]59.071.669.468.765.376.473.674.6
SphMask [41]60.572.371.376.364.072.378.177.7
IKNE (Ours)61.173.072.077.065.876.979.478.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.397, + 0.907, + 0.426 + ], + "angle": 0, + "content": "Table 2. Comparison of 3DIS performance with state-of-the-art approaches on the S3DIS [1] Area 5 and 6-fold cross validation." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.436, + 0.9, + 0.543 + ], + "angle": 0, + "content": "
MethodmAP\\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{25} \\)
SPFormer [43]25.233.839.6
Mask3D [39]27.437.042.3
Q-Former [31]28.137.143.4
MAFT [29]29.238.243.3
IKNE (Ours)29.938.944.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.547, + 0.907, + 0.576 + ], + "angle": 0, + "content": "Table 3. Comparison of 3DIS performance with state-of-the-art approaches on the ScanNet200 [38] validation (with 200 classes)." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.585, + 0.918, + 0.692 + ], + "angle": 0, + "content": "
MethodmAP\\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{25} \\)
PointGroup [24]23.338.548.6
HAIS [6]35.146.752.8
SoftGroup [44]47.363.171.4
Mask3D [39]63.479.285.6
IKNE (Ours)64.981.287.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.697, + 0.907, + 0.726 + ], + "angle": 0, + "content": "Table 4. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the STPLS3D [5] test dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "ing \\(61.1 / 73.0\\) (Area 5) and \\(65.8 / 76.9\\) (6-fold) in \\(\\mathrm{mAP} / \\mathrm{mAP}_{50}\\). Also, in Tab. 3, ours achieves the highest scores on the ScanNet200 [38] dataset containing fine-grained 200 categories. Notably, as shown in Tab. 4, IKNE surpasses existing methods on the STPLS3D [5] dataset, including outdoor 3D scenes, with improvements of up to \\(+1.5 / 2.0 / 2.0\\) in \\(\\mathrm{mAP} / \\mathrm{mAP}_{50} / \\mathrm{mAP}_{25}\\). These results highlight that we successfully handle instance candidates, improving their informativeness for understanding real-world instances from both indoor and outdoor environments. Note that “-” in each table indicates unreported scores from published papers." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14062" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.079, + 0.483, + 0.171 + ], + "angle": 0, + "content": "
Method (Kernel-based)ScanNet ValS3DIS Area 5
Baseline (K)IKAISGmAP / mAP50mAP / mAP50
✓--56.8 / 73.356.3 / 67.5
✓✓-62.7 / 80.860.4 / 72.6
✓-✓62.8 / 81.160.6 / 71.9
✓✓✓62.9 / 81.861.1 / 73.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.173, + 0.483, + 0.203 + ], + "angle": 0, + "content": "Table 5. Ablation study to see the effect of two main modules (IKA and ISG) based on the kernel-based pipeline (K)." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.209, + 0.485, + 0.302 + ], + "angle": 0, + "content": "
Method (Transformer)ScanNet ValS3DIS Area 5
Baseline (T)IKAISGmAP / mAP50mAP / mAP50
✓--58.4 / 75.956.6 / 68.4
✓✓-59.1 / 76.258.8 / 70.7
✓-✓59.4 / 77.258.4 / 70.1
✓✓✓60.8 / 77.959.0 / 71.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.305, + 0.483, + 0.334 + ], + "angle": 0, + "content": "Table 6. Ablation study to see the effect of two main modules (IKA and ISG) based on the transformer-based pipeline (T)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.34, + 0.485, + 0.438 + ], + "angle": 0, + "content": "
BaselineMethodKernel-basedTransformer
IKAISGAverage IoUAverage IoU
✓--0.86270.8486
✓✓-0.88280.8728
✓-✓0.88250.8730
✓✓✓0.89040.8803
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.441, + 0.483, + 0.47 + ], + "angle": 0, + "content": "Table 7. Ablation study to see the effect of two main modules across both pipelines, using IoU scores on the ScanNetV2 [7] val." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.482, + 0.255, + 0.497 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Effect of IKA and ISG Networks. We evaluate variants of our method w/ and w/o IKA and ISG modules based on kernel (K) and transformer-based (T) architectures. As shown in Tab. 5 and 6, the addition of each module surpasses the score of baseline models [29, 32, 39] across all experiments, regardless of model structure. Specifically, IKA boosts performance, with gains of up to \\(+5.9 / 7.5\\) for ScanNetV2 [7] and \\(+4.1 / 5.1\\) for S3DIS [1] in \\(\\mathrm{mAP} / \\mathrm{mAP}_{50}\\), underscoring the importance of integrating scattered clues from multiple candidates. Also, ISG contributes to the spatial understanding of instance candidates via intelligent structural guidance from clarity-enhanced features, improving up to \\(+6.0 / 7.8\\) and \\(+4.3 / 4.4\\). Eventually, utilizing all modules together gains further advances, resulting in the best performance with notable progress on both structures. These results demonstrate each module plays a meaningful role in improving the discriminative capability of instance features." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.766, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Furthermore, to assess how each module influences the accurate determination of instance mask coverage, we use the IoU metric, which is closely associated with segmentation performance. IoU calculates the ratio of the intersection to the union of predicted and ground truth masks. In Tab. 7, we compare the average IoU of our predicted instance masks and those of the baselines for both architectures. As shown in Fig. 1 (a) and Fig.1 (a) (in Supp.), for previous works, a single instance is often separated into sev-" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.079, + 0.906, + 0.253 + ], + "angle": 0, + "content": "
ArchitectureThreshold τScanNet ValS3DIS Area 5
mAP / mAP50mAP / mAP50
Baseline (K)-56.8 / 73.356.3 / 67.5
Kernel-based0.653.6 / 73.455.3 / 65.5
Kernel-based0.757.2 / 76.459.1 / 69.2
Kernel-based0.861.1 / 79.960.7 / 72.3
Kernel-based0.962.9 / 81.861.1 / 73.0
Baseline (T)-58.4 / 75.956.6 / 68.4
Transformer0.657.9 / 75.856.3 / 67.5
Transformer0.759.2 / 76.257.4 / 68.6
Transformer0.859.9 / 77.158.6 / 70.4
Transformer0.960.8 / 77.959.0 / 71.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.258, + 0.905, + 0.287 + ], + "angle": 0, + "content": "Table 8. Ablation study to investigate the threshold \\( \\tau \\) for determining instance-wise candidates representing the same instance." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.295, + 0.712, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.296, + 0.904, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.489, + 0.907, + 0.518 + ], + "angle": 0, + "content": "Figure 5. Comparison to explore the optimal top \\( t \\) for reducing ambiguity while keeping important cues of instance features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.907, + 0.716 + ], + "angle": 0, + "content": "eral fragments. These fragments potentially cause low IoU scores. The IoU improvements with IKA across both architectures confirm its effectiveness in reducing the negative impact of multiple fragments and enhancing the accuracy of mask coverage. Also, the IoU tends to be low, particularly when instance masks include the surrounding backgrounds or when adjacent instances are not clearly distinguished, as in Fig. 1 (b) and Fig.1 (b) (in Supp.), leading to overestimated or underestimated instance masks. In addition to IKA, ISG enhances the IoU scores, verifying our guidance is sufficiently practical. We further validate the significance of ours with additional experiments in Supp. (see Sec.4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Thresholds \\(\\tau\\) for Instance-wise Identification. We specify candidates representing the same instance using predefined threshold \\(\\tau\\) from the correlation matrices \\(I_{ka}\\) and \\(I_{sg}\\). In Tab. 8, based on kernel (K) and transformer (T) architectures [29, 32, 39], we explore the impact of \\(\\tau\\) for useful knowledge interaction. When \\(\\tau\\) is lower than 0.7, the scores degrade due to the inclusion of inaccurate information from different instance representative features. Yet, with more precise identification using the higher \\(\\tau\\) (\\(\\geqslant 0.8\\)), our knowledge interaction yields considerable improvements. This result emphasizes that clues from accurately identified candidates are precious for enhancing instance interpretation. We" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14063" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.066, + 0.905, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.278, + 0.908, + 0.307 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of instance (Inst.) and semantic (Sem.) masks between the baselines (ISBNet (K) [32], MAFT (T) [29]) and ours with Ground Truth masks on the ScanNetV2. The critical differences are highlighted using yellow boxes. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.461, + 0.33 + ], + "angle": 0, + "content": "also examine \\(\\tau\\) for the STPLS3D [5] in Supp. (Sec.4.3)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.332, + 0.484, + 0.605 + ], + "angle": 0, + "content": "Analysis of the Top t values. In this work, we truncate the top \\(t\\) vectors of \\((U,\\Sigma ,V)\\) from \\(F_{inst}^{\\prime}\\) and reconstruct \\(\\tilde{F}_{inst}^{\\prime}\\) using \\((\\tilde{U},\\tilde{\\Sigma},\\tilde{V})\\) based on a truncated SVD algorithm [17]. Here, the \\(t\\) value impacts the balance between data compression and information preservation. Therefore, it is crucial to find the optimal value of \\(t\\) for proper balance. To this end, we conduct experiments using various \\(t\\) for both kernel (blue) and transformer (orange) architectures in Fig. 5. We represent \\(t\\) as a percentage, so the x-axis indicates the proportion of vectors retained during truncation across three matrices. We observe that the trends of all experiments are similar. In settings with a lower \\(t\\) percentage (0.1), the scores consistently decrease due to large information loss. However, with a higher \\(t\\) (0.7), the effect of noise reduction is insufficient; thereby, the performance is comparable to baselines [29, 32]. We ultimately set \\(t\\) as 0.5, resulting in the highest accuracy while allowing us to preserve meaningful information and effectively filter out inherent noises." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.615, + 0.274, + 0.631 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.867 + ], + "angle": 0, + "content": "Visual Comparison. In Fig. 6, we qualitatively validate the effectiveness of our novel framework IKNE. We visualize the predicted semantic (Sem.) and instance (Inst.) masks of the kernel (K) [32] and transformer (T) [29] baselines with ours on the ScanNetV2 [7]. We also provide the corresponding ground truth for solid comparisons. In Scene 1, ours accurately segments the cabinet as a single instance, whereas the baselines separate it into multiple fragments. In Scene 2, ours correctly distinguishes neighboring instances (four chairs and a table), while baselines erroneously recognize two chairs as one (K) or a table and a chair as one (T). Also, in Scene 3, where objects are disorderly adjacent, ours clearly captures the spatial range of the desk and chairs compared to the baselines. These qualitative results verify that our instance features contain highly informative cues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Structural Guidance from \\(\\tilde{\\mathbf{F}}_{\\mathrm{inst}}\\). In Fig. 7, we qualitatively confirm the significance of structural cues from the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.313, + 0.901, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.51, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Figure 7. Instance feature visualizations of the original features \\( F_{inst} \\) (blue) and ambiguity-reduced features \\( \\bar{F}_{inst} \\) (red)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.907, + 0.685 + ], + "angle": 0, + "content": "noise-reduced features \\(\\tilde{F}_{inst}\\) in smartly guiding the original features \\(F_{inst}\\). As shown in Scene 1 and 2, instance features (red) from \\(\\tilde{F}_{inst}\\) accurately capture the spatial range of instances (chair and desk), whereas instance features (blue) from \\(F_{inst}\\) struggle to cover the complete coverage. Also, in Scene 3, where the distinction is unclear (blue) from surroundings (curtain and wall), the candidate (red) from \\(\\tilde{F}_{inst}\\) clearly segments such instances. These findings highlight guidance from clarity-enhanced features is worthwhile." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.694, + 0.634, + 0.709 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this paper, we introduce a novel instance-wise knowledge enhancement approach, IKNE, for 3D instance segmentation. We focus on optimizing the efficiency of hundreds of instance candidates by effectively handling those representing the same instance. To address these candidates, we propose an IKA to integrate spread clues based on correlation regularization. Also, we present an ISG, which enhances instance features with fundamental cues for object understanding from ambiguity-reduced instance features using the truncated SVD. Our extensive experiments demonstrate the effectiveness of our methods, achieving new state-of-the-art results on both kernel and transformer architectures." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14064" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.486, + 0.395 + ], + "angle": 0, + "content": "This work was supported by the Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2024 (RS-2023-00227409, Development of sketch-based semantic 3D modeling technology for creating user-centric Metaverse content spaces for indoor spaces, \\(90\\%\\); RS-2024-00345025, International Collaborative Research and Global Talent Development for the Development of Copyright Management and Protection Technologies for Generative AI, \\(5\\%\\)), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2019-II190079, Artificial Intelligence Graduate School Program (Korea University), \\(1\\%\\); IITP-2025-RS-2024-00436857, \\(2\\%\\)), the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2025-00521602, \\(2\\%\\)), and Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT (MSIT, Korea) & Gwangju Metropolitan City." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.421, + 0.188, + 0.437 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.446, + 0.485, + 0.516 + ], + "angle": 0, + "content": "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.517, + 0.484, + 0.572 + ], + "angle": 0, + "content": "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[3] Ori Bryt and Michael Elad. Compression of facial images using the k-svd algorithm. Journal of Visual Communication and Image Representation, 19(4):270-282, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.658 + ], + "angle": 0, + "content": "[4] Chin-Chen Chang, Piyu Tsai, and Chia-Chen Lin. Svd-based digital image watermarking scheme. Pattern Recognition Letters, 26(10):1577-1586, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[5] Meida Chen, Qingyong Hu, Zifan Yu, Hugues Thomas, Andrew Feng, Yu Hou, Kyle McCullough, Fengbo Ren, and Lucio Soibelman. Stpls3d: A large-scale synthetic and real aerial photogrammetry 3d point cloud dataset. arXiv preprint arXiv:2203.09065, 2022. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[6] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[7] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[8] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.446, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the European conference on computer vision (ECCV), pages 562-578, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.121, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[9] Renwei Dian, Shutao Li, and Xudong Kang. Regularizing hyperspectral and multispectral image fusion by cnn denoiser. IEEE transactions on neural networks and learning systems, 32(3):1124-1135, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[10] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. IEEE transactions on image processing, 6(9):1305-1315, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[11] Cathrin Elich, Francis Engelmann, Theodora Kontogianni, and Bastian Leibe. 3d bird's-eye-view instance segmentation. In German Conference on Pattern Recognition, pages 48-61. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[12] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[13] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3722-3731, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[14] Gene H Golub and Christian Reinsch. Singular value decomposition and least squares solutions. In Handbook for Automatic Computation: Volume II: Linear Algebra, pages 134-151. Springer, 1971. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[15] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[16] Qiang Guo, Caiming Zhang, Yunfeng Zhang, and Hui Liu. An efficient svd-based method for image denoising. IEEE transactions on Circuits and Systems for Video Technology, 26(5):868-880, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[17] Per Christian Hansen. The truncated svd as a method for regularization. BIT Numerical Mathematics, 27:534-553, 1987. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[18] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.908, + 0.789 + ], + "angle": 0, + "content": "[19] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3d: Robust instance segmentation of 3d point clouds through dynamic convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 354-363, 2021. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[20] Tong He, Wei Yin, Chunhua Shen, and Anton Van den Hengel. Pointinst3d: Segmenting 3d instances by points. In European Conference on Computer Vision, pages 286-302. Springer, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14065" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[22] Peg Howland, Moongu Jeon, and Haesun Park. Structure preserving dimension reduction for clustered text data based on the generalized singular value decomposition. SIAM Journal on Matrix Analysis and Applications, 25(1):165-179, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[23] Sujin Jang, Dae Ung Jo, Sung Ju Hwang, Dongwook Lee, and Daehyun Ji. Stxd: Structural and temporal cross-modal distillation for multi-view 3d object detection. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[24] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.354 + ], + "angle": 0, + "content": "[25] XiaoBing Kang and ShengMin Wei. Identifying tampered regions using singular value decomposition in digital image forensics. In 2008 International conference on computer science and software engineering, pages 926-930. IEEE, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[26] Sungjune Kim, Hyung-gun Chi, Hyerin Lim, Karthik Ramani, Jinkyu Kim, and Sangpil Kim. Higher-order relational reasoning for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15251-15260, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[27] Virginia Klema and Alan Laub. The singular value decomposition: Its computation and some applications. IEEE Transactions on automatic control, 25(2):164-176, 1980. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[28] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9256-9266, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[29] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-attention-free transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3693–3703, 2023. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[30] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[31] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18516-18526, 2023. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[32] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. Isbnet: a 3d point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13550-13559, 2023. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[33] HS Prasantha, HL Shashidhara, and KN Balasubramanya Murthy. Image compression using svd. In International conference on computational intelligence and multimedia applications (ICCIMA 2007), pages 143-145. IEEE, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[34] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[35] Ajit Rajwade, Anand Rangarajan, and Arunava Banerjee. Image denoising using the higher order singular value decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(4):849-862, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[36] Zhongzheng Ren, Ishan Misra, Alexander G Schwing, and Rohit Girdhar. 3d spatial recognition without spatially labeled 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13213, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[37] Wonseok Roh, Hwanhee Jung, Giljoo Nam, Jinseop Yeom, Hyunjie Park, Sang Ho Yoon, and Sanggil Kim. Edge-aware 3d instance segmentation network with intelligent semantic prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20644–20653, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[38] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[39] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d for 3d semantic instance segmentation. arXiv preprint arXiv:2210.03105, 2022. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[40] Qiquan Shi, Yiu-Ming Cheung, and Jian Lou. Robust tensor svd and recovery with rank estimation. IEEE Transactions on Cybernetics, 52(10):10667-10682, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[41] Sangyun Shin, Kaichen Zhou, Madhu Vankadari, Andrew Markham, and Niki Trigoni. Spherical mask: Coarse-to-fine 3d point cloud instance segmentation with spherical representation. arXiv preprint arXiv:2312.11269, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[42] Gilbert W Stewart. On the early history of the singular value decomposition. SIAM review, 35(4):551-566, 1993. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[43] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[44] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[45] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2569-2578, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[46] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scenegraphfusion: Incremental 3d scene graph prediction from rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7515-7525, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14066" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[48] Yizheng Wu, Min Shi, Shuaiyuan Du, Hao Lu, Zhiguo Cao, and Weicai Zhong. 3d instances as 1d kernels. In European Conference on Computer Vision, pages 235-252. Springer, 2022. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[49] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Dening Lu, Mingqiang Wei, and Jun Wang. Venet: Voting enhancement network for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3712-3721, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[50] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds. Advances in neural information processing systems, 32, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[51] Dan Yang, Zongming Ma, and Andreas Buja. A sparse singular value decomposition method for high-dimensional data. Journal of Computational and Graphical Statistics, 23 (4):923-942, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[52] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3947-3956, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pages 12310-12320. PMLR, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[54] Min Zhong, Xinghao Chen, Xiaokang Chen, Gang Zeng, and Yunhe Wang. Maskgroup: Hierarchical point grouping and masking for 3d instance segmentation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14067" + } + ] +] \ No newline at end of file diff --git a/2025/Insightful Instance Features for 3D Instance Segmentation/275c7e3b-7085-4434-b01d-e129139897fa_origin.pdf b/2025/Insightful Instance Features for 3D Instance Segmentation/275c7e3b-7085-4434-b01d-e129139897fa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e7e296c6e2aea32609b23b4464c4334d56db5bbd --- /dev/null +++ b/2025/Insightful Instance Features for 3D Instance Segmentation/275c7e3b-7085-4434-b01d-e129139897fa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d13eb724c6012b310dbc3b0438e54eb1d65346b337207234707a6a942ef00488 +size 7843298 diff --git a/2025/Insightful Instance Features for 3D Instance Segmentation/full.md b/2025/Insightful Instance Features for 3D Instance Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..db2303407d4bc7794cbaa8e8b4b1e166ae3a494f --- /dev/null +++ b/2025/Insightful Instance Features for 3D Instance Segmentation/full.md @@ -0,0 +1,303 @@ +# Insightful Instance Features for 3D Instance Segmentation + +Wonseok Roh $^{1}$ Hwanhee Jung $^{1}$ Giljoo Nam $^{2}$ Dong In Lee $^{1}$ +Hyeongcheol Park $^{1}$ Sang Ho Yoon $^{3}$ Jungseock Joo $^{4}$ Sangpil Kim $^{1*}$ $^{1}$ Korea University $^{2}$ Meta $^{3}$ KAIST $^{4}$ UCLA + +# Abstract + +Recent 3D Instance Segmentation methods typically encode hundreds of instance-wise candidates with instancespecific information in various ways and refine them into final masks. However, they have yet to fully explore the benefit of these candidates. They overlook the valuable cues encoded in multiple candidates that represent different parts of the same instance, resulting in fragments. Also, they often fail to capture the precise spatial range of 3D instances, primarily due to inherent noises from sparse and unordered point clouds. In this work, to address these challenges, we propose IKNE, a novel instance-wise knowledge enhancement approach. We first introduce an Instance-wise Knowledge Aggregation (IKA) to associate scattered single instance details by optimizing correlations among candidates representing the same instance. Moreover, we present an Instance-wise Structural Guidance (ISG) to enhance the spatial understanding of candidates using structural cues from ambiguity-reduced features. Here, we utilize a simple yet effective truncated singular value decomposition algorithm to minimize inherent noises of 3D features. In our extensive experiments on large-scale datasets, ScanNetV2, ScanNet200, S3DIS, and STPLS3D, IKNE outperforms existing works. We validate the effectiveness of our modules in both kernel-based and transformer-based architectures. + +# 1. Introduction + +3D Instance Segmentation (3DIS) is a fundamental 3D computer vision task that significantly contributes to real-world applications. Given 3D point cloud scenes, the 3DIS task identifies respective instances and assigns semantic class labels with the goal of comprehensively understanding entire spatial environments. In real-life 3D scenarios, substantial occlusion and truncation commonly arise, especially when objects are overlapped or obscured by others. To tackle these challenges, early approaches mainly focused on accurately generating region proposals (top-down) [21, 50, 52] or effectively grouping points with clus + +tering algorithms (bottom-up) [6, 24, 45]. However, they presuppose that the intermediate processes, such as bounding box detection [18] or heuristic voting mechanism [34], generate near perfect results, which is often not the case. + +Recently, kernel-based [19, 20, 32, 48] and transformer-based [29, 31, 39, 43] 3DIS approaches have been proposed, aiming to overcome the limitations of traditional frameworks. Kernel-based methods leverage instance-aware kernels for dynamic convolution, which decodes instance masks. They represent instances as kernels, replacing clustering algorithms with point sampling processes. On the other hand, transformer-based methods train instance queries to identify features about individual objects using attention mechanisms. These queries enable the direct prediction of per-point categories and instance labels. In both architectures, models utilize hundreds of instance-wise candidate features, to estimate the final instance mask. For example, ISBNet [32] (kernel-based) produces 256 instance kernels, and MAFT [29] (transformer-based) leverages 400 instance queries. However, despite the significant improvements based on their architectural advantages, they have yet to fully optimize the handling of these numerous candidates and still face several challenges, as illustrated in Fig. 1. To tackle these challenges, we introduce two main modules, IKA and ISG, each thoughtfully designed to enhance the informativeness of instance candidate features. + +First, we empirically observe that multiple candidates from previous works often encode cues from the same single instance, as shown in Fig. 1 (a). These candidates are then typically decoded into considerable fragment masks representing different components of each instance, such as the backrest or armrest of the sofa. In this work, to estimate the complete coverage of each instance, we introduce an Instance-wise Knowledge Aggregation (IKA) network, which integrates the scattered intelligence from multiple fragments beyond relying solely on a few fragments based on confidence scores. Specifically, we optimize the correlations among instance-wise candidate features, encouraging the network to be aware of the distributed instance-specific knowledge. We initially identify candidates representing the same instance based on the similarity of their features, as + +![](images/30a9a2e2005cbac10a94477050abc7efac50968382e685ce4335538ed8967ea2.jpg) +(a) Per-point masks of instance-wise candidates + +![](images/4f5a6da64acca8cbfad76fc499301524148b1bec0e9c30e97c2c73d325ad47a4.jpg) +(b) Instance misinterpretation cases +Figure 1. Examples of two challenging cases of 3DIS. (a) Per-point masks of instance-wise candidates: highly correlated instance-wise candidates usually represent incomplete fragments of the same single instance, such as the backrest or armrest of a sofa. And (b) Instance misinterpretation cases: existing studies often confuse instances from backgrounds (Scene1) or misunderstand the spatial range and appearance of instances (Scene2). Note that we also provide additional examples for both cases in the supplementary material (Fig. 1). + +highly correlated candidates usually capture the information from the same instance. Then, inspired by correlation regularization mechanisms [2, 23, 53], we softly guide identified candidates to be closer to each other in the latent space, aiming to enhance solidarity among them. As a result, IKA facilitates the network to learn relationships between candidates and aggregate their comprehensive details, maximizing the advantages of using hundreds of candidates. + +In addition to IKA, we explore additional avenues to enrich the knowledge of instance-wise candidates. Unlike the human visual cognitive system, which perceives instances with sufficient prior knowledge, the network understands instances depending on their external properties from point features. However, point features are frequently unstable, primarily due to ambiguous inherent noises [13, 36, 47, 49] caused by sparse and incomplete point clouds. These noises are likely to confuse the models, making it hard to precisely interpret the specific spatial range and unique appearance of instances, as shown in Fig. 1 (b). To mitigate the negative impact of such noises, we propose an Instance-wise Structural Guidance (ISG) network specifically designed to improve candidates' structural understanding of 3D instances from point features. Here, we utilize the truncated singular value decomposition (SVD) [14, 42] algorithm to strengthen essential cues (e.g., shape) of instance features while explicitly reducing fuzzy noises. Then, we effectively transfer fundamental clues for object cognition from ambiguity-minimized features to corresponding original features based on the cross-correlations between them. Ultimately, our instance features are now highly informative across complex 3D instances with aggregated and structural knowledge from IKA and ISG. Our novel modules are applicable to frameworks employing considerable instance candidates, regardless of their underlying model structures, including kernel-based and transformer-based architectures. + +Given landmark datasets for 3DIS, ScanNetV2 [7], ScanNet200 [38], S3DIS [1], and STPLS3D [5], we vali + +date the effectiveness of our novel framework. Our method outperforms the existing state-of-the-art methods. To summarize, our main contributions are listed as follows: + +- To extend the benefits of producing multiple instance-wise candidates, we introduce the Instance-wise Knowledge Aggregation (IKA) network, which associates scattered instance-specific information of the same single instance by optimizing correlations among them. +- We design the Instance-wise Structural Guidance (ISG) network to improve the structural knowledge of candidates. Specifically, we use a simple yet strong truncated singular value decomposition process to emphasize essential cues from point features while filtering out noises. +- We analyze the effectiveness of our framework, IKNE, on four challenging datasets. Comprehensive experiments in various 3D scenarios demonstrate that our novel method achieves new State-of-the-Art performance on 3DIS. + +# 2. Related Work + +3D Instance Segmentation (3DIS) aims to distinguish individual instances within 3D scenes and assign corresponding semantic categories. Typically, 3DIS approaches can be categorized into four types: proposal-based [21, 50, 52], grouping-based [6, 11, 12, 24, 45], kernel-based [19, 20, 32, 48], and transformer-based [29, 31, 37, 39, 43]. Based on the impressive achievements of detection methods [18, 46], proposal-based approaches first detect instances and employ proposed regions as hard references to predict masks. However, these strategies, such as 3D-SIS [21] or 3D-BoNet [50], heavily rely on the quality of the outputs from the detection process. On the other hand, grouping-based methods catch 3D instances through a bottom-up pipeline, aggregating closely related points into instances using predicted semantic categories and center offsets. Yet, they still depend on intermediate manually tuned processing, such as point grouping [24] or heuristic voting mechanisms [34]. + +![](images/919a97be83018fe598f7553d2c6c20071c8295fe45b50e6b4b650bc5b690841e.jpg) +Figure 2. An overview of our framework IKNE. Built upon the modern kernel-based structure, our model consists of four main modules: (1) Sparse Convolutional 3D Backbone; (2) Instance-wise Knowledge Aggregation (IKA, Sec. 3.2), which softly associates scattered cues from highly correlated instance features of $F_{inst}$ ; (3) Instance-wise Structural Guidance (ISG, Sec. 3.3), which smartly instructs $F_{inst}$ with essential clues from $\tilde{F}_{inst}$ , using an effective SVD algorithm; and (4) Dynamic Convolution, which produces $\hat{M}_f$ from kernels $K$ . + +To address these limitations, kernel-based approaches have been introduced. For example, DyCo3D [19] employs a clustering algorithm to generate kernels for dynamic convolutions to predict instance masks, while ISBNet [32] presents a cluster-free method using instance-wise kernels. Recently, Spherical Mask [41], based on ISBNet, has overcome the low-quality results of coarse-to-fine strategies by using spherical representation to mitigate the propagation of false negatives and overestimation of instance size. Furthermore, transformer-based methods, Mask3D [39], MAFT [29], and EASE [37] utilize instance-wise queries to encode individual instance information through attention mechanisms. In this work, we focus on improving the quality of numerous instance-wise candidates introduced from recent studies with two novel approaches, IKA and ISG. + +Singular Value Decomposition (SVD) is a fundamental algorithm in linear algebra, widely used for matrix factorization [14, 27, 42], dimensionality reduction [22, 51], and 2D image processing [9, 16, 26, 35, 40]. SVD decomposes a matrix $A \in \mathbb{R}^{m \times n}$ into the dot product of three matrices, $U \in \mathbb{R}^{m \times m}$ , $\Sigma \in \mathbb{R}^{m \times n}$ , and $V \in \mathbb{R}^{n \times n}$ , each containing meaningful vectors and scalar values, as follows: + +$$ +A = U \cdot \Sigma \cdot V ^ {T} \tag {1} +$$ + +where $U$ and $V$ are the matrices with orthonormal columns, satisfying $U^T U = V^T V = I$ , and $\Sigma$ is a diagonal matrix with nonnegative singular values $\sigma$ in descending order as $\sigma_1 \geqslant \sigma_2 \geqslant \dots \geqslant \sigma_r$ . Based on SVD, truncated SVD [17] is an insightful dimension-reduction technique that minimizes nonessential noises while highlighting essential cues from the original matrix $A$ . It produces a low-rank approximation $\tilde{A} \in \mathbb{R}^{m \times n}$ by retaining the most important top $k$ singular values and their corresponding vectors from $U$ , $\Sigma$ , and $V$ : + +$$ +\tilde {A} = U _ {k} \cdot \Sigma_ {k} \cdot V _ {k} ^ {T} \tag {2} +$$ + +where $U_{k}\in \mathbb{R}^{m\times k}$ $\Sigma_{k}\in \mathbb{R}^{k\times k}$ , and $V_{k}\in \mathbb{R}^{n\times k}$ are the + +truncated matrices with the top $k$ singular value vectors. + +In the field of 2D image processing, SVD has been extensively utilized for low-rank approximation. [16, 35] employ SVD to factorize and estimate low-rank patches, effectively reducing noise and enhancing image quality. Similarly, [4, 25] operate SVD to detect tampering by extracting and analyzing low-dimensional representations of images for image forensics. Also, SVD showcases robust performance in image compression [3, 33] and recovery [40]. Further, [9] leverages SVD within CNNs to learn subspace information for hyperspectral image reconstruction. However, despite its potential benefits, the introduction of SVD has received limited attention in 3D vision, especially for point clouds. Based on these observations, we apply SVD to address point features that contain inherent noises due to the unordered properties of point clouds, facilitating the model to become more robust against complex ambiguities. + +# 3. Method + +In this section, we introduce a novel 3D Instance Segmentation framework IKNE. We provide an overview of the whole pipeline (Sec. 3.1 and Fig. 2), followed by detailed technical discussions: Instance-wise Knowledge Aggregation (Sec. 3.2) and Instance-wise Structural Guidance (Sec. 3.3). + +# 3.1. Overview + +Recent 3DIS methods [19, 20, 32, 37, 39] commonly follow a similar paradigm; they spread hundreds of instance candidates across 3D scenes to capture instance-specific information in various ways and refine them into final instance masks. In this work, we encourage these numerous instance candidates to be highly informative, improving their understanding of diverse real-world instances. We describe our instance-wise knowledge enhancement approaches based on the modern kernel-based architecture [32]. As shown in + +Fig. 2, our model consists of four main modules: (1) Sparse Convolutional 3D Backbone; (2) Instance-wise Knowledge Aggregation, which relates the scattered information of individual instances; (3) Instance-wise Structural Guidance, which effectively instructs candidates with spatial cues from augmented instance features; and (4) Dynamic Convolution network. Note that we also describe our transformer-based architecture in the supplementary material (Sec.3). + +Kernel-based 3DIS Architecture. First, sparse convolutional U-Net backbone [15] takes a colored point cloud $P \in \mathbb{R}^{N_p \times 6}$ as input and voxelizes $P$ into voxels to extract point-wise feature maps $F_{p} \in \mathbb{R}^{N_{p} \times D}$ . Given the point features, the point aggregator samples a set of instance candidate features, referring to point positions based on the Farthest Point Sampling (FPS) [10]. Specifically, we employ two-stage point aggregator blocks to produce optimal instance features across 3D scenes. In the first stage, $N_{k}^{\prime}$ number of candidates $F_{inst}^{\prime} \in \mathbb{R}^{N_{k}^{\prime} \times D}$ are sampled from full-resolution feature $F_{p}$ , while in the second stage, $N_{k}$ ( $< N_{k}^{\prime}$ ) instance features $F_{inst} \in \mathbb{R}^{N_{k} \times D}$ are sampled based on the $F_{inst}^{\prime}$ . We then classify the category $C_{i}$ for $i = \{1,2,\dots,N_{c}\}$ of each instance from $F_{inst}$ via linear classification head $f_{\mathrm{cls}}$ using cross-entropy loss as follows: + +$$ +\mathcal {L} _ {c l s} = - \mathbb {E} _ {c, w _ {c} \sim \mathbb {D}} \left[ \sum_ {r \in N _ {c}} w _ {c} [ r ] \log f _ {\mathrm {c l s}} \left(F _ {\text {i n s t}}\right) [ r ] \right] \tag {3} +$$ + +where $w_{c}$ indicates one-hot encoded category labels and $\mathbb{D}$ represents (input) data distribution. Also, we predict instance kernels $K\in \mathbb{R}^{N_k\times D}$ based on $F_{inst}$ through MLPs. Furthermore, each point-wise prediction head processes $F_{p}$ to generate mask features $F_{mask}\in \mathbb{R}^{N_p\times D}$ and axis-aligned bounding boxes $F_{box}\in \mathbb{R}^{N_p\times 6}$ as auxiliary information for mask prediction. Finally, the dynamic convolution network produces the final instance mask $\hat{M}_f$ using instance kernels $K$ and $F_{mask}$ . To optimize $\hat{M}_f$ , we formulate the following loss $\mathcal{L}_{mask}$ as a sum of the two losses: + +$$ +\mathcal {L} _ {\text {m a s k}} = \lambda_ {\mathrm {B C E}} \mathcal {L} _ {\mathrm {B C E}} \left(M _ {f}, \hat {M} _ {f}\right) + \lambda_ {\mathrm {d i c e}} \mathcal {L} _ {\mathrm {d i c e}} \left(M _ {f}, \hat {M} _ {f}\right) \tag {4} +$$ + +where $M_{f}$ is the ground-truth mask, and $\mathcal{L}_{\mathrm{BCE}}$ and $\mathcal{L}_{\mathrm{dice}}$ denote binary cross-entropy and Dice losses [8], respectively. + +# 3.2. Instance-wise Knowledge Aggregation (IKA) + +Recent studies [19, 20, 29, 31, 32, 39, 48] in 3DIS deal with instance-wise features $F_{inst}$ as essential components, aiming to encode rich semantic and geometric information about instances. They spread hundreds of candidates over expansive 3D scenes to cover as many instances as possible; this number $N_{k}$ far exceeds the total number $N_{i}$ of instances present in any single scene (see Tab.4 in Supp.). Although these candidates provide practical information for mask prediction, they often represent separate parts of each instance, + +![](images/300b9787ccc7195a5588487da0f6cc05465952974eb66aa20a386dc31e0d56e9.jpg) +Figure 3. In IKA, we specify highly correlated instance candidates $(I_{ka} > \tau)$ representing identical instances based on the correlation matrix $I_{ka}$ . Then, we softly aggregate scattered clues of related candidates, optimizing $I_{ka}$ with pseudo-binary label $I_{ka}^{*}$ . + +resulting in fragmented instance masks. With these insights, we focus on associating scattered knowledge to fully leverage the advantages of multiple instance candidates. + +To handle multiple candidates, we first identify those representing the same instance based on their pairwise affinity. Specifically, we compute a correlation matrix $I_{ka} \in \mathbb{R}^{N_k \times N_k}$ using the dot product of $F_{inst}$ and its transpose $F_{inst}^T$ , then normalize $I_{ka}$ to a range between 0 and 1. Here, we note that candidates representing the same instance yield relatively high correlation values. Thus, based on $I_{ka}$ , we consider pairwise candidates with correlation values surpassing a predefined threshold $\tau$ as fragments from the same instance. Then, as shown in Fig. 3, we softly guide these candidates by optimizing their correlations in $I_{ka}$ , motivated by the success of recent self-supervised learning strategies [2, 53]. We extend this approach to further enrich coherence among highly correlated candidate features while also encouraging the network to be aware of scattered cues for each instance. In particular, we dynamically construct a pseudo-binary correlation label matrix $I_{ka}^* \in \{0,1\}^{N_k \times N_k}$ , where elements of highly correlated pairs $(\tau)$ in $I_{ka}$ are 1, while the rest $(\leqslant \tau)$ are 0 for all candidate pairs. Given $I_{ka}$ and $I_{ka}^*$ , we formulate the element-wise correlation regularization loss $\mathcal{L}_{KA}$ as follows: + +$$ +I _ {k a} = F _ {i n s t} \cdot F _ {i n s t} ^ {T}, I _ {k a} ^ {*} (i, j) = \left\{ \begin{array}{l l} 1 & \text {i f} I _ {k a} (i, j) > \tau \\ 0 & \text {o t h e r w i s e ,} \end{array} \right. \tag {5} +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {K A} = \sum_ {I _ {k a} > \tau} \| I _ {k a} ^ {*} (i, j) - I _ {k a} (i, j) \| _ {2} \\ + w \cdot \sum_ {I _ {k a} \leqslant \tau} \| I _ {k a} ^ {*} (i, j) - I _ {k a} (i, j) \| _ {2} \tag {6} \\ \end{array} +$$ + +where $\| \cdot \| _2$ denotes the $L_{2}$ norm, and $w$ is the weight value used to balance between highly correlated pairs and others. In Eq. 6, the first term facilitates candidates representing the identical instance closer together in the latent space, + +whereas the second term decreases correlations between unrelated candidates, thereby reducing confusion from irrelevant knowledge. In conclusion, IKA establishes practical connections that strengthen solidarity across multiple candidates encoding scattered cues by regularizing loss $\mathcal{L}_{KA}$ . + +# 3.3. Instance-wise Structural Guidance (ISG) + +In real-world scenarios, 3D instances are usually placed near each other or overlapping, arranged in inconsistent patterns without following conventional standards. Also, these instances are represented as sparse point clouds, which often induce fuzzy inherent noises in point-wise features. Hence, the instance-wise features $F_{inst}$ from these settings are unclear, causing models to misinterpret the spatial extent of each instance. In this work, we explicitly reduce such ambiguity of $F_{inst}$ by leveraging intelligent structural guidance from clarity-enhanced instance-wise feature $\tilde{F}_{inst}$ . + +For accurate segmentation, it is crucial to capture the clear morphological shape of complex instances. From this intuition, we aim to encourage the model to better learn the structural appearance of instances. Thus, we implement effective structural guidance based on a truncated singular value decomposition (SVD) [14, 42] algorithm primarily used for noise reduction. As illustrated in Fig. 4, given $F_{inst}^{\prime} \in \mathbb{R}^{N_k^{\prime} \times D}$ from the first point aggregator, we initially decompose $F_{inst}^{\prime}$ into three matrices based on SVD: (1) $U \in \mathbb{R}^{N_k^{\prime} \times N_k^{\prime}}$ , (2) $\Sigma \in \mathbb{R}^{N_k^{\prime} \times D}$ , and (3) $V \in \mathbb{R}^{D \times D}$ , as: + +$$ +F _ {i n s t} ^ {\prime} = U \cdot \Sigma \cdot V ^ {T} = \sum_ {i} ^ {N _ {k} ^ {\prime}} u _ {i} \cdot \sigma_ {i} \cdot v _ {i} ^ {T} \tag {7} +$$ + +where $U = (u_{1},\ldots ,u_{n})$ and $V = (v_{1},\dots ,v_{n})$ are the matrices with orthonormal columns, and $\Sigma = diag(\sigma_1,\dots,\sigma_n)$ is a diagonal matrix containing singular values, arranged in the descending order. Typically, vectors with higher singular values in $\Sigma$ include meaningful information for object cognition, while the lower-rank vectors are regarded as less important, potentially causing ambiguity. Then, we preserve only the top $t$ singular values from $\Sigma$ , along with their corresponding singular vectors from $U$ and $V$ , facilitating an ambiguity-reduced yet information-rich representation of the original features $F_{inst}^{\prime}$ as follows: + +$$ +\tilde {F} _ {i n s t} ^ {\prime} \in \mathbb {R} ^ {N _ {k} ^ {\prime} \times D} = \tilde {U} \cdot \tilde {\Sigma} \cdot \tilde {V} ^ {T} = \sum_ {i} ^ {N _ {k} ^ {\prime}} \tilde {u} _ {i} \cdot \tilde {\sigma} _ {i} \cdot \tilde {v} _ {i} ^ {T} \tag {8} +$$ + +where $\tilde{U} \in \mathbb{R}^{N_k' \times t}$ , $\tilde{\Sigma} \in \mathbb{R}^{t \times t}$ , and $\tilde{V} \in \mathbb{R}^{D \times t}$ are truncated matrices using only $t$ columns from $U$ , $\Sigma$ , and $V$ , then $\tilde{F}_{inst}'$ is reconstructed from them. Based on both $F_{inst}'$ and $\tilde{F}_{inst}'$ , we sample instance-wise features $F_{inst}$ and $F_{inst}$ , respectively, employing the shared second stage point aggregator. Given $F_{inst}$ and $\tilde{F}_{inst}$ , we calculate the cross-correlation matrix $I_{sg} \in \mathbb{R}^{N_k \times N_k}$ and also generate a pseudo-binary correlation label $I_{sg}^* \in \{0, 1\}^{N_k \times N_k}$ , as in the IKA network. + +![](images/888d7c58af97bbd2c756c3920ad5fb856cd7a5755b7c55dc218a84f94a591b1c.jpg) +Original Instance Feats. +Figure 4. In ISG, we decompose $F_{inst}^{\prime}$ into three matrices $(U, \Sigma, V)$ using the SVD algorithm and keep top $t$ singular vectors to filter out less important noises. Then, we effectively enhance $F_{inst}$ with highlighted essential cues from $\tilde{F}_{inst}$ by optimizing $\mathcal{L}_{SG}$ . + +We then effectively transfer structural clues from $\tilde{F}_{inst}$ , utilizing the correlation regularization loss $\mathcal{L}_{SG}$ as follows: + +$$ +I _ {s g} = F _ {i n s t} \cdot \tilde {F} _ {i n s t} ^ {T}, I _ {s g} ^ {*} (i, j) = \left\{ \begin{array}{l l} 1 & \text {i f} I _ {s g} (i, j) > \tau \\ 0 & \text {o t h e r w i s e ,} \end{array} \right. \tag {9} +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {S G} = \sum_ {I _ {s g} > \tau} \left| \left| I _ {s g} ^ {*} (i, j) - I _ {s g} (i, j) \right| \right| _ {2} \\ + \tilde {w} \cdot \sum_ {I _ {s g} \leqslant \tau} \left\| I _ {s g} ^ {*} (i, j) - I _ {s g} (i, j) \right\| _ {2} \tag {10} \\ \end{array} +$$ + +where $\tilde{w}$ represents the balancing hyper-parameter, and the role of each term in Eq. 10 follows Eq. 6. By optimizing $\mathcal{L}_{SG}$ , ISG conditionally instructs instance features $F_{inst}$ to incorporate spatial configuration details from clarity-enhanced features $\tilde{F}_{inst}$ , improving discriminative power. Loss Function. Finally, our full network minimizes $\mathcal{L}_{total}$ : + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t o t a l}} = \lambda_ {\text {m a s k}} \mathcal {L} _ {\text {m a s k}} + \lambda_ {\text {c l s}} \mathcal {L} _ {\text {c l s}} \\ + \lambda_ {K A} \mathcal {L} _ {K A} + \lambda_ {S G} \mathcal {L} _ {S G} \tag {11} \\ \end{array} +$$ + +where each $\lambda$ is a hyper-parameter decided via extensive grid searches to balance the impact of respective loss terms. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We train and evaluate the overall performance using four landmark benchmarks for 3D Instance Segmentation (3DIS): ScanNetV2 [7], ScanNet200 [38], S3DIS [1], and STPLS3D [5]. These four datasets provide 3D point cloud data collected in diverse real-world environments. Note that detailed descriptions of each dataset and all implementation details are provided in Supplementary Material. + +Evaluation Metrics. We evaluate the 3DIS performance using the Average Precision (AP), a conventional metric in computer vision tasks. We report the mean average precision (mAP) across IoU (Intersection of Union) thresholds + +
MethodScanNetV2 ValScanNetV2 Test
mAP\( mAP_{50} \)\( mAP_{25} \)mAP\( mAP_{50} \)\( mAP_{25} \)
3D-SIS [21]-18.735.716.138.255.8
MTML [28]20.340.255.428.254.973.1
PointGroup [24]34.856.771.340.763.677.8
DyCo3D [19]35.457.672.939.564.176.1
3D-MPA [12]35.559.172.435.561.173.7
MaskGroup [54]42.063.363.343.466.479.2
SoftGroup [44]46.067.678.950.476.186.5
SSTNet [30]49.464.374.050.669.878.9
Mask3D [39]55.273.782.956.678.082.9
SPFormer [43]56.373.982.954.977.085.1
Q-Former [31]56.574.283.358.378.787.4
ISBNet [32]56.873.381.355.976.387.0
MAFT [29]58.475.984.559.678.686.0
EASE [37]60.277.284.259.878.986.7
SphMask [41]62.379.988.261.681.287.5
IKNE (Ours)62.981.888.761.781.288.0
+ +incremented by $5\%$ , ranging from $50\%$ to $95\%$ . Also, we assess $\mathrm{mAP}_{50}$ and $\mathrm{mAP}_{25}$ , which measure accuracy based on IoU thresholds of $50\%$ and $25\%$ . For the S3DIS [1], we further provide mean precision (mPrec) and mean recall (mRec) at a $50\%$ IoU threshold, following previous works. + +# 4.2. Performance Comparison with SOTA Methods + +In this section, we quantitatively compare our proposed framework IKNE with previous methods. Despite significant advances with high scores, these methods, regardless of architecture, still waste considerable instance candidates that may contain valuable information. To maximize the benefits of using numerous candidates, we present novel instance-wise knowledge aggregation (IKA) and guidance (ISG) networks. In Tab. 1, we first evaluate mean average precision (mAP) with different IoU thresholds across 18 classes on the ScanNetV2 [7] dataset. IKNE generally outperforms prior works, achieving new state-of-the-art scores in terms of mAP, $\mathrm{mAP}_{50}$ , and $\mathrm{mAP}_{25}$ on both validation and hidden test sets. While our method slightly exceeds the latest coarse-to-fine approach, Spherical Mask [41], ours transcends our primary baselines, which rely on hundreds of instance candidates, such as ISBNet [32] (kernel-based) and MAFT [29] (transformer), by large margins across all metrics. In Tab. 2, we assess 3DIS performance on Area 5 and the 6-fold cross-validation set of the S3DIS [1] dataset. For Area 5 evaluation, we train with Areas 1 to 6, excluding 5, and validate with Area 5; for 6-fold cross-validation, we average the validation scores across all 6 areas. IKNE demonstrates impressive performance in both evaluations, reaching + +Table 1. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the ScanNetV2 [7] dataset. + +
MethodS3DIS Area 5S3DIS 6-fold CV
AP\( AP_{50} \)\( Prec_{50} \)\( Rec_{50} \)AP\( AP_{50} \)\( Prec_{50} \)\( Rec_{50} \)
3D-Bonet [50]--57.540.2--65.647.6
PointGroup [24]-57.861.962.1-64.069.669.2
3D-MPA [12]--63.158.0--66.764.1
DyCo3D [19]--64.364.2----
MaskGroup [54]-65.062.964.7-69.966.669.6
SSTNet [30]42.759.365.564.254.167.873.573.4
SoftGroup [44]51.666.173.666.654.468.975.369.8
SPFormer [43]-66.872.867.1-69.274.071.1
ISBNet [32]56.367.570.572.060.870.577.577.1
DKNet [48]--70.865.3--75.371.1
Mask3D [39]56.668.468.766.364.575.572.874.5
Q-Former [31]57.769.970.572.262.073.372.773.4
MAFT [29]-69.1------
EASE [37]59.071.669.468.765.376.473.674.6
SphMask [41]60.572.371.376.364.072.378.177.7
IKNE (Ours)61.173.072.077.065.876.979.478.9
+ +Table 2. Comparison of 3DIS performance with state-of-the-art approaches on the S3DIS [1] Area 5 and 6-fold cross validation. + +
MethodmAP\( \mathrm{mAP}_{50} \)\( \mathrm{mAP}_{25} \)
SPFormer [43]25.233.839.6
Mask3D [39]27.437.042.3
Q-Former [31]28.137.143.4
MAFT [29]29.238.243.3
IKNE (Ours)29.938.944.9
+ +Table 3. Comparison of 3DIS performance with state-of-the-art approaches on the ScanNet200 [38] validation (with 200 classes). + +
MethodmAP\( \mathrm{mAP}_{50} \)\( \mathrm{mAP}_{25} \)
PointGroup [24]23.338.548.6
HAIS [6]35.146.752.8
SoftGroup [44]47.363.171.4
Mask3D [39]63.479.285.6
IKNE (Ours)64.981.287.6
+ +Table 4. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the STPLS3D [5] test dataset. + +ing $61.1 / 73.0$ (Area 5) and $65.8 / 76.9$ (6-fold) in $\mathrm{mAP} / \mathrm{mAP}_{50}$ . Also, in Tab. 3, ours achieves the highest scores on the ScanNet200 [38] dataset containing fine-grained 200 categories. Notably, as shown in Tab. 4, IKNE surpasses existing methods on the STPLS3D [5] dataset, including outdoor 3D scenes, with improvements of up to $+1.5 / 2.0 / 2.0$ in $\mathrm{mAP} / \mathrm{mAP}_{50} / \mathrm{mAP}_{25}$ . These results highlight that we successfully handle instance candidates, improving their informativeness for understanding real-world instances from both indoor and outdoor environments. Note that “-” in each table indicates unreported scores from published papers. + +
Method (Kernel-based)ScanNet ValS3DIS Area 5
Baseline (K)IKAISGmAP / mAP50mAP / mAP50
✓--56.8 / 73.356.3 / 67.5
✓✓-62.7 / 80.860.4 / 72.6
✓-✓62.8 / 81.160.6 / 71.9
✓✓✓62.9 / 81.861.1 / 73.0
+ +Table 5. Ablation study to see the effect of two main modules (IKA and ISG) based on the kernel-based pipeline (K). + +
Method (Transformer)ScanNet ValS3DIS Area 5
Baseline (T)IKAISGmAP / mAP50mAP / mAP50
✓--58.4 / 75.956.6 / 68.4
✓✓-59.1 / 76.258.8 / 70.7
✓-✓59.4 / 77.258.4 / 70.1
✓✓✓60.8 / 77.959.0 / 71.6
+ +Table 6. Ablation study to see the effect of two main modules (IKA and ISG) based on the transformer-based pipeline (T). + +
BaselineMethodKernel-basedTransformer
IKAISGAverage IoUAverage IoU
✓--0.86270.8486
✓✓-0.88280.8728
✓-✓0.88250.8730
✓✓✓0.89040.8803
+ +# 4.3. Ablation Studies + +Effect of IKA and ISG Networks. We evaluate variants of our method w/ and w/o IKA and ISG modules based on kernel (K) and transformer-based (T) architectures. As shown in Tab. 5 and 6, the addition of each module surpasses the score of baseline models [29, 32, 39] across all experiments, regardless of model structure. Specifically, IKA boosts performance, with gains of up to $+5.9 / 7.5$ for ScanNetV2 [7] and $+4.1 / 5.1$ for S3DIS [1] in $\mathrm{mAP} / \mathrm{mAP}_{50}$ , underscoring the importance of integrating scattered clues from multiple candidates. Also, ISG contributes to the spatial understanding of instance candidates via intelligent structural guidance from clarity-enhanced features, improving up to $+6.0 / 7.8$ and $+4.3 / 4.4$ . Eventually, utilizing all modules together gains further advances, resulting in the best performance with notable progress on both structures. These results demonstrate each module plays a meaningful role in improving the discriminative capability of instance features. + +Furthermore, to assess how each module influences the accurate determination of instance mask coverage, we use the IoU metric, which is closely associated with segmentation performance. IoU calculates the ratio of the intersection to the union of predicted and ground truth masks. In Tab. 7, we compare the average IoU of our predicted instance masks and those of the baselines for both architectures. As shown in Fig. 1 (a) and Fig.1 (a) (in Supp.), for previous works, a single instance is often separated into sev- + +Table 7. Ablation study to see the effect of two main modules across both pipelines, using IoU scores on the ScanNetV2 [7] val. + +
ArchitectureThreshold τScanNet ValS3DIS Area 5
mAP / mAP50mAP / mAP50
Baseline (K)-56.8 / 73.356.3 / 67.5
Kernel-based0.653.6 / 73.455.3 / 65.5
Kernel-based0.757.2 / 76.459.1 / 69.2
Kernel-based0.861.1 / 79.960.7 / 72.3
Kernel-based0.962.9 / 81.861.1 / 73.0
Baseline (T)-58.4 / 75.956.6 / 68.4
Transformer0.657.9 / 75.856.3 / 67.5
Transformer0.759.2 / 76.257.4 / 68.6
Transformer0.859.9 / 77.158.6 / 70.4
Transformer0.960.8 / 77.959.0 / 71.6
+ +Table 8. Ablation study to investigate the threshold $\tau$ for determining instance-wise candidates representing the same instance. + +![](images/1a966361c95c2551373d737051f9b0fec27598f3f8fad077fe92df7ddef9edc1.jpg) +Figure 5. Comparison to explore the optimal top $t$ for reducing ambiguity while keeping important cues of instance features. + +![](images/868289e5a6cf919f05113b6dcd340f90a55d7574a560f2e6ddffad8e790aa8d4.jpg) + +eral fragments. These fragments potentially cause low IoU scores. The IoU improvements with IKA across both architectures confirm its effectiveness in reducing the negative impact of multiple fragments and enhancing the accuracy of mask coverage. Also, the IoU tends to be low, particularly when instance masks include the surrounding backgrounds or when adjacent instances are not clearly distinguished, as in Fig. 1 (b) and Fig.1 (b) (in Supp.), leading to overestimated or underestimated instance masks. In addition to IKA, ISG enhances the IoU scores, verifying our guidance is sufficiently practical. We further validate the significance of ours with additional experiments in Supp. (see Sec.4). + +Thresholds $\tau$ for Instance-wise Identification. We specify candidates representing the same instance using predefined threshold $\tau$ from the correlation matrices $I_{ka}$ and $I_{sg}$ . In Tab. 8, based on kernel (K) and transformer (T) architectures [29, 32, 39], we explore the impact of $\tau$ for useful knowledge interaction. When $\tau$ is lower than 0.7, the scores degrade due to the inclusion of inaccurate information from different instance representative features. Yet, with more precise identification using the higher $\tau$ ( $\geqslant 0.8$ ), our knowledge interaction yields considerable improvements. This result emphasizes that clues from accurately identified candidates are precious for enhancing instance interpretation. We + +![](images/368dcd2e9bd0908366bd5301ac51392c02a22e2858876da82abdf0f6f7933ce5.jpg) +Figure 6. Qualitative comparison of instance (Inst.) and semantic (Sem.) masks between the baselines (ISBNet (K) [32], MAFT (T) [29]) and ours with Ground Truth masks on the ScanNetV2. The critical differences are highlighted using yellow boxes. Best viewed in color. + +also examine $\tau$ for the STPLS3D [5] in Supp. (Sec.4.3). + +Analysis of the Top t values. In this work, we truncate the top $t$ vectors of $(U,\Sigma ,V)$ from $F_{inst}^{\prime}$ and reconstruct $\tilde{F}_{inst}^{\prime}$ using $(\tilde{U},\tilde{\Sigma},\tilde{V})$ based on a truncated SVD algorithm [17]. Here, the $t$ value impacts the balance between data compression and information preservation. Therefore, it is crucial to find the optimal value of $t$ for proper balance. To this end, we conduct experiments using various $t$ for both kernel (blue) and transformer (orange) architectures in Fig. 5. We represent $t$ as a percentage, so the x-axis indicates the proportion of vectors retained during truncation across three matrices. We observe that the trends of all experiments are similar. In settings with a lower $t$ percentage (0.1), the scores consistently decrease due to large information loss. However, with a higher $t$ (0.7), the effect of noise reduction is insufficient; thereby, the performance is comparable to baselines [29, 32]. We ultimately set $t$ as 0.5, resulting in the highest accuracy while allowing us to preserve meaningful information and effectively filter out inherent noises. + +# 4.4. Qualitative Results + +Visual Comparison. In Fig. 6, we qualitatively validate the effectiveness of our novel framework IKNE. We visualize the predicted semantic (Sem.) and instance (Inst.) masks of the kernel (K) [32] and transformer (T) [29] baselines with ours on the ScanNetV2 [7]. We also provide the corresponding ground truth for solid comparisons. In Scene 1, ours accurately segments the cabinet as a single instance, whereas the baselines separate it into multiple fragments. In Scene 2, ours correctly distinguishes neighboring instances (four chairs and a table), while baselines erroneously recognize two chairs as one (K) or a table and a chair as one (T). Also, in Scene 3, where objects are disorderly adjacent, ours clearly captures the spatial range of the desk and chairs compared to the baselines. These qualitative results verify that our instance features contain highly informative cues. + +Structural Guidance from $\tilde{\mathbf{F}}_{\mathrm{inst}}$ . In Fig. 7, we qualitatively confirm the significance of structural cues from the + +![](images/dc58b0ff2a13755be704cf1499da2153084d4d029327480edc6d57100f8932af.jpg) +Figure 7. Instance feature visualizations of the original features $F_{inst}$ (blue) and ambiguity-reduced features $\bar{F}_{inst}$ (red). + +noise-reduced features $\tilde{F}_{inst}$ in smartly guiding the original features $F_{inst}$ . As shown in Scene 1 and 2, instance features (red) from $\tilde{F}_{inst}$ accurately capture the spatial range of instances (chair and desk), whereas instance features (blue) from $F_{inst}$ struggle to cover the complete coverage. Also, in Scene 3, where the distinction is unclear (blue) from surroundings (curtain and wall), the candidate (red) from $\tilde{F}_{inst}$ clearly segments such instances. These findings highlight guidance from clarity-enhanced features is worthwhile. + +# 5. Conclusion + +In this paper, we introduce a novel instance-wise knowledge enhancement approach, IKNE, for 3D instance segmentation. We focus on optimizing the efficiency of hundreds of instance candidates by effectively handling those representing the same instance. To address these candidates, we propose an IKA to integrate spread clues based on correlation regularization. Also, we present an ISG, which enhances instance features with fundamental cues for object understanding from ambiguity-reduced instance features using the truncated SVD. Our extensive experiments demonstrate the effectiveness of our methods, achieving new state-of-the-art results on both kernel and transformer architectures. + +# Acknowledgement + +This work was supported by the Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2024 (RS-2023-00227409, Development of sketch-based semantic 3D modeling technology for creating user-centric Metaverse content spaces for indoor spaces, $90\%$ ; RS-2024-00345025, International Collaborative Research and Global Talent Development for the Development of Copyright Management and Protection Technologies for Generative AI, $5\%$ ), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2019-II190079, Artificial Intelligence Graduate School Program (Korea University), $1\%$ ; IITP-2025-RS-2024-00436857, $2\%$ ), the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2025-00521602, $2\%$ ), and Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT (MSIT, Korea) & Gwangju Metropolitan City. + +# References + +[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 5, 6, 7 +[2] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. 2, 4 +[3] Ori Bryt and Michael Elad. Compression of facial images using the k-svd algorithm. Journal of Visual Communication and Image Representation, 19(4):270-282, 2008. 3 +[4] Chin-Chen Chang, Piyu Tsai, and Chia-Chen Lin. Svd-based digital image watermarking scheme. Pattern Recognition Letters, 26(10):1577-1586, 2005. 3 +[5] Meida Chen, Qingyong Hu, Zifan Yu, Hugues Thomas, Andrew Feng, Yu Hou, Kyle McCullough, Fengbo Ren, and Lucio Soibelman. Stpls3d: A large-scale synthetic and real aerial photogrammetry 3d point cloud dataset. arXiv preprint arXiv:2203.09065, 2022. 2, 5, 6, 8 +[6] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 1, 2, 6 +[7] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5, 6, 7, 8 +[8] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In + +Proceedings of the European conference on computer vision (ECCV), pages 562-578, 2018. 4 +[9] Renwei Dian, Shutao Li, and Xudong Kang. Regularizing hyperspectral and multispectral image fusion by cnn denoiser. IEEE transactions on neural networks and learning systems, 32(3):1124-1135, 2020. 3 +[10] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. IEEE transactions on image processing, 6(9):1305-1315, 1997. 4 +[11] Cathrin Elich, Francis Engelmann, Theodora Kontogianni, and Bastian Leibe. 3d bird's-eye-view instance segmentation. In German Conference on Pattern Recognition, pages 48-61. Springer, 2019. 2 +[12] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2, 6 +[13] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3722-3731, 2021. 2 +[14] Gene H Golub and Christian Reinsch. Singular value decomposition and least squares solutions. In Handbook for Automatic Computation: Volume II: Linear Algebra, pages 134-151. Springer, 1971. 2, 3, 5 +[15] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 4 +[16] Qiang Guo, Caiming Zhang, Yunfeng Zhang, and Hui Liu. An efficient svd-based method for image denoising. IEEE transactions on Circuits and Systems for Video Technology, 26(5):868-880, 2015. 3 +[17] Per Christian Hansen. The truncated svd as a method for regularization. BIT Numerical Mathematics, 27:534-553, 1987. 3, 8 +[18] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2 +[19] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3d: Robust instance segmentation of 3d point clouds through dynamic convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 354-363, 2021. 1, 2, 3, 4, 6 +[20] Tong He, Wei Yin, Chunhua Shen, and Anton Van den Hengel. Pointinst3d: Segmenting 3d instances by points. In European Conference on Computer Vision, pages 286-302. Springer, 2022. 1, 2, 3, 4 +[21] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 1, 2, 6 + +[22] Peg Howland, Moongu Jeon, and Haesun Park. Structure preserving dimension reduction for clustered text data based on the generalized singular value decomposition. SIAM Journal on Matrix Analysis and Applications, 25(1):165-179, 2003. 3 +[23] Sujin Jang, Dae Ung Jo, Sung Ju Hwang, Dongwook Lee, and Daehyun Ji. Stxd: Structural and temporal cross-modal distillation for multi-view 3d object detection. Advances in Neural Information Processing Systems, 36, 2024. 2 +[24] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 1, 2, 6 +[25] XiaoBing Kang and ShengMin Wei. Identifying tampered regions using singular value decomposition in digital image forensics. In 2008 International conference on computer science and software engineering, pages 926-930. IEEE, 2008. 3 +[26] Sungjune Kim, Hyung-gun Chi, Hyerin Lim, Karthik Ramani, Jinkyu Kim, and Sangpil Kim. Higher-order relational reasoning for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15251-15260, 2024. 3 +[27] Virginia Klema and Alan Laub. The singular value decomposition: Its computation and some applications. IEEE Transactions on automatic control, 25(2):164-176, 1980. 3 +[28] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9256-9266, 2019. 6 +[29] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-attention-free transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3693–3703, 2023. 1, 2, 3, 4, 6, 7, 8 +[30] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 6 +[31] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18516-18526, 2023. 1, 2, 4, 6 +[32] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. Isbnet: a 3d point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13550-13559, 2023. 1, 2, 3, 4, 6, 7, 8 +[33] HS Prasantha, HL Shashidhara, and KN Balasubramanya Murthy. Image compression using svd. In International conference on computational intelligence and multimedia applications (ICCIMA 2007), pages 143-145. IEEE, 2007. 3 +[34] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point + +clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1, 2 +[35] Ajit Rajwade, Anand Rangarajan, and Arunava Banerjee. Image denoising using the higher order singular value decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(4):849-862, 2012. 3 +[36] Zhongzheng Ren, Ishan Misra, Alexander G Schwing, and Rohit Girdhar. 3d spatial recognition without spatially labeled 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13213, 2021. 2 +[37] Wonseok Roh, Hwanhee Jung, Giljoo Nam, Jinseop Yeom, Hyunjie Park, Sang Ho Yoon, and Sanggil Kim. Edge-aware 3d instance segmentation network with intelligent semantic prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20644–20653, 2024. 2, 3, 6 +[38] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2, 5, 6 +[39] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d for 3d semantic instance segmentation. arXiv preprint arXiv:2210.03105, 2022. 1, 2, 3, 4, 6, 7 +[40] Qiquan Shi, Yiu-Ming Cheung, and Jian Lou. Robust tensor svd and recovery with rank estimation. IEEE Transactions on Cybernetics, 52(10):10667-10682, 2021. 3 +[41] Sangyun Shin, Kaichen Zhou, Madhu Vankadari, Andrew Markham, and Niki Trigoni. Spherical mask: Coarse-to-fine 3d point cloud instance segmentation with spherical representation. arXiv preprint arXiv:2312.11269, 2023. 3, 6 +[42] Gilbert W Stewart. On the early history of the singular value decomposition. SIAM review, 35(4):551-566, 1993. 2, 3, 5 +[43] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 1, 2, 6 +[44] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 6 +[45] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2569-2578, 2018. 1, 2 +[46] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2 +[47] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scenegraphfusion: Incremental 3d scene graph prediction from rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7515-7525, 2021. 2 + +[48] Yizheng Wu, Min Shi, Shuaiyuan Du, Hao Lu, Zhiguo Cao, and Weicai Zhong. 3d instances as 1d kernels. In European Conference on Computer Vision, pages 235-252. Springer, 2022. 1, 2, 4, 6 +[49] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Dening Lu, Mingqiang Wei, and Jun Wang. Venet: Voting enhancement network for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3712-3721, 2021. 2 +[50] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds. Advances in neural information processing systems, 32, 2019. 1, 2, 6 +[51] Dan Yang, Zongming Ma, and Andreas Buja. A sparse singular value decomposition method for high-dimensional data. Journal of Computational and Graphical Statistics, 23 (4):923-942, 2014. 3 +[52] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3947-3956, 2019. 1, 2 +[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pages 12310-12320. PMLR, 2021. 2, 4 +[54] Min Zhong, Xinghao Chen, Xiaokang Chen, Gang Zeng, and Yunhe Wang. Maskgroup: Hierarchical point grouping and masking for 3d instance segmentation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 6 \ No newline at end of file diff --git a/2025/Insightful Instance Features for 3D Instance Segmentation/images.zip b/2025/Insightful Instance Features for 3D Instance Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aa848d36cd55af8a309d3732fae8346a55d1887b --- /dev/null +++ b/2025/Insightful Instance Features for 3D Instance Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fdbc81cea2c5197df27cdefd0990506324eb8afb1da9264b5232e45316fa57d +size 873660 diff --git a/2025/Insightful Instance Features for 3D Instance Segmentation/layout.json b/2025/Insightful Instance Features for 3D Instance Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..28bb90d3f5c3e15f0fb53c7cc1786385300cca8b --- /dev/null +++ b/2025/Insightful Instance Features for 3D Instance Segmentation/layout.json @@ -0,0 +1,10432 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 125, + 103, + 486, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 103, + 486, + 121 + ], + "spans": [ + { + "bbox": [ + 125, + 103, + 486, + 121 + ], + "type": "text", + "content": "Insightful Instance Features for 3D Instance Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "spans": [ + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "text", + "content": "Wonseok Roh" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "text", + "content": " Hwanhee Jung" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "text", + "content": " Giljoo Nam" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "text", + "content": " Dong In Lee" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "text", + "content": " \nHyeongcheol Park" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "text", + "content": " Sang Ho Yoon" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "text", + "content": " Jungseock Joo" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "text", + "content": " Sangpil Kim" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "text", + "content": "Korea University " + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "text", + "content": "Meta " + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "text", + "content": "KAIST " + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 127, + 142, + 487, + 194 + ], + "type": "text", + "content": "UCLA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 217, + 201, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 201, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 201, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 241, + 297, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 241, + 297, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 241, + 297, + 529 + ], + "type": "text", + "content": "Recent 3D Instance Segmentation methods typically encode hundreds of instance-wise candidates with instancespecific information in various ways and refine them into final masks. However, they have yet to fully explore the benefit of these candidates. They overlook the valuable cues encoded in multiple candidates that represent different parts of the same instance, resulting in fragments. Also, they often fail to capture the precise spatial range of 3D instances, primarily due to inherent noises from sparse and unordered point clouds. In this work, to address these challenges, we propose IKNE, a novel instance-wise knowledge enhancement approach. We first introduce an Instance-wise Knowledge Aggregation (IKA) to associate scattered single instance details by optimizing correlations among candidates representing the same instance. Moreover, we present an Instance-wise Structural Guidance (ISG) to enhance the spatial understanding of candidates using structural cues from ambiguity-reduced features. Here, we utilize a simple yet effective truncated singular value decomposition algorithm to minimize inherent noises of 3D features. In our extensive experiments on large-scale datasets, ScanNetV2, ScanNet200, S3DIS, and STPLS3D, IKNE outperforms existing works. We validate the effectiveness of our modules in both kernel-based and transformer-based architectures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 137, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 137, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 137, + 555 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "content": "3D Instance Segmentation (3DIS) is a fundamental 3D computer vision task that significantly contributes to real-world applications. Given 3D point cloud scenes, the 3DIS task identifies respective instances and assigns semantic class labels with the goal of comprehensively understanding entire spatial environments. In real-life 3D scenarios, substantial occlusion and truncation commonly arise, especially when objects are overlapped or obscured by others. To tackle these challenges, early approaches mainly focused on accurately generating region proposals (top-down) [21, 50, 52] or effectively grouping points with clus" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 217, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 266 + ], + "type": "text", + "content": "tering algorithms (bottom-up) [6, 24, 45]. However, they presuppose that the intermediate processes, such as bounding box detection [18] or heuristic voting mechanism [34], generate near perfect results, which is often not the case." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 268, + 556, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 556, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 556, + 532 + ], + "type": "text", + "content": "Recently, kernel-based [19, 20, 32, 48] and transformer-based [29, 31, 39, 43] 3DIS approaches have been proposed, aiming to overcome the limitations of traditional frameworks. Kernel-based methods leverage instance-aware kernels for dynamic convolution, which decodes instance masks. They represent instances as kernels, replacing clustering algorithms with point sampling processes. On the other hand, transformer-based methods train instance queries to identify features about individual objects using attention mechanisms. These queries enable the direct prediction of per-point categories and instance labels. In both architectures, models utilize hundreds of instance-wise candidate features, to estimate the final instance mask. For example, ISBNet [32] (kernel-based) produces 256 instance kernels, and MAFT [29] (transformer-based) leverages 400 instance queries. However, despite the significant improvements based on their architectural advantages, they have yet to fully optimize the handling of these numerous candidates and still face several challenges, as illustrated in Fig. 1. To tackle these challenges, we introduce two main modules, IKA and ISG, each thoughtfully designed to enhance the informativeness of instance candidate features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": "First, we empirically observe that multiple candidates from previous works often encode cues from the same single instance, as shown in Fig. 1 (a). These candidates are then typically decoded into considerable fragment masks representing different components of each instance, such as the backrest or armrest of the sofa. In this work, to estimate the complete coverage of each instance, we introduce an Instance-wise Knowledge Aggregation (IKA) network, which integrates the scattered intelligence from multiple fragments beyond relying solely on a few fragments based on confidence scores. Specifically, we optimize the correlations among instance-wise candidate features, encouraging the network to be aware of the distributed instance-specific knowledge. We initially identify candidates representing the same instance based on the similarity of their features, as" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 208, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 208, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 208, + 713 + ], + "type": "text", + "content": "*Corresponding author (spk7@korea.ac.kr)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14057" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 59, + 306, + 195 + ], + "blocks": [ + { + "bbox": [ + 58, + 59, + 306, + 195 + ], + "lines": [ + { + "bbox": [ + 58, + 59, + 306, + 195 + ], + "spans": [ + { + "bbox": [ + 58, + 59, + 306, + 195 + ], + "type": "image", + "image_path": "30a9a2e2005cbac10a94477050abc7efac50968382e685ce4335538ed8967ea2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 196, + 265, + 206 + ], + "lines": [ + { + "bbox": [ + 104, + 196, + 265, + 206 + ], + "spans": [ + { + "bbox": [ + 104, + 196, + 265, + 206 + ], + "type": "text", + "content": "(a) Per-point masks of instance-wise candidates" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 309, + 59, + 553, + 195 + ], + "blocks": [ + { + "bbox": [ + 309, + 59, + 553, + 195 + ], + "lines": [ + { + "bbox": [ + 309, + 59, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 309, + 59, + 553, + 195 + ], + "type": "image", + "image_path": "4f5a6da64acca8cbfad76fc499301524148b1bec0e9c30e97c2c73d325ad47a4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 196, + 490, + 207 + ], + "lines": [ + { + "bbox": [ + 371, + 196, + 490, + 207 + ], + "spans": [ + { + "bbox": [ + 371, + 196, + 490, + 207 + ], + "type": "text", + "content": "(b) Instance misinterpretation cases" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 215, + 555, + 261 + ], + "lines": [ + { + "bbox": [ + 54, + 215, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 54, + 215, + 555, + 261 + ], + "type": "text", + "content": "Figure 1. Examples of two challenging cases of 3DIS. (a) Per-point masks of instance-wise candidates: highly correlated instance-wise candidates usually represent incomplete fragments of the same single instance, such as the backrest or armrest of a sofa. And (b) Instance misinterpretation cases: existing studies often confuse instances from backgrounds (Scene1) or misunderstand the spatial range and appearance of instances (Scene2). Note that we also provide additional examples for both cases in the supplementary material (Fig. 1)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 270, + 294, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 294, + 366 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 294, + 366 + ], + "type": "text", + "content": "highly correlated candidates usually capture the information from the same instance. Then, inspired by correlation regularization mechanisms [2, 23, 53], we softly guide identified candidates to be closer to each other in the latent space, aiming to enhance solidarity among them. As a result, IKA facilitates the network to learn relationships between candidates and aggregate their comprehensive details, maximizing the advantages of using hundreds of candidates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 368, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 368, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 57, + 368, + 295, + 688 + ], + "type": "text", + "content": "In addition to IKA, we explore additional avenues to enrich the knowledge of instance-wise candidates. Unlike the human visual cognitive system, which perceives instances with sufficient prior knowledge, the network understands instances depending on their external properties from point features. However, point features are frequently unstable, primarily due to ambiguous inherent noises [13, 36, 47, 49] caused by sparse and incomplete point clouds. These noises are likely to confuse the models, making it hard to precisely interpret the specific spatial range and unique appearance of instances, as shown in Fig. 1 (b). To mitigate the negative impact of such noises, we propose an Instance-wise Structural Guidance (ISG) network specifically designed to improve candidates' structural understanding of 3D instances from point features. Here, we utilize the truncated singular value decomposition (SVD) [14, 42] algorithm to strengthen essential cues (e.g., shape) of instance features while explicitly reducing fuzzy noises. Then, we effectively transfer fundamental clues for object cognition from ambiguity-minimized features to corresponding original features based on the cross-correlations between them. Ultimately, our instance features are now highly informative across complex 3D instances with aggregated and structural knowledge from IKA and ISG. Our novel modules are applicable to frameworks employing considerable instance candidates, regardless of their underlying model structures, including kernel-based and transformer-based architectures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Given landmark datasets for 3DIS, ScanNetV2 [7], ScanNet200 [38], S3DIS [1], and STPLS3D [5], we vali" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 270, + 553, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 306 + ], + "type": "text", + "content": "date the effectiveness of our novel framework. Our method outperforms the existing state-of-the-art methods. To summarize, our main contributions are listed as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 308, + 553, + 475 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 308, + 553, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 308, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 314, + 308, + 553, + 369 + ], + "type": "text", + "content": "- To extend the benefits of producing multiple instance-wise candidates, we introduce the Instance-wise Knowledge Aggregation (IKA) network, which associates scattered instance-specific information of the same single instance by optimizing correlations among them." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 369, + 553, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 369, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 314, + 369, + 553, + 429 + ], + "type": "text", + "content": "- We design the Instance-wise Structural Guidance (ISG) network to improve the structural knowledge of candidates. Specifically, we use a simple yet strong truncated singular value decomposition process to emphasize essential cues from point features while filtering out noises." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 429, + 553, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 429, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 314, + 429, + 553, + 475 + ], + "type": "text", + "content": "- We analyze the effectiveness of our framework, IKNE, on four challenging datasets. Comprehensive experiments in various 3D scenarios demonstrate that our novel method achieves new State-of-the-Art performance on 3DIS." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 487, + 400, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 400, + 500 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 400, + 500 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "type": "text", + "content": "3D Instance Segmentation (3DIS) aims to distinguish individual instances within 3D scenes and assign corresponding semantic categories. Typically, 3DIS approaches can be categorized into four types: proposal-based [21, 50, 52], grouping-based [6, 11, 12, 24, 45], kernel-based [19, 20, 32, 48], and transformer-based [29, 31, 37, 39, 43]. Based on the impressive achievements of detection methods [18, 46], proposal-based approaches first detect instances and employ proposed regions as hard references to predict masks. However, these strategies, such as 3D-SIS [21] or 3D-BoNet [50], heavily rely on the quality of the outputs from the detection process. On the other hand, grouping-based methods catch 3D instances through a bottom-up pipeline, aggregating closely related points into instances using predicted semantic categories and center offsets. Yet, they still depend on intermediate manually tuned processing, such as point grouping [24] or heuristic voting mechanisms [34]." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14058" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 59, + 555, + 225 + ], + "blocks": [ + { + "bbox": [ + 59, + 59, + 555, + 225 + ], + "lines": [ + { + "bbox": [ + 59, + 59, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 59, + 59, + 555, + 225 + ], + "type": "image", + "image_path": "919a97be83018fe598f7553d2c6c20071c8295fe45b50e6b4b650bc5b690841e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "type": "text", + "content": "Figure 2. An overview of our framework IKNE. Built upon the modern kernel-based structure, our model consists of four main modules: (1) Sparse Convolutional 3D Backbone; (2) Instance-wise Knowledge Aggregation (IKA, Sec. 3.2), which softly associates scattered cues from highly correlated instance features of " + }, + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "type": "text", + "content": "; (3) Instance-wise Structural Guidance (ISG, Sec. 3.3), which smartly instructs " + }, + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "type": "text", + "content": " with essential clues from " + }, + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{inst}" + }, + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "type": "text", + "content": ", using an effective SVD algorithm; and (4) Dynamic Convolution, which produces " + }, + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\hat{M}_f" + }, + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "type": "text", + "content": " from kernels " + }, + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 285, + 294, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 294, + 465 + ], + "type": "text", + "content": "To address these limitations, kernel-based approaches have been introduced. For example, DyCo3D [19] employs a clustering algorithm to generate kernels for dynamic convolutions to predict instance masks, while ISBNet [32] presents a cluster-free method using instance-wise kernels. Recently, Spherical Mask [41], based on ISBNet, has overcome the low-quality results of coarse-to-fine strategies by using spherical representation to mitigate the propagation of false negatives and overestimation of instance size. Furthermore, transformer-based methods, Mask3D [39], MAFT [29], and EASE [37] utilize instance-wise queries to encode individual instance information through attention mechanisms. In this work, we focus on improving the quality of numerous instance-wise candidates introduced from recent studies with two novel approaches, IKA and ISG." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 467, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 295, + 551 + ], + "type": "text", + "content": "Singular Value Decomposition (SVD) is a fundamental algorithm in linear algebra, widely used for matrix factorization [14, 27, 42], dimensionality reduction [22, 51], and 2D image processing [9, 16, 26, 35, 40]. SVD decomposes a matrix " + }, + { + "bbox": [ + 55, + 467, + 295, + 551 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 55, + 467, + 295, + 551 + ], + "type": "text", + "content": " into the dot product of three matrices, " + }, + { + "bbox": [ + 55, + 467, + 295, + 551 + ], + "type": "inline_equation", + "content": "U \\in \\mathbb{R}^{m \\times m}" + }, + { + "bbox": [ + 55, + 467, + 295, + 551 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 467, + 295, + 551 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 55, + 467, + 295, + 551 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 467, + 295, + 551 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 55, + 467, + 295, + 551 + ], + "type": "text", + "content": ", each containing meaningful vectors and scalar values, as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 554, + 294, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 554, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 141, + 554, + 294, + 567 + ], + "type": "interline_equation", + "content": "A = U \\cdot \\Sigma \\cdot V ^ {T} \\tag {1}", + "image_path": "4875cfe4dafb7583349b283aaec0dc15f16d40fc60f17ed4bec8ab275b0c236a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": " are the matrices with orthonormal columns, satisfying " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "U^T U = V^T V = I" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": " is a diagonal matrix with nonnegative singular values " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": " in descending order as " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\sigma_1 \\geqslant \\sigma_2 \\geqslant \\dots \\geqslant \\sigma_r" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": ". Based on SVD, truncated SVD [17] is an insightful dimension-reduction technique that minimizes nonessential noises while highlighting essential cues from the original matrix " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": ". It produces a low-rank approximation " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{A} \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": " by retaining the most important top " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": " singular values and their corresponding vectors from " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 680, + 294, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 680, + 294, + 695 + ], + "spans": [ + { + "bbox": [ + 137, + 680, + 294, + 695 + ], + "type": "interline_equation", + "content": "\\tilde {A} = U _ {k} \\cdot \\Sigma_ {k} \\cdot V _ {k} ^ {T} \\tag {2}", + "image_path": "4bcfec880237237687c154a0636cb63c5c2407acd125587036e840359df0e431.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 700, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 700, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 700, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 700, + 295, + 714 + ], + "type": "inline_equation", + "content": "U_{k}\\in \\mathbb{R}^{m\\times k}" + }, + { + "bbox": [ + 55, + 700, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\Sigma_{k}\\in \\mathbb{R}^{k\\times k}" + }, + { + "bbox": [ + 55, + 700, + 295, + 714 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 55, + 700, + 295, + 714 + ], + "type": "inline_equation", + "content": "V_{k}\\in \\mathbb{R}^{n\\times k}" + }, + { + "bbox": [ + 55, + 700, + 295, + 714 + ], + "type": "text", + "content": " are the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 285, + 541, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 541, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 541, + 297 + ], + "type": "text", + "content": "truncated matrices with the top " + }, + { + "bbox": [ + 313, + 285, + 541, + 297 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 285, + 541, + 297 + ], + "type": "text", + "content": " singular value vectors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 298, + 555, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 490 + ], + "type": "text", + "content": "In the field of 2D image processing, SVD has been extensively utilized for low-rank approximation. [16, 35] employ SVD to factorize and estimate low-rank patches, effectively reducing noise and enhancing image quality. Similarly, [4, 25] operate SVD to detect tampering by extracting and analyzing low-dimensional representations of images for image forensics. Also, SVD showcases robust performance in image compression [3, 33] and recovery [40]. Further, [9] leverages SVD within CNNs to learn subspace information for hyperspectral image reconstruction. However, despite its potential benefits, the introduction of SVD has received limited attention in 3D vision, especially for point clouds. Based on these observations, we apply SVD to address point features that contain inherent noises due to the unordered properties of point clouds, facilitating the model to become more robust against complex ambiguities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 499, + 370, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 499, + 370, + 510 + ], + "spans": [ + { + "bbox": [ + 314, + 499, + 370, + 510 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 519, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 554, + 579 + ], + "type": "text", + "content": "In this section, we introduce a novel 3D Instance Segmentation framework IKNE. We provide an overview of the whole pipeline (Sec. 3.1 and Fig. 2), followed by detailed technical discussions: Instance-wise Knowledge Aggregation (Sec. 3.2) and Instance-wise Structural Guidance (Sec. 3.3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 587, + 382, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 382, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 382, + 598 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Recent 3DIS methods [19, 20, 32, 37, 39] commonly follow a similar paradigm; they spread hundreds of instance candidates across 3D scenes to capture instance-specific information in various ways and refine them into final instance masks. In this work, we encourage these numerous instance candidates to be highly informative, improving their understanding of diverse real-world instances. We describe our instance-wise knowledge enhancement approaches based on the modern kernel-based architecture [32]. As shown in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14059" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "Fig. 2, our model consists of four main modules: (1) Sparse Convolutional 3D Backbone; (2) Instance-wise Knowledge Aggregation, which relates the scattered information of individual instances; (3) Instance-wise Structural Guidance, which effectively instructs candidates with spatial cues from augmented instance features; and (4) Dynamic Convolution network. Note that we also describe our transformer-based architecture in the supplementary material (Sec.3)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": "Kernel-based 3DIS Architecture. First, sparse convolutional U-Net backbone [15] takes a colored point cloud " + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N_p \\times 6}" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": " as input and voxelizes " + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": " into voxels to extract point-wise feature maps " + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "F_{p} \\in \\mathbb{R}^{N_{p} \\times D}" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": ". Given the point features, the point aggregator samples a set of instance candidate features, referring to point positions based on the Farthest Point Sampling (FPS) [10]. Specifically, we employ two-stage point aggregator blocks to produce optimal instance features across 3D scenes. In the first stage, " + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "N_{k}^{\\prime}" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": " number of candidates " + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "F_{inst}^{\\prime} \\in \\mathbb{R}^{N_{k}^{\\prime} \\times D}" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": " are sampled from full-resolution feature " + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": ", while in the second stage, " + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "< N_{k}^{\\prime}" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": ") instance features " + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "F_{inst} \\in \\mathbb{R}^{N_{k} \\times D}" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": " are sampled based on the " + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "F_{inst}^{\\prime}" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": ". We then classify the category " + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "C_{i}" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "i = \\{1,2,\\dots,N_{c}\\}" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": " of each instance from " + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": " via linear classification head " + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 172, + 296, + 354 + ], + "type": "text", + "content": " using cross-entropy loss as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 361, + 296, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 361, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 68, + 361, + 296, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = - \\mathbb {E} _ {c, w _ {c} \\sim \\mathbb {D}} \\left[ \\sum_ {r \\in N _ {c}} w _ {c} [ r ] \\log f _ {\\mathrm {c l s}} \\left(F _ {\\text {i n s t}}\\right) [ r ] \\right] \\tag {3}", + "image_path": "4f3328039a9a8492694c010102b6c297d78be6ca8659bd9e9d79a79ec545229d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "inline_equation", + "content": "w_{c}" + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "content": " indicates one-hot encoded category labels and " + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "content": " represents (input) data distribution. Also, we predict instance kernels " + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "inline_equation", + "content": "K\\in \\mathbb{R}^{N_k\\times D}" + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "content": " through MLPs. Furthermore, each point-wise prediction head processes " + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "content": " to generate mask features " + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "inline_equation", + "content": "F_{mask}\\in \\mathbb{R}^{N_p\\times D}" + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "content": " and axis-aligned bounding boxes " + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "inline_equation", + "content": "F_{box}\\in \\mathbb{R}^{N_p\\times 6}" + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "content": " as auxiliary information for mask prediction. Finally, the dynamic convolution network produces the final instance mask " + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\hat{M}_f" + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "content": " using instance kernels " + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "inline_equation", + "content": "F_{mask}" + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "content": ". To optimize " + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\hat{M}_f" + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "content": ", we formulate the following loss " + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 55, + 402, + 296, + 524 + ], + "type": "text", + "content": " as a sum of the two losses:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 533, + 295, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 533, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 61, + 533, + 295, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {\\mathrm {B C E}} \\mathcal {L} _ {\\mathrm {B C E}} \\left(M _ {f}, \\hat {M} _ {f}\\right) + \\lambda_ {\\mathrm {d i c e}} \\mathcal {L} _ {\\mathrm {d i c e}} \\left(M _ {f}, \\hat {M} _ {f}\\right) \\tag {4}", + "image_path": "266e864ae0193d738f02b8db7cb32ef93d290d08f40805289717a6a9d46601f4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 555, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 555, + 295, + 581 + ], + "type": "inline_equation", + "content": "M_{f}" + }, + { + "bbox": [ + 55, + 555, + 295, + 581 + ], + "type": "text", + "content": " is the ground-truth mask, and " + }, + { + "bbox": [ + 55, + 555, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BCE}}" + }, + { + "bbox": [ + 55, + 555, + 295, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 555, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dice}}" + }, + { + "bbox": [ + 55, + 555, + 295, + 581 + ], + "type": "text", + "content": " denote binary cross-entropy and Dice losses [8], respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 587, + 288, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 288, + 601 + ], + "type": "text", + "content": "3.2. Instance-wise Knowledge Aggregation (IKA)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Recent studies [19, 20, 29, 31, 32, 39, 48] in 3DIS deal with instance-wise features " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " as essential components, aiming to encode rich semantic and geometric information about instances. They spread hundreds of candidates over expansive 3D scenes to cover as many instances as possible; this number " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " far exceeds the total number " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " of instances present in any single scene (see Tab.4 in Supp.). Although these candidates provide practical information for mask prediction, they often represent separate parts of each instance," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 214 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 214 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 214 + ], + "type": "image", + "image_path": "300b9787ccc7195a5588487da0f6cc05465952974eb66aa20a386dc31e0d56e9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 220, + 555, + 265 + ], + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 265 + ], + "type": "text", + "content": "Figure 3. In IKA, we specify highly correlated instance candidates " + }, + { + "bbox": [ + 313, + 220, + 555, + 265 + ], + "type": "inline_equation", + "content": "(I_{ka} > \\tau)" + }, + { + "bbox": [ + 313, + 220, + 555, + 265 + ], + "type": "text", + "content": " representing identical instances based on the correlation matrix " + }, + { + "bbox": [ + 313, + 220, + 555, + 265 + ], + "type": "inline_equation", + "content": "I_{ka}" + }, + { + "bbox": [ + 313, + 220, + 555, + 265 + ], + "type": "text", + "content": ". Then, we softly aggregate scattered clues of related candidates, optimizing " + }, + { + "bbox": [ + 313, + 220, + 555, + 265 + ], + "type": "inline_equation", + "content": "I_{ka}" + }, + { + "bbox": [ + 313, + 220, + 555, + 265 + ], + "type": "text", + "content": " with pseudo-binary label " + }, + { + "bbox": [ + 313, + 220, + 555, + 265 + ], + "type": "inline_equation", + "content": "I_{ka}^{*}" + }, + { + "bbox": [ + 313, + 220, + 555, + 265 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 273, + 553, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 553, + 309 + ], + "type": "text", + "content": "resulting in fragmented instance masks. With these insights, we focus on associating scattered knowledge to fully leverage the advantages of multiple instance candidates." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": "To handle multiple candidates, we first identify those representing the same instance based on their pairwise affinity. Specifically, we compute a correlation matrix " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "I_{ka} \\in \\mathbb{R}^{N_k \\times N_k}" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": " using the dot product of " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": " and its transpose " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "F_{inst}^T" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": ", then normalize " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "I_{ka}" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": " to a range between 0 and 1. Here, we note that candidates representing the same instance yield relatively high correlation values. Thus, based on " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "I_{ka}" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": ", we consider pairwise candidates with correlation values surpassing a predefined threshold " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": " as fragments from the same instance. Then, as shown in Fig. 3, we softly guide these candidates by optimizing their correlations in " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "I_{ka}" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": ", motivated by the success of recent self-supervised learning strategies [2, 53]. We extend this approach to further enrich coherence among highly correlated candidate features while also encouraging the network to be aware of scattered cues for each instance. In particular, we dynamically construct a pseudo-binary correlation label matrix " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "I_{ka}^* \\in \\{0,1\\}^{N_k \\times N_k}" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": ", where elements of highly correlated pairs " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "I_{ka}" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": " are 1, while the rest " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "(\\leqslant \\tau)" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": " are 0 for all candidate pairs. Given " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "I_{ka}" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "I_{ka}^*" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": ", we formulate the element-wise correlation regularization loss " + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KA}" + }, + { + "bbox": [ + 313, + 309, + 555, + 561 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 562, + 555, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 562, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 319, + 562, + 555, + 594 + ], + "type": "interline_equation", + "content": "I _ {k a} = F _ {i n s t} \\cdot F _ {i n s t} ^ {T}, I _ {k a} ^ {*} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} I _ {k a} (i, j) > \\tau \\\\ 0 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {5}", + "image_path": "f375c56e38050c27358e7c149ade403a98434c09b154ed82cbf4366e7273b0d0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 602, + 554, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 602, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 324, + 602, + 554, + 657 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {K A} = \\sum_ {I _ {k a} > \\tau} \\| I _ {k a} ^ {*} (i, j) - I _ {k a} (i, j) \\| _ {2} \\\\ + w \\cdot \\sum_ {I _ {k a} \\leqslant \\tau} \\| I _ {k a} ^ {*} (i, j) - I _ {k a} (i, j) \\| _ {2} \\tag {6} \\\\ \\end{array}", + "image_path": "84c6d8fcf0ad17406dd57bbe5a451432d454141551769f457e95d80335f0cd3d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " norm, and " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " is the weight value used to balance between highly correlated pairs and others. In Eq. 6, the first term facilitates candidates representing the identical instance closer together in the latent space," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14060" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "whereas the second term decreases correlations between unrelated candidates, thereby reducing confusion from irrelevant knowledge. In conclusion, IKA establishes practical connections that strengthen solidarity across multiple candidates encoding scattered cues by regularizing loss " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KA}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 138, + 270, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 270, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 270, + 152 + ], + "type": "text", + "content": "3.3. Instance-wise Structural Guidance (ISG)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 157, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 296, + 277 + ], + "type": "text", + "content": "In real-world scenarios, 3D instances are usually placed near each other or overlapping, arranged in inconsistent patterns without following conventional standards. Also, these instances are represented as sparse point clouds, which often induce fuzzy inherent noises in point-wise features. Hence, the instance-wise features " + }, + { + "bbox": [ + 55, + 157, + 296, + 277 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 55, + 157, + 296, + 277 + ], + "type": "text", + "content": " from these settings are unclear, causing models to misinterpret the spatial extent of each instance. In this work, we explicitly reduce such ambiguity of " + }, + { + "bbox": [ + 55, + 157, + 296, + 277 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 55, + 157, + 296, + 277 + ], + "type": "text", + "content": " by leveraging intelligent structural guidance from clarity-enhanced instance-wise feature " + }, + { + "bbox": [ + 55, + 157, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{inst}" + }, + { + "bbox": [ + 55, + 157, + 296, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 399 + ], + "type": "text", + "content": "For accurate segmentation, it is crucial to capture the clear morphological shape of complex instances. From this intuition, we aim to encourage the model to better learn the structural appearance of instances. Thus, we implement effective structural guidance based on a truncated singular value decomposition (SVD) [14, 42] algorithm primarily used for noise reduction. As illustrated in Fig. 4, given " + }, + { + "bbox": [ + 55, + 277, + 296, + 399 + ], + "type": "inline_equation", + "content": "F_{inst}^{\\prime} \\in \\mathbb{R}^{N_k^{\\prime} \\times D}" + }, + { + "bbox": [ + 55, + 277, + 296, + 399 + ], + "type": "text", + "content": " from the first point aggregator, we initially decompose " + }, + { + "bbox": [ + 55, + 277, + 296, + 399 + ], + "type": "inline_equation", + "content": "F_{inst}^{\\prime}" + }, + { + "bbox": [ + 55, + 277, + 296, + 399 + ], + "type": "text", + "content": " into three matrices based on SVD: (1) " + }, + { + "bbox": [ + 55, + 277, + 296, + 399 + ], + "type": "inline_equation", + "content": "U \\in \\mathbb{R}^{N_k^{\\prime} \\times N_k^{\\prime}}" + }, + { + "bbox": [ + 55, + 277, + 296, + 399 + ], + "type": "text", + "content": ", (2) " + }, + { + "bbox": [ + 55, + 277, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^{N_k^{\\prime} \\times D}" + }, + { + "bbox": [ + 55, + 277, + 296, + 399 + ], + "type": "text", + "content": ", and (3) " + }, + { + "bbox": [ + 55, + 277, + 296, + 399 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{D \\times D}" + }, + { + "bbox": [ + 55, + 277, + 296, + 399 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 403, + 295, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 403, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 99, + 403, + 295, + 435 + ], + "type": "interline_equation", + "content": "F _ {i n s t} ^ {\\prime} = U \\cdot \\Sigma \\cdot V ^ {T} = \\sum_ {i} ^ {N _ {k} ^ {\\prime}} u _ {i} \\cdot \\sigma_ {i} \\cdot v _ {i} ^ {T} \\tag {7}", + "image_path": "680a09d74476f0549503f6ead4167c1570fbda92b4ea31b1805afc0cfb1bc4bb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "inline_equation", + "content": "U = (u_{1},\\ldots ,u_{n})" + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "inline_equation", + "content": "V = (v_{1},\\dots ,v_{n})" + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "text", + "content": " are the matrices with orthonormal columns, and " + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "inline_equation", + "content": "\\Sigma = diag(\\sigma_1,\\dots,\\sigma_n)" + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "text", + "content": " is a diagonal matrix containing singular values, arranged in the descending order. Typically, vectors with higher singular values in " + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "text", + "content": " include meaningful information for object cognition, while the lower-rank vectors are regarded as less important, potentially causing ambiguity. Then, we preserve only the top " + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "text", + "content": " singular values from " + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "text", + "content": ", along with their corresponding singular vectors from " + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "text", + "content": ", facilitating an ambiguity-reduced yet information-rich representation of the original features " + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "inline_equation", + "content": "F_{inst}^{\\prime}" + }, + { + "bbox": [ + 55, + 440, + 296, + 573 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 578, + 295, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 578, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 73, + 578, + 295, + 611 + ], + "type": "interline_equation", + "content": "\\tilde {F} _ {i n s t} ^ {\\prime} \\in \\mathbb {R} ^ {N _ {k} ^ {\\prime} \\times D} = \\tilde {U} \\cdot \\tilde {\\Sigma} \\cdot \\tilde {V} ^ {T} = \\sum_ {i} ^ {N _ {k} ^ {\\prime}} \\tilde {u} _ {i} \\cdot \\tilde {\\sigma} _ {i} \\cdot \\tilde {v} _ {i} ^ {T} \\tag {8}", + "image_path": "479360f69bb2bcbb3fa5a80c4e39e77d39d492d8de3ae4af9276b61e40bbc1c5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{U} \\in \\mathbb{R}^{N_k' \\times t}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\Sigma} \\in \\mathbb{R}^{t \\times t}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{V} \\in \\mathbb{R}^{D \\times t}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " are truncated matrices using only " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " columns from " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{inst}'" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " is reconstructed from them. Based on both " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_{inst}'" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{inst}'" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", we sample instance-wise features " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", respectively, employing the shared second stage point aggregator. Given " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{inst}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", we calculate the cross-correlation matrix " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "I_{sg} \\in \\mathbb{R}^{N_k \\times N_k}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " and also generate a pseudo-binary correlation label " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "I_{sg}^* \\in \\{0, 1\\}^{N_k \\times N_k}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", as in the IKA network." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 75, + 555, + 209 + ], + "blocks": [ + { + "bbox": [ + 317, + 66, + 395, + 75 + ], + "lines": [ + { + "bbox": [ + 317, + 66, + 395, + 75 + ], + "spans": [ + { + "bbox": [ + 317, + 66, + 395, + 75 + ], + "type": "text", + "content": "Original Instance Feats." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 75, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 316, + 75, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 75, + 555, + 209 + ], + "type": "image", + "image_path": "888d7c58af97bbd2c756c3920ad5fb856cd7a5755b7c55dc218a84f94a591b1c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "text", + "content": "Figure 4. In ISG, we decompose " + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "inline_equation", + "content": "F_{inst}^{\\prime}" + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "text", + "content": " into three matrices " + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "inline_equation", + "content": "(U, \\Sigma, V)" + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "text", + "content": " using the SVD algorithm and keep top " + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "text", + "content": " singular vectors to filter out less important noises. Then, we effectively enhance " + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "text", + "content": " with highlighted essential cues from " + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{inst}" + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "text", + "content": " by optimizing " + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SG}" + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 264, + 554, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 554, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 554, + 289 + ], + "type": "text", + "content": "We then effectively transfer structural clues from " + }, + { + "bbox": [ + 313, + 264, + 554, + 289 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{inst}" + }, + { + "bbox": [ + 313, + 264, + 554, + 289 + ], + "type": "text", + "content": ", utilizing the correlation regularization loss " + }, + { + "bbox": [ + 313, + 264, + 554, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SG}" + }, + { + "bbox": [ + 313, + 264, + 554, + 289 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 291, + 555, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 291, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 319, + 291, + 555, + 323 + ], + "type": "interline_equation", + "content": "I _ {s g} = F _ {i n s t} \\cdot \\tilde {F} _ {i n s t} ^ {T}, I _ {s g} ^ {*} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} I _ {s g} (i, j) > \\tau \\\\ 0 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {9}", + "image_path": "877814e21c32cbacf878afe4236c72684a903342279bb4dbdc829d8a4bfeb1c1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 327, + 554, + 382 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 327, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 324, + 327, + 554, + 382 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {S G} = \\sum_ {I _ {s g} > \\tau} \\left| \\left| I _ {s g} ^ {*} (i, j) - I _ {s g} (i, j) \\right| \\right| _ {2} \\\\ + \\tilde {w} \\cdot \\sum_ {I _ {s g} \\leqslant \\tau} \\left\\| I _ {s g} ^ {*} (i, j) - I _ {s g} (i, j) \\right\\| _ {2} \\tag {10} \\\\ \\end{array}", + "image_path": "b7ed7e6d782b093bf78c1b62019e7e7c135675304ebed3102de9aa64b475c556.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\tilde{w}" + }, + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "text", + "content": " represents the balancing hyper-parameter, and the role of each term in Eq. 10 follows Eq. 6. By optimizing " + }, + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SG}" + }, + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "text", + "content": ", ISG conditionally instructs instance features " + }, + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "text", + "content": " to incorporate spatial configuration details from clarity-enhanced features " + }, + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{inst}" + }, + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "text", + "content": ", improving discriminative power. Loss Function. Finally, our full network minimizes " + }, + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{total}" + }, + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 470, + 553, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 470, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 324, + 470, + 553, + 498 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} \\\\ + \\lambda_ {K A} \\mathcal {L} _ {K A} + \\lambda_ {S G} \\mathcal {L} _ {S G} \\tag {11} \\\\ \\end{array}", + "image_path": "a46e2ad270825a704df14acf733493f4757280b125edbeb8b62e48d04a5c5cbe.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 506, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 530 + ], + "type": "text", + "content": "where each " + }, + { + "bbox": [ + 313, + 506, + 554, + 530 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 506, + 554, + 530 + ], + "type": "text", + "content": " is a hyper-parameter decided via extensive grid searches to balance the impact of respective loss terms." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 540, + 395, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 395, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 395, + 553 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 559, + 430, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 430, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 430, + 572 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 579, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 555, + 664 + ], + "type": "text", + "content": "Datasets. We train and evaluate the overall performance using four landmark benchmarks for 3D Instance Segmentation (3DIS): ScanNetV2 [7], ScanNet200 [38], S3DIS [1], and STPLS3D [5]. These four datasets provide 3D point cloud data collected in diverse real-world environments. Note that detailed descriptions of each dataset and all implementation details are provided in Supplementary Material." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate the 3DIS performance using the Average Precision (AP), a conventional metric in computer vision tasks. We report the mean average precision (mAP) across IoU (Intersection of Union) thresholds" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14061" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 62, + 298, + 309 + ], + "blocks": [ + { + "bbox": [ + 56, + 62, + 298, + 309 + ], + "lines": [ + { + "bbox": [ + 56, + 62, + 298, + 309 + ], + "spans": [ + { + "bbox": [ + 56, + 62, + 298, + 309 + ], + "type": "table", + "html": "
MethodScanNetV2 ValScanNetV2 Test
mAP\\( mAP_{50} \\)\\( mAP_{25} \\)mAP\\( mAP_{50} \\)\\( mAP_{25} \\)
3D-SIS [21]-18.735.716.138.255.8
MTML [28]20.340.255.428.254.973.1
PointGroup [24]34.856.771.340.763.677.8
DyCo3D [19]35.457.672.939.564.176.1
3D-MPA [12]35.559.172.435.561.173.7
MaskGroup [54]42.063.363.343.466.479.2
SoftGroup [44]46.067.678.950.476.186.5
SSTNet [30]49.464.374.050.669.878.9
Mask3D [39]55.273.782.956.678.082.9
SPFormer [43]56.373.982.954.977.085.1
Q-Former [31]56.574.283.358.378.787.4
ISBNet [32]56.873.381.355.976.387.0
MAFT [29]58.475.984.559.678.686.0
EASE [37]60.277.284.259.878.986.7
SphMask [41]62.379.988.261.681.287.5
IKNE (Ours)62.981.888.761.781.288.0
", + "image_path": "cf96c15cb1c6a51e8bd48c1b1d0685f0d2ab2c75606cfd6f0de9914985fbd37f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": "incremented by " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": ", ranging from " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": ". Also, we assess " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{25}" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": ", which measure accuracy based on IoU thresholds of " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": ". For the S3DIS [1], we further provide mean precision (mPrec) and mean recall (mRec) at a " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": " IoU threshold, following previous works." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 418, + 296, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 432 + ], + "type": "text", + "content": "4.2. Performance Comparison with SOTA Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 438, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 438, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 438, + 298, + 715 + ], + "type": "text", + "content": "In this section, we quantitatively compare our proposed framework IKNE with previous methods. Despite significant advances with high scores, these methods, regardless of architecture, still waste considerable instance candidates that may contain valuable information. To maximize the benefits of using numerous candidates, we present novel instance-wise knowledge aggregation (IKA) and guidance (ISG) networks. In Tab. 1, we first evaluate mean average precision (mAP) with different IoU thresholds across 18 classes on the ScanNetV2 [7] dataset. IKNE generally outperforms prior works, achieving new state-of-the-art scores in terms of mAP, " + }, + { + "bbox": [ + 54, + 438, + 298, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 54, + 438, + 298, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 438, + 298, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{25}" + }, + { + "bbox": [ + 54, + 438, + 298, + 715 + ], + "type": "text", + "content": " on both validation and hidden test sets. While our method slightly exceeds the latest coarse-to-fine approach, Spherical Mask [41], ours transcends our primary baselines, which rely on hundreds of instance candidates, such as ISBNet [32] (kernel-based) and MAFT [29] (transformer), by large margins across all metrics. In Tab. 2, we assess 3DIS performance on Area 5 and the 6-fold cross-validation set of the S3DIS [1] dataset. For Area 5 evaluation, we train with Areas 1 to 6, excluding 5, and validate with Area 5; for 6-fold cross-validation, we average the validation scores across all 6 areas. IKNE demonstrates impressive performance in both evaluations, reaching" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 315, + 62, + 558, + 309 + ], + "blocks": [ + { + "bbox": [ + 55, + 314, + 296, + 337 + ], + "lines": [ + { + "bbox": [ + 55, + 314, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 296, + 337 + ], + "type": "text", + "content": "Table 1. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the ScanNetV2 [7] dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 62, + 558, + 309 + ], + "lines": [ + { + "bbox": [ + 315, + 62, + 558, + 309 + ], + "spans": [ + { + "bbox": [ + 315, + 62, + 558, + 309 + ], + "type": "table", + "html": "
MethodS3DIS Area 5S3DIS 6-fold CV
AP\\( AP_{50} \\)\\( Prec_{50} \\)\\( Rec_{50} \\)AP\\( AP_{50} \\)\\( Prec_{50} \\)\\( Rec_{50} \\)
3D-Bonet [50]--57.540.2--65.647.6
PointGroup [24]-57.861.962.1-64.069.669.2
3D-MPA [12]--63.158.0--66.764.1
DyCo3D [19]--64.364.2----
MaskGroup [54]-65.062.964.7-69.966.669.6
SSTNet [30]42.759.365.564.254.167.873.573.4
SoftGroup [44]51.666.173.666.654.468.975.369.8
SPFormer [43]-66.872.867.1-69.274.071.1
ISBNet [32]56.367.570.572.060.870.577.577.1
DKNet [48]--70.865.3--75.371.1
Mask3D [39]56.668.468.766.364.575.572.874.5
Q-Former [31]57.769.970.572.262.073.372.773.4
MAFT [29]-69.1------
EASE [37]59.071.669.468.765.376.473.674.6
SphMask [41]60.572.371.376.364.072.378.177.7
IKNE (Ours)61.173.072.077.065.876.979.478.9
", + "image_path": "dfcde9732cfbaea6622f2c9f5792988cd665fb3117a36a0a9626b7077e0b9e4d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 345, + 550, + 430 + ], + "blocks": [ + { + "bbox": [ + 313, + 314, + 555, + 337 + ], + "lines": [ + { + "bbox": [ + 313, + 314, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 555, + 337 + ], + "type": "text", + "content": "Table 2. Comparison of 3DIS performance with state-of-the-art approaches on the S3DIS [1] Area 5 and 6-fold cross validation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 345, + 550, + 430 + ], + "lines": [ + { + "bbox": [ + 317, + 345, + 550, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 345, + 550, + 430 + ], + "type": "table", + "html": "
MethodmAP\\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{25} \\)
SPFormer [43]25.233.839.6
Mask3D [39]27.437.042.3
Q-Former [31]28.137.143.4
MAFT [29]29.238.243.3
IKNE (Ours)29.938.944.9
", + "image_path": "274f2619bd16d4d5826dfb08c93dce4a990f093ef21b90da5bf1dfce20c3f535.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 463, + 561, + 548 + ], + "blocks": [ + { + "bbox": [ + 313, + 433, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 456 + ], + "type": "text", + "content": "Table 3. Comparison of 3DIS performance with state-of-the-art approaches on the ScanNet200 [38] validation (with 200 classes)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 463, + 561, + 548 + ], + "lines": [ + { + "bbox": [ + 315, + 463, + 561, + 548 + ], + "spans": [ + { + "bbox": [ + 315, + 463, + 561, + 548 + ], + "type": "table", + "html": "
MethodmAP\\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{25} \\)
PointGroup [24]23.338.548.6
HAIS [6]35.146.752.8
SoftGroup [44]47.363.171.4
Mask3D [39]63.479.285.6
IKNE (Ours)64.981.287.6
", + "image_path": "455f2596ccc90686327de63480b344fc07ff919772433b61e2e9ade19d0b2ded.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 552, + 555, + 574 + ], + "lines": [ + { + "bbox": [ + 313, + 552, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 555, + 574 + ], + "type": "text", + "content": "Table 4. Comparison of 3D Instance Segmentation performance with state-of-the-art approaches on the STPLS3D [5] test dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "ing " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "61.1 / 73.0" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " (Area 5) and " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "65.8 / 76.9" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " (6-fold) in " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP} / \\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": ". Also, in Tab. 3, ours achieves the highest scores on the ScanNet200 [38] dataset containing fine-grained 200 categories. Notably, as shown in Tab. 4, IKNE surpasses existing methods on the STPLS3D [5] dataset, including outdoor 3D scenes, with improvements of up to " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "+1.5 / 2.0 / 2.0" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP} / \\mathrm{mAP}_{50} / \\mathrm{mAP}_{25}" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": ". These results highlight that we successfully handle instance candidates, improving their informativeness for understanding real-world instances from both indoor and outdoor environments. Note that “-” in each table indicates unreported scores from published papers." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14062" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 62, + 295, + 135 + ], + "blocks": [ + { + "bbox": [ + 58, + 62, + 295, + 135 + ], + "lines": [ + { + "bbox": [ + 58, + 62, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 58, + 62, + 295, + 135 + ], + "type": "table", + "html": "
Method (Kernel-based)ScanNet ValS3DIS Area 5
Baseline (K)IKAISGmAP / mAP50mAP / mAP50
✓--56.8 / 73.356.3 / 67.5
✓✓-62.7 / 80.860.4 / 72.6
✓-✓62.8 / 81.160.6 / 71.9
✓✓✓62.9 / 81.861.1 / 73.0
", + "image_path": "30486668d0d33bed999d583ba27cb4e630bcb91612e0b75f200192fa7035cd0e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 165, + 296, + 239 + ], + "blocks": [ + { + "bbox": [ + 56, + 137, + 295, + 160 + ], + "lines": [ + { + "bbox": [ + 56, + 137, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 295, + 160 + ], + "type": "text", + "content": "Table 5. Ablation study to see the effect of two main modules (IKA and ISG) based on the kernel-based pipeline (K)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 165, + 296, + 239 + ], + "lines": [ + { + "bbox": [ + 57, + 165, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 57, + 165, + 296, + 239 + ], + "type": "table", + "html": "
Method (Transformer)ScanNet ValS3DIS Area 5
Baseline (T)IKAISGmAP / mAP50mAP / mAP50
✓--58.4 / 75.956.6 / 68.4
✓✓-59.1 / 76.258.8 / 70.7
✓-✓59.4 / 77.258.4 / 70.1
✓✓✓60.8 / 77.959.0 / 71.6
", + "image_path": "c20976d700703fec1449f3d521e5bdfd72eb7cd4f12f1961559ce8a55de311ad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 269, + 296, + 346 + ], + "blocks": [ + { + "bbox": [ + 56, + 241, + 295, + 264 + ], + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 264 + ], + "type": "text", + "content": "Table 6. Ablation study to see the effect of two main modules (IKA and ISG) based on the transformer-based pipeline (T)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 269, + 296, + 346 + ], + "lines": [ + { + "bbox": [ + 58, + 269, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 58, + 269, + 296, + 346 + ], + "type": "table", + "html": "
BaselineMethodKernel-basedTransformer
IKAISGAverage IoUAverage IoU
✓--0.86270.8486
✓✓-0.88280.8728
✓-✓0.88250.8730
✓✓✓0.89040.8803
", + "image_path": "7ddc6c0d7561408e9c1c2e835a22363ab73347923fb2f706f4682f10f5cc656f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 381, + 156, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 156, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 156, + 393 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 402, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 295, + 605 + ], + "type": "text", + "content": "Effect of IKA and ISG Networks. We evaluate variants of our method w/ and w/o IKA and ISG modules based on kernel (K) and transformer-based (T) architectures. As shown in Tab. 5 and 6, the addition of each module surpasses the score of baseline models [29, 32, 39] across all experiments, regardless of model structure. Specifically, IKA boosts performance, with gains of up to " + }, + { + "bbox": [ + 54, + 402, + 295, + 605 + ], + "type": "inline_equation", + "content": "+5.9 / 7.5" + }, + { + "bbox": [ + 54, + 402, + 295, + 605 + ], + "type": "text", + "content": " for ScanNetV2 [7] and " + }, + { + "bbox": [ + 54, + 402, + 295, + 605 + ], + "type": "inline_equation", + "content": "+4.1 / 5.1" + }, + { + "bbox": [ + 54, + 402, + 295, + 605 + ], + "type": "text", + "content": " for S3DIS [1] in " + }, + { + "bbox": [ + 54, + 402, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP} / \\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 54, + 402, + 295, + 605 + ], + "type": "text", + "content": ", underscoring the importance of integrating scattered clues from multiple candidates. Also, ISG contributes to the spatial understanding of instance candidates via intelligent structural guidance from clarity-enhanced features, improving up to " + }, + { + "bbox": [ + 54, + 402, + 295, + 605 + ], + "type": "inline_equation", + "content": "+6.0 / 7.8" + }, + { + "bbox": [ + 54, + 402, + 295, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 402, + 295, + 605 + ], + "type": "inline_equation", + "content": "+4.3 / 4.4" + }, + { + "bbox": [ + 54, + 402, + 295, + 605 + ], + "type": "text", + "content": ". Eventually, utilizing all modules together gains further advances, resulting in the best performance with notable progress on both structures. These results demonstrate each module plays a meaningful role in improving the discriminative capability of instance features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 606, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 606, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 606, + 295, + 714 + ], + "type": "text", + "content": "Furthermore, to assess how each module influences the accurate determination of instance mask coverage, we use the IoU metric, which is closely associated with segmentation performance. IoU calculates the ratio of the intersection to the union of predicted and ground truth masks. In Tab. 7, we compare the average IoU of our predicted instance masks and those of the baselines for both architectures. As shown in Fig. 1 (a) and Fig.1 (a) (in Supp.), for previous works, a single instance is often separated into sev-" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 62, + 554, + 200 + ], + "blocks": [ + { + "bbox": [ + 55, + 349, + 295, + 372 + ], + "lines": [ + { + "bbox": [ + 55, + 349, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 295, + 372 + ], + "type": "text", + "content": "Table 7. Ablation study to see the effect of two main modules across both pipelines, using IoU scores on the ScanNetV2 [7] val." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 62, + 554, + 200 + ], + "lines": [ + { + "bbox": [ + 317, + 62, + 554, + 200 + ], + "spans": [ + { + "bbox": [ + 317, + 62, + 554, + 200 + ], + "type": "table", + "html": "
ArchitectureThreshold τScanNet ValS3DIS Area 5
mAP / mAP50mAP / mAP50
Baseline (K)-56.8 / 73.356.3 / 67.5
Kernel-based0.653.6 / 73.455.3 / 65.5
Kernel-based0.757.2 / 76.459.1 / 69.2
Kernel-based0.861.1 / 79.960.7 / 72.3
Kernel-based0.962.9 / 81.861.1 / 73.0
Baseline (T)-58.4 / 75.956.6 / 68.4
Transformer0.657.9 / 75.856.3 / 67.5
Transformer0.759.2 / 76.257.4 / 68.6
Transformer0.859.9 / 77.158.6 / 70.4
Transformer0.960.8 / 77.959.0 / 71.6
", + "image_path": "79e068e56c904992c6d4d42c005aa1331e2eb4bafd421a2efb0292a3086faa22.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 204, + 553, + 227 + ], + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 227 + ], + "type": "text", + "content": "Table 8. Ablation study to investigate the threshold " + }, + { + "bbox": [ + 313, + 204, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 204, + 553, + 227 + ], + "type": "text", + "content": " for determining instance-wise candidates representing the same instance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 233, + 435, + 382 + ], + "blocks": [ + { + "bbox": [ + 317, + 233, + 435, + 382 + ], + "lines": [ + { + "bbox": [ + 317, + 233, + 435, + 382 + ], + "spans": [ + { + "bbox": [ + 317, + 233, + 435, + 382 + ], + "type": "image", + "image_path": "1a966361c95c2551373d737051f9b0fec27598f3f8fad077fe92df7ddef9edc1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 387, + 555, + 410 + ], + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 410 + ], + "type": "text", + "content": "Figure 5. Comparison to explore the optimal top " + }, + { + "bbox": [ + 313, + 387, + 555, + 410 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 387, + 555, + 410 + ], + "type": "text", + "content": " for reducing ambiguity while keeping important cues of instance features." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 437, + 234, + 553, + 376 + ], + "blocks": [ + { + "bbox": [ + 437, + 234, + 553, + 376 + ], + "lines": [ + { + "bbox": [ + 437, + 234, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 437, + 234, + 553, + 376 + ], + "type": "image", + "image_path": "868289e5a6cf919f05113b6dcd340f90a55d7574a560f2e6ddffad8e790aa8d4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 422, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 567 + ], + "type": "text", + "content": "eral fragments. These fragments potentially cause low IoU scores. The IoU improvements with IKA across both architectures confirm its effectiveness in reducing the negative impact of multiple fragments and enhancing the accuracy of mask coverage. Also, the IoU tends to be low, particularly when instance masks include the surrounding backgrounds or when adjacent instances are not clearly distinguished, as in Fig. 1 (b) and Fig.1 (b) (in Supp.), leading to overestimated or underestimated instance masks. In addition to IKA, ISG enhances the IoU scores, verifying our guidance is sufficiently practical. We further validate the significance of ours with additional experiments in Supp. (see Sec.4)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Thresholds " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " for Instance-wise Identification. We specify candidates representing the same instance using predefined threshold " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " from the correlation matrices " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "I_{ka}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "I_{sg}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": ". In Tab. 8, based on kernel (K) and transformer (T) architectures [29, 32, 39], we explore the impact of " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " for useful knowledge interaction. When " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " is lower than 0.7, the scores degrade due to the inclusion of inaccurate information from different instance representative features. Yet, with more precise identification using the higher " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\geqslant 0.8" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "), our knowledge interaction yields considerable improvements. This result emphasizes that clues from accurately identified candidates are precious for enhancing instance interpretation. We" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14063" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 52, + 553, + 216 + ], + "blocks": [ + { + "bbox": [ + 57, + 52, + 553, + 216 + ], + "lines": [ + { + "bbox": [ + 57, + 52, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 57, + 52, + 553, + 216 + ], + "type": "image", + "image_path": "368dcd2e9bd0908366bd5301ac51392c02a22e2858876da82abdf0f6f7933ce5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of instance (Inst.) and semantic (Sem.) masks between the baselines (ISBNet (K) [32], MAFT (T) [29]) and ours with Ground Truth masks on the ScanNetV2. The critical differences are highlighted using yellow boxes. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 249, + 282, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 282, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 282, + 261 + ], + "type": "text", + "content": "also examine " + }, + { + "bbox": [ + 55, + 249, + 282, + 261 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 249, + 282, + 261 + ], + "type": "text", + "content": " for the STPLS3D [5] in Supp. (Sec.4.3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "content": "Analysis of the Top t values. In this work, we truncate the top " + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "content": " vectors of " + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "inline_equation", + "content": "(U,\\Sigma ,V)" + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "inline_equation", + "content": "F_{inst}^{\\prime}" + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "content": " and reconstruct " + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{inst}^{\\prime}" + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "inline_equation", + "content": "(\\tilde{U},\\tilde{\\Sigma},\\tilde{V})" + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "content": " based on a truncated SVD algorithm [17]. Here, the " + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "content": " value impacts the balance between data compression and information preservation. Therefore, it is crucial to find the optimal value of " + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "content": " for proper balance. To this end, we conduct experiments using various " + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "content": " for both kernel (blue) and transformer (orange) architectures in Fig. 5. We represent " + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "content": " as a percentage, so the x-axis indicates the proportion of vectors retained during truncation across three matrices. We observe that the trends of all experiments are similar. In settings with a lower " + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "content": " percentage (0.1), the scores consistently decrease due to large information loss. However, with a higher " + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "content": " (0.7), the effect of noise reduction is insufficient; thereby, the performance is comparable to baselines [29, 32]. We ultimately set " + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 262, + 296, + 479 + ], + "type": "text", + "content": " as 0.5, resulting in the highest accuracy while allowing us to preserve meaningful information and effectively filter out inherent noises." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 487, + 167, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 167, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 167, + 499 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 507, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 686 + ], + "type": "text", + "content": "Visual Comparison. In Fig. 6, we qualitatively validate the effectiveness of our novel framework IKNE. We visualize the predicted semantic (Sem.) and instance (Inst.) masks of the kernel (K) [32] and transformer (T) [29] baselines with ours on the ScanNetV2 [7]. We also provide the corresponding ground truth for solid comparisons. In Scene 1, ours accurately segments the cabinet as a single instance, whereas the baselines separate it into multiple fragments. In Scene 2, ours correctly distinguishes neighboring instances (four chairs and a table), while baselines erroneously recognize two chairs as one (K) or a table and a chair as one (T). Also, in Scene 3, where objects are disorderly adjacent, ours clearly captures the spatial range of the desk and chairs compared to the baselines. These qualitative results verify that our instance features contain highly informative cues." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Structural Guidance from " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_{\\mathrm{inst}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ". In Fig. 7, we qualitatively confirm the significance of structural cues from the" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 247, + 551, + 399 + ], + "blocks": [ + { + "bbox": [ + 316, + 247, + 551, + 399 + ], + "lines": [ + { + "bbox": [ + 316, + 247, + 551, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 247, + 551, + 399 + ], + "type": "image", + "image_path": "dc58b0ff2a13755be704cf1499da2153084d4d029327480edc6d57100f8932af.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 403, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 314, + 403, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 314, + 403, + 555, + 427 + ], + "type": "text", + "content": "Figure 7. Instance feature visualizations of the original features " + }, + { + "bbox": [ + 314, + 403, + 555, + 427 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 314, + 403, + 555, + 427 + ], + "type": "text", + "content": " (blue) and ambiguity-reduced features " + }, + { + "bbox": [ + 314, + 403, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\bar{F}_{inst}" + }, + { + "bbox": [ + 314, + 403, + 555, + 427 + ], + "type": "text", + "content": " (red)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "text", + "content": "noise-reduced features " + }, + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{inst}" + }, + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "text", + "content": " in smartly guiding the original features " + }, + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "text", + "content": ". As shown in Scene 1 and 2, instance features (red) from " + }, + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{inst}" + }, + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "text", + "content": " accurately capture the spatial range of instances (chair and desk), whereas instance features (blue) from " + }, + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "inline_equation", + "content": "F_{inst}" + }, + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "text", + "content": " struggle to cover the complete coverage. Also, in Scene 3, where the distinction is unclear (blue) from surroundings (curtain and wall), the candidate (red) from " + }, + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{inst}" + }, + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "text", + "content": " clearly segments such instances. These findings highlight guidance from clarity-enhanced features is worthwhile." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 549, + 388, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 549, + 388, + 561 + ], + "spans": [ + { + "bbox": [ + 314, + 549, + 388, + 561 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "text", + "content": "In this paper, we introduce a novel instance-wise knowledge enhancement approach, IKNE, for 3D instance segmentation. We focus on optimizing the efficiency of hundreds of instance candidates by effectively handling those representing the same instance. To address these candidates, we propose an IKA to integrate spread clues based on correlation regularization. Also, we present an ISG, which enhances instance features with fundamental cues for object understanding from ambiguity-reduced instance features using the truncated SVD. Our extensive experiments demonstrate the effectiveness of our methods, achieving new state-of-the-art results on both kernel and transformer architectures." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14064" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 312 + ], + "type": "text", + "content": "This work was supported by the Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2024 (RS-2023-00227409, Development of sketch-based semantic 3D modeling technology for creating user-centric Metaverse content spaces for indoor spaces, " + }, + { + "bbox": [ + 55, + 90, + 297, + 312 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 90, + 297, + 312 + ], + "type": "text", + "content": "; RS-2024-00345025, International Collaborative Research and Global Talent Development for the Development of Copyright Management and Protection Technologies for Generative AI, " + }, + { + "bbox": [ + 55, + 90, + 297, + 312 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 90, + 297, + 312 + ], + "type": "text", + "content": "), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2019-II190079, Artificial Intelligence Graduate School Program (Korea University), " + }, + { + "bbox": [ + 55, + 90, + 297, + 312 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 90, + 297, + 312 + ], + "type": "text", + "content": "; IITP-2025-RS-2024-00436857, " + }, + { + "bbox": [ + 55, + 90, + 297, + 312 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 90, + 297, + 312 + ], + "type": "text", + "content": "), the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2025-00521602, " + }, + { + "bbox": [ + 55, + 90, + 297, + 312 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 90, + 297, + 312 + ], + "type": "text", + "content": "), and Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT (MSIT, Korea) & Gwangju Metropolitan City." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 333, + 115, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 115, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 115, + 346 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 353, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 353, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 353, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 61, + 353, + 296, + 408 + ], + "type": "text", + "content": "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 409, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 409, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 61, + 409, + 296, + 453 + ], + "type": "text", + "content": "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 454, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 296, + 487 + ], + "type": "text", + "content": "[3] Ori Bryt and Michael Elad. Compression of facial images using the k-svd algorithm. Journal of Visual Communication and Image Representation, 19(4):270-282, 2008. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 488, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 521 + ], + "type": "text", + "content": "[4] Chin-Chen Chang, Piyu Tsai, and Chia-Chen Lin. Svd-based digital image watermarking scheme. Pattern Recognition Letters, 26(10):1577-1586, 2005. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 522, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 578 + ], + "type": "text", + "content": "[5] Meida Chen, Qingyong Hu, Zifan Yu, Hugues Thomas, Andrew Feng, Yu Hou, Kyle McCullough, Fengbo Ren, and Lucio Soibelman. Stpls3d: A large-scale synthetic and real aerial photogrammetry 3d point cloud dataset. arXiv preprint arXiv:2203.09065, 2022. 2, 5, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 579, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 296, + 633 + ], + "type": "text", + "content": "[6] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "type": "text", + "content": "[7] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 296, + 714 + ], + "type": "text", + "content": "[8] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "Proceedings of the European conference on computer vision (ECCV), pages 562-578, 2018. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 95, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 95, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 319, + 95, + 554, + 139 + ], + "type": "text", + "content": "[9] Renwei Dian, Shutao Li, and Xudong Kang. Regularizing hyperspectral and multispectral image fusion by cnn denoiser. IEEE transactions on neural networks and learning systems, 32(3):1124-1135, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "text", + "content": "[10] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. IEEE transactions on image processing, 6(9):1305-1315, 1997. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 184, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 228 + ], + "type": "text", + "content": "[11] Cathrin Elich, Francis Engelmann, Theodora Kontogianni, and Bastian Leibe. 3d bird's-eye-view instance segmentation. In German Conference on Pattern Recognition, pages 48-61. Springer, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "text", + "content": "[12] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 284, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 348 + ], + "type": "text", + "content": "[13] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3722-3731, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[14] Gene H Golub and Christian Reinsch. Singular value decomposition and least squares solutions. In Handbook for Automatic Computation: Volume II: Linear Algebra, pages 134-151. Springer, 1971. 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "text", + "content": "[15] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "text", + "content": "[16] Qiang Guo, Caiming Zhang, Yunfeng Zhang, and Hui Liu. An efficient svd-based method for image denoising. IEEE transactions on Circuits and Systems for Video Technology, 26(5):868-880, 2015. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "type": "text", + "content": "[17] Per Christian Hansen. The truncated svd as a method for regularization. BIT Numerical Mathematics, 27:534-553, 1987. 3, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 526, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 568 + ], + "type": "text", + "content": "[18] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 570, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 555, + 624 + ], + "type": "text", + "content": "[19] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3d: Robust instance segmentation of 3d point clouds through dynamic convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 354-363, 2021. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[20] Tong He, Wei Yin, Chunhua Shen, and Anton Van den Hengel. Pointinst3d: Segmenting 3d instances by points. In European Conference on Computer Vision, pages 286-302. Springer, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[21] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 1, 2, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14065" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[22] Peg Howland, Moongu Jeon, and Haesun Park. Structure preserving dimension reduction for clustered text data based on the generalized singular value decomposition. SIAM Journal on Matrix Analysis and Applications, 25(1):165-179, 2003. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[23] Sujin Jang, Dae Ung Jo, Sung Ju Hwang, Dongwook Lee, and Daehyun Ji. Stxd: Structural and temporal cross-modal distillation for multi-view 3d object detection. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 295, + 227 + ], + "type": "text", + "content": "[24] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 280 + ], + "type": "text", + "content": "[25] XiaoBing Kang and ShengMin Wei. Identifying tampered regions using singular value decomposition in digital image forensics. In 2008 International conference on computer science and software engineering, pages 926-930. IEEE, 2008. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 294, + 337 + ], + "type": "text", + "content": "[26] Sungjune Kim, Hyung-gun Chi, Hyerin Lim, Karthik Ramani, Jinkyu Kim, and Sangpil Kim. Higher-order relational reasoning for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15251-15260, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "text", + "content": "[27] Virginia Klema and Alan Laub. The singular value decomposition: Its computation and some applications. IEEE Transactions on automatic control, 25(2):164-176, 1980. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "text", + "content": "[28] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9256-9266, 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 469 + ], + "type": "text", + "content": "[29] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-attention-free transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3693–3703, 2023. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 524 + ], + "type": "text", + "content": "[30] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "type": "text", + "content": "[31] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18516-18526, 2023. 1, 2, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 646 + ], + "type": "text", + "content": "[32] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. Isbnet: a 3d point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13550-13559, 2023. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[33] HS Prasantha, HL Shashidhara, and KN Balasubramanya Murthy. Image compression using svd. In International conference on computational intelligence and multimedia applications (ICCIMA 2007), pages 143-145. IEEE, 2007. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[34] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[35] Ajit Rajwade, Anand Rangarajan, and Arunava Banerjee. Image denoising using the higher order singular value decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(4):849-862, 2012. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "type": "text", + "content": "[36] Zhongzheng Ren, Ishan Misra, Alexander G Schwing, and Rohit Girdhar. 3d spatial recognition without spatially labeled 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13213, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "type": "text", + "content": "[37] Wonseok Roh, Hwanhee Jung, Giljoo Nam, Jinseop Yeom, Hyunjie Park, Sang Ho Yoon, and Sanggil Kim. Edge-aware 3d instance segmentation network with intelligent semantic prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20644–20653, 2024. 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "text", + "content": "[38] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 306, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 348 + ], + "type": "text", + "content": "[39] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d for 3d semantic instance segmentation. arXiv preprint arXiv:2210.03105, 2022. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[40] Qiquan Shi, Yiu-Ming Cheung, and Jian Lou. Robust tensor svd and recovery with rank estimation. IEEE Transactions on Cybernetics, 52(10):10667-10682, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[41] Sangyun Shin, Kaichen Zhou, Madhu Vankadari, Andrew Markham, and Niki Trigoni. Spherical mask: Coarse-to-fine 3d point cloud instance segmentation with spherical representation. arXiv preprint arXiv:2312.11269, 2023. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 427, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 448 + ], + "type": "text", + "content": "[42] Gilbert W Stewart. On the early history of the singular value decomposition. SIAM review, 35(4):551-566, 1993. 2, 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "text", + "content": "[43] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 494, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 553, + 547 + ], + "type": "text", + "content": "[44] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "type": "text", + "content": "[45] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2569-2578, 2018. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 604, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 657 + ], + "type": "text", + "content": "[46] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[47] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scenegraphfusion: Incremental 3d scene graph prediction from rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7515-7525, 2021. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14066" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 429 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[48] Yizheng Wu, Min Shi, Shuaiyuan Du, Hao Lu, Zhiguo Cao, and Weicai Zhong. 3d instances as 1d kernels. In European Conference on Computer Vision, pages 235-252. Springer, 2022. 1, 2, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "type": "text", + "content": "[49] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Dening Lu, Mingqiang Wei, and Jun Wang. Venet: Voting enhancement network for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3712-3721, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[50] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds. Advances in neural information processing systems, 32, 2019. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 272 + ], + "type": "text", + "content": "[51] Dan Yang, Zongming Ma, and Andreas Buja. A sparse singular value decomposition method for high-dimensional data. Journal of Computational and Graphical Statistics, 23 (4):923-942, 2014. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 295, + 328 + ], + "type": "text", + "content": "[52] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3947-3956, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "text", + "content": "[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pages 12310-12320. PMLR, 2021. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 429 + ], + "type": "text", + "content": "[54] Min Zhong, Xinghao Chen, Xiaokang Chen, Gang Zeng, and Yunhe Wang. Maskgroup: Hierarchical point grouping and masking for 3d instance segmentation. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 6" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14067" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/ada2a42d-606c-400d-85d5-a68b46a5f62b_content_list.json b/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/ada2a42d-606c-400d-85d5-a68b46a5f62b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1515885c7779efed47878f3d5fa7283e89117083 --- /dev/null +++ b/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/ada2a42d-606c-400d-85d5-a68b46a5f62b_content_list.json @@ -0,0 +1,1506 @@ +[ + { + "type": "text", + "text": "Inst3D-LMM: Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning", + "text_level": 1, + "bbox": [ + 217, + 130, + 782, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hanxun Yu $^{1*}$ , Wentong Li $^{2*}$ , Song Wang $^{1}$ , Junbo Chen $^{3\\dagger}$ , Jianke Zhu $^{1\\dagger}$ $^{1}$ Zhejiang University $^{2}$ Nanjing University of Aeronautics and Astronautics $^{3}$ Udeer.ai {hanxun.yu, songw, jkzhu}@zju.edu.cn, wentong.li@nuaa.edu.cn, junbo@udeer.ai", + "bbox": [ + 153, + 202, + 849, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 294, + 328, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite encouraging progress in 3D scene understanding, it remains challenging to develop an effective Large Multi-modal Model (LMM) that is capable of understanding and reasoning in complex 3D environments. Most previous methods typically encode 3D point and 2D image features separately, neglecting interactions between 2D semantics and 3D object properties, as well as the spatial relationships within the 3D environment. This limitation not only hinders comprehensive representations of 3D scene, but also compromises training and inference efficiency. To address these challenges, we propose a unified Instance-aware 3D Large Multi-modal Model (Inst3D-LMM) to deal with multiple 3D scene understanding tasks simultaneously. To obtain the fine-grained instance-level visual tokens, we first introduce a novel Multi-view Cross-Modal Fusion (MCMF) module to inject the multi-view 2D semantics into their corresponding 3D geometric features. For scene-level relation-aware tokens, we further present a 3D Instance Spatial Relation (3D-ISR) module to capture the intricate pairwise spatial relationships among objects. Additionally, we perform end-to-end multi-task instruction tuning simultaneously without the subsequent task-specific fine-tuning. Extensive experiments demonstrate that our approach outperforms the state-of-the-art methods across 3D scene understanding, reasoning and grounding tasks. Source code is available at: https://github.com/hanxunyu/Inst3D-LMM.", + "bbox": [ + 89, + 327, + 485, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 763, + 223, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building Large Multi-modal Models (LMMs) for 3D scene understanding becomes an emerging research topic with significant potential for advancing autonomous robotics [36]. For example, interactive embodied agents [39] are expected to interpret 3D layouts and predict", + "bbox": [ + 89, + 787, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "object locations based on human instructions.", + "bbox": [ + 511, + 296, + 815, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional 3D scene understanding methods [30, 46] are typically tailored for individual downstream tasks, such as 3D Visual Grounding (3D-VG), 3D Question Answering (3D-QA) and 3D Dense Captioning (3D-DC). In contrast, LMMs are able to handle various 3D perception tasks within a single model. Some methods [16, 43] primarily focus on translating 3D points into the space of 2D Vision Language Models (VLMs) or directly leveraging multiview 2D features as 3D representations. Alternatively, other approaches [6, 45, 58] directly encode the features of 3D points and facilitate alignment with LLM using 3D-text instruction data. However, they often require multi-stage alignment or language-scene pre-training, complicating the development of a versatile model capable of handling multiple tasks. To enable a unified 3D LMM framework, recent work [19] decomposes the input 3D scene into a set of individual object proposals, each identified by unique tokens to explicitly capture 3D object features at the instance level. While this approach exhibits promising results, it neglects the interactions between the 2D semantic features and the properties of 3D objects, as well as the spatial relationship modeling among objects in 3D environments. This oversight further results in substantial token costs for the LLM, thus hindering both training and inference efficiency.", + "bbox": [ + 511, + 311, + 906, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose Inst3D-LMM, an effective Instance-aware 3D Large Multi-modal Model that tackles multiple 3D-language tasks without resorting to task-specific fine-tuning. Our approach fully leverages the powerful 2D Vision Foundation Models (VFMs) and 3D specialist models to extract enriched 2D and 3D features at the instance level, respectively. As shown in Figure 1, in contrast to previous methods, our approach is able to generate fine-grained instance-level representations that encapsulate both geometric and semantic properties, and scene-level representations that capture intricate pairwise spatial relationships among objects in a 3D scene. Moreover, our method results in minor token costs for LLM, thereby enhancing both training and inference efficiency. By leveraging this instance-aware methodology, our Inst3D-LMM", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † Corresponding authors.", + "bbox": [ + 107, + 875, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14147", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8fefee1109f574bdeac6eea50d061c38485d98b8381a99f38aa3e2f0b78da53d.jpg", + "image_caption": [ + "Figure 1. Comparisons between previous 3D LMMs and our proposed Inst3D-LMM. (a) Previous methods [6, 16, 19, 20] typically encode the features of 3D points or 2D images separately and concatenate them directly, often requiring task-specific fine-tuning for different tasks. (b) Illustration of Inst3D-LMM. Our method integrates 2D/3D cross-modal information and captures the intricate spatial relations among objects within 3D environments to generate tight but informative instance/scene-level tokens for the LMM. (c) Compared with other 3D LMMs, our Inst3D-LMM requires fewer computational resources, while offering faster training and inference speeds." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 617, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/59392d37dc948e2cecfddbeef78a690e045fc039d2a2b20cc70ad0b4e050f104.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 89, + 903, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "significantly improves the LLM's ability to comprehend 3D scenes with respect to both efficiency and accuracy.", + "bbox": [ + 88, + 369, + 482, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we introduce a novel Multi-view CrossModal Fusion (MCMF) module that effectively infuses enriched multi-view 2D features into the original 3D features with coarse semantics. A learnable [CLS] token is introduced to aggregate the characteristics of each 2D view, enabling efficient multi-view 2D-to-3D cross-modal transformation. To capture intricate pairwise spatial relationships among objects in a 3D scene, we then propose a 3D Instance Spatial Relation (3D-ISR) module. A spatial condition self-attention between manifold position embeddings and instance-level tokens is presented to produce relation-aware scene-level representations. The resulting instance-level and scene-level representations are subsequently fed into the LLM for end-to-end multi-task instruction tuning.", + "bbox": [ + 88, + 401, + 482, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments across various tasks, including 3D-VG, 3D-QA and 3D-DC, demonstrate that our approach outperforms previous state-of-the-art methods with leading 3D scene understanding, grounding and reasoning capabilities. Unlike most existing methods that focus on close-set scene understanding or require per-task fine-tuning, our Inst3D-LMM operates as a generalist model. We believe this work lays a fundamental step towards unifying diverse 3D vision-language tasks in generative language modeling. To summarize, our contributions are as follows:", + "bbox": [ + 89, + 613, + 482, + 762 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a unified and efficient instance-aware LLM-based framework, called Inst3D-LMM for various 3D scene understanding tasks with end-to-end multi-modal instruction tuning. Serving as a generalist model, our approach demonstrates superior performance across 3D scene understanding, reasoning and spatial localization.", + "- We utilize 2D VFMs to extract multi-view contextual features for each 3D instance and then devise a Multi-view Cross-Modal Fusion (MCMF) module to effectively en" + ], + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "hance instance-level feature representations by jointly integrating 3D geometry and 2D semantic priors.", + "bbox": [ + 524, + 369, + 903, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- A 3D Instance Spatial Relation (3D-ISR) module is introduced to boost the capability of LMM in understanding the complex spatial details within 3D scenes.", + "bbox": [ + 513, + 401, + 903, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 464, + 661, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Scene Understanding with Language. In 3D scene understanding, there is a surge of interest in making use of language queries to capture user intentions for various downstream tasks, such as 3D Visual Grounding [5, 44], 3D Question Answering [30, 33] and 3D Dense Captioning [8, 21]. Specifically, 3D Visual Grounding entails localizing target objects based on language queries. Moreover, 3D Question Answering demands robust 3D spatial perception and reasoning. 3D Dense Captioning involves localizing and describing objects in 3D scenes. Conventional methods typically focus on a specific task. Instead, 3D visual grounding and dense captioning tasks are combined by leveraging their complementary aspects [2, 9, 31, 47]. Recent efforts like 3D-VLP [22] and 3D-VisTA [60] attempt to establish a universal framework by pre-aligning 3D scenes with their corresponding textual descriptions. In contrast to our Inst3D-LMM, most existing methods still focus on close-set scene understanding, which requires either task-specific fine-tuning or striving to build specialized models.", + "bbox": [ + 511, + 491, + 906, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Large Multi-modal Models. Inspired by the significant advancements in Large Language Models (LLMs), researchers extend LLM's knowledge to encompass 3D modality [29, 49, 57, 59]. Point-LLM [14] and Imagebind-LLM [15] have succeeded in bridging the gap between 3D visuals and text by utilizing extensive 3D object datasets. However, these models struggle with interpreting complex spatial relationships in 3D scenes. Another promising di", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14148", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fd36b538b11330d06979505d777b82f51104af180a9302c9d3e26b2ef8929019.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed Inst3D-LMM. Our pipeline takes as input point clouds of a 3D indoor scene, along with RGB-D images. We first employ the pre-trained 3D specialist models and 2D VFMs to extract 3D proposals and multi-view 2D semantic features, respectively. We then suggest the MCMF module to generate fine-grained instance-level tokens. A 3D-ISR module is further introduced to create relation-aware scene-level tokens based on spatial distances. By leveraging the constructed 3D-language prompts, we conduct multi-task instruction tuning to simultaneously handle various 3D tasks." + ], + "image_footnote": [], + "bbox": [ + 98, + 87, + 903, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rection focuses on developing the scene-level 3D LMMs. Hong et al. [16] encode projected 3D features using a 2D vision encoder and incorporate location tokens to augment LLM vocabulary. Other methods [6, 45, 58] directly encode point clouds and utilize 3D scene-text data for better visual interaction through pre-alignment. Huang et al. [18] employ a three-stage training scheme and adopt object identifiers to learn individual object attributes. Chen et al. [7] use special referent tokens for precise referencing and grounding. Wang et al. [43] encode point clouds and RGB images separately. However, some crucial semantic pixels are often lost during their sparse fusion processes, resulting in a coarse visual semantic representation. In this work, we propose an effective instance-aware framework to fuse fine-grained cross-modal information and encode spatial relations, which achieves promising results on multiple 3D-language tasks as a generalist model.", + "bbox": [ + 88, + 565, + 485, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 842, + 225, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to enable LLM to understand the 3D environment and perform various visual interaction tasks based on", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "human instructions. Figure 2 illustrates the architecture of our framework. In this section, we first introduce how to extract features at the instance level using pre-trained 3D models [38, 57] and 2D VFs [23, 35], respectively. Secondly, we present a novel Multi-view Cross-Modal Fusion (MCMF) module to obtain fine-grained instance-level tokens, which are specially designed to effectively integrate 3D geometric features with their corresponding multi-view 2D semantic features. Thirdly, we introduce the 3D Instance Spatial Relation (3D-ISR) module to enhance LLM's ability to capture spatial information at the scene level, which generates relation-aware tokens through attention-based analysis of spatial relationships among different 3D proposals. Under our proposed framework, the MCMF and 3D-ISR modules are jointly optimized, allowing mutual enhancement. Finally, we conduct end-to-end multi-task instruction tuning to address a range of 3D scene understanding tasks simultaneously.", + "bbox": [ + 511, + 565, + 906, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Instance-Level Feature Extraction", + "text_level": 1, + "bbox": [ + 511, + 847, + 812, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Feature Extraction. We first segment the 3D point clouds of each individual instance in a class-agnostic man", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14149", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ner by leveraging a pre-trained 3D instance segmentation model [38]. We only retain the predicted binary 3D instance masks while ignoring their closed-vocabulary class labels. All instance proposals in one scene are represented by $O^{3D} = (O_1^{3D}, \\ldots, O_N^{3D})$ , where $O_i^{3D} = [coordinate, color]$ consists of the attributes of each instance. We obtain all instance proposals of a scene, and the pre-trained 3D encoder $\\mathbf{E}_{\\mathbf{o}}$ [57] is used to extract their instance-level features, i.e., $\\mathbf{f}_{\\mathbf{o}}^{3D} = \\mathbf{E}_{\\mathbf{o}}(O^{3D})$ .", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2D Feature Extraction. Due to the inherent sparsity of 3D point cloud data, previous methods have difficulties in generating discriminative features of each object. In this work, we employ powerful 2D VFMs to extract 2D semantic features for each 3D instance. As in [3, 40], we firstly project the point cloud of each instance $O_{i}^{3D}$ onto the image plane and then select the top $K$ views according to the number of visible points. To obtain accurate 2D masks, we randomly sample $k_{\\mathrm{sample}} = 5$ points as input prompts for SAM [23] to deal with noisy bounding boxes with outliers. Hereby, we select the high-quality mask with the highest confidence score. To enrich features with contextual information, these masks are employed to generate multi-level $(L)$ crops of the selected images that are further fed into the pre-trained CLIP vision encoder [35] to extract features with language-aligned embedding space. Finally, we aggregate multi-level features of sub-images from the same frame to form the 2D multi-view features $O^{2D}$ .", + "bbox": [ + 89, + 227, + 482, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Multi-view Cross-Modal Fusion", + "text_level": 1, + "bbox": [ + 89, + 508, + 370, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To better fuse 3D geometry priors with 2D multi-view semantic priors, we introduce an effective Multi-view CrossModal Fusion (MCMF) module that generates the enriched token representations for each 3D instance before being fed into LLM. The architecture of MCMF is crafted with a coarse-to-fine framework, as shown in Figure 3. To map 3D object features $O^{3D} \\in \\mathbb{R}^{1 \\times N \\times D^{3d}}$ and 2D CLIP features $O^{2D} \\in \\mathbb{R}^{K \\times N \\times D^{2d}}$ into the embedding space of LLM with the dimensionality of $D$ , we utilize a simple two-layer MLP with a LayerNorm and GELU in between, yielding $O^{3D'}$ and $O^{2D'}$ , respectively. $N$ represents the number of instances, and $K$ is the number of images for each object.", + "bbox": [ + 89, + 532, + 483, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, we introduce a Cross-Modal Injection Block to transform the enriched semantic priors from 2D multi-view representations into 3D instance features. Moreover, we adopt a straightforward self-attention layer to further enhance such 3D instance features. For 2D multi-view features, we append a learnable [CLS] token $t_k$ to the flattened CLIP feature maps in order to adaptively encapsulate the global semantic representation of the $k$ -th view. Then, we apply a self-attention layer to multi-view features, enabling semantic gathering to derive $t_k'$ for $k$ -th view as:", + "bbox": [ + 89, + 718, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {k} ^ {\\prime} = \\phi (\\text {S e l f A t t n} ([ t _ {k}, O _ {\\text {v i e w} _ {k}} ^ {2 D ^ {\\prime}} ])), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 882, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e26d593e3821672fd474eed8c2f89be47e2793c7ee27458b3983d6ffe30c7784.jpg", + "image_caption": [ + "Figure 3. Architecture of the proposed Multi-view Cross-Modal Fusion (MCMF) module." + ], + "image_footnote": [], + "bbox": [ + 519, + 89, + 903, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $1 \\leq k \\leq K$ , $O_{\\mathrm{view}_k}^{2D'} \\in \\mathbb{R}^{N \\times D}$ . $\\phi$ is an indicator function that outputs the updated [CLS] token as the first one from the token list.", + "bbox": [ + 511, + 349, + 903, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We obtain the processed 3D token embedding $O^{3D'} \\in \\mathbb{R}^{1 \\times N \\times D}$ containing coarse semantics, where each instance in $O^{3D'}$ corresponds to multiple views of the processed 2D CLIP features $O^{2D'}$ . It can be represented as follows:", + "bbox": [ + 511, + 396, + 903, + 458 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nO _ {i} ^ {2 D ^ {\\prime}} = \\left[ t _ {1} ^ {\\prime}, \\dots , t _ {K} ^ {\\prime} \\right], \\quad i \\in [ 1, N ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 467, + 903, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To infuse the enriched semantic 2D features into 3D geometric feature, we construct the 3D-to-2D-Multi-view pairs, i.e. each 3D instance feature in $O^{3D'} \\in \\mathbb{R}^{1 \\times N \\times D}$ corresponds to 2D features across $K$ views in $O^{2D'} \\in \\mathbb{R}^{K \\times N \\times D}$ . In particular, we utilize the preliminary $O^{3D'}$ as queries, while the 2D multi-view visual features $O^{2D'}$ as enriched reference keys and values. Inspired by [25], the injection process is conducted via a cross-attention layer, which encourages 3D instance queries to absorb the fine-grained semantics of keys and values. This results in the enhanced 3D instance features $O_f^{3D} \\in \\mathbb{R}^{1 \\times N \\times D}$ as below:", + "bbox": [ + 511, + 496, + 905, + 664 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nO _ {f} ^ {3 D} = \\operatorname {C r o s s A t t n} \\left(O ^ {3 D ^ {\\prime}}, O ^ {2 D ^ {\\prime}}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 672, + 903, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, we leverage a residual operation by adding the input $O^{3D'}$ to retain the basic characteristics in generating the instance level 3D visual tokens, $O_{I}^{3D} = O_{f}^{3D} + O^{3D'}$ .", + "bbox": [ + 511, + 702, + 903, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. 3D Instance Spatial Relation", + "text_level": 1, + "bbox": [ + 511, + 756, + 769, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by [5, 27, 60], we develop an effective 3D Instance Spatial Relation (3D-ISR) module to boost LLM's capabilities in assimilating spatial information within the 3D scene, as shown in Figure 4. 3D-ISR utilizes the instance-level 3D visual tokens $\\{O_{I1}^{3D}, O_{I2}^{3D}, \\ldots, O_{IN}^{3D}\\}$ derived from MCMF module along with the corresponding center coordinates $\\{C_1, C_2, \\ldots, C_N\\}$ of all instances as input. For the $i$ -th instance, we define its center coordinates", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14150", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1db6d1c3d65a4d75503fbbfd2fd99a7a85fcd5a99457318ba57bd6bfdd3874da.jpg", + "image_caption": [ + "Figure 4. Illustration of the 3D Instance Spatial Relation (3D-ISR) module in our framework." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as $C_i = (x_i, y_i, z_i)$ . For each pair of instance-level tokens $\\{O_{II}^{3D}, O_{Ij}^{3D}\\}$ , we calculate their Euclidean distance $d_{ij} = \\| C_i - C_j\\|_2$ as well as their horizontal angle $\\theta_h$ and vertical angle $\\theta_v$ . Specially, $\\theta_h = \\arctan 2((y_j - y_i) / (x_j - x_i))$ and $\\theta_v = \\arcsin ((z_j - z_i) / d_{ij})$ . By making use of these positional parameters, we generate the pairwise spatial features $S = \\{s_{ij}\\} \\in \\mathbb{R}^{N \\times N \\times 5}$ as:", + "bbox": [ + 89, + 311, + 483, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i j} = \\left[ \\sin (\\theta_ {h}), \\cos (\\theta_ {h}), \\sin (\\theta_ {v}), \\cos (\\theta_ {v}), d _ {i j} \\right]. (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 428, + 482, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by the language-conditioned self-attention presented in ViL3DRel [5], we suggest a spatial-conditioned self-attention module. Specially, we create position embeddings $\\mathcal{P}$ via the absolution positional encoding (PE), i.e. $\\mathcal{P} = \\mathrm{PE}(C_1,C_2,\\dots ,C_N)$ . A spatial conditioned attention weight $l_{i}$ is computed to select the relevant spatial relations for each instance $O_{II}^{3D}$ , which is formulated as below:", + "bbox": [ + 89, + 454, + 483, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {i} = W _ {P} ^ {\\top} \\left(\\mathcal {P} _ {i} + O _ {I i} ^ {3 D}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 568, + 482, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{P}\\in \\mathbb{R}^{D\\times 5}$ is a learnable parameter and the bias term is omitted for simplicity. The spatial-conditioned attention map $\\omega_{ij}$ is computed by combining $s_{ij},l_i$ and $l_j$ , i.e. $\\omega_{ij} = l_i\\cdot s_{ij}\\cdot l_j$", + "bbox": [ + 89, + 597, + 482, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through this positional transformation, the spatial-conditioned attention map $\\omega_{ij}$ encapsulates the pairwise spatial relationship across the entire 3D scene for each instance. We further integrate the attention map with instance-level visual tokens, which is formulated as below:", + "bbox": [ + 89, + 659, + 483, + 734 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} = \\sum_ {j = 1} ^ {N} \\omega_ {i j} O _ {I i} ^ {3 D}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 742, + 482, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the instance-level representations are concatenated and further processed by a transformer-based encoder $\\Gamma$ , followed by a max-pooling layer (Pool) and a simple two-layer MLP. The scene-level relation-aware tokens $F_{scene} \\in \\mathbb{R}^{D}$ are generated for the entire 3D scene as follows:", + "bbox": [ + 89, + 794, + 483, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {s c e n e} = \\operatorname {M L P} (\\operatorname {P o o l} (\\Gamma ([ F _ {1}, F _ {2}, \\dots , F _ {N} ]))). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 880, + 482, + 897 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, we obtain the final token representations for LLM processing, i.e. instance-level visual tokens $O_{Ii}^{3D}$ and spatial relation-aware scene-level tokens $F_{scene}$ , respectively.", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. End-to-End Multi-task Instruction Tuning", + "text_level": 1, + "bbox": [ + 511, + 147, + 875, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D-Language Prompts. Given an LMM, clear and explicit system messages and instructions are essential to support a range of downstream 3D vision-language tasks. For different tasks, we adopt various task-specific instruction templates with 3D features to generate uniform instruction data, enabling multi-task training. Please refer to the Supplementary Material for more detailed information.", + "bbox": [ + 511, + 171, + 906, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction Tuning. We conduct end-to-end multi-task fine-tuning to fully leverage the capabilities of our Inst3D-LMM framework. The MCMF and 3D-ISR modules enable the LLM with robust 3D scene understanding, grounding and reasoning abilities. We freeze the pre-trained 3D object encoder while updating both MCMF and 3D-ISR modules, as well as LLM. Upon completion of end-to-end multi-task instruction tuning, Inst3D-LMM can effectively handle various 3D vision-language tasks simultaneously, without fine-tuning on specific tasks, as illustrated in Figure 7.", + "bbox": [ + 511, + 277, + 908, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 443, + 648, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 511, + 468, + 714, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Benchmarks. In this work, we conducted our experiments on the ScanNetV2 dataset [11], an extensive indoor 3D scene dataset comprising 1,513 scenes. This dataset includes 3D point clouds, RGB-D frames, and detailed point-level instance segmentation annotations. The whole dataset is divided into 1,201 scenes for training and 312 scenes for validation, with all subsequent benchmarks adhering to these training/validation splits. Our evaluation encompasses a range of 3D scene understanding benchmarks, including ScanRefer [4] and Multi3DRefer [55] for single- and multi-object 3D Visual Grounding, respectively, ScanQA [1] for 3D Question Answering, and Scan2Cap [8] for 3D Dense Captioning. These datasets are converted into a uniform instruction format for multi-task instruction tuning and performance assessment.", + "bbox": [ + 511, + 491, + 908, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For 3D feature extraction, we leverage the 3D instance segmentation model Mask3D [38] pre-trained on ScanNet200 dataset [37], alongside the 3D object encoder Ulip2 [48]/Uni3D [57] based on ViT-L/14 [13]. For 2D feature extraction, we adopt ViT-H-based SAM [23] to obtain high-quality masks. Moreover, we extract 2D semantic features using the CLIP-ViT-L/14-336px vision encoder [35]. These pre-trained models are kept frozen. In our method, we apply multi-view selection and multi-level crops to 2D images, setting $K = 5$ views and $L = 3$ levels. We use Vicuna1.5-7B [10] as our basic LLM, which is fine-tuned from LLaMA2 [41]. Our fine-tuning", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14151", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2dfe96fa395f31888adaf3b5ea21555fbf42099d624bf9179277c8a362d44df9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodOverallUniqueMultipleMulti3DRefer
Acc@0.25↑Acc@0.50↑Acc@0.25↑Acc@0.50↑Acc@0.25↑Acc@0.50↑F1@0.25↑F1@0.50↑
Closed-set, full sup.
ScanRefer [4]37.324.365.043.330.619.8--
3DVG-Trans [56]45.934.577.258.538.428.730.225.5
M3DRef-CLIP [55]51.944.7-77.2-36.842.838.4
ConcreteNet [42]56.149.586.179.247.540.9--
CORE-3DVG [51]56.843.885.067.151.839.8--
Zero-shot
LLM-Grounder [50]17.15.330.822.616.312.9--
Visual-Programming [54]36.432.763.858.427.724.6--
Specialist
OpenScene [34]13.05.120.113.111.14.4--
3D-LLM(Flamingo) [16]21.2-------
3D-LLM(BLIP2-flant5) [16]30.3-------
Chat-3D v2 [18]35.930.461.257.625.222.645.141.6
ReGround3D [58]53.141.1------
Generalist
LAMM [52]-3.4------
3DMIT [26]10.77.2------
Grounded 3D-LLM [7]47.944.1----45.240.6
Chat-Scene [19]55.550.2----57.152.4
Inst3D-LMM57.851.688.681.548.743.258.353.5
", + "bbox": [ + 91, + 88, + 908, + 397 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/50c1ebbffe00d4cedb42b3af21240db958e390a8422efd1ae42943bf3c91a601.jpg", + "table_caption": [ + "Table 1. Quantitative results for 3D Visual Grounding on ScanRefer and Multi3DRefer validation sets. In the ScanRefer dataset, scenes are labeled as \"unique\" (one object per class) or \"multiple\" (more than one). Closed-set methods are fully supervised for specific datasets. \"Zero-shot\" refers to methods that directly use LLMs without fine-tuning. \"Specialist\" and \"Generalist\" categorize methods fine-tuned for specific tasks versus those trained jointly. Bold and underlined numbers indicate the best and the second-best results, respectively." + ], + "table_footnote": [], + "table_body": "
Method# 3D Data for AlignmentScanQAScan2Cap@0.50
B-1↑B-4↑METEOR↑ROUGE↑CIDER↑EM↑B-4↑METEOR↑ROUGE↑CIDER↑
Closed-set, full sup.
VoteNet [12] + MCAN [53]-28.06.211.429.854.717.3----
ScanRefer [4] + MCAN [53]-26.97.911.530.055.418.6----
ScanQA [8]-30.210.113.133.364.921.0----
Scan2Cap [1]-------22.421.443.535.2
3D-VisTA [60]-34.213.115.238.676.627.034.027.154.366.9
LLM-based Methods
LLaVA (zero-shot) [28]-7.10.310.512.35.70.21.58.319.63.2
LAMM [52]25K26.85.810.023.642.49.8----
3D-LLM(Flamingo) [16]675K30.37.212.232.359.220.45.911.429.9-
3D-LLM(BLIP2-flant5) [16]675K39.312.014.535.769.420.58.113.133.2-
Chat-3D v2 [18]38K38.47.316.140.177.121.131.822.350.263.9
LL3DA [6]38K-13.315.437.075.7-35.925.654.665.2
Grounded 3D-LLM [7]107K-13.4--72.7-35.5--70.6
Chat-Scene [19]38K-14.3--87.7-36.3--77.1
Inst3D-LMM38K43.514.918.442.688.624.638.327.557.279.7
", + "bbox": [ + 91, + 467, + 910, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative results for 3D Question Answering and 3D Dense Captioning on the ScanQA and Scan2Cap datasets.", + "bbox": [ + 133, + 714, + 861, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "process utilizes LoRA [17]. We adopt the AdamW optimizer with a weight decay of 0.02. All experiments were performed on 8 NVIDIA A100 GPUs.", + "bbox": [ + 89, + 744, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 801, + 230, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Visual Grounding. We first report visual grounding performance on the ScanRefer and Multi3DRefer validation datasets. As shown in Table 1, our approach outperforms the state-of-the-art model, Chat-Scene [19], by $+2.3\\%$ Acc@0.25 and $+1.2\\%$ F1@0.25 on ScanRefer and", + "bbox": [ + 89, + 824, + 483, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi3DRefer, respectively. Compared to specialist and closed-set methods, Inst3D-LMM, trained with a generalist approach, achieves competitive performance. Figure 5 displays typical visual comparison results.", + "bbox": [ + 511, + 744, + 906, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Question Answering. We then compare Inst3D-LMM with previous leading methods on the ScanQA validation set. Table 2 reports the results. Our Inst3D-LMM consistently outperforms these methods, including recent LL3DA [6], Grounded 3D-LLM [7] and Chat-Scene [19].", + "bbox": [ + 511, + 805, + 908, + 880 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Dense Captioning. This task involves the localization", + "bbox": [ + 511, + 881, + 903, + 896 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14152", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e79d7be4241944c89c7fd56282d186f51e719e79cb1117808871a239f184654c.jpg", + "image_caption": [ + "7-6" + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 285, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2ffa5e1d1ccf3d270cc18ec24aa88e6cefb988e6d4cbb05a12f18bf4bb49592.jpg", + "image_caption": [ + "Figure 6. A visualization of the similarity score between text query and segmented 3D proposals. We compare sentence tokens used in ViL3DRel [5] with position embeddings employed in 3D-ISR." + ], + "image_footnote": [], + "bbox": [ + 305, + 90, + 473, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da796e3ac14e3d8d8653a051d01f783d1382f17972b86e7869d1386965f8c17f.jpg", + "image_caption": [ + "Figure 5. Visual comparisons in 3D Visual Grounding. The rendered images show the ground-truth (blue), incorrectly identified objects (red), and correctly identified objects (green). The colored text indicates the results of text decoupling." + ], + "image_footnote": [ + "Text Query: The kitchen cabinets are under the sink. They are the furthest left cabinets and to the left of the dishwasher." + ], + "bbox": [ + 96, + 172, + 285, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ac875c81967689bd998ab13e86d61467bee8f80b724dda833c9354f04e9a5a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 172, + 473, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and description of instances. As shown in Table 2, our approach achieves $38.3\\%$ B-4@0.50 and $79.7\\%$ C@0.50 on Scan2Cap, which exceeds the closed-set expert model 3D-VisTA [60] by $+4.3\\%$ and $+12.8\\%$ , and still outperforms the LLM-based method LL3DA by $+2.4\\%$ and $+14.5\\%$ .", + "bbox": [ + 89, + 359, + 483, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 445, + 243, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform ablation experiments to thoroughly evaluate the effectiveness of the components. Please refer to the Supplementary Material for more analysis.", + "bbox": [ + 89, + 469, + 482, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of Multi-view Cross-Modal Fusion (MCMF). The MCMF module aims to augment LLM's understanding of 3D instances' geometric and semantic attributes. We explore several methods to combine these features, including direct concatenation for projection ('Concat.'), parallel projection followed by combination ('Parallel'), and vanilla cross-attention ('Cross-Attention'). To encode basic spatial information, we concatenate the results of absolution positional encoding (PE) with 3D features. As shown in Table 3, MCMF outperforms all these methods on ScanRefer, ScanQA and Scan2Cap datasets. We also observe that the performance of 3D LMM is significantly improved by incorporating multi-view 2D CLIP information compared to using only 3D geometric features. These results demonstrate the effectiveness of our MCMF approach.", + "bbox": [ + 89, + 513, + 483, + 741 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a1cb7ef30dd34424a04f2d957ec4aef6efd0425e6f05ba9710e8e2feb78a833c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodScanReferScanQAScan2Cap@0.50
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑B-4↑CIDER↑
w/o Multi-view 2D36.031.931.557.319.954.3
Concat.38.834.436.665.425.263.3
Parallel37.633.535.363.924.862.1
Cross-Attention39.236.537.466.125.564.3
MCMF46.741.941.578.632.768.2
", + "bbox": [ + 91, + 746, + 480, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impacts of 3D Instance Spatial Relation (3D-ISR). We further investigate the 3D-ISR module to analyze its impact", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84619ebb3073cac7a658d3902b7cec84c0c8913562255e76a605b7354176e437.jpg", + "image_caption": [ + "Input 3D Scene" + ], + "image_footnote": [ + "Text Query: There is a brown reclined in the corner of the room. The adjacent walls have two green boards. It is between the intersection of the 2 boards." + ], + "bbox": [ + 517, + 92, + 635, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82cda2238fe073390724c2790d2b6d6ad084d0bee731b4dca4601284be186bb9.jpg", + "image_caption": [ + "Sentence Token" + ], + "image_footnote": [], + "bbox": [ + 637, + 90, + 756, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5126e0323f8535c82d39a60f31c68c2da22c225d69108025743ed435d1d262fc.jpg", + "image_caption": [ + "Position Embedding", + "A" + ], + "image_footnote": [], + "bbox": [ + 759, + 90, + 874, + 148 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on 3D spatial understanding using only 3D geometry priors. We compare 3D-ISR with previous spatial relation modeling methods, such as the 3D localization mechanism in ViL3DRel [5] and 3D-LLM [16], and relation-aware token generation in Chat-3D-v2 [18]. Table 4 reports the results of the comparison. It can be observed that our 3D-ISR consistently surpasses these approaches. Specifically, 3D-ISR significantly enhances performance on grounding tasks (ScanRefer and Multi3dRefer). This aligns with the motivation of our design. We also provide a visualization, as shown in Figure 6, of the similarity between the text query and segmented 3D proposals after processing by LLM. Compared to ViL3DRel [5] using sentence tokens, our 3D-ISR more accurately captures spatial relationships within the scene employing position embeddings of objects.", + "bbox": [ + 511, + 237, + 906, + 465 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/85839773fcded001d376c0c5b75049f8da028908e53ab5fda745b6b1c0244677.jpg", + "table_caption": [ + "Table 3. Ablation evaluations of the proposed Multi-view CrossModal Fusion (MCMF) module." + ], + "table_footnote": [], + "table_body": "
MethodScanReferMulti3DReferScanQA
Acc@0.25↑Acc@0.50↑F1@0.25↑F1@0.50↑B-1↑CIDER↑
w/o spatial relation35.629.834.827.329.255.6
ViL3DRel [5]36.130.538.435.532.858.9
3D-LLM [16]39.236.842.435.232.560.2
Chat-3D-v2 [18]40.837.541.637.835.164.6
3D-ISR48.344.146.241.339.172.3
", + "bbox": [ + 514, + 467, + 903, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study to verify the effectiveness of our 3D Instance Spatial Relation (3D-ISR) module.", + "bbox": [ + 511, + 550, + 903, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablations within MCMF and 3D-ISR modules. We also explore the effectiveness of key designs in MCMF and 3D-ISR modules. The learnable [CLS] token in the MCMF module is introduced to aggregate 2D multi-view features. We compare it with alternative methods, such as token max pooling and Q-Former [24]. Table 5 reports the comparison results, which demonstrate the efficacy of the learnable [CLS] token design. Besides, to assess the impact of different pairwise spatial features in the 3D-ISR module, we evaluate the model using only distance or orientation information to compute spatial features in Eq 4. The results in Table 6 reveal that distance has a greater impact on the grounding capability of the model, while orientation is more crucial in handling Q&A tasks. Combining both pairwise distance and orientation yields the best overall performance. Mutual Benefits of MCMF and 3D-ISR. As shown in Table 7, the model that integrates MCMF and 3D-ISR consistently outperforms those that use either module alone for multiple tasks. To further verify its effectiveness, we combine MCMF with absolution positional encoding (PE) for the basic spatial relation, and directly adopt feature con", + "bbox": [ + 511, + 583, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14153", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b61b337457c3526896f0ddfb41b73ea6b1b582b6ad061616390b2bd3d6b24cf4.jpg", + "image_caption": [ + "Figure 7. Qualitative illustration of Inst3D-LMM across various 3D-language tasks in diverse 3D environments." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 893, + 372 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2dfb0972536ec0f7a91aac782121feea0bf9a9c46a5fd021f08cceb663e6f23c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodScanReferScanQAScan2Cap@0.50
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑B-4↑CIDER↑
Max Pooling48.744.639.274.830.665.5
Q-Former [24]53.247.944.387.935.478.6
CLS Token57.851.643.588.638.379.7
", + "bbox": [ + 91, + 404, + 480, + 473 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/512c717cee5aaca1da4468520964b9ca338277c02721b44c50847d54bfe09623.jpg", + "table_caption": [ + "Table 5. Ablation study of learnable CLS token in MCMF module." + ], + "table_footnote": [], + "table_body": "
MethodScanReferMulti3DReferScanQA
Acc@0.25↑Acc@0.50↑F1@0.25↑F1@0.50↑B-1↑CIDER↑
w/ Dist (only)43.238.042.437.831.265.7
w/ Ort (only)39.733.640.535.932.566.8
Dist + Ort48.344.146.241.339.172.3
", + "bbox": [ + 91, + 494, + 480, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study of 3D-ISR module. Dist (Ort) means only using distance (orientation) to compute pairwise spatial features.", + "bbox": [ + 89, + 566, + 482, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "catenation (Concat.) to integrate 3D object features and 2D multi-view features along with 3D-ISR. These results indicate the synergy efforts between MCMF and 3D-ISR.", + "bbox": [ + 89, + 602, + 482, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different LLMs and Foundation Models. Table 8 presents the results of various LLMs and pre-trained 2D/3D models, including Vicuna-7B [10] vs Vicuna-13B, Ulip2 [48] vs Uni3D [57], CLIP [35] vs SigLIP [32], and ViT-H-based SAM [23] vs ViT-L-based SAM. We find that the performance of grounding and reasoning increases along with the total number of parameters in the foundation models (i.e. ViT-L vs Vit-H and 7B vs 13B). These results indicate that our framework's capability can be improved in tandem with the performance of foundation models.", + "bbox": [ + 89, + 648, + 482, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 814, + 209, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed Inst3D-LMM, an effective instance-aware framework to leverage the potential of Large Multi-modal Models (LMMs) for 3D scene understanding. To improve instance-level representations, we developed", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9819281c250251bf3898edbe6de27b7477e70dbf232342931a12eaf5a20407ee.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodScanReferScanQAScan2Cap@0.50
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑B-4↑CIDER↑
w/ MCMF (only)38.234.639.476.830.566.9
w/ 3D-ISR (only)48.344.139.172.329.864.3
MCMF + PE46.741.941.578.632.768.2
3D-ISR + Concat.49.845.040.675.433.366.5
MCMF + 3D-ISR (full)57.851.643.588.638.379.7
", + "bbox": [ + 514, + 404, + 898, + 484 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/19a79e0b7c13050671d5315d0860f9851121ef24265d880f79473056cb462d50.jpg", + "table_caption": [ + "Table 7. Comparison results of the collaboration between MCMF and 3D-ISR modules." + ], + "table_footnote": [], + "table_body": "
LLM3D Encoder2D VFsScanReferScanQA
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑
Vicuna-7BUlip2CLIP+SAM (ViT-L)49.244.839.579.6
Vicuna-7BUlip2CLIP+SAM (ViT-H)53.646.543.787.5
Vicuna-7BUlip2SigLIP+SAM (ViT-H)54.347.042.284.5
Vicuna-7BUni3DCLIP+SAM (ViT-H)57.851.643.588.6
Vicuna-13BUni3DCLIP+SAM (ViT-H)56.052.042.883.1
Vicuna-13BUni3DSigLIP+SAM (ViT-H)55.247.843.385.4
", + "bbox": [ + 516, + 522, + 898, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Ablation study of different LLMs and pre-trained foundation models.", + "bbox": [ + 511, + 612, + 905, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a novel Multi-view Cross-Modal Fusion (MCMF) module, which injects multi-view 2D semantic open-vocabulary priors into 3D geometry features to generate fine-grained instance-level tokens. Furthermore, we introduced a 3D Instance Spatial Relation (3D-ISR) module that employs the spatial condition attention mechanism to capture pairwise spatial relations. Experimental results demonstrate that our approach achieves promising performance in understanding and reasoning across various 3D vision-language tasks.", + "bbox": [ + 511, + 656, + 906, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 513, + 813, + 671, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China under Grants (62376244). It is also supported by Information Technology Center and State Key Lab of CAD&CG, Zhejiang University.", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14154", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, pages 19129-19139, 2022. 5, 6", + "[2] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In CVPR, pages 16464-16473, 2022. 2", + "[3] Yang Cao, Zeng Yihan, Hang Xu, and Dan Xu. Coda: Collaborative novel box discovery and cross-modal alignment for open-vocabulary 3d object detection. In NeurIPS, pages 71862-71873, 2024. 4", + "[4] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221. Springer, 2020. 5, 6", + "[5] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. In NeurIPS, pages 20522-20535, 2022. 2, 4, 5, 7", + "[6] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In CVPR, pages 26428-26438, 2024. 1, 2, 3, 6", + "[7] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 3, 6", + "[8] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In CVPR, pages 3193-3203, 2021. 2, 5, 6", + "[9] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In ICCV, pages 18109-18119, 2023. 2", + "[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 5, 8", + "[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 2432–2443, 2017. 5", + "[12] Zhipeng Ding, Xu Han, and Marc Niethammer. Volenet: A deep learning label fusion method for multi-atlas segmentation. In MICCAI, pages 202-210, 2019. 6", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5", + "[14] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm:" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2", + "[15] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905, 2023. 2", + "[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. In NeurIPS, pages 20482-20494, 2023. 1, 2, 3, 6, 7", + "[17] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6", + "[18] Haifeng Huang, Zehan Wang, Rongjie Huang, Luping Liu, Xize Cheng, Yang Zhao, Tao Jin, and Zhou Zhao. Chat-3d v2: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023. 3, 6, 7", + "[19] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. NeuIPS, 2024. 1, 2, 6", + "[20] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 2", + "[21] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. In ECCV, pages 528-545. Springer, 2022. 2", + "[22] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In CVPR, pages 10984-10994, 2023. 2", + "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3, 4, 5, 8", + "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 7, 8", + "[25] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 4", + "[26] Zeju Li, Chao Zhang, Xiaoyan Wang, Ruilong Ren, Yifan Xu, Ruifei Ma, and Xiangde Liu. 3dmit: 3d multi-modal instruction tuning for scene understanding. In ICMEW, 2024. 6", + "[27] Cai Liang, Bo Li, Zhengming Zhou, Longlong Wang, Pengfei He, Liang Hu, and Haoxing Wang. Spatioawaregrouding3d: A spatio aware model for improving 3d vision grousing. In CVPRW, 2024. 4" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14155", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, pages 34892-34916, 2023. 6", + "[29] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In NeurIPS, pages 44860-44879, 2023. 2", + "[30] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. In ICLR, 2023. 1, 2", + "[31] Yunze Man, Shuhong Zheng, Zhipeng Bao, Martial Hebert, Liang-Yan Gui, and Yu-Xiong Wang. Lexicon3d: Probing visual foundation models for complex 3d scene understanding. arXiv preprint arXiv:2409.03757, 2024. 2", + "[32] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, pages 529-544. Springer, 2022. 8", + "[33] Maria Parelli, Alexandros Delitzas, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Clip-guided vision-language pre-training for question answering in 3d scenes. In CVPR, pages 5606-5611, 2023. 2", + "[34] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In CVPR, pages 815-824, 2023. 6", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 3, 4, 5, 8", + "[36] Krishan Rana, Jesse Haviland, Sourav Garg, Jad Abou-Chakra, Ian Reid, and Niko Suenderhauf. Sayplan: Grounding large language models using 3d scene graphs for scalable robot task planning. In CoRL, pages 23-72, 2023. 1", + "[37] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, pages 125-141. Springer, 2022. 5", + "[38] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In ICRA, pages 8216-8223, 2023. 3, 4, 5", + "[39] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M. Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In ICCV, pages 2998-3009, 2023. 1", + "[40] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In NeurIPS, pages 68367-68390, 2023. 4", + "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5", + "[42] Ozan Unal, Christos Sakaridis, Suman Saha, Fisher Yu, and Luc Van Gool. Three ways to improve verbo-visual" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fusion for dense 3d visual grounding. arXiv preprint arXiv:2309.04561, 2023. 6", + "[43] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, et al. Embodiedscan: A holistic multimodal 3d perception suite towards embodied ai. In CVPR, pages 19757-19767, 2024. 1, 3", + "[44] Zehan Wang, Haifeng Huang, Yang Zhao, Linjun Li, Xize Cheng, Yichen Zhu, Aoxiong Yin, and Zhou Zhao. 3drp-net: 3d relative position-aware network for 3d visual grounding. In EMNLP, page 10612-10625, 2023. 2", + "[45] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 1, 3", + "[46] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In CVPR, pages 19231-19242, 2023. 1", + "[47] Runsen Xu, Zhiwei Huang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Vlm-grounder: A vlm agent for zero-shot 3d visual grounding. arXiv preprint arXiv:2410.13860, 2024. 2", + "[48] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In CVPR, pages 27091–27101, 2024. 5, 8", + "[49] Fan Yang, Sicheng Zhao, Yanhao Zhang, Haoxiang Chen, Hui Chen, Wenbo Tang, Haonan Lu, Pengfei Xu, Zhenyu Yang, Jungong Han, et al. Lmi3d: Empowering llm with 3d perception from a single 2d image. arXiv preprint arXiv:2408.07422, 2024. 2", + "[50] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. In ICRA, pages 7694-7701, 2024. 6", + "[51] Li Yang, Ziqi Zhang, Zhongang Qi, Yan Xu, Wei Liu, Ying Shan, Bing Li, Weiping Yang, Peng Li, Yan Wang, et al. Exploiting contextual objects and relations for 3d visual grounding. In NeurIPS, pages 49542-49554, 2024. 6", + "[52] Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Xiaoshui Huang, Zhiyong Wang, Lu Sheng, Lei Bai, et al. Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. In NeurIPS, pages 26650-26685, 2023. 6", + "[53] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In CVPR, pages 6281–6290, 2019. 6", + "[54] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3d visual grounding. In CVPR, pages 20623-20633, 2024. 6", + "[55] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, pages 15225-15236, 2023. 5, 6" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14156", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2928-2937, 2021. 6", + "[57] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. In ICLR, 2024. 2, 3, 4, 5, 8", + "[58] Chenming Zhu, Tai Wang, Wenwei Zhang, Kai Chen, and Xihui Liu. Empowering 3d visual grounding with reasoning capabilities. arXiv preprint arXiv:2407.01525, 2024. 1, 3, 6", + "[59] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmms with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 2", + "[60] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 2, 4, 6, 7" + ], + "bbox": [ + 91, + 92, + 480, + 330 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14157", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/ada2a42d-606c-400d-85d5-a68b46a5f62b_model.json b/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/ada2a42d-606c-400d-85d5-a68b46a5f62b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8789b2b1c032be2b248848ac2ce224bb6a85b771 --- /dev/null +++ b/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/ada2a42d-606c-400d-85d5-a68b46a5f62b_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.131, + 0.783, + 0.177 + ], + "angle": 0, + "content": "Inst3D-LMM: Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.203, + 0.851, + 0.263 + ], + "angle": 0, + "content": "Hanxun Yu\\(^{1*}\\), Wentong Li\\(^{2*}\\), Song Wang\\(^{1}\\), Junbo Chen\\(^{3\\dagger}\\), Jianke Zhu\\(^{1\\dagger}\\) \n\\(^{1}\\)Zhejiang University \\(^{2}\\)Nanjing University of Aeronautics and Astronautics \\(^{3}\\)Udeer.ai {hanxun.yu, songw, jkzhu}@zju.edu.cn, wentong.li@nuaa.edu.cn, junbo@udeer.ai" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.295, + 0.33, + 0.312 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.486, + 0.721 + ], + "angle": 0, + "content": "Despite encouraging progress in 3D scene understanding, it remains challenging to develop an effective Large Multi-modal Model (LMM) that is capable of understanding and reasoning in complex 3D environments. Most previous methods typically encode 3D point and 2D image features separately, neglecting interactions between 2D semantics and 3D object properties, as well as the spatial relationships within the 3D environment. This limitation not only hinders comprehensive representations of 3D scene, but also compromises training and inference efficiency. To address these challenges, we propose a unified Instance-aware 3D Large Multi-modal Model (Inst3D-LMM) to deal with multiple 3D scene understanding tasks simultaneously. To obtain the fine-grained instance-level visual tokens, we first introduce a novel Multi-view Cross-Modal Fusion (MCMF) module to inject the multi-view 2D semantics into their corresponding 3D geometric features. For scene-level relation-aware tokens, we further present a 3D Instance Spatial Relation (3D-ISR) module to capture the intricate pairwise spatial relationships among objects. Additionally, we perform end-to-end multi-task instruction tuning simultaneously without the subsequent task-specific fine-tuning. Extensive experiments demonstrate that our approach outperforms the state-of-the-art methods across 3D scene understanding, reasoning and grounding tasks. Source code is available at: https://github.com/hanxunyu/Inst3D-LMM." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.764, + 0.224, + 0.78 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.485, + 0.868 + ], + "angle": 0, + "content": "Building Large Multi-modal Models (LMMs) for 3D scene understanding becomes an emerging research topic with significant potential for advancing autonomous robotics [36]. For example, interactive embodied agents [39] are expected to interpret 3D layouts and predict" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.297, + 0.816, + 0.311 + ], + "angle": 0, + "content": "object locations based on human instructions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.908, + 0.675 + ], + "angle": 0, + "content": "Traditional 3D scene understanding methods [30, 46] are typically tailored for individual downstream tasks, such as 3D Visual Grounding (3D-VG), 3D Question Answering (3D-QA) and 3D Dense Captioning (3D-DC). In contrast, LMMs are able to handle various 3D perception tasks within a single model. Some methods [16, 43] primarily focus on translating 3D points into the space of 2D Vision Language Models (VLMs) or directly leveraging multiview 2D features as 3D representations. Alternatively, other approaches [6, 45, 58] directly encode the features of 3D points and facilitate alignment with LLM using 3D-text instruction data. However, they often require multi-stage alignment or language-scene pre-training, complicating the development of a versatile model capable of handling multiple tasks. To enable a unified 3D LMM framework, recent work [19] decomposes the input 3D scene into a set of individual object proposals, each identified by unique tokens to explicitly capture 3D object features at the instance level. While this approach exhibits promising results, it neglects the interactions between the 2D semantic features and the properties of 3D objects, as well as the spatial relationship modeling among objects in 3D environments. This oversight further results in substantial token costs for the LLM, thus hindering both training and inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose Inst3D-LMM, an effective Instance-aware 3D Large Multi-modal Model that tackles multiple 3D-language tasks without resorting to task-specific fine-tuning. Our approach fully leverages the powerful 2D Vision Foundation Models (VFMs) and 3D specialist models to extract enriched 2D and 3D features at the instance level, respectively. As shown in Figure 1, in contrast to previous methods, our approach is able to generate fine-grained instance-level representations that encapsulate both geometric and semantic properties, and scene-level representations that capture intricate pairwise spatial relationships among objects in a 3D scene. Moreover, our method results in minor token costs for LLM, thereby enhancing both training and inference efficiency. By leveraging this instance-aware methodology, our Inst3D-LMM" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.244, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. † Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14147" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.619, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.09, + 0.905, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.286, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Figure 1. Comparisons between previous 3D LMMs and our proposed Inst3D-LMM. (a) Previous methods [6, 16, 19, 20] typically encode the features of 3D points or 2D images separately and concatenate them directly, often requiring task-specific fine-tuning for different tasks. (b) Illustration of Inst3D-LMM. Our method integrates 2D/3D cross-modal information and captures the intricate spatial relations among objects within 3D environments to generate tight but informative instance/scene-level tokens for the LMM. (c) Compared with other 3D LMMs, our Inst3D-LMM requires fewer computational resources, while offering faster training and inference speeds." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.371, + 0.483, + 0.401 + ], + "angle": 0, + "content": "significantly improves the LLM's ability to comprehend 3D scenes with respect to both efficiency and accuracy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.402, + 0.483, + 0.612 + ], + "angle": 0, + "content": "Specifically, we introduce a novel Multi-view CrossModal Fusion (MCMF) module that effectively infuses enriched multi-view 2D features into the original 3D features with coarse semantics. A learnable [CLS] token is introduced to aggregate the characteristics of each 2D view, enabling efficient multi-view 2D-to-3D cross-modal transformation. To capture intricate pairwise spatial relationships among objects in a 3D scene, we then propose a 3D Instance Spatial Relation (3D-ISR) module. A spatial condition self-attention between manifold position embeddings and instance-level tokens is presented to produce relation-aware scene-level representations. The resulting instance-level and scene-level representations are subsequently fed into the LLM for end-to-end multi-task instruction tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Extensive experiments across various tasks, including 3D-VG, 3D-QA and 3D-DC, demonstrate that our approach outperforms previous state-of-the-art methods with leading 3D scene understanding, grounding and reasoning capabilities. Unlike most existing methods that focus on close-set scene understanding or require per-task fine-tuning, our Inst3D-LMM operates as a generalist model. We believe this work lays a fundamental step towards unifying diverse 3D vision-language tasks in generative language modeling. To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- We propose a unified and efficient instance-aware LLM-based framework, called Inst3D-LMM for various 3D scene understanding tasks with end-to-end multi-modal instruction tuning. Serving as a generalist model, our approach demonstrates superior performance across 3D scene understanding, reasoning and spatial localization." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- We utilize 2D VFMs to extract multi-view contextual features for each 3D instance and then devise a Multi-view Cross-Modal Fusion (MCMF) module to effectively en" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.371, + 0.905, + 0.401 + ], + "angle": 0, + "content": "hance instance-level feature representations by jointly integrating 3D geometry and 2D semantic priors." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.402, + 0.905, + 0.447 + ], + "angle": 0, + "content": "- A 3D Instance Spatial Relation (3D-ISR) module is introduced to boost the capability of LMM in understanding the complex spatial details within 3D scenes." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.465, + 0.663, + 0.48 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.492, + 0.907, + 0.779 + ], + "angle": 0, + "content": "3D Scene Understanding with Language. In 3D scene understanding, there is a surge of interest in making use of language queries to capture user intentions for various downstream tasks, such as 3D Visual Grounding [5, 44], 3D Question Answering [30, 33] and 3D Dense Captioning [8, 21]. Specifically, 3D Visual Grounding entails localizing target objects based on language queries. Moreover, 3D Question Answering demands robust 3D spatial perception and reasoning. 3D Dense Captioning involves localizing and describing objects in 3D scenes. Conventional methods typically focus on a specific task. Instead, 3D visual grounding and dense captioning tasks are combined by leveraging their complementary aspects [2, 9, 31, 47]. Recent efforts like 3D-VLP [22] and 3D-VisTA [60] attempt to establish a universal framework by pre-aligning 3D scenes with their corresponding textual descriptions. In contrast to our Inst3D-LMM, most existing methods still focus on close-set scene understanding, which requires either task-specific fine-tuning or striving to build specialized models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "3D Large Multi-modal Models. Inspired by the significant advancements in Large Language Models (LLMs), researchers extend LLM's knowledge to encompass 3D modality [29, 49, 57, 59]. Point-LLM [14] and Imagebind-LLM [15] have succeeded in bridging the gap between 3D visuals and text by utilizing extensive 3D object datasets. However, these models struggle with interpreting complex spatial relationships in 3D scenes. Another promising di" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14148" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.088, + 0.904, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.48, + 0.907, + 0.551 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed Inst3D-LMM. Our pipeline takes as input point clouds of a 3D indoor scene, along with RGB-D images. We first employ the pre-trained 3D specialist models and 2D VFMs to extract 3D proposals and multi-view 2D semantic features, respectively. We then suggest the MCMF module to generate fine-grained instance-level tokens. A 3D-ISR module is further introduced to create relation-aware scene-level tokens based on spatial distances. By leveraging the constructed 3D-language prompts, we conduct multi-task instruction tuning to simultaneously handle various 3D tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.566, + 0.486, + 0.825 + ], + "angle": 0, + "content": "rection focuses on developing the scene-level 3D LMMs. Hong et al. [16] encode projected 3D features using a 2D vision encoder and incorporate location tokens to augment LLM vocabulary. Other methods [6, 45, 58] directly encode point clouds and utilize 3D scene-text data for better visual interaction through pre-alignment. Huang et al. [18] employ a three-stage training scheme and adopt object identifiers to learn individual object attributes. Chen et al. [7] use special referent tokens for precise referencing and grounding. Wang et al. [43] encode point clouds and RGB images separately. However, some crucial semantic pixels are often lost during their sparse fusion processes, resulting in a coarse visual semantic representation. In this work, we propose an effective instance-aware framework to fuse fine-grained cross-modal information and encode spatial relations, which achieves promising results on multiple 3D-language tasks as a generalist model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.843, + 0.226, + 0.86 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Our goal is to enable LLM to understand the 3D environment and perform various visual interaction tasks based on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.907, + 0.839 + ], + "angle": 0, + "content": "human instructions. Figure 2 illustrates the architecture of our framework. In this section, we first introduce how to extract features at the instance level using pre-trained 3D models [38, 57] and 2D VFs [23, 35], respectively. Secondly, we present a novel Multi-view Cross-Modal Fusion (MCMF) module to obtain fine-grained instance-level tokens, which are specially designed to effectively integrate 3D geometric features with their corresponding multi-view 2D semantic features. Thirdly, we introduce the 3D Instance Spatial Relation (3D-ISR) module to enhance LLM's ability to capture spatial information at the scene level, which generates relation-aware tokens through attention-based analysis of spatial relationships among different 3D proposals. Under our proposed framework, the MCMF and 3D-ISR modules are jointly optimized, allowing mutual enhancement. Finally, we conduct end-to-end multi-task instruction tuning to address a range of 3D scene understanding tasks simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.813, + 0.862 + ], + "angle": 0, + "content": "3.1. Instance-Level Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "3D Feature Extraction. We first segment the 3D point clouds of each individual instance in a class-agnostic man" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14149" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "ner by leveraging a pre-trained 3D instance segmentation model [38]. We only retain the predicted binary 3D instance masks while ignoring their closed-vocabulary class labels. All instance proposals in one scene are represented by \\( O^{3D} = (O_1^{3D}, \\ldots, O_N^{3D}) \\), where \\( O_i^{3D} = [coordinate, color] \\) consists of the attributes of each instance. We obtain all instance proposals of a scene, and the pre-trained 3D encoder \\( \\mathbf{E}_{\\mathbf{o}} \\) [57] is used to extract their instance-level features, i.e., \\( \\mathbf{f}_{\\mathbf{o}}^{3D} = \\mathbf{E}_{\\mathbf{o}}(O^{3D}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.483, + 0.499 + ], + "angle": 0, + "content": "2D Feature Extraction. Due to the inherent sparsity of 3D point cloud data, previous methods have difficulties in generating discriminative features of each object. In this work, we employ powerful 2D VFMs to extract 2D semantic features for each 3D instance. As in [3, 40], we firstly project the point cloud of each instance \\(O_{i}^{3D}\\) onto the image plane and then select the top \\(K\\) views according to the number of visible points. To obtain accurate 2D masks, we randomly sample \\(k_{\\mathrm{sample}} = 5\\) points as input prompts for SAM [23] to deal with noisy bounding boxes with outliers. Hereby, we select the high-quality mask with the highest confidence score. To enrich features with contextual information, these masks are employed to generate multi-level \\((L)\\) crops of the selected images that are further fed into the pre-trained CLIP vision encoder [35] to extract features with language-aligned embedding space. Finally, we aggregate multi-level features of sub-images from the same frame to form the 2D multi-view features \\(O^{2D}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.51, + 0.371, + 0.525 + ], + "angle": 0, + "content": "3.2. Multi-view Cross-Modal Fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.718 + ], + "angle": 0, + "content": "To better fuse 3D geometry priors with 2D multi-view semantic priors, we introduce an effective Multi-view CrossModal Fusion (MCMF) module that generates the enriched token representations for each 3D instance before being fed into LLM. The architecture of MCMF is crafted with a coarse-to-fine framework, as shown in Figure 3. To map 3D object features \\( O^{3D} \\in \\mathbb{R}^{1 \\times N \\times D^{3d}} \\) and 2D CLIP features \\( O^{2D} \\in \\mathbb{R}^{K \\times N \\times D^{2d}} \\) into the embedding space of LLM with the dimensionality of \\( D \\), we utilize a simple two-layer MLP with a LayerNorm and GELU in between, yielding \\( O^{3D'} \\) and \\( O^{2D'} \\), respectively. \\( N \\) represents the number of instances, and \\( K \\) is the number of images for each object." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.484, + 0.871 + ], + "angle": 0, + "content": "Subsequently, we introduce a Cross-Modal Injection Block to transform the enriched semantic priors from 2D multi-view representations into 3D instance features. Moreover, we adopt a straightforward self-attention layer to further enhance such 3D instance features. For 2D multi-view features, we append a learnable [CLS] token \\( t_k \\) to the flattened CLIP feature maps in order to adaptively encapsulate the global semantic representation of the \\( k \\)-th view. Then, we apply a self-attention layer to multi-view features, enabling semantic gathering to derive \\( t_k' \\) for \\( k \\)-th view as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.883, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\nt _ {k} ^ {\\prime} = \\phi (\\text {S e l f A t t n} ([ t _ {k}, O _ {\\text {v i e w} _ {k}} ^ {2 D ^ {\\prime}} ])), \\tag {1}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.09, + 0.905, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.317, + 0.905, + 0.345 + ], + "angle": 0, + "content": "Figure 3. Architecture of the proposed Multi-view Cross-Modal Fusion (MCMF) module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.905, + 0.397 + ], + "angle": 0, + "content": "where \\(1 \\leq k \\leq K\\), \\(O_{\\mathrm{view}_k}^{2D'} \\in \\mathbb{R}^{N \\times D}\\). \\(\\phi\\) is an indicator function that outputs the updated [CLS] token as the first one from the token list." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.397, + 0.905, + 0.459 + ], + "angle": 0, + "content": "We obtain the processed 3D token embedding \\(O^{3D'} \\in \\mathbb{R}^{1 \\times N \\times D}\\) containing coarse semantics, where each instance in \\(O^{3D'}\\) corresponds to multiple views of the processed 2D CLIP features \\(O^{2D'}\\). It can be represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.468, + 0.905, + 0.488 + ], + "angle": 0, + "content": "\\[\nO _ {i} ^ {2 D ^ {\\prime}} = \\left[ t _ {1} ^ {\\prime}, \\dots , t _ {K} ^ {\\prime} \\right], \\quad i \\in [ 1, N ]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.665 + ], + "angle": 0, + "content": "To infuse the enriched semantic 2D features into 3D geometric feature, we construct the 3D-to-2D-Multi-view pairs, i.e. each 3D instance feature in \\(O^{3D'} \\in \\mathbb{R}^{1 \\times N \\times D}\\) corresponds to 2D features across \\(K\\) views in \\(O^{2D'} \\in \\mathbb{R}^{K \\times N \\times D}\\). In particular, we utilize the preliminary \\(O^{3D'}\\) as queries, while the 2D multi-view visual features \\(O^{2D'}\\) as enriched reference keys and values. Inspired by [25], the injection process is conducted via a cross-attention layer, which encourages 3D instance queries to absorb the fine-grained semantics of keys and values. This results in the enhanced 3D instance features \\(O_f^{3D} \\in \\mathbb{R}^{1 \\times N \\times D}\\) as below:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.674, + 0.905, + 0.695 + ], + "angle": 0, + "content": "\\[\nO _ {f} ^ {3 D} = \\operatorname {C r o s s A t t n} \\left(O ^ {3 D ^ {\\prime}}, O ^ {2 D ^ {\\prime}}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.905, + 0.751 + ], + "angle": 0, + "content": "Furthermore, we leverage a residual operation by adding the input \\( O^{3D'} \\) to retain the basic characteristics in generating the instance level 3D visual tokens, \\( O_{I}^{3D} = O_{f}^{3D} + O^{3D'} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.771, + 0.773 + ], + "angle": 0, + "content": "3.3. 3D Instance Spatial Relation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Motivated by [5, 27, 60], we develop an effective 3D Instance Spatial Relation (3D-ISR) module to boost LLM's capabilities in assimilating spatial information within the 3D scene, as shown in Figure 4. 3D-ISR utilizes the instance-level 3D visual tokens \\(\\{O_{I1}^{3D}, O_{I2}^{3D}, \\ldots, O_{IN}^{3D}\\}\\) derived from MCMF module along with the corresponding center coordinates \\(\\{C_1, C_2, \\ldots, C_N\\}\\) of all instances as input. For the \\(i\\)-th instance, we define its center coordinates" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14150" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.273, + 0.483, + 0.302 + ], + "angle": 0, + "content": "Figure 4. Illustration of the 3D Instance Spatial Relation (3D-ISR) module in our framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.484, + 0.419 + ], + "angle": 0, + "content": "as \\(C_i = (x_i, y_i, z_i)\\). For each pair of instance-level tokens \\(\\{O_{II}^{3D}, O_{Ij}^{3D}\\}\\), we calculate their Euclidean distance \\(d_{ij} = \\| C_i - C_j\\|_2\\) as well as their horizontal angle \\(\\theta_h\\) and vertical angle \\(\\theta_v\\). Specially, \\(\\theta_h = \\arctan 2((y_j - y_i) / (x_j - x_i))\\) and \\(\\theta_v = \\arcsin ((z_j - z_i) / d_{ij})\\). By making use of these positional parameters, we generate the pairwise spatial features \\(S = \\{s_{ij}\\} \\in \\mathbb{R}^{N \\times N \\times 5}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.429, + 0.483, + 0.448 + ], + "angle": 0, + "content": "\\[\ns _ {i j} = \\left[ \\sin (\\theta_ {h}), \\cos (\\theta_ {h}), \\sin (\\theta_ {v}), \\cos (\\theta_ {v}), d _ {i j} \\right]. (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.562 + ], + "angle": 0, + "content": "Inspired by the language-conditioned self-attention presented in ViL3DRel [5], we suggest a spatial-conditioned self-attention module. Specially, we create position embeddings \\(\\mathcal{P}\\) via the absolution positional encoding (PE), i.e. \\(\\mathcal{P} = \\mathrm{PE}(C_1,C_2,\\dots ,C_N)\\). A spatial conditioned attention weight \\(l_{i}\\) is computed to select the relevant spatial relations for each instance \\(O_{II}^{3D}\\), which is formulated as below:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.569, + 0.483, + 0.589 + ], + "angle": 0, + "content": "\\[\nl _ {i} = W _ {P} ^ {\\top} \\left(\\mathcal {P} _ {i} + O _ {I i} ^ {3 D}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.66 + ], + "angle": 0, + "content": "where \\(W_{P}\\in \\mathbb{R}^{D\\times 5}\\) is a learnable parameter and the bias term is omitted for simplicity. The spatial-conditioned attention map \\(\\omega_{ij}\\) is computed by combining \\(s_{ij},l_i\\) and \\(l_j\\), i.e. \\(\\omega_{ij} = l_i\\cdot s_{ij}\\cdot l_j\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.735 + ], + "angle": 0, + "content": "Through this positional transformation, the spatial-conditioned attention map \\(\\omega_{ij}\\) encapsulates the pairwise spatial relationship across the entire 3D scene for each instance. We further integrate the attention map with instance-level visual tokens, which is formulated as below:" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.743, + 0.483, + 0.786 + ], + "angle": 0, + "content": "\\[\nF _ {i} = \\sum_ {j = 1} ^ {N} \\omega_ {i j} O _ {I i} ^ {3 D}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.871 + ], + "angle": 0, + "content": "Finally, the instance-level representations are concatenated and further processed by a transformer-based encoder \\(\\Gamma\\), followed by a max-pooling layer (Pool) and a simple two-layer MLP. The scene-level relation-aware tokens \\(F_{scene} \\in \\mathbb{R}^{D}\\) are generated for the entire 3D scene as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.881, + 0.483, + 0.898 + ], + "angle": 0, + "content": "\\[\nF _ {s c e n e} = \\operatorname {M L P} (\\operatorname {P o o l} (\\Gamma ([ F _ {1}, F _ {2}, \\dots , F _ {N} ]))). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "Thus, we obtain the final token representations for LLM processing, i.e. instance-level visual tokens \\(O_{Ii}^{3D}\\) and spatial relation-aware scene-level tokens \\(F_{scene}\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.148, + 0.877, + 0.166 + ], + "angle": 0, + "content": "3.4. End-to-End Multi-task Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.172, + 0.907, + 0.277 + ], + "angle": 0, + "content": "3D-Language Prompts. Given an LMM, clear and explicit system messages and instructions are essential to support a range of downstream 3D vision-language tasks. For different tasks, we adopt various task-specific instruction templates with 3D features to generate uniform instruction data, enabling multi-task training. Please refer to the Supplementary Material for more detailed information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.909, + 0.43 + ], + "angle": 0, + "content": "Instruction Tuning. We conduct end-to-end multi-task fine-tuning to fully leverage the capabilities of our Inst3D-LMM framework. The MCMF and 3D-ISR modules enable the LLM with robust 3D scene understanding, grounding and reasoning abilities. We freeze the pre-trained 3D object encoder while updating both MCMF and 3D-ISR modules, as well as LLM. Upon completion of end-to-end multi-task instruction tuning, Inst3D-LMM can effectively handle various 3D vision-language tasks simultaneously, without fine-tuning on specific tasks, as illustrated in Figure 7." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.444, + 0.649, + 0.461 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.715, + 0.485 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.909, + 0.719 + ], + "angle": 0, + "content": "Datasets and Benchmarks. In this work, we conducted our experiments on the ScanNetV2 dataset [11], an extensive indoor 3D scene dataset comprising 1,513 scenes. This dataset includes 3D point clouds, RGB-D frames, and detailed point-level instance segmentation annotations. The whole dataset is divided into 1,201 scenes for training and 312 scenes for validation, with all subsequent benchmarks adhering to these training/validation splits. Our evaluation encompasses a range of 3D scene understanding benchmarks, including ScanRefer [4] and Multi3DRefer [55] for single- and multi-object 3D Visual Grounding, respectively, ScanQA [1] for 3D Question Answering, and Scan2Cap [8] for 3D Dense Captioning. These datasets are converted into a uniform instruction format for multi-task instruction tuning and performance assessment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. For 3D feature extraction, we leverage the 3D instance segmentation model Mask3D [38] pre-trained on ScanNet200 dataset [37], alongside the 3D object encoder Ulip2 [48]/Uni3D [57] based on ViT-L/14 [13]. For 2D feature extraction, we adopt ViT-H-based SAM [23] to obtain high-quality masks. Moreover, we extract 2D semantic features using the CLIP-ViT-L/14-336px vision encoder [35]. These pre-trained models are kept frozen. In our method, we apply multi-view selection and multi-level crops to 2D images, setting \\( K = 5 \\) views and \\( L = 3 \\) levels. We use Vicuna1.5-7B [10] as our basic LLM, which is fine-tuned from LLaMA2 [41]. Our fine-tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14151" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.91, + 0.398 + ], + "angle": 0, + "content": "
MethodOverallUniqueMultipleMulti3DRefer
Acc@0.25↑Acc@0.50↑Acc@0.25↑Acc@0.50↑Acc@0.25↑Acc@0.50↑F1@0.25↑F1@0.50↑
Closed-set, full sup.
ScanRefer [4]37.324.365.043.330.619.8--
3DVG-Trans [56]45.934.577.258.538.428.730.225.5
M3DRef-CLIP [55]51.944.7-77.2-36.842.838.4
ConcreteNet [42]56.149.586.179.247.540.9--
CORE-3DVG [51]56.843.885.067.151.839.8--
Zero-shot
LLM-Grounder [50]17.15.330.822.616.312.9--
Visual-Programming [54]36.432.763.858.427.724.6--
Specialist
OpenScene [34]13.05.120.113.111.14.4--
3D-LLM(Flamingo) [16]21.2-------
3D-LLM(BLIP2-flant5) [16]30.3-------
Chat-3D v2 [18]35.930.461.257.625.222.645.141.6
ReGround3D [58]53.141.1------
Generalist
LAMM [52]-3.4------
3DMIT [26]10.77.2------
Grounded 3D-LLM [7]47.944.1----45.240.6
Chat-Scene [19]55.550.2----57.152.4
Inst3D-LMM57.851.688.681.548.743.258.353.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.402, + 0.907, + 0.459 + ], + "angle": 0, + "content": "Table 1. Quantitative results for 3D Visual Grounding on ScanRefer and Multi3DRefer validation sets. In the ScanRefer dataset, scenes are labeled as \"unique\" (one object per class) or \"multiple\" (more than one). Closed-set methods are fully supervised for specific datasets. \"Zero-shot\" refers to methods that directly use LLMs without fine-tuning. \"Specialist\" and \"Generalist\" categorize methods fine-tuned for specific tasks versus those trained jointly. Bold and underlined numbers indicate the best and the second-best results, respectively." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.468, + 0.911, + 0.709 + ], + "angle": 0, + "content": "
Method# 3D Data for AlignmentScanQAScan2Cap@0.50
B-1↑B-4↑METEOR↑ROUGE↑CIDER↑EM↑B-4↑METEOR↑ROUGE↑CIDER↑
Closed-set, full sup.
VoteNet [12] + MCAN [53]-28.06.211.429.854.717.3----
ScanRefer [4] + MCAN [53]-26.97.911.530.055.418.6----
ScanQA [8]-30.210.113.133.364.921.0----
Scan2Cap [1]-------22.421.443.535.2
3D-VisTA [60]-34.213.115.238.676.627.034.027.154.366.9
LLM-based Methods
LLaVA (zero-shot) [28]-7.10.310.512.35.70.21.58.319.63.2
LAMM [52]25K26.85.810.023.642.49.8----
3D-LLM(Flamingo) [16]675K30.37.212.232.359.220.45.911.429.9-
3D-LLM(BLIP2-flant5) [16]675K39.312.014.535.769.420.58.113.133.2-
Chat-3D v2 [18]38K38.47.316.140.177.121.131.822.350.263.9
LL3DA [6]38K-13.315.437.075.7-35.925.654.665.2
Grounded 3D-LLM [7]107K-13.4--72.7-35.5--70.6
Chat-Scene [19]38K-14.3--87.7-36.3--77.1
Inst3D-LMM38K43.514.918.442.688.624.638.327.557.279.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.715, + 0.862, + 0.731 + ], + "angle": 0, + "content": "Table 2. Quantitative results for 3D Question Answering and 3D Dense Captioning on the ScanQA and Scan2Cap datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.483, + 0.793 + ], + "angle": 0, + "content": "process utilizes LoRA [17]. We adopt the AdamW optimizer with a weight decay of 0.02. All experiments were performed on 8 NVIDIA A100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.231, + 0.817 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.9 + ], + "angle": 0, + "content": "3D Visual Grounding. We first report visual grounding performance on the ScanRefer and Multi3DRefer validation datasets. As shown in Table 1, our approach outperforms the state-of-the-art model, Chat-Scene [19], by \\(+2.3\\%\\) Acc@0.25 and \\(+1.2\\%\\) F1@0.25 on ScanRefer and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.907, + 0.806 + ], + "angle": 0, + "content": "Multi3DRefer, respectively. Compared to specialist and closed-set methods, Inst3D-LMM, trained with a generalist approach, achieves competitive performance. Figure 5 displays typical visual comparison results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.806, + 0.909, + 0.881 + ], + "angle": 0, + "content": "3D Question Answering. We then compare Inst3D-LMM with previous leading methods on the ScanQA validation set. Table 2 reports the results. Our Inst3D-LMM consistently outperforms these methods, including recent LL3DA [6], Grounded 3D-LLM [7] and Chat-Scene [19]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.882, + 0.905, + 0.897 + ], + "angle": 0, + "content": "3D Dense Captioning. This task involves the localization" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14152" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.091, + 0.287, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.108, + 0.304, + 0.148 + ], + "angle": 0, + "content": "7-6" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.091, + 0.475, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.174, + 0.286, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.174, + 0.475, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.094, + 0.259, + 0.482, + 0.282 + ], + "angle": 0, + "content": "Text Query: The kitchen cabinets are under the sink. They are the furthest left cabinets and to the left of the dishwasher." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.295, + 0.483, + 0.352 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons in 3D Visual Grounding. The rendered images show the ground-truth (blue), incorrectly identified objects (red), and correctly identified objects (green). The colored text indicates the results of text decoupling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.484, + 0.437 + ], + "angle": 0, + "content": "and description of instances. As shown in Table 2, our approach achieves \\(38.3\\%\\) B-4@0.50 and \\(79.7\\%\\) C@0.50 on Scan2Cap, which exceeds the closed-set expert model 3D-VisTA [60] by \\(+4.3\\%\\) and \\(+12.8\\%\\), and still outperforms the LLM-based method LL3DA by \\(+2.4\\%\\) and \\(+14.5\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.446, + 0.245, + 0.463 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.514 + ], + "angle": 0, + "content": "In this section, we perform ablation experiments to thoroughly evaluate the effectiveness of the components. Please refer to the Supplementary Material for more analysis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.742 + ], + "angle": 0, + "content": "Effects of Multi-view Cross-Modal Fusion (MCMF). The MCMF module aims to augment LLM's understanding of 3D instances' geometric and semantic attributes. We explore several methods to combine these features, including direct concatenation for projection ('Concat.'), parallel projection followed by combination ('Parallel'), and vanilla cross-attention ('Cross-Attention'). To encode basic spatial information, we concatenate the results of absolution positional encoding (PE) with 3D features. As shown in Table 3, MCMF outperforms all these methods on ScanRefer, ScanQA and Scan2Cap datasets. We also observe that the performance of 3D LMM is significantly improved by incorporating multi-view 2D CLIP information compared to using only 3D geometric features. These results demonstrate the effectiveness of our MCMF approach." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.833 + ], + "angle": 0, + "content": "
MethodScanReferScanQAScan2Cap@0.50
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑B-4↑CIDER↑
w/o Multi-view 2D36.031.931.557.319.954.3
Concat.38.834.436.665.425.263.3
Parallel37.633.535.363.924.862.1
Cross-Attention39.236.537.466.125.564.3
MCMF46.741.941.578.632.768.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Table 3. Ablation evaluations of the proposed Multi-view CrossModal Fusion (MCMF) module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Impacts of 3D Instance Spatial Relation (3D-ISR). We further investigate the 3D-ISR module to analyze its impact" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.636, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.15, + 0.609, + 0.158 + ], + "angle": 0, + "content": "Input 3D Scene" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.092, + 0.757, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.15, + 0.728, + 0.158 + ], + "angle": 0, + "content": "Sentence Token" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.091, + 0.875, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.15, + 0.859, + 0.158 + ], + "angle": 0, + "content": "Position Embedding" + }, + { + "type": "image_caption", + "bbox": [ + 0.881, + 0.106, + 0.895, + 0.158 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image_footnote", + "bbox": [ + 0.516, + 0.16, + 0.89, + 0.177 + ], + "angle": 0, + "content": "Text Query: There is a brown reclined in the corner of the room. The adjacent walls have two green boards. It is between the intersection of the 2 boards." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.189, + 0.905, + 0.231 + ], + "angle": 0, + "content": "Figure 6. A visualization of the similarity score between text query and segmented 3D proposals. We compare sentence tokens used in ViL3DRel [5] with position embeddings employed in 3D-ISR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.907, + 0.466 + ], + "angle": 0, + "content": "on 3D spatial understanding using only 3D geometry priors. We compare 3D-ISR with previous spatial relation modeling methods, such as the 3D localization mechanism in ViL3DRel [5] and 3D-LLM [16], and relation-aware token generation in Chat-3D-v2 [18]. Table 4 reports the results of the comparison. It can be observed that our 3D-ISR consistently surpasses these approaches. Specifically, 3D-ISR significantly enhances performance on grounding tasks (ScanRefer and Multi3dRefer). This aligns with the motivation of our design. We also provide a visualization, as shown in Figure 6, of the similarity between the text query and segmented 3D proposals after processing by LLM. Compared to ViL3DRel [5] using sentence tokens, our 3D-ISR more accurately captures spatial relationships within the scene employing position embeddings of objects." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.468, + 0.905, + 0.548 + ], + "angle": 0, + "content": "
MethodScanReferMulti3DReferScanQA
Acc@0.25↑Acc@0.50↑F1@0.25↑F1@0.50↑B-1↑CIDER↑
w/o spatial relation35.629.834.827.329.255.6
ViL3DRel [5]36.130.538.435.532.858.9
3D-LLM [16]39.236.842.435.232.560.2
Chat-3D-v2 [18]40.837.541.637.835.164.6
3D-ISR48.344.146.241.339.172.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.551, + 0.905, + 0.579 + ], + "angle": 0, + "content": "Table 4. Ablation study to verify the effectiveness of our 3D Instance Spatial Relation (3D-ISR) module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Ablations within MCMF and 3D-ISR modules. We also explore the effectiveness of key designs in MCMF and 3D-ISR modules. The learnable [CLS] token in the MCMF module is introduced to aggregate 2D multi-view features. We compare it with alternative methods, such as token max pooling and Q-Former [24]. Table 5 reports the comparison results, which demonstrate the efficacy of the learnable [CLS] token design. Besides, to assess the impact of different pairwise spatial features in the 3D-ISR module, we evaluate the model using only distance or orientation information to compute spatial features in Eq 4. The results in Table 6 reveal that distance has a greater impact on the grounding capability of the model, while orientation is more crucial in handling Q&A tasks. Combining both pairwise distance and orientation yields the best overall performance. Mutual Benefits of MCMF and 3D-ISR. As shown in Table 7, the model that integrates MCMF and 3D-ISR consistently outperforms those that use either module alone for multiple tasks. To further verify its effectiveness, we combine MCMF with absolution positional encoding (PE) for the basic spatial relation, and directly adopt feature con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14153" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.895, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.381, + 0.829, + 0.395 + ], + "angle": 0, + "content": "Figure 7. Qualitative illustration of Inst3D-LMM across various 3D-language tasks in diverse 3D environments." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.474 + ], + "angle": 0, + "content": "
MethodScanReferScanQAScan2Cap@0.50
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑B-4↑CIDER↑
Max Pooling48.744.639.274.830.665.5
Q-Former [24]53.247.944.387.935.478.6
CLS Token57.851.643.588.638.379.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.479, + 0.482, + 0.492 + ], + "angle": 0, + "content": "Table 5. Ablation study of learnable CLS token in MCMF module." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.495, + 0.482, + 0.56 + ], + "angle": 0, + "content": "
MethodScanReferMulti3DReferScanQA
Acc@0.25↑Acc@0.50↑F1@0.25↑F1@0.50↑B-1↑CIDER↑
w/ Dist (only)43.238.042.437.831.265.7
w/ Ort (only)39.733.640.535.932.566.8
Dist + Ort48.344.146.241.339.172.3
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.595 + ], + "angle": 0, + "content": "Table 6. Ablation study of 3D-ISR module. Dist (Ort) means only using distance (orientation) to compute pairwise spatial features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.483, + 0.648 + ], + "angle": 0, + "content": "catenation (Concat.) to integrate 3D object features and 2D multi-view features along with 3D-ISR. These results indicate the synergy efforts between MCMF and 3D-ISR." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.483, + 0.8 + ], + "angle": 0, + "content": "Different LLMs and Foundation Models. Table 8 presents the results of various LLMs and pre-trained 2D/3D models, including Vicuna-7B [10] vs Vicuna-13B, Ulip2 [48] vs Uni3D [57], CLIP [35] vs SigLIP [32], and ViT-H-based SAM [23] vs ViT-L-based SAM. We find that the performance of grounding and reasoning increases along with the total number of parameters in the foundation models (i.e. ViT-L vs Vit-H and 7B vs 13B). These results indicate that our framework's capability can be improved in tandem with the performance of foundation models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.21, + 0.83 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In this paper, we proposed Inst3D-LMM, an effective instance-aware framework to leverage the potential of Large Multi-modal Models (LMMs) for 3D scene understanding. To improve instance-level representations, we developed" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.405, + 0.9, + 0.485 + ], + "angle": 0, + "content": "
MethodScanReferScanQAScan2Cap@0.50
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑B-4↑CIDER↑
w/ MCMF (only)38.234.639.476.830.566.9
w/ 3D-ISR (only)48.344.139.172.329.864.3
MCMF + PE46.741.941.578.632.768.2
3D-ISR + Concat.49.845.040.675.433.366.5
MCMF + 3D-ISR (full)57.851.643.588.638.379.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.491, + 0.906, + 0.517 + ], + "angle": 0, + "content": "Table 7. Comparison results of the collaboration between MCMF and 3D-ISR modules." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.523, + 0.9, + 0.607 + ], + "angle": 0, + "content": "
LLM3D Encoder2D VFsScanReferScanQA
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑
Vicuna-7BUlip2CLIP+SAM (ViT-L)49.244.839.579.6
Vicuna-7BUlip2CLIP+SAM (ViT-H)53.646.543.787.5
Vicuna-7BUlip2SigLIP+SAM (ViT-H)54.347.042.284.5
Vicuna-7BUni3DCLIP+SAM (ViT-H)57.851.643.588.6
Vicuna-13BUni3DCLIP+SAM (ViT-H)56.052.042.883.1
Vicuna-13BUni3DSigLIP+SAM (ViT-H)55.247.843.385.4
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.613, + 0.906, + 0.64 + ], + "angle": 0, + "content": "Table 8. Ablation study of different LLMs and pre-trained foundation models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.794 + ], + "angle": 0, + "content": "a novel Multi-view Cross-Modal Fusion (MCMF) module, which injects multi-view 2D semantic open-vocabulary priors into 3D geometry features to generate fine-grained instance-level tokens. Furthermore, we introduced a 3D Instance Spatial Relation (3D-ISR) module that employs the spatial condition attention mechanism to capture pairwise spatial relations. Experimental results demonstrate that our approach achieves promising performance in understanding and reasoning across various 3D vision-language tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.814, + 0.673, + 0.83 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China under Grants (62376244). It is also supported by Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14154" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.483, + 0.157 + ], + "angle": 0, + "content": "[1] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, pages 19129-19139, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.212 + ], + "angle": 0, + "content": "[2] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In CVPR, pages 16464-16473, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.483, + 0.268 + ], + "angle": 0, + "content": "[3] Yang Cao, Zeng Yihan, Hang Xu, and Dan Xu. Coda: Collaborative novel box discovery and cross-modal alignment for open-vocabulary 3d object detection. In NeurIPS, pages 71862-71873, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.482, + 0.312 + ], + "angle": 0, + "content": "[4] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221. Springer, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.314, + 0.482, + 0.368 + ], + "angle": 0, + "content": "[5] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. In NeurIPS, pages 20522-20535, 2022. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[6] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In CVPR, pages 26428-26438, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[7] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.483, + 0.536 + ], + "angle": 0, + "content": "[8] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In CVPR, pages 3193-3203, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[9] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In ICCV, pages 18109-18119, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.595, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.678, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 2432–2443, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.774 + ], + "angle": 0, + "content": "[12] Zhipeng Ding, Xu Han, and Marc Niethammer. Volenet: A deep learning label fusion method for multi-atlas segmentation. In MICCAI, pages 202-210, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.777, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm:" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[15] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. In NeurIPS, pages 20482-20494, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[17] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[18] Haifeng Huang, Zehan Wang, Rongjie Huang, Luping Liu, Xize Cheng, Yang Zhao, Tao Jin, and Zhou Zhao. Chat-3d v2: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.364, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[19] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. NeuIPS, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[20] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[21] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. In ECCV, pages 528-545. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[22] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In CVPR, pages 10984-10994, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[25] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[26] Zeju Li, Chao Zhang, Xiaoyan Wang, Ruilong Ren, Yifan Xu, Ruifei Ma, and Xiangde Liu. 3dmit: 3d multi-modal instruction tuning for scene understanding. In ICMEW, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Cai Liang, Bo Li, Zhengming Zhou, Longlong Wang, Pengfei He, Liang Hu, and Haoxing Wang. Spatioawaregrouding3d: A spatio aware model for improving 3d vision grousing. In CVPRW, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14155" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, pages 34892-34916, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[29] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In NeurIPS, pages 44860-44879, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[30] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[31] Yunze Man, Shuhong Zheng, Zhipeng Bao, Martial Hebert, Liang-Yan Gui, and Yu-Xiong Wang. Lexicon3d: Probing visual foundation models for complex 3d scene understanding. arXiv preprint arXiv:2409.03757, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[32] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, pages 529-544. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[33] Maria Parelli, Alexandros Delitzas, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Clip-guided vision-language pre-training for question answering in 3d scenes. In CVPR, pages 5606-5611, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[34] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In CVPR, pages 815-824, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[36] Krishan Rana, Jesse Haviland, Sourav Garg, Jad Abou-Chakra, Ian Reid, and Niko Suenderhauf. Sayplan: Grounding large language models using 3d scene graphs for scalable robot task planning. In CoRL, pages 23-72, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.637 + ], + "angle": 0, + "content": "[37] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, pages 125-141. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[38] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In ICRA, pages 8216-8223, 2023. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[39] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M. Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In ICCV, pages 2998-3009, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[40] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In NeurIPS, pages 68367-68390, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[42] Ozan Unal, Christos Sakaridis, Suman Saha, Fisher Yu, and Luc Van Gool. Three ways to improve verbo-visual" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "fusion for dense 3d visual grounding. arXiv preprint arXiv:2309.04561, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[43] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, et al. Embodiedscan: A holistic multimodal 3d perception suite towards embodied ai. In CVPR, pages 19757-19767, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[44] Zehan Wang, Haifeng Huang, Yang Zhao, Linjun Li, Xize Cheng, Yichen Zhu, Aoxiong Yin, and Zhou Zhao. 3drp-net: 3d relative position-aware network for 3d visual grounding. In EMNLP, page 10612-10625, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[45] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[46] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In CVPR, pages 19231-19242, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[47] Runsen Xu, Zhiwei Huang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Vlm-grounder: A vlm agent for zero-shot 3d visual grounding. arXiv preprint arXiv:2410.13860, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[48] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In CVPR, pages 27091–27101, 2024. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[49] Fan Yang, Sicheng Zhao, Yanhao Zhang, Haoxiang Chen, Hui Chen, Wenbo Tang, Haonan Lu, Pengfei Xu, Zhenyu Yang, Jungong Han, et al. Lmi3d: Empowering llm with 3d perception from a single 2d image. arXiv preprint arXiv:2408.07422, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[50] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. In ICRA, pages 7694-7701, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[51] Li Yang, Ziqi Zhang, Zhongang Qi, Yan Xu, Wei Liu, Ying Shan, Bing Li, Weiping Yang, Peng Li, Yan Wang, et al. Exploiting contextual objects and relations for 3d visual grounding. In NeurIPS, pages 49542-49554, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[52] Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Xiaoshui Huang, Zhiyong Wang, Lu Sheng, Lei Bai, et al. Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. In NeurIPS, pages 26650-26685, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[53] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In CVPR, pages 6281–6290, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[54] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3d visual grounding. In CVPR, pages 20623-20633, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[55] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, pages 15225-15236, 2023. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14156" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[56] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2928-2937, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[57] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. In ICLR, 2024. 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[58] Chenming Zhu, Tai Wang, Wenwei Zhang, Kai Chen, and Xihui Liu. Empowering 3d visual grounding with reasoning capabilities. arXiv preprint arXiv:2407.01525, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[59] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmms with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[60] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 2, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14157" + } + ] +] \ No newline at end of file diff --git a/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/ada2a42d-606c-400d-85d5-a68b46a5f62b_origin.pdf b/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/ada2a42d-606c-400d-85d5-a68b46a5f62b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d261e465bf623ccd0228d8ead3f32bd777d3965e --- /dev/null +++ b/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/ada2a42d-606c-400d-85d5-a68b46a5f62b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8bd9f5db929442b59d1f9017cc54f5aa8906bbd10be884fa79ece8127545b80c +size 2945988 diff --git a/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/full.md b/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7995ea26f809aa306edf7511b733efa37f8e106e --- /dev/null +++ b/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/full.md @@ -0,0 +1,305 @@ +# Inst3D-LMM: Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning + +Hanxun Yu $^{1*}$ , Wentong Li $^{2*}$ , Song Wang $^{1}$ , Junbo Chen $^{3\dagger}$ , Jianke Zhu $^{1\dagger}$ $^{1}$ Zhejiang University $^{2}$ Nanjing University of Aeronautics and Astronautics $^{3}$ Udeer.ai {hanxun.yu, songw, jkzhu}@zju.edu.cn, wentong.li@nuaa.edu.cn, junbo@udeer.ai + +# Abstract + +Despite encouraging progress in 3D scene understanding, it remains challenging to develop an effective Large Multi-modal Model (LMM) that is capable of understanding and reasoning in complex 3D environments. Most previous methods typically encode 3D point and 2D image features separately, neglecting interactions between 2D semantics and 3D object properties, as well as the spatial relationships within the 3D environment. This limitation not only hinders comprehensive representations of 3D scene, but also compromises training and inference efficiency. To address these challenges, we propose a unified Instance-aware 3D Large Multi-modal Model (Inst3D-LMM) to deal with multiple 3D scene understanding tasks simultaneously. To obtain the fine-grained instance-level visual tokens, we first introduce a novel Multi-view Cross-Modal Fusion (MCMF) module to inject the multi-view 2D semantics into their corresponding 3D geometric features. For scene-level relation-aware tokens, we further present a 3D Instance Spatial Relation (3D-ISR) module to capture the intricate pairwise spatial relationships among objects. Additionally, we perform end-to-end multi-task instruction tuning simultaneously without the subsequent task-specific fine-tuning. Extensive experiments demonstrate that our approach outperforms the state-of-the-art methods across 3D scene understanding, reasoning and grounding tasks. Source code is available at: https://github.com/hanxunyu/Inst3D-LMM. + +# 1. Introduction + +Building Large Multi-modal Models (LMMs) for 3D scene understanding becomes an emerging research topic with significant potential for advancing autonomous robotics [36]. For example, interactive embodied agents [39] are expected to interpret 3D layouts and predict + +object locations based on human instructions. + +Traditional 3D scene understanding methods [30, 46] are typically tailored for individual downstream tasks, such as 3D Visual Grounding (3D-VG), 3D Question Answering (3D-QA) and 3D Dense Captioning (3D-DC). In contrast, LMMs are able to handle various 3D perception tasks within a single model. Some methods [16, 43] primarily focus on translating 3D points into the space of 2D Vision Language Models (VLMs) or directly leveraging multiview 2D features as 3D representations. Alternatively, other approaches [6, 45, 58] directly encode the features of 3D points and facilitate alignment with LLM using 3D-text instruction data. However, they often require multi-stage alignment or language-scene pre-training, complicating the development of a versatile model capable of handling multiple tasks. To enable a unified 3D LMM framework, recent work [19] decomposes the input 3D scene into a set of individual object proposals, each identified by unique tokens to explicitly capture 3D object features at the instance level. While this approach exhibits promising results, it neglects the interactions between the 2D semantic features and the properties of 3D objects, as well as the spatial relationship modeling among objects in 3D environments. This oversight further results in substantial token costs for the LLM, thus hindering both training and inference efficiency. + +In this paper, we propose Inst3D-LMM, an effective Instance-aware 3D Large Multi-modal Model that tackles multiple 3D-language tasks without resorting to task-specific fine-tuning. Our approach fully leverages the powerful 2D Vision Foundation Models (VFMs) and 3D specialist models to extract enriched 2D and 3D features at the instance level, respectively. As shown in Figure 1, in contrast to previous methods, our approach is able to generate fine-grained instance-level representations that encapsulate both geometric and semantic properties, and scene-level representations that capture intricate pairwise spatial relationships among objects in a 3D scene. Moreover, our method results in minor token costs for LLM, thereby enhancing both training and inference efficiency. By leveraging this instance-aware methodology, our Inst3D-LMM + +![](images/8fefee1109f574bdeac6eea50d061c38485d98b8381a99f38aa3e2f0b78da53d.jpg) +Figure 1. Comparisons between previous 3D LMMs and our proposed Inst3D-LMM. (a) Previous methods [6, 16, 19, 20] typically encode the features of 3D points or 2D images separately and concatenate them directly, often requiring task-specific fine-tuning for different tasks. (b) Illustration of Inst3D-LMM. Our method integrates 2D/3D cross-modal information and captures the intricate spatial relations among objects within 3D environments to generate tight but informative instance/scene-level tokens for the LMM. (c) Compared with other 3D LMMs, our Inst3D-LMM requires fewer computational resources, while offering faster training and inference speeds. + +![](images/59392d37dc948e2cecfddbeef78a690e045fc039d2a2b20cc70ad0b4e050f104.jpg) + +significantly improves the LLM's ability to comprehend 3D scenes with respect to both efficiency and accuracy. + +Specifically, we introduce a novel Multi-view CrossModal Fusion (MCMF) module that effectively infuses enriched multi-view 2D features into the original 3D features with coarse semantics. A learnable [CLS] token is introduced to aggregate the characteristics of each 2D view, enabling efficient multi-view 2D-to-3D cross-modal transformation. To capture intricate pairwise spatial relationships among objects in a 3D scene, we then propose a 3D Instance Spatial Relation (3D-ISR) module. A spatial condition self-attention between manifold position embeddings and instance-level tokens is presented to produce relation-aware scene-level representations. The resulting instance-level and scene-level representations are subsequently fed into the LLM for end-to-end multi-task instruction tuning. + +Extensive experiments across various tasks, including 3D-VG, 3D-QA and 3D-DC, demonstrate that our approach outperforms previous state-of-the-art methods with leading 3D scene understanding, grounding and reasoning capabilities. Unlike most existing methods that focus on close-set scene understanding or require per-task fine-tuning, our Inst3D-LMM operates as a generalist model. We believe this work lays a fundamental step towards unifying diverse 3D vision-language tasks in generative language modeling. To summarize, our contributions are as follows: + +- We propose a unified and efficient instance-aware LLM-based framework, called Inst3D-LMM for various 3D scene understanding tasks with end-to-end multi-modal instruction tuning. Serving as a generalist model, our approach demonstrates superior performance across 3D scene understanding, reasoning and spatial localization. +- We utilize 2D VFMs to extract multi-view contextual features for each 3D instance and then devise a Multi-view Cross-Modal Fusion (MCMF) module to effectively en + +hance instance-level feature representations by jointly integrating 3D geometry and 2D semantic priors. + +- A 3D Instance Spatial Relation (3D-ISR) module is introduced to boost the capability of LMM in understanding the complex spatial details within 3D scenes. + +# 2. Related Works + +3D Scene Understanding with Language. In 3D scene understanding, there is a surge of interest in making use of language queries to capture user intentions for various downstream tasks, such as 3D Visual Grounding [5, 44], 3D Question Answering [30, 33] and 3D Dense Captioning [8, 21]. Specifically, 3D Visual Grounding entails localizing target objects based on language queries. Moreover, 3D Question Answering demands robust 3D spatial perception and reasoning. 3D Dense Captioning involves localizing and describing objects in 3D scenes. Conventional methods typically focus on a specific task. Instead, 3D visual grounding and dense captioning tasks are combined by leveraging their complementary aspects [2, 9, 31, 47]. Recent efforts like 3D-VLP [22] and 3D-VisTA [60] attempt to establish a universal framework by pre-aligning 3D scenes with their corresponding textual descriptions. In contrast to our Inst3D-LMM, most existing methods still focus on close-set scene understanding, which requires either task-specific fine-tuning or striving to build specialized models. + +3D Large Multi-modal Models. Inspired by the significant advancements in Large Language Models (LLMs), researchers extend LLM's knowledge to encompass 3D modality [29, 49, 57, 59]. Point-LLM [14] and Imagebind-LLM [15] have succeeded in bridging the gap between 3D visuals and text by utilizing extensive 3D object datasets. However, these models struggle with interpreting complex spatial relationships in 3D scenes. Another promising di + +![](images/fd36b538b11330d06979505d777b82f51104af180a9302c9d3e26b2ef8929019.jpg) +Figure 2. Overview of our proposed Inst3D-LMM. Our pipeline takes as input point clouds of a 3D indoor scene, along with RGB-D images. We first employ the pre-trained 3D specialist models and 2D VFMs to extract 3D proposals and multi-view 2D semantic features, respectively. We then suggest the MCMF module to generate fine-grained instance-level tokens. A 3D-ISR module is further introduced to create relation-aware scene-level tokens based on spatial distances. By leveraging the constructed 3D-language prompts, we conduct multi-task instruction tuning to simultaneously handle various 3D tasks. + +rection focuses on developing the scene-level 3D LMMs. Hong et al. [16] encode projected 3D features using a 2D vision encoder and incorporate location tokens to augment LLM vocabulary. Other methods [6, 45, 58] directly encode point clouds and utilize 3D scene-text data for better visual interaction through pre-alignment. Huang et al. [18] employ a three-stage training scheme and adopt object identifiers to learn individual object attributes. Chen et al. [7] use special referent tokens for precise referencing and grounding. Wang et al. [43] encode point clouds and RGB images separately. However, some crucial semantic pixels are often lost during their sparse fusion processes, resulting in a coarse visual semantic representation. In this work, we propose an effective instance-aware framework to fuse fine-grained cross-modal information and encode spatial relations, which achieves promising results on multiple 3D-language tasks as a generalist model. + +# 3. Methodology + +Our goal is to enable LLM to understand the 3D environment and perform various visual interaction tasks based on + +human instructions. Figure 2 illustrates the architecture of our framework. In this section, we first introduce how to extract features at the instance level using pre-trained 3D models [38, 57] and 2D VFs [23, 35], respectively. Secondly, we present a novel Multi-view Cross-Modal Fusion (MCMF) module to obtain fine-grained instance-level tokens, which are specially designed to effectively integrate 3D geometric features with their corresponding multi-view 2D semantic features. Thirdly, we introduce the 3D Instance Spatial Relation (3D-ISR) module to enhance LLM's ability to capture spatial information at the scene level, which generates relation-aware tokens through attention-based analysis of spatial relationships among different 3D proposals. Under our proposed framework, the MCMF and 3D-ISR modules are jointly optimized, allowing mutual enhancement. Finally, we conduct end-to-end multi-task instruction tuning to address a range of 3D scene understanding tasks simultaneously. + +# 3.1. Instance-Level Feature Extraction + +3D Feature Extraction. We first segment the 3D point clouds of each individual instance in a class-agnostic man + +ner by leveraging a pre-trained 3D instance segmentation model [38]. We only retain the predicted binary 3D instance masks while ignoring their closed-vocabulary class labels. All instance proposals in one scene are represented by $O^{3D} = (O_1^{3D}, \ldots, O_N^{3D})$ , where $O_i^{3D} = [coordinate, color]$ consists of the attributes of each instance. We obtain all instance proposals of a scene, and the pre-trained 3D encoder $\mathbf{E}_{\mathbf{o}}$ [57] is used to extract their instance-level features, i.e., $\mathbf{f}_{\mathbf{o}}^{3D} = \mathbf{E}_{\mathbf{o}}(O^{3D})$ . + +2D Feature Extraction. Due to the inherent sparsity of 3D point cloud data, previous methods have difficulties in generating discriminative features of each object. In this work, we employ powerful 2D VFMs to extract 2D semantic features for each 3D instance. As in [3, 40], we firstly project the point cloud of each instance $O_{i}^{3D}$ onto the image plane and then select the top $K$ views according to the number of visible points. To obtain accurate 2D masks, we randomly sample $k_{\mathrm{sample}} = 5$ points as input prompts for SAM [23] to deal with noisy bounding boxes with outliers. Hereby, we select the high-quality mask with the highest confidence score. To enrich features with contextual information, these masks are employed to generate multi-level $(L)$ crops of the selected images that are further fed into the pre-trained CLIP vision encoder [35] to extract features with language-aligned embedding space. Finally, we aggregate multi-level features of sub-images from the same frame to form the 2D multi-view features $O^{2D}$ . + +# 3.2. Multi-view Cross-Modal Fusion + +To better fuse 3D geometry priors with 2D multi-view semantic priors, we introduce an effective Multi-view CrossModal Fusion (MCMF) module that generates the enriched token representations for each 3D instance before being fed into LLM. The architecture of MCMF is crafted with a coarse-to-fine framework, as shown in Figure 3. To map 3D object features $O^{3D} \in \mathbb{R}^{1 \times N \times D^{3d}}$ and 2D CLIP features $O^{2D} \in \mathbb{R}^{K \times N \times D^{2d}}$ into the embedding space of LLM with the dimensionality of $D$ , we utilize a simple two-layer MLP with a LayerNorm and GELU in between, yielding $O^{3D'}$ and $O^{2D'}$ , respectively. $N$ represents the number of instances, and $K$ is the number of images for each object. + +Subsequently, we introduce a Cross-Modal Injection Block to transform the enriched semantic priors from 2D multi-view representations into 3D instance features. Moreover, we adopt a straightforward self-attention layer to further enhance such 3D instance features. For 2D multi-view features, we append a learnable [CLS] token $t_k$ to the flattened CLIP feature maps in order to adaptively encapsulate the global semantic representation of the $k$ -th view. Then, we apply a self-attention layer to multi-view features, enabling semantic gathering to derive $t_k'$ for $k$ -th view as: + +$$ +t _ {k} ^ {\prime} = \phi (\text {S e l f A t t n} ([ t _ {k}, O _ {\text {v i e w} _ {k}} ^ {2 D ^ {\prime}} ])), \tag {1} +$$ + +![](images/e26d593e3821672fd474eed8c2f89be47e2793c7ee27458b3983d6ffe30c7784.jpg) +Figure 3. Architecture of the proposed Multi-view Cross-Modal Fusion (MCMF) module. + +where $1 \leq k \leq K$ , $O_{\mathrm{view}_k}^{2D'} \in \mathbb{R}^{N \times D}$ . $\phi$ is an indicator function that outputs the updated [CLS] token as the first one from the token list. + +We obtain the processed 3D token embedding $O^{3D'} \in \mathbb{R}^{1 \times N \times D}$ containing coarse semantics, where each instance in $O^{3D'}$ corresponds to multiple views of the processed 2D CLIP features $O^{2D'}$ . It can be represented as follows: + +$$ +O _ {i} ^ {2 D ^ {\prime}} = \left[ t _ {1} ^ {\prime}, \dots , t _ {K} ^ {\prime} \right], \quad i \in [ 1, N ]. \tag {2} +$$ + +To infuse the enriched semantic 2D features into 3D geometric feature, we construct the 3D-to-2D-Multi-view pairs, i.e. each 3D instance feature in $O^{3D'} \in \mathbb{R}^{1 \times N \times D}$ corresponds to 2D features across $K$ views in $O^{2D'} \in \mathbb{R}^{K \times N \times D}$ . In particular, we utilize the preliminary $O^{3D'}$ as queries, while the 2D multi-view visual features $O^{2D'}$ as enriched reference keys and values. Inspired by [25], the injection process is conducted via a cross-attention layer, which encourages 3D instance queries to absorb the fine-grained semantics of keys and values. This results in the enhanced 3D instance features $O_f^{3D} \in \mathbb{R}^{1 \times N \times D}$ as below: + +$$ +O _ {f} ^ {3 D} = \operatorname {C r o s s A t t n} \left(O ^ {3 D ^ {\prime}}, O ^ {2 D ^ {\prime}}\right). \tag {3} +$$ + +Furthermore, we leverage a residual operation by adding the input $O^{3D'}$ to retain the basic characteristics in generating the instance level 3D visual tokens, $O_{I}^{3D} = O_{f}^{3D} + O^{3D'}$ . + +# 3.3. 3D Instance Spatial Relation + +Motivated by [5, 27, 60], we develop an effective 3D Instance Spatial Relation (3D-ISR) module to boost LLM's capabilities in assimilating spatial information within the 3D scene, as shown in Figure 4. 3D-ISR utilizes the instance-level 3D visual tokens $\{O_{I1}^{3D}, O_{I2}^{3D}, \ldots, O_{IN}^{3D}\}$ derived from MCMF module along with the corresponding center coordinates $\{C_1, C_2, \ldots, C_N\}$ of all instances as input. For the $i$ -th instance, we define its center coordinates + +![](images/1db6d1c3d65a4d75503fbbfd2fd99a7a85fcd5a99457318ba57bd6bfdd3874da.jpg) +Figure 4. Illustration of the 3D Instance Spatial Relation (3D-ISR) module in our framework. + +as $C_i = (x_i, y_i, z_i)$ . For each pair of instance-level tokens $\{O_{II}^{3D}, O_{Ij}^{3D}\}$ , we calculate their Euclidean distance $d_{ij} = \| C_i - C_j\|_2$ as well as their horizontal angle $\theta_h$ and vertical angle $\theta_v$ . Specially, $\theta_h = \arctan 2((y_j - y_i) / (x_j - x_i))$ and $\theta_v = \arcsin ((z_j - z_i) / d_{ij})$ . By making use of these positional parameters, we generate the pairwise spatial features $S = \{s_{ij}\} \in \mathbb{R}^{N \times N \times 5}$ as: + +$$ +s _ {i j} = \left[ \sin (\theta_ {h}), \cos (\theta_ {h}), \sin (\theta_ {v}), \cos (\theta_ {v}), d _ {i j} \right]. (4) +$$ + +Inspired by the language-conditioned self-attention presented in ViL3DRel [5], we suggest a spatial-conditioned self-attention module. Specially, we create position embeddings $\mathcal{P}$ via the absolution positional encoding (PE), i.e. $\mathcal{P} = \mathrm{PE}(C_1,C_2,\dots ,C_N)$ . A spatial conditioned attention weight $l_{i}$ is computed to select the relevant spatial relations for each instance $O_{II}^{3D}$ , which is formulated as below: + +$$ +l _ {i} = W _ {P} ^ {\top} \left(\mathcal {P} _ {i} + O _ {I i} ^ {3 D}\right), \tag {5} +$$ + +where $W_{P}\in \mathbb{R}^{D\times 5}$ is a learnable parameter and the bias term is omitted for simplicity. The spatial-conditioned attention map $\omega_{ij}$ is computed by combining $s_{ij},l_i$ and $l_j$ , i.e. $\omega_{ij} = l_i\cdot s_{ij}\cdot l_j$ + +Through this positional transformation, the spatial-conditioned attention map $\omega_{ij}$ encapsulates the pairwise spatial relationship across the entire 3D scene for each instance. We further integrate the attention map with instance-level visual tokens, which is formulated as below: + +$$ +F _ {i} = \sum_ {j = 1} ^ {N} \omega_ {i j} O _ {I i} ^ {3 D}. \tag {6} +$$ + +Finally, the instance-level representations are concatenated and further processed by a transformer-based encoder $\Gamma$ , followed by a max-pooling layer (Pool) and a simple two-layer MLP. The scene-level relation-aware tokens $F_{scene} \in \mathbb{R}^{D}$ are generated for the entire 3D scene as follows: + +$$ +F _ {s c e n e} = \operatorname {M L P} (\operatorname {P o o l} (\Gamma ([ F _ {1}, F _ {2}, \dots , F _ {N} ]))). \tag {7} +$$ + +Thus, we obtain the final token representations for LLM processing, i.e. instance-level visual tokens $O_{Ii}^{3D}$ and spatial relation-aware scene-level tokens $F_{scene}$ , respectively. + +# 3.4. End-to-End Multi-task Instruction Tuning + +3D-Language Prompts. Given an LMM, clear and explicit system messages and instructions are essential to support a range of downstream 3D vision-language tasks. For different tasks, we adopt various task-specific instruction templates with 3D features to generate uniform instruction data, enabling multi-task training. Please refer to the Supplementary Material for more detailed information. + +Instruction Tuning. We conduct end-to-end multi-task fine-tuning to fully leverage the capabilities of our Inst3D-LMM framework. The MCMF and 3D-ISR modules enable the LLM with robust 3D scene understanding, grounding and reasoning abilities. We freeze the pre-trained 3D object encoder while updating both MCMF and 3D-ISR modules, as well as LLM. Upon completion of end-to-end multi-task instruction tuning, Inst3D-LMM can effectively handle various 3D vision-language tasks simultaneously, without fine-tuning on specific tasks, as illustrated in Figure 7. + +# 4. Experiments + +# 4.1. Experimental Setting + +Datasets and Benchmarks. In this work, we conducted our experiments on the ScanNetV2 dataset [11], an extensive indoor 3D scene dataset comprising 1,513 scenes. This dataset includes 3D point clouds, RGB-D frames, and detailed point-level instance segmentation annotations. The whole dataset is divided into 1,201 scenes for training and 312 scenes for validation, with all subsequent benchmarks adhering to these training/validation splits. Our evaluation encompasses a range of 3D scene understanding benchmarks, including ScanRefer [4] and Multi3DRefer [55] for single- and multi-object 3D Visual Grounding, respectively, ScanQA [1] for 3D Question Answering, and Scan2Cap [8] for 3D Dense Captioning. These datasets are converted into a uniform instruction format for multi-task instruction tuning and performance assessment. + +Implementation Details. For 3D feature extraction, we leverage the 3D instance segmentation model Mask3D [38] pre-trained on ScanNet200 dataset [37], alongside the 3D object encoder Ulip2 [48]/Uni3D [57] based on ViT-L/14 [13]. For 2D feature extraction, we adopt ViT-H-based SAM [23] to obtain high-quality masks. Moreover, we extract 2D semantic features using the CLIP-ViT-L/14-336px vision encoder [35]. These pre-trained models are kept frozen. In our method, we apply multi-view selection and multi-level crops to 2D images, setting $K = 5$ views and $L = 3$ levels. We use Vicuna1.5-7B [10] as our basic LLM, which is fine-tuned from LLaMA2 [41]. Our fine-tuning + +
MethodOverallUniqueMultipleMulti3DRefer
Acc@0.25↑Acc@0.50↑Acc@0.25↑Acc@0.50↑Acc@0.25↑Acc@0.50↑F1@0.25↑F1@0.50↑
Closed-set, full sup.
ScanRefer [4]37.324.365.043.330.619.8--
3DVG-Trans [56]45.934.577.258.538.428.730.225.5
M3DRef-CLIP [55]51.944.7-77.2-36.842.838.4
ConcreteNet [42]56.149.586.179.247.540.9--
CORE-3DVG [51]56.843.885.067.151.839.8--
Zero-shot
LLM-Grounder [50]17.15.330.822.616.312.9--
Visual-Programming [54]36.432.763.858.427.724.6--
Specialist
OpenScene [34]13.05.120.113.111.14.4--
3D-LLM(Flamingo) [16]21.2-------
3D-LLM(BLIP2-flant5) [16]30.3-------
Chat-3D v2 [18]35.930.461.257.625.222.645.141.6
ReGround3D [58]53.141.1------
Generalist
LAMM [52]-3.4------
3DMIT [26]10.77.2------
Grounded 3D-LLM [7]47.944.1----45.240.6
Chat-Scene [19]55.550.2----57.152.4
Inst3D-LMM57.851.688.681.548.743.258.353.5
+ +Table 1. Quantitative results for 3D Visual Grounding on ScanRefer and Multi3DRefer validation sets. In the ScanRefer dataset, scenes are labeled as "unique" (one object per class) or "multiple" (more than one). Closed-set methods are fully supervised for specific datasets. "Zero-shot" refers to methods that directly use LLMs without fine-tuning. "Specialist" and "Generalist" categorize methods fine-tuned for specific tasks versus those trained jointly. Bold and underlined numbers indicate the best and the second-best results, respectively. + +
Method# 3D Data for AlignmentScanQAScan2Cap@0.50
B-1↑B-4↑METEOR↑ROUGE↑CIDER↑EM↑B-4↑METEOR↑ROUGE↑CIDER↑
Closed-set, full sup.
VoteNet [12] + MCAN [53]-28.06.211.429.854.717.3----
ScanRefer [4] + MCAN [53]-26.97.911.530.055.418.6----
ScanQA [8]-30.210.113.133.364.921.0----
Scan2Cap [1]-------22.421.443.535.2
3D-VisTA [60]-34.213.115.238.676.627.034.027.154.366.9
LLM-based Methods
LLaVA (zero-shot) [28]-7.10.310.512.35.70.21.58.319.63.2
LAMM [52]25K26.85.810.023.642.49.8----
3D-LLM(Flamingo) [16]675K30.37.212.232.359.220.45.911.429.9-
3D-LLM(BLIP2-flant5) [16]675K39.312.014.535.769.420.58.113.133.2-
Chat-3D v2 [18]38K38.47.316.140.177.121.131.822.350.263.9
LL3DA [6]38K-13.315.437.075.7-35.925.654.665.2
Grounded 3D-LLM [7]107K-13.4--72.7-35.5--70.6
Chat-Scene [19]38K-14.3--87.7-36.3--77.1
Inst3D-LMM38K43.514.918.442.688.624.638.327.557.279.7
+ +Table 2. Quantitative results for 3D Question Answering and 3D Dense Captioning on the ScanQA and Scan2Cap datasets. + +process utilizes LoRA [17]. We adopt the AdamW optimizer with a weight decay of 0.02. All experiments were performed on 8 NVIDIA A100 GPUs. + +# 4.2. Main Results + +3D Visual Grounding. We first report visual grounding performance on the ScanRefer and Multi3DRefer validation datasets. As shown in Table 1, our approach outperforms the state-of-the-art model, Chat-Scene [19], by $+2.3\%$ Acc@0.25 and $+1.2\%$ F1@0.25 on ScanRefer and + +Multi3DRefer, respectively. Compared to specialist and closed-set methods, Inst3D-LMM, trained with a generalist approach, achieves competitive performance. Figure 5 displays typical visual comparison results. + +3D Question Answering. We then compare Inst3D-LMM with previous leading methods on the ScanQA validation set. Table 2 reports the results. Our Inst3D-LMM consistently outperforms these methods, including recent LL3DA [6], Grounded 3D-LLM [7] and Chat-Scene [19]. + +3D Dense Captioning. This task involves the localization + +![](images/e79d7be4241944c89c7fd56282d186f51e719e79cb1117808871a239f184654c.jpg) +7-6 + +![](images/d2ffa5e1d1ccf3d270cc18ec24aa88e6cefb988e6d4cbb05a12f18bf4bb49592.jpg) +Figure 6. A visualization of the similarity score between text query and segmented 3D proposals. We compare sentence tokens used in ViL3DRel [5] with position embeddings employed in 3D-ISR. + +Figure 5. Visual comparisons in 3D Visual Grounding. The rendered images show the ground-truth (blue), incorrectly identified objects (red), and correctly identified objects (green). The colored text indicates the results of text decoupling. +![](images/da796e3ac14e3d8d8653a051d01f783d1382f17972b86e7869d1386965f8c17f.jpg) +Text Query: The kitchen cabinets are under the sink. They are the furthest left cabinets and to the left of the dishwasher. + +![](images/2ac875c81967689bd998ab13e86d61467bee8f80b724dda833c9354f04e9a5a1.jpg) + +and description of instances. As shown in Table 2, our approach achieves $38.3\%$ B-4@0.50 and $79.7\%$ C@0.50 on Scan2Cap, which exceeds the closed-set expert model 3D-VisTA [60] by $+4.3\%$ and $+12.8\%$ , and still outperforms the LLM-based method LL3DA by $+2.4\%$ and $+14.5\%$ . + +# 4.3. Ablation Study + +In this section, we perform ablation experiments to thoroughly evaluate the effectiveness of the components. Please refer to the Supplementary Material for more analysis. + +Effects of Multi-view Cross-Modal Fusion (MCMF). The MCMF module aims to augment LLM's understanding of 3D instances' geometric and semantic attributes. We explore several methods to combine these features, including direct concatenation for projection ('Concat.'), parallel projection followed by combination ('Parallel'), and vanilla cross-attention ('Cross-Attention'). To encode basic spatial information, we concatenate the results of absolution positional encoding (PE) with 3D features. As shown in Table 3, MCMF outperforms all these methods on ScanRefer, ScanQA and Scan2Cap datasets. We also observe that the performance of 3D LMM is significantly improved by incorporating multi-view 2D CLIP information compared to using only 3D geometric features. These results demonstrate the effectiveness of our MCMF approach. + +
MethodScanReferScanQAScan2Cap@0.50
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑B-4↑CIDER↑
w/o Multi-view 2D36.031.931.557.319.954.3
Concat.38.834.436.665.425.263.3
Parallel37.633.535.363.924.862.1
Cross-Attention39.236.537.466.125.564.3
MCMF46.741.941.578.632.768.2
+ +Impacts of 3D Instance Spatial Relation (3D-ISR). We further investigate the 3D-ISR module to analyze its impact + +Input 3D Scene +![](images/84619ebb3073cac7a658d3902b7cec84c0c8913562255e76a605b7354176e437.jpg) +Text Query: There is a brown reclined in the corner of the room. The adjacent walls have two green boards. It is between the intersection of the 2 boards. + +![](images/82cda2238fe073390724c2790d2b6d6ad084d0bee731b4dca4601284be186bb9.jpg) +Sentence Token + +![](images/5126e0323f8535c82d39a60f31c68c2da22c225d69108025743ed435d1d262fc.jpg) +Position Embedding +A + +on 3D spatial understanding using only 3D geometry priors. We compare 3D-ISR with previous spatial relation modeling methods, such as the 3D localization mechanism in ViL3DRel [5] and 3D-LLM [16], and relation-aware token generation in Chat-3D-v2 [18]. Table 4 reports the results of the comparison. It can be observed that our 3D-ISR consistently surpasses these approaches. Specifically, 3D-ISR significantly enhances performance on grounding tasks (ScanRefer and Multi3dRefer). This aligns with the motivation of our design. We also provide a visualization, as shown in Figure 6, of the similarity between the text query and segmented 3D proposals after processing by LLM. Compared to ViL3DRel [5] using sentence tokens, our 3D-ISR more accurately captures spatial relationships within the scene employing position embeddings of objects. + +Table 3. Ablation evaluations of the proposed Multi-view CrossModal Fusion (MCMF) module. + +
MethodScanReferMulti3DReferScanQA
Acc@0.25↑Acc@0.50↑F1@0.25↑F1@0.50↑B-1↑CIDER↑
w/o spatial relation35.629.834.827.329.255.6
ViL3DRel [5]36.130.538.435.532.858.9
3D-LLM [16]39.236.842.435.232.560.2
Chat-3D-v2 [18]40.837.541.637.835.164.6
3D-ISR48.344.146.241.339.172.3
+ +Table 4. Ablation study to verify the effectiveness of our 3D Instance Spatial Relation (3D-ISR) module. + +Ablations within MCMF and 3D-ISR modules. We also explore the effectiveness of key designs in MCMF and 3D-ISR modules. The learnable [CLS] token in the MCMF module is introduced to aggregate 2D multi-view features. We compare it with alternative methods, such as token max pooling and Q-Former [24]. Table 5 reports the comparison results, which demonstrate the efficacy of the learnable [CLS] token design. Besides, to assess the impact of different pairwise spatial features in the 3D-ISR module, we evaluate the model using only distance or orientation information to compute spatial features in Eq 4. The results in Table 6 reveal that distance has a greater impact on the grounding capability of the model, while orientation is more crucial in handling Q&A tasks. Combining both pairwise distance and orientation yields the best overall performance. Mutual Benefits of MCMF and 3D-ISR. As shown in Table 7, the model that integrates MCMF and 3D-ISR consistently outperforms those that use either module alone for multiple tasks. To further verify its effectiveness, we combine MCMF with absolution positional encoding (PE) for the basic spatial relation, and directly adopt feature con + +![](images/b61b337457c3526896f0ddfb41b73ea6b1b582b6ad061616390b2bd3d6b24cf4.jpg) +Figure 7. Qualitative illustration of Inst3D-LMM across various 3D-language tasks in diverse 3D environments. + +
MethodScanReferScanQAScan2Cap@0.50
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑B-4↑CIDER↑
Max Pooling48.744.639.274.830.665.5
Q-Former [24]53.247.944.387.935.478.6
CLS Token57.851.643.588.638.379.7
+ +Table 5. Ablation study of learnable CLS token in MCMF module. + +
MethodScanReferMulti3DReferScanQA
Acc@0.25↑Acc@0.50↑F1@0.25↑F1@0.50↑B-1↑CIDER↑
w/ Dist (only)43.238.042.437.831.265.7
w/ Ort (only)39.733.640.535.932.566.8
Dist + Ort48.344.146.241.339.172.3
+ +Table 6. Ablation study of 3D-ISR module. Dist (Ort) means only using distance (orientation) to compute pairwise spatial features. + +catenation (Concat.) to integrate 3D object features and 2D multi-view features along with 3D-ISR. These results indicate the synergy efforts between MCMF and 3D-ISR. + +Different LLMs and Foundation Models. Table 8 presents the results of various LLMs and pre-trained 2D/3D models, including Vicuna-7B [10] vs Vicuna-13B, Ulip2 [48] vs Uni3D [57], CLIP [35] vs SigLIP [32], and ViT-H-based SAM [23] vs ViT-L-based SAM. We find that the performance of grounding and reasoning increases along with the total number of parameters in the foundation models (i.e. ViT-L vs Vit-H and 7B vs 13B). These results indicate that our framework's capability can be improved in tandem with the performance of foundation models. + +# 5. Conclusion + +In this paper, we proposed Inst3D-LMM, an effective instance-aware framework to leverage the potential of Large Multi-modal Models (LMMs) for 3D scene understanding. To improve instance-level representations, we developed + +
MethodScanReferScanQAScan2Cap@0.50
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑B-4↑CIDER↑
w/ MCMF (only)38.234.639.476.830.566.9
w/ 3D-ISR (only)48.344.139.172.329.864.3
MCMF + PE46.741.941.578.632.768.2
3D-ISR + Concat.49.845.040.675.433.366.5
MCMF + 3D-ISR (full)57.851.643.588.638.379.7
+ +Table 7. Comparison results of the collaboration between MCMF and 3D-ISR modules. + +
LLM3D Encoder2D VFsScanReferScanQA
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑
Vicuna-7BUlip2CLIP+SAM (ViT-L)49.244.839.579.6
Vicuna-7BUlip2CLIP+SAM (ViT-H)53.646.543.787.5
Vicuna-7BUlip2SigLIP+SAM (ViT-H)54.347.042.284.5
Vicuna-7BUni3DCLIP+SAM (ViT-H)57.851.643.588.6
Vicuna-13BUni3DCLIP+SAM (ViT-H)56.052.042.883.1
Vicuna-13BUni3DSigLIP+SAM (ViT-H)55.247.843.385.4
+ +Table 8. Ablation study of different LLMs and pre-trained foundation models. + +a novel Multi-view Cross-Modal Fusion (MCMF) module, which injects multi-view 2D semantic open-vocabulary priors into 3D geometry features to generate fine-grained instance-level tokens. Furthermore, we introduced a 3D Instance Spatial Relation (3D-ISR) module that employs the spatial condition attention mechanism to capture pairwise spatial relations. Experimental results demonstrate that our approach achieves promising performance in understanding and reasoning across various 3D vision-language tasks. + +# Acknowledgments + +This work is supported by National Natural Science Foundation of China under Grants (62376244). It is also supported by Information Technology Center and State Key Lab of CAD&CG, Zhejiang University. + +# References + +[1] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, pages 19129-19139, 2022. 5, 6 +[2] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In CVPR, pages 16464-16473, 2022. 2 +[3] Yang Cao, Zeng Yihan, Hang Xu, and Dan Xu. Coda: Collaborative novel box discovery and cross-modal alignment for open-vocabulary 3d object detection. In NeurIPS, pages 71862-71873, 2024. 4 +[4] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221. Springer, 2020. 5, 6 +[5] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. In NeurIPS, pages 20522-20535, 2022. 2, 4, 5, 7 +[6] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In CVPR, pages 26428-26438, 2024. 1, 2, 3, 6 +[7] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 3, 6 +[8] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In CVPR, pages 3193-3203, 2021. 2, 5, 6 +[9] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In ICCV, pages 18109-18119, 2023. 2 +[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 5, 8 +[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 2432–2443, 2017. 5 +[12] Zhipeng Ding, Xu Han, and Marc Niethammer. Volenet: A deep learning label fusion method for multi-atlas segmentation. In MICCAI, pages 202-210, 2019. 6 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5 +[14] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: + +Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2 +[15] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905, 2023. 2 +[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. In NeurIPS, pages 20482-20494, 2023. 1, 2, 3, 6, 7 +[17] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6 +[18] Haifeng Huang, Zehan Wang, Rongjie Huang, Luping Liu, Xize Cheng, Yang Zhao, Tao Jin, and Zhou Zhao. Chat-3d v2: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023. 3, 6, 7 +[19] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. NeuIPS, 2024. 1, 2, 6 +[20] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 2 +[21] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. In ECCV, pages 528-545. Springer, 2022. 2 +[22] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In CVPR, pages 10984-10994, 2023. 2 +[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3, 4, 5, 8 +[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 7, 8 +[25] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 4 +[26] Zeju Li, Chao Zhang, Xiaoyan Wang, Ruilong Ren, Yifan Xu, Ruifei Ma, and Xiangde Liu. 3dmit: 3d multi-modal instruction tuning for scene understanding. In ICMEW, 2024. 6 +[27] Cai Liang, Bo Li, Zhengming Zhou, Longlong Wang, Pengfei He, Liang Hu, and Haoxing Wang. Spatioawaregrouding3d: A spatio aware model for improving 3d vision grousing. In CVPRW, 2024. 4 + +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, pages 34892-34916, 2023. 6 +[29] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In NeurIPS, pages 44860-44879, 2023. 2 +[30] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. In ICLR, 2023. 1, 2 +[31] Yunze Man, Shuhong Zheng, Zhipeng Bao, Martial Hebert, Liang-Yan Gui, and Yu-Xiong Wang. Lexicon3d: Probing visual foundation models for complex 3d scene understanding. arXiv preprint arXiv:2409.03757, 2024. 2 +[32] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, pages 529-544. Springer, 2022. 8 +[33] Maria Parelli, Alexandros Delitzas, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Clip-guided vision-language pre-training for question answering in 3d scenes. In CVPR, pages 5606-5611, 2023. 2 +[34] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In CVPR, pages 815-824, 2023. 6 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 3, 4, 5, 8 +[36] Krishan Rana, Jesse Haviland, Sourav Garg, Jad Abou-Chakra, Ian Reid, and Niko Suenderhauf. Sayplan: Grounding large language models using 3d scene graphs for scalable robot task planning. In CoRL, pages 23-72, 2023. 1 +[37] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, pages 125-141. Springer, 2022. 5 +[38] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In ICRA, pages 8216-8223, 2023. 3, 4, 5 +[39] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M. Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In ICCV, pages 2998-3009, 2023. 1 +[40] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In NeurIPS, pages 68367-68390, 2023. 4 +[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5 +[42] Ozan Unal, Christos Sakaridis, Suman Saha, Fisher Yu, and Luc Van Gool. Three ways to improve verbo-visual + +fusion for dense 3d visual grounding. arXiv preprint arXiv:2309.04561, 2023. 6 +[43] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, et al. Embodiedscan: A holistic multimodal 3d perception suite towards embodied ai. In CVPR, pages 19757-19767, 2024. 1, 3 +[44] Zehan Wang, Haifeng Huang, Yang Zhao, Linjun Li, Xize Cheng, Yichen Zhu, Aoxiong Yin, and Zhou Zhao. 3drp-net: 3d relative position-aware network for 3d visual grounding. In EMNLP, page 10612-10625, 2023. 2 +[45] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 1, 3 +[46] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In CVPR, pages 19231-19242, 2023. 1 +[47] Runsen Xu, Zhiwei Huang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Vlm-grounder: A vlm agent for zero-shot 3d visual grounding. arXiv preprint arXiv:2410.13860, 2024. 2 +[48] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In CVPR, pages 27091–27101, 2024. 5, 8 +[49] Fan Yang, Sicheng Zhao, Yanhao Zhang, Haoxiang Chen, Hui Chen, Wenbo Tang, Haonan Lu, Pengfei Xu, Zhenyu Yang, Jungong Han, et al. Lmi3d: Empowering llm with 3d perception from a single 2d image. arXiv preprint arXiv:2408.07422, 2024. 2 +[50] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. In ICRA, pages 7694-7701, 2024. 6 +[51] Li Yang, Ziqi Zhang, Zhongang Qi, Yan Xu, Wei Liu, Ying Shan, Bing Li, Weiping Yang, Peng Li, Yan Wang, et al. Exploiting contextual objects and relations for 3d visual grounding. In NeurIPS, pages 49542-49554, 2024. 6 +[52] Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Xiaoshui Huang, Zhiyong Wang, Lu Sheng, Lei Bai, et al. Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. In NeurIPS, pages 26650-26685, 2023. 6 +[53] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In CVPR, pages 6281–6290, 2019. 6 +[54] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3d visual grounding. In CVPR, pages 20623-20633, 2024. 6 +[55] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, pages 15225-15236, 2023. 5, 6 + +[56] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2928-2937, 2021. 6 +[57] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. In ICLR, 2024. 2, 3, 4, 5, 8 +[58] Chenming Zhu, Tai Wang, Wenwei Zhang, Kai Chen, and Xihui Liu. Empowering 3d visual grounding with reasoning capabilities. arXiv preprint arXiv:2407.01525, 2024. 1, 3, 6 +[59] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmms with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 2 +[60] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 2, 4, 6, 7 \ No newline at end of file diff --git a/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/images.zip b/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..92b9848a5100d85f3f2ffad7d7a410b862dcd825 --- /dev/null +++ b/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d450fe371702f6db079438faeb4002d6a558fdaf38ae42ab39acb669da8a96b9 +size 1002661 diff --git a/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/layout.json b/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..505a28096fc7dddbd421f92025fa38071d64eeda --- /dev/null +++ b/2025/Inst3D-LMM_ Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning/layout.json @@ -0,0 +1,8486 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 103, + 479, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 103, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 133, + 103, + 479, + 140 + ], + "type": "text", + "content": "Inst3D-LMM: Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "spans": [ + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "text", + "content": "Hanxun Yu" + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "text", + "content": ", Wentong Li" + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "text", + "content": ", Song Wang" + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "text", + "content": ", Junbo Chen" + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "text", + "content": ", Jianke Zhu" + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "text", + "content": "Nanjing University of Aeronautics and Astronautics " + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 94, + 160, + 520, + 208 + ], + "type": "text", + "content": "Udeer.ai {hanxun.yu, songw, jkzhu}@zju.edu.cn, wentong.li@nuaa.edu.cn, junbo@udeer.ai" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 233, + 201, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 233, + 201, + 247 + ], + "spans": [ + { + "bbox": [ + 151, + 233, + 201, + 247 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 259, + 297, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 297, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 297, + 571 + ], + "type": "text", + "content": "Despite encouraging progress in 3D scene understanding, it remains challenging to develop an effective Large Multi-modal Model (LMM) that is capable of understanding and reasoning in complex 3D environments. Most previous methods typically encode 3D point and 2D image features separately, neglecting interactions between 2D semantics and 3D object properties, as well as the spatial relationships within the 3D environment. This limitation not only hinders comprehensive representations of 3D scene, but also compromises training and inference efficiency. To address these challenges, we propose a unified Instance-aware 3D Large Multi-modal Model (Inst3D-LMM) to deal with multiple 3D scene understanding tasks simultaneously. To obtain the fine-grained instance-level visual tokens, we first introduce a novel Multi-view Cross-Modal Fusion (MCMF) module to inject the multi-view 2D semantics into their corresponding 3D geometric features. For scene-level relation-aware tokens, we further present a 3D Instance Spatial Relation (3D-ISR) module to capture the intricate pairwise spatial relationships among objects. Additionally, we perform end-to-end multi-task instruction tuning simultaneously without the subsequent task-specific fine-tuning. Extensive experiments demonstrate that our approach outperforms the state-of-the-art methods across 3D scene understanding, reasoning and grounding tasks. Source code is available at: https://github.com/hanxunyu/Inst3D-LMM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 605, + 137, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 605, + 137, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 605, + 137, + 617 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 687 + ], + "type": "text", + "content": "Building Large Multi-modal Models (LMMs) for 3D scene understanding becomes an emerging research topic with significant potential for advancing autonomous robotics [36]. For example, interactive embodied agents [39] are expected to interpret 3D layouts and predict" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 235, + 499, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 499, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 499, + 246 + ], + "type": "text", + "content": "object locations based on human instructions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 247, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 534 + ], + "type": "text", + "content": "Traditional 3D scene understanding methods [30, 46] are typically tailored for individual downstream tasks, such as 3D Visual Grounding (3D-VG), 3D Question Answering (3D-QA) and 3D Dense Captioning (3D-DC). In contrast, LMMs are able to handle various 3D perception tasks within a single model. Some methods [16, 43] primarily focus on translating 3D points into the space of 2D Vision Language Models (VLMs) or directly leveraging multiview 2D features as 3D representations. Alternatively, other approaches [6, 45, 58] directly encode the features of 3D points and facilitate alignment with LLM using 3D-text instruction data. However, they often require multi-stage alignment or language-scene pre-training, complicating the development of a versatile model capable of handling multiple tasks. To enable a unified 3D LMM framework, recent work [19] decomposes the input 3D scene into a set of individual object proposals, each identified by unique tokens to explicitly capture 3D object features at the instance level. While this approach exhibits promising results, it neglects the interactions between the 2D semantic features and the properties of 3D objects, as well as the spatial relationship modeling among objects in 3D environments. This oversight further results in substantial token costs for the LLM, thus hindering both training and inference efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we propose Inst3D-LMM, an effective Instance-aware 3D Large Multi-modal Model that tackles multiple 3D-language tasks without resorting to task-specific fine-tuning. Our approach fully leverages the powerful 2D Vision Foundation Models (VFMs) and 3D specialist models to extract enriched 2D and 3D features at the instance level, respectively. As shown in Figure 1, in contrast to previous methods, our approach is able to generate fine-grained instance-level representations that encapsulate both geometric and semantic properties, and scene-level representations that capture intricate pairwise spatial relationships among objects in a 3D scene. Moreover, our method results in minor token costs for LLM, thereby enhancing both training and inference efficiency. By leveraging this instance-aware methodology, our Inst3D-LMM" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 149, + 713 + ], + "type": "text", + "content": "*Equal contribution. † Corresponding authors." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14147" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 378, + 220 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 378, + 220 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 378, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 378, + 220 + ], + "type": "image", + "image_path": "8fefee1109f574bdeac6eea50d061c38485d98b8381a99f38aa3e2f0b78da53d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 226, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 54, + 226, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 226, + 555, + 281 + ], + "type": "text", + "content": "Figure 1. Comparisons between previous 3D LMMs and our proposed Inst3D-LMM. (a) Previous methods [6, 16, 19, 20] typically encode the features of 3D points or 2D images separately and concatenate them directly, often requiring task-specific fine-tuning for different tasks. (b) Illustration of Inst3D-LMM. Our method integrates 2D/3D cross-modal information and captures the intricate spatial relations among objects within 3D environments to generate tight but informative instance/scene-level tokens for the LMM. (c) Compared with other 3D LMMs, our Inst3D-LMM requires fewer computational resources, while offering faster training and inference speeds." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 379, + 71, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 379, + 71, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 379, + 71, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 379, + 71, + 553, + 220 + ], + "type": "image", + "image_path": "59392d37dc948e2cecfddbeef78a690e045fc039d2a2b20cc70ad0b4e050f104.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 293, + 295, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 295, + 317 + ], + "type": "text", + "content": "significantly improves the LLM's ability to comprehend 3D scenes with respect to both efficiency and accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 318, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 318, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 295, + 484 + ], + "type": "text", + "content": "Specifically, we introduce a novel Multi-view CrossModal Fusion (MCMF) module that effectively infuses enriched multi-view 2D features into the original 3D features with coarse semantics. A learnable [CLS] token is introduced to aggregate the characteristics of each 2D view, enabling efficient multi-view 2D-to-3D cross-modal transformation. To capture intricate pairwise spatial relationships among objects in a 3D scene, we then propose a 3D Instance Spatial Relation (3D-ISR) module. A spatial condition self-attention between manifold position embeddings and instance-level tokens is presented to produce relation-aware scene-level representations. The resulting instance-level and scene-level representations are subsequently fed into the LLM for end-to-end multi-task instruction tuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 604 + ], + "type": "text", + "content": "Extensive experiments across various tasks, including 3D-VG, 3D-QA and 3D-DC, demonstrate that our approach outperforms previous state-of-the-art methods with leading 3D scene understanding, grounding and reasoning capabilities. Unlike most existing methods that focus on close-set scene understanding or require per-task fine-tuning, our Inst3D-LMM operates as a generalist model. We believe this work lays a fundamental step towards unifying diverse 3D vision-language tasks in generative language modeling. To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": "- We propose a unified and efficient instance-aware LLM-based framework, called Inst3D-LMM for various 3D scene understanding tasks with end-to-end multi-modal instruction tuning. Serving as a generalist model, our approach demonstrates superior performance across 3D scene understanding, reasoning and spatial localization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "- We utilize 2D VFMs to extract multi-view contextual features for each 3D instance and then devise a Multi-view Cross-Modal Fusion (MCMF) module to effectively en" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 293, + 553, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 293, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 321, + 293, + 553, + 317 + ], + "type": "text", + "content": "hance instance-level feature representations by jointly integrating 3D geometry and 2D semantic priors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 318, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 318, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 314, + 318, + 553, + 354 + ], + "type": "text", + "content": "- A 3D Instance Spatial Relation (3D-ISR) module is introduced to boost the capability of LMM in understanding the complex spatial details within 3D scenes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 368, + 405, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 368, + 405, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 368, + 405, + 380 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 389, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 616 + ], + "type": "text", + "content": "3D Scene Understanding with Language. In 3D scene understanding, there is a surge of interest in making use of language queries to capture user intentions for various downstream tasks, such as 3D Visual Grounding [5, 44], 3D Question Answering [30, 33] and 3D Dense Captioning [8, 21]. Specifically, 3D Visual Grounding entails localizing target objects based on language queries. Moreover, 3D Question Answering demands robust 3D spatial perception and reasoning. 3D Dense Captioning involves localizing and describing objects in 3D scenes. Conventional methods typically focus on a specific task. Instead, 3D visual grounding and dense captioning tasks are combined by leveraging their complementary aspects [2, 9, 31, 47]. Recent efforts like 3D-VLP [22] and 3D-VisTA [60] attempt to establish a universal framework by pre-aligning 3D scenes with their corresponding textual descriptions. In contrast to our Inst3D-LMM, most existing methods still focus on close-set scene understanding, which requires either task-specific fine-tuning or striving to build specialized models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "3D Large Multi-modal Models. Inspired by the significant advancements in Large Language Models (LLMs), researchers extend LLM's knowledge to encompass 3D modality [29, 49, 57, 59]. Point-LLM [14] and Imagebind-LLM [15] have succeeded in bridging the gap between 3D visuals and text by utilizing extensive 3D object datasets. However, these models struggle with interpreting complex spatial relationships in 3D scenes. Another promising di" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14148" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 69, + 553, + 373 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 553, + 373 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 553, + 373 + ], + "type": "image", + "image_path": "fd36b538b11330d06979505d777b82f51104af180a9302c9d3e26b2ef8929019.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 380, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 555, + 436 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed Inst3D-LMM. Our pipeline takes as input point clouds of a 3D indoor scene, along with RGB-D images. We first employ the pre-trained 3D specialist models and 2D VFMs to extract 3D proposals and multi-view 2D semantic features, respectively. We then suggest the MCMF module to generate fine-grained instance-level tokens. A 3D-ISR module is further introduced to create relation-aware scene-level tokens based on spatial distances. By leveraging the constructed 3D-language prompts, we conduct multi-task instruction tuning to simultaneously handle various 3D tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 448, + 297, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 448, + 297, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 297, + 653 + ], + "type": "text", + "content": "rection focuses on developing the scene-level 3D LMMs. Hong et al. [16] encode projected 3D features using a 2D vision encoder and incorporate location tokens to augment LLM vocabulary. Other methods [6, 45, 58] directly encode point clouds and utilize 3D scene-text data for better visual interaction through pre-alignment. Huang et al. [18] employ a three-stage training scheme and adopt object identifiers to learn individual object attributes. Chen et al. [7] use special referent tokens for precise referencing and grounding. Wang et al. [43] encode point clouds and RGB images separately. However, some crucial semantic pixels are often lost during their sparse fusion processes, resulting in a coarse visual semantic representation. In this work, we propose an effective instance-aware framework to fuse fine-grained cross-modal information and encode spatial relations, which achieves promising results on multiple 3D-language tasks as a generalist model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 667, + 138, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 667, + 138, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 667, + 138, + 681 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Our goal is to enable LLM to understand the 3D environment and perform various visual interaction tasks based on" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 448, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 664 + ], + "type": "text", + "content": "human instructions. Figure 2 illustrates the architecture of our framework. In this section, we first introduce how to extract features at the instance level using pre-trained 3D models [38, 57] and 2D VFs [23, 35], respectively. Secondly, we present a novel Multi-view Cross-Modal Fusion (MCMF) module to obtain fine-grained instance-level tokens, which are specially designed to effectively integrate 3D geometric features with their corresponding multi-view 2D semantic features. Thirdly, we introduce the 3D Instance Spatial Relation (3D-ISR) module to enhance LLM's ability to capture spatial information at the scene level, which generates relation-aware tokens through attention-based analysis of spatial relationships among different 3D proposals. Under our proposed framework, the MCMF and 3D-ISR modules are jointly optimized, allowing mutual enhancement. Finally, we conduct end-to-end multi-task instruction tuning to address a range of 3D scene understanding tasks simultaneously." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 671, + 497, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 497, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 497, + 682 + ], + "type": "text", + "content": "3.1. Instance-Level Feature Extraction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "3D Feature Extraction. We first segment the 3D point clouds of each individual instance in a class-agnostic man" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14149" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "ner by leveraging a pre-trained 3D instance segmentation model [38]. We only retain the predicted binary 3D instance masks while ignoring their closed-vocabulary class labels. All instance proposals in one scene are represented by " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "O^{3D} = (O_1^{3D}, \\ldots, O_N^{3D})" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "O_i^{3D} = [coordinate, color]" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " consists of the attributes of each instance. We obtain all instance proposals of a scene, and the pre-trained 3D encoder " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{o}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " [57] is used to extract their instance-level features, i.e., " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{o}}^{3D} = \\mathbf{E}_{\\mathbf{o}}(O^{3D})" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 295, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 295, + 395 + ], + "type": "text", + "content": "2D Feature Extraction. Due to the inherent sparsity of 3D point cloud data, previous methods have difficulties in generating discriminative features of each object. In this work, we employ powerful 2D VFMs to extract 2D semantic features for each 3D instance. As in [3, 40], we firstly project the point cloud of each instance " + }, + { + "bbox": [ + 55, + 180, + 295, + 395 + ], + "type": "inline_equation", + "content": "O_{i}^{3D}" + }, + { + "bbox": [ + 55, + 180, + 295, + 395 + ], + "type": "text", + "content": " onto the image plane and then select the top " + }, + { + "bbox": [ + 55, + 180, + 295, + 395 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 180, + 295, + 395 + ], + "type": "text", + "content": " views according to the number of visible points. To obtain accurate 2D masks, we randomly sample " + }, + { + "bbox": [ + 55, + 180, + 295, + 395 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{sample}} = 5" + }, + { + "bbox": [ + 55, + 180, + 295, + 395 + ], + "type": "text", + "content": " points as input prompts for SAM [23] to deal with noisy bounding boxes with outliers. Hereby, we select the high-quality mask with the highest confidence score. To enrich features with contextual information, these masks are employed to generate multi-level " + }, + { + "bbox": [ + 55, + 180, + 295, + 395 + ], + "type": "inline_equation", + "content": "(L)" + }, + { + "bbox": [ + 55, + 180, + 295, + 395 + ], + "type": "text", + "content": " crops of the selected images that are further fed into the pre-trained CLIP vision encoder [35] to extract features with language-aligned embedding space. Finally, we aggregate multi-level features of sub-images from the same frame to form the 2D multi-view features " + }, + { + "bbox": [ + 55, + 180, + 295, + 395 + ], + "type": "inline_equation", + "content": "O^{2D}" + }, + { + "bbox": [ + 55, + 180, + 295, + 395 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 403, + 227, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 227, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 227, + 415 + ], + "type": "text", + "content": "3.2. Multi-view Cross-Modal Fusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "text", + "content": "To better fuse 3D geometry priors with 2D multi-view semantic priors, we introduce an effective Multi-view CrossModal Fusion (MCMF) module that generates the enriched token representations for each 3D instance before being fed into LLM. The architecture of MCMF is crafted with a coarse-to-fine framework, as shown in Figure 3. To map 3D object features " + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "inline_equation", + "content": "O^{3D} \\in \\mathbb{R}^{1 \\times N \\times D^{3d}}" + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "text", + "content": " and 2D CLIP features " + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "inline_equation", + "content": "O^{2D} \\in \\mathbb{R}^{K \\times N \\times D^{2d}}" + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "text", + "content": " into the embedding space of LLM with the dimensionality of " + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "text", + "content": ", we utilize a simple two-layer MLP with a LayerNorm and GELU in between, yielding " + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "inline_equation", + "content": "O^{3D'}" + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "inline_equation", + "content": "O^{2D'}" + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "text", + "content": " represents the number of instances, and " + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 422, + 296, + 568 + ], + "type": "text", + "content": " is the number of images for each object." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "type": "text", + "content": "Subsequently, we introduce a Cross-Modal Injection Block to transform the enriched semantic priors from 2D multi-view representations into 3D instance features. Moreover, we adopt a straightforward self-attention layer to further enhance such 3D instance features. For 2D multi-view features, we append a learnable [CLS] token " + }, + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "type": "text", + "content": " to the flattened CLIP feature maps in order to adaptively encapsulate the global semantic representation of the " + }, + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "type": "text", + "content": "-th view. Then, we apply a self-attention layer to multi-view features, enabling semantic gathering to derive " + }, + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "type": "inline_equation", + "content": "t_k'" + }, + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "type": "text", + "content": "-th view as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 699, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 699, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 110, + 699, + 295, + 715 + ], + "type": "interline_equation", + "content": "t _ {k} ^ {\\prime} = \\phi (\\text {S e l f A t t n} ([ t _ {k}, O _ {\\text {v i e w} _ {k}} ^ {2 D ^ {\\prime}} ])), \\tag {1}", + "image_path": "32ff06e40e47e1eb0b04ee323cbfd8c5639afdf5721f9090d93fd94b0dc37b09.jpg" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 71, + 553, + 242 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 553, + 242 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 553, + 242 + ], + "type": "image", + "image_path": "e26d593e3821672fd474eed8c2f89be47e2793c7ee27458b3983d6ffe30c7784.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 251, + 553, + 273 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 273 + ], + "type": "text", + "content": "Figure 3. Architecture of the proposed Multi-view Cross-Modal Fusion (MCMF) module." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 277, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 314 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 277, + 553, + 314 + ], + "type": "inline_equation", + "content": "1 \\leq k \\leq K" + }, + { + "bbox": [ + 313, + 277, + 553, + 314 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 277, + 553, + 314 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{view}_k}^{2D'} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 313, + 277, + 553, + 314 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 277, + 553, + 314 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 277, + 553, + 314 + ], + "type": "text", + "content": " is an indicator function that outputs the updated [CLS] token as the first one from the token list." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 314, + 553, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 363 + ], + "type": "text", + "content": "We obtain the processed 3D token embedding " + }, + { + "bbox": [ + 313, + 314, + 553, + 363 + ], + "type": "inline_equation", + "content": "O^{3D'} \\in \\mathbb{R}^{1 \\times N \\times D}" + }, + { + "bbox": [ + 313, + 314, + 553, + 363 + ], + "type": "text", + "content": " containing coarse semantics, where each instance in " + }, + { + "bbox": [ + 313, + 314, + 553, + 363 + ], + "type": "inline_equation", + "content": "O^{3D'}" + }, + { + "bbox": [ + 313, + 314, + 553, + 363 + ], + "type": "text", + "content": " corresponds to multiple views of the processed 2D CLIP features " + }, + { + "bbox": [ + 313, + 314, + 553, + 363 + ], + "type": "inline_equation", + "content": "O^{2D'}" + }, + { + "bbox": [ + 313, + 314, + 553, + 363 + ], + "type": "text", + "content": ". It can be represented as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 361, + 370, + 553, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 370, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 361, + 370, + 553, + 386 + ], + "type": "interline_equation", + "content": "O _ {i} ^ {2 D ^ {\\prime}} = \\left[ t _ {1} ^ {\\prime}, \\dots , t _ {K} ^ {\\prime} \\right], \\quad i \\in [ 1, N ]. \\tag {2}", + "image_path": "2a25715a652d28c666a8a34e72a6649702b43b1b35943a609e12a0e7a4a4aa21.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "text", + "content": "To infuse the enriched semantic 2D features into 3D geometric feature, we construct the 3D-to-2D-Multi-view pairs, i.e. each 3D instance feature in " + }, + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "inline_equation", + "content": "O^{3D'} \\in \\mathbb{R}^{1 \\times N \\times D}" + }, + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "text", + "content": " corresponds to 2D features across " + }, + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "text", + "content": " views in " + }, + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "inline_equation", + "content": "O^{2D'} \\in \\mathbb{R}^{K \\times N \\times D}" + }, + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "text", + "content": ". In particular, we utilize the preliminary " + }, + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "inline_equation", + "content": "O^{3D'}" + }, + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "text", + "content": " as queries, while the 2D multi-view visual features " + }, + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "inline_equation", + "content": "O^{2D'}" + }, + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "text", + "content": " as enriched reference keys and values. Inspired by [25], the injection process is conducted via a cross-attention layer, which encourages 3D instance queries to absorb the fine-grained semantics of keys and values. This results in the enhanced 3D instance features " + }, + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "inline_equation", + "content": "O_f^{3D} \\in \\mathbb{R}^{1 \\times N \\times D}" + }, + { + "bbox": [ + 313, + 393, + 554, + 526 + ], + "type": "text", + "content": " as below:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 366, + 533, + 553, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 533, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 366, + 533, + 553, + 550 + ], + "type": "interline_equation", + "content": "O _ {f} ^ {3 D} = \\operatorname {C r o s s A t t n} \\left(O ^ {3 D ^ {\\prime}}, O ^ {2 D ^ {\\prime}}\\right). \\tag {3}", + "image_path": "638111007749a8ffc7dae0ee74e1bd73fd13bea63d926a382a91f8b7e144905e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 556, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 553, + 594 + ], + "type": "text", + "content": "Furthermore, we leverage a residual operation by adding the input " + }, + { + "bbox": [ + 313, + 556, + 553, + 594 + ], + "type": "inline_equation", + "content": "O^{3D'}" + }, + { + "bbox": [ + 313, + 556, + 553, + 594 + ], + "type": "text", + "content": " to retain the basic characteristics in generating the instance level 3D visual tokens, " + }, + { + "bbox": [ + 313, + 556, + 553, + 594 + ], + "type": "inline_equation", + "content": "O_{I}^{3D} = O_{f}^{3D} + O^{3D'}" + }, + { + "bbox": [ + 313, + 556, + 553, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 599, + 471, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 471, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 471, + 612 + ], + "type": "text", + "content": "3.3. 3D Instance Spatial Relation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Motivated by [5, 27, 60], we develop an effective 3D Instance Spatial Relation (3D-ISR) module to boost LLM's capabilities in assimilating spatial information within the 3D scene, as shown in Figure 4. 3D-ISR utilizes the instance-level 3D visual tokens " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\{O_{I1}^{3D}, O_{I2}^{3D}, \\ldots, O_{IN}^{3D}\\}" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " derived from MCMF module along with the corresponding center coordinates " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\{C_1, C_2, \\ldots, C_N\\}" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " of all instances as input. For the " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "-th instance, we define its center coordinates" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14150" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 207 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 207 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 207 + ], + "type": "image", + "image_path": "1db6d1c3d65a4d75503fbbfd2fd99a7a85fcd5a99457318ba57bd6bfdd3874da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 216, + 295, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 216, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 295, + 239 + ], + "type": "text", + "content": "Figure 4. Illustration of the 3D Instance Spatial Relation (3D-ISR) module in our framework." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "inline_equation", + "content": "C_i = (x_i, y_i, z_i)" + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "content": ". For each pair of instance-level tokens " + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\{O_{II}^{3D}, O_{Ij}^{3D}\\}" + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "content": ", we calculate their Euclidean distance " + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "inline_equation", + "content": "d_{ij} = \\| C_i - C_j\\|_2" + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "content": " as well as their horizontal angle " + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\theta_h" + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "content": " and vertical angle " + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\theta_v" + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "content": ". Specially, " + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\theta_h = \\arctan 2((y_j - y_i) / (x_j - x_i))" + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\theta_v = \\arcsin ((z_j - z_i) / d_{ij})" + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "content": ". By making use of these positional parameters, we generate the pairwise spatial features " + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "inline_equation", + "content": "S = \\{s_{ij}\\} \\in \\mathbb{R}^{N \\times N \\times 5}" + }, + { + "bbox": [ + 55, + 247, + 296, + 331 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 339, + 295, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 339, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 81, + 339, + 295, + 354 + ], + "type": "interline_equation", + "content": "s _ {i j} = \\left[ \\sin (\\theta_ {h}), \\cos (\\theta_ {h}), \\sin (\\theta_ {v}), \\cos (\\theta_ {v}), d _ {i j} \\right]. (4)", + "image_path": "aae28827ea839849952c72c329c678dc62b0aa3d09b108e5e26d0b3c17e91909.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 360, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 445 + ], + "type": "text", + "content": "Inspired by the language-conditioned self-attention presented in ViL3DRel [5], we suggest a spatial-conditioned self-attention module. Specially, we create position embeddings " + }, + { + "bbox": [ + 55, + 360, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 360, + 296, + 445 + ], + "type": "text", + "content": " via the absolution positional encoding (PE), i.e. " + }, + { + "bbox": [ + 55, + 360, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\mathrm{PE}(C_1,C_2,\\dots ,C_N)" + }, + { + "bbox": [ + 55, + 360, + 296, + 445 + ], + "type": "text", + "content": ". A spatial conditioned attention weight " + }, + { + "bbox": [ + 55, + 360, + 296, + 445 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 55, + 360, + 296, + 445 + ], + "type": "text", + "content": " is computed to select the relevant spatial relations for each instance " + }, + { + "bbox": [ + 55, + 360, + 296, + 445 + ], + "type": "inline_equation", + "content": "O_{II}^{3D}" + }, + { + "bbox": [ + 55, + 360, + 296, + 445 + ], + "type": "text", + "content": ", which is formulated as below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 450, + 295, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 450, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 129, + 450, + 295, + 466 + ], + "type": "interline_equation", + "content": "l _ {i} = W _ {P} ^ {\\top} \\left(\\mathcal {P} _ {i} + O _ {I i} ^ {3 D}\\right), \\tag {5}", + "image_path": "b3e9bf2c0e64304e42a481ac781eeeafa9154752499b3e3685b8c9ef10db3960.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 473, + 295, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 473, + 295, + 522 + ], + "type": "inline_equation", + "content": "W_{P}\\in \\mathbb{R}^{D\\times 5}" + }, + { + "bbox": [ + 55, + 473, + 295, + 522 + ], + "type": "text", + "content": " is a learnable parameter and the bias term is omitted for simplicity. The spatial-conditioned attention map " + }, + { + "bbox": [ + 55, + 473, + 295, + 522 + ], + "type": "inline_equation", + "content": "\\omega_{ij}" + }, + { + "bbox": [ + 55, + 473, + 295, + 522 + ], + "type": "text", + "content": " is computed by combining " + }, + { + "bbox": [ + 55, + 473, + 295, + 522 + ], + "type": "inline_equation", + "content": "s_{ij},l_i" + }, + { + "bbox": [ + 55, + 473, + 295, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 473, + 295, + 522 + ], + "type": "inline_equation", + "content": "l_j" + }, + { + "bbox": [ + 55, + 473, + 295, + 522 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 55, + 473, + 295, + 522 + ], + "type": "inline_equation", + "content": "\\omega_{ij} = l_i\\cdot s_{ij}\\cdot l_j" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "text", + "content": "Through this positional transformation, the spatial-conditioned attention map " + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\omega_{ij}" + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "text", + "content": " encapsulates the pairwise spatial relationship across the entire 3D scene for each instance. We further integrate the attention map with instance-level visual tokens, which is formulated as below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 588, + 295, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 588, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 137, + 588, + 295, + 622 + ], + "type": "interline_equation", + "content": "F _ {i} = \\sum_ {j = 1} ^ {N} \\omega_ {i j} O _ {I i} ^ {3 D}. \\tag {6}", + "image_path": "7cc5f22a407a2bd9e2a5c6d23c39caf8132ba9d2e1ab72f0270c6874efc3a149.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 689 + ], + "type": "text", + "content": "Finally, the instance-level representations are concatenated and further processed by a transformer-based encoder " + }, + { + "bbox": [ + 55, + 629, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 55, + 629, + 296, + 689 + ], + "type": "text", + "content": ", followed by a max-pooling layer (Pool) and a simple two-layer MLP. The scene-level relation-aware tokens " + }, + { + "bbox": [ + 55, + 629, + 296, + 689 + ], + "type": "inline_equation", + "content": "F_{scene} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 55, + 629, + 296, + 689 + ], + "type": "text", + "content": " are generated for the entire 3D scene as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 84, + 697, + 295, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 697, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 84, + 697, + 295, + 711 + ], + "type": "interline_equation", + "content": "F _ {s c e n e} = \\operatorname {M L P} (\\operatorname {P o o l} (\\Gamma ([ F _ {1}, F _ {2}, \\dots , F _ {N} ]))). \\tag {7}", + "image_path": "0e38f1b96ff48f52422d90104d2e2652a79faa8262ac4a8ebb16c5093f035a93.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "Thus, we obtain the final token representations for LLM processing, i.e. instance-level visual tokens " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "O_{Ii}^{3D}" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": " and spatial relation-aware scene-level tokens " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "F_{scene}" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 117, + 536, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 536, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 536, + 131 + ], + "type": "text", + "content": "3.4. End-to-End Multi-task Instruction Tuning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 136, + 555, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 136, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 136, + 555, + 219 + ], + "type": "text", + "content": "3D-Language Prompts. Given an LMM, clear and explicit system messages and instructions are essential to support a range of downstream 3D vision-language tasks. For different tasks, we adopt various task-specific instruction templates with 3D features to generate uniform instruction data, enabling multi-task training. Please refer to the Supplementary Material for more detailed information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 220, + 556, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 556, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 556, + 340 + ], + "type": "text", + "content": "Instruction Tuning. We conduct end-to-end multi-task fine-tuning to fully leverage the capabilities of our Inst3D-LMM framework. The MCMF and 3D-ISR modules enable the LLM with robust 3D scene understanding, grounding and reasoning abilities. We freeze the pre-trained 3D object encoder while updating both MCMF and 3D-ISR modules, as well as LLM. Upon completion of end-to-end multi-task instruction tuning, Inst3D-LMM can effectively handle various 3D vision-language tasks simultaneously, without fine-tuning on specific tasks, as illustrated in Figure 7." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 351, + 397, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 397, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 397, + 365 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 371, + 437, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 437, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 437, + 384 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 389, + 556, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 556, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 556, + 569 + ], + "type": "text", + "content": "Datasets and Benchmarks. In this work, we conducted our experiments on the ScanNetV2 dataset [11], an extensive indoor 3D scene dataset comprising 1,513 scenes. This dataset includes 3D point clouds, RGB-D frames, and detailed point-level instance segmentation annotations. The whole dataset is divided into 1,201 scenes for training and 312 scenes for validation, with all subsequent benchmarks adhering to these training/validation splits. Our evaluation encompasses a range of 3D scene understanding benchmarks, including ScanRefer [4] and Multi3DRefer [55] for single- and multi-object 3D Visual Grounding, respectively, ScanQA [1] for 3D Question Answering, and Scan2Cap [8] for 3D Dense Captioning. These datasets are converted into a uniform instruction format for multi-task instruction tuning and performance assessment." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Implementation Details. For 3D feature extraction, we leverage the 3D instance segmentation model Mask3D [38] pre-trained on ScanNet200 dataset [37], alongside the 3D object encoder Ulip2 [48]/Uni3D [57] based on ViT-L/14 [13]. For 2D feature extraction, we adopt ViT-H-based SAM [23] to obtain high-quality masks. Moreover, we extract 2D semantic features using the CLIP-ViT-L/14-336px vision encoder [35]. These pre-trained models are kept frozen. In our method, we apply multi-view selection and multi-level crops to 2D images, setting " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " views and " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " levels. We use Vicuna1.5-7B [10] as our basic LLM, which is fine-tuned from LLaMA2 [41]. Our fine-tuning" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14151" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 556, + 315 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 556, + 315 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 556, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 556, + 315 + ], + "type": "table", + "html": "
MethodOverallUniqueMultipleMulti3DRefer
Acc@0.25↑Acc@0.50↑Acc@0.25↑Acc@0.50↑Acc@0.25↑Acc@0.50↑F1@0.25↑F1@0.50↑
Closed-set, full sup.
ScanRefer [4]37.324.365.043.330.619.8--
3DVG-Trans [56]45.934.577.258.538.428.730.225.5
M3DRef-CLIP [55]51.944.7-77.2-36.842.838.4
ConcreteNet [42]56.149.586.179.247.540.9--
CORE-3DVG [51]56.843.885.067.151.839.8--
Zero-shot
LLM-Grounder [50]17.15.330.822.616.312.9--
Visual-Programming [54]36.432.763.858.427.724.6--
Specialist
OpenScene [34]13.05.120.113.111.14.4--
3D-LLM(Flamingo) [16]21.2-------
3D-LLM(BLIP2-flant5) [16]30.3-------
Chat-3D v2 [18]35.930.461.257.625.222.645.141.6
ReGround3D [58]53.141.1------
Generalist
LAMM [52]-3.4------
3DMIT [26]10.77.2------
Grounded 3D-LLM [7]47.944.1----45.240.6
Chat-Scene [19]55.550.2----57.152.4
Inst3D-LMM57.851.688.681.548.743.258.353.5
", + "image_path": "2dfe96fa395f31888adaf3b5ea21555fbf42099d624bf9179277c8a362d44df9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 370, + 557, + 561 + ], + "blocks": [ + { + "bbox": [ + 55, + 318, + 555, + 363 + ], + "lines": [ + { + "bbox": [ + 55, + 318, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 555, + 363 + ], + "type": "text", + "content": "Table 1. Quantitative results for 3D Visual Grounding on ScanRefer and Multi3DRefer validation sets. In the ScanRefer dataset, scenes are labeled as \"unique\" (one object per class) or \"multiple\" (more than one). Closed-set methods are fully supervised for specific datasets. \"Zero-shot\" refers to methods that directly use LLMs without fine-tuning. \"Specialist\" and \"Generalist\" categorize methods fine-tuned for specific tasks versus those trained jointly. Bold and underlined numbers indicate the best and the second-best results, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 370, + 557, + 561 + ], + "lines": [ + { + "bbox": [ + 56, + 370, + 557, + 561 + ], + "spans": [ + { + "bbox": [ + 56, + 370, + 557, + 561 + ], + "type": "table", + "html": "
Method# 3D Data for AlignmentScanQAScan2Cap@0.50
B-1↑B-4↑METEOR↑ROUGE↑CIDER↑EM↑B-4↑METEOR↑ROUGE↑CIDER↑
Closed-set, full sup.
VoteNet [12] + MCAN [53]-28.06.211.429.854.717.3----
ScanRefer [4] + MCAN [53]-26.97.911.530.055.418.6----
ScanQA [8]-30.210.113.133.364.921.0----
Scan2Cap [1]-------22.421.443.535.2
3D-VisTA [60]-34.213.115.238.676.627.034.027.154.366.9
LLM-based Methods
LLaVA (zero-shot) [28]-7.10.310.512.35.70.21.58.319.63.2
LAMM [52]25K26.85.810.023.642.49.8----
3D-LLM(Flamingo) [16]675K30.37.212.232.359.220.45.911.429.9-
3D-LLM(BLIP2-flant5) [16]675K39.312.014.535.769.420.58.113.133.2-
Chat-3D v2 [18]38K38.47.316.140.177.121.131.822.350.263.9
LL3DA [6]38K-13.315.437.075.7-35.925.654.665.2
Grounded 3D-LLM [7]107K-13.4--72.7-35.5--70.6
Chat-Scene [19]38K-14.3--87.7-36.3--77.1
Inst3D-LMM38K43.514.918.442.688.624.638.327.557.279.7
", + "image_path": "50c1ebbffe00d4cedb42b3af21240db958e390a8422efd1ae42943bf3c91a601.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 566, + 527, + 578 + ], + "lines": [ + { + "bbox": [ + 82, + 566, + 527, + 578 + ], + "spans": [ + { + "bbox": [ + 82, + 566, + 527, + 578 + ], + "type": "text", + "content": "Table 2. Quantitative results for 3D Question Answering and 3D Dense Captioning on the ScanQA and Scan2Cap datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 590, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 628 + ], + "type": "text", + "content": "process utilizes LoRA [17]. We adopt the AdamW optimizer with a weight decay of 0.02. All experiments were performed on 8 NVIDIA A100 GPUs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 635, + 141, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 141, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 141, + 647 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "content": "3D Visual Grounding. We first report visual grounding performance on the ScanRefer and Multi3DRefer validation datasets. As shown in Table 1, our approach outperforms the state-of-the-art model, Chat-Scene [19], by " + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "inline_equation", + "content": "+2.3\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "content": " Acc@0.25 and " + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "inline_equation", + "content": "+1.2\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "content": " F1@0.25 on ScanRefer and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 590, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 638 + ], + "type": "text", + "content": "Multi3DRefer, respectively. Compared to specialist and closed-set methods, Inst3D-LMM, trained with a generalist approach, achieves competitive performance. Figure 5 displays typical visual comparison results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 638, + 556, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 556, + 697 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 556, + 697 + ], + "type": "text", + "content": "3D Question Answering. We then compare Inst3D-LMM with previous leading methods on the ScanQA validation set. Table 2 reports the results. Our Inst3D-LMM consistently outperforms these methods, including recent LL3DA [6], Grounded 3D-LLM [7] and Chat-Scene [19]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 698, + 553, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 698, + 553, + 710 + ], + "spans": [ + { + "bbox": [ + 313, + 698, + 553, + 710 + ], + "type": "text", + "content": "3D Dense Captioning. This task involves the localization" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14152" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 175, + 135 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 175, + 135 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 175, + 135 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 175, + 135 + ], + "type": "image", + "image_path": "e79d7be4241944c89c7fd56282d186f51e719e79cb1117808871a239f184654c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 85, + 186, + 117 + ], + "lines": [ + { + "bbox": [ + 176, + 85, + 186, + 117 + ], + "spans": [ + { + "bbox": [ + 176, + 85, + 186, + 117 + ], + "type": "text", + "content": "7-6" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 187, + 72, + 290, + 135 + ], + "blocks": [ + { + "bbox": [ + 187, + 72, + 290, + 135 + ], + "lines": [ + { + "bbox": [ + 187, + 72, + 290, + 135 + ], + "spans": [ + { + "bbox": [ + 187, + 72, + 290, + 135 + ], + "type": "image", + "image_path": "d2ffa5e1d1ccf3d270cc18ec24aa88e6cefb988e6d4cbb05a12f18bf4bb49592.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 149, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 313, + 149, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 553, + 182 + ], + "type": "text", + "content": "Figure 6. A visualization of the similarity score between text query and segmented 3D proposals. We compare sentence tokens used in ViL3DRel [5] with position embeddings employed in 3D-ISR." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 137, + 175, + 201 + ], + "blocks": [ + { + "bbox": [ + 59, + 137, + 175, + 201 + ], + "lines": [ + { + "bbox": [ + 59, + 137, + 175, + 201 + ], + "spans": [ + { + "bbox": [ + 59, + 137, + 175, + 201 + ], + "type": "image", + "image_path": "da796e3ac14e3d8d8653a051d01f783d1382f17972b86e7869d1386965f8c17f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 205, + 294, + 223 + ], + "lines": [ + { + "bbox": [ + 57, + 205, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 57, + 205, + 294, + 223 + ], + "type": "text", + "content": "Text Query: The kitchen cabinets are under the sink. They are the furthest left cabinets and to the left of the dishwasher." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 233, + 295, + 278 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 295, + 278 + ], + "type": "text", + "content": "Figure 5. Visual comparisons in 3D Visual Grounding. The rendered images show the ground-truth (blue), incorrectly identified objects (red), and correctly identified objects (green). The colored text indicates the results of text decoupling." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 177, + 137, + 290, + 201 + ], + "blocks": [ + { + "bbox": [ + 177, + 137, + 290, + 201 + ], + "lines": [ + { + "bbox": [ + 177, + 137, + 290, + 201 + ], + "spans": [ + { + "bbox": [ + 177, + 137, + 290, + 201 + ], + "type": "image", + "image_path": "2ac875c81967689bd998ab13e86d61467bee8f80b724dda833c9354f04e9a5a1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": "and description of instances. As shown in Table 2, our approach achieves " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "38.3\\%" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": " B-4@0.50 and " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "79.7\\%" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": " C@0.50 on Scan2Cap, which exceeds the closed-set expert model 3D-VisTA [60] by " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "+4.3\\%" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "+12.8\\%" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": ", and still outperforms the LLM-based method LL3DA by " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "+2.4\\%" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "+14.5\\%" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 353, + 149, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 149, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 149, + 366 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 372, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 407 + ], + "type": "text", + "content": "In this section, we perform ablation experiments to thoroughly evaluate the effectiveness of the components. Please refer to the Supplementary Material for more analysis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 407, + 296, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 587 + ], + "type": "text", + "content": "Effects of Multi-view Cross-Modal Fusion (MCMF). The MCMF module aims to augment LLM's understanding of 3D instances' geometric and semantic attributes. We explore several methods to combine these features, including direct concatenation for projection ('Concat.'), parallel projection followed by combination ('Parallel'), and vanilla cross-attention ('Cross-Attention'). To encode basic spatial information, we concatenate the results of absolution positional encoding (PE) with 3D features. As shown in Table 3, MCMF outperforms all these methods on ScanRefer, ScanQA and Scan2Cap datasets. We also observe that the performance of 3D LMM is significantly improved by incorporating multi-view 2D CLIP information compared to using only 3D geometric features. These results demonstrate the effectiveness of our MCMF approach." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 56, + 591, + 294, + 659 + ], + "blocks": [ + { + "bbox": [ + 56, + 591, + 294, + 659 + ], + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 659 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 659 + ], + "type": "table", + "html": "
MethodScanReferScanQAScan2Cap@0.50
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑B-4↑CIDER↑
w/o Multi-view 2D36.031.931.557.319.954.3
Concat.38.834.436.665.425.263.3
Parallel37.633.535.363.924.862.1
Cross-Attention39.236.537.466.125.564.3
MCMF46.741.941.578.632.768.2
", + "image_path": "a1cb7ef30dd34424a04f2d957ec4aef6efd0425e6f05ba9710e8e2feb78a833c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Impacts of 3D Instance Spatial Relation (3D-ISR). We further investigate the 3D-ISR module to analyze its impact" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 389, + 118 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 389, + 118 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 389, + 118 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 389, + 118 + ], + "type": "image", + "image_path": "84619ebb3073cac7a658d3902b7cec84c0c8913562255e76a605b7354176e437.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 118, + 372, + 125 + ], + "lines": [ + { + "bbox": [ + 334, + 118, + 372, + 125 + ], + "spans": [ + { + "bbox": [ + 334, + 118, + 372, + 125 + ], + "type": "text", + "content": "Input 3D Scene" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 126, + 544, + 140 + ], + "lines": [ + { + "bbox": [ + 315, + 126, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 315, + 126, + 544, + 140 + ], + "type": "text", + "content": "Text Query: There is a brown reclined in the corner of the room. The adjacent walls have two green boards. It is between the intersection of the 2 boards." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 390, + 72, + 463, + 118 + ], + "blocks": [ + { + "bbox": [ + 390, + 72, + 463, + 118 + ], + "lines": [ + { + "bbox": [ + 390, + 72, + 463, + 118 + ], + "spans": [ + { + "bbox": [ + 390, + 72, + 463, + 118 + ], + "type": "image", + "image_path": "82cda2238fe073390724c2790d2b6d6ad084d0bee731b4dca4601284be186bb9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 118, + 445, + 125 + ], + "lines": [ + { + "bbox": [ + 407, + 118, + 445, + 125 + ], + "spans": [ + { + "bbox": [ + 407, + 118, + 445, + 125 + ], + "type": "text", + "content": "Sentence Token" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 465, + 72, + 535, + 118 + ], + "blocks": [ + { + "bbox": [ + 465, + 72, + 535, + 118 + ], + "lines": [ + { + "bbox": [ + 465, + 72, + 535, + 118 + ], + "spans": [ + { + "bbox": [ + 465, + 72, + 535, + 118 + ], + "type": "image", + "image_path": "5126e0323f8535c82d39a60f31c68c2da22c225d69108025743ed435d1d262fc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 118, + 525, + 125 + ], + "lines": [ + { + "bbox": [ + 475, + 118, + 525, + 125 + ], + "spans": [ + { + "bbox": [ + 475, + 118, + 525, + 125 + ], + "type": "text", + "content": "Position Embedding" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 539, + 83, + 547, + 125 + ], + "lines": [ + { + "bbox": [ + 539, + 83, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 539, + 83, + 547, + 125 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 188, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 555, + 369 + ], + "type": "text", + "content": "on 3D spatial understanding using only 3D geometry priors. We compare 3D-ISR with previous spatial relation modeling methods, such as the 3D localization mechanism in ViL3DRel [5] and 3D-LLM [16], and relation-aware token generation in Chat-3D-v2 [18]. Table 4 reports the results of the comparison. It can be observed that our 3D-ISR consistently surpasses these approaches. Specifically, 3D-ISR significantly enhances performance on grounding tasks (ScanRefer and Multi3dRefer). This aligns with the motivation of our design. We also provide a visualization, as shown in Figure 6, of the similarity between the text query and segmented 3D proposals after processing by LLM. Compared to ViL3DRel [5] using sentence tokens, our 3D-ISR more accurately captures spatial relationships within the scene employing position embeddings of objects." + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 315, + 370, + 553, + 434 + ], + "blocks": [ + { + "bbox": [ + 55, + 665, + 295, + 687 + ], + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 687 + ], + "type": "text", + "content": "Table 3. Ablation evaluations of the proposed Multi-view CrossModal Fusion (MCMF) module." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 370, + 553, + 434 + ], + "lines": [ + { + "bbox": [ + 315, + 370, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 315, + 370, + 553, + 434 + ], + "type": "table", + "html": "
MethodScanReferMulti3DReferScanQA
Acc@0.25↑Acc@0.50↑F1@0.25↑F1@0.50↑B-1↑CIDER↑
w/o spatial relation35.629.834.827.329.255.6
ViL3DRel [5]36.130.538.435.532.858.9
3D-LLM [16]39.236.842.435.232.560.2
Chat-3D-v2 [18]40.837.541.637.835.164.6
3D-ISR48.344.146.241.339.172.3
", + "image_path": "85839773fcded001d376c0c5b75049f8da028908e53ab5fda745b6b1c0244677.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 436, + 553, + 458 + ], + "lines": [ + { + "bbox": [ + 313, + 436, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 553, + 458 + ], + "type": "text", + "content": "Table 4. Ablation study to verify the effectiveness of our 3D Instance Spatial Relation (3D-ISR) module." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 462, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 714 + ], + "type": "text", + "content": "Ablations within MCMF and 3D-ISR modules. We also explore the effectiveness of key designs in MCMF and 3D-ISR modules. The learnable [CLS] token in the MCMF module is introduced to aggregate 2D multi-view features. We compare it with alternative methods, such as token max pooling and Q-Former [24]. Table 5 reports the comparison results, which demonstrate the efficacy of the learnable [CLS] token design. Besides, to assess the impact of different pairwise spatial features in the 3D-ISR module, we evaluate the model using only distance or orientation information to compute spatial features in Eq 4. The results in Table 6 reveal that distance has a greater impact on the grounding capability of the model, while orientation is more crucial in handling Q&A tasks. Combining both pairwise distance and orientation yields the best overall performance. Mutual Benefits of MCMF and 3D-ISR. As shown in Table 7, the model that integrates MCMF and 3D-ISR consistently outperforms those that use either module alone for multiple tasks. To further verify its effectiveness, we combine MCMF with absolution positional encoding (PE) for the basic spatial relation, and directly adopt feature con" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14153" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 547, + 295 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 547, + 295 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 547, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 547, + 295 + ], + "type": "image", + "image_path": "b61b337457c3526896f0ddfb41b73ea6b1b582b6ad061616390b2bd3d6b24cf4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 301, + 507, + 312 + ], + "lines": [ + { + "bbox": [ + 102, + 301, + 507, + 312 + ], + "spans": [ + { + "bbox": [ + 102, + 301, + 507, + 312 + ], + "type": "text", + "content": "Figure 7. Qualitative illustration of Inst3D-LMM across various 3D-language tasks in diverse 3D environments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 320, + 294, + 375 + ], + "blocks": [ + { + "bbox": [ + 56, + 320, + 294, + 375 + ], + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 375 + ], + "type": "table", + "html": "
MethodScanReferScanQAScan2Cap@0.50
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑B-4↑CIDER↑
Max Pooling48.744.639.274.830.665.5
Q-Former [24]53.247.944.387.935.478.6
CLS Token57.851.643.588.638.379.7
", + "image_path": "2dfb0972536ec0f7a91aac782121feea0bf9a9c46a5fd021f08cceb663e6f23c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 392, + 294, + 443 + ], + "blocks": [ + { + "bbox": [ + 56, + 379, + 294, + 389 + ], + "lines": [ + { + "bbox": [ + 56, + 379, + 294, + 389 + ], + "spans": [ + { + "bbox": [ + 56, + 379, + 294, + 389 + ], + "type": "text", + "content": "Table 5. Ablation study of learnable CLS token in MCMF module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 392, + 294, + 443 + ], + "lines": [ + { + "bbox": [ + 56, + 392, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 294, + 443 + ], + "type": "table", + "html": "
MethodScanReferMulti3DReferScanQA
Acc@0.25↑Acc@0.50↑F1@0.25↑F1@0.50↑B-1↑CIDER↑
w/ Dist (only)43.238.042.437.831.265.7
w/ Ort (only)39.733.640.535.932.566.8
Dist + Ort48.344.146.241.339.172.3
", + "image_path": "512c717cee5aaca1da4468520964b9ca338277c02721b44c50847d54bfe09623.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 295, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 471 + ], + "type": "text", + "content": "Table 6. Ablation study of 3D-ISR module. Dist (Ort) means only using distance (orientation) to compute pairwise spatial features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 477, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 295, + 513 + ], + "type": "text", + "content": "catenation (Concat.) to integrate 3D object features and 2D multi-view features along with 3D-ISR. These results indicate the synergy efforts between MCMF and 3D-ISR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 514, + 295, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 633 + ], + "type": "text", + "content": "Different LLMs and Foundation Models. Table 8 presents the results of various LLMs and pre-trained 2D/3D models, including Vicuna-7B [10] vs Vicuna-13B, Ulip2 [48] vs Uni3D [57], CLIP [35] vs SigLIP [32], and ViT-H-based SAM [23] vs ViT-L-based SAM. We find that the performance of grounding and reasoning increases along with the total number of parameters in the foundation models (i.e. ViT-L vs Vit-H and 7B vs 13B). These results indicate that our framework's capability can be improved in tandem with the performance of foundation models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 645, + 128, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 128, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 128, + 657 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "In this paper, we proposed Inst3D-LMM, an effective instance-aware framework to leverage the potential of Large Multi-modal Models (LMMs) for 3D scene understanding. To improve instance-level representations, we developed" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 320, + 550, + 384 + ], + "blocks": [ + { + "bbox": [ + 315, + 320, + 550, + 384 + ], + "lines": [ + { + "bbox": [ + 315, + 320, + 550, + 384 + ], + "spans": [ + { + "bbox": [ + 315, + 320, + 550, + 384 + ], + "type": "table", + "html": "
MethodScanReferScanQAScan2Cap@0.50
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑B-4↑CIDER↑
w/ MCMF (only)38.234.639.476.830.566.9
w/ 3D-ISR (only)48.344.139.172.329.864.3
MCMF + PE46.741.941.578.632.768.2
3D-ISR + Concat.49.845.040.675.433.366.5
MCMF + 3D-ISR (full)57.851.643.588.638.379.7
", + "image_path": "9819281c250251bf3898edbe6de27b7477e70dbf232342931a12eaf5a20407ee.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 316, + 414, + 550, + 480 + ], + "blocks": [ + { + "bbox": [ + 313, + 388, + 554, + 409 + ], + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 409 + ], + "type": "text", + "content": "Table 7. Comparison results of the collaboration between MCMF and 3D-ISR modules." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 414, + 550, + 480 + ], + "lines": [ + { + "bbox": [ + 316, + 414, + 550, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 414, + 550, + 480 + ], + "type": "table", + "html": "
LLM3D Encoder2D VFsScanReferScanQA
Acc@0.25↑Acc@0.50↑B-1↑CIDER↑
Vicuna-7BUlip2CLIP+SAM (ViT-L)49.244.839.579.6
Vicuna-7BUlip2CLIP+SAM (ViT-H)53.646.543.787.5
Vicuna-7BUlip2SigLIP+SAM (ViT-H)54.347.042.284.5
Vicuna-7BUni3DCLIP+SAM (ViT-H)57.851.643.588.6
Vicuna-13BUni3DCLIP+SAM (ViT-H)56.052.042.883.1
Vicuna-13BUni3DSigLIP+SAM (ViT-H)55.247.843.385.4
", + "image_path": "19a79e0b7c13050671d5315d0860f9851121ef24265d880f79473056cb462d50.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 485, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 506 + ], + "type": "text", + "content": "Table 8. Ablation study of different LLMs and pre-trained foundation models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 520, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 628 + ], + "type": "text", + "content": "a novel Multi-view Cross-Modal Fusion (MCMF) module, which injects multi-view 2D semantic open-vocabulary priors into 3D geometry features to generate fine-grained instance-level tokens. Furthermore, we introduced a 3D Instance Spatial Relation (3D-ISR) module that employs the spatial condition attention mechanism to capture pairwise spatial relations. Experimental results demonstrate that our approach achieves promising performance in understanding and reasoning across various 3D vision-language tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 644, + 411, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 644, + 411, + 657 + ], + "spans": [ + { + "bbox": [ + 314, + 644, + 411, + 657 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China under Grants (62376244). It is also supported by Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14154" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 124 + ], + "type": "text", + "content": "[1] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, pages 19129-19139, 2022. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 167 + ], + "type": "text", + "content": "[2] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In CVPR, pages 16464-16473, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 295, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 295, + 212 + ], + "type": "text", + "content": "[3] Yang Cao, Zeng Yihan, Hang Xu, and Dan Xu. Coda: Collaborative novel box discovery and cross-modal alignment for open-vocabulary 3d object detection. In NeurIPS, pages 71862-71873, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 294, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 247 + ], + "type": "text", + "content": "[4] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221. Springer, 2020. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 248, + 294, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 248, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 248, + 294, + 291 + ], + "type": "text", + "content": "[5] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. In NeurIPS, pages 20522-20535, 2022. 2, 4, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 293, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 294, + 345 + ], + "type": "text", + "content": "[6] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In CVPR, pages 26428-26438, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 390 + ], + "type": "text", + "content": "[7] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 392, + 295, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 424 + ], + "type": "text", + "content": "[8] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In CVPR, pages 3193-3203, 2021. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 426, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 295, + 468 + ], + "type": "text", + "content": "[9] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In ICCV, pages 18109-18119, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 471, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 471, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 57, + 471, + 295, + 534 + ], + "type": "text", + "content": "[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 57, + 471, + 295, + 534 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 57, + 471, + 295, + 534 + ], + "type": "text", + "content": " chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 5, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "type": "text", + "content": "[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 2432–2443, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 581, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 613 + ], + "type": "text", + "content": "[12] Zhipeng Ding, Xu Han, and Marc Niethammer. Volenet: A deep learning label fusion method for multi-atlas segmentation. In MICCAI, pages 202-210, 2019. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 615, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 615, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 615, + 295, + 678 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[14] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm:" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "text", + "content": "Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 151 + ], + "type": "text", + "content": "[15] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 153, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 196 + ], + "type": "text", + "content": "[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. In NeurIPS, pages 20482-20494, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 241 + ], + "type": "text", + "content": "[17] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 285 + ], + "type": "text", + "content": "[18] Haifeng Huang, Zehan Wang, Rongjie Huang, Luping Liu, Xize Cheng, Yang Zhao, Tao Jin, and Zhou Zhao. Chat-3d v2: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023. 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 288, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 288, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 288, + 553, + 341 + ], + "type": "text", + "content": "[19] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. NeuIPS, 2024. 1, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 343, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 396 + ], + "type": "text", + "content": "[20] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "text", + "content": "[21] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. In ECCV, pages 528-545. Springer, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "text", + "content": "[22] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In CVPR, pages 10984-10994, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "text", + "content": "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3, 4, 5, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 580, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 622 + ], + "type": "text", + "content": "[25] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 624, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 666 + ], + "type": "text", + "content": "[26] Zeju Li, Chao Zhang, Xiaoyan Wang, Ruilong Ren, Yifan Xu, Ruifei Ma, and Xiangde Liu. 3dmit: 3d multi-modal instruction tuning for scene understanding. In ICMEW, 2024. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[27] Cai Liang, Bo Li, Zhengming Zhou, Longlong Wang, Pengfei He, Liang Hu, and Haoxing Wang. Spatioawaregrouding3d: A spatio aware model for improving 3d vision grousing. In CVPRW, 2024. 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14155" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, pages 34892-34916, 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 160 + ], + "type": "text", + "content": "[29] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In NeurIPS, pages 44860-44879, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 195 + ], + "type": "text", + "content": "[30] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "text", + "content": "[31] Yunze Man, Shuhong Zheng, Zhipeng Bao, Martial Hebert, Liang-Yan Gui, and Yu-Xiong Wang. Lexicon3d: Probing visual foundation models for complex 3d scene understanding. arXiv preprint arXiv:2409.03757, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "type": "text", + "content": "[32] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, pages 529-544. Springer, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "text", + "content": "[33] Maria Parelli, Alexandros Delitzas, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Clip-guided vision-language pre-training for question answering in 3d scenes. In CVPR, pages 5606-5611, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "text", + "content": "[34] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In CVPR, pages 815-824, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 3, 4, 5, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[36] Krishan Rana, Jesse Haviland, Sourav Garg, Jad Abou-Chakra, Ian Reid, and Niko Suenderhauf. Sayplan: Grounding large language models using 3d scene graphs for scalable robot task planning. In CoRL, pages 23-72, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 504 + ], + "type": "text", + "content": "[37] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, pages 125-141. Springer, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "text", + "content": "[38] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In ICRA, pages 8216-8223, 2023. 3, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "type": "text", + "content": "[39] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M. Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In ICCV, pages 2998-3009, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "type": "text", + "content": "[40] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In NeurIPS, pages 68367-68390, 2023. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "type": "text", + "content": "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[42] Ozan Unal, Christos Sakaridis, Suman Saha, Fisher Yu, and Luc Van Gool. Three ways to improve verbo-visual" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "fusion for dense 3d visual grounding. arXiv preprint arXiv:2309.04561, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 152 + ], + "type": "text", + "content": "[43] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, et al. Embodiedscan: A holistic multimodal 3d perception suite towards embodied ai. In CVPR, pages 19757-19767, 2024. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[44] Zehan Wang, Haifeng Huang, Yang Zhao, Linjun Li, Xize Cheng, Yichen Zhu, Aoxiong Yin, and Zhou Zhao. 3drp-net: 3d relative position-aware network for 3d visual grounding. In EMNLP, page 10612-10625, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "text", + "content": "[45] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "type": "text", + "content": "[46] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In CVPR, pages 19231-19242, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 331 + ], + "type": "text", + "content": "[47] Runsen Xu, Zhiwei Huang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Vlm-grounder: A vlm agent for zero-shot 3d visual grounding. arXiv preprint arXiv:2410.13860, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 388 + ], + "type": "text", + "content": "[48] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In CVPR, pages 27091–27101, 2024. 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "text", + "content": "[49] Fan Yang, Sicheng Zhao, Yanhao Zhang, Haoxiang Chen, Hui Chen, Wenbo Tang, Haonan Lu, Pengfei Xu, Zhenyu Yang, Jungong Han, et al. Lmi3d: Empowering llm with 3d perception from a single 2d image. arXiv preprint arXiv:2408.07422, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 443, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 498 + ], + "type": "text", + "content": "[50] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. In ICRA, pages 7694-7701, 2024. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 499, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 544 + ], + "type": "text", + "content": "[51] Li Yang, Ziqi Zhang, Zhongang Qi, Yan Xu, Wei Liu, Ying Shan, Bing Li, Weiping Yang, Peng Li, Yan Wang, et al. Exploiting contextual objects and relations for 3d visual grounding. In NeurIPS, pages 49542-49554, 2024. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "type": "text", + "content": "[52] Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Xiaoshui Huang, Zhiyong Wang, Lu Sheng, Lei Bai, et al. Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. In NeurIPS, pages 26650-26685, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "text", + "content": "[53] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In CVPR, pages 6281–6290, 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "text", + "content": "[54] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3d visual grounding. In CVPR, pages 20623-20633, 2024. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[55] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, pages 15225-15236, 2023. 5, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14156" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 262 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[56] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2928-2937, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[57] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. In ICLR, 2024. 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[58] Chenming Zhu, Tai Wang, Wenwei Zhang, Kai Chen, and Xihui Liu. Empowering 3d visual grounding with reasoning capabilities. arXiv preprint arXiv:2407.01525, 2024. 1, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[59] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmms with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[60] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 2, 4, 6, 7" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14157" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Instance-wise Supervision-level Optimization in Active Learning/6ab274ce-c563-45f8-9f14-82dfa36b95f4_content_list.json b/2025/Instance-wise Supervision-level Optimization in Active Learning/6ab274ce-c563-45f8-9f14-82dfa36b95f4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1be8858f2442c0453424efd2b6f760a6f466fd77 --- /dev/null +++ b/2025/Instance-wise Supervision-level Optimization in Active Learning/6ab274ce-c563-45f8-9f14-82dfa36b95f4_content_list.json @@ -0,0 +1,1406 @@ +[ + { + "type": "text", + "text": "Instance-wise Supervision-level Optimization in Active Learning", + "text_level": 1, + "bbox": [ + 174, + 130, + 823, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shinnosuke Matsuo $^{1}$ , Riku Togashi $^{2}$ , Ryoma Bise $^{1}$ , Seiichi Uchida $^{1}$ , Masahiro Nomura $^{2}$ $^{1}$ Kyushu University, Japan $^{2}$ CyberAgent, Inc., Japan", + "bbox": [ + 148, + 179, + 844, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 251, + 326, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Active learning (AL) is a label-efficient machine learning paradigm that focuses on selectively annotating high-value instances to maximize learning efficiency. Its effectiveness can be further enhanced by incorporating weak supervision, which uses rough yet cost-effective annotations instead of exact (i.e., full) but expensive annotations. We introduce a novel AL framework, Instance-wise Supervision-Level Optimization (ISO), which not only selects the instances to annotate but also determines their optimal annotation level within a fixed annotation budget. Its optimization criterion leverages the value-to-cost ratio (VCR) of each instance while ensuring diversity among the selected instances. In classification experiments, ISO consistently outperforms traditional AL methods and surpasses a state-of-the-art AL approach that combines full and weak supervision, achieving higher accuracy at a lower overall cost. This code is available at https://github.com/matsuoshinnosuke/ISOAL.", + "bbox": [ + 89, + 282, + 485, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 585, + 220, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Machine learning technology has achieved remarkable success in recent years across various fields, including computer vision, natural language processing, and robotics. However, a major challenge remains: creating labeled datasets for training models often requires significant time and cost, especially in specialized domains.", + "bbox": [ + 89, + 612, + 482, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this issue, active learning [1, 3, 6, 8, 11, 16, 18-21, 24] has been widely studied to improve model performance under limited budgets. In active learning, instances (data samples) for annotation are selected iteratively from an unlabeled dataset, followed by annotation and model retraining. This repeated cycle improves the model's accuracy while progressively minimizing annotation costs. A common approach in active learning is uncertainty-based sampling [3, 5, 8, 9, 18, 22, 24, 25], where instances with high uncertainty, such as low confident of the model's prediction, are selected for annotation, as these are expected to contribute most effectively to model performance.", + "bbox": [ + 89, + 703, + 482, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Another approach to reduce annotation costs is weak su", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pervision [17, 26, 27]. For classification tasks, a weak supervision approach is often designed to just attach rough class labels to individual instances. For example, instead of attaching the exact class label \"song sparrow\" or \"house sparrow\" to an image instance, annotators can attach a rough class label \"sparrows.\" This approach drastically reduces annotation costs (budgets) because the fee paid to a non-expert annotator, who can only assign rough class labels rather than exact ones via a crowdsourcing service, is lower than that of an expert annotator. Hereafter, we call rough classes as superclasses; therefore, \"sparrows\" is a superclass of the exact class \"song sparrow.\"", + "bbox": [ + 511, + 252, + 906, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In general, weak supervision approaches still need instances with full supervision, that is, instances with the exact class labels. This is simply because it is impossible to determine the class boundary between, \"sparrows\" is a superclass of \"song sparrow\" without the instances with these exact class labels. Consequently, we need to balance weak supervision and full supervision under a pre-specified total annotation budget. In other words, we need to optimize the supervision levels of individual instances. Since the total annotation budget is pre-specified as a constraint, the number of instances with full supervision will be limited. More formally, this supervision-level optimization task can be seen as an optimal resource allocation problem with the constraint on available resources.", + "bbox": [ + 511, + 440, + 908, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes a label-efficient machine learning approach, called Instance-wise Supervision-level Optimization (ISO) for active learning, as shown in Fig. 1. Roughly speaking, the proposed approach is a sophisticated hybrid of weak supervision and active learning. Different from the state-of-the-art hybrid approach [22] (reviewed in Section 2), our approach has a novel mechanism that automatically determines the optimal supervision levels of individual instances. Specifically, as shown in Fig. 1, all unlabeled instances at a certain round are automatically divided into three categories: instances for full supervision (with exact class annotation), instances for weak supervision (with superclass annotation), and instances that remain unlabeled. The instances selected in the first two categories are considered to be selected for active learning; annotators, therefore, attach appropriate labels to these instances according", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4939", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff300c4be666d44a776b9ea022ac3ca2c19016ddf4b63f783033e9cf76876faa.jpg", + "image_caption": [ + "Figure 1. Overview of Instance-wise Supervision-level Optimization (ISO) for active learning. The proposed approach combines weak and full supervision within an active learning framework to maximize annotation efficiency under a fixed budget. First, each instance $\\pmb{x}$ in the unlabeled data pool is evaluated for its value-to-cost ratio $(v_{\\mathrm{f}}(\\pmb{x}))$ for full supervision and $v_{\\mathrm{w}}(\\pmb{x})$ for weak supervision). Based on these values and data diversity, instances are selected within the budget constraints for weak supervision (with superclass labels) or full supervision (with exact class labels). This adaptive allocation of supervision-level enables optimal use of resources within the budget constraints in active learning." + ], + "image_footnote": [], + "bbox": [ + 161, + 93, + 848, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to their supervision levels.", + "bbox": [ + 89, + 393, + 266, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our experiments, the proposed method consistently outperformed existing approaches across multiple datasets, demonstrating a more efficient use of the annotation budget. Our approach achieved the same accuracy as the comparative methods while using only three-fifths of the budget by incorporating weak labels when the cost of weak label annotation is half that of fully supervised annotation. Furthermore, in comparison with the state-of-the-art method that combines active learning and multiple supervision, our approach achieved significantly better cost efficiency by leveraging instance-specific information.", + "bbox": [ + 89, + 407, + 483, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are summarized as follows:", + "bbox": [ + 89, + 574, + 483, + 602 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel active learning framework that optimizes the supervision level for each instance, allowing for dynamic allocation of weak and full annotations based on budget constraints.", + "- We propose an efficient algorithm that assigns the optimal supervision level to each instance by considering value-to-cost ratios and data diversity, enhancing the cost-effectiveness of the annotation process.", + "- We validate the effectiveness of our approach on classification tasks, using superclasses as weak supervision, and demonstrate its superiority over conventional AL methods and the state-of-the-art AL method that incorporates both full and weak supervision." + ], + "bbox": [ + 89, + 604, + 483, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 814, + 228, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Active learning (AL) [1, 16, 20] is a technique aimed at maximizing model performance under a limited annotation budget by strategically selecting informative instances for annotation. A widely adopted approach in active learning is", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "uncertainty-based sampling [3, 5, 8, 9, 18, 22, 24, 25]. For example, [18] selects instances with small margins, prioritizing instances for which the model has low confidence. Another common approach focuses on maximizing diversity within selected instances [6, 15, 19, 21], which aims to cover a broad range of the data distribution. For instance, [19] selects a representative core set of instances to maximize diversity. Moreover, there are methods that combine uncertainty and diversity-based strategies to leverage the advantages of both approaches [3, 11]. Additionally, some AL methods [4, 12] address the setting where different annotators have varying accuracy levels. These methods aim to jointly optimize both the selection of data for annotation and the choice of annotators to minimize label noise. However, all of these approaches assume a single supervision-level and do not incorporate multiple supervision-levels.", + "bbox": [ + 511, + 392, + 906, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another approach to reduce annotation costs is weakly supervised learning (WSL) [17, 26, 27], where less detailed or approximate labels instead of fully supervised labels are utilized. In many real-world applications, obtaining fully supervised labels can be expensive and time-consuming. To address this, WSL methods use weak supervision, which is typically easier and cheaper to obtain. For example, in classification tasks, weak supervision can be provided by superclasses such as categories, and in segmentation tasks, it can be provided by bounding boxes, allowing faster labeling with lower costs.", + "bbox": [ + 511, + 638, + 908, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A hybrid approach between AL and WSL, where the proportion of full and weak supervision is adaptively determined, has been proposed in [22]. Although this approach aims to optimize budget use by setting an ideal balance of weak and full supervision, it applies random sampling after setting the proportion without instance-level considera", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4940", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/04698a57d1ab3c979758e1406b61f2134dc551f551548c29875613d0d1680e03.jpg", + "table_caption": [ + "Table 1. Comparison of conventional active learning methods. \"Uncertainty\" (Unc.) refers to whether a method incorporates uncertainty-based instance selection, such as selecting instances with minimal margins. \"Diversity\" (Div.) indicates if the method considers data diversity to maximize the representativeness of selected instances. \"Multiple supervision\" (Mul.) refers to the capability to utilize multiple levels of supervision, such as both weak and full supervision. \"Instance-wise\" (Ins.) represents the ability to choose supervision for each instance adaptively." + ], + "table_footnote": [], + "table_body": "
Margin, etc. [8, 18, 24]Coreset, etc. [6, 19, 21]BADGE, etc. [3, 11]APFWA [22]Ours
Unc.✓×✓×✓
Div.×✓✓×✓
Mul.×××✓✓
Ins.✓✓✓×✓
", + "bbox": [ + 89, + 219, + 486, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions. To our knowledge, no method exists that dynamically assigns full and weak supervision for each instance.", + "bbox": [ + 89, + 345, + 482, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1 summarizes a comparison between popular active learning methods and our instance-wise supervision-level optimization. Each method is evaluated on its ability to incorporate uncertainty, diversity, multiple levels of supervision, and instance-wise adaptability, as defined in the caption. Most uncertainty-based and diversity-based AL methods do not consider multiple levels of supervision, focusing instead on selecting instances based on a single type of annotation. Although APFWA incorporates multiple levels of supervision, it lacks instance-wise optimization. In contrast, our approach is unique in that it optimizes the supervision level on an instance-wise basis, simultaneously considering data uncertainty, diversity, and the annotation cost associated with multiple levels of supervision.", + "bbox": [ + 89, + 377, + 483, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem setting", + "text_level": 1, + "bbox": [ + 89, + 602, + 248, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{D}_{\\mathrm{u}}$ represent the unlabeled data pool, and let $T$ denote the number of rounds in the active learning process, where $B$ is the available budget in each round. In this setting, a set of instances is selected to be annotated with either full supervision or weak supervision, depending on the required level of detail and the associated cost. Specifically, we define $C_{\\mathrm{f}}$ as the annotation cost per instance for full supervision, where precise labels are assigned, and $C_{\\mathrm{w}}$ as the annotation cost per instance for weak supervision, which involves less precise, lower-cost labeling. These annotation costs, $C_{\\mathrm{f}}$ and $C_{\\mathrm{w}}$ , are provided in advance.", + "bbox": [ + 89, + 628, + 482, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The process of batch selection and training is repeated for $T$ rounds, allowing the model to progressively improve by iteratively selecting and annotating the most informative instances within each budget-constrained round. Once an instance is annotated, it is moved from the unlabeled pool $\\mathcal{D}_{\\mathrm{u}}$ to either the fully supervised data pool $\\mathcal{D}_{\\mathrm{f}}$ or the weakly supervised data pool $\\mathcal{D}_{\\mathrm{w}}$ , depending on the supervi", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 ISO: Instance-wise Supervision-level Optimization" + ], + "code_body": "1: Inputs: Round $T$ , budget of each round $B$ , annotation cost $C_{\\mathrm{f}}, C_{\\mathrm{w}}$ , unlabeled data pool $\\mathcal{D}_{\\mathrm{u}}$ , feature extractor $f$ , classification head $h_{\\mathrm{f}}, h_{\\mathrm{w}}$ . \n2: Outputs: trained $f^{T}, h_{\\mathrm{f}}^{T}$ . \n3: for round $t = 1, \\ldots, T$ do \n4: if $t$ is 1 then \n5: Sample randomly instances from $\\mathcal{D}_{\\mathrm{u}}$ within budget $B$ to create an initial fully supervised dataset $\\mathcal{D}_{\\mathrm{f}}$ and weakly supervised dataset $\\mathcal{D}_{\\mathrm{w}}$ with an equal number of instances. \n6: else if $t > 1$ then \n7: Compute the model improvements, $M_{\\mathrm{f}}^{t}, M_{\\mathrm{w}}^{t}$ , for each supervision using $\\mathcal{D}_{\\mathrm{f}}$ and $\\mathcal{D}_{\\mathrm{w}}$ by Eq. (3) and (4). \n8: for $\\pmb{x}$ in $\\mathcal{D}_{\\mathrm{u}}$ do \n9: 1. Compute the uncertainties, $u_{\\mathrm{f}}^{t}(\\pmb{x}), u_{\\mathrm{f}}^{t}(\\pmb{x})$ , for each supervision. \n10: 2. Compute the VCRs, $v_{\\mathrm{f}}^{t}(\\pmb{x}), v_{\\mathrm{w}}^{t}(\\pmb{x})$ , for each supervision by Eq. (1) and (2). \n11: 3. Compute the normalized features $\\tilde{f}^{t}(\\pmb{x})$ . \n12: end for \n13: Select the batch $\\mathcal{D}_{\\mathrm{f}}^{+}, \\mathcal{D}_{\\mathrm{w}}^{+}$ using the VCRs, $v_{\\mathrm{f}}^{t}(\\pmb{x}), v_{\\mathrm{w}}^{t}(\\pmb{x})$ , and the normalized features $\\tilde{f}^{t}(\\pmb{x})$ by Sec. 4.3. \n14: end if \n15: $\\mathcal{D}_{\\mathrm{f}} \\gets \\mathcal{D}_{\\mathrm{f}} \\cup \\mathcal{D}_{\\mathrm{f}}^{+}, \\mathcal{D}_{\\mathrm{u}} \\gets \\mathcal{D}_{\\mathrm{u}} \\setminus \\mathcal{D}_{\\mathrm{f}}^{+}$ \n16: $\\mathcal{D}_{\\mathrm{w}} \\gets \\mathcal{D}_{\\mathrm{w}} \\cup \\mathcal{D}_{\\mathrm{w}}^{+}, \\mathcal{D}_{\\mathrm{u}} \\gets \\mathcal{D}_{\\mathrm{u}} \\setminus \\mathcal{D}_{\\mathrm{w}}^{+}$ \n17: Train $f^{t}, h_{\\mathrm{f}}^{t}$ , and $h_{\\mathrm{w}}^{t}$ using $\\mathcal{D}_{\\mathrm{f}}$ and $\\mathcal{D}_{\\mathrm{w}}$ . \n18: end for", + "bbox": [ + 521, + 123, + 906, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sion level chosen. These annotated instances are then used to train a classifier.", + "bbox": [ + 511, + 593, + 906, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The ultimate goal in this setting is to obtain the classifier's parameters to classify the exact class of input data. To train the classification network, both fully supervised data and weakly supervised data contribute to training the feature extractor in the network. Therefore, balancing fully and weakly supervised data in a cost-effective manner is crucial for optimizing the learning process within budget constraints.", + "bbox": [ + 511, + 623, + 906, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed method: Instance-wise Supervision-level Optimization (ISO)", + "text_level": 1, + "bbox": [ + 511, + 758, + 916, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 801, + 624, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose an efficient algorithm to address the instance-wise supervision-level optimization problem within a fixed budget. Our approach comprises two primary components, as shown in Fig. 1. The first component calculates the value-to-cost ratio (VCR) for each instance-supervision", + "bbox": [ + 511, + 824, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4941", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "level pair, which indicates the expected improvement in model accuracy per unit of cost achieved through annotation for each instance. The VCR for each instance is determined by both the value of the supervision level and its uncertainty. This allows us to evaluate the cost-effectiveness of assigning different supervision levels to each instance individually. The second component performs batch selection within the budget $B$ , optimizing the set of instances based on their VCRs and diversity. This two-part approach enables our approach to maximize the value of annotations within a limited budget by balancing cost-efficiency and diversity.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Value-to-cost ratio (VCR) estimation", + "text_level": 1, + "bbox": [ + 89, + 281, + 408, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we explain the calculation of the value-to-cost ratio (VCR) for each instance-supervision pair. An overview of the algorithm is provided in Algorithm 1. For an instance $\\pmb{x}$ in the unlabeled pool $\\mathcal{D}_{\\mathrm{u}}$ , the VCRs $v_{\\mathrm{f}}^{t}(\\pmb{x})$ for full supervision and $v_{\\mathrm{w}}^{t}(\\pmb{x})$ for weak supervision at round $t$ are determined by the uncertainty of the instance, the value, and cost for the two different-level labels, which are defined as:", + "bbox": [ + 89, + 304, + 483, + 424 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {\\mathrm {f}} ^ {t} (\\boldsymbol {x}) = \\frac {M _ {\\mathrm {f}} ^ {t} u _ {\\mathrm {f}} ^ {t} (\\boldsymbol {x})}{C _ {\\mathrm {f}}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 431, + 480, + 464 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {\\mathrm {w}} ^ {t} (\\boldsymbol {x}) = \\frac {M _ {\\mathrm {w}} ^ {t} u _ {\\mathrm {w}} ^ {t} (\\boldsymbol {x})}{C _ {\\mathrm {w}}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 465, + 480, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $C_{\\mathrm{f}}, C_{\\mathrm{w}}$ represent the annotation costs, and $M_{\\mathrm{f}}^{t}, M_{\\mathrm{w}}^{t}$ are the expected model improvement, where $*_{\\mathrm{f}}, *_{\\mathrm{w}}$ correspond to full supervision and weak supervision, respectively. $u_{\\mathrm{f}}^{t}(\\boldsymbol{x}), u_{\\mathrm{w}}^{t}(\\boldsymbol{x})$ represents the uncertainty for each instance $\\boldsymbol{x}$ in the unlabeled pool $D_{\\mathrm{u}}$ at round $t$ , calculated based on model predictions. This formulation allows us to evaluate the cost-effectiveness of assigning either full or weak supervision to each instance, providing a basis for the batch selection strategy within a fixed budget.", + "bbox": [ + 89, + 507, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The expected model improvements $M_{\\mathrm{f}}^{t}$ for full supervision and $M_{\\mathrm{w}}^{t}$ for weak supervision (i.e., the values for each supervision) are estimated by evaluating the improvement in model performance as the training data is incrementally increased. Consider the $t$ -th annotation round, where fully supervised data $\\mathcal{D}_{\\mathrm{f}}^{t}$ and weakly supervised data $\\mathcal{D}_{\\mathrm{w}}^{t}$ are available at that time. To evaluate accuracy improvement, we divide the labeled data into $K$ subsets and incrementally add training data. For fully supervised data, the weakly supervised data is kept fixed while subsets of the fully supervised data are added. The model performance $m_{\\mathrm{f},k}$ represents the evaluation using the first $k$ subsets of fully supervised training data, with $m_{\\mathrm{f},1}$ being the performance when using just one subset. Similarly, we compute $m_{\\mathrm{w},k}$ for the weakly supervised data, where the fully supervised data is fixed, and subsets of the weakly supervised data are incrementally added. Here, model performance is evaluated us", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing a small validation set randomly sampled within the budget of a single round. Then, a weighted average is applied for the improvement at each step, $m_{k+1}^t - m_k^t$ to smooth the sequence of improvements. The reason for the weighted approach is to focus on the latest model behavior. To obtain the improvement per data point, each improvement value is normalized by dividing by $|\\mathcal{D}| / K$ , where $|\\mathcal{D}|$ represents the number of instances. That is, $M_{\\mathrm{f}}^t$ and $M_{\\mathrm{w}}^t$ are expressed as follows:", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {\\mathrm {f}} ^ {t} = \\frac {\\sum_ {k = 1} ^ {K - 1} k \\cdot \\left(m _ {\\mathrm {f} , k + 1} ^ {t} - m _ {\\mathrm {f} , k} ^ {t}\\right)}{\\sum_ {k = 1} ^ {K - 1} k} \\cdot \\frac {K}{\\left| \\mathcal {D} _ {\\mathrm {f}} \\right|} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 237, + 903, + 277 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {\\mathrm {w}} ^ {t} = \\frac {\\sum_ {k = 1} ^ {K - 1} k \\cdot \\left(m _ {\\mathrm {w} , k + 1} ^ {t} - m _ {\\mathrm {w} , k} ^ {t}\\right)}{\\sum_ {k = 1} ^ {K - 1} k} \\cdot \\frac {K}{\\left| \\mathcal {D} _ {\\mathrm {w}} \\right|} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 280, + 903, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The uncertainty $u_{\\mathrm{f}}^{t}(\\pmb{x})$ and $u_{\\mathrm{w}}^{t}(\\pmb{x})$ , corresponding to full and weak supervision, can be flexibly defined based on the specific task and dataset. For example, in the classification task, the entropy of the class prediction probabilities is adopted. If different functions are used to calculate uncertainty for full and weak supervision, there is a risk that the scales of $u_{\\mathrm{f}}^{t}(\\pmb{x})$ and $u_{\\mathrm{w}}^{t}(\\pmb{x})$ may differ significantly. To address this, we normalize each uncertainty value by converting it to a percentile score. In this transformation, the instance with the highest uncertainty is assigned a score of 1, the instance with the lowest uncertainty receives a score of 0, and the instance ranked at the 10th percentile in terms of uncertainty is assigned a score of 0.9. We use percentiles rather than dividing by the maximum uncertainty value because extreme outliers could result in an overly large maximum, causing most values to be close to zero. This percentile-based normalization ensures that the distribution of uncertainty values remains well-scaled, preserving relative differences across instances without distortion due to outliers.", + "bbox": [ + 511, + 334, + 906, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Valuable and diverse batch selection using VCR", + "text_level": 1, + "bbox": [ + 511, + 648, + 903, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we describe the batch selection using the value-to-cost ratio (VCR) defined in the previous section. The most straightforward approach for batch selection is to greedily select instances with the highest VCR values. However, as noted in previous studies (e.g. [3, 19]), this approach can lead to selecting similar instances that have high VCR values, resulting in a loss of data diversity within the batch.", + "bbox": [ + 511, + 688, + 903, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, we propose a batch selection algorithm to select a diverse and valuable batch. In other words, we select a group of instances with high VCR values and diverse features. Specifically, we represent each instance as a vector using its VCR and feature vector. The vector differs depending on the type of label: $v_{\\mathrm{f}}(\\boldsymbol{x}) \\tilde{f}(\\boldsymbol{x})$ for full labels", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4942", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $v_{\\mathrm{w}}(\\pmb{x}) \\tilde{f}(\\pmb{x})$ for weak labels. Here, $\\tilde{f}(\\pmb{x})$ is a normalized feature vector (with a magnitude of 1) output by the feature extractor $f$ . The optimal set of instances is selected based on the vectors.", + "bbox": [ + 89, + 89, + 480, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We aim to select a set of instances that maximizes the area enclosed by the vertices of the selected vectors, as shown in the center of Figure 1. Here, the optimization parameter is which vectors (instances) to select and their label types. When the selected instances have high diversity and value, the area enclosed by the selected vectors becomes larger. We solve this optimization problem subject to budget constraints (i.e., the sum of the costs of the optimized supervision levels for instances is less than $B$ ). The optimization problem of selecting vectors and determining their supervision levels can be reformulated as a selection problem, where the goal is to choose the optimal set of vectors from all possible vectors, including both those with weak labels $\\{v_{\\mathrm{w}}(\\pmb{x})\\tilde{f} (\\pmb{x})|\\pmb{x}\\in \\mathcal{D}_{\\mathrm{u}}\\}$ and those with full labels $\\{v_{\\mathrm{f}}(\\pmb {x})\\tilde{f} (\\pmb {x})|\\pmb {x}\\in \\mathcal{D}_{\\mathrm{u}}\\}$ . This allows the problem of choosing the type of label to be integrated into the vector selection process. This optimization problem can be solved by maximizing the determinant of a set of vectors subject to budget constraints.", + "bbox": [ + 91, + 152, + 483, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To perform the batch selection that maximizes the determinant of a matrix, one approach is to sample from a $k$ -Determinantal Point Process ( $k$ -DPP) [13]. This method selects a batch containing $k$ vectors with a probability proportional to the determinant of their Gram matrix. However, this differs from the batch selection in our approach in that we choose a set of vectors within a budget $B$ rather than choosing $k$ vectors. Therefore, we propose a new selection method, which is inspired by the work [3] using the $k$ -means++ seeding algorithm [2] as an approximation algorithm for $k$ -DPP.", + "bbox": [ + 89, + 440, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our batch selection method, the vector of the batch is sampled sequentially with a probability proportional to the square of the distance to the vector closest to the one selected so far until the budget $B$ is exhausted. Specifically, if the selected vector corresponds to full supervision (i.e., a vector in the set $\\{v_{\\mathrm{f}}(\\pmb{x})\\tilde{f}(\\pmb{x})|\\pmb{x} \\in \\mathcal{D}_{\\mathrm{u}}\\}$ , the cost $C_{\\mathrm{f}}$ is deducted from the budget, and it is added to the fully supervised data. If it corresponds to weak supervision (i.e., a vector in the set $\\{v_{\\mathrm{w}}(\\pmb{x})\\tilde{f}(\\pmb{x})|\\pmb{x} \\in \\mathcal{D}_{\\mathrm{u}}\\}$ ), the cost $C_{\\mathrm{w}}$ is deducted from the budget, and it is added to the weakly supervised data. This process is repeated until the budget $B$ is exhausted. By doing this, we can obtain a diverse and valuable batch.", + "bbox": [ + 89, + 607, + 483, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 820, + 222, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Dataset", + "text_level": 1, + "bbox": [ + 89, + 845, + 186, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the proposed method, we utilized two datasets: CIFAR100 and Caltech-UCSD", + "bbox": [ + 89, + 869, + 480, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Birds-200-2011 (CUB200). The task is a multi-class classification problem, where each class (full label) belongs to a superclass, and the superclass label is provided as a weak label.", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CIFAR-100 is a dataset of natural images containing 100 classes that span a wide range of everyday objects and animals. Each class in CIFAR-100 is associated with a predefined superclass, grouping similar classes together under broader categories, resulting in 20 superclasses in total. The dataset consists of 60,000 images in total, with 50,000 images for training and 10,000 images for testing", + "bbox": [ + 511, + 151, + 903, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CUB200 [23] consists of 200 classes of bird species, with each class representing a specific bird. Following previous study [14], we defined superclasses based on the suffixes in the class names, resulting in 70 unique superclasses. The dataset contains 11,788 images, with 5,994 images designated for training and 5,794 images for testing.", + "bbox": [ + 511, + 257, + 903, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Comparison methods", + "text_level": 1, + "bbox": [ + 511, + 358, + 714, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compared our approach with eight conventional AL methods, ranging from standard single-supervision AL methods to the state-of-the-art AL method incorporating multiple supervision, as follows:", + "bbox": [ + 511, + 382, + 903, + 441 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Random: The simplest baseline that samples instances randomly from the unlabeled pool.", + "- Margin [18]: A method that selects instances in ascending order of the difference between the highest and second-highest predicted class probabilities, prioritizing instances with minimal margins.", + "- MaxConf [24]: An approach that selects instances in order of the lowest maximum predicted class probability, focusing on instances with the least confidence.", + "- Entropy [24]: A method that selects instances according to the highest entropy in the predicted class probabilities, targeting instances with the greatest uncertainty.", + "- Coreset [19]: A diversity-based sampling method that selects instances farthest from the labeled data using a greedy algorithm to enhance diversity.", + "- ALBL (Active Learning by Learning) [9]: A multiarmed bandit approach that selects between AL methods at each round. In this experiment, we utilized MaxConf and Coreset.", + "- BADGE (Batch Active learning by Diverse Gradient Embeddings) [3]: A method that considers both predictive uncertainty and diversity by selecting instances with diverse and high-magnitude gradients in the gradient space.", + "- APFWA (Adaptive Proportion of Full and Weak Annotations) [22]: A state-of-the-art AL method that incorporates both full and weak supervision, which dynamically determines the proportion of full and weak annotations within each batch." + ], + "bbox": [ + 513, + 443, + 903, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4943", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since AL with both full and weak label annotation is a novel task, there is only one existing method for this task. Instead, we compare our approach with standard single-supervision AL methods as well as the state-of-the-art AL method. This comparison highlights the effectiveness of AL using weak labels. Additionally, a comparison with APFWA demonstrates the advantages of our approach for AL with both fully and weakly supervised labels. Furthermore, in Sec. 6.3, we also compare our approach with a simple combination of weak supervision and conventional AL methods to ensure a fair evaluation.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These methods offer a range of approaches, from simple random sampling to advanced techniques that balance uncertainty and diversity, as well as methods leveraging weak supervision, enabling a comprehensive evaluation of our proposed approach.", + "bbox": [ + 89, + 262, + 482, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 363, + 305, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The network consists of a shared feature extractor $f$ and two classification heads, $h_{\\mathrm{f}}$ for full supervision and $h_{\\mathrm{w}}$ for weak supervision. The encoder of ResNet18 [7] is used for the feature extractor $f$ , and a single linear layer is used for each classification head. For the network trained using both fully and weakly labeled data, we employ a two-stage training approach in which the feature extractor and each classification head are trained first with weakly supervised data and then with fully supervised data. Since both fully supervised and weakly supervised learning are classification tasks in this experiment, we used the cross-entropy loss. The learning rate was set to 0.001, and the optimizer was Adam [10].", + "bbox": [ + 89, + 391, + 482, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The settings for active learning were as follows. The number of rounds $T$ was set to 5. The budget $B$ of each round was set to 1,000 for CIFAR100 and to 500 for CUB200 since CUB200 had approximately 5,000 in training data, which was less than that of CIFAR100. The validation data with full supervision were prepared from training data, 1,000 for CIFAR1000 and 500 for CUB200, which are the same number and small amount as in budget $B$ . The cost of full supervision $C_{\\mathrm{f}}$ was set to 1, and the cost of weak supervision was set to $\\frac{1}{2}$ . In addition, in Sec. 6.2, we conducted the experiments when changing $C_{\\mathrm{w}}$ to $\\frac{1}{4}$ and $\\frac{1}{8}$ to investigate the impact of the difference on the cost of weak supervision.", + "bbox": [ + 89, + 578, + 482, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The number of partitions $K$ for estimating the expected model improvements $M_{\\mathrm{f}}^{t}$ and $M_{\\mathrm{w}}^{t}$ was set to 5. As the performance metric of $m_{\\mathrm{f},k}^{t}, m_{\\mathrm{w},k}^{t}$ , we used the classification accuracy. To enhance stability, we repeated the estimation process three times with different random seeds and calculated the average. For uncertainty measurement, we adopted the simplest approach, Margin [18], to evaluate instance uncertainty.", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experimental results", + "text_level": 1, + "bbox": [ + 513, + 90, + 714, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Comparison with conventional active learning methods", + "text_level": 1, + "bbox": [ + 511, + 114, + 903, + 145 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The comparison results with conventional AL methods on CIFAR100 and CUB200 are shown in Fig. 2(a), (b), respectively. The graphs show the classification accuracy $[\\%]$ for each round $t$ .", + "bbox": [ + 511, + 154, + 903, + 214 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Figure 2(a), the results on the CIFAR100 dataset demonstrate that our approach archives higher accuracy across all rounds compared to conventional methods, including random sampling, uncertainty-based AL method (Margin, MaxConf, Entropy), diversity-based AL method (Coreset), and other AL methods (ALBL, BADGE), which use only full supervision. This demonstrates that our approach effectively leverages weak supervision, achieving higher performance within the same budget constraints. For example, at round $t = 5$ , our approach achieved over $10\\%$ higher accuracy compared to conventional AL methods that use only full supervision. Furthermore, while these methods require five rounds (with a budget of 5000) to reach an accuracy of approximately $30\\%$ , our approach achieved the same or even higher accuracy in just three rounds, using only three-fifths of the budget.", + "bbox": [ + 511, + 215, + 903, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach also surpasses APFWA, another AL approach that incorporates weak supervision by optimizing the proportions of weak and full supervision within the batch but does not focus on selecting instances. This result suggests that our approach more effectively selects cost-effective instances by dynamically considering instance-wise uncertainty and diversity, thereby maximizing the learning impact of weak supervision.", + "bbox": [ + 511, + 457, + 903, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarly, Figure 2(b) presents the results on the CUB200 dataset, where our approach also outperforms other AL methods across all rounds. The consistent improvement observed across both datasets indicates that our approach can dynamically optimize annotation strategies according to the dataset characteristics, achieving high adaptability.", + "bbox": [ + 511, + 579, + 903, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Analysis for weak supervision costs", + "text_level": 1, + "bbox": [ + 511, + 696, + 820, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate that our approach works at any cost of weak supervision and that the cheaper the cost of weak supervision, the greater the effect, we conducted experiments when changing the cost of weak supervision. Figure 3 shows the classification accuracy $[\\%]$ on CIFAR100 and CUB200 at different costs of weak supervision. In this experiment, the cost of full supervision $C_{\\mathrm{f}}$ is fixed at 1, meaning that $C_{\\mathrm{w}}$ represents the relative cost to full supervision. It is important to note that the cost $C_{\\mathrm{w}}$ is defined before active learning begins and is not a hyperparameter of the proposed method. Imagine a scenario where the fee paid to a non-expert annotator, who can only assign rough class labels (weak la", + "bbox": [ + 511, + 719, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4944", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4dc4e1235dcb03007319bdab6df23db603cd0ab30ba6cb085757df61fff12b32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 89, + 433, + 132 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8f670618b8ff430ba3888a047569a930906347cef3afd0a13a6789333a04e0b5.jpg", + "image_caption": [ + "(a) CIFAR100" + ], + "image_footnote": [], + "bbox": [ + 96, + 143, + 284, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa75b502bca009497eff13610fa83b8cd18c3a779816ea4068bef82b553d938f.jpg", + "image_caption": [ + "(b) CUB200" + ], + "image_footnote": [], + "bbox": [ + 297, + 143, + 467, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5802bbf2b719fe0d35273cf6b22a7a2a2b356ba433d9e89d03ac7dbf752fd9d.jpg", + "image_caption": [ + "Figure 2. Comparison of our approach with the conventional active learning methods. The figures show the classification accuracy $[\\% ]$ ( $\\uparrow$ ) for each round $t$ on CIFAR100 and CUB200. The cost of full supervision $C_{\\mathrm{f}}$ is 1, and that of weak supervision is $\\frac{1}{2}$ ." + ], + "image_footnote": [], + "bbox": [ + 127, + 381, + 447, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71c1a593d67beb675f60daa048442032da4b52e0087a56d896f50c6cb59cd633.jpg", + "image_caption": [ + "(a) CIFAR100", + "Figure 3. Classification accuracy $[\\% ]$ ( $\\uparrow$ ) of the proposed method on CIFAR100 and CUB200 when the cost of weak supervision, $C_{\\mathrm{w}}$ is $\\frac{1}{2}$ , $\\frac{1}{4}$ , and $\\frac{1}{8}$ . Note that the cost $C_{\\mathrm{w}}$ is a value set by the user before active learning begins and is not a hyperparameter of the proposed method. This figure illustrates the effect of different weak supervision costs. For reference, we also include the results of conventional active learning methods (gray lines)." + ], + "image_footnote": [], + "bbox": [ + 96, + 420, + 284, + 559 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2641baf29215fa0eb1b4e08ce2ed818bab1e110aec97f1962978198a7fb09631.jpg", + "image_caption": [ + "(b) CUB200" + ], + "image_footnote": [], + "bbox": [ + 297, + 421, + 467, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bels), is lower than that of an expert annotator. The cost $C_{\\mathrm{w}}$ varies for each task or dataset, corresponding to the rewards offered by crowdsourcing, etc.", + "bbox": [ + 89, + 703, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In previous sections, we set the cost of weak supervision to $C_{\\mathrm{w}} = \\frac{1}{2}$ . Here, we investigate the effects of our approach when further reducing $C_{\\mathrm{w}}$ to $\\frac{1}{4}$ and $\\frac{1}{8}$ . As shown in Figure 3, for both datasets, the classification accuracy improves as the cost of weak supervision decreases. This improvement is due to the fact that the lower cost of weak supervision allows more data to be annotated within the same budget, thereby enhancing the training process. In other words, the cheaper the cost of weak supervision, the more impactful weak supervision and our approach become.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Comparison with combined active learning and multiple supervision baseline", + "text_level": 1, + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to comparing our approach with conventional active learning approaches, we conduct experiments to evaluate our approach against a straightforward combination of weak supervision and conventional active learning techniques. This baseline approach applies a fixed proportion of weak supervision within the budget, utilizing conventional active learning strategies within each supervision type.", + "bbox": [ + 511, + 128, + 903, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, in this combined method, weak supervision is allocated at a constant rate— $20\\%$ , $40\\%$ , $60\\%$ , and $80\\%$ in the batch. Within each full and weak supervision, conventional AL methods are employed to select data. In this experiment, the active learning method for each full or weak supervision employed was sampling based on margin, which is used for the uncertainty of the proposed method.", + "bbox": [ + 511, + 234, + 903, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 shows the comparison between our approach and baseline methods that combine active learning with weak supervision on the CIFAR100 dataset under different costs for weak supervision ( $C_{\\mathrm{w}} = \\frac{1}{2}, \\frac{1}{4}$ , and $\\frac{1}{8}$ ). In each case, the baseline methods use fixed proportions of full and weak supervision within each batch, denoted as $\\mathrm{F}(x\\%) + \\mathrm{W}(y\\%)$ , where $x$ and $y$ represent the percentages of full and weak supervision, respectively.", + "bbox": [ + 511, + 340, + 903, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Across all weak supervision costs, our approach consistently achieves high accuracy. An important point to note is that the optimal proportion of full and weak supervision is unknown beforehand and varies depending on the cost of weak supervision. For example, when the cost of weak supervision $C_{\\mathrm{w}}$ is $\\frac{1}{2}$ , the configuration with 60% full supervision and 40% weak supervision achieves the best accuracy among the baseline methods. However, when the cost $C_{\\mathrm{w}}$ is reduced to $\\frac{1}{8}$ , a configuration with 20% full supervision and 80% weak supervision performs best. In contrast, our approach dynamically constructs optimal batches by optimizing both the supervision level and instance selection based on the cost-effectiveness obtained from the trained model at each round.", + "bbox": [ + 511, + 462, + 906, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 681, + 663, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted ablation studies to demonstrate the effectiveness of instance-wise value (i.e., uncertainty) and the batch selection with diversity on CIFAR100 and CUB200, as shown in Fig. 5.", + "bbox": [ + 511, + 705, + 903, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The first ablation variant, referred to as \"Ours w/o uncertainty,\" removes the instance-wise uncertainty component from the VCR calculation. In this variant, the VCR is determined solely based on the annotation cost and expected model improvement without taking into account the", + "bbox": [ + 511, + 766, + 903, + 842 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1APFWA [22] is a method that dynamically determines the proportion of full and weak annotations. It must use random instance sampling to estimate the optimal proportion, and as a result, it cannot be combined with the conventional instance-wise AL methods (e.g., margin-based sampling).", + "bbox": [ + 511, + 849, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4945", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a1532eafe25d4978b5d52b2fb40a9a0c1995d7914597d91293290f0150e9490.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 88, + 727, + 118 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4908f98fbfb49a5ebfafcb60bb112c38f91b198a3328660338f322d9636bbced.jpg", + "image_caption": [ + "(a) $C_{\\mathrm{w}} = 1 / 2$" + ], + "image_footnote": [], + "bbox": [ + 119, + 125, + 339, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4fd101f3ca5a897278ba8136280a8a590d2d9ef2ff65d07bc22a9adc8092b519.jpg", + "image_caption": [ + "(b) $C_{\\mathrm{w}} = 1 / 4$" + ], + "image_footnote": [], + "bbox": [ + 388, + 125, + 609, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/369cef2477ab6be0f60a42a1302d9b874452722766e5c6cbbaf71af6fa9bb421.jpg", + "image_caption": [ + "(c) $C_{\\mathrm{w}} = 1 / 8$" + ], + "image_footnote": [], + "bbox": [ + 661, + 125, + 879, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8cc7e75bf52f191c52751370dd796d214f3108d02126ef517aafc2081360c780.jpg", + "image_caption": [ + "Figure 4. Comparison with combined active learning and multiple supervision baseline. The classification accuracy $[\\%] (\\uparrow)$ on CIFAR100 with fixed proportions of full and weak supervision within each batch. Each batch was sampled with fixed proportions, and conventional active learning methods were applied separately for full and weak supervision.", + "(a) CIFAR100", + "(b) CUB200", + "Figure 5. Ablation study comparing the proposed method without instance-wise uncertainty $u_{\\mathrm{f}}(\\pmb{x})$ , $u_{\\mathrm{w}}(\\pmb{x})$ (Ours w/o uncertainty) and without batch selection considering diversity (Ours w/o diversity). The figures show the classification accuracy $[\\%] (\\uparrow)$ for each round $t$ on CIFAR100 and CUB200." + ], + "image_footnote": [], + "bbox": [ + 96, + 381, + 472, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "uncertainty of each instance. Consequently, the supervision level is not optimized for each instance individually. The results show that our approach outperforms this variant on both datasets. This effect is particularly prominent in the CUB200 dataset, demonstrating the benefit of optimizing \"instance-wise\" supervision-level.", + "bbox": [ + 89, + 670, + 482, + 761 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The second ablation variant, called 'Ours w/o diversity,\" involves a greedy approach to batch selection, where instances are selected in descending order of VCR without considering data diversity. Our approach, which incorporates diversity into the batch selection process, consistently outperforms this variant on both datasets. The impact of diversity-aware selection is especially evident in the CIFAR100 dataset, demonstrating that considering diversity in batch selection significantly enhances model performance.", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion, limitation, and future work", + "text_level": 1, + "bbox": [ + 511, + 383, + 870, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed a novel active learning (AL) approach, called Instance-wise Supervision-level Optimization (ISO), which optimizes the supervision level (either full or weak annotation) for each instance to maximize annotation efficiency within a budget. By evaluating the uncertainty, diversity, and value-to-cost ratio (VCR) of each instance for full and weak supervision, our approach automatically determines the instances to be annotated and their optimal annotation level. Experiments on classification tasks demonstrated that ISO outperforms not only standard AL methods but also state-of-the-art AL methods with weak supervision, while effectively leveraging instance-specific information to improve both accuracy and cost-efficiency.", + "bbox": [ + 511, + 409, + 906, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A current limitation of this study is that the effectiveness of our ISO has been validated only in classification experiments. The concept of ISO is applicable to various tasks beyond classification. For example, ISO can be applied to object segmentation tasks, where precise yet costly object boundaries are assumed for full supervision, while simpler rectangular bounding boxes are used for weak supervision. Another current limitation is the assumption that supervision levels are only two. We can expand to deal with more supervision levels by introducing class levels other than exact classes and superclasses. In future work, we aim to demonstrate the full versatility of our approach by exploring these extensions across a broader range of tasks, reinforcing its potential as a scalable, cost-effective solution for real-world applications.", + "bbox": [ + 511, + 606, + 908, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work was supported by JSPS KAKENHI Grant Number JP23KJ1723, and JST, ACT-X Grant Number JPMJAX23CR, Japan.", + "bbox": [ + 511, + 852, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4946", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Charu C Aggarwal, Xiangnan Kong, Quanquan Gu, Jiawei Han, and S Yu Philip. Active learning: A survey. In Data classification, pages 599-634. Chapman and Hall/CRC, 2014. 1, 2", + "[2] D ARTHUR. k-means++: the advantages of careful seeding. In ACM-SIAM symposium on Discrete algorithms, pages 1027-1035, 2007. 5", + "[3] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. In International Conference on Learning Representations. 1, 2, 3, 4, 5", + "[4] Shayok Chakraborty. Asking the right questions to the right users: Active learning with imperfect oracles. In Association for the Advancement of Artificial Intelligence, pages 3365-3372, 2020. 2", + "[5] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International Conference on Machine Learning, pages 1183-1192, 2017. 1, 2", + "[6] Guy Hacohen, Avihu Dekel, and Daphna Weinshall. Active learning on a budget: Opposite strategies suit high and low budgets. In International Conference on Machine Learning, pages 8175-8195, 2022. 1, 2, 3", + "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770-778, 2016. 6", + "[8] Neil Houlsby, Ferenc Huszár, Zoubin Ghahramani, and Máté Lengyel. Bayesian active learning for classification and preference learning. arXiv preprint arXiv:1112.5745, 2011. 1, 2, 3", + "[9] Wei-Ning Hsu and Hsuan-Tien Lin. Active learning by learning. In Association for the Advancement of Artificial Intelligence, 2015. 1, 2, 5", + "[10] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[11] Andreas Kirsch, Joost Van Amersfoort, and Yarin Gal. Batchbald: Efficient and diverse batch acquisition for deep bayesian active learning. Neural Information Processing Systems, 32, 2019. 1, 2, 3", + "[12] Dani Kiyasseh, Tingting Zhu, and David A Clifton. Soqal: Selective oracle questioning for consistency based active learning of cardiac signals. In International Conference on Machine Learning, pages 11302-11340, 2022. 2", + "[13] Alex Kulesza and Ben Taskar. k-dpps: Fixed-size determinantal point processes. In International Conference on Machine Learning, pages 1193-1200, 2011. 5", + "[14] Chao Lu and Yuexian Zou. Using coarse label constraint for fine-grained visual classification. In Conference on Multimedia Modeling, 2018. 5", + "[15] Andrew Kachites McCallum, Kamal Nigam, et al. Employing em and pool-based active learning for text classification. In International Conference on Machine Learning, pages 350-358, 1998. 2", + "[16] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang." + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "A survey of deep active learning. ACM computing surveys (CSUR), 54(9):1-40, 2021. 1, 2", + "[17] Zeyu Ren, Shuihua Wang, and Yudong Zhang. Weakly supervised machine learning. CAAI Transactions on Intelligence Technology, 8(3):549-580, 2023. 1, 2", + "[18] Dan Roth and Kevin Small. Margin-based active learning for structured output spaces. In European Conference on Machine Learning, pages 413-424, 2006. 1, 2, 3, 5, 6", + "[19] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 2, 3, 4, 5", + "[20] Burr Settles. Active learning literature survey. 2009. 2", + "[21] Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell. Variational adversarial active learning. In International Conference on Computer Vision, pages 5972-5981, 2019. 1, 2, 3", + "[22] Javier Gamazo Tejero, Martin S Zinkernagel, Sebastian Wolf, Raphael Sznitman, and Pablo Márquez-Neila. Full or weak annotations? an adaptive strategy for budget-constrained annotation campaigns. In Conference on Computer Vision and Pattern Recognition, pages 11381-11391, 2023. 1, 2, 3, 5, 7", + "[23] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Technical report, 2011. 5", + "[24] Dan Wang and Yi Shang. A new active labeling method for deep learning. In International Joint Conference on Neural Networks, pages 112-119, 2014. 1, 2, 3, 5", + "[25] Donggeun Yoo and In So Kweon. Learning loss for active learning. In Computer Vision and Pattern Recognition, pages 93-102, 2019. 1, 2", + "[26] Man Zhang, Yong Zhou, Jiaqi Zhao, Yiyun Man, Bing Liu, and Rui Yao. A survey of semi-and weakly supervised semantic segmentation of images. Artificial Intelligence Review, 53:4259-4288, 2020. 1, 2", + "[27] Zhi-Hua Zhou. A brief introduction to weakly supervised learning. National science review, 5(1):44-53, 2018. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 604 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4947", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2025/Instance-wise Supervision-level Optimization in Active Learning/6ab274ce-c563-45f8-9f14-82dfa36b95f4_model.json b/2025/Instance-wise Supervision-level Optimization in Active Learning/6ab274ce-c563-45f8-9f14-82dfa36b95f4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..24859b3154f089ac99513356a0943ddc4776b21c --- /dev/null +++ b/2025/Instance-wise Supervision-level Optimization in Active Learning/6ab274ce-c563-45f8-9f14-82dfa36b95f4_model.json @@ -0,0 +1,1901 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.175, + 0.131, + 0.825, + 0.154 + ], + "angle": 0, + "content": "Instance-wise Supervision-level Optimization in Active Learning" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.18, + 0.846, + 0.219 + ], + "angle": 0, + "content": "Shinnosuke Matsuo\\(^{1}\\), Riku Togashi\\(^{2}\\), Ryoma Bise\\(^{1}\\), Seiichi Uchida\\(^{1}\\), Masahiro Nomura\\(^{2}\\) \n\\(^{1}\\)Kyushu University, Japan \\(^{2}\\)CyberAgent, Inc., Japan" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.252, + 0.327, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.284, + 0.486, + 0.555 + ], + "angle": 0, + "content": "Active learning (AL) is a label-efficient machine learning paradigm that focuses on selectively annotating high-value instances to maximize learning efficiency. Its effectiveness can be further enhanced by incorporating weak supervision, which uses rough yet cost-effective annotations instead of exact (i.e., full) but expensive annotations. We introduce a novel AL framework, Instance-wise Supervision-Level Optimization (ISO), which not only selects the instances to annotate but also determines their optimal annotation level within a fixed annotation budget. Its optimization criterion leverages the value-to-cost ratio (VCR) of each instance while ensuring diversity among the selected instances. In classification experiments, ISO consistently outperforms traditional AL methods and surpasses a state-of-the-art AL approach that combines full and weak supervision, achieving higher accuracy at a lower overall cost. This code is available at https://github.com/matsuoshinnosuke/ISOAL." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.587, + 0.222, + 0.603 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Machine learning technology has achieved remarkable success in recent years across various fields, including computer vision, natural language processing, and robotics. However, a major challenge remains: creating labeled datasets for training models often requires significant time and cost, especially in specialized domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.885 + ], + "angle": 0, + "content": "To address this issue, active learning [1, 3, 6, 8, 11, 16, 18-21, 24] has been widely studied to improve model performance under limited budgets. In active learning, instances (data samples) for annotation are selected iteratively from an unlabeled dataset, followed by annotation and model retraining. This repeated cycle improves the model's accuracy while progressively minimizing annotation costs. A common approach in active learning is uncertainty-based sampling [3, 5, 8, 9, 18, 22, 24, 25], where instances with high uncertainty, such as low confident of the model's prediction, are selected for annotation, as these are expected to contribute most effectively to model performance." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Another approach to reduce annotation costs is weak su" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.253, + 0.908, + 0.435 + ], + "angle": 0, + "content": "pervision [17, 26, 27]. For classification tasks, a weak supervision approach is often designed to just attach rough class labels to individual instances. For example, instead of attaching the exact class label \"song sparrow\" or \"house sparrow\" to an image instance, annotators can attach a rough class label \"sparrows.\" This approach drastically reduces annotation costs (budgets) because the fee paid to a non-expert annotator, who can only assign rough class labels rather than exact ones via a crowdsourcing service, is lower than that of an expert annotator. Hereafter, we call rough classes as superclasses; therefore, \"sparrows\" is a superclass of the exact class \"song sparrow.\"" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.909, + 0.652 + ], + "angle": 0, + "content": "In general, weak supervision approaches still need instances with full supervision, that is, instances with the exact class labels. This is simply because it is impossible to determine the class boundary between, \"sparrows\" is a superclass of \"song sparrow\" without the instances with these exact class labels. Consequently, we need to balance weak supervision and full supervision under a pre-specified total annotation budget. In other words, we need to optimize the supervision levels of individual instances. Since the total annotation budget is pre-specified as a constraint, the number of instances with full supervision will be limited. More formally, this supervision-level optimization task can be seen as an optimal resource allocation problem with the constraint on available resources." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This paper proposes a label-efficient machine learning approach, called Instance-wise Supervision-level Optimization (ISO) for active learning, as shown in Fig. 1. Roughly speaking, the proposed approach is a sophisticated hybrid of weak supervision and active learning. Different from the state-of-the-art hybrid approach [22] (reviewed in Section 2), our approach has a novel mechanism that automatically determines the optimal supervision levels of individual instances. Specifically, as shown in Fig. 1, all unlabeled instances at a certain round are automatically divided into three categories: instances for full supervision (with exact class annotation), instances for weak supervision (with superclass annotation), and instances that remain unlabeled. The instances selected in the first two categories are considered to be selected for active learning; annotators, therefore, attach appropriate labels to these instances according" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4939" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.162, + 0.094, + 0.849, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.283, + 0.908, + 0.368 + ], + "angle": 0, + "content": "Figure 1. Overview of Instance-wise Supervision-level Optimization (ISO) for active learning. The proposed approach combines weak and full supervision within an active learning framework to maximize annotation efficiency under a fixed budget. First, each instance \\( \\pmb{x} \\) in the unlabeled data pool is evaluated for its value-to-cost ratio \\( (v_{\\mathrm{f}}(\\pmb{x})) \\) for full supervision and \\( v_{\\mathrm{w}}(\\pmb{x}) \\) for weak supervision). Based on these values and data diversity, instances are selected within the budget constraints for weak supervision (with superclass labels) or full supervision (with exact class labels). This adaptive allocation of supervision-level enables optimal use of resources within the budget constraints in active learning." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.394, + 0.267, + 0.407 + ], + "angle": 0, + "content": "to their supervision levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.575 + ], + "angle": 0, + "content": "In our experiments, the proposed method consistently outperformed existing approaches across multiple datasets, demonstrating a more efficient use of the annotation budget. Our approach achieved the same accuracy as the comparative methods while using only three-fifths of the budget by incorporating weak labels when the cost of weak label annotation is half that of fully supervised annotation. Furthermore, in comparison with the state-of-the-art method that combines active learning and multiple supervision, our approach achieved significantly better cost efficiency by leveraging instance-specific information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.603 + ], + "angle": 0, + "content": "The main contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.605, + 0.483, + 0.665 + ], + "angle": 0, + "content": "- We introduce a novel active learning framework that optimizes the supervision level for each instance, allowing for dynamic allocation of weak and full annotations based on budget constraints." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.666, + 0.483, + 0.725 + ], + "angle": 0, + "content": "- We propose an efficient algorithm that assigns the optimal supervision level to each instance by considering value-to-cost ratios and data diversity, enhancing the cost-effectiveness of the annotation process." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.726, + 0.484, + 0.801 + ], + "angle": 0, + "content": "- We validate the effectiveness of our approach on classification tasks, using superclasses as weak supervision, and demonstrate its superiority over conventional AL methods and the state-of-the-art AL method that incorporates both full and weak supervision." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.605, + 0.484, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.229, + 0.831 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Active learning (AL) [1, 16, 20] is a technique aimed at maximizing model performance under a limited annotation budget by strategically selecting informative instances for annotation. A widely adopted approach in active learning is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.635 + ], + "angle": 0, + "content": "uncertainty-based sampling [3, 5, 8, 9, 18, 22, 24, 25]. For example, [18] selects instances with small margins, prioritizing instances for which the model has low confidence. Another common approach focuses on maximizing diversity within selected instances [6, 15, 19, 21], which aims to cover a broad range of the data distribution. For instance, [19] selects a representative core set of instances to maximize diversity. Moreover, there are methods that combine uncertainty and diversity-based strategies to leverage the advantages of both approaches [3, 11]. Additionally, some AL methods [4, 12] address the setting where different annotators have varying accuracy levels. These methods aim to jointly optimize both the selection of data for annotation and the choice of annotators to minimize label noise. However, all of these approaches assume a single supervision-level and do not incorporate multiple supervision-levels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.909, + 0.805 + ], + "angle": 0, + "content": "Another approach to reduce annotation costs is weakly supervised learning (WSL) [17, 26, 27], where less detailed or approximate labels instead of fully supervised labels are utilized. In many real-world applications, obtaining fully supervised labels can be expensive and time-consuming. To address this, WSL methods use weak supervision, which is typically easier and cheaper to obtain. For example, in classification tasks, weak supervision can be provided by superclasses such as categories, and in segmentation tasks, it can be provided by bounding boxes, allowing faster labeling with lower costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "A hybrid approach between AL and WSL, where the proportion of full and weak supervision is adaptively determined, has been proposed in [22]. Although this approach aims to optimize budget use by setting an ideal balance of weak and full supervision, it applies random sampling after setting the proportion without instance-level considera" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4940" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.485, + 0.215 + ], + "angle": 0, + "content": "Table 1. Comparison of conventional active learning methods. \"Uncertainty\" (Unc.) refers to whether a method incorporates uncertainty-based instance selection, such as selecting instances with minimal margins. \"Diversity\" (Div.) indicates if the method considers data diversity to maximize the representativeness of selected instances. \"Multiple supervision\" (Mul.) refers to the capability to utilize multiple levels of supervision, such as both weak and full supervision. \"Instance-wise\" (Ins.) represents the ability to choose supervision for each instance adaptively." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.22, + 0.487, + 0.324 + ], + "angle": 0, + "content": "
Margin, etc. [8, 18, 24]Coreset, etc. [6, 19, 21]BADGE, etc. [3, 11]APFWA [22]Ours
Unc.✓×✓×✓
Div.×✓✓×✓
Mul.×××✓✓
Ins.✓✓✓×✓
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.483, + 0.377 + ], + "angle": 0, + "content": "tions. To our knowledge, no method exists that dynamically assigns full and weak supervision for each instance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.485, + 0.591 + ], + "angle": 0, + "content": "Table 1 summarizes a comparison between popular active learning methods and our instance-wise supervision-level optimization. Each method is evaluated on its ability to incorporate uncertainty, diversity, multiple levels of supervision, and instance-wise adaptability, as defined in the caption. Most uncertainty-based and diversity-based AL methods do not consider multiple levels of supervision, focusing instead on selecting instances based on a single type of annotation. Although APFWA incorporates multiple levels of supervision, it lacks instance-wise optimization. In contrast, our approach is unique in that it optimizes the supervision level on an instance-wise basis, simultaneously considering data uncertainty, diversity, and the annotation cost associated with multiple levels of supervision." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.603, + 0.249, + 0.621 + ], + "angle": 0, + "content": "3. Problem setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Let \\( \\mathcal{D}_{\\mathrm{u}} \\) represent the unlabeled data pool, and let \\( T \\) denote the number of rounds in the active learning process, where \\( B \\) is the available budget in each round. In this setting, a set of instances is selected to be annotated with either full supervision or weak supervision, depending on the required level of detail and the associated cost. Specifically, we define \\( C_{\\mathrm{f}} \\) as the annotation cost per instance for full supervision, where precise labels are assigned, and \\( C_{\\mathrm{w}} \\) as the annotation cost per instance for weak supervision, which involves less precise, lower-cost labeling. These annotation costs, \\( C_{\\mathrm{f}} \\) and \\( C_{\\mathrm{w}} \\), are provided in advance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The process of batch selection and training is repeated for \\( T \\) rounds, allowing the model to progressively improve by iteratively selecting and annotating the most informative instances within each budget-constrained round. Once an instance is annotated, it is moved from the unlabeled pool \\( \\mathcal{D}_{\\mathrm{u}} \\) to either the fully supervised data pool \\( \\mathcal{D}_{\\mathrm{f}} \\) or the weakly supervised data pool \\( \\mathcal{D}_{\\mathrm{w}} \\), depending on the supervi" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.091, + 0.907, + 0.121 + ], + "angle": 0, + "content": "Algorithm 1 ISO: Instance-wise Supervision-level Optimization" + }, + { + "type": "algorithm", + "bbox": [ + 0.522, + 0.124, + 0.907, + 0.568 + ], + "angle": 0, + "content": "1: Inputs: Round \\(T\\), budget of each round \\(B\\), annotation cost \\(C_{\\mathrm{f}}, C_{\\mathrm{w}}\\), unlabeled data pool \\(\\mathcal{D}_{\\mathrm{u}}\\), feature extractor \\(f\\), classification head \\(h_{\\mathrm{f}}, h_{\\mathrm{w}}\\). \n2: Outputs: trained \\(f^{T}, h_{\\mathrm{f}}^{T}\\). \n3: for round \\(t = 1, \\ldots, T\\) do \n4: if \\(t\\) is 1 then \n5: Sample randomly instances from \\(\\mathcal{D}_{\\mathrm{u}}\\) within budget \\(B\\) to create an initial fully supervised dataset \\(\\mathcal{D}_{\\mathrm{f}}\\) and weakly supervised dataset \\(\\mathcal{D}_{\\mathrm{w}}\\) with an equal number of instances. \n6: else if \\(t > 1\\) then \n7: Compute the model improvements, \\(M_{\\mathrm{f}}^{t}, M_{\\mathrm{w}}^{t}\\), for each supervision using \\(\\mathcal{D}_{\\mathrm{f}}\\) and \\(\\mathcal{D}_{\\mathrm{w}}\\) by Eq. (3) and (4). \n8: for \\(\\pmb{x}\\) in \\(\\mathcal{D}_{\\mathrm{u}}\\) do \n9: 1. Compute the uncertainties, \\(u_{\\mathrm{f}}^{t}(\\pmb{x}), u_{\\mathrm{f}}^{t}(\\pmb{x})\\), for each supervision. \n10: 2. Compute the VCRs, \\(v_{\\mathrm{f}}^{t}(\\pmb{x}), v_{\\mathrm{w}}^{t}(\\pmb{x})\\), for each supervision by Eq. (1) and (2). \n11: 3. Compute the normalized features \\(\\tilde{f}^{t}(\\pmb{x})\\). \n12: end for \n13: Select the batch \\(\\mathcal{D}_{\\mathrm{f}}^{+}, \\mathcal{D}_{\\mathrm{w}}^{+}\\) using the VCRs, \\(v_{\\mathrm{f}}^{t}(\\pmb{x}), v_{\\mathrm{w}}^{t}(\\pmb{x})\\), and the normalized features \\(\\tilde{f}^{t}(\\pmb{x})\\) by Sec. 4.3. \n14: end if \n15: \\(\\mathcal{D}_{\\mathrm{f}} \\gets \\mathcal{D}_{\\mathrm{f}} \\cup \\mathcal{D}_{\\mathrm{f}}^{+}, \\mathcal{D}_{\\mathrm{u}} \\gets \\mathcal{D}_{\\mathrm{u}} \\setminus \\mathcal{D}_{\\mathrm{f}}^{+}\\) \n16: \\(\\mathcal{D}_{\\mathrm{w}} \\gets \\mathcal{D}_{\\mathrm{w}} \\cup \\mathcal{D}_{\\mathrm{w}}^{+}, \\mathcal{D}_{\\mathrm{u}} \\gets \\mathcal{D}_{\\mathrm{u}} \\setminus \\mathcal{D}_{\\mathrm{w}}^{+}\\) \n17: Train \\(f^{t}, h_{\\mathrm{f}}^{t}\\), and \\(h_{\\mathrm{w}}^{t}\\) using \\(\\mathcal{D}_{\\mathrm{f}}\\) and \\(\\mathcal{D}_{\\mathrm{w}}\\). \n18: end for" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.594, + 0.907, + 0.622 + ], + "angle": 0, + "content": "sion level chosen. These annotated instances are then used to train a classifier." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.907, + 0.744 + ], + "angle": 0, + "content": "The ultimate goal in this setting is to obtain the classifier's parameters to classify the exact class of input data. To train the classification network, both fully supervised data and weakly supervised data contribute to training the feature extractor in the network. Therefore, balancing fully and weakly supervised data in a cost-effective manner is crucial for optimizing the learning process within budget constraints." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.759, + 0.918, + 0.795 + ], + "angle": 0, + "content": "4. Proposed method: Instance-wise Supervision-level Optimization (ISO)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.625, + 0.817 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We propose an efficient algorithm to address the instance-wise supervision-level optimization problem within a fixed budget. Our approach comprises two primary components, as shown in Fig. 1. The first component calculates the value-to-cost ratio (VCR) for each instance-supervision" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4941" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "level pair, which indicates the expected improvement in model accuracy per unit of cost achieved through annotation for each instance. The VCR for each instance is determined by both the value of the supervision level and its uncertainty. This allows us to evaluate the cost-effectiveness of assigning different supervision levels to each instance individually. The second component performs batch selection within the budget \\( B \\), optimizing the set of instances based on their VCRs and diversity. This two-part approach enables our approach to maximize the value of annotations within a limited budget by balancing cost-efficiency and diversity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.282, + 0.409, + 0.297 + ], + "angle": 0, + "content": "4.2. Value-to-cost ratio (VCR) estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.305, + 0.484, + 0.425 + ], + "angle": 0, + "content": "In this section, we explain the calculation of the value-to-cost ratio (VCR) for each instance-supervision pair. An overview of the algorithm is provided in Algorithm 1. For an instance \\( \\pmb{x} \\) in the unlabeled pool \\( \\mathcal{D}_{\\mathrm{u}} \\), the VCRs \\( v_{\\mathrm{f}}^{t}(\\pmb{x}) \\) for full supervision and \\( v_{\\mathrm{w}}^{t}(\\pmb{x}) \\) for weak supervision at round \\( t \\) are determined by the uncertainty of the instance, the value, and cost for the two different-level labels, which are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.432, + 0.482, + 0.465 + ], + "angle": 0, + "content": "\\[\nv _ {\\mathrm {f}} ^ {t} (\\boldsymbol {x}) = \\frac {M _ {\\mathrm {f}} ^ {t} u _ {\\mathrm {f}} ^ {t} (\\boldsymbol {x})}{C _ {\\mathrm {f}}} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.466, + 0.482, + 0.5 + ], + "angle": 0, + "content": "\\[\nv _ {\\mathrm {w}} ^ {t} (\\boldsymbol {x}) = \\frac {M _ {\\mathrm {w}} ^ {t} u _ {\\mathrm {w}} ^ {t} (\\boldsymbol {x})}{C _ {\\mathrm {w}}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.483, + 0.644 + ], + "angle": 0, + "content": "where \\( C_{\\mathrm{f}}, C_{\\mathrm{w}} \\) represent the annotation costs, and \\( M_{\\mathrm{f}}^{t}, M_{\\mathrm{w}}^{t} \\) are the expected model improvement, where \\( *_{\\mathrm{f}}, *_{\\mathrm{w}} \\) correspond to full supervision and weak supervision, respectively. \\( u_{\\mathrm{f}}^{t}(\\boldsymbol{x}), u_{\\mathrm{w}}^{t}(\\boldsymbol{x}) \\) represents the uncertainty for each instance \\( \\boldsymbol{x} \\) in the unlabeled pool \\( D_{\\mathrm{u}} \\) at round \\( t \\), calculated based on model predictions. This formulation allows us to evaluate the cost-effectiveness of assigning either full or weak supervision to each instance, providing a basis for the batch selection strategy within a fixed budget." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The expected model improvements \\( M_{\\mathrm{f}}^{t} \\) for full supervision and \\( M_{\\mathrm{w}}^{t} \\) for weak supervision (i.e., the values for each supervision) are estimated by evaluating the improvement in model performance as the training data is incrementally increased. Consider the \\( t \\)-th annotation round, where fully supervised data \\( \\mathcal{D}_{\\mathrm{f}}^{t} \\) and weakly supervised data \\( \\mathcal{D}_{\\mathrm{w}}^{t} \\) are available at that time. To evaluate accuracy improvement, we divide the labeled data into \\( K \\) subsets and incrementally add training data. For fully supervised data, the weakly supervised data is kept fixed while subsets of the fully supervised data are added. The model performance \\( m_{\\mathrm{f},k} \\) represents the evaluation using the first \\( k \\) subsets of fully supervised training data, with \\( m_{\\mathrm{f},1} \\) being the performance when using just one subset. Similarly, we compute \\( m_{\\mathrm{w},k} \\) for the weakly supervised data, where the fully supervised data is fixed, and subsets of the weakly supervised data are incrementally added. Here, model performance is evaluated us" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "ing a small validation set randomly sampled within the budget of a single round. Then, a weighted average is applied for the improvement at each step, \\( m_{k+1}^t - m_k^t \\) to smooth the sequence of improvements. The reason for the weighted approach is to focus on the latest model behavior. To obtain the improvement per data point, each improvement value is normalized by dividing by \\( |\\mathcal{D}| / K \\), where \\( |\\mathcal{D}| \\) represents the number of instances. That is, \\( M_{\\mathrm{f}}^t \\) and \\( M_{\\mathrm{w}}^t \\) are expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.238, + 0.905, + 0.279 + ], + "angle": 0, + "content": "\\[\nM _ {\\mathrm {f}} ^ {t} = \\frac {\\sum_ {k = 1} ^ {K - 1} k \\cdot \\left(m _ {\\mathrm {f} , k + 1} ^ {t} - m _ {\\mathrm {f} , k} ^ {t}\\right)}{\\sum_ {k = 1} ^ {K - 1} k} \\cdot \\frac {K}{\\left| \\mathcal {D} _ {\\mathrm {f}} \\right|} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.281, + 0.905, + 0.323 + ], + "angle": 0, + "content": "\\[\nM _ {\\mathrm {w}} ^ {t} = \\frac {\\sum_ {k = 1} ^ {K - 1} k \\cdot \\left(m _ {\\mathrm {w} , k + 1} ^ {t} - m _ {\\mathrm {w} , k} ^ {t}\\right)}{\\sum_ {k = 1} ^ {K - 1} k} \\cdot \\frac {K}{\\left| \\mathcal {D} _ {\\mathrm {w}} \\right|} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.636 + ], + "angle": 0, + "content": "The uncertainty \\( u_{\\mathrm{f}}^{t}(\\pmb{x}) \\) and \\( u_{\\mathrm{w}}^{t}(\\pmb{x}) \\), corresponding to full and weak supervision, can be flexibly defined based on the specific task and dataset. For example, in the classification task, the entropy of the class prediction probabilities is adopted. If different functions are used to calculate uncertainty for full and weak supervision, there is a risk that the scales of \\( u_{\\mathrm{f}}^{t}(\\pmb{x}) \\) and \\( u_{\\mathrm{w}}^{t}(\\pmb{x}) \\) may differ significantly. To address this, we normalize each uncertainty value by converting it to a percentile score. In this transformation, the instance with the highest uncertainty is assigned a score of 1, the instance with the lowest uncertainty receives a score of 0, and the instance ranked at the 10th percentile in terms of uncertainty is assigned a score of 0.9. We use percentiles rather than dividing by the maximum uncertainty value because extreme outliers could result in an overly large maximum, causing most values to be close to zero. This percentile-based normalization ensures that the distribution of uncertainty values remains well-scaled, preserving relative differences across instances without distortion due to outliers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.905, + 0.666 + ], + "angle": 0, + "content": "4.3. Valuable and diverse batch selection using VCR" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.809 + ], + "angle": 0, + "content": "In this section, we describe the batch selection using the value-to-cost ratio (VCR) defined in the previous section. The most straightforward approach for batch selection is to greedily select instances with the highest VCR values. However, as noted in previous studies (e.g. [3, 19]), this approach can lead to selecting similar instances that have high VCR values, resulting in a loss of data diversity within the batch." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To address this, we propose a batch selection algorithm to select a diverse and valuable batch. In other words, we select a group of instances with high VCR values and diverse features. Specifically, we represent each instance as a vector using its VCR and feature vector. The vector differs depending on the type of label: \\( v_{\\mathrm{f}}(\\boldsymbol{x}) \\tilde{f}(\\boldsymbol{x}) \\) for full labels" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4942" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.482, + 0.151 + ], + "angle": 0, + "content": "and \\( v_{\\mathrm{w}}(\\pmb{x}) \\tilde{f}(\\pmb{x}) \\) for weak labels. Here, \\( \\tilde{f}(\\pmb{x}) \\) is a normalized feature vector (with a magnitude of 1) output by the feature extractor \\( f \\). The optimal set of instances is selected based on the vectors." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.153, + 0.484, + 0.439 + ], + "angle": 0, + "content": "We aim to select a set of instances that maximizes the area enclosed by the vertices of the selected vectors, as shown in the center of Figure 1. Here, the optimization parameter is which vectors (instances) to select and their label types. When the selected instances have high diversity and value, the area enclosed by the selected vectors becomes larger. We solve this optimization problem subject to budget constraints (i.e., the sum of the costs of the optimized supervision levels for instances is less than \\( B \\)). The optimization problem of selecting vectors and determining their supervision levels can be reformulated as a selection problem, where the goal is to choose the optimal set of vectors from all possible vectors, including both those with weak labels \\(\\{v_{\\mathrm{w}}(\\pmb{x})\\tilde{f} (\\pmb{x})|\\pmb{x}\\in \\mathcal{D}_{\\mathrm{u}}\\}\\) and those with full labels \\(\\{v_{\\mathrm{f}}(\\pmb {x})\\tilde{f} (\\pmb {x})|\\pmb {x}\\in \\mathcal{D}_{\\mathrm{u}}\\}\\). This allows the problem of choosing the type of label to be integrated into the vector selection process. This optimization problem can be solved by maximizing the determinant of a set of vectors subject to budget constraints." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.606 + ], + "angle": 0, + "content": "To perform the batch selection that maximizes the determinant of a matrix, one approach is to sample from a \\(k\\)-Determinantal Point Process (\\(k\\)-DPP) [13]. This method selects a batch containing \\(k\\) vectors with a probability proportional to the determinant of their Gram matrix. However, this differs from the batch selection in our approach in that we choose a set of vectors within a budget \\(B\\) rather than choosing \\(k\\) vectors. Therefore, we propose a new selection method, which is inspired by the work [3] using the \\(k\\)-means++ seeding algorithm [2] as an approximation algorithm for \\(k\\)-DPP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.484, + 0.804 + ], + "angle": 0, + "content": "In our batch selection method, the vector of the batch is sampled sequentially with a probability proportional to the square of the distance to the vector closest to the one selected so far until the budget \\( B \\) is exhausted. Specifically, if the selected vector corresponds to full supervision (i.e., a vector in the set \\( \\{v_{\\mathrm{f}}(\\pmb{x})\\tilde{f}(\\pmb{x})|\\pmb{x} \\in \\mathcal{D}_{\\mathrm{u}}\\} \\), the cost \\( C_{\\mathrm{f}} \\) is deducted from the budget, and it is added to the fully supervised data. If it corresponds to weak supervision (i.e., a vector in the set \\( \\{v_{\\mathrm{w}}(\\pmb{x})\\tilde{f}(\\pmb{x})|\\pmb{x} \\in \\mathcal{D}_{\\mathrm{u}}\\} \\)), the cost \\( C_{\\mathrm{w}} \\) is deducted from the budget, and it is added to the weakly supervised data. This process is repeated until the budget \\( B \\) is exhausted. By doing this, we can obtain a diverse and valuable batch." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.821, + 0.223, + 0.837 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.187, + 0.86 + ], + "angle": 0, + "content": "5.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.482, + 0.9 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of the proposed method, we utilized two datasets: CIFAR100 and Caltech-UCSD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "Birds-200-2011 (CUB200). The task is a multi-class classification problem, where each class (full label) belongs to a superclass, and the superclass label is provided as a weak label." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.905, + 0.258 + ], + "angle": 0, + "content": "CIFAR-100 is a dataset of natural images containing 100 classes that span a wide range of everyday objects and animals. Each class in CIFAR-100 is associated with a predefined superclass, grouping similar classes together under broader categories, resulting in 20 superclasses in total. The dataset consists of 60,000 images in total, with 50,000 images for training and 10,000 images for testing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.35 + ], + "angle": 0, + "content": "CUB200 [23] consists of 200 classes of bird species, with each class representing a specific bird. Following previous study [14], we defined superclasses based on the suffixes in the class names, resulting in 70 unique superclasses. The dataset contains 11,788 images, with 5,994 images designated for training and 5,794 images for testing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.359, + 0.715, + 0.374 + ], + "angle": 0, + "content": "5.2. Comparison methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.905, + 0.442 + ], + "angle": 0, + "content": "We compared our approach with eight conventional AL methods, ranging from standard single-supervision AL methods to the state-of-the-art AL method incorporating multiple supervision, as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.444, + 0.904, + 0.473 + ], + "angle": 0, + "content": "- Random: The simplest baseline that samples instances randomly from the unlabeled pool." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.477, + 0.905, + 0.537 + ], + "angle": 0, + "content": "- Margin [18]: A method that selects instances in ascending order of the difference between the highest and second-highest predicted class probabilities, prioritizing instances with minimal margins." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.54, + 0.905, + 0.584 + ], + "angle": 0, + "content": "- MaxConf [24]: An approach that selects instances in order of the lowest maximum predicted class probability, focusing on instances with the least confidence." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.588, + 0.904, + 0.633 + ], + "angle": 0, + "content": "- Entropy [24]: A method that selects instances according to the highest entropy in the predicted class probabilities, targeting instances with the greatest uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.636, + 0.905, + 0.681 + ], + "angle": 0, + "content": "- Coreset [19]: A diversity-based sampling method that selects instances farthest from the labeled data using a greedy algorithm to enhance diversity." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.684, + 0.905, + 0.743 + ], + "angle": 0, + "content": "- ALBL (Active Learning by Learning) [9]: A multiarmed bandit approach that selects between AL methods at each round. In this experiment, we utilized MaxConf and Coreset." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.747, + 0.905, + 0.822 + ], + "angle": 0, + "content": "- BADGE (Batch Active learning by Diverse Gradient Embeddings) [3]: A method that considers both predictive uncertainty and diversity by selecting instances with diverse and high-magnitude gradients in the gradient space." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.825, + 0.905, + 0.899 + ], + "angle": 0, + "content": "- APFWA (Adaptive Proportion of Full and Weak Annotations) [22]: A state-of-the-art AL method that incorporates both full and weak supervision, which dynamically determines the proportion of full and weak annotations within each batch." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.444, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4943" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "Since AL with both full and weak label annotation is a novel task, there is only one existing method for this task. Instead, we compare our approach with standard single-supervision AL methods as well as the state-of-the-art AL method. This comparison highlights the effectiveness of AL using weak labels. Additionally, a comparison with APFWA demonstrates the advantages of our approach for AL with both fully and weakly supervised labels. Furthermore, in Sec. 6.3, we also compare our approach with a simple combination of weak supervision and conventional AL methods to ensure a fair evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.263, + 0.483, + 0.34 + ], + "angle": 0, + "content": "These methods offer a range of approaches, from simple random sampling to advanced techniques that balance uncertainty and diversity, as well as methods leveraging weak supervision, enabling a comprehensive evaluation of our proposed approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.364, + 0.307, + 0.38 + ], + "angle": 0, + "content": "5.3. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.483, + 0.573 + ], + "angle": 0, + "content": "The network consists of a shared feature extractor \\( f \\) and two classification heads, \\( h_{\\mathrm{f}} \\) for full supervision and \\( h_{\\mathrm{w}} \\) for weak supervision. The encoder of ResNet18 [7] is used for the feature extractor \\( f \\), and a single linear layer is used for each classification head. For the network trained using both fully and weakly labeled data, we employ a two-stage training approach in which the feature extractor and each classification head are trained first with weakly supervised data and then with fully supervised data. Since both fully supervised and weakly supervised learning are classification tasks in this experiment, we used the cross-entropy loss. The learning rate was set to 0.001, and the optimizer was Adam [10]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.483, + 0.777 + ], + "angle": 0, + "content": "The settings for active learning were as follows. The number of rounds \\( T \\) was set to 5. The budget \\( B \\) of each round was set to 1,000 for CIFAR100 and to 500 for CUB200 since CUB200 had approximately 5,000 in training data, which was less than that of CIFAR100. The validation data with full supervision were prepared from training data, 1,000 for CIFAR1000 and 500 for CUB200, which are the same number and small amount as in budget \\( B \\). The cost of full supervision \\( C_{\\mathrm{f}} \\) was set to 1, and the cost of weak supervision was set to \\( \\frac{1}{2} \\). In addition, in Sec. 6.2, we conducted the experiments when changing \\( C_{\\mathrm{w}} \\) to \\( \\frac{1}{4} \\) and \\( \\frac{1}{8} \\) to investigate the impact of the difference on the cost of weak supervision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The number of partitions \\( K \\) for estimating the expected model improvements \\( M_{\\mathrm{f}}^{t} \\) and \\( M_{\\mathrm{w}}^{t} \\) was set to 5. As the performance metric of \\( m_{\\mathrm{f},k}^{t}, m_{\\mathrm{w},k}^{t} \\), we used the classification accuracy. To enhance stability, we repeated the estimation process three times with different random seeds and calculated the average. For uncertainty measurement, we adopted the simplest approach, Margin [18], to evaluate instance uncertainty." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.715, + 0.108 + ], + "angle": 0, + "content": "6. Experimental results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.146 + ], + "angle": 0, + "content": "6.1. Comparison with conventional active learning methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.155, + 0.905, + 0.215 + ], + "angle": 0, + "content": "The comparison results with conventional AL methods on CIFAR100 and CUB200 are shown in Fig. 2(a), (b), respectively. The graphs show the classification accuracy \\([\\%]\\) for each round \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.216, + 0.905, + 0.457 + ], + "angle": 0, + "content": "In Figure 2(a), the results on the CIFAR100 dataset demonstrate that our approach archives higher accuracy across all rounds compared to conventional methods, including random sampling, uncertainty-based AL method (Margin, MaxConf, Entropy), diversity-based AL method (Coreset), and other AL methods (ALBL, BADGE), which use only full supervision. This demonstrates that our approach effectively leverages weak supervision, achieving higher performance within the same budget constraints. For example, at round \\( t = 5 \\), our approach achieved over \\( 10\\% \\) higher accuracy compared to conventional AL methods that use only full supervision. Furthermore, while these methods require five rounds (with a budget of 5000) to reach an accuracy of approximately \\( 30\\% \\), our approach achieved the same or even higher accuracy in just three rounds, using only three-fifths of the budget." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.458, + 0.905, + 0.578 + ], + "angle": 0, + "content": "Our approach also surpasses APFWA, another AL approach that incorporates weak supervision by optimizing the proportions of weak and full supervision within the batch but does not focus on selecting instances. This result suggests that our approach more effectively selects cost-effective instances by dynamically considering instance-wise uncertainty and diversity, thereby maximizing the learning impact of weak supervision." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.58, + 0.905, + 0.685 + ], + "angle": 0, + "content": "Similarly, Figure 2(b) presents the results on the CUB200 dataset, where our approach also outperforms other AL methods across all rounds. The consistent improvement observed across both datasets indicates that our approach can dynamically optimize annotation strategies according to the dataset characteristics, achieving high adaptability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.821, + 0.713 + ], + "angle": 0, + "content": "6.2. Analysis for weak supervision costs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To demonstrate that our approach works at any cost of weak supervision and that the cheaper the cost of weak supervision, the greater the effect, we conducted experiments when changing the cost of weak supervision. Figure 3 shows the classification accuracy \\([\\%]\\) on CIFAR100 and CUB200 at different costs of weak supervision. In this experiment, the cost of full supervision \\(C_{\\mathrm{f}}\\) is fixed at 1, meaning that \\(C_{\\mathrm{w}}\\) represents the relative cost to full supervision. It is important to note that the cost \\(C_{\\mathrm{w}}\\) is defined before active learning begins and is not a hyperparameter of the proposed method. Imagine a scenario where the fee paid to a non-expert annotator, who can only assign rough class labels (weak la" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4944" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.09, + 0.434, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.144, + 0.285, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.29, + 0.229, + 0.302 + ], + "angle": 0, + "content": "(a) CIFAR100" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.144, + 0.468, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.29, + 0.417, + 0.302 + ], + "angle": 0, + "content": "(b) CUB200" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.308, + 0.484, + 0.366 + ], + "angle": 0, + "content": "Figure 2. Comparison of our approach with the conventional active learning methods. The figures show the classification accuracy \\([\\% ]\\) (\\(\\uparrow\\)) for each round \\(t\\) on CIFAR100 and CUB200. The cost of full supervision \\(C_{\\mathrm{f}}\\) is 1, and that of weak supervision is \\(\\frac{1}{2}\\)." + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.382, + 0.449, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.421, + 0.285, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.569, + 0.228, + 0.58 + ], + "angle": 0, + "content": "(a) CIFAR100" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.422, + 0.468, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.569, + 0.416, + 0.58 + ], + "angle": 0, + "content": "(b) CUB200" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.586, + 0.484, + 0.684 + ], + "angle": 0, + "content": "Figure 3. Classification accuracy \\([\\% ]\\) (\\(\\uparrow\\)) of the proposed method on CIFAR100 and CUB200 when the cost of weak supervision, \\(C_{\\mathrm{w}}\\) is \\(\\frac{1}{2}\\), \\(\\frac{1}{4}\\), and \\(\\frac{1}{8}\\). Note that the cost \\(C_{\\mathrm{w}}\\) is a value set by the user before active learning begins and is not a hyperparameter of the proposed method. This figure illustrates the effect of different weak supervision costs. For reference, we also include the results of conventional active learning methods (gray lines)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.749 + ], + "angle": 0, + "content": "bels), is lower than that of an expert annotator. The cost \\( C_{\\mathrm{w}} \\) varies for each task or dataset, corresponding to the rewards offered by crowdsourcing, etc." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In previous sections, we set the cost of weak supervision to \\( C_{\\mathrm{w}} = \\frac{1}{2} \\). Here, we investigate the effects of our approach when further reducing \\( C_{\\mathrm{w}} \\) to \\( \\frac{1}{4} \\) and \\( \\frac{1}{8} \\). As shown in Figure 3, for both datasets, the classification accuracy improves as the cost of weak supervision decreases. This improvement is due to the fact that the lower cost of weak supervision allows more data to be annotated within the same budget, thereby enhancing the training process. In other words, the cheaper the cost of weak supervision, the more impactful weak supervision and our approach become." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.905, + 0.122 + ], + "angle": 0, + "content": "6.3. Comparison with combined active learning and multiple supervision baseline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.13, + 0.905, + 0.235 + ], + "angle": 0, + "content": "In addition to comparing our approach with conventional active learning approaches, we conduct experiments to evaluate our approach against a straightforward combination of weak supervision and conventional active learning techniques. This baseline approach applies a fixed proportion of weak supervision within the budget, utilizing conventional active learning strategies within each supervision type." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.236, + 0.905, + 0.341 + ], + "angle": 0, + "content": "Specifically, in this combined method, weak supervision is allocated at a constant rate—\\(20\\%\\), \\(40\\%\\), \\(60\\%\\), and \\(80\\%\\) in the batch. Within each full and weak supervision, conventional AL methods are employed to select data. In this experiment, the active learning method for each full or weak supervision employed was sampling based on margin, which is used for the uncertainty of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.905, + 0.462 + ], + "angle": 0, + "content": "Figure 4 shows the comparison between our approach and baseline methods that combine active learning with weak supervision on the CIFAR100 dataset under different costs for weak supervision (\\( C_{\\mathrm{w}} = \\frac{1}{2}, \\frac{1}{4} \\), and \\( \\frac{1}{8} \\)). In each case, the baseline methods use fixed proportions of full and weak supervision within each batch, denoted as \\( \\mathrm{F}(x\\%) + \\mathrm{W}(y\\%) \\), where \\( x \\) and \\( y \\) represent the percentages of full and weak supervision, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.673 + ], + "angle": 0, + "content": "Across all weak supervision costs, our approach consistently achieves high accuracy. An important point to note is that the optimal proportion of full and weak supervision is unknown beforehand and varies depending on the cost of weak supervision. For example, when the cost of weak supervision \\( C_{\\mathrm{w}} \\) is \\( \\frac{1}{2} \\), the configuration with 60% full supervision and 40% weak supervision achieves the best accuracy among the baseline methods. However, when the cost \\( C_{\\mathrm{w}} \\) is reduced to \\( \\frac{1}{8} \\), a configuration with 20% full supervision and 80% weak supervision performs best. In contrast, our approach dynamically constructs optimal batches by optimizing both the supervision level and instance selection based on the cost-effectiveness obtained from the trained model at each round." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.664, + 0.699 + ], + "angle": 0, + "content": "6.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.905, + 0.766 + ], + "angle": 0, + "content": "We conducted ablation studies to demonstrate the effectiveness of instance-wise value (i.e., uncertainty) and the batch selection with diversity on CIFAR100 and CUB200, as shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.905, + 0.843 + ], + "angle": 0, + "content": "The first ablation variant, referred to as \"Ours w/o uncertainty,\" removes the instance-wise uncertainty component from the VCR calculation. In this variant, the VCR is determined solely based on the annotation cost and expected model improvement without taking into account the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.851, + 0.905, + 0.901 + ], + "angle": 0, + "content": "1APFWA [22] is a method that dynamically determines the proportion of full and weak annotations. It must use random instance sampling to estimate the optimal proportion, and as a result, it cannot be combined with the conventional instance-wise AL methods (e.g., margin-based sampling)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4945" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.271, + 0.089, + 0.728, + 0.119 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.125, + 0.341, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.292, + 0.268, + 0.306 + ], + "angle": 0, + "content": "(a) \\(C_{\\mathrm{w}} = 1 / 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.125, + 0.61, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.292, + 0.54, + 0.306 + ], + "angle": 0, + "content": "(b) \\(C_{\\mathrm{w}} = 1 / 4\\)" + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.125, + 0.88, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.292, + 0.811, + 0.306 + ], + "angle": 0, + "content": "(c) \\(C_{\\mathrm{w}} = 1 / 8\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 4. Comparison with combined active learning and multiple supervision baseline. The classification accuracy \\([\\%] (\\uparrow)\\) on CIFAR100 with fixed proportions of full and weak supervision within each batch. Each batch was sampled with fixed proportions, and conventional active learning methods were applied separately for full and weak supervision." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.382, + 0.473, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.551, + 0.229, + 0.563 + ], + "angle": 0, + "content": "(a) CIFAR100" + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.551, + 0.416, + 0.562 + ], + "angle": 0, + "content": "(b) CUB200" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.637 + ], + "angle": 0, + "content": "Figure 5. Ablation study comparing the proposed method without instance-wise uncertainty \\( u_{\\mathrm{f}}(\\pmb{x}) \\), \\( u_{\\mathrm{w}}(\\pmb{x}) \\) (Ours w/o uncertainty) and without batch selection considering diversity (Ours w/o diversity). The figures show the classification accuracy \\([\\%] (\\uparrow)\\) for each round \\( t \\) on CIFAR100 and CUB200." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.762 + ], + "angle": 0, + "content": "uncertainty of each instance. Consequently, the supervision level is not optimized for each instance individually. The results show that our approach outperforms this variant on both datasets. This effect is particularly prominent in the CUB200 dataset, demonstrating the benefit of optimizing \"instance-wise\" supervision-level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The second ablation variant, called 'Ours w/o diversity,\" involves a greedy approach to batch selection, where instances are selected in descending order of VCR without considering data diversity. Our approach, which incorporates diversity into the batch selection process, consistently outperforms this variant on both datasets. The impact of diversity-aware selection is especially evident in the CIFAR100 dataset, demonstrating that considering diversity in batch selection significantly enhances model performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.384, + 0.871, + 0.401 + ], + "angle": 0, + "content": "7. Conclusion, limitation, and future work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.606 + ], + "angle": 0, + "content": "In this paper, we proposed a novel active learning (AL) approach, called Instance-wise Supervision-level Optimization (ISO), which optimizes the supervision level (either full or weak annotation) for each instance to maximize annotation efficiency within a budget. By evaluating the uncertainty, diversity, and value-to-cost ratio (VCR) of each instance for full and weak supervision, our approach automatically determines the instances to be annotated and their optimal annotation level. Experiments on classification tasks demonstrated that ISO outperforms not only standard AL methods but also state-of-the-art AL methods with weak supervision, while effectively leveraging instance-specific information to improve both accuracy and cost-efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.909, + 0.833 + ], + "angle": 0, + "content": "A current limitation of this study is that the effectiveness of our ISO has been validated only in classification experiments. The concept of ISO is applicable to various tasks beyond classification. For example, ISO can be applied to object segmentation tasks, where precise yet costly object boundaries are assumed for full supervision, while simpler rectangular bounding boxes are used for weak supervision. Another current limitation is the assumption that supervision levels are only two. We can expand to deal with more supervision levels by introducing class levels other than exact classes and superclasses. In future work, we aim to demonstrate the full versatility of our approach by exploring these extensions across a broader range of tasks, reinforcing its potential as a scalable, cost-effective solution for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.853, + 0.908, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements: This work was supported by JSPS KAKENHI Grant Number JP23KJ1723, and JST, ACT-X Grant Number JPMJAX23CR, Japan." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4946" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Charu C Aggarwal, Xiangnan Kong, Quanquan Gu, Jiawei Han, and S Yu Philip. Active learning: A survey. In Data classification, pages 599-634. Chapman and Hall/CRC, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[2] D ARTHUR. k-means++: the advantages of careful seeding. In ACM-SIAM symposium on Discrete algorithms, pages 1027-1035, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[3] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. In International Conference on Learning Representations. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.283, + 0.484, + 0.338 + ], + "angle": 0, + "content": "[4] Shayok Chakraborty. Asking the right questions to the right users: Active learning with imperfect oracles. In Association for the Advancement of Artificial Intelligence, pages 3365-3372, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.339, + 0.484, + 0.394 + ], + "angle": 0, + "content": "[5] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International Conference on Machine Learning, pages 1183-1192, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.395, + 0.484, + 0.451 + ], + "angle": 0, + "content": "[6] Guy Hacohen, Avihu Dekel, and Daphna Weinshall. Active learning on a budget: Opposite strategies suit high and low budgets. In International Conference on Machine Learning, pages 8175-8195, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.484, + 0.493 + ], + "angle": 0, + "content": "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.484, + 0.548 + ], + "angle": 0, + "content": "[8] Neil Houlsby, Ferenc Huszár, Zoubin Ghahramani, and Máté Lengyel. Bayesian active learning for classification and preference learning. arXiv preprint arXiv:1112.5745, 2011. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.55, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[9] Wei-Ning Hsu and Hsuan-Tien Lin. Active learning by learning. In Association for the Advancement of Artificial Intelligence, 2015. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.593, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[10] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[11] Andreas Kirsch, Joost Van Amersfoort, and Yarin Gal. Batchbald: Efficient and diverse batch acquisition for deep bayesian active learning. Neural Information Processing Systems, 32, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.733 + ], + "angle": 0, + "content": "[12] Dani Kiyasseh, Tingting Zhu, and David A Clifton. Soqal: Selective oracle questioning for consistency based active learning of cardiac signals. In International Conference on Machine Learning, pages 11302-11340, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.774 + ], + "angle": 0, + "content": "[13] Alex Kulesza and Ben Taskar. k-dpps: Fixed-size determinantal point processes. In International Conference on Machine Learning, pages 1193-1200, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[14] Chao Lu and Yuexian Zou. Using coarse label constraint for fine-grained visual classification. In Conference on Multimedia Modeling, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[15] Andrew Kachites McCallum, Kamal Nigam, et al. Employing em and pool-based active learning for text classification. In International Conference on Machine Learning, pages 350-358, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[16] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "A survey of deep active learning. ACM computing surveys (CSUR), 54(9):1-40, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[17] Zeyu Ren, Shuihua Wang, and Yudong Zhang. Weakly supervised machine learning. CAAI Transactions on Intelligence Technology, 8(3):549-580, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[18] Dan Roth and Kevin Small. Margin-based active learning for structured output spaces. In European Conference on Machine Learning, pages 413-424, 2006. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[19] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.874, + 0.278 + ], + "angle": 0, + "content": "[20] Burr Settles. Active learning literature survey. 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.321 + ], + "angle": 0, + "content": "[21] Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell. Variational adversarial active learning. In International Conference on Computer Vision, pages 5972-5981, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[22] Javier Gamazo Tejero, Martin S Zinkernagel, Sebastian Wolf, Raphael Sznitman, and Pablo Márquez-Neila. Full or weak annotations? an adaptive strategy for budget-constrained annotation campaigns. In Conference on Computer Vision and Pattern Recognition, pages 11381-11391, 2023. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.904, + 0.433 + ], + "angle": 0, + "content": "[23] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Technical report, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.477 + ], + "angle": 0, + "content": "[24] Dan Wang and Yi Shang. A new active labeling method for deep learning. In International Joint Conference on Neural Networks, pages 112-119, 2014. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.478, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[25] Donggeun Yoo and In So Kweon. Learning loss for active learning. In Computer Vision and Pattern Recognition, pages 93-102, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[26] Man Zhang, Yong Zhou, Jiaqi Zhao, Yiyun Man, Bing Liu, and Rui Yao. A survey of semi-and weakly supervised semantic segmentation of images. Artificial Intelligence Review, 53:4259-4288, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.905, + 0.605 + ], + "angle": 0, + "content": "[27] Zhi-Hua Zhou. A brief introduction to weakly supervised learning. National science review, 5(1):44-53, 2018. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4947" + } + ] +] \ No newline at end of file diff --git a/2025/Instance-wise Supervision-level Optimization in Active Learning/6ab274ce-c563-45f8-9f14-82dfa36b95f4_origin.pdf b/2025/Instance-wise Supervision-level Optimization in Active Learning/6ab274ce-c563-45f8-9f14-82dfa36b95f4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e91756608e3acc71d399018082f07b590213a45d --- /dev/null +++ b/2025/Instance-wise Supervision-level Optimization in Active Learning/6ab274ce-c563-45f8-9f14-82dfa36b95f4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c133a1b1e4ce51d37825cf6401618c34b08be61c4ea2e97b4b66469483b3fd2 +size 672936 diff --git a/2025/Instance-wise Supervision-level Optimization in Active Learning/full.md b/2025/Instance-wise Supervision-level Optimization in Active Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7ea79a991ba7e459aa0517c1ee01215d84aa5800 --- /dev/null +++ b/2025/Instance-wise Supervision-level Optimization in Active Learning/full.md @@ -0,0 +1,284 @@ +# Instance-wise Supervision-level Optimization in Active Learning + +Shinnosuke Matsuo $^{1}$ , Riku Togashi $^{2}$ , Ryoma Bise $^{1}$ , Seiichi Uchida $^{1}$ , Masahiro Nomura $^{2}$ $^{1}$ Kyushu University, Japan $^{2}$ CyberAgent, Inc., Japan + +# Abstract + +Active learning (AL) is a label-efficient machine learning paradigm that focuses on selectively annotating high-value instances to maximize learning efficiency. Its effectiveness can be further enhanced by incorporating weak supervision, which uses rough yet cost-effective annotations instead of exact (i.e., full) but expensive annotations. We introduce a novel AL framework, Instance-wise Supervision-Level Optimization (ISO), which not only selects the instances to annotate but also determines their optimal annotation level within a fixed annotation budget. Its optimization criterion leverages the value-to-cost ratio (VCR) of each instance while ensuring diversity among the selected instances. In classification experiments, ISO consistently outperforms traditional AL methods and surpasses a state-of-the-art AL approach that combines full and weak supervision, achieving higher accuracy at a lower overall cost. This code is available at https://github.com/matsuoshinnosuke/ISOAL. + +# 1. Introduction + +Machine learning technology has achieved remarkable success in recent years across various fields, including computer vision, natural language processing, and robotics. However, a major challenge remains: creating labeled datasets for training models often requires significant time and cost, especially in specialized domains. + +To address this issue, active learning [1, 3, 6, 8, 11, 16, 18-21, 24] has been widely studied to improve model performance under limited budgets. In active learning, instances (data samples) for annotation are selected iteratively from an unlabeled dataset, followed by annotation and model retraining. This repeated cycle improves the model's accuracy while progressively minimizing annotation costs. A common approach in active learning is uncertainty-based sampling [3, 5, 8, 9, 18, 22, 24, 25], where instances with high uncertainty, such as low confident of the model's prediction, are selected for annotation, as these are expected to contribute most effectively to model performance. + +Another approach to reduce annotation costs is weak su + +pervision [17, 26, 27]. For classification tasks, a weak supervision approach is often designed to just attach rough class labels to individual instances. For example, instead of attaching the exact class label "song sparrow" or "house sparrow" to an image instance, annotators can attach a rough class label "sparrows." This approach drastically reduces annotation costs (budgets) because the fee paid to a non-expert annotator, who can only assign rough class labels rather than exact ones via a crowdsourcing service, is lower than that of an expert annotator. Hereafter, we call rough classes as superclasses; therefore, "sparrows" is a superclass of the exact class "song sparrow." + +In general, weak supervision approaches still need instances with full supervision, that is, instances with the exact class labels. This is simply because it is impossible to determine the class boundary between, "sparrows" is a superclass of "song sparrow" without the instances with these exact class labels. Consequently, we need to balance weak supervision and full supervision under a pre-specified total annotation budget. In other words, we need to optimize the supervision levels of individual instances. Since the total annotation budget is pre-specified as a constraint, the number of instances with full supervision will be limited. More formally, this supervision-level optimization task can be seen as an optimal resource allocation problem with the constraint on available resources. + +This paper proposes a label-efficient machine learning approach, called Instance-wise Supervision-level Optimization (ISO) for active learning, as shown in Fig. 1. Roughly speaking, the proposed approach is a sophisticated hybrid of weak supervision and active learning. Different from the state-of-the-art hybrid approach [22] (reviewed in Section 2), our approach has a novel mechanism that automatically determines the optimal supervision levels of individual instances. Specifically, as shown in Fig. 1, all unlabeled instances at a certain round are automatically divided into three categories: instances for full supervision (with exact class annotation), instances for weak supervision (with superclass annotation), and instances that remain unlabeled. The instances selected in the first two categories are considered to be selected for active learning; annotators, therefore, attach appropriate labels to these instances according + +![](images/ff300c4be666d44a776b9ea022ac3ca2c19016ddf4b63f783033e9cf76876faa.jpg) +Figure 1. Overview of Instance-wise Supervision-level Optimization (ISO) for active learning. The proposed approach combines weak and full supervision within an active learning framework to maximize annotation efficiency under a fixed budget. First, each instance $\pmb{x}$ in the unlabeled data pool is evaluated for its value-to-cost ratio $(v_{\mathrm{f}}(\pmb{x}))$ for full supervision and $v_{\mathrm{w}}(\pmb{x})$ for weak supervision). Based on these values and data diversity, instances are selected within the budget constraints for weak supervision (with superclass labels) or full supervision (with exact class labels). This adaptive allocation of supervision-level enables optimal use of resources within the budget constraints in active learning. + +to their supervision levels. + +In our experiments, the proposed method consistently outperformed existing approaches across multiple datasets, demonstrating a more efficient use of the annotation budget. Our approach achieved the same accuracy as the comparative methods while using only three-fifths of the budget by incorporating weak labels when the cost of weak label annotation is half that of fully supervised annotation. Furthermore, in comparison with the state-of-the-art method that combines active learning and multiple supervision, our approach achieved significantly better cost efficiency by leveraging instance-specific information. + +The main contributions of this paper are summarized as follows: + +- We introduce a novel active learning framework that optimizes the supervision level for each instance, allowing for dynamic allocation of weak and full annotations based on budget constraints. +- We propose an efficient algorithm that assigns the optimal supervision level to each instance by considering value-to-cost ratios and data diversity, enhancing the cost-effectiveness of the annotation process. +- We validate the effectiveness of our approach on classification tasks, using superclasses as weak supervision, and demonstrate its superiority over conventional AL methods and the state-of-the-art AL method that incorporates both full and weak supervision. + +# 2. Related work + +Active learning (AL) [1, 16, 20] is a technique aimed at maximizing model performance under a limited annotation budget by strategically selecting informative instances for annotation. A widely adopted approach in active learning is + +uncertainty-based sampling [3, 5, 8, 9, 18, 22, 24, 25]. For example, [18] selects instances with small margins, prioritizing instances for which the model has low confidence. Another common approach focuses on maximizing diversity within selected instances [6, 15, 19, 21], which aims to cover a broad range of the data distribution. For instance, [19] selects a representative core set of instances to maximize diversity. Moreover, there are methods that combine uncertainty and diversity-based strategies to leverage the advantages of both approaches [3, 11]. Additionally, some AL methods [4, 12] address the setting where different annotators have varying accuracy levels. These methods aim to jointly optimize both the selection of data for annotation and the choice of annotators to minimize label noise. However, all of these approaches assume a single supervision-level and do not incorporate multiple supervision-levels. + +Another approach to reduce annotation costs is weakly supervised learning (WSL) [17, 26, 27], where less detailed or approximate labels instead of fully supervised labels are utilized. In many real-world applications, obtaining fully supervised labels can be expensive and time-consuming. To address this, WSL methods use weak supervision, which is typically easier and cheaper to obtain. For example, in classification tasks, weak supervision can be provided by superclasses such as categories, and in segmentation tasks, it can be provided by bounding boxes, allowing faster labeling with lower costs. + +A hybrid approach between AL and WSL, where the proportion of full and weak supervision is adaptively determined, has been proposed in [22]. Although this approach aims to optimize budget use by setting an ideal balance of weak and full supervision, it applies random sampling after setting the proportion without instance-level considera + +Table 1. Comparison of conventional active learning methods. "Uncertainty" (Unc.) refers to whether a method incorporates uncertainty-based instance selection, such as selecting instances with minimal margins. "Diversity" (Div.) indicates if the method considers data diversity to maximize the representativeness of selected instances. "Multiple supervision" (Mul.) refers to the capability to utilize multiple levels of supervision, such as both weak and full supervision. "Instance-wise" (Ins.) represents the ability to choose supervision for each instance adaptively. + +
Margin, etc. [8, 18, 24]Coreset, etc. [6, 19, 21]BADGE, etc. [3, 11]APFWA [22]Ours
Unc.✓×✓×✓
Div.×✓✓×✓
Mul.×××✓✓
Ins.✓✓✓×✓
+ +tions. To our knowledge, no method exists that dynamically assigns full and weak supervision for each instance. + +Table 1 summarizes a comparison between popular active learning methods and our instance-wise supervision-level optimization. Each method is evaluated on its ability to incorporate uncertainty, diversity, multiple levels of supervision, and instance-wise adaptability, as defined in the caption. Most uncertainty-based and diversity-based AL methods do not consider multiple levels of supervision, focusing instead on selecting instances based on a single type of annotation. Although APFWA incorporates multiple levels of supervision, it lacks instance-wise optimization. In contrast, our approach is unique in that it optimizes the supervision level on an instance-wise basis, simultaneously considering data uncertainty, diversity, and the annotation cost associated with multiple levels of supervision. + +# 3. Problem setting + +Let $\mathcal{D}_{\mathrm{u}}$ represent the unlabeled data pool, and let $T$ denote the number of rounds in the active learning process, where $B$ is the available budget in each round. In this setting, a set of instances is selected to be annotated with either full supervision or weak supervision, depending on the required level of detail and the associated cost. Specifically, we define $C_{\mathrm{f}}$ as the annotation cost per instance for full supervision, where precise labels are assigned, and $C_{\mathrm{w}}$ as the annotation cost per instance for weak supervision, which involves less precise, lower-cost labeling. These annotation costs, $C_{\mathrm{f}}$ and $C_{\mathrm{w}}$ , are provided in advance. + +The process of batch selection and training is repeated for $T$ rounds, allowing the model to progressively improve by iteratively selecting and annotating the most informative instances within each budget-constrained round. Once an instance is annotated, it is moved from the unlabeled pool $\mathcal{D}_{\mathrm{u}}$ to either the fully supervised data pool $\mathcal{D}_{\mathrm{f}}$ or the weakly supervised data pool $\mathcal{D}_{\mathrm{w}}$ , depending on the supervi + +Algorithm 1 ISO: Instance-wise Supervision-level Optimization +1: Inputs: Round $T$ , budget of each round $B$ , annotation cost $C_{\mathrm{f}}, C_{\mathrm{w}}$ , unlabeled data pool $\mathcal{D}_{\mathrm{u}}$ , feature extractor $f$ , classification head $h_{\mathrm{f}}, h_{\mathrm{w}}$ . +2: Outputs: trained $f^{T}, h_{\mathrm{f}}^{T}$ . +3: for round $t = 1, \ldots, T$ do +4: if $t$ is 1 then +5: Sample randomly instances from $\mathcal{D}_{\mathrm{u}}$ within budget $B$ to create an initial fully supervised dataset $\mathcal{D}_{\mathrm{f}}$ and weakly supervised dataset $\mathcal{D}_{\mathrm{w}}$ with an equal number of instances. +6: else if $t > 1$ then +7: Compute the model improvements, $M_{\mathrm{f}}^{t}, M_{\mathrm{w}}^{t}$ , for each supervision using $\mathcal{D}_{\mathrm{f}}$ and $\mathcal{D}_{\mathrm{w}}$ by Eq. (3) and (4). +8: for $\pmb{x}$ in $\mathcal{D}_{\mathrm{u}}$ do +9: 1. Compute the uncertainties, $u_{\mathrm{f}}^{t}(\pmb{x}), u_{\mathrm{f}}^{t}(\pmb{x})$ , for each supervision. +10: 2. Compute the VCRs, $v_{\mathrm{f}}^{t}(\pmb{x}), v_{\mathrm{w}}^{t}(\pmb{x})$ , for each supervision by Eq. (1) and (2). +11: 3. Compute the normalized features $\tilde{f}^{t}(\pmb{x})$ . +12: end for +13: Select the batch $\mathcal{D}_{\mathrm{f}}^{+}, \mathcal{D}_{\mathrm{w}}^{+}$ using the VCRs, $v_{\mathrm{f}}^{t}(\pmb{x}), v_{\mathrm{w}}^{t}(\pmb{x})$ , and the normalized features $\tilde{f}^{t}(\pmb{x})$ by Sec. 4.3. +14: end if +15: $\mathcal{D}_{\mathrm{f}} \gets \mathcal{D}_{\mathrm{f}} \cup \mathcal{D}_{\mathrm{f}}^{+}, \mathcal{D}_{\mathrm{u}} \gets \mathcal{D}_{\mathrm{u}} \setminus \mathcal{D}_{\mathrm{f}}^{+}$ +16: $\mathcal{D}_{\mathrm{w}} \gets \mathcal{D}_{\mathrm{w}} \cup \mathcal{D}_{\mathrm{w}}^{+}, \mathcal{D}_{\mathrm{u}} \gets \mathcal{D}_{\mathrm{u}} \setminus \mathcal{D}_{\mathrm{w}}^{+}$ +17: Train $f^{t}, h_{\mathrm{f}}^{t}$ , and $h_{\mathrm{w}}^{t}$ using $\mathcal{D}_{\mathrm{f}}$ and $\mathcal{D}_{\mathrm{w}}$ . +18: end for + +sion level chosen. These annotated instances are then used to train a classifier. + +The ultimate goal in this setting is to obtain the classifier's parameters to classify the exact class of input data. To train the classification network, both fully supervised data and weakly supervised data contribute to training the feature extractor in the network. Therefore, balancing fully and weakly supervised data in a cost-effective manner is crucial for optimizing the learning process within budget constraints. + +# 4. Proposed method: Instance-wise Supervision-level Optimization (ISO) + +# 4.1. Overview + +We propose an efficient algorithm to address the instance-wise supervision-level optimization problem within a fixed budget. Our approach comprises two primary components, as shown in Fig. 1. The first component calculates the value-to-cost ratio (VCR) for each instance-supervision + +level pair, which indicates the expected improvement in model accuracy per unit of cost achieved through annotation for each instance. The VCR for each instance is determined by both the value of the supervision level and its uncertainty. This allows us to evaluate the cost-effectiveness of assigning different supervision levels to each instance individually. The second component performs batch selection within the budget $B$ , optimizing the set of instances based on their VCRs and diversity. This two-part approach enables our approach to maximize the value of annotations within a limited budget by balancing cost-efficiency and diversity. + +# 4.2. Value-to-cost ratio (VCR) estimation + +In this section, we explain the calculation of the value-to-cost ratio (VCR) for each instance-supervision pair. An overview of the algorithm is provided in Algorithm 1. For an instance $\pmb{x}$ in the unlabeled pool $\mathcal{D}_{\mathrm{u}}$ , the VCRs $v_{\mathrm{f}}^{t}(\pmb{x})$ for full supervision and $v_{\mathrm{w}}^{t}(\pmb{x})$ for weak supervision at round $t$ are determined by the uncertainty of the instance, the value, and cost for the two different-level labels, which are defined as: + +$$ +v _ {\mathrm {f}} ^ {t} (\boldsymbol {x}) = \frac {M _ {\mathrm {f}} ^ {t} u _ {\mathrm {f}} ^ {t} (\boldsymbol {x})}{C _ {\mathrm {f}}} \tag {1} +$$ + +$$ +v _ {\mathrm {w}} ^ {t} (\boldsymbol {x}) = \frac {M _ {\mathrm {w}} ^ {t} u _ {\mathrm {w}} ^ {t} (\boldsymbol {x})}{C _ {\mathrm {w}}} \tag {2} +$$ + +where $C_{\mathrm{f}}, C_{\mathrm{w}}$ represent the annotation costs, and $M_{\mathrm{f}}^{t}, M_{\mathrm{w}}^{t}$ are the expected model improvement, where $*_{\mathrm{f}}, *_{\mathrm{w}}$ correspond to full supervision and weak supervision, respectively. $u_{\mathrm{f}}^{t}(\boldsymbol{x}), u_{\mathrm{w}}^{t}(\boldsymbol{x})$ represents the uncertainty for each instance $\boldsymbol{x}$ in the unlabeled pool $D_{\mathrm{u}}$ at round $t$ , calculated based on model predictions. This formulation allows us to evaluate the cost-effectiveness of assigning either full or weak supervision to each instance, providing a basis for the batch selection strategy within a fixed budget. + +The expected model improvements $M_{\mathrm{f}}^{t}$ for full supervision and $M_{\mathrm{w}}^{t}$ for weak supervision (i.e., the values for each supervision) are estimated by evaluating the improvement in model performance as the training data is incrementally increased. Consider the $t$ -th annotation round, where fully supervised data $\mathcal{D}_{\mathrm{f}}^{t}$ and weakly supervised data $\mathcal{D}_{\mathrm{w}}^{t}$ are available at that time. To evaluate accuracy improvement, we divide the labeled data into $K$ subsets and incrementally add training data. For fully supervised data, the weakly supervised data is kept fixed while subsets of the fully supervised data are added. The model performance $m_{\mathrm{f},k}$ represents the evaluation using the first $k$ subsets of fully supervised training data, with $m_{\mathrm{f},1}$ being the performance when using just one subset. Similarly, we compute $m_{\mathrm{w},k}$ for the weakly supervised data, where the fully supervised data is fixed, and subsets of the weakly supervised data are incrementally added. Here, model performance is evaluated us + +ing a small validation set randomly sampled within the budget of a single round. Then, a weighted average is applied for the improvement at each step, $m_{k+1}^t - m_k^t$ to smooth the sequence of improvements. The reason for the weighted approach is to focus on the latest model behavior. To obtain the improvement per data point, each improvement value is normalized by dividing by $|\mathcal{D}| / K$ , where $|\mathcal{D}|$ represents the number of instances. That is, $M_{\mathrm{f}}^t$ and $M_{\mathrm{w}}^t$ are expressed as follows: + +$$ +M _ {\mathrm {f}} ^ {t} = \frac {\sum_ {k = 1} ^ {K - 1} k \cdot \left(m _ {\mathrm {f} , k + 1} ^ {t} - m _ {\mathrm {f} , k} ^ {t}\right)}{\sum_ {k = 1} ^ {K - 1} k} \cdot \frac {K}{\left| \mathcal {D} _ {\mathrm {f}} \right|} \tag {3} +$$ + +$$ +M _ {\mathrm {w}} ^ {t} = \frac {\sum_ {k = 1} ^ {K - 1} k \cdot \left(m _ {\mathrm {w} , k + 1} ^ {t} - m _ {\mathrm {w} , k} ^ {t}\right)}{\sum_ {k = 1} ^ {K - 1} k} \cdot \frac {K}{\left| \mathcal {D} _ {\mathrm {w}} \right|} \tag {4} +$$ + +The uncertainty $u_{\mathrm{f}}^{t}(\pmb{x})$ and $u_{\mathrm{w}}^{t}(\pmb{x})$ , corresponding to full and weak supervision, can be flexibly defined based on the specific task and dataset. For example, in the classification task, the entropy of the class prediction probabilities is adopted. If different functions are used to calculate uncertainty for full and weak supervision, there is a risk that the scales of $u_{\mathrm{f}}^{t}(\pmb{x})$ and $u_{\mathrm{w}}^{t}(\pmb{x})$ may differ significantly. To address this, we normalize each uncertainty value by converting it to a percentile score. In this transformation, the instance with the highest uncertainty is assigned a score of 1, the instance with the lowest uncertainty receives a score of 0, and the instance ranked at the 10th percentile in terms of uncertainty is assigned a score of 0.9. We use percentiles rather than dividing by the maximum uncertainty value because extreme outliers could result in an overly large maximum, causing most values to be close to zero. This percentile-based normalization ensures that the distribution of uncertainty values remains well-scaled, preserving relative differences across instances without distortion due to outliers. + +# 4.3. Valuable and diverse batch selection using VCR + +In this section, we describe the batch selection using the value-to-cost ratio (VCR) defined in the previous section. The most straightforward approach for batch selection is to greedily select instances with the highest VCR values. However, as noted in previous studies (e.g. [3, 19]), this approach can lead to selecting similar instances that have high VCR values, resulting in a loss of data diversity within the batch. + +To address this, we propose a batch selection algorithm to select a diverse and valuable batch. In other words, we select a group of instances with high VCR values and diverse features. Specifically, we represent each instance as a vector using its VCR and feature vector. The vector differs depending on the type of label: $v_{\mathrm{f}}(\boldsymbol{x}) \tilde{f}(\boldsymbol{x})$ for full labels + +and $v_{\mathrm{w}}(\pmb{x}) \tilde{f}(\pmb{x})$ for weak labels. Here, $\tilde{f}(\pmb{x})$ is a normalized feature vector (with a magnitude of 1) output by the feature extractor $f$ . The optimal set of instances is selected based on the vectors. + +We aim to select a set of instances that maximizes the area enclosed by the vertices of the selected vectors, as shown in the center of Figure 1. Here, the optimization parameter is which vectors (instances) to select and their label types. When the selected instances have high diversity and value, the area enclosed by the selected vectors becomes larger. We solve this optimization problem subject to budget constraints (i.e., the sum of the costs of the optimized supervision levels for instances is less than $B$ ). The optimization problem of selecting vectors and determining their supervision levels can be reformulated as a selection problem, where the goal is to choose the optimal set of vectors from all possible vectors, including both those with weak labels $\{v_{\mathrm{w}}(\pmb{x})\tilde{f} (\pmb{x})|\pmb{x}\in \mathcal{D}_{\mathrm{u}}\}$ and those with full labels $\{v_{\mathrm{f}}(\pmb {x})\tilde{f} (\pmb {x})|\pmb {x}\in \mathcal{D}_{\mathrm{u}}\}$ . This allows the problem of choosing the type of label to be integrated into the vector selection process. This optimization problem can be solved by maximizing the determinant of a set of vectors subject to budget constraints. + +To perform the batch selection that maximizes the determinant of a matrix, one approach is to sample from a $k$ -Determinantal Point Process ( $k$ -DPP) [13]. This method selects a batch containing $k$ vectors with a probability proportional to the determinant of their Gram matrix. However, this differs from the batch selection in our approach in that we choose a set of vectors within a budget $B$ rather than choosing $k$ vectors. Therefore, we propose a new selection method, which is inspired by the work [3] using the $k$ -means++ seeding algorithm [2] as an approximation algorithm for $k$ -DPP. + +In our batch selection method, the vector of the batch is sampled sequentially with a probability proportional to the square of the distance to the vector closest to the one selected so far until the budget $B$ is exhausted. Specifically, if the selected vector corresponds to full supervision (i.e., a vector in the set $\{v_{\mathrm{f}}(\pmb{x})\tilde{f}(\pmb{x})|\pmb{x} \in \mathcal{D}_{\mathrm{u}}\}$ , the cost $C_{\mathrm{f}}$ is deducted from the budget, and it is added to the fully supervised data. If it corresponds to weak supervision (i.e., a vector in the set $\{v_{\mathrm{w}}(\pmb{x})\tilde{f}(\pmb{x})|\pmb{x} \in \mathcal{D}_{\mathrm{u}}\}$ ), the cost $C_{\mathrm{w}}$ is deducted from the budget, and it is added to the weakly supervised data. This process is repeated until the budget $B$ is exhausted. By doing this, we can obtain a diverse and valuable batch. + +# 5. Experiments + +# 5.1. Dataset + +To demonstrate the effectiveness of the proposed method, we utilized two datasets: CIFAR100 and Caltech-UCSD + +Birds-200-2011 (CUB200). The task is a multi-class classification problem, where each class (full label) belongs to a superclass, and the superclass label is provided as a weak label. + +CIFAR-100 is a dataset of natural images containing 100 classes that span a wide range of everyday objects and animals. Each class in CIFAR-100 is associated with a predefined superclass, grouping similar classes together under broader categories, resulting in 20 superclasses in total. The dataset consists of 60,000 images in total, with 50,000 images for training and 10,000 images for testing + +CUB200 [23] consists of 200 classes of bird species, with each class representing a specific bird. Following previous study [14], we defined superclasses based on the suffixes in the class names, resulting in 70 unique superclasses. The dataset contains 11,788 images, with 5,994 images designated for training and 5,794 images for testing. + +# 5.2. Comparison methods + +We compared our approach with eight conventional AL methods, ranging from standard single-supervision AL methods to the state-of-the-art AL method incorporating multiple supervision, as follows: + +- Random: The simplest baseline that samples instances randomly from the unlabeled pool. +- Margin [18]: A method that selects instances in ascending order of the difference between the highest and second-highest predicted class probabilities, prioritizing instances with minimal margins. +- MaxConf [24]: An approach that selects instances in order of the lowest maximum predicted class probability, focusing on instances with the least confidence. +- Entropy [24]: A method that selects instances according to the highest entropy in the predicted class probabilities, targeting instances with the greatest uncertainty. +- Coreset [19]: A diversity-based sampling method that selects instances farthest from the labeled data using a greedy algorithm to enhance diversity. +- ALBL (Active Learning by Learning) [9]: A multiarmed bandit approach that selects between AL methods at each round. In this experiment, we utilized MaxConf and Coreset. +- BADGE (Batch Active learning by Diverse Gradient Embeddings) [3]: A method that considers both predictive uncertainty and diversity by selecting instances with diverse and high-magnitude gradients in the gradient space. +- APFWA (Adaptive Proportion of Full and Weak Annotations) [22]: A state-of-the-art AL method that incorporates both full and weak supervision, which dynamically determines the proportion of full and weak annotations within each batch. + +Since AL with both full and weak label annotation is a novel task, there is only one existing method for this task. Instead, we compare our approach with standard single-supervision AL methods as well as the state-of-the-art AL method. This comparison highlights the effectiveness of AL using weak labels. Additionally, a comparison with APFWA demonstrates the advantages of our approach for AL with both fully and weakly supervised labels. Furthermore, in Sec. 6.3, we also compare our approach with a simple combination of weak supervision and conventional AL methods to ensure a fair evaluation. + +These methods offer a range of approaches, from simple random sampling to advanced techniques that balance uncertainty and diversity, as well as methods leveraging weak supervision, enabling a comprehensive evaluation of our proposed approach. + +# 5.3. Implementation details + +The network consists of a shared feature extractor $f$ and two classification heads, $h_{\mathrm{f}}$ for full supervision and $h_{\mathrm{w}}$ for weak supervision. The encoder of ResNet18 [7] is used for the feature extractor $f$ , and a single linear layer is used for each classification head. For the network trained using both fully and weakly labeled data, we employ a two-stage training approach in which the feature extractor and each classification head are trained first with weakly supervised data and then with fully supervised data. Since both fully supervised and weakly supervised learning are classification tasks in this experiment, we used the cross-entropy loss. The learning rate was set to 0.001, and the optimizer was Adam [10]. + +The settings for active learning were as follows. The number of rounds $T$ was set to 5. The budget $B$ of each round was set to 1,000 for CIFAR100 and to 500 for CUB200 since CUB200 had approximately 5,000 in training data, which was less than that of CIFAR100. The validation data with full supervision were prepared from training data, 1,000 for CIFAR1000 and 500 for CUB200, which are the same number and small amount as in budget $B$ . The cost of full supervision $C_{\mathrm{f}}$ was set to 1, and the cost of weak supervision was set to $\frac{1}{2}$ . In addition, in Sec. 6.2, we conducted the experiments when changing $C_{\mathrm{w}}$ to $\frac{1}{4}$ and $\frac{1}{8}$ to investigate the impact of the difference on the cost of weak supervision. + +The number of partitions $K$ for estimating the expected model improvements $M_{\mathrm{f}}^{t}$ and $M_{\mathrm{w}}^{t}$ was set to 5. As the performance metric of $m_{\mathrm{f},k}^{t}, m_{\mathrm{w},k}^{t}$ , we used the classification accuracy. To enhance stability, we repeated the estimation process three times with different random seeds and calculated the average. For uncertainty measurement, we adopted the simplest approach, Margin [18], to evaluate instance uncertainty. + +# 6. Experimental results + +# 6.1. Comparison with conventional active learning methods + +The comparison results with conventional AL methods on CIFAR100 and CUB200 are shown in Fig. 2(a), (b), respectively. The graphs show the classification accuracy $[\%]$ for each round $t$ . + +In Figure 2(a), the results on the CIFAR100 dataset demonstrate that our approach archives higher accuracy across all rounds compared to conventional methods, including random sampling, uncertainty-based AL method (Margin, MaxConf, Entropy), diversity-based AL method (Coreset), and other AL methods (ALBL, BADGE), which use only full supervision. This demonstrates that our approach effectively leverages weak supervision, achieving higher performance within the same budget constraints. For example, at round $t = 5$ , our approach achieved over $10\%$ higher accuracy compared to conventional AL methods that use only full supervision. Furthermore, while these methods require five rounds (with a budget of 5000) to reach an accuracy of approximately $30\%$ , our approach achieved the same or even higher accuracy in just three rounds, using only three-fifths of the budget. + +Our approach also surpasses APFWA, another AL approach that incorporates weak supervision by optimizing the proportions of weak and full supervision within the batch but does not focus on selecting instances. This result suggests that our approach more effectively selects cost-effective instances by dynamically considering instance-wise uncertainty and diversity, thereby maximizing the learning impact of weak supervision. + +Similarly, Figure 2(b) presents the results on the CUB200 dataset, where our approach also outperforms other AL methods across all rounds. The consistent improvement observed across both datasets indicates that our approach can dynamically optimize annotation strategies according to the dataset characteristics, achieving high adaptability. + +# 6.2. Analysis for weak supervision costs + +To demonstrate that our approach works at any cost of weak supervision and that the cheaper the cost of weak supervision, the greater the effect, we conducted experiments when changing the cost of weak supervision. Figure 3 shows the classification accuracy $[\%]$ on CIFAR100 and CUB200 at different costs of weak supervision. In this experiment, the cost of full supervision $C_{\mathrm{f}}$ is fixed at 1, meaning that $C_{\mathrm{w}}$ represents the relative cost to full supervision. It is important to note that the cost $C_{\mathrm{w}}$ is defined before active learning begins and is not a hyperparameter of the proposed method. Imagine a scenario where the fee paid to a non-expert annotator, who can only assign rough class labels (weak la + +![](images/4dc4e1235dcb03007319bdab6df23db603cd0ab30ba6cb085757df61fff12b32.jpg) + +![](images/8f670618b8ff430ba3888a047569a930906347cef3afd0a13a6789333a04e0b5.jpg) +(a) CIFAR100 + +![](images/aa75b502bca009497eff13610fa83b8cd18c3a779816ea4068bef82b553d938f.jpg) +(b) CUB200 + +![](images/f5802bbf2b719fe0d35273cf6b22a7a2a2b356ba433d9e89d03ac7dbf752fd9d.jpg) +Figure 2. Comparison of our approach with the conventional active learning methods. The figures show the classification accuracy $[\% ]$ ( $\uparrow$ ) for each round $t$ on CIFAR100 and CUB200. The cost of full supervision $C_{\mathrm{f}}$ is 1, and that of weak supervision is $\frac{1}{2}$ . + +![](images/71c1a593d67beb675f60daa048442032da4b52e0087a56d896f50c6cb59cd633.jpg) +(a) CIFAR100 +Figure 3. Classification accuracy $[\% ]$ ( $\uparrow$ ) of the proposed method on CIFAR100 and CUB200 when the cost of weak supervision, $C_{\mathrm{w}}$ is $\frac{1}{2}$ , $\frac{1}{4}$ , and $\frac{1}{8}$ . Note that the cost $C_{\mathrm{w}}$ is a value set by the user before active learning begins and is not a hyperparameter of the proposed method. This figure illustrates the effect of different weak supervision costs. For reference, we also include the results of conventional active learning methods (gray lines). + +![](images/2641baf29215fa0eb1b4e08ce2ed818bab1e110aec97f1962978198a7fb09631.jpg) +(b) CUB200 + +bels), is lower than that of an expert annotator. The cost $C_{\mathrm{w}}$ varies for each task or dataset, corresponding to the rewards offered by crowdsourcing, etc. + +In previous sections, we set the cost of weak supervision to $C_{\mathrm{w}} = \frac{1}{2}$ . Here, we investigate the effects of our approach when further reducing $C_{\mathrm{w}}$ to $\frac{1}{4}$ and $\frac{1}{8}$ . As shown in Figure 3, for both datasets, the classification accuracy improves as the cost of weak supervision decreases. This improvement is due to the fact that the lower cost of weak supervision allows more data to be annotated within the same budget, thereby enhancing the training process. In other words, the cheaper the cost of weak supervision, the more impactful weak supervision and our approach become. + +# 6.3. Comparison with combined active learning and multiple supervision baseline + +In addition to comparing our approach with conventional active learning approaches, we conduct experiments to evaluate our approach against a straightforward combination of weak supervision and conventional active learning techniques. This baseline approach applies a fixed proportion of weak supervision within the budget, utilizing conventional active learning strategies within each supervision type. + +Specifically, in this combined method, weak supervision is allocated at a constant rate— $20\%$ , $40\%$ , $60\%$ , and $80\%$ in the batch. Within each full and weak supervision, conventional AL methods are employed to select data. In this experiment, the active learning method for each full or weak supervision employed was sampling based on margin, which is used for the uncertainty of the proposed method. + +Figure 4 shows the comparison between our approach and baseline methods that combine active learning with weak supervision on the CIFAR100 dataset under different costs for weak supervision ( $C_{\mathrm{w}} = \frac{1}{2}, \frac{1}{4}$ , and $\frac{1}{8}$ ). In each case, the baseline methods use fixed proportions of full and weak supervision within each batch, denoted as $\mathrm{F}(x\%) + \mathrm{W}(y\%)$ , where $x$ and $y$ represent the percentages of full and weak supervision, respectively. + +Across all weak supervision costs, our approach consistently achieves high accuracy. An important point to note is that the optimal proportion of full and weak supervision is unknown beforehand and varies depending on the cost of weak supervision. For example, when the cost of weak supervision $C_{\mathrm{w}}$ is $\frac{1}{2}$ , the configuration with 60% full supervision and 40% weak supervision achieves the best accuracy among the baseline methods. However, when the cost $C_{\mathrm{w}}$ is reduced to $\frac{1}{8}$ , a configuration with 20% full supervision and 80% weak supervision performs best. In contrast, our approach dynamically constructs optimal batches by optimizing both the supervision level and instance selection based on the cost-effectiveness obtained from the trained model at each round. + +# 6.4. Ablation study + +We conducted ablation studies to demonstrate the effectiveness of instance-wise value (i.e., uncertainty) and the batch selection with diversity on CIFAR100 and CUB200, as shown in Fig. 5. + +The first ablation variant, referred to as "Ours w/o uncertainty," removes the instance-wise uncertainty component from the VCR calculation. In this variant, the VCR is determined solely based on the annotation cost and expected model improvement without taking into account the + +![](images/8a1532eafe25d4978b5d52b2fb40a9a0c1995d7914597d91293290f0150e9490.jpg) + +![](images/4908f98fbfb49a5ebfafcb60bb112c38f91b198a3328660338f322d9636bbced.jpg) +(a) $C_{\mathrm{w}} = 1 / 2$ + +![](images/4fd101f3ca5a897278ba8136280a8a590d2d9ef2ff65d07bc22a9adc8092b519.jpg) +(b) $C_{\mathrm{w}} = 1 / 4$ + +![](images/369cef2477ab6be0f60a42a1302d9b874452722766e5c6cbbaf71af6fa9bb421.jpg) +(c) $C_{\mathrm{w}} = 1 / 8$ + +![](images/8cc7e75bf52f191c52751370dd796d214f3108d02126ef517aafc2081360c780.jpg) +Figure 4. Comparison with combined active learning and multiple supervision baseline. The classification accuracy $[\%] (\uparrow)$ on CIFAR100 with fixed proportions of full and weak supervision within each batch. Each batch was sampled with fixed proportions, and conventional active learning methods were applied separately for full and weak supervision. +(a) CIFAR100 +(b) CUB200 +Figure 5. Ablation study comparing the proposed method without instance-wise uncertainty $u_{\mathrm{f}}(\pmb{x})$ , $u_{\mathrm{w}}(\pmb{x})$ (Ours w/o uncertainty) and without batch selection considering diversity (Ours w/o diversity). The figures show the classification accuracy $[\%] (\uparrow)$ for each round $t$ on CIFAR100 and CUB200. + +uncertainty of each instance. Consequently, the supervision level is not optimized for each instance individually. The results show that our approach outperforms this variant on both datasets. This effect is particularly prominent in the CUB200 dataset, demonstrating the benefit of optimizing "instance-wise" supervision-level. + +The second ablation variant, called 'Ours w/o diversity," involves a greedy approach to batch selection, where instances are selected in descending order of VCR without considering data diversity. Our approach, which incorporates diversity into the batch selection process, consistently outperforms this variant on both datasets. The impact of diversity-aware selection is especially evident in the CIFAR100 dataset, demonstrating that considering diversity in batch selection significantly enhances model performance. + +# 7. Conclusion, limitation, and future work + +In this paper, we proposed a novel active learning (AL) approach, called Instance-wise Supervision-level Optimization (ISO), which optimizes the supervision level (either full or weak annotation) for each instance to maximize annotation efficiency within a budget. By evaluating the uncertainty, diversity, and value-to-cost ratio (VCR) of each instance for full and weak supervision, our approach automatically determines the instances to be annotated and their optimal annotation level. Experiments on classification tasks demonstrated that ISO outperforms not only standard AL methods but also state-of-the-art AL methods with weak supervision, while effectively leveraging instance-specific information to improve both accuracy and cost-efficiency. + +A current limitation of this study is that the effectiveness of our ISO has been validated only in classification experiments. The concept of ISO is applicable to various tasks beyond classification. For example, ISO can be applied to object segmentation tasks, where precise yet costly object boundaries are assumed for full supervision, while simpler rectangular bounding boxes are used for weak supervision. Another current limitation is the assumption that supervision levels are only two. We can expand to deal with more supervision levels by introducing class levels other than exact classes and superclasses. In future work, we aim to demonstrate the full versatility of our approach by exploring these extensions across a broader range of tasks, reinforcing its potential as a scalable, cost-effective solution for real-world applications. + +Acknowledgements: This work was supported by JSPS KAKENHI Grant Number JP23KJ1723, and JST, ACT-X Grant Number JPMJAX23CR, Japan. + +# References + +[1] Charu C Aggarwal, Xiangnan Kong, Quanquan Gu, Jiawei Han, and S Yu Philip. Active learning: A survey. In Data classification, pages 599-634. Chapman and Hall/CRC, 2014. 1, 2 +[2] D ARTHUR. k-means++: the advantages of careful seeding. In ACM-SIAM symposium on Discrete algorithms, pages 1027-1035, 2007. 5 +[3] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. In International Conference on Learning Representations. 1, 2, 3, 4, 5 +[4] Shayok Chakraborty. Asking the right questions to the right users: Active learning with imperfect oracles. In Association for the Advancement of Artificial Intelligence, pages 3365-3372, 2020. 2 +[5] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International Conference on Machine Learning, pages 1183-1192, 2017. 1, 2 +[6] Guy Hacohen, Avihu Dekel, and Daphna Weinshall. Active learning on a budget: Opposite strategies suit high and low budgets. In International Conference on Machine Learning, pages 8175-8195, 2022. 1, 2, 3 +[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770-778, 2016. 6 +[8] Neil Houlsby, Ferenc Huszár, Zoubin Ghahramani, and Máté Lengyel. Bayesian active learning for classification and preference learning. arXiv preprint arXiv:1112.5745, 2011. 1, 2, 3 +[9] Wei-Ning Hsu and Hsuan-Tien Lin. Active learning by learning. In Association for the Advancement of Artificial Intelligence, 2015. 1, 2, 5 +[10] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[11] Andreas Kirsch, Joost Van Amersfoort, and Yarin Gal. Batchbald: Efficient and diverse batch acquisition for deep bayesian active learning. Neural Information Processing Systems, 32, 2019. 1, 2, 3 +[12] Dani Kiyasseh, Tingting Zhu, and David A Clifton. Soqal: Selective oracle questioning for consistency based active learning of cardiac signals. In International Conference on Machine Learning, pages 11302-11340, 2022. 2 +[13] Alex Kulesza and Ben Taskar. k-dpps: Fixed-size determinantal point processes. In International Conference on Machine Learning, pages 1193-1200, 2011. 5 +[14] Chao Lu and Yuexian Zou. Using coarse label constraint for fine-grained visual classification. In Conference on Multimedia Modeling, 2018. 5 +[15] Andrew Kachites McCallum, Kamal Nigam, et al. Employing em and pool-based active learning for text classification. In International Conference on Machine Learning, pages 350-358, 1998. 2 +[16] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. + +A survey of deep active learning. ACM computing surveys (CSUR), 54(9):1-40, 2021. 1, 2 +[17] Zeyu Ren, Shuihua Wang, and Yudong Zhang. Weakly supervised machine learning. CAAI Transactions on Intelligence Technology, 8(3):549-580, 2023. 1, 2 +[18] Dan Roth and Kevin Small. Margin-based active learning for structured output spaces. In European Conference on Machine Learning, pages 413-424, 2006. 1, 2, 3, 5, 6 +[19] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 2, 3, 4, 5 +[20] Burr Settles. Active learning literature survey. 2009. 2 +[21] Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell. Variational adversarial active learning. In International Conference on Computer Vision, pages 5972-5981, 2019. 1, 2, 3 +[22] Javier Gamazo Tejero, Martin S Zinkernagel, Sebastian Wolf, Raphael Sznitman, and Pablo Márquez-Neila. Full or weak annotations? an adaptive strategy for budget-constrained annotation campaigns. In Conference on Computer Vision and Pattern Recognition, pages 11381-11391, 2023. 1, 2, 3, 5, 7 +[23] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Technical report, 2011. 5 +[24] Dan Wang and Yi Shang. A new active labeling method for deep learning. In International Joint Conference on Neural Networks, pages 112-119, 2014. 1, 2, 3, 5 +[25] Donggeun Yoo and In So Kweon. Learning loss for active learning. In Computer Vision and Pattern Recognition, pages 93-102, 2019. 1, 2 +[26] Man Zhang, Yong Zhou, Jiaqi Zhao, Yiyun Man, Bing Liu, and Rui Yao. A survey of semi-and weakly supervised semantic segmentation of images. Artificial Intelligence Review, 53:4259-4288, 2020. 1, 2 +[27] Zhi-Hua Zhou. A brief introduction to weakly supervised learning. National science review, 5(1):44-53, 2018. 1, 2 \ No newline at end of file diff --git a/2025/Instance-wise Supervision-level Optimization in Active Learning/images.zip b/2025/Instance-wise Supervision-level Optimization in Active Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..40205a968d01a4d62760aeff3197dc6ff87e9344 --- /dev/null +++ b/2025/Instance-wise Supervision-level Optimization in Active Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24162d55b2040fcc315adc279e63b6c310ef868602edf1dd28821a83288671e9 +size 241697 diff --git a/2025/Instance-wise Supervision-level Optimization in Active Learning/layout.json b/2025/Instance-wise Supervision-level Optimization in Active Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..87b9aba0d535b06c1fed108fe975a0a1c8d7aa02 --- /dev/null +++ b/2025/Instance-wise Supervision-level Optimization in Active Learning/layout.json @@ -0,0 +1,9146 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 504, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 504, + 121 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 504, + 121 + ], + "type": "text", + "content": "Instance-wise Supervision-level Optimization in Active Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "spans": [ + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "text", + "content": "Shinnosuke Matsuo" + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "text", + "content": ", Riku Togashi" + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "text", + "content": ", Ryoma Bise" + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "text", + "content": ", Seiichi Uchida" + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "text", + "content": ", Masahiro Nomura" + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "text", + "content": "Kyushu University, Japan " + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 142, + 517, + 173 + ], + "type": "text", + "content": "CyberAgent, Inc., Japan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "spans": [ + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 224, + 297, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 297, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 297, + 439 + ], + "type": "text", + "content": "Active learning (AL) is a label-efficient machine learning paradigm that focuses on selectively annotating high-value instances to maximize learning efficiency. Its effectiveness can be further enhanced by incorporating weak supervision, which uses rough yet cost-effective annotations instead of exact (i.e., full) but expensive annotations. We introduce a novel AL framework, Instance-wise Supervision-Level Optimization (ISO), which not only selects the instances to annotate but also determines their optimal annotation level within a fixed annotation budget. Its optimization criterion leverages the value-to-cost ratio (VCR) of each instance while ensuring diversity among the selected instances. In classification experiments, ISO consistently outperforms traditional AL methods and surpasses a state-of-the-art AL approach that combines full and weak supervision, achieving higher accuracy at a lower overall cost. This code is available at https://github.com/matsuoshinnosuke/ISOAL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 464, + 135, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 135, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 135, + 477 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "text", + "content": "Machine learning technology has achieved remarkable success in recent years across various fields, including computer vision, natural language processing, and robotics. However, a major challenge remains: creating labeled datasets for training models often requires significant time and cost, especially in specialized domains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 700 + ], + "type": "text", + "content": "To address this issue, active learning [1, 3, 6, 8, 11, 16, 18-21, 24] has been widely studied to improve model performance under limited budgets. In active learning, instances (data samples) for annotation are selected iteratively from an unlabeled dataset, followed by annotation and model retraining. This repeated cycle improves the model's accuracy while progressively minimizing annotation costs. A common approach in active learning is uncertainty-based sampling [3, 5, 8, 9, 18, 22, 24, 25], where instances with high uncertainty, such as low confident of the model's prediction, are selected for annotation, as these are expected to contribute most effectively to model performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Another approach to reduce annotation costs is weak su" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 200, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 555, + 344 + ], + "type": "text", + "content": "pervision [17, 26, 27]. For classification tasks, a weak supervision approach is often designed to just attach rough class labels to individual instances. For example, instead of attaching the exact class label \"song sparrow\" or \"house sparrow\" to an image instance, annotators can attach a rough class label \"sparrows.\" This approach drastically reduces annotation costs (budgets) because the fee paid to a non-expert annotator, who can only assign rough class labels rather than exact ones via a crowdsourcing service, is lower than that of an expert annotator. Hereafter, we call rough classes as superclasses; therefore, \"sparrows\" is a superclass of the exact class \"song sparrow.\"" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 349, + 556, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 556, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 556, + 516 + ], + "type": "text", + "content": "In general, weak supervision approaches still need instances with full supervision, that is, instances with the exact class labels. This is simply because it is impossible to determine the class boundary between, \"sparrows\" is a superclass of \"song sparrow\" without the instances with these exact class labels. Consequently, we need to balance weak supervision and full supervision under a pre-specified total annotation budget. In other words, we need to optimize the supervision levels of individual instances. Since the total annotation budget is pre-specified as a constraint, the number of instances with full supervision will be limited. More formally, this supervision-level optimization task can be seen as an optimal resource allocation problem with the constraint on available resources." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "This paper proposes a label-efficient machine learning approach, called Instance-wise Supervision-level Optimization (ISO) for active learning, as shown in Fig. 1. Roughly speaking, the proposed approach is a sophisticated hybrid of weak supervision and active learning. Different from the state-of-the-art hybrid approach [22] (reviewed in Section 2), our approach has a novel mechanism that automatically determines the optimal supervision levels of individual instances. Specifically, as shown in Fig. 1, all unlabeled instances at a certain round are automatically divided into three categories: instances for full supervision (with exact class annotation), instances for weak supervision (with superclass annotation), and instances that remain unlabeled. The instances selected in the first two categories are considered to be selected for active learning; annotators, therefore, attach appropriate labels to these instances according" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4939" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 74, + 519, + 216 + ], + "blocks": [ + { + "bbox": [ + 99, + 74, + 519, + 216 + ], + "lines": [ + { + "bbox": [ + 99, + 74, + 519, + 216 + ], + "spans": [ + { + "bbox": [ + 99, + 74, + 519, + 216 + ], + "type": "image", + "image_path": "ff300c4be666d44a776b9ea022ac3ca2c19016ddf4b63f783033e9cf76876faa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 224, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 54, + 224, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 555, + 291 + ], + "type": "text", + "content": "Figure 1. Overview of Instance-wise Supervision-level Optimization (ISO) for active learning. The proposed approach combines weak and full supervision within an active learning framework to maximize annotation efficiency under a fixed budget. First, each instance " + }, + { + "bbox": [ + 54, + 224, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 54, + 224, + 555, + 291 + ], + "type": "text", + "content": " in the unlabeled data pool is evaluated for its value-to-cost ratio " + }, + { + "bbox": [ + 54, + 224, + 555, + 291 + ], + "type": "inline_equation", + "content": "(v_{\\mathrm{f}}(\\pmb{x}))" + }, + { + "bbox": [ + 54, + 224, + 555, + 291 + ], + "type": "text", + "content": " for full supervision and " + }, + { + "bbox": [ + 54, + 224, + 555, + 291 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{w}}(\\pmb{x})" + }, + { + "bbox": [ + 54, + 224, + 555, + 291 + ], + "type": "text", + "content": " for weak supervision). Based on these values and data diversity, instances are selected within the budget constraints for weak supervision (with superclass labels) or full supervision (with exact class labels). This adaptive allocation of supervision-level enables optimal use of resources within the budget constraints in active learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 312, + 163, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 163, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 163, + 322 + ], + "type": "text", + "content": "to their supervision levels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 323, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 455 + ], + "type": "text", + "content": "In our experiments, the proposed method consistently outperformed existing approaches across multiple datasets, demonstrating a more efficient use of the annotation budget. Our approach achieved the same accuracy as the comparative methods while using only three-fifths of the budget by incorporating weak labels when the cost of weak label annotation is half that of fully supervised annotation. Furthermore, in comparison with the state-of-the-art method that combines active learning and multiple supervision, our approach achieved significantly better cost efficiency by leveraging instance-specific information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 455, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 477 + ], + "type": "text", + "content": "The main contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 479, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 526 + ], + "type": "text", + "content": "- We introduce a novel active learning framework that optimizes the supervision level for each instance, allowing for dynamic allocation of weak and full annotations based on budget constraints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 527, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 527, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 527, + 295, + 574 + ], + "type": "text", + "content": "- We propose an efficient algorithm that assigns the optimal supervision level to each instance by considering value-to-cost ratios and data diversity, enhancing the cost-effectiveness of the annotation process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 574, + 296, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 574, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 574, + 296, + 634 + ], + "type": "text", + "content": "- We validate the effectiveness of our approach on classification tasks, using superclasses as weak supervision, and demonstrate its superiority over conventional AL methods and the state-of-the-art AL method that incorporates both full and weak supervision." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 645, + 140, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 140, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 140, + 658 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Active learning (AL) [1, 16, 20] is a technique aimed at maximizing model performance under a limited annotation budget by strategically selecting informative instances for annotation. A widely adopted approach in active learning is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 311, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 502 + ], + "type": "text", + "content": "uncertainty-based sampling [3, 5, 8, 9, 18, 22, 24, 25]. For example, [18] selects instances with small margins, prioritizing instances for which the model has low confidence. Another common approach focuses on maximizing diversity within selected instances [6, 15, 19, 21], which aims to cover a broad range of the data distribution. For instance, [19] selects a representative core set of instances to maximize diversity. Moreover, there are methods that combine uncertainty and diversity-based strategies to leverage the advantages of both approaches [3, 11]. Additionally, some AL methods [4, 12] address the setting where different annotators have varying accuracy levels. These methods aim to jointly optimize both the selection of data for annotation and the choice of annotators to minimize label noise. However, all of these approaches assume a single supervision-level and do not incorporate multiple supervision-levels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 506, + 556, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 637 + ], + "type": "text", + "content": "Another approach to reduce annotation costs is weakly supervised learning (WSL) [17, 26, 27], where less detailed or approximate labels instead of fully supervised labels are utilized. In many real-world applications, obtaining fully supervised labels can be expensive and time-consuming. To address this, WSL methods use weak supervision, which is typically easier and cheaper to obtain. For example, in classification tasks, weak supervision can be provided by superclasses such as categories, and in segmentation tasks, it can be provided by bounding boxes, allowing faster labeling with lower costs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "A hybrid approach between AL and WSL, where the proportion of full and weak supervision is adaptively determined, has been proposed in [22]. Although this approach aims to optimize budget use by setting an ideal balance of weak and full supervision, it applies random sampling after setting the proportion without instance-level considera" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4940" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 174, + 298, + 256 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "type": "text", + "content": "Table 1. Comparison of conventional active learning methods. \"Uncertainty\" (Unc.) refers to whether a method incorporates uncertainty-based instance selection, such as selecting instances with minimal margins. \"Diversity\" (Div.) indicates if the method considers data diversity to maximize the representativeness of selected instances. \"Multiple supervision\" (Mul.) refers to the capability to utilize multiple levels of supervision, such as both weak and full supervision. \"Instance-wise\" (Ins.) represents the ability to choose supervision for each instance adaptively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 174, + 298, + 256 + ], + "lines": [ + { + "bbox": [ + 55, + 174, + 298, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 298, + 256 + ], + "type": "table", + "html": "
Margin, etc. [8, 18, 24]Coreset, etc. [6, 19, 21]BADGE, etc. [3, 11]APFWA [22]Ours
Unc.✓×✓×✓
Div.×✓✓×✓
Mul.×××✓✓
Ins.✓✓✓×✓
", + "image_path": "04698a57d1ab3c979758e1406b61f2134dc551f551548c29875613d0d1680e03.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 274, + 295, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 298 + ], + "type": "text", + "content": "tions. To our knowledge, no method exists that dynamically assigns full and weak supervision for each instance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 299, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 468 + ], + "type": "text", + "content": "Table 1 summarizes a comparison between popular active learning methods and our instance-wise supervision-level optimization. Each method is evaluated on its ability to incorporate uncertainty, diversity, multiple levels of supervision, and instance-wise adaptability, as defined in the caption. Most uncertainty-based and diversity-based AL methods do not consider multiple levels of supervision, focusing instead on selecting instances based on a single type of annotation. Although APFWA incorporates multiple levels of supervision, it lacks instance-wise optimization. In contrast, our approach is unique in that it optimizes the supervision level on an instance-wise basis, simultaneously considering data uncertainty, diversity, and the annotation cost associated with multiple levels of supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 477, + 152, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 152, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 152, + 491 + ], + "type": "text", + "content": "3. Problem setting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{u}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "text", + "content": " represent the unlabeled data pool, and let " + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "text", + "content": " denote the number of rounds in the active learning process, where " + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "text", + "content": " is the available budget in each round. In this setting, a set of instances is selected to be annotated with either full supervision or weak supervision, depending on the required level of detail and the associated cost. Specifically, we define " + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "text", + "content": " as the annotation cost per instance for full supervision, where precise labels are assigned, and " + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "text", + "content": " as the annotation cost per instance for weak supervision, which involves less precise, lower-cost labeling. These annotation costs, " + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "text", + "content": ", are provided in advance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "The process of batch selection and training is repeated for " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " rounds, allowing the model to progressively improve by iteratively selecting and annotating the most informative instances within each budget-constrained round. Once an instance is annotated, it is moved from the unlabeled pool " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{u}}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " to either the fully supervised data pool " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " or the weakly supervised data pool " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{w}}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ", depending on the supervi" + } + ] + } + ], + "index": 6 + }, + { + "type": "code", + "bbox": [ + 319, + 98, + 555, + 449 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 95 + ], + "type": "text", + "content": "Algorithm 1 ISO: Instance-wise Supervision-level Optimization" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": "1: Inputs: Round " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ", budget of each round " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ", annotation cost " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{f}}, C_{\\mathrm{w}}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ", unlabeled data pool " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{u}}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ", feature extractor " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ", classification head " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{f}}, h_{\\mathrm{w}}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ". \n2: Outputs: trained " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "f^{T}, h_{\\mathrm{f}}^{T}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ". \n3: for round " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "t = 1, \\ldots, T" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " do \n4: if " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " is 1 then \n5: Sample randomly instances from " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{u}}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " within budget " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " to create an initial fully supervised dataset " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{f}}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " and weakly supervised dataset " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{w}}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " with an equal number of instances. \n6: else if " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " then \n7: Compute the model improvements, " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{f}}^{t}, M_{\\mathrm{w}}^{t}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ", for each supervision using " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{f}}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{w}}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " by Eq. (3) and (4). \n8: for " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{u}}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " do \n9: 1. Compute the uncertainties, " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{f}}^{t}(\\pmb{x}), u_{\\mathrm{f}}^{t}(\\pmb{x})" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ", for each supervision. \n10: 2. Compute the VCRs, " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{f}}^{t}(\\pmb{x}), v_{\\mathrm{w}}^{t}(\\pmb{x})" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ", for each supervision by Eq. (1) and (2). \n11: 3. Compute the normalized features " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\tilde{f}^{t}(\\pmb{x})" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ". \n12: end for \n13: Select the batch " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{f}}^{+}, \\mathcal{D}_{\\mathrm{w}}^{+}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " using the VCRs, " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{f}}^{t}(\\pmb{x}), v_{\\mathrm{w}}^{t}(\\pmb{x})" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ", and the normalized features " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\tilde{f}^{t}(\\pmb{x})" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " by Sec. 4.3. \n14: end if \n15: " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{f}} \\gets \\mathcal{D}_{\\mathrm{f}} \\cup \\mathcal{D}_{\\mathrm{f}}^{+}, \\mathcal{D}_{\\mathrm{u}} \\gets \\mathcal{D}_{\\mathrm{u}} \\setminus \\mathcal{D}_{\\mathrm{f}}^{+}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " \n16: " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{w}} \\gets \\mathcal{D}_{\\mathrm{w}} \\cup \\mathcal{D}_{\\mathrm{w}}^{+}, \\mathcal{D}_{\\mathrm{u}} \\gets \\mathcal{D}_{\\mathrm{u}} \\setminus \\mathcal{D}_{\\mathrm{w}}^{+}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " \n17: Train " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "f^{t}, h_{\\mathrm{f}}^{t}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{w}}^{t}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{f}}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{w}}" + }, + { + "bbox": [ + 319, + 98, + 555, + 449 + ], + "type": "text", + "content": ". \n18: end for" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_body" + } + ], + "index": 8, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 470, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 492 + ], + "type": "text", + "content": "sion level chosen. These annotated instances are then used to train a classifier." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 494, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 555, + 589 + ], + "type": "text", + "content": "The ultimate goal in this setting is to obtain the classifier's parameters to classify the exact class of input data. To train the classification network, both fully supervised data and weakly supervised data contribute to training the feature extractor in the network. Therefore, balancing fully and weakly supervised data in a cost-effective manner is crucial for optimizing the learning process within budget constraints." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 601, + 561, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 561, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 561, + 629 + ], + "type": "text", + "content": "4. Proposed method: Instance-wise Supervision-level Optimization (ISO)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 635, + 382, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 382, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 382, + 647 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": "We propose an efficient algorithm to address the instance-wise supervision-level optimization problem within a fixed budget. Our approach comprises two primary components, as shown in Fig. 1. The first component calculates the value-to-cost ratio (VCR) for each instance-supervision" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4941" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "level pair, which indicates the expected improvement in model accuracy per unit of cost achieved through annotation for each instance. The VCR for each instance is determined by both the value of the supervision level and its uncertainty. This allows us to evaluate the cost-effectiveness of assigning different supervision levels to each instance individually. The second component performs batch selection within the budget " + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": ", optimizing the set of instances based on their VCRs and diversity. This two-part approach enables our approach to maximize the value of annotations within a limited budget by balancing cost-efficiency and diversity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 223, + 250, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 250, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 250, + 235 + ], + "type": "text", + "content": "4.2. Value-to-cost ratio (VCR) estimation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 241, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 296, + 336 + ], + "type": "text", + "content": "In this section, we explain the calculation of the value-to-cost ratio (VCR) for each instance-supervision pair. An overview of the algorithm is provided in Algorithm 1. For an instance " + }, + { + "bbox": [ + 55, + 241, + 296, + 336 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 241, + 296, + 336 + ], + "type": "text", + "content": " in the unlabeled pool " + }, + { + "bbox": [ + 55, + 241, + 296, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{u}}" + }, + { + "bbox": [ + 55, + 241, + 296, + 336 + ], + "type": "text", + "content": ", the VCRs " + }, + { + "bbox": [ + 55, + 241, + 296, + 336 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{f}}^{t}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 241, + 296, + 336 + ], + "type": "text", + "content": " for full supervision and " + }, + { + "bbox": [ + 55, + 241, + 296, + 336 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{w}}^{t}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 241, + 296, + 336 + ], + "type": "text", + "content": " for weak supervision at round " + }, + { + "bbox": [ + 55, + 241, + 296, + 336 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 241, + 296, + 336 + ], + "type": "text", + "content": " are determined by the uncertainty of the instance, the value, and cost for the two different-level labels, which are defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 342, + 294, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 294, + 368 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 294, + 368 + ], + "type": "interline_equation", + "content": "v _ {\\mathrm {f}} ^ {t} (\\boldsymbol {x}) = \\frac {M _ {\\mathrm {f}} ^ {t} u _ {\\mathrm {f}} ^ {t} (\\boldsymbol {x})}{C _ {\\mathrm {f}}} \\tag {1}", + "image_path": "17845fd946046d8c84959d5a01646da76d96cb9df4df72e5c8f748f5c3a6c229.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 369, + 294, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 294, + 396 + ], + "type": "interline_equation", + "content": "v _ {\\mathrm {w}} ^ {t} (\\boldsymbol {x}) = \\frac {M _ {\\mathrm {w}} ^ {t} u _ {\\mathrm {w}} ^ {t} (\\boldsymbol {x})}{C _ {\\mathrm {w}}} \\tag {2}", + "image_path": "8820042bda5ed8656960c9c6fe4676086eadf70e5f4311546d5364b0400306ad.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{f}}, C_{\\mathrm{w}}" + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "text", + "content": " represent the annotation costs, and " + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{f}}^{t}, M_{\\mathrm{w}}^{t}" + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "text", + "content": " are the expected model improvement, where " + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "inline_equation", + "content": "*_{\\mathrm{f}}, *_{\\mathrm{w}}" + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "text", + "content": " correspond to full supervision and weak supervision, respectively. " + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{f}}^{t}(\\boldsymbol{x}), u_{\\mathrm{w}}^{t}(\\boldsymbol{x})" + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "text", + "content": " represents the uncertainty for each instance " + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "text", + "content": " in the unlabeled pool " + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{u}}" + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "text", + "content": " at round " + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 402, + 295, + 510 + ], + "type": "text", + "content": ", calculated based on model predictions. This formulation allows us to evaluate the cost-effectiveness of assigning either full or weak supervision to each instance, providing a basis for the batch selection strategy within a fixed budget." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "The expected model improvements " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{f}}^{t}" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " for full supervision and " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{w}}^{t}" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " for weak supervision (i.e., the values for each supervision) are estimated by evaluating the improvement in model performance as the training data is incrementally increased. Consider the " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "-th annotation round, where fully supervised data " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{f}}^{t}" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " and weakly supervised data " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{w}}^{t}" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " are available at that time. To evaluate accuracy improvement, we divide the labeled data into " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " subsets and incrementally add training data. For fully supervised data, the weakly supervised data is kept fixed while subsets of the fully supervised data are added. The model performance " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{f},k}" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " represents the evaluation using the first " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " subsets of fully supervised training data, with " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{f},1}" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " being the performance when using just one subset. Similarly, we compute " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{w},k}" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " for the weakly supervised data, where the fully supervised data is fixed, and subsets of the weakly supervised data are incrementally added. Here, model performance is evaluated us" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "ing a small validation set randomly sampled within the budget of a single round. Then, a weighted average is applied for the improvement at each step, " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "m_{k+1}^t - m_k^t" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " to smooth the sequence of improvements. The reason for the weighted approach is to focus on the latest model behavior. To obtain the improvement per data point, each improvement value is normalized by dividing by " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}| / K" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}|" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " represents the number of instances. That is, " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{f}}^t" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{w}}^t" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " are expressed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 354, + 188, + 553, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 188, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 354, + 188, + 553, + 220 + ], + "type": "interline_equation", + "content": "M _ {\\mathrm {f}} ^ {t} = \\frac {\\sum_ {k = 1} ^ {K - 1} k \\cdot \\left(m _ {\\mathrm {f} , k + 1} ^ {t} - m _ {\\mathrm {f} , k} ^ {t}\\right)}{\\sum_ {k = 1} ^ {K - 1} k} \\cdot \\frac {K}{\\left| \\mathcal {D} _ {\\mathrm {f}} \\right|} \\tag {3}", + "image_path": "bc70517f69863056e332616ca6ec8bf254b8eda88bad0714a65febcf80584c05.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 344, + 222, + 553, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 222, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 344, + 222, + 553, + 255 + ], + "type": "interline_equation", + "content": "M _ {\\mathrm {w}} ^ {t} = \\frac {\\sum_ {k = 1} ^ {K - 1} k \\cdot \\left(m _ {\\mathrm {w} , k + 1} ^ {t} - m _ {\\mathrm {w} , k} ^ {t}\\right)}{\\sum_ {k = 1} ^ {K - 1} k} \\cdot \\frac {K}{\\left| \\mathcal {D} _ {\\mathrm {w}} \\right|} \\tag {4}", + "image_path": "6c25e23fa66e540ede83e1ef6e9edfb3f837f670ce8a9d9df517a41b3911a323.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 265, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 503 + ], + "type": "text", + "content": "The uncertainty " + }, + { + "bbox": [ + 313, + 265, + 555, + 503 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{f}}^{t}(\\pmb{x})" + }, + { + "bbox": [ + 313, + 265, + 555, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 265, + 555, + 503 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{w}}^{t}(\\pmb{x})" + }, + { + "bbox": [ + 313, + 265, + 555, + 503 + ], + "type": "text", + "content": ", corresponding to full and weak supervision, can be flexibly defined based on the specific task and dataset. For example, in the classification task, the entropy of the class prediction probabilities is adopted. If different functions are used to calculate uncertainty for full and weak supervision, there is a risk that the scales of " + }, + { + "bbox": [ + 313, + 265, + 555, + 503 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{f}}^{t}(\\pmb{x})" + }, + { + "bbox": [ + 313, + 265, + 555, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 265, + 555, + 503 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{w}}^{t}(\\pmb{x})" + }, + { + "bbox": [ + 313, + 265, + 555, + 503 + ], + "type": "text", + "content": " may differ significantly. To address this, we normalize each uncertainty value by converting it to a percentile score. In this transformation, the instance with the highest uncertainty is assigned a score of 1, the instance with the lowest uncertainty receives a score of 0, and the instance ranked at the 10th percentile in terms of uncertainty is assigned a score of 0.9. We use percentiles rather than dividing by the maximum uncertainty value because extreme outliers could result in an overly large maximum, causing most values to be close to zero. This percentile-based normalization ensures that the distribution of uncertainty values remains well-scaled, preserving relative differences across instances without distortion due to outliers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 514, + 553, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 553, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 553, + 527 + ], + "type": "text", + "content": "4.3. Valuable and diverse batch selection using VCR" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 640 + ], + "type": "text", + "content": "In this section, we describe the batch selection using the value-to-cost ratio (VCR) defined in the previous section. The most straightforward approach for batch selection is to greedily select instances with the highest VCR values. However, as noted in previous studies (e.g. [3, 19]), this approach can lead to selecting similar instances that have high VCR values, resulting in a loss of data diversity within the batch." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "To address this, we propose a batch selection algorithm to select a diverse and valuable batch. In other words, we select a group of instances with high VCR values and diverse features. Specifically, we represent each instance as a vector using its VCR and feature vector. The vector differs depending on the type of label: " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{f}}(\\boldsymbol{x}) \\tilde{f}(\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " for full labels" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4942" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 294, + 119 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 55, + 71, + 294, + 119 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{w}}(\\pmb{x}) \\tilde{f}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 71, + 294, + 119 + ], + "type": "text", + "content": " for weak labels. Here, " + }, + { + "bbox": [ + 55, + 71, + 294, + 119 + ], + "type": "inline_equation", + "content": "\\tilde{f}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 71, + 294, + 119 + ], + "type": "text", + "content": " is a normalized feature vector (with a magnitude of 1) output by the feature extractor " + }, + { + "bbox": [ + 55, + 71, + 294, + 119 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 71, + 294, + 119 + ], + "type": "text", + "content": ". The optimal set of instances is selected based on the vectors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 121, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 121, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 296, + 347 + ], + "type": "text", + "content": "We aim to select a set of instances that maximizes the area enclosed by the vertices of the selected vectors, as shown in the center of Figure 1. Here, the optimization parameter is which vectors (instances) to select and their label types. When the selected instances have high diversity and value, the area enclosed by the selected vectors becomes larger. We solve this optimization problem subject to budget constraints (i.e., the sum of the costs of the optimized supervision levels for instances is less than " + }, + { + "bbox": [ + 56, + 121, + 296, + 347 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 56, + 121, + 296, + 347 + ], + "type": "text", + "content": "). The optimization problem of selecting vectors and determining their supervision levels can be reformulated as a selection problem, where the goal is to choose the optimal set of vectors from all possible vectors, including both those with weak labels " + }, + { + "bbox": [ + 56, + 121, + 296, + 347 + ], + "type": "inline_equation", + "content": "\\{v_{\\mathrm{w}}(\\pmb{x})\\tilde{f} (\\pmb{x})|\\pmb{x}\\in \\mathcal{D}_{\\mathrm{u}}\\}" + }, + { + "bbox": [ + 56, + 121, + 296, + 347 + ], + "type": "text", + "content": " and those with full labels " + }, + { + "bbox": [ + 56, + 121, + 296, + 347 + ], + "type": "inline_equation", + "content": "\\{v_{\\mathrm{f}}(\\pmb {x})\\tilde{f} (\\pmb {x})|\\pmb {x}\\in \\mathcal{D}_{\\mathrm{u}}\\}" + }, + { + "bbox": [ + 56, + 121, + 296, + 347 + ], + "type": "text", + "content": ". This allows the problem of choosing the type of label to be integrated into the vector selection process. This optimization problem can be solved by maximizing the determinant of a set of vectors subject to budget constraints." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "text", + "content": "To perform the batch selection that maximizes the determinant of a matrix, one approach is to sample from a " + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "text", + "content": "-Determinantal Point Process (" + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "text", + "content": "-DPP) [13]. This method selects a batch containing " + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "text", + "content": " vectors with a probability proportional to the determinant of their Gram matrix. However, this differs from the batch selection in our approach in that we choose a set of vectors within a budget " + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "text", + "content": " rather than choosing " + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "text", + "content": " vectors. Therefore, we propose a new selection method, which is inspired by the work [3] using the " + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "text", + "content": "-means++ seeding algorithm [2] as an approximation algorithm for " + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 349, + 296, + 479 + ], + "type": "text", + "content": "-DPP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "text", + "content": "In our batch selection method, the vector of the batch is sampled sequentially with a probability proportional to the square of the distance to the vector closest to the one selected so far until the budget " + }, + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "text", + "content": " is exhausted. Specifically, if the selected vector corresponds to full supervision (i.e., a vector in the set " + }, + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\{v_{\\mathrm{f}}(\\pmb{x})\\tilde{f}(\\pmb{x})|\\pmb{x} \\in \\mathcal{D}_{\\mathrm{u}}\\}" + }, + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "text", + "content": ", the cost " + }, + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "text", + "content": " is deducted from the budget, and it is added to the fully supervised data. If it corresponds to weak supervision (i.e., a vector in the set " + }, + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\{v_{\\mathrm{w}}(\\pmb{x})\\tilde{f}(\\pmb{x})|\\pmb{x} \\in \\mathcal{D}_{\\mathrm{u}}\\}" + }, + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "text", + "content": "), the cost " + }, + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}}" + }, + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "text", + "content": " is deducted from the budget, and it is added to the weakly supervised data. This process is repeated until the budget " + }, + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 481, + 296, + 636 + ], + "type": "text", + "content": " is exhausted. By doing this, we can obtain a diverse and valuable batch." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 650, + 136, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 136, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 136, + 662 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 670, + 114, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 114, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 114, + 681 + ], + "type": "text", + "content": "5.1. Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 712 + ], + "type": "text", + "content": "To demonstrate the effectiveness of the proposed method, we utilized two datasets: CIFAR100 and Caltech-UCSD" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "Birds-200-2011 (CUB200). The task is a multi-class classification problem, where each class (full label) belongs to a superclass, and the superclass label is provided as a weak label." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 120, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 553, + 204 + ], + "type": "text", + "content": "CIFAR-100 is a dataset of natural images containing 100 classes that span a wide range of everyday objects and animals. Each class in CIFAR-100 is associated with a predefined superclass, grouping similar classes together under broader categories, resulting in 20 superclasses in total. The dataset consists of 60,000 images in total, with 50,000 images for training and 10,000 images for testing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 204, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 277 + ], + "type": "text", + "content": "CUB200 [23] consists of 200 classes of bird species, with each class representing a specific bird. Following previous study [14], we defined superclasses based on the suffixes in the class names, resulting in 70 unique superclasses. The dataset contains 11,788 images, with 5,994 images designated for training and 5,794 images for testing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 284, + 437, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 437, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 437, + 296 + ], + "type": "text", + "content": "5.2. Comparison methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 303, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 553, + 350 + ], + "type": "text", + "content": "We compared our approach with eight conventional AL methods, ranging from standard single-supervision AL methods to the state-of-the-art AL method incorporating multiple supervision, as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 351, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 351, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 351, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 314, + 351, + 553, + 374 + ], + "type": "text", + "content": "- Random: The simplest baseline that samples instances randomly from the unlabeled pool." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 377, + 553, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 377, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 315, + 377, + 553, + 425 + ], + "type": "text", + "content": "- Margin [18]: A method that selects instances in ascending order of the difference between the highest and second-highest predicted class probabilities, prioritizing instances with minimal margins." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 427, + 553, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 427, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 315, + 427, + 553, + 462 + ], + "type": "text", + "content": "- MaxConf [24]: An approach that selects instances in order of the lowest maximum predicted class probability, focusing on instances with the least confidence." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 465, + 553, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 465, + 553, + 501 + ], + "spans": [ + { + "bbox": [ + 315, + 465, + 553, + 501 + ], + "type": "text", + "content": "- Entropy [24]: A method that selects instances according to the highest entropy in the predicted class probabilities, targeting instances with the greatest uncertainty." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 503, + 553, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 503, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 315, + 503, + 553, + 539 + ], + "type": "text", + "content": "- Coreset [19]: A diversity-based sampling method that selects instances farthest from the labeled data using a greedy algorithm to enhance diversity." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 541, + 553, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 541, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 315, + 541, + 553, + 588 + ], + "type": "text", + "content": "- ALBL (Active Learning by Learning) [9]: A multiarmed bandit approach that selects between AL methods at each round. In this experiment, we utilized MaxConf and Coreset." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 591, + 553, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 591, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 315, + 591, + 553, + 651 + ], + "type": "text", + "content": "- BADGE (Batch Active learning by Diverse Gradient Embeddings) [3]: A method that considers both predictive uncertainty and diversity by selecting instances with diverse and high-magnitude gradients in the gradient space." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 653, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 653, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 653, + 553, + 712 + ], + "type": "text", + "content": "- APFWA (Adaptive Proportion of Full and Weak Annotations) [22]: A state-of-the-art AL method that incorporates both full and weak supervision, which dynamically determines the proportion of full and weak annotations within each batch." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4943" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "Since AL with both full and weak label annotation is a novel task, there is only one existing method for this task. Instead, we compare our approach with standard single-supervision AL methods as well as the state-of-the-art AL method. This comparison highlights the effectiveness of AL using weak labels. Additionally, a comparison with APFWA demonstrates the advantages of our approach for AL with both fully and weakly supervised labels. Furthermore, in Sec. 6.3, we also compare our approach with a simple combination of weak supervision and conventional AL methods to ensure a fair evaluation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 208, + 295, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 269 + ], + "type": "text", + "content": "These methods offer a range of approaches, from simple random sampling to advanced techniques that balance uncertainty and diversity, as well as methods leveraging weak supervision, enabling a comprehensive evaluation of our proposed approach." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 288, + 187, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 187, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 187, + 300 + ], + "type": "text", + "content": "5.3. Implementation details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 310, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 295, + 453 + ], + "type": "text", + "content": "The network consists of a shared feature extractor " + }, + { + "bbox": [ + 55, + 310, + 295, + 453 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 310, + 295, + 453 + ], + "type": "text", + "content": " and two classification heads, " + }, + { + "bbox": [ + 55, + 310, + 295, + 453 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 310, + 295, + 453 + ], + "type": "text", + "content": " for full supervision and " + }, + { + "bbox": [ + 55, + 310, + 295, + 453 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{w}}" + }, + { + "bbox": [ + 55, + 310, + 295, + 453 + ], + "type": "text", + "content": " for weak supervision. The encoder of ResNet18 [7] is used for the feature extractor " + }, + { + "bbox": [ + 55, + 310, + 295, + 453 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 310, + 295, + 453 + ], + "type": "text", + "content": ", and a single linear layer is used for each classification head. For the network trained using both fully and weakly labeled data, we employ a two-stage training approach in which the feature extractor and each classification head are trained first with weakly supervised data and then with fully supervised data. Since both fully supervised and weakly supervised learning are classification tasks in this experiment, we used the cross-entropy loss. The learning rate was set to 0.001, and the optimizer was Adam [10]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "text", + "content": "The settings for active learning were as follows. The number of rounds " + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "text", + "content": " was set to 5. The budget " + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "text", + "content": " of each round was set to 1,000 for CIFAR100 and to 500 for CUB200 since CUB200 had approximately 5,000 in training data, which was less than that of CIFAR100. The validation data with full supervision were prepared from training data, 1,000 for CIFAR1000 and 500 for CUB200, which are the same number and small amount as in budget " + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "text", + "content": ". The cost of full supervision " + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "text", + "content": " was set to 1, and the cost of weak supervision was set to " + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}" + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "text", + "content": ". In addition, in Sec. 6.2, we conducted the experiments when changing " + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}}" + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 55, + 458, + 295, + 615 + ], + "type": "text", + "content": " to investigate the impact of the difference on the cost of weak supervision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "The number of partitions " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": " for estimating the expected model improvements " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{f}}^{t}" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{w}}^{t}" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": " was set to 5. As the performance metric of " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{f},k}^{t}, m_{\\mathrm{w},k}^{t}" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": ", we used the classification accuracy. To enhance stability, we repeated the estimation process three times with different random seeds and calculated the average. For uncertainty measurement, we adopted the simplest approach, Margin [18], to evaluate instance uncertainty." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 437, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 437, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 437, + 85 + ], + "type": "text", + "content": "6. Experimental results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 91, + 553, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 115 + ], + "type": "text", + "content": "6.1. Comparison with conventional active learning methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 122, + 553, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 553, + 170 + ], + "type": "text", + "content": "The comparison results with conventional AL methods on CIFAR100 and CUB200 are shown in Fig. 2(a), (b), respectively. The graphs show the classification accuracy " + }, + { + "bbox": [ + 313, + 122, + 553, + 170 + ], + "type": "inline_equation", + "content": "[\\%]" + }, + { + "bbox": [ + 313, + 122, + 553, + 170 + ], + "type": "text", + "content": " for each round " + }, + { + "bbox": [ + 313, + 122, + 553, + 170 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 122, + 553, + 170 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 171, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 553, + 361 + ], + "type": "text", + "content": "In Figure 2(a), the results on the CIFAR100 dataset demonstrate that our approach archives higher accuracy across all rounds compared to conventional methods, including random sampling, uncertainty-based AL method (Margin, MaxConf, Entropy), diversity-based AL method (Coreset), and other AL methods (ALBL, BADGE), which use only full supervision. This demonstrates that our approach effectively leverages weak supervision, achieving higher performance within the same budget constraints. For example, at round " + }, + { + "bbox": [ + 313, + 171, + 553, + 361 + ], + "type": "inline_equation", + "content": "t = 5" + }, + { + "bbox": [ + 313, + 171, + 553, + 361 + ], + "type": "text", + "content": ", our approach achieved over " + }, + { + "bbox": [ + 313, + 171, + 553, + 361 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 171, + 553, + 361 + ], + "type": "text", + "content": " higher accuracy compared to conventional AL methods that use only full supervision. Furthermore, while these methods require five rounds (with a budget of 5000) to reach an accuracy of approximately " + }, + { + "bbox": [ + 313, + 171, + 553, + 361 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 171, + 553, + 361 + ], + "type": "text", + "content": ", our approach achieved the same or even higher accuracy in just three rounds, using only three-fifths of the budget." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 362, + 553, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 553, + 457 + ], + "type": "text", + "content": "Our approach also surpasses APFWA, another AL approach that incorporates weak supervision by optimizing the proportions of weak and full supervision within the batch but does not focus on selecting instances. This result suggests that our approach more effectively selects cost-effective instances by dynamically considering instance-wise uncertainty and diversity, thereby maximizing the learning impact of weak supervision." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 459, + 553, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 553, + 542 + ], + "type": "text", + "content": "Similarly, Figure 2(b) presents the results on the CUB200 dataset, where our approach also outperforms other AL methods across all rounds. The consistent improvement observed across both datasets indicates that our approach can dynamically optimize annotation strategies according to the dataset characteristics, achieving high adaptability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 552, + 502, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 502, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 502, + 564 + ], + "type": "text", + "content": "6.2. Analysis for weak supervision costs" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": "To demonstrate that our approach works at any cost of weak supervision and that the cheaper the cost of weak supervision, the greater the effect, we conducted experiments when changing the cost of weak supervision. Figure 3 shows the classification accuracy " + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "inline_equation", + "content": "[\\%]" + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": " on CIFAR100 and CUB200 at different costs of weak supervision. In this experiment, the cost of full supervision " + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{f}}" + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": " is fixed at 1, meaning that " + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}}" + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": " represents the relative cost to full supervision. It is important to note that the cost " + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}}" + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": " is defined before active learning begins and is not a hyperparameter of the proposed method. Imagine a scenario where the fee paid to a non-expert annotator, who can only assign rough class labels (weak la" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4944" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 71, + 265, + 105 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 265, + 105 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 265, + 105 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 265, + 105 + ], + "type": "image", + "image_path": "4dc4e1235dcb03007319bdab6df23db603cd0ab30ba6cb085757df61fff12b32.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 114, + 174, + 223 + ], + "blocks": [ + { + "bbox": [ + 59, + 114, + 174, + 223 + ], + "lines": [ + { + "bbox": [ + 59, + 114, + 174, + 223 + ], + "spans": [ + { + "bbox": [ + 59, + 114, + 174, + 223 + ], + "type": "image", + "image_path": "8f670618b8ff430ba3888a047569a930906347cef3afd0a13a6789333a04e0b5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 229, + 140, + 239 + ], + "lines": [ + { + "bbox": [ + 91, + 229, + 140, + 239 + ], + "spans": [ + { + "bbox": [ + 91, + 229, + 140, + 239 + ], + "type": "text", + "content": "(a) CIFAR100" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 182, + 114, + 286, + 223 + ], + "blocks": [ + { + "bbox": [ + 182, + 114, + 286, + 223 + ], + "lines": [ + { + "bbox": [ + 182, + 114, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 182, + 114, + 286, + 223 + ], + "type": "image", + "image_path": "aa75b502bca009497eff13610fa83b8cd18c3a779816ea4068bef82b553d938f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 229, + 255, + 239 + ], + "lines": [ + { + "bbox": [ + 212, + 229, + 255, + 239 + ], + "spans": [ + { + "bbox": [ + 212, + 229, + 255, + 239 + ], + "type": "text", + "content": "(b) CUB200" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 78, + 302, + 274, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 243, + 296, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 289 + ], + "type": "text", + "content": "Figure 2. Comparison of our approach with the conventional active learning methods. The figures show the classification accuracy " + }, + { + "bbox": [ + 55, + 243, + 296, + 289 + ], + "type": "inline_equation", + "content": "[\\% ]" + }, + { + "bbox": [ + 55, + 243, + 296, + 289 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 243, + 296, + 289 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 243, + 296, + 289 + ], + "type": "text", + "content": ") for each round " + }, + { + "bbox": [ + 55, + 243, + 296, + 289 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 243, + 296, + 289 + ], + "type": "text", + "content": " on CIFAR100 and CUB200. The cost of full supervision " + }, + { + "bbox": [ + 55, + 243, + 296, + 289 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 243, + 296, + 289 + ], + "type": "text", + "content": " is 1, and that of weak supervision is " + }, + { + "bbox": [ + 55, + 243, + 296, + 289 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}" + }, + { + "bbox": [ + 55, + 243, + 296, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 302, + 274, + 327 + ], + "lines": [ + { + "bbox": [ + 78, + 302, + 274, + 327 + ], + "spans": [ + { + "bbox": [ + 78, + 302, + 274, + 327 + ], + "type": "image", + "image_path": "f5802bbf2b719fe0d35273cf6b22a7a2a2b356ba433d9e89d03ac7dbf752fd9d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 59, + 333, + 174, + 443 + ], + "blocks": [ + { + "bbox": [ + 59, + 333, + 174, + 443 + ], + "lines": [ + { + "bbox": [ + 59, + 333, + 174, + 443 + ], + "spans": [ + { + "bbox": [ + 59, + 333, + 174, + 443 + ], + "type": "image", + "image_path": "71c1a593d67beb675f60daa048442032da4b52e0087a56d896f50c6cb59cd633.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 450, + 139, + 459 + ], + "lines": [ + { + "bbox": [ + 92, + 450, + 139, + 459 + ], + "spans": [ + { + "bbox": [ + 92, + 450, + 139, + 459 + ], + "type": "text", + "content": "(a) CIFAR100" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "text", + "content": "Figure 3. Classification accuracy " + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "inline_equation", + "content": "[\\% ]" + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "text", + "content": ") of the proposed method on CIFAR100 and CUB200 when the cost of weak supervision, " + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}}" + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}" + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "text", + "content": ". Note that the cost " + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}}" + }, + { + "bbox": [ + 55, + 464, + 296, + 541 + ], + "type": "text", + "content": " is a value set by the user before active learning begins and is not a hyperparameter of the proposed method. This figure illustrates the effect of different weak supervision costs. For reference, we also include the results of conventional active learning methods (gray lines)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 182, + 334, + 286, + 443 + ], + "blocks": [ + { + "bbox": [ + 182, + 334, + 286, + 443 + ], + "lines": [ + { + "bbox": [ + 182, + 334, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 182, + 334, + 286, + 443 + ], + "type": "image", + "image_path": "2641baf29215fa0eb1b4e08ce2ed818bab1e110aec97f1962978198a7fb09631.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 450, + 254, + 459 + ], + "lines": [ + { + "bbox": [ + 212, + 450, + 254, + 459 + ], + "spans": [ + { + "bbox": [ + 212, + 450, + 254, + 459 + ], + "type": "text", + "content": "(b) CUB200" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 557, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 593 + ], + "type": "text", + "content": "bels), is lower than that of an expert annotator. The cost " + }, + { + "bbox": [ + 55, + 557, + 295, + 593 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}}" + }, + { + "bbox": [ + 55, + 557, + 295, + 593 + ], + "type": "text", + "content": " varies for each task or dataset, corresponding to the rewards offered by crowdsourcing, etc." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "In previous sections, we set the cost of weak supervision to " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}} = \\frac{1}{2}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ". Here, we investigate the effects of our approach when further reducing " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ". As shown in Figure 3, for both datasets, the classification accuracy improves as the cost of weak supervision decreases. This improvement is due to the fact that the lower cost of weak supervision allows more data to be annotated within the same budget, thereby enhancing the training process. In other words, the cheaper the cost of weak supervision, the more impactful weak supervision and our approach become." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "6.3. Comparison with combined active learning and multiple supervision baseline" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 102, + 553, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 102, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 553, + 186 + ], + "type": "text", + "content": "In addition to comparing our approach with conventional active learning approaches, we conduct experiments to evaluate our approach against a straightforward combination of weak supervision and conventional active learning techniques. This baseline approach applies a fixed proportion of weak supervision within the budget, utilizing conventional active learning strategies within each supervision type." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "text", + "content": "Specifically, in this combined method, weak supervision is allocated at a constant rate—" + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "text", + "content": " in the batch. Within each full and weak supervision, conventional AL methods are employed to select data. In this experiment, the active learning method for each full or weak supervision employed was sampling based on margin, which is used for the uncertainty of the proposed method." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 270, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 365 + ], + "type": "text", + "content": "Figure 4 shows the comparison between our approach and baseline methods that combine active learning with weak supervision on the CIFAR100 dataset under different costs for weak supervision (" + }, + { + "bbox": [ + 313, + 270, + 553, + 365 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}} = \\frac{1}{2}, \\frac{1}{4}" + }, + { + "bbox": [ + 313, + 270, + 553, + 365 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 270, + 553, + 365 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 313, + 270, + 553, + 365 + ], + "type": "text", + "content": "). In each case, the baseline methods use fixed proportions of full and weak supervision within each batch, denoted as " + }, + { + "bbox": [ + 313, + 270, + 553, + 365 + ], + "type": "inline_equation", + "content": "\\mathrm{F}(x\\%) + \\mathrm{W}(y\\%)" + }, + { + "bbox": [ + 313, + 270, + 553, + 365 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 270, + 553, + 365 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 270, + 553, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 270, + 553, + 365 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 270, + 553, + 365 + ], + "type": "text", + "content": " represent the percentages of full and weak supervision, respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 366, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 533 + ], + "type": "text", + "content": "Across all weak supervision costs, our approach consistently achieves high accuracy. An important point to note is that the optimal proportion of full and weak supervision is unknown beforehand and varies depending on the cost of weak supervision. For example, when the cost of weak supervision " + }, + { + "bbox": [ + 313, + 366, + 555, + 533 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}}" + }, + { + "bbox": [ + 313, + 366, + 555, + 533 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 366, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}" + }, + { + "bbox": [ + 313, + 366, + 555, + 533 + ], + "type": "text", + "content": ", the configuration with 60% full supervision and 40% weak supervision achieves the best accuracy among the baseline methods. However, when the cost " + }, + { + "bbox": [ + 313, + 366, + 555, + 533 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}}" + }, + { + "bbox": [ + 313, + 366, + 555, + 533 + ], + "type": "text", + "content": " is reduced to " + }, + { + "bbox": [ + 313, + 366, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 313, + 366, + 555, + 533 + ], + "type": "text", + "content": ", a configuration with 20% full supervision and 80% weak supervision performs best. In contrast, our approach dynamically constructs optimal batches by optimizing both the supervision level and instance selection based on the cost-effectiveness obtained from the trained model at each round." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 540, + 406, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 406, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 406, + 553 + ], + "type": "text", + "content": "6.4. Ablation study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 559, + 553, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 553, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 553, + 606 + ], + "type": "text", + "content": "We conducted ablation studies to demonstrate the effectiveness of instance-wise value (i.e., uncertainty) and the batch selection with diversity on CIFAR100 and CUB200, as shown in Fig. 5." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 607, + 553, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 553, + 667 + ], + "type": "text", + "content": "The first ablation variant, referred to as \"Ours w/o uncertainty,\" removes the instance-wise uncertainty component from the VCR calculation. In this variant, the VCR is determined solely based on the annotation cost and expected model improvement without taking into account the" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 673, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 553, + 713 + ], + "type": "text", + "content": "1APFWA [22] is a method that dynamically determines the proportion of full and weak annotations. It must use random instance sampling to estimate the optimal proportion, and as a result, it cannot be combined with the conventional instance-wise AL methods (e.g., margin-based sampling)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4945" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 165, + 70, + 445, + 94 + ], + "blocks": [ + { + "bbox": [ + 165, + 70, + 445, + 94 + ], + "lines": [ + { + "bbox": [ + 165, + 70, + 445, + 94 + ], + "spans": [ + { + "bbox": [ + 165, + 70, + 445, + 94 + ], + "type": "image", + "image_path": "8a1532eafe25d4978b5d52b2fb40a9a0c1995d7914597d91293290f0150e9490.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 73, + 99, + 208, + 225 + ], + "blocks": [ + { + "bbox": [ + 73, + 99, + 208, + 225 + ], + "lines": [ + { + "bbox": [ + 73, + 99, + 208, + 225 + ], + "spans": [ + { + "bbox": [ + 73, + 99, + 208, + 225 + ], + "type": "image", + "image_path": "4908f98fbfb49a5ebfafcb60bb112c38f91b198a3328660338f322d9636bbced.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 231, + 164, + 242 + ], + "lines": [ + { + "bbox": [ + 115, + 231, + 164, + 242 + ], + "spans": [ + { + "bbox": [ + 115, + 231, + 164, + 242 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 115, + 231, + 164, + 242 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}} = 1 / 2" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 238, + 99, + 373, + 225 + ], + "blocks": [ + { + "bbox": [ + 238, + 99, + 373, + 225 + ], + "lines": [ + { + "bbox": [ + 238, + 99, + 373, + 225 + ], + "spans": [ + { + "bbox": [ + 238, + 99, + 373, + 225 + ], + "type": "image", + "image_path": "4fd101f3ca5a897278ba8136280a8a590d2d9ef2ff65d07bc22a9adc8092b519.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 231, + 330, + 242 + ], + "lines": [ + { + "bbox": [ + 280, + 231, + 330, + 242 + ], + "spans": [ + { + "bbox": [ + 280, + 231, + 330, + 242 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 280, + 231, + 330, + 242 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}} = 1 / 4" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 405, + 99, + 538, + 225 + ], + "blocks": [ + { + "bbox": [ + 405, + 99, + 538, + 225 + ], + "lines": [ + { + "bbox": [ + 405, + 99, + 538, + 225 + ], + "spans": [ + { + "bbox": [ + 405, + 99, + 538, + 225 + ], + "type": "image", + "image_path": "369cef2477ab6be0f60a42a1302d9b874452722766e5c6cbbaf71af6fa9bb421.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 231, + 496, + 242 + ], + "lines": [ + { + "bbox": [ + 446, + 231, + 496, + 242 + ], + "spans": [ + { + "bbox": [ + 446, + 231, + 496, + 242 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 446, + 231, + 496, + 242 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{w}} = 1 / 8" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 302, + 289, + 430 + ], + "blocks": [ + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "type": "text", + "content": "Figure 4. Comparison with combined active learning and multiple supervision baseline. The classification accuracy " + }, + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "type": "inline_equation", + "content": "[\\%] (\\uparrow)" + }, + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "type": "text", + "content": " on CIFAR100 with fixed proportions of full and weak supervision within each batch. Each batch was sampled with fixed proportions, and conventional active learning methods were applied separately for full and weak supervision." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 302, + 289, + 430 + ], + "lines": [ + { + "bbox": [ + 59, + 302, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 59, + 302, + 289, + 430 + ], + "type": "image", + "image_path": "8cc7e75bf52f191c52751370dd796d214f3108d02126ef517aafc2081360c780.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 436, + 140, + 445 + ], + "lines": [ + { + "bbox": [ + 92, + 436, + 140, + 445 + ], + "spans": [ + { + "bbox": [ + 92, + 436, + 140, + 445 + ], + "type": "text", + "content": "(a) CIFAR100" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 212, + 436, + 254, + 445 + ], + "lines": [ + { + "bbox": [ + 212, + 436, + 254, + 445 + ], + "spans": [ + { + "bbox": [ + 212, + 436, + 254, + 445 + ], + "type": "text", + "content": "(b) CUB200" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 449, + 295, + 504 + ], + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 504 + ], + "type": "text", + "content": "Figure 5. Ablation study comparing the proposed method without instance-wise uncertainty " + }, + { + "bbox": [ + 55, + 449, + 295, + 504 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{f}}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 449, + 295, + 504 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 449, + 295, + 504 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{w}}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 449, + 295, + 504 + ], + "type": "text", + "content": " (Ours w/o uncertainty) and without batch selection considering diversity (Ours w/o diversity). The figures show the classification accuracy " + }, + { + "bbox": [ + 55, + 449, + 295, + 504 + ], + "type": "inline_equation", + "content": "[\\%] (\\uparrow)" + }, + { + "bbox": [ + 55, + 449, + 295, + 504 + ], + "type": "text", + "content": " for each round " + }, + { + "bbox": [ + 55, + 449, + 295, + 504 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 449, + 295, + 504 + ], + "type": "text", + "content": " on CIFAR100 and CUB200." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 531, + 295, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 603 + ], + "type": "text", + "content": "uncertainty of each instance. Consequently, the supervision level is not optimized for each instance individually. The results show that our approach outperforms this variant on both datasets. This effect is particularly prominent in the CUB200 dataset, demonstrating the benefit of optimizing \"instance-wise\" supervision-level." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "The second ablation variant, called 'Ours w/o diversity,\" involves a greedy approach to batch selection, where instances are selected in descending order of VCR without considering data diversity. Our approach, which incorporates diversity into the batch selection process, consistently outperforms this variant on both datasets. The impact of diversity-aware selection is especially evident in the CIFAR100 dataset, demonstrating that considering diversity in batch selection significantly enhances model performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 304, + 533, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 533, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 533, + 317 + ], + "type": "text", + "content": "7. Conclusion, limitation, and future work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 324, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 479 + ], + "type": "text", + "content": "In this paper, we proposed a novel active learning (AL) approach, called Instance-wise Supervision-level Optimization (ISO), which optimizes the supervision level (either full or weak annotation) for each instance to maximize annotation efficiency within a budget. By evaluating the uncertainty, diversity, and value-to-cost ratio (VCR) of each instance for full and weak supervision, our approach automatically determines the instances to be annotated and their optimal annotation level. Experiments on classification tasks demonstrated that ISO outperforms not only standard AL methods but also state-of-the-art AL methods with weak supervision, while effectively leveraging instance-specific information to improve both accuracy and cost-efficiency." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 480, + 556, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 556, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 556, + 659 + ], + "type": "text", + "content": "A current limitation of this study is that the effectiveness of our ISO has been validated only in classification experiments. The concept of ISO is applicable to various tasks beyond classification. For example, ISO can be applied to object segmentation tasks, where precise yet costly object boundaries are assumed for full supervision, while simpler rectangular bounding boxes are used for weak supervision. Another current limitation is the assumption that supervision levels are only two. We can expand to deal with more supervision levels by introducing class levels other than exact classes and superclasses. In future work, we aim to demonstrate the full versatility of our approach by exploring these extensions across a broader range of tasks, reinforcing its potential as a scalable, cost-effective solution for real-world applications." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 675, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 675, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 675, + 555, + 712 + ], + "type": "text", + "content": "Acknowledgements: This work was supported by JSPS KAKENHI Grant Number JP23KJ1723, and JST, ACT-X Grant Number JPMJAX23CR, Japan." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4946" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Charu C Aggarwal, Xiangnan Kong, Quanquan Gu, Jiawei Han, and S Yu Philip. Active learning: A survey. In Data classification, pages 599-634. Chapman and Hall/CRC, 2014. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "type": "text", + "content": "[2] D ARTHUR. k-means++: the advantages of careful seeding. In ACM-SIAM symposium on Discrete algorithms, pages 1027-1035, 2007. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 296, + 223 + ], + "type": "text", + "content": "[3] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. In International Conference on Learning Representations. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 296, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 296, + 267 + ], + "type": "text", + "content": "[4] Shayok Chakraborty. Asking the right questions to the right users: Active learning with imperfect oracles. In Association for the Advancement of Artificial Intelligence, pages 3365-3372, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 268, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 268, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 268, + 296, + 312 + ], + "type": "text", + "content": "[5] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International Conference on Machine Learning, pages 1183-1192, 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 312, + 296, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 312, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 312, + 296, + 357 + ], + "type": "text", + "content": "[6] Guy Hacohen, Avihu Dekel, and Daphna Weinshall. Active learning on a budget: Opposite strategies suit high and low budgets. In International Conference on Machine Learning, pages 8175-8195, 2022. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 357, + 296, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 296, + 390 + ], + "type": "text", + "content": "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 391, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 296, + 434 + ], + "type": "text", + "content": "[8] Neil Houlsby, Ferenc Huszár, Zoubin Ghahramani, and Máté Lengyel. Bayesian active learning for classification and preference learning. arXiv preprint arXiv:1112.5745, 2011. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 435, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 435, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 435, + 296, + 468 + ], + "type": "text", + "content": "[9] Wei-Ning Hsu and Hsuan-Tien Lin. Active learning by learning. In Association for the Advancement of Artificial Intelligence, 2015. 1, 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 469, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 469, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 469, + 295, + 491 + ], + "type": "text", + "content": "[10] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 491, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 295, + 534 + ], + "type": "text", + "content": "[11] Andreas Kirsch, Joost Van Amersfoort, and Yarin Gal. Batchbald: Efficient and diverse batch acquisition for deep bayesian active learning. Neural Information Processing Systems, 32, 2019. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 536, + 295, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 580 + ], + "type": "text", + "content": "[12] Dani Kiyasseh, Tingting Zhu, and David A Clifton. Soqal: Selective oracle questioning for consistency based active learning of cardiac signals. In International Conference on Machine Learning, pages 11302-11340, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 581, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 613 + ], + "type": "text", + "content": "[13] Alex Kulesza and Ben Taskar. k-dpps: Fixed-size determinantal point processes. In International Conference on Machine Learning, pages 1193-1200, 2011. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 613, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 646 + ], + "type": "text", + "content": "[14] Chao Lu and Yuexian Zou. Using coarse label constraint for fine-grained visual classification. In Conference on Multimedia Modeling, 2018. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[15] Andrew Kachites McCallum, Kamal Nigam, et al. Employing em and pool-based active learning for text classification. In International Conference on Machine Learning, pages 350-358, 1998. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "type": "text", + "content": "[16] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 479 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "A survey of deep active learning. ACM computing surveys (CSUR), 54(9):1-40, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[17] Zeyu Ren, Shuihua Wang, and Yudong Zhang. Weakly supervised machine learning. CAAI Transactions on Intelligence Technology, 8(3):549-580, 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 129, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 163 + ], + "type": "text", + "content": "[18] Dan Roth and Kevin Small. Margin-based active learning for structured output spaces. In European Conference on Machine Learning, pages 413-424, 2006. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 164, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 206 + ], + "type": "text", + "content": "[19] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 2, 3, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 208, + 534, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 534, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 534, + 220 + ], + "type": "text", + "content": "[20] Burr Settles. Active learning literature survey. 2009. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 220, + 553, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 254 + ], + "type": "text", + "content": "[21] Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell. Variational adversarial active learning. In International Conference on Computer Vision, pages 5972-5981, 2019. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 255, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 319 + ], + "type": "text", + "content": "[22] Javier Gamazo Tejero, Martin S Zinkernagel, Sebastian Wolf, Raphael Sznitman, and Pablo Márquez-Neila. Full or weak annotations? an adaptive strategy for budget-constrained annotation campaigns. In Conference on Computer Vision and Pattern Recognition, pages 11381-11391, 2023. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 321, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 342 + ], + "type": "text", + "content": "[23] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Technical report, 2011. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 344, + 553, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 377 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 377 + ], + "type": "text", + "content": "[24] Dan Wang and Yi Shang. A new active labeling method for deep learning. In International Joint Conference on Neural Networks, pages 112-119, 2014. 1, 2, 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 378, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 378, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 378, + 553, + 410 + ], + "type": "text", + "content": "[25] Donggeun Yoo and In So Kweon. Learning loss for active learning. In Computer Vision and Pattern Recognition, pages 93-102, 2019. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 411, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 455 + ], + "type": "text", + "content": "[26] Man Zhang, Yong Zhou, Jiaqi Zhao, Yiyun Man, Bing Liu, and Rui Yao. A survey of semi-and weakly supervised semantic segmentation of images. Artificial Intelligence Review, 53:4259-4288, 2020. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 456, + 553, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 479 + ], + "type": "text", + "content": "[27] Zhi-Hua Zhou. A brief introduction to weakly supervised learning. National science review, 5(1):44-53, 2018. 1, 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4947" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/63b6efae-0d11-487a-b293-b1b579ef5966_content_list.json b/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/63b6efae-0d11-487a-b293-b1b579ef5966_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2fabfb050ee53b0a5614227382efec36db3a3c5d --- /dev/null +++ b/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/63b6efae-0d11-487a-b293-b1b579ef5966_content_list.json @@ -0,0 +1,1490 @@ +[ + { + "type": "text", + "text": "InstanceCap: Improving Text-to-Video Generation via Instance-aware Structured Caption", + "text_level": 1, + "bbox": [ + 143, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tiehan Fan $^{1*}$ Kepan Nan $^{1*}$ Rui Xie $^{1}$ Penghao Zhou $^{2}$ Zhenheng Yang $^{2}$ Chaoyou Fu $^{1}$ Xiang Li $^{3}$ Jian Yang $^{1}$ Ying Tai $^{1\\boxtimes}$ $^{1}$ Nanjing University ${}^{2}$ ByteDance ${}^{3}$ Nankai University \nhttps://github.com/NJU-PCALab/InstanceCap", + "bbox": [ + 200, + 202, + 795, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 323, + 326, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-video generation has evolved rapidly in recent years, delivering remarkable results. Training typically relies on video-caption paired data, which plays a crucial role in enhancing generation performance. However, current video captions often suffer from insufficient details, hallucinations and imprecise motion depiction, affecting the fidelity and consistency of generated videos. In this work, we propose a novel instance-aware structured caption framework, termed InstanceCap, to achieve instance-level and fine-grained video caption for the first time. Based on this scheme, we design an auxiliary models cluster to convert original video into instances to enhance instance fidelity. Video instances are further used to refine dense prompts into structured phrases, achieving concise yet precise descriptions. Furthermore, a 22K InstanceVid dataset is curated for training, and an enhancement pipeline that tailored to InstanceCap structure is proposed for inference. Experimental results demonstrate that our proposed InstanceCap significantly outperform previous models, ensuring high fidelity between captions and videos while reducing hallucinations.", + "bbox": [ + 88, + 354, + 485, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 700, + 220, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, text-to-video (T2V) generation with advanced diffusion transformers (DiT) [2, 10, 12-14, 17, 19, 21, 24, 26, 34, 38] have attracted significant attention for the ability to generate realistic, long-duration videos based on text prompts. Video-caption paired data is typically used in training and plays a crucial role in enhancing generation performance [23, 32]. Current video recaption methods often incorporate multimodal large language models to produce detailed captions, which however usually suffer from hallucinations, leading to inconsistencies between captions", + "bbox": [ + 89, + 726, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and video content. Consequently, creating consistent videocaption pairs with accurate details and precise motion depiction for T2V generation remains a significant challenge.", + "bbox": [ + 511, + 310, + 906, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As shown in Figure 1, current video recaption methods can be broadly categorized into three types: 1) Short captions, such as Panda-70M [4], lack sufficient coverage of video content, leading to low fidelity. 2) Dense captions, like ShareGPT4Video [3], enrich textual content but suffer from hallucination issues, often generating meaningless or inaccurate video content. 3) Coarse-level structured captions, exemplified by MiraData [11], improve video quality but provide coarse-level details. Moreover, the redundancy introduced by MLLM across structures diminishes its overall effectiveness. To this end, achieving accurate captions remains two crucial challenges: 1) High fidelity between caption and video: Retain as much of the original video's objects, textures, and motion information as possible. 2) Accurate content in caption: Enable MLLM model to generate precise content, minimizing hallucinations and repetition.", + "bbox": [ + 511, + 356, + 908, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the challenges, we propose a novel instance-aware structured caption framework, termed InstanceCap, to achieve instance-level and fine-grained video caption for the first time. Our structure is specifically designed to incorporate instances, background, and camera movement. For each instance, we specify class, appearance, actions, motion, and position. To enhance the fidelity and accuracy of video captions, we focus on two key aspects: 1) From Global Video to Local Instances: For each instance, we propose an auxiliary models cluster (AMC) to isolate it from the original video and obtain the corresponding position and category information. This operation minimizes interference from unrelated regions while retaining as much of the original video's information as possible. 2) From Dense prompt to Structured Phrases: We leverage multimodal large language models (MLLMs) in an improved Chain-of-Thought (CoT) process to obtain concise yet accurate descriptions of textures, camera movement, actions and motion for each instance. This reduces the probability of hallucinations and irrelevant content produced by the lan", + "bbox": [ + 511, + 599, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contributions. Ying Tai is the corresponding author.", + "bbox": [ + 114, + 886, + 431, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28974", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a5f3c5616adef5084ff90e77d498ddf70b3e7df04b9f716f92f6dd2add08c58.jpg", + "image_caption": [ + "[InstanceCap](Ours)", + "Wrong Ambiguity Detailed", + "Figure 1. Top: Comparison of the reconstruction via-recaption results between InstanceCap and state-of-the-art captioning methods for annotating the ground truth video. InstanceCap produces results that more closely resemble the original video, showing greater detail fidelity (highlighted by the red circle). Bottom: The corresponding captions generated by InstanceCap and others. Red denotes incorrect captions, blue represents ambiguous captions, and green indicates detailed and accurate descriptions of video. Specific visual hints are marked as A, B, and C for clarity. All videos are generated using the same video generation product, Hailuo AI1, which has robust prompt-following capabilities, clearly highlighting the effectiveness of InstanceCap." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "\"Global Description\": A woman with blonde hair, wearing a dark blue shirt with white stripes, is speaking and gesturing with her hands.", + "bbox": [ + 94, + 282, + 380, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Structured Description\": {", + "bbox": [ + 102, + 297, + 186, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Main Obj", + "bbox": [ + 102, + 305, + 133, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "No.0-1 Class:\"person\",", + "bbox": [ + 102, + 313, + 165, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "\"Appearance.\" The person in the video is a woman with blonde hair, wearing glasses, her face is to the right of the frame. She is dressed in a dark blue blouse with white stripes. The woman appears to be middle-aged and is of Caucasian race. She is wearing a necklace $①$ and a black dress. She has a thick skirt and a facial expression throughout the video, indicating that she is engaged in a conversation or presentation.\"", + "bbox": [ + 94, + 323, + 398, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "\"Actions and Motion\": \"Her gestures are expressive, with her hands moving fluidly to emphasize her points and she is holding a pen. The way she speaks suggests she is passionate about the topic at hand, her body language open and inviting.\"", + "bbox": [ + 94, + 354, + 385, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "\"Position\": \"person stays at bottom-middle.\"", + "bbox": [ + 102, + 376, + 250, + 382 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "\"Background Detail\": \"The environment in the video is a stage or presentation area with a grand and elegant atmosphere. The backdrop consists of a red curtain, which adds a touch of drama and sophistication to the setting. Text square 'ROVO' can be seen in the lower right corner. The lighting is bright and evenly distributed, illuminating the stage and creating a clear and focused environment for the speaker. The overall ambiance is professional and formal, indicative of a well-organized and prestigious event.\"", + "bbox": [ + 94, + 393, + 401, + 433 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "'Camera Movement': The camera is static, maintaining a short distance shot throughout the video. The shooting angle remains consistent, capturing the speaker from the waist up.' \n1", + "bbox": [ + 94, + 433, + 395, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[PANDA-70M](CVPR 2024)", + "text_level": 1, + "bbox": [ + 410, + 271, + 500, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A woman wearing glasses and a blue tie is sitting on a red couch.", + "bbox": [ + 410, + 281, + 625, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[ShareGPT4Video](NeuIPS 2024)", + "text_level": 1, + "bbox": [ + 408, + 296, + 522, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The video features an individual standing in front of a red background, engaging in what appears to be a presentation or discussion. Throughout the video, the person is seen making various gestures with their hands, indicating a dynamic and expressive delivery. Initially, the person's hands are raised, palms facing upwards, suggesting an explanatory or emphatic gesture. As the video progresses, the individual's gestures evolve; the hands are lowered and then raised again, with the fingers pointing upwards, indicating a shift in the emphasis or point of discussion. Subsequently, the hands are brought closer together, with the fingers interlocked, suggesting a moment of emphasis or a transition in the presentation. Finally, the hands are raised again, with the fingers spread apart, indicating a continuation of the expressive gestures. Throughout the video, the individual's attire remains consistent, featuring a dark-colored top with a lighter-colored collar, and the background remains a solid red, providing a stark contrast that keeps the focus on the speaker. The camera angle and framing stay constant, with no discernible movement, maintaining a steady focus on the individual and their gestures. The lighting is even, ensuring that the subject is clearly visible against the red backdrop. There are no changes in the environment, the individual's attire, or the camera's perspective, indicating a continuous and uninterrupted presentation or discussion.", + "bbox": [ + 408, + 304, + 630, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[MiraData][NeurIPS 2024]", + "text_level": 1, + "bbox": [ + 638, + 281, + 725, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this video, a person with blond hair is engaging in an animated conversation while seated. They are wearing glasses and a striped shirt and express themselves through a series of dynamic hand gestures and head movements. The individual appears to be deeply focused on the dialogue, occasionally looking down or away, possibly to emphasize their points or formulate thoughts. The person's gestures escalate throughout the video, suggesting an increasing intensity or passion about the topic. The main subject, identifiable by their blond hair and glasses, transitions from initially looking down to making various expressive hand gestures and head movements. Their body language involves considerable arm movement, indicating engagement and enthusiasm in the conversation. The sequence begins with the individual in a somewhat reserved posture, escalating to more expansive and assertive gestures, before returning to a composed state. The background of the video is a studio setting with prominent red and white color blocks, suggesting a professional and possibly formal environment. Part of a sign or text that reads \"RM\" is visible, indicating a possible connection to a broadcast or presentation context. The exact location isn't specified, but the controlled indoor setting implies a purposeful and staged production. The view shot remains consistent, focused on the upper body of the main subject, ensuring their gestures and facial expressions are clearly visible. The camera maintains a mid-shot with minimal movement, emphasizing the individual's actions and expressions. There are no noticeable changes in the shooting angle, keeping the subject centered throughout the frames to convey their animated discourse. The video employs a straightforward, television broadcast style, emphasizing clear and focused framing to capture the subject's expressive communication in a professional studio environment.", + "bbox": [ + 638, + 289, + 900, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "guage model compared to traditional caption methods that directly describe video content in a complex, dense caption.", + "bbox": [ + 89, + 569, + 482, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate the effectiveness of InstanceCap in T2V generation, we constructed a high-definition video dataset comprising 22K samples to create a training dataset with our instance-aware structured captions, named InstanceVid. At the inference stage, we also implemented a prompt enhancement pipeline tailored to our structured captioning method, enabling the generation of concise captions that better align with user needs. Our InstanceCap integrates seamlessly with existing diffusion models. Experimental results demonstrate that after finetuned with our InstanceVid, the T2V model exhibits better ability with prompt following on details and motion actions. In summary, our main contributions are as follows:", + "bbox": [ + 89, + 603, + 483, + 800 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose InstanceCap, the first instance-aware structured caption method for text-to-video generation. We also constructed a 22K InstanceVid dataset during training, and developed an enhancement pipeline that tailored to InstanceCap structure during inference.", + "- We design the AMC paradigm to convert global video" + ], + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "into instances, enhancing instance fidelity. Additionally, we propose an improved CoT pipeline with MLLMs to refine dense prompts into structured phrases, achieving concise yet precise descriptions.", + "bbox": [ + 524, + 569, + 906, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments on video reconstruction demonstrate that our InstanceCap significantly enhances the fidelity between captions and videos. T2V models fine-tuned on our InstanceVid further achieve more precise generation on instance details and motion actions.", + "bbox": [ + 513, + 630, + 906, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 717, + 663, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video recaptioning. Advancements in text-to-video generation demand high-quality video-text datasets to build robust foundational video models for visual-language alignment. Current video recaption methods fall into two main categories: manual annotation [1, 25, 39] and end-to-end recaption using multimodal large language models (MLLM). Although manual annotation provides higher accuracy, scaling datasets to meet the needs of high-quality video generation models remains a substantial challenge. Recent", + "bbox": [ + 511, + 742, + 906, + 878 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 https://hailuoai.com/video", + "bbox": [ + 531, + 887, + 679, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28975", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "advances in MLLM have demonstrated impressive capabilities in video understanding and description generation. Panda [4] and InternVid [27] are with short captions, offering computational efficiency but frequently omitting crucial content and exhibit low video fidelity. OpenVid-1M [18], Vript [30] and ShareGPT4Video [3] are with dense captions, which provide richer content but face challenges: Hallucinations due to complexity, inclusion of redundant information, and text encoder truncation caused by excessively long text. MiraData [11] achieves coarse-level structured captions that attempts to mediate these issues but struggle with fine detail and redundancy across structures. Different from the existing video recaption methods, InstanceCap is the first instance-aware structured caption approach for text-to-video generation, ensuring high fidelity between caption and video while reducing hallucinations and repetition.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-video generation. Despite numerous high-quality video generation models [2, 10, 12, 13, 19, 32] perform well with simple directives, they often falter with complex prompts requiring precise instance-level details or intricate camera movements. Analyzing current video-text datasets suggests these limitations may stem from suboptimal training data quality. Traditional recaption methods have not sufficiently captured instance-specific detail granularity or provided comprehensive descriptions of camera movements. To enhance instance-level details and motion consistency, we construct a 22K InstanceVid dataset for training/finetuning, and develop an enhancement pipeline InstanceEnhancer that tailored to the proposed instance-aware structure during inference.", + "bbox": [ + 89, + 352, + 482, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 577, + 181, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Section 3.1, we first present the InstanceCap pipeline, as shown in Figure 2. Based on this pipeline, we recapture the carefully curated dataset InstanceVid in Section 3.2, enhancing T2V models' instance generation. Additionally, in Section 3.3, we introduce InstanceEnhancer to convert short prompts into our proposed instance-aware structured caption format during inference.", + "bbox": [ + 89, + 603, + 482, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. InstanceCap", + "text_level": 1, + "bbox": [ + 89, + 719, + 223, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video preprocessing with auxiliary model cluster. For continuous video processing, we implemented uniform sampling using $\\text{decord}^2$ , following the methodology established in LLaVA-Video [36]. This approach enables us to extract essential temporal metadata, including duration, frame count, and timestamps, allowing MLLMs to better interpret temporal sequences in recaptioning tasks. Additionally, to enhance MLLMs' capabilities through structured guidance, we incorporate several auxiliary models to", + "bbox": [ + 89, + 741, + 482, + 876 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "achieve accurate object detection, video instance segmentation and camera motion prediction, providing precise prior information to the MLLMs.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Global description, background detail and camera movement. When describing video content, a high-quality global description should capture primary elements, environmental context, camera movements, angles, and tonal qualities. MLLMs excel in generating high-level video summaries using Chain-of-Thought methodology. By employing carefully designed prompts with CoT, we can guide MLLMs to produce detailed background descriptions while minimizing references to foreground elements.", + "bbox": [ + 511, + 140, + 903, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, MLLMs' limitations in processing discrete frames rather than continuous video segments make it challenging to distinguish camera motion from instance action. To address this, we achieve camera annotations from OpenSora [37] for basic movements (e.g., zoom, rotation) and rely on MLLMs to capture subtle motion attributes (e.g., intensity, speed). The integration of camera movement indicators with MLLM capabilities provides comprehensive annotations, as illustrated in Figure 9 (a).", + "bbox": [ + 511, + 277, + 903, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Structured Description on Instances. In this subsection, we introduce the details to achieve our instance-aware structured description. To address MLLMs' limitations in instance annotation and the suboptimal results of directly adapting weak visual prompts from images to videos [22, 33], we make full use of the auxiliary model cluster, including initial object detection [6, 40], video instance segmentation with SAM2 [20], and blur non-instance regions to achieve blur background, resulting in better recaptioning outcomes compared to alternative visual prompt methods in video, as shown in Figure 10. To this end, we decompose the global videos into local instances.", + "bbox": [ + 511, + 417, + 903, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next, we describe how to achieve detailed and accurate description of each instance (Figure 3). To maintain instance-level precision, we deliberately constrain the information accessible to MLLMs during instance annotation. Crucially, the global video remains invisible due to our designed blurred backgrounds, preventing MLLMs from confusing information across multiple instances. This approach allows InstanceCap to focus on local instances identified through auxiliary model cluster. Furthermore, to avoid the potential limitation of MLLMs seeing only isolated instances, which could lead to overlooking inter-instance interactions and subsequent misinterpretations, we incorporate the global description mentioned in previous subsection. Specifically, we inject the global description into the instance-annotation MLLMs. This strategic mitigates potential biases in instance action descriptions while maintaining instance-specific accuracy.", + "bbox": [ + 511, + 598, + 903, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance the capability of InstanceCap in capturing instance-level details, we introduce novel insights into the improved CoT process. Our analysis of current MLLM-", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/dmcl/decord", + "bbox": [ + 107, + 886, + 282, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28976", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/35e2ebb0d3939a7c4875f0e451dc28ac4d52f26a3aee144daa15278081862f8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 89, + 883, + 383 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/855d4717d08bc26500404d5bec9a3fccd0368c2d5cd5d32deea05cacc31581f5.jpg", + "image_caption": [ + "Figure 2. Overview of InstanceCap pipeline. Details of \"from dense prompts to structured phrases\" design are shown in Figure 3.", + "Figure 3. Details on \"from dense prompts to structured phrases\" design. We propose an improved CoT pipeline with carefully designed information interactions (red arrow), which facilitates MLLMs to accurately capture instances with precise descriptions on attributes." + ], + "image_footnote": [], + "bbox": [ + 155, + 404, + 839, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "based video recaption methods shows that simple prompts, like “Please provide a detailed description of this video” or Chain-of-Thought prompts “Let’s think step by step... First, please note... Finally, summarize the video content...” fail to capture precise instance details. Additional experiments reveal that MLLMs can effectively annotate these details when given fine-grained prompts, such as “Please note if the characters have any accessories” or “Please observe whether there are spots on the bananas”. To enhance the details of instances, we develop Human-designed Class Hints, crafting specific prompts for about 80 detectable categories using our auxiliary models cluster. Specifically, we present the “person” class prompt here: “Please focus primarily on the person’s facial expressions, attire, age, gender, and race in the video and give a detailed description. Please mention if there are any necklaces, watches, hat or other decoration; otherwise, there’s no need to bring them up.” Besides, we also developed a curated Posi", + "bbox": [ + 88, + 625, + 485, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tive/Negative Lexicon to guide MLLMs in generating more aesthetically refined captions. More details can be found in our supplementary material.", + "bbox": [ + 511, + 625, + 906, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. InstanceVid", + "text_level": 1, + "bbox": [ + 511, + 680, + 643, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "InstanceVid is curated via refining a subset from the high-aesthetic, high-consistency videos from OpenVid-1M [18]. To showcase our method's high-fidelity labeling of instance details and motion, we selected video samples that included at least one instance exhibiting high motion intensity during dataset filtering. More statistical information of InstanceVidis provided in the supplementary material.", + "bbox": [ + 511, + 702, + 906, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. InstanceEnhancer", + "text_level": 1, + "bbox": [ + 511, + 818, + 691, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the caption distribution of training data differs from that of inference text, it may result in poor instruction-following performance or even problematic outputs. This issue is evident in T2V generation, particularly when long", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28977", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/627a420102a76c64088e89fb60a929e73622418d1f94b825207ac3ced85ed05a.jpg", + "image_caption": [ + "Figure 4. High-level overview of InstanceEnhancer, illustrating the data flow and the partitioning of stages. For a detailed implementation, refer to the supplemental materials, which provide an in-depth description of the enhancer pipeline design and the interdependencies between the stages." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 480, + 156 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "captions are used for training but short captions for inference, leading to subpar results. Since users typically prefer short captions, it is essential to enhance short caption effectively to better align with our proposed instance-aware structured caption during training.", + "bbox": [ + 89, + 243, + 483, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 4, we introduce a tuning-free approach called InstanceEnhancer that achieves this by strictly limiting the generated formats to match the caption corresponding to the training input we used. Our method differs from existing tuning-free caption enhancement approaches, such as those presented in RPG and RAG [5, 31]. Instead of directly enhancing short captions, which we found can introduce inconsistencies between multiple instances' actions and their environmental context in video generation, we employ a two-stage enhancement strategy. In Stage A, short prompts are expanded into detailed long prompts. Stage B(I)&(II) uses both expanded and original captions to segment and enhance specific instances, preserving contextual coherence while ensuring precise instance identification. The quantitative evaluation results are presented in Table 2. Due to the space limitation, more details of our enhancement pipeline can be found in the supplementary material.", + "bbox": [ + 91, + 320, + 483, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 594, + 223, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 619, + 281, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video reconstruction with recaptions. To comprehensively evaluate InstanceCap, we conducted a series of experiments, benchmarking against state-of-the-art methods including Panda-70M [4], ShareGPT4Video [3], and Mi-raData [11]. To this end, we carefully selected 500 video clips from OpenVid-1M [18]. For each video, we generated one caption using various caption models, which were then input into the advanced T2V model CogvideoX-5b [32] for video generation. We calculated the differences between the generated videos and the ground truth videos to evaluate each caption model's performance, where smaller visual differences indicate more accurate captions and higher fidelity.", + "bbox": [ + 89, + 642, + 482, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduced several metrics to evaluate the video reconstruction performance: 1) $3\\mathrm{DVAE}_{score}$ : Using 3DVAE from CogVideoX [29, 32] as the backbone, we extract hidden-space representations from both the original videos", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9c5c8a17c65e52a276d1bc18f7f784cab315da6ea606313f8a65912570e59848.jpg", + "image_caption": [ + "Figure 5. Comparison on reconstruction via-recapiction between InstanceCap and MiraData. Corresponding 3DVAE scores are also indicated. Similar semantics shared between InstanceCap and GT are indicated by red circles and lines." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 903, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and their recaption-reconstructed counterparts. These representations quantify the perceptual distance between them. 2) VQAScore: In order to obtain text-video consistency evaluation results that match human preferences, we use VQAScore [15] for evaluation. Specifically, we sample the video uniformly, and then calculate the VQA value for this set of sampled frames and their corresponding caption. This score is positively correlated with caption quality, providing an objective measure of text-video alignment. 3) Human Evaluation: We conducted a user study with a panel of evaluators to assess caption quality across two aspects: Instance Detail (ID) and Hallucination Scores (HS).", + "bbox": [ + 511, + 393, + 906, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "T2V generation. To thoroughly evaluate the T2V generation performance of our InstanceCap, we utilize the InstanceVid dataset to finetune the state-of-the-art DiT-based T2V generation model Open-Sora [37]. In our evaluation, we compare with Open-Sora, CogVideoX-5b [32], Pyramid-Flow [10], and Open-Sora-Plan [13]. To enable fine-grained, instance-level assessment, we construct a highly challenging evaluation benchmark called Inseval, inspired by recent advancements in T2I and T2V evaluation [7, 8, 15, 16, 28]. Specifically, we curate a diverse evaluation dataset of over 200 carefully crafted instance-level prompt-answer pairs, covering both single-object and multi-object scenarios systematically across five key dimensions: Action, Color, Shape, Texture, and Detail. Motivated by the previous evaluation benchmarks [8], we implement a CoT reasoning framework for generating structured QA responses to ensure objective and consistent evaluation, allowing us to derive instance-level evaluation scores that align closely with human perception and preferences. This approach provides a more nuanced and reliable assessment of instance-level generation quality.", + "bbox": [ + 511, + 583, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28978", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/25ded0840f93a18ca00a23ebbe851b39708636a38609b95ba0162698cad33ff2.jpg", + "image_caption": [ + "Single Action: A swift lion sprints across the plain." + ], + "image_footnote": [], + "bbox": [ + 112, + 89, + 493, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42613abec6fcd1b50fc62ddf9a7cba2ed3d632bcc2b462bed4b62fc3496e9e91.jpg", + "image_caption": [ + "Multiple Action: A fish swims and a turtle moves on the seabed.", + "Figure 6. Visual comparison of InstanceCap and Opensa on Single and Multiple Action Score. In terms of the dynamic degree of video generation, we show better consistency and enhanced multi-instance dynamic generation effect." + ], + "image_footnote": [], + "bbox": [ + 498, + 90, + 883, + 229 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7b714acb84caed20dd7219684ce867a9e36865ae47d07a3a9af2c589467b2767.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Captioning Methods3DVAEscore↓VQAScore↑Avg. Length
Panda-70M [4]146.630.79913 words
ShareGPT4Video [3]147.000.866174 words
LLaVA-Video-72B [36]147.630.904109 words
MiraData(GPT-4V) [11]144.750.905197 words
InstanceCap(Ours)143.000.916165 words
Ablation Study
InstanceCap w/o camera hints144.380.906166 words
InstanceCap w/o human hints143.870.908165 words
InstanceCap w/ red circle144.120.906162 words
InstanceCap w/ blue background144.380.903165 words
", + "bbox": [ + 109, + 291, + 478, + 419 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/483c46ded3d890e815aebb61b3444e2ae94beb85ab2986e5e13c7ea053be48a1.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on reconstruction-via-recaption results. The best results are marked in **bold**, and the second-best are **underscored**. As a reference, CogVideoX-5b accepts 226 text tokens, with any excess being truncated. Meanwhile, the results of the ablation study are shown in the lower part of the table." + ], + "table_footnote": [], + "table_body": "
T2V ModelSingle†Multiple†Average†
ActionColorShapeTextureDetailActionColorTexture
CogVideoX-5B [32]64%60%44%60%20%8%48%40%43.00%
Pyramid-Flow-2B [10]44%68%32%32%7%4%24%16%28.38%
Open-Sora Plan v1.3-2.7B [13]64%44%36%32%27%20%32%12%33.38%
Open-Sora v1.2-1.1B [37]40%56%36%40%13%12%16%16%28.63%
+ Ours w/o InstanceEnhancer56%60%40%48%27%16%32%24%37.88%
+ Ours w/ InstanceEnhancer48%60%56%64%20%24%48%40%45.00%
+ Panda-capitalizer [4]40%48%28%40%20%8%20%12%27.00%
+ ShareGPT4Video [3]40%44%32%24%13%16%8%20%24.63%
+ MiraData [11]44%48%40%44%13%12%28%16%30.63%
+ LLaVA [18]52%52%28%28%20%12%28%16%29.50%
", + "bbox": [ + 91, + 526, + 483, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison between InstanceCap and SOTA video captioning models, all based on the popular T2V model Open-Sora. Additionally, we also compare three powerful T2V models, including CogVideoX-5B, Pyramid-Flow, and Open-Sora Plan. The best results of video captioning methods and Open-Sora are marked in bold, and the second-best are underscored.", + "bbox": [ + 89, + 633, + 483, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with SOTA Caption Models", + "text_level": 1, + "bbox": [ + 89, + 726, + 439, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative evaluation. CogVideoX-5b [32] is a latent text-to-video generation model known for its capability to generate realistic, long-duration videos based on text prompts. With the integration of our InstanceCap into CogVideoX-5b, as substantiated by Figure 5, the model exhibits a notable enhancement in the video reconstruction capacity. This demonstrates that our instance-aware structured captions retain more of the original video's information, leading to higher fidelity. For instance, our InstanceCap can retain information such as \"glasses\",", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b5bb88d6c323125e35aaac5276fd94defd246c28fe0d2f01d5154a3fd6f37c9e.jpg", + "image_caption": [ + "Figure 7. User study on instance detail and hallucination scores. Ours instance-aware structured caption shows clear advantages compared to the coarse-structured MiraData [11]." + ], + "image_footnote": [], + "bbox": [ + 524, + 311, + 684, + 426 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1156a94ae26aefb36404a47728809950623b8e6e6c6f13753d75ff969ccc1992.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 311, + 897, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"grey sweater\", and \"relative position of two people\", whereas MiraData [11] almost completely loses these important details. A similar conclusion can be drawn from Figure 1. These results underscore the significant improvements achieved by our InstanceCap, resulting in high-quality reconstruction characterized by rich detail and high fidelity between our captions and the original videos.", + "bbox": [ + 511, + 536, + 906, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative evaluation. Table 1 presents quality comparisons between our InstanceCap and other caption methods across two metrics. Based on the results, we make the following observations: 1) Our method delivers comparable or superior quality to the four baselines, demonstrating its ability to enhance fidelity between videos and captions. This strong alignment with human perceptual judgments and preferences is evident in Figures 1 and 5. 2) Our method consistently excels across all metrics for captions under 200 words, an acceptable length for most T2V models, illustrating its generalizability. Figure 7 compares MiraData's captions with our instance-aware structured ones. We randomly selected videos from an open-domain dataset, and a panel of evaluators assessed caption quality using standardized criteria. Results indicate that our captions offer significantly richer and more accurate de", + "bbox": [ + 511, + 659, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28979", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dbb6b3c0296763b112dc945e5bebe9768f568cf0fd9592cbfd229f08bf5f9ad2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
T2V ModelQuality Metrics↑Temporal Metrics↑Motion MetricsTV Align Metrics
Action Score↑Motion AC Score↑Flow Score↑CLIP Score↑Blip-Blue↑SD Score↑
Open-Sora v1.2-1.1B53.2258.4867.87440.6520.5522.8468.19
+ Panda-captioner54.9857.2856.85440.6719.8119.5867.91
+ ShareGPT4Video52.7560.6768.55440.6620.2823.1067.95
+ LLaVA54.7960.1769.96440.5720.5323.1868.18
+ MiraData54.7160.2665.00440.3720.5322.9868.23
+ InstanceCap(Ours)54.7156.5772.04460.6820.5423.1968.88
", + "bbox": [ + 91, + 88, + 906, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results on Evalcrafter. Whereas dataset quality determines aesthetic metrics, InstanceCap significantly enhances dynamism and text-video alignment, with acceptable trade-offs in temporal consistency. Consistent with Inseval's performance.", + "bbox": [ + 89, + 205, + 906, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c8be9b24fc73d5653bdf8d9a39dee3f156d0ff2a8c92dc057dbf50875754167.jpg", + "image_caption": [ + "Multiple Color: A sapphire blue keychain hangs from a light brown bag." + ], + "image_footnote": [], + "bbox": [ + 112, + 250, + 455, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d151e9f76d61556484c759f8370b745af50fba03b537e9287a875b244a8750c2.jpg", + "image_caption": [ + "Single Shape: A HEXAGONAL HELMET rests on the seat of the bench, near the window." + ], + "image_footnote": [], + "bbox": [ + 112, + 356, + 455, + 449 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6c47b03e838fa780652532d91699f187ba8f1a15d5b9b964380a005359704f0.jpg", + "image_caption": [ + "Single Texture: A smooth paper bag rests on the counter near the coffee machine.", + "Figure 8. Visual comparison of InstanceCap and Open-Sora on instance-level attributes. InstanceCap excels in precise instance detail fidelity and instruction-following capabilities, even with complex multi-instance and multi-attribute scenarios." + ], + "image_footnote": [], + "bbox": [ + 112, + 459, + 454, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "scriptions while reducing hallucination artifacts compared to MiraData's output.", + "bbox": [ + 89, + 638, + 482, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Text-to-Video Generation", + "text_level": 1, + "bbox": [ + 89, + 681, + 323, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative evaluation. Figures 6 and 8 provide visual comparisons of the T2V generation results. It can be observed that the infusion of our InstanceVid dataset into Open-Sora [37] serves to further enhance its video synthesis capabilities across four fundamental dimensions (Action, Color, Shape, and Texture). These four different aspects correspond to information in our instance-aware structured captions such as \"Actions and Motion\", \"Appearance\", etc. For instance, our model accurately generates the \"sprints\" action of the lion in Figure 6, as opposed to Open-Sora [37]. In Figure 8, benefiting from our instance-aware caption, our model generates the accurate \"light brown bag\" instance described in caption, where Open-Sora [37]", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "completely loses this instance. These results indicate that our InstanceVid can provide accurate and instance-level guiding information for video generation models.", + "bbox": [ + 511, + 250, + 906, + 295 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative evaluation. We conduct the quantitative evaluation for InstanceCap using the proposed Inseval metrics in Table 2. We can draw the following conclusions: 1) Fine-tuning with InstanceVid consistently improves all metrics over the base model Open-Sora, demonstrating the effectiveness of InstanceCap. In particular, our Detail score ranks first, justifying the capacity of InstanceCap to capture complex instance detail in video. 2) Compared to other video captioning models finetuned based on Open-Sora, InstanceCap shows clear advantages in video generation tasks. 3) Compared to larger models like CogVideoX or Pyramid-Flow, our approach achieves a higher average metric than Pyramid-Flow, and performs comparably to CogVideoX in several specific metrics like 'Single-Color/Shape/Detail' and 'Multiple-Action', but with much fewer parameters. While primarily focused on enhancing text-to-video instance generation, our method achieves competitive results across EvalCrafter [16] (Table 3).", + "bbox": [ + 511, + 301, + 908, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 585, + 666, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of human design class hints and camera movement hints. We discuss the impact of incorporating human-designed class hints and camera movement hints on annotation outcomes and provide relevant caption visualizations in Figure 9. These annotations aid MLLMs in focusing more precisely on key elements, resulting in richer and more accurate annotations.", + "bbox": [ + 511, + 608, + 905, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablations on different video visual prompts. Comparative results for various video visual prompt methods used in caption generation are shown in Figure 10. As illustrated in (a), weak visual prompts derived from static image techniques, such as red circles, bounding boxes, or selective grayscale manipulation of non-target areas [22, 33], limit MLLMs in distinguishing and describe specific targets in multi-instance scenes, leading to attribute blending and vague annotations across instances. In contrast, our method excels in instance-specific feature extraction, accurately differentiating figures like the coach and players. Figure 10 (b) illustrates strong visual prompts that involve complete oc-", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28980", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20f4ae0b14ce1faef2e636b99076c61ad9dd9f2bab56ced8d0cd0a04c3c9139d.jpg", + "image_caption": [ + "zoom in" + ], + "image_footnote": [], + "bbox": [ + 112, + 109, + 171, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04eae64f693c8f9f0af39ad090606112daf0355965c6e6ea5c0eeb3de511b485.jpg", + "image_caption": [ + "w/o hints", + "(a)" + ], + "image_footnote": [ + "The camera movement in the video is smooth and steady, providing a clear and stable view of the scene." + ], + "bbox": [ + 200, + 94, + 815, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab579a622954ef5de8953e60bf25d0e154f1cbe575c0672ba44ad2e61d2f9939.jpg", + "image_caption": [], + "image_footnote": [ + "The camera movement in the video is a steady and gradual zoom-in, focusing on a specific area of the cityscape." + ], + "bbox": [ + 714, + 175, + 736, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c2951ba09b70e18b1442d1a4258c8a785f21f53c0dc756faae08077646b1097.jpg", + "image_caption": [ + "w/o hints" + ], + "image_footnote": [], + "bbox": [ + 124, + 223, + 165, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2211f46f60fea9f26f1dd992a7649436013730f91fac7c12a16fb5e7118262e.jpg", + "image_caption": [ + "w/ hints(ours)", + "(b)" + ], + "image_footnote": [ + "An older man, his hair white, is wearing a blue denim shirt and jeans." + ], + "bbox": [ + 197, + 210, + 815, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/596a0a559e3b180bfb8fd320b6873e5a5a39ccc4d85240d3111a1f9c35944d6f.jpg", + "image_caption": [], + "image_footnote": [ + "The person in the video is a middle-aged male with white hair. He is wearing a denim shirt and jeans, and he has a watch on his left wrist." + ], + "bbox": [ + 539, + 292, + 560, + 305 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a0bcf6c8a3d6ad6290fe4f44b91ed65c8a16982825923393e8e2b41e4e4cf062.jpg", + "image_caption": [ + "w/ hints(ours)", + "Figure 10. (a) Comparison against the weak visual prompt for reconstruction via caption visualization on multi-instance targets. (b) Comparison against color screen backgrounds (red), which may negatively affect MLLM labeling performance." + ], + "image_footnote": [], + "bbox": [ + 91, + 351, + 488, + 565 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/15f48e79a35e34bba57712e0a20d6a721176110edf77ebb22a8a83333359bfed.jpg", + "image_caption": [ + "Figure 9. (a) Ablation study on the effect of camera movement hints on the accuracy of MLLM labeling. (b) Impact of human-designed class hints on the details of instance labeling." + ], + "image_footnote": [], + "bbox": [ + 500, + 351, + 900, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "clusion of non-target regions to eliminate MLLM interactions with irrelevant instances. Conventional methods use primary color screens, but this often misguides MLLMs, causing them to incorporate incorrect context in captions. Our designed blur background masking approach, however, preserves visual consistency with natural scenes, enabling MLLMs to generate accurate and contextually relevant annotations with minimal prompting guidance. In Table 1, we present the quantitative results of the ablation study.", + "bbox": [ + 89, + 604, + 482, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions and Limitations", + "text_level": 1, + "bbox": [ + 89, + 746, + 354, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce InstanceCap, the first instance-aware structured caption method for text-to-video generation. We design an Auxiliary Models Cluster (AMC) to convert global video into instances, enhancing instance fidelity. we also propose an improved CoT pipeline with MLLMs to refine dense prompts into structured phrases, achieving concise yet precise instance descriptions compared to the previous video caption models. Additionally, based on InstanceCap, we curated InstanceVid dataset", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for training and InstanceEnhancer during inference, significantly enhancing T2V models' generation capabilities on instance details and actions.", + "bbox": [ + 511, + 604, + 905, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Since the precision of InstanceCap partly depends on object detection methods, requiring fine-tuning of the detection model for domain-specific instances, and its benefits decrease in instance-free scenes. Furthermore, the scale of InstanceVid limits its use as a large-scale pre-training dataset. Moving forward, we plan to apply InstanceCap to a larger video dataset and train more powerful T2V models to amplify its impact.", + "bbox": [ + 511, + 660, + 906, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 790, + 692, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by Natural Science Foundation of China: No. 62406135 and Natural Science Foundation of Jiangsu Province: BK20241198. Supported by the Gusu Innovation and Entrepreneur leading Talents: No. ZXL2024362. Supported by the AI & AI for Science Project of Nanjing University: No. 14380007.", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28981", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] David L. Chen and William B. Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics (ACL-2011), Portland, OR, 2011. 2", + "[2] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, pages 7310-7320, 2024. 1, 3", + "[3] Lin Chen, Xin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 1, 3, 5, 6, 2", + "[4] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. arXiv preprint arXiv:2402.19479, 2024. 1,3,5,6", + "[5] Zhennan Chen, Yajie Li, Haofan Wang, Zhibo Chen, Zhengkai Jiang, Jun Li, Qian Wang, Jian Yang, and Ying Tai. Region-aware text-to-image generation via hard binding and soft refinement. arXiv preprint arXiv:2411.06558, 2024. 5", + "[6] Xiantao Hu, Ying Tai, Xu Zhao, Chen Zhao, Zhenyu Zhang, Jun Li, Bineng Zhong, and Jian Yang. Exploiting multimodal spatial-temporal patterns for video object tracking, 2024. 3", + "[7] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. arXiv preprint arXiv:2307.06350, 2023. 5", + "[8] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5", + "[9] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1-38, 2023. 3", + "[10] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 1, 3, 5, 6", + "[11] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. Miradata: A large-scale video dataset with long durations and structured captions. arXiv preprint arXiv:2407.06358, 2024. 1, 3, 5, 6", + "[12] Kuaishou. Kling. https://kling.kuaishou.com, 2024. 1, 3", + "[13] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 3, 5, 6" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. arXiv preprint arXiv:2410.05677, 2024. 1", + "[15] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation, 2024. 5", + "[16] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. arXiv preprint arXiv:2310.11440, 2023.5.7", + "[17] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 1", + "[18] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 3, 4, 5, 6", + "[19] Pika. Pika 1.0. https://pika.art, 2023. 1, 3", + "[20] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dollar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3", + "[21] Runway. Gen-2. https://research.runwayml.com/gen2, 2023.1", + "[22] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11987-11997, 2023. 3, 7", + "[23] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 1", + "[24] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. ModelScope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1", + "[25] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research, 2020. 2", + "[26] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. LaVie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 1", + "[27] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations, 2023. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28982", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Yinwei Wu, Xianpan Zhou, Bing Ma, Xuefeng Su, Kai Ma, and Xinchao Wang. Ifadapter: Instance feature control for grounded text-to-image generation, 2024. 5", + "[29] Rui Xie, Yinhong Liu, Penghao Zhou, Chen Zhao, Jun Zhou, Kai Zhang, Zhenyu Zhang, Jian Yang, Zhenheng Yang, and Ying Tai. Star: Spatial-temporal augmentation with text-to-video models for real-world video super-resolution. arXiv preprint arXiv:2501.02976, 2025. 5", + "[30] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words, 2024. 3", + "[31] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In Forty-first International Conference on Machine Learning, 2024. 5", + "[32] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 3, 5, 6, 2", + "[33] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. AI Open, 5:30-38, 2024. 3, 7", + "[34] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. IJCV, 2024. 1", + "[35] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 2", + "[36] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data, 2024. 3, 6", + "[37] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 3, 5, 6, 7", + "[38] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-Sora: Democratizing efficient video production for all. https://github.com/hpcaitech/Open-Sora, 2024.1", + "[39] Luowei Zhou, Nathan Louis, and Jason J. Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction, 2018. 2", + "[40] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6748-6758, 2023. 3" + ], + "bbox": [ + 91, + 92, + 480, + 813 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28983", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/63b6efae-0d11-487a-b293-b1b579ef5966_model.json b/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/63b6efae-0d11-487a-b293-b1b579ef5966_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d7246932a4d14c5147d06168a89b6a86c026d237 --- /dev/null +++ b/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/63b6efae-0d11-487a-b293-b1b579ef5966_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.13, + 0.856, + 0.178 + ], + "angle": 0, + "content": "InstanceCap: Improving Text-to-Video Generation via Instance-aware Structured Caption" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.203, + 0.796, + 0.277 + ], + "angle": 0, + "content": "Tiehan Fan\\(^{1*}\\) Kepan Nan\\(^{1*}\\) Rui Xie\\(^{1}\\) Penghao Zhou\\(^{2}\\) Zhenheng Yang\\(^{2}\\) Chaoyou Fu\\(^{1}\\) Xiang Li\\(^{3}\\) Jian Yang\\(^{1}\\) Ying Tai\\(^{1\\boxtimes}\\) \n\\(^{1}\\) Nanjing University \\({}^{2}\\) ByteDance \\({}^{3}\\) Nankai University \nhttps://github.com/NJU-PCALab/InstanceCap" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.324, + 0.327, + 0.34 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.356, + 0.486, + 0.675 + ], + "angle": 0, + "content": "Text-to-video generation has evolved rapidly in recent years, delivering remarkable results. Training typically relies on video-caption paired data, which plays a crucial role in enhancing generation performance. However, current video captions often suffer from insufficient details, hallucinations and imprecise motion depiction, affecting the fidelity and consistency of generated videos. In this work, we propose a novel instance-aware structured caption framework, termed InstanceCap, to achieve instance-level and fine-grained video caption for the first time. Based on this scheme, we design an auxiliary models cluster to convert original video into instances to enhance instance fidelity. Video instances are further used to refine dense prompts into structured phrases, achieving concise yet precise descriptions. Furthermore, a 22K InstanceVid dataset is curated for training, and an enhancement pipeline that tailored to InstanceCap structure is proposed for inference. Experimental results demonstrate that our proposed InstanceCap significantly outperform previous models, ensuring high fidelity between captions and videos while reducing hallucinations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.702, + 0.222, + 0.718 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Recently, text-to-video (T2V) generation with advanced diffusion transformers (DiT) [2, 10, 12-14, 17, 19, 21, 24, 26, 34, 38] have attracted significant attention for the ability to generate realistic, long-duration videos based on text prompts. Video-caption paired data is typically used in training and plays a crucial role in enhancing generation performance [23, 32]. Current video recaption methods often incorporate multimodal large language models to produce detailed captions, which however usually suffer from hallucinations, leading to inconsistencies between captions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.907, + 0.356 + ], + "angle": 0, + "content": "and video content. Consequently, creating consistent videocaption pairs with accurate details and precise motion depiction for T2V generation remains a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.909, + 0.599 + ], + "angle": 0, + "content": "As shown in Figure 1, current video recaption methods can be broadly categorized into three types: 1) Short captions, such as Panda-70M [4], lack sufficient coverage of video content, leading to low fidelity. 2) Dense captions, like ShareGPT4Video [3], enrich textual content but suffer from hallucination issues, often generating meaningless or inaccurate video content. 3) Coarse-level structured captions, exemplified by MiraData [11], improve video quality but provide coarse-level details. Moreover, the redundancy introduced by MLLM across structures diminishes its overall effectiveness. To this end, achieving accurate captions remains two crucial challenges: 1) High fidelity between caption and video: Retain as much of the original video's objects, textures, and motion information as possible. 2) Accurate content in caption: Enable MLLM model to generate precise content, minimizing hallucinations and repetition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To address the challenges, we propose a novel instance-aware structured caption framework, termed InstanceCap, to achieve instance-level and fine-grained video caption for the first time. Our structure is specifically designed to incorporate instances, background, and camera movement. For each instance, we specify class, appearance, actions, motion, and position. To enhance the fidelity and accuracy of video captions, we focus on two key aspects: 1) From Global Video to Local Instances: For each instance, we propose an auxiliary models cluster (AMC) to isolate it from the original video and obtain the corresponding position and category information. This operation minimizes interference from unrelated regions while retaining as much of the original video's information as possible. 2) From Dense prompt to Structured Phrases: We leverage multimodal large language models (MLLMs) in an improved Chain-of-Thought (CoT) process to obtain concise yet accurate descriptions of textures, camera movement, actions and motion for each instance. This reduces the probability of hallucinations and irrelevant content produced by the lan" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.432, + 0.901 + ], + "angle": 0, + "content": "* Equal contributions. Ying Tai is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28974" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.269, + 0.165, + 0.279 + ], + "angle": 0, + "content": "[InstanceCap](Ours)" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.284, + 0.382, + 0.298 + ], + "angle": 0, + "content": "\"Global Description\": A woman with blonde hair, wearing a dark blue shirt with white stripes, is speaking and gesturing with her hands." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.298, + 0.187, + 0.305 + ], + "angle": 0, + "content": "Structured Description\": {" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.306, + 0.135, + 0.313 + ], + "angle": 0, + "content": "Main Obj" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.314, + 0.166, + 0.323 + ], + "angle": 0, + "content": "No.0-1 Class:\"person\"," + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.324, + 0.4, + 0.356 + ], + "angle": 0, + "content": "\"Appearance.\" The person in the video is a woman with blonde hair, wearing glasses, her face is to the right of the frame. She is dressed in a dark blue blouse with white stripes. The woman appears to be middle-aged and is of Caucasian race. She is wearing a necklace \\(①\\) and a black dress. She has a thick skirt and a facial expression throughout the video, indicating that she is engaged in a conversation or presentation.\"" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.356, + 0.387, + 0.377 + ], + "angle": 0, + "content": "\"Actions and Motion\": \"Her gestures are expressive, with her hands moving fluidly to emphasize her points and she is holding a pen. The way she speaks suggests she is passionate about the topic at hand, her body language open and inviting.\"" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.377, + 0.25, + 0.383 + ], + "angle": 0, + "content": "\"Position\": \"person stays at bottom-middle.\"" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.395, + 0.402, + 0.434 + ], + "angle": 0, + "content": "\"Background Detail\": \"The environment in the video is a stage or presentation area with a grand and elegant atmosphere. The backdrop consists of a red curtain, which adds a touch of drama and sophistication to the setting. Text square 'ROVO' can be seen in the lower right corner. The lighting is bright and evenly distributed, illuminating the stage and creating a clear and focused environment for the speaker. The overall ambiance is professional and formal, indicative of a well-organized and prestigious event.\"" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.434, + 0.396, + 0.459 + ], + "angle": 0, + "content": "'Camera Movement': The camera is static, maintaining a short distance shot throughout the video. The shooting angle remains consistent, capturing the speaker from the waist up.' \n1" + }, + { + "type": "title", + "bbox": [ + 0.411, + 0.272, + 0.501, + 0.282 + ], + "angle": 0, + "content": "[PANDA-70M](CVPR 2024)" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.282, + 0.627, + 0.289 + ], + "angle": 0, + "content": "A woman wearing glasses and a blue tie is sitting on a red couch." + }, + { + "type": "title", + "bbox": [ + 0.41, + 0.297, + 0.523, + 0.305 + ], + "angle": 0, + "content": "[ShareGPT4Video](NeuIPS 2024)" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.305, + 0.632, + 0.459 + ], + "angle": 0, + "content": "The video features an individual standing in front of a red background, engaging in what appears to be a presentation or discussion. Throughout the video, the person is seen making various gestures with their hands, indicating a dynamic and expressive delivery. Initially, the person's hands are raised, palms facing upwards, suggesting an explanatory or emphatic gesture. As the video progresses, the individual's gestures evolve; the hands are lowered and then raised again, with the fingers pointing upwards, indicating a shift in the emphasis or point of discussion. Subsequently, the hands are brought closer together, with the fingers interlocked, suggesting a moment of emphasis or a transition in the presentation. Finally, the hands are raised again, with the fingers spread apart, indicating a continuation of the expressive gestures. Throughout the video, the individual's attire remains consistent, featuring a dark-colored top with a lighter-colored collar, and the background remains a solid red, providing a stark contrast that keeps the focus on the speaker. The camera angle and framing stay constant, with no discernible movement, maintaining a steady focus on the individual and their gestures. The lighting is even, ensuring that the subject is clearly visible against the red backdrop. There are no changes in the environment, the individual's attire, or the camera's perspective, indicating a continuous and uninterrupted presentation or discussion." + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.27, + 0.813, + 0.28 + ], + "angle": 0, + "content": "Wrong Ambiguity Detailed" + }, + { + "type": "title", + "bbox": [ + 0.639, + 0.282, + 0.727, + 0.29 + ], + "angle": 0, + "content": "[MiraData][NeurIPS 2024]" + }, + { + "type": "text", + "bbox": [ + 0.639, + 0.29, + 0.901, + 0.459 + ], + "angle": 0, + "content": "In this video, a person with blond hair is engaging in an animated conversation while seated. They are wearing glasses and a striped shirt and express themselves through a series of dynamic hand gestures and head movements. The individual appears to be deeply focused on the dialogue, occasionally looking down or away, possibly to emphasize their points or formulate thoughts. The person's gestures escalate throughout the video, suggesting an increasing intensity or passion about the topic. The main subject, identifiable by their blond hair and glasses, transitions from initially looking down to making various expressive hand gestures and head movements. Their body language involves considerable arm movement, indicating engagement and enthusiasm in the conversation. The sequence begins with the individual in a somewhat reserved posture, escalating to more expansive and assertive gestures, before returning to a composed state. The background of the video is a studio setting with prominent red and white color blocks, suggesting a professional and possibly formal environment. Part of a sign or text that reads \"RM\" is visible, indicating a possible connection to a broadcast or presentation context. The exact location isn't specified, but the controlled indoor setting implies a purposeful and staged production. The view shot remains consistent, focused on the upper body of the main subject, ensuring their gestures and facial expressions are clearly visible. The camera maintains a mid-shot with minimal movement, emphasizing the individual's actions and expressions. There are no noticeable changes in the shooting angle, keeping the subject centered throughout the frames to convey their animated discourse. The video employs a straightforward, television broadcast style, emphasizing clear and focused framing to capture the subject's expressive communication in a professional studio environment." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.467, + 0.907, + 0.551 + ], + "angle": 0, + "content": "Figure 1. Top: Comparison of the reconstruction via-recaption results between InstanceCap and state-of-the-art captioning methods for annotating the ground truth video. InstanceCap produces results that more closely resemble the original video, showing greater detail fidelity (highlighted by the red circle). Bottom: The corresponding captions generated by InstanceCap and others. Red denotes incorrect captions, blue represents ambiguous captions, and green indicates detailed and accurate descriptions of video. Specific visual hints are marked as A, B, and C for clarity. All videos are generated using the same video generation product, Hailuo AI1, which has robust prompt-following capabilities, clearly highlighting the effectiveness of InstanceCap." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.6 + ], + "angle": 0, + "content": "guage model compared to traditional caption methods that directly describe video content in a complex, dense caption." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.484, + 0.801 + ], + "angle": 0, + "content": "To validate the effectiveness of InstanceCap in T2V generation, we constructed a high-definition video dataset comprising 22K samples to create a training dataset with our instance-aware structured captions, named InstanceVid. At the inference stage, we also implemented a prompt enhancement pipeline tailored to our structured captioning method, enabling the generation of concise captions that better align with user needs. Our InstanceCap integrates seamlessly with existing diffusion models. Experimental results demonstrate that after finetuned with our InstanceVid, the T2V model exhibits better ability with prompt following on details and motion actions. In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.886 + ], + "angle": 0, + "content": "- We propose InstanceCap, the first instance-aware structured caption method for text-to-video generation. We also constructed a 22K InstanceVid dataset during training, and developed an enhancement pipeline that tailored to InstanceCap structure during inference." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We design the AMC paradigm to convert global video" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.57, + 0.907, + 0.63 + ], + "angle": 0, + "content": "into instances, enhancing instance fidelity. Additionally, we propose an improved CoT pipeline with MLLMs to refine dense prompts into structured phrases, achieving concise yet precise descriptions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.631, + 0.907, + 0.705 + ], + "angle": 0, + "content": "- Extensive experiments on video reconstruction demonstrate that our InstanceCap significantly enhances the fidelity between captions and videos. T2V models fine-tuned on our InstanceVid further achieve more precise generation on instance details and motion actions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.718, + 0.664, + 0.734 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.743, + 0.907, + 0.88 + ], + "angle": 0, + "content": "Video recaptioning. Advancements in text-to-video generation demand high-quality video-text datasets to build robust foundational video models for visual-language alignment. Current video recaption methods fall into two main categories: manual annotation [1, 25, 39] and end-to-end recaption using multimodal large language models (MLLM). Although manual annotation provides higher accuracy, scaling datasets to meet the needs of high-quality video generation models remains a substantial challenge. Recent" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.888, + 0.68, + 0.9 + ], + "angle": 0, + "content": "1 https://hailuoai.com/video" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28975" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "advances in MLLM have demonstrated impressive capabilities in video understanding and description generation. Panda [4] and InternVid [27] are with short captions, offering computational efficiency but frequently omitting crucial content and exhibit low video fidelity. OpenVid-1M [18], Vript [30] and ShareGPT4Video [3] are with dense captions, which provide richer content but face challenges: Hallucinations due to complexity, inclusion of redundant information, and text encoder truncation caused by excessively long text. MiraData [11] achieves coarse-level structured captions that attempts to mediate these issues but struggle with fine detail and redundancy across structures. Different from the existing video recaption methods, InstanceCap is the first instance-aware structured caption approach for text-to-video generation, ensuring high fidelity between caption and video while reducing hallucinations and repetition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.565 + ], + "angle": 0, + "content": "Text-to-video generation. Despite numerous high-quality video generation models [2, 10, 12, 13, 19, 32] perform well with simple directives, they often falter with complex prompts requiring precise instance-level details or intricate camera movements. Analyzing current video-text datasets suggests these limitations may stem from suboptimal training data quality. Traditional recaption methods have not sufficiently captured instance-specific detail granularity or provided comprehensive descriptions of camera movements. To enhance instance-level details and motion consistency, we construct a 22K InstanceVid dataset for training/finetuning, and develop an enhancement pipeline InstanceEnhancer that tailored to the proposed instance-aware structure during inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.578, + 0.182, + 0.594 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.483, + 0.71 + ], + "angle": 0, + "content": "In Section 3.1, we first present the InstanceCap pipeline, as shown in Figure 2. Based on this pipeline, we recapture the carefully curated dataset InstanceVid in Section 3.2, enhancing T2V models' instance generation. Additionally, in Section 3.3, we introduce InstanceEnhancer to convert short prompts into our proposed instance-aware structured caption format during inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.72, + 0.225, + 0.736 + ], + "angle": 0, + "content": "3.1. InstanceCap" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Video preprocessing with auxiliary model cluster. For continuous video processing, we implemented uniform sampling using \\( \\text{decord}^2 \\), following the methodology established in LLaVA-Video [36]. This approach enables us to extract essential temporal metadata, including duration, frame count, and timestamps, allowing MLLMs to better interpret temporal sequences in recaptioning tasks. Additionally, to enhance MLLMs' capabilities through structured guidance, we incorporate several auxiliary models to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "achieve accurate object detection, video instance segmentation and camera motion prediction, providing precise prior information to the MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.141, + 0.905, + 0.277 + ], + "angle": 0, + "content": "Global description, background detail and camera movement. When describing video content, a high-quality global description should capture primary elements, environmental context, camera movements, angles, and tonal qualities. MLLMs excel in generating high-level video summaries using Chain-of-Thought methodology. By employing carefully designed prompts with CoT, we can guide MLLMs to produce detailed background descriptions while minimizing references to foreground elements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.905, + 0.413 + ], + "angle": 0, + "content": "However, MLLMs' limitations in processing discrete frames rather than continuous video segments make it challenging to distinguish camera motion from instance action. To address this, we achieve camera annotations from OpenSora [37] for basic movements (e.g., zoom, rotation) and rely on MLLMs to capture subtle motion attributes (e.g., intensity, speed). The integration of camera movement indicators with MLLM capabilities provides comprehensive annotations, as illustrated in Figure 9 (a)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.905, + 0.598 + ], + "angle": 0, + "content": "Structured Description on Instances. In this subsection, we introduce the details to achieve our instance-aware structured description. To address MLLMs' limitations in instance annotation and the suboptimal results of directly adapting weak visual prompts from images to videos [22, 33], we make full use of the auxiliary model cluster, including initial object detection [6, 40], video instance segmentation with SAM2 [20], and blur non-instance regions to achieve blur background, resulting in better recaptioning outcomes compared to alternative visual prompt methods in video, as shown in Figure 10. To this end, we decompose the global videos into local instances." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.905, + 0.856 + ], + "angle": 0, + "content": "Next, we describe how to achieve detailed and accurate description of each instance (Figure 3). To maintain instance-level precision, we deliberately constrain the information accessible to MLLMs during instance annotation. Crucially, the global video remains invisible due to our designed blurred backgrounds, preventing MLLMs from confusing information across multiple instances. This approach allows InstanceCap to focus on local instances identified through auxiliary model cluster. Furthermore, to avoid the potential limitation of MLLMs seeing only isolated instances, which could lead to overlooking inter-instance interactions and subsequent misinterpretations, we incorporate the global description mentioned in previous subsection. Specifically, we inject the global description into the instance-annotation MLLMs. This strategic mitigates potential biases in instance action descriptions while maintaining instance-specific accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To enhance the capability of InstanceCap in capturing instance-level details, we introduce novel insights into the improved CoT process. Our analysis of current MLLM-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.283, + 0.9 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/dmcl/decord" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28976" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.09, + 0.885, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.388, + 0.887, + 0.403 + ], + "angle": 0, + "content": "Figure 2. Overview of InstanceCap pipeline. Details of \"from dense prompts to structured phrases\" design are shown in Figure 3." + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.405, + 0.841, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.572, + 0.908, + 0.601 + ], + "angle": 0, + "content": "Figure 3. Details on \"from dense prompts to structured phrases\" design. We propose an improved CoT pipeline with carefully designed information interactions (red arrow), which facilitates MLLMs to accurately capture instances with precise descriptions on attributes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.626, + 0.486, + 0.899 + ], + "angle": 0, + "content": "based video recaption methods shows that simple prompts, like “Please provide a detailed description of this video” or Chain-of-Thought prompts “Let’s think step by step... First, please note... Finally, summarize the video content...” fail to capture precise instance details. Additional experiments reveal that MLLMs can effectively annotate these details when given fine-grained prompts, such as “Please note if the characters have any accessories” or “Please observe whether there are spots on the bananas”. To enhance the details of instances, we develop Human-designed Class Hints, crafting specific prompts for about 80 detectable categories using our auxiliary models cluster. Specifically, we present the “person” class prompt here: “Please focus primarily on the person’s facial expressions, attire, age, gender, and race in the video and give a detailed description. Please mention if there are any necklaces, watches, hat or other decoration; otherwise, there’s no need to bring them up.” Besides, we also developed a curated Posi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.907, + 0.673 + ], + "angle": 0, + "content": "tive/Negative Lexicon to guide MLLMs in generating more aesthetically refined captions. More details can be found in our supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.644, + 0.696 + ], + "angle": 0, + "content": "3.2. InstanceVid" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.81 + ], + "angle": 0, + "content": "InstanceVid is curated via refining a subset from the high-aesthetic, high-consistency videos from OpenVid-1M [18]. To showcase our method's high-fidelity labeling of instance details and motion, we selected video samples that included at least one instance exhibiting high motion intensity during dataset filtering. More statistical information of InstanceVidis provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.692, + 0.833 + ], + "angle": 0, + "content": "3.3. InstanceEnhancer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "When the caption distribution of training data differs from that of inference text, it may result in poor instruction-following performance or even problematic outputs. This issue is evident in T2V generation, particularly when long" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28977" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.091, + 0.482, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.16, + 0.483, + 0.23 + ], + "angle": 0, + "content": "Figure 4. High-level overview of InstanceEnhancer, illustrating the data flow and the partitioning of stages. For a detailed implementation, refer to the supplemental materials, which provide an in-depth description of the enhancer pipeline design and the interdependencies between the stages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.484, + 0.32 + ], + "angle": 0, + "content": "captions are used for training but short captions for inference, leading to subpar results. Since users typically prefer short captions, it is essential to enhance short caption effectively to better align with our proposed instance-aware structured caption during training." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.321, + 0.485, + 0.579 + ], + "angle": 0, + "content": "As shown in Figure 4, we introduce a tuning-free approach called InstanceEnhancer that achieves this by strictly limiting the generated formats to match the caption corresponding to the training input we used. Our method differs from existing tuning-free caption enhancement approaches, such as those presented in RPG and RAG [5, 31]. Instead of directly enhancing short captions, which we found can introduce inconsistencies between multiple instances' actions and their environmental context in video generation, we employ a two-stage enhancement strategy. In Stage A, short prompts are expanded into detailed long prompts. Stage B(I)&(II) uses both expanded and original captions to segment and enhance specific instances, preserving contextual coherence while ensuring precise instance identification. The quantitative evaluation results are presented in Table 2. Due to the space limitation, more details of our enhancement pipeline can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.595, + 0.224, + 0.612 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.62, + 0.282, + 0.637 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Video reconstruction with recaptions. To comprehensively evaluate InstanceCap, we conducted a series of experiments, benchmarking against state-of-the-art methods including Panda-70M [4], ShareGPT4Video [3], and Mi-raData [11]. To this end, we carefully selected 500 video clips from OpenVid-1M [18]. For each video, we generated one caption using various caption models, which were then input into the advanced T2V model CogvideoX-5b [32] for video generation. We calculated the differences between the generated videos and the ground truth videos to evaluate each caption model's performance, where smaller visual differences indicate more accurate captions and higher fidelity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We introduced several metrics to evaluate the video reconstruction performance: 1) \\(3\\mathrm{DVAE}_{score}\\): Using 3DVAE from CogVideoX [29, 32] as the backbone, we extract hidden-space representations from both the original videos" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.319, + 0.907, + 0.375 + ], + "angle": 0, + "content": "Figure 5. Comparison on reconstruction via-recapiction between InstanceCap and MiraData. Corresponding 3DVAE scores are also indicated. Similar semantics shared between InstanceCap and GT are indicated by red circles and lines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.577 + ], + "angle": 0, + "content": "and their recaption-reconstructed counterparts. These representations quantify the perceptual distance between them. 2) VQAScore: In order to obtain text-video consistency evaluation results that match human preferences, we use VQAScore [15] for evaluation. Specifically, we sample the video uniformly, and then calculate the VQA value for this set of sampled frames and their corresponding caption. This score is positively correlated with caption quality, providing an objective measure of text-video alignment. 3) Human Evaluation: We conducted a user study with a panel of evaluators to assess caption quality across two aspects: Instance Detail (ID) and Hallucination Scores (HS)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.903 + ], + "angle": 0, + "content": "T2V generation. To thoroughly evaluate the T2V generation performance of our InstanceCap, we utilize the InstanceVid dataset to finetune the state-of-the-art DiT-based T2V generation model Open-Sora [37]. In our evaluation, we compare with Open-Sora, CogVideoX-5b [32], Pyramid-Flow [10], and Open-Sora-Plan [13]. To enable fine-grained, instance-level assessment, we construct a highly challenging evaluation benchmark called Inseval, inspired by recent advancements in T2I and T2V evaluation [7, 8, 15, 16, 28]. Specifically, we curate a diverse evaluation dataset of over 200 carefully crafted instance-level prompt-answer pairs, covering both single-object and multi-object scenarios systematically across five key dimensions: Action, Color, Shape, Texture, and Detail. Motivated by the previous evaluation benchmarks [8], we implement a CoT reasoning framework for generating structured QA responses to ensure objective and consistent evaluation, allowing us to derive instance-level evaluation scores that align closely with human perception and preferences. This approach provides a more nuanced and reliable assessment of instance-level generation quality." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28978" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.09, + 0.495, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.232, + 0.421, + 0.243 + ], + "angle": 0, + "content": "Single Action: A swift lion sprints across the plain." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.091, + 0.885, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.232, + 0.884, + 0.243 + ], + "angle": 0, + "content": "Multiple Action: A fish swims and a turtle moves on the seabed." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.246, + 0.907, + 0.275 + ], + "angle": 0, + "content": "Figure 6. Visual comparison of InstanceCap and Opensa on Single and Multiple Action Score. In terms of the dynamic degree of video generation, we show better consistency and enhanced multi-instance dynamic generation effect." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.292, + 0.48, + 0.42 + ], + "angle": 0, + "content": "
Captioning Methods3DVAEscore↓VQAScore↑Avg. Length
Panda-70M [4]146.630.79913 words
ShareGPT4Video [3]147.000.866174 words
LLaVA-Video-72B [36]147.630.904109 words
MiraData(GPT-4V) [11]144.750.905197 words
InstanceCap(Ours)143.000.916165 words
Ablation Study
InstanceCap w/o camera hints144.380.906166 words
InstanceCap w/o human hints143.870.908165 words
InstanceCap w/ red circle144.120.906162 words
InstanceCap w/ blue background144.380.903165 words
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.441, + 0.502, + 0.511 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on reconstruction-via-recaption results. The best results are marked in **bold**, and the second-best are **underscored**. As a reference, CogVideoX-5b accepts 226 text tokens, with any excess being truncated. Meanwhile, the results of the ablation study are shown in the lower part of the table." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.527, + 0.485, + 0.63 + ], + "angle": 0, + "content": "
T2V ModelSingle†Multiple†Average†
ActionColorShapeTextureDetailActionColorTexture
CogVideoX-5B [32]64%60%44%60%20%8%48%40%43.00%
Pyramid-Flow-2B [10]44%68%32%32%7%4%24%16%28.38%
Open-Sora Plan v1.3-2.7B [13]64%44%36%32%27%20%32%12%33.38%
Open-Sora v1.2-1.1B [37]40%56%36%40%13%12%16%16%28.63%
+ Ours w/o InstanceEnhancer56%60%40%48%27%16%32%24%37.88%
+ Ours w/ InstanceEnhancer48%60%56%64%20%24%48%40%45.00%
+ Panda-capitalizer [4]40%48%28%40%20%8%20%12%27.00%
+ ShareGPT4Video [3]40%44%32%24%13%16%8%20%24.63%
+ MiraData [11]44%48%40%44%13%12%28%16%30.63%
+ LLaVA [18]52%52%28%28%20%12%28%16%29.50%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.635, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison between InstanceCap and SOTA video captioning models, all based on the popular T2V model Open-Sora. Additionally, we also compare three powerful T2V models, including CogVideoX-5B, Pyramid-Flow, and Open-Sora Plan. The best results of video captioning methods and Open-Sora are marked in bold, and the second-best are underscored." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.44, + 0.745 + ], + "angle": 0, + "content": "4.2. Comparison with SOTA Caption Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Qualitative evaluation. CogVideoX-5b [32] is a latent text-to-video generation model known for its capability to generate realistic, long-duration videos based on text prompts. With the integration of our InstanceCap into CogVideoX-5b, as substantiated by Figure 5, the model exhibits a notable enhancement in the video reconstruction capacity. This demonstrates that our instance-aware structured captions retain more of the original video's information, leading to higher fidelity. For instance, our InstanceCap can retain information such as \"glasses\"," + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.313, + 0.685, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.313, + 0.898, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Figure 7. User study on instance detail and hallucination scores. Ours instance-aware structured caption shows clear advantages compared to the coarse-structured MiraData [11]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.645 + ], + "angle": 0, + "content": "\"grey sweater\", and \"relative position of two people\", whereas MiraData [11] almost completely loses these important details. A similar conclusion can be drawn from Figure 1. These results underscore the significant improvements achieved by our InstanceCap, resulting in high-quality reconstruction characterized by rich detail and high fidelity between our captions and the original videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Quantitative evaluation. Table 1 presents quality comparisons between our InstanceCap and other caption methods across two metrics. Based on the results, we make the following observations: 1) Our method delivers comparable or superior quality to the four baselines, demonstrating its ability to enhance fidelity between videos and captions. This strong alignment with human perceptual judgments and preferences is evident in Figures 1 and 5. 2) Our method consistently excels across all metrics for captions under 200 words, an acceptable length for most T2V models, illustrating its generalizability. Figure 7 compares MiraData's captions with our instance-aware structured ones. We randomly selected videos from an open-domain dataset, and a panel of evaluators assessed caption quality using standardized criteria. Results indicate that our captions offer significantly richer and more accurate de" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28979" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.196 + ], + "angle": 0, + "content": "
T2V ModelQuality Metrics↑Temporal Metrics↑Motion MetricsTV Align Metrics
Action Score↑Motion AC Score↑Flow Score↑CLIP Score↑Blip-Blue↑SD Score↑
Open-Sora v1.2-1.1B53.2258.4867.87440.6520.5522.8468.19
+ Panda-captioner54.9857.2856.85440.6719.8119.5867.91
+ ShareGPT4Video52.7560.6768.55440.6620.2823.1067.95
+ LLaVA54.7960.1769.96440.5720.5323.1868.18
+ MiraData54.7160.2665.00440.3720.5322.9868.23
+ InstanceCap(Ours)54.7156.5772.04460.6820.5423.1968.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.206, + 0.907, + 0.234 + ], + "angle": 0, + "content": "Table 3. Results on Evalcrafter. Whereas dataset quality determines aesthetic metrics, InstanceCap significantly enhances dynamism and text-video alignment, with acceptable trade-offs in temporal consistency. Consistent with Inseval's performance." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.25, + 0.457, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.346, + 0.418, + 0.356 + ], + "angle": 0, + "content": "Multiple Color: A sapphire blue keychain hangs from a light brown bag." + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.357, + 0.456, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.45, + 0.464, + 0.459 + ], + "angle": 0, + "content": "Single Shape: A HEXAGONAL HELMET rests on the seat of the bench, near the window." + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.46, + 0.455, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.555, + 0.462, + 0.565 + ], + "angle": 0, + "content": "Single Texture: A smooth paper bag rests on the counter near the coffee machine." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.625 + ], + "angle": 0, + "content": "Figure 8. Visual comparison of InstanceCap and Open-Sora on instance-level attributes. InstanceCap excels in precise instance detail fidelity and instruction-following capabilities, even with complex multi-instance and multi-attribute scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.483, + 0.671 + ], + "angle": 0, + "content": "scriptions while reducing hallucination artifacts compared to MiraData's output." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.325, + 0.696 + ], + "angle": 0, + "content": "4.3. Text-to-Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Qualitative evaluation. Figures 6 and 8 provide visual comparisons of the T2V generation results. It can be observed that the infusion of our InstanceVid dataset into Open-Sora [37] serves to further enhance its video synthesis capabilities across four fundamental dimensions (Action, Color, Shape, and Texture). These four different aspects correspond to information in our instance-aware structured captions such as \"Actions and Motion\", \"Appearance\", etc. For instance, our model accurately generates the \"sprints\" action of the lion in Figure 6, as opposed to Open-Sora [37]. In Figure 8, benefiting from our instance-aware caption, our model generates the accurate \"light brown bag\" instance described in caption, where Open-Sora [37]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.907, + 0.296 + ], + "angle": 0, + "content": "completely loses this instance. These results indicate that our InstanceVid can provide accurate and instance-level guiding information for video generation models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.909, + 0.575 + ], + "angle": 0, + "content": "Quantitative evaluation. We conduct the quantitative evaluation for InstanceCap using the proposed Inseval metrics in Table 2. We can draw the following conclusions: 1) Fine-tuning with InstanceVid consistently improves all metrics over the base model Open-Sora, demonstrating the effectiveness of InstanceCap. In particular, our Detail score ranks first, justifying the capacity of InstanceCap to capture complex instance detail in video. 2) Compared to other video captioning models finetuned based on Open-Sora, InstanceCap shows clear advantages in video generation tasks. 3) Compared to larger models like CogVideoX or Pyramid-Flow, our approach achieves a higher average metric than Pyramid-Flow, and performs comparably to CogVideoX in several specific metrics like 'Single-Color/Shape/Detail' and 'Multiple-Action', but with much fewer parameters. While primarily focused on enhancing text-to-video instance generation, our method achieves competitive results across EvalCrafter [16] (Table 3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.586, + 0.667, + 0.602 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.906, + 0.713 + ], + "angle": 0, + "content": "Effects of human design class hints and camera movement hints. We discuss the impact of incorporating human-designed class hints and camera movement hints on annotation outcomes and provide relevant caption visualizations in Figure 9. These annotations aid MLLMs in focusing more precisely on key elements, resulting in richer and more accurate annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Ablations on different video visual prompts. Comparative results for various video visual prompt methods used in caption generation are shown in Figure 10. As illustrated in (a), weak visual prompts derived from static image techniques, such as red circles, bounding boxes, or selective grayscale manipulation of non-target areas [22, 33], limit MLLMs in distinguishing and describe specific targets in multi-instance scenes, leading to attribute blending and vague annotations across instances. In contrast, our method excels in instance-specific feature extraction, accurately differentiating figures like the coach and players. Figure 10 (b) illustrates strong visual prompts that involve complete oc-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28980" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.11, + 0.173, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.156, + 0.168, + 0.164 + ], + "angle": 0, + "content": "zoom in" + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.176, + 0.186, + 0.185 + ], + "angle": 0, + "content": "w/o hints" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.19, + 0.187, + 0.2 + ], + "angle": 0, + "content": "w/ hints(ours)" + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.095, + 0.816, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.201, + 0.177, + 0.712, + 0.187 + ], + "angle": 0, + "content": "The camera movement in the video is smooth and steady, providing a clear and stable view of the scene." + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.176, + 0.737, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.201, + 0.192, + 0.775, + 0.203 + ], + "angle": 0, + "content": "The camera movement in the video is a steady and gradual zoom-in, focusing on a specific area of the cityscape." + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.103, + 0.858, + 0.114 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.224, + 0.166, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.293, + 0.18, + 0.302 + ], + "angle": 0, + "content": "w/o hints" + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.307, + 0.18, + 0.317 + ], + "angle": 0, + "content": "w/ hints(ours)" + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.211, + 0.816, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.2, + 0.293, + 0.538, + 0.304 + ], + "angle": 0, + "content": "An older man, his hair white, is wearing a blue denim shirt and jeans." + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.293, + 0.562, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.2, + 0.306, + 0.873, + 0.317 + ], + "angle": 0, + "content": "The person in the video is a middle-aged male with white hair. He is wearing a denim shirt and jeans, and he has a watch on his left wrist." + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.215, + 0.858, + 0.226 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.32, + 0.907, + 0.347 + ], + "angle": 0, + "content": "Figure 9. (a) Ablation study on the effect of camera movement hints on the accuracy of MLLM labeling. (b) Impact of human-designed class hints on the details of instance labeling." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.352, + 0.49, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.352, + 0.901, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.57, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Figure 10. (a) Comparison against the weak visual prompt for reconstruction via caption visualization on multi-instance targets. (b) Comparison against color screen backgrounds (red), which may negatively affect MLLM labeling performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.483, + 0.742 + ], + "angle": 0, + "content": "clusion of non-target regions to eliminate MLLM interactions with irrelevant instances. Conventional methods use primary color screens, but this often misguides MLLMs, causing them to incorporate incorrect context in captions. Our designed blur background masking approach, however, preserves visual consistency with natural scenes, enabling MLLMs to generate accurate and contextually relevant annotations with minimal prompting guidance. In Table 1, we present the quantitative results of the ablation study." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.356, + 0.761 + ], + "angle": 0, + "content": "5. Conclusions and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce InstanceCap, the first instance-aware structured caption method for text-to-video generation. We design an Auxiliary Models Cluster (AMC) to convert global video into instances, enhancing instance fidelity. we also propose an improved CoT pipeline with MLLMs to refine dense prompts into structured phrases, achieving concise yet precise instance descriptions compared to the previous video caption models. Additionally, based on InstanceCap, we curated InstanceVid dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.906, + 0.651 + ], + "angle": 0, + "content": "for training and InstanceEnhancer during inference, significantly enhancing T2V models' generation capabilities on instance details and actions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.907, + 0.783 + ], + "angle": 0, + "content": "Limitations. Since the precision of InstanceCap partly depends on object detection methods, requiring fine-tuning of the detection model for domain-specific instances, and its benefits decrease in instance-free scenes. Furthermore, the scale of InstanceVid limits its use as a large-scale pre-training dataset. Moving forward, we plan to apply InstanceCap to a larger video dataset and train more powerful T2V models to amplify its impact." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.791, + 0.693, + 0.807 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This work was supported by Natural Science Foundation of China: No. 62406135 and Natural Science Foundation of Jiangsu Province: BK20241198. Supported by the Gusu Innovation and Entrepreneur leading Talents: No. ZXL2024362. Supported by the AI & AI for Science Project of Nanjing University: No. 14380007." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28981" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] David L. Chen and William B. Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics (ACL-2011), Portland, OR, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, pages 7310-7320, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.296 + ], + "angle": 0, + "content": "[3] Lin Chen, Xin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 1, 3, 5, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.482, + 0.38 + ], + "angle": 0, + "content": "[4] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. arXiv preprint arXiv:2402.19479, 2024. 1,3,5,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.482, + 0.449 + ], + "angle": 0, + "content": "[5] Zhennan Chen, Yajie Li, Haofan Wang, Zhibo Chen, Zhengkai Jiang, Jun Li, Qian Wang, Jian Yang, and Ying Tai. Region-aware text-to-image generation via hard binding and soft refinement. arXiv preprint arXiv:2411.06558, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[6] Xiantao Hu, Ying Tai, Xu Zhao, Chen Zhao, Zhenyu Zhang, Jun Li, Bineng Zhong, and Jian Yang. Exploiting multimodal spatial-temporal patterns for video object tracking, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.482, + 0.549 + ], + "angle": 0, + "content": "[7] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. arXiv preprint arXiv:2307.06350, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.647 + ], + "angle": 0, + "content": "[8] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[9] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1-38, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[10] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[11] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. Miradata: A large-scale video dataset with long durations and structured captions. arXiv preprint arXiv:2407.06358, 2024. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[12] Kuaishou. Kling. https://kling.kuaishou.com, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[13] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[14] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. arXiv preprint arXiv:2410.05677, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[15] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[16] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. arXiv preprint arXiv:2310.11440, 2023.5.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[17] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.411 + ], + "angle": 0, + "content": "[18] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.853, + 0.427 + ], + "angle": 0, + "content": "[19] Pika. Pika 1.0. https://pika.art, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.429, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[20] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dollar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.902, + 0.551 + ], + "angle": 0, + "content": "[21] Runway. Gen-2. https://research.runwayml.com/gen2, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.553, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[22] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11987-11997, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[23] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[24] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. ModelScope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[25] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[26] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. LaVie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[27] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "28982" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[28] Yinwei Wu, Xianpan Zhou, Bing Ma, Xuefeng Su, Kai Ma, and Xinchao Wang. Ifadapter: Instance feature control for grounded text-to-image generation, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[29] Rui Xie, Yinhong Liu, Penghao Zhou, Chen Zhao, Jun Zhou, Kai Zhang, Zhenyu Zhang, Jian Yang, Zhenheng Yang, and Ying Tai. Star: Spatial-temporal augmentation with text-to-video models for real-world video super-resolution. arXiv preprint arXiv:2501.02976, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[30] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[31] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In Forty-first International Conference on Machine Learning, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[32] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 3, 5, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.443 + ], + "angle": 0, + "content": "[33] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. AI Open, 5:30-38, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[34] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. IJCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[35] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[36] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[37] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[38] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-Sora: Democratizing efficient video production for all. https://github.com/hpcaitech/Open-Sora, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[39] Luowei Zhou, Nathan Louis, and Jason J. Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[40] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6748-6758, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "28983" + } + ] +] \ No newline at end of file diff --git a/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/63b6efae-0d11-487a-b293-b1b579ef5966_origin.pdf b/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/63b6efae-0d11-487a-b293-b1b579ef5966_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c350efd075d92618521f99c371d7d07fcd5768f2 --- /dev/null +++ b/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/63b6efae-0d11-487a-b293-b1b579ef5966_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05134bf462ff664c5766b029f3913716814a68ccc2d44706f2416457367ae668 +size 3356228 diff --git a/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/full.md b/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/full.md new file mode 100644 index 0000000000000000000000000000000000000000..13e9ea1e2b95729f6fcaa4d607787fadc9652975 --- /dev/null +++ b/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/full.md @@ -0,0 +1,279 @@ +# InstanceCap: Improving Text-to-Video Generation via Instance-aware Structured Caption + +Tiehan Fan $^{1*}$ Kepan Nan $^{1*}$ Rui Xie $^{1}$ Penghao Zhou $^{2}$ Zhenheng Yang $^{2}$ Chaoyou Fu $^{1}$ Xiang Li $^{3}$ Jian Yang $^{1}$ Ying Tai $^{1\boxtimes}$ $^{1}$ Nanjing University ${}^{2}$ ByteDance ${}^{3}$ Nankai University +https://github.com/NJU-PCALab/InstanceCap + +# Abstract + +Text-to-video generation has evolved rapidly in recent years, delivering remarkable results. Training typically relies on video-caption paired data, which plays a crucial role in enhancing generation performance. However, current video captions often suffer from insufficient details, hallucinations and imprecise motion depiction, affecting the fidelity and consistency of generated videos. In this work, we propose a novel instance-aware structured caption framework, termed InstanceCap, to achieve instance-level and fine-grained video caption for the first time. Based on this scheme, we design an auxiliary models cluster to convert original video into instances to enhance instance fidelity. Video instances are further used to refine dense prompts into structured phrases, achieving concise yet precise descriptions. Furthermore, a 22K InstanceVid dataset is curated for training, and an enhancement pipeline that tailored to InstanceCap structure is proposed for inference. Experimental results demonstrate that our proposed InstanceCap significantly outperform previous models, ensuring high fidelity between captions and videos while reducing hallucinations. + +# 1. Introduction + +Recently, text-to-video (T2V) generation with advanced diffusion transformers (DiT) [2, 10, 12-14, 17, 19, 21, 24, 26, 34, 38] have attracted significant attention for the ability to generate realistic, long-duration videos based on text prompts. Video-caption paired data is typically used in training and plays a crucial role in enhancing generation performance [23, 32]. Current video recaption methods often incorporate multimodal large language models to produce detailed captions, which however usually suffer from hallucinations, leading to inconsistencies between captions + +and video content. Consequently, creating consistent videocaption pairs with accurate details and precise motion depiction for T2V generation remains a significant challenge. + +As shown in Figure 1, current video recaption methods can be broadly categorized into three types: 1) Short captions, such as Panda-70M [4], lack sufficient coverage of video content, leading to low fidelity. 2) Dense captions, like ShareGPT4Video [3], enrich textual content but suffer from hallucination issues, often generating meaningless or inaccurate video content. 3) Coarse-level structured captions, exemplified by MiraData [11], improve video quality but provide coarse-level details. Moreover, the redundancy introduced by MLLM across structures diminishes its overall effectiveness. To this end, achieving accurate captions remains two crucial challenges: 1) High fidelity between caption and video: Retain as much of the original video's objects, textures, and motion information as possible. 2) Accurate content in caption: Enable MLLM model to generate precise content, minimizing hallucinations and repetition. + +To address the challenges, we propose a novel instance-aware structured caption framework, termed InstanceCap, to achieve instance-level and fine-grained video caption for the first time. Our structure is specifically designed to incorporate instances, background, and camera movement. For each instance, we specify class, appearance, actions, motion, and position. To enhance the fidelity and accuracy of video captions, we focus on two key aspects: 1) From Global Video to Local Instances: For each instance, we propose an auxiliary models cluster (AMC) to isolate it from the original video and obtain the corresponding position and category information. This operation minimizes interference from unrelated regions while retaining as much of the original video's information as possible. 2) From Dense prompt to Structured Phrases: We leverage multimodal large language models (MLLMs) in an improved Chain-of-Thought (CoT) process to obtain concise yet accurate descriptions of textures, camera movement, actions and motion for each instance. This reduces the probability of hallucinations and irrelevant content produced by the lan + +![](images/3a5f3c5616adef5084ff90e77d498ddf70b3e7df04b9f716f92f6dd2add08c58.jpg) +[InstanceCap](Ours) +Wrong Ambiguity Detailed +Figure 1. Top: Comparison of the reconstruction via-recaption results between InstanceCap and state-of-the-art captioning methods for annotating the ground truth video. InstanceCap produces results that more closely resemble the original video, showing greater detail fidelity (highlighted by the red circle). Bottom: The corresponding captions generated by InstanceCap and others. Red denotes incorrect captions, blue represents ambiguous captions, and green indicates detailed and accurate descriptions of video. Specific visual hints are marked as A, B, and C for clarity. All videos are generated using the same video generation product, Hailuo AI1, which has robust prompt-following capabilities, clearly highlighting the effectiveness of InstanceCap. + +"Global Description": A woman with blonde hair, wearing a dark blue shirt with white stripes, is speaking and gesturing with her hands. + +Structured Description": { + +Main Obj + +No.0-1 Class:"person", + +"Appearance." The person in the video is a woman with blonde hair, wearing glasses, her face is to the right of the frame. She is dressed in a dark blue blouse with white stripes. The woman appears to be middle-aged and is of Caucasian race. She is wearing a necklace $①$ and a black dress. She has a thick skirt and a facial expression throughout the video, indicating that she is engaged in a conversation or presentation." + +"Actions and Motion": "Her gestures are expressive, with her hands moving fluidly to emphasize her points and she is holding a pen. The way she speaks suggests she is passionate about the topic at hand, her body language open and inviting." + +"Position": "person stays at bottom-middle." + +"Background Detail": "The environment in the video is a stage or presentation area with a grand and elegant atmosphere. The backdrop consists of a red curtain, which adds a touch of drama and sophistication to the setting. Text square 'ROVO' can be seen in the lower right corner. The lighting is bright and evenly distributed, illuminating the stage and creating a clear and focused environment for the speaker. The overall ambiance is professional and formal, indicative of a well-organized and prestigious event." + +'Camera Movement': The camera is static, maintaining a short distance shot throughout the video. The shooting angle remains consistent, capturing the speaker from the waist up.' +1 + +# [PANDA-70M](CVPR 2024) + +A woman wearing glasses and a blue tie is sitting on a red couch. + +# [ShareGPT4Video](NeuIPS 2024) + +The video features an individual standing in front of a red background, engaging in what appears to be a presentation or discussion. Throughout the video, the person is seen making various gestures with their hands, indicating a dynamic and expressive delivery. Initially, the person's hands are raised, palms facing upwards, suggesting an explanatory or emphatic gesture. As the video progresses, the individual's gestures evolve; the hands are lowered and then raised again, with the fingers pointing upwards, indicating a shift in the emphasis or point of discussion. Subsequently, the hands are brought closer together, with the fingers interlocked, suggesting a moment of emphasis or a transition in the presentation. Finally, the hands are raised again, with the fingers spread apart, indicating a continuation of the expressive gestures. Throughout the video, the individual's attire remains consistent, featuring a dark-colored top with a lighter-colored collar, and the background remains a solid red, providing a stark contrast that keeps the focus on the speaker. The camera angle and framing stay constant, with no discernible movement, maintaining a steady focus on the individual and their gestures. The lighting is even, ensuring that the subject is clearly visible against the red backdrop. There are no changes in the environment, the individual's attire, or the camera's perspective, indicating a continuous and uninterrupted presentation or discussion. + +# [MiraData][NeurIPS 2024] + +In this video, a person with blond hair is engaging in an animated conversation while seated. They are wearing glasses and a striped shirt and express themselves through a series of dynamic hand gestures and head movements. The individual appears to be deeply focused on the dialogue, occasionally looking down or away, possibly to emphasize their points or formulate thoughts. The person's gestures escalate throughout the video, suggesting an increasing intensity or passion about the topic. The main subject, identifiable by their blond hair and glasses, transitions from initially looking down to making various expressive hand gestures and head movements. Their body language involves considerable arm movement, indicating engagement and enthusiasm in the conversation. The sequence begins with the individual in a somewhat reserved posture, escalating to more expansive and assertive gestures, before returning to a composed state. The background of the video is a studio setting with prominent red and white color blocks, suggesting a professional and possibly formal environment. Part of a sign or text that reads "RM" is visible, indicating a possible connection to a broadcast or presentation context. The exact location isn't specified, but the controlled indoor setting implies a purposeful and staged production. The view shot remains consistent, focused on the upper body of the main subject, ensuring their gestures and facial expressions are clearly visible. The camera maintains a mid-shot with minimal movement, emphasizing the individual's actions and expressions. There are no noticeable changes in the shooting angle, keeping the subject centered throughout the frames to convey their animated discourse. The video employs a straightforward, television broadcast style, emphasizing clear and focused framing to capture the subject's expressive communication in a professional studio environment. + +guage model compared to traditional caption methods that directly describe video content in a complex, dense caption. + +To validate the effectiveness of InstanceCap in T2V generation, we constructed a high-definition video dataset comprising 22K samples to create a training dataset with our instance-aware structured captions, named InstanceVid. At the inference stage, we also implemented a prompt enhancement pipeline tailored to our structured captioning method, enabling the generation of concise captions that better align with user needs. Our InstanceCap integrates seamlessly with existing diffusion models. Experimental results demonstrate that after finetuned with our InstanceVid, the T2V model exhibits better ability with prompt following on details and motion actions. In summary, our main contributions are as follows: + +- We propose InstanceCap, the first instance-aware structured caption method for text-to-video generation. We also constructed a 22K InstanceVid dataset during training, and developed an enhancement pipeline that tailored to InstanceCap structure during inference. +- We design the AMC paradigm to convert global video + +into instances, enhancing instance fidelity. Additionally, we propose an improved CoT pipeline with MLLMs to refine dense prompts into structured phrases, achieving concise yet precise descriptions. + +- Extensive experiments on video reconstruction demonstrate that our InstanceCap significantly enhances the fidelity between captions and videos. T2V models fine-tuned on our InstanceVid further achieve more precise generation on instance details and motion actions. + +# 2. Related Works + +Video recaptioning. Advancements in text-to-video generation demand high-quality video-text datasets to build robust foundational video models for visual-language alignment. Current video recaption methods fall into two main categories: manual annotation [1, 25, 39] and end-to-end recaption using multimodal large language models (MLLM). Although manual annotation provides higher accuracy, scaling datasets to meet the needs of high-quality video generation models remains a substantial challenge. Recent + +advances in MLLM have demonstrated impressive capabilities in video understanding and description generation. Panda [4] and InternVid [27] are with short captions, offering computational efficiency but frequently omitting crucial content and exhibit low video fidelity. OpenVid-1M [18], Vript [30] and ShareGPT4Video [3] are with dense captions, which provide richer content but face challenges: Hallucinations due to complexity, inclusion of redundant information, and text encoder truncation caused by excessively long text. MiraData [11] achieves coarse-level structured captions that attempts to mediate these issues but struggle with fine detail and redundancy across structures. Different from the existing video recaption methods, InstanceCap is the first instance-aware structured caption approach for text-to-video generation, ensuring high fidelity between caption and video while reducing hallucinations and repetition. + +Text-to-video generation. Despite numerous high-quality video generation models [2, 10, 12, 13, 19, 32] perform well with simple directives, they often falter with complex prompts requiring precise instance-level details or intricate camera movements. Analyzing current video-text datasets suggests these limitations may stem from suboptimal training data quality. Traditional recaption methods have not sufficiently captured instance-specific detail granularity or provided comprehensive descriptions of camera movements. To enhance instance-level details and motion consistency, we construct a 22K InstanceVid dataset for training/finetuning, and develop an enhancement pipeline InstanceEnhancer that tailored to the proposed instance-aware structure during inference. + +# 3. Method + +In Section 3.1, we first present the InstanceCap pipeline, as shown in Figure 2. Based on this pipeline, we recapture the carefully curated dataset InstanceVid in Section 3.2, enhancing T2V models' instance generation. Additionally, in Section 3.3, we introduce InstanceEnhancer to convert short prompts into our proposed instance-aware structured caption format during inference. + +# 3.1. InstanceCap + +Video preprocessing with auxiliary model cluster. For continuous video processing, we implemented uniform sampling using $\text{decord}^2$ , following the methodology established in LLaVA-Video [36]. This approach enables us to extract essential temporal metadata, including duration, frame count, and timestamps, allowing MLLMs to better interpret temporal sequences in recaptioning tasks. Additionally, to enhance MLLMs' capabilities through structured guidance, we incorporate several auxiliary models to + +achieve accurate object detection, video instance segmentation and camera motion prediction, providing precise prior information to the MLLMs. + +Global description, background detail and camera movement. When describing video content, a high-quality global description should capture primary elements, environmental context, camera movements, angles, and tonal qualities. MLLMs excel in generating high-level video summaries using Chain-of-Thought methodology. By employing carefully designed prompts with CoT, we can guide MLLMs to produce detailed background descriptions while minimizing references to foreground elements. + +However, MLLMs' limitations in processing discrete frames rather than continuous video segments make it challenging to distinguish camera motion from instance action. To address this, we achieve camera annotations from OpenSora [37] for basic movements (e.g., zoom, rotation) and rely on MLLMs to capture subtle motion attributes (e.g., intensity, speed). The integration of camera movement indicators with MLLM capabilities provides comprehensive annotations, as illustrated in Figure 9 (a). + +Structured Description on Instances. In this subsection, we introduce the details to achieve our instance-aware structured description. To address MLLMs' limitations in instance annotation and the suboptimal results of directly adapting weak visual prompts from images to videos [22, 33], we make full use of the auxiliary model cluster, including initial object detection [6, 40], video instance segmentation with SAM2 [20], and blur non-instance regions to achieve blur background, resulting in better recaptioning outcomes compared to alternative visual prompt methods in video, as shown in Figure 10. To this end, we decompose the global videos into local instances. + +Next, we describe how to achieve detailed and accurate description of each instance (Figure 3). To maintain instance-level precision, we deliberately constrain the information accessible to MLLMs during instance annotation. Crucially, the global video remains invisible due to our designed blurred backgrounds, preventing MLLMs from confusing information across multiple instances. This approach allows InstanceCap to focus on local instances identified through auxiliary model cluster. Furthermore, to avoid the potential limitation of MLLMs seeing only isolated instances, which could lead to overlooking inter-instance interactions and subsequent misinterpretations, we incorporate the global description mentioned in previous subsection. Specifically, we inject the global description into the instance-annotation MLLMs. This strategic mitigates potential biases in instance action descriptions while maintaining instance-specific accuracy. + +To enhance the capability of InstanceCap in capturing instance-level details, we introduce novel insights into the improved CoT process. Our analysis of current MLLM- + +![](images/35e2ebb0d3939a7c4875f0e451dc28ac4d52f26a3aee144daa15278081862f8b.jpg) + +![](images/855d4717d08bc26500404d5bec9a3fccd0368c2d5cd5d32deea05cacc31581f5.jpg) +Figure 2. Overview of InstanceCap pipeline. Details of "from dense prompts to structured phrases" design are shown in Figure 3. +Figure 3. Details on "from dense prompts to structured phrases" design. We propose an improved CoT pipeline with carefully designed information interactions (red arrow), which facilitates MLLMs to accurately capture instances with precise descriptions on attributes. + +based video recaption methods shows that simple prompts, like “Please provide a detailed description of this video” or Chain-of-Thought prompts “Let’s think step by step... First, please note... Finally, summarize the video content...” fail to capture precise instance details. Additional experiments reveal that MLLMs can effectively annotate these details when given fine-grained prompts, such as “Please note if the characters have any accessories” or “Please observe whether there are spots on the bananas”. To enhance the details of instances, we develop Human-designed Class Hints, crafting specific prompts for about 80 detectable categories using our auxiliary models cluster. Specifically, we present the “person” class prompt here: “Please focus primarily on the person’s facial expressions, attire, age, gender, and race in the video and give a detailed description. Please mention if there are any necklaces, watches, hat or other decoration; otherwise, there’s no need to bring them up.” Besides, we also developed a curated Posi + +tive/Negative Lexicon to guide MLLMs in generating more aesthetically refined captions. More details can be found in our supplementary material. + +# 3.2. InstanceVid + +InstanceVid is curated via refining a subset from the high-aesthetic, high-consistency videos from OpenVid-1M [18]. To showcase our method's high-fidelity labeling of instance details and motion, we selected video samples that included at least one instance exhibiting high motion intensity during dataset filtering. More statistical information of InstanceVidis provided in the supplementary material. + +# 3.3. InstanceEnhancer + +When the caption distribution of training data differs from that of inference text, it may result in poor instruction-following performance or even problematic outputs. This issue is evident in T2V generation, particularly when long + +![](images/627a420102a76c64088e89fb60a929e73622418d1f94b825207ac3ced85ed05a.jpg) +Figure 4. High-level overview of InstanceEnhancer, illustrating the data flow and the partitioning of stages. For a detailed implementation, refer to the supplemental materials, which provide an in-depth description of the enhancer pipeline design and the interdependencies between the stages. + +captions are used for training but short captions for inference, leading to subpar results. Since users typically prefer short captions, it is essential to enhance short caption effectively to better align with our proposed instance-aware structured caption during training. + +As shown in Figure 4, we introduce a tuning-free approach called InstanceEnhancer that achieves this by strictly limiting the generated formats to match the caption corresponding to the training input we used. Our method differs from existing tuning-free caption enhancement approaches, such as those presented in RPG and RAG [5, 31]. Instead of directly enhancing short captions, which we found can introduce inconsistencies between multiple instances' actions and their environmental context in video generation, we employ a two-stage enhancement strategy. In Stage A, short prompts are expanded into detailed long prompts. Stage B(I)&(II) uses both expanded and original captions to segment and enhance specific instances, preserving contextual coherence while ensuring precise instance identification. The quantitative evaluation results are presented in Table 2. Due to the space limitation, more details of our enhancement pipeline can be found in the supplementary material. + +# 4. Experiments + +# 4.1. Experimental Setup + +Video reconstruction with recaptions. To comprehensively evaluate InstanceCap, we conducted a series of experiments, benchmarking against state-of-the-art methods including Panda-70M [4], ShareGPT4Video [3], and Mi-raData [11]. To this end, we carefully selected 500 video clips from OpenVid-1M [18]. For each video, we generated one caption using various caption models, which were then input into the advanced T2V model CogvideoX-5b [32] for video generation. We calculated the differences between the generated videos and the ground truth videos to evaluate each caption model's performance, where smaller visual differences indicate more accurate captions and higher fidelity. + +We introduced several metrics to evaluate the video reconstruction performance: 1) $3\mathrm{DVAE}_{score}$ : Using 3DVAE from CogVideoX [29, 32] as the backbone, we extract hidden-space representations from both the original videos + +![](images/9c5c8a17c65e52a276d1bc18f7f784cab315da6ea606313f8a65912570e59848.jpg) +Figure 5. Comparison on reconstruction via-recapiction between InstanceCap and MiraData. Corresponding 3DVAE scores are also indicated. Similar semantics shared between InstanceCap and GT are indicated by red circles and lines. + +and their recaption-reconstructed counterparts. These representations quantify the perceptual distance between them. 2) VQAScore: In order to obtain text-video consistency evaluation results that match human preferences, we use VQAScore [15] for evaluation. Specifically, we sample the video uniformly, and then calculate the VQA value for this set of sampled frames and their corresponding caption. This score is positively correlated with caption quality, providing an objective measure of text-video alignment. 3) Human Evaluation: We conducted a user study with a panel of evaluators to assess caption quality across two aspects: Instance Detail (ID) and Hallucination Scores (HS). + +T2V generation. To thoroughly evaluate the T2V generation performance of our InstanceCap, we utilize the InstanceVid dataset to finetune the state-of-the-art DiT-based T2V generation model Open-Sora [37]. In our evaluation, we compare with Open-Sora, CogVideoX-5b [32], Pyramid-Flow [10], and Open-Sora-Plan [13]. To enable fine-grained, instance-level assessment, we construct a highly challenging evaluation benchmark called Inseval, inspired by recent advancements in T2I and T2V evaluation [7, 8, 15, 16, 28]. Specifically, we curate a diverse evaluation dataset of over 200 carefully crafted instance-level prompt-answer pairs, covering both single-object and multi-object scenarios systematically across five key dimensions: Action, Color, Shape, Texture, and Detail. Motivated by the previous evaluation benchmarks [8], we implement a CoT reasoning framework for generating structured QA responses to ensure objective and consistent evaluation, allowing us to derive instance-level evaluation scores that align closely with human perception and preferences. This approach provides a more nuanced and reliable assessment of instance-level generation quality. + +![](images/25ded0840f93a18ca00a23ebbe851b39708636a38609b95ba0162698cad33ff2.jpg) +Single Action: A swift lion sprints across the plain. + +![](images/42613abec6fcd1b50fc62ddf9a7cba2ed3d632bcc2b462bed4b62fc3496e9e91.jpg) +Multiple Action: A fish swims and a turtle moves on the seabed. +Figure 6. Visual comparison of InstanceCap and Opensa on Single and Multiple Action Score. In terms of the dynamic degree of video generation, we show better consistency and enhanced multi-instance dynamic generation effect. + +
Captioning Methods3DVAEscore↓VQAScore↑Avg. Length
Panda-70M [4]146.630.79913 words
ShareGPT4Video [3]147.000.866174 words
LLaVA-Video-72B [36]147.630.904109 words
MiraData(GPT-4V) [11]144.750.905197 words
InstanceCap(Ours)143.000.916165 words
Ablation Study
InstanceCap w/o camera hints144.380.906166 words
InstanceCap w/o human hints143.870.908165 words
InstanceCap w/ red circle144.120.906162 words
InstanceCap w/ blue background144.380.903165 words
+ +Table 1. Quantitative comparisons on reconstruction-via-recaption results. The best results are marked in **bold**, and the second-best are **underscored**. As a reference, CogVideoX-5b accepts 226 text tokens, with any excess being truncated. Meanwhile, the results of the ablation study are shown in the lower part of the table. + +
T2V ModelSingle†Multiple†Average†
ActionColorShapeTextureDetailActionColorTexture
CogVideoX-5B [32]64%60%44%60%20%8%48%40%43.00%
Pyramid-Flow-2B [10]44%68%32%32%7%4%24%16%28.38%
Open-Sora Plan v1.3-2.7B [13]64%44%36%32%27%20%32%12%33.38%
Open-Sora v1.2-1.1B [37]40%56%36%40%13%12%16%16%28.63%
+ Ours w/o InstanceEnhancer56%60%40%48%27%16%32%24%37.88%
+ Ours w/ InstanceEnhancer48%60%56%64%20%24%48%40%45.00%
+ Panda-capitalizer [4]40%48%28%40%20%8%20%12%27.00%
+ ShareGPT4Video [3]40%44%32%24%13%16%8%20%24.63%
+ MiraData [11]44%48%40%44%13%12%28%16%30.63%
+ LLaVA [18]52%52%28%28%20%12%28%16%29.50%
+ +Table 2. Quantitative comparison between InstanceCap and SOTA video captioning models, all based on the popular T2V model Open-Sora. Additionally, we also compare three powerful T2V models, including CogVideoX-5B, Pyramid-Flow, and Open-Sora Plan. The best results of video captioning methods and Open-Sora are marked in bold, and the second-best are underscored. + +# 4.2. Comparison with SOTA Caption Models + +Qualitative evaluation. CogVideoX-5b [32] is a latent text-to-video generation model known for its capability to generate realistic, long-duration videos based on text prompts. With the integration of our InstanceCap into CogVideoX-5b, as substantiated by Figure 5, the model exhibits a notable enhancement in the video reconstruction capacity. This demonstrates that our instance-aware structured captions retain more of the original video's information, leading to higher fidelity. For instance, our InstanceCap can retain information such as "glasses", + +![](images/b5bb88d6c323125e35aaac5276fd94defd246c28fe0d2f01d5154a3fd6f37c9e.jpg) +Figure 7. User study on instance detail and hallucination scores. Ours instance-aware structured caption shows clear advantages compared to the coarse-structured MiraData [11]. + +![](images/1156a94ae26aefb36404a47728809950623b8e6e6c6f13753d75ff969ccc1992.jpg) + +"grey sweater", and "relative position of two people", whereas MiraData [11] almost completely loses these important details. A similar conclusion can be drawn from Figure 1. These results underscore the significant improvements achieved by our InstanceCap, resulting in high-quality reconstruction characterized by rich detail and high fidelity between our captions and the original videos. + +Quantitative evaluation. Table 1 presents quality comparisons between our InstanceCap and other caption methods across two metrics. Based on the results, we make the following observations: 1) Our method delivers comparable or superior quality to the four baselines, demonstrating its ability to enhance fidelity between videos and captions. This strong alignment with human perceptual judgments and preferences is evident in Figures 1 and 5. 2) Our method consistently excels across all metrics for captions under 200 words, an acceptable length for most T2V models, illustrating its generalizability. Figure 7 compares MiraData's captions with our instance-aware structured ones. We randomly selected videos from an open-domain dataset, and a panel of evaluators assessed caption quality using standardized criteria. Results indicate that our captions offer significantly richer and more accurate de + +
T2V ModelQuality Metrics↑Temporal Metrics↑Motion MetricsTV Align Metrics
Action Score↑Motion AC Score↑Flow Score↑CLIP Score↑Blip-Blue↑SD Score↑
Open-Sora v1.2-1.1B53.2258.4867.87440.6520.5522.8468.19
+ Panda-captioner54.9857.2856.85440.6719.8119.5867.91
+ ShareGPT4Video52.7560.6768.55440.6620.2823.1067.95
+ LLaVA54.7960.1769.96440.5720.5323.1868.18
+ MiraData54.7160.2665.00440.3720.5322.9868.23
+ InstanceCap(Ours)54.7156.5772.04460.6820.5423.1968.88
+ +Table 3. Results on Evalcrafter. Whereas dataset quality determines aesthetic metrics, InstanceCap significantly enhances dynamism and text-video alignment, with acceptable trade-offs in temporal consistency. Consistent with Inseval's performance. + +![](images/2c8be9b24fc73d5653bdf8d9a39dee3f156d0ff2a8c92dc057dbf50875754167.jpg) +Multiple Color: A sapphire blue keychain hangs from a light brown bag. + +![](images/d151e9f76d61556484c759f8370b745af50fba03b537e9287a875b244a8750c2.jpg) +Single Shape: A HEXAGONAL HELMET rests on the seat of the bench, near the window. + +![](images/d6c47b03e838fa780652532d91699f187ba8f1a15d5b9b964380a005359704f0.jpg) +Single Texture: A smooth paper bag rests on the counter near the coffee machine. +Figure 8. Visual comparison of InstanceCap and Open-Sora on instance-level attributes. InstanceCap excels in precise instance detail fidelity and instruction-following capabilities, even with complex multi-instance and multi-attribute scenarios. + +scriptions while reducing hallucination artifacts compared to MiraData's output. + +# 4.3. Text-to-Video Generation + +Qualitative evaluation. Figures 6 and 8 provide visual comparisons of the T2V generation results. It can be observed that the infusion of our InstanceVid dataset into Open-Sora [37] serves to further enhance its video synthesis capabilities across four fundamental dimensions (Action, Color, Shape, and Texture). These four different aspects correspond to information in our instance-aware structured captions such as "Actions and Motion", "Appearance", etc. For instance, our model accurately generates the "sprints" action of the lion in Figure 6, as opposed to Open-Sora [37]. In Figure 8, benefiting from our instance-aware caption, our model generates the accurate "light brown bag" instance described in caption, where Open-Sora [37] + +completely loses this instance. These results indicate that our InstanceVid can provide accurate and instance-level guiding information for video generation models. + +Quantitative evaluation. We conduct the quantitative evaluation for InstanceCap using the proposed Inseval metrics in Table 2. We can draw the following conclusions: 1) Fine-tuning with InstanceVid consistently improves all metrics over the base model Open-Sora, demonstrating the effectiveness of InstanceCap. In particular, our Detail score ranks first, justifying the capacity of InstanceCap to capture complex instance detail in video. 2) Compared to other video captioning models finetuned based on Open-Sora, InstanceCap shows clear advantages in video generation tasks. 3) Compared to larger models like CogVideoX or Pyramid-Flow, our approach achieves a higher average metric than Pyramid-Flow, and performs comparably to CogVideoX in several specific metrics like 'Single-Color/Shape/Detail' and 'Multiple-Action', but with much fewer parameters. While primarily focused on enhancing text-to-video instance generation, our method achieves competitive results across EvalCrafter [16] (Table 3). + +# 4.4. Ablation Study + +Effects of human design class hints and camera movement hints. We discuss the impact of incorporating human-designed class hints and camera movement hints on annotation outcomes and provide relevant caption visualizations in Figure 9. These annotations aid MLLMs in focusing more precisely on key elements, resulting in richer and more accurate annotations. + +Ablations on different video visual prompts. Comparative results for various video visual prompt methods used in caption generation are shown in Figure 10. As illustrated in (a), weak visual prompts derived from static image techniques, such as red circles, bounding boxes, or selective grayscale manipulation of non-target areas [22, 33], limit MLLMs in distinguishing and describe specific targets in multi-instance scenes, leading to attribute blending and vague annotations across instances. In contrast, our method excels in instance-specific feature extraction, accurately differentiating figures like the coach and players. Figure 10 (b) illustrates strong visual prompts that involve complete oc- + +![](images/20f4ae0b14ce1faef2e636b99076c61ad9dd9f2bab56ced8d0cd0a04c3c9139d.jpg) +zoom in + +w/o hints +(a) +![](images/04eae64f693c8f9f0af39ad090606112daf0355965c6e6ea5c0eeb3de511b485.jpg) +The camera movement in the video is smooth and steady, providing a clear and stable view of the scene. + +![](images/ab579a622954ef5de8953e60bf25d0e154f1cbe575c0672ba44ad2e61d2f9939.jpg) +The camera movement in the video is a steady and gradual zoom-in, focusing on a specific area of the cityscape. + +![](images/8c2951ba09b70e18b1442d1a4258c8a785f21f53c0dc756faae08077646b1097.jpg) +w/o hints + +w/ hints(ours) +(b) +![](images/d2211f46f60fea9f26f1dd992a7649436013730f91fac7c12a16fb5e7118262e.jpg) +An older man, his hair white, is wearing a blue denim shirt and jeans. + +![](images/596a0a559e3b180bfb8fd320b6873e5a5a39ccc4d85240d3111a1f9c35944d6f.jpg) +The person in the video is a middle-aged male with white hair. He is wearing a denim shirt and jeans, and he has a watch on his left wrist. + +![](images/a0bcf6c8a3d6ad6290fe4f44b91ed65c8a16982825923393e8e2b41e4e4cf062.jpg) +w/ hints(ours) +Figure 10. (a) Comparison against the weak visual prompt for reconstruction via caption visualization on multi-instance targets. (b) Comparison against color screen backgrounds (red), which may negatively affect MLLM labeling performance. + +![](images/15f48e79a35e34bba57712e0a20d6a721176110edf77ebb22a8a83333359bfed.jpg) +Figure 9. (a) Ablation study on the effect of camera movement hints on the accuracy of MLLM labeling. (b) Impact of human-designed class hints on the details of instance labeling. + +clusion of non-target regions to eliminate MLLM interactions with irrelevant instances. Conventional methods use primary color screens, but this often misguides MLLMs, causing them to incorporate incorrect context in captions. Our designed blur background masking approach, however, preserves visual consistency with natural scenes, enabling MLLMs to generate accurate and contextually relevant annotations with minimal prompting guidance. In Table 1, we present the quantitative results of the ablation study. + +# 5. Conclusions and Limitations + +In this paper, we introduce InstanceCap, the first instance-aware structured caption method for text-to-video generation. We design an Auxiliary Models Cluster (AMC) to convert global video into instances, enhancing instance fidelity. we also propose an improved CoT pipeline with MLLMs to refine dense prompts into structured phrases, achieving concise yet precise instance descriptions compared to the previous video caption models. Additionally, based on InstanceCap, we curated InstanceVid dataset + +for training and InstanceEnhancer during inference, significantly enhancing T2V models' generation capabilities on instance details and actions. + +Limitations. Since the precision of InstanceCap partly depends on object detection methods, requiring fine-tuning of the detection model for domain-specific instances, and its benefits decrease in instance-free scenes. Furthermore, the scale of InstanceVid limits its use as a large-scale pre-training dataset. Moving forward, we plan to apply InstanceCap to a larger video dataset and train more powerful T2V models to amplify its impact. + +# 6. Acknowledgments + +This work was supported by Natural Science Foundation of China: No. 62406135 and Natural Science Foundation of Jiangsu Province: BK20241198. Supported by the Gusu Innovation and Entrepreneur leading Talents: No. ZXL2024362. Supported by the AI & AI for Science Project of Nanjing University: No. 14380007. + +# References + +[1] David L. Chen and William B. Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics (ACL-2011), Portland, OR, 2011. 2 +[2] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, pages 7310-7320, 2024. 1, 3 +[3] Lin Chen, Xin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 1, 3, 5, 6, 2 +[4] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. arXiv preprint arXiv:2402.19479, 2024. 1,3,5,6 +[5] Zhennan Chen, Yajie Li, Haofan Wang, Zhibo Chen, Zhengkai Jiang, Jun Li, Qian Wang, Jian Yang, and Ying Tai. Region-aware text-to-image generation via hard binding and soft refinement. arXiv preprint arXiv:2411.06558, 2024. 5 +[6] Xiantao Hu, Ying Tai, Xu Zhao, Chen Zhao, Zhenyu Zhang, Jun Li, Bineng Zhong, and Jian Yang. Exploiting multimodal spatial-temporal patterns for video object tracking, 2024. 3 +[7] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. arXiv preprint arXiv:2307.06350, 2023. 5 +[8] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5 +[9] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1-38, 2023. 3 +[10] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 1, 3, 5, 6 +[11] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. Miradata: A large-scale video dataset with long durations and structured captions. arXiv preprint arXiv:2407.06358, 2024. 1, 3, 5, 6 +[12] Kuaishou. Kling. https://kling.kuaishou.com, 2024. 1, 3 +[13] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 3, 5, 6 + +[14] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. arXiv preprint arXiv:2410.05677, 2024. 1 +[15] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation, 2024. 5 +[16] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. arXiv preprint arXiv:2310.11440, 2023.5.7 +[17] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 1 +[18] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 3, 4, 5, 6 +[19] Pika. Pika 1.0. https://pika.art, 2023. 1, 3 +[20] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dollar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3 +[21] Runway. Gen-2. https://research.runwayml.com/gen2, 2023.1 +[22] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11987-11997, 2023. 3, 7 +[23] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 1 +[24] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. ModelScope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1 +[25] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research, 2020. 2 +[26] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. LaVie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 1 +[27] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations, 2023. 3 + +[28] Yinwei Wu, Xianpan Zhou, Bing Ma, Xuefeng Su, Kai Ma, and Xinchao Wang. Ifadapter: Instance feature control for grounded text-to-image generation, 2024. 5 +[29] Rui Xie, Yinhong Liu, Penghao Zhou, Chen Zhao, Jun Zhou, Kai Zhang, Zhenyu Zhang, Jian Yang, Zhenheng Yang, and Ying Tai. Star: Spatial-temporal augmentation with text-to-video models for real-world video super-resolution. arXiv preprint arXiv:2501.02976, 2025. 5 +[30] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words, 2024. 3 +[31] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In Forty-first International Conference on Machine Learning, 2024. 5 +[32] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 3, 5, 6, 2 +[33] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. AI Open, 5:30-38, 2024. 3, 7 +[34] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. IJCV, 2024. 1 +[35] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 2 +[36] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data, 2024. 3, 6 +[37] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 3, 5, 6, 7 +[38] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-Sora: Democratizing efficient video production for all. https://github.com/hpcaitech/Open-Sora, 2024.1 +[39] Luowei Zhou, Nathan Louis, and Jason J. Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction, 2018. 2 +[40] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6748-6758, 2023. 3 \ No newline at end of file diff --git a/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/images.zip b/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..486ea05ef0bc72e5c6b2ababcc6b60a5dcfcfa7e --- /dev/null +++ b/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08a051fd80bb249d2b0376f6ac65b2ab03dc0a90fadc354f249aeddf3f6a07f3 +size 873711 diff --git a/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/layout.json b/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bc034f1a288bf54b86bd1d57917d2f08a4660525 --- /dev/null +++ b/2025/InstanceCap_ Improving Text-to-Video Generation via Instance-aware Structured Caption/layout.json @@ -0,0 +1,6922 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 88, + 102, + 523, + 140 + ], + "type": "text", + "content": "InstanceCap: Improving Text-to-Video Generation via Instance-aware Structured Caption" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "spans": [ + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "text", + "content": "Tiehan Fan" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "text", + "content": " Kepan Nan" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "text", + "content": " Rui Xie" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "text", + "content": " Penghao Zhou" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "text", + "content": " Zhenheng Yang" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "text", + "content": " Chaoyou Fu" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "text", + "content": " Xiang Li" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "text", + "content": " Jian Yang" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "text", + "content": " Ying Tai" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "text", + "content": " Nanjing University " + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "text", + "content": " ByteDance " + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 123, + 160, + 487, + 219 + ], + "type": "text", + "content": " Nankai University \nhttps://github.com/NJU-PCALab/InstanceCap" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 256, + 200, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 256, + 200, + 269 + ], + "spans": [ + { + "bbox": [ + 152, + 256, + 200, + 269 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 281, + 297, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 297, + 534 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 297, + 534 + ], + "type": "text", + "content": "Text-to-video generation has evolved rapidly in recent years, delivering remarkable results. Training typically relies on video-caption paired data, which plays a crucial role in enhancing generation performance. However, current video captions often suffer from insufficient details, hallucinations and imprecise motion depiction, affecting the fidelity and consistency of generated videos. In this work, we propose a novel instance-aware structured caption framework, termed InstanceCap, to achieve instance-level and fine-grained video caption for the first time. Based on this scheme, we design an auxiliary models cluster to convert original video into instances to enhance instance fidelity. Video instances are further used to refine dense prompts into structured phrases, achieving concise yet precise descriptions. Furthermore, a 22K InstanceVid dataset is curated for training, and an enhancement pipeline that tailored to InstanceCap structure is proposed for inference. Experimental results demonstrate that our proposed InstanceCap significantly outperform previous models, ensuring high fidelity between captions and videos while reducing hallucinations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 555, + 135, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 135, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 135, + 568 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 575, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 296, + 696 + ], + "type": "text", + "content": "Recently, text-to-video (T2V) generation with advanced diffusion transformers (DiT) [2, 10, 12-14, 17, 19, 21, 24, 26, 34, 38] have attracted significant attention for the ability to generate realistic, long-duration videos based on text prompts. Video-caption paired data is typically used in training and plays a crucial role in enhancing generation performance [23, 32]. Current video recaption methods often incorporate multimodal large language models to produce detailed captions, which however usually suffer from hallucinations, leading to inconsistencies between captions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 246, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 281 + ], + "type": "text", + "content": "and video content. Consequently, creating consistent videocaption pairs with accurate details and precise motion depiction for T2V generation remains a significant challenge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 556, + 474 + ], + "type": "text", + "content": "As shown in Figure 1, current video recaption methods can be broadly categorized into three types: 1) Short captions, such as Panda-70M [4], lack sufficient coverage of video content, leading to low fidelity. 2) Dense captions, like ShareGPT4Video [3], enrich textual content but suffer from hallucination issues, often generating meaningless or inaccurate video content. 3) Coarse-level structured captions, exemplified by MiraData [11], improve video quality but provide coarse-level details. Moreover, the redundancy introduced by MLLM across structures diminishes its overall effectiveness. To this end, achieving accurate captions remains two crucial challenges: 1) High fidelity between caption and video: Retain as much of the original video's objects, textures, and motion information as possible. 2) Accurate content in caption: Enable MLLM model to generate precise content, minimizing hallucinations and repetition." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "text", + "content": "To address the challenges, we propose a novel instance-aware structured caption framework, termed InstanceCap, to achieve instance-level and fine-grained video caption for the first time. Our structure is specifically designed to incorporate instances, background, and camera movement. For each instance, we specify class, appearance, actions, motion, and position. To enhance the fidelity and accuracy of video captions, we focus on two key aspects: 1) From Global Video to Local Instances: For each instance, we propose an auxiliary models cluster (AMC) to isolate it from the original video and obtain the corresponding position and category information. This operation minimizes interference from unrelated regions while retaining as much of the original video's information as possible. 2) From Dense prompt to Structured Phrases: We leverage multimodal large language models (MLLMs) in an improved Chain-of-Thought (CoT) process to obtain concise yet accurate descriptions of textures, camera movement, actions and motion for each instance. This reduces the probability of hallucinations and irrelevant content produced by the lan" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 264, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 264, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 264, + 713 + ], + "type": "text", + "content": "* Equal contributions. Ying Tai is the corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28974" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 210 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 210 + ], + "type": "image", + "image_path": "3a5f3c5616adef5084ff90e77d498ddf70b3e7df04b9f716f92f6dd2add08c58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 213, + 100, + 220 + ], + "lines": [ + { + "bbox": [ + 58, + 213, + 100, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 213, + 100, + 220 + ], + "type": "text", + "content": "[InstanceCap](Ours)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 213, + 497, + 221 + ], + "lines": [ + { + "bbox": [ + 391, + 213, + 497, + 221 + ], + "spans": [ + { + "bbox": [ + 391, + 213, + 497, + 221 + ], + "type": "text", + "content": "Wrong Ambiguity Detailed" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 369, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 436 + ], + "type": "text", + "content": "Figure 1. Top: Comparison of the reconstruction via-recaption results between InstanceCap and state-of-the-art captioning methods for annotating the ground truth video. InstanceCap produces results that more closely resemble the original video, showing greater detail fidelity (highlighted by the red circle). Bottom: The corresponding captions generated by InstanceCap and others. Red denotes incorrect captions, blue represents ambiguous captions, and green indicates detailed and accurate descriptions of video. Specific visual hints are marked as A, B, and C for clarity. All videos are generated using the same video generation product, Hailuo AI1, which has robust prompt-following capabilities, clearly highlighting the effectiveness of InstanceCap." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 224, + 233, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 224, + 233, + 236 + ], + "spans": [ + { + "bbox": [ + 58, + 224, + 233, + 236 + ], + "type": "text", + "content": "\"Global Description\": A woman with blonde hair, wearing a dark blue shirt with white stripes, is speaking and gesturing with her hands." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 236, + 114, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 236, + 114, + 241 + ], + "spans": [ + { + "bbox": [ + 63, + 236, + 114, + 241 + ], + "type": "text", + "content": "Structured Description\": {" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 242, + 82, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 242, + 82, + 247 + ], + "spans": [ + { + "bbox": [ + 63, + 242, + 82, + 247 + ], + "type": "text", + "content": "Main Obj" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 248, + 101, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 248, + 101, + 255 + ], + "spans": [ + { + "bbox": [ + 63, + 248, + 101, + 255 + ], + "type": "text", + "content": "No.0-1 Class:\"person\"," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 256, + 244, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 256, + 244, + 281 + ], + "spans": [ + { + "bbox": [ + 58, + 256, + 244, + 281 + ], + "type": "text", + "content": "\"Appearance.\" The person in the video is a woman with blonde hair, wearing glasses, her face is to the right of the frame. She is dressed in a dark blue blouse with white stripes. The woman appears to be middle-aged and is of Caucasian race. She is wearing a necklace " + }, + { + "bbox": [ + 58, + 256, + 244, + 281 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 58, + 256, + 244, + 281 + ], + "type": "text", + "content": " and a black dress. She has a thick skirt and a facial expression throughout the video, indicating that she is engaged in a conversation or presentation.\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 281, + 236, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 281, + 236, + 298 + ], + "spans": [ + { + "bbox": [ + 58, + 281, + 236, + 298 + ], + "type": "text", + "content": "\"Actions and Motion\": \"Her gestures are expressive, with her hands moving fluidly to emphasize her points and she is holding a pen. The way she speaks suggests she is passionate about the topic at hand, her body language open and inviting.\"" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 298, + 153, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 298, + 153, + 303 + ], + "spans": [ + { + "bbox": [ + 63, + 298, + 153, + 303 + ], + "type": "text", + "content": "\"Position\": \"person stays at bottom-middle.\"" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 312, + 246, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 312, + 246, + 343 + ], + "spans": [ + { + "bbox": [ + 58, + 312, + 246, + 343 + ], + "type": "text", + "content": "\"Background Detail\": \"The environment in the video is a stage or presentation area with a grand and elegant atmosphere. The backdrop consists of a red curtain, which adds a touch of drama and sophistication to the setting. Text square 'ROVO' can be seen in the lower right corner. The lighting is bright and evenly distributed, illuminating the stage and creating a clear and focused environment for the speaker. The overall ambiance is professional and formal, indicative of a well-organized and prestigious event.\"" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 343, + 242, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 343, + 242, + 363 + ], + "spans": [ + { + "bbox": [ + 58, + 343, + 242, + 363 + ], + "type": "text", + "content": "'Camera Movement': The camera is static, maintaining a short distance shot throughout the video. The shooting angle remains consistent, capturing the speaker from the waist up.' \n1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 251, + 215, + 306, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 215, + 306, + 223 + ], + "spans": [ + { + "bbox": [ + 251, + 215, + 306, + 223 + ], + "type": "text", + "content": "[PANDA-70M](CVPR 2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 251, + 223, + 383, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 223, + 383, + 228 + ], + "spans": [ + { + "bbox": [ + 251, + 223, + 383, + 228 + ], + "type": "text", + "content": "A woman wearing glasses and a blue tie is sitting on a red couch." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 250, + 235, + 320, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 235, + 320, + 241 + ], + "spans": [ + { + "bbox": [ + 250, + 235, + 320, + 241 + ], + "type": "text", + "content": "[ShareGPT4Video](NeuIPS 2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 250, + 241, + 386, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 241, + 386, + 363 + ], + "spans": [ + { + "bbox": [ + 250, + 241, + 386, + 363 + ], + "type": "text", + "content": "The video features an individual standing in front of a red background, engaging in what appears to be a presentation or discussion. Throughout the video, the person is seen making various gestures with their hands, indicating a dynamic and expressive delivery. Initially, the person's hands are raised, palms facing upwards, suggesting an explanatory or emphatic gesture. As the video progresses, the individual's gestures evolve; the hands are lowered and then raised again, with the fingers pointing upwards, indicating a shift in the emphasis or point of discussion. Subsequently, the hands are brought closer together, with the fingers interlocked, suggesting a moment of emphasis or a transition in the presentation. Finally, the hands are raised again, with the fingers spread apart, indicating a continuation of the expressive gestures. Throughout the video, the individual's attire remains consistent, featuring a dark-colored top with a lighter-colored collar, and the background remains a solid red, providing a stark contrast that keeps the focus on the speaker. The camera angle and framing stay constant, with no discernible movement, maintaining a steady focus on the individual and their gestures. The lighting is even, ensuring that the subject is clearly visible against the red backdrop. There are no changes in the environment, the individual's attire, or the camera's perspective, indicating a continuous and uninterrupted presentation or discussion." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 391, + 223, + 444, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 223, + 444, + 229 + ], + "spans": [ + { + "bbox": [ + 391, + 223, + 444, + 229 + ], + "type": "text", + "content": "[MiraData][NeurIPS 2024]" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 391, + 229, + 551, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 229, + 551, + 363 + ], + "spans": [ + { + "bbox": [ + 391, + 229, + 551, + 363 + ], + "type": "text", + "content": "In this video, a person with blond hair is engaging in an animated conversation while seated. They are wearing glasses and a striped shirt and express themselves through a series of dynamic hand gestures and head movements. The individual appears to be deeply focused on the dialogue, occasionally looking down or away, possibly to emphasize their points or formulate thoughts. The person's gestures escalate throughout the video, suggesting an increasing intensity or passion about the topic. The main subject, identifiable by their blond hair and glasses, transitions from initially looking down to making various expressive hand gestures and head movements. Their body language involves considerable arm movement, indicating engagement and enthusiasm in the conversation. The sequence begins with the individual in a somewhat reserved posture, escalating to more expansive and assertive gestures, before returning to a composed state. The background of the video is a studio setting with prominent red and white color blocks, suggesting a professional and possibly formal environment. Part of a sign or text that reads \"RM\" is visible, indicating a possible connection to a broadcast or presentation context. The exact location isn't specified, but the controlled indoor setting implies a purposeful and staged production. The view shot remains consistent, focused on the upper body of the main subject, ensuring their gestures and facial expressions are clearly visible. The camera maintains a mid-shot with minimal movement, emphasizing the individual's actions and expressions. There are no noticeable changes in the shooting angle, keeping the subject centered throughout the frames to convey their animated discourse. The video employs a straightforward, television broadcast style, emphasizing clear and focused framing to capture the subject's expressive communication in a professional studio environment." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 451, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 475 + ], + "type": "text", + "content": "guage model compared to traditional caption methods that directly describe video content in a complex, dense caption." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 478, + 296, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 296, + 634 + ], + "type": "text", + "content": "To validate the effectiveness of InstanceCap in T2V generation, we constructed a high-definition video dataset comprising 22K samples to create a training dataset with our instance-aware structured captions, named InstanceVid. At the inference stage, we also implemented a prompt enhancement pipeline tailored to our structured captioning method, enabling the generation of concise captions that better align with user needs. Our InstanceCap integrates seamlessly with existing diffusion models. Experimental results demonstrate that after finetuned with our InstanceVid, the T2V model exhibits better ability with prompt following on details and motion actions. In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 55, + 642, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 701 + ], + "type": "text", + "content": "- We propose InstanceCap, the first instance-aware structured caption method for text-to-video generation. We also constructed a 22K InstanceVid dataset during training, and developed an enhancement pipeline that tailored to InstanceCap structure during inference." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "- We design the AMC paradigm to convert global video" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 451, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 451, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 321, + 451, + 555, + 498 + ], + "type": "text", + "content": "into instances, enhancing instance fidelity. Additionally, we propose an improved CoT pipeline with MLLMs to refine dense prompts into structured phrases, achieving concise yet precise descriptions." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 499, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 499, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 314, + 499, + 555, + 558 + ], + "type": "text", + "content": "- Extensive experiments on video reconstruction demonstrate that our InstanceCap significantly enhances the fidelity between captions and videos. T2V models fine-tuned on our InstanceVid further achieve more precise generation on instance details and motion actions." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 568, + 406, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 568, + 406, + 581 + ], + "spans": [ + { + "bbox": [ + 314, + 568, + 406, + 581 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 588, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 555, + 696 + ], + "type": "text", + "content": "Video recaptioning. Advancements in text-to-video generation demand high-quality video-text datasets to build robust foundational video models for visual-language alignment. Current video recaption methods fall into two main categories: manual annotation [1, 25, 39] and end-to-end recaption using multimodal large language models (MLLM). Although manual annotation provides higher accuracy, scaling datasets to meet the needs of high-quality video generation models remains a substantial challenge. Recent" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 703, + 416, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 703, + 416, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 703, + 416, + 712 + ], + "type": "text", + "content": "1 https://hailuoai.com/video" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28975" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "advances in MLLM have demonstrated impressive capabilities in video understanding and description generation. Panda [4] and InternVid [27] are with short captions, offering computational efficiency but frequently omitting crucial content and exhibit low video fidelity. OpenVid-1M [18], Vript [30] and ShareGPT4Video [3] are with dense captions, which provide richer content but face challenges: Hallucinations due to complexity, inclusion of redundant information, and text encoder truncation caused by excessively long text. MiraData [11] achieves coarse-level structured captions that attempts to mediate these issues but struggle with fine detail and redundancy across structures. Different from the existing video recaption methods, InstanceCap is the first instance-aware structured caption approach for text-to-video generation, ensuring high fidelity between caption and video while reducing hallucinations and repetition." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 279, + 295, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 447 + ], + "type": "text", + "content": "Text-to-video generation. Despite numerous high-quality video generation models [2, 10, 12, 13, 19, 32] perform well with simple directives, they often falter with complex prompts requiring precise instance-level details or intricate camera movements. Analyzing current video-text datasets suggests these limitations may stem from suboptimal training data quality. Traditional recaption methods have not sufficiently captured instance-specific detail granularity or provided comprehensive descriptions of camera movements. To enhance instance-level details and motion consistency, we construct a 22K InstanceVid dataset for training/finetuning, and develop an enhancement pipeline InstanceEnhancer that tailored to the proposed instance-aware structure during inference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 457, + 111, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 111, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 111, + 470 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 478, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 295, + 562 + ], + "type": "text", + "content": "In Section 3.1, we first present the InstanceCap pipeline, as shown in Figure 2. Based on this pipeline, we recapture the carefully curated dataset InstanceVid in Section 3.2, enhancing T2V models' instance generation. Additionally, in Section 3.3, we introduce InstanceEnhancer to convert short prompts into our proposed instance-aware structured caption format during inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 137, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 137, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 137, + 582 + ], + "type": "text", + "content": "3.1. InstanceCap" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 587, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 295, + 694 + ], + "type": "text", + "content": "Video preprocessing with auxiliary model cluster. For continuous video processing, we implemented uniform sampling using " + }, + { + "bbox": [ + 55, + 587, + 295, + 694 + ], + "type": "inline_equation", + "content": "\\text{decord}^2" + }, + { + "bbox": [ + 55, + 587, + 295, + 694 + ], + "type": "text", + "content": ", following the methodology established in LLaVA-Video [36]. This approach enables us to extract essential temporal metadata, including duration, frame count, and timestamps, allowing MLLMs to better interpret temporal sequences in recaptioning tasks. Additionally, to enhance MLLMs' capabilities through structured guidance, we incorporate several auxiliary models to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "achieve accurate object detection, video instance segmentation and camera motion prediction, providing precise prior information to the MLLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 111, + 553, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 111, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 553, + 219 + ], + "type": "text", + "content": "Global description, background detail and camera movement. When describing video content, a high-quality global description should capture primary elements, environmental context, camera movements, angles, and tonal qualities. MLLMs excel in generating high-level video summaries using Chain-of-Thought methodology. By employing carefully designed prompts with CoT, we can guide MLLMs to produce detailed background descriptions while minimizing references to foreground elements." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 220, + 553, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 327 + ], + "type": "text", + "content": "However, MLLMs' limitations in processing discrete frames rather than continuous video segments make it challenging to distinguish camera motion from instance action. To address this, we achieve camera annotations from OpenSora [37] for basic movements (e.g., zoom, rotation) and rely on MLLMs to capture subtle motion attributes (e.g., intensity, speed). The integration of camera movement indicators with MLLM capabilities provides comprehensive annotations, as illustrated in Figure 9 (a)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 331, + 553, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 473 + ], + "type": "text", + "content": "Structured Description on Instances. In this subsection, we introduce the details to achieve our instance-aware structured description. To address MLLMs' limitations in instance annotation and the suboptimal results of directly adapting weak visual prompts from images to videos [22, 33], we make full use of the auxiliary model cluster, including initial object detection [6, 40], video instance segmentation with SAM2 [20], and blur non-instance regions to achieve blur background, resulting in better recaptioning outcomes compared to alternative visual prompt methods in video, as shown in Figure 10. To this end, we decompose the global videos into local instances." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 474, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 553, + 677 + ], + "type": "text", + "content": "Next, we describe how to achieve detailed and accurate description of each instance (Figure 3). To maintain instance-level precision, we deliberately constrain the information accessible to MLLMs during instance annotation. Crucially, the global video remains invisible due to our designed blurred backgrounds, preventing MLLMs from confusing information across multiple instances. This approach allows InstanceCap to focus on local instances identified through auxiliary model cluster. Furthermore, to avoid the potential limitation of MLLMs seeing only isolated instances, which could lead to overlooking inter-instance interactions and subsequent misinterpretations, we incorporate the global description mentioned in previous subsection. Specifically, we inject the global description into the instance-annotation MLLMs. This strategic mitigates potential biases in instance action descriptions while maintaining instance-specific accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "To enhance the capability of InstanceCap in capturing instance-level details, we introduce novel insights into the improved CoT process. Our analysis of current MLLM-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 173, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 173, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 173, + 712 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 702, + 173, + 712 + ], + "type": "text", + "content": "https://github.com/dmcl/decord" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28976" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 541, + 304 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 541, + 304 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 541, + 304 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 541, + 304 + ], + "type": "image", + "image_path": "35e2ebb0d3939a7c4875f0e451dc28ac4d52f26a3aee144daa15278081862f8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 95, + 320, + 514, + 449 + ], + "blocks": [ + { + "bbox": [ + 67, + 307, + 542, + 319 + ], + "lines": [ + { + "bbox": [ + 67, + 307, + 542, + 319 + ], + "spans": [ + { + "bbox": [ + 67, + 307, + 542, + 319 + ], + "type": "text", + "content": "Figure 2. Overview of InstanceCap pipeline. Details of \"from dense prompts to structured phrases\" design are shown in Figure 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 95, + 320, + 514, + 449 + ], + "lines": [ + { + "bbox": [ + 95, + 320, + 514, + 449 + ], + "spans": [ + { + "bbox": [ + 95, + 320, + 514, + 449 + ], + "type": "image", + "image_path": "855d4717d08bc26500404d5bec9a3fccd0368c2d5cd5d32deea05cacc31581f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 453, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 453, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 555, + 475 + ], + "type": "text", + "content": "Figure 3. Details on \"from dense prompts to structured phrases\" design. We propose an improved CoT pipeline with carefully designed information interactions (red arrow), which facilitates MLLMs to accurately capture instances with precise descriptions on attributes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 495, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 495, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 495, + 297, + 712 + ], + "type": "text", + "content": "based video recaption methods shows that simple prompts, like “Please provide a detailed description of this video” or Chain-of-Thought prompts “Let’s think step by step... First, please note... Finally, summarize the video content...” fail to capture precise instance details. Additional experiments reveal that MLLMs can effectively annotate these details when given fine-grained prompts, such as “Please note if the characters have any accessories” or “Please observe whether there are spots on the bananas”. To enhance the details of instances, we develop Human-designed Class Hints, crafting specific prompts for about 80 detectable categories using our auxiliary models cluster. Specifically, we present the “person” class prompt here: “Please focus primarily on the person’s facial expressions, attire, age, gender, and race in the video and give a detailed description. Please mention if there are any necklaces, watches, hat or other decoration; otherwise, there’s no need to bring them up.” Besides, we also developed a curated Posi" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 495, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 533 + ], + "type": "text", + "content": "tive/Negative Lexicon to guide MLLMs in generating more aesthetically refined captions. More details can be found in our supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 539, + 394, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 394, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 394, + 551 + ], + "type": "text", + "content": "3.2. InstanceVid" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 556, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 641 + ], + "type": "text", + "content": "InstanceVid is curated via refining a subset from the high-aesthetic, high-consistency videos from OpenVid-1M [18]. To showcase our method's high-fidelity labeling of instance details and motion, we selected video samples that included at least one instance exhibiting high motion intensity during dataset filtering. More statistical information of InstanceVidis provided in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 648, + 423, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 423, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 423, + 659 + ], + "type": "text", + "content": "3.3. InstanceEnhancer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "When the caption distribution of training data differs from that of inference text, it may result in poor instruction-following performance or even problematic outputs. This issue is evident in T2V generation, particularly when long" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28977" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 294, + 124 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 294, + 124 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 294, + 124 + ], + "type": "image", + "image_path": "627a420102a76c64088e89fb60a929e73622418d1f94b825207ac3ced85ed05a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 126, + 295, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 126, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 295, + 182 + ], + "type": "text", + "content": "Figure 4. High-level overview of InstanceEnhancer, illustrating the data flow and the partitioning of stages. For a detailed implementation, refer to the supplemental materials, which provide an in-depth description of the enhancer pipeline design and the interdependencies between the stages." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 193, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 253 + ], + "type": "text", + "content": "captions are used for training but short captions for inference, leading to subpar results. Since users typically prefer short captions, it is essential to enhance short caption effectively to better align with our proposed instance-aware structured caption during training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 254, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 458 + ], + "type": "text", + "content": "As shown in Figure 4, we introduce a tuning-free approach called InstanceEnhancer that achieves this by strictly limiting the generated formats to match the caption corresponding to the training input we used. Our method differs from existing tuning-free caption enhancement approaches, such as those presented in RPG and RAG [5, 31]. Instead of directly enhancing short captions, which we found can introduce inconsistencies between multiple instances' actions and their environmental context in video generation, we employ a two-stage enhancement strategy. In Stage A, short prompts are expanded into detailed long prompts. Stage B(I)&(II) uses both expanded and original captions to segment and enhance specific instances, preserving contextual coherence while ensuring precise instance identification. The quantitative evaluation results are presented in Table 2. Due to the space limitation, more details of our enhancement pipeline can be found in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 471, + 137, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 137, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 137, + 484 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 491, + 172, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 172, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 172, + 504 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "text", + "content": "Video reconstruction with recaptions. To comprehensively evaluate InstanceCap, we conducted a series of experiments, benchmarking against state-of-the-art methods including Panda-70M [4], ShareGPT4Video [3], and Mi-raData [11]. To this end, we carefully selected 500 video clips from OpenVid-1M [18]. For each video, we generated one caption using various caption models, which were then input into the advanced T2V model CogvideoX-5b [32] for video generation. We calculated the differences between the generated videos and the ground truth videos to evaluate each caption model's performance, where smaller visual differences indicate more accurate captions and higher fidelity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We introduced several metrics to evaluate the video reconstruction performance: 1) " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "3\\mathrm{DVAE}_{score}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ": Using 3DVAE from CogVideoX [29, 32] as the backbone, we extract hidden-space representations from both the original videos" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 553, + 250 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 250 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 250 + ], + "type": "image", + "image_path": "9c5c8a17c65e52a276d1bc18f7f784cab315da6ea606313f8a65912570e59848.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 252, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 252, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 555, + 297 + ], + "type": "text", + "content": "Figure 5. Comparison on reconstruction via-recapiction between InstanceCap and MiraData. Corresponding 3DVAE scores are also indicated. Similar semantics shared between InstanceCap and GT are indicated by red circles and lines." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": "and their recaption-reconstructed counterparts. These representations quantify the perceptual distance between them. 2) VQAScore: In order to obtain text-video consistency evaluation results that match human preferences, we use VQAScore [15] for evaluation. Specifically, we sample the video uniformly, and then calculate the VQA value for this set of sampled frames and their corresponding caption. This score is positively correlated with caption quality, providing an objective measure of text-video alignment. 3) Human Evaluation: We conducted a user study with a panel of evaluators to assess caption quality across two aspects: Instance Detail (ID) and Hallucination Scores (HS)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "content": "T2V generation. To thoroughly evaluate the T2V generation performance of our InstanceCap, we utilize the InstanceVid dataset to finetune the state-of-the-art DiT-based T2V generation model Open-Sora [37]. In our evaluation, we compare with Open-Sora, CogVideoX-5b [32], Pyramid-Flow [10], and Open-Sora-Plan [13]. To enable fine-grained, instance-level assessment, we construct a highly challenging evaluation benchmark called Inseval, inspired by recent advancements in T2I and T2V evaluation [7, 8, 15, 16, 28]. Specifically, we curate a diverse evaluation dataset of over 200 carefully crafted instance-level prompt-answer pairs, covering both single-object and multi-object scenarios systematically across five key dimensions: Action, Color, Shape, Texture, and Detail. Motivated by the previous evaluation benchmarks [8], we implement a CoT reasoning framework for generating structured QA responses to ensure objective and consistent evaluation, allowing us to derive instance-level evaluation scores that align closely with human perception and preferences. This approach provides a more nuanced and reliable assessment of instance-level generation quality." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28978" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 302, + 182 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 302, + 182 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 302, + 182 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 302, + 182 + ], + "type": "image", + "image_path": "25ded0840f93a18ca00a23ebbe851b39708636a38609b95ba0162698cad33ff2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 183, + 257, + 192 + ], + "lines": [ + { + "bbox": [ + 70, + 183, + 257, + 192 + ], + "spans": [ + { + "bbox": [ + 70, + 183, + 257, + 192 + ], + "type": "text", + "content": "Single Action: A swift lion sprints across the plain." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 305, + 72, + 541, + 182 + ], + "blocks": [ + { + "bbox": [ + 305, + 72, + 541, + 182 + ], + "lines": [ + { + "bbox": [ + 305, + 72, + 541, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 541, + 182 + ], + "type": "image", + "image_path": "42613abec6fcd1b50fc62ddf9a7cba2ed3d632bcc2b462bed4b62fc3496e9e91.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 183, + 541, + 192 + ], + "lines": [ + { + "bbox": [ + 305, + 183, + 541, + 192 + ], + "spans": [ + { + "bbox": [ + 305, + 183, + 541, + 192 + ], + "type": "text", + "content": "Multiple Action: A fish swims and a turtle moves on the seabed." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "type": "text", + "content": "Figure 6. Visual comparison of InstanceCap and Opensa on Single and Multiple Action Score. In terms of the dynamic degree of video generation, we show better consistency and enhanced multi-instance dynamic generation effect." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 231, + 293, + 332 + ], + "blocks": [ + { + "bbox": [ + 67, + 231, + 293, + 332 + ], + "lines": [ + { + "bbox": [ + 67, + 231, + 293, + 332 + ], + "spans": [ + { + "bbox": [ + 67, + 231, + 293, + 332 + ], + "type": "table", + "html": "
Captioning Methods3DVAEscore↓VQAScore↑Avg. Length
Panda-70M [4]146.630.79913 words
ShareGPT4Video [3]147.000.866174 words
LLaVA-Video-72B [36]147.630.904109 words
MiraData(GPT-4V) [11]144.750.905197 words
InstanceCap(Ours)143.000.916165 words
Ablation Study
InstanceCap w/o camera hints144.380.906166 words
InstanceCap w/o human hints143.870.908165 words
InstanceCap w/ red circle144.120.906162 words
InstanceCap w/ blue background144.380.903165 words
", + "image_path": "7b714acb84caed20dd7219684ce867a9e36865ae47d07a3a9af2c589467b2767.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 417, + 296, + 498 + ], + "blocks": [ + { + "bbox": [ + 55, + 349, + 307, + 404 + ], + "lines": [ + { + "bbox": [ + 55, + 349, + 307, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 307, + 404 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on reconstruction-via-recaption results. The best results are marked in **bold**, and the second-best are **underscored**. As a reference, CogVideoX-5b accepts 226 text tokens, with any excess being truncated. Meanwhile, the results of the ablation study are shown in the lower part of the table." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 417, + 296, + 498 + ], + "lines": [ + { + "bbox": [ + 56, + 417, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 296, + 498 + ], + "type": "table", + "html": "
T2V ModelSingle†Multiple†Average†
ActionColorShapeTextureDetailActionColorTexture
CogVideoX-5B [32]64%60%44%60%20%8%48%40%43.00%
Pyramid-Flow-2B [10]44%68%32%32%7%4%24%16%28.38%
Open-Sora Plan v1.3-2.7B [13]64%44%36%32%27%20%32%12%33.38%
Open-Sora v1.2-1.1B [37]40%56%36%40%13%12%16%16%28.63%
+ Ours w/o InstanceEnhancer56%60%40%48%27%16%32%24%37.88%
+ Ours w/ InstanceEnhancer48%60%56%64%20%24%48%40%45.00%
+ Panda-capitalizer [4]40%48%28%40%20%8%20%12%27.00%
+ ShareGPT4Video [3]40%44%32%24%13%16%8%20%24.63%
+ MiraData [11]44%48%40%44%13%12%28%16%30.63%
+ LLaVA [18]52%52%28%28%20%12%28%16%29.50%
", + "image_path": "483c46ded3d890e815aebb61b3444e2ae94beb85ab2986e5e13c7ea053be48a1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 502, + 296, + 568 + ], + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 568 + ], + "type": "text", + "content": "Table 2. Quantitative comparison between InstanceCap and SOTA video captioning models, all based on the popular T2V model Open-Sora. Additionally, we also compare three powerful T2V models, including CogVideoX-5B, Pyramid-Flow, and Open-Sora Plan. The best results of video captioning methods and Open-Sora are marked in bold, and the second-best are underscored." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 575, + 269, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 269, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 269, + 590 + ], + "type": "text", + "content": "4.2. Comparison with SOTA Caption Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "Qualitative evaluation. CogVideoX-5b [32] is a latent text-to-video generation model known for its capability to generate realistic, long-duration videos based on text prompts. With the integration of our InstanceCap into CogVideoX-5b, as substantiated by Figure 5, the model exhibits a notable enhancement in the video reconstruction capacity. This demonstrates that our instance-aware structured captions retain more of the original video's information, leading to higher fidelity. For instance, our InstanceCap can retain information such as \"glasses\"," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 247, + 419, + 338 + ], + "blocks": [ + { + "bbox": [ + 321, + 247, + 419, + 338 + ], + "lines": [ + { + "bbox": [ + 321, + 247, + 419, + 338 + ], + "spans": [ + { + "bbox": [ + 321, + 247, + 419, + 338 + ], + "type": "image", + "image_path": "b5bb88d6c323125e35aaac5276fd94defd246c28fe0d2f01d5154a3fd6f37c9e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 349, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 384 + ], + "type": "text", + "content": "Figure 7. User study on instance detail and hallucination scores. Ours instance-aware structured caption shows clear advantages compared to the coarse-structured MiraData [11]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 452, + 247, + 549, + 338 + ], + "blocks": [ + { + "bbox": [ + 452, + 247, + 549, + 338 + ], + "lines": [ + { + "bbox": [ + 452, + 247, + 549, + 338 + ], + "spans": [ + { + "bbox": [ + 452, + 247, + 549, + 338 + ], + "type": "image", + "image_path": "1156a94ae26aefb36404a47728809950623b8e6e6c6f13753d75ff969ccc1992.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 425, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 510 + ], + "type": "text", + "content": "\"grey sweater\", and \"relative position of two people\", whereas MiraData [11] almost completely loses these important details. A similar conclusion can be drawn from Figure 1. These results underscore the significant improvements achieved by our InstanceCap, resulting in high-quality reconstruction characterized by rich detail and high fidelity between our captions and the original videos." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": "Quantitative evaluation. Table 1 presents quality comparisons between our InstanceCap and other caption methods across two metrics. Based on the results, we make the following observations: 1) Our method delivers comparable or superior quality to the four baselines, demonstrating its ability to enhance fidelity between videos and captions. This strong alignment with human perceptual judgments and preferences is evident in Figures 1 and 5. 2) Our method consistently excels across all metrics for captions under 200 words, an acceptable length for most T2V models, illustrating its generalizability. Figure 7 compares MiraData's captions with our instance-aware structured ones. We randomly selected videos from an open-domain dataset, and a panel of evaluators assessed caption quality using standardized criteria. Results indicate that our captions offer significantly richer and more accurate de" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28979" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 155 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 155 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 155 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 155 + ], + "type": "table", + "html": "
T2V ModelQuality Metrics↑Temporal Metrics↑Motion MetricsTV Align Metrics
Action Score↑Motion AC Score↑Flow Score↑CLIP Score↑Blip-Blue↑SD Score↑
Open-Sora v1.2-1.1B53.2258.4867.87440.6520.5522.8468.19
+ Panda-captioner54.9857.2856.85440.6719.8119.5867.91
+ ShareGPT4Video52.7560.6768.55440.6620.2823.1067.95
+ LLaVA54.7960.1769.96440.5720.5323.1868.18
+ MiraData54.7160.2665.00440.3720.5322.9868.23
+ InstanceCap(Ours)54.7156.5772.04460.6820.5423.1968.88
", + "image_path": "dbb6b3c0296763b112dc945e5bebe9768f568cf0fd9592cbfd229f08bf5f9ad2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 163, + 555, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 555, + 185 + ], + "type": "text", + "content": "Table 3. Results on Evalcrafter. Whereas dataset quality determines aesthetic metrics, InstanceCap significantly enhances dynamism and text-video alignment, with acceptable trade-offs in temporal consistency. Consistent with Inseval's performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 69, + 198, + 279, + 272 + ], + "blocks": [ + { + "bbox": [ + 69, + 198, + 279, + 272 + ], + "lines": [ + { + "bbox": [ + 69, + 198, + 279, + 272 + ], + "spans": [ + { + "bbox": [ + 69, + 198, + 279, + 272 + ], + "type": "image", + "image_path": "2c8be9b24fc73d5653bdf8d9a39dee3f156d0ff2a8c92dc057dbf50875754167.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 274, + 255, + 281 + ], + "lines": [ + { + "bbox": [ + 69, + 274, + 255, + 281 + ], + "spans": [ + { + "bbox": [ + 69, + 274, + 255, + 281 + ], + "type": "text", + "content": "Multiple Color: A sapphire blue keychain hangs from a light brown bag." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 69, + 282, + 279, + 356 + ], + "blocks": [ + { + "bbox": [ + 69, + 282, + 279, + 356 + ], + "lines": [ + { + "bbox": [ + 69, + 282, + 279, + 356 + ], + "spans": [ + { + "bbox": [ + 69, + 282, + 279, + 356 + ], + "type": "image", + "image_path": "d151e9f76d61556484c759f8370b745af50fba03b537e9287a875b244a8750c2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 356, + 283, + 363 + ], + "lines": [ + { + "bbox": [ + 69, + 356, + 283, + 363 + ], + "spans": [ + { + "bbox": [ + 69, + 356, + 283, + 363 + ], + "type": "text", + "content": "Single Shape: A HEXAGONAL HELMET rests on the seat of the bench, near the window." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 69, + 364, + 278, + 437 + ], + "blocks": [ + { + "bbox": [ + 69, + 364, + 278, + 437 + ], + "lines": [ + { + "bbox": [ + 69, + 364, + 278, + 437 + ], + "spans": [ + { + "bbox": [ + 69, + 364, + 278, + 437 + ], + "type": "image", + "image_path": "d6c47b03e838fa780652532d91699f187ba8f1a15d5b9b964380a005359704f0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 439, + 282, + 447 + ], + "lines": [ + { + "bbox": [ + 67, + 439, + 282, + 447 + ], + "spans": [ + { + "bbox": [ + 67, + 439, + 282, + 447 + ], + "type": "text", + "content": "Single Texture: A smooth paper bag rests on the counter near the coffee machine." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 450, + 295, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 495 + ], + "type": "text", + "content": "Figure 8. Visual comparison of InstanceCap and Open-Sora on instance-level attributes. InstanceCap excels in precise instance detail fidelity and instruction-following capabilities, even with complex multi-instance and multi-attribute scenarios." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 506, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 531 + ], + "type": "text", + "content": "scriptions while reducing hallucination artifacts compared to MiraData's output." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 540, + 198, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 198, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 198, + 551 + ], + "type": "text", + "content": "4.3. Text-to-Video Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "Qualitative evaluation. Figures 6 and 8 provide visual comparisons of the T2V generation results. It can be observed that the infusion of our InstanceVid dataset into Open-Sora [37] serves to further enhance its video synthesis capabilities across four fundamental dimensions (Action, Color, Shape, and Texture). These four different aspects correspond to information in our instance-aware structured captions such as \"Actions and Motion\", \"Appearance\", etc. For instance, our model accurately generates the \"sprints\" action of the lion in Figure 6, as opposed to Open-Sora [37]. In Figure 8, benefiting from our instance-aware caption, our model generates the accurate \"light brown bag\" instance described in caption, where Open-Sora [37]" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 198, + 555, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 234 + ], + "type": "text", + "content": "completely loses this instance. These results indicate that our InstanceVid can provide accurate and instance-level guiding information for video generation models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 239, + 556, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 556, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 556, + 455 + ], + "type": "text", + "content": "Quantitative evaluation. We conduct the quantitative evaluation for InstanceCap using the proposed Inseval metrics in Table 2. We can draw the following conclusions: 1) Fine-tuning with InstanceVid consistently improves all metrics over the base model Open-Sora, demonstrating the effectiveness of InstanceCap. In particular, our Detail score ranks first, justifying the capacity of InstanceCap to capture complex instance detail in video. 2) Compared to other video captioning models finetuned based on Open-Sora, InstanceCap shows clear advantages in video generation tasks. 3) Compared to larger models like CogVideoX or Pyramid-Flow, our approach achieves a higher average metric than Pyramid-Flow, and performs comparably to CogVideoX in several specific metrics like 'Single-Color/Shape/Detail' and 'Multiple-Action', but with much fewer parameters. While primarily focused on enhancing text-to-video instance generation, our method achieves competitive results across EvalCrafter [16] (Table 3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 464, + 408, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 408, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 408, + 476 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 482, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 554, + 564 + ], + "type": "text", + "content": "Effects of human design class hints and camera movement hints. We discuss the impact of incorporating human-designed class hints and camera movement hints on annotation outcomes and provide relevant caption visualizations in Figure 9. These annotations aid MLLMs in focusing more precisely on key elements, resulting in richer and more accurate annotations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Ablations on different video visual prompts. Comparative results for various video visual prompt methods used in caption generation are shown in Figure 10. As illustrated in (a), weak visual prompts derived from static image techniques, such as red circles, bounding boxes, or selective grayscale manipulation of non-target areas [22, 33], limit MLLMs in distinguishing and describe specific targets in multi-instance scenes, leading to attribute blending and vague annotations across instances. In contrast, our method excels in instance-specific feature extraction, accurately differentiating figures like the coach and players. Figure 10 (b) illustrates strong visual prompts that involve complete oc-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28980" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 87, + 105, + 121 + ], + "blocks": [ + { + "bbox": [ + 69, + 87, + 105, + 121 + ], + "lines": [ + { + "bbox": [ + 69, + 87, + 105, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 87, + 105, + 121 + ], + "type": "image", + "image_path": "20f4ae0b14ce1faef2e636b99076c61ad9dd9f2bab56ced8d0cd0a04c3c9139d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 123, + 102, + 129 + ], + "lines": [ + { + "bbox": [ + 76, + 123, + 102, + 129 + ], + "spans": [ + { + "bbox": [ + 76, + 123, + 102, + 129 + ], + "type": "text", + "content": "zoom in" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 123, + 75, + 499, + 140 + ], + "blocks": [ + { + "bbox": [ + 82, + 139, + 113, + 146 + ], + "lines": [ + { + "bbox": [ + 82, + 139, + 113, + 146 + ], + "spans": [ + { + "bbox": [ + 82, + 139, + 113, + 146 + ], + "type": "text", + "content": "w/o hints" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 123, + 75, + 499, + 140 + ], + "lines": [ + { + "bbox": [ + 123, + 75, + 499, + 140 + ], + "spans": [ + { + "bbox": [ + 123, + 75, + 499, + 140 + ], + "type": "image", + "image_path": "04eae64f693c8f9f0af39ad090606112daf0355965c6e6ea5c0eeb3de511b485.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 140, + 435, + 148 + ], + "lines": [ + { + "bbox": [ + 123, + 140, + 435, + 148 + ], + "spans": [ + { + "bbox": [ + 123, + 140, + 435, + 148 + ], + "type": "text", + "content": "The camera movement in the video is smooth and steady, providing a clear and stable view of the scene." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 512, + 81, + 525, + 90 + ], + "lines": [ + { + "bbox": [ + 512, + 81, + 525, + 90 + ], + "spans": [ + { + "bbox": [ + 512, + 81, + 525, + 90 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 437, + 139, + 451, + 150 + ], + "blocks": [ + { + "bbox": [ + 437, + 139, + 451, + 150 + ], + "lines": [ + { + "bbox": [ + 437, + 139, + 451, + 150 + ], + "spans": [ + { + "bbox": [ + 437, + 139, + 451, + 150 + ], + "type": "image", + "image_path": "ab579a622954ef5de8953e60bf25d0e154f1cbe575c0672ba44ad2e61d2f9939.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 152, + 474, + 160 + ], + "lines": [ + { + "bbox": [ + 123, + 152, + 474, + 160 + ], + "spans": [ + { + "bbox": [ + 123, + 152, + 474, + 160 + ], + "type": "text", + "content": "The camera movement in the video is a steady and gradual zoom-in, focusing on a specific area of the cityscape." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 76, + 177, + 101, + 215 + ], + "blocks": [ + { + "bbox": [ + 76, + 177, + 101, + 215 + ], + "lines": [ + { + "bbox": [ + 76, + 177, + 101, + 215 + ], + "spans": [ + { + "bbox": [ + 76, + 177, + 101, + 215 + ], + "type": "image", + "image_path": "8c2951ba09b70e18b1442d1a4258c8a785f21f53c0dc756faae08077646b1097.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 232, + 110, + 239 + ], + "lines": [ + { + "bbox": [ + 79, + 232, + 110, + 239 + ], + "spans": [ + { + "bbox": [ + 79, + 232, + 110, + 239 + ], + "type": "text", + "content": "w/o hints" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 121, + 167, + 499, + 231 + ], + "blocks": [ + { + "bbox": [ + 69, + 150, + 114, + 158 + ], + "lines": [ + { + "bbox": [ + 69, + 150, + 114, + 158 + ], + "spans": [ + { + "bbox": [ + 69, + 150, + 114, + 158 + ], + "type": "text", + "content": "w/ hints(ours)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 121, + 167, + 499, + 231 + ], + "lines": [ + { + "bbox": [ + 121, + 167, + 499, + 231 + ], + "spans": [ + { + "bbox": [ + 121, + 167, + 499, + 231 + ], + "type": "image", + "image_path": "d2211f46f60fea9f26f1dd992a7649436013730f91fac7c12a16fb5e7118262e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 232, + 329, + 240 + ], + "lines": [ + { + "bbox": [ + 122, + 232, + 329, + 240 + ], + "spans": [ + { + "bbox": [ + 122, + 232, + 329, + 240 + ], + "type": "text", + "content": "An older man, his hair white, is wearing a blue denim shirt and jeans." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 512, + 170, + 525, + 178 + ], + "lines": [ + { + "bbox": [ + 512, + 170, + 525, + 178 + ], + "spans": [ + { + "bbox": [ + 512, + 170, + 525, + 178 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 330, + 232, + 343, + 242 + ], + "blocks": [ + { + "bbox": [ + 330, + 232, + 343, + 242 + ], + "lines": [ + { + "bbox": [ + 330, + 232, + 343, + 242 + ], + "spans": [ + { + "bbox": [ + 330, + 232, + 343, + 242 + ], + "type": "image", + "image_path": "596a0a559e3b180bfb8fd320b6873e5a5a39ccc4d85240d3111a1f9c35944d6f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 242, + 534, + 251 + ], + "lines": [ + { + "bbox": [ + 122, + 242, + 534, + 251 + ], + "spans": [ + { + "bbox": [ + 122, + 242, + 534, + 251 + ], + "type": "text", + "content": "The person in the video is a middle-aged male with white hair. He is wearing a denim shirt and jeans, and he has a watch on his left wrist." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 56, + 278, + 299, + 448 + ], + "blocks": [ + { + "bbox": [ + 65, + 243, + 110, + 251 + ], + "lines": [ + { + "bbox": [ + 65, + 243, + 110, + 251 + ], + "spans": [ + { + "bbox": [ + 65, + 243, + 110, + 251 + ], + "type": "text", + "content": "w/ hints(ours)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 278, + 299, + 448 + ], + "lines": [ + { + "bbox": [ + 56, + 278, + 299, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 299, + 448 + ], + "type": "image", + "image_path": "a0bcf6c8a3d6ad6290fe4f44b91ed65c8a16982825923393e8e2b41e4e4cf062.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 451, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 55, + 451, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 555, + 473 + ], + "type": "text", + "content": "Figure 10. (a) Comparison against the weak visual prompt for reconstruction via caption visualization on multi-instance targets. (b) Comparison against color screen backgrounds (red), which may negatively affect MLLM labeling performance." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 306, + 278, + 551, + 448 + ], + "blocks": [ + { + "bbox": [ + 55, + 253, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 555, + 274 + ], + "type": "text", + "content": "Figure 9. (a) Ablation study on the effect of camera movement hints on the accuracy of MLLM labeling. (b) Impact of human-designed class hints on the details of instance labeling." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 278, + 551, + 448 + ], + "lines": [ + { + "bbox": [ + 306, + 278, + 551, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 278, + 551, + 448 + ], + "type": "image", + "image_path": "15f48e79a35e34bba57712e0a20d6a721176110edf77ebb22a8a83333359bfed.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 479, + 295, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 587 + ], + "type": "text", + "content": "clusion of non-target regions to eliminate MLLM interactions with irrelevant instances. Conventional methods use primary color screens, but this often misguides MLLMs, causing them to incorporate incorrect context in captions. Our designed blur background masking approach, however, preserves visual consistency with natural scenes, enabling MLLMs to generate accurate and contextually relevant annotations with minimal prompting guidance. In Table 1, we present the quantitative results of the ablation study." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 591, + 217, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 217, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 217, + 602 + ], + "type": "text", + "content": "5. Conclusions and Limitations" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "In this paper, we introduce InstanceCap, the first instance-aware structured caption method for text-to-video generation. We design an Auxiliary Models Cluster (AMC) to convert global video into instances, enhancing instance fidelity. we also propose an improved CoT pipeline with MLLMs to refine dense prompts into structured phrases, achieving concise yet precise instance descriptions compared to the previous video caption models. Additionally, based on InstanceCap, we curated InstanceVid dataset" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 479, + 554, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 515 + ], + "type": "text", + "content": "for training and InstanceEnhancer during inference, significantly enhancing T2V models' generation capabilities on instance details and actions." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 523, + 555, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 555, + 620 + ], + "type": "text", + "content": "Limitations. Since the precision of InstanceCap partly depends on object detection methods, requiring fine-tuning of the detection model for domain-specific instances, and its benefits decrease in instance-free scenes. Furthermore, the scale of InstanceVid limits its use as a large-scale pre-training dataset. Moving forward, we plan to apply InstanceCap to a larger video dataset and train more powerful T2V models to amplify its impact." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 626, + 424, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 424, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 424, + 639 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "This work was supported by Natural Science Foundation of China: No. 62406135 and Natural Science Foundation of Jiangsu Province: BK20241198. Supported by the Gusu Innovation and Entrepreneur leading Talents: No. ZXL2024362. Supported by the AI & AI for Science Project of Nanjing University: No. 14380007." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28981" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] David L. Chen and William B. Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics (ACL-2011), Portland, OR, 2011. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 179 + ], + "type": "text", + "content": "[2] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, pages 7310-7320, 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "text", + "content": "[3] Lin Chen, Xin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 1, 3, 5, 6, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 300 + ], + "type": "text", + "content": "[4] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. arXiv preprint arXiv:2402.19479, 2024. 1,3,5,6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 294, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 294, + 355 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 294, + 355 + ], + "type": "text", + "content": "[5] Zhennan Chen, Yajie Li, Haofan Wang, Zhibo Chen, Zhengkai Jiang, Jun Li, Qian Wang, Jian Yang, and Ying Tai. Region-aware text-to-image generation via hard binding and soft refinement. arXiv preprint arXiv:2411.06558, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 390 + ], + "type": "text", + "content": "[6] Xiantao Hu, Ying Tai, Xu Zhao, Chen Zhao, Zhenyu Zhang, Jun Li, Bineng Zhong, and Jian Yang. Exploiting multimodal spatial-temporal patterns for video object tracking, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "type": "text", + "content": "[7] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. arXiv preprint arXiv:2307.06350, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 294, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 512 + ], + "type": "text", + "content": "[8] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 294, + 556 + ], + "type": "text", + "content": "[9] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1-38, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 294, + 612 + ], + "type": "text", + "content": "[10] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 1, 3, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 667 + ], + "type": "text", + "content": "[11] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. Miradata: A large-scale video dataset with long durations and structured captions. arXiv preprint arXiv:2407.06358, 2024. 1, 3, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 669, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 294, + 689 + ], + "type": "text", + "content": "[12] Kuaishou. Kling. https://kling.kuaishou.com, 2024. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 712 + ], + "type": "text", + "content": "[13] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 3, 5, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[14] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. arXiv preprint arXiv:2410.05677, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 171 + ], + "type": "text", + "content": "[15] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 173, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 226 + ], + "type": "text", + "content": "[16] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. arXiv preprint arXiv:2310.11440, 2023.5.7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[17] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 272, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 325 + ], + "type": "text", + "content": "[18] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 3, 4, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 522, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 522, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 522, + 338 + ], + "type": "text", + "content": "[19] Pika. Pika 1.0. https://pika.art, 2023. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 339, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 339, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 553, + 415 + ], + "type": "text", + "content": "[20] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dollar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 552, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 552, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 552, + 436 + ], + "type": "text", + "content": "[21] Runway. Gen-2. https://research.runwayml.com/gen2, 2023.1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 437, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 553, + 491 + ], + "type": "text", + "content": "[22] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11987-11997, 2023. 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 493, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 514 + ], + "type": "text", + "content": "[23] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "text", + "content": "[24] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. ModelScope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 548, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 590 + ], + "type": "text", + "content": "[25] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "type": "text", + "content": "[26] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. LaVie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "text", + "content": "[27] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28982" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 644 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[28] Yinwei Wu, Xianpan Zhou, Bing Ma, Xuefeng Su, Kai Ma, and Xinchao Wang. Ifadapter: Instance feature control for grounded text-to-image generation, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "text", + "content": "[29] Rui Xie, Yinhong Liu, Penghao Zhou, Chen Zhao, Jun Zhou, Kai Zhang, Zhenyu Zhang, Jian Yang, Zhenheng Yang, and Ying Tai. Star: Spatial-temporal augmentation with text-to-video models for real-world video super-resolution. arXiv preprint arXiv:2501.02976, 2025. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "text", + "content": "[30] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[31] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In Forty-first International Conference on Machine Learning, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "text", + "content": "[32] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 3, 5, 6, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 350 + ], + "type": "text", + "content": "[33] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. AI Open, 5:30-38, 2024. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "text", + "content": "[34] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. IJCV, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 430 + ], + "type": "text", + "content": "[35] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 432, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 294, + 464 + ], + "type": "text", + "content": "[36] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data, 2024. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 508 + ], + "type": "text", + "content": "[37] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 511, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 564 + ], + "type": "text", + "content": "[38] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-Sora: Democratizing efficient video production for all. https://github.com/hpcaitech/Open-Sora, 2024.1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "type": "text", + "content": "[39] Luowei Zhou, Nathan Louis, and Jason J. Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "text", + "content": "[40] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6748-6758, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28983" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/e6f2e48d-8efc-4d47-81e4-de4b789d0ecb_content_list.json b/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/e6f2e48d-8efc-4d47-81e4-de4b789d0ecb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..53e39d3b2befafc7839de4111a130b4db89d0f29 --- /dev/null +++ b/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/e6f2e48d-8efc-4d47-81e4-de4b789d0ecb_content_list.json @@ -0,0 +1,1856 @@ +[ + { + "type": "text", + "text": "InstanceGaussian: Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception", + "text_level": 1, + "bbox": [ + 133, + 128, + 864, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haijie Li $^{1}$ Yanmin Wu $^{1}$ Jiarui Meng $^{1}$ Qiankun Gao $^{1}$ Zhiyao Zhang $^{3}$ Ronggang Wang $^{1,2}$ Jian Zhang $^{1,2*}$", + "bbox": [ + 200, + 200, + 795, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Electronic and Computer Engineering, Peking University, China \n $^{2}$ Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University \n $^{3}$ College of Information Science and Engineering, Northeastern University, China", + "bbox": [ + 119, + 244, + 875, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 351, + 326, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D scene understanding is vital for applications in autonomous driving, robotics, and augmented reality. However, scene understanding based on 3D Gaussian Splating faces three key challenges: (i) an imbalance between appearance and semantics, (ii) inconsistencies in object boundaries, and (iii) difficulties with top-down instance segmentation. To address these challenges, we propose InstanceGaussian, a method that jointly learns appearance and semantic features while adaptively aggregating instances. Our contributions are as follows: (i) a new Semantic-Scaffold-GS representation to improve feature representation and boundary delineation, (ii) a progressive training strategy for enhanced stability and segmentation, and (iii) a category-agnostic, bottom-up instance aggregation approach for better segmentation. Experimental results demonstrate that our approach achieves state-of-the-art performance in category-agnostic, open-vocabulary 3D point-level segmentation, validating the effectiveness of our proposed method. Project page: https://lhj-git.github.io/InstanceGaussian/", + "bbox": [ + 88, + 383, + 485, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 712, + 220, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D scene understanding aims to comprehensively capture a scene's content, structure, and semantics, addressing tasks like instance segmentation, object detection, and open-vocabulary queries. This is crucial for applications in autonomous driving, robotics, and augmented reality, where detailed scene comprehension can significantly improve performance and safety. Traditional 3D representations,", + "bbox": [ + 89, + 737, + 483, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "such as voxels, point clouds, and meshes, provide structured models but often involve trade-offs between spatial resolution and computational efficiency. 3D Gaussian Splating (3DGS) [16] offers a novel approach that combines the explicitness of traditional representations with neural adaptability, enabling efficient and detailed scene modeling. 3DGS is particularly promising for advancing 3D scene understanding by effectively balancing quality and efficiency in capturing complex scene geometry.", + "bbox": [ + 511, + 352, + 906, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the realm of 3D scene understanding utilizing 3DGS, three critical challenges emerge. 1) The first is an inherent imbalance in the representation of appearance versus semantics. Capturing the fine-grained texture details of a given object or region necessitates multiple Gaussians, each with varied appearance attributes. However, these Gaussians collectively require only a single, shared semantic attribute to accurately represent the semantics of the region. This imbalance results in a scenario where the number of Gaussians may be sufficient for appearance representation yet redundant for semantic expression. 2) The second challenge is an inconsistency between appearance and semantics. Without semantic constraints, i.e., in pure appearance reconstruction, a single Gaussian can represent different objects or regions. At object boundaries, for instance, a single Gaussian may simultaneously represent both the object's foreground and background. Recent approaches, such as language-embedded Gaussian Splatting techniques [25, 34], employ a decoupled learning strategy that overlooks the interdependence between color and semantics, leading to inconsistencies between appearance and semantics and posing significant challenges for both 3D point segmentation and 2D image segmentation. 3) Thirdly, previous approaches are predominantly designed in a top-down manner, often relying on prior category information. For instance, GaussianGrouping [38] defines the number of instances based on 2D tracking results, FastLGS [14] uses", + "bbox": [ + 511, + 493, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 112, + 851, + 243, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work is supported by Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology(Grant No. 2024B1212010006)", + "bbox": [ + 89, + 863, + 482, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14078", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the count of matches in cross-view settings to determine the number of objects, and OpenGaussian [34] relies on a pre-defined number of codebook entries. Such methods are prone to issues arising from uneven category distributions, often resulting in over-segmentation or under-segmentation when handling fine-grained instances in complex scenes.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose InstanceGaussian, a method capable of jointly learning object appearance and instance features while adaptively aggregating objects. Our contributions are summarized as follows:", + "bbox": [ + 89, + 181, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Semantic-Scaffold-GS Representation: We introduce a novel representation method that balances appearance and semantics. By more flexibly allocating the semantic and appearance attributes of Gaussians, this method enables the learning of more accurate object geometric boundaries and yields improved feature representations.", + "2. Progressive Appearance-Semantic Joint Training Strategy: We propose a progressive training strategy that optimizes the joint representation of appearance and semantics* in a step-by-step manner. This ensures their consistency throughout the training process, thereby enhancing the model's stability and improving the accuracy of downstream tasks such as 3D point segmentation and 2D image segmentation.", + "3. Bottom-Up, Category-Agnostic Instance Aggregation: To further enhance the capability of 3D instance segmentation, we introduce a bottom-up, category-agnostic instance aggregation method. This method employs a clustering algorithm based on farthest point sampling and connected component analysis, effectively avoiding over-segmentation or under-segmentation." + ], + "bbox": [ + 89, + 244, + 482, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 579, + 232, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1.3D Representation", + "text_level": 1, + "bbox": [ + 89, + 604, + 269, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D representation involves the digital description and modeling of objects and scenes in three-dimensional space using techniques such as voxels, point clouds, meshes, signed distance functions, and neural representations. Neural representations have made substantial progress with techniques like NeRF [22], which has notably enhanced the quality of novel view synthesis through learning-based optimization. Various methods [2-4, 15] have been proposed to further enhance NeRF's expressive capabilities, but they often face challenges with slow training and rendering speeds. To address these issues, alternative approaches focusing on explicit representations have been developed, which have been effective in reducing the computational burden and decreasing training and inference times. These methods include voxels [10, 27, 30], hash grids [23] and point", + "bbox": [ + 89, + 627, + 482, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "clouds [1, 28, 31].", + "bbox": [ + 511, + 90, + 637, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussian Splitting (3DGS) [16] is an innovative representation technique for modeling and rendering 3D scenes using Gaussian distributions. In recent advancements [11, 12, 20, 21, 36, 40], Scaffold-GS [20] forms a hierarchical and region-aware scene representation that respects the underlying geometric structure of the scene. This strategy exploits scene structure to guide and constrain the distribution of 3D Gaussians, allowing them to locally adapt to varying viewing angles and distances. This innovation allows for enhanced adaptation to various viewpoints and distances while preserving the scene's underlying structure, showcasing the ongoing advancements in 3DGS and its role in shaping the future of neural rendering.", + "bbox": [ + 511, + 107, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. 3D Perception via 2D-3D Correlation", + "text_level": 1, + "bbox": [ + 511, + 316, + 834, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in 3D perception have seen the integration of 2D Vision-Language Models (VLMs) with 3D point cloud processing, leading to significant progress in the field [13, 32, 33, 35, 42]. These methods primarily focus on aligning features and projecting 3D data into 2D to enhance zero-shot learning capabilities. Existing approaches that exploit the 2D-3D correlation can be broadly categorized into two types: one leverages feature information from 2D images, while the other utilizes segmentation masks derived from 2D images.", + "bbox": [ + 511, + 339, + 903, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Feature Distillation. The significant advancements in 2D scene understanding, initiated by SAM [18] and its variants, have inspired integrating semantic features into NeRF. Various methods have been developed to incorporate semantic features from models like CLIP [26] and DINO [5] into NeRF, enhancing 3D segmentation, understanding, and editing tasks. LeRF [17] distills features from accessible VLMs such as CLIP into a 3D scene represented by NeRF. Additionally, [19] presents a 3D open-vocabulary segmentation pipeline using NeRF. Within the 3DGS framework, LEGaussians [29] introduces uncertainty and semantic feature attributes to each Gaussian, rendering a semantic map with corresponding uncertainties. This map is compared with quantized CLIP and DINO dense features extracted from the ground truth image. LangSplat [25] employs a scene-wise language autoencoder to learn language features in the scene-specific latent space, effectively discerning clear boundaries between objects in rendered feature images. Feature3DGS [41] distills SAM's encoder features into 3D Gaussians and employs SAM's decoder to interpret 2D rendered feature maps for segmentation. However, feature distillation-based approaches face challenges such as high training costs, significant storage requirements, and slow inference times, limiting their practical application.", + "bbox": [ + 511, + 492, + 903, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Mask Lifting. Recent works have explored lifting 2D segmentation masks into 3D space [6, 7, 24, 34, 37-39]. SAI3D [39] and Open3DIS [24] propose methods", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "* The term \"semantic\" is used here to distinguish it from \"appearance\". In subsequent sections and code implementations, this concept will be explicitly referred to as \"instance features\".", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14079", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/de6857c57f26014baee394b6ea4a63791380e0d3275d1dc47ca667c06d7d9b04.jpg", + "image_caption": [ + "Figure 1. Top row: Appearance-semantic joint Gaussian representation avoids the imbalance and inconsistency in appearance-semantic learning. Bottom row: Bottom-up instantiation: Over-segmentation is achieved via FPS sampling and clustering, followed by instantiation through graph-connectivity-based aggregation." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 859, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for merging 3D superpoints [9] guided by predictions from SAM. MaskClustering [37] introduces a mask graph clustering approach based on the view consensus rate, targeting open-vocabulary 3D instance segmentation. To achieve the 2D mask consistency across views, GaussianGrouping [38] performs simultaneous reconstruction and segmentation of open-world 3D objects, guided by 2D mask predictions obtained from SAM and 3D spatial consistency constraints. SAGA [6] utilizes a contrastive learning approach with masks generated by SAM. It maps SAM's features into a low-dimensional space using a trainable MLP, replicating these features to address inconsistency issues. Click-Gaussian [7] introduces an efficient method for interactive segmentation of 3D Gaussians. It uses two-level granularity feature fields from 2D masks and incorporates Global Feature-guided Learning (GFL) to resolve cross-view inconsistencies. OpenGaussian [34] employs SAM masks to train instance features with 3D consistency, proposing a two-stage codebook for feature discretization and an instance-level 3D-2D feature association method that links 3D points to 2D masks and CLIP features, demonstrating effectiveness in open vocabulary-based 3D object selection and understanding. However, OpenGaussian uses a fixed codebook entry count and does not fully exploit the continuity of 3D space, which constrains its performance.", + "bbox": [ + 91, + 380, + 483, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 775, + 183, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 801, + 220, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting [16]. Given a set of $n$ 3D Gaussians $\\mathcal{G}$ , each with center $\\mu$ , color $\\pmb{c}$ , opacity $\\sigma$ , rotation $\\pmb{R}$ and scale $\\pmb{S}$ , they are rendered as an image $\\pmb{I}_{ren}$ at a specific viewpoint via differentiable rasterization $\\mathcal{D}$ . The reconstruction loss $\\mathcal{L}_{rgb}$ is then calculated by comparing", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$I_{ren}$ with the GT image $I_{gt}$ . This process can be simplified as shown below, where blue variables are estimizable parameters.", + "bbox": [ + 511, + 380, + 906, + 425 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} (\\boldsymbol {\\mu}, \\boldsymbol {c}, \\boldsymbol {\\sigma}, \\boldsymbol {R}, \\boldsymbol {S}) _ {n} \\xrightarrow {\\mathcal {D}} \\boldsymbol {I} _ {r e n} \\longrightarrow \\mathcal {L} _ {r g b} \\longleftarrow \\boldsymbol {I} _ {g t}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 429, + 892, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OpenGaussian [34]. Based on 3DGS, OpenGaussian adds an instance feature $\\pmb{f} \\in \\mathbb{R}^{6}$ for each Gaussian, passes it through differentiable rasterization $\\mathcal{D}$ to render feature map $M \\in \\mathbb{R}^{H \\times W \\times 6}$ , and computes intra-mask smoothness loss $\\mathcal{L}_{s}$ and inter-mask contrastive loss $\\mathcal{L}_{c}$ with SAM masks $B \\in \\mathbb{R}^{H \\times W \\times 1}$ . This process simplifies as follows, where only the feature $\\pmb{f}$ is optimized while other appearance attributes are frozen (shown in gray).", + "bbox": [ + 511, + 455, + 906, + 575 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} (\\mu , c, \\sigma , R, S, f) _ {n} \\xrightarrow {\\mathcal {D}} M \\longrightarrow \\mathcal {L} _ {s}, \\mathcal {L} _ {c} \\longleftarrow B\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 579, + 898, + 603 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\sum_ {i = 1} ^ {m} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\boldsymbol {B} _ {i, h, w} \\cdot \\left\\| \\boldsymbol {M} _ {:, h, w} - \\bar {\\boldsymbol {M}} _ {i} \\right\\| ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 614, + 905, + 656 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\frac {1}{m (m - 1)} \\sum_ {i = 1} ^ {m} \\sum_ {j = 1, j \\neq i} ^ {m} \\frac {1}{\\left\\| \\bar {\\boldsymbol {M}} _ {i} - \\bar {\\boldsymbol {M}} _ {j} \\right\\| ^ {2}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 665, + 905, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $H, W$ represents the image size and $m$ corresponds to the number of SAM masks in the current view. The $\\mathcal{L}_s$ aims to make each pixel feature $M_{:,h,w}$ converge towards the average feature $\\bar{M}_i$ of its associated mask for intramask smoothness. The $\\mathcal{L}_c$ seeks to maximize the separation between the average features $\\bar{M}_i$ , $\\bar{M}_j$ of different masks.", + "bbox": [ + 511, + 718, + 905, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scaffold-GS [20]. It derives $n$ neural Gaussian points $\\mathcal{G}$ from $n^{\\prime}$ anchor points $\\mathcal{A}$ ( $n = 10n^{\\prime}$ in their implementation). The Gaussian centers $\\pmb{\\mu}$ are obtained by shifting its anchor centers $\\pmb{\\mu}^{\\prime}$ by an offset $\\pmb{o}$ , while other appearance attributes are decoded from the appearance embedding $\\pmb{e}$ through MLPs. As shown below, the blue parts are learn", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14080", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/839492095401407be55dc1fbbc448e1d60a616576b5e911b29dc25eec1fead63.jpg", + "image_caption": [ + "Figure 2. Progressive appearance-semantic joint training. (a) Train appearance only; (b) Independent appearance-semantic training; (c) Joint appearance-semantic training." + ], + "image_footnote": [], + "bbox": [ + 109, + 89, + 465, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "able parameters, and their optimization process is similar to 3DGS.", + "bbox": [ + 89, + 243, + 482, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c} \\overbrace {\\mathcal {A} (\\boldsymbol {\\mu} ^ {\\prime} , \\boldsymbol {e}) _ {n ^ {\\prime}}} ^ {+ o} \\quad \\mathcal {G} (\\vec {\\boldsymbol {\\mu}}, \\underbrace {\\boldsymbol {c} , \\boldsymbol {\\sigma} , \\boldsymbol {R} , \\boldsymbol {S}} _ {\\mathrm {M L P s}}) _ {n} \\xrightarrow {\\mathcal {D}} \\boldsymbol {I} _ {\\mathrm {r e n}} \\to \\mathcal {L} _ {\\mathrm {r g b}} \\leftarrow \\boldsymbol {I} _ {\\mathrm {g t}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 275, + 478, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Appearance-semantic joint Gaussian Representation", + "text_level": 1, + "bbox": [ + 89, + 347, + 482, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(1) Appearance-semantic balanced representation. For the same object or region, the texture details may vary while the semantics remain consistent. Consequently, many Gaussians with diverse appearance attributes are required to represent the texture details of the region. These Gaussians require only a single semantic attribute to effectively capture the region's semantics, meaning the same number of Gaussians may under-represent appearance while overrepresenting semantics. We refer to this phenomenon as the imbalance in the representation of appearance and semantics by Gaussians. To this end, we propose a Semantic-Scaffold-GS representation to balance appearance and semantics:", + "bbox": [ + 89, + 386, + 483, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overbrace {\\underbrace {\\mathcal {A} (\\boldsymbol {\\mu} ^ {\\prime} , \\boldsymbol {e} , \\boldsymbol {f}) _ {n ^ {\\prime}}} _ {\\text {M L P s}} \\overbrace {\\mathcal {G} (\\boldsymbol {\\mu} , \\boldsymbol {c} , \\boldsymbol {\\sigma} , \\boldsymbol {R} , \\boldsymbol {S} , \\boldsymbol {\\tilde {f}}) _ {n}} ^ {\\text {s h a r e}}} ^ {+ o} \\xrightarrow {\\mathcal {I} _ {\\text {r e n}},} \\underset {M} {\\longrightarrow} \\overset {\\mathcal {L} _ {\\text {r g b}},} {\\mathcal {L} _ {s}, \\mathcal {L} _ {c}} \\leftarrow \\overset {\\boldsymbol {I} _ {\\text {g t}},} {\\boldsymbol {B}}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 584, + 483, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The blue parts are the estimizable parameters; see Sec. 3.1 for variable definitions. Similar to Scaffold-GS, we derive $n$ neural Gaussian points $\\mathcal{G}$ from $n'$ anchor points $\\mathcal{A}$ ( $n = 5n'$ in our implementation). Specifically, each group of 5 child Gaussians with different appearance attributes is decoded from the appearance embedding $e$ of their common parent anchor point using MLPs. However, at the semantic level, these 5 child Gaussians share the same instance feature $f$ as their parent anchor point. On the one hand, this representation strengthens the learning of appearance, especially in scenes represented by sparse points (e.g., ScanNet and SLAM-reconstructed scenes); on the other hand, sharing features among neighboring points significantly enhances the efficiency of semantic learning. The process is described in the top row of Fig. 1.", + "bbox": [ + 89, + 642, + 482, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(2) Appearance-semantic consistency learning. Without semantic constraints, i.e., pure appearance reconstruc", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion, a Gaussian can represent across objects/regions. For instance, at boundaries, a single Gaussian may simultaneously represent both the object and the background. Recent language-embedded Gaussian Splatting works [25, 34] follow a paradigm of pre-training appearance reconstruction, freezing the appearance properties, and subsequently learning semantic attributes. This tends to cause inconsistencies between appearance and semantics, posing challenges for 3D point segmentation and 2D image segmentation. Therefore, we propose a progressive appearance-seismic joint training strategy to ensure their consistency as shown in Fig. 2: during the 0-10k steps, only the appearance attributes are trained, then appearance and semantics are trained independently in the 10k-20k steps, and finally, appearance and semantics are jointly trained in the 20k-30k steps. Here, we employ the loss (Eq. (1) (2)) of OpenGaussian to train instance features, but its inter-mask contrastive loss value is too large, causing instability of joint training. We overcome this problem through the following improvements:", + "bbox": [ + 511, + 90, + 906, + 392 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\frac {1}{m (m - 1)} \\sum_ {i = 1} ^ {m} \\sum_ {j = 1, j \\neq i} ^ {m} \\frac {\\mathbf {1} _ {\\| \\bar {M} _ {i} - \\bar {M} _ {j} \\| ^ {2} < \\tau}}{\\left\\| \\bar {M} _ {i} - \\bar {M} _ {j} \\right\\| ^ {2}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 401, + 906, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{1}_{\\| \\bar{M}_i - \\bar{M}_j\\|^2 < \\tau}$ takes values of 1 or 0, aiming to truncate the loss between two objects whose dissimilarity (quantified by the L2 distance) exceeds a threshold $\\tau$ (set to 0.4 in our implementation). This improvement is intuitive because there is no need to enforce that the features of two different objects are infinitely distant from each other.", + "bbox": [ + 511, + 453, + 906, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Bottom-up Category-Agnostic Instantiation", + "text_level": 1, + "bbox": [ + 511, + 553, + 883, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ a bottom-up method for category-agnostic instantiation. Specifically, we first over-segment the scene into numerous sub-objects. Then, we construct a graph of these sub-objects and perform aggregation based on the graph connectivity, adaptively obtaining several independent and complete objects. The bottom row of Fig. 1 describes this process.", + "bbox": [ + 511, + 575, + 905, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(1) Over-segmentation. Inspired by point cloud perception methods, which utilize farthest point sampling (FPS) algorithms to sample seed points for candidate proposals from dense point clouds, we achieve over-segmentation of the scene through FPS-based clustering.", + "bbox": [ + 511, + 681, + 905, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first sample $s$ seed points from all $n$ Gaussians through FPS and use them as the initial centers of $s$ clusters, where $s$ is much greater than the number of objects in the scene, with a default value of 1000. Then we start the clustering process, updating the instance feature of all Gaussians and the $s$ clusters. This process can be formulated as:", + "bbox": [ + 511, + 757, + 905, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {f}, \\bar {\\mu}, \\bar {\\mathbf {I}} = \\phi (\\mathbf {X}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 861, + 903, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} = \\left[ \\operatorname {P E} (\\boldsymbol {\\mu}) \\in \\mathbb {R} ^ {n \\times d}; \\boldsymbol {f} \\in \\mathbb {R} ^ {n \\times 6} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 883, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14081", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Bottom-up category-agnostic instantiation" + ], + "code_body": "Input: $\\mu, f$ - point position and feature, $s$ - the number of over-segmented sub-objects, $r$ -voxel resolution, $\\gamma$ - connectivity threshold.", + "bbox": [ + 91, + 109, + 483, + 155 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: well-segmented complete object $\\hat{\\mathbf{l}},\\hat{\\mathbf{f}}$", + "bbox": [ + 93, + 155, + 400, + 170 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: procedure OVERSEGMENTATION $(\\mu, f, s, r, \\gamma)$", + "2: $\\bar{f},\\bar{\\mu}\\gets FPS(\\pmb {f},\\pmb {\\mu},s)$ ▷initialize s centers", + "3: $\\mathbf{X} \\gets \\operatorname{concat} PE(\\boldsymbol{\\mu})$ and $\\pmb{f}$", + "4: $\\overline{\\mathbf{I}},\\overline{\\pmb{f}},\\overline{\\pmb{\\mu}}\\gets k - means(\\mathbf{X},s)$ $\\triangleright$ Eq. (4)", + "5: return SUBOBJECTAGG $(\\mu, \\overline{\\mathbf{I}}, \\overline{\\mathbf{f}}, s, r, \\gamma)$", + "6: end procedure", + "7: procedure SUBOBJECTAGG $(\\mu, \\overline{\\mathbf{I}}, \\overline{f}, s, r, \\gamma)$", + "8: $\\mathbf{G} \\in \\mathbb{R}^{s \\times s} \\gets$ initialize the graph with zeros", + "9: $\\mathbf{V}\\mathbf{o}\\mathbf{x} \\gets \\text{Voxelization}(\\boldsymbol{\\mu},\\bar{\\mathbf{I}},r) \\quad \\triangleright$ sub-object. voxels", + "10: for $i$ in $[0,s)$ do", + "11: for $j$ in $[0,s)$ do", + "12: $dis(i,j)\\gets L2distance(\\bar{\\pmb{f}}_i,\\bar{\\pmb{f}}_j)$", + "13: $\\mathbb{1}(i,j)\\gets near(\\mathbf{Vox}_i,\\mathbf{Vox}_j)$", + "14: $\\mathbf{G}_{ij}\\gets dis(i,j)\\cdot \\mathbb{1}(i,j)$ $\\triangleright$ Eq. (6)", + "15: end for", + "16: end for", + "17: $\\hat{\\mathbf{l}},\\hat{\\pmb{f}}\\gets Aggrecation(\\mathbf{G},\\gamma)$", + "18: return $\\bar{\\mathbf{I}},\\hat{f}$", + "19: end procedure" + ], + "bbox": [ + 101, + 175, + 480, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi(\\cdot)$ denotes the k-means, $\\mathbf{X} \\in \\mathbb{R}^{n \\times (d + 6)}$ is the attribute of the $n$ Gaussians participating in the clustering, which is the concatenation of the positional encoding $(\\mathrm{PE}(\\cdot))$ for position $\\pmb{\\mu}$ and the instance feature $\\pmb{f}$ . Finally, $s$ sub-object clusters are obtained, $\\bar{\\pmb{f}} \\in \\mathbb{R}^{s \\times 6}$ is their feature, $\\bar{\\pmb{\\mu}} \\in \\mathbb{R}^{s \\times 3}$ is their centers, $\\bar{\\mathbf{I}} \\in \\mathbb{Z}_{[0,s]}^n$ represents the label indicating which subobject each point belongs to. During this process, only the feature $\\pmb{f}$ needs to be optimized, while the position $\\pmb{\\mu}$ is frozen. This process is shown in lines 1-4 of Alg. 1.", + "bbox": [ + 89, + 489, + 483, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(2) Graph connectivity-based sub-objects aggregation. This process is shown in lines 7-19 of Alg. 1. We use the $s$ sub-objects obtained from the over-segmentation step as nodes to construct a graph and aggregate the sub-objects into complete objects by evaluating node connectivity. This graph can be represented by an $s \\times s$ adjacency matrix $\\mathbf{G}$ . $\\mathbf{G}_{ij}$ represents the connectivity (edge) between the $i$ -th and $j$ -th sub-objects (nodes):", + "bbox": [ + 89, + 643, + 483, + 763 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {i j} = \\left\\| \\bar {\\boldsymbol {f}} _ {i} - \\bar {\\boldsymbol {f}} _ {j} \\right\\| _ {2} \\cdot \\mathbb {1} (i, j), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 768, + 482, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we design two connectivity criteria: 1) the L2 distance of the node features, a closer distance indicates that the two nodes are more likely to be merged; 2) the spatial relationship between the nodes is represented by the indicator function $\\mathbb{1}(\\cdot)$ , which determines whether two subobjects are adjacent or share voxels by voxelizing the 3D space with a resolution $r$ , and returns 0 or 1.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After constructing the graph, we perform aggregation on the nodes with connectivity values $\\mathbf{G}_{ij}$ in the range $(0,\\gamma]$ using the standard graph connected component algorithm (such as breadth-first search), where $\\gamma$ is set to 0.1 in our implementation. Due to the indicator function, our graph is sparse and can be accelerated through the union-find method. Finally, by aggregating the feature-spatial similar sub-objects, we adaptively obtain several $(e.g.,m)$ independent and complete objects with $\\hat{\\pmb{f}}\\in \\mathbb{R}^{m\\times 6}$ representing their feature, $\\hat{\\mathbf{I}}\\in \\mathbb{Z}_{[0,m)}^n$ representing point label.", + "bbox": [ + 511, + 90, + 906, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments*", + "text_level": 1, + "bbox": [ + 511, + 261, + 653, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Category-Agnostic 3D Instance Segmentation", + "text_level": 1, + "bbox": [ + 511, + 287, + 898, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Settings. 1) Task: Category-agnostic 3D instance segmentation aims to segment independent 3D objects/instances in a given point cloud (Gaussians can naturally be treated as points) scene without considering their semantic categories, ensuring that each point ultimately belongs to only one instance. 2) Dataset and Metrics: Following OpenGaussian, we evaluate using the ScanNet [8] dataset with GT instance-level annotations. More specifically, we calculate the instance-level mIoU and mAcc metrics for the 10 scenes selected by OpenGaussian. 3) Baseline: We mainly compare with Gaussian-based methods OpenGaussian [34] and GaussianGrouping [38] with RGB images and SAM masks as input. Additionally, we provide the performance of the depth-assisted back-projection method MaskClustering [37] as a reference.", + "bbox": [ + 511, + 311, + 906, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results. Quantitative results, as shown in Tab. 1, demonstrate that InstanceGaussian significantly outperforms Gaussian-based methods and achieves comparable results to MaskClustering which uses depth to assist. GaussianGrouping relies on a tracking-based model for cross-view object association, which is prone to tracking failures in complex scenes like ScanNet, characterized by significant viewpoint changes and occlusions, consequently leading to lower performance. Although OpenGaussian achieves instance feature learning with cross-view 3D consistency by the smooth-contrast losses, avoiding explicit tracking-based association, its performance bottleneck lies in the brute-force, naive $k$ -means clustering, lacking parameter adaptability. Our method enables more discriminative feature learning through appearance-semantic joint representation and training, achieving more accurate and generalizable 3D instance segmentation through oversegmentation followed by connectivity-based adaptive aggregation. The visualization results are shown in Fig. 3, where the well-segmented instances faithfully reflect the performance advantages of InstanceGaussian.", + "bbox": [ + 511, + 541, + 908, + 859 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "* Please refer to the supplementary materials for additional 3D/2D visualization results and comparisons with more methods.", + "bbox": [ + 511, + 875, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14082", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3af07ffd268086c9bf3e0506605d324885af5b91db9534728775974547d20ab7.jpg", + "image_caption": [ + "Reference Mesh" + ], + "image_footnote": [], + "bbox": [ + 148, + 102, + 259, + 184 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f2f18bb75e9419f29346ee470695f8a7f01309d89abda4cb53d6d9469fe3d876.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 166, + 186, + 238, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1c7e1330adcaa7ed5e4161e5048e0194c5cad4b33eabd7e4fe4d909cf6b29569.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 148, + 273, + 258, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c13b28042ac9c66b8a675bfe0ec516eaa2d8ad0fb5d5a3cb3aca0ebc5a3b446.jpg", + "image_caption": [ + "Figure 3. Visualization comparison of category-agnostic 3D instance segmentation result. InstanceGaussian outperforms OpenGaussian and GaussinGrouping in accurately distinguishing different 3D objects." + ], + "image_footnote": [], + "bbox": [ + 130, + 353, + 274, + 431 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7263f0a7df430b451156408669dbd06d1d6754a414c9447adb9bd0a4cb78e539.jpg", + "image_caption": [ + "Gaussian Grouping" + ], + "image_footnote": [], + "bbox": [ + 300, + 103, + 408, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/539b2081e77fa5c461b170b022908583f25cf08c7a8435819a65baeadee21c55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 188, + 390, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/249c27bd3b1c7b91d5eb4536319bcdcad1100ac3cddf671a5061ba56c459e4f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 273, + 410, + 353 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ecbf7b5f7d3c5fab2a1d654f3a67c324c7e44a72c5fe5bae8e456c788f9c81c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 353, + 426, + 430 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7de0dfaef2c03219e00d7e4b8fe5c32c65d4741c4404a170d4d10e950c599f57.jpg", + "image_caption": [ + "OpenGaussian" + ], + "image_footnote": [], + "bbox": [ + 447, + 103, + 555, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cf9277f340f954ce4fcd9f8ee5f46b58c4ff7f09106c6e8958651cf1f67eb172.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 188, + 535, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c4d2b95b9b52ef4a8ef9ff0a19667bbfdd84114b63e7ede518013cb838aff075.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 273, + 557, + 352 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf146717520e69aa1ba5a9f450d6228cb449ae0f67fce93bd5deed32db76586a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 352, + 571, + 430 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4cd182aa0de18c5cb4b32ac7e8070f296a0ca01a0511ec0c924124ec4d9b5157.jpg", + "image_caption": [ + "InstanceGaussian" + ], + "image_footnote": [], + "bbox": [ + 589, + 103, + 699, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/78ec5a7c517262a98b088049bb56585829fdce2b6c46ff0ee704d32148b69515.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 188, + 679, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/400cfa6eb70c179ebec5fcb175f9d5e359078958962f695a92dc91a59459db9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 273, + 700, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2afbd7a9d6401a61eee41357a6206de00336e036ed86215fe8780b18ebe29d54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 353, + 718, + 430 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/87bdb03b1f6d06b757198b57013ebf80d46965ebcc59cbb9e83b55cc51ed39a0.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 736, + 103, + 844, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5355a039ad2bde57f15b3d4d8a6e778da8582e7286d6ad2728182351813fa1d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 189, + 825, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82a5ce080e1d9b939af1a284eead8c1843f05d1ba126ece630e296b36ee1057e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 732, + 273, + 846, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/54e3e10d6e4eeb3b8ff7c46d959d37c37cc1f9cf23c1bd97f25c01139d643d05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 353, + 861, + 430 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d91acfb6f72fefcf48a828fa39681c8391986cd83309d3a8a17267672774a9d2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodsmIoU ↑mAcc. ↑
Depth-aidedMaskClustering [37]54.1780.95
Gaussian-basedGaussianGrouping [38]22.5530.54
OpenGaussian [34]27.3252.44
InstanceGaussian (ours)50.2780.22
", + "bbox": [ + 94, + 473, + 478, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Open-Vocabulary Query Point Cloud Understanding", + "text_level": 1, + "bbox": [ + 89, + 604, + 482, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After 3D instantiation as mentioned in Sec. 3.3, we adopt the 2D mask-3D instance association method proposed by OpenGaussian to associate each 3D instance with a 512-dim CLIP image feature. This allows us to support text-based understanding by extracting 512-dim CLIP text features and calculating cosine similarity.", + "bbox": [ + 89, + 643, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings. 1) Task: This task takes text as input and finds the matching point cloud by calculating the cosine similarity between text features and point features. In theory, it supports arbitrary open vocabulary input, but to compute quantitative metrics with the annotated GT point cloud, we use text that corresponds to the pre-defined categories as queries. 2) Dataset and Metrics: Following OpenGaussian, we use 19, 15, and 10 categories from ScanNet as text queries, assigning the closest text to the point cloud to compute the mIoU and mAcc for each category. 3) Baseline: We mainly compare with the recent Gaussian-based meth", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e679f53e0bcb69cd48f17972e3be122604a75149d179f01b5e355a980f9534e1.jpg", + "table_caption": [ + "Table 1. Performance of category-agnostic 3D instance segmentation on the Scannet dataset. Accurate is measured by mAcc@0.25." + ], + "table_footnote": [], + "table_body": "
TypeMethods19 classes15 classes10 classes
mIoUmAcc.mIoUmAcc.mIoUmAcc.
DepthMaskClustering [37]34.5954.3234.0045.0738.1056.74
Gaussian-basedLangSplat [25]3.789.115.3513.208.4022.06
LEGaussians [29]3.8410.879.0122.2212.8228.62
OpenGaussian [34]24.7341.5430.1348.2538.2955.19
InstanceGaussian40.6654.0142.5159.1547.9464.01
", + "bbox": [ + 516, + 474, + 905, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance of open vocabulary semantic segmentation on the Scannet dataset. Accurate is measured by mAcc@0.25.", + "bbox": [ + 511, + 561, + 905, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ods LangSplat, LEGaussians, and OpenGaussian. We also provide the performance of a SOTA method MaskClustering utilizing depth information to assist back projection as a comparison.", + "bbox": [ + 511, + 598, + 905, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. The quantitative results are shown in Tab. 2, where InstanceGaussian achieves SOTA performance under the settings of 19, 15, and 10 categories, significantly outperforming Gaussian-based methods. The poor performance of LangSplat and LEGaussians stems from ambiguous Gaussian feature learning and detrimental language feature distillation. The performance of OpenGaussian is limited by its coupled appearance-semantic representation and fixed cluster count. GaussianGrouping is not included in the comparison because it does not support 3D-level text queries. Surprisingly, our method even surpasses the depth-assisted MaskClustering, although it outperforms our method in the category-agnostic instance segmentation task. We analyze that the reason lies in its heavy reliance on high-quality SAM masks. Fig. 4 shows the results of relevance point cloud retrieval using text queries, in which", + "bbox": [ + 511, + 659, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14083", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7b18f8ec1534019b3a592b2618ff47398f4ee7fe7a7deaafa68ee68863bf8b43.jpg", + "image_caption": [ + "Reference mesh" + ], + "image_footnote": [], + "bbox": [ + 133, + 90, + 254, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a34e5d21ec986b8b67087ee9a99191b021ab5aefe10acfa8e8934de43bad8399.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 90, + 372, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1310c2b48c0b0782144e0f2c3d1a5f32433f2599d3ed449450917a4e807957a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 90, + 493, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9acc604a453ab387d82045f7251341967cd78a1f2fb6def3a12dcd3b94dc425b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 90, + 616, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb6adc43521290ee44b41898aeb7dd644976ac076d1c6716fe1a3e6b9dcf4cfa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 90, + 736, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9cdc8b599f3e3c45bc64cfd1ee313adc71f5f37ed0b4901a955e901acec22eba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 90, + 859, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8dc98c7b5d9042a590bcbb2037e8df7751ef62faa2cc9e8970c8dd638018e12a.jpg", + "image_caption": [ + "Figure 4. Open-vocabulary query point cloud Understanding on ScanNet dataset. InstanceGaussian shows advanced text query capabilities." + ], + "image_footnote": [], + "bbox": [ + 145, + 171, + 243, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b556a074f72f0b3286ab90d5c1edf380d369594dd46c56d6a01a13213907d801.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 160, + 364, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d99224d7a64178e9deae598d2d5764272cec06b9c02ec221d2a44e1e917d148.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 160, + 485, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/405581978adba94bba124a188845a900727c01bd4acce8a26f477caf9917ed68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 160, + 604, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/011ede4acb607a7cb30305db3037ad7f7a990d34222686b64a1761ab647f55bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 160, + 728, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8576814350eab78109db68e8929182e8bc3b670a3459f273111e6e65e571f333.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 748, + 160, + 846, + 250 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ece2992c62c5fa6ca89864b99474e9fc02045252ab0bbb83bc5b1fb28c1ca969.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsmIoU ↑mAcc. ↑
LangSplat [25]9.6612.41
LEGaussians [29]16.2123.82
OpenGaussian [34]38.3651.34
InstanceGaussian (ours)45.3058.44
", + "bbox": [ + 130, + 277, + 442, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "both common categories (such as \"bed\", \"curtain\") are used as queries, and the ability to understand open-vocabulary (such as \"ride\", \"music\") is also demonstrated.", + "bbox": [ + 89, + 407, + 483, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Open Vocabulary 3D Object Selection and Rendering", + "text_level": 1, + "bbox": [ + 89, + 467, + 450, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings. 1) Task: Similar to Sec. 4.2, we first use open vocabulary text to select matching 3D Gaussians, and then render the 3D Gaussians into multi-view 2D images through the rasterization pipeline of 3DGS. 2) Dataset and Metrics: Following LangSplat and OpenGaussian, we use the LeRF [17] dataset for evaluation, and after rendering the selected 3D objects into 2D images, we calculate the mIoU and mAcc between the rendered 2D objects and the GT 2D target objects. 3) Baseline: Only Gaussian-based methods possess 3D point-level perception and rendering capabilities, thus we only compare against Gaussian-based methods Langsplat, Legaussian, and OpenGaussian.", + "bbox": [ + 89, + 506, + 483, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Result. Quantitative results are displayed in the Tab. 3, which demonstrates that even when rendering 3D Gaussians onto 2D images for evaluation. As previously analyzed, LangSplat and LEGaussian, due to their poor 3D perception capabilities, fail to accurately select 3D Gaussians relevant to the text query, resulting in low-quality rendering results. The performance advantage of our method stems from two aspects. First, the bottom-up adaptive instantiation ensures good discrimination between objects. Second, the joint representation and learning of appearance and semantics avoid the appearance-semantic inconsistency present in OpenGaussian. The qualitative results in Fig. 5 more intuitively demonstrate this point. Our method renders objects with clearer boundaries, unlike OpenGaussian,", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1830d36a10bad6918e9b038fb72fb72f7fa10eba94d830fb15040e0a7424764d.jpg", + "table_caption": [ + "Table 3. Performance of open vocabulary 3D object selection and rendering on Lerf dataset. Accurate is measured by mAcc@0.25." + ], + "table_footnote": [], + "table_body": "
CaseJotin-repres.Joint-trainSematic Seg.Instance Seg.
#1✓30.7147.40
#2✓33.1549.57
#3✓✓40.6650.27
", + "bbox": [ + 517, + 277, + 903, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Performance of instance and semantic segmentation with different representation and training strategies on Scannet. Evaluated by mIoU.", + "bbox": [ + 511, + 349, + 906, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which due to its appearance-semantic ambiguity, requires some Gaussians to represent multiple objects/regions, especially evident in the boundary areas.", + "bbox": [ + 511, + 414, + 905, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 482, + 666, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1. Gaussian Representation and Training Strategy", + "text_level": 1, + "bbox": [ + 511, + 508, + 890, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct the ablation studies as shown in Tab. 4 to demonstrate the effectiveness of our Semantic-Scaffold-GS joint representation and progressive joint training strategy. The experiments are evaluated using the semantic and instance segmentation tasks on the ScanNet dataset. The joint representation achieves a balance between appearance and semantics, allowing the same anchor point to express different appearances while sharing the same semantics. The progressive joint training avoids inconsistencies in appearance and semantics, enhances the distinguishability between objects, and thus leads to improved instance segmentation and semantic understanding performance.", + "bbox": [ + 511, + 532, + 906, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 6 more intuitively demonstrates our avoidance of inconsistency. The first and second rows respectively visualize the appearance color and instance features of the 3D Gaussian ellipsoid in the 3D viewer, where the ellipse represents the boundary of the Gaussian. Fig. 6 (a)(c) shows the results of our joint training, which achieve clearer and more distinctive expressions in both appearance and semantics compared to Fig. 6 (b)(d). This improvement not only brings benefits to the understanding of the 3D level (i.e., the semantic level, as proven by our quantitative ablation studies), but also enhances the quality of object rendering (i.e., the appearance level, as shown in Sec. 4.3).", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14084", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/301ba5096efc9a3ca427731558995e14442424ead17651bb1a5d240f725647ee.jpg", + "image_caption": [ + "Figure 5. Open-vocabulary 3D object selection and rendering on the LeRF dataset. InstanceGaussian outperforms OpenGaussian in accurately identifying the objects' boundaries by text queries." + ], + "image_footnote": [], + "bbox": [ + 130, + 85, + 864, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc587546c2592cda3251bc002d4a9c2713bc44890d43879d9e21f0a79aa153dc.jpg", + "image_caption": [ + "Figure 6. Visualization results of the appearance color and instance features of the Gaussians ellipsoid in the 3D viewer demonstrate that our method achieves superior appearance-semantic consistency. (a)(c): Results of progressive joint training. (b)(d): Without joint training." + ], + "image_footnote": [], + "bbox": [ + 130, + 304, + 859, + 508 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6dae5ac9d105e4e46affe25a123911ad41734c90b71ab45e18c41594370a6bd4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CasefeaturevoxelizeSemantic Seg.Instance Seg.
#1✓21.7527.50
#2✓28.9843.41
#3✓✓40.6650.27
", + "bbox": [ + 94, + 554, + 483, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Performance of instance and semantic segmentation with different connectivity condition strategies. Evaluated by mIoU.", + "bbox": [ + 89, + 630, + 482, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.2. Aggregation Condition", + "text_level": 1, + "bbox": [ + 89, + 669, + 295, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the process of aggregating over-segmented sub-objects based on a connectivity graph, we considered two connectivity conditions: the L2 distance of instance features, and the proximity judgment after voxelization.", + "bbox": [ + 89, + 688, + 482, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an ablation study on its configuration as shown in Tab. 5. In case #1, we use the color of the Gaussians as a feature and consider their positions for aggregation, which results in poor performance. In case #2, we only consider instance features and achieve a significant performance improvement, demonstrating the effectiveness of our feature learning. In case #3, we take both conditions into account and achieve the best performance, which is an intuitive idea that in aggregation, we need to consider both semantic similarity and spatial proximity.", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 511, + 555, + 763, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce an appearance-semantic compact Gaussian representation to address the imbalance between appearance and instance features using Gaussians. On one hand, this representation enhances the learning of appearance; on the other hand, sharing features among neighboring points significantly boosts the efficiency of semantic learning. The bottom-up category-agnostic instantiation adaptively determines the number of categories, cleverly avoiding the issue of over-segmentation or under-segmentation. In category-agnostic segmentation tasks, we significantly outperform previous models, achieving state-of-the-art (SOTA) performance in open vocabulary understanding for 3D point-level tasks.", + "bbox": [ + 511, + 578, + 906, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite achieving novel performance on multiple tasks, our method still has the following limitations: 1) Longer training time for slower rendering, since the appearance attributes need to be calculated through MLP. 2) Segmentation accuracy is still affected when most of the images in the scene are incorrectly segmented by SAM. Failure case analyses can be found in the supplementary materials. We hope those drawbacks can be addressed in future works.", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14085", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conference on Computer Vision (ECCV), 2020. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), 2021. 2", + "[6] Jiazhong Cen, Jiemin Fang, Chen Yang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Segment any 3d gaussians. arXiv preprint arXiv:2312.00860, 2023. 2, 3", + "[7] Seokhun Choi, Hyeonseop Song, Jaechul Kim, Taehyeong Kim, and Hoseok Do. Click-gaussian: Interactive segmentation to any 3d gaussians. In European Conference on Computer Vision (ECCV), 2024. 2, 3", + "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5", + "[9] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 2004. 3", + "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qin-hong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[11] Qiankun Gao, Jiarui Meng, Chengxiang Wen, Jie Chen, and Jian Zhang. Hicom: Hierarchical coherent" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "motion for dynamic streamable scenes with 3d gaussian splatting. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2", + "[12] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2", + "[13] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[14] Yuzhou Ji, He Zhu, Junshu Tang, Wuyi Liu, Zhizhong Zhang, Xin Tan, and Yuan Xie. Fastlgs: Speeding up language embedded gaussians with feature grid mapping. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 1", + "[15] Yifan Jiang, Peter Hedman, Ben Mildenhall, Dejia Xu, Jonathan T Barron, Zhangyang Wang, and Tianfan Xue. Alignerf: High-fidelity neural radiance fields via alignment-aware training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023. 1, 2, 3", + "[17] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7", + "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[19] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In NeurIPS, 2023. 2", + "[20] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffolds: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3", + "[21] Jiarui Meng, Haijie Li, Yanmin Wu, Qiankun Gao, Shuzhou Yang, Jian Zhang, and Siwei Ma. Mirror-3dgs: Incorporating mirror reflections into 3d gaus" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14086", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sian splatting. In 2024 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2024. 2", + "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2", + "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 2022. 2", + "[24] Phuc Nguyen, Tuan Duc Ngo, Evangelos Kalogerakis, Chuang Gan, Anh Tran, Cuong Pham, and Khoi Nguyen. Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[25] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 1, 2, 4, 6, 7", + "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 2", + "[27] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 2023. 2", + "[28] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM TOG, 2022. 2", + "[29] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 2, 6, 7", + "[30] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[31] Yifan Wang, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 2019. 2", + "[32] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and dense alignment for 3d visual grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[33] Yanmin Wu, Qiankun Gao, Renrui Zhang, Haijie Li, and Jian Zhang. Language-assisted 3d scene understanding. IEEE Transactions on Multimedia, 2025. 2", + "[34] Yanmin Wu, Jiarui Meng, Hajie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Jian Zhang. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. In Advances in Neural Information Processing Systems, volume 37, pages 19114-19138, 2024. 1, 2, 3, 4, 5, 6, 7", + "[35] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[36] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[37] Mi Yan, Jiazhao Zhang, Yan Zhu, and He Wang. Maskclustering: View consensus based mask graph clustering for open-vocabulary 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5, 6", + "[38] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In European Conference on Computer Vision (ECCV), 2024. 1, 3, 5, 6", + "[39] Yingda Yin, Yuzheng Liu, Yang Xiao, Daniel Cohen-Or, Jingwei Huang, and Baoquan Chen. Sai3d: Segment any instance in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[40] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[41] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14087", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[42] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "bbox": [ + 91, + 90, + 480, + 181 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14088", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/e6f2e48d-8efc-4d47-81e4-de4b789d0ecb_model.json b/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/e6f2e48d-8efc-4d47-81e4-de4b789d0ecb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c71be0d4d96ba2050ce0ecb8240dd4f02f6b6038 --- /dev/null +++ b/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/e6f2e48d-8efc-4d47-81e4-de4b789d0ecb_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.135, + 0.13, + 0.865, + 0.177 + ], + "angle": 0, + "content": "InstanceGaussian: Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.202, + 0.796, + 0.241 + ], + "angle": 0, + "content": "Haijie Li\\(^{1}\\) Yanmin Wu\\(^{1}\\) Jiarui Meng\\(^{1}\\) Qiankun Gao\\(^{1}\\) Zhiyao Zhang\\(^{3}\\) Ronggang Wang\\(^{1,2}\\) Jian Zhang\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.245, + 0.876, + 0.318 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Electronic and Computer Engineering, Peking University, China \n\\(^{2}\\)Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University \n\\(^{3}\\)College of Information Science and Engineering, Northeastern University, China" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.352, + 0.327, + 0.368 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.384, + 0.486, + 0.687 + ], + "angle": 0, + "content": "3D scene understanding is vital for applications in autonomous driving, robotics, and augmented reality. However, scene understanding based on 3D Gaussian Splating faces three key challenges: (i) an imbalance between appearance and semantics, (ii) inconsistencies in object boundaries, and (iii) difficulties with top-down instance segmentation. To address these challenges, we propose InstanceGaussian, a method that jointly learns appearance and semantic features while adaptively aggregating instances. Our contributions are as follows: (i) a new Semantic-Scaffold-GS representation to improve feature representation and boundary delineation, (ii) a progressive training strategy for enhanced stability and segmentation, and (iii) a category-agnostic, bottom-up instance aggregation approach for better segmentation. Experimental results demonstrate that our approach achieves state-of-the-art performance in category-agnostic, open-vocabulary 3D point-level segmentation, validating the effectiveness of our proposed method. Project page: https://lhj-git.github.io/InstanceGaussian/" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.222, + 0.729 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.484, + 0.845 + ], + "angle": 0, + "content": "3D scene understanding aims to comprehensively capture a scene's content, structure, and semantics, addressing tasks like instance segmentation, object detection, and open-vocabulary queries. This is crucial for applications in autonomous driving, robotics, and augmented reality, where detailed scene comprehension can significantly improve performance and safety. Traditional 3D representations," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.908, + 0.49 + ], + "angle": 0, + "content": "such as voxels, point clouds, and meshes, provide structured models but often involve trade-offs between spatial resolution and computational efficiency. 3D Gaussian Splating (3DGS) [16] offers a novel approach that combines the explicitness of traditional representations with neural adaptability, enabling efficient and detailed scene modeling. 3DGS is particularly promising for advancing 3D scene understanding by effectively balancing quality and efficiency in capturing complex scene geometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In the realm of 3D scene understanding utilizing 3DGS, three critical challenges emerge. 1) The first is an inherent imbalance in the representation of appearance versus semantics. Capturing the fine-grained texture details of a given object or region necessitates multiple Gaussians, each with varied appearance attributes. However, these Gaussians collectively require only a single, shared semantic attribute to accurately represent the semantics of the region. This imbalance results in a scenario where the number of Gaussians may be sufficient for appearance representation yet redundant for semantic expression. 2) The second challenge is an inconsistency between appearance and semantics. Without semantic constraints, i.e., in pure appearance reconstruction, a single Gaussian can represent different objects or regions. At object boundaries, for instance, a single Gaussian may simultaneously represent both the object's foreground and background. Recent approaches, such as language-embedded Gaussian Splatting techniques [25, 34], employ a decoupled learning strategy that overlooks the interdependence between color and semantics, leading to inconsistencies between appearance and semantics and posing significant challenges for both 3D point segmentation and 2D image segmentation. 3) Thirdly, previous approaches are predominantly designed in a top-down manner, often relying on prior category information. For instance, GaussianGrouping [38] defines the number of instances based on 2D tracking results, FastLGS [14] uses" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.852, + 0.245, + 0.864 + ], + "angle": 0, + "content": "* Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.864, + 0.483, + 0.9 + ], + "angle": 0, + "content": "This work is supported by Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology(Grant No. 2024B1212010006)" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.852, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14078" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "the count of matches in cross-view settings to determine the number of objects, and OpenGaussian [34] relies on a pre-defined number of codebook entries. Such methods are prone to issues arising from uneven category distributions, often resulting in over-segmentation or under-segmentation when handling fine-grained instances in complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.482, + 0.243 + ], + "angle": 0, + "content": "To address these challenges, we propose InstanceGaussian, a method capable of jointly learning object appearance and instance features while adaptively aggregating objects. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.246, + 0.483, + 0.337 + ], + "angle": 0, + "content": "1. Semantic-Scaffold-GS Representation: We introduce a novel representation method that balances appearance and semantics. By more flexibly allocating the semantic and appearance attributes of Gaussians, this method enables the learning of more accurate object geometric boundaries and yields improved feature representations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.337, + 0.483, + 0.457 + ], + "angle": 0, + "content": "2. Progressive Appearance-Semantic Joint Training Strategy: We propose a progressive training strategy that optimizes the joint representation of appearance and semantics* in a step-by-step manner. This ensures their consistency throughout the training process, thereby enhancing the model's stability and improving the accuracy of downstream tasks such as 3D point segmentation and 2D image segmentation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.458, + 0.483, + 0.563 + ], + "angle": 0, + "content": "3. Bottom-Up, Category-Agnostic Instance Aggregation: To further enhance the capability of 3D instance segmentation, we introduce a bottom-up, category-agnostic instance aggregation method. This method employs a clustering algorithm based on farthest point sampling and connected component analysis, effectively avoiding over-segmentation or under-segmentation." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.246, + 0.483, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.58, + 0.233, + 0.595 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.605, + 0.271, + 0.621 + ], + "angle": 0, + "content": "2.1.3D Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.855 + ], + "angle": 0, + "content": "3D representation involves the digital description and modeling of objects and scenes in three-dimensional space using techniques such as voxels, point clouds, meshes, signed distance functions, and neural representations. Neural representations have made substantial progress with techniques like NeRF [22], which has notably enhanced the quality of novel view synthesis through learning-based optimization. Various methods [2-4, 15] have been proposed to further enhance NeRF's expressive capabilities, but they often face challenges with slow training and rendering speeds. To address these issues, alternative approaches focusing on explicit representations have been developed, which have been effective in reducing the computational burden and decreasing training and inference times. These methods include voxels [10, 27, 30], hash grids [23] and point" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.638, + 0.106 + ], + "angle": 0, + "content": "clouds [1, 28, 31]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.905, + 0.304 + ], + "angle": 0, + "content": "3D Gaussian Splitting (3DGS) [16] is an innovative representation technique for modeling and rendering 3D scenes using Gaussian distributions. In recent advancements [11, 12, 20, 21, 36, 40], Scaffold-GS [20] forms a hierarchical and region-aware scene representation that respects the underlying geometric structure of the scene. This strategy exploits scene structure to guide and constrain the distribution of 3D Gaussians, allowing them to locally adapt to varying viewing angles and distances. This innovation allows for enhanced adaptation to various viewpoints and distances while preserving the scene's underlying structure, showcasing the ongoing advancements in 3DGS and its role in shaping the future of neural rendering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.317, + 0.835, + 0.333 + ], + "angle": 0, + "content": "2.2. 3D Perception via 2D-3D Correlation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.905, + 0.49 + ], + "angle": 0, + "content": "Recent advancements in 3D perception have seen the integration of 2D Vision-Language Models (VLMs) with 3D point cloud processing, leading to significant progress in the field [13, 32, 33, 35, 42]. These methods primarily focus on aligning features and projecting 3D data into 2D to enhance zero-shot learning capabilities. Existing approaches that exploit the 2D-3D correlation can be broadly categorized into two types: one leverages feature information from 2D images, while the other utilizes segmentation masks derived from 2D images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.905, + 0.854 + ], + "angle": 0, + "content": "Feature Distillation. The significant advancements in 2D scene understanding, initiated by SAM [18] and its variants, have inspired integrating semantic features into NeRF. Various methods have been developed to incorporate semantic features from models like CLIP [26] and DINO [5] into NeRF, enhancing 3D segmentation, understanding, and editing tasks. LeRF [17] distills features from accessible VLMs such as CLIP into a 3D scene represented by NeRF. Additionally, [19] presents a 3D open-vocabulary segmentation pipeline using NeRF. Within the 3DGS framework, LEGaussians [29] introduces uncertainty and semantic feature attributes to each Gaussian, rendering a semantic map with corresponding uncertainties. This map is compared with quantized CLIP and DINO dense features extracted from the ground truth image. LangSplat [25] employs a scene-wise language autoencoder to learn language features in the scene-specific latent space, effectively discerning clear boundaries between objects in rendered feature images. Feature3DGS [41] distills SAM's encoder features into 3D Gaussians and employs SAM's decoder to interpret 2D rendered feature maps for segmentation. However, feature distillation-based approaches face challenges such as high training costs, significant storage requirements, and slow inference times, limiting their practical application." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "2D Mask Lifting. Recent works have explored lifting 2D segmentation masks into 3D space [6, 7, 24, 34, 37-39]. SAI3D [39] and Open3DIS [24] propose methods" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "* The term \"semantic\" is used here to distinguish it from \"appearance\". In subsequent sections and code implementations, this concept will be explicitly referred to as \"instance features\"." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14079" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.089, + 0.861, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.327, + 0.908, + 0.37 + ], + "angle": 0, + "content": "Figure 1. Top row: Appearance-semantic joint Gaussian representation avoids the imbalance and inconsistency in appearance-semantic learning. Bottom row: Bottom-up instantiation: Over-segmentation is achieved via FPS sampling and clustering, followed by instantiation through graph-connectivity-based aggregation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.381, + 0.485, + 0.76 + ], + "angle": 0, + "content": "for merging 3D superpoints [9] guided by predictions from SAM. MaskClustering [37] introduces a mask graph clustering approach based on the view consensus rate, targeting open-vocabulary 3D instance segmentation. To achieve the 2D mask consistency across views, GaussianGrouping [38] performs simultaneous reconstruction and segmentation of open-world 3D objects, guided by 2D mask predictions obtained from SAM and 3D spatial consistency constraints. SAGA [6] utilizes a contrastive learning approach with masks generated by SAM. It maps SAM's features into a low-dimensional space using a trainable MLP, replicating these features to address inconsistency issues. Click-Gaussian [7] introduces an efficient method for interactive segmentation of 3D Gaussians. It uses two-level granularity feature fields from 2D masks and incorporates Global Feature-guided Learning (GFL) to resolve cross-view inconsistencies. OpenGaussian [34] employs SAM masks to train instance features with 3D consistency, proposing a two-stage codebook for feature discretization and an instance-level 3D-2D feature association method that links 3D points to 2D masks and CLIP features, demonstrating effectiveness in open vocabulary-based 3D object selection and understanding. However, OpenGaussian uses a fixed codebook entry count and does not fully exploit the continuity of 3D space, which constrains its performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.776, + 0.184, + 0.791 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.221, + 0.818 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "3D Gaussian Splatting [16]. Given a set of \\(n\\) 3D Gaussians \\(\\mathcal{G}\\), each with center \\(\\mu\\), color \\(\\pmb{c}\\), opacity \\(\\sigma\\), rotation \\(\\pmb{R}\\) and scale \\(\\pmb{S}\\), they are rendered as an image \\(\\pmb{I}_{ren}\\) at a specific viewpoint via differentiable rasterization \\(\\mathcal{D}\\). The reconstruction loss \\(\\mathcal{L}_{rgb}\\) is then calculated by comparing" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.381, + 0.907, + 0.426 + ], + "angle": 0, + "content": "\\(I_{ren}\\) with the GT image \\(I_{gt}\\). This process can be simplified as shown below, where blue variables are estimizable parameters." + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.43, + 0.893, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} (\\boldsymbol {\\mu}, \\boldsymbol {c}, \\boldsymbol {\\sigma}, \\boldsymbol {R}, \\boldsymbol {S}) _ {n} \\xrightarrow {\\mathcal {D}} \\boldsymbol {I} _ {r e n} \\longrightarrow \\mathcal {L} _ {r g b} \\longleftarrow \\boldsymbol {I} _ {g t}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.577 + ], + "angle": 0, + "content": "OpenGaussian [34]. Based on 3DGS, OpenGaussian adds an instance feature \\( \\pmb{f} \\in \\mathbb{R}^{6} \\) for each Gaussian, passes it through differentiable rasterization \\( \\mathcal{D} \\) to render feature map \\( M \\in \\mathbb{R}^{H \\times W \\times 6} \\), and computes intra-mask smoothness loss \\( \\mathcal{L}_{s} \\) and inter-mask contrastive loss \\( \\mathcal{L}_{c} \\) with SAM masks \\( B \\in \\mathbb{R}^{H \\times W \\times 1} \\). This process simplifies as follows, where only the feature \\( \\pmb{f} \\) is optimized while other appearance attributes are frozen (shown in gray)." + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.58, + 0.9, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} (\\mu , c, \\sigma , R, S, f) _ {n} \\xrightarrow {\\mathcal {D}} M \\longrightarrow \\mathcal {L} _ {s}, \\mathcal {L} _ {c} \\longleftarrow B\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.615, + 0.906, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\sum_ {i = 1} ^ {m} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\boldsymbol {B} _ {i, h, w} \\cdot \\left\\| \\boldsymbol {M} _ {:, h, w} - \\bar {\\boldsymbol {M}} _ {i} \\right\\| ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.666, + 0.906, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\frac {1}{m (m - 1)} \\sum_ {i = 1} ^ {m} \\sum_ {j = 1, j \\neq i} ^ {m} \\frac {1}{\\left\\| \\bar {\\boldsymbol {M}} _ {i} - \\bar {\\boldsymbol {M}} _ {j} \\right\\| ^ {2}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.809 + ], + "angle": 0, + "content": "where \\(H, W\\) represents the image size and \\(m\\) corresponds to the number of SAM masks in the current view. The \\(\\mathcal{L}_s\\) aims to make each pixel feature \\(M_{:,h,w}\\) converge towards the average feature \\(\\bar{M}_i\\) of its associated mask for intramask smoothness. The \\(\\mathcal{L}_c\\) seeks to maximize the separation between the average features \\(\\bar{M}_i\\), \\(\\bar{M}_j\\) of different masks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Scaffold-GS [20]. It derives \\(n\\) neural Gaussian points \\(\\mathcal{G}\\) from \\(n^{\\prime}\\) anchor points \\(\\mathcal{A}\\) (\\(n = 10n^{\\prime}\\) in their implementation). The Gaussian centers \\(\\pmb{\\mu}\\) are obtained by shifting its anchor centers \\(\\pmb{\\mu}^{\\prime}\\) by an offset \\(\\pmb{o}\\), while other appearance attributes are decoded from the appearance embedding \\(\\pmb{e}\\) through MLPs. As shown below, the blue parts are learn" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14080" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.09, + 0.466, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.19, + 0.483, + 0.233 + ], + "angle": 0, + "content": "Figure 2. Progressive appearance-semantic joint training. (a) Train appearance only; (b) Independent appearance-semantic training; (c) Joint appearance-semantic training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.483, + 0.273 + ], + "angle": 0, + "content": "able parameters, and their optimization process is similar to 3DGS." + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.276, + 0.48, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c} \\overbrace {\\mathcal {A} (\\boldsymbol {\\mu} ^ {\\prime} , \\boldsymbol {e}) _ {n ^ {\\prime}}} ^ {+ o} \\quad \\mathcal {G} (\\vec {\\boldsymbol {\\mu}}, \\underbrace {\\boldsymbol {c} , \\boldsymbol {\\sigma} , \\boldsymbol {R} , \\boldsymbol {S}} _ {\\mathrm {M L P s}}) _ {n} \\xrightarrow {\\mathcal {D}} \\boldsymbol {I} _ {\\mathrm {r e n}} \\to \\mathcal {L} _ {\\mathrm {r g b}} \\leftarrow \\boldsymbol {I} _ {\\mathrm {g t}} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.348, + 0.483, + 0.379 + ], + "angle": 0, + "content": "3.2. Appearance-semantic joint Gaussian Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.484, + 0.581 + ], + "angle": 0, + "content": "(1) Appearance-semantic balanced representation. For the same object or region, the texture details may vary while the semantics remain consistent. Consequently, many Gaussians with diverse appearance attributes are required to represent the texture details of the region. These Gaussians require only a single semantic attribute to effectively capture the region's semantics, meaning the same number of Gaussians may under-represent appearance while overrepresenting semantics. We refer to this phenomenon as the imbalance in the representation of appearance and semantics by Gaussians. To this end, we propose a Semantic-Scaffold-GS representation to balance appearance and semantics:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.585, + 0.484, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\overbrace {\\underbrace {\\mathcal {A} (\\boldsymbol {\\mu} ^ {\\prime} , \\boldsymbol {e} , \\boldsymbol {f}) _ {n ^ {\\prime}}} _ {\\text {M L P s}} \\overbrace {\\mathcal {G} (\\boldsymbol {\\mu} , \\boldsymbol {c} , \\boldsymbol {\\sigma} , \\boldsymbol {R} , \\boldsymbol {S} , \\boldsymbol {\\tilde {f}}) _ {n}} ^ {\\text {s h a r e}}} ^ {+ o} \\xrightarrow {\\mathcal {I} _ {\\text {r e n}},} \\underset {M} {\\longrightarrow} \\overset {\\mathcal {L} _ {\\text {r g b}},} {\\mathcal {L} _ {s}, \\mathcal {L} _ {c}} \\leftarrow \\overset {\\boldsymbol {I} _ {\\text {g t}},} {\\boldsymbol {B}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.869 + ], + "angle": 0, + "content": "The blue parts are the estimizable parameters; see Sec. 3.1 for variable definitions. Similar to Scaffold-GS, we derive \\( n \\) neural Gaussian points \\( \\mathcal{G} \\) from \\( n' \\) anchor points \\( \\mathcal{A} \\) (\\( n = 5n' \\) in our implementation). Specifically, each group of 5 child Gaussians with different appearance attributes is decoded from the appearance embedding \\( e \\) of their common parent anchor point using MLPs. However, at the semantic level, these 5 child Gaussians share the same instance feature \\( f \\) as their parent anchor point. On the one hand, this representation strengthens the learning of appearance, especially in scenes represented by sparse points (e.g., ScanNet and SLAM-reconstructed scenes); on the other hand, sharing features among neighboring points significantly enhances the efficiency of semantic learning. The process is described in the top row of Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "(2) Appearance-semantic consistency learning. Without semantic constraints, i.e., pure appearance reconstruc" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.393 + ], + "angle": 0, + "content": "tion, a Gaussian can represent across objects/regions. For instance, at boundaries, a single Gaussian may simultaneously represent both the object and the background. Recent language-embedded Gaussian Splatting works [25, 34] follow a paradigm of pre-training appearance reconstruction, freezing the appearance properties, and subsequently learning semantic attributes. This tends to cause inconsistencies between appearance and semantics, posing challenges for 3D point segmentation and 2D image segmentation. Therefore, we propose a progressive appearance-seismic joint training strategy to ensure their consistency as shown in Fig. 2: during the 0-10k steps, only the appearance attributes are trained, then appearance and semantics are trained independently in the 10k-20k steps, and finally, appearance and semantics are jointly trained in the 20k-30k steps. Here, we employ the loss (Eq. (1) (2)) of OpenGaussian to train instance features, but its inter-mask contrastive loss value is too large, causing instability of joint training. We overcome this problem through the following improvements:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.402, + 0.907, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\frac {1}{m (m - 1)} \\sum_ {i = 1} ^ {m} \\sum_ {j = 1, j \\neq i} ^ {m} \\frac {\\mathbf {1} _ {\\| \\bar {M} _ {i} - \\bar {M} _ {j} \\| ^ {2} < \\tau}}{\\left\\| \\bar {M} _ {i} - \\bar {M} _ {j} \\right\\| ^ {2}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.907, + 0.545 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}_{\\| \\bar{M}_i - \\bar{M}_j\\|^2 < \\tau}\\) takes values of 1 or 0, aiming to truncate the loss between two objects whose dissimilarity (quantified by the L2 distance) exceeds a threshold \\(\\tau\\) (set to 0.4 in our implementation). This improvement is intuitive because there is no need to enforce that the features of two different objects are infinitely distant from each other." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.554, + 0.885, + 0.571 + ], + "angle": 0, + "content": "3.3. Bottom-up Category-Agnostic Instantiation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.906, + 0.681 + ], + "angle": 0, + "content": "We employ a bottom-up method for category-agnostic instantiation. Specifically, we first over-segment the scene into numerous sub-objects. Then, we construct a graph of these sub-objects and perform aggregation based on the graph connectivity, adaptively obtaining several independent and complete objects. The bottom row of Fig. 1 describes this process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.906, + 0.757 + ], + "angle": 0, + "content": "(1) Over-segmentation. Inspired by point cloud perception methods, which utilize farthest point sampling (FPS) algorithms to sample seed points for candidate proposals from dense point clouds, we achieve over-segmentation of the scene through FPS-based clustering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.906, + 0.861 + ], + "angle": 0, + "content": "We first sample \\( s \\) seed points from all \\( n \\) Gaussians through FPS and use them as the initial centers of \\( s \\) clusters, where \\( s \\) is much greater than the number of objects in the scene, with a default value of 1000. Then we start the clustering process, updating the instance feature of all Gaussians and the \\( s \\) clusters. This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.654, + 0.862, + 0.905, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\bar {f}, \\bar {\\mu}, \\bar {\\mathbf {I}} = \\phi (\\mathbf {X}), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.884, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} = \\left[ \\operatorname {P E} (\\boldsymbol {\\mu}) \\in \\mathbb {R} ^ {n \\times d}; \\boldsymbol {f} \\in \\mathbb {R} ^ {n \\times 6} \\right], \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14081" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.092, + 0.461, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Bottom-up category-agnostic instantiation" + }, + { + "type": "algorithm", + "bbox": [ + 0.093, + 0.111, + 0.484, + 0.156 + ], + "angle": 0, + "content": "Input: \\(\\mu, f\\) - point position and feature, \\(s\\) - the number of over-segmented sub-objects, \\(r\\)-voxel resolution, \\(\\gamma\\) - connectivity threshold." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.156, + 0.401, + 0.171 + ], + "angle": 0, + "content": "Output: well-segmented complete object \\(\\hat{\\mathbf{l}},\\hat{\\mathbf{f}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.176, + 0.434, + 0.19 + ], + "angle": 0, + "content": "1: procedure OVERSEGMENTATION \\((\\mu, f, s, r, \\gamma)\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.19, + 0.482, + 0.205 + ], + "angle": 0, + "content": "2: \\(\\bar{f},\\bar{\\mu}\\gets FPS(\\pmb {f},\\pmb {\\mu},s)\\) ▷initialize s centers" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.205, + 0.482, + 0.22 + ], + "angle": 0, + "content": "3: \\(\\mathbf{X} \\gets \\operatorname{concat} PE(\\boldsymbol{\\mu})\\) and \\(\\pmb{f}\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.22, + 0.482, + 0.235 + ], + "angle": 0, + "content": "4: \\(\\overline{\\mathbf{I}},\\overline{\\pmb{f}},\\overline{\\pmb{\\mu}}\\gets k - means(\\mathbf{X},s)\\) \\(\\triangleright\\) Eq. (4)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.235, + 0.387, + 0.251 + ], + "angle": 0, + "content": "5: return SUBOBJECTAGG \\((\\mu, \\overline{\\mathbf{I}}, \\overline{\\mathbf{f}}, s, r, \\gamma)\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.252, + 0.225, + 0.265 + ], + "angle": 0, + "content": "6: end procedure" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.269, + 0.414, + 0.285 + ], + "angle": 0, + "content": "7: procedure SUBOBJECTAGG \\((\\mu, \\overline{\\mathbf{I}}, \\overline{f}, s, r, \\gamma)\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.285, + 0.436, + 0.299 + ], + "angle": 0, + "content": "8: \\(\\mathbf{G} \\in \\mathbb{R}^{s \\times s} \\gets\\) initialize the graph with zeros" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.3, + 0.482, + 0.315 + ], + "angle": 0, + "content": "9: \\(\\mathbf{V}\\mathbf{o}\\mathbf{x} \\gets \\text{Voxelization}(\\boldsymbol{\\mu},\\bar{\\mathbf{I}},r) \\quad \\triangleright\\) sub-object. voxels" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.315, + 0.255, + 0.329 + ], + "angle": 0, + "content": "10: for \\(i\\) in \\([0,s)\\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.33, + 0.281, + 0.344 + ], + "angle": 0, + "content": "11: for \\(j\\) in \\([0,s)\\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.344, + 0.41, + 0.361 + ], + "angle": 0, + "content": "12: \\(dis(i,j)\\gets L2distance(\\bar{\\pmb{f}}_i,\\bar{\\pmb{f}}_j)\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.361, + 0.382, + 0.376 + ], + "angle": 0, + "content": "13: \\(\\mathbb{1}(i,j)\\gets near(\\mathbf{Vox}_i,\\mathbf{Vox}_j)\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.376, + 0.482, + 0.391 + ], + "angle": 0, + "content": "14: \\(\\mathbf{G}_{ij}\\gets dis(i,j)\\cdot \\mathbb{1}(i,j)\\) \\(\\triangleright\\) Eq. (6)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.392, + 0.223, + 0.403 + ], + "angle": 0, + "content": "15: end for" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.406, + 0.198, + 0.418 + ], + "angle": 0, + "content": "16: end for" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.419, + 0.331, + 0.436 + ], + "angle": 0, + "content": "17: \\(\\hat{\\mathbf{l}},\\hat{\\pmb{f}}\\gets Aggrecation(\\mathbf{G},\\gamma)\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.436, + 0.199, + 0.451 + ], + "angle": 0, + "content": "18: return \\(\\bar{\\mathbf{I}},\\hat{f}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.452, + 0.224, + 0.465 + ], + "angle": 0, + "content": "19: end procedure" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.176, + 0.482, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.643 + ], + "angle": 0, + "content": "where \\(\\phi(\\cdot)\\) denotes the k-means, \\(\\mathbf{X} \\in \\mathbb{R}^{n \\times (d + 6)}\\) is the attribute of the \\(n\\) Gaussians participating in the clustering, which is the concatenation of the positional encoding \\((\\mathrm{PE}(\\cdot))\\) for position \\(\\pmb{\\mu}\\) and the instance feature \\(\\pmb{f}\\). Finally, \\(s\\) sub-object clusters are obtained, \\(\\bar{\\pmb{f}} \\in \\mathbb{R}^{s \\times 6}\\) is their feature, \\(\\bar{\\pmb{\\mu}} \\in \\mathbb{R}^{s \\times 3}\\) is their centers, \\(\\bar{\\mathbf{I}} \\in \\mathbb{Z}_{[0,s]}^n\\) represents the label indicating which subobject each point belongs to. During this process, only the feature \\(\\pmb{f}\\) needs to be optimized, while the position \\(\\pmb{\\mu}\\) is frozen. This process is shown in lines 1-4 of Alg. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.764 + ], + "angle": 0, + "content": "(2) Graph connectivity-based sub-objects aggregation. This process is shown in lines 7-19 of Alg. 1. We use the \\(s\\) sub-objects obtained from the over-segmentation step as nodes to construct a graph and aggregate the sub-objects into complete objects by evaluating node connectivity. This graph can be represented by an \\(s \\times s\\) adjacency matrix \\(\\mathbf{G}\\). \\(\\mathbf{G}_{ij}\\) represents the connectivity (edge) between the \\(i\\)-th and \\(j\\)-th sub-objects (nodes):" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.77, + 0.483, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {i j} = \\left\\| \\bar {\\boldsymbol {f}} _ {i} - \\bar {\\boldsymbol {f}} _ {j} \\right\\| _ {2} \\cdot \\mathbb {1} (i, j), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where we design two connectivity criteria: 1) the L2 distance of the node features, a closer distance indicates that the two nodes are more likely to be merged; 2) the spatial relationship between the nodes is represented by the indicator function \\(\\mathbb{1}(\\cdot)\\), which determines whether two subobjects are adjacent or share voxels by voxelizing the 3D space with a resolution \\(r\\), and returns 0 or 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.246 + ], + "angle": 0, + "content": "After constructing the graph, we perform aggregation on the nodes with connectivity values \\(\\mathbf{G}_{ij}\\) in the range \\((0,\\gamma]\\) using the standard graph connected component algorithm (such as breadth-first search), where \\(\\gamma\\) is set to 0.1 in our implementation. Due to the indicator function, our graph is sparse and can be accelerated through the union-find method. Finally, by aggregating the feature-spatial similar sub-objects, we adaptively obtain several \\((e.g.,m)\\) independent and complete objects with \\(\\hat{\\pmb{f}}\\in \\mathbb{R}^{m\\times 6}\\) representing their feature, \\(\\hat{\\mathbf{I}}\\in \\mathbb{Z}_{[0,m)}^n\\) representing point label." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.262, + 0.655, + 0.279 + ], + "angle": 0, + "content": "4. Experiments*" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.289, + 0.899, + 0.306 + ], + "angle": 0, + "content": "4.1. Category-Agnostic 3D Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Settings. 1) Task: Category-agnostic 3D instance segmentation aims to segment independent 3D objects/instances in a given point cloud (Gaussians can naturally be treated as points) scene without considering their semantic categories, ensuring that each point ultimately belongs to only one instance. 2) Dataset and Metrics: Following OpenGaussian, we evaluate using the ScanNet [8] dataset with GT instance-level annotations. More specifically, we calculate the instance-level mIoU and mAcc metrics for the 10 scenes selected by OpenGaussian. 3) Baseline: We mainly compare with Gaussian-based methods OpenGaussian [34] and GaussianGrouping [38] with RGB images and SAM masks as input. Additionally, we provide the performance of the depth-assisted back-projection method MaskClustering [37] as a reference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.909, + 0.86 + ], + "angle": 0, + "content": "Results. Quantitative results, as shown in Tab. 1, demonstrate that InstanceGaussian significantly outperforms Gaussian-based methods and achieves comparable results to MaskClustering which uses depth to assist. GaussianGrouping relies on a tracking-based model for cross-view object association, which is prone to tracking failures in complex scenes like ScanNet, characterized by significant viewpoint changes and occlusions, consequently leading to lower performance. Although OpenGaussian achieves instance feature learning with cross-view 3D consistency by the smooth-contrast losses, avoiding explicit tracking-based association, its performance bottleneck lies in the brute-force, naive \\(k\\)-means clustering, lacking parameter adaptability. Our method enables more discriminative feature learning through appearance-semantic joint representation and training, achieving more accurate and generalizable 3D instance segmentation through oversegmentation followed by connectivity-based adaptive aggregation. The visualization results are shown in Fig. 3, where the well-segmented instances faithfully reflect the performance advantages of InstanceGaussian." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.907, + 0.901 + ], + "angle": 0, + "content": "* Please refer to the supplementary materials for additional 3D/2D visualization results and comparisons with more methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14082" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.088, + 0.258, + 0.102 + ], + "angle": 0, + "content": "Reference Mesh" + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.103, + 0.26, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.188, + 0.24, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.15, + 0.274, + 0.259, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.354, + 0.276, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.089, + 0.415, + 0.103 + ], + "angle": 0, + "content": "Gaussian Grouping" + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.104, + 0.41, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.189, + 0.391, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.275, + 0.411, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.354, + 0.427, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.089, + 0.547, + 0.102 + ], + "angle": 0, + "content": "OpenGaussian" + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.104, + 0.557, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.189, + 0.537, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.274, + 0.558, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.353, + 0.573, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.089, + 0.704, + 0.102 + ], + "angle": 0, + "content": "InstanceGaussian" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.104, + 0.7, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.189, + 0.681, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.275, + 0.702, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.354, + 0.719, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.089, + 0.836, + 0.101 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.104, + 0.845, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.19, + 0.826, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.275, + 0.848, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.724, + 0.354, + 0.862, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.433, + 0.907, + 0.461 + ], + "angle": 0, + "content": "Figure 3. Visualization comparison of category-agnostic 3D instance segmentation result. InstanceGaussian outperforms OpenGaussian and GaussinGrouping in accurately distinguishing different 3D objects." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.474, + 0.48, + 0.561 + ], + "angle": 0, + "content": "
TypeMethodsmIoU ↑mAcc. ↑
Depth-aidedMaskClustering [37]54.1780.95
Gaussian-basedGaussianGrouping [38]22.5530.54
OpenGaussian [34]27.3252.44
InstanceGaussian (ours)50.2780.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.565, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Table 1. Performance of category-agnostic 3D instance segmentation on the Scannet dataset. Accurate is measured by mAcc@0.25." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.606, + 0.483, + 0.638 + ], + "angle": 0, + "content": "4.2. Open-Vocabulary Query Point Cloud Understanding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.734 + ], + "angle": 0, + "content": "After 3D instantiation as mentioned in Sec. 3.3, we adopt the 2D mask-3D instance association method proposed by OpenGaussian to associate each 3D instance with a 512-dim CLIP image feature. This allows us to support text-based understanding by extracting 512-dim CLIP text features and calculating cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Settings. 1) Task: This task takes text as input and finds the matching point cloud by calculating the cosine similarity between text features and point features. In theory, it supports arbitrary open vocabulary input, but to compute quantitative metrics with the annotated GT point cloud, we use text that corresponds to the pre-defined categories as queries. 2) Dataset and Metrics: Following OpenGaussian, we use 19, 15, and 10 categories from ScanNet as text queries, assigning the closest text to the point cloud to compute the mIoU and mAcc for each category. 3) Baseline: We mainly compare with the recent Gaussian-based meth" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.475, + 0.906, + 0.557 + ], + "angle": 0, + "content": "
TypeMethods19 classes15 classes10 classes
mIoUmAcc.mIoUmAcc.mIoUmAcc.
DepthMaskClustering [37]34.5954.3234.0045.0738.1056.74
Gaussian-basedLangSplat [25]3.789.115.3513.208.4022.06
LEGaussians [29]3.8410.879.0122.2212.8228.62
OpenGaussian [34]24.7341.5430.1348.2538.2955.19
InstanceGaussian40.6654.0142.5159.1547.9464.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.563, + 0.906, + 0.591 + ], + "angle": 0, + "content": "Table 2. Performance of open vocabulary semantic segmentation on the Scannet dataset. Accurate is measured by mAcc@0.25." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.659 + ], + "angle": 0, + "content": "ods LangSplat, LEGaussians, and OpenGaussian. We also provide the performance of a SOTA method MaskClustering utilizing depth information to assist back projection as a comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Results. The quantitative results are shown in Tab. 2, where InstanceGaussian achieves SOTA performance under the settings of 19, 15, and 10 categories, significantly outperforming Gaussian-based methods. The poor performance of LangSplat and LEGaussians stems from ambiguous Gaussian feature learning and detrimental language feature distillation. The performance of OpenGaussian is limited by its coupled appearance-semantic representation and fixed cluster count. GaussianGrouping is not included in the comparison because it does not support 3D-level text queries. Surprisingly, our method even surpasses the depth-assisted MaskClustering, although it outperforms our method in the category-agnostic instance segmentation task. We analyze that the reason lies in its heavy reliance on high-quality SAM masks. Fig. 4 shows the results of relevance point cloud retrieval using text queries, in which" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14083" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.092, + 0.255, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.092, + 0.374, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.092, + 0.495, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.092, + 0.617, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.092, + 0.738, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.092, + 0.86, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.161, + 0.24, + 0.171 + ], + "angle": 0, + "content": "Reference mesh" + }, + { + "type": "image", + "bbox": [ + 0.146, + 0.172, + 0.245, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.161, + 0.365, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.161, + 0.486, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.161, + 0.605, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.628, + 0.161, + 0.729, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.161, + 0.848, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.256, + 0.904, + 0.271 + ], + "angle": 0, + "content": "Figure 4. Open-vocabulary query point cloud Understanding on ScanNet dataset. InstanceGaussian shows advanced text query capabilities." + }, + { + "type": "table", + "bbox": [ + 0.131, + 0.279, + 0.443, + 0.364 + ], + "angle": 0, + "content": "
MethodsmIoU ↑mAcc. ↑
LangSplat [25]9.6612.41
LEGaussians [29]16.2123.82
OpenGaussian [34]38.3651.34
InstanceGaussian (ours)45.3058.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.367, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Table 3. Performance of open vocabulary 3D object selection and rendering on Lerf dataset. Accurate is measured by mAcc@0.25." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.453 + ], + "angle": 0, + "content": "both common categories (such as \"bed\", \"curtain\") are used as queries, and the ability to understand open-vocabulary (such as \"ride\", \"music\") is also demonstrated." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.468, + 0.452, + 0.5 + ], + "angle": 0, + "content": "4.3. Open Vocabulary 3D Object Selection and Rendering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.688 + ], + "angle": 0, + "content": "Settings. 1) Task: Similar to Sec. 4.2, we first use open vocabulary text to select matching 3D Gaussians, and then render the 3D Gaussians into multi-view 2D images through the rasterization pipeline of 3DGS. 2) Dataset and Metrics: Following LangSplat and OpenGaussian, we use the LeRF [17] dataset for evaluation, and after rendering the selected 3D objects into 2D images, we calculate the mIoU and mAcc between the rendered 2D objects and the GT 2D target objects. 3) Baseline: Only Gaussian-based methods possess 3D point-level perception and rendering capabilities, thus we only compare against Gaussian-based methods Langsplat, Legaussian, and OpenGaussian." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Result. Quantitative results are displayed in the Tab. 3, which demonstrates that even when rendering 3D Gaussians onto 2D images for evaluation. As previously analyzed, LangSplat and LEGaussian, due to their poor 3D perception capabilities, fail to accurately select 3D Gaussians relevant to the text query, resulting in low-quality rendering results. The performance advantage of our method stems from two aspects. First, the bottom-up adaptive instantiation ensures good discrimination between objects. Second, the joint representation and learning of appearance and semantics avoid the appearance-semantic inconsistency present in OpenGaussian. The qualitative results in Fig. 5 more intuitively demonstrate this point. Our method renders objects with clearer boundaries, unlike OpenGaussian," + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.279, + 0.905, + 0.346 + ], + "angle": 0, + "content": "
CaseJotin-repres.Joint-trainSematic Seg.Instance Seg.
#1✓30.7147.40
#2✓33.1549.57
#3✓✓40.6650.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.35, + 0.907, + 0.392 + ], + "angle": 0, + "content": "Table 4. Performance of instance and semantic segmentation with different representation and training strategies on Scannet. Evaluated by mIoU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.906, + 0.46 + ], + "angle": 0, + "content": "which due to its appearance-semantic ambiguity, requires some Gaussians to represent multiple objects/regions, especially evident in the boundary areas." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.483, + 0.667, + 0.5 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.51, + 0.891, + 0.527 + ], + "angle": 0, + "content": "4.4.1. Gaussian Representation and Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.907, + 0.716 + ], + "angle": 0, + "content": "We conduct the ablation studies as shown in Tab. 4 to demonstrate the effectiveness of our Semantic-Scaffold-GS joint representation and progressive joint training strategy. The experiments are evaluated using the semantic and instance segmentation tasks on the ScanNet dataset. The joint representation achieves a balance between appearance and semantics, allowing the same anchor point to express different appearances while sharing the same semantics. The progressive joint training avoids inconsistencies in appearance and semantics, enhances the distinguishability between objects, and thus leads to improved instance segmentation and semantic understanding performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Fig. 6 more intuitively demonstrates our avoidance of inconsistency. The first and second rows respectively visualize the appearance color and instance features of the 3D Gaussian ellipsoid in the 3D viewer, where the ellipse represents the boundary of the Gaussian. Fig. 6 (a)(c) shows the results of our joint training, which achieve clearer and more distinctive expressions in both appearance and semantics compared to Fig. 6 (b)(d). This improvement not only brings benefits to the understanding of the 3D level (i.e., the semantic level, as proven by our quantitative ablation studies), but also enhances the quality of object rendering (i.e., the appearance level, as shown in Sec. 4.3)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14084" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.086, + 0.865, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.27, + 0.908, + 0.298 + ], + "angle": 0, + "content": "Figure 5. Open-vocabulary 3D object selection and rendering on the LeRF dataset. InstanceGaussian outperforms OpenGaussian in accurately identifying the objects' boundaries by text queries." + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.305, + 0.861, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.513, + 0.908, + 0.543 + ], + "angle": 0, + "content": "Figure 6. Visualization results of the appearance color and instance features of the Gaussians ellipsoid in the 3D viewer demonstrate that our method achieves superior appearance-semantic consistency. (a)(c): Results of progressive joint training. (b)(d): Without joint training." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.555, + 0.484, + 0.627 + ], + "angle": 0, + "content": "
CasefeaturevoxelizeSemantic Seg.Instance Seg.
#1✓21.7527.50
#2✓28.9843.41
#3✓✓40.6650.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.631, + 0.483, + 0.66 + ], + "angle": 0, + "content": "Table 5. Performance of instance and semantic segmentation with different connectivity condition strategies. Evaluated by mIoU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.67, + 0.297, + 0.685 + ], + "angle": 0, + "content": "4.4.2. Aggregation Condition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.749 + ], + "angle": 0, + "content": "In the process of aggregating over-segmented sub-objects based on a connectivity graph, we considered two connectivity conditions: the L2 distance of instance features, and the proximity judgment after voxelization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We conduct an ablation study on its configuration as shown in Tab. 5. In case #1, we use the color of the Gaussians as a feature and consider their positions for aggregation, which results in poor performance. In case #2, we only consider instance features and achieve a significant performance improvement, demonstrating the effectiveness of our feature learning. In case #3, we take both conditions into account and achieve the best performance, which is an intuitive idea that in aggregation, we need to consider both semantic similarity and spatial proximity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.556, + 0.764, + 0.572 + ], + "angle": 0, + "content": "5. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.907, + 0.776 + ], + "angle": 0, + "content": "In this paper, we introduce an appearance-semantic compact Gaussian representation to address the imbalance between appearance and instance features using Gaussians. On one hand, this representation enhances the learning of appearance; on the other hand, sharing features among neighboring points significantly boosts the efficiency of semantic learning. The bottom-up category-agnostic instantiation adaptively determines the number of categories, cleverly avoiding the issue of over-segmentation or under-segmentation. In category-agnostic segmentation tasks, we significantly outperform previous models, achieving state-of-the-art (SOTA) performance in open vocabulary understanding for 3D point-level tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Despite achieving novel performance on multiple tasks, our method still has the following limitations: 1) Longer training time for slower rendering, since the appearance attributes need to be calculated through MLP. 2) Segmentation accuracy is still affected when most of the images in the scene are incorrectly segmented by SAM. Failure case analyses can be found in the supplementary materials. We hope those drawbacks can be addressed in future works." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14085" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.178, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.345 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.513 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.483, + 0.576 + ], + "angle": 0, + "content": "[6] Jiazhong Cen, Jiemin Fang, Chen Yang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Segment any 3d gaussians. arXiv preprint arXiv:2312.00860, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.578, + 0.483, + 0.638 + ], + "angle": 0, + "content": "[7] Seokhun Choi, Hyeonseop Song, Jaechul Kim, Taehyeong Kim, and Hoseok Do. Click-gaussian: Interactive segmentation to any 3d gaussians. In European Conference on Computer Vision (ECCV), 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.64, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.483, + 0.777 + ], + "angle": 0, + "content": "[9] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.779, + 0.483, + 0.867 + ], + "angle": 0, + "content": "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qin-hong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Qiankun Gao, Jiarui Meng, Chengxiang Wen, Jie Chen, and Jian Zhang. Hicom: Hierarchical coherent" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.55, + 0.093, + 0.905, + 0.137 + ], + "angle": 0, + "content": "motion for dynamic streamable scenes with 3d gaussian splatting. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.139, + 0.905, + 0.198 + ], + "angle": 0, + "content": "[12] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.2, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[13] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.366 + ], + "angle": 0, + "content": "[14] Yuzhou Ji, He Zhu, Junshu Tang, Wuyi Liu, Zhizhong Zhang, Xin Tan, and Yuan Xie. Fastlgs: Speeding up language embedded gaussians with feature grid mapping. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[15] Yifan Jiang, Peter Hedman, Ben Mildenhall, Dejia Xu, Jonathan T Barron, Zhangyang Wang, and Tianfan Xue. Alignerf: High-fidelity neural radiance fields via alignment-aware training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.905, + 0.52 + ], + "angle": 0, + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.905, + 0.595 + ], + "angle": 0, + "content": "[17] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.597, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.688, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[19] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.765, + 0.905, + 0.853 + ], + "angle": 0, + "content": "[20] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffolds: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[21] Jiarui Meng, Haijie Li, Yanmin Wu, Qiankun Gao, Shuzhou Yang, Jian Zhang, and Siwei Ma. Mirror-3dgs: Incorporating mirror reflections into 3d gaus" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14086" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.128, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "sian splatting. In 2024 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.138, + 0.483, + 0.198 + ], + "angle": 0, + "content": "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.2, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.483, + 0.35 + ], + "angle": 0, + "content": "[24] Phuc Nguyen, Tuan Duc Ngo, Evangelos Kalogerakis, Chuang Gan, Anh Tran, Cuong Pham, and Khoi Nguyen. Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.352, + 0.483, + 0.427 + ], + "angle": 0, + "content": "[25] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.429, + 0.483, + 0.518 + ], + "angle": 0, + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.521, + 0.483, + 0.609 + ], + "angle": 0, + "content": "[27] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.612, + 0.483, + 0.655 + ], + "angle": 0, + "content": "[28] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM TOG, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[29] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.483, + 0.808 + ], + "angle": 0, + "content": "[30] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.81, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[31] Yifan Wang, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[32] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.55, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "and dense alignment for 3d visual grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.14, + 0.905, + 0.185 + ], + "angle": 0, + "content": "[33] Yanmin Wu, Qiankun Gao, Renrui Zhang, Haijie Li, and Jian Zhang. Language-assisted 3d scene understanding. IEEE Transactions on Multimedia, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.188, + 0.905, + 0.293 + ], + "angle": 0, + "content": "[34] Yanmin Wu, Jiarui Meng, Hajie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Jian Zhang. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. In Advances in Neural Information Processing Systems, volume 37, pages 19114-19138, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.296, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[35] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.905, + 0.48 + ], + "angle": 0, + "content": "[36] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.483, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[37] Mi Yan, Jiazhao Zhang, Yan Zhu, and He Wang. Maskclustering: View consensus based mask graph clustering for open-vocabulary 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[38] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In European Conference on Computer Vision (ECCV), 2024. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.639, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[39] Yingda Yin, Yuzheng Liu, Yang Xiao, Daniel Cohen-Or, Jingwei Huang, and Baoquan Chen. Sai3d: Segment any instance in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.792 + ], + "angle": 0, + "content": "[40] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[41] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14087" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "[42] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14088" + } + ] +] \ No newline at end of file diff --git a/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/e6f2e48d-8efc-4d47-81e4-de4b789d0ecb_origin.pdf b/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/e6f2e48d-8efc-4d47-81e4-de4b789d0ecb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..67af3a6f12b9326187de24c013a1df442fac9e92 --- /dev/null +++ b/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/e6f2e48d-8efc-4d47-81e4-de4b789d0ecb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d247d9280f91f7886da36db264e4ed83b34e03fbfc217e2bb5162b6b7dacac2f +size 8842257 diff --git a/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/full.md b/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9be74354582e340cfd378df1488bdd0e6d1e28aa --- /dev/null +++ b/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/full.md @@ -0,0 +1,360 @@ +# InstanceGaussian: Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception + +Haijie Li $^{1}$ Yanmin Wu $^{1}$ Jiarui Meng $^{1}$ Qiankun Gao $^{1}$ Zhiyao Zhang $^{3}$ Ronggang Wang $^{1,2}$ Jian Zhang $^{1,2*}$ + +$^{1}$ School of Electronic and Computer Engineering, Peking University, China + $^{2}$ Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University + $^{3}$ College of Information Science and Engineering, Northeastern University, China + +# Abstract + +3D scene understanding is vital for applications in autonomous driving, robotics, and augmented reality. However, scene understanding based on 3D Gaussian Splating faces three key challenges: (i) an imbalance between appearance and semantics, (ii) inconsistencies in object boundaries, and (iii) difficulties with top-down instance segmentation. To address these challenges, we propose InstanceGaussian, a method that jointly learns appearance and semantic features while adaptively aggregating instances. Our contributions are as follows: (i) a new Semantic-Scaffold-GS representation to improve feature representation and boundary delineation, (ii) a progressive training strategy for enhanced stability and segmentation, and (iii) a category-agnostic, bottom-up instance aggregation approach for better segmentation. Experimental results demonstrate that our approach achieves state-of-the-art performance in category-agnostic, open-vocabulary 3D point-level segmentation, validating the effectiveness of our proposed method. Project page: https://lhj-git.github.io/InstanceGaussian/ + +# 1. Introduction + +3D scene understanding aims to comprehensively capture a scene's content, structure, and semantics, addressing tasks like instance segmentation, object detection, and open-vocabulary queries. This is crucial for applications in autonomous driving, robotics, and augmented reality, where detailed scene comprehension can significantly improve performance and safety. Traditional 3D representations, + +such as voxels, point clouds, and meshes, provide structured models but often involve trade-offs between spatial resolution and computational efficiency. 3D Gaussian Splating (3DGS) [16] offers a novel approach that combines the explicitness of traditional representations with neural adaptability, enabling efficient and detailed scene modeling. 3DGS is particularly promising for advancing 3D scene understanding by effectively balancing quality and efficiency in capturing complex scene geometry. + +In the realm of 3D scene understanding utilizing 3DGS, three critical challenges emerge. 1) The first is an inherent imbalance in the representation of appearance versus semantics. Capturing the fine-grained texture details of a given object or region necessitates multiple Gaussians, each with varied appearance attributes. However, these Gaussians collectively require only a single, shared semantic attribute to accurately represent the semantics of the region. This imbalance results in a scenario where the number of Gaussians may be sufficient for appearance representation yet redundant for semantic expression. 2) The second challenge is an inconsistency between appearance and semantics. Without semantic constraints, i.e., in pure appearance reconstruction, a single Gaussian can represent different objects or regions. At object boundaries, for instance, a single Gaussian may simultaneously represent both the object's foreground and background. Recent approaches, such as language-embedded Gaussian Splatting techniques [25, 34], employ a decoupled learning strategy that overlooks the interdependence between color and semantics, leading to inconsistencies between appearance and semantics and posing significant challenges for both 3D point segmentation and 2D image segmentation. 3) Thirdly, previous approaches are predominantly designed in a top-down manner, often relying on prior category information. For instance, GaussianGrouping [38] defines the number of instances based on 2D tracking results, FastLGS [14] uses + +the count of matches in cross-view settings to determine the number of objects, and OpenGaussian [34] relies on a pre-defined number of codebook entries. Such methods are prone to issues arising from uneven category distributions, often resulting in over-segmentation or under-segmentation when handling fine-grained instances in complex scenes. + +To address these challenges, we propose InstanceGaussian, a method capable of jointly learning object appearance and instance features while adaptively aggregating objects. Our contributions are summarized as follows: + +1. Semantic-Scaffold-GS Representation: We introduce a novel representation method that balances appearance and semantics. By more flexibly allocating the semantic and appearance attributes of Gaussians, this method enables the learning of more accurate object geometric boundaries and yields improved feature representations. +2. Progressive Appearance-Semantic Joint Training Strategy: We propose a progressive training strategy that optimizes the joint representation of appearance and semantics* in a step-by-step manner. This ensures their consistency throughout the training process, thereby enhancing the model's stability and improving the accuracy of downstream tasks such as 3D point segmentation and 2D image segmentation. +3. Bottom-Up, Category-Agnostic Instance Aggregation: To further enhance the capability of 3D instance segmentation, we introduce a bottom-up, category-agnostic instance aggregation method. This method employs a clustering algorithm based on farthest point sampling and connected component analysis, effectively avoiding over-segmentation or under-segmentation. + +# 2. Related Work + +# 2.1.3D Representation + +3D representation involves the digital description and modeling of objects and scenes in three-dimensional space using techniques such as voxels, point clouds, meshes, signed distance functions, and neural representations. Neural representations have made substantial progress with techniques like NeRF [22], which has notably enhanced the quality of novel view synthesis through learning-based optimization. Various methods [2-4, 15] have been proposed to further enhance NeRF's expressive capabilities, but they often face challenges with slow training and rendering speeds. To address these issues, alternative approaches focusing on explicit representations have been developed, which have been effective in reducing the computational burden and decreasing training and inference times. These methods include voxels [10, 27, 30], hash grids [23] and point + +clouds [1, 28, 31]. + +3D Gaussian Splitting (3DGS) [16] is an innovative representation technique for modeling and rendering 3D scenes using Gaussian distributions. In recent advancements [11, 12, 20, 21, 36, 40], Scaffold-GS [20] forms a hierarchical and region-aware scene representation that respects the underlying geometric structure of the scene. This strategy exploits scene structure to guide and constrain the distribution of 3D Gaussians, allowing them to locally adapt to varying viewing angles and distances. This innovation allows for enhanced adaptation to various viewpoints and distances while preserving the scene's underlying structure, showcasing the ongoing advancements in 3DGS and its role in shaping the future of neural rendering. + +# 2.2. 3D Perception via 2D-3D Correlation + +Recent advancements in 3D perception have seen the integration of 2D Vision-Language Models (VLMs) with 3D point cloud processing, leading to significant progress in the field [13, 32, 33, 35, 42]. These methods primarily focus on aligning features and projecting 3D data into 2D to enhance zero-shot learning capabilities. Existing approaches that exploit the 2D-3D correlation can be broadly categorized into two types: one leverages feature information from 2D images, while the other utilizes segmentation masks derived from 2D images. + +Feature Distillation. The significant advancements in 2D scene understanding, initiated by SAM [18] and its variants, have inspired integrating semantic features into NeRF. Various methods have been developed to incorporate semantic features from models like CLIP [26] and DINO [5] into NeRF, enhancing 3D segmentation, understanding, and editing tasks. LeRF [17] distills features from accessible VLMs such as CLIP into a 3D scene represented by NeRF. Additionally, [19] presents a 3D open-vocabulary segmentation pipeline using NeRF. Within the 3DGS framework, LEGaussians [29] introduces uncertainty and semantic feature attributes to each Gaussian, rendering a semantic map with corresponding uncertainties. This map is compared with quantized CLIP and DINO dense features extracted from the ground truth image. LangSplat [25] employs a scene-wise language autoencoder to learn language features in the scene-specific latent space, effectively discerning clear boundaries between objects in rendered feature images. Feature3DGS [41] distills SAM's encoder features into 3D Gaussians and employs SAM's decoder to interpret 2D rendered feature maps for segmentation. However, feature distillation-based approaches face challenges such as high training costs, significant storage requirements, and slow inference times, limiting their practical application. + +2D Mask Lifting. Recent works have explored lifting 2D segmentation masks into 3D space [6, 7, 24, 34, 37-39]. SAI3D [39] and Open3DIS [24] propose methods + +![](images/de6857c57f26014baee394b6ea4a63791380e0d3275d1dc47ca667c06d7d9b04.jpg) +Figure 1. Top row: Appearance-semantic joint Gaussian representation avoids the imbalance and inconsistency in appearance-semantic learning. Bottom row: Bottom-up instantiation: Over-segmentation is achieved via FPS sampling and clustering, followed by instantiation through graph-connectivity-based aggregation. + +for merging 3D superpoints [9] guided by predictions from SAM. MaskClustering [37] introduces a mask graph clustering approach based on the view consensus rate, targeting open-vocabulary 3D instance segmentation. To achieve the 2D mask consistency across views, GaussianGrouping [38] performs simultaneous reconstruction and segmentation of open-world 3D objects, guided by 2D mask predictions obtained from SAM and 3D spatial consistency constraints. SAGA [6] utilizes a contrastive learning approach with masks generated by SAM. It maps SAM's features into a low-dimensional space using a trainable MLP, replicating these features to address inconsistency issues. Click-Gaussian [7] introduces an efficient method for interactive segmentation of 3D Gaussians. It uses two-level granularity feature fields from 2D masks and incorporates Global Feature-guided Learning (GFL) to resolve cross-view inconsistencies. OpenGaussian [34] employs SAM masks to train instance features with 3D consistency, proposing a two-stage codebook for feature discretization and an instance-level 3D-2D feature association method that links 3D points to 2D masks and CLIP features, demonstrating effectiveness in open vocabulary-based 3D object selection and understanding. However, OpenGaussian uses a fixed codebook entry count and does not fully exploit the continuity of 3D space, which constrains its performance. + +# 3. Method + +# 3.1. Preliminary + +3D Gaussian Splatting [16]. Given a set of $n$ 3D Gaussians $\mathcal{G}$ , each with center $\mu$ , color $\pmb{c}$ , opacity $\sigma$ , rotation $\pmb{R}$ and scale $\pmb{S}$ , they are rendered as an image $\pmb{I}_{ren}$ at a specific viewpoint via differentiable rasterization $\mathcal{D}$ . The reconstruction loss $\mathcal{L}_{rgb}$ is then calculated by comparing + +$I_{ren}$ with the GT image $I_{gt}$ . This process can be simplified as shown below, where blue variables are estimizable parameters. + +$$ +\mathcal {G} (\boldsymbol {\mu}, \boldsymbol {c}, \boldsymbol {\sigma}, \boldsymbol {R}, \boldsymbol {S}) _ {n} \xrightarrow {\mathcal {D}} \boldsymbol {I} _ {r e n} \longrightarrow \mathcal {L} _ {r g b} \longleftarrow \boldsymbol {I} _ {g t} +$$ + +OpenGaussian [34]. Based on 3DGS, OpenGaussian adds an instance feature $\pmb{f} \in \mathbb{R}^{6}$ for each Gaussian, passes it through differentiable rasterization $\mathcal{D}$ to render feature map $M \in \mathbb{R}^{H \times W \times 6}$ , and computes intra-mask smoothness loss $\mathcal{L}_{s}$ and inter-mask contrastive loss $\mathcal{L}_{c}$ with SAM masks $B \in \mathbb{R}^{H \times W \times 1}$ . This process simplifies as follows, where only the feature $\pmb{f}$ is optimized while other appearance attributes are frozen (shown in gray). + +$$ +\mathcal {G} (\mu , c, \sigma , R, S, f) _ {n} \xrightarrow {\mathcal {D}} M \longrightarrow \mathcal {L} _ {s}, \mathcal {L} _ {c} \longleftarrow B +$$ + +$$ +\mathcal {L} _ {s} = \sum_ {i = 1} ^ {m} \sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} \boldsymbol {B} _ {i, h, w} \cdot \left\| \boldsymbol {M} _ {:, h, w} - \bar {\boldsymbol {M}} _ {i} \right\| ^ {2}, \tag {1} +$$ + +$$ +\mathcal {L} _ {c} = \frac {1}{m (m - 1)} \sum_ {i = 1} ^ {m} \sum_ {j = 1, j \neq i} ^ {m} \frac {1}{\left\| \bar {\boldsymbol {M}} _ {i} - \bar {\boldsymbol {M}} _ {j} \right\| ^ {2}}, \tag {2} +$$ + +where $H, W$ represents the image size and $m$ corresponds to the number of SAM masks in the current view. The $\mathcal{L}_s$ aims to make each pixel feature $M_{:,h,w}$ converge towards the average feature $\bar{M}_i$ of its associated mask for intramask smoothness. The $\mathcal{L}_c$ seeks to maximize the separation between the average features $\bar{M}_i$ , $\bar{M}_j$ of different masks. + +Scaffold-GS [20]. It derives $n$ neural Gaussian points $\mathcal{G}$ from $n^{\prime}$ anchor points $\mathcal{A}$ ( $n = 10n^{\prime}$ in their implementation). The Gaussian centers $\pmb{\mu}$ are obtained by shifting its anchor centers $\pmb{\mu}^{\prime}$ by an offset $\pmb{o}$ , while other appearance attributes are decoded from the appearance embedding $\pmb{e}$ through MLPs. As shown below, the blue parts are learn + +![](images/839492095401407be55dc1fbbc448e1d60a616576b5e911b29dc25eec1fead63.jpg) +Figure 2. Progressive appearance-semantic joint training. (a) Train appearance only; (b) Independent appearance-semantic training; (c) Joint appearance-semantic training. + +able parameters, and their optimization process is similar to 3DGS. + +$$ +\begin{array}{c} \overbrace {\mathcal {A} (\boldsymbol {\mu} ^ {\prime} , \boldsymbol {e}) _ {n ^ {\prime}}} ^ {+ o} \quad \mathcal {G} (\vec {\boldsymbol {\mu}}, \underbrace {\boldsymbol {c} , \boldsymbol {\sigma} , \boldsymbol {R} , \boldsymbol {S}} _ {\mathrm {M L P s}}) _ {n} \xrightarrow {\mathcal {D}} \boldsymbol {I} _ {\mathrm {r e n}} \to \mathcal {L} _ {\mathrm {r g b}} \leftarrow \boldsymbol {I} _ {\mathrm {g t}} \\ \end{array} +$$ + +# 3.2. Appearance-semantic joint Gaussian Representation + +(1) Appearance-semantic balanced representation. For the same object or region, the texture details may vary while the semantics remain consistent. Consequently, many Gaussians with diverse appearance attributes are required to represent the texture details of the region. These Gaussians require only a single semantic attribute to effectively capture the region's semantics, meaning the same number of Gaussians may under-represent appearance while overrepresenting semantics. We refer to this phenomenon as the imbalance in the representation of appearance and semantics by Gaussians. To this end, we propose a Semantic-Scaffold-GS representation to balance appearance and semantics: + +$$ +\overbrace {\underbrace {\mathcal {A} (\boldsymbol {\mu} ^ {\prime} , \boldsymbol {e} , \boldsymbol {f}) _ {n ^ {\prime}}} _ {\text {M L P s}} \overbrace {\mathcal {G} (\boldsymbol {\mu} , \boldsymbol {c} , \boldsymbol {\sigma} , \boldsymbol {R} , \boldsymbol {S} , \boldsymbol {\tilde {f}}) _ {n}} ^ {\text {s h a r e}}} ^ {+ o} \xrightarrow {\mathcal {I} _ {\text {r e n}},} \underset {M} {\longrightarrow} \overset {\mathcal {L} _ {\text {r g b}},} {\mathcal {L} _ {s}, \mathcal {L} _ {c}} \leftarrow \overset {\boldsymbol {I} _ {\text {g t}},} {\boldsymbol {B}} +$$ + +The blue parts are the estimizable parameters; see Sec. 3.1 for variable definitions. Similar to Scaffold-GS, we derive $n$ neural Gaussian points $\mathcal{G}$ from $n'$ anchor points $\mathcal{A}$ ( $n = 5n'$ in our implementation). Specifically, each group of 5 child Gaussians with different appearance attributes is decoded from the appearance embedding $e$ of their common parent anchor point using MLPs. However, at the semantic level, these 5 child Gaussians share the same instance feature $f$ as their parent anchor point. On the one hand, this representation strengthens the learning of appearance, especially in scenes represented by sparse points (e.g., ScanNet and SLAM-reconstructed scenes); on the other hand, sharing features among neighboring points significantly enhances the efficiency of semantic learning. The process is described in the top row of Fig. 1. + +(2) Appearance-semantic consistency learning. Without semantic constraints, i.e., pure appearance reconstruc + +tion, a Gaussian can represent across objects/regions. For instance, at boundaries, a single Gaussian may simultaneously represent both the object and the background. Recent language-embedded Gaussian Splatting works [25, 34] follow a paradigm of pre-training appearance reconstruction, freezing the appearance properties, and subsequently learning semantic attributes. This tends to cause inconsistencies between appearance and semantics, posing challenges for 3D point segmentation and 2D image segmentation. Therefore, we propose a progressive appearance-seismic joint training strategy to ensure their consistency as shown in Fig. 2: during the 0-10k steps, only the appearance attributes are trained, then appearance and semantics are trained independently in the 10k-20k steps, and finally, appearance and semantics are jointly trained in the 20k-30k steps. Here, we employ the loss (Eq. (1) (2)) of OpenGaussian to train instance features, but its inter-mask contrastive loss value is too large, causing instability of joint training. We overcome this problem through the following improvements: + +$$ +\mathcal {L} _ {c} = \frac {1}{m (m - 1)} \sum_ {i = 1} ^ {m} \sum_ {j = 1, j \neq i} ^ {m} \frac {\mathbf {1} _ {\| \bar {M} _ {i} - \bar {M} _ {j} \| ^ {2} < \tau}}{\left\| \bar {M} _ {i} - \bar {M} _ {j} \right\| ^ {2}}, \tag {3} +$$ + +where $\mathbb{1}_{\| \bar{M}_i - \bar{M}_j\|^2 < \tau}$ takes values of 1 or 0, aiming to truncate the loss between two objects whose dissimilarity (quantified by the L2 distance) exceeds a threshold $\tau$ (set to 0.4 in our implementation). This improvement is intuitive because there is no need to enforce that the features of two different objects are infinitely distant from each other. + +# 3.3. Bottom-up Category-Agnostic Instantiation + +We employ a bottom-up method for category-agnostic instantiation. Specifically, we first over-segment the scene into numerous sub-objects. Then, we construct a graph of these sub-objects and perform aggregation based on the graph connectivity, adaptively obtaining several independent and complete objects. The bottom row of Fig. 1 describes this process. + +(1) Over-segmentation. Inspired by point cloud perception methods, which utilize farthest point sampling (FPS) algorithms to sample seed points for candidate proposals from dense point clouds, we achieve over-segmentation of the scene through FPS-based clustering. + +We first sample $s$ seed points from all $n$ Gaussians through FPS and use them as the initial centers of $s$ clusters, where $s$ is much greater than the number of objects in the scene, with a default value of 1000. Then we start the clustering process, updating the instance feature of all Gaussians and the $s$ clusters. This process can be formulated as: + +$$ +\bar {f}, \bar {\mu}, \bar {\mathbf {I}} = \phi (\mathbf {X}), \tag {4} +$$ + +$$ +\mathbf {X} = \left[ \operatorname {P E} (\boldsymbol {\mu}) \in \mathbb {R} ^ {n \times d}; \boldsymbol {f} \in \mathbb {R} ^ {n \times 6} \right], \tag {5} +$$ + +Algorithm 1 Bottom-up category-agnostic instantiation +Input: $\mu, f$ - point position and feature, $s$ - the number of over-segmented sub-objects, $r$ -voxel resolution, $\gamma$ - connectivity threshold. + +Output: well-segmented complete object $\hat{\mathbf{l}},\hat{\mathbf{f}}$ + +1: procedure OVERSEGMENTATION $(\mu, f, s, r, \gamma)$ +2: $\bar{f},\bar{\mu}\gets FPS(\pmb {f},\pmb {\mu},s)$ ▷initialize s centers +3: $\mathbf{X} \gets \operatorname{concat} PE(\boldsymbol{\mu})$ and $\pmb{f}$ +4: $\overline{\mathbf{I}},\overline{\pmb{f}},\overline{\pmb{\mu}}\gets k - means(\mathbf{X},s)$ $\triangleright$ Eq. (4) +5: return SUBOBJECTAGG $(\mu, \overline{\mathbf{I}}, \overline{\mathbf{f}}, s, r, \gamma)$ +6: end procedure +7: procedure SUBOBJECTAGG $(\mu, \overline{\mathbf{I}}, \overline{f}, s, r, \gamma)$ +8: $\mathbf{G} \in \mathbb{R}^{s \times s} \gets$ initialize the graph with zeros +9: $\mathbf{V}\mathbf{o}\mathbf{x} \gets \text{Voxelization}(\boldsymbol{\mu},\bar{\mathbf{I}},r) \quad \triangleright$ sub-object. voxels +10: for $i$ in $[0,s)$ do +11: for $j$ in $[0,s)$ do +12: $dis(i,j)\gets L2distance(\bar{\pmb{f}}_i,\bar{\pmb{f}}_j)$ +13: $\mathbb{1}(i,j)\gets near(\mathbf{Vox}_i,\mathbf{Vox}_j)$ +14: $\mathbf{G}_{ij}\gets dis(i,j)\cdot \mathbb{1}(i,j)$ $\triangleright$ Eq. (6) +15: end for +16: end for +17: $\hat{\mathbf{l}},\hat{\pmb{f}}\gets Aggrecation(\mathbf{G},\gamma)$ +18: return $\bar{\mathbf{I}},\hat{f}$ +19: end procedure + +where $\phi(\cdot)$ denotes the k-means, $\mathbf{X} \in \mathbb{R}^{n \times (d + 6)}$ is the attribute of the $n$ Gaussians participating in the clustering, which is the concatenation of the positional encoding $(\mathrm{PE}(\cdot))$ for position $\pmb{\mu}$ and the instance feature $\pmb{f}$ . Finally, $s$ sub-object clusters are obtained, $\bar{\pmb{f}} \in \mathbb{R}^{s \times 6}$ is their feature, $\bar{\pmb{\mu}} \in \mathbb{R}^{s \times 3}$ is their centers, $\bar{\mathbf{I}} \in \mathbb{Z}_{[0,s]}^n$ represents the label indicating which subobject each point belongs to. During this process, only the feature $\pmb{f}$ needs to be optimized, while the position $\pmb{\mu}$ is frozen. This process is shown in lines 1-4 of Alg. 1. + +(2) Graph connectivity-based sub-objects aggregation. This process is shown in lines 7-19 of Alg. 1. We use the $s$ sub-objects obtained from the over-segmentation step as nodes to construct a graph and aggregate the sub-objects into complete objects by evaluating node connectivity. This graph can be represented by an $s \times s$ adjacency matrix $\mathbf{G}$ . $\mathbf{G}_{ij}$ represents the connectivity (edge) between the $i$ -th and $j$ -th sub-objects (nodes): + +$$ +\mathbf {G} _ {i j} = \left\| \bar {\boldsymbol {f}} _ {i} - \bar {\boldsymbol {f}} _ {j} \right\| _ {2} \cdot \mathbb {1} (i, j), \tag {6} +$$ + +where we design two connectivity criteria: 1) the L2 distance of the node features, a closer distance indicates that the two nodes are more likely to be merged; 2) the spatial relationship between the nodes is represented by the indicator function $\mathbb{1}(\cdot)$ , which determines whether two subobjects are adjacent or share voxels by voxelizing the 3D space with a resolution $r$ , and returns 0 or 1. + +After constructing the graph, we perform aggregation on the nodes with connectivity values $\mathbf{G}_{ij}$ in the range $(0,\gamma]$ using the standard graph connected component algorithm (such as breadth-first search), where $\gamma$ is set to 0.1 in our implementation. Due to the indicator function, our graph is sparse and can be accelerated through the union-find method. Finally, by aggregating the feature-spatial similar sub-objects, we adaptively obtain several $(e.g.,m)$ independent and complete objects with $\hat{\pmb{f}}\in \mathbb{R}^{m\times 6}$ representing their feature, $\hat{\mathbf{I}}\in \mathbb{Z}_{[0,m)}^n$ representing point label. + +# 4. Experiments* + +# 4.1. Category-Agnostic 3D Instance Segmentation + +Settings. 1) Task: Category-agnostic 3D instance segmentation aims to segment independent 3D objects/instances in a given point cloud (Gaussians can naturally be treated as points) scene without considering their semantic categories, ensuring that each point ultimately belongs to only one instance. 2) Dataset and Metrics: Following OpenGaussian, we evaluate using the ScanNet [8] dataset with GT instance-level annotations. More specifically, we calculate the instance-level mIoU and mAcc metrics for the 10 scenes selected by OpenGaussian. 3) Baseline: We mainly compare with Gaussian-based methods OpenGaussian [34] and GaussianGrouping [38] with RGB images and SAM masks as input. Additionally, we provide the performance of the depth-assisted back-projection method MaskClustering [37] as a reference. + +Results. Quantitative results, as shown in Tab. 1, demonstrate that InstanceGaussian significantly outperforms Gaussian-based methods and achieves comparable results to MaskClustering which uses depth to assist. GaussianGrouping relies on a tracking-based model for cross-view object association, which is prone to tracking failures in complex scenes like ScanNet, characterized by significant viewpoint changes and occlusions, consequently leading to lower performance. Although OpenGaussian achieves instance feature learning with cross-view 3D consistency by the smooth-contrast losses, avoiding explicit tracking-based association, its performance bottleneck lies in the brute-force, naive $k$ -means clustering, lacking parameter adaptability. Our method enables more discriminative feature learning through appearance-semantic joint representation and training, achieving more accurate and generalizable 3D instance segmentation through oversegmentation followed by connectivity-based adaptive aggregation. The visualization results are shown in Fig. 3, where the well-segmented instances faithfully reflect the performance advantages of InstanceGaussian. + +![](images/3af07ffd268086c9bf3e0506605d324885af5b91db9534728775974547d20ab7.jpg) +Reference Mesh + +![](images/f2f18bb75e9419f29346ee470695f8a7f01309d89abda4cb53d6d9469fe3d876.jpg) + +![](images/1c7e1330adcaa7ed5e4161e5048e0194c5cad4b33eabd7e4fe4d909cf6b29569.jpg) + +![](images/7c13b28042ac9c66b8a675bfe0ec516eaa2d8ad0fb5d5a3cb3aca0ebc5a3b446.jpg) +Figure 3. Visualization comparison of category-agnostic 3D instance segmentation result. InstanceGaussian outperforms OpenGaussian and GaussinGrouping in accurately distinguishing different 3D objects. + +![](images/7263f0a7df430b451156408669dbd06d1d6754a414c9447adb9bd0a4cb78e539.jpg) +Gaussian Grouping + +![](images/539b2081e77fa5c461b170b022908583f25cf08c7a8435819a65baeadee21c55.jpg) + +![](images/249c27bd3b1c7b91d5eb4536319bcdcad1100ac3cddf671a5061ba56c459e4f3.jpg) + +![](images/4ecbf7b5f7d3c5fab2a1d654f3a67c324c7e44a72c5fe5bae8e456c788f9c81c.jpg) + +![](images/7de0dfaef2c03219e00d7e4b8fe5c32c65d4741c4404a170d4d10e950c599f57.jpg) +OpenGaussian + +![](images/cf9277f340f954ce4fcd9f8ee5f46b58c4ff7f09106c6e8958651cf1f67eb172.jpg) + +![](images/c4d2b95b9b52ef4a8ef9ff0a19667bbfdd84114b63e7ede518013cb838aff075.jpg) + +![](images/bf146717520e69aa1ba5a9f450d6228cb449ae0f67fce93bd5deed32db76586a.jpg) + +![](images/4cd182aa0de18c5cb4b32ac7e8070f296a0ca01a0511ec0c924124ec4d9b5157.jpg) +InstanceGaussian + +![](images/78ec5a7c517262a98b088049bb56585829fdce2b6c46ff0ee704d32148b69515.jpg) + +![](images/400cfa6eb70c179ebec5fcb175f9d5e359078958962f695a92dc91a59459db9e.jpg) + +![](images/2afbd7a9d6401a61eee41357a6206de00336e036ed86215fe8780b18ebe29d54.jpg) + +![](images/87bdb03b1f6d06b757198b57013ebf80d46965ebcc59cbb9e83b55cc51ed39a0.jpg) +Ground Truth + +![](images/5355a039ad2bde57f15b3d4d8a6e778da8582e7286d6ad2728182351813fa1d5.jpg) + +![](images/82a5ce080e1d9b939af1a284eead8c1843f05d1ba126ece630e296b36ee1057e.jpg) + +![](images/54e3e10d6e4eeb3b8ff7c46d959d37c37cc1f9cf23c1bd97f25c01139d643d05.jpg) + +
TypeMethodsmIoU ↑mAcc. ↑
Depth-aidedMaskClustering [37]54.1780.95
Gaussian-basedGaussianGrouping [38]22.5530.54
OpenGaussian [34]27.3252.44
InstanceGaussian (ours)50.2780.22
+ +# 4.2. Open-Vocabulary Query Point Cloud Understanding + +After 3D instantiation as mentioned in Sec. 3.3, we adopt the 2D mask-3D instance association method proposed by OpenGaussian to associate each 3D instance with a 512-dim CLIP image feature. This allows us to support text-based understanding by extracting 512-dim CLIP text features and calculating cosine similarity. + +Settings. 1) Task: This task takes text as input and finds the matching point cloud by calculating the cosine similarity between text features and point features. In theory, it supports arbitrary open vocabulary input, but to compute quantitative metrics with the annotated GT point cloud, we use text that corresponds to the pre-defined categories as queries. 2) Dataset and Metrics: Following OpenGaussian, we use 19, 15, and 10 categories from ScanNet as text queries, assigning the closest text to the point cloud to compute the mIoU and mAcc for each category. 3) Baseline: We mainly compare with the recent Gaussian-based meth + +Table 1. Performance of category-agnostic 3D instance segmentation on the Scannet dataset. Accurate is measured by mAcc@0.25. + +
TypeMethods19 classes15 classes10 classes
mIoUmAcc.mIoUmAcc.mIoUmAcc.
DepthMaskClustering [37]34.5954.3234.0045.0738.1056.74
Gaussian-basedLangSplat [25]3.789.115.3513.208.4022.06
LEGaussians [29]3.8410.879.0122.2212.8228.62
OpenGaussian [34]24.7341.5430.1348.2538.2955.19
InstanceGaussian40.6654.0142.5159.1547.9464.01
+ +Table 2. Performance of open vocabulary semantic segmentation on the Scannet dataset. Accurate is measured by mAcc@0.25. + +ods LangSplat, LEGaussians, and OpenGaussian. We also provide the performance of a SOTA method MaskClustering utilizing depth information to assist back projection as a comparison. + +Results. The quantitative results are shown in Tab. 2, where InstanceGaussian achieves SOTA performance under the settings of 19, 15, and 10 categories, significantly outperforming Gaussian-based methods. The poor performance of LangSplat and LEGaussians stems from ambiguous Gaussian feature learning and detrimental language feature distillation. The performance of OpenGaussian is limited by its coupled appearance-semantic representation and fixed cluster count. GaussianGrouping is not included in the comparison because it does not support 3D-level text queries. Surprisingly, our method even surpasses the depth-assisted MaskClustering, although it outperforms our method in the category-agnostic instance segmentation task. We analyze that the reason lies in its heavy reliance on high-quality SAM masks. Fig. 4 shows the results of relevance point cloud retrieval using text queries, in which + +![](images/7b18f8ec1534019b3a592b2618ff47398f4ee7fe7a7deaafa68ee68863bf8b43.jpg) +Reference mesh + +![](images/a34e5d21ec986b8b67087ee9a99191b021ab5aefe10acfa8e8934de43bad8399.jpg) + +![](images/f1310c2b48c0b0782144e0f2c3d1a5f32433f2599d3ed449450917a4e807957a.jpg) + +![](images/9acc604a453ab387d82045f7251341967cd78a1f2fb6def3a12dcd3b94dc425b.jpg) + +![](images/cb6adc43521290ee44b41898aeb7dd644976ac076d1c6716fe1a3e6b9dcf4cfa.jpg) + +![](images/9cdc8b599f3e3c45bc64cfd1ee313adc71f5f37ed0b4901a955e901acec22eba.jpg) + +![](images/8dc98c7b5d9042a590bcbb2037e8df7751ef62faa2cc9e8970c8dd638018e12a.jpg) +Figure 4. Open-vocabulary query point cloud Understanding on ScanNet dataset. InstanceGaussian shows advanced text query capabilities. + +![](images/b556a074f72f0b3286ab90d5c1edf380d369594dd46c56d6a01a13213907d801.jpg) + +![](images/0d99224d7a64178e9deae598d2d5764272cec06b9c02ec221d2a44e1e917d148.jpg) + +![](images/405581978adba94bba124a188845a900727c01bd4acce8a26f477caf9917ed68.jpg) + +![](images/011ede4acb607a7cb30305db3037ad7f7a990d34222686b64a1761ab647f55bd.jpg) + +![](images/8576814350eab78109db68e8929182e8bc3b670a3459f273111e6e65e571f333.jpg) + +
MethodsmIoU ↑mAcc. ↑
LangSplat [25]9.6612.41
LEGaussians [29]16.2123.82
OpenGaussian [34]38.3651.34
InstanceGaussian (ours)45.3058.44
+ +both common categories (such as "bed", "curtain") are used as queries, and the ability to understand open-vocabulary (such as "ride", "music") is also demonstrated. + +# 4.3. Open Vocabulary 3D Object Selection and Rendering + +Settings. 1) Task: Similar to Sec. 4.2, we first use open vocabulary text to select matching 3D Gaussians, and then render the 3D Gaussians into multi-view 2D images through the rasterization pipeline of 3DGS. 2) Dataset and Metrics: Following LangSplat and OpenGaussian, we use the LeRF [17] dataset for evaluation, and after rendering the selected 3D objects into 2D images, we calculate the mIoU and mAcc between the rendered 2D objects and the GT 2D target objects. 3) Baseline: Only Gaussian-based methods possess 3D point-level perception and rendering capabilities, thus we only compare against Gaussian-based methods Langsplat, Legaussian, and OpenGaussian. + +Result. Quantitative results are displayed in the Tab. 3, which demonstrates that even when rendering 3D Gaussians onto 2D images for evaluation. As previously analyzed, LangSplat and LEGaussian, due to their poor 3D perception capabilities, fail to accurately select 3D Gaussians relevant to the text query, resulting in low-quality rendering results. The performance advantage of our method stems from two aspects. First, the bottom-up adaptive instantiation ensures good discrimination between objects. Second, the joint representation and learning of appearance and semantics avoid the appearance-semantic inconsistency present in OpenGaussian. The qualitative results in Fig. 5 more intuitively demonstrate this point. Our method renders objects with clearer boundaries, unlike OpenGaussian, + +Table 3. Performance of open vocabulary 3D object selection and rendering on Lerf dataset. Accurate is measured by mAcc@0.25. + +
CaseJotin-repres.Joint-trainSematic Seg.Instance Seg.
#1✓30.7147.40
#2✓33.1549.57
#3✓✓40.6650.27
+ +Table 4. Performance of instance and semantic segmentation with different representation and training strategies on Scannet. Evaluated by mIoU. + +which due to its appearance-semantic ambiguity, requires some Gaussians to represent multiple objects/regions, especially evident in the boundary areas. + +# 4.4. Ablation Study + +# 4.4.1. Gaussian Representation and Training Strategy + +We conduct the ablation studies as shown in Tab. 4 to demonstrate the effectiveness of our Semantic-Scaffold-GS joint representation and progressive joint training strategy. The experiments are evaluated using the semantic and instance segmentation tasks on the ScanNet dataset. The joint representation achieves a balance between appearance and semantics, allowing the same anchor point to express different appearances while sharing the same semantics. The progressive joint training avoids inconsistencies in appearance and semantics, enhances the distinguishability between objects, and thus leads to improved instance segmentation and semantic understanding performance. + +Fig. 6 more intuitively demonstrates our avoidance of inconsistency. The first and second rows respectively visualize the appearance color and instance features of the 3D Gaussian ellipsoid in the 3D viewer, where the ellipse represents the boundary of the Gaussian. Fig. 6 (a)(c) shows the results of our joint training, which achieve clearer and more distinctive expressions in both appearance and semantics compared to Fig. 6 (b)(d). This improvement not only brings benefits to the understanding of the 3D level (i.e., the semantic level, as proven by our quantitative ablation studies), but also enhances the quality of object rendering (i.e., the appearance level, as shown in Sec. 4.3). + +![](images/301ba5096efc9a3ca427731558995e14442424ead17651bb1a5d240f725647ee.jpg) +Figure 5. Open-vocabulary 3D object selection and rendering on the LeRF dataset. InstanceGaussian outperforms OpenGaussian in accurately identifying the objects' boundaries by text queries. + +![](images/dc587546c2592cda3251bc002d4a9c2713bc44890d43879d9e21f0a79aa153dc.jpg) +Figure 6. Visualization results of the appearance color and instance features of the Gaussians ellipsoid in the 3D viewer demonstrate that our method achieves superior appearance-semantic consistency. (a)(c): Results of progressive joint training. (b)(d): Without joint training. + +
CasefeaturevoxelizeSemantic Seg.Instance Seg.
#1✓21.7527.50
#2✓28.9843.41
#3✓✓40.6650.27
+ +Table 5. Performance of instance and semantic segmentation with different connectivity condition strategies. Evaluated by mIoU. + +# 4.4.2. Aggregation Condition + +In the process of aggregating over-segmented sub-objects based on a connectivity graph, we considered two connectivity conditions: the L2 distance of instance features, and the proximity judgment after voxelization. + +We conduct an ablation study on its configuration as shown in Tab. 5. In case #1, we use the color of the Gaussians as a feature and consider their positions for aggregation, which results in poor performance. In case #2, we only consider instance features and achieve a significant performance improvement, demonstrating the effectiveness of our feature learning. In case #3, we take both conditions into account and achieve the best performance, which is an intuitive idea that in aggregation, we need to consider both semantic similarity and spatial proximity. + +# 5. Conclusion and Limitation + +In this paper, we introduce an appearance-semantic compact Gaussian representation to address the imbalance between appearance and instance features using Gaussians. On one hand, this representation enhances the learning of appearance; on the other hand, sharing features among neighboring points significantly boosts the efficiency of semantic learning. The bottom-up category-agnostic instantiation adaptively determines the number of categories, cleverly avoiding the issue of over-segmentation or under-segmentation. In category-agnostic segmentation tasks, we significantly outperform previous models, achieving state-of-the-art (SOTA) performance in open vocabulary understanding for 3D point-level tasks. + +Despite achieving novel performance on multiple tasks, our method still has the following limitations: 1) Longer training time for slower rendering, since the appearance attributes need to be calculated through MLP. 2) Segmentation accuracy is still affected when most of the images in the scene are incorrectly segmented by SAM. Failure case analyses can be found in the supplementary materials. We hope those drawbacks can be addressed in future works. + +# References + +[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conference on Computer Vision (ECCV), 2020. 2 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), 2021. 2 +[6] Jiazhong Cen, Jiemin Fang, Chen Yang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Segment any 3d gaussians. arXiv preprint arXiv:2312.00860, 2023. 2, 3 +[7] Seokhun Choi, Hyeonseop Song, Jaechul Kim, Taehyeong Kim, and Hoseok Do. Click-gaussian: Interactive segmentation to any 3d gaussians. In European Conference on Computer Vision (ECCV), 2024. 2, 3 +[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5 +[9] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 2004. 3 +[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qin-hong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[11] Qiankun Gao, Jiarui Meng, Chengxiang Wen, Jie Chen, and Jian Zhang. Hicom: Hierarchical coherent + +motion for dynamic streamable scenes with 3d gaussian splatting. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2 +[12] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2 +[13] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[14] Yuzhou Ji, He Zhu, Junshu Tang, Wuyi Liu, Zhizhong Zhang, Xin Tan, and Yuan Xie. Fastlgs: Speeding up language embedded gaussians with feature grid mapping. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 1 +[15] Yifan Jiang, Peter Hedman, Ben Mildenhall, Dejia Xu, Jonathan T Barron, Zhangyang Wang, and Tianfan Xue. Alignerf: High-fidelity neural radiance fields via alignment-aware training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023. 1, 2, 3 +[17] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7 +[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[19] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In NeurIPS, 2023. 2 +[20] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffolds: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3 +[21] Jiarui Meng, Haijie Li, Yanmin Wu, Qiankun Gao, Shuzhou Yang, Jian Zhang, and Siwei Ma. Mirror-3dgs: Incorporating mirror reflections into 3d gaus + +sian splatting. In 2024 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2024. 2 +[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2 +[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 2022. 2 +[24] Phuc Nguyen, Tuan Duc Ngo, Evangelos Kalogerakis, Chuang Gan, Anh Tran, Cuong Pham, and Khoi Nguyen. Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[25] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 1, 2, 4, 6, 7 +[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 2 +[27] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 2023. 2 +[28] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM TOG, 2022. 2 +[29] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 2, 6, 7 +[30] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[31] Yifan Wang, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 2019. 2 +[32] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling + +and dense alignment for 3d visual grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[33] Yanmin Wu, Qiankun Gao, Renrui Zhang, Haijie Li, and Jian Zhang. Language-assisted 3d scene understanding. IEEE Transactions on Multimedia, 2025. 2 +[34] Yanmin Wu, Jiarui Meng, Hajie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Jian Zhang. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. In Advances in Neural Information Processing Systems, volume 37, pages 19114-19138, 2024. 1, 2, 3, 4, 5, 6, 7 +[35] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[36] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[37] Mi Yan, Jiazhao Zhang, Yan Zhu, and He Wang. Maskclustering: View consensus based mask graph clustering for open-vocabulary 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5, 6 +[38] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In European Conference on Computer Vision (ECCV), 2024. 1, 3, 5, 6 +[39] Yingda Yin, Yuzheng Liu, Yang Xiao, Daniel Cohen-Or, Jingwei Huang, and Baoquan Chen. Sai3d: Segment any instance in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[40] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[41] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 + +[42] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 \ No newline at end of file diff --git a/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/images.zip b/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d2d5798f902677f8a9920e70987a765456329352 --- /dev/null +++ b/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9ea2552bb71c448ce0b367641de27a8cdbe9617f9a811345a7147afdd1ecc4f +size 696557 diff --git a/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/layout.json b/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..09fb39e976097ddcc33e55dd014ae7c0b8790346 --- /dev/null +++ b/2025/InstanceGaussian_ Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception/layout.json @@ -0,0 +1,10190 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 102, + 529, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 102, + 529, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 529, + 140 + ], + "type": "text", + "content": "InstanceGaussian: Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "spans": [ + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "text", + "content": "Haijie Li" + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "text", + "content": " Yanmin Wu" + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "text", + "content": " Jiarui Meng" + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "text", + "content": " Qiankun Gao" + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "text", + "content": " Zhiyao Zhang" + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "text", + "content": " Ronggang Wang" + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "text", + "content": " Jian Zhang" + }, + { + "bbox": [ + 123, + 159, + 487, + 190 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 194, + 536, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 194, + 536, + 251 + ], + "spans": [ + { + "bbox": [ + 73, + 194, + 536, + 251 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 194, + 536, + 251 + ], + "type": "text", + "content": "School of Electronic and Computer Engineering, Peking University, China \n" + }, + { + "bbox": [ + 73, + 194, + 536, + 251 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 194, + 536, + 251 + ], + "type": "text", + "content": "Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University \n" + }, + { + "bbox": [ + 73, + 194, + 536, + 251 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 73, + 194, + 536, + 251 + ], + "type": "text", + "content": "College of Information Science and Engineering, Northeastern University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 278, + 200, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 278, + 200, + 291 + ], + "spans": [ + { + "bbox": [ + 151, + 278, + 200, + 291 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 304, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 297, + 544 + ], + "type": "text", + "content": "3D scene understanding is vital for applications in autonomous driving, robotics, and augmented reality. However, scene understanding based on 3D Gaussian Splating faces three key challenges: (i) an imbalance between appearance and semantics, (ii) inconsistencies in object boundaries, and (iii) difficulties with top-down instance segmentation. To address these challenges, we propose InstanceGaussian, a method that jointly learns appearance and semantic features while adaptively aggregating instances. Our contributions are as follows: (i) a new Semantic-Scaffold-GS representation to improve feature representation and boundary delineation, (ii) a progressive training strategy for enhanced stability and segmentation, and (iii) a category-agnostic, bottom-up instance aggregation approach for better segmentation. Experimental results demonstrate that our approach achieves state-of-the-art performance in category-agnostic, open-vocabulary 3D point-level segmentation, validating the effectiveness of our proposed method. Project page: https://lhj-git.github.io/InstanceGaussian/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 564, + 135, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 135, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 135, + 577 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 584, + 296, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 296, + 669 + ], + "type": "text", + "content": "3D scene understanding aims to comprehensively capture a scene's content, structure, and semantics, addressing tasks like instance segmentation, object detection, and open-vocabulary queries. This is crucial for applications in autonomous driving, robotics, and augmented reality, where detailed scene comprehension can significantly improve performance and safety. Traditional 3D representations," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 279, + 555, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 388 + ], + "type": "text", + "content": "such as voxels, point clouds, and meshes, provide structured models but often involve trade-offs between spatial resolution and computational efficiency. 3D Gaussian Splating (3DGS) [16] offers a novel approach that combines the explicitness of traditional representations with neural adaptability, enabling efficient and detailed scene modeling. 3DGS is particularly promising for advancing 3D scene understanding by effectively balancing quality and efficiency in capturing complex scene geometry." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "text", + "content": "In the realm of 3D scene understanding utilizing 3DGS, three critical challenges emerge. 1) The first is an inherent imbalance in the representation of appearance versus semantics. Capturing the fine-grained texture details of a given object or region necessitates multiple Gaussians, each with varied appearance attributes. However, these Gaussians collectively require only a single, shared semantic attribute to accurately represent the semantics of the region. This imbalance results in a scenario where the number of Gaussians may be sufficient for appearance representation yet redundant for semantic expression. 2) The second challenge is an inconsistency between appearance and semantics. Without semantic constraints, i.e., in pure appearance reconstruction, a single Gaussian can represent different objects or regions. At object boundaries, for instance, a single Gaussian may simultaneously represent both the object's foreground and background. Recent approaches, such as language-embedded Gaussian Splatting techniques [25, 34], employ a decoupled learning strategy that overlooks the interdependence between color and semantics, leading to inconsistencies between appearance and semantics and posing significant challenges for both 3D point segmentation and 2D image segmentation. 3) Thirdly, previous approaches are predominantly designed in a top-down manner, often relying on prior category information. For instance, GaussianGrouping [38] defines the number of instances based on 2D tracking results, FastLGS [14] uses" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 674, + 149, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 674, + 149, + 684 + ], + "spans": [ + { + "bbox": [ + 69, + 674, + 149, + 684 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 684, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 684, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 684, + 295, + 712 + ], + "type": "text", + "content": "This work is supported by Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology(Grant No. 2024B1212010006)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14078" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "the count of matches in cross-view settings to determine the number of objects, and OpenGaussian [34] relies on a pre-defined number of codebook entries. Such methods are prone to issues arising from uneven category distributions, often resulting in over-segmentation or under-segmentation when handling fine-grained instances in complex scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 294, + 192 + ], + "type": "text", + "content": "To address these challenges, we propose InstanceGaussian, a method capable of jointly learning object appearance and instance features while adaptively aggregating objects. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 295, + 445 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 194, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 295, + 266 + ], + "type": "text", + "content": "1. Semantic-Scaffold-GS Representation: We introduce a novel representation method that balances appearance and semantics. By more flexibly allocating the semantic and appearance attributes of Gaussians, this method enables the learning of more accurate object geometric boundaries and yields improved feature representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 266, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 361 + ], + "type": "text", + "content": "2. Progressive Appearance-Semantic Joint Training Strategy: We propose a progressive training strategy that optimizes the joint representation of appearance and semantics* in a step-by-step manner. This ensures their consistency throughout the training process, thereby enhancing the model's stability and improving the accuracy of downstream tasks such as 3D point segmentation and 2D image segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 362, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 295, + 445 + ], + "type": "text", + "content": "3. Bottom-Up, Category-Agnostic Instance Aggregation: To further enhance the capability of 3D instance segmentation, we introduce a bottom-up, category-agnostic instance aggregation method. This method employs a clustering algorithm based on farthest point sampling and connected component analysis, effectively avoiding over-segmentation or under-segmentation." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 459, + 142, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 142, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 142, + 471 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 479, + 165, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 165, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 165, + 491 + ], + "type": "text", + "content": "2.1.3D Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 497, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 677 + ], + "type": "text", + "content": "3D representation involves the digital description and modeling of objects and scenes in three-dimensional space using techniques such as voxels, point clouds, meshes, signed distance functions, and neural representations. Neural representations have made substantial progress with techniques like NeRF [22], which has notably enhanced the quality of novel view synthesis through learning-based optimization. Various methods [2-4, 15] have been proposed to further enhance NeRF's expressive capabilities, but they often face challenges with slow training and rendering speeds. To address these issues, alternative approaches focusing on explicit representations have been developed, which have been effective in reducing the computational burden and decreasing training and inference times. These methods include voxels [10, 27, 30], hash grids [23] and point" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 390, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 390, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 390, + 83 + ], + "type": "text", + "content": "clouds [1, 28, 31]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 85, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 553, + 240 + ], + "type": "text", + "content": "3D Gaussian Splitting (3DGS) [16] is an innovative representation technique for modeling and rendering 3D scenes using Gaussian distributions. In recent advancements [11, 12, 20, 21, 36, 40], Scaffold-GS [20] forms a hierarchical and region-aware scene representation that respects the underlying geometric structure of the scene. This strategy exploits scene structure to guide and constrain the distribution of 3D Gaussians, allowing them to locally adapt to varying viewing angles and distances. This innovation allows for enhanced adaptation to various viewpoints and distances while preserving the scene's underlying structure, showcasing the ongoing advancements in 3DGS and its role in shaping the future of neural rendering." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 251, + 511, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 511, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 511, + 263 + ], + "type": "text", + "content": "2.2. 3D Perception via 2D-3D Correlation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 269, + 553, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 553, + 388 + ], + "type": "text", + "content": "Recent advancements in 3D perception have seen the integration of 2D Vision-Language Models (VLMs) with 3D point cloud processing, leading to significant progress in the field [13, 32, 33, 35, 42]. These methods primarily focus on aligning features and projecting 3D data into 2D to enhance zero-shot learning capabilities. Existing approaches that exploit the 2D-3D correlation can be broadly categorized into two types: one leverages feature information from 2D images, while the other utilizes segmentation masks derived from 2D images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 390, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 553, + 676 + ], + "type": "text", + "content": "Feature Distillation. The significant advancements in 2D scene understanding, initiated by SAM [18] and its variants, have inspired integrating semantic features into NeRF. Various methods have been developed to incorporate semantic features from models like CLIP [26] and DINO [5] into NeRF, enhancing 3D segmentation, understanding, and editing tasks. LeRF [17] distills features from accessible VLMs such as CLIP into a 3D scene represented by NeRF. Additionally, [19] presents a 3D open-vocabulary segmentation pipeline using NeRF. Within the 3DGS framework, LEGaussians [29] introduces uncertainty and semantic feature attributes to each Gaussian, rendering a semantic map with corresponding uncertainties. This map is compared with quantized CLIP and DINO dense features extracted from the ground truth image. LangSplat [25] employs a scene-wise language autoencoder to learn language features in the scene-specific latent space, effectively discerning clear boundaries between objects in rendered feature images. Feature3DGS [41] distills SAM's encoder features into 3D Gaussians and employs SAM's decoder to interpret 2D rendered feature maps for segmentation. However, feature distillation-based approaches face challenges such as high training costs, significant storage requirements, and slow inference times, limiting their practical application." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "2D Mask Lifting. Recent works have explored lifting 2D segmentation masks into 3D space [6, 7, 24, 34, 37-39]. SAI3D [39] and Open3DIS [24] propose methods" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "* The term \"semantic\" is used here to distinguish it from \"appearance\". In subsequent sections and code implementations, this concept will be explicitly referred to as \"instance features\"." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14079" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 526, + 255 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 526, + 255 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 526, + 255 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 526, + 255 + ], + "type": "image", + "image_path": "de6857c57f26014baee394b6ea4a63791380e0d3275d1dc47ca667c06d7d9b04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 258, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 258, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 555, + 293 + ], + "type": "text", + "content": "Figure 1. Top row: Appearance-semantic joint Gaussian representation avoids the imbalance and inconsistency in appearance-semantic learning. Bottom row: Bottom-up instantiation: Over-segmentation is achieved via FPS sampling and clustering, followed by instantiation through graph-connectivity-based aggregation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 301, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 301, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 296, + 601 + ], + "type": "text", + "content": "for merging 3D superpoints [9] guided by predictions from SAM. MaskClustering [37] introduces a mask graph clustering approach based on the view consensus rate, targeting open-vocabulary 3D instance segmentation. To achieve the 2D mask consistency across views, GaussianGrouping [38] performs simultaneous reconstruction and segmentation of open-world 3D objects, guided by 2D mask predictions obtained from SAM and 3D spatial consistency constraints. SAGA [6] utilizes a contrastive learning approach with masks generated by SAM. It maps SAM's features into a low-dimensional space using a trainable MLP, replicating these features to address inconsistency issues. Click-Gaussian [7] introduces an efficient method for interactive segmentation of 3D Gaussians. It uses two-level granularity feature fields from 2D masks and incorporates Global Feature-guided Learning (GFL) to resolve cross-view inconsistencies. OpenGaussian [34] employs SAM masks to train instance features with 3D consistency, proposing a two-stage codebook for feature discretization and an instance-level 3D-2D feature association method that links 3D points to 2D masks and CLIP features, demonstrating effectiveness in open vocabulary-based 3D object selection and understanding. However, OpenGaussian uses a fixed codebook entry count and does not fully exploit the continuity of 3D space, which constrains its performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 614, + 112, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 112, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 112, + 626 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "3D Gaussian Splatting [16]. Given a set of " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " 3D Gaussians " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", each with center " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", color " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{S}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", they are rendered as an image " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{I}_{ren}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " at a specific viewpoint via differentiable rasterization " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ". The reconstruction loss " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " is then calculated by comparing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 301, + 555, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 337 + ], + "type": "inline_equation", + "content": "I_{ren}" + }, + { + "bbox": [ + 313, + 301, + 555, + 337 + ], + "type": "text", + "content": " with the GT image " + }, + { + "bbox": [ + 313, + 301, + 555, + 337 + ], + "type": "inline_equation", + "content": "I_{gt}" + }, + { + "bbox": [ + 313, + 301, + 555, + 337 + ], + "type": "text", + "content": ". This process can be simplified as shown below, where blue variables are estimizable parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 329, + 340, + 546, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 340, + 546, + 358 + ], + "spans": [ + { + "bbox": [ + 329, + 340, + 546, + 358 + ], + "type": "interline_equation", + "content": "\\mathcal {G} (\\boldsymbol {\\mu}, \\boldsymbol {c}, \\boldsymbol {\\sigma}, \\boldsymbol {R}, \\boldsymbol {S}) _ {n} \\xrightarrow {\\mathcal {D}} \\boldsymbol {I} _ {r e n} \\longrightarrow \\mathcal {L} _ {r g b} \\longleftarrow \\boldsymbol {I} _ {g t}", + "image_path": "d4cd9c7ba48f5eb60e2c55baec6abaec3b5c365aa661f81069ba3997ef14f00c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "text", + "content": "OpenGaussian [34]. Based on 3DGS, OpenGaussian adds an instance feature " + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{f} \\in \\mathbb{R}^{6}" + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "text", + "content": " for each Gaussian, passes it through differentiable rasterization " + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "text", + "content": " to render feature map " + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{H \\times W \\times 6}" + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "text", + "content": ", and computes intra-mask smoothness loss " + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{s}" + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "text", + "content": " and inter-mask contrastive loss " + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{c}" + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "text", + "content": " with SAM masks " + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "text", + "content": ". This process simplifies as follows, where only the feature " + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "text", + "content": " is optimized while other appearance attributes are frozen (shown in gray)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 459, + 550, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 459, + 550, + 478 + ], + "spans": [ + { + "bbox": [ + 328, + 459, + 550, + 478 + ], + "type": "interline_equation", + "content": "\\mathcal {G} (\\mu , c, \\sigma , R, S, f) _ {n} \\xrightarrow {\\mathcal {D}} M \\longrightarrow \\mathcal {L} _ {s}, \\mathcal {L} _ {c} \\longleftarrow B", + "image_path": "baa5fa7c838f5dc99c0cb4e06f8465ddd0a9db7b2cde250aaf9cc471905b215f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 338, + 487, + 554, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 487, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 338, + 487, + 554, + 520 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\sum_ {i = 1} ^ {m} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\boldsymbol {B} _ {i, h, w} \\cdot \\left\\| \\boldsymbol {M} _ {:, h, w} - \\bar {\\boldsymbol {M}} _ {i} \\right\\| ^ {2}, \\tag {1}", + "image_path": "e74b0762f4a2c4d1f57d54c8d4a6c22e5bd8195d5cad50d4bf294d50c6130901.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 341, + 527, + 554, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 527, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 341, + 527, + 554, + 559 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\frac {1}{m (m - 1)} \\sum_ {i = 1} ^ {m} \\sum_ {j = 1, j \\neq i} ^ {m} \\frac {1}{\\left\\| \\bar {\\boldsymbol {M}} _ {i} - \\bar {\\boldsymbol {M}} _ {j} \\right\\| ^ {2}}, \\tag {2}", + "image_path": "b49d55c3c3b02226c3af735e85605c948cb1d9416eacc8d8cf9f27685503470b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": " represents the image size and " + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": " corresponds to the number of SAM masks in the current view. The " + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": " aims to make each pixel feature " + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "inline_equation", + "content": "M_{:,h,w}" + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": " converge towards the average feature " + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\bar{M}_i" + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": " of its associated mask for intramask smoothness. The " + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": " seeks to maximize the separation between the average features " + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\bar{M}_i" + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\bar{M}_j" + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": " of different masks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "Scaffold-GS [20]. It derives " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " neural Gaussian points " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "n^{\\prime}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " anchor points " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "n = 10n^{\\prime}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " in their implementation). The Gaussian centers " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " are obtained by shifting its anchor centers " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{\\prime}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " by an offset " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": ", while other appearance attributes are decoded from the appearance embedding " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{e}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " through MLPs. As shown below, the blue parts are learn" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14080" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 71, + 285, + 148 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 285, + 148 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 285, + 148 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 285, + 148 + ], + "type": "image", + "image_path": "839492095401407be55dc1fbbc448e1d60a616576b5e911b29dc25eec1fead63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 150, + 295, + 184 + ], + "lines": [ + { + "bbox": [ + 55, + 150, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 295, + 184 + ], + "type": "text", + "content": "Figure 2. Progressive appearance-semantic joint training. (a) Train appearance only; (b) Independent appearance-semantic training; (c) Joint appearance-semantic training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 193, + 295, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 295, + 216 + ], + "type": "text", + "content": "able parameters, and their optimization process is similar to 3DGS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 218, + 293, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 218, + 293, + 259 + ], + "spans": [ + { + "bbox": [ + 70, + 218, + 293, + 259 + ], + "type": "interline_equation", + "content": "\\begin{array}{c} \\overbrace {\\mathcal {A} (\\boldsymbol {\\mu} ^ {\\prime} , \\boldsymbol {e}) _ {n ^ {\\prime}}} ^ {+ o} \\quad \\mathcal {G} (\\vec {\\boldsymbol {\\mu}}, \\underbrace {\\boldsymbol {c} , \\boldsymbol {\\sigma} , \\boldsymbol {R} , \\boldsymbol {S}} _ {\\mathrm {M L P s}}) _ {n} \\xrightarrow {\\mathcal {D}} \\boldsymbol {I} _ {\\mathrm {r e n}} \\to \\mathcal {L} _ {\\mathrm {r g b}} \\leftarrow \\boldsymbol {I} _ {\\mathrm {g t}} \\\\ \\end{array}", + "image_path": "f8922af86c1b25668f17ae6f092a167b1023c092ea6282867dd8a4e48492b6f6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 275, + 295, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 300 + ], + "type": "text", + "content": "3.2. Appearance-semantic joint Gaussian Representation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 306, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 460 + ], + "type": "text", + "content": "(1) Appearance-semantic balanced representation. For the same object or region, the texture details may vary while the semantics remain consistent. Consequently, many Gaussians with diverse appearance attributes are required to represent the texture details of the region. These Gaussians require only a single semantic attribute to effectively capture the region's semantics, meaning the same number of Gaussians may under-represent appearance while overrepresenting semantics. We refer to this phenomenon as the imbalance in the representation of appearance and semantics by Gaussians. To this end, we propose a Semantic-Scaffold-GS representation to balance appearance and semantics:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 463, + 296, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 463, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 57, + 463, + 296, + 503 + ], + "type": "interline_equation", + "content": "\\overbrace {\\underbrace {\\mathcal {A} (\\boldsymbol {\\mu} ^ {\\prime} , \\boldsymbol {e} , \\boldsymbol {f}) _ {n ^ {\\prime}}} _ {\\text {M L P s}} \\overbrace {\\mathcal {G} (\\boldsymbol {\\mu} , \\boldsymbol {c} , \\boldsymbol {\\sigma} , \\boldsymbol {R} , \\boldsymbol {S} , \\boldsymbol {\\tilde {f}}) _ {n}} ^ {\\text {s h a r e}}} ^ {+ o} \\xrightarrow {\\mathcal {I} _ {\\text {r e n}},} \\underset {M} {\\longrightarrow} \\overset {\\mathcal {L} _ {\\text {r g b}},} {\\mathcal {L} _ {s}, \\mathcal {L} _ {c}} \\leftarrow \\overset {\\boldsymbol {I} _ {\\text {g t}},} {\\boldsymbol {B}}", + "image_path": "e813fc485a77bf19642c9f93bd91e4945fe28675a0d59f026fc4da1e3436c2be.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "text", + "content": "The blue parts are the estimizable parameters; see Sec. 3.1 for variable definitions. Similar to Scaffold-GS, we derive " + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "text", + "content": " neural Gaussian points " + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "inline_equation", + "content": "n'" + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "text", + "content": " anchor points " + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "inline_equation", + "content": "n = 5n'" + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "text", + "content": " in our implementation). Specifically, each group of 5 child Gaussians with different appearance attributes is decoded from the appearance embedding " + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "text", + "content": " of their common parent anchor point using MLPs. However, at the semantic level, these 5 child Gaussians share the same instance feature " + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 509, + 295, + 688 + ], + "type": "text", + "content": " as their parent anchor point. On the one hand, this representation strengthens the learning of appearance, especially in scenes represented by sparse points (e.g., ScanNet and SLAM-reconstructed scenes); on the other hand, sharing features among neighboring points significantly enhances the efficiency of semantic learning. The process is described in the top row of Fig. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "(2) Appearance-semantic consistency learning. Without semantic constraints, i.e., pure appearance reconstruc" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 311 + ], + "type": "text", + "content": "tion, a Gaussian can represent across objects/regions. For instance, at boundaries, a single Gaussian may simultaneously represent both the object and the background. Recent language-embedded Gaussian Splatting works [25, 34] follow a paradigm of pre-training appearance reconstruction, freezing the appearance properties, and subsequently learning semantic attributes. This tends to cause inconsistencies between appearance and semantics, posing challenges for 3D point segmentation and 2D image segmentation. Therefore, we propose a progressive appearance-seismic joint training strategy to ensure their consistency as shown in Fig. 2: during the 0-10k steps, only the appearance attributes are trained, then appearance and semantics are trained independently in the 10k-20k steps, and finally, appearance and semantics are jointly trained in the 20k-30k steps. Here, we employ the loss (Eq. (1) (2)) of OpenGaussian to train instance features, but its inter-mask contrastive loss value is too large, causing instability of joint training. We overcome this problem through the following improvements:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 318, + 555, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 318, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 340, + 318, + 555, + 352 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\frac {1}{m (m - 1)} \\sum_ {i = 1} ^ {m} \\sum_ {j = 1, j \\neq i} ^ {m} \\frac {\\mathbf {1} _ {\\| \\bar {M} _ {i} - \\bar {M} _ {j} \\| ^ {2} < \\tau}}{\\left\\| \\bar {M} _ {i} - \\bar {M} _ {j} \\right\\| ^ {2}}, \\tag {3}", + "image_path": "b95b5711f5b965f5679b9055e139c13eb964d83bcd6da40ba4bb5377ecabf49f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 359, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 555, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 359, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{\\| \\bar{M}_i - \\bar{M}_j\\|^2 < \\tau}" + }, + { + "bbox": [ + 313, + 359, + 555, + 431 + ], + "type": "text", + "content": " takes values of 1 or 0, aiming to truncate the loss between two objects whose dissimilarity (quantified by the L2 distance) exceeds a threshold " + }, + { + "bbox": [ + 313, + 359, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 359, + 555, + 431 + ], + "type": "text", + "content": " (set to 0.4 in our implementation). This improvement is intuitive because there is no need to enforce that the features of two different objects are infinitely distant from each other." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 438, + 541, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 541, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 541, + 452 + ], + "type": "text", + "content": "3.3. Bottom-up Category-Agnostic Instantiation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 456, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 539 + ], + "type": "text", + "content": "We employ a bottom-up method for category-agnostic instantiation. Specifically, we first over-segment the scene into numerous sub-objects. Then, we construct a graph of these sub-objects and perform aggregation based on the graph connectivity, adaptively obtaining several independent and complete objects. The bottom row of Fig. 1 describes this process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "text", + "content": "(1) Over-segmentation. Inspired by point cloud perception methods, which utilize farthest point sampling (FPS) algorithms to sample seed points for candidate proposals from dense point clouds, we achieve over-segmentation of the scene through FPS-based clustering." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 600, + 554, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 554, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 554, + 681 + ], + "type": "text", + "content": "We first sample " + }, + { + "bbox": [ + 313, + 600, + 554, + 681 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 600, + 554, + 681 + ], + "type": "text", + "content": " seed points from all " + }, + { + "bbox": [ + 313, + 600, + 554, + 681 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 600, + 554, + 681 + ], + "type": "text", + "content": " Gaussians through FPS and use them as the initial centers of " + }, + { + "bbox": [ + 313, + 600, + 554, + 681 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 600, + 554, + 681 + ], + "type": "text", + "content": " clusters, where " + }, + { + "bbox": [ + 313, + 600, + 554, + 681 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 600, + 554, + 681 + ], + "type": "text", + "content": " is much greater than the number of objects in the scene, with a default value of 1000. Then we start the clustering process, updating the instance feature of all Gaussians and the " + }, + { + "bbox": [ + 313, + 600, + 554, + 681 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 600, + 554, + 681 + ], + "type": "text", + "content": " clusters. This process can be formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 400, + 682, + 553, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 682, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 400, + 682, + 553, + 696 + ], + "type": "interline_equation", + "content": "\\bar {f}, \\bar {\\mu}, \\bar {\\mathbf {I}} = \\phi (\\mathbf {X}), \\tag {4}", + "image_path": "6c3ec9e964dffbd57e546000806683a1df97f6e3b3fd1ec22e8534046a395a1c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 362, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 362, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {X} = \\left[ \\operatorname {P E} (\\boldsymbol {\\mu}) \\in \\mathbb {R} ^ {n \\times d}; \\boldsymbol {f} \\in \\mathbb {R} ^ {n \\times 6} \\right], \\tag {5}", + "image_path": "edd1ac03a48caa91f357ec559ed44ebf1e36b8566223ec1a0469d2c9d69c9c4c.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14081" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 87, + 296, + 123 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 282, + 83 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 282, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 282, + 83 + ], + "type": "text", + "content": "Algorithm 1 Bottom-up category-agnostic instantiation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 87, + 296, + 123 + ], + "lines": [ + { + "bbox": [ + 56, + 87, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 296, + 123 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 56, + 87, + 296, + 123 + ], + "type": "inline_equation", + "content": "\\mu, f" + }, + { + "bbox": [ + 56, + 87, + 296, + 123 + ], + "type": "text", + "content": " - point position and feature, " + }, + { + "bbox": [ + 56, + 87, + 296, + 123 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 56, + 87, + 296, + 123 + ], + "type": "text", + "content": " - the number of over-segmented sub-objects, " + }, + { + "bbox": [ + 56, + 87, + 296, + 123 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 56, + 87, + 296, + 123 + ], + "type": "text", + "content": "-voxel resolution, " + }, + { + "bbox": [ + 56, + 87, + 296, + 123 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 87, + 296, + 123 + ], + "type": "text", + "content": " - connectivity threshold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 57, + 123, + 245, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 123, + 245, + 135 + ], + "spans": [ + { + "bbox": [ + 57, + 123, + 245, + 135 + ], + "type": "text", + "content": "Output: well-segmented complete object " + }, + { + "bbox": [ + 57, + 123, + 245, + 135 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{l}},\\hat{\\mathbf{f}}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 139, + 294, + 368 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 63, + 139, + 265, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 139, + 265, + 150 + ], + "spans": [ + { + "bbox": [ + 63, + 139, + 265, + 150 + ], + "type": "text", + "content": "1: procedure OVERSEGMENTATION " + }, + { + "bbox": [ + 63, + 139, + 265, + 150 + ], + "type": "inline_equation", + "content": "(\\mu, f, s, r, \\gamma)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 150, + 294, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 150, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 63, + 150, + 294, + 162 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 63, + 150, + 294, + 162 + ], + "type": "inline_equation", + "content": "\\bar{f},\\bar{\\mu}\\gets FPS(\\pmb {f},\\pmb {\\mu},s)" + }, + { + "bbox": [ + 63, + 150, + 294, + 162 + ], + "type": "text", + "content": " ▷initialize s centers" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 162, + 294, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 162, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 63, + 162, + 294, + 174 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 63, + 162, + 294, + 174 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\gets \\operatorname{concat} PE(\\boldsymbol{\\mu})" + }, + { + "bbox": [ + 63, + 162, + 294, + 174 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 63, + 162, + 294, + 174 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 174, + 294, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 174, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 63, + 174, + 294, + 186 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 63, + 174, + 294, + 186 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{I}},\\overline{\\pmb{f}},\\overline{\\pmb{\\mu}}\\gets k - means(\\mathbf{X},s)" + }, + { + "bbox": [ + 63, + 174, + 294, + 186 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 63, + 174, + 294, + 186 + ], + "type": "text", + "content": " Eq. (4)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 186, + 236, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 186, + 236, + 198 + ], + "spans": [ + { + "bbox": [ + 63, + 186, + 236, + 198 + ], + "type": "text", + "content": "5: return SUBOBJECTAGG " + }, + { + "bbox": [ + 63, + 186, + 236, + 198 + ], + "type": "inline_equation", + "content": "(\\mu, \\overline{\\mathbf{I}}, \\overline{\\mathbf{f}}, s, r, \\gamma)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 199, + 137, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 199, + 137, + 209 + ], + "spans": [ + { + "bbox": [ + 63, + 199, + 137, + 209 + ], + "type": "text", + "content": "6: end procedure" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 213, + 253, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 213, + 253, + 225 + ], + "spans": [ + { + "bbox": [ + 63, + 213, + 253, + 225 + ], + "type": "text", + "content": "7: procedure SUBOBJECTAGG " + }, + { + "bbox": [ + 63, + 213, + 253, + 225 + ], + "type": "inline_equation", + "content": "(\\mu, \\overline{\\mathbf{I}}, \\overline{f}, s, r, \\gamma)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 225, + 266, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 225, + 266, + 236 + ], + "spans": [ + { + "bbox": [ + 63, + 225, + 266, + 236 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 63, + 225, + 266, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{G} \\in \\mathbb{R}^{s \\times s} \\gets" + }, + { + "bbox": [ + 63, + 225, + 266, + 236 + ], + "type": "text", + "content": " initialize the graph with zeros" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 237, + 294, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 237, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 63, + 237, + 294, + 249 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 63, + 237, + 294, + 249 + ], + "type": "inline_equation", + "content": "\\mathbf{V}\\mathbf{o}\\mathbf{x} \\gets \\text{Voxelization}(\\boldsymbol{\\mu},\\bar{\\mathbf{I}},r) \\quad \\triangleright" + }, + { + "bbox": [ + 63, + 237, + 294, + 249 + ], + "type": "text", + "content": " sub-object. voxels" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 249, + 156, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 249, + 156, + 260 + ], + "spans": [ + { + "bbox": [ + 62, + 249, + 156, + 260 + ], + "type": "text", + "content": "10: for " + }, + { + "bbox": [ + 62, + 249, + 156, + 260 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 62, + 249, + 156, + 260 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 62, + 249, + 156, + 260 + ], + "type": "inline_equation", + "content": "[0,s)" + }, + { + "bbox": [ + 62, + 249, + 156, + 260 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 261, + 171, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 171, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 171, + 272 + ], + "type": "text", + "content": "11: for " + }, + { + "bbox": [ + 62, + 261, + 171, + 272 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 62, + 261, + 171, + 272 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 62, + 261, + 171, + 272 + ], + "type": "inline_equation", + "content": "[0,s)" + }, + { + "bbox": [ + 62, + 261, + 171, + 272 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 272, + 250, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 250, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 250, + 285 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 62, + 272, + 250, + 285 + ], + "type": "inline_equation", + "content": "dis(i,j)\\gets L2distance(\\bar{\\pmb{f}}_i,\\bar{\\pmb{f}}_j)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 285, + 233, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 233, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 233, + 297 + ], + "type": "text", + "content": "13: " + }, + { + "bbox": [ + 62, + 285, + 233, + 297 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(i,j)\\gets near(\\mathbf{Vox}_i,\\mathbf{Vox}_j)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 62, + 297, + 294, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 294, + 309 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 294, + 309 + ], + "type": "text", + "content": "14: " + }, + { + "bbox": [ + 62, + 297, + 294, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{ij}\\gets dis(i,j)\\cdot \\mathbb{1}(i,j)" + }, + { + "bbox": [ + 62, + 297, + 294, + 309 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 62, + 297, + 294, + 309 + ], + "type": "text", + "content": " Eq. (6)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 62, + 310, + 136, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 310, + 136, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 310, + 136, + 319 + ], + "type": "text", + "content": "15: end for" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 62, + 321, + 121, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 321, + 121, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 321, + 121, + 331 + ], + "type": "text", + "content": "16: end for" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 62, + 331, + 202, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 202, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 202, + 345 + ], + "type": "text", + "content": "17: " + }, + { + "bbox": [ + 62, + 331, + 202, + 345 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{l}},\\hat{\\pmb{f}}\\gets Aggrecation(\\mathbf{G},\\gamma)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 62, + 345, + 121, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 345, + 121, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 345, + 121, + 357 + ], + "type": "text", + "content": "18: return " + }, + { + "bbox": [ + 62, + 345, + 121, + 357 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{I}},\\hat{f}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 62, + 357, + 137, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 137, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 137, + 368 + ], + "type": "text", + "content": "19: end procedure" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "content": " denotes the k-means, " + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{n \\times (d + 6)}" + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "content": " is the attribute of the " + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "content": " Gaussians participating in the clustering, which is the concatenation of the positional encoding " + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "inline_equation", + "content": "(\\mathrm{PE}(\\cdot))" + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "content": " for position " + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "content": " and the instance feature " + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "content": " sub-object clusters are obtained, " + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{f}} \\in \\mathbb{R}^{s \\times 6}" + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "content": " is their feature, " + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{\\mu}} \\in \\mathbb{R}^{s \\times 3}" + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "content": " is their centers, " + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{I}} \\in \\mathbb{Z}_{[0,s]}^n" + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "content": " represents the label indicating which subobject each point belongs to. During this process, only the feature " + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "content": " needs to be optimized, while the position " + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 55, + 388, + 296, + 509 + ], + "type": "text", + "content": " is frozen. This process is shown in lines 1-4 of Alg. 1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "content": "(2) Graph connectivity-based sub-objects aggregation. This process is shown in lines 7-19 of Alg. 1. We use the " + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "content": " sub-objects obtained from the over-segmentation step as nodes to construct a graph and aggregate the sub-objects into complete objects by evaluating node connectivity. This graph can be represented by an " + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "inline_equation", + "content": "s \\times s" + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "content": " adjacency matrix " + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{ij}" + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "content": " represents the connectivity (edge) between the " + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "content": "-th sub-objects (nodes):" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 609, + 295, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 609, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 116, + 609, + 295, + 625 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {i j} = \\left\\| \\bar {\\boldsymbol {f}} _ {i} - \\bar {\\boldsymbol {f}} _ {j} \\right\\| _ {2} \\cdot \\mathbb {1} (i, j), \\tag {6}", + "image_path": "22e3d99a72c9cfd12a4982e456a7326fbbe06de3fe3fdc564ebce79a2655e347.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "where we design two connectivity criteria: 1) the L2 distance of the node features, a closer distance indicates that the two nodes are more likely to be merged; 2) the spatial relationship between the nodes is represented by the indicator function " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ", which determines whether two subobjects are adjacent or share voxels by voxelizing the 3D space with a resolution " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ", and returns 0 or 1." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": "After constructing the graph, we perform aggregation on the nodes with connectivity values " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{ij}" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " in the range " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "(0,\\gamma]" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " using the standard graph connected component algorithm (such as breadth-first search), where " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " is set to 0.1 in our implementation. Due to the indicator function, our graph is sparse and can be accelerated through the union-find method. Finally, by aggregating the feature-spatial similar sub-objects, we adaptively obtain several " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "(e.g.,m)" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " independent and complete objects with " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}\\in \\mathbb{R}^{m\\times 6}" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " representing their feature, " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}\\in \\mathbb{Z}_{[0,m)}^n" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " representing point label." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 207, + 400, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 400, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 400, + 220 + ], + "type": "text", + "content": "4. Experiments*" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 228, + 550, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 550, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 550, + 242 + ], + "type": "text", + "content": "4.1. Category-Agnostic 3D Instance Segmentation" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 247, + 555, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 427 + ], + "type": "text", + "content": "Settings. 1) Task: Category-agnostic 3D instance segmentation aims to segment independent 3D objects/instances in a given point cloud (Gaussians can naturally be treated as points) scene without considering their semantic categories, ensuring that each point ultimately belongs to only one instance. 2) Dataset and Metrics: Following OpenGaussian, we evaluate using the ScanNet [8] dataset with GT instance-level annotations. More specifically, we calculate the instance-level mIoU and mAcc metrics for the 10 scenes selected by OpenGaussian. 3) Baseline: We mainly compare with Gaussian-based methods OpenGaussian [34] and GaussianGrouping [38] with RGB images and SAM masks as input. Additionally, we provide the performance of the depth-assisted back-projection method MaskClustering [37] as a reference." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 429, + 556, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 556, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 556, + 681 + ], + "type": "text", + "content": "Results. Quantitative results, as shown in Tab. 1, demonstrate that InstanceGaussian significantly outperforms Gaussian-based methods and achieves comparable results to MaskClustering which uses depth to assist. GaussianGrouping relies on a tracking-based model for cross-view object association, which is prone to tracking failures in complex scenes like ScanNet, characterized by significant viewpoint changes and occlusions, consequently leading to lower performance. Although OpenGaussian achieves instance feature learning with cross-view 3D consistency by the smooth-contrast losses, avoiding explicit tracking-based association, its performance bottleneck lies in the brute-force, naive " + }, + { + "bbox": [ + 313, + 429, + 556, + 681 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 429, + 556, + 681 + ], + "type": "text", + "content": "-means clustering, lacking parameter adaptability. Our method enables more discriminative feature learning through appearance-semantic joint representation and training, achieving more accurate and generalizable 3D instance segmentation through oversegmentation followed by connectivity-based adaptive aggregation. The visualization results are shown in Fig. 3, where the well-segmented instances faithfully reflect the performance advantages of InstanceGaussian." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": "* Please refer to the supplementary materials for additional 3D/2D visualization results and comparisons with more methods." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14082" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 81, + 159, + 146 + ], + "blocks": [ + { + "bbox": [ + 91, + 69, + 157, + 80 + ], + "lines": [ + { + "bbox": [ + 91, + 69, + 157, + 80 + ], + "spans": [ + { + "bbox": [ + 91, + 69, + 157, + 80 + ], + "type": "text", + "content": "Reference Mesh" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 81, + 159, + 146 + ], + "lines": [ + { + "bbox": [ + 91, + 81, + 159, + 146 + ], + "spans": [ + { + "bbox": [ + 91, + 81, + 159, + 146 + ], + "type": "image", + "image_path": "3af07ffd268086c9bf3e0506605d324885af5b91db9534728775974547d20ab7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 102, + 148, + 146, + 216 + ], + "blocks": [ + { + "bbox": [ + 102, + 148, + 146, + 216 + ], + "lines": [ + { + "bbox": [ + 102, + 148, + 146, + 216 + ], + "spans": [ + { + "bbox": [ + 102, + 148, + 146, + 216 + ], + "type": "image", + "image_path": "f2f18bb75e9419f29346ee470695f8a7f01309d89abda4cb53d6d9469fe3d876.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 91, + 217, + 158, + 281 + ], + "blocks": [ + { + "bbox": [ + 91, + 217, + 158, + 281 + ], + "lines": [ + { + "bbox": [ + 91, + 217, + 158, + 281 + ], + "spans": [ + { + "bbox": [ + 91, + 217, + 158, + 281 + ], + "type": "image", + "image_path": "1c7e1330adcaa7ed5e4161e5048e0194c5cad4b33eabd7e4fe4d909cf6b29569.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 80, + 280, + 168, + 342 + ], + "blocks": [ + { + "bbox": [ + 80, + 280, + 168, + 342 + ], + "lines": [ + { + "bbox": [ + 80, + 280, + 168, + 342 + ], + "spans": [ + { + "bbox": [ + 80, + 280, + 168, + 342 + ], + "type": "image", + "image_path": "7c13b28042ac9c66b8a675bfe0ec516eaa2d8ad0fb5d5a3cb3aca0ebc5a3b446.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 342, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 555, + 365 + ], + "type": "text", + "content": "Figure 3. Visualization comparison of category-agnostic 3D instance segmentation result. InstanceGaussian outperforms OpenGaussian and GaussinGrouping in accurately distinguishing different 3D objects." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 184, + 82, + 250, + 148 + ], + "blocks": [ + { + "bbox": [ + 181, + 70, + 253, + 81 + ], + "lines": [ + { + "bbox": [ + 181, + 70, + 253, + 81 + ], + "spans": [ + { + "bbox": [ + 181, + 70, + 253, + 81 + ], + "type": "text", + "content": "Gaussian Grouping" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 184, + 82, + 250, + 148 + ], + "lines": [ + { + "bbox": [ + 184, + 82, + 250, + 148 + ], + "spans": [ + { + "bbox": [ + 184, + 82, + 250, + 148 + ], + "type": "image", + "image_path": "7263f0a7df430b451156408669dbd06d1d6754a414c9447adb9bd0a4cb78e539.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 196, + 149, + 239, + 216 + ], + "blocks": [ + { + "bbox": [ + 196, + 149, + 239, + 216 + ], + "lines": [ + { + "bbox": [ + 196, + 149, + 239, + 216 + ], + "spans": [ + { + "bbox": [ + 196, + 149, + 239, + 216 + ], + "type": "image", + "image_path": "539b2081e77fa5c461b170b022908583f25cf08c7a8435819a65baeadee21c55.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 185, + 217, + 251, + 280 + ], + "blocks": [ + { + "bbox": [ + 185, + 217, + 251, + 280 + ], + "lines": [ + { + "bbox": [ + 185, + 217, + 251, + 280 + ], + "spans": [ + { + "bbox": [ + 185, + 217, + 251, + 280 + ], + "type": "image", + "image_path": "249c27bd3b1c7b91d5eb4536319bcdcad1100ac3cddf671a5061ba56c459e4f3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 173, + 280, + 261, + 341 + ], + "blocks": [ + { + "bbox": [ + 173, + 280, + 261, + 341 + ], + "lines": [ + { + "bbox": [ + 173, + 280, + 261, + 341 + ], + "spans": [ + { + "bbox": [ + 173, + 280, + 261, + 341 + ], + "type": "image", + "image_path": "4ecbf7b5f7d3c5fab2a1d654f3a67c324c7e44a72c5fe5bae8e456c788f9c81c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 274, + 82, + 340, + 148 + ], + "blocks": [ + { + "bbox": [ + 280, + 70, + 334, + 80 + ], + "lines": [ + { + "bbox": [ + 280, + 70, + 334, + 80 + ], + "spans": [ + { + "bbox": [ + 280, + 70, + 334, + 80 + ], + "type": "text", + "content": "OpenGaussian" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 274, + 82, + 340, + 148 + ], + "lines": [ + { + "bbox": [ + 274, + 82, + 340, + 148 + ], + "spans": [ + { + "bbox": [ + 274, + 82, + 340, + 148 + ], + "type": "image", + "image_path": "7de0dfaef2c03219e00d7e4b8fe5c32c65d4741c4404a170d4d10e950c599f57.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 286, + 149, + 328, + 216 + ], + "blocks": [ + { + "bbox": [ + 286, + 149, + 328, + 216 + ], + "lines": [ + { + "bbox": [ + 286, + 149, + 328, + 216 + ], + "spans": [ + { + "bbox": [ + 286, + 149, + 328, + 216 + ], + "type": "image", + "image_path": "cf9277f340f954ce4fcd9f8ee5f46b58c4ff7f09106c6e8958651cf1f67eb172.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 274, + 217, + 341, + 279 + ], + "blocks": [ + { + "bbox": [ + 274, + 217, + 341, + 279 + ], + "lines": [ + { + "bbox": [ + 274, + 217, + 341, + 279 + ], + "spans": [ + { + "bbox": [ + 274, + 217, + 341, + 279 + ], + "type": "image", + "image_path": "c4d2b95b9b52ef4a8ef9ff0a19667bbfdd84114b63e7ede518013cb838aff075.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 263, + 279, + 350, + 341 + ], + "blocks": [ + { + "bbox": [ + 263, + 279, + 350, + 341 + ], + "lines": [ + { + "bbox": [ + 263, + 279, + 350, + 341 + ], + "spans": [ + { + "bbox": [ + 263, + 279, + 350, + 341 + ], + "type": "image", + "image_path": "bf146717520e69aa1ba5a9f450d6228cb449ae0f67fce93bd5deed32db76586a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 361, + 82, + 428, + 148 + ], + "blocks": [ + { + "bbox": [ + 360, + 70, + 430, + 80 + ], + "lines": [ + { + "bbox": [ + 360, + 70, + 430, + 80 + ], + "spans": [ + { + "bbox": [ + 360, + 70, + 430, + 80 + ], + "type": "text", + "content": "InstanceGaussian" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 82, + 428, + 148 + ], + "lines": [ + { + "bbox": [ + 361, + 82, + 428, + 148 + ], + "spans": [ + { + "bbox": [ + 361, + 82, + 428, + 148 + ], + "type": "image", + "image_path": "4cd182aa0de18c5cb4b32ac7e8070f296a0ca01a0511ec0c924124ec4d9b5157.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 372, + 149, + 416, + 216 + ], + "blocks": [ + { + "bbox": [ + 372, + 149, + 416, + 216 + ], + "lines": [ + { + "bbox": [ + 372, + 149, + 416, + 216 + ], + "spans": [ + { + "bbox": [ + 372, + 149, + 416, + 216 + ], + "type": "image", + "image_path": "78ec5a7c517262a98b088049bb56585829fdce2b6c46ff0ee704d32148b69515.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 363, + 217, + 429, + 281 + ], + "blocks": [ + { + "bbox": [ + 363, + 217, + 429, + 281 + ], + "lines": [ + { + "bbox": [ + 363, + 217, + 429, + 281 + ], + "spans": [ + { + "bbox": [ + 363, + 217, + 429, + 281 + ], + "type": "image", + "image_path": "400cfa6eb70c179ebec5fcb175f9d5e359078958962f695a92dc91a59459db9e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 352, + 280, + 440, + 341 + ], + "blocks": [ + { + "bbox": [ + 352, + 280, + 440, + 341 + ], + "lines": [ + { + "bbox": [ + 352, + 280, + 440, + 341 + ], + "spans": [ + { + "bbox": [ + 352, + 280, + 440, + 341 + ], + "type": "image", + "image_path": "2afbd7a9d6401a61eee41357a6206de00336e036ed86215fe8780b18ebe29d54.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 451, + 82, + 517, + 148 + ], + "blocks": [ + { + "bbox": [ + 457, + 70, + 511, + 79 + ], + "lines": [ + { + "bbox": [ + 457, + 70, + 511, + 79 + ], + "spans": [ + { + "bbox": [ + 457, + 70, + 511, + 79 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 451, + 82, + 517, + 148 + ], + "lines": [ + { + "bbox": [ + 451, + 82, + 517, + 148 + ], + "spans": [ + { + "bbox": [ + 451, + 82, + 517, + 148 + ], + "type": "image", + "image_path": "87bdb03b1f6d06b757198b57013ebf80d46965ebcc59cbb9e83b55cc51ed39a0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 462, + 150, + 505, + 217 + ], + "blocks": [ + { + "bbox": [ + 462, + 150, + 505, + 217 + ], + "lines": [ + { + "bbox": [ + 462, + 150, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 462, + 150, + 505, + 217 + ], + "type": "image", + "image_path": "5355a039ad2bde57f15b3d4d8a6e778da8582e7286d6ad2728182351813fa1d5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 448, + 217, + 518, + 281 + ], + "blocks": [ + { + "bbox": [ + 448, + 217, + 518, + 281 + ], + "lines": [ + { + "bbox": [ + 448, + 217, + 518, + 281 + ], + "spans": [ + { + "bbox": [ + 448, + 217, + 518, + 281 + ], + "type": "image", + "image_path": "82a5ce080e1d9b939af1a284eead8c1843f05d1ba126ece630e296b36ee1057e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 443, + 280, + 527, + 341 + ], + "blocks": [ + { + "bbox": [ + 443, + 280, + 527, + 341 + ], + "lines": [ + { + "bbox": [ + 443, + 280, + 527, + 341 + ], + "spans": [ + { + "bbox": [ + 443, + 280, + 527, + 341 + ], + "type": "image", + "image_path": "54e3e10d6e4eeb3b8ff7c46d959d37c37cc1f9cf23c1bd97f25c01139d643d05.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 58, + 375, + 293, + 444 + ], + "blocks": [ + { + "bbox": [ + 58, + 375, + 293, + 444 + ], + "lines": [ + { + "bbox": [ + 58, + 375, + 293, + 444 + ], + "spans": [ + { + "bbox": [ + 58, + 375, + 293, + 444 + ], + "type": "table", + "html": "
TypeMethodsmIoU ↑mAcc. ↑
Depth-aidedMaskClustering [37]54.1780.95
Gaussian-basedGaussianGrouping [38]22.5530.54
OpenGaussian [34]27.3252.44
InstanceGaussian (ours)50.2780.22
", + "image_path": "d91acfb6f72fefcf48a828fa39681c8391986cd83309d3a8a17267672774a9d2.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 479, + 295, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 505 + ], + "type": "text", + "content": "4.2. Open-Vocabulary Query Point Cloud Understanding" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "content": "After 3D instantiation as mentioned in Sec. 3.3, we adopt the 2D mask-3D instance association method proposed by OpenGaussian to associate each 3D instance with a 512-dim CLIP image feature. This allows us to support text-based understanding by extracting 512-dim CLIP text features and calculating cosine similarity." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Settings. 1) Task: This task takes text as input and finds the matching point cloud by calculating the cosine similarity between text features and point features. In theory, it supports arbitrary open vocabulary input, but to compute quantitative metrics with the annotated GT point cloud, we use text that corresponds to the pre-defined categories as queries. 2) Dataset and Metrics: Following OpenGaussian, we use 19, 15, and 10 categories from ScanNet as text queries, assigning the closest text to the point cloud to compute the mIoU and mAcc for each category. 3) Baseline: We mainly compare with the recent Gaussian-based meth" + } + ] + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 316, + 376, + 554, + 441 + ], + "blocks": [ + { + "bbox": [ + 55, + 447, + 295, + 469 + ], + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 469 + ], + "type": "text", + "content": "Table 1. Performance of category-agnostic 3D instance segmentation on the Scannet dataset. Accurate is measured by mAcc@0.25." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 376, + 554, + 441 + ], + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 441 + ], + "type": "table", + "html": "
TypeMethods19 classes15 classes10 classes
mIoUmAcc.mIoUmAcc.mIoUmAcc.
DepthMaskClustering [37]34.5954.3234.0045.0738.1056.74
Gaussian-basedLangSplat [25]3.789.115.3513.208.4022.06
LEGaussians [29]3.8410.879.0122.2212.8228.62
OpenGaussian [34]24.7341.5430.1348.2538.2955.19
InstanceGaussian40.6654.0142.5159.1547.9464.01
", + "image_path": "e679f53e0bcb69cd48f17972e3be122604a75149d179f01b5e355a980f9534e1.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 445, + 554, + 468 + ], + "lines": [ + { + "bbox": [ + 313, + 445, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 554, + 468 + ], + "type": "text", + "content": "Table 2. Performance of open vocabulary semantic segmentation on the Scannet dataset. Accurate is measured by mAcc@0.25." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 474, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 521 + ], + "type": "text", + "content": "ods LangSplat, LEGaussians, and OpenGaussian. We also provide the performance of a SOTA method MaskClustering utilizing depth information to assist back projection as a comparison." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": "Results. The quantitative results are shown in Tab. 2, where InstanceGaussian achieves SOTA performance under the settings of 19, 15, and 10 categories, significantly outperforming Gaussian-based methods. The poor performance of LangSplat and LEGaussians stems from ambiguous Gaussian feature learning and detrimental language feature distillation. The performance of OpenGaussian is limited by its coupled appearance-semantic representation and fixed cluster count. GaussianGrouping is not included in the comparison because it does not support 3D-level text queries. Surprisingly, our method even surpasses the depth-assisted MaskClustering, although it outperforms our method in the category-agnostic instance segmentation task. We analyze that the reason lies in its heavy reliance on high-quality SAM masks. Fig. 4 shows the results of relevance point cloud retrieval using text queries, in which" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14083" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 72, + 156, + 125 + ], + "blocks": [ + { + "bbox": [ + 82, + 72, + 156, + 125 + ], + "lines": [ + { + "bbox": [ + 82, + 72, + 156, + 125 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 156, + 125 + ], + "type": "image", + "image_path": "7b18f8ec1534019b3a592b2618ff47398f4ee7fe7a7deaafa68ee68863bf8b43.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 127, + 146, + 135 + ], + "lines": [ + { + "bbox": [ + 89, + 127, + 146, + 135 + ], + "spans": [ + { + "bbox": [ + 89, + 127, + 146, + 135 + ], + "type": "text", + "content": "Reference mesh" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 72, + 228, + 125 + ], + "blocks": [ + { + "bbox": [ + 156, + 72, + 228, + 125 + ], + "lines": [ + { + "bbox": [ + 156, + 72, + 228, + 125 + ], + "spans": [ + { + "bbox": [ + 156, + 72, + 228, + 125 + ], + "type": "image", + "image_path": "a34e5d21ec986b8b67087ee9a99191b021ab5aefe10acfa8e8934de43bad8399.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 230, + 72, + 302, + 125 + ], + "blocks": [ + { + "bbox": [ + 230, + 72, + 302, + 125 + ], + "lines": [ + { + "bbox": [ + 230, + 72, + 302, + 125 + ], + "spans": [ + { + "bbox": [ + 230, + 72, + 302, + 125 + ], + "type": "image", + "image_path": "f1310c2b48c0b0782144e0f2c3d1a5f32433f2599d3ed449450917a4e807957a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 377, + 125 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 377, + 125 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 377, + 125 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 377, + 125 + ], + "type": "image", + "image_path": "9acc604a453ab387d82045f7251341967cd78a1f2fb6def3a12dcd3b94dc425b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 378, + 72, + 451, + 125 + ], + "blocks": [ + { + "bbox": [ + 378, + 72, + 451, + 125 + ], + "lines": [ + { + "bbox": [ + 378, + 72, + 451, + 125 + ], + "spans": [ + { + "bbox": [ + 378, + 72, + 451, + 125 + ], + "type": "image", + "image_path": "cb6adc43521290ee44b41898aeb7dd644976ac076d1c6716fe1a3e6b9dcf4cfa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 452, + 72, + 526, + 125 + ], + "blocks": [ + { + "bbox": [ + 452, + 72, + 526, + 125 + ], + "lines": [ + { + "bbox": [ + 452, + 72, + 526, + 125 + ], + "spans": [ + { + "bbox": [ + 452, + 72, + 526, + 125 + ], + "type": "image", + "image_path": "9cdc8b599f3e3c45bc64cfd1ee313adc71f5f37ed0b4901a955e901acec22eba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 89, + 136, + 149, + 198 + ], + "blocks": [ + { + "bbox": [ + 89, + 136, + 149, + 198 + ], + "lines": [ + { + "bbox": [ + 89, + 136, + 149, + 198 + ], + "spans": [ + { + "bbox": [ + 89, + 136, + 149, + 198 + ], + "type": "image", + "image_path": "8dc98c7b5d9042a590bcbb2037e8df7751ef62faa2cc9e8970c8dd638018e12a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 202, + 553, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 553, + 214 + ], + "type": "text", + "content": "Figure 4. Open-vocabulary query point cloud Understanding on ScanNet dataset. InstanceGaussian shows advanced text query capabilities." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 162, + 127, + 223, + 198 + ], + "blocks": [ + { + "bbox": [ + 162, + 127, + 223, + 198 + ], + "lines": [ + { + "bbox": [ + 162, + 127, + 223, + 198 + ], + "spans": [ + { + "bbox": [ + 162, + 127, + 223, + 198 + ], + "type": "image", + "image_path": "b556a074f72f0b3286ab90d5c1edf380d369594dd46c56d6a01a13213907d801.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 236, + 127, + 297, + 198 + ], + "blocks": [ + { + "bbox": [ + 236, + 127, + 297, + 198 + ], + "lines": [ + { + "bbox": [ + 236, + 127, + 297, + 198 + ], + "spans": [ + { + "bbox": [ + 236, + 127, + 297, + 198 + ], + "type": "image", + "image_path": "0d99224d7a64178e9deae598d2d5764272cec06b9c02ec221d2a44e1e917d148.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 127, + 370, + 198 + ], + "blocks": [ + { + "bbox": [ + 310, + 127, + 370, + 198 + ], + "lines": [ + { + "bbox": [ + 310, + 127, + 370, + 198 + ], + "spans": [ + { + "bbox": [ + 310, + 127, + 370, + 198 + ], + "type": "image", + "image_path": "405581978adba94bba124a188845a900727c01bd4acce8a26f477caf9917ed68.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 384, + 127, + 446, + 198 + ], + "blocks": [ + { + "bbox": [ + 384, + 127, + 446, + 198 + ], + "lines": [ + { + "bbox": [ + 384, + 127, + 446, + 198 + ], + "spans": [ + { + "bbox": [ + 384, + 127, + 446, + 198 + ], + "type": "image", + "image_path": "011ede4acb607a7cb30305db3037ad7f7a990d34222686b64a1761ab647f55bd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 458, + 127, + 518, + 198 + ], + "blocks": [ + { + "bbox": [ + 458, + 127, + 518, + 198 + ], + "lines": [ + { + "bbox": [ + 458, + 127, + 518, + 198 + ], + "spans": [ + { + "bbox": [ + 458, + 127, + 518, + 198 + ], + "type": "image", + "image_path": "8576814350eab78109db68e8929182e8bc3b670a3459f273111e6e65e571f333.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 80, + 220, + 271, + 288 + ], + "blocks": [ + { + "bbox": [ + 80, + 220, + 271, + 288 + ], + "lines": [ + { + "bbox": [ + 80, + 220, + 271, + 288 + ], + "spans": [ + { + "bbox": [ + 80, + 220, + 271, + 288 + ], + "type": "table", + "html": "
MethodsmIoU ↑mAcc. ↑
LangSplat [25]9.6612.41
LEGaussians [29]16.2123.82
OpenGaussian [34]38.3651.34
InstanceGaussian (ours)45.3058.44
", + "image_path": "ece2992c62c5fa6ca89864b99474e9fc02045252ab0bbb83bc5b1fb28c1ca969.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 323, + 296, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 358 + ], + "type": "text", + "content": "both common categories (such as \"bed\", \"curtain\") are used as queries, and the ability to understand open-vocabulary (such as \"ride\", \"music\") is also demonstrated." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 370, + 276, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 276, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 276, + 396 + ], + "type": "text", + "content": "4.3. Open Vocabulary 3D Object Selection and Rendering" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 401, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 544 + ], + "type": "text", + "content": "Settings. 1) Task: Similar to Sec. 4.2, we first use open vocabulary text to select matching 3D Gaussians, and then render the 3D Gaussians into multi-view 2D images through the rasterization pipeline of 3DGS. 2) Dataset and Metrics: Following LangSplat and OpenGaussian, we use the LeRF [17] dataset for evaluation, and after rendering the selected 3D objects into 2D images, we calculate the mIoU and mAcc between the rendered 2D objects and the GT 2D target objects. 3) Baseline: Only Gaussian-based methods possess 3D point-level perception and rendering capabilities, thus we only compare against Gaussian-based methods Langsplat, Legaussian, and OpenGaussian." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "Result. Quantitative results are displayed in the Tab. 3, which demonstrates that even when rendering 3D Gaussians onto 2D images for evaluation. As previously analyzed, LangSplat and LEGaussian, due to their poor 3D perception capabilities, fail to accurately select 3D Gaussians relevant to the text query, resulting in low-quality rendering results. The performance advantage of our method stems from two aspects. First, the bottom-up adaptive instantiation ensures good discrimination between objects. Second, the joint representation and learning of appearance and semantics avoid the appearance-semantic inconsistency present in OpenGaussian. The qualitative results in Fig. 5 more intuitively demonstrate this point. Our method renders objects with clearer boundaries, unlike OpenGaussian," + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 317, + 220, + 553, + 274 + ], + "blocks": [ + { + "bbox": [ + 55, + 290, + 295, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 314 + ], + "type": "text", + "content": "Table 3. Performance of open vocabulary 3D object selection and rendering on Lerf dataset. Accurate is measured by mAcc@0.25." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "type": "table", + "html": "
CaseJotin-repres.Joint-trainSematic Seg.Instance Seg.
#1✓30.7147.40
#2✓33.1549.57
#3✓✓40.6650.27
", + "image_path": "1830d36a10bad6918e9b038fb72fb72f7fa10eba94d830fb15040e0a7424764d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 277, + 555, + 310 + ], + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 310 + ], + "type": "text", + "content": "Table 4. Performance of instance and semantic segmentation with different representation and training strategies on Scannet. Evaluated by mIoU." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 328, + 554, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 554, + 364 + ], + "type": "text", + "content": "which due to its appearance-semantic ambiguity, requires some Gaussians to represent multiple objects/regions, especially evident in the boundary areas." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 382, + 408, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 382, + 408, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 382, + 408, + 396 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 403, + 545, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 545, + 417 + ], + "type": "text", + "content": "4.4.1. Gaussian Representation and Training Strategy" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 422, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 567 + ], + "type": "text", + "content": "We conduct the ablation studies as shown in Tab. 4 to demonstrate the effectiveness of our Semantic-Scaffold-GS joint representation and progressive joint training strategy. The experiments are evaluated using the semantic and instance segmentation tasks on the ScanNet dataset. The joint representation achieves a balance between appearance and semantics, allowing the same anchor point to express different appearances while sharing the same semantics. The progressive joint training avoids inconsistencies in appearance and semantics, enhances the distinguishability between objects, and thus leads to improved instance segmentation and semantic understanding performance." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Fig. 6 more intuitively demonstrates our avoidance of inconsistency. The first and second rows respectively visualize the appearance color and instance features of the 3D Gaussian ellipsoid in the 3D viewer, where the ellipse represents the boundary of the Gaussian. Fig. 6 (a)(c) shows the results of our joint training, which achieve clearer and more distinctive expressions in both appearance and semantics compared to Fig. 6 (b)(d). This improvement not only brings benefits to the understanding of the 3D level (i.e., the semantic level, as proven by our quantitative ablation studies), but also enhances the quality of object rendering (i.e., the appearance level, as shown in Sec. 4.3)." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14084" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 68, + 529, + 210 + ], + "blocks": [ + { + "bbox": [ + 80, + 68, + 529, + 210 + ], + "lines": [ + { + "bbox": [ + 80, + 68, + 529, + 210 + ], + "spans": [ + { + "bbox": [ + 80, + 68, + 529, + 210 + ], + "type": "image", + "image_path": "301ba5096efc9a3ca427731558995e14442424ead17651bb1a5d240f725647ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "text", + "content": "Figure 5. Open-vocabulary 3D object selection and rendering on the LeRF dataset. InstanceGaussian outperforms OpenGaussian in accurately identifying the objects' boundaries by text queries." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 80, + 241, + 526, + 403 + ], + "blocks": [ + { + "bbox": [ + 80, + 241, + 526, + 403 + ], + "lines": [ + { + "bbox": [ + 80, + 241, + 526, + 403 + ], + "spans": [ + { + "bbox": [ + 80, + 241, + 526, + 403 + ], + "type": "image", + "image_path": "dc587546c2592cda3251bc002d4a9c2713bc44890d43879d9e21f0a79aa153dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 406, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 406, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 555, + 430 + ], + "type": "text", + "content": "Figure 6. Visualization results of the appearance color and instance features of the Gaussians ellipsoid in the 3D viewer demonstrate that our method achieves superior appearance-semantic consistency. (a)(c): Results of progressive joint training. (b)(d): Without joint training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 439, + 296, + 496 + ], + "blocks": [ + { + "bbox": [ + 58, + 439, + 296, + 496 + ], + "lines": [ + { + "bbox": [ + 58, + 439, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 58, + 439, + 296, + 496 + ], + "type": "table", + "html": "
CasefeaturevoxelizeSemantic Seg.Instance Seg.
#1✓21.7527.50
#2✓28.9843.41
#3✓✓40.6650.27
", + "image_path": "6dae5ac9d105e4e46affe25a123911ad41734c90b71ab45e18c41594370a6bd4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 499, + 295, + 522 + ], + "lines": [ + { + "bbox": [ + 55, + 499, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 295, + 522 + ], + "type": "text", + "content": "Table 5. Performance of instance and semantic segmentation with different connectivity condition strategies. Evaluated by mIoU." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 530, + 181, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 181, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 181, + 542 + ], + "type": "text", + "content": "4.4.2. Aggregation Condition" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 593 + ], + "type": "text", + "content": "In the process of aggregating over-segmented sub-objects based on a connectivity graph, we considered two connectivity conditions: the L2 distance of instance features, and the proximity judgment after voxelization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "We conduct an ablation study on its configuration as shown in Tab. 5. In case #1, we use the color of the Gaussians as a feature and consider their positions for aggregation, which results in poor performance. In case #2, we only consider instance features and achieve a significant performance improvement, demonstrating the effectiveness of our feature learning. In case #3, we take both conditions into account and achieve the best performance, which is an intuitive idea that in aggregation, we need to consider both semantic similarity and spatial proximity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 440, + 467, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 467, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 467, + 453 + ], + "type": "text", + "content": "5. Conclusion and Limitation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 458, + 555, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 614 + ], + "type": "text", + "content": "In this paper, we introduce an appearance-semantic compact Gaussian representation to address the imbalance between appearance and instance features using Gaussians. On one hand, this representation enhances the learning of appearance; on the other hand, sharing features among neighboring points significantly boosts the efficiency of semantic learning. The bottom-up category-agnostic instantiation adaptively determines the number of categories, cleverly avoiding the issue of over-segmentation or under-segmentation. In category-agnostic segmentation tasks, we significantly outperform previous models, achieving state-of-the-art (SOTA) performance in open vocabulary understanding for 3D point-level tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Despite achieving novel performance on multiple tasks, our method still has the following limitations: 1) Longer training time for slower rendering, since the appearance attributes need to be calculated through MLP. 2) Segmentation accuracy is still affected when most of the images in the scene are incorrectly segmented by SAM. Failure case analyses can be found in the supplementary materials. We hope those drawbacks can be addressed in future works." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14085" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 139 + ], + "type": "text", + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 140, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 296, + 212 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 295, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 273 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 273 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 274, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 334 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 335, + 295, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 406 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 409, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 295, + 456 + ], + "type": "text", + "content": "[6] Jiazhong Cen, Jiemin Fang, Chen Yang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Segment any 3d gaussians. arXiv preprint arXiv:2312.00860, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 457, + 295, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 457, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 62, + 457, + 295, + 505 + ], + "type": "text", + "content": "[7] Seokhun Choi, Hyeonseop Song, Jaechul Kim, Taehyeong Kim, and Hoseok Do. Click-gaussian: Interactive segmentation to any 3d gaussians. In European Conference on Computer Vision (ECCV), 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 506, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 506, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 506, + 295, + 577 + ], + "type": "text", + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 579, + 295, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 295, + 615 + ], + "type": "text", + "content": "[9] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 2004. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 616, + 295, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 616, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 57, + 616, + 295, + 686 + ], + "type": "text", + "content": "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qin-hong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 689, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 689, + 295, + 713 + ], + "type": "text", + "content": "[11] Qiankun Gao, Jiarui Meng, Chengxiang Wen, Jie Chen, and Jian Zhang. Hicom: Hierarchical coherent" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 336, + 73, + 553, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 73, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 336, + 73, + 553, + 108 + ], + "type": "text", + "content": "motion for dynamic streamable scenes with 3d gaussian splatting. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 110, + 553, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 110, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 316, + 110, + 553, + 156 + ], + "type": "text", + "content": "[12] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 158, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 158, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 158, + 553, + 229 + ], + "type": "text", + "content": "[13] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 231, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 289 + ], + "type": "text", + "content": "[14] Yuzhou Ji, He Zhu, Junshu Tang, Wuyi Liu, Zhizhong Zhang, Xin Tan, and Yuan Xie. Fastlgs: Speeding up language embedded gaussians with feature grid mapping. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 291, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 361 + ], + "type": "text", + "content": "[15] Yifan Jiang, Peter Hedman, Ben Mildenhall, Dejia Xu, Jonathan T Barron, Zhangyang Wang, and Tianfan Xue. Alignerf: High-fidelity neural radiance fields via alignment-aware training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 363, + 553, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 553, + 411 + ], + "type": "text", + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 412, + 553, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 553, + 471 + ], + "type": "text", + "content": "[17] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 472, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 543 + ], + "type": "text", + "content": "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 544, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 544, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 553, + 603 + ], + "type": "text", + "content": "[19] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 605, + 553, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 605, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 316, + 605, + 553, + 675 + ], + "type": "text", + "content": "[20] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffolds: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 677, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 553, + 714 + ], + "type": "text", + "content": "[21] Jiarui Meng, Haijie Li, Yanmin Wu, Qiankun Gao, Shuzhou Yang, Jian Zhang, and Siwei Ma. Mirror-3dgs: Incorporating mirror reflections into 3d gaus" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14086" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 78, + 72, + 295, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 295, + 108 + ], + "type": "text", + "content": "sian splatting. In 2024 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 109, + 295, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 109, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 109, + 295, + 156 + ], + "type": "text", + "content": "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 158, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 158, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 158, + 294, + 205 + ], + "type": "text", + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 295, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 295, + 277 + ], + "type": "text", + "content": "[24] Phuc Nguyen, Tuan Duc Ngo, Evangelos Kalogerakis, Chuang Gan, Anh Tran, Cuong Pham, and Khoi Nguyen. Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 278, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 295, + 338 + ], + "type": "text", + "content": "[25] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 339, + 295, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 339, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 339, + 295, + 410 + ], + "type": "text", + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 412, + 295, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 412, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 412, + 295, + 482 + ], + "type": "text", + "content": "[27] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 484, + 295, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 484, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 484, + 295, + 518 + ], + "type": "text", + "content": "[28] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM TOG, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 520, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 295, + 579 + ], + "type": "text", + "content": "[29] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 581, + 295, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 295, + 639 + ], + "type": "text", + "content": "[30] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 641, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 641, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 641, + 295, + 689 + ], + "type": "text", + "content": "[31] Yifan Wang, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 689, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 689, + 295, + 714 + ], + "type": "text", + "content": "[32] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 336, + 72, + 553, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 336, + 72, + 553, + 108 + ], + "type": "text", + "content": "and dense alignment for 3d visual grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 110, + 553, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 110, + 553, + 146 + ], + "spans": [ + { + "bbox": [ + 316, + 110, + 553, + 146 + ], + "type": "text", + "content": "[33] Yanmin Wu, Qiankun Gao, Renrui Zhang, Haijie Li, and Jian Zhang. Language-assisted 3d scene understanding. IEEE Transactions on Multimedia, 2025. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 148, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 148, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 148, + 553, + 232 + ], + "type": "text", + "content": "[34] Yanmin Wu, Jiarui Meng, Hajie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Jian Zhang. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. In Advances in Neural Information Processing Systems, volume 37, pages 19114-19138, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 234, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 234, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 553, + 317 + ], + "type": "text", + "content": "[35] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 320, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 553, + 380 + ], + "type": "text", + "content": "[36] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 382, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 382, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 382, + 553, + 453 + ], + "type": "text", + "content": "[37] Mi Yan, Jiazhao Zhang, Yan Zhu, and He Wang. Maskclustering: View consensus based mask graph clustering for open-vocabulary 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 456, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 503 + ], + "type": "text", + "content": "[38] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In European Conference on Computer Vision (ECCV), 2024. 1, 3, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 506, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 506, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 317, + 506, + 553, + 565 + ], + "type": "text", + "content": "[39] Yingda Yin, Yuzheng Liu, Yang Xiao, Daniel Cohen-Or, Jingwei Huang, and Baoquan Chen. Sai3d: Segment any instance in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 568, + 553, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 627 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 627 + ], + "type": "text", + "content": "[40] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 630, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 630, + 553, + 713 + ], + "type": "text", + "content": "[41] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14087" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 144 + ], + "type": "text", + "content": "[42] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14088" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/661ff35f-282c-423b-9185-bb6d8ff5da4b_content_list.json b/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/661ff35f-282c-423b-9185-bb6d8ff5da4b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8d47dc449f884ac66418cf2d0f49f7e3345f8b60 --- /dev/null +++ b/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/661ff35f-282c-423b-9185-bb6d8ff5da4b_content_list.json @@ -0,0 +1,2030 @@ +[ + { + "type": "text", + "text": "Instant Adversarial Purification with Adversarial Consistency Distillation", + "text_level": 1, + "bbox": [ + 127, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chun Tong Lei $^{1}$ , Hon Ming Yam $^{1}$ , Zhongliang Guo $^{2}$ , Yifei Qian $^{3}$ , Chun Pong Lau $^{1*}$", + "bbox": [ + 169, + 179, + 833, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ City University of Hong Kong $^{2}$ University of St Andrews $^{3}$ University of Nottingham", + "bbox": [ + 192, + 199, + 803, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ctlei2, hmyam4, cplau27}@cityu.edu.hk, zg34@st-andrews.ac.uk, yifei.qian@nottingham.ac.uk", + "bbox": [ + 207, + 218, + 785, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural networks have revolutionized numerous fields with their exceptional performance, yet they remain susceptible to adversarial attacks through subtle perturbations. While diffusion-based purification methods like DiffPure offer promising defense mechanisms, their computational overhead presents a significant practical limitation. In this paper, we introduce One Step Control Purification (OSCP), a novel defense framework that achieves robust adversarial purification in a single Neural Function Evaluation (NFE) within diffusion models. We propose Gaussian Adversarial Noise Distillation (GAND) as the distillation objective and Controlled Adversarial Purification (CAP) as the inference pipeline, which makes OSCP demonstrate remarkable efficiency while maintaining defense efficacy. Our proposed GAND addresses a fundamental tension between consistency distillation and adversarial perturbation, bridging the gap between natural and adversarial manifolds in the latent space, while remaining computationally efficient through Parameter-Efficient Fine-Tuning (PEFT) methods such as LoRA, eliminating the high computational budget request from full parameter fine-tuning. The CAP guides the purification process through the unlearnable edge detection operator calculated by the input image as an extra prompt, effectively preventing the purified images from deviating from their original appearance when large purification steps are used. Our experimental results on ImageNet showcase OSCP's superior performance, achieving a $74.19\\%$ defense success rate with merely 0.1s per purification — a 100-fold speedup compared to conventional approaches.", + "bbox": [ + 88, + 301, + 485, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 771, + 220, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Neural Networks (DNNs) have fundamentally transformed the landscape of computer vision, achieving unprecedented performance across diverse applications. Despite these extraordinary achievements, a critical vulnerability lurks beneath their impressive performance, one that", + "bbox": [ + 89, + 797, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9060e72e6b0fbe0a4a50907d2cce57241bd40c88a2eb66661c84d54cfbc717c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 270, + 903, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9510eaf7efb66e7691b017b12feb96bf99d6c59c29975ea7a6a6180ca8cc3bd4.jpg", + "image_caption": [ + "Figure 1. Comparison between existing methods and our proposed approach. Our method achieves superior performance with just a single inference step, significantly reducing computational cost. Through adversarial noise-adapted U-Net fine-tuning, we demonstrate better detail preservation after denoising, as evident in the zoomed-in regions for circles (bottom). This makes OSCP an efficient and practical solution for adversarial purification." + ], + "image_footnote": [], + "bbox": [ + 517, + 383, + 901, + 443 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "has emerged as a paramount concern in the research community [14, 23, 31].", + "bbox": [ + 511, + 549, + 905, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Achilles' heel of these sophisticated neural architectures lies in their remarkable vulnerability to adversarial attacks [24, 43]. Through the injection of meticulously engineered perturbations that are imperceptible to human observers yet devastating to model performance, adversaries can systematically manipulate DNNs into catastrophic misclassifications [15]. While this vulnerability has found constructive applications in safeguarding against AI misuse, such as protecting intellectual property and preventing unauthorized model exploitation [13], its implications extend far beyond these beneficial uses. The susceptibility to adversarial manipulation poses a critical threat to the deployment of DNNs in high-stakes domains, where system reliability and robustness are not merely desirable but essential for public safety and security [42].", + "bbox": [ + 509, + 580, + 906, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In response to these security challenges, two principal defense paradigms have emerged: adversarial training [22, 46] and adversarial purification [32, 36, 49]. Adversarial training, despite its widespread adoption, exhibits a critical limitation — it requires prior knowledge of attack methods, inherently constraining its effectiveness against", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24331", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d13f6ed46c2324c61849c6357e587026def3804b72f1629f995ea477162c4f4d.jpg", + "image_caption": [ + "Figure 2. The pipeline of Our proposed OSCP. (a) the left figure shows that the adversarial images, which are crafted through intentional attacks, exhibit a shifted distribution after the diffusion process that deviates from the Standard Normal distribution. In response, our proposed GAND can learn to recover the attacked images by modeling this additional adversarial noise with LoRA. (b) the right figure illustrates the pipeline that our proposed CAP leverages non-learnable edge detection operators to guide the purification of adversarial samples, avoiding potential inductive bias introduced by neural networks. It is worth noting that our method achieves remarkable performance by just running a single U-Net inference step." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 419, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f7550585d24860142267a8d7be6a28a7ef5562b43e323557e37b913ad8f70505.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 88, + 903, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "unknown threats. In contrast, adversarial purification offers a more versatile defense framework by focusing on perturbation removal rather than attack-specific training, making it inherently more adaptable to emerging adversarial threats.", + "bbox": [ + 88, + 353, + 480, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial purification methods, particularly those leveraging diffusion models [17, 39], have demonstrated remarkable effectiveness in neutralizing adversarial attacks by mapping perturbed samples back to the natural distribution [32, 44]. Diffusion models, with their unique denoising capabilities and superior training stability, offer significant advantages over traditional generative approaches such as GANs. Their ability to systematically restore corrupted images through iterative denoising has established them as a compelling framework for adversarial defense.", + "bbox": [ + 88, + 414, + 482, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their effectiveness, diffusion models face a critical limitation: their multi-step denoising process incurs substantial computational overhead, making them impractical for real-time defense scenarios. This inherent inefficiency poses a significant challenge in applications demanding rapid response times, such as real-time security systems. Consequently, there is an urgent need for efficient purification methods that can maintain robust defense while minimizing computational costs.", + "bbox": [ + 88, + 566, + 482, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we introduce One Step Control Purification (OSCP), a novel diffusion-based framework that achieves robust adversarial defense in a single inference step. At its core, OSCP first leverages consistency distillation for fast inference while explicitly addressing the fundamental discrepancy between adversarial and clean sample distributions. This insight leads to our first key innovation, the distillation objective — Gaussian Adversarial Noise Distillation (GAND). As shown in Fig. 1 and Fig. 2 (left), GAND enables efficient purification while maintaining strong defense performance.", + "bbox": [ + 88, + 703, + 482, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work also addresses another fundamental challenge in diffusion-based purification — the loss of semantic in-", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "formation in large-step inference. While Latent Consistency Models offer acceleration, they often sacrifice image quality, resulting in semantic degradation and blurry outputs. To preserve both efficiency and fidelity, as shown in Fig. 2 (right), we propose Controlled Adversarial Purification (CAP) as the inference pipeline to empower GAND, which integrates ControlNet with non-learnable edge detection operators, enabling high-quality purification while maintaining semantic integrity.", + "bbox": [ + 511, + 353, + 906, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed OSCP framework, composed of GAND as the distillation objective and CAP as the inference pipeline, achieves state-of-the-art performance with $74.19\\%$ robust accuracy on ImageNet while requiring merely 0.1s per purification. This breakthrough in efficiency and effectiveness not only advances the field of adversarial defense but also enables the practical deployment of robust neural networks in real-world, time-sensitive applications.", + "bbox": [ + 511, + 508, + 908, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our key contributions are as follows:", + "bbox": [ + 532, + 651, + 861, + 666 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose GAND, a novel consistency distillation objective for adversarial training on Latent Consistency Models. Our empirical results demonstrate its exceptional robustness and transferability against unknown adversarial attacks.", + "- We introduce CAP, an innovative inference pipeline that leverages non-learnable edge detection operators to enhance the controllability and semantic preservation of adversarial purification.", + "- We develop OSCP, an integrated framework combining GAND and CAP, that achieves real-time adversarial purification while maintaining robust defense performance, making diffusion-based defenses practical for time-critical applications." + ], + "bbox": [ + 513, + 686, + 903, + 897 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24332", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 89, + 233, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Adversarial training", + "text_level": 1, + "bbox": [ + 89, + 114, + 284, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial training [53] has established itself as a cornerstone defense strategy against adversarial attacks by incorporating perturbed examples into the training process [21]. Numerous studies have shown its effectiveness [2, 11], showing significant improvements in model robustness against known attack patterns. However, this approach exhibits an inherent limitation: models tend to overfit known attacks, compromising their resilience against novel attack vectors [22]. Recent advances attempt to address this limitation by leveraging diffusion models to generate diverse adversarial samples [12, 46], aiming to enhance generalization and prevent overfitting.", + "bbox": [ + 89, + 137, + 483, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Adversarial Purification", + "text_level": 1, + "bbox": [ + 89, + 328, + 313, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial purification has evolved significantly in its approach to defending against attacks through input restoration. Initially pioneered with GAN-based methods [36], the field underwent a paradigm shift with the advent of diffusion models [17, 39, 40]. These diffusion-based approaches have demonstrated superior purification capabilities [32, 44, 49]. Notably, DiffPure [32] simultaneously removes adversarial perturbations and Gaussian noise from the forward process, theoretically justified by the reduced KL divergence between clean and adversarial image distributions. Despite these advances, the computational overhead of multiple denoising steps remains a critical bottleneck [44], limiting practical deployment.", + "bbox": [ + 89, + 351, + 483, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 556, + 259, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models have demonstrated remarkable capabilities across diverse domains, including text-to-image synthesis [34, 35], video generation [3, 18], and 3D content creation [28, 33]. The foundational Denoising Diffusion Probabilistic Models (DDPM) [17] established a two-phase framework: a forward process that gradually applies Gaussian noise following Markov properties, and a reverse process that learns to reconstruct the original image through a reverse Markov chain. However, the sequential nature of this process results in prohibitively long inference times, as computational cost scales linearly with the number of denoising steps.", + "bbox": [ + 89, + 580, + 483, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Efficient Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 771, + 328, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the computational bottleneck, Consistency Models [41] introduced a novel consistency training paradigm, enabling image generation in significantly fewer steps. This advancement led to the development of Latent Consistency Models (LCM) [29], which further accelerated the generation process. LCM-LoRA [30] enhanced efficiency through Low-Rank Adaptation, a Parameter-Efficient Fine-Tuning", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(PEFT) [19] approach that substantially reduces computational requirements. Additionally, ControlNet [51] introduced flexible conditioning mechanisms for Stable Diffusion, enabling precise control over generation through various visual guides including edge maps, pose estimation, sketches, and depth information.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 196, + 651, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffusion model", + "text_level": 1, + "bbox": [ + 511, + 220, + 671, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising Diffusion Probabilistic Models (DDPM) [17] generate images by learning from the reverse Markov chain with Gaussian noise added to the original image. The forward process can be formulated as linear combination of original image $\\pmb{x}_0$ and standard Gaussian noise $\\pmb{\\epsilon}$ , $\\bar{\\alpha}_t$ denoted cumulative product from $\\alpha_1$ to $\\alpha_t$ , $\\alpha_t = 1 - \\beta_t$ for any $t$ , $\\beta_t$ is predefined variance schedule of diffusion process:", + "bbox": [ + 511, + 244, + 906, + 363 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, \\quad \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 368, + 906, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and the denoising step can be expressed as:", + "bbox": [ + 511, + 388, + 802, + 402 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {x}} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta} ^ {*}} \\left(\\boldsymbol {x} _ {t}, t\\right)\\right) + \\sqrt {\\beta_ {t}} \\boldsymbol {\\epsilon}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 405, + 906, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where model parameter $\\theta^{*}$ minimize the loss between actual noise and predict noise:", + "bbox": [ + 511, + 440, + 905, + 470 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} ^ {*} = \\arg \\min _ {\\boldsymbol {\\theta}} \\mathbb {E} _ {\\boldsymbol {x} _ {0}, t, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 472, + 905, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Diffusion-Base Purification (DBP)", + "text_level": 1, + "bbox": [ + 511, + 505, + 808, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DiffPure [32] proposes that diffusion models can remove adversarial noise by performing a sub-process of the normal reverse process ( $t = 0$ to $t = T$ ). By adding predefined $t^*$ ( $t^* < T$ ) of noise to the adversarial image $\\boldsymbol{x}_{adv}$ , which is formed by the sum of original image $\\boldsymbol{x}$ and adversarial noise $\\delta$ , $\\delta$ can be generated by $L_p$ attack [31] or AutoAttack [6]:", + "bbox": [ + 511, + 527, + 905, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\delta} = \\arg \\max _ {\\delta} \\mathcal {L} (C (\\boldsymbol {x} + \\boldsymbol {\\delta}), y), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 622, + 905, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $C$ is the classifier, $y$ is the true label. The forward process of diffusion purification method uses:", + "bbox": [ + 511, + 646, + 905, + 676 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} \\left(t ^ {*}\\right) = \\sqrt {\\bar {\\alpha} \\left(t ^ {*}\\right)} \\boldsymbol {x} + \\sqrt {1 - \\bar {\\alpha} \\left(t ^ {*}\\right)} \\boldsymbol {\\epsilon}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 679, + 903, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and solve the reverse process of DDPM from time step $t^*$ to 0, to get the purified $\\hat{\\pmb{x}}_{adv}^{0}$ that is close to the original image $\\pmb{x}$ , allowing the classifier to classify the image with the correct label.", + "bbox": [ + 511, + 700, + 905, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Consistency Model", + "text_level": 1, + "bbox": [ + 511, + 771, + 694, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models are known to have long inference time, which limits their usage in the real world. To tackle this situation, Consistency model [41] has been proposed, which makes it possible to generate images within 2 to 4 U-Net inference steps, by distilling a Consistency Model from a pretrianed diffusion model.", + "bbox": [ + 511, + 794, + 905, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the efficiency of latent space models compared", + "bbox": [ + 531, + 885, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24333", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to pixel-based models, Latent Consistency Model [29] has been proposed, making use of pretrained encoder and decoder to transform images from pixel space to latent space. Latent Consistency Model's definition is", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t\\right) = f _ {\\theta} \\left(\\boldsymbol {z} _ {t ^ {\\prime}}, t ^ {\\prime}\\right) \\quad \\forall t, t ^ {\\prime} \\in [ 0, T ] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 154, + 482, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $f_{\\theta}$ can be parameterized as:", + "bbox": [ + 89, + 174, + 305, + 189 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} (\\boldsymbol {z}, \\boldsymbol {c}, t) = c _ {\\mathrm {s k i p}} (t) \\boldsymbol {z} + c _ {\\mathrm {o u t}} (t) \\left(\\frac {\\boldsymbol {z} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {\\epsilon} _ {\\theta} (\\boldsymbol {z} , \\boldsymbol {c} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 191, + 483, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{\\mathrm{skip}}(t)$ and $c_{\\mathrm{out}}(t)$ are differentiable functions such that $c_{\\mathrm{skip}}(\\epsilon) = 1$ and $c_{\\mathrm{out}}(\\epsilon) = 0$ . Distilled with LoRA is introduced in LCM-LoRA [30], which dramatically reduces training time and computation cost in distillation by hugely reducing trainable parameters.", + "bbox": [ + 89, + 237, + 483, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 328, + 181, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we aim to solve the slow purification problem in DBP by leveraging the Latent Consistency Distillation to speed up the purification backbone, enabling single-step adversarial purification.", + "bbox": [ + 89, + 354, + 483, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recognizing that adversarial noise differs from the Gaussian noise that diffusion models are typically designed to remove, we propose Gaussian Adversarial Noise Distillation (GAND). This novel distillation objective specifically targets adversarial noise, enhancing the purification performance. Our approach builds upon the insight that combining adversarial purification with adversarial training can yield superior results [25], effectively addressing the distinct distributions of Gaussian and adversarial noise.", + "bbox": [ + 89, + 415, + 483, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, diffusion models tend to produce images deviating from the originals when purification steps $t^*$ are large [44]. To address this, we introduce Controlled Adversarial Purification (CAP), shown in Fig. 2 (right), which utilizes the unlearnable edge detection operator computed on the input to guide the purification process.", + "bbox": [ + 89, + 551, + 483, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. One Step Control Purification", + "text_level": 1, + "bbox": [ + 89, + 652, + 356, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Definition Our goal can be formulated as:", + "bbox": [ + 89, + 676, + 444, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {x} _ {g t} \\simeq \\hat {\\boldsymbol {x}} _ {a d v} ^ {0} = \\mathcal {D} (f (\\mathcal {E} (\\boldsymbol {x} _ {a d v}), t ^ {*})) \\tag {8} \\\\ \\mathrm {s . t .} \\quad C (\\boldsymbol {x} _ {g t}) = C (\\hat {\\boldsymbol {x}} _ {a d v} ^ {0}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 703, + 482, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{x}_{gt}$ is groundtruth image, $\\pmb{x}_{adv}$ is the adversarial image, $f$ is a model with the purification function in latent space, $t^*$ is predefined purification step, $\\mathcal{E}$ and $\\mathcal{D}$ are pretrained image encoder and decoder respectively. For simplifying symbolic, we denote $\\mathcal{E}(\\pmb{x}_{adv})$ as $\\pmb{z}_{adv}$ .", + "bbox": [ + 89, + 750, + 483, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An overview of our method OSCP is illustrated in Fig. 2; OSCP can be separated into two components: a) training a sped-up backbone model with noise and denoise function; b) pipeline for purification process with non-text guidance. We propose Gaussian Adversarial Noise Distillation", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 GAND", + "text_level": 1, + "bbox": [ + 514, + 90, + 655, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Require: dataset $\\mathcal{X}$ , class label of the image $\\pmb{y}$ , classifier $C$ , LCM $f_{\\theta}$ and its parameter $\\theta$ , ODE solver $\\Psi$ and distance metric $d$ . $\\mathcal{L}_C$ and $\\mathcal{L}_G$ indicate $\\mathcal{L}_{CIG}$ (Eq. (14)) and $\\mathcal{L}_{GAND}$ (Eq. (15)) respectively. $t^\\prime$ denote time step $t_{n + k}$ in diffusion process.", + "bbox": [ + 514, + 109, + 903, + 185 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1: while not convergence do", + "bbox": [ + 524, + 186, + 715, + 199 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2: Sample $\\pmb{x} \\sim \\mathcal{X}$ , $n \\sim \\mathcal{U}[1, (N - k) / 2]$", + "3: $\\pmb{z} = \\mathcal{E}(\\pmb{x})$", + "4: $\\pmb{\\delta}_{adv} = \\arg \\max_{\\delta} L(C(\\mathcal{D}(\\pmb{z} + \\pmb{\\delta})), \\pmb{y})$ Eq. (12)", + "5: $z_{t^{\\prime}}^{*} = \\sqrt{\\bar{\\alpha}_{t^{\\prime}}} z + \\sqrt{1 - \\bar{\\alpha}_{t^{\\prime}}} (\\epsilon +\\delta_{adv})$ Eq. (11)", + "6: $\\hat{z}_{t_n}^{\\Psi}\\gets z_{t'}^* +\\Psi (z_{t'}^*,t',t_n,\\emptyset)$ Eq. (13)", + "7: $\\mathcal{L}_G(\\pmb {\\theta},\\pmb {\\theta}^{-})\\gets d(f_\\theta (\\pmb{z}_{t'}^*,\\varnothing ,t'),f_{\\theta^{-}}(\\hat{z}_{t_n}^\\Psi ,\\varnothing ,t_n))$", + "8: $\\mathcal{L}_C(\\pmb {\\theta})\\gets d(f_\\theta (\\pmb {z}_{t'}^*,\\varnothing ,t'),\\pmb {z})$", + "9: $\\mathcal{L}_T(\\pmb {\\theta},\\pmb{\\theta}^-)\\gets \\mathcal{L}_G(\\pmb {\\theta},\\pmb{\\theta}^-) + \\lambda_{CIG}\\mathcal{L}_C(\\pmb {\\theta})$ Eq. (16)", + "10: $\\pmb{\\theta} \\gets \\pmb{\\theta} - \\eta \\nabla_{\\theta} \\mathcal{L}(\\pmb{\\theta}, \\pmb{\\theta}^{-})$", + "11: end while" + ], + "bbox": [ + 519, + 200, + 900, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(GAND) and Controlled Adversarial Purification (CAP) to respond to those two unique requirements.", + "bbox": [ + 511, + 354, + 903, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Gaussian Adversarial Noise Distillation", + "text_level": 1, + "bbox": [ + 511, + 397, + 851, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To respond to the first component we defined, we propose GAND as illustrated in Fig. 2 (left) and detailed in Alg. 1, a novel adversarial distillation objective, which combines adversarial training, Latent Consistency Distillation, and learning how to remove the adversarial and Gaussian noise from the shifted standard Gaussian distribution.", + "bbox": [ + 511, + 421, + 903, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The original latent consistency distillation first encodes the image to latent code by encoder $\\mathcal{E}$ , such that $z = \\mathcal{E}(\\boldsymbol{x})$ and trains the model base on loss function:", + "bbox": [ + 511, + 512, + 903, + 555 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\theta , \\theta^ {-}; \\Psi) = \\tag {9} \\\\ \\mathbb {E} _ {\\boldsymbol {z}, \\boldsymbol {c}, n} [ d (f _ {\\theta} (\\boldsymbol {z} _ {t _ {n + k}}, \\boldsymbol {c}, t _ {n + k}), f _ {\\theta^ {-}} (\\hat {\\boldsymbol {z}} _ {t _ {n}} ^ {\\Psi , \\omega}, \\boldsymbol {c}, t _ {n})) ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 556, + 903, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d(\\cdot, \\cdot)$ is distance metric, $\\Psi(\\cdot, \\cdot, \\cdot, \\cdot)$ is DDIM PF-ODE solver $\\Psi_{\\mathrm{DDIM}}$ [29], $f(\\cdot, \\cdot, \\cdot)$ is latent consistency function in E.q. 7 and $\\hat{z}_{t_n}^{\\Psi, \\omega}$ is an estimation of the solution of the PF-ODE from $t_{n+k} \\to t_n$ using the DDIM PF-ODE solver $\\Psi$ :", + "bbox": [ + 511, + 598, + 905, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {z} _ {t _ {n}} ^ {\\Psi , \\omega} \\leftarrow z _ {t _ {n + k}} + (1 + \\omega) \\Psi \\left(z _ {t _ {n + k}}, t _ {n + k}, t _ {n}, c\\right) \\tag {10} \\\\ - \\omega \\Psi \\left(\\mathbf {z} _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \\varnothing\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 674, + 903, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\omega$ is guidance scale and it is sampled from $[\\omega_{\\mathrm{min}},\\omega_{\\mathrm{max}}]$", + "bbox": [ + 511, + 715, + 883, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to learn the denoise trajectory from a combination of Gaussian noise and adversarial noise to natural image; we need a model satisfying that $f_{\\theta}(z_{adv}(t), \\emptyset, t) = f_{\\theta}(z(t), \\emptyset, t) = z_{\\epsilon}$ for all $t$ . $z_{\\epsilon}$ is the limit of $z_{t}$ when $t \\to 0$ where $z_{\\epsilon} \\simeq z$ . However, if we look closely at the formulation of $f_{\\theta}$ in Eq. (7), $f_{\\theta}(z_{adv}(t), \\emptyset, t)$ is converging to $z_{adv}$ when $t \\to 0$ . $f_{\\theta}(z_{adv}(t), \\emptyset, t) - f_{\\theta}(z(t), \\emptyset, t) \\to \\delta_{adv}$ when $t \\to 0$ , where $\\delta_{adv} = z_{adv} - z$ . This violates the definition of a Latent Consistency Model mentioned in Eq. (6). To address this disharmony, we introduce $z_{t}^{*}$", + "bbox": [ + 511, + 731, + 905, + 881 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {t} ^ {*} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {z} + \\sqrt {1 - \\bar {\\alpha} _ {t}} (\\boldsymbol {\\epsilon} + \\boldsymbol {\\delta} _ {a d v}), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 883, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24334", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by borrowing the idea of forward diffusion process, making $z_{t}^{*}$ be an linear combination of $z$ and $z_{adv}$ , worth to mention that these $\\delta_{adv}$ used in GAND are generated from latent space attack.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\delta} _ {a d v} = \\arg \\max _ {\\delta} \\mathcal {L} \\left(C \\left(\\mathcal {D} (\\mathcal {E} (\\boldsymbol {x}) + \\boldsymbol {\\delta})\\right), \\boldsymbol {y}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 154, + 482, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The $z_{t}^{*}\\rightarrow z$ when $t\\to 0$ , and $z_{t}^{*}\\rightarrow \\epsilon +\\delta_{adv}$ when $t\\rightarrow T$ which converges to the natural image when $t$ is small and matches with our goal, removing the adversarial noise and Gaussian noise concurrently from the shifted normal distribution. Hence, we eliminate disharmony between adversarial training and the original latent consistency distillation objective.", + "bbox": [ + 89, + 178, + 482, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, as we do not give condition embeddings, we remove the $\\omega$ in the formulation of estimation $\\hat{z}_{t_n}^{\\Psi,\\omega}$ to be:", + "bbox": [ + 89, + 285, + 483, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {t _ {n}} ^ {\\Psi} \\leftarrow z _ {t _ {n + k}} + \\Psi \\left(z _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \\varnothing\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 319, + 482, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varnothing$ means null conditional embedding.", + "bbox": [ + 89, + 339, + 387, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, our goal is purifying images instead of image generation, we can train our LCM in a weaker constraint, where we only need the consistency function satisfying on $[0,t], t < T$ . Hence, we only simulate the $n$ of time step $t_n$ in $\\mathcal{U}[1, (N - k) / 2]$ . This means that our LCM only satisfies the consistency function in the first half of the time steps.", + "bbox": [ + 89, + 354, + 483, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, inspired by [20], we add a Clear Image Guide (CIG) loss to further ensure our distillation process is training towards the proposed purification goal. CIG loss is given by:", + "bbox": [ + 89, + 460, + 483, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C I G} (\\boldsymbol {\\theta}) = \\mathbb {E} _ {z, n} [ d (f _ {\\theta} (\\boldsymbol {z} _ {t _ {n + k}} ^ {*}, \\varnothing , t _ {n + k}), \\boldsymbol {z}) ]. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 525, + 482, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, our total distillation loss is the combination of GAND loss", + "bbox": [ + 89, + 544, + 483, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {G A N D} \\left(\\boldsymbol {\\theta}, \\boldsymbol {\\theta} ^ {-}\\right) = \\tag {15} \\\\ \\mathbb {E} _ {z, n} \\left[ d \\left(f _ {\\theta} \\left(\\boldsymbol {z} _ {t _ {n + k}} ^ {*}, \\varnothing , t _ {n + k}\\right), f _ {\\theta^ {-}} \\left(\\hat {\\boldsymbol {z}} _ {t _ {n}} ^ {\\Psi}, \\varnothing , t _ {n}\\right)\\right) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 575, + 482, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and CIG loss as:", + "bbox": [ + 89, + 614, + 202, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {T o t a l}} (\\boldsymbol {\\theta}, \\boldsymbol {\\theta} ^ {-}) = \\mathcal {L} _ {\\text {G A N D}} (\\boldsymbol {\\theta}, \\boldsymbol {\\theta} ^ {-}) + \\lambda_ {C I G} \\mathcal {L} _ {C I G} (\\boldsymbol {\\theta}), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 633, + 482, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\lambda_{CIG}$ is used to ensure two losses are on the same scale.", + "bbox": [ + 89, + 652, + 462, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The following theorem shows if the loss of GAND converges to an arbitrarily small number, the difference between the purified image and the clean image will be arbitrarily small. We can use this as guidance on how to choose our purification $t$ in the inference step since our method is reaching the theoretical optimal performance at a specific $t$ .", + "bbox": [ + 89, + 669, + 483, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Controlled Adversarial Purification", + "text_level": 1, + "bbox": [ + 89, + 770, + 398, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the first component we defined, we propose CAP (as shown in Fig. 3) - a purification inference pipeline that utilizes visual prompts instead of text guidance. We chose visual prompts because text-based guidance can be vulnerable to caption semantic attacks [47], which may compromise the accuracy of generated text prompts. Therefore, we opt for a more traditional and robust approach by using an", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f6755442a0008389179161a9a360e07432b1789be02605ff407073e312598976.jpg", + "image_caption": [ + "Figure 3. CAP used edge image of the adversarial image to control our purification process, maximizing the remaining semantic information of the purified image." + ], + "image_footnote": [], + "bbox": [ + 514, + 90, + 901, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "unlearnable edge detection operator to generate edge guidance of the input adversarial sample.", + "bbox": [ + 511, + 276, + 903, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In purification process, we first encode the adversarial image $\\mathbf{z}_{adv} = \\mathcal{E}(\\mathbf{x}_{adv})$ to latent space using pre-trained image encoder $\\mathcal{E}$ and sample a random noise $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ in the dimension of the latent space. Then, we diffuse the $\\mathbf{z}_{adv}$ with predefined strength $t^*$ , using forward latent diffusion process:", + "bbox": [ + 511, + 306, + 905, + 396 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {a d v} \\left(t ^ {*}\\right) = \\sqrt {\\bar {\\alpha} \\left(t ^ {*}\\right)} \\boldsymbol {z} _ {a d v} + \\sqrt {1 - \\bar {\\alpha} \\left(t ^ {*}\\right)} \\epsilon . \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 400, + 903, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we purify $\\mathbf{z}_{adv}(t^{*})$ using our LCM trained by GAND (in Alg. 1), $f_{\\theta}(\\mathbf{z},\\mathbf{c},t)$ , $\\mathbf{z}$ is image latent, $\\mathbf{c}$ is the condition embedding (e.g., text, canny edge image) and $t$ is time step. The Latent Consistency Model has been introduced in Eq. (7). The purified image latent comes from the latent consistency function, $\\hat{\\mathbf{z}}_{adv}^{0} = f_{\\theta}(\\mathbf{z}_{adv}(t^{*}),\\mathbf{c}_{ce},t^{*})$ , where $\\mathbf{c}_{ce}$ means canny edge images which are provided by edge detection operators [4]. Although ControlNet is an extra plug-in tool for Stable Diffusion, $\\mathbf{c}_{ce}$ is not exactly an input of $f_{\\theta}$ , we treat it as condition embedding here for simplifying the equation. To further reduce the effect of the adversarial image, we remove the $c_{skip}(t)\\mathbf{z}_{adv}(t)$ in LCM $f_{\\theta}$ and denote this LCM as $f_{\\theta}^{-}$ , since this term will maintain most of the adversarial noise. Therefore, our purified image latent is actually:", + "bbox": [ + 511, + 421, + 905, + 647 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {z} _ {a d v} ^ {0} = f _ {\\theta} ^ {-} \\left(\\mathbf {z} _ {a d v} \\left(t ^ {*}\\right), \\mathbf {c} _ {c e}, t ^ {*}\\right) \\\\ = c _ {\\text {o u t}} (t) \\left(\\frac {\\boldsymbol {z} _ {a d v} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {\\epsilon} _ {\\theta} (\\boldsymbol {z} _ {a d v} , \\boldsymbol {c} _ {c e} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right). \\tag {18} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 648, + 903, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Same as setting $c_{skip}(t) \\equiv 0$ . Finally, We reconstruct the purified image by the image decoder $\\mathcal{D}$ , $\\hat{\\pmb{x}}_{adv}^{0} = \\mathcal{D}(\\hat{\\pmb{z}}_{adv}^{0})$ .", + "bbox": [ + 511, + 705, + 903, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments and Results", + "text_level": 1, + "bbox": [ + 513, + 753, + 746, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training setting Our adversarial latent consistency model is distilled from Stable Diffusion v1.5 [34], which utilizes $\\epsilon$ -prediction [17]. We split the first 10,000 images and the remaining 40,000 images in the ImageNet [8] valset as our testing set and training set, respectively. The GAND framework is trained on our training set, with all images resized to $512 \\times 512$ resolution. Training proceeds for 20,000 iterations with a batch size of 4, employing an initial learning", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24335", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/596b7fa0d0f097f479b8a16713c1ec36cc424403ed8075f1a2fc229eb47b2300.jpg", + "table_caption": [ + "Table 1. Accuracy (%) results for ImageNet. The best models against different attack methods are in bold. The methods marked with * mean the data are borrowed from the original paper." + ], + "table_footnote": [], + "table_body": "
DefenseAttack MethodStandard AccuracyRobust AccuracyArchitecture
Without defenseuntargeted PGD-10080.55%0.01%ResNet-50
Without defenseAutoAttack80.55%0.00%ResNet-50
Without defenserandom-targeted PGD-4082.33%0.04%ResNet-152
Diffusion Based Purification
Wang et al. [44]untargeted PGD-10073.53%72.97%ResNet-50
Wang et al. [44]random-targeted PGD-4078.10%77.86%ResNet-152
DiffPure [32]AutoAttack75.77%73.02%ResNet-50
Adversarial Training
Amini et al. [1]*AutoAttack77.96%59.64%ConvNeXt-L
Singh et al. [38]*AutoAttack77.00%57.70%ConvNeXt-L
Hybrid
OSCP (ours)untargeted PGD-10077.63%73.89%ResNet-50
OSCP (ours)AutoAttack77.63%74.19%ResNet-50
OSCP (ours)random-targeted PGD-4079.81%78.78%ResNet-152
", + "bbox": [ + 102, + 121, + 890, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rate of 8e-6 and a 500-step warm-up period. For the diffusion process, we employ DDIM-Solver [39] as the PF-ODE solver $\\Psi$ with a skipping step $k = 20$ in Equation 13. Adversarial latents are generated using PGD-10 with $\\epsilon = 0.03$ against a ResNet50 victim model, and the consistency improvement gradient weight $\\lambda_{CIG}$ is set to 0.001. Notably, our experiments demonstrate that PEFT via LoRA is sufficient for achieving optimal performance.", + "bbox": [ + 88, + 405, + 480, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attack Settings We conduct our evaluations on our testing set. Our evaluation encompasses diverse architectures, spanning traditional CNNs (ResNet 50 and 152 [16], WideResNet [50]) and modern Vision Transformers (ViT-B [10], Swin-B [27]). For adversarial attacks, we employ standard $L_{p}$ norm-based methods including PGD [31] and AutoAttack [6]. The PGD attacks are configured with $L_{\\infty}$ bounds $\\gamma \\in \\{4/255, 16/255\\}$ and corresponding step sizes $\\eta \\in \\{1/255, 0.025 \\cdot 16/255\\}$ , while AutoAttack uses $L_{\\infty}$ bound $\\gamma = 4/255$ . Here, PGD- $n$ denotes PGD attack with $n$ iterations. In our evaluation metrics, we define standard accuracy as performance on clean data with our defense framework, robust accuracy as performance on adversarial data with our defense, and clean accuracy as performance on clean data without any defense. We fix the random seed of every experiment to avoid the randomness.", + "bbox": [ + 89, + 527, + 482, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Main Result", + "text_level": 1, + "bbox": [ + 89, + 784, + 220, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method is regarded as a hybrid method, consisting of adversarial training and purification. Hence we compare our method with both categories of the defenses. Based on the provided attack setting, our method demonstrates significant computational efficiency, completing evaluations in 3 hours on an NVIDIA F40 GPU compared to GDMP's", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2d8403619d46e3715b788089c82c8b15c24f82a5b3a5b73f241e07db75a78cf7.jpg", + "table_caption": [ + "Table 2. Accuracy (\\%) and attack success rate (ASR %) results in various architectures, where the attack method is PGD-100, with budget 4/255, attack step size 1/255." + ], + "table_footnote": [], + "table_body": "
ArchitectureCleanASRStandardRobust
WRN-50-282.6%100%77.6%75.2%
Vit-b-1681.6%100%78.2%71.6%
Swin-b83.6%100%79.2%77.8%
", + "bbox": [ + 522, + 449, + 890, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[44] 48 hours. As shown in Tab. 1, diffusion based purification achieves satisfactory robust accuracy while sacrificing standard accuracy. On the other hand, adversarial training has little standard accuracy loss but the robust accuracy is unsatisfactory. Our method achieves promising results on both standard and robust accuracy with three different attack methods. We achieve robust accuracy improvements of $0.92\\%$ and $1.17\\%$ against PGD-100 ( $\\gamma = 4/255$ ) and AutoAttack respectively, while improving from $77.86\\%$ to $78.78\\%$ against random targeted PGD-40 ( $\\gamma = 16/255$ ). We empirically determine $t^{*} = 200$ as the optimal purification strength.", + "bbox": [ + 511, + 551, + 906, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Analysis With Different Architectures", + "text_level": 1, + "bbox": [ + 511, + 741, + 841, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While our GAND-trained LCM is initially optimized using adversarial examples generated for ResNet50, we evaluate its generalization capability across different architectures to address potential overfitting concerns. As shown in Tab. 2, our method maintains strong robust accuracy across architectures, with even the lowest performance of $71.6\\%$ on $\\mathrm{ViT}_{b-16}$ demonstrating effective transfer. These experiments, along with subsequent ImageNet evaluations, are conducted on a representative subset of 500 images from", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24336", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/91fedb18906819faaafa0993a6f639672aac24f0f9d910697536c296b3dd2c42.jpg", + "table_caption": [ + "Table 3. Robust Accuracy (%) on our method under Diff-PGD-10 attack $\\gamma = 8/{255}\\left( {\\eta = 2/{255}}\\right)$ on ImageNet [8]." + ], + "table_footnote": [], + "table_body": "
Defence methodResNet-50ResNet-152WideResNet-50-2Vit-b-16Swin-bConvNeXt-b
DiffPure (t* = 100)53.8%49.4%52.2%16.6%45.1%42.9%
Ours (t* = 250)59.0%56.5%57.9%34.1%53.9%49.1%
", + "bbox": [ + 135, + 107, + 859, + 175 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3a6675548748c9e43a88d2b684b7755679f4e642174df5148fda6e5224a64694.jpg", + "table_caption": [ + "Table 4. Inference time of purification models to purify one image on an NVIDIA F40 GPU. Notably, GDP refers to Wang et al. [44], DiffPure refers to Nie et al. [32]." + ], + "table_footnote": [], + "table_body": "
MethodDatasett* ∈ [0,1000]runtime (s)
GDMPImageNet250~ 9
DiffPureImageNet150~ 11
OSCP(Ours)ImageNet20/200/500/1000~ 0.1
OSCP(Ours)CelabA-HQ20/200/500/1000~ 0.5
", + "bbox": [ + 91, + 232, + 478, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our previously mentioned 10,000 images test set, as we observe consistent performance between these sample sizes.", + "bbox": [ + 89, + 327, + 482, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Defend Adaptive Attack", + "text_level": 1, + "bbox": [ + 89, + 368, + 313, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 3, we test our method under version 2 of Diff-PGD attack [48], which is a SOTA attack method against diffusion based purification. Our method boost the robust accuracy in Vit-b-16 from $16.6\\%$ to $34.1\\%$ . In ResNet-50, ResNet-152, WideResNet-50-2, the robust accuracy increase $5.2\\%$ , $7.1\\%$ and $5.7\\%$ respectively. For robust accuracy in Swin-b and ConvNeXt, our method achieve $53.9\\%$ and $49.1\\%$ , which is $8.8\\%$ and $6.2\\%$ higher than DiffPure. Worth to mention that our method uses LCM model which is deterministic, and hence, Expectation over Transformation, which is usually use in other robust evaluation for DBP, is not needed here.", + "bbox": [ + 88, + 392, + 482, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Inference Time", + "text_level": 1, + "bbox": [ + 89, + 584, + 243, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conventional diffusion-based purification methods are computationally intensive, making them impractical for real-time applications. As shown in Tab. 4, our method achieves purification in just $\\mathbf{0.1s}$ for ImageNet and $\\mathbf{0.5s}$ for CelebA-HQ (1024 × 1024) on a NVIDIA F40 GPU, independent of $t^*$ where the inference time of other method scales up depending on $t^*$ . More detailed experiments can be found in the supplementary materials. Our method significantly speeds up the purification process, and this acceleration enables deployment in time-critical scenarios such as autonomous driving and purification in video.", + "bbox": [ + 88, + 608, + 482, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Face Recognition", + "text_level": 1, + "bbox": [ + 89, + 785, + 259, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further test our model on a subset of 1000 images from CelebA-HQ [26], choosing purification step $t^*$ as 200, control scale 0.8, and utilizing the GAND weights trained on ImageNet. We defend against targeted PGD attacks that attempt to manipulate Arcface [9] (AF), FaceNet [37] (FN), and MobileFaceNet [5] (MFN) to misidentify any input face", + "bbox": [ + 88, + 809, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/838b58af723cecc3df2f8f4e4c753a39b51db2a0372da66ef626111cc41da27a.jpg", + "table_caption": [ + "Table 5. Robust accuracy (%) for CelebA-HQ under targeted PGD-10 $L_{\\infty} \\gamma (\\gamma = 4 / 255)$ , $\\eta = 0.5 * 5 / 255$ . The best model is in bold." + ], + "table_footnote": [], + "table_body": "
Defense MethodAFFNMFN
Without defense0%0.3%0%
GaussianBlur (σ = 7.0)2.8%51.4%2.8%
Das et al. [7] (n = 60)17.3%84.1%27.6%
CAP(Ours)83.4%97.8%82.8%
OSCP(Ours)86.8%97.8%84.9%
", + "bbox": [ + 524, + 232, + 887, + 345 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a2160f5dd6caf266a5e3cbd523a79f78258e299459c0d271985e1728f9608213.jpg", + "table_caption": [ + "Table 6. IQAs for $\\epsilon = 8 / 255$ . $\\uparrow/\\downarrow$ indicate higher/ lower value consists the better image quality. Notably, IQA $(\\cdot)$ indicates one of IQA we leverage, $x$ refers to the clean image, $x_{\\mathrm{adv}}$ refers to the adversarial sample, $x_{\\mathrm{ours}}$ refers to the purified image by our method and $x_{\\mathrm{DiffPure}}$ refers to purified image by DiffPure." + ], + "table_footnote": [], + "table_body": "
LPIPS↓PSNR↑SSIM↑
IQA(x, xadv)0.154734.210.8906
IQA(x, xours)0.237024.130.7343
IQA(x, xDiffPure)0.261624.110.7155
", + "bbox": [ + 537, + 426, + 880, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as a specific target person. Our defense goal is to prevent this misidentification by ensuring the cosine similarity between the purified image embedding and the target image remains below the recognition threshold. To validate the effectiveness of ImageNet-trained GAND weights on CelebA-HQ, we compare OSCP with CAP (OSCP without GAND) as shown in Tab. 5. The GAND weights significantly improve robust accuracy from $83.4\\%$ to $86.8\\%$ on Arcface and from $82.8\\%$ to $84.9\\%$ on MobileFaceNet. Notably, both our methods (CAP and OSCP) achieve an exceptional robust accuracy of $97.8\\%$ on FaceNet.", + "bbox": [ + 511, + 494, + 906, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.6. Image Quality Assessment", + "text_level": 1, + "bbox": [ + 511, + 666, + 754, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regarding image quality assessments, we compare our method with DiffPure using a subset of 1000 ImageNet images from our testing set. We use PSNR to reflect the overall distortion level of images; SSIM [45] to indicate how human eyes perceive image structures; LPIPS [52] to assesses perceptual similarity. Both methods are configured with $t^{*} = 250$ for fair comparison. As shown in Tab. 6, our method demonstrates superior performance across multiple metrics: LPIPS decreases from 0.2616 to 0.2370, while PSNR and SSIM improve from 24.11 to 24.13 and 0.7155 to 0.7343. These improvements consistently indicate that our purified images maintain greater fidelity to the original inputs, validating the effectiveness of our purification approach. The visualization is shown in Fig. 4.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24337", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0bf874e9c64cd2ac0b736733c729e46dd27809c8f0cadc6e9baf2097b446b88f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 88, + 192, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9f9335b1abfeaaf6928a89c694c34289a754062db12d25ce0064989252c797b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 88, + 287, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4cefdc3bdf64550a3a4bdbcad813403cc10290705f15c03107b6ee83e172a2cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 88, + 382, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/876422d0e6b470b9a35c3544ebcce78147441d3836ba718f6c325b006be042cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 88, + 478, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2eb80cc15bc8e46f5898b03f567979ed2929e276dd718cf4ee6c1df364a4e88b.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 98, + 164, + 191, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf0271471374b6d6053683b36950c62a53437ec4f8e62ab167335dd5537dbef9.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 192, + 164, + 287, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c49710e3ec5a8efabe4ec99e2ea39baacf6eac4d1484e064395764e2f9dc93d0.jpg", + "image_caption": [ + "Figure 4. Visualization of the IQA experiment which compares with DiffPure and the proposed method. (a) Input image. (b) Adversarial image. (c) DiffPure. (d) Ours." + ], + "image_footnote": [], + "bbox": [ + 289, + 164, + 383, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2bcf63c853ad961c7988b57cdc2f2b9bd755f9086a941507d011a8401b9090a9.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 383, + 164, + 478, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d68a87a8c723768a763eb9d64b79f9b258edc6503a45b8b2a0c0981aec53293e.jpg", + "image_caption": [ + "Figure 5. Performance of our method on different $t^*$ under PGD-100 $L_{\\infty}\\gamma$ ( $\\gamma = 4/255$ ), $\\eta = 0.01^{*}4/255$ , where we evaluate on ResNet50 on ImageNet." + ], + "image_footnote": [], + "bbox": [ + 114, + 316, + 455, + 443 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/059909d9f4621f9b462962c1545362d7423ea740877a894eee5adf4cdefb32bb.jpg", + "table_caption": [ + "Table 7. Accuracy (%) under PGD-100 $L_{\\infty}\\gamma$ ( $\\gamma = 4/255$ ), $\\eta = 0.01*4/255$ . The best result is in **bold**, the 2nd best is **underlined**." + ], + "table_footnote": [], + "table_body": "
CAPGANDStandardRobust
XX75.0%72.4%
X✓74.0%72.6%
✓X77.0%73.2%
✓✓76.8%75.0%
", + "bbox": [ + 161, + 536, + 411, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.7. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 650, + 254, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct comprehensive ablation studies on 500 images subset from our testing set. First, as shown in Fig. 5, we optimize CAP's performance across different ControlNet conditioning scales, identifying 0.8 as the optimal value. Tab. 7 demonstrates the individual and combined effects of our components. The baseline LCM-LoRA already shows promising performance, indicating its potential for adversarial purification. GAND-trained weights yield marginal improvements, while CAP integration increases standard and robust accuracy by $2\\%$ and $0.8\\%$ respectively. The full OSCP framework (combining both) significantly improves robust accuracy from $72.4\\%$ to $75\\%$ , validating our design choices. In Tab. 8, we analyze the impact of the $c_{skip}(t)\\pmb{z}_t$ term from Eq. (18). While standard accuracy remains stable, removing this term improves robust accuracy,", + "bbox": [ + 89, + 674, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/52616e538aa15c902b9494bc65b07e6c5c68d0e612c64181ee3b7895f73ba62f.jpg", + "table_caption": [ + "Table 8. Accuracy with and without $c_{skip}(t)z_t$ term in LCM in our method under PGD-100 ( $\\gamma = 8/255$ ), $\\eta = 2/255$ on ResNet50 on a subset of 500 images from our ImageNet testing set." + ], + "table_footnote": [], + "table_body": "
\\(c_{skip}(t)z_t\\)Acc.\\(t^* = 20\\)\\(t^* = 60\\)\\(t^* = 100\\)\\(t^* = 200\\)\\(t^* = 250\\)
XRobust43.4%56.0%59.4%66.4%68.4%
XStandard81.2%81.4%81.2%79.2%76.6%
✓Robust44.2%54.8%60.4%66.0%67.8%
✓Standard81.2%81.4%81.2%79.2%76.6%
", + "bbox": [ + 517, + 133, + 906, + 223 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b9536e1c72c323099b2276038273e71c5c5ea72ba06d8f5938759562abaaef3a.jpg", + "table_caption": [ + "Table 9. Accuracy of our method under AutoAttack ( $\\gamma = 8/255$ ) on ResNet50 on a subset of 1000 images from ImageNet on different sets of training time steps. ( $t^{*} = 200$ )" + ], + "table_footnote": [], + "table_body": "
Acc.N-k/4N-k/23(N-k)/4N-k
Robust74.0%74.3%73.7%73.6%
Standard78.8%79.4%79.1%79.1%
", + "bbox": [ + 550, + 281, + 870, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "likely because $z_{t}$ retains adversarial noise from the input image. This observation leads us to exclude this term from our LCM implementation.", + "bbox": [ + 511, + 349, + 905, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, Tab. 9 explores different training time step ranges: $\\mathcal{U}[1,(N - k) / 4],\\mathcal{U}[1,(N - k) / 2],\\mathcal{U}[1,3(N - k) / 4]$ , and $\\mathcal{U}[1,N - k]$ (original). Results show that smaller time step ranges maintain better performance, suggesting that fine-tuning LCM with smaller time steps is sufficient for purification tasks.", + "bbox": [ + 511, + 393, + 905, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 511, + 494, + 761, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present One Step Control Purification (OSCP), a novel framework that addresses two critical challenges in diffusion-based adversarial purification: computational efficiency and semantic preservation. Our approach integrates the knowledge of pre-trained Diffusion Models with the proposed Gaussian Adversarial Noise Distillation (GAND) to achieve robust defense in a single inference step. Then as the inference framework, the Controlled Adversarial Purification (CAP), utilizing the unlearnable edge detection operator, effectively preserves semantic information of the input during purification. Extensive experiments validate OSPC's effectiveness across diverse architectures and datasets. On ImageNet, we achieve $74.19\\%$ robust accuracy against strong attacks while requiring only 0.1s per purification—a 100-fold speedup compared to existing methods. The framework demonstrates strong cross-domain generalization, from natural images to face recognition tasks, while our ablation studies confirm the synergistic benefits of combining GAND and CAP.", + "bbox": [ + 511, + 518, + 906, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our work represents significant progress in real-time adversarial defense, several promising directions remain, including the adaptive purification strength mechanisms and alternative control conditions. These advances could further enhance the applicability of diffusion-based defenses in time-critical scenarios.", + "bbox": [ + 511, + 806, + 908, + 896 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24338", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sajjad Amini, Mohammadreza Teymoorianfard, Shiqing Ma, and Amir Houmansadr. Meansparse: Post-training robustness enhancement through mean-centered feature sparsification. CoRR, 2024. 6", + "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, pages 274-283, 2018. 3", + "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22563-22575, 2023. 3", + "[4] John Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-8(6):679-698, 1986. 5", + "[5] Sheng Chen, Yang Liu, Xiang Gao, and Zhen Han. Mobilefacenet: Efficient cnns for accurate real-time face verification on mobile devices. In Biometric Recognition, pages 428-438. Springer International Publishing, 2018. 7", + "[6] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International Conference on Machine Learning, pages 2206-2216, 2020. 3, 6", + "[7] Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen, Fred Hohman, Siwei Li, Li Chen, Michael E Kounavis, and Duen Horng Chau. Shield: Fast, practical defense and vaccination for deep learning usingJPEG compression. In ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 196-204, 2018. 7", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 5, 7", + "[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 7", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6", + "[11] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv e-prints, pages arXiv-2010, 2020. 3", + "[12] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Zhongliang Guo, Junhao Dong, Yifei Qian, Kaixuan Wang, Weiye Li, Ziheng Guo, Yuheng Wang, Yanli Li, Ognjen Arandjelovic, and Lei Fang. Artwork protection against neural style transfer using locally adaptive adversarial color attack. In ECAI 2024, pages 1414-1421. IOS Press, 2024. 1", + "[14] Zhongliang Guo, Chun Tong Lei, Lei Fang, Shuai Zhao, Yifei Qian, Jingyu Lin, Zeyu Wang, Cunjian Chen, Ognjen Arandjelovic, and Chun Pong Lau. A grey-box attack against latent diffusion model-based image editing by posterior collapse. arXiv preprint arXiv:2408.10901, 2024. 1", + "[15] Zhongliang Guo, Weiye Li, Yifei Qian, Ognjen Arandjelovic, and Lei Fang. A white-box false positive adversarial attack method on contrastive loss based offline handwritten signature verification models. In International Conference on Artificial Intelligence and Statistics, pages 901-909, 2024. 1", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 5", + "[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3", + "[19] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799, 2019. 3", + "[20] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ODE trajectory of diffusion. In International Conference on Learning Representations, 2024. 5", + "[21] Chun Pong Lau, Jiang Liu, Hossein Souri, Wei-An Lin, Soheil Feizi, and Rama Chellappa. Interpolated joint space adversarial training for robust and generalizable defenses. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):13054-13067, 2023. 3", + "[22] Wei-An Lin, Chun Pong Lau, Alexander Levine, Rama Chellappa, and Soheil Feizi. Dual manifold adversarial robustness: Defense against lp and non-lp adversarial attacks. In Advances in Neural Information Processing Systems, pages 3487-3498, 2020. 1, 3", + "[23] Jiang Liu, Chun Pong Lau, Hossein Souri, Soheil Feizi, and Rama Chellappa. Mutual adversarial training: Learning together is better than going alone. IEEE Transactions on Information Forensics and Security, 17:2364-2377, 2022. 1", + "[24] Jiang Liu, Chen Wei, Yuxiang Guo, Heng Yu, Alan Yuille, Soheil Feizi, Chun Pong Lau, and Rama Chellappa. Instruct2attack: Language-guided semantic adversarial attacks. arXiv preprint arXiv:2311.15551, 2023. 1", + "[25] Yiming Liu, Kezhao Liu, Yao Xiao, Ziyi Dong, Xiaogang Xu, Pengxu Wei, and Liang Lin. Towards better adversaries" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24339", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ial purification via adversarial denoising diffusion training. arXiv preprint arXiv:2404.14309, 2024. 4", + "[26] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 7", + "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 6", + "[28] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2837-2845, 2021. 3", + "[29] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv e-prints, pages arXiv-2310, 2023. 3, 4", + "[30] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv e-prints, pages arXiv-2311, 2023. 3, 4", + "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 3, 6", + "[32] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Animashree Anandkumar. Diffusion models for adversarial purification. In International Conference on Machine Learning, pages 16805-16827, 2022. 1, 2, 3, 6, 7", + "[33] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations, 2023. 3", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685. IEEE, 2022. 3, 5", + "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3", + "[36] Pouya Samangouei, Maya Kabbab, and Rama Chellappa. Defense-gan: Protecting classifiers against adversarial attacks using generative models. In International Conference on Learning Representations, 2018. 1, 3", + "[37] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 7", + "[38] Naman Deep Singh, Francesco Croce, and Matthias Hein. Revisiting adversarial training for imagenet: Architectures, training and generalization across threat models. Advances in Neural Information Processing Systems, 36, 2024. 6" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 3, 6", + "[40] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[41] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In International Conference on Machine Learning, pages 32211-32252, 2023. 3", + "[42] Hossein Souri, Pirazh Khorramshahi, Chun Pong Lau, Micah Goldblum, and Rama Chellappa. Identification of attack-specific signatures in adversarial examples. arXiv preprint arXiv:2110.06802, 2021. 1", + "[43] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 1", + "[44] Jinyi Wang, Zhaoyang Lyu, Dahua Lin, Bo Dai, and Hongfei Fu. Guided diffusion model for adversarial purification. arXiv preprint arXiv:2205.14969, 2022. 2, 3, 4, 6, 7", + "[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 7", + "[46] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve adversarial training. In International Conference on Machine Learning, pages 36246-36263, 2023. 1, 3", + "[47] Yan Xu, Baoyuan Wu, Fumin Shen, Yanbo Fan, Yong Zhang, Heng Tao Shen, and Wei Liu. Exact adversarial attack to image captioning via structured output learning with latent variables. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5", + "[48] Haotian Xue, Alexandre Araujo, Bin Hu, and Yongxin Chen. Diffusion-based adversarial sample generation for improved stealthiness and controllability. Advances in Neural Information Processing Systems, 36, 2024. 7", + "[49] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, pages 12062-12072, 2021. 1, 3", + "[50] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 6", + "[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3", + "[52] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7", + "[53] Shuai Zhao, Meihuizi Jia, Zhongliang Guo, Leilei Gan, Xiaoyu Xu, Xiaobao Wu, Jie Fu, Feng Yichao, Fengjun Pan, and Anh Tuan Luu. A survey of recent backdoor attacks and defenses in large language models. Transactions on Machine Learning Research, 2025. Survey Certification. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24340", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/661ff35f-282c-423b-9185-bb6d8ff5da4b_model.json b/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/661ff35f-282c-423b-9185-bb6d8ff5da4b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6f377386b7548a2e114dd89f4e8378d68a69b32e --- /dev/null +++ b/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/661ff35f-282c-423b-9185-bb6d8ff5da4b_model.json @@ -0,0 +1,2761 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.131, + 0.872, + 0.154 + ], + "angle": 0, + "content": "Instant Adversarial Purification with Adversarial Consistency Distillation" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.18, + 0.834, + 0.2 + ], + "angle": 0, + "content": "Chun Tong Lei\\(^{1}\\), Hon Ming Yam\\(^{1}\\), Zhongliang Guo\\(^{2}\\), Yifei Qian\\(^{3}\\), Chun Pong Lau\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.2, + 0.804, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)City University of Hong Kong \\(^{2}\\)University of St Andrews \\(^{3}\\)University of Nottingham" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.219, + 0.787, + 0.234 + ], + "angle": 0, + "content": "{ctlei2, hmyam4, cplau27}@cityu.edu.hk, zg34@st-andrews.ac.uk, yifei.qian@nottingham.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.302, + 0.486, + 0.741 + ], + "angle": 0, + "content": "Neural networks have revolutionized numerous fields with their exceptional performance, yet they remain susceptible to adversarial attacks through subtle perturbations. While diffusion-based purification methods like DiffPure offer promising defense mechanisms, their computational overhead presents a significant practical limitation. In this paper, we introduce One Step Control Purification (OSCP), a novel defense framework that achieves robust adversarial purification in a single Neural Function Evaluation (NFE) within diffusion models. We propose Gaussian Adversarial Noise Distillation (GAND) as the distillation objective and Controlled Adversarial Purification (CAP) as the inference pipeline, which makes OSCP demonstrate remarkable efficiency while maintaining defense efficacy. Our proposed GAND addresses a fundamental tension between consistency distillation and adversarial perturbation, bridging the gap between natural and adversarial manifolds in the latent space, while remaining computationally efficient through Parameter-Efficient Fine-Tuning (PEFT) methods such as LoRA, eliminating the high computational budget request from full parameter fine-tuning. The CAP guides the purification process through the unlearnable edge detection operator calculated by the input image as an extra prompt, effectively preventing the purified images from deviating from their original appearance when large purification steps are used. Our experimental results on ImageNet showcase OSCP's superior performance, achieving a \\(74.19\\%\\) defense success rate with merely 0.1s per purification — a 100-fold speedup compared to conventional approaches." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.772, + 0.222, + 0.788 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.798, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Deep Neural Networks (DNNs) have fundamentally transformed the landscape of computer vision, achieving unprecedented performance across diverse applications. Despite these extraordinary achievements, a critical vulnerability lurks beneath their impressive performance, one that" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.271, + 0.905, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.385, + 0.903, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.445, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Figure 1. Comparison between existing methods and our proposed approach. Our method achieves superior performance with just a single inference step, significantly reducing computational cost. Through adversarial noise-adapted U-Net fine-tuning, we demonstrate better detail preservation after denoising, as evident in the zoomed-in regions for circles (bottom). This makes OSCP an efficient and practical solution for adversarial purification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.906, + 0.581 + ], + "angle": 0, + "content": "has emerged as a paramount concern in the research community [14, 23, 31]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.582, + 0.907, + 0.809 + ], + "angle": 0, + "content": "The Achilles' heel of these sophisticated neural architectures lies in their remarkable vulnerability to adversarial attacks [24, 43]. Through the injection of meticulously engineered perturbations that are imperceptible to human observers yet devastating to model performance, adversaries can systematically manipulate DNNs into catastrophic misclassifications [15]. While this vulnerability has found constructive applications in safeguarding against AI misuse, such as protecting intellectual property and preventing unauthorized model exploitation [13], its implications extend far beyond these beneficial uses. The susceptibility to adversarial manipulation poses a critical threat to the deployment of DNNs in high-stakes domains, where system reliability and robustness are not merely desirable but essential for public safety and security [42]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In response to these security challenges, two principal defense paradigms have emerged: adversarial training [22, 46] and adversarial purification [32, 36, 49]. Adversarial training, despite its widespread adoption, exhibits a critical limitation — it requires prior knowledge of attack methods, inherently constraining its effectiveness against" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.24, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24331" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.42, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.089, + 0.904, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.242, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Figure 2. The pipeline of Our proposed OSCP. (a) the left figure shows that the adversarial images, which are crafted through intentional attacks, exhibit a shifted distribution after the diffusion process that deviates from the Standard Normal distribution. In response, our proposed GAND can learn to recover the attacked images by modeling this additional adversarial noise with LoRA. (b) the right figure illustrates the pipeline that our proposed CAP leverages non-learnable edge detection operators to guide the purification of adversarial samples, avoiding potential inductive bias introduced by neural networks. It is worth noting that our method achieves remarkable performance by just running a single U-Net inference step." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.482, + 0.415 + ], + "angle": 0, + "content": "unknown threats. In contrast, adversarial purification offers a more versatile defense framework by focusing on perturbation removal rather than attack-specific training, making it inherently more adaptable to emerging adversarial threats." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.415, + 0.483, + 0.566 + ], + "angle": 0, + "content": "Adversarial purification methods, particularly those leveraging diffusion models [17, 39], have demonstrated remarkable effectiveness in neutralizing adversarial attacks by mapping perturbed samples back to the natural distribution [32, 44]. Diffusion models, with their unique denoising capabilities and superior training stability, offer significant advantages over traditional generative approaches such as GANs. Their ability to systematically restore corrupted images through iterative denoising has established them as a compelling framework for adversarial defense." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.567, + 0.483, + 0.702 + ], + "angle": 0, + "content": "Despite their effectiveness, diffusion models face a critical limitation: their multi-step denoising process incurs substantial computational overhead, making them impractical for real-time defense scenarios. This inherent inefficiency poses a significant challenge in applications demanding rapid response times, such as real-time security systems. Consequently, there is an urgent need for efficient purification methods that can maintain robust defense while minimizing computational costs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.704, + 0.483, + 0.87 + ], + "angle": 0, + "content": "To address these challenges, we introduce One Step Control Purification (OSCP), a novel diffusion-based framework that achieves robust adversarial defense in a single inference step. At its core, OSCP first leverages consistency distillation for fast inference while explicitly addressing the fundamental discrepancy between adversarial and clean sample distributions. This insight leads to our first key innovation, the distillation objective — Gaussian Adversarial Noise Distillation (GAND). As shown in Fig. 1 and Fig. 2 (left), GAND enables efficient purification while maintaining strong defense performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our work also addresses another fundamental challenge in diffusion-based purification — the loss of semantic in-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.492 + ], + "angle": 0, + "content": "formation in large-step inference. While Latent Consistency Models offer acceleration, they often sacrifice image quality, resulting in semantic degradation and blurry outputs. To preserve both efficiency and fidelity, as shown in Fig. 2 (right), we propose Controlled Adversarial Purification (CAP) as the inference pipeline to empower GAND, which integrates ControlNet with non-learnable edge detection operators, enabling high-quality purification while maintaining semantic integrity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.909, + 0.633 + ], + "angle": 0, + "content": "Our proposed OSCP framework, composed of GAND as the distillation objective and CAP as the inference pipeline, achieves state-of-the-art performance with \\(74.19\\%\\) robust accuracy on ImageNet while requiring merely 0.1s per purification. This breakthrough in efficiency and effectiveness not only advances the field of adversarial defense but also enables the practical deployment of robust neural networks in real-world, time-sensitive applications." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.652, + 0.862, + 0.667 + ], + "angle": 0, + "content": "In summary, our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.687, + 0.905, + 0.76 + ], + "angle": 0, + "content": "- We propose GAND, a novel consistency distillation objective for adversarial training on Latent Consistency Models. Our empirical results demonstrate its exceptional robustness and transferability against unknown adversarial attacks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.762, + 0.905, + 0.822 + ], + "angle": 0, + "content": "- We introduce CAP, an innovative inference pipeline that leverages non-learnable edge detection operators to enhance the controllability and semantic preservation of adversarial purification." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.822, + 0.905, + 0.898 + ], + "angle": 0, + "content": "- We develop OSCP, an integrated framework combining GAND and CAP, that achieves real-time adversarial purification while maintaining robust defense performance, making diffusion-based defenses practical for time-critical applications." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.687, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24332" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.234, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.285, + 0.132 + ], + "angle": 0, + "content": "2.1. Adversarial training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.484, + 0.32 + ], + "angle": 0, + "content": "Adversarial training [53] has established itself as a cornerstone defense strategy against adversarial attacks by incorporating perturbed examples into the training process [21]. Numerous studies have shown its effectiveness [2, 11], showing significant improvements in model robustness against known attack patterns. However, this approach exhibits an inherent limitation: models tend to overfit known attacks, compromising their resilience against novel attack vectors [22]. Recent advances attempt to address this limitation by leveraging diffusion models to generate diverse adversarial samples [12, 46], aiming to enhance generalization and prevent overfitting." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.329, + 0.314, + 0.344 + ], + "angle": 0, + "content": "2.2. Adversarial Purification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.485, + 0.549 + ], + "angle": 0, + "content": "Adversarial purification has evolved significantly in its approach to defending against attacks through input restoration. Initially pioneered with GAN-based methods [36], the field underwent a paradigm shift with the advent of diffusion models [17, 39, 40]. These diffusion-based approaches have demonstrated superior purification capabilities [32, 44, 49]. Notably, DiffPure [32] simultaneously removes adversarial perturbations and Gaussian noise from the forward process, theoretically justified by the reduced KL divergence between clean and adversarial image distributions. Despite these advances, the computational overhead of multiple denoising steps remains a critical bottleneck [44], limiting practical deployment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.558, + 0.26, + 0.573 + ], + "angle": 0, + "content": "2.3. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.484, + 0.763 + ], + "angle": 0, + "content": "Diffusion models have demonstrated remarkable capabilities across diverse domains, including text-to-image synthesis [34, 35], video generation [3, 18], and 3D content creation [28, 33]. The foundational Denoising Diffusion Probabilistic Models (DDPM) [17] established a two-phase framework: a forward process that gradually applies Gaussian noise following Markov properties, and a reverse process that learns to reconstruct the original image through a reverse Markov chain. However, the sequential nature of this process results in prohibitively long inference times, as computational cost scales linearly with the number of denoising steps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.329, + 0.787 + ], + "angle": 0, + "content": "2.4. Efficient Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To address the computational bottleneck, Consistency Models [41] introduced a novel consistency training paradigm, enabling image generation in significantly fewer steps. This advancement led to the development of Latent Consistency Models (LCM) [29], which further accelerated the generation process. LCM-LoRA [30] enhanced efficiency through Low-Rank Adaptation, a Parameter-Efficient Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "(PEFT) [19] approach that substantially reduces computational requirements. Additionally, ControlNet [51] introduced flexible conditioning mechanisms for Stable Diffusion, enabling precise control over generation through various visual guides including edge maps, pose estimation, sketches, and depth information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.197, + 0.652, + 0.212 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.222, + 0.673, + 0.236 + ], + "angle": 0, + "content": "3.1. Diffusion model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.245, + 0.907, + 0.364 + ], + "angle": 0, + "content": "Denoising Diffusion Probabilistic Models (DDPM) [17] generate images by learning from the reverse Markov chain with Gaussian noise added to the original image. The forward process can be formulated as linear combination of original image \\( \\pmb{x}_0 \\) and standard Gaussian noise \\( \\pmb{\\epsilon} \\), \\( \\bar{\\alpha}_t \\) denoted cumulative product from \\( \\alpha_1 \\) to \\( \\alpha_t \\), \\( \\alpha_t = 1 - \\beta_t \\) for any \\( t \\), \\( \\beta_t \\) is predefined variance schedule of diffusion process:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.369, + 0.907, + 0.386 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, \\quad \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.389, + 0.803, + 0.404 + ], + "angle": 0, + "content": "and the denoising step can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.406, + 0.907, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {x}} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta} ^ {*}} \\left(\\boldsymbol {x} _ {t}, t\\right)\\right) + \\sqrt {\\beta_ {t}} \\boldsymbol {\\epsilon}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.906, + 0.471 + ], + "angle": 0, + "content": "where model parameter \\(\\theta^{*}\\) minimize the loss between actual noise and predict noise:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.473, + 0.906, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} ^ {*} = \\arg \\min _ {\\boldsymbol {\\theta}} \\mathbb {E} _ {\\boldsymbol {x} _ {0}, t, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\right]. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.506, + 0.81, + 0.521 + ], + "angle": 0, + "content": "3.2. Diffusion-Base Purification (DBP)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.529, + 0.906, + 0.621 + ], + "angle": 0, + "content": "DiffPure [32] proposes that diffusion models can remove adversarial noise by performing a sub-process of the normal reverse process (\\(t = 0\\) to \\(t = T\\)). By adding predefined \\(t^*\\) (\\(t^* < T\\)) of noise to the adversarial image \\(\\boldsymbol{x}_{adv}\\), which is formed by the sum of original image \\(\\boldsymbol{x}\\) and adversarial noise \\(\\delta\\), \\(\\delta\\) can be generated by \\(L_p\\) attack [31] or AutoAttack [6]:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.623, + 0.906, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\delta} = \\arg \\max _ {\\delta} \\mathcal {L} (C (\\boldsymbol {x} + \\boldsymbol {\\delta}), y), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.906, + 0.677 + ], + "angle": 0, + "content": "where \\(C\\) is the classifier, \\(y\\) is the true label. The forward process of diffusion purification method uses:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.68, + 0.905, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} \\left(t ^ {*}\\right) = \\sqrt {\\bar {\\alpha} \\left(t ^ {*}\\right)} \\boldsymbol {x} + \\sqrt {1 - \\bar {\\alpha} \\left(t ^ {*}\\right)} \\boldsymbol {\\epsilon}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.906, + 0.76 + ], + "angle": 0, + "content": "and solve the reverse process of DDPM from time step \\( t^* \\) to 0, to get the purified \\( \\hat{\\pmb{x}}_{adv}^{0} \\) that is close to the original image \\( \\pmb{x} \\), allowing the classifier to classify the image with the correct label." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.696, + 0.787 + ], + "angle": 0, + "content": "3.3. Consistency Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.885 + ], + "angle": 0, + "content": "Diffusion models are known to have long inference time, which limits their usage in the real world. To tackle this situation, Consistency model [41] has been proposed, which makes it possible to generate images within 2 to 4 U-Net inference steps, by distilling a Consistency Model from a pretrianed diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Due to the efficiency of latent space models compared" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24333" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "to pixel-based models, Latent Consistency Model [29] has been proposed, making use of pretrained encoder and decoder to transform images from pixel space to latent space. Latent Consistency Model's definition is" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.155, + 0.483, + 0.173 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t\\right) = f _ {\\theta} \\left(\\boldsymbol {z} _ {t ^ {\\prime}}, t ^ {\\prime}\\right) \\quad \\forall t, t ^ {\\prime} \\in [ 0, T ] \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.175, + 0.306, + 0.19 + ], + "angle": 0, + "content": "and \\(f_{\\theta}\\) can be parameterized as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.192, + 0.484, + 0.237 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} (\\boldsymbol {z}, \\boldsymbol {c}, t) = c _ {\\mathrm {s k i p}} (t) \\boldsymbol {z} + c _ {\\mathrm {o u t}} (t) \\left(\\frac {\\boldsymbol {z} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {\\epsilon} _ {\\theta} (\\boldsymbol {z} , \\boldsymbol {c} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.238, + 0.484, + 0.314 + ], + "angle": 0, + "content": "where \\( c_{\\mathrm{skip}}(t) \\) and \\( c_{\\mathrm{out}}(t) \\) are differentiable functions such that \\( c_{\\mathrm{skip}}(\\epsilon) = 1 \\) and \\( c_{\\mathrm{out}}(\\epsilon) = 0 \\). Distilled with LoRA is introduced in LCM-LoRA [30], which dramatically reduces training time and computation cost in distillation by hugely reducing trainable parameters." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.329, + 0.182, + 0.344 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.484, + 0.415 + ], + "angle": 0, + "content": "In this paper, we aim to solve the slow purification problem in DBP by leveraging the Latent Consistency Distillation to speed up the purification backbone, enabling single-step adversarial purification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.55 + ], + "angle": 0, + "content": "Recognizing that adversarial noise differs from the Gaussian noise that diffusion models are typically designed to remove, we propose Gaussian Adversarial Noise Distillation (GAND). This novel distillation objective specifically targets adversarial noise, enhancing the purification performance. Our approach builds upon the insight that combining adversarial purification with adversarial training can yield superior results [25], effectively addressing the distinct distributions of Gaussian and adversarial noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.643 + ], + "angle": 0, + "content": "However, diffusion models tend to produce images deviating from the originals when purification steps \\( t^* \\) are large [44]. To address this, we introduce Controlled Adversarial Purification (CAP), shown in Fig. 2 (right), which utilizes the unlearnable edge detection operator computed on the input to guide the purification process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.357, + 0.669 + ], + "angle": 0, + "content": "4.1. One Step Control Purification" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.677, + 0.446, + 0.692 + ], + "angle": 0, + "content": "Problem Definition Our goal can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.704, + 0.483, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {x} _ {g t} \\simeq \\hat {\\boldsymbol {x}} _ {a d v} ^ {0} = \\mathcal {D} (f (\\mathcal {E} (\\boldsymbol {x} _ {a d v}), t ^ {*})) \\tag {8} \\\\ \\mathrm {s . t .} \\quad C (\\boldsymbol {x} _ {g t}) = C (\\hat {\\boldsymbol {x}} _ {a d v} ^ {0}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.825 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_{gt}\\) is groundtruth image, \\(\\pmb{x}_{adv}\\) is the adversarial image, \\(f\\) is a model with the purification function in latent space, \\(t^*\\) is predefined purification step, \\(\\mathcal{E}\\) and \\(\\mathcal{D}\\) are pretrained image encoder and decoder respectively. For simplifying symbolic, we denote \\(\\mathcal{E}(\\pmb{x}_{adv})\\) as \\(\\pmb{z}_{adv}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "An overview of our method OSCP is illustrated in Fig. 2; OSCP can be separated into two components: a) training a sped-up backbone model with noise and denoise function; b) pipeline for purification process with non-text guidance. We propose Gaussian Adversarial Noise Distillation" + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.092, + 0.656, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 GAND" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.111, + 0.905, + 0.186 + ], + "angle": 0, + "content": "Require: dataset \\(\\mathcal{X}\\), class label of the image \\(\\pmb{y}\\), classifier \\(C\\), LCM \\(f_{\\theta}\\) and its parameter \\(\\theta\\), ODE solver \\(\\Psi\\) and distance metric \\(d\\). \\(\\mathcal{L}_C\\) and \\(\\mathcal{L}_G\\) indicate \\(\\mathcal{L}_{CIG}\\) (Eq. (14)) and \\(\\mathcal{L}_{GAND}\\) (Eq. (15)) respectively. \\(t^\\prime\\) denote time step \\(t_{n + k}\\) in diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.187, + 0.716, + 0.2 + ], + "angle": 0, + "content": "1: while not convergence do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.201, + 0.81, + 0.216 + ], + "angle": 0, + "content": "2: Sample \\(\\pmb{x} \\sim \\mathcal{X}\\), \\(n \\sim \\mathcal{U}[1, (N - k) / 2]\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.217, + 0.627, + 0.231 + ], + "angle": 0, + "content": "3: \\( \\pmb{z} = \\mathcal{E}(\\pmb{x}) \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.232, + 0.901, + 0.247 + ], + "angle": 0, + "content": "4: \\(\\pmb{\\delta}_{adv} = \\arg \\max_{\\delta} L(C(\\mathcal{D}(\\pmb{z} + \\pmb{\\delta})), \\pmb{y})\\) Eq. (12)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.247, + 0.901, + 0.262 + ], + "angle": 0, + "content": "5: \\(z_{t^{\\prime}}^{*} = \\sqrt{\\bar{\\alpha}_{t^{\\prime}}} z + \\sqrt{1 - \\bar{\\alpha}_{t^{\\prime}}} (\\epsilon +\\delta_{adv})\\) Eq. (11)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.263, + 0.901, + 0.277 + ], + "angle": 0, + "content": "6: \\(\\hat{z}_{t_n}^{\\Psi}\\gets z_{t'}^* +\\Psi (z_{t'}^*,t',t_n,\\emptyset)\\) Eq. (13)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.277, + 0.875, + 0.293 + ], + "angle": 0, + "content": "7: \\(\\mathcal{L}_G(\\pmb {\\theta},\\pmb {\\theta}^{-})\\gets d(f_\\theta (\\pmb{z}_{t'}^*,\\varnothing ,t'),f_{\\theta^{-}}(\\hat{z}_{t_n}^\\Psi ,\\varnothing ,t_n))\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.293, + 0.754, + 0.307 + ], + "angle": 0, + "content": "8: \\(\\mathcal{L}_C(\\pmb {\\theta})\\gets d(f_\\theta (\\pmb {z}_{t'}^*,\\varnothing ,t'),\\pmb {z})\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.308, + 0.901, + 0.323 + ], + "angle": 0, + "content": "9: \\(\\mathcal{L}_T(\\pmb {\\theta},\\pmb{\\theta}^-)\\gets \\mathcal{L}_G(\\pmb {\\theta},\\pmb{\\theta}^-) + \\lambda_{CIG}\\mathcal{L}_C(\\pmb {\\theta})\\) Eq. (16)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.323, + 0.719, + 0.337 + ], + "angle": 0, + "content": "10: \\(\\pmb{\\theta} \\gets \\pmb{\\theta} - \\eta \\nabla_{\\theta} \\mathcal{L}(\\pmb{\\theta}, \\pmb{\\theta}^{-})\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.338, + 0.614, + 0.351 + ], + "angle": 0, + "content": "11: end while" + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.201, + 0.901, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.356, + 0.905, + 0.387 + ], + "angle": 0, + "content": "(GAND) and Controlled Adversarial Purification (CAP) to respond to those two unique requirements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.398, + 0.852, + 0.413 + ], + "angle": 0, + "content": "4.2. Gaussian Adversarial Noise Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.905, + 0.511 + ], + "angle": 0, + "content": "To respond to the first component we defined, we propose GAND as illustrated in Fig. 2 (left) and detailed in Alg. 1, a novel adversarial distillation objective, which combines adversarial training, Latent Consistency Distillation, and learning how to remove the adversarial and Gaussian noise from the shifted standard Gaussian distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.905, + 0.556 + ], + "angle": 0, + "content": "The original latent consistency distillation first encodes the image to latent code by encoder \\(\\mathcal{E}\\), such that \\(z = \\mathcal{E}(\\boldsymbol{x})\\) and trains the model base on loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.558, + 0.905, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\theta , \\theta^ {-}; \\Psi) = \\tag {9} \\\\ \\mathbb {E} _ {\\boldsymbol {z}, \\boldsymbol {c}, n} [ d (f _ {\\theta} (\\boldsymbol {z} _ {t _ {n + k}}, \\boldsymbol {c}, t _ {n + k}), f _ {\\theta^ {-}} (\\hat {\\boldsymbol {z}} _ {t _ {n}} ^ {\\Psi , \\omega}, \\boldsymbol {c}, t _ {n})) ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.674 + ], + "angle": 0, + "content": "where \\(d(\\cdot, \\cdot)\\) is distance metric, \\(\\Psi(\\cdot, \\cdot, \\cdot, \\cdot)\\) is DDIM PF-ODE solver \\(\\Psi_{\\mathrm{DDIM}}\\) [29], \\(f(\\cdot, \\cdot, \\cdot)\\) is latent consistency function in E.q. 7 and \\(\\hat{z}_{t_n}^{\\Psi, \\omega}\\) is an estimation of the solution of the PF-ODE from \\(t_{n+k} \\to t_n\\) using the DDIM PF-ODE solver \\(\\Psi\\):" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.675, + 0.905, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {z} _ {t _ {n}} ^ {\\Psi , \\omega} \\leftarrow z _ {t _ {n + k}} + (1 + \\omega) \\Psi \\left(z _ {t _ {n + k}}, t _ {n + k}, t _ {n}, c\\right) \\tag {10} \\\\ - \\omega \\Psi \\left(\\mathbf {z} _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \\varnothing\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.716, + 0.885, + 0.731 + ], + "angle": 0, + "content": "\\(\\omega\\) is guidance scale and it is sampled from \\([\\omega_{\\mathrm{min}},\\omega_{\\mathrm{max}}]\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.906, + 0.882 + ], + "angle": 0, + "content": "We aim to learn the denoise trajectory from a combination of Gaussian noise and adversarial noise to natural image; we need a model satisfying that \\( f_{\\theta}(z_{adv}(t), \\emptyset, t) = f_{\\theta}(z(t), \\emptyset, t) = z_{\\epsilon} \\) for all \\( t \\). \\( z_{\\epsilon} \\) is the limit of \\( z_{t} \\) when \\( t \\to 0 \\) where \\( z_{\\epsilon} \\simeq z \\). However, if we look closely at the formulation of \\( f_{\\theta} \\) in Eq. (7), \\( f_{\\theta}(z_{adv}(t), \\emptyset, t) \\) is converging to \\( z_{adv} \\) when \\( t \\to 0 \\). \\( f_{\\theta}(z_{adv}(t), \\emptyset, t) - f_{\\theta}(z(t), \\emptyset, t) \\to \\delta_{adv} \\) when \\( t \\to 0 \\), where \\( \\delta_{adv} = z_{adv} - z \\). This violates the definition of a Latent Consistency Model mentioned in Eq. (6). To address this disharmony, we introduce \\( z_{t}^{*} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.885, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {t} ^ {*} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {z} + \\sqrt {1 - \\bar {\\alpha} _ {t}} (\\boldsymbol {\\epsilon} + \\boldsymbol {\\delta} _ {a d v}), \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24334" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "by borrowing the idea of forward diffusion process, making \\( z_{t}^{*} \\) be an linear combination of \\( z \\) and \\( z_{adv} \\), worth to mention that these \\( \\delta_{adv} \\) used in GAND are generated from latent space attack." + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.155, + 0.483, + 0.177 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\delta} _ {a d v} = \\arg \\max _ {\\delta} \\mathcal {L} \\left(C \\left(\\mathcal {D} (\\mathcal {E} (\\boldsymbol {x}) + \\boldsymbol {\\delta})\\right), \\boldsymbol {y}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.179, + 0.483, + 0.284 + ], + "angle": 0, + "content": "The \\(z_{t}^{*}\\rightarrow z\\) when \\(t\\to 0\\) , and \\(z_{t}^{*}\\rightarrow \\epsilon +\\delta_{adv}\\) when \\(t\\rightarrow T\\) which converges to the natural image when \\(t\\) is small and matches with our goal, removing the adversarial noise and Gaussian noise concurrently from the shifted normal distribution. Hence, we eliminate disharmony between adversarial training and the original latent consistency distillation objective." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.484, + 0.318 + ], + "angle": 0, + "content": "Then, as we do not give condition embeddings, we remove the \\(\\omega\\) in the formulation of estimation \\(\\hat{z}_{t_n}^{\\Psi,\\omega}\\) to be:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.32, + 0.483, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {t _ {n}} ^ {\\Psi} \\leftarrow z _ {t _ {n + k}} + \\Psi \\left(z _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \\varnothing\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.34, + 0.388, + 0.355 + ], + "angle": 0, + "content": "where \\(\\varnothing\\) means null conditional embedding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.484, + 0.461 + ], + "angle": 0, + "content": "Furthermore, our goal is purifying images instead of image generation, we can train our LCM in a weaker constraint, where we only need the consistency function satisfying on \\([0,t], t < T\\). Hence, we only simulate the \\(n\\) of time step \\(t_n\\) in \\(\\mathcal{U}[1, (N - k) / 2]\\). This means that our LCM only satisfies the consistency function in the first half of the time steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.522 + ], + "angle": 0, + "content": "In addition, inspired by [20], we add a Clear Image Guide (CIG) loss to further ensure our distillation process is training towards the proposed purification goal. CIG loss is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.526, + 0.483, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C I G} (\\boldsymbol {\\theta}) = \\mathbb {E} _ {z, n} [ d (f _ {\\theta} (\\boldsymbol {z} _ {t _ {n + k}} ^ {*}, \\varnothing , t _ {n + k}), \\boldsymbol {z}) ]. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.484, + 0.573 + ], + "angle": 0, + "content": "Therefore, our total distillation loss is the combination of GAND loss" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.576, + 0.483, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {G A N D} \\left(\\boldsymbol {\\theta}, \\boldsymbol {\\theta} ^ {-}\\right) = \\tag {15} \\\\ \\mathbb {E} _ {z, n} \\left[ d \\left(f _ {\\theta} \\left(\\boldsymbol {z} _ {t _ {n + k}} ^ {*}, \\varnothing , t _ {n + k}\\right), f _ {\\theta^ {-}} \\left(\\hat {\\boldsymbol {z}} _ {t _ {n}} ^ {\\Psi}, \\varnothing , t _ {n}\\right)\\right) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.616, + 0.204, + 0.63 + ], + "angle": 0, + "content": "and CIG loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.634, + 0.483, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {T o t a l}} (\\boldsymbol {\\theta}, \\boldsymbol {\\theta} ^ {-}) = \\mathcal {L} _ {\\text {G A N D}} (\\boldsymbol {\\theta}, \\boldsymbol {\\theta} ^ {-}) + \\lambda_ {C I G} \\mathcal {L} _ {C I G} (\\boldsymbol {\\theta}), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.654, + 0.464, + 0.668 + ], + "angle": 0, + "content": "\\(\\lambda_{CIG}\\) is used to ensure two losses are on the same scale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.484, + 0.76 + ], + "angle": 0, + "content": "The following theorem shows if the loss of GAND converges to an arbitrarily small number, the difference between the purified image and the clean image will be arbitrarily small. We can use this as guidance on how to choose our purification \\( t \\) in the inference step since our method is reaching the theoretical optimal performance at a specific \\( t \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.4, + 0.787 + ], + "angle": 0, + "content": "4.3. Controlled Adversarial Purification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To address the first component we defined, we propose CAP (as shown in Fig. 3) - a purification inference pipeline that utilizes visual prompts instead of text guidance. We chose visual prompts because text-based guidance can be vulnerable to caption semantic attacks [47], which may compromise the accuracy of generated text prompts. Therefore, we opt for a more traditional and robust approach by using an" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.092, + 0.903, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.23, + 0.905, + 0.272 + ], + "angle": 0, + "content": "Figure 3. CAP used edge image of the adversarial image to control our purification process, maximizing the remaining semantic information of the purified image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.905, + 0.307 + ], + "angle": 0, + "content": "unlearnable edge detection operator to generate edge guidance of the input adversarial sample." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.906, + 0.397 + ], + "angle": 0, + "content": "In purification process, we first encode the adversarial image \\( \\mathbf{z}_{adv} = \\mathcal{E}(\\mathbf{x}_{adv}) \\) to latent space using pre-trained image encoder \\( \\mathcal{E} \\) and sample a random noise \\( \\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\) in the dimension of the latent space. Then, we diffuse the \\( \\mathbf{z}_{adv} \\) with predefined strength \\( t^* \\), using forward latent diffusion process:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.401, + 0.905, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {a d v} \\left(t ^ {*}\\right) = \\sqrt {\\bar {\\alpha} \\left(t ^ {*}\\right)} \\boldsymbol {z} _ {a d v} + \\sqrt {1 - \\bar {\\alpha} \\left(t ^ {*}\\right)} \\epsilon . \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.906, + 0.648 + ], + "angle": 0, + "content": "Then, we purify \\( \\mathbf{z}_{adv}(t^{*}) \\) using our LCM trained by GAND (in Alg. 1), \\( f_{\\theta}(\\mathbf{z},\\mathbf{c},t) \\), \\( \\mathbf{z} \\) is image latent, \\( \\mathbf{c} \\) is the condition embedding (e.g., text, canny edge image) and \\( t \\) is time step. The Latent Consistency Model has been introduced in Eq. (7). The purified image latent comes from the latent consistency function, \\( \\hat{\\mathbf{z}}_{adv}^{0} = f_{\\theta}(\\mathbf{z}_{adv}(t^{*}),\\mathbf{c}_{ce},t^{*}) \\), where \\( \\mathbf{c}_{ce} \\) means canny edge images which are provided by edge detection operators [4]. Although ControlNet is an extra plug-in tool for Stable Diffusion, \\( \\mathbf{c}_{ce} \\) is not exactly an input of \\( f_{\\theta} \\), we treat it as condition embedding here for simplifying the equation. To further reduce the effect of the adversarial image, we remove the \\( c_{skip}(t)\\mathbf{z}_{adv}(t) \\) in LCM \\( f_{\\theta} \\) and denote this LCM as \\( f_{\\theta}^{-} \\), since this term will maintain most of the adversarial noise. Therefore, our purified image latent is actually:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.65, + 0.905, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {z} _ {a d v} ^ {0} = f _ {\\theta} ^ {-} \\left(\\mathbf {z} _ {a d v} \\left(t ^ {*}\\right), \\mathbf {c} _ {c e}, t ^ {*}\\right) \\\\ = c _ {\\text {o u t}} (t) \\left(\\frac {\\boldsymbol {z} _ {a d v} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {\\epsilon} _ {\\theta} (\\boldsymbol {z} _ {a d v} , \\boldsymbol {c} _ {c e} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right). \\tag {18} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.707, + 0.905, + 0.739 + ], + "angle": 0, + "content": "Same as setting \\( c_{skip}(t) \\equiv 0 \\). Finally, We reconstruct the purified image by the image decoder \\( \\mathcal{D} \\), \\( \\hat{\\pmb{x}}_{adv}^{0} = \\mathcal{D}(\\hat{\\pmb{z}}_{adv}^{0}) \\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.754, + 0.748, + 0.771 + ], + "angle": 0, + "content": "5. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Training setting Our adversarial latent consistency model is distilled from Stable Diffusion v1.5 [34], which utilizes \\(\\epsilon\\)-prediction [17]. We split the first 10,000 images and the remaining 40,000 images in the ImageNet [8] valset as our testing set and training set, respectively. The GAND framework is trained on our training set, with all images resized to \\(512 \\times 512\\) resolution. Training proceeds for 20,000 iterations with a batch size of 4, employing an initial learning" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24335" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.091, + 0.907, + 0.12 + ], + "angle": 0, + "content": "Table 1. Accuracy (%) results for ImageNet. The best models against different attack methods are in bold. The methods marked with * mean the data are borrowed from the original paper." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.122, + 0.892, + 0.384 + ], + "angle": 0, + "content": "
DefenseAttack MethodStandard AccuracyRobust AccuracyArchitecture
Without defenseuntargeted PGD-10080.55%0.01%ResNet-50
Without defenseAutoAttack80.55%0.00%ResNet-50
Without defenserandom-targeted PGD-4082.33%0.04%ResNet-152
Diffusion Based Purification
Wang et al. [44]untargeted PGD-10073.53%72.97%ResNet-50
Wang et al. [44]random-targeted PGD-4078.10%77.86%ResNet-152
DiffPure [32]AutoAttack75.77%73.02%ResNet-50
Adversarial Training
Amini et al. [1]*AutoAttack77.96%59.64%ConvNeXt-L
Singh et al. [38]*AutoAttack77.00%57.70%ConvNeXt-L
Hybrid
OSCP (ours)untargeted PGD-10077.63%73.89%ResNet-50
OSCP (ours)AutoAttack77.63%74.19%ResNet-50
OSCP (ours)random-targeted PGD-4079.81%78.78%ResNet-152
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.406, + 0.482, + 0.527 + ], + "angle": 0, + "content": "rate of 8e-6 and a 500-step warm-up period. For the diffusion process, we employ DDIM-Solver [39] as the PF-ODE solver \\(\\Psi\\) with a skipping step \\(k = 20\\) in Equation 13. Adversarial latents are generated using PGD-10 with \\(\\epsilon = 0.03\\) against a ResNet50 victim model, and the consistency improvement gradient weight \\(\\lambda_{CIG}\\) is set to 0.001. Notably, our experiments demonstrate that PEFT via LoRA is sufficient for achieving optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.483, + 0.771 + ], + "angle": 0, + "content": "Attack Settings We conduct our evaluations on our testing set. Our evaluation encompasses diverse architectures, spanning traditional CNNs (ResNet 50 and 152 [16], WideResNet [50]) and modern Vision Transformers (ViT-B [10], Swin-B [27]). For adversarial attacks, we employ standard \\( L_{p} \\) norm-based methods including PGD [31] and AutoAttack [6]. The PGD attacks are configured with \\( L_{\\infty} \\) bounds \\( \\gamma \\in \\{4/255, 16/255\\} \\) and corresponding step sizes \\( \\eta \\in \\{1/255, 0.025 \\cdot 16/255\\} \\), while AutoAttack uses \\( L_{\\infty} \\) bound \\( \\gamma = 4/255 \\). Here, PGD-\\( n \\) denotes PGD attack with \\( n \\) iterations. In our evaluation metrics, we define standard accuracy as performance on clean data with our defense framework, robust accuracy as performance on adversarial data with our defense, and clean accuracy as performance on clean data without any defense. We fix the random seed of every experiment to avoid the randomness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.222, + 0.8 + ], + "angle": 0, + "content": "5.1. Main Result" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our method is regarded as a hybrid method, consisting of adversarial training and purification. Hence we compare our method with both categories of the defenses. Based on the provided attack setting, our method demonstrates significant computational efficiency, completing evaluations in 3 hours on an NVIDIA F40 GPU compared to GDMP's" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.406, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Table 2. Accuracy (\\%) and attack success rate (ASR %) results in various architectures, where the attack method is PGD-100, with budget 4/255, attack step size 1/255." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.45, + 0.892, + 0.54 + ], + "angle": 0, + "content": "
ArchitectureCleanASRStandardRobust
WRN-50-282.6%100%77.6%75.2%
Vit-b-1681.6%100%78.2%71.6%
Swin-b83.6%100%79.2%77.8%
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.736 + ], + "angle": 0, + "content": "[44] 48 hours. As shown in Tab. 1, diffusion based purification achieves satisfactory robust accuracy while sacrificing standard accuracy. On the other hand, adversarial training has little standard accuracy loss but the robust accuracy is unsatisfactory. Our method achieves promising results on both standard and robust accuracy with three different attack methods. We achieve robust accuracy improvements of \\(0.92\\%\\) and \\(1.17\\%\\) against PGD-100 (\\(\\gamma = 4/255\\)) and AutoAttack respectively, while improving from \\(77.86\\%\\) to \\(78.78\\%\\) against random targeted PGD-40 (\\(\\gamma = 16/255\\)). We empirically determine \\(t^{*} = 200\\) as the optimal purification strength." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.842, + 0.758 + ], + "angle": 0, + "content": "5.2. Analysis With Different Architectures" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "While our GAND-trained LCM is initially optimized using adversarial examples generated for ResNet50, we evaluate its generalization capability across different architectures to address potential overfitting concerns. As shown in Tab. 2, our method maintains strong robust accuracy across architectures, with even the lowest performance of \\(71.6\\%\\) on \\(\\mathrm{ViT}_{b-16}\\) demonstrating effective transfer. These experiments, along with subsequent ImageNet evaluations, are conducted on a representative subset of 500 images from" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24336" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.156, + 0.09, + 0.842, + 0.107 + ], + "angle": 0, + "content": "Table 3. Robust Accuracy (%) on our method under Diff-PGD-10 attack \\( \\gamma = 8/{255}\\left( {\\eta = 2/{255}}\\right) \\) on ImageNet [8]." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.108, + 0.86, + 0.176 + ], + "angle": 0, + "content": "
Defence methodResNet-50ResNet-152WideResNet-50-2Vit-b-16Swin-bConvNeXt-b
DiffPure (t* = 100)53.8%49.4%52.2%16.6%45.1%42.9%
Ours (t* = 250)59.0%56.5%57.9%34.1%53.9%49.1%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.188, + 0.483, + 0.23 + ], + "angle": 0, + "content": "Table 4. Inference time of purification models to purify one image on an NVIDIA F40 GPU. Notably, GDP refers to Wang et al. [44], DiffPure refers to Nie et al. [32]." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.233, + 0.479, + 0.326 + ], + "angle": 0, + "content": "
MethodDatasett* ∈ [0,1000]runtime (s)
GDMPImageNet250~ 9
DiffPureImageNet150~ 11
OSCP(Ours)ImageNet20/200/500/1000~ 0.1
OSCP(Ours)CelabA-HQ20/200/500/1000~ 0.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.483, + 0.358 + ], + "angle": 0, + "content": "our previously mentioned 10,000 images test set, as we observe consistent performance between these sample sizes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.369, + 0.315, + 0.386 + ], + "angle": 0, + "content": "5.3. Defend Adaptive Attack" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.393, + 0.483, + 0.573 + ], + "angle": 0, + "content": "In Tab. 3, we test our method under version 2 of Diff-PGD attack [48], which is a SOTA attack method against diffusion based purification. Our method boost the robust accuracy in Vit-b-16 from \\(16.6\\%\\) to \\(34.1\\%\\). In ResNet-50, ResNet-152, WideResNet-50-2, the robust accuracy increase \\(5.2\\%\\), \\(7.1\\%\\) and \\(5.7\\%\\) respectively. For robust accuracy in Swin-b and ConvNeXt, our method achieve \\(53.9\\%\\) and \\(49.1\\%\\), which is \\(8.8\\%\\) and \\(6.2\\%\\) higher than DiffPure. Worth to mention that our method uses LCM model which is deterministic, and hence, Expectation over Transformation, which is usually use in other robust evaluation for DBP, is not needed here." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.585, + 0.245, + 0.601 + ], + "angle": 0, + "content": "5.4. Inference Time" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.609, + 0.483, + 0.776 + ], + "angle": 0, + "content": "Conventional diffusion-based purification methods are computationally intensive, making them impractical for real-time applications. As shown in Tab. 4, our method achieves purification in just \\( \\mathbf{0.1s} \\) for ImageNet and \\( \\mathbf{0.5s} \\) for CelebA-HQ (1024 × 1024) on a NVIDIA F40 GPU, independent of \\( t^* \\) where the inference time of other method scales up depending on \\( t^* \\). More detailed experiments can be found in the supplementary materials. Our method significantly speeds up the purification process, and this acceleration enables deployment in time-critical scenarios such as autonomous driving and purification in video." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.786, + 0.26, + 0.803 + ], + "angle": 0, + "content": "5.5. Face Recognition" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We further test our model on a subset of 1000 images from CelebA-HQ [26], choosing purification step \\( t^* \\) as 200, control scale 0.8, and utilizing the GAND weights trained on ImageNet. We defend against targeted PGD attacks that attempt to manipulate Arcface [9] (AF), FaceNet [37] (FN), and MobileFaceNet [5] (MFN) to misidentify any input face" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.188, + 0.907, + 0.228 + ], + "angle": 0, + "content": "Table 5. Robust accuracy (%) for CelebA-HQ under targeted PGD-10 \\( L_{\\infty} \\gamma (\\gamma = 4 / 255) \\), \\( \\eta = 0.5 * 5 / 255 \\). The best model is in bold." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.233, + 0.888, + 0.346 + ], + "angle": 0, + "content": "
Defense MethodAFFNMFN
Without defense0%0.3%0%
GaussianBlur (σ = 7.0)2.8%51.4%2.8%
Das et al. [7] (n = 60)17.3%84.1%27.6%
CAP(Ours)83.4%97.8%82.8%
OSCP(Ours)86.8%97.8%84.9%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.356, + 0.907, + 0.426 + ], + "angle": 0, + "content": "Table 6. IQAs for \\(\\epsilon = 8 / 255\\). \\(\\uparrow/\\downarrow\\) indicate higher/ lower value consists the better image quality. Notably, IQA \\((\\cdot)\\) indicates one of IQA we leverage, \\(x\\) refers to the clean image, \\(x_{\\mathrm{adv}}\\) refers to the adversarial sample, \\(x_{\\mathrm{ours}}\\) refers to the purified image by our method and \\(x_{\\mathrm{DiffPure}}\\) refers to purified image by DiffPure." + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.428, + 0.882, + 0.494 + ], + "angle": 0, + "content": "
LPIPS↓PSNR↑SSIM↑
IQA(x, xadv)0.154734.210.8906
IQA(x, xours)0.237024.130.7343
IQA(x, xDiffPure)0.261624.110.7155
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.907, + 0.661 + ], + "angle": 0, + "content": "as a specific target person. Our defense goal is to prevent this misidentification by ensuring the cosine similarity between the purified image embedding and the target image remains below the recognition threshold. To validate the effectiveness of ImageNet-trained GAND weights on CelebA-HQ, we compare OSCP with CAP (OSCP without GAND) as shown in Tab. 5. The GAND weights significantly improve robust accuracy from \\(83.4\\%\\) to \\(86.8\\%\\) on Arcface and from \\(82.8\\%\\) to \\(84.9\\%\\) on MobileFaceNet. Notably, both our methods (CAP and OSCP) achieve an exceptional robust accuracy of \\(97.8\\%\\) on FaceNet." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.755, + 0.683 + ], + "angle": 0, + "content": "5.6. Image Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Regarding image quality assessments, we compare our method with DiffPure using a subset of 1000 ImageNet images from our testing set. We use PSNR to reflect the overall distortion level of images; SSIM [45] to indicate how human eyes perceive image structures; LPIPS [52] to assesses perceptual similarity. Both methods are configured with \\( t^{*} = 250 \\) for fair comparison. As shown in Tab. 6, our method demonstrates superior performance across multiple metrics: LPIPS decreases from 0.2616 to 0.2370, while PSNR and SSIM improve from 24.11 to 24.13 and 0.7155 to 0.7343. These improvements consistently indicate that our purified images maintain greater fidelity to the original inputs, validating the effectiveness of our purification approach. The visualization is shown in Fig. 4." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24337" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.193, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.195, + 0.089, + 0.288, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.089, + 0.383, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.089, + 0.48, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.165, + 0.192, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.241, + 0.156, + 0.253 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.165, + 0.288, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.241, + 0.251, + 0.253 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.165, + 0.384, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.165, + 0.48, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.241, + 0.442, + 0.253 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.257, + 0.483, + 0.299 + ], + "angle": 0, + "content": "Figure 4. Visualization of the IQA experiment which compares with DiffPure and the proposed method. (a) Input image. (b) Adversarial image. (c) DiffPure. (d) Ours." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.318, + 0.457, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.451, + 0.483, + 0.493 + ], + "angle": 0, + "content": "Figure 5. Performance of our method on different \\(t^*\\) under PGD-100 \\(L_{\\infty}\\gamma\\) (\\(\\gamma = 4/255\\)), \\(\\eta = 0.01^{*}4/255\\), where we evaluate on ResNet50 on ImageNet." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.535 + ], + "angle": 0, + "content": "Table 7. Accuracy (%) under PGD-100 \\(L_{\\infty}\\gamma\\) (\\(\\gamma = 4/255\\)), \\(\\eta = 0.01*4/255\\). The best result is in **bold**, the 2nd best is **underlined**." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.537, + 0.413, + 0.643 + ], + "angle": 0, + "content": "
CAPGANDStandardRobust
XX75.0%72.4%
X✓74.0%72.6%
✓X77.0%73.2%
✓✓76.8%75.0%
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.651, + 0.255, + 0.667 + ], + "angle": 0, + "content": "5.7. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We conduct comprehensive ablation studies on 500 images subset from our testing set. First, as shown in Fig. 5, we optimize CAP's performance across different ControlNet conditioning scales, identifying 0.8 as the optimal value. Tab. 7 demonstrates the individual and combined effects of our components. The baseline LCM-LoRA already shows promising performance, indicating its potential for adversarial purification. GAND-trained weights yield marginal improvements, while CAP integration increases standard and robust accuracy by \\(2\\%\\) and \\(0.8\\%\\) respectively. The full OSCP framework (combining both) significantly improves robust accuracy from \\(72.4\\%\\) to \\(75\\%\\), validating our design choices. In Tab. 8, we analyze the impact of the \\(c_{skip}(t)\\pmb{z}_t\\) term from Eq. (18). While standard accuracy remains stable, removing this term improves robust accuracy," + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.091, + 0.906, + 0.133 + ], + "angle": 0, + "content": "Table 8. Accuracy with and without \\( c_{skip}(t)z_t \\) term in LCM in our method under PGD-100 (\\( \\gamma = 8/255 \\)), \\( \\eta = 2/255 \\) on ResNet50 on a subset of 500 images from our ImageNet testing set." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.135, + 0.907, + 0.224 + ], + "angle": 0, + "content": "
\\(c_{skip}(t)z_t\\)Acc.\\(t^* = 20\\)\\(t^* = 60\\)\\(t^* = 100\\)\\(t^* = 200\\)\\(t^* = 250\\)
XRobust43.4%56.0%59.4%66.4%68.4%
XStandard81.2%81.4%81.2%79.2%76.6%
✓Robust44.2%54.8%60.4%66.0%67.8%
✓Standard81.2%81.4%81.2%79.2%76.6%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.237, + 0.906, + 0.279 + ], + "angle": 0, + "content": "Table 9. Accuracy of our method under AutoAttack (\\(\\gamma = 8/255\\)) on ResNet50 on a subset of 1000 images from ImageNet on different sets of training time steps. (\\(t^{*} = 200\\))" + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.282, + 0.871, + 0.348 + ], + "angle": 0, + "content": "
Acc.N-k/4N-k/23(N-k)/4N-k
Robust74.0%74.3%73.7%73.6%
Standard78.8%79.4%79.1%79.1%
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.906, + 0.394 + ], + "angle": 0, + "content": "likely because \\( z_{t} \\) retains adversarial noise from the input image. This observation leads us to exclude this term from our LCM implementation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.906, + 0.485 + ], + "angle": 0, + "content": "Finally, Tab. 9 explores different training time step ranges: \\(\\mathcal{U}[1,(N - k) / 4],\\mathcal{U}[1,(N - k) / 2],\\mathcal{U}[1,3(N - k) / 4]\\), and \\(\\mathcal{U}[1,N - k]\\) (original). Results show that smaller time step ranges maintain better performance, suggesting that fine-tuning LCM with smaller time steps is sufficient for purification tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.495, + 0.763, + 0.51 + ], + "angle": 0, + "content": "6. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.806 + ], + "angle": 0, + "content": "In this paper, we present One Step Control Purification (OSCP), a novel framework that addresses two critical challenges in diffusion-based adversarial purification: computational efficiency and semantic preservation. Our approach integrates the knowledge of pre-trained Diffusion Models with the proposed Gaussian Adversarial Noise Distillation (GAND) to achieve robust defense in a single inference step. Then as the inference framework, the Controlled Adversarial Purification (CAP), utilizing the unlearnable edge detection operator, effectively preserves semantic information of the input during purification. Extensive experiments validate OSPC's effectiveness across diverse architectures and datasets. On ImageNet, we achieve \\(74.19\\%\\) robust accuracy against strong attacks while requiring only 0.1s per purification—a 100-fold speedup compared to existing methods. The framework demonstrates strong cross-domain generalization, from natural images to face recognition tasks, while our ablation studies confirm the synergistic benefits of combining GAND and CAP." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.909, + 0.897 + ], + "angle": 0, + "content": "While our work represents significant progress in real-time adversarial defense, several promising directions remain, including the adaptive purification strength mechanisms and alternative control conditions. These advances could further enhance the applicability of diffusion-based defenses in time-critical scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24338" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Sajjad Amini, Mohammadreza Teymoorianfard, Shiqing Ma, and Amir Houmansadr. Meansparse: Post-training robustness enhancement through mean-centered feature sparsification. CoRR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, pages 274-283, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.484, + 0.313 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22563-22575, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[4] John Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-8(6):679-698, 1986. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[5] Sheng Chen, Yang Liu, Xiang Gao, and Zhen Han. Mobilefacenet: Efficient cnns for accurate real-time face verification on mobile devices. In Biometric Recognition, pages 428-438. Springer International Publishing, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.483, + 0.472 + ], + "angle": 0, + "content": "[6] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International Conference on Machine Learning, pages 2206-2216, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.475, + 0.484, + 0.558 + ], + "angle": 0, + "content": "[7] Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen, Fred Hohman, Siwei Li, Li Chen, Michael E Kounavis, and Duen Horng Chau. Shield: Fast, practical defense and vaccination for deep learning usingJPEG compression. In ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 196-204, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.689, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[11] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv e-prints, pages arXiv-2010, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[13] Zhongliang Guo, Junhao Dong, Yifei Qian, Kaixuan Wang, Weiye Li, Ziheng Guo, Yuheng Wang, Yanli Li, Ognjen Arandjelovic, and Lei Fang. Artwork protection against neural style transfer using locally adaptive adversarial color attack. In ECAI 2024, pages 1414-1421. IOS Press, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[14] Zhongliang Guo, Chun Tong Lei, Lei Fang, Shuai Zhao, Yifei Qian, Jingyu Lin, Zeyu Wang, Cunjian Chen, Ognjen Arandjelovic, and Chun Pong Lau. A grey-box attack against latent diffusion model-based image editing by posterior collapse. arXiv preprint arXiv:2408.10901, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.313 + ], + "angle": 0, + "content": "[15] Zhongliang Guo, Weiye Li, Yifei Qian, Ognjen Arandjelovic, and Lei Fang. A white-box false positive adversarial attack method on contrastive loss based offline handwritten signature verification models. In International Conference on Artificial Intelligence and Statistics, pages 901-909, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.316, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.414, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[19] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.905, + 0.62 + ], + "angle": 0, + "content": "[20] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ODE trajectory of diffusion. In International Conference on Learning Representations, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[21] Chun Pong Lau, Jiang Liu, Hossein Souri, Wei-An Lin, Soheil Feizi, and Rama Chellappa. Interpolated joint space adversarial training for robust and generalizable defenses. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):13054-13067, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[22] Wei-An Lin, Chun Pong Lau, Alexander Levine, Rama Chellappa, and Soheil Feizi. Dual manifold adversarial robustness: Defense against lp and non-lp adversarial attacks. In Advances in Neural Information Processing Systems, pages 3487-3498, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[23] Jiang Liu, Chun Pong Lau, Hossein Souri, Soheil Feizi, and Rama Chellappa. Mutual adversarial training: Learning together is better than going alone. IEEE Transactions on Information Forensics and Security, 17:2364-2377, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[24] Jiang Liu, Chen Wei, Yuxiang Guo, Heng Yu, Alan Yuille, Soheil Feizi, Chun Pong Lau, and Rama Chellappa. Instruct2attack: Language-guided semantic adversarial attacks. arXiv preprint arXiv:2311.15551, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Yiming Liu, Kezhao Liu, Yao Xiao, Ziyi Dong, Xiaogang Xu, Pengxu Wei, and Liang Lin. Towards better adversaries" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24339" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.481, + 0.12 + ], + "angle": 0, + "content": "ial purification via adversarial denoising diffusion training. arXiv preprint arXiv:2404.14309, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[26] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.25, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[28] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2837-2845, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.307, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[29] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv e-prints, pages arXiv-2310, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.363, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[30] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv e-prints, pages arXiv-2311, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.483, + 0.476 + ], + "angle": 0, + "content": "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.478, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[32] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Animashree Anandkumar. Diffusion models for adversarial purification. In International Conference on Machine Learning, pages 16805-16827, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.535, + 0.482, + 0.576 + ], + "angle": 0, + "content": "[33] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.578, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685. IEEE, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[36] Pouya Samangouei, Maya Kabbab, and Rama Chellappa. Defense-gan: Protecting classifiers against adversarial attacks using generative models. In International Conference on Learning Representations, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[37] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Naman Deep Singh, Francesco Croce, and Matthias Hein. Revisiting adversarial training for imagenet: Architectures, training and generalization across threat models. Advances in Neural Information Processing Systems, 36, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[40] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[41] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In International Conference on Machine Learning, pages 32211-32252, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.22, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[42] Hossein Souri, Pirazh Khorramshahi, Chun Pong Lau, Micah Goldblum, and Rama Chellappa. Identification of attack-specific signatures in adversarial examples. arXiv preprint arXiv:2110.06802, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.904, + 0.328 + ], + "angle": 0, + "content": "[43] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.904, + 0.371 + ], + "angle": 0, + "content": "[44] Jinyi Wang, Zhaoyang Lyu, Dahua Lin, Bo Dai, and Hongfei Fu. Guided diffusion model for adversarial purification. arXiv preprint arXiv:2205.14969, 2022. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.904, + 0.427 + ], + "angle": 0, + "content": "[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.904, + 0.483 + ], + "angle": 0, + "content": "[46] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve adversarial training. In International Conference on Machine Learning, pages 36246-36263, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.485, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[47] Yan Xu, Baoyuan Wu, Fumin Shen, Yanbo Fan, Yong Zhang, Heng Tao Shen, and Wei Liu. Exact adversarial attack to image captioning via structured output learning with latent variables. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.555, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[48] Haotian Xue, Alexandre Araujo, Bin Hu, and Yongxin Chen. Diffusion-based adversarial sample generation for improved stealthiness and controllability. Advances in Neural Information Processing Systems, 36, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.61, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[49] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, pages 12062-12072, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.904, + 0.705 + ], + "angle": 0, + "content": "[50] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[52] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.833, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Shuai Zhao, Meihuizi Jia, Zhongliang Guo, Leilei Gan, Xiaoyu Xu, Xiaobao Wu, Jie Fu, Feng Yichao, Fengjun Pan, and Anh Tuan Luu. A survey of recent backdoor attacks and defenses in large language models. Transactions on Machine Learning Research, 2025. Survey Certification. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24340" + } + ] +] \ No newline at end of file diff --git a/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/661ff35f-282c-423b-9185-bb6d8ff5da4b_origin.pdf b/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/661ff35f-282c-423b-9185-bb6d8ff5da4b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..739952a13641cfb366be1a850b50a63d830f0fcb --- /dev/null +++ b/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/661ff35f-282c-423b-9185-bb6d8ff5da4b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ede262ca5ec29d14494e8e25bf50e67765f59a98ecb7b814d955dcf915c4967c +size 3499714 diff --git a/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/full.md b/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fad77043a43acd1778c43af72d99174428b6101e --- /dev/null +++ b/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/full.md @@ -0,0 +1,426 @@ +# Instant Adversarial Purification with Adversarial Consistency Distillation + +Chun Tong Lei $^{1}$ , Hon Ming Yam $^{1}$ , Zhongliang Guo $^{2}$ , Yifei Qian $^{3}$ , Chun Pong Lau $^{1*}$ + +$^{1}$ City University of Hong Kong $^{2}$ University of St Andrews $^{3}$ University of Nottingham + +{ctlei2, hmyam4, cplau27}@cityu.edu.hk, zg34@st-andrews.ac.uk, yifei.qian@nottingham.ac.uk + +# Abstract + +Neural networks have revolutionized numerous fields with their exceptional performance, yet they remain susceptible to adversarial attacks through subtle perturbations. While diffusion-based purification methods like DiffPure offer promising defense mechanisms, their computational overhead presents a significant practical limitation. In this paper, we introduce One Step Control Purification (OSCP), a novel defense framework that achieves robust adversarial purification in a single Neural Function Evaluation (NFE) within diffusion models. We propose Gaussian Adversarial Noise Distillation (GAND) as the distillation objective and Controlled Adversarial Purification (CAP) as the inference pipeline, which makes OSCP demonstrate remarkable efficiency while maintaining defense efficacy. Our proposed GAND addresses a fundamental tension between consistency distillation and adversarial perturbation, bridging the gap between natural and adversarial manifolds in the latent space, while remaining computationally efficient through Parameter-Efficient Fine-Tuning (PEFT) methods such as LoRA, eliminating the high computational budget request from full parameter fine-tuning. The CAP guides the purification process through the unlearnable edge detection operator calculated by the input image as an extra prompt, effectively preventing the purified images from deviating from their original appearance when large purification steps are used. Our experimental results on ImageNet showcase OSCP's superior performance, achieving a $74.19\%$ defense success rate with merely 0.1s per purification — a 100-fold speedup compared to conventional approaches. + +# 1. Introduction + +Deep Neural Networks (DNNs) have fundamentally transformed the landscape of computer vision, achieving unprecedented performance across diverse applications. Despite these extraordinary achievements, a critical vulnerability lurks beneath their impressive performance, one that + +![](images/9060e72e6b0fbe0a4a50907d2cce57241bd40c88a2eb66661c84d54cfbc717c3.jpg) + +![](images/9510eaf7efb66e7691b017b12feb96bf99d6c59c29975ea7a6a6180ca8cc3bd4.jpg) +Figure 1. Comparison between existing methods and our proposed approach. Our method achieves superior performance with just a single inference step, significantly reducing computational cost. Through adversarial noise-adapted U-Net fine-tuning, we demonstrate better detail preservation after denoising, as evident in the zoomed-in regions for circles (bottom). This makes OSCP an efficient and practical solution for adversarial purification. + +has emerged as a paramount concern in the research community [14, 23, 31]. + +The Achilles' heel of these sophisticated neural architectures lies in their remarkable vulnerability to adversarial attacks [24, 43]. Through the injection of meticulously engineered perturbations that are imperceptible to human observers yet devastating to model performance, adversaries can systematically manipulate DNNs into catastrophic misclassifications [15]. While this vulnerability has found constructive applications in safeguarding against AI misuse, such as protecting intellectual property and preventing unauthorized model exploitation [13], its implications extend far beyond these beneficial uses. The susceptibility to adversarial manipulation poses a critical threat to the deployment of DNNs in high-stakes domains, where system reliability and robustness are not merely desirable but essential for public safety and security [42]. + +In response to these security challenges, two principal defense paradigms have emerged: adversarial training [22, 46] and adversarial purification [32, 36, 49]. Adversarial training, despite its widespread adoption, exhibits a critical limitation — it requires prior knowledge of attack methods, inherently constraining its effectiveness against + +![](images/d13f6ed46c2324c61849c6357e587026def3804b72f1629f995ea477162c4f4d.jpg) +Figure 2. The pipeline of Our proposed OSCP. (a) the left figure shows that the adversarial images, which are crafted through intentional attacks, exhibit a shifted distribution after the diffusion process that deviates from the Standard Normal distribution. In response, our proposed GAND can learn to recover the attacked images by modeling this additional adversarial noise with LoRA. (b) the right figure illustrates the pipeline that our proposed CAP leverages non-learnable edge detection operators to guide the purification of adversarial samples, avoiding potential inductive bias introduced by neural networks. It is worth noting that our method achieves remarkable performance by just running a single U-Net inference step. + +![](images/f7550585d24860142267a8d7be6a28a7ef5562b43e323557e37b913ad8f70505.jpg) + +unknown threats. In contrast, adversarial purification offers a more versatile defense framework by focusing on perturbation removal rather than attack-specific training, making it inherently more adaptable to emerging adversarial threats. + +Adversarial purification methods, particularly those leveraging diffusion models [17, 39], have demonstrated remarkable effectiveness in neutralizing adversarial attacks by mapping perturbed samples back to the natural distribution [32, 44]. Diffusion models, with their unique denoising capabilities and superior training stability, offer significant advantages over traditional generative approaches such as GANs. Their ability to systematically restore corrupted images through iterative denoising has established them as a compelling framework for adversarial defense. + +Despite their effectiveness, diffusion models face a critical limitation: their multi-step denoising process incurs substantial computational overhead, making them impractical for real-time defense scenarios. This inherent inefficiency poses a significant challenge in applications demanding rapid response times, such as real-time security systems. Consequently, there is an urgent need for efficient purification methods that can maintain robust defense while minimizing computational costs. + +To address these challenges, we introduce One Step Control Purification (OSCP), a novel diffusion-based framework that achieves robust adversarial defense in a single inference step. At its core, OSCP first leverages consistency distillation for fast inference while explicitly addressing the fundamental discrepancy between adversarial and clean sample distributions. This insight leads to our first key innovation, the distillation objective — Gaussian Adversarial Noise Distillation (GAND). As shown in Fig. 1 and Fig. 2 (left), GAND enables efficient purification while maintaining strong defense performance. + +Our work also addresses another fundamental challenge in diffusion-based purification — the loss of semantic in- + +formation in large-step inference. While Latent Consistency Models offer acceleration, they often sacrifice image quality, resulting in semantic degradation and blurry outputs. To preserve both efficiency and fidelity, as shown in Fig. 2 (right), we propose Controlled Adversarial Purification (CAP) as the inference pipeline to empower GAND, which integrates ControlNet with non-learnable edge detection operators, enabling high-quality purification while maintaining semantic integrity. + +Our proposed OSCP framework, composed of GAND as the distillation objective and CAP as the inference pipeline, achieves state-of-the-art performance with $74.19\%$ robust accuracy on ImageNet while requiring merely 0.1s per purification. This breakthrough in efficiency and effectiveness not only advances the field of adversarial defense but also enables the practical deployment of robust neural networks in real-world, time-sensitive applications. + +In summary, our key contributions are as follows: + +- We propose GAND, a novel consistency distillation objective for adversarial training on Latent Consistency Models. Our empirical results demonstrate its exceptional robustness and transferability against unknown adversarial attacks. +- We introduce CAP, an innovative inference pipeline that leverages non-learnable edge detection operators to enhance the controllability and semantic preservation of adversarial purification. +- We develop OSCP, an integrated framework combining GAND and CAP, that achieves real-time adversarial purification while maintaining robust defense performance, making diffusion-based defenses practical for time-critical applications. + +# 2. Related Work + +# 2.1. Adversarial training + +Adversarial training [53] has established itself as a cornerstone defense strategy against adversarial attacks by incorporating perturbed examples into the training process [21]. Numerous studies have shown its effectiveness [2, 11], showing significant improvements in model robustness against known attack patterns. However, this approach exhibits an inherent limitation: models tend to overfit known attacks, compromising their resilience against novel attack vectors [22]. Recent advances attempt to address this limitation by leveraging diffusion models to generate diverse adversarial samples [12, 46], aiming to enhance generalization and prevent overfitting. + +# 2.2. Adversarial Purification + +Adversarial purification has evolved significantly in its approach to defending against attacks through input restoration. Initially pioneered with GAN-based methods [36], the field underwent a paradigm shift with the advent of diffusion models [17, 39, 40]. These diffusion-based approaches have demonstrated superior purification capabilities [32, 44, 49]. Notably, DiffPure [32] simultaneously removes adversarial perturbations and Gaussian noise from the forward process, theoretically justified by the reduced KL divergence between clean and adversarial image distributions. Despite these advances, the computational overhead of multiple denoising steps remains a critical bottleneck [44], limiting practical deployment. + +# 2.3. Diffusion Models + +Diffusion models have demonstrated remarkable capabilities across diverse domains, including text-to-image synthesis [34, 35], video generation [3, 18], and 3D content creation [28, 33]. The foundational Denoising Diffusion Probabilistic Models (DDPM) [17] established a two-phase framework: a forward process that gradually applies Gaussian noise following Markov properties, and a reverse process that learns to reconstruct the original image through a reverse Markov chain. However, the sequential nature of this process results in prohibitively long inference times, as computational cost scales linearly with the number of denoising steps. + +# 2.4. Efficient Diffusion Models + +To address the computational bottleneck, Consistency Models [41] introduced a novel consistency training paradigm, enabling image generation in significantly fewer steps. This advancement led to the development of Latent Consistency Models (LCM) [29], which further accelerated the generation process. LCM-LoRA [30] enhanced efficiency through Low-Rank Adaptation, a Parameter-Efficient Fine-Tuning + +(PEFT) [19] approach that substantially reduces computational requirements. Additionally, ControlNet [51] introduced flexible conditioning mechanisms for Stable Diffusion, enabling precise control over generation through various visual guides including edge maps, pose estimation, sketches, and depth information. + +# 3. Preliminaries + +# 3.1. Diffusion model + +Denoising Diffusion Probabilistic Models (DDPM) [17] generate images by learning from the reverse Markov chain with Gaussian noise added to the original image. The forward process can be formulated as linear combination of original image $\pmb{x}_0$ and standard Gaussian noise $\pmb{\epsilon}$ , $\bar{\alpha}_t$ denoted cumulative product from $\alpha_1$ to $\alpha_t$ , $\alpha_t = 1 - \beta_t$ for any $t$ , $\beta_t$ is predefined variance schedule of diffusion process: + +$$ +\boldsymbol {x} _ {t} = \sqrt {\bar {\alpha} _ {t}} \boldsymbol {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon}, \quad \boldsymbol {\epsilon} \sim \mathcal {N} (0, \mathbf {I}), \tag {1} +$$ + +and the denoising step can be expressed as: + +$$ +\hat {\boldsymbol {x}} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\boldsymbol {x} _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \boldsymbol {\epsilon} _ {\boldsymbol {\theta} ^ {*}} \left(\boldsymbol {x} _ {t}, t\right)\right) + \sqrt {\beta_ {t}} \boldsymbol {\epsilon}, \tag {2} +$$ + +where model parameter $\theta^{*}$ minimize the loss between actual noise and predict noise: + +$$ +\boldsymbol {\theta} ^ {*} = \arg \min _ {\boldsymbol {\theta}} \mathbb {E} _ {\boldsymbol {x} _ {0}, t, \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} (\boldsymbol {x} _ {t}, t) \| _ {2} ^ {2} \right]. \tag {3} +$$ + +# 3.2. Diffusion-Base Purification (DBP) + +DiffPure [32] proposes that diffusion models can remove adversarial noise by performing a sub-process of the normal reverse process ( $t = 0$ to $t = T$ ). By adding predefined $t^*$ ( $t^* < T$ ) of noise to the adversarial image $\boldsymbol{x}_{adv}$ , which is formed by the sum of original image $\boldsymbol{x}$ and adversarial noise $\delta$ , $\delta$ can be generated by $L_p$ attack [31] or AutoAttack [6]: + +$$ +\boldsymbol {\delta} = \arg \max _ {\delta} \mathcal {L} (C (\boldsymbol {x} + \boldsymbol {\delta}), y), \tag {4} +$$ + +where $C$ is the classifier, $y$ is the true label. The forward process of diffusion purification method uses: + +$$ +\boldsymbol {x} \left(t ^ {*}\right) = \sqrt {\bar {\alpha} \left(t ^ {*}\right)} \boldsymbol {x} + \sqrt {1 - \bar {\alpha} \left(t ^ {*}\right)} \boldsymbol {\epsilon}, \tag {5} +$$ + +and solve the reverse process of DDPM from time step $t^*$ to 0, to get the purified $\hat{\pmb{x}}_{adv}^{0}$ that is close to the original image $\pmb{x}$ , allowing the classifier to classify the image with the correct label. + +# 3.3. Consistency Model + +Diffusion models are known to have long inference time, which limits their usage in the real world. To tackle this situation, Consistency model [41] has been proposed, which makes it possible to generate images within 2 to 4 U-Net inference steps, by distilling a Consistency Model from a pretrianed diffusion model. + +Due to the efficiency of latent space models compared + +to pixel-based models, Latent Consistency Model [29] has been proposed, making use of pretrained encoder and decoder to transform images from pixel space to latent space. Latent Consistency Model's definition is + +$$ +f _ {\theta} \left(\boldsymbol {z} _ {t}, t\right) = f _ {\theta} \left(\boldsymbol {z} _ {t ^ {\prime}}, t ^ {\prime}\right) \quad \forall t, t ^ {\prime} \in [ 0, T ] \tag {6} +$$ + +and $f_{\theta}$ can be parameterized as: + +$$ +f _ {\theta} (\boldsymbol {z}, \boldsymbol {c}, t) = c _ {\mathrm {s k i p}} (t) \boldsymbol {z} + c _ {\mathrm {o u t}} (t) \left(\frac {\boldsymbol {z} - \sqrt {1 - \bar {\alpha} _ {t}} \hat {\epsilon} _ {\theta} (\boldsymbol {z} , \boldsymbol {c} , t)}{\sqrt {\bar {\alpha} _ {t}}}\right), \tag {7} +$$ + +where $c_{\mathrm{skip}}(t)$ and $c_{\mathrm{out}}(t)$ are differentiable functions such that $c_{\mathrm{skip}}(\epsilon) = 1$ and $c_{\mathrm{out}}(\epsilon) = 0$ . Distilled with LoRA is introduced in LCM-LoRA [30], which dramatically reduces training time and computation cost in distillation by hugely reducing trainable parameters. + +# 4. Method + +In this paper, we aim to solve the slow purification problem in DBP by leveraging the Latent Consistency Distillation to speed up the purification backbone, enabling single-step adversarial purification. + +Recognizing that adversarial noise differs from the Gaussian noise that diffusion models are typically designed to remove, we propose Gaussian Adversarial Noise Distillation (GAND). This novel distillation objective specifically targets adversarial noise, enhancing the purification performance. Our approach builds upon the insight that combining adversarial purification with adversarial training can yield superior results [25], effectively addressing the distinct distributions of Gaussian and adversarial noise. + +However, diffusion models tend to produce images deviating from the originals when purification steps $t^*$ are large [44]. To address this, we introduce Controlled Adversarial Purification (CAP), shown in Fig. 2 (right), which utilizes the unlearnable edge detection operator computed on the input to guide the purification process. + +# 4.1. One Step Control Purification + +Problem Definition Our goal can be formulated as: + +$$ +\begin{array}{l} \boldsymbol {x} _ {g t} \simeq \hat {\boldsymbol {x}} _ {a d v} ^ {0} = \mathcal {D} (f (\mathcal {E} (\boldsymbol {x} _ {a d v}), t ^ {*})) \tag {8} \\ \mathrm {s . t .} \quad C (\boldsymbol {x} _ {g t}) = C (\hat {\boldsymbol {x}} _ {a d v} ^ {0}), \\ \end{array} +$$ + +where $\pmb{x}_{gt}$ is groundtruth image, $\pmb{x}_{adv}$ is the adversarial image, $f$ is a model with the purification function in latent space, $t^*$ is predefined purification step, $\mathcal{E}$ and $\mathcal{D}$ are pretrained image encoder and decoder respectively. For simplifying symbolic, we denote $\mathcal{E}(\pmb{x}_{adv})$ as $\pmb{z}_{adv}$ . + +An overview of our method OSCP is illustrated in Fig. 2; OSCP can be separated into two components: a) training a sped-up backbone model with noise and denoise function; b) pipeline for purification process with non-text guidance. We propose Gaussian Adversarial Noise Distillation + +# Algorithm 1 GAND + +Require: dataset $\mathcal{X}$ , class label of the image $\pmb{y}$ , classifier $C$ , LCM $f_{\theta}$ and its parameter $\theta$ , ODE solver $\Psi$ and distance metric $d$ . $\mathcal{L}_C$ and $\mathcal{L}_G$ indicate $\mathcal{L}_{CIG}$ (Eq. (14)) and $\mathcal{L}_{GAND}$ (Eq. (15)) respectively. $t^\prime$ denote time step $t_{n + k}$ in diffusion process. + +1: while not convergence do + +2: Sample $\pmb{x} \sim \mathcal{X}$ , $n \sim \mathcal{U}[1, (N - k) / 2]$ +3: $\pmb{z} = \mathcal{E}(\pmb{x})$ +4: $\pmb{\delta}_{adv} = \arg \max_{\delta} L(C(\mathcal{D}(\pmb{z} + \pmb{\delta})), \pmb{y})$ Eq. (12) +5: $z_{t^{\prime}}^{*} = \sqrt{\bar{\alpha}_{t^{\prime}}} z + \sqrt{1 - \bar{\alpha}_{t^{\prime}}} (\epsilon +\delta_{adv})$ Eq. (11) +6: $\hat{z}_{t_n}^{\Psi}\gets z_{t'}^* +\Psi (z_{t'}^*,t',t_n,\emptyset)$ Eq. (13) +7: $\mathcal{L}_G(\pmb {\theta},\pmb {\theta}^{-})\gets d(f_\theta (\pmb{z}_{t'}^*,\varnothing ,t'),f_{\theta^{-}}(\hat{z}_{t_n}^\Psi ,\varnothing ,t_n))$ +8: $\mathcal{L}_C(\pmb {\theta})\gets d(f_\theta (\pmb {z}_{t'}^*,\varnothing ,t'),\pmb {z})$ +9: $\mathcal{L}_T(\pmb {\theta},\pmb{\theta}^-)\gets \mathcal{L}_G(\pmb {\theta},\pmb{\theta}^-) + \lambda_{CIG}\mathcal{L}_C(\pmb {\theta})$ Eq. (16) +10: $\pmb{\theta} \gets \pmb{\theta} - \eta \nabla_{\theta} \mathcal{L}(\pmb{\theta}, \pmb{\theta}^{-})$ +11: end while + +(GAND) and Controlled Adversarial Purification (CAP) to respond to those two unique requirements. + +# 4.2. Gaussian Adversarial Noise Distillation + +To respond to the first component we defined, we propose GAND as illustrated in Fig. 2 (left) and detailed in Alg. 1, a novel adversarial distillation objective, which combines adversarial training, Latent Consistency Distillation, and learning how to remove the adversarial and Gaussian noise from the shifted standard Gaussian distribution. + +The original latent consistency distillation first encodes the image to latent code by encoder $\mathcal{E}$ , such that $z = \mathcal{E}(\boldsymbol{x})$ and trains the model base on loss function: + +$$ +\begin{array}{l} \mathcal {L} (\theta , \theta^ {-}; \Psi) = \tag {9} \\ \mathbb {E} _ {\boldsymbol {z}, \boldsymbol {c}, n} [ d (f _ {\theta} (\boldsymbol {z} _ {t _ {n + k}}, \boldsymbol {c}, t _ {n + k}), f _ {\theta^ {-}} (\hat {\boldsymbol {z}} _ {t _ {n}} ^ {\Psi , \omega}, \boldsymbol {c}, t _ {n})) ], \\ \end{array} +$$ + +where $d(\cdot, \cdot)$ is distance metric, $\Psi(\cdot, \cdot, \cdot, \cdot)$ is DDIM PF-ODE solver $\Psi_{\mathrm{DDIM}}$ [29], $f(\cdot, \cdot, \cdot)$ is latent consistency function in E.q. 7 and $\hat{z}_{t_n}^{\Psi, \omega}$ is an estimation of the solution of the PF-ODE from $t_{n+k} \to t_n$ using the DDIM PF-ODE solver $\Psi$ : + +$$ +\begin{array}{l} \hat {z} _ {t _ {n}} ^ {\Psi , \omega} \leftarrow z _ {t _ {n + k}} + (1 + \omega) \Psi \left(z _ {t _ {n + k}}, t _ {n + k}, t _ {n}, c\right) \tag {10} \\ - \omega \Psi \left(\mathbf {z} _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \varnothing\right), \\ \end{array} +$$ + +$\omega$ is guidance scale and it is sampled from $[\omega_{\mathrm{min}},\omega_{\mathrm{max}}]$ + +We aim to learn the denoise trajectory from a combination of Gaussian noise and adversarial noise to natural image; we need a model satisfying that $f_{\theta}(z_{adv}(t), \emptyset, t) = f_{\theta}(z(t), \emptyset, t) = z_{\epsilon}$ for all $t$ . $z_{\epsilon}$ is the limit of $z_{t}$ when $t \to 0$ where $z_{\epsilon} \simeq z$ . However, if we look closely at the formulation of $f_{\theta}$ in Eq. (7), $f_{\theta}(z_{adv}(t), \emptyset, t)$ is converging to $z_{adv}$ when $t \to 0$ . $f_{\theta}(z_{adv}(t), \emptyset, t) - f_{\theta}(z(t), \emptyset, t) \to \delta_{adv}$ when $t \to 0$ , where $\delta_{adv} = z_{adv} - z$ . This violates the definition of a Latent Consistency Model mentioned in Eq. (6). To address this disharmony, we introduce $z_{t}^{*}$ + +$$ +\boldsymbol {z} _ {t} ^ {*} = \sqrt {\bar {\alpha} _ {t}} \boldsymbol {z} + \sqrt {1 - \bar {\alpha} _ {t}} (\boldsymbol {\epsilon} + \boldsymbol {\delta} _ {a d v}), \tag {11} +$$ + +by borrowing the idea of forward diffusion process, making $z_{t}^{*}$ be an linear combination of $z$ and $z_{adv}$ , worth to mention that these $\delta_{adv}$ used in GAND are generated from latent space attack. + +$$ +\boldsymbol {\delta} _ {a d v} = \arg \max _ {\delta} \mathcal {L} \left(C \left(\mathcal {D} (\mathcal {E} (\boldsymbol {x}) + \boldsymbol {\delta})\right), \boldsymbol {y}\right). \tag {12} +$$ + +The $z_{t}^{*}\rightarrow z$ when $t\to 0$ , and $z_{t}^{*}\rightarrow \epsilon +\delta_{adv}$ when $t\rightarrow T$ which converges to the natural image when $t$ is small and matches with our goal, removing the adversarial noise and Gaussian noise concurrently from the shifted normal distribution. Hence, we eliminate disharmony between adversarial training and the original latent consistency distillation objective. + +Then, as we do not give condition embeddings, we remove the $\omega$ in the formulation of estimation $\hat{z}_{t_n}^{\Psi,\omega}$ to be: + +$$ +\hat {z} _ {t _ {n}} ^ {\Psi} \leftarrow z _ {t _ {n + k}} + \Psi \left(z _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \varnothing\right), \tag {13} +$$ + +where $\varnothing$ means null conditional embedding. + +Furthermore, our goal is purifying images instead of image generation, we can train our LCM in a weaker constraint, where we only need the consistency function satisfying on $[0,t], t < T$ . Hence, we only simulate the $n$ of time step $t_n$ in $\mathcal{U}[1, (N - k) / 2]$ . This means that our LCM only satisfies the consistency function in the first half of the time steps. + +In addition, inspired by [20], we add a Clear Image Guide (CIG) loss to further ensure our distillation process is training towards the proposed purification goal. CIG loss is given by: + +$$ +\mathcal {L} _ {C I G} (\boldsymbol {\theta}) = \mathbb {E} _ {z, n} [ d (f _ {\theta} (\boldsymbol {z} _ {t _ {n + k}} ^ {*}, \varnothing , t _ {n + k}), \boldsymbol {z}) ]. \tag {14} +$$ + +Therefore, our total distillation loss is the combination of GAND loss + +$$ +\begin{array}{l} \mathcal {L} _ {G A N D} \left(\boldsymbol {\theta}, \boldsymbol {\theta} ^ {-}\right) = \tag {15} \\ \mathbb {E} _ {z, n} \left[ d \left(f _ {\theta} \left(\boldsymbol {z} _ {t _ {n + k}} ^ {*}, \varnothing , t _ {n + k}\right), f _ {\theta^ {-}} \left(\hat {\boldsymbol {z}} _ {t _ {n}} ^ {\Psi}, \varnothing , t _ {n}\right)\right) \right], \\ \end{array} +$$ + +and CIG loss as: + +$$ +\mathcal {L} _ {\text {T o t a l}} (\boldsymbol {\theta}, \boldsymbol {\theta} ^ {-}) = \mathcal {L} _ {\text {G A N D}} (\boldsymbol {\theta}, \boldsymbol {\theta} ^ {-}) + \lambda_ {C I G} \mathcal {L} _ {C I G} (\boldsymbol {\theta}), \tag {16} +$$ + +$\lambda_{CIG}$ is used to ensure two losses are on the same scale. + +The following theorem shows if the loss of GAND converges to an arbitrarily small number, the difference between the purified image and the clean image will be arbitrarily small. We can use this as guidance on how to choose our purification $t$ in the inference step since our method is reaching the theoretical optimal performance at a specific $t$ . + +# 4.3. Controlled Adversarial Purification + +To address the first component we defined, we propose CAP (as shown in Fig. 3) - a purification inference pipeline that utilizes visual prompts instead of text guidance. We chose visual prompts because text-based guidance can be vulnerable to caption semantic attacks [47], which may compromise the accuracy of generated text prompts. Therefore, we opt for a more traditional and robust approach by using an + +![](images/f6755442a0008389179161a9a360e07432b1789be02605ff407073e312598976.jpg) +Figure 3. CAP used edge image of the adversarial image to control our purification process, maximizing the remaining semantic information of the purified image. + +unlearnable edge detection operator to generate edge guidance of the input adversarial sample. + +In purification process, we first encode the adversarial image $\mathbf{z}_{adv} = \mathcal{E}(\mathbf{x}_{adv})$ to latent space using pre-trained image encoder $\mathcal{E}$ and sample a random noise $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ in the dimension of the latent space. Then, we diffuse the $\mathbf{z}_{adv}$ with predefined strength $t^*$ , using forward latent diffusion process: + +$$ +\boldsymbol {z} _ {a d v} \left(t ^ {*}\right) = \sqrt {\bar {\alpha} \left(t ^ {*}\right)} \boldsymbol {z} _ {a d v} + \sqrt {1 - \bar {\alpha} \left(t ^ {*}\right)} \epsilon . \tag {17} +$$ + +Then, we purify $\mathbf{z}_{adv}(t^{*})$ using our LCM trained by GAND (in Alg. 1), $f_{\theta}(\mathbf{z},\mathbf{c},t)$ , $\mathbf{z}$ is image latent, $\mathbf{c}$ is the condition embedding (e.g., text, canny edge image) and $t$ is time step. The Latent Consistency Model has been introduced in Eq. (7). The purified image latent comes from the latent consistency function, $\hat{\mathbf{z}}_{adv}^{0} = f_{\theta}(\mathbf{z}_{adv}(t^{*}),\mathbf{c}_{ce},t^{*})$ , where $\mathbf{c}_{ce}$ means canny edge images which are provided by edge detection operators [4]. Although ControlNet is an extra plug-in tool for Stable Diffusion, $\mathbf{c}_{ce}$ is not exactly an input of $f_{\theta}$ , we treat it as condition embedding here for simplifying the equation. To further reduce the effect of the adversarial image, we remove the $c_{skip}(t)\mathbf{z}_{adv}(t)$ in LCM $f_{\theta}$ and denote this LCM as $f_{\theta}^{-}$ , since this term will maintain most of the adversarial noise. Therefore, our purified image latent is actually: + +$$ +\begin{array}{l} \hat {z} _ {a d v} ^ {0} = f _ {\theta} ^ {-} \left(\mathbf {z} _ {a d v} \left(t ^ {*}\right), \mathbf {c} _ {c e}, t ^ {*}\right) \\ = c _ {\text {o u t}} (t) \left(\frac {\boldsymbol {z} _ {a d v} - \sqrt {1 - \bar {\alpha} _ {t}} \hat {\epsilon} _ {\theta} (\boldsymbol {z} _ {a d v} , \boldsymbol {c} _ {c e} , t)}{\sqrt {\bar {\alpha} _ {t}}}\right). \tag {18} \\ \end{array} +$$ + +Same as setting $c_{skip}(t) \equiv 0$ . Finally, We reconstruct the purified image by the image decoder $\mathcal{D}$ , $\hat{\pmb{x}}_{adv}^{0} = \mathcal{D}(\hat{\pmb{z}}_{adv}^{0})$ . + +# 5. Experiments and Results + +Training setting Our adversarial latent consistency model is distilled from Stable Diffusion v1.5 [34], which utilizes $\epsilon$ -prediction [17]. We split the first 10,000 images and the remaining 40,000 images in the ImageNet [8] valset as our testing set and training set, respectively. The GAND framework is trained on our training set, with all images resized to $512 \times 512$ resolution. Training proceeds for 20,000 iterations with a batch size of 4, employing an initial learning + +Table 1. Accuracy (%) results for ImageNet. The best models against different attack methods are in bold. The methods marked with * mean the data are borrowed from the original paper. + +
DefenseAttack MethodStandard AccuracyRobust AccuracyArchitecture
Without defenseuntargeted PGD-10080.55%0.01%ResNet-50
Without defenseAutoAttack80.55%0.00%ResNet-50
Without defenserandom-targeted PGD-4082.33%0.04%ResNet-152
Diffusion Based Purification
Wang et al. [44]untargeted PGD-10073.53%72.97%ResNet-50
Wang et al. [44]random-targeted PGD-4078.10%77.86%ResNet-152
DiffPure [32]AutoAttack75.77%73.02%ResNet-50
Adversarial Training
Amini et al. [1]*AutoAttack77.96%59.64%ConvNeXt-L
Singh et al. [38]*AutoAttack77.00%57.70%ConvNeXt-L
Hybrid
OSCP (ours)untargeted PGD-10077.63%73.89%ResNet-50
OSCP (ours)AutoAttack77.63%74.19%ResNet-50
OSCP (ours)random-targeted PGD-4079.81%78.78%ResNet-152
+ +rate of 8e-6 and a 500-step warm-up period. For the diffusion process, we employ DDIM-Solver [39] as the PF-ODE solver $\Psi$ with a skipping step $k = 20$ in Equation 13. Adversarial latents are generated using PGD-10 with $\epsilon = 0.03$ against a ResNet50 victim model, and the consistency improvement gradient weight $\lambda_{CIG}$ is set to 0.001. Notably, our experiments demonstrate that PEFT via LoRA is sufficient for achieving optimal performance. + +Attack Settings We conduct our evaluations on our testing set. Our evaluation encompasses diverse architectures, spanning traditional CNNs (ResNet 50 and 152 [16], WideResNet [50]) and modern Vision Transformers (ViT-B [10], Swin-B [27]). For adversarial attacks, we employ standard $L_{p}$ norm-based methods including PGD [31] and AutoAttack [6]. The PGD attacks are configured with $L_{\infty}$ bounds $\gamma \in \{4/255, 16/255\}$ and corresponding step sizes $\eta \in \{1/255, 0.025 \cdot 16/255\}$ , while AutoAttack uses $L_{\infty}$ bound $\gamma = 4/255$ . Here, PGD- $n$ denotes PGD attack with $n$ iterations. In our evaluation metrics, we define standard accuracy as performance on clean data with our defense framework, robust accuracy as performance on adversarial data with our defense, and clean accuracy as performance on clean data without any defense. We fix the random seed of every experiment to avoid the randomness. + +# 5.1. Main Result + +Our method is regarded as a hybrid method, consisting of adversarial training and purification. Hence we compare our method with both categories of the defenses. Based on the provided attack setting, our method demonstrates significant computational efficiency, completing evaluations in 3 hours on an NVIDIA F40 GPU compared to GDMP's + +Table 2. Accuracy (\%) and attack success rate (ASR %) results in various architectures, where the attack method is PGD-100, with budget 4/255, attack step size 1/255. + +
ArchitectureCleanASRStandardRobust
WRN-50-282.6%100%77.6%75.2%
Vit-b-1681.6%100%78.2%71.6%
Swin-b83.6%100%79.2%77.8%
+ +[44] 48 hours. As shown in Tab. 1, diffusion based purification achieves satisfactory robust accuracy while sacrificing standard accuracy. On the other hand, adversarial training has little standard accuracy loss but the robust accuracy is unsatisfactory. Our method achieves promising results on both standard and robust accuracy with three different attack methods. We achieve robust accuracy improvements of $0.92\%$ and $1.17\%$ against PGD-100 ( $\gamma = 4/255$ ) and AutoAttack respectively, while improving from $77.86\%$ to $78.78\%$ against random targeted PGD-40 ( $\gamma = 16/255$ ). We empirically determine $t^{*} = 200$ as the optimal purification strength. + +# 5.2. Analysis With Different Architectures + +While our GAND-trained LCM is initially optimized using adversarial examples generated for ResNet50, we evaluate its generalization capability across different architectures to address potential overfitting concerns. As shown in Tab. 2, our method maintains strong robust accuracy across architectures, with even the lowest performance of $71.6\%$ on $\mathrm{ViT}_{b-16}$ demonstrating effective transfer. These experiments, along with subsequent ImageNet evaluations, are conducted on a representative subset of 500 images from + +Table 3. Robust Accuracy (%) on our method under Diff-PGD-10 attack $\gamma = 8/{255}\left( {\eta = 2/{255}}\right)$ on ImageNet [8]. + +
Defence methodResNet-50ResNet-152WideResNet-50-2Vit-b-16Swin-bConvNeXt-b
DiffPure (t* = 100)53.8%49.4%52.2%16.6%45.1%42.9%
Ours (t* = 250)59.0%56.5%57.9%34.1%53.9%49.1%
+ +Table 4. Inference time of purification models to purify one image on an NVIDIA F40 GPU. Notably, GDP refers to Wang et al. [44], DiffPure refers to Nie et al. [32]. + +
MethodDatasett* ∈ [0,1000]runtime (s)
GDMPImageNet250~ 9
DiffPureImageNet150~ 11
OSCP(Ours)ImageNet20/200/500/1000~ 0.1
OSCP(Ours)CelabA-HQ20/200/500/1000~ 0.5
+ +our previously mentioned 10,000 images test set, as we observe consistent performance between these sample sizes. + +# 5.3. Defend Adaptive Attack + +In Tab. 3, we test our method under version 2 of Diff-PGD attack [48], which is a SOTA attack method against diffusion based purification. Our method boost the robust accuracy in Vit-b-16 from $16.6\%$ to $34.1\%$ . In ResNet-50, ResNet-152, WideResNet-50-2, the robust accuracy increase $5.2\%$ , $7.1\%$ and $5.7\%$ respectively. For robust accuracy in Swin-b and ConvNeXt, our method achieve $53.9\%$ and $49.1\%$ , which is $8.8\%$ and $6.2\%$ higher than DiffPure. Worth to mention that our method uses LCM model which is deterministic, and hence, Expectation over Transformation, which is usually use in other robust evaluation for DBP, is not needed here. + +# 5.4. Inference Time + +Conventional diffusion-based purification methods are computationally intensive, making them impractical for real-time applications. As shown in Tab. 4, our method achieves purification in just $\mathbf{0.1s}$ for ImageNet and $\mathbf{0.5s}$ for CelebA-HQ (1024 × 1024) on a NVIDIA F40 GPU, independent of $t^*$ where the inference time of other method scales up depending on $t^*$ . More detailed experiments can be found in the supplementary materials. Our method significantly speeds up the purification process, and this acceleration enables deployment in time-critical scenarios such as autonomous driving and purification in video. + +# 5.5. Face Recognition + +We further test our model on a subset of 1000 images from CelebA-HQ [26], choosing purification step $t^*$ as 200, control scale 0.8, and utilizing the GAND weights trained on ImageNet. We defend against targeted PGD attacks that attempt to manipulate Arcface [9] (AF), FaceNet [37] (FN), and MobileFaceNet [5] (MFN) to misidentify any input face + +Table 5. Robust accuracy (%) for CelebA-HQ under targeted PGD-10 $L_{\infty} \gamma (\gamma = 4 / 255)$ , $\eta = 0.5 * 5 / 255$ . The best model is in bold. + +
Defense MethodAFFNMFN
Without defense0%0.3%0%
GaussianBlur (σ = 7.0)2.8%51.4%2.8%
Das et al. [7] (n = 60)17.3%84.1%27.6%
CAP(Ours)83.4%97.8%82.8%
OSCP(Ours)86.8%97.8%84.9%
+ +Table 6. IQAs for $\epsilon = 8 / 255$ . $\uparrow/\downarrow$ indicate higher/ lower value consists the better image quality. Notably, IQA $(\cdot)$ indicates one of IQA we leverage, $x$ refers to the clean image, $x_{\mathrm{adv}}$ refers to the adversarial sample, $x_{\mathrm{ours}}$ refers to the purified image by our method and $x_{\mathrm{DiffPure}}$ refers to purified image by DiffPure. + +
LPIPS↓PSNR↑SSIM↑
IQA(x, xadv)0.154734.210.8906
IQA(x, xours)0.237024.130.7343
IQA(x, xDiffPure)0.261624.110.7155
+ +as a specific target person. Our defense goal is to prevent this misidentification by ensuring the cosine similarity between the purified image embedding and the target image remains below the recognition threshold. To validate the effectiveness of ImageNet-trained GAND weights on CelebA-HQ, we compare OSCP with CAP (OSCP without GAND) as shown in Tab. 5. The GAND weights significantly improve robust accuracy from $83.4\%$ to $86.8\%$ on Arcface and from $82.8\%$ to $84.9\%$ on MobileFaceNet. Notably, both our methods (CAP and OSCP) achieve an exceptional robust accuracy of $97.8\%$ on FaceNet. + +# 5.6. Image Quality Assessment + +Regarding image quality assessments, we compare our method with DiffPure using a subset of 1000 ImageNet images from our testing set. We use PSNR to reflect the overall distortion level of images; SSIM [45] to indicate how human eyes perceive image structures; LPIPS [52] to assesses perceptual similarity. Both methods are configured with $t^{*} = 250$ for fair comparison. As shown in Tab. 6, our method demonstrates superior performance across multiple metrics: LPIPS decreases from 0.2616 to 0.2370, while PSNR and SSIM improve from 24.11 to 24.13 and 0.7155 to 0.7343. These improvements consistently indicate that our purified images maintain greater fidelity to the original inputs, validating the effectiveness of our purification approach. The visualization is shown in Fig. 4. + +![](images/0bf874e9c64cd2ac0b736733c729e46dd27809c8f0cadc6e9baf2097b446b88f.jpg) + +![](images/9f9335b1abfeaaf6928a89c694c34289a754062db12d25ce0064989252c797b3.jpg) + +![](images/4cefdc3bdf64550a3a4bdbcad813403cc10290705f15c03107b6ee83e172a2cd.jpg) + +![](images/876422d0e6b470b9a35c3544ebcce78147441d3836ba718f6c325b006be042cb.jpg) + +![](images/2eb80cc15bc8e46f5898b03f567979ed2929e276dd718cf4ee6c1df364a4e88b.jpg) +(a) + +![](images/bf0271471374b6d6053683b36950c62a53437ec4f8e62ab167335dd5537dbef9.jpg) +(b) + +![](images/c49710e3ec5a8efabe4ec99e2ea39baacf6eac4d1484e064395764e2f9dc93d0.jpg) +Figure 4. Visualization of the IQA experiment which compares with DiffPure and the proposed method. (a) Input image. (b) Adversarial image. (c) DiffPure. (d) Ours. + +![](images/2bcf63c853ad961c7988b57cdc2f2b9bd755f9086a941507d011a8401b9090a9.jpg) +(d) + +![](images/d68a87a8c723768a763eb9d64b79f9b258edc6503a45b8b2a0c0981aec53293e.jpg) +Figure 5. Performance of our method on different $t^*$ under PGD-100 $L_{\infty}\gamma$ ( $\gamma = 4/255$ ), $\eta = 0.01^{*}4/255$ , where we evaluate on ResNet50 on ImageNet. + +Table 7. Accuracy (%) under PGD-100 $L_{\infty}\gamma$ ( $\gamma = 4/255$ ), $\eta = 0.01*4/255$ . The best result is in **bold**, the 2nd best is **underlined**. + +
CAPGANDStandardRobust
XX75.0%72.4%
X✓74.0%72.6%
✓X77.0%73.2%
✓✓76.8%75.0%
+ +# 5.7. Ablation Studies + +We conduct comprehensive ablation studies on 500 images subset from our testing set. First, as shown in Fig. 5, we optimize CAP's performance across different ControlNet conditioning scales, identifying 0.8 as the optimal value. Tab. 7 demonstrates the individual and combined effects of our components. The baseline LCM-LoRA already shows promising performance, indicating its potential for adversarial purification. GAND-trained weights yield marginal improvements, while CAP integration increases standard and robust accuracy by $2\%$ and $0.8\%$ respectively. The full OSCP framework (combining both) significantly improves robust accuracy from $72.4\%$ to $75\%$ , validating our design choices. In Tab. 8, we analyze the impact of the $c_{skip}(t)\pmb{z}_t$ term from Eq. (18). While standard accuracy remains stable, removing this term improves robust accuracy, + +Table 8. Accuracy with and without $c_{skip}(t)z_t$ term in LCM in our method under PGD-100 ( $\gamma = 8/255$ ), $\eta = 2/255$ on ResNet50 on a subset of 500 images from our ImageNet testing set. + +
\(c_{skip}(t)z_t\)Acc.\(t^* = 20\)\(t^* = 60\)\(t^* = 100\)\(t^* = 200\)\(t^* = 250\)
XRobust43.4%56.0%59.4%66.4%68.4%
XStandard81.2%81.4%81.2%79.2%76.6%
✓Robust44.2%54.8%60.4%66.0%67.8%
✓Standard81.2%81.4%81.2%79.2%76.6%
+ +Table 9. Accuracy of our method under AutoAttack ( $\gamma = 8/255$ ) on ResNet50 on a subset of 1000 images from ImageNet on different sets of training time steps. ( $t^{*} = 200$ ) + +
Acc.N-k/4N-k/23(N-k)/4N-k
Robust74.0%74.3%73.7%73.6%
Standard78.8%79.4%79.1%79.1%
+ +likely because $z_{t}$ retains adversarial noise from the input image. This observation leads us to exclude this term from our LCM implementation. + +Finally, Tab. 9 explores different training time step ranges: $\mathcal{U}[1,(N - k) / 4],\mathcal{U}[1,(N - k) / 2],\mathcal{U}[1,3(N - k) / 4]$ , and $\mathcal{U}[1,N - k]$ (original). Results show that smaller time step ranges maintain better performance, suggesting that fine-tuning LCM with smaller time steps is sufficient for purification tasks. + +# 6. Conclusion and Discussion + +In this paper, we present One Step Control Purification (OSCP), a novel framework that addresses two critical challenges in diffusion-based adversarial purification: computational efficiency and semantic preservation. Our approach integrates the knowledge of pre-trained Diffusion Models with the proposed Gaussian Adversarial Noise Distillation (GAND) to achieve robust defense in a single inference step. Then as the inference framework, the Controlled Adversarial Purification (CAP), utilizing the unlearnable edge detection operator, effectively preserves semantic information of the input during purification. Extensive experiments validate OSPC's effectiveness across diverse architectures and datasets. On ImageNet, we achieve $74.19\%$ robust accuracy against strong attacks while requiring only 0.1s per purification—a 100-fold speedup compared to existing methods. The framework demonstrates strong cross-domain generalization, from natural images to face recognition tasks, while our ablation studies confirm the synergistic benefits of combining GAND and CAP. + +While our work represents significant progress in real-time adversarial defense, several promising directions remain, including the adaptive purification strength mechanisms and alternative control conditions. These advances could further enhance the applicability of diffusion-based defenses in time-critical scenarios. + +# References + +[1] Sajjad Amini, Mohammadreza Teymoorianfard, Shiqing Ma, and Amir Houmansadr. Meansparse: Post-training robustness enhancement through mean-centered feature sparsification. CoRR, 2024. 6 +[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, pages 274-283, 2018. 3 +[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22563-22575, 2023. 3 +[4] John Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-8(6):679-698, 1986. 5 +[5] Sheng Chen, Yang Liu, Xiang Gao, and Zhen Han. Mobilefacenet: Efficient cnns for accurate real-time face verification on mobile devices. In Biometric Recognition, pages 428-438. Springer International Publishing, 2018. 7 +[6] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International Conference on Machine Learning, pages 2206-2216, 2020. 3, 6 +[7] Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen, Fred Hohman, Siwei Li, Li Chen, Michael E Kounavis, and Duen Horng Chau. Shield: Fast, practical defense and vaccination for deep learning usingJPEG compression. In ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 196-204, 2018. 7 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 5, 7 +[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 7 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6 +[11] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv e-prints, pages arXiv-2010, 2020. 3 +[12] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 3 + +[13] Zhongliang Guo, Junhao Dong, Yifei Qian, Kaixuan Wang, Weiye Li, Ziheng Guo, Yuheng Wang, Yanli Li, Ognjen Arandjelovic, and Lei Fang. Artwork protection against neural style transfer using locally adaptive adversarial color attack. In ECAI 2024, pages 1414-1421. IOS Press, 2024. 1 +[14] Zhongliang Guo, Chun Tong Lei, Lei Fang, Shuai Zhao, Yifei Qian, Jingyu Lin, Zeyu Wang, Cunjian Chen, Ognjen Arandjelovic, and Chun Pong Lau. A grey-box attack against latent diffusion model-based image editing by posterior collapse. arXiv preprint arXiv:2408.10901, 2024. 1 +[15] Zhongliang Guo, Weiye Li, Yifei Qian, Ognjen Arandjelovic, and Lei Fang. A white-box false positive adversarial attack method on contrastive loss based offline handwritten signature verification models. In International Conference on Artificial Intelligence and Statistics, pages 901-909, 2024. 1 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 5 +[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3 +[19] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799, 2019. 3 +[20] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ODE trajectory of diffusion. In International Conference on Learning Representations, 2024. 5 +[21] Chun Pong Lau, Jiang Liu, Hossein Souri, Wei-An Lin, Soheil Feizi, and Rama Chellappa. Interpolated joint space adversarial training for robust and generalizable defenses. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):13054-13067, 2023. 3 +[22] Wei-An Lin, Chun Pong Lau, Alexander Levine, Rama Chellappa, and Soheil Feizi. Dual manifold adversarial robustness: Defense against lp and non-lp adversarial attacks. In Advances in Neural Information Processing Systems, pages 3487-3498, 2020. 1, 3 +[23] Jiang Liu, Chun Pong Lau, Hossein Souri, Soheil Feizi, and Rama Chellappa. Mutual adversarial training: Learning together is better than going alone. IEEE Transactions on Information Forensics and Security, 17:2364-2377, 2022. 1 +[24] Jiang Liu, Chen Wei, Yuxiang Guo, Heng Yu, Alan Yuille, Soheil Feizi, Chun Pong Lau, and Rama Chellappa. Instruct2attack: Language-guided semantic adversarial attacks. arXiv preprint arXiv:2311.15551, 2023. 1 +[25] Yiming Liu, Kezhao Liu, Yao Xiao, Ziyi Dong, Xiaogang Xu, Pengxu Wei, and Liang Lin. Towards better adversaries + +ial purification via adversarial denoising diffusion training. arXiv preprint arXiv:2404.14309, 2024. 4 +[26] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 7 +[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 6 +[28] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2837-2845, 2021. 3 +[29] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv e-prints, pages arXiv-2310, 2023. 3, 4 +[30] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv e-prints, pages arXiv-2311, 2023. 3, 4 +[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 3, 6 +[32] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Animashree Anandkumar. Diffusion models for adversarial purification. In International Conference on Machine Learning, pages 16805-16827, 2022. 1, 2, 3, 6, 7 +[33] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations, 2023. 3 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685. IEEE, 2022. 3, 5 +[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3 +[36] Pouya Samangouei, Maya Kabbab, and Rama Chellappa. Defense-gan: Protecting classifiers against adversarial attacks using generative models. In International Conference on Learning Representations, 2018. 1, 3 +[37] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 7 +[38] Naman Deep Singh, Francesco Croce, and Matthias Hein. Revisiting adversarial training for imagenet: Architectures, training and generalization across threat models. Advances in Neural Information Processing Systems, 36, 2024. 6 + +[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 3, 6 +[40] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 3 +[41] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In International Conference on Machine Learning, pages 32211-32252, 2023. 3 +[42] Hossein Souri, Pirazh Khorramshahi, Chun Pong Lau, Micah Goldblum, and Rama Chellappa. Identification of attack-specific signatures in adversarial examples. arXiv preprint arXiv:2110.06802, 2021. 1 +[43] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 1 +[44] Jinyi Wang, Zhaoyang Lyu, Dahua Lin, Bo Dai, and Hongfei Fu. Guided diffusion model for adversarial purification. arXiv preprint arXiv:2205.14969, 2022. 2, 3, 4, 6, 7 +[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 7 +[46] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve adversarial training. In International Conference on Machine Learning, pages 36246-36263, 2023. 1, 3 +[47] Yan Xu, Baoyuan Wu, Fumin Shen, Yanbo Fan, Yong Zhang, Heng Tao Shen, and Wei Liu. Exact adversarial attack to image captioning via structured output learning with latent variables. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5 +[48] Haotian Xue, Alexandre Araujo, Bin Hu, and Yongxin Chen. Diffusion-based adversarial sample generation for improved stealthiness and controllability. Advances in Neural Information Processing Systems, 36, 2024. 7 +[49] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, pages 12062-12072, 2021. 1, 3 +[50] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 6 +[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3 +[52] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7 +[53] Shuai Zhao, Meihuizi Jia, Zhongliang Guo, Leilei Gan, Xiaoyu Xu, Xiaobao Wu, Jie Fu, Feng Yichao, Fengjun Pan, and Anh Tuan Luu. A survey of recent backdoor attacks and defenses in large language models. Transactions on Machine Learning Research, 2025. Survey Certification. 3 \ No newline at end of file diff --git a/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/images.zip b/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7d1316f566d41a92959776dd44bd9b4aaabe25b2 --- /dev/null +++ b/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:398674a13037dfb7f58d403445f26459acc81cb56fd8238fb1eb5f32e6636fc2 +size 596767 diff --git a/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/layout.json b/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1129a86b4ae49acf57ad37e704d80ca1b3719a48 --- /dev/null +++ b/2025/Instant Adversarial Purification with Adversarial Consistency Distillation/layout.json @@ -0,0 +1,12572 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "type": "text", + "content": "Instant Adversarial Purification with Adversarial Consistency Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 142, + 510, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 142, + 510, + 158 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 510, + 158 + ], + "type": "text", + "content": "Chun Tong Lei" + }, + { + "bbox": [ + 104, + 142, + 510, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 142, + 510, + 158 + ], + "type": "text", + "content": ", Hon Ming Yam" + }, + { + "bbox": [ + 104, + 142, + 510, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 142, + 510, + 158 + ], + "type": "text", + "content": ", Zhongliang Guo" + }, + { + "bbox": [ + 104, + 142, + 510, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 142, + 510, + 158 + ], + "type": "text", + "content": ", Yifei Qian" + }, + { + "bbox": [ + 104, + 142, + 510, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 104, + 142, + 510, + 158 + ], + "type": "text", + "content": ", Chun Pong Lau" + }, + { + "bbox": [ + 104, + 142, + 510, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 158, + 492, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 158, + 492, + 171 + ], + "spans": [ + { + "bbox": [ + 118, + 158, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 158, + 492, + 171 + ], + "type": "text", + "content": "City University of Hong Kong " + }, + { + "bbox": [ + 118, + 158, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 158, + 492, + 171 + ], + "type": "text", + "content": "University of St Andrews " + }, + { + "bbox": [ + 118, + 158, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 118, + 158, + 492, + 171 + ], + "type": "text", + "content": "University of Nottingham" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 173, + 481, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 173, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 127, + 173, + 481, + 185 + ], + "type": "text", + "content": "{ctlei2, hmyam4, cplau27}@cityu.edu.hk, zg34@st-andrews.ac.uk, yifei.qian@nottingham.ac.uk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 239, + 297, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 586 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 586 + ], + "type": "text", + "content": "Neural networks have revolutionized numerous fields with their exceptional performance, yet they remain susceptible to adversarial attacks through subtle perturbations. While diffusion-based purification methods like DiffPure offer promising defense mechanisms, their computational overhead presents a significant practical limitation. In this paper, we introduce One Step Control Purification (OSCP), a novel defense framework that achieves robust adversarial purification in a single Neural Function Evaluation (NFE) within diffusion models. We propose Gaussian Adversarial Noise Distillation (GAND) as the distillation objective and Controlled Adversarial Purification (CAP) as the inference pipeline, which makes OSCP demonstrate remarkable efficiency while maintaining defense efficacy. Our proposed GAND addresses a fundamental tension between consistency distillation and adversarial perturbation, bridging the gap between natural and adversarial manifolds in the latent space, while remaining computationally efficient through Parameter-Efficient Fine-Tuning (PEFT) methods such as LoRA, eliminating the high computational budget request from full parameter fine-tuning. The CAP guides the purification process through the unlearnable edge detection operator calculated by the input image as an extra prompt, effectively preventing the purified images from deviating from their original appearance when large purification steps are used. Our experimental results on ImageNet showcase OSCP's superior performance, achieving a " + }, + { + "bbox": [ + 54, + 239, + 297, + 586 + ], + "type": "inline_equation", + "content": "74.19\\%" + }, + { + "bbox": [ + 54, + 239, + 297, + 586 + ], + "type": "text", + "content": " defense success rate with merely 0.1s per purification — a 100-fold speedup compared to conventional approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 611, + 135, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 611, + 135, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 611, + 135, + 624 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": "Deep Neural Networks (DNNs) have fundamentally transformed the landscape of computer vision, achieving unprecedented performance across diverse applications. Despite these extraordinary achievements, a critical vulnerability lurks beneath their impressive performance, one that" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 214, + 553, + 297 + ], + "blocks": [ + { + "bbox": [ + 317, + 214, + 553, + 297 + ], + "lines": [ + { + "bbox": [ + 317, + 214, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 214, + 553, + 297 + ], + "type": "image", + "image_path": "9060e72e6b0fbe0a4a50907d2cce57241bd40c88a2eb66661c84d54cfbc717c3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 304, + 552, + 351 + ], + "blocks": [ + { + "bbox": [ + 317, + 304, + 552, + 351 + ], + "lines": [ + { + "bbox": [ + 317, + 304, + 552, + 351 + ], + "spans": [ + { + "bbox": [ + 317, + 304, + 552, + 351 + ], + "type": "image", + "image_path": "9510eaf7efb66e7691b017b12feb96bf99d6c59c29975ea7a6a6180ca8cc3bd4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 352, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 430 + ], + "type": "text", + "content": "Figure 1. Comparison between existing methods and our proposed approach. Our method achieves superior performance with just a single inference step, significantly reducing computational cost. Through adversarial noise-adapted U-Net fine-tuning, we demonstrate better detail preservation after denoising, as evident in the zoomed-in regions for circles (bottom). This makes OSCP an efficient and practical solution for adversarial purification." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 435, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 554, + 460 + ], + "type": "text", + "content": "has emerged as a paramount concern in the research community [14, 23, 31]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 460, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 555, + 640 + ], + "type": "text", + "content": "The Achilles' heel of these sophisticated neural architectures lies in their remarkable vulnerability to adversarial attacks [24, 43]. Through the injection of meticulously engineered perturbations that are imperceptible to human observers yet devastating to model performance, adversaries can systematically manipulate DNNs into catastrophic misclassifications [15]. While this vulnerability has found constructive applications in safeguarding against AI misuse, such as protecting intellectual property and preventing unauthorized model exploitation [13], its implications extend far beyond these beneficial uses. The susceptibility to adversarial manipulation poses a critical threat to the deployment of DNNs in high-stakes domains, where system reliability and robustness are not merely desirable but essential for public safety and security [42]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "In response to these security challenges, two principal defense paradigms have emerged: adversarial training [22, 46] and adversarial purification [32, 36, 49]. Adversarial training, despite its widespread adoption, exhibits a critical limitation — it requires prior knowledge of attack methods, inherently constraining its effectiveness against" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24331" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 257, + 190 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 257, + 190 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 257, + 190 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 257, + 190 + ], + "type": "image", + "image_path": "d13f6ed46c2324c61849c6357e587026def3804b72f1629f995ea477162c4f4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": "Figure 2. The pipeline of Our proposed OSCP. (a) the left figure shows that the adversarial images, which are crafted through intentional attacks, exhibit a shifted distribution after the diffusion process that deviates from the Standard Normal distribution. In response, our proposed GAND can learn to recover the attacked images by modeling this additional adversarial noise with LoRA. (b) the right figure illustrates the pipeline that our proposed CAP leverages non-learnable edge detection operators to guide the purification of adversarial samples, avoiding potential inductive bias introduced by neural networks. It is worth noting that our method achieves remarkable performance by just running a single U-Net inference step." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 261, + 70, + 553, + 190 + ], + "blocks": [ + { + "bbox": [ + 261, + 70, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 261, + 70, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 261, + 70, + 553, + 190 + ], + "type": "image", + "image_path": "f7550585d24860142267a8d7be6a28a7ef5562b43e323557e37b913ad8f70505.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 280, + 294, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 294, + 328 + ], + "type": "text", + "content": "unknown threats. In contrast, adversarial purification offers a more versatile defense framework by focusing on perturbation removal rather than attack-specific training, making it inherently more adaptable to emerging adversarial threats." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 328, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 328, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 295, + 448 + ], + "type": "text", + "content": "Adversarial purification methods, particularly those leveraging diffusion models [17, 39], have demonstrated remarkable effectiveness in neutralizing adversarial attacks by mapping perturbed samples back to the natural distribution [32, 44]. Diffusion models, with their unique denoising capabilities and superior training stability, offer significant advantages over traditional generative approaches such as GANs. Their ability to systematically restore corrupted images through iterative denoising has established them as a compelling framework for adversarial defense." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 449, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 449, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 295, + 555 + ], + "type": "text", + "content": "Despite their effectiveness, diffusion models face a critical limitation: their multi-step denoising process incurs substantial computational overhead, making them impractical for real-time defense scenarios. This inherent inefficiency poses a significant challenge in applications demanding rapid response times, such as real-time security systems. Consequently, there is an urgent need for efficient purification methods that can maintain robust defense while minimizing computational costs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 557, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 557, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 557, + 295, + 689 + ], + "type": "text", + "content": "To address these challenges, we introduce One Step Control Purification (OSCP), a novel diffusion-based framework that achieves robust adversarial defense in a single inference step. At its core, OSCP first leverages consistency distillation for fast inference while explicitly addressing the fundamental discrepancy between adversarial and clean sample distributions. This insight leads to our first key innovation, the distillation objective — Gaussian Adversarial Noise Distillation (GAND). As shown in Fig. 1 and Fig. 2 (left), GAND enables efficient purification while maintaining strong defense performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Our work also addresses another fundamental challenge in diffusion-based purification — the loss of semantic in-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 280, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 389 + ], + "type": "text", + "content": "formation in large-step inference. While Latent Consistency Models offer acceleration, they often sacrifice image quality, resulting in semantic degradation and blurry outputs. To preserve both efficiency and fidelity, as shown in Fig. 2 (right), we propose Controlled Adversarial Purification (CAP) as the inference pipeline to empower GAND, which integrates ControlNet with non-learnable edge detection operators, enabling high-quality purification while maintaining semantic integrity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 403, + 556, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 501 + ], + "type": "text", + "content": "Our proposed OSCP framework, composed of GAND as the distillation objective and CAP as the inference pipeline, achieves state-of-the-art performance with " + }, + { + "bbox": [ + 313, + 403, + 556, + 501 + ], + "type": "inline_equation", + "content": "74.19\\%" + }, + { + "bbox": [ + 313, + 403, + 556, + 501 + ], + "type": "text", + "content": " robust accuracy on ImageNet while requiring merely 0.1s per purification. This breakthrough in efficiency and effectiveness not only advances the field of adversarial defense but also enables the practical deployment of robust neural networks in real-world, time-sensitive applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 516, + 527, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 516, + 527, + 528 + ], + "spans": [ + { + "bbox": [ + 326, + 516, + 527, + 528 + ], + "type": "text", + "content": "In summary, our key contributions are as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 544, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 314, + 544, + 553, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 544, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 314, + 544, + 553, + 601 + ], + "type": "text", + "content": "- We propose GAND, a novel consistency distillation objective for adversarial training on Latent Consistency Models. Our empirical results demonstrate its exceptional robustness and transferability against unknown adversarial attacks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 603, + 553, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 603, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 314, + 603, + 553, + 651 + ], + "type": "text", + "content": "- We introduce CAP, an innovative inference pipeline that leverages non-learnable edge detection operators to enhance the controllability and semantic preservation of adversarial purification." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 651, + 553, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 651, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 314, + 651, + 553, + 711 + ], + "type": "text", + "content": "- We develop OSCP, an integrated framework combining GAND and CAP, that achieves real-time adversarial purification while maintaining robust defense performance, making diffusion-based defenses practical for time-critical applications." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24332" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 143, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 143, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 143, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 174, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 174, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 174, + 104 + ], + "type": "text", + "content": "2.1. Adversarial training" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 109, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 296, + 253 + ], + "type": "text", + "content": "Adversarial training [53] has established itself as a cornerstone defense strategy against adversarial attacks by incorporating perturbed examples into the training process [21]. Numerous studies have shown its effectiveness [2, 11], showing significant improvements in model robustness against known attack patterns. However, this approach exhibits an inherent limitation: models tend to overfit known attacks, compromising their resilience against novel attack vectors [22]. Recent advances attempt to address this limitation by leveraging diffusion models to generate diverse adversarial samples [12, 46], aiming to enhance generalization and prevent overfitting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 260, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 192, + 272 + ], + "type": "text", + "content": "2.2. Adversarial Purification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "type": "text", + "content": "Adversarial purification has evolved significantly in its approach to defending against attacks through input restoration. Initially pioneered with GAN-based methods [36], the field underwent a paradigm shift with the advent of diffusion models [17, 39, 40]. These diffusion-based approaches have demonstrated superior purification capabilities [32, 44, 49]. Notably, DiffPure [32] simultaneously removes adversarial perturbations and Gaussian noise from the forward process, theoretically justified by the reduced KL divergence between clean and adversarial image distributions. Despite these advances, the computational overhead of multiple denoising steps remains a critical bottleneck [44], limiting practical deployment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 441, + 159, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 159, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 159, + 453 + ], + "type": "text", + "content": "2.3. Diffusion Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 604 + ], + "type": "text", + "content": "Diffusion models have demonstrated remarkable capabilities across diverse domains, including text-to-image synthesis [34, 35], video generation [3, 18], and 3D content creation [28, 33]. The foundational Denoising Diffusion Probabilistic Models (DDPM) [17] established a two-phase framework: a forward process that gradually applies Gaussian noise following Markov properties, and a reverse process that learns to reconstruct the original image through a reverse Markov chain. However, the sequential nature of this process results in prohibitively long inference times, as computational cost scales linearly with the number of denoising steps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 611, + 201, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 201, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 201, + 623 + ], + "type": "text", + "content": "2.4. Efficient Diffusion Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "To address the computational bottleneck, Consistency Models [41] introduced a novel consistency training paradigm, enabling image generation in significantly fewer steps. This advancement led to the development of Latent Consistency Models (LCM) [29], which further accelerated the generation process. LCM-LoRA [30] enhanced efficiency through Low-Rank Adaptation, a Parameter-Efficient Fine-Tuning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "(PEFT) [19] approach that substantially reduces computational requirements. Additionally, ControlNet [51] introduced flexible conditioning mechanisms for Stable Diffusion, enabling precise control over generation through various visual guides including edge maps, pose estimation, sketches, and depth information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 156, + 399, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 399, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 399, + 167 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 175, + 411, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 411, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 411, + 186 + ], + "type": "text", + "content": "3.1. Diffusion model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "text", + "content": "Denoising Diffusion Probabilistic Models (DDPM) [17] generate images by learning from the reverse Markov chain with Gaussian noise added to the original image. The forward process can be formulated as linear combination of original image " + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "text", + "content": " and standard Gaussian noise " + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}" + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "text", + "content": " denoted cumulative product from " + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\alpha_1" + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\alpha_t" + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\alpha_t = 1 - \\beta_t" + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 313, + 194, + 555, + 288 + ], + "type": "text", + "content": " is predefined variance schedule of diffusion process:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 292, + 555, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 292, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 347, + 292, + 555, + 305 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, \\quad \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {1}", + "image_path": "46c5e8985064a8cf10e1a88c0713a9204e668c8d6934e24be92164224fe789cf.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 308, + 491, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 491, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 491, + 319 + ], + "type": "text", + "content": "and the denoising step can be expressed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 321, + 555, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 321, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 321, + 321, + 555, + 348 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {x}} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta} ^ {*}} \\left(\\boldsymbol {x} _ {t}, t\\right)\\right) + \\sqrt {\\beta_ {t}} \\boldsymbol {\\epsilon}, \\tag {2}", + "image_path": "b9181d2ed96117bd737513d51aafce6bf08ce0c346ace17c3686f7e3b7a71407.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 349, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 554, + 373 + ], + "type": "text", + "content": "where model parameter " + }, + { + "bbox": [ + 313, + 349, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 313, + 349, + 554, + 373 + ], + "type": "text", + "content": " minimize the loss between actual noise and predict noise:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 348, + 374, + 554, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 374, + 554, + 393 + ], + "spans": [ + { + "bbox": [ + 348, + 374, + 554, + 393 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} ^ {*} = \\arg \\min _ {\\boldsymbol {\\theta}} \\mathbb {E} _ {\\boldsymbol {x} _ {0}, t, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\right]. \\tag {3}", + "image_path": "2828c6e7855762cc61ab25f3be254eaa4bebe6312f3a85e710ae6819d23d9253.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 400, + 495, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 495, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 495, + 412 + ], + "type": "text", + "content": "3.2. Diffusion-Base Purification (DBP)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": "DiffPure [32] proposes that diffusion models can remove adversarial noise by performing a sub-process of the normal reverse process (" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": "). By adding predefined " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "t^* < T" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": ") of noise to the adversarial image " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{adv}" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": ", which is formed by the sum of original image " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": " and adversarial noise " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": " can be generated by " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "L_p" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": " attack [31] or AutoAttack [6]:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 371, + 493, + 554, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 493, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 371, + 493, + 554, + 510 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\delta} = \\arg \\max _ {\\delta} \\mathcal {L} (C (\\boldsymbol {x} + \\boldsymbol {\\delta}), y), \\tag {4}", + "image_path": "292f182578bd2de45888b6c7289597ceff4e374bb8d1ce0d3d23db3d3ef56c95.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 512, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 554, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 512, + 554, + 536 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 512, + 554, + 536 + ], + "type": "text", + "content": " is the classifier, " + }, + { + "bbox": [ + 313, + 512, + 554, + 536 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 512, + 554, + 536 + ], + "type": "text", + "content": " is the true label. The forward process of diffusion purification method uses:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 359, + 538, + 553, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 538, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 359, + 538, + 553, + 553 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} \\left(t ^ {*}\\right) = \\sqrt {\\bar {\\alpha} \\left(t ^ {*}\\right)} \\boldsymbol {x} + \\sqrt {1 - \\bar {\\alpha} \\left(t ^ {*}\\right)} \\boldsymbol {\\epsilon}, \\tag {5}", + "image_path": "0cecfed1d27b0c328ae9cbd843bf289807636ef171aa92a910a06e21612ab686.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 555, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 601 + ], + "type": "text", + "content": "and solve the reverse process of DDPM from time step " + }, + { + "bbox": [ + 313, + 555, + 554, + 601 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 313, + 555, + 554, + 601 + ], + "type": "text", + "content": " to 0, to get the purified " + }, + { + "bbox": [ + 313, + 555, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{adv}^{0}" + }, + { + "bbox": [ + 313, + 555, + 554, + 601 + ], + "type": "text", + "content": " that is close to the original image " + }, + { + "bbox": [ + 313, + 555, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 555, + 554, + 601 + ], + "type": "text", + "content": ", allowing the classifier to classify the image with the correct label." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 611, + 425, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 425, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 425, + 623 + ], + "type": "text", + "content": "3.3. Consistency Model" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": "Diffusion models are known to have long inference time, which limits their usage in the real world. To tackle this situation, Consistency model [41] has been proposed, which makes it possible to generate images within 2 to 4 U-Net inference steps, by distilling a Consistency Model from a pretrianed diffusion model." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 325, + 701, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 554, + 713 + ], + "type": "text", + "content": "Due to the efficiency of latent space models compared" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24333" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "to pixel-based models, Latent Consistency Model [29] has been proposed, making use of pretrained encoder and decoder to transform images from pixel space to latent space. Latent Consistency Model's definition is" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 122, + 295, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 122, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 97, + 122, + 295, + 137 + ], + "type": "interline_equation", + "content": "f _ {\\theta} \\left(\\boldsymbol {z} _ {t}, t\\right) = f _ {\\theta} \\left(\\boldsymbol {z} _ {t ^ {\\prime}}, t ^ {\\prime}\\right) \\quad \\forall t, t ^ {\\prime} \\in [ 0, T ] \\tag {6}", + "image_path": "27d5b8a77f6b91b9a08cc1ca7e1ecb68554e57330aadac13ecbfd338085f2f5b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 138, + 187, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 187, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 187, + 150 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 55, + 138, + 187, + 150 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 138, + 187, + 150 + ], + "type": "text", + "content": " can be parameterized as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 152, + 296, + 187 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 296, + 187 + ], + "type": "interline_equation", + "content": "f _ {\\theta} (\\boldsymbol {z}, \\boldsymbol {c}, t) = c _ {\\mathrm {s k i p}} (t) \\boldsymbol {z} + c _ {\\mathrm {o u t}} (t) \\left(\\frac {\\boldsymbol {z} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {\\epsilon} _ {\\theta} (\\boldsymbol {z} , \\boldsymbol {c} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right), \\tag {7}", + "image_path": "4d162c9b5ed31ed6446d4674c4d0bc59ae6e613127c45697ee6d537dd7a2486b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{skip}}(t)" + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{out}}(t)" + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": " are differentiable functions such that " + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{skip}}(\\epsilon) = 1" + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{out}}(\\epsilon) = 0" + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": ". Distilled with LoRA is introduced in LCM-LoRA [30], which dramatically reduces training time and computation cost in distillation by hugely reducing trainable parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 260, + 111, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 111, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 111, + 272 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 281, + 296, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 328 + ], + "type": "text", + "content": "In this paper, we aim to solve the slow purification problem in DBP by leveraging the Latent Consistency Distillation to speed up the purification backbone, enabling single-step adversarial purification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 329, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 435 + ], + "type": "text", + "content": "Recognizing that adversarial noise differs from the Gaussian noise that diffusion models are typically designed to remove, we propose Gaussian Adversarial Noise Distillation (GAND). This novel distillation objective specifically targets adversarial noise, enhancing the purification performance. Our approach builds upon the insight that combining adversarial purification with adversarial training can yield superior results [25], effectively addressing the distinct distributions of Gaussian and adversarial noise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 437, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 509 + ], + "type": "text", + "content": "However, diffusion models tend to produce images deviating from the originals when purification steps " + }, + { + "bbox": [ + 55, + 437, + 296, + 509 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 55, + 437, + 296, + 509 + ], + "type": "text", + "content": " are large [44]. To address this, we introduce Controlled Adversarial Purification (CAP), shown in Fig. 2 (right), which utilizes the unlearnable edge detection operator computed on the input to guide the purification process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 517, + 218, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 218, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 218, + 529 + ], + "type": "text", + "content": "4.1. One Step Control Purification" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 536, + 272, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 272, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 272, + 548 + ], + "type": "text", + "content": "Problem Definition Our goal can be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 557, + 295, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 557, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 104, + 557, + 295, + 588 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {x} _ {g t} \\simeq \\hat {\\boldsymbol {x}} _ {a d v} ^ {0} = \\mathcal {D} (f (\\mathcal {E} (\\boldsymbol {x} _ {a d v}), t ^ {*})) \\tag {8} \\\\ \\mathrm {s . t .} \\quad C (\\boldsymbol {x} _ {g t}) = C (\\hat {\\boldsymbol {x}} _ {a d v} ^ {0}), \\\\ \\end{array}", + "image_path": "37668f9dadde8163ca83183baa97346ae71adbbdd1766886034e16499e81ab9d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{gt}" + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": " is groundtruth image, " + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{adv}" + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": " is the adversarial image, " + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": " is a model with the purification function in latent space, " + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": " is predefined purification step, " + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": " are pretrained image encoder and decoder respectively. For simplifying symbolic, we denote " + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\pmb{x}_{adv})" + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{adv}" + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "An overview of our method OSCP is illustrated in Fig. 2; OSCP can be separated into two components: a) training a sped-up backbone model with noise and denoise function; b) pipeline for purification process with non-text guidance. We propose Gaussian Adversarial Noise Distillation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 72, + 401, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 401, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 401, + 83 + ], + "type": "text", + "content": "Algorithm 1 GAND" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "spans": [ + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": "Require: dataset " + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": ", class label of the image " + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": ", classifier " + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": ", LCM " + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": " and its parameter " + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": ", ODE solver " + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": " and distance metric " + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_G" + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": " indicate " + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CIG}" + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": " (Eq. (14)) and " + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GAND}" + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": " (Eq. (15)) respectively. " + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "t^\\prime" + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": " denote time step " + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "t_{n + k}" + }, + { + "bbox": [ + 315, + 87, + 553, + 147 + ], + "type": "text", + "content": " in diffusion process." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 148, + 438, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 148, + 438, + 158 + ], + "spans": [ + { + "bbox": [ + 321, + 148, + 438, + 158 + ], + "type": "text", + "content": "1: while not convergence do" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 159, + 551, + 277 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 321, + 159, + 495, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 159, + 495, + 171 + ], + "spans": [ + { + "bbox": [ + 321, + 159, + 495, + 171 + ], + "type": "text", + "content": "2: Sample " + }, + { + "bbox": [ + 321, + 159, + 495, + 171 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\sim \\mathcal{X}" + }, + { + "bbox": [ + 321, + 159, + 495, + 171 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 321, + 159, + 495, + 171 + ], + "type": "inline_equation", + "content": "n \\sim \\mathcal{U}[1, (N - k) / 2]" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 171, + 383, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 171, + 383, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 171, + 383, + 182 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 321, + 171, + 383, + 182 + ], + "type": "inline_equation", + "content": "\\pmb{z} = \\mathcal{E}(\\pmb{x})" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 183, + 551, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 183, + 551, + 195 + ], + "spans": [ + { + "bbox": [ + 321, + 183, + 551, + 195 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 321, + 183, + 551, + 195 + ], + "type": "inline_equation", + "content": "\\pmb{\\delta}_{adv} = \\arg \\max_{\\delta} L(C(\\mathcal{D}(\\pmb{z} + \\pmb{\\delta})), \\pmb{y})" + }, + { + "bbox": [ + 321, + 183, + 551, + 195 + ], + "type": "text", + "content": " Eq. (12)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 195, + 551, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 195, + 551, + 207 + ], + "spans": [ + { + "bbox": [ + 321, + 195, + 551, + 207 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 321, + 195, + 551, + 207 + ], + "type": "inline_equation", + "content": "z_{t^{\\prime}}^{*} = \\sqrt{\\bar{\\alpha}_{t^{\\prime}}} z + \\sqrt{1 - \\bar{\\alpha}_{t^{\\prime}}} (\\epsilon +\\delta_{adv})" + }, + { + "bbox": [ + 321, + 195, + 551, + 207 + ], + "type": "text", + "content": " Eq. (11)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 208, + 551, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 208, + 551, + 219 + ], + "spans": [ + { + "bbox": [ + 321, + 208, + 551, + 219 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 321, + 208, + 551, + 219 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t_n}^{\\Psi}\\gets z_{t'}^* +\\Psi (z_{t'}^*,t',t_n,\\emptyset)" + }, + { + "bbox": [ + 321, + 208, + 551, + 219 + ], + "type": "text", + "content": " Eq. (13)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 219, + 535, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 219, + 535, + 232 + ], + "spans": [ + { + "bbox": [ + 321, + 219, + 535, + 232 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 321, + 219, + 535, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_G(\\pmb {\\theta},\\pmb {\\theta}^{-})\\gets d(f_\\theta (\\pmb{z}_{t'}^*,\\varnothing ,t'),f_{\\theta^{-}}(\\hat{z}_{t_n}^\\Psi ,\\varnothing ,t_n))" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 232, + 461, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 232, + 461, + 243 + ], + "spans": [ + { + "bbox": [ + 321, + 232, + 461, + 243 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 321, + 232, + 461, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C(\\pmb {\\theta})\\gets d(f_\\theta (\\pmb {z}_{t'}^*,\\varnothing ,t'),\\pmb {z})" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 243, + 551, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 243, + 551, + 255 + ], + "spans": [ + { + "bbox": [ + 321, + 243, + 551, + 255 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 321, + 243, + 551, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_T(\\pmb {\\theta},\\pmb{\\theta}^-)\\gets \\mathcal{L}_G(\\pmb {\\theta},\\pmb{\\theta}^-) + \\lambda_{CIG}\\mathcal{L}_C(\\pmb {\\theta})" + }, + { + "bbox": [ + 321, + 243, + 551, + 255 + ], + "type": "text", + "content": " Eq. (16)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 255, + 440, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 255, + 440, + 266 + ], + "spans": [ + { + "bbox": [ + 318, + 255, + 440, + 266 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 318, + 255, + 440, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} \\gets \\pmb{\\theta} - \\eta \\nabla_{\\theta} \\mathcal{L}(\\pmb{\\theta}, \\pmb{\\theta}^{-})" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 267, + 375, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 267, + 375, + 277 + ], + "spans": [ + { + "bbox": [ + 318, + 267, + 375, + 277 + ], + "type": "text", + "content": "11: end while" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 281, + 553, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 306 + ], + "type": "text", + "content": "(GAND) and Controlled Adversarial Purification (CAP) to respond to those two unique requirements." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 315, + 521, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 521, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 521, + 327 + ], + "type": "text", + "content": "4.2. Gaussian Adversarial Noise Distillation" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 334, + 553, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 553, + 404 + ], + "type": "text", + "content": "To respond to the first component we defined, we propose GAND as illustrated in Fig. 2 (left) and detailed in Alg. 1, a novel adversarial distillation objective, which combines adversarial training, Latent Consistency Distillation, and learning how to remove the adversarial and Gaussian noise from the shifted standard Gaussian distribution." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 406, + 553, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 553, + 440 + ], + "type": "text", + "content": "The original latent consistency distillation first encodes the image to latent code by encoder " + }, + { + "bbox": [ + 313, + 406, + 553, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 406, + 553, + 440 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 313, + 406, + 553, + 440 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 406, + 553, + 440 + ], + "type": "text", + "content": " and trains the model base on loss function:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 322, + 441, + 553, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 441, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 322, + 441, + 553, + 472 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\theta , \\theta^ {-}; \\Psi) = \\tag {9} \\\\ \\mathbb {E} _ {\\boldsymbol {z}, \\boldsymbol {c}, n} [ d (f _ {\\theta} (\\boldsymbol {z} _ {t _ {n + k}}, \\boldsymbol {c}, t _ {n + k}), f _ {\\theta^ {-}} (\\hat {\\boldsymbol {z}} _ {t _ {n}} ^ {\\Psi , \\omega}, \\boldsymbol {c}, t _ {n})) ], \\\\ \\end{array}", + "image_path": "762a512bb4408f298b5461101d9e773ce55e70d26d1b2d990d6c6da1158de6be.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "text", + "content": " is distance metric, " + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\Psi(\\cdot, \\cdot, \\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "text", + "content": " is DDIM PF-ODE solver " + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{DDIM}}" + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "text", + "content": " [29], " + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "inline_equation", + "content": "f(\\cdot, \\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "text", + "content": " is latent consistency function in E.q. 7 and " + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t_n}^{\\Psi, \\omega}" + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "text", + "content": " is an estimation of the solution of the PF-ODE from " + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "inline_equation", + "content": "t_{n+k} \\to t_n" + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "text", + "content": " using the DDIM PF-ODE solver " + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 474, + 554, + 533 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 324, + 534, + 553, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 534, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 324, + 534, + 553, + 564 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {z} _ {t _ {n}} ^ {\\Psi , \\omega} \\leftarrow z _ {t _ {n + k}} + (1 + \\omega) \\Psi \\left(z _ {t _ {n + k}}, t _ {n + k}, t _ {n}, c\\right) \\tag {10} \\\\ - \\omega \\Psi \\left(\\mathbf {z} _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \\varnothing\\right), \\\\ \\end{array}", + "image_path": "e5e16e581b684158b3271f96f584e379aa22e4ebf9d0d7844f5039d539e82d5e.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 567, + 541, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 541, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 541, + 578 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 567, + 541, + 578 + ], + "type": "text", + "content": " is guidance scale and it is sampled from " + }, + { + "bbox": [ + 313, + 567, + 541, + 578 + ], + "type": "inline_equation", + "content": "[\\omega_{\\mathrm{min}},\\omega_{\\mathrm{max}}]" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": "We aim to learn the denoise trajectory from a combination of Gaussian noise and adversarial noise to natural image; we need a model satisfying that " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "f_{\\theta}(z_{adv}(t), \\emptyset, t) = f_{\\theta}(z(t), \\emptyset, t) = z_{\\epsilon}" + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "z_{\\epsilon}" + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": " is the limit of " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "t \\to 0" + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "z_{\\epsilon} \\simeq z" + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": ". However, if we look closely at the formulation of " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": " in Eq. (7), " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "f_{\\theta}(z_{adv}(t), \\emptyset, t)" + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": " is converging to " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "z_{adv}" + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "t \\to 0" + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "f_{\\theta}(z_{adv}(t), \\emptyset, t) - f_{\\theta}(z(t), \\emptyset, t) \\to \\delta_{adv}" + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "t \\to 0" + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "\\delta_{adv} = z_{adv} - z" + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "text", + "content": ". This violates the definition of a Latent Consistency Model mentioned in Eq. (6). To address this disharmony, we introduce " + }, + { + "bbox": [ + 313, + 579, + 554, + 698 + ], + "type": "inline_equation", + "content": "z_{t}^{*}" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 361, + 700, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 361, + 700, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {t} ^ {*} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {z} + \\sqrt {1 - \\bar {\\alpha} _ {t}} (\\boldsymbol {\\epsilon} + \\boldsymbol {\\delta} _ {a d v}), \\tag {11}", + "image_path": "3ab5dbc8e0b722f8fb3e2829d177e89a179c739b1ee22eba6edcfa553275e573.jpg" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24334" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "by borrowing the idea of forward diffusion process, making " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "z_{t}^{*}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " be an linear combination of " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "z_{adv}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ", worth to mention that these " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\delta_{adv}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " used in GAND are generated from latent space attack." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 122, + 295, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 122, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 91, + 122, + 295, + 140 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\delta} _ {a d v} = \\arg \\max _ {\\delta} \\mathcal {L} \\left(C \\left(\\mathcal {D} (\\mathcal {E} (\\boldsymbol {x}) + \\boldsymbol {\\delta})\\right), \\boldsymbol {y}\\right). \\tag {12}", + "image_path": "1f9d722ad85cb21b6551972d94070c4892b39c75b32c95745ca1b993f85b8f39.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 141, + 295, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 141, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 295, + 224 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 55, + 141, + 295, + 224 + ], + "type": "inline_equation", + "content": "z_{t}^{*}\\rightarrow z" + }, + { + "bbox": [ + 55, + 141, + 295, + 224 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 55, + 141, + 295, + 224 + ], + "type": "inline_equation", + "content": "t\\to 0" + }, + { + "bbox": [ + 55, + 141, + 295, + 224 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 55, + 141, + 295, + 224 + ], + "type": "inline_equation", + "content": "z_{t}^{*}\\rightarrow \\epsilon +\\delta_{adv}" + }, + { + "bbox": [ + 55, + 141, + 295, + 224 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 55, + 141, + 295, + 224 + ], + "type": "inline_equation", + "content": "t\\rightarrow T" + }, + { + "bbox": [ + 55, + 141, + 295, + 224 + ], + "type": "text", + "content": " which converges to the natural image when " + }, + { + "bbox": [ + 55, + 141, + 295, + 224 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 141, + 295, + 224 + ], + "type": "text", + "content": " is small and matches with our goal, removing the adversarial noise and Gaussian noise concurrently from the shifted normal distribution. Hence, we eliminate disharmony between adversarial training and the original latent consistency distillation objective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 226, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 251 + ], + "type": "text", + "content": "Then, as we do not give condition embeddings, we remove the " + }, + { + "bbox": [ + 55, + 226, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 226, + 296, + 251 + ], + "type": "text", + "content": " in the formulation of estimation " + }, + { + "bbox": [ + 55, + 226, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t_n}^{\\Psi,\\omega}" + }, + { + "bbox": [ + 55, + 226, + 296, + 251 + ], + "type": "text", + "content": " to be:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 253, + 295, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 253, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 96, + 253, + 295, + 268 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {t _ {n}} ^ {\\Psi} \\leftarrow z _ {t _ {n + k}} + \\Psi \\left(z _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \\varnothing\\right), \\tag {13}", + "image_path": "217092e4991592b02409273aca2c9dbf7980e2ac5fa512bcfe59b16da458d83e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 269, + 237, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 237, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 237, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 269, + 237, + 281 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 55, + 269, + 237, + 281 + ], + "type": "text", + "content": " means null conditional embedding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "type": "text", + "content": "Furthermore, our goal is purifying images instead of image generation, we can train our LCM in a weaker constraint, where we only need the consistency function satisfying on " + }, + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "type": "inline_equation", + "content": "[0,t], t < T" + }, + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "type": "text", + "content": ". Hence, we only simulate the " + }, + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "type": "text", + "content": " of time step " + }, + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{U}[1, (N - k) / 2]" + }, + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "type": "text", + "content": ". This means that our LCM only satisfies the consistency function in the first half of the time steps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 365, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 413 + ], + "type": "text", + "content": "In addition, inspired by [20], we add a Clear Image Guide (CIG) loss to further ensure our distillation process is training towards the proposed purification goal. CIG loss is given by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 416, + 295, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 416, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 77, + 416, + 295, + 430 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C I G} (\\boldsymbol {\\theta}) = \\mathbb {E} _ {z, n} [ d (f _ {\\theta} (\\boldsymbol {z} _ {t _ {n + k}} ^ {*}, \\varnothing , t _ {n + k}), \\boldsymbol {z}) ]. \\tag {14}", + "image_path": "71ce584a25d071b41181301cef06fb6c6ed1231a85519d66397a449244a28aae.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 431, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 296, + 453 + ], + "type": "text", + "content": "Therefore, our total distillation loss is the combination of GAND loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 456, + 295, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 456, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 61, + 456, + 295, + 486 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {G A N D} \\left(\\boldsymbol {\\theta}, \\boldsymbol {\\theta} ^ {-}\\right) = \\tag {15} \\\\ \\mathbb {E} _ {z, n} \\left[ d \\left(f _ {\\theta} \\left(\\boldsymbol {z} _ {t _ {n + k}} ^ {*}, \\varnothing , t _ {n + k}\\right), f _ {\\theta^ {-}} \\left(\\hat {\\boldsymbol {z}} _ {t _ {n}} ^ {\\Psi}, \\varnothing , t _ {n}\\right)\\right) \\right], \\\\ \\end{array}", + "image_path": "0a5c45f64958141b7fe0b847693f3470771e79fc0712d9eb7b4538dad50561cb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 487, + 124, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 124, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 124, + 498 + ], + "type": "text", + "content": "and CIG loss as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 60, + 502, + 295, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 502, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 60, + 502, + 295, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {T o t a l}} (\\boldsymbol {\\theta}, \\boldsymbol {\\theta} ^ {-}) = \\mathcal {L} _ {\\text {G A N D}} (\\boldsymbol {\\theta}, \\boldsymbol {\\theta} ^ {-}) + \\lambda_ {C I G} \\mathcal {L} _ {C I G} (\\boldsymbol {\\theta}), \\tag {16}", + "image_path": "dd7d6287ad8532a1ed1c7e1cea8c809671b1c992b9bd52da34a5b376bd6e950c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 517, + 283, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 283, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 283, + 529 + ], + "type": "inline_equation", + "content": "\\lambda_{CIG}" + }, + { + "bbox": [ + 55, + 517, + 283, + 529 + ], + "type": "text", + "content": " is used to ensure two losses are on the same scale." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 530, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 296, + 601 + ], + "type": "text", + "content": "The following theorem shows if the loss of GAND converges to an arbitrarily small number, the difference between the purified image and the clean image will be arbitrarily small. We can use this as guidance on how to choose our purification " + }, + { + "bbox": [ + 55, + 530, + 296, + 601 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 530, + 296, + 601 + ], + "type": "text", + "content": " in the inference step since our method is reaching the theoretical optimal performance at a specific " + }, + { + "bbox": [ + 55, + 530, + 296, + 601 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 530, + 296, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 610, + 244, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 244, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 244, + 623 + ], + "type": "text", + "content": "4.3. Controlled Adversarial Purification" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "To address the first component we defined, we propose CAP (as shown in Fig. 3) - a purification inference pipeline that utilizes visual prompts instead of text guidance. We chose visual prompts because text-based guidance can be vulnerable to caption semantic attacks [47], which may compromise the accuracy of generated text prompts. Therefore, we opt for a more traditional and robust approach by using an" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 552, + 180 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 552, + 180 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 552, + 180 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 552, + 180 + ], + "type": "image", + "image_path": "f6755442a0008389179161a9a360e07432b1789be02605ff407073e312598976.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 182, + 553, + 215 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 215 + ], + "type": "text", + "content": "Figure 3. CAP used edge image of the adversarial image to control our purification process, maximizing the remaining semantic information of the purified image." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 219, + 553, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 553, + 243 + ], + "type": "text", + "content": "unlearnable edge detection operator to generate edge guidance of the input adversarial sample." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 243, + 554, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 554, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 554, + 314 + ], + "type": "text", + "content": "In purification process, we first encode the adversarial image " + }, + { + "bbox": [ + 313, + 243, + 554, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{adv} = \\mathcal{E}(\\mathbf{x}_{adv})" + }, + { + "bbox": [ + 313, + 243, + 554, + 314 + ], + "type": "text", + "content": " to latent space using pre-trained image encoder " + }, + { + "bbox": [ + 313, + 243, + 554, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 243, + 554, + 314 + ], + "type": "text", + "content": " and sample a random noise " + }, + { + "bbox": [ + 313, + 243, + 554, + 314 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 243, + 554, + 314 + ], + "type": "text", + "content": " in the dimension of the latent space. Then, we diffuse the " + }, + { + "bbox": [ + 313, + 243, + 554, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{adv}" + }, + { + "bbox": [ + 313, + 243, + 554, + 314 + ], + "type": "text", + "content": " with predefined strength " + }, + { + "bbox": [ + 313, + 243, + 554, + 314 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 313, + 243, + 554, + 314 + ], + "type": "text", + "content": ", using forward latent diffusion process:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 348, + 317, + 553, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 317, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 348, + 317, + 553, + 332 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {a d v} \\left(t ^ {*}\\right) = \\sqrt {\\bar {\\alpha} \\left(t ^ {*}\\right)} \\boldsymbol {z} _ {a d v} + \\sqrt {1 - \\bar {\\alpha} \\left(t ^ {*}\\right)} \\epsilon . \\tag {17}", + "image_path": "730a1891fc9a5d05b563c1b86e109303f0bde9df7138cafbe3dcedcfa5290b07.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "content": "Then, we purify " + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{adv}(t^{*})" + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "content": " using our LCM trained by GAND (in Alg. 1), " + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{z},\\mathbf{c},t)" + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "content": " is image latent, " + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "content": " is the condition embedding (e.g., text, canny edge image) and " + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "content": " is time step. The Latent Consistency Model has been introduced in Eq. (7). The purified image latent comes from the latent consistency function, " + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{adv}^{0} = f_{\\theta}(\\mathbf{z}_{adv}(t^{*}),\\mathbf{c}_{ce},t^{*})" + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{ce}" + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "content": " means canny edge images which are provided by edge detection operators [4]. Although ControlNet is an extra plug-in tool for Stable Diffusion, " + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{ce}" + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "content": " is not exactly an input of " + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "content": ", we treat it as condition embedding here for simplifying the equation. To further reduce the effect of the adversarial image, we remove the " + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "inline_equation", + "content": "c_{skip}(t)\\mathbf{z}_{adv}(t)" + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "content": " in LCM " + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "content": " and denote this LCM as " + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{-}" + }, + { + "bbox": [ + 313, + 334, + 554, + 513 + ], + "type": "text", + "content": ", since this term will maintain most of the adversarial noise. Therefore, our purified image latent is actually:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 514, + 553, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 514, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 320, + 514, + 553, + 558 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {z} _ {a d v} ^ {0} = f _ {\\theta} ^ {-} \\left(\\mathbf {z} _ {a d v} \\left(t ^ {*}\\right), \\mathbf {c} _ {c e}, t ^ {*}\\right) \\\\ = c _ {\\text {o u t}} (t) \\left(\\frac {\\boldsymbol {z} _ {a d v} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {\\epsilon} _ {\\theta} (\\boldsymbol {z} _ {a d v} , \\boldsymbol {c} _ {c e} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right). \\tag {18} \\\\ \\end{array}", + "image_path": "c5be35a694adaf10d7dc4498bdd6a78a776665060c4ba237dbc24f15e2b1d34d.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 559, + 553, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 553, + 585 + ], + "type": "text", + "content": "Same as setting " + }, + { + "bbox": [ + 313, + 559, + 553, + 585 + ], + "type": "inline_equation", + "content": "c_{skip}(t) \\equiv 0" + }, + { + "bbox": [ + 313, + 559, + 553, + 585 + ], + "type": "text", + "content": ". Finally, We reconstruct the purified image by the image decoder " + }, + { + "bbox": [ + 313, + 559, + 553, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 559, + 553, + 585 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 559, + 553, + 585 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{adv}^{0} = \\mathcal{D}(\\hat{\\pmb{z}}_{adv}^{0})" + }, + { + "bbox": [ + 313, + 559, + 553, + 585 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 597, + 457, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 457, + 610 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 457, + 610 + ], + "type": "text", + "content": "5. Experiments and Results" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "Training setting Our adversarial latent consistency model is distilled from Stable Diffusion v1.5 [34], which utilizes " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "-prediction [17]. We split the first 10,000 images and the remaining 40,000 images in the ImageNet [8] valset as our testing set and training set, respectively. The GAND framework is trained on our training set, with all images resized to " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": " resolution. Training proceeds for 20,000 iterations with a batch size of 4, employing an initial learning" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24335" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 96, + 545, + 304 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 555, + 95 + ], + "type": "text", + "content": "Table 1. Accuracy (%) results for ImageNet. The best models against different attack methods are in bold. The methods marked with * mean the data are borrowed from the original paper." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 96, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 63, + 96, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 63, + 96, + 545, + 304 + ], + "type": "table", + "html": "
DefenseAttack MethodStandard AccuracyRobust AccuracyArchitecture
Without defenseuntargeted PGD-10080.55%0.01%ResNet-50
Without defenseAutoAttack80.55%0.00%ResNet-50
Without defenserandom-targeted PGD-4082.33%0.04%ResNet-152
Diffusion Based Purification
Wang et al. [44]untargeted PGD-10073.53%72.97%ResNet-50
Wang et al. [44]random-targeted PGD-4078.10%77.86%ResNet-152
DiffPure [32]AutoAttack75.77%73.02%ResNet-50
Adversarial Training
Amini et al. [1]*AutoAttack77.96%59.64%ConvNeXt-L
Singh et al. [38]*AutoAttack77.00%57.70%ConvNeXt-L
Hybrid
OSCP (ours)untargeted PGD-10077.63%73.89%ResNet-50
OSCP (ours)AutoAttack77.63%74.19%ResNet-50
OSCP (ours)random-targeted PGD-4079.81%78.78%ResNet-152
", + "image_path": "596b7fa0d0f097f479b8a16713c1ec36cc424403ed8075f1a2fc229eb47b2300.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 321, + 294, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 321, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 54, + 321, + 294, + 417 + ], + "type": "text", + "content": "rate of 8e-6 and a 500-step warm-up period. For the diffusion process, we employ DDIM-Solver [39] as the PF-ODE solver " + }, + { + "bbox": [ + 54, + 321, + 294, + 417 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 54, + 321, + 294, + 417 + ], + "type": "text", + "content": " with a skipping step " + }, + { + "bbox": [ + 54, + 321, + 294, + 417 + ], + "type": "inline_equation", + "content": "k = 20" + }, + { + "bbox": [ + 54, + 321, + 294, + 417 + ], + "type": "text", + "content": " in Equation 13. Adversarial latents are generated using PGD-10 with " + }, + { + "bbox": [ + 54, + 321, + 294, + 417 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.03" + }, + { + "bbox": [ + 54, + 321, + 294, + 417 + ], + "type": "text", + "content": " against a ResNet50 victim model, and the consistency improvement gradient weight " + }, + { + "bbox": [ + 54, + 321, + 294, + 417 + ], + "type": "inline_equation", + "content": "\\lambda_{CIG}" + }, + { + "bbox": [ + 54, + 321, + 294, + 417 + ], + "type": "text", + "content": " is set to 0.001. Notably, our experiments demonstrate that PEFT via LoRA is sufficient for achieving optimal performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "text", + "content": "Attack Settings We conduct our evaluations on our testing set. Our evaluation encompasses diverse architectures, spanning traditional CNNs (ResNet 50 and 152 [16], WideResNet [50]) and modern Vision Transformers (ViT-B [10], Swin-B [27]). For adversarial attacks, we employ standard " + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "text", + "content": " norm-based methods including PGD [31] and AutoAttack [6]. The PGD attacks are configured with " + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "text", + "content": " bounds " + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\{4/255, 16/255\\}" + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "text", + "content": " and corresponding step sizes " + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\eta \\in \\{1/255, 0.025 \\cdot 16/255\\}" + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "text", + "content": ", while AutoAttack uses " + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "text", + "content": " bound " + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\gamma = 4/255" + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "text", + "content": ". Here, PGD-" + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "text", + "content": " denotes PGD attack with " + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 418, + 295, + 610 + ], + "type": "text", + "content": " iterations. In our evaluation metrics, we define standard accuracy as performance on clean data with our defense framework, robust accuracy as performance on adversarial data with our defense, and clean accuracy as performance on clean data without any defense. We fix the random seed of every experiment to avoid the randomness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 621, + 135, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 135, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 135, + 633 + ], + "type": "text", + "content": "5.1. Main Result" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Our method is regarded as a hybrid method, consisting of adversarial training and purification. Hence we compare our method with both categories of the defenses. Based on the provided attack setting, our method demonstrates significant computational efficiency, completing evaluations in 3 hours on an NVIDIA F40 GPU compared to GDMP's" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 320, + 356, + 545, + 427 + ], + "blocks": [ + { + "bbox": [ + 313, + 321, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 354 + ], + "type": "text", + "content": "Table 2. Accuracy (\\%) and attack success rate (ASR %) results in various architectures, where the attack method is PGD-100, with budget 4/255, attack step size 1/255." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 356, + 545, + 427 + ], + "lines": [ + { + "bbox": [ + 320, + 356, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 320, + 356, + 545, + 427 + ], + "type": "table", + "html": "
ArchitectureCleanASRStandardRobust
WRN-50-282.6%100%77.6%75.2%
Vit-b-1681.6%100%78.2%71.6%
Swin-b83.6%100%79.2%77.8%
", + "image_path": "2d8403619d46e3715b788089c82c8b15c24f82a5b3a5b73f241e07db75a78cf7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "text", + "content": "[44] 48 hours. As shown in Tab. 1, diffusion based purification achieves satisfactory robust accuracy while sacrificing standard accuracy. On the other hand, adversarial training has little standard accuracy loss but the robust accuracy is unsatisfactory. Our method achieves promising results on both standard and robust accuracy with three different attack methods. We achieve robust accuracy improvements of " + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "inline_equation", + "content": "0.92\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "inline_equation", + "content": "1.17\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "text", + "content": " against PGD-100 (" + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\gamma = 4/255" + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "text", + "content": ") and AutoAttack respectively, while improving from " + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "inline_equation", + "content": "77.86\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "inline_equation", + "content": "78.78\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "text", + "content": " against random targeted PGD-40 (" + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\gamma = 16/255" + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "text", + "content": "). We empirically determine " + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "inline_equation", + "content": "t^{*} = 200" + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "text", + "content": " as the optimal purification strength." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 587, + 515, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 515, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 515, + 600 + ], + "type": "text", + "content": "5.2. Analysis With Different Architectures" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "While our GAND-trained LCM is initially optimized using adversarial examples generated for ResNet50, we evaluate its generalization capability across different architectures to address potential overfitting concerns. As shown in Tab. 2, our method maintains strong robust accuracy across architectures, with even the lowest performance of " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "71.6\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{ViT}_{b-16}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " demonstrating effective transfer. These experiments, along with subsequent ImageNet evaluations, are conducted on a representative subset of 500 images from" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24336" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 85, + 526, + 139 + ], + "blocks": [ + { + "bbox": [ + 95, + 71, + 515, + 84 + ], + "lines": [ + { + "bbox": [ + 95, + 71, + 515, + 84 + ], + "spans": [ + { + "bbox": [ + 95, + 71, + 515, + 84 + ], + "type": "text", + "content": "Table 3. Robust Accuracy (%) on our method under Diff-PGD-10 attack " + }, + { + "bbox": [ + 95, + 71, + 515, + 84 + ], + "type": "inline_equation", + "content": "\\gamma = 8/{255}\\left( {\\eta = 2/{255}}\\right)" + }, + { + "bbox": [ + 95, + 71, + 515, + 84 + ], + "type": "text", + "content": " on ImageNet [8]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 85, + 526, + 139 + ], + "lines": [ + { + "bbox": [ + 83, + 85, + 526, + 139 + ], + "spans": [ + { + "bbox": [ + 83, + 85, + 526, + 139 + ], + "type": "table", + "html": "
Defence methodResNet-50ResNet-152WideResNet-50-2Vit-b-16Swin-bConvNeXt-b
DiffPure (t* = 100)53.8%49.4%52.2%16.6%45.1%42.9%
Ours (t* = 250)59.0%56.5%57.9%34.1%53.9%49.1%
", + "image_path": "91fedb18906819faaafa0993a6f639672aac24f0f9d910697536c296b3dd2c42.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 184, + 293, + 258 + ], + "blocks": [ + { + "bbox": [ + 55, + 148, + 295, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 148, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 295, + 182 + ], + "type": "text", + "content": "Table 4. Inference time of purification models to purify one image on an NVIDIA F40 GPU. Notably, GDP refers to Wang et al. [44], DiffPure refers to Nie et al. [32]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 184, + 293, + 258 + ], + "lines": [ + { + "bbox": [ + 56, + 184, + 293, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 293, + 258 + ], + "type": "table", + "html": "
MethodDatasett* ∈ [0,1000]runtime (s)
GDMPImageNet250~ 9
DiffPureImageNet150~ 11
OSCP(Ours)ImageNet20/200/500/1000~ 0.1
OSCP(Ours)CelabA-HQ20/200/500/1000~ 0.5
", + "image_path": "3a6675548748c9e43a88d2b684b7755679f4e642174df5148fda6e5224a64694.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 259, + 295, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 295, + 283 + ], + "type": "text", + "content": "our previously mentioned 10,000 images test set, as we observe consistent performance between these sample sizes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 292, + 192, + 305 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 192, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 192, + 305 + ], + "type": "text", + "content": "5.3. Defend Adaptive Attack" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "text", + "content": "In Tab. 3, we test our method under version 2 of Diff-PGD attack [48], which is a SOTA attack method against diffusion based purification. Our method boost the robust accuracy in Vit-b-16 from " + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "inline_equation", + "content": "16.6\\%" + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "inline_equation", + "content": "34.1\\%" + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "text", + "content": ". In ResNet-50, ResNet-152, WideResNet-50-2, the robust accuracy increase " + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "inline_equation", + "content": "7.1\\%" + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "text", + "content": " respectively. For robust accuracy in Swin-b and ConvNeXt, our method achieve " + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "inline_equation", + "content": "53.9\\%" + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "inline_equation", + "content": "49.1\\%" + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "inline_equation", + "content": "8.8\\%" + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 54, + 311, + 295, + 453 + ], + "type": "text", + "content": " higher than DiffPure. Worth to mention that our method uses LCM model which is deterministic, and hence, Expectation over Transformation, which is usually use in other robust evaluation for DBP, is not needed here." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 463, + 149, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 149, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 149, + 475 + ], + "type": "text", + "content": "5.4. Inference Time" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 482, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 482, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 54, + 482, + 295, + 614 + ], + "type": "text", + "content": "Conventional diffusion-based purification methods are computationally intensive, making them impractical for real-time applications. As shown in Tab. 4, our method achieves purification in just " + }, + { + "bbox": [ + 54, + 482, + 295, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{0.1s}" + }, + { + "bbox": [ + 54, + 482, + 295, + 614 + ], + "type": "text", + "content": " for ImageNet and " + }, + { + "bbox": [ + 54, + 482, + 295, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{0.5s}" + }, + { + "bbox": [ + 54, + 482, + 295, + 614 + ], + "type": "text", + "content": " for CelebA-HQ (1024 × 1024) on a NVIDIA F40 GPU, independent of " + }, + { + "bbox": [ + 54, + 482, + 295, + 614 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 54, + 482, + 295, + 614 + ], + "type": "text", + "content": " where the inference time of other method scales up depending on " + }, + { + "bbox": [ + 54, + 482, + 295, + 614 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 54, + 482, + 295, + 614 + ], + "type": "text", + "content": ". More detailed experiments can be found in the supplementary materials. Our method significantly speeds up the purification process, and this acceleration enables deployment in time-critical scenarios such as autonomous driving and purification in video." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 622, + 159, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 159, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 159, + 635 + ], + "type": "text", + "content": "5.5. Face Recognition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": "We further test our model on a subset of 1000 images from CelebA-HQ [26], choosing purification step " + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": " as 200, control scale 0.8, and utilizing the GAND weights trained on ImageNet. We defend against targeted PGD attacks that attempt to manipulate Arcface [9] (AF), FaceNet [37] (FN), and MobileFaceNet [5] (MFN) to misidentify any input face" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 321, + 184, + 543, + 274 + ], + "blocks": [ + { + "bbox": [ + 313, + 148, + 555, + 180 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 180 + ], + "type": "text", + "content": "Table 5. Robust accuracy (%) for CelebA-HQ under targeted PGD-10 " + }, + { + "bbox": [ + 313, + 148, + 555, + 180 + ], + "type": "inline_equation", + "content": "L_{\\infty} \\gamma (\\gamma = 4 / 255)" + }, + { + "bbox": [ + 313, + 148, + 555, + 180 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 148, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\eta = 0.5 * 5 / 255" + }, + { + "bbox": [ + 313, + 148, + 555, + 180 + ], + "type": "text", + "content": ". The best model is in bold." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 184, + 543, + 274 + ], + "lines": [ + { + "bbox": [ + 321, + 184, + 543, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 184, + 543, + 274 + ], + "type": "table", + "html": "
Defense MethodAFFNMFN
Without defense0%0.3%0%
GaussianBlur (σ = 7.0)2.8%51.4%2.8%
Das et al. [7] (n = 60)17.3%84.1%27.6%
CAP(Ours)83.4%97.8%82.8%
OSCP(Ours)86.8%97.8%84.9%
", + "image_path": "838b58af723cecc3df2f8f4e4c753a39b51db2a0372da66ef626111cc41da27a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 329, + 338, + 539, + 391 + ], + "blocks": [ + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "text", + "content": "Table 6. IQAs for " + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "inline_equation", + "content": "\\uparrow/\\downarrow" + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "text", + "content": " indicate higher/ lower value consists the better image quality. Notably, IQA " + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "text", + "content": " indicates one of IQA we leverage, " + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "text", + "content": " refers to the clean image, " + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{adv}}" + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "text", + "content": " refers to the adversarial sample, " + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{ours}}" + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "text", + "content": " refers to the purified image by our method and " + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{DiffPure}}" + }, + { + "bbox": [ + 313, + 281, + 555, + 337 + ], + "type": "text", + "content": " refers to purified image by DiffPure." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 338, + 539, + 391 + ], + "lines": [ + { + "bbox": [ + 329, + 338, + 539, + 391 + ], + "spans": [ + { + "bbox": [ + 329, + 338, + 539, + 391 + ], + "type": "table", + "html": "
LPIPS↓PSNR↑SSIM↑
IQA(x, xadv)0.154734.210.8906
IQA(x, xours)0.237024.130.7343
IQA(x, xDiffPure)0.261624.110.7155
", + "image_path": "a2160f5dd6caf266a5e3cbd523a79f78258e299459c0d271985e1728f9608213.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 392, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 523 + ], + "type": "text", + "content": "as a specific target person. Our defense goal is to prevent this misidentification by ensuring the cosine similarity between the purified image embedding and the target image remains below the recognition threshold. To validate the effectiveness of ImageNet-trained GAND weights on CelebA-HQ, we compare OSCP with CAP (OSCP without GAND) as shown in Tab. 5. The GAND weights significantly improve robust accuracy from " + }, + { + "bbox": [ + 313, + 392, + 555, + 523 + ], + "type": "inline_equation", + "content": "83.4\\%" + }, + { + "bbox": [ + 313, + 392, + 555, + 523 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 392, + 555, + 523 + ], + "type": "inline_equation", + "content": "86.8\\%" + }, + { + "bbox": [ + 313, + 392, + 555, + 523 + ], + "type": "text", + "content": " on Arcface and from " + }, + { + "bbox": [ + 313, + 392, + 555, + 523 + ], + "type": "inline_equation", + "content": "82.8\\%" + }, + { + "bbox": [ + 313, + 392, + 555, + 523 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 392, + 555, + 523 + ], + "type": "inline_equation", + "content": "84.9\\%" + }, + { + "bbox": [ + 313, + 392, + 555, + 523 + ], + "type": "text", + "content": " on MobileFaceNet. Notably, both our methods (CAP and OSCP) achieve an exceptional robust accuracy of " + }, + { + "bbox": [ + 313, + 392, + 555, + 523 + ], + "type": "inline_equation", + "content": "97.8\\%" + }, + { + "bbox": [ + 313, + 392, + 555, + 523 + ], + "type": "text", + "content": " on FaceNet." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 528, + 462, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 462, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 462, + 540 + ], + "type": "text", + "content": "5.6. Image Quality Assessment" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "Regarding image quality assessments, we compare our method with DiffPure using a subset of 1000 ImageNet images from our testing set. We use PSNR to reflect the overall distortion level of images; SSIM [45] to indicate how human eyes perceive image structures; LPIPS [52] to assesses perceptual similarity. Both methods are configured with " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "t^{*} = 250" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " for fair comparison. As shown in Tab. 6, our method demonstrates superior performance across multiple metrics: LPIPS decreases from 0.2616 to 0.2370, while PSNR and SSIM improve from 24.11 to 24.13 and 0.7155 to 0.7343. These improvements consistently indicate that our purified images maintain greater fidelity to the original inputs, validating the effectiveness of our purification approach. The visualization is shown in Fig. 4." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24337" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 118, + 129 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 118, + 129 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 118, + 129 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 118, + 129 + ], + "type": "image", + "image_path": "0bf874e9c64cd2ac0b736733c729e46dd27809c8f0cadc6e9baf2097b446b88f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 119, + 70, + 176, + 129 + ], + "blocks": [ + { + "bbox": [ + 119, + 70, + 176, + 129 + ], + "lines": [ + { + "bbox": [ + 119, + 70, + 176, + 129 + ], + "spans": [ + { + "bbox": [ + 119, + 70, + 176, + 129 + ], + "type": "image", + "image_path": "9f9335b1abfeaaf6928a89c694c34289a754062db12d25ce0064989252c797b3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 177, + 70, + 234, + 129 + ], + "blocks": [ + { + "bbox": [ + 177, + 70, + 234, + 129 + ], + "lines": [ + { + "bbox": [ + 177, + 70, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 177, + 70, + 234, + 129 + ], + "type": "image", + "image_path": "4cefdc3bdf64550a3a4bdbcad813403cc10290705f15c03107b6ee83e172a2cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 70, + 293, + 129 + ], + "blocks": [ + { + "bbox": [ + 235, + 70, + 293, + 129 + ], + "lines": [ + { + "bbox": [ + 235, + 70, + 293, + 129 + ], + "spans": [ + { + "bbox": [ + 235, + 70, + 293, + 129 + ], + "type": "image", + "image_path": "876422d0e6b470b9a35c3544ebcce78147441d3836ba718f6c325b006be042cb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 130, + 117, + 188 + ], + "blocks": [ + { + "bbox": [ + 60, + 130, + 117, + 188 + ], + "lines": [ + { + "bbox": [ + 60, + 130, + 117, + 188 + ], + "spans": [ + { + "bbox": [ + 60, + 130, + 117, + 188 + ], + "type": "image", + "image_path": "2eb80cc15bc8e46f5898b03f567979ed2929e276dd718cf4ee6c1df364a4e88b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 190, + 95, + 200 + ], + "lines": [ + { + "bbox": [ + 84, + 190, + 95, + 200 + ], + "spans": [ + { + "bbox": [ + 84, + 190, + 95, + 200 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 118, + 130, + 176, + 188 + ], + "blocks": [ + { + "bbox": [ + 118, + 130, + 176, + 188 + ], + "lines": [ + { + "bbox": [ + 118, + 130, + 176, + 188 + ], + "spans": [ + { + "bbox": [ + 118, + 130, + 176, + 188 + ], + "type": "image", + "image_path": "bf0271471374b6d6053683b36950c62a53437ec4f8e62ab167335dd5537dbef9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 190, + 153, + 200 + ], + "lines": [ + { + "bbox": [ + 141, + 190, + 153, + 200 + ], + "spans": [ + { + "bbox": [ + 141, + 190, + 153, + 200 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 177, + 130, + 235, + 188 + ], + "blocks": [ + { + "bbox": [ + 177, + 130, + 235, + 188 + ], + "lines": [ + { + "bbox": [ + 177, + 130, + 235, + 188 + ], + "spans": [ + { + "bbox": [ + 177, + 130, + 235, + 188 + ], + "type": "image", + "image_path": "c49710e3ec5a8efabe4ec99e2ea39baacf6eac4d1484e064395764e2f9dc93d0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 203, + 295, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 203, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 295, + 236 + ], + "type": "text", + "content": "Figure 4. Visualization of the IQA experiment which compares with DiffPure and the proposed method. (a) Input image. (b) Adversarial image. (c) DiffPure. (d) Ours." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 235, + 130, + 293, + 188 + ], + "blocks": [ + { + "bbox": [ + 235, + 130, + 293, + 188 + ], + "lines": [ + { + "bbox": [ + 235, + 130, + 293, + 188 + ], + "spans": [ + { + "bbox": [ + 235, + 130, + 293, + 188 + ], + "type": "image", + "image_path": "2bcf63c853ad961c7988b57cdc2f2b9bd755f9086a941507d011a8401b9090a9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 190, + 270, + 200 + ], + "lines": [ + { + "bbox": [ + 258, + 190, + 270, + 200 + ], + "spans": [ + { + "bbox": [ + 258, + 190, + 270, + 200 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 70, + 251, + 279, + 351 + ], + "blocks": [ + { + "bbox": [ + 70, + 251, + 279, + 351 + ], + "lines": [ + { + "bbox": [ + 70, + 251, + 279, + 351 + ], + "spans": [ + { + "bbox": [ + 70, + 251, + 279, + 351 + ], + "type": "image", + "image_path": "d68a87a8c723768a763eb9d64b79f9b258edc6503a45b8b2a0c0981aec53293e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 357, + 295, + 390 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 390 + ], + "type": "text", + "content": "Figure 5. Performance of our method on different " + }, + { + "bbox": [ + 55, + 357, + 295, + 390 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 55, + 357, + 295, + 390 + ], + "type": "text", + "content": " under PGD-100 " + }, + { + "bbox": [ + 55, + 357, + 295, + 390 + ], + "type": "inline_equation", + "content": "L_{\\infty}\\gamma" + }, + { + "bbox": [ + 55, + 357, + 295, + 390 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 357, + 295, + 390 + ], + "type": "inline_equation", + "content": "\\gamma = 4/255" + }, + { + "bbox": [ + 55, + 357, + 295, + 390 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 55, + 357, + 295, + 390 + ], + "type": "inline_equation", + "content": "\\eta = 0.01^{*}4/255" + }, + { + "bbox": [ + 55, + 357, + 295, + 390 + ], + "type": "text", + "content": ", where we evaluate on ResNet50 on ImageNet." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 99, + 425, + 252, + 509 + ], + "blocks": [ + { + "bbox": [ + 55, + 401, + 295, + 423 + ], + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 423 + ], + "type": "text", + "content": "Table 7. Accuracy (%) under PGD-100 " + }, + { + "bbox": [ + 55, + 401, + 295, + 423 + ], + "type": "inline_equation", + "content": "L_{\\infty}\\gamma" + }, + { + "bbox": [ + 55, + 401, + 295, + 423 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 401, + 295, + 423 + ], + "type": "inline_equation", + "content": "\\gamma = 4/255" + }, + { + "bbox": [ + 55, + 401, + 295, + 423 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 55, + 401, + 295, + 423 + ], + "type": "inline_equation", + "content": "\\eta = 0.01*4/255" + }, + { + "bbox": [ + 55, + 401, + 295, + 423 + ], + "type": "text", + "content": ". The best result is in **bold**, the 2nd best is **underlined**." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 425, + 252, + 509 + ], + "lines": [ + { + "bbox": [ + 99, + 425, + 252, + 509 + ], + "spans": [ + { + "bbox": [ + 99, + 425, + 252, + 509 + ], + "type": "table", + "html": "
CAPGANDStandardRobust
XX75.0%72.4%
X✓74.0%72.6%
✓X77.0%73.2%
✓✓76.8%75.0%
", + "image_path": "059909d9f4621f9b462962c1545362d7423ea740877a894eee5adf4cdefb32bb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 515, + 156, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 156, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 156, + 528 + ], + "type": "text", + "content": "5.7. Ablation Studies" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": "We conduct comprehensive ablation studies on 500 images subset from our testing set. First, as shown in Fig. 5, we optimize CAP's performance across different ControlNet conditioning scales, identifying 0.8 as the optimal value. Tab. 7 demonstrates the individual and combined effects of our components. The baseline LCM-LoRA already shows promising performance, indicating its potential for adversarial purification. GAND-trained weights yield marginal improvements, while CAP integration increases standard and robust accuracy by " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": " respectively. The full OSCP framework (combining both) significantly improves robust accuracy from " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "72.4\\%" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": ", validating our design choices. In Tab. 8, we analyze the impact of the " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{skip}(t)\\pmb{z}_t" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": " term from Eq. (18). While standard accuracy remains stable, removing this term improves robust accuracy," + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 317, + 106, + 555, + 177 + ], + "blocks": [ + { + "bbox": [ + 313, + 72, + 554, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 105 + ], + "type": "text", + "content": "Table 8. Accuracy with and without " + }, + { + "bbox": [ + 313, + 72, + 554, + 105 + ], + "type": "inline_equation", + "content": "c_{skip}(t)z_t" + }, + { + "bbox": [ + 313, + 72, + 554, + 105 + ], + "type": "text", + "content": " term in LCM in our method under PGD-100 (" + }, + { + "bbox": [ + 313, + 72, + 554, + 105 + ], + "type": "inline_equation", + "content": "\\gamma = 8/255" + }, + { + "bbox": [ + 313, + 72, + 554, + 105 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 313, + 72, + 554, + 105 + ], + "type": "inline_equation", + "content": "\\eta = 2/255" + }, + { + "bbox": [ + 313, + 72, + 554, + 105 + ], + "type": "text", + "content": " on ResNet50 on a subset of 500 images from our ImageNet testing set." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 106, + 555, + 177 + ], + "lines": [ + { + "bbox": [ + 317, + 106, + 555, + 177 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 555, + 177 + ], + "type": "table", + "html": "
\\(c_{skip}(t)z_t\\)Acc.\\(t^* = 20\\)\\(t^* = 60\\)\\(t^* = 100\\)\\(t^* = 200\\)\\(t^* = 250\\)
XRobust43.4%56.0%59.4%66.4%68.4%
XStandard81.2%81.4%81.2%79.2%76.6%
✓Robust44.2%54.8%60.4%66.0%67.8%
✓Standard81.2%81.4%81.2%79.2%76.6%
", + "image_path": "52616e538aa15c902b9494bc65b07e6c5c68d0e612c64181ee3b7895f73ba62f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 337, + 223, + 533, + 275 + ], + "blocks": [ + { + "bbox": [ + 313, + 187, + 554, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 187, + 554, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 554, + 220 + ], + "type": "text", + "content": "Table 9. Accuracy of our method under AutoAttack (" + }, + { + "bbox": [ + 313, + 187, + 554, + 220 + ], + "type": "inline_equation", + "content": "\\gamma = 8/255" + }, + { + "bbox": [ + 313, + 187, + 554, + 220 + ], + "type": "text", + "content": ") on ResNet50 on a subset of 1000 images from ImageNet on different sets of training time steps. (" + }, + { + "bbox": [ + 313, + 187, + 554, + 220 + ], + "type": "inline_equation", + "content": "t^{*} = 200" + }, + { + "bbox": [ + 313, + 187, + 554, + 220 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 223, + 533, + 275 + ], + "lines": [ + { + "bbox": [ + 337, + 223, + 533, + 275 + ], + "spans": [ + { + "bbox": [ + 337, + 223, + 533, + 275 + ], + "type": "table", + "html": "
Acc.N-k/4N-k/23(N-k)/4N-k
Robust74.0%74.3%73.7%73.6%
Standard78.8%79.4%79.1%79.1%
", + "image_path": "b9536e1c72c323099b2276038273e71c5c5ea72ba06d8f5938759562abaaef3a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 277, + 554, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 554, + 312 + ], + "type": "text", + "content": "likely because " + }, + { + "bbox": [ + 313, + 277, + 554, + 312 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 277, + 554, + 312 + ], + "type": "text", + "content": " retains adversarial noise from the input image. This observation leads us to exclude this term from our LCM implementation." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 312, + 554, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 384 + ], + "type": "text", + "content": "Finally, Tab. 9 explores different training time step ranges: " + }, + { + "bbox": [ + 313, + 312, + 554, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{U}[1,(N - k) / 4],\\mathcal{U}[1,(N - k) / 2],\\mathcal{U}[1,3(N - k) / 4]" + }, + { + "bbox": [ + 313, + 312, + 554, + 384 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 312, + 554, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{U}[1,N - k]" + }, + { + "bbox": [ + 313, + 312, + 554, + 384 + ], + "type": "text", + "content": " (original). Results show that smaller time step ranges maintain better performance, suggesting that fine-tuning LCM with smaller time steps is sufficient for purification tasks." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 392, + 466, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 466, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 466, + 403 + ], + "type": "text", + "content": "6. Conclusion and Discussion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 411, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 638 + ], + "type": "text", + "content": "In this paper, we present One Step Control Purification (OSCP), a novel framework that addresses two critical challenges in diffusion-based adversarial purification: computational efficiency and semantic preservation. Our approach integrates the knowledge of pre-trained Diffusion Models with the proposed Gaussian Adversarial Noise Distillation (GAND) to achieve robust defense in a single inference step. Then as the inference framework, the Controlled Adversarial Purification (CAP), utilizing the unlearnable edge detection operator, effectively preserves semantic information of the input during purification. Extensive experiments validate OSPC's effectiveness across diverse architectures and datasets. On ImageNet, we achieve " + }, + { + "bbox": [ + 313, + 411, + 555, + 638 + ], + "type": "inline_equation", + "content": "74.19\\%" + }, + { + "bbox": [ + 313, + 411, + 555, + 638 + ], + "type": "text", + "content": " robust accuracy against strong attacks while requiring only 0.1s per purification—a 100-fold speedup compared to existing methods. The framework demonstrates strong cross-domain generalization, from natural images to face recognition tasks, while our ablation studies confirm the synergistic benefits of combining GAND and CAP." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 639, + 556, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 556, + 710 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 556, + 710 + ], + "type": "text", + "content": "While our work represents significant progress in real-time adversarial defense, several promising directions remain, including the adaptive purification strength mechanisms and alternative control conditions. These advances could further enhance the applicability of diffusion-based defenses in time-critical scenarios." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24338" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Sajjad Amini, Mohammadreza Teymoorianfard, Shiqing Ma, and Amir Houmansadr. Meansparse: Post-training robustness enhancement through mean-centered feature sparsification. CoRR, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "text", + "content": "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, pages 274-283, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 296, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 296, + 247 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22563-22575, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 282 + ], + "type": "text", + "content": "[4] John Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-8(6):679-698, 1986. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "type": "text", + "content": "[5] Sheng Chen, Yang Liu, Xiang Gao, and Zhen Han. Mobilefacenet: Efficient cnns for accurate real-time face verification on mobile devices. In Biometric Recognition, pages 428-438. Springer International Publishing, 2018. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 331, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 295, + 373 + ], + "type": "text", + "content": "[6] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International Conference on Machine Learning, pages 2206-2216, 2020. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 376, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 296, + 441 + ], + "type": "text", + "content": "[7] Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen, Fred Hohman, Siwei Li, Li Chen, Michael E Kounavis, and Duen Horng Chau. Shield: Fast, practical defense and vaccination for deep learning usingJPEG compression. In ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 196-204, 2018. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 444, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 295, + 487 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 489, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 295, + 542 + ], + "type": "text", + "content": "[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 545, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 295, + 621 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "text", + "content": "[11] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv e-prints, pages arXiv-2010, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[13] Zhongliang Guo, Junhao Dong, Yifei Qian, Kaixuan Wang, Weiye Li, Ziheng Guo, Yuheng Wang, Yanli Li, Ognjen Arandjelovic, and Lei Fang. Artwork protection against neural style transfer using locally adaptive adversarial color attack. In ECAI 2024, pages 1414-1421. IOS Press, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "text", + "content": "[14] Zhongliang Guo, Chun Tong Lei, Lei Fang, Shuai Zhao, Yifei Qian, Jingyu Lin, Zeyu Wang, Cunjian Chen, Ognjen Arandjelovic, and Chun Pong Lau. A grey-box attack against latent diffusion model-based image editing by posterior collapse. arXiv preprint arXiv:2408.10901, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 185, + 553, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 247 + ], + "type": "text", + "content": "[15] Zhongliang Guo, Weiye Li, Yifei Qian, Ognjen Arandjelovic, and Lei Fang. A white-box false positive adversarial attack method on contrastive loss based offline handwritten signature verification models. In International Conference on Artificial Intelligence and Statistics, pages 901-909, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 250, + 553, + 293 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 295, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 326 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 327, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 553, + 370 + ], + "type": "text", + "content": "[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 372, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 553, + 426 + ], + "type": "text", + "content": "[19] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 427, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 553, + 491 + ], + "type": "text", + "content": "[20] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ODE trajectory of diffusion. In International Conference on Learning Representations, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "type": "text", + "content": "[21] Chun Pong Lau, Jiang Liu, Hossein Souri, Wei-An Lin, Soheil Feizi, and Rama Chellappa. Interpolated joint space adversarial training for robust and generalizable defenses. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):13054-13067, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 548, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 601 + ], + "type": "text", + "content": "[22] Wei-An Lin, Chun Pong Lau, Alexander Levine, Rama Chellappa, and Soheil Feizi. Dual manifold adversarial robustness: Defense against lp and non-lp adversarial attacks. In Advances in Neural Information Processing Systems, pages 3487-3498, 2020. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 604, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 646 + ], + "type": "text", + "content": "[23] Jiang Liu, Chun Pong Lau, Hossein Souri, Soheil Feizi, and Rama Chellappa. Mutual adversarial training: Learning together is better than going alone. IEEE Transactions on Information Forensics and Security, 17:2364-2377, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "type": "text", + "content": "[24] Jiang Liu, Chen Wei, Yuxiang Guo, Heng Yu, Alan Yuille, Soheil Feizi, Chun Pong Lau, and Rama Chellappa. Instruct2attack: Language-guided semantic adversarial attacks. arXiv preprint arXiv:2311.15551, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Yiming Liu, Kezhao Liu, Yao Xiao, Ziyi Dong, Xiaogang Xu, Pengxu Wei, and Liang Lin. Towards better adversaries" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24339" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "ial purification via adversarial denoising diffusion training. arXiv preprint arXiv:2404.14309, 2024. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "text", + "content": "[26] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 198, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 198, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 57, + 198, + 295, + 241 + ], + "type": "text", + "content": "[28] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2837-2845, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 243, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 243, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 57, + 243, + 294, + 285 + ], + "type": "text", + "content": "[29] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv e-prints, pages arXiv-2310, 2023. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 287, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 287, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 57, + 287, + 294, + 331 + ], + "type": "text", + "content": "[30] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv e-prints, pages arXiv-2311, 2023. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 295, + 376 + ], + "type": "text", + "content": "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 378, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 378, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 57, + 378, + 294, + 422 + ], + "type": "text", + "content": "[32] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Animashree Anandkumar. Diffusion models for adversarial purification. In International Conference on Machine Learning, pages 16805-16827, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 423, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 423, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 57, + 423, + 294, + 456 + ], + "type": "text", + "content": "[33] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 457, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 457, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 57, + 457, + 294, + 511 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685. IEEE, 2022. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 513, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 295, + 578 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 624 + ], + "type": "text", + "content": "[36] Pouya Samangouei, Maya Kabbab, and Rama Chellappa. Defense-gan: Protecting classifiers against adversarial attacks using generative models. In International Conference on Learning Representations, 2018. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "text", + "content": "[37] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[38] Naman Deep Singh, Francesco Croce, and Matthias Hein. Revisiting adversarial training for imagenet: Architectures, training and generalization across threat models. Advances in Neural Information Processing Systems, 36, 2024. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "text", + "content": "[40] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "text", + "content": "[41] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In International Conference on Machine Learning, pages 32211-32252, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 174, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 553, + 216 + ], + "type": "text", + "content": "[42] Hossein Souri, Pirazh Khorramshahi, Chun Pong Lau, Micah Goldblum, and Rama Chellappa. Identification of attack-specific signatures in adversarial examples. arXiv preprint arXiv:2110.06802, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "text", + "content": "[43] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 262, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 293 + ], + "type": "text", + "content": "[44] Jinyi Wang, Zhaoyang Lyu, Dahua Lin, Bo Dai, and Hongfei Fu. Guided diffusion model for adversarial purification. arXiv preprint arXiv:2205.14969, 2022. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 295, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 553, + 338 + ], + "type": "text", + "content": "[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 339, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 382 + ], + "type": "text", + "content": "[46] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve adversarial training. In International Conference on Machine Learning, pages 36246-36263, 2023. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 384, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 384, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 384, + 553, + 437 + ], + "type": "text", + "content": "[47] Yan Xu, Baoyuan Wu, Fumin Shen, Yanbo Fan, Yong Zhang, Heng Tao Shen, and Wei Liu. Exact adversarial attack to image captioning via structured output learning with latent variables. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 439, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 439, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 439, + 553, + 481 + ], + "type": "text", + "content": "[48] Haotian Xue, Alexandre Araujo, Bin Hu, and Yongxin Chen. Diffusion-based adversarial sample generation for improved stealthiness and controllability. Advances in Neural Information Processing Systems, 36, 2024. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 483, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 483, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 483, + 553, + 525 + ], + "type": "text", + "content": "[49] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, pages 12062-12072, 2021. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "text", + "content": "[50] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "type": "text", + "content": "[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 604, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 658 + ], + "type": "text", + "content": "[52] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 659, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 659, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 659, + 553, + 713 + ], + "type": "text", + "content": "[53] Shuai Zhao, Meihuizi Jia, Zhongliang Guo, Leilei Gan, Xiaoyu Xu, Xiaobao Wu, Jie Fu, Feng Yichao, Fengjun Pan, and Anh Tuan Luu. A survey of recent backdoor attacks and defenses in large language models. Transactions on Machine Learning Research, 2025. Survey Certification. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24340" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/c9f2cc0e-b9b5-4129-a1a3-c1fc202a4666_content_list.json b/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/c9f2cc0e-b9b5-4129-a1a3-c1fc202a4666_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c30329dfd8a2f9c8b6c9d7a9e213d20ac04db94c --- /dev/null +++ b/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/c9f2cc0e-b9b5-4129-a1a3-c1fc202a4666_content_list.json @@ -0,0 +1,1804 @@ +[ + { + "type": "text", + "text": "Instant Gaussian Stream: Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting", + "text_level": 1, + "bbox": [ + 101, + 128, + 897, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinbo Yan $^{1}$ , Rui Peng $^{1,2}$ , Zhiyan Wang $^{1}$ , Luyang Tang $^{1,2}$ , Jiayu Yang $^{1,2}$ , Jie Liang $^{1}$ , Jiahao Wu $^{1}$ , Ronggang Wang $^{1,2*}$", + "bbox": [ + 218, + 200, + 776, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University 2Pengcheng Laboratory", + "bbox": [ + 122, + 239, + 874, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yjb, ruipeng, zywang23, tly926, liangjie, wjh0616}@stu.pku.edu.cn jiayuyang@pku.edu.cn rgwang@pkusz.edu.cn", + "bbox": [ + 207, + 294, + 789, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/299dc7953d51ad4b4f952107a0ca8d82f08eae69a4bd1323b6241a4f63801438.jpg", + "image_caption": [ + "(a)High Quality Result" + ], + "image_footnote": [], + "bbox": [ + 94, + 363, + 205, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c7b9e0e5d6ffd37a29677ac0bb5b95520aa6f1b8ad81ff68a0c1e9321f880d3.jpg", + "image_caption": [ + "Figure 1. Performance comparison with pervious SOTA[17, 29, 31, 51, 62]. Our method achieves a per-frame reconstruction time of 2.67s, delivering high-quality rendering results in a streaming fashion (a)(b), with a noticeable improvement in performance (c). * denotes a streamable method." + ], + "image_footnote": [], + "bbox": [ + 210, + 363, + 320, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a4ad20a46aa7c0e85e0d8b10bad9a058b73eb4360228a253d01580d0d75be50.jpg", + "image_caption": [ + "(b)Realtime Rendering" + ], + "image_footnote": [], + "bbox": [ + 338, + 364, + 607, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6972288d4c0311de77821dbf47436ee544ccae8d0189aa5bfb2b5a6799786b48.jpg", + "image_caption": [ + "(c)Comparison on Quality and Train Speed" + ], + "image_footnote": [], + "bbox": [ + 620, + 361, + 906, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 595, + 326, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building Free-Viewpoint Videos in a streaming manner offers the advantage of rapid responsiveness compared to offline training methods, greatly enhancing user experience. However, current streaming approaches face challenges of high per-frame reconstruction time $(10s+)$ and error accumulation, limiting their broader application. In this paper, we propose Instant Gaussian Stream (IGS), a fast and generalizable streaming framework, to address these issues. First, we introduce a generalized Anchor-driven Gaussian Motion Network, which projects multi-view 2D motion features into 3D space, using anchor points to drive the motion of all Gaussians. This generalized Network generates the motion of Gaussians for each target frame in the time required for a single inference. Second, we propose a Key-frame-guided Streaming Strategy that refines each key frame, enabling accurate reconstruction of temporally complex scenes while mitigating error accumulation. We conducted extensive in-domain and cross-domain evalua", + "bbox": [ + 88, + 628, + 485, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions, demonstrating that our approach can achieve streaming with a average per-frame reconstruction time of $2s+$ , alongside a enhancement in view synthesis quality.", + "bbox": [ + 511, + 598, + 906, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 676, + 645, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing Free-Viewpoint Videos (FVV) from multiview images is a valuable area of research, with applications spanning immersive media such as VR, AR, and sports broadcasting. By enabling interactive, photorealistic visuals, FVVs from dynamic scenes hold the potential to become a next-generation visual medium, offering experiences that go beyond traditional video formats. To enhance user experience, streaming-based FVV construction—where dynamic scenes are reconstructed frame by frame—offers a low-delay response compared to traditional offline training approaches, making it better suited for real-time, interactive applications.", + "bbox": [ + 509, + 702, + 906, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With advancements in real-time rendering and high-", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16520", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quality view synthesis powered by 3D Gaussian Splatting (3DGS)[26], dynamic scene reconstruction has seen rapid progress. Some offline training methods[23, 31, 62, 66, 69, 71] achieve high-quality view synthesis but require collecting all frames before training can begin. This limitation makes them less suitable for scenarios that demand fast response times, such as live streaming and virtual meetings. To address these challenges, some methods[29, 51] adopt a streaming framework that reconstructs dynamic scenes frame by frame by modeling inter-frame differences. However, streaming-based dynamic scene reconstruction still faces significant challenges. First, current methods typically require per-frame optimization, resulting in high per-frame latencies $(10\\mathrm{s}+)$ , which severely impact the real-time usability of these systems. Additionally, error accumulation across frames degrades the reconstruction quality of later frames, making it difficult for streaming methods to scale effectively to longer video sequences.", + "bbox": [ + 89, + 90, + 480, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To promote the streaming framework to be more practical, we introduce Instant Gaussian Stream (IGS), a streaming approach for dynamic scene reconstruction that achieves a per-frame reconstruction time of $2s+$ , mitigates error accumulation, and enhances view synthesis quality. First, to tackle the issue of high per-frame reconstruction time, we developed a generalized Anchor-driven Gaussian Motion Network (AGM-Net). This network utilizes a set of key points, called anchor points, to carry motion features that guide Gaussian transformations. This design allows the inference process to compute the motion of Gaussian primitives between frames in a single feedforward pass, eliminating the need for per-frame optimization. Second, to further improve view synthesis quality and minimize error accumulation, we propose a Key-frame-guided Streaming strategy. By establishing key-frame sequences and performing maxpoint-bounded refinement on key frames, our method mitigates the impact of error accumulation and enhances rendering quality in temporally complex scenes.", + "bbox": [ + 91, + 364, + 482, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conducted extensive validation in both in-domain and cross-domain scenarios, and the experimental results demonstrate the strong generalization capability of our model, with significant improvements over current state-of-the-art methods in terms of per-frame reconstruction time and rendering quality. To the best of our knowledge, this is the first approach to use a generalized method for streaming reconstruction of dynamic scenes. Our contributions are summarized below.", + "bbox": [ + 89, + 656, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a generalized Anchor-driven Gaussian Motion Network that captures Gaussian motion between adjacent frames with a single inference, eliminating the need for frame-by-frame optimization.", + "- We designed a Key-frame-guided Streaming strategy to enhance our method's capability in handling temporally complex scenes, improving overall view synthesis qual" + ], + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ity within the streaming framework and mitigating error accumulation.", + "bbox": [ + 526, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- The evaluation results in both in-domain and cross-domain scenarios demonstrate the generalization capability of our method and its state-of-the-art performance. We achieve a 2.7 s per-frame reconstruction time for streaming, representing a significant improvement over previous methods. Additionally, we improve view synthesis quality, enabling real-time rendering at 204 FPS while maintaining comparable storage overhead.", + "bbox": [ + 513, + 121, + 906, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 273, + 651, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1.3D Reconstruction and View Synthesis", + "text_level": 1, + "bbox": [ + 511, + 304, + 844, + 320 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel view synthesis (NVS) has always been a hot topic in the field of computer vision. By using MLP to implicitly represent the scene, Neural Radiance Fields (NeRF) [38] achieves realistic rendering. Subsequent works have imporced NeRF to enhance rendering quality [1, 2, 60], reduce the number of training views [40, 58, 63, 67], lessen dependence on camera poses [4, 10, 32, 56], and improve both training and inference speeds [3, 8, 16, 20, 21, 39, 44, 45, 50]. 3D Gaussian Splitting (3DGS) [26] employs anisotropic Gaussian primitives to represent scenes and introduces rasterization-based splatting rendering algorithm, enhancing both speed and rendering quality. Some methods focus on various aspects of improving Gaussian field representations, including rendering quality[28, 37, 46, 70, 74, 79], enhancing geometric accuracy[22, 75, 76], and increasing compression efficiency, [11, 13, 37, 68], joint optimization of camera pose and gaussian fields [14, 18, 47], as well 3D generation [9, 54, 55, 81].", + "bbox": [ + 511, + 330, + 906, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Generalizable 3D Reconstruction for Acceleration", + "text_level": 1, + "bbox": [ + 511, + 631, + 903, + 660 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3DGS requires per-scene optimization to achieve realistic rendering results. To accelerate this time-consuming process, some works [24, 27, 52, 53, 77, 80], inspired by generalizable NeRF [7, 25, 61, 65, 73], have proposed to train generalizable Gaussian models on large-scale datasets to enable fast reconstruction. PixelSplat [6] utilizes an Transformer to encode features and decode them into Gaussian attributes. Other generalizable models [12, 15, 35, 78] utilize Transformers or Multi-View Stereo (MVS) [72] techniques to construct cost volumes followed by a decoder, achieving real-time rendering speeds and excellent generalizability. To the best of our knowledge, our work is the first to apply generalizable models to dynamic streaming scenes, utilizing their rapid inference capabilities to accelerate the processing of dynamic scenes reconstruction.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16521", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Dynamic Scene Reconstruction and View Synthesis", + "text_level": 1, + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There have been numerous efforts to extend static scene reconstruction to dynamic scenes based on NeRF[5, 17, 19, 33, 34, 41, 43, 48, 57]. Since the advent of 3D Gaussian Splatting (3DGS)[26], researchers have explored incorporating its real-time rendering capabilities into dynamic scene reconstruction[23, 31, 62, 66, 69, 71].", + "bbox": [ + 89, + 128, + 483, + 218 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, these approaches rely on offline training with full video sequences, making them unsuitable for applications requiring real-time interaction and fast response. To address this issue, existing methods such as StreamRF[29], NeRFPlayer[49], ReRF[59], and 3DGStream[51] reformulate the dynamic modeling problem using a Streaming method. Notably, 3DGStream[51], based on Gaussian Splatting, optimizes a Neural Transformation Cache to model Gaussian movements between frames, further improving the performance. Although these methods achieve promising results, they still rely on per-frame optimization, resulting in significant delays (with current SOTA methods requiring over 10 seconds per frame[51]). Our approach offers a new perspective for streaming dynamic scene modeling: by training a generalized network, we eliminate the need for per-frame optimization, achieving low per-frame reconstruction time alongside high rendering quality.", + "bbox": [ + 89, + 219, + 483, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 489, + 181, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we begin with an overview of the pipeline in Sec.3.1. Then, in Sec. 3.3, we introduce the Anchordriven Gaussian Motion Network (AGM-Net), a generalized model that drives Gaussian motion from the previous frame using anchor points, which serve as key points in the 3D scene. Following this, we present our Key-frame-guided Streaming strategy in Sec.3.4. Finally, in Sec.3.5, we outline the loss function used in our training.", + "bbox": [ + 89, + 516, + 483, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 648, + 200, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to model dynamic scenes in a streaming manner with minimal per-frame reconstruction time. To achieve this, we adopt a generalized AGM-Net that extracts 3D motion features from the scene using anchor points and drives the motion of Gaussian primitives between frames in a single inference step. And we propose a key-frame-guided Streaming strategy to further improve view synthesis quality and handle temporally complex scenes while addressing error accumulation in streaming reconstruction. The overall pipeline is illustrated in Fig. 2.", + "bbox": [ + 89, + 671, + 482, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 832, + 218, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gaussian splatting[26] represents static scenes as a collection of anisotropic 3D Gaussians. The color of each pixel is obtained through point-based alpha blending rendering,", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "enabling high-fidelity real-time novel view synthesis.", + "bbox": [ + 511, + 90, + 864, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, each Gaussian primitive $\\mathcal{G}_i$ is parameterized by a center $\\mu \\in \\mathbb{R}^3$ , 3D covariance matrix $\\Sigma \\in \\mathbb{R}^{3\\times 3}$ , opacity $\\alpha \\in \\mathbb{R}$ , and color $c\\in \\mathbb{R}^{3(n + 1)^2}$ :", + "bbox": [ + 511, + 106, + 903, + 152 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} (x) = e ^ {- \\frac {1}{2} (x - \\mu) ^ {T} \\Sigma^ {- 1} (x - \\mu)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 162, + 903, + 184 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During rendering, the 3D Gaussian is first projected onto 2D space. Subsequently, the Gaussians covering a pixel are sorted based on depth. The color of the pixel $\\mathbf{c}$ is obtained using point-based alpha blending rendering:", + "bbox": [ + 511, + 195, + 906, + 257 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} = \\sum_ {i = 1} ^ {n} c _ {i} \\alpha_ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {i} ^ {\\prime}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 267, + 903, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\alpha^{\\prime}$ represents the opacity after projection onto the 2D space.", + "bbox": [ + 511, + 321, + 903, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Anchor-driven Gaussian Motion Network", + "text_level": 1, + "bbox": [ + 511, + 361, + 867, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion Feature Maps: Given multi-view images of current frames $\\mathbf{I}' = (I_1', \\dots, I_V')$ with camera parameters, We can first construct a multi-view image pair, which contains the current frame and the previous frame $\\mathbf{I}$ from corresponding viewpoints. Then, we use a optical flow model to obtain the intermediate flow embeddings. Next, a modulation layer[9, 42] is applied to inject the viewpoint and depth information into the embeddings, ultimately resulting in 2D motion feature maps $\\mathbf{F} \\in \\mathbb{R}^{V \\times C \\times H \\times W}$ .", + "bbox": [ + 511, + 383, + 905, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Anchor Sampling: To deform the Gaussian primitives $\\mathcal{G}$ from the previous frame, we need to compute the motion of each Gaussian. However, directly computing the motion for each Gaussian is computationally expensive and memory-intensive due to the large number of Gaussian points. To address this, we employ an anchor-point-based approach to represent the motion features of the entire scene in 3D space. The anchor-driven approach supports batch processing during training, reducing computational overhead while preserving the geometric information of the Gaussian primitives. Specifically, we use Farthest Point Sampling (FPS) to sample $M$ anchor points from the $N$ Gaussian primitives", + "bbox": [ + 511, + 518, + 906, + 700 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} = \\mathbf {F P S} \\left(\\left\\{\\mu_ {i} \\right\\} _ {i \\in N}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 714, + 903, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{C} \\in \\mathbb{R}^{M \\times 3}$ represents the sampled anchor points with $M$ set to 8192 in our experiments, and $\\mu_{i}$ denotes the position of $\\mathcal{G}_{i}$", + "bbox": [ + 511, + 734, + 903, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Projection-aware 3D Motion Feature Lift: We adopt a projection-aware approach to lift multiview 2D motion features into 3D space. Specifically, we project sampled anchor points onto each motion feature map based on the camera poses, obtaining high-resolution motion features:", + "bbox": [ + 511, + 780, + 905, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} = \\frac {1}{V} \\sum_ {j \\in V} \\Psi \\left(\\Pi_ {j} \\left(\\mathcal {C} _ {i}\\right), F _ {j}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 866, + 903, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16522", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4d097330a1a915bfa0a1a13b2ecb645a1908cd8f782eb7e1fb2d81b94390a9a0.jpg", + "image_caption": [ + "Figure 2. The overall pipeline of IGS. (a) Starting from the key frame and moving towards the target frame, we extract the 2D Motion Feature Map. (b) Then we sample M anchor points from the Gaussian primitives of the key frame, (c) and the anchor points are projected onto these feature maps to obtain 3D motion features through Projection-aware Motion Feature Lift. (d) Each Gaussian point interpolates its own motion feature from neighboring anchors and applies a weighted aggregation of features, which is then decoded into the motion of the Gaussian between the key frame and the target frame. (e) The entire streaming reconstruction process is guided by the Key-frame-guided Streaming strategy, where the key frame directly infers subsequent candidate frames until the next key-frame is reached, at which point max-point bounded refinement is applied to the key-frame." + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 908, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Pi_j(\\mathcal{C}_i)$ represents the projection of $\\mathcal{C}_i$ onto the image plane of $F_{j}$ using the camera parameters of $F_{j}$ , and $\\Psi$ denotes bilinear interpolation. By projection, each anchor point can accurately obtain its feature $f_{i}\\in \\mathbb{R}^{C}$ from the multi-view feature map, effectively lifting the 2D motion map into 3D space.", + "bbox": [ + 88, + 507, + 480, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then use these features $\\{f_i\\}_{i\\in M}$ , stored at each anchor point, as input to a Transformer block using self-attention to further capture motion information within the 3D scene.", + "bbox": [ + 89, + 598, + 482, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{z _ {i}: z _ {i} \\in \\mathbb {R} ^ {C} \\right\\} _ {i \\in M} = \\operatorname {T r a n s f o r m e r} \\left(\\left\\{f _ {i} \\right\\} _ {i \\in M}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 670, + 482, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The output of the Transformer block $\\{z_i\\}_{i\\in M}$ represents the final 3D motion features we obtain. Now, we can use these 3D motion features to represent the motion information of an anchor and its neighborhood, and drive the motion of the neighboring Gaussian points based on these motion features.", + "bbox": [ + 89, + 700, + 482, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interpolate and Motion Decode: Using the 3D motion features stored at anchor points, we can assign each Gaussian point a motion feature by interpolating from its K nearest anchors in the neighborhood:", + "bbox": [ + 89, + 792, + 482, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i} = \\frac {\\sum_ {k \\in \\mathcal {N} (i)} e ^ {- d _ {k}} z _ {k}}{\\sum_ {k \\in \\mathcal {N} (i)} e ^ {- d _ {k}}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 864, + 482, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(i)$ represents the set of neighboring anchor points of Gaussian point $\\mathcal{G}_i$ , and $d_k$ represents the Euclidean distance from Gaussian point $G_i$ to anchor $\\mathcal{C}_k$ . Then we can use a Linear head to decode the Motion feature to the movement of a Gaussian primitive:", + "bbox": [ + 511, + 507, + 906, + 583 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd \\mu_ {i}, d r o t _ {i} = \\mathbf {L i n e a r} \\left(z _ {i}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 595, + 906, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "here, we use the deformation of the Gaussian's position $d\\mu_{i}$ and the deformation of the rotation $drot_{i}$ , to represent the movement of a Gaussian primitive. The new position and rotation of the Gaussian are as follows:", + "bbox": [ + 511, + 623, + 906, + 683 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {i} ^ {\\prime} = \\mu_ {i} + d \\mu_ {i}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 656, + 694, + 906, + 714 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {r o t} _ {i} ^ {\\prime} = \\operatorname {n o r m} \\left(\\operatorname {r o t} _ {i}\\right) \\times \\operatorname {n o r m} \\left(\\operatorname {d r o t} _ {i}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 738, + 903, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "here ' refers to the new attributes. norm denotes to quaternion normalization and $\\times$ represents quaternion multiplication, as used in previous work[51].", + "bbox": [ + 511, + 763, + 906, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Key-frame-guided Streaming", + "text_level": 1, + "bbox": [ + 511, + 816, + 774, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using AGM-Net, we can transition Gaussian primitives from the previous frame to the current frame within a single forward inference pass. However, this process only adjusts the position and rotation of Gaussian primitives, making it", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16523", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "effective for capturing rigid motion but inadequate for accurately representing non-rigid motion. Furthermore, the number of Gaussian points remains constant, limiting its capacity to model temporally dynamic scenes where objects may appear or disappear. These limitations result in challenges in capturing scene dynamics and can lead to error accumulation across frames.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To better model object changes and reduce error accumulation, we propose a Key-frame-guided Streaming strategy that uses key frames as the initial state for deforming Gaussians in subsequent frames. We also introduce a Max points bounded refinement strategy, enabling efficient key frame reconstruction without redundant points and preventing point count growth across frames. This approach helps avoid overfitting in sparse-viewpoint scenes by effectively managing point density.", + "bbox": [ + 89, + 198, + 480, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Key-frame-guided strategy: Starting from frame 0, we designate a key frame every $w$ frames, forming a key-frame sequence $\\{K_0, K_w, \\dots, K_{nw}\\}$ . The remaining frames serve as candidate frames. During streaming reconstruction, for example, beginning with a key frame $K_{iw}$ , we deform the Gaussians forward across successive candidate frames using AGM-Net until reaching the next key frame $K_{(i+1)w}$ . At this point, we refine the deformed Gaussians of key frame $K_{(i+1)w}$ . Then, we continue deforming from key frame $K_{(i+1)w}$ to process subsequent frames.", + "bbox": [ + 89, + 335, + 482, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This key-frame-guided strategy offers several advantages. First, when AGM-Net is applied to candidate frames, it is always start from the most recent key frame, preventing error propagation across candidate frames between key frames and eliminating cumulative error. Second, candidate frames do not require optimization-based refinement, as their Gaussians are generated through a single model inference with AGM-Net, ensuring low per-frame reconstruction time. Additionally, we can batch process up to $w$ frames following each key frame, which further accelerates our pipeline.", + "bbox": [ + 89, + 489, + 482, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Max points bounded Key-frame Refinement: During the refinement of each key frame, we optimize all parameters of the Gaussians and support cloning, splitting, and filtering, which is same to 3DGS[26]. This approach allows us to handle object deformations as well as the appearance and disappearance of objects in temporally complex scenes, effectively preventing error accumulation from key frame to subsequent frames. However, this optimization strategy can lead to a gradual increase in Gaussian primitives at each key frame, which not only raises computational complexity and storage requirements but also risks overfitting in sparse viewpoints, particularly in dynamic scenes where viewpoints are generally limited.", + "bbox": [ + 89, + 657, + 482, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this, we adopt a Max Points Bounded Refine method. When densifying Gaussian points, we control the number of Gaussians allowed to densify by adjusting each", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "point's gradient, ensuring that the total number of points does not exceed a predefined maximum.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss Function", + "text_level": 1, + "bbox": [ + 513, + 131, + 660, + 145 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our training process consists of two parts: offline training the generalized AGM-Net and performing online training for the key frames. The generalized AGM-Net only needs to be trained once, and it can generalize to multiple scenes. We train the AGM-Net across scenes using gradient descent, relying solely on a view synthesis loss between our predicted views and the ground truth views, which includes an $\\mathcal{L}_1$ term and an $\\mathcal{L}_{D - SSIM}$ term and can be formulated as:", + "bbox": [ + 511, + 152, + 903, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {D - S S I M} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 286, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When performing online training on the Gaussians in key frames, we use the same loss function as in Eq.10. However, this time, we optimize the attributes of the Gaussian primitives rather than the parameters of the neural network.", + "bbox": [ + 511, + 314, + 903, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Implementation details", + "text_level": 1, + "bbox": [ + 511, + 388, + 732, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this Section, we first introduce the datasets we used, along with the partitioning and preprocessing of training data, in Sec. 4.1. Next, we provide a detailed explanation of the configuration of the AGM network and the training hyperparameters in Sec. 4.2. Finally, we describe the detailed setup for streaming in Sec. 4.3.", + "bbox": [ + 511, + 414, + 903, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 513, + 614, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Neural 3D Video Datasets (N3DV)[30] includes 6 dynamic scenes recorded using a multi-view setup featuring 21 cameras, with a resolution of $2704 \\times 2028$ . Each multiview video comprises 300 frames.", + "bbox": [ + 511, + 536, + 903, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meeting Room Datasets[29] includes 3 dynamic scenes recorded with 13 cameras at a resolution of $1280 \\times 720$ . Each multi-view video also contains 300 frames.", + "bbox": [ + 511, + 597, + 903, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Preparation: We split four sequences from the N3DV dataset into the training set, with the remaining two sequences, {cut roasted beef, sear steak}, used as the test set. For the training set, we constructed 3D Gaussians for all frames in the four training sequences, totaling 1200 frames, which required 192 GPU hours. For each frame's 3D Gaussian, we performed motions forward and backward for five frames, creating 12,000 pairs for training. For testing, we selected one viewpoint for evaluation for both datasets, consistent with previous methods.", + "bbox": [ + 511, + 642, + 905, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.AGM Network", + "text_level": 1, + "bbox": [ + 511, + 803, + 663, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use GM-Flow[64] to extract optical flow embeddings and add a Swin-Transformer[36] block for fine-tuning while keeping the other parameters of GM-Flow fixed. Our AGM model accepts an arbitrary number of input views. To balance computational complexity and performance, we use", + "bbox": [ + 511, + 824, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16524", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$V = 4$ views, each producing a motion map with $C = 128$ channels and a resolution of $128 \\times 128$ . We sample $M = 8192$ anchor points from Gaussian Points, which sufficiently captures dynamic details. The Transformer block in 3D motion feature lift module comprises 4 layers, yielding a 3D motion feature with $C = 128$ channels. For rendering, we adopt a variant of Gaussian Splitting Rasterization from Rade-GS[76] to obtain more accurate depth maps and geometry.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training, we randomly select 4 views as input and use 8 views for supervision. Training is conducted on four A100 GPUs with 40GB of memory each, running for a total of 15 epochs with a batch size of 16. The parameter $\\gamma$ in Eq. 10 is set to 0.2. We use the Adam optimizer with a weight decay of 0.05, and $\\beta$ values of (0.9, 0.95). The learning rate is set to $4\\times 10^{-4}$ for training on the N3DV dataset.", + "bbox": [ + 89, + 227, + 480, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Streaming Inference", + "text_level": 1, + "bbox": [ + 89, + 345, + 284, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We set $w = 5$ to construct key frame sequences, resulting in 60 keyframes from a 300-frame video, and conduct an ablation study to assess the impact of different $w$ values in Sec. 5.3. We designed two versions for keyframe optimization: a smaller version IGS-s (Ours-s) with 50 iterations refinement for Key frames, providing lower per-frame latency, and a larger version IGS- $l$ (Ours-1) with 100 iterations, which achieves higher reconstruction quality. In both versions, densification and pruning are performed every 20 iterations. For the test sequences, we construct the Gaussians for the 0th frame using the compression method provided by Lightgaussian[13], which reduces storage usage and mitigates overfitting due to sparse viewpoints. We employ 6,000 iterations for training the first frame of the N3DV dataset, compressing the number of Gaussians at 5,000 iterations. For the Meeting Room dataset, we train the Gaussians of the first frame using 15,000 iterations, compressing the number of Gaussians at 7,000 iterations. For more details, please refer to the Supp..", + "bbox": [ + 89, + 368, + 482, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 671, + 223, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Metrics and Baselines", + "text_level": 1, + "bbox": [ + 89, + 696, + 294, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines: We compare our approach to current state-of-the-art methods for dynamic scene reconstruction, covering both offline and online training methods. Offline methods[17, 31, 62, 66, 71] rely on a set of Gaussian primitives or Hex-planes to represent entire dynamic scenes. Online training methods[29, 51] employ per-frame optimization to support streaming reconstruction. Specifically, 3DGStream[51] models the movement of Gaussian points across frames by optimizing a Neural Transform Cache, creating a 3DGS-based pipeline for free-viewpoint video streaming that enables high-quality, real-time reconstruction of dynamic scenes.", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/53af74a9e5ce1d914f2908b1a9c11c5357c190fb4b2be56a897c2a63140525b6.jpg", + "table_caption": [ + "Table 1. Comparison on the N3DV dataset, with results measured at a resolution of $1352 \\times 1014$ . $\\dagger$ indicates that the evaluation was performed using the official code in the same experimental environment as ours, including the same initial point cloud. Highlights denote the best and second best results." + ], + "table_footnote": [], + "table_body": "
MethodPSNR† (dB)Train ↓ (s)Render† (FPS)Storage↓ (MB)
Offline training
Kplanes[17]32.17480.151.0
Realtime-4DGS[71]33.68-114-
4DGS[62]32.707.8300.3
Spacetime-GS[31]33.71481400.7
Saro-GS[66]33.90-401.0
Online training
StreamRF[29]32.09158.331.4
3DGStream[51]33.11122157.8
3DGStream[51]†32.7516.932047.69
Ours-s33.892.672047.90
Ours-l34.153.352047.90
", + "bbox": [ + 521, + 169, + 890, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics: Following prior work, we evaluate and report PSNR, Storage usage, Train time, and Render Speed to compare with previous methods. All metrics are averaged over the full 300-frame sequence, including frame 0.", + "bbox": [ + 511, + 381, + 903, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 453, + 653, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In-domain evaluation: We present our in-domain evaluation on two test sequences from the N3DV dataset, with results shown in Tab. 1. For a fair comparison of performance, we tested 3DGStream using the same Gaussians from the 0th frame and applied the same variant of Gaussian Splatting Rasterization as used in our approach (denoted with $\\dagger$ in the table). Compared to 3DGStream and StreamRF, our method achieves a 6x reduction in train time, with an average delay of 2.67 seconds per frame, while maintaining comparable rendering speed and storage usage. Our approach also achieves enhanced rendering quality. Compared to offline training methods, our approach provides low-delay streaming capabilities while achieving state-of-the-art rendering quality and reducing training time. A qualitative comparison of rendering quality can be seen in Fig. 5. It is evident that our method outperforms others in rendering details, such as the transition between the knife and fork, and in modeling complex dynamic scenes, like the moving hand and the shifting reflection on the wall.", + "bbox": [ + 511, + 476, + 903, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also conducted a PSNR trend comparison with 3DGStream to verify the effectiveness of our method in mitigating error accumulation. The comparison results and the smoothed trends are shown in Fig. 2. As seen, our rendering quality does not degrade with increasing frame number, while 3DGStream suffers from error accumulation, with a noticeable decline in quality as the frame number increases. This confirms the effectiveness of our approach in address", + "bbox": [ + 511, + 779, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16525", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing error accumulation. However, it is also apparent that our method exhibits more fluctuation in per-frame PSNR. This is because 3DGStream assumes small inter-frame motion, leading to smaller adjustments and smoother differences between frames.", + "bbox": [ + 89, + 90, + 483, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/018f82a665c02b8e9cf1200d87fdc023eeea23c51f98134a2b56129176979dcc.jpg", + "image_caption": [ + "Figure 3. The PSNR trend comparison on the sear steak." + ], + "image_footnote": [], + "bbox": [ + 132, + 176, + 434, + 353 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e92e7b3a22c8b3f6cd4837bbf0636d1203292a8b67010ca65794d9654d967505.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison from the Meeting Room dataset. Cross-domain evaluation: We performed a cross-domain evaluation on the Meeting Room Dataset using a model trained on N3DV. The evaluation results are presented in Tab. 2. Our method outperforms 3DGStream in rendering quality, train time, and storage efficiency, achieving streaming with just 2.77s of per-frame reconstruction time, a significant improvement over 3DGStream. This demonstrates the effectiveness and generalizability of our approach, as it enables efficient dynamic scene modeling with streaming capabilities in new environments, without requiring per-frame optimization. A qualitative comparison of rendering quality can be seen in Fig. 4. Compared to 3DGStream, which produces artifacts near moving objects, our method yields more accurate motion during large displacements, resulting in improved performance in temporally complex scenes." + ], + "image_footnote": [], + "bbox": [ + 112, + 392, + 460, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 848, + 243, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The use of the pretrained optical flow model: We used a pretrained optical flow model to extract flow embeddings", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/90180acbbbddd8f046f89e8e52cb2d58e455172d4a1eda33c7476df457f37384.jpg", + "table_caption": [ + "Table 2. Comparison on the Meeting Room dataset. $\\dagger$ indicates that the evaluation was performed using the official code in the same experimental environment as ours, including the same initial point cloud." + ], + "table_footnote": [], + "table_body": "
MethodPSNR†(dB)Train ↓(s)Render↑(FPS)Storage↓(MB)
3DGStream[51]†28.3611.512527.59
Ours-s29.242.772521.26
Ours-130.133.202521.26
", + "bbox": [ + 522, + 143, + 893, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7ee517e43fa0e06adf98bed9a8b17a3a224c684a4eeac97718cc8a2ec71c85cf.jpg", + "table_caption": [ + "Table 3. Ablation Study Results" + ], + "table_footnote": [], + "table_body": "
MethodPSNR†(dB)Train↓(s)Storage↓(MB)
No-pretrained optical flow model31.072.657.90
No-projection-aware feature lift32.952.387.90
No-points bounded refinement33.233.02110.26
Ours-s(full)33.622.677.90
", + "bbox": [ + 521, + 255, + 890, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from image pairs, which are then lifted into 3D space. To validate its effectiveness, we replaced the pretrained model with a 4-layer UNet without pretrained parameters and trained it jointly with the overall model. The results in Tab. 3 highlight the benefit of using the 2D prior.", + "bbox": [ + 511, + 349, + 905, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Projection-aware 3D Motion Feature Lift: We use a projection-based approach to lift multi-view 2D motion feature maps into 3D space, accurately linking 3D anchor points to 2D features. To evaluate its effectiveness, we replaced this method with a Transformer-based approach using cross-attention between image features and anchor points, enhanced with positional embeddings through a 4-layer Transformer block. As shown in Tab. 3, Projection-aware Feature Lift is crucial for IGS performance, with only a slight increase in training time.", + "bbox": [ + 511, + 429, + 905, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key-frame guided Streaming: We employ a key-frame-guided strategy to address error accumulation in streaming and to enhance reconstruction quality. Keyframes are selected and refined through Max-points-bounded Refine. Without this refinement, AGM-Net would rely solely on Gaussians propagated from the last keyframe, resulting in accumulated errors that significantly impact performance, as shown in Fig. 6 (a). We also evaluate the effect of max-points bounding during refinement, as shown in Tab. 3. Without point limits, storage requirements increase substantially, and overfitting causes a decline in view quality.", + "bbox": [ + 511, + 580, + 905, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key-frame selection: We conducted an ablation study on the interval $w$ for setting keyframes, testing values of $w = 1$ , $w = 5$ , and $w = 10$ , with results shown in Tab. 4. When $w = 1$ , every frame becomes a keyframe, leading to excessive optimization that overfits Gaussians to training views, degrading test view quality and increasing training time and storage. Conversely, with $w = 10$ , each keyframe drives the next 10 frames, but this distance weakens model performance, as it relies on assumptions about adjacent-frame", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16526", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7bffae3ad7ed61269bbb878e43661d24901c5be646a5c65dedc77bcc27675dec.jpg", + "image_caption": [ + "SaRo-GS" + ], + "image_footnote": [], + "bbox": [ + 133, + 90, + 277, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2bf7d0683a3665f5dabf7338b770aa73a1c163ca8c17811a8f9cab80d867413.jpg", + "image_caption": [ + "4DGS" + ], + "image_footnote": [], + "bbox": [ + 277, + 90, + 424, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e73f22f4004fbdd5f540f8e860143bdd06c71824407239939089a199cb29bd3.jpg", + "image_caption": [ + "3DGStream" + ], + "image_footnote": [], + "bbox": [ + 424, + 90, + 570, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bcca0dd79eb747ab28c2aea8646ac6954ca9d5dea01cdfa5f34dff3fe35d6a36.jpg", + "image_caption": [ + "IGS" + ], + "image_footnote": [], + "bbox": [ + 570, + 90, + 714, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9c3521f2e7e087e78bf59eae62ec64021b6bd52d2b37045789320d45a7fbb557.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 715, + 90, + 859, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/72bc753b4ee5e0508e6a0643a4a359055bc2e65c9c962d279c1e6dcc69d40f2a.jpg", + "image_caption": [ + "(a) Ablation Study on Key-frame Refinement" + ], + "image_footnote": [], + "bbox": [ + 93, + 417, + 285, + 531 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/014072f71d96268000634e44821552aacfa85e0c308b1760a85161fb961a991e.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison from the N3DV dataset.", + "(b) Per-frame reconstruction time", + "Figure 6. (a)Ablation Study on Key-frame Refinement. (b)Per-frame reconstruction time." + ], + "image_footnote": [], + "bbox": [ + 285, + 417, + 480, + 532 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e00156d7868b5024f035889961f7a71108f0966bb4d64470dc209a0a6e0c6c38.jpg", + "table_caption": [ + "Table 4. The impact of different keyframe intervals $w$ ." + ], + "table_footnote": [], + "table_body": "
MethodPSNR(dB)↑Train(s)↓Storage(MB)↓
w=133.556.3836.0
w=533.622.677.90
w=1030.142.751.26
", + "bbox": [ + 122, + 603, + 450, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "similarity. The setting $w = 5$ strikes the best balance across view synthesis quality, train time, and storage, and is thus our final choice.", + "bbox": [ + 89, + 681, + 482, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 89, + 744, + 204, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1. Independent per-frame reconstruction time", + "text_level": 1, + "bbox": [ + 89, + 771, + 459, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further evaluate the performance of IGS by analyzing the independent per-frame reconstruction time, as shown in Fig. 6 (b). The reconstruction time for each frame exhibits a periodic pattern: for candidate frames, it takes $0.8\\mathrm{s}$ , while for key frames, it takes $4\\mathrm{s}$ and $7.5\\mathrm{s}$ for the small and large versions, respectively, which are significantly smaller than the $16\\mathrm{s}$ required by 3DGStream.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Limitation", + "text_level": 1, + "bbox": [ + 513, + 417, + 632, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "IGS is the first to use a generalized method for streaming dynamic scene reconstruction, but it has limitations that can be addressed in future work. As shown in Fig. 3, our results exhibit jitter between adjacent frames, caused by the lack of temporal dependencies in the current framework. This makes the model more sensitive to noise. In contrast, 3DGStream assumes minimal motion between frames, yielding smoother results, but it fails in scenes with large motion (Fig. 4). To reduce jitter, we plan to incorporate temporal dependencies into IGS, modeling them as a time series for more robust performance.", + "bbox": [ + 511, + 439, + 906, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 618, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose IGS as a novel streaming-based method for modeling dynamic scenes. With a generalized approach, IGS achieves frame-by-frame modeling in just over 2 seconds per frame, maintaining state-of-the-art rendering quality while keeping storage low. We introduce a generalized AGM-Net that lifts 2D multi-view motion features to 3D anchor points, using these anchors to drive Gaussian motion. This allows the model to infer the motion of Gaussians between adjacent frames in a single step. Additionally, we propose a Key-frame-guided Streaming strategy, where key frame sequences are selected and refined to mitigate error accumulation, further enhancing rendering quality. Extensive in-domain and cross-domain experiments demonstrate the strong generalization capabilities of our model, reducing significant streaming average cost while achieving state-of-the-art rendering quality, render speed, and storage efficiency.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16527", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is financially supported for Outstanding Talents Training Fund in Shenzhen, this work is also financially supported by Shenzhen Science and Technology Program-Shenzhen Cultivation of Excellent Scientific and Technological Innovation Talents project(Grant No. RCJC20200714114435057), Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology(Grant No. 2024B1212010006), National Natural Science Foundation of China U21B2012.", + "bbox": [ + 89, + 114, + 483, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 266, + 187, + 282 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855-5864, 2021. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2", + "[4] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 2", + "[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 3", + "[6] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2", + "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2", + "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2", + "[9] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In European Conference on Computer Vision (ECCV), 2024. 2, 3" + ], + "bbox": [ + 99, + 291, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Yue Chen, Xingyu Chen, Xuan Wang, Qi Zhang, Yu Guo, Ying Shan, and Fei Wang. Local-to-global registration for bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8264-8273, 2023. 2", + "[11] Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. Hac: Hash-grid assisted context for 3d gaussian splatting compression. In European Conference on Computer Vision, pages 422-438. Springer, 2025. 2", + "[12] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 2", + "[13] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with $15\\mathrm{x}$ reduction and $200+$ fps. arXiv preprint arXiv:2311.17245, 2023. 2, 6", + "[14] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2, 2024. 2", + "[15] Xin Fei, Wenzhao Zheng, Yueqi Duan, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Jiwen Lu. Pixelgaussian: Generalizable 3d gaussian reconstruction from arbitrary views. arXiv preprint arXiv:2410.18979, 2024. 2", + "[16] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2", + "[17] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbaek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 1, 3, 6", + "[18] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 2", + "[19] Xiang Guo, Jiadai Sun, Yuchao Dai, Guanying Chen, Xiaoting Ye, Xiao Tan, Errui Ding, Yumeng Zhang, and Jingdong Wang. Forward flow for novel view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16022-16033, 2023. 3", + "[20] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. ICCV, 2021. 2", + "[21] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19774–19783, 2023. 2", + "[22] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac" + ], + "bbox": [ + 516, + 92, + 906, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16528", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "curate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2", + "[23] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. arXiv preprint arXiv:2312.14937, 2023. 2, 3", + "[24] Haian Jin, Hanwen Jiang, Hao Tan, Kai Zhang, Sai Bi, Tianyuan Zhang, Fujun Luan, Noah Snavely, and Zexiang Xu. Lvsm: A large view synthesis model with minimal 3d inductive bias. arXiv preprint arXiv:2410.17242, 2024. 2", + "[25] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. 2", + "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 5", + "[27] Hao Li, Yuanyuan Gao, Dingwen Zhang, Chenming Wu, Yalun Dai, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Junwei Han. Gprt: Towards generalizable 3d gaussians without pose priors in real-time. arXiv preprint arXiv:2403.10147, 2024. 2", + "[28] Haolin Li, Jinyang Liu, Mario Sznaier, and Octavia Camps. 3d-hgs: 3d half-gaussian splatting. arXiv preprint arXiv:2406.02720, 2024. 2", + "[29] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 1, 2, 3, 5, 6", + "[30] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 5", + "[31] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. arXiv preprint arXiv:2312.16812, 2023. 1, 2, 3, 6", + "[32] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741-5751, 2021. 2", + "[33] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. High-fidelity and real-time novel view synthesis for dynamic scenes. In SIGGRAPH Asia 2023 Conference Papers, pages 1-9, 2023. 3", + "[34] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. Advances in Neural Information Processing Systems, 35:36762-36775, 2022. 3", + "[35] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splattering reconstruction from multi-view stereo. In European Conference on Computer Vision, pages 37-53. Springer, 2025. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 5", + "[37] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2", + "[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2", + "[40] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 2", + "[41] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 3", + "[42] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022.3", + "[43] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 3", + "[44] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 2", + "[45] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2", + "[46] Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. arXiv preprint arXiv:2403.17898, 2024. 2", + "[47] Christian Schmidt, Jens Piekenbrinck, and Bastian Leibe. Look gauss, no pose: Novel view synthesis using gaussian splatting without accurate pose initialization. In IROS, 2024. 2", + "[48] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16529", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 3", + "[49] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 3", + "[50] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2", + "[51] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20675-20685, 2024. 1, 2, 3, 4, 6, 7", + "[52] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, João F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. arXiv preprint arXiv:2406.04343, 2024. 2", + "[53] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2", + "[54] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2", + "[55] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 2", + "[56] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2", + "[57] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis, 2021. 3", + "[58] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 2", + "[59] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 76-87, 2023. 3", + "[60] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4150-4159, 2023. 2", + "[61] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2021. 2", + "[62] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 1, 2, 3, 6", + "[63] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2", + "[64] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 5", + "[65] Han Xu, Jiteng Yuan, and Jiayi Ma. Murf: Mutually reinforcing multi-modal image registration and fusion. IEEE transactions on pattern analysis and machine intelligence, 45(10):12148-12166, 2023. 2", + "[66] Jinbo Yan, Rui Peng, Luyang Tang, and Ronggang Wang. 4d gaussian splatting with scale-aware residual field and adaptive optimization for real-time rendering of temporally complex dynamic scenes. In Proceedings of the 32nd ACM International Conference on Multimedia, page 7871-7880, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3, 6", + "[67] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. 2", + "[68] Runyi Yang, Zhenxin Zhu, Zhou Jiang, Baijun Ye, Xiaoxue Chen, Yifei Zhang, Yuantao Chen, Jian Zhao, and Hao Zhao. Spectrally pruned gaussian fields with neural compensation. arXiv preprint arXiv:2405.00676, 2024. 2", + "[69] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3", + "[70] Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. arXiv preprint arXiv:2402.15870, 2024. 2", + "[71] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and render-" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16530", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing with 4d gaussian splatting. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 6", + "[72] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2", + "[73] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 2", + "[74] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2", + "[75] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2", + "[76] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2, 6", + "[77] Chuanrui Zhang, Yingshuang Zou, Zhuoling Li, Minmin Yi, and Haoqian Wang. Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. arXiv preprint arXiv:2408.13770, 2024. 2", + "[78] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2", + "[79] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. In ECCV, 2024. 2", + "[80] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 2", + "[81] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2" + ], + "bbox": [ + 91, + 92, + 482, + 727 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "16531", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/c9f2cc0e-b9b5-4129-a1a3-c1fc202a4666_model.json b/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/c9f2cc0e-b9b5-4129-a1a3-c1fc202a4666_model.json new file mode 100644 index 0000000000000000000000000000000000000000..699f9e76210901261faab01bdb8c8fe7f9e0d0d4 --- /dev/null +++ b/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/c9f2cc0e-b9b5-4129-a1a3-c1fc202a4666_model.json @@ -0,0 +1,2787 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.13, + 0.898, + 0.178 + ], + "angle": 0, + "content": "Instant Gaussian Stream: Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.202, + 0.777, + 0.24 + ], + "angle": 0, + "content": "Jinbo Yan\\(^{1}\\), Rui Peng\\(^{1,2}\\), Zhiyan Wang\\(^{1}\\), Luyang Tang\\(^{1,2}\\), Jiayu Yang\\(^{1,2}\\), Jie Liang\\(^{1}\\), Jiahao Wu\\(^{1}\\), Ronggang Wang\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.24, + 0.875, + 0.294 + ], + "angle": 0, + "content": "1Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University 2Pengcheng Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.295, + 0.79, + 0.327 + ], + "angle": 0, + "content": "{yjb, ruipeng, zywang23, tly926, liangjie, wjh0616}@stu.pku.edu.cn jiayuyang@pku.edu.cn rgwang@pkusz.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.364, + 0.206, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.531, + 0.264, + 0.544 + ], + "angle": 0, + "content": "(a)High Quality Result" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.364, + 0.321, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.365, + 0.609, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.531, + 0.547, + 0.544 + ], + "angle": 0, + "content": "(b)Realtime Rendering" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.362, + 0.907, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.53, + 0.905, + 0.543 + ], + "angle": 0, + "content": "(c)Comparison on Quality and Train Speed" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.555, + 0.908, + 0.596 + ], + "angle": 0, + "content": "Figure 1. Performance comparison with pervious SOTA[17, 29, 31, 51, 62]. Our method achieves a per-frame reconstruction time of 2.67s, delivering high-quality rendering results in a streaming fashion (a)(b), with a noticeable improvement in performance (c). * denotes a streamable method." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.597, + 0.327, + 0.613 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Building Free-Viewpoint Videos in a streaming manner offers the advantage of rapid responsiveness compared to offline training methods, greatly enhancing user experience. However, current streaming approaches face challenges of high per-frame reconstruction time \\((10s+)\\) and error accumulation, limiting their broader application. In this paper, we propose Instant Gaussian Stream (IGS), a fast and generalizable streaming framework, to address these issues. First, we introduce a generalized Anchor-driven Gaussian Motion Network, which projects multi-view 2D motion features into 3D space, using anchor points to drive the motion of all Gaussians. This generalized Network generates the motion of Gaussians for each target frame in the time required for a single inference. Second, we propose a Key-frame-guided Streaming Strategy that refines each key frame, enabling accurate reconstruction of temporally complex scenes while mitigating error accumulation. We conducted extensive in-domain and cross-domain evalua" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.645 + ], + "angle": 0, + "content": "tions, demonstrating that our approach can achieve streaming with a average per-frame reconstruction time of \\(2s+\\), alongside a enhancement in view synthesis quality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.677, + 0.646, + 0.692 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.703, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Reconstructing Free-Viewpoint Videos (FVV) from multiview images is a valuable area of research, with applications spanning immersive media such as VR, AR, and sports broadcasting. By enabling interactive, photorealistic visuals, FVVs from dynamic scenes hold the potential to become a next-generation visual medium, offering experiences that go beyond traditional video formats. To enhance user experience, streaming-based FVV construction—where dynamic scenes are reconstructed frame by frame—offers a low-delay response compared to traditional offline training approaches, making it better suited for real-time, interactive applications." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "With advancements in real-time rendering and high-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16520" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.364 + ], + "angle": 0, + "content": "quality view synthesis powered by 3D Gaussian Splatting (3DGS)[26], dynamic scene reconstruction has seen rapid progress. Some offline training methods[23, 31, 62, 66, 69, 71] achieve high-quality view synthesis but require collecting all frames before training can begin. This limitation makes them less suitable for scenarios that demand fast response times, such as live streaming and virtual meetings. To address these challenges, some methods[29, 51] adopt a streaming framework that reconstructs dynamic scenes frame by frame by modeling inter-frame differences. However, streaming-based dynamic scene reconstruction still faces significant challenges. First, current methods typically require per-frame optimization, resulting in high per-frame latencies \\((10\\mathrm{s}+)\\), which severely impact the real-time usability of these systems. Additionally, error accumulation across frames degrades the reconstruction quality of later frames, making it difficult for streaming methods to scale effectively to longer video sequences." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.366, + 0.483, + 0.655 + ], + "angle": 0, + "content": "To promote the streaming framework to be more practical, we introduce Instant Gaussian Stream (IGS), a streaming approach for dynamic scene reconstruction that achieves a per-frame reconstruction time of \\(2s+\\), mitigates error accumulation, and enhances view synthesis quality. First, to tackle the issue of high per-frame reconstruction time, we developed a generalized Anchor-driven Gaussian Motion Network (AGM-Net). This network utilizes a set of key points, called anchor points, to carry motion features that guide Gaussian transformations. This design allows the inference process to compute the motion of Gaussian primitives between frames in a single feedforward pass, eliminating the need for per-frame optimization. Second, to further improve view synthesis quality and minimize error accumulation, we propose a Key-frame-guided Streaming strategy. By establishing key-frame sequences and performing maxpoint-bounded refinement on key frames, our method mitigates the impact of error accumulation and enhances rendering quality in temporally complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.793 + ], + "angle": 0, + "content": "We conducted extensive validation in both in-domain and cross-domain scenarios, and the experimental results demonstrate the strong generalization capability of our model, with significant improvements over current state-of-the-art methods in terms of per-frame reconstruction time and rendering quality. To the best of our knowledge, this is the first approach to use a generalized method for streaming reconstruction of dynamic scenes. Our contributions are summarized below." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.796, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- We propose a generalized Anchor-driven Gaussian Motion Network that captures Gaussian motion between adjacent frames with a single inference, eliminating the need for frame-by-frame optimization." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.857, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- We designed a Key-frame-guided Streaming strategy to enhance our method's capability in handling temporally complex scenes, improving overall view synthesis qual" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "ity within the streaming framework and mitigating error accumulation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.122, + 0.907, + 0.244 + ], + "angle": 0, + "content": "- The evaluation results in both in-domain and cross-domain scenarios demonstrate the generalization capability of our method and its state-of-the-art performance. We achieve a 2.7 s per-frame reconstruction time for streaming, representing a significant improvement over previous methods. Additionally, we improve view synthesis quality, enabling real-time rendering at 204 FPS while maintaining comparable storage overhead." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.274, + 0.652, + 0.29 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.305, + 0.846, + 0.321 + ], + "angle": 0, + "content": "2.1.3D Reconstruction and View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.908, + 0.605 + ], + "angle": 0, + "content": "Novel view synthesis (NVS) has always been a hot topic in the field of computer vision. By using MLP to implicitly represent the scene, Neural Radiance Fields (NeRF) [38] achieves realistic rendering. Subsequent works have imporced NeRF to enhance rendering quality [1, 2, 60], reduce the number of training views [40, 58, 63, 67], lessen dependence on camera poses [4, 10, 32, 56], and improve both training and inference speeds [3, 8, 16, 20, 21, 39, 44, 45, 50]. 3D Gaussian Splitting (3DGS) [26] employs anisotropic Gaussian primitives to represent scenes and introduces rasterization-based splatting rendering algorithm, enhancing both speed and rendering quality. Some methods focus on various aspects of improving Gaussian field representations, including rendering quality[28, 37, 46, 70, 74, 79], enhancing geometric accuracy[22, 75, 76], and increasing compression efficiency, [11, 13, 37, 68], joint optimization of camera pose and gaussian fields [14, 18, 47], as well 3D generation [9, 54, 55, 81]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.632, + 0.905, + 0.661 + ], + "angle": 0, + "content": "2.2. Generalizable 3D Reconstruction for Acceleration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "3DGS requires per-scene optimization to achieve realistic rendering results. To accelerate this time-consuming process, some works [24, 27, 52, 53, 77, 80], inspired by generalizable NeRF [7, 25, 61, 65, 73], have proposed to train generalizable Gaussian models on large-scale datasets to enable fast reconstruction. PixelSplat [6] utilizes an Transformer to encode features and decode them into Gaussian attributes. Other generalizable models [12, 15, 35, 78] utilize Transformers or Multi-View Stereo (MVS) [72] techniques to construct cost volumes followed by a decoder, achieving real-time rendering speeds and excellent generalizability. To the best of our knowledge, our work is the first to apply generalizable models to dynamic streaming scenes, utilizing their rapid inference capabilities to accelerate the processing of dynamic scenes reconstruction." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16521" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.483, + 0.121 + ], + "angle": 0, + "content": "2.3. Dynamic Scene Reconstruction and View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.129, + 0.484, + 0.219 + ], + "angle": 0, + "content": "There have been numerous efforts to extend static scene reconstruction to dynamic scenes based on NeRF[5, 17, 19, 33, 34, 41, 43, 48, 57]. Since the advent of 3D Gaussian Splatting (3DGS)[26], researchers have explored incorporating its real-time rendering capabilities into dynamic scene reconstruction[23, 31, 62, 66, 69, 71]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.22, + 0.484, + 0.477 + ], + "angle": 0, + "content": "However, these approaches rely on offline training with full video sequences, making them unsuitable for applications requiring real-time interaction and fast response. To address this issue, existing methods such as StreamRF[29], NeRFPlayer[49], ReRF[59], and 3DGStream[51] reformulate the dynamic modeling problem using a Streaming method. Notably, 3DGStream[51], based on Gaussian Splatting, optimizes a Neural Transformation Cache to model Gaussian movements between frames, further improving the performance. Although these methods achieve promising results, they still rely on per-frame optimization, resulting in significant delays (with current SOTA methods requiring over 10 seconds per frame[51]). Our approach offers a new perspective for streaming dynamic scene modeling: by training a generalized network, we eliminate the need for per-frame optimization, achieving low per-frame reconstruction time alongside high rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.491, + 0.182, + 0.507 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.639 + ], + "angle": 0, + "content": "In this section, we begin with an overview of the pipeline in Sec.3.1. Then, in Sec. 3.3, we introduce the Anchordriven Gaussian Motion Network (AGM-Net), a generalized model that drives Gaussian motion from the previous frame using anchor points, which serve as key points in the 3D scene. Following this, we present our Key-frame-guided Streaming strategy in Sec.3.4. Finally, in Sec.3.5, we outline the loss function used in our training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.201, + 0.663 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Our goal is to model dynamic scenes in a streaming manner with minimal per-frame reconstruction time. To achieve this, we adopt a generalized AGM-Net that extracts 3D motion features from the scene using anchor points and drives the motion of Gaussian primitives between frames in a single inference step. And we propose a key-frame-guided Streaming strategy to further improve view synthesis quality and handle temporally complex scenes while addressing error accumulation in streaming reconstruction. The overall pipeline is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.22, + 0.849 + ], + "angle": 0, + "content": "3.2. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Gaussian splatting[26] represents static scenes as a collection of anisotropic 3D Gaussians. The color of each pixel is obtained through point-based alpha blending rendering," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.866, + 0.107 + ], + "angle": 0, + "content": "enabling high-fidelity real-time novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.153 + ], + "angle": 0, + "content": "Specifically, each Gaussian primitive \\(\\mathcal{G}_i\\) is parameterized by a center \\(\\mu \\in \\mathbb{R}^3\\), 3D covariance matrix \\(\\Sigma \\in \\mathbb{R}^{3\\times 3}\\), opacity \\(\\alpha \\in \\mathbb{R}\\), and color \\(c\\in \\mathbb{R}^{3(n + 1)^2}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.164, + 0.905, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} (x) = e ^ {- \\frac {1}{2} (x - \\mu) ^ {T} \\Sigma^ {- 1} (x - \\mu)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.196, + 0.907, + 0.258 + ], + "angle": 0, + "content": "During rendering, the 3D Gaussian is first projected onto 2D space. Subsequently, the Gaussians covering a pixel are sorted based on depth. The color of the pixel \\(\\mathbf{c}\\) is obtained using point-based alpha blending rendering:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.268, + 0.905, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} = \\sum_ {i = 1} ^ {n} c _ {i} \\alpha_ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {i} ^ {\\prime}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.322, + 0.905, + 0.352 + ], + "angle": 0, + "content": "Here, \\(\\alpha^{\\prime}\\) represents the opacity after projection onto the 2D space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.362, + 0.869, + 0.376 + ], + "angle": 0, + "content": "3.3. Anchor-driven Gaussian Motion Network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.906, + 0.519 + ], + "angle": 0, + "content": "Motion Feature Maps: Given multi-view images of current frames \\(\\mathbf{I}' = (I_1', \\dots, I_V')\\) with camera parameters, We can first construct a multi-view image pair, which contains the current frame and the previous frame \\(\\mathbf{I}\\) from corresponding viewpoints. Then, we use a optical flow model to obtain the intermediate flow embeddings. Next, a modulation layer[9, 42] is applied to inject the viewpoint and depth information into the embeddings, ultimately resulting in 2D motion feature maps \\(\\mathbf{F} \\in \\mathbb{R}^{V \\times C \\times H \\times W}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.702 + ], + "angle": 0, + "content": "Anchor Sampling: To deform the Gaussian primitives \\(\\mathcal{G}\\) from the previous frame, we need to compute the motion of each Gaussian. However, directly computing the motion for each Gaussian is computationally expensive and memory-intensive due to the large number of Gaussian points. To address this, we employ an anchor-point-based approach to represent the motion features of the entire scene in 3D space. The anchor-driven approach supports batch processing during training, reducing computational overhead while preserving the geometric information of the Gaussian primitives. Specifically, we use Farthest Point Sampling (FPS) to sample \\(M\\) anchor points from the \\(N\\) Gaussian primitives" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.715, + 0.905, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} = \\mathbf {F P S} \\left(\\left\\{\\mu_ {i} \\right\\} _ {i \\in N}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.78 + ], + "angle": 0, + "content": "where \\(\\mathcal{C} \\in \\mathbb{R}^{M \\times 3}\\) represents the sampled anchor points with \\(M\\) set to 8192 in our experiments, and \\(\\mu_{i}\\) denotes the position of \\(\\mathcal{G}_{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Projection-aware 3D Motion Feature Lift: We adopt a projection-aware approach to lift multiview 2D motion features into 3D space. Specifically, we project sampled anchor points onto each motion feature map based on the camera poses, obtaining high-resolution motion features:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.867, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\nf _ {i} = \\frac {1}{V} \\sum_ {j \\in V} \\Psi \\left(\\Pi_ {j} \\left(\\mathcal {C} _ {i}\\right), F _ {j}\\right) \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16522" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.91, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.402, + 0.908, + 0.501 + ], + "angle": 0, + "content": "Figure 2. The overall pipeline of IGS. (a) Starting from the key frame and moving towards the target frame, we extract the 2D Motion Feature Map. (b) Then we sample M anchor points from the Gaussian primitives of the key frame, (c) and the anchor points are projected onto these feature maps to obtain 3D motion features through Projection-aware Motion Feature Lift. (d) Each Gaussian point interpolates its own motion feature from neighboring anchors and applies a weighted aggregation of features, which is then decoded into the motion of the Gaussian between the key frame and the target frame. (e) The entire streaming reconstruction process is guided by the Key-frame-guided Streaming strategy, where the key frame directly infers subsequent candidate frames until the next key-frame is reached, at which point max-point bounded refinement is applied to the key-frame." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.482, + 0.599 + ], + "angle": 0, + "content": "where \\(\\Pi_j(\\mathcal{C}_i)\\) represents the projection of \\(\\mathcal{C}_i\\) onto the image plane of \\(F_{j}\\) using the camera parameters of \\(F_{j}\\), and \\(\\Psi\\) denotes bilinear interpolation. By projection, each anchor point can accurately obtain its feature \\(f_{i}\\in \\mathbb{R}^{C}\\) from the multi-view feature map, effectively lifting the 2D motion map into 3D space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.659 + ], + "angle": 0, + "content": "We then use these features \\(\\{f_i\\}_{i\\in M}\\), stored at each anchor point, as input to a Transformer block using self-attention to further capture motion information within the 3D scene." + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.671, + 0.483, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\left\\{z _ {i}: z _ {i} \\in \\mathbb {R} ^ {C} \\right\\} _ {i \\in M} = \\operatorname {T r a n s f o r m e r} \\left(\\left\\{f _ {i} \\right\\} _ {i \\in M}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.791 + ], + "angle": 0, + "content": "The output of the Transformer block \\(\\{z_i\\}_{i\\in M}\\) represents the final 3D motion features we obtain. Now, we can use these 3D motion features to represent the motion information of an anchor and its neighborhood, and drive the motion of the neighboring Gaussian points based on these motion features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Interpolate and Motion Decode: Using the 3D motion features stored at anchor points, we can assign each Gaussian point a motion feature by interpolating from its K nearest anchors in the neighborhood:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.865, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\nz _ {i} = \\frac {\\sum_ {k \\in \\mathcal {N} (i)} e ^ {- d _ {k}} z _ {k}}{\\sum_ {k \\in \\mathcal {N} (i)} e ^ {- d _ {k}}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.584 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(i)\\) represents the set of neighboring anchor points of Gaussian point \\(\\mathcal{G}_i\\), and \\(d_k\\) represents the Euclidean distance from Gaussian point \\(G_i\\) to anchor \\(\\mathcal{C}_k\\). Then we can use a Linear head to decode the Motion feature to the movement of a Gaussian primitive:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.597, + 0.907, + 0.613 + ], + "angle": 0, + "content": "\\[\nd \\mu_ {i}, d r o t _ {i} = \\mathbf {L i n e a r} \\left(z _ {i}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.907, + 0.684 + ], + "angle": 0, + "content": "here, we use the deformation of the Gaussian's position \\(d\\mu_{i}\\) and the deformation of the rotation \\(drot_{i}\\), to represent the movement of a Gaussian primitive. The new position and rotation of the Gaussian are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.657, + 0.695, + 0.907, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\mu_ {i} ^ {\\prime} = \\mu_ {i} + d \\mu_ {i}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.739, + 0.905, + 0.757 + ], + "angle": 0, + "content": "\\[\n\\operatorname {r o t} _ {i} ^ {\\prime} = \\operatorname {n o r m} \\left(\\operatorname {r o t} _ {i}\\right) \\times \\operatorname {n o r m} \\left(\\operatorname {d r o t} _ {i}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.809 + ], + "angle": 0, + "content": "here ' refers to the new attributes. norm denotes to quaternion normalization and \\(\\times\\) represents quaternion multiplication, as used in previous work[51]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.776, + 0.835 + ], + "angle": 0, + "content": "3.4. Key-frame-guided Streaming" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Using AGM-Net, we can transition Gaussian primitives from the previous frame to the current frame within a single forward inference pass. However, this process only adjusts the position and rotation of Gaussian primitives, making it" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16523" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "effective for capturing rigid motion but inadequate for accurately representing non-rigid motion. Furthermore, the number of Gaussian points remains constant, limiting its capacity to model temporally dynamic scenes where objects may appear or disappear. These limitations result in challenges in capturing scene dynamics and can lead to error accumulation across frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.482, + 0.336 + ], + "angle": 0, + "content": "To better model object changes and reduce error accumulation, we propose a Key-frame-guided Streaming strategy that uses key frames as the initial state for deforming Gaussians in subsequent frames. We also introduce a Max points bounded refinement strategy, enabling efficient key frame reconstruction without redundant points and preventing point count growth across frames. This approach helps avoid overfitting in sparse-viewpoint scenes by effectively managing point density." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.483, + 0.489 + ], + "angle": 0, + "content": "Key-frame-guided strategy: Starting from frame 0, we designate a key frame every \\( w \\) frames, forming a key-frame sequence \\( \\{K_0, K_w, \\dots, K_{nw}\\} \\). The remaining frames serve as candidate frames. During streaming reconstruction, for example, beginning with a key frame \\( K_{iw} \\), we deform the Gaussians forward across successive candidate frames using AGM-Net until reaching the next key frame \\( K_{(i+1)w} \\). At this point, we refine the deformed Gaussians of key frame \\( K_{(i+1)w} \\). Then, we continue deforming from key frame \\( K_{(i+1)w} \\) to process subsequent frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.655 + ], + "angle": 0, + "content": "This key-frame-guided strategy offers several advantages. First, when AGM-Net is applied to candidate frames, it is always start from the most recent key frame, preventing error propagation across candidate frames between key frames and eliminating cumulative error. Second, candidate frames do not require optimization-based refinement, as their Gaussians are generated through a single model inference with AGM-Net, ensuring low per-frame reconstruction time. Additionally, we can batch process up to \\( w \\) frames following each key frame, which further accelerates our pipeline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Max points bounded Key-frame Refinement: During the refinement of each key frame, we optimize all parameters of the Gaussians and support cloning, splitting, and filtering, which is same to 3DGS[26]. This approach allows us to handle object deformations as well as the appearance and disappearance of objects in temporally complex scenes, effectively preventing error accumulation from key frame to subsequent frames. However, this optimization strategy can lead to a gradual increase in Gaussian primitives at each key frame, which not only raises computational complexity and storage requirements but also risks overfitting in sparse viewpoints, particularly in dynamic scenes where viewpoints are generally limited." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address this, we adopt a Max Points Bounded Refine method. When densifying Gaussian points, we control the number of Gaussians allowed to densify by adjusting each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "point's gradient, ensuring that the total number of points does not exceed a predefined maximum." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.132, + 0.661, + 0.146 + ], + "angle": 0, + "content": "3.5. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.154, + 0.905, + 0.274 + ], + "angle": 0, + "content": "Our training process consists of two parts: offline training the generalized AGM-Net and performing online training for the key frames. The generalized AGM-Net only needs to be trained once, and it can generalize to multiple scenes. We train the AGM-Net across scenes using gradient descent, relying solely on a view synthesis loss between our predicted views and the ground truth views, which includes an \\(\\mathcal{L}_1\\) term and an \\(\\mathcal{L}_{D - SSIM}\\) term and can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.287, + 0.905, + 0.303 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {D - S S I M} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.315, + 0.905, + 0.376 + ], + "angle": 0, + "content": "When performing online training on the Gaussians in key frames, we use the same loss function as in Eq.10. However, this time, we optimize the attributes of the Gaussian primitives rather than the parameters of the neural network." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.389, + 0.733, + 0.406 + ], + "angle": 0, + "content": "4. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.905, + 0.505 + ], + "angle": 0, + "content": "In this Section, we first introduce the datasets we used, along with the partitioning and preprocessing of training data, in Sec. 4.1. Next, we provide a detailed explanation of the configuration of the AGM network and the training hyperparameters in Sec. 4.2. Finally, we describe the detailed setup for streaming in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.616, + 0.529 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.905, + 0.597 + ], + "angle": 0, + "content": "The Neural 3D Video Datasets (N3DV)[30] includes 6 dynamic scenes recorded using a multi-view setup featuring 21 cameras, with a resolution of \\(2704 \\times 2028\\). Each multiview video comprises 300 frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.905, + 0.642 + ], + "angle": 0, + "content": "Meeting Room Datasets[29] includes 3 dynamic scenes recorded with 13 cameras at a resolution of \\(1280 \\times 720\\). Each multi-view video also contains 300 frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.794 + ], + "angle": 0, + "content": "Dataset Preparation: We split four sequences from the N3DV dataset into the training set, with the remaining two sequences, {cut roasted beef, sear steak}, used as the test set. For the training set, we constructed 3D Gaussians for all frames in the four training sequences, totaling 1200 frames, which required 192 GPU hours. For each frame's 3D Gaussian, we performed motions forward and backward for five frames, creating 12,000 pairs for training. For testing, we selected one viewpoint for evaluation for both datasets, consistent with previous methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.665, + 0.818 + ], + "angle": 0, + "content": "4.2.AGM Network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We use GM-Flow[64] to extract optical flow embeddings and add a Swin-Transformer[36] block for fine-tuning while keeping the other parameters of GM-Flow fixed. Our AGM model accepts an arbitrary number of input views. To balance computational complexity and performance, we use" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16524" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "\\(V = 4\\) views, each producing a motion map with \\(C = 128\\) channels and a resolution of \\(128 \\times 128\\). We sample \\(M = 8192\\) anchor points from Gaussian Points, which sufficiently captures dynamic details. The Transformer block in 3D motion feature lift module comprises 4 layers, yielding a 3D motion feature with \\(C = 128\\) channels. For rendering, we adopt a variant of Gaussian Splitting Rasterization from Rade-GS[76] to obtain more accurate depth maps and geometry." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.482, + 0.336 + ], + "angle": 0, + "content": "During training, we randomly select 4 views as input and use 8 views for supervision. Training is conducted on four A100 GPUs with 40GB of memory each, running for a total of 15 epochs with a batch size of 16. The parameter \\(\\gamma\\) in Eq. 10 is set to 0.2. We use the Adam optimizer with a weight decay of 0.05, and \\(\\beta\\) values of (0.9, 0.95). The learning rate is set to \\(4\\times 10^{-4}\\) for training on the N3DV dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.346, + 0.285, + 0.362 + ], + "angle": 0, + "content": "4.3. Streaming Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.483, + 0.656 + ], + "angle": 0, + "content": "We set \\( w = 5 \\) to construct key frame sequences, resulting in 60 keyframes from a 300-frame video, and conduct an ablation study to assess the impact of different \\( w \\) values in Sec. 5.3. We designed two versions for keyframe optimization: a smaller version IGS-s (Ours-s) with 50 iterations refinement for Key frames, providing lower per-frame latency, and a larger version IGS- \\( l \\) (Ours-1) with 100 iterations, which achieves higher reconstruction quality. In both versions, densification and pruning are performed every 20 iterations. For the test sequences, we construct the Gaussians for the 0th frame using the compression method provided by Lightgaussian[13], which reduces storage usage and mitigates overfitting due to sparse viewpoints. We employ 6,000 iterations for training the first frame of the N3DV dataset, compressing the number of Gaussians at 5,000 iterations. For the Meeting Room dataset, we train the Gaussians of the first frame using 15,000 iterations, compressing the number of Gaussians at 7,000 iterations. For more details, please refer to the Supp.." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.224, + 0.689 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.295, + 0.712 + ], + "angle": 0, + "content": "5.1. Metrics and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Baselines: We compare our approach to current state-of-the-art methods for dynamic scene reconstruction, covering both offline and online training methods. Offline methods[17, 31, 62, 66, 71] rely on a set of Gaussian primitives or Hex-planes to represent entire dynamic scenes. Online training methods[29, 51] employ per-frame optimization to support streaming reconstruction. Specifically, 3DGStream[51] models the movement of Gaussian points across frames by optimizing a Neural Transform Cache, creating a 3DGS-based pipeline for free-viewpoint video streaming that enables high-quality, real-time reconstruction of dynamic scenes." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.16 + ], + "angle": 0, + "content": "Table 1. Comparison on the N3DV dataset, with results measured at a resolution of \\(1352 \\times 1014\\). \\(\\dagger\\) indicates that the evaluation was performed using the official code in the same experimental environment as ours, including the same initial point cloud. Highlights denote the best and second best results." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.17, + 0.892, + 0.376 + ], + "angle": 0, + "content": "
MethodPSNR† (dB)Train ↓ (s)Render† (FPS)Storage↓ (MB)
Offline training
Kplanes[17]32.17480.151.0
Realtime-4DGS[71]33.68-114-
4DGS[62]32.707.8300.3
Spacetime-GS[31]33.71481400.7
Saro-GS[66]33.90-401.0
Online training
StreamRF[29]32.09158.331.4
3DGStream[51]33.11122157.8
3DGStream[51]†32.7516.932047.69
Ours-s33.892.672047.90
Ours-l34.153.352047.90
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.382, + 0.905, + 0.444 + ], + "angle": 0, + "content": "Metrics: Following prior work, we evaluate and report PSNR, Storage usage, Train time, and Render Speed to compare with previous methods. All metrics are averaged over the full 300-frame sequence, including frame 0." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.454, + 0.654, + 0.47 + ], + "angle": 0, + "content": "5.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.477, + 0.905, + 0.779 + ], + "angle": 0, + "content": "In-domain evaluation: We present our in-domain evaluation on two test sequences from the N3DV dataset, with results shown in Tab. 1. For a fair comparison of performance, we tested 3DGStream using the same Gaussians from the 0th frame and applied the same variant of Gaussian Splatting Rasterization as used in our approach (denoted with \\(\\dagger\\) in the table). Compared to 3DGStream and StreamRF, our method achieves a 6x reduction in train time, with an average delay of 2.67 seconds per frame, while maintaining comparable rendering speed and storage usage. Our approach also achieves enhanced rendering quality. Compared to offline training methods, our approach provides low-delay streaming capabilities while achieving state-of-the-art rendering quality and reducing training time. A qualitative comparison of rendering quality can be seen in Fig. 5. It is evident that our method outperforms others in rendering details, such as the transition between the knife and fork, and in modeling complex dynamic scenes, like the moving hand and the shifting reflection on the wall." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We also conducted a PSNR trend comparison with 3DGStream to verify the effectiveness of our method in mitigating error accumulation. The comparison results and the smoothed trends are shown in Fig. 2. As seen, our rendering quality does not degrade with increasing frame number, while 3DGStream suffers from error accumulation, with a noticeable decline in quality as the frame number increases. This confirms the effectiveness of our approach in address" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16525" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.166 + ], + "angle": 0, + "content": "ing error accumulation. However, it is also apparent that our method exhibits more fluctuation in per-frame PSNR. This is because 3DGStream assumes small inter-frame motion, leading to smaller adjustments and smoother differences between frames." + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.177, + 0.436, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.354, + 0.456, + 0.366 + ], + "angle": 0, + "content": "Figure 3. The PSNR trend comparison on the sear steak." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.393, + 0.462, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.587, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison from the Meeting Room dataset. Cross-domain evaluation: We performed a cross-domain evaluation on the Meeting Room Dataset using a model trained on N3DV. The evaluation results are presented in Tab. 2. Our method outperforms 3DGStream in rendering quality, train time, and storage efficiency, achieving streaming with just 2.77s of per-frame reconstruction time, a significant improvement over 3DGStream. This demonstrates the effectiveness and generalizability of our approach, as it enables efficient dynamic scene modeling with streaming capabilities in new environments, without requiring per-frame optimization. A qualitative comparison of rendering quality can be seen in Fig. 4. Compared to 3DGStream, which produces artifacts near moving objects, our method yields more accurate motion during large displacements, resulting in improved performance in temporally complex scenes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.245, + 0.866 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The use of the pretrained optical flow model: We used a pretrained optical flow model to extract flow embeddings" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.906, + 0.144 + ], + "angle": 0, + "content": "Table 2. Comparison on the Meeting Room dataset. \\(\\dagger\\) indicates that the evaluation was performed using the official code in the same experimental environment as ours, including the same initial point cloud." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.145, + 0.894, + 0.226 + ], + "angle": 0, + "content": "
MethodPSNR†(dB)Train ↓(s)Render↑(FPS)Storage↓(MB)
3DGStream[51]†28.3611.512527.59
Ours-s29.242.772521.26
Ours-130.133.202521.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.614, + 0.242, + 0.807, + 0.256 + ], + "angle": 0, + "content": "Table 3. Ablation Study Results" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.256, + 0.891, + 0.344 + ], + "angle": 0, + "content": "
MethodPSNR†(dB)Train↓(s)Storage↓(MB)
No-pretrained optical flow model31.072.657.90
No-projection-aware feature lift32.952.387.90
No-points bounded refinement33.233.02110.26
Ours-s(full)33.622.677.90
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.906, + 0.427 + ], + "angle": 0, + "content": "from image pairs, which are then lifted into 3D space. To validate its effectiveness, we replaced the pretrained model with a 4-layer UNet without pretrained parameters and trained it jointly with the overall model. The results in Tab. 3 highlight the benefit of using the 2D prior." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.906, + 0.581 + ], + "angle": 0, + "content": "Projection-aware 3D Motion Feature Lift: We use a projection-based approach to lift multi-view 2D motion feature maps into 3D space, accurately linking 3D anchor points to 2D features. To evaluate its effectiveness, we replaced this method with a Transformer-based approach using cross-attention between image features and anchor points, enhanced with positional embeddings through a 4-layer Transformer block. As shown in Tab. 3, Projection-aware Feature Lift is crucial for IGS performance, with only a slight increase in training time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.75 + ], + "angle": 0, + "content": "Key-frame guided Streaming: We employ a key-frame-guided strategy to address error accumulation in streaming and to enhance reconstruction quality. Keyframes are selected and refined through Max-points-bounded Refine. Without this refinement, AGM-Net would rely solely on Gaussians propagated from the last keyframe, resulting in accumulated errors that significantly impact performance, as shown in Fig. 6 (a). We also evaluate the effect of max-points bounding during refinement, as shown in Tab. 3. Without point limits, storage requirements increase substantially, and overfitting causes a decline in view quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Key-frame selection: We conducted an ablation study on the interval \\( w \\) for setting keyframes, testing values of \\( w = 1 \\), \\( w = 5 \\), and \\( w = 10 \\), with results shown in Tab. 4. When \\( w = 1 \\), every frame becomes a keyframe, leading to excessive optimization that overfits Gaussians to training views, degrading test view quality and increasing training time and storage. Conversely, with \\( w = 10 \\), each keyframe drives the next 10 frames, but this distance weakens model performance, as it relies on assumptions about adjacent-frame" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16526" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.092, + 0.278, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.386, + 0.23, + 0.397 + ], + "angle": 0, + "content": "SaRo-GS" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.092, + 0.425, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.386, + 0.356, + 0.396 + ], + "angle": 0, + "content": "4DGS" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.092, + 0.571, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.386, + 0.537, + 0.396 + ], + "angle": 0, + "content": "3DGStream" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.092, + 0.715, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.386, + 0.643, + 0.396 + ], + "angle": 0, + "content": "IGS" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.092, + 0.861, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.386, + 0.792, + 0.396 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.397, + 0.669, + 0.41 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison from the N3DV dataset." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.418, + 0.287, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.535, + 0.276, + 0.544 + ], + "angle": 0, + "content": "(a) Ablation Study on Key-frame Refinement" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.419, + 0.482, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.535, + 0.451, + 0.544 + ], + "angle": 0, + "content": "(b) Per-frame reconstruction time" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.545, + 0.483, + 0.571 + ], + "angle": 0, + "content": "Figure 6. (a)Ablation Study on Key-frame Refinement. (b)Per-frame reconstruction time." + }, + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.589, + 0.447, + 0.603 + ], + "angle": 0, + "content": "Table 4. The impact of different keyframe intervals \\( w \\) ." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.604, + 0.452, + 0.673 + ], + "angle": 0, + "content": "
MethodPSNR(dB)↑Train(s)↓Storage(MB)↓
w=133.556.3836.0
w=533.622.677.90
w=1030.142.751.26
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.483, + 0.728 + ], + "angle": 0, + "content": "similarity. The setting \\( w = 5 \\) strikes the best balance across view synthesis quality, train time, and storage, and is thus our final choice." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.205, + 0.761 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.772, + 0.46, + 0.789 + ], + "angle": 0, + "content": "6.1. Independent per-frame reconstruction time" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We further evaluate the performance of IGS by analyzing the independent per-frame reconstruction time, as shown in Fig. 6 (b). The reconstruction time for each frame exhibits a periodic pattern: for candidate frames, it takes \\(0.8\\mathrm{s}\\), while for key frames, it takes \\(4\\mathrm{s}\\) and \\(7.5\\mathrm{s}\\) for the small and large versions, respectively, which are significantly smaller than the \\(16\\mathrm{s}\\) required by 3DGStream." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.419, + 0.633, + 0.433 + ], + "angle": 0, + "content": "6.2. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.607 + ], + "angle": 0, + "content": "IGS is the first to use a generalized method for streaming dynamic scene reconstruction, but it has limitations that can be addressed in future work. As shown in Fig. 3, our results exhibit jitter between adjacent frames, caused by the lack of temporal dependencies in the current framework. This makes the model more sensitive to noise. In contrast, 3DGStream assumes minimal motion between frames, yielding smoother results, but it fails in scenes with large motion (Fig. 4). To reduce jitter, we plan to incorporate temporal dependencies into IGS, modeling them as a time series for more robust performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.619, + 0.634, + 0.635 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose IGS as a novel streaming-based method for modeling dynamic scenes. With a generalized approach, IGS achieves frame-by-frame modeling in just over 2 seconds per frame, maintaining state-of-the-art rendering quality while keeping storage low. We introduce a generalized AGM-Net that lifts 2D multi-view motion features to 3D anchor points, using these anchors to drive Gaussian motion. This allows the model to infer the motion of Gaussians between adjacent frames in a single step. Additionally, we propose a Key-frame-guided Streaming strategy, where key frame sequences are selected and refined to mitigate error accumulation, further enhancing rendering quality. Extensive in-domain and cross-domain experiments demonstrate the strong generalization capabilities of our model, reducing significant streaming average cost while achieving state-of-the-art rendering quality, render speed, and storage efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16527" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "8. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.252 + ], + "angle": 0, + "content": "This work is financially supported for Outstanding Talents Training Fund in Shenzhen, this work is also financially supported by Shenzhen Science and Technology Program-Shenzhen Cultivation of Excellent Scientific and Technological Innovation Talents project(Grant No. RCJC20200714114435057), Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology(Grant No. 2024B1212010006), National Natural Science Foundation of China U21B2012." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.267, + 0.188, + 0.283 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.292, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855-5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.378, + 0.484, + 0.447 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.448, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[4] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[6] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.484, + 0.842 + ], + "angle": 0, + "content": "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[9] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In European Conference on Computer Vision (ECCV), 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.292, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[10] Yue Chen, Xingyu Chen, Xuan Wang, Qi Zhang, Yu Guo, Ying Shan, and Fei Wang. Local-to-global registration for bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8264-8273, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[11] Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. Hac: Hash-grid assisted context for 3d gaussian splatting compression. In European Conference on Computer Vision, pages 422-438. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.907, + 0.288 + ], + "angle": 0, + "content": "[12] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.906, + 0.343 + ], + "angle": 0, + "content": "[13] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with \\(15\\mathrm{x}\\) reduction and \\(200+\\) fps. arXiv preprint arXiv:2311.17245, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[14] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[15] Xin Fei, Wenzhao Zheng, Yueqi Duan, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Jiwen Lu. Pixelgaussian: Generalizable 3d gaussian reconstruction from arbitrary views. arXiv preprint arXiv:2410.18979, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.907, + 0.538 + ], + "angle": 0, + "content": "[16] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.907, + 0.621 + ], + "angle": 0, + "content": "[17] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbaek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.906, + 0.69 + ], + "angle": 0, + "content": "[18] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[19] Xiang Guo, Jiadai Sun, Yuchao Dai, Guanying Chen, Xiaoting Ye, Xiao Tan, Errui Ding, Yumeng Zhang, and Jingdong Wang. Forward flow for novel view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16022-16033, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.803 + ], + "angle": 0, + "content": "[20] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.906, + 0.873 + ], + "angle": 0, + "content": "[21] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19774–19783, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.903 + ], + "angle": 0, + "content": "[22] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16528" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "curate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[23] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. arXiv preprint arXiv:2312.14937, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[24] Haian Jin, Hanwen Jiang, Hao Tan, Kai Zhang, Sai Bi, Tianyuan Zhang, Fujun Luan, Noah Snavely, and Zexiang Xu. Lvsm: A large view synthesis model with minimal 3d inductive bias. arXiv preprint arXiv:2410.17242, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[25] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[27] Hao Li, Yuanyuan Gao, Dingwen Zhang, Chenming Wu, Yalun Dai, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Junwei Han. Gprt: Towards generalizable 3d gaussians without pose priors in real-time. arXiv preprint arXiv:2403.10147, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[28] Haolin Li, Jinyang Liu, Mario Sznaier, and Octavia Camps. 3d-hgs: 3d half-gaussian splatting. arXiv preprint arXiv:2406.02720, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[29] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[30] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.61, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[31] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. arXiv preprint arXiv:2312.16812, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[32] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741-5751, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[33] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. High-fidelity and real-time novel view synthesis for dynamic scenes. In SIGGRAPH Asia 2023 Conference Papers, pages 1-9, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[34] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. Advances in Neural Information Processing Systems, 35:36762-36775, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splattering reconstruction from multi-view stereo. In European Conference on Computer Vision, pages 37-53. Springer, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[37] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.299 + ], + "angle": 0, + "content": "[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.301, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[40] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[41] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[42] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.552, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[43] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[44] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[45] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[46] Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. arXiv preprint arXiv:2403.17898, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[47] Christian Schmidt, Jens Piekenbrinck, and Bastian Leibe. Look gauss, no pose: Novel view synthesis using gaussian splatting without accurate pose initialization. In IROS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16529" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.231 + ], + "angle": 0, + "content": "[49] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[50] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.483, + 0.356 + ], + "angle": 0, + "content": "[51] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20675-20685, 2024. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[52] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, João F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. arXiv preprint arXiv:2406.04343, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[53] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[54] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[55] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.483, + 0.678 + ], + "angle": 0, + "content": "[56] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[57] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[58] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[59] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 76-87, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[60] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4150-4159, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[61] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[62] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[63] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[64] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[65] Han Xu, Jiteng Yuan, and Jiayi Ma. Murf: Mutually reinforcing multi-modal image registration and fusion. IEEE transactions on pattern analysis and machine intelligence, 45(10):12148-12166, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[66] Jinbo Yan, Rui Peng, Luyang Tang, and Ronggang Wang. 4d gaussian splatting with scale-aware residual field and adaptive optimization for real-time rendering of temporally complex dynamic scenes. In Proceedings of the 32nd ACM International Conference on Multimedia, page 7871-7880, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[67] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[68] Runyi Yang, Zhenxin Zhu, Zhou Jiang, Baijun Ye, Xiaoxue Chen, Yifei Zhang, Yuantao Chen, Jian Zhao, and Hao Zhao. Spectrally pruned gaussian fields with neural compensation. arXiv preprint arXiv:2405.00676, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[69] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[70] Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. arXiv preprint arXiv:2402.15870, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[71] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and render-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16530" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "ing with 4d gaussian splatting. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[72] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[73] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[74] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[75] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[76] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[77] Chuanrui Zhang, Yingshuang Zou, Zhuoling Li, Minmin Yi, and Haoqian Wang. Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. arXiv preprint arXiv:2408.13770, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[78] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[79] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[80] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[81] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.728 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16531" + } + ] +] \ No newline at end of file diff --git a/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/c9f2cc0e-b9b5-4129-a1a3-c1fc202a4666_origin.pdf b/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/c9f2cc0e-b9b5-4129-a1a3-c1fc202a4666_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..946b99249565fcab2538c076f46a04ce124eca60 --- /dev/null +++ b/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/c9f2cc0e-b9b5-4129-a1a3-c1fc202a4666_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0443e200a12c0695e0ddb7e4666404546ff32374da11760819fff543c4d48a55 +size 2036293 diff --git a/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/full.md b/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d2865e3db2fd137d5cb2ba102cc22a4df68beabd --- /dev/null +++ b/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/full.md @@ -0,0 +1,382 @@ +# Instant Gaussian Stream: Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting + +Jinbo Yan $^{1}$ , Rui Peng $^{1,2}$ , Zhiyan Wang $^{1}$ , Luyang Tang $^{1,2}$ , Jiayu Yang $^{1,2}$ , Jie Liang $^{1}$ , Jiahao Wu $^{1}$ , Ronggang Wang $^{1,2*}$ + +1Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University 2Pengcheng Laboratory + +{yjb, ruipeng, zywang23, tly926, liangjie, wjh0616}@stu.pku.edu.cn jiayuyang@pku.edu.cn rgwang@pkusz.edu.cn + +![](images/299dc7953d51ad4b4f952107a0ca8d82f08eae69a4bd1323b6241a4f63801438.jpg) +(a)High Quality Result + +![](images/8c7b9e0e5d6ffd37a29677ac0bb5b95520aa6f1b8ad81ff68a0c1e9321f880d3.jpg) +Figure 1. Performance comparison with pervious SOTA[17, 29, 31, 51, 62]. Our method achieves a per-frame reconstruction time of 2.67s, delivering high-quality rendering results in a streaming fashion (a)(b), with a noticeable improvement in performance (c). * denotes a streamable method. + +![](images/5a4ad20a46aa7c0e85e0d8b10bad9a058b73eb4360228a253d01580d0d75be50.jpg) +(b)Realtime Rendering + +![](images/6972288d4c0311de77821dbf47436ee544ccae8d0189aa5bfb2b5a6799786b48.jpg) +(c)Comparison on Quality and Train Speed + +# Abstract + +Building Free-Viewpoint Videos in a streaming manner offers the advantage of rapid responsiveness compared to offline training methods, greatly enhancing user experience. However, current streaming approaches face challenges of high per-frame reconstruction time $(10s+)$ and error accumulation, limiting their broader application. In this paper, we propose Instant Gaussian Stream (IGS), a fast and generalizable streaming framework, to address these issues. First, we introduce a generalized Anchor-driven Gaussian Motion Network, which projects multi-view 2D motion features into 3D space, using anchor points to drive the motion of all Gaussians. This generalized Network generates the motion of Gaussians for each target frame in the time required for a single inference. Second, we propose a Key-frame-guided Streaming Strategy that refines each key frame, enabling accurate reconstruction of temporally complex scenes while mitigating error accumulation. We conducted extensive in-domain and cross-domain evalua + +tions, demonstrating that our approach can achieve streaming with a average per-frame reconstruction time of $2s+$ , alongside a enhancement in view synthesis quality. + +# 1. Introduction + +Reconstructing Free-Viewpoint Videos (FVV) from multiview images is a valuable area of research, with applications spanning immersive media such as VR, AR, and sports broadcasting. By enabling interactive, photorealistic visuals, FVVs from dynamic scenes hold the potential to become a next-generation visual medium, offering experiences that go beyond traditional video formats. To enhance user experience, streaming-based FVV construction—where dynamic scenes are reconstructed frame by frame—offers a low-delay response compared to traditional offline training approaches, making it better suited for real-time, interactive applications. + +With advancements in real-time rendering and high- + +quality view synthesis powered by 3D Gaussian Splatting (3DGS)[26], dynamic scene reconstruction has seen rapid progress. Some offline training methods[23, 31, 62, 66, 69, 71] achieve high-quality view synthesis but require collecting all frames before training can begin. This limitation makes them less suitable for scenarios that demand fast response times, such as live streaming and virtual meetings. To address these challenges, some methods[29, 51] adopt a streaming framework that reconstructs dynamic scenes frame by frame by modeling inter-frame differences. However, streaming-based dynamic scene reconstruction still faces significant challenges. First, current methods typically require per-frame optimization, resulting in high per-frame latencies $(10\mathrm{s}+)$ , which severely impact the real-time usability of these systems. Additionally, error accumulation across frames degrades the reconstruction quality of later frames, making it difficult for streaming methods to scale effectively to longer video sequences. + +To promote the streaming framework to be more practical, we introduce Instant Gaussian Stream (IGS), a streaming approach for dynamic scene reconstruction that achieves a per-frame reconstruction time of $2s+$ , mitigates error accumulation, and enhances view synthesis quality. First, to tackle the issue of high per-frame reconstruction time, we developed a generalized Anchor-driven Gaussian Motion Network (AGM-Net). This network utilizes a set of key points, called anchor points, to carry motion features that guide Gaussian transformations. This design allows the inference process to compute the motion of Gaussian primitives between frames in a single feedforward pass, eliminating the need for per-frame optimization. Second, to further improve view synthesis quality and minimize error accumulation, we propose a Key-frame-guided Streaming strategy. By establishing key-frame sequences and performing maxpoint-bounded refinement on key frames, our method mitigates the impact of error accumulation and enhances rendering quality in temporally complex scenes. + +We conducted extensive validation in both in-domain and cross-domain scenarios, and the experimental results demonstrate the strong generalization capability of our model, with significant improvements over current state-of-the-art methods in terms of per-frame reconstruction time and rendering quality. To the best of our knowledge, this is the first approach to use a generalized method for streaming reconstruction of dynamic scenes. Our contributions are summarized below. + +- We propose a generalized Anchor-driven Gaussian Motion Network that captures Gaussian motion between adjacent frames with a single inference, eliminating the need for frame-by-frame optimization. +- We designed a Key-frame-guided Streaming strategy to enhance our method's capability in handling temporally complex scenes, improving overall view synthesis qual + +ity within the streaming framework and mitigating error accumulation. + +- The evaluation results in both in-domain and cross-domain scenarios demonstrate the generalization capability of our method and its state-of-the-art performance. We achieve a 2.7 s per-frame reconstruction time for streaming, representing a significant improvement over previous methods. Additionally, we improve view synthesis quality, enabling real-time rendering at 204 FPS while maintaining comparable storage overhead. + +# 2. Related work + +# 2.1.3D Reconstruction and View Synthesis + +Novel view synthesis (NVS) has always been a hot topic in the field of computer vision. By using MLP to implicitly represent the scene, Neural Radiance Fields (NeRF) [38] achieves realistic rendering. Subsequent works have imporced NeRF to enhance rendering quality [1, 2, 60], reduce the number of training views [40, 58, 63, 67], lessen dependence on camera poses [4, 10, 32, 56], and improve both training and inference speeds [3, 8, 16, 20, 21, 39, 44, 45, 50]. 3D Gaussian Splitting (3DGS) [26] employs anisotropic Gaussian primitives to represent scenes and introduces rasterization-based splatting rendering algorithm, enhancing both speed and rendering quality. Some methods focus on various aspects of improving Gaussian field representations, including rendering quality[28, 37, 46, 70, 74, 79], enhancing geometric accuracy[22, 75, 76], and increasing compression efficiency, [11, 13, 37, 68], joint optimization of camera pose and gaussian fields [14, 18, 47], as well 3D generation [9, 54, 55, 81]. + +# 2.2. Generalizable 3D Reconstruction for Acceleration + +3DGS requires per-scene optimization to achieve realistic rendering results. To accelerate this time-consuming process, some works [24, 27, 52, 53, 77, 80], inspired by generalizable NeRF [7, 25, 61, 65, 73], have proposed to train generalizable Gaussian models on large-scale datasets to enable fast reconstruction. PixelSplat [6] utilizes an Transformer to encode features and decode them into Gaussian attributes. Other generalizable models [12, 15, 35, 78] utilize Transformers or Multi-View Stereo (MVS) [72] techniques to construct cost volumes followed by a decoder, achieving real-time rendering speeds and excellent generalizability. To the best of our knowledge, our work is the first to apply generalizable models to dynamic streaming scenes, utilizing their rapid inference capabilities to accelerate the processing of dynamic scenes reconstruction. + +# 2.3. Dynamic Scene Reconstruction and View Synthesis + +There have been numerous efforts to extend static scene reconstruction to dynamic scenes based on NeRF[5, 17, 19, 33, 34, 41, 43, 48, 57]. Since the advent of 3D Gaussian Splatting (3DGS)[26], researchers have explored incorporating its real-time rendering capabilities into dynamic scene reconstruction[23, 31, 62, 66, 69, 71]. + +However, these approaches rely on offline training with full video sequences, making them unsuitable for applications requiring real-time interaction and fast response. To address this issue, existing methods such as StreamRF[29], NeRFPlayer[49], ReRF[59], and 3DGStream[51] reformulate the dynamic modeling problem using a Streaming method. Notably, 3DGStream[51], based on Gaussian Splatting, optimizes a Neural Transformation Cache to model Gaussian movements between frames, further improving the performance. Although these methods achieve promising results, they still rely on per-frame optimization, resulting in significant delays (with current SOTA methods requiring over 10 seconds per frame[51]). Our approach offers a new perspective for streaming dynamic scene modeling: by training a generalized network, we eliminate the need for per-frame optimization, achieving low per-frame reconstruction time alongside high rendering quality. + +# 3. Method + +In this section, we begin with an overview of the pipeline in Sec.3.1. Then, in Sec. 3.3, we introduce the Anchordriven Gaussian Motion Network (AGM-Net), a generalized model that drives Gaussian motion from the previous frame using anchor points, which serve as key points in the 3D scene. Following this, we present our Key-frame-guided Streaming strategy in Sec.3.4. Finally, in Sec.3.5, we outline the loss function used in our training. + +# 3.1. Overview + +Our goal is to model dynamic scenes in a streaming manner with minimal per-frame reconstruction time. To achieve this, we adopt a generalized AGM-Net that extracts 3D motion features from the scene using anchor points and drives the motion of Gaussian primitives between frames in a single inference step. And we propose a key-frame-guided Streaming strategy to further improve view synthesis quality and handle temporally complex scenes while addressing error accumulation in streaming reconstruction. The overall pipeline is illustrated in Fig. 2. + +# 3.2. Preliminary + +Gaussian splatting[26] represents static scenes as a collection of anisotropic 3D Gaussians. The color of each pixel is obtained through point-based alpha blending rendering, + +enabling high-fidelity real-time novel view synthesis. + +Specifically, each Gaussian primitive $\mathcal{G}_i$ is parameterized by a center $\mu \in \mathbb{R}^3$ , 3D covariance matrix $\Sigma \in \mathbb{R}^{3\times 3}$ , opacity $\alpha \in \mathbb{R}$ , and color $c\in \mathbb{R}^{3(n + 1)^2}$ : + +$$ +\mathcal {G} (x) = e ^ {- \frac {1}{2} (x - \mu) ^ {T} \Sigma^ {- 1} (x - \mu)} \tag {1} +$$ + +During rendering, the 3D Gaussian is first projected onto 2D space. Subsequently, the Gaussians covering a pixel are sorted based on depth. The color of the pixel $\mathbf{c}$ is obtained using point-based alpha blending rendering: + +$$ +\mathbf {c} = \sum_ {i = 1} ^ {n} c _ {i} \alpha_ {i} ^ {\prime} \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {i} ^ {\prime}\right) \tag {2} +$$ + +Here, $\alpha^{\prime}$ represents the opacity after projection onto the 2D space. + +# 3.3. Anchor-driven Gaussian Motion Network + +Motion Feature Maps: Given multi-view images of current frames $\mathbf{I}' = (I_1', \dots, I_V')$ with camera parameters, We can first construct a multi-view image pair, which contains the current frame and the previous frame $\mathbf{I}$ from corresponding viewpoints. Then, we use a optical flow model to obtain the intermediate flow embeddings. Next, a modulation layer[9, 42] is applied to inject the viewpoint and depth information into the embeddings, ultimately resulting in 2D motion feature maps $\mathbf{F} \in \mathbb{R}^{V \times C \times H \times W}$ . + +Anchor Sampling: To deform the Gaussian primitives $\mathcal{G}$ from the previous frame, we need to compute the motion of each Gaussian. However, directly computing the motion for each Gaussian is computationally expensive and memory-intensive due to the large number of Gaussian points. To address this, we employ an anchor-point-based approach to represent the motion features of the entire scene in 3D space. The anchor-driven approach supports batch processing during training, reducing computational overhead while preserving the geometric information of the Gaussian primitives. Specifically, we use Farthest Point Sampling (FPS) to sample $M$ anchor points from the $N$ Gaussian primitives + +$$ +\mathcal {C} = \mathbf {F P S} \left(\left\{\mu_ {i} \right\} _ {i \in N}\right) \tag {3} +$$ + +where $\mathcal{C} \in \mathbb{R}^{M \times 3}$ represents the sampled anchor points with $M$ set to 8192 in our experiments, and $\mu_{i}$ denotes the position of $\mathcal{G}_{i}$ + +Projection-aware 3D Motion Feature Lift: We adopt a projection-aware approach to lift multiview 2D motion features into 3D space. Specifically, we project sampled anchor points onto each motion feature map based on the camera poses, obtaining high-resolution motion features: + +$$ +f _ {i} = \frac {1}{V} \sum_ {j \in V} \Psi \left(\Pi_ {j} \left(\mathcal {C} _ {i}\right), F _ {j}\right) \tag {4} +$$ + +![](images/4d097330a1a915bfa0a1a13b2ecb645a1908cd8f782eb7e1fb2d81b94390a9a0.jpg) +Figure 2. The overall pipeline of IGS. (a) Starting from the key frame and moving towards the target frame, we extract the 2D Motion Feature Map. (b) Then we sample M anchor points from the Gaussian primitives of the key frame, (c) and the anchor points are projected onto these feature maps to obtain 3D motion features through Projection-aware Motion Feature Lift. (d) Each Gaussian point interpolates its own motion feature from neighboring anchors and applies a weighted aggregation of features, which is then decoded into the motion of the Gaussian between the key frame and the target frame. (e) The entire streaming reconstruction process is guided by the Key-frame-guided Streaming strategy, where the key frame directly infers subsequent candidate frames until the next key-frame is reached, at which point max-point bounded refinement is applied to the key-frame. + +where $\Pi_j(\mathcal{C}_i)$ represents the projection of $\mathcal{C}_i$ onto the image plane of $F_{j}$ using the camera parameters of $F_{j}$ , and $\Psi$ denotes bilinear interpolation. By projection, each anchor point can accurately obtain its feature $f_{i}\in \mathbb{R}^{C}$ from the multi-view feature map, effectively lifting the 2D motion map into 3D space. + +We then use these features $\{f_i\}_{i\in M}$ , stored at each anchor point, as input to a Transformer block using self-attention to further capture motion information within the 3D scene. + +$$ +\left\{z _ {i}: z _ {i} \in \mathbb {R} ^ {C} \right\} _ {i \in M} = \operatorname {T r a n s f o r m e r} \left(\left\{f _ {i} \right\} _ {i \in M}\right) \tag {5} +$$ + +The output of the Transformer block $\{z_i\}_{i\in M}$ represents the final 3D motion features we obtain. Now, we can use these 3D motion features to represent the motion information of an anchor and its neighborhood, and drive the motion of the neighboring Gaussian points based on these motion features. + +Interpolate and Motion Decode: Using the 3D motion features stored at anchor points, we can assign each Gaussian point a motion feature by interpolating from its K nearest anchors in the neighborhood: + +$$ +z _ {i} = \frac {\sum_ {k \in \mathcal {N} (i)} e ^ {- d _ {k}} z _ {k}}{\sum_ {k \in \mathcal {N} (i)} e ^ {- d _ {k}}} \tag {6} +$$ + +where $\mathcal{N}(i)$ represents the set of neighboring anchor points of Gaussian point $\mathcal{G}_i$ , and $d_k$ represents the Euclidean distance from Gaussian point $G_i$ to anchor $\mathcal{C}_k$ . Then we can use a Linear head to decode the Motion feature to the movement of a Gaussian primitive: + +$$ +d \mu_ {i}, d r o t _ {i} = \mathbf {L i n e a r} \left(z _ {i}\right) \tag {7} +$$ + +here, we use the deformation of the Gaussian's position $d\mu_{i}$ and the deformation of the rotation $drot_{i}$ , to represent the movement of a Gaussian primitive. The new position and rotation of the Gaussian are as follows: + +$$ +\mu_ {i} ^ {\prime} = \mu_ {i} + d \mu_ {i}, \tag {8} +$$ + +$$ +\operatorname {r o t} _ {i} ^ {\prime} = \operatorname {n o r m} \left(\operatorname {r o t} _ {i}\right) \times \operatorname {n o r m} \left(\operatorname {d r o t} _ {i}\right). \tag {9} +$$ + +here ' refers to the new attributes. norm denotes to quaternion normalization and $\times$ represents quaternion multiplication, as used in previous work[51]. + +# 3.4. Key-frame-guided Streaming + +Using AGM-Net, we can transition Gaussian primitives from the previous frame to the current frame within a single forward inference pass. However, this process only adjusts the position and rotation of Gaussian primitives, making it + +effective for capturing rigid motion but inadequate for accurately representing non-rigid motion. Furthermore, the number of Gaussian points remains constant, limiting its capacity to model temporally dynamic scenes where objects may appear or disappear. These limitations result in challenges in capturing scene dynamics and can lead to error accumulation across frames. + +To better model object changes and reduce error accumulation, we propose a Key-frame-guided Streaming strategy that uses key frames as the initial state for deforming Gaussians in subsequent frames. We also introduce a Max points bounded refinement strategy, enabling efficient key frame reconstruction without redundant points and preventing point count growth across frames. This approach helps avoid overfitting in sparse-viewpoint scenes by effectively managing point density. + +Key-frame-guided strategy: Starting from frame 0, we designate a key frame every $w$ frames, forming a key-frame sequence $\{K_0, K_w, \dots, K_{nw}\}$ . The remaining frames serve as candidate frames. During streaming reconstruction, for example, beginning with a key frame $K_{iw}$ , we deform the Gaussians forward across successive candidate frames using AGM-Net until reaching the next key frame $K_{(i+1)w}$ . At this point, we refine the deformed Gaussians of key frame $K_{(i+1)w}$ . Then, we continue deforming from key frame $K_{(i+1)w}$ to process subsequent frames. + +This key-frame-guided strategy offers several advantages. First, when AGM-Net is applied to candidate frames, it is always start from the most recent key frame, preventing error propagation across candidate frames between key frames and eliminating cumulative error. Second, candidate frames do not require optimization-based refinement, as their Gaussians are generated through a single model inference with AGM-Net, ensuring low per-frame reconstruction time. Additionally, we can batch process up to $w$ frames following each key frame, which further accelerates our pipeline. + +Max points bounded Key-frame Refinement: During the refinement of each key frame, we optimize all parameters of the Gaussians and support cloning, splitting, and filtering, which is same to 3DGS[26]. This approach allows us to handle object deformations as well as the appearance and disappearance of objects in temporally complex scenes, effectively preventing error accumulation from key frame to subsequent frames. However, this optimization strategy can lead to a gradual increase in Gaussian primitives at each key frame, which not only raises computational complexity and storage requirements but also risks overfitting in sparse viewpoints, particularly in dynamic scenes where viewpoints are generally limited. + +To address this, we adopt a Max Points Bounded Refine method. When densifying Gaussian points, we control the number of Gaussians allowed to densify by adjusting each + +point's gradient, ensuring that the total number of points does not exceed a predefined maximum. + +# 3.5. Loss Function + +Our training process consists of two parts: offline training the generalized AGM-Net and performing online training for the key frames. The generalized AGM-Net only needs to be trained once, and it can generalize to multiple scenes. We train the AGM-Net across scenes using gradient descent, relying solely on a view synthesis loss between our predicted views and the ground truth views, which includes an $\mathcal{L}_1$ term and an $\mathcal{L}_{D - SSIM}$ term and can be formulated as: + +$$ +\mathcal {L} = (1 - \lambda) \mathcal {L} _ {1} + \lambda \mathcal {L} _ {D - S S I M} \tag {10} +$$ + +When performing online training on the Gaussians in key frames, we use the same loss function as in Eq.10. However, this time, we optimize the attributes of the Gaussian primitives rather than the parameters of the neural network. + +# 4. Implementation details + +In this Section, we first introduce the datasets we used, along with the partitioning and preprocessing of training data, in Sec. 4.1. Next, we provide a detailed explanation of the configuration of the AGM network and the training hyperparameters in Sec. 4.2. Finally, we describe the detailed setup for streaming in Sec. 4.3. + +# 4.1. Datasets + +The Neural 3D Video Datasets (N3DV)[30] includes 6 dynamic scenes recorded using a multi-view setup featuring 21 cameras, with a resolution of $2704 \times 2028$ . Each multiview video comprises 300 frames. + +Meeting Room Datasets[29] includes 3 dynamic scenes recorded with 13 cameras at a resolution of $1280 \times 720$ . Each multi-view video also contains 300 frames. + +Dataset Preparation: We split four sequences from the N3DV dataset into the training set, with the remaining two sequences, {cut roasted beef, sear steak}, used as the test set. For the training set, we constructed 3D Gaussians for all frames in the four training sequences, totaling 1200 frames, which required 192 GPU hours. For each frame's 3D Gaussian, we performed motions forward and backward for five frames, creating 12,000 pairs for training. For testing, we selected one viewpoint for evaluation for both datasets, consistent with previous methods. + +# 4.2.AGM Network + +We use GM-Flow[64] to extract optical flow embeddings and add a Swin-Transformer[36] block for fine-tuning while keeping the other parameters of GM-Flow fixed. Our AGM model accepts an arbitrary number of input views. To balance computational complexity and performance, we use + +$V = 4$ views, each producing a motion map with $C = 128$ channels and a resolution of $128 \times 128$ . We sample $M = 8192$ anchor points from Gaussian Points, which sufficiently captures dynamic details. The Transformer block in 3D motion feature lift module comprises 4 layers, yielding a 3D motion feature with $C = 128$ channels. For rendering, we adopt a variant of Gaussian Splitting Rasterization from Rade-GS[76] to obtain more accurate depth maps and geometry. + +During training, we randomly select 4 views as input and use 8 views for supervision. Training is conducted on four A100 GPUs with 40GB of memory each, running for a total of 15 epochs with a batch size of 16. The parameter $\gamma$ in Eq. 10 is set to 0.2. We use the Adam optimizer with a weight decay of 0.05, and $\beta$ values of (0.9, 0.95). The learning rate is set to $4\times 10^{-4}$ for training on the N3DV dataset. + +# 4.3. Streaming Inference + +We set $w = 5$ to construct key frame sequences, resulting in 60 keyframes from a 300-frame video, and conduct an ablation study to assess the impact of different $w$ values in Sec. 5.3. We designed two versions for keyframe optimization: a smaller version IGS-s (Ours-s) with 50 iterations refinement for Key frames, providing lower per-frame latency, and a larger version IGS- $l$ (Ours-1) with 100 iterations, which achieves higher reconstruction quality. In both versions, densification and pruning are performed every 20 iterations. For the test sequences, we construct the Gaussians for the 0th frame using the compression method provided by Lightgaussian[13], which reduces storage usage and mitigates overfitting due to sparse viewpoints. We employ 6,000 iterations for training the first frame of the N3DV dataset, compressing the number of Gaussians at 5,000 iterations. For the Meeting Room dataset, we train the Gaussians of the first frame using 15,000 iterations, compressing the number of Gaussians at 7,000 iterations. For more details, please refer to the Supp.. + +# 5. Experiments + +# 5.1. Metrics and Baselines + +Baselines: We compare our approach to current state-of-the-art methods for dynamic scene reconstruction, covering both offline and online training methods. Offline methods[17, 31, 62, 66, 71] rely on a set of Gaussian primitives or Hex-planes to represent entire dynamic scenes. Online training methods[29, 51] employ per-frame optimization to support streaming reconstruction. Specifically, 3DGStream[51] models the movement of Gaussian points across frames by optimizing a Neural Transform Cache, creating a 3DGS-based pipeline for free-viewpoint video streaming that enables high-quality, real-time reconstruction of dynamic scenes. + +Table 1. Comparison on the N3DV dataset, with results measured at a resolution of $1352 \times 1014$ . $\dagger$ indicates that the evaluation was performed using the official code in the same experimental environment as ours, including the same initial point cloud. Highlights denote the best and second best results. + +
MethodPSNR† (dB)Train ↓ (s)Render† (FPS)Storage↓ (MB)
Offline training
Kplanes[17]32.17480.151.0
Realtime-4DGS[71]33.68-114-
4DGS[62]32.707.8300.3
Spacetime-GS[31]33.71481400.7
Saro-GS[66]33.90-401.0
Online training
StreamRF[29]32.09158.331.4
3DGStream[51]33.11122157.8
3DGStream[51]†32.7516.932047.69
Ours-s33.892.672047.90
Ours-l34.153.352047.90
+ +Metrics: Following prior work, we evaluate and report PSNR, Storage usage, Train time, and Render Speed to compare with previous methods. All metrics are averaged over the full 300-frame sequence, including frame 0. + +# 5.2. Comparisons + +In-domain evaluation: We present our in-domain evaluation on two test sequences from the N3DV dataset, with results shown in Tab. 1. For a fair comparison of performance, we tested 3DGStream using the same Gaussians from the 0th frame and applied the same variant of Gaussian Splatting Rasterization as used in our approach (denoted with $\dagger$ in the table). Compared to 3DGStream and StreamRF, our method achieves a 6x reduction in train time, with an average delay of 2.67 seconds per frame, while maintaining comparable rendering speed and storage usage. Our approach also achieves enhanced rendering quality. Compared to offline training methods, our approach provides low-delay streaming capabilities while achieving state-of-the-art rendering quality and reducing training time. A qualitative comparison of rendering quality can be seen in Fig. 5. It is evident that our method outperforms others in rendering details, such as the transition between the knife and fork, and in modeling complex dynamic scenes, like the moving hand and the shifting reflection on the wall. + +We also conducted a PSNR trend comparison with 3DGStream to verify the effectiveness of our method in mitigating error accumulation. The comparison results and the smoothed trends are shown in Fig. 2. As seen, our rendering quality does not degrade with increasing frame number, while 3DGStream suffers from error accumulation, with a noticeable decline in quality as the frame number increases. This confirms the effectiveness of our approach in address + +ing error accumulation. However, it is also apparent that our method exhibits more fluctuation in per-frame PSNR. This is because 3DGStream assumes small inter-frame motion, leading to smaller adjustments and smoother differences between frames. + +![](images/018f82a665c02b8e9cf1200d87fdc023eeea23c51f98134a2b56129176979dcc.jpg) +Figure 3. The PSNR trend comparison on the sear steak. + +![](images/e92e7b3a22c8b3f6cd4837bbf0636d1203292a8b67010ca65794d9654d967505.jpg) +Figure 4. Qualitative comparison from the Meeting Room dataset. Cross-domain evaluation: We performed a cross-domain evaluation on the Meeting Room Dataset using a model trained on N3DV. The evaluation results are presented in Tab. 2. Our method outperforms 3DGStream in rendering quality, train time, and storage efficiency, achieving streaming with just 2.77s of per-frame reconstruction time, a significant improvement over 3DGStream. This demonstrates the effectiveness and generalizability of our approach, as it enables efficient dynamic scene modeling with streaming capabilities in new environments, without requiring per-frame optimization. A qualitative comparison of rendering quality can be seen in Fig. 4. Compared to 3DGStream, which produces artifacts near moving objects, our method yields more accurate motion during large displacements, resulting in improved performance in temporally complex scenes. + +# 5.3. Ablation Study + +The use of the pretrained optical flow model: We used a pretrained optical flow model to extract flow embeddings + +Table 2. Comparison on the Meeting Room dataset. $\dagger$ indicates that the evaluation was performed using the official code in the same experimental environment as ours, including the same initial point cloud. + +
MethodPSNR†(dB)Train ↓(s)Render↑(FPS)Storage↓(MB)
3DGStream[51]†28.3611.512527.59
Ours-s29.242.772521.26
Ours-130.133.202521.26
+ +Table 3. Ablation Study Results + +
MethodPSNR†(dB)Train↓(s)Storage↓(MB)
No-pretrained optical flow model31.072.657.90
No-projection-aware feature lift32.952.387.90
No-points bounded refinement33.233.02110.26
Ours-s(full)33.622.677.90
+ +from image pairs, which are then lifted into 3D space. To validate its effectiveness, we replaced the pretrained model with a 4-layer UNet without pretrained parameters and trained it jointly with the overall model. The results in Tab. 3 highlight the benefit of using the 2D prior. + +Projection-aware 3D Motion Feature Lift: We use a projection-based approach to lift multi-view 2D motion feature maps into 3D space, accurately linking 3D anchor points to 2D features. To evaluate its effectiveness, we replaced this method with a Transformer-based approach using cross-attention between image features and anchor points, enhanced with positional embeddings through a 4-layer Transformer block. As shown in Tab. 3, Projection-aware Feature Lift is crucial for IGS performance, with only a slight increase in training time. + +Key-frame guided Streaming: We employ a key-frame-guided strategy to address error accumulation in streaming and to enhance reconstruction quality. Keyframes are selected and refined through Max-points-bounded Refine. Without this refinement, AGM-Net would rely solely on Gaussians propagated from the last keyframe, resulting in accumulated errors that significantly impact performance, as shown in Fig. 6 (a). We also evaluate the effect of max-points bounding during refinement, as shown in Tab. 3. Without point limits, storage requirements increase substantially, and overfitting causes a decline in view quality. + +Key-frame selection: We conducted an ablation study on the interval $w$ for setting keyframes, testing values of $w = 1$ , $w = 5$ , and $w = 10$ , with results shown in Tab. 4. When $w = 1$ , every frame becomes a keyframe, leading to excessive optimization that overfits Gaussians to training views, degrading test view quality and increasing training time and storage. Conversely, with $w = 10$ , each keyframe drives the next 10 frames, but this distance weakens model performance, as it relies on assumptions about adjacent-frame + +![](images/7bffae3ad7ed61269bbb878e43661d24901c5be646a5c65dedc77bcc27675dec.jpg) +SaRo-GS + +![](images/a2bf7d0683a3665f5dabf7338b770aa73a1c163ca8c17811a8f9cab80d867413.jpg) +4DGS + +![](images/2e73f22f4004fbdd5f540f8e860143bdd06c71824407239939089a199cb29bd3.jpg) +3DGStream + +![](images/bcca0dd79eb747ab28c2aea8646ac6954ca9d5dea01cdfa5f34dff3fe35d6a36.jpg) +IGS + +![](images/9c3521f2e7e087e78bf59eae62ec64021b6bd52d2b37045789320d45a7fbb557.jpg) +GT + +![](images/72bc753b4ee5e0508e6a0643a4a359055bc2e65c9c962d279c1e6dcc69d40f2a.jpg) +(a) Ablation Study on Key-frame Refinement + +![](images/014072f71d96268000634e44821552aacfa85e0c308b1760a85161fb961a991e.jpg) +Figure 5. Qualitative comparison from the N3DV dataset. +(b) Per-frame reconstruction time +Figure 6. (a)Ablation Study on Key-frame Refinement. (b)Per-frame reconstruction time. + +Table 4. The impact of different keyframe intervals $w$ . + +
MethodPSNR(dB)↑Train(s)↓Storage(MB)↓
w=133.556.3836.0
w=533.622.677.90
w=1030.142.751.26
+ +similarity. The setting $w = 5$ strikes the best balance across view synthesis quality, train time, and storage, and is thus our final choice. + +# 6. Discussion + +# 6.1. Independent per-frame reconstruction time + +We further evaluate the performance of IGS by analyzing the independent per-frame reconstruction time, as shown in Fig. 6 (b). The reconstruction time for each frame exhibits a periodic pattern: for candidate frames, it takes $0.8\mathrm{s}$ , while for key frames, it takes $4\mathrm{s}$ and $7.5\mathrm{s}$ for the small and large versions, respectively, which are significantly smaller than the $16\mathrm{s}$ required by 3DGStream. + +# 6.2. Limitation + +IGS is the first to use a generalized method for streaming dynamic scene reconstruction, but it has limitations that can be addressed in future work. As shown in Fig. 3, our results exhibit jitter between adjacent frames, caused by the lack of temporal dependencies in the current framework. This makes the model more sensitive to noise. In contrast, 3DGStream assumes minimal motion between frames, yielding smoother results, but it fails in scenes with large motion (Fig. 4). To reduce jitter, we plan to incorporate temporal dependencies into IGS, modeling them as a time series for more robust performance. + +# 7. Conclusion + +In this paper, we propose IGS as a novel streaming-based method for modeling dynamic scenes. With a generalized approach, IGS achieves frame-by-frame modeling in just over 2 seconds per frame, maintaining state-of-the-art rendering quality while keeping storage low. We introduce a generalized AGM-Net that lifts 2D multi-view motion features to 3D anchor points, using these anchors to drive Gaussian motion. This allows the model to infer the motion of Gaussians between adjacent frames in a single step. Additionally, we propose a Key-frame-guided Streaming strategy, where key frame sequences are selected and refined to mitigate error accumulation, further enhancing rendering quality. Extensive in-domain and cross-domain experiments demonstrate the strong generalization capabilities of our model, reducing significant streaming average cost while achieving state-of-the-art rendering quality, render speed, and storage efficiency. + +# 8. Acknowledgments + +This work is financially supported for Outstanding Talents Training Fund in Shenzhen, this work is also financially supported by Shenzhen Science and Technology Program-Shenzhen Cultivation of Excellent Scientific and Technological Innovation Talents project(Grant No. RCJC20200714114435057), Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology(Grant No. 2024B1212010006), National Natural Science Foundation of China U21B2012. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855-5864, 2021. 2 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2 +[4] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 2 +[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 3 +[6] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2 +[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2 +[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2 +[9] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In European Conference on Computer Vision (ECCV), 2024. 2, 3 + +[10] Yue Chen, Xingyu Chen, Xuan Wang, Qi Zhang, Yu Guo, Ying Shan, and Fei Wang. Local-to-global registration for bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8264-8273, 2023. 2 +[11] Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. Hac: Hash-grid assisted context for 3d gaussian splatting compression. In European Conference on Computer Vision, pages 422-438. Springer, 2025. 2 +[12] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 2 +[13] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with $15\mathrm{x}$ reduction and $200+$ fps. arXiv preprint arXiv:2311.17245, 2023. 2, 6 +[14] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2, 2024. 2 +[15] Xin Fei, Wenzhao Zheng, Yueqi Duan, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Jiwen Lu. Pixelgaussian: Generalizable 3d gaussian reconstruction from arbitrary views. arXiv preprint arXiv:2410.18979, 2024. 2 +[16] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2 +[17] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbaek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 1, 3, 6 +[18] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 2 +[19] Xiang Guo, Jiadai Sun, Yuchao Dai, Guanying Chen, Xiaoting Ye, Xiao Tan, Errui Ding, Yumeng Zhang, and Jingdong Wang. Forward flow for novel view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16022-16033, 2023. 3 +[20] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. ICCV, 2021. 2 +[21] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19774–19783, 2023. 2 +[22] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac + +curate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2 +[23] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. arXiv preprint arXiv:2312.14937, 2023. 2, 3 +[24] Haian Jin, Hanwen Jiang, Hao Tan, Kai Zhang, Sai Bi, Tianyuan Zhang, Fujun Luan, Noah Snavely, and Zexiang Xu. Lvsm: A large view synthesis model with minimal 3d inductive bias. arXiv preprint arXiv:2410.17242, 2024. 2 +[25] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. 2 +[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 5 +[27] Hao Li, Yuanyuan Gao, Dingwen Zhang, Chenming Wu, Yalun Dai, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Junwei Han. Gprt: Towards generalizable 3d gaussians without pose priors in real-time. arXiv preprint arXiv:2403.10147, 2024. 2 +[28] Haolin Li, Jinyang Liu, Mario Sznaier, and Octavia Camps. 3d-hgs: 3d half-gaussian splatting. arXiv preprint arXiv:2406.02720, 2024. 2 +[29] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 1, 2, 3, 5, 6 +[30] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 5 +[31] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. arXiv preprint arXiv:2312.16812, 2023. 1, 2, 3, 6 +[32] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741-5751, 2021. 2 +[33] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. High-fidelity and real-time novel view synthesis for dynamic scenes. In SIGGRAPH Asia 2023 Conference Papers, pages 1-9, 2023. 3 +[34] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. Advances in Neural Information Processing Systems, 35:36762-36775, 2022. 3 +[35] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splattering reconstruction from multi-view stereo. In European Conference on Computer Vision, pages 37-53. Springer, 2025. 2 + +[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 5 +[37] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2 +[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2 +[40] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 2 +[41] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 3 +[42] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022.3 +[43] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 3 +[44] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 2 +[45] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2 +[46] Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. arXiv preprint arXiv:2403.17898, 2024. 2 +[47] Christian Schmidt, Jens Piekenbrinck, and Bastian Leibe. Look gauss, no pose: Novel view synthesis using gaussian splatting without accurate pose initialization. In IROS, 2024. 2 +[48] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural + +4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 3 +[49] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 3 +[50] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2 +[51] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20675-20685, 2024. 1, 2, 3, 4, 6, 7 +[52] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, João F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. arXiv preprint arXiv:2406.04343, 2024. 2 +[53] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2 +[54] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2 +[55] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 2 +[56] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2 +[57] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis, 2021. 3 +[58] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 2 +[59] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 76-87, 2023. 3 +[60] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera + +trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4150-4159, 2023. 2 +[61] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2021. 2 +[62] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 1, 2, 3, 6 +[63] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2 +[64] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 5 +[65] Han Xu, Jiteng Yuan, and Jiayi Ma. Murf: Mutually reinforcing multi-modal image registration and fusion. IEEE transactions on pattern analysis and machine intelligence, 45(10):12148-12166, 2023. 2 +[66] Jinbo Yan, Rui Peng, Luyang Tang, and Ronggang Wang. 4d gaussian splatting with scale-aware residual field and adaptive optimization for real-time rendering of temporally complex dynamic scenes. In Proceedings of the 32nd ACM International Conference on Multimedia, page 7871-7880, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3, 6 +[67] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. 2 +[68] Runyi Yang, Zhenxin Zhu, Zhou Jiang, Baijun Ye, Xiaoxue Chen, Yifei Zhang, Yuantao Chen, Jian Zhao, and Hao Zhao. Spectrally pruned gaussian fields with neural compensation. arXiv preprint arXiv:2405.00676, 2024. 2 +[69] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3 +[70] Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. arXiv preprint arXiv:2402.15870, 2024. 2 +[71] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and render- + +ing with 4d gaussian splatting. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 6 +[72] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2 +[73] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 2 +[74] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2 +[75] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2 +[76] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2, 6 +[77] Chuanrui Zhang, Yingshuang Zou, Zhuoling Li, Minmin Yi, and Haoqian Wang. Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. arXiv preprint arXiv:2408.13770, 2024. 2 +[78] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2 +[79] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. In ECCV, 2024. 2 +[80] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 2 +[81] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2 \ No newline at end of file diff --git a/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/images.zip b/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f8e7a3ea3511fe35c2c6b4e6631eda58f567656b --- /dev/null +++ b/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:595b33be46d4192579c4a9604e14b77e70741e858ac9b3ae9e19ed132e6e118d +size 514011 diff --git a/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/layout.json b/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..66bc9e5c11e27f4be70be4787ce33caf63bdbc93 --- /dev/null +++ b/2025/Instant Gaussian Stream_ Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting/layout.json @@ -0,0 +1,10169 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 102, + 549, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 102, + 549, + 140 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 549, + 140 + ], + "type": "text", + "content": "Instant Gaussian Stream: Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "text", + "content": "Jinbo Yan" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "text", + "content": ", Rui Peng" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "text", + "content": ", Zhiyan Wang" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "text", + "content": ", Luyang Tang" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "text", + "content": ", Jiayu Yang" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "text", + "content": ", Jie Liang" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "text", + "content": ", Jiahao Wu" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "text", + "content": ", Ronggang Wang" + }, + { + "bbox": [ + 134, + 159, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 190, + 535, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 190, + 535, + 232 + ], + "spans": [ + { + "bbox": [ + 75, + 190, + 535, + 232 + ], + "type": "text", + "content": "1Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University 2Pengcheng Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 233, + 483, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 233, + 483, + 258 + ], + "spans": [ + { + "bbox": [ + 127, + 233, + 483, + 258 + ], + "type": "text", + "content": "{yjb, ruipeng, zywang23, tly926, liangjie, wjh0616}@stu.pku.edu.cn jiayuyang@pku.edu.cn rgwang@pkusz.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 288, + 126, + 413 + ], + "blocks": [ + { + "bbox": [ + 58, + 288, + 126, + 413 + ], + "lines": [ + { + "bbox": [ + 58, + 288, + 126, + 413 + ], + "spans": [ + { + "bbox": [ + 58, + 288, + 126, + 413 + ], + "type": "image", + "image_path": "299dc7953d51ad4b4f952107a0ca8d82f08eae69a4bd1323b6241a4f63801438.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 420, + 161, + 430 + ], + "lines": [ + { + "bbox": [ + 83, + 420, + 161, + 430 + ], + "spans": [ + { + "bbox": [ + 83, + 420, + 161, + 430 + ], + "type": "text", + "content": "(a)High Quality Result" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 129, + 288, + 196, + 413 + ], + "blocks": [ + { + "bbox": [ + 129, + 288, + 196, + 413 + ], + "lines": [ + { + "bbox": [ + 129, + 288, + 196, + 413 + ], + "spans": [ + { + "bbox": [ + 129, + 288, + 196, + 413 + ], + "type": "image", + "image_path": "8c7b9e0e5d6ffd37a29677ac0bb5b95520aa6f1b8ad81ff68a0c1e9321f880d3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 439, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 54, + 439, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 439, + 555, + 472 + ], + "type": "text", + "content": "Figure 1. Performance comparison with pervious SOTA[17, 29, 31, 51, 62]. Our method achieves a per-frame reconstruction time of 2.67s, delivering high-quality rendering results in a streaming fashion (a)(b), with a noticeable improvement in performance (c). * denotes a streamable method." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 207, + 289, + 372, + 413 + ], + "blocks": [ + { + "bbox": [ + 207, + 289, + 372, + 413 + ], + "lines": [ + { + "bbox": [ + 207, + 289, + 372, + 413 + ], + "spans": [ + { + "bbox": [ + 207, + 289, + 372, + 413 + ], + "type": "image", + "image_path": "5a4ad20a46aa7c0e85e0d8b10bad9a058b73eb4360228a253d01580d0d75be50.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 420, + 334, + 430 + ], + "lines": [ + { + "bbox": [ + 256, + 420, + 334, + 430 + ], + "spans": [ + { + "bbox": [ + 256, + 420, + 334, + 430 + ], + "type": "text", + "content": "(b)Realtime Rendering" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 380, + 286, + 555, + 418 + ], + "blocks": [ + { + "bbox": [ + 380, + 286, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 380, + 286, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 380, + 286, + 555, + 418 + ], + "type": "image", + "image_path": "6972288d4c0311de77821dbf47436ee544ccae8d0189aa5bfb2b5a6799786b48.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 419, + 553, + 430 + ], + "lines": [ + { + "bbox": [ + 408, + 419, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 408, + 419, + 553, + 430 + ], + "type": "text", + "content": "(c)Comparison on Quality and Train Speed" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 472, + 200, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 472, + 200, + 485 + ], + "spans": [ + { + "bbox": [ + 151, + 472, + 200, + 485 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 498, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 297, + 713 + ], + "type": "text", + "content": "Building Free-Viewpoint Videos in a streaming manner offers the advantage of rapid responsiveness compared to offline training methods, greatly enhancing user experience. However, current streaming approaches face challenges of high per-frame reconstruction time " + }, + { + "bbox": [ + 54, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "(10s+)" + }, + { + "bbox": [ + 54, + 498, + 297, + 713 + ], + "type": "text", + "content": " and error accumulation, limiting their broader application. In this paper, we propose Instant Gaussian Stream (IGS), a fast and generalizable streaming framework, to address these issues. First, we introduce a generalized Anchor-driven Gaussian Motion Network, which projects multi-view 2D motion features into 3D space, using anchor points to drive the motion of all Gaussians. This generalized Network generates the motion of Gaussians for each target frame in the time required for a single inference. Second, we propose a Key-frame-guided Streaming Strategy that refines each key frame, enabling accurate reconstruction of temporally complex scenes while mitigating error accumulation. We conducted extensive in-domain and cross-domain evalua" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 474, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 510 + ], + "type": "text", + "content": "tions, demonstrating that our approach can achieve streaming with a average per-frame reconstruction time of " + }, + { + "bbox": [ + 313, + 474, + 555, + 510 + ], + "type": "inline_equation", + "content": "2s+" + }, + { + "bbox": [ + 313, + 474, + 555, + 510 + ], + "type": "text", + "content": ", alongside a enhancement in view synthesis quality." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 536, + 395, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 536, + 395, + 548 + ], + "spans": [ + { + "bbox": [ + 314, + 536, + 395, + 548 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 556, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 556, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 312, + 556, + 555, + 700 + ], + "type": "text", + "content": "Reconstructing Free-Viewpoint Videos (FVV) from multiview images is a valuable area of research, with applications spanning immersive media such as VR, AR, and sports broadcasting. By enabling interactive, photorealistic visuals, FVVs from dynamic scenes hold the potential to become a next-generation visual medium, offering experiences that go beyond traditional video formats. To enhance user experience, streaming-based FVV construction—where dynamic scenes are reconstructed frame by frame—offers a low-delay response compared to traditional offline training approaches, making it better suited for real-time, interactive applications." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "With advancements in real-time rendering and high-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16520" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": "quality view synthesis powered by 3D Gaussian Splatting (3DGS)[26], dynamic scene reconstruction has seen rapid progress. Some offline training methods[23, 31, 62, 66, 69, 71] achieve high-quality view synthesis but require collecting all frames before training can begin. This limitation makes them less suitable for scenarios that demand fast response times, such as live streaming and virtual meetings. To address these challenges, some methods[29, 51] adopt a streaming framework that reconstructs dynamic scenes frame by frame by modeling inter-frame differences. However, streaming-based dynamic scene reconstruction still faces significant challenges. First, current methods typically require per-frame optimization, resulting in high per-frame latencies " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "(10\\mathrm{s}+)" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": ", which severely impact the real-time usability of these systems. Additionally, error accumulation across frames degrades the reconstruction quality of later frames, making it difficult for streaming methods to scale effectively to longer video sequences." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 289, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 295, + 518 + ], + "type": "text", + "content": "To promote the streaming framework to be more practical, we introduce Instant Gaussian Stream (IGS), a streaming approach for dynamic scene reconstruction that achieves a per-frame reconstruction time of " + }, + { + "bbox": [ + 56, + 289, + 295, + 518 + ], + "type": "inline_equation", + "content": "2s+" + }, + { + "bbox": [ + 56, + 289, + 295, + 518 + ], + "type": "text", + "content": ", mitigates error accumulation, and enhances view synthesis quality. First, to tackle the issue of high per-frame reconstruction time, we developed a generalized Anchor-driven Gaussian Motion Network (AGM-Net). This network utilizes a set of key points, called anchor points, to carry motion features that guide Gaussian transformations. This design allows the inference process to compute the motion of Gaussian primitives between frames in a single feedforward pass, eliminating the need for per-frame optimization. Second, to further improve view synthesis quality and minimize error accumulation, we propose a Key-frame-guided Streaming strategy. By establishing key-frame sequences and performing maxpoint-bounded refinement on key frames, our method mitigates the impact of error accumulation and enhances rendering quality in temporally complex scenes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 520, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 628 + ], + "type": "text", + "content": "We conducted extensive validation in both in-domain and cross-domain scenarios, and the experimental results demonstrate the strong generalization capability of our model, with significant improvements over current state-of-the-art methods in terms of per-frame reconstruction time and rendering quality. To the best of our knowledge, this is the first approach to use a generalized method for streaming reconstruction of dynamic scenes. Our contributions are summarized below." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "content": "- We propose a generalized Anchor-driven Gaussian Motion Network that captures Gaussian motion between adjacent frames with a single inference, eliminating the need for frame-by-frame optimization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 678, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 678, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 678, + 296, + 714 + ], + "type": "text", + "content": "- We designed a Key-frame-guided Streaming strategy to enhance our method's capability in handling temporally complex scenes, improving overall view synthesis qual" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "content": "ity within the streaming framework and mitigating error accumulation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 96, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 555, + 193 + ], + "type": "text", + "content": "- The evaluation results in both in-domain and cross-domain scenarios demonstrate the generalization capability of our method and its state-of-the-art performance. We achieve a 2.7 s per-frame reconstruction time for streaming, representing a significant improvement over previous methods. Additionally, we improve view synthesis quality, enabling real-time rendering at 204 FPS while maintaining comparable storage overhead." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 217, + 399, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 217, + 399, + 229 + ], + "spans": [ + { + "bbox": [ + 314, + 217, + 399, + 229 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 241, + 517, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 517, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 517, + 254 + ], + "type": "text", + "content": "2.1.3D Reconstruction and View Synthesis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 262, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 479 + ], + "type": "text", + "content": "Novel view synthesis (NVS) has always been a hot topic in the field of computer vision. By using MLP to implicitly represent the scene, Neural Radiance Fields (NeRF) [38] achieves realistic rendering. Subsequent works have imporced NeRF to enhance rendering quality [1, 2, 60], reduce the number of training views [40, 58, 63, 67], lessen dependence on camera poses [4, 10, 32, 56], and improve both training and inference speeds [3, 8, 16, 20, 21, 39, 44, 45, 50]. 3D Gaussian Splitting (3DGS) [26] employs anisotropic Gaussian primitives to represent scenes and introduces rasterization-based splatting rendering algorithm, enhancing both speed and rendering quality. Some methods focus on various aspects of improving Gaussian field representations, including rendering quality[28, 37, 46, 70, 74, 79], enhancing geometric accuracy[22, 75, 76], and increasing compression efficiency, [11, 13, 37, 68], joint optimization of camera pose and gaussian fields [14, 18, 47], as well 3D generation [9, 54, 55, 81]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 500, + 553, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 553, + 523 + ], + "type": "text", + "content": "2.2. Generalizable 3D Reconstruction for Acceleration" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "3DGS requires per-scene optimization to achieve realistic rendering results. To accelerate this time-consuming process, some works [24, 27, 52, 53, 77, 80], inspired by generalizable NeRF [7, 25, 61, 65, 73], have proposed to train generalizable Gaussian models on large-scale datasets to enable fast reconstruction. PixelSplat [6] utilizes an Transformer to encode features and decode them into Gaussian attributes. Other generalizable models [12, 15, 35, 78] utilize Transformers or Multi-View Stereo (MVS) [72] techniques to construct cost volumes followed by a decoder, achieving real-time rendering speeds and excellent generalizability. To the best of our knowledge, our work is the first to apply generalizable models to dynamic streaming scenes, utilizing their rapid inference capabilities to accelerate the processing of dynamic scenes reconstruction." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16521" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "2.3. Dynamic Scene Reconstruction and View Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 296, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 296, + 173 + ], + "type": "text", + "content": "There have been numerous efforts to extend static scene reconstruction to dynamic scenes based on NeRF[5, 17, 19, 33, 34, 41, 43, 48, 57]. Since the advent of 3D Gaussian Splatting (3DGS)[26], researchers have explored incorporating its real-time rendering capabilities into dynamic scene reconstruction[23, 31, 62, 66, 69, 71]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 174, + 296, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 296, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 296, + 377 + ], + "type": "text", + "content": "However, these approaches rely on offline training with full video sequences, making them unsuitable for applications requiring real-time interaction and fast response. To address this issue, existing methods such as StreamRF[29], NeRFPlayer[49], ReRF[59], and 3DGStream[51] reformulate the dynamic modeling problem using a Streaming method. Notably, 3DGStream[51], based on Gaussian Splatting, optimizes a Neural Transformation Cache to model Gaussian movements between frames, further improving the performance. Although these methods achieve promising results, they still rely on per-frame optimization, resulting in significant delays (with current SOTA methods requiring over 10 seconds per frame[51]). Our approach offers a new perspective for streaming dynamic scene modeling: by training a generalized network, we eliminate the need for per-frame optimization, achieving low per-frame reconstruction time alongside high rendering quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 111, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 111, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 111, + 401 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 409, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 506 + ], + "type": "text", + "content": "In this section, we begin with an overview of the pipeline in Sec.3.1. Then, in Sec. 3.3, we introduce the Anchordriven Gaussian Motion Network (AGM-Net), a generalized model that drives Gaussian motion from the previous frame using anchor points, which serve as key points in the 3D scene. Following this, we present our Key-frame-guided Streaming strategy in Sec.3.4. Finally, in Sec.3.5, we outline the loss function used in our training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 514, + 123, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 123, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 123, + 525 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 651 + ], + "type": "text", + "content": "Our goal is to model dynamic scenes in a streaming manner with minimal per-frame reconstruction time. To achieve this, we adopt a generalized AGM-Net that extracts 3D motion features from the scene using anchor points and drives the motion of Gaussian primitives between frames in a single inference step. And we propose a key-frame-guided Streaming strategy to further improve view synthesis quality and handle temporally complex scenes while addressing error accumulation in streaming reconstruction. The overall pipeline is illustrated in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 134, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 134, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 134, + 672 + ], + "type": "text", + "content": "3.2. Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Gaussian splatting[26] represents static scenes as a collection of anisotropic 3D Gaussians. The color of each pixel is obtained through point-based alpha blending rendering," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 529, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 529, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 529, + 84 + ], + "type": "text", + "content": "enabling high-fidelity real-time novel view synthesis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 84, + 553, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 121 + ], + "type": "text", + "content": "Specifically, each Gaussian primitive " + }, + { + "bbox": [ + 313, + 84, + 553, + 121 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 313, + 84, + 553, + 121 + ], + "type": "text", + "content": " is parameterized by a center " + }, + { + "bbox": [ + 313, + 84, + 553, + 121 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 84, + 553, + 121 + ], + "type": "text", + "content": ", 3D covariance matrix " + }, + { + "bbox": [ + 313, + 84, + 553, + 121 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 313, + 84, + 553, + 121 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 313, + 84, + 553, + 121 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 84, + 553, + 121 + ], + "type": "text", + "content": ", and color " + }, + { + "bbox": [ + 313, + 84, + 553, + 121 + ], + "type": "inline_equation", + "content": "c\\in \\mathbb{R}^{3(n + 1)^2}" + }, + { + "bbox": [ + 313, + 84, + 553, + 121 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 375, + 129, + 553, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 129, + 553, + 146 + ], + "spans": [ + { + "bbox": [ + 375, + 129, + 553, + 146 + ], + "type": "interline_equation", + "content": "\\mathcal {G} (x) = e ^ {- \\frac {1}{2} (x - \\mu) ^ {T} \\Sigma^ {- 1} (x - \\mu)} \\tag {1}", + "image_path": "a9c84696d04119e085ed3443f9d4c0ed98fe294c198f422780d623638fc223aa.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 155, + 555, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 155, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 555, + 204 + ], + "type": "text", + "content": "During rendering, the 3D Gaussian is first projected onto 2D space. Subsequently, the Gaussians covering a pixel are sorted based on depth. The color of the pixel " + }, + { + "bbox": [ + 313, + 155, + 555, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 155, + 555, + 204 + ], + "type": "text", + "content": " is obtained using point-based alpha blending rendering:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 381, + 212, + 553, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 212, + 553, + 245 + ], + "spans": [ + { + "bbox": [ + 381, + 212, + 553, + 245 + ], + "type": "interline_equation", + "content": "\\mathbf {c} = \\sum_ {i = 1} ^ {n} c _ {i} \\alpha_ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {i} ^ {\\prime}\\right) \\tag {2}", + "image_path": "0f9aa08738838379c5929e6fbee30b16b39829eec38820ebd36a61e22a5d0f7e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 255, + 553, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 278 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 255, + 553, + 278 + ], + "type": "inline_equation", + "content": "\\alpha^{\\prime}" + }, + { + "bbox": [ + 313, + 255, + 553, + 278 + ], + "type": "text", + "content": " represents the opacity after projection onto the 2D space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 286, + 531, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 531, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 531, + 297 + ], + "type": "text", + "content": "3.3. Anchor-driven Gaussian Motion Network" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "type": "text", + "content": "Motion Feature Maps: Given multi-view images of current frames " + }, + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{I}' = (I_1', \\dots, I_V')" + }, + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "type": "text", + "content": " with camera parameters, We can first construct a multi-view image pair, which contains the current frame and the previous frame " + }, + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "type": "text", + "content": " from corresponding viewpoints. Then, we use a optical flow model to obtain the intermediate flow embeddings. Next, a modulation layer[9, 42] is applied to inject the viewpoint and depth information into the embeddings, ultimately resulting in 2D motion feature maps " + }, + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{V \\times C \\times H \\times W}" + }, + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "type": "text", + "content": "Anchor Sampling: To deform the Gaussian primitives " + }, + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "type": "text", + "content": " from the previous frame, we need to compute the motion of each Gaussian. However, directly computing the motion for each Gaussian is computationally expensive and memory-intensive due to the large number of Gaussian points. To address this, we employ an anchor-point-based approach to represent the motion features of the entire scene in 3D space. The anchor-driven approach supports batch processing during training, reducing computational overhead while preserving the geometric information of the Gaussian primitives. Specifically, we use Farthest Point Sampling (FPS) to sample " + }, + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "type": "text", + "content": " anchor points from the " + }, + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "type": "text", + "content": " Gaussian primitives" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 390, + 566, + 553, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 566, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 390, + 566, + 553, + 580 + ], + "type": "interline_equation", + "content": "\\mathcal {C} = \\mathbf {F P S} \\left(\\left\\{\\mu_ {i} \\right\\} _ {i \\in N}\\right) \\tag {3}", + "image_path": "e16cf8fb92f5a3587ded01e9aa28a74d3e23f1f7c13f8fab035e4aab8a16b93c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\in \\mathbb{R}^{M \\times 3}" + }, + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "text", + "content": " represents the sampled anchor points with " + }, + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "text", + "content": " set to 8192 in our experiments, and " + }, + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "text", + "content": " denotes the position of " + }, + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{i}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 618, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 677 + ], + "type": "text", + "content": "Projection-aware 3D Motion Feature Lift: We adopt a projection-aware approach to lift multiview 2D motion features into 3D space. Specifically, we project sampled anchor points onto each motion feature map based on the camera poses, obtaining high-resolution motion features:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 377, + 686, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 686, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 377, + 686, + 553, + 715 + ], + "type": "interline_equation", + "content": "f _ {i} = \\frac {1}{V} \\sum_ {j \\in V} \\Psi \\left(\\Pi_ {j} \\left(\\mathcal {C} _ {i}\\right), F _ {j}\\right) \\tag {4}", + "image_path": "b5f519fc7c43ee0c5047ef5788a1211acaa126ef7daa6ded9382adb1fb1f3e5e.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16522" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 556, + 309 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 556, + 309 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 556, + 309 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 556, + 309 + ], + "type": "image", + "image_path": "4d097330a1a915bfa0a1a13b2ecb645a1908cd8f782eb7e1fb2d81b94390a9a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 318, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 54, + 318, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 555, + 396 + ], + "type": "text", + "content": "Figure 2. The overall pipeline of IGS. (a) Starting from the key frame and moving towards the target frame, we extract the 2D Motion Feature Map. (b) Then we sample M anchor points from the Gaussian primitives of the key frame, (c) and the anchor points are projected onto these feature maps to obtain 3D motion features through Projection-aware Motion Feature Lift. (d) Each Gaussian point interpolates its own motion feature from neighboring anchors and applies a weighted aggregation of features, which is then decoded into the motion of the Gaussian between the key frame and the target frame. (e) The entire streaming reconstruction process is guided by the Key-frame-guided Streaming strategy, where the key frame directly infers subsequent candidate frames until the next key-frame is reached, at which point max-point bounded refinement is applied to the key-frame." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "inline_equation", + "content": "\\Pi_j(\\mathcal{C}_i)" + }, + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "text", + "content": " represents the projection of " + }, + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "text", + "content": " onto the image plane of " + }, + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "inline_equation", + "content": "F_{j}" + }, + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "text", + "content": " using the camera parameters of " + }, + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "inline_equation", + "content": "F_{j}" + }, + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "text", + "content": " denotes bilinear interpolation. By projection, each anchor point can accurately obtain its feature " + }, + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "inline_equation", + "content": "f_{i}\\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 54, + 402, + 294, + 474 + ], + "type": "text", + "content": " from the multi-view feature map, effectively lifting the 2D motion map into 3D space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 474, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 521 + ], + "type": "text", + "content": "We then use these features " + }, + { + "bbox": [ + 55, + 474, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\{f_i\\}_{i\\in M}" + }, + { + "bbox": [ + 55, + 474, + 295, + 521 + ], + "type": "text", + "content": ", stored at each anchor point, as input to a Transformer block using self-attention to further capture motion information within the 3D scene." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 531, + 295, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 531, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 70, + 531, + 295, + 547 + ], + "type": "interline_equation", + "content": "\\left\\{z _ {i}: z _ {i} \\in \\mathbb {R} ^ {C} \\right\\} _ {i \\in M} = \\operatorname {T r a n s f o r m e r} \\left(\\left\\{f _ {i} \\right\\} _ {i \\in M}\\right) \\tag {5}", + "image_path": "1e5ec525e39228276c239158b27f677b1a8a8ad80c53daa64a39fddf63c97b85.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 555, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 626 + ], + "type": "text", + "content": "The output of the Transformer block " + }, + { + "bbox": [ + 55, + 555, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\{z_i\\}_{i\\in M}" + }, + { + "bbox": [ + 55, + 555, + 295, + 626 + ], + "type": "text", + "content": " represents the final 3D motion features we obtain. Now, we can use these 3D motion features to represent the motion information of an anchor and its neighborhood, and drive the motion of the neighboring Gaussian points based on these motion features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 628, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 676 + ], + "type": "text", + "content": "Interpolate and Motion Decode: Using the 3D motion features stored at anchor points, we can assign each Gaussian point a motion feature by interpolating from its K nearest anchors in the neighborhood:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 685, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 685, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 127, + 685, + 295, + 717 + ], + "type": "interline_equation", + "content": "z _ {i} = \\frac {\\sum_ {k \\in \\mathcal {N} (i)} e ^ {- d _ {k}} z _ {k}}{\\sum_ {k \\in \\mathcal {N} (i)} e ^ {- d _ {k}}} \\tag {6}", + "image_path": "1241e8c05c9f2d143e2dcb7511be6297cf10b8b9676d1c89d6b03fd54eb33c4b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 402, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 402, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(i)" + }, + { + "bbox": [ + 313, + 402, + 555, + 462 + ], + "type": "text", + "content": " represents the set of neighboring anchor points of Gaussian point " + }, + { + "bbox": [ + 313, + 402, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 313, + 402, + 555, + 462 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 402, + 555, + 462 + ], + "type": "inline_equation", + "content": "d_k" + }, + { + "bbox": [ + 313, + 402, + 555, + 462 + ], + "type": "text", + "content": " represents the Euclidean distance from Gaussian point " + }, + { + "bbox": [ + 313, + 402, + 555, + 462 + ], + "type": "inline_equation", + "content": "G_i" + }, + { + "bbox": [ + 313, + 402, + 555, + 462 + ], + "type": "text", + "content": " to anchor " + }, + { + "bbox": [ + 313, + 402, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_k" + }, + { + "bbox": [ + 313, + 402, + 555, + 462 + ], + "type": "text", + "content": ". Then we can use a Linear head to decode the Motion feature to the movement of a Gaussian primitive:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 381, + 472, + 555, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 472, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 381, + 472, + 555, + 485 + ], + "type": "interline_equation", + "content": "d \\mu_ {i}, d r o t _ {i} = \\mathbf {L i n e a r} \\left(z _ {i}\\right) \\tag {7}", + "image_path": "f00b64a0464d3a949808971b9cab0c24b0e1827a1e6998aba49df298379a5eb3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 494, + 555, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 555, + 541 + ], + "type": "text", + "content": "here, we use the deformation of the Gaussian's position " + }, + { + "bbox": [ + 313, + 494, + 555, + 541 + ], + "type": "inline_equation", + "content": "d\\mu_{i}" + }, + { + "bbox": [ + 313, + 494, + 555, + 541 + ], + "type": "text", + "content": " and the deformation of the rotation " + }, + { + "bbox": [ + 313, + 494, + 555, + 541 + ], + "type": "inline_equation", + "content": "drot_{i}" + }, + { + "bbox": [ + 313, + 494, + 555, + 541 + ], + "type": "text", + "content": ", to represent the movement of a Gaussian primitive. The new position and rotation of the Gaussian are as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 402, + 550, + 555, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 550, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 402, + 550, + 555, + 566 + ], + "type": "interline_equation", + "content": "\\mu_ {i} ^ {\\prime} = \\mu_ {i} + d \\mu_ {i}, \\tag {8}", + "image_path": "39f8a2ef878daa1b03f6e2a5796f1da4d1858b6394919590fa94afecbb31e092.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 585, + 553, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 585, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 359, + 585, + 553, + 599 + ], + "type": "interline_equation", + "content": "\\operatorname {r o t} _ {i} ^ {\\prime} = \\operatorname {n o r m} \\left(\\operatorname {r o t} _ {i}\\right) \\times \\operatorname {n o r m} \\left(\\operatorname {d r o t} _ {i}\\right). \\tag {9}", + "image_path": "e9f4d8f2e62ffc43c0565245ae44e5d6ee3c3a6f47ef7a6c7711cf43602055a0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "text", + "content": "here ' refers to the new attributes. norm denotes to quaternion normalization and " + }, + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "text", + "content": " represents quaternion multiplication, as used in previous work[51]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 647, + 474, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 474, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 474, + 661 + ], + "type": "text", + "content": "3.4. Key-frame-guided Streaming" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Using AGM-Net, we can transition Gaussian primitives from the previous frame to the current frame within a single forward inference pass. However, this process only adjusts the position and rotation of Gaussian primitives, making it" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16523" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "effective for capturing rigid motion but inadequate for accurately representing non-rigid motion. Furthermore, the number of Gaussian points remains constant, limiting its capacity to model temporally dynamic scenes where objects may appear or disappear. These limitations result in challenges in capturing scene dynamics and can lead to error accumulation across frames." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "type": "text", + "content": "To better model object changes and reduce error accumulation, we propose a Key-frame-guided Streaming strategy that uses key frames as the initial state for deforming Gaussians in subsequent frames. We also introduce a Max points bounded refinement strategy, enabling efficient key frame reconstruction without redundant points and preventing point count growth across frames. This approach helps avoid overfitting in sparse-viewpoint scenes by effectively managing point density." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "text", + "content": "Key-frame-guided strategy: Starting from frame 0, we designate a key frame every " + }, + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "text", + "content": " frames, forming a key-frame sequence " + }, + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "inline_equation", + "content": "\\{K_0, K_w, \\dots, K_{nw}\\}" + }, + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "text", + "content": ". The remaining frames serve as candidate frames. During streaming reconstruction, for example, beginning with a key frame " + }, + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "inline_equation", + "content": "K_{iw}" + }, + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "text", + "content": ", we deform the Gaussians forward across successive candidate frames using AGM-Net until reaching the next key frame " + }, + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "inline_equation", + "content": "K_{(i+1)w}" + }, + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "text", + "content": ". At this point, we refine the deformed Gaussians of key frame " + }, + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "inline_equation", + "content": "K_{(i+1)w}" + }, + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "text", + "content": ". Then, we continue deforming from key frame " + }, + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "inline_equation", + "content": "K_{(i+1)w}" + }, + { + "bbox": [ + 55, + 266, + 295, + 387 + ], + "type": "text", + "content": " to process subsequent frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 518 + ], + "type": "text", + "content": "This key-frame-guided strategy offers several advantages. First, when AGM-Net is applied to candidate frames, it is always start from the most recent key frame, preventing error propagation across candidate frames between key frames and eliminating cumulative error. Second, candidate frames do not require optimization-based refinement, as their Gaussians are generated through a single model inference with AGM-Net, ensuring low per-frame reconstruction time. Additionally, we can batch process up to " + }, + { + "bbox": [ + 55, + 388, + 295, + 518 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 388, + 295, + 518 + ], + "type": "text", + "content": " frames following each key frame, which further accelerates our pipeline." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 675 + ], + "type": "text", + "content": "Max points bounded Key-frame Refinement: During the refinement of each key frame, we optimize all parameters of the Gaussians and support cloning, splitting, and filtering, which is same to 3DGS[26]. This approach allows us to handle object deformations as well as the appearance and disappearance of objects in temporally complex scenes, effectively preventing error accumulation from key frame to subsequent frames. However, this optimization strategy can lead to a gradual increase in Gaussian primitives at each key frame, which not only raises computational complexity and storage requirements but also risks overfitting in sparse viewpoints, particularly in dynamic scenes where viewpoints are generally limited." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "To address this, we adopt a Max Points Bounded Refine method. When densifying Gaussian points, we control the number of Gaussians allowed to densify by adjusting each" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "point's gradient, ensuring that the total number of points does not exceed a predefined maximum." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 104, + 404, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 104, + 404, + 115 + ], + "spans": [ + { + "bbox": [ + 314, + 104, + 404, + 115 + ], + "type": "text", + "content": "3.5. Loss Function" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "text", + "content": "Our training process consists of two parts: offline training the generalized AGM-Net and performing online training for the key frames. The generalized AGM-Net only needs to be trained once, and it can generalize to multiple scenes. We train the AGM-Net across scenes using gradient descent, relying solely on a view synthesis loss between our predicted views and the ground truth views, which includes an " + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "text", + "content": " term and an " + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{D - SSIM}" + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "text", + "content": " term and can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 370, + 227, + 553, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 227, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 370, + 227, + 553, + 239 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {D - S S I M} \\tag {10}", + "image_path": "dcbc2171c9f2010e2e5767b8b7b904884e375050ff70ded9efe0cbf708af82f5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 249, + 553, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 553, + 297 + ], + "type": "text", + "content": "When performing online training on the Gaussians in key frames, we use the same loss function as in Eq.10. However, this time, we optimize the attributes of the Gaussian primitives rather than the parameters of the neural network." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 308, + 448, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 448, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 448, + 321 + ], + "type": "text", + "content": "4. Implementation details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 328, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 553, + 399 + ], + "type": "text", + "content": "In this Section, we first introduce the datasets we used, along with the partitioning and preprocessing of training data, in Sec. 4.1. Next, we provide a detailed explanation of the configuration of the AGM network and the training hyperparameters in Sec. 4.2. Finally, we describe the detailed setup for streaming in Sec. 4.3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 407, + 376, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 376, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 376, + 418 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 425, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 553, + 472 + ], + "type": "text", + "content": "The Neural 3D Video Datasets (N3DV)[30] includes 6 dynamic scenes recorded using a multi-view setup featuring 21 cameras, with a resolution of " + }, + { + "bbox": [ + 313, + 425, + 553, + 472 + ], + "type": "inline_equation", + "content": "2704 \\times 2028" + }, + { + "bbox": [ + 313, + 425, + 553, + 472 + ], + "type": "text", + "content": ". Each multiview video comprises 300 frames." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 473, + 553, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 553, + 508 + ], + "type": "text", + "content": "Meeting Room Datasets[29] includes 3 dynamic scenes recorded with 13 cameras at a resolution of " + }, + { + "bbox": [ + 313, + 473, + 553, + 508 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 313, + 473, + 553, + 508 + ], + "type": "text", + "content": ". Each multi-view video also contains 300 frames." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 509, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 628 + ], + "type": "text", + "content": "Dataset Preparation: We split four sequences from the N3DV dataset into the training set, with the remaining two sequences, {cut roasted beef, sear steak}, used as the test set. For the training set, we constructed 3D Gaussians for all frames in the four training sequences, totaling 1200 frames, which required 192 GPU hours. For each frame's 3D Gaussian, we performed motions forward and backward for five frames, creating 12,000 pairs for training. For testing, we selected one viewpoint for evaluation for both datasets, consistent with previous methods." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 636, + 406, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 406, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 406, + 647 + ], + "type": "text", + "content": "4.2.AGM Network" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": "We use GM-Flow[64] to extract optical flow embeddings and add a Swin-Transformer[36] block for fine-tuning while keeping the other parameters of GM-Flow fixed. Our AGM model accepts an arbitrary number of input views. To balance computational complexity and performance, we use" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16524" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "V = 4" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " views, each producing a motion map with " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "C = 128" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " channels and a resolution of " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": ". We sample " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "M = 8192" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " anchor points from Gaussian Points, which sufficiently captures dynamic details. The Transformer block in 3D motion feature lift module comprises 4 layers, yielding a 3D motion feature with " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "C = 128" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " channels. For rendering, we adopt a variant of Gaussian Splitting Rasterization from Rade-GS[76] to obtain more accurate depth maps and geometry." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 294, + 266 + ], + "type": "text", + "content": "During training, we randomly select 4 views as input and use 8 views for supervision. Training is conducted on four A100 GPUs with 40GB of memory each, running for a total of 15 epochs with a batch size of 16. The parameter " + }, + { + "bbox": [ + 55, + 180, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 180, + 294, + 266 + ], + "type": "text", + "content": " in Eq. 10 is set to 0.2. We use the Adam optimizer with a weight decay of 0.05, and " + }, + { + "bbox": [ + 55, + 180, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 180, + 294, + 266 + ], + "type": "text", + "content": " values of (0.9, 0.95). The learning rate is set to " + }, + { + "bbox": [ + 55, + 180, + 294, + 266 + ], + "type": "inline_equation", + "content": "4\\times 10^{-4}" + }, + { + "bbox": [ + 55, + 180, + 294, + 266 + ], + "type": "text", + "content": " for training on the N3DV dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 274, + 174, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 174, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 174, + 286 + ], + "type": "text", + "content": "4.3. Streaming Inference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 292, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 295, + 519 + ], + "type": "text", + "content": "We set " + }, + { + "bbox": [ + 55, + 292, + 295, + 519 + ], + "type": "inline_equation", + "content": "w = 5" + }, + { + "bbox": [ + 55, + 292, + 295, + 519 + ], + "type": "text", + "content": " to construct key frame sequences, resulting in 60 keyframes from a 300-frame video, and conduct an ablation study to assess the impact of different " + }, + { + "bbox": [ + 55, + 292, + 295, + 519 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 292, + 295, + 519 + ], + "type": "text", + "content": " values in Sec. 5.3. We designed two versions for keyframe optimization: a smaller version IGS-s (Ours-s) with 50 iterations refinement for Key frames, providing lower per-frame latency, and a larger version IGS- " + }, + { + "bbox": [ + 55, + 292, + 295, + 519 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 292, + 295, + 519 + ], + "type": "text", + "content": " (Ours-1) with 100 iterations, which achieves higher reconstruction quality. In both versions, densification and pruning are performed every 20 iterations. For the test sequences, we construct the Gaussians for the 0th frame using the compression method provided by Lightgaussian[13], which reduces storage usage and mitigates overfitting due to sparse viewpoints. We employ 6,000 iterations for training the first frame of the N3DV dataset, compressing the number of Gaussians at 5,000 iterations. For the Meeting Room dataset, we train the Gaussians of the first frame using 15,000 iterations, compressing the number of Gaussians at 7,000 iterations. For more details, please refer to the Supp.." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 532, + 137, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 137, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 137, + 545 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 180, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 180, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 180, + 563 + ], + "type": "text", + "content": "5.1. Metrics and Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Baselines: We compare our approach to current state-of-the-art methods for dynamic scene reconstruction, covering both offline and online training methods. Offline methods[17, 31, 62, 66, 71] rely on a set of Gaussian primitives or Hex-planes to represent entire dynamic scenes. Online training methods[29, 51] employ per-frame optimization to support streaming reconstruction. Specifically, 3DGStream[51] models the movement of Gaussian points across frames by optimizing a Neural Transform Cache, creating a 3DGS-based pipeline for free-viewpoint video streaming that enables high-quality, real-time reconstruction of dynamic scenes." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 134, + 545, + 297 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 126 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 126 + ], + "type": "text", + "content": "Table 1. Comparison on the N3DV dataset, with results measured at a resolution of " + }, + { + "bbox": [ + 313, + 71, + 553, + 126 + ], + "type": "inline_equation", + "content": "1352 \\times 1014" + }, + { + "bbox": [ + 313, + 71, + 553, + 126 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 71, + 553, + 126 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 71, + 553, + 126 + ], + "type": "text", + "content": " indicates that the evaluation was performed using the official code in the same experimental environment as ours, including the same initial point cloud. Highlights denote the best and second best results." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 134, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 319, + 134, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 319, + 134, + 545, + 297 + ], + "type": "table", + "html": "
MethodPSNR† (dB)Train ↓ (s)Render† (FPS)Storage↓ (MB)
Offline training
Kplanes[17]32.17480.151.0
Realtime-4DGS[71]33.68-114-
4DGS[62]32.707.8300.3
Spacetime-GS[31]33.71481400.7
Saro-GS[66]33.90-401.0
Online training
StreamRF[29]32.09158.331.4
3DGStream[51]33.11122157.8
3DGStream[51]†32.7516.932047.69
Ours-s33.892.672047.90
Ours-l34.153.352047.90
", + "image_path": "53af74a9e5ce1d914f2908b1a9c11c5357c190fb4b2be56a897c2a63140525b6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 302, + 553, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 553, + 351 + ], + "type": "text", + "content": "Metrics: Following prior work, we evaluate and report PSNR, Storage usage, Train time, and Render Speed to compare with previous methods. All metrics are averaged over the full 300-frame sequence, including frame 0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 359, + 400, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 400, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 400, + 372 + ], + "type": "text", + "content": "5.2. Comparisons" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 377, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 553, + 616 + ], + "type": "text", + "content": "In-domain evaluation: We present our in-domain evaluation on two test sequences from the N3DV dataset, with results shown in Tab. 1. For a fair comparison of performance, we tested 3DGStream using the same Gaussians from the 0th frame and applied the same variant of Gaussian Splatting Rasterization as used in our approach (denoted with " + }, + { + "bbox": [ + 313, + 377, + 553, + 616 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 377, + 553, + 616 + ], + "type": "text", + "content": " in the table). Compared to 3DGStream and StreamRF, our method achieves a 6x reduction in train time, with an average delay of 2.67 seconds per frame, while maintaining comparable rendering speed and storage usage. Our approach also achieves enhanced rendering quality. Compared to offline training methods, our approach provides low-delay streaming capabilities while achieving state-of-the-art rendering quality and reducing training time. A qualitative comparison of rendering quality can be seen in Fig. 5. It is evident that our method outperforms others in rendering details, such as the transition between the knife and fork, and in modeling complex dynamic scenes, like the moving hand and the shifting reflection on the wall." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "text", + "content": "We also conducted a PSNR trend comparison with 3DGStream to verify the effectiveness of our method in mitigating error accumulation. The comparison results and the smoothed trends are shown in Fig. 2. As seen, our rendering quality does not degrade with increasing frame number, while 3DGStream suffers from error accumulation, with a noticeable decline in quality as the frame number increases. This confirms the effectiveness of our approach in address" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16525" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "text", + "content": "ing error accumulation. However, it is also apparent that our method exhibits more fluctuation in per-frame PSNR. This is because 3DGStream assumes small inter-frame motion, leading to smaller adjustments and smoother differences between frames." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 81, + 140, + 266, + 280 + ], + "blocks": [ + { + "bbox": [ + 81, + 140, + 266, + 280 + ], + "lines": [ + { + "bbox": [ + 81, + 140, + 266, + 280 + ], + "spans": [ + { + "bbox": [ + 81, + 140, + 266, + 280 + ], + "type": "image", + "image_path": "018f82a665c02b8e9cf1200d87fdc023eeea23c51f98134a2b56129176979dcc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 280, + 279, + 289 + ], + "lines": [ + { + "bbox": [ + 71, + 280, + 279, + 289 + ], + "spans": [ + { + "bbox": [ + 71, + 280, + 279, + 289 + ], + "type": "text", + "content": "Figure 3. The PSNR trend comparison on the sear steak." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 69, + 311, + 282, + 456 + ], + "blocks": [ + { + "bbox": [ + 69, + 311, + 282, + 456 + ], + "lines": [ + { + "bbox": [ + 69, + 311, + 282, + 456 + ], + "spans": [ + { + "bbox": [ + 69, + 311, + 282, + 456 + ], + "type": "image", + "image_path": "e92e7b3a22c8b3f6cd4837bbf0636d1203292a8b67010ca65794d9654d967505.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 464, + 295, + 666 + ], + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 666 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison from the Meeting Room dataset. Cross-domain evaluation: We performed a cross-domain evaluation on the Meeting Room Dataset using a model trained on N3DV. The evaluation results are presented in Tab. 2. Our method outperforms 3DGStream in rendering quality, train time, and storage efficiency, achieving streaming with just 2.77s of per-frame reconstruction time, a significant improvement over 3DGStream. This demonstrates the effectiveness and generalizability of our approach, as it enables efficient dynamic scene modeling with streaming capabilities in new environments, without requiring per-frame optimization. A qualitative comparison of rendering quality can be seen in Fig. 4. Compared to 3DGStream, which produces artifacts near moving objects, our method yields more accurate motion during large displacements, resulting in improved performance in temporally complex scenes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 672, + 149, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 149, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 149, + 685 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "The use of the pretrained optical flow model: We used a pretrained optical flow model to extract flow embeddings" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 320, + 114, + 547, + 178 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 554, + 114 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 554, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 554, + 114 + ], + "type": "text", + "content": "Table 2. Comparison on the Meeting Room dataset. " + }, + { + "bbox": [ + 313, + 71, + 554, + 114 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 71, + 554, + 114 + ], + "type": "text", + "content": " indicates that the evaluation was performed using the official code in the same experimental environment as ours, including the same initial point cloud." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 114, + 547, + 178 + ], + "lines": [ + { + "bbox": [ + 320, + 114, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 320, + 114, + 547, + 178 + ], + "type": "table", + "html": "
MethodPSNR†(dB)Train ↓(s)Render↑(FPS)Storage↓(MB)
3DGStream[51]†28.3611.512527.59
Ours-s29.242.772521.26
Ours-130.133.202521.26
", + "image_path": "90180acbbbddd8f046f89e8e52cb2d58e455172d4a1eda33c7476df457f37384.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 319, + 202, + 545, + 272 + ], + "blocks": [ + { + "bbox": [ + 375, + 191, + 493, + 202 + ], + "lines": [ + { + "bbox": [ + 375, + 191, + 493, + 202 + ], + "spans": [ + { + "bbox": [ + 375, + 191, + 493, + 202 + ], + "type": "text", + "content": "Table 3. Ablation Study Results" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 202, + 545, + 272 + ], + "lines": [ + { + "bbox": [ + 319, + 202, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 319, + 202, + 545, + 272 + ], + "type": "table", + "html": "
MethodPSNR†(dB)Train↓(s)Storage↓(MB)
No-pretrained optical flow model31.072.657.90
No-projection-aware feature lift32.952.387.90
No-points bounded refinement33.233.02110.26
Ours-s(full)33.622.677.90
", + "image_path": "7ee517e43fa0e06adf98bed9a8b17a3a224c684a4eeac97718cc8a2ec71c85cf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 277, + 554, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 554, + 338 + ], + "type": "text", + "content": "from image pairs, which are then lifted into 3D space. To validate its effectiveness, we replaced the pretrained model with a 4-layer UNet without pretrained parameters and trained it jointly with the overall model. The results in Tab. 3 highlight the benefit of using the 2D prior." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 340, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 554, + 460 + ], + "type": "text", + "content": "Projection-aware 3D Motion Feature Lift: We use a projection-based approach to lift multi-view 2D motion feature maps into 3D space, accurately linking 3D anchor points to 2D features. To evaluate its effectiveness, we replaced this method with a Transformer-based approach using cross-attention between image features and anchor points, enhanced with positional embeddings through a 4-layer Transformer block. As shown in Tab. 3, Projection-aware Feature Lift is crucial for IGS performance, with only a slight increase in training time." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 460, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 594 + ], + "type": "text", + "content": "Key-frame guided Streaming: We employ a key-frame-guided strategy to address error accumulation in streaming and to enhance reconstruction quality. Keyframes are selected and refined through Max-points-bounded Refine. Without this refinement, AGM-Net would rely solely on Gaussians propagated from the last keyframe, resulting in accumulated errors that significantly impact performance, as shown in Fig. 6 (a). We also evaluate the effect of max-points bounding during refinement, as shown in Tab. 3. Without point limits, storage requirements increase substantially, and overfitting causes a decline in view quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Key-frame selection: We conducted an ablation study on the interval " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " for setting keyframes, testing values of " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "w = 1" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "w = 5" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "w = 10" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ", with results shown in Tab. 4. When " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "w = 1" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ", every frame becomes a keyframe, leading to excessive optimization that overfits Gaussians to training views, degrading test view quality and increasing training time and storage. Conversely, with " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "w = 10" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ", each keyframe drives the next 10 frames, but this distance weakens model performance, as it relies on assumptions about adjacent-frame" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16526" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 72, + 170, + 303 + ], + "blocks": [ + { + "bbox": [ + 82, + 72, + 170, + 303 + ], + "lines": [ + { + "bbox": [ + 82, + 72, + 170, + 303 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 170, + 303 + ], + "type": "image", + "image_path": "7bffae3ad7ed61269bbb878e43661d24901c5be646a5c65dedc77bcc27675dec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 305, + 140, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 305, + 140, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 140, + 314 + ], + "type": "text", + "content": "SaRo-GS" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 72, + 260, + 303 + ], + "blocks": [ + { + "bbox": [ + 170, + 72, + 260, + 303 + ], + "lines": [ + { + "bbox": [ + 170, + 72, + 260, + 303 + ], + "spans": [ + { + "bbox": [ + 170, + 72, + 260, + 303 + ], + "type": "image", + "image_path": "a2bf7d0683a3665f5dabf7338b770aa73a1c163ca8c17811a8f9cab80d867413.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 305, + 217, + 313 + ], + "lines": [ + { + "bbox": [ + 194, + 305, + 217, + 313 + ], + "spans": [ + { + "bbox": [ + 194, + 305, + 217, + 313 + ], + "type": "text", + "content": "4DGS" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 260, + 72, + 349, + 303 + ], + "blocks": [ + { + "bbox": [ + 260, + 72, + 349, + 303 + ], + "lines": [ + { + "bbox": [ + 260, + 72, + 349, + 303 + ], + "spans": [ + { + "bbox": [ + 260, + 72, + 349, + 303 + ], + "type": "image", + "image_path": "2e73f22f4004fbdd5f540f8e860143bdd06c71824407239939089a199cb29bd3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 305, + 328, + 313 + ], + "lines": [ + { + "bbox": [ + 286, + 305, + 328, + 313 + ], + "spans": [ + { + "bbox": [ + 286, + 305, + 328, + 313 + ], + "type": "text", + "content": "3DGStream" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 349, + 72, + 437, + 303 + ], + "blocks": [ + { + "bbox": [ + 349, + 72, + 437, + 303 + ], + "lines": [ + { + "bbox": [ + 349, + 72, + 437, + 303 + ], + "spans": [ + { + "bbox": [ + 349, + 72, + 437, + 303 + ], + "type": "image", + "image_path": "bcca0dd79eb747ab28c2aea8646ac6954ca9d5dea01cdfa5f34dff3fe35d6a36.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 305, + 393, + 313 + ], + "lines": [ + { + "bbox": [ + 378, + 305, + 393, + 313 + ], + "spans": [ + { + "bbox": [ + 378, + 305, + 393, + 313 + ], + "type": "text", + "content": "IGS" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 438, + 72, + 526, + 303 + ], + "blocks": [ + { + "bbox": [ + 438, + 72, + 526, + 303 + ], + "lines": [ + { + "bbox": [ + 438, + 72, + 526, + 303 + ], + "spans": [ + { + "bbox": [ + 438, + 72, + 526, + 303 + ], + "type": "image", + "image_path": "9c3521f2e7e087e78bf59eae62ec64021b6bd52d2b37045789320d45a7fbb557.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 305, + 484, + 313 + ], + "lines": [ + { + "bbox": [ + 471, + 305, + 484, + 313 + ], + "spans": [ + { + "bbox": [ + 471, + 305, + 484, + 313 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 57, + 331, + 175, + 421 + ], + "blocks": [ + { + "bbox": [ + 57, + 331, + 175, + 421 + ], + "lines": [ + { + "bbox": [ + 57, + 331, + 175, + 421 + ], + "spans": [ + { + "bbox": [ + 57, + 331, + 175, + 421 + ], + "type": "image", + "image_path": "72bc753b4ee5e0508e6a0643a4a359055bc2e65c9c962d279c1e6dcc69d40f2a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 423, + 168, + 430 + ], + "lines": [ + { + "bbox": [ + 70, + 423, + 168, + 430 + ], + "spans": [ + { + "bbox": [ + 70, + 423, + 168, + 430 + ], + "type": "text", + "content": "(a) Ablation Study on Key-frame Refinement" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 175, + 331, + 294, + 422 + ], + "blocks": [ + { + "bbox": [ + 200, + 314, + 409, + 324 + ], + "lines": [ + { + "bbox": [ + 200, + 314, + 409, + 324 + ], + "spans": [ + { + "bbox": [ + 200, + 314, + 409, + 324 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison from the N3DV dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 331, + 294, + 422 + ], + "lines": [ + { + "bbox": [ + 175, + 331, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 175, + 331, + 294, + 422 + ], + "type": "image", + "image_path": "014072f71d96268000634e44821552aacfa85e0c308b1760a85161fb961a991e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 423, + 276, + 430 + ], + "lines": [ + { + "bbox": [ + 203, + 423, + 276, + 430 + ], + "spans": [ + { + "bbox": [ + 203, + 423, + 276, + 430 + ], + "type": "text", + "content": "(b) Per-frame reconstruction time" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 431, + 295, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 431, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 295, + 452 + ], + "type": "text", + "content": "Figure 6. (a)Ablation Study on Key-frame Refinement. (b)Per-frame reconstruction time." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 75, + 478, + 276, + 533 + ], + "blocks": [ + { + "bbox": [ + 76, + 466, + 273, + 477 + ], + "lines": [ + { + "bbox": [ + 76, + 466, + 273, + 477 + ], + "spans": [ + { + "bbox": [ + 76, + 466, + 273, + 477 + ], + "type": "text", + "content": "Table 4. The impact of different keyframe intervals " + }, + { + "bbox": [ + 76, + 466, + 273, + 477 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 76, + 466, + 273, + 477 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 478, + 276, + 533 + ], + "lines": [ + { + "bbox": [ + 75, + 478, + 276, + 533 + ], + "spans": [ + { + "bbox": [ + 75, + 478, + 276, + 533 + ], + "type": "table", + "html": "
MethodPSNR(dB)↑Train(s)↓Storage(MB)↓
w=133.556.3836.0
w=533.622.677.90
w=1030.142.751.26
", + "image_path": "e00156d7868b5024f035889961f7a71108f0966bb4d64470dc209a0a6e0c6c38.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 540, + 295, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 576 + ], + "type": "text", + "content": "similarity. The setting " + }, + { + "bbox": [ + 55, + 540, + 295, + 576 + ], + "type": "inline_equation", + "content": "w = 5" + }, + { + "bbox": [ + 55, + 540, + 295, + 576 + ], + "type": "text", + "content": " strikes the best balance across view synthesis quality, train time, and storage, and is thus our final choice." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 590, + 125, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 125, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 125, + 602 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 611, + 281, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 281, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 281, + 624 + ], + "type": "text", + "content": "6.1. Independent per-frame reconstruction time" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "We further evaluate the performance of IGS by analyzing the independent per-frame reconstruction time, as shown in Fig. 6 (b). The reconstruction time for each frame exhibits a periodic pattern: for candidate frames, it takes " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.8\\mathrm{s}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ", while for key frames, it takes " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "4\\mathrm{s}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "7.5\\mathrm{s}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " for the small and large versions, respectively, which are significantly smaller than the " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "16\\mathrm{s}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " required by 3DGStream." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 331, + 387, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 331, + 387, + 342 + ], + "spans": [ + { + "bbox": [ + 314, + 331, + 387, + 342 + ], + "type": "text", + "content": "6.2. Limitation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 348, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 480 + ], + "type": "text", + "content": "IGS is the first to use a generalized method for streaming dynamic scene reconstruction, but it has limitations that can be addressed in future work. As shown in Fig. 3, our results exhibit jitter between adjacent frames, caused by the lack of temporal dependencies in the current framework. This makes the model more sensitive to noise. In contrast, 3DGStream assumes minimal motion between frames, yielding smoother results, but it fails in scenes with large motion (Fig. 4). To reduce jitter, we plan to incorporate temporal dependencies into IGS, modeling them as a time series for more robust performance." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 490, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 490, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 490, + 388, + 502 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose IGS as a novel streaming-based method for modeling dynamic scenes. With a generalized approach, IGS achieves frame-by-frame modeling in just over 2 seconds per frame, maintaining state-of-the-art rendering quality while keeping storage low. We introduce a generalized AGM-Net that lifts 2D multi-view motion features to 3D anchor points, using these anchors to drive Gaussian motion. This allows the model to infer the motion of Gaussians between adjacent frames in a single step. Additionally, we propose a Key-frame-guided Streaming strategy, where key frame sequences are selected and refined to mitigate error accumulation, further enhancing rendering quality. Extensive in-domain and cross-domain experiments demonstrate the strong generalization capabilities of our model, reducing significant streaming average cost while achieving state-of-the-art rendering quality, render speed, and storage efficiency." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16527" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "8. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "type": "text", + "content": "This work is financially supported for Outstanding Talents Training Fund in Shenzhen, this work is also financially supported by Shenzhen Science and Technology Program-Shenzhen Cultivation of Excellent Scientific and Technological Innovation Talents project(Grant No. RCJC20200714114435057), Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology(Grant No. 2024B1212010006), National Natural Science Foundation of China U21B2012." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 211, + 115, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 115, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 115, + 224 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 231, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 231, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 231, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 231, + 296, + 297 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855-5864, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 299, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 299, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 61, + 299, + 296, + 354 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 354, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 354, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 61, + 354, + 296, + 409 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 410, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 296, + 465 + ], + "type": "text", + "content": "[4] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 467, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 296, + 510 + ], + "type": "text", + "content": "[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 512, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 567 + ], + "type": "text", + "content": "[6] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 568, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 296, + 623 + ], + "type": "text", + "content": "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 624, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 666 + ], + "type": "text", + "content": "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "text", + "content": "[9] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In European Conference on Computer Vision (ECCV), 2024. 2, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[10] Yue Chen, Xingyu Chen, Xuan Wang, Qi Zhang, Yu Guo, Ying Shan, and Fei Wang. Local-to-global registration for bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8264-8273, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 172 + ], + "type": "text", + "content": "[11] Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. Hac: Hash-grid assisted context for 3d gaussian splatting compression. In European Conference on Computer Vision, pages 422-438. Springer, 2025. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 173, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 555, + 228 + ], + "type": "text", + "content": "[12] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "type": "text", + "content": "[13] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with " + }, + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "type": "inline_equation", + "content": "15\\mathrm{x}" + }, + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "type": "text", + "content": " reduction and " + }, + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "type": "inline_equation", + "content": "200+" + }, + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "type": "text", + "content": " fps. arXiv preprint arXiv:2311.17245, 2023. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 272, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 554, + 327 + ], + "type": "text", + "content": "[14] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 327, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 370 + ], + "type": "text", + "content": "[15] Xin Fei, Wenzhao Zheng, Yueqi Duan, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Jiwen Lu. Pixelgaussian: Generalizable 3d gaussian reconstruction from arbitrary views. arXiv preprint arXiv:2410.18979, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 372, + 555, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 555, + 426 + ], + "type": "text", + "content": "[16] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 426, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 555, + 491 + ], + "type": "text", + "content": "[17] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbaek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 1, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 493, + 554, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 554, + 546 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 554, + 546 + ], + "type": "text", + "content": "[18] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 548, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 554, + 602 + ], + "type": "text", + "content": "[19] Xiang Guo, Jiadai Sun, Yuchao Dai, Guanying Chen, Xiaoting Ye, Xiao Tan, Errui Ding, Yumeng Zhang, and Jingdong Wang. Forward flow for novel view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16022-16033, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "type": "text", + "content": "[20] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. ICCV, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 636, + 554, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 554, + 691 + ], + "type": "text", + "content": "[21] Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, and Yuewen Ma. Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19774–19783, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 692, + 554, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 715 + ], + "type": "text", + "content": "[22] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16528" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "curate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 295, + 139 + ], + "type": "text", + "content": "[23] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. arXiv preprint arXiv:2312.14937, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 57, + 140, + 294, + 183 + ], + "type": "text", + "content": "[24] Haian Jin, Hanwen Jiang, Hao Tan, Kai Zhang, Sai Bi, Tianyuan Zhang, Fujun Luan, Noah Snavely, and Zexiang Xu. Lvsm: A large view synthesis model with minimal 3d inductive bias. arXiv preprint arXiv:2410.17242, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 237 + ], + "type": "text", + "content": "[25] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365-18375, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[26] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "text", + "content": "[27] Hao Li, Yuanyuan Gao, Dingwen Zhang, Chenming Wu, Yalun Dai, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Junwei Han. Gprt: Towards generalizable 3d gaussians without pose priors in real-time. arXiv preprint arXiv:2403.10147, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "text", + "content": "[28] Haolin Li, Jinyang Liu, Mario Sznaier, and Octavia Camps. 3d-hgs: 3d half-gaussian splatting. arXiv preprint arXiv:2406.02720, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "text", + "content": "[29] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 482 + ], + "type": "text", + "content": "[30] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 483, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 483, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 294, + 514 + ], + "type": "text", + "content": "[31] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. arXiv preprint arXiv:2312.16812, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 559 + ], + "type": "text", + "content": "[32] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741-5751, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "text", + "content": "[33] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. High-fidelity and real-time novel view synthesis for dynamic scenes. In SIGGRAPH Asia 2023 Conference Papers, pages 1-9, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 604, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 658 + ], + "type": "text", + "content": "[34] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. Advances in Neural Information Processing Systems, 35:36762-36775, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[35] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splattering reconstruction from multi-view stereo. In European Conference on Computer Vision, pages 37-53. Springer, 2025. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "text", + "content": "[37] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 553, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 236 + ], + "type": "text", + "content": "[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 238, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 238, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 553, + 282 + ], + "type": "text", + "content": "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "type": "text", + "content": "[40] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 404 + ], + "type": "text", + "content": "[41] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 435 + ], + "type": "text", + "content": "[42] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "text", + "content": "[43] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 492, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 547 + ], + "type": "text", + "content": "[44] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "type": "text", + "content": "[45] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "type": "text", + "content": "[46] Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. arXiv preprint arXiv:2403.17898, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[47] Christian Schmidt, Jens Piekenbrinck, and Bastian Leibe. Look gauss, no pose: Novel view synthesis using gaussian splatting without accurate pose initialization. In IROS, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[48] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16529" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 182 + ], + "type": "text", + "content": "[49] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 295, + 216 + ], + "type": "text", + "content": "[50] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 295, + 281 + ], + "type": "text", + "content": "[51] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20675-20685, 2024. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 295, + 337 + ], + "type": "text", + "content": "[52] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, João F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. arXiv preprint arXiv:2406.04343, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "text", + "content": "[53] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 435 + ], + "type": "text", + "content": "[54] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 480 + ], + "type": "text", + "content": "[55] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 295, + 536 + ], + "type": "text", + "content": "[56] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 568 + ], + "type": "text", + "content": "[57] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "text", + "content": "[58] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "type": "text", + "content": "[59] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 76-87, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[60] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4150-4159, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "text", + "content": "[61] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "text", + "content": "[62] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 307 + ], + "type": "text", + "content": "[63] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "text", + "content": "[64] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "text", + "content": "[65] Han Xu, Jiteng Yuan, and Jiayi Ma. Murf: Mutually reinforcing multi-modal image registration and fusion. IEEE transactions on pattern analysis and machine intelligence, 45(10):12148-12166, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 410, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 487 + ], + "type": "text", + "content": "[66] Jinbo Yan, Rui Peng, Luyang Tang, and Ronggang Wang. 4d gaussian splatting with scale-aware residual field and adaptive optimization for real-time rendering of temporally complex dynamic scenes. In Proceedings of the 32nd ACM International Conference on Multimedia, page 7871-7880, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 488, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 542 + ], + "type": "text", + "content": "[67] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 544, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 589 + ], + "type": "text", + "content": "[68] Runyi Yang, Zhenxin Zhu, Zhou Jiang, Baijun Ye, Xiaoxue Chen, Yifei Zhang, Yuantao Chen, Jian Zhao, and Hao Zhao. Spectrally pruned gaussian fields with neural compensation. arXiv preprint arXiv:2405.00676, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "text", + "content": "[69] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[70] Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. arXiv preprint arXiv:2402.15870, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "text", + "content": "[71] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and render-" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16530" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 576 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "ing with 4d gaussian splatting. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[72] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[73] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "text", + "content": "[74] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "text", + "content": "[75] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[76] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 364 + ], + "type": "text", + "content": "[77] Chuanrui Zhang, Yingshuang Zou, Zhuoling Li, Minmin Yi, and Haoqian Wang. Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. arXiv preprint arXiv:2408.13770, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "text", + "content": "[78] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 410, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 443 + ], + "type": "text", + "content": "[79] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. In ECCV, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 444, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 509 + ], + "type": "text", + "content": "[80] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 576 + ], + "type": "text", + "content": "[81] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16531" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/671f9dcf-f80a-4cc0-9e9d-72da3c00b249_content_list.json b/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/671f9dcf-f80a-4cc0-9e9d-72da3c00b249_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..280701dcfdac39188dabb008cb3898c3d700be70 --- /dev/null +++ b/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/671f9dcf-f80a-4cc0-9e9d-72da3c00b249_content_list.json @@ -0,0 +1,2009 @@ +[ + { + "type": "text", + "text": "Instant3dit: Multiview Inpainting for Fast Editing of 3D Objects", + "text_level": 1, + "bbox": [ + 158, + 59, + 810, + 82 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amir Barda", + "text_level": 1, + "bbox": [ + 209, + 109, + 305, + 125 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tel Aviv University, Israel", + "bbox": [ + 151, + 127, + 362, + 143 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "amirbarda@mail.tau.ac.il", + "bbox": [ + 151, + 148, + 364, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Noam Aigerman", + "text_level": 1, + "bbox": [ + 179, + 170, + 315, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Université de Montréal, Canada", + "bbox": [ + 122, + 188, + 375, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "noam.aigerman@umontreal.ca", + "bbox": [ + 132, + 209, + 364, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matheus Gadelha", + "text_level": 1, + "bbox": [ + 431, + 109, + 573, + 125 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adobe Research, USA", + "bbox": [ + 411, + 127, + 591, + 143 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gadelha@adobe.com", + "bbox": [ + 424, + 148, + 578, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amit H. Bermano", + "text_level": 1, + "bbox": [ + 447, + 170, + 591, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tel Aviv University, Israel", + "bbox": [ + 416, + 188, + 624, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "amberman@mail.tau.ac.il", + "bbox": [ + 418, + 209, + 622, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vladimir G. Kim", + "text_level": 1, + "bbox": [ + 661, + 109, + 797, + 125 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adobe Research, USA", + "bbox": [ + 638, + 127, + 818, + 143 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vokim@adobe.com", + "bbox": [ + 660, + 148, + 795, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thibault Groueix", + "text_level": 1, + "bbox": [ + 686, + 170, + 826, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adobe Research, USA", + "bbox": [ + 666, + 188, + 846, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "grouelix@adobe.com", + "bbox": [ + 678, + 209, + 831, + 220 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79fe105ef7f700b9cc59e95bb49270ebf78853dbc93c5115b641fe8a874a30a2.jpg", + "image_caption": [ + "\"A bear holding" + ], + "image_footnote": [], + "bbox": [ + 127, + 244, + 205, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/573220bdcc1e193dac273bf085b351ecf0925f79c9993c3c473de19cb0aab423.jpg", + "image_caption": [ + "Gaussian Splat", + "a honey pot\"" + ], + "image_footnote": [], + "bbox": [ + 225, + 263, + 295, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e8893b8f51d4c62d1103167b473eef4d5fdd284a56c5502ab3dab40cbd945a4.jpg", + "image_caption": [ + "${a}_{i} \\in a\\;{a}_{i}$ 分母最小值为 $k$ ." + ], + "image_footnote": [], + "bbox": [ + 295, + 252, + 364, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8166ce8aa3678494c82b7abe09509ff13aaf6f6226417d54cbe271d41d4e19b1.jpg", + "image_caption": [ + "Original Mesh" + ], + "image_footnote": [], + "bbox": [ + 364, + 242, + 434, + 343 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa68e11ac21493e12a835275d69d2f2240cc443dae875f17d71275535d7d0d74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 306, + 504, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0cd4098025325c8447cfd88f9767fb4eec25b0ab51ef50b386aee663f4bd9c85.jpg", + "image_caption": [ + "\"Bear with wings\"" + ], + "image_footnote": [], + "bbox": [ + 119, + 361, + 210, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7283166adb068aa4322e2940d9f1f1459d8aef7b230a3c57018e27d82e2d378f.jpg", + "image_caption": [ + "3 sec." + ], + "image_footnote": [], + "bbox": [ + 212, + 363, + 297, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86146188ed847b276b674be09883f54afc2f63dfac6932aedb148c0450e0ac88.jpg", + "image_caption": [ + "6 sec." + ], + "image_footnote": [], + "bbox": [ + 299, + 361, + 372, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6833eb127da64f346283d66a86ea40f273c5f2273fdb501e30c76d622ee1688b.jpg", + "image_caption": [ + "25 sec.", + "Figure 1. Our method takes as input a 3D object along with a 3D mask (first column) and a text prompt, and uses our multiview inpainting diffusion model to consistently paint the mask in four rendered views of the object. Off-the-shelf reconstructors can be used on the multiview output to give an NeRF, a Gaussian Splat (second column), or a mesh (third column) that can be used along with adaptive remeshing to ensure the unmasked region is exactly preserved e.g. topology, uvs, (fourth and fifth column). This feedforward approach is orders of magnitude faster than previous works in generative 3D editing, taking just $\\approx$ 3 seconds per multiview edit, then 0.7 seconds to reconstruct a GS or a NeRF, 3 seconds for a mesh, and $\\approx$ 20 seconds for optional mesh post-processing." + ], + "image_footnote": [], + "bbox": [ + 374, + 361, + 452, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e3c073f6479ad717196dd72a625a85c59f5bc636234bdfc93d7c0047ce77ed4.jpg", + "image_caption": [ + "Input mesh $^+$ mask", + "\"Man wearing" + ], + "image_footnote": [], + "bbox": [ + 514, + 271, + 578, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f9dbba604eb1f012e8c8e51b8023335f002cc44f35f012cfce1d11369f1d4b6d.jpg", + "image_caption": [ + "NeRF", + "a medieval" + ], + "image_footnote": [], + "bbox": [ + 598, + 271, + 656, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/98a0dc3427157cc12496097e2c71f9df2bf1a37e816943067548ccd4f12c7771.jpg", + "image_caption": [ + "Mesh", + "helmet\"" + ], + "image_footnote": [], + "bbox": [ + 666, + 271, + 725, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d33af5dc58501f133fd22e8fc69c3f544ed82f13e0b07c8e030478f384e9128f.jpg", + "image_caption": [ + "Adaptive Remeshing", + "Original Mesh" + ], + "image_footnote": [], + "bbox": [ + 733, + 270, + 794, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9a33730c73c7d788ded54a7717ed274888d07e540f9fc5eeeb305de90b2bfd01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 315, + 851, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ccbdc4bb131a513338bb237c9ee25eaf6ada74274fb4028fd50d3d3015f637d.jpg", + "image_caption": [ + "\"An elven warrior\"" + ], + "image_footnote": [], + "bbox": [ + 516, + 359, + 581, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5390b3477b5e65ffdcb04a3bf3d0ddee8a8e7809ed868552402e0c4af66a3d00.jpg", + "image_caption": [ + "3 sec." + ], + "image_footnote": [], + "bbox": [ + 596, + 359, + 658, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb7b766dfa6acb755c9f5c869ead7f66dce696fa76710231f9bc638cafe9161b.jpg", + "image_caption": [ + "6 sec." + ], + "image_footnote": [], + "bbox": [ + 669, + 359, + 730, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/44a256c8841cd2daf8ede1930ecb7c0b69843be9836553a987c11e6b0d870922.jpg", + "image_caption": [ + "25 sec." + ], + "image_footnote": [], + "bbox": [ + 736, + 359, + 795, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 570, + 313, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a generative technique to edit 3D shapes, represented as meshes, NeRFs, or Gaussian Splats, in $\\approx 3$ seconds, without the need for running an SDS type of optimization. Our key insight is to cast 3D editing as a multiview image inpainting problem, as this representation is generic and can be mapped back to any 3D representation using the bank of available Large Reconstruction Models. We explore different fine-tuning strategies to obtain both multiview generation and inpainting capabilities within the same diffusion model. In particular, the design of the inpainting mask is an important factor of training an inpainting model, and we propose several masking strategies to mimic the types of edits a user would perform on a 3D shape. Our approach takes 3D generative editing from hours to seconds and produces higher-quality results compared to previous works.", + "bbox": [ + 75, + 590, + 470, + 833 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 849, + 209, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have seen an explosion in AI-guided 3D generative techniques for creation and editing of 3D content, al", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lowing the user to control the generation process with often as little as a single text prompt [4, 9, 10, 42, 50, 55]. While text prompts provide an easy interface for novice users, they lack fine control over the generation, such as generating a specific object in a specific location over a pre-existing 3D model, e.g., make a bear hold a previously non-existent honey pot (see Figure 1).", + "bbox": [ + 496, + 571, + 892, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thus, several recent works have focused on such localized generation of 3D objects, namely 3D inpainting [6, 42, 50], i.e., filling-in masked-out content in a 3D object, conditioned on a textual description of the desired fill-in, similarly to 2D inpainting [8, 50]. Figure 1 shows several examples of such inpaintings produced by our method, where the mask (made up of coarse 3D primitives) is highlighted in purple color, the text prompt describing the target inpainting is written below, and the resulting inpainted 3D reconstructions are shown next to it.", + "bbox": [ + 496, + 683, + 892, + 834 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current 3D inpainting methods cannot be integrated into production pipelines, as they suffer from two fundamental issues: 1) long runtimes, and 2) low quality. We observe these two issues can be explained by previous works' re", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 20 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 17, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 45 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16273", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liacence on optimization of the 3D model by distilling knowledge from a generative model for 2D images via some variant of Score Distillation Sampling (SDS) [37]. Such an optimization process is extremely slow, as it relies on running an image diffusion model over multiple renderings of the 3D object and back-propagating gradients, thus explaining the first issue of slow runtimes. However, we argue that it also lies at the root of the second issue, of degradation in quality: as we show through extensive comparisons (see Figure 6), this class of approaches significantly harms the quality of the inpainting, as SDS-like optimization produces inaccurate and fuzzy results. We attribute it to the fact observed in the original DreamFision paper [37], which stated that \"2D image samples produced using SDS tend to lack diversity\" and \"3D results exhibit few differences across random seeds,\" explaining this behavior with the tendency to seek specific distribution modes regardless of the seed. Thus, it stands to reason it will struggle to inpaint a region of an existing object that is not one of those seeked modes.", + "bbox": [ + 75, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose an alternative approach to 3D inpainting that significantly improves the quality of the inpainted results and the runtime of the process. Drawing inspiration from previous works for fast 3D generation [26, 31, 52, 58], we eliminate both the above issues, by turning the problem on its head: instead of directly optimizing a 3D object using, e.g., SDS, we instead train an image generator to create 2D images of the inpainted 3D object from canonical viewpoints, and then reconstruct the newly-inpainted 3D object in a post-process, via either a feed-forward prediction [26, 54, 58], or lightweight optimization [31, 52]. By doing so we show we avoid both the slow runtimes as well as issues with masking that arise from approaches such as SDS when applied to 3D inpainting.", + "bbox": [ + 75, + 380, + 472, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adapting this multiview generation technique to the inpainting problem leads to a core challenge: how to ensure that the inpainting of masked region in the different 2D generated image is 3D-consistent from different viewpoints (otherwise, it does not represent a coherent 3D object that can be reconstructed), while at the same time adhering to the 3D mask painted over the original 3D object. Thus, following our observation above, our main technical contribution is designing a scheme to obtain a diffusion model that provides multiview-consistent inpainting.", + "bbox": [ + 75, + 594, + 472, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to achieve a multiview-consistent inpainting diffuser, we devise a custom training strategy, along with a novel dataset of 3D masks for 3D inpainting. Namely, we produce multiview consistent masks, avoiding problems resulting from occlusion (that our experiments validate is critical for performance). We design the dataset to support three designated editing modes with different levels of granularity (see Section 3.2)", + "bbox": [ + 75, + 747, + 472, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we craft a custom training strategy to leverage the priors learned by pretrained text-conditioned image gen", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "erators. We propose to start from a pretrained image inpainter, and use our dataset and training regime to make it multiview consistent. We provide extensive empirical evidence to the specific design choices of our approach, e.g., compare our strategy with the opposite route – fine-tune a multiview diffuser to perform inpainting.", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show through extensive experiments that our approach enables performing elaborate 3D editing using simple masking of a 3D region, rendering the object and the mask from multiple views, running the inpainting model, and using fast multiview reconstruction techniques to obtain the edited 3D object (see Figure 1). Our approach is thus agnostic to the underlying 3D representation, and we show it supports meshes [54], Gaussian Splats [58] and Radiance Fields [54]. Our experiments validate the speed of our multiview image-based representation and the higher quality of the resulting inpainting compared to previous techniques. To summarize, our contributions are:", + "bbox": [ + 496, + 183, + 893, + 362 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The first high-quality, fast method for 3D inpainting, enabling fast and localized generation on NeRFs, Gaussian Splats and meshes.", + "- A 3D masking approach along with a novel training dataset of masks for multiview inpainting, with three types of masks corresponding to different edit modes.", + "- A finetuning strategy to efficiently leverage pretrained representations, along with an empirical study of alternative training strategies for multiview inpainting." + ], + "bbox": [ + 500, + 364, + 890, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 516, + 640, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We position our work against other generative 3D editing approaches and separately discuss multiview 3D generation and inpainting. Closest to our approach are NeRFiller [50] and the concurrent work MVInpainter [6].", + "bbox": [ + 496, + 542, + 893, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Generative Editing. As we aim for a general tool, we focus the discussion on category-agnostic approaches. The literature can be organized by the core technique used to distill the 2D generative prior - CLIP optimization [16, 22, 34], Score-Distillation Sampling (SDS) [4, 8-11, 23, 24, 27, 33, 35, 42, 60], Iterative Dataset Update (IDU) [17, 50] - or by the type of representation targeted - meshes [4, 16, 22, 23, 34], NeRFs [8, 17, 33, 35, 42, 50], or SDF [9, 10, 27, 60]. All generative editing approaches take advantage of Text-to-Image (T2I) models, whose prowess stems from the billions of text-image pairs available as training data. This abundance of data does not extend to the 3d domain, with the largest public datasets, Objaverse and Objaverse-XL [12, 13], containing millions of meshes, with the vast majority being very low-quality data.", + "bbox": [ + 496, + 612, + 893, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Score Distillation Sampling (SDS) loss, first defined in [37], use T2I models as guidance for 3d generation, by iteratively denoising renderings. SDS has multiple downsides: the denoised images lack consistency at different", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16274", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7e2b0d1b6a5a6335c1600677baf045201fad3dc255011e892b8434eeb034a23d.jpg", + "image_caption": [ + "Figure 2. Overview. Given a NeRF, a Gaussian Splat, or a mesh, the user draws a 3D mask to mark a region to be filled and provides a text prompt to guide the generation. Instant3dit renders four canonical views of the masked object and uses our multiview inpainting network to fill the mask. We use off-the-shelf 3D reconstructors to convert the multiview representation into a NeRF, a Gaussian Splat, or a mesh." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 872, + 204 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "iterations, and the optimization is therefore very noisy, requiring careful parameter tuning to converge. Second, the optimization typically takes around one hour to converge, as it requires a lot of steps [4]. Combining SDS with Gaussian Splatting partially alleviate this problem by making the rendering operation in each step much faster, as shown in GaussianDreamer [55], but is still not interactive (about 15 minutes per generation). Lastly, SDS produces saturated, low-quality textures due to the high classifier-free guidance used, even though recent variations alleviate this issue [32].", + "bbox": [ + 75, + 277, + 468, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Iterative Dataset Update, first introduced in [17], is a variant of SDS, taking multiple denoising steps to produce a clean image, used for several optimization steps in a row. NeRFiller [50] extends this approach for editing of NeRFs.", + "bbox": [ + 76, + 431, + 468, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to prior work, our approach does not require an optimization to leverage the generative prior. We propose to perform the edit using a multiview image grid to represent the 3D asset, with a single diffusion inference. As a result, editing is orders of magnitude faster (i.e. from hours to seconds) and more stable (i.e. our approach always robustly produces a result and does not require careful tuning of hyperparameters). Lastly, several off-the-shelf LRM transformers [19] can instantly convert our inpainted image grid to meshes [51], NeRFs [26], or Gaussian Splats [58], making our editing technique representation agnostic.", + "bbox": [ + 75, + 496, + 470, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multiview generative 3D. To encourage multiview 3D consistency, several works have explored modifying the attention layers with cross-attention blocks to facilitate the exchange of information between views [29, 30, 44]. A simpler approach consists of directly generating a $2 \\times 2$ image grid, without modifying the attention layers. In this case, the exchange of information between view is done via self-attention. This effectively trades resolution with consistency, and can be achieved by conditioning a T2I model on multiview depth [7, 46], or by fine-tuning on a dataset of 3D multiview renderings [26, 43, 47]. Unique3D [52] and direct 2.5d [31] extend this approach to image grids of normals, and CRM [49] to canonical coordinate map. These works have shown that adopting this representation to generate 3D shape is orders of magnitude faster than SDS, from", + "bbox": [ + 75, + 674, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "hours to seconds. In this work, we extend these ideas to 3D editing, and explore fine-tuning strategies to get inpainting and multiview generation in the same model. Of note, NeRFiller [50] proposes to inpaint a grid of $2 \\times 2$ images with a single-view inpainting model without fine-tuning it for multiview consistency. We compare against this baseline and show the importance of fine-tuning the inpainting model. Furthermore, NeRFiller relies on a more costly IDU optimization, which requires about $30 \\mathrm{k}$ steps.", + "bbox": [ + 496, + 277, + 893, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image inpainting. The ability to collect and annotate images at scale has fueled major progress in image inpainting. Current inpainting tools have a wide range of applications ranging from real photographs to illustrative drawing [1, 41]. We refer the reader to the recent survey of Xiang et al. [53]. In this paper, we are interested in leveraging these pre-trained priors to edit 3D assets. As discussed in the previous paragraph, the main challenge is achieving multiview consistent results, as open-source models like Stable Diffusion [41] inpaint each frame independently without 3D consistency. In this work, we propose to teach multiview consistency to a single-view inpainting network. As noted in Zeng et al. [57], the design of the inpainting masks is a critical part of training an inpainting model. The masks should resemble the type of masks that users will draw at inference as closely as possible. However, how to apply that insight to 3D editing is not straightforward. We propose three masking strategies that address different workflows in Section 3.2.", + "bbox": [ + 496, + 422, + 892, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concurrent work. Similarly observing that IDU and SDS are unstable and lengthy optimizations, the recent approach MVInpainter [6] adds a video priors to a single-view inpainting model, via a LoRA [20], to encourage multiview consistency. Their focus in on object insertion and removal in captured 3D scenes, while ours in on 3D generative editing of objects, which leads us to different masking strategies to train the inpainting model.", + "bbox": [ + 496, + 709, + 893, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 844, + 591, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The input to our system consists of tuples $\\langle S, M, y \\rangle$ representing a 3D shape $S$ , a 3D mask region $M$ , and a text de", + "bbox": [ + 500, + 869, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16275", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/626ed0dc9f9e7733cbb7793b7115543aca752947efa38c0deab26323e6de0a93.jpg", + "image_caption": [ + "Figure 3. Multiview representation. We represent 3D shapes multiview renderings. Editing is done using an image-based diffusion model that operates on $I_{c}(S,M)$ and $I_{b}(M,S)$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 236, + 193 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7b22beea185b5ec6467bc039bfe39641ed64f7893b7860268a55ffcceefc7a7d.jpg", + "image_caption": [ + "$I(S)$" + ], + "image_footnote": [], + "bbox": [ + 238, + 107, + 325, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c761dc83475868ce14d3342c4e1dde46cc4b4e22d345622a9a41e03d468d2c76.jpg", + "image_caption": [ + "$I(M)$" + ], + "image_footnote": [], + "bbox": [ + 326, + 108, + 410, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "scription $y$ - henceforth referred to as prompt. The goal of our method is to create a new shape $S'$ whose areas covered by $M$ are modified to follow $y$ .", + "bbox": [ + 75, + 270, + 470, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multiview representation. We propose to perform 3D inpainting by inpainting multiple renderings of an object in a view-consistent manner. Consider a rendering operator $\\mathcal{R}$ that renders a set of shapes (a scene) $S$ , from a viewpoint $\\pi$ . We refer to $\\mathcal{R}_c$ and $\\mathcal{R}_b$ to indicate a rendering to a color RGB image and a binary mask, respectively. Finally, we refer to $\\mathcal{R}^U$ to indicate that we only render the visible pixels belonging to the shape $U$ while assuming $U \\in S$ . Using this operator, we define the multiview representations as:", + "bbox": [ + 75, + 324, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {k} (U, V) := \\bigoplus_ {\\pi \\in \\Pi} \\mathcal {R} _ {k} ^ {U} [ \\{U \\} \\cup \\{V \\}; \\pi ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 470, + 468, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bigoplus$ concatenates the images in a $2\\times 2$ grid, $k$ is rendering modality (color or binary), and $\\Pi = \\{C(\\alpha ,\\frac{\\pi}{4})|\\alpha \\in$ $\\{0,\\frac{\\pi}{2},\\pi ,3\\frac{\\pi}{2}\\} \\}$ . $C(\\alpha ,\\beta)$ is a function that returns a viewpoint configuration corresponding to a camera pointing at the origin of the coordinate system and positioned on the surface of a canonical sphere according to azimuth $\\alpha$ and elevation $\\beta$ . Using this operator, the input to our multiview diffusion model is defined as $I_{c}(S,M)$ and $I_{b}(M,S)$ . Intuitively, $I_{c}(S,M)$ is just an image containing the visible pixels of $S$ in the scene $\\{S\\} \\cup \\{M\\}$ rendered from multiple views organized in a grid, and $I_{b}(M,S)$ is binary rendering of the visible pixels of $M$ . An illustration of this representation is presented in Figure 3.", + "bbox": [ + 75, + 513, + 468, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D editing. Given $I_{c}(S,M)$ and $I_{b}(M,S)$ , we propose to use a diffusion model $\\epsilon_{\\theta}$ to generate an inpainted multiview representation $\\hat{I}_{c}$ . From that, we can then use a posed multiview reconstruction process $\\Phi$ to obtain the edited shape $\\hat{S} = \\Phi (\\hat{I}_c)$ . Different choices for $\\Phi$ yield different applications and tradeoffs. For example, we can have very fast learning-based reconstruction from posed multiview images using various representations, like NeRFs [26], meshes [51] and gaussian splats [58]. We can also apply lightweight optimization routines based on differentiable rendering such as ROAR [3], ISOMER [52] or Direct2.5d [31]. This class of methods is usually slightly slower but has the additional", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "benefit of allowing us to carefully craft the optimization procedure in order to achieve several desirable properties like geometric regularization and preservation of the original asset attributes – color, connectivity, UVs, and so on. Figure 1 shows results using various types of reconstructors. We use NeRF-LRM [26] to compare against other NeRF editing approaches in Table 1 and Figure 4. Generating an image with $\\epsilon_{\\theta}$ takes about 3 seconds on an A100, and the various reconstructors range from a few milliseconds [26, 58] to a few seconds [31, 51, 52].", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the rest of this section, we explain how to train the diffusion model $\\epsilon_{\\theta}$ for multiview generation (Section 3.1) and inpainting (Section 3.2).", + "bbox": [ + 496, + 242, + 890, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Background on Diffusion Models.", + "text_level": 1, + "bbox": [ + 500, + 299, + 794, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training. During training, we sample an image $x$ from the dataset, with condition $c$ (e.g. text, mask, or depth), a time step $t$ between 0 and $T$ , and a noise $\\epsilon \\sim \\mathcal{N}(0,1)$ , injected to $x$ to create a noisy image $\\tilde{x}(t)$ :", + "bbox": [ + 496, + 332, + 890, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {x} (t) = \\sqrt {\\alpha (t)} \\cdot x + \\sqrt {1 - \\alpha (t)} \\cdot \\epsilon , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 409, + 890, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha(t)$ controls the amount of noise to inject i.e. $\\alpha(0) = 1$ is no noise and $\\alpha(T) = 0$ is pure noise. A denoising unet $\\epsilon_{\\theta}$ is trained to denoise $\\tilde{x}(t)$ by minimizing:", + "bbox": [ + 498, + 435, + 890, + 482 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d i f f}} = w (t) \\left\\| \\epsilon_ {\\theta} (\\tilde {x} (t); t, c) - x \\right\\| ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 494, + 890, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w(t)$ a scheme to scale the gradients according to $t$ . Once $\\epsilon_{\\theta}$ is trained, $\\epsilon_{\\theta}(\\tilde{x}(t); t, c)$ is the projection of $\\tilde{x}(t)$ to the manifold of images defined by the training dataset. Details about the training procedure can be found in the supplemental material.", + "bbox": [ + 496, + 523, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inference. We start from pure noise $\\tilde{x}(T)$ and follow the direction of the manifold defined by $\\epsilon_{\\theta}(\\tilde{x}(t); t, c)$ . There exists multiple samplers to discretize this trajectory into a discrete number of steps. In practice, we use the Euler scheduler [21] and 29 steps.", + "bbox": [ + 496, + 609, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent Diffusion. In this work, we use latent diffusion models i.e. the diffusion happens in a 4-dimensional latent space instead of RGB space, and a pretrained VQ-VAE [40] encodes and decodes images from that space. Our method is agnostic to this, so we keep the presentation general.", + "bbox": [ + 496, + 694, + 890, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multiview diffusion. To generate $2 \\times 2$ consistent views, we follow [26] and simply replace the training distribution with a distribution of $2 \\times 2$ images. While this approach would yield poor results if it were trained from scratch, given the scarcity of high-quality 3D data, it performs well if the models are fine-tune from \"foundation\" text-to-image models, i.e. pre-trained on millions of images. To create the dataset, we render a curated list [26] of 5K objects from", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16276", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Objaverse [13], filtered for high-quality, and generate high-quality captions $y_{\\mathrm{n}}$ for each 3D object with LLaVa [28]. We now explain how to train for inpainting jointly.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inpainting. The main specificity is that the condition $c$ is composed on the text prompt $y$ , the base image with holes, and the inpainting mask i.e. $c = \\{y, I_c(S, M), I_b(M, S)\\}$ . In practice, for latent models, $I_c(S, M)$ is passed through the encoder $\\mathcal{E}$ of the VQ-VAE, concatenated with the downsampled version of the mask $I_b(M, S)$ , and the noisy latents $\\tilde{x}(t)$ , leading to a 9-channel tensor input to the denoising unet $\\epsilon_{\\theta}$ , along with the encoding of the text condition. During training, we randomly drop the mask $10\\%$ of the time, falling back to multiview diffusion training.", + "bbox": [ + 75, + 145, + 470, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Multi-View Inpainting Masks", + "text_level": 1, + "bbox": [ + 76, + 305, + 341, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset creation. A key part of our method consists in generating multiview masks that are 3D consistent. Thus, the binary masks used for training our multiview inpainting model are obtained by rendering 3D shapes; i.e. $I_{b}(M,S)$ . Consider the example in Figure 3. Even though $M$ is simply an ellipsoid, its multiview representation $I_{b}(M,S)$ is not - it has the occlusions from its interaction with $S$ . We empirically demonstrate that 3D-aware masks and multiview consistent images are crucial for better performance (see Table 2). Our model is trained based on a set of shapes $\\mathcal{D}$ . For every shape $S \\in \\mathcal{D}$ , we create a set of 3D masks $\\mathcal{M}_S$ . Using those, we can define our training dataset $\\mathcal{I}$ as:", + "bbox": [ + 75, + 325, + 468, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} := \\left\\{ \\begin{array}{c} \\left\\langle I _ {c} (S, \\varnothing), I _ {c} (S, M), I _ {b} (M, S), y _ {S} \\right\\rangle \\\\ \\text {w h e r e} S \\in \\mathcal {D}, M \\in \\mathcal {M} _ {S} \\end{array} \\right\\} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 517, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I_{c}(S, \\varnothing), I_{c}(S, M), I_{b}(M, S)$ are the color ground-truth image, color input image, and binary input mask, respectively. $y_{S}$ is a text prompt describing the shape $S$ obtained from a VLM model [28]. Since generating the masks online would slow down training, we preprocess them offline. In practice, we have $|\\mathcal{M}_S| = 30$ , containing equal portions of 3 different kinds of masks. Considering that $|\\mathcal{D}| \\approx 5K$ , our multiview dataset $\\mathcal{I}$ has $\\approx 150K$ data points. We will release our dataset upon publication.", + "bbox": [ + 75, + 560, + 468, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A key remaining issue is how $\\mathcal{M}_S$ is created. As noted in [57], the design of the training masks plays a central role in training an inpainting model. The best performances are naturally obtained when the distribution of training masks closely follows the distribution of edits that users will make at test time. We confirm this in our experiments with a simple ablation, where we naively use random 2D mask (see Table 2). We thus propose three types of masks, corresponding to three types of editing. The remaining of this section will explain how each type of mask is created.", + "bbox": [ + 75, + 696, + 468, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Type I: coarse edit. In this setting, the inpainted part of $S$ is fully contained", + "bbox": [ + 76, + 854, + 267, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7a542f470ab3eb8a79488bfb5d39b1288edd350105e276c7e09712689b655007.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 852, + 468, + 912 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "inside $M$ . $M$ is computed", + "bbox": [ + 496, + 90, + 692, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by randomly sampling a part of $S$ and taking its convex hull. To select such part, we randomly sample a plane passing through $S$ , effectively splitting $S$ into two parts, and we select one part at random. More precisely, we start by sampling a random point $p$ inside the bounding box of $S$ and a random direction $n$ . The plane $P$ passing through $p$ with normal $n$ is defined by $\\{x \\in \\mathbb{R}^3 | x \\cdot p = p \\cdot n\\}$ . $M$ is defined as the convex hull of all the face midpoints that are above $P$ i.e. $\\{f = (v_1, v_2, v_3) | f \\in F, \\frac{v_1 + v_2 + v_3}{3} \\cdot p >= p \\cdot n\\}$ , where $F$ denotes the list of faces. To avoid Z-fighting during rendering between $S$ and $M$ , we scale $M$ by 20% while keeping its center of mass the same, ensuring it completely envelopes the part of $S$ above $P$ .", + "bbox": [ + 496, + 106, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Type II: mesh sculpting.", + "text_level": 1, + "bbox": [ + 498, + 311, + 689, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This type of mask is designed to represent a more precise edit where the user", + "bbox": [ + 498, + 328, + 689, + 372 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af72ce7a09671870435d0637e91efa18a5a4f45a20d6f2e79494298a7ff6d688.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 313, + 890, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "expects content to be created in a portion of space similar to the mask. It requires more expertise and time from the user than Type I, but also provides more precise control over the generated content. As we can see in the illustration to the right, $M$ is a tight fit over $S -$ no volume inside $M$ is not also inside $S$ . To generate these masks, we sample a plane $P$ , as in Type I masks, and select all the faces that have their midpoint above $P$ i.e. $M = \\{f = (v_{1}, v_{2}, v_{3}) | f \\in F, \\frac{v_{1} + v_{2} + v_{3}}{3} \\cdot p > = p \\cdot n\\}$ .", + "bbox": [ + 496, + 373, + 890, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Type III: surface editing.", + "text_level": 1, + "bbox": [ + 498, + 518, + 689, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We aim to support local texture modifications, where the user selects a surface", + "bbox": [ + 498, + 534, + 689, + 579 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/949faf04bdd4cb2f8ebfaf692a32b0b3496c49819e31a8ba68937a3f65a07338.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 517, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "patch and prompts Instant3dit to modify its texture. There are various ways to generate this type of mask if we assume that the meshes have good triangulation. Unfortunately, this is typically not the case for shape datasets. Thus, propose a simple heuristic that works for any triangle soup. We sample a vertex $p$ in $S$ , then several cylinders with elliptical bases of varying sizes, all centered on $p$ (depicted in purple). The number of cylinders is uniformly sampled between 3 and 6, the revolution axis is sampled on the unit sphere, the height and radii are sampled between 0.1 and 0.3. We call this volume $C$ . Finally select all the faces whose midpoint falls within $C$ i.e. $M = \\{f = (v_{1}, v_{2}, v_{3}) | f \\in F, \\frac{v_{1} + v_{2} + v_{3}}{3} \\in C\\}$ (depicted in red).", + "bbox": [ + 496, + 580, + 890, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 806, + 633, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe our novel multiview inpainting benchmark, which we use to compare several baselines and run ablations.", + "bbox": [ + 496, + 828, + 890, + 872 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark. We hold out 500 2x2 multiview images and", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16277", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/169028abaf099140377ae6f8b823c15d0e6bf03b031f1404693c6c0428cc188b.jpg", + "image_caption": [ + "Figure 4. Comparison to baselines. We show different inpainting results from different baselines; our multiview inpainting method offers the highest quality while maintaining consistency." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 468, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "their synthetic multiview masks from the training set to create our benchmark. We condition all inpainting on BLIP [25] captions, and ensure no training object is present in the benchmark. We then evaluate how well different methods inpaint all the views.", + "bbox": [ + 75, + 261, + 470, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate the quality of multiview inpainting, we use three types of metrics: measuring prompt adherence, multiview consistency, and visual quality. To measure prompt adherence, we use CLIP [38] similarity score between the generated 2x2 multiview image and the text prompt. We follow previous work [36] and use two models, CLIP-ViT-L-14 (ClipL) and CLIP-ViT-BigG-14 (ClipG), for encoding. To measure 3D consistency we reconstruct NeRF from the sparse inpainted views [26], and re-render from same camera angles as the input sparse views. We then compare the NeRF renderings to generated inpainted images using various image similarity scores: SSIM [48], LPIPS [59] and DreamSim [14]. Finally, we measure visual quality using FID score [18] comparing to a distribution of held-out images.", + "bbox": [ + 75, + 348, + 472, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison to Baselines. We use these metrics to compare our method (Table 1, bottom) to several baseline alternatives (Table 1, top). As there are no off-the-shelf multiview consistent inpainting methods, we couple different image diffusion techniques with blended diffusion [2] to inpaint the multiview image. We try two different architectures from the commonly used SDXL diffusion model [36], SDXL (as-is) and SDXL-inpainting (fine-tuned on 2D inpainting task). Both SDXL baselines perform poorly on multiview consistency since they are not explicitly trained to complete images consistently and, in the case of SDXL-inpainting, are only trained with 2D image masks. Finally we test Instant3D [26] coupled with blended diffusion. Since this architecture is trained for 3D reconstruction it achieves a better multiview consistency, however, our method still outperforms this baseline with respect to all other metrics. See Figure 4 for some example results produced with baselines and our method.", + "bbox": [ + 75, + 585, + 470, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablating Diffusion Backbones. Our method can be fine-tuned from any image diffusion model, and thus we", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3191bc7c6e1681e0b94c70dd69dc7dde83c951e2d42f5fb0d03f183e774ad128.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPrompt AdherenceMulti-view consistencyVisual quality
ClipL↑ClipG↑SSIM↑LPIPS↓DreamSim↓FID↓
SDXL [36]28.6342.580.8740.0650.134127.0
SDXL-inpainting [36]27.5741.330.8570.0640.137159.5
Instant3D [26]28.7843.060.8920.0440.102120.0
ablations from different diffusion backbones (our final method is at the bottom)
Ours (SD 1.5-inpainting)27.7941.330.7190.090.599128.3
Ours (SD 2.0-inpainting)27.5941.230.7290.0960.589124.5
Ours (Instant3D)28.5742.500.8940.0430.097121.1
Ours (SDXL-inpainting)29.0143.480.8940.0450.100118.4
", + "bbox": [ + 501, + 87, + 890, + 205 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/744674b4de2b929bea86cdfae192175ee404b9b05505a7a3ab8f65c8855964c0.jpg", + "table_caption": [ + "Table 1. Multiview text-to-image inpainting. This table demonstrates quantitative performance of several baselines and ablations of our method, with respect to metrics that capture prompt adherence, multiview consistency, and visual quality. Results are color-coded worst and best, with best highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodType IType IIType IIII+II+IIIUser Generated
FID↓ClipG↑
Random 2D145.4129.6102.3131.124.22
Type I131.2128.6102.3121.326.24
Type II170.9122.6101.7128.225.53
Type III193.0148.199.05142.224.2
I+II+III130.0124.9100.0118.426.5
", + "bbox": [ + 522, + 308, + 867, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Mask ablation. We ablate the choice of 3D masks used during training (rows), and evaluate on different subsets of the benchmark containing only some types of masks (columns).", + "bbox": [ + 496, + 412, + 892, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "further evaluate how performance changes with respect to different backbones (Table 1, bottom section). Unsurprisingly, weaker backbones with fewer parameters (SD 1.5, SD 2.0) [40] lead to inferior performance and very poor multiview consistency. Using Instant3D [26] gives the best multiview consistency, but fine-tuning does not seem to improve visual quality and prompt adherence. Our method fine-tunes from SDXL-inpainting [36], and generally learns to better adhere to the prompt with high visual fidelity. We believe this observation is consistent with insights from prior works [26, 50] that show that existing diffusion models already have some implicit understanding of multiview images and can be fine-tuned for 3D consistency with little training data. Inpainting, on the other hand, is a more challenging task, and thus, it is easier to fine-tune a model that was trained at scale for an inpainting task (SDXL-inpainting) to create a multiview consistent image. See supplemental material for qualitative examples.", + "bbox": [ + 496, + 482, + 893, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablating Masks. The key ingredient in our training data is the three types of multiview consistent masks. To evaluate the importance of training masks, we ablate on different types of masks, and fine-tune SDXL inpainting model using only one mask type. Each row in Table 2 is trained with a different type of mask: Random 2D follows prior 2D inpainting work [40] sampling mask independently per image (see inset),", + "bbox": [ + 496, + 786, + 893, + 907 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16278", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Type I only uses coarse 3D blobs, Type II uses large surface selections, Type III uses local surface patches (see Section 3.2). Bottom row is our method trained with all types of masks.", + "bbox": [ + 76, + 90, + 333, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e4ef81960025b4d550bd465cef7cdbf601ba96bb9ada7060adeca94d776d6c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 77, + 444, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To compare these different techniques, we split our benchmark into single-mask vs all-masks data, and compute FID score for each test subset (see columns in Table 2). Unsurprisingly, there is a strong bias in the diagonal, indicating that training on a particular type of mask helps inpainting it at inference time. We note, however, that simple random rectangular masks perform substantially worse. Our finetuning on all mask types offers the best performance on arbitrary masks at inference time.", + "bbox": [ + 75, + 169, + 468, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further take 15 user-generated masks and inpaint them to evaluate how well different types of training masks generalize to real use-cases. We found that our multi-mask training offers the highest Clip-similarity of inpainted results to the user prompts, unfortunately FID score is not meaningful for such a small sample size. We refer the reader to supplemental material where we provide qualitative inpainting results for user-generated masks, demonstrating that multi-mask training data yields the highest inpainting quality.", + "bbox": [ + 75, + 306, + 470, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization to Novel Camera Angles. We note that unlike Instant3d [26], the inpainting model adapts its output to the orientation of the non-masked region, indicating an ability to generalize to unseen azimuth angles.", + "bbox": [ + 75, + 474, + 468, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We quantify this by offsetting the 2x2 camera positions by 16 equally spaced azimuth offsets from 0 to 337.5 degrees, generating 500 images for each offset (using the same models as in our benchmark). See how evaluation metrics change for different azimuth offsets in the inset.", + "bbox": [ + 76, + 535, + 276, + 684 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d4dfc53a5315e4834937e8837ad73246a56abc9c3565442952d3fe348f94b0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 535, + 468, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The FID, ClipL, and ClipG scores remain consistent throughout. This adaptability is crucial for inpainting, as the orientation of the unmasked portion is unknown beforehand. Fine-tuning follows [26], with multiview renders and masks at fixed angles for each object. Remarkably, the fine-tuned model generalizes to unseen camera orientations and FOV angles, likely due to the strong inductive bias from the masked image and prior knowledge from the baseline diffusion model. We show some visual examples in the supplementary.", + "bbox": [ + 75, + 686, + 468, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Limitations. Our multiview inpainting might ignore thin masks (a behavior also observed in 2D inpainting models) (Figure 5, left). Additionally, due to training renders having", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c12462c3fd0cb7f764fe6f5ab5a807da10fc15a7ff83dc08ab3abf729b00ef63.jpg", + "image_caption": [ + "Figure 5. Failure Cases. Typical failure cases include failure to adhere to the prompt for thin masks or large masks, which have little inductive bias from the unmasked area." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 890, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "only a white background, we find that in some cases without existing inductive bias for the mask, the network prefers to generate a white background at the expense of aligning with the prompt (Figure 5, right). These cases can usually be resolved by choosing a different random seed.", + "bbox": [ + 498, + 227, + 890, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Applications", + "text_level": 1, + "bbox": [ + 500, + 316, + 630, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While the main contribution of our work is the multiview consistent inpainting method, which can be trivially complemented with a large reconstruction model (LRM) to infer the full 3D shape from the completed views. In this section we explore applications of our technique for editing various neural and traditional 3D representations. We also developed an interactive application where users can load their own 3D shapes, create masks with basic primitives and visualize the results of our method. More information about this application can be found in supplemental.", + "bbox": [ + 498, + 340, + 890, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "NeRF Editing. Given an input NeRF, user's mask and a prompt, we use our multiview inpainting to generate the edited views. These views are used as input to the LRM from Instant3D [26] to create the modified representation. We show a few edits in Figure 2 (our result is on the right). In addition to our method we show results for several alternative techniques that allow editing NeRFs. Vox-E [42] and MVEdit [8] do not allow for a user-provided mask, and instead infer the area to be edited using the attention weights from the prompt, which may result in undesirable changes in texture and geometry throughout the object in areas the user would like to keep unchanged. Progressive3d [10] and NeRFFiller [50] struggle with multiview consistency, as they rely on 2D diffusion models for SDS and IDU, respectively. This may result in incorrectly positioned edits, as in the case of the centaur example. Note that our approach is also significantly faster than all existing timelines and produces results in seconds rather than minutes.", + "bbox": [ + 496, + 507, + 890, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further conduct an informal user preference study between our method and the closest approach that can also inpaint multiview masks, NeRFiller [50]. We showed the masked input and the prompt to 15 users and asked them to choose between two outputs (see supplemental for exact verbiage). Out of 208 pairs, the users preferred our method in 180 cases $(86\\%)$ vs 28 $(14\\%)$ for NeRFiller.", + "bbox": [ + 496, + 779, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since our multiview inpainting is not tied to any par", + "bbox": [ + 517, + 886, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16279", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc35e1c1b29a9775d3be2e4caf18a1d21c62ff6203e93537d3d9362a9150be2b.jpg", + "image_caption": [ + "Figure 6. Application: NeRF editing. We compare our method (rightmost column) to other methods that allow editing NeRFs, by rendering a front-facing view of the 3D edited asset for all methods. Note that the only method that takes arbitrary masks as input is NeRFiller [50], we also conduct an informal user study with 15 users choosing between our output and NeRFiller for 208 pairs of results. We found that users prefer our method in $86\\%$ of the cases. We report average timing for all methods, and confirm Instant3dit is substantially faster than other alternatives." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 460, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/270f1bb8804106b615ae5669f15109f604f6711de10d5fbb9991844bc261cd0b.jpg", + "image_caption": [ + "Figure 7. Application: mesh editing. Each of the 3 edits shown is performed locally on the original mesh, only modifying regions selected by the user. Locality is achieved by running the ROAR [3] geometry optimization, which takes 20 extra seconds per edit." + ], + "image_footnote": [], + "bbox": [ + 76, + 431, + 467, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ticular 3D representation, our method can also be used to edit Gaussian Splats (GS) by simply using an appropriate LRM [58], see the supplementary for GS editing examples.", + "bbox": [ + 75, + 656, + 468, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mesh Editing. For the users that work on traditional mesh representation, we propose using MeshLRM [51] with an adaptive remeshing layer [3] that uses the LRM result as a guidance. This leads to a fast and fully controllable mesh editing pipeline that is guaranteed to preserve the original mesh attributes (e.g., UVs, rigging) and triangulation in the unedited parts of the mesh [4] (Figure 7). To maintain fine details, we use a normal estimator [15] directly on the diffusion output and use the normals as targets in the vertex optimization. More examples in supplemental material.", + "bbox": [ + 75, + 717, + 470, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Texture Editing. The user can use our method to edit sur", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c21ec592231fa40a2b824211c5c1e6a601fda9e6d96a516ac15b4d419913187a.jpg", + "image_caption": [ + "Figure 8. Application: texture editing. Our method can be used to modify texture on a user-selected region. In this case, we run through our NeRF editing pipeline, but only sample colors from the NeRF in the selected region." + ], + "image_footnote": [], + "bbox": [ + 540, + 87, + 851, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "face texture details. We use NeRF-LRM to reconstruct a 3D representation and back-project the colors to the user-selected mesh region. Starting with an auto-generated texture from previous work [39] (Fig. 8, first column), the user labels a mask (second column) to either add new texture elements (facemask, saddle) or fix artifacts in the texture (goldfish exhibiting inconsistent left-right colors due to lack of multiview consistency in the previous texturing method).", + "bbox": [ + 498, + 339, + 893, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 479, + 619, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced Instant3dit, a multiview inpainting diffusion model, and show its application to fast and localized editing of 3D assets. We propose three types of inpainting masks to train the model, corresponding to three types of user edits, and plan to release this dataset. We justify our training strategy and choice of 3d masks by measuring a comprehensive set of metrics: prompt adherence, 3d consistency of the generated content, and generation quality. Instant3dit shows superior quality to the contemporary 3D editing pipelines and is orders of magnitude faster, running as fast as a single image generation using diffusion.", + "bbox": [ + 496, + 506, + 890, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We see clear opportunities to improve this tool further. First, SDXL limits our input resolution to $1024^2$ . We follow in the steps of previous works [26, 31, 52] which found that a $2 \\times 2$ grid of $512^2$ images strikes a good balance between resolution and coverage. An interesting avenue for future work is to explore increasing the number of views (i.e. to $3 \\times 3$ ) in addition to using multiview-aware super-resolution. Additionally, several approaches such as DMD [56] distill diffusion models into one-step or few-step models, which would make Instant3dit run in a fraction of a second and bring a truly interactive editing experience. Second, video diffusion model [5, 45] have shown remarkable 3D consistency and visual quality, which can be harnessed to improve the quality of 3D generation further, although, in terms of speed, they are currently slower than text-to-image models.", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16280", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adobe Inc. Adobe photoshop. 3", + "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. CVPR, 2022. 6", + "[3] Amir Barda, Yotam Erel, Yoni Kasten, and Amit H. Bermano. Roar: Robust adaptive reconstruction of shapes using planar projections. arXiv preprint arXiv:2307.00690, 2023. 4, 8", + "[4] Amir Barda, Vladimir G. Kim, Noam Algerman, Amit H. Bermano, and Thibault Groueix. Magicclay: Sculpting meshes with generative neural fields. SIGGRAPH Asia, 2024. 1, 2, 3, 8", + "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. https://openai.com/research/, 2024.8", + "[6] Chenjie Cao, Chaohui Yu, Yanwei Fu, Fan Wang, and Xi-angyang Xue. Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing. arXiv preprint arXiv:2408.08000, 2024.1,2,3", + "[7] Duygu Ceylan, Valentin Deschaintre, Thibault Groueix, Rosalie Martin, Chun-Hao Huang, Romain Rouffet, Vladimir Kim, and Gaetan Lassagne. Matatlas: Text-driven consistent geometry texturing and material assignment. arXiv preprint arXiv:2404.02899, 2024.3", + "[8] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032, 2024. 1, 2, 7", + "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. ICCV, 2023. 1, 2", + "[10] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. ICLR, 2024. 1, 2, 7", + "[11] Dale Decatur, Itai Lang, Kfir Aberman, and Rana Hanocka. 3d paintbrush: Local stylization of 3d shapes with cascaded score distillation. CVPR, 2024. 2", + "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. NeurIPS, 2023. 2", + "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. CVPR, 2023. 2, 5", + "[14] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. NeurIPS, 2023. 6", + "[15] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowiz-" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ard: Unleashing the diffusion priors for 3d geometry estimation from a single image. ECCV, 2024. 8", + "[16] William Gao, Noam Aigerman, Groueix Thibault, Vladimir Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. SIGGRAPH, 2023. 2", + "[17] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. CVPR, 2023. 2, 3", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2018. 6", + "[19] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. ICLR, 2024. 3", + "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2021. 3", + "[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 2022. 4", + "[22] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Popa Tiberiu. Clip-mesh: Generating textured meshes from text using pretrained image-text models. SIGGRAPH Asia, 2022. 2", + "[23] Hyunwoo Kim, Itai Lang, Noam Aigerman, Thibault Groueix, Vladimir G. Kim, and Rana Hanocka. Meshup: Multi-target mesh deformation via blended distillation. 3DV, 2024. 2", + "[24] Hyunwoo Kim, Itai Lang, Noam Aigerman, Thibault Groueix, Vladimir G. Kim, and Rana Hanocka. Meshup: Multi-target mesh deformation via blended score distillation. 3DV, 2024. 2", + "[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICLR, 2022. 6", + "[26] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. ICLR, 2024. 2, 3, 4, 6, 7, 8", + "[27] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. AAAI, 2024. 2", + "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 5", + "[29] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. ICLR, 2023. 3", + "[30] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. CVPR, 2024. 3" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16281", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Yuanxun Lu, Jingyang Zhang, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, Long Quan, Xun Cao, and Yao Yao. Direct2.5: Diverse text-to-3d generation via multi-view 2.5d diffusion. CVPR, 2024. 2, 3, 4, 8", + "[32] Artem Lukoianov, Haitz Sáez de Ocariz Borde, Kristjan Greenewald, Vitor Campagnolo Guizilini, Timur Bagautdinov, Vincent Sitzmann, and Justin Solomon. Score distillation via reparametrized ddim. NeurIPS, 2024. 3", + "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. CVPR, 2023. 2", + "[34] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. CVPR, 2022. 2", + "[35] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. ICCV, 2023. 2", + "[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. *ICLR*, 2024. 6", + "[37] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. *ICLR*, 2023. 2", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. ICML, 2021. 6", + "[39] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. SIGGRAPH, 2023. 8", + "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2022. 4, 6", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2022. 3", + "[42] Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. Vox-e: Text-guided voxel editing of 3d objects. CVPR, 2023. 1, 2, 7", + "[43] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3", + "[44] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. ICLR, 2024. 3", + "[45] Genmo Team. Mochi 1. GitHub repository, 2024. 8", + "[46] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. 3DV, 2024. 3", + "[47] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 2004. 6", + "[49] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. CoRR, 2024. 3", + "[50] Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Nerfill: Completing scenes via generative 3d inpainting. CVPR, 2024. 1, 2, 3, 6, 7, 8", + "[51] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 3, 4, 8", + "[52] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 2, 3, 4, 8", + "[53] Hanyu Xiang, Qin Zou, Muhammad Ali Nawaz, Xianfeng Huang, Fan Zhang, and Hongkai Yu. Deep learning for image inpainting: A survey. Pattern Recognition, 2023. 3", + "[54] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.2", + "[55] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. CVPR, 2024. 1, 3", + "[56] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. CVPR, 2024. 8", + "[57] Yu Zeng, Zhe L. Lin, Jimei Yang, Jianming Zhang, Eli Shechtman, and Huchuan Lu. High-resolution image inpainting with iterative confidence feedback and guided upsampling. ECCV, 2020. 3, 5", + "[58] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. ECCV, 2024. 2, 3, 4, 8", + "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. CVPR, 2018. 6", + "[60] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. SIGGRAPH Asia, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 811 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16282", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/671f9dcf-f80a-4cc0-9e9d-72da3c00b249_model.json b/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/671f9dcf-f80a-4cc0-9e9d-72da3c00b249_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fe2964e8be4ab263655b97d1f41ddfdfd73fd03f --- /dev/null +++ b/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/671f9dcf-f80a-4cc0-9e9d-72da3c00b249_model.json @@ -0,0 +1,2860 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.021 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.018, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.046 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.06, + 0.812, + 0.083 + ], + "angle": 0, + "content": "Instant3dit: Multiview Inpainting for Fast Editing of 3D Objects" + }, + { + "type": "title", + "bbox": [ + 0.21, + 0.111, + 0.307, + 0.126 + ], + "angle": 0, + "content": "Amir Barda" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.128, + 0.364, + 0.145 + ], + "angle": 0, + "content": "Tel Aviv University, Israel" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.149, + 0.365, + 0.162 + ], + "angle": 0, + "content": "amirbarda@mail.tau.ac.il" + }, + { + "type": "title", + "bbox": [ + 0.181, + 0.171, + 0.316, + 0.188 + ], + "angle": 0, + "content": "Noam Aigerman" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.189, + 0.376, + 0.205 + ], + "angle": 0, + "content": "Université de Montréal, Canada" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.21, + 0.366, + 0.222 + ], + "angle": 0, + "content": "noam.aigerman@umontreal.ca" + }, + { + "type": "title", + "bbox": [ + 0.432, + 0.111, + 0.574, + 0.126 + ], + "angle": 0, + "content": "Matheus Gadelha" + }, + { + "type": "text", + "bbox": [ + 0.412, + 0.128, + 0.593, + 0.145 + ], + "angle": 0, + "content": "Adobe Research, USA" + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.149, + 0.58, + 0.162 + ], + "angle": 0, + "content": "gadelha@adobe.com" + }, + { + "type": "title", + "bbox": [ + 0.448, + 0.171, + 0.593, + 0.186 + ], + "angle": 0, + "content": "Amit H. Bermano" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.189, + 0.625, + 0.206 + ], + "angle": 0, + "content": "Tel Aviv University, Israel" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.21, + 0.623, + 0.222 + ], + "angle": 0, + "content": "amberman@mail.tau.ac.il" + }, + { + "type": "title", + "bbox": [ + 0.662, + 0.11, + 0.798, + 0.126 + ], + "angle": 0, + "content": "Vladimir G. Kim" + }, + { + "type": "text", + "bbox": [ + 0.64, + 0.128, + 0.82, + 0.144 + ], + "angle": 0, + "content": "Adobe Research, USA" + }, + { + "type": "text", + "bbox": [ + 0.661, + 0.149, + 0.797, + 0.162 + ], + "angle": 0, + "content": "vokim@adobe.com" + }, + { + "type": "title", + "bbox": [ + 0.687, + 0.171, + 0.828, + 0.186 + ], + "angle": 0, + "content": "Thibault Groueix" + }, + { + "type": "text", + "bbox": [ + 0.667, + 0.189, + 0.848, + 0.205 + ], + "angle": 0, + "content": "Adobe Research, USA" + }, + { + "type": "text", + "bbox": [ + 0.679, + 0.21, + 0.832, + 0.222 + ], + "angle": 0, + "content": "grouelix@adobe.com" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.246, + 0.207, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.343, + 0.224, + 0.354 + ], + "angle": 0, + "content": "\"A bear holding" + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.253, + 0.297, + 0.264 + ], + "angle": 0, + "content": "Gaussian Splat" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.264, + 0.296, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.343, + 0.307, + 0.354 + ], + "angle": 0, + "content": "a honey pot\"" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.253, + 0.365, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.345, + 0.365, + 0.354 + ], + "angle": 0, + "content": "\\( {a}_{i} \\in a\\;{a}_{i} \\) 分母最小值为 \\( k \\) ." + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.243, + 0.435, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.285, + 0.482, + 0.308 + ], + "angle": 0, + "content": "Original Mesh" + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.308, + 0.506, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.362, + 0.212, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.431, + 0.191, + 0.455 + ], + "angle": 0, + "content": "\"Bear with wings\"" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.364, + 0.299, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.434, + 0.291, + 0.451 + ], + "angle": 0, + "content": "3 sec." + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.362, + 0.374, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.433, + 0.366, + 0.45 + ], + "angle": 0, + "content": "6 sec." + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.362, + 0.454, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.437, + 0.446, + 0.45 + ], + "angle": 0, + "content": "25 sec." + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.246, + 0.584, + 0.268 + ], + "angle": 0, + "content": "Input mesh \\(^+\\) mask" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.272, + 0.58, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.343, + 0.595, + 0.354 + ], + "angle": 0, + "content": "\"Man wearing" + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.254, + 0.648, + 0.264 + ], + "angle": 0, + "content": "NeRF" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.272, + 0.658, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.343, + 0.665, + 0.353 + ], + "angle": 0, + "content": "a medieval" + }, + { + "type": "image_caption", + "bbox": [ + 0.684, + 0.254, + 0.715, + 0.264 + ], + "angle": 0, + "content": "Mesh" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.272, + 0.727, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.343, + 0.718, + 0.353 + ], + "angle": 0, + "content": "helmet\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.244, + 0.797, + 0.268 + ], + "angle": 0, + "content": "Adaptive Remeshing" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.271, + 0.795, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.292, + 0.846, + 0.314 + ], + "angle": 0, + "content": "Original Mesh" + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.316, + 0.852, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.36, + 0.582, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.432, + 0.578, + 0.453 + ], + "angle": 0, + "content": "\"An elven warrior\"" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.36, + 0.659, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.433, + 0.666, + 0.452 + ], + "angle": 0, + "content": "3 sec." + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.361, + 0.731, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.433, + 0.74, + 0.451 + ], + "angle": 0, + "content": "6 sec." + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.36, + 0.797, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.748, + 0.433, + 0.812, + 0.45 + ], + "angle": 0, + "content": "25 sec." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.474, + 0.895, + 0.558 + ], + "angle": 0, + "content": "Figure 1. Our method takes as input a 3D object along with a 3D mask (first column) and a text prompt, and uses our multiview inpainting diffusion model to consistently paint the mask in four rendered views of the object. Off-the-shelf reconstructors can be used on the multiview output to give an NeRF, a Gaussian Splat (second column), or a mesh (third column) that can be used along with adaptive remeshing to ensure the unmasked region is exactly preserved e.g. topology, uvs, (fourth and fifth column). This feedforward approach is orders of magnitude faster than previous works in generative 3D editing, taking just \\(\\approx\\) 3 seconds per multiview edit, then 0.7 seconds to reconstruct a GS or a NeRF, 3 seconds for a mesh, and \\(\\approx\\) 20 seconds for optional mesh post-processing." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.571, + 0.314, + 0.587 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.471, + 0.834 + ], + "angle": 0, + "content": "We propose a generative technique to edit 3D shapes, represented as meshes, NeRFs, or Gaussian Splats, in \\(\\approx 3\\) seconds, without the need for running an SDS type of optimization. Our key insight is to cast 3D editing as a multiview image inpainting problem, as this representation is generic and can be mapped back to any 3D representation using the bank of available Large Reconstruction Models. We explore different fine-tuning strategies to obtain both multiview generation and inpainting capabilities within the same diffusion model. In particular, the design of the inpainting mask is an important factor of training an inpainting model, and we propose several masking strategies to mimic the types of edits a user would perform on a 3D shape. Our approach takes 3D generative editing from hours to seconds and produces higher-quality results compared to previous works." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.851, + 0.21, + 0.866 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Recent years have seen an explosion in AI-guided 3D generative techniques for creation and editing of 3D content, al" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.893, + 0.679 + ], + "angle": 0, + "content": "lowing the user to control the generation process with often as little as a single text prompt [4, 9, 10, 42, 50, 55]. While text prompts provide an easy interface for novice users, they lack fine control over the generation, such as generating a specific object in a specific location over a pre-existing 3D model, e.g., make a bear hold a previously non-existent honey pot (see Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.893, + 0.835 + ], + "angle": 0, + "content": "Thus, several recent works have focused on such localized generation of 3D objects, namely 3D inpainting [6, 42, 50], i.e., filling-in masked-out content in a 3D object, conditioned on a textual description of the desired fill-in, similarly to 2D inpainting [8, 50]. Figure 1 shows several examples of such inpaintings produced by our method, where the mask (made up of coarse 3D primitives) is highlighted in purple color, the text prompt describing the target inpainting is written below, and the resulting inpainted 3D reconstructions are shown next to it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Current 3D inpainting methods cannot be integrated into production pipelines, as they suffer from two fundamental issues: 1) long runtimes, and 2) low quality. We observe these two issues can be explained by previous works' re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16273" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.379 + ], + "angle": 0, + "content": "liacence on optimization of the 3D model by distilling knowledge from a generative model for 2D images via some variant of Score Distillation Sampling (SDS) [37]. Such an optimization process is extremely slow, as it relies on running an image diffusion model over multiple renderings of the 3D object and back-propagating gradients, thus explaining the first issue of slow runtimes. However, we argue that it also lies at the root of the second issue, of degradation in quality: as we show through extensive comparisons (see Figure 6), this class of approaches significantly harms the quality of the inpainting, as SDS-like optimization produces inaccurate and fuzzy results. We attribute it to the fact observed in the original DreamFision paper [37], which stated that \"2D image samples produced using SDS tend to lack diversity\" and \"3D results exhibit few differences across random seeds,\" explaining this behavior with the tendency to seek specific distribution modes regardless of the seed. Thus, it stands to reason it will struggle to inpaint a region of an existing object that is not one of those seeked modes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.473, + 0.593 + ], + "angle": 0, + "content": "In this paper, we propose an alternative approach to 3D inpainting that significantly improves the quality of the inpainted results and the runtime of the process. Drawing inspiration from previous works for fast 3D generation [26, 31, 52, 58], we eliminate both the above issues, by turning the problem on its head: instead of directly optimizing a 3D object using, e.g., SDS, we instead train an image generator to create 2D images of the inpainted 3D object from canonical viewpoints, and then reconstruct the newly-inpainted 3D object in a post-process, via either a feed-forward prediction [26, 54, 58], or lightweight optimization [31, 52]. By doing so we show we avoid both the slow runtimes as well as issues with masking that arise from approaches such as SDS when applied to 3D inpainting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.473, + 0.746 + ], + "angle": 0, + "content": "Adapting this multiview generation technique to the inpainting problem leads to a core challenge: how to ensure that the inpainting of masked region in the different 2D generated image is 3D-consistent from different viewpoints (otherwise, it does not represent a coherent 3D object that can be reconstructed), while at the same time adhering to the 3D mask painted over the original 3D object. Thus, following our observation above, our main technical contribution is designing a scheme to obtain a diffusion model that provides multiview-consistent inpainting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.473, + 0.869 + ], + "angle": 0, + "content": "In order to achieve a multiview-consistent inpainting diffuser, we devise a custom training strategy, along with a novel dataset of 3D masks for 3D inpainting. Namely, we produce multiview consistent masks, avoiding problems resulting from occlusion (that our experiments validate is critical for performance). We design the dataset to support three designated editing modes with different levels of granularity (see Section 3.2)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Finally, we craft a custom training strategy to leverage the priors learned by pretrained text-conditioned image gen" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.895, + 0.183 + ], + "angle": 0, + "content": "erators. We propose to start from a pretrained image inpainter, and use our dataset and training regime to make it multiview consistent. We provide extensive empirical evidence to the specific design choices of our approach, e.g., compare our strategy with the opposite route – fine-tune a multiview diffuser to perform inpainting." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.184, + 0.895, + 0.363 + ], + "angle": 0, + "content": "We show through extensive experiments that our approach enables performing elaborate 3D editing using simple masking of a 3D region, rendering the object and the mask from multiple views, running the inpainting model, and using fast multiview reconstruction techniques to obtain the edited 3D object (see Figure 1). Our approach is thus agnostic to the underlying 3D representation, and we show it supports meshes [54], Gaussian Splats [58] and Radiance Fields [54]. Our experiments validate the speed of our multiview image-based representation and the higher quality of the resulting inpainting compared to previous techniques. To summarize, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.365, + 0.892, + 0.41 + ], + "angle": 0, + "content": "- The first high-quality, fast method for 3D inpainting, enabling fast and localized generation on NeRFs, Gaussian Splats and meshes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.411, + 0.892, + 0.455 + ], + "angle": 0, + "content": "- A 3D masking approach along with a novel training dataset of masks for multiview inpainting, with three types of masks corresponding to different edit modes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.456, + 0.892, + 0.501 + ], + "angle": 0, + "content": "- A finetuning strategy to efficiently leverage pretrained representations, along with an empirical study of alternative training strategies for multiview inpainting." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.365, + 0.892, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.517, + 0.642, + 0.533 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.543, + 0.894, + 0.605 + ], + "angle": 0, + "content": "We position our work against other generative 3D editing approaches and separately discuss multiview 3D generation and inpainting. Closest to our approach are NeRFiller [50] and the concurrent work MVInpainter [6]." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.613, + 0.895, + 0.84 + ], + "angle": 0, + "content": "3D Generative Editing. As we aim for a general tool, we focus the discussion on category-agnostic approaches. The literature can be organized by the core technique used to distill the 2D generative prior - CLIP optimization [16, 22, 34], Score-Distillation Sampling (SDS) [4, 8-11, 23, 24, 27, 33, 35, 42, 60], Iterative Dataset Update (IDU) [17, 50] - or by the type of representation targeted - meshes [4, 16, 22, 23, 34], NeRFs [8, 17, 33, 35, 42, 50], or SDF [9, 10, 27, 60]. All generative editing approaches take advantage of Text-to-Image (T2I) models, whose prowess stems from the billions of text-image pairs available as training data. This abundance of data does not extend to the 3d domain, with the largest public datasets, Objaverse and Objaverse-XL [12, 13], containing millions of meshes, with the vast majority being very low-quality data." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The Score Distillation Sampling (SDS) loss, first defined in [37], use T2I models as guidance for 3d generation, by iteratively denoising renderings. SDS has multiple downsides: the denoised images lack consistency at different" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16274" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.874, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.215, + 0.895, + 0.259 + ], + "angle": 0, + "content": "Figure 2. Overview. Given a NeRF, a Gaussian Splat, or a mesh, the user draws a 3D mask to mark a region to be filled and provides a text prompt to guide the generation. Instant3dit renders four canonical views of the masked object and uses our multiview inpainting network to fill the mask. We use off-the-shelf 3D reconstructors to convert the multiview representation into a NeRF, a Gaussian Splat, or a mesh." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.47, + 0.43 + ], + "angle": 0, + "content": "iterations, and the optimization is therefore very noisy, requiring careful parameter tuning to converge. Second, the optimization typically takes around one hour to converge, as it requires a lot of steps [4]. Combining SDS with Gaussian Splatting partially alleviate this problem by making the rendering operation in each step much faster, as shown in GaussianDreamer [55], but is still not interactive (about 15 minutes per generation). Lastly, SDS produces saturated, low-quality textures due to the high classifier-free guidance used, even though recent variations alleviate this issue [32]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.433, + 0.47, + 0.495 + ], + "angle": 0, + "content": "Iterative Dataset Update, first introduced in [17], is a variant of SDS, taking multiple denoising steps to produce a clean image, used for several optimization steps in a row. NeRFiller [50] extends this approach for editing of NeRFs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.471, + 0.664 + ], + "angle": 0, + "content": "Compared to prior work, our approach does not require an optimization to leverage the generative prior. We propose to perform the edit using a multiview image grid to represent the 3D asset, with a single diffusion inference. As a result, editing is orders of magnitude faster (i.e. from hours to seconds) and more stable (i.e. our approach always robustly produces a result and does not require careful tuning of hyperparameters). Lastly, several off-the-shelf LRM transformers [19] can instantly convert our inpainted image grid to meshes [51], NeRFs [26], or Gaussian Splats [58], making our editing technique representation agnostic." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Multiview generative 3D. To encourage multiview 3D consistency, several works have explored modifying the attention layers with cross-attention blocks to facilitate the exchange of information between views [29, 30, 44]. A simpler approach consists of directly generating a \\(2 \\times 2\\) image grid, without modifying the attention layers. In this case, the exchange of information between view is done via self-attention. This effectively trades resolution with consistency, and can be achieved by conditioning a T2I model on multiview depth [7, 46], or by fine-tuning on a dataset of 3D multiview renderings [26, 43, 47]. Unique3D [52] and direct 2.5d [31] extend this approach to image grids of normals, and CRM [49] to canonical coordinate map. These works have shown that adopting this representation to generate 3D shape is orders of magnitude faster than SDS, from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.895, + 0.416 + ], + "angle": 0, + "content": "hours to seconds. In this work, we extend these ideas to 3D editing, and explore fine-tuning strategies to get inpainting and multiview generation in the same model. Of note, NeRFiller [50] proposes to inpaint a grid of \\(2 \\times 2\\) images with a single-view inpainting model without fine-tuning it for multiview consistency. We compare against this baseline and show the importance of fine-tuning the inpainting model. Furthermore, NeRFiller relies on a more costly IDU optimization, which requires about \\(30 \\mathrm{k}\\) steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.893, + 0.709 + ], + "angle": 0, + "content": "Image inpainting. The ability to collect and annotate images at scale has fueled major progress in image inpainting. Current inpainting tools have a wide range of applications ranging from real photographs to illustrative drawing [1, 41]. We refer the reader to the recent survey of Xiang et al. [53]. In this paper, we are interested in leveraging these pre-trained priors to edit 3D assets. As discussed in the previous paragraph, the main challenge is achieving multiview consistent results, as open-source models like Stable Diffusion [41] inpaint each frame independently without 3D consistency. In this work, we propose to teach multiview consistency to a single-view inpainting network. As noted in Zeng et al. [57], the design of the inpainting masks is a critical part of training an inpainting model. The masks should resemble the type of masks that users will draw at inference as closely as possible. However, how to apply that insight to 3D editing is not straightforward. We propose three masking strategies that address different workflows in Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.895, + 0.832 + ], + "angle": 0, + "content": "Concurrent work. Similarly observing that IDU and SDS are unstable and lengthy optimizations, the recent approach MVInpainter [6] adds a video priors to a single-view inpainting model, via a LoRA [20], to encourage multiview consistency. Their focus in on object insertion and removal in captured 3D scenes, while ours in on 3D generative editing of objects, which leads us to different masking strategies to train the inpainting model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.593, + 0.861 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "The input to our system consists of tuples \\(\\langle S, M, y \\rangle\\) representing a 3D shape \\(S\\), a 3D mask region \\(M\\), and a text de" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16275" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.237, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.108, + 0.326, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.175, + 0.292, + 0.186 + ], + "angle": 0, + "content": "\\(I(S)\\)" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.109, + 0.411, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.175, + 0.383, + 0.186 + ], + "angle": 0, + "content": "\\(I(M)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.213, + 0.47, + 0.256 + ], + "angle": 0, + "content": "Figure 3. Multiview representation. We represent 3D shapes multiview renderings. Editing is done using an image-based diffusion model that operates on \\( I_{c}(S,M) \\) and \\( I_{b}(M,S) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.271, + 0.471, + 0.318 + ], + "angle": 0, + "content": "scription \\( y \\) - henceforth referred to as prompt. The goal of our method is to create a new shape \\( S' \\) whose areas covered by \\( M \\) are modified to follow \\( y \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.47, + 0.462 + ], + "angle": 0, + "content": "Multiview representation. We propose to perform 3D inpainting by inpainting multiple renderings of an object in a view-consistent manner. Consider a rendering operator \\(\\mathcal{R}\\) that renders a set of shapes (a scene) \\(S\\), from a viewpoint \\(\\pi\\). We refer to \\(\\mathcal{R}_c\\) and \\(\\mathcal{R}_b\\) to indicate a rendering to a color RGB image and a binary mask, respectively. Finally, we refer to \\(\\mathcal{R}^U\\) to indicate that we only render the visible pixels belonging to the shape \\(U\\) while assuming \\(U \\in S\\). Using this operator, we define the multiview representations as:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.472, + 0.47, + 0.504 + ], + "angle": 0, + "content": "\\[\nI _ {k} (U, V) := \\bigoplus_ {\\pi \\in \\Pi} \\mathcal {R} _ {k} ^ {U} [ \\{U \\} \\cup \\{V \\}; \\pi ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.713 + ], + "angle": 0, + "content": "where \\(\\bigoplus\\) concatenates the images in a \\(2\\times 2\\) grid, \\(k\\) is rendering modality (color or binary), and \\(\\Pi = \\{C(\\alpha ,\\frac{\\pi}{4})|\\alpha \\in\\) \\(\\{0,\\frac{\\pi}{2},\\pi ,3\\frac{\\pi}{2}\\} \\}\\) . \\(C(\\alpha ,\\beta)\\) is a function that returns a viewpoint configuration corresponding to a camera pointing at the origin of the coordinate system and positioned on the surface of a canonical sphere according to azimuth \\(\\alpha\\) and elevation \\(\\beta\\) . Using this operator, the input to our multiview diffusion model is defined as \\(I_{c}(S,M)\\) and \\(I_{b}(M,S)\\) . Intuitively, \\(I_{c}(S,M)\\) is just an image containing the visible pixels of \\(S\\) in the scene \\(\\{S\\} \\cup \\{M\\}\\) rendered from multiple views organized in a grid, and \\(I_{b}(M,S)\\) is binary rendering of the visible pixels of \\(M\\) . An illustration of this representation is presented in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "3D editing. Given \\( I_{c}(S,M) \\) and \\( I_{b}(M,S) \\), we propose to use a diffusion model \\( \\epsilon_{\\theta} \\) to generate an inpainted multiview representation \\( \\hat{I}_{c} \\). From that, we can then use a posed multiview reconstruction process \\( \\Phi \\) to obtain the edited shape \\( \\hat{S} = \\Phi (\\hat{I}_c) \\). Different choices for \\( \\Phi \\) yield different applications and tradeoffs. For example, we can have very fast learning-based reconstruction from posed multiview images using various representations, like NeRFs [26], meshes [51] and gaussian splats [58]. We can also apply lightweight optimization routines based on differentiable rendering such as ROAR [3], ISOMER [52] or Direct2.5d [31]. This class of methods is usually slightly slower but has the additional" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "benefit of allowing us to carefully craft the optimization procedure in order to achieve several desirable properties like geometric regularization and preservation of the original asset attributes – color, connectivity, UVs, and so on. Figure 1 shows results using various types of reconstructors. We use NeRF-LRM [26] to compare against other NeRF editing approaches in Table 1 and Figure 4. Generating an image with \\(\\epsilon_{\\theta}\\) takes about 3 seconds on an A100, and the various reconstructors range from a few milliseconds [26, 58] to a few seconds [31, 51, 52]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.289 + ], + "angle": 0, + "content": "In the rest of this section, we explain how to train the diffusion model \\(\\epsilon_{\\theta}\\) for multiview generation (Section 3.1) and inpainting (Section 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.3, + 0.795, + 0.317 + ], + "angle": 0, + "content": "3.1. Background on Diffusion Models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.395 + ], + "angle": 0, + "content": "Training. During training, we sample an image \\( x \\) from the dataset, with condition \\( c \\) (e.g. text, mask, or depth), a time step \\( t \\) between 0 and \\( T \\), and a noise \\( \\epsilon \\sim \\mathcal{N}(0,1) \\), injected to \\( x \\) to create a noisy image \\( \\tilde{x}(t) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.41, + 0.891, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\tilde {x} (t) = \\sqrt {\\alpha (t)} \\cdot x + \\sqrt {1 - \\alpha (t)} \\cdot \\epsilon , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.436, + 0.892, + 0.483 + ], + "angle": 0, + "content": "where \\(\\alpha(t)\\) controls the amount of noise to inject i.e. \\(\\alpha(0) = 1\\) is no noise and \\(\\alpha(T) = 0\\) is pure noise. A denoising unet \\(\\epsilon_{\\theta}\\) is trained to denoise \\(\\tilde{x}(t)\\) by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.495, + 0.891, + 0.512 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d i f f}} = w (t) \\left\\| \\epsilon_ {\\theta} (\\tilde {x} (t); t, c) - x \\right\\| ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.892, + 0.601 + ], + "angle": 0, + "content": "where \\( w(t) \\) a scheme to scale the gradients according to \\( t \\). Once \\( \\epsilon_{\\theta} \\) is trained, \\( \\epsilon_{\\theta}(\\tilde{x}(t); t, c) \\) is the projection of \\( \\tilde{x}(t) \\) to the manifold of images defined by the training dataset. Details about the training procedure can be found in the supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.687 + ], + "angle": 0, + "content": "Inference. We start from pure noise \\(\\tilde{x}(T)\\) and follow the direction of the manifold defined by \\(\\epsilon_{\\theta}(\\tilde{x}(t); t, c)\\). There exists multiple samplers to discretize this trajectory into a discrete number of steps. In practice, we use the Euler scheduler [21] and 29 steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.772 + ], + "angle": 0, + "content": "Latent Diffusion. In this work, we use latent diffusion models i.e. the diffusion happens in a 4-dimensional latent space instead of RGB space, and a pretrained VQ-VAE [40] encodes and decodes images from that space. Our method is agnostic to this, so we keep the presentation general." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multiview diffusion. To generate \\(2 \\times 2\\) consistent views, we follow [26] and simply replace the training distribution with a distribution of \\(2 \\times 2\\) images. While this approach would yield poor results if it were trained from scratch, given the scarcity of high-quality 3D data, it performs well if the models are fine-tune from \"foundation\" text-to-image models, i.e. pre-trained on millions of images. To create the dataset, we render a curated list [26] of 5K objects from" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16276" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "Objaverse [13], filtered for high-quality, and generate high-quality captions \\( y_{\\mathrm{n}} \\) for each 3D object with LLaVa [28]. We now explain how to train for inpainting jointly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.146, + 0.471, + 0.298 + ], + "angle": 0, + "content": "Inpainting. The main specificity is that the condition \\( c \\) is composed on the text prompt \\( y \\), the base image with holes, and the inpainting mask i.e. \\( c = \\{y, I_c(S, M), I_b(M, S)\\} \\). In practice, for latent models, \\( I_c(S, M) \\) is passed through the encoder \\( \\mathcal{E} \\) of the VQ-VAE, concatenated with the downsampled version of the mask \\( I_b(M, S) \\), and the noisy latents \\( \\tilde{x}(t) \\), leading to a 9-channel tensor input to the denoising unet \\( \\epsilon_{\\theta} \\), along with the encoding of the text condition. During training, we randomly drop the mask \\( 10\\% \\) of the time, falling back to multiview diffusion training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.306, + 0.342, + 0.323 + ], + "angle": 0, + "content": "3.2. Multi-View Inpainting Masks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.47, + 0.509 + ], + "angle": 0, + "content": "Dataset creation. A key part of our method consists in generating multiview masks that are 3D consistent. Thus, the binary masks used for training our multiview inpainting model are obtained by rendering 3D shapes; i.e. \\( I_{b}(M,S) \\). Consider the example in Figure 3. Even though \\( M \\) is simply an ellipsoid, its multiview representation \\( I_{b}(M,S) \\) is not - it has the occlusions from its interaction with \\( S \\). We empirically demonstrate that 3D-aware masks and multiview consistent images are crucial for better performance (see Table 2). Our model is trained based on a set of shapes \\( \\mathcal{D} \\). For every shape \\( S \\in \\mathcal{D} \\), we create a set of 3D masks \\( \\mathcal{M}_S \\). Using those, we can define our training dataset \\( \\mathcal{I} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.518, + 0.47, + 0.552 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} := \\left\\{ \\begin{array}{c} \\left\\langle I _ {c} (S, \\varnothing), I _ {c} (S, M), I _ {b} (M, S), y _ {S} \\right\\rangle \\\\ \\text {w h e r e} S \\in \\mathcal {D}, M \\in \\mathcal {M} _ {S} \\end{array} \\right\\} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.47, + 0.698 + ], + "angle": 0, + "content": "where \\( I_{c}(S, \\varnothing), I_{c}(S, M), I_{b}(M, S) \\) are the color ground-truth image, color input image, and binary input mask, respectively. \\( y_{S} \\) is a text prompt describing the shape \\( S \\) obtained from a VLM model [28]. Since generating the masks online would slow down training, we preprocess them offline. In practice, we have \\( |\\mathcal{M}_S| = 30 \\), containing equal portions of 3 different kinds of masks. Considering that \\( |\\mathcal{D}| \\approx 5K \\), our multiview dataset \\( \\mathcal{I} \\) has \\( \\approx 150K \\) data points. We will release our dataset upon publication." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.47, + 0.849 + ], + "angle": 0, + "content": "A key remaining issue is how \\(\\mathcal{M}_S\\) is created. As noted in [57], the design of the training masks plays a central role in training an inpainting model. The best performances are naturally obtained when the distribution of training masks closely follows the distribution of edits that users will make at test time. We confirm this in our experiments with a simple ablation, where we naively use random 2D mask (see Table 2). We thus propose three types of masks, corresponding to three types of editing. The remaining of this section will explain how each type of mask is created." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.269, + 0.903 + ], + "angle": 0, + "content": "Type I: coarse edit. In this setting, the inpainted part of \\( S \\) is fully contained" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.853, + 0.47, + 0.913 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.693, + 0.107 + ], + "angle": 0, + "content": "inside \\(M\\). \\(M\\) is computed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.303 + ], + "angle": 0, + "content": "by randomly sampling a part of \\( S \\) and taking its convex hull. To select such part, we randomly sample a plane passing through \\( S \\), effectively splitting \\( S \\) into two parts, and we select one part at random. More precisely, we start by sampling a random point \\( p \\) inside the bounding box of \\( S \\) and a random direction \\( n \\). The plane \\( P \\) passing through \\( p \\) with normal \\( n \\) is defined by \\( \\{x \\in \\mathbb{R}^3 | x \\cdot p = p \\cdot n\\} \\). \\( M \\) is defined as the convex hull of all the face midpoints that are above \\( P \\) i.e. \\( \\{f = (v_1, v_2, v_3) | f \\in F, \\frac{v_1 + v_2 + v_3}{3} \\cdot p >= p \\cdot n\\} \\), where \\( F \\) denotes the list of faces. To avoid Z-fighting during rendering between \\( S \\) and \\( M \\), we scale \\( M \\) by 20% while keeping its center of mass the same, ensuring it completely envelopes the part of \\( S \\) above \\( P \\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.313, + 0.691, + 0.329 + ], + "angle": 0, + "content": "Type II: mesh sculpting." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.329, + 0.691, + 0.373 + ], + "angle": 0, + "content": "This type of mask is designed to represent a more precise edit where the user" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.314, + 0.892, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.892, + 0.513 + ], + "angle": 0, + "content": "expects content to be created in a portion of space similar to the mask. It requires more expertise and time from the user than Type I, but also provides more precise control over the generated content. As we can see in the illustration to the right, \\( M \\) is a tight fit over \\( S - \\) no volume inside \\( M \\) is not also inside \\( S \\). To generate these masks, we sample a plane \\( P \\), as in Type I masks, and select all the faces that have their midpoint above \\( P \\) i.e. \\( M = \\{f = (v_{1}, v_{2}, v_{3}) | f \\in F, \\frac{v_{1} + v_{2} + v_{3}}{3} \\cdot p > = p \\cdot n\\} \\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.52, + 0.691, + 0.535 + ], + "angle": 0, + "content": "Type III: surface editing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.535, + 0.691, + 0.58 + ], + "angle": 0, + "content": "We aim to support local texture modifications, where the user selects a surface" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.518, + 0.892, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.791 + ], + "angle": 0, + "content": "patch and prompts Instant3dit to modify its texture. There are various ways to generate this type of mask if we assume that the meshes have good triangulation. Unfortunately, this is typically not the case for shape datasets. Thus, propose a simple heuristic that works for any triangle soup. We sample a vertex \\( p \\) in \\( S \\), then several cylinders with elliptical bases of varying sizes, all centered on \\( p \\) (depicted in purple). The number of cylinders is uniformly sampled between 3 and 6, the revolution axis is sampled on the unit sphere, the height and radii are sampled between 0.1 and 0.3. We call this volume \\( C \\). Finally select all the faces whose midpoint falls within \\( C \\) i.e. \\( M = \\{f = (v_{1}, v_{2}, v_{3}) | f \\in F, \\frac{v_{1} + v_{2} + v_{3}}{3} \\in C\\} \\) (depicted in red)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.634, + 0.824 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.829, + 0.892, + 0.873 + ], + "angle": 0, + "content": "In this section, we describe our novel multiview inpainting benchmark, which we use to compare several baselines and run ablations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Benchmark. We hold out 500 2x2 multiview images and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16277" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.47, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.201, + 0.471, + 0.245 + ], + "angle": 0, + "content": "Figure 4. Comparison to baselines. We show different inpainting results from different baselines; our multiview inpainting method offers the highest quality while maintaining consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.471, + 0.337 + ], + "angle": 0, + "content": "their synthetic multiview masks from the training set to create our benchmark. We condition all inpainting on BLIP [25] captions, and ensure no training object is present in the benchmark. We then evaluate how well different methods inpaint all the views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.473, + 0.576 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate the quality of multiview inpainting, we use three types of metrics: measuring prompt adherence, multiview consistency, and visual quality. To measure prompt adherence, we use CLIP [38] similarity score between the generated 2x2 multiview image and the text prompt. We follow previous work [36] and use two models, CLIP-ViT-L-14 (ClipL) and CLIP-ViT-BigG-14 (ClipG), for encoding. To measure 3D consistency we reconstruct NeRF from the sparse inpainted views [26], and re-render from same camera angles as the input sparse views. We then compare the NeRF renderings to generated inpainted images using various image similarity scores: SSIM [48], LPIPS [59] and DreamSim [14]. Finally, we measure visual quality using FID score [18] comparing to a distribution of held-out images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.471, + 0.86 + ], + "angle": 0, + "content": "Comparison to Baselines. We use these metrics to compare our method (Table 1, bottom) to several baseline alternatives (Table 1, top). As there are no off-the-shelf multiview consistent inpainting methods, we couple different image diffusion techniques with blended diffusion [2] to inpaint the multiview image. We try two different architectures from the commonly used SDXL diffusion model [36], SDXL (as-is) and SDXL-inpainting (fine-tuned on 2D inpainting task). Both SDXL baselines perform poorly on multiview consistency since they are not explicitly trained to complete images consistently and, in the case of SDXL-inpainting, are only trained with 2D image masks. Finally we test Instant3D [26] coupled with blended diffusion. Since this architecture is trained for 3D reconstruction it achieves a better multiview consistency, however, our method still outperforms this baseline with respect to all other metrics. See Figure 4 for some example results produced with baselines and our method." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Ablating Diffusion Backbones. Our method can be fine-tuned from any image diffusion model, and thus we" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.088, + 0.892, + 0.206 + ], + "angle": 0, + "content": "
MethodPrompt AdherenceMulti-view consistencyVisual quality
ClipL↑ClipG↑SSIM↑LPIPS↓DreamSim↓FID↓
SDXL [36]28.6342.580.8740.0650.134127.0
SDXL-inpainting [36]27.5741.330.8570.0640.137159.5
Instant3D [26]28.7843.060.8920.0440.102120.0
ablations from different diffusion backbones (our final method is at the bottom)
Ours (SD 1.5-inpainting)27.7941.330.7190.090.599128.3
Ours (SD 2.0-inpainting)27.5941.230.7290.0960.589124.5
Ours (Instant3D)28.5742.500.8940.0430.097121.1
Ours (SDXL-inpainting)29.0143.480.8940.0450.100118.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.216, + 0.894, + 0.288 + ], + "angle": 0, + "content": "Table 1. Multiview text-to-image inpainting. This table demonstrates quantitative performance of several baselines and ablations of our method, with respect to metrics that capture prompt adherence, multiview consistency, and visual quality. Results are color-coded worst and best, with best highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.309, + 0.868, + 0.404 + ], + "angle": 0, + "content": "
MethodType IType IIType IIII+II+IIIUser Generated
FID↓ClipG↑
Random 2D145.4129.6102.3131.124.22
Type I131.2128.6102.3121.326.24
Type II170.9122.6101.7128.225.53
Type III193.0148.199.05142.224.2
I+II+III130.0124.9100.0118.426.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.413, + 0.893, + 0.456 + ], + "angle": 0, + "content": "Table 2. Mask ablation. We ablate the choice of 3D masks used during training (rows), and evaluate on different subsets of the benchmark containing only some types of masks (columns)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.895, + 0.771 + ], + "angle": 0, + "content": "further evaluate how performance changes with respect to different backbones (Table 1, bottom section). Unsurprisingly, weaker backbones with fewer parameters (SD 1.5, SD 2.0) [40] lead to inferior performance and very poor multiview consistency. Using Instant3D [26] gives the best multiview consistency, but fine-tuning does not seem to improve visual quality and prompt adherence. Our method fine-tunes from SDXL-inpainting [36], and generally learns to better adhere to the prompt with high visual fidelity. We believe this observation is consistent with insights from prior works [26, 50] that show that existing diffusion models already have some implicit understanding of multiview images and can be fine-tuned for 3D consistency with little training data. Inpainting, on the other hand, is a more challenging task, and thus, it is easier to fine-tune a model that was trained at scale for an inpainting task (SDXL-inpainting) to create a multiview consistent image. See supplemental material for qualitative examples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.787, + 0.895, + 0.909 + ], + "angle": 0, + "content": "Ablating Masks. The key ingredient in our training data is the three types of multiview consistent masks. To evaluate the importance of training masks, we ablate on different types of masks, and fine-tune SDXL inpainting model using only one mask type. Each row in Table 2 is trained with a different type of mask: Random 2D follows prior 2D inpainting work [40] sampling mask independently per image (see inset)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16278" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.334, + 0.167 + ], + "angle": 0, + "content": "Type I only uses coarse 3D blobs, Type II uses large surface selections, Type III uses local surface patches (see Section 3.2). Bottom row is our method trained with all types of masks." + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.078, + 0.445, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.47, + 0.304 + ], + "angle": 0, + "content": "To compare these different techniques, we split our benchmark into single-mask vs all-masks data, and compute FID score for each test subset (see columns in Table 2). Unsurprisingly, there is a strong bias in the diagonal, indicating that training on a particular type of mask helps inpainting it at inference time. We note, however, that simple random rectangular masks perform substantially worse. Our finetuning on all mask types offers the best performance on arbitrary masks at inference time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.471, + 0.459 + ], + "angle": 0, + "content": "We further take 15 user-generated masks and inpaint them to evaluate how well different types of training masks generalize to real use-cases. We found that our multi-mask training offers the highest Clip-similarity of inpainted results to the user prompts, unfortunately FID score is not meaningful for such a small sample size. We refer the reader to supplemental material where we provide qualitative inpainting results for user-generated masks, demonstrating that multi-mask training data yields the highest inpainting quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.535 + ], + "angle": 0, + "content": "Generalization to Novel Camera Angles. We note that unlike Instant3d [26], the inpainting model adapts its output to the orientation of the non-masked region, indicating an ability to generalize to unseen azimuth angles." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.536, + 0.277, + 0.685 + ], + "angle": 0, + "content": "We quantify this by offsetting the 2x2 camera positions by 16 equally spaced azimuth offsets from 0 to 337.5 degrees, generating 500 images for each offset (using the same models as in our benchmark). See how evaluation metrics change for different azimuth offsets in the inset." + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.536, + 0.47, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.84 + ], + "angle": 0, + "content": "The FID, ClipL, and ClipG scores remain consistent throughout. This adaptability is crucial for inpainting, as the orientation of the unmasked portion is unknown beforehand. Fine-tuning follows [26], with multiview renders and masks at fixed angles for each object. Remarkably, the fine-tuned model generalizes to unseen camera orientations and FOV angles, likely due to the strong inductive bias from the masked image and prior knowledge from the baseline diffusion model. We show some visual examples in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Limitations. Our multiview inpainting might ignore thin masks (a behavior also observed in 2D inpainting models) (Figure 5, left). Additionally, due to training renders having" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.892, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.18, + 0.892, + 0.221 + ], + "angle": 0, + "content": "Figure 5. Failure Cases. Typical failure cases include failure to adhere to the prompt for thin masks or large masks, which have little inductive bias from the unmasked area." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.228, + 0.892, + 0.304 + ], + "angle": 0, + "content": "only a white background, we find that in some cases without existing inductive bias for the mask, the network prefers to generate a white background at the expense of aligning with the prompt (Figure 5, right). These cases can usually be resolved by choosing a different random seed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.317, + 0.631, + 0.334 + ], + "angle": 0, + "content": "5. Applications" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.342, + 0.892, + 0.493 + ], + "angle": 0, + "content": "While the main contribution of our work is the multiview consistent inpainting method, which can be trivially complemented with a large reconstruction model (LRM) to infer the full 3D shape from the completed views. In this section we explore applications of our technique for editing various neural and traditional 3D representations. We also developed an interactive application where users can load their own 3D shapes, create masks with basic primitives and visualize the results of our method. More information about this application can be found in supplemental." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.78 + ], + "angle": 0, + "content": "NeRF Editing. Given an input NeRF, user's mask and a prompt, we use our multiview inpainting to generate the edited views. These views are used as input to the LRM from Instant3D [26] to create the modified representation. We show a few edits in Figure 2 (our result is on the right). In addition to our method we show results for several alternative techniques that allow editing NeRFs. Vox-E [42] and MVEdit [8] do not allow for a user-provided mask, and instead infer the area to be edited using the attention weights from the prompt, which may result in undesirable changes in texture and geometry throughout the object in areas the user would like to keep unchanged. Progressive3d [10] and NeRFFiller [50] struggle with multiview consistency, as they rely on 2D diffusion models for SDS and IDU, respectively. This may result in incorrectly positioned edits, as in the case of the centaur example. Note that our approach is also significantly faster than all existing timelines and produces results in seconds rather than minutes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.885 + ], + "angle": 0, + "content": "We further conduct an informal user preference study between our method and the closest approach that can also inpaint multiview masks, NeRFiller [50]. We showed the masked input and the prompt to 15 users and asked them to choose between two outputs (see supplemental for exact verbiage). Out of 208 pairs, the users preferred our method in 180 cases \\((86\\%)\\) vs 28 \\((14\\%)\\) for NeRFiller." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.887, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Since our multiview inpainting is not tied to any par" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16279" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.087, + 0.462, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.293, + 0.474, + 0.419 + ], + "angle": 0, + "content": "Figure 6. Application: NeRF editing. We compare our method (rightmost column) to other methods that allow editing NeRFs, by rendering a front-facing view of the 3D edited asset for all methods. Note that the only method that takes arbitrary masks as input is NeRFiller [50], we also conduct an informal user study with 15 users choosing between our output and NeRFiller for 208 pairs of results. We found that users prefer our method in \\(86\\%\\) of the cases. We report average timing for all methods, and confirm Instant3dit is substantially faster than other alternatives." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.433, + 0.468, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.643 + ], + "angle": 0, + "content": "Figure 7. Application: mesh editing. Each of the 3 edits shown is performed locally on the original mesh, only modifying regions selected by the user. Locality is achieved by running the ROAR [3] geometry optimization, which takes 20 extra seconds per edit." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.703 + ], + "angle": 0, + "content": "ticular 3D representation, our method can also be used to edit Gaussian Splats (GS) by simply using an appropriate LRM [58], see the supplementary for GS editing examples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.471, + 0.871 + ], + "angle": 0, + "content": "Mesh Editing. For the users that work on traditional mesh representation, we propose using MeshLRM [51] with an adaptive remeshing layer [3] that uses the LRM result as a guidance. This leads to a fast and fully controllable mesh editing pipeline that is guaranteed to preserve the original mesh attributes (e.g., UVs, rigging) and triangulation in the unedited parts of the mesh [4] (Figure 7). To maintain fine details, we use a normal estimator [15] directly on the diffusion output and use the normals as targets in the vertex optimization. More examples in supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Texture Editing. The user can use our method to edit sur" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.088, + 0.852, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.266, + 0.895, + 0.323 + ], + "angle": 0, + "content": "Figure 8. Application: texture editing. Our method can be used to modify texture on a user-selected region. In this case, we run through our NeRF editing pipeline, but only sample colors from the NeRF in the selected region." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.34, + 0.895, + 0.462 + ], + "angle": 0, + "content": "face texture details. We use NeRF-LRM to reconstruct a 3D representation and back-project the colors to the user-selected mesh region. Starting with an auto-generated texture from previous work [39] (Fig. 8, first column), the user labels a mask (second column) to either add new texture elements (facemask, saddle) or fix artifacts in the texture (goldfish exhibiting inconsistent left-right colors due to lack of multiview consistency in the previous texturing method)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.48, + 0.62, + 0.495 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.674 + ], + "angle": 0, + "content": "We introduced Instant3dit, a multiview inpainting diffusion model, and show its application to fast and localized editing of 3D assets. We propose three types of inpainting masks to train the model, corresponding to three types of user edits, and plan to release this dataset. We justify our training strategy and choice of 3d masks by measuring a comprehensive set of metrics: prompt adherence, 3d consistency of the generated content, and generation quality. Instant3dit shows superior quality to the contemporary 3D editing pipelines and is orders of magnitude faster, running as fast as a single image generation using diffusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We see clear opportunities to improve this tool further. First, SDXL limits our input resolution to \\(1024^2\\). We follow in the steps of previous works [26, 31, 52] which found that a \\(2 \\times 2\\) grid of \\(512^2\\) images strikes a good balance between resolution and coverage. An interesting avenue for future work is to explore increasing the number of views (i.e. to \\(3 \\times 3\\)) in addition to using multiview-aware super-resolution. Additionally, several approaches such as DMD [56] distill diffusion models into one-step or few-step models, which would make Instant3dit run in a fraction of a second and bring a truly interactive editing experience. Second, video diffusion model [5, 45] have shown remarkable 3D consistency and visual quality, which can be harnessed to improve the quality of 3D generation further, although, in terms of speed, they are currently slower than text-to-image models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16280" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.305, + 0.129 + ], + "angle": 0, + "content": "[1] Adobe Inc. Adobe photoshop. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.131, + 0.469, + 0.171 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.174, + 0.469, + 0.228 + ], + "angle": 0, + "content": "[3] Amir Barda, Yotam Erel, Yoni Kasten, and Amit H. Bermano. Roar: Robust adaptive reconstruction of shapes using planar projections. arXiv preprint arXiv:2307.00690, 2023. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.232, + 0.469, + 0.286 + ], + "angle": 0, + "content": "[4] Amir Barda, Vladimir G. Kim, Noam Algerman, Amit H. Bermano, and Thibault Groueix. Magicclay: Sculpting meshes with generative neural fields. SIGGRAPH Asia, 2024. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. https://openai.com/research/, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.469, + 0.4 + ], + "angle": 0, + "content": "[6] Chenjie Cao, Chaohui Yu, Yanwei Fu, Fan Wang, and Xi-angyang Xue. Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing. arXiv preprint arXiv:2408.08000, 2024.1,2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.469, + 0.471 + ], + "angle": 0, + "content": "[7] Duygu Ceylan, Valentin Deschaintre, Thibault Groueix, Rosalie Martin, Chun-Hao Huang, Romain Rouffet, Vladimir Kim, and Gaetan Lassagne. Matatlas: Text-driven consistent geometry texturing and material assignment. arXiv preprint arXiv:2404.02899, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[8] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[10] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. ICLR, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[11] Dale Decatur, Itai Lang, Kfir Aberman, and Rana Hanocka. 3d paintbrush: Local stylization of 3d shapes with cascaded score distillation. CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. CVPR, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[14] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. NeurIPS, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowiz-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ard: Unleashing the diffusion priors for 3d geometry estimation from a single image. ECCV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[16] William Gao, Noam Aigerman, Groueix Thibault, Vladimir Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. SIGGRAPH, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[17] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[19] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.407, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[22] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Popa Tiberiu. Clip-mesh: Generating textured meshes from text using pretrained image-text models. SIGGRAPH Asia, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[23] Hyunwoo Kim, Itai Lang, Noam Aigerman, Thibault Groueix, Vladimir G. Kim, and Rana Hanocka. Meshup: Multi-target mesh deformation via blended distillation. 3DV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[24] Hyunwoo Kim, Itai Lang, Noam Aigerman, Thibault Groueix, Vladimir G. Kim, and Rana Hanocka. Meshup: Multi-target mesh deformation via blended score distillation. 3DV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[26] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. ICLR, 2024. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[27] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[29] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[30] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. CVPR, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16281" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] Yuanxun Lu, Jingyang Zhang, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, Long Quan, Xun Cao, and Yao Yao. Direct2.5: Diverse text-to-3d generation via multi-view 2.5d diffusion. CVPR, 2024. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[32] Artem Lukoianov, Haitz Sáez de Ocariz Borde, Kristjan Greenewald, Vitor Campagnolo Guizilini, Timur Bagautdinov, Vincent Sitzmann, and Justin Solomon. Score distillation via reparametrized ddim. NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[34] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[35] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. *ICLR*, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[37] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. *ICLR*, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.545 + ], + "angle": 0, + "content": "[39] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. SIGGRAPH, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[42] Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. Vox-e: Text-guided voxel editing of 3d objects. CVPR, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[43] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[44] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.424, + 0.802 + ], + "angle": 0, + "content": "[45] Genmo Team. Mochi 1. GitHub repository, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[46] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. 3DV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[47] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[48] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[49] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. CoRR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[50] Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Nerfill: Completing scenes via generative 3d inpainting. CVPR, 2024. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[51] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[52] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[53] Hanyu Xiang, Qin Zou, Muhammad Ali Nawaz, Xianfeng Huang, Fan Zhang, and Hongkai Yu. Deep learning for image inpainting: A survey. Pattern Recognition, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.486 + ], + "angle": 0, + "content": "[54] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[55] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. CVPR, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.559, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[56] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.616, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[57] Yu Zeng, Zhe L. Lin, Jimei Yang, Jianming Zhang, Eli Shechtman, and Huchuan Lu. High-resolution image inpainting with iterative confidence feedback and guided upsampling. ECCV, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[58] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. ECCV, 2024. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[60] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. SIGGRAPH Asia, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16282" + } + ] +] \ No newline at end of file diff --git a/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/671f9dcf-f80a-4cc0-9e9d-72da3c00b249_origin.pdf b/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/671f9dcf-f80a-4cc0-9e9d-72da3c00b249_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fdeffe8658bf16a6f99f1c2745c36eee5c4b2c6f --- /dev/null +++ b/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/671f9dcf-f80a-4cc0-9e9d-72da3c00b249_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4669e6af67176c14a484129b7b1909addc0db5a43cb1be3188a112305cb6b1ec +size 3113483 diff --git a/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/full.md b/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7ed1f1b88be8663806d70d377eb617205267a2da --- /dev/null +++ b/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/full.md @@ -0,0 +1,395 @@ +# Instant3dit: Multiview Inpainting for Fast Editing of 3D Objects + +# Amir Barda + +Tel Aviv University, Israel + +amirbarda@mail.tau.ac.il + +# Noam Aigerman + +Université de Montréal, Canada + +noam.aigerman@umontreal.ca + +# Matheus Gadelha + +Adobe Research, USA + +gadelha@adobe.com + +# Amit H. Bermano + +Tel Aviv University, Israel + +amberman@mail.tau.ac.il + +# Vladimir G. Kim + +Adobe Research, USA + +vokim@adobe.com + +# Thibault Groueix + +Adobe Research, USA + +grouelix@adobe.com + +![](images/79fe105ef7f700b9cc59e95bb49270ebf78853dbc93c5115b641fe8a874a30a2.jpg) +"A bear holding + +![](images/573220bdcc1e193dac273bf085b351ecf0925f79c9993c3c473de19cb0aab423.jpg) +Gaussian Splat +a honey pot" + +![](images/3e8893b8f51d4c62d1103167b473eef4d5fdd284a56c5502ab3dab40cbd945a4.jpg) +${a}_{i} \in a\;{a}_{i}$ 分母最小值为 $k$ . + +![](images/8166ce8aa3678494c82b7abe09509ff13aaf6f6226417d54cbe271d41d4e19b1.jpg) +Original Mesh + +![](images/aa68e11ac21493e12a835275d69d2f2240cc443dae875f17d71275535d7d0d74.jpg) + +![](images/0cd4098025325c8447cfd88f9767fb4eec25b0ab51ef50b386aee663f4bd9c85.jpg) +"Bear with wings" + +![](images/7283166adb068aa4322e2940d9f1f1459d8aef7b230a3c57018e27d82e2d378f.jpg) +3 sec. + +![](images/86146188ed847b276b674be09883f54afc2f63dfac6932aedb148c0450e0ac88.jpg) +6 sec. + +![](images/6833eb127da64f346283d66a86ea40f273c5f2273fdb501e30c76d622ee1688b.jpg) +25 sec. +Figure 1. Our method takes as input a 3D object along with a 3D mask (first column) and a text prompt, and uses our multiview inpainting diffusion model to consistently paint the mask in four rendered views of the object. Off-the-shelf reconstructors can be used on the multiview output to give an NeRF, a Gaussian Splat (second column), or a mesh (third column) that can be used along with adaptive remeshing to ensure the unmasked region is exactly preserved e.g. topology, uvs, (fourth and fifth column). This feedforward approach is orders of magnitude faster than previous works in generative 3D editing, taking just $\approx$ 3 seconds per multiview edit, then 0.7 seconds to reconstruct a GS or a NeRF, 3 seconds for a mesh, and $\approx$ 20 seconds for optional mesh post-processing. + +![](images/6e3c073f6479ad717196dd72a625a85c59f5bc636234bdfc93d7c0047ce77ed4.jpg) +Input mesh $^+$ mask +"Man wearing + +![](images/f9dbba604eb1f012e8c8e51b8023335f002cc44f35f012cfce1d11369f1d4b6d.jpg) +NeRF +a medieval + +![](images/98a0dc3427157cc12496097e2c71f9df2bf1a37e816943067548ccd4f12c7771.jpg) +Mesh +helmet" + +![](images/d33af5dc58501f133fd22e8fc69c3f544ed82f13e0b07c8e030478f384e9128f.jpg) +Adaptive Remeshing +Original Mesh + +![](images/9a33730c73c7d788ded54a7717ed274888d07e540f9fc5eeeb305de90b2bfd01.jpg) + +![](images/1ccbdc4bb131a513338bb237c9ee25eaf6ada74274fb4028fd50d3d3015f637d.jpg) +"An elven warrior" + +![](images/5390b3477b5e65ffdcb04a3bf3d0ddee8a8e7809ed868552402e0c4af66a3d00.jpg) +3 sec. + +![](images/fb7b766dfa6acb755c9f5c869ead7f66dce696fa76710231f9bc638cafe9161b.jpg) +6 sec. + +![](images/44a256c8841cd2daf8ede1930ecb7c0b69843be9836553a987c11e6b0d870922.jpg) +25 sec. + +# Abstract + +We propose a generative technique to edit 3D shapes, represented as meshes, NeRFs, or Gaussian Splats, in $\approx 3$ seconds, without the need for running an SDS type of optimization. Our key insight is to cast 3D editing as a multiview image inpainting problem, as this representation is generic and can be mapped back to any 3D representation using the bank of available Large Reconstruction Models. We explore different fine-tuning strategies to obtain both multiview generation and inpainting capabilities within the same diffusion model. In particular, the design of the inpainting mask is an important factor of training an inpainting model, and we propose several masking strategies to mimic the types of edits a user would perform on a 3D shape. Our approach takes 3D generative editing from hours to seconds and produces higher-quality results compared to previous works. + +# 1. Introduction + +Recent years have seen an explosion in AI-guided 3D generative techniques for creation and editing of 3D content, al + +lowing the user to control the generation process with often as little as a single text prompt [4, 9, 10, 42, 50, 55]. While text prompts provide an easy interface for novice users, they lack fine control over the generation, such as generating a specific object in a specific location over a pre-existing 3D model, e.g., make a bear hold a previously non-existent honey pot (see Figure 1). + +Thus, several recent works have focused on such localized generation of 3D objects, namely 3D inpainting [6, 42, 50], i.e., filling-in masked-out content in a 3D object, conditioned on a textual description of the desired fill-in, similarly to 2D inpainting [8, 50]. Figure 1 shows several examples of such inpaintings produced by our method, where the mask (made up of coarse 3D primitives) is highlighted in purple color, the text prompt describing the target inpainting is written below, and the resulting inpainted 3D reconstructions are shown next to it. + +Current 3D inpainting methods cannot be integrated into production pipelines, as they suffer from two fundamental issues: 1) long runtimes, and 2) low quality. We observe these two issues can be explained by previous works' re + +liacence on optimization of the 3D model by distilling knowledge from a generative model for 2D images via some variant of Score Distillation Sampling (SDS) [37]. Such an optimization process is extremely slow, as it relies on running an image diffusion model over multiple renderings of the 3D object and back-propagating gradients, thus explaining the first issue of slow runtimes. However, we argue that it also lies at the root of the second issue, of degradation in quality: as we show through extensive comparisons (see Figure 6), this class of approaches significantly harms the quality of the inpainting, as SDS-like optimization produces inaccurate and fuzzy results. We attribute it to the fact observed in the original DreamFision paper [37], which stated that "2D image samples produced using SDS tend to lack diversity" and "3D results exhibit few differences across random seeds," explaining this behavior with the tendency to seek specific distribution modes regardless of the seed. Thus, it stands to reason it will struggle to inpaint a region of an existing object that is not one of those seeked modes. + +In this paper, we propose an alternative approach to 3D inpainting that significantly improves the quality of the inpainted results and the runtime of the process. Drawing inspiration from previous works for fast 3D generation [26, 31, 52, 58], we eliminate both the above issues, by turning the problem on its head: instead of directly optimizing a 3D object using, e.g., SDS, we instead train an image generator to create 2D images of the inpainted 3D object from canonical viewpoints, and then reconstruct the newly-inpainted 3D object in a post-process, via either a feed-forward prediction [26, 54, 58], or lightweight optimization [31, 52]. By doing so we show we avoid both the slow runtimes as well as issues with masking that arise from approaches such as SDS when applied to 3D inpainting. + +Adapting this multiview generation technique to the inpainting problem leads to a core challenge: how to ensure that the inpainting of masked region in the different 2D generated image is 3D-consistent from different viewpoints (otherwise, it does not represent a coherent 3D object that can be reconstructed), while at the same time adhering to the 3D mask painted over the original 3D object. Thus, following our observation above, our main technical contribution is designing a scheme to obtain a diffusion model that provides multiview-consistent inpainting. + +In order to achieve a multiview-consistent inpainting diffuser, we devise a custom training strategy, along with a novel dataset of 3D masks for 3D inpainting. Namely, we produce multiview consistent masks, avoiding problems resulting from occlusion (that our experiments validate is critical for performance). We design the dataset to support three designated editing modes with different levels of granularity (see Section 3.2) + +Finally, we craft a custom training strategy to leverage the priors learned by pretrained text-conditioned image gen + +erators. We propose to start from a pretrained image inpainter, and use our dataset and training regime to make it multiview consistent. We provide extensive empirical evidence to the specific design choices of our approach, e.g., compare our strategy with the opposite route – fine-tune a multiview diffuser to perform inpainting. + +We show through extensive experiments that our approach enables performing elaborate 3D editing using simple masking of a 3D region, rendering the object and the mask from multiple views, running the inpainting model, and using fast multiview reconstruction techniques to obtain the edited 3D object (see Figure 1). Our approach is thus agnostic to the underlying 3D representation, and we show it supports meshes [54], Gaussian Splats [58] and Radiance Fields [54]. Our experiments validate the speed of our multiview image-based representation and the higher quality of the resulting inpainting compared to previous techniques. To summarize, our contributions are: + +- The first high-quality, fast method for 3D inpainting, enabling fast and localized generation on NeRFs, Gaussian Splats and meshes. +- A 3D masking approach along with a novel training dataset of masks for multiview inpainting, with three types of masks corresponding to different edit modes. +- A finetuning strategy to efficiently leverage pretrained representations, along with an empirical study of alternative training strategies for multiview inpainting. + +# 2. Related Work + +We position our work against other generative 3D editing approaches and separately discuss multiview 3D generation and inpainting. Closest to our approach are NeRFiller [50] and the concurrent work MVInpainter [6]. + +3D Generative Editing. As we aim for a general tool, we focus the discussion on category-agnostic approaches. The literature can be organized by the core technique used to distill the 2D generative prior - CLIP optimization [16, 22, 34], Score-Distillation Sampling (SDS) [4, 8-11, 23, 24, 27, 33, 35, 42, 60], Iterative Dataset Update (IDU) [17, 50] - or by the type of representation targeted - meshes [4, 16, 22, 23, 34], NeRFs [8, 17, 33, 35, 42, 50], or SDF [9, 10, 27, 60]. All generative editing approaches take advantage of Text-to-Image (T2I) models, whose prowess stems from the billions of text-image pairs available as training data. This abundance of data does not extend to the 3d domain, with the largest public datasets, Objaverse and Objaverse-XL [12, 13], containing millions of meshes, with the vast majority being very low-quality data. + +The Score Distillation Sampling (SDS) loss, first defined in [37], use T2I models as guidance for 3d generation, by iteratively denoising renderings. SDS has multiple downsides: the denoised images lack consistency at different + +![](images/7e2b0d1b6a5a6335c1600677baf045201fad3dc255011e892b8434eeb034a23d.jpg) +Figure 2. Overview. Given a NeRF, a Gaussian Splat, or a mesh, the user draws a 3D mask to mark a region to be filled and provides a text prompt to guide the generation. Instant3dit renders four canonical views of the masked object and uses our multiview inpainting network to fill the mask. We use off-the-shelf 3D reconstructors to convert the multiview representation into a NeRF, a Gaussian Splat, or a mesh. + +iterations, and the optimization is therefore very noisy, requiring careful parameter tuning to converge. Second, the optimization typically takes around one hour to converge, as it requires a lot of steps [4]. Combining SDS with Gaussian Splatting partially alleviate this problem by making the rendering operation in each step much faster, as shown in GaussianDreamer [55], but is still not interactive (about 15 minutes per generation). Lastly, SDS produces saturated, low-quality textures due to the high classifier-free guidance used, even though recent variations alleviate this issue [32]. + +Iterative Dataset Update, first introduced in [17], is a variant of SDS, taking multiple denoising steps to produce a clean image, used for several optimization steps in a row. NeRFiller [50] extends this approach for editing of NeRFs. + +Compared to prior work, our approach does not require an optimization to leverage the generative prior. We propose to perform the edit using a multiview image grid to represent the 3D asset, with a single diffusion inference. As a result, editing is orders of magnitude faster (i.e. from hours to seconds) and more stable (i.e. our approach always robustly produces a result and does not require careful tuning of hyperparameters). Lastly, several off-the-shelf LRM transformers [19] can instantly convert our inpainted image grid to meshes [51], NeRFs [26], or Gaussian Splats [58], making our editing technique representation agnostic. + +Multiview generative 3D. To encourage multiview 3D consistency, several works have explored modifying the attention layers with cross-attention blocks to facilitate the exchange of information between views [29, 30, 44]. A simpler approach consists of directly generating a $2 \times 2$ image grid, without modifying the attention layers. In this case, the exchange of information between view is done via self-attention. This effectively trades resolution with consistency, and can be achieved by conditioning a T2I model on multiview depth [7, 46], or by fine-tuning on a dataset of 3D multiview renderings [26, 43, 47]. Unique3D [52] and direct 2.5d [31] extend this approach to image grids of normals, and CRM [49] to canonical coordinate map. These works have shown that adopting this representation to generate 3D shape is orders of magnitude faster than SDS, from + +hours to seconds. In this work, we extend these ideas to 3D editing, and explore fine-tuning strategies to get inpainting and multiview generation in the same model. Of note, NeRFiller [50] proposes to inpaint a grid of $2 \times 2$ images with a single-view inpainting model without fine-tuning it for multiview consistency. We compare against this baseline and show the importance of fine-tuning the inpainting model. Furthermore, NeRFiller relies on a more costly IDU optimization, which requires about $30 \mathrm{k}$ steps. + +Image inpainting. The ability to collect and annotate images at scale has fueled major progress in image inpainting. Current inpainting tools have a wide range of applications ranging from real photographs to illustrative drawing [1, 41]. We refer the reader to the recent survey of Xiang et al. [53]. In this paper, we are interested in leveraging these pre-trained priors to edit 3D assets. As discussed in the previous paragraph, the main challenge is achieving multiview consistent results, as open-source models like Stable Diffusion [41] inpaint each frame independently without 3D consistency. In this work, we propose to teach multiview consistency to a single-view inpainting network. As noted in Zeng et al. [57], the design of the inpainting masks is a critical part of training an inpainting model. The masks should resemble the type of masks that users will draw at inference as closely as possible. However, how to apply that insight to 3D editing is not straightforward. We propose three masking strategies that address different workflows in Section 3.2. + +Concurrent work. Similarly observing that IDU and SDS are unstable and lengthy optimizations, the recent approach MVInpainter [6] adds a video priors to a single-view inpainting model, via a LoRA [20], to encourage multiview consistency. Their focus in on object insertion and removal in captured 3D scenes, while ours in on 3D generative editing of objects, which leads us to different masking strategies to train the inpainting model. + +# 3. Method + +The input to our system consists of tuples $\langle S, M, y \rangle$ representing a 3D shape $S$ , a 3D mask region $M$ , and a text de + +![](images/626ed0dc9f9e7733cbb7793b7115543aca752947efa38c0deab26323e6de0a93.jpg) +Figure 3. Multiview representation. We represent 3D shapes multiview renderings. Editing is done using an image-based diffusion model that operates on $I_{c}(S,M)$ and $I_{b}(M,S)$ . + +![](images/7b22beea185b5ec6467bc039bfe39641ed64f7893b7860268a55ffcceefc7a7d.jpg) +$I(S)$ + +![](images/c761dc83475868ce14d3342c4e1dde46cc4b4e22d345622a9a41e03d468d2c76.jpg) +$I(M)$ + +scription $y$ - henceforth referred to as prompt. The goal of our method is to create a new shape $S'$ whose areas covered by $M$ are modified to follow $y$ . + +Multiview representation. We propose to perform 3D inpainting by inpainting multiple renderings of an object in a view-consistent manner. Consider a rendering operator $\mathcal{R}$ that renders a set of shapes (a scene) $S$ , from a viewpoint $\pi$ . We refer to $\mathcal{R}_c$ and $\mathcal{R}_b$ to indicate a rendering to a color RGB image and a binary mask, respectively. Finally, we refer to $\mathcal{R}^U$ to indicate that we only render the visible pixels belonging to the shape $U$ while assuming $U \in S$ . Using this operator, we define the multiview representations as: + +$$ +I _ {k} (U, V) := \bigoplus_ {\pi \in \Pi} \mathcal {R} _ {k} ^ {U} [ \{U \} \cup \{V \}; \pi ] \tag {1} +$$ + +where $\bigoplus$ concatenates the images in a $2\times 2$ grid, $k$ is rendering modality (color or binary), and $\Pi = \{C(\alpha ,\frac{\pi}{4})|\alpha \in$ $\{0,\frac{\pi}{2},\pi ,3\frac{\pi}{2}\} \}$ . $C(\alpha ,\beta)$ is a function that returns a viewpoint configuration corresponding to a camera pointing at the origin of the coordinate system and positioned on the surface of a canonical sphere according to azimuth $\alpha$ and elevation $\beta$ . Using this operator, the input to our multiview diffusion model is defined as $I_{c}(S,M)$ and $I_{b}(M,S)$ . Intuitively, $I_{c}(S,M)$ is just an image containing the visible pixels of $S$ in the scene $\{S\} \cup \{M\}$ rendered from multiple views organized in a grid, and $I_{b}(M,S)$ is binary rendering of the visible pixels of $M$ . An illustration of this representation is presented in Figure 3. + +3D editing. Given $I_{c}(S,M)$ and $I_{b}(M,S)$ , we propose to use a diffusion model $\epsilon_{\theta}$ to generate an inpainted multiview representation $\hat{I}_{c}$ . From that, we can then use a posed multiview reconstruction process $\Phi$ to obtain the edited shape $\hat{S} = \Phi (\hat{I}_c)$ . Different choices for $\Phi$ yield different applications and tradeoffs. For example, we can have very fast learning-based reconstruction from posed multiview images using various representations, like NeRFs [26], meshes [51] and gaussian splats [58]. We can also apply lightweight optimization routines based on differentiable rendering such as ROAR [3], ISOMER [52] or Direct2.5d [31]. This class of methods is usually slightly slower but has the additional + +benefit of allowing us to carefully craft the optimization procedure in order to achieve several desirable properties like geometric regularization and preservation of the original asset attributes – color, connectivity, UVs, and so on. Figure 1 shows results using various types of reconstructors. We use NeRF-LRM [26] to compare against other NeRF editing approaches in Table 1 and Figure 4. Generating an image with $\epsilon_{\theta}$ takes about 3 seconds on an A100, and the various reconstructors range from a few milliseconds [26, 58] to a few seconds [31, 51, 52]. + +In the rest of this section, we explain how to train the diffusion model $\epsilon_{\theta}$ for multiview generation (Section 3.1) and inpainting (Section 3.2). + +# 3.1. Background on Diffusion Models. + +Training. During training, we sample an image $x$ from the dataset, with condition $c$ (e.g. text, mask, or depth), a time step $t$ between 0 and $T$ , and a noise $\epsilon \sim \mathcal{N}(0,1)$ , injected to $x$ to create a noisy image $\tilde{x}(t)$ : + +$$ +\tilde {x} (t) = \sqrt {\alpha (t)} \cdot x + \sqrt {1 - \alpha (t)} \cdot \epsilon , \tag {2} +$$ + +where $\alpha(t)$ controls the amount of noise to inject i.e. $\alpha(0) = 1$ is no noise and $\alpha(T) = 0$ is pure noise. A denoising unet $\epsilon_{\theta}$ is trained to denoise $\tilde{x}(t)$ by minimizing: + +$$ +L _ {\text {d i f f}} = w (t) \left\| \epsilon_ {\theta} (\tilde {x} (t); t, c) - x \right\| ^ {2}, \tag {3} +$$ + +where $w(t)$ a scheme to scale the gradients according to $t$ . Once $\epsilon_{\theta}$ is trained, $\epsilon_{\theta}(\tilde{x}(t); t, c)$ is the projection of $\tilde{x}(t)$ to the manifold of images defined by the training dataset. Details about the training procedure can be found in the supplemental material. + +Inference. We start from pure noise $\tilde{x}(T)$ and follow the direction of the manifold defined by $\epsilon_{\theta}(\tilde{x}(t); t, c)$ . There exists multiple samplers to discretize this trajectory into a discrete number of steps. In practice, we use the Euler scheduler [21] and 29 steps. + +Latent Diffusion. In this work, we use latent diffusion models i.e. the diffusion happens in a 4-dimensional latent space instead of RGB space, and a pretrained VQ-VAE [40] encodes and decodes images from that space. Our method is agnostic to this, so we keep the presentation general. + +Multiview diffusion. To generate $2 \times 2$ consistent views, we follow [26] and simply replace the training distribution with a distribution of $2 \times 2$ images. While this approach would yield poor results if it were trained from scratch, given the scarcity of high-quality 3D data, it performs well if the models are fine-tune from "foundation" text-to-image models, i.e. pre-trained on millions of images. To create the dataset, we render a curated list [26] of 5K objects from + +Objaverse [13], filtered for high-quality, and generate high-quality captions $y_{\mathrm{n}}$ for each 3D object with LLaVa [28]. We now explain how to train for inpainting jointly. + +Inpainting. The main specificity is that the condition $c$ is composed on the text prompt $y$ , the base image with holes, and the inpainting mask i.e. $c = \{y, I_c(S, M), I_b(M, S)\}$ . In practice, for latent models, $I_c(S, M)$ is passed through the encoder $\mathcal{E}$ of the VQ-VAE, concatenated with the downsampled version of the mask $I_b(M, S)$ , and the noisy latents $\tilde{x}(t)$ , leading to a 9-channel tensor input to the denoising unet $\epsilon_{\theta}$ , along with the encoding of the text condition. During training, we randomly drop the mask $10\%$ of the time, falling back to multiview diffusion training. + +# 3.2. Multi-View Inpainting Masks + +Dataset creation. A key part of our method consists in generating multiview masks that are 3D consistent. Thus, the binary masks used for training our multiview inpainting model are obtained by rendering 3D shapes; i.e. $I_{b}(M,S)$ . Consider the example in Figure 3. Even though $M$ is simply an ellipsoid, its multiview representation $I_{b}(M,S)$ is not - it has the occlusions from its interaction with $S$ . We empirically demonstrate that 3D-aware masks and multiview consistent images are crucial for better performance (see Table 2). Our model is trained based on a set of shapes $\mathcal{D}$ . For every shape $S \in \mathcal{D}$ , we create a set of 3D masks $\mathcal{M}_S$ . Using those, we can define our training dataset $\mathcal{I}$ as: + +$$ +\mathcal {I} := \left\{ \begin{array}{c} \left\langle I _ {c} (S, \varnothing), I _ {c} (S, M), I _ {b} (M, S), y _ {S} \right\rangle \\ \text {w h e r e} S \in \mathcal {D}, M \in \mathcal {M} _ {S} \end{array} \right\} \tag {4} +$$ + +where $I_{c}(S, \varnothing), I_{c}(S, M), I_{b}(M, S)$ are the color ground-truth image, color input image, and binary input mask, respectively. $y_{S}$ is a text prompt describing the shape $S$ obtained from a VLM model [28]. Since generating the masks online would slow down training, we preprocess them offline. In practice, we have $|\mathcal{M}_S| = 30$ , containing equal portions of 3 different kinds of masks. Considering that $|\mathcal{D}| \approx 5K$ , our multiview dataset $\mathcal{I}$ has $\approx 150K$ data points. We will release our dataset upon publication. + +A key remaining issue is how $\mathcal{M}_S$ is created. As noted in [57], the design of the training masks plays a central role in training an inpainting model. The best performances are naturally obtained when the distribution of training masks closely follows the distribution of edits that users will make at test time. We confirm this in our experiments with a simple ablation, where we naively use random 2D mask (see Table 2). We thus propose three types of masks, corresponding to three types of editing. The remaining of this section will explain how each type of mask is created. + +Type I: coarse edit. In this setting, the inpainted part of $S$ is fully contained + +![](images/7a542f470ab3eb8a79488bfb5d39b1288edd350105e276c7e09712689b655007.jpg) + +inside $M$ . $M$ is computed + +by randomly sampling a part of $S$ and taking its convex hull. To select such part, we randomly sample a plane passing through $S$ , effectively splitting $S$ into two parts, and we select one part at random. More precisely, we start by sampling a random point $p$ inside the bounding box of $S$ and a random direction $n$ . The plane $P$ passing through $p$ with normal $n$ is defined by $\{x \in \mathbb{R}^3 | x \cdot p = p \cdot n\}$ . $M$ is defined as the convex hull of all the face midpoints that are above $P$ i.e. $\{f = (v_1, v_2, v_3) | f \in F, \frac{v_1 + v_2 + v_3}{3} \cdot p >= p \cdot n\}$ , where $F$ denotes the list of faces. To avoid Z-fighting during rendering between $S$ and $M$ , we scale $M$ by 20% while keeping its center of mass the same, ensuring it completely envelopes the part of $S$ above $P$ . + +# Type II: mesh sculpting. + +This type of mask is designed to represent a more precise edit where the user + +![](images/af72ce7a09671870435d0637e91efa18a5a4f45a20d6f2e79494298a7ff6d688.jpg) + +expects content to be created in a portion of space similar to the mask. It requires more expertise and time from the user than Type I, but also provides more precise control over the generated content. As we can see in the illustration to the right, $M$ is a tight fit over $S -$ no volume inside $M$ is not also inside $S$ . To generate these masks, we sample a plane $P$ , as in Type I masks, and select all the faces that have their midpoint above $P$ i.e. $M = \{f = (v_{1}, v_{2}, v_{3}) | f \in F, \frac{v_{1} + v_{2} + v_{3}}{3} \cdot p > = p \cdot n\}$ . + +# Type III: surface editing. + +We aim to support local texture modifications, where the user selects a surface + +![](images/949faf04bdd4cb2f8ebfaf692a32b0b3496c49819e31a8ba68937a3f65a07338.jpg) + +patch and prompts Instant3dit to modify its texture. There are various ways to generate this type of mask if we assume that the meshes have good triangulation. Unfortunately, this is typically not the case for shape datasets. Thus, propose a simple heuristic that works for any triangle soup. We sample a vertex $p$ in $S$ , then several cylinders with elliptical bases of varying sizes, all centered on $p$ (depicted in purple). The number of cylinders is uniformly sampled between 3 and 6, the revolution axis is sampled on the unit sphere, the height and radii are sampled between 0.1 and 0.3. We call this volume $C$ . Finally select all the faces whose midpoint falls within $C$ i.e. $M = \{f = (v_{1}, v_{2}, v_{3}) | f \in F, \frac{v_{1} + v_{2} + v_{3}}{3} \in C\}$ (depicted in red). + +# 4. Experiments + +In this section, we describe our novel multiview inpainting benchmark, which we use to compare several baselines and run ablations. + +Benchmark. We hold out 500 2x2 multiview images and + +![](images/169028abaf099140377ae6f8b823c15d0e6bf03b031f1404693c6c0428cc188b.jpg) +Figure 4. Comparison to baselines. We show different inpainting results from different baselines; our multiview inpainting method offers the highest quality while maintaining consistency. + +their synthetic multiview masks from the training set to create our benchmark. We condition all inpainting on BLIP [25] captions, and ensure no training object is present in the benchmark. We then evaluate how well different methods inpaint all the views. + +Evaluation Metrics. To evaluate the quality of multiview inpainting, we use three types of metrics: measuring prompt adherence, multiview consistency, and visual quality. To measure prompt adherence, we use CLIP [38] similarity score between the generated 2x2 multiview image and the text prompt. We follow previous work [36] and use two models, CLIP-ViT-L-14 (ClipL) and CLIP-ViT-BigG-14 (ClipG), for encoding. To measure 3D consistency we reconstruct NeRF from the sparse inpainted views [26], and re-render from same camera angles as the input sparse views. We then compare the NeRF renderings to generated inpainted images using various image similarity scores: SSIM [48], LPIPS [59] and DreamSim [14]. Finally, we measure visual quality using FID score [18] comparing to a distribution of held-out images. + +Comparison to Baselines. We use these metrics to compare our method (Table 1, bottom) to several baseline alternatives (Table 1, top). As there are no off-the-shelf multiview consistent inpainting methods, we couple different image diffusion techniques with blended diffusion [2] to inpaint the multiview image. We try two different architectures from the commonly used SDXL diffusion model [36], SDXL (as-is) and SDXL-inpainting (fine-tuned on 2D inpainting task). Both SDXL baselines perform poorly on multiview consistency since they are not explicitly trained to complete images consistently and, in the case of SDXL-inpainting, are only trained with 2D image masks. Finally we test Instant3D [26] coupled with blended diffusion. Since this architecture is trained for 3D reconstruction it achieves a better multiview consistency, however, our method still outperforms this baseline with respect to all other metrics. See Figure 4 for some example results produced with baselines and our method. + +Ablating Diffusion Backbones. Our method can be fine-tuned from any image diffusion model, and thus we + +
MethodPrompt AdherenceMulti-view consistencyVisual quality
ClipL↑ClipG↑SSIM↑LPIPS↓DreamSim↓FID↓
SDXL [36]28.6342.580.8740.0650.134127.0
SDXL-inpainting [36]27.5741.330.8570.0640.137159.5
Instant3D [26]28.7843.060.8920.0440.102120.0
ablations from different diffusion backbones (our final method is at the bottom)
Ours (SD 1.5-inpainting)27.7941.330.7190.090.599128.3
Ours (SD 2.0-inpainting)27.5941.230.7290.0960.589124.5
Ours (Instant3D)28.5742.500.8940.0430.097121.1
Ours (SDXL-inpainting)29.0143.480.8940.0450.100118.4
+ +Table 1. Multiview text-to-image inpainting. This table demonstrates quantitative performance of several baselines and ablations of our method, with respect to metrics that capture prompt adherence, multiview consistency, and visual quality. Results are color-coded worst and best, with best highlighted in bold. + +
MethodType IType IIType IIII+II+IIIUser Generated
FID↓ClipG↑
Random 2D145.4129.6102.3131.124.22
Type I131.2128.6102.3121.326.24
Type II170.9122.6101.7128.225.53
Type III193.0148.199.05142.224.2
I+II+III130.0124.9100.0118.426.5
+ +Table 2. Mask ablation. We ablate the choice of 3D masks used during training (rows), and evaluate on different subsets of the benchmark containing only some types of masks (columns). + +further evaluate how performance changes with respect to different backbones (Table 1, bottom section). Unsurprisingly, weaker backbones with fewer parameters (SD 1.5, SD 2.0) [40] lead to inferior performance and very poor multiview consistency. Using Instant3D [26] gives the best multiview consistency, but fine-tuning does not seem to improve visual quality and prompt adherence. Our method fine-tunes from SDXL-inpainting [36], and generally learns to better adhere to the prompt with high visual fidelity. We believe this observation is consistent with insights from prior works [26, 50] that show that existing diffusion models already have some implicit understanding of multiview images and can be fine-tuned for 3D consistency with little training data. Inpainting, on the other hand, is a more challenging task, and thus, it is easier to fine-tune a model that was trained at scale for an inpainting task (SDXL-inpainting) to create a multiview consistent image. See supplemental material for qualitative examples. + +Ablating Masks. The key ingredient in our training data is the three types of multiview consistent masks. To evaluate the importance of training masks, we ablate on different types of masks, and fine-tune SDXL inpainting model using only one mask type. Each row in Table 2 is trained with a different type of mask: Random 2D follows prior 2D inpainting work [40] sampling mask independently per image (see inset), + +Type I only uses coarse 3D blobs, Type II uses large surface selections, Type III uses local surface patches (see Section 3.2). Bottom row is our method trained with all types of masks. + +![](images/0e4ef81960025b4d550bd465cef7cdbf601ba96bb9ada7060adeca94d776d6c3.jpg) + +To compare these different techniques, we split our benchmark into single-mask vs all-masks data, and compute FID score for each test subset (see columns in Table 2). Unsurprisingly, there is a strong bias in the diagonal, indicating that training on a particular type of mask helps inpainting it at inference time. We note, however, that simple random rectangular masks perform substantially worse. Our finetuning on all mask types offers the best performance on arbitrary masks at inference time. + +We further take 15 user-generated masks and inpaint them to evaluate how well different types of training masks generalize to real use-cases. We found that our multi-mask training offers the highest Clip-similarity of inpainted results to the user prompts, unfortunately FID score is not meaningful for such a small sample size. We refer the reader to supplemental material where we provide qualitative inpainting results for user-generated masks, demonstrating that multi-mask training data yields the highest inpainting quality. + +Generalization to Novel Camera Angles. We note that unlike Instant3d [26], the inpainting model adapts its output to the orientation of the non-masked region, indicating an ability to generalize to unseen azimuth angles. + +We quantify this by offsetting the 2x2 camera positions by 16 equally spaced azimuth offsets from 0 to 337.5 degrees, generating 500 images for each offset (using the same models as in our benchmark). See how evaluation metrics change for different azimuth offsets in the inset. + +![](images/3d4dfc53a5315e4834937e8837ad73246a56abc9c3565442952d3fe348f94b0a.jpg) + +The FID, ClipL, and ClipG scores remain consistent throughout. This adaptability is crucial for inpainting, as the orientation of the unmasked portion is unknown beforehand. Fine-tuning follows [26], with multiview renders and masks at fixed angles for each object. Remarkably, the fine-tuned model generalizes to unseen camera orientations and FOV angles, likely due to the strong inductive bias from the masked image and prior knowledge from the baseline diffusion model. We show some visual examples in the supplementary. + +Limitations. Our multiview inpainting might ignore thin masks (a behavior also observed in 2D inpainting models) (Figure 5, left). Additionally, due to training renders having + +![](images/c12462c3fd0cb7f764fe6f5ab5a807da10fc15a7ff83dc08ab3abf729b00ef63.jpg) +Figure 5. Failure Cases. Typical failure cases include failure to adhere to the prompt for thin masks or large masks, which have little inductive bias from the unmasked area. + +only a white background, we find that in some cases without existing inductive bias for the mask, the network prefers to generate a white background at the expense of aligning with the prompt (Figure 5, right). These cases can usually be resolved by choosing a different random seed. + +# 5. Applications + +While the main contribution of our work is the multiview consistent inpainting method, which can be trivially complemented with a large reconstruction model (LRM) to infer the full 3D shape from the completed views. In this section we explore applications of our technique for editing various neural and traditional 3D representations. We also developed an interactive application where users can load their own 3D shapes, create masks with basic primitives and visualize the results of our method. More information about this application can be found in supplemental. + +NeRF Editing. Given an input NeRF, user's mask and a prompt, we use our multiview inpainting to generate the edited views. These views are used as input to the LRM from Instant3D [26] to create the modified representation. We show a few edits in Figure 2 (our result is on the right). In addition to our method we show results for several alternative techniques that allow editing NeRFs. Vox-E [42] and MVEdit [8] do not allow for a user-provided mask, and instead infer the area to be edited using the attention weights from the prompt, which may result in undesirable changes in texture and geometry throughout the object in areas the user would like to keep unchanged. Progressive3d [10] and NeRFFiller [50] struggle with multiview consistency, as they rely on 2D diffusion models for SDS and IDU, respectively. This may result in incorrectly positioned edits, as in the case of the centaur example. Note that our approach is also significantly faster than all existing timelines and produces results in seconds rather than minutes. + +We further conduct an informal user preference study between our method and the closest approach that can also inpaint multiview masks, NeRFiller [50]. We showed the masked input and the prompt to 15 users and asked them to choose between two outputs (see supplemental for exact verbiage). Out of 208 pairs, the users preferred our method in 180 cases $(86\%)$ vs 28 $(14\%)$ for NeRFiller. + +Since our multiview inpainting is not tied to any par + +![](images/dc35e1c1b29a9775d3be2e4caf18a1d21c62ff6203e93537d3d9362a9150be2b.jpg) +Figure 6. Application: NeRF editing. We compare our method (rightmost column) to other methods that allow editing NeRFs, by rendering a front-facing view of the 3D edited asset for all methods. Note that the only method that takes arbitrary masks as input is NeRFiller [50], we also conduct an informal user study with 15 users choosing between our output and NeRFiller for 208 pairs of results. We found that users prefer our method in $86\%$ of the cases. We report average timing for all methods, and confirm Instant3dit is substantially faster than other alternatives. + +![](images/270f1bb8804106b615ae5669f15109f604f6711de10d5fbb9991844bc261cd0b.jpg) +Figure 7. Application: mesh editing. Each of the 3 edits shown is performed locally on the original mesh, only modifying regions selected by the user. Locality is achieved by running the ROAR [3] geometry optimization, which takes 20 extra seconds per edit. + +ticular 3D representation, our method can also be used to edit Gaussian Splats (GS) by simply using an appropriate LRM [58], see the supplementary for GS editing examples. + +Mesh Editing. For the users that work on traditional mesh representation, we propose using MeshLRM [51] with an adaptive remeshing layer [3] that uses the LRM result as a guidance. This leads to a fast and fully controllable mesh editing pipeline that is guaranteed to preserve the original mesh attributes (e.g., UVs, rigging) and triangulation in the unedited parts of the mesh [4] (Figure 7). To maintain fine details, we use a normal estimator [15] directly on the diffusion output and use the normals as targets in the vertex optimization. More examples in supplemental material. + +Texture Editing. The user can use our method to edit sur + +![](images/c21ec592231fa40a2b824211c5c1e6a601fda9e6d96a516ac15b4d419913187a.jpg) +Figure 8. Application: texture editing. Our method can be used to modify texture on a user-selected region. In this case, we run through our NeRF editing pipeline, but only sample colors from the NeRF in the selected region. + +face texture details. We use NeRF-LRM to reconstruct a 3D representation and back-project the colors to the user-selected mesh region. Starting with an auto-generated texture from previous work [39] (Fig. 8, first column), the user labels a mask (second column) to either add new texture elements (facemask, saddle) or fix artifacts in the texture (goldfish exhibiting inconsistent left-right colors due to lack of multiview consistency in the previous texturing method). + +# 6. Conclusion + +We introduced Instant3dit, a multiview inpainting diffusion model, and show its application to fast and localized editing of 3D assets. We propose three types of inpainting masks to train the model, corresponding to three types of user edits, and plan to release this dataset. We justify our training strategy and choice of 3d masks by measuring a comprehensive set of metrics: prompt adherence, 3d consistency of the generated content, and generation quality. Instant3dit shows superior quality to the contemporary 3D editing pipelines and is orders of magnitude faster, running as fast as a single image generation using diffusion. + +We see clear opportunities to improve this tool further. First, SDXL limits our input resolution to $1024^2$ . We follow in the steps of previous works [26, 31, 52] which found that a $2 \times 2$ grid of $512^2$ images strikes a good balance between resolution and coverage. An interesting avenue for future work is to explore increasing the number of views (i.e. to $3 \times 3$ ) in addition to using multiview-aware super-resolution. Additionally, several approaches such as DMD [56] distill diffusion models into one-step or few-step models, which would make Instant3dit run in a fraction of a second and bring a truly interactive editing experience. Second, video diffusion model [5, 45] have shown remarkable 3D consistency and visual quality, which can be harnessed to improve the quality of 3D generation further, although, in terms of speed, they are currently slower than text-to-image models. + +# References + +[1] Adobe Inc. Adobe photoshop. 3 +[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. CVPR, 2022. 6 +[3] Amir Barda, Yotam Erel, Yoni Kasten, and Amit H. Bermano. Roar: Robust adaptive reconstruction of shapes using planar projections. arXiv preprint arXiv:2307.00690, 2023. 4, 8 +[4] Amir Barda, Vladimir G. Kim, Noam Algerman, Amit H. Bermano, and Thibault Groueix. Magicclay: Sculpting meshes with generative neural fields. SIGGRAPH Asia, 2024. 1, 2, 3, 8 +[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. https://openai.com/research/, 2024.8 +[6] Chenjie Cao, Chaohui Yu, Yanwei Fu, Fan Wang, and Xi-angyang Xue. Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing. arXiv preprint arXiv:2408.08000, 2024.1,2,3 +[7] Duygu Ceylan, Valentin Deschaintre, Thibault Groueix, Rosalie Martin, Chun-Hao Huang, Romain Rouffet, Vladimir Kim, and Gaetan Lassagne. Matatlas: Text-driven consistent geometry texturing and material assignment. arXiv preprint arXiv:2404.02899, 2024.3 +[8] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032, 2024. 1, 2, 7 +[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. ICCV, 2023. 1, 2 +[10] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. ICLR, 2024. 1, 2, 7 +[11] Dale Decatur, Itai Lang, Kfir Aberman, and Rana Hanocka. 3d paintbrush: Local stylization of 3d shapes with cascaded score distillation. CVPR, 2024. 2 +[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. NeurIPS, 2023. 2 +[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. CVPR, 2023. 2, 5 +[14] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. NeurIPS, 2023. 6 +[15] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowiz- + +ard: Unleashing the diffusion priors for 3d geometry estimation from a single image. ECCV, 2024. 8 +[16] William Gao, Noam Aigerman, Groueix Thibault, Vladimir Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. SIGGRAPH, 2023. 2 +[17] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. CVPR, 2023. 2, 3 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2018. 6 +[19] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. ICLR, 2024. 3 +[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2021. 3 +[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 2022. 4 +[22] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Popa Tiberiu. Clip-mesh: Generating textured meshes from text using pretrained image-text models. SIGGRAPH Asia, 2022. 2 +[23] Hyunwoo Kim, Itai Lang, Noam Aigerman, Thibault Groueix, Vladimir G. Kim, and Rana Hanocka. Meshup: Multi-target mesh deformation via blended distillation. 3DV, 2024. 2 +[24] Hyunwoo Kim, Itai Lang, Noam Aigerman, Thibault Groueix, Vladimir G. Kim, and Rana Hanocka. Meshup: Multi-target mesh deformation via blended score distillation. 3DV, 2024. 2 +[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICLR, 2022. 6 +[26] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. ICLR, 2024. 2, 3, 4, 6, 7, 8 +[27] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. AAAI, 2024. 2 +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 5 +[29] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. ICLR, 2023. 3 +[30] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. CVPR, 2024. 3 + +[31] Yuanxun Lu, Jingyang Zhang, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, Long Quan, Xun Cao, and Yao Yao. Direct2.5: Diverse text-to-3d generation via multi-view 2.5d diffusion. CVPR, 2024. 2, 3, 4, 8 +[32] Artem Lukoianov, Haitz Sáez de Ocariz Borde, Kristjan Greenewald, Vitor Campagnolo Guizilini, Timur Bagautdinov, Vincent Sitzmann, and Justin Solomon. Score distillation via reparametrized ddim. NeurIPS, 2024. 3 +[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. CVPR, 2023. 2 +[34] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. CVPR, 2022. 2 +[35] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. ICCV, 2023. 2 +[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. *ICLR*, 2024. 6 +[37] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. *ICLR*, 2023. 2 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. ICML, 2021. 6 +[39] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. SIGGRAPH, 2023. 8 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2022. 4, 6 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2022. 3 +[42] Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. Vox-e: Text-guided voxel editing of 3d objects. CVPR, 2023. 1, 2, 7 +[43] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3 +[44] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. ICLR, 2024. 3 +[45] Genmo Team. Mochi 1. GitHub repository, 2024. 8 +[46] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. 3DV, 2024. 3 +[47] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. 3 + +[48] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 2004. 6 +[49] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. CoRR, 2024. 3 +[50] Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Nerfill: Completing scenes via generative 3d inpainting. CVPR, 2024. 1, 2, 3, 6, 7, 8 +[51] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 3, 4, 8 +[52] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 2, 3, 4, 8 +[53] Hanyu Xiang, Qin Zou, Muhammad Ali Nawaz, Xianfeng Huang, Fan Zhang, and Hongkai Yu. Deep learning for image inpainting: A survey. Pattern Recognition, 2023. 3 +[54] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.2 +[55] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. CVPR, 2024. 1, 3 +[56] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. CVPR, 2024. 8 +[57] Yu Zeng, Zhe L. Lin, Jimei Yang, Jianming Zhang, Eli Shechtman, and Huchuan Lu. High-resolution image inpainting with iterative confidence feedback and guided upsampling. ECCV, 2020. 3, 5 +[58] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. ECCV, 2024. 2, 3, 4, 8 +[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. CVPR, 2018. 6 +[60] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. SIGGRAPH Asia, 2023. 2 \ No newline at end of file diff --git a/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/images.zip b/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..73cc258cc40b639516835d44fc56a6b1af0c3980 --- /dev/null +++ b/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74739ba96342b2bc9b043affab5997739c803374df89d04c87268569f1aeda7d +size 423918 diff --git a/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/layout.json b/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..82e5eeb4955a11dc9adee2e96a62b81680dfbba8 --- /dev/null +++ b/2025/Instant3dit_ Multiview Inpainting for Fast Editing of 3D Objects/layout.json @@ -0,0 +1,11729 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 47, + 496, + 65 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 47, + 496, + 65 + ], + "spans": [ + { + "bbox": [ + 97, + 47, + 496, + 65 + ], + "type": "text", + "content": "Instant3dit: Multiview Inpainting for Fast Editing of 3D Objects" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 128, + 87, + 187, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 87, + 187, + 99 + ], + "spans": [ + { + "bbox": [ + 128, + 87, + 187, + 99 + ], + "type": "text", + "content": "Amir Barda" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 101, + 222, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 101, + 222, + 114 + ], + "spans": [ + { + "bbox": [ + 93, + 101, + 222, + 114 + ], + "type": "text", + "content": "Tel Aviv University, Israel" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 118, + 223, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 118, + 223, + 128 + ], + "spans": [ + { + "bbox": [ + 93, + 118, + 223, + 128 + ], + "type": "text", + "content": "amirbarda@mail.tau.ac.il" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 135, + 193, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 135, + 193, + 148 + ], + "spans": [ + { + "bbox": [ + 110, + 135, + 193, + 148 + ], + "type": "text", + "content": "Noam Aigerman" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 149, + 230, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 149, + 230, + 162 + ], + "spans": [ + { + "bbox": [ + 75, + 149, + 230, + 162 + ], + "type": "text", + "content": "Université de Montréal, Canada" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 81, + 166, + 223, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 166, + 223, + 175 + ], + "spans": [ + { + "bbox": [ + 81, + 166, + 223, + 175 + ], + "type": "text", + "content": "noam.aigerman@umontreal.ca" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 264, + 87, + 351, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 87, + 351, + 99 + ], + "spans": [ + { + "bbox": [ + 264, + 87, + 351, + 99 + ], + "type": "text", + "content": "Matheus Gadelha" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 252, + 101, + 362, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 101, + 362, + 114 + ], + "spans": [ + { + "bbox": [ + 252, + 101, + 362, + 114 + ], + "type": "text", + "content": "Adobe Research, USA" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 260, + 118, + 354, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 118, + 354, + 128 + ], + "spans": [ + { + "bbox": [ + 260, + 118, + 354, + 128 + ], + "type": "text", + "content": "gadelha@adobe.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 274, + 135, + 362, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 135, + 362, + 147 + ], + "spans": [ + { + "bbox": [ + 274, + 135, + 362, + 147 + ], + "type": "text", + "content": "Amit H. Bermano" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 255, + 149, + 382, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 149, + 382, + 163 + ], + "spans": [ + { + "bbox": [ + 255, + 149, + 382, + 163 + ], + "type": "text", + "content": "Tel Aviv University, Israel" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 256, + 166, + 381, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 166, + 381, + 175 + ], + "spans": [ + { + "bbox": [ + 256, + 166, + 381, + 175 + ], + "type": "text", + "content": "amberman@mail.tau.ac.il" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 405, + 87, + 488, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 87, + 488, + 99 + ], + "spans": [ + { + "bbox": [ + 405, + 87, + 488, + 99 + ], + "type": "text", + "content": "Vladimir G. Kim" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 391, + 101, + 501, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 101, + 501, + 114 + ], + "spans": [ + { + "bbox": [ + 391, + 101, + 501, + 114 + ], + "type": "text", + "content": "Adobe Research, USA" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 404, + 118, + 487, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 118, + 487, + 128 + ], + "spans": [ + { + "bbox": [ + 404, + 118, + 487, + 128 + ], + "type": "text", + "content": "vokim@adobe.com" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 420, + 135, + 506, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 135, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 420, + 135, + 506, + 147 + ], + "type": "text", + "content": "Thibault Groueix" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 408, + 149, + 518, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 149, + 518, + 162 + ], + "spans": [ + { + "bbox": [ + 408, + 149, + 518, + 162 + ], + "type": "text", + "content": "Adobe Research, USA" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 415, + 166, + 509, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 166, + 509, + 175 + ], + "spans": [ + { + "bbox": [ + 415, + 166, + 509, + 175 + ], + "type": "text", + "content": "grouelix@adobe.com" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 78, + 194, + 126, + 270 + ], + "blocks": [ + { + "bbox": [ + 78, + 194, + 126, + 270 + ], + "lines": [ + { + "bbox": [ + 78, + 194, + 126, + 270 + ], + "spans": [ + { + "bbox": [ + 78, + 194, + 126, + 270 + ], + "type": "image", + "image_path": "79fe105ef7f700b9cc59e95bb49270ebf78853dbc93c5115b641fe8a874a30a2.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 271, + 137, + 280 + ], + "lines": [ + { + "bbox": [ + 75, + 271, + 137, + 280 + ], + "spans": [ + { + "bbox": [ + 75, + 271, + 137, + 280 + ], + "type": "text", + "content": "\"A bear holding" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 138, + 209, + 181, + 271 + ], + "blocks": [ + { + "bbox": [ + 133, + 200, + 181, + 209 + ], + "lines": [ + { + "bbox": [ + 133, + 200, + 181, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 200, + 181, + 209 + ], + "type": "text", + "content": "Gaussian Splat" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 209, + 181, + 271 + ], + "lines": [ + { + "bbox": [ + 138, + 209, + 181, + 271 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 181, + 271 + ], + "type": "image", + "image_path": "573220bdcc1e193dac273bf085b351ecf0925f79c9993c3c473de19cb0aab423.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 271, + 187, + 280 + ], + "lines": [ + { + "bbox": [ + 137, + 271, + 187, + 280 + ], + "spans": [ + { + "bbox": [ + 137, + 271, + 187, + 280 + ], + "type": "text", + "content": "a honey pot\"" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 181, + 200, + 223, + 273 + ], + "blocks": [ + { + "bbox": [ + 181, + 200, + 223, + 273 + ], + "lines": [ + { + "bbox": [ + 181, + 200, + 223, + 273 + ], + "spans": [ + { + "bbox": [ + 181, + 200, + 223, + 273 + ], + "type": "image", + "image_path": "3e8893b8f51d4c62d1103167b473eef4d5fdd284a56c5502ab3dab40cbd945a4.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 273, + 223, + 280 + ], + "lines": [ + { + "bbox": [ + 192, + 273, + 223, + 280 + ], + "spans": [ + { + "bbox": [ + 192, + 273, + 223, + 280 + ], + "type": "inline_equation", + "content": "{a}_{i} \\in a\\;{a}_{i}" + }, + { + "bbox": [ + 192, + 273, + 223, + 280 + ], + "type": "text", + "content": " 分母最小值为 " + }, + { + "bbox": [ + 192, + 273, + 223, + 280 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 192, + 273, + 223, + 280 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 223, + 192, + 266, + 272 + ], + "blocks": [ + { + "bbox": [ + 223, + 192, + 266, + 272 + ], + "lines": [ + { + "bbox": [ + 223, + 192, + 266, + 272 + ], + "spans": [ + { + "bbox": [ + 223, + 192, + 266, + 272 + ], + "type": "image", + "image_path": "8166ce8aa3678494c82b7abe09509ff13aaf6f6226417d54cbe271d41d4e19b1.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 225, + 294, + 243 + ], + "lines": [ + { + "bbox": [ + 266, + 225, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 266, + 225, + 294, + 243 + ], + "type": "text", + "content": "Original Mesh" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 265, + 243, + 309, + 307 + ], + "blocks": [ + { + "bbox": [ + 265, + 243, + 309, + 307 + ], + "lines": [ + { + "bbox": [ + 265, + 243, + 309, + 307 + ], + "spans": [ + { + "bbox": [ + 265, + 243, + 309, + 307 + ], + "type": "image", + "image_path": "aa68e11ac21493e12a835275d69d2f2240cc443dae875f17d71275535d7d0d74.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 73, + 286, + 129, + 338 + ], + "blocks": [ + { + "bbox": [ + 73, + 286, + 129, + 338 + ], + "lines": [ + { + "bbox": [ + 73, + 286, + 129, + 338 + ], + "spans": [ + { + "bbox": [ + 73, + 286, + 129, + 338 + ], + "type": "image", + "image_path": "0cd4098025325c8447cfd88f9767fb4eec25b0ab51ef50b386aee663f4bd9c85.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 341, + 116, + 360 + ], + "lines": [ + { + "bbox": [ + 74, + 341, + 116, + 360 + ], + "spans": [ + { + "bbox": [ + 74, + 341, + 116, + 360 + ], + "type": "text", + "content": "\"Bear with wings\"" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 130, + 288, + 182, + 342 + ], + "blocks": [ + { + "bbox": [ + 130, + 288, + 182, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 288, + 182, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 182, + 342 + ], + "type": "image", + "image_path": "7283166adb068aa4322e2940d9f1f1459d8aef7b230a3c57018e27d82e2d378f.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 343, + 178, + 357 + ], + "lines": [ + { + "bbox": [ + 143, + 343, + 178, + 357 + ], + "spans": [ + { + "bbox": [ + 143, + 343, + 178, + 357 + ], + "type": "text", + "content": "3 sec." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 183, + 286, + 228, + 342 + ], + "blocks": [ + { + "bbox": [ + 183, + 286, + 228, + 342 + ], + "lines": [ + { + "bbox": [ + 183, + 286, + 228, + 342 + ], + "spans": [ + { + "bbox": [ + 183, + 286, + 228, + 342 + ], + "type": "image", + "image_path": "86146188ed847b276b674be09883f54afc2f63dfac6932aedb148c0450e0ac88.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 342, + 223, + 356 + ], + "lines": [ + { + "bbox": [ + 188, + 342, + 223, + 356 + ], + "spans": [ + { + "bbox": [ + 188, + 342, + 223, + 356 + ], + "type": "text", + "content": "6 sec." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 229, + 286, + 277, + 346 + ], + "blocks": [ + { + "bbox": [ + 229, + 286, + 277, + 346 + ], + "lines": [ + { + "bbox": [ + 229, + 286, + 277, + 346 + ], + "spans": [ + { + "bbox": [ + 229, + 286, + 277, + 346 + ], + "type": "image", + "image_path": "6833eb127da64f346283d66a86ea40f273c5f2273fdb501e30c76d622ee1688b.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 346, + 272, + 356 + ], + "lines": [ + { + "bbox": [ + 234, + 346, + 272, + 356 + ], + "spans": [ + { + "bbox": [ + 234, + 346, + 272, + 356 + ], + "type": "text", + "content": "25 sec." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 375, + 547, + 441 + ], + "lines": [ + { + "bbox": [ + 46, + 375, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 547, + 441 + ], + "type": "text", + "content": "Figure 1. Our method takes as input a 3D object along with a 3D mask (first column) and a text prompt, and uses our multiview inpainting diffusion model to consistently paint the mask in four rendered views of the object. Off-the-shelf reconstructors can be used on the multiview output to give an NeRF, a Gaussian Splat (second column), or a mesh (third column) that can be used along with adaptive remeshing to ensure the unmasked region is exactly preserved e.g. topology, uvs, (fourth and fifth column). This feedforward approach is orders of magnitude faster than previous works in generative 3D editing, taking just " + }, + { + "bbox": [ + 46, + 375, + 547, + 441 + ], + "type": "inline_equation", + "content": "\\approx" + }, + { + "bbox": [ + 46, + 375, + 547, + 441 + ], + "type": "text", + "content": " 3 seconds per multiview edit, then 0.7 seconds to reconstruct a GS or a NeRF, 3 seconds for a mesh, and " + }, + { + "bbox": [ + 46, + 375, + 547, + 441 + ], + "type": "inline_equation", + "content": "\\approx" + }, + { + "bbox": [ + 46, + 375, + 547, + 441 + ], + "type": "text", + "content": " 20 seconds for optional mesh post-processing." + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 315, + 215, + 354, + 270 + ], + "blocks": [ + { + "bbox": [ + 312, + 194, + 357, + 212 + ], + "lines": [ + { + "bbox": [ + 312, + 194, + 357, + 212 + ], + "spans": [ + { + "bbox": [ + 312, + 194, + 357, + 212 + ], + "type": "text", + "content": "Input mesh " + }, + { + "bbox": [ + 312, + 194, + 357, + 212 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 312, + 194, + 357, + 212 + ], + "type": "text", + "content": " mask" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 215, + 354, + 270 + ], + "lines": [ + { + "bbox": [ + 315, + 215, + 354, + 270 + ], + "spans": [ + { + "bbox": [ + 315, + 215, + 354, + 270 + ], + "type": "image", + "image_path": "6e3c073f6479ad717196dd72a625a85c59f5bc636234bdfc93d7c0047ce77ed4.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 271, + 364, + 280 + ], + "lines": [ + { + "bbox": [ + 315, + 271, + 364, + 280 + ], + "spans": [ + { + "bbox": [ + 315, + 271, + 364, + 280 + ], + "type": "text", + "content": "\"Man wearing" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 366, + 215, + 402, + 270 + ], + "blocks": [ + { + "bbox": [ + 375, + 201, + 396, + 209 + ], + "lines": [ + { + "bbox": [ + 375, + 201, + 396, + 209 + ], + "spans": [ + { + "bbox": [ + 375, + 201, + 396, + 209 + ], + "type": "text", + "content": "NeRF" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 215, + 402, + 270 + ], + "lines": [ + { + "bbox": [ + 366, + 215, + 402, + 270 + ], + "spans": [ + { + "bbox": [ + 366, + 215, + 402, + 270 + ], + "type": "image", + "image_path": "f9dbba604eb1f012e8c8e51b8023335f002cc44f35f012cfce1d11369f1d4b6d.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 271, + 406, + 279 + ], + "lines": [ + { + "bbox": [ + 366, + 271, + 406, + 279 + ], + "spans": [ + { + "bbox": [ + 366, + 271, + 406, + 279 + ], + "type": "text", + "content": "a medieval" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 408, + 215, + 444, + 270 + ], + "blocks": [ + { + "bbox": [ + 418, + 201, + 437, + 209 + ], + "lines": [ + { + "bbox": [ + 418, + 201, + 437, + 209 + ], + "spans": [ + { + "bbox": [ + 418, + 201, + 437, + 209 + ], + "type": "text", + "content": "Mesh" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 408, + 215, + 444, + 270 + ], + "lines": [ + { + "bbox": [ + 408, + 215, + 444, + 270 + ], + "spans": [ + { + "bbox": [ + 408, + 215, + 444, + 270 + ], + "type": "image", + "image_path": "98a0dc3427157cc12496097e2c71f9df2bf1a37e816943067548ccd4f12c7771.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 271, + 439, + 279 + ], + "lines": [ + { + "bbox": [ + 408, + 271, + 439, + 279 + ], + "spans": [ + { + "bbox": [ + 408, + 271, + 439, + 279 + ], + "type": "text", + "content": "helmet\"" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 449, + 214, + 486, + 273 + ], + "blocks": [ + { + "bbox": [ + 453, + 193, + 487, + 212 + ], + "lines": [ + { + "bbox": [ + 453, + 193, + 487, + 212 + ], + "spans": [ + { + "bbox": [ + 453, + 193, + 487, + 212 + ], + "type": "text", + "content": "Adaptive Remeshing" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 214, + 486, + 273 + ], + "lines": [ + { + "bbox": [ + 449, + 214, + 486, + 273 + ], + "spans": [ + { + "bbox": [ + 449, + 214, + 486, + 273 + ], + "type": "image", + "image_path": "d33af5dc58501f133fd22e8fc69c3f544ed82f13e0b07c8e030478f384e9128f.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 231, + 517, + 248 + ], + "lines": [ + { + "bbox": [ + 488, + 231, + 517, + 248 + ], + "spans": [ + { + "bbox": [ + 488, + 231, + 517, + 248 + ], + "type": "text", + "content": "Original Mesh" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 484, + 250, + 521, + 305 + ], + "blocks": [ + { + "bbox": [ + 484, + 250, + 521, + 305 + ], + "lines": [ + { + "bbox": [ + 484, + 250, + 521, + 305 + ], + "spans": [ + { + "bbox": [ + 484, + 250, + 521, + 305 + ], + "type": "image", + "image_path": "9a33730c73c7d788ded54a7717ed274888d07e540f9fc5eeeb305de90b2bfd01.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 316, + 285, + 356, + 340 + ], + "blocks": [ + { + "bbox": [ + 316, + 285, + 356, + 340 + ], + "lines": [ + { + "bbox": [ + 316, + 285, + 356, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 356, + 340 + ], + "type": "image", + "image_path": "1ccbdc4bb131a513338bb237c9ee25eaf6ada74274fb4028fd50d3d3015f637d.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 342, + 353, + 358 + ], + "lines": [ + { + "bbox": [ + 316, + 342, + 353, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 353, + 358 + ], + "type": "text", + "content": "\"An elven warrior\"" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 365, + 285, + 403, + 342 + ], + "blocks": [ + { + "bbox": [ + 365, + 285, + 403, + 342 + ], + "lines": [ + { + "bbox": [ + 365, + 285, + 403, + 342 + ], + "spans": [ + { + "bbox": [ + 365, + 285, + 403, + 342 + ], + "type": "image", + "image_path": "5390b3477b5e65ffdcb04a3bf3d0ddee8a8e7809ed868552402e0c4af66a3d00.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 342, + 407, + 357 + ], + "lines": [ + { + "bbox": [ + 373, + 342, + 407, + 357 + ], + "spans": [ + { + "bbox": [ + 373, + 342, + 407, + 357 + ], + "type": "text", + "content": "3 sec." + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 410, + 285, + 447, + 342 + ], + "blocks": [ + { + "bbox": [ + 410, + 285, + 447, + 342 + ], + "lines": [ + { + "bbox": [ + 410, + 285, + 447, + 342 + ], + "spans": [ + { + "bbox": [ + 410, + 285, + 447, + 342 + ], + "type": "image", + "image_path": "fb7b766dfa6acb755c9f5c869ead7f66dce696fa76710231f9bc638cafe9161b.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 342, + 452, + 357 + ], + "lines": [ + { + "bbox": [ + 417, + 342, + 452, + 357 + ], + "spans": [ + { + "bbox": [ + 417, + 342, + 452, + 357 + ], + "type": "text", + "content": "6 sec." + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 451, + 285, + 487, + 342 + ], + "blocks": [ + { + "bbox": [ + 451, + 285, + 487, + 342 + ], + "lines": [ + { + "bbox": [ + 451, + 285, + 487, + 342 + ], + "spans": [ + { + "bbox": [ + 451, + 285, + 487, + 342 + ], + "type": "image", + "image_path": "44a256c8841cd2daf8ede1930ecb7c0b69843be9836553a987c11e6b0d870922.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 342, + 496, + 356 + ], + "lines": [ + { + "bbox": [ + 457, + 342, + 496, + 356 + ], + "spans": [ + { + "bbox": [ + 457, + 342, + 496, + 356 + ], + "type": "text", + "content": "25 sec." + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + } + ], + "index": 60 + }, + { + "bbox": [ + 143, + 452, + 192, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 452, + 192, + 464 + ], + "spans": [ + { + "bbox": [ + 143, + 452, + 192, + 464 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 46, + 468, + 288, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 288, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 288, + 660 + ], + "type": "text", + "content": "We propose a generative technique to edit 3D shapes, represented as meshes, NeRFs, or Gaussian Splats, in " + }, + { + "bbox": [ + 46, + 468, + 288, + 660 + ], + "type": "inline_equation", + "content": "\\approx 3" + }, + { + "bbox": [ + 46, + 468, + 288, + 660 + ], + "type": "text", + "content": " seconds, without the need for running an SDS type of optimization. Our key insight is to cast 3D editing as a multiview image inpainting problem, as this representation is generic and can be mapped back to any 3D representation using the bank of available Large Reconstruction Models. We explore different fine-tuning strategies to obtain both multiview generation and inpainting capabilities within the same diffusion model. In particular, the design of the inpainting mask is an important factor of training an inpainting model, and we propose several masking strategies to mimic the types of edits a user would perform on a 3D shape. Our approach takes 3D generative editing from hours to seconds and produces higher-quality results compared to previous works." + } + ] + } + ], + "index": 64 + }, + { + "bbox": [ + 47, + 673, + 128, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 128, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 128, + 685 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 65 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Recent years have seen an explosion in AI-guided 3D generative techniques for creation and editing of 3D content, al" + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 304, + 453, + 546, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 546, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 546, + 537 + ], + "type": "text", + "content": "lowing the user to control the generation process with often as little as a single text prompt [4, 9, 10, 42, 50, 55]. While text prompts provide an easy interface for novice users, they lack fine control over the generation, such as generating a specific object in a specific location over a pre-existing 3D model, e.g., make a bear hold a previously non-existent honey pot (see Figure 1)." + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 541, + 546, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 546, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 546, + 661 + ], + "type": "text", + "content": "Thus, several recent works have focused on such localized generation of 3D objects, namely 3D inpainting [6, 42, 50], i.e., filling-in masked-out content in a 3D object, conditioned on a textual description of the desired fill-in, similarly to 2D inpainting [8, 50]. Figure 1 shows several examples of such inpaintings produced by our method, where the mask (made up of coarse 3D primitives) is highlighted in purple color, the text prompt describing the target inpainting is written below, and the resulting inpainted 3D reconstructions are shown next to it." + } + ] + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Current 3D inpainting methods cannot be integrated into production pipelines, as they suffer from two fundamental issues: 1) long runtimes, and 2) low quality. We observe these two issues can be explained by previous works' re" + } + ] + } + ], + "index": 69 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 36 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 36 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 36 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16273" + } + ] + } + ], + "index": 70 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "content": "liacence on optimization of the 3D model by distilling knowledge from a generative model for 2D images via some variant of Score Distillation Sampling (SDS) [37]. Such an optimization process is extremely slow, as it relies on running an image diffusion model over multiple renderings of the 3D object and back-propagating gradients, thus explaining the first issue of slow runtimes. However, we argue that it also lies at the root of the second issue, of degradation in quality: as we show through extensive comparisons (see Figure 6), this class of approaches significantly harms the quality of the inpainting, as SDS-like optimization produces inaccurate and fuzzy results. We attribute it to the fact observed in the original DreamFision paper [37], which stated that \"2D image samples produced using SDS tend to lack diversity\" and \"3D results exhibit few differences across random seeds,\" explaining this behavior with the tendency to seek specific distribution modes regardless of the seed. Thus, it stands to reason it will struggle to inpaint a region of an existing object that is not one of those seeked modes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 301, + 289, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 289, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 289, + 469 + ], + "type": "text", + "content": "In this paper, we propose an alternative approach to 3D inpainting that significantly improves the quality of the inpainted results and the runtime of the process. Drawing inspiration from previous works for fast 3D generation [26, 31, 52, 58], we eliminate both the above issues, by turning the problem on its head: instead of directly optimizing a 3D object using, e.g., SDS, we instead train an image generator to create 2D images of the inpainted 3D object from canonical viewpoints, and then reconstruct the newly-inpainted 3D object in a post-process, via either a feed-forward prediction [26, 54, 58], or lightweight optimization [31, 52]. By doing so we show we avoid both the slow runtimes as well as issues with masking that arise from approaches such as SDS when applied to 3D inpainting." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 471, + 289, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 289, + 590 + ], + "type": "text", + "content": "Adapting this multiview generation technique to the inpainting problem leads to a core challenge: how to ensure that the inpainting of masked region in the different 2D generated image is 3D-consistent from different viewpoints (otherwise, it does not represent a coherent 3D object that can be reconstructed), while at the same time adhering to the 3D mask painted over the original 3D object. Thus, following our observation above, our main technical contribution is designing a scheme to obtain a diffusion model that provides multiview-consistent inpainting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 592, + 289, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 289, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 289, + 688 + ], + "type": "text", + "content": "In order to achieve a multiview-consistent inpainting diffuser, we devise a custom training strategy, along with a novel dataset of 3D masks for 3D inpainting. Namely, we produce multiview consistent masks, avoiding problems resulting from occlusion (that our experiments validate is critical for performance). We design the dataset to support three designated editing modes with different levels of granularity (see Section 3.2)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Finally, we craft a custom training strategy to leverage the priors learned by pretrained text-conditioned image gen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "erators. We propose to start from a pretrained image inpainter, and use our dataset and training regime to make it multiview consistent. We provide extensive empirical evidence to the specific design choices of our approach, e.g., compare our strategy with the opposite route – fine-tune a multiview diffuser to perform inpainting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 145, + 547, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 547, + 287 + ], + "type": "text", + "content": "We show through extensive experiments that our approach enables performing elaborate 3D editing using simple masking of a 3D region, rendering the object and the mask from multiple views, running the inpainting model, and using fast multiview reconstruction techniques to obtain the edited 3D object (see Figure 1). Our approach is thus agnostic to the underlying 3D representation, and we show it supports meshes [54], Gaussian Splats [58] and Radiance Fields [54]. Our experiments validate the speed of our multiview image-based representation and the higher quality of the resulting inpainting compared to previous techniques. To summarize, our contributions are:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 289, + 545, + 396 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 289, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 545, + 324 + ], + "type": "text", + "content": "- The first high-quality, fast method for 3D inpainting, enabling fast and localized generation on NeRFs, Gaussian Splats and meshes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 325, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 325, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 325, + 545, + 360 + ], + "type": "text", + "content": "- A 3D masking approach along with a novel training dataset of masks for multiview inpainting, with three types of masks corresponding to different edit modes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 361, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 545, + 396 + ], + "type": "text", + "content": "- A finetuning strategy to efficiently leverage pretrained representations, along with an empirical study of alternative training strategies for multiview inpainting." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 409, + 392, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 409, + 392, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 409, + 392, + 422 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 430, + 547, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 479 + ], + "type": "text", + "content": "We position our work against other generative 3D editing approaches and separately discuss multiview 3D generation and inpainting. Closest to our approach are NeRFiller [50] and the concurrent work MVInpainter [6]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 665 + ], + "type": "text", + "content": "3D Generative Editing. As we aim for a general tool, we focus the discussion on category-agnostic approaches. The literature can be organized by the core technique used to distill the 2D generative prior - CLIP optimization [16, 22, 34], Score-Distillation Sampling (SDS) [4, 8-11, 23, 24, 27, 33, 35, 42, 60], Iterative Dataset Update (IDU) [17, 50] - or by the type of representation targeted - meshes [4, 16, 22, 23, 34], NeRFs [8, 17, 33, 35, 42, 50], or SDF [9, 10, 27, 60]. All generative editing approaches take advantage of Text-to-Image (T2I) models, whose prowess stems from the billions of text-image pairs available as training data. This abundance of data does not extend to the 3d domain, with the largest public datasets, Objaverse and Objaverse-XL [12, 13], containing millions of meshes, with the vast majority being very low-quality data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "The Score Distillation Sampling (SDS) loss, first defined in [37], use T2I models as guidance for 3d generation, by iteratively denoising renderings. SDS has multiple downsides: the denoised images lack consistency at different" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16274" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 534, + 162 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 534, + 162 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 534, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 534, + 162 + ], + "type": "image", + "image_path": "7e2b0d1b6a5a6335c1600677baf045201fad3dc255011e892b8434eeb034a23d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 170, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 170, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 547, + 205 + ], + "type": "text", + "content": "Figure 2. Overview. Given a NeRF, a Gaussian Splat, or a mesh, the user draws a 3D mask to mark a region to be filled and provides a text prompt to guide the generation. Instant3dit renders four canonical views of the masked object and uses our multiview inpainting network to fill the mask. We use off-the-shelf 3D reconstructors to convert the multiview representation into a NeRF, a Gaussian Splat, or a mesh." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 340 + ], + "type": "text", + "content": "iterations, and the optimization is therefore very noisy, requiring careful parameter tuning to converge. Second, the optimization typically takes around one hour to converge, as it requires a lot of steps [4]. Combining SDS with Gaussian Splatting partially alleviate this problem by making the rendering operation in each step much faster, as shown in GaussianDreamer [55], but is still not interactive (about 15 minutes per generation). Lastly, SDS produces saturated, low-quality textures due to the high classifier-free guidance used, even though recent variations alleviate this issue [32]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 342, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 287, + 392 + ], + "type": "text", + "content": "Iterative Dataset Update, first introduced in [17], is a variant of SDS, taking multiple denoising steps to produce a clean image, used for several optimization steps in a row. NeRFiller [50] extends this approach for editing of NeRFs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 393, + 288, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 288, + 525 + ], + "type": "text", + "content": "Compared to prior work, our approach does not require an optimization to leverage the generative prior. We propose to perform the edit using a multiview image grid to represent the 3D asset, with a single diffusion inference. As a result, editing is orders of magnitude faster (i.e. from hours to seconds) and more stable (i.e. our approach always robustly produces a result and does not require careful tuning of hyperparameters). Lastly, several off-the-shelf LRM transformers [19] can instantly convert our inpainted image grid to meshes [51], NeRFs [26], or Gaussian Splats [58], making our editing technique representation agnostic." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": "Multiview generative 3D. To encourage multiview 3D consistency, several works have explored modifying the attention layers with cross-attention blocks to facilitate the exchange of information between views [29, 30, 44]. A simpler approach consists of directly generating a " + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": " image grid, without modifying the attention layers. In this case, the exchange of information between view is done via self-attention. This effectively trades resolution with consistency, and can be achieved by conditioning a T2I model on multiview depth [7, 46], or by fine-tuning on a dataset of 3D multiview renderings [26, 43, 47]. Unique3D [52] and direct 2.5d [31] extend this approach to image grids of normals, and CRM [49] to canonical coordinate map. These works have shown that adopting this representation to generate 3D shape is orders of magnitude faster than SDS, from" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 220, + 547, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 547, + 329 + ], + "type": "text", + "content": "hours to seconds. In this work, we extend these ideas to 3D editing, and explore fine-tuning strategies to get inpainting and multiview generation in the same model. Of note, NeRFiller [50] proposes to inpaint a grid of " + }, + { + "bbox": [ + 304, + 220, + 547, + 329 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 220, + 547, + 329 + ], + "type": "text", + "content": " images with a single-view inpainting model without fine-tuning it for multiview consistency. We compare against this baseline and show the importance of fine-tuning the inpainting model. Furthermore, NeRFiller relies on a more costly IDU optimization, which requires about " + }, + { + "bbox": [ + 304, + 220, + 547, + 329 + ], + "type": "inline_equation", + "content": "30 \\mathrm{k}" + }, + { + "bbox": [ + 304, + 220, + 547, + 329 + ], + "type": "text", + "content": " steps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 335, + 546, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 546, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 546, + 561 + ], + "type": "text", + "content": "Image inpainting. The ability to collect and annotate images at scale has fueled major progress in image inpainting. Current inpainting tools have a wide range of applications ranging from real photographs to illustrative drawing [1, 41]. We refer the reader to the recent survey of Xiang et al. [53]. In this paper, we are interested in leveraging these pre-trained priors to edit 3D assets. As discussed in the previous paragraph, the main challenge is achieving multiview consistent results, as open-source models like Stable Diffusion [41] inpaint each frame independently without 3D consistency. In this work, we propose to teach multiview consistency to a single-view inpainting network. As noted in Zeng et al. [57], the design of the inpainting masks is a critical part of training an inpainting model. The masks should resemble the type of masks that users will draw at inference as closely as possible. However, how to apply that insight to 3D editing is not straightforward. We propose three masking strategies that address different workflows in Section 3.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 562, + 547, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 547, + 658 + ], + "type": "text", + "content": "Concurrent work. Similarly observing that IDU and SDS are unstable and lengthy optimizations, the recent approach MVInpainter [6] adds a video priors to a single-view inpainting model, via a LoRA [20], to encourage multiview consistency. Their focus in on object insertion and removal in captured 3D scenes, while ours in on 3D generative editing of objects, which leads us to different masking strategies to train the inpainting model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 669, + 362, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 362, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 362, + 681 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": "The input to our system consists of tuples " + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\langle S, M, y \\rangle" + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": " representing a 3D shape " + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": ", a 3D mask region " + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": ", and a text de" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16275" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 145, + 153 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 145, + 153 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 145, + 153 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 145, + 153 + ], + "type": "image", + "image_path": "626ed0dc9f9e7733cbb7793b7115543aca752947efa38c0deab26323e6de0a93.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 168, + 287, + 202 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 287, + 202 + ], + "type": "text", + "content": "Figure 3. Multiview representation. We represent 3D shapes multiview renderings. Editing is done using an image-based diffusion model that operates on " + }, + { + "bbox": [ + 46, + 168, + 287, + 202 + ], + "type": "inline_equation", + "content": "I_{c}(S,M)" + }, + { + "bbox": [ + 46, + 168, + 287, + 202 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 168, + 287, + 202 + ], + "type": "inline_equation", + "content": "I_{b}(M,S)" + }, + { + "bbox": [ + 46, + 168, + 287, + 202 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 146, + 85, + 199, + 137 + ], + "blocks": [ + { + "bbox": [ + 146, + 85, + 199, + 137 + ], + "lines": [ + { + "bbox": [ + 146, + 85, + 199, + 137 + ], + "spans": [ + { + "bbox": [ + 146, + 85, + 199, + 137 + ], + "type": "image", + "image_path": "7b22beea185b5ec6467bc039bfe39641ed64f7893b7860268a55ffcceefc7a7d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 138, + 178, + 147 + ], + "lines": [ + { + "bbox": [ + 165, + 138, + 178, + 147 + ], + "spans": [ + { + "bbox": [ + 165, + 138, + 178, + 147 + ], + "type": "inline_equation", + "content": "I(S)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 200, + 86, + 251, + 137 + ], + "blocks": [ + { + "bbox": [ + 200, + 86, + 251, + 137 + ], + "lines": [ + { + "bbox": [ + 200, + 86, + 251, + 137 + ], + "spans": [ + { + "bbox": [ + 200, + 86, + 251, + 137 + ], + "type": "image", + "image_path": "c761dc83475868ce14d3342c4e1dde46cc4b4e22d345622a9a41e03d468d2c76.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 138, + 234, + 147 + ], + "lines": [ + { + "bbox": [ + 220, + 138, + 234, + 147 + ], + "spans": [ + { + "bbox": [ + 220, + 138, + 234, + 147 + ], + "type": "inline_equation", + "content": "I(M)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 214, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 214, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 288, + 251 + ], + "type": "text", + "content": "scription " + }, + { + "bbox": [ + 46, + 214, + 288, + 251 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 214, + 288, + 251 + ], + "type": "text", + "content": " - henceforth referred to as prompt. The goal of our method is to create a new shape " + }, + { + "bbox": [ + 46, + 214, + 288, + 251 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 46, + 214, + 288, + 251 + ], + "type": "text", + "content": " whose areas covered by " + }, + { + "bbox": [ + 46, + 214, + 288, + 251 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 214, + 288, + 251 + ], + "type": "text", + "content": " are modified to follow " + }, + { + "bbox": [ + 46, + 214, + 288, + 251 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 214, + 288, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": "Multiview representation. We propose to perform 3D inpainting by inpainting multiple renderings of an object in a view-consistent manner. Consider a rendering operator " + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": " that renders a set of shapes (a scene) " + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": ", from a viewpoint " + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": ". We refer to " + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_c" + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_b" + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": " to indicate a rendering to a color RGB image and a binary mask, respectively. Finally, we refer to " + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^U" + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": " to indicate that we only render the visible pixels belonging to the shape " + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": " while assuming " + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "inline_equation", + "content": "U \\in S" + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": ". Using this operator, we define the multiview representations as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 373, + 287, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 373, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 91, + 373, + 287, + 399 + ], + "type": "interline_equation", + "content": "I _ {k} (U, V) := \\bigoplus_ {\\pi \\in \\Pi} \\mathcal {R} _ {k} ^ {U} [ \\{U \\} \\cup \\{V \\}; \\pi ] \\tag {1}", + "image_path": "f6123a3fa4aeaad578975d1f716a016d191a7db1b33e144c28f948d0eaf6d0a9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\bigoplus" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " concatenates the images in a " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "2\\times 2" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " grid, " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " is rendering modality (color or binary), and " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\Pi = \\{C(\\alpha ,\\frac{\\pi}{4})|\\alpha \\in" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\{0,\\frac{\\pi}{2},\\pi ,3\\frac{\\pi}{2}\\} \\}" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "C(\\alpha ,\\beta)" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " is a function that returns a viewpoint configuration corresponding to a camera pointing at the origin of the coordinate system and positioned on the surface of a canonical sphere according to azimuth " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " and elevation " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " . Using this operator, the input to our multiview diffusion model is defined as " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "I_{c}(S,M)" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "I_{b}(M,S)" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " . Intuitively, " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "I_{c}(S,M)" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " is just an image containing the visible pixels of " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " in the scene " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\{S\\} \\cup \\{M\\}" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " rendered from multiple views organized in a grid, and " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "I_{b}(M,S)" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " is binary rendering of the visible pixels of " + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 407, + 287, + 564 + ], + "type": "text", + "content": " . An illustration of this representation is presented in Figure 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "3D editing. Given " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "I_{c}(S,M)" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "I_{b}(M,S)" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", we propose to use a diffusion model " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " to generate an inpainted multiview representation " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{I}_{c}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ". From that, we can then use a posed multiview reconstruction process " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " to obtain the edited shape " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{S} = \\Phi (\\hat{I}_c)" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ". Different choices for " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " yield different applications and tradeoffs. For example, we can have very fast learning-based reconstruction from posed multiview images using various representations, like NeRFs [26], meshes [51] and gaussian splats [58]. We can also apply lightweight optimization routines based on differentiable rendering such as ROAR [3], ISOMER [52] or Direct2.5d [31]. This class of methods is usually slightly slower but has the additional" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "benefit of allowing us to carefully craft the optimization procedure in order to achieve several desirable properties like geometric regularization and preservation of the original asset attributes – color, connectivity, UVs, and so on. Figure 1 shows results using various types of reconstructors. We use NeRF-LRM [26] to compare against other NeRF editing approaches in Table 1 and Figure 4. Generating an image with " + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": " takes about 3 seconds on an A100, and the various reconstructors range from a few milliseconds [26, 58] to a few seconds [31, 51, 52]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 192, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 228 + ], + "type": "text", + "content": "In the rest of this section, we explain how to train the diffusion model " + }, + { + "bbox": [ + 304, + 192, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 192, + 545, + 228 + ], + "type": "text", + "content": " for multiview generation (Section 3.1) and inpainting (Section 3.2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 237, + 486, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 237, + 486, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 237, + 486, + 251 + ], + "type": "text", + "content": "3.1. Background on Diffusion Models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "text", + "content": "Training. During training, we sample an image " + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "text", + "content": " from the dataset, with condition " + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "text", + "content": " (e.g. text, mask, or depth), a time step " + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "text", + "content": " between 0 and " + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "text", + "content": ", and a noise " + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "text", + "content": ", injected to " + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "text", + "content": " to create a noisy image " + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\tilde{x}(t)" + }, + { + "bbox": [ + 304, + 263, + 545, + 312 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 352, + 324, + 545, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 324, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 352, + 324, + 545, + 339 + ], + "type": "interline_equation", + "content": "\\tilde {x} (t) = \\sqrt {\\alpha (t)} \\cdot x + \\sqrt {1 - \\alpha (t)} \\cdot \\epsilon , \\tag {2}", + "image_path": "23aaa08050be8ac5fb0d565f1d8b851d350e823fff9aad45a0a604c7e4e20394.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 345, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 345, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 305, + 345, + 545, + 382 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 345, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\alpha(t)" + }, + { + "bbox": [ + 305, + 345, + 545, + 382 + ], + "type": "text", + "content": " controls the amount of noise to inject i.e. " + }, + { + "bbox": [ + 305, + 345, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\alpha(0) = 1" + }, + { + "bbox": [ + 305, + 345, + 545, + 382 + ], + "type": "text", + "content": " is no noise and " + }, + { + "bbox": [ + 305, + 345, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\alpha(T) = 0" + }, + { + "bbox": [ + 305, + 345, + 545, + 382 + ], + "type": "text", + "content": " is pure noise. A denoising unet " + }, + { + "bbox": [ + 305, + 345, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 305, + 345, + 545, + 382 + ], + "type": "text", + "content": " is trained to denoise " + }, + { + "bbox": [ + 305, + 345, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\tilde{x}(t)" + }, + { + "bbox": [ + 305, + 345, + 545, + 382 + ], + "type": "text", + "content": " by minimizing:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 392, + 545, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 392, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 356, + 392, + 545, + 405 + ], + "type": "interline_equation", + "content": "L _ {\\text {d i f f}} = w (t) \\left\\| \\epsilon_ {\\theta} (\\tilde {x} (t); t, c) - x \\right\\| ^ {2}, \\tag {3}", + "image_path": "c65ac229c0ea01a57598185cfdd38fee39c8ab09c4ec0af7896cfa7bd820ab48.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "text", + "content": " a scheme to scale the gradients according to " + }, + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "text", + "content": ". Once " + }, + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "text", + "content": " is trained, " + }, + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\tilde{x}(t); t, c)" + }, + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "text", + "content": " is the projection of " + }, + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{x}(t)" + }, + { + "bbox": [ + 304, + 415, + 545, + 475 + ], + "type": "text", + "content": " to the manifold of images defined by the training dataset. Details about the training procedure can be found in the supplemental material." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 483, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 544 + ], + "type": "text", + "content": "Inference. We start from pure noise " + }, + { + "bbox": [ + 304, + 483, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\tilde{x}(T)" + }, + { + "bbox": [ + 304, + 483, + 545, + 544 + ], + "type": "text", + "content": " and follow the direction of the manifold defined by " + }, + { + "bbox": [ + 304, + 483, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\tilde{x}(t); t, c)" + }, + { + "bbox": [ + 304, + 483, + 545, + 544 + ], + "type": "text", + "content": ". There exists multiple samplers to discretize this trajectory into a discrete number of steps. In practice, we use the Euler scheduler [21] and 29 steps." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 550, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 611 + ], + "type": "text", + "content": "Latent Diffusion. In this work, we use latent diffusion models i.e. the diffusion happens in a 4-dimensional latent space instead of RGB space, and a pretrained VQ-VAE [40] encodes and decodes images from that space. Our method is agnostic to this, so we keep the presentation general." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Multiview diffusion. To generate " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " consistent views, we follow [26] and simply replace the training distribution with a distribution of " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " images. While this approach would yield poor results if it were trained from scratch, given the scarcity of high-quality 3D data, it performs well if the models are fine-tune from \"foundation\" text-to-image models, i.e. pre-trained on millions of images. To create the dataset, we render a curated list [26] of 5K objects from" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16276" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "Objaverse [13], filtered for high-quality, and generate high-quality captions " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{n}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": " for each 3D object with LLaVa [28]. We now explain how to train for inpainting jointly." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "text", + "content": "Inpainting. The main specificity is that the condition " + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "text", + "content": " is composed on the text prompt " + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "text", + "content": ", the base image with holes, and the inpainting mask i.e. " + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "inline_equation", + "content": "c = \\{y, I_c(S, M), I_b(M, S)\\}" + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "text", + "content": ". In practice, for latent models, " + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "inline_equation", + "content": "I_c(S, M)" + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "text", + "content": " is passed through the encoder " + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "text", + "content": " of the VQ-VAE, concatenated with the downsampled version of the mask " + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "inline_equation", + "content": "I_b(M, S)" + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "text", + "content": ", and the noisy latents " + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "inline_equation", + "content": "\\tilde{x}(t)" + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "text", + "content": ", leading to a 9-channel tensor input to the denoising unet " + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "text", + "content": ", along with the encoding of the text condition. During training, we randomly drop the mask " + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 115, + 288, + 236 + ], + "type": "text", + "content": " of the time, falling back to multiview diffusion training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 242, + 209, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 242, + 209, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 209, + 255 + ], + "type": "text", + "content": "3.2. Multi-View Inpainting Masks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "text", + "content": "Dataset creation. A key part of our method consists in generating multiview masks that are 3D consistent. Thus, the binary masks used for training our multiview inpainting model are obtained by rendering 3D shapes; i.e. " + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "inline_equation", + "content": "I_{b}(M,S)" + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "text", + "content": ". Consider the example in Figure 3. Even though " + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "text", + "content": " is simply an ellipsoid, its multiview representation " + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "inline_equation", + "content": "I_{b}(M,S)" + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "text", + "content": " is not - it has the occlusions from its interaction with " + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "text", + "content": ". We empirically demonstrate that 3D-aware masks and multiview consistent images are crucial for better performance (see Table 2). Our model is trained based on a set of shapes " + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "text", + "content": ". For every shape " + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "inline_equation", + "content": "S \\in \\mathcal{D}" + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "text", + "content": ", we create a set of 3D masks " + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_S" + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "text", + "content": ". Using those, we can define our training dataset " + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 258, + 287, + 403 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 410, + 287, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 410, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 63, + 410, + 287, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {I} := \\left\\{ \\begin{array}{c} \\left\\langle I _ {c} (S, \\varnothing), I _ {c} (S, M), I _ {b} (M, S), y _ {S} \\right\\rangle \\\\ \\text {w h e r e} S \\in \\mathcal {D}, M \\in \\mathcal {M} _ {S} \\end{array} \\right\\} \\tag {4}", + "image_path": "89d7acecfd5320df2543f0d6869aa096ad42a51f8bc8310d2dcec12d0971c9e0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "inline_equation", + "content": "I_{c}(S, \\varnothing), I_{c}(S, M), I_{b}(M, S)" + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "text", + "content": " are the color ground-truth image, color input image, and binary input mask, respectively. " + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "inline_equation", + "content": "y_{S}" + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "text", + "content": " is a text prompt describing the shape " + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "text", + "content": " obtained from a VLM model [28]. Since generating the masks online would slow down training, we preprocess them offline. In practice, we have " + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "inline_equation", + "content": "|\\mathcal{M}_S| = 30" + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "text", + "content": ", containing equal portions of 3 different kinds of masks. Considering that " + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}| \\approx 5K" + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "text", + "content": ", our multiview dataset " + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\approx 150K" + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "text", + "content": " data points. We will release our dataset upon publication." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "type": "text", + "content": "A key remaining issue is how " + }, + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_S" + }, + { + "bbox": [ + 46, + 552, + 287, + 672 + ], + "type": "text", + "content": " is created. As noted in [57], the design of the training masks plays a central role in training an inpainting model. The best performances are naturally obtained when the distribution of training masks closely follows the distribution of edits that users will make at test time. We confirm this in our experiments with a simple ablation, where we naively use random 2D mask (see Table 2). We thus propose three types of masks, corresponding to three types of editing. The remaining of this section will explain how each type of mask is created." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 164, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 164, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 164, + 715 + ], + "type": "text", + "content": "Type I: coarse edit. In this setting, the inpainted part of " + }, + { + "bbox": [ + 47, + 677, + 164, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 677, + 164, + 715 + ], + "type": "text", + "content": " is fully contained" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 185, + 675, + 287, + 723 + ], + "blocks": [ + { + "bbox": [ + 185, + 675, + 287, + 723 + ], + "lines": [ + { + "bbox": [ + 185, + 675, + 287, + 723 + ], + "spans": [ + { + "bbox": [ + 185, + 675, + 287, + 723 + ], + "type": "image", + "image_path": "7a542f470ab3eb8a79488bfb5d39b1288edd350105e276c7e09712689b655007.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 424, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 424, + 84 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 424, + 84 + ], + "type": "text", + "content": "inside " + }, + { + "bbox": [ + 304, + 72, + 424, + 84 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 424, + 84 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 72, + 424, + 84 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 424, + 84 + ], + "type": "text", + "content": " is computed" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": "by randomly sampling a part of " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": " and taking its convex hull. To select such part, we randomly sample a plane passing through " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": ", effectively splitting " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": " into two parts, and we select one part at random. More precisely, we start by sampling a random point " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": " inside the bounding box of " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": " and a random direction " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": ". The plane " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": " passing through " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": " with normal " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": " is defined by " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\{x \\in \\mathbb{R}^3 | x \\cdot p = p \\cdot n\\}" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": " is defined as the convex hull of all the face midpoints that are above " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": " i.e. " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\{f = (v_1, v_2, v_3) | f \\in F, \\frac{v_1 + v_2 + v_3}{3} \\cdot p >= p \\cdot n\\}" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": " denotes the list of faces. To avoid Z-fighting during rendering between " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": ", we scale " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": " by 20% while keeping its center of mass the same, ensuring it completely envelopes the part of " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": " above " + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 84, + 545, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 247, + 422, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 422, + 260 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 422, + 260 + ], + "type": "text", + "content": "Type II: mesh sculpting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 260, + 422, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 260, + 422, + 295 + ], + "spans": [ + { + "bbox": [ + 305, + 260, + 422, + 295 + ], + "type": "text", + "content": "This type of mask is designed to represent a more precise edit where the user" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 442, + 248, + 545, + 292 + ], + "blocks": [ + { + "bbox": [ + 442, + 248, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 442, + 248, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 442, + 248, + 545, + 292 + ], + "type": "image", + "image_path": "af72ce7a09671870435d0637e91efa18a5a4f45a20d6f2e79494298a7ff6d688.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "text", + "content": "expects content to be created in a portion of space similar to the mask. It requires more expertise and time from the user than Type I, but also provides more precise control over the generated content. As we can see in the illustration to the right, " + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "text", + "content": " is a tight fit over " + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "inline_equation", + "content": "S -" + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "text", + "content": " no volume inside " + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "text", + "content": " is not also inside " + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "text", + "content": ". To generate these masks, we sample a plane " + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "text", + "content": ", as in Type I masks, and select all the faces that have their midpoint above " + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "text", + "content": " i.e. " + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "inline_equation", + "content": "M = \\{f = (v_{1}, v_{2}, v_{3}) | f \\in F, \\frac{v_{1} + v_{2} + v_{3}}{3} \\cdot p > = p \\cdot n\\}" + }, + { + "bbox": [ + 304, + 296, + 545, + 406 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 411, + 422, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 422, + 423 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 422, + 423 + ], + "type": "text", + "content": "Type III: surface editing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 423, + 422, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 423, + 422, + 459 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 422, + 459 + ], + "type": "text", + "content": "We aim to support local texture modifications, where the user selects a surface" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 443, + 410, + 545, + 458 + ], + "blocks": [ + { + "bbox": [ + 443, + 410, + 545, + 458 + ], + "lines": [ + { + "bbox": [ + 443, + 410, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 443, + 410, + 545, + 458 + ], + "type": "image", + "image_path": "949faf04bdd4cb2f8ebfaf692a32b0b3496c49819e31a8ba68937a3f65a07338.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "text", + "content": "patch and prompts Instant3dit to modify its texture. There are various ways to generate this type of mask if we assume that the meshes have good triangulation. Unfortunately, this is typically not the case for shape datasets. Thus, propose a simple heuristic that works for any triangle soup. We sample a vertex " + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "text", + "content": ", then several cylinders with elliptical bases of varying sizes, all centered on " + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "text", + "content": " (depicted in purple). The number of cylinders is uniformly sampled between 3 and 6, the revolution axis is sampled on the unit sphere, the height and radii are sampled between 0.1 and 0.3. We call this volume " + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "text", + "content": ". Finally select all the faces whose midpoint falls within " + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "text", + "content": " i.e. " + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "inline_equation", + "content": "M = \\{f = (v_{1}, v_{2}, v_{3}) | f \\in F, \\frac{v_{1} + v_{2} + v_{3}}{3} \\in C\\}" + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "text", + "content": " (depicted in red)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 639, + 388, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 388, + 652 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 388, + 652 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 656, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 656, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 656, + 545, + 691 + ], + "type": "text", + "content": "In this section, we describe our novel multiview inpainting benchmark, which we use to compare several baselines and run ablations." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Benchmark. We hold out 500 2x2 multiview images and" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16277" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 287, + 152 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 152 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 152 + ], + "type": "image", + "image_path": "169028abaf099140377ae6f8b823c15d0e6bf03b031f1404693c6c0428cc188b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 159, + 288, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 288, + 194 + ], + "type": "text", + "content": "Figure 4. Comparison to baselines. We show different inpainting results from different baselines; our multiview inpainting method offers the highest quality while maintaining consistency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 288, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 288, + 266 + ], + "type": "text", + "content": "their synthetic multiview masks from the training set to create our benchmark. We condition all inpainting on BLIP [25] captions, and ensure no training object is present in the benchmark. We then evaluate how well different methods inpaint all the views." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 276, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 289, + 456 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate the quality of multiview inpainting, we use three types of metrics: measuring prompt adherence, multiview consistency, and visual quality. To measure prompt adherence, we use CLIP [38] similarity score between the generated 2x2 multiview image and the text prompt. We follow previous work [36] and use two models, CLIP-ViT-L-14 (ClipL) and CLIP-ViT-BigG-14 (ClipG), for encoding. To measure 3D consistency we reconstruct NeRF from the sparse inpainted views [26], and re-render from same camera angles as the input sparse views. We then compare the NeRF renderings to generated inpainted images using various image similarity scores: SSIM [48], LPIPS [59] and DreamSim [14]. Finally, we measure visual quality using FID score [18] comparing to a distribution of held-out images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 464, + 288, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 288, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 288, + 681 + ], + "type": "text", + "content": "Comparison to Baselines. We use these metrics to compare our method (Table 1, bottom) to several baseline alternatives (Table 1, top). As there are no off-the-shelf multiview consistent inpainting methods, we couple different image diffusion techniques with blended diffusion [2] to inpaint the multiview image. We try two different architectures from the commonly used SDXL diffusion model [36], SDXL (as-is) and SDXL-inpainting (fine-tuned on 2D inpainting task). Both SDXL baselines perform poorly on multiview consistency since they are not explicitly trained to complete images consistently and, in the case of SDXL-inpainting, are only trained with 2D image masks. Finally we test Instant3D [26] coupled with blended diffusion. Since this architecture is trained for 3D reconstruction it achieves a better multiview consistency, however, our method still outperforms this baseline with respect to all other metrics. See Figure 4 for some example results produced with baselines and our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Ablating Diffusion Backbones. Our method can be fine-tuned from any image diffusion model, and thus we" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 69, + 545, + 163 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 545, + 163 + ], + "type": "table", + "html": "
MethodPrompt AdherenceMulti-view consistencyVisual quality
ClipL↑ClipG↑SSIM↑LPIPS↓DreamSim↓FID↓
SDXL [36]28.6342.580.8740.0650.134127.0
SDXL-inpainting [36]27.5741.330.8570.0640.137159.5
Instant3D [26]28.7843.060.8920.0440.102120.0
ablations from different diffusion backbones (our final method is at the bottom)
Ours (SD 1.5-inpainting)27.7941.330.7190.090.599128.3
Ours (SD 2.0-inpainting)27.5941.230.7290.0960.589124.5
Ours (Instant3D)28.5742.500.8940.0430.097121.1
Ours (SDXL-inpainting)29.0143.480.8940.0450.100118.4
", + "image_path": "3191bc7c6e1681e0b94c70dd69dc7dde83c951e2d42f5fb0d03f183e774ad128.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 320, + 244, + 531, + 319 + ], + "blocks": [ + { + "bbox": [ + 304, + 171, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 304, + 171, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 547, + 228 + ], + "type": "text", + "content": "Table 1. Multiview text-to-image inpainting. This table demonstrates quantitative performance of several baselines and ablations of our method, with respect to metrics that capture prompt adherence, multiview consistency, and visual quality. Results are color-coded worst and best, with best highlighted in bold." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 244, + 531, + 319 + ], + "lines": [ + { + "bbox": [ + 320, + 244, + 531, + 319 + ], + "spans": [ + { + "bbox": [ + 320, + 244, + 531, + 319 + ], + "type": "table", + "html": "
MethodType IType IIType IIII+II+IIIUser Generated
FID↓ClipG↑
Random 2D145.4129.6102.3131.124.22
Type I131.2128.6102.3121.326.24
Type II170.9122.6101.7128.225.53
Type III193.0148.199.05142.224.2
I+II+III130.0124.9100.0118.426.5
", + "image_path": "744674b4de2b929bea86cdfae192175ee404b9b05505a7a3ab8f65c8855964c0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 327, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 304, + 327, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 546, + 361 + ], + "type": "text", + "content": "Table 2. Mask ablation. We ablate the choice of 3D masks used during training (rows), and evaluate on different subsets of the benchmark containing only some types of masks (columns)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 382, + 547, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 547, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 547, + 610 + ], + "type": "text", + "content": "further evaluate how performance changes with respect to different backbones (Table 1, bottom section). Unsurprisingly, weaker backbones with fewer parameters (SD 1.5, SD 2.0) [40] lead to inferior performance and very poor multiview consistency. Using Instant3D [26] gives the best multiview consistency, but fine-tuning does not seem to improve visual quality and prompt adherence. Our method fine-tunes from SDXL-inpainting [36], and generally learns to better adhere to the prompt with high visual fidelity. We believe this observation is consistent with insights from prior works [26, 50] that show that existing diffusion models already have some implicit understanding of multiview images and can be fine-tuned for 3D consistency with little training data. Inpainting, on the other hand, is a more challenging task, and thus, it is easier to fine-tune a model that was trained at scale for an inpainting task (SDXL-inpainting) to create a multiview consistent image. See supplemental material for qualitative examples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 623, + 547, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 623, + 547, + 719 + ], + "spans": [ + { + "bbox": [ + 304, + 623, + 547, + 719 + ], + "type": "text", + "content": "Ablating Masks. The key ingredient in our training data is the three types of multiview consistent masks. To evaluate the importance of training masks, we ablate on different types of masks, and fine-tune SDXL inpainting model using only one mask type. Each row in Table 2 is trained with a different type of mask: Random 2D follows prior 2D inpainting work [40] sampling mask independently per image (see inset)," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16278" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 204, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 204, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 204, + 132 + ], + "type": "text", + "content": "Type I only uses coarse 3D blobs, Type II uses large surface selections, Type III uses local surface patches (see Section 3.2). Bottom row is our method trained with all types of masks." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 61, + 272, + 120 + ], + "blocks": [ + { + "bbox": [ + 220, + 61, + 272, + 120 + ], + "lines": [ + { + "bbox": [ + 220, + 61, + 272, + 120 + ], + "spans": [ + { + "bbox": [ + 220, + 61, + 272, + 120 + ], + "type": "image", + "image_path": "0e4ef81960025b4d550bd465cef7cdbf601ba96bb9ada7060adeca94d776d6c3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 134, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 287, + 240 + ], + "type": "text", + "content": "To compare these different techniques, we split our benchmark into single-mask vs all-masks data, and compute FID score for each test subset (see columns in Table 2). Unsurprisingly, there is a strong bias in the diagonal, indicating that training on a particular type of mask helps inpainting it at inference time. We note, however, that simple random rectangular masks perform substantially worse. Our finetuning on all mask types offers the best performance on arbitrary masks at inference time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 243, + 288, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 363 + ], + "type": "text", + "content": "We further take 15 user-generated masks and inpaint them to evaluate how well different types of training masks generalize to real use-cases. We found that our multi-mask training offers the highest Clip-similarity of inpainted results to the user prompts, unfortunately FID score is not meaningful for such a small sample size. We refer the reader to supplemental material where we provide qualitative inpainting results for user-generated masks, demonstrating that multi-mask training data yields the highest inpainting quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 287, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 423 + ], + "type": "text", + "content": "Generalization to Novel Camera Angles. We note that unlike Instant3d [26], the inpainting model adapts its output to the orientation of the non-masked region, indicating an ability to generalize to unseen azimuth angles." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 424, + 169, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 169, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 169, + 542 + ], + "type": "text", + "content": "We quantify this by offsetting the 2x2 camera positions by 16 equally spaced azimuth offsets from 0 to 337.5 degrees, generating 500 images for each offset (using the same models as in our benchmark). See how evaluation metrics change for different azimuth offsets in the inset." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 171, + 424, + 287, + 531 + ], + "blocks": [ + { + "bbox": [ + 171, + 424, + 287, + 531 + ], + "lines": [ + { + "bbox": [ + 171, + 424, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 171, + 424, + 287, + 531 + ], + "type": "image", + "image_path": "3d4dfc53a5315e4834937e8837ad73246a56abc9c3565442952d3fe348f94b0a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 665 + ], + "type": "text", + "content": "The FID, ClipL, and ClipG scores remain consistent throughout. This adaptability is crucial for inpainting, as the orientation of the unmasked portion is unknown beforehand. Fine-tuning follows [26], with multiview renders and masks at fixed angles for each object. Remarkably, the fine-tuned model generalizes to unseen camera orientations and FOV angles, likely due to the strong inductive bias from the masked image and prior knowledge from the baseline diffusion model. We show some visual examples in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Limitations. Our multiview inpainting might ignore thin masks (a behavior also observed in 2D inpainting models) (Figure 5, left). Additionally, due to training renders having" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 545, + 133 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 133 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 133 + ], + "type": "image", + "image_path": "c12462c3fd0cb7f764fe6f5ab5a807da10fc15a7ff83dc08ab3abf729b00ef63.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 142, + 545, + 175 + ], + "lines": [ + { + "bbox": [ + 305, + 142, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 305, + 142, + 545, + 175 + ], + "type": "text", + "content": "Figure 5. Failure Cases. Typical failure cases include failure to adhere to the prompt for thin masks or large masks, which have little inductive bias from the unmasked area." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 180, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 545, + 240 + ], + "type": "text", + "content": "only a white background, we find that in some cases without existing inductive bias for the mask, the network prefers to generate a white background at the expense of aligning with the prompt (Figure 5, right). These cases can usually be resolved by choosing a different random seed." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 251, + 386, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 386, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 386, + 264 + ], + "type": "text", + "content": "5. Applications" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 270, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 390 + ], + "type": "text", + "content": "While the main contribution of our work is the multiview consistent inpainting method, which can be trivially complemented with a large reconstruction model (LRM) to infer the full 3D shape from the completed views. In this section we explore applications of our technique for editing various neural and traditional 3D representations. We also developed an interactive application where users can load their own 3D shapes, create masks with basic primitives and visualize the results of our method. More information about this application can be found in supplemental." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 617 + ], + "type": "text", + "content": "NeRF Editing. Given an input NeRF, user's mask and a prompt, we use our multiview inpainting to generate the edited views. These views are used as input to the LRM from Instant3D [26] to create the modified representation. We show a few edits in Figure 2 (our result is on the right). In addition to our method we show results for several alternative techniques that allow editing NeRFs. Vox-E [42] and MVEdit [8] do not allow for a user-provided mask, and instead infer the area to be edited using the attention weights from the prompt, which may result in undesirable changes in texture and geometry throughout the object in areas the user would like to keep unchanged. Progressive3d [10] and NeRFFiller [50] struggle with multiview consistency, as they rely on 2D diffusion models for SDS and IDU, respectively. This may result in incorrectly positioned edits, as in the case of the centaur example. Note that our approach is also significantly faster than all existing timelines and produces results in seconds rather than minutes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 700 + ], + "type": "text", + "content": "We further conduct an informal user preference study between our method and the closest approach that can also inpaint multiview masks, NeRFiller [50]. We showed the masked input and the prompt to 15 users and asked them to choose between two outputs (see supplemental for exact verbiage). Out of 208 pairs, the users preferred our method in 180 cases " + }, + { + "bbox": [ + 304, + 617, + 545, + 700 + ], + "type": "inline_equation", + "content": "(86\\%)" + }, + { + "bbox": [ + 304, + 617, + 545, + 700 + ], + "type": "text", + "content": " vs 28 " + }, + { + "bbox": [ + 304, + 617, + 545, + 700 + ], + "type": "inline_equation", + "content": "(14\\%)" + }, + { + "bbox": [ + 304, + 617, + 545, + 700 + ], + "type": "text", + "content": " for NeRFiller." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 702, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 545, + 714 + ], + "type": "text", + "content": "Since our multiview inpainting is not tied to any par" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16279" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 282, + 223 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 282, + 223 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 282, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 282, + 223 + ], + "type": "image", + "image_path": "dc35e1c1b29a9775d3be2e4caf18a1d21c62ff6203e93537d3d9362a9150be2b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 290, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 290, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 290, + 331 + ], + "type": "text", + "content": "Figure 6. Application: NeRF editing. We compare our method (rightmost column) to other methods that allow editing NeRFs, by rendering a front-facing view of the 3D edited asset for all methods. Note that the only method that takes arbitrary masks as input is NeRFiller [50], we also conduct an informal user study with 15 users choosing between our output and NeRFiller for 208 pairs of results. We found that users prefer our method in " + }, + { + "bbox": [ + 46, + 232, + 290, + 331 + ], + "type": "inline_equation", + "content": "86\\%" + }, + { + "bbox": [ + 46, + 232, + 290, + 331 + ], + "type": "text", + "content": " of the cases. We report average timing for all methods, and confirm Instant3dit is substantially faster than other alternatives." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 342, + 286, + 453 + ], + "blocks": [ + { + "bbox": [ + 47, + 342, + 286, + 453 + ], + "lines": [ + { + "bbox": [ + 47, + 342, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 286, + 453 + ], + "type": "image", + "image_path": "270f1bb8804106b615ae5669f15109f604f6711de10d5fbb9991844bc261cd0b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 462, + 288, + 509 + ], + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 509 + ], + "type": "text", + "content": "Figure 7. Application: mesh editing. Each of the 3 edits shown is performed locally on the original mesh, only modifying regions selected by the user. Locality is achieved by running the ROAR [3] geometry optimization, which takes 20 extra seconds per edit." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 520, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 556 + ], + "type": "text", + "content": "ticular 3D representation, our method can also be used to edit Gaussian Splats (GS) by simply using an appropriate LRM [58], see the supplementary for GS editing examples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 568, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 288, + 689 + ], + "type": "text", + "content": "Mesh Editing. For the users that work on traditional mesh representation, we propose using MeshLRM [51] with an adaptive remeshing layer [3] that uses the LRM result as a guidance. This leads to a fast and fully controllable mesh editing pipeline that is guaranteed to preserve the original mesh attributes (e.g., UVs, rigging) and triangulation in the unedited parts of the mesh [4] (Figure 7). To maintain fine details, we use a normal estimator [15] directly on the diffusion output and use the normals as targets in the vertex optimization. More examples in supplemental material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Texture Editing. The user can use our method to edit sur" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 331, + 69, + 521, + 203 + ], + "blocks": [ + { + "bbox": [ + 331, + 69, + 521, + 203 + ], + "lines": [ + { + "bbox": [ + 331, + 69, + 521, + 203 + ], + "spans": [ + { + "bbox": [ + 331, + 69, + 521, + 203 + ], + "type": "image", + "image_path": "c21ec592231fa40a2b824211c5c1e6a601fda9e6d96a516ac15b4d419913187a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 210, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 305, + 210, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 210, + 547, + 255 + ], + "type": "text", + "content": "Figure 8. Application: texture editing. Our method can be used to modify texture on a user-selected region. In this case, we run through our NeRF editing pipeline, but only sample colors from the NeRF in the selected region." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 269, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 269, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 269, + 547, + 365 + ], + "type": "text", + "content": "face texture details. We use NeRF-LRM to reconstruct a 3D representation and back-project the colors to the user-selected mesh region. Starting with an auto-generated texture from previous work [39] (Fig. 8, first column), the user labels a mask (second column) to either add new texture elements (facemask, saddle) or fix artifacts in the texture (goldfish exhibiting inconsistent left-right colors due to lack of multiview consistency in the previous texturing method)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 380, + 379, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 379, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 379, + 392 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 401, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 533 + ], + "type": "text", + "content": "We introduced Instant3dit, a multiview inpainting diffusion model, and show its application to fast and localized editing of 3D assets. We propose three types of inpainting masks to train the model, corresponding to three types of user edits, and plan to release this dataset. We justify our training strategy and choice of 3d masks by measuring a comprehensive set of metrics: prompt adherence, 3d consistency of the generated content, and generation quality. Instant3dit shows superior quality to the contemporary 3D editing pipelines and is orders of magnitude faster, running as fast as a single image generation using diffusion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "We see clear opportunities to improve this tool further. First, SDXL limits our input resolution to " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "1024^2" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ". We follow in the steps of previous works [26, 31, 52] which found that a " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " grid of " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "512^2" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " images strikes a good balance between resolution and coverage. An interesting avenue for future work is to explore increasing the number of views (i.e. to " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ") in addition to using multiview-aware super-resolution. Additionally, several approaches such as DMD [56] distill diffusion models into one-step or few-step models, which would make Instant3dit run in a fraction of a second and bring a truly interactive editing experience. Second, video diffusion model [5, 45] have shown remarkable 3D consistency and visual quality, which can be harnessed to improve the quality of 3D generation further, although, in terms of speed, they are currently slower than text-to-image models." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16280" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 186, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 186, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 186, + 102 + ], + "type": "text", + "content": "[1] Adobe Inc. Adobe photoshop. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 287, + 135 + ], + "type": "text", + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. CVPR, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 137, + 287, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 180 + ], + "type": "text", + "content": "[3] Amir Barda, Yotam Erel, Yoni Kasten, and Amit H. Bermano. Roar: Robust adaptive reconstruction of shapes using planar projections. arXiv preprint arXiv:2307.00690, 2023. 4, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 183, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 183, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 183, + 287, + 226 + ], + "type": "text", + "content": "[4] Amir Barda, Vladimir G. Kim, Noam Algerman, Amit H. Bermano, and Thibault Groueix. Magicclay: Sculpting meshes with generative neural fields. SIGGRAPH Asia, 2024. 1, 2, 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 271 + ], + "type": "text", + "content": "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. https://openai.com/research/, 2024.8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 274, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 287, + 316 + ], + "type": "text", + "content": "[6] Chenjie Cao, Chaohui Yu, Yanwei Fu, Fan Wang, and Xi-angyang Xue. Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing. arXiv preprint arXiv:2408.08000, 2024.1,2,3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 318, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 373 + ], + "type": "text", + "content": "[7] Duygu Ceylan, Valentin Deschaintre, Thibault Groueix, Rosalie Martin, Chun-Hao Huang, Romain Rouffet, Vladimir Kim, and Gaetan Lassagne. Matatlas: Text-driven consistent geometry texturing and material assignment. arXiv preprint arXiv:2404.02899, 2024.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "type": "text", + "content": "[8] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032, 2024. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 420, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 453 + ], + "type": "text", + "content": "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 454, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 498 + ], + "type": "text", + "content": "[10] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. ICLR, 2024. 1, 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 499, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 531 + ], + "type": "text", + "content": "[11] Dale Decatur, Itai Lang, Kfir Aberman, and Rana Hanocka. 3d paintbrush: Local stylization of 3d shapes with cascaded score distillation. CVPR, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "type": "text", + "content": "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of " + }, + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "type": "text", + "content": " objects. NeurIPS, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. CVPR, 2023. 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[14] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. NeurIPS, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowiz-" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ard: Unleashing the diffusion priors for 3d geometry estimation from a single image. ECCV, 2024. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[16] William Gao, Noam Aigerman, Groueix Thibault, Vladimir Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. SIGGRAPH, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 130, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 162 + ], + "type": "text", + "content": "[17] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 164, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 207 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2018. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "text", + "content": "[19] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. ICLR, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 254, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 286 + ], + "type": "text", + "content": "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 288, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 545, + 319 + ], + "type": "text", + "content": "[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 322, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 322, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 322, + 545, + 364 + ], + "type": "text", + "content": "[22] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Popa Tiberiu. Clip-mesh: Generating textured meshes from text using pretrained image-text models. SIGGRAPH Asia, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 366, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 408 + ], + "type": "text", + "content": "[23] Hyunwoo Kim, Itai Lang, Noam Aigerman, Thibault Groueix, Vladimir G. Kim, and Rana Hanocka. Meshup: Multi-target mesh deformation via blended distillation. 3DV, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "text", + "content": "[24] Hyunwoo Kim, Itai Lang, Noam Aigerman, Thibault Groueix, Vladimir G. Kim, and Rana Hanocka. Meshup: Multi-target mesh deformation via blended score distillation. 3DV, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICLR, 2022. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "text", + "content": "[26] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. ICLR, 2024. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 557, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 590 + ], + "type": "text", + "content": "[27] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. AAAI, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 591, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 612 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 614, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 657 + ], + "type": "text", + "content": "[29] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. ICLR, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[30] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. CVPR, 2024. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16281" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[31] Yuanxun Lu, Jingyang Zhang, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, Long Quan, Xun Cao, and Yao Yao. Direct2.5: Diverse text-to-3d generation via multi-view 2.5d diffusion. CVPR, 2024. 2, 3, 4, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[32] Artem Lukoianov, Haitz Sáez de Ocariz Borde, Kristjan Greenewald, Vitor Campagnolo Guizilini, Timur Bagautdinov, Vincent Sitzmann, and Justin Solomon. Score distillation via reparametrized ddim. NeurIPS, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 195 + ], + "type": "text", + "content": "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. CVPR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 228 + ], + "type": "text", + "content": "[34] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. CVPR, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "text", + "content": "[35] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. ICCV, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. *ICLR*, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "text", + "content": "[37] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. *ICLR*, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 397 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. ICML, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 431 + ], + "type": "text", + "content": "[39] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. SIGGRAPH, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 433, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 464 + ], + "type": "text", + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2022. 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "text", + "content": "[42] Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. Vox-e: Text-guided voxel editing of 3d objects. CVPR, 2023. 1, 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 534, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 588 + ], + "type": "text", + "content": "[43] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "text", + "content": "[44] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. ICLR, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 259, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 259, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 259, + 635 + ], + "type": "text", + "content": "[45] Genmo Team. Mochi 1. GitHub repository, 2024. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 677 + ], + "type": "text", + "content": "[46] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. 3DV, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[47] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 643 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[48] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 2004. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[49] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. CoRR, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[50] Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Nerfill: Completing scenes via generative 3d inpainting. CVPR, 2024. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 197, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 250 + ], + "type": "text", + "content": "[51] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 3, 4, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 296 + ], + "type": "text", + "content": "[52] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 2, 3, 4, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 297, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 330 + ], + "type": "text", + "content": "[53] Hanyu Xiang, Qin Zou, Muhammad Ali Nawaz, Xianfeng Huang, Fan Zhang, and Hongkai Yu. Deep learning for image inpainting: A survey. Pattern Recognition, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 331, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 384 + ], + "type": "text", + "content": "[54] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 387, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 441 + ], + "type": "text", + "content": "[55] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. CVPR, 2024. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 442, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 442, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 442, + 545, + 485 + ], + "type": "text", + "content": "[56] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. CVPR, 2024. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 487, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 531 + ], + "type": "text", + "content": "[57] Yu Zeng, Zhe L. Lin, Jimei Yang, Jianming Zhang, Eli Shechtman, and Huchuan Lu. High-resolution image inpainting with iterative confidence feedback and guided upsampling. ECCV, 2020. 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 533, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 575 + ], + "type": "text", + "content": "[58] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. ECCV, 2024. 2, 3, 4, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "type": "text", + "content": "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. CVPR, 2018. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 611, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 643 + ], + "type": "text", + "content": "[60] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. SIGGRAPH Asia, 2023. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16282" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/a04fb8db-d9e4-4e5e-a8c0-00327c1bd422_content_list.json b/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/a04fb8db-d9e4-4e5e-a8c0-00327c1bd422_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..98f31d29bd09cfab9f82574e1215250b1df35eeb --- /dev/null +++ b/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/a04fb8db-d9e4-4e5e-a8c0-00327c1bd422_content_list.json @@ -0,0 +1,2942 @@ +[ + { + "type": "text", + "text": "Instruct-CLIP: Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning", + "text_level": 1, + "bbox": [ + 187, + 130, + 816, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sherry X. Chen", + "bbox": [ + 276, + 215, + 403, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Misha Sra", + "bbox": [ + 468, + 215, + 553, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pradeep Sen", + "bbox": [ + 617, + 215, + 720, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of California, Santa Barbara", + "bbox": [ + 341, + 246, + 653, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xchen774, sra, psen}@ucsb.edu", + "bbox": [ + 364, + 266, + 630, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although natural language instructions offer an intuitive way to guide automated image editing, deep-learning models often struggle to achieve high-quality results, largely due to the difficulty of creating large, high-quality training datasets. To do this, previous approaches have typically relied on text-to-image (T2I) generative models to produce pairs of original and edited images that simulate the input/output of an instruction-guided image-editing model. However, these image pairs often fail to align with the specified edit instructions due to the limitations of T2I models, which negatively impacts models trained on such datasets. To address this, we present Instruct-CLIP (I-CLIP), a self-supervised method that learns the semantic changes between original and edited images to refine and better align the instructions in existing datasets. Furthermore, we adapt Instruct-CLIP to handle noisy latent images and diffusion timesteps so that it can be used to train latent diffusion models (LDMs) and efficiently enforce alignment between the edit instruction and the image changes in latent space at any step of the diffusion pipeline. We use Instruct-CLIP to correct the InstructPix2Pix dataset and get over 120K refined samples we then use to fine-tune their model, guided by our novel I-CLIP-based loss function. The resulting model can produce edits that are more aligned with the given instructions. Our code and dataset are available at https://github.com/SherryXTChen/Instruct-CLIP.git.", + "bbox": [ + 89, + 359, + 483, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 782, + 220, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in instruction-guided image editing have introduced intuitive and powerful tools that require only a single text instruction for image editing [1, 24-27]. In particular, InstructPix2Pix (IP2P) [1] pioneered diffusion-based instruction-driven image editing, serving as the foundation for numerous subsequent works [8, 11, 25]. These", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4064cbdbe74a34237954bc693f8b510a892fa479a3708327c059f926106889ee.jpg", + "image_caption": [ + "Original", + "\"Change the panda to a monkey\"" + ], + "image_footnote": [], + "bbox": [ + 527, + 340, + 647, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8780646092415adacdf94000231248dd14457fc025d2e19dde7e7f88439ab3f.jpg", + "image_caption": [ + "IP2P" + ], + "image_footnote": [], + "bbox": [ + 647, + 340, + 766, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/46c9334dc4c13157fa99244bf8dd8ea965351b7cf6fa0e94ae3e85f51609fd88.jpg", + "image_caption": [ + "I-CLIP (Ours)" + ], + "image_footnote": [], + "bbox": [ + 767, + 340, + 887, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e143abed567f69ba12f36627f005a62faf7bcedbc1be7290396a325fb841b93.jpg", + "image_caption": [ + "\"Let the old man be dressed in a military uniform\"" + ], + "image_footnote": [], + "bbox": [ + 527, + 410, + 647, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f02c93c2bdb28c512dfe34789bb85af4167fcf070c3e7247544740ea7ccc0580.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 410, + 766, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a26de27ca92f9dacfe10dafa91843a707a0860b84e8bcbaad2d0acc39beb3717.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 410, + 887, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3b0ea394ed1ecb2d2ee52e6bbd7fb303ac3cebf2ab5e8b9e051dc16a017ff95b.jpg", + "image_caption": [ + "\"Change the background to a classroom\"" + ], + "image_footnote": [], + "bbox": [ + 527, + 511, + 647, + 564 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73db55f64e3a6f8c390a3653e874c27b73501d75af6fc082036c8e883641c467.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 511, + 766, + 564 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c128f79eca84f99b119be5e19901d71c36a22cd0cb823af09bb97356ed9a55fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 511, + 887, + 564 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e88fc350301b685886516f25da5eb468528559cdc1c1f6010633428ab642a00.jpg", + "image_caption": [ + "\"The cat needs to be black\"" + ], + "image_footnote": [], + "bbox": [ + 527, + 580, + 647, + 648 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/936477d796fc15f5c908ac607c9ee1e1a1c94caba67c2bcb6fdce16c9f58a45a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 580, + 766, + 648 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3fa513c733c17c20876c10f325c9723ef221f2af50db1ba1dcbd92bb60948d6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 580, + 887, + 648 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/98bfd82049da715b9ff632e31b1875e64fb5d15b350f34951f148e8dde9c3ba5.jpg", + "image_caption": [ + "\"Make it a chocolate cake\"", + "Figure 1. Results showcasing the strength of Instruct-CLIP (I-CLIP) compared to state-of-the-art InstructPix2Pix (IP2P) [1]." + ], + "image_footnote": [], + "bbox": [ + 527, + 664, + 647, + 757 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c9c36dfcf2aea554d3901e78253d552d91192ac51ad1e31cec3d2aa82c757035.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 665, + 766, + 757 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dea6f4ddbe6f812772756c5434f1eb30a14c056b574f2e724512cb72b3da6372.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 665, + 887, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods leverage pre-trained text-to-image (T2I) models [3, 10, 15, 18, 19] and condition the generation process on edit instructions instead of the usual prompts. However, these models still need to be fine-tuned on appropriate datasets to learn how to make instruction-guided edits.", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28513", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/09d638d9d0ba49a534a4105e9a8ef8f61a1809c31300fce056aec719c13bf2d8.jpg", + "image_caption": [ + "Original" + ], + "image_footnote": [], + "bbox": [ + 93, + 103, + 184, + 175 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f4527b57e63dbba0757e278c3e5fab77810a05b1895ebc7e2072606b6de95a80.jpg", + "image_caption": [ + "Edited", + "ORIGINAL: \"by andy warhol\" (0.130) \nREFINED: \"make the sky pink\" (0.263)" + ], + "image_footnote": [], + "bbox": [ + 186, + 103, + 279, + 175 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8c55a7127e7266c2c581611c409d0afcc460377236f8c70b432c7a00f83dfb9f.jpg", + "image_caption": [ + "Original", + "ORIGINAL: \"make the waves into a hurricane\" (0.090) \nREFINED: \"add a lightning storm to the sky\" (0.289)" + ], + "image_footnote": [], + "bbox": [ + 297, + 104, + 385, + 174 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/176d5acdd320062248080f7549f87d837b35d9b864efe44e61fd36ab8457d13c.jpg", + "image_caption": [ + "Edited" + ], + "image_footnote": [], + "bbox": [ + 390, + 104, + 478, + 174 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5b3651168b82d2b723ea14240d4ff8d0c5c39a4bae60ad87cc5374ff50fae067.jpg", + "image_caption": [ + "ORIGINAL:\"make it morning\"(0.179) \nREFINED:\"make it a foggy morning\"(0.324)" + ], + "image_footnote": [], + "bbox": [ + 93, + 199, + 184, + 270 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e24e9330e453ab7aa0f016124871ae34d3760fa6666eaee31f26dc1dbe94e86c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 186, + 199, + 279, + 270 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/66aea714ea8979115bdf574b4c60a40e7387f9d69cf1b056aebf7cd025be2596.jpg", + "image_caption": [ + "ORIGINAL: \"make it a landscape\" (0.090) \nREFINED: \"a photograph\" (0.233)" + ], + "image_footnote": [], + "bbox": [ + 295, + 198, + 388, + 270 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f53887aa8d2ce70146eb91f7448816ee12ada20f103b1285545512d8b03220c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 198, + 480, + 270 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1cdf2f5294b4d86f51f1998406058bacfe1d404ef93b51d91fd69be4abcb047f.jpg", + "image_caption": [ + "ORIGINAL: \"make it a daydream\" (0.113) \nREFINED: \"make it spring\" (0.300)", + "Figure 2. Problems with existing instruction-guided image-editing datasets [1]. As shown, there are many examples where the dataset's original edit instruction does not match the actual changes in the images. Our I-CLIP approach refines edit instructions to match the visual change better and allows us to train a system that produces better outputs. The values in parentheses are the cosine similarity between the visual change from the original to the edited image and the edit instruction from I-CLIP." + ], + "image_footnote": [], + "bbox": [ + 93, + 294, + 184, + 366 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/63dc5c118c824db7085957c91b1589d77b6becf405792d9ace3cde3170893c81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 186, + 294, + 279, + 366 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7d56f08041ea4297a320ee0bcb98d4d4642b9461d51b265dc0a035d978a375d2.jpg", + "image_caption": [ + "ORIGINAL: \"make it a video game character\" (0.097) \nREFINED: \"as a 3 d sculpture\" (0.357)" + ], + "image_footnote": [], + "bbox": [ + 290, + 292, + 380, + 364 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ac214b009a2003d44d3e81e943e6df22cc94d267a28e359603db28e34cfcb883.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 294, + 475, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some have created these datasets by using T2I models to approximate the desired behavior of an instruction-guided image editing system. For example, several approaches [1, 26, 27] leverage Prompt-to-Prompt [6] to generate a pair of images that look like one is an edited version of the other by using a pair of appropriately modified prompts. A separate, fine-tuned large-language model (LLM), such as GPT3 [2], is then used to generate plausible edit instructions from these prompts.", + "bbox": [ + 89, + 534, + 482, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the generality of Prompt-to-Prompt and LLMs, such datasets can cover a wide range of instructions. However, the quality of image editing is often limited by the capabilities of the generative methods, resulting in problems when the changes between the original and the edited image are misaligned with the edit instruction (Fig. 2). This affects the performance on models trained on these datasets, as we can see with the IP2P results in Fig. 1.", + "bbox": [ + 89, + 671, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although improving image-editing datasets would significantly enhance the performance of models trained on them, doing so is difficult. On one hand, manually creating edit instructions and corresponding outputs is extremely labor-intensive and impractical. On the other, it is nontrivial to construct a pipeline that can perform effective edits in order to create an instruction-guided, image-editing", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dataset. After all, if such a pipeline existed, it would be a successful instruction-guided, image-editing method itself.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, creating a brand new dataset from scratch may not be necessary. Despite the inherent limitations of previous datasets, we observe that they still provide some \"signal\" for training image editing models. After all, as shown in Fig. 1, we see that models trained on them produce results that, while imperfect, show that the system is understanding something about the desired edit. Can we harness this signal to refine the dataset and address its issues?", + "bbox": [ + 511, + 122, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by CLIP [17], which learns rich semantic alignments through contrastive language-image pre-training, we embed original-edited image pairs and their corresponding edit instructions into the same feature space in a similar manner with a neural network, which we call Instruct-CLIP. We then follow this with a separate text decoder to generate more precise instructions from the learned features [9].", + "bbox": [ + 511, + 243, + 903, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instruct-CLIP is equipped with a modified DINOv2 [13] backbone capable of handling noisy latent images in Stable Diffusion (SD) [18] so it can not only enhance training data quality in the pre-processing stage but also serve as an efficient learning objective during training. We use Instruct-CLIP to refine instructions in the IP2P dataset [1], resulting in $120\\mathrm{K}+$ unique, enhanced samples (see Fig. 2). The IP2P model is then fine-tuned on this dataset with an Instruct-CLIP-guided loss, enabling the trained instruction-guided editing model to produce results more faithfully aligned with the instructions (see Fig. 5).", + "bbox": [ + 511, + 349, + 903, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 531, + 518, + 763, + 532 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A dual-purpose model that refines dataset instructions and enhances editing model training.", + "- A large-scale dataset with over 120K samples featuring more accurate and enriched instructions.", + "- An instruction-guided image editing method trained on above dataset." + ], + "bbox": [ + 513, + 534, + 903, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 640, + 653, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-image diffusion-based image editing", + "text_level": 1, + "bbox": [ + 511, + 666, + 885, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [3, 4, 10, 15, 18, 19] have set a new standard for image generation quality, and form the basis of various text-to-image (T2I) editing applications. A range of methods [6, 12, 14, 20-22] aim to edit a given image to semantically align with a target prompt, using the input prompt as an anchor. However, traditional T2I models like Stable Diffusion (SD) [18] often yield inconsistent results even with similar prompts, where both subject and context can change significantly. Prompt-to-Prompt [6] addresses this by preserving attention maps for the shared components in the input and the target prompt to improve image visual alignment, but both images are synthesized so it cannot handle arbitrary input images. Null-text inversion [12] overcomes this by reconstructing an input image in SD based on", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28514", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the input prompt. They do this by optimizing the empty-text as the negative prompt, enabling controlled edits when combined with Prompt-to-Prompt.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other methods similarly leverage diffusion inversion techniques paired with unique editing mechanisms: Diffusion Disentanglement [22] aims to find the optimal blend of input and target prompt features to find the compromise between visual alignment with the input image and semantic alignment with the target prompt; EDICT [21] uses coupling inversion for improved reconstruction, which in turn enhance editing performance; pix2pix-zero [14] optimizes model attention maps for the target image to match the ones for the input image for visual consistency between the two; and Plug-and-Play [20] integrates self-attention maps from the input image to guide the edited image generation. While each method has its own advantages, a major shared shortcoming is they require prompts for both input and output images, which can be cumbersome for users.", + "bbox": [ + 89, + 137, + 482, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Instruction-guided image editing", + "text_level": 1, + "bbox": [ + 89, + 375, + 377, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "New approaches have recently emerged that leverage priors from Stable Diffusion and which offer a more user-friendly approach to image editing, requiring only a single edit instruction [1, 8, 11, 24-26]. For example, InstructPix2Pix (IP2P) [1] replaces the prompt with an edit instruction that conditions the generation of the edited image, where the input image is concatenated as an extra condition as well. Among these instruction-guided image editing methods, HIVE [26] collects user feedback on outputs to train a reward model, which iteratively improves the model's output. Watch Your Step [11] and ZONE [8] both use a masking mechanism – either by measuring the discrepancy between IP2P predictions with and without the instruction, or by employing a region intersection-over-union (IoU) scheme with a segmentation model. This helps them avoid editing instruction-irrelevant regions, resulting in better localized editing capabilities.", + "bbox": [ + 89, + 398, + 482, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Instruction-guided datasets", + "text_level": 1, + "bbox": [ + 89, + 666, + 336, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regardless what additional control mechanism is introduced to improve instruction-guided, image-editing results, the development of instruction-guided image editing models fundamentally relies on suitable training datasets. However, creating large-scale, high-quality datasets automatically presents a significant challenge, as such a data creation system would effectively constitute an editing method itself. Prior works have addressed this challenge by combining existing synthesis models to approximate the behavior of desired instruction-guided image editing methods, primarily following two distinct approaches.", + "bbox": [ + 89, + 688, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several works [1, 26, 27] adopted the first approach, which leverages Prompt-to-Prompt [6] to generate pairs of original/edited images using corresponding prompt pairs", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that describe the content of each image. These systems typically fine-tune a large language model (LLM) like GPT-3 [2] to produce edit instructions from prompt pairs. While this approach generates diverse samples, the dataset quality is inherently constrained by the limitations of both Prompt-to-Prompt and the LLM, often resulting in misalignment between the edit instruction and the actual transformation from the original to the edited image (see Fig. 2).", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second approach attempts to address these limitations by creating datasets through inpainting [24, 25]. Here, edited ground truth images are generated by inpainting selected regions of original images. Although this technique improves the quality of localized image editing samples, it has two significant drawbacks: 1) it cannot generate samples with global modifications (such as style transfer), and 2) it often requires manual creation of inpainting regions. These limitations not only increase the cost of dataset creation but also substantially restrict the dataset size.", + "bbox": [ + 511, + 212, + 903, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now describe our approach to address the limitations in previous work.", + "bbox": [ + 511, + 363, + 903, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Instruct-CLIP", + "text_level": 1, + "bbox": [ + 513, + 407, + 660, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to improve previous instruction-guided image editing methods by enhancing the quality of available datasets [1, 7, 23, 26], which we propose to do by correcting their edit instructions. However, while instructions in these datasets sometimes do not reflect the actual changes between the original and edited images, we observe that the datasets overall still provide enough \"signal\" to train editing models. The main challenge is therefore how to properly leverage this \"signal\" to refine the given edit instructions and consequently improve the overall dataset quality.", + "bbox": [ + 511, + 431, + 903, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we propose Instruct-CLIP (I-CLIP, for short), which embeds the visual change between the original image $I^o$ and the edited result $I^e$ into the same feature space as the corresponding edit instruction $p$ through contrastive learning with a neural network. Our approach is inspired by CLIP [17], which learns the semantic alignment between images and their captions. Likewise, Instruct-CLIP learns the relationship between the visual changes in the original-edit image pair and its edit instruction.", + "bbox": [ + 511, + 584, + 905, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Just like in CLIP, our approach has an image encoder that in our case encodes the visual change between the input and edited images (denoted as I-CLIP $_{\\mathrm{vis}}(I^o, I^e))$ , and a text encoder that encodes the edit instruction $(\\mathrm{I - CLIP}_{\\mathrm{txt}}(p))$ . However, unlike the original CLIP image encoder, ours takes both the original and edited images as input so that it can encode their visual difference. As shown in Fig. 3a, we fine-tune the two encoders by computing and minimizing the contrastive loss between them (Eq. 2), similar to CLIP.", + "bbox": [ + 511, + 719, + 903, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, while the architecture of I-CLIP $_{\\mathrm{txt}}$ is the same as CLIP, our I-CLIP $_{\\mathrm{vis}}$ adds two shared-weighted DINOv2 [13] modules in front of CLIP $_{\\mathrm{vis}}$ , as shown in", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28515", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a0be5bc45527e087197317b873cf6ba85f5653dd2f1e24de6a8b44d2dc88aefc.jpg", + "image_caption": [ + "(a) Overall Instruct-CLIP system architecture" + ], + "image_footnote": [], + "bbox": [ + 148, + 85, + 477, + 208 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d7b49a1efa24227fcfc76507870139ba80ed00b2d61f89d720b6b2887a8d0d31.jpg", + "image_caption": [ + "(b) Architecture of visual-change encoder I-CLIP VIS $(I^o, I^e)$", + "Figure 3. Instruct-CLIP architectures. (a) Overview of Instruct-CLIP (I-CLIP), which embeds the visual change in the original/edited images $I^o$ and $I^e$ and the edit instruction $p$ into the same feature space through contrastive loss, $\\mathcal{L}_{\\mathrm{contrast}}$ (Eq. 2). To obtain refined instruction $p$ from its I-CLIP embedding $z^{\\mathrm{txt}}$ , we adopt the same approach in DeCap [9] to decode $z^{\\mathrm{txt}}$ back to $p$ using cross-entropy loss, $\\mathcal{L}_{\\mathrm{DeCap}}$ (Eq. 4). At inference time, the text decoder takes the embedded visual change from the original to the edited image $(z^{\\mathrm{vis}})$ and decodes it to produce a new instruction. Due to the significant cosine similarity gap between $z^{\\mathrm{vis}}$ and $z^{\\mathrm{txt}}$ even when they are well aligned, directly decoding $z^{\\mathrm{vis}}$ leads to suboptimal results. To achieve a representation of $z^{\\mathrm{vis}}$ closer to the text features that the instruction decoder learned during training, we compute $(z^{\\mathrm{vis}})'$ with Eq. 6 and decode it to obtain the refined instruction $p'$ , which is used to improve the dataset. (b) The architecture of image encoder I-CLIP $\\mathrm{vis}$ includes two shared-weighted DINOv2 [13] modules in front of a standard CLIP $\\mathrm{vis}$ encoder." + ], + "image_footnote": [], + "bbox": [ + 552, + 89, + 841, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3b. These DINOv2 units extract rich, robust visual features from the input images, allowing $\\mathrm{CLIP}_{\\mathrm{vis}}$ to focus on encoding the difference between the original and edited images. This is essential when we introduce Stable Diffusion's (SD) latent image encoding and diffusion into the network, as explained in Sec. 3.2.", + "bbox": [ + 88, + 372, + 480, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To describe the architecture of I-CLIP $_{\\mathrm{vis}}$ , we write the output of DINOv2 for a given image $I$ as $d = \\mathrm{DINO}_{\\mathrm{v2}}(I)$ and the intermediate features from each of its $l$ layers as $\\{d_i \\mid i \\in [1, l]\\}$ . So in I-CLIP $_{\\mathrm{vis}}$ , we first we compute $d^o$ and $d^e$ by passing the input image $I^o$ and the edited image $I^e$ through DINOv2, respectively, then compute the DINOv2 feature difference $d^e - d^o$ that we input to CLIP $_{\\mathrm{vis}}$ . The intermediate feature differences $d_i^e - d_i^o$ are also added to input of the $i$ th layer of CLIP $_{\\mathrm{vis}}$ before being processed by that layer. We find that this improves model performance in our experiments by providing additional information that may be lost from $d^e - d^o$ when being processed from the $(i + 1)^{\\mathrm{th}}$ to $l^{\\mathrm{th}}$ layer, similar to skip connections [5].", + "bbox": [ + 89, + 463, + 482, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train Instruct-CLIP, we first initialize $\\mathrm{CLIP}_{\\mathrm{vis}}$ and I-CLIP $_{\\mathrm{txt}}$ with the respective blocks from the pre-trained CLIP model. Then, given a batch of $n$ data samples $\\{(I_i^o,I_i^e,p_i)\\mid i\\in [1,n]\\}$ from an existing instruction-guided image-editing dataset like InstructPix2Pix [1], we first compute the Instruct-CLIP features for both the visual difference and edit instruction:", + "bbox": [ + 89, + 659, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {i} ^ {\\text {v i s}} = \\mathrm {I} - \\operatorname {C L I P} _ {\\text {v i s}} \\left(I _ {i} ^ {o}, I _ {i} ^ {e}\\right) \\\\ z _ {i} ^ {\\mathrm {t x t}} = \\mathrm {I} - \\mathrm {C L I P} _ {\\mathrm {t x t}} (p _ {i}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 768, + 482, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and compute the contrastive loss $\\mathcal{L}_{\\mathrm{contrast}}$ between them:", + "bbox": [ + 89, + 809, + 460, + 824 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n t r a s t}} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\log \\frac {\\exp \\left(\\sin \\left(z _ {i} ^ {\\text {v i s}}, z _ {i} ^ {\\text {t x t}}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(\\sin \\left(z _ {i} ^ {\\text {v i s}}, z _ {j} ^ {\\text {t x t}}\\right) / \\tau\\right)} \\right. \\tag {2} \\\\ \\left. + \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(z _ {i} ^ {\\mathrm {t x t}} , z _ {i} ^ {\\mathrm {v i s}}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(\\operatorname {s i m} \\left(z _ {i} ^ {\\mathrm {t x t}}, z _ {j} ^ {\\mathrm {v i s}}\\right) / \\tau\\right)}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 828, + 482, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where scalar $\\tau$ is a learnable \"temperature\" parameter that controls the sharpness of the similarity distribution and function $\\mathrm{sim}(\\cdot, \\cdot)$ measures normalized cosine similarity:", + "bbox": [ + 511, + 372, + 906, + 419 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s i m} \\left(z _ {1}, z _ {2}\\right) = \\frac {z _ {1} \\cdot z _ {2}}{\\left\\| z _ {1} \\right\\| \\left\\| z _ {2} \\right\\|}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 433, + 906, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first use this contrastive loss to fine-tune the I-CLIP $_{\\mathrm{vis}}$ and I-CLIP $_{\\mathrm{txt}}$ modules in Fig. 3a. Once they are trained to embed the visual change and the instruction into the same feature space, we then use the approach of DeCap [9] to translate the features into text instructions as described next.", + "bbox": [ + 511, + 478, + 905, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Predicting instructions by decoding features", + "text_level": 1, + "bbox": [ + 511, + 571, + 885, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning the alignment between original/edited image pairs and instructions does not automatically result in better instructions, since a decoder is required to translate I-CLIP latent-space features into actual text instructions. To do this, we leverage the approach of DeCap [9], where we finetune pre-trained CLIP and GPT-2 decoding head [16] for our application to match the original edit instruction $p$ by either decoding $z^{\\mathrm{vis}} = \\mathrm{I - CLIP}_{\\mathrm{vis}}(I^{o},I^{e})$ or $z^{\\mathrm{txt}} = \\mathrm{I - CLIP}_{\\mathrm{txt}}(p)$ .", + "bbox": [ + 511, + 595, + 905, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the original edit instructions $p$ in the datasets may not match the visible changes encoded in $z^{\\mathrm{vis}}$ , we do not want to force the model to decode $z^{\\mathrm{vis}}$ directly to $p$ during training. Instead, we train it by decoding $z^{\\mathrm{txt}}$ back to $p$ , using the same approach as DeCap [9]. Formally, let $p = [c_1, c_2, \\dots, c_n]$ be an edit instruction's token representation of length $n$ , where each token $c_i$ is a one-hot vector. Similarly, the decoded/predicted instruction is $p' = [c_1', c_2', \\dots, c_n']$ (instructions can be represented by the same number of tokens through padding or truncation). This is basically a classification problem that requires identifying the correct \"element\" for each token, so the training", + "bbox": [ + 511, + 719, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28516", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "objective of the text decoder is a cross-entropy loss:", + "bbox": [ + 89, + 90, + 434, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D e C a p}} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\sum_ {c \\in C} y _ {i} ^ {p} (c) \\log y _ {i} ^ {p ^ {\\prime}} (c), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 116, + 482, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C$ is the set of all possible tokens, $y_{i}^{p}(c)$ is the ground-truth for each token $c$ , and $y_{i}^{p^{\\prime}}(c)$ is the probability of token $c$ at position $i$ for $p^{\\prime}$ as predicted by the model.", + "bbox": [ + 89, + 166, + 482, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At inference time, our goal is to generate improved instructions that accurately capture the visual changes from the original image to the edited one by decoding $z^{\\mathrm{vis}}$ . However, since the average cosine similarity between visual and text CLIP features is relatively low, even when well matched (around 0.2) [9], a text decoder that has been trained on text features $z^{\\mathrm{txt}}$ will struggle to handle $z^{\\mathrm{vis}}$ due to such differences, producing sub-optimal decoded results.", + "bbox": [ + 89, + 215, + 483, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this, we would like the decoder to get as input a feature representation of $z^{\\mathrm{vis}}$ that is more similar to the text features the instruction decoder was trained on. A simple approach would be to compute the cosine similarity between $z^{\\mathrm{vis}}$ and each instruction's text feature in the dataset, selecting the feature with the highest similarity. However, this method would simply retrieve the most similar instruction, limiting both the diversity and accuracy of refined instructions. Instead, we follow the approach of DeCap [9], which leverages information from all text features by weighting their influence by their cosine similarity to $z^{\\mathrm{vis}}$ . Essentially, instructions with text features more similar to $z^{\\mathrm{vis}}$ should contribute more to the refined instruction.", + "bbox": [ + 89, + 335, + 483, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, we calculate the probability of each instruction feature contributing to the refined representation using the softmax function over the cosine similarities between all text features and $z^{\\mathrm{vis}}$ . Formally, for a dataset of size $n$ , let $\\{z_i^{\\mathrm{txt}} = \\mathrm{I - CLIP}_{\\mathrm{txt}}(p_i) \\mid i \\in [1,n]\\}$ be the set of text features, each corresponding to an instruction $p_i$ in the dataset. The probability $w_i$ that instruction $p_i$ 's text feature influences the refined instruction for $z^{\\mathrm{vis}}$ is defined as:", + "bbox": [ + 89, + 532, + 483, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} = \\frac {\\exp \\left(\\operatorname {s i m} \\left(z _ {i} ^ {\\mathrm {t x t}} , z ^ {\\mathrm {v i s}}\\right)\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(\\operatorname {s i m} \\left(z _ {j} ^ {\\mathrm {t x t}}, z ^ {\\mathrm {v i s}}\\right)\\right)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 660, + 482, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use these weights to project $z^{\\mathrm{vis}}$ to the text feature space:", + "bbox": [ + 89, + 707, + 482, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(z ^ {\\text {v i s}}\\right) ^ {\\prime} = \\sum_ {i = 1} ^ {n} \\left(w _ {i} \\cdot z _ {i} ^ {\\text {t x t}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 734, + 482, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the end, the refined instruction for $(I^o,I^e)$ is $p^\\prime =$ DeCap $\\left((z^{\\mathrm{vis}})^{\\prime}\\right)$ , where DeCap(·) denotes the instruction decoder, and so the refined dataset sample is $(I^{o},I^{e},p^{\\prime})$", + "bbox": [ + 89, + 777, + 483, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Working in the latent-diffusion domain", + "text_level": 1, + "bbox": [ + 89, + 833, + 428, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With I-CLIP and DeCap, we can now generate datasets with improved semantic alignment between original/edited image changes and their edit instructions (see results in Fig. 2),", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/baa4c8b08e2749f695ecee1cf8df1addad361b10a07932bde8371db131c74e59.jpg", + "image_caption": [ + "Figure 4. Training our LD-DINOv2 model. To use I-CLIP as part of the training objective for Stable Diffusion [18], it needs to handle noisy latent images. Therefore, we replace the original DINOv2 backbone in Fig. 3b with a latent-diffusion version of it we call LD-DINOv2, which takes both the noisy latent image $\\tilde{L}_k$ from SD VAE encoding and forward-diffusion (FD) timestep $t_k$ . We then train LD-DINOv2 to \"ignore\" the noise and the latent-space compression and to extract the original DINOv2 features using the training objective $\\mathcal{L}_{\\mathrm{LD - DINO_{v2}}}$ (Eq. 7)." + ], + "image_footnote": [], + "bbox": [ + 514, + 90, + 901, + 164 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "thereby providing a better \"signal\" for training instruction-guided, image-editing models. So one might try to simply train models like InstructPix2Pix [1] directly as-is on our improved dataset. While this does lead to some improvement (see Sec. 4.5), we find we get even more improvement if we reinforce the alignment between the visual change and the edit instruction during the training of the stable diffusion (SD) model itself. This means using I-CLIP as an integral part of the training objective beyond just dataset refinement.", + "bbox": [ + 511, + 333, + 905, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To do this, I-CLIP must directly handle noisy, latent images in SD, defined as $\\tilde{L}_k = \\mathrm{FD}(\\mathrm{VAE}_{\\mathrm{enc}}(I), N, t_k)$ . Here, $\\mathrm{FD}(\\cdot)$ is the $k^{\\mathrm{th}}$ step of the forward-diffusion process [19] which takes in image $I$ , embeds it in the latent domain with SD's variational autoencoder $\\mathrm{VAE}_{\\mathrm{enc}}(I)$ , and then adds randomly sampled noise $N \\sim \\mathcal{N}(0, \\mathbf{I})$ with respect to timestep $t_k$ . Our DINOv2 feature backbone in I-CLIP, which we call LD-DINOv2 (see Fig. 4), works with $\\tilde{L}_k$ directly by also taking the corresponding timestep $t_k$ as input and is trained to \"ignore\" the noise and the latent-space compression to extract the original DINOv2 features with the following training objective:", + "bbox": [ + 511, + 468, + 906, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L D - D I N O} _ {\\mathrm {v} 2}} = 1 - \\operatorname {s i m} \\left(d ^ {\\mathrm {L D}}, d\\right) + \\frac {1}{l} \\sum_ {i = 1} ^ {l} \\left(1 - \\operatorname {s i m} \\left(d _ {i} ^ {\\mathrm {L D}}, d _ {i}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 657, + 906, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d^{\\mathrm{LD}}$ is the output of LD-DINOv2, $d$ is the original DINOv2 output, and the intermediate features from each of its $l$ layers are $\\{d_i^{\\mathrm{LD}}\\mid i\\in [1,l]\\}$ . The visual feature of the original-edited image pair $(I^{o},I^{e})$ then becomes:", + "bbox": [ + 511, + 702, + 906, + 765 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz ^ {\\mathrm {v i s}} = \\mathrm {I - C L I P} _ {\\mathrm {v i s}} \\left(\\tilde {L} _ {k _ {1}} ^ {o}, t _ {k _ {1}}, \\tilde {L} _ {k _ {2}} ^ {e}, t _ {k _ {2}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 772, + 906, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where:", + "bbox": [ + 513, + 800, + 563, + 813 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {L} _ {k _ {1}} ^ {o} = \\mathrm {F D} \\left(\\mathrm {V A E} _ {\\text {e n c}} \\left(I ^ {o}\\right), N _ {1}, t _ {k _ {1}}\\right) \\\\ \\tilde {\\sim} \\end{array} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 821, + 903, + 848 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {L} _ {k _ {2}} ^ {e} = \\mathrm {F D} \\left(\\mathrm {V A E} _ {\\mathrm {e n c}} \\left(I ^ {e}\\right), N _ {2}, t _ {k _ {2}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 844, + 818, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are generated using randomly sampled Gaussian noise terms $N_{1}$ and $N_{2}$ . Note that although for diffusion-based", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28517", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "image editing training, no noise is added to the original image, we noisify the original image to train I-CLIP $_{\\mathrm{vis}}$ to make it more robust. For further details on the LD-DINOv2 training process, please refer to the supplementary. Note that we set $k_{1} = k_{2} = 0$ during dataset refinement stage so LD-DINOv2 will process the pristine latent images.", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Training our instruction-guided editing model", + "text_level": 1, + "bbox": [ + 89, + 195, + 480, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At this point, we would like to leverage both the refined dataset and I-CLIP to enhance SD-based instruction-guided image editing methods, such as InstructPix2Pix (IP2P) [1].", + "bbox": [ + 89, + 218, + 483, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given data sample $(I^o, I^e, p)$ , IP2P usually trains a denoising UNet to predict the noise added to the edited image:", + "bbox": [ + 89, + 265, + 483, + 296 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {N} _ {k} = \\mathrm {U N e t} _ {\\mathrm {I P 2 P}} \\left(L ^ {o}, \\tilde {L} _ {k} ^ {e}, t _ {k}, p\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 311, + 483, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which is conditioned on the original image in the latent domain $(L^o)$ and the edit instruction $p$ . The traditional training loss in this setup is the mean squared error (MSE) with respect to the actual noise $N$ added to the GT edited image:", + "bbox": [ + 89, + 340, + 483, + 402 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M S E}} = \\left\\| N - \\tilde {N} _ {k} \\right\\| _ {2} ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 417, + 483, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given refined sample $(I^o, I^e, p')$ , we design a complementary loss function to incorporate our Instruct-CLIP guidance into the objective and force the visual change to match the refined edit instruction:", + "bbox": [ + 89, + 455, + 483, + 517 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {I - C L I P}} = 1 - \\operatorname {s i m} \\left(\\mathrm {I - C L I P} _ {\\text {v i s}} \\left(L ^ {o}, t _ {0}, \\tilde {L} _ {k - 1} ^ {e}, t _ {k - 1}\\right), \\right. \\tag {12} \\\\ \\mathrm {I} - \\operatorname {C L I P} _ {\\text {t x t}} \\left(p ^ {\\prime}\\right)), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 527, + 483, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the intermediate, denoised latent output using the predicted noise is calculated as:", + "bbox": [ + 89, + 580, + 483, + 612 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {L} _ {k - 1} ^ {e} = \\mathrm {R D} _ {k, k - 1} \\left(\\tilde {L} _ {k} ^ {e}, \\tilde {N} _ {k}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 625, + 482, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\mathrm{RD}_{k,k - 1}(\\tilde{L}_k^e,\\tilde{N}_k)$ is the reverse-diffusion process. Thus, our final training objective is:", + "bbox": [ + 89, + 657, + 482, + 689 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M S E}} + \\lambda \\mathcal {L} _ {\\mathrm {I - C L I P}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 704, + 482, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is set to 0.1 to balance $\\mathcal{L}_{\\mathrm{I - CLIP}}$ with the MSE loss.", + "bbox": [ + 89, + 734, + 477, + 751 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 768, + 223, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implemented and trained Instruct-CLIP (I-CLIP) as described above and used it to refine the InstructPix2Pix (IP2P) dataset [1] to get $120\\mathrm{K}+$ refined instructions, which took around 10 hours on two A6000 GPUs. We then used this refined dataset to fine-tune the InstructPix2Pix model. Please refer to the supplementary for more details. We now describe the experiments to test our method.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Baselines and benchmarks", + "text_level": 1, + "bbox": [ + 513, + 90, + 756, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our editing model with several baselines: InstInpaint (I-Inp) [24], MagicBrush (MagBr), HIVE [26], InstructPix2Pix (IP2P) [1], Watch Your Steps (WYS) [11], and ZONE [8] (see Sec. 2). We evaluate methods on two instruction-guided image-editing benchmarks: MagicBrush (MagBr) [25] and ZONE [8]. The former contains multi-turn edits where multiple instructions are used to edit one image iteratively, as opposed to single-turn edit where the image is edited once. We compare results quantitatively with the following metrics used in prior work [1, 8, 11, 25]:", + "bbox": [ + 511, + 112, + 906, + 263 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CLIP-T = sim(CLIPvis $(I^{e})^{\\prime}$ , CLIPtxt $(p^{e})$ )", + "- CLIP-I = sim(CLIPvis((Ie)'), CLIPvis(Ie))", + "DINO-I $=$ sim(DINOv2((Ie)'),DINOv2(Ie))" + ], + "bbox": [ + 513, + 263, + 826, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $(I^e)'$ denotes the method output corresponding to a benchmark sample $(I^o, I^e, p^o, p^e)$ , where $I^o$ is the original image, $I^e$ is the ground-truth edited image (available for MagBr, not for ZONE), $p^o$ is the caption for the original image, and $p^e$ is the intended caption for the edited image. The CLIP-T score assesses semantic alignment between the result and its intended caption in the benchmark while the CLIP and DINO similarity scores (CLIP-I and DINO-I, respectively) evaluate the visual alignment between the result and the ground-truth, if available. However, we note that as shown in Fig. 5 and in the supplemental material, these metrics do not always match the visual quality of the image edits, but are presented here for completeness.", + "bbox": [ + 511, + 309, + 906, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Instruction refinement results", + "text_level": 1, + "bbox": [ + 513, + 515, + 777, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present samples from our refined dataset in Fig. 2 (see supplementary for more samples). As we can see, Instruct-CLIP is able to correct wrong instructions such as \"make the waves into a hurricane\" to \"add a lightning storm to the sky\" ( $1^{st}$ row right). By correcting these samples, we reduce the noise due to instructions not reflecting the actual edits, which in turn helps improve the performance of the model trained on this dataset as we will see next.", + "bbox": [ + 511, + 537, + 906, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Image-editing results", + "text_level": 1, + "bbox": [ + 513, + 667, + 712, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We found a diverse set of samples that showcase the strength of our method and compared it to baselines in Figs. 1 and 5 (see supplemental for more). Our model not only addresses many issues present in IP2P – such as unintended changes to regions irrelevant to the instructions – but also generates results aligned more accurately with the edit instructions. This is particularly important for multi-turn editing to avoid results diverse further from the desired outcome across edit turns as shown in the supplementary. Notably, this is achieved without manually created datasets or masking mechanisms to explicitly locate the editing region, as used in other approaches.", + "bbox": [ + 511, + 689, + 906, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we present the CLIP-T value of each output in Fig. 5, with the best value per sample underlined. As", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28518", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ed0387408d761004085dcd7a34b369491737693aba35da7da49fcfcc1b204dd4.jpg", + "image_caption": [ + "Figure 5. Comparison with state-of-the-art approaches for instruction-guided image editing, including HIVE [26], Inst-Inpaint (I-Inp) [24], Watch Your Steps (WYS) [11], ZONE [8], MagicBrush (MagBr) [25], InstructPix2Pix (IP2P) [1] showcasing the strength of our approach. CLIP-T value of each output is shown at its top-left corner, with the best value per row underlined. Note that the image with the best CLIP-T score is not necessarily the visually best result, underscoring the deficiencies of conventional metrics (including CLIP-I and DINO-I shown in the supplemental) for measuring the quality of image edits." + ], + "image_footnote": [], + "bbox": [ + 130, + 89, + 859, + 815 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28519", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cb82873f1f933285add237eb110bd952d67866c74d3c5ca67f695196523455f2.jpg", + "table_caption": [ + "MagBr data (single-turn)", + "MagBr data (multi-turn) ZONE data" + ], + "table_footnote": [], + "table_body": "
CLIP-T ↑CLIP-I ↑DINO-I ↑CLIP-T ↑CLIP-I ↑DINO-I ↑CLIP-T ↑
HIVE0.3030.8920.7460.2990.8520.6590.297
I-Inp0.2850.8870.7290.2770.8590.6540.267
MagBr0.3070.9290.8360.3030.8960.7590.292
WYS0.3130.9240.8150.3130.8870.7270.301
ZONE0.3010.9290.8240.3070.8960.7500.296
IP2P0.3000.8540.6450.2980.8240.5730.296
Ours0.3050.9110.8030.3010.8710.7210.297
(+1.67%)(+6.67%)(+24.5%)(+1.01%)(+5.70%)(+25.83%)(+0.34%)
", + "bbox": [ + 94, + 99, + 478, + 198 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/344ed3b82f10c7b9fc8ff1449b792bc736f359f3cbe464d6455ba466da4c9018.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with baselines. An up arrow (↑) indicates higher values are better. The percentage improvements over baselines are in parentheses." + ], + "table_footnote": [], + "table_body": "
vs. MagBr#%vs. IP2P#%IP2P vs. MagBr#%
ours win56654.16%ours win31029.67%IP2P win48546.41%
tie16816.08%tie61158.47%tie14814.16%
MagBr win31129.76%IP2P win12411.87%MagBr win41239.43%
", + "bbox": [ + 94, + 258, + 478, + 306 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7675a1b14c64964c8309a1e9e6e7778d5ce4d72518a5a9c4d07ad9aff7ea53b7.jpg", + "table_caption": [ + "Table 2. Method pairwise comparison responses from user study." + ], + "table_footnote": [], + "table_body": "
MagBr dataZONE data
CLIP-T ↑CLIP-I ↑DINO-I ↑CLIP-T ↑
IP2P0.3000.8540.6450.296
Ours (w/o data)0.2990.8620.6710.295
Ours (w/o loss)0.3030.9030.7820.298
Ours0.3050.9110.8030.297
", + "bbox": [ + 102, + 337, + 465, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study. An up arrow $(\\uparrow)$ means higher values are better. We trained two variants for our model: one trained on the original IP2P dataset without refined instructions (\"Ours (w/o data)\") and the other trained on our refined dataset without the Instruct-CLIP guidance loss (\"Ours (w/o loss)\"). We can see both the refined data and the loss help improve results.", + "bbox": [ + 89, + 431, + 483, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shown, these values – as well as the CLIP-I and DINO-I values provided in the supplementary material – are not indicative of visual quality, but are included for completeness. Table 1 provides a quantitative comparison of our model in both single-turn and multi-turn edit scenarios.", + "bbox": [ + 89, + 532, + 483, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. User study", + "text_level": 1, + "bbox": [ + 89, + 619, + 210, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also conducted a user study with 104 participants (60 male, 43 female, 1 non-binary) to compare our method with MagicBrush (MagBr) [25] and InstructPix2Pix (IP2P) [1] on the InstructBrush benchmark [28], ensuring fairness since no method was trained on similar data. Each person evaluated 33 randomly sampled output pairs based on the input image and edit instruction, and select which was better or if they were similar. We balanced the study across all pairwise comparisons, with 1,045 responses per pair. As shown in Table 2, participants overall found our method better than IP2P (17.8% more wins) and MagBr (24.4%).", + "bbox": [ + 89, + 641, + 483, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 818, + 240, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess the impact of refined instructions and our Instruct-CLIP guidance loss, we trained two model variants: one on the original IP2P dataset with the Instruct-CLIP guidance loss (\"Ours (w/o data)) and another on our re", + "bbox": [ + 89, + 840, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Original", + "bbox": [ + 535, + 89, + 576, + 98 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "IP2P", + "bbox": [ + 617, + 89, + 642, + 98 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ours (w/o data)", + "bbox": [ + 673, + 89, + 746, + 98 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ours (w/o loss)", + "bbox": [ + 750, + 89, + 823, + 98 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ours", + "bbox": [ + 852, + 89, + 877, + 97 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/114f37cd7382922a5459e3bb0cab72b89c4279caeff9c7aca7fc3a73c313a867.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 99, + 593, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1164a49dd8c5332dc26641af3fc7b1865d55e7ddf7ed43eb74d28d2085fa818e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 99, + 669, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e158995cd06d2fdb580edc5b8ab12456ffd74b2a3a908eb2173167d601f6c2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 99, + 746, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96a48d934089c6de468a5f8221de8bab1a4f9f0813dca88a2c5b4a0f2b431074.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 99, + 825, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ce8d12ebb8fa2a63cfa40cbde8b24f0b8ff34ff41334cfc0e8a465118c2c06f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 99, + 901, + 157 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"Place a floor lamp next to the chair.\"", + "bbox": [ + 514, + 159, + 712, + 171 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f11d58e9637cd11db9cd21a914e49b9da1cdcc52d259a00446426cada90f0ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 174, + 591, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f5d11a7dd568828f689e01b9f6eea4683d09a4133f3048193e944d15c1e2b94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 174, + 669, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6e99f324cebd2924f4db5dd6da33ce11110896a658f163d3ec6ea30162872b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 174, + 746, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2809d23df25649f9773a2c22c7db7f57488f17d2dd9f6ebf60096e0306244184.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 174, + 825, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd15791da0232bba74907688794d3148c1ca36913d7369576f610384f3b4089e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 174, + 901, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"Replace the truck with a tractor\"", + "bbox": [ + 514, + 234, + 694, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6. Effect of refined instructions and Instruct-CLIP guidance loss. Compared to variants of our model trained without refined instructions (\"Ours (w/o data)\") or without the guidance loss (\"Ours (w/o loss)\"), our model produces superior results.", + "bbox": [ + 511, + 250, + 903, + 304 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ori", + "bbox": [ + 516, + 318, + 555, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "IP2P", + "bbox": [ + 576, + 320, + 607, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MagBr", + "bbox": [ + 607, + 320, + 660, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ori", + "bbox": [ + 718, + 320, + 741, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "IP2P", + "bbox": [ + 764, + 320, + 794, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "IagBr Ours", + "bbox": [ + 815, + 320, + 887, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 519, + 333, + 535, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "②", + "bbox": [ + 535, + 333, + 555, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 558, + 333, + 580, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "${13}/{14}$", + "bbox": [ + 581, + 333, + 604, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "${13}/{14}$", + "bbox": [ + 606, + 333, + 637, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "③", + "bbox": [ + 638, + 333, + 658, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "② ${\\mathrm{{CO}}}_{2}$ : 溶液", + "bbox": [ + 660, + 333, + 681, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "③", + "bbox": [ + 683, + 333, + 704, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1", + "bbox": [ + 705, + 333, + 727, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\frac{3}{1} + u + {4q} = 1 + {uq}$ dH", + "bbox": [ + 728, + 333, + 746, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 750, + 333, + 771, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": ".", + "bbox": [ + 772, + 333, + 803, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "${10} : \\frac{{a}_{n + 1}}{{b}_{n + 1}} = \\frac{2{a}_{n + 1}}{2{b}_{n + 1}} = \\frac{{a}_{1} + {a}_{{2n} + 1}}{{b}_{1} + {b}_{{2n} + 1}} =$", + "bbox": [ + 805, + 333, + 826, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4", + "bbox": [ + 828, + 333, + 849, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "。", + "bbox": [ + 851, + 333, + 870, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Y", + "bbox": [ + 519, + 343, + 535, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 535, + 343, + 555, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Y", + "bbox": [ + 558, + 343, + 580, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 581, + 343, + 604, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m : x = 1$ 或 ${3x} + {4y} + 1 = 0$", + "bbox": [ + 606, + 343, + 637, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A", + "bbox": [ + 638, + 343, + 658, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "${12}/{12}$", + "bbox": [ + 660, + 343, + 681, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "${14} \\times {25}$", + "bbox": [ + 702, + 343, + 723, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 728, + 343, + 746, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\Rightarrow \\;{3}^{x} = - 2 + 2\\sqrt{2} + 2\\sqrt{2} = - 4$", + "bbox": [ + 750, + 343, + 771, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 772, + 343, + 794, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 797, + 343, + 826, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 828, + 343, + 849, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\downarrow x$", + "bbox": [ + 852, + 343, + 872, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "${13}/{14}$", + "bbox": [ + 519, + 353, + 535, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 535, + 353, + 555, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "${13}/{14}$", + "bbox": [ + 558, + 353, + 580, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$ ;", + "bbox": [ + 581, + 353, + 604, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$ ;", + "bbox": [ + 606, + 353, + 637, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 638, + 353, + 658, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "${13}/{14}$", + "bbox": [ + 660, + 353, + 681, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 702, + 353, + 723, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 728, + 353, + 746, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 750, + 353, + 771, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 772, + 353, + 794, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 797, + 353, + 826, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(1)", + "bbox": [ + 828, + 353, + 849, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 852, + 353, + 872, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"Turn orange shirt into blue\"", + "bbox": [ + 519, + 369, + 684, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"Ac", + "bbox": [ + 709, + 369, + 728, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ld a", + "bbox": [ + 730, + 369, + 751, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "n ea", + "bbox": [ + 753, + 369, + 772, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gle", + "bbox": [ + 774, + 369, + 795, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "above", + "bbox": [ + 797, + 369, + 828, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the", + "bbox": [ + 830, + 369, + 848, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 519, + 383, + 535, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 535, + 383, + 555, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 558, + 383, + 580, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 581, + 383, + 604, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 606, + 383, + 637, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 638, + 383, + 658, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 660, + 383, + 681, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 683, + 383, + 704, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\left\\{ {2 : {x}^{\\prime }\\left( x\\right) = {y}^{\\prime }\\left( x\\right) }\\right\\} : y = {13}\\left( x\\right) = {13}{y}^{\\prime }\\left( x\\right)$", + "bbox": [ + 705, + 383, + 736, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 738, + 383, + 759, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\frac{1}{2} =$", + "bbox": [ + 761, + 383, + 782, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$ ;", + "bbox": [ + 785, + 383, + 816, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\therefore m = \\frac{3}{11}$", + "bbox": [ + 818, + 383, + 839, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"Delete the wooden boat\"", + "bbox": [ + 519, + 422, + 669, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"Remove the white plane\"", + "bbox": [ + 709, + 422, + 861, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7. Examples of failure cases compared with Instruct-Pix2Pix (IP2P) [1] and MagicBrush (MagBr) [25] performed on the original (Ori) images.", + "bbox": [ + 511, + 440, + 903, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fined dataset without the guidance loss (\"Ours (w/o loss)\"). We compare these with IP2P and our full method in Table 3, and visual examples are provided in Fig. 6. Results show that refined data significantly improve visual alignment with ground truth, evidenced by higher CLIP and DINO scores, while the guidance loss further enhances this alignment.", + "bbox": [ + 511, + 508, + 905, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Limitations", + "text_level": 1, + "bbox": [ + 511, + 611, + 638, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite being able to correct a lot of inaccurate edit instructions in the training set, our method is still heavily influenced by the limitations of the generative method [6] these training images come from, which include color bleeding outside the intended edit region (Fig. 7, top left), and incomplete object addition (Fig. 7, top right). While our model respects the original images better, it sometimes struggles to remove objects in the original images (Fig. 7, bottom).", + "bbox": [ + 511, + 633, + 905, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 768, + 632, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented Instruct-CLIP, a self-supervised method for instruction-guided image editing that learns the semantic changes between original and edited images to refine edit instructions in datasets. Applying I-CLIP to the InstructPix2Pix dataset yields over 120K refined samples, which we use to fine-tune its model with our I-CLIP-guided loss function and generate better edit results.", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28520", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2Pix: Learning to follow image editing instructions. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 2, 3", + "[3] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 34:8780-8794, 2021. 1, 2", + "[4] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis, march 2024. URL http://arxiv.org/abs/2403.03206, 2024. 2", + "[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4", + "[6] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3, 8", + "[7] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-Edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. 3", + "[8] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, Jianzhuang Liu, et al. ZONE: Zero-shot instruction-guided local editing. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 6254-6263, 2024. 1, 3, 6, 7", + "[9] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. DeCap: Decoding CLIP latents for zero-shot captioning via text-only training. arXiv preprint arXiv:2303.03032, 2023. 2, 4, 5", + "[10] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent Consistency Models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 2", + "[11] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Marcus A Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G Derpanis, and Igor Gilitschenski. Watch Your Steps: Local image and scene editing by text instructions. In European Conference on Computer Vision (ECCV), pages 111-129. Springer, 2025. 1, 3, 6, 7", + "[12] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text Inversion for editing real images using guided diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 6038–6047, 2023. 2", + "[13] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez," + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 3, 4", + "[14] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In SIGGRAPH, pages 1-11, 2023. 2, 3", + "[15] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2", + "[16] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 4", + "[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2, 3", + "[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 5", + "[19] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2, 5", + "[20] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-Play diffusion features for text-driven image-to-image translation. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921–1930, 2023. 2, 3", + "[21] Bram Wallace, Akash Gokul, and Nikhil Naik. EDICT: Exact diffusion inversion via coupled transformations. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 22532-22541, 2023. 3", + "[22] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 1900-1910, 2023. 2, 3", + "[23] Ling Yang, Bohan Zeng, Jiaming Liu, Hong Li, Minghao Xu, Wentao Zhang, and Shuicheng Yan. EditWorld: Simulating world dynamics for instruction-following image editing. arXiv preprint arXiv:2405.14785, 2024. 3", + "[24] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-Inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246, 2023. 1, 3, 6, 7", + "[25] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. MagicBrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1, 3, 6, 7, 8", + "[26] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28521", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Stefano Ermon, Caiming Xiong, and Ran Xu. HIVE: Harnessing human feedback for instructional visual editing. arXiv preprint arXiv:2303.09618, 2023. 2, 3, 6, 7", + "[27] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. 1, 2, 3", + "[28] Ruoyu Zhao, Qingnan Fan, Fei Kou, Shuai Qin, Hong Gu, Wei Wu, Pengcheng Xu, Mingrui Zhu, Nannan Wang, and Xinbo Gao. InstructBrush: Learning attention-based instruction optimization for image editing. arXiv preprint arXiv:2403.18660, 2024. 8" + ], + "bbox": [ + 91, + 90, + 482, + 273 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28522", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/a04fb8db-d9e4-4e5e-a8c0-00327c1bd422_model.json b/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/a04fb8db-d9e4-4e5e-a8c0-00327c1bd422_model.json new file mode 100644 index 0000000000000000000000000000000000000000..65944fa78daa0d1fe41bc19e06a5947acada387c --- /dev/null +++ b/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/a04fb8db-d9e4-4e5e-a8c0-00327c1bd422_model.json @@ -0,0 +1,3454 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.188, + 0.131, + 0.817, + 0.177 + ], + "angle": 0, + "content": "Instruct-CLIP: Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.217, + 0.404, + 0.234 + ], + "angle": 0, + "content": "Sherry X. Chen" + }, + { + "type": "text", + "bbox": [ + 0.469, + 0.217, + 0.554, + 0.233 + ], + "angle": 0, + "content": "Misha Sra" + }, + { + "type": "text", + "bbox": [ + 0.619, + 0.217, + 0.721, + 0.234 + ], + "angle": 0, + "content": "Pradeep Sen" + }, + { + "type": "text", + "bbox": [ + 0.343, + 0.247, + 0.655, + 0.265 + ], + "angle": 0, + "content": "University of California, Santa Barbara" + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.267, + 0.631, + 0.283 + ], + "angle": 0, + "content": "{xchen774, sra, psen}@ucsb.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.485, + 0.754 + ], + "angle": 0, + "content": "Although natural language instructions offer an intuitive way to guide automated image editing, deep-learning models often struggle to achieve high-quality results, largely due to the difficulty of creating large, high-quality training datasets. To do this, previous approaches have typically relied on text-to-image (T2I) generative models to produce pairs of original and edited images that simulate the input/output of an instruction-guided image-editing model. However, these image pairs often fail to align with the specified edit instructions due to the limitations of T2I models, which negatively impacts models trained on such datasets. To address this, we present Instruct-CLIP (I-CLIP), a self-supervised method that learns the semantic changes between original and edited images to refine and better align the instructions in existing datasets. Furthermore, we adapt Instruct-CLIP to handle noisy latent images and diffusion timesteps so that it can be used to train latent diffusion models (LDMs) and efficiently enforce alignment between the edit instruction and the image changes in latent space at any step of the diffusion pipeline. We use Instruct-CLIP to correct the InstructPix2Pix dataset and get over 120K refined samples we then use to fine-tune their model, guided by our novel I-CLIP-based loss function. The resulting model can produce edits that are more aligned with the given instructions. Our code and dataset are available at https://github.com/SherryXTChen/Instruct-CLIP.git." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.784, + 0.222, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Recent advances in instruction-guided image editing have introduced intuitive and powerful tools that require only a single text instruction for image editing [1, 24-27]. In particular, InstructPix2Pix (IP2P) [1] pioneered diffusion-based instruction-driven image editing, serving as the foundation for numerous subsequent works [8, 11, 25]. These" + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.327, + 0.613, + 0.339 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.341, + 0.648, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.327, + 0.721, + 0.339 + ], + "angle": 0, + "content": "IP2P" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.341, + 0.767, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.327, + 0.867, + 0.339 + ], + "angle": 0, + "content": "I-CLIP (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.341, + 0.888, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.395, + 0.731, + 0.409 + ], + "angle": 0, + "content": "\"Change the panda to a monkey\"" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.411, + 0.648, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.411, + 0.767, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.411, + 0.888, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.496, + 0.837, + 0.509 + ], + "angle": 0, + "content": "\"Let the old man be dressed in a military uniform\"" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.512, + 0.648, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.512, + 0.767, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.512, + 0.888, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.566, + 0.779, + 0.58 + ], + "angle": 0, + "content": "\"Change the background to a classroom\"" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.582, + 0.648, + 0.649 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.582, + 0.767, + 0.649 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.582, + 0.888, + 0.649 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.651, + 0.698, + 0.664 + ], + "angle": 0, + "content": "\"The cat needs to be black\"" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.665, + 0.648, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.666, + 0.767, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.666, + 0.888, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.76, + 0.694, + 0.773 + ], + "angle": 0, + "content": "\"Make it a chocolate cake\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.778, + 0.907, + 0.807 + ], + "angle": 0, + "content": "Figure 1. Results showcasing the strength of Instruct-CLIP (I-CLIP) compared to state-of-the-art InstructPix2Pix (IP2P) [1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "methods leverage pre-trained text-to-image (T2I) models [3, 10, 15, 18, 19] and condition the generation process on edit instructions instead of the usual prompts. However, these models still need to be fine-tuned on appropriate datasets to learn how to make instruction-guided edits." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28513" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.09, + 0.168, + 0.103 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image_caption", + "bbox": [ + 0.213, + 0.09, + 0.254, + 0.102 + ], + "angle": 0, + "content": "Edited" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.104, + 0.186, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.187, + 0.104, + 0.28, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.177, + 0.223, + 0.193 + ], + "angle": 0, + "content": "ORIGINAL: \"by andy warhol\" (0.130) \nREFINED: \"make the sky pink\" (0.263)" + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.09, + 0.371, + 0.103 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.105, + 0.387, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.09, + 0.456, + 0.102 + ], + "angle": 0, + "content": "Edited" + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.105, + 0.48, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.177, + 0.478, + 0.193 + ], + "angle": 0, + "content": "ORIGINAL: \"make the waves into a hurricane\" (0.090) \nREFINED: \"add a lightning storm to the sky\" (0.289)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.2, + 0.186, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.187, + 0.2, + 0.28, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.272, + 0.246, + 0.288 + ], + "angle": 0, + "content": "ORIGINAL:\"make it morning\"(0.179) \nREFINED:\"make it a foggy morning\"(0.324)" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.199, + 0.389, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.199, + 0.482, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.272, + 0.437, + 0.288 + ], + "angle": 0, + "content": "ORIGINAL: \"make it a landscape\" (0.090) \nREFINED: \"a photograph\" (0.233)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.295, + 0.186, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.187, + 0.295, + 0.28, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.368, + 0.235, + 0.384 + ], + "angle": 0, + "content": "ORIGINAL: \"make it a daydream\" (0.113) \nREFINED: \"make it spring\" (0.300)" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.294, + 0.382, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.295, + 0.476, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.368, + 0.468, + 0.384 + ], + "angle": 0, + "content": "ORIGINAL: \"make it a video game character\" (0.097) \nREFINED: \"as a 3 d sculpture\" (0.357)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.504 + ], + "angle": 0, + "content": "Figure 2. Problems with existing instruction-guided image-editing datasets [1]. As shown, there are many examples where the dataset's original edit instruction does not match the actual changes in the images. Our I-CLIP approach refines edit instructions to match the visual change better and allows us to train a system that produces better outputs. The values in parentheses are the cosine similarity between the visual change from the original to the edited image and the edit instruction from I-CLIP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.67 + ], + "angle": 0, + "content": "Some have created these datasets by using T2I models to approximate the desired behavior of an instruction-guided image editing system. For example, several approaches [1, 26, 27] leverage Prompt-to-Prompt [6] to generate a pair of images that look like one is an edited version of the other by using a pair of appropriately modified prompts. A separate, fine-tuned large-language model (LLM), such as GPT3 [2], is then used to generate plausible edit instructions from these prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Given the generality of Prompt-to-Prompt and LLMs, such datasets can cover a wide range of instructions. However, the quality of image editing is often limited by the capabilities of the generative methods, resulting in problems when the changes between the original and the edited image are misaligned with the edit instruction (Fig. 2). This affects the performance on models trained on these datasets, as we can see with the IP2P results in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Although improving image-editing datasets would significantly enhance the performance of models trained on them, doing so is difficult. On one hand, manually creating edit instructions and corresponding outputs is extremely labor-intensive and impractical. On the other, it is nontrivial to construct a pipeline that can perform effective edits in order to create an instruction-guided, image-editing" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "dataset. After all, if such a pipeline existed, it would be a successful instruction-guided, image-editing method itself." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.905, + 0.243 + ], + "angle": 0, + "content": "Furthermore, creating a brand new dataset from scratch may not be necessary. Despite the inherent limitations of previous datasets, we observe that they still provide some \"signal\" for training image editing models. After all, as shown in Fig. 1, we see that models trained on them produce results that, while imperfect, show that the system is understanding something about the desired edit. Can we harness this signal to refine the dataset and address its issues?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.905, + 0.349 + ], + "angle": 0, + "content": "Inspired by CLIP [17], which learns rich semantic alignments through contrastive language-image pre-training, we embed original-edited image pairs and their corresponding edit instructions into the same feature space in a similar manner with a neural network, which we call Instruct-CLIP. We then follow this with a separate text decoder to generate more precise instructions from the learned features [9]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.905, + 0.517 + ], + "angle": 0, + "content": "Instruct-CLIP is equipped with a modified DINOv2 [13] backbone capable of handling noisy latent images in Stable Diffusion (SD) [18] so it can not only enhance training data quality in the pre-processing stage but also serve as an efficient learning objective during training. We use Instruct-CLIP to refine instructions in the IP2P dataset [1], resulting in \\(120\\mathrm{K}+\\) unique, enhanced samples (see Fig. 2). The IP2P model is then fine-tuned on this dataset with an Instruct-CLIP-guided loss, enabling the trained instruction-guided editing model to produce results more faithfully aligned with the instructions (see Fig. 5)." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.519, + 0.764, + 0.533 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.535, + 0.905, + 0.564 + ], + "angle": 0, + "content": "- A dual-purpose model that refines dataset instructions and enhances editing model training." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.565, + 0.905, + 0.593 + ], + "angle": 0, + "content": "- A large-scale dataset with over 120K samples featuring more accurate and enriched instructions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.595, + 0.905, + 0.623 + ], + "angle": 0, + "content": "- An instruction-guided image editing method trained on above dataset." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.535, + 0.905, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.641, + 0.655, + 0.657 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.886, + 0.683 + ], + "angle": 0, + "content": "2.1. Text-to-image diffusion-based image editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Diffusion models [3, 4, 10, 15, 18, 19] have set a new standard for image generation quality, and form the basis of various text-to-image (T2I) editing applications. A range of methods [6, 12, 14, 20-22] aim to edit a given image to semantically align with a target prompt, using the input prompt as an anchor. However, traditional T2I models like Stable Diffusion (SD) [18] often yield inconsistent results even with similar prompts, where both subject and context can change significantly. Prompt-to-Prompt [6] addresses this by preserving attention maps for the shared components in the input and the target prompt to improve image visual alignment, but both images are synthesized so it cannot handle arbitrary input images. Null-text inversion [12] overcomes this by reconstructing an input image in SD based on" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28514" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "the input prompt. They do this by optimizing the empty-text as the negative prompt, enabling controlled edits when combined with Prompt-to-Prompt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.483, + 0.364 + ], + "angle": 0, + "content": "Other methods similarly leverage diffusion inversion techniques paired with unique editing mechanisms: Diffusion Disentanglement [22] aims to find the optimal blend of input and target prompt features to find the compromise between visual alignment with the input image and semantic alignment with the target prompt; EDICT [21] uses coupling inversion for improved reconstruction, which in turn enhance editing performance; pix2pix-zero [14] optimizes model attention maps for the target image to match the ones for the input image for visual consistency between the two; and Plug-and-Play [20] integrates self-attention maps from the input image to guide the edited image generation. While each method has its own advantages, a major shared shortcoming is they require prompts for both input and output images, which can be cumbersome for users." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.376, + 0.379, + 0.392 + ], + "angle": 0, + "content": "2.2. Instruction-guided image editing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.483, + 0.655 + ], + "angle": 0, + "content": "New approaches have recently emerged that leverage priors from Stable Diffusion and which offer a more user-friendly approach to image editing, requiring only a single edit instruction [1, 8, 11, 24-26]. For example, InstructPix2Pix (IP2P) [1] replaces the prompt with an edit instruction that conditions the generation of the edited image, where the input image is concatenated as an extra condition as well. Among these instruction-guided image editing methods, HIVE [26] collects user feedback on outputs to train a reward model, which iteratively improves the model's output. Watch Your Step [11] and ZONE [8] both use a masking mechanism – either by measuring the discrepancy between IP2P predictions with and without the instruction, or by employing a region intersection-over-union (IoU) scheme with a segmentation model. This helps them avoid editing instruction-irrelevant regions, resulting in better localized editing capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.338, + 0.683 + ], + "angle": 0, + "content": "2.3. Instruction-guided datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.855 + ], + "angle": 0, + "content": "Regardless what additional control mechanism is introduced to improve instruction-guided, image-editing results, the development of instruction-guided image editing models fundamentally relies on suitable training datasets. However, creating large-scale, high-quality datasets automatically presents a significant challenge, as such a data creation system would effectively constitute an editing method itself. Prior works have addressed this challenge by combining existing synthesis models to approximate the behavior of desired instruction-guided image editing methods, primarily following two distinct approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Several works [1, 26, 27] adopted the first approach, which leverages Prompt-to-Prompt [6] to generate pairs of original/edited images using corresponding prompt pairs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "that describe the content of each image. These systems typically fine-tune a large language model (LLM) like GPT-3 [2] to produce edit instructions from prompt pairs. While this approach generates diverse samples, the dataset quality is inherently constrained by the limitations of both Prompt-to-Prompt and the LLM, often resulting in misalignment between the edit instruction and the actual transformation from the original to the edited image (see Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.363 + ], + "angle": 0, + "content": "The second approach attempts to address these limitations by creating datasets through inpainting [24, 25]. Here, edited ground truth images are generated by inpainting selected regions of original images. Although this technique improves the quality of localized image editing samples, it has two significant drawbacks: 1) it cannot generate samples with global modifications (such as style transfer), and 2) it often requires manual creation of inpainting regions. These limitations not only increase the cost of dataset creation but also substantially restrict the dataset size." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.364, + 0.905, + 0.394 + ], + "angle": 0, + "content": "We now describe our approach to address the limitations in previous work." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.408, + 0.661, + 0.423 + ], + "angle": 0, + "content": "3. Instruct-CLIP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.905, + 0.583 + ], + "angle": 0, + "content": "Our goal is to improve previous instruction-guided image editing methods by enhancing the quality of available datasets [1, 7, 23, 26], which we propose to do by correcting their edit instructions. However, while instructions in these datasets sometimes do not reflect the actual changes between the original and edited images, we observe that the datasets overall still provide enough \"signal\" to train editing models. The main challenge is therefore how to properly leverage this \"signal\" to refine the given edit instructions and consequently improve the overall dataset quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.719 + ], + "angle": 0, + "content": "To this end, we propose Instruct-CLIP (I-CLIP, for short), which embeds the visual change between the original image \\( I^o \\) and the edited result \\( I^e \\) into the same feature space as the corresponding edit instruction \\( p \\) through contrastive learning with a neural network. Our approach is inspired by CLIP [17], which learns the semantic alignment between images and their captions. Likewise, Instruct-CLIP learns the relationship between the visual changes in the original-edit image pair and its edit instruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.855 + ], + "angle": 0, + "content": "Just like in CLIP, our approach has an image encoder that in our case encodes the visual change between the input and edited images (denoted as I-CLIP\\(_{\\mathrm{vis}}(I^o, I^e))\\), and a text encoder that encodes the edit instruction \\((\\mathrm{I - CLIP}_{\\mathrm{txt}}(p))\\). However, unlike the original CLIP image encoder, ours takes both the original and edited images as input so that it can encode their visual difference. As shown in Fig. 3a, we fine-tune the two encoders by computing and minimizing the contrastive loss between them (Eq. 2), similar to CLIP." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Furthermore, while the architecture of I-CLIP\\(_{\\mathrm{txt}}\\) is the same as CLIP, our I-CLIP\\(_{\\mathrm{vis}}\\) adds two shared-weighted DINOv2 [13] modules in front of CLIP\\(_{\\mathrm{vis}}\\), as shown in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28515" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.087, + 0.478, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.215, + 0.456, + 0.228 + ], + "angle": 0, + "content": "(a) Overall Instruct-CLIP system architecture" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.09, + 0.843, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.215, + 0.872, + 0.228 + ], + "angle": 0, + "content": "(b) Architecture of visual-change encoder I-CLIP VIS \\((I^o, I^e)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.235, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Figure 3. Instruct-CLIP architectures. (a) Overview of Instruct-CLIP (I-CLIP), which embeds the visual change in the original/edited images \\( I^o \\) and \\( I^e \\) and the edit instruction \\( p \\) into the same feature space through contrastive loss, \\( \\mathcal{L}_{\\mathrm{contrast}} \\) (Eq. 2). To obtain refined instruction \\( p \\) from its I-CLIP embedding \\( z^{\\mathrm{txt}} \\), we adopt the same approach in DeCap [9] to decode \\( z^{\\mathrm{txt}} \\) back to \\( p \\) using cross-entropy loss, \\( \\mathcal{L}_{\\mathrm{DeCap}} \\) (Eq. 4). At inference time, the text decoder takes the embedded visual change from the original to the edited image \\( (z^{\\mathrm{vis}}) \\) and decodes it to produce a new instruction. Due to the significant cosine similarity gap between \\( z^{\\mathrm{vis}} \\) and \\( z^{\\mathrm{txt}} \\) even when they are well aligned, directly decoding \\( z^{\\mathrm{vis}} \\) leads to suboptimal results. To achieve a representation of \\( z^{\\mathrm{vis}} \\) closer to the text features that the instruction decoder learned during training, we compute \\( (z^{\\mathrm{vis}})' \\) with Eq. 6 and decode it to obtain the refined instruction \\( p' \\), which is used to improve the dataset. (b) The architecture of image encoder I-CLIP \\( \\mathrm{vis} \\) includes two shared-weighted DINOv2 [13] modules in front of a standard CLIP \\( \\mathrm{vis} \\) encoder." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.373, + 0.482, + 0.463 + ], + "angle": 0, + "content": "Fig. 3b. These DINOv2 units extract rich, robust visual features from the input images, allowing \\(\\mathrm{CLIP}_{\\mathrm{vis}}\\) to focus on encoding the difference between the original and edited images. This is essential when we introduce Stable Diffusion's (SD) latent image encoding and diffusion into the network, as explained in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.66 + ], + "angle": 0, + "content": "To describe the architecture of I-CLIP\\(_{\\mathrm{vis}}\\), we write the output of DINOv2 for a given image \\(I\\) as \\(d = \\mathrm{DINO}_{\\mathrm{v2}}(I)\\) and the intermediate features from each of its \\(l\\) layers as \\(\\{d_i \\mid i \\in [1, l]\\}\\). So in I-CLIP\\(_{\\mathrm{vis}}\\), we first we compute \\(d^o\\) and \\(d^e\\) by passing the input image \\(I^o\\) and the edited image \\(I^e\\) through DINOv2, respectively, then compute the DINOv2 feature difference \\(d^e - d^o\\) that we input to CLIP\\(_{\\mathrm{vis}}\\). The intermediate feature differences \\(d_i^e - d_i^o\\) are also added to input of the \\(i\\)th layer of CLIP\\(_{\\mathrm{vis}}\\) before being processed by that layer. We find that this improves model performance in our experiments by providing additional information that may be lost from \\(d^e - d^o\\) when being processed from the \\((i + 1)^{\\mathrm{th}}\\) to \\(l^{\\mathrm{th}}\\) layer, similar to skip connections [5]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.764 + ], + "angle": 0, + "content": "To train Instruct-CLIP, we first initialize \\(\\mathrm{CLIP}_{\\mathrm{vis}}\\) and I-CLIP\\(_{\\mathrm{txt}}\\) with the respective blocks from the pre-trained CLIP model. Then, given a batch of \\(n\\) data samples \\(\\{(I_i^o,I_i^e,p_i)\\mid i\\in [1,n]\\}\\) from an existing instruction-guided image-editing dataset like InstructPix2Pix [1], we first compute the Instruct-CLIP features for both the visual difference and edit instruction:" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.769, + 0.483, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {i} ^ {\\text {v i s}} = \\mathrm {I} - \\operatorname {C L I P} _ {\\text {v i s}} \\left(I _ {i} ^ {o}, I _ {i} ^ {e}\\right) \\\\ z _ {i} ^ {\\mathrm {t x t}} = \\mathrm {I} - \\mathrm {C L I P} _ {\\mathrm {t x t}} (p _ {i}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.462, + 0.825 + ], + "angle": 0, + "content": "and compute the contrastive loss \\(\\mathcal{L}_{\\mathrm{contrast}}\\) between them:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.829, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n t r a s t}} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\log \\frac {\\exp \\left(\\sin \\left(z _ {i} ^ {\\text {v i s}}, z _ {i} ^ {\\text {t x t}}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(\\sin \\left(z _ {i} ^ {\\text {v i s}}, z _ {j} ^ {\\text {t x t}}\\right) / \\tau\\right)} \\right. \\tag {2} \\\\ \\left. + \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(z _ {i} ^ {\\mathrm {t x t}} , z _ {i} ^ {\\mathrm {v i s}}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(\\operatorname {s i m} \\left(z _ {i} ^ {\\mathrm {t x t}}, z _ {j} ^ {\\mathrm {v i s}}\\right) / \\tau\\right)}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.907, + 0.42 + ], + "angle": 0, + "content": "where scalar \\(\\tau\\) is a learnable \"temperature\" parameter that controls the sharpness of the similarity distribution and function \\(\\mathrm{sim}(\\cdot, \\cdot)\\) measures normalized cosine similarity:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.434, + 0.907, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s i m} \\left(z _ {1}, z _ {2}\\right) = \\frac {z _ {1} \\cdot z _ {2}}{\\left\\| z _ {1} \\right\\| \\left\\| z _ {2} \\right\\|}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.906, + 0.554 + ], + "angle": 0, + "content": "We first use this contrastive loss to fine-tune the I-CLIP\\(_{\\mathrm{vis}}\\) and I-CLIP\\(_{\\mathrm{txt}}\\) modules in Fig. 3a. Once they are trained to embed the visual change and the instruction into the same feature space, we then use the approach of DeCap [9] to translate the features into text instructions as described next." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.887, + 0.589 + ], + "angle": 0, + "content": "3.1. Predicting instructions by decoding features" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.718 + ], + "angle": 0, + "content": "Learning the alignment between original/edited image pairs and instructions does not automatically result in better instructions, since a decoder is required to translate I-CLIP latent-space features into actual text instructions. To do this, we leverage the approach of DeCap [9], where we finetune pre-trained CLIP and GPT-2 decoding head [16] for our application to match the original edit instruction \\( p \\) by either decoding \\( z^{\\mathrm{vis}} = \\mathrm{I - CLIP}_{\\mathrm{vis}}(I^{o},I^{e}) \\) or \\( z^{\\mathrm{txt}} = \\mathrm{I - CLIP}_{\\mathrm{txt}}(p) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Since the original edit instructions \\( p \\) in the datasets may not match the visible changes encoded in \\( z^{\\mathrm{vis}} \\), we do not want to force the model to decode \\( z^{\\mathrm{vis}} \\) directly to \\( p \\) during training. Instead, we train it by decoding \\( z^{\\mathrm{txt}} \\) back to \\( p \\), using the same approach as DeCap [9]. Formally, let \\( p = [c_1, c_2, \\dots, c_n] \\) be an edit instruction's token representation of length \\( n \\), where each token \\( c_i \\) is a one-hot vector. Similarly, the decoded/predicted instruction is \\( p' = [c_1', c_2', \\dots, c_n'] \\) (instructions can be represented by the same number of tokens through padding or truncation). This is basically a classification problem that requires identifying the correct \"element\" for each token, so the training" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28516" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.435, + 0.108 + ], + "angle": 0, + "content": "objective of the text decoder is a cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.117, + 0.483, + 0.158 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D e C a p}} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\sum_ {c \\in C} y _ {i} ^ {p} (c) \\log y _ {i} ^ {p ^ {\\prime}} (c), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.167, + 0.483, + 0.216 + ], + "angle": 0, + "content": "where \\( C \\) is the set of all possible tokens, \\( y_{i}^{p}(c) \\) is the ground-truth for each token \\( c \\), and \\( y_{i}^{p^{\\prime}}(c) \\) is the probability of token \\( c \\) at position \\( i \\) for \\( p^{\\prime} \\) as predicted by the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.484, + 0.337 + ], + "angle": 0, + "content": "At inference time, our goal is to generate improved instructions that accurately capture the visual changes from the original image to the edited one by decoding \\( z^{\\mathrm{vis}} \\). However, since the average cosine similarity between visual and text CLIP features is relatively low, even when well matched (around 0.2) [9], a text decoder that has been trained on text features \\( z^{\\mathrm{txt}} \\) will struggle to handle \\( z^{\\mathrm{vis}} \\) due to such differences, producing sub-optimal decoded results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.484, + 0.532 + ], + "angle": 0, + "content": "To address this, we would like the decoder to get as input a feature representation of \\( z^{\\mathrm{vis}} \\) that is more similar to the text features the instruction decoder was trained on. A simple approach would be to compute the cosine similarity between \\( z^{\\mathrm{vis}} \\) and each instruction's text feature in the dataset, selecting the feature with the highest similarity. However, this method would simply retrieve the most similar instruction, limiting both the diversity and accuracy of refined instructions. Instead, we follow the approach of DeCap [9], which leverages information from all text features by weighting their influence by their cosine similarity to \\( z^{\\mathrm{vis}} \\). Essentially, instructions with text features more similar to \\( z^{\\mathrm{vis}} \\) should contribute more to the refined instruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.653 + ], + "angle": 0, + "content": "Therefore, we calculate the probability of each instruction feature contributing to the refined representation using the softmax function over the cosine similarities between all text features and \\( z^{\\mathrm{vis}} \\). Formally, for a dataset of size \\( n \\), let \\( \\{z_i^{\\mathrm{txt}} = \\mathrm{I - CLIP}_{\\mathrm{txt}}(p_i) \\mid i \\in [1,n]\\} \\) be the set of text features, each corresponding to an instruction \\( p_i \\) in the dataset. The probability \\( w_i \\) that instruction \\( p_i \\)'s text feature influences the refined instruction for \\( z^{\\mathrm{vis}} \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.661, + 0.483, + 0.699 + ], + "angle": 0, + "content": "\\[\nw _ {i} = \\frac {\\exp \\left(\\operatorname {s i m} \\left(z _ {i} ^ {\\mathrm {t x t}} , z ^ {\\mathrm {v i s}}\\right)\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(\\operatorname {s i m} \\left(z _ {j} ^ {\\mathrm {t x t}}, z ^ {\\mathrm {v i s}}\\right)\\right)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.708, + 0.483, + 0.725 + ], + "angle": 0, + "content": "We use these weights to project \\( z^{\\mathrm{vis}} \\) to the text feature space:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.735, + 0.483, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\left(z ^ {\\text {v i s}}\\right) ^ {\\prime} = \\sum_ {i = 1} ^ {n} \\left(w _ {i} \\cdot z _ {i} ^ {\\text {t x t}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.484, + 0.827 + ], + "angle": 0, + "content": "In the end, the refined instruction for \\((I^o,I^e)\\) is \\(p^\\prime =\\) DeCap \\(\\left((z^{\\mathrm{vis}})^{\\prime}\\right)\\) , where DeCap(·) denotes the instruction decoder, and so the refined dataset sample is \\((I^{o},I^{e},p^{\\prime})\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.429, + 0.851 + ], + "angle": 0, + "content": "3.2. Working in the latent-diffusion domain" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "With I-CLIP and DeCap, we can now generate datasets with improved semantic alignment between original/edited image changes and their edit instructions (see results in Fig. 2)," + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.091, + 0.903, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.31 + ], + "angle": 0, + "content": "Figure 4. Training our LD-DINOv2 model. To use I-CLIP as part of the training objective for Stable Diffusion [18], it needs to handle noisy latent images. Therefore, we replace the original DINOv2 backbone in Fig. 3b with a latent-diffusion version of it we call LD-DINOv2, which takes both the noisy latent image \\(\\tilde{L}_k\\) from SD VAE encoding and forward-diffusion (FD) timestep \\(t_k\\). We then train LD-DINOv2 to \"ignore\" the noise and the latent-space compression and to extract the original DINOv2 features using the training objective \\(\\mathcal{L}_{\\mathrm{LD - DINO_{v2}}}\\) (Eq. 7)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.906, + 0.469 + ], + "angle": 0, + "content": "thereby providing a better \"signal\" for training instruction-guided, image-editing models. So one might try to simply train models like InstructPix2Pix [1] directly as-is on our improved dataset. While this does lead to some improvement (see Sec. 4.5), we find we get even more improvement if we reinforce the alignment between the visual change and the edit instruction during the training of the stable diffusion (SD) model itself. This means using I-CLIP as an integral part of the training objective beyond just dataset refinement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.651 + ], + "angle": 0, + "content": "To do this, I-CLIP must directly handle noisy, latent images in SD, defined as \\(\\tilde{L}_k = \\mathrm{FD}(\\mathrm{VAE}_{\\mathrm{enc}}(I), N, t_k)\\). Here, \\(\\mathrm{FD}(\\cdot)\\) is the \\(k^{\\mathrm{th}}\\) step of the forward-diffusion process [19] which takes in image \\(I\\), embeds it in the latent domain with SD's variational autoencoder \\(\\mathrm{VAE}_{\\mathrm{enc}}(I)\\), and then adds randomly sampled noise \\(N \\sim \\mathcal{N}(0, \\mathbf{I})\\) with respect to timestep \\(t_k\\). Our DINOv2 feature backbone in I-CLIP, which we call LD-DINOv2 (see Fig. 4), works with \\(\\tilde{L}_k\\) directly by also taking the corresponding timestep \\(t_k\\) as input and is trained to \"ignore\" the noise and the latent-space compression to extract the original DINOv2 features with the following training objective:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.659, + 0.907, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L D - D I N O} _ {\\mathrm {v} 2}} = 1 - \\operatorname {s i m} \\left(d ^ {\\mathrm {L D}}, d\\right) + \\frac {1}{l} \\sum_ {i = 1} ^ {l} \\left(1 - \\operatorname {s i m} \\left(d _ {i} ^ {\\mathrm {L D}}, d _ {i}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.703, + 0.907, + 0.766 + ], + "angle": 0, + "content": "where \\(d^{\\mathrm{LD}}\\) is the output of LD-DINOv2, \\(d\\) is the original DINOv2 output, and the intermediate features from each of its \\(l\\) layers are \\(\\{d_i^{\\mathrm{LD}}\\mid i\\in [1,l]\\}\\). The visual feature of the original-edited image pair \\((I^{o},I^{e})\\) then becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.773, + 0.907, + 0.793 + ], + "angle": 0, + "content": "\\[\nz ^ {\\mathrm {v i s}} = \\mathrm {I - C L I P} _ {\\mathrm {v i s}} \\left(\\tilde {L} _ {k _ {1}} ^ {o}, t _ {k _ {1}}, \\tilde {L} _ {k _ {2}} ^ {e}, t _ {k _ {2}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.801, + 0.564, + 0.814 + ], + "angle": 0, + "content": "where:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.823, + 0.905, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {L} _ {k _ {1}} ^ {o} = \\mathrm {F D} \\left(\\mathrm {V A E} _ {\\text {e n c}} \\left(I ^ {o}\\right), N _ {1}, t _ {k _ {1}}\\right) \\\\ \\tilde {\\sim} \\end{array} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.845, + 0.82, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\tilde {L} _ {k _ {2}} ^ {e} = \\mathrm {F D} \\left(\\mathrm {V A E} _ {\\mathrm {e n c}} \\left(I ^ {e}\\right), N _ {2}, t _ {k _ {2}}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "are generated using randomly sampled Gaussian noise terms \\(N_{1}\\) and \\(N_{2}\\). Note that although for diffusion-based" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28517" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "image editing training, no noise is added to the original image, we noisify the original image to train I-CLIP\\(_{\\mathrm{vis}}\\) to make it more robust. For further details on the LD-DINOv2 training process, please refer to the supplementary. Note that we set \\(k_{1} = k_{2} = 0\\) during dataset refinement stage so LD-DINOv2 will process the pristine latent images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.196, + 0.481, + 0.213 + ], + "angle": 0, + "content": "3.3. Training our instruction-guided editing model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.219, + 0.484, + 0.265 + ], + "angle": 0, + "content": "At this point, we would like to leverage both the refined dataset and I-CLIP to enhance SD-based instruction-guided image editing methods, such as InstructPix2Pix (IP2P) [1]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.266, + 0.484, + 0.297 + ], + "angle": 0, + "content": "Given data sample \\((I^o, I^e, p)\\), IP2P usually trains a denoising UNet to predict the noise added to the edited image:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.313, + 0.484, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\tilde {N} _ {k} = \\mathrm {U N e t} _ {\\mathrm {I P 2 P}} \\left(L ^ {o}, \\tilde {L} _ {k} ^ {e}, t _ {k}, p\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.404 + ], + "angle": 0, + "content": "which is conditioned on the original image in the latent domain \\((L^o)\\) and the edit instruction \\(p\\). The traditional training loss in this setup is the mean squared error (MSE) with respect to the actual noise \\(N\\) added to the GT edited image:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.418, + 0.484, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M S E}} = \\left\\| N - \\tilde {N} _ {k} \\right\\| _ {2} ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.485, + 0.518 + ], + "angle": 0, + "content": "Given refined sample \\((I^o, I^e, p')\\), we design a complementary loss function to incorporate our Instruct-CLIP guidance into the objective and force the visual change to match the refined edit instruction:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.529, + 0.484, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {I - C L I P}} = 1 - \\operatorname {s i m} \\left(\\mathrm {I - C L I P} _ {\\text {v i s}} \\left(L ^ {o}, t _ {0}, \\tilde {L} _ {k - 1} ^ {e}, t _ {k - 1}\\right), \\right. \\tag {12} \\\\ \\mathrm {I} - \\operatorname {C L I P} _ {\\text {t x t}} \\left(p ^ {\\prime}\\right)), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.613 + ], + "angle": 0, + "content": "where the intermediate, denoised latent output using the predicted noise is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.625, + 0.483, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\tilde {L} _ {k - 1} ^ {e} = \\mathrm {R D} _ {k, k - 1} \\left(\\tilde {L} _ {k} ^ {e}, \\tilde {N} _ {k}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.69 + ], + "angle": 0, + "content": "Here, \\(\\mathrm{RD}_{k,k - 1}(\\tilde{L}_k^e,\\tilde{N}_k)\\) is the reverse-diffusion process. Thus, our final training objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.705, + 0.483, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M S E}} + \\lambda \\mathcal {L} _ {\\mathrm {I - C L I P}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.735, + 0.478, + 0.752 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is set to 0.1 to balance \\(\\mathcal{L}_{\\mathrm{I - CLIP}}\\) with the MSE loss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.769, + 0.224, + 0.786 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We implemented and trained Instruct-CLIP (I-CLIP) as described above and used it to refine the InstructPix2Pix (IP2P) dataset [1] to get \\(120\\mathrm{K}+\\) refined instructions, which took around 10 hours on two A6000 GPUs. We then used this refined dataset to fine-tune the InstructPix2Pix model. Please refer to the supplementary for more details. We now describe the experiments to test our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.757, + 0.106 + ], + "angle": 0, + "content": "4.1. Baselines and benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.907, + 0.264 + ], + "angle": 0, + "content": "We compare our editing model with several baselines: InstInpaint (I-Inp) [24], MagicBrush (MagBr), HIVE [26], InstructPix2Pix (IP2P) [1], Watch Your Steps (WYS) [11], and ZONE [8] (see Sec. 2). We evaluate methods on two instruction-guided image-editing benchmarks: MagicBrush (MagBr) [25] and ZONE [8]. The former contains multi-turn edits where multiple instructions are used to edit one image iteratively, as opposed to single-turn edit where the image is edited once. We compare results quantitatively with the following metrics used in prior work [1, 8, 11, 25]:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.264, + 0.818, + 0.28 + ], + "angle": 0, + "content": "- CLIP-T = sim(CLIPvis \\((I^{e})^{\\prime}\\), CLIPtxt \\((p^{e})\\) )" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.28, + 0.815, + 0.294 + ], + "angle": 0, + "content": "- CLIP-I = sim(CLIPvis((Ie)'), CLIPvis(Ie))" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.295, + 0.828, + 0.31 + ], + "angle": 0, + "content": "DINO-I \\(=\\) sim(DINOv2((Ie)'),DINOv2(Ie))" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.264, + 0.828, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.31, + 0.907, + 0.506 + ], + "angle": 0, + "content": "Here, \\((I^e)'\\) denotes the method output corresponding to a benchmark sample \\((I^o, I^e, p^o, p^e)\\), where \\(I^o\\) is the original image, \\(I^e\\) is the ground-truth edited image (available for MagBr, not for ZONE), \\(p^o\\) is the caption for the original image, and \\(p^e\\) is the intended caption for the edited image. The CLIP-T score assesses semantic alignment between the result and its intended caption in the benchmark while the CLIP and DINO similarity scores (CLIP-I and DINO-I, respectively) evaluate the visual alignment between the result and the ground-truth, if available. However, we note that as shown in Fig. 5 and in the supplemental material, these metrics do not always match the visual quality of the image edits, but are presented here for completeness." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.516, + 0.779, + 0.53 + ], + "angle": 0, + "content": "4.2. Instruction refinement results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.538, + 0.907, + 0.658 + ], + "angle": 0, + "content": "We present samples from our refined dataset in Fig. 2 (see supplementary for more samples). As we can see, Instruct-CLIP is able to correct wrong instructions such as \"make the waves into a hurricane\" to \"add a lightning storm to the sky\" (\\(1^{st}\\) row right). By correcting these samples, we reduce the noise due to instructions not reflecting the actual edits, which in turn helps improve the performance of the model trained on this dataset as we will see next." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.668, + 0.714, + 0.684 + ], + "angle": 0, + "content": "4.3. Image-editing results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.87 + ], + "angle": 0, + "content": "We found a diverse set of samples that showcase the strength of our method and compared it to baselines in Figs. 1 and 5 (see supplemental for more). Our model not only addresses many issues present in IP2P – such as unintended changes to regions irrelevant to the instructions – but also generates results aligned more accurately with the edit instructions. This is particularly important for multi-turn editing to avoid results diverse further from the desired outcome across edit turns as shown in the supplementary. Notably, this is achieved without manually created datasets or masking mechanisms to explicitly locate the editing region, as used in other approaches." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Furthermore, we present the CLIP-T value of each output in Fig. 5, with the best value per sample underlined. As" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28518" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.09, + 0.86, + 0.816 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.824, + 0.908, + 0.896 + ], + "angle": 0, + "content": "Figure 5. Comparison with state-of-the-art approaches for instruction-guided image editing, including HIVE [26], Inst-Inpaint (I-Inp) [24], Watch Your Steps (WYS) [11], ZONE [8], MagicBrush (MagBr) [25], InstructPix2Pix (IP2P) [1] showcasing the strength of our approach. CLIP-T value of each output is shown at its top-left corner, with the best value per row underlined. Note that the image with the best CLIP-T score is not necessarily the visually best result, underscoring the deficiencies of conventional metrics (including CLIP-I and DINO-I shown in the supplemental) for measuring the quality of image edits." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "28519" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.147, + 0.09, + 0.262, + 0.099 + ], + "angle": 0, + "content": "MagBr data (single-turn)" + }, + { + "type": "table_caption", + "bbox": [ + 0.294, + 0.09, + 0.478, + 0.099 + ], + "angle": 0, + "content": "MagBr data (multi-turn) ZONE data" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.1, + 0.48, + 0.199 + ], + "angle": 0, + "content": "
CLIP-T ↑CLIP-I ↑DINO-I ↑CLIP-T ↑CLIP-I ↑DINO-I ↑CLIP-T ↑
HIVE0.3030.8920.7460.2990.8520.6590.297
I-Inp0.2850.8870.7290.2770.8590.6540.267
MagBr0.3070.9290.8360.3030.8960.7590.292
WYS0.3130.9240.8150.3130.8870.7270.301
ZONE0.3010.9290.8240.3070.8960.7500.296
IP2P0.3000.8540.6450.2980.8240.5730.296
Ours0.3050.9110.8030.3010.8710.7210.297
(+1.67%)(+6.67%)(+24.5%)(+1.01%)(+5.70%)(+25.83%)(+0.34%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.202, + 0.483, + 0.245 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with baselines. An up arrow (↑) indicates higher values are better. The percentage improvements over baselines are in parentheses." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.259, + 0.48, + 0.307 + ], + "angle": 0, + "content": "
vs. MagBr#%vs. IP2P#%IP2P vs. MagBr#%
ours win56654.16%ours win31029.67%IP2P win48546.41%
tie16816.08%tie61158.47%tie14814.16%
MagBr win31129.76%IP2P win12411.87%MagBr win41239.43%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.31, + 0.48, + 0.324 + ], + "angle": 0, + "content": "Table 2. Method pairwise comparison responses from user study." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.338, + 0.467, + 0.43 + ], + "angle": 0, + "content": "
MagBr dataZONE data
CLIP-T ↑CLIP-I ↑DINO-I ↑CLIP-T ↑
IP2P0.3000.8540.6450.296
Ours (w/o data)0.2990.8620.6710.295
Ours (w/o loss)0.3030.9030.7820.298
Ours0.3050.9110.8030.297
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.515 + ], + "angle": 0, + "content": "Table 3. Ablation study. An up arrow \\((\\uparrow)\\) means higher values are better. We trained two variants for our model: one trained on the original IP2P dataset without refined instructions (\"Ours (w/o data)\") and the other trained on our refined dataset without the Instruct-CLIP guidance loss (\"Ours (w/o loss)\"). We can see both the refined data and the loss help improve results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.609 + ], + "angle": 0, + "content": "shown, these values – as well as the CLIP-I and DINO-I values provided in the supplementary material – are not indicative of visual quality, but are included for completeness. Table 1 provides a quantitative comparison of our model in both single-turn and multi-turn edit scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.62, + 0.212, + 0.635 + ], + "angle": 0, + "content": "4.4. User study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.807 + ], + "angle": 0, + "content": "We also conducted a user study with 104 participants (60 male, 43 female, 1 non-binary) to compare our method with MagicBrush (MagBr) [25] and InstructPix2Pix (IP2P) [1] on the InstructBrush benchmark [28], ensuring fairness since no method was trained on similar data. Each person evaluated 33 randomly sampled output pairs based on the input image and edit instruction, and select which was better or if they were similar. We balanced the study across all pairwise comparisons, with 1,045 responses per pair. As shown in Table 2, participants overall found our method better than IP2P (17.8% more wins) and MagBr (24.4%)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.241, + 0.835 + ], + "angle": 0, + "content": "4.5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To assess the impact of refined instructions and our Instruct-CLIP guidance loss, we trained two model variants: one on the original IP2P dataset with the Instruct-CLIP guidance loss (\"Ours (w/o data)) and another on our re" + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.09, + 0.577, + 0.099 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "table_caption", + "bbox": [ + 0.618, + 0.09, + 0.643, + 0.099 + ], + "angle": 0, + "content": "IP2P" + }, + { + "type": "table_caption", + "bbox": [ + 0.674, + 0.09, + 0.748, + 0.099 + ], + "angle": 0, + "content": "Ours (w/o data)" + }, + { + "type": "table_caption", + "bbox": [ + 0.751, + 0.09, + 0.825, + 0.099 + ], + "angle": 0, + "content": "Ours (w/o loss)" + }, + { + "type": "table_caption", + "bbox": [ + 0.854, + 0.09, + 0.878, + 0.098 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.1, + 0.594, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.1, + 0.671, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.1, + 0.748, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.1, + 0.826, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.828, + 0.1, + 0.903, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.16, + 0.714, + 0.172 + ], + "angle": 0, + "content": "\"Place a floor lamp next to the chair.\"" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.175, + 0.593, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.175, + 0.67, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.175, + 0.748, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.175, + 0.826, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.828, + 0.175, + 0.903, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.235, + 0.695, + 0.247 + ], + "angle": 0, + "content": "\"Replace the truck with a tractor\"" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.25, + 0.905, + 0.305 + ], + "angle": 0, + "content": "Figure 6. Effect of refined instructions and Instruct-CLIP guidance loss. Compared to variants of our model trained without refined instructions (\"Ours (w/o data)\") or without the guidance loss (\"Ours (w/o loss)\"), our model produces superior results." + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.319, + 0.556, + 0.334 + ], + "angle": 0, + "content": "Ori" + }, + { + "type": "table_caption", + "bbox": [ + 0.578, + 0.321, + 0.608, + 0.333 + ], + "angle": 0, + "content": "IP2P" + }, + { + "type": "table_caption", + "bbox": [ + 0.609, + 0.321, + 0.661, + 0.333 + ], + "angle": 0, + "content": "MagBr" + }, + { + "type": "table_caption", + "bbox": [ + 0.72, + 0.321, + 0.743, + 0.333 + ], + "angle": 0, + "content": "Ori" + }, + { + "type": "table_caption", + "bbox": [ + 0.765, + 0.321, + 0.795, + 0.333 + ], + "angle": 0, + "content": "IP2P" + }, + { + "type": "table_caption", + "bbox": [ + 0.816, + 0.321, + 0.888, + 0.333 + ], + "angle": 0, + "content": "IagBr Ours" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.334, + 0.536, + 0.343 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.334, + 0.557, + 0.343 + ], + "angle": 0, + "content": "②" + }, + { + "type": "table_caption", + "bbox": [ + 0.56, + 0.334, + 0.581, + 0.343 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.583, + 0.334, + 0.605, + 0.343 + ], + "angle": 0, + "content": "\\( {13}/{14} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.607, + 0.334, + 0.638, + 0.343 + ], + "angle": 0, + "content": "\\( {13}/{14} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.64, + 0.334, + 0.659, + 0.343 + ], + "angle": 0, + "content": "③" + }, + { + "type": "table_caption", + "bbox": [ + 0.661, + 0.334, + 0.682, + 0.343 + ], + "angle": 0, + "content": "② \\( {\\mathrm{{CO}}}_{2} \\) : 溶液" + }, + { + "type": "table_caption", + "bbox": [ + 0.684, + 0.334, + 0.705, + 0.343 + ], + "angle": 0, + "content": "③" + }, + { + "type": "table_caption", + "bbox": [ + 0.707, + 0.334, + 0.728, + 0.343 + ], + "angle": 0, + "content": "1" + }, + { + "type": "table_caption", + "bbox": [ + 0.73, + 0.334, + 0.748, + 0.343 + ], + "angle": 0, + "content": "\\( \\frac{3}{1} + u + {4q} = 1 + {uq} \\) dH" + }, + { + "type": "table_caption", + "bbox": [ + 0.751, + 0.334, + 0.772, + 0.343 + ], + "angle": 0, + "content": "#" + }, + { + "type": "table_caption", + "bbox": [ + 0.774, + 0.334, + 0.805, + 0.343 + ], + "angle": 0, + "content": "." + }, + { + "type": "table_caption", + "bbox": [ + 0.807, + 0.334, + 0.827, + 0.343 + ], + "angle": 0, + "content": "\\( {10} : \\frac{{a}_{n + 1}}{{b}_{n + 1}} = \\frac{2{a}_{n + 1}}{2{b}_{n + 1}} = \\frac{{a}_{1} + {a}_{{2n} + 1}}{{b}_{1} + {b}_{{2n} + 1}} = \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.829, + 0.334, + 0.85, + 0.343 + ], + "angle": 0, + "content": "4" + }, + { + "type": "table_caption", + "bbox": [ + 0.852, + 0.334, + 0.872, + 0.343 + ], + "angle": 0, + "content": "。" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.344, + 0.536, + 0.353 + ], + "angle": 0, + "content": "Y" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.344, + 0.557, + 0.353 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.56, + 0.344, + 0.581, + 0.353 + ], + "angle": 0, + "content": "Y" + }, + { + "type": "table_caption", + "bbox": [ + 0.583, + 0.344, + 0.605, + 0.353 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.607, + 0.344, + 0.638, + 0.353 + ], + "angle": 0, + "content": "\\( \\therefore m : x = 1 \\) 或 \\( {3x} + {4y} + 1 = 0 \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.64, + 0.344, + 0.659, + 0.353 + ], + "angle": 0, + "content": "A" + }, + { + "type": "table_caption", + "bbox": [ + 0.661, + 0.344, + 0.682, + 0.353 + ], + "angle": 0, + "content": "\\( {12}/{12} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.704, + 0.344, + 0.724, + 0.353 + ], + "angle": 0, + "content": "\\( {14} \\times {25} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.73, + 0.344, + 0.748, + 0.353 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.751, + 0.344, + 0.772, + 0.353 + ], + "angle": 0, + "content": "\\( \\Rightarrow \\;{3}^{x} = - 2 + 2\\sqrt{2} + 2\\sqrt{2} = - 4 \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.774, + 0.344, + 0.795, + 0.353 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.798, + 0.344, + 0.828, + 0.353 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.83, + 0.344, + 0.85, + 0.353 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.853, + 0.344, + 0.873, + 0.353 + ], + "angle": 0, + "content": "\\( \\downarrow x \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.354, + 0.536, + 0.363 + ], + "angle": 0, + "content": "\\( {13}/{14} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.354, + 0.557, + 0.363 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.56, + 0.354, + 0.581, + 0.363 + ], + "angle": 0, + "content": "\\( {13}/{14} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.583, + 0.354, + 0.605, + 0.363 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\) ;" + }, + { + "type": "table_caption", + "bbox": [ + 0.607, + 0.354, + 0.638, + 0.363 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\) ;" + }, + { + "type": "table_caption", + "bbox": [ + 0.64, + 0.354, + 0.659, + 0.363 + ], + "angle": 0, + "content": "#" + }, + { + "type": "table_caption", + "bbox": [ + 0.661, + 0.354, + 0.682, + 0.363 + ], + "angle": 0, + "content": "\\( {13}/{14} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.704, + 0.354, + 0.724, + 0.363 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.73, + 0.354, + 0.748, + 0.363 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.751, + 0.354, + 0.772, + 0.363 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.774, + 0.354, + 0.795, + 0.363 + ], + "angle": 0, + "content": "#" + }, + { + "type": "table_caption", + "bbox": [ + 0.798, + 0.354, + 0.828, + 0.363 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.83, + 0.354, + 0.85, + 0.363 + ], + "angle": 0, + "content": "(1)" + }, + { + "type": "table_caption", + "bbox": [ + 0.853, + 0.354, + 0.873, + 0.363 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.371, + 0.685, + 0.383 + ], + "angle": 0, + "content": "\"Turn orange shirt into blue\"" + }, + { + "type": "table_caption", + "bbox": [ + 0.71, + 0.371, + 0.729, + 0.382 + ], + "angle": 0, + "content": "\"Ac" + }, + { + "type": "table_caption", + "bbox": [ + 0.731, + 0.371, + 0.752, + 0.382 + ], + "angle": 0, + "content": "ld a" + }, + { + "type": "table_caption", + "bbox": [ + 0.754, + 0.371, + 0.774, + 0.382 + ], + "angle": 0, + "content": "n ea" + }, + { + "type": "table_caption", + "bbox": [ + 0.776, + 0.371, + 0.796, + 0.382 + ], + "angle": 0, + "content": "gle" + }, + { + "type": "table_caption", + "bbox": [ + 0.798, + 0.371, + 0.829, + 0.382 + ], + "angle": 0, + "content": "above" + }, + { + "type": "table_caption", + "bbox": [ + 0.831, + 0.371, + 0.849, + 0.382 + ], + "angle": 0, + "content": "the" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.385, + 0.536, + 0.394 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.385, + 0.557, + 0.394 + ], + "angle": 0, + "content": "#" + }, + { + "type": "table_caption", + "bbox": [ + 0.56, + 0.385, + 0.581, + 0.394 + ], + "angle": 0, + "content": "#" + }, + { + "type": "table_caption", + "bbox": [ + 0.583, + 0.385, + 0.605, + 0.394 + ], + "angle": 0, + "content": "#" + }, + { + "type": "table_caption", + "bbox": [ + 0.607, + 0.385, + 0.638, + 0.394 + ], + "angle": 0, + "content": "" + }, + { + "type": "table_caption", + "bbox": [ + 0.64, + 0.385, + 0.659, + 0.394 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.661, + 0.385, + 0.682, + 0.394 + ], + "angle": 0, + "content": "#" + }, + { + "type": "table_caption", + "bbox": [ + 0.684, + 0.385, + 0.705, + 0.394 + ], + "angle": 0, + "content": "#" + }, + { + "type": "table_caption", + "bbox": [ + 0.707, + 0.385, + 0.738, + 0.394 + ], + "angle": 0, + "content": "\\( \\left\\{ {2 : {x}^{\\prime }\\left( x\\right) = {y}^{\\prime }\\left( x\\right) }\\right\\} : y = {13}\\left( x\\right) = {13}{y}^{\\prime }\\left( x\\right) \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.739, + 0.385, + 0.76, + 0.394 + ], + "angle": 0, + "content": "#" + }, + { + "type": "table_caption", + "bbox": [ + 0.763, + 0.385, + 0.784, + 0.394 + ], + "angle": 0, + "content": "\\( \\frac{1}{2} = \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.786, + 0.385, + 0.817, + 0.394 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\) ;" + }, + { + "type": "table_caption", + "bbox": [ + 0.819, + 0.385, + 0.84, + 0.394 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.424, + 0.671, + 0.435 + ], + "angle": 0, + "content": "\"Delete the wooden boat\"" + }, + { + "type": "table_caption", + "bbox": [ + 0.71, + 0.424, + 0.862, + 0.436 + ], + "angle": 0, + "content": "\"Remove the white plane\"" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.441, + 0.905, + 0.482 + ], + "angle": 0, + "content": "Figure 7. Examples of failure cases compared with Instruct-Pix2Pix (IP2P) [1] and MagicBrush (MagBr) [25] performed on the original (Ori) images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.906, + 0.601 + ], + "angle": 0, + "content": "fined dataset without the guidance loss (\"Ours (w/o loss)\"). We compare these with IP2P and our full method in Table 3, and visual examples are provided in Fig. 6. Results show that refined data significantly improve visual alignment with ground truth, evidenced by higher CLIP and DINO scores, while the guidance loss further enhances this alignment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.612, + 0.64, + 0.626 + ], + "angle": 0, + "content": "4.6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.906, + 0.755 + ], + "angle": 0, + "content": "Despite being able to correct a lot of inaccurate edit instructions in the training set, our method is still heavily influenced by the limitations of the generative method [6] these training images come from, which include color bleeding outside the intended edit region (Fig. 7, top left), and incomplete object addition (Fig. 7, top right). While our model respects the original images better, it sometimes struggles to remove objects in the original images (Fig. 7, bottom)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.633, + 0.785 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We have presented Instruct-CLIP, a self-supervised method for instruction-guided image editing that learns the semantic changes between original and edited images to refine edit instructions in datasets. Applying I-CLIP to the InstructPix2Pix dataset yields over 120K refined samples, which we use to fine-tune its model with our I-CLIP-guided loss function and generate better edit results." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28520" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2Pix: Learning to follow image editing instructions. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.253 + ], + "angle": 0, + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[3] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 34:8780-8794, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.382 + ], + "angle": 0, + "content": "[4] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis, march 2024. URL http://arxiv.org/abs/2403.03206, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.492 + ], + "angle": 0, + "content": "[6] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.483, + 0.549 + ], + "angle": 0, + "content": "[7] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-Edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[8] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, Jianzhuang Liu, et al. ZONE: Zero-shot instruction-guided local editing. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 6254-6263, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[9] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. DeCap: Decoding CLIP latents for zero-shot captioning via text-only training. arXiv preprint arXiv:2303.03032, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[10] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent Consistency Models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[11] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Marcus A Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G Derpanis, and Igor Gilitschenski. Watch Your Steps: Local image and scene editing by text instructions. In European Conference on Computer Vision (ECCV), pages 111-129. Springer, 2025. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[12] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text Inversion for editing real images using guided diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 6038–6047, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.872, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[14] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In SIGGRAPH, pages 1-11, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[15] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[16] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[19] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[20] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-Play diffusion features for text-driven image-to-image translation. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921–1930, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[21] Bram Wallace, Akash Gokul, and Nikhil Naik. EDICT: Exact diffusion inversion via coupled transformations. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 22532-22541, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[22] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 1900-1910, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[23] Ling Yang, Bohan Zeng, Jiaming Liu, Hong Li, Minghao Xu, Wentao Zhang, and Shuicheng Yan. EditWorld: Simulating world dynamics for instruction-following image editing. arXiv preprint arXiv:2405.14785, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.759, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[24] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-Inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.816, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[25] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. MagicBrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28521" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Stefano Ermon, Caiming Xiong, and Ran Xu. HIVE: Harnessing human feedback for instructional visual editing. arXiv preprint arXiv:2303.09618, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[27] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[28] Ruoyu Zhao, Qingnan Fan, Fei Kou, Shuai Qin, Hong Gu, Wei Wu, Pengcheng Xu, Mingrui Zhu, Nannan Wang, and Xinbo Gao. InstructBrush: Learning attention-based instruction optimization for image editing. arXiv preprint arXiv:2403.18660, 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28522" + } + ] +] \ No newline at end of file diff --git a/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/a04fb8db-d9e4-4e5e-a8c0-00327c1bd422_origin.pdf b/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/a04fb8db-d9e4-4e5e-a8c0-00327c1bd422_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d4b3049550affedec1d7911f14bb51fc9f45c0fa --- /dev/null +++ b/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/a04fb8db-d9e4-4e5e-a8c0-00327c1bd422_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15825018cf74600ad8e33c69d9d7853802be1beac15c424e158792b69a594a88 +size 4368828 diff --git a/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/full.md b/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2e0f22d5280b1486080987fadd0efa8c79a2aeed --- /dev/null +++ b/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/full.md @@ -0,0 +1,566 @@ +# Instruct-CLIP: Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning + +Sherry X. Chen + +Misha Sra + +Pradeep Sen + +University of California, Santa Barbara + +{xchen774, sra, psen}@ucsb.edu + +# Abstract + +Although natural language instructions offer an intuitive way to guide automated image editing, deep-learning models often struggle to achieve high-quality results, largely due to the difficulty of creating large, high-quality training datasets. To do this, previous approaches have typically relied on text-to-image (T2I) generative models to produce pairs of original and edited images that simulate the input/output of an instruction-guided image-editing model. However, these image pairs often fail to align with the specified edit instructions due to the limitations of T2I models, which negatively impacts models trained on such datasets. To address this, we present Instruct-CLIP (I-CLIP), a self-supervised method that learns the semantic changes between original and edited images to refine and better align the instructions in existing datasets. Furthermore, we adapt Instruct-CLIP to handle noisy latent images and diffusion timesteps so that it can be used to train latent diffusion models (LDMs) and efficiently enforce alignment between the edit instruction and the image changes in latent space at any step of the diffusion pipeline. We use Instruct-CLIP to correct the InstructPix2Pix dataset and get over 120K refined samples we then use to fine-tune their model, guided by our novel I-CLIP-based loss function. The resulting model can produce edits that are more aligned with the given instructions. Our code and dataset are available at https://github.com/SherryXTChen/Instruct-CLIP.git. + +# 1. Introduction + +Recent advances in instruction-guided image editing have introduced intuitive and powerful tools that require only a single text instruction for image editing [1, 24-27]. In particular, InstructPix2Pix (IP2P) [1] pioneered diffusion-based instruction-driven image editing, serving as the foundation for numerous subsequent works [8, 11, 25]. These + +![](images/4064cbdbe74a34237954bc693f8b510a892fa479a3708327c059f926106889ee.jpg) +Original +"Change the panda to a monkey" + +![](images/b8780646092415adacdf94000231248dd14457fc025d2e19dde7e7f88439ab3f.jpg) +IP2P + +![](images/46c9334dc4c13157fa99244bf8dd8ea965351b7cf6fa0e94ae3e85f51609fd88.jpg) +I-CLIP (Ours) + +![](images/6e143abed567f69ba12f36627f005a62faf7bcedbc1be7290396a325fb841b93.jpg) +"Let the old man be dressed in a military uniform" + +![](images/f02c93c2bdb28c512dfe34789bb85af4167fcf070c3e7247544740ea7ccc0580.jpg) + +![](images/a26de27ca92f9dacfe10dafa91843a707a0860b84e8bcbaad2d0acc39beb3717.jpg) + +![](images/3b0ea394ed1ecb2d2ee52e6bbd7fb303ac3cebf2ab5e8b9e051dc16a017ff95b.jpg) +"Change the background to a classroom" + +![](images/73db55f64e3a6f8c390a3653e874c27b73501d75af6fc082036c8e883641c467.jpg) + +![](images/c128f79eca84f99b119be5e19901d71c36a22cd0cb823af09bb97356ed9a55fe.jpg) + +![](images/5e88fc350301b685886516f25da5eb468528559cdc1c1f6010633428ab642a00.jpg) +"The cat needs to be black" + +![](images/936477d796fc15f5c908ac607c9ee1e1a1c94caba67c2bcb6fdce16c9f58a45a.jpg) + +![](images/3fa513c733c17c20876c10f325c9723ef221f2af50db1ba1dcbd92bb60948d6b.jpg) + +![](images/98bfd82049da715b9ff632e31b1875e64fb5d15b350f34951f148e8dde9c3ba5.jpg) +"Make it a chocolate cake" +Figure 1. Results showcasing the strength of Instruct-CLIP (I-CLIP) compared to state-of-the-art InstructPix2Pix (IP2P) [1]. + +![](images/c9c36dfcf2aea554d3901e78253d552d91192ac51ad1e31cec3d2aa82c757035.jpg) + +![](images/dea6f4ddbe6f812772756c5434f1eb30a14c056b574f2e724512cb72b3da6372.jpg) + +methods leverage pre-trained text-to-image (T2I) models [3, 10, 15, 18, 19] and condition the generation process on edit instructions instead of the usual prompts. However, these models still need to be fine-tuned on appropriate datasets to learn how to make instruction-guided edits. + +![](images/09d638d9d0ba49a534a4105e9a8ef8f61a1809c31300fce056aec719c13bf2d8.jpg) +Original + +![](images/f4527b57e63dbba0757e278c3e5fab77810a05b1895ebc7e2072606b6de95a80.jpg) +Edited +ORIGINAL: "by andy warhol" (0.130) +REFINED: "make the sky pink" (0.263) + +![](images/8c55a7127e7266c2c581611c409d0afcc460377236f8c70b432c7a00f83dfb9f.jpg) +Original +ORIGINAL: "make the waves into a hurricane" (0.090) +REFINED: "add a lightning storm to the sky" (0.289) + +![](images/176d5acdd320062248080f7549f87d837b35d9b864efe44e61fd36ab8457d13c.jpg) +Edited + +![](images/5b3651168b82d2b723ea14240d4ff8d0c5c39a4bae60ad87cc5374ff50fae067.jpg) +ORIGINAL:"make it morning"(0.179) +REFINED:"make it a foggy morning"(0.324) + +![](images/e24e9330e453ab7aa0f016124871ae34d3760fa6666eaee31f26dc1dbe94e86c.jpg) + +![](images/66aea714ea8979115bdf574b4c60a40e7387f9d69cf1b056aebf7cd025be2596.jpg) +ORIGINAL: "make it a landscape" (0.090) +REFINED: "a photograph" (0.233) + +![](images/f53887aa8d2ce70146eb91f7448816ee12ada20f103b1285545512d8b03220c4.jpg) + +![](images/1cdf2f5294b4d86f51f1998406058bacfe1d404ef93b51d91fd69be4abcb047f.jpg) +ORIGINAL: "make it a daydream" (0.113) +REFINED: "make it spring" (0.300) +Figure 2. Problems with existing instruction-guided image-editing datasets [1]. As shown, there are many examples where the dataset's original edit instruction does not match the actual changes in the images. Our I-CLIP approach refines edit instructions to match the visual change better and allows us to train a system that produces better outputs. The values in parentheses are the cosine similarity between the visual change from the original to the edited image and the edit instruction from I-CLIP. + +![](images/63dc5c118c824db7085957c91b1589d77b6becf405792d9ace3cde3170893c81.jpg) + +![](images/7d56f08041ea4297a320ee0bcb98d4d4642b9461d51b265dc0a035d978a375d2.jpg) +ORIGINAL: "make it a video game character" (0.097) +REFINED: "as a 3 d sculpture" (0.357) + +![](images/ac214b009a2003d44d3e81e943e6df22cc94d267a28e359603db28e34cfcb883.jpg) + +Some have created these datasets by using T2I models to approximate the desired behavior of an instruction-guided image editing system. For example, several approaches [1, 26, 27] leverage Prompt-to-Prompt [6] to generate a pair of images that look like one is an edited version of the other by using a pair of appropriately modified prompts. A separate, fine-tuned large-language model (LLM), such as GPT3 [2], is then used to generate plausible edit instructions from these prompts. + +Given the generality of Prompt-to-Prompt and LLMs, such datasets can cover a wide range of instructions. However, the quality of image editing is often limited by the capabilities of the generative methods, resulting in problems when the changes between the original and the edited image are misaligned with the edit instruction (Fig. 2). This affects the performance on models trained on these datasets, as we can see with the IP2P results in Fig. 1. + +Although improving image-editing datasets would significantly enhance the performance of models trained on them, doing so is difficult. On one hand, manually creating edit instructions and corresponding outputs is extremely labor-intensive and impractical. On the other, it is nontrivial to construct a pipeline that can perform effective edits in order to create an instruction-guided, image-editing + +dataset. After all, if such a pipeline existed, it would be a successful instruction-guided, image-editing method itself. + +Furthermore, creating a brand new dataset from scratch may not be necessary. Despite the inherent limitations of previous datasets, we observe that they still provide some "signal" for training image editing models. After all, as shown in Fig. 1, we see that models trained on them produce results that, while imperfect, show that the system is understanding something about the desired edit. Can we harness this signal to refine the dataset and address its issues? + +Inspired by CLIP [17], which learns rich semantic alignments through contrastive language-image pre-training, we embed original-edited image pairs and their corresponding edit instructions into the same feature space in a similar manner with a neural network, which we call Instruct-CLIP. We then follow this with a separate text decoder to generate more precise instructions from the learned features [9]. + +Instruct-CLIP is equipped with a modified DINOv2 [13] backbone capable of handling noisy latent images in Stable Diffusion (SD) [18] so it can not only enhance training data quality in the pre-processing stage but also serve as an efficient learning objective during training. We use Instruct-CLIP to refine instructions in the IP2P dataset [1], resulting in $120\mathrm{K}+$ unique, enhanced samples (see Fig. 2). The IP2P model is then fine-tuned on this dataset with an Instruct-CLIP-guided loss, enabling the trained instruction-guided editing model to produce results more faithfully aligned with the instructions (see Fig. 5). + +In summary, our contributions are: + +- A dual-purpose model that refines dataset instructions and enhances editing model training. +- A large-scale dataset with over 120K samples featuring more accurate and enriched instructions. +- An instruction-guided image editing method trained on above dataset. + +# 2. Related Work + +# 2.1. Text-to-image diffusion-based image editing + +Diffusion models [3, 4, 10, 15, 18, 19] have set a new standard for image generation quality, and form the basis of various text-to-image (T2I) editing applications. A range of methods [6, 12, 14, 20-22] aim to edit a given image to semantically align with a target prompt, using the input prompt as an anchor. However, traditional T2I models like Stable Diffusion (SD) [18] often yield inconsistent results even with similar prompts, where both subject and context can change significantly. Prompt-to-Prompt [6] addresses this by preserving attention maps for the shared components in the input and the target prompt to improve image visual alignment, but both images are synthesized so it cannot handle arbitrary input images. Null-text inversion [12] overcomes this by reconstructing an input image in SD based on + +the input prompt. They do this by optimizing the empty-text as the negative prompt, enabling controlled edits when combined with Prompt-to-Prompt. + +Other methods similarly leverage diffusion inversion techniques paired with unique editing mechanisms: Diffusion Disentanglement [22] aims to find the optimal blend of input and target prompt features to find the compromise between visual alignment with the input image and semantic alignment with the target prompt; EDICT [21] uses coupling inversion for improved reconstruction, which in turn enhance editing performance; pix2pix-zero [14] optimizes model attention maps for the target image to match the ones for the input image for visual consistency between the two; and Plug-and-Play [20] integrates self-attention maps from the input image to guide the edited image generation. While each method has its own advantages, a major shared shortcoming is they require prompts for both input and output images, which can be cumbersome for users. + +# 2.2. Instruction-guided image editing + +New approaches have recently emerged that leverage priors from Stable Diffusion and which offer a more user-friendly approach to image editing, requiring only a single edit instruction [1, 8, 11, 24-26]. For example, InstructPix2Pix (IP2P) [1] replaces the prompt with an edit instruction that conditions the generation of the edited image, where the input image is concatenated as an extra condition as well. Among these instruction-guided image editing methods, HIVE [26] collects user feedback on outputs to train a reward model, which iteratively improves the model's output. Watch Your Step [11] and ZONE [8] both use a masking mechanism – either by measuring the discrepancy between IP2P predictions with and without the instruction, or by employing a region intersection-over-union (IoU) scheme with a segmentation model. This helps them avoid editing instruction-irrelevant regions, resulting in better localized editing capabilities. + +# 2.3. Instruction-guided datasets + +Regardless what additional control mechanism is introduced to improve instruction-guided, image-editing results, the development of instruction-guided image editing models fundamentally relies on suitable training datasets. However, creating large-scale, high-quality datasets automatically presents a significant challenge, as such a data creation system would effectively constitute an editing method itself. Prior works have addressed this challenge by combining existing synthesis models to approximate the behavior of desired instruction-guided image editing methods, primarily following two distinct approaches. + +Several works [1, 26, 27] adopted the first approach, which leverages Prompt-to-Prompt [6] to generate pairs of original/edited images using corresponding prompt pairs + +that describe the content of each image. These systems typically fine-tune a large language model (LLM) like GPT-3 [2] to produce edit instructions from prompt pairs. While this approach generates diverse samples, the dataset quality is inherently constrained by the limitations of both Prompt-to-Prompt and the LLM, often resulting in misalignment between the edit instruction and the actual transformation from the original to the edited image (see Fig. 2). + +The second approach attempts to address these limitations by creating datasets through inpainting [24, 25]. Here, edited ground truth images are generated by inpainting selected regions of original images. Although this technique improves the quality of localized image editing samples, it has two significant drawbacks: 1) it cannot generate samples with global modifications (such as style transfer), and 2) it often requires manual creation of inpainting regions. These limitations not only increase the cost of dataset creation but also substantially restrict the dataset size. + +We now describe our approach to address the limitations in previous work. + +# 3. Instruct-CLIP + +Our goal is to improve previous instruction-guided image editing methods by enhancing the quality of available datasets [1, 7, 23, 26], which we propose to do by correcting their edit instructions. However, while instructions in these datasets sometimes do not reflect the actual changes between the original and edited images, we observe that the datasets overall still provide enough "signal" to train editing models. The main challenge is therefore how to properly leverage this "signal" to refine the given edit instructions and consequently improve the overall dataset quality. + +To this end, we propose Instruct-CLIP (I-CLIP, for short), which embeds the visual change between the original image $I^o$ and the edited result $I^e$ into the same feature space as the corresponding edit instruction $p$ through contrastive learning with a neural network. Our approach is inspired by CLIP [17], which learns the semantic alignment between images and their captions. Likewise, Instruct-CLIP learns the relationship between the visual changes in the original-edit image pair and its edit instruction. + +Just like in CLIP, our approach has an image encoder that in our case encodes the visual change between the input and edited images (denoted as I-CLIP $_{\mathrm{vis}}(I^o, I^e))$ , and a text encoder that encodes the edit instruction $(\mathrm{I - CLIP}_{\mathrm{txt}}(p))$ . However, unlike the original CLIP image encoder, ours takes both the original and edited images as input so that it can encode their visual difference. As shown in Fig. 3a, we fine-tune the two encoders by computing and minimizing the contrastive loss between them (Eq. 2), similar to CLIP. + +Furthermore, while the architecture of I-CLIP $_{\mathrm{txt}}$ is the same as CLIP, our I-CLIP $_{\mathrm{vis}}$ adds two shared-weighted DINOv2 [13] modules in front of CLIP $_{\mathrm{vis}}$ , as shown in + +![](images/a0be5bc45527e087197317b873cf6ba85f5653dd2f1e24de6a8b44d2dc88aefc.jpg) +(a) Overall Instruct-CLIP system architecture + +![](images/d7b49a1efa24227fcfc76507870139ba80ed00b2d61f89d720b6b2887a8d0d31.jpg) +(b) Architecture of visual-change encoder I-CLIP VIS $(I^o, I^e)$ +Figure 3. Instruct-CLIP architectures. (a) Overview of Instruct-CLIP (I-CLIP), which embeds the visual change in the original/edited images $I^o$ and $I^e$ and the edit instruction $p$ into the same feature space through contrastive loss, $\mathcal{L}_{\mathrm{contrast}}$ (Eq. 2). To obtain refined instruction $p$ from its I-CLIP embedding $z^{\mathrm{txt}}$ , we adopt the same approach in DeCap [9] to decode $z^{\mathrm{txt}}$ back to $p$ using cross-entropy loss, $\mathcal{L}_{\mathrm{DeCap}}$ (Eq. 4). At inference time, the text decoder takes the embedded visual change from the original to the edited image $(z^{\mathrm{vis}})$ and decodes it to produce a new instruction. Due to the significant cosine similarity gap between $z^{\mathrm{vis}}$ and $z^{\mathrm{txt}}$ even when they are well aligned, directly decoding $z^{\mathrm{vis}}$ leads to suboptimal results. To achieve a representation of $z^{\mathrm{vis}}$ closer to the text features that the instruction decoder learned during training, we compute $(z^{\mathrm{vis}})'$ with Eq. 6 and decode it to obtain the refined instruction $p'$ , which is used to improve the dataset. (b) The architecture of image encoder I-CLIP $\mathrm{vis}$ includes two shared-weighted DINOv2 [13] modules in front of a standard CLIP $\mathrm{vis}$ encoder. + +Fig. 3b. These DINOv2 units extract rich, robust visual features from the input images, allowing $\mathrm{CLIP}_{\mathrm{vis}}$ to focus on encoding the difference between the original and edited images. This is essential when we introduce Stable Diffusion's (SD) latent image encoding and diffusion into the network, as explained in Sec. 3.2. + +To describe the architecture of I-CLIP $_{\mathrm{vis}}$ , we write the output of DINOv2 for a given image $I$ as $d = \mathrm{DINO}_{\mathrm{v2}}(I)$ and the intermediate features from each of its $l$ layers as $\{d_i \mid i \in [1, l]\}$ . So in I-CLIP $_{\mathrm{vis}}$ , we first we compute $d^o$ and $d^e$ by passing the input image $I^o$ and the edited image $I^e$ through DINOv2, respectively, then compute the DINOv2 feature difference $d^e - d^o$ that we input to CLIP $_{\mathrm{vis}}$ . The intermediate feature differences $d_i^e - d_i^o$ are also added to input of the $i$ th layer of CLIP $_{\mathrm{vis}}$ before being processed by that layer. We find that this improves model performance in our experiments by providing additional information that may be lost from $d^e - d^o$ when being processed from the $(i + 1)^{\mathrm{th}}$ to $l^{\mathrm{th}}$ layer, similar to skip connections [5]. + +To train Instruct-CLIP, we first initialize $\mathrm{CLIP}_{\mathrm{vis}}$ and I-CLIP $_{\mathrm{txt}}$ with the respective blocks from the pre-trained CLIP model. Then, given a batch of $n$ data samples $\{(I_i^o,I_i^e,p_i)\mid i\in [1,n]\}$ from an existing instruction-guided image-editing dataset like InstructPix2Pix [1], we first compute the Instruct-CLIP features for both the visual difference and edit instruction: + +$$ +\begin{array}{l} z _ {i} ^ {\text {v i s}} = \mathrm {I} - \operatorname {C L I P} _ {\text {v i s}} \left(I _ {i} ^ {o}, I _ {i} ^ {e}\right) \\ z _ {i} ^ {\mathrm {t x t}} = \mathrm {I} - \mathrm {C L I P} _ {\mathrm {t x t}} (p _ {i}), \\ \end{array} +$$ + +and compute the contrastive loss $\mathcal{L}_{\mathrm{contrast}}$ between them: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {c o n t r a s t}} = - \frac {1}{n} \sum_ {i = 1} ^ {n} \left(\log \frac {\exp \left(\sin \left(z _ {i} ^ {\text {v i s}}, z _ {i} ^ {\text {t x t}}\right) / \tau\right)}{\sum_ {j = 1} ^ {n} \exp \left(\sin \left(z _ {i} ^ {\text {v i s}}, z _ {j} ^ {\text {t x t}}\right) / \tau\right)} \right. \tag {2} \\ \left. + \log \frac {\exp \left(\operatorname {s i m} \left(z _ {i} ^ {\mathrm {t x t}} , z _ {i} ^ {\mathrm {v i s}}\right) / \tau\right)}{\sum_ {j = 1} ^ {n} \exp \left(\operatorname {s i m} \left(z _ {i} ^ {\mathrm {t x t}}, z _ {j} ^ {\mathrm {v i s}}\right) / \tau\right)}\right), \\ \end{array} +$$ + +where scalar $\tau$ is a learnable "temperature" parameter that controls the sharpness of the similarity distribution and function $\mathrm{sim}(\cdot, \cdot)$ measures normalized cosine similarity: + +$$ +\operatorname {s i m} \left(z _ {1}, z _ {2}\right) = \frac {z _ {1} \cdot z _ {2}}{\left\| z _ {1} \right\| \left\| z _ {2} \right\|}. \tag {3} +$$ + +We first use this contrastive loss to fine-tune the I-CLIP $_{\mathrm{vis}}$ and I-CLIP $_{\mathrm{txt}}$ modules in Fig. 3a. Once they are trained to embed the visual change and the instruction into the same feature space, we then use the approach of DeCap [9] to translate the features into text instructions as described next. + +# 3.1. Predicting instructions by decoding features + +Learning the alignment between original/edited image pairs and instructions does not automatically result in better instructions, since a decoder is required to translate I-CLIP latent-space features into actual text instructions. To do this, we leverage the approach of DeCap [9], where we finetune pre-trained CLIP and GPT-2 decoding head [16] for our application to match the original edit instruction $p$ by either decoding $z^{\mathrm{vis}} = \mathrm{I - CLIP}_{\mathrm{vis}}(I^{o},I^{e})$ or $z^{\mathrm{txt}} = \mathrm{I - CLIP}_{\mathrm{txt}}(p)$ . + +Since the original edit instructions $p$ in the datasets may not match the visible changes encoded in $z^{\mathrm{vis}}$ , we do not want to force the model to decode $z^{\mathrm{vis}}$ directly to $p$ during training. Instead, we train it by decoding $z^{\mathrm{txt}}$ back to $p$ , using the same approach as DeCap [9]. Formally, let $p = [c_1, c_2, \dots, c_n]$ be an edit instruction's token representation of length $n$ , where each token $c_i$ is a one-hot vector. Similarly, the decoded/predicted instruction is $p' = [c_1', c_2', \dots, c_n']$ (instructions can be represented by the same number of tokens through padding or truncation). This is basically a classification problem that requires identifying the correct "element" for each token, so the training + +objective of the text decoder is a cross-entropy loss: + +$$ +\mathcal {L} _ {\mathrm {D e C a p}} = - \frac {1}{n} \sum_ {i = 1} ^ {n} \sum_ {c \in C} y _ {i} ^ {p} (c) \log y _ {i} ^ {p ^ {\prime}} (c), \tag {4} +$$ + +where $C$ is the set of all possible tokens, $y_{i}^{p}(c)$ is the ground-truth for each token $c$ , and $y_{i}^{p^{\prime}}(c)$ is the probability of token $c$ at position $i$ for $p^{\prime}$ as predicted by the model. + +At inference time, our goal is to generate improved instructions that accurately capture the visual changes from the original image to the edited one by decoding $z^{\mathrm{vis}}$ . However, since the average cosine similarity between visual and text CLIP features is relatively low, even when well matched (around 0.2) [9], a text decoder that has been trained on text features $z^{\mathrm{txt}}$ will struggle to handle $z^{\mathrm{vis}}$ due to such differences, producing sub-optimal decoded results. + +To address this, we would like the decoder to get as input a feature representation of $z^{\mathrm{vis}}$ that is more similar to the text features the instruction decoder was trained on. A simple approach would be to compute the cosine similarity between $z^{\mathrm{vis}}$ and each instruction's text feature in the dataset, selecting the feature with the highest similarity. However, this method would simply retrieve the most similar instruction, limiting both the diversity and accuracy of refined instructions. Instead, we follow the approach of DeCap [9], which leverages information from all text features by weighting their influence by their cosine similarity to $z^{\mathrm{vis}}$ . Essentially, instructions with text features more similar to $z^{\mathrm{vis}}$ should contribute more to the refined instruction. + +Therefore, we calculate the probability of each instruction feature contributing to the refined representation using the softmax function over the cosine similarities between all text features and $z^{\mathrm{vis}}$ . Formally, for a dataset of size $n$ , let $\{z_i^{\mathrm{txt}} = \mathrm{I - CLIP}_{\mathrm{txt}}(p_i) \mid i \in [1,n]\}$ be the set of text features, each corresponding to an instruction $p_i$ in the dataset. The probability $w_i$ that instruction $p_i$ 's text feature influences the refined instruction for $z^{\mathrm{vis}}$ is defined as: + +$$ +w _ {i} = \frac {\exp \left(\operatorname {s i m} \left(z _ {i} ^ {\mathrm {t x t}} , z ^ {\mathrm {v i s}}\right)\right)}{\sum_ {j = 1} ^ {n} \exp \left(\operatorname {s i m} \left(z _ {j} ^ {\mathrm {t x t}}, z ^ {\mathrm {v i s}}\right)\right)}. \tag {5} +$$ + +We use these weights to project $z^{\mathrm{vis}}$ to the text feature space: + +$$ +\left(z ^ {\text {v i s}}\right) ^ {\prime} = \sum_ {i = 1} ^ {n} \left(w _ {i} \cdot z _ {i} ^ {\text {t x t}}\right). \tag {6} +$$ + +In the end, the refined instruction for $(I^o,I^e)$ is $p^\prime =$ DeCap $\left((z^{\mathrm{vis}})^{\prime}\right)$ , where DeCap(·) denotes the instruction decoder, and so the refined dataset sample is $(I^{o},I^{e},p^{\prime})$ + +# 3.2. Working in the latent-diffusion domain + +With I-CLIP and DeCap, we can now generate datasets with improved semantic alignment between original/edited image changes and their edit instructions (see results in Fig. 2), + +![](images/baa4c8b08e2749f695ecee1cf8df1addad361b10a07932bde8371db131c74e59.jpg) +Figure 4. Training our LD-DINOv2 model. To use I-CLIP as part of the training objective for Stable Diffusion [18], it needs to handle noisy latent images. Therefore, we replace the original DINOv2 backbone in Fig. 3b with a latent-diffusion version of it we call LD-DINOv2, which takes both the noisy latent image $\tilde{L}_k$ from SD VAE encoding and forward-diffusion (FD) timestep $t_k$ . We then train LD-DINOv2 to "ignore" the noise and the latent-space compression and to extract the original DINOv2 features using the training objective $\mathcal{L}_{\mathrm{LD - DINO_{v2}}}$ (Eq. 7). + +thereby providing a better "signal" for training instruction-guided, image-editing models. So one might try to simply train models like InstructPix2Pix [1] directly as-is on our improved dataset. While this does lead to some improvement (see Sec. 4.5), we find we get even more improvement if we reinforce the alignment between the visual change and the edit instruction during the training of the stable diffusion (SD) model itself. This means using I-CLIP as an integral part of the training objective beyond just dataset refinement. + +To do this, I-CLIP must directly handle noisy, latent images in SD, defined as $\tilde{L}_k = \mathrm{FD}(\mathrm{VAE}_{\mathrm{enc}}(I), N, t_k)$ . Here, $\mathrm{FD}(\cdot)$ is the $k^{\mathrm{th}}$ step of the forward-diffusion process [19] which takes in image $I$ , embeds it in the latent domain with SD's variational autoencoder $\mathrm{VAE}_{\mathrm{enc}}(I)$ , and then adds randomly sampled noise $N \sim \mathcal{N}(0, \mathbf{I})$ with respect to timestep $t_k$ . Our DINOv2 feature backbone in I-CLIP, which we call LD-DINOv2 (see Fig. 4), works with $\tilde{L}_k$ directly by also taking the corresponding timestep $t_k$ as input and is trained to "ignore" the noise and the latent-space compression to extract the original DINOv2 features with the following training objective: + +$$ +\mathcal {L} _ {\mathrm {L D - D I N O} _ {\mathrm {v} 2}} = 1 - \operatorname {s i m} \left(d ^ {\mathrm {L D}}, d\right) + \frac {1}{l} \sum_ {i = 1} ^ {l} \left(1 - \operatorname {s i m} \left(d _ {i} ^ {\mathrm {L D}}, d _ {i}\right)\right), \tag {7} +$$ + +where $d^{\mathrm{LD}}$ is the output of LD-DINOv2, $d$ is the original DINOv2 output, and the intermediate features from each of its $l$ layers are $\{d_i^{\mathrm{LD}}\mid i\in [1,l]\}$ . The visual feature of the original-edited image pair $(I^{o},I^{e})$ then becomes: + +$$ +z ^ {\mathrm {v i s}} = \mathrm {I - C L I P} _ {\mathrm {v i s}} \left(\tilde {L} _ {k _ {1}} ^ {o}, t _ {k _ {1}}, \tilde {L} _ {k _ {2}} ^ {e}, t _ {k _ {2}}\right), \tag {8} +$$ + +where: + +$$ +\begin{array}{l} \tilde {L} _ {k _ {1}} ^ {o} = \mathrm {F D} \left(\mathrm {V A E} _ {\text {e n c}} \left(I ^ {o}\right), N _ {1}, t _ {k _ {1}}\right) \\ \tilde {\sim} \end{array} \tag {9} +$$ + +$$ +\tilde {L} _ {k _ {2}} ^ {e} = \mathrm {F D} \left(\mathrm {V A E} _ {\mathrm {e n c}} \left(I ^ {e}\right), N _ {2}, t _ {k _ {2}}\right), +$$ + +are generated using randomly sampled Gaussian noise terms $N_{1}$ and $N_{2}$ . Note that although for diffusion-based + +image editing training, no noise is added to the original image, we noisify the original image to train I-CLIP $_{\mathrm{vis}}$ to make it more robust. For further details on the LD-DINOv2 training process, please refer to the supplementary. Note that we set $k_{1} = k_{2} = 0$ during dataset refinement stage so LD-DINOv2 will process the pristine latent images. + +# 3.3. Training our instruction-guided editing model + +At this point, we would like to leverage both the refined dataset and I-CLIP to enhance SD-based instruction-guided image editing methods, such as InstructPix2Pix (IP2P) [1]. + +Given data sample $(I^o, I^e, p)$ , IP2P usually trains a denoising UNet to predict the noise added to the edited image: + +$$ +\tilde {N} _ {k} = \mathrm {U N e t} _ {\mathrm {I P 2 P}} \left(L ^ {o}, \tilde {L} _ {k} ^ {e}, t _ {k}, p\right), \tag {10} +$$ + +which is conditioned on the original image in the latent domain $(L^o)$ and the edit instruction $p$ . The traditional training loss in this setup is the mean squared error (MSE) with respect to the actual noise $N$ added to the GT edited image: + +$$ +\mathcal {L} _ {\mathrm {M S E}} = \left\| N - \tilde {N} _ {k} \right\| _ {2} ^ {2}, \tag {11} +$$ + +Given refined sample $(I^o, I^e, p')$ , we design a complementary loss function to incorporate our Instruct-CLIP guidance into the objective and force the visual change to match the refined edit instruction: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {I - C L I P}} = 1 - \operatorname {s i m} \left(\mathrm {I - C L I P} _ {\text {v i s}} \left(L ^ {o}, t _ {0}, \tilde {L} _ {k - 1} ^ {e}, t _ {k - 1}\right), \right. \tag {12} \\ \mathrm {I} - \operatorname {C L I P} _ {\text {t x t}} \left(p ^ {\prime}\right)), \\ \end{array} +$$ + +where the intermediate, denoised latent output using the predicted noise is calculated as: + +$$ +\tilde {L} _ {k - 1} ^ {e} = \mathrm {R D} _ {k, k - 1} \left(\tilde {L} _ {k} ^ {e}, \tilde {N} _ {k}\right). \tag {13} +$$ + +Here, $\mathrm{RD}_{k,k - 1}(\tilde{L}_k^e,\tilde{N}_k)$ is the reverse-diffusion process. Thus, our final training objective is: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {M S E}} + \lambda \mathcal {L} _ {\mathrm {I - C L I P}}, \tag {14} +$$ + +where $\lambda$ is set to 0.1 to balance $\mathcal{L}_{\mathrm{I - CLIP}}$ with the MSE loss. + +# 4. Experiments + +We implemented and trained Instruct-CLIP (I-CLIP) as described above and used it to refine the InstructPix2Pix (IP2P) dataset [1] to get $120\mathrm{K}+$ refined instructions, which took around 10 hours on two A6000 GPUs. We then used this refined dataset to fine-tune the InstructPix2Pix model. Please refer to the supplementary for more details. We now describe the experiments to test our method. + +# 4.1. Baselines and benchmarks + +We compare our editing model with several baselines: InstInpaint (I-Inp) [24], MagicBrush (MagBr), HIVE [26], InstructPix2Pix (IP2P) [1], Watch Your Steps (WYS) [11], and ZONE [8] (see Sec. 2). We evaluate methods on two instruction-guided image-editing benchmarks: MagicBrush (MagBr) [25] and ZONE [8]. The former contains multi-turn edits where multiple instructions are used to edit one image iteratively, as opposed to single-turn edit where the image is edited once. We compare results quantitatively with the following metrics used in prior work [1, 8, 11, 25]: + +- CLIP-T = sim(CLIPvis $(I^{e})^{\prime}$ , CLIPtxt $(p^{e})$ ) +- CLIP-I = sim(CLIPvis((Ie)'), CLIPvis(Ie)) +DINO-I $=$ sim(DINOv2((Ie)'),DINOv2(Ie)) + +Here, $(I^e)'$ denotes the method output corresponding to a benchmark sample $(I^o, I^e, p^o, p^e)$ , where $I^o$ is the original image, $I^e$ is the ground-truth edited image (available for MagBr, not for ZONE), $p^o$ is the caption for the original image, and $p^e$ is the intended caption for the edited image. The CLIP-T score assesses semantic alignment between the result and its intended caption in the benchmark while the CLIP and DINO similarity scores (CLIP-I and DINO-I, respectively) evaluate the visual alignment between the result and the ground-truth, if available. However, we note that as shown in Fig. 5 and in the supplemental material, these metrics do not always match the visual quality of the image edits, but are presented here for completeness. + +# 4.2. Instruction refinement results + +We present samples from our refined dataset in Fig. 2 (see supplementary for more samples). As we can see, Instruct-CLIP is able to correct wrong instructions such as "make the waves into a hurricane" to "add a lightning storm to the sky" ( $1^{st}$ row right). By correcting these samples, we reduce the noise due to instructions not reflecting the actual edits, which in turn helps improve the performance of the model trained on this dataset as we will see next. + +# 4.3. Image-editing results + +We found a diverse set of samples that showcase the strength of our method and compared it to baselines in Figs. 1 and 5 (see supplemental for more). Our model not only addresses many issues present in IP2P – such as unintended changes to regions irrelevant to the instructions – but also generates results aligned more accurately with the edit instructions. This is particularly important for multi-turn editing to avoid results diverse further from the desired outcome across edit turns as shown in the supplementary. Notably, this is achieved without manually created datasets or masking mechanisms to explicitly locate the editing region, as used in other approaches. + +Furthermore, we present the CLIP-T value of each output in Fig. 5, with the best value per sample underlined. As + +![](images/ed0387408d761004085dcd7a34b369491737693aba35da7da49fcfcc1b204dd4.jpg) +Figure 5. Comparison with state-of-the-art approaches for instruction-guided image editing, including HIVE [26], Inst-Inpaint (I-Inp) [24], Watch Your Steps (WYS) [11], ZONE [8], MagicBrush (MagBr) [25], InstructPix2Pix (IP2P) [1] showcasing the strength of our approach. CLIP-T value of each output is shown at its top-left corner, with the best value per row underlined. Note that the image with the best CLIP-T score is not necessarily the visually best result, underscoring the deficiencies of conventional metrics (including CLIP-I and DINO-I shown in the supplemental) for measuring the quality of image edits. + +MagBr data (single-turn) +MagBr data (multi-turn) ZONE data + +
CLIP-T ↑CLIP-I ↑DINO-I ↑CLIP-T ↑CLIP-I ↑DINO-I ↑CLIP-T ↑
HIVE0.3030.8920.7460.2990.8520.6590.297
I-Inp0.2850.8870.7290.2770.8590.6540.267
MagBr0.3070.9290.8360.3030.8960.7590.292
WYS0.3130.9240.8150.3130.8870.7270.301
ZONE0.3010.9290.8240.3070.8960.7500.296
IP2P0.3000.8540.6450.2980.8240.5730.296
Ours0.3050.9110.8030.3010.8710.7210.297
(+1.67%)(+6.67%)(+24.5%)(+1.01%)(+5.70%)(+25.83%)(+0.34%)
+ +Table 1. Quantitative comparison with baselines. An up arrow (↑) indicates higher values are better. The percentage improvements over baselines are in parentheses. + +
vs. MagBr#%vs. IP2P#%IP2P vs. MagBr#%
ours win56654.16%ours win31029.67%IP2P win48546.41%
tie16816.08%tie61158.47%tie14814.16%
MagBr win31129.76%IP2P win12411.87%MagBr win41239.43%
+ +Table 2. Method pairwise comparison responses from user study. + +
MagBr dataZONE data
CLIP-T ↑CLIP-I ↑DINO-I ↑CLIP-T ↑
IP2P0.3000.8540.6450.296
Ours (w/o data)0.2990.8620.6710.295
Ours (w/o loss)0.3030.9030.7820.298
Ours0.3050.9110.8030.297
+ +Table 3. Ablation study. An up arrow $(\uparrow)$ means higher values are better. We trained two variants for our model: one trained on the original IP2P dataset without refined instructions ("Ours (w/o data)") and the other trained on our refined dataset without the Instruct-CLIP guidance loss ("Ours (w/o loss)"). We can see both the refined data and the loss help improve results. + +shown, these values – as well as the CLIP-I and DINO-I values provided in the supplementary material – are not indicative of visual quality, but are included for completeness. Table 1 provides a quantitative comparison of our model in both single-turn and multi-turn edit scenarios. + +# 4.4. User study + +We also conducted a user study with 104 participants (60 male, 43 female, 1 non-binary) to compare our method with MagicBrush (MagBr) [25] and InstructPix2Pix (IP2P) [1] on the InstructBrush benchmark [28], ensuring fairness since no method was trained on similar data. Each person evaluated 33 randomly sampled output pairs based on the input image and edit instruction, and select which was better or if they were similar. We balanced the study across all pairwise comparisons, with 1,045 responses per pair. As shown in Table 2, participants overall found our method better than IP2P (17.8% more wins) and MagBr (24.4%). + +# 4.5. Ablation study + +To assess the impact of refined instructions and our Instruct-CLIP guidance loss, we trained two model variants: one on the original IP2P dataset with the Instruct-CLIP guidance loss ("Ours (w/o data)) and another on our re + +Original + +IP2P + +Ours (w/o data) + +Ours (w/o loss) + +Ours + +![](images/114f37cd7382922a5459e3bb0cab72b89c4279caeff9c7aca7fc3a73c313a867.jpg) + +![](images/1164a49dd8c5332dc26641af3fc7b1865d55e7ddf7ed43eb74d28d2085fa818e.jpg) + +![](images/7e158995cd06d2fdb580edc5b8ab12456ffd74b2a3a908eb2173167d601f6c2d.jpg) + +![](images/96a48d934089c6de468a5f8221de8bab1a4f9f0813dca88a2c5b4a0f2b431074.jpg) + +![](images/8ce8d12ebb8fa2a63cfa40cbde8b24f0b8ff34ff41334cfc0e8a465118c2c06f.jpg) + +"Place a floor lamp next to the chair." + +![](images/3f11d58e9637cd11db9cd21a914e49b9da1cdcc52d259a00446426cada90f0ce.jpg) + +![](images/1f5d11a7dd568828f689e01b9f6eea4683d09a4133f3048193e944d15c1e2b94.jpg) + +![](images/d6e99f324cebd2924f4db5dd6da33ce11110896a658f163d3ec6ea30162872b8.jpg) + +![](images/2809d23df25649f9773a2c22c7db7f57488f17d2dd9f6ebf60096e0306244184.jpg) + +![](images/fd15791da0232bba74907688794d3148c1ca36913d7369576f610384f3b4089e.jpg) + +"Replace the truck with a tractor" + +Figure 6. Effect of refined instructions and Instruct-CLIP guidance loss. Compared to variants of our model trained without refined instructions ("Ours (w/o data)") or without the guidance loss ("Ours (w/o loss)"), our model produces superior results. + +Ori + +IP2P + +MagBr + +Ori + +IP2P + +IagBr Ours + +$\therefore m = \frac{3}{11}$ + +② + +$\therefore m = \frac{3}{11}$ + +${13}/{14}$ + +${13}/{14}$ + +③ + +② ${\mathrm{{CO}}}_{2}$ : 溶液 + +③ + +1 + +$\frac{3}{1} + u + {4q} = 1 + {uq}$ dH + +# + +. + +${10} : \frac{{a}_{n + 1}}{{b}_{n + 1}} = \frac{2{a}_{n + 1}}{2{b}_{n + 1}} = \frac{{a}_{1} + {a}_{{2n} + 1}}{{b}_{1} + {b}_{{2n} + 1}} =$ + +4 + +。 + +Y + +$\therefore m = \frac{3}{11}$ + +Y + +$\therefore m = \frac{3}{11}$ + +$\therefore m : x = 1$ 或 ${3x} + {4y} + 1 = 0$ + +A + +${12}/{12}$ + +${14} \times {25}$ + +$\therefore m = \frac{3}{11}$ + +$\Rightarrow \;{3}^{x} = - 2 + 2\sqrt{2} + 2\sqrt{2} = - 4$ + +$\therefore m = \frac{3}{11}$ + +$\therefore m = \frac{3}{11}$ + +$\therefore m = \frac{3}{11}$ + +$\downarrow x$ + +${13}/{14}$ + +$\therefore m = \frac{3}{11}$ + +${13}/{14}$ + +$\therefore m = \frac{3}{11}$ ; + +$\therefore m = \frac{3}{11}$ ; + +# + +${13}/{14}$ + +$\therefore m = \frac{3}{11}$ + +$\therefore m = \frac{3}{11}$ + +$\therefore m = \frac{3}{11}$ + +# + +$\therefore m = \frac{3}{11}$ + +(1) + +$\therefore m = \frac{3}{11}$ + +"Turn orange shirt into blue" + +"Ac + +ld a + +n ea + +gle + +above + +the + +$\therefore m = \frac{3}{11}$ + +# + +# + +# + +$\therefore m = \frac{3}{11}$ + +# + +# + +$\left\{ {2 : {x}^{\prime }\left( x\right) = {y}^{\prime }\left( x\right) }\right\} : y = {13}\left( x\right) = {13}{y}^{\prime }\left( x\right)$ + +# + +$\frac{1}{2} =$ + +$\therefore m = \frac{3}{11}$ ; + +$\therefore m = \frac{3}{11}$ + +"Delete the wooden boat" + +"Remove the white plane" + +Figure 7. Examples of failure cases compared with Instruct-Pix2Pix (IP2P) [1] and MagicBrush (MagBr) [25] performed on the original (Ori) images. + +fined dataset without the guidance loss ("Ours (w/o loss)"). We compare these with IP2P and our full method in Table 3, and visual examples are provided in Fig. 6. Results show that refined data significantly improve visual alignment with ground truth, evidenced by higher CLIP and DINO scores, while the guidance loss further enhances this alignment. + +# 4.6. Limitations + +Despite being able to correct a lot of inaccurate edit instructions in the training set, our method is still heavily influenced by the limitations of the generative method [6] these training images come from, which include color bleeding outside the intended edit region (Fig. 7, top left), and incomplete object addition (Fig. 7, top right). While our model respects the original images better, it sometimes struggles to remove objects in the original images (Fig. 7, bottom). + +# 5. Conclusion + +We have presented Instruct-CLIP, a self-supervised method for instruction-guided image editing that learns the semantic changes between original and edited images to refine edit instructions in datasets. Applying I-CLIP to the InstructPix2Pix dataset yields over 120K refined samples, which we use to fine-tune its model with our I-CLIP-guided loss function and generate better edit results. + +# References + +[1] Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2Pix: Learning to follow image editing instructions. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 2, 3 +[3] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 34:8780-8794, 2021. 1, 2 +[4] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis, march 2024. URL http://arxiv.org/abs/2403.03206, 2024. 2 +[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4 +[6] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3, 8 +[7] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-Edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. 3 +[8] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, Jianzhuang Liu, et al. ZONE: Zero-shot instruction-guided local editing. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 6254-6263, 2024. 1, 3, 6, 7 +[9] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. DeCap: Decoding CLIP latents for zero-shot captioning via text-only training. arXiv preprint arXiv:2303.03032, 2023. 2, 4, 5 +[10] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent Consistency Models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 2 +[11] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Marcus A Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G Derpanis, and Igor Gilitschenski. Watch Your Steps: Local image and scene editing by text instructions. In European Conference on Computer Vision (ECCV), pages 111-129. Springer, 2025. 1, 3, 6, 7 +[12] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text Inversion for editing real images using guided diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 6038–6047, 2023. 2 +[13] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, + +Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 3, 4 +[14] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In SIGGRAPH, pages 1-11, 2023. 2, 3 +[15] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2 +[16] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 4 +[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2, 3 +[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 5 +[19] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2, 5 +[20] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-Play diffusion features for text-driven image-to-image translation. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921–1930, 2023. 2, 3 +[21] Bram Wallace, Akash Gokul, and Nikhil Naik. EDICT: Exact diffusion inversion via coupled transformations. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 22532-22541, 2023. 3 +[22] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 1900-1910, 2023. 2, 3 +[23] Ling Yang, Bohan Zeng, Jiaming Liu, Hong Li, Minghao Xu, Wentao Zhang, and Shuicheng Yan. EditWorld: Simulating world dynamics for instruction-following image editing. arXiv preprint arXiv:2405.14785, 2024. 3 +[24] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-Inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246, 2023. 1, 3, 6, 7 +[25] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. MagicBrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1, 3, 6, 7, 8 +[26] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, + +Stefano Ermon, Caiming Xiong, and Ran Xu. HIVE: Harnessing human feedback for instructional visual editing. arXiv preprint arXiv:2303.09618, 2023. 2, 3, 6, 7 +[27] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. 1, 2, 3 +[28] Ruoyu Zhao, Qingnan Fan, Fei Kou, Shuai Qin, Hong Gu, Wei Wu, Pengcheng Xu, Mingrui Zhu, Nannan Wang, and Xinbo Gao. InstructBrush: Learning attention-based instruction optimization for image editing. arXiv preprint arXiv:2403.18660, 2024. 8 \ No newline at end of file diff --git a/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/images.zip b/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f5b2f9e4e139eb217e8402100730106f56c23c68 --- /dev/null +++ b/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24756bfea7a5f5c50320b61a00e77595183a17c7f0e22782e4387f00936882be +size 827394 diff --git a/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/layout.json b/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2ebc9c9a311a36930397548711bdc70dbc2e6030 --- /dev/null +++ b/2025/Instruct-CLIP_ Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning/layout.json @@ -0,0 +1,13643 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 500, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 500, + 140 + ], + "type": "text", + "content": "Instruct-CLIP: Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 169, + 171, + 247, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 171, + 247, + 185 + ], + "spans": [ + { + "bbox": [ + 169, + 171, + 247, + 185 + ], + "type": "text", + "content": "Sherry X. Chen" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 287, + 171, + 339, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 171, + 339, + 184 + ], + "spans": [ + { + "bbox": [ + 287, + 171, + 339, + 184 + ], + "type": "text", + "content": "Misha Sra" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 378, + 171, + 441, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 171, + 441, + 185 + ], + "spans": [ + { + "bbox": [ + 378, + 171, + 441, + 185 + ], + "type": "text", + "content": "Pradeep Sen" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 209, + 195, + 400, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 195, + 400, + 209 + ], + "spans": [ + { + "bbox": [ + 209, + 195, + 400, + 209 + ], + "type": "text", + "content": "University of California, Santa Barbara" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 223, + 211, + 386, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 211, + 386, + 224 + ], + "spans": [ + { + "bbox": [ + 223, + 211, + 386, + 224 + ], + "type": "text", + "content": "{xchen774, sra, psen}@ucsb.edu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 152, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 285, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 597 + ], + "type": "text", + "content": "Although natural language instructions offer an intuitive way to guide automated image editing, deep-learning models often struggle to achieve high-quality results, largely due to the difficulty of creating large, high-quality training datasets. To do this, previous approaches have typically relied on text-to-image (T2I) generative models to produce pairs of original and edited images that simulate the input/output of an instruction-guided image-editing model. However, these image pairs often fail to align with the specified edit instructions due to the limitations of T2I models, which negatively impacts models trained on such datasets. To address this, we present Instruct-CLIP (I-CLIP), a self-supervised method that learns the semantic changes between original and edited images to refine and better align the instructions in existing datasets. Furthermore, we adapt Instruct-CLIP to handle noisy latent images and diffusion timesteps so that it can be used to train latent diffusion models (LDMs) and efficiently enforce alignment between the edit instruction and the image changes in latent space at any step of the diffusion pipeline. We use Instruct-CLIP to correct the InstructPix2Pix dataset and get over 120K refined samples we then use to fine-tune their model, guided by our novel I-CLIP-based loss function. The resulting model can produce edits that are more aligned with the given instructions. Our code and dataset are available at https://github.com/SherryXTChen/Instruct-CLIP.git." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 620, + 135, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 620, + 135, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 135, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Recent advances in instruction-guided image editing have introduced intuitive and powerful tools that require only a single text instruction for image editing [1, 24-27]. In particular, InstructPix2Pix (IP2P) [1] pioneered diffusion-based instruction-driven image editing, serving as the foundation for numerous subsequent works [8, 11, 25]. These" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 323, + 270, + 396, + 312 + ], + "blocks": [ + { + "bbox": [ + 345, + 258, + 375, + 268 + ], + "lines": [ + { + "bbox": [ + 345, + 258, + 375, + 268 + ], + "spans": [ + { + "bbox": [ + 345, + 258, + 375, + 268 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 270, + 396, + 312 + ], + "lines": [ + { + "bbox": [ + 323, + 270, + 396, + 312 + ], + "spans": [ + { + "bbox": [ + 323, + 270, + 396, + 312 + ], + "type": "image", + "image_path": "4064cbdbe74a34237954bc693f8b510a892fa479a3708327c059f926106889ee.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 312, + 447, + 323 + ], + "lines": [ + { + "bbox": [ + 322, + 312, + 447, + 323 + ], + "spans": [ + { + "bbox": [ + 322, + 312, + 447, + 323 + ], + "type": "text", + "content": "\"Change the panda to a monkey\"" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 396, + 270, + 469, + 312 + ], + "blocks": [ + { + "bbox": [ + 422, + 258, + 441, + 268 + ], + "lines": [ + { + "bbox": [ + 422, + 258, + 441, + 268 + ], + "spans": [ + { + "bbox": [ + 422, + 258, + 441, + 268 + ], + "type": "text", + "content": "IP2P" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 396, + 270, + 469, + 312 + ], + "lines": [ + { + "bbox": [ + 396, + 270, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 396, + 270, + 469, + 312 + ], + "type": "image", + "image_path": "b8780646092415adacdf94000231248dd14457fc025d2e19dde7e7f88439ab3f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 470, + 270, + 543, + 312 + ], + "blocks": [ + { + "bbox": [ + 481, + 258, + 530, + 268 + ], + "lines": [ + { + "bbox": [ + 481, + 258, + 530, + 268 + ], + "spans": [ + { + "bbox": [ + 481, + 258, + 530, + 268 + ], + "type": "text", + "content": "I-CLIP (Ours)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 270, + 543, + 312 + ], + "lines": [ + { + "bbox": [ + 470, + 270, + 543, + 312 + ], + "spans": [ + { + "bbox": [ + 470, + 270, + 543, + 312 + ], + "type": "image", + "image_path": "46c9334dc4c13157fa99244bf8dd8ea965351b7cf6fa0e94ae3e85f51609fd88.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 323, + 325, + 396, + 392 + ], + "blocks": [ + { + "bbox": [ + 323, + 325, + 396, + 392 + ], + "lines": [ + { + "bbox": [ + 323, + 325, + 396, + 392 + ], + "spans": [ + { + "bbox": [ + 323, + 325, + 396, + 392 + ], + "type": "image", + "image_path": "6e143abed567f69ba12f36627f005a62faf7bcedbc1be7290396a325fb841b93.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 392, + 512, + 403 + ], + "lines": [ + { + "bbox": [ + 322, + 392, + 512, + 403 + ], + "spans": [ + { + "bbox": [ + 322, + 392, + 512, + 403 + ], + "type": "text", + "content": "\"Let the old man be dressed in a military uniform\"" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 396, + 325, + 469, + 392 + ], + "blocks": [ + { + "bbox": [ + 396, + 325, + 469, + 392 + ], + "lines": [ + { + "bbox": [ + 396, + 325, + 469, + 392 + ], + "spans": [ + { + "bbox": [ + 396, + 325, + 469, + 392 + ], + "type": "image", + "image_path": "f02c93c2bdb28c512dfe34789bb85af4167fcf070c3e7247544740ea7ccc0580.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 470, + 325, + 543, + 392 + ], + "blocks": [ + { + "bbox": [ + 470, + 325, + 543, + 392 + ], + "lines": [ + { + "bbox": [ + 470, + 325, + 543, + 392 + ], + "spans": [ + { + "bbox": [ + 470, + 325, + 543, + 392 + ], + "type": "image", + "image_path": "a26de27ca92f9dacfe10dafa91843a707a0860b84e8bcbaad2d0acc39beb3717.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 323, + 405, + 396, + 447 + ], + "blocks": [ + { + "bbox": [ + 323, + 405, + 396, + 447 + ], + "lines": [ + { + "bbox": [ + 323, + 405, + 396, + 447 + ], + "spans": [ + { + "bbox": [ + 323, + 405, + 396, + 447 + ], + "type": "image", + "image_path": "3b0ea394ed1ecb2d2ee52e6bbd7fb303ac3cebf2ab5e8b9e051dc16a017ff95b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 448, + 476, + 459 + ], + "lines": [ + { + "bbox": [ + 322, + 448, + 476, + 459 + ], + "spans": [ + { + "bbox": [ + 322, + 448, + 476, + 459 + ], + "type": "text", + "content": "\"Change the background to a classroom\"" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 396, + 405, + 469, + 447 + ], + "blocks": [ + { + "bbox": [ + 396, + 405, + 469, + 447 + ], + "lines": [ + { + "bbox": [ + 396, + 405, + 469, + 447 + ], + "spans": [ + { + "bbox": [ + 396, + 405, + 469, + 447 + ], + "type": "image", + "image_path": "73db55f64e3a6f8c390a3653e874c27b73501d75af6fc082036c8e883641c467.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 470, + 405, + 543, + 447 + ], + "blocks": [ + { + "bbox": [ + 470, + 405, + 543, + 447 + ], + "lines": [ + { + "bbox": [ + 470, + 405, + 543, + 447 + ], + "spans": [ + { + "bbox": [ + 470, + 405, + 543, + 447 + ], + "type": "image", + "image_path": "c128f79eca84f99b119be5e19901d71c36a22cd0cb823af09bb97356ed9a55fe.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 323, + 460, + 396, + 514 + ], + "blocks": [ + { + "bbox": [ + 323, + 460, + 396, + 514 + ], + "lines": [ + { + "bbox": [ + 323, + 460, + 396, + 514 + ], + "spans": [ + { + "bbox": [ + 323, + 460, + 396, + 514 + ], + "type": "image", + "image_path": "5e88fc350301b685886516f25da5eb468528559cdc1c1f6010633428ab642a00.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 515, + 427, + 525 + ], + "lines": [ + { + "bbox": [ + 322, + 515, + 427, + 525 + ], + "spans": [ + { + "bbox": [ + 322, + 515, + 427, + 525 + ], + "type": "text", + "content": "\"The cat needs to be black\"" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 396, + 460, + 469, + 514 + ], + "blocks": [ + { + "bbox": [ + 396, + 460, + 469, + 514 + ], + "lines": [ + { + "bbox": [ + 396, + 460, + 469, + 514 + ], + "spans": [ + { + "bbox": [ + 396, + 460, + 469, + 514 + ], + "type": "image", + "image_path": "936477d796fc15f5c908ac607c9ee1e1a1c94caba67c2bcb6fdce16c9f58a45a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 470, + 460, + 543, + 514 + ], + "blocks": [ + { + "bbox": [ + 470, + 460, + 543, + 514 + ], + "lines": [ + { + "bbox": [ + 470, + 460, + 543, + 514 + ], + "spans": [ + { + "bbox": [ + 470, + 460, + 543, + 514 + ], + "type": "image", + "image_path": "3fa513c733c17c20876c10f325c9723ef221f2af50db1ba1dcbd92bb60948d6b.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 323, + 526, + 396, + 600 + ], + "blocks": [ + { + "bbox": [ + 323, + 526, + 396, + 600 + ], + "lines": [ + { + "bbox": [ + 323, + 526, + 396, + 600 + ], + "spans": [ + { + "bbox": [ + 323, + 526, + 396, + 600 + ], + "type": "image", + "image_path": "98bfd82049da715b9ff632e31b1875e64fb5d15b350f34951f148e8dde9c3ba5.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 601, + 424, + 612 + ], + "lines": [ + { + "bbox": [ + 322, + 601, + 424, + 612 + ], + "spans": [ + { + "bbox": [ + 322, + 601, + 424, + 612 + ], + "type": "text", + "content": "\"Make it a chocolate cake\"" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 616, + 555, + 639 + ], + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 639 + ], + "type": "text", + "content": "Figure 1. Results showcasing the strength of Instruct-CLIP (I-CLIP) compared to state-of-the-art InstructPix2Pix (IP2P) [1]." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 396, + 527, + 469, + 600 + ], + "blocks": [ + { + "bbox": [ + 396, + 527, + 469, + 600 + ], + "lines": [ + { + "bbox": [ + 396, + 527, + 469, + 600 + ], + "spans": [ + { + "bbox": [ + 396, + 527, + 469, + 600 + ], + "type": "image", + "image_path": "c9c36dfcf2aea554d3901e78253d552d91192ac51ad1e31cec3d2aa82c757035.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 470, + 527, + 543, + 600 + ], + "blocks": [ + { + "bbox": [ + 470, + 527, + 543, + 600 + ], + "lines": [ + { + "bbox": [ + 470, + 527, + 543, + 600 + ], + "spans": [ + { + "bbox": [ + 470, + 527, + 543, + 600 + ], + "type": "image", + "image_path": "dea6f4ddbe6f812772756c5434f1eb30a14c056b574f2e724512cb72b3da6372.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "methods leverage pre-trained text-to-image (T2I) models [3, 10, 15, 18, 19] and condition the generation process on edit instructions instead of the usual prompts. However, these models still need to be fine-tuned on appropriate datasets to learn how to make instruction-guided edits." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28513" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 82, + 113, + 139 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 102, + 81 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 102, + 81 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 102, + 81 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 82, + 113, + 139 + ], + "lines": [ + { + "bbox": [ + 57, + 82, + 113, + 139 + ], + "spans": [ + { + "bbox": [ + 57, + 82, + 113, + 139 + ], + "type": "image", + "image_path": "09d638d9d0ba49a534a4105e9a8ef8f61a1809c31300fce056aec719c13bf2d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 114, + 82, + 171, + 139 + ], + "blocks": [ + { + "bbox": [ + 130, + 71, + 155, + 80 + ], + "lines": [ + { + "bbox": [ + 130, + 71, + 155, + 80 + ], + "spans": [ + { + "bbox": [ + 130, + 71, + 155, + 80 + ], + "type": "text", + "content": "Edited" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 114, + 82, + 171, + 139 + ], + "lines": [ + { + "bbox": [ + 114, + 82, + 171, + 139 + ], + "spans": [ + { + "bbox": [ + 114, + 82, + 171, + 139 + ], + "type": "image", + "image_path": "f4527b57e63dbba0757e278c3e5fab77810a05b1895ebc7e2072606b6de95a80.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 140, + 136, + 152 + ], + "lines": [ + { + "bbox": [ + 57, + 140, + 136, + 152 + ], + "spans": [ + { + "bbox": [ + 57, + 140, + 136, + 152 + ], + "type": "text", + "content": "ORIGINAL: \"by andy warhol\" (0.130) \nREFINED: \"make the sky pink\" (0.263)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 182, + 83, + 236, + 138 + ], + "blocks": [ + { + "bbox": [ + 195, + 71, + 227, + 81 + ], + "lines": [ + { + "bbox": [ + 195, + 71, + 227, + 81 + ], + "spans": [ + { + "bbox": [ + 195, + 71, + 227, + 81 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 83, + 236, + 138 + ], + "lines": [ + { + "bbox": [ + 182, + 83, + 236, + 138 + ], + "spans": [ + { + "bbox": [ + 182, + 83, + 236, + 138 + ], + "type": "image", + "image_path": "8c55a7127e7266c2c581611c409d0afcc460377236f8c70b432c7a00f83dfb9f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 140, + 292, + 152 + ], + "lines": [ + { + "bbox": [ + 181, + 140, + 292, + 152 + ], + "spans": [ + { + "bbox": [ + 181, + 140, + 292, + 152 + ], + "type": "text", + "content": "ORIGINAL: \"make the waves into a hurricane\" (0.090) \nREFINED: \"add a lightning storm to the sky\" (0.289)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 239, + 83, + 293, + 138 + ], + "blocks": [ + { + "bbox": [ + 254, + 71, + 279, + 80 + ], + "lines": [ + { + "bbox": [ + 254, + 71, + 279, + 80 + ], + "spans": [ + { + "bbox": [ + 254, + 71, + 279, + 80 + ], + "type": "text", + "content": "Edited" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 239, + 83, + 293, + 138 + ], + "lines": [ + { + "bbox": [ + 239, + 83, + 293, + 138 + ], + "spans": [ + { + "bbox": [ + 239, + 83, + 293, + 138 + ], + "type": "image", + "image_path": "176d5acdd320062248080f7549f87d837b35d9b864efe44e61fd36ab8457d13c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 57, + 158, + 113, + 214 + ], + "blocks": [ + { + "bbox": [ + 57, + 158, + 113, + 214 + ], + "lines": [ + { + "bbox": [ + 57, + 158, + 113, + 214 + ], + "spans": [ + { + "bbox": [ + 57, + 158, + 113, + 214 + ], + "type": "image", + "image_path": "5b3651168b82d2b723ea14240d4ff8d0c5c39a4bae60ad87cc5374ff50fae067.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 215, + 150, + 228 + ], + "lines": [ + { + "bbox": [ + 57, + 215, + 150, + 228 + ], + "spans": [ + { + "bbox": [ + 57, + 215, + 150, + 228 + ], + "type": "text", + "content": "ORIGINAL:\"make it morning\"(0.179) \nREFINED:\"make it a foggy morning\"(0.324)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 114, + 158, + 171, + 214 + ], + "blocks": [ + { + "bbox": [ + 114, + 158, + 171, + 214 + ], + "lines": [ + { + "bbox": [ + 114, + 158, + 171, + 214 + ], + "spans": [ + { + "bbox": [ + 114, + 158, + 171, + 214 + ], + "type": "image", + "image_path": "e24e9330e453ab7aa0f016124871ae34d3760fa6666eaee31f26dc1dbe94e86c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 181, + 157, + 238, + 214 + ], + "blocks": [ + { + "bbox": [ + 181, + 157, + 238, + 214 + ], + "lines": [ + { + "bbox": [ + 181, + 157, + 238, + 214 + ], + "spans": [ + { + "bbox": [ + 181, + 157, + 238, + 214 + ], + "type": "image", + "image_path": "66aea714ea8979115bdf574b4c60a40e7387f9d69cf1b056aebf7cd025be2596.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 215, + 267, + 228 + ], + "lines": [ + { + "bbox": [ + 181, + 215, + 267, + 228 + ], + "spans": [ + { + "bbox": [ + 181, + 215, + 267, + 228 + ], + "type": "text", + "content": "ORIGINAL: \"make it a landscape\" (0.090) \nREFINED: \"a photograph\" (0.233)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 239, + 157, + 294, + 214 + ], + "blocks": [ + { + "bbox": [ + 239, + 157, + 294, + 214 + ], + "lines": [ + { + "bbox": [ + 239, + 157, + 294, + 214 + ], + "spans": [ + { + "bbox": [ + 239, + 157, + 294, + 214 + ], + "type": "image", + "image_path": "f53887aa8d2ce70146eb91f7448816ee12ada20f103b1285545512d8b03220c4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 57, + 233, + 113, + 290 + ], + "blocks": [ + { + "bbox": [ + 57, + 233, + 113, + 290 + ], + "lines": [ + { + "bbox": [ + 57, + 233, + 113, + 290 + ], + "spans": [ + { + "bbox": [ + 57, + 233, + 113, + 290 + ], + "type": "image", + "image_path": "1cdf2f5294b4d86f51f1998406058bacfe1d404ef93b51d91fd69be4abcb047f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 291, + 143, + 304 + ], + "lines": [ + { + "bbox": [ + 57, + 291, + 143, + 304 + ], + "spans": [ + { + "bbox": [ + 57, + 291, + 143, + 304 + ], + "type": "text", + "content": "ORIGINAL: \"make it a daydream\" (0.113) \nREFINED: \"make it spring\" (0.300)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 312, + 296, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 399 + ], + "type": "text", + "content": "Figure 2. Problems with existing instruction-guided image-editing datasets [1]. As shown, there are many examples where the dataset's original edit instruction does not match the actual changes in the images. Our I-CLIP approach refines edit instructions to match the visual change better and allows us to train a system that produces better outputs. The values in parentheses are the cosine similarity between the visual change from the original to the edited image and the edit instruction from I-CLIP." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 114, + 233, + 171, + 290 + ], + "blocks": [ + { + "bbox": [ + 114, + 233, + 171, + 290 + ], + "lines": [ + { + "bbox": [ + 114, + 233, + 171, + 290 + ], + "spans": [ + { + "bbox": [ + 114, + 233, + 171, + 290 + ], + "type": "image", + "image_path": "63dc5c118c824db7085957c91b1589d77b6becf405792d9ace3cde3170893c81.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 178, + 232, + 233, + 289 + ], + "blocks": [ + { + "bbox": [ + 178, + 232, + 233, + 289 + ], + "lines": [ + { + "bbox": [ + 178, + 232, + 233, + 289 + ], + "spans": [ + { + "bbox": [ + 178, + 232, + 233, + 289 + ], + "type": "image", + "image_path": "7d56f08041ea4297a320ee0bcb98d4d4642b9461d51b265dc0a035d978a375d2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 291, + 286, + 304 + ], + "lines": [ + { + "bbox": [ + 176, + 291, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 176, + 291, + 286, + 304 + ], + "type": "text", + "content": "ORIGINAL: \"make it a video game character\" (0.097) \nREFINED: \"as a 3 d sculpture\" (0.357)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 235, + 233, + 291, + 289 + ], + "blocks": [ + { + "bbox": [ + 235, + 233, + 291, + 289 + ], + "lines": [ + { + "bbox": [ + 235, + 233, + 291, + 289 + ], + "spans": [ + { + "bbox": [ + 235, + 233, + 291, + 289 + ], + "type": "image", + "image_path": "ac214b009a2003d44d3e81e943e6df22cc94d267a28e359603db28e34cfcb883.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 423, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 530 + ], + "type": "text", + "content": "Some have created these datasets by using T2I models to approximate the desired behavior of an instruction-guided image editing system. For example, several approaches [1, 26, 27] leverage Prompt-to-Prompt [6] to generate a pair of images that look like one is an edited version of the other by using a pair of appropriately modified prompts. A separate, fine-tuned large-language model (LLM), such as GPT3 [2], is then used to generate plausible edit instructions from these prompts." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 532, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 628 + ], + "type": "text", + "content": "Given the generality of Prompt-to-Prompt and LLMs, such datasets can cover a wide range of instructions. However, the quality of image editing is often limited by the capabilities of the generative methods, resulting in problems when the changes between the original and the edited image are misaligned with the edit instruction (Fig. 2). This affects the performance on models trained on these datasets, as we can see with the IP2P results in Fig. 1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Although improving image-editing datasets would significantly enhance the performance of models trained on them, doing so is difficult. On one hand, manually creating edit instructions and corresponding outputs is extremely labor-intensive and impractical. On the other, it is nontrivial to construct a pipeline that can perform effective edits in order to create an instruction-guided, image-editing" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "dataset. After all, if such a pipeline existed, it would be a successful instruction-guided, image-editing method itself." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 97, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 192 + ], + "type": "text", + "content": "Furthermore, creating a brand new dataset from scratch may not be necessary. Despite the inherent limitations of previous datasets, we observe that they still provide some \"signal\" for training image editing models. After all, as shown in Fig. 1, we see that models trained on them produce results that, while imperfect, show that the system is understanding something about the desired edit. Can we harness this signal to refine the dataset and address its issues?" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 193, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 276 + ], + "type": "text", + "content": "Inspired by CLIP [17], which learns rich semantic alignments through contrastive language-image pre-training, we embed original-edited image pairs and their corresponding edit instructions into the same feature space in a similar manner with a neural network, which we call Instruct-CLIP. We then follow this with a separate text decoder to generate more precise instructions from the learned features [9]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 277, + 553, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 409 + ], + "type": "text", + "content": "Instruct-CLIP is equipped with a modified DINOv2 [13] backbone capable of handling noisy latent images in Stable Diffusion (SD) [18] so it can not only enhance training data quality in the pre-processing stage but also serve as an efficient learning objective during training. We use Instruct-CLIP to refine instructions in the IP2P dataset [1], resulting in " + }, + { + "bbox": [ + 313, + 277, + 553, + 409 + ], + "type": "inline_equation", + "content": "120\\mathrm{K}+" + }, + { + "bbox": [ + 313, + 277, + 553, + 409 + ], + "type": "text", + "content": " unique, enhanced samples (see Fig. 2). The IP2P model is then fine-tuned on this dataset with an Instruct-CLIP-guided loss, enabling the trained instruction-guided editing model to produce results more faithfully aligned with the instructions (see Fig. 5)." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 325, + 411, + 467, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 411, + 467, + 422 + ], + "spans": [ + { + "bbox": [ + 325, + 411, + 467, + 422 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 314, + 423, + 553, + 493 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 314, + 423, + 553, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 423, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 314, + 423, + 553, + 446 + ], + "type": "text", + "content": "- A dual-purpose model that refines dataset instructions and enhances editing model training." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 314, + 447, + 553, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 447, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 314, + 447, + 553, + 469 + ], + "type": "text", + "content": "- A large-scale dataset with over 120K samples featuring more accurate and enriched instructions." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 314, + 471, + 553, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 471, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 314, + 471, + 553, + 493 + ], + "type": "text", + "content": "- An instruction-guided image editing method trained on above dataset." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 507, + 400, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 507, + 400, + 520 + ], + "spans": [ + { + "bbox": [ + 314, + 507, + 400, + 520 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 528, + 542, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 542, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 542, + 540 + ], + "type": "text", + "content": "2.1. Text-to-image diffusion-based image editing" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Diffusion models [3, 4, 10, 15, 18, 19] have set a new standard for image generation quality, and form the basis of various text-to-image (T2I) editing applications. A range of methods [6, 12, 14, 20-22] aim to edit a given image to semantically align with a target prompt, using the input prompt as an anchor. However, traditional T2I models like Stable Diffusion (SD) [18] often yield inconsistent results even with similar prompts, where both subject and context can change significantly. Prompt-to-Prompt [6] addresses this by preserving attention maps for the shared components in the input and the target prompt to improve image visual alignment, but both images are synthesized so it cannot handle arbitrary input images. Null-text inversion [12] overcomes this by reconstructing an input image in SD based on" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28514" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "the input prompt. They do this by optimizing the empty-text as the negative prompt, enabling controlled edits when combined with Prompt-to-Prompt." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 295, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 295, + 288 + ], + "type": "text", + "content": "Other methods similarly leverage diffusion inversion techniques paired with unique editing mechanisms: Diffusion Disentanglement [22] aims to find the optimal blend of input and target prompt features to find the compromise between visual alignment with the input image and semantic alignment with the target prompt; EDICT [21] uses coupling inversion for improved reconstruction, which in turn enhance editing performance; pix2pix-zero [14] optimizes model attention maps for the target image to match the ones for the input image for visual consistency between the two; and Plug-and-Play [20] integrates self-attention maps from the input image to guide the edited image generation. While each method has its own advantages, a major shared shortcoming is they require prompts for both input and output images, which can be cumbersome for users." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 297, + 231, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 231, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 231, + 310 + ], + "type": "text", + "content": "2.2. Instruction-guided image editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 518 + ], + "type": "text", + "content": "New approaches have recently emerged that leverage priors from Stable Diffusion and which offer a more user-friendly approach to image editing, requiring only a single edit instruction [1, 8, 11, 24-26]. For example, InstructPix2Pix (IP2P) [1] replaces the prompt with an edit instruction that conditions the generation of the edited image, where the input image is concatenated as an extra condition as well. Among these instruction-guided image editing methods, HIVE [26] collects user feedback on outputs to train a reward model, which iteratively improves the model's output. Watch Your Step [11] and ZONE [8] both use a masking mechanism – either by measuring the discrepancy between IP2P predictions with and without the instruction, or by employing a region intersection-over-union (IoU) scheme with a segmentation model. This helps them avoid editing instruction-irrelevant regions, resulting in better localized editing capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 528, + 206, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 206, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 206, + 540 + ], + "type": "text", + "content": "2.3. Instruction-guided datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 677 + ], + "type": "text", + "content": "Regardless what additional control mechanism is introduced to improve instruction-guided, image-editing results, the development of instruction-guided image editing models fundamentally relies on suitable training datasets. However, creating large-scale, high-quality datasets automatically presents a significant challenge, as such a data creation system would effectively constitute an editing method itself. Prior works have addressed this challenge by combining existing synthesis models to approximate the behavior of desired instruction-guided image editing methods, primarily following two distinct approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Several works [1, 26, 27] adopted the first approach, which leverages Prompt-to-Prompt [6] to generate pairs of original/edited images using corresponding prompt pairs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "that describe the content of each image. These systems typically fine-tune a large language model (LLM) like GPT-3 [2] to produce edit instructions from prompt pairs. While this approach generates diverse samples, the dataset quality is inherently constrained by the limitations of both Prompt-to-Prompt and the LLM, often resulting in misalignment between the edit instruction and the actual transformation from the original to the edited image (see Fig. 2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 168, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 287 + ], + "type": "text", + "content": "The second approach attempts to address these limitations by creating datasets through inpainting [24, 25]. Here, edited ground truth images are generated by inpainting selected regions of original images. Although this technique improves the quality of localized image editing samples, it has two significant drawbacks: 1) it cannot generate samples with global modifications (such as style transfer), and 2) it often requires manual creation of inpainting regions. These limitations not only increase the cost of dataset creation but also substantially restrict the dataset size." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 288, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 312 + ], + "type": "text", + "content": "We now describe our approach to address the limitations in previous work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 323, + 404, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 404, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 404, + 335 + ], + "type": "text", + "content": "3. Instruct-CLIP" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 553, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 461 + ], + "type": "text", + "content": "Our goal is to improve previous instruction-guided image editing methods by enhancing the quality of available datasets [1, 7, 23, 26], which we propose to do by correcting their edit instructions. However, while instructions in these datasets sometimes do not reflect the actual changes between the original and edited images, we observe that the datasets overall still provide enough \"signal\" to train editing models. The main challenge is therefore how to properly leverage this \"signal\" to refine the given edit instructions and consequently improve the overall dataset quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 463, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 569 + ], + "type": "text", + "content": "To this end, we propose Instruct-CLIP (I-CLIP, for short), which embeds the visual change between the original image " + }, + { + "bbox": [ + 313, + 463, + 554, + 569 + ], + "type": "inline_equation", + "content": "I^o" + }, + { + "bbox": [ + 313, + 463, + 554, + 569 + ], + "type": "text", + "content": " and the edited result " + }, + { + "bbox": [ + 313, + 463, + 554, + 569 + ], + "type": "inline_equation", + "content": "I^e" + }, + { + "bbox": [ + 313, + 463, + 554, + 569 + ], + "type": "text", + "content": " into the same feature space as the corresponding edit instruction " + }, + { + "bbox": [ + 313, + 463, + 554, + 569 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 463, + 554, + 569 + ], + "type": "text", + "content": " through contrastive learning with a neural network. Our approach is inspired by CLIP [17], which learns the semantic alignment between images and their captions. Likewise, Instruct-CLIP learns the relationship between the visual changes in the original-edit image pair and its edit instruction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 677 + ], + "type": "text", + "content": "Just like in CLIP, our approach has an image encoder that in our case encodes the visual change between the input and edited images (denoted as I-CLIP" + }, + { + "bbox": [ + 313, + 570, + 553, + 677 + ], + "type": "inline_equation", + "content": "_{\\mathrm{vis}}(I^o, I^e))" + }, + { + "bbox": [ + 313, + 570, + 553, + 677 + ], + "type": "text", + "content": ", and a text encoder that encodes the edit instruction " + }, + { + "bbox": [ + 313, + 570, + 553, + 677 + ], + "type": "inline_equation", + "content": "(\\mathrm{I - CLIP}_{\\mathrm{txt}}(p))" + }, + { + "bbox": [ + 313, + 570, + 553, + 677 + ], + "type": "text", + "content": ". However, unlike the original CLIP image encoder, ours takes both the original and edited images as input so that it can encode their visual difference. As shown in Fig. 3a, we fine-tune the two encoders by computing and minimizing the contrastive loss between them (Eq. 2), similar to CLIP." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Furthermore, while the architecture of I-CLIP" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "_{\\mathrm{txt}}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " is the same as CLIP, our I-CLIP" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "_{\\mathrm{vis}}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " adds two shared-weighted DINOv2 [13] modules in front of CLIP" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "_{\\mathrm{vis}}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": ", as shown in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28515" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 68, + 292, + 165 + ], + "blocks": [ + { + "bbox": [ + 91, + 68, + 292, + 165 + ], + "lines": [ + { + "bbox": [ + 91, + 68, + 292, + 165 + ], + "spans": [ + { + "bbox": [ + 91, + 68, + 292, + 165 + ], + "type": "image", + "image_path": "a0be5bc45527e087197317b873cf6ba85f5653dd2f1e24de6a8b44d2dc88aefc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 170, + 279, + 180 + ], + "lines": [ + { + "bbox": [ + 131, + 170, + 279, + 180 + ], + "spans": [ + { + "bbox": [ + 131, + 170, + 279, + 180 + ], + "type": "text", + "content": "(a) Overall Instruct-CLIP system architecture" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 338, + 71, + 515, + 168 + ], + "blocks": [ + { + "bbox": [ + 338, + 71, + 515, + 168 + ], + "lines": [ + { + "bbox": [ + 338, + 71, + 515, + 168 + ], + "spans": [ + { + "bbox": [ + 338, + 71, + 515, + 168 + ], + "type": "image", + "image_path": "d7b49a1efa24227fcfc76507870139ba80ed00b2d61f89d720b6b2887a8d0d31.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 170, + 533, + 180 + ], + "lines": [ + { + "bbox": [ + 340, + 170, + 533, + 180 + ], + "spans": [ + { + "bbox": [ + 340, + 170, + 533, + 180 + ], + "type": "text", + "content": "(b) Architecture of visual-change encoder I-CLIP VIS " + }, + { + "bbox": [ + 340, + 170, + 533, + 180 + ], + "type": "inline_equation", + "content": "(I^o, I^e)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": "Figure 3. Instruct-CLIP architectures. (a) Overview of Instruct-CLIP (I-CLIP), which embeds the visual change in the original/edited images " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "I^o" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "I^e" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " and the edit instruction " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " into the same feature space through contrastive loss, " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contrast}}" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " (Eq. 2). To obtain refined instruction " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " from its I-CLIP embedding " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{txt}}" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": ", we adopt the same approach in DeCap [9] to decode " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{txt}}" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " back to " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " using cross-entropy loss, " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DeCap}}" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " (Eq. 4). At inference time, the text decoder takes the embedded visual change from the original to the edited image " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "(z^{\\mathrm{vis}})" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " and decodes it to produce a new instruction. Due to the significant cosine similarity gap between " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{txt}}" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " even when they are well aligned, directly decoding " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " leads to suboptimal results. To achieve a representation of " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " closer to the text features that the instruction decoder learned during training, we compute " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "(z^{\\mathrm{vis}})'" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " with Eq. 6 and decode it to obtain the refined instruction " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": ", which is used to improve the dataset. (b) The architecture of image encoder I-CLIP " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{vis}" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " includes two shared-weighted DINOv2 [13] modules in front of a standard CLIP " + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{vis}" + }, + { + "bbox": [ + 55, + 186, + 555, + 275 + ], + "type": "text", + "content": " encoder." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 295, + 294, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 295, + 294, + 366 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 294, + 366 + ], + "type": "text", + "content": "Fig. 3b. These DINOv2 units extract rich, robust visual features from the input images, allowing " + }, + { + "bbox": [ + 54, + 295, + 294, + 366 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{vis}}" + }, + { + "bbox": [ + 54, + 295, + 294, + 366 + ], + "type": "text", + "content": " to focus on encoding the difference between the original and edited images. This is essential when we introduce Stable Diffusion's (SD) latent image encoding and diffusion into the network, as explained in Sec. 3.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": "To describe the architecture of I-CLIP" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "_{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": ", we write the output of DINOv2 for a given image " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "d = \\mathrm{DINO}_{\\mathrm{v2}}(I)" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": " and the intermediate features from each of its " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": " layers as " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "\\{d_i \\mid i \\in [1, l]\\}" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": ". So in I-CLIP" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "_{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": ", we first we compute " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "d^o" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "d^e" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": " by passing the input image " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "I^o" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": " and the edited image " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "I^e" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": " through DINOv2, respectively, then compute the DINOv2 feature difference " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "d^e - d^o" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": " that we input to CLIP" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "_{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": ". The intermediate feature differences " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "d_i^e - d_i^o" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": " are also added to input of the " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": "th layer of CLIP" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "_{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": " before being processed by that layer. We find that this improves model performance in our experiments by providing additional information that may be lost from " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "d^e - d^o" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": " when being processed from the " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "(i + 1)^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "inline_equation", + "content": "l^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 367, + 295, + 522 + ], + "type": "text", + "content": " layer, similar to skip connections [5]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": "To train Instruct-CLIP, we first initialize " + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": " and I-CLIP" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "inline_equation", + "content": "_{\\mathrm{txt}}" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": " with the respective blocks from the pre-trained CLIP model. Then, given a batch of " + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": " data samples " + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\{(I_i^o,I_i^e,p_i)\\mid i\\in [1,n]\\}" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": " from an existing instruction-guided image-editing dataset like InstructPix2Pix [1], we first compute the Instruct-CLIP features for both the visual difference and edit instruction:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 609, + 295, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 609, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 126, + 609, + 295, + 637 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {i} ^ {\\text {v i s}} = \\mathrm {I} - \\operatorname {C L I P} _ {\\text {v i s}} \\left(I _ {i} ^ {o}, I _ {i} ^ {e}\\right) \\\\ z _ {i} ^ {\\mathrm {t x t}} = \\mathrm {I} - \\mathrm {C L I P} _ {\\mathrm {t x t}} (p _ {i}), \\\\ \\end{array}", + "image_path": "d6f62c2f61b878984f622d5e7bbcc4213125058c695da55555358d5ccf03e439.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 282, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 282, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 282, + 653 + ], + "type": "text", + "content": "and compute the contrastive loss " + }, + { + "bbox": [ + 55, + 641, + 282, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contrast}}" + }, + { + "bbox": [ + 55, + 641, + 282, + 653 + ], + "type": "text", + "content": " between them:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 656, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 656, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 63, + 656, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {c o n t r a s t}} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\log \\frac {\\exp \\left(\\sin \\left(z _ {i} ^ {\\text {v i s}}, z _ {i} ^ {\\text {t x t}}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(\\sin \\left(z _ {i} ^ {\\text {v i s}}, z _ {j} ^ {\\text {t x t}}\\right) / \\tau\\right)} \\right. \\tag {2} \\\\ \\left. + \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(z _ {i} ^ {\\mathrm {t x t}} , z _ {i} ^ {\\mathrm {v i s}}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(\\operatorname {s i m} \\left(z _ {i} ^ {\\mathrm {t x t}}, z _ {j} ^ {\\mathrm {v i s}}\\right) / \\tau\\right)}\\right), \\\\ \\end{array}", + "image_path": "c240960768a8691dc8f593f3963c41581a8316ba56564435ad632a43294e6bf0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 295, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 555, + 332 + ], + "type": "text", + "content": "where scalar " + }, + { + "bbox": [ + 313, + 295, + 555, + 332 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 295, + 555, + 332 + ], + "type": "text", + "content": " is a learnable \"temperature\" parameter that controls the sharpness of the similarity distribution and function " + }, + { + "bbox": [ + 313, + 295, + 555, + 332 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 295, + 555, + 332 + ], + "type": "text", + "content": " measures normalized cosine similarity:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 381, + 343, + 555, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 343, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 381, + 343, + 555, + 369 + ], + "type": "interline_equation", + "content": "\\operatorname {s i m} \\left(z _ {1}, z _ {2}\\right) = \\frac {z _ {1} \\cdot z _ {2}}{\\left\\| z _ {1} \\right\\| \\left\\| z _ {2} \\right\\|}. \\tag {3}", + "image_path": "3b55b05477b0646f0e4162c198f37d702b03c34ca209c7dc35117fe456e8d09e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 379, + 554, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 554, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 554, + 438 + ], + "type": "text", + "content": "We first use this contrastive loss to fine-tune the I-CLIP" + }, + { + "bbox": [ + 313, + 379, + 554, + 438 + ], + "type": "inline_equation", + "content": "_{\\mathrm{vis}}" + }, + { + "bbox": [ + 313, + 379, + 554, + 438 + ], + "type": "text", + "content": " and I-CLIP" + }, + { + "bbox": [ + 313, + 379, + 554, + 438 + ], + "type": "inline_equation", + "content": "_{\\mathrm{txt}}" + }, + { + "bbox": [ + 313, + 379, + 554, + 438 + ], + "type": "text", + "content": " modules in Fig. 3a. Once they are trained to embed the visual change and the instruction into the same feature space, we then use the approach of DeCap [9] to translate the features into text instructions as described next." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 453, + 542, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 542, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 542, + 466 + ], + "type": "text", + "content": "3.1. Predicting instructions by decoding features" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "type": "text", + "content": "Learning the alignment between original/edited image pairs and instructions does not automatically result in better instructions, since a decoder is required to translate I-CLIP latent-space features into actual text instructions. To do this, we leverage the approach of DeCap [9], where we finetune pre-trained CLIP and GPT-2 decoding head [16] for our application to match the original edit instruction " + }, + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "type": "text", + "content": " by either decoding " + }, + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}} = \\mathrm{I - CLIP}_{\\mathrm{vis}}(I^{o},I^{e})" + }, + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{txt}} = \\mathrm{I - CLIP}_{\\mathrm{txt}}(p)" + }, + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": "Since the original edit instructions " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " in the datasets may not match the visible changes encoded in " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": ", we do not want to force the model to decode " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " directly to " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " during training. Instead, we train it by decoding " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{txt}}" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " back to " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": ", using the same approach as DeCap [9]. Formally, let " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "p = [c_1, c_2, \\dots, c_n]" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " be an edit instruction's token representation of length " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": ", where each token " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " is a one-hot vector. Similarly, the decoded/predicted instruction is " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "p' = [c_1', c_2', \\dots, c_n']" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " (instructions can be represented by the same number of tokens through padding or truncation). This is basically a classification problem that requires identifying the correct \"element\" for each token, so the training" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28516" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 266, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 266, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 266, + 85 + ], + "type": "text", + "content": "objective of the text decoder is a cross-entropy loss:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 96, + 92, + 295, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 92, + 295, + 125 + ], + "spans": [ + { + "bbox": [ + 96, + 92, + 295, + 125 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D e C a p}} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\sum_ {c \\in C} y _ {i} ^ {p} (c) \\log y _ {i} ^ {p ^ {\\prime}} (c), \\tag {4}", + "image_path": "a99ca59ebbf9a1b59d3d6c82c8316396543ec8cbfcd36b27cc780621b0a1fa05.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "text", + "content": " is the set of all possible tokens, " + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "inline_equation", + "content": "y_{i}^{p}(c)" + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "text", + "content": " is the ground-truth for each token " + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "inline_equation", + "content": "y_{i}^{p^{\\prime}}(c)" + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "text", + "content": " is the probability of token " + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "text", + "content": " at position " + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "inline_equation", + "content": "p^{\\prime}" + }, + { + "bbox": [ + 55, + 132, + 295, + 171 + ], + "type": "text", + "content": " as predicted by the model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "text", + "content": "At inference time, our goal is to generate improved instructions that accurately capture the visual changes from the original image to the edited one by decoding " + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "text", + "content": ". However, since the average cosine similarity between visual and text CLIP features is relatively low, even when well matched (around 0.2) [9], a text decoder that has been trained on text features " + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{txt}}" + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "text", + "content": " will struggle to handle " + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "text", + "content": " due to such differences, producing sub-optimal decoded results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 266, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 421 + ], + "type": "text", + "content": "To address this, we would like the decoder to get as input a feature representation of " + }, + { + "bbox": [ + 55, + 266, + 296, + 421 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 421 + ], + "type": "text", + "content": " that is more similar to the text features the instruction decoder was trained on. A simple approach would be to compute the cosine similarity between " + }, + { + "bbox": [ + 55, + 266, + 296, + 421 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 421 + ], + "type": "text", + "content": " and each instruction's text feature in the dataset, selecting the feature with the highest similarity. However, this method would simply retrieve the most similar instruction, limiting both the diversity and accuracy of refined instructions. Instead, we follow the approach of DeCap [9], which leverages information from all text features by weighting their influence by their cosine similarity to " + }, + { + "bbox": [ + 55, + 266, + 296, + 421 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 421 + ], + "type": "text", + "content": ". Essentially, instructions with text features more similar to " + }, + { + "bbox": [ + 55, + 266, + 296, + 421 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 421 + ], + "type": "text", + "content": " should contribute more to the refined instruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "text", + "content": "Therefore, we calculate the probability of each instruction feature contributing to the refined representation using the softmax function over the cosine similarities between all text features and " + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "text", + "content": ". Formally, for a dataset of size " + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\{z_i^{\\mathrm{txt}} = \\mathrm{I - CLIP}_{\\mathrm{txt}}(p_i) \\mid i \\in [1,n]\\}" + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "text", + "content": " be the set of text features, each corresponding to an instruction " + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "text", + "content": " in the dataset. The probability " + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "text", + "content": " that instruction " + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "text", + "content": "'s text feature influences the refined instruction for " + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 422, + 296, + 517 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 523, + 295, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 523, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 107, + 523, + 295, + 553 + ], + "type": "interline_equation", + "content": "w _ {i} = \\frac {\\exp \\left(\\operatorname {s i m} \\left(z _ {i} ^ {\\mathrm {t x t}} , z ^ {\\mathrm {v i s}}\\right)\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(\\operatorname {s i m} \\left(z _ {j} ^ {\\mathrm {t x t}}, z ^ {\\mathrm {v i s}}\\right)\\right)}. \\tag {5}", + "image_path": "7e9bef0f177cbfd51ea0aebfadf22b2bf3d07b045f6d2961a425df32b32c5656.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 560, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 295, + 574 + ], + "type": "text", + "content": "We use these weights to project " + }, + { + "bbox": [ + 55, + 560, + 295, + 574 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 560, + 295, + 574 + ], + "type": "text", + "content": " to the text feature space:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 582, + 295, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 582, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 127, + 582, + 295, + 613 + ], + "type": "interline_equation", + "content": "\\left(z ^ {\\text {v i s}}\\right) ^ {\\prime} = \\sum_ {i = 1} ^ {n} \\left(w _ {i} \\cdot z _ {i} ^ {\\text {t x t}}\\right). \\tag {6}", + "image_path": "4bb808054d2377c8821c5bb29ead139765e6710562416ecee98acd0ebf7041d5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "text", + "content": "In the end, the refined instruction for " + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "inline_equation", + "content": "(I^o,I^e)" + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "inline_equation", + "content": "p^\\prime =" + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "text", + "content": " DeCap " + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\left((z^{\\mathrm{vis}})^{\\prime}\\right)" + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "text", + "content": " , where DeCap(·) denotes the instruction decoder, and so the refined dataset sample is " + }, + { + "bbox": [ + 55, + 616, + 296, + 654 + ], + "type": "inline_equation", + "content": "(I^{o},I^{e},p^{\\prime})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 660, + 262, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 262, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 262, + 673 + ], + "type": "text", + "content": "3.2. Working in the latent-diffusion domain" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "With I-CLIP and DeCap, we can now generate datasets with improved semantic alignment between original/edited image changes and their edit instructions (see results in Fig. 2)," + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 552, + 130 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 552, + 130 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 552, + 130 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 552, + 130 + ], + "type": "image", + "image_path": "baa4c8b08e2749f695ecee1cf8df1addad361b10a07932bde8371db131c74e59.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 144, + 555, + 245 + ], + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 245 + ], + "type": "text", + "content": "Figure 4. Training our LD-DINOv2 model. To use I-CLIP as part of the training objective for Stable Diffusion [18], it needs to handle noisy latent images. Therefore, we replace the original DINOv2 backbone in Fig. 3b with a latent-diffusion version of it we call LD-DINOv2, which takes both the noisy latent image " + }, + { + "bbox": [ + 313, + 144, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\tilde{L}_k" + }, + { + "bbox": [ + 313, + 144, + 555, + 245 + ], + "type": "text", + "content": " from SD VAE encoding and forward-diffusion (FD) timestep " + }, + { + "bbox": [ + 313, + 144, + 555, + 245 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 313, + 144, + 555, + 245 + ], + "type": "text", + "content": ". We then train LD-DINOv2 to \"ignore\" the noise and the latent-space compression and to extract the original DINOv2 features using the training objective " + }, + { + "bbox": [ + 313, + 144, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LD - DINO_{v2}}}" + }, + { + "bbox": [ + 313, + 144, + 555, + 245 + ], + "type": "text", + "content": " (Eq. 7)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 264, + 554, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 554, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 554, + 371 + ], + "type": "text", + "content": "thereby providing a better \"signal\" for training instruction-guided, image-editing models. So one might try to simply train models like InstructPix2Pix [1] directly as-is on our improved dataset. While this does lead to some improvement (see Sec. 4.5), we find we get even more improvement if we reinforce the alignment between the visual change and the edit instruction during the training of the stable diffusion (SD) model itself. This means using I-CLIP as an integral part of the training objective beyond just dataset refinement." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "text", + "content": "To do this, I-CLIP must directly handle noisy, latent images in SD, defined as " + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\tilde{L}_k = \\mathrm{FD}(\\mathrm{VAE}_{\\mathrm{enc}}(I), N, t_k)" + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathrm{FD}(\\cdot)" + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "inline_equation", + "content": "k^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "text", + "content": " step of the forward-diffusion process [19] which takes in image " + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "text", + "content": ", embeds it in the latent domain with SD's variational autoencoder " + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathrm{VAE}_{\\mathrm{enc}}(I)" + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "text", + "content": ", and then adds randomly sampled noise " + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "inline_equation", + "content": "N \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "text", + "content": " with respect to timestep " + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "text", + "content": ". Our DINOv2 feature backbone in I-CLIP, which we call LD-DINOv2 (see Fig. 4), works with " + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\tilde{L}_k" + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "text", + "content": " directly by also taking the corresponding timestep " + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "text", + "content": " as input and is trained to \"ignore\" the noise and the latent-space compression to extract the original DINOv2 features with the following training objective:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 521, + 555, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 521, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 318, + 521, + 555, + 552 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L D - D I N O} _ {\\mathrm {v} 2}} = 1 - \\operatorname {s i m} \\left(d ^ {\\mathrm {L D}}, d\\right) + \\frac {1}{l} \\sum_ {i = 1} ^ {l} \\left(1 - \\operatorname {s i m} \\left(d _ {i} ^ {\\mathrm {L D}}, d _ {i}\\right)\\right), \\tag {7}", + "image_path": "7c5d9aef9dd5faf2f1a0b7e143091d1e4727ff6580976cd89ec78f47b1bba1d5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 556, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 556, + 555, + 606 + ], + "type": "inline_equation", + "content": "d^{\\mathrm{LD}}" + }, + { + "bbox": [ + 313, + 556, + 555, + 606 + ], + "type": "text", + "content": " is the output of LD-DINOv2, " + }, + { + "bbox": [ + 313, + 556, + 555, + 606 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 556, + 555, + 606 + ], + "type": "text", + "content": " is the original DINOv2 output, and the intermediate features from each of its " + }, + { + "bbox": [ + 313, + 556, + 555, + 606 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 556, + 555, + 606 + ], + "type": "text", + "content": " layers are " + }, + { + "bbox": [ + 313, + 556, + 555, + 606 + ], + "type": "inline_equation", + "content": "\\{d_i^{\\mathrm{LD}}\\mid i\\in [1,l]\\}" + }, + { + "bbox": [ + 313, + 556, + 555, + 606 + ], + "type": "text", + "content": ". The visual feature of the original-edited image pair " + }, + { + "bbox": [ + 313, + 556, + 555, + 606 + ], + "type": "inline_equation", + "content": "(I^{o},I^{e})" + }, + { + "bbox": [ + 313, + 556, + 555, + 606 + ], + "type": "text", + "content": " then becomes:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 612, + 555, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 612, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 361, + 612, + 555, + 628 + ], + "type": "interline_equation", + "content": "z ^ {\\mathrm {v i s}} = \\mathrm {I - C L I P} _ {\\mathrm {v i s}} \\left(\\tilde {L} _ {k _ {1}} ^ {o}, t _ {k _ {1}}, \\tilde {L} _ {k _ {2}} ^ {e}, t _ {k _ {2}}\\right), \\tag {8}", + "image_path": "eedc41fdd5be812bad49e7df6750d3204c36476eea4923af024d3f7c137c8547.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 634, + 345, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 634, + 345, + 644 + ], + "spans": [ + { + "bbox": [ + 314, + 634, + 345, + 644 + ], + "type": "text", + "content": "where:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 367, + 651, + 553, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 651, + 553, + 672 + ], + "spans": [ + { + "bbox": [ + 367, + 651, + 553, + 672 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {L} _ {k _ {1}} ^ {o} = \\mathrm {F D} \\left(\\mathrm {V A E} _ {\\text {e n c}} \\left(I ^ {o}\\right), N _ {1}, t _ {k _ {1}}\\right) \\\\ \\tilde {\\sim} \\end{array} \\tag {9}", + "image_path": "04e0e4b866234298f7798fd9d2b31c83049b9405be7f5b18312bdb4a41e49875.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 368, + 669, + 501, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 669, + 501, + 683 + ], + "spans": [ + { + "bbox": [ + 368, + 669, + 501, + 683 + ], + "type": "interline_equation", + "content": "\\tilde {L} _ {k _ {2}} ^ {e} = \\mathrm {F D} \\left(\\mathrm {V A E} _ {\\mathrm {e n c}} \\left(I ^ {e}\\right), N _ {2}, t _ {k _ {2}}\\right),", + "image_path": "e1c27c757359ca7ffd1a1d6fc3bb5b498cce1fe2ccab0a1f100b25f1e88bde2c.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "are generated using randomly sampled Gaussian noise terms " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": ". Note that although for diffusion-based" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28517" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "image editing training, no noise is added to the original image, we noisify the original image to train I-CLIP" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "_{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " to make it more robust. For further details on the LD-DINOv2 training process, please refer to the supplementary. Note that we set " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "k_{1} = k_{2} = 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " during dataset refinement stage so LD-DINOv2 will process the pristine latent images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 155, + 294, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 294, + 168 + ], + "type": "text", + "content": "3.3. Training our instruction-guided editing model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 173, + 296, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 173, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 296, + 209 + ], + "type": "text", + "content": "At this point, we would like to leverage both the refined dataset and I-CLIP to enhance SD-based instruction-guided image editing methods, such as InstructPix2Pix (IP2P) [1]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 210, + 296, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 296, + 235 + ], + "type": "text", + "content": "Given data sample " + }, + { + "bbox": [ + 55, + 210, + 296, + 235 + ], + "type": "inline_equation", + "content": "(I^o, I^e, p)" + }, + { + "bbox": [ + 55, + 210, + 296, + 235 + ], + "type": "text", + "content": ", IP2P usually trains a denoising UNet to predict the noise added to the edited image:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 247, + 296, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 247, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 113, + 247, + 296, + 262 + ], + "type": "interline_equation", + "content": "\\tilde {N} _ {k} = \\mathrm {U N e t} _ {\\mathrm {I P 2 P}} \\left(L ^ {o}, \\tilde {L} _ {k} ^ {e}, t _ {k}, p\\right), \\tag {10}", + "image_path": "7886bc73b47c60340ed45d61a0850b1bac570b3d1c4d5176e1e1e17090e576dc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 270, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 319 + ], + "type": "text", + "content": "which is conditioned on the original image in the latent domain " + }, + { + "bbox": [ + 55, + 270, + 296, + 319 + ], + "type": "inline_equation", + "content": "(L^o)" + }, + { + "bbox": [ + 55, + 270, + 296, + 319 + ], + "type": "text", + "content": " and the edit instruction " + }, + { + "bbox": [ + 55, + 270, + 296, + 319 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 270, + 296, + 319 + ], + "type": "text", + "content": ". The traditional training loss in this setup is the mean squared error (MSE) with respect to the actual noise " + }, + { + "bbox": [ + 55, + 270, + 296, + 319 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 270, + 296, + 319 + ], + "type": "text", + "content": " added to the GT edited image:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 331, + 296, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 331, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 129, + 331, + 296, + 354 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M S E}} = \\left\\| N - \\tilde {N} _ {k} \\right\\| _ {2} ^ {2}, \\tag {11}", + "image_path": "b26a89e27ca1f65c1d4246476bf4b26aa1a2231247d2868d0724e6e2cb6dd4b4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 361, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 410 + ], + "type": "text", + "content": "Given refined sample " + }, + { + "bbox": [ + 55, + 361, + 296, + 410 + ], + "type": "inline_equation", + "content": "(I^o, I^e, p')" + }, + { + "bbox": [ + 55, + 361, + 296, + 410 + ], + "type": "text", + "content": ", we design a complementary loss function to incorporate our Instruct-CLIP guidance into the objective and force the visual change to match the refined edit instruction:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 418, + 296, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 418, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 67, + 418, + 296, + 449 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {I - C L I P}} = 1 - \\operatorname {s i m} \\left(\\mathrm {I - C L I P} _ {\\text {v i s}} \\left(L ^ {o}, t _ {0}, \\tilde {L} _ {k - 1} ^ {e}, t _ {k - 1}\\right), \\right. \\tag {12} \\\\ \\mathrm {I} - \\operatorname {C L I P} _ {\\text {t x t}} \\left(p ^ {\\prime}\\right)), \\\\ \\end{array}", + "image_path": "7e20b6694338820706edb585ce181fa0c19291fb008fdd6c5ab9a9cff6bc5475.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 460, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 485 + ], + "type": "text", + "content": "where the intermediate, denoised latent output using the predicted noise is calculated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 118, + 495, + 295, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 495, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 118, + 495, + 295, + 510 + ], + "type": "interline_equation", + "content": "\\tilde {L} _ {k - 1} ^ {e} = \\mathrm {R D} _ {k, k - 1} \\left(\\tilde {L} _ {k} ^ {e}, \\tilde {N} _ {k}\\right). \\tag {13}", + "image_path": "04d7e081fbb9cfdad5bc26882d7dc996f28b2f9f11493de0240689c4d513a22f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 521, + 295, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 546 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 521, + 295, + 546 + ], + "type": "inline_equation", + "content": "\\mathrm{RD}_{k,k - 1}(\\tilde{L}_k^e,\\tilde{N}_k)" + }, + { + "bbox": [ + 55, + 521, + 295, + 546 + ], + "type": "text", + "content": " is the reverse-diffusion process. Thus, our final training objective is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 558, + 295, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 558, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 129, + 558, + 295, + 571 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M S E}} + \\lambda \\mathcal {L} _ {\\mathrm {I - C L I P}}, \\tag {14}", + "image_path": "da61601dcba9df843bdc106b2517d699a3a97b9a88270561e0054f24950773a1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 582, + 292, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 292, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 292, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 582, + 292, + 595 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 582, + 292, + 595 + ], + "type": "text", + "content": " is set to 0.1 to balance " + }, + { + "bbox": [ + 55, + 582, + 292, + 595 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{I - CLIP}}" + }, + { + "bbox": [ + 55, + 582, + 292, + 595 + ], + "type": "text", + "content": " with the MSE loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 609, + 137, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 137, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 137, + 622 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "We implemented and trained Instruct-CLIP (I-CLIP) as described above and used it to refine the InstructPix2Pix (IP2P) dataset [1] to get " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "120\\mathrm{K}+" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " refined instructions, which took around 10 hours on two A6000 GPUs. We then used this refined dataset to fine-tune the InstructPix2Pix model. Please refer to the supplementary for more details. We now describe the experiments to test our method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 72, + 463, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 463, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 463, + 83 + ], + "type": "text", + "content": "4.1. Baselines and benchmarks" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "content": "We compare our editing model with several baselines: InstInpaint (I-Inp) [24], MagicBrush (MagBr), HIVE [26], InstructPix2Pix (IP2P) [1], Watch Your Steps (WYS) [11], and ZONE [8] (see Sec. 2). We evaluate methods on two instruction-guided image-editing benchmarks: MagicBrush (MagBr) [25] and ZONE [8]. The former contains multi-turn edits where multiple instructions are used to edit one image iteratively, as opposed to single-turn edit where the image is edited once. We compare results quantitatively with the following metrics used in prior work [1, 8, 11, 25]:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 209, + 506, + 245 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 314, + 209, + 500, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 209, + 500, + 221 + ], + "spans": [ + { + "bbox": [ + 314, + 209, + 500, + 221 + ], + "type": "text", + "content": "- CLIP-T = sim(CLIPvis " + }, + { + "bbox": [ + 314, + 209, + 500, + 221 + ], + "type": "inline_equation", + "content": "(I^{e})^{\\prime}" + }, + { + "bbox": [ + 314, + 209, + 500, + 221 + ], + "type": "text", + "content": ", CLIPtxt " + }, + { + "bbox": [ + 314, + 209, + 500, + 221 + ], + "type": "inline_equation", + "content": "(p^{e})" + }, + { + "bbox": [ + 314, + 209, + 500, + 221 + ], + "type": "text", + "content": " )" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 221, + 498, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 221, + 498, + 232 + ], + "spans": [ + { + "bbox": [ + 315, + 221, + 498, + 232 + ], + "type": "text", + "content": "- CLIP-I = sim(CLIPvis((Ie)'), CLIPvis(Ie))" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 233, + 506, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 233, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 315, + 233, + 506, + 245 + ], + "type": "text", + "content": "DINO-I " + }, + { + "bbox": [ + 315, + 233, + 506, + 245 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 315, + 233, + 506, + 245 + ], + "type": "text", + "content": " sim(DINOv2((Ie)'),DINOv2(Ie))" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "inline_equation", + "content": "(I^e)'" + }, + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "text", + "content": " denotes the method output corresponding to a benchmark sample " + }, + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "inline_equation", + "content": "(I^o, I^e, p^o, p^e)" + }, + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "inline_equation", + "content": "I^o" + }, + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "text", + "content": " is the original image, " + }, + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "inline_equation", + "content": "I^e" + }, + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "text", + "content": " is the ground-truth edited image (available for MagBr, not for ZONE), " + }, + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "inline_equation", + "content": "p^o" + }, + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "text", + "content": " is the caption for the original image, and " + }, + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "inline_equation", + "content": "p^e" + }, + { + "bbox": [ + 313, + 245, + 555, + 400 + ], + "type": "text", + "content": " is the intended caption for the edited image. The CLIP-T score assesses semantic alignment between the result and its intended caption in the benchmark while the CLIP and DINO similarity scores (CLIP-I and DINO-I, respectively) evaluate the visual alignment between the result and the ground-truth, if available. However, we note that as shown in Fig. 5 and in the supplemental material, these metrics do not always match the visual quality of the image edits, but are presented here for completeness." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 408, + 476, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 408, + 476, + 419 + ], + "spans": [ + { + "bbox": [ + 314, + 408, + 476, + 419 + ], + "type": "text", + "content": "4.2. Instruction refinement results" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": "We present samples from our refined dataset in Fig. 2 (see supplementary for more samples). As we can see, Instruct-CLIP is able to correct wrong instructions such as \"make the waves into a hurricane\" to \"add a lightning storm to the sky\" (" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": " row right). By correcting these samples, we reduce the noise due to instructions not reflecting the actual edits, which in turn helps improve the performance of the model trained on this dataset as we will see next." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 529, + 436, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 529, + 436, + 541 + ], + "spans": [ + { + "bbox": [ + 314, + 529, + 436, + 541 + ], + "type": "text", + "content": "4.3. Image-editing results" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 546, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 689 + ], + "type": "text", + "content": "We found a diverse set of samples that showcase the strength of our method and compared it to baselines in Figs. 1 and 5 (see supplemental for more). Our model not only addresses many issues present in IP2P – such as unintended changes to regions irrelevant to the instructions – but also generates results aligned more accurately with the edit instructions. This is particularly important for multi-turn editing to avoid results diverse further from the desired outcome across edit turns as shown in the supplementary. Notably, this is achieved without manually created datasets or masking mechanisms to explicitly locate the editing region, as used in other approaches." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Furthermore, we present the CLIP-T value of each output in Fig. 5, with the best value per sample underlined. As" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28518" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 71, + 526, + 646 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 526, + 646 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 526, + 646 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 526, + 646 + ], + "type": "image", + "image_path": "ed0387408d761004085dcd7a34b369491737693aba35da7da49fcfcc1b204dd4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 652, + 555, + 709 + ], + "lines": [ + { + "bbox": [ + 55, + 652, + 555, + 709 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 555, + 709 + ], + "type": "text", + "content": "Figure 5. Comparison with state-of-the-art approaches for instruction-guided image editing, including HIVE [26], Inst-Inpaint (I-Inp) [24], Watch Your Steps (WYS) [11], ZONE [8], MagicBrush (MagBr) [25], InstructPix2Pix (IP2P) [1] showcasing the strength of our approach. CLIP-T value of each output is shown at its top-left corner, with the best value per row underlined. Note that the image with the best CLIP-T score is not necessarily the visually best result, underscoring the deficiencies of conventional metrics (including CLIP-I and DINO-I shown in the supplemental) for measuring the quality of image edits." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28519" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 79, + 293, + 157 + ], + "blocks": [ + { + "bbox": [ + 89, + 71, + 160, + 78 + ], + "lines": [ + { + "bbox": [ + 89, + 71, + 160, + 78 + ], + "spans": [ + { + "bbox": [ + 89, + 71, + 160, + 78 + ], + "type": "text", + "content": "MagBr data (single-turn)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 179, + 71, + 292, + 78 + ], + "lines": [ + { + "bbox": [ + 179, + 71, + 292, + 78 + ], + "spans": [ + { + "bbox": [ + 179, + 71, + 292, + 78 + ], + "type": "text", + "content": "MagBr data (multi-turn) ZONE data" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 79, + 293, + 157 + ], + "lines": [ + { + "bbox": [ + 58, + 79, + 293, + 157 + ], + "spans": [ + { + "bbox": [ + 58, + 79, + 293, + 157 + ], + "type": "table", + "html": "
CLIP-T ↑CLIP-I ↑DINO-I ↑CLIP-T ↑CLIP-I ↑DINO-I ↑CLIP-T ↑
HIVE0.3030.8920.7460.2990.8520.6590.297
I-Inp0.2850.8870.7290.2770.8590.6540.267
MagBr0.3070.9290.8360.3030.8960.7590.292
WYS0.3130.9240.8150.3130.8870.7270.301
ZONE0.3010.9290.8240.3070.8960.7500.296
IP2P0.3000.8540.6450.2980.8240.5730.296
Ours0.3050.9110.8030.3010.8710.7210.297
(+1.67%)(+6.67%)(+24.5%)(+1.01%)(+5.70%)(+25.83%)(+0.34%)
", + "image_path": "cb82873f1f933285add237eb110bd952d67866c74d3c5ca67f695196523455f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 205, + 293, + 243 + ], + "blocks": [ + { + "bbox": [ + 55, + 159, + 295, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 295, + 194 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with baselines. An up arrow (↑) indicates higher values are better. The percentage improvements over baselines are in parentheses." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 205, + 293, + 243 + ], + "lines": [ + { + "bbox": [ + 58, + 205, + 293, + 243 + ], + "spans": [ + { + "bbox": [ + 58, + 205, + 293, + 243 + ], + "type": "table", + "html": "
vs. MagBr#%vs. IP2P#%IP2P vs. MagBr#%
ours win56654.16%ours win31029.67%IP2P win48546.41%
tie16816.08%tie61158.47%tie14814.16%
MagBr win31129.76%IP2P win12411.87%MagBr win41239.43%
", + "image_path": "344ed3b82f10c7b9fc8ff1449b792bc736f359f3cbe464d6455ba466da4c9018.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 63, + 267, + 285, + 340 + ], + "blocks": [ + { + "bbox": [ + 56, + 245, + 293, + 256 + ], + "lines": [ + { + "bbox": [ + 56, + 245, + 293, + 256 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 293, + 256 + ], + "type": "text", + "content": "Table 2. Method pairwise comparison responses from user study." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 267, + 285, + 340 + ], + "lines": [ + { + "bbox": [ + 63, + 267, + 285, + 340 + ], + "spans": [ + { + "bbox": [ + 63, + 267, + 285, + 340 + ], + "type": "table", + "html": "
MagBr dataZONE data
CLIP-T ↑CLIP-I ↑DINO-I ↑CLIP-T ↑
IP2P0.3000.8540.6450.296
Ours (w/o data)0.2990.8620.6710.295
Ours (w/o loss)0.3030.9030.7820.298
Ours0.3050.9110.8030.297
", + "image_path": "7675a1b14c64964c8309a1e9e6e7778d5ce4d72518a5a9c4d07ad9aff7ea53b7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 342, + 296, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 407 + ], + "type": "text", + "content": "Table 3. Ablation study. An up arrow " + }, + { + "bbox": [ + 55, + 342, + 296, + 407 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 55, + 342, + 296, + 407 + ], + "type": "text", + "content": " means higher values are better. We trained two variants for our model: one trained on the original IP2P dataset without refined instructions (\"Ours (w/o data)\") and the other trained on our refined dataset without the Instruct-CLIP guidance loss (\"Ours (w/o loss)\"). We can see both the refined data and the loss help improve results." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 422, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 482 + ], + "type": "text", + "content": "shown, these values – as well as the CLIP-I and DINO-I values provided in the supplementary material – are not indicative of visual quality, but are included for completeness. Table 1 provides a quantitative comparison of our model in both single-turn and multi-turn edit scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 491, + 129, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 129, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 129, + 502 + ], + "type": "text", + "content": "4.4. User study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 508, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 639 + ], + "type": "text", + "content": "We also conducted a user study with 104 participants (60 male, 43 female, 1 non-binary) to compare our method with MagicBrush (MagBr) [25] and InstructPix2Pix (IP2P) [1] on the InstructBrush benchmark [28], ensuring fairness since no method was trained on similar data. Each person evaluated 33 randomly sampled output pairs based on the input image and edit instruction, and select which was better or if they were similar. We balanced the study across all pairwise comparisons, with 1,045 responses per pair. As shown in Table 2, participants overall found our method better than IP2P (17.8% more wins) and MagBr (24.4%)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 648, + 147, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 147, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 147, + 661 + ], + "type": "text", + "content": "4.5. Ablation study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": "To assess the impact of refined instructions and our Instruct-CLIP guidance loss, we trained two model variants: one on the original IP2P dataset with the Instruct-CLIP guidance loss (\"Ours (w/o data)) and another on our re" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 71, + 353, + 78 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 353, + 78 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 353, + 78 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 378, + 71, + 393, + 78 + ], + "lines": [ + { + "bbox": [ + 378, + 71, + 393, + 78 + ], + "spans": [ + { + "bbox": [ + 378, + 71, + 393, + 78 + ], + "type": "text", + "content": "IP2P" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 412, + 71, + 457, + 78 + ], + "lines": [ + { + "bbox": [ + 412, + 71, + 457, + 78 + ], + "spans": [ + { + "bbox": [ + 412, + 71, + 457, + 78 + ], + "type": "text", + "content": "Ours (w/o data)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 459, + 71, + 504, + 78 + ], + "lines": [ + { + "bbox": [ + 459, + 71, + 504, + 78 + ], + "spans": [ + { + "bbox": [ + 459, + 71, + 504, + 78 + ], + "type": "text", + "content": "Ours (w/o loss)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 522, + 71, + 537, + 77 + ], + "lines": [ + { + "bbox": [ + 522, + 71, + 537, + 77 + ], + "spans": [ + { + "bbox": [ + 522, + 71, + 537, + 77 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 79, + 363, + 125 + ], + "blocks": [ + { + "bbox": [ + 316, + 79, + 363, + 125 + ], + "lines": [ + { + "bbox": [ + 316, + 79, + 363, + 125 + ], + "spans": [ + { + "bbox": [ + 316, + 79, + 363, + 125 + ], + "type": "image", + "image_path": "114f37cd7382922a5459e3bb0cab72b89c4279caeff9c7aca7fc3a73c313a867.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 364, + 79, + 410, + 125 + ], + "blocks": [ + { + "bbox": [ + 364, + 79, + 410, + 125 + ], + "lines": [ + { + "bbox": [ + 364, + 79, + 410, + 125 + ], + "spans": [ + { + "bbox": [ + 364, + 79, + 410, + 125 + ], + "type": "image", + "image_path": "1164a49dd8c5332dc26641af3fc7b1865d55e7ddf7ed43eb74d28d2085fa818e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 411, + 79, + 457, + 125 + ], + "blocks": [ + { + "bbox": [ + 411, + 79, + 457, + 125 + ], + "lines": [ + { + "bbox": [ + 411, + 79, + 457, + 125 + ], + "spans": [ + { + "bbox": [ + 411, + 79, + 457, + 125 + ], + "type": "image", + "image_path": "7e158995cd06d2fdb580edc5b8ab12456ffd74b2a3a908eb2173167d601f6c2d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 460, + 79, + 505, + 125 + ], + "blocks": [ + { + "bbox": [ + 460, + 79, + 505, + 125 + ], + "lines": [ + { + "bbox": [ + 460, + 79, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 460, + 79, + 505, + 125 + ], + "type": "image", + "image_path": "96a48d934089c6de468a5f8221de8bab1a4f9f0813dca88a2c5b4a0f2b431074.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 506, + 79, + 552, + 125 + ], + "blocks": [ + { + "bbox": [ + 506, + 79, + 552, + 125 + ], + "lines": [ + { + "bbox": [ + 506, + 79, + 552, + 125 + ], + "spans": [ + { + "bbox": [ + 506, + 79, + 552, + 125 + ], + "type": "image", + "image_path": "8ce8d12ebb8fa2a63cfa40cbde8b24f0b8ff34ff41334cfc0e8a465118c2c06f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 126, + 436, + 136 + ], + "lines": [ + { + "bbox": [ + 315, + 126, + 436, + 136 + ], + "spans": [ + { + "bbox": [ + 315, + 126, + 436, + 136 + ], + "type": "text", + "content": "\"Place a floor lamp next to the chair.\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 138, + 362, + 184 + ], + "blocks": [ + { + "bbox": [ + 316, + 138, + 362, + 184 + ], + "lines": [ + { + "bbox": [ + 316, + 138, + 362, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 138, + 362, + 184 + ], + "type": "image", + "image_path": "3f11d58e9637cd11db9cd21a914e49b9da1cdcc52d259a00446426cada90f0ce.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 363, + 138, + 410, + 184 + ], + "blocks": [ + { + "bbox": [ + 363, + 138, + 410, + 184 + ], + "lines": [ + { + "bbox": [ + 363, + 138, + 410, + 184 + ], + "spans": [ + { + "bbox": [ + 363, + 138, + 410, + 184 + ], + "type": "image", + "image_path": "1f5d11a7dd568828f689e01b9f6eea4683d09a4133f3048193e944d15c1e2b94.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 411, + 138, + 457, + 184 + ], + "blocks": [ + { + "bbox": [ + 411, + 138, + 457, + 184 + ], + "lines": [ + { + "bbox": [ + 411, + 138, + 457, + 184 + ], + "spans": [ + { + "bbox": [ + 411, + 138, + 457, + 184 + ], + "type": "image", + "image_path": "d6e99f324cebd2924f4db5dd6da33ce11110896a658f163d3ec6ea30162872b8.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 459, + 138, + 505, + 184 + ], + "blocks": [ + { + "bbox": [ + 459, + 138, + 505, + 184 + ], + "lines": [ + { + "bbox": [ + 459, + 138, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 459, + 138, + 505, + 184 + ], + "type": "image", + "image_path": "2809d23df25649f9773a2c22c7db7f57488f17d2dd9f6ebf60096e0306244184.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 506, + 138, + 552, + 184 + ], + "blocks": [ + { + "bbox": [ + 506, + 138, + 552, + 184 + ], + "lines": [ + { + "bbox": [ + 506, + 138, + 552, + 184 + ], + "spans": [ + { + "bbox": [ + 506, + 138, + 552, + 184 + ], + "type": "image", + "image_path": "fd15791da0232bba74907688794d3148c1ca36913d7369576f610384f3b4089e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 186, + 425, + 195 + ], + "lines": [ + { + "bbox": [ + 315, + 186, + 425, + 195 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 425, + 195 + ], + "type": "text", + "content": "\"Replace the truck with a tractor\"" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 198, + 553, + 241 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 241 + ], + "type": "text", + "content": "Figure 6. Effect of refined instructions and Instruct-CLIP guidance loss. Compared to variants of our model trained without refined instructions (\"Ours (w/o data)\") or without the guidance loss (\"Ours (w/o loss)\"), our model produces superior results." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 316, + 252, + 340, + 264 + ], + "lines": [ + { + "bbox": [ + 316, + 252, + 340, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 340, + 264 + ], + "type": "text", + "content": "Ori" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 353, + 254, + 372, + 263 + ], + "lines": [ + { + "bbox": [ + 353, + 254, + 372, + 263 + ], + "spans": [ + { + "bbox": [ + 353, + 254, + 372, + 263 + ], + "type": "text", + "content": "IP2P" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 372, + 254, + 404, + 263 + ], + "lines": [ + { + "bbox": [ + 372, + 254, + 404, + 263 + ], + "spans": [ + { + "bbox": [ + 372, + 254, + 404, + 263 + ], + "type": "text", + "content": "MagBr" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 440, + 254, + 454, + 263 + ], + "lines": [ + { + "bbox": [ + 440, + 254, + 454, + 263 + ], + "spans": [ + { + "bbox": [ + 440, + 254, + 454, + 263 + ], + "type": "text", + "content": "Ori" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 468, + 254, + 486, + 263 + ], + "lines": [ + { + "bbox": [ + 468, + 254, + 486, + 263 + ], + "spans": [ + { + "bbox": [ + 468, + 254, + 486, + 263 + ], + "type": "text", + "content": "IP2P" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 499, + 254, + 543, + 263 + ], + "lines": [ + { + "bbox": [ + 499, + 254, + 543, + 263 + ], + "spans": [ + { + "bbox": [ + 499, + 254, + 543, + 263 + ], + "type": "text", + "content": "IagBr Ours" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 318, + 264, + 328, + 271 + ], + "lines": [ + { + "bbox": [ + 318, + 264, + 328, + 271 + ], + "spans": [ + { + "bbox": [ + 318, + 264, + 328, + 271 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 328, + 264, + 340, + 271 + ], + "lines": [ + { + "bbox": [ + 328, + 264, + 340, + 271 + ], + "spans": [ + { + "bbox": [ + 328, + 264, + 340, + 271 + ], + "type": "text", + "content": "②" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 342, + 264, + 355, + 271 + ], + "lines": [ + { + "bbox": [ + 342, + 264, + 355, + 271 + ], + "spans": [ + { + "bbox": [ + 342, + 264, + 355, + 271 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 356, + 264, + 370, + 271 + ], + "lines": [ + { + "bbox": [ + 356, + 264, + 370, + 271 + ], + "spans": [ + { + "bbox": [ + 356, + 264, + 370, + 271 + ], + "type": "inline_equation", + "content": "{13}/{14}" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 371, + 264, + 390, + 271 + ], + "lines": [ + { + "bbox": [ + 371, + 264, + 390, + 271 + ], + "spans": [ + { + "bbox": [ + 371, + 264, + 390, + 271 + ], + "type": "inline_equation", + "content": "{13}/{14}" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 391, + 264, + 403, + 271 + ], + "lines": [ + { + "bbox": [ + 391, + 264, + 403, + 271 + ], + "spans": [ + { + "bbox": [ + 391, + 264, + 403, + 271 + ], + "type": "text", + "content": "③" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 404, + 264, + 417, + 271 + ], + "lines": [ + { + "bbox": [ + 404, + 264, + 417, + 271 + ], + "spans": [ + { + "bbox": [ + 404, + 264, + 417, + 271 + ], + "type": "text", + "content": "② " + }, + { + "bbox": [ + 404, + 264, + 417, + 271 + ], + "type": "inline_equation", + "content": "{\\mathrm{{CO}}}_{2}" + }, + { + "bbox": [ + 404, + 264, + 417, + 271 + ], + "type": "text", + "content": " : 溶液" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 418, + 264, + 431, + 271 + ], + "lines": [ + { + "bbox": [ + 418, + 264, + 431, + 271 + ], + "spans": [ + { + "bbox": [ + 418, + 264, + 431, + 271 + ], + "type": "text", + "content": "③" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 432, + 264, + 445, + 271 + ], + "lines": [ + { + "bbox": [ + 432, + 264, + 445, + 271 + ], + "spans": [ + { + "bbox": [ + 432, + 264, + 445, + 271 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 446, + 264, + 457, + 271 + ], + "lines": [ + { + "bbox": [ + 446, + 264, + 457, + 271 + ], + "spans": [ + { + "bbox": [ + 446, + 264, + 457, + 271 + ], + "type": "inline_equation", + "content": "\\frac{3}{1} + u + {4q} = 1 + {uq}" + }, + { + "bbox": [ + 446, + 264, + 457, + 271 + ], + "type": "text", + "content": " dH" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 459, + 264, + 472, + 271 + ], + "lines": [ + { + "bbox": [ + 459, + 264, + 472, + 271 + ], + "spans": [ + { + "bbox": [ + 459, + 264, + 472, + 271 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 473, + 264, + 492, + 271 + ], + "lines": [ + { + "bbox": [ + 473, + 264, + 492, + 271 + ], + "spans": [ + { + "bbox": [ + 473, + 264, + 492, + 271 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 493, + 264, + 506, + 271 + ], + "lines": [ + { + "bbox": [ + 493, + 264, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 493, + 264, + 506, + 271 + ], + "type": "inline_equation", + "content": "{10} : \\frac{{a}_{n + 1}}{{b}_{n + 1}} = \\frac{2{a}_{n + 1}}{2{b}_{n + 1}} = \\frac{{a}_{1} + {a}_{{2n} + 1}}{{b}_{1} + {b}_{{2n} + 1}} =" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 507, + 264, + 520, + 271 + ], + "lines": [ + { + "bbox": [ + 507, + 264, + 520, + 271 + ], + "spans": [ + { + "bbox": [ + 507, + 264, + 520, + 271 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 521, + 264, + 533, + 271 + ], + "lines": [ + { + "bbox": [ + 521, + 264, + 533, + 271 + ], + "spans": [ + { + "bbox": [ + 521, + 264, + 533, + 271 + ], + "type": "text", + "content": "。" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 318, + 272, + 328, + 279 + ], + "lines": [ + { + "bbox": [ + 318, + 272, + 328, + 279 + ], + "spans": [ + { + "bbox": [ + 318, + 272, + 328, + 279 + ], + "type": "text", + "content": "Y" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 328, + 272, + 340, + 279 + ], + "lines": [ + { + "bbox": [ + 328, + 272, + 340, + 279 + ], + "spans": [ + { + "bbox": [ + 328, + 272, + 340, + 279 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 342, + 272, + 355, + 279 + ], + "lines": [ + { + "bbox": [ + 342, + 272, + 355, + 279 + ], + "spans": [ + { + "bbox": [ + 342, + 272, + 355, + 279 + ], + "type": "text", + "content": "Y" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 356, + 272, + 370, + 279 + ], + "lines": [ + { + "bbox": [ + 356, + 272, + 370, + 279 + ], + "spans": [ + { + "bbox": [ + 356, + 272, + 370, + 279 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 371, + 272, + 390, + 279 + ], + "lines": [ + { + "bbox": [ + 371, + 272, + 390, + 279 + ], + "spans": [ + { + "bbox": [ + 371, + 272, + 390, + 279 + ], + "type": "inline_equation", + "content": "\\therefore m : x = 1" + }, + { + "bbox": [ + 371, + 272, + 390, + 279 + ], + "type": "text", + "content": " 或 " + }, + { + "bbox": [ + 371, + 272, + 390, + 279 + ], + "type": "inline_equation", + "content": "{3x} + {4y} + 1 = 0" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 391, + 272, + 403, + 279 + ], + "lines": [ + { + "bbox": [ + 391, + 272, + 403, + 279 + ], + "spans": [ + { + "bbox": [ + 391, + 272, + 403, + 279 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 404, + 272, + 417, + 279 + ], + "lines": [ + { + "bbox": [ + 404, + 272, + 417, + 279 + ], + "spans": [ + { + "bbox": [ + 404, + 272, + 417, + 279 + ], + "type": "inline_equation", + "content": "{12}/{12}" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 430, + 272, + 443, + 279 + ], + "lines": [ + { + "bbox": [ + 430, + 272, + 443, + 279 + ], + "spans": [ + { + "bbox": [ + 430, + 272, + 443, + 279 + ], + "type": "inline_equation", + "content": "{14} \\times {25}" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 446, + 272, + 457, + 279 + ], + "lines": [ + { + "bbox": [ + 446, + 272, + 457, + 279 + ], + "spans": [ + { + "bbox": [ + 446, + 272, + 457, + 279 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 459, + 272, + 472, + 279 + ], + "lines": [ + { + "bbox": [ + 459, + 272, + 472, + 279 + ], + "spans": [ + { + "bbox": [ + 459, + 272, + 472, + 279 + ], + "type": "inline_equation", + "content": "\\Rightarrow \\;{3}^{x} = - 2 + 2\\sqrt{2} + 2\\sqrt{2} = - 4" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 473, + 272, + 486, + 279 + ], + "lines": [ + { + "bbox": [ + 473, + 272, + 486, + 279 + ], + "spans": [ + { + "bbox": [ + 473, + 272, + 486, + 279 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 488, + 272, + 506, + 279 + ], + "lines": [ + { + "bbox": [ + 488, + 272, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 488, + 272, + 506, + 279 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 507, + 272, + 520, + 279 + ], + "lines": [ + { + "bbox": [ + 507, + 272, + 520, + 279 + ], + "spans": [ + { + "bbox": [ + 507, + 272, + 520, + 279 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 522, + 272, + 534, + 279 + ], + "lines": [ + { + "bbox": [ + 522, + 272, + 534, + 279 + ], + "spans": [ + { + "bbox": [ + 522, + 272, + 534, + 279 + ], + "type": "inline_equation", + "content": "\\downarrow x" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 318, + 280, + 328, + 287 + ], + "lines": [ + { + "bbox": [ + 318, + 280, + 328, + 287 + ], + "spans": [ + { + "bbox": [ + 318, + 280, + 328, + 287 + ], + "type": "inline_equation", + "content": "{13}/{14}" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 328, + 280, + 340, + 287 + ], + "lines": [ + { + "bbox": [ + 328, + 280, + 340, + 287 + ], + "spans": [ + { + "bbox": [ + 328, + 280, + 340, + 287 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 342, + 280, + 355, + 287 + ], + "lines": [ + { + "bbox": [ + 342, + 280, + 355, + 287 + ], + "spans": [ + { + "bbox": [ + 342, + 280, + 355, + 287 + ], + "type": "inline_equation", + "content": "{13}/{14}" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 356, + 280, + 370, + 287 + ], + "lines": [ + { + "bbox": [ + 356, + 280, + 370, + 287 + ], + "spans": [ + { + "bbox": [ + 356, + 280, + 370, + 287 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + }, + { + "bbox": [ + 356, + 280, + 370, + 287 + ], + "type": "text", + "content": " ;" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 371, + 280, + 390, + 287 + ], + "lines": [ + { + "bbox": [ + 371, + 280, + 390, + 287 + ], + "spans": [ + { + "bbox": [ + 371, + 280, + 390, + 287 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + }, + { + "bbox": [ + 371, + 280, + 390, + 287 + ], + "type": "text", + "content": " ;" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 391, + 280, + 403, + 287 + ], + "lines": [ + { + "bbox": [ + 391, + 280, + 403, + 287 + ], + "spans": [ + { + "bbox": [ + 391, + 280, + 403, + 287 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 404, + 280, + 417, + 287 + ], + "lines": [ + { + "bbox": [ + 404, + 280, + 417, + 287 + ], + "spans": [ + { + "bbox": [ + 404, + 280, + 417, + 287 + ], + "type": "inline_equation", + "content": "{13}/{14}" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 430, + 280, + 443, + 287 + ], + "lines": [ + { + "bbox": [ + 430, + 280, + 443, + 287 + ], + "spans": [ + { + "bbox": [ + 430, + 280, + 443, + 287 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 446, + 280, + 457, + 287 + ], + "lines": [ + { + "bbox": [ + 446, + 280, + 457, + 287 + ], + "spans": [ + { + "bbox": [ + 446, + 280, + 457, + 287 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 459, + 280, + 472, + 287 + ], + "lines": [ + { + "bbox": [ + 459, + 280, + 472, + 287 + ], + "spans": [ + { + "bbox": [ + 459, + 280, + 472, + 287 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 473, + 280, + 486, + 287 + ], + "lines": [ + { + "bbox": [ + 473, + 280, + 486, + 287 + ], + "spans": [ + { + "bbox": [ + 473, + 280, + 486, + 287 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 488, + 280, + 506, + 287 + ], + "lines": [ + { + "bbox": [ + 488, + 280, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 488, + 280, + 506, + 287 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 507, + 280, + 520, + 287 + ], + "lines": [ + { + "bbox": [ + 507, + 280, + 520, + 287 + ], + "spans": [ + { + "bbox": [ + 507, + 280, + 520, + 287 + ], + "type": "text", + "content": "(1)" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 522, + 280, + 534, + 287 + ], + "lines": [ + { + "bbox": [ + 522, + 280, + 534, + 287 + ], + "spans": [ + { + "bbox": [ + 522, + 280, + 534, + 287 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 318, + 293, + 419, + 303 + ], + "lines": [ + { + "bbox": [ + 318, + 293, + 419, + 303 + ], + "spans": [ + { + "bbox": [ + 318, + 293, + 419, + 303 + ], + "type": "text", + "content": "\"Turn orange shirt into blue\"" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 434, + 293, + 446, + 302 + ], + "lines": [ + { + "bbox": [ + 434, + 293, + 446, + 302 + ], + "spans": [ + { + "bbox": [ + 434, + 293, + 446, + 302 + ], + "type": "text", + "content": "\"Ac" + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 447, + 293, + 460, + 302 + ], + "lines": [ + { + "bbox": [ + 447, + 293, + 460, + 302 + ], + "spans": [ + { + "bbox": [ + 447, + 293, + 460, + 302 + ], + "type": "text", + "content": "ld a" + } + ] + } + ], + "index": 82, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 461, + 293, + 473, + 302 + ], + "lines": [ + { + "bbox": [ + 461, + 293, + 473, + 302 + ], + "spans": [ + { + "bbox": [ + 461, + 293, + 473, + 302 + ], + "type": "text", + "content": "n ea" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 474, + 293, + 487, + 302 + ], + "lines": [ + { + "bbox": [ + 474, + 293, + 487, + 302 + ], + "spans": [ + { + "bbox": [ + 474, + 293, + 487, + 302 + ], + "type": "text", + "content": "gle" + } + ] + } + ], + "index": 84, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 488, + 293, + 507, + 302 + ], + "lines": [ + { + "bbox": [ + 488, + 293, + 507, + 302 + ], + "spans": [ + { + "bbox": [ + 488, + 293, + 507, + 302 + ], + "type": "text", + "content": "above" + } + ] + } + ], + "index": 85, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 508, + 293, + 519, + 302 + ], + "lines": [ + { + "bbox": [ + 508, + 293, + 519, + 302 + ], + "spans": [ + { + "bbox": [ + 508, + 293, + 519, + 302 + ], + "type": "text", + "content": "the" + } + ] + } + ], + "index": 86, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 318, + 304, + 328, + 312 + ], + "lines": [ + { + "bbox": [ + 318, + 304, + 328, + 312 + ], + "spans": [ + { + "bbox": [ + 318, + 304, + 328, + 312 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 87, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 328, + 304, + 340, + 312 + ], + "lines": [ + { + "bbox": [ + 328, + 304, + 340, + 312 + ], + "spans": [ + { + "bbox": [ + 328, + 304, + 340, + 312 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 88, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 342, + 304, + 355, + 312 + ], + "lines": [ + { + "bbox": [ + 342, + 304, + 355, + 312 + ], + "spans": [ + { + "bbox": [ + 342, + 304, + 355, + 312 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 89, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 356, + 304, + 370, + 312 + ], + "lines": [ + { + "bbox": [ + 356, + 304, + 370, + 312 + ], + "spans": [ + { + "bbox": [ + 356, + 304, + 370, + 312 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 90, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 371, + 304, + 390, + 312 + ], + "lines": [ + { + "bbox": [ + 371, + 304, + 390, + 312 + ], + "spans": [ + { + "bbox": [ + 371, + 304, + 390, + 312 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 91, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 391, + 304, + 403, + 312 + ], + "lines": [ + { + "bbox": [ + 391, + 304, + 403, + 312 + ], + "spans": [ + { + "bbox": [ + 391, + 304, + 403, + 312 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 92, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 404, + 304, + 417, + 312 + ], + "lines": [ + { + "bbox": [ + 404, + 304, + 417, + 312 + ], + "spans": [ + { + "bbox": [ + 404, + 304, + 417, + 312 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 93, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 418, + 304, + 431, + 312 + ], + "lines": [ + { + "bbox": [ + 418, + 304, + 431, + 312 + ], + "spans": [ + { + "bbox": [ + 418, + 304, + 431, + 312 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 94, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 432, + 304, + 451, + 312 + ], + "lines": [ + { + "bbox": [ + 432, + 304, + 451, + 312 + ], + "spans": [ + { + "bbox": [ + 432, + 304, + 451, + 312 + ], + "type": "inline_equation", + "content": "\\left\\{ {2 : {x}^{\\prime }\\left( x\\right) = {y}^{\\prime }\\left( x\\right) }\\right\\} : y = {13}\\left( x\\right) = {13}{y}^{\\prime }\\left( x\\right)" + } + ] + } + ], + "index": 95, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 452, + 304, + 465, + 312 + ], + "lines": [ + { + "bbox": [ + 452, + 304, + 465, + 312 + ], + "spans": [ + { + "bbox": [ + 452, + 304, + 465, + 312 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 96, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 466, + 304, + 479, + 312 + ], + "lines": [ + { + "bbox": [ + 466, + 304, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 466, + 304, + 479, + 312 + ], + "type": "inline_equation", + "content": "\\frac{1}{2} =" + } + ] + } + ], + "index": 97, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 481, + 304, + 500, + 312 + ], + "lines": [ + { + "bbox": [ + 481, + 304, + 500, + 312 + ], + "spans": [ + { + "bbox": [ + 481, + 304, + 500, + 312 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + }, + { + "bbox": [ + 481, + 304, + 500, + 312 + ], + "type": "text", + "content": " ;" + } + ] + } + ], + "index": 98, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 501, + 304, + 514, + 312 + ], + "lines": [ + { + "bbox": [ + 501, + 304, + 514, + 312 + ], + "spans": [ + { + "bbox": [ + 501, + 304, + 514, + 312 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + } + ] + } + ], + "index": 99, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 318, + 335, + 410, + 344 + ], + "lines": [ + { + "bbox": [ + 318, + 335, + 410, + 344 + ], + "spans": [ + { + "bbox": [ + 318, + 335, + 410, + 344 + ], + "type": "text", + "content": "\"Delete the wooden boat\"" + } + ] + } + ], + "index": 100, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 434, + 335, + 527, + 345 + ], + "lines": [ + { + "bbox": [ + 434, + 335, + 527, + 345 + ], + "spans": [ + { + "bbox": [ + 434, + 335, + 527, + 345 + ], + "type": "text", + "content": "\"Remove the white plane\"" + } + ] + } + ], + "index": 101, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 349, + 553, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 349, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 553, + 381 + ], + "type": "text", + "content": "Figure 7. Examples of failure cases compared with Instruct-Pix2Pix (IP2P) [1] and MagicBrush (MagBr) [25] performed on the original (Ori) images." + } + ] + } + ], + "index": 102, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "content": "fined dataset without the guidance loss (\"Ours (w/o loss)\"). We compare these with IP2P and our full method in Table 3, and visual examples are provided in Fig. 6. Results show that refined data significantly improve visual alignment with ground truth, evidenced by higher CLIP and DINO scores, while the guidance loss further enhances this alignment." + } + ] + } + ], + "index": 103 + }, + { + "bbox": [ + 313, + 484, + 391, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 391, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 391, + 495 + ], + "type": "text", + "content": "4.6. Limitations" + } + ] + } + ], + "index": 104 + }, + { + "bbox": [ + 313, + 502, + 554, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 597 + ], + "type": "text", + "content": "Despite being able to correct a lot of inaccurate edit instructions in the training set, our method is still heavily influenced by the limitations of the generative method [6] these training images come from, which include color bleeding outside the intended edit region (Fig. 7, top left), and incomplete object addition (Fig. 7, top right). While our model respects the original images better, it sometimes struggles to remove objects in the original images (Fig. 7, bottom)." + } + ] + } + ], + "index": 105 + }, + { + "bbox": [ + 313, + 609, + 387, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 387, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 387, + 621 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 106 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "We have presented Instruct-CLIP, a self-supervised method for instruction-guided image editing that learns the semantic changes between original and edited images to refine edit instructions in datasets. Applying I-CLIP to the InstructPix2Pix dataset yields over 120K refined samples, which we use to fine-tune its model with our I-CLIP-guided loss function and generate better edit results." + } + ] + } + ], + "index": 107 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28520" + } + ] + } + ], + "index": 108 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2Pix: Learning to follow image editing instructions. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 200 + ], + "type": "text", + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "type": "text", + "content": "[3] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 34:8780-8794, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "type": "text", + "content": "[4] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis, march 2024. URL http://arxiv.org/abs/2403.03206, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 303, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 294, + 345 + ], + "type": "text", + "content": "[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 389 + ], + "type": "text", + "content": "[6] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 391, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 295, + 434 + ], + "type": "text", + "content": "[7] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-Edit: A high-quality dataset for instruction-based image editing. arXiv preprint arXiv:2404.09990, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 490 + ], + "type": "text", + "content": "[8] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, Jianzhuang Liu, et al. ZONE: Zero-shot instruction-guided local editing. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 6254-6263, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 524 + ], + "type": "text", + "content": "[9] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. DeCap: Decoding CLIP latents for zero-shot captioning via text-only training. arXiv preprint arXiv:2303.03032, 2023. 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "type": "text", + "content": "[10] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent Consistency Models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 635 + ], + "type": "text", + "content": "[11] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Marcus A Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G Derpanis, and Igor Gilitschenski. Watch Your Steps: Local image and scene editing by text instructions. In European Conference on Computer Vision (ECCV), pages 111-129. Springer, 2025. 1, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "text", + "content": "[12] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text Inversion for editing real images using guided diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 6038–6047, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 690, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 690, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 690, + 295, + 713 + ], + "type": "text", + "content": "[13] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "type": "text", + "content": "[14] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In SIGGRAPH, pages 1-11, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 195 + ], + "type": "text", + "content": "[15] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "text", + "content": "[16] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 232, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 297 + ], + "type": "text", + "content": "[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 298, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 553, + 352 + ], + "type": "text", + "content": "[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 387 + ], + "type": "text", + "content": "[19] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "type": "text", + "content": "[20] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-Play diffusion features for text-driven image-to-image translation. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921–1930, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "text", + "content": "[21] Bram Wallace, Akash Gokul, and Nikhil Naik. EDICT: Exact diffusion inversion via coupled transformations. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 22532-22541, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 489, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 553 + ], + "type": "text", + "content": "[22] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 1900-1910, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 556, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 553, + 599 + ], + "type": "text", + "content": "[23] Ling Yang, Bohan Zeng, Jiaming Liu, Hong Li, Minghao Xu, Wentao Zhang, and Shuicheng Yan. EditWorld: Simulating world dynamics for instruction-following image editing. arXiv preprint arXiv:2405.14785, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "type": "text", + "content": "[24] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-Inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 646, + 553, + 689 + ], + "type": "text", + "content": "[25] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. MagicBrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[26] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28521" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 217 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "Stefano Ermon, Caiming Xiong, and Ran Xu. HIVE: Harnessing human feedback for instructional visual editing. arXiv preprint arXiv:2303.09618, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[27] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "text", + "content": "[28] Ruoyu Zhao, Qingnan Fan, Fei Kou, Shuai Qin, Hong Gu, Wei Wu, Pengcheng Xu, Mingrui Zhu, Nannan Wang, and Xinbo Gao. InstructBrush: Learning attention-based instruction optimization for image editing. arXiv preprint arXiv:2403.18660, 2024. 8" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28522" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Instruction-based Image Manipulation by Watching How Things Move/6db53ffc-36dc-4f4f-8a58-b9c5bb9ce24b_content_list.json b/2025/Instruction-based Image Manipulation by Watching How Things Move/6db53ffc-36dc-4f4f-8a58-b9c5bb9ce24b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5bb70b5a8864a6167d74ec3f64b171aa485b3b1b --- /dev/null +++ b/2025/Instruction-based Image Manipulation by Watching How Things Move/6db53ffc-36dc-4f4f-8a58-b9c5bb9ce24b_content_list.json @@ -0,0 +1,2381 @@ +[ + { + "type": "text", + "text": "Instruction-based Image Manipulation by Watching How Things Move", + "text_level": 1, + "bbox": [ + 140, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingdeng Cao $^{1\\dagger}$ Xuaner Zhang $^{2}$ Yinqiang Zheng $^{1}$ Zhihao Xia $^{2}$ $^{1}$ The University of Tokyo $^{2}$ Adobe", + "bbox": [ + 225, + 179, + 772, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d25058744bf9235de49aceef260d561d3ab706146174d6e4fdb099546b211f0b.jpg", + "image_caption": [ + "Lower the horse's head" + ], + "image_footnote": [], + "bbox": [ + 93, + 268, + 359, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a14f0a26670d007471a58159de2ed31c450ced8c865ffe825d81222bba32aa8.jpg", + "image_caption": [ + "Change the view to the side", + "Figure 1. We propose InstructMove, an instruction-based image editing model trained on frame pairs from videos with instructions generated by Multimodal LLMs. Our model excels at non-rigid editing, such as adjusting subject poses, expressions, and altering viewpoints, while maintaining content consistency. Additionally, our method supports precise, localized edits through the integration of masks, human poses, and other control mechanisms." + ], + "image_footnote": [], + "bbox": [ + 93, + 382, + 359, + 486 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17f7a0c35cb4815bc62ff01c8a3c211d75d86329308db058b578c5e8da90ac7d.jpg", + "image_caption": [ + "Make the man look angry" + ], + "image_footnote": [], + "bbox": [ + 364, + 268, + 633, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/15b0aa691a8b3eb16dd0081078756242750a2239fec0d998aa817a8478e3e9d9.jpg", + "image_caption": [ + "Move the camera to the left" + ], + "image_footnote": [], + "bbox": [ + 366, + 382, + 633, + 486 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d5fc7c1e13b1318b8f3033055211036a43f71df900efcf963af1a150e8db68d7.jpg", + "image_caption": [ + "Close the dog's eye" + ], + "image_footnote": [], + "bbox": [ + 638, + 268, + 903, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c153208fc770eae9915e8b12dfe87edef1bc4e8d1392b4613053c479e14a449c.jpg", + "image_caption": [ + "Have the man look at the side." + ], + "image_footnote": [], + "bbox": [ + 638, + 382, + 903, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 593, + 326, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces a novel dataset construction pipeline that samples pairs of frames from videos and uses multimodal large language models (MLLMs) to generate editing instructions for training instruction-based image manipulation models. Video frames inherently preserve the identity of subjects and scenes, ensuring consistent content preservation during editing. Additionally, video data captures diverse, natural dynamics—such as non-rigid subject motion and complex camera movements—that are difficult to model otherwise, making it an ideal source for scalable dataset construction. Using this approach, we create a new dataset to train InstructMove, a model capable of instruction-based complex manipulations that are difficult to achieve with synthetically generated datasets. Our model demonstrates state-of-the-art performance in tasks such as adjusting subject poses, rearranging elements, and altering", + "bbox": [ + 88, + 628, + 485, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "camera perspectives. The project page is available here.", + "bbox": [ + 513, + 580, + 883, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 628, + 643, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have seen remarkable progress in text-to-image (T2I) generation [10, 28, 31-33], with state-of-the-art methods now capable of producing high-quality and hyper-realistic images. The key to this success lies in training on vast datasets containing billions of images from the internet, paired with text descriptions generated by image captioning models [30]. In contrast, editing an existing image based on textual instructions remains challenging, with output quality still falling short and a limited range of editing types available. A major bottleneck in this area is the difficulty of obtaining a comparably large dataset of source-target-instruction triplets for training.", + "bbox": [ + 509, + 657, + 906, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous approaches [7, 8, 14, 20, 25, 26, 40] attempted to adapt pretrained T2I models without fine-tuning, using noise inversion to effectively \"regenerate\" the target image with modified textual prompts. However, these tuning-", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Work done during an internship at Adobe.", + "bbox": [ + 114, + 887, + 346, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2704", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "free methods are often slow and lack robustness. InstructPix2Pix [5] introduced a method to bootstrap a dataset of source images into editing triplets by generating editing instructions using a language model and creating target images using tuning-free techniques. Subsequent works [12, 16, 34, 42] have aimed to refine this process or filter the generated outputs. Nonetheless, a fundamental limitation remains: target images are synthetically generated, which constrains the model's potential due to data quality issues. Consequently, existing methods often struggle to preserve the appearance of edited subjects or to perform precise, detailed edits, as shown in Fig. 2.", + "bbox": [ + 89, + 90, + 482, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a novel approach for creating large-scale, instruction-based image-editing datasets using real images sampled from internet videos, capturing naturally realistic transformations from source to target. Our key insight is that video frames capture rich information on how things move — pose changes, element shifts, and camera movements — making them an ideal data source for training models for image manipulation. A similar insight was demonstrated in [3] to show that video frames are effective training data to fix coarse manual edits. We then leverage powerful multi-modal large language models (MLLMs), such as GPT-4o [1] and LLaVA [24] to generate descriptive text at or above human-level quality.", + "bbox": [ + 89, + 273, + 482, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our data construction starts with extracting video frames that undergo realistic yet complex transformations. Next, we use MLLMs to analyze these transformations and generate precise editing instructions. Additionally, we introduce a novel spatial conditioning strategy: instead of conventional channel concatenation, we concatenate the reference image with the noise input along the spatial dimension. This approach enables the model to access the reference image through cross-attention across the network, greatly enhancing its ability to perform flexible edits while preserving the appearance of the source image.", + "bbox": [ + 89, + 472, + 482, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By fine-tuning pretrained text-to-image models (e.g., Stable Diffusion [32]) on our constructed dataset with spatial conditioning strategy, we enable natural-language-driven image manipulations that were previously challenging. These include adjusting subject poses, rearranging elements, or altering camera perspectives, all while maintaining the integrity and details of the original image (see Fig. 1). Our approach also seamlessly integrates with mask-based guidance for local editing, and also with other control types [27, 43], allowing for versatile modifications.", + "bbox": [ + 89, + 640, + 482, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 792, + 434, + 806 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a data construction pipeline that uses video frames and MLLMs to create source-target-instruction triplets for training image manipulation models.", + "- We introduce a spatial conditioning strategy that conditions the reference image alongside the noise map to improve content preservation while enabling flexible edits." + ], + "bbox": [ + 91, + 810, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e140039e28e0c8e774d7413822af56ee66e52cb53dfc83bd44d16a777dcf5768.jpg", + "image_caption": [ + "Put the toy's legs together.", + "Figure 2. Existing methods struggle with complex edits on real photos, such as non-rigid transformations. They often either fail to follow the editing instructions or produce inconsistent outputs." + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 616, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3f334bb4adfccb5ca4d059758215446f601bae958d21de2e071ee82ec2b03d1c.jpg", + "image_caption": [ + "InstructPix2Pix" + ], + "image_footnote": [], + "bbox": [ + 617, + 89, + 712, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e88f24291cba54cf29b21ae285c78bd73ed96c02ede776ef1d5f1689f812c7e9.jpg", + "image_caption": [ + "MagicBrush" + ], + "image_footnote": [], + "bbox": [ + 712, + 89, + 807, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fce7696f8aa4f0a655e76491d9b7ce7b87181991db0d8f933ec90b7c1f3d9313.jpg", + "image_caption": [ + "UltraEdit" + ], + "image_footnote": [], + "bbox": [ + 808, + 89, + 903, + 162 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our approach enables a wider range of text-based image manipulations, including pose adjustments, element rearrangements, and perspective changes, while also supporting precise, localized edits through integration with additional control mechanisms.", + "bbox": [ + 513, + 256, + 906, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 348, + 653, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Zero-shot Image Editing with Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 373, + 905, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [15, 35, 37] have become the leading approach in text-to-image generation due to their ability to effectively translate noise into structured data distributions. Leveraging these pretrained models, several image editing techniques have been developed by manipulating the sampling process [4, 8, 17, 20, 22, 25, 26, 29], modifying the internal architecture of the denoising network [7, 14, 40], or incorporating additional optimization steps [19, 23, 44].", + "bbox": [ + 511, + 396, + 905, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For example, SDEdit adds noise to the image and then generates the desired edit by conditioning on a target prompt during the reverse denoising process. Other approaches, such as Prompt-to-Prompt [14], Plug-and-Play [40], and MasaCtrl [7], manipulate the attention mechanism after inverting the image to noise using DDIM inversion [36], allowing for edits that align with the desired prompt. More recent works [9, 13, 18, 26, 41] have refined the inversion process to enhance editing robustness and accuracy. While these zero-shot methods offer flexibility in editing images, they are often too slow and lack robustness, making them less suitable for practical applications.", + "bbox": [ + 511, + 518, + 906, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Training Instruction-based Editing Models", + "text_level": 1, + "bbox": [ + 511, + 710, + 879, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Comparing to modifying target image descriptions, a more intuitive and user-friendly editing method is to use direct instructions (e.g., “Have the man look at the camera”) to guide the editing process. InstructPix2Pix [5] pioneered this concept by fine-tuning Stable Diffusion [32] with editing instructions. To handle more complex instructions, methods like MGIE [11] and SmartEdit [16] integrate Multimodal LLMs to improve the understanding of editing instructions.", + "bbox": [ + 511, + 733, + 905, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A major challenge in training instruction-based editing models is constructing paired image data with corresponding instructions. InstructPix2Pix tackles this by using lan", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2705", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0a94d051d110361c181678232f814b24ae75d2a0ce179497a7ae00e0bf408311.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 118, + 166, + 138 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/99702fb5f7dfcc4e2a6d09418336b4b8c442d640b8e9c2b50e618ea164bf10ab.jpg", + "image_caption": [ + "(a) Frame selection: Sample frame pairs from internet videos." + ], + "image_footnote": [], + "bbox": [ + 174, + 101, + 318, + 155 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ff7d6e9debac14f82b95cc9e3634b44273ff853f832f718ab70a2c9a089567a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 99, + 408, + 154 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/23c9f03916ed4a7b487a74da9fabfe52ad1661f092bc5d203b4e2df067a642ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 99, + 478, + 154 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f80bd892402f7922c35661a1c28d859ca4edbd6c187374bcdbc36b9f89c3263c.jpg", + "image_caption": [ + "(b) Generate instructions:" + ], + "image_footnote": [], + "bbox": [ + 99, + 176, + 197, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/564018e6a8caf1ca93bb8e87c33fad7a4368fd63aa94dbb1c80169428fd53d99.jpg", + "image_caption": [ + "Adjust the woman's gaze from the book to looking directly at the camera.", + "Have the man give a thumbs up gesture." + ], + "image_footnote": [], + "bbox": [ + 210, + 178, + 462, + 194 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/754e56634d8add88a000ff2bb25be7dfdd335c9d820e9683e8518f9163b40aea.jpg", + "image_caption": [ + "(c) Generated source-target-instruction triplets for training:", + "\"Remove the smile from the woman's face.\"", + "Figure 3. Our data construction pipeline. (a) We begin by sampling suitable frame pairs from videos, ensuring realistic and moderate transformations. (b) These frame pairs are used to prompt Multimodal Large Language Models (MLLMs) to generate detailed editing instructions. (c) This process results in a large-scale dataset with realistic image pairs and precise editing instructions." + ], + "image_footnote": [], + "bbox": [ + 96, + 241, + 220, + 289 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/77bd7bb8f363f7b55723263591af664694625aa193d6313197c63311f99e053a.jpg", + "image_caption": [ + "\"Rotate the bird to the left and make its tail visible.\"" + ], + "image_footnote": [], + "bbox": [ + 225, + 241, + 349, + 289 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a5a764da79a68f9e7b2794bddbe68c5d30d43b5e38c80844239894c482feafc7.jpg", + "image_caption": [ + "\"Make the person looking at the laptop screen.\"" + ], + "image_footnote": [], + "bbox": [ + 354, + 241, + 477, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "guage models, such as GPT-3 [6], to generate both an edited image caption and an editing instruction. These captions are converted into images using Prompt-to-Prompt [14]. MagicBrush [42] constructs a small dataset by performing edits using DALL-E 2 [31]. InstructDiffusion [12] augments datasets with synthetic pairs for tasks like object removal and replacement. EmuEdit [34] improves data pair construction using Llama 2 [39] and local masks with Prompt-to-Prompt, combining these with other vision datasets to build a large-scale dataset.", + "bbox": [ + 88, + 407, + 482, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, in existing large-scale editing datasets, most target images are generated using synthetic methods, which often introduces artifacts and significant appearance deviations from the source images. These datasets are typically limited to high-level edits such as object overlay and style transfer. As a result, models trained on these datasets struggle with complex edits on real photos, such as non-rigid transformations and viewpoint changes. In contrast, our approach leverages real video frames as source and target images, using Multimodal LLMs to generate instructions directly from these frames. This allows for the construction of a more realistic dataset, capturing complex transformations and overcoming the limitations of previous methods.", + "bbox": [ + 89, + 559, + 483, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset Construction", + "text_level": 1, + "bbox": [ + 89, + 768, + 292, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike existing instruction-based editing datasets, which rely on tuning-free methods to generate target images, we propose a novel data construction pipeline that leverages real video frame pairs. By extracting frame pairs that capture meaningful transformations and using Multimodal Large Language Models (MLLMs) to generate detailed editing instructions, we create a more realistic and high-", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "quality editing dataset. In Sec. 3.1, we outline our approach for extracting suitable frame pairs from videos, followed by Sec. 3.2, where we describe how we use MLLMs to generate precise editing instructions. Finally, Sec. 3.3 provides an in-depth analysis of the resulting dataset. An overview of the data construction pipeline is illustrated in Fig. 3.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Sampling Frame Pairs from Videos", + "text_level": 1, + "bbox": [ + 511, + 191, + 820, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In a video, the same subjects and elements can appear in different ways, move naturally, or be shown from different angles. For example, a person might change their posture, expression, or movement over time. These changes provide useful clues about how a scene or perspective might look with small adjustments, making videos a great source for picking frame pairs to use in image editing tasks. However, it's important to sample the frame pairs with proper criteria to make sure they are suitable for training.", + "bbox": [ + 511, + 213, + 906, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Initial sampling of paired frames. We begin by captioning each video to generate an overall description. Videos containing specific keywords (e.g., landscape, abstract, still) that are not suitable for image editing are filtered out. We then sample two frames $(I^s, I^e)$ with a fixed time interval (i.e., 3-5 seconds) to ensure sufficient transformations occur in the objects of interest.", + "bbox": [ + 511, + 356, + 905, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion-based filtering. To refine the selection, we apply motion filtering to exclude frames with either too little or too much movement. Specifically, we use RAFT [38] to compute the optical flow $O_{I^s \\to I^e}$ from $I^s$ to $I^e$ and derive the flow magnitude $M = \\sqrt{O_x^2 + O_y^2}$ . Frame pairs with moderate movement are retained, while those with excessive or minimal motion are discarded. To further ensure consistency, we assess background changes by computing a background occlusion mask through backward warping of $I^e$ to $I^s$ using the estimated flow. By measuring the occlusion mask ratio, we filter out frames with significant background changes (i.e., large masked regions). This motion filtering pipeline allows us to obtain image pairs with realistic, moderate transformations.", + "bbox": [ + 511, + 469, + 906, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Instruction Generation with MLLMs", + "text_level": 1, + "bbox": [ + 511, + 696, + 834, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next, we describe how we generate accurate editing instructions $C$ given frame pairs $(I^s, I^e)$ . Previous approaches [5, 34] rely solely on text-based methods, using Large Language Models (LLMs) to generate instructions and captions of target images purely from the source caption. This approach significantly limits the variety of edits that can be captured in the dataset. As a result, these methods are often constrained to high-level edits, such as adding, removing, or replacing elements, or changing the style.", + "bbox": [ + 511, + 719, + 905, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this limitation, we utilize multimodal LLMs, such as GPT-4o [1] or Pixtral-12B [2], to directly analyze frame pairs and generate editing instructions. Specifically,", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2706", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/214dfe6241514c38fd63af0b6cfd1b1b898f82706750b591051dc5298266e883.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsReal Target?# EditsEditing Types
Non-rigidViewpointAdd/RemoveStyle
InstructPix2Pix [5]X0.3MXX✓✓
MagicBrush [42]X10KXX✓✓
EmuEdit [34]X10MXX✓✓
UltraEdit [45]X4.1MXX✓✓
Ours✓6M✓✓XX
", + "bbox": [ + 94, + 88, + 480, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Comparison with existing image-editing datasets. Our dataset is the first large-scale dataset with real target images while supporting complex editing types, such as non-rigid transformations and camera viewpoint adjustments.", + "bbox": [ + 89, + 199, + 482, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MLLMs are prompted to examine the differences between the source and target images, focusing on changes in subjects, relative positions, camera angles, and background. Based on these observations, the models generate detailed, context-specific instructions.", + "bbox": [ + 89, + 268, + 482, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To improve clarity, we instruct the MLLMs to use absolute terms (e.g., \"Move the bee to the center of the flower\" instead of \"Move the bee to the position in the target image\") and to begin instructions with an action verb (e.g., Change, Move, Adjust) for better usability in real-world applications. Additionally, to ensure high-quality instructions, we allow the MLLMs to reject frame pairs where the edits are too complex to describe accurately, and we exclude such pairs from our dataset.", + "bbox": [ + 88, + 344, + 482, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dataset Analysis", + "text_level": 1, + "bbox": [ + 89, + 489, + 254, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In total, after filtering, we sampled 6 million image pairs from videos, each with appropriate levels of motion and transformation. All frames are captioned using Pixtral-12B [2]. The edits captured in our dataset focus on realistic transformations, such as non-rigid changes of subjects/objects, rearranging elements, and altering camera perspectives—types of edits that are largely absent in existing datasets. Examples can be seen in Fig. 3, with comparisons to prior datasets presented in Table 1.", + "bbox": [ + 88, + 512, + 482, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our dataset can be combined with existing ones to train models that handle both realistic edits introduced in this paper, and the artistic edits explored in previous works (e.g., converting a photo into a painting), as shown in the supplementary material.", + "bbox": [ + 89, + 648, + 482, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, some video frames may include subtle changes that are not fully captured in the generated instructions. However, our experiments show that these small discrepancies do not significantly impact the training process, as the model focuses on learning the major transformations between frame pairs.", + "bbox": [ + 89, + 724, + 482, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 829, + 181, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using the constructed dataset, we fine-tune a pretrained T2I model for instruction-based image manipulation, enabling edits such as non-rigid transformations of subjects, object", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8f37a491d91ff19762ed23456a1a5cb3fe7319f0361dcd2f7319f5aaae16607d.jpg", + "image_caption": [ + "Figure 4. Overview of the proposed model architecture for instruction-based image editing. The source and target images are first encoded into latent representations $z^s$ and $z^e$ using a pretrained encoder. The target latent $z^e$ is then transformed into a noisy latent $z_t^e$ through the forward diffusion process. We concatenate the source image latent and the noisy target latent along the width dimension to form the model input, which is fed into the denoising U-Net $\\epsilon_{\\theta}$ to predict a noise map. The right half of the output, corresponding to the noisy target input, is cropped and compared with the original noise map." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 903, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "repositioning, and changes in viewpoint. Unlike previous methods, in Sec. 4.1 we introduce a novel spatial conditioning approach that conditions the pretrained T2I models on reference images without requiring any architectural modifications. This technique allows instructions to control the target structure while preserving the identity and content of the original image. Since our method does not alter the underlying architecture of the pretrained T2I models, it seamlessly integrates with additional control mechanisms, such as masks or ControlNets [43], to enable even more precise and localized edits (Sec. 4.2).", + "bbox": [ + 511, + 393, + 906, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Finetuning T2I Models for Image Manipulation", + "text_level": 1, + "bbox": [ + 511, + 573, + 903, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing instruction-based editing models [5, 11, 16, 34] typically condition T2I models on a reference image by concatenating it with the noise map along the channel dimension as input to the denoising network. However, this channel-based conditioning method spatially aligns the reference and target images, which limits the model's flexibility in making structural changes, such as non-rigid transformations or viewpoint adjustments.", + "bbox": [ + 511, + 595, + 905, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, we introduce Spatial Conditioning, which conditions the model by concatenating the source image with the noise map along the spatial dimension.", + "bbox": [ + 511, + 718, + 903, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a triplet $(I^s, I^e, C)$ , we first encode both images into latent representations $z^s$ and $z^e$ using a pretrained variational autoencoder [21]. The target latent $z^e$ is then transformed into a noisy latent $z_t^e$ via the forward diffusion process. We concatenate the source image latent with the noisy target latent along the width dimension to create the model input $z_t = \\mathrm{Concat}_{width}([z^s, z_t^e])$ . This input, which doubles the width of the noisy latent, is fed into the denoising U-Net $\\epsilon_{\\theta}$ to predict a noise map of the same width. We then", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2707", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "crop the right half of the output corresponding to the noisy input and compute the denoising loss:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {E d i t}} = \\mathbb {E} _ {z _ {t}, C, t, \\epsilon} \\left[ \\| \\epsilon - \\operatorname {C r o p} _ {\\text {w i d t h}} \\left(\\epsilon_ {\\theta} \\left(z _ {t}, C, t\\right)\\right) \\| ^ {2} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 132, + 482, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This spatial conditioning approach, while seemingly similar to channel conditioning, offers critical advantages for image editing tasks. First, it does not spatially align the reference image with the noisy input, allowing the denoising network to process the condition and the noisy input in parallel. This enables the network to attend to any region of the condition image at each cross-attention layer, facilitating structural changes in the target image. Moreover, since the network has access to the features of the condition image throughout all layers, it can better preserve the details and identity of the source image, leading to more accurate and realistic edits.", + "bbox": [ + 89, + 155, + 483, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Incorporating Additional Controls", + "text_level": 1, + "bbox": [ + 89, + 345, + 392, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once trained, our editing model can perform a wide range of edits based on user-provided instructions. However, text instructions alone are often insufficiently precise, making additional controls essential for accurate image editing.", + "bbox": [ + 89, + 367, + 483, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Integration with masks. Similar to zero-shot inpainting in T2I models, our approach supports mask-based localization to edit specific regions of an image. During inference, we blend the updated latent with the reference latent using a mask to control the area of modification:", + "bbox": [ + 89, + 435, + 483, + 510 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} ^ {*} = (1 - m) \\cdot z _ {t - 1} ^ {s} + m \\cdot z _ {t - 1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 522, + 482, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z_{t-1}^{s}$ is the noisy latent from the forward diffusion of the source image latent $z^{s}$ , and $m$ is the mask resized to match the latent resolution.", + "bbox": [ + 89, + 544, + 482, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Integration with other spatial controls. For more complex edits, additional visual cues beyond text can greatly enhance control. Users can provide sketches, or adjust key points of a skeleton to modify poses, for example. To enable this, we integrate controllable diffusion models like ControlNet [43] and T2I-Adapter [27]. Crucially, because our model retains the architecture of the original T2I model, we can directly use these control techniques with our trained image-editing models, incorporating spatial maps for finer control. This compatibility is not possible with previous instruction-based editing models like InstructPix2Pix.", + "bbox": [ + 89, + 589, + 483, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 768, + 223, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Sec. 5.1, we outline the experimental setup, including the evaluation benchmark and quantitative metrics. In Sec. 5.2, we compare with state-of-the-art text-guided image editing methods both quantitatively and qualitatively. Sec. 5.3 demonstrates our model's ability to incorporate additional controls for more precise editing. Finally, we assess the impact of our dataset and conditioning strategy in Sec. 5.4.", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 90, + 702, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We fine-tune Stable Diffusion V1.5 [32] with the proposed spatial conditioning strategy on our newly constructed dataset to enable various types of complex image edits. Input images have a resolution of $512 \\times 512$ . The model is trained for 100,000 iterations with a constant learning rate of $1 \\times 10^{-4}$ , using the Adam optimizer to update model parameters. Training is conducted on 8 NVIDIA A100 GPUs with a total batch size of 256. During sampling, we employ the DDIM [36] scheduler with 50 steps to generate the final edited images. Note that mask and additional spatial controls are optional; when unspecified, results are generated solely based on textual instructions.", + "bbox": [ + 511, + 122, + 906, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation benchmark. Existing instruction-based image editing benchmarks, such as MagicBrush [42] and EmuEdit [34], primarily focus on editing tasks that maintain the original image structure, such as style modification, object replacement, and color or texture changes. These benchmarks are therefore unsuitable for evaluating non-rigid, consistent image editing tasks that alter the image structure while preserving identity and appearance details. To address this gap, we created a specialized test set tailored for the non-rigid editing tasks presented. Our benchmark consists of 50 images, each paired with human-curated editing instructions that specify modifications in pose, expression, viewpoint, shape, position, and other structural transformations. The benchmark excludes any local or global style or appearance edits. Examples of input images and their corresponding instructions are shown in Fig. 5. More details are in the supplementary material.", + "bbox": [ + 511, + 313, + 906, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. For evaluation, we employ metrics to assess two main aspects of image editing: alignment with text instructions and fidelity to the source image. To evaluate instruction adherence, following previous works, we first generate captions for the target image using a large language model (LLM) based on the source image's caption and the given instructions. We then assess the agreement between caption changes and actual image changes (CLIP-D). However, this metric relies on generating captions for target images, which may not directly measure alignment with the original instructions. To address this, we propose a new metric: we prompt MLLMs to generate instructions based on the input and output images for each method, then compute the CLIP distance between these generated instructions and the original ones, denoted as CLIP-Inst. For evaluating faithfulness, we calculate the CLIP feature distance between the input and output images (CLIP-I).", + "bbox": [ + 511, + 580, + 906, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also collect human feedback from 40 participants to evaluate the quality of the edited images. Each participant is shown a subset of 20 examples, including the outputs from each method, alongside the source image and the editing", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2708", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f0a1aaeae56160944eabb84aef40d511b119435d525c628586640b7201f8ce78.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCLIP-D ↑CLIP-Inst↑CLIP-I↑
NullTextInversion [26]0.06600.76480.9063
MasaCtrl [7]0.04360.85270.9160
Imagic [19]0.05080.86450.8379
InstructPix2Pix [5]0.08870.85690.9380
MagicBrush [42]0.09720.86480.9318
UltraEdit [45]0.08240.85710.9184
Ours0.13610.87240.9275
", + "bbox": [ + 96, + 88, + 478, + 231 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5bf3599b096b29c43b447b8317eb1ea38c0fa26d4e6a87208d0a1cf8e07df6ac.jpg", + "table_caption": [ + "Table 2. Quantitative comparison with state-of-the-art text-guided image editing methods on our collected benchmark." + ], + "table_footnote": [], + "table_body": "
Imagic [19]InstructPix2Pix [5]MagicBrush [42]Ours
Preference5.0%3.25%4.13%87.62%
", + "bbox": [ + 94, + 285, + 478, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "instruction. Participants are asked to choose the image that \"best follows the editing instruction while maintaining the highest quality and preserving the original content.\"", + "bbox": [ + 89, + 388, + 483, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Baseline Comparisons", + "text_level": 1, + "bbox": [ + 89, + 446, + 297, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Methods for comparison. We compare our proposed method against state-of-the-art text-guided image editing approaches with publicly available models, including zero-shot description-based methods: NullTextInversion [26], MasaCtrl [7], and Imagic [19], as well as instruction-based methods: InstructPix2Pix [5], MagicBrush [42], and UltraEdit [45]. For description-guided methods, we generate the target description by captioning the input image and deriving the description from the instructions using an LLM.", + "bbox": [ + 88, + 477, + 482, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. Table 2 presents the quantitative evaluation of both description- and instruction-based methods on our editing benchmark. A method that perfectly preserves content but fails to follow the editing instructions is still inadequate. For instance, InstructPix2Pix and MagicBrush show higher CLIP-I scores than our approach because they struggle with non-rigid edits and viewpoint changes. As a result, they often leave the input image mostly unchanged (see Fig. 5), leading to high content preservation scores. In contrast, our method achieves the highest scores in instruction alignment while still maintaining strong content preservation. This is further validated by the user study in Table 3, where in over $87.62\\%$ of cases, users preferred our method over the other four approaches.", + "bbox": [ + 89, + 619, + 482, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results. We present a qualitative comparison of various methods in Fig. 5. Since prior approaches are primarily designed for editing tasks that preserve the original image structure, they struggle with non-rigid edits (such", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4b880a1d00ee27d1de85ad58fdd3a7e2b9141c799aa77fabe90f9cabe416ec8d.jpg", + "table_caption": [ + "Table 3. Human preference results. Percentage indicates how often each method was selected as the best result." + ], + "table_footnote": [], + "table_body": "
MethodCLIP-D↑CLIP-Inst ↑CLIP-I ↑
SC + IP2P data0.12770.84140.9094
CC + Our data0.08530.86790.8552
SC + Our data0.13610.87240.9275
", + "bbox": [ + 519, + 88, + 903, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Ablation on the training dataset and conditioning approach. Training on the InstructPix2Pix dataset [5] leads to poorer instruction alignment and content preservation. Additionally, our Spatial Conditioning (SC) method outperforms the standard Channel Conditioning (CC) approach.", + "bbox": [ + 511, + 181, + 906, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as changes in pose and expression), object repositioning, or viewpoint adjustments. These methods often fail to follow such editing instructions, either producing images identical to the source or generating edits that do not align with the specified instructions. Even in cases where the instructions are somewhat followed, they frequently fail to maintain content consistency. In contrast, our method, trained on real video frames with naturalistic transformations, is capable of handling these complex edits while preserving consistency with the input images.", + "bbox": [ + 511, + 266, + 906, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Results with Additional Controls", + "text_level": 1, + "bbox": [ + 511, + 426, + 802, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We demonstrate the incorporation of additional controls for precise editing in Fig. 6. Instructions alone can sometimes be ambiguous regarding which part of the image to edit; in such cases, a mask can be used for localized adjustments (Fig. 6 (a)). In other scenarios, instructions alone may not suffice, especially when precise control is needed for adjustments like altering a subject's pose. Users can address this by directly adjusting control points to define a skeleton pose or providing a rough sketch of the desired composition. We show that our approach can seamlessly integrate ControlNet to utilize such controls for precise editing (Fig. 6 (b)).", + "bbox": [ + 511, + 448, + 908, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Ablations", + "text_level": 1, + "bbox": [ + 511, + 622, + 624, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We argue that our proposed video + MLLM dataset construction pipeline, along with the Spatial Conditioning (SC) approach, enables our model to perform complex image edits while maintaining appearance consistency. We evaluate the impact of each component in Table 4 and Fig. 7.", + "bbox": [ + 511, + 643, + 906, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. To demonstrate the importance of using real video frame pairs for training image editing models, we fine-tune a model with the same architecture as ours on the dataset proposed in InstructPix2Pix [5], where target images are synthetically generated using Prompt-to-Prompt [14]. Compared to our model trained on InstructPix2Pix's dataset (denoted as \"SC + IP2P data\"), our method not only better follows editing instructions but also preserves the original image content more effectively.", + "bbox": [ + 511, + 727, + 908, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial Conditioning. We also fine-tune our model using our dataset, but with the Channel Conditioning (CC)", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2709", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/91cacfe3ed31cf4d5b1705a4f881f250717646e5a869be5cf993b9a7a0090971.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 88, + 205, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f440677de63a429f8cea8c1dea8f381b5201112a0b8067b9677a8c316193e0ac.jpg", + "image_caption": [ + "A bear toy wearing a sweater with arm raised" + ], + "image_footnote": [], + "bbox": [ + 220, + 90, + 330, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d903e7caf616179af1997e7dcc565caf3b1b78e0d84aeffb915d5d1430a8b141.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 90, + 442, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ffd30630ce5c97525268855b530a31448d46a038a0368b2f91b2076b937a63fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 198, + 204, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6ce94ac85d29e6938ece0e0e41475a0d6eb86f51380907f416aaf440a974897.jpg", + "image_caption": [ + "A dog sitting on a ledge looking at the camera" + ], + "image_footnote": [], + "bbox": [ + 220, + 198, + 330, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/433d728936984dafb5928bb122c8efc9e514027ff7e8fff710c8ae2f02b700d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 198, + 442, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56b40da103c56a33d1f156a4616774885d7777c3bca6917ecc46d2b60e801898.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 294, + 204, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a82ed22aace129490f3e54c84e5bd69b36248213831d4d12ecea3eed040c296.jpg", + "image_caption": [ + "A straight tower against the sky" + ], + "image_footnote": [], + "bbox": [ + 220, + 294, + 330, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7347f24ff28b4bdee459c835c1d10b724a00ec845a673927d7ab9201382a8b7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 294, + 442, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/09fc6ff167cf0b45b2dbf3b5698e4c7802a67f192e9005400fdfefdce5aa6855.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 390, + 204, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d7773ca3e36dfcc3a15632bc4b701424d1b5840cb52d3f4f94993890b277603.jpg", + "image_caption": [ + "A man meditating and giving a thumbs up" + ], + "image_footnote": [], + "bbox": [ + 220, + 390, + 330, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df977abb63050a56ddc7b28d8baf2fa1b386a1f3d7bba373604fb67cebc0a2d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 390, + 442, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e50f9ae78483a179daab2d8f9fdae63a53ed86ff0410cfad130e10e0df303fb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 486, + 202, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1977b9db7dcf1c57475a88f83e1c6c6c0e74f281f3af7a210de3e2cc185667c.jpg", + "image_caption": [ + "A woman with blonde braids smiling at camera." + ], + "image_footnote": [], + "bbox": [ + 220, + 486, + 330, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3385714f27dff57efe9bf754708ac6f3b380c3217152fd17e211f71a3c0d6741.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 486, + 442, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e4cf758af8fe081e1081f7dc5d74e8000ca2c65e74f473ec458824520d1e0fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 582, + 204, + 667 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e121415176c4350688b35f9c8c08d06402d2bcd83b80fb8dfc8de78dff7708a0.jpg", + "image_caption": [ + "A brown dog jumping in the air" + ], + "image_footnote": [], + "bbox": [ + 220, + 583, + 330, + 667 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51673d34f7e67828b59880980327f9253c3c1cc6150259c9b4fd57b5ac8bf5cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 583, + 442, + 667 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c10f52907241c2000243e3bd1a7cb92c1c3e28c1796b80988e836f989b7b5ca1.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison with state-of-the-art image editing methods, including both description-based and instruction-based approaches. Existing methods struggle with complex edits such as non-rigid transformations (e.g., changes in pose and expression), object repositioning, or viewpoint adjustments. They often either fail to follow the editing instructions or produce images with inconsistencies, such as identity shifts. In contrast, our method, trained on real video frames with naturalistic transformations, successfully handles these edits while maintaining consistency with the original input images." + ], + "image_footnote": [], + "bbox": [ + 94, + 679, + 202, + 763 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/462d286b69b1a02d6707d3b1ca48038c3c6db8af7de42f849b598ad044be3fae.jpg", + "image_caption": [ + "A home office, viewed from the left" + ], + "image_footnote": [], + "bbox": [ + 220, + 679, + 330, + 763 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ec98dc2cad1b21476df5658cb2df93ed8507037f3eb57b4f87c94997a3d55ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 679, + 442, + 763 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75b104fc8d4b1779f91273a2fe1304860e95afcce2fbcb0c2b82cdee28504780.jpg", + "image_caption": [ + "InstructPix2Pix" + ], + "image_footnote": [], + "bbox": [ + 457, + 103, + 566, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3466b4087e60390eae434a62ff7ee109cbc107aa668f807e677b2959bb68a160.jpg", + "image_caption": [ + "MagicBrush", + "Raise the bear doll's arm." + ], + "image_footnote": [], + "bbox": [ + 570, + 103, + 678, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c941c4dbe61749aa24e27487c0ca8c1c3af30df7f915b813d160517f264411e.jpg", + "image_caption": [ + "UltraEdit" + ], + "image_footnote": [], + "bbox": [ + 681, + 103, + 790, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3739f64bc9621d60f34a99cb6e83ddbe626e9163d4eac6ea21da37aeeba543b.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 792, + 103, + 901, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cbc76734170b8fbcf023e88c8887b8cc1c0970c2efae808c2062f7a3c8adddf1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 198, + 566, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb4a6988d62e144719af3bd0643400b3ade30b8cecc42f398ca8578854b60fd7.jpg", + "image_caption": [ + "Make the dog look at the camera" + ], + "image_footnote": [], + "bbox": [ + 570, + 198, + 679, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79f211ee573eda276b6daf7d78c9b7231a41f7f6743b7c196ac09c6f2f1c2645.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 198, + 790, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de8da886337501fb3feca3903078b04a37686b15fc7e029bedf2a3f15df31233.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 198, + 901, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/083bbac3e50923e2e50f6130325becfd0131ae613ffc041d733ed56fd47fe3e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 294, + 566, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/07ade9e662f26855a26cbcee48c5da85d78ea04932a8e8cdc062400e33bf1df1.jpg", + "image_caption": [ + "Make the building straight" + ], + "image_footnote": [], + "bbox": [ + 570, + 294, + 679, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf4b1f3c544180ad5995c49bc604eeb8a0d5c5be21209c8458d81a3b8e1cf02a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 294, + 790, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e33fce7a1d6d356d0a9a56a820f886ef28239ec2bae7fb7090bd139ff40933fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 294, + 901, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3f3a404b919828ebf83709039220bce10d30cd47d4e3bafd125184facb6411d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 390, + 568, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46c9da0f00d6894ead964e35dd1bb8f2eb2d41fa120b1207ca5afbede10e042d.jpg", + "image_caption": [ + "Have the man give a thumbs-up" + ], + "image_footnote": [], + "bbox": [ + 570, + 390, + 679, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/141851b06ac8d87c308211aa65a80d868a141fe97e21edef2dec16f50ea71380.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 390, + 790, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/459387dc18997e3249b56fe07a87e29d70705506d05ee84d74f231730e9f51ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 390, + 901, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e17669b5536ff5720a470d00c854d25c792943b23f12e32bff5af515826cf2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 487, + 568, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c9a47a3aed6288d61566178cfeebe8e98b69f0d04890c8187eb379ef864a954a.jpg", + "image_caption": [ + "Make the girl smile at the camera" + ], + "image_footnote": [], + "bbox": [ + 570, + 487, + 679, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a60ddb8e718980f52377fdfac40c892df93cc65ddc9a0688cd44ecf1c6c5ff71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 487, + 790, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5823349de2693976a327192d7f05ba0b3cbce849e60ac0e7591e6d16b74325f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 487, + 901, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c60eb66624b317f02e212292f8c7aa215142787c2a989991327826f018807320.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 583, + 568, + 667 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/693268b8d5cb7f23dd9169200df79263817b1d03982ec269e1c7fc6cadbc79f0.jpg", + "image_caption": [ + "Make the dog jump in the air" + ], + "image_footnote": [], + "bbox": [ + 570, + 583, + 679, + 667 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e59dc1877d9ed207f6b700ecbdcc902c5eef36749dc36899bcafc43ad336900.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 583, + 790, + 667 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/156b9f24f58bfae6178e0b120568ed2d3f541d264371ff60d266911f7f1eb00c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 583, + 901, + 667 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/583fed326c5485b24f6e57cc836dcd43adefd5114a199fa2f6c345ffdb1f30dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 679, + 568, + 763 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98eda8ae1494c9ddc2a8d40fba797c5dfd665b0c01cadc145156c1f5bb2f07b7.jpg", + "image_caption": [ + "Move the camera to the right" + ], + "image_footnote": [], + "bbox": [ + 570, + 679, + 679, + 763 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7baf980a2b2f5fde432f67c01ff754db57f1ac2c11f2d8e1b09fe7fbec16ceaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 679, + 790, + 763 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0eef6f7683b4b64d65484b00340c4b0caf24a02990c3b40020651216f09a2cff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 679, + 901, + 763 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2710", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca76addccb123eaaed823414d5bcaafc3127279df901d3dadb05cd0c42e61604.jpg", + "image_caption": [ + "Make the girl look up", + "(a) Localized image editing" + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 491, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7922f8ed3b9bc4f85d4f57638d5bc80f32cdd7605fcdfbd5f8e1a62b344a66d5.jpg", + "image_caption": [ + "Rotate the banana 45 degrees anti-clockwise. + Scribble", + "(b) Results with additional controls" + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 903, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e2f30d9c498906001d218aeaf9d72f56df715ffcbbb3c06e5a2ebe305e8f93d.jpg", + "image_caption": [ + "Figure 6. Qualitative results of our method with additional controls. (a) Our model can utilize a mask to specify which part of the image to edit, enabling localized adjustments and resolving ambiguities in the instructions. (b) When combined with ControlNet, our model can accept additional inputs, such as human poses or rough sketches, to achieve precise edits in subject poses or object positioning. This level of control is not possible with previous methods.", + "Make the woman smile", + "at the camera" + ], + "image_footnote": [], + "bbox": [ + 93, + 416, + 191, + 492 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77cec6bb74a0a99c5a0447086cb9367dda8bace4db1c667ae324afb8f71d9cc7.jpg", + "image_caption": [ + "SC + IP2P data" + ], + "image_footnote": [], + "bbox": [ + 191, + 416, + 289, + 492 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd3bfc7762341cfa7f5fb152a1ef38278047c583c10f28f1aaea05034d0e7120.jpg", + "image_caption": [ + "CC + Our data" + ], + "image_footnote": [], + "bbox": [ + 290, + 416, + 385, + 492 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f8bf1c09c1f9e9451aced58975f0516d09fca58f9093a3d0ff267064afd5fe0.jpg", + "image_caption": [ + "SC + Ours data" + ], + "image_footnote": [], + "bbox": [ + 385, + 416, + 483, + 492 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/80e8596f28541697564e9aa8b956f3e9815159d0ffcd2501ee3748628a64829c.jpg", + "image_caption": [ + "Figure 7. Qualitative ablation on the training dataset and conditioning approach. Compared to models trained on the Instruct-Pix2Pix dataset (IP2P) [5] or using the standard Channel Conditioning (CC) method, our model trained on our dataset with the Spatial Conditioning (SC) approach demonstrates superior instruction alignment and content preservation.", + "Input Image", + "Make them look to the side", + "Figure 8. Limitations. Left: Occasional unintended slight viewpoint changes. Right: Inability to accurately isolate specific objects based on instructions." + ], + "image_footnote": [], + "bbox": [ + 93, + 638, + 290, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 191, + 638, + 287, + 714 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18a83c05ecaa656cd804d83d7ce9612680548a1feecf4e264b6f441c437ea115.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 300, + 646, + 374, + 713 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f1da111493e24878a20c8be53481520429c75af2cd9cdc4672b4000645814bfe.jpg", + "image_caption": [ + "Lower the taller dog's head" + ], + "image_footnote": [], + "bbox": [ + 392, + 647, + 470, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "method, where the reference latent is concatenated with the target latent along the channel dimension rather than the spatial dimension. As shown in Table 4 and Fig. 7, the proposed Spatial Conditioning approach significantly improves the model's ability to apply complex edits while achieving higher quality and better content preservation.", + "bbox": [ + 89, + 814, + 483, + 905 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitation", + "text_level": 1, + "bbox": [ + 513, + 417, + 627, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our approach introduces a novel data construction pipeline, but it still has some limitations. The quality of the dataset is influenced by the frame filtering process and the capabilities of the MLLMs. Occasionally, MLLMs may generate inaccurate instructions or fail to capture all transformations between frames, leading our editing model to apply unintended changes, such as slight viewpoint shifts that deviate from the original instructions. The effectiveness of our model also depends on the accuracy of training instructions and the comprehension abilities of pretrained T2I models, which can affect its performance on complex edits. Examples of these limitations are shown in Fig. 8. Moreover, since the transformations captured from video frames are restricted to realistic changes, our model cannot handle artistic edits like style transfer or object replacement. This limitation can be mitigated by integrating our dataset with others, as demonstrated in the supplementary material.", + "bbox": [ + 511, + 436, + 906, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 696, + 633, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present an approach for sampling video frames and utilizing MLLMs to generate editing instructions for training instruction-based image manipulation models. Unlike existing datasets, which rely on synthetically generated target images, our method leverages supervision signals from videos and MLLMs to support complex edits, such as nonrigid transformations and viewpoint changes, while preserving content consistency. Future work could focus on refining filtering techniques, either by improving MLLMs or incorporating human-in-the-loop processes, as well as integrating video data with other datasets to further enhance image editing capabilities.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2711", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3", + "[2] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Devendra Chaplot, Jessica Chudnovsky, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliiou, Paul Jacob, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 3, 4", + "[3] Hadi Alzayer, Zhihao Xia, Xuaner Zhang, Eli Shechtman, Jia-Bin Huang, and Michael Gharbi. Magic fixup: Streamlining photo editing by watching dynamic videos. arXiv preprint arXiv:2403.13044, 2024. 2", + "[4] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2", + "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2, 3, 4, 6, 8", + "[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3", + "[7] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 1, 2, 6", + "[8] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 1, 2", + "[9] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. arXiv preprint arXiv:2408.00735, 2024. 2", + "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1", + "[11] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 2, 4" + ], + "bbox": [ + 93, + 227, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "We would like to thank Nanxuan Zhao and Yilin Wang for their valuable feedback. This research was supported in part by JSPS KAKENHI with Grant Numbers 24KK0209, 24K22318, 22H00529, and JST-Mirai Program JPMJMI23G1.", + "[12] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024. 2, 3", + "[13] Ligong Han, Song Wen, Qi Chen, Zhixing Zhang, Kunpeng Song, Mengwei Ren, Ruijiang Gao, Anastasis Stathopoulos, Xiaoxiao He, Yuxiao Chen, et al. Proxedit: Improving tuning-free real image editing with proximal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4291-4301, 2024. 2", + "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 3, 6", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2", + "[16] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024. 2, 4", + "[17] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 2", + "[18] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2", + "[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2, 6", + "[20] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 1, 2", + "[21] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4", + "[22] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 2", + "[23] Pengzhi Li, Qinxuan Huang, Yikang Ding, and Zhiheng Li. Layerdiffusion: Layered controlled image editing with diffusion models. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 2", + "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 2" + ], + "bbox": [ + 93, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2712", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 1, 2", + "[26] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 1, 2, 6", + "[27] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 2, 5", + "[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1", + "[29] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH Conference Proceedings, pages 1-11, 2023. 2", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1", + "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 3", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 5", + "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1", + "[34] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871-8879, 2024. 2, 3, 4, 5", + "[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2", + "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 5" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[38] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision, pages 402-419. Springer, 2020. 3", + "[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3", + "[40] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2", + "[41] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. Turboedit: Instant text-based image editing. arXiv preprint arXiv:2408.08332, 2024. 2", + "[42] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4, 5, 6", + "[43] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4, 5", + "[44] Shiwen Zhang, Shuai Xiao, and Weilin Huang. Forgedit: Text guided image editing via learning and forgetting. arXiv preprint arXiv:2309.10556, 2023. 2", + "[45] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. 4, 6" + ], + "bbox": [ + 516, + 92, + 903, + 598 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2713", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Instruction-based Image Manipulation by Watching How Things Move/6db53ffc-36dc-4f4f-8a58-b9c5bb9ce24b_model.json b/2025/Instruction-based Image Manipulation by Watching How Things Move/6db53ffc-36dc-4f4f-8a58-b9c5bb9ce24b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4b621bb0ec8ab36daae33af7a6dc6ac37ab075f6 --- /dev/null +++ b/2025/Instruction-based Image Manipulation by Watching How Things Move/6db53ffc-36dc-4f4f-8a58-b9c5bb9ce24b_model.json @@ -0,0 +1,3223 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.142, + 0.131, + 0.856, + 0.153 + ], + "angle": 0, + "content": "Instruction-based Image Manipulation by Watching How Things Move" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.18, + 0.773, + 0.235 + ], + "angle": 0, + "content": "Mingdeng Cao\\(^{1\\dagger}\\) Xuaner Zhang\\(^{2}\\) Yinqiang Zheng\\(^{1}\\) Zhihao Xia\\(^{2}\\) \n\\(^{1}\\)The University of Tokyo \\(^{2}\\)Adobe" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.269, + 0.361, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.374, + 0.274, + 0.382 + ], + "angle": 0, + "content": "Lower the horse's head" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.383, + 0.361, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.487, + 0.282, + 0.495 + ], + "angle": 0, + "content": "Change the view to the side" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.269, + 0.634, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.374, + 0.548, + 0.383 + ], + "angle": 0, + "content": "Make the man look angry" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.383, + 0.634, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.487, + 0.556, + 0.495 + ], + "angle": 0, + "content": "Move the camera to the left" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.27, + 0.905, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.374, + 0.808, + 0.383 + ], + "angle": 0, + "content": "Close the dog's eye" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.383, + 0.904, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.487, + 0.83, + 0.494 + ], + "angle": 0, + "content": "Have the man look at the side." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.511, + 0.907, + 0.566 + ], + "angle": 0, + "content": "Figure 1. We propose InstructMove, an instruction-based image editing model trained on frame pairs from videos with instructions generated by Multimodal LLMs. Our model excels at non-rigid editing, such as adjusting subject poses, expressions, and altering viewpoints, while maintaining content consistency. Additionally, our method supports precise, localized edits through the integration of masks, human poses, and other control mechanisms." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.594, + 0.327, + 0.61 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.63, + 0.486, + 0.873 + ], + "angle": 0, + "content": "This paper introduces a novel dataset construction pipeline that samples pairs of frames from videos and uses multimodal large language models (MLLMs) to generate editing instructions for training instruction-based image manipulation models. Video frames inherently preserve the identity of subjects and scenes, ensuring consistent content preservation during editing. Additionally, video data captures diverse, natural dynamics—such as non-rigid subject motion and complex camera movements—that are difficult to model otherwise, making it an ideal source for scalable dataset construction. Using this approach, we create a new dataset to train InstructMove, a model capable of instruction-based complex manipulations that are difficult to achieve with synthetically generated datasets. Our model demonstrates state-of-the-art performance in tasks such as adjusting subject poses, rearranging elements, and altering" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.581, + 0.885, + 0.596 + ], + "angle": 0, + "content": "camera perspectives. The project page is available here." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.63, + 0.645, + 0.645 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.658, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Recent years have seen remarkable progress in text-to-image (T2I) generation [10, 28, 31-33], with state-of-the-art methods now capable of producing high-quality and hyper-realistic images. The key to this success lies in training on vast datasets containing billions of images from the internet, paired with text descriptions generated by image captioning models [30]. In contrast, editing an existing image based on textual instructions remains challenging, with output quality still falling short and a limited range of editing types available. A major bottleneck in this area is the difficulty of obtaining a comparably large dataset of source-target-instruction triplets for training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Previous approaches [7, 8, 14, 20, 25, 26, 40] attempted to adapt pretrained T2I models without fine-tuning, using noise inversion to effectively \"regenerate\" the target image with modified textual prompts. However, these tuning-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.348, + 0.9 + ], + "angle": 0, + "content": "† Work done during an internship at Adobe." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2704" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.273 + ], + "angle": 0, + "content": "free methods are often slow and lack robustness. InstructPix2Pix [5] introduced a method to bootstrap a dataset of source images into editing triplets by generating editing instructions using a language model and creating target images using tuning-free techniques. Subsequent works [12, 16, 34, 42] have aimed to refine this process or filter the generated outputs. Nonetheless, a fundamental limitation remains: target images are synthetically generated, which constrains the model's potential due to data quality issues. Consequently, existing methods often struggle to preserve the appearance of edited subjects or to perform precise, detailed edits, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.275, + 0.483, + 0.472 + ], + "angle": 0, + "content": "In this paper, we present a novel approach for creating large-scale, instruction-based image-editing datasets using real images sampled from internet videos, capturing naturally realistic transformations from source to target. Our key insight is that video frames capture rich information on how things move — pose changes, element shifts, and camera movements — making them an ideal data source for training models for image manipulation. A similar insight was demonstrated in [3] to show that video frames are effective training data to fix coarse manual edits. We then leverage powerful multi-modal large language models (MLLMs), such as GPT-4o [1] and LLaVA [24] to generate descriptive text at or above human-level quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.639 + ], + "angle": 0, + "content": "Our data construction starts with extracting video frames that undergo realistic yet complex transformations. Next, we use MLLMs to analyze these transformations and generate precise editing instructions. Additionally, we introduce a novel spatial conditioning strategy: instead of conventional channel concatenation, we concatenate the reference image with the noise input along the spatial dimension. This approach enables the model to access the reference image through cross-attention across the network, greatly enhancing its ability to perform flexible edits while preserving the appearance of the source image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.792 + ], + "angle": 0, + "content": "By fine-tuning pretrained text-to-image models (e.g., Stable Diffusion [32]) on our constructed dataset with spatial conditioning strategy, we enable natural-language-driven image manipulations that were previously challenging. These include adjusting subject poses, rearranging elements, or altering camera perspectives, all while maintaining the integrity and details of the original image (see Fig. 1). Our approach also seamlessly integrates with mask-based guidance for local editing, and also with other control types [27, 43], allowing for versatile modifications." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.794, + 0.435, + 0.808 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.811, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- We introduce a data construction pipeline that uses video frames and MLLMs to create source-target-instruction triplets for training image manipulation models." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- We introduce a spatial conditioning strategy that conditions the reference image alongside the noise map to improve content preservation while enabling flexible edits." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.089, + 0.617, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.164, + 0.618, + 0.172 + ], + "angle": 0, + "content": "Put the toy's legs together." + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.09, + 0.714, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.164, + 0.696, + 0.172 + ], + "angle": 0, + "content": "InstructPix2Pix" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.09, + 0.808, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.164, + 0.784, + 0.172 + ], + "angle": 0, + "content": "MagicBrush" + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.09, + 0.905, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.164, + 0.874, + 0.171 + ], + "angle": 0, + "content": "UltraEdit" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.187, + 0.908, + 0.23 + ], + "angle": 0, + "content": "Figure 2. Existing methods struggle with complex edits on real photos, such as non-rigid transformations. They often either fail to follow the editing instructions or produce inconsistent outputs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.257, + 0.907, + 0.332 + ], + "angle": 0, + "content": "- Our approach enables a wider range of text-based image manipulations, including pose adjustments, element rearrangements, and perspective changes, while also supporting precise, localized edits through integration with additional control mechanisms." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.349, + 0.655, + 0.365 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.374, + 0.906, + 0.39 + ], + "angle": 0, + "content": "2.1. Zero-shot Image Editing with Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.906, + 0.518 + ], + "angle": 0, + "content": "Diffusion models [15, 35, 37] have become the leading approach in text-to-image generation due to their ability to effectively translate noise into structured data distributions. Leveraging these pretrained models, several image editing techniques have been developed by manipulating the sampling process [4, 8, 17, 20, 22, 25, 26, 29], modifying the internal architecture of the denoising network [7, 14, 40], or incorporating additional optimization steps [19, 23, 44]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.701 + ], + "angle": 0, + "content": "For example, SDEdit adds noise to the image and then generates the desired edit by conditioning on a target prompt during the reverse denoising process. Other approaches, such as Prompt-to-Prompt [14], Plug-and-Play [40], and MasaCtrl [7], manipulate the attention mechanism after inverting the image to noise using DDIM inversion [36], allowing for edits that align with the desired prompt. More recent works [9, 13, 18, 26, 41] have refined the inversion process to enhance editing robustness and accuracy. While these zero-shot methods offer flexibility in editing images, they are often too slow and lack robustness, making them less suitable for practical applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.88, + 0.728 + ], + "angle": 0, + "content": "2.2. Training Instruction-based Editing Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Comparing to modifying target image descriptions, a more intuitive and user-friendly editing method is to use direct instructions (e.g., “Have the man look at the camera”) to guide the editing process. InstructPix2Pix [5] pioneered this concept by fine-tuning Stable Diffusion [32] with editing instructions. To handle more complex instructions, methods like MGIE [11] and SmartEdit [16] integrate Multimodal LLMs to improve the understanding of editing instructions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "A major challenge in training instruction-based editing models is constructing paired image data with corresponding instructions. InstructPix2Pix tackles this by using lan" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2705" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.092, + 0.307, + 0.1 + ], + "angle": 0, + "content": "(a) Frame selection: Sample frame pairs from internet videos." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.119, + 0.168, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.175, + 0.102, + 0.32, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.1, + 0.409, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.1, + 0.48, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.162, + 0.191, + 0.17 + ], + "angle": 0, + "content": "(b) Generate instructions:" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.177, + 0.199, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.179, + 0.464, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.212, + 0.335, + 0.226 + ], + "angle": 0, + "content": "Adjust the woman's gaze from the book to looking directly at the camera." + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.212, + 0.468, + 0.221 + ], + "angle": 0, + "content": "Have the man give a thumbs up gesture." + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.234, + 0.31, + 0.242 + ], + "angle": 0, + "content": "(c) Generated source-target-instruction triplets for training:" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.242, + 0.222, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.291, + 0.209, + 0.297 + ], + "angle": 0, + "content": "\"Remove the smile from the woman's face.\"" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.242, + 0.35, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.291, + 0.347, + 0.297 + ], + "angle": 0, + "content": "\"Rotate the bird to the left and make its tail visible.\"" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.242, + 0.478, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.291, + 0.469, + 0.297 + ], + "angle": 0, + "content": "\"Make the person looking at the laptop screen.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.305, + 0.484, + 0.389 + ], + "angle": 0, + "content": "Figure 3. Our data construction pipeline. (a) We begin by sampling suitable frame pairs from videos, ensuring realistic and moderate transformations. (b) These frame pairs are used to prompt Multimodal Large Language Models (MLLMs) to generate detailed editing instructions. (c) This process results in a large-scale dataset with realistic image pairs and precise editing instructions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.409, + 0.483, + 0.558 + ], + "angle": 0, + "content": "guage models, such as GPT-3 [6], to generate both an edited image caption and an editing instruction. These captions are converted into images using Prompt-to-Prompt [14]. MagicBrush [42] constructs a small dataset by performing edits using DALL-E 2 [31]. InstructDiffusion [12] augments datasets with synthetic pairs for tasks like object removal and replacement. EmuEdit [34] improves data pair construction using Llama 2 [39] and local masks with Prompt-to-Prompt, combining these with other vision datasets to build a large-scale dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.484, + 0.756 + ], + "angle": 0, + "content": "However, in existing large-scale editing datasets, most target images are generated using synthetic methods, which often introduces artifacts and significant appearance deviations from the source images. These datasets are typically limited to high-level edits such as object overlay and style transfer. As a result, models trained on these datasets struggle with complex edits on real photos, such as non-rigid transformations and viewpoint changes. In contrast, our approach leverages real video frames as source and target images, using Multimodal LLMs to generate instructions directly from these frames. This allows for the construction of a more realistic dataset, capturing complex transformations and overcoming the limitations of previous methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.294, + 0.785 + ], + "angle": 0, + "content": "3. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Unlike existing instruction-based editing datasets, which rely on tuning-free methods to generate target images, we propose a novel data construction pipeline that leverages real video frame pairs. By extracting frame pairs that capture meaningful transformations and using Multimodal Large Language Models (MLLMs) to generate detailed editing instructions, we create a more realistic and high-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "quality editing dataset. In Sec. 3.1, we outline our approach for extracting suitable frame pairs from videos, followed by Sec. 3.2, where we describe how we use MLLMs to generate precise editing instructions. Finally, Sec. 3.3 provides an in-depth analysis of the resulting dataset. An overview of the data construction pipeline is illustrated in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.192, + 0.821, + 0.208 + ], + "angle": 0, + "content": "3.1. Sampling Frame Pairs from Videos" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.907, + 0.35 + ], + "angle": 0, + "content": "In a video, the same subjects and elements can appear in different ways, move naturally, or be shown from different angles. For example, a person might change their posture, expression, or movement over time. These changes provide useful clues about how a scene or perspective might look with small adjustments, making videos a great source for picking frame pairs to use in image editing tasks. However, it's important to sample the frame pairs with proper criteria to make sure they are suitable for training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.906, + 0.463 + ], + "angle": 0, + "content": "Initial sampling of paired frames. We begin by captioning each video to generate an overall description. Videos containing specific keywords (e.g., landscape, abstract, still) that are not suitable for image editing are filtered out. We then sample two frames \\((I^s, I^e)\\) with a fixed time interval (i.e., 3-5 seconds) to ensure sufficient transformations occur in the objects of interest." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.907, + 0.688 + ], + "angle": 0, + "content": "Motion-based filtering. To refine the selection, we apply motion filtering to exclude frames with either too little or too much movement. Specifically, we use RAFT [38] to compute the optical flow \\( O_{I^s \\to I^e} \\) from \\( I^s \\) to \\( I^e \\) and derive the flow magnitude \\( M = \\sqrt{O_x^2 + O_y^2} \\). Frame pairs with moderate movement are retained, while those with excessive or minimal motion are discarded. To further ensure consistency, we assess background changes by computing a background occlusion mask through backward warping of \\( I^e \\) to \\( I^s \\) using the estimated flow. By measuring the occlusion mask ratio, we filter out frames with significant background changes (i.e., large masked regions). This motion filtering pipeline allows us to obtain image pairs with realistic, moderate transformations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.836, + 0.712 + ], + "angle": 0, + "content": "3.2. Instruction Generation with MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Next, we describe how we generate accurate editing instructions \\( C \\) given frame pairs \\( (I^s, I^e) \\). Previous approaches [5, 34] rely solely on text-based methods, using Large Language Models (LLMs) to generate instructions and captions of target images purely from the source caption. This approach significantly limits the variety of edits that can be captured in the dataset. As a result, these methods are often constrained to high-level edits, such as adding, removing, or replacing elements, or changing the style." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To address this limitation, we utilize multimodal LLMs, such as GPT-4o [1] or Pixtral-12B [2], to directly analyze frame pairs and generate editing instructions. Specifically," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2706" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.19 + ], + "angle": 0, + "content": "
DatasetsReal Target?# EditsEditing Types
Non-rigidViewpointAdd/RemoveStyle
InstructPix2Pix [5]X0.3MXX✓✓
MagicBrush [42]X10KXX✓✓
EmuEdit [34]X10MXX✓✓
UltraEdit [45]X4.1MXX✓✓
Ours✓6M✓✓XX
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.2, + 0.483, + 0.256 + ], + "angle": 0, + "content": "Table 1. Comparison with existing image-editing datasets. Our dataset is the first large-scale dataset with real target images while supporting complex editing types, such as non-rigid transformations and camera viewpoint adjustments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.269, + 0.483, + 0.344 + ], + "angle": 0, + "content": "MLLMs are prompted to examine the differences between the source and target images, focusing on changes in subjects, relative positions, camera angles, and background. Based on these observations, the models generate detailed, context-specific instructions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.345, + 0.483, + 0.481 + ], + "angle": 0, + "content": "To improve clarity, we instruct the MLLMs to use absolute terms (e.g., \"Move the bee to the center of the flower\" instead of \"Move the bee to the position in the target image\") and to begin instructions with an action verb (e.g., Change, Move, Adjust) for better usability in real-world applications. Additionally, to ensure high-quality instructions, we allow the MLLMs to reject frame pairs where the edits are too complex to describe accurately, and we exclude such pairs from our dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.491, + 0.255, + 0.507 + ], + "angle": 0, + "content": "3.3. Dataset Analysis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.513, + 0.483, + 0.649 + ], + "angle": 0, + "content": "In total, after filtering, we sampled 6 million image pairs from videos, each with appropriate levels of motion and transformation. All frames are captioned using Pixtral-12B [2]. The edits captured in our dataset focus on realistic transformations, such as non-rigid changes of subjects/objects, rearranging elements, and altering camera perspectives—types of edits that are largely absent in existing datasets. Examples can be seen in Fig. 3, with comparisons to prior datasets presented in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.483, + 0.724 + ], + "angle": 0, + "content": "Our dataset can be combined with existing ones to train models that handle both realistic edits introduced in this paper, and the artistic edits explored in previous works (e.g., converting a photo into a painting), as shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.725, + 0.483, + 0.815 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, some video frames may include subtle changes that are not fully captured in the generated instructions. However, our experiments show that these small discrepancies do not significantly impact the training process, as the model focuses on learning the major transformations between frame pairs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.182, + 0.845 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Using the constructed dataset, we fine-tune a pretrained T2I model for instruction-based image manipulation, enabling edits such as non-rigid transformations of subjects, object" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.905, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.241, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 4. Overview of the proposed model architecture for instruction-based image editing. The source and target images are first encoded into latent representations \\( z^s \\) and \\( z^e \\) using a pretrained encoder. The target latent \\( z^e \\) is then transformed into a noisy latent \\( z_t^e \\) through the forward diffusion process. We concatenate the source image latent and the noisy target latent along the width dimension to form the model input, which is fed into the denoising U-Net \\( \\epsilon_{\\theta} \\) to predict a noise map. The right half of the output, corresponding to the noisy target input, is cropped and compared with the original noise map." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.561 + ], + "angle": 0, + "content": "repositioning, and changes in viewpoint. Unlike previous methods, in Sec. 4.1 we introduce a novel spatial conditioning approach that conditions the pretrained T2I models on reference images without requiring any architectural modifications. This technique allows instructions to control the target structure while preserving the identity and content of the original image. Since our method does not alter the underlying architecture of the pretrained T2I models, it seamlessly integrates with additional control mechanisms, such as masks or ControlNets [43], to enable even more precise and localized edits (Sec. 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.905, + 0.59 + ], + "angle": 0, + "content": "4.1. Finetuning T2I Models for Image Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.906, + 0.718 + ], + "angle": 0, + "content": "Existing instruction-based editing models [5, 11, 16, 34] typically condition T2I models on a reference image by concatenating it with the noise map along the channel dimension as input to the denoising network. However, this channel-based conditioning method spatially aligns the reference and target images, which limits the model's flexibility in making structural changes, such as non-rigid transformations or viewpoint adjustments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.905, + 0.764 + ], + "angle": 0, + "content": "To address this, we introduce Spatial Conditioning, which conditions the model by concatenating the source image with the noise map along the spatial dimension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Given a triplet \\((I^s, I^e, C)\\), we first encode both images into latent representations \\(z^s\\) and \\(z^e\\) using a pretrained variational autoencoder [21]. The target latent \\(z^e\\) is then transformed into a noisy latent \\(z_t^e\\) via the forward diffusion process. We concatenate the source image latent with the noisy target latent along the width dimension to create the model input \\(z_t = \\mathrm{Concat}_{width}([z^s, z_t^e])\\). This input, which doubles the width of the noisy latent, is fed into the denoising U-Net \\(\\epsilon_{\\theta}\\) to predict a noise map of the same width. We then" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2707" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "crop the right half of the output corresponding to the noisy input and compute the denoising loss:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.133, + 0.483, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {E d i t}} = \\mathbb {E} _ {z _ {t}, C, t, \\epsilon} \\left[ \\| \\epsilon - \\operatorname {C r o p} _ {\\text {w i d t h}} \\left(\\epsilon_ {\\theta} \\left(z _ {t}, C, t\\right)\\right) \\| ^ {2} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.156, + 0.484, + 0.336 + ], + "angle": 0, + "content": "This spatial conditioning approach, while seemingly similar to channel conditioning, offers critical advantages for image editing tasks. First, it does not spatially align the reference image with the noisy input, allowing the denoising network to process the condition and the noisy input in parallel. This enables the network to attend to any region of the condition image at each cross-attention layer, facilitating structural changes in the target image. Moreover, since the network has access to the features of the condition image throughout all layers, it can better preserve the details and identity of the source image, leading to more accurate and realistic edits." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.346, + 0.393, + 0.363 + ], + "angle": 0, + "content": "4.2. Incorporating Additional Controls" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.484, + 0.43 + ], + "angle": 0, + "content": "Once trained, our editing model can perform a wide range of edits based on user-provided instructions. However, text instructions alone are often insufficiently precise, making additional controls essential for accurate image editing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.511 + ], + "angle": 0, + "content": "Integration with masks. Similar to zero-shot inpainting in T2I models, our approach supports mask-based localization to edit specific regions of an image. During inference, we blend the updated latent with the reference latent using a mask to control the area of modification:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.523, + 0.483, + 0.541 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} ^ {*} = (1 - m) \\cdot z _ {t - 1} ^ {s} + m \\cdot z _ {t - 1}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.483, + 0.589 + ], + "angle": 0, + "content": "where \\( z_{t-1}^{s} \\) is the noisy latent from the forward diffusion of the source image latent \\( z^{s} \\), and \\( m \\) is the mask resized to match the latent resolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.756 + ], + "angle": 0, + "content": "Integration with other spatial controls. For more complex edits, additional visual cues beyond text can greatly enhance control. Users can provide sketches, or adjust key points of a skeleton to modify poses, for example. To enable this, we integrate controllable diffusion models like ControlNet [43] and T2I-Adapter [27]. Crucially, because our model retains the architecture of the original T2I model, we can directly use these control techniques with our trained image-editing models, incorporating spatial maps for finer control. This compatibility is not possible with previous instruction-based editing models like InstructPix2Pix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.224, + 0.788 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In Sec. 5.1, we outline the experimental setup, including the evaluation benchmark and quantitative metrics. In Sec. 5.2, we compare with state-of-the-art text-guided image editing methods both quantitatively and qualitatively. Sec. 5.3 demonstrates our model's ability to incorporate additional controls for more precise editing. Finally, we assess the impact of our dataset and conditioning strategy in Sec. 5.4." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.704, + 0.108 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.907, + 0.304 + ], + "angle": 0, + "content": "Implementation Details. We fine-tune Stable Diffusion V1.5 [32] with the proposed spatial conditioning strategy on our newly constructed dataset to enable various types of complex image edits. Input images have a resolution of \\(512 \\times 512\\). The model is trained for 100,000 iterations with a constant learning rate of \\(1 \\times 10^{-4}\\), using the Adam optimizer to update model parameters. Training is conducted on 8 NVIDIA A100 GPUs with a total batch size of 256. During sampling, we employ the DDIM [36] scheduler with 50 steps to generate the final edited images. Note that mask and additional spatial controls are optional; when unspecified, results are generated solely based on textual instructions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.908, + 0.572 + ], + "angle": 0, + "content": "Evaluation benchmark. Existing instruction-based image editing benchmarks, such as MagicBrush [42] and EmuEdit [34], primarily focus on editing tasks that maintain the original image structure, such as style modification, object replacement, and color or texture changes. These benchmarks are therefore unsuitable for evaluating non-rigid, consistent image editing tasks that alter the image structure while preserving identity and appearance details. To address this gap, we created a specialized test set tailored for the non-rigid editing tasks presented. Our benchmark consists of 50 images, each paired with human-curated editing instructions that specify modifications in pose, expression, viewpoint, shape, position, and other structural transformations. The benchmark excludes any local or global style or appearance edits. Examples of input images and their corresponding instructions are shown in Fig. 5. More details are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Evaluation metrics. For evaluation, we employ metrics to assess two main aspects of image editing: alignment with text instructions and fidelity to the source image. To evaluate instruction adherence, following previous works, we first generate captions for the target image using a large language model (LLM) based on the source image's caption and the given instructions. We then assess the agreement between caption changes and actual image changes (CLIP-D). However, this metric relies on generating captions for target images, which may not directly measure alignment with the original instructions. To address this, we propose a new metric: we prompt MLLMs to generate instructions based on the input and output images for each method, then compute the CLIP distance between these generated instructions and the original ones, denoted as CLIP-Inst. For evaluating faithfulness, we calculate the CLIP feature distance between the input and output images (CLIP-I)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We also collect human feedback from 40 participants to evaluate the quality of the edited images. Each participant is shown a subset of 20 examples, including the outputs from each method, alongside the source image and the editing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2708" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.232 + ], + "angle": 0, + "content": "
MethodCLIP-D ↑CLIP-Inst↑CLIP-I↑
NullTextInversion [26]0.06600.76480.9063
MasaCtrl [7]0.04360.85270.9160
Imagic [19]0.05080.86450.8379
InstructPix2Pix [5]0.08870.85690.9380
MagicBrush [42]0.09720.86480.9318
UltraEdit [45]0.08240.85710.9184
Ours0.13610.87240.9275
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.242, + 0.483, + 0.271 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison with state-of-the-art text-guided image editing methods on our collected benchmark." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.286, + 0.48, + 0.324 + ], + "angle": 0, + "content": "
Imagic [19]InstructPix2Pix [5]MagicBrush [42]Ours
Preference5.0%3.25%4.13%87.62%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.333, + 0.483, + 0.361 + ], + "angle": 0, + "content": "Table 3. Human preference results. Percentage indicates how often each method was selected as the best result." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.39, + 0.484, + 0.437 + ], + "angle": 0, + "content": "instruction. Participants are asked to choose the image that \"best follows the editing instruction while maintaining the highest quality and preserving the original content.\"" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.447, + 0.299, + 0.464 + ], + "angle": 0, + "content": "5.2. Baseline Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.483, + 0.614 + ], + "angle": 0, + "content": "Methods for comparison. We compare our proposed method against state-of-the-art text-guided image editing approaches with publicly available models, including zero-shot description-based methods: NullTextInversion [26], MasaCtrl [7], and Imagic [19], as well as instruction-based methods: InstructPix2Pix [5], MagicBrush [42], and UltraEdit [45]. For description-guided methods, we generate the target description by captioning the input image and deriving the description from the instructions using an LLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.833 + ], + "angle": 0, + "content": "Quantitative results. Table 2 presents the quantitative evaluation of both description- and instruction-based methods on our editing benchmark. A method that perfectly preserves content but fails to follow the editing instructions is still inadequate. For instance, InstructPix2Pix and MagicBrush show higher CLIP-I scores than our approach because they struggle with non-rigid edits and viewpoint changes. As a result, they often leave the input image mostly unchanged (see Fig. 5), leading to high content preservation scores. In contrast, our method achieves the highest scores in instruction alignment while still maintaining strong content preservation. This is further validated by the user study in Table 3, where in over \\(87.62\\%\\) of cases, users preferred our method over the other four approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Qualitative results. We present a qualitative comparison of various methods in Fig. 5. Since prior approaches are primarily designed for editing tasks that preserve the original image structure, they struggle with non-rigid edits (such" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.089, + 0.905, + 0.173 + ], + "angle": 0, + "content": "
MethodCLIP-D↑CLIP-Inst ↑CLIP-I ↑
SC + IP2P data0.12770.84140.9094
CC + Our data0.08530.86790.8552
SC + Our data0.13610.87240.9275
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.183, + 0.907, + 0.253 + ], + "angle": 0, + "content": "Table 4. Ablation on the training dataset and conditioning approach. Training on the InstructPix2Pix dataset [5] leads to poorer instruction alignment and content preservation. Additionally, our Spatial Conditioning (SC) method outperforms the standard Channel Conditioning (CC) approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.907, + 0.419 + ], + "angle": 0, + "content": "as changes in pose and expression), object repositioning, or viewpoint adjustments. These methods often fail to follow such editing instructions, either producing images identical to the source or generating edits that do not align with the specified instructions. Even in cases where the instructions are somewhat followed, they frequently fail to maintain content consistency. In contrast, our method, trained on real video frames with naturalistic transformations, is capable of handling these complex edits while preserving consistency with the input images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.427, + 0.803, + 0.442 + ], + "angle": 0, + "content": "5.3. Results with Additional Controls" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.909, + 0.616 + ], + "angle": 0, + "content": "We demonstrate the incorporation of additional controls for precise editing in Fig. 6. Instructions alone can sometimes be ambiguous regarding which part of the image to edit; in such cases, a mask can be used for localized adjustments (Fig. 6 (a)). In other scenarios, instructions alone may not suffice, especially when precise control is needed for adjustments like altering a subject's pose. Users can address this by directly adjusting control points to define a skeleton pose or providing a rough sketch of the desired composition. We show that our approach can seamlessly integrate ControlNet to utilize such controls for precise editing (Fig. 6 (b))." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.625, + 0.637 + ], + "angle": 0, + "content": "5.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.721 + ], + "angle": 0, + "content": "We argue that our proposed video + MLLM dataset construction pipeline, along with the Spatial Conditioning (SC) approach, enables our model to perform complex image edits while maintaining appearance consistency. We evaluate the impact of each component in Table 4 and Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.909, + 0.865 + ], + "angle": 0, + "content": "Dataset. To demonstrate the importance of using real video frame pairs for training image editing models, we fine-tune a model with the same architecture as ours on the dataset proposed in InstructPix2Pix [5], where target images are synthetically generated using Prompt-to-Prompt [14]. Compared to our model trained on InstructPix2Pix's dataset (denoted as \"SC + IP2P data\"), our method not only better follows editing instructions but also preserves the original image content more effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Spatial Conditioning. We also fine-tune our model using our dataset, but with the Channel Conditioning (CC)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2709" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.207, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.092, + 0.331, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.092, + 0.444, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.188, + 0.438, + 0.197 + ], + "angle": 0, + "content": "A bear toy wearing a sweater with arm raised" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.199, + 0.205, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.199, + 0.331, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.199, + 0.444, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.284, + 0.437, + 0.293 + ], + "angle": 0, + "content": "A dog sitting on a ledge looking at the camera" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.295, + 0.205, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.295, + 0.331, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.295, + 0.444, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.381, + 0.408, + 0.39 + ], + "angle": 0, + "content": "A straight tower against the sky" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.391, + 0.205, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.391, + 0.331, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.391, + 0.443, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.477, + 0.426, + 0.486 + ], + "angle": 0, + "content": "A man meditating and giving a thumbs up" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.487, + 0.204, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.487, + 0.331, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.487, + 0.443, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.573, + 0.438, + 0.582 + ], + "angle": 0, + "content": "A woman with blonde braids smiling at camera." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.583, + 0.205, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.584, + 0.331, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.584, + 0.443, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.669, + 0.402, + 0.678 + ], + "angle": 0, + "content": "A brown dog jumping in the air" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.68, + 0.204, + 0.764 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.68, + 0.331, + 0.764 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.68, + 0.443, + 0.764 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.767, + 0.416, + 0.776 + ], + "angle": 0, + "content": "A home office, viewed from the left" + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.092, + 0.553, + 0.102 + ], + "angle": 0, + "content": "InstructPix2Pix" + }, + { + "type": "image", + "bbox": [ + 0.459, + 0.104, + 0.568, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.093, + 0.655, + 0.102 + ], + "angle": 0, + "content": "MagicBrush" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.104, + 0.679, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.093, + 0.759, + 0.102 + ], + "angle": 0, + "content": "UltraEdit" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.104, + 0.791, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.835, + 0.093, + 0.86, + 0.102 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.104, + 0.902, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.188, + 0.739, + 0.196 + ], + "angle": 0, + "content": "Raise the bear doll's arm." + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.199, + 0.568, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.199, + 0.68, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.199, + 0.791, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.199, + 0.902, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.284, + 0.756, + 0.293 + ], + "angle": 0, + "content": "Make the dog look at the camera" + }, + { + "type": "image", + "bbox": [ + 0.459, + 0.295, + 0.568, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.295, + 0.68, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.295, + 0.791, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.295, + 0.902, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.38, + 0.742, + 0.39 + ], + "angle": 0, + "content": "Make the building straight" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.391, + 0.569, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.391, + 0.68, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.391, + 0.791, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.391, + 0.902, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.477, + 0.752, + 0.486 + ], + "angle": 0, + "content": "Have the man give a thumbs-up" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.488, + 0.569, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.488, + 0.68, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.488, + 0.791, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.488, + 0.902, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.573, + 0.759, + 0.582 + ], + "angle": 0, + "content": "Make the girl smile at the camera" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.584, + 0.569, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.584, + 0.68, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.584, + 0.791, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.584, + 0.902, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.669, + 0.748, + 0.678 + ], + "angle": 0, + "content": "Make the dog jump in the air" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.68, + 0.569, + 0.764 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.68, + 0.68, + 0.764 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.68, + 0.791, + 0.764 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.68, + 0.902, + 0.764 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.766, + 0.747, + 0.775 + ], + "angle": 0, + "content": "Move the camera to the right" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.793, + 0.907, + 0.864 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison with state-of-the-art image editing methods, including both description-based and instruction-based approaches. Existing methods struggle with complex edits such as non-rigid transformations (e.g., changes in pose and expression), object repositioning, or viewpoint adjustments. They often either fail to follow the editing instructions or produce images with inconsistencies, such as identity shifts. In contrast, our method, trained on real video frames with naturalistic transformations, successfully handles these edits while maintaining consistency with the original input images." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2710" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.493, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.314, + 0.404, + 0.323 + ], + "angle": 0, + "content": "Make the girl look up" + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.328, + 0.353, + 0.338 + ], + "angle": 0, + "content": "(a) Localized image editing" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.091, + 0.905, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.315, + 0.889, + 0.324 + ], + "angle": 0, + "content": "Rotate the banana 45 degrees anti-clockwise. + Scribble" + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.328, + 0.782, + 0.337 + ], + "angle": 0, + "content": "(b) Results with additional controls" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.345, + 0.907, + 0.4 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of our method with additional controls. (a) Our model can utilize a mask to specify which part of the image to edit, enabling localized adjustments and resolving ambiguities in the instructions. (b) When combined with ControlNet, our model can accept additional inputs, such as human poses or rough sketches, to achieve precise edits in subject poses or object positioning. This level of control is not possible with previous methods." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.417, + 0.192, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.494, + 0.183, + 0.501 + ], + "angle": 0, + "content": "Make the woman smile" + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.501, + 0.169, + 0.508 + ], + "angle": 0, + "content": "at the camera" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.417, + 0.29, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.494, + 0.269, + 0.501 + ], + "angle": 0, + "content": "SC + IP2P data" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.417, + 0.387, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.494, + 0.365, + 0.501 + ], + "angle": 0, + "content": "CC + Our data" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.417, + 0.484, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.494, + 0.463, + 0.501 + ], + "angle": 0, + "content": "SC + Ours data" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.611 + ], + "angle": 0, + "content": "Figure 7. Qualitative ablation on the training dataset and conditioning approach. Compared to models trained on the Instruct-Pix2Pix dataset (IP2P) [5] or using the standard Channel Conditioning (CC) method, our model trained on our dataset with the Spatial Conditioning (SC) approach demonstrates superior instruction alignment and content preservation." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.64, + 0.291, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.717, + 0.164, + 0.724 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.64, + 0.289, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.201, + 0.716, + 0.279, + 0.723 + ], + "angle": 0, + "content": "Make them look to the side" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.647, + 0.375, + 0.714 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.716, + 0.359, + 0.724 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.648, + 0.472, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.716, + 0.473, + 0.724 + ], + "angle": 0, + "content": "Lower the taller dog's head" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.739, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Figure 8. Limitations. Left: Occasional unintended slight viewpoint changes. Right: Inability to accurately isolate specific objects based on instructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.484, + 0.906 + ], + "angle": 0, + "content": "method, where the reference latent is concatenated with the target latent along the channel dimension rather than the spatial dimension. As shown in Table 4 and Fig. 7, the proposed Spatial Conditioning approach significantly improves the model's ability to apply complex edits while achieving higher quality and better content preservation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.419, + 0.628, + 0.433 + ], + "angle": 0, + "content": "6. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.907, + 0.694 + ], + "angle": 0, + "content": "Our approach introduces a novel data construction pipeline, but it still has some limitations. The quality of the dataset is influenced by the frame filtering process and the capabilities of the MLLMs. Occasionally, MLLMs may generate inaccurate instructions or fail to capture all transformations between frames, leading our editing model to apply unintended changes, such as slight viewpoint shifts that deviate from the original instructions. The effectiveness of our model also depends on the accuracy of training instructions and the comprehension abilities of pretrained T2I models, which can affect its performance on complex edits. Examples of these limitations are shown in Fig. 8. Moreover, since the transformations captured from video frames are restricted to realistic changes, our model cannot handle artistic edits like style transfer or object replacement. This limitation can be mitigated by integrating our dataset with others, as demonstrated in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.698, + 0.634, + 0.713 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We present an approach for sampling video frames and utilizing MLLMs to generate editing instructions for training instruction-based image manipulation models. Unlike existing datasets, which rely on synthetically generated target images, our method leverages supervision signals from videos and MLLMs to support complex edits, such as nonrigid transformations and viewpoint changes, while preserving content consistency. Future work could focus on refining filtering techniques, either by improving MLLMs or incorporating human-in-the-loop processes, as well as integrating video data with other datasets to further enhance image editing capabilities." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2711" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.19 + ], + "angle": 0, + "content": "We would like to thank Nanxuan Zhao and Yilin Wang for their valuable feedback. This research was supported in part by JSPS KAKENHI with Grant Numbers 24KK0209, 24K22318, 22H00529, and JST-Mirai Program JPMJMI23G1." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.484, + 0.298 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.299, + 0.484, + 0.367 + ], + "angle": 0, + "content": "[2] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Devendra Chaplot, Jessica Chudnovsky, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliiou, Paul Jacob, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.369, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[3] Hadi Alzayer, Zhihao Xia, Xuaner Zhang, Eli Shechtman, Jia-Bin Huang, and Michael Gharbi. Magic fixup: Streamlining photo editing by watching dynamic videos. arXiv preprint arXiv:2403.13044, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.425, + 0.484, + 0.507 + ], + "angle": 0, + "content": "[4] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.484, + 0.577 + ], + "angle": 0, + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.484, + 0.689 + ], + "angle": 0, + "content": "[7] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[8] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[9] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. arXiv preprint arXiv:2408.00735, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.175 + ], + "angle": 0, + "content": "[12] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[13] Ligong Han, Song Wen, Qi Chen, Zhixing Zhang, Kunpeng Song, Mengwei Ren, Ruijiang Gao, Anastasis Stathopoulos, Xiaoxiao He, Yuxiao Chen, et al. Proxedit: Improving tuning-free real image editing with proximal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4291-4301, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.262, + 0.906, + 0.317 + ], + "angle": 0, + "content": "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.319, + 0.906, + 0.36 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.362, + 0.906, + 0.458 + ], + "angle": 0, + "content": "[16] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.461, + 0.906, + 0.529 + ], + "angle": 0, + "content": "[17] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.906, + 0.585 + ], + "angle": 0, + "content": "[18] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.588, + 0.906, + 0.658 + ], + "angle": 0, + "content": "[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.66, + 0.906, + 0.728 + ], + "angle": 0, + "content": "[20] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[21] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[22] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[23] Pengzhi Li, Qinxuan Huang, Yikang Ding, and Zhiheng Li. Layerdiffusion: Layered controlled image editing with diffusion models. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2712" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[26] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[27] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[29] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH Conference Proceedings, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.621 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[34] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871-8879, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[37] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[38] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision, pages 402-419. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[40] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[41] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. Turboedit: Instant text-based image editing. arXiv preprint arXiv:2408.08332, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[42] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[43] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[44] Shiwen Zhang, Shuai Xiao, and Weilin Huang. Forgedit: Text guided image editing via learning and forgetting. arXiv preprint arXiv:2309.10556, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[45] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2713" + } + ] +] \ No newline at end of file diff --git a/2025/Instruction-based Image Manipulation by Watching How Things Move/6db53ffc-36dc-4f4f-8a58-b9c5bb9ce24b_origin.pdf b/2025/Instruction-based Image Manipulation by Watching How Things Move/6db53ffc-36dc-4f4f-8a58-b9c5bb9ce24b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bb3f80af21e8383a209834306fcb7c2d012af3fa --- /dev/null +++ b/2025/Instruction-based Image Manipulation by Watching How Things Move/6db53ffc-36dc-4f4f-8a58-b9c5bb9ce24b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a14b68126387c608912fa40b8357e4b2a5eb9ac02b6afed679a6894fe153842 +size 8097962 diff --git a/2025/Instruction-based Image Manipulation by Watching How Things Move/full.md b/2025/Instruction-based Image Manipulation by Watching How Things Move/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5e38ba1ef10805b0d94976bbaec45e647d40ca55 --- /dev/null +++ b/2025/Instruction-based Image Manipulation by Watching How Things Move/full.md @@ -0,0 +1,447 @@ +# Instruction-based Image Manipulation by Watching How Things Move + +Mingdeng Cao $^{1\dagger}$ Xuaner Zhang $^{2}$ Yinqiang Zheng $^{1}$ Zhihao Xia $^{2}$ $^{1}$ The University of Tokyo $^{2}$ Adobe + +![](images/d25058744bf9235de49aceef260d561d3ab706146174d6e4fdb099546b211f0b.jpg) +Lower the horse's head + +![](images/5a14f0a26670d007471a58159de2ed31c450ced8c865ffe825d81222bba32aa8.jpg) +Change the view to the side +Figure 1. We propose InstructMove, an instruction-based image editing model trained on frame pairs from videos with instructions generated by Multimodal LLMs. Our model excels at non-rigid editing, such as adjusting subject poses, expressions, and altering viewpoints, while maintaining content consistency. Additionally, our method supports precise, localized edits through the integration of masks, human poses, and other control mechanisms. + +![](images/17f7a0c35cb4815bc62ff01c8a3c211d75d86329308db058b578c5e8da90ac7d.jpg) +Make the man look angry + +![](images/15b0aa691a8b3eb16dd0081078756242750a2239fec0d998aa817a8478e3e9d9.jpg) +Move the camera to the left + +![](images/d5fc7c1e13b1318b8f3033055211036a43f71df900efcf963af1a150e8db68d7.jpg) +Close the dog's eye + +![](images/c153208fc770eae9915e8b12dfe87edef1bc4e8d1392b4613053c479e14a449c.jpg) +Have the man look at the side. + +# Abstract + +This paper introduces a novel dataset construction pipeline that samples pairs of frames from videos and uses multimodal large language models (MLLMs) to generate editing instructions for training instruction-based image manipulation models. Video frames inherently preserve the identity of subjects and scenes, ensuring consistent content preservation during editing. Additionally, video data captures diverse, natural dynamics—such as non-rigid subject motion and complex camera movements—that are difficult to model otherwise, making it an ideal source for scalable dataset construction. Using this approach, we create a new dataset to train InstructMove, a model capable of instruction-based complex manipulations that are difficult to achieve with synthetically generated datasets. Our model demonstrates state-of-the-art performance in tasks such as adjusting subject poses, rearranging elements, and altering + +camera perspectives. The project page is available here. + +# 1. Introduction + +Recent years have seen remarkable progress in text-to-image (T2I) generation [10, 28, 31-33], with state-of-the-art methods now capable of producing high-quality and hyper-realistic images. The key to this success lies in training on vast datasets containing billions of images from the internet, paired with text descriptions generated by image captioning models [30]. In contrast, editing an existing image based on textual instructions remains challenging, with output quality still falling short and a limited range of editing types available. A major bottleneck in this area is the difficulty of obtaining a comparably large dataset of source-target-instruction triplets for training. + +Previous approaches [7, 8, 14, 20, 25, 26, 40] attempted to adapt pretrained T2I models without fine-tuning, using noise inversion to effectively "regenerate" the target image with modified textual prompts. However, these tuning- + +free methods are often slow and lack robustness. InstructPix2Pix [5] introduced a method to bootstrap a dataset of source images into editing triplets by generating editing instructions using a language model and creating target images using tuning-free techniques. Subsequent works [12, 16, 34, 42] have aimed to refine this process or filter the generated outputs. Nonetheless, a fundamental limitation remains: target images are synthetically generated, which constrains the model's potential due to data quality issues. Consequently, existing methods often struggle to preserve the appearance of edited subjects or to perform precise, detailed edits, as shown in Fig. 2. + +In this paper, we present a novel approach for creating large-scale, instruction-based image-editing datasets using real images sampled from internet videos, capturing naturally realistic transformations from source to target. Our key insight is that video frames capture rich information on how things move — pose changes, element shifts, and camera movements — making them an ideal data source for training models for image manipulation. A similar insight was demonstrated in [3] to show that video frames are effective training data to fix coarse manual edits. We then leverage powerful multi-modal large language models (MLLMs), such as GPT-4o [1] and LLaVA [24] to generate descriptive text at or above human-level quality. + +Our data construction starts with extracting video frames that undergo realistic yet complex transformations. Next, we use MLLMs to analyze these transformations and generate precise editing instructions. Additionally, we introduce a novel spatial conditioning strategy: instead of conventional channel concatenation, we concatenate the reference image with the noise input along the spatial dimension. This approach enables the model to access the reference image through cross-attention across the network, greatly enhancing its ability to perform flexible edits while preserving the appearance of the source image. + +By fine-tuning pretrained text-to-image models (e.g., Stable Diffusion [32]) on our constructed dataset with spatial conditioning strategy, we enable natural-language-driven image manipulations that were previously challenging. These include adjusting subject poses, rearranging elements, or altering camera perspectives, all while maintaining the integrity and details of the original image (see Fig. 1). Our approach also seamlessly integrates with mask-based guidance for local editing, and also with other control types [27, 43], allowing for versatile modifications. + +Our contributions can be summarized as follows: + +- We introduce a data construction pipeline that uses video frames and MLLMs to create source-target-instruction triplets for training image manipulation models. +- We introduce a spatial conditioning strategy that conditions the reference image alongside the noise map to improve content preservation while enabling flexible edits. + +![](images/e140039e28e0c8e774d7413822af56ee66e52cb53dfc83bd44d16a777dcf5768.jpg) +Put the toy's legs together. +Figure 2. Existing methods struggle with complex edits on real photos, such as non-rigid transformations. They often either fail to follow the editing instructions or produce inconsistent outputs. + +![](images/3f334bb4adfccb5ca4d059758215446f601bae958d21de2e071ee82ec2b03d1c.jpg) +InstructPix2Pix + +![](images/e88f24291cba54cf29b21ae285c78bd73ed96c02ede776ef1d5f1689f812c7e9.jpg) +MagicBrush + +![](images/fce7696f8aa4f0a655e76491d9b7ce7b87181991db0d8f933ec90b7c1f3d9313.jpg) +UltraEdit + +- Our approach enables a wider range of text-based image manipulations, including pose adjustments, element rearrangements, and perspective changes, while also supporting precise, localized edits through integration with additional control mechanisms. + +# 2. Related Work + +# 2.1. Zero-shot Image Editing with Diffusion Models + +Diffusion models [15, 35, 37] have become the leading approach in text-to-image generation due to their ability to effectively translate noise into structured data distributions. Leveraging these pretrained models, several image editing techniques have been developed by manipulating the sampling process [4, 8, 17, 20, 22, 25, 26, 29], modifying the internal architecture of the denoising network [7, 14, 40], or incorporating additional optimization steps [19, 23, 44]. + +For example, SDEdit adds noise to the image and then generates the desired edit by conditioning on a target prompt during the reverse denoising process. Other approaches, such as Prompt-to-Prompt [14], Plug-and-Play [40], and MasaCtrl [7], manipulate the attention mechanism after inverting the image to noise using DDIM inversion [36], allowing for edits that align with the desired prompt. More recent works [9, 13, 18, 26, 41] have refined the inversion process to enhance editing robustness and accuracy. While these zero-shot methods offer flexibility in editing images, they are often too slow and lack robustness, making them less suitable for practical applications. + +# 2.2. Training Instruction-based Editing Models + +Comparing to modifying target image descriptions, a more intuitive and user-friendly editing method is to use direct instructions (e.g., “Have the man look at the camera”) to guide the editing process. InstructPix2Pix [5] pioneered this concept by fine-tuning Stable Diffusion [32] with editing instructions. To handle more complex instructions, methods like MGIE [11] and SmartEdit [16] integrate Multimodal LLMs to improve the understanding of editing instructions. + +A major challenge in training instruction-based editing models is constructing paired image data with corresponding instructions. InstructPix2Pix tackles this by using lan + +![](images/0a94d051d110361c181678232f814b24ae75d2a0ce179497a7ae00e0bf408311.jpg) + +![](images/99702fb5f7dfcc4e2a6d09418336b4b8c442d640b8e9c2b50e618ea164bf10ab.jpg) +(a) Frame selection: Sample frame pairs from internet videos. + +![](images/ff7d6e9debac14f82b95cc9e3634b44273ff853f832f718ab70a2c9a089567a7.jpg) + +![](images/23c9f03916ed4a7b487a74da9fabfe52ad1661f092bc5d203b4e2df067a642ca.jpg) + +![](images/f80bd892402f7922c35661a1c28d859ca4edbd6c187374bcdbc36b9f89c3263c.jpg) +(b) Generate instructions: + +![](images/564018e6a8caf1ca93bb8e87c33fad7a4368fd63aa94dbb1c80169428fd53d99.jpg) +Adjust the woman's gaze from the book to looking directly at the camera. +Have the man give a thumbs up gesture. + +![](images/754e56634d8add88a000ff2bb25be7dfdd335c9d820e9683e8518f9163b40aea.jpg) +(c) Generated source-target-instruction triplets for training: +"Remove the smile from the woman's face." +Figure 3. Our data construction pipeline. (a) We begin by sampling suitable frame pairs from videos, ensuring realistic and moderate transformations. (b) These frame pairs are used to prompt Multimodal Large Language Models (MLLMs) to generate detailed editing instructions. (c) This process results in a large-scale dataset with realistic image pairs and precise editing instructions. + +![](images/77bd7bb8f363f7b55723263591af664694625aa193d6313197c63311f99e053a.jpg) +"Rotate the bird to the left and make its tail visible." + +![](images/a5a764da79a68f9e7b2794bddbe68c5d30d43b5e38c80844239894c482feafc7.jpg) +"Make the person looking at the laptop screen." + +guage models, such as GPT-3 [6], to generate both an edited image caption and an editing instruction. These captions are converted into images using Prompt-to-Prompt [14]. MagicBrush [42] constructs a small dataset by performing edits using DALL-E 2 [31]. InstructDiffusion [12] augments datasets with synthetic pairs for tasks like object removal and replacement. EmuEdit [34] improves data pair construction using Llama 2 [39] and local masks with Prompt-to-Prompt, combining these with other vision datasets to build a large-scale dataset. + +However, in existing large-scale editing datasets, most target images are generated using synthetic methods, which often introduces artifacts and significant appearance deviations from the source images. These datasets are typically limited to high-level edits such as object overlay and style transfer. As a result, models trained on these datasets struggle with complex edits on real photos, such as non-rigid transformations and viewpoint changes. In contrast, our approach leverages real video frames as source and target images, using Multimodal LLMs to generate instructions directly from these frames. This allows for the construction of a more realistic dataset, capturing complex transformations and overcoming the limitations of previous methods. + +# 3. Dataset Construction + +Unlike existing instruction-based editing datasets, which rely on tuning-free methods to generate target images, we propose a novel data construction pipeline that leverages real video frame pairs. By extracting frame pairs that capture meaningful transformations and using Multimodal Large Language Models (MLLMs) to generate detailed editing instructions, we create a more realistic and high- + +quality editing dataset. In Sec. 3.1, we outline our approach for extracting suitable frame pairs from videos, followed by Sec. 3.2, where we describe how we use MLLMs to generate precise editing instructions. Finally, Sec. 3.3 provides an in-depth analysis of the resulting dataset. An overview of the data construction pipeline is illustrated in Fig. 3. + +# 3.1. Sampling Frame Pairs from Videos + +In a video, the same subjects and elements can appear in different ways, move naturally, or be shown from different angles. For example, a person might change their posture, expression, or movement over time. These changes provide useful clues about how a scene or perspective might look with small adjustments, making videos a great source for picking frame pairs to use in image editing tasks. However, it's important to sample the frame pairs with proper criteria to make sure they are suitable for training. + +Initial sampling of paired frames. We begin by captioning each video to generate an overall description. Videos containing specific keywords (e.g., landscape, abstract, still) that are not suitable for image editing are filtered out. We then sample two frames $(I^s, I^e)$ with a fixed time interval (i.e., 3-5 seconds) to ensure sufficient transformations occur in the objects of interest. + +Motion-based filtering. To refine the selection, we apply motion filtering to exclude frames with either too little or too much movement. Specifically, we use RAFT [38] to compute the optical flow $O_{I^s \to I^e}$ from $I^s$ to $I^e$ and derive the flow magnitude $M = \sqrt{O_x^2 + O_y^2}$ . Frame pairs with moderate movement are retained, while those with excessive or minimal motion are discarded. To further ensure consistency, we assess background changes by computing a background occlusion mask through backward warping of $I^e$ to $I^s$ using the estimated flow. By measuring the occlusion mask ratio, we filter out frames with significant background changes (i.e., large masked regions). This motion filtering pipeline allows us to obtain image pairs with realistic, moderate transformations. + +# 3.2. Instruction Generation with MLLMs + +Next, we describe how we generate accurate editing instructions $C$ given frame pairs $(I^s, I^e)$ . Previous approaches [5, 34] rely solely on text-based methods, using Large Language Models (LLMs) to generate instructions and captions of target images purely from the source caption. This approach significantly limits the variety of edits that can be captured in the dataset. As a result, these methods are often constrained to high-level edits, such as adding, removing, or replacing elements, or changing the style. + +To address this limitation, we utilize multimodal LLMs, such as GPT-4o [1] or Pixtral-12B [2], to directly analyze frame pairs and generate editing instructions. Specifically, + +
DatasetsReal Target?# EditsEditing Types
Non-rigidViewpointAdd/RemoveStyle
InstructPix2Pix [5]X0.3MXX✓✓
MagicBrush [42]X10KXX✓✓
EmuEdit [34]X10MXX✓✓
UltraEdit [45]X4.1MXX✓✓
Ours✓6M✓✓XX
+ +Table 1. Comparison with existing image-editing datasets. Our dataset is the first large-scale dataset with real target images while supporting complex editing types, such as non-rigid transformations and camera viewpoint adjustments. + +MLLMs are prompted to examine the differences between the source and target images, focusing on changes in subjects, relative positions, camera angles, and background. Based on these observations, the models generate detailed, context-specific instructions. + +To improve clarity, we instruct the MLLMs to use absolute terms (e.g., "Move the bee to the center of the flower" instead of "Move the bee to the position in the target image") and to begin instructions with an action verb (e.g., Change, Move, Adjust) for better usability in real-world applications. Additionally, to ensure high-quality instructions, we allow the MLLMs to reject frame pairs where the edits are too complex to describe accurately, and we exclude such pairs from our dataset. + +# 3.3. Dataset Analysis + +In total, after filtering, we sampled 6 million image pairs from videos, each with appropriate levels of motion and transformation. All frames are captioned using Pixtral-12B [2]. The edits captured in our dataset focus on realistic transformations, such as non-rigid changes of subjects/objects, rearranging elements, and altering camera perspectives—types of edits that are largely absent in existing datasets. Examples can be seen in Fig. 3, with comparisons to prior datasets presented in Table 1. + +Our dataset can be combined with existing ones to train models that handle both realistic edits introduced in this paper, and the artistic edits explored in previous works (e.g., converting a photo into a painting), as shown in the supplementary material. + +As illustrated in Fig. 3, some video frames may include subtle changes that are not fully captured in the generated instructions. However, our experiments show that these small discrepancies do not significantly impact the training process, as the model focuses on learning the major transformations between frame pairs. + +# 4. Method + +Using the constructed dataset, we fine-tune a pretrained T2I model for instruction-based image manipulation, enabling edits such as non-rigid transformations of subjects, object + +![](images/8f37a491d91ff19762ed23456a1a5cb3fe7319f0361dcd2f7319f5aaae16607d.jpg) +Figure 4. Overview of the proposed model architecture for instruction-based image editing. The source and target images are first encoded into latent representations $z^s$ and $z^e$ using a pretrained encoder. The target latent $z^e$ is then transformed into a noisy latent $z_t^e$ through the forward diffusion process. We concatenate the source image latent and the noisy target latent along the width dimension to form the model input, which is fed into the denoising U-Net $\epsilon_{\theta}$ to predict a noise map. The right half of the output, corresponding to the noisy target input, is cropped and compared with the original noise map. + +repositioning, and changes in viewpoint. Unlike previous methods, in Sec. 4.1 we introduce a novel spatial conditioning approach that conditions the pretrained T2I models on reference images without requiring any architectural modifications. This technique allows instructions to control the target structure while preserving the identity and content of the original image. Since our method does not alter the underlying architecture of the pretrained T2I models, it seamlessly integrates with additional control mechanisms, such as masks or ControlNets [43], to enable even more precise and localized edits (Sec. 4.2). + +# 4.1. Finetuning T2I Models for Image Manipulation + +Existing instruction-based editing models [5, 11, 16, 34] typically condition T2I models on a reference image by concatenating it with the noise map along the channel dimension as input to the denoising network. However, this channel-based conditioning method spatially aligns the reference and target images, which limits the model's flexibility in making structural changes, such as non-rigid transformations or viewpoint adjustments. + +To address this, we introduce Spatial Conditioning, which conditions the model by concatenating the source image with the noise map along the spatial dimension. + +Given a triplet $(I^s, I^e, C)$ , we first encode both images into latent representations $z^s$ and $z^e$ using a pretrained variational autoencoder [21]. The target latent $z^e$ is then transformed into a noisy latent $z_t^e$ via the forward diffusion process. We concatenate the source image latent with the noisy target latent along the width dimension to create the model input $z_t = \mathrm{Concat}_{width}([z^s, z_t^e])$ . This input, which doubles the width of the noisy latent, is fed into the denoising U-Net $\epsilon_{\theta}$ to predict a noise map of the same width. We then + +crop the right half of the output corresponding to the noisy input and compute the denoising loss: + +$$ +\mathcal {L} _ {\text {E d i t}} = \mathbb {E} _ {z _ {t}, C, t, \epsilon} \left[ \| \epsilon - \operatorname {C r o p} _ {\text {w i d t h}} \left(\epsilon_ {\theta} \left(z _ {t}, C, t\right)\right) \| ^ {2} \right]. \tag {1} +$$ + +This spatial conditioning approach, while seemingly similar to channel conditioning, offers critical advantages for image editing tasks. First, it does not spatially align the reference image with the noisy input, allowing the denoising network to process the condition and the noisy input in parallel. This enables the network to attend to any region of the condition image at each cross-attention layer, facilitating structural changes in the target image. Moreover, since the network has access to the features of the condition image throughout all layers, it can better preserve the details and identity of the source image, leading to more accurate and realistic edits. + +# 4.2. Incorporating Additional Controls + +Once trained, our editing model can perform a wide range of edits based on user-provided instructions. However, text instructions alone are often insufficiently precise, making additional controls essential for accurate image editing. + +Integration with masks. Similar to zero-shot inpainting in T2I models, our approach supports mask-based localization to edit specific regions of an image. During inference, we blend the updated latent with the reference latent using a mask to control the area of modification: + +$$ +z _ {t - 1} ^ {*} = (1 - m) \cdot z _ {t - 1} ^ {s} + m \cdot z _ {t - 1}, \tag {2} +$$ + +where $z_{t-1}^{s}$ is the noisy latent from the forward diffusion of the source image latent $z^{s}$ , and $m$ is the mask resized to match the latent resolution. + +Integration with other spatial controls. For more complex edits, additional visual cues beyond text can greatly enhance control. Users can provide sketches, or adjust key points of a skeleton to modify poses, for example. To enable this, we integrate controllable diffusion models like ControlNet [43] and T2I-Adapter [27]. Crucially, because our model retains the architecture of the original T2I model, we can directly use these control techniques with our trained image-editing models, incorporating spatial maps for finer control. This compatibility is not possible with previous instruction-based editing models like InstructPix2Pix. + +# 5. Experiments + +In Sec. 5.1, we outline the experimental setup, including the evaluation benchmark and quantitative metrics. In Sec. 5.2, we compare with state-of-the-art text-guided image editing methods both quantitatively and qualitatively. Sec. 5.3 demonstrates our model's ability to incorporate additional controls for more precise editing. Finally, we assess the impact of our dataset and conditioning strategy in Sec. 5.4. + +# 5.1. Experimental Setup + +Implementation Details. We fine-tune Stable Diffusion V1.5 [32] with the proposed spatial conditioning strategy on our newly constructed dataset to enable various types of complex image edits. Input images have a resolution of $512 \times 512$ . The model is trained for 100,000 iterations with a constant learning rate of $1 \times 10^{-4}$ , using the Adam optimizer to update model parameters. Training is conducted on 8 NVIDIA A100 GPUs with a total batch size of 256. During sampling, we employ the DDIM [36] scheduler with 50 steps to generate the final edited images. Note that mask and additional spatial controls are optional; when unspecified, results are generated solely based on textual instructions. + +Evaluation benchmark. Existing instruction-based image editing benchmarks, such as MagicBrush [42] and EmuEdit [34], primarily focus on editing tasks that maintain the original image structure, such as style modification, object replacement, and color or texture changes. These benchmarks are therefore unsuitable for evaluating non-rigid, consistent image editing tasks that alter the image structure while preserving identity and appearance details. To address this gap, we created a specialized test set tailored for the non-rigid editing tasks presented. Our benchmark consists of 50 images, each paired with human-curated editing instructions that specify modifications in pose, expression, viewpoint, shape, position, and other structural transformations. The benchmark excludes any local or global style or appearance edits. Examples of input images and their corresponding instructions are shown in Fig. 5. More details are in the supplementary material. + +Evaluation metrics. For evaluation, we employ metrics to assess two main aspects of image editing: alignment with text instructions and fidelity to the source image. To evaluate instruction adherence, following previous works, we first generate captions for the target image using a large language model (LLM) based on the source image's caption and the given instructions. We then assess the agreement between caption changes and actual image changes (CLIP-D). However, this metric relies on generating captions for target images, which may not directly measure alignment with the original instructions. To address this, we propose a new metric: we prompt MLLMs to generate instructions based on the input and output images for each method, then compute the CLIP distance between these generated instructions and the original ones, denoted as CLIP-Inst. For evaluating faithfulness, we calculate the CLIP feature distance between the input and output images (CLIP-I). + +We also collect human feedback from 40 participants to evaluate the quality of the edited images. Each participant is shown a subset of 20 examples, including the outputs from each method, alongside the source image and the editing + +
MethodCLIP-D ↑CLIP-Inst↑CLIP-I↑
NullTextInversion [26]0.06600.76480.9063
MasaCtrl [7]0.04360.85270.9160
Imagic [19]0.05080.86450.8379
InstructPix2Pix [5]0.08870.85690.9380
MagicBrush [42]0.09720.86480.9318
UltraEdit [45]0.08240.85710.9184
Ours0.13610.87240.9275
+ +Table 2. Quantitative comparison with state-of-the-art text-guided image editing methods on our collected benchmark. + +
Imagic [19]InstructPix2Pix [5]MagicBrush [42]Ours
Preference5.0%3.25%4.13%87.62%
+ +instruction. Participants are asked to choose the image that "best follows the editing instruction while maintaining the highest quality and preserving the original content." + +# 5.2. Baseline Comparisons + +Methods for comparison. We compare our proposed method against state-of-the-art text-guided image editing approaches with publicly available models, including zero-shot description-based methods: NullTextInversion [26], MasaCtrl [7], and Imagic [19], as well as instruction-based methods: InstructPix2Pix [5], MagicBrush [42], and UltraEdit [45]. For description-guided methods, we generate the target description by captioning the input image and deriving the description from the instructions using an LLM. + +Quantitative results. Table 2 presents the quantitative evaluation of both description- and instruction-based methods on our editing benchmark. A method that perfectly preserves content but fails to follow the editing instructions is still inadequate. For instance, InstructPix2Pix and MagicBrush show higher CLIP-I scores than our approach because they struggle with non-rigid edits and viewpoint changes. As a result, they often leave the input image mostly unchanged (see Fig. 5), leading to high content preservation scores. In contrast, our method achieves the highest scores in instruction alignment while still maintaining strong content preservation. This is further validated by the user study in Table 3, where in over $87.62\%$ of cases, users preferred our method over the other four approaches. + +Qualitative results. We present a qualitative comparison of various methods in Fig. 5. Since prior approaches are primarily designed for editing tasks that preserve the original image structure, they struggle with non-rigid edits (such + +Table 3. Human preference results. Percentage indicates how often each method was selected as the best result. + +
MethodCLIP-D↑CLIP-Inst ↑CLIP-I ↑
SC + IP2P data0.12770.84140.9094
CC + Our data0.08530.86790.8552
SC + Our data0.13610.87240.9275
+ +Table 4. Ablation on the training dataset and conditioning approach. Training on the InstructPix2Pix dataset [5] leads to poorer instruction alignment and content preservation. Additionally, our Spatial Conditioning (SC) method outperforms the standard Channel Conditioning (CC) approach. + +as changes in pose and expression), object repositioning, or viewpoint adjustments. These methods often fail to follow such editing instructions, either producing images identical to the source or generating edits that do not align with the specified instructions. Even in cases where the instructions are somewhat followed, they frequently fail to maintain content consistency. In contrast, our method, trained on real video frames with naturalistic transformations, is capable of handling these complex edits while preserving consistency with the input images. + +# 5.3. Results with Additional Controls + +We demonstrate the incorporation of additional controls for precise editing in Fig. 6. Instructions alone can sometimes be ambiguous regarding which part of the image to edit; in such cases, a mask can be used for localized adjustments (Fig. 6 (a)). In other scenarios, instructions alone may not suffice, especially when precise control is needed for adjustments like altering a subject's pose. Users can address this by directly adjusting control points to define a skeleton pose or providing a rough sketch of the desired composition. We show that our approach can seamlessly integrate ControlNet to utilize such controls for precise editing (Fig. 6 (b)). + +# 5.4. Ablations + +We argue that our proposed video + MLLM dataset construction pipeline, along with the Spatial Conditioning (SC) approach, enables our model to perform complex image edits while maintaining appearance consistency. We evaluate the impact of each component in Table 4 and Fig. 7. + +Dataset. To demonstrate the importance of using real video frame pairs for training image editing models, we fine-tune a model with the same architecture as ours on the dataset proposed in InstructPix2Pix [5], where target images are synthetically generated using Prompt-to-Prompt [14]. Compared to our model trained on InstructPix2Pix's dataset (denoted as "SC + IP2P data"), our method not only better follows editing instructions but also preserves the original image content more effectively. + +Spatial Conditioning. We also fine-tune our model using our dataset, but with the Channel Conditioning (CC) + +![](images/91cacfe3ed31cf4d5b1705a4f881f250717646e5a869be5cf993b9a7a0090971.jpg) + +![](images/f440677de63a429f8cea8c1dea8f381b5201112a0b8067b9677a8c316193e0ac.jpg) +A bear toy wearing a sweater with arm raised + +![](images/d903e7caf616179af1997e7dcc565caf3b1b78e0d84aeffb915d5d1430a8b141.jpg) + +![](images/ffd30630ce5c97525268855b530a31448d46a038a0368b2f91b2076b937a63fc.jpg) + +![](images/f6ce94ac85d29e6938ece0e0e41475a0d6eb86f51380907f416aaf440a974897.jpg) +A dog sitting on a ledge looking at the camera + +![](images/433d728936984dafb5928bb122c8efc9e514027ff7e8fff710c8ae2f02b700d3.jpg) + +![](images/56b40da103c56a33d1f156a4616774885d7777c3bca6917ecc46d2b60e801898.jpg) + +![](images/9a82ed22aace129490f3e54c84e5bd69b36248213831d4d12ecea3eed040c296.jpg) +A straight tower against the sky + +![](images/7347f24ff28b4bdee459c835c1d10b724a00ec845a673927d7ab9201382a8b7e.jpg) + +![](images/09fc6ff167cf0b45b2dbf3b5698e4c7802a67f192e9005400fdfefdce5aa6855.jpg) + +![](images/7d7773ca3e36dfcc3a15632bc4b701424d1b5840cb52d3f4f94993890b277603.jpg) +A man meditating and giving a thumbs up + +![](images/df977abb63050a56ddc7b28d8baf2fa1b386a1f3d7bba373604fb67cebc0a2d6.jpg) + +![](images/e50f9ae78483a179daab2d8f9fdae63a53ed86ff0410cfad130e10e0df303fb2.jpg) + +![](images/e1977b9db7dcf1c57475a88f83e1c6c6c0e74f281f3af7a210de3e2cc185667c.jpg) +A woman with blonde braids smiling at camera. + +![](images/3385714f27dff57efe9bf754708ac6f3b380c3217152fd17e211f71a3c0d6741.jpg) + +![](images/0e4cf758af8fe081e1081f7dc5d74e8000ca2c65e74f473ec458824520d1e0fd.jpg) + +![](images/e121415176c4350688b35f9c8c08d06402d2bcd83b80fb8dfc8de78dff7708a0.jpg) +A brown dog jumping in the air + +![](images/51673d34f7e67828b59880980327f9253c3c1cc6150259c9b4fd57b5ac8bf5cc.jpg) + +![](images/c10f52907241c2000243e3bd1a7cb92c1c3e28c1796b80988e836f989b7b5ca1.jpg) +Figure 5. Qualitative comparison with state-of-the-art image editing methods, including both description-based and instruction-based approaches. Existing methods struggle with complex edits such as non-rigid transformations (e.g., changes in pose and expression), object repositioning, or viewpoint adjustments. They often either fail to follow the editing instructions or produce images with inconsistencies, such as identity shifts. In contrast, our method, trained on real video frames with naturalistic transformations, successfully handles these edits while maintaining consistency with the original input images. + +![](images/462d286b69b1a02d6707d3b1ca48038c3c6db8af7de42f849b598ad044be3fae.jpg) +A home office, viewed from the left + +![](images/5ec98dc2cad1b21476df5658cb2df93ed8507037f3eb57b4f87c94997a3d55ac.jpg) + +![](images/75b104fc8d4b1779f91273a2fe1304860e95afcce2fbcb0c2b82cdee28504780.jpg) +InstructPix2Pix + +![](images/3466b4087e60390eae434a62ff7ee109cbc107aa668f807e677b2959bb68a160.jpg) +MagicBrush +Raise the bear doll's arm. + +![](images/5c941c4dbe61749aa24e27487c0ca8c1c3af30df7f915b813d160517f264411e.jpg) +UltraEdit + +![](images/e3739f64bc9621d60f34a99cb6e83ddbe626e9163d4eac6ea21da37aeeba543b.jpg) +Ours + +![](images/cbc76734170b8fbcf023e88c8887b8cc1c0970c2efae808c2062f7a3c8adddf1.jpg) + +![](images/bb4a6988d62e144719af3bd0643400b3ade30b8cecc42f398ca8578854b60fd7.jpg) +Make the dog look at the camera + +![](images/79f211ee573eda276b6daf7d78c9b7231a41f7f6743b7c196ac09c6f2f1c2645.jpg) + +![](images/de8da886337501fb3feca3903078b04a37686b15fc7e029bedf2a3f15df31233.jpg) + +![](images/083bbac3e50923e2e50f6130325becfd0131ae613ffc041d733ed56fd47fe3e7.jpg) + +![](images/07ade9e662f26855a26cbcee48c5da85d78ea04932a8e8cdc062400e33bf1df1.jpg) +Make the building straight + +![](images/cf4b1f3c544180ad5995c49bc604eeb8a0d5c5be21209c8458d81a3b8e1cf02a.jpg) + +![](images/e33fce7a1d6d356d0a9a56a820f886ef28239ec2bae7fb7090bd139ff40933fb.jpg) + +![](images/c3f3a404b919828ebf83709039220bce10d30cd47d4e3bafd125184facb6411d.jpg) + +![](images/46c9da0f00d6894ead964e35dd1bb8f2eb2d41fa120b1207ca5afbede10e042d.jpg) +Have the man give a thumbs-up + +![](images/141851b06ac8d87c308211aa65a80d868a141fe97e21edef2dec16f50ea71380.jpg) + +![](images/459387dc18997e3249b56fe07a87e29d70705506d05ee84d74f231730e9f51ec.jpg) + +![](images/2e17669b5536ff5720a470d00c854d25c792943b23f12e32bff5af515826cf2a.jpg) + +![](images/c9a47a3aed6288d61566178cfeebe8e98b69f0d04890c8187eb379ef864a954a.jpg) +Make the girl smile at the camera + +![](images/a60ddb8e718980f52377fdfac40c892df93cc65ddc9a0688cd44ecf1c6c5ff71.jpg) + +![](images/5823349de2693976a327192d7f05ba0b3cbce849e60ac0e7591e6d16b74325f7.jpg) + +![](images/c60eb66624b317f02e212292f8c7aa215142787c2a989991327826f018807320.jpg) + +![](images/693268b8d5cb7f23dd9169200df79263817b1d03982ec269e1c7fc6cadbc79f0.jpg) +Make the dog jump in the air + +![](images/7e59dc1877d9ed207f6b700ecbdcc902c5eef36749dc36899bcafc43ad336900.jpg) + +![](images/156b9f24f58bfae6178e0b120568ed2d3f541d264371ff60d266911f7f1eb00c.jpg) + +![](images/583fed326c5485b24f6e57cc836dcd43adefd5114a199fa2f6c345ffdb1f30dd.jpg) + +![](images/98eda8ae1494c9ddc2a8d40fba797c5dfd665b0c01cadc145156c1f5bb2f07b7.jpg) +Move the camera to the right + +![](images/7baf980a2b2f5fde432f67c01ff754db57f1ac2c11f2d8e1b09fe7fbec16ceaf.jpg) + +![](images/0eef6f7683b4b64d65484b00340c4b0caf24a02990c3b40020651216f09a2cff.jpg) + +![](images/ca76addccb123eaaed823414d5bcaafc3127279df901d3dadb05cd0c42e61604.jpg) +Make the girl look up +(a) Localized image editing + +![](images/7922f8ed3b9bc4f85d4f57638d5bc80f32cdd7605fcdfbd5f8e1a62b344a66d5.jpg) +Rotate the banana 45 degrees anti-clockwise. + Scribble +(b) Results with additional controls + +![](images/6e2f30d9c498906001d218aeaf9d72f56df715ffcbbb3c06e5a2ebe305e8f93d.jpg) +Figure 6. Qualitative results of our method with additional controls. (a) Our model can utilize a mask to specify which part of the image to edit, enabling localized adjustments and resolving ambiguities in the instructions. (b) When combined with ControlNet, our model can accept additional inputs, such as human poses or rough sketches, to achieve precise edits in subject poses or object positioning. This level of control is not possible with previous methods. +Make the woman smile +at the camera + +![](images/77cec6bb74a0a99c5a0447086cb9367dda8bace4db1c667ae324afb8f71d9cc7.jpg) +SC + IP2P data + +![](images/cd3bfc7762341cfa7f5fb152a1ef38278047c583c10f28f1aaea05034d0e7120.jpg) +CC + Our data + +![](images/7f8bf1c09c1f9e9451aced58975f0516d09fca58f9093a3d0ff267064afd5fe0.jpg) +SC + Ours data + +![](images/80e8596f28541697564e9aa8b956f3e9815159d0ffcd2501ee3748628a64829c.jpg) +Figure 7. Qualitative ablation on the training dataset and conditioning approach. Compared to models trained on the Instruct-Pix2Pix dataset (IP2P) [5] or using the standard Channel Conditioning (CC) method, our model trained on our dataset with the Spatial Conditioning (SC) approach demonstrates superior instruction alignment and content preservation. +Input Image +Make them look to the side +Figure 8. Limitations. Left: Occasional unintended slight viewpoint changes. Right: Inability to accurately isolate specific objects based on instructions. + +![](images/18a83c05ecaa656cd804d83d7ce9612680548a1feecf4e264b6f441c437ea115.jpg) +Input Image + +![](images/f1da111493e24878a20c8be53481520429c75af2cd9cdc4672b4000645814bfe.jpg) +Lower the taller dog's head + +method, where the reference latent is concatenated with the target latent along the channel dimension rather than the spatial dimension. As shown in Table 4 and Fig. 7, the proposed Spatial Conditioning approach significantly improves the model's ability to apply complex edits while achieving higher quality and better content preservation. + +# 6. Limitation + +Our approach introduces a novel data construction pipeline, but it still has some limitations. The quality of the dataset is influenced by the frame filtering process and the capabilities of the MLLMs. Occasionally, MLLMs may generate inaccurate instructions or fail to capture all transformations between frames, leading our editing model to apply unintended changes, such as slight viewpoint shifts that deviate from the original instructions. The effectiveness of our model also depends on the accuracy of training instructions and the comprehension abilities of pretrained T2I models, which can affect its performance on complex edits. Examples of these limitations are shown in Fig. 8. Moreover, since the transformations captured from video frames are restricted to realistic changes, our model cannot handle artistic edits like style transfer or object replacement. This limitation can be mitigated by integrating our dataset with others, as demonstrated in the supplementary material. + +# 7. Conclusion + +We present an approach for sampling video frames and utilizing MLLMs to generate editing instructions for training instruction-based image manipulation models. Unlike existing datasets, which rely on synthetically generated target images, our method leverages supervision signals from videos and MLLMs to support complex edits, such as nonrigid transformations and viewpoint changes, while preserving content consistency. Future work could focus on refining filtering techniques, either by improving MLLMs or incorporating human-in-the-loop processes, as well as integrating video data with other datasets to further enhance image editing capabilities. + +# Acknowledgment + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3 +[2] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Devendra Chaplot, Jessica Chudnovsky, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliiou, Paul Jacob, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 3, 4 +[3] Hadi Alzayer, Zhihao Xia, Xuaner Zhang, Eli Shechtman, Jia-Bin Huang, and Michael Gharbi. Magic fixup: Streamlining photo editing by watching dynamic videos. arXiv preprint arXiv:2403.13044, 2024. 2 +[4] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2 +[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2, 3, 4, 6, 8 +[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3 +[7] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 1, 2, 6 +[8] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 1, 2 +[9] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. arXiv preprint arXiv:2408.00735, 2024. 2 +[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1 +[11] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 2, 4 + +We would like to thank Nanxuan Zhao and Yilin Wang for their valuable feedback. This research was supported in part by JSPS KAKENHI with Grant Numbers 24KK0209, 24K22318, 22H00529, and JST-Mirai Program JPMJMI23G1. +[12] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024. 2, 3 +[13] Ligong Han, Song Wen, Qi Chen, Zhixing Zhang, Kunpeng Song, Mengwei Ren, Ruijiang Gao, Anastasis Stathopoulos, Xiaoxiao He, Yuxiao Chen, et al. Proxedit: Improving tuning-free real image editing with proximal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4291-4301, 2024. 2 +[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 3, 6 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2 +[16] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024. 2, 4 +[17] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 2 +[18] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2 +[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2, 6 +[20] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 1, 2 +[21] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4 +[22] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 2 +[23] Pengzhi Li, Qinxuan Huang, Yikang Ding, and Zhiheng Li. Layerdiffusion: Layered controlled image editing with diffusion models. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 2 +[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 2 + +[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 1, 2 +[26] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 1, 2, 6 +[27] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 2, 5 +[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1 +[29] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH Conference Proceedings, pages 1-11, 2023. 2 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1 +[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 3 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 5 +[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1 +[34] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871-8879, 2024. 2, 3, 4, 5 +[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2 +[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 5 + +[37] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 2 +[38] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision, pages 402-419. Springer, 2020. 3 +[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3 +[40] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2 +[41] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. Turboedit: Instant text-based image editing. arXiv preprint arXiv:2408.08332, 2024. 2 +[42] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4, 5, 6 +[43] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4, 5 +[44] Shiwen Zhang, Shuai Xiao, and Weilin Huang. Forgedit: Text guided image editing via learning and forgetting. arXiv preprint arXiv:2309.10556, 2023. 2 +[45] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. 4, 6 \ No newline at end of file diff --git a/2025/Instruction-based Image Manipulation by Watching How Things Move/images.zip b/2025/Instruction-based Image Manipulation by Watching How Things Move/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e9e8446c8b5594fe35501900cc3914cbaf72ce12 --- /dev/null +++ b/2025/Instruction-based Image Manipulation by Watching How Things Move/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d6ca9b0ce2d55ba68c21feb0b26492f81cca75226a15c355b6844179c5e5400 +size 807235 diff --git a/2025/Instruction-based Image Manipulation by Watching How Things Move/layout.json b/2025/Instruction-based Image Manipulation by Watching How Things Move/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9bf05cbb13dfbb5a55e7933b2f5aa7cf41cc5c2d --- /dev/null +++ b/2025/Instruction-based Image Manipulation by Watching How Things Move/layout.json @@ -0,0 +1,11264 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 523, + 121 + ], + "type": "text", + "content": "Instruction-based Image Manipulation by Watching How Things Move" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "type": "text", + "content": "Mingdeng Cao" + }, + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "type": "text", + "content": " Xuaner Zhang" + }, + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "type": "text", + "content": " Yinqiang Zheng" + }, + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "type": "text", + "content": " Zhihao Xia" + }, + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "type": "text", + "content": "The University of Tokyo " + }, + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 142, + 473, + 186 + ], + "type": "text", + "content": "Adobe" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 213, + 220, + 296 + ], + "blocks": [ + { + "bbox": [ + 57, + 213, + 220, + 296 + ], + "lines": [ + { + "bbox": [ + 57, + 213, + 220, + 296 + ], + "spans": [ + { + "bbox": [ + 57, + 213, + 220, + 296 + ], + "type": "image", + "image_path": "d25058744bf9235de49aceef260d561d3ab706146174d6e4fdb099546b211f0b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 296, + 167, + 302 + ], + "lines": [ + { + "bbox": [ + 111, + 296, + 167, + 302 + ], + "spans": [ + { + "bbox": [ + 111, + 296, + 167, + 302 + ], + "type": "text", + "content": "Lower the horse's head" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 303, + 220, + 385 + ], + "blocks": [ + { + "bbox": [ + 57, + 303, + 220, + 385 + ], + "lines": [ + { + "bbox": [ + 57, + 303, + 220, + 385 + ], + "spans": [ + { + "bbox": [ + 57, + 303, + 220, + 385 + ], + "type": "image", + "image_path": "5a14f0a26670d007471a58159de2ed31c450ced8c865ffe825d81222bba32aa8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 385, + 172, + 392 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 172, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 172, + 392 + ], + "type": "text", + "content": "Change the view to the side" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 404, + 555, + 448 + ], + "lines": [ + { + "bbox": [ + 55, + 404, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 555, + 448 + ], + "type": "text", + "content": "Figure 1. We propose InstructMove, an instruction-based image editing model trained on frame pairs from videos with instructions generated by Multimodal LLMs. Our model excels at non-rigid editing, such as adjusting subject poses, expressions, and altering viewpoints, while maintaining content consistency. Additionally, our method supports precise, localized edits through the integration of masks, human poses, and other control mechanisms." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 213, + 388, + 296 + ], + "blocks": [ + { + "bbox": [ + 223, + 213, + 388, + 296 + ], + "lines": [ + { + "bbox": [ + 223, + 213, + 388, + 296 + ], + "spans": [ + { + "bbox": [ + 223, + 213, + 388, + 296 + ], + "type": "image", + "image_path": "17f7a0c35cb4815bc62ff01c8a3c211d75d86329308db058b578c5e8da90ac7d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 296, + 335, + 303 + ], + "lines": [ + { + "bbox": [ + 276, + 296, + 335, + 303 + ], + "spans": [ + { + "bbox": [ + 276, + 296, + 335, + 303 + ], + "type": "text", + "content": "Make the man look angry" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 224, + 303, + 388, + 385 + ], + "blocks": [ + { + "bbox": [ + 224, + 303, + 388, + 385 + ], + "lines": [ + { + "bbox": [ + 224, + 303, + 388, + 385 + ], + "spans": [ + { + "bbox": [ + 224, + 303, + 388, + 385 + ], + "type": "image", + "image_path": "15b0aa691a8b3eb16dd0081078756242750a2239fec0d998aa817a8478e3e9d9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 385, + 340, + 392 + ], + "lines": [ + { + "bbox": [ + 271, + 385, + 340, + 392 + ], + "spans": [ + { + "bbox": [ + 271, + 385, + 340, + 392 + ], + "type": "text", + "content": "Move the camera to the left" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 391, + 213, + 553, + 296 + ], + "blocks": [ + { + "bbox": [ + 391, + 213, + 553, + 296 + ], + "lines": [ + { + "bbox": [ + 391, + 213, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 391, + 213, + 553, + 296 + ], + "type": "image", + "image_path": "d5fc7c1e13b1318b8f3033055211036a43f71df900efcf963af1a150e8db68d7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 296, + 494, + 303 + ], + "lines": [ + { + "bbox": [ + 449, + 296, + 494, + 303 + ], + "spans": [ + { + "bbox": [ + 449, + 296, + 494, + 303 + ], + "type": "text", + "content": "Close the dog's eye" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 391, + 303, + 553, + 384 + ], + "blocks": [ + { + "bbox": [ + 391, + 303, + 553, + 384 + ], + "lines": [ + { + "bbox": [ + 391, + 303, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 391, + 303, + 553, + 384 + ], + "type": "image", + "image_path": "c153208fc770eae9915e8b12dfe87edef1bc4e8d1392b4613053c479e14a449c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 385, + 507, + 391 + ], + "lines": [ + { + "bbox": [ + 436, + 385, + 507, + 391 + ], + "spans": [ + { + "bbox": [ + 436, + 385, + 507, + 391 + ], + "type": "text", + "content": "Have the man look at the side." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 152, + 470, + 200, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 470, + 200, + 483 + ], + "spans": [ + { + "bbox": [ + 152, + 470, + 200, + 483 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 54, + 498, + 297, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 297, + 691 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 297, + 691 + ], + "type": "text", + "content": "This paper introduces a novel dataset construction pipeline that samples pairs of frames from videos and uses multimodal large language models (MLLMs) to generate editing instructions for training instruction-based image manipulation models. Video frames inherently preserve the identity of subjects and scenes, ensuring consistent content preservation during editing. Additionally, video data captures diverse, natural dynamics—such as non-rigid subject motion and complex camera movements—that are difficult to model otherwise, making it an ideal source for scalable dataset construction. Using this approach, we create a new dataset to train InstructMove, a model capable of instruction-based complex manipulations that are difficult to achieve with synthetically generated datasets. Our model demonstrates state-of-the-art performance in tasks such as adjusting subject poses, rearranging elements, and altering" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 460, + 541, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 460, + 541, + 472 + ], + "spans": [ + { + "bbox": [ + 314, + 460, + 541, + 472 + ], + "type": "text", + "content": "camera perspectives. The project page is available here." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 498, + 394, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 498, + 394, + 510 + ], + "spans": [ + { + "bbox": [ + 314, + 498, + 394, + 510 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 521, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 555, + 664 + ], + "type": "text", + "content": "Recent years have seen remarkable progress in text-to-image (T2I) generation [10, 28, 31-33], with state-of-the-art methods now capable of producing high-quality and hyper-realistic images. The key to this success lies in training on vast datasets containing billions of images from the internet, paired with text descriptions generated by image captioning models [30]. In contrast, editing an existing image based on textual instructions remains challenging, with output quality still falling short and a limited range of editing types available. A major bottleneck in this area is the difficulty of obtaining a comparably large dataset of source-target-instruction triplets for training." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Previous approaches [7, 8, 14, 20, 25, 26, 40] attempted to adapt pretrained T2I models without fine-tuning, using noise inversion to effectively \"regenerate\" the target image with modified textual prompts. However, these tuning-" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 212, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 212, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 212, + 712 + ], + "type": "text", + "content": "† Work done during an internship at Adobe." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2704" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 216 + ], + "type": "text", + "content": "free methods are often slow and lack robustness. InstructPix2Pix [5] introduced a method to bootstrap a dataset of source images into editing triplets by generating editing instructions using a language model and creating target images using tuning-free techniques. Subsequent works [12, 16, 34, 42] have aimed to refine this process or filter the generated outputs. Nonetheless, a fundamental limitation remains: target images are synthetically generated, which constrains the model's potential due to data quality issues. Consequently, existing methods often struggle to preserve the appearance of edited subjects or to perform precise, detailed edits, as shown in Fig. 2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "type": "text", + "content": "In this paper, we present a novel approach for creating large-scale, instruction-based image-editing datasets using real images sampled from internet videos, capturing naturally realistic transformations from source to target. Our key insight is that video frames capture rich information on how things move — pose changes, element shifts, and camera movements — making them an ideal data source for training models for image manipulation. A similar insight was demonstrated in [3] to show that video frames are effective training data to fix coarse manual edits. We then leverage powerful multi-modal large language models (MLLMs), such as GPT-4o [1] and LLaVA [24] to generate descriptive text at or above human-level quality." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 374, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 506 + ], + "type": "text", + "content": "Our data construction starts with extracting video frames that undergo realistic yet complex transformations. Next, we use MLLMs to analyze these transformations and generate precise editing instructions. Additionally, we introduce a novel spatial conditioning strategy: instead of conventional channel concatenation, we concatenate the reference image with the noise input along the spatial dimension. This approach enables the model to access the reference image through cross-attention across the network, greatly enhancing its ability to perform flexible edits while preserving the appearance of the source image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 507, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 627 + ], + "type": "text", + "content": "By fine-tuning pretrained text-to-image models (e.g., Stable Diffusion [32]) on our constructed dataset with spatial conditioning strategy, we enable natural-language-driven image manipulations that were previously challenging. These include adjusting subject poses, rearranging elements, or altering camera perspectives, all while maintaining the integrity and details of the original image (see Fig. 1). Our approach also seamlessly integrates with mask-based guidance for local editing, and also with other control types [27, 43], allowing for versatile modifications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 628, + 266, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 628, + 266, + 639 + ], + "spans": [ + { + "bbox": [ + 67, + 628, + 266, + 639 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 642, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 642, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 642, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 642, + 295, + 677 + ], + "type": "text", + "content": "- We introduce a data construction pipeline that uses video frames and MLLMs to create source-target-instruction triplets for training image manipulation models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 296, + 714 + ], + "type": "text", + "content": "- We introduce a spatial conditioning strategy that conditions the reference image alongside the noise map to improve content preservation while enabling flexible edits." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 377, + 129 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 377, + 129 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 377, + 129 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 377, + 129 + ], + "type": "image", + "image_path": "e140039e28e0c8e774d7413822af56ee66e52cb53dfc83bd44d16a777dcf5768.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 129, + 378, + 136 + ], + "lines": [ + { + "bbox": [ + 319, + 129, + 378, + 136 + ], + "spans": [ + { + "bbox": [ + 319, + 129, + 378, + 136 + ], + "type": "text", + "content": "Put the toy's legs together." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 148, + 555, + 182 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 182 + ], + "type": "text", + "content": "Figure 2. Existing methods struggle with complex edits on real photos, such as non-rigid transformations. They often either fail to follow the editing instructions or produce inconsistent outputs." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 378, + 71, + 436, + 129 + ], + "blocks": [ + { + "bbox": [ + 378, + 71, + 436, + 129 + ], + "lines": [ + { + "bbox": [ + 378, + 71, + 436, + 129 + ], + "spans": [ + { + "bbox": [ + 378, + 71, + 436, + 129 + ], + "type": "image", + "image_path": "3f334bb4adfccb5ca4d059758215446f601bae958d21de2e071ee82ec2b03d1c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 129, + 425, + 136 + ], + "lines": [ + { + "bbox": [ + 389, + 129, + 425, + 136 + ], + "spans": [ + { + "bbox": [ + 389, + 129, + 425, + 136 + ], + "type": "text", + "content": "InstructPix2Pix" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 436, + 71, + 494, + 129 + ], + "blocks": [ + { + "bbox": [ + 436, + 71, + 494, + 129 + ], + "lines": [ + { + "bbox": [ + 436, + 71, + 494, + 129 + ], + "spans": [ + { + "bbox": [ + 436, + 71, + 494, + 129 + ], + "type": "image", + "image_path": "e88f24291cba54cf29b21ae285c78bd73ed96c02ede776ef1d5f1689f812c7e9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 129, + 479, + 136 + ], + "lines": [ + { + "bbox": [ + 451, + 129, + 479, + 136 + ], + "spans": [ + { + "bbox": [ + 451, + 129, + 479, + 136 + ], + "type": "text", + "content": "MagicBrush" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 495, + 71, + 553, + 129 + ], + "blocks": [ + { + "bbox": [ + 495, + 71, + 553, + 129 + ], + "lines": [ + { + "bbox": [ + 495, + 71, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 495, + 71, + 553, + 129 + ], + "type": "image", + "image_path": "fce7696f8aa4f0a655e76491d9b7ce7b87181991db0d8f933ec90b7c1f3d9313.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 129, + 534, + 135 + ], + "lines": [ + { + "bbox": [ + 512, + 129, + 534, + 135 + ], + "spans": [ + { + "bbox": [ + 512, + 129, + 534, + 135 + ], + "type": "text", + "content": "UltraEdit" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 203, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 203, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 203, + 555, + 262 + ], + "type": "text", + "content": "- Our approach enables a wider range of text-based image manipulations, including pose adjustments, element rearrangements, and perspective changes, while also supporting precise, localized edits through integration with additional control mechanisms." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 276, + 400, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 276, + 400, + 289 + ], + "spans": [ + { + "bbox": [ + 314, + 276, + 400, + 289 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 296, + 554, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 554, + 308 + ], + "type": "text", + "content": "2.1. Zero-shot Image Editing with Diffusion Models" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 314, + 554, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 410 + ], + "type": "text", + "content": "Diffusion models [15, 35, 37] have become the leading approach in text-to-image generation due to their ability to effectively translate noise into structured data distributions. Leveraging these pretrained models, several image editing techniques have been developed by manipulating the sampling process [4, 8, 17, 20, 22, 25, 26, 29], modifying the internal architecture of the denoising network [7, 14, 40], or incorporating additional optimization steps [19, 23, 44]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "type": "text", + "content": "For example, SDEdit adds noise to the image and then generates the desired edit by conditioning on a target prompt during the reverse denoising process. Other approaches, such as Prompt-to-Prompt [14], Plug-and-Play [40], and MasaCtrl [7], manipulate the attention mechanism after inverting the image to noise using DDIM inversion [36], allowing for edits that align with the desired prompt. More recent works [9, 13, 18, 26, 41] have refined the inversion process to enhance editing robustness and accuracy. While these zero-shot methods offer flexibility in editing images, they are often too slow and lack robustness, making them less suitable for practical applications." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 563, + 538, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 538, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 538, + 576 + ], + "type": "text", + "content": "2.2. Training Instruction-based Editing Models" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 581, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 554, + 677 + ], + "type": "text", + "content": "Comparing to modifying target image descriptions, a more intuitive and user-friendly editing method is to use direct instructions (e.g., “Have the man look at the camera”) to guide the editing process. InstructPix2Pix [5] pioneered this concept by fine-tuning Stable Diffusion [32] with editing instructions. To handle more complex instructions, methods like MGIE [11] and SmartEdit [16] integrate Multimodal LLMs to improve the understanding of editing instructions." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "A major challenge in training instruction-based editing models is constructing paired image data with corresponding instructions. InstructPix2Pix tackles this by using lan" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2705" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 94, + 102, + 110 + ], + "blocks": [ + { + "bbox": [ + 65, + 94, + 102, + 110 + ], + "lines": [ + { + "bbox": [ + 65, + 94, + 102, + 110 + ], + "spans": [ + { + "bbox": [ + 65, + 94, + 102, + 110 + ], + "type": "image", + "image_path": "0a94d051d110361c181678232f814b24ae75d2a0ce179497a7ae00e0bf408311.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 107, + 80, + 195, + 123 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 187, + 79 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 187, + 79 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 187, + 79 + ], + "type": "text", + "content": "(a) Frame selection: Sample frame pairs from internet videos." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 107, + 80, + 195, + 123 + ], + "lines": [ + { + "bbox": [ + 107, + 80, + 195, + 123 + ], + "spans": [ + { + "bbox": [ + 107, + 80, + 195, + 123 + ], + "type": "image", + "image_path": "99702fb5f7dfcc4e2a6d09418336b4b8c442d640b8e9c2b50e618ea164bf10ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 79, + 250, + 122 + ], + "blocks": [ + { + "bbox": [ + 205, + 79, + 250, + 122 + ], + "lines": [ + { + "bbox": [ + 205, + 79, + 250, + 122 + ], + "spans": [ + { + "bbox": [ + 205, + 79, + 250, + 122 + ], + "type": "image", + "image_path": "ff7d6e9debac14f82b95cc9e3634b44273ff853f832f718ab70a2c9a089567a7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 251, + 79, + 293, + 122 + ], + "blocks": [ + { + "bbox": [ + 251, + 79, + 293, + 122 + ], + "lines": [ + { + "bbox": [ + 251, + 79, + 293, + 122 + ], + "spans": [ + { + "bbox": [ + 251, + 79, + 293, + 122 + ], + "type": "image", + "image_path": "23c9f03916ed4a7b487a74da9fabfe52ad1661f092bc5d203b4e2df067a642ca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 140, + 121, + 178 + ], + "blocks": [ + { + "bbox": [ + 61, + 128, + 116, + 134 + ], + "lines": [ + { + "bbox": [ + 61, + 128, + 116, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 116, + 134 + ], + "type": "text", + "content": "(b) Generate instructions:" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 140, + 121, + 178 + ], + "lines": [ + { + "bbox": [ + 61, + 140, + 121, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 121, + 178 + ], + "type": "image", + "image_path": "f80bd892402f7922c35661a1c28d859ca4edbd6c187374bcdbc36b9f89c3263c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 129, + 141, + 283, + 154 + ], + "blocks": [ + { + "bbox": [ + 129, + 141, + 283, + 154 + ], + "lines": [ + { + "bbox": [ + 129, + 141, + 283, + 154 + ], + "spans": [ + { + "bbox": [ + 129, + 141, + 283, + 154 + ], + "type": "image", + "image_path": "564018e6a8caf1ca93bb8e87c33fad7a4368fd63aa94dbb1c80169428fd53d99.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 167, + 205, + 178 + ], + "lines": [ + { + "bbox": [ + 132, + 167, + 205, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 205, + 178 + ], + "type": "text", + "content": "Adjust the woman's gaze from the book to looking directly at the camera." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 212, + 167, + 286, + 175 + ], + "lines": [ + { + "bbox": [ + 212, + 167, + 286, + 175 + ], + "spans": [ + { + "bbox": [ + 212, + 167, + 286, + 175 + ], + "type": "text", + "content": "Have the man give a thumbs up gesture." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 191, + 135, + 229 + ], + "blocks": [ + { + "bbox": [ + 59, + 185, + 189, + 191 + ], + "lines": [ + { + "bbox": [ + 59, + 185, + 189, + 191 + ], + "spans": [ + { + "bbox": [ + 59, + 185, + 189, + 191 + ], + "type": "text", + "content": "(c) Generated source-target-instruction triplets for training:" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 191, + 135, + 229 + ], + "lines": [ + { + "bbox": [ + 59, + 191, + 135, + 229 + ], + "spans": [ + { + "bbox": [ + 59, + 191, + 135, + 229 + ], + "type": "image", + "image_path": "754e56634d8add88a000ff2bb25be7dfdd335c9d820e9683e8518f9163b40aea.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 230, + 127, + 235 + ], + "lines": [ + { + "bbox": [ + 69, + 230, + 127, + 235 + ], + "spans": [ + { + "bbox": [ + 69, + 230, + 127, + 235 + ], + "type": "text", + "content": "\"Remove the smile from the woman's face.\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 241, + 296, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 241, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 296, + 308 + ], + "type": "text", + "content": "Figure 3. Our data construction pipeline. (a) We begin by sampling suitable frame pairs from videos, ensuring realistic and moderate transformations. (b) These frame pairs are used to prompt Multimodal Large Language Models (MLLMs) to generate detailed editing instructions. (c) This process results in a large-scale dataset with realistic image pairs and precise editing instructions." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 138, + 191, + 214, + 229 + ], + "blocks": [ + { + "bbox": [ + 138, + 191, + 214, + 229 + ], + "lines": [ + { + "bbox": [ + 138, + 191, + 214, + 229 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 214, + 229 + ], + "type": "image", + "image_path": "77bd7bb8f363f7b55723263591af664694625aa193d6313197c63311f99e053a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 230, + 212, + 235 + ], + "lines": [ + { + "bbox": [ + 140, + 230, + 212, + 235 + ], + "spans": [ + { + "bbox": [ + 140, + 230, + 212, + 235 + ], + "type": "text", + "content": "\"Rotate the bird to the left and make its tail visible.\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 217, + 191, + 292, + 229 + ], + "blocks": [ + { + "bbox": [ + 217, + 191, + 292, + 229 + ], + "lines": [ + { + "bbox": [ + 217, + 191, + 292, + 229 + ], + "spans": [ + { + "bbox": [ + 217, + 191, + 292, + 229 + ], + "type": "image", + "image_path": "a5a764da79a68f9e7b2794bddbe68c5d30d43b5e38c80844239894c482feafc7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 230, + 287, + 235 + ], + "lines": [ + { + "bbox": [ + 220, + 230, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 220, + 230, + 287, + 235 + ], + "type": "text", + "content": "\"Make the person looking at the laptop screen.\"" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 323, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 323, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 295, + 441 + ], + "type": "text", + "content": "guage models, such as GPT-3 [6], to generate both an edited image caption and an editing instruction. These captions are converted into images using Prompt-to-Prompt [14]. MagicBrush [42] constructs a small dataset by performing edits using DALL-E 2 [31]. InstructDiffusion [12] augments datasets with synthetic pairs for tasks like object removal and replacement. EmuEdit [34] improves data pair construction using Llama 2 [39] and local masks with Prompt-to-Prompt, combining these with other vision datasets to build a large-scale dataset." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 598 + ], + "type": "text", + "content": "However, in existing large-scale editing datasets, most target images are generated using synthetic methods, which often introduces artifacts and significant appearance deviations from the source images. These datasets are typically limited to high-level edits such as object overlay and style transfer. As a result, models trained on these datasets struggle with complex edits on real photos, such as non-rigid transformations and viewpoint changes. In contrast, our approach leverages real video frames as source and target images, using Multimodal LLMs to generate instructions directly from these frames. This allows for the construction of a more realistic dataset, capturing complex transformations and overcoming the limitations of previous methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 609, + 179, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 179, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 179, + 621 + ], + "type": "text", + "content": "3. Dataset Construction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Unlike existing instruction-based editing datasets, which rely on tuning-free methods to generate target images, we propose a novel data construction pipeline that leverages real video frame pairs. By extracting frame pairs that capture meaningful transformations and using Multimodal Large Language Models (MLLMs) to generate detailed editing instructions, we create a more realistic and high-" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "quality editing dataset. In Sec. 3.1, we outline our approach for extracting suitable frame pairs from videos, followed by Sec. 3.2, where we describe how we use MLLMs to generate precise editing instructions. Finally, Sec. 3.3 provides an in-depth analysis of the resulting dataset. An overview of the data construction pipeline is illustrated in Fig. 3." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 152, + 502, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 502, + 164 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 502, + 164 + ], + "type": "text", + "content": "3.1. Sampling Frame Pairs from Videos" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 169, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 555, + 277 + ], + "type": "text", + "content": "In a video, the same subjects and elements can appear in different ways, move naturally, or be shown from different angles. For example, a person might change their posture, expression, or movement over time. These changes provide useful clues about how a scene or perspective might look with small adjustments, making videos a great source for picking frame pairs to use in image editing tasks. However, it's important to sample the frame pairs with proper criteria to make sure they are suitable for training." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 282, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 554, + 366 + ], + "type": "text", + "content": "Initial sampling of paired frames. We begin by captioning each video to generate an overall description. Videos containing specific keywords (e.g., landscape, abstract, still) that are not suitable for image editing are filtered out. We then sample two frames " + }, + { + "bbox": [ + 313, + 282, + 554, + 366 + ], + "type": "inline_equation", + "content": "(I^s, I^e)" + }, + { + "bbox": [ + 313, + 282, + 554, + 366 + ], + "type": "text", + "content": " with a fixed time interval (i.e., 3-5 seconds) to ensure sufficient transformations occur in the objects of interest." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "text", + "content": "Motion-based filtering. To refine the selection, we apply motion filtering to exclude frames with either too little or too much movement. Specifically, we use RAFT [38] to compute the optical flow " + }, + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "inline_equation", + "content": "O_{I^s \\to I^e}" + }, + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "inline_equation", + "content": "I^s" + }, + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "inline_equation", + "content": "I^e" + }, + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "text", + "content": " and derive the flow magnitude " + }, + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "inline_equation", + "content": "M = \\sqrt{O_x^2 + O_y^2}" + }, + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "text", + "content": ". Frame pairs with moderate movement are retained, while those with excessive or minimal motion are discarded. To further ensure consistency, we assess background changes by computing a background occlusion mask through backward warping of " + }, + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "inline_equation", + "content": "I^e" + }, + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "inline_equation", + "content": "I^s" + }, + { + "bbox": [ + 313, + 372, + 555, + 544 + ], + "type": "text", + "content": " using the estimated flow. By measuring the occlusion mask ratio, we filter out frames with significant background changes (i.e., large masked regions). This motion filtering pipeline allows us to obtain image pairs with realistic, moderate transformations." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 552, + 511, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 511, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 511, + 563 + ], + "type": "text", + "content": "3.2. Instruction Generation with MLLMs" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "text", + "content": "Next, we describe how we generate accurate editing instructions " + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "text", + "content": " given frame pairs " + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "inline_equation", + "content": "(I^s, I^e)" + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "text", + "content": ". Previous approaches [5, 34] rely solely on text-based methods, using Large Language Models (LLMs) to generate instructions and captions of target images purely from the source caption. This approach significantly limits the variety of edits that can be captured in the dataset. As a result, these methods are often constrained to high-level edits, such as adding, removing, or replacing elements, or changing the style." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "To address this limitation, we utilize multimodal LLMs, such as GPT-4o [1] or Pixtral-12B [2], to directly analyze frame pairs and generate editing instructions. Specifically," + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2706" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 150 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 150 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 150 + ], + "type": "table", + "html": "
DatasetsReal Target?# EditsEditing Types
Non-rigidViewpointAdd/RemoveStyle
InstructPix2Pix [5]X0.3MXX✓✓
MagicBrush [42]X10KXX✓✓
EmuEdit [34]X10MXX✓✓
UltraEdit [45]X4.1MXX✓✓
Ours✓6M✓✓XX
", + "image_path": "214dfe6241514c38fd63af0b6cfd1b1b898f82706750b591051dc5298266e883.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 295, + 202 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 202 + ], + "type": "text", + "content": "Table 1. Comparison with existing image-editing datasets. Our dataset is the first large-scale dataset with real target images while supporting complex editing types, such as non-rigid transformations and camera viewpoint adjustments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 213, + 295, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 272 + ], + "type": "text", + "content": "MLLMs are prompted to examine the differences between the source and target images, focusing on changes in subjects, relative positions, camera angles, and background. Based on these observations, the models generate detailed, context-specific instructions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 273, + 295, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 273, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 295, + 380 + ], + "type": "text", + "content": "To improve clarity, we instruct the MLLMs to use absolute terms (e.g., \"Move the bee to the center of the flower\" instead of \"Move the bee to the position in the target image\") and to begin instructions with an action verb (e.g., Change, Move, Adjust) for better usability in real-world applications. Additionally, to ensure high-quality instructions, we allow the MLLMs to reject frame pairs where the edits are too complex to describe accurately, and we exclude such pairs from our dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 388, + 156, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 156, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 156, + 401 + ], + "type": "text", + "content": "3.3. Dataset Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 406, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 406, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 54, + 406, + 295, + 514 + ], + "type": "text", + "content": "In total, after filtering, we sampled 6 million image pairs from videos, each with appropriate levels of motion and transformation. All frames are captioned using Pixtral-12B [2]. The edits captured in our dataset focus on realistic transformations, such as non-rigid changes of subjects/objects, rearranging elements, and altering camera perspectives—types of edits that are largely absent in existing datasets. Examples can be seen in Fig. 3, with comparisons to prior datasets presented in Table 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 514, + 295, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 573 + ], + "type": "text", + "content": "Our dataset can be combined with existing ones to train models that handle both realistic edits introduced in this paper, and the artistic edits explored in previous works (e.g., converting a photo into a painting), as shown in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 574, + 295, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 645 + ], + "type": "text", + "content": "As illustrated in Fig. 3, some video frames may include subtle changes that are not fully captured in the generated instructions. However, our experiments show that these small discrepancies do not significantly impact the training process, as the model focuses on learning the major transformations between frame pairs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 657, + 111, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 111, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 111, + 669 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Using the constructed dataset, we fine-tune a pretrained T2I model for instruction-based image manipulation, enabling edits such as non-rigid transformations of subjects, object" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 553, + 187 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 187 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 187 + ], + "type": "image", + "image_path": "8f37a491d91ff19762ed23456a1a5cb3fe7319f0361dcd2f7319f5aaae16607d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 190, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 313, + 190, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 555, + 301 + ], + "type": "text", + "content": "Figure 4. Overview of the proposed model architecture for instruction-based image editing. The source and target images are first encoded into latent representations " + }, + { + "bbox": [ + 313, + 190, + 555, + 301 + ], + "type": "inline_equation", + "content": "z^s" + }, + { + "bbox": [ + 313, + 190, + 555, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 190, + 555, + 301 + ], + "type": "inline_equation", + "content": "z^e" + }, + { + "bbox": [ + 313, + 190, + 555, + 301 + ], + "type": "text", + "content": " using a pretrained encoder. The target latent " + }, + { + "bbox": [ + 313, + 190, + 555, + 301 + ], + "type": "inline_equation", + "content": "z^e" + }, + { + "bbox": [ + 313, + 190, + 555, + 301 + ], + "type": "text", + "content": " is then transformed into a noisy latent " + }, + { + "bbox": [ + 313, + 190, + 555, + 301 + ], + "type": "inline_equation", + "content": "z_t^e" + }, + { + "bbox": [ + 313, + 190, + 555, + 301 + ], + "type": "text", + "content": " through the forward diffusion process. We concatenate the source image latent and the noisy target latent along the width dimension to form the model input, which is fed into the denoising U-Net " + }, + { + "bbox": [ + 313, + 190, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 190, + 555, + 301 + ], + "type": "text", + "content": " to predict a noise map. The right half of the output, corresponding to the noisy target input, is cropped and compared with the original noise map." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 312, + 555, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 444 + ], + "type": "text", + "content": "repositioning, and changes in viewpoint. Unlike previous methods, in Sec. 4.1 we introduce a novel spatial conditioning approach that conditions the pretrained T2I models on reference images without requiring any architectural modifications. This technique allows instructions to control the target structure while preserving the identity and content of the original image. Since our method does not alter the underlying architecture of the pretrained T2I models, it seamlessly integrates with additional control mechanisms, such as masks or ControlNets [43], to enable even more precise and localized edits (Sec. 4.2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 454, + 553, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 553, + 467 + ], + "type": "text", + "content": "4.1. Finetuning T2I Models for Image Manipulation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "type": "text", + "content": "Existing instruction-based editing models [5, 11, 16, 34] typically condition T2I models on a reference image by concatenating it with the noise map along the channel dimension as input to the denoising network. However, this channel-based conditioning method spatially aligns the reference and target images, which limits the model's flexibility in making structural changes, such as non-rigid transformations or viewpoint adjustments." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 569, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 553, + 605 + ], + "type": "text", + "content": "To address this, we introduce Spatial Conditioning, which conditions the model by concatenating the source image with the noise map along the spatial dimension." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Given a triplet " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "(I^s, I^e, C)" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ", we first encode both images into latent representations " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "z^s" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "z^e" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " using a pretrained variational autoencoder [21]. The target latent " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "z^e" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " is then transformed into a noisy latent " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "z_t^e" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " via the forward diffusion process. We concatenate the source image latent with the noisy target latent along the width dimension to create the model input " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "z_t = \\mathrm{Concat}_{width}([z^s, z_t^e])" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ". This input, which doubles the width of the noisy latent, is fed into the denoising U-Net " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " to predict a noise map of the same width. We then" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2707" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "crop the right half of the output corresponding to the noisy input and compute the denoising loss:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 105, + 295, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 105, + 295, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 105, + 295, + 121 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {E d i t}} = \\mathbb {E} _ {z _ {t}, C, t, \\epsilon} \\left[ \\| \\epsilon - \\operatorname {C r o p} _ {\\text {w i d t h}} \\left(\\epsilon_ {\\theta} \\left(z _ {t}, C, t\\right)\\right) \\| ^ {2} \\right]. \\tag {1}", + "image_path": "a2355222e020d4e490d5d70be0c5acaf4b4b2a13f4e56235050e9b9be17191ef.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 123, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 123, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 296, + 266 + ], + "type": "text", + "content": "This spatial conditioning approach, while seemingly similar to channel conditioning, offers critical advantages for image editing tasks. First, it does not spatially align the reference image with the noisy input, allowing the denoising network to process the condition and the noisy input in parallel. This enables the network to attend to any region of the condition image at each cross-attention layer, facilitating structural changes in the target image. Moreover, since the network has access to the features of the condition image throughout all layers, it can better preserve the details and identity of the source image, leading to more accurate and realistic edits." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 274, + 240, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 240, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 240, + 287 + ], + "type": "text", + "content": "4.2. Incorporating Additional Controls" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 291, + 296, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 296, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 296, + 340 + ], + "type": "text", + "content": "Once trained, our editing model can perform a wide range of edits based on user-provided instructions. However, text instructions alone are often insufficiently precise, making additional controls essential for accurate image editing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "content": "Integration with masks. Similar to zero-shot inpainting in T2I models, our approach supports mask-based localization to edit specific regions of an image. During inference, we blend the updated latent with the reference latent using a mask to control the area of modification:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 414, + 295, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 414, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 104, + 414, + 295, + 428 + ], + "type": "interline_equation", + "content": "z _ {t - 1} ^ {*} = (1 - m) \\cdot z _ {t - 1} ^ {s} + m \\cdot z _ {t - 1}, \\tag {2}", + "image_path": "25ff9b0fcade43391adac4ed4fa3c36a1ad93a1c4410e1675c62d8ce4848d3ff.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 431, + 295, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 295, + 466 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 431, + 295, + 466 + ], + "type": "inline_equation", + "content": "z_{t-1}^{s}" + }, + { + "bbox": [ + 55, + 431, + 295, + 466 + ], + "type": "text", + "content": " is the noisy latent from the forward diffusion of the source image latent " + }, + { + "bbox": [ + 55, + 431, + 295, + 466 + ], + "type": "inline_equation", + "content": "z^{s}" + }, + { + "bbox": [ + 55, + 431, + 295, + 466 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 431, + 295, + 466 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 431, + 295, + 466 + ], + "type": "text", + "content": " is the mask resized to match the latent resolution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 467, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 598 + ], + "type": "text", + "content": "Integration with other spatial controls. For more complex edits, additional visual cues beyond text can greatly enhance control. Users can provide sketches, or adjust key points of a skeleton to modify poses, for example. To enable this, we integrate controllable diffusion models like ControlNet [43] and T2I-Adapter [27]. Crucially, because our model retains the architecture of the original T2I model, we can directly use these control techniques with our trained image-editing models, incorporating spatial maps for finer control. This compatibility is not possible with previous instruction-based editing models like InstructPix2Pix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 609, + 137, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 137, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 137, + 624 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "In Sec. 5.1, we outline the experimental setup, including the evaluation benchmark and quantitative metrics. In Sec. 5.2, we compare with state-of-the-art text-guided image editing methods both quantitatively and qualitatively. Sec. 5.3 demonstrates our model's ability to incorporate additional controls for more precise editing. Finally, we assess the impact of our dataset and conditioning strategy in Sec. 5.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 430, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 430, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 430, + 85 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 97, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 555, + 240 + ], + "type": "text", + "content": "Implementation Details. We fine-tune Stable Diffusion V1.5 [32] with the proposed spatial conditioning strategy on our newly constructed dataset to enable various types of complex image edits. Input images have a resolution of " + }, + { + "bbox": [ + 313, + 97, + 555, + 240 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 97, + 555, + 240 + ], + "type": "text", + "content": ". The model is trained for 100,000 iterations with a constant learning rate of " + }, + { + "bbox": [ + 313, + 97, + 555, + 240 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 97, + 555, + 240 + ], + "type": "text", + "content": ", using the Adam optimizer to update model parameters. Training is conducted on 8 NVIDIA A100 GPUs with a total batch size of 256. During sampling, we employ the DDIM [36] scheduler with 50 steps to generate the final edited images. Note that mask and additional spatial controls are optional; when unspecified, results are generated solely based on textual instructions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 248, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 555, + 453 + ], + "type": "text", + "content": "Evaluation benchmark. Existing instruction-based image editing benchmarks, such as MagicBrush [42] and EmuEdit [34], primarily focus on editing tasks that maintain the original image structure, such as style modification, object replacement, and color or texture changes. These benchmarks are therefore unsuitable for evaluating non-rigid, consistent image editing tasks that alter the image structure while preserving identity and appearance details. To address this gap, we created a specialized test set tailored for the non-rigid editing tasks presented. Our benchmark consists of 50 images, each paired with human-curated editing instructions that specify modifications in pose, expression, viewpoint, shape, position, and other structural transformations. The benchmark excludes any local or global style or appearance edits. Examples of input images and their corresponding instructions are shown in Fig. 5. More details are in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 663 + ], + "type": "text", + "content": "Evaluation metrics. For evaluation, we employ metrics to assess two main aspects of image editing: alignment with text instructions and fidelity to the source image. To evaluate instruction adherence, following previous works, we first generate captions for the target image using a large language model (LLM) based on the source image's caption and the given instructions. We then assess the agreement between caption changes and actual image changes (CLIP-D). However, this metric relies on generating captions for target images, which may not directly measure alignment with the original instructions. To address this, we propose a new metric: we prompt MLLMs to generate instructions based on the input and output images for each method, then compute the CLIP distance between these generated instructions and the original ones, denoted as CLIP-Inst. For evaluating faithfulness, we calculate the CLIP feature distance between the input and output images (CLIP-I)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "We also collect human feedback from 40 participants to evaluate the quality of the edited images. Each participant is shown a subset of 20 examples, including the outputs from each method, alongside the source image and the editing" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2708" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 293, + 183 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 183 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 183 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 183 + ], + "type": "table", + "html": "
MethodCLIP-D ↑CLIP-Inst↑CLIP-I↑
NullTextInversion [26]0.06600.76480.9063
MasaCtrl [7]0.04360.85270.9160
Imagic [19]0.05080.86450.8379
InstructPix2Pix [5]0.08870.85690.9380
MagicBrush [42]0.09720.86480.9318
UltraEdit [45]0.08240.85710.9184
Ours0.13610.87240.9275
", + "image_path": "f0a1aaeae56160944eabb84aef40d511b119435d525c628586640b7201f8ce78.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 226, + 293, + 256 + ], + "blocks": [ + { + "bbox": [ + 55, + 191, + 295, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 295, + 214 + ], + "type": "text", + "content": "Table 2. Quantitative comparison with state-of-the-art text-guided image editing methods on our collected benchmark." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 226, + 293, + 256 + ], + "lines": [ + { + "bbox": [ + 58, + 226, + 293, + 256 + ], + "spans": [ + { + "bbox": [ + 58, + 226, + 293, + 256 + ], + "type": "table", + "html": "
Imagic [19]InstructPix2Pix [5]MagicBrush [42]Ours
Preference5.0%3.25%4.13%87.62%
", + "image_path": "5bf3599b096b29c43b447b8317eb1ea38c0fa26d4e6a87208d0a1cf8e07df6ac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 346 + ], + "type": "text", + "content": "instruction. Participants are asked to choose the image that \"best follows the editing instruction while maintaining the highest quality and preserving the original content.\"" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 354, + 182, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 182, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 182, + 367 + ], + "type": "text", + "content": "5.2. Baseline Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 378, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 295, + 486 + ], + "type": "text", + "content": "Methods for comparison. We compare our proposed method against state-of-the-art text-guided image editing approaches with publicly available models, including zero-shot description-based methods: NullTextInversion [26], MasaCtrl [7], and Imagic [19], as well as instruction-based methods: InstructPix2Pix [5], MagicBrush [42], and UltraEdit [45]. For description-guided methods, we generate the target description by captioning the input image and deriving the description from the instructions using an LLM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 491, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 659 + ], + "type": "text", + "content": "Quantitative results. Table 2 presents the quantitative evaluation of both description- and instruction-based methods on our editing benchmark. A method that perfectly preserves content but fails to follow the editing instructions is still inadequate. For instance, InstructPix2Pix and MagicBrush show higher CLIP-I scores than our approach because they struggle with non-rigid edits and viewpoint changes. As a result, they often leave the input image mostly unchanged (see Fig. 5), leading to high content preservation scores. In contrast, our method achieves the highest scores in instruction alignment while still maintaining strong content preservation. This is further validated by the user study in Table 3, where in over " + }, + { + "bbox": [ + 55, + 491, + 295, + 659 + ], + "type": "inline_equation", + "content": "87.62\\%" + }, + { + "bbox": [ + 55, + 491, + 295, + 659 + ], + "type": "text", + "content": " of cases, users preferred our method over the other four approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Qualitative results. We present a qualitative comparison of various methods in Fig. 5. Since prior approaches are primarily designed for editing tasks that preserve the original image structure, they struggle with non-rigid edits (such" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 553, + 137 + ], + "blocks": [ + { + "bbox": [ + 55, + 263, + 295, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 263, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 295, + 285 + ], + "type": "text", + "content": "Table 3. Human preference results. Percentage indicates how often each method was selected as the best result." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 70, + 553, + 137 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 553, + 137 + ], + "type": "table", + "html": "
MethodCLIP-D↑CLIP-Inst ↑CLIP-I ↑
SC + IP2P data0.12770.84140.9094
CC + Our data0.08530.86790.8552
SC + Our data0.13610.87240.9275
", + "image_path": "4b880a1d00ee27d1de85ad58fdd3a7e2b9141c799aa77fabe90f9cabe416ec8d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 144, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 200 + ], + "type": "text", + "content": "Table 4. Ablation on the training dataset and conditioning approach. Training on the InstructPix2Pix dataset [5] leads to poorer instruction alignment and content preservation. Additionally, our Spatial Conditioning (SC) method outperforms the standard Channel Conditioning (CC) approach." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 211, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 331 + ], + "type": "text", + "content": "as changes in pose and expression), object repositioning, or viewpoint adjustments. These methods often fail to follow such editing instructions, either producing images identical to the source or generating edits that do not align with the specified instructions. Even in cases where the instructions are somewhat followed, they frequently fail to maintain content consistency. In contrast, our method, trained on real video frames with naturalistic transformations, is capable of handling these complex edits while preserving consistency with the input images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 338, + 491, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 491, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 491, + 350 + ], + "type": "text", + "content": "5.3. Results with Additional Controls" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 355, + 556, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 556, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 556, + 487 + ], + "type": "text", + "content": "We demonstrate the incorporation of additional controls for precise editing in Fig. 6. Instructions alone can sometimes be ambiguous regarding which part of the image to edit; in such cases, a mask can be used for localized adjustments (Fig. 6 (a)). In other scenarios, instructions alone may not suffice, especially when precise control is needed for adjustments like altering a subject's pose. Users can address this by directly adjusting control points to define a skeleton pose or providing a rough sketch of the desired composition. We show that our approach can seamlessly integrate ControlNet to utilize such controls for precise editing (Fig. 6 (b))." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 493, + 382, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 382, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 382, + 504 + ], + "type": "text", + "content": "5.4. Ablations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 555, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 571 + ], + "type": "text", + "content": "We argue that our proposed video + MLLM dataset construction pipeline, along with the Spatial Conditioning (SC) approach, enables our model to perform complex image edits while maintaining appearance consistency. We evaluate the impact of each component in Table 4 and Fig. 7." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 576, + 556, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 556, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 556, + 685 + ], + "type": "text", + "content": "Dataset. To demonstrate the importance of using real video frame pairs for training image editing models, we fine-tune a model with the same architecture as ours on the dataset proposed in InstructPix2Pix [5], where target images are synthetically generated using Prompt-to-Prompt [14]. Compared to our model trained on InstructPix2Pix's dataset (denoted as \"SC + IP2P data\"), our method not only better follows editing instructions but also preserves the original image content more effectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Spatial Conditioning. We also fine-tune our model using our dataset, but with the Channel Conditioning (CC)" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2709" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 126, + 148 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 126, + 148 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 126, + 148 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 126, + 148 + ], + "type": "image", + "image_path": "91cacfe3ed31cf4d5b1705a4f881f250717646e5a869be5cf993b9a7a0090971.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 135, + 72, + 202, + 147 + ], + "blocks": [ + { + "bbox": [ + 135, + 72, + 202, + 147 + ], + "lines": [ + { + "bbox": [ + 135, + 72, + 202, + 147 + ], + "spans": [ + { + "bbox": [ + 135, + 72, + 202, + 147 + ], + "type": "image", + "image_path": "f440677de63a429f8cea8c1dea8f381b5201112a0b8067b9677a8c316193e0ac.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 148, + 268, + 156 + ], + "lines": [ + { + "bbox": [ + 138, + 148, + 268, + 156 + ], + "spans": [ + { + "bbox": [ + 138, + 148, + 268, + 156 + ], + "type": "text", + "content": "A bear toy wearing a sweater with arm raised" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 203, + 72, + 271, + 147 + ], + "blocks": [ + { + "bbox": [ + 203, + 72, + 271, + 147 + ], + "lines": [ + { + "bbox": [ + 203, + 72, + 271, + 147 + ], + "spans": [ + { + "bbox": [ + 203, + 72, + 271, + 147 + ], + "type": "image", + "image_path": "d903e7caf616179af1997e7dcc565caf3b1b78e0d84aeffb915d5d1430a8b141.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 157, + 125, + 224 + ], + "blocks": [ + { + "bbox": [ + 57, + 157, + 125, + 224 + ], + "lines": [ + { + "bbox": [ + 57, + 157, + 125, + 224 + ], + "spans": [ + { + "bbox": [ + 57, + 157, + 125, + 224 + ], + "type": "image", + "image_path": "ffd30630ce5c97525268855b530a31448d46a038a0368b2f91b2076b937a63fc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 157, + 202, + 224 + ], + "blocks": [ + { + "bbox": [ + 135, + 157, + 202, + 224 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 202, + 224 + ], + "type": "image", + "image_path": "f6ce94ac85d29e6938ece0e0e41475a0d6eb86f51380907f416aaf440a974897.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 224, + 267, + 232 + ], + "lines": [ + { + "bbox": [ + 138, + 224, + 267, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 267, + 232 + ], + "type": "text", + "content": "A dog sitting on a ledge looking at the camera" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 203, + 157, + 271, + 224 + ], + "blocks": [ + { + "bbox": [ + 203, + 157, + 271, + 224 + ], + "lines": [ + { + "bbox": [ + 203, + 157, + 271, + 224 + ], + "spans": [ + { + "bbox": [ + 203, + 157, + 271, + 224 + ], + "type": "image", + "image_path": "433d728936984dafb5928bb122c8efc9e514027ff7e8fff710c8ae2f02b700d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 233, + 125, + 301 + ], + "blocks": [ + { + "bbox": [ + 58, + 233, + 125, + 301 + ], + "lines": [ + { + "bbox": [ + 58, + 233, + 125, + 301 + ], + "spans": [ + { + "bbox": [ + 58, + 233, + 125, + 301 + ], + "type": "image", + "image_path": "56b40da103c56a33d1f156a4616774885d7777c3bca6917ecc46d2b60e801898.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 135, + 233, + 202, + 300 + ], + "blocks": [ + { + "bbox": [ + 135, + 233, + 202, + 300 + ], + "lines": [ + { + "bbox": [ + 135, + 233, + 202, + 300 + ], + "spans": [ + { + "bbox": [ + 135, + 233, + 202, + 300 + ], + "type": "image", + "image_path": "9a82ed22aace129490f3e54c84e5bd69b36248213831d4d12ecea3eed040c296.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 301, + 249, + 308 + ], + "lines": [ + { + "bbox": [ + 157, + 301, + 249, + 308 + ], + "spans": [ + { + "bbox": [ + 157, + 301, + 249, + 308 + ], + "type": "text", + "content": "A straight tower against the sky" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 203, + 233, + 271, + 300 + ], + "blocks": [ + { + "bbox": [ + 203, + 233, + 271, + 300 + ], + "lines": [ + { + "bbox": [ + 203, + 233, + 271, + 300 + ], + "spans": [ + { + "bbox": [ + 203, + 233, + 271, + 300 + ], + "type": "image", + "image_path": "7347f24ff28b4bdee459c835c1d10b724a00ec845a673927d7ab9201382a8b7e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 58, + 309, + 125, + 377 + ], + "blocks": [ + { + "bbox": [ + 58, + 309, + 125, + 377 + ], + "lines": [ + { + "bbox": [ + 58, + 309, + 125, + 377 + ], + "spans": [ + { + "bbox": [ + 58, + 309, + 125, + 377 + ], + "type": "image", + "image_path": "09fc6ff167cf0b45b2dbf3b5698e4c7802a67f192e9005400fdfefdce5aa6855.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 135, + 309, + 202, + 376 + ], + "blocks": [ + { + "bbox": [ + 135, + 309, + 202, + 376 + ], + "lines": [ + { + "bbox": [ + 135, + 309, + 202, + 376 + ], + "spans": [ + { + "bbox": [ + 135, + 309, + 202, + 376 + ], + "type": "image", + "image_path": "7d7773ca3e36dfcc3a15632bc4b701424d1b5840cb52d3f4f94993890b277603.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 377, + 260, + 384 + ], + "lines": [ + { + "bbox": [ + 146, + 377, + 260, + 384 + ], + "spans": [ + { + "bbox": [ + 146, + 377, + 260, + 384 + ], + "type": "text", + "content": "A man meditating and giving a thumbs up" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 203, + 309, + 271, + 376 + ], + "blocks": [ + { + "bbox": [ + 203, + 309, + 271, + 376 + ], + "lines": [ + { + "bbox": [ + 203, + 309, + 271, + 376 + ], + "spans": [ + { + "bbox": [ + 203, + 309, + 271, + 376 + ], + "type": "image", + "image_path": "df977abb63050a56ddc7b28d8baf2fa1b386a1f3d7bba373604fb67cebc0a2d6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 58, + 385, + 124, + 453 + ], + "blocks": [ + { + "bbox": [ + 58, + 385, + 124, + 453 + ], + "lines": [ + { + "bbox": [ + 58, + 385, + 124, + 453 + ], + "spans": [ + { + "bbox": [ + 58, + 385, + 124, + 453 + ], + "type": "image", + "image_path": "e50f9ae78483a179daab2d8f9fdae63a53ed86ff0410cfad130e10e0df303fb2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 135, + 385, + 202, + 453 + ], + "blocks": [ + { + "bbox": [ + 135, + 385, + 202, + 453 + ], + "lines": [ + { + "bbox": [ + 135, + 385, + 202, + 453 + ], + "spans": [ + { + "bbox": [ + 135, + 385, + 202, + 453 + ], + "type": "image", + "image_path": "e1977b9db7dcf1c57475a88f83e1c6c6c0e74f281f3af7a210de3e2cc185667c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 453, + 268, + 460 + ], + "lines": [ + { + "bbox": [ + 138, + 453, + 268, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 453, + 268, + 460 + ], + "type": "text", + "content": "A woman with blonde braids smiling at camera." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 203, + 385, + 271, + 453 + ], + "blocks": [ + { + "bbox": [ + 203, + 385, + 271, + 453 + ], + "lines": [ + { + "bbox": [ + 203, + 385, + 271, + 453 + ], + "spans": [ + { + "bbox": [ + 203, + 385, + 271, + 453 + ], + "type": "image", + "image_path": "3385714f27dff57efe9bf754708ac6f3b380c3217152fd17e211f71a3c0d6741.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 58, + 461, + 125, + 529 + ], + "blocks": [ + { + "bbox": [ + 58, + 461, + 125, + 529 + ], + "lines": [ + { + "bbox": [ + 58, + 461, + 125, + 529 + ], + "spans": [ + { + "bbox": [ + 58, + 461, + 125, + 529 + ], + "type": "image", + "image_path": "0e4cf758af8fe081e1081f7dc5d74e8000ca2c65e74f473ec458824520d1e0fd.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 135, + 462, + 202, + 529 + ], + "blocks": [ + { + "bbox": [ + 135, + 462, + 202, + 529 + ], + "lines": [ + { + "bbox": [ + 135, + 462, + 202, + 529 + ], + "spans": [ + { + "bbox": [ + 135, + 462, + 202, + 529 + ], + "type": "image", + "image_path": "e121415176c4350688b35f9c8c08d06402d2bcd83b80fb8dfc8de78dff7708a0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 529, + 246, + 536 + ], + "lines": [ + { + "bbox": [ + 160, + 529, + 246, + 536 + ], + "spans": [ + { + "bbox": [ + 160, + 529, + 246, + 536 + ], + "type": "text", + "content": "A brown dog jumping in the air" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 203, + 462, + 271, + 529 + ], + "blocks": [ + { + "bbox": [ + 203, + 462, + 271, + 529 + ], + "lines": [ + { + "bbox": [ + 203, + 462, + 271, + 529 + ], + "spans": [ + { + "bbox": [ + 203, + 462, + 271, + 529 + ], + "type": "image", + "image_path": "51673d34f7e67828b59880980327f9253c3c1cc6150259c9b4fd57b5ac8bf5cc.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 58, + 538, + 124, + 605 + ], + "blocks": [ + { + "bbox": [ + 58, + 538, + 124, + 605 + ], + "lines": [ + { + "bbox": [ + 58, + 538, + 124, + 605 + ], + "spans": [ + { + "bbox": [ + 58, + 538, + 124, + 605 + ], + "type": "image", + "image_path": "c10f52907241c2000243e3bd1a7cb92c1c3e28c1796b80988e836f989b7b5ca1.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 628, + 555, + 684 + ], + "lines": [ + { + "bbox": [ + 55, + 628, + 555, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 555, + 684 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison with state-of-the-art image editing methods, including both description-based and instruction-based approaches. Existing methods struggle with complex edits such as non-rigid transformations (e.g., changes in pose and expression), object repositioning, or viewpoint adjustments. They often either fail to follow the editing instructions or produce images with inconsistencies, such as identity shifts. In contrast, our method, trained on real video frames with naturalistic transformations, successfully handles these edits while maintaining consistency with the original input images." + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 135, + 538, + 202, + 605 + ], + "blocks": [ + { + "bbox": [ + 135, + 538, + 202, + 605 + ], + "lines": [ + { + "bbox": [ + 135, + 538, + 202, + 605 + ], + "spans": [ + { + "bbox": [ + 135, + 538, + 202, + 605 + ], + "type": "image", + "image_path": "462d286b69b1a02d6707d3b1ca48038c3c6db8af7de42f849b598ad044be3fae.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 607, + 254, + 614 + ], + "lines": [ + { + "bbox": [ + 151, + 607, + 254, + 614 + ], + "spans": [ + { + "bbox": [ + 151, + 607, + 254, + 614 + ], + "type": "text", + "content": "A home office, viewed from the left" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 203, + 538, + 271, + 605 + ], + "blocks": [ + { + "bbox": [ + 203, + 538, + 271, + 605 + ], + "lines": [ + { + "bbox": [ + 203, + 538, + 271, + 605 + ], + "spans": [ + { + "bbox": [ + 203, + 538, + 271, + 605 + ], + "type": "image", + "image_path": "5ec98dc2cad1b21476df5658cb2df93ed8507037f3eb57b4f87c94997a3d55ac.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 280, + 82, + 347, + 147 + ], + "blocks": [ + { + "bbox": [ + 279, + 72, + 338, + 80 + ], + "lines": [ + { + "bbox": [ + 279, + 72, + 338, + 80 + ], + "spans": [ + { + "bbox": [ + 279, + 72, + 338, + 80 + ], + "type": "text", + "content": "InstructPix2Pix" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 280, + 82, + 347, + 147 + ], + "lines": [ + { + "bbox": [ + 280, + 82, + 347, + 147 + ], + "spans": [ + { + "bbox": [ + 280, + 82, + 347, + 147 + ], + "type": "image", + "image_path": "75b104fc8d4b1779f91273a2fe1304860e95afcce2fbcb0c2b82cdee28504780.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 349, + 82, + 415, + 147 + ], + "blocks": [ + { + "bbox": [ + 364, + 73, + 400, + 80 + ], + "lines": [ + { + "bbox": [ + 364, + 73, + 400, + 80 + ], + "spans": [ + { + "bbox": [ + 364, + 73, + 400, + 80 + ], + "type": "text", + "content": "MagicBrush" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 349, + 82, + 415, + 147 + ], + "lines": [ + { + "bbox": [ + 349, + 82, + 415, + 147 + ], + "spans": [ + { + "bbox": [ + 349, + 82, + 415, + 147 + ], + "type": "image", + "image_path": "3466b4087e60390eae434a62ff7ee109cbc107aa668f807e677b2959bb68a160.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 148, + 452, + 155 + ], + "lines": [ + { + "bbox": [ + 380, + 148, + 452, + 155 + ], + "spans": [ + { + "bbox": [ + 380, + 148, + 452, + 155 + ], + "type": "text", + "content": "Raise the bear doll's arm." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 417, + 82, + 484, + 147 + ], + "blocks": [ + { + "bbox": [ + 436, + 73, + 464, + 80 + ], + "lines": [ + { + "bbox": [ + 436, + 73, + 464, + 80 + ], + "spans": [ + { + "bbox": [ + 436, + 73, + 464, + 80 + ], + "type": "text", + "content": "UltraEdit" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 82, + 484, + 147 + ], + "lines": [ + { + "bbox": [ + 417, + 82, + 484, + 147 + ], + "spans": [ + { + "bbox": [ + 417, + 82, + 484, + 147 + ], + "type": "image", + "image_path": "5c941c4dbe61749aa24e27487c0ca8c1c3af30df7f915b813d160517f264411e.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 485, + 82, + 552, + 147 + ], + "blocks": [ + { + "bbox": [ + 511, + 73, + 526, + 80 + ], + "lines": [ + { + "bbox": [ + 511, + 73, + 526, + 80 + ], + "spans": [ + { + "bbox": [ + 511, + 73, + 526, + 80 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 485, + 82, + 552, + 147 + ], + "lines": [ + { + "bbox": [ + 485, + 82, + 552, + 147 + ], + "spans": [ + { + "bbox": [ + 485, + 82, + 552, + 147 + ], + "type": "image", + "image_path": "e3739f64bc9621d60f34a99cb6e83ddbe626e9163d4eac6ea21da37aeeba543b.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 281, + 157, + 347, + 224 + ], + "blocks": [ + { + "bbox": [ + 281, + 157, + 347, + 224 + ], + "lines": [ + { + "bbox": [ + 281, + 157, + 347, + 224 + ], + "spans": [ + { + "bbox": [ + 281, + 157, + 347, + 224 + ], + "type": "image", + "image_path": "cbc76734170b8fbcf023e88c8887b8cc1c0970c2efae808c2062f7a3c8adddf1.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 349, + 157, + 416, + 224 + ], + "blocks": [ + { + "bbox": [ + 349, + 157, + 416, + 224 + ], + "lines": [ + { + "bbox": [ + 349, + 157, + 416, + 224 + ], + "spans": [ + { + "bbox": [ + 349, + 157, + 416, + 224 + ], + "type": "image", + "image_path": "bb4a6988d62e144719af3bd0643400b3ade30b8cecc42f398ca8578854b60fd7.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 224, + 462, + 232 + ], + "lines": [ + { + "bbox": [ + 369, + 224, + 462, + 232 + ], + "spans": [ + { + "bbox": [ + 369, + 224, + 462, + 232 + ], + "type": "text", + "content": "Make the dog look at the camera" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 417, + 157, + 484, + 224 + ], + "blocks": [ + { + "bbox": [ + 417, + 157, + 484, + 224 + ], + "lines": [ + { + "bbox": [ + 417, + 157, + 484, + 224 + ], + "spans": [ + { + "bbox": [ + 417, + 157, + 484, + 224 + ], + "type": "image", + "image_path": "79f211ee573eda276b6daf7d78c9b7231a41f7f6743b7c196ac09c6f2f1c2645.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 485, + 157, + 552, + 224 + ], + "blocks": [ + { + "bbox": [ + 485, + 157, + 552, + 224 + ], + "lines": [ + { + "bbox": [ + 485, + 157, + 552, + 224 + ], + "spans": [ + { + "bbox": [ + 485, + 157, + 552, + 224 + ], + "type": "image", + "image_path": "de8da886337501fb3feca3903078b04a37686b15fc7e029bedf2a3f15df31233.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 280, + 233, + 347, + 300 + ], + "blocks": [ + { + "bbox": [ + 280, + 233, + 347, + 300 + ], + "lines": [ + { + "bbox": [ + 280, + 233, + 347, + 300 + ], + "spans": [ + { + "bbox": [ + 280, + 233, + 347, + 300 + ], + "type": "image", + "image_path": "083bbac3e50923e2e50f6130325becfd0131ae613ffc041d733ed56fd47fe3e7.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 349, + 233, + 416, + 300 + ], + "blocks": [ + { + "bbox": [ + 349, + 233, + 416, + 300 + ], + "lines": [ + { + "bbox": [ + 349, + 233, + 416, + 300 + ], + "spans": [ + { + "bbox": [ + 349, + 233, + 416, + 300 + ], + "type": "image", + "image_path": "07ade9e662f26855a26cbcee48c5da85d78ea04932a8e8cdc062400e33bf1df1.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 300, + 454, + 308 + ], + "lines": [ + { + "bbox": [ + 378, + 300, + 454, + 308 + ], + "spans": [ + { + "bbox": [ + 378, + 300, + 454, + 308 + ], + "type": "text", + "content": "Make the building straight" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 417, + 233, + 484, + 300 + ], + "blocks": [ + { + "bbox": [ + 417, + 233, + 484, + 300 + ], + "lines": [ + { + "bbox": [ + 417, + 233, + 484, + 300 + ], + "spans": [ + { + "bbox": [ + 417, + 233, + 484, + 300 + ], + "type": "image", + "image_path": "cf4b1f3c544180ad5995c49bc604eeb8a0d5c5be21209c8458d81a3b8e1cf02a.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 485, + 233, + 552, + 300 + ], + "blocks": [ + { + "bbox": [ + 485, + 233, + 552, + 300 + ], + "lines": [ + { + "bbox": [ + 485, + 233, + 552, + 300 + ], + "spans": [ + { + "bbox": [ + 485, + 233, + 552, + 300 + ], + "type": "image", + "image_path": "e33fce7a1d6d356d0a9a56a820f886ef28239ec2bae7fb7090bd139ff40933fb.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 281, + 309, + 348, + 376 + ], + "blocks": [ + { + "bbox": [ + 281, + 309, + 348, + 376 + ], + "lines": [ + { + "bbox": [ + 281, + 309, + 348, + 376 + ], + "spans": [ + { + "bbox": [ + 281, + 309, + 348, + 376 + ], + "type": "image", + "image_path": "c3f3a404b919828ebf83709039220bce10d30cd47d4e3bafd125184facb6411d.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 349, + 309, + 416, + 376 + ], + "blocks": [ + { + "bbox": [ + 349, + 309, + 416, + 376 + ], + "lines": [ + { + "bbox": [ + 349, + 309, + 416, + 376 + ], + "spans": [ + { + "bbox": [ + 349, + 309, + 416, + 376 + ], + "type": "image", + "image_path": "46c9da0f00d6894ead964e35dd1bb8f2eb2d41fa120b1207ca5afbede10e042d.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 377, + 460, + 384 + ], + "lines": [ + { + "bbox": [ + 372, + 377, + 460, + 384 + ], + "spans": [ + { + "bbox": [ + 372, + 377, + 460, + 384 + ], + "type": "text", + "content": "Have the man give a thumbs-up" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 417, + 309, + 484, + 376 + ], + "blocks": [ + { + "bbox": [ + 417, + 309, + 484, + 376 + ], + "lines": [ + { + "bbox": [ + 417, + 309, + 484, + 376 + ], + "spans": [ + { + "bbox": [ + 417, + 309, + 484, + 376 + ], + "type": "image", + "image_path": "141851b06ac8d87c308211aa65a80d868a141fe97e21edef2dec16f50ea71380.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 485, + 309, + 552, + 376 + ], + "blocks": [ + { + "bbox": [ + 485, + 309, + 552, + 376 + ], + "lines": [ + { + "bbox": [ + 485, + 309, + 552, + 376 + ], + "spans": [ + { + "bbox": [ + 485, + 309, + 552, + 376 + ], + "type": "image", + "image_path": "459387dc18997e3249b56fe07a87e29d70705506d05ee84d74f231730e9f51ec.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 281, + 386, + 348, + 453 + ], + "blocks": [ + { + "bbox": [ + 281, + 386, + 348, + 453 + ], + "lines": [ + { + "bbox": [ + 281, + 386, + 348, + 453 + ], + "spans": [ + { + "bbox": [ + 281, + 386, + 348, + 453 + ], + "type": "image", + "image_path": "2e17669b5536ff5720a470d00c854d25c792943b23f12e32bff5af515826cf2a.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 349, + 386, + 416, + 453 + ], + "blocks": [ + { + "bbox": [ + 349, + 386, + 416, + 453 + ], + "lines": [ + { + "bbox": [ + 349, + 386, + 416, + 453 + ], + "spans": [ + { + "bbox": [ + 349, + 386, + 416, + 453 + ], + "type": "image", + "image_path": "c9a47a3aed6288d61566178cfeebe8e98b69f0d04890c8187eb379ef864a954a.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 453, + 464, + 460 + ], + "lines": [ + { + "bbox": [ + 369, + 453, + 464, + 460 + ], + "spans": [ + { + "bbox": [ + 369, + 453, + 464, + 460 + ], + "type": "text", + "content": "Make the girl smile at the camera" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 417, + 386, + 484, + 453 + ], + "blocks": [ + { + "bbox": [ + 417, + 386, + 484, + 453 + ], + "lines": [ + { + "bbox": [ + 417, + 386, + 484, + 453 + ], + "spans": [ + { + "bbox": [ + 417, + 386, + 484, + 453 + ], + "type": "image", + "image_path": "a60ddb8e718980f52377fdfac40c892df93cc65ddc9a0688cd44ecf1c6c5ff71.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 485, + 386, + 552, + 453 + ], + "blocks": [ + { + "bbox": [ + 485, + 386, + 552, + 453 + ], + "lines": [ + { + "bbox": [ + 485, + 386, + 552, + 453 + ], + "spans": [ + { + "bbox": [ + 485, + 386, + 552, + 453 + ], + "type": "image", + "image_path": "5823349de2693976a327192d7f05ba0b3cbce849e60ac0e7591e6d16b74325f7.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 281, + 462, + 348, + 529 + ], + "blocks": [ + { + "bbox": [ + 281, + 462, + 348, + 529 + ], + "lines": [ + { + "bbox": [ + 281, + 462, + 348, + 529 + ], + "spans": [ + { + "bbox": [ + 281, + 462, + 348, + 529 + ], + "type": "image", + "image_path": "c60eb66624b317f02e212292f8c7aa215142787c2a989991327826f018807320.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 349, + 462, + 416, + 529 + ], + "blocks": [ + { + "bbox": [ + 349, + 462, + 416, + 529 + ], + "lines": [ + { + "bbox": [ + 349, + 462, + 416, + 529 + ], + "spans": [ + { + "bbox": [ + 349, + 462, + 416, + 529 + ], + "type": "image", + "image_path": "693268b8d5cb7f23dd9169200df79263817b1d03982ec269e1c7fc6cadbc79f0.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 529, + 457, + 536 + ], + "lines": [ + { + "bbox": [ + 375, + 529, + 457, + 536 + ], + "spans": [ + { + "bbox": [ + 375, + 529, + 457, + 536 + ], + "type": "text", + "content": "Make the dog jump in the air" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 417, + 462, + 484, + 529 + ], + "blocks": [ + { + "bbox": [ + 417, + 462, + 484, + 529 + ], + "lines": [ + { + "bbox": [ + 417, + 462, + 484, + 529 + ], + "spans": [ + { + "bbox": [ + 417, + 462, + 484, + 529 + ], + "type": "image", + "image_path": "7e59dc1877d9ed207f6b700ecbdcc902c5eef36749dc36899bcafc43ad336900.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 485, + 462, + 552, + 529 + ], + "blocks": [ + { + "bbox": [ + 485, + 462, + 552, + 529 + ], + "lines": [ + { + "bbox": [ + 485, + 462, + 552, + 529 + ], + "spans": [ + { + "bbox": [ + 485, + 462, + 552, + 529 + ], + "type": "image", + "image_path": "156b9f24f58bfae6178e0b120568ed2d3f541d264371ff60d266911f7f1eb00c.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 281, + 538, + 348, + 605 + ], + "blocks": [ + { + "bbox": [ + 281, + 538, + 348, + 605 + ], + "lines": [ + { + "bbox": [ + 281, + 538, + 348, + 605 + ], + "spans": [ + { + "bbox": [ + 281, + 538, + 348, + 605 + ], + "type": "image", + "image_path": "583fed326c5485b24f6e57cc836dcd43adefd5114a199fa2f6c345ffdb1f30dd.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 349, + 538, + 416, + 605 + ], + "blocks": [ + { + "bbox": [ + 349, + 538, + 416, + 605 + ], + "lines": [ + { + "bbox": [ + 349, + 538, + 416, + 605 + ], + "spans": [ + { + "bbox": [ + 349, + 538, + 416, + 605 + ], + "type": "image", + "image_path": "98eda8ae1494c9ddc2a8d40fba797c5dfd665b0c01cadc145156c1f5bb2f07b7.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 606, + 457, + 613 + ], + "lines": [ + { + "bbox": [ + 372, + 606, + 457, + 613 + ], + "spans": [ + { + "bbox": [ + 372, + 606, + 457, + 613 + ], + "type": "text", + "content": "Move the camera to the right" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_caption" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 417, + 538, + 484, + 605 + ], + "blocks": [ + { + "bbox": [ + 417, + 538, + 484, + 605 + ], + "lines": [ + { + "bbox": [ + 417, + 538, + 484, + 605 + ], + "spans": [ + { + "bbox": [ + 417, + 538, + 484, + 605 + ], + "type": "image", + "image_path": "7baf980a2b2f5fde432f67c01ff754db57f1ac2c11f2d8e1b09fe7fbec16ceaf.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 485, + 538, + 552, + 605 + ], + "blocks": [ + { + "bbox": [ + 485, + 538, + 552, + 605 + ], + "lines": [ + { + "bbox": [ + 485, + 538, + 552, + 605 + ], + "spans": [ + { + "bbox": [ + 485, + 538, + 552, + 605 + ], + "type": "image", + "image_path": "0eef6f7683b4b64d65484b00340c4b0caf24a02990c3b40020651216f09a2cff.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + } + ], + "index": 65 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2710" + } + ] + } + ], + "index": 68 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 301, + 248 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 301, + 248 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 301, + 248 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 301, + 248 + ], + "type": "image", + "image_path": "ca76addccb123eaaed823414d5bcaafc3127279df901d3dadb05cd0c42e61604.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 248, + 247, + 255 + ], + "lines": [ + { + "bbox": [ + 192, + 248, + 247, + 255 + ], + "spans": [ + { + "bbox": [ + 192, + 248, + 247, + 255 + ], + "type": "text", + "content": "Make the girl look up" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 259, + 216, + 267 + ], + "lines": [ + { + "bbox": [ + 141, + 259, + 216, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 259, + 216, + 267 + ], + "type": "text", + "content": "(a) Localized image editing" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 553, + 248 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 553, + 248 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 553, + 248 + ], + "type": "image", + "image_path": "7922f8ed3b9bc4f85d4f57638d5bc80f32cdd7605fcdfbd5f8e1a62b344a66d5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 249, + 544, + 256 + ], + "lines": [ + { + "bbox": [ + 398, + 249, + 544, + 256 + ], + "spans": [ + { + "bbox": [ + 398, + 249, + 544, + 256 + ], + "type": "text", + "content": "Rotate the banana 45 degrees anti-clockwise. + Scribble" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 384, + 259, + 478, + 266 + ], + "lines": [ + { + "bbox": [ + 384, + 259, + 478, + 266 + ], + "spans": [ + { + "bbox": [ + 384, + 259, + 478, + 266 + ], + "type": "text", + "content": "(b) Results with additional controls" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 330, + 117, + 390 + ], + "blocks": [ + { + "bbox": [ + 55, + 273, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 273, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 555, + 316 + ], + "type": "text", + "content": "Figure 6. Qualitative results of our method with additional controls. (a) Our model can utilize a mask to specify which part of the image to edit, enabling localized adjustments and resolving ambiguities in the instructions. (b) When combined with ControlNet, our model can accept additional inputs, such as human poses or rough sketches, to achieve precise edits in subject poses or object positioning. This level of control is not possible with previous methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 330, + 117, + 390 + ], + "lines": [ + { + "bbox": [ + 57, + 330, + 117, + 390 + ], + "spans": [ + { + "bbox": [ + 57, + 330, + 117, + 390 + ], + "type": "image", + "image_path": "6e2f30d9c498906001d218aeaf9d72f56df715ffcbbb3c06e5a2ebe305e8f93d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 391, + 111, + 396 + ], + "lines": [ + { + "bbox": [ + 62, + 391, + 111, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 111, + 396 + ], + "type": "text", + "content": "Make the woman smile" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 396, + 103, + 402 + ], + "lines": [ + { + "bbox": [ + 71, + 396, + 103, + 402 + ], + "spans": [ + { + "bbox": [ + 71, + 396, + 103, + 402 + ], + "type": "text", + "content": "at the camera" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 117, + 330, + 177, + 390 + ], + "blocks": [ + { + "bbox": [ + 117, + 330, + 177, + 390 + ], + "lines": [ + { + "bbox": [ + 117, + 330, + 177, + 390 + ], + "spans": [ + { + "bbox": [ + 117, + 330, + 177, + 390 + ], + "type": "image", + "image_path": "77cec6bb74a0a99c5a0447086cb9367dda8bace4db1c667ae324afb8f71d9cc7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 391, + 164, + 396 + ], + "lines": [ + { + "bbox": [ + 127, + 391, + 164, + 396 + ], + "spans": [ + { + "bbox": [ + 127, + 391, + 164, + 396 + ], + "type": "text", + "content": "SC + IP2P data" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 178, + 330, + 236, + 390 + ], + "blocks": [ + { + "bbox": [ + 178, + 330, + 236, + 390 + ], + "lines": [ + { + "bbox": [ + 178, + 330, + 236, + 390 + ], + "spans": [ + { + "bbox": [ + 178, + 330, + 236, + 390 + ], + "type": "image", + "image_path": "cd3bfc7762341cfa7f5fb152a1ef38278047c583c10f28f1aaea05034d0e7120.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 391, + 223, + 396 + ], + "lines": [ + { + "bbox": [ + 187, + 391, + 223, + 396 + ], + "spans": [ + { + "bbox": [ + 187, + 391, + 223, + 396 + ], + "type": "text", + "content": "CC + Our data" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 236, + 330, + 296, + 390 + ], + "blocks": [ + { + "bbox": [ + 236, + 330, + 296, + 390 + ], + "lines": [ + { + "bbox": [ + 236, + 330, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 236, + 330, + 296, + 390 + ], + "type": "image", + "image_path": "7f8bf1c09c1f9e9451aced58975f0516d09fca58f9093a3d0ff267064afd5fe0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 391, + 283, + 396 + ], + "lines": [ + { + "bbox": [ + 246, + 391, + 283, + 396 + ], + "spans": [ + { + "bbox": [ + 246, + 391, + 283, + 396 + ], + "type": "text", + "content": "SC + Ours data" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 57, + 506, + 178, + 567 + ], + "blocks": [ + { + "bbox": [ + 55, + 418, + 296, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 483 + ], + "type": "text", + "content": "Figure 7. Qualitative ablation on the training dataset and conditioning approach. Compared to models trained on the Instruct-Pix2Pix dataset (IP2P) [5] or using the standard Channel Conditioning (CC) method, our model trained on our dataset with the Spatial Conditioning (SC) approach demonstrates superior instruction alignment and content preservation." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 506, + 178, + 567 + ], + "lines": [ + { + "bbox": [ + 57, + 506, + 178, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 506, + 178, + 567 + ], + "type": "image", + "image_path": "80e8596f28541697564e9aa8b956f3e9815159d0ffcd2501ee3748628a64829c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 567, + 100, + 573 + ], + "lines": [ + { + "bbox": [ + 74, + 567, + 100, + 573 + ], + "spans": [ + { + "bbox": [ + 74, + 567, + 100, + 573 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 123, + 567, + 170, + 572 + ], + "lines": [ + { + "bbox": [ + 123, + 567, + 170, + 572 + ], + "spans": [ + { + "bbox": [ + 123, + 567, + 170, + 572 + ], + "type": "text", + "content": "Make them look to the side" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 585, + 295, + 617 + ], + "lines": [ + { + "bbox": [ + 55, + 585, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 295, + 617 + ], + "type": "text", + "content": "Figure 8. Limitations. Left: Occasional unintended slight viewpoint changes. Right: Inability to accurately isolate specific objects based on instructions." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 117, + 506, + 176, + 566 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 506, + 176, + 566 + ], + "spans": [ + { + "bbox": [ + 117, + 506, + 176, + 566 + ], + "type": "image", + "image_path": "d62b019dc677522bb27ced2da330bb5959c98aec2093d20d9b8919e797bb95c2.jpg" + } + ] + } + ], + "index": 19, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 184, + 512, + 229, + 565 + ], + "blocks": [ + { + "bbox": [ + 184, + 512, + 229, + 565 + ], + "lines": [ + { + "bbox": [ + 184, + 512, + 229, + 565 + ], + "spans": [ + { + "bbox": [ + 184, + 512, + 229, + 565 + ], + "type": "image", + "image_path": "18a83c05ecaa656cd804d83d7ce9612680548a1feecf4e264b6f441c437ea115.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 567, + 219, + 573 + ], + "lines": [ + { + "bbox": [ + 192, + 567, + 219, + 573 + ], + "spans": [ + { + "bbox": [ + 192, + 567, + 219, + 573 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 240, + 513, + 288, + 566 + ], + "blocks": [ + { + "bbox": [ + 240, + 513, + 288, + 566 + ], + "lines": [ + { + "bbox": [ + 240, + 513, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 240, + 513, + 288, + 566 + ], + "type": "image", + "image_path": "f1da111493e24878a20c8be53481520429c75af2cd9cdc4672b4000645814bfe.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 567, + 289, + 573 + ], + "lines": [ + { + "bbox": [ + 241, + 567, + 289, + 573 + ], + "spans": [ + { + "bbox": [ + 241, + 567, + 289, + 573 + ], + "type": "text", + "content": "Lower the taller dog's head" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 645, + 296, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 296, + 717 + ], + "type": "text", + "content": "method, where the reference latent is concatenated with the target latent along the channel dimension rather than the spatial dimension. As shown in Table 4 and Fig. 7, the proposed Spatial Conditioning approach significantly improves the model's ability to apply complex edits while achieving higher quality and better content preservation." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 314, + 331, + 384, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 331, + 384, + 342 + ], + "spans": [ + { + "bbox": [ + 314, + 331, + 384, + 342 + ], + "type": "text", + "content": "6. Limitation" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 346, + 555, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 549 + ], + "type": "text", + "content": "Our approach introduces a novel data construction pipeline, but it still has some limitations. The quality of the dataset is influenced by the frame filtering process and the capabilities of the MLLMs. Occasionally, MLLMs may generate inaccurate instructions or fail to capture all transformations between frames, leading our editing model to apply unintended changes, such as slight viewpoint shifts that deviate from the original instructions. The effectiveness of our model also depends on the accuracy of training instructions and the comprehension abilities of pretrained T2I models, which can affect its performance on complex edits. Examples of these limitations are shown in Fig. 8. Moreover, since the transformations captured from video frames are restricted to realistic changes, our model cannot handle artistic edits like style transfer or object replacement. This limitation can be mitigated by integrating our dataset with others, as demonstrated in the supplementary material." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 314, + 552, + 388, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 388, + 564 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 388, + 564 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We present an approach for sampling video frames and utilizing MLLMs to generate editing instructions for training instruction-based image manipulation models. Unlike existing datasets, which rely on synthetically generated target images, our method leverages supervision signals from videos and MLLMs to support complex edits, such as nonrigid transformations and viewpoint changes, while preserving content consistency. Future work could focus on refining filtering techniques, either by improving MLLMs or incorporating human-in-the-loop processes, as well as integrating video data with other datasets to further enhance image editing capabilities." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "2711" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 180, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 296, + 236 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 236, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 236, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 61, + 236, + 296, + 290 + ], + "type": "text", + "content": "[2] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Devendra Chaplot, Jessica Chudnovsky, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliiou, Paul Jacob, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 292, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 292, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 292, + 295, + 335 + ], + "type": "text", + "content": "[3] Hadi Alzayer, Zhihao Xia, Xuaner Zhang, Eli Shechtman, Jia-Bin Huang, and Michael Gharbi. Magic fixup: Streamlining photo editing by watching dynamic videos. arXiv preprint arXiv:2403.13044, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 336, + 296, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 336, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 61, + 336, + 296, + 401 + ], + "type": "text", + "content": "[4] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 456 + ], + "type": "text", + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 294, + 479 + ], + "type": "text", + "content": "[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 481, + 296, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 296, + 545 + ], + "type": "text", + "content": "[7] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 547, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 296, + 590 + ], + "type": "text", + "content": "[8] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 591, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 296, + 635 + ], + "type": "text", + "content": "[9] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. arXiv preprint arXiv:2408.00735, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "type": "text", + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[11] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 2, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "type": "text", + "content": "We would like to thank Nanxuan Zhao and Yilin Wang for their valuable feedback. This research was supported in part by JSPS KAKENHI with Grant Numbers 24KK0209, 24K22318, 22H00529, and JST-Mirai Program JPMJMI23G1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "type": "text", + "content": "[12] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 206 + ], + "type": "text", + "content": "[13] Ligong Han, Song Wen, Qi Chen, Zhixing Zhang, Kunpeng Song, Mengwei Ren, Ruijiang Gao, Anastasis Stathopoulos, Xiaoxiao He, Yuxiao Chen, et al. Proxedit: Improving tuning-free real image editing with proximal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4291-4301, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 207, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 207, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 207, + 554, + 251 + ], + "type": "text", + "content": "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 252, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 252, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 252, + 554, + 285 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 286, + 554, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 286, + 554, + 362 + ], + "spans": [ + { + "bbox": [ + 317, + 286, + 554, + 362 + ], + "type": "text", + "content": "[16] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 365, + 554, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 554, + 418 + ], + "type": "text", + "content": "[17] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 421, + 554, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 554, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 554, + 463 + ], + "type": "text", + "content": "[18] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 465, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 465, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 465, + 554, + 521 + ], + "type": "text", + "content": "[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 522, + 554, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 522, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 554, + 576 + ], + "type": "text", + "content": "[20] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 578, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 554, + 600 + ], + "type": "text", + "content": "[21] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 633 + ], + "type": "text", + "content": "[22] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 554, + 679 + ], + "type": "text", + "content": "[23] Pengzhi Li, Qinxuan Huang, Yikang Ding, and Zhiheng Li. Layerdiffusion: Layered controlled image editing with diffusion models. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2712" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "type": "text", + "content": "[26] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "type": "text", + "content": "[27] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "text", + "content": "[28] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "type": "text", + "content": "[29] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH Conference Proceedings, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 394, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 436 + ], + "type": "text", + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 558 + ], + "type": "text", + "content": "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 623 + ], + "type": "text", + "content": "[34] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871-8879, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 625, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 677 + ], + "type": "text", + "content": "[35] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 474 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[37] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "text", + "content": "[38] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision, pages 402-419. Springer, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[40] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "type": "text", + "content": "[41] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. Turboedit: Instant text-based image editing. arXiv preprint arXiv:2408.08332, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "text", + "content": "[42] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "text", + "content": "[43] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 387, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 419 + ], + "type": "text", + "content": "[44] Shiwen Zhang, Shuai Xiao, and Weilin Huang. Forgedit: Text guided image editing via learning and forgetting. arXiv preprint arXiv:2309.10556, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 421, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 474 + ], + "type": "text", + "content": "[45] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. 4, 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2713" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Integral Fast Fourier Color Constancy/9fd420e6-b4fa-48bd-940e-f8dcc99e9a7d_content_list.json b/2025/Integral Fast Fourier Color Constancy/9fd420e6-b4fa-48bd-940e-f8dcc99e9a7d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..79fd18a7a0c9440f8551d0ad3efe7d490b3b8691 --- /dev/null +++ b/2025/Integral Fast Fourier Color Constancy/9fd420e6-b4fa-48bd-940e-f8dcc99e9a7d_content_list.json @@ -0,0 +1,2088 @@ +[ + { + "type": "text", + "text": "Integral Fast Fourier Color Constancy", + "text_level": 1, + "bbox": [ + 302, + 130, + 694, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenjun Wei $^{1}$ * Yanlin Qian $^{2}$ * Huaian Chen $^{1\\dagger}$ Junkang Dai $^{1}$ Yi Jin $^{1\\dagger}$ $^{1}$ University of Science and Technology of China $^{2}$ DJI Technology Co., Ltd \n{natstand, anchen, junkangdai}@mail.ustc.edu.cn honza.qian@dji.com", + "bbox": [ + 192, + 178, + 805, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional auto white balance (AWB) algorithms typically assume a single global illuminant source, which leads to color distortions in multi-illuminant scenes. While recent neural network-based methods have shown excellent accuracy in such scenarios, their high parameter count and computational demands limit their practicality for real-time video applications. The Fast Fourier Color Constancy (FFCC) algorithm was proposed for single-illuminant-source scenes, predicting a global illuminant source with high efficiency. However, it cannot be directly applied to multi-illuminant scenarios unless specifically modified. To address this, we propose Integral Fast Fourier Color Constancy (IFFCC), an extension of FFCC tailored for multi-illuminant scenes. IFFCC leverages the proposed integral UV histogram to accelerate histogram computations across all possible regions in Cartesian space and parallelizes Fourier-based convolution operations, resulting in a spatially-smooth illumination map. This approach enables high-accuracy, real-time AWB in multi-illuminant scenes. Extensive experiments show that IFFCC achieves accuracy that is on par with or surpasses that of pixel-level neural networks, while reducing the parameter count by over $400\\times$ and processing speed by $20 - 100\\times$ faster than network-based approaches.", + "bbox": [ + 89, + 300, + 483, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 690, + 220, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Color constancy is a fundamental property of the human visual system, allowing for consistent color perception despite changes in illumination chromaticity. While this ability may seem \"natural\" to the human eye and visual cortex, replicating it in computational imaging systems is a complex challenge [19, 22]. Auto white balance (AWB) in image signal processing (ISP) is thus essential for approximating human-like color constancy by estimating ambient illumination chromaticity, enabling accurate color reproduction across varied lighting conditions.", + "bbox": [ + 89, + 715, + 482, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b156077960eef34f511ecb6e49ee0bd3c4e88c4c761d57de9dbc6749278691a.jpg", + "image_caption": [ + "Figure 1. Frameworks comparison between IFFCC and FFCC (top). FFCC uses a global histogram to predict a single illuminant, limiting its application in multi-illuminant scenes. In contrast, IFFCC utilizes an integral UV histogram for efficient region-specific histogram computation, generating spatially continuous illumination map through parallelized Fast Fourier convolution and smoothing to achieve high-accuracy, real-time auto white balance in multi-illuminant scenarios." + ], + "image_footnote": [], + "bbox": [ + 516, + 266, + 906, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional white balance methods often assume the presence of a single dominant illumination and rely on statistical algorithms [18, 21, 23, 45], such as the gray world algorithm [10] and the white patch algorithm [33] etc., to estimate the scene illumination. These algorithms may also be machine learning-based, trained on AWB benchmarks [3, 27]. However, they frequently bear local color distortion in complex multi-illuminant environments, as a single white balance adjustment is inadequate to handle varying lighting conditions. For instance, a room illuminated by both warm high pressure sodium lighting and cool window light may still fail to accurately reproduce the scene's true colors.", + "bbox": [ + 511, + 595, + 906, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, deep neural network-based algorithms have been widely applied to pixel-level multi-illuminant prediction tasks [11, 31, 34, 45], utilizing deep learning to achieve accurate illumination map estimation in complex scenes. These methods are especially effective in capturing edge-aware, spatial illumination variations. However, the high parameter counts and computational demands of these models create substantial challenges for deployment in real-", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26420", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "time, resource-limited applications, such as mobile photography devices. Beyond achieving high pixel-level prediction accuracy, a practical local AWB solution from an industrial perspective should prioritize several key requirements:", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Effectiveness - The algorithm can accomplish the AWB task up to a high accuracy, in any given spatial region, ensuring robust performance in a wide range of scenarios and including different light intensities and color temperatures.", + "bbox": [ + 89, + 152, + 480, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficiency - White balance adjustments must be fast enough to provide immediate feedback during view-finding and capture, particularly in rapidly changing lighting. Limited camera resources require AWB, along with other ISP functions like AE, AF, and color correction, to operate in real time. On resource-constrained devices like smartphones, low computational complexity and memory usage are essential, with AWB processing times ideally under 10 ms.", + "bbox": [ + 89, + 213, + 480, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thumbnail Input or Stat - Many color constancy algorithms require full-resolution, high-bit images, which is resource-intensive on pervasive mobile devices. For the purpose of real-time preview, the module for generating AWB stat is often realized in hardware (like DSP), outputting a thumbnail-sized input (e.g., $64 \\times 48$ , 8-bit).", + "bbox": [ + 89, + 335, + 480, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Smoothness - The algorithm should maintain stable performance under varying lighting conditions, avoiding artifacts or abrupt transitions, and ensure smooth, accurate white balance across entire images or video frames.", + "bbox": [ + 89, + 426, + 480, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The FFCC algorithm [3] effectively meets the requirements for global illumination scenarios but faces challenges in multi-illuminant scenes. Specifically, FFCC requires repeated histogram extraction for each target region, significantly increasing processing time and limiting its suitability for real-time applications, especially on resource-constrained devices like smartphones. Additionally, FFCC lacks mechanisms to ensure smooth spatial transitions under rapidly changing lighting, which can result in abrupt shifts in white balance.", + "bbox": [ + 89, + 487, + 480, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome FFCC's limitations in multi-illuminant scenarios, we propose Integral Fast Fourier Color Constancy (IFFCC), a local extension of FFCC [3] designed for white balance in complex lighting conditions, as shown in Fig.", + "bbox": [ + 89, + 638, + 480, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. IFFCC achieves speeds $20 - 100 \\times$ faster than other local methods, with a runtime of $5.8\\mathrm{ms}$ for $64 \\times 48$ preview images on a CPU. Building upon FFCC, IFFCC introduces several key contributions:", + "bbox": [ + 89, + 700, + 480, + 760 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. IFFCC outperforms prior-arts traditional algorithms and most network-based methods in multi-light source scenarios, achieving performance comparable to the best network-based methods;", + "2. IFFCC leverages the proposed integral UV histogram to accelerate histogram extraction across all possible regions in Cartesian space. It efficiently computes the histogram for any region using only a few simple arithmetic operations, ensuring both high accuracy and real-time" + ], + "bbox": [ + 89, + 763, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performance for multi-illuminant white balance.", + "bbox": [ + 532, + 90, + 852, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. IFFCC is spatially adaptive: spatial interpolation and filtering on local estimates ensure smooth transitions while preserving distinct light boundaries and sharp gradients.", + "bbox": [ + 514, + 106, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 167, + 653, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Global AWB Methods", + "text_level": 1, + "bbox": [ + 513, + 193, + 718, + 207 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conventional methods have long served as foundational approaches for uniform color constancy. These techniques estimate the illuminant's color based on the statistical characteristics of a single lighting source in the image [9, 17, 18, 21, 23, 47]. Notable single-illumination examples include the Gray World assumption [10], which posits that the average color of a scene under equal-energy white illumination should appear gray. Similarly, White Patch [33] assumes that the brightest pixel serves as a hint for illumination chromaticity. Refining these assumptions to focus on local patches or higher-order gradients has led to more robust statistics-based methods, optimized for more-challenging global ambient illuminations. These methods include General Gray World [1], Gray Edge [47], Shades-of-Gray [16], and LSRS [20], among others [11, 42].", + "bbox": [ + 511, + 215, + 903, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the context of machine learning-based methods, more machine learning-based methods have been proposed, such as learning-based kernels [2, 3], convolutional feature extraction [6, 27], and other machine learning approaches have been devised specifically for handling single-illumination scenarios [35, 36, 41, 43, 48]. The most closely related methods to ours are Barron's CCC [2] and FFCC [3]. CCC reformulates color constancy as a color localization task in log-chromaticity space, using convolutional kernels to accurately locate the correct color. FFCC builds on this by applying the Fast Fourier Transform on a chromaticity torus, greatly improving the speed and accuracy of chromaticity estimation. However, both are global methods and lack direct applicability to local scenes unless explicitly provided with a local window.", + "bbox": [ + 511, + 443, + 905, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Local AWB Methods", + "text_level": 1, + "bbox": [ + 513, + 666, + 709, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several multi-illuminant benchmarks have been introduced [4, 5, 8, 24, 31, 35, 39], with various methods proposed to tackle local white balance. These methods often rely on additional prior information or metadata, such as the number of light sources [4, 26] or facial regions of interest [5]. For instance, [28] uses flash photography to achieve local white balance. A novel framework in [4] treats the problem as an energy minimization task within a conditional random field, leveraging local light source estimates. [8] employs a global method with spatial variation, while [7] introduces a CNN framework based on sliding windows. Additionally, [46] applies a generative adversarial network (GAN) to correct images using synthetic datasets, adjusting the color based on the lighting information. More recently, Domislovic et", + "bbox": [ + 511, + 688, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26421", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "al. [14] proposed extracting spatial features for illumination prediction on each patch, and AID [32] introduces a deep model utilizing slot attention [37] to generate chromaticity and weight maps for each light source, with the added benefit of allowing user edits.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Among these approaches, most focus on pixel-level predictions, often leveraging encoder-decoder architectures (e.g., U-Net) to capture fine-grained illumination chromaticity details. However, this strategy presents several practical challenges. First, pixel-level prediction is computationally intensive, especially for high-resolution images, which significantly limits real-time applicability. Second, due to the lack of spatial context, pixel-wise color correction can overlook critical illumination chromaticity differences, leading to the loss of smooth illumination transitions and distortions in color edges, both of which are essential for accurate human visual perception. By treating each pixel independently, these methods risk introducing inconsistencies across the image and erasing subtle but important color shifts. As a result, their effectiveness in real-world applications, particularly in resource-constrained environments such as mobile devices, remains significantly limited.", + "bbox": [ + 91, + 167, + 482, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 435, + 181, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose the Integral Fast Fourier Color Constancy (IFFCC) model, which builds upon the foundations of Fast Fourier Color Constancy (FFCC) [3]. Drawing inspiration from [40], we implement an histogram search for all possible target regions in image data within Cartesian coordinate space. This enables efficient retrieval of any region's histogram in linear computation time, avoiding redundant summation operations.", + "bbox": [ + 89, + 460, + 482, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FFCC Foundation. Consider a linear image stat $I$ obtained from the camera, where there are no saturated pixels and the black level has been subtracted. Assuming the RGB value of any pixel is the product of its true albedo RGB value $W$ and the uniform illumination value shared across all pixels: $I = W \\times L$ . Although this assumption disregards shadows, dichromatic reflections, and spatial variations, it remains effective and widely accepted. Our aim is to estimate the $L$ given $I$ , and then $W = I / L$ .", + "bbox": [ + 89, + 580, + 482, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For an input RGB image $I$ , CCC [2] defines two log-chroma measures for each pixel $i$ :", + "bbox": [ + 89, + 717, + 482, + 747 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nu ^ {(i)} = \\log \\left(I _ {g} ^ {(i)} / I _ {r} ^ {(i)}\\right) v ^ {(i)} = \\log \\left(I _ {g} ^ {(i)} / I _ {b} ^ {(i)}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 753, + 482, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neglecting the absolute scaling of $W$ , the estimation of $L$ is then simplified as $L_{u} = \\log (L_{g} / L_{r}) \\quad L_{v} = \\log (L_{g} / L_{b})$ . Given the unknown absolute scale, the inverse mapping from RGB to UV is challenging. To simplify the recovery process, we assume $L$ is the unit-norm:", + "bbox": [ + 89, + 781, + 482, + 857 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {r} = \\exp (- L _ {u}) / z \\quad L _ {g} = 1 / z \\quad L _ {b} = \\exp (- L _ {v}) / z \\tag {2} \\\\ z = \\sqrt {\\exp (- L _ {u}) ^ {2} + \\exp (- L _ {v}) ^ {2} + 1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 864, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This logarithmic chromatic space method has several advantages over the RGB approach. First, it involves only two unknowns instead of three, which enhances numerical stability. Additionally, this framework transforms the multiplicative constraints related to $W$ and $I$ into additive constraints [15], effectively converting the color constancy task into a positioning task in a two-dimensional space.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following FFCC, we use FFT-based convolution in the log-chroma space, leveraging its periodicity to accelerate computations. Given the histogram's limited size, which does not fully capture the range of colors in natural images, a modular algorithm is needed to allow pixels to \"wrap around\" in this space:", + "bbox": [ + 511, + 196, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} H (x, y) = \\sum_ {i} \\left(\\mod \\left((u ^ {(i)} - u _ {s}) / b - x, n\\right) < 1 \\right. \\tag {3} \\\\ \\wedge \\mod \\left(\\left(v ^ {(i)} - v _ {s}\\right) / b - y, n\\right) < 1\\bigg) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 299, + 903, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $H(x,y)$ represents the count of pixels in $I$ whose log-chroma coordinates are close to the $(u,v)$ coordinates associated with the histogram position $(x,y)$ , $n$ is the number of bins, and $b$ is the bin size. $u_{s}$ and $v_{s}$ denote the starting positions of $u$ and $v$ , respectively. The operations here differ from those in CCC. In this case, the $(x,y)$ positions in the histogram no longer represent absolute colors, but rather a set of $(u,v)$ color coordinates. The filtered results produce a set of light sources, from which we apply the \"gray light de-aliasing\" method to resolve the resulting light source clusters. This method assumes that the light sources are as close as possible to the center of the histogram.", + "bbox": [ + 511, + 372, + 903, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FFCC employs the Differentiable Bivariate von Mises (BVM) [38] method to address the single estimation of light sources from the toroidal PDF. Unlike FFCC, we need to predict multiple light sources for each image. Therefore, the BVM method is parallelized across multiple windows to predict several light sources simultaneously, as described in Section 3.3. The final convolutional structure is as follows:", + "bbox": [ + 511, + 554, + 905, + 659 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP = \\operatorname {s o f t m a x} \\left(B + M \\circ \\sum_ {s} H _ {s} * F _ {s}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 672, + 903, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B$ is the bias map, $M$ is the learnable gain map, and $\\{F_s\\}$ is the set of learnable filters convolved with the edge-enhanced histogram group $\\{H_s\\}$ .", + "bbox": [ + 511, + 715, + 903, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Integral UV Histogram", + "text_level": 1, + "bbox": [ + 511, + 772, + 728, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FFCC predicts a global illumination for the image. However, when multiple illuminations are anticipated for different target regions, the process requires FFCC to repeatedly extract histograms from these various areas, significantly increasing computational load and processing time. Inspired by the work of [40], we efficiently compute the histograms for all potential regions within a given image", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26422", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9fa8d7fe5135bcbc8551dc46a1cfd948d3d409523a7929c5db461a3bcc871ded.jpg", + "image_caption": [ + "(a) Integral UV histogram" + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 253, + 204 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c22d7a3a899c0ae6c6527a8836033d10d8d3519e3fa7f1d558c0f4b712741192.jpg", + "image_caption": [ + "(b) Overlapping grid" + ], + "image_footnote": [], + "bbox": [ + 254, + 98, + 379, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d0adb3ec75ce2066cbdd29ef80d6c335fa17045106cb0f51e895a248ddc2f1da.jpg", + "image_caption": [ + "(c) Local histogram" + ], + "image_footnote": [], + "bbox": [ + 382, + 99, + 504, + 199 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e9c2b23ee513c98e8c3e89e53bf83048f53676751e8e385a4d05a766e6230922.jpg", + "image_caption": [ + "(d) Parallel prediction", + "Figure 2. Overview of the IFFCC process for predicting multi-illuminant scenes. We precompute the integral UV histogram of the input image in the log-chroma space (a). Multiple target regions are extracted on the integrated UV histogram with the overlapping grid (b), and the local log-chroma histogram (c) corresponding to each region is computed using simple arithmetic operations on the UV histogram. Each histogram forms a torus that can be unfolded, producing a cluster of potential light sources. A de-aliasing step selects the most likely light source for each region (d), represented by a point within an ellipse, where the ellipse indicates IFFCC's covariance output. Final light estimates are smoothed through interpolation and guided filtering, creating a continuous illumination map (e) that combines with the input to produce a white-balanced image (f) for multi-illuminant scenes." + ], + "image_footnote": [], + "bbox": [ + 511, + 88, + 624, + 207 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a158bceeb65724cc355c40269d2aaab2c1c4edd3fad9f2126ce093515cd1e9cc.jpg", + "image_caption": [ + "(e) Spatial smoothing" + ], + "image_footnote": [], + "bbox": [ + 627, + 99, + 751, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dc2acfde9366f71c35a851c6a2567d933b4de6f20ffaea42c249a67e0f61159c.jpg", + "image_caption": [ + "(f) Output image" + ], + "image_footnote": [], + "bbox": [ + 756, + 99, + 882, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "using Cartesian coordinates. By leveraging the spatial positions of data in this coordinate system, we propagate a function starting from the origin to navigate through the remaining positions. The histogram for the current point is then updated based on the histograms of previously processed points. Once the integral histogram for each data point is obtained, the histogram for the target region can be derived using simple addition and subtraction operations. This method eliminates the need to reconstruct separate histograms for each region individually. As a result, this efficient approach significantly reduces computational overhead and improves processing time, enabling rapid analysis of multiple areas within the image.", + "bbox": [ + 88, + 354, + 483, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After transforming the image into log-chroma space by Eq.1, we define the integral uv histogram as follows:", + "bbox": [ + 89, + 551, + 482, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {\\text {I n t e g r a l}} \\left(x ^ {n}, b\\right) = \\bigcup_ {m = 0} ^ {n} B (u, v) ^ {m} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 590, + 482, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nB (u, v) = \\left( \\begin{array}{l} \\mod \\left(\\operatorname {r o u n d} \\left(\\frac {u - u _ {s}}{b}\\right), n\\right), \\\\ \\mod \\left(\\operatorname {r o u n d} \\left(\\frac {v - v _ {s}}{b}\\right), n\\right) \\end{array} \\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 637, + 482, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x^n$ is a sequence of points $x^0, x^1, \\ldots, x^n$ . $B(\\cdot)$ represents the corresponding bin for the current data point, $\\bigcup$ denotes the union operator, and the value of bin $b$ in $H_{\\mathrm{Integral}}(x^n, b)$ is the sum of the histogram values of bin $b$ from all previously traversed data points. The integral histogram can also be expressed recursively as follows:", + "bbox": [ + 89, + 710, + 483, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {\\text {I n t e g r a l}} \\left(x ^ {m}, b\\right) = H _ {\\text {I n t e g r a l}} \\left(x ^ {m - 1}, b\\right) \\bigcup B (u, v) ^ {m} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 815, + 482, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where at the initial data point, $H_{\\mathrm{Integral}}(0,b) = 0$ .", + "bbox": [ + 89, + 854, + 413, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In log-chroma space, updating the histogram at the current point requires the histograms of the data points above,", + "bbox": [ + 89, + 871, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the left, and the top-left of the current point. These three histograms have already been calculated in previous scans, so only three simple arithmetic operations are needed to update the histogram for the current point. Afterward, the bin value $b$ for the current point is calculated, and the histogram is updated accordingly. Before each propagation step, all scanned bin values are copied to the current bin values, which can be efficiently handled through fast hardware or pointer operations.", + "bbox": [ + 511, + 354, + 906, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For an RGB image converted to $N_{1} \\times N_{2} \\times 2$ log-chroma space data, the integral histogram propagated via wavefront scanning from the top-left starting point can be written as:", + "bbox": [ + 511, + 489, + 905, + 535 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} H _ {\\text {I n t e g r a l}} (u, v, b) = H _ {\\text {I n t e g r a l}} (u - 1, v, b) \\\\ + H _ {\\text {I n t e g r a l}} (u, v - 1, b) \\tag {8} \\\\ - H _ {\\text {I n t e g r a l}} (u - 1, v - 1, b) + B (u, v) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 541, + 903, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(u,v)$ is the coordinate of the log-chroma space. In our method, the data in log-chroma space can be represented by a two-dimensional floating point tensor of a two-dimensional scene [40], allowing us to calculate the computational load ratio between the traditional histogram method and the integral histogram method as follows:", + "bbox": [ + 511, + 601, + 905, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR = \\frac {\\left[ 2 1 7 M _ {1} M _ {2} + 4 B ^ {2} \\right] S _ {1} S _ {2}}{2 7 4 + 1 1 (6 8 + 4 B ^ {2}) S _ {1} S _ {2}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 696, + 903, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_{1}$ and $M_{2}$ represent the size of the target region in the two-dimensional scene, $B$ is the number of bins, and $S_{1}$ and $S_{2}$ are the different scales used when extracting histograms. The result of the calculation for $R$ when $M_{1} = M_{2} = 128$ , $B = 64$ , and $S_{1} = S_{2} = 1$ is approximately $R = 19.71$ .", + "bbox": [ + 511, + 734, + 905, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overlapping Grid", + "text_level": 1, + "bbox": [ + 511, + 833, + 689, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During inference, we use a sliding window with overlap to efficiently retrieve histograms for sub-regions in the log-chroma space, as illustrated in Fig.2. By pre-computing an", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26423", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "integral histogram over the entire scene, histograms for any arbitrary sub-region can be directly retrieved using Eq.8, without reconstructing each one from scratch, which significantly reduces computational costs. This overlapping grid method captures local chromatic variations by sliding a defined window across the scene and calculating histograms incrementally. The overlapping regions between windows maintain continuity and minimize edge effects, enhancing the accuracy of chromatic analysis throughout the scene.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Parallel Prediction", + "text_level": 1, + "bbox": [ + 89, + 234, + 272, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To speed up inference, our method requires only one inference to predict several illumination across multiple target regions within the image. Given an image's integral UV histogram in log-chroma space, denoted as $H_{\\mathrm{integral}}(H,W,B,B,N)$ , where $H$ and $W$ represent the image size, $B$ is the number of bins, and $N$ is the number of augmented channel's histogram. The overlapping grid approach can be applied to obtain a collection $X$ of $k$ target region histograms. This collection $X$ , represented as $X(k,B,B,N)$ , includes histograms for each specific target region in the log-chroma space. The process of calculating the PDF of $X$ in the frequency domain is as follows:", + "bbox": [ + 89, + 256, + 483, + 438 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} H = \\operatorname {I F F T} \\left(\\sum_ {k} \\operatorname {F F T} (X) \\cdot \\operatorname {F F T} (F)\\right) + B (10) \\\\ P = \\frac {\\exp^ {(H - \\max (H))}}{\\sum_ {\\mathrm {d i m s}} \\exp^ {(H - \\max (H))}} (11) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 445, + 482, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{FFT}(.)$ and $\\mathrm{IFFT}(.)$ are the fast Fourier transform and its inverse, respectively, $F$ is the convolution kernel, and $B$ is the bias map.", + "bbox": [ + 89, + 532, + 482, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To parallelize the estimation of the mean of the Bivariate Von Mises (BVM) distribution from a histogram, we calculate circular mean values for the histogram coordinates $i$ and $j$ across multiple regions concurrently. This process enables effective mean direction estimation within the toroidal data space, essential for capturing periodic patterns in wrapped grid data. For each region $k$ , we define $\\mu = \\{\\mu_k\\}_{k=1}^K$ , where each $\\mu_k$ is given by:", + "bbox": [ + 89, + 577, + 483, + 699 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {k} = \\left[ \\begin{array}{l} u _ {s} \\\\ v _ {s} \\end{array} \\right] + b \\left[ \\begin{array}{l} \\bmod \\left(\\frac {n}{2 \\pi} \\arctan 2 \\left(y _ {1} ^ {k}, x _ {1} ^ {k}\\right), n\\right) \\\\ \\bmod \\left(\\frac {n}{2 \\pi} \\arctan 2 \\left(y _ {2} ^ {k}, x _ {2} ^ {k}\\right), n\\right) \\end{array} \\right] \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 704, + 482, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the components $y_1^k, x_1^k, y_2^k$ , and $x_2^k$ are computed in parallel for each $k$ :", + "bbox": [ + 89, + 746, + 482, + 776 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {1} ^ {k} = \\sum_ {i} P _ {i} ^ {k} \\sin \\left(\\frac {2 \\pi i}{n}\\right), x _ {1} ^ {k} = \\sum_ {i} P _ {i} ^ {k} \\cos \\left(\\frac {2 \\pi i}{n}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 782, + 482, + 828 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {2} ^ {k} = \\sum_ {j} P _ {j} ^ {k} \\sin \\left(\\frac {2 \\pi j}{n}\\right), x _ {2} ^ {k} = \\sum_ {j} P _ {j} ^ {k} \\cos \\left(\\frac {2 \\pi j}{n}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 823, + 444, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P_{i}^{k}$ and $P_{j}^{k}$ represent the marginal distributions for rows and columns of each target region's 2D histogram.", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After concurrently obtaining illumination values for all $k$ target regions, we reconstruct the full illumination map by interpolating or applying edge-preserving filtering methods.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Spatial Smoothing", + "text_level": 1, + "bbox": [ + 511, + 145, + 692, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In multi-light scenarios, patch-wise illumination estimates often exhibit discrete, locally discontinuous characteristics. To reconstruct a smooth and seamless illumination map across the entire image, it is vital to postprocess these patch-level estimates using interpolation.", + "bbox": [ + 511, + 167, + 905, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ linear interpolation due to its simplicity and ability to create smooth, piecewise-linear transitions for estimating intermediate values. Although higher-degree polynomial interpolation could be used, it risks introducing oscillations. After getting estimates for $k$ target regions, interpolation adjusts these values based on their relative locations, resulting in a smooth illumination map. Although interpolation enhances overall smoothness, it may blur local details, particularly in regions with strong light edges or steep illumination gradients, potentially leading to a loss of edge information.", + "bbox": [ + 511, + 242, + 906, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Guided filtering [25] is then applied to preserve details at the boundaries and structural features of the illumination map, preventing blur or distortion during transitions. Overall, combining these techniques results in a globally coherent illumination distribution that maintains edge clarity while appearing smoother and more natural. Given the interpolated illumination map $L$ and the raw image $G$ as the guiding image, the output illumination map $O$ at each pixel $O_{i}$ can be expressed as:", + "bbox": [ + 511, + 407, + 905, + 544 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO _ {i} = a _ {k} G _ {i} + b _ {k}, \\quad \\forall i \\in \\omega_ {k} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 554, + 903, + 570 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\na _ {k} = \\frac {\\frac {1}{| \\omega |} \\sum_ {i \\in \\omega_ {k}} G _ {i} L _ {i} - \\bar {G} _ {k} \\bar {L} _ {k}}{\\sigma_ {k} ^ {2} + \\epsilon}, \\quad b _ {k} = \\bar {L} _ {k} - a _ {k} \\bar {G} _ {k} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 577, + 903, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "here, $\\bar{G}_k$ and $\\bar{L}_k$ are the mean values of $G$ and $L$ in window $\\omega_k$ , with $\\sigma_k^2$ as the variance of $G$ . The constant $\\epsilon$ prevents division by zero. The final illumination map $O$ is obtained by weighted averaging.", + "bbox": [ + 511, + 619, + 905, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO = \\frac {1}{| \\omega |} \\sum_ {k \\in \\omega_ {i}} \\left(a _ {k} G _ {i} + b _ {k}\\right) \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 690, + 903, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This post-processing step is essential: it creates a coherent illumination map that accurately reflects the lighting conditions across the entire image while preserving critical features, thus the reconstructed illumination aligns well with the underlying content of the image.", + "bbox": [ + 511, + 734, + 905, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 823, + 638, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 848, + 702, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluated our proposed IFFCC on two widely used multi-light source datasets: the LSMI dataset [31] and the", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26424", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3888c9b6297409194c15f6e6a164c3be83ea2fc049b03b43214a78ec21b38cd9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCanon_5dCanon_550dMotoPanasonicSonyUse all
meanmedianmeanmedianmeanmedianmeanmedianmeanmedianmeanmedian
Huassin and Akbari [29]13.8813.7613.7313.5413.0612.8913.6213.7314.2213.6813.6713.46
CRF(White-Patch)[4]7.665.967.695.776.945.327.585.697.905.867.195.44
Patch-based[24]4.853.114.612.984.172.494.582.954.963.164.302.89
Keypoint-based[24]5.593.885.483.944.973.285.423.795.763.825.463.59
Superpixel-based[24]4.883.684.793.564.192.894.553.544.973.824.203.10
IFFCC2.061.542.161.771.901.362.061.592.661.962.191.56
Bianco et al.[7]4.634.024.594.164.273.894.984.565.344.768.015.69
HypNet/SelNet[44]5.424.935.164.835.094.745.234.985.825.346.313.95
LSMI-U[31]3.142.332.922.462.792.043.172.703.582.883.722.40
Domislovic et al.[12]2.632.182.772.062.071.492.381.732.922.242.281.60
", + "bbox": [ + 122, + 87, + 883, + 247 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/395e39980f52a0318150e975873d56dcecdac2c67fb2c1f3d02dfaf68d257b43.jpg", + "table_caption": [ + "Table 1. The performance results on the Shadow dataset are presented. The dataset is segmented by images captured from different cameras, with \"use all\" representing the entire dataset. We report the mean and median angular errors for each subset. IFFCC employs the default settings, with a window size of 128 and an overlap size of 64. The model is trained using the blended illumination within each window as the ground truth (GT), and the corresponding error is computed." + ], + "table_footnote": [], + "table_body": "
MethodOutdoorIndoorNighttime
meanmedianb25w25meanmedianb25w25meanmedianb25w25
Huassin and Akbari[29]13.1912.897.1019.1214.8413.228.4320.1816.3514.969.7522.33
CRF(White-Patch)[4]7.315.662.3113.368.786.894.3614.406.965.162.3815.85
Patch-based[24]4.533.291.3810.475.584.432.0911.393.482.171.8912.53
Keypoint-based[24]5.773.881.2411.786.254.041.7912.603.994.101.4913.08
Superpixel-based[24]4.733.011.338.215.183.491.929.244.632.981.349.82
IFFCC1.741.280.423.833.042.140.666.712.851.530.547.43
Bianco et al.[7]4.862.361.9811.757.425.623.3812.508.225.783.1013.68
HypNet/SelNet[44]6.094.081.7712.317.075.272.9811.494.763.212.2012.59
LSMI-U[31]2.122.031.025.485.724.621.498.544.812.471.338.89
Domislovic et al.[12]1.781.460.654.344.453.310.987.463.681.890.838.22
", + "bbox": [ + 122, + 325, + 883, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The performance results on the Shadow dataset are presented. The dataset is segmented by images captured in different scenes, with each scene including images taken by five different cameras. We report the mean and median angular errors for each subset. IFFCC uses the default settings, with a window size of 128 and an overlap size of 64. The model is trained using the blended illumination within each window as the ground truth (GT), and the corresponding error is computed.", + "bbox": [ + 89, + 494, + 906, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Shadow dataset [13]. The LSMI dataset includes over 7,486 images captured by three different cameras in a variety of scenes, each containing multiple light sources. The Shadow dataset consists of 2,500 images taken in both indoor and outdoor settings, with each image accompanied by a binary segmentation mask, where each region is illuminated by a single light source. In all experiments, full-sized training images were randomly cropped to $128 \\times 128$ , and testing was conducted on $256 \\times 256$ images. All images were black-level corrected, and calibration objects were masked.", + "bbox": [ + 89, + 577, + 482, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "IFFCC's computational efficiency enables CPU-only training and inference, eliminating the need for a GPU. All experiments were conducted on an Intel Xeon Gold 6258R CPU. Performance was evaluated using standard error metrics, including mean and median angular errors and the arithmetic means of the first and third quartiles (\"best $25\\%$ \" and \"worst $25\\%$ ). Additionally, we compared model parameter counts and CPU inference times for $256\\times 256$ images. Since our method operates on patches, we calculate blended illumination values from the ground truth illumination map within a sliding window, using these as the ground", + "bbox": [ + 89, + 734, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "truth for each patch. The impact of different training approaches is discussed in the ablation study.", + "bbox": [ + 511, + 577, + 903, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training, we input a set of enhanced histograms derived from filtered images that ensure non-negativity and intensity scaling consistency. Each image includes a local absolute deviation measure. Using the preconditioned frequency-domain optimization from FFCC [3], training runs for 64 iterations with a histogram size of $64 \\times 64$ .", + "bbox": [ + 511, + 608, + 905, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results and Comparisons", + "text_level": 1, + "bbox": [ + 511, + 710, + 746, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison We evaluated IFFCC against three traditional methods [4, 24, 29] and three learning-based methods [7, 12, 44] on the Shadow dataset. The experimental results are presented in Tables 1 and 2. Table 1 organizes the Shadow dataset into subsets based on images captured by different cameras, while Table 2 categorizes the dataset by various scene types, each containing photos taken with five distinct cameras. Despite comparisons with advanced network-based methods, IFFCC achieves state-of-the-art performance.", + "bbox": [ + 511, + 733, + 905, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results in Table 2 highlight the strong general", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26425", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08a937f70aae03599c928fd24a5b0ba0e383f9427fbb59c3c4026447a14e8ee5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 89, + 228, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1bec36dacc91776241920f0dbea216be1b771b850927083ecc5d11d2b00c5edb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 172, + 227, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56888799e5407b56c099a688485b917d9b3e861459154d8da45c002770b68bf6.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 94, + 255, + 227, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d6c1f4d1c60575cb7891b98fd9494aaa5d858a2d4b95da0ac267ae4af85e675.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 90, + 362, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/af6be4370513d5bd7350b9d0bcd3315bcd3ee941292a55f3242fb07f087f3c81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 172, + 362, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bde56dc819d1defdbb5541be2a7f288b84702c48bbcde735bb6b4b59fd5f7b39.jpg", + "image_caption": [ + "Gijsenij et al." + ], + "image_footnote": [], + "bbox": [ + 230, + 253, + 362, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a352477d17e3c42cdedc1870e49f0d90d6890147c66e8827500ec030b1e26f9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 90, + 496, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55a775a8a6912962afa985239c7bf9bd4221cc3ae75832540de61b1f5c67221a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 172, + 496, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7be33c47aac92232fd67ea8b0bc19b2208ad45f5877d437c76e2fb134c111f3c.jpg", + "image_caption": [ + "Bianoco et al." + ], + "image_footnote": [], + "bbox": [ + 364, + 253, + 496, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0dd66e408b39582773835b0802eaccfb3dee27a61af14c164757c3f2567c832c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 90, + 630, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/692b60b28a4e8c3437472f690710de1a647822075bbf7efd1ea832814eef9025.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 172, + 630, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d1b057fdc16654def0802470ac38a2a866c9c9c4401edbe5cc9883e6564e733.jpg", + "image_caption": [ + "LSMI-U" + ], + "image_footnote": [], + "bbox": [ + 500, + 253, + 630, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5f3ef5d045df532156854e852da0238fefc914336b67d0f33af1da1da554168.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 90, + 764, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/578fa221ec4854d1bee0d613c308525238a53ee3919aae231ab6dd78f2fc4955.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 172, + 764, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c7360708af7d35414bd03dcf2b61b266b35729567a32baf242b1c8ecec6aade1.jpg", + "image_caption": [ + "IFFCC" + ], + "image_footnote": [], + "bbox": [ + 633, + 253, + 764, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b011d90704d890e79777db34f213bb54497172c95b761511802578ff7a2d10b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 90, + 900, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d577927f0c885928d6cb4d026c736fdbd2a8c232ba6488c7ce0b9ddc1aaa7a32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 172, + 900, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a538d4fffeae53398142dc16c7a2d39c8ca8917514bb2c6236a4308133f96165.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 767, + 253, + 900, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0ad81bbb352626a96769bff158d51f6a6ff903c8928395347a2ffc805588863d.jpg", + "image_caption": [ + "Figure 3. Visualization of the comparison results on the LSMI dataset. For IFFCC, the window size is set to 128 with an overlap of 64. In other patch-based methods, the window size is set to the default value of 32." + ], + "image_footnote": [], + "bbox": [ + 96, + 407, + 254, + 486 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba26574610174747eb5ae94244c0ed0028e0ffd54ed1c48f554b9e3f8a4270b9.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 96, + 487, + 254, + 563 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59f71e4f4ddd0ad0117802a538f575941be9f2e3911c7f9027a9777d2d017bc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 407, + 416, + 486 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bd6ee5f171edd36fd445a4186ea3a8821cf802b79ea5cf9d44af44020dcf580.jpg", + "image_caption": [ + "White Patch" + ], + "image_footnote": [], + "bbox": [ + 258, + 487, + 416, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac70c95c817f221dfc4bd0d450aca068649ca126eebb87fcef44bcabf9f288ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 407, + 576, + 486 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/33380cca317a2900d5dfb00acc3485cbe98211611d31c1440736d4f80a6e7d42.jpg", + "image_caption": [ + "IFFCC" + ], + "image_footnote": [], + "bbox": [ + 419, + 487, + 576, + 563 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c533d90204ff306db7966fe16e36c4ae3c872714971aae4d8272aca8202eb307.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 407, + 738, + 486 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be30854263048d763a57a87751f48c3eaaf6127646d44cc082414eda1d18e041.jpg", + "image_caption": [ + "Illumination map", + "Figure 4. Visualization of illumination comparison for different patch-based methods. The window size is 32, with an overlap size of 16." + ], + "image_footnote": [], + "bbox": [ + 580, + 487, + 738, + 563 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/72b93274c4df36efdfd8fd403d295b30e491785bb00de686dd3a2a11099b491c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 407, + 901, + 486 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04bc55f93adc5ba4720e9b02f5a2a0ea3e4c3e23ec3bb8a4e214a9b4039d7dd1.jpg", + "image_caption": [ + "White-balanced image" + ], + "image_footnote": [], + "bbox": [ + 741, + 487, + 901, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ization ability of IFFCC. Network-based methods generally perform worse than IFFCC on multi-camera datasets, where each camera's unique spectral sensitivities complicate learning a single mapping across different devices. IFFCC generalizes well across multiple cameras because it leverages an integral log-chroma histogram approach, which directly models chromaticity distributions without being overly reliant on the specific spectral sensitivities of individual cameras. This histogram-based method effectively captures illumination variations across different scenes and cameras, allowing IFFCC to produce robust predictions without requiring extensive fine-tuning for each camera's unique spectral characteristics.", + "bbox": [ + 88, + 633, + 482, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The experimental results on the LSMI dataset, presented in Tables 3, and 4, provide a comparative analysis of patch-based and pixel-level methods. The test set is categorized into Multi (multiple light sources), and Mixed (compris", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing both single and multiple light sources). Among patch-based approaches, IFFCC demonstrates state-of-the-art performance, achieving results comparable to those of leading pixel-level methods, including AID [32]. Notably, IFFCC's compact parameter count and rapid processing speed render it highly suitable for deployment on mobile and resource-constrained devices.", + "bbox": [ + 511, + 633, + 906, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compared the processing time for a $256 \\times 256$ image in a multi-illuminant scene using FFCC and IFFCC, as shown in Table 5. As the number of windows increases, IF-FCC significantly outperforms FFCC in processing speed.", + "bbox": [ + 511, + 744, + 906, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison Fig.3 visually compares our method against alternative approaches on the LSMI dataset. To enhance visualization quality, color correction, gamma correction, and tone mapping were applied to the white-balanced images. The results indicate that IFFCC achieves more accurate white balance restoration than competing", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26426", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1dae61065cc8e0656026aba8cf2e4424ce6d03d57053b118b814f8b1216a5032.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMultiParams (M)Time (s)
Galaxy mean med.Nikon mean med.Sony mean med.
Patch -levelGijsenij[23]12.389.579.316.5911.198.14-1.3
Bianoco[7]5.564.344.654.194.383.940.16-
IFFCC2.481.902.301.502.481.970.0120.03
Pix2Pix[30]4.282.634.182.764.373.265.40.48
Pixel LSMI-H[31]3.132.703.203.013.653.330.520.28
-levelLSMI-U[31]2.852.552.361.842.762.385.70.51
AID[32]2.031.432.261.392.161.646.4>1
", + "bbox": [ + 101, + 88, + 472, + 234 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5c889eceed5f07f00df441ff46890d1341ef40fefc7c8f4d94382f4e7ac1ed49.jpg", + "table_caption": [ + "Table 3. The performance results on the multi-illumination images from the LSMI dataset are presented, with testing time corresponding to the processing time for $256\\times 256$ images on a CPU." + ], + "table_footnote": [], + "table_body": "
MethodMixedParams (M)Time (s)
Galaxy mean med.Nikon mean med.Sony mean med.
Patch -levelGijsenij[23]10.097.439.215.839.467.13-1.3
Bianoco[7]4.893.833.933.483.843.210.16-
IFFCC1.981.792.111.481.931.720.0120.03
Pix2Pix[30]5.662.445.412.494.202.205.40.48
Pixel LSMI-H[31]3.062.542.922.633.212.850.520.28
-level LSMI-U[31]2.631.912.161.552.682.295.70.51
AID[32]1.631.321.761.141.681.446.4>1
", + "bbox": [ + 101, + 299, + 470, + 445 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/121045d95f5e8995edb90824e37a7ec859129aec0713cfb90000eca93144eb4f.jpg", + "table_caption": [ + "Table 4. The performance results on the mixed-illumination images (containing both single and multiple light sources) from the LSMI dataset are presented, with testing time corresponding to the processing time for $256 \\times 256$ images on a CPU." + ], + "table_footnote": [], + "table_body": "
[32,0][32,16][64,32][64,48][128,64][128,96]
FFCC162ms450ms203ms689ms88ms234ms
IFFCC34ms46ms29ms40ms27ms33ms
", + "bbox": [ + 94, + 523, + 478, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "methods. We also present a comparison of illumination maps between our method and patch-based approaches, as shown in Fig.4. The window size is set to $32 \\times 32$ with an overlap of 16. The results demonstrate that IFFCC achieves a more precise and smoother illumination prediction.", + "bbox": [ + 89, + 642, + 483, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 727, + 243, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6 presents results from ablation experiments analyzing the impact of various training strategies, window sizes, and overlap regions. Here, \"B\" represents training with a blend of illumination maps as the ground truth (GT), and \"M\" uses the dominant light source as GT. Parameters \"ws\" and \"os\" indicate window and overlap sizes, respectively. Findings reveal that smaller window sizes, with limited information, reduce prediction accuracy. However, the blended illumination strategy provides improved accuracy compared to using a single dominant light source. Visual", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/170639a8111eda06b64c9933611c1e45499fe109289f25eeaf9f308ea695c33a.jpg", + "table_caption": [ + "Table 5. Comparison of average testing time for FFCC and IF-FCC on the Shadow dataset, evaluated across different window and overlap sizes. FFCC employs a conventional histogram computation approach." + ], + "table_footnote": [], + "table_body": "
type[ws, os]meanmedianb25w25
B[32, 16]2.652.171.374.65
B[64, 0]2.352.011.134.31
M[64, 32]2.532.041.254.83
B[64, 32]2.371.871.164.38
M[128, 0]2.462.031.044.78
B[128, 0]2.161.680.914.23
M[128, 32]2.501.881.044.91
B[128, 32]2.141.640.854.30
M[128, 64]2.421.950.934.71
B[128, 64]2.061.540.814.19
B[128, 96]2.131.590.834.25
B[192, 128]1.981.340.724.08
", + "bbox": [ + 563, + 88, + 864, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study on the Shadow dataset (Canon5d), reporting the impact of different window sizes, overlap sizes, and training strategies on accuracy.", + "bbox": [ + 511, + 239, + 906, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1af028746080cd044c89b20f5f01098b51ecb64f29ab673ec13a63830e9752ec.jpg", + "image_caption": [ + "Figure 5. (a)-(d) show illumination maps generated with different window sizes ([32, 16], [64, 32], [128, 64], [192, 128]), each processed with smooth transition adjustments. (e) presents the ground truth (GT) illumination map, and (f) displays the white-balanced image obtained by combining the input image with the GT." + ], + "image_footnote": [], + "bbox": [ + 517, + 295, + 645, + 438 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a75bba796124fc67a525818245f86163fea8d2a4ec33384501ef87c136c35c17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 296, + 772, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b36b0d9b85f3ccc80032ba5388a5177d38152dcefb2d893600aac85cdda397a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 296, + 901, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "comparisons (Fig. 5) show that smaller windows enhance texture details in the predicted illumination maps but with lower accuracy. Conversely, larger windows (e.g., [192, 128]) reduce error metrics by capturing more information, yet they may blur boundaries in multi-illuminant regions, compromising edge clarity.", + "bbox": [ + 511, + 532, + 906, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 637, + 633, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose IFFCC, a practical solution for achieving accurate, real-time auto white balance (AWB) in multi-illuminant environments. Built upon the Fast Fourier Color Constancy (FFCC) algorithm, IFFCC leverages an integral UV histogram and parallelized Fourier-based convolution to rapidly estimate illumination across different image regions, addressing the need for localized, adaptive AWB. Results on multiple datasets demonstrate that IFFCC delivers accuracy comparable to or surpassing pixel-level neural network methods, while reducing the parameter count by over 400 times and speeding up processing by a factor of 20–100. These advantages make IFFCC highly suitable for resource-limited, real-time applications such as video streaming.", + "bbox": [ + 511, + 662, + 906, + 858 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We believe that FFCC-based methods, including IFFCC, will continue to enrich the field and attract increased attention across a broader range of ISP applications.", + "bbox": [ + 511, + 857, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26427", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work was supported in part by the Anhui Provincial Natural Science Foundation under Grant 2408085J024, the National Natural Science Foundation of China under Grant 62401532, and the Anhui Provincial Key Research and Development Plan under Grant 202304a05020072.", + "bbox": [ + 89, + 90, + 485, + 175 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 205, + 187, + 220 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kobus Barnard, Vlad Cardei, and Brian Funt. A comparison of computational color constancy algorithms. i: Methodology and experiments with synthesized data. IEEE transactions on Image Processing, 11(9):972-984, 2002. 2", + "[2] Jonathan T Barron. Convolutional color constancy. In Proceedings of the IEEE International Conference on Computer Vision, pages 379-387, 2015. 2, 3", + "[3] Jonathan T Barron and Yun-Ta Tsai. Fast fourier color constancy. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 886-894, 2017. 1, 2, 3, 6", + "[4] Shida Beigpour, Christian Riess, Joost Van De Weijer, and Eli Angelopoulou. Multi-illuminant estimation with conditional random fields. IEEE Transactions on Image Processing, 23(1):83–96, 2013. 2, 6", + "[5] Simone Bianco and Raimondo Schettini. Adaptive color constancy using faces. IEEE transactions on pattern analysis and machine intelligence, 36(8):1505-1518, 2014. 2", + "[6] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Color constancy using cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 81-89, 2015. 2", + "[7] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Single and multiple illuminant estimation using convolutional neural networks. IEEE Transactions on Image Processing, 26(9):4347-4362, 2017. 2, 6, 8", + "[8] Michael Bleier, Christian Riess, Shida Beigpour, Eva Eibenberger, Eli Angelopoulou, Tobias Tröger, and André Kaup. Color constancy and non-uniform illumination: Can existing algorithms work? In 2011 IEEE international conference on computer vision workshops (ICCV Workshops), pages 774-781. IEEE, 2011. 2", + "[9] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 2", + "[10] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 1, 2", + "[11] Dongliang Cheng, Dilip K Prasad, and Michael S Brown. Illuminant estimation for color constancy: why spatial-domain methods work and the role of the color distribution. JOSA A, 31(5):1049-1058, 2014. 1, 2", + "[12] Ilija Domislovic, Donik Vrsnak, Marko Subasic, and Sven Loncarić. Outdoor daytime multi-illuminant color constancy. In 2021 12th International Symposium on Image and Signal Processing and Analysis (ISPA), pages 270–275. IEEE, 2021. 6" + ], + "bbox": [ + 93, + 229, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Iija Domislovic, Donik Vršnak, Marko Subašić, and Sven Lončarić. Shadows & lumination: Two-illuminant multiple cameras color constancy dataset. Expert systems with applications, 224:120045, 2023. 6", + "[14] Ilija Domislovic, Donik Vrsnjak, Marko Subasic, and Sven Loncarić. Color constancy for non-uniform illumination estimation with variable number of illuminants. Neural Computing and Applications, 35(20):14825-14835, 2023. 3", + "[15] Graham D Finlayson and Steven D Hordley. Color constancy at a pixel. JOSA A, 18(2):253-264, 2001. 3", + "[16] Graham D Finlayson and Elisabetta Trezzi. Shades of gray and colour constancy. In Color and Imaging Conference, pages 37-41. Society of Imaging Science and Technology, 2004. 2", + "[17] Graham D. Finlayson, Steven D. Hordley, and Paul M. Hubel. Color by correlation: A simple, unifying framework for color constancy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(11):1209-1221, 2001. 2", + "[18] David A Forsyth. A novel algorithm for color constancy. International Journal of Computer Vision, 5(1):5-35, 1990. 1, 2", + "[19] David H Foster. Color constancy. Vision research, 51(7): 674-700, 2011. 1", + "[20] Shaobing Gao, Wangwang Han, Kaifu Yang, Chaoyi Li, and Yongjie Li. Efficient color constancy with local surface reflectance statistics. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 158-173. Springer, 2014. 2", + "[21] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Generalized gamut mapping using image derivative structures for color constancy. International Journal of Computer Vision, 86:127-139, 2010. 1, 2", + "[22] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Computational color constancy: Survey and experiments. IEEE transactions on image processing, 20(9):2475-2489, 2011. 1", + "[23] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):918-929, 2011. 1, 2, 8", + "[24] Arjan Gijsenij, Rui Lu, and Theo Gevers. Color constancy for multiple light sources. IEEE Transactions on image processing, 21(2):697-707, 2011. 2, 6", + "[25] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE transactions on pattern analysis and machine intelligence, 35(6):1397-1409, 2012. 5", + "[26] Eugene Hsu, Tom Mertens, Sylvain Paris, Shai Avidan, and Frédo Durand. Light mixture estimation for spatially varying white balance. In ACM SIGGRAPH 2008 papers, pages 1-7. 2008. 2", + "[27] Yuanming Hu, Baoyuan Wang, and Stephen Lin. Fc4: Fully convolutional color constancy with confidence-weighted pooling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4085-4094, 2017. 1, 2", + "[28] Zhuo Hui, Aswin C Sankaranarayanan, Kalyan Sunkavalli, and Sunil Hadap. White balance under mixed illumination" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26428", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "using flash photography. In 2016 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2016. 2", + "[29] Md Akmol Hussain and Akbar Sheikh Akbari. Color constancy algorithm for mixed-illuminant scene images. IEEE Access, 6:8964–8976, 2018. 6", + "[30] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 8", + "[31] Dongyoung Kim, Jinwoo Kim, Seonghyeon Nam, Dongwoo Lee, Yeonkyung Lee, Nahiup Kang, Hyong-Euk Lee, ByungIn Yoo, Jae-Joon Han, and Seon Joo Kim. Large scale multi-illuminant (lsmi) dataset for developing white balance algorithm under mixed illumination. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2410–2419, 2021. 1, 2, 5, 6, 8", + "[32] Dongyoung Kim, Jinwoo Kim, Junsang Yu, and Seon Joo Kim. Attentive illumination decomposition model for multi-illuminant white balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25512-25521, 2024. 3, 7, 8", + "[33] Edwin H Land. The retinax theory of color vision. In Proceedings of the Royal Institution of Great Britain, page 23, 1974. 1, 2", + "[34] Shuwei Li, Jikai Wang, Michael S Brown, and Robby T Tan. Transc: Transformer-based multiple illuminant color constancy using multitask learning. arXiv preprint arXiv:2211.08772, 1(2):3, 2022. 1", + "[35] Yongjie Liu and Sijie Shen. Self-adaptive single and multi-illuminant estimation framework based on deep learning. arXiv preprint arXiv:1902.04705, 2019. 2", + "[36] Yi-Chen Lo, Chia-Che Chang, Hsuan-Chao Chiu, Yu-Hao Huang, Chia-Ping Chen, Yu-Lin Chang, and Kevin Jou. Clcc: Contrastive learning for color constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8053-8063, 2021. 2", + "[37] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33:11525-11538, 2020. 3", + "[38] Kantilal Varichand Mardia. Statistics of directional data. Journal of the Royal Statistical Society Series B: Statistical Methodology, 37(3):349-371, 1975. 3", + "[39] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A dataset of multi-illumination images in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4080–4089, 2019. 2", + "[40] Fatih Porikli. Integral histogram: A fast way to extract histograms in cartesian spaces. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 829-836. IEEE, 2005. 3, 4", + "[41] Yanlin Qian, Ke Chen, Jarno Nikkanen, Joni-Kristian Kamarainen, and Jiri Matas. Recurrent color constancy. In Proceedings of the IEEE international conference on computer vision, pages 5458-5466, 2017. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Yanlin Qian, Joni-Kristian Kamarainen, Jarno Nikkanen, and Jiri Matas. On finding gray pixels. In CVPR, 2019. 2", + "[43] Yanlin Qian, Jani Käpylä, Joni-Kristian Kämäräinen, Samu Koskinen, and Jiri Matas. A benchmark for burst color constancy. In ECCVW, 2020. 2", + "[44] Wu Shi, Chen Change Loy, and Xiaou Tang. Deep specialized network for illuminant estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 371-387. Springer, 2016. 6", + "[45] Oleksii Sidorov. Conditional gans for multi-illuminant color constancy: Revolution or yet another approach? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 1", + "[46] Oleksii Sidorov. Conditional gans for multi-illuminant color constancy: Revolution or yet another approach? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2", + "[47] Joost Van De Weijer, Theo Gevers, and Arjan Gijsenij. Edge-based color constancy. IEEE Transactions on image processing, 16(9):2207-2214, 2007. 2", + "[48] Bolei Xu, Jingxin Liu, Xianxu Hou, Bozhi Liu, and Guoping Qiu. End-to-end illuminant estimation based on deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3616-3625, 2020. 2" + ], + "bbox": [ + 514, + 90, + 903, + 458 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26429", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Integral Fast Fourier Color Constancy/9fd420e6-b4fa-48bd-940e-f8dcc99e9a7d_model.json b/2025/Integral Fast Fourier Color Constancy/9fd420e6-b4fa-48bd-940e-f8dcc99e9a7d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..23f7cc5df5840d8b0a6ae82203eeec0de34da766 --- /dev/null +++ b/2025/Integral Fast Fourier Color Constancy/9fd420e6-b4fa-48bd-940e-f8dcc99e9a7d_model.json @@ -0,0 +1,2728 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.303, + 0.131, + 0.695, + 0.154 + ], + "angle": 0, + "content": "Integral Fast Fourier Color Constancy" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.179, + 0.806, + 0.235 + ], + "angle": 0, + "content": "Wenjun Wei\\(^{1}\\) * Yanlin Qian\\(^{2}\\) * Huaian Chen\\(^{1\\dagger}\\) Junkang Dai\\(^{1}\\) Yi Jin\\(^{1\\dagger}\\) \n\\(^{1}\\)University of Science and Technology of China \\(^{2}\\)DJI Technology Co., Ltd \n{natstand, anchen, junkangdai}@mail.ustc.edu.cn honza.qian@dji.com" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.485, + 0.664 + ], + "angle": 0, + "content": "Traditional auto white balance (AWB) algorithms typically assume a single global illuminant source, which leads to color distortions in multi-illuminant scenes. While recent neural network-based methods have shown excellent accuracy in such scenarios, their high parameter count and computational demands limit their practicality for real-time video applications. The Fast Fourier Color Constancy (FFCC) algorithm was proposed for single-illuminant-source scenes, predicting a global illuminant source with high efficiency. However, it cannot be directly applied to multi-illuminant scenarios unless specifically modified. To address this, we propose Integral Fast Fourier Color Constancy (IFFCC), an extension of FFCC tailored for multi-illuminant scenes. IFFCC leverages the proposed integral UV histogram to accelerate histogram computations across all possible regions in Cartesian space and parallelizes Fourier-based convolution operations, resulting in a spatially-smooth illumination map. This approach enables high-accuracy, real-time AWB in multi-illuminant scenes. Extensive experiments show that IFFCC achieves accuracy that is on par with or surpasses that of pixel-level neural networks, while reducing the parameter count by over \\(400\\times\\) and processing speed by \\(20 - 100\\times\\) faster than network-based approaches." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.691, + 0.222, + 0.707 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.716, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Color constancy is a fundamental property of the human visual system, allowing for consistent color perception despite changes in illumination chromaticity. While this ability may seem \"natural\" to the human eye and visual cortex, replicating it in computational imaging systems is a complex challenge [19, 22]. Auto white balance (AWB) in image signal processing (ISP) is thus essential for approximating human-like color constancy by estimating ambient illumination chromaticity, enabling accurate color reproduction across varied lighting conditions." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.267, + 0.907, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.565 + ], + "angle": 0, + "content": "Figure 1. Frameworks comparison between IFFCC and FFCC (top). FFCC uses a global histogram to predict a single illuminant, limiting its application in multi-illuminant scenes. In contrast, IFFCC utilizes an integral UV histogram for efficient region-specific histogram computation, generating spatially continuous illumination map through parallelized Fast Fourier convolution and smoothing to achieve high-accuracy, real-time auto white balance in multi-illuminant scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Traditional white balance methods often assume the presence of a single dominant illumination and rely on statistical algorithms [18, 21, 23, 45], such as the gray world algorithm [10] and the white patch algorithm [33] etc., to estimate the scene illumination. These algorithms may also be machine learning-based, trained on AWB benchmarks [3, 27]. However, they frequently bear local color distortion in complex multi-illuminant environments, as a single white balance adjustment is inadequate to handle varying lighting conditions. For instance, a room illuminated by both warm high pressure sodium lighting and cool window light may still fail to accurately reproduce the scene's true colors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recently, deep neural network-based algorithms have been widely applied to pixel-level multi-illuminant prediction tasks [11, 31, 34, 45], utilizing deep learning to achieve accurate illumination map estimation in complex scenes. These methods are especially effective in capturing edge-aware, spatial illumination variations. However, the high parameter counts and computational demands of these models create substantial challenges for deployment in real-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26420" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "time, resource-limited applications, such as mobile photography devices. Beyond achieving high pixel-level prediction accuracy, a practical local AWB solution from an industrial perspective should prioritize several key requirements:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.482, + 0.213 + ], + "angle": 0, + "content": "Effectiveness - The algorithm can accomplish the AWB task up to a high accuracy, in any given spatial region, ensuring robust performance in a wide range of scenarios and including different light intensities and color temperatures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.482, + 0.335 + ], + "angle": 0, + "content": "Efficiency - White balance adjustments must be fast enough to provide immediate feedback during view-finding and capture, particularly in rapidly changing lighting. Limited camera resources require AWB, along with other ISP functions like AE, AF, and color correction, to operate in real time. On resource-constrained devices like smartphones, low computational complexity and memory usage are essential, with AWB processing times ideally under 10 ms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.482, + 0.426 + ], + "angle": 0, + "content": "Thumbnail Input or Stat - Many color constancy algorithms require full-resolution, high-bit images, which is resource-intensive on pervasive mobile devices. For the purpose of real-time preview, the module for generating AWB stat is often realized in hardware (like DSP), outputting a thumbnail-sized input (e.g., \\(64 \\times 48\\), 8-bit)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.482, + 0.487 + ], + "angle": 0, + "content": "Smoothness - The algorithm should maintain stable performance under varying lighting conditions, avoiding artifacts or abrupt transitions, and ensure smooth, accurate white balance across entire images or video frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.482, + 0.638 + ], + "angle": 0, + "content": "The FFCC algorithm [3] effectively meets the requirements for global illumination scenarios but faces challenges in multi-illuminant scenes. Specifically, FFCC requires repeated histogram extraction for each target region, significantly increasing processing time and limiting its suitability for real-time applications, especially on resource-constrained devices like smartphones. Additionally, FFCC lacks mechanisms to ensure smooth spatial transitions under rapidly changing lighting, which can result in abrupt shifts in white balance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.482, + 0.701 + ], + "angle": 0, + "content": "To overcome FFCC's limitations in multi-illuminant scenarios, we propose Integral Fast Fourier Color Constancy (IFFCC), a local extension of FFCC [3] designed for white balance in complex lighting conditions, as shown in Fig." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.482, + 0.761 + ], + "angle": 0, + "content": "1. IFFCC achieves speeds \\(20 - 100 \\times\\) faster than other local methods, with a runtime of \\(5.8\\mathrm{ms}\\) for \\(64 \\times 48\\) preview images on a CPU. Building upon FFCC, IFFCC introduces several key contributions:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.765, + 0.482, + 0.825 + ], + "angle": 0, + "content": "1. IFFCC outperforms prior-arts traditional algorithms and most network-based methods in multi-light source scenarios, achieving performance comparable to the best network-based methods;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.482, + 0.901 + ], + "angle": 0, + "content": "2. IFFCC leverages the proposed integral UV histogram to accelerate histogram extraction across all possible regions in Cartesian space. It efficiently computes the histogram for any region using only a few simple arithmetic operations, ensuring both high accuracy and real-time" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.765, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.853, + 0.106 + ], + "angle": 0, + "content": "performance for multi-illuminant white balance." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.107, + 0.905, + 0.152 + ], + "angle": 0, + "content": "3. IFFCC is spatially adaptive: spatial interpolation and filtering on local estimates ensure smooth transitions while preserving distinct light boundaries and sharp gradients." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.168, + 0.655, + 0.184 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.194, + 0.719, + 0.208 + ], + "angle": 0, + "content": "2.1. Global AWB Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.216, + 0.905, + 0.442 + ], + "angle": 0, + "content": "Conventional methods have long served as foundational approaches for uniform color constancy. These techniques estimate the illuminant's color based on the statistical characteristics of a single lighting source in the image [9, 17, 18, 21, 23, 47]. Notable single-illumination examples include the Gray World assumption [10], which posits that the average color of a scene under equal-energy white illumination should appear gray. Similarly, White Patch [33] assumes that the brightest pixel serves as a hint for illumination chromaticity. Refining these assumptions to focus on local patches or higher-order gradients has led to more robust statistics-based methods, optimized for more-challenging global ambient illuminations. These methods include General Gray World [1], Gray Edge [47], Shades-of-Gray [16], and LSRS [20], among others [11, 42]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.444, + 0.906, + 0.655 + ], + "angle": 0, + "content": "In the context of machine learning-based methods, more machine learning-based methods have been proposed, such as learning-based kernels [2, 3], convolutional feature extraction [6, 27], and other machine learning approaches have been devised specifically for handling single-illumination scenarios [35, 36, 41, 43, 48]. The most closely related methods to ours are Barron's CCC [2] and FFCC [3]. CCC reformulates color constancy as a color localization task in log-chromaticity space, using convolutional kernels to accurately locate the correct color. FFCC builds on this by applying the Fast Fourier Transform on a chromaticity torus, greatly improving the speed and accuracy of chromaticity estimation. However, both are global methods and lack direct applicability to local scenes unless explicitly provided with a local window." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.667, + 0.71, + 0.682 + ], + "angle": 0, + "content": "2.2. Local AWB Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.689, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Several multi-illuminant benchmarks have been introduced [4, 5, 8, 24, 31, 35, 39], with various methods proposed to tackle local white balance. These methods often rely on additional prior information or metadata, such as the number of light sources [4, 26] or facial regions of interest [5]. For instance, [28] uses flash photography to achieve local white balance. A novel framework in [4] treats the problem as an energy minimization task within a conditional random field, leveraging local light source estimates. [8] employs a global method with spatial variation, while [7] introduces a CNN framework based on sliding windows. Additionally, [46] applies a generative adversarial network (GAN) to correct images using synthetic datasets, adjusting the color based on the lighting information. More recently, Domislovic et" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26421" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "al. [14] proposed extracting spatial features for illumination prediction on each patch, and AID [32] introduces a deep model utilizing slot attention [37] to generate chromaticity and weight maps for each light source, with the added benefit of allowing user edits." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.169, + 0.483, + 0.424 + ], + "angle": 0, + "content": "Among these approaches, most focus on pixel-level predictions, often leveraging encoder-decoder architectures (e.g., U-Net) to capture fine-grained illumination chromaticity details. However, this strategy presents several practical challenges. First, pixel-level prediction is computationally intensive, especially for high-resolution images, which significantly limits real-time applicability. Second, due to the lack of spatial context, pixel-wise color correction can overlook critical illumination chromaticity differences, leading to the loss of smooth illumination transitions and distortions in color edges, both of which are essential for accurate human visual perception. By treating each pixel independently, these methods risk introducing inconsistencies across the image and erasing subtle but important color shifts. As a result, their effectiveness in real-world applications, particularly in resource-constrained environments such as mobile devices, remains significantly limited." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.436, + 0.182, + 0.451 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.581 + ], + "angle": 0, + "content": "We propose the Integral Fast Fourier Color Constancy (IFFCC) model, which builds upon the foundations of Fast Fourier Color Constancy (FFCC) [3]. Drawing inspiration from [40], we implement an histogram search for all possible target regions in image data within Cartesian coordinate space. This enables efficient retrieval of any region's histogram in linear computation time, avoiding redundant summation operations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.717 + ], + "angle": 0, + "content": "FFCC Foundation. Consider a linear image stat \\( I \\) obtained from the camera, where there are no saturated pixels and the black level has been subtracted. Assuming the RGB value of any pixel is the product of its true albedo RGB value \\( W \\) and the uniform illumination value shared across all pixels: \\( I = W \\times L \\). Although this assumption disregards shadows, dichromatic reflections, and spatial variations, it remains effective and widely accepted. Our aim is to estimate the \\( L \\) given \\( I \\), and then \\( W = I / L \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.748 + ], + "angle": 0, + "content": "For an input RGB image \\(I\\), CCC [2] defines two log-chroma measures for each pixel \\(i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.755, + 0.483, + 0.777 + ], + "angle": 0, + "content": "\\[\nu ^ {(i)} = \\log \\left(I _ {g} ^ {(i)} / I _ {r} ^ {(i)}\\right) v ^ {(i)} = \\log \\left(I _ {g} ^ {(i)} / I _ {b} ^ {(i)}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.782, + 0.483, + 0.858 + ], + "angle": 0, + "content": "Neglecting the absolute scaling of \\( W \\), the estimation of \\( L \\) is then simplified as \\( L_{u} = \\log (L_{g} / L_{r}) \\quad L_{v} = \\log (L_{g} / L_{b}) \\). Given the unknown absolute scale, the inverse mapping from RGB to UV is challenging. To simplify the recovery process, we assume \\( L \\) is the unit-norm:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.865, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {r} = \\exp (- L _ {u}) / z \\quad L _ {g} = 1 / z \\quad L _ {b} = \\exp (- L _ {v}) / z \\tag {2} \\\\ z = \\sqrt {\\exp (- L _ {u}) ^ {2} + \\exp (- L _ {v}) ^ {2} + 1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "This logarithmic chromatic space method has several advantages over the RGB approach. First, it involves only two unknowns instead of three, which enhances numerical stability. Additionally, this framework transforms the multiplicative constraints related to \\( W \\) and \\( I \\) into additive constraints [15], effectively converting the color constancy task into a positioning task in a two-dimensional space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.905, + 0.289 + ], + "angle": 0, + "content": "Following FFCC, we use FFT-based convolution in the log-chroma space, leveraging its periodicity to accelerate computations. Given the histogram's limited size, which does not fully capture the range of colors in natural images, a modular algorithm is needed to allow pixels to \"wrap around\" in this space:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.3, + 0.905, + 0.362 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} H (x, y) = \\sum_ {i} \\left(\\mod \\left((u ^ {(i)} - u _ {s}) / b - x, n\\right) < 1 \\right. \\tag {3} \\\\ \\wedge \\mod \\left(\\left(v ^ {(i)} - v _ {s}\\right) / b - y, n\\right) < 1\\bigg) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.905, + 0.554 + ], + "angle": 0, + "content": "where \\( H(x,y) \\) represents the count of pixels in \\( I \\) whose log-chroma coordinates are close to the \\( (u,v) \\) coordinates associated with the histogram position \\( (x,y) \\), \\( n \\) is the number of bins, and \\( b \\) is the bin size. \\( u_{s} \\) and \\( v_{s} \\) denote the starting positions of \\( u \\) and \\( v \\), respectively. The operations here differ from those in CCC. In this case, the \\( (x,y) \\) positions in the histogram no longer represent absolute colors, but rather a set of \\( (u,v) \\) color coordinates. The filtered results produce a set of light sources, from which we apply the \"gray light de-aliasing\" method to resolve the resulting light source clusters. This method assumes that the light sources are as close as possible to the center of the histogram." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.906, + 0.66 + ], + "angle": 0, + "content": "FFCC employs the Differentiable Bivariate von Mises (BVM) [38] method to address the single estimation of light sources from the toroidal PDF. Unlike FFCC, we need to predict multiple light sources for each image. Therefore, the BVM method is parallelized across multiple windows to predict several light sources simultaneously, as described in Section 3.3. The final convolutional structure is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.674, + 0.905, + 0.705 + ], + "angle": 0, + "content": "\\[\nP = \\operatorname {s o f t m a x} \\left(B + M \\circ \\sum_ {s} H _ {s} * F _ {s}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.717, + 0.905, + 0.763 + ], + "angle": 0, + "content": "where \\(B\\) is the bias map, \\(M\\) is the learnable gain map, and \\(\\{F_s\\}\\) is the set of learnable filters convolved with the edge-enhanced histogram group \\(\\{H_s\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.729, + 0.789 + ], + "angle": 0, + "content": "3.1. Integral UV Histogram" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "FFCC predicts a global illumination for the image. However, when multiple illuminations are anticipated for different target regions, the process requires FFCC to repeatedly extract histograms from these various areas, significantly increasing computational load and processing time. Inspired by the work of [40], we efficiently compute the histograms for all potential regions within a given image" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26422" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.089, + 0.254, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.206, + 0.253, + 0.216 + ], + "angle": 0, + "content": "(a) Integral UV histogram" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.099, + 0.38, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.204, + 0.373, + 0.216 + ], + "angle": 0, + "content": "(b) Overlapping grid" + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.101, + 0.505, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.204, + 0.494, + 0.216 + ], + "angle": 0, + "content": "(c) Local histogram" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.089, + 0.625, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.204, + 0.622, + 0.216 + ], + "angle": 0, + "content": "(d) Parallel prediction" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.1, + 0.752, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.204, + 0.749, + 0.216 + ], + "angle": 0, + "content": "(e) Spatial smoothing" + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.1, + 0.883, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.204, + 0.869, + 0.216 + ], + "angle": 0, + "content": "(f) Output image" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.23, + 0.907, + 0.329 + ], + "angle": 0, + "content": "Figure 2. Overview of the IFFCC process for predicting multi-illuminant scenes. We precompute the integral UV histogram of the input image in the log-chroma space (a). Multiple target regions are extracted on the integrated UV histogram with the overlapping grid (b), and the local log-chroma histogram (c) corresponding to each region is computed using simple arithmetic operations on the UV histogram. Each histogram forms a torus that can be unfolded, producing a cluster of potential light sources. A de-aliasing step selects the most likely light source for each region (d), represented by a point within an ellipse, where the ellipse indicates IFFCC's covariance output. Final light estimates are smoothed through interpolation and guided filtering, creating a continuous illumination map (e) that combines with the input to produce a white-balanced image (f) for multi-illuminant scenes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.355, + 0.484, + 0.55 + ], + "angle": 0, + "content": "using Cartesian coordinates. By leveraging the spatial positions of data in this coordinate system, we propagate a function starting from the origin to navigate through the remaining positions. The histogram for the current point is then updated based on the histograms of previously processed points. Once the integral histogram for each data point is obtained, the histogram for the target region can be derived using simple addition and subtraction operations. This method eliminates the need to reconstruct separate histograms for each region individually. As a result, this efficient approach significantly reduces computational overhead and improves processing time, enabling rapid analysis of multiple areas within the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.582 + ], + "angle": 0, + "content": "After transforming the image into log-chroma space by Eq.1, we define the integral uv histogram as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.591, + 0.483, + 0.63 + ], + "angle": 0, + "content": "\\[\nH _ {\\text {I n t e g r a l}} \\left(x ^ {n}, b\\right) = \\bigcup_ {m = 0} ^ {n} B (u, v) ^ {m} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.638, + 0.483, + 0.707 + ], + "angle": 0, + "content": "\\[\nB (u, v) = \\left( \\begin{array}{l} \\mod \\left(\\operatorname {r o u n d} \\left(\\frac {u - u _ {s}}{b}\\right), n\\right), \\\\ \\mod \\left(\\operatorname {r o u n d} \\left(\\frac {v - v _ {s}}{b}\\right), n\\right) \\end{array} \\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.484, + 0.803 + ], + "angle": 0, + "content": "where \\( x^n \\) is a sequence of points \\( x^0, x^1, \\ldots, x^n \\). \\( B(\\cdot) \\) represents the corresponding bin for the current data point, \\( \\bigcup \\) denotes the union operator, and the value of bin \\( b \\) in \\( H_{\\mathrm{Integral}}(x^n, b) \\) is the sum of the histogram values of bin \\( b \\) from all previously traversed data points. The integral histogram can also be expressed recursively as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.816, + 0.483, + 0.838 + ], + "angle": 0, + "content": "\\[\nH _ {\\text {I n t e g r a l}} \\left(x ^ {m}, b\\right) = H _ {\\text {I n t e g r a l}} \\left(x ^ {m - 1}, b\\right) \\bigcup B (u, v) ^ {m} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.414, + 0.871 + ], + "angle": 0, + "content": "where at the initial data point, \\( H_{\\mathrm{Integral}}(0,b) = 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.872, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In log-chroma space, updating the histogram at the current point requires the histograms of the data points above," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.49 + ], + "angle": 0, + "content": "the left, and the top-left of the current point. These three histograms have already been calculated in previous scans, so only three simple arithmetic operations are needed to update the histogram for the current point. Afterward, the bin value \\( b \\) for the current point is calculated, and the histogram is updated accordingly. Before each propagation step, all scanned bin values are copied to the current bin values, which can be efficiently handled through fast hardware or pointer operations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.906, + 0.536 + ], + "angle": 0, + "content": "For an RGB image converted to \\( N_{1} \\times N_{2} \\times 2 \\) log-chroma space data, the integral histogram propagated via wavefront scanning from the top-left starting point can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.542, + 0.905, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} H _ {\\text {I n t e g r a l}} (u, v, b) = H _ {\\text {I n t e g r a l}} (u - 1, v, b) \\\\ + H _ {\\text {I n t e g r a l}} (u, v - 1, b) \\tag {8} \\\\ - H _ {\\text {I n t e g r a l}} (u - 1, v - 1, b) + B (u, v) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.906, + 0.693 + ], + "angle": 0, + "content": "where \\((u,v)\\) is the coordinate of the log-chroma space. In our method, the data in log-chroma space can be represented by a two-dimensional floating point tensor of a two-dimensional scene [40], allowing us to calculate the computational load ratio between the traditional histogram method and the integral histogram method as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.698, + 0.905, + 0.731 + ], + "angle": 0, + "content": "\\[\nR = \\frac {\\left[ 2 1 7 M _ {1} M _ {2} + 4 B ^ {2} \\right] S _ {1} S _ {2}}{2 7 4 + 1 1 (6 8 + 4 B ^ {2}) S _ {1} S _ {2}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.906, + 0.827 + ], + "angle": 0, + "content": "where \\( M_{1} \\) and \\( M_{2} \\) represent the size of the target region in the two-dimensional scene, \\( B \\) is the number of bins, and \\( S_{1} \\) and \\( S_{2} \\) are the different scales used when extracting histograms. The result of the calculation for \\( R \\) when \\( M_{1} = M_{2} = 128 \\), \\( B = 64 \\), and \\( S_{1} = S_{2} = 1 \\) is approximately \\( R = 19.71 \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.691, + 0.85 + ], + "angle": 0, + "content": "3.2. Overlapping Grid" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "During inference, we use a sliding window with overlap to efficiently retrieve histograms for sub-regions in the log-chroma space, as illustrated in Fig.2. By pre-computing an" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26423" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.228 + ], + "angle": 0, + "content": "integral histogram over the entire scene, histograms for any arbitrary sub-region can be directly retrieved using Eq.8, without reconstructing each one from scratch, which significantly reduces computational costs. This overlapping grid method captures local chromatic variations by sliding a defined window across the scene and calculating histograms incrementally. The overlapping regions between windows maintain continuity and minimize edge effects, enhancing the accuracy of chromatic analysis throughout the scene." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.235, + 0.273, + 0.25 + ], + "angle": 0, + "content": "3.3. Parallel Prediction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.257, + 0.484, + 0.439 + ], + "angle": 0, + "content": "To speed up inference, our method requires only one inference to predict several illumination across multiple target regions within the image. Given an image's integral UV histogram in log-chroma space, denoted as \\( H_{\\mathrm{integral}}(H,W,B,B,N) \\), where \\( H \\) and \\( W \\) represent the image size, \\( B \\) is the number of bins, and \\( N \\) is the number of augmented channel's histogram. The overlapping grid approach can be applied to obtain a collection \\( X \\) of \\( k \\) target region histograms. This collection \\( X \\), represented as \\( X(k,B,B,N) \\), includes histograms for each specific target region in the log-chroma space. The process of calculating the PDF of \\( X \\) in the frequency domain is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.446, + 0.483, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} H = \\operatorname {I F F T} \\left(\\sum_ {k} \\operatorname {F F T} (X) \\cdot \\operatorname {F F T} (F)\\right) + B (10) \\\\ P = \\frac {\\exp^ {(H - \\max (H))}}{\\sum_ {\\mathrm {d i m s}} \\exp^ {(H - \\max (H))}} (11) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.483, + 0.578 + ], + "angle": 0, + "content": "where \\(\\mathrm{FFT}(.)\\) and \\(\\mathrm{IFFT}(.)\\) are the fast Fourier transform and its inverse, respectively, \\(F\\) is the convolution kernel, and \\(B\\) is the bias map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.484, + 0.7 + ], + "angle": 0, + "content": "To parallelize the estimation of the mean of the Bivariate Von Mises (BVM) distribution from a histogram, we calculate circular mean values for the histogram coordinates \\(i\\) and \\(j\\) across multiple regions concurrently. This process enables effective mean direction estimation within the toroidal data space, essential for capturing periodic patterns in wrapped grid data. For each region \\(k\\), we define \\(\\mu = \\{\\mu_k\\}_{k=1}^K\\), where each \\(\\mu_k\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.705, + 0.483, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mu_ {k} = \\left[ \\begin{array}{l} u _ {s} \\\\ v _ {s} \\end{array} \\right] + b \\left[ \\begin{array}{l} \\bmod \\left(\\frac {n}{2 \\pi} \\arctan 2 \\left(y _ {1} ^ {k}, x _ {1} ^ {k}\\right), n\\right) \\\\ \\bmod \\left(\\frac {n}{2 \\pi} \\arctan 2 \\left(y _ {2} ^ {k}, x _ {2} ^ {k}\\right), n\\right) \\end{array} \\right] \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.483, + 0.777 + ], + "angle": 0, + "content": "where the components \\( y_1^k, x_1^k, y_2^k \\), and \\( x_2^k \\) are computed in parallel for each \\( k \\):" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.784, + 0.483, + 0.829 + ], + "angle": 0, + "content": "\\[\ny _ {1} ^ {k} = \\sum_ {i} P _ {i} ^ {k} \\sin \\left(\\frac {2 \\pi i}{n}\\right), x _ {1} ^ {k} = \\sum_ {i} P _ {i} ^ {k} \\cos \\left(\\frac {2 \\pi i}{n}\\right) \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.824, + 0.445, + 0.862 + ], + "angle": 0, + "content": "\\[\ny _ {2} ^ {k} = \\sum_ {j} P _ {j} ^ {k} \\sin \\left(\\frac {2 \\pi j}{n}\\right), x _ {2} ^ {k} = \\sum_ {j} P _ {j} ^ {k} \\cos \\left(\\frac {2 \\pi j}{n}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\( P_{i}^{k} \\) and \\( P_{j}^{k} \\) represent the marginal distributions for rows and columns of each target region's 2D histogram." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "After concurrently obtaining illumination values for all \\(k\\) target regions, we reconstruct the full illumination map by interpolating or applying edge-preserving filtering methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.146, + 0.693, + 0.163 + ], + "angle": 0, + "content": "3.4. Spatial Smoothing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.906, + 0.243 + ], + "angle": 0, + "content": "In multi-light scenarios, patch-wise illumination estimates often exhibit discrete, locally discontinuous characteristics. To reconstruct a smooth and seamless illumination map across the entire image, it is vital to postprocess these patch-level estimates using interpolation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.907, + 0.409 + ], + "angle": 0, + "content": "We employ linear interpolation due to its simplicity and ability to create smooth, piecewise-linear transitions for estimating intermediate values. Although higher-degree polynomial interpolation could be used, it risks introducing oscillations. After getting estimates for \\( k \\) target regions, interpolation adjusts these values based on their relative locations, resulting in a smooth illumination map. Although interpolation enhances overall smoothness, it may blur local details, particularly in regions with strong light edges or steep illumination gradients, potentially leading to a loss of edge information." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.409, + 0.906, + 0.545 + ], + "angle": 0, + "content": "Guided filtering [25] is then applied to preserve details at the boundaries and structural features of the illumination map, preventing blur or distortion during transitions. Overall, combining these techniques results in a globally coherent illumination distribution that maintains edge clarity while appearing smoother and more natural. Given the interpolated illumination map \\( L \\) and the raw image \\( G \\) as the guiding image, the output illumination map \\( O \\) at each pixel \\( O_{i} \\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.555, + 0.905, + 0.571 + ], + "angle": 0, + "content": "\\[\nO _ {i} = a _ {k} G _ {i} + b _ {k}, \\quad \\forall i \\in \\omega_ {k} \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.578, + 0.905, + 0.617 + ], + "angle": 0, + "content": "\\[\na _ {k} = \\frac {\\frac {1}{| \\omega |} \\sum_ {i \\in \\omega_ {k}} G _ {i} L _ {i} - \\bar {G} _ {k} \\bar {L} _ {k}}{\\sigma_ {k} ^ {2} + \\epsilon}, \\quad b _ {k} = \\bar {L} _ {k} - a _ {k} \\bar {G} _ {k} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.621, + 0.906, + 0.682 + ], + "angle": 0, + "content": "here, \\(\\bar{G}_k\\) and \\(\\bar{L}_k\\) are the mean values of \\(G\\) and \\(L\\) in window \\(\\omega_k\\), with \\(\\sigma_k^2\\) as the variance of \\(G\\). The constant \\(\\epsilon\\) prevents division by zero. The final illumination map \\(O\\) is obtained by weighted averaging." + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.691, + 0.905, + 0.728 + ], + "angle": 0, + "content": "\\[\nO = \\frac {1}{| \\omega |} \\sum_ {k \\in \\omega_ {i}} \\left(a _ {k} G _ {i} + b _ {k}\\right) \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.906, + 0.812 + ], + "angle": 0, + "content": "This post-processing step is essential: it creates a coherent illumination map that accurately reflects the lighting conditions across the entire image while preserving critical features, thus the reconstructed illumination aligns well with the underlying content of the image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.64, + 0.841 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.704, + 0.866 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We evaluated our proposed IFFCC on two widely used multi-light source datasets: the LSMI dataset [31] and the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26424" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.088, + 0.885, + 0.248 + ], + "angle": 0, + "content": "
MethodCanon_5dCanon_550dMotoPanasonicSonyUse all
meanmedianmeanmedianmeanmedianmeanmedianmeanmedianmeanmedian
Huassin and Akbari [29]13.8813.7613.7313.5413.0612.8913.6213.7314.2213.6813.6713.46
CRF(White-Patch)[4]7.665.967.695.776.945.327.585.697.905.867.195.44
Patch-based[24]4.853.114.612.984.172.494.582.954.963.164.302.89
Keypoint-based[24]5.593.885.483.944.973.285.423.795.763.825.463.59
Superpixel-based[24]4.883.684.793.564.192.894.553.544.973.824.203.10
IFFCC2.061.542.161.771.901.362.061.592.661.962.191.56
Bianco et al.[7]4.634.024.594.164.273.894.984.565.344.768.015.69
HypNet/SelNet[44]5.424.935.164.835.094.745.234.985.825.346.313.95
LSMI-U[31]3.142.332.922.462.792.043.172.703.582.883.722.40
Domislovic et al.[12]2.632.182.772.062.071.492.381.732.922.242.281.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.257, + 0.908, + 0.314 + ], + "angle": 0, + "content": "Table 1. The performance results on the Shadow dataset are presented. The dataset is segmented by images captured from different cameras, with \"use all\" representing the entire dataset. We report the mean and median angular errors for each subset. IFFCC employs the default settings, with a window size of 128 and an overlap size of 64. The model is trained using the blended illumination within each window as the ground truth (GT), and the corresponding error is computed." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.326, + 0.885, + 0.487 + ], + "angle": 0, + "content": "
MethodOutdoorIndoorNighttime
meanmedianb25w25meanmedianb25w25meanmedianb25w25
Huassin and Akbari[29]13.1912.897.1019.1214.8413.228.4320.1816.3514.969.7522.33
CRF(White-Patch)[4]7.315.662.3113.368.786.894.3614.406.965.162.3815.85
Patch-based[24]4.533.291.3810.475.584.432.0911.393.482.171.8912.53
Keypoint-based[24]5.773.881.2411.786.254.041.7912.603.994.101.4913.08
Superpixel-based[24]4.733.011.338.215.183.491.929.244.632.981.349.82
IFFCC1.741.280.423.833.042.140.666.712.851.530.547.43
Bianco et al.[7]4.862.361.9811.757.425.623.3812.508.225.783.1013.68
HypNet/SelNet[44]6.094.081.7712.317.075.272.9811.494.763.212.2012.59
LSMI-U[31]2.122.031.025.485.724.621.498.544.812.471.338.89
Domislovic et al.[12]1.781.460.654.344.453.310.987.463.681.890.838.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.496, + 0.908, + 0.553 + ], + "angle": 0, + "content": "Table 2. The performance results on the Shadow dataset are presented. The dataset is segmented by images captured in different scenes, with each scene including images taken by five different cameras. We report the mean and median angular errors for each subset. IFFCC uses the default settings, with a window size of 128 and an overlap size of 64. The model is trained using the blended illumination within each window as the ground truth (GT), and the corresponding error is computed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.483, + 0.73 + ], + "angle": 0, + "content": "Shadow dataset [13]. The LSMI dataset includes over 7,486 images captured by three different cameras in a variety of scenes, each containing multiple light sources. The Shadow dataset consists of 2,500 images taken in both indoor and outdoor settings, with each image accompanied by a binary segmentation mask, where each region is illuminated by a single light source. In all experiments, full-sized training images were randomly cropped to \\(128 \\times 128\\), and testing was conducted on \\(256 \\times 256\\) images. All images were black-level corrected, and calibration objects were masked." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.903 + ], + "angle": 0, + "content": "IFFCC's computational efficiency enables CPU-only training and inference, eliminating the need for a GPU. All experiments were conducted on an Intel Xeon Gold 6258R CPU. Performance was evaluated using standard error metrics, including mean and median angular errors and the arithmetic means of the first and third quartiles (\"best \\(25\\%\\) \" and \"worst \\(25\\%\\) ). Additionally, we compared model parameter counts and CPU inference times for \\(256\\times 256\\) images. Since our method operates on patches, we calculate blended illumination values from the ground truth illumination map within a sliding window, using these as the ground" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.905, + 0.608 + ], + "angle": 0, + "content": "truth for each patch. The impact of different training approaches is discussed in the ablation study." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.906, + 0.7 + ], + "angle": 0, + "content": "During training, we input a set of enhanced histograms derived from filtered images that ensure non-negativity and intensity scaling consistency. Each image includes a local absolute deviation measure. Using the preconditioned frequency-domain optimization from FFCC [3], training runs for 64 iterations with a histogram size of \\(64 \\times 64\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.747, + 0.728 + ], + "angle": 0, + "content": "4.2. Results and Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.906, + 0.884 + ], + "angle": 0, + "content": "Quantitative Comparison We evaluated IFFCC against three traditional methods [4, 24, 29] and three learning-based methods [7, 12, 44] on the Shadow dataset. The experimental results are presented in Tables 1 and 2. Table 1 organizes the Shadow dataset into subsets based on images captured by different cameras, while Table 2 categorizes the dataset by various scene types, each containing photos taken with five distinct cameras. Despite comparisons with advanced network-based methods, IFFCC achieves state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "The results in Table 2 highlight the strong general" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "26425" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.229, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.174, + 0.228, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.256, + 0.228, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.339, + 0.182, + 0.353 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.091, + 0.363, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.174, + 0.363, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.255, + 0.363, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.339, + 0.337, + 0.353 + ], + "angle": 0, + "content": "Gijsenij et al." + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.091, + 0.498, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.174, + 0.498, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.255, + 0.498, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.339, + 0.477, + 0.352 + ], + "angle": 0, + "content": "Bianoco et al." + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.091, + 0.632, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.174, + 0.632, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.255, + 0.632, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.339, + 0.598, + 0.351 + ], + "angle": 0, + "content": "LSMI-U" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.091, + 0.766, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.174, + 0.766, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.255, + 0.766, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.339, + 0.728, + 0.351 + ], + "angle": 0, + "content": "IFFCC" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.091, + 0.901, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.174, + 0.901, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.255, + 0.901, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.339, + 0.846, + 0.351 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.364, + 0.908, + 0.394 + ], + "angle": 0, + "content": "Figure 3. Visualization of the comparison results on the LSMI dataset. For IFFCC, the window size is set to 128 with an overlap of 64. In other patch-based methods, the window size is set to the default value of 32." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.409, + 0.256, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.488, + 0.256, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.568, + 0.195, + 0.581 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.409, + 0.418, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.488, + 0.417, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.567, + 0.377, + 0.58 + ], + "angle": 0, + "content": "White Patch" + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.409, + 0.578, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.488, + 0.578, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.567, + 0.521, + 0.58 + ], + "angle": 0, + "content": "IFFCC" + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.409, + 0.74, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.488, + 0.74, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.567, + 0.721, + 0.581 + ], + "angle": 0, + "content": "Illumination map" + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.409, + 0.902, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.488, + 0.902, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.567, + 0.892, + 0.581 + ], + "angle": 0, + "content": "White-balanced image" + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.593, + 0.901, + 0.607 + ], + "angle": 0, + "content": "Figure 4. Visualization of illumination comparison for different patch-based methods. The window size is 32, with an overlap size of 16." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.634, + 0.483, + 0.83 + ], + "angle": 0, + "content": "ization ability of IFFCC. Network-based methods generally perform worse than IFFCC on multi-camera datasets, where each camera's unique spectral sensitivities complicate learning a single mapping across different devices. IFFCC generalizes well across multiple cameras because it leverages an integral log-chroma histogram approach, which directly models chromaticity distributions without being overly reliant on the specific spectral sensitivities of individual cameras. This histogram-based method effectively captures illumination variations across different scenes and cameras, allowing IFFCC to produce robust predictions without requiring extensive fine-tuning for each camera's unique spectral characteristics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The experimental results on the LSMI dataset, presented in Tables 3, and 4, provide a comparative analysis of patch-based and pixel-level methods. The test set is categorized into Multi (multiple light sources), and Mixed (compris" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.738 + ], + "angle": 0, + "content": "ing both single and multiple light sources). Among patch-based approaches, IFFCC demonstrates state-of-the-art performance, achieving results comparable to those of leading pixel-level methods, including AID [32]. Notably, IFFCC's compact parameter count and rapid processing speed render it highly suitable for deployment on mobile and resource-constrained devices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.907, + 0.806 + ], + "angle": 0, + "content": "We also compared the processing time for a \\(256 \\times 256\\) image in a multi-illuminant scene using FFCC and IFFCC, as shown in Table 5. As the number of windows increases, IF-FCC significantly outperforms FFCC in processing speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Qualitative Comparison Fig.3 visually compares our method against alternative approaches on the LSMI dataset. To enhance visualization quality, color correction, gamma correction, and tone mapping were applied to the white-balanced images. The results indicate that IFFCC achieves more accurate white balance restoration than competing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26426" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.473, + 0.235 + ], + "angle": 0, + "content": "
MethodMultiParams (M)Time (s)
Galaxy mean med.Nikon mean med.Sony mean med.
Patch -levelGijsenij[23]12.389.579.316.5911.198.14-1.3
Bianoco[7]5.564.344.654.194.383.940.16-
IFFCC2.481.902.301.502.481.970.0120.03
Pix2Pix[30]4.282.634.182.764.373.265.40.48
Pixel LSMI-H[31]3.132.703.203.013.653.330.520.28
-levelLSMI-U[31]2.852.552.361.842.762.385.70.51
AID[32]2.031.432.261.392.161.646.4>1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.244, + 0.483, + 0.287 + ], + "angle": 0, + "content": "Table 3. The performance results on the multi-illumination images from the LSMI dataset are presented, with testing time corresponding to the processing time for \\(256\\times 256\\) images on a CPU." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.3, + 0.472, + 0.446 + ], + "angle": 0, + "content": "
MethodMixedParams (M)Time (s)
Galaxy mean med.Nikon mean med.Sony mean med.
Patch -levelGijsenij[23]10.097.439.215.839.467.13-1.3
Bianoco[7]4.893.833.933.483.843.210.16-
IFFCC1.981.792.111.481.931.720.0120.03
Pix2Pix[30]5.662.445.412.494.202.205.40.48
Pixel LSMI-H[31]3.062.542.922.633.212.850.520.28
-level LSMI-U[31]2.631.912.161.552.682.295.70.51
AID[32]1.631.321.761.141.681.446.4>1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.514 + ], + "angle": 0, + "content": "Table 4. The performance results on the mixed-illumination images (containing both single and multiple light sources) from the LSMI dataset are presented, with testing time corresponding to the processing time for \\(256 \\times 256\\) images on a CPU." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.525, + 0.48, + 0.566 + ], + "angle": 0, + "content": "
[32,0][32,16][64,32][64,48][128,64][128,96]
FFCC162ms450ms203ms689ms88ms234ms
IFFCC34ms46ms29ms40ms27ms33ms
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.633 + ], + "angle": 0, + "content": "Table 5. Comparison of average testing time for FFCC and IF-FCC on the Shadow dataset, evaluated across different window and overlap sizes. FFCC employs a conventional histogram computation approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.72 + ], + "angle": 0, + "content": "methods. We also present a comparison of illumination maps between our method and patch-based approaches, as shown in Fig.4. The window size is set to \\(32 \\times 32\\) with an overlap of 16. The results demonstrate that IFFCC achieves a more precise and smoother illumination prediction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.245, + 0.745 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Table 6 presents results from ablation experiments analyzing the impact of various training strategies, window sizes, and overlap regions. Here, \"B\" represents training with a blend of illumination maps as the ground truth (GT), and \"M\" uses the dominant light source as GT. Parameters \"ws\" and \"os\" indicate window and overlap sizes, respectively. Findings reveal that smaller window sizes, with limited information, reduce prediction accuracy. However, the blended illumination strategy provides improved accuracy compared to using a single dominant light source. Visual" + }, + { + "type": "table", + "bbox": [ + 0.564, + 0.089, + 0.865, + 0.23 + ], + "angle": 0, + "content": "
type[ws, os]meanmedianb25w25
B[32, 16]2.652.171.374.65
B[64, 0]2.352.011.134.31
M[64, 32]2.532.041.254.83
B[64, 32]2.371.871.164.38
M[128, 0]2.462.031.044.78
B[128, 0]2.161.680.914.23
M[128, 32]2.501.881.044.91
B[128, 32]2.141.640.854.30
M[128, 64]2.421.950.934.71
B[128, 64]2.061.540.814.19
B[128, 96]2.131.590.834.25
B[192, 128]1.981.340.724.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.24, + 0.907, + 0.283 + ], + "angle": 0, + "content": "Table 6. Ablation study on the Shadow dataset (Canon5d), reporting the impact of different window sizes, overlap sizes, and training strategies on accuracy." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.296, + 0.646, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.297, + 0.773, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.297, + 0.903, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.45, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Figure 5. (a)-(d) show illumination maps generated with different window sizes ([32, 16], [64, 32], [128, 64], [192, 128]), each processed with smooth transition adjustments. (e) presents the ground truth (GT) illumination map, and (f) displays the white-balanced image obtained by combining the input image with the GT." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.907, + 0.625 + ], + "angle": 0, + "content": "comparisons (Fig. 5) show that smaller windows enhance texture details in the predicted illumination maps but with lower accuracy. Conversely, larger windows (e.g., [192, 128]) reduce error metrics by capturing more information, yet they may blur boundaries in multi-illuminant regions, compromising edge clarity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.638, + 0.634, + 0.654 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.907, + 0.859 + ], + "angle": 0, + "content": "We propose IFFCC, a practical solution for achieving accurate, real-time auto white balance (AWB) in multi-illuminant environments. Built upon the Fast Fourier Color Constancy (FFCC) algorithm, IFFCC leverages an integral UV histogram and parallelized Fourier-based convolution to rapidly estimate illumination across different image regions, addressing the need for localized, adaptive AWB. Results on multiple datasets demonstrate that IFFCC delivers accuracy comparable to or surpassing pixel-level neural network methods, while reducing the parameter count by over 400 times and speeding up processing by a factor of 20–100. These advantages make IFFCC highly suitable for resource-limited, real-time applications such as video streaming." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.858, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We believe that FFCC-based methods, including IFFCC, will continue to enrich the field and attract increased attention across a broader range of ISP applications." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26427" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "Acknowledgments This work was supported in part by the Anhui Provincial Natural Science Foundation under Grant 2408085J024, the National Natural Science Foundation of China under Grant 62401532, and the Anhui Provincial Key Research and Development Plan under Grant 202304a05020072." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.206, + 0.188, + 0.221 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.23, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[1] Kobus Barnard, Vlad Cardei, and Brian Funt. A comparison of computational color constancy algorithms. i: Methodology and experiments with synthesized data. IEEE transactions on Image Processing, 11(9):972-984, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.288, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[2] Jonathan T Barron. Convolutional color constancy. In Proceedings of the IEEE International Conference on Computer Vision, pages 379-387, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[3] Jonathan T Barron and Yun-Ta Tsai. Fast fourier color constancy. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 886-894, 2017. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[4] Shida Beigpour, Christian Riess, Joost Van De Weijer, and Eli Angelopoulou. Multi-illuminant estimation with conditional random fields. IEEE Transactions on Image Processing, 23(1):83–96, 2013. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[5] Simone Bianco and Raimondo Schettini. Adaptive color constancy using faces. IEEE transactions on pattern analysis and machine intelligence, 36(8):1505-1518, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[6] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Color constancy using cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 81-89, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[7] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Single and multiple illuminant estimation using convolutional neural networks. IEEE Transactions on Image Processing, 26(9):4347-4362, 2017. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[8] Michael Bleier, Christian Riess, Shida Beigpour, Eva Eibenberger, Eli Angelopoulou, Tobias Tröger, and André Kaup. Color constancy and non-uniform illumination: Can existing algorithms work? In 2011 IEEE international conference on computer vision workshops (ICCV Workshops), pages 774-781. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[9] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[10] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[11] Dongliang Cheng, Dilip K Prasad, and Michael S Brown. Illuminant estimation for color constancy: why spatial-domain methods work and the role of the color distribution. JOSA A, 31(5):1049-1058, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[12] Ilija Domislovic, Donik Vrsnak, Marko Subasic, and Sven Loncarić. Outdoor daytime multi-illuminant color constancy. In 2021 12th International Symposium on Image and Signal Processing and Analysis (ISPA), pages 270–275. IEEE, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.23, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Iija Domislovic, Donik Vršnak, Marko Subašić, and Sven Lončarić. Shadows & lumination: Two-illuminant multiple cameras color constancy dataset. Expert systems with applications, 224:120045, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[14] Ilija Domislovic, Donik Vrsnjak, Marko Subasic, and Sven Loncarić. Color constancy for non-uniform illumination estimation with variable number of illuminants. Neural Computing and Applications, 35(20):14825-14835, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[15] Graham D Finlayson and Steven D Hordley. Color constancy at a pixel. JOSA A, 18(2):253-264, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[16] Graham D Finlayson and Elisabetta Trezzi. Shades of gray and colour constancy. In Color and Imaging Conference, pages 37-41. Society of Imaging Science and Technology, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[17] Graham D. Finlayson, Steven D. Hordley, and Paul M. Hubel. Color by correlation: A simple, unifying framework for color constancy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(11):1209-1221, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[18] David A Forsyth. A novel algorithm for color constancy. International Journal of Computer Vision, 5(1):5-35, 1990. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.904, + 0.417 + ], + "angle": 0, + "content": "[19] David H Foster. Color constancy. Vision research, 51(7): 674-700, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[20] Shaobing Gao, Wangwang Han, Kaifu Yang, Chaoyi Li, and Yongjie Li. Efficient color constancy with local surface reflectance statistics. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 158-173. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[21] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Generalized gamut mapping using image derivative structures for color constancy. International Journal of Computer Vision, 86:127-139, 2010. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[22] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Computational color constancy: Survey and experiments. IEEE transactions on image processing, 20(9):2475-2489, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[23] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):918-929, 2011. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[24] Arjan Gijsenij, Rui Lu, and Theo Gevers. Color constancy for multiple light sources. IEEE Transactions on image processing, 21(2):697-707, 2011. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[25] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE transactions on pattern analysis and machine intelligence, 35(6):1397-1409, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[26] Eugene Hsu, Tom Mertens, Sylvain Paris, Shai Avidan, and Frédo Durand. Light mixture estimation for spatially varying white balance. In ACM SIGGRAPH 2008 papers, pages 1-7. 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[27] Yuanming Hu, Baoyuan Wang, and Stephen Lin. Fc4: Fully convolutional color constancy with confidence-weighted pooling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4085-4094, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[28] Zhuo Hui, Aswin C Sankaranarayanan, Kalyan Sunkavalli, and Sunil Hadap. White balance under mixed illumination" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26428" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "using flash photography. In 2016 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.175 + ], + "angle": 0, + "content": "[29] Md Akmol Hussain and Akbar Sheikh Akbari. Color constancy algorithm for mixed-illuminant scene images. IEEE Access, 6:8964–8976, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.244 + ], + "angle": 0, + "content": "[30] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[31] Dongyoung Kim, Jinwoo Kim, Seonghyeon Nam, Dongwoo Lee, Yeonkyung Lee, Nahiup Kang, Hyong-Euk Lee, ByungIn Yoo, Jae-Joon Han, and Seon Joo Kim. Large scale multi-illuminant (lsmi) dataset for developing white balance algorithm under mixed illumination. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2410–2419, 2021. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[32] Dongyoung Kim, Jinwoo Kim, Junsang Yu, and Seon Joo Kim. Attentive illumination decomposition model for multi-illuminant white balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25512-25521, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[33] Edwin H Land. The retinax theory of color vision. In Proceedings of the Royal Institution of Great Britain, page 23, 1974. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[34] Shuwei Li, Jikai Wang, Michael S Brown, and Robby T Tan. Transc: Transformer-based multiple illuminant color constancy using multitask learning. arXiv preprint arXiv:2211.08772, 1(2):3, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[35] Yongjie Liu and Sijie Shen. Self-adaptive single and multi-illuminant estimation framework based on deep learning. arXiv preprint arXiv:1902.04705, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[36] Yi-Chen Lo, Chia-Che Chang, Hsuan-Chao Chiu, Yu-Hao Huang, Chia-Ping Chen, Yu-Lin Chang, and Kevin Jou. Clcc: Contrastive learning for color constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8053-8063, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[37] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33:11525-11538, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[38] Kantilal Varichand Mardia. Statistics of directional data. Journal of the Royal Statistical Society Series B: Statistical Methodology, 37(3):349-371, 1975. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[39] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A dataset of multi-illumination images in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4080–4089, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[40] Fatih Porikli. Integral histogram: A fast way to extract histograms in cartesian spaces. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 829-836. IEEE, 2005. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[41] Yanlin Qian, Ke Chen, Jarno Nikkanen, Joni-Kristian Kamarainen, and Jiri Matas. Recurrent color constancy. In Proceedings of the IEEE international conference on computer vision, pages 5458-5466, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[42] Yanlin Qian, Joni-Kristian Kamarainen, Jarno Nikkanen, and Jiri Matas. On finding gray pixels. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[43] Yanlin Qian, Jani Käpylä, Joni-Kristian Kämäräinen, Samu Koskinen, and Jiri Matas. A benchmark for burst color constancy. In ECCVW, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[44] Wu Shi, Chen Change Loy, and Xiaou Tang. Deep specialized network for illuminant estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 371-387. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[45] Oleksii Sidorov. Conditional gans for multi-illuminant color constancy: Revolution or yet another approach? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[46] Oleksii Sidorov. Conditional gans for multi-illuminant color constancy: Revolution or yet another approach? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[47] Joost Van De Weijer, Theo Gevers, and Arjan Gijsenij. Edge-based color constancy. IEEE Transactions on image processing, 16(9):2207-2214, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[48] Bolei Xu, Jingxin Liu, Xianxu Hou, Bozhi Liu, and Guoping Qiu. End-to-end illuminant estimation based on deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3616-3625, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.092, + 0.905, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26429" + } + ] +] \ No newline at end of file diff --git a/2025/Integral Fast Fourier Color Constancy/9fd420e6-b4fa-48bd-940e-f8dcc99e9a7d_origin.pdf b/2025/Integral Fast Fourier Color Constancy/9fd420e6-b4fa-48bd-940e-f8dcc99e9a7d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7b8ac452b150ce79fe4057df7d1ffdd79740d789 --- /dev/null +++ b/2025/Integral Fast Fourier Color Constancy/9fd420e6-b4fa-48bd-940e-f8dcc99e9a7d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f73bad3c57429c18637109ed037b7308278d804be35f7eb498fb6318e253cd73 +size 1337964 diff --git a/2025/Integral Fast Fourier Color Constancy/full.md b/2025/Integral Fast Fourier Color Constancy/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cffd92465ef04bbefe4bb8362cba8ae5281b7fbe --- /dev/null +++ b/2025/Integral Fast Fourier Color Constancy/full.md @@ -0,0 +1,416 @@ +# Integral Fast Fourier Color Constancy + +Wenjun Wei $^{1}$ * Yanlin Qian $^{2}$ * Huaian Chen $^{1\dagger}$ Junkang Dai $^{1}$ Yi Jin $^{1\dagger}$ $^{1}$ University of Science and Technology of China $^{2}$ DJI Technology Co., Ltd +{natstand, anchen, junkangdai}@mail.ustc.edu.cn honza.qian@dji.com + +# Abstract + +Traditional auto white balance (AWB) algorithms typically assume a single global illuminant source, which leads to color distortions in multi-illuminant scenes. While recent neural network-based methods have shown excellent accuracy in such scenarios, their high parameter count and computational demands limit their practicality for real-time video applications. The Fast Fourier Color Constancy (FFCC) algorithm was proposed for single-illuminant-source scenes, predicting a global illuminant source with high efficiency. However, it cannot be directly applied to multi-illuminant scenarios unless specifically modified. To address this, we propose Integral Fast Fourier Color Constancy (IFFCC), an extension of FFCC tailored for multi-illuminant scenes. IFFCC leverages the proposed integral UV histogram to accelerate histogram computations across all possible regions in Cartesian space and parallelizes Fourier-based convolution operations, resulting in a spatially-smooth illumination map. This approach enables high-accuracy, real-time AWB in multi-illuminant scenes. Extensive experiments show that IFFCC achieves accuracy that is on par with or surpasses that of pixel-level neural networks, while reducing the parameter count by over $400\times$ and processing speed by $20 - 100\times$ faster than network-based approaches. + +# 1. Introduction + +Color constancy is a fundamental property of the human visual system, allowing for consistent color perception despite changes in illumination chromaticity. While this ability may seem "natural" to the human eye and visual cortex, replicating it in computational imaging systems is a complex challenge [19, 22]. Auto white balance (AWB) in image signal processing (ISP) is thus essential for approximating human-like color constancy by estimating ambient illumination chromaticity, enabling accurate color reproduction across varied lighting conditions. + +![](images/6b156077960eef34f511ecb6e49ee0bd3c4e88c4c761d57de9dbc6749278691a.jpg) +Figure 1. Frameworks comparison between IFFCC and FFCC (top). FFCC uses a global histogram to predict a single illuminant, limiting its application in multi-illuminant scenes. In contrast, IFFCC utilizes an integral UV histogram for efficient region-specific histogram computation, generating spatially continuous illumination map through parallelized Fast Fourier convolution and smoothing to achieve high-accuracy, real-time auto white balance in multi-illuminant scenarios. + +Traditional white balance methods often assume the presence of a single dominant illumination and rely on statistical algorithms [18, 21, 23, 45], such as the gray world algorithm [10] and the white patch algorithm [33] etc., to estimate the scene illumination. These algorithms may also be machine learning-based, trained on AWB benchmarks [3, 27]. However, they frequently bear local color distortion in complex multi-illuminant environments, as a single white balance adjustment is inadequate to handle varying lighting conditions. For instance, a room illuminated by both warm high pressure sodium lighting and cool window light may still fail to accurately reproduce the scene's true colors. + +Recently, deep neural network-based algorithms have been widely applied to pixel-level multi-illuminant prediction tasks [11, 31, 34, 45], utilizing deep learning to achieve accurate illumination map estimation in complex scenes. These methods are especially effective in capturing edge-aware, spatial illumination variations. However, the high parameter counts and computational demands of these models create substantial challenges for deployment in real- + +time, resource-limited applications, such as mobile photography devices. Beyond achieving high pixel-level prediction accuracy, a practical local AWB solution from an industrial perspective should prioritize several key requirements: + +Effectiveness - The algorithm can accomplish the AWB task up to a high accuracy, in any given spatial region, ensuring robust performance in a wide range of scenarios and including different light intensities and color temperatures. + +Efficiency - White balance adjustments must be fast enough to provide immediate feedback during view-finding and capture, particularly in rapidly changing lighting. Limited camera resources require AWB, along with other ISP functions like AE, AF, and color correction, to operate in real time. On resource-constrained devices like smartphones, low computational complexity and memory usage are essential, with AWB processing times ideally under 10 ms. + +Thumbnail Input or Stat - Many color constancy algorithms require full-resolution, high-bit images, which is resource-intensive on pervasive mobile devices. For the purpose of real-time preview, the module for generating AWB stat is often realized in hardware (like DSP), outputting a thumbnail-sized input (e.g., $64 \times 48$ , 8-bit). + +Smoothness - The algorithm should maintain stable performance under varying lighting conditions, avoiding artifacts or abrupt transitions, and ensure smooth, accurate white balance across entire images or video frames. + +The FFCC algorithm [3] effectively meets the requirements for global illumination scenarios but faces challenges in multi-illuminant scenes. Specifically, FFCC requires repeated histogram extraction for each target region, significantly increasing processing time and limiting its suitability for real-time applications, especially on resource-constrained devices like smartphones. Additionally, FFCC lacks mechanisms to ensure smooth spatial transitions under rapidly changing lighting, which can result in abrupt shifts in white balance. + +To overcome FFCC's limitations in multi-illuminant scenarios, we propose Integral Fast Fourier Color Constancy (IFFCC), a local extension of FFCC [3] designed for white balance in complex lighting conditions, as shown in Fig. + +1. IFFCC achieves speeds $20 - 100 \times$ faster than other local methods, with a runtime of $5.8\mathrm{ms}$ for $64 \times 48$ preview images on a CPU. Building upon FFCC, IFFCC introduces several key contributions: + +1. IFFCC outperforms prior-arts traditional algorithms and most network-based methods in multi-light source scenarios, achieving performance comparable to the best network-based methods; +2. IFFCC leverages the proposed integral UV histogram to accelerate histogram extraction across all possible regions in Cartesian space. It efficiently computes the histogram for any region using only a few simple arithmetic operations, ensuring both high accuracy and real-time + +performance for multi-illuminant white balance. + +3. IFFCC is spatially adaptive: spatial interpolation and filtering on local estimates ensure smooth transitions while preserving distinct light boundaries and sharp gradients. + +# 2. Related Work + +# 2.1. Global AWB Methods + +Conventional methods have long served as foundational approaches for uniform color constancy. These techniques estimate the illuminant's color based on the statistical characteristics of a single lighting source in the image [9, 17, 18, 21, 23, 47]. Notable single-illumination examples include the Gray World assumption [10], which posits that the average color of a scene under equal-energy white illumination should appear gray. Similarly, White Patch [33] assumes that the brightest pixel serves as a hint for illumination chromaticity. Refining these assumptions to focus on local patches or higher-order gradients has led to more robust statistics-based methods, optimized for more-challenging global ambient illuminations. These methods include General Gray World [1], Gray Edge [47], Shades-of-Gray [16], and LSRS [20], among others [11, 42]. + +In the context of machine learning-based methods, more machine learning-based methods have been proposed, such as learning-based kernels [2, 3], convolutional feature extraction [6, 27], and other machine learning approaches have been devised specifically for handling single-illumination scenarios [35, 36, 41, 43, 48]. The most closely related methods to ours are Barron's CCC [2] and FFCC [3]. CCC reformulates color constancy as a color localization task in log-chromaticity space, using convolutional kernels to accurately locate the correct color. FFCC builds on this by applying the Fast Fourier Transform on a chromaticity torus, greatly improving the speed and accuracy of chromaticity estimation. However, both are global methods and lack direct applicability to local scenes unless explicitly provided with a local window. + +# 2.2. Local AWB Methods + +Several multi-illuminant benchmarks have been introduced [4, 5, 8, 24, 31, 35, 39], with various methods proposed to tackle local white balance. These methods often rely on additional prior information or metadata, such as the number of light sources [4, 26] or facial regions of interest [5]. For instance, [28] uses flash photography to achieve local white balance. A novel framework in [4] treats the problem as an energy minimization task within a conditional random field, leveraging local light source estimates. [8] employs a global method with spatial variation, while [7] introduces a CNN framework based on sliding windows. Additionally, [46] applies a generative adversarial network (GAN) to correct images using synthetic datasets, adjusting the color based on the lighting information. More recently, Domislovic et + +al. [14] proposed extracting spatial features for illumination prediction on each patch, and AID [32] introduces a deep model utilizing slot attention [37] to generate chromaticity and weight maps for each light source, with the added benefit of allowing user edits. + +Among these approaches, most focus on pixel-level predictions, often leveraging encoder-decoder architectures (e.g., U-Net) to capture fine-grained illumination chromaticity details. However, this strategy presents several practical challenges. First, pixel-level prediction is computationally intensive, especially for high-resolution images, which significantly limits real-time applicability. Second, due to the lack of spatial context, pixel-wise color correction can overlook critical illumination chromaticity differences, leading to the loss of smooth illumination transitions and distortions in color edges, both of which are essential for accurate human visual perception. By treating each pixel independently, these methods risk introducing inconsistencies across the image and erasing subtle but important color shifts. As a result, their effectiveness in real-world applications, particularly in resource-constrained environments such as mobile devices, remains significantly limited. + +# 3. Method + +We propose the Integral Fast Fourier Color Constancy (IFFCC) model, which builds upon the foundations of Fast Fourier Color Constancy (FFCC) [3]. Drawing inspiration from [40], we implement an histogram search for all possible target regions in image data within Cartesian coordinate space. This enables efficient retrieval of any region's histogram in linear computation time, avoiding redundant summation operations. + +FFCC Foundation. Consider a linear image stat $I$ obtained from the camera, where there are no saturated pixels and the black level has been subtracted. Assuming the RGB value of any pixel is the product of its true albedo RGB value $W$ and the uniform illumination value shared across all pixels: $I = W \times L$ . Although this assumption disregards shadows, dichromatic reflections, and spatial variations, it remains effective and widely accepted. Our aim is to estimate the $L$ given $I$ , and then $W = I / L$ . + +For an input RGB image $I$ , CCC [2] defines two log-chroma measures for each pixel $i$ : + +$$ +u ^ {(i)} = \log \left(I _ {g} ^ {(i)} / I _ {r} ^ {(i)}\right) v ^ {(i)} = \log \left(I _ {g} ^ {(i)} / I _ {b} ^ {(i)}\right) \tag {1} +$$ + +Neglecting the absolute scaling of $W$ , the estimation of $L$ is then simplified as $L_{u} = \log (L_{g} / L_{r}) \quad L_{v} = \log (L_{g} / L_{b})$ . Given the unknown absolute scale, the inverse mapping from RGB to UV is challenging. To simplify the recovery process, we assume $L$ is the unit-norm: + +$$ +\begin{array}{l} L _ {r} = \exp (- L _ {u}) / z \quad L _ {g} = 1 / z \quad L _ {b} = \exp (- L _ {v}) / z \tag {2} \\ z = \sqrt {\exp (- L _ {u}) ^ {2} + \exp (- L _ {v}) ^ {2} + 1} \\ \end{array} +$$ + +This logarithmic chromatic space method has several advantages over the RGB approach. First, it involves only two unknowns instead of three, which enhances numerical stability. Additionally, this framework transforms the multiplicative constraints related to $W$ and $I$ into additive constraints [15], effectively converting the color constancy task into a positioning task in a two-dimensional space. + +Following FFCC, we use FFT-based convolution in the log-chroma space, leveraging its periodicity to accelerate computations. Given the histogram's limited size, which does not fully capture the range of colors in natural images, a modular algorithm is needed to allow pixels to "wrap around" in this space: + +$$ +\begin{array}{l} H (x, y) = \sum_ {i} \left(\mod \left((u ^ {(i)} - u _ {s}) / b - x, n\right) < 1 \right. \tag {3} \\ \wedge \mod \left(\left(v ^ {(i)} - v _ {s}\right) / b - y, n\right) < 1\bigg) \\ \end{array} +$$ + +where $H(x,y)$ represents the count of pixels in $I$ whose log-chroma coordinates are close to the $(u,v)$ coordinates associated with the histogram position $(x,y)$ , $n$ is the number of bins, and $b$ is the bin size. $u_{s}$ and $v_{s}$ denote the starting positions of $u$ and $v$ , respectively. The operations here differ from those in CCC. In this case, the $(x,y)$ positions in the histogram no longer represent absolute colors, but rather a set of $(u,v)$ color coordinates. The filtered results produce a set of light sources, from which we apply the "gray light de-aliasing" method to resolve the resulting light source clusters. This method assumes that the light sources are as close as possible to the center of the histogram. + +FFCC employs the Differentiable Bivariate von Mises (BVM) [38] method to address the single estimation of light sources from the toroidal PDF. Unlike FFCC, we need to predict multiple light sources for each image. Therefore, the BVM method is parallelized across multiple windows to predict several light sources simultaneously, as described in Section 3.3. The final convolutional structure is as follows: + +$$ +P = \operatorname {s o f t m a x} \left(B + M \circ \sum_ {s} H _ {s} * F _ {s}\right) \tag {4} +$$ + +where $B$ is the bias map, $M$ is the learnable gain map, and $\{F_s\}$ is the set of learnable filters convolved with the edge-enhanced histogram group $\{H_s\}$ . + +# 3.1. Integral UV Histogram + +FFCC predicts a global illumination for the image. However, when multiple illuminations are anticipated for different target regions, the process requires FFCC to repeatedly extract histograms from these various areas, significantly increasing computational load and processing time. Inspired by the work of [40], we efficiently compute the histograms for all potential regions within a given image + +![](images/9fa8d7fe5135bcbc8551dc46a1cfd948d3d409523a7929c5db461a3bcc871ded.jpg) +(a) Integral UV histogram + +![](images/c22d7a3a899c0ae6c6527a8836033d10d8d3519e3fa7f1d558c0f4b712741192.jpg) +(b) Overlapping grid + +![](images/d0adb3ec75ce2066cbdd29ef80d6c335fa17045106cb0f51e895a248ddc2f1da.jpg) +(c) Local histogram + +![](images/e9c2b23ee513c98e8c3e89e53bf83048f53676751e8e385a4d05a766e6230922.jpg) +(d) Parallel prediction +Figure 2. Overview of the IFFCC process for predicting multi-illuminant scenes. We precompute the integral UV histogram of the input image in the log-chroma space (a). Multiple target regions are extracted on the integrated UV histogram with the overlapping grid (b), and the local log-chroma histogram (c) corresponding to each region is computed using simple arithmetic operations on the UV histogram. Each histogram forms a torus that can be unfolded, producing a cluster of potential light sources. A de-aliasing step selects the most likely light source for each region (d), represented by a point within an ellipse, where the ellipse indicates IFFCC's covariance output. Final light estimates are smoothed through interpolation and guided filtering, creating a continuous illumination map (e) that combines with the input to produce a white-balanced image (f) for multi-illuminant scenes. + +![](images/a158bceeb65724cc355c40269d2aaab2c1c4edd3fad9f2126ce093515cd1e9cc.jpg) +(e) Spatial smoothing + +![](images/dc2acfde9366f71c35a851c6a2567d933b4de6f20ffaea42c249a67e0f61159c.jpg) +(f) Output image + +using Cartesian coordinates. By leveraging the spatial positions of data in this coordinate system, we propagate a function starting from the origin to navigate through the remaining positions. The histogram for the current point is then updated based on the histograms of previously processed points. Once the integral histogram for each data point is obtained, the histogram for the target region can be derived using simple addition and subtraction operations. This method eliminates the need to reconstruct separate histograms for each region individually. As a result, this efficient approach significantly reduces computational overhead and improves processing time, enabling rapid analysis of multiple areas within the image. + +After transforming the image into log-chroma space by Eq.1, we define the integral uv histogram as follows: + +$$ +H _ {\text {I n t e g r a l}} \left(x ^ {n}, b\right) = \bigcup_ {m = 0} ^ {n} B (u, v) ^ {m} \tag {5} +$$ + +$$ +B (u, v) = \left( \begin{array}{l} \mod \left(\operatorname {r o u n d} \left(\frac {u - u _ {s}}{b}\right), n\right), \\ \mod \left(\operatorname {r o u n d} \left(\frac {v - v _ {s}}{b}\right), n\right) \end{array} \right) \tag {6} +$$ + +where $x^n$ is a sequence of points $x^0, x^1, \ldots, x^n$ . $B(\cdot)$ represents the corresponding bin for the current data point, $\bigcup$ denotes the union operator, and the value of bin $b$ in $H_{\mathrm{Integral}}(x^n, b)$ is the sum of the histogram values of bin $b$ from all previously traversed data points. The integral histogram can also be expressed recursively as follows: + +$$ +H _ {\text {I n t e g r a l}} \left(x ^ {m}, b\right) = H _ {\text {I n t e g r a l}} \left(x ^ {m - 1}, b\right) \bigcup B (u, v) ^ {m} \tag {7} +$$ + +where at the initial data point, $H_{\mathrm{Integral}}(0,b) = 0$ . + +In log-chroma space, updating the histogram at the current point requires the histograms of the data points above, + +the left, and the top-left of the current point. These three histograms have already been calculated in previous scans, so only three simple arithmetic operations are needed to update the histogram for the current point. Afterward, the bin value $b$ for the current point is calculated, and the histogram is updated accordingly. Before each propagation step, all scanned bin values are copied to the current bin values, which can be efficiently handled through fast hardware or pointer operations. + +For an RGB image converted to $N_{1} \times N_{2} \times 2$ log-chroma space data, the integral histogram propagated via wavefront scanning from the top-left starting point can be written as: + +$$ +\begin{array}{l} H _ {\text {I n t e g r a l}} (u, v, b) = H _ {\text {I n t e g r a l}} (u - 1, v, b) \\ + H _ {\text {I n t e g r a l}} (u, v - 1, b) \tag {8} \\ - H _ {\text {I n t e g r a l}} (u - 1, v - 1, b) + B (u, v) \\ \end{array} +$$ + +where $(u,v)$ is the coordinate of the log-chroma space. In our method, the data in log-chroma space can be represented by a two-dimensional floating point tensor of a two-dimensional scene [40], allowing us to calculate the computational load ratio between the traditional histogram method and the integral histogram method as follows: + +$$ +R = \frac {\left[ 2 1 7 M _ {1} M _ {2} + 4 B ^ {2} \right] S _ {1} S _ {2}}{2 7 4 + 1 1 (6 8 + 4 B ^ {2}) S _ {1} S _ {2}} \tag {9} +$$ + +where $M_{1}$ and $M_{2}$ represent the size of the target region in the two-dimensional scene, $B$ is the number of bins, and $S_{1}$ and $S_{2}$ are the different scales used when extracting histograms. The result of the calculation for $R$ when $M_{1} = M_{2} = 128$ , $B = 64$ , and $S_{1} = S_{2} = 1$ is approximately $R = 19.71$ . + +# 3.2. Overlapping Grid + +During inference, we use a sliding window with overlap to efficiently retrieve histograms for sub-regions in the log-chroma space, as illustrated in Fig.2. By pre-computing an + +integral histogram over the entire scene, histograms for any arbitrary sub-region can be directly retrieved using Eq.8, without reconstructing each one from scratch, which significantly reduces computational costs. This overlapping grid method captures local chromatic variations by sliding a defined window across the scene and calculating histograms incrementally. The overlapping regions between windows maintain continuity and minimize edge effects, enhancing the accuracy of chromatic analysis throughout the scene. + +# 3.3. Parallel Prediction + +To speed up inference, our method requires only one inference to predict several illumination across multiple target regions within the image. Given an image's integral UV histogram in log-chroma space, denoted as $H_{\mathrm{integral}}(H,W,B,B,N)$ , where $H$ and $W$ represent the image size, $B$ is the number of bins, and $N$ is the number of augmented channel's histogram. The overlapping grid approach can be applied to obtain a collection $X$ of $k$ target region histograms. This collection $X$ , represented as $X(k,B,B,N)$ , includes histograms for each specific target region in the log-chroma space. The process of calculating the PDF of $X$ in the frequency domain is as follows: + +$$ +\begin{array}{l} H = \operatorname {I F F T} \left(\sum_ {k} \operatorname {F F T} (X) \cdot \operatorname {F F T} (F)\right) + B (10) \\ P = \frac {\exp^ {(H - \max (H))}}{\sum_ {\mathrm {d i m s}} \exp^ {(H - \max (H))}} (11) \\ \end{array} +$$ + +where $\mathrm{FFT}(.)$ and $\mathrm{IFFT}(.)$ are the fast Fourier transform and its inverse, respectively, $F$ is the convolution kernel, and $B$ is the bias map. + +To parallelize the estimation of the mean of the Bivariate Von Mises (BVM) distribution from a histogram, we calculate circular mean values for the histogram coordinates $i$ and $j$ across multiple regions concurrently. This process enables effective mean direction estimation within the toroidal data space, essential for capturing periodic patterns in wrapped grid data. For each region $k$ , we define $\mu = \{\mu_k\}_{k=1}^K$ , where each $\mu_k$ is given by: + +$$ +\mu_ {k} = \left[ \begin{array}{l} u _ {s} \\ v _ {s} \end{array} \right] + b \left[ \begin{array}{l} \bmod \left(\frac {n}{2 \pi} \arctan 2 \left(y _ {1} ^ {k}, x _ {1} ^ {k}\right), n\right) \\ \bmod \left(\frac {n}{2 \pi} \arctan 2 \left(y _ {2} ^ {k}, x _ {2} ^ {k}\right), n\right) \end{array} \right] \tag {12} +$$ + +where the components $y_1^k, x_1^k, y_2^k$ , and $x_2^k$ are computed in parallel for each $k$ : + +$$ +y _ {1} ^ {k} = \sum_ {i} P _ {i} ^ {k} \sin \left(\frac {2 \pi i}{n}\right), x _ {1} ^ {k} = \sum_ {i} P _ {i} ^ {k} \cos \left(\frac {2 \pi i}{n}\right) \tag {13} +$$ + +$$ +y _ {2} ^ {k} = \sum_ {j} P _ {j} ^ {k} \sin \left(\frac {2 \pi j}{n}\right), x _ {2} ^ {k} = \sum_ {j} P _ {j} ^ {k} \cos \left(\frac {2 \pi j}{n}\right) +$$ + +where $P_{i}^{k}$ and $P_{j}^{k}$ represent the marginal distributions for rows and columns of each target region's 2D histogram. + +After concurrently obtaining illumination values for all $k$ target regions, we reconstruct the full illumination map by interpolating or applying edge-preserving filtering methods. + +# 3.4. Spatial Smoothing + +In multi-light scenarios, patch-wise illumination estimates often exhibit discrete, locally discontinuous characteristics. To reconstruct a smooth and seamless illumination map across the entire image, it is vital to postprocess these patch-level estimates using interpolation. + +We employ linear interpolation due to its simplicity and ability to create smooth, piecewise-linear transitions for estimating intermediate values. Although higher-degree polynomial interpolation could be used, it risks introducing oscillations. After getting estimates for $k$ target regions, interpolation adjusts these values based on their relative locations, resulting in a smooth illumination map. Although interpolation enhances overall smoothness, it may blur local details, particularly in regions with strong light edges or steep illumination gradients, potentially leading to a loss of edge information. + +Guided filtering [25] is then applied to preserve details at the boundaries and structural features of the illumination map, preventing blur or distortion during transitions. Overall, combining these techniques results in a globally coherent illumination distribution that maintains edge clarity while appearing smoother and more natural. Given the interpolated illumination map $L$ and the raw image $G$ as the guiding image, the output illumination map $O$ at each pixel $O_{i}$ can be expressed as: + +$$ +O _ {i} = a _ {k} G _ {i} + b _ {k}, \quad \forall i \in \omega_ {k} \tag {14} +$$ + +$$ +a _ {k} = \frac {\frac {1}{| \omega |} \sum_ {i \in \omega_ {k}} G _ {i} L _ {i} - \bar {G} _ {k} \bar {L} _ {k}}{\sigma_ {k} ^ {2} + \epsilon}, \quad b _ {k} = \bar {L} _ {k} - a _ {k} \bar {G} _ {k} \tag {15} +$$ + +here, $\bar{G}_k$ and $\bar{L}_k$ are the mean values of $G$ and $L$ in window $\omega_k$ , with $\sigma_k^2$ as the variance of $G$ . The constant $\epsilon$ prevents division by zero. The final illumination map $O$ is obtained by weighted averaging. + +$$ +O = \frac {1}{| \omega |} \sum_ {k \in \omega_ {i}} \left(a _ {k} G _ {i} + b _ {k}\right) \tag {16} +$$ + +This post-processing step is essential: it creates a coherent illumination map that accurately reflects the lighting conditions across the entire image while preserving critical features, thus the reconstructed illumination aligns well with the underlying content of the image. + +# 4. Experiment + +# 4.1. Experimental Setup + +We evaluated our proposed IFFCC on two widely used multi-light source datasets: the LSMI dataset [31] and the + +
MethodCanon_5dCanon_550dMotoPanasonicSonyUse all
meanmedianmeanmedianmeanmedianmeanmedianmeanmedianmeanmedian
Huassin and Akbari [29]13.8813.7613.7313.5413.0612.8913.6213.7314.2213.6813.6713.46
CRF(White-Patch)[4]7.665.967.695.776.945.327.585.697.905.867.195.44
Patch-based[24]4.853.114.612.984.172.494.582.954.963.164.302.89
Keypoint-based[24]5.593.885.483.944.973.285.423.795.763.825.463.59
Superpixel-based[24]4.883.684.793.564.192.894.553.544.973.824.203.10
IFFCC2.061.542.161.771.901.362.061.592.661.962.191.56
Bianco et al.[7]4.634.024.594.164.273.894.984.565.344.768.015.69
HypNet/SelNet[44]5.424.935.164.835.094.745.234.985.825.346.313.95
LSMI-U[31]3.142.332.922.462.792.043.172.703.582.883.722.40
Domislovic et al.[12]2.632.182.772.062.071.492.381.732.922.242.281.60
+ +Table 1. The performance results on the Shadow dataset are presented. The dataset is segmented by images captured from different cameras, with "use all" representing the entire dataset. We report the mean and median angular errors for each subset. IFFCC employs the default settings, with a window size of 128 and an overlap size of 64. The model is trained using the blended illumination within each window as the ground truth (GT), and the corresponding error is computed. + +
MethodOutdoorIndoorNighttime
meanmedianb25w25meanmedianb25w25meanmedianb25w25
Huassin and Akbari[29]13.1912.897.1019.1214.8413.228.4320.1816.3514.969.7522.33
CRF(White-Patch)[4]7.315.662.3113.368.786.894.3614.406.965.162.3815.85
Patch-based[24]4.533.291.3810.475.584.432.0911.393.482.171.8912.53
Keypoint-based[24]5.773.881.2411.786.254.041.7912.603.994.101.4913.08
Superpixel-based[24]4.733.011.338.215.183.491.929.244.632.981.349.82
IFFCC1.741.280.423.833.042.140.666.712.851.530.547.43
Bianco et al.[7]4.862.361.9811.757.425.623.3812.508.225.783.1013.68
HypNet/SelNet[44]6.094.081.7712.317.075.272.9811.494.763.212.2012.59
LSMI-U[31]2.122.031.025.485.724.621.498.544.812.471.338.89
Domislovic et al.[12]1.781.460.654.344.453.310.987.463.681.890.838.22
+ +Table 2. The performance results on the Shadow dataset are presented. The dataset is segmented by images captured in different scenes, with each scene including images taken by five different cameras. We report the mean and median angular errors for each subset. IFFCC uses the default settings, with a window size of 128 and an overlap size of 64. The model is trained using the blended illumination within each window as the ground truth (GT), and the corresponding error is computed. + +Shadow dataset [13]. The LSMI dataset includes over 7,486 images captured by three different cameras in a variety of scenes, each containing multiple light sources. The Shadow dataset consists of 2,500 images taken in both indoor and outdoor settings, with each image accompanied by a binary segmentation mask, where each region is illuminated by a single light source. In all experiments, full-sized training images were randomly cropped to $128 \times 128$ , and testing was conducted on $256 \times 256$ images. All images were black-level corrected, and calibration objects were masked. + +IFFCC's computational efficiency enables CPU-only training and inference, eliminating the need for a GPU. All experiments were conducted on an Intel Xeon Gold 6258R CPU. Performance was evaluated using standard error metrics, including mean and median angular errors and the arithmetic means of the first and third quartiles ("best $25\%$ " and "worst $25\%$ ). Additionally, we compared model parameter counts and CPU inference times for $256\times 256$ images. Since our method operates on patches, we calculate blended illumination values from the ground truth illumination map within a sliding window, using these as the ground + +truth for each patch. The impact of different training approaches is discussed in the ablation study. + +During training, we input a set of enhanced histograms derived from filtered images that ensure non-negativity and intensity scaling consistency. Each image includes a local absolute deviation measure. Using the preconditioned frequency-domain optimization from FFCC [3], training runs for 64 iterations with a histogram size of $64 \times 64$ . + +# 4.2. Results and Comparisons + +Quantitative Comparison We evaluated IFFCC against three traditional methods [4, 24, 29] and three learning-based methods [7, 12, 44] on the Shadow dataset. The experimental results are presented in Tables 1 and 2. Table 1 organizes the Shadow dataset into subsets based on images captured by different cameras, while Table 2 categorizes the dataset by various scene types, each containing photos taken with five distinct cameras. Despite comparisons with advanced network-based methods, IFFCC achieves state-of-the-art performance. + +The results in Table 2 highlight the strong general + +![](images/08a937f70aae03599c928fd24a5b0ba0e383f9427fbb59c3c4026447a14e8ee5.jpg) + +![](images/1bec36dacc91776241920f0dbea216be1b771b850927083ecc5d11d2b00c5edb.jpg) + +![](images/56888799e5407b56c099a688485b917d9b3e861459154d8da45c002770b68bf6.jpg) +Input + +![](images/0d6c1f4d1c60575cb7891b98fd9494aaa5d858a2d4b95da0ac267ae4af85e675.jpg) + +![](images/af6be4370513d5bd7350b9d0bcd3315bcd3ee941292a55f3242fb07f087f3c81.jpg) + +![](images/bde56dc819d1defdbb5541be2a7f288b84702c48bbcde735bb6b4b59fd5f7b39.jpg) +Gijsenij et al. + +![](images/a352477d17e3c42cdedc1870e49f0d90d6890147c66e8827500ec030b1e26f9b.jpg) + +![](images/55a775a8a6912962afa985239c7bf9bd4221cc3ae75832540de61b1f5c67221a.jpg) + +![](images/7be33c47aac92232fd67ea8b0bc19b2208ad45f5877d437c76e2fb134c111f3c.jpg) +Bianoco et al. + +![](images/0dd66e408b39582773835b0802eaccfb3dee27a61af14c164757c3f2567c832c.jpg) + +![](images/692b60b28a4e8c3437472f690710de1a647822075bbf7efd1ea832814eef9025.jpg) + +![](images/6d1b057fdc16654def0802470ac38a2a866c9c9c4401edbe5cc9883e6564e733.jpg) +LSMI-U + +![](images/c5f3ef5d045df532156854e852da0238fefc914336b67d0f33af1da1da554168.jpg) + +![](images/578fa221ec4854d1bee0d613c308525238a53ee3919aae231ab6dd78f2fc4955.jpg) + +![](images/c7360708af7d35414bd03dcf2b61b266b35729567a32baf242b1c8ecec6aade1.jpg) +IFFCC + +![](images/b011d90704d890e79777db34f213bb54497172c95b761511802578ff7a2d10b5.jpg) + +![](images/d577927f0c885928d6cb4d026c736fdbd2a8c232ba6488c7ce0b9ddc1aaa7a32.jpg) + +![](images/a538d4fffeae53398142dc16c7a2d39c8ca8917514bb2c6236a4308133f96165.jpg) +GT + +![](images/0ad81bbb352626a96769bff158d51f6a6ff903c8928395347a2ffc805588863d.jpg) +Figure 3. Visualization of the comparison results on the LSMI dataset. For IFFCC, the window size is set to 128 with an overlap of 64. In other patch-based methods, the window size is set to the default value of 32. + +![](images/ba26574610174747eb5ae94244c0ed0028e0ffd54ed1c48f554b9e3f8a4270b9.jpg) +Input + +![](images/59f71e4f4ddd0ad0117802a538f575941be9f2e3911c7f9027a9777d2d017bc8.jpg) + +![](images/3bd6ee5f171edd36fd445a4186ea3a8821cf802b79ea5cf9d44af44020dcf580.jpg) +White Patch + +![](images/ac70c95c817f221dfc4bd0d450aca068649ca126eebb87fcef44bcabf9f288ee.jpg) + +![](images/33380cca317a2900d5dfb00acc3485cbe98211611d31c1440736d4f80a6e7d42.jpg) +IFFCC + +![](images/c533d90204ff306db7966fe16e36c4ae3c872714971aae4d8272aca8202eb307.jpg) + +![](images/be30854263048d763a57a87751f48c3eaaf6127646d44cc082414eda1d18e041.jpg) +Illumination map +Figure 4. Visualization of illumination comparison for different patch-based methods. The window size is 32, with an overlap size of 16. + +![](images/72b93274c4df36efdfd8fd403d295b30e491785bb00de686dd3a2a11099b491c.jpg) + +![](images/04bc55f93adc5ba4720e9b02f5a2a0ea3e4c3e23ec3bb8a4e214a9b4039d7dd1.jpg) +White-balanced image + +ization ability of IFFCC. Network-based methods generally perform worse than IFFCC on multi-camera datasets, where each camera's unique spectral sensitivities complicate learning a single mapping across different devices. IFFCC generalizes well across multiple cameras because it leverages an integral log-chroma histogram approach, which directly models chromaticity distributions without being overly reliant on the specific spectral sensitivities of individual cameras. This histogram-based method effectively captures illumination variations across different scenes and cameras, allowing IFFCC to produce robust predictions without requiring extensive fine-tuning for each camera's unique spectral characteristics. + +The experimental results on the LSMI dataset, presented in Tables 3, and 4, provide a comparative analysis of patch-based and pixel-level methods. The test set is categorized into Multi (multiple light sources), and Mixed (compris + +ing both single and multiple light sources). Among patch-based approaches, IFFCC demonstrates state-of-the-art performance, achieving results comparable to those of leading pixel-level methods, including AID [32]. Notably, IFFCC's compact parameter count and rapid processing speed render it highly suitable for deployment on mobile and resource-constrained devices. + +We also compared the processing time for a $256 \times 256$ image in a multi-illuminant scene using FFCC and IFFCC, as shown in Table 5. As the number of windows increases, IF-FCC significantly outperforms FFCC in processing speed. + +Qualitative Comparison Fig.3 visually compares our method against alternative approaches on the LSMI dataset. To enhance visualization quality, color correction, gamma correction, and tone mapping were applied to the white-balanced images. The results indicate that IFFCC achieves more accurate white balance restoration than competing + +
MethodMultiParams (M)Time (s)
Galaxy mean med.Nikon mean med.Sony mean med.
Patch -levelGijsenij[23]12.389.579.316.5911.198.14-1.3
Bianoco[7]5.564.344.654.194.383.940.16-
IFFCC2.481.902.301.502.481.970.0120.03
Pix2Pix[30]4.282.634.182.764.373.265.40.48
Pixel LSMI-H[31]3.132.703.203.013.653.330.520.28
-levelLSMI-U[31]2.852.552.361.842.762.385.70.51
AID[32]2.031.432.261.392.161.646.4>1
+ +Table 3. The performance results on the multi-illumination images from the LSMI dataset are presented, with testing time corresponding to the processing time for $256\times 256$ images on a CPU. + +
MethodMixedParams (M)Time (s)
Galaxy mean med.Nikon mean med.Sony mean med.
Patch -levelGijsenij[23]10.097.439.215.839.467.13-1.3
Bianoco[7]4.893.833.933.483.843.210.16-
IFFCC1.981.792.111.481.931.720.0120.03
Pix2Pix[30]5.662.445.412.494.202.205.40.48
Pixel LSMI-H[31]3.062.542.922.633.212.850.520.28
-level LSMI-U[31]2.631.912.161.552.682.295.70.51
AID[32]1.631.321.761.141.681.446.4>1
+ +Table 4. The performance results on the mixed-illumination images (containing both single and multiple light sources) from the LSMI dataset are presented, with testing time corresponding to the processing time for $256 \times 256$ images on a CPU. + +
[32,0][32,16][64,32][64,48][128,64][128,96]
FFCC162ms450ms203ms689ms88ms234ms
IFFCC34ms46ms29ms40ms27ms33ms
+ +methods. We also present a comparison of illumination maps between our method and patch-based approaches, as shown in Fig.4. The window size is set to $32 \times 32$ with an overlap of 16. The results demonstrate that IFFCC achieves a more precise and smoother illumination prediction. + +# 4.3. Ablation Study + +Table 6 presents results from ablation experiments analyzing the impact of various training strategies, window sizes, and overlap regions. Here, "B" represents training with a blend of illumination maps as the ground truth (GT), and "M" uses the dominant light source as GT. Parameters "ws" and "os" indicate window and overlap sizes, respectively. Findings reveal that smaller window sizes, with limited information, reduce prediction accuracy. However, the blended illumination strategy provides improved accuracy compared to using a single dominant light source. Visual + +Table 5. Comparison of average testing time for FFCC and IF-FCC on the Shadow dataset, evaluated across different window and overlap sizes. FFCC employs a conventional histogram computation approach. + +
type[ws, os]meanmedianb25w25
B[32, 16]2.652.171.374.65
B[64, 0]2.352.011.134.31
M[64, 32]2.532.041.254.83
B[64, 32]2.371.871.164.38
M[128, 0]2.462.031.044.78
B[128, 0]2.161.680.914.23
M[128, 32]2.501.881.044.91
B[128, 32]2.141.640.854.30
M[128, 64]2.421.950.934.71
B[128, 64]2.061.540.814.19
B[128, 96]2.131.590.834.25
B[192, 128]1.981.340.724.08
+ +Table 6. Ablation study on the Shadow dataset (Canon5d), reporting the impact of different window sizes, overlap sizes, and training strategies on accuracy. + +![](images/1af028746080cd044c89b20f5f01098b51ecb64f29ab673ec13a63830e9752ec.jpg) +Figure 5. (a)-(d) show illumination maps generated with different window sizes ([32, 16], [64, 32], [128, 64], [192, 128]), each processed with smooth transition adjustments. (e) presents the ground truth (GT) illumination map, and (f) displays the white-balanced image obtained by combining the input image with the GT. + +![](images/a75bba796124fc67a525818245f86163fea8d2a4ec33384501ef87c136c35c17.jpg) + +![](images/7b36b0d9b85f3ccc80032ba5388a5177d38152dcefb2d893600aac85cdda397a.jpg) + +comparisons (Fig. 5) show that smaller windows enhance texture details in the predicted illumination maps but with lower accuracy. Conversely, larger windows (e.g., [192, 128]) reduce error metrics by capturing more information, yet they may blur boundaries in multi-illuminant regions, compromising edge clarity. + +# 5. Conclusion + +We propose IFFCC, a practical solution for achieving accurate, real-time auto white balance (AWB) in multi-illuminant environments. Built upon the Fast Fourier Color Constancy (FFCC) algorithm, IFFCC leverages an integral UV histogram and parallelized Fourier-based convolution to rapidly estimate illumination across different image regions, addressing the need for localized, adaptive AWB. Results on multiple datasets demonstrate that IFFCC delivers accuracy comparable to or surpassing pixel-level neural network methods, while reducing the parameter count by over 400 times and speeding up processing by a factor of 20–100. These advantages make IFFCC highly suitable for resource-limited, real-time applications such as video streaming. + +We believe that FFCC-based methods, including IFFCC, will continue to enrich the field and attract increased attention across a broader range of ISP applications. + +Acknowledgments This work was supported in part by the Anhui Provincial Natural Science Foundation under Grant 2408085J024, the National Natural Science Foundation of China under Grant 62401532, and the Anhui Provincial Key Research and Development Plan under Grant 202304a05020072. + +# References + +[1] Kobus Barnard, Vlad Cardei, and Brian Funt. A comparison of computational color constancy algorithms. i: Methodology and experiments with synthesized data. IEEE transactions on Image Processing, 11(9):972-984, 2002. 2 +[2] Jonathan T Barron. Convolutional color constancy. In Proceedings of the IEEE International Conference on Computer Vision, pages 379-387, 2015. 2, 3 +[3] Jonathan T Barron and Yun-Ta Tsai. Fast fourier color constancy. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 886-894, 2017. 1, 2, 3, 6 +[4] Shida Beigpour, Christian Riess, Joost Van De Weijer, and Eli Angelopoulou. Multi-illuminant estimation with conditional random fields. IEEE Transactions on Image Processing, 23(1):83–96, 2013. 2, 6 +[5] Simone Bianco and Raimondo Schettini. Adaptive color constancy using faces. IEEE transactions on pattern analysis and machine intelligence, 36(8):1505-1518, 2014. 2 +[6] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Color constancy using cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 81-89, 2015. 2 +[7] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Single and multiple illuminant estimation using convolutional neural networks. IEEE Transactions on Image Processing, 26(9):4347-4362, 2017. 2, 6, 8 +[8] Michael Bleier, Christian Riess, Shida Beigpour, Eva Eibenberger, Eli Angelopoulou, Tobias Tröger, and André Kaup. Color constancy and non-uniform illumination: Can existing algorithms work? In 2011 IEEE international conference on computer vision workshops (ICCV Workshops), pages 774-781. IEEE, 2011. 2 +[9] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 2 +[10] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 1, 2 +[11] Dongliang Cheng, Dilip K Prasad, and Michael S Brown. Illuminant estimation for color constancy: why spatial-domain methods work and the role of the color distribution. JOSA A, 31(5):1049-1058, 2014. 1, 2 +[12] Ilija Domislovic, Donik Vrsnak, Marko Subasic, and Sven Loncarić. Outdoor daytime multi-illuminant color constancy. In 2021 12th International Symposium on Image and Signal Processing and Analysis (ISPA), pages 270–275. IEEE, 2021. 6 + +[13] Iija Domislovic, Donik Vršnak, Marko Subašić, and Sven Lončarić. Shadows & lumination: Two-illuminant multiple cameras color constancy dataset. Expert systems with applications, 224:120045, 2023. 6 +[14] Ilija Domislovic, Donik Vrsnjak, Marko Subasic, and Sven Loncarić. Color constancy for non-uniform illumination estimation with variable number of illuminants. Neural Computing and Applications, 35(20):14825-14835, 2023. 3 +[15] Graham D Finlayson and Steven D Hordley. Color constancy at a pixel. JOSA A, 18(2):253-264, 2001. 3 +[16] Graham D Finlayson and Elisabetta Trezzi. Shades of gray and colour constancy. In Color and Imaging Conference, pages 37-41. Society of Imaging Science and Technology, 2004. 2 +[17] Graham D. Finlayson, Steven D. Hordley, and Paul M. Hubel. Color by correlation: A simple, unifying framework for color constancy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(11):1209-1221, 2001. 2 +[18] David A Forsyth. A novel algorithm for color constancy. International Journal of Computer Vision, 5(1):5-35, 1990. 1, 2 +[19] David H Foster. Color constancy. Vision research, 51(7): 674-700, 2011. 1 +[20] Shaobing Gao, Wangwang Han, Kaifu Yang, Chaoyi Li, and Yongjie Li. Efficient color constancy with local surface reflectance statistics. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 158-173. Springer, 2014. 2 +[21] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Generalized gamut mapping using image derivative structures for color constancy. International Journal of Computer Vision, 86:127-139, 2010. 1, 2 +[22] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Computational color constancy: Survey and experiments. IEEE transactions on image processing, 20(9):2475-2489, 2011. 1 +[23] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):918-929, 2011. 1, 2, 8 +[24] Arjan Gijsenij, Rui Lu, and Theo Gevers. Color constancy for multiple light sources. IEEE Transactions on image processing, 21(2):697-707, 2011. 2, 6 +[25] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE transactions on pattern analysis and machine intelligence, 35(6):1397-1409, 2012. 5 +[26] Eugene Hsu, Tom Mertens, Sylvain Paris, Shai Avidan, and Frédo Durand. Light mixture estimation for spatially varying white balance. In ACM SIGGRAPH 2008 papers, pages 1-7. 2008. 2 +[27] Yuanming Hu, Baoyuan Wang, and Stephen Lin. Fc4: Fully convolutional color constancy with confidence-weighted pooling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4085-4094, 2017. 1, 2 +[28] Zhuo Hui, Aswin C Sankaranarayanan, Kalyan Sunkavalli, and Sunil Hadap. White balance under mixed illumination + +using flash photography. In 2016 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2016. 2 +[29] Md Akmol Hussain and Akbar Sheikh Akbari. Color constancy algorithm for mixed-illuminant scene images. IEEE Access, 6:8964–8976, 2018. 6 +[30] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 8 +[31] Dongyoung Kim, Jinwoo Kim, Seonghyeon Nam, Dongwoo Lee, Yeonkyung Lee, Nahiup Kang, Hyong-Euk Lee, ByungIn Yoo, Jae-Joon Han, and Seon Joo Kim. Large scale multi-illuminant (lsmi) dataset for developing white balance algorithm under mixed illumination. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2410–2419, 2021. 1, 2, 5, 6, 8 +[32] Dongyoung Kim, Jinwoo Kim, Junsang Yu, and Seon Joo Kim. Attentive illumination decomposition model for multi-illuminant white balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25512-25521, 2024. 3, 7, 8 +[33] Edwin H Land. The retinax theory of color vision. In Proceedings of the Royal Institution of Great Britain, page 23, 1974. 1, 2 +[34] Shuwei Li, Jikai Wang, Michael S Brown, and Robby T Tan. Transc: Transformer-based multiple illuminant color constancy using multitask learning. arXiv preprint arXiv:2211.08772, 1(2):3, 2022. 1 +[35] Yongjie Liu and Sijie Shen. Self-adaptive single and multi-illuminant estimation framework based on deep learning. arXiv preprint arXiv:1902.04705, 2019. 2 +[36] Yi-Chen Lo, Chia-Che Chang, Hsuan-Chao Chiu, Yu-Hao Huang, Chia-Ping Chen, Yu-Lin Chang, and Kevin Jou. Clcc: Contrastive learning for color constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8053-8063, 2021. 2 +[37] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33:11525-11538, 2020. 3 +[38] Kantilal Varichand Mardia. Statistics of directional data. Journal of the Royal Statistical Society Series B: Statistical Methodology, 37(3):349-371, 1975. 3 +[39] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A dataset of multi-illumination images in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4080–4089, 2019. 2 +[40] Fatih Porikli. Integral histogram: A fast way to extract histograms in cartesian spaces. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 829-836. IEEE, 2005. 3, 4 +[41] Yanlin Qian, Ke Chen, Jarno Nikkanen, Joni-Kristian Kamarainen, and Jiri Matas. Recurrent color constancy. In Proceedings of the IEEE international conference on computer vision, pages 5458-5466, 2017. 2 + +[42] Yanlin Qian, Joni-Kristian Kamarainen, Jarno Nikkanen, and Jiri Matas. On finding gray pixels. In CVPR, 2019. 2 +[43] Yanlin Qian, Jani Käpylä, Joni-Kristian Kämäräinen, Samu Koskinen, and Jiri Matas. A benchmark for burst color constancy. In ECCVW, 2020. 2 +[44] Wu Shi, Chen Change Loy, and Xiaou Tang. Deep specialized network for illuminant estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 371-387. Springer, 2016. 6 +[45] Oleksii Sidorov. Conditional gans for multi-illuminant color constancy: Revolution or yet another approach? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 1 +[46] Oleksii Sidorov. Conditional gans for multi-illuminant color constancy: Revolution or yet another approach? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2 +[47] Joost Van De Weijer, Theo Gevers, and Arjan Gijsenij. Edge-based color constancy. IEEE Transactions on image processing, 16(9):2207-2214, 2007. 2 +[48] Bolei Xu, Jingxin Liu, Xianxu Hou, Bozhi Liu, and Guoping Qiu. End-to-end illuminant estimation based on deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3616-3625, 2020. 2 \ No newline at end of file diff --git a/2025/Integral Fast Fourier Color Constancy/images.zip b/2025/Integral Fast Fourier Color Constancy/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..343874f0cd388de7c68c1acbad1dbaaf4b45b5e3 --- /dev/null +++ b/2025/Integral Fast Fourier Color Constancy/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4b200ecca561f7401911c38afec465edba3f33e59aa3b18126ae514972198e6 +size 708412 diff --git a/2025/Integral Fast Fourier Color Constancy/layout.json b/2025/Integral Fast Fourier Color Constancy/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f80afd370148330c082a5f5daed8401b946374a2 --- /dev/null +++ b/2025/Integral Fast Fourier Color Constancy/layout.json @@ -0,0 +1,10906 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 185, + 103, + 425, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 103, + 425, + 121 + ], + "spans": [ + { + "bbox": [ + 185, + 103, + 425, + 121 + ], + "type": "text", + "content": "Integral Fast Fourier Color Constancy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "spans": [ + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "text", + "content": "Wenjun Wei" + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "text", + "content": " * Yanlin Qian" + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "text", + "content": " * Huaian Chen" + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "text", + "content": " Junkang Dai" + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "text", + "content": " Yi Jin" + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "text", + "content": "University of Science and Technology of China " + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 141, + 493, + 186 + ], + "type": "text", + "content": "DJI Technology Co., Ltd \n{natstand, anchen, junkangdai}@mail.ustc.edu.cn honza.qian@dji.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 238, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 525 + ], + "type": "text", + "content": "Traditional auto white balance (AWB) algorithms typically assume a single global illuminant source, which leads to color distortions in multi-illuminant scenes. While recent neural network-based methods have shown excellent accuracy in such scenarios, their high parameter count and computational demands limit their practicality for real-time video applications. The Fast Fourier Color Constancy (FFCC) algorithm was proposed for single-illuminant-source scenes, predicting a global illuminant source with high efficiency. However, it cannot be directly applied to multi-illuminant scenarios unless specifically modified. To address this, we propose Integral Fast Fourier Color Constancy (IFFCC), an extension of FFCC tailored for multi-illuminant scenes. IFFCC leverages the proposed integral UV histogram to accelerate histogram computations across all possible regions in Cartesian space and parallelizes Fourier-based convolution operations, resulting in a spatially-smooth illumination map. This approach enables high-accuracy, real-time AWB in multi-illuminant scenes. Extensive experiments show that IFFCC achieves accuracy that is on par with or surpasses that of pixel-level neural networks, while reducing the parameter count by over " + }, + { + "bbox": [ + 55, + 238, + 296, + 525 + ], + "type": "inline_equation", + "content": "400\\times" + }, + { + "bbox": [ + 55, + 238, + 296, + 525 + ], + "type": "text", + "content": " and processing speed by " + }, + { + "bbox": [ + 55, + 238, + 296, + 525 + ], + "type": "inline_equation", + "content": "20 - 100\\times" + }, + { + "bbox": [ + 55, + 238, + 296, + 525 + ], + "type": "text", + "content": " faster than network-based approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 547, + 135, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 547, + 135, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 547, + 135, + 559 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 687 + ], + "type": "text", + "content": "Color constancy is a fundamental property of the human visual system, allowing for consistent color perception despite changes in illumination chromaticity. While this ability may seem \"natural\" to the human eye and visual cortex, replicating it in computational imaging systems is a complex challenge [19, 22]. Auto white balance (AWB) in image signal processing (ISP) is thus essential for approximating human-like color constancy by estimating ambient illumination chromaticity, enabling accurate color reproduction across varied lighting conditions." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 211, + 555, + 353 + ], + "blocks": [ + { + "bbox": [ + 316, + 211, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 316, + 211, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 211, + 555, + 353 + ], + "type": "image", + "image_path": "6b156077960eef34f511ecb6e49ee0bd3c4e88c4c761d57de9dbc6749278691a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 358, + 555, + 447 + ], + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 447 + ], + "type": "text", + "content": "Figure 1. Frameworks comparison between IFFCC and FFCC (top). FFCC uses a global histogram to predict a single illuminant, limiting its application in multi-illuminant scenes. In contrast, IFFCC utilizes an integral UV histogram for efficient region-specific histogram computation, generating spatially continuous illumination map through parallelized Fast Fourier convolution and smoothing to achieve high-accuracy, real-time auto white balance in multi-illuminant scenarios." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 472, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 616 + ], + "type": "text", + "content": "Traditional white balance methods often assume the presence of a single dominant illumination and rely on statistical algorithms [18, 21, 23, 45], such as the gray world algorithm [10] and the white patch algorithm [33] etc., to estimate the scene illumination. These algorithms may also be machine learning-based, trained on AWB benchmarks [3, 27]. However, they frequently bear local color distortion in complex multi-illuminant environments, as a single white balance adjustment is inadequate to handle varying lighting conditions. For instance, a room illuminated by both warm high pressure sodium lighting and cool window light may still fail to accurately reproduce the scene's true colors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Recently, deep neural network-based algorithms have been widely applied to pixel-level multi-illuminant prediction tasks [11, 31, 34, 45], utilizing deep learning to achieve accurate illumination map estimation in complex scenes. These methods are especially effective in capturing edge-aware, spatial illumination variations. However, the high parameter counts and computational demands of these models create substantial challenges for deployment in real-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26420" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "time, resource-limited applications, such as mobile photography devices. Beyond achieving high pixel-level prediction accuracy, a practical local AWB solution from an industrial perspective should prioritize several key requirements:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 294, + 168 + ], + "type": "text", + "content": "Effectiveness - The algorithm can accomplish the AWB task up to a high accuracy, in any given spatial region, ensuring robust performance in a wide range of scenarios and including different light intensities and color temperatures." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 169, + 294, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 294, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 294, + 265 + ], + "type": "text", + "content": "Efficiency - White balance adjustments must be fast enough to provide immediate feedback during view-finding and capture, particularly in rapidly changing lighting. Limited camera resources require AWB, along with other ISP functions like AE, AF, and color correction, to operate in real time. On resource-constrained devices like smartphones, low computational complexity and memory usage are essential, with AWB processing times ideally under 10 ms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 266, + 294, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 294, + 337 + ], + "type": "text", + "content": "Thumbnail Input or Stat - Many color constancy algorithms require full-resolution, high-bit images, which is resource-intensive on pervasive mobile devices. For the purpose of real-time preview, the module for generating AWB stat is often realized in hardware (like DSP), outputting a thumbnail-sized input (e.g., " + }, + { + "bbox": [ + 55, + 266, + 294, + 337 + ], + "type": "inline_equation", + "content": "64 \\times 48" + }, + { + "bbox": [ + 55, + 266, + 294, + 337 + ], + "type": "text", + "content": ", 8-bit)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 338, + 294, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 294, + 385 + ], + "type": "text", + "content": "Smoothness - The algorithm should maintain stable performance under varying lighting conditions, avoiding artifacts or abrupt transitions, and ensure smooth, accurate white balance across entire images or video frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 386, + 294, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 294, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 294, + 505 + ], + "type": "text", + "content": "The FFCC algorithm [3] effectively meets the requirements for global illumination scenarios but faces challenges in multi-illuminant scenes. Specifically, FFCC requires repeated histogram extraction for each target region, significantly increasing processing time and limiting its suitability for real-time applications, especially on resource-constrained devices like smartphones. Additionally, FFCC lacks mechanisms to ensure smooth spatial transitions under rapidly changing lighting, which can result in abrupt shifts in white balance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 506, + 294, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 294, + 555 + ], + "type": "text", + "content": "To overcome FFCC's limitations in multi-illuminant scenarios, we propose Integral Fast Fourier Color Constancy (IFFCC), a local extension of FFCC [3] designed for white balance in complex lighting conditions, as shown in Fig." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 555, + 294, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 294, + 602 + ], + "type": "text", + "content": "1. IFFCC achieves speeds " + }, + { + "bbox": [ + 55, + 555, + 294, + 602 + ], + "type": "inline_equation", + "content": "20 - 100 \\times" + }, + { + "bbox": [ + 55, + 555, + 294, + 602 + ], + "type": "text", + "content": " faster than other local methods, with a runtime of " + }, + { + "bbox": [ + 55, + 555, + 294, + 602 + ], + "type": "inline_equation", + "content": "5.8\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 555, + 294, + 602 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 555, + 294, + 602 + ], + "type": "inline_equation", + "content": "64 \\times 48" + }, + { + "bbox": [ + 55, + 555, + 294, + 602 + ], + "type": "text", + "content": " preview images on a CPU. Building upon FFCC, IFFCC introduces several key contributions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 605, + 294, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 605, + 294, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 605, + 294, + 653 + ], + "type": "text", + "content": "1. IFFCC outperforms prior-arts traditional algorithms and most network-based methods in multi-light source scenarios, achieving performance comparable to the best network-based methods;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "content": "2. IFFCC leverages the proposed integral UV histogram to accelerate histogram extraction across all possible regions in Cartesian space. It efficiently computes the histogram for any region using only a few simple arithmetic operations, ensuring both high accuracy and real-time" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 72, + 522, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 522, + 83 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 522, + 83 + ], + "type": "text", + "content": "performance for multi-illuminant white balance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 84, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 84, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 315, + 84, + 553, + 120 + ], + "type": "text", + "content": "3. IFFCC is spatially adaptive: spatial interpolation and filtering on local estimates ensure smooth transitions while preserving distinct light boundaries and sharp gradients." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 133, + 400, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 400, + 145 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 400, + 145 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 153, + 440, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 153, + 440, + 164 + ], + "spans": [ + { + "bbox": [ + 314, + 153, + 440, + 164 + ], + "type": "text", + "content": "2.1. Global AWB Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 171, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 553, + 350 + ], + "type": "text", + "content": "Conventional methods have long served as foundational approaches for uniform color constancy. These techniques estimate the illuminant's color based on the statistical characteristics of a single lighting source in the image [9, 17, 18, 21, 23, 47]. Notable single-illumination examples include the Gray World assumption [10], which posits that the average color of a scene under equal-energy white illumination should appear gray. Similarly, White Patch [33] assumes that the brightest pixel serves as a hint for illumination chromaticity. Refining these assumptions to focus on local patches or higher-order gradients has led to more robust statistics-based methods, optimized for more-challenging global ambient illuminations. These methods include General Gray World [1], Gray Edge [47], Shades-of-Gray [16], and LSRS [20], among others [11, 42]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 351, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 554, + 518 + ], + "type": "text", + "content": "In the context of machine learning-based methods, more machine learning-based methods have been proposed, such as learning-based kernels [2, 3], convolutional feature extraction [6, 27], and other machine learning approaches have been devised specifically for handling single-illumination scenarios [35, 36, 41, 43, 48]. The most closely related methods to ours are Barron's CCC [2] and FFCC [3]. CCC reformulates color constancy as a color localization task in log-chromaticity space, using convolutional kernels to accurately locate the correct color. FFCC builds on this by applying the Fast Fourier Transform on a chromaticity torus, greatly improving the speed and accuracy of chromaticity estimation. However, both are global methods and lack direct applicability to local scenes unless explicitly provided with a local window." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 528, + 434, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 528, + 434, + 540 + ], + "spans": [ + { + "bbox": [ + 314, + 528, + 434, + 540 + ], + "type": "text", + "content": "2.2. Local AWB Methods" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "content": "Several multi-illuminant benchmarks have been introduced [4, 5, 8, 24, 31, 35, 39], with various methods proposed to tackle local white balance. These methods often rely on additional prior information or metadata, such as the number of light sources [4, 26] or facial regions of interest [5]. For instance, [28] uses flash photography to achieve local white balance. A novel framework in [4] treats the problem as an energy minimization task within a conditional random field, leveraging local light source estimates. [8] employs a global method with spatial variation, while [7] introduces a CNN framework based on sliding windows. Additionally, [46] applies a generative adversarial network (GAN) to correct images using synthetic datasets, adjusting the color based on the lighting information. More recently, Domislovic et" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26421" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "al. [14] proposed extracting spatial features for illumination prediction on each patch, and AID [32] introduces a deep model utilizing slot attention [37] to generate chromaticity and weight maps for each light source, with the added benefit of allowing user edits." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 133, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 295, + 335 + ], + "type": "text", + "content": "Among these approaches, most focus on pixel-level predictions, often leveraging encoder-decoder architectures (e.g., U-Net) to capture fine-grained illumination chromaticity details. However, this strategy presents several practical challenges. First, pixel-level prediction is computationally intensive, especially for high-resolution images, which significantly limits real-time applicability. Second, due to the lack of spatial context, pixel-wise color correction can overlook critical illumination chromaticity differences, leading to the loss of smooth illumination transitions and distortions in color edges, both of which are essential for accurate human visual perception. By treating each pixel independently, these methods risk introducing inconsistencies across the image and erasing subtle but important color shifts. As a result, their effectiveness in real-world applications, particularly in resource-constrained environments such as mobile devices, remains significantly limited." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 345, + 111, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 111, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 111, + 357 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 460 + ], + "type": "text", + "content": "We propose the Integral Fast Fourier Color Constancy (IFFCC) model, which builds upon the foundations of Fast Fourier Color Constancy (FFCC) [3]. Drawing inspiration from [40], we implement an histogram search for all possible target regions in image data within Cartesian coordinate space. This enables efficient retrieval of any region's histogram in linear computation time, avoiding redundant summation operations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "content": "FFCC Foundation. Consider a linear image stat " + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "content": " obtained from the camera, where there are no saturated pixels and the black level has been subtracted. Assuming the RGB value of any pixel is the product of its true albedo RGB value " + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "content": " and the uniform illumination value shared across all pixels: " + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "inline_equation", + "content": "I = W \\times L" + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "content": ". Although this assumption disregards shadows, dichromatic reflections, and spatial variations, it remains effective and widely accepted. Our aim is to estimate the " + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "content": ", and then " + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "inline_equation", + "content": "W = I / L" + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 568, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 592 + ], + "type": "text", + "content": "For an input RGB image " + }, + { + "bbox": [ + 55, + 568, + 295, + 592 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 568, + 295, + 592 + ], + "type": "text", + "content": ", CCC [2] defines two log-chroma measures for each pixel " + }, + { + "bbox": [ + 55, + 568, + 295, + 592 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 568, + 295, + 592 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 597, + 295, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 597, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 85, + 597, + 295, + 615 + ], + "type": "interline_equation", + "content": "u ^ {(i)} = \\log \\left(I _ {g} ^ {(i)} / I _ {r} ^ {(i)}\\right) v ^ {(i)} = \\log \\left(I _ {g} ^ {(i)} / I _ {b} ^ {(i)}\\right) \\tag {1}", + "image_path": "9c1845afa40b416d2f0d576f99e54e53eeac80ba5e23dd99bc87db118708daa5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 619, + 295, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 295, + 679 + ], + "type": "text", + "content": "Neglecting the absolute scaling of " + }, + { + "bbox": [ + 55, + 619, + 295, + 679 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 619, + 295, + 679 + ], + "type": "text", + "content": ", the estimation of " + }, + { + "bbox": [ + 55, + 619, + 295, + 679 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 619, + 295, + 679 + ], + "type": "text", + "content": " is then simplified as " + }, + { + "bbox": [ + 55, + 619, + 295, + 679 + ], + "type": "inline_equation", + "content": "L_{u} = \\log (L_{g} / L_{r}) \\quad L_{v} = \\log (L_{g} / L_{b})" + }, + { + "bbox": [ + 55, + 619, + 295, + 679 + ], + "type": "text", + "content": ". Given the unknown absolute scale, the inverse mapping from RGB to UV is challenging. To simplify the recovery process, we assume " + }, + { + "bbox": [ + 55, + 619, + 295, + 679 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 619, + 295, + 679 + ], + "type": "text", + "content": " is the unit-norm:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 685, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 685, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 62, + 685, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {r} = \\exp (- L _ {u}) / z \\quad L _ {g} = 1 / z \\quad L _ {b} = \\exp (- L _ {v}) / z \\tag {2} \\\\ z = \\sqrt {\\exp (- L _ {u}) ^ {2} + \\exp (- L _ {v}) ^ {2} + 1} \\\\ \\end{array}", + "image_path": "efedf18bf6d227d7f2a40a9c391b39f00379eb7de067747c817e2510614b20d0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "This logarithmic chromatic space method has several advantages over the RGB approach. First, it involves only two unknowns instead of three, which enhances numerical stability. Additionally, this framework transforms the multiplicative constraints related to " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " into additive constraints [15], effectively converting the color constancy task into a positioning task in a two-dimensional space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 156, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 228 + ], + "type": "text", + "content": "Following FFCC, we use FFT-based convolution in the log-chroma space, leveraging its periodicity to accelerate computations. Given the histogram's limited size, which does not fully capture the range of colors in natural images, a modular algorithm is needed to allow pixels to \"wrap around\" in this space:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 237, + 553, + 286 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 237, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 319, + 237, + 553, + 286 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} H (x, y) = \\sum_ {i} \\left(\\mod \\left((u ^ {(i)} - u _ {s}) / b - x, n\\right) < 1 \\right. \\tag {3} \\\\ \\wedge \\mod \\left(\\left(v ^ {(i)} - v _ {s}\\right) / b - y, n\\right) < 1\\bigg) \\\\ \\end{array}", + "image_path": "57d70d671043278db1593d1c4e7080e638de7439ce845b851a118b45b0e8bc3d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "inline_equation", + "content": "H(x,y)" + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "content": " represents the count of pixels in " + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "content": " whose log-chroma coordinates are close to the " + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "content": " coordinates associated with the histogram position " + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "content": " is the number of bins, and " + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "content": " is the bin size. " + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "inline_equation", + "content": "u_{s}" + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "content": " denote the starting positions of " + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "content": ", respectively. The operations here differ from those in CCC. In this case, the " + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "content": " positions in the histogram no longer represent absolute colors, but rather a set of " + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 295, + 553, + 438 + ], + "type": "text", + "content": " color coordinates. The filtered results produce a set of light sources, from which we apply the \"gray light de-aliasing\" method to resolve the resulting light source clusters. This method assumes that the light sources are as close as possible to the center of the histogram." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 439, + 554, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 554, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 554, + 522 + ], + "type": "text", + "content": "FFCC employs the Differentiable Bivariate von Mises (BVM) [38] method to address the single estimation of light sources from the toroidal PDF. Unlike FFCC, we need to predict multiple light sources for each image. Therefore, the BVM method is parallelized across multiple windows to predict several light sources simultaneously, as described in Section 3.3. The final convolutional structure is as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 533, + 553, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 533, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 358, + 533, + 553, + 558 + ], + "type": "interline_equation", + "content": "P = \\operatorname {s o f t m a x} \\left(B + M \\circ \\sum_ {s} H _ {s} * F _ {s}\\right) \\tag {4}", + "image_path": "a176a962c20a8a3a6d2a5f39d6304abf273d7ed3b2e3ddf3b705490f4d3bd5ff.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "text", + "content": " is the bias map, " + }, + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "text", + "content": " is the learnable gain map, and " + }, + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "inline_equation", + "content": "\\{F_s\\}" + }, + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "text", + "content": " is the set of learnable filters convolved with the edge-enhanced histogram group " + }, + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "inline_equation", + "content": "\\{H_s\\}" + }, + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 612, + 446, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 446, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 446, + 624 + ], + "type": "text", + "content": "3.1. Integral UV Histogram" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "FFCC predicts a global illumination for the image. However, when multiple illuminations are anticipated for different target regions, the process requires FFCC to repeatedly extract histograms from these various areas, significantly increasing computational load and processing time. Inspired by the work of [40], we efficiently compute the histograms for all potential regions within a given image" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26422" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 155, + 162 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 155, + 162 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 155, + 162 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 155, + 162 + ], + "type": "image", + "image_path": "9fa8d7fe5135bcbc8551dc46a1cfd948d3d409523a7929c5db461a3bcc871ded.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 163, + 154, + 171 + ], + "lines": [ + { + "bbox": [ + 69, + 163, + 154, + 171 + ], + "spans": [ + { + "bbox": [ + 69, + 163, + 154, + 171 + ], + "type": "text", + "content": "(a) Integral UV histogram" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 78, + 232, + 156 + ], + "blocks": [ + { + "bbox": [ + 156, + 78, + 232, + 156 + ], + "lines": [ + { + "bbox": [ + 156, + 78, + 232, + 156 + ], + "spans": [ + { + "bbox": [ + 156, + 78, + 232, + 156 + ], + "type": "image", + "image_path": "c22d7a3a899c0ae6c6527a8836033d10d8d3519e3fa7f1d558c0f4b712741192.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 161, + 228, + 171 + ], + "lines": [ + { + "bbox": [ + 160, + 161, + 228, + 171 + ], + "spans": [ + { + "bbox": [ + 160, + 161, + 228, + 171 + ], + "type": "text", + "content": "(b) Overlapping grid" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 234, + 79, + 309, + 158 + ], + "blocks": [ + { + "bbox": [ + 234, + 79, + 309, + 158 + ], + "lines": [ + { + "bbox": [ + 234, + 79, + 309, + 158 + ], + "spans": [ + { + "bbox": [ + 234, + 79, + 309, + 158 + ], + "type": "image", + "image_path": "d0adb3ec75ce2066cbdd29ef80d6c335fa17045106cb0f51e895a248ddc2f1da.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 161, + 302, + 171 + ], + "lines": [ + { + "bbox": [ + 238, + 161, + 302, + 171 + ], + "spans": [ + { + "bbox": [ + 238, + 161, + 302, + 171 + ], + "type": "text", + "content": "(c) Local histogram" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 70, + 382, + 164 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 382, + 164 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 382, + 164 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 382, + 164 + ], + "type": "image", + "image_path": "e9c2b23ee513c98e8c3e89e53bf83048f53676751e8e385a4d05a766e6230922.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 161, + 380, + 171 + ], + "lines": [ + { + "bbox": [ + 309, + 161, + 380, + 171 + ], + "spans": [ + { + "bbox": [ + 309, + 161, + 380, + 171 + ], + "type": "text", + "content": "(d) Parallel prediction" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 182, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 54, + 182, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 54, + 182, + 555, + 260 + ], + "type": "text", + "content": "Figure 2. Overview of the IFFCC process for predicting multi-illuminant scenes. We precompute the integral UV histogram of the input image in the log-chroma space (a). Multiple target regions are extracted on the integrated UV histogram with the overlapping grid (b), and the local log-chroma histogram (c) corresponding to each region is computed using simple arithmetic operations on the UV histogram. Each histogram forms a torus that can be unfolded, producing a cluster of potential light sources. A de-aliasing step selects the most likely light source for each region (d), represented by a point within an ellipse, where the ellipse indicates IFFCC's covariance output. Final light estimates are smoothed through interpolation and guided filtering, creating a continuous illumination map (e) that combines with the input to produce a white-balanced image (f) for multi-illuminant scenes." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 384, + 79, + 460, + 159 + ], + "blocks": [ + { + "bbox": [ + 384, + 79, + 460, + 159 + ], + "lines": [ + { + "bbox": [ + 384, + 79, + 460, + 159 + ], + "spans": [ + { + "bbox": [ + 384, + 79, + 460, + 159 + ], + "type": "image", + "image_path": "a158bceeb65724cc355c40269d2aaab2c1c4edd3fad9f2126ce093515cd1e9cc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 161, + 458, + 171 + ], + "lines": [ + { + "bbox": [ + 387, + 161, + 458, + 171 + ], + "spans": [ + { + "bbox": [ + 387, + 161, + 458, + 171 + ], + "type": "text", + "content": "(e) Spatial smoothing" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 463, + 79, + 540, + 159 + ], + "blocks": [ + { + "bbox": [ + 463, + 79, + 540, + 159 + ], + "lines": [ + { + "bbox": [ + 463, + 79, + 540, + 159 + ], + "spans": [ + { + "bbox": [ + 463, + 79, + 540, + 159 + ], + "type": "image", + "image_path": "dc2acfde9366f71c35a851c6a2567d933b4de6f20ffaea42c249a67e0f61159c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 161, + 531, + 171 + ], + "lines": [ + { + "bbox": [ + 476, + 161, + 531, + 171 + ], + "spans": [ + { + "bbox": [ + 476, + 161, + 531, + 171 + ], + "type": "text", + "content": "(f) Output image" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 281, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 296, + 435 + ], + "type": "text", + "content": "using Cartesian coordinates. By leveraging the spatial positions of data in this coordinate system, we propagate a function starting from the origin to navigate through the remaining positions. The histogram for the current point is then updated based on the histograms of previously processed points. Once the integral histogram for each data point is obtained, the histogram for the target region can be derived using simple addition and subtraction operations. This method eliminates the need to reconstruct separate histograms for each region individually. As a result, this efficient approach significantly reduces computational overhead and improves processing time, enabling rapid analysis of multiple areas within the image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 437, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 460 + ], + "type": "text", + "content": "After transforming the image into log-chroma space by Eq.1, we define the integral uv histogram as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 110, + 468, + 295, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 468, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 110, + 468, + 295, + 498 + ], + "type": "interline_equation", + "content": "H _ {\\text {I n t e g r a l}} \\left(x ^ {n}, b\\right) = \\bigcup_ {m = 0} ^ {n} B (u, v) ^ {m} \\tag {5}", + "image_path": "989b8c67c9da709a7c49435bcfedb8ab575b3792ce92f8baebd18a0c41042e88.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 69, + 505, + 295, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 505, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 69, + 505, + 295, + 559 + ], + "type": "interline_equation", + "content": "B (u, v) = \\left( \\begin{array}{l} \\mod \\left(\\operatorname {r o u n d} \\left(\\frac {u - u _ {s}}{b}\\right), n\\right), \\\\ \\mod \\left(\\operatorname {r o u n d} \\left(\\frac {v - v _ {s}}{b}\\right), n\\right) \\end{array} \\right) \\tag {6}", + "image_path": "ae953c2b1fff81b9fcc8c5a5c0e412d0124c3e4bc478f46191a5855f6e1b37ce.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "inline_equation", + "content": "x^n" + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "text", + "content": " is a sequence of points " + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "inline_equation", + "content": "x^0, x^1, \\ldots, x^n" + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "inline_equation", + "content": "B(\\cdot)" + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "text", + "content": " represents the corresponding bin for the current data point, " + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\bigcup" + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "text", + "content": " denotes the union operator, and the value of bin " + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{Integral}}(x^n, b)" + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "text", + "content": " is the sum of the histogram values of bin " + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 563, + 296, + 635 + ], + "type": "text", + "content": " from all previously traversed data points. The integral histogram can also be expressed recursively as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 69, + 646, + 295, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 646, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 69, + 646, + 295, + 663 + ], + "type": "interline_equation", + "content": "H _ {\\text {I n t e g r a l}} \\left(x ^ {m}, b\\right) = H _ {\\text {I n t e g r a l}} \\left(x ^ {m - 1}, b\\right) \\bigcup B (u, v) ^ {m} \\tag {7}", + "image_path": "72a4fcf8cc672712478293e3b17f944eb44fa6a88ca5a222df9ec995fdb06b4f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 677, + 253, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 253, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 253, + 689 + ], + "type": "text", + "content": "where at the initial data point, " + }, + { + "bbox": [ + 55, + 677, + 253, + 689 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{Integral}}(0,b) = 0" + }, + { + "bbox": [ + 55, + 677, + 253, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 690, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 690, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 690, + 295, + 713 + ], + "type": "text", + "content": "In log-chroma space, updating the histogram at the current point requires the histograms of the data points above," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 281, + 555, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 388 + ], + "type": "text", + "content": "the left, and the top-left of the current point. These three histograms have already been calculated in previous scans, so only three simple arithmetic operations are needed to update the histogram for the current point. Afterward, the bin value " + }, + { + "bbox": [ + 313, + 281, + 555, + 388 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 281, + 555, + 388 + ], + "type": "text", + "content": " for the current point is calculated, and the histogram is updated accordingly. Before each propagation step, all scanned bin values are copied to the current bin values, which can be efficiently handled through fast hardware or pointer operations." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "text", + "content": "For an RGB image converted to " + }, + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "inline_equation", + "content": "N_{1} \\times N_{2} \\times 2" + }, + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "text", + "content": " log-chroma space data, the integral histogram propagated via wavefront scanning from the top-left starting point can be written as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 329, + 429, + 553, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 429, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 329, + 429, + 553, + 472 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} H _ {\\text {I n t e g r a l}} (u, v, b) = H _ {\\text {I n t e g r a l}} (u - 1, v, b) \\\\ + H _ {\\text {I n t e g r a l}} (u, v - 1, b) \\tag {8} \\\\ - H _ {\\text {I n t e g r a l}} (u - 1, v - 1, b) + B (u, v) \\\\ \\end{array}", + "image_path": "f7605d3a9ce3dd90051b3d97470d6700928a09b0ef2e2370edc0e64e94b4bd1f.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 476, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 554, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 476, + 554, + 548 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 476, + 554, + 548 + ], + "type": "text", + "content": " is the coordinate of the log-chroma space. In our method, the data in log-chroma space can be represented by a two-dimensional floating point tensor of a two-dimensional scene [40], allowing us to calculate the computational load ratio between the traditional histogram method and the integral histogram method as follows:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 367, + 552, + 553, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 552, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 367, + 552, + 553, + 578 + ], + "type": "interline_equation", + "content": "R = \\frac {\\left[ 2 1 7 M _ {1} M _ {2} + 4 B ^ {2} \\right] S _ {1} S _ {2}}{2 7 4 + 1 1 (6 8 + 4 B ^ {2}) S _ {1} S _ {2}} \\tag {9}", + "image_path": "7348e8f914a95d7061d55db35a266c84955450236e5d68c87bee7a1b83695a99.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "inline_equation", + "content": "M_{2}" + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "text", + "content": " represent the size of the target region in the two-dimensional scene, " + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "text", + "content": " is the number of bins, and " + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "text", + "content": " are the different scales used when extracting histograms. The result of the calculation for " + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "inline_equation", + "content": "M_{1} = M_{2} = 128" + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "inline_equation", + "content": "B = 64" + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "inline_equation", + "content": "S_{1} = S_{2} = 1" + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "text", + "content": " is approximately " + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "inline_equation", + "content": "R = 19.71" + }, + { + "bbox": [ + 313, + 582, + 554, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 660, + 422, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 422, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 422, + 673 + ], + "type": "text", + "content": "3.2. Overlapping Grid" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "During inference, we use a sliding window with overlap to efficiently retrieve histograms for sub-regions in the log-chroma space, as illustrated in Fig.2. By pre-computing an" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26423" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "integral histogram over the entire scene, histograms for any arbitrary sub-region can be directly retrieved using Eq.8, without reconstructing each one from scratch, which significantly reduces computational costs. This overlapping grid method captures local chromatic variations by sliding a defined window across the scene and calculating histograms incrementally. The overlapping regions between windows maintain continuity and minimize edge effects, enhancing the accuracy of chromatic analysis throughout the scene." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 186, + 167, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 167, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 167, + 198 + ], + "type": "text", + "content": "3.3. Parallel Prediction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "text", + "content": "To speed up inference, our method requires only one inference to predict several illumination across multiple target regions within the image. Given an image's integral UV histogram in log-chroma space, denoted as " + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{integral}}(H,W,B,B,N)" + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "text", + "content": " represent the image size, " + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "text", + "content": " is the number of bins, and " + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "text", + "content": " is the number of augmented channel's histogram. The overlapping grid approach can be applied to obtain a collection " + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "text", + "content": " target region histograms. This collection " + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "text", + "content": ", represented as " + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "inline_equation", + "content": "X(k,B,B,N)" + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "text", + "content": ", includes histograms for each specific target region in the log-chroma space. The process of calculating the PDF of " + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 203, + 296, + 347 + ], + "type": "text", + "content": " in the frequency domain is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 353, + 295, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 353, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 80, + 353, + 295, + 419 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} H = \\operatorname {I F F T} \\left(\\sum_ {k} \\operatorname {F F T} (X) \\cdot \\operatorname {F F T} (F)\\right) + B (10) \\\\ P = \\frac {\\exp^ {(H - \\max (H))}}{\\sum_ {\\mathrm {d i m s}} \\exp^ {(H - \\max (H))}} (11) \\\\ \\end{array}", + "image_path": "e64d720821705bad9915784cb32e687cc8d6bf1bb40356a1a90819af5aa30abe.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 422, + 295, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 457 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 422, + 295, + 457 + ], + "type": "inline_equation", + "content": "\\mathrm{FFT}(.)" + }, + { + "bbox": [ + 55, + 422, + 295, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 422, + 295, + 457 + ], + "type": "inline_equation", + "content": "\\mathrm{IFFT}(.)" + }, + { + "bbox": [ + 55, + 422, + 295, + 457 + ], + "type": "text", + "content": " are the fast Fourier transform and its inverse, respectively, " + }, + { + "bbox": [ + 55, + 422, + 295, + 457 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 422, + 295, + 457 + ], + "type": "text", + "content": " is the convolution kernel, and " + }, + { + "bbox": [ + 55, + 422, + 295, + 457 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 422, + 295, + 457 + ], + "type": "text", + "content": " is the bias map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "text", + "content": "To parallelize the estimation of the mean of the Bivariate Von Mises (BVM) distribution from a histogram, we calculate circular mean values for the histogram coordinates " + }, + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "text", + "content": " across multiple regions concurrently. This process enables effective mean direction estimation within the toroidal data space, essential for capturing periodic patterns in wrapped grid data. For each region " + }, + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "inline_equation", + "content": "\\mu = \\{\\mu_k\\}_{k=1}^K" + }, + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "inline_equation", + "content": "\\mu_k" + }, + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 558, + 295, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 558, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 69, + 558, + 295, + 586 + ], + "type": "interline_equation", + "content": "\\mu_ {k} = \\left[ \\begin{array}{l} u _ {s} \\\\ v _ {s} \\end{array} \\right] + b \\left[ \\begin{array}{l} \\bmod \\left(\\frac {n}{2 \\pi} \\arctan 2 \\left(y _ {1} ^ {k}, x _ {1} ^ {k}\\right), n\\right) \\\\ \\bmod \\left(\\frac {n}{2 \\pi} \\arctan 2 \\left(y _ {2} ^ {k}, x _ {2} ^ {k}\\right), n\\right) \\end{array} \\right] \\tag {12}", + "image_path": "7a2c37f1285adc4e8fa7f75c14c722d34fbbe1dda9aa2510823d23f9ff49bec4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "text", + "content": "where the components " + }, + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "inline_equation", + "content": "y_1^k, x_1^k, y_2^k" + }, + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "inline_equation", + "content": "x_2^k" + }, + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "text", + "content": " are computed in parallel for each " + }, + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 60, + 620, + 295, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 620, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 60, + 620, + 295, + 656 + ], + "type": "interline_equation", + "content": "y _ {1} ^ {k} = \\sum_ {i} P _ {i} ^ {k} \\sin \\left(\\frac {2 \\pi i}{n}\\right), x _ {1} ^ {k} = \\sum_ {i} P _ {i} ^ {k} \\cos \\left(\\frac {2 \\pi i}{n}\\right) \\tag {13}", + "image_path": "b509477d5c9ee6df161c111bf6a36bf35ee0d46a3d2d7136a82c690f0a1fc85e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 652, + 272, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 652, + 272, + 682 + ], + "spans": [ + { + "bbox": [ + 62, + 652, + 272, + 682 + ], + "type": "interline_equation", + "content": "y _ {2} ^ {k} = \\sum_ {j} P _ {j} ^ {k} \\sin \\left(\\frac {2 \\pi j}{n}\\right), x _ {2} ^ {k} = \\sum_ {j} P _ {j} ^ {k} \\cos \\left(\\frac {2 \\pi j}{n}\\right)", + "image_path": "1d108dbad4e661ac36c3147bfd4832d18c9de7692907ed064534f38562aeb248.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "P_{i}^{k}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "P_{j}^{k}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " represent the marginal distributions for rows and columns of each target region's 2D histogram." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "After concurrently obtaining illumination values for all " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " target regions, we reconstruct the full illumination map by interpolating or applying edge-preserving filtering methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 115, + 424, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 115, + 424, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 115, + 424, + 129 + ], + "type": "text", + "content": "3.4. Spatial Smoothing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 133, + 554, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 554, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 554, + 192 + ], + "type": "text", + "content": "In multi-light scenarios, patch-wise illumination estimates often exhibit discrete, locally discontinuous characteristics. To reconstruct a smooth and seamless illumination map across the entire image, it is vital to postprocess these patch-level estimates using interpolation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 192, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 323 + ], + "type": "text", + "content": "We employ linear interpolation due to its simplicity and ability to create smooth, piecewise-linear transitions for estimating intermediate values. Although higher-degree polynomial interpolation could be used, it risks introducing oscillations. After getting estimates for " + }, + { + "bbox": [ + 313, + 192, + 555, + 323 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 192, + 555, + 323 + ], + "type": "text", + "content": " target regions, interpolation adjusts these values based on their relative locations, resulting in a smooth illumination map. Although interpolation enhances overall smoothness, it may blur local details, particularly in regions with strong light edges or steep illumination gradients, potentially leading to a loss of edge information." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 323, + 554, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 554, + 431 + ], + "type": "text", + "content": "Guided filtering [25] is then applied to preserve details at the boundaries and structural features of the illumination map, preventing blur or distortion during transitions. Overall, combining these techniques results in a globally coherent illumination distribution that maintains edge clarity while appearing smoother and more natural. Given the interpolated illumination map " + }, + { + "bbox": [ + 313, + 323, + 554, + 431 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 323, + 554, + 431 + ], + "type": "text", + "content": " and the raw image " + }, + { + "bbox": [ + 313, + 323, + 554, + 431 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 323, + 554, + 431 + ], + "type": "text", + "content": " as the guiding image, the output illumination map " + }, + { + "bbox": [ + 313, + 323, + 554, + 431 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 323, + 554, + 431 + ], + "type": "text", + "content": " at each pixel " + }, + { + "bbox": [ + 313, + 323, + 554, + 431 + ], + "type": "inline_equation", + "content": "O_{i}" + }, + { + "bbox": [ + 313, + 323, + 554, + 431 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 439, + 553, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 439, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 375, + 439, + 553, + 452 + ], + "type": "interline_equation", + "content": "O _ {i} = a _ {k} G _ {i} + b _ {k}, \\quad \\forall i \\in \\omega_ {k} \\tag {14}", + "image_path": "3994282430bd67279f94ceb26b3d15eb983ea17aea3f8b1925b9d97268c0467d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 457, + 553, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 457, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 318, + 457, + 553, + 488 + ], + "type": "interline_equation", + "content": "a _ {k} = \\frac {\\frac {1}{| \\omega |} \\sum_ {i \\in \\omega_ {k}} G _ {i} L _ {i} - \\bar {G} _ {k} \\bar {L} _ {k}}{\\sigma_ {k} ^ {2} + \\epsilon}, \\quad b _ {k} = \\bar {L} _ {k} - a _ {k} \\bar {G} _ {k} \\tag {15}", + "image_path": "434b404ef6b642a31472909131f8d18c52b7facaa0dd993b6c860e5a6b9abd67.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": "here, " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\bar{G}_k" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\bar{L}_k" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": " are the mean values of " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": " in window " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\omega_k" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\sigma_k^2" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": " as the variance of " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": ". The constant " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": " prevents division by zero. The final illumination map " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": " is obtained by weighted averaging." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 379, + 547, + 553, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 547, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 379, + 547, + 553, + 576 + ], + "type": "interline_equation", + "content": "O = \\frac {1}{| \\omega |} \\sum_ {k \\in \\omega_ {i}} \\left(a _ {k} G _ {i} + b _ {k}\\right) \\tag {16}", + "image_path": "b12c30bf3c1e9439e36fa96920cb8a6335adab5c576b98b012c7b15bc938c188.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 582, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 643 + ], + "type": "text", + "content": "This post-processing step is essential: it creates a coherent illumination map that accurately reflects the lighting conditions across the entire image while preserving critical features, thus the reconstructed illumination aligns well with the underlying content of the image." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 652, + 391, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 391, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 391, + 666 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 672, + 430, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 430, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 430, + 685 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "We evaluated our proposed IFFCC on two widely used multi-light source datasets: the LSMI dataset [31] and the" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26424" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 69, + 541, + 196 + ], + "blocks": [ + { + "bbox": [ + 75, + 69, + 541, + 196 + ], + "lines": [ + { + "bbox": [ + 75, + 69, + 541, + 196 + ], + "spans": [ + { + "bbox": [ + 75, + 69, + 541, + 196 + ], + "type": "table", + "html": "
MethodCanon_5dCanon_550dMotoPanasonicSonyUse all
meanmedianmeanmedianmeanmedianmeanmedianmeanmedianmeanmedian
Huassin and Akbari [29]13.8813.7613.7313.5413.0612.8913.6213.7314.2213.6813.6713.46
CRF(White-Patch)[4]7.665.967.695.776.945.327.585.697.905.867.195.44
Patch-based[24]4.853.114.612.984.172.494.582.954.963.164.302.89
Keypoint-based[24]5.593.885.483.944.973.285.423.795.763.825.463.59
Superpixel-based[24]4.883.684.793.564.192.894.553.544.973.824.203.10
IFFCC2.061.542.161.771.901.362.061.592.661.962.191.56
Bianco et al.[7]4.634.024.594.164.273.894.984.565.344.768.015.69
HypNet/SelNet[44]5.424.935.164.835.094.745.234.985.825.346.313.95
LSMI-U[31]3.142.332.922.462.792.043.172.703.582.883.722.40
Domislovic et al.[12]2.632.182.772.062.071.492.381.732.922.242.281.60
", + "image_path": "3888c9b6297409194c15f6e6a164c3be83ea2fc049b03b43214a78ec21b38cd9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 75, + 258, + 541, + 385 + ], + "blocks": [ + { + "bbox": [ + 55, + 203, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 55, + 203, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 555, + 248 + ], + "type": "text", + "content": "Table 1. The performance results on the Shadow dataset are presented. The dataset is segmented by images captured from different cameras, with \"use all\" representing the entire dataset. We report the mean and median angular errors for each subset. IFFCC employs the default settings, with a window size of 128 and an overlap size of 64. The model is trained using the blended illumination within each window as the ground truth (GT), and the corresponding error is computed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 258, + 541, + 385 + ], + "lines": [ + { + "bbox": [ + 75, + 258, + 541, + 385 + ], + "spans": [ + { + "bbox": [ + 75, + 258, + 541, + 385 + ], + "type": "table", + "html": "
MethodOutdoorIndoorNighttime
meanmedianb25w25meanmedianb25w25meanmedianb25w25
Huassin and Akbari[29]13.1912.897.1019.1214.8413.228.4320.1816.3514.969.7522.33
CRF(White-Patch)[4]7.315.662.3113.368.786.894.3614.406.965.162.3815.85
Patch-based[24]4.533.291.3810.475.584.432.0911.393.482.171.8912.53
Keypoint-based[24]5.773.881.2411.786.254.041.7912.603.994.101.4913.08
Superpixel-based[24]4.733.011.338.215.183.491.929.244.632.981.349.82
IFFCC1.741.280.423.833.042.140.666.712.851.530.547.43
Bianco et al.[7]4.862.361.9811.757.425.623.3812.508.225.783.1013.68
HypNet/SelNet[44]6.094.081.7712.317.075.272.9811.494.763.212.2012.59
LSMI-U[31]2.122.031.025.485.724.621.498.544.812.471.338.89
Domislovic et al.[12]1.781.460.654.344.453.310.987.463.681.890.838.22
", + "image_path": "395e39980f52a0318150e975873d56dcecdac2c67fb2c1f3d02dfaf68d257b43.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 392, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 555, + 437 + ], + "type": "text", + "content": "Table 2. The performance results on the Shadow dataset are presented. The dataset is segmented by images captured in different scenes, with each scene including images taken by five different cameras. We report the mean and median angular errors for each subset. IFFCC uses the default settings, with a window size of 128 and an overlap size of 64. The model is trained using the blended illumination within each window as the ground truth (GT), and the corresponding error is computed." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "text", + "content": "Shadow dataset [13]. The LSMI dataset includes over 7,486 images captured by three different cameras in a variety of scenes, each containing multiple light sources. The Shadow dataset consists of 2,500 images taken in both indoor and outdoor settings, with each image accompanied by a binary segmentation mask, where each region is illuminated by a single light source. In all experiments, full-sized training images were randomly cropped to " + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "text", + "content": ", and testing was conducted on " + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "text", + "content": " images. All images were black-level corrected, and calibration objects were masked." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": "IFFCC's computational efficiency enables CPU-only training and inference, eliminating the need for a GPU. All experiments were conducted on an Intel Xeon Gold 6258R CPU. Performance was evaluated using standard error metrics, including mean and median angular errors and the arithmetic means of the first and third quartiles (\"best " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": " \" and \"worst " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": " ). Additionally, we compared model parameter counts and CPU inference times for " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": " images. Since our method operates on patches, we calculate blended illumination values from the ground truth illumination map within a sliding window, using these as the ground" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 457, + 553, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 553, + 481 + ], + "type": "text", + "content": "truth for each patch. The impact of different training approaches is discussed in the ablation study." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 482, + 554, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 554, + 554 + ], + "type": "text", + "content": "During training, we input a set of enhanced histograms derived from filtered images that ensure non-negativity and intensity scaling consistency. Each image includes a local absolute deviation measure. Using the preconditioned frequency-domain optimization from FFCC [3], training runs for 64 iterations with a histogram size of " + }, + { + "bbox": [ + 313, + 482, + 554, + 554 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 482, + 554, + 554 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 563, + 457, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 457, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 457, + 576 + ], + "type": "text", + "content": "4.2. Results and Comparisons" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 581, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 554, + 700 + ], + "type": "text", + "content": "Quantitative Comparison We evaluated IFFCC against three traditional methods [4, 24, 29] and three learning-based methods [7, 12, 44] on the Shadow dataset. The experimental results are presented in Tables 1 and 2. Table 1 organizes the Shadow dataset into subsets based on images captured by different cameras, while Table 2 categorizes the dataset by various scene types, each containing photos taken with five distinct cameras. Despite comparisons with advanced network-based methods, IFFCC achieves state-of-the-art performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "The results in Table 2 highlight the strong general" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "26425" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 140, + 137 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 140, + 137 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 140, + 137 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 140, + 137 + ], + "type": "image", + "image_path": "08a937f70aae03599c928fd24a5b0ba0e383f9427fbb59c3c4026447a14e8ee5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 137, + 139, + 201 + ], + "blocks": [ + { + "bbox": [ + 58, + 137, + 139, + 201 + ], + "lines": [ + { + "bbox": [ + 58, + 137, + 139, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 137, + 139, + 201 + ], + "type": "image", + "image_path": "1bec36dacc91776241920f0dbea216be1b771b850927083ecc5d11d2b00c5edb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 202, + 139, + 266 + ], + "blocks": [ + { + "bbox": [ + 58, + 202, + 139, + 266 + ], + "lines": [ + { + "bbox": [ + 58, + 202, + 139, + 266 + ], + "spans": [ + { + "bbox": [ + 58, + 202, + 139, + 266 + ], + "type": "image", + "image_path": "56888799e5407b56c099a688485b917d9b3e861459154d8da45c002770b68bf6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 268, + 111, + 279 + ], + "lines": [ + { + "bbox": [ + 88, + 268, + 111, + 279 + ], + "spans": [ + { + "bbox": [ + 88, + 268, + 111, + 279 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 141, + 72, + 222, + 137 + ], + "blocks": [ + { + "bbox": [ + 141, + 72, + 222, + 137 + ], + "lines": [ + { + "bbox": [ + 141, + 72, + 222, + 137 + ], + "spans": [ + { + "bbox": [ + 141, + 72, + 222, + 137 + ], + "type": "image", + "image_path": "0d6c1f4d1c60575cb7891b98fd9494aaa5d858a2d4b95da0ac267ae4af85e675.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 141, + 137, + 222, + 201 + ], + "blocks": [ + { + "bbox": [ + 141, + 137, + 222, + 201 + ], + "lines": [ + { + "bbox": [ + 141, + 137, + 222, + 201 + ], + "spans": [ + { + "bbox": [ + 141, + 137, + 222, + 201 + ], + "type": "image", + "image_path": "af6be4370513d5bd7350b9d0bcd3315bcd3ee941292a55f3242fb07f087f3c81.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 141, + 201, + 222, + 266 + ], + "blocks": [ + { + "bbox": [ + 141, + 201, + 222, + 266 + ], + "lines": [ + { + "bbox": [ + 141, + 201, + 222, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 201, + 222, + 266 + ], + "type": "image", + "image_path": "bde56dc819d1defdbb5541be2a7f288b84702c48bbcde735bb6b4b59fd5f7b39.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 268, + 206, + 279 + ], + "lines": [ + { + "bbox": [ + 154, + 268, + 206, + 279 + ], + "spans": [ + { + "bbox": [ + 154, + 268, + 206, + 279 + ], + "type": "text", + "content": "Gijsenij et al." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 72, + 304, + 137 + ], + "blocks": [ + { + "bbox": [ + 223, + 72, + 304, + 137 + ], + "lines": [ + { + "bbox": [ + 223, + 72, + 304, + 137 + ], + "spans": [ + { + "bbox": [ + 223, + 72, + 304, + 137 + ], + "type": "image", + "image_path": "a352477d17e3c42cdedc1870e49f0d90d6890147c66e8827500ec030b1e26f9b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 223, + 137, + 304, + 201 + ], + "blocks": [ + { + "bbox": [ + 223, + 137, + 304, + 201 + ], + "lines": [ + { + "bbox": [ + 223, + 137, + 304, + 201 + ], + "spans": [ + { + "bbox": [ + 223, + 137, + 304, + 201 + ], + "type": "image", + "image_path": "55a775a8a6912962afa985239c7bf9bd4221cc3ae75832540de61b1f5c67221a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 223, + 201, + 304, + 266 + ], + "blocks": [ + { + "bbox": [ + 223, + 201, + 304, + 266 + ], + "lines": [ + { + "bbox": [ + 223, + 201, + 304, + 266 + ], + "spans": [ + { + "bbox": [ + 223, + 201, + 304, + 266 + ], + "type": "image", + "image_path": "7be33c47aac92232fd67ea8b0bc19b2208ad45f5877d437c76e2fb134c111f3c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 268, + 291, + 278 + ], + "lines": [ + { + "bbox": [ + 238, + 268, + 291, + 278 + ], + "spans": [ + { + "bbox": [ + 238, + 268, + 291, + 278 + ], + "type": "text", + "content": "Bianoco et al." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 386, + 137 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 386, + 137 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 386, + 137 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 386, + 137 + ], + "type": "image", + "image_path": "0dd66e408b39582773835b0802eaccfb3dee27a61af14c164757c3f2567c832c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 137, + 386, + 200 + ], + "blocks": [ + { + "bbox": [ + 306, + 137, + 386, + 200 + ], + "lines": [ + { + "bbox": [ + 306, + 137, + 386, + 200 + ], + "spans": [ + { + "bbox": [ + 306, + 137, + 386, + 200 + ], + "type": "image", + "image_path": "692b60b28a4e8c3437472f690710de1a647822075bbf7efd1ea832814eef9025.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 201, + 386, + 266 + ], + "blocks": [ + { + "bbox": [ + 306, + 201, + 386, + 266 + ], + "lines": [ + { + "bbox": [ + 306, + 201, + 386, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 386, + 266 + ], + "type": "image", + "image_path": "6d1b057fdc16654def0802470ac38a2a866c9c9c4401edbe5cc9883e6564e733.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 268, + 365, + 277 + ], + "lines": [ + { + "bbox": [ + 331, + 268, + 365, + 277 + ], + "spans": [ + { + "bbox": [ + 331, + 268, + 365, + 277 + ], + "type": "text", + "content": "LSMI-U" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 388, + 72, + 468, + 137 + ], + "blocks": [ + { + "bbox": [ + 388, + 72, + 468, + 137 + ], + "lines": [ + { + "bbox": [ + 388, + 72, + 468, + 137 + ], + "spans": [ + { + "bbox": [ + 388, + 72, + 468, + 137 + ], + "type": "image", + "image_path": "c5f3ef5d045df532156854e852da0238fefc914336b67d0f33af1da1da554168.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 388, + 137, + 468, + 200 + ], + "blocks": [ + { + "bbox": [ + 388, + 137, + 468, + 200 + ], + "lines": [ + { + "bbox": [ + 388, + 137, + 468, + 200 + ], + "spans": [ + { + "bbox": [ + 388, + 137, + 468, + 200 + ], + "type": "image", + "image_path": "578fa221ec4854d1bee0d613c308525238a53ee3919aae231ab6dd78f2fc4955.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 388, + 201, + 468, + 266 + ], + "blocks": [ + { + "bbox": [ + 388, + 201, + 468, + 266 + ], + "lines": [ + { + "bbox": [ + 388, + 201, + 468, + 266 + ], + "spans": [ + { + "bbox": [ + 388, + 201, + 468, + 266 + ], + "type": "image", + "image_path": "c7360708af7d35414bd03dcf2b61b266b35729567a32baf242b1c8ecec6aade1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 268, + 445, + 277 + ], + "lines": [ + { + "bbox": [ + 416, + 268, + 445, + 277 + ], + "spans": [ + { + "bbox": [ + 416, + 268, + 445, + 277 + ], + "type": "text", + "content": "IFFCC" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 470, + 72, + 551, + 137 + ], + "blocks": [ + { + "bbox": [ + 470, + 72, + 551, + 137 + ], + "lines": [ + { + "bbox": [ + 470, + 72, + 551, + 137 + ], + "spans": [ + { + "bbox": [ + 470, + 72, + 551, + 137 + ], + "type": "image", + "image_path": "b011d90704d890e79777db34f213bb54497172c95b761511802578ff7a2d10b5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 470, + 137, + 551, + 200 + ], + "blocks": [ + { + "bbox": [ + 470, + 137, + 551, + 200 + ], + "lines": [ + { + "bbox": [ + 470, + 137, + 551, + 200 + ], + "spans": [ + { + "bbox": [ + 470, + 137, + 551, + 200 + ], + "type": "image", + "image_path": "d577927f0c885928d6cb4d026c736fdbd2a8c232ba6488c7ce0b9ddc1aaa7a32.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 470, + 201, + 551, + 266 + ], + "blocks": [ + { + "bbox": [ + 470, + 201, + 551, + 266 + ], + "lines": [ + { + "bbox": [ + 470, + 201, + 551, + 266 + ], + "spans": [ + { + "bbox": [ + 470, + 201, + 551, + 266 + ], + "type": "image", + "image_path": "a538d4fffeae53398142dc16c7a2d39c8ca8917514bb2c6236a4308133f96165.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 268, + 517, + 277 + ], + "lines": [ + { + "bbox": [ + 503, + 268, + 517, + 277 + ], + "spans": [ + { + "bbox": [ + 503, + 268, + 517, + 277 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 59, + 323, + 156, + 385 + ], + "blocks": [ + { + "bbox": [ + 55, + 288, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 288, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 555, + 312 + ], + "type": "text", + "content": "Figure 3. Visualization of the comparison results on the LSMI dataset. For IFFCC, the window size is set to 128 with an overlap of 64. In other patch-based methods, the window size is set to the default value of 32." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 323, + 156, + 385 + ], + "lines": [ + { + "bbox": [ + 59, + 323, + 156, + 385 + ], + "spans": [ + { + "bbox": [ + 59, + 323, + 156, + 385 + ], + "type": "image", + "image_path": "0ad81bbb352626a96769bff158d51f6a6ff903c8928395347a2ffc805588863d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 59, + 386, + 156, + 446 + ], + "blocks": [ + { + "bbox": [ + 59, + 386, + 156, + 446 + ], + "lines": [ + { + "bbox": [ + 59, + 386, + 156, + 446 + ], + "spans": [ + { + "bbox": [ + 59, + 386, + 156, + 446 + ], + "type": "image", + "image_path": "ba26574610174747eb5ae94244c0ed0028e0ffd54ed1c48f554b9e3f8a4270b9.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 449, + 119, + 460 + ], + "lines": [ + { + "bbox": [ + 97, + 449, + 119, + 460 + ], + "spans": [ + { + "bbox": [ + 97, + 449, + 119, + 460 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 158, + 323, + 255, + 385 + ], + "blocks": [ + { + "bbox": [ + 158, + 323, + 255, + 385 + ], + "lines": [ + { + "bbox": [ + 158, + 323, + 255, + 385 + ], + "spans": [ + { + "bbox": [ + 158, + 323, + 255, + 385 + ], + "type": "image", + "image_path": "59f71e4f4ddd0ad0117802a538f575941be9f2e3911c7f9027a9777d2d017bc8.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 158, + 386, + 255, + 445 + ], + "blocks": [ + { + "bbox": [ + 158, + 386, + 255, + 445 + ], + "lines": [ + { + "bbox": [ + 158, + 386, + 255, + 445 + ], + "spans": [ + { + "bbox": [ + 158, + 386, + 255, + 445 + ], + "type": "image", + "image_path": "3bd6ee5f171edd36fd445a4186ea3a8821cf802b79ea5cf9d44af44020dcf580.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 449, + 230, + 459 + ], + "lines": [ + { + "bbox": [ + 182, + 449, + 230, + 459 + ], + "spans": [ + { + "bbox": [ + 182, + 449, + 230, + 459 + ], + "type": "text", + "content": "White Patch" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 257, + 323, + 353, + 385 + ], + "blocks": [ + { + "bbox": [ + 257, + 323, + 353, + 385 + ], + "lines": [ + { + "bbox": [ + 257, + 323, + 353, + 385 + ], + "spans": [ + { + "bbox": [ + 257, + 323, + 353, + 385 + ], + "type": "image", + "image_path": "ac70c95c817f221dfc4bd0d450aca068649ca126eebb87fcef44bcabf9f288ee.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 257, + 386, + 353, + 446 + ], + "blocks": [ + { + "bbox": [ + 257, + 386, + 353, + 446 + ], + "lines": [ + { + "bbox": [ + 257, + 386, + 353, + 446 + ], + "spans": [ + { + "bbox": [ + 257, + 386, + 353, + 446 + ], + "type": "image", + "image_path": "33380cca317a2900d5dfb00acc3485cbe98211611d31c1440736d4f80a6e7d42.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 449, + 318, + 459 + ], + "lines": [ + { + "bbox": [ + 290, + 449, + 318, + 459 + ], + "spans": [ + { + "bbox": [ + 290, + 449, + 318, + 459 + ], + "type": "text", + "content": "IFFCC" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 355, + 323, + 452, + 385 + ], + "blocks": [ + { + "bbox": [ + 355, + 323, + 452, + 385 + ], + "lines": [ + { + "bbox": [ + 355, + 323, + 452, + 385 + ], + "spans": [ + { + "bbox": [ + 355, + 323, + 452, + 385 + ], + "type": "image", + "image_path": "c533d90204ff306db7966fe16e36c4ae3c872714971aae4d8272aca8202eb307.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 355, + 386, + 452, + 446 + ], + "blocks": [ + { + "bbox": [ + 355, + 386, + 452, + 446 + ], + "lines": [ + { + "bbox": [ + 355, + 386, + 452, + 446 + ], + "spans": [ + { + "bbox": [ + 355, + 386, + 452, + 446 + ], + "type": "image", + "image_path": "be30854263048d763a57a87751f48c3eaaf6127646d44cc082414eda1d18e041.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 449, + 441, + 460 + ], + "lines": [ + { + "bbox": [ + 374, + 449, + 441, + 460 + ], + "spans": [ + { + "bbox": [ + 374, + 449, + 441, + 460 + ], + "type": "text", + "content": "Illumination map" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 469, + 551, + 480 + ], + "lines": [ + { + "bbox": [ + 56, + 469, + 551, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 469, + 551, + 480 + ], + "type": "text", + "content": "Figure 4. Visualization of illumination comparison for different patch-based methods. The window size is 32, with an overlap size of 16." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 454, + 323, + 552, + 385 + ], + "blocks": [ + { + "bbox": [ + 454, + 323, + 552, + 385 + ], + "lines": [ + { + "bbox": [ + 454, + 323, + 552, + 385 + ], + "spans": [ + { + "bbox": [ + 454, + 323, + 552, + 385 + ], + "type": "image", + "image_path": "72b93274c4df36efdfd8fd403d295b30e491785bb00de686dd3a2a11099b491c.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 454, + 386, + 552, + 446 + ], + "blocks": [ + { + "bbox": [ + 454, + 386, + 552, + 446 + ], + "lines": [ + { + "bbox": [ + 454, + 386, + 552, + 446 + ], + "spans": [ + { + "bbox": [ + 454, + 386, + 552, + 446 + ], + "type": "image", + "image_path": "04bc55f93adc5ba4720e9b02f5a2a0ea3e4c3e23ec3bb8a4e214a9b4039d7dd1.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 449, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 459, + 449, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 459, + 449, + 545, + 460 + ], + "type": "text", + "content": "White-balanced image" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "bbox": [ + 54, + 502, + 295, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 502, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 54, + 502, + 295, + 657 + ], + "type": "text", + "content": "ization ability of IFFCC. Network-based methods generally perform worse than IFFCC on multi-camera datasets, where each camera's unique spectral sensitivities complicate learning a single mapping across different devices. IFFCC generalizes well across multiple cameras because it leverages an integral log-chroma histogram approach, which directly models chromaticity distributions without being overly reliant on the specific spectral sensitivities of individual cameras. This histogram-based method effectively captures illumination variations across different scenes and cameras, allowing IFFCC to produce robust predictions without requiring extensive fine-tuning for each camera's unique spectral characteristics." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "The experimental results on the LSMI dataset, presented in Tables 3, and 4, provide a comparative analysis of patch-based and pixel-level methods. The test set is categorized into Multi (multiple light sources), and Mixed (compris" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 502, + 555, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 584 + ], + "type": "text", + "content": "ing both single and multiple light sources). Among patch-based approaches, IFFCC demonstrates state-of-the-art performance, achieving results comparable to those of leading pixel-level methods, including AID [32]. Notably, IFFCC's compact parameter count and rapid processing speed render it highly suitable for deployment on mobile and resource-constrained devices." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 590, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 638 + ], + "type": "text", + "content": "We also compared the processing time for a " + }, + { + "bbox": [ + 313, + 590, + 555, + 638 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 590, + 555, + 638 + ], + "type": "text", + "content": " image in a multi-illuminant scene using FFCC and IFFCC, as shown in Table 5. As the number of windows increases, IF-FCC significantly outperforms FFCC in processing speed." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Qualitative Comparison Fig.3 visually compares our method against alternative approaches on the LSMI dataset. To enhance visualization quality, color correction, gamma correction, and tone mapping were applied to the white-balanced images. The results indicate that IFFCC achieves more accurate white balance restoration than competing" + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26426" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 289, + 186 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 289, + 186 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 289, + 186 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 289, + 186 + ], + "type": "table", + "html": "
MethodMultiParams (M)Time (s)
Galaxy mean med.Nikon mean med.Sony mean med.
Patch -levelGijsenij[23]12.389.579.316.5911.198.14-1.3
Bianoco[7]5.564.344.654.194.383.940.16-
IFFCC2.481.902.301.502.481.970.0120.03
Pix2Pix[30]4.282.634.182.764.373.265.40.48
Pixel LSMI-H[31]3.132.703.203.013.653.330.520.28
-levelLSMI-U[31]2.852.552.361.842.762.385.70.51
AID[32]2.031.432.261.392.161.646.4>1
", + "image_path": "1dae61065cc8e0656026aba8cf2e4424ce6d03d57053b118b814f8b1216a5032.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 237, + 288, + 353 + ], + "blocks": [ + { + "bbox": [ + 55, + 193, + 295, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 193, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 295, + 227 + ], + "type": "text", + "content": "Table 3. The performance results on the multi-illumination images from the LSMI dataset are presented, with testing time corresponding to the processing time for " + }, + { + "bbox": [ + 55, + 193, + 295, + 227 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 55, + 193, + 295, + 227 + ], + "type": "text", + "content": " images on a CPU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 237, + 288, + 353 + ], + "lines": [ + { + "bbox": [ + 62, + 237, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 62, + 237, + 288, + 353 + ], + "type": "table", + "html": "
MethodMixedParams (M)Time (s)
Galaxy mean med.Nikon mean med.Sony mean med.
Patch -levelGijsenij[23]10.097.439.215.839.467.13-1.3
Bianoco[7]4.893.833.933.483.843.210.16-
IFFCC1.981.792.111.481.931.720.0120.03
Pix2Pix[30]5.662.445.412.494.202.205.40.48
Pixel LSMI-H[31]3.062.542.922.633.212.850.520.28
-level LSMI-U[31]2.631.912.161.552.682.295.70.51
AID[32]1.631.321.761.141.681.446.4>1
", + "image_path": "5c889eceed5f07f00df441ff46890d1341ef40fefc7c8f4d94382f4e7ac1ed49.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 415, + 293, + 448 + ], + "blocks": [ + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "type": "text", + "content": "Table 4. The performance results on the mixed-illumination images (containing both single and multiple light sources) from the LSMI dataset are presented, with testing time corresponding to the processing time for " + }, + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "type": "text", + "content": " images on a CPU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 415, + 293, + 448 + ], + "lines": [ + { + "bbox": [ + 58, + 415, + 293, + 448 + ], + "spans": [ + { + "bbox": [ + 58, + 415, + 293, + 448 + ], + "type": "table", + "html": "
[32,0][32,16][64,32][64,48][128,64][128,96]
FFCC162ms450ms203ms689ms88ms234ms
IFFCC34ms46ms29ms40ms27ms33ms
", + "image_path": "121045d95f5e8995edb90824e37a7ec859129aec0713cfb90000eca93144eb4f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 509, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 570 + ], + "type": "text", + "content": "methods. We also present a comparison of illumination maps between our method and patch-based approaches, as shown in Fig.4. The window size is set to " + }, + { + "bbox": [ + 55, + 509, + 296, + 570 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 509, + 296, + 570 + ], + "type": "text", + "content": " with an overlap of 16. The results demonstrate that IFFCC achieves a more precise and smoother illumination prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 149, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 149, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 149, + 590 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Table 6 presents results from ablation experiments analyzing the impact of various training strategies, window sizes, and overlap regions. Here, \"B\" represents training with a blend of illumination maps as the ground truth (GT), and \"M\" uses the dominant light source as GT. Parameters \"ws\" and \"os\" indicate window and overlap sizes, respectively. Findings reveal that smaller window sizes, with limited information, reduce prediction accuracy. However, the blended illumination strategy provides improved accuracy compared to using a single dominant light source. Visual" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 345, + 70, + 529, + 182 + ], + "blocks": [ + { + "bbox": [ + 55, + 456, + 295, + 501 + ], + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 501 + ], + "type": "text", + "content": "Table 5. Comparison of average testing time for FFCC and IF-FCC on the Shadow dataset, evaluated across different window and overlap sizes. FFCC employs a conventional histogram computation approach." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 70, + 529, + 182 + ], + "lines": [ + { + "bbox": [ + 345, + 70, + 529, + 182 + ], + "spans": [ + { + "bbox": [ + 345, + 70, + 529, + 182 + ], + "type": "table", + "html": "
type[ws, os]meanmedianb25w25
B[32, 16]2.652.171.374.65
B[64, 0]2.352.011.134.31
M[64, 32]2.532.041.254.83
B[64, 32]2.371.871.164.38
M[128, 0]2.462.031.044.78
B[128, 0]2.161.680.914.23
M[128, 32]2.501.881.044.91
B[128, 32]2.141.640.854.30
M[128, 64]2.421.950.934.71
B[128, 64]2.061.540.814.19
B[128, 96]2.131.590.834.25
B[192, 128]1.981.340.724.08
", + "image_path": "170639a8111eda06b64c9933611c1e45499fe109289f25eeaf9f308ea695c33a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 190, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 313, + 190, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 555, + 224 + ], + "type": "text", + "content": "Table 6. Ablation study on the Shadow dataset (Canon5d), reporting the impact of different window sizes, overlap sizes, and training strategies on accuracy." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 234, + 395, + 347 + ], + "blocks": [ + { + "bbox": [ + 317, + 234, + 395, + 347 + ], + "lines": [ + { + "bbox": [ + 317, + 234, + 395, + 347 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 395, + 347 + ], + "type": "image", + "image_path": "1af028746080cd044c89b20f5f01098b51ecb64f29ab673ec13a63830e9752ec.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 356, + 555, + 411 + ], + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 411 + ], + "type": "text", + "content": "Figure 5. (a)-(d) show illumination maps generated with different window sizes ([32, 16], [64, 32], [128, 64], [192, 128]), each processed with smooth transition adjustments. (e) presents the ground truth (GT) illumination map, and (f) displays the white-balanced image obtained by combining the input image with the GT." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 395, + 235, + 473, + 346 + ], + "blocks": [ + { + "bbox": [ + 395, + 235, + 473, + 346 + ], + "lines": [ + { + "bbox": [ + 395, + 235, + 473, + 346 + ], + "spans": [ + { + "bbox": [ + 395, + 235, + 473, + 346 + ], + "type": "image", + "image_path": "a75bba796124fc67a525818245f86163fea8d2a4ec33384501ef87c136c35c17.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 474, + 235, + 552, + 346 + ], + "blocks": [ + { + "bbox": [ + 474, + 235, + 552, + 346 + ], + "lines": [ + { + "bbox": [ + 474, + 235, + 552, + 346 + ], + "spans": [ + { + "bbox": [ + 474, + 235, + 552, + 346 + ], + "type": "image", + "image_path": "7b36b0d9b85f3ccc80032ba5388a5177d38152dcefb2d893600aac85cdda397a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 422, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 495 + ], + "type": "text", + "content": "comparisons (Fig. 5) show that smaller windows enhance texture details in the predicted illumination maps but with lower accuracy. Conversely, larger windows (e.g., [192, 128]) reduce error metrics by capturing more information, yet they may blur boundaries in multi-illuminant regions, compromising edge clarity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 505, + 388, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 388, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 388, + 517 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 525, + 555, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 680 + ], + "type": "text", + "content": "We propose IFFCC, a practical solution for achieving accurate, real-time auto white balance (AWB) in multi-illuminant environments. Built upon the Fast Fourier Color Constancy (FFCC) algorithm, IFFCC leverages an integral UV histogram and parallelized Fourier-based convolution to rapidly estimate illumination across different image regions, addressing the need for localized, adaptive AWB. Results on multiple datasets demonstrate that IFFCC delivers accuracy comparable to or surpassing pixel-level neural network methods, while reducing the parameter count by over 400 times and speeding up processing by a factor of 20–100. These advantages make IFFCC highly suitable for resource-limited, real-time applications such as video streaming." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 679, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 679, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 679, + 555, + 714 + ], + "type": "text", + "content": "We believe that FFCC-based methods, including IFFCC, will continue to enrich the field and attract increased attention across a broader range of ISP applications." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26427" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 139 + ], + "type": "text", + "content": "Acknowledgments This work was supported in part by the Anhui Provincial Natural Science Foundation under Grant 2408085J024, the National Natural Science Foundation of China under Grant 62401532, and the Anhui Provincial Key Research and Development Plan under Grant 202304a05020072." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 182, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 182, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 182, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 294, + 226 + ], + "type": "text", + "content": "[1] Kobus Barnard, Vlad Cardei, and Brian Funt. A comparison of computational color constancy algorithms. i: Methodology and experiments with synthesized data. IEEE transactions on Image Processing, 11(9):972-984, 2002. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 228, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 295, + 259 + ], + "type": "text", + "content": "[2] Jonathan T Barron. Convolutional color constancy. In Proceedings of the IEEE International Conference on Computer Vision, pages 379-387, 2015. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 262, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 295, + 304 + ], + "type": "text", + "content": "[3] Jonathan T Barron and Yun-Ta Tsai. Fast fourier color constancy. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 886-894, 2017. 1, 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "text", + "content": "[4] Shida Beigpour, Christian Riess, Joost Van De Weijer, and Eli Angelopoulou. Multi-illuminant estimation with conditional random fields. IEEE Transactions on Image Processing, 23(1):83–96, 2013. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 353, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 385 + ], + "type": "text", + "content": "[5] Simone Bianco and Raimondo Schettini. Adaptive color constancy using faces. IEEE transactions on pattern analysis and machine intelligence, 36(8):1505-1518, 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 387, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 294, + 430 + ], + "type": "text", + "content": "[6] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Color constancy using cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 81-89, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 432, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 294, + 475 + ], + "type": "text", + "content": "[7] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Single and multiple illuminant estimation using convolutional neural networks. IEEE Transactions on Image Processing, 26(9):4347-4362, 2017. 2, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 477, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 294, + 542 + ], + "type": "text", + "content": "[8] Michael Bleier, Christian Riess, Shida Beigpour, Eva Eibenberger, Eli Angelopoulou, Tobias Tröger, and André Kaup. Color constancy and non-uniform illumination: Can existing algorithms work? In 2011 IEEE international conference on computer vision workshops (ICCV Workshops), pages 774-781. IEEE, 2011. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 544, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 294, + 576 + ], + "type": "text", + "content": "[9] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 579, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 294, + 610 + ], + "type": "text", + "content": "[10] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "type": "text", + "content": "[11] Dongliang Cheng, Dilip K Prasad, and Michael S Brown. Illuminant estimation for color constancy: why spatial-domain methods work and the role of the color distribution. JOSA A, 31(5):1049-1058, 2014. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[12] Ilija Domislovic, Donik Vrsnak, Marko Subasic, and Sven Loncarić. Outdoor daytime multi-illuminant color constancy. In 2021 12th International Symposium on Image and Signal Processing and Analysis (ISPA), pages 270–275. IEEE, 2021. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Iija Domislovic, Donik Vršnak, Marko Subašić, and Sven Lončarić. Shadows & lumination: Two-illuminant multiple cameras color constancy dataset. Expert systems with applications, 224:120045, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "text", + "content": "[14] Ilija Domislovic, Donik Vrsnjak, Marko Subasic, and Sven Loncarić. Color constancy for non-uniform illumination estimation with variable number of illuminants. Neural Computing and Applications, 35(20):14825-14835, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 185 + ], + "type": "text", + "content": "[15] Graham D Finlayson and Steven D Hordley. Color constancy at a pixel. JOSA A, 18(2):253-264, 2001. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 228 + ], + "type": "text", + "content": "[16] Graham D Finlayson and Elisabetta Trezzi. Shades of gray and colour constancy. In Color and Imaging Conference, pages 37-41. Society of Imaging Science and Technology, 2004. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "text", + "content": "[17] Graham D. Finlayson, Steven D. Hordley, and Paul M. Hubel. Color by correlation: A simple, unifying framework for color constancy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(11):1209-1221, 2001. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 276, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 307 + ], + "type": "text", + "content": "[18] David A Forsyth. A novel algorithm for color constancy. International Journal of Computer Vision, 5(1):5-35, 1990. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 310, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 330 + ], + "type": "text", + "content": "[19] David H Foster. Color constancy. Vision research, 51(7): 674-700, 2011. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 333, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 553, + 396 + ], + "type": "text", + "content": "[20] Shaobing Gao, Wangwang Han, Kaifu Yang, Chaoyi Li, and Yongjie Li. Efficient color constancy with local surface reflectance statistics. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 158-173. Springer, 2014. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 399, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 442 + ], + "type": "text", + "content": "[21] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Generalized gamut mapping using image derivative structures for color constancy. International Journal of Computer Vision, 86:127-139, 2010. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 445, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 485 + ], + "type": "text", + "content": "[22] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Computational color constancy: Survey and experiments. IEEE transactions on image processing, 20(9):2475-2489, 2011. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 489, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 532 + ], + "type": "text", + "content": "[23] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):918-929, 2011. 1, 2, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 534, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 566 + ], + "type": "text", + "content": "[24] Arjan Gijsenij, Rui Lu, and Theo Gevers. Color constancy for multiple light sources. IEEE Transactions on image processing, 21(2):697-707, 2011. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 568, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 600 + ], + "type": "text", + "content": "[25] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE transactions on pattern analysis and machine intelligence, 35(6):1397-1409, 2012. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 602, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 643 + ], + "type": "text", + "content": "[26] Eugene Hsu, Tom Mertens, Sylvain Paris, Shai Avidan, and Frédo Durand. Light mixture estimation for spatially varying white balance. In ACM SIGGRAPH 2008 papers, pages 1-7. 2008. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[27] Yuanming Hu, Baoyuan Wang, and Stephen Lin. Fc4: Fully convolutional color constancy with confidence-weighted pooling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4085-4094, 2017. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "text", + "content": "[28] Zhuo Hui, Aswin C Sankaranarayanan, Kalyan Sunkavalli, and Sunil Hadap. White balance under mixed illumination" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26428" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 105 + ], + "type": "text", + "content": "using flash photography. In 2016 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 138 + ], + "type": "text", + "content": "[29] Md Akmol Hussain and Akbar Sheikh Akbari. Color constancy algorithm for mixed-illuminant scene images. IEEE Access, 6:8964–8976, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 193 + ], + "type": "text", + "content": "[30] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 270 + ], + "type": "text", + "content": "[31] Dongyoung Kim, Jinwoo Kim, Seonghyeon Nam, Dongwoo Lee, Yeonkyung Lee, Nahiup Kang, Hyong-Euk Lee, ByungIn Yoo, Jae-Joon Han, and Seon Joo Kim. Large scale multi-illuminant (lsmi) dataset for developing white balance algorithm under mixed illumination. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2410–2419, 2021. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 271, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 294, + 326 + ], + "type": "text", + "content": "[32] Dongyoung Kim, Jinwoo Kim, Junsang Yu, and Seon Joo Kim. Attentive illumination decomposition model for multi-illuminant white balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25512-25521, 2024. 3, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "text", + "content": "[33] Edwin H Land. The retinax theory of color vision. In Proceedings of the Royal Institution of Great Britain, page 23, 1974. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "text", + "content": "[34] Shuwei Li, Jikai Wang, Michael S Brown, and Robby T Tan. Transc: Transformer-based multiple illuminant color constancy using multitask learning. arXiv preprint arXiv:2211.08772, 1(2):3, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "type": "text", + "content": "[35] Yongjie Liu and Sijie Shen. Self-adaptive single and multi-illuminant estimation framework based on deep learning. arXiv preprint arXiv:1902.04705, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[36] Yi-Chen Lo, Chia-Che Chang, Hsuan-Chao Chiu, Yu-Hao Huang, Chia-Ping Chen, Yu-Lin Chang, and Kevin Jou. Clcc: Contrastive learning for color constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8053-8063, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 492, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 547 + ], + "type": "text", + "content": "[37] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33:11525-11538, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 580 + ], + "type": "text", + "content": "[38] Kantilal Varichand Mardia. Statistics of directional data. Journal of the Royal Statistical Society Series B: Statistical Methodology, 37(3):349-371, 1975. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "text", + "content": "[39] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A dataset of multi-illumination images in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4080–4089, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "text", + "content": "[40] Fatih Porikli. Integral histogram: A fast way to extract histograms in cartesian spaces. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 829-836. IEEE, 2005. 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[41] Yanlin Qian, Ke Chen, Jarno Nikkanen, Joni-Kristian Kamarainen, and Jiri Matas. Recurrent color constancy. In Proceedings of the IEEE international conference on computer vision, pages 5458-5466, 2017. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 363 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "type": "text", + "content": "[42] Yanlin Qian, Joni-Kristian Kamarainen, Jarno Nikkanen, and Jiri Matas. On finding gray pixels. In CVPR, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 315, + 96, + 553, + 128 + ], + "type": "text", + "content": "[43] Yanlin Qian, Jani Käpylä, Joni-Kristian Kämäräinen, Samu Koskinen, and Jiri Matas. A benchmark for burst color constancy. In ECCVW, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 553, + 184 + ], + "type": "text", + "content": "[44] Wu Shi, Chen Change Loy, and Xiaou Tang. Deep specialized network for illuminant estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 371-387. Springer, 2016. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[45] Oleksii Sidorov. Conditional gans for multi-illuminant color constancy: Revolution or yet another approach? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[46] Oleksii Sidorov. Conditional gans for multi-illuminant color constancy: Revolution or yet another approach? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 275, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 308 + ], + "type": "text", + "content": "[47] Joost Van De Weijer, Theo Gevers, and Arjan Gijsenij. Edge-based color constancy. IEEE Transactions on image processing, 16(9):2207-2214, 2007. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "text", + "content": "[48] Bolei Xu, Jingxin Liu, Xianxu Hou, Bozhi Liu, and Guoping Qiu. End-to-end illuminant estimation based on deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3616-3625, 2020. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26429" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/4511c287-17f9-4b70-a6d5-d77c48a5192a_content_list.json b/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/4511c287-17f9-4b70-a6d5-d77c48a5192a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2d010ff018d887067db895e7e3a5c64e18297262 --- /dev/null +++ b/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/4511c287-17f9-4b70-a6d5-d77c48a5192a_content_list.json @@ -0,0 +1,1626 @@ +[ + { + "type": "text", + "text": "InterAct: Advancing Large-Scale Versatile 3D Human-Object Interaction Generation", + "text_level": 1, + "bbox": [ + 127, + 128, + 870, + 174 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee9ab7a9c8f7ee5ca8269e301a95769d9c61b2836b596fb89c0929f95e464738.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 200, + 890, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ebbdf565494c4000173240ed91cfb3804bece795c306c69fdf547df490d31b7.jpg", + "image_caption": [ + "Figure 1. An overview of InterAct, our large-scale 3D human-object interaction (HOI) benchmark, covering six HOI generation tasks." + ], + "image_footnote": [], + "bbox": [ + 91, + 305, + 908, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 598, + 326, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While large-scale human motion capture datasets have advanced human motion generation, modeling and generating dynamic 3D human-object interactions (HOIs) remain challenging due to dataset limitations. Existing datasets often lack extensive, high-quality motion and annotation and exhibit artifacts such as contact penetration, floating, and incorrect hand motions. To address these issues, we introduce InterAct, a large-scale 3D HOI benchmark featuring dataset and methodological advancements. First, we consolidate and standardize 21.81 hours of HOI data from diverse sources, enriching it with detailed textual annotations. Second, we propose a unified optimization framework to enhance data quality by reducing artifacts and correcting hand motions. Leveraging the principle of contact invariance, we maintain human-object relationships while introducing motion variations, expanding the dataset to 30.70 hours. Third, we define six benchmarking tasks and develop a unified HOI generative modeling perspective, achieving state-of-the-art", + "bbox": [ + 89, + 628, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "performance. Extensive experiments validate the utility of our dataset as a foundational resource for advancing 3D human-object interaction generation. The dataset will be publicly accessible to support further research in the field.", + "bbox": [ + 511, + 599, + 908, + 661 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 691, + 645, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in human motion modeling have significantly benefited from extensive motion capture (MoCap) datasets [22, 41, 50, 67, 69], enabling the creation of scalable generative models for diverse human movements. Building upon this foundation, researchers are increasingly turning to the more intricate challenge of generating human-object interactions (HOIs) [102-104]. This emerging area holds considerable promise for applications in robotics, animation, and computer vision.", + "bbox": [ + 511, + 719, + 908, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, high-quality HOI generation faces notable obstacles due to factors such as increased degrees of freedom introduced by objects, varied object geometries, dynamic in", + "bbox": [ + 511, + 856, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7048", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/94e1c5bbfb41468236a164856099830bc7868c53c740f66c203fc26dedee41f2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetClipHourTextHandObject
GRAB [77]1,3353.76X✓51
BEHAVE [4]2994.13XX18
InterCap [28]2330.62X✓10
Chairs [29]1,0412.37X✓92
HODome [115]1762.82X✓21
OMOMO [38]4,8388.274,838X15
IMHD [123]1640.97X✓10
InterAct (Ours)11,35021.8134,050✓217
InterAct-X (Ours)16,20130.7048,630✓217
", + "bbox": [ + 94, + 88, + 480, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison between InterAct, InterAct-X, and human-object interaction datasets we collect. Beyond a substantially larger scale, our dataset introduces comprehensive textual annotations and enhances interaction quality, offering a more versatile foundation for large-scale HOI generation.", + "bbox": [ + 89, + 253, + 483, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "teractions, and the necessity for physically accurate contact modeling. Current methods often struggle to achieve realism primarily because existing datasets lack scalability and comprehensive annotations, which are crucial for models to effectively understand interaction dynamics and link them to related domains such as natural language.", + "bbox": [ + 89, + 352, + 483, + 443 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, these challenges underscore the need for comprehensive, high-quality HOI datasets: (1) Limited and Inconsistent Datasets: Existing methods typically depend on small datasets with limited hours of data, difficult to consolidate due to inconsistent human representations, object types, coordinate systems, and annotations. Available annotations [38, 63] are frequently coarse and incomplete, lacking detailed descriptions of human states, object interactions, and involved body parts. (2) Prevalent Artifacts: Current datasets often contain artifacts from MoCap limitations and occlusions, including unnatural penetrations, floating contacts, inaccurate hand poses [4, 38], and significant motion jitter [28]. These issues compromise the models' capacity to learn realistic human-object dynamics.", + "bbox": [ + 88, + 443, + 483, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we present InterAct, a benchmark designed to systematically overcome current limitations and drive advancements in 3D HOI modeling. As shown in Table 1, InterAct offers a large-scale, standardized dataset of carefully curated interactions from existing resources1, enriched by detailed textual annotations.", + "bbox": [ + 89, + 655, + 483, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further enhance dataset quality and scope, we introduce a unified optimization approach, addressing major penetration and floating artifacts first in whole-body interactions, followed by refined corrections for nuanced hand-object interactions. Additionally, we propose the concept of contact invariance, inspired by motion mirroring techniques, to generate realistic synthetic data by varying human motions while", + "bbox": [ + 89, + 746, + 483, + 852 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2dcd34ca991cba8ca1dbdeb3209270981a842cf323ec4e95ba21022e4f047a6f.jpg", + "image_caption": [ + "Figure 2. (a) Our data processing pipeline consolidating data, annotations via foundation models, corrections, and interaction illustrations. (b) Statistics on motion and text annotations." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "maintaining consistent object contacts. This augmentation expands InterAct into InterAct-X, providing approximately 9 additional hours of data and substantially improving generative model performance.", + "bbox": [ + 511, + 297, + 906, + 358 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging this comprehensive, richly annotated dataset, we define benchmarks across six key HOI generation tasks, Text-to-Interaction, Action-to-Interaction, Object-to-Human, Human-to-Object, Interaction Prediction, and Interaction Imitation, as shown in Figure 1, and propose a unified modeling and representation for kinematic generative tasks. Our method utilizes multi-task learning to jointly model motion and contact, achieving state-of-the-art performance as validated by comprehensive evaluations.", + "bbox": [ + 511, + 359, + 908, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are: (1) InterAct, the most extensive 3D HOI benchmark to date, facilitating large-scale generative modeling. (2) A unified optimization-based framework to correct and augment MoCap data, addressing common artifacts and significantly enhancing dataset quality. We believe this will aid future research in overcoming data scarcity before capturing potentially imperfect data. (3) Comprehensive benchmarks across six HOI generation tasks, establishing standardized metrics and demonstrating superior performance over existing approaches. This benchmark lays a strong foundation for future research, encouraging advancements across multiple facets of 3D HOI generation.", + "bbox": [ + 511, + 496, + 908, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 691, + 653, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic 3D HOI Dataset. Many large-scale datasets with sequential human motion data have established benchmarks for the task of 3D human motion generation. However, human actions are influenced not only by individual intents but also by interactions with the surrounding environment. To address this complexity, new datasets have been developed to capture the dynamics between humans and their environments, including interactions with other humans [1, 40, 54, 101] and scenes [8, 24, 25]. Though there are fruitful hand-object interaction datasets [11, 23, 44, 48, 56, 58, 80, 91, 91, 112, 113, 121], our focus is specifically on whole-body interactions with dynamic objects, ranging from", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1We integrate publicly available datasets [4, 28, 29, 38, 77, 115, 123] focusing on single-human interactions with rigid and dynamic objects. Certain datasets were selectively included based on relevance and data type.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7049", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "low-dynamic interactions such as approaching and manipulation [77] to highly dynamic interactions involving multiple body parts [4, 20, 28-30, 32, 38, 107, 115, 123].", + "bbox": [ + 89, + 90, + 485, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to address the limitations of these datasets and open new possibilities for future research. Our InterAct dataset maintains advantages in motion quality, fine-grained textual annotations, detailed hand gestures, and comprehensive annotation modalities. We provide quantitative comparisons of InterAct and existing datasets in Table 1, demonstrating the superiority of our dataset in these aspects.", + "bbox": [ + 89, + 136, + 485, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic 3D HOI Generation. Existing human-object interaction (HOI) datasets have laid a robust foundation for generating dynamic, whole-body interactions. Extensive research has explored the generation of hand-object interactions [9, 13, 39, 43, 49, 84, 111, 114, 118, 126, 127] and static human-object interactions [27, 33, 64, 87, 96, 97, 106, 108, 119, 124]. Meanwhile, full-body dynamic interactions have also been studied extensively [14, 21, 34-36, 38, 52, 53, 71, 75, 76, 78, 85, 94, 105, 120, 125], though these often face significant limitations, including narrow action repertoires and dependence on static objects. Recent advancements, such as InterDiff [102], have introduced diverse interactions involving dynamic objects and multiple body parts. Building upon this, subsequent approaches like InterDreamer [103] and other contemporary studies [18, 37, 63, 74, 93, 95, 103, 117] further demonstrate the feasibility of converting textual descriptions into realistic 3D human-object interaction sequences. Despite these advances, current methods remain constrained by a shortage of high-quality, large-scale datasets, often encountering issues related to physical inaccuracies, such as floating contacts or interpenetration. In parallel, physics-based methods leveraging deep reinforcement learning (RL) [3, 6, 10, 15, 26, 42, 55, 61, 81, 83, 86, 89, 90, 92, 98, 100, 109, 110] successfully generate physically accurate interactions, with applications in sports [47] like basketball [88, 89] and soccer [99]. Nonetheless, these methods typically produce rigid interaction patterns from limited datasets, while a recent work, InterMimic [104], illustrates that physics-based approaches can digest effectively across diverse and dynamic object interactions. Our work addresses these fundamental limitations at the dataset level by providing enhanced diversity and comprehensive sequences of human-object interactions. This facilitates multiple generative tasks, supports better contact modeling, and improves the capability to synthesize realistic and generalized human-object interactions.", + "bbox": [ + 89, + 241, + 486, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. InterAct Dataset", + "text_level": 1, + "bbox": [ + 89, + 799, + 256, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. We introduce InterAct, the first unified benchmark tailored explicitly for sequential 3D human-object interaction (HOI) generative modeling. Distinguished by its unprecedented scale and comprehensiveness, InterAct significantly surpasses existing datasets, as summarized in Table 1.", + "bbox": [ + 89, + 824, + 485, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3402e9787fb5232f3e63ae47ab08ab60b7b23f220afe834df2b23d3b4d539b96.jpg", + "image_caption": [ + "(a) Markers in SMPL-H", + "Figure 3. Marker-based representation for human." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 700, + 162 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4b0e57fb1ec483f4a17a14b96fa900f73192d6e0f30a08ebe44b45703a1b009c.jpg", + "image_caption": [ + "(b) Markers in SMPL-X" + ], + "image_footnote": [], + "bbox": [ + 712, + 88, + 903, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "InterAct is available in two versions: (1) A basic version consolidating seven existing datasets, providing 21.81 hours of annotated 3D whole-body interactions with corresponding semantic descriptions. (2) An advanced version, InterAct-X, extending the basic dataset through synthetic data generated via our unified optimization framework. Figure 2 presents examples of motion and text annotations. To ensure high-quality standards, we employ a multifaceted annotation strategy combining human expertise, automated foundation models, and advanced HOI modeling techniques, all validated through rigorous manual quality checks.", + "bbox": [ + 511, + 228, + 908, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Collection, Annotation, and Unification", + "text_level": 1, + "bbox": [ + 511, + 407, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We compile data from seven datasets [4, 28, 29, 38, 77, 115, 123], featuring motion capture of a single human interacting with a single dynamic 3D object, where humans are annotated with SMPL [45, 62, 73]. We address their heterogeneity in two key aspects: annotations and representations.", + "bbox": [ + 511, + 431, + 906, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unifying Textual Annotations. Since most datasets either lack textual descriptions or provide only very coarse text descriptions [38], we implement a two-phase annotation procedure involving human annotators and GPT-4 [60] to generate consistent and detailed annotations across all subsets. In the first phase, human annotators provided detailed and precise descriptions of the interactions, adhering to the following guidelines: (i) Split motion sequences into clips averaging 300 frames (approximately 10 seconds) but no longer than 400 frames each; (ii) Clearly describe the actions and the body parts involved in the interactions. For example, a typical annotation is: \"A person sits on a stool and touches the ground with their left hand, then their right hand.\" For the subset derived from OMOMO [38], we skip this phase and directly utilize their annotations. In the second phase, we use GPT-4 to rephrase and simplify human annotations to enhance diversity and consistency. For example, the rephrased version is \"A person perches on a stool, touching the ground with their left hand, then their right hand,\" and the simplified version is \"A person sits on a stool, touching the ground with each hand alternately.\" Next, we employ GPT-4 to classify each description into one of our predefined 15 action labels with in-context learning [7]. The action label for the above sequence is \"Sit.\" We meticulously review all generated texts and action labels to ensure high quality and alignment across the dataset.", + "bbox": [ + 511, + 508, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7050", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unifying Human Representations. Different datasets employ varying human models (e.g., SMPL-H [73], SMPL-X [62]) and diverse shapes. A straightforward solution can be to convert different humans from SMPL-H and SMPL-X to a consistent SMPL version and encode shape parameters into the generative modeling. However, although SMPL is widely used in various human-related tasks, it is fundamentally a rotation-based representation. In the context of human motion generation, Cartesian features like joint positions and velocities are more commonly used, as seen in the integration with the HumanML3D representation [22] and in most text-to-motion work [65, 82, 116]. This is still suboptimal because joints are located beneath the body's surface and do not explicitly participate in interactions. To overcome the limitations, we use markers - specific sets of human vertices representing human motion and interactions - as a simple and unified representation capable of effectively inferring contact, evaluated in Table 5. Similar approaches are discussed in [94, 102]. Then we need to select a marker set that is consistent between SMPL-H and SMPL-X models.", + "bbox": [ + 89, + 90, + 485, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given two human body models, SMPL-H and SMPL-X, sharing the same shape, we establish marker correspondences in two steps. First, we index the markers on the SMPL-H surface as defined in prior work [94, 102]. Second, we locate the corresponding vertices on SMPL-X by selecting the closest points to these SMPL-H markers, leveraging the official SMPL conversion, which maps each SMPL-H vertex to the nearest point on the SMPL-X mesh. We extensively evaluated the approximation error of these marker correspondences across a broad range of poses. Our results show that the maximum error consistently remains below $1\\mathrm{cm}$ , a deviation unlikely to affect the overall performance of HOI generation. This high consistency arises because the markers are rigidly attached to the body, and soft deformations are disabled. As a result, the identical rigid transformations in SMPL-H and SMPL-X preserve the correspondence of the markers. Additional details on such correspondence-preserving conditions can be found in [31]. Figure 3 illustrates the marker sets for SMPL-H and SMPL-X. We use this marker-based representation to train the generative models for the tasks outlined in Sec. 4, while still relying on the original SMPL-H or SMPL-X representations for the interaction correction and augmentation methods described in Sec. 3.2.", + "bbox": [ + 91, + 395, + 486, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Interaction Correction and Augmentation", + "text_level": 1, + "bbox": [ + 89, + 770, + 447, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present a unified optimization framework that addresses both the correction of MoCap artifacts and the augmentation of the dataset by introducing more synthetic data. The process takes as input the motions and geometries of humans and objects, then compares them against predefined standards to define loss functions. Using gradient-based optimization, we iteratively adjust human and object", + "bbox": [ + 89, + 795, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "motions to minimize these losses, thereby refining the data to meet the desired quality criteria. The key challenge lies in formulating learning objectives that not only rectify existing data but also facilitate the generation of new synthetic data.", + "bbox": [ + 511, + 90, + 906, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our optimization is carried out in three sequential steps: (1) full-body correction; (2) hand correction; and (3) interaction augmentation. Hand correction is handled separately because, although hand poses occupy considerable space in the SMPL representation, they contribute relatively little to the overall scale of the learning objectives. By decoupling hand correction from full-body correction, we can better balance these two processes and define more targeted objectives. In what follows, we first introduce the hand correction stage. Hand Correction. Given that many existing datasets contain inaccurate hand poses [4, 38], our approach selectively promotes contact only in regions where ground-truth data indicates hand-object interaction, while ensuring the hand motion remains natural, in spirit to InterMimic [104] but relying on predefined optimization instead of RL. This approach is effective for the whole-body interaction datasets we utilize, which generally do not require high dexterity and typically only involve the hand conforming to the object for grasping, as a common assumption in existing work [79, 127], while we distinguish our approach from those that employ multi-stage, learning-based methods for the same purpose.", + "bbox": [ + 511, + 152, + 908, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We divide our hand correction objectives into two categories: contact promotion and hand constraints. Contact promotion is guided by the following contact loss:", + "bbox": [ + 511, + 469, + 908, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathrm {c o n t}} = \\sum_ {i = 1} ^ {L} c _ {i} \\sum_ {j} d _ {j} [ i ],\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 518, + 794, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{j}[i]$ is the distance between the $j$ -th hand vertex and its nearest point on the object's surface at the $i$ -th frame, and $c_{i}$ indicates whether the object and the hand are in contact at frame $i$ . The contact indicator $c_{i}$ , inferred from ground truth data, is a function based on hand-object distance $\\min_j d_j[i]$ , which we provide details in supplementary. The hand constraint objectives are introduced to preserve naturalness and temporal smoothness in the hand motions. These constraints include: (1) penetration loss, which penalizes intersections between the hand and the object. (2) smoothness loss, which promotes consistent contact and reduces jittering. (3) prior loss, which constrains the range of motion (RoM) of the fingers to maintain realism. Without this constraint, contact promotion could inadvertently drive fingers into biologically impossible poses. Detailed formulations of these loss functions are provided in the supplementary.", + "bbox": [ + 511, + 568, + 908, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Full-Body Correction. In this stage, all human and object poses can be updated via gradient descent. We add a reconstruction loss to ensure that the optimized interactions closely match the ground truth. Other losses mirror those used in hand correction, with two key differences: (1) Contact and penetration losses are computed for the entire body", + "bbox": [ + 511, + 810, + 910, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7051", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rather than just the hands. (2) Prior loss is omitted because the reconstruction loss alone suffices to maintain plausible human motion. Detailed formulations of these losses are provided in supplementary.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Interaction Augmentation. Synthetic data has become increasingly important in computer vision and generative modeling [5, 12, 19, 57], prompting a key question in the context of HOI animation: Can we scale up datasets without collecting additional MoCap data? Does existing interaction data offer information beyond its observable motions? Consider a scenario where a person grasps a box and walks, as illustrated in Figure 5. Even if their gait changes slightly, hand-box contact should remain consistent to preserve the semantics of the interaction. This illustrates the principle of interaction invariance: the core interaction persists despite minor variations in motion. Leveraging this principle, we can augment our dataset by injecting new human motions while preserving consistent object interactions. Training neural networks on such augmented data enables them to naturally learn this invariance, a common strategy in symmetric learning [17], and ultimately enhances model performance.", + "bbox": [ + 89, + 152, + 483, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our augmentation pipeline consists of three steps: (1) Object Displacement: We apply a random displacement to the object's trajectory, uniformly across all timesteps. (2) Interaction Alignment: We optimize the human motion to maintain interaction with the displaced object, using both contact consistency and reconstruction objectives. (3) Interaction Filtering: We remove low-quality augmentations, those with unreasonable initial displacements, significant penetrations (human-object or self-penetration), or alignment failures indicated by high optimization losses (e.g., excessive jitter).", + "bbox": [ + 89, + 407, + 483, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the alignment phase, the primary objective is the contact consistency loss. We first compute a distance matrix $\\mathbf{D}$ , where each element $\\mathbf{D}_{jk} = \\|\\mathbf{v}_h^j - \\mathbf{v}_o^k\\|$ denotes the Euclidean distance between the $j$ -th human vertex and the $k$ -th object vertex, before or after displacement. With a reference matrix $\\hat{\\mathbf{D}}$ from the original (pre-displacement) setup, we optimize the human motion using:", + "bbox": [ + 89, + 574, + 483, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathrm {a l i g n}} = \\sum_ {i = 1} ^ {L} \\sum_ {j, k} \\frac {1}{(\\hat {\\mathbf {D}} _ {j k} + \\epsilon) ^ {2}} \\left| \\hat {\\mathbf {D}} _ {j k} - \\mathbf {D} _ {j k} \\right| ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 689, + 437, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a small constant to prevent division by zero. This formulation preserves distances between vertex pairs that were initially close, while de-emphasizing pairs that were farther apart. Additional terms in the objective enforce naturalness and stability in non-interactive regions of the human pose, as detailed in the supplementary material.", + "bbox": [ + 89, + 741, + 483, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Tasks and Methods", + "text_level": 1, + "bbox": [ + 89, + 844, + 277, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we formally define six distinct tasks featured in our benchmark. We use unified representation across", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "five kinematic generative tasks, where each human-object interaction sequence is represented as $\\langle h,o\\rangle$ , annotated with an action category $a$ and a text description $t$ . The human $h$ includes marker coordinates, marker velocities, signed distance vectors from each marker to the object, and foot-ground contact labels. The object $o$ represents object motion, including object rotation angles, object translations. Object geometry is described by Basis Point Set (BPS) [68].", + "bbox": [ + 511, + 90, + 906, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(1) Text-Conditioned Interaction Generation. Initially in [18, 63], the task learns a function to generate the interaction sequence based on text: $\\mathcal{G}_{\\mathrm{t2i}}(t)\\mapsto \\langle h,o\\rangle$", + "bbox": [ + 511, + 212, + 906, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(2) Action-Conditioned Interaction Generation. The objective is to learn a function that maps an action label to the corresponding interaction sequence: $\\mathcal{G}_{\\mathrm{a2i}}(\\pmb {a})\\mapsto \\langle \\pmb {h},\\pmb {o}\\rangle$", + "bbox": [ + 511, + 257, + 905, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(3) Object-Conditioned Human Generation. Initially in [38], the task generates human motion based on object sequences through a function $\\mathcal{G}_{\\mathrm{o2h}}(\\pmb {o})\\mapsto \\pmb{h}$ .", + "bbox": [ + 511, + 303, + 905, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(4) Human-Conditioned Object Generation. Conversely, this task focuses on generating object motion sequences from human motion sequences via a function $\\mathcal{G}_{\\mathrm{h2o}}(\\pmb {h})\\mapsto \\pmb{o}$ .", + "bbox": [ + 511, + 347, + 905, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(5) Interaction Prediction. Initially in [102], the task aims to predict future human-object interactions based on past. Let $\\langle \\pmb{h}_p, \\pmb{o}_p \\rangle$ denote the past interaction and $\\langle \\pmb{h}_f, \\pmb{o}_f \\rangle$ the future. The goal is to learn: $\\mathcal{G}_{\\mathrm{p2f}}(\\langle \\pmb{h}_p, \\pmb{o}_p \\rangle) \\mapsto \\langle \\pmb{h}_f, \\pmb{o}_f \\rangle$ .", + "bbox": [ + 511, + 393, + 905, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(6) Interaction Imitation. Following [3, 88, 122], this task focuses on learning physics-based control policies to reproduce human-object interactions in a physics simulator. The output is an action sequence $\\pmb{f}$ , specified as joint Proportional-Derivative (PD) targets. The goal is to learn a function $\\mathcal{G}_{\\mathrm{i2f}}$ that maps the reference interaction sequences to the PD actuation sequences: $\\mathcal{G}_{\\mathrm{i2f}}(\\langle h,o\\rangle)\\mapsto f$ .", + "bbox": [ + 511, + 454, + 906, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unifying Multi-Task HOI Generation. We introduce an additional feature, $\\eta$ , which encodes human-object relationships through vectors extending from each human marker to its nearest point on the object's surface. The specific configuration of $\\eta$ for each generative task is described in Sec. 5. Using this feature, we can unify first five kinematic generative tasks into a multi-task learning framework by treating $\\eta$ as an additional output. For example, we redefine the text-conditioned interaction task as $\\mathcal{G}_{\\mathrm{t2i}}(t) \\mapsto \\langle h, o, \\eta \\rangle$ , where $\\mathcal{G}$ is a transformer-based diffusion model. This formulation compels the model to learn spatial relationships inherent to the interactions. In our experiments, we observe that this simple strategy, enhanced by large-scale data, consistently outperforms existing methods. Similar ideas are explored in [18, 37, 38, 63, 74, 93, 102].", + "bbox": [ + 511, + 559, + 908, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 799, + 643, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin by evaluating the effectiveness of our data correction and augmentation methods. Following this, we benchmark existing work and our proposed method on the tasks using our dataset. We standardize the evaluation metrics and present extensive results, including ablation studies. We in", + "bbox": [ + 511, + 824, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7052", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d89e67ccd06be2a8a616659f932b4ce6f45b3376d1003fdc4854b5639a6d8d6e.jpg", + "image_caption": [ + "Figure 4. Qualitative evaluation of interaction correction (bottom) on the OMOMO [38] dataset shows hand recovery compared to the ground truth interaction (top). Zoom in to see details of the hand recovery." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 262 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/906cd0675528ace98b05fd0e024ade7ad10acdb400441dc699a9e4419ff4222f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetCorrectionAugmentationPene (m)↓Cont RatioUser Study (%)
BEHAVE [4]××0.0170.04822.3
✓×0.0160.07139.7
✓✓0.0160.06938.0
OMOMO [38]××0.0090.07123.9
✓×0.0070.13139.4
✓✓0.0110.13736.7
", + "bbox": [ + 93, + 321, + 483, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative evaluation and user study on the quality of data from interaction correction and augmentation.", + "bbox": [ + 89, + 415, + 483, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "clude additional implementation details, such as the train-test split, in supplementary.", + "bbox": [ + 89, + 470, + 483, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Correction and Augmentation", + "text_level": 1, + "bbox": [ + 89, + 513, + 357, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We use the following two metrics: Penetration refers to the intersection depth – maximum of negative sign distances from human vertices to the object's surface – average across the sequence. Contact Ratio represents the average ratio of human vertices where their distances to object are under a threshold.", + "bbox": [ + 89, + 537, + 483, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Evaluations. Table 2 shows that our correction process significantly improves the quality of the original MoCap data by enhancing human-object contact and reducing penetration artifacts. Moreover, the quality of the augmented data is comparable to that of the corrected data and exceeds the quality of the original dataset.", + "bbox": [ + 89, + 628, + 483, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Evaluations. Recognizing that quantitative metrics may not fully capture data quality, we conducted a double-blind user study. We randomly selected sequences from each of the raw, corrected, and augmented data for the subset from BEHAV [4] and OMOMO [38] datasets. Human judges were presented with 30 tuple of interactions and asked to rank the quality of three sequences. According to Table 2, over $39\\%$ of judges select the corrected data as having the highest quality, significantly outperforming the original data. This confirms that our correction process effectively enhances data realism. Moreover, the augmented data receive ratings comparable to the corrected data, indicating", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that our synthetic data is of high quality. In Figure 4, we visualize our correction results. Despite the original data lacking detailed hand information, we successfully recover vivid and accurate hand interactions. Figure 5 showcases our augmentation, which introduces new high-quality synthetic data while maintaining consistent contact in interactions.", + "bbox": [ + 511, + 325, + 906, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Language Conditioned HOI Generation", + "text_level": 1, + "bbox": [ + 511, + 424, + 856, + 440 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Following the literature on text-to-motion generation [22], we develop five metrics for evaluation. The Fréchet Inception Distance (FID) quantifies the similarity between generated HOI features and the ground truth. The Multimodality and Diversity metrics assess the variety within the generated HOI. R-Precision measures the alignment between the textual descriptions and the generated HOI. The Multimodal Distance (MM Dist) evaluates the disparity between HOI features and corresponding text features. To obtain human-object interaction (HOI) and text features for calculating these metrics, existing methods often train their feature extractors on very limited data [18, 63, 74, 93], which can degrade the quality of the evaluation. We address this limitation by incorporating our larger-scale data with marker and BPS representations. Instead of formulating a classification task to train the feature extractor [22], we follow [46, 66] and employ sequence-level contrastive learning with an InfoNCE loss [59] to train a text encoder and an HOI encoder, integrating Sentence-BERT [72] into the text encoder.", + "bbox": [ + 511, + 446, + 908, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and Implementation Details. We adopt HOI-Diff [63] as our base model because it is the only publicly available option compatible with our requirements. For example, CHOIS [37] requires additional conditions beyond text input. HOI-Diff utilizes a transformer-based diffusion model [82] as the backbone, and integrates an affordable model as the classifier guidance [16]. We develop several baseline variants towards our final method by implementing three key modifications: (i) Text Encoder: We replace HOI-Diff's CLIP-based text encoder, where the latent space is not structured for human-object interactions, with our pre", + "bbox": [ + 511, + 734, + 910, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7053", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29409087b827a1c9874414c9ae315a5a2b83ac7e362ff30c3bae9bd0b77c2c17.jpg", + "image_caption": [ + "Figure 5. Qualitative evaluation of interaction augmentation (bottom) shows high-quality synthetic data varied from original (top)." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 911, + 295 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/75af318df00a5f26ca47b8f500265af0afb652a46cefa6e8dc1e2484e40bbc38.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
HOI-Aware Object Enc.HOI-Aware Text Enc.Contact GenerationContact GuidanceR-Precision†FID↓MM Dist↓Multimodality↑Diversity→
Top 1Top 2Top 3
Ground Truth0.852±0.0000.966±0.0010.989±0.0010.000±0.0002.810±0.002-11.489±0.011
XXXX0.733±0.0070.909±0.0020.957±0.0023.192±0.1914.950±0.0233.149±0.45211.192±0.019
XX✓X0.730±0.0070.913±0.0040.958±0.0051.997±0.0924.752±0.0654.171±0.02711.501±0.037
✓X✓X0.737±0.0110.912±0.0020.963±0.0081.837±0.1264.631±0.0782.836±0.58311.369±0.096
✓✓✓X0.784±0.0040.940±0.0020.980±0.0031.570±0.1394.414±0.0642.677±0.56211.409±0.005
✓✓✓✓0.784±0.0040.940±0.0000.977±0.0021.567±0.1444.412±0.0653.842±0.00511.518±0.178
", + "bbox": [ + 94, + 334, + 906, + 450 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/38576889299891bd193ae8bbee1d339009e76cab230f9099184332f6bc31edab.jpg", + "table_caption": [ + "Table 3. Quantitative evaluation on the task of text-conditioned interaction generation. A batch size of 64 is used for R-Precision." + ], + "table_footnote": [], + "table_body": "
MethodFID↓Multimodality↑Diversity→
Ground Truth0.000±0.000-11.489±0.011
HOI-Diff [63]3.566±0.0985.321±0.14310.989±0.112
Ours2.161±0.0375.792±0.05911.291±0.261
", + "bbox": [ + 132, + 497, + 444, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trained interaction-aware text encoder. (ii) Object Shape Encoding: We substitute the original PointNet++ [70], also not pretrained for capturing interaction, with BPS [68] representations for encoding object shapes. (iii) Instead of using affordance as guidance, we regress the contact representation $\\eta$ through a multi-task learning. (iv) We further incorporate the contact prediction as classifier guidance [16] within the denoising process, which we detail in supplementary.", + "bbox": [ + 89, + 628, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. Table 3 presents the evaluation results for the text-conditioned interaction generation task. We assess the impact of four design choices introduced above. Notably, incorporating contact modeling and BPS encoding significantly improves the quality of generated HOI, substantially enhancing the FID score. Furthermore, using our interaction-aware text and HOI encoder enhances the quality of the generated interactions and improves the alignment between the generated results and the text input. Lastly, classifier guidance provides a slight overall performance improve", + "bbox": [ + 88, + 750, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/57ea761ff5a0bae2b673ee01760d95890209231d2489d983b7a7b565ec63cf24.jpg", + "table_caption": [ + "Table 4. Quantitative evaluation on the task of action-conditioned interaction generation on the entire InterAct testset." + ], + "table_footnote": [], + "table_body": "
RepresentationPene (m)↓Cont Ratio
SMPL0.0300.025
Joint0.0270.032
Marker0.0250.028
", + "bbox": [ + 593, + 497, + 828, + 571 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8f056b0be00c8891467c87955215b6094c562fe5279c5049701af42cd9e82f57.jpg", + "table_caption": [ + "Table 5. Ablation study on different representation for text-to-interaction task, evaluated under BEHAVE [4] subset." + ], + "table_footnote": [], + "table_body": "
ModelMPMPE↓FS↓Cprec↑Crec↑Cacc↑F1 Score↑
2-stage [38]36.500.270.810.850.770.80
1-stage36.940.290.840.820.800.81
Ours (Disc.)36.950.280.850.840.830.82
Ours (Cont.)35.690.280.850.890.850.85
", + "bbox": [ + 516, + 630, + 903, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Quantitative evaluation on object-conditioned human motion generation, with novel objects unseen from training.", + "bbox": [ + 511, + 714, + 906, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ment. Table 4 benchmarks the performance improvements of these design choices similarly for the action-conditioned interaction generation task.", + "bbox": [ + 511, + 776, + 906, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Marker-Based Representation. As shown in Table 5, we compare different human representations on text-to-interaction generation. Without altering contact modeling, marker-based representation produces interactions with fewer artifacts compared to other representations.", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7054", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. HOI Inpainting", + "text_level": 1, + "bbox": [ + 89, + 90, + 248, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metrics. Following OMOMO [38], we develop six metrics tailored for evaluating our marker-based representation. The Mean Per-Marker Position Error (MPMPE) are used to measure the similarity between the generated marker motion and the ground truth. Foot Sliding (FS) is employed to assess the skating effect, reflecting the plausibility of the motion. Additionally, we use a set of contact metrics, including precision $(C_{\\mathrm{prec}})$ , recall $(C_{\\mathrm{rec}})$ , accuracy $(C_{\\mathrm{acc}})$ , and the F1 Score, to evaluate the quality of human-object contact compared to the ground truth. For the human-conditioned object generation task, we include $T_{\\mathrm{err}}$ and $O_{\\mathrm{err}}$ , which measure discrepancies in object translation and orientation between the generated and the ground truth.", + "bbox": [ + 89, + 112, + 483, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines and Implementation Details. We adopt OMOMO [38] as our base model since it is the only publicly available option. OMOMO employs a two-stage generation process: it first generates hand motions and then uses these to guide full-body generations. This strategy is particularly effective for the OMOMO dataset, where interactions primarily involve hand contact. However, it is less effective when applied to our InterAct data, which features more versatile whole-body engagement. Motivated by this, we use a single-stage pipeline with multi-task learning, as introduced in Sec. 4. We investigate two different choices for contact regression: $\\eta_{\\mathrm{Cont.}}$ , which encodes the nearest vector and distance between human markers and the object. $\\eta_{\\mathrm{Disc.}}$ , which encodes the contact labels for each marker.", + "bbox": [ + 89, + 310, + 483, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Results. Tables 6 and 7 illustrate the effectiveness of our single-stage pipeline, with notable performance improvements when incorporating multi-task modeling. These results provide additional evidence, complementing the evaluation of text-conditioned generation, that multitask learning significantly enhances model performance.", + "bbox": [ + 89, + 521, + 483, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Interaction Prediction", + "text_level": 1, + "bbox": [ + 89, + 619, + 297, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Metrics. We compare the generated poses to the ground truth motion data using MPMPE (mean perimeter errors), measured in meters. Second, we assess object motion accuracy using Trans.Err., the average $l_{2}$ distance between the predicted and ground truth object translations, and Rot.Err., the average $l_{1}$ distance between the predicted and ground truth object quaternions, following [102].", + "bbox": [ + 89, + 643, + 482, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. We adapt InterDiff [102] to utilize a marker-based representation and evaluate it at various scales to assess whether it benefits from scaling laws.", + "bbox": [ + 89, + 750, + 483, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Interaction Imitation", + "text_level": 1, + "bbox": [ + 89, + 801, + 289, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. We use PhysHOI [88] to imitate sequences from our InterAct dataset, selecting four sequences shown in Figure 6. The IsaacGym [51] simulator is used, following the same architecture, reward, and representation design as PhysHOI. Training for each sequence,", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e350045dfc5cb0e9e37f275fc81131d1e67cd8a444090bbbe9321483a08eb200.jpg", + "image_caption": [ + "Figure 6. Qualitative results demonstrate the successful imitation of our corrected data using PhysHOI [88]." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 165 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d8c398260c1b48792661244a82eb8c1e0b82e9dd2095dea527da4683cf6c3116.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTerr↓Oerr↓Cprec↑Crec↑Cacc↑F1 Score↑
1-stage25.920.910.830.660.720.69
Ours (multi-task)23.980.830.840.680.740.72
", + "bbox": [ + 517, + 218, + 903, + 265 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8bb755a294568ee58aaf5a142265370f4cae115e26127550191fc8b12e633254.jpg", + "table_caption": [ + "Table 7. Quantitative evaluation on human-to-object." + ], + "table_footnote": [], + "table_body": "
Training dataModel sizeGlobal MPMPE↓Local MPMPE↓Trans. Err.↓\\( \\text{Rot.Err.}^{\\downarrow } \\)
BEHAV×10.1200.1030.1330.352
×20.1050.0920.1090.312
×30.1130.1000.1180.343
InterAct (Ours)×10.1060.0950.1060.297
×20.0940.0830.1030.286
×30.0910.0790.0940.264
", + "bbox": [ + 517, + 305, + 903, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Quantitative evaluation on interaction prediction.", + "bbox": [ + 529, + 409, + 883, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with separate evaluations for both raw and corrected data, is performed on a single NVIDIA A40 GPU over the course of one day.", + "bbox": [ + 511, + 449, + 903, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Evaluation. In addition to Figure 6, which presents the qualitative results, we evaluate the imitation policy by training on both corrected and raw data, reporting the success rate as defined in [88]. Evaluating four examples from Figure 6 and averaging the success rates over 2048 environments, training on the corrected data achieves a success rate of $90.7\\%$ , surpassing the $84.4\\%$ achieved with raw data. This demonstrates that our interaction correction provides better data for the motion imitation task.", + "bbox": [ + 511, + 494, + 905, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Results. Table 8 demonstrates that our dataset, with its larger volume of data, supports the improved performance of the trained model with larger scale. In contrast, training the model on limited data leads to overfitting.", + "bbox": [ + 511, + 631, + 905, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 705, + 633, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce InterAct, a large-scale 3D whole-body human-object interaction benchmark. We employ a unified optimization framework that performs interaction correction and augmentation, enhancing data quality and augment the dataset with synthetic data, scaling to 30.70 hours of interactions and 48,630 textual descriptions. We introduce a simple yet effective multi-task learning approach for unified HOI modeling, enabling models to be trained more effectively across multiple tasks. Our comprehensive experiments highlight the significant advantages of InterAct and our methodologies, resulting in more expressive interaction generation.", + "bbox": [ + 511, + 729, + 906, + 896 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7055", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Amazon-Illinois Center on AI for Interactive Conversational Experiences, the Toyota Research Institute, the IBM-Illinois Discovery Accelerator Institute, and Snap Inc. This work used computational resources, including the NCSA Delta and DeltaAI and the PTI Jetstream2 supercomputers through allocations CIS230012, CIS230013, and CIS240311 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, as well as the TACC Frontera supercomputer, Amazon Web Services (AWS), and OpenAI API through the National Artificial Intelligence Research Resource (NAIRR) Pilot.", + "bbox": [ + 89, + 90, + 485, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 301, + 187, + 318 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] CMU graphics lab motion capture database. http://mocap.cs.cmu.edu/.2", + "[2] Easymocap - make human motion capture easier. Github, 2021. 2", + "[3] Jinseok Bae, Jungdam Won, Donggeun Lim, Cheol-Hui Min, and Young Min Kim. Pmp: Learning to physically interact with environments using part-wise motion priors. In SIGGRAPH, 2023. 3, 5", + "[4] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAV: Dataset and method for tracking human object interactions. In CVPR, 2022. 2, 3, 4, 6, 7, 5", + "[5] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, 2023. 5", + "[6] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. arXiv preprint arXiv:2309.07907, 2023. 3", + "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 3", + "[8] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In ECCV, 2020. 2", + "[9] Junuk Cha, Jihyeon Kim, Jae Shin Yoon, and Seungryul Baek. Text2hoi: Text-guided 3d motion generation for hand-object interaction. In CVPR, 2024. 3", + "[10] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In AAAI, 2021. 3", + "[11] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In CVPR, 2021. 2", + "[12] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing" + ], + "bbox": [ + 99, + 327, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gpt4v-synthesized data for a lite vision-language model. arXiv preprint arXiv:2402.11684, 2024.5", + "[13] Sammy Christen, Shreyas Hampali, Fadime Sener, Edoardo Remelli, Tomas Hodan, Eric Sauser, Shugao Ma, and Bugra Tekin. Diffh2o: Diffusion-based synthesis of handobject interactions from textual descriptions. arXiv preprint arXiv:2403.17827, 2024. 3", + "[14] Enric Corona, Albert Pumarola, Guillem Alenya, and Francesc Moreno-Noguer. Context-aware human motion prediction. In CVPR, 2020. 3", + "[15] Jieming Cui, Tengyu Liu, Nian Liu, Yaodong Yang, Yixin Zhu, and Siyuan Huang. AnySkill: Learning open-vocabulary physical skill for interactive agents. In CVPR, 2024. 3", + "[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 6, 7", + "[17] Sander Dieleman, Jeffrey De Fauw, and Koray Kavukcuoglu. Exploiting cyclic symmetry in convolutional neural networks. In ICML, 2016. 5", + "[18] Christian Diller and Angela Dai. CG-HOI: Contact-guided 3d human-object interaction generation. In CVPR, 2024. 3, 5, 6", + "[19] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, 2024. 5", + "[20] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In CVPR, 2023. 3", + "[21] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. IMoS: Intent-driven full-body motion synthesis for human-object interactions. arXiv preprint arXiv:2212.07555, 2022. 3", + "[22] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 1, 4, 6", + "[23] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In CVPR, 2020. 2", + "[24] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In ICCV, 2019. 2", + "[25] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael Black. Stochastic scene-aware motion prediction. In ICCV, 2021. 2", + "[26] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 3", + "[27] Zhi Hou, Baosheng Yu, and Dacheng Tao. Compositional 3d human-object neural animation. arXiv preprint arXiv:2304.14070, 2023. 3", + "[28] Yinghao Huang, Omid Taheri, Michael J. Black, and Dimitrios Tzionas. InterCap: Joint markerless 3D tracking of humans and objects in interaction. In GCPR, 2022. 2, 3, 5", + "[29] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. CHAIRS: Towards full-body articulated human-object interaction. In ICCV, 2023. 2, 3, 5" + ], + "bbox": [ + 524, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7056", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In CVPR, 2024. 3", + "[31] Marilyn Keller, Keenon Werling, Soyong Shin, Scott Delp, Sergi Pujades, C Karen Liu, and Michael J Black. From skin to skeleton: Towards biomechanically accurate 3d digital humans. ACM Transactions on Graphics (TOG), 42(6):1-12, 2023. 4", + "[32] Jeonghwan Kim, Jisoo Kim, Jeonghyeon Na, and Hanbyul Joo. ParaHome: Parameterizing everyday home activities towards 3d generative modeling of human-object interactions. arXiv preprint arXiv:2401.10232, 2024. 3", + "[33] Taeksoo Kim, Shunsuke Saito, and Hanbyul Joo. NCHO: Unsupervised learning for neural 3d composition of humans and objects. In ICCV, 2023. 3", + "[34] Franziska Krebs, Andre Meixner, Isabel Patzer, and Tamim Asfour. The kit bimanual manipulation dataset. In Humanoids, 2021. 3", + "[35] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. NIFTY: Neural object interaction fields for guided human motion synthesis. arXiv preprint arXiv:2307.07511, 2023.", + "[36] Jiye Lee and Hanbyul Joo. Locomotion-Action-Manipulation: Synthesizing human-scene interactions in complex 3d environments. In ICCV, 2023. 3", + "[37] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913, 2023. 3, 5, 6", + "[38] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2, 3, 4, 5, 6, 7, 8", + "[39] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. arXiv preprint arXiv:2303.13129, 2023. 3", + "[40] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. InterGen: Diffusion-based multi-human motion generation under complex interactions. arXiv preprint arXiv:2304.05684, 2023. 2", + "[41] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-X: A large-scale 3d expressive whole-body human motion dataset. In NeurIPS, 2023. 1", + "[42] Libin Liu and Jessica Hodgins. Learning basketball dribbling skills using trajectory optimization and deep reinforcement learning. ACM Transactions on Graphics (TOG), 37 (4):1-14, 2018. 3", + "[43] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In ICCV, 2023. 3", + "[44] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In CVPR, 2022. 2" + ], + "bbox": [ + 99, + 92, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics, 2015. 3, 5", + "[46] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humanomatopo: Text-aligned whole-body motion generation. arxiv:2310.12978, 2023. 6", + "[47] Zhengyi Luo, Jiashun Wang, Kangni Liu, Haotian Zhang, Chen Tessler, Jingbo Wang, Ye Yuan, Jinkun Cao, Zihui Lin, Fengyi Wang, et al. Smpolympics: Sports environments for physically simulated humanoids. arXiv preprint arXiv:2407.00187, 2024. 3", + "[48] Xintao Lv, Liang Xu, Yichao Yan, Xin Jin, Congsheng Xu, Shuwen Wu, Yifan Liu, Lincheng Li, Mengxiao Bi, Wenjun Zeng, et al. Himo: A new benchmark for full-body human interacting with multiple objects. In ECCV. Springer, 2025. 2", + "[49] Junyi Ma, Jingyi Xu, Xieyuanli Chen, and Hesheng Wang. Diff-ip2d: Diffusion-based hand-object interaction prediction on egocentric videos. arXiv preprint arXiv:2405.04370, 2024. 3", + "[50] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. AMASS: Archive of motion capture as surface shapes. In ICCV, 2019. 1", + "[51] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 8", + "[52] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. The kit whole-body human motion database. In ICAR, 2015. 3", + "[53] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. Unifying representations and large-scale whole-body motion databases for studying human motion. IEEE Transactions on Robotics, 32(4):796-809, 2016. 3", + "[54] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3D pose estimation from monocular RGB. In 3DV, 2018. 2", + "[55] Josh Merel, Saran Tunyasuvunakool, Arun Ahuja, Yuval Tassa, Leonard Hasenclever, Vu Pham, Tom Erez, Greg Wayne, and Nicolas Heess. Catch & carry: reusable neural controllers for vision-guided whole-body tasks. ACM Transactions on Graphics (TOG), 39(4):39-1, 2020. 3", + "[56] Gyeongsik Moon, Shouu-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In ECCV, 2020. 2", + "[57] Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. In NeurIPS, 2024. 5", + "[58] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In CVPR, 2023. 2" + ], + "bbox": [ + 522, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7057", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 6", + "[60] OpenAI. ChatGPT. https://chat.openai.com/, 2023.3,1", + "[61] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. arXiv preprint arXiv:2308.09036, 2023. 3", + "[62] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 3, 4, 1, 2", + "[63] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. HOI-Diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 2, 3, 5, 6, 7", + "[64] Ilya A Petrov, Riccardo Marin, Julian Chibane, and Gerard Pons-Moll. Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In CVPR, 2023. 3", + "[65] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In ECCV, 2022. 4", + "[66] Mathis Petrovich, Michael J Black, and Gül Varol. TMR: Text-to-motion retrieval using contrastive 3d human motion synthesis. In ICCV, 2023. 6", + "[67] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236–252, 2016. 1", + "[68] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In ICCV, 2019. 5, 7, 1", + "[69] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english labels. In CVPR, 2021. 1", + "[70] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 7", + "[71] Haziq Razali and Yiannis Demiris. Action-conditioned generation of bimanual object manipulation sequences. In AAAI, 2023. 3", + "[72] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 6", + "[73] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6), 2017. 3, 4, 1, 2", + "[74] Wenfeng Song, Xinyu Zhang, Shuai Li, Yang Gao, Aimin Hao, Xia Hou, Chenglizhao Chen, Ning Li, and Hong Qin. Hoianimator: Generating text-prompt human-object animations using novel perceptive diffusion models. In CVPR, 2024. 3, 5, 6" + ], + "bbox": [ + 99, + 92, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 3", + "[76] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39(4):54-1, 2020. 3", + "[77] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In ECCV, 2020. 2, 3, 1, 5", + "[78] Omid Taheri, Vasileios Choutas, Michael J Black, and Dimitrios Tzionas. GOAL: Generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 3", + "[79] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J Black. Grip: Generating interaction poses using spatial cues and latent consistency. In 3DV, 2024. 4", + "[80] Purva Tendulkar, Dídac Surís, and Carl Vondrick. Flex: Full-body grasping without full-body grasps. In ICCV, 2023. 2", + "[81] Chen Tessler, Yunrong Guo, Ofir Nabati, Gal Chechik, and Xue Bin Peng. Maskedmimic: Unified physics-based character control through masked motion inpainting. arXiv preprint arXiv:2409.14393, 2024. 3", + "[82] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 4, 6", + "[83] Guy Tevet, Signal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. Closd: Closing the loop between simulation and diffusion for multi-task character control. arXiv preprint arXiv:2410.03441, 2024. 3", + "[84] Jie Tian, Lingxiao Yang, Ran Ji, Yuexin Ma, Lan Xu, Jingyi Yu, Ye Shi, and Jingya Wang. Gaze-guided hand-object interaction synthesis: Benchmark and method. arXiv preprint arXiv:2403.16169, 2024. 3", + "[85] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 2022. 3", + "[86] Jiashun Wang, Jessica Hodgins, and Jungdam Won. Strategy and skill learning for physics-based table tennis animation. In SIGGRAPH, 2024. 3", + "[87] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In 3DV, 2022. 3", + "[88] Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. PhysHOI: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393, 2023. 3, 5, 8", + "[89] Yinhuai Wang, Qihan Zhao, Runyi Yu, Ailing Zeng, Jing Lin, Zhengyi Luo, Hok Wai Tsui, Jiwen Yu, Xiu Li, Qifeng Chen, et al. Skillmimic: Learning reusable basketball skills from demonstrations. arXiv preprint arXiv:2408.15270, 2024.3" + ], + "bbox": [ + 522, + 92, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7058", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[90] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. InterControl: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 3", + "[91] Noah Wiederhold, Ava Megyeri, DiMaggio Paris, Sean Banerjee, and Natasha Banerjee. Hoh: Markerless multimodal human-object-human handover dataset with large object count. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[92] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In CoRL, 2023. 3", + "[93] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. THOR: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 3, 5, 6", + "[94] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. SAGA: Stochastic whole-body grasping with contact. In ECCV, 2022. 3, 4", + "[95] Zhen Wu, Jiaman Li, and C Karen Liu. Human-object interaction from human-level instructions. arXiv preprint arXiv:2406.17840, 2024. 3", + "[96] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. In ECCV, 2022. 3", + "[97] Xianghui Xie, Jan Eric Lenssen, and Gerard Pons-Moll. InterTrack: Tracking human object interaction without object templates. arXiv preprint arXiv:2408.13953, 2024. 3", + "[98] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. OmniControl: Control any joint at any time for human motion generation. arXiv preprint arXiv:2310.08580, 2023. 3", + "[99] Zhaoming Xie, Sebastian Starke, Hung Yu Ling, and Michiel van de Panne. Learning soccer juggling skills with layerwise mixture-of-experts. In SIGGRAPH, 2022. 3", + "[100] Zhaoming Xie, Jonathan Tseng, Sebastian Starke, Michiel van de Panne, and C Karen Liu. Hierarchical planning and control for box loco-manipulation. arXiv preprint arXiv:2306.09532, 2023. 3", + "[101] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter-x: Towards versatile human-human interaction analysis. arXiv preprint arXiv:2312.16051, 2023. 2", + "[102] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 1, 3, 4, 5, 8", + "[103] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. arXiv preprint arXiv:2403.19652, 2024. 3", + "[104] Sirui Xu, Hung Yu Ling, Yu-Xiong Wang, and Liang-Yan Gui. InterMimic: Towards universal whole-body control for physics-based human-object interactions. In CVPR, 2025. 1, 3, 4", + "[105] Xiang Xu, Hanbyul Joo, Greg Mori, and Manolis Savva. D3D-HOI: Dynamic 3d human-object interactions from videos. arXiv preprint arXiv:2108.08420, 2021. 3" + ], + "bbox": [ + 93, + 92, + 483, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[106] ChangHee Yang, ChanHee Kang, Kyeongbo Kong, Hanni Oh, and Suk-Ju Kang. Person in place: Generating associative skeleton-guidance maps for human-object interaction image editing. In CVPR, 2024. 3", + "[107] Jie Yang, Xuesong Niu, Nan Jiang, Ruimao Zhang, and Siyuan Huang. F-HOI: Toward fine-grained semantic-aligned 3d human-object interactions. In ECCV, 2024. 3", + "[108] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, and Zheng-Jun Zha. Lemon: Learning 3d human-object interaction relation from 2d images. In CVPR, 2024. 3", + "[109] Zeshi Yang, Kangkang Yin, and Libin Liu. Learning to use chopsticks in diverse gripping styles. ACM Transactions on Graphics (TOG), 41(4):1-17, 2022. 3", + "[110] Heyuan Yao, Zhenhua Song, Yuyang Zhou, Tenglong Ao, Baoquan Chen, and Libin Liu. Moconvq: Unified physics-based motion control via scalable discrete representations. ACM Transactions on Graphics (TOG), 43(4):1-21, 2024. 3", + "[111] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance diffusion: Synthesizing hand-object interactions. In CVPR, 2023. 3", + "[112] Xinyu Zhan, Lixin Yang, Yifei Zhao, Kangrui Mao, Hanlin Xu, Zenan Lin, Kailin Li, and Cewu Lu. Oakink2: A dataset of bimanual hands-object manipulation in complex task completion. arXiv preprint arXiv:2403.19417, 2024. 2", + "[113] Chengwen Zhang, Yun Liu, Ruofan Xing, Bingda Tang, and Li Yi. Core4d: A 4d human-object-human interaction dataset for collaborative object rearrangement. arXiv preprint arXiv:2406.19353, 2024. 2", + "[114] Hui Zhang, Sammy Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. ArtiGrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv preprint arXiv:2309.03891, 2023. 3", + "[115] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. NeuralDome: A neural modeling pipeline on multi-view humanobject interactions. In CVPR, 2023. 2, 3, 5", + "[116] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2M-GPT: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 4", + "[117] Juze Zhang, Jingyan Zhang, Zining Song, Zhanhe Shi, Chengfeng Zhao, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Hoi-m^3: Capture multiple humans and objects interaction within contextual environment. In CVPR, 2024. 3", + "[118] Jiajun Zhang, Yuxiang Zhang, Liang An, Mengcheng Li, Hongwen Zhang, Zonghai Hu, and Yebin Liu. Manidext: Hand-object manipulation synthesis via continuous correspondence embeddings and residual-guided diffusion. arXiv preprint arXiv:2409.09300, 2024. 3", + "[119] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In ECCV, 2020. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "7059", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[120] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. COUCH: Towards controllable human-chair interactions. In ECCV, 2022. 3", + "[121] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Ilya Petrov, Vladimir Guzov, Helisa Dhamo, Eduardo Pérez-Pellitero, and Gerard Pons-Moll. FORCE: Dataset and method for intuitive physics guided human-object interaction. arXiv preprint arXiv:2403.11237, 2024. 2", + "[122] Yunbo Zhang, Deepak Gopinath, Yuting Ye, Jessica Hodgins, Greg Turk, and Jungdam Won. Simulation and retargeting of complex multi-character interactions. In SIGGRAPH, 2023. 5", + "[123] Chengfeng Zhao, Juze Zhang, Jiashen Du, Ziwei Shan, Junye Wang, Jingyi Yu, Jingya Wang, and Lan Xu. I'M HOI: Inertia-aware monocular capture of 3d human-object interactions. In CVPR, 2024. 2, 3, 5", + "[124] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. In ECCV, 2022. 3", + "[125] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In ICCV, 2023. 3", + "[126] Juntian Zheng, Qingyuan Zheng, Lixing Fang, Yun Liu, and Li Yi. CAMS: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis. In CVPR, 2023. 3", + "[127] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In ECCV. Springer, 2022. 3, 4" + ], + "bbox": [ + 91, + 92, + 483, + 515 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "7060", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/4511c287-17f9-4b70-a6d5-d77c48a5192a_model.json b/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/4511c287-17f9-4b70-a6d5-d77c48a5192a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..774e084b803cb0ad6022633d912670d933516ad7 --- /dev/null +++ b/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/4511c287-17f9-4b70-a6d5-d77c48a5192a_model.json @@ -0,0 +1,2965 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.13, + 0.872, + 0.175 + ], + "angle": 0, + "content": "InterAct: Advancing Large-Scale Versatile 3D Human-Object Interaction Generation" + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.202, + 0.891, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.306, + 0.909, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.564, + 0.892, + 0.58 + ], + "angle": 0, + "content": "Figure 1. An overview of InterAct, our large-scale 3D human-object interaction (HOI) benchmark, covering six HOI generation tasks." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.599, + 0.327, + 0.615 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.486, + 0.904 + ], + "angle": 0, + "content": "While large-scale human motion capture datasets have advanced human motion generation, modeling and generating dynamic 3D human-object interactions (HOIs) remain challenging due to dataset limitations. Existing datasets often lack extensive, high-quality motion and annotation and exhibit artifacts such as contact penetration, floating, and incorrect hand motions. To address these issues, we introduce InterAct, a large-scale 3D HOI benchmark featuring dataset and methodological advancements. First, we consolidate and standardize 21.81 hours of HOI data from diverse sources, enriching it with detailed textual annotations. Second, we propose a unified optimization framework to enhance data quality by reducing artifacts and correcting hand motions. Leveraging the principle of contact invariance, we maintain human-object relationships while introducing motion variations, expanding the dataset to 30.70 hours. Third, we define six benchmarking tasks and develop a unified HOI generative modeling perspective, achieving state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.909, + 0.662 + ], + "angle": 0, + "content": "performance. Extensive experiments validate the utility of our dataset as a foundational resource for advancing 3D human-object interaction generation. The dataset will be publicly accessible to support further research in the field." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.693, + 0.646, + 0.709 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.856 + ], + "angle": 0, + "content": "Recent advances in human motion modeling have significantly benefited from extensive motion capture (MoCap) datasets [22, 41, 50, 67, 69], enabling the creation of scalable generative models for diverse human movements. Building upon this foundation, researchers are increasingly turning to the more intricate challenge of generating human-object interactions (HOIs) [102-104]. This emerging area holds considerable promise for applications in robotics, animation, and computer vision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.857, + 0.91, + 0.903 + ], + "angle": 0, + "content": "However, high-quality HOI generation faces notable obstacles due to factors such as increased degrees of freedom introduced by objects, varied object geometries, dynamic in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7048" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.246 + ], + "angle": 0, + "content": "
DatasetClipHourTextHandObject
GRAB [77]1,3353.76X✓51
BEHAVE [4]2994.13XX18
InterCap [28]2330.62X✓10
Chairs [29]1,0412.37X✓92
HODome [115]1762.82X✓21
OMOMO [38]4,8388.274,838X15
IMHD [123]1640.97X✓10
InterAct (Ours)11,35021.8134,050✓217
InterAct-X (Ours)16,20130.7048,630✓217
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.255, + 0.485, + 0.326 + ], + "angle": 0, + "content": "Table 1. Comparison between InterAct, InterAct-X, and human-object interaction datasets we collect. Beyond a substantially larger scale, our dataset introduces comprehensive textual annotations and enhances interaction quality, offering a more versatile foundation for large-scale HOI generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.484, + 0.444 + ], + "angle": 0, + "content": "teractions, and the necessity for physically accurate contact modeling. Current methods often struggle to achieve realism primarily because existing datasets lack scalability and comprehensive annotations, which are crucial for models to effectively understand interaction dynamics and link them to related domains such as natural language." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.444, + 0.484, + 0.655 + ], + "angle": 0, + "content": "Specifically, these challenges underscore the need for comprehensive, high-quality HOI datasets: (1) Limited and Inconsistent Datasets: Existing methods typically depend on small datasets with limited hours of data, difficult to consolidate due to inconsistent human representations, object types, coordinate systems, and annotations. Available annotations [38, 63] are frequently coarse and incomplete, lacking detailed descriptions of human states, object interactions, and involved body parts. (2) Prevalent Artifacts: Current datasets often contain artifacts from MoCap limitations and occlusions, including unnatural penetrations, floating contacts, inaccurate hand poses [4, 38], and significant motion jitter [28]. These issues compromise the models' capacity to learn realistic human-object dynamics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.746 + ], + "angle": 0, + "content": "To address these challenges, we present InterAct, a benchmark designed to systematically overcome current limitations and drive advancements in 3D HOI modeling. As shown in Table 1, InterAct offers a large-scale, standardized dataset of carefully curated interactions from existing resources1, enriched by detailed textual annotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.484, + 0.853 + ], + "angle": 0, + "content": "To further enhance dataset quality and scope, we introduce a unified optimization approach, addressing major penetration and floating artifacts first in whole-body interactions, followed by refined corrections for nuanced hand-object interactions. Additionally, we propose the concept of contact invariance, inspired by motion mirroring techniques, to generate realistic synthetic data by varying human motions while" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.228, + 0.908, + 0.27 + ], + "angle": 0, + "content": "Figure 2. (a) Our data processing pipeline consolidating data, annotations via foundation models, corrections, and interaction illustrations. (b) Statistics on motion and text annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.359 + ], + "angle": 0, + "content": "maintaining consistent object contacts. This augmentation expands InterAct into InterAct-X, providing approximately 9 additional hours of data and substantially improving generative model performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.909, + 0.496 + ], + "angle": 0, + "content": "Leveraging this comprehensive, richly annotated dataset, we define benchmarks across six key HOI generation tasks, Text-to-Interaction, Action-to-Interaction, Object-to-Human, Human-to-Object, Interaction Prediction, and Interaction Imitation, as shown in Figure 1, and propose a unified modeling and representation for kinematic generative tasks. Our method utilizes multi-task learning to jointly model motion and contact, achieving state-of-the-art performance as validated by comprehensive evaluations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.909, + 0.679 + ], + "angle": 0, + "content": "In summary, our contributions are: (1) InterAct, the most extensive 3D HOI benchmark to date, facilitating large-scale generative modeling. (2) A unified optimization-based framework to correct and augment MoCap data, addressing common artifacts and significantly enhancing dataset quality. We believe this will aid future research in overcoming data scarcity before capturing potentially imperfect data. (3) Comprehensive benchmarks across six HOI generation tasks, establishing standardized metrics and demonstrating superior performance over existing approaches. This benchmark lays a strong foundation for future research, encouraging advancements across multiple facets of 3D HOI generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.693, + 0.655, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Dynamic 3D HOI Dataset. Many large-scale datasets with sequential human motion data have established benchmarks for the task of 3D human motion generation. However, human actions are influenced not only by individual intents but also by interactions with the surrounding environment. To address this complexity, new datasets have been developed to capture the dynamics between humans and their environments, including interactions with other humans [1, 40, 54, 101] and scenes [8, 24, 25]. Though there are fruitful hand-object interaction datasets [11, 23, 44, 48, 56, 58, 80, 91, 91, 112, 113, 121], our focus is specifically on whole-body interactions with dynamic objects, ranging from" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1We integrate publicly available datasets [4, 28, 29, 38, 77, 115, 123] focusing on single-human interactions with rigid and dynamic objects. Certain datasets were selectively included based on relevance and data type." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7049" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.137 + ], + "angle": 0, + "content": "low-dynamic interactions such as approaching and manipulation [77] to highly dynamic interactions involving multiple body parts [4, 20, 28-30, 32, 38, 107, 115, 123]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.486, + 0.242 + ], + "angle": 0, + "content": "We aim to address the limitations of these datasets and open new possibilities for future research. Our InterAct dataset maintains advantages in motion quality, fine-grained textual annotations, detailed hand gestures, and comprehensive annotation modalities. We provide quantitative comparisons of InterAct and existing datasets in Table 1, demonstrating the superiority of our dataset in these aspects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.487, + 0.788 + ], + "angle": 0, + "content": "Dynamic 3D HOI Generation. Existing human-object interaction (HOI) datasets have laid a robust foundation for generating dynamic, whole-body interactions. Extensive research has explored the generation of hand-object interactions [9, 13, 39, 43, 49, 84, 111, 114, 118, 126, 127] and static human-object interactions [27, 33, 64, 87, 96, 97, 106, 108, 119, 124]. Meanwhile, full-body dynamic interactions have also been studied extensively [14, 21, 34-36, 38, 52, 53, 71, 75, 76, 78, 85, 94, 105, 120, 125], though these often face significant limitations, including narrow action repertoires and dependence on static objects. Recent advancements, such as InterDiff [102], have introduced diverse interactions involving dynamic objects and multiple body parts. Building upon this, subsequent approaches like InterDreamer [103] and other contemporary studies [18, 37, 63, 74, 93, 95, 103, 117] further demonstrate the feasibility of converting textual descriptions into realistic 3D human-object interaction sequences. Despite these advances, current methods remain constrained by a shortage of high-quality, large-scale datasets, often encountering issues related to physical inaccuracies, such as floating contacts or interpenetration. In parallel, physics-based methods leveraging deep reinforcement learning (RL) [3, 6, 10, 15, 26, 42, 55, 61, 81, 83, 86, 89, 90, 92, 98, 100, 109, 110] successfully generate physically accurate interactions, with applications in sports [47] like basketball [88, 89] and soccer [99]. Nonetheless, these methods typically produce rigid interaction patterns from limited datasets, while a recent work, InterMimic [104], illustrates that physics-based approaches can digest effectively across diverse and dynamic object interactions. Our work addresses these fundamental limitations at the dataset level by providing enhanced diversity and comprehensive sequences of human-object interactions. This facilitates multiple generative tasks, supports better contact modeling, and improves the capability to synthesize realistic and generalized human-object interactions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.258, + 0.815 + ], + "angle": 0, + "content": "3. InterAct Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Overview. We introduce InterAct, the first unified benchmark tailored explicitly for sequential 3D human-object interaction (HOI) generative modeling. Distinguished by its unprecedented scale and comprehensiveness, InterAct significantly surpasses existing datasets, as summarized in Table 1." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.701, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.164, + 0.666, + 0.175 + ], + "angle": 0, + "content": "(a) Markers in SMPL-H" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.089, + 0.905, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.745, + 0.164, + 0.865, + 0.175 + ], + "angle": 0, + "content": "(b) Markers in SMPL-X" + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.186, + 0.86, + 0.201 + ], + "angle": 0, + "content": "Figure 3. Marker-based representation for human." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.909, + 0.396 + ], + "angle": 0, + "content": "InterAct is available in two versions: (1) A basic version consolidating seven existing datasets, providing 21.81 hours of annotated 3D whole-body interactions with corresponding semantic descriptions. (2) An advanced version, InterAct-X, extending the basic dataset through synthetic data generated via our unified optimization framework. Figure 2 presents examples of motion and text annotations. To ensure high-quality standards, we employ a multifaceted annotation strategy combining human expertise, automated foundation models, and advanced HOI modeling techniques, all validated through rigorous manual quality checks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.408, + 0.892, + 0.423 + ], + "angle": 0, + "content": "3.1. Data Collection, Annotation, and Unification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.908, + 0.507 + ], + "angle": 0, + "content": "We compile data from seven datasets [4, 28, 29, 38, 77, 115, 123], featuring motion capture of a single human interacting with a single dynamic 3D object, where humans are annotated with SMPL [45, 62, 73]. We address their heterogeneity in two key aspects: annotations and representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Unifying Textual Annotations. Since most datasets either lack textual descriptions or provide only very coarse text descriptions [38], we implement a two-phase annotation procedure involving human annotators and GPT-4 [60] to generate consistent and detailed annotations across all subsets. In the first phase, human annotators provided detailed and precise descriptions of the interactions, adhering to the following guidelines: (i) Split motion sequences into clips averaging 300 frames (approximately 10 seconds) but no longer than 400 frames each; (ii) Clearly describe the actions and the body parts involved in the interactions. For example, a typical annotation is: \"A person sits on a stool and touches the ground with their left hand, then their right hand.\" For the subset derived from OMOMO [38], we skip this phase and directly utilize their annotations. In the second phase, we use GPT-4 to rephrase and simplify human annotations to enhance diversity and consistency. For example, the rephrased version is \"A person perches on a stool, touching the ground with their left hand, then their right hand,\" and the simplified version is \"A person sits on a stool, touching the ground with each hand alternately.\" Next, we employ GPT-4 to classify each description into one of our predefined 15 action labels with in-context learning [7]. The action label for the above sequence is \"Sit.\" We meticulously review all generated texts and action labels to ensure high quality and alignment across the dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7050" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.394 + ], + "angle": 0, + "content": "Unifying Human Representations. Different datasets employ varying human models (e.g., SMPL-H [73], SMPL-X [62]) and diverse shapes. A straightforward solution can be to convert different humans from SMPL-H and SMPL-X to a consistent SMPL version and encode shape parameters into the generative modeling. However, although SMPL is widely used in various human-related tasks, it is fundamentally a rotation-based representation. In the context of human motion generation, Cartesian features like joint positions and velocities are more commonly used, as seen in the integration with the HumanML3D representation [22] and in most text-to-motion work [65, 82, 116]. This is still suboptimal because joints are located beneath the body's surface and do not explicitly participate in interactions. To overcome the limitations, we use markers - specific sets of human vertices representing human motion and interactions - as a simple and unified representation capable of effectively inferring contact, evaluated in Table 5. Similar approaches are discussed in [94, 102]. Then we need to select a marker set that is consistent between SMPL-H and SMPL-X models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.396, + 0.487, + 0.759 + ], + "angle": 0, + "content": "Given two human body models, SMPL-H and SMPL-X, sharing the same shape, we establish marker correspondences in two steps. First, we index the markers on the SMPL-H surface as defined in prior work [94, 102]. Second, we locate the corresponding vertices on SMPL-X by selecting the closest points to these SMPL-H markers, leveraging the official SMPL conversion, which maps each SMPL-H vertex to the nearest point on the SMPL-X mesh. We extensively evaluated the approximation error of these marker correspondences across a broad range of poses. Our results show that the maximum error consistently remains below \\(1\\mathrm{cm}\\), a deviation unlikely to affect the overall performance of HOI generation. This high consistency arises because the markers are rigidly attached to the body, and soft deformations are disabled. As a result, the identical rigid transformations in SMPL-H and SMPL-X preserve the correspondence of the markers. Additional details on such correspondence-preserving conditions can be found in [31]. Figure 3 illustrates the marker sets for SMPL-H and SMPL-X. We use this marker-based representation to train the generative models for the tasks outlined in Sec. 4, while still relying on the original SMPL-H or SMPL-X representations for the interaction correction and augmentation methods described in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.449, + 0.787 + ], + "angle": 0, + "content": "3.2. Interaction Correction and Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.486, + 0.902 + ], + "angle": 0, + "content": "In this section, we present a unified optimization framework that addresses both the correction of MoCap artifacts and the augmentation of the dataset by introducing more synthetic data. The process takes as input the motions and geometries of humans and objects, then compares them against predefined standards to define loss functions. Using gradient-based optimization, we iteratively adjust human and object" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.152 + ], + "angle": 0, + "content": "motions to minimize these losses, thereby refining the data to meet the desired quality criteria. The key challenge lies in formulating learning objectives that not only rectify existing data but also facilitate the generation of new synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.909, + 0.469 + ], + "angle": 0, + "content": "Our optimization is carried out in three sequential steps: (1) full-body correction; (2) hand correction; and (3) interaction augmentation. Hand correction is handled separately because, although hand poses occupy considerable space in the SMPL representation, they contribute relatively little to the overall scale of the learning objectives. By decoupling hand correction from full-body correction, we can better balance these two processes and define more targeted objectives. In what follows, we first introduce the hand correction stage. Hand Correction. Given that many existing datasets contain inaccurate hand poses [4, 38], our approach selectively promotes contact only in regions where ground-truth data indicates hand-object interaction, while ensuring the hand motion remains natural, in spirit to InterMimic [104] but relying on predefined optimization instead of RL. This approach is effective for the whole-body interaction datasets we utilize, which generally do not require high dexterity and typically only involve the hand conforming to the object for grasping, as a common assumption in existing work [79, 127], while we distinguish our approach from those that employ multi-stage, learning-based methods for the same purpose." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.47, + 0.909, + 0.515 + ], + "angle": 0, + "content": "We divide our hand correction objectives into two categories: contact promotion and hand constraints. Contact promotion is guided by the following contact loss:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.52, + 0.795, + 0.563 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathrm {c o n t}} = \\sum_ {i = 1} ^ {L} c _ {i} \\sum_ {j} d _ {j} [ i ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.811 + ], + "angle": 0, + "content": "where \\(d_{j}[i]\\) is the distance between the \\(j\\)-th hand vertex and its nearest point on the object's surface at the \\(i\\)-th frame, and \\(c_{i}\\) indicates whether the object and the hand are in contact at frame \\(i\\). The contact indicator \\(c_{i}\\), inferred from ground truth data, is a function based on hand-object distance \\(\\min_j d_j[i]\\), which we provide details in supplementary. The hand constraint objectives are introduced to preserve naturalness and temporal smoothness in the hand motions. These constraints include: (1) penetration loss, which penalizes intersections between the hand and the object. (2) smoothness loss, which promotes consistent contact and reduces jittering. (3) prior loss, which constrains the range of motion (RoM) of the fingers to maintain realism. Without this constraint, contact promotion could inadvertently drive fingers into biologically impossible poses. Detailed formulations of these loss functions are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Full-Body Correction. In this stage, all human and object poses can be updated via gradient descent. We add a reconstruction loss to ensure that the optimized interactions closely match the ground truth. Other losses mirror those used in hand correction, with two key differences: (1) Contact and penetration losses are computed for the entire body" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.958 + ], + "angle": 0, + "content": "7051" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "rather than just the hands. (2) Prior loss is omitted because the reconstruction loss alone suffices to maintain plausible human motion. Detailed formulations of these losses are provided in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.484, + 0.408 + ], + "angle": 0, + "content": "Interaction Augmentation. Synthetic data has become increasingly important in computer vision and generative modeling [5, 12, 19, 57], prompting a key question in the context of HOI animation: Can we scale up datasets without collecting additional MoCap data? Does existing interaction data offer information beyond its observable motions? Consider a scenario where a person grasps a box and walks, as illustrated in Figure 5. Even if their gait changes slightly, hand-box contact should remain consistent to preserve the semantics of the interaction. This illustrates the principle of interaction invariance: the core interaction persists despite minor variations in motion. Leveraging this principle, we can augment our dataset by injecting new human motions while preserving consistent object interactions. Training neural networks on such augmented data enables them to naturally learn this invariance, a common strategy in symmetric learning [17], and ultimately enhances model performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.573 + ], + "angle": 0, + "content": "Our augmentation pipeline consists of three steps: (1) Object Displacement: We apply a random displacement to the object's trajectory, uniformly across all timesteps. (2) Interaction Alignment: We optimize the human motion to maintain interaction with the displaced object, using both contact consistency and reconstruction objectives. (3) Interaction Filtering: We remove low-quality augmentations, those with unreasonable initial displacements, significant penetrations (human-object or self-penetration), or alignment failures indicated by high optimization losses (e.g., excessive jitter)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.681 + ], + "angle": 0, + "content": "During the alignment phase, the primary objective is the contact consistency loss. We first compute a distance matrix \\(\\mathbf{D}\\), where each element \\(\\mathbf{D}_{jk} = \\|\\mathbf{v}_h^j - \\mathbf{v}_o^k\\|\\) denotes the Euclidean distance between the \\(j\\)-th human vertex and the \\(k\\)-th object vertex, before or after displacement. With a reference matrix \\(\\hat{\\mathbf{D}}\\) from the original (pre-displacement) setup, we optimize the human motion using:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.69, + 0.438, + 0.734 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathrm {a l i g n}} = \\sum_ {i = 1} ^ {L} \\sum_ {j, k} \\frac {1}{(\\hat {\\mathbf {D}} _ {j k} + \\epsilon) ^ {2}} \\left| \\hat {\\mathbf {D}} _ {j k} - \\mathbf {D} _ {j k} \\right| ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.484, + 0.833 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a small constant to prevent division by zero. This formulation preserves distances between vertex pairs that were initially close, while de-emphasizing pairs that were farther apart. Additional terms in the objective enforce naturalness and stability in non-interactive regions of the human pose, as detailed in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.279, + 0.862 + ], + "angle": 0, + "content": "4. Tasks and Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we formally define six distinct tasks featured in our benchmark. We use unified representation across" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.212 + ], + "angle": 0, + "content": "five kinematic generative tasks, where each human-object interaction sequence is represented as \\(\\langle h,o\\rangle\\), annotated with an action category \\(a\\) and a text description \\(t\\). The human \\(h\\) includes marker coordinates, marker velocities, signed distance vectors from each marker to the object, and foot-ground contact labels. The object \\(o\\) represents object motion, including object rotation angles, object translations. Object geometry is described by Basis Point Set (BPS) [68]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.907, + 0.258 + ], + "angle": 0, + "content": "(1) Text-Conditioned Interaction Generation. Initially in [18, 63], the task learns a function to generate the interaction sequence based on text: \\(\\mathcal{G}_{\\mathrm{t2i}}(t)\\mapsto \\langle h,o\\rangle\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.906, + 0.304 + ], + "angle": 0, + "content": "(2) Action-Conditioned Interaction Generation. The objective is to learn a function that maps an action label to the corresponding interaction sequence: \\(\\mathcal{G}_{\\mathrm{a2i}}(\\pmb {a})\\mapsto \\langle \\pmb {h},\\pmb {o}\\rangle\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.304, + 0.906, + 0.348 + ], + "angle": 0, + "content": "(3) Object-Conditioned Human Generation. Initially in [38], the task generates human motion based on object sequences through a function \\(\\mathcal{G}_{\\mathrm{o2h}}(\\pmb {o})\\mapsto \\pmb{h}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.348, + 0.906, + 0.394 + ], + "angle": 0, + "content": "(4) Human-Conditioned Object Generation. Conversely, this task focuses on generating object motion sequences from human motion sequences via a function \\(\\mathcal{G}_{\\mathrm{h2o}}(\\pmb {h})\\mapsto \\pmb{o}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.394, + 0.906, + 0.455 + ], + "angle": 0, + "content": "(5) Interaction Prediction. Initially in [102], the task aims to predict future human-object interactions based on past. Let \\(\\langle \\pmb{h}_p, \\pmb{o}_p \\rangle\\) denote the past interaction and \\(\\langle \\pmb{h}_f, \\pmb{o}_f \\rangle\\) the future. The goal is to learn: \\(\\mathcal{G}_{\\mathrm{p2f}}(\\langle \\pmb{h}_p, \\pmb{o}_p \\rangle) \\mapsto \\langle \\pmb{h}_f, \\pmb{o}_f \\rangle\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.907, + 0.56 + ], + "angle": 0, + "content": "(6) Interaction Imitation. Following [3, 88, 122], this task focuses on learning physics-based control policies to reproduce human-object interactions in a physics simulator. The output is an action sequence \\( \\pmb{f} \\), specified as joint Proportional-Derivative (PD) targets. The goal is to learn a function \\( \\mathcal{G}_{\\mathrm{i2f}} \\) that maps the reference interaction sequences to the PD actuation sequences: \\( \\mathcal{G}_{\\mathrm{i2f}}(\\langle h,o\\rangle)\\mapsto f \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.909, + 0.787 + ], + "angle": 0, + "content": "Unifying Multi-Task HOI Generation. We introduce an additional feature, \\(\\eta\\), which encodes human-object relationships through vectors extending from each human marker to its nearest point on the object's surface. The specific configuration of \\(\\eta\\) for each generative task is described in Sec. 5. Using this feature, we can unify first five kinematic generative tasks into a multi-task learning framework by treating \\(\\eta\\) as an additional output. For example, we redefine the text-conditioned interaction task as \\(\\mathcal{G}_{\\mathrm{t2i}}(t) \\mapsto \\langle h, o, \\eta \\rangle\\), where \\(\\mathcal{G}\\) is a transformer-based diffusion model. This formulation compels the model to learn spatial relationships inherent to the interactions. In our experiments, we observe that this simple strategy, enhanced by large-scale data, consistently outperforms existing methods. Similar ideas are explored in [18, 37, 38, 63, 74, 93, 102]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.645, + 0.817 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We begin by evaluating the effectiveness of our data correction and augmentation methods. Following this, we benchmark existing work and our proposed method on the tasks using our dataset. We standardize the evaluation metrics and present extensive results, including ablation studies. We in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7052" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.086, + 0.912, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.271, + 0.908, + 0.3 + ], + "angle": 0, + "content": "Figure 4. Qualitative evaluation of interaction correction (bottom) on the OMOMO [38] dataset shows hand recovery compared to the ground truth interaction (top). Zoom in to see details of the hand recovery." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.322, + 0.485, + 0.406 + ], + "angle": 0, + "content": "
DatasetCorrectionAugmentationPene (m)↓Cont RatioUser Study (%)
BEHAVE [4]××0.0170.04822.3
✓×0.0160.07139.7
✓✓0.0160.06938.0
OMOMO [38]××0.0090.07123.9
✓×0.0070.13139.4
✓✓0.0110.13736.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.445 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation and user study on the quality of data from interaction correction and augmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.503 + ], + "angle": 0, + "content": "clude additional implementation details, such as the train-test split, in supplementary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.514, + 0.358, + 0.53 + ], + "angle": 0, + "content": "5.1. Correction and Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.627 + ], + "angle": 0, + "content": "Metrics. We use the following two metrics: Penetration refers to the intersection depth – maximum of negative sign distances from human vertices to the object's surface – average across the sequence. Contact Ratio represents the average ratio of human vertices where their distances to object are under a threshold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.719 + ], + "angle": 0, + "content": "Quantitative Evaluations. Table 2 shows that our correction process significantly improves the quality of the original MoCap data by enhancing human-object contact and reducing penetration artifacts. Moreover, the quality of the augmented data is comparable to that of the corrected data and exceeds the quality of the original dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Qualitative Evaluations. Recognizing that quantitative metrics may not fully capture data quality, we conducted a double-blind user study. We randomly selected sequences from each of the raw, corrected, and augmented data for the subset from BEHAV [4] and OMOMO [38] datasets. Human judges were presented with 30 tuple of interactions and asked to rank the quality of three sequences. According to Table 2, over \\(39\\%\\) of judges select the corrected data as having the highest quality, significantly outperforming the original data. This confirms that our correction process effectively enhances data realism. Moreover, the augmented data receive ratings comparable to the corrected data, indicating" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.907, + 0.417 + ], + "angle": 0, + "content": "that our synthetic data is of high quality. In Figure 4, we visualize our correction results. Despite the original data lacking detailed hand information, we successfully recover vivid and accurate hand interactions. Figure 5 showcases our augmentation, which introduces new high-quality synthetic data while maintaining consistent contact in interactions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.425, + 0.857, + 0.441 + ], + "angle": 0, + "content": "5.2. Language Conditioned HOI Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.909, + 0.735 + ], + "angle": 0, + "content": "Metrics. Following the literature on text-to-motion generation [22], we develop five metrics for evaluation. The Fréchet Inception Distance (FID) quantifies the similarity between generated HOI features and the ground truth. The Multimodality and Diversity metrics assess the variety within the generated HOI. R-Precision measures the alignment between the textual descriptions and the generated HOI. The Multimodal Distance (MM Dist) evaluates the disparity between HOI features and corresponding text features. To obtain human-object interaction (HOI) and text features for calculating these metrics, existing methods often train their feature extractors on very limited data [18, 63, 74, 93], which can degrade the quality of the evaluation. We address this limitation by incorporating our larger-scale data with marker and BPS representations. Instead of formulating a classification task to train the feature extractor [22], we follow [46, 66] and employ sequence-level contrastive learning with an InfoNCE loss [59] to train a text encoder and an HOI encoder, integrating Sentence-BERT [72] into the text encoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Baselines and Implementation Details. We adopt HOI-Diff [63] as our base model because it is the only publicly available option compatible with our requirements. For example, CHOIS [37] requires additional conditions beyond text input. HOI-Diff utilizes a transformer-based diffusion model [82] as the backbone, and integrates an affordable model as the classifier guidance [16]. We develop several baseline variants towards our final method by implementing three key modifications: (i) Text Encoder: We replace HOI-Diff's CLIP-based text encoder, where the latent space is not structured for human-object interactions, with our pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7053" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.912, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.308, + 0.887, + 0.323 + ], + "angle": 0, + "content": "Figure 5. Qualitative evaluation of interaction augmentation (bottom) shows high-quality synthetic data varied from original (top)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.335, + 0.907, + 0.451 + ], + "angle": 0, + "content": "
HOI-Aware Object Enc.HOI-Aware Text Enc.Contact GenerationContact GuidanceR-Precision†FID↓MM Dist↓Multimodality↑Diversity→
Top 1Top 2Top 3
Ground Truth0.852±0.0000.966±0.0010.989±0.0010.000±0.0002.810±0.002-11.489±0.011
XXXX0.733±0.0070.909±0.0020.957±0.0023.192±0.1914.950±0.0233.149±0.45211.192±0.019
XX✓X0.730±0.0070.913±0.0040.958±0.0051.997±0.0924.752±0.0654.171±0.02711.501±0.037
✓X✓X0.737±0.0110.912±0.0020.963±0.0081.837±0.1264.631±0.0782.836±0.58311.369±0.096
✓✓✓X0.784±0.0040.940±0.0020.980±0.0031.570±0.1394.414±0.0642.677±0.56211.409±0.005
✓✓✓✓0.784±0.0040.940±0.0000.977±0.0021.567±0.1444.412±0.0653.842±0.00511.518±0.178
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.461, + 0.885, + 0.475 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation on the task of text-conditioned interaction generation. A batch size of 64 is used for R-Precision." + }, + { + "type": "table", + "bbox": [ + 0.133, + 0.498, + 0.445, + 0.563 + ], + "angle": 0, + "content": "
MethodFID↓Multimodality↑Diversity→
Ground Truth0.000±0.000-11.489±0.011
HOI-Diff [63]3.566±0.0985.321±0.14310.989±0.112
Ours2.161±0.0375.792±0.05911.291±0.261
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.572, + 0.483, + 0.601 + ], + "angle": 0, + "content": "Table 4. Quantitative evaluation on the task of action-conditioned interaction generation on the entire InterAct testset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.749 + ], + "angle": 0, + "content": "trained interaction-aware text encoder. (ii) Object Shape Encoding: We substitute the original PointNet++ [70], also not pretrained for capturing interaction, with BPS [68] representations for encoding object shapes. (iii) Instead of using affordance as guidance, we regress the contact representation \\(\\eta\\) through a multi-task learning. (iv) We further incorporate the contact prediction as classifier guidance [16] within the denoising process, which we detail in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Quantitative Results. Table 3 presents the evaluation results for the text-conditioned interaction generation task. We assess the impact of four design choices introduced above. Notably, incorporating contact modeling and BPS encoding significantly improves the quality of generated HOI, substantially enhancing the FID score. Furthermore, using our interaction-aware text and HOI encoder enhances the quality of the generated interactions and improves the alignment between the generated results and the text input. Lastly, classifier guidance provides a slight overall performance improve" + }, + { + "type": "table", + "bbox": [ + 0.594, + 0.498, + 0.83, + 0.573 + ], + "angle": 0, + "content": "
RepresentationPene (m)↓Cont Ratio
SMPL0.0300.025
Joint0.0270.032
Marker0.0250.028
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.583, + 0.907, + 0.611 + ], + "angle": 0, + "content": "Table 5. Ablation study on different representation for text-to-interaction task, evaluated under BEHAVE [4] subset." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.704 + ], + "angle": 0, + "content": "
ModelMPMPE↓FS↓Cprec↑Crec↑Cacc↑F1 Score↑
2-stage [38]36.500.270.810.850.770.80
1-stage36.940.290.840.820.800.81
Ours (Disc.)36.950.280.850.840.830.82
Ours (Cont.)35.690.280.850.890.850.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.715, + 0.907, + 0.743 + ], + "angle": 0, + "content": "Table 6. Quantitative evaluation on object-conditioned human motion generation, with novel objects unseen from training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.907, + 0.822 + ], + "angle": 0, + "content": "ment. Table 4 benchmarks the performance improvements of these design choices similarly for the action-conditioned interaction generation task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Marker-Based Representation. As shown in Table 5, we compare different human representations on text-to-interaction generation. Without altering contact modeling, marker-based representation produces interactions with fewer artifacts compared to other representations." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7054" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.249, + 0.108 + ], + "angle": 0, + "content": "5.3. HOI Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.31 + ], + "angle": 0, + "content": "Metrics. Following OMOMO [38], we develop six metrics tailored for evaluating our marker-based representation. The Mean Per-Marker Position Error (MPMPE) are used to measure the similarity between the generated marker motion and the ground truth. Foot Sliding (FS) is employed to assess the skating effect, reflecting the plausibility of the motion. Additionally, we use a set of contact metrics, including precision \\((C_{\\mathrm{prec}})\\), recall \\((C_{\\mathrm{rec}})\\), accuracy \\((C_{\\mathrm{acc}})\\), and the F1 Score, to evaluate the quality of human-object contact compared to the ground truth. For the human-conditioned object generation task, we include \\(T_{\\mathrm{err}}\\) and \\(O_{\\mathrm{err}}\\), which measure discrepancies in object translation and orientation between the generated and the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.311, + 0.484, + 0.521 + ], + "angle": 0, + "content": "Baselines and Implementation Details. We adopt OMOMO [38] as our base model since it is the only publicly available option. OMOMO employs a two-stage generation process: it first generates hand motions and then uses these to guide full-body generations. This strategy is particularly effective for the OMOMO dataset, where interactions primarily involve hand contact. However, it is less effective when applied to our InterAct data, which features more versatile whole-body engagement. Motivated by this, we use a single-stage pipeline with multi-task learning, as introduced in Sec. 4. We investigate two different choices for contact regression: \\(\\eta_{\\mathrm{Cont.}}\\), which encodes the nearest vector and distance between human markers and the object. \\(\\eta_{\\mathrm{Disc.}}\\), which encodes the contact labels for each marker." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.485, + 0.613 + ], + "angle": 0, + "content": "Quantitative Results. Tables 6 and 7 illustrate the effectiveness of our single-stage pipeline, with notable performance improvements when incorporating multi-task modeling. These results provide additional evidence, complementing the evaluation of text-conditioned generation, that multitask learning significantly enhances model performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.62, + 0.298, + 0.635 + ], + "angle": 0, + "content": "5.4. Interaction Prediction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Evaluation Metrics. We compare the generated poses to the ground truth motion data using MPMPE (mean perimeter errors), measured in meters. Second, we assess object motion accuracy using Trans.Err., the average \\( l_{2} \\) distance between the predicted and ground truth object translations, and Rot.Err., the average \\( l_{1} \\) distance between the predicted and ground truth object quaternions, following [102]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.795 + ], + "angle": 0, + "content": "Implementation Details. We adapt InterDiff [102] to utilize a marker-based representation and evaluate it at various scales to assess whether it benefits from scaling laws." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.29, + 0.817 + ], + "angle": 0, + "content": "5.5. Interaction Imitation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. We use PhysHOI [88] to imitate sequences from our InterAct dataset, selecting four sequences shown in Figure 6. The IsaacGym [51] simulator is used, following the same architecture, reward, and representation design as PhysHOI. Training for each sequence," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.177, + 0.905, + 0.205 + ], + "angle": 0, + "content": "Figure 6. Qualitative results demonstrate the successful imitation of our corrected data using PhysHOI [88]." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.266 + ], + "angle": 0, + "content": "
ModelTerr↓Oerr↓Cprec↑Crec↑Cacc↑F1 Score↑
1-stage25.920.910.830.660.720.69
Ours (multi-task)23.980.830.840.680.740.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.278, + 0.871, + 0.292 + ], + "angle": 0, + "content": "Table 7. Quantitative evaluation on human-to-object." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.306, + 0.905, + 0.399 + ], + "angle": 0, + "content": "
Training dataModel sizeGlobal MPMPE↓Local MPMPE↓Trans. Err.↓\\( \\text{Rot.Err.}^{\\downarrow } \\)
BEHAV×10.1200.1030.1330.352
×20.1050.0920.1090.312
×30.1130.1000.1180.343
InterAct (Ours)×10.1060.0950.1060.297
×20.0940.0830.1030.286
×30.0910.0790.0940.264
" + }, + { + "type": "table_caption", + "bbox": [ + 0.531, + 0.41, + 0.885, + 0.423 + ], + "angle": 0, + "content": "Table 8. Quantitative evaluation on interaction prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.905, + 0.495 + ], + "angle": 0, + "content": "with separate evaluations for both raw and corrected data, is performed on a single NVIDIA A40 GPU over the course of one day." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.906, + 0.631 + ], + "angle": 0, + "content": "Quantitative Evaluation. In addition to Figure 6, which presents the qualitative results, we evaluate the imitation policy by training on both corrected and raw data, reporting the success rate as defined in [88]. Evaluating four examples from Figure 6 and averaging the success rates over 2048 environments, training on the corrected data achieves a success rate of \\(90.7\\%\\), surpassing the \\(84.4\\%\\) achieved with raw data. This demonstrates that our interaction correction provides better data for the motion imitation task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.632, + 0.906, + 0.692 + ], + "angle": 0, + "content": "Quantitative Results. Table 8 demonstrates that our dataset, with its larger volume of data, supports the improved performance of the trained model with larger scale. In contrast, training the model on limited data leads to overfitting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.706, + 0.634, + 0.721 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.907, + 0.897 + ], + "angle": 0, + "content": "We introduce InterAct, a large-scale 3D whole-body human-object interaction benchmark. We employ a unified optimization framework that performs interaction correction and augmentation, enhancing data quality and augment the dataset with synthetic data, scaling to 30.70 hours of interactions and 48,630 textual descriptions. We introduce a simple yet effective multi-task learning approach for unified HOI modeling, enabling models to be trained more effectively across multiple tasks. Our comprehensive experiments highlight the significant advantages of InterAct and our methodologies, resulting in more expressive interaction generation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7055" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.288 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Amazon-Illinois Center on AI for Interactive Conversational Experiences, the Toyota Research Institute, the IBM-Illinois Discovery Accelerator Institute, and Snap Inc. This work used computational resources, including the NCSA Delta and DeltaAI and the PTI Jetstream2 supercomputers through allocations CIS230012, CIS230013, and CIS240311 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, as well as the TACC Frontera supercomputer, Amazon Web Services (AWS), and OpenAI API through the National Artificial Intelligence Research Resource (NAIRR) Pilot." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.303, + 0.188, + 0.319 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.328, + 0.484, + 0.355 + ], + "angle": 0, + "content": "[1] CMU graphics lab motion capture database. http://mocap.cs.cmu.edu/.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.357, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[2] Easymocap - make human motion capture easier. Github, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.387, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[3] Jinseok Bae, Jungdam Won, Donggeun Lim, Cheol-Hui Min, and Young Min Kim. Pmp: Learning to physically interact with environments using part-wise motion priors. In SIGGRAPH, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.444, + 0.484, + 0.499 + ], + "angle": 0, + "content": "[4] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAV: Dataset and method for tracking human object interactions. In CVPR, 2022. 2, 3, 4, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.501, + 0.484, + 0.543 + ], + "angle": 0, + "content": "[5] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.545, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[6] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. arXiv preprint arXiv:2309.07907, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.602, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.659, + 0.484, + 0.7 + ], + "angle": 0, + "content": "[8] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.702, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[9] Junuk Cha, Jihyeon Kim, Jae Shin Yoon, and Seungryul Baek. Text2hoi: Text-guided 3d motion generation for hand-object interaction. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.746, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[10] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In AAAI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.789, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[11] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[12] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.328, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "gpt4v-synthesized data for a lite vision-language model. arXiv preprint arXiv:2402.11684, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.908, + 0.188 + ], + "angle": 0, + "content": "[13] Sammy Christen, Shreyas Hampali, Fadime Sener, Edoardo Remelli, Tomas Hodan, Eric Sauser, Shugao Ma, and Bugra Tekin. Diffh2o: Diffusion-based synthesis of handobject interactions from textual descriptions. arXiv preprint arXiv:2403.17827, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.19, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[14] Enric Corona, Albert Pumarola, Guillem Alenya, and Francesc Moreno-Noguer. Context-aware human motion prediction. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.233, + 0.908, + 0.286 + ], + "angle": 0, + "content": "[15] Jieming Cui, Tengyu Liu, Nian Liu, Yaodong Yang, Yixin Zhu, and Siyuan Huang. AnySkill: Learning open-vocabulary physical skill for interactive agents. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.288, + 0.907, + 0.315 + ], + "angle": 0, + "content": "[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.317, + 0.908, + 0.356 + ], + "angle": 0, + "content": "[17] Sander Dieleman, Jeffrey De Fauw, and Koray Kavukcuoglu. Exploiting cyclic symmetry in convolutional neural networks. In ICML, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.358, + 0.907, + 0.397 + ], + "angle": 0, + "content": "[18] Christian Diller and Angela Dai. CG-HOI: Contact-guided 3d human-object interaction generation. In CVPR, 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.4, + 0.907, + 0.44 + ], + "angle": 0, + "content": "[19] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.442, + 0.907, + 0.496 + ], + "angle": 0, + "content": "[20] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.498, + 0.908, + 0.551 + ], + "angle": 0, + "content": "[21] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. IMoS: Intent-driven full-body motion synthesis for human-object interactions. arXiv preprint arXiv:2212.07555, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.553, + 0.907, + 0.593 + ], + "angle": 0, + "content": "[22] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.595, + 0.908, + 0.636 + ], + "angle": 0, + "content": "[23] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.637, + 0.907, + 0.677 + ], + "angle": 0, + "content": "[24] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.679, + 0.908, + 0.718 + ], + "angle": 0, + "content": "[25] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael Black. Stochastic scene-aware motion prediction. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.72, + 0.908, + 0.761 + ], + "angle": 0, + "content": "[26] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.763, + 0.908, + 0.802 + ], + "angle": 0, + "content": "[27] Zhi Hou, Baosheng Yu, and Dacheng Tao. Compositional 3d human-object neural animation. arXiv preprint arXiv:2304.14070, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.804, + 0.908, + 0.845 + ], + "angle": 0, + "content": "[28] Yinghao Huang, Omid Taheri, Michael J. Black, and Dimitrios Tzionas. InterCap: Joint markerless 3D tracking of humans and objects in interaction. In GCPR, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.847, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[29] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. CHAIRS: Towards full-body articulated human-object interaction. In ICCV, 2023. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7056" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[30] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.151, + 0.485, + 0.219 + ], + "angle": 0, + "content": "[31] Marilyn Keller, Keenon Werling, Soyong Shin, Scott Delp, Sergi Pujades, C Karen Liu, and Michael J Black. From skin to skeleton: Towards biomechanically accurate 3d digital humans. ACM Transactions on Graphics (TOG), 42(6):1-12, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.222, + 0.484, + 0.277 + ], + "angle": 0, + "content": "[32] Jeonghwan Kim, Jisoo Kim, Jeonghyeon Na, and Hanbyul Joo. ParaHome: Parameterizing everyday home activities towards 3d generative modeling of human-object interactions. arXiv preprint arXiv:2401.10232, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.28, + 0.484, + 0.322 + ], + "angle": 0, + "content": "[33] Taeksoo Kim, Shunsuke Saito, and Hanbyul Joo. NCHO: Unsupervised learning for neural 3d composition of humans and objects. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.324, + 0.484, + 0.365 + ], + "angle": 0, + "content": "[34] Franziska Krebs, Andre Meixner, Isabel Patzer, and Tamim Asfour. The kit bimanual manipulation dataset. In Humanoids, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[35] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. NIFTY: Neural object interaction fields for guided human motion synthesis. arXiv preprint arXiv:2307.07511, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[36] Jiye Lee and Hanbyul Joo. Locomotion-Action-Manipulation: Synthesizing human-scene interactions in complex 3d environments. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.484, + 0.523 + ], + "angle": 0, + "content": "[37] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.527, + 0.482, + 0.569 + ], + "angle": 0, + "content": "[38] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.571, + 0.484, + 0.625 + ], + "angle": 0, + "content": "[39] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. arXiv preprint arXiv:2303.13129, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.629, + 0.482, + 0.683 + ], + "angle": 0, + "content": "[40] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. InterGen: Diffusion-based multi-human motion generation under complex interactions. arXiv preprint arXiv:2304.05684, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.686, + 0.484, + 0.74 + ], + "angle": 0, + "content": "[41] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-X: A large-scale 3d expressive whole-body human motion dataset. In NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.744, + 0.484, + 0.798 + ], + "angle": 0, + "content": "[42] Libin Liu and Jessica Hodgins. Learning basketball dribbling skills using trajectory optimization and deep reinforcement learning. ACM Transactions on Graphics (TOG), 37 (4):1-14, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[43] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[44] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[45] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics, 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.149, + 0.908, + 0.203 + ], + "angle": 0, + "content": "[46] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humanomatopo: Text-aligned whole-body motion generation. arxiv:2310.12978, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.205, + 0.908, + 0.272 + ], + "angle": 0, + "content": "[47] Zhengyi Luo, Jiashun Wang, Kangni Liu, Haotian Zhang, Chen Tessler, Jingbo Wang, Ye Yuan, Jinkun Cao, Zihui Lin, Fengyi Wang, et al. Smpolympics: Sports environments for physically simulated humanoids. arXiv preprint arXiv:2407.00187, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.274, + 0.908, + 0.342 + ], + "angle": 0, + "content": "[48] Xintao Lv, Liang Xu, Yichao Yan, Xin Jin, Congsheng Xu, Shuwen Wu, Yifan Liu, Lincheng Li, Mengxiao Bi, Wenjun Zeng, et al. Himo: A new benchmark for full-body human interacting with multiple objects. In ECCV. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.344, + 0.908, + 0.397 + ], + "angle": 0, + "content": "[49] Junyi Ma, Jingyi Xu, Xieyuanli Chen, and Hesheng Wang. Diff-ip2d: Diffusion-based hand-object interaction prediction on egocentric videos. arXiv preprint arXiv:2405.04370, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.4, + 0.908, + 0.44 + ], + "angle": 0, + "content": "[50] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. AMASS: Archive of motion capture as surface shapes. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.442, + 0.908, + 0.51 + ], + "angle": 0, + "content": "[51] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.511, + 0.908, + 0.551 + ], + "angle": 0, + "content": "[52] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. The kit whole-body human motion database. In ICAR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.553, + 0.908, + 0.62 + ], + "angle": 0, + "content": "[53] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. Unifying representations and large-scale whole-body motion databases for studying human motion. IEEE Transactions on Robotics, 32(4):796-809, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.623, + 0.908, + 0.676 + ], + "angle": 0, + "content": "[54] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3D pose estimation from monocular RGB. In 3DV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.679, + 0.908, + 0.747 + ], + "angle": 0, + "content": "[55] Josh Merel, Saran Tunyasuvunakool, Arun Ahuja, Yuval Tassa, Leonard Hasenclever, Vu Pham, Tom Erez, Greg Wayne, and Nicolas Heess. Catch & carry: reusable neural controllers for vision-guided whole-body tasks. ACM Transactions on Graphics (TOG), 39(4):39-1, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.749, + 0.908, + 0.802 + ], + "angle": 0, + "content": "[56] Gyeongsik Moon, Shouu-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.804, + 0.908, + 0.845 + ], + "angle": 0, + "content": "[57] Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. In NeurIPS, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.846, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[58] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In CVPR, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7057" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.135 + ], + "angle": 0, + "content": "[59] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.484, + 0.163 + ], + "angle": 0, + "content": "[60] OpenAI. ChatGPT. https://chat.openai.com/, 2023.3,1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.165, + 0.484, + 0.22 + ], + "angle": 0, + "content": "[61] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. arXiv preprint arXiv:2308.09036, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.222, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[62] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 3, 4, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.292, + 0.484, + 0.359 + ], + "angle": 0, + "content": "[63] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. HOI-Diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.484, + 0.416 + ], + "angle": 0, + "content": "[64] Ilya A Petrov, Riccardo Marin, Julian Chibane, and Gerard Pons-Moll. Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.419, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[65] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In ECCV, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.462, + 0.484, + 0.503 + ], + "angle": 0, + "content": "[66] Mathis Petrovich, Michael J Black, and Gül Varol. TMR: Text-to-motion retrieval using contrastive 3d human motion synthesis. In ICCV, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.505, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[67] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236–252, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[68] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In ICCV, 2019. 5, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[69] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english labels. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.688 + ], + "angle": 0, + "content": "[70] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[71] Haziq Razali and Yiannis Demiris. Action-conditioned generation of bimanual object manipulation sequences. In AAAI, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[72] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[73] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6), 2017. 3, 4, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[74] Wenfeng Song, Xinyu Zhang, Shuai Li, Yang Gao, Aimin Hao, Xia Hou, Chenglizhao Chen, Ning Li, and Hong Qin. Hoianimator: Generating text-prompt human-object animations using novel perceptive diffusion models. In CVPR, 2024. 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.135 + ], + "angle": 0, + "content": "[75] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.136, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[76] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39(4):54-1, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.191, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[77] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In ECCV, 2020. 2, 3, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.234, + 0.907, + 0.274 + ], + "angle": 0, + "content": "[78] Omid Taheri, Vasileios Choutas, Michael J Black, and Dimitrios Tzionas. GOAL: Generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.275, + 0.907, + 0.328 + ], + "angle": 0, + "content": "[79] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J Black. Grip: Generating interaction poses using spatial cues and latent consistency. In 3DV, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.33, + 0.907, + 0.369 + ], + "angle": 0, + "content": "[80] Purva Tendulkar, Dídac Surís, and Carl Vondrick. Flex: Full-body grasping without full-body grasps. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.372, + 0.907, + 0.426 + ], + "angle": 0, + "content": "[81] Chen Tessler, Yunrong Guo, Ofir Nabati, Gal Chechik, and Xue Bin Peng. Maskedmimic: Unified physics-based character control through masked motion inpainting. arXiv preprint arXiv:2409.14393, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.428, + 0.907, + 0.481 + ], + "angle": 0, + "content": "[82] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.484, + 0.907, + 0.552 + ], + "angle": 0, + "content": "[83] Guy Tevet, Signal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. Closd: Closing the loop between simulation and diffusion for multi-task character control. arXiv preprint arXiv:2410.03441, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.554, + 0.907, + 0.607 + ], + "angle": 0, + "content": "[84] Jie Tian, Lingxiao Yang, Ran Ji, Yuexin Ma, Lan Xu, Jingyi Yu, Ye Shi, and Jingya Wang. Gaze-guided hand-object interaction synthesis: Benchmark and method. arXiv preprint arXiv:2403.16169, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.609, + 0.907, + 0.677 + ], + "angle": 0, + "content": "[85] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.679, + 0.907, + 0.718 + ], + "angle": 0, + "content": "[86] Jiashun Wang, Jessica Hodgins, and Jungdam Won. Strategy and skill learning for physics-based table tennis animation. In SIGGRAPH, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.72, + 0.907, + 0.775 + ], + "angle": 0, + "content": "[87] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In 3DV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.777, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[88] Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. PhysHOI: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393, 2023. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.832, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[89] Yinhuai Wang, Qihan Zhao, Runyi Yu, Ailing Zeng, Jing Lin, Zhengyi Luo, Hok Wai Tsui, Jiwen Yu, Xiu Li, Qifeng Chen, et al. Skillmimic: Learning reusable basketball skills from demonstrations. arXiv preprint arXiv:2408.15270, 2024.3" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7058" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.135 + ], + "angle": 0, + "content": "[90] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. InterControl: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[91] Noah Wiederhold, Ava Megyeri, DiMaggio Paris, Sean Banerjee, and Natasha Banerjee. Hoh: Markerless multimodal human-object-human handover dataset with large object count. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.207, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[92] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In CoRL, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.25, + 0.484, + 0.303 + ], + "angle": 0, + "content": "[93] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. THOR: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.307, + 0.484, + 0.348 + ], + "angle": 0, + "content": "[94] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. SAGA: Stochastic whole-body grasping with contact. In ECCV, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[95] Zhen Wu, Jiaman Li, and C Karen Liu. Human-object interaction from human-level instructions. arXiv preprint arXiv:2406.17840, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.392, + 0.484, + 0.433 + ], + "angle": 0, + "content": "[96] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.435, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[97] Xianghui Xie, Jan Eric Lenssen, and Gerard Pons-Moll. InterTrack: Tracking human object interaction without object templates. arXiv preprint arXiv:2408.13953, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.478, + 0.484, + 0.531 + ], + "angle": 0, + "content": "[98] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. OmniControl: Control any joint at any time for human motion generation. arXiv preprint arXiv:2310.08580, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.484, + 0.575 + ], + "angle": 0, + "content": "[99] Zhaoming Xie, Sebastian Starke, Hung Yu Ling, and Michiel van de Panne. Learning soccer juggling skills with layerwise mixture-of-experts. In SIGGRAPH, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.577, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[100] Zhaoming Xie, Jonathan Tseng, Sebastian Starke, Michiel van de Panne, and C Karen Liu. Hierarchical planning and control for box loco-manipulation. arXiv preprint arXiv:2306.09532, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.633, + 0.484, + 0.699 + ], + "angle": 0, + "content": "[101] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter-x: Towards versatile human-human interaction analysis. arXiv preprint arXiv:2312.16051, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[102] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 1, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[103] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. arXiv preprint arXiv:2403.19652, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[104] Sirui Xu, Hung Yu Ling, Yu-Xiong Wang, and Liang-Yan Gui. InterMimic: Towards universal whole-body control for physics-based human-object interactions. In CVPR, 2025. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[105] Xiang Xu, Hanbyul Joo, Greg Mori, and Manolis Savva. D3D-HOI: Dynamic 3d human-object interactions from videos. arXiv preprint arXiv:2108.08420, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[106] ChangHee Yang, ChanHee Kang, Kyeongbo Kong, Hanni Oh, and Suk-Ju Kang. Person in place: Generating associative skeleton-guidance maps for human-object interaction image editing. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.907, + 0.191 + ], + "angle": 0, + "content": "[107] Jie Yang, Xuesong Niu, Nan Jiang, Ruimao Zhang, and Siyuan Huang. F-HOI: Toward fine-grained semantic-aligned 3d human-object interactions. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.907, + 0.234 + ], + "angle": 0, + "content": "[108] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, and Zheng-Jun Zha. Lemon: Learning 3d human-object interaction relation from 2d images. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.237, + 0.906, + 0.277 + ], + "angle": 0, + "content": "[109] Zeshi Yang, Kangkang Yin, and Libin Liu. Learning to use chopsticks in diverse gripping styles. ACM Transactions on Graphics (TOG), 41(4):1-17, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.28, + 0.907, + 0.333 + ], + "angle": 0, + "content": "[110] Heyuan Yao, Zhenhua Song, Yuyang Zhou, Tenglong Ao, Baoquan Chen, and Libin Liu. Moconvq: Unified physics-based motion control via scalable discrete representations. ACM Transactions on Graphics (TOG), 43(4):1-21, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.336, + 0.907, + 0.39 + ], + "angle": 0, + "content": "[111] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance diffusion: Synthesizing hand-object interactions. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.393, + 0.907, + 0.448 + ], + "angle": 0, + "content": "[112] Xinyu Zhan, Lixin Yang, Yifei Zhao, Kangrui Mao, Hanlin Xu, Zenan Lin, Kailin Li, and Cewu Lu. Oakink2: A dataset of bimanual hands-object manipulation in complex task completion. arXiv preprint arXiv:2403.19417, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.451, + 0.907, + 0.504 + ], + "angle": 0, + "content": "[113] Chengwen Zhang, Yun Liu, Ruofan Xing, Bingda Tang, and Li Yi. Core4d: A 4d human-object-human interaction dataset for collaborative object rearrangement. arXiv preprint arXiv:2406.19353, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.507, + 0.907, + 0.573 + ], + "angle": 0, + "content": "[114] Hui Zhang, Sammy Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. ArtiGrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv preprint arXiv:2309.03891, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.907, + 0.631 + ], + "angle": 0, + "content": "[115] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. NeuralDome: A neural modeling pipeline on multi-view humanobject interactions. In CVPR, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.634, + 0.907, + 0.701 + ], + "angle": 0, + "content": "[116] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2M-GPT: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.705, + 0.907, + 0.771 + ], + "angle": 0, + "content": "[117] Juze Zhang, Jingyan Zhang, Zining Song, Zhanhe Shi, Chengfeng Zhao, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Hoi-m^3: Capture multiple humans and objects interaction within contextual environment. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.907, + 0.843 + ], + "angle": 0, + "content": "[118] Jiajun Zhang, Yuxiang Zhang, Liang An, Mengcheng Li, Hongwen Zhang, Zonghai Hu, and Yebin Liu. Manidext: Hand-object manipulation synthesis via continuous correspondence embeddings and residual-guided diffusion. arXiv preprint arXiv:2409.09300, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.847, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[119] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In ECCV, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7059" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[120] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. COUCH: Towards controllable human-chair interactions. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[121] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Ilya Petrov, Vladimir Guzov, Helisa Dhamo, Eduardo Pérez-Pellitero, and Gerard Pons-Moll. FORCE: Dataset and method for intuitive physics guided human-object interaction. arXiv preprint arXiv:2403.11237, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[122] Yunbo Zhang, Deepak Gopinath, Yuting Ye, Jessica Hodgins, Greg Turk, and Jungdam Won. Simulation and retargeting of complex multi-character interactions. In SIGGRAPH, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[123] Chengfeng Zhao, Juze Zhang, Jiashen Du, Ziwei Shan, Junye Wang, Jingyi Yu, Jingya Wang, and Lan Xu. I'M HOI: Inertia-aware monocular capture of 3d human-object interactions. In CVPR, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[124] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[125] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.484, + 0.459 + ], + "angle": 0, + "content": "[126] Juntian Zheng, Qingyuan Zheng, Lixing Fang, Yun Liu, and Li Yi. CAMS: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[127] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In ECCV. Springer, 2022. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7060" + } + ] +] \ No newline at end of file diff --git a/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/4511c287-17f9-4b70-a6d5-d77c48a5192a_origin.pdf b/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/4511c287-17f9-4b70-a6d5-d77c48a5192a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..aac326ec311f27e69abcb3947a62ba76418a0d0b --- /dev/null +++ b/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/4511c287-17f9-4b70-a6d5-d77c48a5192a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f0dca160221652c6515c07d0510ac61b87b28dcd60db5ea1469c376bbc5f187 +size 7981140 diff --git a/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/full.md b/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..59f28e29d995bdda33b16133658e6049f2c652fa --- /dev/null +++ b/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/full.md @@ -0,0 +1,360 @@ +# InterAct: Advancing Large-Scale Versatile 3D Human-Object Interaction Generation + +![](images/ee9ab7a9c8f7ee5ca8269e301a95769d9c61b2836b596fb89c0929f95e464738.jpg) + +![](images/7ebbdf565494c4000173240ed91cfb3804bece795c306c69fdf547df490d31b7.jpg) +Figure 1. An overview of InterAct, our large-scale 3D human-object interaction (HOI) benchmark, covering six HOI generation tasks. + +# Abstract + +While large-scale human motion capture datasets have advanced human motion generation, modeling and generating dynamic 3D human-object interactions (HOIs) remain challenging due to dataset limitations. Existing datasets often lack extensive, high-quality motion and annotation and exhibit artifacts such as contact penetration, floating, and incorrect hand motions. To address these issues, we introduce InterAct, a large-scale 3D HOI benchmark featuring dataset and methodological advancements. First, we consolidate and standardize 21.81 hours of HOI data from diverse sources, enriching it with detailed textual annotations. Second, we propose a unified optimization framework to enhance data quality by reducing artifacts and correcting hand motions. Leveraging the principle of contact invariance, we maintain human-object relationships while introducing motion variations, expanding the dataset to 30.70 hours. Third, we define six benchmarking tasks and develop a unified HOI generative modeling perspective, achieving state-of-the-art + +performance. Extensive experiments validate the utility of our dataset as a foundational resource for advancing 3D human-object interaction generation. The dataset will be publicly accessible to support further research in the field. + +# 1. Introduction + +Recent advances in human motion modeling have significantly benefited from extensive motion capture (MoCap) datasets [22, 41, 50, 67, 69], enabling the creation of scalable generative models for diverse human movements. Building upon this foundation, researchers are increasingly turning to the more intricate challenge of generating human-object interactions (HOIs) [102-104]. This emerging area holds considerable promise for applications in robotics, animation, and computer vision. + +However, high-quality HOI generation faces notable obstacles due to factors such as increased degrees of freedom introduced by objects, varied object geometries, dynamic in + +
DatasetClipHourTextHandObject
GRAB [77]1,3353.76X✓51
BEHAVE [4]2994.13XX18
InterCap [28]2330.62X✓10
Chairs [29]1,0412.37X✓92
HODome [115]1762.82X✓21
OMOMO [38]4,8388.274,838X15
IMHD [123]1640.97X✓10
InterAct (Ours)11,35021.8134,050✓217
InterAct-X (Ours)16,20130.7048,630✓217
+ +Table 1. Comparison between InterAct, InterAct-X, and human-object interaction datasets we collect. Beyond a substantially larger scale, our dataset introduces comprehensive textual annotations and enhances interaction quality, offering a more versatile foundation for large-scale HOI generation. + +teractions, and the necessity for physically accurate contact modeling. Current methods often struggle to achieve realism primarily because existing datasets lack scalability and comprehensive annotations, which are crucial for models to effectively understand interaction dynamics and link them to related domains such as natural language. + +Specifically, these challenges underscore the need for comprehensive, high-quality HOI datasets: (1) Limited and Inconsistent Datasets: Existing methods typically depend on small datasets with limited hours of data, difficult to consolidate due to inconsistent human representations, object types, coordinate systems, and annotations. Available annotations [38, 63] are frequently coarse and incomplete, lacking detailed descriptions of human states, object interactions, and involved body parts. (2) Prevalent Artifacts: Current datasets often contain artifacts from MoCap limitations and occlusions, including unnatural penetrations, floating contacts, inaccurate hand poses [4, 38], and significant motion jitter [28]. These issues compromise the models' capacity to learn realistic human-object dynamics. + +To address these challenges, we present InterAct, a benchmark designed to systematically overcome current limitations and drive advancements in 3D HOI modeling. As shown in Table 1, InterAct offers a large-scale, standardized dataset of carefully curated interactions from existing resources1, enriched by detailed textual annotations. + +To further enhance dataset quality and scope, we introduce a unified optimization approach, addressing major penetration and floating artifacts first in whole-body interactions, followed by refined corrections for nuanced hand-object interactions. Additionally, we propose the concept of contact invariance, inspired by motion mirroring techniques, to generate realistic synthetic data by varying human motions while + +![](images/2dcd34ca991cba8ca1dbdeb3209270981a842cf323ec4e95ba21022e4f047a6f.jpg) +Figure 2. (a) Our data processing pipeline consolidating data, annotations via foundation models, corrections, and interaction illustrations. (b) Statistics on motion and text annotations. + +maintaining consistent object contacts. This augmentation expands InterAct into InterAct-X, providing approximately 9 additional hours of data and substantially improving generative model performance. + +Leveraging this comprehensive, richly annotated dataset, we define benchmarks across six key HOI generation tasks, Text-to-Interaction, Action-to-Interaction, Object-to-Human, Human-to-Object, Interaction Prediction, and Interaction Imitation, as shown in Figure 1, and propose a unified modeling and representation for kinematic generative tasks. Our method utilizes multi-task learning to jointly model motion and contact, achieving state-of-the-art performance as validated by comprehensive evaluations. + +In summary, our contributions are: (1) InterAct, the most extensive 3D HOI benchmark to date, facilitating large-scale generative modeling. (2) A unified optimization-based framework to correct and augment MoCap data, addressing common artifacts and significantly enhancing dataset quality. We believe this will aid future research in overcoming data scarcity before capturing potentially imperfect data. (3) Comprehensive benchmarks across six HOI generation tasks, establishing standardized metrics and demonstrating superior performance over existing approaches. This benchmark lays a strong foundation for future research, encouraging advancements across multiple facets of 3D HOI generation. + +# 2. Related Work + +Dynamic 3D HOI Dataset. Many large-scale datasets with sequential human motion data have established benchmarks for the task of 3D human motion generation. However, human actions are influenced not only by individual intents but also by interactions with the surrounding environment. To address this complexity, new datasets have been developed to capture the dynamics between humans and their environments, including interactions with other humans [1, 40, 54, 101] and scenes [8, 24, 25]. Though there are fruitful hand-object interaction datasets [11, 23, 44, 48, 56, 58, 80, 91, 91, 112, 113, 121], our focus is specifically on whole-body interactions with dynamic objects, ranging from + +low-dynamic interactions such as approaching and manipulation [77] to highly dynamic interactions involving multiple body parts [4, 20, 28-30, 32, 38, 107, 115, 123]. + +We aim to address the limitations of these datasets and open new possibilities for future research. Our InterAct dataset maintains advantages in motion quality, fine-grained textual annotations, detailed hand gestures, and comprehensive annotation modalities. We provide quantitative comparisons of InterAct and existing datasets in Table 1, demonstrating the superiority of our dataset in these aspects. + +Dynamic 3D HOI Generation. Existing human-object interaction (HOI) datasets have laid a robust foundation for generating dynamic, whole-body interactions. Extensive research has explored the generation of hand-object interactions [9, 13, 39, 43, 49, 84, 111, 114, 118, 126, 127] and static human-object interactions [27, 33, 64, 87, 96, 97, 106, 108, 119, 124]. Meanwhile, full-body dynamic interactions have also been studied extensively [14, 21, 34-36, 38, 52, 53, 71, 75, 76, 78, 85, 94, 105, 120, 125], though these often face significant limitations, including narrow action repertoires and dependence on static objects. Recent advancements, such as InterDiff [102], have introduced diverse interactions involving dynamic objects and multiple body parts. Building upon this, subsequent approaches like InterDreamer [103] and other contemporary studies [18, 37, 63, 74, 93, 95, 103, 117] further demonstrate the feasibility of converting textual descriptions into realistic 3D human-object interaction sequences. Despite these advances, current methods remain constrained by a shortage of high-quality, large-scale datasets, often encountering issues related to physical inaccuracies, such as floating contacts or interpenetration. In parallel, physics-based methods leveraging deep reinforcement learning (RL) [3, 6, 10, 15, 26, 42, 55, 61, 81, 83, 86, 89, 90, 92, 98, 100, 109, 110] successfully generate physically accurate interactions, with applications in sports [47] like basketball [88, 89] and soccer [99]. Nonetheless, these methods typically produce rigid interaction patterns from limited datasets, while a recent work, InterMimic [104], illustrates that physics-based approaches can digest effectively across diverse and dynamic object interactions. Our work addresses these fundamental limitations at the dataset level by providing enhanced diversity and comprehensive sequences of human-object interactions. This facilitates multiple generative tasks, supports better contact modeling, and improves the capability to synthesize realistic and generalized human-object interactions. + +# 3. InterAct Dataset + +Overview. We introduce InterAct, the first unified benchmark tailored explicitly for sequential 3D human-object interaction (HOI) generative modeling. Distinguished by its unprecedented scale and comprehensiveness, InterAct significantly surpasses existing datasets, as summarized in Table 1. + +![](images/3402e9787fb5232f3e63ae47ab08ab60b7b23f220afe834df2b23d3b4d539b96.jpg) +(a) Markers in SMPL-H +Figure 3. Marker-based representation for human. + +![](images/4b0e57fb1ec483f4a17a14b96fa900f73192d6e0f30a08ebe44b45703a1b009c.jpg) +(b) Markers in SMPL-X + +InterAct is available in two versions: (1) A basic version consolidating seven existing datasets, providing 21.81 hours of annotated 3D whole-body interactions with corresponding semantic descriptions. (2) An advanced version, InterAct-X, extending the basic dataset through synthetic data generated via our unified optimization framework. Figure 2 presents examples of motion and text annotations. To ensure high-quality standards, we employ a multifaceted annotation strategy combining human expertise, automated foundation models, and advanced HOI modeling techniques, all validated through rigorous manual quality checks. + +# 3.1. Data Collection, Annotation, and Unification + +We compile data from seven datasets [4, 28, 29, 38, 77, 115, 123], featuring motion capture of a single human interacting with a single dynamic 3D object, where humans are annotated with SMPL [45, 62, 73]. We address their heterogeneity in two key aspects: annotations and representations. + +Unifying Textual Annotations. Since most datasets either lack textual descriptions or provide only very coarse text descriptions [38], we implement a two-phase annotation procedure involving human annotators and GPT-4 [60] to generate consistent and detailed annotations across all subsets. In the first phase, human annotators provided detailed and precise descriptions of the interactions, adhering to the following guidelines: (i) Split motion sequences into clips averaging 300 frames (approximately 10 seconds) but no longer than 400 frames each; (ii) Clearly describe the actions and the body parts involved in the interactions. For example, a typical annotation is: "A person sits on a stool and touches the ground with their left hand, then their right hand." For the subset derived from OMOMO [38], we skip this phase and directly utilize their annotations. In the second phase, we use GPT-4 to rephrase and simplify human annotations to enhance diversity and consistency. For example, the rephrased version is "A person perches on a stool, touching the ground with their left hand, then their right hand," and the simplified version is "A person sits on a stool, touching the ground with each hand alternately." Next, we employ GPT-4 to classify each description into one of our predefined 15 action labels with in-context learning [7]. The action label for the above sequence is "Sit." We meticulously review all generated texts and action labels to ensure high quality and alignment across the dataset. + +Unifying Human Representations. Different datasets employ varying human models (e.g., SMPL-H [73], SMPL-X [62]) and diverse shapes. A straightforward solution can be to convert different humans from SMPL-H and SMPL-X to a consistent SMPL version and encode shape parameters into the generative modeling. However, although SMPL is widely used in various human-related tasks, it is fundamentally a rotation-based representation. In the context of human motion generation, Cartesian features like joint positions and velocities are more commonly used, as seen in the integration with the HumanML3D representation [22] and in most text-to-motion work [65, 82, 116]. This is still suboptimal because joints are located beneath the body's surface and do not explicitly participate in interactions. To overcome the limitations, we use markers - specific sets of human vertices representing human motion and interactions - as a simple and unified representation capable of effectively inferring contact, evaluated in Table 5. Similar approaches are discussed in [94, 102]. Then we need to select a marker set that is consistent between SMPL-H and SMPL-X models. + +Given two human body models, SMPL-H and SMPL-X, sharing the same shape, we establish marker correspondences in two steps. First, we index the markers on the SMPL-H surface as defined in prior work [94, 102]. Second, we locate the corresponding vertices on SMPL-X by selecting the closest points to these SMPL-H markers, leveraging the official SMPL conversion, which maps each SMPL-H vertex to the nearest point on the SMPL-X mesh. We extensively evaluated the approximation error of these marker correspondences across a broad range of poses. Our results show that the maximum error consistently remains below $1\mathrm{cm}$ , a deviation unlikely to affect the overall performance of HOI generation. This high consistency arises because the markers are rigidly attached to the body, and soft deformations are disabled. As a result, the identical rigid transformations in SMPL-H and SMPL-X preserve the correspondence of the markers. Additional details on such correspondence-preserving conditions can be found in [31]. Figure 3 illustrates the marker sets for SMPL-H and SMPL-X. We use this marker-based representation to train the generative models for the tasks outlined in Sec. 4, while still relying on the original SMPL-H or SMPL-X representations for the interaction correction and augmentation methods described in Sec. 3.2. + +# 3.2. Interaction Correction and Augmentation + +In this section, we present a unified optimization framework that addresses both the correction of MoCap artifacts and the augmentation of the dataset by introducing more synthetic data. The process takes as input the motions and geometries of humans and objects, then compares them against predefined standards to define loss functions. Using gradient-based optimization, we iteratively adjust human and object + +motions to minimize these losses, thereby refining the data to meet the desired quality criteria. The key challenge lies in formulating learning objectives that not only rectify existing data but also facilitate the generation of new synthetic data. + +Our optimization is carried out in three sequential steps: (1) full-body correction; (2) hand correction; and (3) interaction augmentation. Hand correction is handled separately because, although hand poses occupy considerable space in the SMPL representation, they contribute relatively little to the overall scale of the learning objectives. By decoupling hand correction from full-body correction, we can better balance these two processes and define more targeted objectives. In what follows, we first introduce the hand correction stage. Hand Correction. Given that many existing datasets contain inaccurate hand poses [4, 38], our approach selectively promotes contact only in regions where ground-truth data indicates hand-object interaction, while ensuring the hand motion remains natural, in spirit to InterMimic [104] but relying on predefined optimization instead of RL. This approach is effective for the whole-body interaction datasets we utilize, which generally do not require high dexterity and typically only involve the hand conforming to the object for grasping, as a common assumption in existing work [79, 127], while we distinguish our approach from those that employ multi-stage, learning-based methods for the same purpose. + +We divide our hand correction objectives into two categories: contact promotion and hand constraints. Contact promotion is guided by the following contact loss: + +$$ +E _ {\mathrm {c o n t}} = \sum_ {i = 1} ^ {L} c _ {i} \sum_ {j} d _ {j} [ i ], +$$ + +where $d_{j}[i]$ is the distance between the $j$ -th hand vertex and its nearest point on the object's surface at the $i$ -th frame, and $c_{i}$ indicates whether the object and the hand are in contact at frame $i$ . The contact indicator $c_{i}$ , inferred from ground truth data, is a function based on hand-object distance $\min_j d_j[i]$ , which we provide details in supplementary. The hand constraint objectives are introduced to preserve naturalness and temporal smoothness in the hand motions. These constraints include: (1) penetration loss, which penalizes intersections between the hand and the object. (2) smoothness loss, which promotes consistent contact and reduces jittering. (3) prior loss, which constrains the range of motion (RoM) of the fingers to maintain realism. Without this constraint, contact promotion could inadvertently drive fingers into biologically impossible poses. Detailed formulations of these loss functions are provided in the supplementary. + +Full-Body Correction. In this stage, all human and object poses can be updated via gradient descent. We add a reconstruction loss to ensure that the optimized interactions closely match the ground truth. Other losses mirror those used in hand correction, with two key differences: (1) Contact and penetration losses are computed for the entire body + +rather than just the hands. (2) Prior loss is omitted because the reconstruction loss alone suffices to maintain plausible human motion. Detailed formulations of these losses are provided in supplementary. + +Interaction Augmentation. Synthetic data has become increasingly important in computer vision and generative modeling [5, 12, 19, 57], prompting a key question in the context of HOI animation: Can we scale up datasets without collecting additional MoCap data? Does existing interaction data offer information beyond its observable motions? Consider a scenario where a person grasps a box and walks, as illustrated in Figure 5. Even if their gait changes slightly, hand-box contact should remain consistent to preserve the semantics of the interaction. This illustrates the principle of interaction invariance: the core interaction persists despite minor variations in motion. Leveraging this principle, we can augment our dataset by injecting new human motions while preserving consistent object interactions. Training neural networks on such augmented data enables them to naturally learn this invariance, a common strategy in symmetric learning [17], and ultimately enhances model performance. + +Our augmentation pipeline consists of three steps: (1) Object Displacement: We apply a random displacement to the object's trajectory, uniformly across all timesteps. (2) Interaction Alignment: We optimize the human motion to maintain interaction with the displaced object, using both contact consistency and reconstruction objectives. (3) Interaction Filtering: We remove low-quality augmentations, those with unreasonable initial displacements, significant penetrations (human-object or self-penetration), or alignment failures indicated by high optimization losses (e.g., excessive jitter). + +During the alignment phase, the primary objective is the contact consistency loss. We first compute a distance matrix $\mathbf{D}$ , where each element $\mathbf{D}_{jk} = \|\mathbf{v}_h^j - \mathbf{v}_o^k\|$ denotes the Euclidean distance between the $j$ -th human vertex and the $k$ -th object vertex, before or after displacement. With a reference matrix $\hat{\mathbf{D}}$ from the original (pre-displacement) setup, we optimize the human motion using: + +$$ +E _ {\mathrm {a l i g n}} = \sum_ {i = 1} ^ {L} \sum_ {j, k} \frac {1}{(\hat {\mathbf {D}} _ {j k} + \epsilon) ^ {2}} \left| \hat {\mathbf {D}} _ {j k} - \mathbf {D} _ {j k} \right| ^ {2}, +$$ + +where $\epsilon$ is a small constant to prevent division by zero. This formulation preserves distances between vertex pairs that were initially close, while de-emphasizing pairs that were farther apart. Additional terms in the objective enforce naturalness and stability in non-interactive regions of the human pose, as detailed in the supplementary material. + +# 4. Tasks and Methods + +In this section, we formally define six distinct tasks featured in our benchmark. We use unified representation across + +five kinematic generative tasks, where each human-object interaction sequence is represented as $\langle h,o\rangle$ , annotated with an action category $a$ and a text description $t$ . The human $h$ includes marker coordinates, marker velocities, signed distance vectors from each marker to the object, and foot-ground contact labels. The object $o$ represents object motion, including object rotation angles, object translations. Object geometry is described by Basis Point Set (BPS) [68]. + +(1) Text-Conditioned Interaction Generation. Initially in [18, 63], the task learns a function to generate the interaction sequence based on text: $\mathcal{G}_{\mathrm{t2i}}(t)\mapsto \langle h,o\rangle$ + +(2) Action-Conditioned Interaction Generation. The objective is to learn a function that maps an action label to the corresponding interaction sequence: $\mathcal{G}_{\mathrm{a2i}}(\pmb {a})\mapsto \langle \pmb {h},\pmb {o}\rangle$ + +(3) Object-Conditioned Human Generation. Initially in [38], the task generates human motion based on object sequences through a function $\mathcal{G}_{\mathrm{o2h}}(\pmb {o})\mapsto \pmb{h}$ . + +(4) Human-Conditioned Object Generation. Conversely, this task focuses on generating object motion sequences from human motion sequences via a function $\mathcal{G}_{\mathrm{h2o}}(\pmb {h})\mapsto \pmb{o}$ . + +(5) Interaction Prediction. Initially in [102], the task aims to predict future human-object interactions based on past. Let $\langle \pmb{h}_p, \pmb{o}_p \rangle$ denote the past interaction and $\langle \pmb{h}_f, \pmb{o}_f \rangle$ the future. The goal is to learn: $\mathcal{G}_{\mathrm{p2f}}(\langle \pmb{h}_p, \pmb{o}_p \rangle) \mapsto \langle \pmb{h}_f, \pmb{o}_f \rangle$ . + +(6) Interaction Imitation. Following [3, 88, 122], this task focuses on learning physics-based control policies to reproduce human-object interactions in a physics simulator. The output is an action sequence $\pmb{f}$ , specified as joint Proportional-Derivative (PD) targets. The goal is to learn a function $\mathcal{G}_{\mathrm{i2f}}$ that maps the reference interaction sequences to the PD actuation sequences: $\mathcal{G}_{\mathrm{i2f}}(\langle h,o\rangle)\mapsto f$ . + +Unifying Multi-Task HOI Generation. We introduce an additional feature, $\eta$ , which encodes human-object relationships through vectors extending from each human marker to its nearest point on the object's surface. The specific configuration of $\eta$ for each generative task is described in Sec. 5. Using this feature, we can unify first five kinematic generative tasks into a multi-task learning framework by treating $\eta$ as an additional output. For example, we redefine the text-conditioned interaction task as $\mathcal{G}_{\mathrm{t2i}}(t) \mapsto \langle h, o, \eta \rangle$ , where $\mathcal{G}$ is a transformer-based diffusion model. This formulation compels the model to learn spatial relationships inherent to the interactions. In our experiments, we observe that this simple strategy, enhanced by large-scale data, consistently outperforms existing methods. Similar ideas are explored in [18, 37, 38, 63, 74, 93, 102]. + +# 5. Experiments + +We begin by evaluating the effectiveness of our data correction and augmentation methods. Following this, we benchmark existing work and our proposed method on the tasks using our dataset. We standardize the evaluation metrics and present extensive results, including ablation studies. We in + +![](images/d89e67ccd06be2a8a616659f932b4ce6f45b3376d1003fdc4854b5639a6d8d6e.jpg) +Figure 4. Qualitative evaluation of interaction correction (bottom) on the OMOMO [38] dataset shows hand recovery compared to the ground truth interaction (top). Zoom in to see details of the hand recovery. + +
DatasetCorrectionAugmentationPene (m)↓Cont RatioUser Study (%)
BEHAVE [4]××0.0170.04822.3
✓×0.0160.07139.7
✓✓0.0160.06938.0
OMOMO [38]××0.0090.07123.9
✓×0.0070.13139.4
✓✓0.0110.13736.7
+ +Table 2. Quantitative evaluation and user study on the quality of data from interaction correction and augmentation. + +clude additional implementation details, such as the train-test split, in supplementary. + +# 5.1. Correction and Augmentation + +Metrics. We use the following two metrics: Penetration refers to the intersection depth – maximum of negative sign distances from human vertices to the object's surface – average across the sequence. Contact Ratio represents the average ratio of human vertices where their distances to object are under a threshold. + +Quantitative Evaluations. Table 2 shows that our correction process significantly improves the quality of the original MoCap data by enhancing human-object contact and reducing penetration artifacts. Moreover, the quality of the augmented data is comparable to that of the corrected data and exceeds the quality of the original dataset. + +Qualitative Evaluations. Recognizing that quantitative metrics may not fully capture data quality, we conducted a double-blind user study. We randomly selected sequences from each of the raw, corrected, and augmented data for the subset from BEHAV [4] and OMOMO [38] datasets. Human judges were presented with 30 tuple of interactions and asked to rank the quality of three sequences. According to Table 2, over $39\%$ of judges select the corrected data as having the highest quality, significantly outperforming the original data. This confirms that our correction process effectively enhances data realism. Moreover, the augmented data receive ratings comparable to the corrected data, indicating + +that our synthetic data is of high quality. In Figure 4, we visualize our correction results. Despite the original data lacking detailed hand information, we successfully recover vivid and accurate hand interactions. Figure 5 showcases our augmentation, which introduces new high-quality synthetic data while maintaining consistent contact in interactions. + +# 5.2. Language Conditioned HOI Generation + +Metrics. Following the literature on text-to-motion generation [22], we develop five metrics for evaluation. The Fréchet Inception Distance (FID) quantifies the similarity between generated HOI features and the ground truth. The Multimodality and Diversity metrics assess the variety within the generated HOI. R-Precision measures the alignment between the textual descriptions and the generated HOI. The Multimodal Distance (MM Dist) evaluates the disparity between HOI features and corresponding text features. To obtain human-object interaction (HOI) and text features for calculating these metrics, existing methods often train their feature extractors on very limited data [18, 63, 74, 93], which can degrade the quality of the evaluation. We address this limitation by incorporating our larger-scale data with marker and BPS representations. Instead of formulating a classification task to train the feature extractor [22], we follow [46, 66] and employ sequence-level contrastive learning with an InfoNCE loss [59] to train a text encoder and an HOI encoder, integrating Sentence-BERT [72] into the text encoder. + +Baselines and Implementation Details. We adopt HOI-Diff [63] as our base model because it is the only publicly available option compatible with our requirements. For example, CHOIS [37] requires additional conditions beyond text input. HOI-Diff utilizes a transformer-based diffusion model [82] as the backbone, and integrates an affordable model as the classifier guidance [16]. We develop several baseline variants towards our final method by implementing three key modifications: (i) Text Encoder: We replace HOI-Diff's CLIP-based text encoder, where the latent space is not structured for human-object interactions, with our pre + +![](images/29409087b827a1c9874414c9ae315a5a2b83ac7e362ff30c3bae9bd0b77c2c17.jpg) +Figure 5. Qualitative evaluation of interaction augmentation (bottom) shows high-quality synthetic data varied from original (top). + +
HOI-Aware Object Enc.HOI-Aware Text Enc.Contact GenerationContact GuidanceR-Precision†FID↓MM Dist↓Multimodality↑Diversity→
Top 1Top 2Top 3
Ground Truth0.852±0.0000.966±0.0010.989±0.0010.000±0.0002.810±0.002-11.489±0.011
XXXX0.733±0.0070.909±0.0020.957±0.0023.192±0.1914.950±0.0233.149±0.45211.192±0.019
XX✓X0.730±0.0070.913±0.0040.958±0.0051.997±0.0924.752±0.0654.171±0.02711.501±0.037
✓X✓X0.737±0.0110.912±0.0020.963±0.0081.837±0.1264.631±0.0782.836±0.58311.369±0.096
✓✓✓X0.784±0.0040.940±0.0020.980±0.0031.570±0.1394.414±0.0642.677±0.56211.409±0.005
✓✓✓✓0.784±0.0040.940±0.0000.977±0.0021.567±0.1444.412±0.0653.842±0.00511.518±0.178
+ +Table 3. Quantitative evaluation on the task of text-conditioned interaction generation. A batch size of 64 is used for R-Precision. + +
MethodFID↓Multimodality↑Diversity→
Ground Truth0.000±0.000-11.489±0.011
HOI-Diff [63]3.566±0.0985.321±0.14310.989±0.112
Ours2.161±0.0375.792±0.05911.291±0.261
+ +trained interaction-aware text encoder. (ii) Object Shape Encoding: We substitute the original PointNet++ [70], also not pretrained for capturing interaction, with BPS [68] representations for encoding object shapes. (iii) Instead of using affordance as guidance, we regress the contact representation $\eta$ through a multi-task learning. (iv) We further incorporate the contact prediction as classifier guidance [16] within the denoising process, which we detail in supplementary. + +Quantitative Results. Table 3 presents the evaluation results for the text-conditioned interaction generation task. We assess the impact of four design choices introduced above. Notably, incorporating contact modeling and BPS encoding significantly improves the quality of generated HOI, substantially enhancing the FID score. Furthermore, using our interaction-aware text and HOI encoder enhances the quality of the generated interactions and improves the alignment between the generated results and the text input. Lastly, classifier guidance provides a slight overall performance improve + +Table 4. Quantitative evaluation on the task of action-conditioned interaction generation on the entire InterAct testset. + +
RepresentationPene (m)↓Cont Ratio
SMPL0.0300.025
Joint0.0270.032
Marker0.0250.028
+ +Table 5. Ablation study on different representation for text-to-interaction task, evaluated under BEHAVE [4] subset. + +
ModelMPMPE↓FS↓Cprec↑Crec↑Cacc↑F1 Score↑
2-stage [38]36.500.270.810.850.770.80
1-stage36.940.290.840.820.800.81
Ours (Disc.)36.950.280.850.840.830.82
Ours (Cont.)35.690.280.850.890.850.85
+ +Table 6. Quantitative evaluation on object-conditioned human motion generation, with novel objects unseen from training. + +ment. Table 4 benchmarks the performance improvements of these design choices similarly for the action-conditioned interaction generation task. + +Effectiveness of Marker-Based Representation. As shown in Table 5, we compare different human representations on text-to-interaction generation. Without altering contact modeling, marker-based representation produces interactions with fewer artifacts compared to other representations. + +# 5.3. HOI Inpainting + +Metrics. Following OMOMO [38], we develop six metrics tailored for evaluating our marker-based representation. The Mean Per-Marker Position Error (MPMPE) are used to measure the similarity between the generated marker motion and the ground truth. Foot Sliding (FS) is employed to assess the skating effect, reflecting the plausibility of the motion. Additionally, we use a set of contact metrics, including precision $(C_{\mathrm{prec}})$ , recall $(C_{\mathrm{rec}})$ , accuracy $(C_{\mathrm{acc}})$ , and the F1 Score, to evaluate the quality of human-object contact compared to the ground truth. For the human-conditioned object generation task, we include $T_{\mathrm{err}}$ and $O_{\mathrm{err}}$ , which measure discrepancies in object translation and orientation between the generated and the ground truth. + +Baselines and Implementation Details. We adopt OMOMO [38] as our base model since it is the only publicly available option. OMOMO employs a two-stage generation process: it first generates hand motions and then uses these to guide full-body generations. This strategy is particularly effective for the OMOMO dataset, where interactions primarily involve hand contact. However, it is less effective when applied to our InterAct data, which features more versatile whole-body engagement. Motivated by this, we use a single-stage pipeline with multi-task learning, as introduced in Sec. 4. We investigate two different choices for contact regression: $\eta_{\mathrm{Cont.}}$ , which encodes the nearest vector and distance between human markers and the object. $\eta_{\mathrm{Disc.}}$ , which encodes the contact labels for each marker. + +Quantitative Results. Tables 6 and 7 illustrate the effectiveness of our single-stage pipeline, with notable performance improvements when incorporating multi-task modeling. These results provide additional evidence, complementing the evaluation of text-conditioned generation, that multitask learning significantly enhances model performance. + +# 5.4. Interaction Prediction + +Evaluation Metrics. We compare the generated poses to the ground truth motion data using MPMPE (mean perimeter errors), measured in meters. Second, we assess object motion accuracy using Trans.Err., the average $l_{2}$ distance between the predicted and ground truth object translations, and Rot.Err., the average $l_{1}$ distance between the predicted and ground truth object quaternions, following [102]. + +Implementation Details. We adapt InterDiff [102] to utilize a marker-based representation and evaluate it at various scales to assess whether it benefits from scaling laws. + +# 5.5. Interaction Imitation + +Implementation Details. We use PhysHOI [88] to imitate sequences from our InterAct dataset, selecting four sequences shown in Figure 6. The IsaacGym [51] simulator is used, following the same architecture, reward, and representation design as PhysHOI. Training for each sequence, + +![](images/e350045dfc5cb0e9e37f275fc81131d1e67cd8a444090bbbe9321483a08eb200.jpg) +Figure 6. Qualitative results demonstrate the successful imitation of our corrected data using PhysHOI [88]. + +
ModelTerr↓Oerr↓Cprec↑Crec↑Cacc↑F1 Score↑
1-stage25.920.910.830.660.720.69
Ours (multi-task)23.980.830.840.680.740.72
+ +Table 7. Quantitative evaluation on human-to-object. + +
Training dataModel sizeGlobal MPMPE↓Local MPMPE↓Trans. Err.↓\( \text{Rot.Err.}^{\downarrow } \)
BEHAV×10.1200.1030.1330.352
×20.1050.0920.1090.312
×30.1130.1000.1180.343
InterAct (Ours)×10.1060.0950.1060.297
×20.0940.0830.1030.286
×30.0910.0790.0940.264
+ +Table 8. Quantitative evaluation on interaction prediction. + +with separate evaluations for both raw and corrected data, is performed on a single NVIDIA A40 GPU over the course of one day. + +Quantitative Evaluation. In addition to Figure 6, which presents the qualitative results, we evaluate the imitation policy by training on both corrected and raw data, reporting the success rate as defined in [88]. Evaluating four examples from Figure 6 and averaging the success rates over 2048 environments, training on the corrected data achieves a success rate of $90.7\%$ , surpassing the $84.4\%$ achieved with raw data. This demonstrates that our interaction correction provides better data for the motion imitation task. + +Quantitative Results. Table 8 demonstrates that our dataset, with its larger volume of data, supports the improved performance of the trained model with larger scale. In contrast, training the model on limited data leads to overfitting. + +# 6. Conclusion + +We introduce InterAct, a large-scale 3D whole-body human-object interaction benchmark. We employ a unified optimization framework that performs interaction correction and augmentation, enhancing data quality and augment the dataset with synthetic data, scaling to 30.70 hours of interactions and 48,630 textual descriptions. We introduce a simple yet effective multi-task learning approach for unified HOI modeling, enabling models to be trained more effectively across multiple tasks. Our comprehensive experiments highlight the significant advantages of InterAct and our methodologies, resulting in more expressive interaction generation. + +Acknowledgments. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Amazon-Illinois Center on AI for Interactive Conversational Experiences, the Toyota Research Institute, the IBM-Illinois Discovery Accelerator Institute, and Snap Inc. This work used computational resources, including the NCSA Delta and DeltaAI and the PTI Jetstream2 supercomputers through allocations CIS230012, CIS230013, and CIS240311 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, as well as the TACC Frontera supercomputer, Amazon Web Services (AWS), and OpenAI API through the National Artificial Intelligence Research Resource (NAIRR) Pilot. + +# References + +[1] CMU graphics lab motion capture database. http://mocap.cs.cmu.edu/.2 +[2] Easymocap - make human motion capture easier. Github, 2021. 2 +[3] Jinseok Bae, Jungdam Won, Donggeun Lim, Cheol-Hui Min, and Young Min Kim. Pmp: Learning to physically interact with environments using part-wise motion priors. In SIGGRAPH, 2023. 3, 5 +[4] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAV: Dataset and method for tracking human object interactions. In CVPR, 2022. 2, 3, 4, 6, 7, 5 +[5] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, 2023. 5 +[6] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. arXiv preprint arXiv:2309.07907, 2023. 3 +[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 3 +[8] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In ECCV, 2020. 2 +[9] Junuk Cha, Jihyeon Kim, Jae Shin Yoon, and Seungryul Baek. Text2hoi: Text-guided 3d motion generation for hand-object interaction. In CVPR, 2024. 3 +[10] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In AAAI, 2021. 3 +[11] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In CVPR, 2021. 2 +[12] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing + +gpt4v-synthesized data for a lite vision-language model. arXiv preprint arXiv:2402.11684, 2024.5 +[13] Sammy Christen, Shreyas Hampali, Fadime Sener, Edoardo Remelli, Tomas Hodan, Eric Sauser, Shugao Ma, and Bugra Tekin. Diffh2o: Diffusion-based synthesis of handobject interactions from textual descriptions. arXiv preprint arXiv:2403.17827, 2024. 3 +[14] Enric Corona, Albert Pumarola, Guillem Alenya, and Francesc Moreno-Noguer. Context-aware human motion prediction. In CVPR, 2020. 3 +[15] Jieming Cui, Tengyu Liu, Nian Liu, Yaodong Yang, Yixin Zhu, and Siyuan Huang. AnySkill: Learning open-vocabulary physical skill for interactive agents. In CVPR, 2024. 3 +[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 6, 7 +[17] Sander Dieleman, Jeffrey De Fauw, and Koray Kavukcuoglu. Exploiting cyclic symmetry in convolutional neural networks. In ICML, 2016. 5 +[18] Christian Diller and Angela Dai. CG-HOI: Contact-guided 3d human-object interaction generation. In CVPR, 2024. 3, 5, 6 +[19] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, 2024. 5 +[20] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In CVPR, 2023. 3 +[21] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. IMoS: Intent-driven full-body motion synthesis for human-object interactions. arXiv preprint arXiv:2212.07555, 2022. 3 +[22] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 1, 4, 6 +[23] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In CVPR, 2020. 2 +[24] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In ICCV, 2019. 2 +[25] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael Black. Stochastic scene-aware motion prediction. In ICCV, 2021. 2 +[26] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 3 +[27] Zhi Hou, Baosheng Yu, and Dacheng Tao. Compositional 3d human-object neural animation. arXiv preprint arXiv:2304.14070, 2023. 3 +[28] Yinghao Huang, Omid Taheri, Michael J. Black, and Dimitrios Tzionas. InterCap: Joint markerless 3D tracking of humans and objects in interaction. In GCPR, 2022. 2, 3, 5 +[29] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. CHAIRS: Towards full-body articulated human-object interaction. In ICCV, 2023. 2, 3, 5 + +[30] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In CVPR, 2024. 3 +[31] Marilyn Keller, Keenon Werling, Soyong Shin, Scott Delp, Sergi Pujades, C Karen Liu, and Michael J Black. From skin to skeleton: Towards biomechanically accurate 3d digital humans. ACM Transactions on Graphics (TOG), 42(6):1-12, 2023. 4 +[32] Jeonghwan Kim, Jisoo Kim, Jeonghyeon Na, and Hanbyul Joo. ParaHome: Parameterizing everyday home activities towards 3d generative modeling of human-object interactions. arXiv preprint arXiv:2401.10232, 2024. 3 +[33] Taeksoo Kim, Shunsuke Saito, and Hanbyul Joo. NCHO: Unsupervised learning for neural 3d composition of humans and objects. In ICCV, 2023. 3 +[34] Franziska Krebs, Andre Meixner, Isabel Patzer, and Tamim Asfour. The kit bimanual manipulation dataset. In Humanoids, 2021. 3 +[35] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. NIFTY: Neural object interaction fields for guided human motion synthesis. arXiv preprint arXiv:2307.07511, 2023. +[36] Jiye Lee and Hanbyul Joo. Locomotion-Action-Manipulation: Synthesizing human-scene interactions in complex 3d environments. In ICCV, 2023. 3 +[37] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913, 2023. 3, 5, 6 +[38] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2, 3, 4, 5, 6, 7, 8 +[39] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. arXiv preprint arXiv:2303.13129, 2023. 3 +[40] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. InterGen: Diffusion-based multi-human motion generation under complex interactions. arXiv preprint arXiv:2304.05684, 2023. 2 +[41] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-X: A large-scale 3d expressive whole-body human motion dataset. In NeurIPS, 2023. 1 +[42] Libin Liu and Jessica Hodgins. Learning basketball dribbling skills using trajectory optimization and deep reinforcement learning. ACM Transactions on Graphics (TOG), 37 (4):1-14, 2018. 3 +[43] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In ICCV, 2023. 3 +[44] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In CVPR, 2022. 2 + +[45] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics, 2015. 3, 5 +[46] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humanomatopo: Text-aligned whole-body motion generation. arxiv:2310.12978, 2023. 6 +[47] Zhengyi Luo, Jiashun Wang, Kangni Liu, Haotian Zhang, Chen Tessler, Jingbo Wang, Ye Yuan, Jinkun Cao, Zihui Lin, Fengyi Wang, et al. Smpolympics: Sports environments for physically simulated humanoids. arXiv preprint arXiv:2407.00187, 2024. 3 +[48] Xintao Lv, Liang Xu, Yichao Yan, Xin Jin, Congsheng Xu, Shuwen Wu, Yifan Liu, Lincheng Li, Mengxiao Bi, Wenjun Zeng, et al. Himo: A new benchmark for full-body human interacting with multiple objects. In ECCV. Springer, 2025. 2 +[49] Junyi Ma, Jingyi Xu, Xieyuanli Chen, and Hesheng Wang. Diff-ip2d: Diffusion-based hand-object interaction prediction on egocentric videos. arXiv preprint arXiv:2405.04370, 2024. 3 +[50] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. AMASS: Archive of motion capture as surface shapes. In ICCV, 2019. 1 +[51] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 8 +[52] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. The kit whole-body human motion database. In ICAR, 2015. 3 +[53] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. Unifying representations and large-scale whole-body motion databases for studying human motion. IEEE Transactions on Robotics, 32(4):796-809, 2016. 3 +[54] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3D pose estimation from monocular RGB. In 3DV, 2018. 2 +[55] Josh Merel, Saran Tunyasuvunakool, Arun Ahuja, Yuval Tassa, Leonard Hasenclever, Vu Pham, Tom Erez, Greg Wayne, and Nicolas Heess. Catch & carry: reusable neural controllers for vision-guided whole-body tasks. ACM Transactions on Graphics (TOG), 39(4):39-1, 2020. 3 +[56] Gyeongsik Moon, Shouu-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In ECCV, 2020. 2 +[57] Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. In NeurIPS, 2024. 5 +[58] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In CVPR, 2023. 2 + +[59] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 6 +[60] OpenAI. ChatGPT. https://chat.openai.com/, 2023.3,1 +[61] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. arXiv preprint arXiv:2308.09036, 2023. 3 +[62] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 3, 4, 1, 2 +[63] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. HOI-Diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 2, 3, 5, 6, 7 +[64] Ilya A Petrov, Riccardo Marin, Julian Chibane, and Gerard Pons-Moll. Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In CVPR, 2023. 3 +[65] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In ECCV, 2022. 4 +[66] Mathis Petrovich, Michael J Black, and Gül Varol. TMR: Text-to-motion retrieval using contrastive 3d human motion synthesis. In ICCV, 2023. 6 +[67] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236–252, 2016. 1 +[68] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In ICCV, 2019. 5, 7, 1 +[69] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english labels. In CVPR, 2021. 1 +[70] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 7 +[71] Haziq Razali and Yiannis Demiris. Action-conditioned generation of bimanual object manipulation sequences. In AAAI, 2023. 3 +[72] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 6 +[73] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6), 2017. 3, 4, 1, 2 +[74] Wenfeng Song, Xinyu Zhang, Shuai Li, Yang Gao, Aimin Hao, Xia Hou, Chenglizhao Chen, Ning Li, and Hong Qin. Hoianimator: Generating text-prompt human-object animations using novel perceptive diffusion models. In CVPR, 2024. 3, 5, 6 + +[75] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 3 +[76] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39(4):54-1, 2020. 3 +[77] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In ECCV, 2020. 2, 3, 1, 5 +[78] Omid Taheri, Vasileios Choutas, Michael J Black, and Dimitrios Tzionas. GOAL: Generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 3 +[79] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J Black. Grip: Generating interaction poses using spatial cues and latent consistency. In 3DV, 2024. 4 +[80] Purva Tendulkar, Dídac Surís, and Carl Vondrick. Flex: Full-body grasping without full-body grasps. In ICCV, 2023. 2 +[81] Chen Tessler, Yunrong Guo, Ofir Nabati, Gal Chechik, and Xue Bin Peng. Maskedmimic: Unified physics-based character control through masked motion inpainting. arXiv preprint arXiv:2409.14393, 2024. 3 +[82] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 4, 6 +[83] Guy Tevet, Signal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. Closd: Closing the loop between simulation and diffusion for multi-task character control. arXiv preprint arXiv:2410.03441, 2024. 3 +[84] Jie Tian, Lingxiao Yang, Ran Ji, Yuexin Ma, Lan Xu, Jingyi Yu, Ye Shi, and Jingya Wang. Gaze-guided hand-object interaction synthesis: Benchmark and method. arXiv preprint arXiv:2403.16169, 2024. 3 +[85] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 2022. 3 +[86] Jiashun Wang, Jessica Hodgins, and Jungdam Won. Strategy and skill learning for physics-based table tennis animation. In SIGGRAPH, 2024. 3 +[87] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In 3DV, 2022. 3 +[88] Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. PhysHOI: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393, 2023. 3, 5, 8 +[89] Yinhuai Wang, Qihan Zhao, Runyi Yu, Ailing Zeng, Jing Lin, Zhengyi Luo, Hok Wai Tsui, Jiwen Yu, Xiu Li, Qifeng Chen, et al. Skillmimic: Learning reusable basketball skills from demonstrations. arXiv preprint arXiv:2408.15270, 2024.3 + +[90] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. InterControl: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 3 +[91] Noah Wiederhold, Ava Megyeri, DiMaggio Paris, Sean Banerjee, and Natasha Banerjee. Hoh: Markerless multimodal human-object-human handover dataset with large object count. Advances in Neural Information Processing Systems, 36, 2024. 2 +[92] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In CoRL, 2023. 3 +[93] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. THOR: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 3, 5, 6 +[94] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. SAGA: Stochastic whole-body grasping with contact. In ECCV, 2022. 3, 4 +[95] Zhen Wu, Jiaman Li, and C Karen Liu. Human-object interaction from human-level instructions. arXiv preprint arXiv:2406.17840, 2024. 3 +[96] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. In ECCV, 2022. 3 +[97] Xianghui Xie, Jan Eric Lenssen, and Gerard Pons-Moll. InterTrack: Tracking human object interaction without object templates. arXiv preprint arXiv:2408.13953, 2024. 3 +[98] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. OmniControl: Control any joint at any time for human motion generation. arXiv preprint arXiv:2310.08580, 2023. 3 +[99] Zhaoming Xie, Sebastian Starke, Hung Yu Ling, and Michiel van de Panne. Learning soccer juggling skills with layerwise mixture-of-experts. In SIGGRAPH, 2022. 3 +[100] Zhaoming Xie, Jonathan Tseng, Sebastian Starke, Michiel van de Panne, and C Karen Liu. Hierarchical planning and control for box loco-manipulation. arXiv preprint arXiv:2306.09532, 2023. 3 +[101] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter-x: Towards versatile human-human interaction analysis. arXiv preprint arXiv:2312.16051, 2023. 2 +[102] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 1, 3, 4, 5, 8 +[103] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. arXiv preprint arXiv:2403.19652, 2024. 3 +[104] Sirui Xu, Hung Yu Ling, Yu-Xiong Wang, and Liang-Yan Gui. InterMimic: Towards universal whole-body control for physics-based human-object interactions. In CVPR, 2025. 1, 3, 4 +[105] Xiang Xu, Hanbyul Joo, Greg Mori, and Manolis Savva. D3D-HOI: Dynamic 3d human-object interactions from videos. arXiv preprint arXiv:2108.08420, 2021. 3 + +[106] ChangHee Yang, ChanHee Kang, Kyeongbo Kong, Hanni Oh, and Suk-Ju Kang. Person in place: Generating associative skeleton-guidance maps for human-object interaction image editing. In CVPR, 2024. 3 +[107] Jie Yang, Xuesong Niu, Nan Jiang, Ruimao Zhang, and Siyuan Huang. F-HOI: Toward fine-grained semantic-aligned 3d human-object interactions. In ECCV, 2024. 3 +[108] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, and Zheng-Jun Zha. Lemon: Learning 3d human-object interaction relation from 2d images. In CVPR, 2024. 3 +[109] Zeshi Yang, Kangkang Yin, and Libin Liu. Learning to use chopsticks in diverse gripping styles. ACM Transactions on Graphics (TOG), 41(4):1-17, 2022. 3 +[110] Heyuan Yao, Zhenhua Song, Yuyang Zhou, Tenglong Ao, Baoquan Chen, and Libin Liu. Moconvq: Unified physics-based motion control via scalable discrete representations. ACM Transactions on Graphics (TOG), 43(4):1-21, 2024. 3 +[111] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance diffusion: Synthesizing hand-object interactions. In CVPR, 2023. 3 +[112] Xinyu Zhan, Lixin Yang, Yifei Zhao, Kangrui Mao, Hanlin Xu, Zenan Lin, Kailin Li, and Cewu Lu. Oakink2: A dataset of bimanual hands-object manipulation in complex task completion. arXiv preprint arXiv:2403.19417, 2024. 2 +[113] Chengwen Zhang, Yun Liu, Ruofan Xing, Bingda Tang, and Li Yi. Core4d: A 4d human-object-human interaction dataset for collaborative object rearrangement. arXiv preprint arXiv:2406.19353, 2024. 2 +[114] Hui Zhang, Sammy Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. ArtiGrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv preprint arXiv:2309.03891, 2023. 3 +[115] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. NeuralDome: A neural modeling pipeline on multi-view humanobject interactions. In CVPR, 2023. 2, 3, 5 +[116] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2M-GPT: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 4 +[117] Juze Zhang, Jingyan Zhang, Zining Song, Zhanhe Shi, Chengfeng Zhao, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Hoi-m^3: Capture multiple humans and objects interaction within contextual environment. In CVPR, 2024. 3 +[118] Jiajun Zhang, Yuxiang Zhang, Liang An, Mengcheng Li, Hongwen Zhang, Zonghai Hu, and Yebin Liu. Manidext: Hand-object manipulation synthesis via continuous correspondence embeddings and residual-guided diffusion. arXiv preprint arXiv:2409.09300, 2024. 3 +[119] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In ECCV, 2020. 3 + +[120] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. COUCH: Towards controllable human-chair interactions. In ECCV, 2022. 3 +[121] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Ilya Petrov, Vladimir Guzov, Helisa Dhamo, Eduardo Pérez-Pellitero, and Gerard Pons-Moll. FORCE: Dataset and method for intuitive physics guided human-object interaction. arXiv preprint arXiv:2403.11237, 2024. 2 +[122] Yunbo Zhang, Deepak Gopinath, Yuting Ye, Jessica Hodgins, Greg Turk, and Jungdam Won. Simulation and retargeting of complex multi-character interactions. In SIGGRAPH, 2023. 5 +[123] Chengfeng Zhao, Juze Zhang, Jiashen Du, Ziwei Shan, Junye Wang, Jingyi Yu, Jingya Wang, and Lan Xu. I'M HOI: Inertia-aware monocular capture of 3d human-object interactions. In CVPR, 2024. 2, 3, 5 +[124] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. In ECCV, 2022. 3 +[125] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In ICCV, 2023. 3 +[126] Juntian Zheng, Qingyuan Zheng, Lixing Fang, Yun Liu, and Li Yi. CAMS: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis. In CVPR, 2023. 3 +[127] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In ECCV. Springer, 2022. 3, 4 \ No newline at end of file diff --git a/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/images.zip b/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7c3dfbc6b5e2d6c20eb3b28692a79f9087e8dc4c --- /dev/null +++ b/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf47af8cff0f77949894f996938e246ff7b987658d2cae9b83a1fbab9516d451 +size 485278 diff --git a/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/layout.json b/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..932de1e5215f0655e301019783a1b4de0a4d994a --- /dev/null +++ b/2025/InterAct_ Advancing Large-Scale Versatile 3D Human-Object Interaction Generation/layout.json @@ -0,0 +1,9911 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 102, + 533, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 102, + 533, + 138 + ], + "spans": [ + { + "bbox": [ + 78, + 102, + 533, + 138 + ], + "type": "text", + "content": "InterAct: Advancing Large-Scale Versatile 3D Human-Object Interaction Generation" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 64, + 159, + 545, + 232 + ], + "blocks": [ + { + "bbox": [ + 64, + 159, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 64, + 159, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 64, + 159, + 545, + 232 + ], + "type": "image", + "image_path": "ee9ab7a9c8f7ee5ca8269e301a95769d9c61b2836b596fb89c0929f95e464738.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 242, + 556, + 439 + ], + "blocks": [ + { + "bbox": [ + 56, + 242, + 556, + 439 + ], + "lines": [ + { + "bbox": [ + 56, + 242, + 556, + 439 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 556, + 439 + ], + "type": "image", + "image_path": "7ebbdf565494c4000173240ed91cfb3804bece795c306c69fdf547df490d31b7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 446, + 545, + 459 + ], + "lines": [ + { + "bbox": [ + 60, + 446, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 60, + 446, + 545, + 459 + ], + "type": "text", + "content": "Figure 1. An overview of InterAct, our large-scale 3D human-object interaction (HOI) benchmark, covering six HOI generation tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 474, + 200, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 474, + 200, + 487 + ], + "spans": [ + { + "bbox": [ + 151, + 474, + 200, + 487 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 297, + 715 + ], + "type": "text", + "content": "While large-scale human motion capture datasets have advanced human motion generation, modeling and generating dynamic 3D human-object interactions (HOIs) remain challenging due to dataset limitations. Existing datasets often lack extensive, high-quality motion and annotation and exhibit artifacts such as contact penetration, floating, and incorrect hand motions. To address these issues, we introduce InterAct, a large-scale 3D HOI benchmark featuring dataset and methodological advancements. First, we consolidate and standardize 21.81 hours of HOI data from diverse sources, enriching it with detailed textual annotations. Second, we propose a unified optimization framework to enhance data quality by reducing artifacts and correcting hand motions. Leveraging the principle of contact invariance, we maintain human-object relationships while introducing motion variations, expanding the dataset to 30.70 hours. Third, we define six benchmarking tasks and develop a unified HOI generative modeling perspective, achieving state-of-the-art" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 475, + 556, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 524 + ], + "type": "text", + "content": "performance. Extensive experiments validate the utility of our dataset as a foundational resource for advancing 3D human-object interaction generation. The dataset will be publicly accessible to support further research in the field." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 548, + 395, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 548, + 395, + 561 + ], + "spans": [ + { + "bbox": [ + 314, + 548, + 395, + 561 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 677 + ], + "type": "text", + "content": "Recent advances in human motion modeling have significantly benefited from extensive motion capture (MoCap) datasets [22, 41, 50, 67, 69], enabling the creation of scalable generative models for diverse human movements. Building upon this foundation, researchers are increasingly turning to the more intricate challenge of generating human-object interactions (HOIs) [102-104]. This emerging area holds considerable promise for applications in robotics, animation, and computer vision." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 678, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 556, + 715 + ], + "type": "text", + "content": "However, high-quality HOI generation faces notable obstacles due to factors such as increased degrees of freedom introduced by objects, varied object geometries, dynamic in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7048" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 194 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 194 + ], + "type": "table", + "html": "
DatasetClipHourTextHandObject
GRAB [77]1,3353.76X✓51
BEHAVE [4]2994.13XX18
InterCap [28]2330.62X✓10
Chairs [29]1,0412.37X✓92
HODome [115]1762.82X✓21
OMOMO [38]4,8388.274,838X15
IMHD [123]1640.97X✓10
InterAct (Ours)11,35021.8134,050✓217
InterAct-X (Ours)16,20130.7048,630✓217
", + "image_path": "94e1c5bbfb41468236a164856099830bc7868c53c740f66c203fc26dedee41f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 296, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 296, + 258 + ], + "type": "text", + "content": "Table 1. Comparison between InterAct, InterAct-X, and human-object interaction datasets we collect. Beyond a substantially larger scale, our dataset introduces comprehensive textual annotations and enhances interaction quality, offering a more versatile foundation for large-scale HOI generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 279, + 296, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 351 + ], + "type": "text", + "content": "teractions, and the necessity for physically accurate contact modeling. Current methods often struggle to achieve realism primarily because existing datasets lack scalability and comprehensive annotations, which are crucial for models to effectively understand interaction dynamics and link them to related domains such as natural language." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 351, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 351, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 351, + 296, + 518 + ], + "type": "text", + "content": "Specifically, these challenges underscore the need for comprehensive, high-quality HOI datasets: (1) Limited and Inconsistent Datasets: Existing methods typically depend on small datasets with limited hours of data, difficult to consolidate due to inconsistent human representations, object types, coordinate systems, and annotations. Available annotations [38, 63] are frequently coarse and incomplete, lacking detailed descriptions of human states, object interactions, and involved body parts. (2) Prevalent Artifacts: Current datasets often contain artifacts from MoCap limitations and occlusions, including unnatural penetrations, floating contacts, inaccurate hand poses [4, 38], and significant motion jitter [28]. These issues compromise the models' capacity to learn realistic human-object dynamics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 519, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 590 + ], + "type": "text", + "content": "To address these challenges, we present InterAct, a benchmark designed to systematically overcome current limitations and drive advancements in 3D HOI modeling. As shown in Table 1, InterAct offers a large-scale, standardized dataset of carefully curated interactions from existing resources1, enriched by detailed textual annotations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 591, + 296, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 296, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 296, + 675 + ], + "type": "text", + "content": "To further enhance dataset quality and scope, we introduce a unified optimization approach, addressing major penetration and floating artifacts first in whole-body interactions, followed by refined corrections for nuanced hand-object interactions. Additionally, we propose the concept of contact invariance, inspired by motion mirroring techniques, to generate realistic synthetic data by varying human motions while" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 173 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 173 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 173 + ], + "type": "image", + "image_path": "2dcd34ca991cba8ca1dbdeb3209270981a842cf323ec4e95ba21022e4f047a6f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "type": "text", + "content": "Figure 2. (a) Our data processing pipeline consolidating data, annotations via foundation models, corrections, and interaction illustrations. (b) Statistics on motion and text annotations." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 236, + 555, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 284 + ], + "type": "text", + "content": "maintaining consistent object contacts. This augmentation expands InterAct into InterAct-X, providing approximately 9 additional hours of data and substantially improving generative model performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 285, + 556, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 556, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 556, + 392 + ], + "type": "text", + "content": "Leveraging this comprehensive, richly annotated dataset, we define benchmarks across six key HOI generation tasks, Text-to-Interaction, Action-to-Interaction, Object-to-Human, Human-to-Object, Interaction Prediction, and Interaction Imitation, as shown in Figure 1, and propose a unified modeling and representation for kinematic generative tasks. Our method utilizes multi-task learning to jointly model motion and contact, achieving state-of-the-art performance as validated by comprehensive evaluations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 393, + 556, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 556, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 556, + 537 + ], + "type": "text", + "content": "In summary, our contributions are: (1) InterAct, the most extensive 3D HOI benchmark to date, facilitating large-scale generative modeling. (2) A unified optimization-based framework to correct and augment MoCap data, addressing common artifacts and significantly enhancing dataset quality. We believe this will aid future research in overcoming data scarcity before capturing potentially imperfect data. (3) Comprehensive benchmarks across six HOI generation tasks, establishing standardized metrics and demonstrating superior performance over existing approaches. This benchmark lays a strong foundation for future research, encouraging advancements across multiple facets of 3D HOI generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 548, + 400, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 548, + 400, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 548, + 400, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Dynamic 3D HOI Dataset. Many large-scale datasets with sequential human motion data have established benchmarks for the task of 3D human motion generation. However, human actions are influenced not only by individual intents but also by interactions with the surrounding environment. To address this complexity, new datasets have been developed to capture the dynamics between humans and their environments, including interactions with other humans [1, 40, 54, 101] and scenes [8, 24, 25]. Though there are fruitful hand-object interaction datasets [11, 23, 44, 48, 56, 58, 80, 91, 91, 112, 113, 121], our focus is specifically on whole-body interactions with dynamic objects, ranging from" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "1We integrate publicly available datasets [4, 28, 29, 38, 77, 115, 123] focusing on single-human interactions with rigid and dynamic objects. Certain datasets were selectively included based on relevance and data type." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7049" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 108 + ], + "type": "text", + "content": "low-dynamic interactions such as approaching and manipulation [77] to highly dynamic interactions involving multiple body parts [4, 20, 28-30, 32, 38, 107, 115, 123]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 297, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 297, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 297, + 191 + ], + "type": "text", + "content": "We aim to address the limitations of these datasets and open new possibilities for future research. Our InterAct dataset maintains advantages in motion quality, fine-grained textual annotations, detailed hand gestures, and comprehensive annotation modalities. We provide quantitative comparisons of InterAct and existing datasets in Table 1, demonstrating the superiority of our dataset in these aspects." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 191, + 298, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 298, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 298, + 624 + ], + "type": "text", + "content": "Dynamic 3D HOI Generation. Existing human-object interaction (HOI) datasets have laid a robust foundation for generating dynamic, whole-body interactions. Extensive research has explored the generation of hand-object interactions [9, 13, 39, 43, 49, 84, 111, 114, 118, 126, 127] and static human-object interactions [27, 33, 64, 87, 96, 97, 106, 108, 119, 124]. Meanwhile, full-body dynamic interactions have also been studied extensively [14, 21, 34-36, 38, 52, 53, 71, 75, 76, 78, 85, 94, 105, 120, 125], though these often face significant limitations, including narrow action repertoires and dependence on static objects. Recent advancements, such as InterDiff [102], have introduced diverse interactions involving dynamic objects and multiple body parts. Building upon this, subsequent approaches like InterDreamer [103] and other contemporary studies [18, 37, 63, 74, 93, 95, 103, 117] further demonstrate the feasibility of converting textual descriptions into realistic 3D human-object interaction sequences. Despite these advances, current methods remain constrained by a shortage of high-quality, large-scale datasets, often encountering issues related to physical inaccuracies, such as floating contacts or interpenetration. In parallel, physics-based methods leveraging deep reinforcement learning (RL) [3, 6, 10, 15, 26, 42, 55, 61, 81, 83, 86, 89, 90, 92, 98, 100, 109, 110] successfully generate physically accurate interactions, with applications in sports [47] like basketball [88, 89] and soccer [99]. Nonetheless, these methods typically produce rigid interaction patterns from limited datasets, while a recent work, InterMimic [104], illustrates that physics-based approaches can digest effectively across diverse and dynamic object interactions. Our work addresses these fundamental limitations at the dataset level by providing enhanced diversity and comprehensive sequences of human-object interactions. This facilitates multiple generative tasks, supports better contact modeling, and improves the capability to synthesize realistic and generalized human-object interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 633, + 157, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 157, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 157, + 645 + ], + "type": "text", + "content": "3. InterAct Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 297, + 713 + ], + "type": "text", + "content": "Overview. We introduce InterAct, the first unified benchmark tailored explicitly for sequential 3D human-object interaction (HOI) generative modeling. Distinguished by its unprecedented scale and comprehensiveness, InterAct significantly surpasses existing datasets, as summarized in Table 1." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 429, + 129 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 429, + 129 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 429, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 429, + 129 + ], + "type": "image", + "image_path": "3402e9787fb5232f3e63ae47ab08ab60b7b23f220afe834df2b23d3b4d539b96.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 129, + 407, + 138 + ], + "lines": [ + { + "bbox": [ + 333, + 129, + 407, + 138 + ], + "spans": [ + { + "bbox": [ + 333, + 129, + 407, + 138 + ], + "type": "text", + "content": "(a) Markers in SMPL-H" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 147, + 526, + 159 + ], + "lines": [ + { + "bbox": [ + 342, + 147, + 526, + 159 + ], + "spans": [ + { + "bbox": [ + 342, + 147, + 526, + 159 + ], + "type": "text", + "content": "Figure 3. Marker-based representation for human." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 436, + 70, + 553, + 129 + ], + "blocks": [ + { + "bbox": [ + 436, + 70, + 553, + 129 + ], + "lines": [ + { + "bbox": [ + 436, + 70, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 436, + 70, + 553, + 129 + ], + "type": "image", + "image_path": "4b0e57fb1ec483f4a17a14b96fa900f73192d6e0f30a08ebe44b45703a1b009c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 129, + 529, + 138 + ], + "lines": [ + { + "bbox": [ + 455, + 129, + 529, + 138 + ], + "spans": [ + { + "bbox": [ + 455, + 129, + 529, + 138 + ], + "type": "text", + "content": "(b) Markers in SMPL-X" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 181, + 556, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 556, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 556, + 313 + ], + "type": "text", + "content": "InterAct is available in two versions: (1) A basic version consolidating seven existing datasets, providing 21.81 hours of annotated 3D whole-body interactions with corresponding semantic descriptions. (2) An advanced version, InterAct-X, extending the basic dataset through synthetic data generated via our unified optimization framework. Figure 2 presents examples of motion and text annotations. To ensure high-quality standards, we employ a multifaceted annotation strategy combining human expertise, automated foundation models, and advanced HOI modeling techniques, all validated through rigorous manual quality checks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 323, + 545, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 545, + 335 + ], + "type": "text", + "content": "3.1. Data Collection, Annotation, and Unification" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 342, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 401 + ], + "type": "text", + "content": "We compile data from seven datasets [4, 28, 29, 38, 77, 115, 123], featuring motion capture of a single human interacting with a single dynamic 3D object, where humans are annotated with SMPL [45, 62, 73]. We address their heterogeneity in two key aspects: annotations and representations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 714 + ], + "type": "text", + "content": "Unifying Textual Annotations. Since most datasets either lack textual descriptions or provide only very coarse text descriptions [38], we implement a two-phase annotation procedure involving human annotators and GPT-4 [60] to generate consistent and detailed annotations across all subsets. In the first phase, human annotators provided detailed and precise descriptions of the interactions, adhering to the following guidelines: (i) Split motion sequences into clips averaging 300 frames (approximately 10 seconds) but no longer than 400 frames each; (ii) Clearly describe the actions and the body parts involved in the interactions. For example, a typical annotation is: \"A person sits on a stool and touches the ground with their left hand, then their right hand.\" For the subset derived from OMOMO [38], we skip this phase and directly utilize their annotations. In the second phase, we use GPT-4 to rephrase and simplify human annotations to enhance diversity and consistency. For example, the rephrased version is \"A person perches on a stool, touching the ground with their left hand, then their right hand,\" and the simplified version is \"A person sits on a stool, touching the ground with each hand alternately.\" Next, we employ GPT-4 to classify each description into one of our predefined 15 action labels with in-context learning [7]. The action label for the above sequence is \"Sit.\" We meticulously review all generated texts and action labels to ensure high quality and alignment across the dataset." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7050" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 312 + ], + "type": "text", + "content": "Unifying Human Representations. Different datasets employ varying human models (e.g., SMPL-H [73], SMPL-X [62]) and diverse shapes. A straightforward solution can be to convert different humans from SMPL-H and SMPL-X to a consistent SMPL version and encode shape parameters into the generative modeling. However, although SMPL is widely used in various human-related tasks, it is fundamentally a rotation-based representation. In the context of human motion generation, Cartesian features like joint positions and velocities are more commonly used, as seen in the integration with the HumanML3D representation [22] and in most text-to-motion work [65, 82, 116]. This is still suboptimal because joints are located beneath the body's surface and do not explicitly participate in interactions. To overcome the limitations, we use markers - specific sets of human vertices representing human motion and interactions - as a simple and unified representation capable of effectively inferring contact, evaluated in Table 5. Similar approaches are discussed in [94, 102]. Then we need to select a marker set that is consistent between SMPL-H and SMPL-X models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 313, + 298, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 313, + 298, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 313, + 298, + 601 + ], + "type": "text", + "content": "Given two human body models, SMPL-H and SMPL-X, sharing the same shape, we establish marker correspondences in two steps. First, we index the markers on the SMPL-H surface as defined in prior work [94, 102]. Second, we locate the corresponding vertices on SMPL-X by selecting the closest points to these SMPL-H markers, leveraging the official SMPL conversion, which maps each SMPL-H vertex to the nearest point on the SMPL-X mesh. We extensively evaluated the approximation error of these marker correspondences across a broad range of poses. Our results show that the maximum error consistently remains below " + }, + { + "bbox": [ + 56, + 313, + 298, + 601 + ], + "type": "inline_equation", + "content": "1\\mathrm{cm}" + }, + { + "bbox": [ + 56, + 313, + 298, + 601 + ], + "type": "text", + "content": ", a deviation unlikely to affect the overall performance of HOI generation. This high consistency arises because the markers are rigidly attached to the body, and soft deformations are disabled. As a result, the identical rigid transformations in SMPL-H and SMPL-X preserve the correspondence of the markers. Additional details on such correspondence-preserving conditions can be found in [31]. Figure 3 illustrates the marker sets for SMPL-H and SMPL-X. We use this marker-based representation to train the generative models for the tasks outlined in Sec. 4, while still relying on the original SMPL-H or SMPL-X representations for the interaction correction and augmentation methods described in Sec. 3.2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 610, + 274, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 274, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 274, + 623 + ], + "type": "text", + "content": "3.2. Interaction Correction and Augmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 297, + 714 + ], + "type": "text", + "content": "In this section, we present a unified optimization framework that addresses both the correction of MoCap artifacts and the augmentation of the dataset by introducing more synthetic data. The process takes as input the motions and geometries of humans and objects, then compares them against predefined standards to define loss functions. Using gradient-based optimization, we iteratively adjust human and object" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "content": "motions to minimize these losses, thereby refining the data to meet the desired quality criteria. The key challenge lies in formulating learning objectives that not only rectify existing data but also facilitate the generation of new synthetic data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 121, + 556, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 556, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 556, + 371 + ], + "type": "text", + "content": "Our optimization is carried out in three sequential steps: (1) full-body correction; (2) hand correction; and (3) interaction augmentation. Hand correction is handled separately because, although hand poses occupy considerable space in the SMPL representation, they contribute relatively little to the overall scale of the learning objectives. By decoupling hand correction from full-body correction, we can better balance these two processes and define more targeted objectives. In what follows, we first introduce the hand correction stage. Hand Correction. Given that many existing datasets contain inaccurate hand poses [4, 38], our approach selectively promotes contact only in regions where ground-truth data indicates hand-object interaction, while ensuring the hand motion remains natural, in spirit to InterMimic [104] but relying on predefined optimization instead of RL. This approach is effective for the whole-body interaction datasets we utilize, which generally do not require high dexterity and typically only involve the hand conforming to the object for grasping, as a common assumption in existing work [79, 127], while we distinguish our approach from those that employ multi-stage, learning-based methods for the same purpose." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 372, + 556, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 556, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 556, + 407 + ], + "type": "text", + "content": "We divide our hand correction objectives into two categories: contact promotion and hand constraints. Contact promotion is guided by the following contact loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 383, + 411, + 486, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 411, + 486, + 445 + ], + "spans": [ + { + "bbox": [ + 383, + 411, + 486, + 445 + ], + "type": "interline_equation", + "content": "E _ {\\mathrm {c o n t}} = \\sum_ {i = 1} ^ {L} c _ {i} \\sum_ {j} d _ {j} [ i ],", + "image_path": "6ecd9735e9d48267e6a94494c7fc7be08bc58c28944ca5a610db5d7bac3eab4e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "inline_equation", + "content": "d_{j}[i]" + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "text", + "content": " is the distance between the " + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "text", + "content": "-th hand vertex and its nearest point on the object's surface at the " + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "text", + "content": "-th frame, and " + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "text", + "content": " indicates whether the object and the hand are in contact at frame " + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "text", + "content": ". The contact indicator " + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "text", + "content": ", inferred from ground truth data, is a function based on hand-object distance " + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "inline_equation", + "content": "\\min_j d_j[i]" + }, + { + "bbox": [ + 313, + 450, + 556, + 642 + ], + "type": "text", + "content": ", which we provide details in supplementary. The hand constraint objectives are introduced to preserve naturalness and temporal smoothness in the hand motions. These constraints include: (1) penetration loss, which penalizes intersections between the hand and the object. (2) smoothness loss, which promotes consistent contact and reduces jittering. (3) prior loss, which constrains the range of motion (RoM) of the fingers to maintain realism. Without this constraint, contact promotion could inadvertently drive fingers into biologically impossible poses. Detailed formulations of these loss functions are provided in the supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "content": "Full-Body Correction. In this stage, all human and object poses can be updated via gradient descent. We add a reconstruction loss to ensure that the optimized interactions closely match the ground truth. Other losses mirror those used in hand correction, with two key differences: (1) Contact and penetration losses are computed for the entire body" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7051" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "rather than just the hands. (2) Prior loss is omitted because the reconstruction loss alone suffices to maintain plausible human motion. Detailed formulations of these losses are provided in supplementary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 296, + 323 + ], + "type": "text", + "content": "Interaction Augmentation. Synthetic data has become increasingly important in computer vision and generative modeling [5, 12, 19, 57], prompting a key question in the context of HOI animation: Can we scale up datasets without collecting additional MoCap data? Does existing interaction data offer information beyond its observable motions? Consider a scenario where a person grasps a box and walks, as illustrated in Figure 5. Even if their gait changes slightly, hand-box contact should remain consistent to preserve the semantics of the interaction. This illustrates the principle of interaction invariance: the core interaction persists despite minor variations in motion. Leveraging this principle, we can augment our dataset by injecting new human motions while preserving consistent object interactions. Training neural networks on such augmented data enables them to naturally learn this invariance, a common strategy in symmetric learning [17], and ultimately enhances model performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 323, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 453 + ], + "type": "text", + "content": "Our augmentation pipeline consists of three steps: (1) Object Displacement: We apply a random displacement to the object's trajectory, uniformly across all timesteps. (2) Interaction Alignment: We optimize the human motion to maintain interaction with the displaced object, using both contact consistency and reconstruction objectives. (3) Interaction Filtering: We remove low-quality augmentations, those with unreasonable initial displacements, significant penetrations (human-object or self-penetration), or alignment failures indicated by high optimization losses (e.g., excessive jitter)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "text", + "content": "During the alignment phase, the primary objective is the contact consistency loss. We first compute a distance matrix " + }, + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "text", + "content": ", where each element " + }, + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{jk} = \\|\\mathbf{v}_h^j - \\mathbf{v}_o^k\\|" + }, + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "text", + "content": " denotes the Euclidean distance between the " + }, + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "text", + "content": "-th human vertex and the " + }, + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "text", + "content": "-th object vertex, before or after displacement. With a reference matrix " + }, + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}" + }, + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "text", + "content": " from the original (pre-displacement) setup, we optimize the human motion using:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 546, + 268, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 546, + 268, + 581 + ], + "spans": [ + { + "bbox": [ + 83, + 546, + 268, + 581 + ], + "type": "interline_equation", + "content": "E _ {\\mathrm {a l i g n}} = \\sum_ {i = 1} ^ {L} \\sum_ {j, k} \\frac {1}{(\\hat {\\mathbf {D}} _ {j k} + \\epsilon) ^ {2}} \\left| \\hat {\\mathbf {D}} _ {j k} - \\mathbf {D} _ {j k} \\right| ^ {2},", + "image_path": "b408af74a4207a0f393f910a10504a891b983ac27aa61ab8b599a23ee005b639.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 587, + 296, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 296, + 659 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 587, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 587, + 296, + 659 + ], + "type": "text", + "content": " is a small constant to prevent division by zero. This formulation preserves distances between vertex pairs that were initially close, while de-emphasizing pairs that were farther apart. Additional terms in the objective enforce naturalness and stability in non-interactive regions of the human pose, as detailed in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 669, + 170, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 170, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 170, + 682 + ], + "type": "text", + "content": "4. Tasks and Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "In this section, we formally define six distinct tasks featured in our benchmark. We use unified representation across" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": "five kinematic generative tasks, where each human-object interaction sequence is represented as " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "\\langle h,o\\rangle" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ", annotated with an action category " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " and a text description " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ". The human " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " includes marker coordinates, marker velocities, signed distance vectors from each marker to the object, and foot-ground contact labels. The object " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " represents object motion, including object rotation angles, object translations. Object geometry is described by Basis Point Set (BPS) [68]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 168, + 555, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 204 + ], + "type": "text", + "content": "(1) Text-Conditioned Interaction Generation. Initially in [18, 63], the task learns a function to generate the interaction sequence based on text: " + }, + { + "bbox": [ + 313, + 168, + 555, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{t2i}}(t)\\mapsto \\langle h,o\\rangle" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 204, + 554, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 554, + 240 + ], + "type": "text", + "content": "(2) Action-Conditioned Interaction Generation. The objective is to learn a function that maps an action label to the corresponding interaction sequence: " + }, + { + "bbox": [ + 313, + 204, + 554, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{a2i}}(\\pmb {a})\\mapsto \\langle \\pmb {h},\\pmb {o}\\rangle" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 240, + 554, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 554, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 554, + 275 + ], + "type": "text", + "content": "(3) Object-Conditioned Human Generation. Initially in [38], the task generates human motion based on object sequences through a function " + }, + { + "bbox": [ + 313, + 240, + 554, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{o2h}}(\\pmb {o})\\mapsto \\pmb{h}" + }, + { + "bbox": [ + 313, + 240, + 554, + 275 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 275, + 554, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 554, + 312 + ], + "type": "text", + "content": "(4) Human-Conditioned Object Generation. Conversely, this task focuses on generating object motion sequences from human motion sequences via a function " + }, + { + "bbox": [ + 313, + 275, + 554, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{h2o}}(\\pmb {h})\\mapsto \\pmb{o}" + }, + { + "bbox": [ + 313, + 275, + 554, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 312, + 554, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 360 + ], + "type": "text", + "content": "(5) Interaction Prediction. Initially in [102], the task aims to predict future human-object interactions based on past. Let " + }, + { + "bbox": [ + 313, + 312, + 554, + 360 + ], + "type": "inline_equation", + "content": "\\langle \\pmb{h}_p, \\pmb{o}_p \\rangle" + }, + { + "bbox": [ + 313, + 312, + 554, + 360 + ], + "type": "text", + "content": " denote the past interaction and " + }, + { + "bbox": [ + 313, + 312, + 554, + 360 + ], + "type": "inline_equation", + "content": "\\langle \\pmb{h}_f, \\pmb{o}_f \\rangle" + }, + { + "bbox": [ + 313, + 312, + 554, + 360 + ], + "type": "text", + "content": " the future. The goal is to learn: " + }, + { + "bbox": [ + 313, + 312, + 554, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{p2f}}(\\langle \\pmb{h}_p, \\pmb{o}_p \\rangle) \\mapsto \\langle \\pmb{h}_f, \\pmb{o}_f \\rangle" + }, + { + "bbox": [ + 313, + 312, + 554, + 360 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 360, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 443 + ], + "type": "text", + "content": "(6) Interaction Imitation. Following [3, 88, 122], this task focuses on learning physics-based control policies to reproduce human-object interactions in a physics simulator. The output is an action sequence " + }, + { + "bbox": [ + 313, + 360, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 313, + 360, + 555, + 443 + ], + "type": "text", + "content": ", specified as joint Proportional-Derivative (PD) targets. The goal is to learn a function " + }, + { + "bbox": [ + 313, + 360, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{i2f}}" + }, + { + "bbox": [ + 313, + 360, + 555, + 443 + ], + "type": "text", + "content": " that maps the reference interaction sequences to the PD actuation sequences: " + }, + { + "bbox": [ + 313, + 360, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{i2f}}(\\langle h,o\\rangle)\\mapsto f" + }, + { + "bbox": [ + 313, + 360, + 555, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 443, + 556, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 556, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 556, + 623 + ], + "type": "text", + "content": "Unifying Multi-Task HOI Generation. We introduce an additional feature, " + }, + { + "bbox": [ + 313, + 443, + 556, + 623 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 443, + 556, + 623 + ], + "type": "text", + "content": ", which encodes human-object relationships through vectors extending from each human marker to its nearest point on the object's surface. The specific configuration of " + }, + { + "bbox": [ + 313, + 443, + 556, + 623 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 443, + 556, + 623 + ], + "type": "text", + "content": " for each generative task is described in Sec. 5. Using this feature, we can unify first five kinematic generative tasks into a multi-task learning framework by treating " + }, + { + "bbox": [ + 313, + 443, + 556, + 623 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 443, + 556, + 623 + ], + "type": "text", + "content": " as an additional output. For example, we redefine the text-conditioned interaction task as " + }, + { + "bbox": [ + 313, + 443, + 556, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{t2i}}(t) \\mapsto \\langle h, o, \\eta \\rangle" + }, + { + "bbox": [ + 313, + 443, + 556, + 623 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 443, + 556, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 443, + 556, + 623 + ], + "type": "text", + "content": " is a transformer-based diffusion model. This formulation compels the model to learn spatial relationships inherent to the interactions. In our experiments, we observe that this simple strategy, enhanced by large-scale data, consistently outperforms existing methods. Similar ideas are explored in [18, 37, 38, 63, 74, 93, 102]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 633, + 394, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 394, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 394, + 647 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "content": "We begin by evaluating the effectiveness of our data correction and augmentation methods. Following this, we benchmark existing work and our proposed method on the tasks using our dataset. We standardize the evaluation metrics and present extensive results, including ablation studies. We in" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7052" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 208 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 208 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 208 + ], + "type": "image", + "image_path": "d89e67ccd06be2a8a616659f932b4ce6f45b3376d1003fdc4854b5639a6d8d6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 214, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 214, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 555, + 237 + ], + "type": "text", + "content": "Figure 4. Qualitative evaluation of interaction correction (bottom) on the OMOMO [38] dataset shows hand recovery compared to the ground truth interaction (top). Zoom in to see details of the hand recovery." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 255, + 296, + 321 + ], + "blocks": [ + { + "bbox": [ + 57, + 255, + 296, + 321 + ], + "lines": [ + { + "bbox": [ + 57, + 255, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 57, + 255, + 296, + 321 + ], + "type": "table", + "html": "
DatasetCorrectionAugmentationPene (m)↓Cont RatioUser Study (%)
BEHAVE [4]××0.0170.04822.3
✓×0.0160.07139.7
✓✓0.0160.06938.0
OMOMO [38]××0.0090.07123.9
✓×0.0070.13139.4
✓✓0.0110.13736.7
", + "image_path": "906cd0675528ace98b05fd0e024ade7ad10acdb400441dc699a9e4419ff4222f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 329, + 296, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 352 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation and user study on the quality of data from interaction correction and augmentation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 373, + 296, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 398 + ], + "type": "text", + "content": "clude additional implementation details, such as the train-test split, in supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 219, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 219, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 219, + 419 + ], + "type": "text", + "content": "5.1. Correction and Augmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 496 + ], + "type": "text", + "content": "Metrics. We use the following two metrics: Penetration refers to the intersection depth – maximum of negative sign distances from human vertices to the object's surface – average across the sequence. Contact Ratio represents the average ratio of human vertices where their distances to object are under a threshold." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 569 + ], + "type": "text", + "content": "Quantitative Evaluations. Table 2 shows that our correction process significantly improves the quality of the original MoCap data by enhancing human-object contact and reducing penetration artifacts. Moreover, the quality of the augmented data is comparable to that of the corrected data and exceeds the quality of the original dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Qualitative Evaluations. Recognizing that quantitative metrics may not fully capture data quality, we conducted a double-blind user study. We randomly selected sequences from each of the raw, corrected, and augmented data for the subset from BEHAV [4] and OMOMO [38] datasets. Human judges were presented with 30 tuple of interactions and asked to rank the quality of three sequences. According to Table 2, over " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "39\\%" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " of judges select the corrected data as having the highest quality, significantly outperforming the original data. This confirms that our correction process effectively enhances data realism. Moreover, the augmented data receive ratings comparable to the corrected data, indicating" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 258, + 555, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 330 + ], + "type": "text", + "content": "that our synthetic data is of high quality. In Figure 4, we visualize our correction results. Despite the original data lacking detailed hand information, we successfully recover vivid and accurate hand interactions. Figure 5 showcases our augmentation, which introduces new high-quality synthetic data while maintaining consistent contact in interactions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 336, + 524, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 524, + 349 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 524, + 349 + ], + "type": "text", + "content": "5.2. Language Conditioned HOI Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 354, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 582 + ], + "type": "text", + "content": "Metrics. Following the literature on text-to-motion generation [22], we develop five metrics for evaluation. The Fréchet Inception Distance (FID) quantifies the similarity between generated HOI features and the ground truth. The Multimodality and Diversity metrics assess the variety within the generated HOI. R-Precision measures the alignment between the textual descriptions and the generated HOI. The Multimodal Distance (MM Dist) evaluates the disparity between HOI features and corresponding text features. To obtain human-object interaction (HOI) and text features for calculating these metrics, existing methods often train their feature extractors on very limited data [18, 63, 74, 93], which can degrade the quality of the evaluation. We address this limitation by incorporating our larger-scale data with marker and BPS representations. Instead of formulating a classification task to train the feature extractor [22], we follow [46, 66] and employ sequence-level contrastive learning with an InfoNCE loss [59] to train a text encoder and an HOI encoder, integrating Sentence-BERT [72] into the text encoder." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "text", + "content": "Baselines and Implementation Details. We adopt HOI-Diff [63] as our base model because it is the only publicly available option compatible with our requirements. For example, CHOIS [37] requires additional conditions beyond text input. HOI-Diff utilizes a transformer-based diffusion model [82] as the backbone, and integrates an affordable model as the classifier guidance [16]. We develop several baseline variants towards our final method by implementing three key modifications: (i) Text Encoder: We replace HOI-Diff's CLIP-based text encoder, where the latent space is not structured for human-object interactions, with our pre" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7053" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 558, + 234 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 558, + 234 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 558, + 234 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 558, + 234 + ], + "type": "image", + "image_path": "29409087b827a1c9874414c9ae315a5a2b83ac7e362ff30c3bae9bd0b77c2c17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 243, + 542, + 255 + ], + "lines": [ + { + "bbox": [ + 66, + 243, + 542, + 255 + ], + "spans": [ + { + "bbox": [ + 66, + 243, + 542, + 255 + ], + "type": "text", + "content": "Figure 5. Qualitative evaluation of interaction augmentation (bottom) shows high-quality synthetic data varied from original (top)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 265, + 555, + 357 + ], + "blocks": [ + { + "bbox": [ + 58, + 265, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 58, + 265, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 58, + 265, + 555, + 357 + ], + "type": "table", + "html": "
HOI-Aware Object Enc.HOI-Aware Text Enc.Contact GenerationContact GuidanceR-Precision†FID↓MM Dist↓Multimodality↑Diversity→
Top 1Top 2Top 3
Ground Truth0.852±0.0000.966±0.0010.989±0.0010.000±0.0002.810±0.002-11.489±0.011
XXXX0.733±0.0070.909±0.0020.957±0.0023.192±0.1914.950±0.0233.149±0.45211.192±0.019
XX✓X0.730±0.0070.913±0.0040.958±0.0051.997±0.0924.752±0.0654.171±0.02711.501±0.037
✓X✓X0.737±0.0110.912±0.0020.963±0.0081.837±0.1264.631±0.0782.836±0.58311.369±0.096
✓✓✓X0.784±0.0040.940±0.0020.980±0.0031.570±0.1394.414±0.0642.677±0.56211.409±0.005
✓✓✓✓0.784±0.0040.940±0.0000.977±0.0021.567±0.1444.412±0.0653.842±0.00511.518±0.178
", + "image_path": "75af318df00a5f26ca47b8f500265af0afb652a46cefa6e8dc1e2484e40bbc38.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 81, + 394, + 272, + 445 + ], + "blocks": [ + { + "bbox": [ + 67, + 365, + 541, + 376 + ], + "lines": [ + { + "bbox": [ + 67, + 365, + 541, + 376 + ], + "spans": [ + { + "bbox": [ + 67, + 365, + 541, + 376 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation on the task of text-conditioned interaction generation. A batch size of 64 is used for R-Precision." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 81, + 394, + 272, + 445 + ], + "lines": [ + { + "bbox": [ + 81, + 394, + 272, + 445 + ], + "spans": [ + { + "bbox": [ + 81, + 394, + 272, + 445 + ], + "type": "table", + "html": "
MethodFID↓Multimodality↑Diversity→
Ground Truth0.000±0.000-11.489±0.011
HOI-Diff [63]3.566±0.0985.321±0.14310.989±0.112
Ours2.161±0.0375.792±0.05911.291±0.261
", + "image_path": "38576889299891bd193ae8bbee1d339009e76cab230f9099184332f6bc31edab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 593 + ], + "type": "text", + "content": "trained interaction-aware text encoder. (ii) Object Shape Encoding: We substitute the original PointNet++ [70], also not pretrained for capturing interaction, with BPS [68] representations for encoding object shapes. (iii) Instead of using affordance as guidance, we regress the contact representation " + }, + { + "bbox": [ + 55, + 498, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 498, + 295, + 593 + ], + "type": "text", + "content": " through a multi-task learning. (iv) We further incorporate the contact prediction as classifier guidance [16] within the denoising process, which we detail in supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": "Quantitative Results. Table 3 presents the evaluation results for the text-conditioned interaction generation task. We assess the impact of four design choices introduced above. Notably, incorporating contact modeling and BPS encoding significantly improves the quality of generated HOI, substantially enhancing the FID score. Furthermore, using our interaction-aware text and HOI encoder enhances the quality of the generated interactions and improves the alignment between the generated results and the text input. Lastly, classifier guidance provides a slight overall performance improve" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 363, + 394, + 507, + 453 + ], + "blocks": [ + { + "bbox": [ + 55, + 453, + 295, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 475 + ], + "type": "text", + "content": "Table 4. Quantitative evaluation on the task of action-conditioned interaction generation on the entire InterAct testset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 363, + 394, + 507, + 453 + ], + "lines": [ + { + "bbox": [ + 363, + 394, + 507, + 453 + ], + "spans": [ + { + "bbox": [ + 363, + 394, + 507, + 453 + ], + "type": "table", + "html": "
RepresentationPene (m)↓Cont Ratio
SMPL0.0300.025
Joint0.0270.032
Marker0.0250.028
", + "image_path": "57ea761ff5a0bae2b673ee01760d95890209231d2489d983b7a7b565ec63cf24.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 499, + 553, + 557 + ], + "blocks": [ + { + "bbox": [ + 313, + 461, + 555, + 483 + ], + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 483 + ], + "type": "text", + "content": "Table 5. Ablation study on different representation for text-to-interaction task, evaluated under BEHAVE [4] subset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 499, + 553, + 557 + ], + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 557 + ], + "type": "table", + "html": "
ModelMPMPE↓FS↓Cprec↑Crec↑Cacc↑F1 Score↑
2-stage [38]36.500.270.810.850.770.80
1-stage36.940.290.840.820.800.81
Ours (Disc.)36.950.280.850.840.830.82
Ours (Cont.)35.690.280.850.890.850.85
", + "image_path": "8f056b0be00c8891467c87955215b6094c562fe5279c5049701af42cd9e82f57.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 566, + 555, + 588 + ], + "lines": [ + { + "bbox": [ + 313, + 566, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 555, + 588 + ], + "type": "text", + "content": "Table 6. Quantitative evaluation on object-conditioned human motion generation, with novel objects unseen from training." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 615, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 555, + 651 + ], + "type": "text", + "content": "ment. Table 4 benchmarks the performance improvements of these design choices similarly for the action-conditioned interaction generation task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "Effectiveness of Marker-Based Representation. As shown in Table 5, we compare different human representations on text-to-interaction generation. Without altering contact modeling, marker-based representation produces interactions with fewer artifacts compared to other representations." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7054" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 152, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 152, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 152, + 85 + ], + "type": "text", + "content": "5.3. HOI Inpainting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": "Metrics. Following OMOMO [38], we develop six metrics tailored for evaluating our marker-based representation. The Mean Per-Marker Position Error (MPMPE) are used to measure the similarity between the generated marker motion and the ground truth. Foot Sliding (FS) is employed to assess the skating effect, reflecting the plausibility of the motion. Additionally, we use a set of contact metrics, including precision " + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "inline_equation", + "content": "(C_{\\mathrm{prec}})" + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": ", recall " + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "inline_equation", + "content": "(C_{\\mathrm{rec}})" + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": ", accuracy " + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "inline_equation", + "content": "(C_{\\mathrm{acc}})" + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": ", and the F1 Score, to evaluate the quality of human-object contact compared to the ground truth. For the human-conditioned object generation task, we include " + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{err}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{err}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": ", which measure discrepancies in object translation and orientation between the generated and the ground truth." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 246, + 296, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 296, + 412 + ], + "type": "text", + "content": "Baselines and Implementation Details. We adopt OMOMO [38] as our base model since it is the only publicly available option. OMOMO employs a two-stage generation process: it first generates hand motions and then uses these to guide full-body generations. This strategy is particularly effective for the OMOMO dataset, where interactions primarily involve hand contact. However, it is less effective when applied to our InterAct data, which features more versatile whole-body engagement. Motivated by this, we use a single-stage pipeline with multi-task learning, as introduced in Sec. 4. We investigate two different choices for contact regression: " + }, + { + "bbox": [ + 55, + 246, + 296, + 412 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathrm{Cont.}}" + }, + { + "bbox": [ + 55, + 246, + 296, + 412 + ], + "type": "text", + "content": ", which encodes the nearest vector and distance between human markers and the object. " + }, + { + "bbox": [ + 55, + 246, + 296, + 412 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathrm{Disc.}}" + }, + { + "bbox": [ + 55, + 246, + 296, + 412 + ], + "type": "text", + "content": ", which encodes the contact labels for each marker." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 413, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 485 + ], + "type": "text", + "content": "Quantitative Results. Tables 6 and 7 illustrate the effectiveness of our single-stage pipeline, with notable performance improvements when incorporating multi-task modeling. These results provide additional evidence, complementing the evaluation of text-conditioned generation, that multitask learning significantly enhances model performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 182, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 182, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 182, + 502 + ], + "type": "text", + "content": "5.4. Interaction Prediction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 593 + ], + "type": "text", + "content": "Evaluation Metrics. We compare the generated poses to the ground truth motion data using MPMPE (mean perimeter errors), measured in meters. Second, we assess object motion accuracy using Trans.Err., the average " + }, + { + "bbox": [ + 55, + 510, + 295, + 593 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 55, + 510, + 295, + 593 + ], + "type": "text", + "content": " distance between the predicted and ground truth object translations, and Rot.Err., the average " + }, + { + "bbox": [ + 55, + 510, + 295, + 593 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 55, + 510, + 295, + 593 + ], + "type": "text", + "content": " distance between the predicted and ground truth object quaternions, following [102]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 629 + ], + "type": "text", + "content": "Implementation Details. We adapt InterDiff [102] to utilize a marker-based representation and evaluate it at various scales to assess whether it benefits from scaling laws." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 635, + 177, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 177, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 177, + 647 + ], + "type": "text", + "content": "5.5. Interaction Imitation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Implementation Details. We use PhysHOI [88] to imitate sequences from our InterAct dataset, selecting four sequences shown in Figure 6. The IsaacGym [51] simulator is used, following the same architecture, reward, and representation design as PhysHOI. Training for each sequence," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 131 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 131 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 131 + ], + "type": "image", + "image_path": "e350045dfc5cb0e9e37f275fc81131d1e67cd8a444090bbbe9321483a08eb200.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 140, + 553, + 162 + ], + "lines": [ + { + "bbox": [ + 313, + 140, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 553, + 162 + ], + "type": "text", + "content": "Figure 6. Qualitative results demonstrate the successful imitation of our corrected data using PhysHOI [88]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 173, + 553, + 210 + ], + "blocks": [ + { + "bbox": [ + 317, + 173, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 210 + ], + "type": "table", + "html": "
ModelTerr↓Oerr↓Cprec↑Crec↑Cacc↑F1 Score↑
1-stage25.920.910.830.660.720.69
Ours (multi-task)23.980.830.840.680.740.72
", + "image_path": "d8c398260c1b48792661244a82eb8c1e0b82e9dd2095dea527da4683cf6c3116.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 242, + 553, + 316 + ], + "blocks": [ + { + "bbox": [ + 334, + 220, + 533, + 231 + ], + "lines": [ + { + "bbox": [ + 334, + 220, + 533, + 231 + ], + "spans": [ + { + "bbox": [ + 334, + 220, + 533, + 231 + ], + "type": "text", + "content": "Table 7. Quantitative evaluation on human-to-object." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 242, + 553, + 316 + ], + "lines": [ + { + "bbox": [ + 317, + 242, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 242, + 553, + 316 + ], + "type": "table", + "html": "
Training dataModel sizeGlobal MPMPE↓Local MPMPE↓Trans. Err.↓\\( \\text{Rot.Err.}^{\\downarrow } \\)
BEHAV×10.1200.1030.1330.352
×20.1050.0920.1090.312
×30.1130.1000.1180.343
InterAct (Ours)×10.1060.0950.1060.297
×20.0940.0830.1030.286
×30.0910.0790.0940.264
", + "image_path": "8bb755a294568ee58aaf5a142265370f4cae115e26127550191fc8b12e633254.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 324, + 541, + 335 + ], + "lines": [ + { + "bbox": [ + 324, + 324, + 541, + 335 + ], + "spans": [ + { + "bbox": [ + 324, + 324, + 541, + 335 + ], + "type": "text", + "content": "Table 8. Quantitative evaluation on interaction prediction." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 356, + 553, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 553, + 392 + ], + "type": "text", + "content": "with separate evaluations for both raw and corrected data, is performed on a single NVIDIA A40 GPU over the course of one day." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 392, + 554, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 499 + ], + "type": "text", + "content": "Quantitative Evaluation. In addition to Figure 6, which presents the qualitative results, we evaluate the imitation policy by training on both corrected and raw data, reporting the success rate as defined in [88]. Evaluating four examples from Figure 6 and averaging the success rates over 2048 environments, training on the corrected data achieves a success rate of " + }, + { + "bbox": [ + 313, + 392, + 554, + 499 + ], + "type": "inline_equation", + "content": "90.7\\%" + }, + { + "bbox": [ + 313, + 392, + 554, + 499 + ], + "type": "text", + "content": ", surpassing the " + }, + { + "bbox": [ + 313, + 392, + 554, + 499 + ], + "type": "inline_equation", + "content": "84.4\\%" + }, + { + "bbox": [ + 313, + 392, + 554, + 499 + ], + "type": "text", + "content": " achieved with raw data. This demonstrates that our interaction correction provides better data for the motion imitation task." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 500, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 554, + 548 + ], + "type": "text", + "content": "Quantitative Results. Table 8 demonstrates that our dataset, with its larger volume of data, supports the improved performance of the trained model with larger scale. In contrast, training the model on limited data leads to overfitting." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 578, + 555, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 710 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 710 + ], + "type": "text", + "content": "We introduce InterAct, a large-scale 3D whole-body human-object interaction benchmark. We employ a unified optimization framework that performs interaction correction and augmentation, enhancing data quality and augment the dataset with synthetic data, scaling to 30.70 hours of interactions and 48,630 textual descriptions. We introduce a simple yet effective multi-task learning approach for unified HOI modeling, enabling models to be trained more effectively across multiple tasks. Our comprehensive experiments highlight the significant advantages of InterAct and our methodologies, resulting in more expressive interaction generation." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7055" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Amazon-Illinois Center on AI for Interactive Conversational Experiences, the Toyota Research Institute, the IBM-Illinois Discovery Accelerator Institute, and Snap Inc. This work used computational resources, including the NCSA Delta and DeltaAI and the PTI Jetstream2 supercomputers through allocations CIS230012, CIS230013, and CIS240311 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, as well as the TACC Frontera supercomputer, Amazon Web Services (AWS), and OpenAI API through the National Artificial Intelligence Research Resource (NAIRR) Pilot." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 239, + 115, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 115, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 115, + 252 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 259, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 259, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 259, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 66, + 259, + 296, + 281 + ], + "type": "text", + "content": "[1] CMU graphics lab motion capture database. http://mocap.cs.cmu.edu/.2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 282, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 282, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 66, + 282, + 296, + 304 + ], + "type": "text", + "content": "[2] Easymocap - make human motion capture easier. Github, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 306, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 306, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 66, + 306, + 296, + 350 + ], + "type": "text", + "content": "[3] Jinseok Bae, Jungdam Won, Donggeun Lim, Cheol-Hui Min, and Young Min Kim. Pmp: Learning to physically interact with environments using part-wise motion priors. In SIGGRAPH, 2023. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 351, + 296, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 351, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 66, + 351, + 296, + 395 + ], + "type": "text", + "content": "[4] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAV: Dataset and method for tracking human object interactions. In CVPR, 2022. 2, 3, 4, 6, 7, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 396, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 396, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 66, + 396, + 296, + 430 + ], + "type": "text", + "content": "[5] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 431, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 431, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 66, + 431, + 296, + 474 + ], + "type": "text", + "content": "[6] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. arXiv preprint arXiv:2309.07907, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 476, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 476, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 66, + 476, + 296, + 520 + ], + "type": "text", + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 521, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 521, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 66, + 521, + 296, + 554 + ], + "type": "text", + "content": "[8] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In ECCV, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 555, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 555, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 66, + 555, + 296, + 588 + ], + "type": "text", + "content": "[9] Junuk Cha, Jihyeon Kim, Jae Shin Yoon, and Seungryul Baek. Text2hoi: Text-guided 3d motion generation for hand-object interaction. In CVPR, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 590, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 590, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 61, + 590, + 296, + 622 + ], + "type": "text", + "content": "[10] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In AAAI, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 624, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 624, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 61, + 624, + 296, + 677 + ], + "type": "text", + "content": "[11] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In CVPR, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 61, + 680, + 296, + 714 + ], + "type": "text", + "content": "[12] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "type": "text", + "content": "gpt4v-synthesized data for a lite vision-language model. arXiv preprint arXiv:2402.11684, 2024.5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 96, + 555, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 555, + 148 + ], + "type": "text", + "content": "[13] Sammy Christen, Shreyas Hampali, Fadime Sener, Edoardo Remelli, Tomas Hodan, Eric Sauser, Shugao Ma, and Bugra Tekin. Diffh2o: Diffusion-based synthesis of handobject interactions from textual descriptions. arXiv preprint arXiv:2403.17827, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 150, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 150, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 321, + 150, + 555, + 183 + ], + "type": "text", + "content": "[14] Enric Corona, Albert Pumarola, Guillem Alenya, and Francesc Moreno-Noguer. Context-aware human motion prediction. In CVPR, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 184, + 555, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 184, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 321, + 184, + 555, + 226 + ], + "type": "text", + "content": "[15] Jieming Cui, Tengyu Liu, Nian Liu, Yaodong Yang, Yixin Zhu, and Siyuan Huang. AnySkill: Learning open-vocabulary physical skill for interactive agents. In CVPR, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 228, + 555, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 228, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 321, + 228, + 555, + 249 + ], + "type": "text", + "content": "[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 251, + 555, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 251, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 321, + 251, + 555, + 281 + ], + "type": "text", + "content": "[17] Sander Dieleman, Jeffrey De Fauw, and Koray Kavukcuoglu. Exploiting cyclic symmetry in convolutional neural networks. In ICML, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 283, + 555, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 283, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 321, + 283, + 555, + 314 + ], + "type": "text", + "content": "[18] Christian Diller and Angela Dai. CG-HOI: Contact-guided 3d human-object interaction generation. In CVPR, 2024. 3, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 316, + 555, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 316, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 321, + 316, + 555, + 348 + ], + "type": "text", + "content": "[19] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 350, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 350, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 321, + 350, + 555, + 392 + ], + "type": "text", + "content": "[20] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 394, + 555, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 394, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 321, + 394, + 555, + 436 + ], + "type": "text", + "content": "[21] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. IMoS: Intent-driven full-body motion synthesis for human-object interactions. arXiv preprint arXiv:2212.07555, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 437, + 555, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 437, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 321, + 437, + 555, + 469 + ], + "type": "text", + "content": "[22] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 1, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 471, + 555, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 471, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 321, + 471, + 555, + 503 + ], + "type": "text", + "content": "[23] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In CVPR, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 504, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 504, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 321, + 504, + 555, + 536 + ], + "type": "text", + "content": "[24] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In ICCV, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 537, + 555, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 537, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 321, + 537, + 555, + 568 + ], + "type": "text", + "content": "[25] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael Black. Stochastic scene-aware motion prediction. In ICCV, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 570, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 570, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 321, + 570, + 555, + 602 + ], + "type": "text", + "content": "[26] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 604, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 604, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 321, + 604, + 555, + 635 + ], + "type": "text", + "content": "[27] Zhi Hou, Baosheng Yu, and Dacheng Tao. Compositional 3d human-object neural animation. arXiv preprint arXiv:2304.14070, 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 321, + 636, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 636, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 321, + 636, + 555, + 669 + ], + "type": "text", + "content": "[28] Yinghao Huang, Omid Taheri, Michael J. Black, and Dimitrios Tzionas. InterCap: Joint markerless 3D tracking of humans and objects in interaction. In GCPR, 2022. 2, 3, 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 321, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 670, + 555, + 712 + ], + "type": "text", + "content": "[29] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. CHAIRS: Towards full-body articulated human-object interaction. In ICCV, 2023. 2, 3, 5" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7056" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 117 + ], + "type": "text", + "content": "[30] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In CVPR, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 119, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 119, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 119, + 296, + 173 + ], + "type": "text", + "content": "[31] Marilyn Keller, Keenon Werling, Soyong Shin, Scott Delp, Sergi Pujades, C Karen Liu, and Michael J Black. From skin to skeleton: Towards biomechanically accurate 3d digital humans. ACM Transactions on Graphics (TOG), 42(6):1-12, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 175, + 296, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 296, + 219 + ], + "type": "text", + "content": "[32] Jeonghwan Kim, Jisoo Kim, Jeonghyeon Na, and Hanbyul Joo. ParaHome: Parameterizing everyday home activities towards 3d generative modeling of human-object interactions. arXiv preprint arXiv:2401.10232, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 221, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 221, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 221, + 296, + 255 + ], + "type": "text", + "content": "[33] Taeksoo Kim, Shunsuke Saito, and Hanbyul Joo. NCHO: Unsupervised learning for neural 3d composition of humans and objects. In ICCV, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 256, + 296, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 256, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 256, + 296, + 289 + ], + "type": "text", + "content": "[34] Franziska Krebs, Andre Meixner, Isabel Patzer, and Tamim Asfour. The kit bimanual manipulation dataset. In Humanoids, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 334 + ], + "type": "text", + "content": "[35] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. NIFTY: Neural object interaction fields for guided human motion synthesis. arXiv preprint arXiv:2307.07511, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 336, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 296, + 369 + ], + "type": "text", + "content": "[36] Jiye Lee and Hanbyul Joo. Locomotion-Action-Manipulation: Synthesizing human-scene interactions in complex 3d environments. In ICCV, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 371, + 296, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 296, + 414 + ], + "type": "text", + "content": "[37] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913, 2023. 3, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 417, + 294, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 417, + 294, + 450 + ], + "spans": [ + { + "bbox": [ + 62, + 417, + 294, + 450 + ], + "type": "text", + "content": "[38] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 452, + 296, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 452, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 452, + 296, + 495 + ], + "type": "text", + "content": "[39] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. arXiv preprint arXiv:2303.13129, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 498, + 294, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 294, + 540 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 294, + 540 + ], + "type": "text", + "content": "[40] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. InterGen: Diffusion-based multi-human motion generation under complex interactions. arXiv preprint arXiv:2304.05684, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 543, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 543, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 543, + 296, + 586 + ], + "type": "text", + "content": "[41] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-X: A large-scale 3d expressive whole-body human motion dataset. In NeurIPS, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 589, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 589, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 589, + 296, + 632 + ], + "type": "text", + "content": "[42] Libin Liu and Jessica Hodgins. Learning basketball dribbling skills using trajectory optimization and deep reinforcement learning. ACM Transactions on Graphics (TOG), 37 (4):1-14, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 294, + 667 + ], + "type": "text", + "content": "[43] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In ICCV, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "text", + "content": "[44] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In CVPR, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "text", + "content": "[45] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics, 2015. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 118, + 555, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 555, + 160 + ], + "type": "text", + "content": "[46] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humanomatopo: Text-aligned whole-body motion generation. arxiv:2310.12978, 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 162, + 555, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 162, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 320, + 162, + 555, + 215 + ], + "type": "text", + "content": "[47] Zhengyi Luo, Jiashun Wang, Kangni Liu, Haotian Zhang, Chen Tessler, Jingbo Wang, Ye Yuan, Jinkun Cao, Zihui Lin, Fengyi Wang, et al. Smpolympics: Sports environments for physically simulated humanoids. arXiv preprint arXiv:2407.00187, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 217, + 555, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 217, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 320, + 217, + 555, + 270 + ], + "type": "text", + "content": "[48] Xintao Lv, Liang Xu, Yichao Yan, Xin Jin, Congsheng Xu, Shuwen Wu, Yifan Liu, Lincheng Li, Mengxiao Bi, Wenjun Zeng, et al. Himo: A new benchmark for full-body human interacting with multiple objects. In ECCV. Springer, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 272, + 555, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 272, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 320, + 272, + 555, + 314 + ], + "type": "text", + "content": "[49] Junyi Ma, Jingyi Xu, Xieyuanli Chen, and Hesheng Wang. Diff-ip2d: Diffusion-based hand-object interaction prediction on egocentric videos. arXiv preprint arXiv:2405.04370, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 316, + 555, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 316, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 320, + 316, + 555, + 348 + ], + "type": "text", + "content": "[50] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. AMASS: Archive of motion capture as surface shapes. In ICCV, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 350, + 555, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 350, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 320, + 350, + 555, + 403 + ], + "type": "text", + "content": "[51] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 404, + 555, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 404, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 320, + 404, + 555, + 436 + ], + "type": "text", + "content": "[52] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. The kit whole-body human motion database. In ICAR, 2015. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 437, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 437, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 320, + 437, + 555, + 491 + ], + "type": "text", + "content": "[53] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. Unifying representations and large-scale whole-body motion databases for studying human motion. IEEE Transactions on Robotics, 32(4):796-809, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 493, + 555, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 493, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 320, + 493, + 555, + 535 + ], + "type": "text", + "content": "[54] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3D pose estimation from monocular RGB. In 3DV, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 537, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 537, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 320, + 537, + 555, + 591 + ], + "type": "text", + "content": "[55] Josh Merel, Saran Tunyasuvunakool, Arun Ahuja, Yuval Tassa, Leonard Hasenclever, Vu Pham, Tom Erez, Greg Wayne, and Nicolas Heess. Catch & carry: reusable neural controllers for vision-guided whole-body tasks. ACM Transactions on Graphics (TOG), 39(4):39-1, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 593, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 593, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 320, + 593, + 555, + 635 + ], + "type": "text", + "content": "[56] Gyeongsik Moon, Shouu-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In ECCV, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 636, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 636, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 320, + 636, + 555, + 669 + ], + "type": "text", + "content": "[57] Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. In NeurIPS, 2024. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 555, + 712 + ], + "type": "text", + "content": "[58] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7057" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 296, + 106 + ], + "type": "text", + "content": "[59] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 296, + 129 + ], + "type": "text", + "content": "[60] OpenAI. ChatGPT. https://chat.openai.com/, 2023.3,1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 130, + 296, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 130, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 62, + 130, + 296, + 174 + ], + "type": "text", + "content": "[61] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. arXiv preprint arXiv:2308.09036, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 175, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 296, + 228 + ], + "type": "text", + "content": "[62] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 3, 4, 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 231, + 296, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 231, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 62, + 231, + 296, + 284 + ], + "type": "text", + "content": "[63] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. HOI-Diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 286, + 296, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 296, + 329 + ], + "type": "text", + "content": "[64] Ilya A Petrov, Riccardo Marin, Julian Chibane, and Gerard Pons-Moll. Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In CVPR, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 331, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 296, + 364 + ], + "type": "text", + "content": "[65] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In ECCV, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 365, + 296, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 296, + 398 + ], + "type": "text", + "content": "[66] Mathis Petrovich, Michael J Black, and Gül Varol. TMR: Text-to-motion retrieval using contrastive 3d human motion synthesis. In ICCV, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 399, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 431 + ], + "type": "text", + "content": "[67] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236–252, 2016. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 434, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 296, + 465 + ], + "type": "text", + "content": "[68] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In ICCV, 2019. 5, 7, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 468, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 296, + 510 + ], + "type": "text", + "content": "[69] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english labels. In CVPR, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 512, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 544 + ], + "type": "text", + "content": "[70] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 546, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 296, + 578 + ], + "type": "text", + "content": "[71] Haziq Razali and Yiannis Demiris. Action-conditioned generation of bimanual object manipulation sequences. In AAAI, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 580, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 612 + ], + "type": "text", + "content": "[72] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 614, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 656 + ], + "type": "text", + "content": "[73] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6), 2017. 3, 4, 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "type": "text", + "content": "[74] Wenfeng Song, Xinyu Zhang, Shuai Li, Yang Gao, Aimin Hao, Xia Hou, Chenglizhao Chen, Ning Li, and Hong Qin. Hoianimator: Generating text-prompt human-object animations using novel perceptive diffusion models. In CVPR, 2024. 3, 5, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 106 + ], + "type": "text", + "content": "[75] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 107, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 107, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 320, + 107, + 555, + 149 + ], + "type": "text", + "content": "[76] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39(4):54-1, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 151, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 151, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 320, + 151, + 555, + 183 + ], + "type": "text", + "content": "[77] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In ECCV, 2020. 2, 3, 1, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 185, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 185, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 185, + 555, + 217 + ], + "type": "text", + "content": "[78] Omid Taheri, Vasileios Choutas, Michael J Black, and Dimitrios Tzionas. GOAL: Generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 217, + 555, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 217, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 320, + 217, + 555, + 259 + ], + "type": "text", + "content": "[79] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J Black. Grip: Generating interaction poses using spatial cues and latent consistency. In 3DV, 2024. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 261, + 555, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 261, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 320, + 261, + 555, + 292 + ], + "type": "text", + "content": "[80] Purva Tendulkar, Dídac Surís, and Carl Vondrick. Flex: Full-body grasping without full-body grasps. In ICCV, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 294, + 555, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 294, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 320, + 294, + 555, + 337 + ], + "type": "text", + "content": "[81] Chen Tessler, Yunrong Guo, Ofir Nabati, Gal Chechik, and Xue Bin Peng. Maskedmimic: Unified physics-based character control through masked motion inpainting. arXiv preprint arXiv:2409.14393, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 338, + 555, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 338, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 320, + 338, + 555, + 380 + ], + "type": "text", + "content": "[82] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 4, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 383, + 555, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 383, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 320, + 383, + 555, + 437 + ], + "type": "text", + "content": "[83] Guy Tevet, Signal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. Closd: Closing the loop between simulation and diffusion for multi-task character control. arXiv preprint arXiv:2410.03441, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 438, + 555, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 438, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 320, + 438, + 555, + 480 + ], + "type": "text", + "content": "[84] Jie Tian, Lingxiao Yang, Ran Ji, Yuexin Ma, Lan Xu, Jingyi Yu, Ye Shi, and Jingya Wang. Gaze-guided hand-object interaction synthesis: Benchmark and method. arXiv preprint arXiv:2403.16169, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 482, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 482, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 320, + 482, + 555, + 536 + ], + "type": "text", + "content": "[85] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 537, + 555, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 537, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 320, + 537, + 555, + 568 + ], + "type": "text", + "content": "[86] Jiashun Wang, Jessica Hodgins, and Jungdam Won. Strategy and skill learning for physics-based table tennis animation. In SIGGRAPH, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 570, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 570, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 320, + 570, + 555, + 613 + ], + "type": "text", + "content": "[87] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In 3DV, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 615, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 615, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 320, + 615, + 555, + 657 + ], + "type": "text", + "content": "[88] Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. PhysHOI: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393, 2023. 3, 5, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 555, + 712 + ], + "type": "text", + "content": "[89] Yinhuai Wang, Qihan Zhao, Runyi Yu, Ailing Zeng, Jing Lin, Zhengyi Luo, Hok Wai Tsui, Jiwen Yu, Xiu Li, Qifeng Chen, et al. Skillmimic: Learning reusable basketball skills from demonstrations. arXiv preprint arXiv:2408.15270, 2024.3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7058" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 296, + 106 + ], + "type": "text", + "content": "[90] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. InterControl: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 296, + 162 + ], + "type": "text", + "content": "[91] Noah Wiederhold, Ava Megyeri, DiMaggio Paris, Sean Banerjee, and Natasha Banerjee. Hoh: Markerless multimodal human-object-human handover dataset with large object count. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 163, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 163, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 296, + 196 + ], + "type": "text", + "content": "[92] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In CoRL, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 198, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 296, + 239 + ], + "type": "text", + "content": "[93] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. THOR: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 243, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 296, + 275 + ], + "type": "text", + "content": "[94] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. SAGA: Stochastic whole-body grasping with contact. In ECCV, 2022. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 277, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 296, + 308 + ], + "type": "text", + "content": "[95] Zhen Wu, Jiaman Li, and C Karen Liu. Human-object interaction from human-level instructions. arXiv preprint arXiv:2406.17840, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 310, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 310, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 310, + 296, + 342 + ], + "type": "text", + "content": "[96] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. In ECCV, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 344, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 344, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 344, + 296, + 376 + ], + "type": "text", + "content": "[97] Xianghui Xie, Jan Eric Lenssen, and Gerard Pons-Moll. InterTrack: Tracking human object interaction without object templates. arXiv preprint arXiv:2408.13953, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 378, + 296, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 378, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 62, + 378, + 296, + 420 + ], + "type": "text", + "content": "[98] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. OmniControl: Control any joint at any time for human motion generation. arXiv preprint arXiv:2310.08580, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 422, + 296, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 455 + ], + "type": "text", + "content": "[99] Zhaoming Xie, Sebastian Starke, Hung Yu Ling, and Michiel van de Panne. Learning soccer juggling skills with layerwise mixture-of-experts. In SIGGRAPH, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 456, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 456, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 57, + 456, + 295, + 498 + ], + "type": "text", + "content": "[100] Zhaoming Xie, Jonathan Tseng, Sebastian Starke, Michiel van de Panne, and C Karen Liu. Hierarchical planning and control for box loco-manipulation. arXiv preprint arXiv:2306.09532, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 501, + 296, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 501, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 501, + 296, + 553 + ], + "type": "text", + "content": "[101] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter-x: Towards versatile human-human interaction analysis. arXiv preprint arXiv:2312.16051, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 557, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 296, + 599 + ], + "type": "text", + "content": "[102] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 1, 3, 4, 5, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "type": "text", + "content": "[103] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. arXiv preprint arXiv:2403.19652, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "text", + "content": "[104] Sirui Xu, Hung Yu Ling, Yu-Xiong Wang, and Liang-Yan Gui. InterMimic: Towards universal whole-body control for physics-based human-object interactions. In CVPR, 2025. 1, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[105] Xiang Xu, Hanbyul Joo, Greg Mori, and Manolis Savva. D3D-HOI: Dynamic 3d human-object interactions from videos. arXiv preprint arXiv:2108.08420, 2021. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[106] ChangHee Yang, ChanHee Kang, Kyeongbo Kong, Hanni Oh, and Suk-Ju Kang. Person in place: Generating associative skeleton-guidance maps for human-object interaction image editing. In CVPR, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 119, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 555, + 151 + ], + "type": "text", + "content": "[107] Jie Yang, Xuesong Niu, Nan Jiang, Ruimao Zhang, and Siyuan Huang. F-HOI: Toward fine-grained semantic-aligned 3d human-object interactions. In ECCV, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 153, + 555, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 555, + 185 + ], + "type": "text", + "content": "[108] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, and Zheng-Jun Zha. Lemon: Learning 3d human-object interaction relation from 2d images. In CVPR, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 187, + 554, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 187, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 317, + 187, + 554, + 219 + ], + "type": "text", + "content": "[109] Zeshi Yang, Kangkang Yin, and Libin Liu. Learning to use chopsticks in diverse gripping styles. ACM Transactions on Graphics (TOG), 41(4):1-17, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 221, + 555, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 221, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 317, + 221, + 555, + 263 + ], + "type": "text", + "content": "[110] Heyuan Yao, Zhenhua Song, Yuyang Zhou, Tenglong Ao, Baoquan Chen, and Libin Liu. Moconvq: Unified physics-based motion control via scalable discrete representations. ACM Transactions on Graphics (TOG), 43(4):1-21, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 266, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 266, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 555, + 308 + ], + "type": "text", + "content": "[111] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance diffusion: Synthesizing hand-object interactions. In CVPR, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 311, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 311, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 311, + 555, + 354 + ], + "type": "text", + "content": "[112] Xinyu Zhan, Lixin Yang, Yifei Zhao, Kangrui Mao, Hanlin Xu, Zenan Lin, Kailin Li, and Cewu Lu. Oakink2: A dataset of bimanual hands-object manipulation in complex task completion. arXiv preprint arXiv:2403.19417, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 357, + 555, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 357, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 317, + 357, + 555, + 399 + ], + "type": "text", + "content": "[113] Chengwen Zhang, Yun Liu, Ruofan Xing, Bingda Tang, and Li Yi. Core4d: A 4d human-object-human interaction dataset for collaborative object rearrangement. arXiv preprint arXiv:2406.19353, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 401, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 401, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 401, + 555, + 453 + ], + "type": "text", + "content": "[114] Hui Zhang, Sammy Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. ArtiGrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv preprint arXiv:2309.03891, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 456, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 555, + 499 + ], + "type": "text", + "content": "[115] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. NeuralDome: A neural modeling pipeline on multi-view humanobject interactions. In CVPR, 2023. 2, 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 502, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 502, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 502, + 555, + 555 + ], + "type": "text", + "content": "[116] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2M-GPT: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 558, + 555, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 555, + 610 + ], + "type": "text", + "content": "[117] Juze Zhang, Jingyan Zhang, Zining Song, Zhanhe Shi, Chengfeng Zhao, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Hoi-m^3: Capture multiple humans and objects interaction within contextual environment. In CVPR, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 614, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 555, + 667 + ], + "type": "text", + "content": "[118] Jiajun Zhang, Yuxiang Zhang, Liang An, Mengcheng Li, Hongwen Zhang, Zonghai Hu, and Yebin Liu. Manidext: Hand-object manipulation synthesis via continuous correspondence embeddings and residual-guided diffusion. arXiv preprint arXiv:2409.09300, 2024. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 555, + 712 + ], + "type": "text", + "content": "[119] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In ECCV, 2020. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7059" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 408 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[120] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. COUCH: Towards controllable human-chair interactions. In ECCV, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[121] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Ilya Petrov, Vladimir Guzov, Helisa Dhamo, Eduardo Pérez-Pellitero, and Gerard Pons-Moll. FORCE: Dataset and method for intuitive physics guided human-object interaction. arXiv preprint arXiv:2403.11237, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 205 + ], + "type": "text", + "content": "[122] Yunbo Zhang, Deepak Gopinath, Yuting Ye, Jessica Hodgins, Greg Turk, and Jungdam Won. Simulation and retargeting of complex multi-character interactions. In SIGGRAPH, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "type": "text", + "content": "[123] Chengfeng Zhao, Juze Zhang, Jiashen Du, Ziwei Shan, Junye Wang, Jingyi Yu, Jingya Wang, and Lan Xu. I'M HOI: Inertia-aware monocular capture of 3d human-object interactions. In CVPR, 2024. 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 285 + ], + "type": "text", + "content": "[124] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. In ECCV, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 296, + 319 + ], + "type": "text", + "content": "[125] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In ICCV, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 296, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 363 + ], + "type": "text", + "content": "[126] Juntian Zheng, Qingyuan Zheng, Lixing Fang, Yun Liu, and Li Yi. CAMS: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis. In CVPR, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 408 + ], + "type": "text", + "content": "[127] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In ECCV. Springer, 2022. 3, 4" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7060" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/1657842e-f939-4237-87d3-49f995d53e68_content_list.json b/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/1657842e-f939-4237-87d3-49f995d53e68_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6a39b13526b4c946188c27c1caa0c5f78d61b90e --- /dev/null +++ b/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/1657842e-f939-4237-87d3-49f995d53e68_content_list.json @@ -0,0 +1,1447 @@ +[ + { + "type": "text", + "text": "InterDyn: Controllable Interactive Dynamics with Video Diffusion Models", + "text_level": 1, + "bbox": [ + 120, + 126, + 870, + 148 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rick Akkerman $^{1,2*}$ Haiwen Feng $^{1*†}$ Michael J. Black $^{1}$ Dimitrios Tzionas $^{2}$ Victoria Fernández Abrevaya $^{1}$", + "bbox": [ + 102, + 172, + 898, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Max Planck Institute for Intelligent Systems, Tübingen, Germany $^{2}$ University of Amsterdam, the Netherlands", + "bbox": [ + 163, + 193, + 836, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{rick.akkerman, haiwen.feng, black, victoria.abrevaya}@tuebingen.mpg.de d.tzionas@uva.nl", + "bbox": [ + 104, + 210, + 895, + 227 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/30b0ac9f68c40219c46eb31fd49f92bbdee1d119f6ed39037577255457f9180b.jpg", + "image_caption": [ + "Input image" + ], + "image_footnote": [], + "bbox": [ + 93, + 263, + 225, + 333 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f263dc19b34cc956bce9bec510d1cc9ba36ef6e51f61f7573a896f776fb1e9a.jpg", + "image_caption": [ + "Video generation by InterDyn using only the hand mask sequence as control signal" + ], + "image_footnote": [], + "bbox": [ + 240, + 263, + 901, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/113443254ca639e65d103d521d5dce89e0f8d67c2dd453fa94c33f7d5b93205a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 345, + 225, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02312659b0585085ead49a474af9732b321279cb8a80bb0c106e2a1d61995635.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 344, + 901, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1aabfaedc7193a34b5fdd250b073f50a42f2a211a73bba63d8952a9bb96e6e4d.jpg", + "image_caption": [ + "Force propagation" + ], + "image_footnote": [], + "bbox": [ + 94, + 453, + 482, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee346ad8c41e11075b1455634b058139e7def2b188313e9387be0341fac73f0e.jpg", + "image_caption": [ + "Counterfactual dynamics", + "Figure 1. We present InterDyn, a framework for synthesizing realistic interactive dynamics without 3D reconstruction and physics simulation. Our core principle is to rely on the implicit physics knowledge embedded in large-scale video generative models. Given an image and a \"driving motion\", our model generates the consequential scene dynamics. We investigate the generated interactive dynamics in a simple object collision scenario (bottom) and complex in-the-wild human-object interaction (top)." + ], + "image_footnote": [ + "Denotes a driving object motion | Tracks indicate an object with generated uncontrolled dynamics" + ], + "bbox": [ + 504, + 453, + 901, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 645, + 326, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Predicting the dynamics of interacting objects is essential for both humans and intelligent systems. However, existing approaches are limited to simplified, toy settings and lack generalizability to complex, real-world environments. Recent advances in generative models have enabled the prediction of state transitions based on interventions, but focus on generating a single future state which neglects the continuous dynamics resulting from the interaction. To address this gap, we propose InterDyn, a novel framework that generates videos of interactive dynamics given an initial frame and a control signal encoding the motion of a driving object or actor. Our key insight is that large video", + "bbox": [ + 86, + 678, + 485, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "generation models can act as both neural renderers and implicit physics \"simulators\", having learned interactive dynamics from large-scale video data. To effectively harness this capability, we introduce an interactive control mechanism that conditions the video generation process on the motion of the driving entity. Qualitative results demonstrate that InterDyn generates plausible, temporally consistent videos of complex object interactions while generalizing to unseen objects. Quantitative evaluations show that InterDyn outperforms baselines that focus on static state transitions. This work highlights the potential of leveraging video generative models as implicit physics engines. Code and trained models will be released at: https://interdyn.is.tue.mpg.de/.", + "bbox": [ + 511, + 647, + 908, + 859 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 112, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Project lead", + "bbox": [ + 116, + 887, + 186, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12467", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Humans have the remarkable ability to predict the future dynamics of observed systems intuitively. With just a single image, we can anticipate and imagine how objects will move over time – not only their motion but also their interactions with the environment and other elements in the scene. Inferring this requires an advanced form of scene-level reasoning beyond merely recognizing the semantics and geometry of static elements; it involves a deep physical and causal understanding of how each object will interact given the environment, object properties, and forces.", + "bbox": [ + 89, + 114, + 483, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There has been a growing interest in developing machine learning systems that emulate similar levels of dynamic understanding given visual observations, such as images or videos. Early work [84] addressed this by first reconstructing a 3D representation from the image, then predicting future states with a physics simulator and finally generating the output video with a rendering engine. This relies heavily on explicit reconstruction and simulation, which is computationally intensive, prone to errors, and may not generalize well. More recent methods [2, 24, 39, 46, 49] leverage keypoint or latent representations within graph relational frameworks; however, they have only been trained and validated in over-simplified, synthetic environments, showing limited generalizability to complex real-world scenarios.", + "bbox": [ + 89, + 266, + 483, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead, the advent of powerful generative models [1, 5, 16, 54, 66] opens new avenues for synthesizing interactions under complex scenarios. For example, Sudhakar et al. [70] recently proposed CosHand, a controllable image-to-image model based on Stable Diffusion [66] that infers state transitions of an object. The task here is defined as follows: given an image of a hand interacting with an object, alongside a hand mask of the current frame and a mask of the hand at a future frame, generate a modified input image that satisfies the mask, with realistic interactions. The challenge, as in early intuitive physics works, lies in accurately modeling how the objects will change after forces are applied. However, we argue that static state transitions are insufficient for this task, as they fail to capture the continuous dynamic processes inherent to the problem, e.g. see Fig. 2. Investigating interactive dynamics within a two-state setting is highly limiting, since dynamics can extend beyond the period of direct contact – for example, predicting the motion occurring while a person pours water requires a physical understanding that goes beyond the state of the hand at a future frame. The driving force, in this case the hand, may interact with the system only briefly, but the system's subsequent dynamics continue according to physical laws and may even influence other parts via force propagation.", + "bbox": [ + 89, + 478, + 483, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we explore controllable synthesis of interactive dynamics—generating a video from an input image and a dynamic control signal (e.g. a moving hand mask) to model realistic object dynamics. In particular, we propose", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ebe545c08fdf1434341070911268a6e0cf3fd3097f7b639364bfe10dc5ebd701.jpg", + "image_caption": [ + "Figure 2. State transition vs. dynamics. Methods that generate static state transitions (i.e. predict a future image) such as CosHand [70] struggle to capture the inherent dynamic processes involved in human-object interactions. Here, we show a video sequence where the motion continues beyond the interaction." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 905, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "InterDyn, a novel framework for synthesizing controllable dynamic interactions that leverages the physical and dynamics \"knowledge\" of a large video generation model [5]. Unlike prior approaches that rely on explicit physical simulation [84] or are constrained to static state transitions [70], we leverage video generation models to generate dynamic processes implicitly, see Fig. 1. Specifically, we extend Stable Video Diffusion (SVD) [5] with a dynamic control branch and fine-tune it on diverse scenes, enabling synthesis of complex interactions aligned with the control signal.", + "bbox": [ + 511, + 321, + 906, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We start our investigation by fine-tuning InterDyn on a simple synthetic scenario of cubes, cylinders, and spheres: the CLEVRER dataset [97]. To control the motion we add a mask driving signal that manipulates the movement of some (but not all) of the objects in the scene. We then evaluate how the synthesized trajectories of uncontrolled objects change under various interactions, including multiple objects colliding with each other. This multi-object collision setting allows us to \"probe\" the physical understanding and causal effects of the video diffusion model, and our qualitative experiments show InterDyn's ability for counterfactual future prediction and physical force propagation.", + "bbox": [ + 511, + 474, + 908, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Next, we evaluate how the system performs in a difficult real-world scenario, such as Human-Object Interaction (HOI). Here, the dexterity of hand motions and the diversity of objects vastly increase the complexity of the problem. We fine-tune the model on a commonly used HOI video dataset [21] and compare with the state-of-the-art baseline CosHand [70], as well as two text-control based interactive dynamics generation methods: Seer [23] and DynamiCrafter [89]. We quantify our investigations using standard image and video metrics, as well as a motion fidelity metric based on point tracking. InterDyn surpasses the previous SOTA over $37.5\\%$ on LPIPS and $77\\%$ on FVD on the Something-Something-v2 (SSV2) dataset [21]. Our experiments also demonstrate diverse and physically plausible generations of interactive dynamics, probing into SVD's \"understanding\" of physics and dynamics.", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12468", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we present InterDyn, a framework that employs video generative models to simulate object dynamics without explicit 3D reconstruction or physical simulation. We demonstrate how the inherent \"knowledge\" within video generation models can be leveraged to predict complex object interactions and movements over time, implicitly modeling physical and causal dynamics. We perform comprehensive experiments on multi-object collision datasets and hand-object manipulation datasets, demonstrating the effectiveness of our approach.", + "bbox": [ + 89, + 90, + 480, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 262, + 232, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modeling human-object interactions (HOI). Human-object interaction has been widely studied within the context of 3D reconstruction [17, 18, 27, 28, 73, 88, 94], where the goal is to recover realistic geometry of hands and objects. The field of 3D HOI synthesis has also received increasing attention, including the generation of static [40, 44, 71, 106] or dynamic [63, 72, 103, 105] hand poses conditioned on 3D objects, whole-body interactions [90], or more recently, hand-object meshes given textual descriptions [10, 15, 57, 86, 96]. Few works address HOI synthesis in the 2D domain. GANHand [14] predicts 3D hand shape and pose given an RGB image of an object, while AffordanceDiffusion [95] estimates a 2D hand using a diffusion model. Kulal et al. [47] take as input an image of a human and a scene separately and generate a composite image that positions the human with correct affordances. Also relevant is HOIDiffusion [101], in which a texture-less rendering of a 3D hand and object is converted to a realistic image using a text description. Most closely related to us is CosHand [70], which takes as input an RGB image of a hand-object interaction, a hand mask at the current state, and the hand mask of the future state, and generates an RGB image of the future state. Unlike us, they cannot generate post-interaction object dynamics. Importantly, none of these works study dynamics, generating instead discrete state transitions that fail to capture the nuanced, temporally coherent behaviors observed in interactions.", + "bbox": [ + 91, + 289, + 483, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthesizing causal physical relations from visual input.", + "text_level": 1, + "bbox": [ + 89, + 704, + 482, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A growing body of work aims to model and predict physical causal effects from visual inputs such as images or videos. For example, research in intuitive physics seeks to replicate the human-like, non-mathematical understanding of physical events, e.g. by predicting future frames given an input video. Early works like [22, 48] train neural networks to assess the stability of block towers, while [24] leverage prior physical knowledge formalized through partial differential equations (PDEs). Other approaches investigate counterfactual reasoning by leveraging graph neural networks [2, 39]. Wu et al. [82-84] explore the use of an inverse rendering approach, extracting geometry and physical properties from", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the video which are then coupled with a physics simulator and a rendering engine to generate the future frames. Other works [81] incorporate Interaction Networks [3] to approximate physical systems from video data. These approaches are often limited to simplified, synthetic datasets and struggle to generalize to real-world scenarios.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent methods have started to combine language models with physical engines. Liu et al. [51] ground a large language model using a computational physics engine while Gao et al. [19] show that fine-tuning a vision-language model (VLM) on annotated datasets of physical concepts improves its understanding of physical interactions. Closely related to our work is PhysGen [53], which trains an imaged-to-video model that conditions the video generation on physics parameters (e.g., force or torque). However, the model relies on a dynamics simulator to generate motion, and its application is limited to rigid objects. A related but tangential line of work focuses on identifying and generating the effects of objects on their surroundings. For example, Omnimatte [56] introduces the problem of identifying all parts of a scene influenced by an object, given a video and a mask of the object. Similarly, Lu et al. [55] propose to re-time the motion of different subjects in a scene while maintaining realistic interactions with the environment. ActAnywhere [62] generates videos with plausible human-scene interactions, taking a masked video of a person and a background image as input. These works address the problem of synthesizing realistic interactions within a scene, however, lack fine-grained control.", + "bbox": [ + 511, + 194, + 906, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controllable video generation. Video generation has advanced significantly in recent years, with diffusion models leading to substantial improvements in unconditional [33, 99], text-based [1, 5, 6, 12, 20, 23, 26, 32, 37, 68, 76, 85, 89, 92, 104] and image-based [1, 5, 20, 25, 78, 89] generation. These advances have raised the question of how to incorporate more nuanced control into video generation. Some text-to-video approaches are trained by \"inflating\" text-to-image (T2V) models [8, 13, 25, 26, 38, 85], and can thus be integrated with conditional T2V models such as ControlNet [100] or T2V-Adapter [60]. Control can also be achieved by conditioning on trajectories [58, 87, 98] or bounding-boxes [77], fine-tuning on appropriate datasets. VideoComposer [78] incorporates multiple condition types, including text, depth, style, and temporal conditions via motion vectors. Camera motion control has also been explored, withAnimateDiff [79] employing LoRA [34] modules to control camera movement, while MotionCtrl [80] and CameraCtrl [29] directly embed the camera information for more precise control. Additionally, several works target human animation from a pose control signal, such as DreamPose [42], MagicPose [91], andAnimateAnyone [35], but do not account for interactions.", + "bbox": [ + 511, + 553, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12469", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Controllable Interactive Dynamics", + "text_level": 1, + "bbox": [ + 89, + 89, + 405, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video diffusion models such as [5, 54] have demonstrated impressive performance in generating videos from text or images, and have even shown potential in tasks that require 3D understanding when properly fine-tuned [36, 75]. Trained on millions of videos, we hypothesize that these models also possess implicit knowledge of complex interactive dynamics, such as those that appear when humans interact with objects. Out of the box, however, they lack a precise control mechanism, often relying solely on textual inputs or requiring careful selection of the starting frame.", + "bbox": [ + 89, + 114, + 483, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task. Given an input image, $\\pmb{x} \\in \\mathbb{R}^{1 \\times H \\times W \\times 3}$ , and a driving motion in the form of a pixel-wise corresponding control signal $\\pmb{c} \\in \\mathbb{R}^{N \\times H \\times W \\times 3}$ , we task InterDyn with generating a video sequence, $\\pmb{y} \\in \\mathbb{R}^{N \\times H \\times W \\times 3}$ , depicting plausible object dynamics. Through this task, we aim to learn the conditional distribution between a driving motion, such as that of a human hand, and the consequent motion of manipulated objects. In other words, the model needs to synthesize plausible object movement and appearance without any indication other than the driving motion, while maintaining physical and visual consistency with the input image.", + "bbox": [ + 89, + 270, + 483, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stable Video Diffusion. We extend Stable Video Diffusion [5] (SVD) to enable controllable interactive dynamics and explore the versatility of this model across a range of scenarios. SVD is a publicly available U-Net-based latent diffusion model [66] that extends Stable Diffusion 2.1 to video generation by interleaving the network with temporal layers. Given a static input image of a scene, SVD denoises a sequence of $N$ frames $\\pmb{y} \\in \\mathbb{R}^{N \\times H \\times W \\times 3}$ to generate a video that follows the initial frame. The input image is fed into the denoising U-Net by concatenating its latent to each of the frames in the noised input, and by supplying its CLIP [64] embedding to the U-Net's cross-attention layers. In addition, SVD is conditioned on the video's FPS and motion ID, where the motion ID represents the amount of motion in the video. We found a motion ID of 40 to align well with our frozen SVD prior.", + "bbox": [ + 89, + 441, + 483, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Control. InterDyn extends SVD with an additional control signal $c \\in \\mathbb{R}^{N \\times H \\times W \\times 3}$ by integrating a ControlNet-like branch [100]. An overview of our pipeline is presented in Fig. 3. The SVD weights remain frozen to preserve its learned dynamics prior. Following [100], we introduce a trainable copy of the SVD encoder $E$ , connected to SVD's frozen decoder via skip connections, and modulated by zero-initialized convolutions. We use a small CNN, $\\mathcal{E}(\\cdot)$ , to encode the control signal $c$ into the latent space, which is then added to the noisy input latent that is passed to the ControlNet encoder. Similar to SVD, the control branch interleaves convolutional, spatial, and temporal blocks, enabling InterDyn to process the control signal in a temporal-aware manner. This helps InterDyn to be robust to", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9046525db769913ab2d00853ff28fe6ab5804a5309900a6555b944e893f40745.jpg", + "image_caption": [ + "Figure 3. Overview of InterDyn. Given an input image depicting a scene, such as a hand holding a remote, and a \"driving motion,\" such as a sequence of binary hand masks, InterDyn generates a video depicting plausible hand and object dynamics. Crucially, InterDyn receives no control signal for the object. Through this setup, we probe and assess the implicit knowledge of large video generation models on complex interactive dynamics. We use Stable Video Diffusion (SVD) as our frozen backbone and fine-tune a separate control signal encoder. Videos are iteratively denoised over $T$ timesteps, starting from Gaussian noise $\\epsilon \\sim \\mathcal{N}(0,I)$ ." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 905, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "noisy control signals, see Fig. 6. We opt for binary masks as conditioning signal due to their accessibility. However, our method can be extended to incorporate diverse types of signals. We find that for hand-object interactions, the type of conditioning signal does not significantly impact performance, see Appendix A and Tab. S1.", + "bbox": [ + 511, + 434, + 906, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inference. During inference, we start from an input image, control signal sequence, and randomly sampled Gaussian noise $\\epsilon \\sim \\mathcal{N}(0,I)$ . Through iteratively applying InterDyn over $T$ denoising timesteps, we generate a video $\\pmb{y}$ depicting plausible hand and object dynamics, aligned with the control signal.", + "bbox": [ + 511, + 529, + 905, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 633, + 645, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The primary goal of this work is to synthesize scene-level interactive dynamics by leveraging the implicit physical understanding of a pre-trained video generative model. We begin by probing the model's ability to predict physically plausible outcomes within simulated environments, specifically using the CLEVRER dataset [97]. We test force propagation amongst uncontrolled objects and examine counterfactual future prediction by generating videos for one input image with different control signals. Motivated by promising results, we extend our investigation to complex, real-world hand-object interaction scenarios using the Something-Something-v2 (SSV2) dataset [21], conducting comprehensive comparisons with existing baselines that pursue similar objectives. Additionally, we showcase diverse physical examples to demonstrate the capabilities of InterDyn in generating realistic interactive dynamics.", + "bbox": [ + 511, + 659, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12470", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e301e10426e190d9050c359053ddaf95d49c1b84a0e16be2b8ff73b60fa5d86c.jpg", + "image_caption": [ + "(a) Force Propagation. InterDyn generates force propagation to and amongst uncontrolled objects (cyan cylinder, pink cylinder, and pink cube)." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 906, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dba23c81a0e46e6701a9bde0c25c5a11a0c1513d3dd0d3519aa619850b75768a.jpg", + "image_caption": [ + "(b) Counterfactual Dynamics. Two generations of InterDyn on the same input image: (1) w/ brown-cylinder control and (2) w/o brown-cylinder control.", + "Figure 4. Qualitative investigation on the CLEVRER dataset. Given an input image and the \"driving\" motion of one or two objects, our model predicts the future interactive dynamics of multiple elements in the scene. The driving motion is given in the form of semantic mask sequences. The generated object motions are highlighted with a red-line trajectory. Note that our model can generate videos with force propagation across multiple uncontrolled objects (top) and can generate multiple futures (bottom). Q Zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 91, + 202, + 906, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation", + "text_level": 1, + "bbox": [ + 89, + 460, + 250, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We initialize InterDyn with the 14-frame image-to-video weights of SVD [5]. During training, we use the Adam optimizer [45] with a learning rate of $1 \\times 10^{-5}$ . We use the EDM framework [43] with noise distribution $\\log \\sigma \\sim \\mathcal{N}(0.7, 1.6^2)$ . We train on two 80GB H100 GPUs, with a per-GPU batch size of 4. Video length and FPS define the temporal resolution of dynamics; to balance short- and long-range events we subsample videos to 7 FPS. To facilitate classifier-free guidance [31], we randomly drop the input image with a probability of $5\\%$ . At inference, we apply the Euler scheduler [43] over 50 denoising timesteps.", + "bbox": [ + 88, + 482, + 483, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Metrics", + "text_level": 1, + "bbox": [ + 89, + 657, + 187, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate InterDyn on image quality, spatio-temporal similarity, and motion fidelity. Image quality metrics are computed frame-wise. All metrics are reported excluding the first frame, as it serves as input conditioning.", + "bbox": [ + 89, + 680, + 483, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image quality. We report the Structural Similarity Index Measure (SSIM) [79], Peak Signal-to-Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS) [102], Fréchet Inception Distance (FID) [30] and unbiased Kernel Inception Distance (KID) [4].", + "bbox": [ + 89, + 744, + 483, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatio-temporal similarity. To assess the spatio-temporal perceptual similarity between the ground truth and the generated video distributions, we use the Fréchet Video Distance (FVD) and unbiased Kernel Video Distance (KVD) proposed in [74]. We use the implementation of [69].", + "bbox": [ + 88, + 824, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion Fidelity. Through InterDyn, we do not have explicit control over object dynamics, which means that the pixel alignment of an object in the generated and ground truth video is only guaranteed in the starting frame. In this case, comparing generated object motion to the ground truth naively might misrepresent the true quality of object motion over time. Therefore, we adapt the Motion Fidelity (MF) metric proposed by Yatim et al. [93], which measures the similarity between point-tracking trajectories.", + "bbox": [ + 511, + 460, + 906, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To compute the metric for any video, we obtain a mask of the object in the starting frame, sample 100 points on the object, and track these throughout both the ground truth and generated video using CoTracker3 [41]. Given the resulting two sets of tracklets $\\mathcal{T} = \\{\\tau_1,\\dots ,\\tau_n\\}$ , $\\tilde{\\mathcal{T}} = \\{\\tilde{\\tau}_1,\\dots ,\\tilde{\\tau}_m\\}$ the motion fidelity metric is defined as:", + "bbox": [ + 511, + 598, + 906, + 689 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{m} \\sum_ {\\tilde {\\tau} \\in \\tilde {\\mathcal {T}}} \\max _ {\\tau \\in \\mathcal {T}} \\mathbf {c o r r} (\\tau , \\tilde {\\tau}) + \\frac {1}{n} \\sum_ {\\tau \\in \\mathcal {T}} \\max _ {\\tilde {\\tau} \\in \\tilde {\\mathcal {T}}} \\mathbf {c o r r} (\\tau , \\tilde {\\tau}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 700, + 906, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with the correlation between two tracklets $\\mathbf{corr}(\\tau, \\tilde{\\tau})$ [50]:", + "bbox": [ + 511, + 753, + 898, + 771 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c o r r} (\\tau , \\tilde {\\tau}) \\frac {1}{F} \\sum_ {k = 1} ^ {F} \\frac {v _ {k} ^ {x} a \\cdot \\tilde {v} _ {k} ^ {x} + v _ {k} ^ {y} \\cdot \\tilde {v} _ {k} ^ {y}}{\\sqrt {(v _ {k} ^ {x}) ^ {2} + (v _ {k} ^ {y}) ^ {2}} \\cdot \\sqrt {(\\tilde {v} _ {k} ^ {x}) ^ {2} + (\\tilde {v} _ {k} ^ {y}) ^ {2}}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 782, + 906, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(v_k^x, v_k^y)$ , $(\\tilde{v}_k^x, \\tilde{v}_k^y)$ are the $k^{th}$ frame displacement of tracklets $\\tau, \\tilde{\\tau}$ respectively. If there are less than 100 points to query on the object due to it being too small, we do not consider the video for the motion fidelity metric.", + "bbox": [ + 511, + 838, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12471", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Probing Dynamics with Object Collision Events", + "text_level": 1, + "bbox": [ + 89, + 90, + 480, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, we fine-tune InterDyn on an object collision dataset to probe its ability to generate realistic object interactions. Qualitatively, we review whether InterDyn can produce plausible object motion for uncontrolled objects, given the motion of objects entering the scene. In addition, we examine whether InterDyn can generate counterfactual videos for the same input image, but different control signals.", + "bbox": [ + 89, + 112, + 482, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We use CLEVRER [97], which provides 20,000 videos of colliding objects with annotated segmentation masks and metadata on collision events. We construct a control signal for objects entering the scene and aim to use InterDyn to generate the motion of the objects that are already present, upon collision. Stationary objects do not receive any form of control signal. Colored masks help the model distinguish unique objects. The frames are cropped and scaled to $320 \\times 448$ , and we only sample input frames before collisions between objects in the scene, to maximize InterDyn's exposure to interactive dynamics.", + "bbox": [ + 89, + 222, + 483, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Force propagation. InterDyn can generate force propagation dynamics between a controlled object and an uncontrolled object, as well as amongst uncontrolled objects, as illustrated in Fig. 4a. Here, the red cylinder at the top left is the driving force. It collides with the uncontrolled blue cylinder, which then collides with the uncontrolled purple cylinder, in turn striking the uncontrolled purple cube on the far right. Point-tracking trajectories display how collisions alter each object's position. This suggests that InterDyn possesses an implicit understanding of physical interactions, enabling it to generate plausible dynamics.", + "bbox": [ + 89, + 393, + 483, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Counterfactual dynamics. By altering the control signal, InterDyn can simulate counterfactual scenarios for the same input image, as shown in Fig. 4b. In \"Generation 1\", the gray cylinder (controlled) collides with the stationary red sphere (uncontrolled), causing it to move; it is later struck by the brown cylinder (controlled), altering its path once again. In \"Generation 2\", removing the brown cylinder lets the red sphere continue along its original trajectory, consistent with expectations. Crucially, there is no control signal for the red sphere throughout the sequence; its motion and trajectory are entirely generated by InterDyn.", + "bbox": [ + 89, + 563, + 482, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Probing InterDyn on CLEVRER highlights its ability to generate interactive dynamics for objects within a simple synthetic environment. We provide additional results in video format on our webpage.", + "bbox": [ + 89, + 734, + 483, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Generating Human-Object Interactions", + "text_level": 1, + "bbox": [ + 89, + 803, + 431, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For this experiment, we fine-tune InterDyn on a human-action dataset, focused on hand-object interaction. We encode human movement as a sequence of pixel-aligned binary hand masks and task InterDyn to generate a video with hand movements and corresponding object dynamics.", + "bbox": [ + 89, + 825, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We use Something-Something-v2 (SSV2), which provides 220,847 videos of humans performing basic actions with everyday objects. It contains actions such as \"pushing [something] from left to right\", \"squeezing [something]\" and \"lifting [something] with [something] on it\". This dataset allows us to train InterDyn at a larger scale and compare with our baseline CosHand [70]. We train one version of InterDyn at the same resolution as CosHand, $256 \\times 256$ , and a second version at $256 \\times 384$ , which aligns better with the dynamic prior of SVD.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We generate binary hand masks by prompting Segment Anything 2 (SAM2) [65] with hand bounding boxes, provided by the Something-Else dataset [59]. Similar to our baseline CosHand, we exclude the \"pretending\" class. Additionally, we remove all \"[something] is done behind [something]\" classes since the object would be out of view of the camera. We remove videos smaller than the target resolution/length, ensure the hand and objects are continuously visible, and crop videos larger than the target resolution while centering the object. We include videos without obvious motion and state transition, such as \"holding [something]\". Since bounding box annotations are only provided for the train split (79,043 samples for $256 \\times 384$ and 104,260 for $256 \\times 256$ ) and validation split (8,667 samples for $256 \\times 384$ and 11,229 for $256 \\times 256$ ), we report all evaluation metrics on the validation split.", + "bbox": [ + 511, + 243, + 906, + 484 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/496d5842234bf144097f5c51c0980be1d898e59e0df055ff34743b8950b74631.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison. A two-state approach such as CosHand [70] struggles with post-interaction object dynamics." + ], + "image_footnote": [], + "bbox": [ + 519, + 503, + 906, + 691 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b57feeb496674aaef718726f580fd586084668415794913ff42ec78d573cb139.jpg", + "image_caption": [ + "Figure 6. Robustness to noise. SAM2 outputs noisy/coarse masks for frames with considerable motion blur (orange/red). Despite this, InterDyn can generate plausible hand and object dynamics." + ], + "image_footnote": [], + "bbox": [ + 519, + 744, + 903, + 849 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12472", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/36c13c41b2208273a87cd005977ac656f1d048b33a94e18c11a2681a793f803e.jpg", + "image_caption": [ + "Figure 7. Diverse generation of interactive dynamics. We show multiple challenging examples, such as (from top to bottom): interacting with articulated objects, pouring liquid, letting an object fall, squeezing a highly deformable or \"collapsible\" object, interacting with reflective objects, tilting a ridged object, or stacking objects. Q Zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 906, + 843 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12473", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ee9e526b149a9ce772020447694030346042c57b4a1787760b1208cd5d52d28f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSSIM ↑PSNR ↑LPIPS ↓FID ↓KID ↓FVD ↓KVD ↓Motion Fidelity ↑ [93]
Seer [23]0.41810.710.58833.350.030287.4681.31—
DynamiCrafter [89]†———17.480.014204.1131.81—
CosHand-Independent [70]0.61516.870.3132.950.00291.1819.240.432
CosHand-Autoregressive [70]0.53114.920.40812.660.01290.3013.680.570
Ours 256×2560.66418.600.2604.950.00419.271.990.633
Ours 256×3840.68019.040.2525.070.00422.222.090.641
", + "bbox": [ + 91, + 88, + 898, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison on Something-Something-v2. We compare against two language-instructed video generation methods, Seer [23] and DynamiCrafter [89] and two video extensions of our baseline CosHand [70]. We report results for InterDyn at two resolutions: $256 \\times 256$ (matching CosHand) and $256 \\times 384$ (matching SVD's prior). Methods denoted with $\\dagger$ do not use SSV2 as their training dataset.", + "bbox": [ + 89, + 210, + 906, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. To generate videos using CosHand [70] rather than state transitions, we run CosHand in a frame-by-frame approach (CosHand-Independent) and an auto-regressive approach (CosHand-Autoregressive). In the frame-by-frame variant, each future frame is independently predicted from the initial frame and its corresponding mask:", + "bbox": [ + 89, + 282, + 482, + 376 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {t + 1} = \\operatorname {C o s H a n d} \\left(x _ {0}, h _ {0}, h _ {t + 1}\\right), \\quad \\forall t \\in [ 0, 1 3 ],\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 383, + 442, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $h_0$ and $h_{t + 1}$ denote masks, and $x_0$ the initial frame. In the auto-regressive variant, we use CosHand to generate video frames sequentially:", + "bbox": [ + 89, + 407, + 483, + 455 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {t + 1} = \\operatorname {C o s H a n d} \\left(\\hat {x} _ {t}, h _ {t}, h _ {t + 1}\\right), \\quad \\forall t \\in [ 0, 1 3 ],\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 464, + 442, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with $\\hat{x}_t$ being the previously generated frame and $\\hat{x}_0 = x_0$ .", + "bbox": [ + 89, + 489, + 480, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also compare against two language-instructed video generation methods: Seer [23] and DynamiCrafter [89], which we prompt with a first frame and its corresponding SSV2 class label as instruction. Note that DynamiCrafter was trained with a random video frame as conditioning, i.e. not always the first frame. At inference, its generated videos are thus not strictly a continuation of the conditioning, which precludes frame-aligned metrics. We compare InterDyn against these baselines in Tab. 1.", + "bbox": [ + 88, + 505, + 482, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative analysis. The frame-by-frame variant of CosHand achieves high image quality but struggles with temporal coherence and motion fidelity. Qualitatively, we notice that object locations are inconsistent across frames. In contrast, the auto-regressive variant improves object motion fidelity but suffers from lower frame-wise image quality due to error propagation. Both variants fail in scenarios requiring accurate post-interaction dynamics, such as when objects continue moving after being released from direct hand contact, as shown in Fig. 5.", + "bbox": [ + 88, + 643, + 482, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method, InterDyn $(256 \\times 384)$ , achieves the best overall performance, surpassing our baseline CosHand in spatio-temporal dynamics, motion fidelity, and all but two image quality metrics. We hypothesize that this might be due to two reasons (1) we use a frozen U-Net, while CosHand fine-tunes its model on SSV2, so CosHand might generate frames closer to the SSV2 distribution, and 2)", + "bbox": [ + 88, + 795, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "when SVD was trained, it was initialized with SD weights as spatial layers, and then fine-tuned over multiple stages; this might have degraded its spatial prior, and by extension the quality of produced images compared to CosHand.", + "bbox": [ + 511, + 282, + 903, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tuned on noisy masks and leveraging its temporal-aware control branch, InterDyn can interpret a noisy control sequence; e.g. when SAM2 produces a coarse and noisy hand mask sequence, InterDyn generates detailed hands including individual fingers, see Fig. 6. Though not always consistent, InterDyn is capable of depicting post-interaction dynamics, such as rolling or sliding objects.", + "bbox": [ + 511, + 343, + 906, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative analysis. We present diverse qualitative results generated by InterDyn in Fig. 7. Row 1 shows how InterDyn generates the articulated motion of an object. Row 2 showcases pouring water into a glass; note how the water level increases over time. Row 3 demonstrates an object being dropped, moving out of frame when falling, and rolling back in frame once hitting the floor, featuring realistic motion blur synthesis. Rows 4 and 5 illustrate how InterDyn handles squeezing interactions—the rubber and the spring are compressed and restored accordingly. Row 6 demonstrates an understanding of physical size and distance to the camera, as the phone moves closer to the viewer. These results highlight the complexity that InterDyn can handle, implying its generalization ability and future potential as an implicit, yet generalized physical simulator and renderer.", + "bbox": [ + 511, + 452, + 908, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 691, + 633, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced InterDyn, a framework that generates videos of interactive dynamics using large video generation models as implicit physics simulators. By incorporating an interactive control mechanism, InterDyn produces plausible, temporally consistent videos of object interactions—including complex human-object interactions—while generalizing to unseen objects. Our evaluations demonstrate that InterDyn effectively captures continuous motion and subsequent dynamics, outperforming baselines that focus on single future states. This work highlights the potential of using video generative models as physics simulators without explicit reconstruction, opening new avenues for future research.", + "bbox": [ + 511, + 717, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12474", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: We thank Benjamin Pellkofer for IT support, Peter Kulits for discussions, Tomasz Niewiadomski for coding support, and Allan D. Jepson, Artur Grigorev, Soubhik Sanyal, Omid Taheri & Angjoo Kanazawa for feedback. DT is supported by the ERC Starting Grant (project STRIPES, 101165317).", + "bbox": [ + 89, + 92, + 480, + 166 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Disclosure: https://files.is.tue.mpg.de/black/CoI_CVPR_2025.txt", + "bbox": [ + 89, + 167, + 482, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "DT has received a research gift fund from Google.", + "bbox": [ + 91, + 183, + 390, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 208, + 187, + 223 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A space-time diffusion model for video generation. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-11, 2024. 2, 3", + "[2] Fabien Baradel, Natalia Neverova, Julien Mille, Greg Mori, and Christian Wolf. CoPhy: Counterfactual learning of physical dynamics. In International Conference on Learning Representations (ICLR), 2020. 2, 3", + "[3] Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, et al. Interaction networks for learning about objects, relations and physics. In Conference on Neural Information Processing Systems (NeurIPS), 2016. 3", + "[4] Mikołaj Binkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In International Conference on Learning Representations (ICLR), 2018. 5", + "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable Video Diffusion: Scaling latent video diffusion models to large datasets. arXiv:2311.15127, 2023. 2, 3, 4, 5", + "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 22563-22575, 2023. 3", + "[7] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning to follow image editing instructions. In Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, 2023. 14", + "[8] Shengqu Cai, Duygu Ceylan, Matheus Gadelha, ChunHao Paul Huang, Tuanfeng Yang Wang, and Gordon Wetzstein. Generative rendering: Controllable 4D-guided video generation with 2D diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 7611-7620, 2024. 3", + "[9] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2D pose estimation using part affinity fields. In Computer Vision and Pattern Recognition (CVPR), pages 7291–7299, 2017. 14", + "[10] Junuk Cha, Jihyeon Kim, Jae Shin Yoon, and Seungryul Baek. Text2HOI: Text-guided 3D motion generation for hand-object interaction. In Computer Vision and Pattern Recognition (CVPR), pages 1577-1585, 2024. 3", + "[11] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S. Narang, Karl" + ], + "bbox": [ + 101, + 233, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Van Wyk, Umar Iqbal, Stan Birchfield, et al. DexYCB: A benchmark for capturing hand grasping of objects. In Computer Vision and Pattern Recognition (CVPR), pages 9044-9053, 2021. 14", + "[12] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv:2310.19512, 2023. 3", + "[13] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Ji-ashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable text-to-video generation with diffusion models. arXiv:2305.13840, 2023. 3", + "[14] Enric Corona, Albert Pumarola, Guillem Alenya, Francesc Moreno-Noguer, and Grégory Rogez. GanHand: Predicting human grasp affordances in multi-object scenes. In Computer Vision and Pattern Recognition (CVPR), pages 5031-5041, 2020. 3", + "[15] Christian Diller and Angela Dai. CG-HOI: Contact-guided 3D human-object interaction generation. In Computer Vision and Pattern Recognition (CVPR), pages 19888-19901, 2024. 3", + "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 2", + "[17] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In Computer Vision and Pattern Recognition (CVPR), pages 12943-12954, 2023. 3", + "[18] Zicong Fan, Maria Parelli, Maria Eleni Kadoglou, Xu Chen, Muhammed Kocabas, Michael J. Black, and Otmar Hilliges. HOLD: Category-agnostic 3D reconstruction of interacting hands and objects from video. In Computer Vision and Pattern Recognition (CVPR), pages 494–504, 2024. 3", + "[19] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. In International Conference on Robotics and Automation (ICRA), pages 12462-12469, 2024. 3", + "[20] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu Video: Factorizing text-to-video generation by explicit image conditioning. In European Conference on Computer Vision (ECCV), 2024. 3", + "[21] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The something something video database for learning and evaluating visual common sense. In International Conference on Computer Vision (ICCV), pages 5842-5850, 2017. 2, 4" + ], + "bbox": [ + 522, + 92, + 903, + 893 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12475", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Oliver Groth, Fabian B. Fuchs, Ingmar Posner, and Andrea Vedaldi. ShapeStacks: Learning vision-based physical intuition for generalised object stacking. In European Conference on Computer Vision (ECCV), pages 702-717, 2018. 3", + "[23] Xianfan Gu, Chuan Wen, Weirui Ye, Jiaming Song, and Yang Gao. Seer: Language instructed video prediction with latent diffusion models. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 8, 17", + "[24] Vincent Le Guen and Nicolas Thome. Disentangling physical dynamics from unknown factors for unsupervised video prediction. In Computer Vision and Pattern Recognition (CVPR), pages 11471-11481, 2020. 2, 3", + "[25] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Pengfei Wan, Di Zhang, Yufan Liu, Weiming Hu, Zhengjun Zha, et al. I2V-Adapter: A general image-to-video adapter for diffusion models. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2024. 3", + "[26] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. *AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning*. arXiv:2307.04725, 2023. 3", + "[27] Yana Hasson, Gül Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Computer Vision and Pattern Recognition (CVPR), pages 11807-11816, 2019. 3", + "[28] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from RGB videos. In International Conference on 3D Vision (3DV), pages 659-668, 2021. 3", + "[29] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. CameraCtrl: Enabling camera control for text-to-video generation. arXiv:2404.02101, 2024. 3", + "[30] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Conference on Neural Information Processing Systems (NeurIPS), pages 6629-6640, 2017. 5", + "[31] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 5", + "[32] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, et al. Image Video: High definition video generation with diffusion models. arXiv:2210.02303, 2022. 3", + "[33] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In Conference on Neural Information Processing Systems (NeurIPS), pages 8633-8646, 2022. 3", + "[34] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 3" + ], + "bbox": [ + 99, + 90, + 485, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate Anyone: Consistent and controllable image-to-video synthesis for character animation. In Computer Vision and Pattern Recognition (CVPR), pages 8153-8163, 2024. 3", + "[36] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. DepthCrafter: Generating consistent long depth sequences for open-world videos. arXiv:2409.02095, 2024. 4", + "[37] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make It Move: Controllable image-to-video generation with text descriptions. In Computer Vision and Pattern Recognition (CVPR), pages 18219-18228, 2022. 3", + "[38] Zhihao Hu and Dong Xu. VideoControlNet: A motion-guided video-to-video translation framework by using diffusion model with ControlNet. arXiv:2307.14073, 2023. 3", + "[39] Steeven Janny, Fabien Baradel, Natalia Neverova, Madiha Nadri, Greg Mori, and Christian Wolf. Filtered-CoPhy: Unsupervised learning of counterfactual physics in pixel space. In International Conference on Learning Representations (ICLR), 2022. 2, 3", + "[40] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In International Conference on Computer Vision (ICCV), pages 11107-11116, 2021. 3", + "[41] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker3: Simpler and better point tracking by pseudolabelling real videos. In European Conference on Computer Vision (ECCV), pages 18-35, 2024. 5", + "[42] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. DreamPose: Fashion image-to-video synthesis via stable diffusion. In International Conference on Computer Vision (ICCV), pages 22623-22633, 2023. 3", + "[43] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Conference on Neural Information Processing Systems (NeurIPS), 35:26565-26577, 2022. 5", + "[44] Hyeonwoo Kim, Sookwan Han, Patrick Kwon, et al. Beyond the Contact: Discovering comprehensive affordance for 3D objects from pre-trained 2D diffusion models. In European Conference on Computer Vision (ECCV), pages 400-419, 2024. 3", + "[45] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 5", + "[46] Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems. In International Conference on Machine Learning (ICML), pages 2693-2702, 2018. 2", + "[47] Sumith Kulal, Tim Brooks, Alex Aiken, Jiajun Wu, Jimei Yang, Jingwan Lu, Alexei A. Efros, and Krishna Kumar Singh. Putting people in their place: Affordance-aware human insertion into scenes. In Computer Vision and Pattern Recognition (CVPR), pages 17089-17099, 2023. 3" + ], + "bbox": [ + 522, + 90, + 906, + 897 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12476", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Adam Lerer, Sam Gross, and Rob Fergus. Learning physical intuition of block towers by example. In International Conference on Machine Learning (ICML), pages 430-438, 2016. 3", + "[49] Yunzhu Li, Antonio Torralba, Anima Anandkumar, Dieter Fox, and Animesh Garg. Causal discovery in physical systems from videos. Conference on Neural Information Processing Systems (NeurIPS), 2020. 2", + "[50] Ce Liu, Antonio Torralba, William T. Freeman, Frédo Durand, and Edward H. Adelson. Motion magnification. Transactions on Graphics (TOG), 24(3):519-526, 2005. 5", + "[51] Ruibo Liu, Jason Wei, Shixiang Shane Gu, Te-Yen Wu, Soroush Vosoughi, Claire Cui, Denny Zhou, and Andrew M Dai. Mind's Eye: Grounded language model reasoning through simulation. In International Conference on Learning Representations (ICLR), 2023. 3", + "[52] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. arXiv:2303.05499, 2023. 14", + "[53] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. PhysGen: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision (ECCV), pages 360-378, 2024. 3", + "[54] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv:2402.17177, 2024. 2, 4", + "[55] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T. Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. Transactions on Graphics (TOG), 39(6): 1-14, 2020. 3", + "[56] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T. Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In Computer Vision and Pattern Recognition (CVPR), pages 4507-4515, 2021. 3", + "[57] Jiaxin Lu, Hao Kang, Haoxiang Li, Bo Liu, Yiding Yang, Qixing Huang, and Gang Hua. UGG: Unified generative grasping. In European Conference on Computer Vision (ECCV), pages 414-433, 2024. 3", + "[58] Wan-Duo Kurt Ma, John P Lewis, and W. Bastiaan Kleijn. TrailBlazer: Trajectory control for diffusion-based video generation. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-11, 2023. 3", + "[59] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-Else: Compositional action recognition with spatial-temporal interaction networks. In Computer Vision and Pattern Recognition (CVPR), pages 1049–1059, 2020. 6", + "[60] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to" + ], + "bbox": [ + 99, + 92, + 485, + 893 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image diffusion models. In AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 3", + "[61] Franziska Mueller, Micah Davis, Florian Bernard, Oleksandr Sotnychenko, Micekal Verschooor, Miguel A. Otaduy, Dan Casas, and Christian Theobalt. Real-time pose and shape reconstruction of two interacting hands with a single depth camera. Transactions on Graphics (TOG), 38(4): 1-13, 2019. 14", + "[62] Boxiao Pan, Zhan Xu, Chun-Hao Paul Huang, Krishna Kumar Singh, Yang Zhou, Leonidas J. Guibas, and Jimei Yang. ActAnywhere: Subject-aware video background generation. In Conference on Neural Information Processing Systems (NeurIPS), 2024. 3", + "[63] Georgios Paschalidis, Romana Wilschut, Dimitrije Antic, Omid Taheri, and Dimitrios Tzionas. 3D whole-body grasp synthesis with directional controllability. In International Conference on 3D Vision (3DV), 2025. 3", + "[64] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 4", + "[65] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Girshick, Piotr Dólar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 6, 14", + "[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 4, 14", + "[67] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. Transactions on Graphics (TOG), 36(6), 2022. 14", + "[68] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-A-Video: Text-to-video generation without text-video data. In International Conference on Learning Representations (ICLR), 2023. 3", + "[69] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. StyleGAN-V: A continuous video generator with the price, image quality and perks of StyleGAN2. In Computer Vision and Pattern Recognition (CVPR), pages 3626-3636, 2022. 5", + "[70] Sruthi Sudhakar, Ruoshi Liu, Basile Van Hoorick, Carl Vondrick, and Richard Zemel. Controlling the world by sleight of hand. In European Conference on Computer Vision (ECCV), pages 414-430, 2024. 2, 3, 6, 8, 14, 17", + "[71] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), pages 581-600, 2020. 3" + ], + "bbox": [ + 522, + 92, + 906, + 895 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12477", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J. Black. GRIP: Generating interaction poses using latent consistency and spatial cues. In International Conference on 3D Vision (3DV), pages 933-943, 2024. 3", + "[73] Tze Ho Elden Tse, Kwang In Kim, Ales Leonardis, and Hyung Jin Chang. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In Computer Vision and Pattern Recognition (CVPR), pages 1664-1674, 2022. 3", + "[74] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation. In International Conference on Learning Representations Workshops (ICLRw), 2019. 5", + "[75] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. In European Conference on Computer Vision (ECCV), pages 439-457, 2025. 4, 14", + "[76] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv:2308.06571, 2023. 3", + "[77] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. In International Conference on Machine Learning (ICML), 2024. 3", + "[78] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional video synthesis with motion controllability. In Conference on Neural Information Processing Systems (NeurIPS), 2024. 3", + "[79] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image Quality Assessment: From error visibility to structural similarity. Transactions on Image Processing (TIP), 13(4):600-612, 2004. 3, 5", + "[80] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A unified and flexible motion controller for video generation. In Transactions on Graphics (TOG), pages 1-11, 2024. 3", + "[81] Nicholas Watters, Daniel Zoran, Theophane Weber, Peter Battaglia, Razvan Pascanu, and Andrea Tacchetti. Visual Interaction Networks: Learning a physics simulator from video. In Conference on Neural Information Processing Systems (NeurIPS), pages 4539-4547, 2017. 3", + "[82] Jiajun Wu, Ilker Yildirim, Joseph J. Lim, Bill Freeman, and Joshua B. Tenenbaum. Galileo: Perceiving physical object properties by integrating a physics engine with deep learning. In Conference on Neural Information Processing Systems (NeurIPS), pages 127-135, 2015. 3", + "[83] Jiajun Wu, Joseph J. Lim, Hongyi Zhang, Joshua B. Tenenbaum, and William T. Freeman. Physics 101: Learning physical object properties from unlabeled videos. In *British Machine Vision Conference (BMVC)*, 2016.", + "[84] Jiajun Wu, Erika Lu, Pushmeet Kohli, Bill Freeman, and Josh Tenenbaum. Learning to see physics via visual de" + ], + "bbox": [ + 99, + 92, + 482, + 909 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "animation. In Conference on Neural Information Processing Systems (NeurIPS), pages 153-164, 2017. 2, 3", + "[85] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-A-Video: One-shot tuning of image diffusion models for text-to-video generation. In International Conference on Computer Vision (ICCV), pages 7623–7633, 2023. 3", + "[86] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. THOR: Text to human-object interaction diffusion via relation intervention. arXiv:2403.11208, 2024. 3", + "[87] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. DragAnything: Motion control for anything using entity representation. In European Conference on Computer Vision (ECCV), pages 331-348, 2025. 3", + "[88] Xianghui Xie, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Template free reconstruction of human-object interaction with procedural interaction generation. In Computer Vision and Pattern Recognition (CVPR), pages 10003-10015, 2024. 3", + "[89] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. DynamiCrafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision (ECCV), pages 399-417, 2025. 2, 3, 8, 17", + "[90] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3D human-object interactions with physics-informed diffusion. In International Conference on Computer Vision (ICCV), pages 14928-14940, 2023. 3", + "[91] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. MagicAnimate: Temporally consistent human image animation using diffusion model. In Computer Vision and Pattern Recognition (CVPR), pages 1481–1490, 2024. 3", + "[92] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. CogVideoX: Text-to-video diffusion models with an expert transformer. arXiv:2408.06072, 2024.3", + "[93] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In Computer Vision and Pattern Recognition (CVPR), pages 8466-8476, 2024. 5, 8, 14, 17", + "[94] Yufei Ye, Poorvi Hebbar, Abhinav Gupta, and Shubham Tulsiani. Diffusion-guided reconstruction of everyday hand-object interaction clips. In International Conference on Computer Vision (ICCV), pages 19717-19728, 2023. 3", + "[95] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance Diffusion: Synthesizing hand-object interactions. In Computer Vision and Pattern Recognition (CVPR), pages 22479-22489, 2023. 3" + ], + "bbox": [ + 524, + 92, + 903, + 893 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12478", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[96] Yufei Ye, Abhinav Gupta, Kris Kitani, and Shubham Tulsiani. G-HOP: Generative hand-object prior for interaction reconstruction and grasp synthesis. In Computer Vision and Pattern Recognition (CVPR), pages 1911–1920, 2024. 3", + "[97] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Ji-jun Wu, Antonio Torralba, and Joshua B. Tenenbaum. CLEVRR: Collision events for video representation and reasoning. In International Conference on Learning Representations (ICLR), 2020. 2, 4, 6", + "[98] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. DragNUWA: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv:2308.08089, 2023. 3", + "[99] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Computer Vision and Pattern Recognition (CVPR), pages 18456-18466, 2023. 3", + "[100] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3, 4", + "[101] Mengqi Zhang, Yang Fu, Zheng Ding, Sifei Liu, Zhuowen Tu, and Xiaolong Wang. HOIDiffusion: Generating realistic 3D hand-object interaction data. In Computer Vision" + ], + "bbox": [ + 93, + 90, + 485, + 424 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Pattern Recognition (CVPR), pages 8521-8531, 2024. 3", + "[102] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5", + "[103] Juntian Zheng, Qingyuan Zheng, Lixing Fang, Yun Liu, and Li Yi. CAMS: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis. In Computer Vision and Pattern Recognition (CVPR), pages 585-594, 2023. 3", + "[104] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient video generation with latent diffusion models. arXiv:2211.11018, 2022.3", + "[105] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. GEARS: Local geometry-aware hand-object interaction synthesis. In Computer Vision and Pattern Recognition (CVPR), pages 20634-20643, 2024. 3", + "[106] Tianqiang Zhu, Rina Wu, Xiangbo Lin, and Yi Sun. Toward human-like grasp: Dexterous grasping via semantic representation of object-hand. In International Conference on Computer Vision (ICCV), pages 15741-15751, 2021. 3" + ], + "bbox": [ + 516, + 92, + 903, + 416 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "12479", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/1657842e-f939-4237-87d3-49f995d53e68_model.json b/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/1657842e-f939-4237-87d3-49f995d53e68_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c3c78b3eea1a1923b0aa826e4dd07f8607b5bab2 --- /dev/null +++ b/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/1657842e-f939-4237-87d3-49f995d53e68_model.json @@ -0,0 +1,2646 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.127, + 0.872, + 0.15 + ], + "angle": 0, + "content": "InterDyn: Controllable Interactive Dynamics with Video Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.174, + 0.899, + 0.192 + ], + "angle": 0, + "content": "Rick Akkerman\\(^{1,2*}\\) Haiwen Feng\\(^{1*†}\\) Michael J. Black\\(^{1}\\) Dimitrios Tzionas\\(^{2}\\) Victoria Fernández Abrevaya\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.194, + 0.837, + 0.209 + ], + "angle": 0, + "content": "\\(^{1}\\)Max Planck Institute for Intelligent Systems, Tübingen, Germany \\(^{2}\\)University of Amsterdam, the Netherlands" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.212, + 0.896, + 0.228 + ], + "angle": 0, + "content": "{rick.akkerman, haiwen.feng, black, victoria.abrevaya}@tuebingen.mpg.de d.tzionas@uva.nl" + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.248, + 0.193, + 0.261 + ], + "angle": 0, + "content": "Input image" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.264, + 0.226, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.248, + 0.806, + 0.262 + ], + "angle": 0, + "content": "Video generation by InterDyn using only the hand mask sequence as control signal" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.264, + 0.903, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.346, + 0.226, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.345, + 0.903, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.441, + 0.34, + 0.454 + ], + "angle": 0, + "content": "Force propagation" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.454, + 0.483, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.44, + 0.776, + 0.452 + ], + "angle": 0, + "content": "Counterfactual dynamics" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.454, + 0.902, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.189, + 0.535, + 0.78, + 0.548 + ], + "angle": 0, + "content": "Denotes a driving object motion | Tracks indicate an object with generated uncontrolled dynamics" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.565, + 0.908, + 0.622 + ], + "angle": 0, + "content": "Figure 1. We present InterDyn, a framework for synthesizing realistic interactive dynamics without 3D reconstruction and physics simulation. Our core principle is to rely on the implicit physics knowledge embedded in large-scale video generative models. Given an image and a \"driving motion\", our model generates the consequential scene dynamics. We investigate the generated interactive dynamics in a simple object collision scenario (bottom) and complex in-the-wild human-object interaction (top)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.646, + 0.327, + 0.661 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.679, + 0.486, + 0.862 + ], + "angle": 0, + "content": "Predicting the dynamics of interacting objects is essential for both humans and intelligent systems. However, existing approaches are limited to simplified, toy settings and lack generalizability to complex, real-world environments. Recent advances in generative models have enabled the prediction of state transitions based on interventions, but focus on generating a single future state which neglects the continuous dynamics resulting from the interaction. To address this gap, we propose InterDyn, a novel framework that generates videos of interactive dynamics given an initial frame and a control signal encoding the motion of a driving object or actor. Our key insight is that large video" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.909, + 0.86 + ], + "angle": 0, + "content": "generation models can act as both neural renderers and implicit physics \"simulators\", having learned interactive dynamics from large-scale video data. To effectively harness this capability, we introduce an interactive control mechanism that conditions the video generation process on the motion of the driving entity. Qualitative results demonstrate that InterDyn generates plausible, temporally consistent videos of complex object interactions while generalizing to unseen objects. Quantitative evaluations show that InterDyn outperforms baselines that focus on static state transitions. This work highlights the potential of leveraging video generative models as implicit physics engines. Code and trained models will be released at: https://interdyn.is.tue.mpg.de/." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.875, + 0.223, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.888, + 0.187, + 0.9 + ], + "angle": 0, + "content": "†Project lead" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.875, + 0.223, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12467" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.266 + ], + "angle": 0, + "content": "Humans have the remarkable ability to predict the future dynamics of observed systems intuitively. With just a single image, we can anticipate and imagine how objects will move over time – not only their motion but also their interactions with the environment and other elements in the scene. Inferring this requires an advanced form of scene-level reasoning beyond merely recognizing the semantics and geometry of static elements; it involves a deep physical and causal understanding of how each object will interact given the environment, object properties, and forces." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.484, + 0.478 + ], + "angle": 0, + "content": "There has been a growing interest in developing machine learning systems that emulate similar levels of dynamic understanding given visual observations, such as images or videos. Early work [84] addressed this by first reconstructing a 3D representation from the image, then predicting future states with a physics simulator and finally generating the output video with a rendering engine. This relies heavily on explicit reconstruction and simulation, which is computationally intensive, prone to errors, and may not generalize well. More recent methods [2, 24, 39, 46, 49] leverage keypoint or latent representations within graph relational frameworks; however, they have only been trained and validated in over-simplified, synthetic environments, showing limited generalizability to complex real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.841 + ], + "angle": 0, + "content": "Instead, the advent of powerful generative models [1, 5, 16, 54, 66] opens new avenues for synthesizing interactions under complex scenarios. For example, Sudhakar et al. [70] recently proposed CosHand, a controllable image-to-image model based on Stable Diffusion [66] that infers state transitions of an object. The task here is defined as follows: given an image of a hand interacting with an object, alongside a hand mask of the current frame and a mask of the hand at a future frame, generate a modified input image that satisfies the mask, with realistic interactions. The challenge, as in early intuitive physics works, lies in accurately modeling how the objects will change after forces are applied. However, we argue that static state transitions are insufficient for this task, as they fail to capture the continuous dynamic processes inherent to the problem, e.g. see Fig. 2. Investigating interactive dynamics within a two-state setting is highly limiting, since dynamics can extend beyond the period of direct contact – for example, predicting the motion occurring while a person pours water requires a physical understanding that goes beyond the state of the hand at a future frame. The driving force, in this case the hand, may interact with the system only briefly, but the system's subsequent dynamics continue according to physical laws and may even influence other parts via force propagation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this paper, we explore controllable synthesis of interactive dynamics—generating a video from an input image and a dynamic control signal (e.g. a moving hand mask) to model realistic object dynamics. In particular, we propose" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.906, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.234, + 0.907, + 0.305 + ], + "angle": 0, + "content": "Figure 2. State transition vs. dynamics. Methods that generate static state transitions (i.e. predict a future image) such as CosHand [70] struggle to capture the inherent dynamic processes involved in human-object interactions. Here, we show a video sequence where the motion continues beyond the interaction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.474 + ], + "angle": 0, + "content": "InterDyn, a novel framework for synthesizing controllable dynamic interactions that leverages the physical and dynamics \"knowledge\" of a large video generation model [5]. Unlike prior approaches that rely on explicit physical simulation [84] or are constrained to static state transitions [70], we leverage video generation models to generate dynamic processes implicitly, see Fig. 1. Specifically, we extend Stable Video Diffusion (SVD) [5] with a dynamic control branch and fine-tune it on diverse scenes, enabling synthesis of complex interactions aligned with the control signal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.909, + 0.657 + ], + "angle": 0, + "content": "We start our investigation by fine-tuning InterDyn on a simple synthetic scenario of cubes, cylinders, and spheres: the CLEVRER dataset [97]. To control the motion we add a mask driving signal that manipulates the movement of some (but not all) of the objects in the scene. We then evaluate how the synthesized trajectories of uncontrolled objects change under various interactions, including multiple objects colliding with each other. This multi-object collision setting allows us to \"probe\" the physical understanding and causal effects of the video diffusion model, and our qualitative experiments show InterDyn's ability for counterfactual future prediction and physical force propagation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Next, we evaluate how the system performs in a difficult real-world scenario, such as Human-Object Interaction (HOI). Here, the dexterity of hand motions and the diversity of objects vastly increase the complexity of the problem. We fine-tune the model on a commonly used HOI video dataset [21] and compare with the state-of-the-art baseline CosHand [70], as well as two text-control based interactive dynamics generation methods: Seer [23] and DynamiCrafter [89]. We quantify our investigations using standard image and video metrics, as well as a motion fidelity metric based on point tracking. InterDyn surpasses the previous SOTA over \\(37.5\\%\\) on LPIPS and \\(77\\%\\) on FVD on the Something-Something-v2 (SSV2) dataset [21]. Our experiments also demonstrate diverse and physically plausible generations of interactive dynamics, probing into SVD's \"understanding\" of physics and dynamics." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12468" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.244 + ], + "angle": 0, + "content": "In summary, we present InterDyn, a framework that employs video generative models to simulate object dynamics without explicit 3D reconstruction or physical simulation. We demonstrate how the inherent \"knowledge\" within video generation models can be leveraged to predict complex object interactions and movements over time, implicitly modeling physical and causal dynamics. We perform comprehensive experiments on multi-object collision datasets and hand-object manipulation datasets, demonstrating the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.263, + 0.233, + 0.279 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.29, + 0.485, + 0.697 + ], + "angle": 0, + "content": "Modeling human-object interactions (HOI). Human-object interaction has been widely studied within the context of 3D reconstruction [17, 18, 27, 28, 73, 88, 94], where the goal is to recover realistic geometry of hands and objects. The field of 3D HOI synthesis has also received increasing attention, including the generation of static [40, 44, 71, 106] or dynamic [63, 72, 103, 105] hand poses conditioned on 3D objects, whole-body interactions [90], or more recently, hand-object meshes given textual descriptions [10, 15, 57, 86, 96]. Few works address HOI synthesis in the 2D domain. GANHand [14] predicts 3D hand shape and pose given an RGB image of an object, while AffordanceDiffusion [95] estimates a 2D hand using a diffusion model. Kulal et al. [47] take as input an image of a human and a scene separately and generate a composite image that positions the human with correct affordances. Also relevant is HOIDiffusion [101], in which a texture-less rendering of a 3D hand and object is converted to a realistic image using a text description. Most closely related to us is CosHand [70], which takes as input an RGB image of a hand-object interaction, a hand mask at the current state, and the hand mask of the future state, and generates an RGB image of the future state. Unlike us, they cannot generate post-interaction object dynamics. Importantly, none of these works study dynamics, generating instead discrete state transitions that fail to capture the nuanced, temporally coherent behaviors observed in interactions." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.72 + ], + "angle": 0, + "content": "Synthesizing causal physical relations from visual input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "A growing body of work aims to model and predict physical causal effects from visual inputs such as images or videos. For example, research in intuitive physics seeks to replicate the human-like, non-mathematical understanding of physical events, e.g. by predicting future frames given an input video. Early works like [22, 48] train neural networks to assess the stability of block towers, while [24] leverage prior physical knowledge formalized through partial differential equations (PDEs). Other approaches investigate counterfactual reasoning by leveraging graph neural networks [2, 39]. Wu et al. [82-84] explore the use of an inverse rendering approach, extracting geometry and physical properties from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "the video which are then coupled with a physics simulator and a rendering engine to generate the future frames. Other works [81] incorporate Interaction Networks [3] to approximate physical systems from video data. These approaches are often limited to simplified, synthetic datasets and struggle to generalize to real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.195, + 0.908, + 0.542 + ], + "angle": 0, + "content": "Recent methods have started to combine language models with physical engines. Liu et al. [51] ground a large language model using a computational physics engine while Gao et al. [19] show that fine-tuning a vision-language model (VLM) on annotated datasets of physical concepts improves its understanding of physical interactions. Closely related to our work is PhysGen [53], which trains an imaged-to-video model that conditions the video generation on physics parameters (e.g., force or torque). However, the model relies on a dynamics simulator to generate motion, and its application is limited to rigid objects. A related but tangential line of work focuses on identifying and generating the effects of objects on their surroundings. For example, Omnimatte [56] introduces the problem of identifying all parts of a scene influenced by an object, given a video and a mask of the object. Similarly, Lu et al. [55] propose to re-time the motion of different subjects in a scene while maintaining realistic interactions with the environment. ActAnywhere [62] generates videos with plausible human-scene interactions, taking a masked video of a person and a background image as input. These works address the problem of synthesizing realistic interactions within a scene, however, lack fine-grained control." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Controllable video generation. Video generation has advanced significantly in recent years, with diffusion models leading to substantial improvements in unconditional [33, 99], text-based [1, 5, 6, 12, 20, 23, 26, 32, 37, 68, 76, 85, 89, 92, 104] and image-based [1, 5, 20, 25, 78, 89] generation. These advances have raised the question of how to incorporate more nuanced control into video generation. Some text-to-video approaches are trained by \"inflating\" text-to-image (T2V) models [8, 13, 25, 26, 38, 85], and can thus be integrated with conditional T2V models such as ControlNet [100] or T2V-Adapter [60]. Control can also be achieved by conditioning on trajectories [58, 87, 98] or bounding-boxes [77], fine-tuning on appropriate datasets. VideoComposer [78] incorporates multiple condition types, including text, depth, style, and temporal conditions via motion vectors. Camera motion control has also been explored, withAnimateDiff [79] employing LoRA [34] modules to control camera movement, while MotionCtrl [80] and CameraCtrl [29] directly embed the camera information for more precise control. Additionally, several works target human animation from a pose control signal, such as DreamPose [42], MagicPose [91], andAnimateAnyone [35], but do not account for interactions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12469" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.406, + 0.108 + ], + "angle": 0, + "content": "3. Controllable Interactive Dynamics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.268 + ], + "angle": 0, + "content": "Video diffusion models such as [5, 54] have demonstrated impressive performance in generating videos from text or images, and have even shown potential in tasks that require 3D understanding when properly fine-tuned [36, 75]. Trained on millions of videos, we hypothesize that these models also possess implicit knowledge of complex interactive dynamics, such as those that appear when humans interact with objects. Out of the box, however, they lack a precise control mechanism, often relying solely on textual inputs or requiring careful selection of the starting frame." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.271, + 0.484, + 0.439 + ], + "angle": 0, + "content": "Task. Given an input image, \\( \\pmb{x} \\in \\mathbb{R}^{1 \\times H \\times W \\times 3} \\), and a driving motion in the form of a pixel-wise corresponding control signal \\( \\pmb{c} \\in \\mathbb{R}^{N \\times H \\times W \\times 3} \\), we task InterDyn with generating a video sequence, \\( \\pmb{y} \\in \\mathbb{R}^{N \\times H \\times W \\times 3} \\), depicting plausible object dynamics. Through this task, we aim to learn the conditional distribution between a driving motion, such as that of a human hand, and the consequent motion of manipulated objects. In other words, the model needs to synthesize plausible object movement and appearance without any indication other than the driving motion, while maintaining physical and visual consistency with the input image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.484, + 0.685 + ], + "angle": 0, + "content": "Stable Video Diffusion. We extend Stable Video Diffusion [5] (SVD) to enable controllable interactive dynamics and explore the versatility of this model across a range of scenarios. SVD is a publicly available U-Net-based latent diffusion model [66] that extends Stable Diffusion 2.1 to video generation by interleaving the network with temporal layers. Given a static input image of a scene, SVD denoises a sequence of \\(N\\) frames \\(\\pmb{y} \\in \\mathbb{R}^{N \\times H \\times W \\times 3}\\) to generate a video that follows the initial frame. The input image is fed into the denoising U-Net by concatenating its latent to each of the frames in the noised input, and by supplying its CLIP [64] embedding to the U-Net's cross-attention layers. In addition, SVD is conditioned on the video's FPS and motion ID, where the motion ID represents the amount of motion in the video. We found a motion ID of 40 to align well with our frozen SVD prior." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Control. InterDyn extends SVD with an additional control signal \\( c \\in \\mathbb{R}^{N \\times H \\times W \\times 3} \\) by integrating a ControlNet-like branch [100]. An overview of our pipeline is presented in Fig. 3. The SVD weights remain frozen to preserve its learned dynamics prior. Following [100], we introduce a trainable copy of the SVD encoder \\( E \\), connected to SVD's frozen decoder via skip connections, and modulated by zero-initialized convolutions. We use a small CNN, \\( \\mathcal{E}(\\cdot) \\), to encode the control signal \\( c \\) into the latent space, which is then added to the noisy input latent that is passed to the ControlNet encoder. Similar to SVD, the control branch interleaves convolutional, spatial, and temporal blocks, enabling InterDyn to process the control signal in a temporal-aware manner. This helps InterDyn to be robust to" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.906, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.27, + 0.908, + 0.41 + ], + "angle": 0, + "content": "Figure 3. Overview of InterDyn. Given an input image depicting a scene, such as a hand holding a remote, and a \"driving motion,\" such as a sequence of binary hand masks, InterDyn generates a video depicting plausible hand and object dynamics. Crucially, InterDyn receives no control signal for the object. Through this setup, we probe and assess the implicit knowledge of large video generation models on complex interactive dynamics. We use Stable Video Diffusion (SVD) as our frozen backbone and fine-tune a separate control signal encoder. Videos are iteratively denoised over \\( T \\) timesteps, starting from Gaussian noise \\( \\epsilon \\sim \\mathcal{N}(0,I) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.435, + 0.907, + 0.526 + ], + "angle": 0, + "content": "noisy control signals, see Fig. 6. We opt for binary masks as conditioning signal due to their accessibility. However, our method can be extended to incorporate diverse types of signals. We find that for hand-object interactions, the type of conditioning signal does not significantly impact performance, see Appendix A and Tab. S1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.906, + 0.622 + ], + "angle": 0, + "content": "Inference. During inference, we start from an input image, control signal sequence, and randomly sampled Gaussian noise \\(\\epsilon \\sim \\mathcal{N}(0,I)\\). Through iteratively applying InterDyn over \\(T\\) denoising timesteps, we generate a video \\(\\pmb{y}\\) depicting plausible hand and object dynamics, aligned with the control signal." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.646, + 0.651 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The primary goal of this work is to synthesize scene-level interactive dynamics by leveraging the implicit physical understanding of a pre-trained video generative model. We begin by probing the model's ability to predict physically plausible outcomes within simulated environments, specifically using the CLEVRER dataset [97]. We test force propagation amongst uncontrolled objects and examine counterfactual future prediction by generating videos for one input image with different control signals. Motivated by promising results, we extend our investigation to complex, real-world hand-object interaction scenarios using the Something-Something-v2 (SSV2) dataset [21], conducting comprehensive comparisons with existing baselines that pursue similar objectives. Additionally, we showcase diverse physical examples to demonstrate the capabilities of InterDyn in generating realistic interactive dynamics." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12470" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.907, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.183, + 0.878, + 0.198 + ], + "angle": 0, + "content": "(a) Force Propagation. InterDyn generates force propagation to and amongst uncontrolled objects (cyan cylinder, pink cylinder, and pink cube)." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.203, + 0.907, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.366, + 0.903, + 0.381 + ], + "angle": 0, + "content": "(b) Counterfactual Dynamics. Two generations of InterDyn on the same input image: (1) w/ brown-cylinder control and (2) w/o brown-cylinder control." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.392, + 0.907, + 0.449 + ], + "angle": 0, + "content": "Figure 4. Qualitative investigation on the CLEVRER dataset. Given an input image and the \"driving\" motion of one or two objects, our model predicts the future interactive dynamics of multiple elements in the scene. The driving motion is given in the form of semantic mask sequences. The generated object motions are highlighted with a red-line trajectory. Note that our model can generate videos with force propagation across multiple uncontrolled objects (top) and can generate multiple futures (bottom). Q Zoom in for details." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.461, + 0.251, + 0.478 + ], + "angle": 0, + "content": "4.1. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.483, + 0.484, + 0.651 + ], + "angle": 0, + "content": "We initialize InterDyn with the 14-frame image-to-video weights of SVD [5]. During training, we use the Adam optimizer [45] with a learning rate of \\(1 \\times 10^{-5}\\). We use the EDM framework [43] with noise distribution \\(\\log \\sigma \\sim \\mathcal{N}(0.7, 1.6^2)\\). We train on two 80GB H100 GPUs, with a per-GPU batch size of 4. Video length and FPS define the temporal resolution of dynamics; to balance short- and long-range events we subsample videos to 7 FPS. To facilitate classifier-free guidance [31], we randomly drop the input image with a probability of \\(5\\%\\). At inference, we apply the Euler scheduler [43] over 50 denoising timesteps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.659, + 0.188, + 0.673 + ], + "angle": 0, + "content": "4.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.484, + 0.743 + ], + "angle": 0, + "content": "We evaluate InterDyn on image quality, spatio-temporal similarity, and motion fidelity. Image quality metrics are computed frame-wise. All metrics are reported excluding the first frame, as it serves as input conditioning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.484, + 0.822 + ], + "angle": 0, + "content": "Image quality. We report the Structural Similarity Index Measure (SSIM) [79], Peak Signal-to-Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS) [102], Fréchet Inception Distance (FID) [30] and unbiased Kernel Inception Distance (KID) [4]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Spatio-temporal similarity. To assess the spatio-temporal perceptual similarity between the ground truth and the generated video distributions, we use the Fréchet Video Distance (FVD) and unbiased Kernel Video Distance (KVD) proposed in [74]. We use the implementation of [69]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Motion Fidelity. Through InterDyn, we do not have explicit control over object dynamics, which means that the pixel alignment of an object in the generated and ground truth video is only guaranteed in the starting frame. In this case, comparing generated object motion to the ground truth naively might misrepresent the true quality of object motion over time. Therefore, we adapt the Motion Fidelity (MF) metric proposed by Yatim et al. [93], which measures the similarity between point-tracking trajectories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.69 + ], + "angle": 0, + "content": "To compute the metric for any video, we obtain a mask of the object in the starting frame, sample 100 points on the object, and track these throughout both the ground truth and generated video using CoTracker3 [41]. Given the resulting two sets of tracklets \\(\\mathcal{T} = \\{\\tau_1,\\dots ,\\tau_n\\}\\), \\(\\tilde{\\mathcal{T}} = \\{\\tilde{\\tau}_1,\\dots ,\\tilde{\\tau}_m\\}\\) the motion fidelity metric is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.702, + 0.907, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{m} \\sum_ {\\tilde {\\tau} \\in \\tilde {\\mathcal {T}}} \\max _ {\\tau \\in \\mathcal {T}} \\mathbf {c o r r} (\\tau , \\tilde {\\tau}) + \\frac {1}{n} \\sum_ {\\tau \\in \\mathcal {T}} \\max _ {\\tilde {\\tau} \\in \\tilde {\\mathcal {T}}} \\mathbf {c o r r} (\\tau , \\tilde {\\tau}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.755, + 0.899, + 0.772 + ], + "angle": 0, + "content": "with the correlation between two tracklets \\(\\mathbf{corr}(\\tau, \\tilde{\\tau})\\) [50]:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.784, + 0.907, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c o r r} (\\tau , \\tilde {\\tau}) \\frac {1}{F} \\sum_ {k = 1} ^ {F} \\frac {v _ {k} ^ {x} a \\cdot \\tilde {v} _ {k} ^ {x} + v _ {k} ^ {y} \\cdot \\tilde {v} _ {k} ^ {y}}{\\sqrt {(v _ {k} ^ {x}) ^ {2} + (v _ {k} ^ {y}) ^ {2}} \\cdot \\sqrt {(\\tilde {v} _ {k} ^ {x}) ^ {2} + (\\tilde {v} _ {k} ^ {y}) ^ {2}}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.839, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\((v_k^x, v_k^y)\\), \\((\\tilde{v}_k^x, \\tilde{v}_k^y)\\) are the \\(k^{th}\\) frame displacement of tracklets \\(\\tau, \\tilde{\\tau}\\) respectively. If there are less than 100 points to query on the object due to it being too small, we do not consider the video for the motion fidelity metric." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12471" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.482, + 0.108 + ], + "angle": 0, + "content": "4.3. Probing Dynamics with Object Collision Events" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.483, + 0.219 + ], + "angle": 0, + "content": "Here, we fine-tune InterDyn on an object collision dataset to probe its ability to generate realistic object interactions. Qualitatively, we review whether InterDyn can produce plausible object motion for uncontrolled objects, given the motion of objects entering the scene. In addition, we examine whether InterDyn can generate counterfactual videos for the same input image, but different control signals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.223, + 0.484, + 0.39 + ], + "angle": 0, + "content": "Dataset. We use CLEVRER [97], which provides 20,000 videos of colliding objects with annotated segmentation masks and metadata on collision events. We construct a control signal for objects entering the scene and aim to use InterDyn to generate the motion of the objects that are already present, upon collision. Stationary objects do not receive any form of control signal. Colored masks help the model distinguish unique objects. The frames are cropped and scaled to \\(320 \\times 448\\), and we only sample input frames before collisions between objects in the scene, to maximize InterDyn's exposure to interactive dynamics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.56 + ], + "angle": 0, + "content": "Force propagation. InterDyn can generate force propagation dynamics between a controlled object and an uncontrolled object, as well as amongst uncontrolled objects, as illustrated in Fig. 4a. Here, the red cylinder at the top left is the driving force. It collides with the uncontrolled blue cylinder, which then collides with the uncontrolled purple cylinder, in turn striking the uncontrolled purple cube on the far right. Point-tracking trajectories display how collisions alter each object's position. This suggests that InterDyn possesses an implicit understanding of physical interactions, enabling it to generate plausible dynamics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Counterfactual dynamics. By altering the control signal, InterDyn can simulate counterfactual scenarios for the same input image, as shown in Fig. 4b. In \"Generation 1\", the gray cylinder (controlled) collides with the stationary red sphere (uncontrolled), causing it to move; it is later struck by the brown cylinder (controlled), altering its path once again. In \"Generation 2\", removing the brown cylinder lets the red sphere continue along its original trajectory, consistent with expectations. Crucially, there is no control signal for the red sphere throughout the sequence; its motion and trajectory are entirely generated by InterDyn." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.797 + ], + "angle": 0, + "content": "Probing InterDyn on CLEVRER highlights its ability to generate interactive dynamics for objects within a simple synthetic environment. We provide additional results in video format on our webpage." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.433, + 0.82 + ], + "angle": 0, + "content": "4.4. Generating Human-Object Interactions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.903 + ], + "angle": 0, + "content": "For this experiment, we fine-tune InterDyn on a human-action dataset, focused on hand-object interaction. We encode human movement as a sequence of pixel-aligned binary hand masks and task InterDyn to generate a video with hand movements and corresponding object dynamics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "Dataset. We use Something-Something-v2 (SSV2), which provides 220,847 videos of humans performing basic actions with everyday objects. It contains actions such as \"pushing [something] from left to right\", \"squeezing [something]\" and \"lifting [something] with [something] on it\". This dataset allows us to train InterDyn at a larger scale and compare with our baseline CosHand [70]. We train one version of InterDyn at the same resolution as CosHand, \\(256 \\times 256\\), and a second version at \\(256 \\times 384\\), which aligns better with the dynamic prior of SVD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.907, + 0.486 + ], + "angle": 0, + "content": "We generate binary hand masks by prompting Segment Anything 2 (SAM2) [65] with hand bounding boxes, provided by the Something-Else dataset [59]. Similar to our baseline CosHand, we exclude the \"pretending\" class. Additionally, we remove all \"[something] is done behind [something]\" classes since the object would be out of view of the camera. We remove videos smaller than the target resolution/length, ensure the hand and objects are continuously visible, and crop videos larger than the target resolution while centering the object. We include videos without obvious motion and state transition, such as \"holding [something]\". Since bounding box annotations are only provided for the train split (79,043 samples for \\(256 \\times 384\\) and 104,260 for \\(256 \\times 256\\)) and validation split (8,667 samples for \\(256 \\times 384\\) and 11,229 for \\(256 \\times 256\\)), we report all evaluation metrics on the validation split." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.505, + 0.907, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.702, + 0.907, + 0.731 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison. A two-state approach such as CosHand [70] struggles with post-interaction object dynamics." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.746, + 0.905, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.861, + 0.907, + 0.904 + ], + "angle": 0, + "content": "Figure 6. Robustness to noise. SAM2 outputs noisy/coarse masks for frames with considerable motion blur (orange/red). Despite this, InterDyn can generate plausible hand and object dynamics." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12472" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.907, + 0.844 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.856, + 0.907, + 0.898 + ], + "angle": 0, + "content": "Figure 7. Diverse generation of interactive dynamics. We show multiple challenging examples, such as (from top to bottom): interacting with articulated objects, pouring liquid, letting an object fall, squeezing a highly deformable or \"collapsible\" object, interacting with reflective objects, tilting a ridged object, or stacking objects. Q Zoom in for details." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12473" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.9, + 0.196 + ], + "angle": 0, + "content": "
MethodSSIM ↑PSNR ↑LPIPS ↓FID ↓KID ↓FVD ↓KVD ↓Motion Fidelity ↑ [93]
Seer [23]0.41810.710.58833.350.030287.4681.31—
DynamiCrafter [89]†———17.480.014204.1131.81—
CosHand-Independent [70]0.61516.870.3132.950.00291.1819.240.432
CosHand-Autoregressive [70]0.53114.920.40812.660.01290.3013.680.570
Ours 256×2560.66418.600.2604.950.00419.271.990.633
Ours 256×3840.68019.040.2525.070.00422.222.090.641
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.212, + 0.907, + 0.257 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on Something-Something-v2. We compare against two language-instructed video generation methods, Seer [23] and DynamiCrafter [89] and two video extensions of our baseline CosHand [70]. We report results for InterDyn at two resolutions: \\(256 \\times 256\\) (matching CosHand) and \\(256 \\times 384\\) (matching SVD's prior). Methods denoted with \\(\\dagger\\) do not use SSV2 as their training dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.483, + 0.377 + ], + "angle": 0, + "content": "Baselines. To generate videos using CosHand [70] rather than state transitions, we run CosHand in a frame-by-frame approach (CosHand-Independent) and an auto-regressive approach (CosHand-Autoregressive). In the frame-by-frame variant, each future frame is independently predicted from the initial frame and its corresponding mask:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.384, + 0.443, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {t + 1} = \\operatorname {C o s H a n d} \\left(x _ {0}, h _ {0}, h _ {t + 1}\\right), \\quad \\forall t \\in [ 0, 1 3 ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.456 + ], + "angle": 0, + "content": "where \\( h_0 \\) and \\( h_{t + 1} \\) denote masks, and \\( x_0 \\) the initial frame. In the auto-regressive variant, we use CosHand to generate video frames sequentially:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.465, + 0.443, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {t + 1} = \\operatorname {C o s H a n d} \\left(\\hat {x} _ {t}, h _ {t}, h _ {t + 1}\\right), \\quad \\forall t \\in [ 0, 1 3 ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.481, + 0.505 + ], + "angle": 0, + "content": "with \\(\\hat{x}_t\\) being the previously generated frame and \\(\\hat{x}_0 = x_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.506, + 0.483, + 0.642 + ], + "angle": 0, + "content": "We also compare against two language-instructed video generation methods: Seer [23] and DynamiCrafter [89], which we prompt with a first frame and its corresponding SSV2 class label as instruction. Note that DynamiCrafter was trained with a random video frame as conditioning, i.e. not always the first frame. At inference, its generated videos are thus not strictly a continuation of the conditioning, which precludes frame-aligned metrics. We compare InterDyn against these baselines in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.645, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Quantitative analysis. The frame-by-frame variant of CosHand achieves high image quality but struggles with temporal coherence and motion fidelity. Qualitatively, we notice that object locations are inconsistent across frames. In contrast, the auto-regressive variant improves object motion fidelity but suffers from lower frame-wise image quality due to error propagation. Both variants fail in scenarios requiring accurate post-interaction dynamics, such as when objects continue moving after being released from direct hand contact, as shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Our method, InterDyn \\((256 \\times 384)\\), achieves the best overall performance, surpassing our baseline CosHand in spatio-temporal dynamics, motion fidelity, and all but two image quality metrics. We hypothesize that this might be due to two reasons (1) we use a frozen U-Net, while CosHand fine-tunes its model on SSV2, so CosHand might generate frames closer to the SSV2 distribution, and 2)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.283, + 0.905, + 0.344 + ], + "angle": 0, + "content": "when SVD was trained, it was initialized with SD weights as spatial layers, and then fine-tuned over multiple stages; this might have degraded its spatial prior, and by extension the quality of produced images compared to CosHand." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.907, + 0.45 + ], + "angle": 0, + "content": "Fine-tuned on noisy masks and leveraging its temporal-aware control branch, InterDyn can interpret a noisy control sequence; e.g. when SAM2 produces a coarse and noisy hand mask sequence, InterDyn generates detailed hands including individual fingers, see Fig. 6. Though not always consistent, InterDyn is capable of depicting post-interaction dynamics, such as rolling or sliding objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.909, + 0.68 + ], + "angle": 0, + "content": "Qualitative analysis. We present diverse qualitative results generated by InterDyn in Fig. 7. Row 1 shows how InterDyn generates the articulated motion of an object. Row 2 showcases pouring water into a glass; note how the water level increases over time. Row 3 demonstrates an object being dropped, moving out of frame when falling, and rolling back in frame once hitting the floor, featuring realistic motion blur synthesis. Rows 4 and 5 illustrate how InterDyn handles squeezing interactions—the rubber and the spring are compressed and restored accordingly. Row 6 demonstrates an understanding of physical size and distance to the camera, as the phone moves closer to the viewer. These results highlight the complexity that InterDyn can handle, implying its generalization ability and future potential as an implicit, yet generalized physical simulator and renderer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.692, + 0.634, + 0.708 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.9 + ], + "angle": 0, + "content": "We introduced InterDyn, a framework that generates videos of interactive dynamics using large video generation models as implicit physics simulators. By incorporating an interactive control mechanism, InterDyn produces plausible, temporally consistent videos of object interactions—including complex human-object interactions—while generalizing to unseen objects. Our evaluations demonstrate that InterDyn effectively captures continuous motion and subsequent dynamics, outperforming baselines that focus on single future states. This work highlights the potential of using video generative models as physics simulators without explicit reconstruction, opening new avenues for future research." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12474" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.093, + 0.482, + 0.167 + ], + "angle": 0, + "content": "Acknowledgements: We thank Benjamin Pellkofer for IT support, Peter Kulits for discussions, Tomasz Niewiadomski for coding support, and Allan D. Jepson, Artur Grigorev, Soubhik Sanyal, Omid Taheri & Angjoo Kanazawa for feedback. DT is supported by the ERC Starting Grant (project STRIPES, 101165317)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.169, + 0.483, + 0.183 + ], + "angle": 0, + "content": "Disclosure: https://files.is.tue.mpg.de/black/CoI_CVPR_2025.txt" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.184, + 0.392, + 0.197 + ], + "angle": 0, + "content": "DT has received a research gift fund from Google." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.209, + 0.188, + 0.224 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.234, + 0.483, + 0.317 + ], + "angle": 0, + "content": "[1] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A space-time diffusion model for video generation. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-11, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.318, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[2] Fabien Baradel, Natalia Neverova, Julien Mille, Greg Mori, and Christian Wolf. CoPhy: Counterfactual learning of physical dynamics. In International Conference on Learning Representations (ICLR), 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.373, + 0.483, + 0.428 + ], + "angle": 0, + "content": "[3] Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, et al. Interaction networks for learning about objects, relations and physics. In Conference on Neural Information Processing Systems (NeurIPS), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.429, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[4] Mikołaj Binkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In International Conference on Learning Representations (ICLR), 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.484, + 0.484, + 0.553 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable Video Diffusion: Scaling latent video diffusion models to large datasets. arXiv:2311.15127, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.554, + 0.483, + 0.622 + ], + "angle": 0, + "content": "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 22563-22575, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.623, + 0.483, + 0.678 + ], + "angle": 0, + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning to follow image editing instructions. In Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, 2023. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.679, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[8] Shengqu Cai, Duygu Ceylan, Matheus Gadelha, ChunHao Paul Huang, Tuanfeng Yang Wang, and Gordon Wetzstein. Generative rendering: Controllable 4D-guided video generation with 2D diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 7611-7620, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.761, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[9] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2D pose estimation using part affinity fields. In Computer Vision and Pattern Recognition (CVPR), pages 7291–7299, 2017. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[10] Junuk Cha, Jihyeon Kim, Jae Shin Yoon, and Seungryul Baek. Text2HOI: Text-guided 3D motion generation for hand-object interaction. In Computer Vision and Pattern Recognition (CVPR), pages 1577-1585, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.872, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S. Narang, Karl" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.234, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Van Wyk, Umar Iqbal, Stan Birchfield, et al. DexYCB: A benchmark for capturing hand grasping of objects. In Computer Vision and Pattern Recognition (CVPR), pages 9044-9053, 2021. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.149, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[12] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv:2310.19512, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.22, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[13] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Ji-ashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable text-to-video generation with diffusion models. arXiv:2305.13840, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.277, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[14] Enric Corona, Albert Pumarola, Guillem Alenya, Francesc Moreno-Noguer, and Grégory Rogez. GanHand: Predicting human grasp affordances in multi-object scenes. In Computer Vision and Pattern Recognition (CVPR), pages 5031-5041, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.346, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[15] Christian Diller and Angela Dai. CG-HOI: Contact-guided 3D human-object interaction generation. In Computer Vision and Pattern Recognition (CVPR), pages 19888-19901, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.403, + 0.905, + 0.485 + ], + "angle": 0, + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.488, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[17] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In Computer Vision and Pattern Recognition (CVPR), pages 12943-12954, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.558, + 0.905, + 0.64 + ], + "angle": 0, + "content": "[18] Zicong Fan, Maria Parelli, Maria Eleni Kadoglou, Xu Chen, Muhammed Kocabas, Michael J. Black, and Otmar Hilliges. HOLD: Category-agnostic 3D reconstruction of interacting hands and objects from video. In Computer Vision and Pattern Recognition (CVPR), pages 494–504, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.642, + 0.905, + 0.712 + ], + "angle": 0, + "content": "[19] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. In International Conference on Robotics and Automation (ICRA), pages 12462-12469, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.713, + 0.905, + 0.795 + ], + "angle": 0, + "content": "[20] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu Video: Factorizing text-to-video generation by explicit image conditioning. In European Conference on Computer Vision (ECCV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.797, + 0.905, + 0.894 + ], + "angle": 0, + "content": "[21] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The something something video database for learning and evaluating visual common sense. In International Conference on Computer Vision (ICCV), pages 5842-5850, 2017. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12475" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.16 + ], + "angle": 0, + "content": "[22] Oliver Groth, Fabian B. Fuchs, Ingmar Posner, and Andrea Vedaldi. ShapeStacks: Learning vision-based physical intuition for generalised object stacking. In European Conference on Computer Vision (ECCV), pages 702-717, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.161, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[23] Xianfan Gu, Chuan Wen, Weirui Ye, Jiaming Song, and Yang Gao. Seer: Language instructed video prediction with latent diffusion models. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 8, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.219, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[24] Vincent Le Guen and Nicolas Thome. Disentangling physical dynamics from unknown factors for unsupervised video prediction. In Computer Vision and Pattern Recognition (CVPR), pages 11471-11481, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.276, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[25] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Pengfei Wan, Di Zhang, Yufan Liu, Weiming Hu, Zhengjun Zha, et al. I2V-Adapter: A general image-to-video adapter for diffusion models. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.484, + 0.427 + ], + "angle": 0, + "content": "[26] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. *AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning*. arXiv:2307.04725, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.428, + 0.484, + 0.497 + ], + "angle": 0, + "content": "[27] Yana Hasson, Gül Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Computer Vision and Pattern Recognition (CVPR), pages 11807-11816, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.498, + 0.484, + 0.553 + ], + "angle": 0, + "content": "[28] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from RGB videos. In International Conference on 3D Vision (3DV), pages 659-668, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.554, + 0.484, + 0.609 + ], + "angle": 0, + "content": "[29] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. CameraCtrl: Enabling camera control for text-to-video generation. arXiv:2404.02101, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.61, + 0.484, + 0.679 + ], + "angle": 0, + "content": "[30] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Conference on Neural Information Processing Systems (NeurIPS), pages 6629-6640, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.68, + 0.484, + 0.707 + ], + "angle": 0, + "content": "[31] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.708, + 0.484, + 0.776 + ], + "angle": 0, + "content": "[32] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, et al. Image Video: High definition video generation with diffusion models. arXiv:2210.02303, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.778, + 0.484, + 0.833 + ], + "angle": 0, + "content": "[33] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In Conference on Neural Information Processing Systems (NeurIPS), pages 8633-8646, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.834, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[34] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.092, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[35] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate Anyone: Consistent and controllable image-to-video synthesis for character animation. In Computer Vision and Pattern Recognition (CVPR), pages 8153-8163, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.162, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[36] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. DepthCrafter: Generating consistent long depth sequences for open-world videos. arXiv:2409.02095, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.22, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[37] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make It Move: Controllable image-to-video generation with text descriptions. In Computer Vision and Pattern Recognition (CVPR), pages 18219-18228, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.277, + 0.906, + 0.33 + ], + "angle": 0, + "content": "[38] Zhihao Hu and Dong Xu. VideoControlNet: A motion-guided video-to-video translation framework by using diffusion model with ControlNet. arXiv:2307.14073, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.332, + 0.906, + 0.402 + ], + "angle": 0, + "content": "[39] Steeven Janny, Fabien Baradel, Natalia Neverova, Madiha Nadri, Greg Mori, and Christian Wolf. Filtered-CoPhy: Unsupervised learning of counterfactual physics in pixel space. In International Conference on Learning Representations (ICLR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.403, + 0.906, + 0.459 + ], + "angle": 0, + "content": "[40] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In International Conference on Computer Vision (ICCV), pages 11107-11116, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.46, + 0.906, + 0.53 + ], + "angle": 0, + "content": "[41] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker3: Simpler and better point tracking by pseudolabelling real videos. In European Conference on Computer Vision (ECCV), pages 18-35, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.531, + 0.906, + 0.599 + ], + "angle": 0, + "content": "[42] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. DreamPose: Fashion image-to-video synthesis via stable diffusion. In International Conference on Computer Vision (ICCV), pages 22623-22633, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.601, + 0.906, + 0.657 + ], + "angle": 0, + "content": "[43] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Conference on Neural Information Processing Systems (NeurIPS), 35:26565-26577, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.658, + 0.906, + 0.727 + ], + "angle": 0, + "content": "[44] Hyeonwoo Kim, Sookwan Han, Patrick Kwon, et al. Beyond the Contact: Discovering comprehensive affordance for 3D objects from pre-trained 2D diffusion models. In European Conference on Computer Vision (ECCV), pages 400-419, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.728, + 0.906, + 0.77 + ], + "angle": 0, + "content": "[45] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.771, + 0.906, + 0.827 + ], + "angle": 0, + "content": "[46] Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems. In International Conference on Machine Learning (ICML), pages 2693-2702, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.828, + 0.906, + 0.898 + ], + "angle": 0, + "content": "[47] Sumith Kulal, Tim Brooks, Alex Aiken, Jiajun Wu, Jimei Yang, Jingwan Lu, Alexei A. Efros, and Krishna Kumar Singh. Putting people in their place: Affordance-aware human insertion into scenes. In Computer Vision and Pattern Recognition (CVPR), pages 17089-17099, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.092, + 0.907, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12476" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[48] Adam Lerer, Sam Gross, and Rob Fergus. Learning physical intuition of block towers by example. In International Conference on Machine Learning (ICML), pages 430-438, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[49] Yunzhu Li, Antonio Torralba, Anima Anandkumar, Dieter Fox, and Animesh Garg. Causal discovery in physical systems from videos. Conference on Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.206, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[50] Ce Liu, Antonio Torralba, William T. Freeman, Frédo Durand, and Edward H. Adelson. Motion magnification. Transactions on Graphics (TOG), 24(3):519-526, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.249, + 0.483, + 0.317 + ], + "angle": 0, + "content": "[51] Ruibo Liu, Jason Wei, Shixiang Shane Gu, Te-Yen Wu, Soroush Vosoughi, Claire Cui, Denny Zhou, and Andrew M Dai. Mind's Eye: Grounded language model reasoning through simulation. In International Conference on Learning Representations (ICLR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.319, + 0.483, + 0.386 + ], + "angle": 0, + "content": "[52] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. arXiv:2303.05499, 2023. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.483, + 0.444 + ], + "angle": 0, + "content": "[53] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. PhysGen: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision (ECCV), pages 360-378, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[54] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv:2402.17177, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.515, + 0.483, + 0.583 + ], + "angle": 0, + "content": "[55] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T. Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. Transactions on Graphics (TOG), 39(6): 1-14, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.585, + 0.483, + 0.653 + ], + "angle": 0, + "content": "[56] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T. Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In Computer Vision and Pattern Recognition (CVPR), pages 4507-4515, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.655, + 0.483, + 0.711 + ], + "angle": 0, + "content": "[57] Jiaxin Lu, Hao Kang, Haoxiang Li, Bo Liu, Yiding Yang, Qixing Huang, and Gang Hua. UGG: Unified generative grasping. In European Conference on Computer Vision (ECCV), pages 414-433, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.712, + 0.483, + 0.78 + ], + "angle": 0, + "content": "[58] Wan-Duo Kurt Ma, John P Lewis, and W. Bastiaan Kleijn. TrailBlazer: Trajectory control for diffusion-based video generation. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.782, + 0.483, + 0.85 + ], + "angle": 0, + "content": "[59] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-Else: Compositional action recognition with spatial-temporal interaction networks. In Computer Vision and Pattern Recognition (CVPR), pages 1049–1059, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.852, + 0.483, + 0.894 + ], + "angle": 0, + "content": "[60] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "image diffusion models. In AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.122, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[61] Franziska Mueller, Micah Davis, Florian Bernard, Oleksandr Sotnychenko, Micekal Verschooor, Miguel A. Otaduy, Dan Casas, and Christian Theobalt. Real-time pose and shape reconstruction of two interacting hands with a single depth camera. Transactions on Graphics (TOG), 38(4): 1-13, 2019. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.206, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[62] Boxiao Pan, Zhan Xu, Chun-Hao Paul Huang, Krishna Kumar Singh, Yang Zhou, Leonidas J. Guibas, and Jimei Yang. ActAnywhere: Subject-aware video background generation. In Conference on Neural Information Processing Systems (NeurIPS), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.277, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[63] Georgios Paschalidis, Romana Wilschut, Dimitrije Antic, Omid Taheri, and Dimitrios Tzionas. 3D whole-body grasp synthesis with directional controllability. In International Conference on 3D Vision (3DV), 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.334, + 0.906, + 0.416 + ], + "angle": 0, + "content": "[64] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.418, + 0.906, + 0.514 + ], + "angle": 0, + "content": "[65] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Girshick, Piotr Dólar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.516, + 0.906, + 0.583 + ], + "angle": 0, + "content": "[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.585, + 0.906, + 0.64 + ], + "angle": 0, + "content": "[67] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. Transactions on Graphics (TOG), 36(6), 2022. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.642, + 0.906, + 0.712 + ], + "angle": 0, + "content": "[68] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-A-Video: Text-to-video generation without text-video data. In International Conference on Learning Representations (ICLR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.713, + 0.906, + 0.781 + ], + "angle": 0, + "content": "[69] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. StyleGAN-V: A continuous video generator with the price, image quality and perks of StyleGAN2. In Computer Vision and Pattern Recognition (CVPR), pages 3626-3636, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.784, + 0.906, + 0.839 + ], + "angle": 0, + "content": "[70] Sruthi Sudhakar, Ruoshi Liu, Basile Van Hoorick, Carl Vondrick, and Richard Zemel. Controlling the world by sleight of hand. In European Conference on Computer Vision (ECCV), pages 414-430, 2024. 2, 3, 6, 8, 14, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.84, + 0.906, + 0.896 + ], + "angle": 0, + "content": "[71] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), pages 581-600, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12477" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[72] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J. Black. GRIP: Generating interaction poses using latent consistency and spatial cues. In International Conference on 3D Vision (3DV), pages 933-943, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[73] Tze Ho Elden Tse, Kwang In Kim, Ales Leonardis, and Hyung Jin Chang. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In Computer Vision and Pattern Recognition (CVPR), pages 1664-1674, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[74] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation. In International Conference on Learning Representations Workshops (ICLRw), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.301, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[75] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. In European Conference on Computer Vision (ECCV), pages 439-457, 2025. 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.385, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[76] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv:2308.06571, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.427, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[77] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. In International Conference on Machine Learning (ICML), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[78] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional video synthesis with motion controllability. In Conference on Neural Information Processing Systems (NeurIPS), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[79] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image Quality Assessment: From error visibility to structural similarity. Transactions on Image Processing (TIP), 13(4):600-612, 2004. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[80] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A unified and flexible motion controller for video generation. In Transactions on Graphics (TOG), pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[81] Nicholas Watters, Daniel Zoran, Theophane Weber, Peter Battaglia, Razvan Pascanu, and Andrea Tacchetti. Visual Interaction Networks: Learning a physics simulator from video. In Conference on Neural Information Processing Systems (NeurIPS), pages 4539-4547, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[82] Jiajun Wu, Ilker Yildirim, Joseph J. Lim, Bill Freeman, and Joshua B. Tenenbaum. Galileo: Perceiving physical object properties by integrating a physics engine with deep learning. In Conference on Neural Information Processing Systems (NeurIPS), pages 127-135, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.829, + 0.482, + 0.883 + ], + "angle": 0, + "content": "[83] Jiajun Wu, Joseph J. Lim, Hongyi Zhang, Joshua B. Tenenbaum, and William T. Freeman. Physics 101: Learning physical object properties from unlabeled videos. In *British Machine Vision Conference (BMVC)*, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.884, + 0.482, + 0.91 + ], + "angle": 0, + "content": "[84] Jiajun Wu, Erika Lu, Pushmeet Kohli, Bill Freeman, and Josh Tenenbaum. Learning to see physics via visual de" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.91 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "animation. In Conference on Neural Information Processing Systems (NeurIPS), pages 153-164, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[85] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-A-Video: One-shot tuning of image diffusion models for text-to-video generation. In International Conference on Computer Vision (ICCV), pages 7623–7633, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.207, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[86] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. THOR: Text to human-object interaction diffusion via relation intervention. arXiv:2403.11208, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.262, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[87] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. DragAnything: Motion control for anything using entity representation. In European Conference on Computer Vision (ECCV), pages 331-348, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.334, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[88] Xianghui Xie, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Template free reconstruction of human-object interaction with procedural interaction generation. In Computer Vision and Pattern Recognition (CVPR), pages 10003-10015, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.403, + 0.905, + 0.485 + ], + "angle": 0, + "content": "[89] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. DynamiCrafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision (ECCV), pages 399-417, 2025. 2, 3, 8, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.488, + 0.905, + 0.555 + ], + "angle": 0, + "content": "[90] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3D human-object interactions with physics-informed diffusion. In International Conference on Computer Vision (ICCV), pages 14928-14940, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.558, + 0.905, + 0.64 + ], + "angle": 0, + "content": "[91] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. MagicAnimate: Temporally consistent human image animation using diffusion model. In Computer Vision and Pattern Recognition (CVPR), pages 1481–1490, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.642, + 0.905, + 0.711 + ], + "angle": 0, + "content": "[92] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. CogVideoX: Text-to-video diffusion models with an expert transformer. arXiv:2408.06072, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.713, + 0.905, + 0.768 + ], + "angle": 0, + "content": "[93] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In Computer Vision and Pattern Recognition (CVPR), pages 8466-8476, 2024. 5, 8, 14, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.77, + 0.905, + 0.825 + ], + "angle": 0, + "content": "[94] Yufei Ye, Poorvi Hebbar, Abhinav Gupta, and Shubham Tulsiani. Diffusion-guided reconstruction of everyday hand-object interaction clips. In International Conference on Computer Vision (ICCV), pages 19717-19728, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.826, + 0.905, + 0.895 + ], + "angle": 0, + "content": "[95] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance Diffusion: Synthesizing hand-object interactions. In Computer Vision and Pattern Recognition (CVPR), pages 22479-22489, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12478" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[96] Yufei Ye, Abhinav Gupta, Kris Kitani, and Shubham Tulsiani. G-HOP: Generative hand-object prior for interaction reconstruction and grasp synthesis. In Computer Vision and Pattern Recognition (CVPR), pages 1911–1920, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.148, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[97] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Ji-jun Wu, Antonio Torralba, and Joshua B. Tenenbaum. CLEVRR: Collision events for video representation and reasoning. In International Conference on Learning Representations (ICLR), 2020. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.218, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[98] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. DragNUWA: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv:2308.08089, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[99] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Computer Vision and Pattern Recognition (CVPR), pages 18456-18466, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.328, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[100] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.384, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[101] Mengqi Zhang, Yang Fu, Zheng Ding, Sifei Liu, Zhuowen Tu, and Xiaolong Wang. HOIDiffusion: Generating realistic 3D hand-object interaction data. In Computer Vision" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.092, + 0.486, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "and Pattern Recognition (CVPR), pages 8521-8531, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[102] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[103] Juntian Zheng, Qingyuan Zheng, Lixing Fang, Yun Liu, and Li Yi. CAMS: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis. In Computer Vision and Pattern Recognition (CVPR), pages 585-594, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[104] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient video generation with latent diffusion models. arXiv:2211.11018, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[105] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. GEARS: Local geometry-aware hand-object interaction synthesis. In Computer Vision and Pattern Recognition (CVPR), pages 20634-20643, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[106] Tianqiang Zhu, Rina Wu, Xiangbo Lin, and Yi Sun. Toward human-like grasp: Dexterous grasping via semantic representation of object-hand. In International Conference on Computer Vision (ICCV), pages 15741-15751, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12479" + } + ] +] \ No newline at end of file diff --git a/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/1657842e-f939-4237-87d3-49f995d53e68_origin.pdf b/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/1657842e-f939-4237-87d3-49f995d53e68_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ed85a168f9a4b37e05741a1d0c20cba4a28efadc --- /dev/null +++ b/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/1657842e-f939-4237-87d3-49f995d53e68_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:caf2b221889aefaddec6247d0b71d1e9547cf7150fa10e2b5edbb22671e5cf33 +size 2675801 diff --git a/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/full.md b/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8eb086841b68a4e507757cd229df86c5bc6a6f0b --- /dev/null +++ b/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/full.md @@ -0,0 +1,313 @@ +# InterDyn: Controllable Interactive Dynamics with Video Diffusion Models + +Rick Akkerman $^{1,2*}$ Haiwen Feng $^{1*†}$ Michael J. Black $^{1}$ Dimitrios Tzionas $^{2}$ Victoria Fernández Abrevaya $^{1}$ + +$^{1}$ Max Planck Institute for Intelligent Systems, Tübingen, Germany $^{2}$ University of Amsterdam, the Netherlands + +{rick.akkerman, haiwen.feng, black, victoria.abrevaya}@tuebingen.mpg.de d.tzionas@uva.nl + +![](images/30b0ac9f68c40219c46eb31fd49f92bbdee1d119f6ed39037577255457f9180b.jpg) +Input image + +![](images/5f263dc19b34cc956bce9bec510d1cc9ba36ef6e51f61f7573a896f776fb1e9a.jpg) +Video generation by InterDyn using only the hand mask sequence as control signal + +![](images/113443254ca639e65d103d521d5dce89e0f8d67c2dd453fa94c33f7d5b93205a.jpg) + +![](images/02312659b0585085ead49a474af9732b321279cb8a80bb0c106e2a1d61995635.jpg) + +![](images/1aabfaedc7193a34b5fdd250b073f50a42f2a211a73bba63d8952a9bb96e6e4d.jpg) +Force propagation + +Counterfactual dynamics +Figure 1. We present InterDyn, a framework for synthesizing realistic interactive dynamics without 3D reconstruction and physics simulation. Our core principle is to rely on the implicit physics knowledge embedded in large-scale video generative models. Given an image and a "driving motion", our model generates the consequential scene dynamics. We investigate the generated interactive dynamics in a simple object collision scenario (bottom) and complex in-the-wild human-object interaction (top). +![](images/ee346ad8c41e11075b1455634b058139e7def2b188313e9387be0341fac73f0e.jpg) +Denotes a driving object motion | Tracks indicate an object with generated uncontrolled dynamics + +# Abstract + +Predicting the dynamics of interacting objects is essential for both humans and intelligent systems. However, existing approaches are limited to simplified, toy settings and lack generalizability to complex, real-world environments. Recent advances in generative models have enabled the prediction of state transitions based on interventions, but focus on generating a single future state which neglects the continuous dynamics resulting from the interaction. To address this gap, we propose InterDyn, a novel framework that generates videos of interactive dynamics given an initial frame and a control signal encoding the motion of a driving object or actor. Our key insight is that large video + +generation models can act as both neural renderers and implicit physics "simulators", having learned interactive dynamics from large-scale video data. To effectively harness this capability, we introduce an interactive control mechanism that conditions the video generation process on the motion of the driving entity. Qualitative results demonstrate that InterDyn generates plausible, temporally consistent videos of complex object interactions while generalizing to unseen objects. Quantitative evaluations show that InterDyn outperforms baselines that focus on static state transitions. This work highlights the potential of leveraging video generative models as implicit physics engines. Code and trained models will be released at: https://interdyn.is.tue.mpg.de/. + +# 1. Introduction + +Humans have the remarkable ability to predict the future dynamics of observed systems intuitively. With just a single image, we can anticipate and imagine how objects will move over time – not only their motion but also their interactions with the environment and other elements in the scene. Inferring this requires an advanced form of scene-level reasoning beyond merely recognizing the semantics and geometry of static elements; it involves a deep physical and causal understanding of how each object will interact given the environment, object properties, and forces. + +There has been a growing interest in developing machine learning systems that emulate similar levels of dynamic understanding given visual observations, such as images or videos. Early work [84] addressed this by first reconstructing a 3D representation from the image, then predicting future states with a physics simulator and finally generating the output video with a rendering engine. This relies heavily on explicit reconstruction and simulation, which is computationally intensive, prone to errors, and may not generalize well. More recent methods [2, 24, 39, 46, 49] leverage keypoint or latent representations within graph relational frameworks; however, they have only been trained and validated in over-simplified, synthetic environments, showing limited generalizability to complex real-world scenarios. + +Instead, the advent of powerful generative models [1, 5, 16, 54, 66] opens new avenues for synthesizing interactions under complex scenarios. For example, Sudhakar et al. [70] recently proposed CosHand, a controllable image-to-image model based on Stable Diffusion [66] that infers state transitions of an object. The task here is defined as follows: given an image of a hand interacting with an object, alongside a hand mask of the current frame and a mask of the hand at a future frame, generate a modified input image that satisfies the mask, with realistic interactions. The challenge, as in early intuitive physics works, lies in accurately modeling how the objects will change after forces are applied. However, we argue that static state transitions are insufficient for this task, as they fail to capture the continuous dynamic processes inherent to the problem, e.g. see Fig. 2. Investigating interactive dynamics within a two-state setting is highly limiting, since dynamics can extend beyond the period of direct contact – for example, predicting the motion occurring while a person pours water requires a physical understanding that goes beyond the state of the hand at a future frame. The driving force, in this case the hand, may interact with the system only briefly, but the system's subsequent dynamics continue according to physical laws and may even influence other parts via force propagation. + +In this paper, we explore controllable synthesis of interactive dynamics—generating a video from an input image and a dynamic control signal (e.g. a moving hand mask) to model realistic object dynamics. In particular, we propose + +![](images/ebe545c08fdf1434341070911268a6e0cf3fd3097f7b639364bfe10dc5ebd701.jpg) +Figure 2. State transition vs. dynamics. Methods that generate static state transitions (i.e. predict a future image) such as CosHand [70] struggle to capture the inherent dynamic processes involved in human-object interactions. Here, we show a video sequence where the motion continues beyond the interaction. + +InterDyn, a novel framework for synthesizing controllable dynamic interactions that leverages the physical and dynamics "knowledge" of a large video generation model [5]. Unlike prior approaches that rely on explicit physical simulation [84] or are constrained to static state transitions [70], we leverage video generation models to generate dynamic processes implicitly, see Fig. 1. Specifically, we extend Stable Video Diffusion (SVD) [5] with a dynamic control branch and fine-tune it on diverse scenes, enabling synthesis of complex interactions aligned with the control signal. + +We start our investigation by fine-tuning InterDyn on a simple synthetic scenario of cubes, cylinders, and spheres: the CLEVRER dataset [97]. To control the motion we add a mask driving signal that manipulates the movement of some (but not all) of the objects in the scene. We then evaluate how the synthesized trajectories of uncontrolled objects change under various interactions, including multiple objects colliding with each other. This multi-object collision setting allows us to "probe" the physical understanding and causal effects of the video diffusion model, and our qualitative experiments show InterDyn's ability for counterfactual future prediction and physical force propagation. + +Next, we evaluate how the system performs in a difficult real-world scenario, such as Human-Object Interaction (HOI). Here, the dexterity of hand motions and the diversity of objects vastly increase the complexity of the problem. We fine-tune the model on a commonly used HOI video dataset [21] and compare with the state-of-the-art baseline CosHand [70], as well as two text-control based interactive dynamics generation methods: Seer [23] and DynamiCrafter [89]. We quantify our investigations using standard image and video metrics, as well as a motion fidelity metric based on point tracking. InterDyn surpasses the previous SOTA over $37.5\%$ on LPIPS and $77\%$ on FVD on the Something-Something-v2 (SSV2) dataset [21]. Our experiments also demonstrate diverse and physically plausible generations of interactive dynamics, probing into SVD's "understanding" of physics and dynamics. + +In summary, we present InterDyn, a framework that employs video generative models to simulate object dynamics without explicit 3D reconstruction or physical simulation. We demonstrate how the inherent "knowledge" within video generation models can be leveraged to predict complex object interactions and movements over time, implicitly modeling physical and causal dynamics. We perform comprehensive experiments on multi-object collision datasets and hand-object manipulation datasets, demonstrating the effectiveness of our approach. + +# 2. Related Work + +Modeling human-object interactions (HOI). Human-object interaction has been widely studied within the context of 3D reconstruction [17, 18, 27, 28, 73, 88, 94], where the goal is to recover realistic geometry of hands and objects. The field of 3D HOI synthesis has also received increasing attention, including the generation of static [40, 44, 71, 106] or dynamic [63, 72, 103, 105] hand poses conditioned on 3D objects, whole-body interactions [90], or more recently, hand-object meshes given textual descriptions [10, 15, 57, 86, 96]. Few works address HOI synthesis in the 2D domain. GANHand [14] predicts 3D hand shape and pose given an RGB image of an object, while AffordanceDiffusion [95] estimates a 2D hand using a diffusion model. Kulal et al. [47] take as input an image of a human and a scene separately and generate a composite image that positions the human with correct affordances. Also relevant is HOIDiffusion [101], in which a texture-less rendering of a 3D hand and object is converted to a realistic image using a text description. Most closely related to us is CosHand [70], which takes as input an RGB image of a hand-object interaction, a hand mask at the current state, and the hand mask of the future state, and generates an RGB image of the future state. Unlike us, they cannot generate post-interaction object dynamics. Importantly, none of these works study dynamics, generating instead discrete state transitions that fail to capture the nuanced, temporally coherent behaviors observed in interactions. + +# Synthesizing causal physical relations from visual input. + +A growing body of work aims to model and predict physical causal effects from visual inputs such as images or videos. For example, research in intuitive physics seeks to replicate the human-like, non-mathematical understanding of physical events, e.g. by predicting future frames given an input video. Early works like [22, 48] train neural networks to assess the stability of block towers, while [24] leverage prior physical knowledge formalized through partial differential equations (PDEs). Other approaches investigate counterfactual reasoning by leveraging graph neural networks [2, 39]. Wu et al. [82-84] explore the use of an inverse rendering approach, extracting geometry and physical properties from + +the video which are then coupled with a physics simulator and a rendering engine to generate the future frames. Other works [81] incorporate Interaction Networks [3] to approximate physical systems from video data. These approaches are often limited to simplified, synthetic datasets and struggle to generalize to real-world scenarios. + +Recent methods have started to combine language models with physical engines. Liu et al. [51] ground a large language model using a computational physics engine while Gao et al. [19] show that fine-tuning a vision-language model (VLM) on annotated datasets of physical concepts improves its understanding of physical interactions. Closely related to our work is PhysGen [53], which trains an imaged-to-video model that conditions the video generation on physics parameters (e.g., force or torque). However, the model relies on a dynamics simulator to generate motion, and its application is limited to rigid objects. A related but tangential line of work focuses on identifying and generating the effects of objects on their surroundings. For example, Omnimatte [56] introduces the problem of identifying all parts of a scene influenced by an object, given a video and a mask of the object. Similarly, Lu et al. [55] propose to re-time the motion of different subjects in a scene while maintaining realistic interactions with the environment. ActAnywhere [62] generates videos with plausible human-scene interactions, taking a masked video of a person and a background image as input. These works address the problem of synthesizing realistic interactions within a scene, however, lack fine-grained control. + +Controllable video generation. Video generation has advanced significantly in recent years, with diffusion models leading to substantial improvements in unconditional [33, 99], text-based [1, 5, 6, 12, 20, 23, 26, 32, 37, 68, 76, 85, 89, 92, 104] and image-based [1, 5, 20, 25, 78, 89] generation. These advances have raised the question of how to incorporate more nuanced control into video generation. Some text-to-video approaches are trained by "inflating" text-to-image (T2V) models [8, 13, 25, 26, 38, 85], and can thus be integrated with conditional T2V models such as ControlNet [100] or T2V-Adapter [60]. Control can also be achieved by conditioning on trajectories [58, 87, 98] or bounding-boxes [77], fine-tuning on appropriate datasets. VideoComposer [78] incorporates multiple condition types, including text, depth, style, and temporal conditions via motion vectors. Camera motion control has also been explored, withAnimateDiff [79] employing LoRA [34] modules to control camera movement, while MotionCtrl [80] and CameraCtrl [29] directly embed the camera information for more precise control. Additionally, several works target human animation from a pose control signal, such as DreamPose [42], MagicPose [91], andAnimateAnyone [35], but do not account for interactions. + +# 3. Controllable Interactive Dynamics + +Video diffusion models such as [5, 54] have demonstrated impressive performance in generating videos from text or images, and have even shown potential in tasks that require 3D understanding when properly fine-tuned [36, 75]. Trained on millions of videos, we hypothesize that these models also possess implicit knowledge of complex interactive dynamics, such as those that appear when humans interact with objects. Out of the box, however, they lack a precise control mechanism, often relying solely on textual inputs or requiring careful selection of the starting frame. + +Task. Given an input image, $\pmb{x} \in \mathbb{R}^{1 \times H \times W \times 3}$ , and a driving motion in the form of a pixel-wise corresponding control signal $\pmb{c} \in \mathbb{R}^{N \times H \times W \times 3}$ , we task InterDyn with generating a video sequence, $\pmb{y} \in \mathbb{R}^{N \times H \times W \times 3}$ , depicting plausible object dynamics. Through this task, we aim to learn the conditional distribution between a driving motion, such as that of a human hand, and the consequent motion of manipulated objects. In other words, the model needs to synthesize plausible object movement and appearance without any indication other than the driving motion, while maintaining physical and visual consistency with the input image. + +Stable Video Diffusion. We extend Stable Video Diffusion [5] (SVD) to enable controllable interactive dynamics and explore the versatility of this model across a range of scenarios. SVD is a publicly available U-Net-based latent diffusion model [66] that extends Stable Diffusion 2.1 to video generation by interleaving the network with temporal layers. Given a static input image of a scene, SVD denoises a sequence of $N$ frames $\pmb{y} \in \mathbb{R}^{N \times H \times W \times 3}$ to generate a video that follows the initial frame. The input image is fed into the denoising U-Net by concatenating its latent to each of the frames in the noised input, and by supplying its CLIP [64] embedding to the U-Net's cross-attention layers. In addition, SVD is conditioned on the video's FPS and motion ID, where the motion ID represents the amount of motion in the video. We found a motion ID of 40 to align well with our frozen SVD prior. + +Control. InterDyn extends SVD with an additional control signal $c \in \mathbb{R}^{N \times H \times W \times 3}$ by integrating a ControlNet-like branch [100]. An overview of our pipeline is presented in Fig. 3. The SVD weights remain frozen to preserve its learned dynamics prior. Following [100], we introduce a trainable copy of the SVD encoder $E$ , connected to SVD's frozen decoder via skip connections, and modulated by zero-initialized convolutions. We use a small CNN, $\mathcal{E}(\cdot)$ , to encode the control signal $c$ into the latent space, which is then added to the noisy input latent that is passed to the ControlNet encoder. Similar to SVD, the control branch interleaves convolutional, spatial, and temporal blocks, enabling InterDyn to process the control signal in a temporal-aware manner. This helps InterDyn to be robust to + +![](images/9046525db769913ab2d00853ff28fe6ab5804a5309900a6555b944e893f40745.jpg) +Figure 3. Overview of InterDyn. Given an input image depicting a scene, such as a hand holding a remote, and a "driving motion," such as a sequence of binary hand masks, InterDyn generates a video depicting plausible hand and object dynamics. Crucially, InterDyn receives no control signal for the object. Through this setup, we probe and assess the implicit knowledge of large video generation models on complex interactive dynamics. We use Stable Video Diffusion (SVD) as our frozen backbone and fine-tune a separate control signal encoder. Videos are iteratively denoised over $T$ timesteps, starting from Gaussian noise $\epsilon \sim \mathcal{N}(0,I)$ . + +noisy control signals, see Fig. 6. We opt for binary masks as conditioning signal due to their accessibility. However, our method can be extended to incorporate diverse types of signals. We find that for hand-object interactions, the type of conditioning signal does not significantly impact performance, see Appendix A and Tab. S1. + +Inference. During inference, we start from an input image, control signal sequence, and randomly sampled Gaussian noise $\epsilon \sim \mathcal{N}(0,I)$ . Through iteratively applying InterDyn over $T$ denoising timesteps, we generate a video $\pmb{y}$ depicting plausible hand and object dynamics, aligned with the control signal. + +# 4. Experiments + +The primary goal of this work is to synthesize scene-level interactive dynamics by leveraging the implicit physical understanding of a pre-trained video generative model. We begin by probing the model's ability to predict physically plausible outcomes within simulated environments, specifically using the CLEVRER dataset [97]. We test force propagation amongst uncontrolled objects and examine counterfactual future prediction by generating videos for one input image with different control signals. Motivated by promising results, we extend our investigation to complex, real-world hand-object interaction scenarios using the Something-Something-v2 (SSV2) dataset [21], conducting comprehensive comparisons with existing baselines that pursue similar objectives. Additionally, we showcase diverse physical examples to demonstrate the capabilities of InterDyn in generating realistic interactive dynamics. + +![](images/e301e10426e190d9050c359053ddaf95d49c1b84a0e16be2b8ff73b60fa5d86c.jpg) +(a) Force Propagation. InterDyn generates force propagation to and amongst uncontrolled objects (cyan cylinder, pink cylinder, and pink cube). + +![](images/dba23c81a0e46e6701a9bde0c25c5a11a0c1513d3dd0d3519aa619850b75768a.jpg) +(b) Counterfactual Dynamics. Two generations of InterDyn on the same input image: (1) w/ brown-cylinder control and (2) w/o brown-cylinder control. +Figure 4. Qualitative investigation on the CLEVRER dataset. Given an input image and the "driving" motion of one or two objects, our model predicts the future interactive dynamics of multiple elements in the scene. The driving motion is given in the form of semantic mask sequences. The generated object motions are highlighted with a red-line trajectory. Note that our model can generate videos with force propagation across multiple uncontrolled objects (top) and can generate multiple futures (bottom). Q Zoom in for details. + +# 4.1. Implementation + +We initialize InterDyn with the 14-frame image-to-video weights of SVD [5]. During training, we use the Adam optimizer [45] with a learning rate of $1 \times 10^{-5}$ . We use the EDM framework [43] with noise distribution $\log \sigma \sim \mathcal{N}(0.7, 1.6^2)$ . We train on two 80GB H100 GPUs, with a per-GPU batch size of 4. Video length and FPS define the temporal resolution of dynamics; to balance short- and long-range events we subsample videos to 7 FPS. To facilitate classifier-free guidance [31], we randomly drop the input image with a probability of $5\%$ . At inference, we apply the Euler scheduler [43] over 50 denoising timesteps. + +# 4.2. Metrics + +We evaluate InterDyn on image quality, spatio-temporal similarity, and motion fidelity. Image quality metrics are computed frame-wise. All metrics are reported excluding the first frame, as it serves as input conditioning. + +Image quality. We report the Structural Similarity Index Measure (SSIM) [79], Peak Signal-to-Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS) [102], Fréchet Inception Distance (FID) [30] and unbiased Kernel Inception Distance (KID) [4]. + +Spatio-temporal similarity. To assess the spatio-temporal perceptual similarity between the ground truth and the generated video distributions, we use the Fréchet Video Distance (FVD) and unbiased Kernel Video Distance (KVD) proposed in [74]. We use the implementation of [69]. + +Motion Fidelity. Through InterDyn, we do not have explicit control over object dynamics, which means that the pixel alignment of an object in the generated and ground truth video is only guaranteed in the starting frame. In this case, comparing generated object motion to the ground truth naively might misrepresent the true quality of object motion over time. Therefore, we adapt the Motion Fidelity (MF) metric proposed by Yatim et al. [93], which measures the similarity between point-tracking trajectories. + +To compute the metric for any video, we obtain a mask of the object in the starting frame, sample 100 points on the object, and track these throughout both the ground truth and generated video using CoTracker3 [41]. Given the resulting two sets of tracklets $\mathcal{T} = \{\tau_1,\dots ,\tau_n\}$ , $\tilde{\mathcal{T}} = \{\tilde{\tau}_1,\dots ,\tilde{\tau}_m\}$ the motion fidelity metric is defined as: + +$$ +\frac {1}{m} \sum_ {\tilde {\tau} \in \tilde {\mathcal {T}}} \max _ {\tau \in \mathcal {T}} \mathbf {c o r r} (\tau , \tilde {\tau}) + \frac {1}{n} \sum_ {\tau \in \mathcal {T}} \max _ {\tilde {\tau} \in \tilde {\mathcal {T}}} \mathbf {c o r r} (\tau , \tilde {\tau}), \tag {1} +$$ + +with the correlation between two tracklets $\mathbf{corr}(\tau, \tilde{\tau})$ [50]: + +$$ +\mathbf {c o r r} (\tau , \tilde {\tau}) \frac {1}{F} \sum_ {k = 1} ^ {F} \frac {v _ {k} ^ {x} a \cdot \tilde {v} _ {k} ^ {x} + v _ {k} ^ {y} \cdot \tilde {v} _ {k} ^ {y}}{\sqrt {(v _ {k} ^ {x}) ^ {2} + (v _ {k} ^ {y}) ^ {2}} \cdot \sqrt {(\tilde {v} _ {k} ^ {x}) ^ {2} + (\tilde {v} _ {k} ^ {y}) ^ {2}}}, \tag {2} +$$ + +where $(v_k^x, v_k^y)$ , $(\tilde{v}_k^x, \tilde{v}_k^y)$ are the $k^{th}$ frame displacement of tracklets $\tau, \tilde{\tau}$ respectively. If there are less than 100 points to query on the object due to it being too small, we do not consider the video for the motion fidelity metric. + +# 4.3. Probing Dynamics with Object Collision Events + +Here, we fine-tune InterDyn on an object collision dataset to probe its ability to generate realistic object interactions. Qualitatively, we review whether InterDyn can produce plausible object motion for uncontrolled objects, given the motion of objects entering the scene. In addition, we examine whether InterDyn can generate counterfactual videos for the same input image, but different control signals. + +Dataset. We use CLEVRER [97], which provides 20,000 videos of colliding objects with annotated segmentation masks and metadata on collision events. We construct a control signal for objects entering the scene and aim to use InterDyn to generate the motion of the objects that are already present, upon collision. Stationary objects do not receive any form of control signal. Colored masks help the model distinguish unique objects. The frames are cropped and scaled to $320 \times 448$ , and we only sample input frames before collisions between objects in the scene, to maximize InterDyn's exposure to interactive dynamics. + +Force propagation. InterDyn can generate force propagation dynamics between a controlled object and an uncontrolled object, as well as amongst uncontrolled objects, as illustrated in Fig. 4a. Here, the red cylinder at the top left is the driving force. It collides with the uncontrolled blue cylinder, which then collides with the uncontrolled purple cylinder, in turn striking the uncontrolled purple cube on the far right. Point-tracking trajectories display how collisions alter each object's position. This suggests that InterDyn possesses an implicit understanding of physical interactions, enabling it to generate plausible dynamics. + +Counterfactual dynamics. By altering the control signal, InterDyn can simulate counterfactual scenarios for the same input image, as shown in Fig. 4b. In "Generation 1", the gray cylinder (controlled) collides with the stationary red sphere (uncontrolled), causing it to move; it is later struck by the brown cylinder (controlled), altering its path once again. In "Generation 2", removing the brown cylinder lets the red sphere continue along its original trajectory, consistent with expectations. Crucially, there is no control signal for the red sphere throughout the sequence; its motion and trajectory are entirely generated by InterDyn. + +Probing InterDyn on CLEVRER highlights its ability to generate interactive dynamics for objects within a simple synthetic environment. We provide additional results in video format on our webpage. + +# 4.4. Generating Human-Object Interactions + +For this experiment, we fine-tune InterDyn on a human-action dataset, focused on hand-object interaction. We encode human movement as a sequence of pixel-aligned binary hand masks and task InterDyn to generate a video with hand movements and corresponding object dynamics. + +Dataset. We use Something-Something-v2 (SSV2), which provides 220,847 videos of humans performing basic actions with everyday objects. It contains actions such as "pushing [something] from left to right", "squeezing [something]" and "lifting [something] with [something] on it". This dataset allows us to train InterDyn at a larger scale and compare with our baseline CosHand [70]. We train one version of InterDyn at the same resolution as CosHand, $256 \times 256$ , and a second version at $256 \times 384$ , which aligns better with the dynamic prior of SVD. + +We generate binary hand masks by prompting Segment Anything 2 (SAM2) [65] with hand bounding boxes, provided by the Something-Else dataset [59]. Similar to our baseline CosHand, we exclude the "pretending" class. Additionally, we remove all "[something] is done behind [something]" classes since the object would be out of view of the camera. We remove videos smaller than the target resolution/length, ensure the hand and objects are continuously visible, and crop videos larger than the target resolution while centering the object. We include videos without obvious motion and state transition, such as "holding [something]". Since bounding box annotations are only provided for the train split (79,043 samples for $256 \times 384$ and 104,260 for $256 \times 256$ ) and validation split (8,667 samples for $256 \times 384$ and 11,229 for $256 \times 256$ ), we report all evaluation metrics on the validation split. + +![](images/496d5842234bf144097f5c51c0980be1d898e59e0df055ff34743b8950b74631.jpg) +Figure 5. Qualitative comparison. A two-state approach such as CosHand [70] struggles with post-interaction object dynamics. + +![](images/b57feeb496674aaef718726f580fd586084668415794913ff42ec78d573cb139.jpg) +Figure 6. Robustness to noise. SAM2 outputs noisy/coarse masks for frames with considerable motion blur (orange/red). Despite this, InterDyn can generate plausible hand and object dynamics. + +![](images/36c13c41b2208273a87cd005977ac656f1d048b33a94e18c11a2681a793f803e.jpg) +Figure 7. Diverse generation of interactive dynamics. We show multiple challenging examples, such as (from top to bottom): interacting with articulated objects, pouring liquid, letting an object fall, squeezing a highly deformable or "collapsible" object, interacting with reflective objects, tilting a ridged object, or stacking objects. Q Zoom in for details. + +
MethodSSIM ↑PSNR ↑LPIPS ↓FID ↓KID ↓FVD ↓KVD ↓Motion Fidelity ↑ [93]
Seer [23]0.41810.710.58833.350.030287.4681.31—
DynamiCrafter [89]†———17.480.014204.1131.81—
CosHand-Independent [70]0.61516.870.3132.950.00291.1819.240.432
CosHand-Autoregressive [70]0.53114.920.40812.660.01290.3013.680.570
Ours 256×2560.66418.600.2604.950.00419.271.990.633
Ours 256×3840.68019.040.2525.070.00422.222.090.641
+ +Table 1. Quantitative comparison on Something-Something-v2. We compare against two language-instructed video generation methods, Seer [23] and DynamiCrafter [89] and two video extensions of our baseline CosHand [70]. We report results for InterDyn at two resolutions: $256 \times 256$ (matching CosHand) and $256 \times 384$ (matching SVD's prior). Methods denoted with $\dagger$ do not use SSV2 as their training dataset. + +Baselines. To generate videos using CosHand [70] rather than state transitions, we run CosHand in a frame-by-frame approach (CosHand-Independent) and an auto-regressive approach (CosHand-Autoregressive). In the frame-by-frame variant, each future frame is independently predicted from the initial frame and its corresponding mask: + +$$ +\hat {x} _ {t + 1} = \operatorname {C o s H a n d} \left(x _ {0}, h _ {0}, h _ {t + 1}\right), \quad \forall t \in [ 0, 1 3 ], +$$ + +where $h_0$ and $h_{t + 1}$ denote masks, and $x_0$ the initial frame. In the auto-regressive variant, we use CosHand to generate video frames sequentially: + +$$ +\hat {x} _ {t + 1} = \operatorname {C o s H a n d} \left(\hat {x} _ {t}, h _ {t}, h _ {t + 1}\right), \quad \forall t \in [ 0, 1 3 ], +$$ + +with $\hat{x}_t$ being the previously generated frame and $\hat{x}_0 = x_0$ . + +We also compare against two language-instructed video generation methods: Seer [23] and DynamiCrafter [89], which we prompt with a first frame and its corresponding SSV2 class label as instruction. Note that DynamiCrafter was trained with a random video frame as conditioning, i.e. not always the first frame. At inference, its generated videos are thus not strictly a continuation of the conditioning, which precludes frame-aligned metrics. We compare InterDyn against these baselines in Tab. 1. + +Quantitative analysis. The frame-by-frame variant of CosHand achieves high image quality but struggles with temporal coherence and motion fidelity. Qualitatively, we notice that object locations are inconsistent across frames. In contrast, the auto-regressive variant improves object motion fidelity but suffers from lower frame-wise image quality due to error propagation. Both variants fail in scenarios requiring accurate post-interaction dynamics, such as when objects continue moving after being released from direct hand contact, as shown in Fig. 5. + +Our method, InterDyn $(256 \times 384)$ , achieves the best overall performance, surpassing our baseline CosHand in spatio-temporal dynamics, motion fidelity, and all but two image quality metrics. We hypothesize that this might be due to two reasons (1) we use a frozen U-Net, while CosHand fine-tunes its model on SSV2, so CosHand might generate frames closer to the SSV2 distribution, and 2) + +when SVD was trained, it was initialized with SD weights as spatial layers, and then fine-tuned over multiple stages; this might have degraded its spatial prior, and by extension the quality of produced images compared to CosHand. + +Fine-tuned on noisy masks and leveraging its temporal-aware control branch, InterDyn can interpret a noisy control sequence; e.g. when SAM2 produces a coarse and noisy hand mask sequence, InterDyn generates detailed hands including individual fingers, see Fig. 6. Though not always consistent, InterDyn is capable of depicting post-interaction dynamics, such as rolling or sliding objects. + +Qualitative analysis. We present diverse qualitative results generated by InterDyn in Fig. 7. Row 1 shows how InterDyn generates the articulated motion of an object. Row 2 showcases pouring water into a glass; note how the water level increases over time. Row 3 demonstrates an object being dropped, moving out of frame when falling, and rolling back in frame once hitting the floor, featuring realistic motion blur synthesis. Rows 4 and 5 illustrate how InterDyn handles squeezing interactions—the rubber and the spring are compressed and restored accordingly. Row 6 demonstrates an understanding of physical size and distance to the camera, as the phone moves closer to the viewer. These results highlight the complexity that InterDyn can handle, implying its generalization ability and future potential as an implicit, yet generalized physical simulator and renderer. + +# 5. Conclusion + +We introduced InterDyn, a framework that generates videos of interactive dynamics using large video generation models as implicit physics simulators. By incorporating an interactive control mechanism, InterDyn produces plausible, temporally consistent videos of object interactions—including complex human-object interactions—while generalizing to unseen objects. Our evaluations demonstrate that InterDyn effectively captures continuous motion and subsequent dynamics, outperforming baselines that focus on single future states. This work highlights the potential of using video generative models as physics simulators without explicit reconstruction, opening new avenues for future research. + +Acknowledgements: We thank Benjamin Pellkofer for IT support, Peter Kulits for discussions, Tomasz Niewiadomski for coding support, and Allan D. Jepson, Artur Grigorev, Soubhik Sanyal, Omid Taheri & Angjoo Kanazawa for feedback. DT is supported by the ERC Starting Grant (project STRIPES, 101165317). + +Disclosure: https://files.is.tue.mpg.de/black/CoI_CVPR_2025.txt + +DT has received a research gift fund from Google. + +# References + +[1] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A space-time diffusion model for video generation. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-11, 2024. 2, 3 +[2] Fabien Baradel, Natalia Neverova, Julien Mille, Greg Mori, and Christian Wolf. CoPhy: Counterfactual learning of physical dynamics. In International Conference on Learning Representations (ICLR), 2020. 2, 3 +[3] Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, et al. Interaction networks for learning about objects, relations and physics. In Conference on Neural Information Processing Systems (NeurIPS), 2016. 3 +[4] Mikołaj Binkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In International Conference on Learning Representations (ICLR), 2018. 5 +[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable Video Diffusion: Scaling latent video diffusion models to large datasets. arXiv:2311.15127, 2023. 2, 3, 4, 5 +[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 22563-22575, 2023. 3 +[7] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning to follow image editing instructions. In Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, 2023. 14 +[8] Shengqu Cai, Duygu Ceylan, Matheus Gadelha, ChunHao Paul Huang, Tuanfeng Yang Wang, and Gordon Wetzstein. Generative rendering: Controllable 4D-guided video generation with 2D diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 7611-7620, 2024. 3 +[9] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2D pose estimation using part affinity fields. In Computer Vision and Pattern Recognition (CVPR), pages 7291–7299, 2017. 14 +[10] Junuk Cha, Jihyeon Kim, Jae Shin Yoon, and Seungryul Baek. Text2HOI: Text-guided 3D motion generation for hand-object interaction. In Computer Vision and Pattern Recognition (CVPR), pages 1577-1585, 2024. 3 +[11] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S. Narang, Karl + +Van Wyk, Umar Iqbal, Stan Birchfield, et al. DexYCB: A benchmark for capturing hand grasping of objects. In Computer Vision and Pattern Recognition (CVPR), pages 9044-9053, 2021. 14 +[12] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv:2310.19512, 2023. 3 +[13] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Ji-ashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable text-to-video generation with diffusion models. arXiv:2305.13840, 2023. 3 +[14] Enric Corona, Albert Pumarola, Guillem Alenya, Francesc Moreno-Noguer, and Grégory Rogez. GanHand: Predicting human grasp affordances in multi-object scenes. In Computer Vision and Pattern Recognition (CVPR), pages 5031-5041, 2020. 3 +[15] Christian Diller and Angela Dai. CG-HOI: Contact-guided 3D human-object interaction generation. In Computer Vision and Pattern Recognition (CVPR), pages 19888-19901, 2024. 3 +[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 2 +[17] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In Computer Vision and Pattern Recognition (CVPR), pages 12943-12954, 2023. 3 +[18] Zicong Fan, Maria Parelli, Maria Eleni Kadoglou, Xu Chen, Muhammed Kocabas, Michael J. Black, and Otmar Hilliges. HOLD: Category-agnostic 3D reconstruction of interacting hands and objects from video. In Computer Vision and Pattern Recognition (CVPR), pages 494–504, 2024. 3 +[19] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. In International Conference on Robotics and Automation (ICRA), pages 12462-12469, 2024. 3 +[20] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu Video: Factorizing text-to-video generation by explicit image conditioning. In European Conference on Computer Vision (ECCV), 2024. 3 +[21] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The something something video database for learning and evaluating visual common sense. In International Conference on Computer Vision (ICCV), pages 5842-5850, 2017. 2, 4 + +[22] Oliver Groth, Fabian B. Fuchs, Ingmar Posner, and Andrea Vedaldi. ShapeStacks: Learning vision-based physical intuition for generalised object stacking. In European Conference on Computer Vision (ECCV), pages 702-717, 2018. 3 +[23] Xianfan Gu, Chuan Wen, Weirui Ye, Jiaming Song, and Yang Gao. Seer: Language instructed video prediction with latent diffusion models. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 8, 17 +[24] Vincent Le Guen and Nicolas Thome. Disentangling physical dynamics from unknown factors for unsupervised video prediction. In Computer Vision and Pattern Recognition (CVPR), pages 11471-11481, 2020. 2, 3 +[25] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Pengfei Wan, Di Zhang, Yufan Liu, Weiming Hu, Zhengjun Zha, et al. I2V-Adapter: A general image-to-video adapter for diffusion models. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2024. 3 +[26] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. *AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning*. arXiv:2307.04725, 2023. 3 +[27] Yana Hasson, Gül Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Computer Vision and Pattern Recognition (CVPR), pages 11807-11816, 2019. 3 +[28] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from RGB videos. In International Conference on 3D Vision (3DV), pages 659-668, 2021. 3 +[29] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. CameraCtrl: Enabling camera control for text-to-video generation. arXiv:2404.02101, 2024. 3 +[30] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Conference on Neural Information Processing Systems (NeurIPS), pages 6629-6640, 2017. 5 +[31] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 5 +[32] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, et al. Image Video: High definition video generation with diffusion models. arXiv:2210.02303, 2022. 3 +[33] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In Conference on Neural Information Processing Systems (NeurIPS), pages 8633-8646, 2022. 3 +[34] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 3 + +[35] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate Anyone: Consistent and controllable image-to-video synthesis for character animation. In Computer Vision and Pattern Recognition (CVPR), pages 8153-8163, 2024. 3 +[36] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. DepthCrafter: Generating consistent long depth sequences for open-world videos. arXiv:2409.02095, 2024. 4 +[37] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make It Move: Controllable image-to-video generation with text descriptions. In Computer Vision and Pattern Recognition (CVPR), pages 18219-18228, 2022. 3 +[38] Zhihao Hu and Dong Xu. VideoControlNet: A motion-guided video-to-video translation framework by using diffusion model with ControlNet. arXiv:2307.14073, 2023. 3 +[39] Steeven Janny, Fabien Baradel, Natalia Neverova, Madiha Nadri, Greg Mori, and Christian Wolf. Filtered-CoPhy: Unsupervised learning of counterfactual physics in pixel space. In International Conference on Learning Representations (ICLR), 2022. 2, 3 +[40] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In International Conference on Computer Vision (ICCV), pages 11107-11116, 2021. 3 +[41] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker3: Simpler and better point tracking by pseudolabelling real videos. In European Conference on Computer Vision (ECCV), pages 18-35, 2024. 5 +[42] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. DreamPose: Fashion image-to-video synthesis via stable diffusion. In International Conference on Computer Vision (ICCV), pages 22623-22633, 2023. 3 +[43] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Conference on Neural Information Processing Systems (NeurIPS), 35:26565-26577, 2022. 5 +[44] Hyeonwoo Kim, Sookwan Han, Patrick Kwon, et al. Beyond the Contact: Discovering comprehensive affordance for 3D objects from pre-trained 2D diffusion models. In European Conference on Computer Vision (ECCV), pages 400-419, 2024. 3 +[45] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 5 +[46] Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems. In International Conference on Machine Learning (ICML), pages 2693-2702, 2018. 2 +[47] Sumith Kulal, Tim Brooks, Alex Aiken, Jiajun Wu, Jimei Yang, Jingwan Lu, Alexei A. Efros, and Krishna Kumar Singh. Putting people in their place: Affordance-aware human insertion into scenes. In Computer Vision and Pattern Recognition (CVPR), pages 17089-17099, 2023. 3 + +[48] Adam Lerer, Sam Gross, and Rob Fergus. Learning physical intuition of block towers by example. In International Conference on Machine Learning (ICML), pages 430-438, 2016. 3 +[49] Yunzhu Li, Antonio Torralba, Anima Anandkumar, Dieter Fox, and Animesh Garg. Causal discovery in physical systems from videos. Conference on Neural Information Processing Systems (NeurIPS), 2020. 2 +[50] Ce Liu, Antonio Torralba, William T. Freeman, Frédo Durand, and Edward H. Adelson. Motion magnification. Transactions on Graphics (TOG), 24(3):519-526, 2005. 5 +[51] Ruibo Liu, Jason Wei, Shixiang Shane Gu, Te-Yen Wu, Soroush Vosoughi, Claire Cui, Denny Zhou, and Andrew M Dai. Mind's Eye: Grounded language model reasoning through simulation. In International Conference on Learning Representations (ICLR), 2023. 3 +[52] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. arXiv:2303.05499, 2023. 14 +[53] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. PhysGen: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision (ECCV), pages 360-378, 2024. 3 +[54] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv:2402.17177, 2024. 2, 4 +[55] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T. Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. Transactions on Graphics (TOG), 39(6): 1-14, 2020. 3 +[56] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T. Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In Computer Vision and Pattern Recognition (CVPR), pages 4507-4515, 2021. 3 +[57] Jiaxin Lu, Hao Kang, Haoxiang Li, Bo Liu, Yiding Yang, Qixing Huang, and Gang Hua. UGG: Unified generative grasping. In European Conference on Computer Vision (ECCV), pages 414-433, 2024. 3 +[58] Wan-Duo Kurt Ma, John P Lewis, and W. Bastiaan Kleijn. TrailBlazer: Trajectory control for diffusion-based video generation. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-11, 2023. 3 +[59] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-Else: Compositional action recognition with spatial-temporal interaction networks. In Computer Vision and Pattern Recognition (CVPR), pages 1049–1059, 2020. 6 +[60] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to + +image diffusion models. In AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 3 +[61] Franziska Mueller, Micah Davis, Florian Bernard, Oleksandr Sotnychenko, Micekal Verschooor, Miguel A. Otaduy, Dan Casas, and Christian Theobalt. Real-time pose and shape reconstruction of two interacting hands with a single depth camera. Transactions on Graphics (TOG), 38(4): 1-13, 2019. 14 +[62] Boxiao Pan, Zhan Xu, Chun-Hao Paul Huang, Krishna Kumar Singh, Yang Zhou, Leonidas J. Guibas, and Jimei Yang. ActAnywhere: Subject-aware video background generation. In Conference on Neural Information Processing Systems (NeurIPS), 2024. 3 +[63] Georgios Paschalidis, Romana Wilschut, Dimitrije Antic, Omid Taheri, and Dimitrios Tzionas. 3D whole-body grasp synthesis with directional controllability. In International Conference on 3D Vision (3DV), 2025. 3 +[64] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 4 +[65] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Girshick, Piotr Dólar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 6, 14 +[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 4, 14 +[67] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. Transactions on Graphics (TOG), 36(6), 2022. 14 +[68] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-A-Video: Text-to-video generation without text-video data. In International Conference on Learning Representations (ICLR), 2023. 3 +[69] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. StyleGAN-V: A continuous video generator with the price, image quality and perks of StyleGAN2. In Computer Vision and Pattern Recognition (CVPR), pages 3626-3636, 2022. 5 +[70] Sruthi Sudhakar, Ruoshi Liu, Basile Van Hoorick, Carl Vondrick, and Richard Zemel. Controlling the world by sleight of hand. In European Conference on Computer Vision (ECCV), pages 414-430, 2024. 2, 3, 6, 8, 14, 17 +[71] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), pages 581-600, 2020. 3 + +[72] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J. Black. GRIP: Generating interaction poses using latent consistency and spatial cues. In International Conference on 3D Vision (3DV), pages 933-943, 2024. 3 +[73] Tze Ho Elden Tse, Kwang In Kim, Ales Leonardis, and Hyung Jin Chang. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In Computer Vision and Pattern Recognition (CVPR), pages 1664-1674, 2022. 3 +[74] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation. In International Conference on Learning Representations Workshops (ICLRw), 2019. 5 +[75] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. In European Conference on Computer Vision (ECCV), pages 439-457, 2025. 4, 14 +[76] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv:2308.06571, 2023. 3 +[77] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. In International Conference on Machine Learning (ICML), 2024. 3 +[78] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional video synthesis with motion controllability. In Conference on Neural Information Processing Systems (NeurIPS), 2024. 3 +[79] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image Quality Assessment: From error visibility to structural similarity. Transactions on Image Processing (TIP), 13(4):600-612, 2004. 3, 5 +[80] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A unified and flexible motion controller for video generation. In Transactions on Graphics (TOG), pages 1-11, 2024. 3 +[81] Nicholas Watters, Daniel Zoran, Theophane Weber, Peter Battaglia, Razvan Pascanu, and Andrea Tacchetti. Visual Interaction Networks: Learning a physics simulator from video. In Conference on Neural Information Processing Systems (NeurIPS), pages 4539-4547, 2017. 3 +[82] Jiajun Wu, Ilker Yildirim, Joseph J. Lim, Bill Freeman, and Joshua B. Tenenbaum. Galileo: Perceiving physical object properties by integrating a physics engine with deep learning. In Conference on Neural Information Processing Systems (NeurIPS), pages 127-135, 2015. 3 +[83] Jiajun Wu, Joseph J. Lim, Hongyi Zhang, Joshua B. Tenenbaum, and William T. Freeman. Physics 101: Learning physical object properties from unlabeled videos. In *British Machine Vision Conference (BMVC)*, 2016. +[84] Jiajun Wu, Erika Lu, Pushmeet Kohli, Bill Freeman, and Josh Tenenbaum. Learning to see physics via visual de + +animation. In Conference on Neural Information Processing Systems (NeurIPS), pages 153-164, 2017. 2, 3 +[85] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-A-Video: One-shot tuning of image diffusion models for text-to-video generation. In International Conference on Computer Vision (ICCV), pages 7623–7633, 2023. 3 +[86] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. THOR: Text to human-object interaction diffusion via relation intervention. arXiv:2403.11208, 2024. 3 +[87] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. DragAnything: Motion control for anything using entity representation. In European Conference on Computer Vision (ECCV), pages 331-348, 2025. 3 +[88] Xianghui Xie, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Template free reconstruction of human-object interaction with procedural interaction generation. In Computer Vision and Pattern Recognition (CVPR), pages 10003-10015, 2024. 3 +[89] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. DynamiCrafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision (ECCV), pages 399-417, 2025. 2, 3, 8, 17 +[90] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3D human-object interactions with physics-informed diffusion. In International Conference on Computer Vision (ICCV), pages 14928-14940, 2023. 3 +[91] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. MagicAnimate: Temporally consistent human image animation using diffusion model. In Computer Vision and Pattern Recognition (CVPR), pages 1481–1490, 2024. 3 +[92] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. CogVideoX: Text-to-video diffusion models with an expert transformer. arXiv:2408.06072, 2024.3 +[93] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In Computer Vision and Pattern Recognition (CVPR), pages 8466-8476, 2024. 5, 8, 14, 17 +[94] Yufei Ye, Poorvi Hebbar, Abhinav Gupta, and Shubham Tulsiani. Diffusion-guided reconstruction of everyday hand-object interaction clips. In International Conference on Computer Vision (ICCV), pages 19717-19728, 2023. 3 +[95] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance Diffusion: Synthesizing hand-object interactions. In Computer Vision and Pattern Recognition (CVPR), pages 22479-22489, 2023. 3 + +[96] Yufei Ye, Abhinav Gupta, Kris Kitani, and Shubham Tulsiani. G-HOP: Generative hand-object prior for interaction reconstruction and grasp synthesis. In Computer Vision and Pattern Recognition (CVPR), pages 1911–1920, 2024. 3 +[97] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Ji-jun Wu, Antonio Torralba, and Joshua B. Tenenbaum. CLEVRR: Collision events for video representation and reasoning. In International Conference on Learning Representations (ICLR), 2020. 2, 4, 6 +[98] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. DragNUWA: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv:2308.08089, 2023. 3 +[99] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Computer Vision and Pattern Recognition (CVPR), pages 18456-18466, 2023. 3 +[100] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3, 4 +[101] Mengqi Zhang, Yang Fu, Zheng Ding, Sifei Liu, Zhuowen Tu, and Xiaolong Wang. HOIDiffusion: Generating realistic 3D hand-object interaction data. In Computer Vision + +and Pattern Recognition (CVPR), pages 8521-8531, 2024. 3 +[102] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5 +[103] Juntian Zheng, Qingyuan Zheng, Lixing Fang, Yun Liu, and Li Yi. CAMS: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis. In Computer Vision and Pattern Recognition (CVPR), pages 585-594, 2023. 3 +[104] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient video generation with latent diffusion models. arXiv:2211.11018, 2022.3 +[105] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. GEARS: Local geometry-aware hand-object interaction synthesis. In Computer Vision and Pattern Recognition (CVPR), pages 20634-20643, 2024. 3 +[106] Tianqiang Zhu, Rina Wu, Xiangbo Lin, and Yi Sun. Toward human-like grasp: Dexterous grasping via semantic representation of object-hand. In International Conference on Computer Vision (ICCV), pages 15741-15751, 2021. 3 \ No newline at end of file diff --git a/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/images.zip b/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..85258852c45921379ca8aaf716e4165aa5a8be32 --- /dev/null +++ b/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8260f3d1748b210994b9e7280e4f76d13b3e3f6164a6de47a881ce426cde8052 +size 607578 diff --git a/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/layout.json b/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..df174384ba952c201c20d98a5350902ce390fe10 --- /dev/null +++ b/2025/InterDyn_ Controllable Interactive Dynamics with Video Diffusion Models/layout.json @@ -0,0 +1,8913 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 100, + 533, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 100, + 533, + 118 + ], + "spans": [ + { + "bbox": [ + 74, + 100, + 533, + 118 + ], + "type": "text", + "content": "InterDyn: Controllable Interactive Dynamics with Video Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 137, + 550, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 137, + 550, + 152 + ], + "spans": [ + { + "bbox": [ + 63, + 137, + 550, + 152 + ], + "type": "text", + "content": "Rick Akkerman" + }, + { + "bbox": [ + 63, + 137, + 550, + 152 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 63, + 137, + 550, + 152 + ], + "type": "text", + "content": " Haiwen Feng" + }, + { + "bbox": [ + 63, + 137, + 550, + 152 + ], + "type": "inline_equation", + "content": "^{1*†}" + }, + { + "bbox": [ + 63, + 137, + 550, + 152 + ], + "type": "text", + "content": " Michael J. Black" + }, + { + "bbox": [ + 63, + 137, + 550, + 152 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 137, + 550, + 152 + ], + "type": "text", + "content": " Dimitrios Tzionas" + }, + { + "bbox": [ + 63, + 137, + 550, + 152 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 137, + 550, + 152 + ], + "type": "text", + "content": " Victoria Fernández Abrevaya" + }, + { + "bbox": [ + 63, + 137, + 550, + 152 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 153, + 512, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 153, + 512, + 165 + ], + "spans": [ + { + "bbox": [ + 100, + 153, + 512, + 165 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 153, + 512, + 165 + ], + "type": "text", + "content": "Max Planck Institute for Intelligent Systems, Tübingen, Germany " + }, + { + "bbox": [ + 100, + 153, + 512, + 165 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 153, + 512, + 165 + ], + "type": "text", + "content": "University of Amsterdam, the Netherlands" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 167, + 548, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 167, + 548, + 180 + ], + "spans": [ + { + "bbox": [ + 64, + 167, + 548, + 180 + ], + "type": "text", + "content": "{rick.akkerman, haiwen.feng, black, victoria.abrevaya}@tuebingen.mpg.de d.tzionas@uva.nl" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 209, + 138, + 264 + ], + "blocks": [ + { + "bbox": [ + 76, + 196, + 118, + 206 + ], + "lines": [ + { + "bbox": [ + 76, + 196, + 118, + 206 + ], + "spans": [ + { + "bbox": [ + 76, + 196, + 118, + 206 + ], + "type": "text", + "content": "Input image" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 209, + 138, + 264 + ], + "lines": [ + { + "bbox": [ + 57, + 209, + 138, + 264 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 138, + 264 + ], + "type": "image", + "image_path": "30b0ac9f68c40219c46eb31fd49f92bbdee1d119f6ed39037577255457f9180b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 147, + 209, + 552, + 271 + ], + "blocks": [ + { + "bbox": [ + 219, + 196, + 493, + 207 + ], + "lines": [ + { + "bbox": [ + 219, + 196, + 493, + 207 + ], + "spans": [ + { + "bbox": [ + 219, + 196, + 493, + 207 + ], + "type": "text", + "content": "Video generation by InterDyn using only the hand mask sequence as control signal" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 147, + 209, + 552, + 271 + ], + "lines": [ + { + "bbox": [ + 147, + 209, + 552, + 271 + ], + "spans": [ + { + "bbox": [ + 147, + 209, + 552, + 271 + ], + "type": "image", + "image_path": "5f263dc19b34cc956bce9bec510d1cc9ba36ef6e51f61f7573a896f776fb1e9a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 58, + 274, + 138, + 328 + ], + "blocks": [ + { + "bbox": [ + 58, + 274, + 138, + 328 + ], + "lines": [ + { + "bbox": [ + 58, + 274, + 138, + 328 + ], + "spans": [ + { + "bbox": [ + 58, + 274, + 138, + 328 + ], + "type": "image", + "image_path": "113443254ca639e65d103d521d5dce89e0f8d67c2dd453fa94c33f7d5b93205a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 147, + 273, + 552, + 335 + ], + "blocks": [ + { + "bbox": [ + 147, + 273, + 552, + 335 + ], + "lines": [ + { + "bbox": [ + 147, + 273, + 552, + 335 + ], + "spans": [ + { + "bbox": [ + 147, + 273, + 552, + 335 + ], + "type": "image", + "image_path": "02312659b0585085ead49a474af9732b321279cb8a80bb0c106e2a1d61995635.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 58, + 359, + 295, + 414 + ], + "blocks": [ + { + "bbox": [ + 142, + 349, + 208, + 359 + ], + "lines": [ + { + "bbox": [ + 142, + 349, + 208, + 359 + ], + "spans": [ + { + "bbox": [ + 142, + 349, + 208, + 359 + ], + "type": "text", + "content": "Force propagation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 359, + 295, + 414 + ], + "lines": [ + { + "bbox": [ + 58, + 359, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 58, + 359, + 295, + 414 + ], + "type": "image", + "image_path": "1aabfaedc7193a34b5fdd250b073f50a42f2a211a73bba63d8952a9bb96e6e4d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 359, + 552, + 421 + ], + "blocks": [ + { + "bbox": [ + 385, + 348, + 474, + 357 + ], + "lines": [ + { + "bbox": [ + 385, + 348, + 474, + 357 + ], + "spans": [ + { + "bbox": [ + 385, + 348, + 474, + 357 + ], + "type": "text", + "content": "Counterfactual dynamics" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 359, + 552, + 421 + ], + "lines": [ + { + "bbox": [ + 309, + 359, + 552, + 421 + ], + "spans": [ + { + "bbox": [ + 309, + 359, + 552, + 421 + ], + "type": "image", + "image_path": "ee346ad8c41e11075b1455634b058139e7def2b188313e9387be0341fac73f0e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 423, + 477, + 434 + ], + "lines": [ + { + "bbox": [ + 115, + 423, + 477, + 434 + ], + "spans": [ + { + "bbox": [ + 115, + 423, + 477, + 434 + ], + "type": "text", + "content": "Denotes a driving object motion | Tracks indicate an object with generated uncontrolled dynamics" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 54, + 447, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 54, + 447, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 54, + 447, + 555, + 492 + ], + "type": "text", + "content": "Figure 1. We present InterDyn, a framework for synthesizing realistic interactive dynamics without 3D reconstruction and physics simulation. Our core principle is to rely on the implicit physics knowledge embedded in large-scale video generative models. Given an image and a \"driving motion\", our model generates the consequential scene dynamics. We investigate the generated interactive dynamics in a simple object collision scenario (bottom) and complex in-the-wild human-object interaction (top)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 151, + 511, + 200, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 511, + 200, + 523 + ], + "spans": [ + { + "bbox": [ + 151, + 511, + 200, + 523 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 53, + 537, + 297, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 537, + 297, + 682 + ], + "spans": [ + { + "bbox": [ + 53, + 537, + 297, + 682 + ], + "type": "text", + "content": "Predicting the dynamics of interacting objects is essential for both humans and intelligent systems. However, existing approaches are limited to simplified, toy settings and lack generalizability to complex, real-world environments. Recent advances in generative models have enabled the prediction of state transitions based on interventions, but focus on generating a single future state which neglects the continuous dynamics resulting from the interaction. To address this gap, we propose InterDyn, a novel framework that generates videos of interactive dynamics given an initial frame and a control signal encoding the motion of a driving object or actor. Our key insight is that large video" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 513, + 556, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 556, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 556, + 681 + ], + "type": "text", + "content": "generation models can act as both neural renderers and implicit physics \"simulators\", having learned interactive dynamics from large-scale video data. To effectively harness this capability, we introduce an interactive control mechanism that conditions the video generation process on the motion of the driving entity. Qualitative results demonstrate that InterDyn generates plausible, temporally consistent videos of complex object interactions while generalizing to unseen objects. Quantitative evaluations show that InterDyn outperforms baselines that focus on static state transitions. This work highlights the potential of leveraging video generative models as implicit physics engines. Code and trained models will be released at: https://interdyn.is.tue.mpg.de/." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 136, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 71, + 703, + 114, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 703, + 114, + 712 + ], + "spans": [ + { + "bbox": [ + 71, + 703, + 114, + 712 + ], + "type": "text", + "content": "†Project lead" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12467" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 210 + ], + "type": "text", + "content": "Humans have the remarkable ability to predict the future dynamics of observed systems intuitively. With just a single image, we can anticipate and imagine how objects will move over time – not only their motion but also their interactions with the environment and other elements in the scene. Inferring this requires an advanced form of scene-level reasoning beyond merely recognizing the semantics and geometry of static elements; it involves a deep physical and causal understanding of how each object will interact given the environment, object properties, and forces." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 211, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 296, + 378 + ], + "type": "text", + "content": "There has been a growing interest in developing machine learning systems that emulate similar levels of dynamic understanding given visual observations, such as images or videos. Early work [84] addressed this by first reconstructing a 3D representation from the image, then predicting future states with a physics simulator and finally generating the output video with a rendering engine. This relies heavily on explicit reconstruction and simulation, which is computationally intensive, prone to errors, and may not generalize well. More recent methods [2, 24, 39, 46, 49] leverage keypoint or latent representations within graph relational frameworks; however, they have only been trained and validated in over-simplified, synthetic environments, showing limited generalizability to complex real-world scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 379, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 666 + ], + "type": "text", + "content": "Instead, the advent of powerful generative models [1, 5, 16, 54, 66] opens new avenues for synthesizing interactions under complex scenarios. For example, Sudhakar et al. [70] recently proposed CosHand, a controllable image-to-image model based on Stable Diffusion [66] that infers state transitions of an object. The task here is defined as follows: given an image of a hand interacting with an object, alongside a hand mask of the current frame and a mask of the hand at a future frame, generate a modified input image that satisfies the mask, with realistic interactions. The challenge, as in early intuitive physics works, lies in accurately modeling how the objects will change after forces are applied. However, we argue that static state transitions are insufficient for this task, as they fail to capture the continuous dynamic processes inherent to the problem, e.g. see Fig. 2. Investigating interactive dynamics within a two-state setting is highly limiting, since dynamics can extend beyond the period of direct contact – for example, predicting the motion occurring while a person pours water requires a physical understanding that goes beyond the state of the hand at a future frame. The driving force, in this case the hand, may interact with the system only briefly, but the system's subsequent dynamics continue according to physical laws and may even influence other parts via force propagation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "In this paper, we explore controllable synthesis of interactive dynamics—generating a video from an input image and a dynamic control signal (e.g. a moving hand mask) to model realistic object dynamics. In particular, we propose" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 554, + 177 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 554, + 177 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 554, + 177 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 554, + 177 + ], + "type": "image", + "image_path": "ebe545c08fdf1434341070911268a6e0cf3fd3097f7b639364bfe10dc5ebd701.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 185, + 555, + 241 + ], + "lines": [ + { + "bbox": [ + 313, + 185, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 555, + 241 + ], + "type": "text", + "content": "Figure 2. State transition vs. dynamics. Methods that generate static state transitions (i.e. predict a future image) such as CosHand [70] struggle to capture the inherent dynamic processes involved in human-object interactions. Here, we show a video sequence where the motion continues beyond the interaction." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 255, + 555, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 375 + ], + "type": "text", + "content": "InterDyn, a novel framework for synthesizing controllable dynamic interactions that leverages the physical and dynamics \"knowledge\" of a large video generation model [5]. Unlike prior approaches that rely on explicit physical simulation [84] or are constrained to static state transitions [70], we leverage video generation models to generate dynamic processes implicitly, see Fig. 1. Specifically, we extend Stable Video Diffusion (SVD) [5] with a dynamic control branch and fine-tune it on diverse scenes, enabling synthesis of complex interactions aligned with the control signal." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 376, + 556, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 556, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 556, + 520 + ], + "type": "text", + "content": "We start our investigation by fine-tuning InterDyn on a simple synthetic scenario of cubes, cylinders, and spheres: the CLEVRER dataset [97]. To control the motion we add a mask driving signal that manipulates the movement of some (but not all) of the objects in the scene. We then evaluate how the synthesized trajectories of uncontrolled objects change under various interactions, including multiple objects colliding with each other. This multi-object collision setting allows us to \"probe\" the physical understanding and causal effects of the video diffusion model, and our qualitative experiments show InterDyn's ability for counterfactual future prediction and physical force propagation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "Next, we evaluate how the system performs in a difficult real-world scenario, such as Human-Object Interaction (HOI). Here, the dexterity of hand motions and the diversity of objects vastly increase the complexity of the problem. We fine-tune the model on a commonly used HOI video dataset [21] and compare with the state-of-the-art baseline CosHand [70], as well as two text-control based interactive dynamics generation methods: Seer [23] and DynamiCrafter [89]. We quantify our investigations using standard image and video metrics, as well as a motion fidelity metric based on point tracking. InterDyn surpasses the previous SOTA over " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "37.5\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " on LPIPS and " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "77\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " on FVD on the Something-Something-v2 (SSV2) dataset [21]. Our experiments also demonstrate diverse and physically plausible generations of interactive dynamics, probing into SVD's \"understanding\" of physics and dynamics." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12468" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": "In summary, we present InterDyn, a framework that employs video generative models to simulate object dynamics without explicit 3D reconstruction or physical simulation. We demonstrate how the inherent \"knowledge\" within video generation models can be leveraged to predict complex object interactions and movements over time, implicitly modeling physical and causal dynamics. We perform comprehensive experiments on multi-object collision datasets and hand-object manipulation datasets, demonstrating the effectiveness of our approach." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 208, + 142, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 142, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 142, + 220 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 229, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 296, + 552 + ], + "type": "text", + "content": "Modeling human-object interactions (HOI). Human-object interaction has been widely studied within the context of 3D reconstruction [17, 18, 27, 28, 73, 88, 94], where the goal is to recover realistic geometry of hands and objects. The field of 3D HOI synthesis has also received increasing attention, including the generation of static [40, 44, 71, 106] or dynamic [63, 72, 103, 105] hand poses conditioned on 3D objects, whole-body interactions [90], or more recently, hand-object meshes given textual descriptions [10, 15, 57, 86, 96]. Few works address HOI synthesis in the 2D domain. GANHand [14] predicts 3D hand shape and pose given an RGB image of an object, while AffordanceDiffusion [95] estimates a 2D hand using a diffusion model. Kulal et al. [47] take as input an image of a human and a scene separately and generate a composite image that positions the human with correct affordances. Also relevant is HOIDiffusion [101], in which a texture-less rendering of a 3D hand and object is converted to a realistic image using a text description. Most closely related to us is CosHand [70], which takes as input an RGB image of a hand-object interaction, a hand mask at the current state, and the hand mask of the future state, and generates an RGB image of the future state. Unlike us, they cannot generate post-interaction object dynamics. Importantly, none of these works study dynamics, generating instead discrete state transitions that fail to capture the nuanced, temporally coherent behaviors observed in interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 570 + ], + "type": "text", + "content": "Synthesizing causal physical relations from visual input." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "A growing body of work aims to model and predict physical causal effects from visual inputs such as images or videos. For example, research in intuitive physics seeks to replicate the human-like, non-mathematical understanding of physical events, e.g. by predicting future frames given an input video. Early works like [22, 48] train neural networks to assess the stability of block towers, while [24] leverage prior physical knowledge formalized through partial differential equations (PDEs). Other approaches investigate counterfactual reasoning by leveraging graph neural networks [2, 39]. Wu et al. [82-84] explore the use of an inverse rendering approach, extracting geometry and physical properties from" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "the video which are then coupled with a physics simulator and a rendering engine to generate the future frames. Other works [81] incorporate Interaction Networks [3] to approximate physical systems from video data. These approaches are often limited to simplified, synthetic datasets and struggle to generalize to real-world scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 154, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 555, + 429 + ], + "type": "text", + "content": "Recent methods have started to combine language models with physical engines. Liu et al. [51] ground a large language model using a computational physics engine while Gao et al. [19] show that fine-tuning a vision-language model (VLM) on annotated datasets of physical concepts improves its understanding of physical interactions. Closely related to our work is PhysGen [53], which trains an imaged-to-video model that conditions the video generation on physics parameters (e.g., force or torque). However, the model relies on a dynamics simulator to generate motion, and its application is limited to rigid objects. A related but tangential line of work focuses on identifying and generating the effects of objects on their surroundings. For example, Omnimatte [56] introduces the problem of identifying all parts of a scene influenced by an object, given a video and a mask of the object. Similarly, Lu et al. [55] propose to re-time the motion of different subjects in a scene while maintaining realistic interactions with the environment. ActAnywhere [62] generates videos with plausible human-scene interactions, taking a masked video of a person and a background image as input. These works address the problem of synthesizing realistic interactions within a scene, however, lack fine-grained control." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": "Controllable video generation. Video generation has advanced significantly in recent years, with diffusion models leading to substantial improvements in unconditional [33, 99], text-based [1, 5, 6, 12, 20, 23, 26, 32, 37, 68, 76, 85, 89, 92, 104] and image-based [1, 5, 20, 25, 78, 89] generation. These advances have raised the question of how to incorporate more nuanced control into video generation. Some text-to-video approaches are trained by \"inflating\" text-to-image (T2V) models [8, 13, 25, 26, 38, 85], and can thus be integrated with conditional T2V models such as ControlNet [100] or T2V-Adapter [60]. Control can also be achieved by conditioning on trajectories [58, 87, 98] or bounding-boxes [77], fine-tuning on appropriate datasets. VideoComposer [78] incorporates multiple condition types, including text, depth, style, and temporal conditions via motion vectors. Camera motion control has also been explored, withAnimateDiff [79] employing LoRA [34] modules to control camera movement, while MotionCtrl [80] and CameraCtrl [29] directly embed the camera information for more precise control. Additionally, several works target human animation from a pose control signal, such as DreamPose [42], MagicPose [91], andAnimateAnyone [35], but do not account for interactions." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12469" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 248, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 248, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 248, + 85 + ], + "type": "text", + "content": "3. Controllable Interactive Dynamics" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 212 + ], + "type": "text", + "content": "Video diffusion models such as [5, 54] have demonstrated impressive performance in generating videos from text or images, and have even shown potential in tasks that require 3D understanding when properly fine-tuned [36, 75]. Trained on millions of videos, we hypothesize that these models also possess implicit knowledge of complex interactive dynamics, such as those that appear when humans interact with objects. Out of the box, however, they lack a precise control mechanism, often relying solely on textual inputs or requiring careful selection of the starting frame." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 214, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 214, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 296, + 347 + ], + "type": "text", + "content": "Task. Given an input image, " + }, + { + "bbox": [ + 55, + 214, + 296, + 347 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^{1 \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 214, + 296, + 347 + ], + "type": "text", + "content": ", and a driving motion in the form of a pixel-wise corresponding control signal " + }, + { + "bbox": [ + 55, + 214, + 296, + 347 + ], + "type": "inline_equation", + "content": "\\pmb{c} \\in \\mathbb{R}^{N \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 214, + 296, + 347 + ], + "type": "text", + "content": ", we task InterDyn with generating a video sequence, " + }, + { + "bbox": [ + 55, + 214, + 296, + 347 + ], + "type": "inline_equation", + "content": "\\pmb{y} \\in \\mathbb{R}^{N \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 214, + 296, + 347 + ], + "type": "text", + "content": ", depicting plausible object dynamics. Through this task, we aim to learn the conditional distribution between a driving motion, such as that of a human hand, and the consequent motion of manipulated objects. In other words, the model needs to synthesize plausible object movement and appearance without any indication other than the driving motion, while maintaining physical and visual consistency with the input image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 542 + ], + "type": "text", + "content": "Stable Video Diffusion. We extend Stable Video Diffusion [5] (SVD) to enable controllable interactive dynamics and explore the versatility of this model across a range of scenarios. SVD is a publicly available U-Net-based latent diffusion model [66] that extends Stable Diffusion 2.1 to video generation by interleaving the network with temporal layers. Given a static input image of a scene, SVD denoises a sequence of " + }, + { + "bbox": [ + 55, + 350, + 296, + 542 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 350, + 296, + 542 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 55, + 350, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\pmb{y} \\in \\mathbb{R}^{N \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 350, + 296, + 542 + ], + "type": "text", + "content": " to generate a video that follows the initial frame. The input image is fed into the denoising U-Net by concatenating its latent to each of the frames in the noised input, and by supplying its CLIP [64] embedding to the U-Net's cross-attention layers. In addition, SVD is conditioned on the video's FPS and motion ID, where the motion ID represents the amount of motion in the video. We found a motion ID of 40 to align well with our frozen SVD prior." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Control. InterDyn extends SVD with an additional control signal " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^{N \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " by integrating a ControlNet-like branch [100]. An overview of our pipeline is presented in Fig. 3. The SVD weights remain frozen to preserve its learned dynamics prior. Following [100], we introduce a trainable copy of the SVD encoder " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": ", connected to SVD's frozen decoder via skip connections, and modulated by zero-initialized convolutions. We use a small CNN, " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": ", to encode the control signal " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " into the latent space, which is then added to the noisy input latent that is passed to the ControlNet encoder. Similar to SVD, the control branch interleaves convolutional, spatial, and temporal blocks, enabling InterDyn to process the control signal in a temporal-aware manner. This helps InterDyn to be robust to" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 554, + 205 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 554, + 205 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 554, + 205 + ], + "type": "image", + "image_path": "9046525db769913ab2d00853ff28fe6ab5804a5309900a6555b944e893f40745.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 213, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 324 + ], + "type": "text", + "content": "Figure 3. Overview of InterDyn. Given an input image depicting a scene, such as a hand holding a remote, and a \"driving motion,\" such as a sequence of binary hand masks, InterDyn generates a video depicting plausible hand and object dynamics. Crucially, InterDyn receives no control signal for the object. Through this setup, we probe and assess the implicit knowledge of large video generation models on complex interactive dynamics. We use Stable Video Diffusion (SVD) as our frozen backbone and fine-tune a separate control signal encoder. Videos are iteratively denoised over " + }, + { + "bbox": [ + 313, + 213, + 555, + 324 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 213, + 555, + 324 + ], + "type": "text", + "content": " timesteps, starting from Gaussian noise " + }, + { + "bbox": [ + 313, + 213, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 213, + 555, + 324 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 344, + 555, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 555, + 416 + ], + "type": "text", + "content": "noisy control signals, see Fig. 6. We opt for binary masks as conditioning signal due to their accessibility. However, our method can be extended to incorporate diverse types of signals. We find that for hand-object interactions, the type of conditioning signal does not significantly impact performance, see Appendix A and Tab. S1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 419, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 554, + 492 + ], + "type": "text", + "content": "Inference. During inference, we start from an input image, control signal sequence, and randomly sampled Gaussian noise " + }, + { + "bbox": [ + 313, + 419, + 554, + 492 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 419, + 554, + 492 + ], + "type": "text", + "content": ". Through iteratively applying InterDyn over " + }, + { + "bbox": [ + 313, + 419, + 554, + 492 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 419, + 554, + 492 + ], + "type": "text", + "content": " denoising timesteps, we generate a video " + }, + { + "bbox": [ + 313, + 419, + 554, + 492 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 313, + 419, + 554, + 492 + ], + "type": "text", + "content": " depicting plausible hand and object dynamics, aligned with the control signal." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 502, + 395, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 395, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 395, + 515 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": "The primary goal of this work is to synthesize scene-level interactive dynamics by leveraging the implicit physical understanding of a pre-trained video generative model. We begin by probing the model's ability to predict physically plausible outcomes within simulated environments, specifically using the CLEVRER dataset [97]. We test force propagation amongst uncontrolled objects and examine counterfactual future prediction by generating videos for one input image with different control signals. Motivated by promising results, we extend our investigation to complex, real-world hand-object interaction scenarios using the Something-Something-v2 (SSV2) dataset [21], conducting comprehensive comparisons with existing baselines that pursue similar objectives. Additionally, we showcase diverse physical examples to demonstrate the capabilities of InterDyn in generating realistic interactive dynamics." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12470" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 555, + 140 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 555, + 140 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 555, + 140 + ], + "type": "image", + "image_path": "e301e10426e190d9050c359053ddaf95d49c1b84a0e16be2b8ff73b60fa5d86c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 144, + 537, + 156 + ], + "lines": [ + { + "bbox": [ + 72, + 144, + 537, + 156 + ], + "spans": [ + { + "bbox": [ + 72, + 144, + 537, + 156 + ], + "type": "text", + "content": "(a) Force Propagation. InterDyn generates force propagation to and amongst uncontrolled objects (cyan cylinder, pink cylinder, and pink cube)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 160, + 555, + 286 + ], + "blocks": [ + { + "bbox": [ + 56, + 160, + 555, + 286 + ], + "lines": [ + { + "bbox": [ + 56, + 160, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 160, + 555, + 286 + ], + "type": "image", + "image_path": "dba23c81a0e46e6701a9bde0c25c5a11a0c1513d3dd0d3519aa619850b75768a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 289, + 552, + 301 + ], + "lines": [ + { + "bbox": [ + 56, + 289, + 552, + 301 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 552, + 301 + ], + "type": "text", + "content": "(b) Counterfactual Dynamics. Two generations of InterDyn on the same input image: (1) w/ brown-cylinder control and (2) w/o brown-cylinder control." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 310, + 555, + 355 + ], + "lines": [ + { + "bbox": [ + 54, + 310, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 54, + 310, + 555, + 355 + ], + "type": "text", + "content": "Figure 4. Qualitative investigation on the CLEVRER dataset. Given an input image and the \"driving\" motion of one or two objects, our model predicts the future interactive dynamics of multiple elements in the scene. The driving motion is given in the form of semantic mask sequences. The generated object motions are highlighted with a red-line trajectory. Note that our model can generate videos with force propagation across multiple uncontrolled objects (top) and can generate multiple futures (bottom). Q Zoom in for details." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 153, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 153, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 153, + 378 + ], + "type": "text", + "content": "4.1. Implementation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 382, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 382, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 382, + 296, + 515 + ], + "type": "text", + "content": "We initialize InterDyn with the 14-frame image-to-video weights of SVD [5]. During training, we use the Adam optimizer [45] with a learning rate of " + }, + { + "bbox": [ + 54, + 382, + 296, + 515 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 54, + 382, + 296, + 515 + ], + "type": "text", + "content": ". We use the EDM framework [43] with noise distribution " + }, + { + "bbox": [ + 54, + 382, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\log \\sigma \\sim \\mathcal{N}(0.7, 1.6^2)" + }, + { + "bbox": [ + 54, + 382, + 296, + 515 + ], + "type": "text", + "content": ". We train on two 80GB H100 GPUs, with a per-GPU batch size of 4. Video length and FPS define the temporal resolution of dynamics; to balance short- and long-range events we subsample videos to 7 FPS. To facilitate classifier-free guidance [31], we randomly drop the input image with a probability of " + }, + { + "bbox": [ + 54, + 382, + 296, + 515 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 382, + 296, + 515 + ], + "type": "text", + "content": ". At inference, we apply the Euler scheduler [43] over 50 denoising timesteps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 521, + 115, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 115, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 115, + 533 + ], + "type": "text", + "content": "4.2. Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 539, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 588 + ], + "type": "text", + "content": "We evaluate InterDyn on image quality, spatio-temporal similarity, and motion fidelity. Image quality metrics are computed frame-wise. All metrics are reported excluding the first frame, as it serves as input conditioning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 590, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 651 + ], + "type": "text", + "content": "Image quality. We report the Structural Similarity Index Measure (SSIM) [79], Peak Signal-to-Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS) [102], Fréchet Inception Distance (FID) [30] and unbiased Kernel Inception Distance (KID) [4]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": "Spatio-temporal similarity. To assess the spatio-temporal perceptual similarity between the ground truth and the generated video distributions, we use the Fréchet Video Distance (FVD) and unbiased Kernel Video Distance (KVD) proposed in [74]. We use the implementation of [69]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 473 + ], + "type": "text", + "content": "Motion Fidelity. Through InterDyn, we do not have explicit control over object dynamics, which means that the pixel alignment of an object in the generated and ground truth video is only guaranteed in the starting frame. In this case, comparing generated object motion to the ground truth naively might misrepresent the true quality of object motion over time. Therefore, we adapt the Motion Fidelity (MF) metric proposed by Yatim et al. [93], which measures the similarity between point-tracking trajectories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 474, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 546 + ], + "type": "text", + "content": "To compute the metric for any video, we obtain a mask of the object in the starting frame, sample 100 points on the object, and track these throughout both the ground truth and generated video using CoTracker3 [41]. Given the resulting two sets of tracklets " + }, + { + "bbox": [ + 313, + 474, + 555, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{\\tau_1,\\dots ,\\tau_n\\}" + }, + { + "bbox": [ + 313, + 474, + 555, + 546 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 555, + 546 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{T}} = \\{\\tilde{\\tau}_1,\\dots ,\\tilde{\\tau}_m\\}" + }, + { + "bbox": [ + 313, + 474, + 555, + 546 + ], + "type": "text", + "content": " the motion fidelity metric is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 329, + 555, + 555, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 555, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 329, + 555, + 555, + 587 + ], + "type": "interline_equation", + "content": "\\frac {1}{m} \\sum_ {\\tilde {\\tau} \\in \\tilde {\\mathcal {T}}} \\max _ {\\tau \\in \\mathcal {T}} \\mathbf {c o r r} (\\tau , \\tilde {\\tau}) + \\frac {1}{n} \\sum_ {\\tau \\in \\mathcal {T}} \\max _ {\\tilde {\\tau} \\in \\tilde {\\mathcal {T}}} \\mathbf {c o r r} (\\tau , \\tilde {\\tau}), \\tag {1}", + "image_path": "1923ec4f3ae1340f2cefdc25105633b2e35dd5e788a17a63acdb0015a1080212.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 597, + 550, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 550, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 550, + 611 + ], + "type": "text", + "content": "with the correlation between two tracklets " + }, + { + "bbox": [ + 313, + 597, + 550, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{corr}(\\tau, \\tilde{\\tau})" + }, + { + "bbox": [ + 313, + 597, + 550, + 611 + ], + "type": "text", + "content": " [50]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 620, + 555, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 620, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 318, + 620, + 555, + 654 + ], + "type": "interline_equation", + "content": "\\mathbf {c o r r} (\\tau , \\tilde {\\tau}) \\frac {1}{F} \\sum_ {k = 1} ^ {F} \\frac {v _ {k} ^ {x} a \\cdot \\tilde {v} _ {k} ^ {x} + v _ {k} ^ {y} \\cdot \\tilde {v} _ {k} ^ {y}}{\\sqrt {(v _ {k} ^ {x}) ^ {2} + (v _ {k} ^ {y}) ^ {2}} \\cdot \\sqrt {(\\tilde {v} _ {k} ^ {x}) ^ {2} + (\\tilde {v} _ {k} ^ {y}) ^ {2}}}, \\tag {2}", + "image_path": "2cd6f09ee21dfc78c3583621231803cd4461d7093946365fdcac107aa531a982.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "(v_k^x, v_k^y)" + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "(\\tilde{v}_k^x, \\tilde{v}_k^y)" + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "text", + "content": " frame displacement of tracklets " + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\tau, \\tilde{\\tau}" + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "text", + "content": " respectively. If there are less than 100 points to query on the object due to it being too small, we do not consider the video for the motion fidelity metric." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12471" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "content": "4.3. Probing Dynamics with Object Collision Events" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 295, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 295, + 173 + ], + "type": "text", + "content": "Here, we fine-tune InterDyn on an object collision dataset to probe its ability to generate realistic object interactions. Qualitatively, we review whether InterDyn can produce plausible object motion for uncontrolled objects, given the motion of objects entering the scene. In addition, we examine whether InterDyn can generate counterfactual videos for the same input image, but different control signals." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 176, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 176, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 296, + 308 + ], + "type": "text", + "content": "Dataset. We use CLEVRER [97], which provides 20,000 videos of colliding objects with annotated segmentation masks and metadata on collision events. We construct a control signal for objects entering the scene and aim to use InterDyn to generate the motion of the objects that are already present, upon collision. Stationary objects do not receive any form of control signal. Colored masks help the model distinguish unique objects. The frames are cropped and scaled to " + }, + { + "bbox": [ + 55, + 176, + 296, + 308 + ], + "type": "inline_equation", + "content": "320 \\times 448" + }, + { + "bbox": [ + 55, + 176, + 296, + 308 + ], + "type": "text", + "content": ", and we only sample input frames before collisions between objects in the scene, to maximize InterDyn's exposure to interactive dynamics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 443 + ], + "type": "text", + "content": "Force propagation. InterDyn can generate force propagation dynamics between a controlled object and an uncontrolled object, as well as amongst uncontrolled objects, as illustrated in Fig. 4a. Here, the red cylinder at the top left is the driving force. It collides with the uncontrolled blue cylinder, which then collides with the uncontrolled purple cylinder, in turn striking the uncontrolled purple cube on the far right. Point-tracking trajectories display how collisions alter each object's position. This suggests that InterDyn possesses an implicit understanding of physical interactions, enabling it to generate plausible dynamics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 446, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 295, + 579 + ], + "type": "text", + "content": "Counterfactual dynamics. By altering the control signal, InterDyn can simulate counterfactual scenarios for the same input image, as shown in Fig. 4b. In \"Generation 1\", the gray cylinder (controlled) collides with the stationary red sphere (uncontrolled), causing it to move; it is later struck by the brown cylinder (controlled), altering its path once again. In \"Generation 2\", removing the brown cylinder lets the red sphere continue along its original trajectory, consistent with expectations. Crucially, there is no control signal for the red sphere throughout the sequence; its motion and trajectory are entirely generated by InterDyn." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 631 + ], + "type": "text", + "content": "Probing InterDyn on CLEVRER highlights its ability to generate interactive dynamics for objects within a simple synthetic environment. We provide additional results in video format on our webpage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 636, + 264, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 264, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 264, + 649 + ], + "type": "text", + "content": "4.4. Generating Human-Object Interactions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "content": "For this experiment, we fine-tune InterDyn on a human-action dataset, focused on hand-object interaction. We encode human movement as a sequence of pixel-aligned binary hand masks and task InterDyn to generate a video with hand movements and corresponding object dynamics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "Dataset. We use Something-Something-v2 (SSV2), which provides 220,847 videos of humans performing basic actions with everyday objects. It contains actions such as \"pushing [something] from left to right\", \"squeezing [something]\" and \"lifting [something] with [something] on it\". This dataset allows us to train InterDyn at a larger scale and compare with our baseline CosHand [70]. We train one version of InterDyn at the same resolution as CosHand, " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": ", and a second version at " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "256 \\times 384" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": ", which aligns better with the dynamic prior of SVD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 193, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 555, + 384 + ], + "type": "text", + "content": "We generate binary hand masks by prompting Segment Anything 2 (SAM2) [65] with hand bounding boxes, provided by the Something-Else dataset [59]. Similar to our baseline CosHand, we exclude the \"pretending\" class. Additionally, we remove all \"[something] is done behind [something]\" classes since the object would be out of view of the camera. We remove videos smaller than the target resolution/length, ensure the hand and objects are continuously visible, and crop videos larger than the target resolution while centering the object. We include videos without obvious motion and state transition, such as \"holding [something]\". Since bounding box annotations are only provided for the train split (79,043 samples for " + }, + { + "bbox": [ + 313, + 193, + 555, + 384 + ], + "type": "inline_equation", + "content": "256 \\times 384" + }, + { + "bbox": [ + 313, + 193, + 555, + 384 + ], + "type": "text", + "content": " and 104,260 for " + }, + { + "bbox": [ + 313, + 193, + 555, + 384 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 193, + 555, + 384 + ], + "type": "text", + "content": ") and validation split (8,667 samples for " + }, + { + "bbox": [ + 313, + 193, + 555, + 384 + ], + "type": "inline_equation", + "content": "256 \\times 384" + }, + { + "bbox": [ + 313, + 193, + 555, + 384 + ], + "type": "text", + "content": " and 11,229 for " + }, + { + "bbox": [ + 313, + 193, + 555, + 384 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 193, + 555, + 384 + ], + "type": "text", + "content": "), we report all evaluation metrics on the validation split." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 399, + 555, + 548 + ], + "blocks": [ + { + "bbox": [ + 318, + 399, + 555, + 548 + ], + "lines": [ + { + "bbox": [ + 318, + 399, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 318, + 399, + 555, + 548 + ], + "type": "image", + "image_path": "496d5842234bf144097f5c51c0980be1d898e59e0df055ff34743b8950b74631.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 555, + 555, + 578 + ], + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 578 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison. A two-state approach such as CosHand [70] struggles with post-interaction object dynamics." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 590, + 553, + 673 + ], + "blocks": [ + { + "bbox": [ + 318, + 590, + 553, + 673 + ], + "lines": [ + { + "bbox": [ + 318, + 590, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 318, + 590, + 553, + 673 + ], + "type": "image", + "image_path": "b57feeb496674aaef718726f580fd586084668415794913ff42ec78d573cb139.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 681, + 555, + 715 + ], + "lines": [ + { + "bbox": [ + 313, + 681, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 681, + 555, + 715 + ], + "type": "text", + "content": "Figure 6. Robustness to noise. SAM2 outputs noisy/coarse masks for frames with considerable motion blur (orange/red). Despite this, InterDyn can generate plausible hand and object dynamics." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12472" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 555, + 668 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 555, + 668 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 555, + 668 + ], + "type": "image", + "image_path": "36c13c41b2208273a87cd005977ac656f1d048b33a94e18c11a2681a793f803e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 677, + 555, + 711 + ], + "lines": [ + { + "bbox": [ + 55, + 677, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 555, + 711 + ], + "type": "text", + "content": "Figure 7. Diverse generation of interactive dynamics. We show multiple challenging examples, such as (from top to bottom): interacting with articulated objects, pouring liquid, letting an object fall, squeezing a highly deformable or \"collapsible\" object, interacting with reflective objects, tilting a ridged object, or stacking objects. Q Zoom in for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12473" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 550, + 155 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 550, + 155 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 550, + 155 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 550, + 155 + ], + "type": "table", + "html": "
MethodSSIM ↑PSNR ↑LPIPS ↓FID ↓KID ↓FVD ↓KVD ↓Motion Fidelity ↑ [93]
Seer [23]0.41810.710.58833.350.030287.4681.31—
DynamiCrafter [89]†———17.480.014204.1131.81—
CosHand-Independent [70]0.61516.870.3132.950.00291.1819.240.432
CosHand-Autoregressive [70]0.53114.920.40812.660.01290.3013.680.570
Ours 256×2560.66418.600.2604.950.00419.271.990.633
Ours 256×3840.68019.040.2525.070.00422.222.090.641
", + "image_path": "ee9e526b149a9ce772020447694030346042c57b4a1787760b1208cd5d52d28f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 555, + 203 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 555, + 203 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on Something-Something-v2. We compare against two language-instructed video generation methods, Seer [23] and DynamiCrafter [89] and two video extensions of our baseline CosHand [70]. We report results for InterDyn at two resolutions: " + }, + { + "bbox": [ + 55, + 167, + 555, + 203 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 167, + 555, + 203 + ], + "type": "text", + "content": " (matching CosHand) and " + }, + { + "bbox": [ + 55, + 167, + 555, + 203 + ], + "type": "inline_equation", + "content": "256 \\times 384" + }, + { + "bbox": [ + 55, + 167, + 555, + 203 + ], + "type": "text", + "content": " (matching SVD's prior). Methods denoted with " + }, + { + "bbox": [ + 55, + 167, + 555, + 203 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 167, + 555, + 203 + ], + "type": "text", + "content": " do not use SSV2 as their training dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 224, + 295, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 298 + ], + "type": "text", + "content": "Baselines. To generate videos using CosHand [70] rather than state transitions, we run CosHand in a frame-by-frame approach (CosHand-Independent) and an auto-regressive approach (CosHand-Autoregressive). In the frame-by-frame variant, each future frame is independently predicted from the initial frame and its corresponding mask:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 304, + 271, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 304, + 271, + 317 + ], + "spans": [ + { + "bbox": [ + 80, + 304, + 271, + 317 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {t + 1} = \\operatorname {C o s H a n d} \\left(x _ {0}, h _ {0}, h _ {t + 1}\\right), \\quad \\forall t \\in [ 0, 1 3 ],", + "image_path": "1d485e9cf21bbfc90a7efa9fb87112f5129124a7b8c064e85d46f3e0d747e7cf.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 323, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 323, + 296, + 361 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 55, + 323, + 296, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 323, + 296, + 361 + ], + "type": "inline_equation", + "content": "h_{t + 1}" + }, + { + "bbox": [ + 55, + 323, + 296, + 361 + ], + "type": "text", + "content": " denote masks, and " + }, + { + "bbox": [ + 55, + 323, + 296, + 361 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 323, + 296, + 361 + ], + "type": "text", + "content": " the initial frame. In the auto-regressive variant, we use CosHand to generate video frames sequentially:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 368, + 271, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 368, + 271, + 381 + ], + "spans": [ + { + "bbox": [ + 80, + 368, + 271, + 381 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {t + 1} = \\operatorname {C o s H a n d} \\left(\\hat {x} _ {t}, h _ {t}, h _ {t + 1}\\right), \\quad \\forall t \\in [ 0, 1 3 ],", + "image_path": "9fd9c3ebeab69c4703628b4589988af788fab4a8ed2462e4a93171f1dd5a0981.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 388, + 294, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 294, + 399 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 388, + 294, + 399 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 55, + 388, + 294, + 399 + ], + "type": "text", + "content": " being the previously generated frame and " + }, + { + "bbox": [ + 55, + 388, + 294, + 399 + ], + "type": "inline_equation", + "content": "\\hat{x}_0 = x_0" + }, + { + "bbox": [ + 55, + 388, + 294, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 400, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 400, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 400, + 295, + 508 + ], + "type": "text", + "content": "We also compare against two language-instructed video generation methods: Seer [23] and DynamiCrafter [89], which we prompt with a first frame and its corresponding SSV2 class label as instruction. Note that DynamiCrafter was trained with a random video frame as conditioning, i.e. not always the first frame. At inference, its generated videos are thus not strictly a continuation of the conditioning, which precludes frame-aligned metrics. We compare InterDyn against these baselines in Tab. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "text", + "content": "Quantitative analysis. The frame-by-frame variant of CosHand achieves high image quality but struggles with temporal coherence and motion fidelity. Qualitatively, we notice that object locations are inconsistent across frames. In contrast, the auto-regressive variant improves object motion fidelity but suffers from lower frame-wise image quality due to error propagation. Both variants fail in scenarios requiring accurate post-interaction dynamics, such as when objects continue moving after being released from direct hand contact, as shown in Fig. 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "text", + "content": "Our method, InterDyn " + }, + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "inline_equation", + "content": "(256 \\times 384)" + }, + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "text", + "content": ", achieves the best overall performance, surpassing our baseline CosHand in spatio-temporal dynamics, motion fidelity, and all but two image quality metrics. We hypothesize that this might be due to two reasons (1) we use a frozen U-Net, while CosHand fine-tunes its model on SSV2, so CosHand might generate frames closer to the SSV2 distribution, and 2)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 224, + 553, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 553, + 272 + ], + "type": "text", + "content": "when SVD was trained, it was initialized with SD weights as spatial layers, and then fine-tuned over multiple stages; this might have degraded its spatial prior, and by extension the quality of produced images compared to CosHand." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 272, + 555, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 555, + 356 + ], + "type": "text", + "content": "Fine-tuned on noisy masks and leveraging its temporal-aware control branch, InterDyn can interpret a noisy control sequence; e.g. when SAM2 produces a coarse and noisy hand mask sequence, InterDyn generates detailed hands including individual fingers, see Fig. 6. Though not always consistent, InterDyn is capable of depicting post-interaction dynamics, such as rolling or sliding objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 358, + 556, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 556, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 556, + 538 + ], + "type": "text", + "content": "Qualitative analysis. We present diverse qualitative results generated by InterDyn in Fig. 7. Row 1 shows how InterDyn generates the articulated motion of an object. Row 2 showcases pouring water into a glass; note how the water level increases over time. Row 3 demonstrates an object being dropped, moving out of frame when falling, and rolling back in frame once hitting the floor, featuring realistic motion blur synthesis. Rows 4 and 5 illustrate how InterDyn handles squeezing interactions—the rubber and the spring are compressed and restored accordingly. Row 6 demonstrates an understanding of physical size and distance to the camera, as the phone moves closer to the viewer. These results highlight the complexity that InterDyn can handle, implying its generalization ability and future potential as an implicit, yet generalized physical simulator and renderer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 568, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 712 + ], + "type": "text", + "content": "We introduced InterDyn, a framework that generates videos of interactive dynamics using large video generation models as implicit physics simulators. By incorporating an interactive control mechanism, InterDyn produces plausible, temporally consistent videos of object interactions—including complex human-object interactions—while generalizing to unseen objects. Our evaluations demonstrate that InterDyn effectively captures continuous motion and subsequent dynamics, outperforming baselines that focus on single future states. This work highlights the potential of using video generative models as physics simulators without explicit reconstruction, opening new avenues for future research." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12474" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 294, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 73, + 294, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 294, + 132 + ], + "type": "text", + "content": "Acknowledgements: We thank Benjamin Pellkofer for IT support, Peter Kulits for discussions, Tomasz Niewiadomski for coding support, and Allan D. Jepson, Artur Grigorev, Soubhik Sanyal, Omid Taheri & Angjoo Kanazawa for feedback. DT is supported by the ERC Starting Grant (project STRIPES, 101165317)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 144 + ], + "type": "text", + "content": "Disclosure: https://files.is.tue.mpg.de/black/CoI_CVPR_2025.txt" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 145, + 239, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 145, + 239, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 145, + 239, + 156 + ], + "type": "text", + "content": "DT has received a research gift fund from Google." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 165, + 115, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 115, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 115, + 177 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 185, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 185, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 185, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 66, + 185, + 295, + 251 + ], + "type": "text", + "content": "[1] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A space-time diffusion model for video generation. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-11, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 251, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 251, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 66, + 251, + 296, + 295 + ], + "type": "text", + "content": "[2] Fabien Baradel, Natalia Neverova, Julien Mille, Greg Mori, and Christian Wolf. CoPhy: Counterfactual learning of physical dynamics. In International Conference on Learning Representations (ICLR), 2020. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 295, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 295, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 66, + 295, + 295, + 338 + ], + "type": "text", + "content": "[3] Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, et al. Interaction networks for learning about objects, relations and physics. In Conference on Neural Information Processing Systems (NeurIPS), 2016. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 339, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 339, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 66, + 339, + 295, + 381 + ], + "type": "text", + "content": "[4] Mikołaj Binkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In International Conference on Learning Representations (ICLR), 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 383, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 383, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 66, + 383, + 296, + 437 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable Video Diffusion: Scaling latent video diffusion models to large datasets. arXiv:2311.15127, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 438, + 295, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 438, + 295, + 492 + ], + "spans": [ + { + "bbox": [ + 66, + 438, + 295, + 492 + ], + "type": "text", + "content": "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 22563-22575, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 493, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 493, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 66, + 493, + 295, + 536 + ], + "type": "text", + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning to follow image editing instructions. In Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, 2023. 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 537, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 537, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 66, + 537, + 295, + 601 + ], + "type": "text", + "content": "[8] Shengqu Cai, Duygu Ceylan, Matheus Gadelha, ChunHao Paul Huang, Tuanfeng Yang Wang, and Gordon Wetzstein. Generative rendering: Controllable 4D-guided video generation with 2D diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 7611-7620, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 66, + 602, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 602, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 66, + 602, + 295, + 646 + ], + "type": "text", + "content": "[9] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2D pose estimation using part affinity fields. In Computer Vision and Pattern Recognition (CVPR), pages 7291–7299, 2017. 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "type": "text", + "content": "[10] Junuk Cha, Jihyeon Kim, Jae Shin Yoon, and Seungryul Baek. Text2HOI: Text-guided 3D motion generation for hand-object interaction. In Computer Vision and Pattern Recognition (CVPR), pages 1577-1585, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 690, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 690, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 690, + 295, + 712 + ], + "type": "text", + "content": "[11] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S. Narang, Karl" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 708 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 116 + ], + "type": "text", + "content": "Van Wyk, Umar Iqbal, Stan Birchfield, et al. DexYCB: A benchmark for capturing hand grasping of objects. In Computer Vision and Pattern Recognition (CVPR), pages 9044-9053, 2021. 14" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 553, + 172 + ], + "type": "text", + "content": "[12] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv:2310.19512, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 174, + 553, + 217 + ], + "type": "text", + "content": "[13] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Ji-ashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable text-to-video generation with diffusion models. arXiv:2305.13840, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 219, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 219, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 320, + 219, + 553, + 272 + ], + "type": "text", + "content": "[14] Enric Corona, Albert Pumarola, Guillem Alenya, Francesc Moreno-Noguer, and Grégory Rogez. GanHand: Predicting human grasp affordances in multi-object scenes. In Computer Vision and Pattern Recognition (CVPR), pages 5031-5041, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 274, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 274, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 321, + 274, + 553, + 317 + ], + "type": "text", + "content": "[15] Christian Diller and Angela Dai. CG-HOI: Contact-guided 3D human-object interaction generation. In Computer Vision and Pattern Recognition (CVPR), pages 19888-19901, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 319, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 319, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 321, + 319, + 553, + 384 + ], + "type": "text", + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 386, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 386, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 321, + 386, + 553, + 441 + ], + "type": "text", + "content": "[17] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In Computer Vision and Pattern Recognition (CVPR), pages 12943-12954, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 441, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 441, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 321, + 441, + 553, + 506 + ], + "type": "text", + "content": "[18] Zicong Fan, Maria Parelli, Maria Eleni Kadoglou, Xu Chen, Muhammed Kocabas, Michael J. Black, and Otmar Hilliges. HOLD: Category-agnostic 3D reconstruction of interacting hands and objects from video. In Computer Vision and Pattern Recognition (CVPR), pages 494–504, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 508, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 508, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 321, + 508, + 553, + 563 + ], + "type": "text", + "content": "[19] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. In International Conference on Robotics and Automation (ICRA), pages 12462-12469, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 564, + 553, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 564, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 321, + 564, + 553, + 629 + ], + "type": "text", + "content": "[20] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu Video: Factorizing text-to-video generation by explicit image conditioning. In European Conference on Computer Vision (ECCV), 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 631, + 553, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 631, + 553, + 708 + ], + "spans": [ + { + "bbox": [ + 321, + 631, + 553, + 708 + ], + "type": "text", + "content": "[21] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The something something video database for learning and evaluating visual common sense. In International Conference on Computer Vision (ICCV), pages 5842-5850, 2017. 2, 4" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12475" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 126 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 126 + ], + "type": "text", + "content": "[22] Oliver Groth, Fabian B. Fuchs, Ingmar Posner, and Andrea Vedaldi. ShapeStacks: Learning vision-based physical intuition for generalised object stacking. In European Conference on Computer Vision (ECCV), pages 702-717, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 127, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 127, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 127, + 296, + 173 + ], + "type": "text", + "content": "[23] Xianfan Gu, Chuan Wen, Weirui Ye, Jiaming Song, and Yang Gao. Seer: Language instructed video prediction with latent diffusion models. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 8, 17" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 173, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 173, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 173, + 296, + 217 + ], + "type": "text", + "content": "[24] Vincent Le Guen and Nicolas Thome. Disentangling physical dynamics from unknown factors for unsupervised video prediction. In Computer Vision and Pattern Recognition (CVPR), pages 11471-11481, 2020. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 218, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 218, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 218, + 296, + 282 + ], + "type": "text", + "content": "[25] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Pengfei Wan, Di Zhang, Yufan Liu, Weiming Hu, Zhengjun Zha, et al. I2V-Adapter: A general image-to-video adapter for diffusion models. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 284, + 296, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 296, + 338 + ], + "type": "text", + "content": "[26] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. *AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning*. arXiv:2307.04725, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 338, + 296, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 296, + 393 + ], + "type": "text", + "content": "[27] Yana Hasson, Gül Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Computer Vision and Pattern Recognition (CVPR), pages 11807-11816, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 394, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 394, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 394, + 296, + 437 + ], + "type": "text", + "content": "[28] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from RGB videos. In International Conference on 3D Vision (3DV), pages 659-668, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 438, + 296, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 438, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 62, + 438, + 296, + 482 + ], + "type": "text", + "content": "[29] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. CameraCtrl: Enabling camera control for text-to-video generation. arXiv:2404.02101, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 483, + 296, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 483, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 62, + 483, + 296, + 537 + ], + "type": "text", + "content": "[30] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Conference on Neural Information Processing Systems (NeurIPS), pages 6629-6640, 2017. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 538, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 538, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 538, + 296, + 559 + ], + "type": "text", + "content": "[31] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 560, + 296, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 560, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 62, + 560, + 296, + 614 + ], + "type": "text", + "content": "[32] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, et al. Image Video: High definition video generation with diffusion models. arXiv:2210.02303, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 616, + 296, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 616, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 62, + 616, + 296, + 659 + ], + "type": "text", + "content": "[33] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In Conference on Neural Information Processing Systems (NeurIPS), pages 8633-8646, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 660, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 660, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 62, + 660, + 296, + 715 + ], + "type": "text", + "content": "[34] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 72, + 555, + 711 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 320, + 72, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 72, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 555, + 127 + ], + "type": "text", + "content": "[35] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate Anyone: Consistent and controllable image-to-video synthesis for character animation. In Computer Vision and Pattern Recognition (CVPR), pages 8153-8163, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 128, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 128, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 320, + 128, + 554, + 173 + ], + "type": "text", + "content": "[36] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. DepthCrafter: Generating consistent long depth sequences for open-world videos. arXiv:2409.02095, 2024. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 174, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 174, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 174, + 554, + 217 + ], + "type": "text", + "content": "[37] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make It Move: Controllable image-to-video generation with text descriptions. In Computer Vision and Pattern Recognition (CVPR), pages 18219-18228, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 219, + 554, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 219, + 554, + 261 + ], + "spans": [ + { + "bbox": [ + 320, + 219, + 554, + 261 + ], + "type": "text", + "content": "[38] Zhihao Hu and Dong Xu. VideoControlNet: A motion-guided video-to-video translation framework by using diffusion model with ControlNet. arXiv:2307.14073, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 262, + 554, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 262, + 554, + 318 + ], + "spans": [ + { + "bbox": [ + 320, + 262, + 554, + 318 + ], + "type": "text", + "content": "[39] Steeven Janny, Fabien Baradel, Natalia Neverova, Madiha Nadri, Greg Mori, and Christian Wolf. Filtered-CoPhy: Unsupervised learning of counterfactual physics in pixel space. In International Conference on Learning Representations (ICLR), 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 319, + 554, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 319, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 320, + 319, + 554, + 363 + ], + "type": "text", + "content": "[40] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In International Conference on Computer Vision (ICCV), pages 11107-11116, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 364, + 554, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 364, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 320, + 364, + 554, + 419 + ], + "type": "text", + "content": "[41] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker3: Simpler and better point tracking by pseudolabelling real videos. In European Conference on Computer Vision (ECCV), pages 18-35, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 420, + 554, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 420, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 320, + 420, + 554, + 474 + ], + "type": "text", + "content": "[42] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. DreamPose: Fashion image-to-video synthesis via stable diffusion. In International Conference on Computer Vision (ICCV), pages 22623-22633, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 475, + 554, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 475, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 320, + 475, + 554, + 520 + ], + "type": "text", + "content": "[43] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Conference on Neural Information Processing Systems (NeurIPS), 35:26565-26577, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 521, + 554, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 521, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 320, + 521, + 554, + 575 + ], + "type": "text", + "content": "[44] Hyeonwoo Kim, Sookwan Han, Patrick Kwon, et al. Beyond the Contact: Discovering comprehensive affordance for 3D objects from pre-trained 2D diffusion models. In European Conference on Computer Vision (ECCV), pages 400-419, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 576, + 554, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 576, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 320, + 576, + 554, + 609 + ], + "type": "text", + "content": "[45] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 610, + 554, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 610, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 320, + 610, + 554, + 654 + ], + "type": "text", + "content": "[46] Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems. In International Conference on Machine Learning (ICML), pages 2693-2702, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 655, + 554, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 655, + 554, + 711 + ], + "spans": [ + { + "bbox": [ + 320, + 655, + 554, + 711 + ], + "type": "text", + "content": "[47] Sumith Kulal, Tim Brooks, Alex Aiken, Jiajun Wu, Jimei Yang, Jingwan Lu, Alexei A. Efros, and Krishna Kumar Singh. Putting people in their place: Affordance-aware human insertion into scenes. In Computer Vision and Pattern Recognition (CVPR), pages 17089-17099, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12476" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 708 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "text", + "content": "[48] Adam Lerer, Sam Gross, and Rob Fergus. Learning physical intuition of block towers by example. In International Conference on Machine Learning (ICML), pages 430-438, 2016. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 296, + 162 + ], + "type": "text", + "content": "[49] Yunzhu Li, Antonio Torralba, Anima Anandkumar, Dieter Fox, and Animesh Garg. Causal discovery in physical systems from videos. Conference on Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 163, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 163, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 296, + 195 + ], + "type": "text", + "content": "[50] Ce Liu, Antonio Torralba, William T. Freeman, Frédo Durand, and Edward H. Adelson. Motion magnification. Transactions on Graphics (TOG), 24(3):519-526, 2005. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 197, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 197, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 62, + 197, + 295, + 251 + ], + "type": "text", + "content": "[51] Ruibo Liu, Jason Wei, Shixiang Shane Gu, Te-Yen Wu, Soroush Vosoughi, Claire Cui, Denny Zhou, and Andrew M Dai. Mind's Eye: Grounded language model reasoning through simulation. In International Conference on Learning Representations (ICLR), 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 252, + 295, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 252, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 252, + 295, + 305 + ], + "type": "text", + "content": "[52] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. arXiv:2303.05499, 2023. 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 295, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 295, + 351 + ], + "type": "text", + "content": "[53] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. PhysGen: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision (ECCV), pages 360-378, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 352, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 295, + 407 + ], + "type": "text", + "content": "[54] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv:2402.17177, 2024. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 407, + 295, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 407, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 62, + 407, + 295, + 461 + ], + "type": "text", + "content": "[55] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T. Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. Transactions on Graphics (TOG), 39(6): 1-14, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 463, + 295, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 463, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 62, + 463, + 295, + 517 + ], + "type": "text", + "content": "[56] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T. Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In Computer Vision and Pattern Recognition (CVPR), pages 4507-4515, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 518, + 295, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 518, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 62, + 518, + 295, + 563 + ], + "type": "text", + "content": "[57] Jiaxin Lu, Hao Kang, Haoxiang Li, Bo Liu, Yiding Yang, Qixing Huang, and Gang Hua. UGG: Unified generative grasping. In European Conference on Computer Vision (ECCV), pages 414-433, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 563, + 295, + 617 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 563, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 62, + 563, + 295, + 617 + ], + "type": "text", + "content": "[58] Wan-Duo Kurt Ma, John P Lewis, and W. Bastiaan Kleijn. TrailBlazer: Trajectory control for diffusion-based video generation. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-11, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 619, + 295, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 619, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 62, + 619, + 295, + 673 + ], + "type": "text", + "content": "[59] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-Else: Compositional action recognition with spatial-temporal interaction networks. In Computer Vision and Pattern Recognition (CVPR), pages 1049–1059, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 674, + 295, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 674, + 295, + 708 + ], + "spans": [ + { + "bbox": [ + 62, + 674, + 295, + 708 + ], + "type": "text", + "content": "[60] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 709 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 338, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 555, + 95 + ], + "type": "text", + "content": "image diffusion models. In AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 96, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 554, + 161 + ], + "type": "text", + "content": "[61] Franziska Mueller, Micah Davis, Florian Bernard, Oleksandr Sotnychenko, Micekal Verschooor, Miguel A. Otaduy, Dan Casas, and Christian Theobalt. Real-time pose and shape reconstruction of two interacting hands with a single depth camera. Transactions on Graphics (TOG), 38(4): 1-13, 2019. 14" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 163, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 163, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 163, + 554, + 217 + ], + "type": "text", + "content": "[62] Boxiao Pan, Zhan Xu, Chun-Hao Paul Huang, Krishna Kumar Singh, Yang Zhou, Leonidas J. Guibas, and Jimei Yang. ActAnywhere: Subject-aware video background generation. In Conference on Neural Information Processing Systems (NeurIPS), 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 219, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 219, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 320, + 219, + 554, + 262 + ], + "type": "text", + "content": "[63] Georgios Paschalidis, Romana Wilschut, Dimitrije Antic, Omid Taheri, and Dimitrios Tzionas. 3D whole-body grasp synthesis with directional controllability. In International Conference on 3D Vision (3DV), 2025. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 264, + 554, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 264, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 320, + 264, + 554, + 329 + ], + "type": "text", + "content": "[64] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 331, + 554, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 331, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 320, + 331, + 554, + 407 + ], + "type": "text", + "content": "[65] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Girshick, Piotr Dólar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 6, 14" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 408, + 554, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 408, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 320, + 408, + 554, + 461 + ], + "type": "text", + "content": "[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 4, 14" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 463, + 554, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 463, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 320, + 463, + 554, + 506 + ], + "type": "text", + "content": "[67] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. Transactions on Graphics (TOG), 36(6), 2022. 14" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 508, + 554, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 508, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 320, + 508, + 554, + 563 + ], + "type": "text", + "content": "[68] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-A-Video: Text-to-video generation without text-video data. In International Conference on Learning Representations (ICLR), 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 564, + 554, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 564, + 554, + 618 + ], + "spans": [ + { + "bbox": [ + 320, + 564, + 554, + 618 + ], + "type": "text", + "content": "[69] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. StyleGAN-V: A continuous video generator with the price, image quality and perks of StyleGAN2. In Computer Vision and Pattern Recognition (CVPR), pages 3626-3636, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 620, + 554, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 620, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 320, + 620, + 554, + 664 + ], + "type": "text", + "content": "[70] Sruthi Sudhakar, Ruoshi Liu, Basile Van Hoorick, Carl Vondrick, and Richard Zemel. Controlling the world by sleight of hand. In European Conference on Computer Vision (ECCV), pages 414-430, 2024. 2, 3, 6, 8, 14, 17" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 665, + 554, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 665, + 554, + 709 + ], + "spans": [ + { + "bbox": [ + 320, + 665, + 554, + 709 + ], + "type": "text", + "content": "[71] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), pages 581-600, 2020. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12477" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 720 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "text", + "content": "[72] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J. Black. GRIP: Generating interaction poses using latent consistency and spatial cues. In International Conference on 3D Vision (3DV), pages 933-943, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 294, + 183 + ], + "type": "text", + "content": "[73] Tze Ho Elden Tse, Kwang In Kim, Ales Leonardis, and Hyung Jin Chang. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In Computer Vision and Pattern Recognition (CVPR), pages 1664-1674, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 184, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 295, + 237 + ], + "type": "text", + "content": "[74] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation. In International Conference on Learning Representations Workshops (ICLRw), 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 238, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 238, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 238, + 294, + 303 + ], + "type": "text", + "content": "[75] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. In European Conference on Computer Vision (ECCV), pages 439-457, 2025. 4, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 304, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 294, + 336 + ], + "type": "text", + "content": "[76] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv:2308.06571, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 338, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 294, + 390 + ], + "type": "text", + "content": "[77] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. In International Conference on Machine Learning (ICML), 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 392, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 446 + ], + "type": "text", + "content": "[78] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional video synthesis with motion controllability. In Conference on Neural Information Processing Systems (NeurIPS), 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 447, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 491 + ], + "type": "text", + "content": "[79] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image Quality Assessment: From error visibility to structural similarity. Transactions on Image Processing (TIP), 13(4):600-612, 2004. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 491, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 294, + 546 + ], + "type": "text", + "content": "[80] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A unified and flexible motion controller for video generation. In Transactions on Graphics (TOG), pages 1-11, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 547, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 294, + 601 + ], + "type": "text", + "content": "[81] Nicholas Watters, Daniel Zoran, Theophane Weber, Peter Battaglia, Razvan Pascanu, and Andrea Tacchetti. Visual Interaction Networks: Learning a physics simulator from video. In Conference on Neural Information Processing Systems (NeurIPS), pages 4539-4547, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 601, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 294, + 655 + ], + "type": "text", + "content": "[82] Jiajun Wu, Ilker Yildirim, Joseph J. Lim, Bill Freeman, and Joshua B. Tenenbaum. Galileo: Perceiving physical object properties by integrating a physics engine with deep learning. In Conference on Neural Information Processing Systems (NeurIPS), pages 127-135, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 656, + 294, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 656, + 294, + 699 + ], + "spans": [ + { + "bbox": [ + 62, + 656, + 294, + 699 + ], + "type": "text", + "content": "[83] Jiajun Wu, Joseph J. Lim, Hongyi Zhang, Joshua B. Tenenbaum, and William T. Freeman. Physics 101: Learning physical object properties from unlabeled videos. In *British Machine Vision Conference (BMVC)*, 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 700, + 294, + 720 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 700, + 294, + 720 + ], + "spans": [ + { + "bbox": [ + 62, + 700, + 294, + 720 + ], + "type": "text", + "content": "[84] Jiajun Wu, Erika Lu, Pushmeet Kohli, Bill Freeman, and Josh Tenenbaum. Learning to see physics via visual de" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 708 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "text", + "content": "animation. In Conference on Neural Information Processing Systems (NeurIPS), pages 153-164, 2017. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 96, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 553, + 162 + ], + "type": "text", + "content": "[85] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-A-Video: One-shot tuning of image diffusion models for text-to-video generation. In International Conference on Computer Vision (ICCV), pages 7623–7633, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 321, + 163, + 553, + 205 + ], + "type": "text", + "content": "[86] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. THOR: Text to human-object interaction diffusion via relation intervention. arXiv:2403.11208, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 207, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 207, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 321, + 207, + 553, + 262 + ], + "type": "text", + "content": "[87] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. DragAnything: Motion control for anything using entity representation. In European Conference on Computer Vision (ECCV), pages 331-348, 2025. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 264, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 264, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 321, + 264, + 553, + 317 + ], + "type": "text", + "content": "[88] Xianghui Xie, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Template free reconstruction of human-object interaction with procedural interaction generation. In Computer Vision and Pattern Recognition (CVPR), pages 10003-10015, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 319, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 319, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 321, + 319, + 553, + 384 + ], + "type": "text", + "content": "[89] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. DynamiCrafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision (ECCV), pages 399-417, 2025. 2, 3, 8, 17" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 386, + 553, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 386, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 321, + 386, + 553, + 439 + ], + "type": "text", + "content": "[90] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3D human-object interactions with physics-informed diffusion. In International Conference on Computer Vision (ICCV), pages 14928-14940, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 441, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 441, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 321, + 441, + 553, + 506 + ], + "type": "text", + "content": "[91] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. MagicAnimate: Temporally consistent human image animation using diffusion model. In Computer Vision and Pattern Recognition (CVPR), pages 1481–1490, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 508, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 508, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 321, + 508, + 553, + 563 + ], + "type": "text", + "content": "[92] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. CogVideoX: Text-to-video diffusion models with an expert transformer. arXiv:2408.06072, 2024.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 564, + 553, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 564, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 321, + 564, + 553, + 608 + ], + "type": "text", + "content": "[93] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In Computer Vision and Pattern Recognition (CVPR), pages 8466-8476, 2024. 5, 8, 14, 17" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 609, + 553, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 609, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 321, + 609, + 553, + 653 + ], + "type": "text", + "content": "[94] Yufei Ye, Poorvi Hebbar, Abhinav Gupta, and Shubham Tulsiani. Diffusion-guided reconstruction of everyday hand-object interaction clips. In International Conference on Computer Vision (ICCV), pages 19717-19728, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 654, + 553, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 654, + 553, + 708 + ], + "spans": [ + { + "bbox": [ + 321, + 654, + 553, + 708 + ], + "type": "text", + "content": "[95] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance Diffusion: Synthesizing hand-object interactions. In Computer Vision and Pattern Recognition (CVPR), pages 22479-22489, 2023. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12478" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 297, + 336 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 116 + ], + "type": "text", + "content": "[96] Yufei Ye, Abhinav Gupta, Kris Kitani, and Shubham Tulsiani. G-HOP: Generative hand-object prior for interaction reconstruction and grasp synthesis. In Computer Vision and Pattern Recognition (CVPR), pages 1911–1920, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 117, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 117, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 117, + 295, + 171 + ], + "type": "text", + "content": "[97] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Ji-jun Wu, Antonio Torralba, and Joshua B. Tenenbaum. CLEVRR: Collision events for video representation and reasoning. In International Conference on Learning Representations (ICLR), 2020. 2, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 172, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 294, + 215 + ], + "type": "text", + "content": "[98] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. DragNUWA: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv:2308.08089, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 216, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 294, + 258 + ], + "type": "text", + "content": "[99] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Computer Vision and Pattern Recognition (CVPR), pages 18456-18466, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 259, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 259, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 57, + 259, + 294, + 303 + ], + "type": "text", + "content": "[100] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 304, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 304, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 57, + 304, + 294, + 336 + ], + "type": "text", + "content": "[101] Mengqi Zhang, Yang Fu, Zheng Ding, Sifei Liu, Zhuowen Tu, and Xiaolong Wang. HOIDiffusion: Generating realistic 3D hand-object interaction data. In Computer Vision" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 330 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "text", + "content": "and Pattern Recognition (CVPR), pages 8521-8531, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[102] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[103] Juntian Zheng, Qingyuan Zheng, Lixing Fang, Yun Liu, and Li Yi. CAMS: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis. In Computer Vision and Pattern Recognition (CVPR), pages 585-594, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[104] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient video generation with latent diffusion models. arXiv:2211.11018, 2022.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[105] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. GEARS: Local geometry-aware hand-object interaction synthesis. In Computer Vision and Pattern Recognition (CVPR), pages 20634-20643, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[106] Tianqiang Zhu, Rina Wu, Xiangbo Lin, and Yi Sun. Toward human-like grasp: Dexterous grasping via semantic representation of object-hand. In International Conference on Computer Vision (ICCV), pages 15741-15751, 2021. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12479" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/c17df58e-4f4a-4403-b950-698942e97f1f_content_list.json b/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/c17df58e-4f4a-4403-b950-698942e97f1f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f3f985844d77c90e6b81e6827ae41268166738fb --- /dev/null +++ b/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/c17df58e-4f4a-4403-b950-698942e97f1f_content_list.json @@ -0,0 +1,1509 @@ +[ + { + "type": "image", + "img_path": "images/a1a2a0a3a33404ae530a2680deabf1baebe50111a29d9931bac3a9f91e9263e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 123, + 145, + 159 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "INTERMIMIC: Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions", + "text_level": 1, + "bbox": [ + 145, + 133, + 888, + 180 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sirui Xu $^{1}$ Hung Yu Ling $^{2}$ Yu-Xiong Wang $^{1\\dagger}$ Liang-Yan Gui $^{1\\dagger}$", + "bbox": [ + 225, + 205, + 772, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Illinois Urbana-Champaign 2 Electronic Arts Equal Advising", + "bbox": [ + 251, + 224, + 746, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://sirui-xu.github.io/InterMimic", + "bbox": [ + 333, + 263, + 663, + 277 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9f6d3db18762132dbe4b9601ef951b5dda1f69372b6152f3831f8dbf55886183.jpg", + "image_caption": [ + "Highly-Dynamic Interaction", + "Multi-Object Interaction", + "Scalable Interaction Skills" + ], + "image_footnote": [], + "bbox": [ + 91, + 304, + 906, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f14c8ec6592615be708d340bdc035cd2408313cb1f97916107a9bc0c28a060d3.jpg", + "image_caption": [ + "Humanoid Robot Interaction", + "MoCap Correction", + "Interaction Prediction", + "Lift the woodchair.", + "Text-to-Interaction", + "Figure 1. InterMimic enables physically simulated humans to perform interactions with dynamic and diverse objects. It supports highly-dynamic, multi-object interactions and scalable skill learning (Top), making it adaptable for versatile downstream applications (Bottom): it can translate whole-body loco-manipulation skills to a humanoid robot [24, 81], perfect interaction MoCap data, and bridge kinematic generation, e.g., predicting future interactions from past (InterDiff [101]) or generating interactions given text prompts (InterDreamer [102])." + ], + "image_footnote": [], + "bbox": [ + 93, + 435, + 901, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 663, + 441, + 901, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 638, + 326, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Achieving realistic simulations of humans interacting with a wide range of objects has long been a fundamental goal. Extending physics-based motion imitation to complex human-object interactions (HOIs) is challenging due to intricate human-object coupling, variability in object geometries, and artifacts in motion capture data, such as inaccurate contacts and limited hand detail. We introduce InterMimic, a framework that enables a single policy to robustly learn from hours of imperfect MoCap data covering diverse full-body interactions with dynamic and varied objects. Our key insight is to employ a curriculum strategy – perfect first, then scale up. We first train subject-specific teacher policies to mimic, retarget, and refine motion capture data. Next, we distill these teachers into a student policy, with the teachers acting as online experts providing direct supervision.", + "bbox": [ + 88, + 675, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as well as high-quality references. Notably, we incorporate RL fine-tuning on the student policy to surpass mere demonstration replication and achieve higher-quality solutions. Our experiments demonstrate that InterMimic produces realistic and diverse interactions across multiple HOI datasets. The learned policy generalizes in a zero-shot manner and seamlessly integrates with kinematic generators, elevating the framework from mere imitation to generative modeling of complex human-object interactions.", + "bbox": [ + 511, + 641, + 908, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 813, + 645, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Animating human-object interactions is a challenging and time-consuming task even for skilled animators. It requires a deep understanding of physics and meticulous attention to detail to create natural and convincing interactions. While", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12266", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motion Capture (MoCap) data provides references, animators often need to correct contact errors caused by sensor limitations and occlusions between humans and objects. However, this process remains unscalable, as refining a single motion demands a delicate balance between preserving the captured data and ensuring its physical plausibility.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Physics-based human motion imitation [38, 63] offers an alternative approach to improving motion fidelity, by training control policies to mimic reference MoCap data within a physics simulator. However, scaling up human-object interaction imitation presents significant challenges: (i) MoCap Imperfection: Contact artifacts are common, causing expected contacts to fluctuate instead of maintaining consistent zero distance, often due to MoCap limitations or missing hand capture [3, 39]. Accurately imitating MoCap kinematics can result in unrealistic dynamics in simulation. Moreover, HOI datasets often include diverse human shapes, requiring motion retargeting to adapt movements across different human models while preserving interaction dynamics. This retargeting process is imperfect and can introduce new contact artifacts or exacerbate existing ones. (ii) Scaling-up: Although large-scale motion imitation has been explored in previous works [50, 79, 92, 106], it remains largely underexplored for whole-body interactions involving dynamic and diverse objects.", + "bbox": [ + 91, + 184, + 486, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim to utilize rich yet imperfect motion capture interaction datasets to train a control policy capable of learning diverse motor skills while enhancing the plausibility of these actions by correcting errors, such as inaccurate hand motions and faulty contacts. Our approach is grounded on the key insight of tackling the challenges of skill perfection and skill integration progressively. We implement a curriculum-based teacher-student distillation framework, where multiple teacher policies focus on imitating and refining small subsets of interactions, and a student policy integrates these skills from the teachers.", + "bbox": [ + 89, + 473, + 482, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead of relying on curated data that covers a limited range of actions [4, 51], we employ multiple teacher policies trained on a diverse set of imperfect interaction data and address two key challenges: retargeting and recovering. First, we unify all training policies to a canonical human model, by embedding HOI retargeting directly into the imitation. This is achieved by reframing the policy learning to optimize both imitation and retargeting objectives. Second, our teacher policies refine interaction motion through learning from it, as accurate contact dynamics enforced by a physics simulator inherently correct inaccuracies in the reference kinematics. To support this, we introduce tailored contact-guided reward and optimize trajectory collection, enabling effective skill imitation despite MoCap errors.", + "bbox": [ + 89, + 642, + 482, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Introducing teacher policies offers several key benefits. By leveraging teacher rollouts, we effectively distill raw MoCap data into refined HOI references with a unified em", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "bodiment and enhanced physical fidelity. These refined references guide the subsequent student policy training, reducing the negative impact of errors in the original MoCap data. A major hurdle in scaling motion imitation is the sample inefficiency of Reinforcement Learning (RL), which can lead to prohibitively long training times. Our teacher-student approach mitigates this through a space-time trade-off: multiple teacher policies are trained in parallel on smaller, more manageable data subsets, and their expertise is then distilled into a single student policy. We begin with demonstration-based distillation to bootstrap PPO [71] updates, reducing reliance on pure trial and error and enabling more effective scaling. As training progresses, the student gradually shifts from heavy demonstration guidance to increased RL updates, ultimately surpassing simple demonstration memorization. This mirrors alignment strategies in Large Language Models (LLMs), where demonstration-based pretraining is refined through RL fine-tuning [56].", + "bbox": [ + 511, + 90, + 903, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows: (i) We introduce InterMimic, which, to the best of our knowledge, is the first framework designed to train physically simulated humans to develop a wide range of whole-body motor skills for interacting with diverse and dynamic objects, extending beyond traditional grasping tasks. (ii) We develop a teacher-student training strategy, where teacher policies provide a unified solution to address the challenges of retargeting and refining in HOI imitation. The student distillation introduces a scalable solution by leveraging a space-time tradeoff. (iii) We demonstrate that our unified framework, InterMimic, as illustrated in Figure 1, effectively handles versatile physics-based interaction animation, recovering motions with realistic and physically plausible details. Notably, by combining kinematic generators with InterMimic, we enable a physics-based agent to achieve tasks such as interaction prediction and text-to-interaction generation.", + "bbox": [ + 511, + 372, + 906, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 669, + 653, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Significant progress has been made in physics-based human interaction animation and control, with advancements in areas such as human-human interactions [45, 90], hand-object interactions [60, 86, 100, 104], human interactions with static scenes [5, 36, 57, 79, 96], and real-world humanoid control for object manipulation [2, 11, 16, 19, 20, 28, 41, 43, 72, 107]. Among these areas, we are the first to achieve universal whole-body loco-manipulation simulation with diverse dynamic objects, beyond pick-and-place and grasping actions – a novel achievement in animation and an idealized reference for real-world humanoid control. Below, we elaborate on recent studies on whole-body interaction animation, particularly involving dynamic objects.", + "bbox": [ + 511, + 704, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12267", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Kinematic Interaction Animation", + "text_level": 1, + "bbox": [ + 89, + 90, + 382, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generating human interactions has been a long-standing topic in animation and computer graphics [15, 37]. Significant advances in character animation have emerged with the advent of deep learning, e.g., including phase-function-based methods [22] that enable object interactions like carrying a box [74] or playing basketball [75]. This is extended to more diverse but static objects approaching [35, 77, 94, 111]. Subsequent efforts [14, 26, 27, 40, 42, 49, 66, 95] integrate object motion into interactions but remain constrained by assuming that interactions occur primarily through the hands. To address this, recent developments [7, 10, 12, 21, 30, 62, 73, 84, 93, 101, 102] introduce interactions in a fashion of whole-body loco-manipulation that engages multiple body parts in contact. However, these methods often suffer from physical inaccuracies, such as floating contacts and penetrations, while they generate only body motion without considering hand dexterity. In this work, we address physical inaccuracies by refining imperfect kinematic generation through physics simulation, with InterDiff [101] and HOI-Diff [62] serving as motion planning for loco-manipulation that bridges high-level decision-making (e.g., text instruction) with low-level execution.", + "bbox": [ + 89, + 113, + 483, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Physics-based Interaction Animation", + "text_level": 1, + "bbox": [ + 89, + 454, + 410, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Physics-based methods generate motion through motor control policies within a physics simulator, e.g., achieved via deep reinforcement learning to track reference motions [63]. These policies are directly applicable for executing simple interactions, such as punching or striking an object [8, 65, 78, 80]. To achieve more complex interactions, early studies focus on specific scenarios, including notable sports-related [52] examples such as basketball [88], skating [44], soccer [98], tennis [108], table tennis [85], and more proposed in [1]. Research also demonstrates flexibility in more general but simpler box carrying tasks [58, 87, 113]. These advancements are achieved through the integration of multiple control policies [55], the use of adversarial motion priors [13, 18, 64], and imitating diverse kinematic generations [95, 99]. However, these methods train their policies in a non-scalable manner, with each policy handling only specific object types or actions. In pursuit of a single, scalable policy to enable multiple interaction skills, existing methods either rely on fixed interaction patterns, such as approaching and grasping objects [4, 51], or extend single-object skills, e.g., interactions involving a basketball [89]. Additionally, they mostly depend on highly curated data from the GRAB dataset [76], which, despite its high quality, primarily features low-dynamic full-body motion and only small-sized objects. More recent datasets [3, 23, 25, 31, 39, 46-48, 53, 97, 103, 109, 110, 112, 114] offer richer full-body interactions with diverse objects but contain noticeable artifacts that chal", + "bbox": [ + 89, + 478, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f5d9c41cf08c4f0fd5f8d0add65355f1e232146e94c43d345381ec12fa00146f.jpg", + "image_caption": [ + "Figure 2. Our two-stage pipeline: (i) training each teacher policy (MLP) on a small data subset with initialization corrected via Physical State Initialization (PSI), and (ii) freezing the teacher policies to provide refined references for training a student policy (Transformer). The student leverages teacher supervision for effective scaling and is fine-tuned through RL." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 214 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lenge existing motion imitation approaches. We process data from OMOMO [39], BEHAVE [3], HODome [109], and IMHD [114] collected in the InterAct [103] dataset, and the multi-object dataset HIMO [53], highlighting InterMimic's scalability to diverse interactions and its robustness against MoCap artifacts.", + "bbox": [ + 511, + 338, + 906, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 446, + 648, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Formulation. The goal of human-object interaction (HOI) imitation is to learn a policy $\\pi$ that produces simulated human-object motion $\\{\\pmb{q}_t\\}_{t=1}^T$ closely matching a ground-truth reference $\\{\\hat{\\pmb{q}}_t\\}_{t=1}^T$ derived from large-scale MoCap data. Given the geometries of the human and objects, the policy should also compensate for missing or inaccurate details in the dataset. Each pose $\\pmb{q}_t$ has two components: the human pose $\\pmb{q}_t^h$ and the object pose $\\pmb{q}_t^o$ . The human pose is defined as $\\pmb{q}_t^h = \\{\\pmb{\\theta}_t^h, \\pmb{p}_t^h\\}$ , where $\\pmb{\\theta}_t^h \\in \\mathbb{R}^{52 \\times 3}$ represents the joint rotations, and $\\pmb{p}_t^h \\in \\mathbb{R}^{52 \\times 3}$ specifies the joint positions. Specifically, our human model includes 30 hand joints and 22 joints for the rest of the body, with one root joint's rotation and position specified in global coordinates, whereas the rotations and positions of all other joints are defined relative to their respective parent joints' coordinate systems. The object pose $\\pmb{q}_t^o$ is represented as $\\{\\pmb{\\theta}_t^o, \\pmb{p}_t^o\\}$ , where $\\pmb{\\theta}_t^o \\in \\mathbb{R}^3$ denotes the object's orientation and $\\pmb{p}_t^o \\in \\mathbb{R}^3$ the position. All simulation states have corresponding ground-truth values, denoted by the hat symbol. For instance, the reference object rotation is $\\{\\hat{\\pmb{\\theta}}_t^o\\}_{t=1}^T$ . The environmental setup for the simulation is detailed in Sec. B. Overview. We formulate interaction imitation as a Markov Decision Process (MDP), defined by states, actions, simulator-provided transition dynamics, and a reward function. Figure 2 illustrates our two-stage framework: (i) training teacher policies $\\pi^{(T)}$ on small skill subsets, and (ii) distilling these teachers into a scalable student policy $\\pi^{(S)}$ for large-scale skill learning. In Sec. 3.1, we define the states $s_t$", + "bbox": [ + 511, + 474, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12268", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and actions $\\pmb{a}_t$ , applicable to both teacher $\\pi^{(T)}$ and student $\\pi^{(S)}$ policies. In Sec. 3.2, we describe how teacher policies are trained via RL, focusing on reward designs that facilitate retargeting, as well as techniques that mitigate the impact of imperfections in the reference data. Sec. 3.3 details the subsequent distillation of teachers into a scalable student policy, leveraging both RL and learning from demonstration.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Policy Representation", + "text_level": 1, + "bbox": [ + 89, + 215, + 294, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "State. The state $s_t$ , which serves as input to the policy, comprises two components $s_t = \\{s_t^s, s_t^g\\}$ . The first part, $s_t^s$ , contains human proprioception and object observations, expressed as, $\\{\\{\\pmb{\\theta}_t^h, \\pmb{p}_t^h, \\pmb{\\omega}_t^h, \\pmb{v}_t^h\\}, \\{\\pmb{\\theta}_t^o, \\pmb{p}_t^o, \\pmb{\\omega}_t^o, \\pmb{v}_t^o\\}, \\{d_t, c_t\\}\\}$ where $\\{\\pmb{\\theta}_t^h, \\pmb{p}_t^h, \\pmb{\\omega}_t^h, \\pmb{v}_t^h\\}$ represent the rotation, position, angular velocity, and velocity of all joints, respectively, while $\\{\\pmb{\\theta}_t^o, \\pmb{p}_t^o, \\pmb{\\omega}_t^o, \\pmb{v}_t^o\\}$ represent the orientation, location, velocity, and angular velocity of the object, respectively. Motivated by [6], we include object geometry and whole-body haptic sensing from two elements: (i) $d_t$ , vectors from human joints to their nearest points on each object surface; and (ii) $c_t$ , contact markers indicating whether the human's rigid body parts experience applied forces; this serves as simplified tactile or force sensing - an important multimodal input in robot manipulation tasks [9]. The goal state $s_t^g = \\{s_{t,t+k}^g\\}_{k \\in K}$ integrates reference poses from the ground truth motion, where $s_{t,t+k}^g$ is defined as,", + "bbox": [ + 89, + 243, + 483, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\{\\hat {\\boldsymbol {\\theta}} _ {t + k} ^ {h} \\ominus \\boldsymbol {\\theta} _ {t} ^ {h}, \\hat {\\boldsymbol {p}} _ {t + k} ^ {h} - \\boldsymbol {p} _ {t} ^ {h} \\}, \\{\\hat {\\boldsymbol {\\theta}} _ {t + k} ^ {o} \\ominus \\boldsymbol {\\theta} _ {t} ^ {o}, \\hat {\\boldsymbol {p}} _ {t + k} ^ {o} - \\boldsymbol {p} _ {t} ^ {o} \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 522, + 480, + 550 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\hat {\\boldsymbol {d}} _ {t + k} - \\boldsymbol {d} _ {t}, \\hat {\\boldsymbol {c}} _ {t + k} - \\boldsymbol {c} _ {t} \\}, \\{\\hat {\\boldsymbol {\\theta}} _ {t + k} ^ {h}, \\hat {\\boldsymbol {p}} _ {t + k} ^ {h}, \\hat {\\boldsymbol {\\theta}} _ {t + k} ^ {o}, \\hat {\\boldsymbol {p}} _ {t + k} ^ {o} \\} \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 547, + 452, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{\\theta}}_{t + k}^{h},\\hat{\\pmb{p}}_{t + k}^{h},\\hat{\\pmb{d}}_{t + k},\\hat{\\pmb{c}}_{t + k}$ represent the reference information at time step $t + k,\\ominus$ denotes the calculation of rotation difference. All continuous elements of $s_t$ are normalized relative to the current direction of view of the human and the position of the root [63].", + "bbox": [ + 89, + 588, + 482, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given that most MoCap data does not provide reference contact or tactile information, we extract reference contact markers $\\hat{c}_{t + k}$ by inferring dynamic information, beyond relying solely on inaccurate contact distances, specifically by analyzing the object's acceleration to detect human-induced forces. To accommodate the variability in contact distances observed in reference motion, we discretize reference contact markers using varying distance thresholds, as illustrated in Fig. 3(i). The neutral areas serve as buffer zones, avoiding the penalization or enforcement of strict contact. See Sec. C of supplementary for details.", + "bbox": [ + 89, + 671, + 482, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Action. Our human model has 51 actuated joints, defining an action space of $\\mathbf{a}_t \\in \\mathbb{R}^{51 \\times 3}$ . These actions are specified as joint PD targets using the exponential map and are converted into torques applied to each of the human joints.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Imitation as Perfecting", + "text_level": 1, + "bbox": [ + 513, + 90, + 725, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The teacher policy $\\pi^{(\\mathrm{T})}$ is trained via RL to maximize the expected discounted reward by comparing simulated states against potentially erroneous reference states. The training involves: (i) trajectory collection, where we explain how trajectories are initialized and terminated. (ii) policy updating, where collected trajectories and their associated rewards are used to refine the policy. In this section, we elaborate on our reward design and how we optimize trajectory collection to mitigate the impact of reference inaccuracies.", + "bbox": [ + 511, + 113, + 903, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Imitation as Retargeting. We tailor teacher policies to each human subject, while all policies share the same base human model. This serves the retargeting purpose by converting HOIs from different human shapes into a unified base shape. Although motion imitation does not necessarily require a unified human model [50, 91], our approach offers two benefits: (i) It enhances integration with kinematic generation methods, which generally perform better on a single, unified shape [17]. (ii) It demonstrates possible integration with real-world humanoid deployment, which requires retargeting to a consistent physical embodiment. In Figure 1, our method translates MoCap data into motor skills on a Unitree G1 [81] with two Inspire hands [24], all without external retargeting in complex contact-rich scenarios. See Sec. F of the supplementary for additional details.", + "bbox": [ + 511, + 251, + 905, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human [83] or HOI [32] retargeting can be formulated as an optimization problem. Inverse Kinematics (IK) methods, such as those based on quadratic programming [34], demonstrate effectiveness in simplified scenarios but remain underexplored for motions featuring intricate object interactions. RL, by contrast, solves the optimization by maximizing an expected cumulative reward, prompting us to investigate whether RL-driven HOI imitation can be used for HOI retargeting. This extends existing physics-based retargeting approaches, which either omit object interactions [67] or are non-scalable with a single reference [113].", + "bbox": [ + 511, + 477, + 905, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the kinematics should differ due to the embodiment gap, we argue that the dynamics between human and object should remain invariant. Thus, we define rewards to include an embodiment-aware component that loosely aligns the simulated kinematics with the reference interaction, and an embodiment-agnostic reward component that encourages dynamics to be close to the reference.", + "bbox": [ + 511, + 643, + 905, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Embodiment-Aware Reward. When the human and object are far apart, retargeting should prioritize capturing rotational motion, whereas when they are close, accurate position tracking becomes crucial for achieving contact. To reflect this, we define the weights $\\pmb{w}_d$ that are inversely proportional to the distances between joints and the object [113]. The reward thus includes cost functions for joint position $E_p^h = \\langle \\pmb{\\Delta}_p^h, \\pmb{w}_d \\rangle$ , rotation $E_\\theta^h = \\langle \\pmb{\\Delta}_\\theta^h, \\mathbf{1} - \\pmb{w}_d \\rangle$ , and interaction tracking $E_d = \\langle \\pmb{\\Delta}_d, \\pmb{w}_d \\rangle$ , where $\\langle \\cdot, \\cdot \\rangle$ is the inner product, $\\pmb{\\Delta}_p^h[i] = \\| \\hat{\\pmb{p}}^h[i] - \\pmb{p}^h[i] \\|$ , $\\pmb{\\Delta}_\\theta^h[i] =$", + "bbox": [ + 511, + 750, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12269", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/111fbf92abb4b6809bda6221ceaa38a3758d1c1710c4dd9b869544587fabd65b.jpg", + "image_caption": [ + "(i) Reference contact markers", + "Figure 3. (i) Visualization of reference contact markers that accommodate varied contact distances: red to promote contact, green for neutral areas where contact is neither promoted nor penalized, and blue to penalize contact. (ii) Initializing the rollout with reference (RSI) or reference corrected via simulation (PSI)." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 277, + 162 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e2e0019fc9d39d2746e6af3c0045298378d8b7396113fc57417d2b0089f67cd7.jpg", + "image_caption": [ + "(ii) RSI vs. PSI" + ], + "image_footnote": [], + "bbox": [ + 281, + 97, + 351, + 159 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/610b8a83552a18149b470ea8f89665350e244f4dbba2121a0575a3857b8fb085.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 98, + 480, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\| \\hat{\\pmb{\\theta}}^h [i]\\ominus \\pmb{\\theta}^h [i]\\|$ , and $\\pmb{\\Delta}_d[i] = \\| \\hat{\\pmb{d}} [i] - \\pmb {d}[i]\\|$ represent the displacement for the variables defined in Sec. 3.1 with timestep $t$ omitted. The formulation of $\\pmb{w}_{d}$ is provided in Sec. D of supplementary. The reward to be maximized can be formulated as $\\exp (-\\lambda E)$ for each cost function $E$ with a specific hyperparameter $\\lambda$ . Details can be found in Sec. D.", + "bbox": [ + 89, + 282, + 483, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Embodiment-Agnostic Reward. The reward includes components for object tracking and contact tracking. The object tracking cost is defined for position $E_{p}^{o} = \\|\\hat{\\boldsymbol{p}}^{o} - \\boldsymbol{p}^{o}\\|$ and rotation $E_{\\theta}^{o} = \\|\\hat{\\boldsymbol{\\theta}}^{o} - \\boldsymbol{\\theta}^{o}\\|$ , with all values normalized to the human's current position and direction. The contact tracking reward comprises two cost functions: body contact promotion $E_{b}^{c}$ and penalty $E_{p}^{c}$ , both aligning the simulated contact $c$ with reference markers $\\hat{c}$ , as shown in Figure 3. We define three contact levels – promotion, penalty, and neutral – to accommodate potential inaccuracies in reference contact distances. The detailed formulation can be found in Sec. D of the supplementary. Since the physics engine does not differentiate between object, ground, and self-contact, we adopt two strategies: (i) we model foot-ground contact promotion and penalty. This ensures proper foot lifting during cyclic walking and mitigates foot hobbling. (ii) We allow self-collision to avoid self-contact promotion but to promote object interaction. This poses minimal risk as the policy is guided by MoCap reference, which, although lacking perfect contact accuracy, rarely shows self-penetration. For humanoid robots with embodiments that differ from the MoCap reference and require real-world applicability, we disable self-collision, as discussed in Sec. F.", + "bbox": [ + 89, + 380, + 482, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce energy consumption rewards [105] to penalize large human or object jitters, with a proposed contact energy penalizing abrupt contact to promote compliant interactions. See Sec. D of supplementary for more details.", + "bbox": [ + 89, + 732, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hand Interaction Discovery. We use data with average or flattened hand poses [3, 39], which makes accurate object manipulation imitation challenging. To address this, we activate a reference contact marker for any hand part when a fingertip or palm is near an object. Given tasks that do not demand high dexterity, employing a contact-promoting reward with this marker enables policies to develop effective", + "bbox": [ + 88, + 795, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "hand interaction strategies, leveraging the exploratory nature of RL. Additionally, we constrain the range of motion (RoM) of the hands to ensure natural movement. See Sec. D and Sec. B of the supplementary for further details.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Policy Learning. Following [63], the control policy $\\pi$ is trained using PPO [71] with the policy gradient $L(\\psi) = \\mathbb{E}_t[\\min (r_t(\\psi)A_t,\\mathrm{clip}(r_t(\\psi),1 - \\epsilon ,1 + \\epsilon)A_t)$ . $\\psi$ are the parameters of $\\pi$ and $r_t(\\psi)$ quantifies the difference in action likelihoods between updated and old policies. $\\epsilon$ is a small constant, and $A_{t}$ is the advantage estimate given by the generalized advantage estimator $\\mathrm{GAE}(\\lambda)$ [70].", + "bbox": [ + 511, + 152, + 906, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Physical State Initialization. Learning later-phase motion can be essential for policies to achieve high rewards during earlier phases, compared to incrementally learning from the starting phase. Thus, Reference State Initialization (RSI) [63] sets the current pose $\\mathbf{q}_t$ to a reference pose $\\hat{\\mathbf{q}}_t$ at a random timestep $t$ , for initializing the rollout. However, initializing with the imperfect reference can introduce critical artifacts, such as contact floating or incorrect hand motion, leading to unrecoverable failures, e.g., object falling, as depicted in Figure 3(ii). These issues render many initializations ineffective, limiting training on certain interaction phases since successful rollouts may not reach them before the maximum length. The problem is exacerbated by the use of prioritized sampling [79, 91], which favors high-failure-rate initializations.", + "bbox": [ + 511, + 261, + 906, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the need for higher-quality reference initialization, we propose Physical State Initialization (PSI). As illustrated in Figure 2, PSI begins by creating an initialization buffer that stores reference states from MoCap and simulation states from prior rollouts. For each new rollout, an initial state is randomly selected from this buffer, which increases the likelihood of starting from advantageous positions. Once a rollout is completed, trajectories are evaluated based on their expected discounted rewards; those above a certain threshold are added to the buffer using a first-in-first-out (FIFO) strategy, while older or lower-quality trajectories are discarded. This selective reintroduction of high-value states for initialization helps maintain stable policy updates. We apply PSI in a sparse manner to ensure training efficiency. As shown in Figure 3(ii), PSI can collect trajectories for policy update that RSI does not effectively utilize. Further details are provided in Sec. E of the supplementary.", + "bbox": [ + 511, + 489, + 906, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Interaction Early Termination. Early Termination (ET) [63] is commonly used in motion imitation, ending an episode when a body part makes unplanned ground contact or when the character deviates significantly from the reference [50], thus stopping the policy from overvaluing invalid transitions. However, additional conditions should be considered for human-object interactions. We propose Interaction Early Termination (IET), which supplements ET with three extra checks: (i) Object points deviate from their references by more than $0.5\\mathrm{m}$ on average. (ii) Weighted av", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12270", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "erage distances between the character's joints and the object surface exceed $0.5\\mathrm{m}$ from the reference. (iii) Any required body-object contact is lost for over 10 consecutive frames. Full conditions are detailed in Sec. E of the supplementary.", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Imitation with Distillation", + "text_level": 1, + "bbox": [ + 89, + 159, + 328, + 174 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Figure 2, after training the teacher policies on data from each subject (Sec. 3.2), we aggregate them to train a student policy $\\pi^{(\\mathrm{S})}$ to master all skills. As outlined in Algorithm 1, the combined teacher policies, denoted by $\\pi^{(\\mathrm{T})}$ for brevity, serves dual roles by providing state-action trajectories $(s^{(\\mathrm{T})}, a^{(\\mathrm{T})})$ : (i) the state $s^{(\\mathrm{T})}$ for reference distillation, and (ii) the action $a^{(\\mathrm{T})}$ for policy distillation.", + "bbox": [ + 89, + 181, + 483, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reference Distillation. Noisy MoCap data can hinder policy learning, especially at larger scales. In contrast, teacher policies trained on smaller-scale data effectively address these issues by correcting contact artifacts, refining hand placements, and recovering missing details (see Figures 1 and 5). To fully leverage teacher policies, we use their rollouts as references for defining the student policy's goal state and reward functions, distinguishing our approach from distillation based on only action output.", + "bbox": [ + 89, + 287, + 483, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Policy Distillation. We apply distillation on action outputs, which we view as crucial for scaling policies to large datasets. In essence, we trade space for time: teacher policies are trained in parallel on smaller data subsets, allowing the student policy to scale through distillation. Following Algorithm 1, we begin with Behavior Cloning (BC) [29] and use RL fine-tuning to go beyond demonstration memorization, an approach common in LLM alignment [56]. We integrate BC into online policy updates with a staged schedule: we start with DAgger [69] and gradually transition to PPO. Throughout, the critic is continuously trained with the reward from Sec. 3.2. This RL fine-tuning phase is crucial as teacher policies may behave differently when performing similar skills, and simple BC can lead to suboptimal \"averaging\" behavior, where RL fine-tuning helps the student converge on optimal solutions.", + "bbox": [ + 89, + 424, + 483, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Architecture", + "text_level": 1, + "bbox": [ + 89, + 674, + 225, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We set the keyframe indices $K$ (Sec. 3.1, Eq. 1) to $\\{1,16\\}$ for the teacher policies and $\\{1,2,4,16\\}$ for the student policy. The broader observation window for the student policy helps it better distinguish different skills with larger-scale data. Teacher policies employ MLPs, common in physics-based animation [63], while the student policy handles higher-dimensional observations, for which MLPs are less effective. Thus, we use a transformer [82] architecture for sequential modeling [79], as shown in Figure 2.", + "bbox": [ + 89, + 696, + 483, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 845, + 223, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate teacher policies on their ability to imitate imperfect HOI references, and assess the entire teacher-", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Distillation with RL Fine-tuning" + ], + "code_body": "1: Input: A composite policy $\\pi^{(\\mathrm{T})}$ integrated from individual teacher policies, student policy parameters $\\psi$ , student value function parameters $\\phi$ , schedule hyperparameter $\\beta$ for DAgger, horizon length $H$ for PPO \n2: for $t = 0,1,2,\\ldots$ do \n3: for $h$ from 1 to $H$ do \n4: Sample a variable $u \\sim \\mathrm{Uniform}(0,1)$ \n5: Collect $s^{(\\mathrm{T})}, a^{(\\mathrm{T})}$ from teacher $\\pi^{(\\mathrm{T})}$ \n6: Obtain the refined reference from $s^{(\\mathrm{T})}$ to define $s^{(\\mathrm{S})}$ and $r(\\cdot)$ , obtain $a^{(\\mathrm{S})}$ from $\\pi_{\\phi}^{(\\mathrm{S})}(a^{(\\mathrm{S})}|s^{(\\mathrm{S})})$ . \n7: if $u \\leq \\frac{t}{\\beta}$ then ▷ Use the teacher \n8: Given $s^{(\\mathrm{S})}$ , execute $a^{(\\mathrm{S})}$ , observe $s^{\\prime (\\mathrm{S})}, r$ \n9: else ▷ Use the student \n10: Given $s^{(\\mathrm{S})}$ , execute $a^{(\\mathrm{T})}$ , observe $s^{\\prime (\\mathrm{S})}, r$ \n11: end if \n12: Store the transition $(s^{(\\mathrm{S})}, s^{\\prime (\\mathrm{S})}, a^{(\\mathrm{S})}, a^{(\\mathrm{T})}, r)$ \n13: end for \n14: Update $\\phi$ with TD( $\\lambda$ ) \n15: Compute PPO objective: $L(\\psi)$ \n16: Compute $J(\\psi) = \\| a^{(\\mathrm{S})} - a^{(\\mathrm{T})}\\|$ \n17: Compute the weight: $w = \\min \\left(\\frac{t}{\\beta}, 1\\right)$ \n18: Update $\\psi$ by gradient: $\\nabla_{\\psi}(wL(\\psi) + (1 - w)J(\\psi))$ \n19: end for", + "bbox": [ + 516, + 109, + 906, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "student framework for scalability to large-scale data and zero-shot generalization across various scenarios. Additional experiments are provided in Sec. G of supplementary.", + "bbox": [ + 511, + 501, + 906, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We use the following datasets from InterAct [103]: OMOMO [39], BEHAV [3], HODome [109], IMHD [114], and HIMO [53]. OMOMO, containing 15 objects and approximately 10 hours of data, is our primary dataset for evaluating the full teacher-student distillation framework for its scale. We train 17 teacher policies, one per subject, with subject 14 reserved as the test set and the remaining data used for training the student policy. A small portion of data is discarded after teacher imitation due to severe MoCap errors that could not be corrected (see Sec. F and Sec. H of the supplementary). Additional datasets are used to evaluate teacher policies in various MoCap scenarios with different error levels and interaction types. We focus on highly dynamic motions (Figure 1) and interactions involving multiple body parts (Figure 4), while excluding scenarios such as carrying a bag with a strap, since the simulator [54] used lacks full support for soft bodies.", + "bbox": [ + 511, + 549, + 908, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We use the following metrics: (i) Success Rate is defined as the proportion of references that the policy successfully imitates at least once, averaged across all references, while (ii) Duration is the time (in seconds) that the imitation is maintained without triggering the interaction early termination conditions introduced in Sec. 3.2. (iii)", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12271", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df088beb273b5d0cccb0de723203bf713b2d524bdd406705ed2e3150b9f9258b.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison between PhysHOI [88] (top), the reference motion (middle) from the BEHAVE [3] dataset, and the interaction refined by our teacher trained on it (bottom). InterMimic faithfully imitates the interactions involving multiple body parts while correcting errors in the original reference." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 486, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human Tracking Error $(E_h)$ measures the per-joint position error (cm) between the simulated and reference human (excluding hand joints for BEHAV [3] and OMOMO [39] due to inaccuracy). (iv) Object Tracking Error $(E_o)$ measures the per-point position error (cm) between the simulated and reference object. Both errors are averaged over the duration of the imitation in the best-performing trial.", + "bbox": [ + 89, + 412, + 483, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. To facilitate fair comparisons, we downgrade our method for teacher policy evaluation to imitate either a single MoCap clip (Figure 4) or multiple clips with a single object (Table 1), enabling direct comparison with PhysHOI [88] and SkillMimic [89] (Sec. 4.1 and 4.2). Due to the lack of established baselines for large-scale HOI imitation, we adapt the following variants for comparison with our student policy (Sec. 4.3): (i) PPO [71] trains an imitation policy from scratch, following [63]. We experiment with both versions, with and without reference distillation; (ii) DAgger [69] distills the student without RL fine-tuning, a process we refer to as policy distillation.", + "bbox": [ + 89, + 520, + 483, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. The control policies operate at $30\\mathrm{Hz}$ and are trained using the Isaac Gym simulator [54]. Teacher policies are implemented as MLPs with hidden layers of sizes 1024, 1024, and 512. The student policy is implemented as a three-layer Transformer encoder with 4 heads, a hidden size of 256, and a feed-forward layer of 512. The critics are also modeled as MLPs with the same architecture as the teacher policies. To integrate the student policy with kinematic generators, including text-to-HOI [62] and future interaction prediction [101], we train these models using reference data distilled by the teacher policies from the OMOMO [39] dataset, following the same train-test split as the student policy training. For the text-to", + "bbox": [ + 89, + 704, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a63d96253bb909ed0735badf9ee59d3011c4179d1bdbe3141f7f0d8f9d0fb973.jpg", + "image_caption": [ + "Figure 5. We recover plausible object rotations (bottom) that are challenging for motion capture due to the equivariant geometries of objects, which result in the object sliding on the ground (top)." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 905, + 258 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/399b547c96adc08b076e973f700ae6b6067520abe5605e3614243ebe8bc1582a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTime↑Eh↓Eo↓
SkillMimic [89]12.27.213.4
InterMimic (Ours) w/o IET40.36.79.9
InterMimic (Ours) w/o PSI36.16.610.2
InterMimic (Ours)42.66.49.2
", + "bbox": [ + 555, + 323, + 867, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison between the teacher policy from InterMimic and SkillMimic [89] to imitate data extracted from the BEHAVE [3] dataset involving a single subject interacting with yogamat. We ablate our proposed approach by removing interaction early termination and physical state initialization.", + "bbox": [ + 511, + 414, + 906, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HOI model, we train it to generate 10 seconds of motion and use 24 generated samples for evaluation, while for future interaction prediction, the model generates 25 future frames given 10 past frames and we use 60 generated samples for evaluation. See Sec. F of the supplementary.", + "bbox": [ + 511, + 503, + 905, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 588, + 735, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 shows that the baseline struggles with MoCap imperfections, e.g., incorrect hand positioning, and thus results in clearly shorter tracking durations. In contrast, our method maintains reference tracking for longer durations and produces interactions that closely match the reference. Table 2 shows that our method consistently outperforms baselines in both training data imitation and out-of-distribution generalization, including interactions from the test set and from kinematic generations. We discuss the effectiveness of specific design choices in Sec. 4.3.", + "bbox": [ + 511, + 611, + 906, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 771, + 723, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 shows a representative sequence from the experiment in Table 1, illustrating how our teacher policy corrects interactions that PhysHOI [88] fails to track robustly – our method effectively withstands and corrects incorrect hand positioning and floating contacts in the reference. Beyond obvious errors, our method also rectifies the rotation of symmetric objects that MoCap inaccurately depicts as slid", + "bbox": [ + 511, + 795, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12272", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a11b1340716ba76f6944bb90c30be16d59941a9ddaa7802e6b1cf251c2e52ce0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PPOReference DistillationPolicy DistillationArchitectureOMOMO-Train [39]OMOMO [39]-TestOMOMO [39]-Test (w × 10)HOI-Diff [62]InterDiff [101]
Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓
✓××MLP23.9101.67.215.69.685.37.516.23.971.27.517.90.00.0--6.711.76.216.4
×✓✓54.5139.97.111.054.3140.27.111.215.591.79.313.74.284.810.19.765.027.47.513.4
✓✓×71.7152.88.912.791.6173.78.513.245.8127.69.114.98.3130.910.113.873.328.96.914.4
✓✓✓90.7168.05.59.795.5173.95.411.962.6140.96.614.512.5121.48.612.175.029.16.213.5
✓✓✓Transformer88.8167.06.010.298.1176.55.911.356.8134.76.613.212.5119.08.512.676.729.36.413.3
", + "bbox": [ + 93, + 88, + 903, + 165 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Quantitative evaluation of large-scale interaction imitation using OMOMO [39], kinematic generations from HOI-Diff [62], and InterDiff [101]. Additionally, we evaluate on test set when objects with weights ten times greater than those used during training.", + "bbox": [ + 89, + 176, + 906, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb3cea12521aa060a57d58b0ef9480c512c9960923375cdf38950155543c5029.jpg", + "image_caption": [ + "Figure 6. Zero-shot integration with a text-to-HOI model HOI-Diff [62] (Top), using 'Kick the large box' as the prompt, and an interaction prediction model InterDiff [101] (Bottom), where gray meshes are past states and colored illustrate future generations." + ], + "image_footnote": [], + "bbox": [ + 93, + 227, + 480, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing along the ground, shown in Figure 5. Figure 6 presents additional examples that complement Figure 1, demonstrating how our approach integrates with kinematic generators for future interaction prediction and text-to-interaction synthesis. This zero-shot generalization extends to novel objects unseen during training (Figure 7), highlighting the effectiveness of our object geometry and contact-encoded representation, as well as the large-scale training.", + "bbox": [ + 88, + 474, + 482, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 604, + 243, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of PSI and IET. We conduct an ablation study, as demonstrated in Table 1, comparing the full approach to \"Ours w/o PSI\". The results validate that Physical State Initialization (PSI) is effective by mitigating inaccuracies in the motion capture data. We also observe reduced effectiveness without our interaction early termination, as training often spends rollouts on irrelevant periods.", + "bbox": [ + 89, + 628, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Reference Distillation. Compared to directly scaling imitation from MoCap with potential imperfections (line 1 in Table 2), using references refined by the teacher policy (line 3) achieves consistently better performance on all metrics. The improvement is even more pronounced on the test set, where, without reference distillation, the policy struggles with unseen shapes, while retargeting by reference distillation eliminates the difficulty.", + "bbox": [ + 89, + 734, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Joint PPO and DAgger Updates. As shown in Table 2, training a policy from scratch (line 3) or relying solely on policy distillation (DAgger, line 2) fails to", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/be2f54312dd20b1d975985e217970ac222668c82f3c43e9a52c986a5ecb9d5ef.jpg", + "image_caption": [ + "Figure 7. Zero-shot generalization of our student policy on novel objects from BEHAV [3] and HODome [109]." + ], + "image_footnote": [], + "bbox": [ + 514, + 227, + 903, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "achieve optimal performance. While supervised skill learning lays the groundwork, additional PPO fine-tuning is crucial for resolving conflicts among teacher policies. This is important because our subject-based clustering may not effectively distinguish between different interaction patterns, and ambiguity arises when multiple teachers produce different actions for similar motions.", + "bbox": [ + 511, + 363, + 903, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Transformer for Policy Learning. From Table 2, we see that using a Transformer policy (line 5) outperforms MLP-based approaches, particularly on the test set and out-of-distribution cases generated by the kinematic model. We attribute this to the Transformer's inductive bias in sequential modeling and its capacity to incorporate longer-term observations, enabling it to handle complex spatio-temporal dependencies more effectively.", + "bbox": [ + 511, + 469, + 905, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 603, + 633, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce a framework for synthesizing realistic human-object interactions that are both physically grounded and generalizable. Unlike previous methods, our approach leverages a rich repository of imperfect MoCap data to facilitate the learning of various interaction skills across a wide variety of objects. To address inaccuracies in the MoCap data, we propose contact-guided rewards and optimize trajectory collection, enabling teacher policies to recover missing physical details in the original data. These teacher policies are used to train student policies within a distillation framework that combines policy distillation and reference distillation, thus enabling efficient skill scaling. Our approach shows zero-shot generalizability, which effectively bridges the gap between imitation and generative capabilities by integrating with kinematic generation. We believe that this framework can be adapted for whole-body loco-manipulation for real-world robots, enabling them to handle objects with human-like dexterity and nuance.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12273", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We thank Wei Yang, Yu-Wei Chao, Arsalan Mousavian, Ankur Handa, Samuel Schulter, Morteza Ziyadi, and Xiaohan Fei for valuable discussions. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Amazon-Illinois Center on AI for Interactive Conversational Experiences, the Toyota Research Institute, the IBM-Illinois Discovery Accelerator Institute, and Snap Inc. This work used computational resources, including the NCSA Delta and DeltaAI and the PTI Jetstream2 supercomputers through allocations CIS230012, CIS230013, and CIS240311 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, as well as the TACC Frontera supercomputer and Amazon Web Services (AWS) through the National Artificial Intelligence Research Resource (NAIRR) Pilot.", + "bbox": [ + 89, + 90, + 485, + 333 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 347, + 187, + 363 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinseok Bae, Jungdam Won, Donggeun Lim, Cheol-Hui Min, and Young Min Kim. Pmp: Learning to physically interact with environments using part-wise motion priors. In SIGGRAPH, 2023. 3", + "[2] Qingwei Ben, Feiyu Jia, Jia Zeng, Junting Dong, Dahua Lin, and Jiangmiao Pang. HOMIE: Humanoid locomap-. manipulation with isomorphic exoskeleton cockpit. arXiv preprint arXiv:2502.13013, 2025. 2", + "[3] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAV: Dataset and method for tracking human object interactions. In CVPR, 2022. 2, 3, 5, 6, 7, 8", + "[4] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. In 3DV, 2024. 2, 3", + "[5] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In AAAI, 2021. 2", + "[6] Sammy Christen, Muhammed Kocabas, Emre Aksan, Jemin Hwangbo, Jie Song, and Otmar Hilliges. D-grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In CVPR, 2022. 4", + "[7] Enric Corona, Albert Pumarola, Guillem Alenya, and Francesc Moreno-Noguer. Context-aware human motion prediction. In CVPR, 2020. 3", + "[8] Jieming Cui, Tengyu Liu, Nian Liu, Yaodong Yang, Yixin Zhu, and Siyuan Huang. AnySkill: Learning open-vocabulary physical skill for interactive agents. In CVPR, 2024. 3", + "[9] Ravinder S Dahiya, Giorgio Metta, Maurizio Valle, and Giulio Sandini. Tactile sensing—from humans to humanoids. IEEE transactions on robotics, 26(1):1-20, 2009. 4", + "[10] Divyanshu Daiya, Damon Conover, and Aniket Bera. COLLAGE: Collaborative human-agent interaction generation using hierarchical latent diffusion and language models. arXiv preprint arXiv:2409.20502, 2024. 3" + ], + "bbox": [ + 101, + 372, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Jeremy Dao, Helei Duan, and Alan Fern. Sim-to-real learning for humanoid box loco-manipulation. In ICRA, 2024. 2", + "[12] Christian Diller and Angela Dai. CG-HOI: Contact-guided 3d human-object interaction generation. In CVPR, 2024. 3", + "[13] Jiawei Gao, Ziqin Wang, Zeqi Xiao, Jingbo Wang, Tai Wang, Jinkun Cao, Xiaolin Hu, Si Liu, Jifeng Dai, and Jiangmiao Pang. CooHOI: Learning cooperative human-object interaction with manipulated object dynamics. In NeurIPS, 2024. 3", + "[14] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. IMoS: Intent-driven full-body motion synthesis for human-object interactions. In Computer Graphics Forum, 2023. 3", + "[15] Michael Gleicher. Motion editing with spacetime constraints. In Proceedings of the 1997 symposium on Interactive 3D graphics, 1997. 3", + "[16] Zhaoyuan Gu, Junheng Li, Wenlan Shen, Wenhao Yu, Zhaoming Xie, Stephen McCrory, Xianyi Cheng, Abdulaziz Shamsah, Robert Griffin, C Karen Liu, et al. Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning. arXiv preprint arXiv:2501.02116, 2025. 2", + "[17] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 4", + "[18] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 3", + "[19] Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. OmniH2O: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. arXiv preprint arXiv:2406.08858, 2024. 2", + "[20] Tairan He, Zhengyi Luo, Wenli Xiao, Chong Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Learning human-to-humanoid real-time whole-body teleoperation. In IROS, 2024. 2", + "[21] Wenkun He, Yun Liu, Ruitao Liu, and Li Yi. SyncDiff: Synchronized motion diffusion for multi-body human-object interaction synthesis. arXiv preprint arXiv:2412.20104, 2024.3", + "[22] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. 3", + "[23] Yinghao Huang, Omid Taheri, Michael J. Black, and Dimitrios Tzionas. InterCap: Joint markerless 3D tracking of humans and objects in interaction. In GCRP, 2022. 3", + "[24] Inspire-robots. Smaller and higher-precision motion control experts. https://inspire-robots/store/.1,4,5", + "[25] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. CHAIRS: Towards full-body articulated human-object interaction. In ICCV, 2023. 3", + "[26] Nan Jiang, Zimo He, Zi Wang, Hongjie Li, Yixin Chen, Siyuan Huang, and Yixin Zhu. Autonomous character" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12274", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "scene interaction synthesis from text instruction. In SIGGRAPH Asia, 2024. 3", + "[27] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In CVPR, 2024. 3", + "[28] Zhenyu Jiang, Yuqi Xie, Kevin Lin, Zhenjia Xu, Weikang Wan, Ajay Mandlekar, Linxi Fan, and Yuke Zhu. DexMimicGen: Automated data generation for bimanual dexterous manipulation via imitation learning. arXiv preprint arXiv:2410.24185, 2024. 2", + "[29] Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. SuperPADL: Scaling language-directed physics-based control with progressive supervised distillation. In SIGGRAPH, 2024. 6", + "[30] Hyeonwoo Kim, Sangwon Beak, and Hanbyul Joo. DAViD: Modeling dynamic affordance of 3d objects using pre-trained video diffusion models. arXiv preprint arXiv:2501.08333, 2025. 3", + "[31] Jeonghwan Kim, Jisoo Kim, Jeonghyeon Na, and Hanbyul Joo. ParaHome: Parameterizing everyday home activities towards 3d generative modeling of human-object interactions. arXiv preprint arXiv:2401.10232, 2024. 3", + "[32] Yeonjoon Kim, Hangil Park, Seungbae Bang, and Sung-Hee Lee. Retargeting human-object interaction to virtual avatars. IEEE transactions on visualization and computer graphics, 22(11):2405-2412, 2016. 4", + "[33] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 5", + "[34] Dieter Kraft. Algorithm 733: Tomp-fortran modules for optimal control calculations. ACM Transactions on Mathematical Software (TOMS), 20(3):262-281, 1994. 4", + "[35] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. NIFTY: Neural object interaction fields for guided human motion synthesis. In CVPR, 2024. 3", + "[36] Jiye Lee and Hanbyul Joo. Locomotion-Action-Manipulation: Synthesizing human-scene interactions in complex 3d environments. In ICCV, 2023. 2", + "[37] Kang Hoon Lee, Myung Geol Choi, and Jehee Lee. Motion patches: building blocks for virtual environments annotated with motion data. In SIGGRAPH. 2006. 3", + "[38] Yoonsang Lee, Sungeun Kim, and Jehee Lee. Data-driven biped control. In SIGGRAPH. 2010. 2", + "[39] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2, 3, 5, 6, 7, 8", + "[40] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. In ECCV, 2024. 3", + "[41] Jinhan Li, Yifeng Zhu, Yuqi Xie, Zhenyu Jiang, Mingyo Seo, Georgios Pavlakos, and Yuke Zhu. OKAMI: Teaching humanoid robots manipulation skills through single video imitation. In CoRL, 2024. 2", + "[42] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. In WACV, 2024. 3" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Fukang Liu, Zhaoyuan Gu, Yilin Cai, Ziyi Zhou, Shijie Zhao, Hyunyoung Jung, Sehoon Ha, Yue Chen, Danfei Xu, and Ye Zhao. Opt2skill: Imitating dynamically-feasible whole-body trajectories for versatile humanoid loco-manipulation. arXiv preprint arXiv:2409.20514, 2024. 2", + "[44] Libin Liu and Jessica Hodgins. Learning to schedule control fragments for physics-based characters using deep q-learning. ACM Transactions on Graphics (TOG), 36(3): 1-14, 2017. 3", + "[45] Yunze Liu, Changxi Chen, Chenjing Ding, and Li Yi. Phys-Reaction: Physically plausible real-time humanoid reaction synthesis via forward dynamics guided 4d imitation. In ACMMM, 2024. 2", + "[46] Yun Liu, Bowen Yang, Licheng Zhong, He Wang, and Li Yi. Mimicking-bench: A benchmark for generalizable humanoid-scene interaction learning via human mimicking. arXiv preprint arXiv:2412.17730, 2024. 3", + "[47] Yun Liu, Haolin Yang, Xu Si, Ling Liu, Zipeng Li, Yuxiang Zhang, Yebin Liu, and Li Yi. Taco: Benchmarking generalizable bimanual tool-action-object understanding. In CVPR, 2024.", + "[48] Yun Liu, Chengwen Zhang, Ruofan Xing, Bingda Tang, Bowen Yang, and Li Yi. Core4d: A 4d human-object-human interaction dataset for collaborative object rearrangement. In CVPR, 2025. 3", + "[49] Jintao Lu, He Zhang, Yuting Ye, Takaaki Shiratori, Sebastian Starke, and Taku Komura. CHOICE: Coordinated human-object interaction in cluttered environments for pick-and-place actions. arXiv preprint arXiv:2412.06702, 2024.3", + "[50] Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In ICCV, 2023. 2, 4, 5, 3", + "[51] Zhengyi Luo, Jinkun Cao, Sammy Christen, Alexander Winkler, Kris Kitani, and Weipeng Xu. Grasping diverse objects with simulated humanoids. In NeurIPS, 2024. 2, 3", + "[52] Zhengyi Luo, Jiashun Wang, Kangni Liu, Haotian Zhang, Chen Tessler, Jingbo Wang, Ye Yuan, Jinkun Cao, Zihui Lin, Fengyi Wang, et al. SMPLOlympics: Sports environments for physically simulated humanoids. arXiv preprint arXiv:2407.00187, 2024.3", + "[53] Xintao Lv, Liang Xu, Yichao Yan, Xin Jin, Congsheng Xu, Shuwen Wu, Yifan Liu, Lincheng Li, Mengxiao Bi, Wenjun Zeng, et al. HIMO: A new benchmark for full-body human interacting with multiple objects. In ECCV, 2024. 3, 6", + "[54] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. In NeurIPS, 2021. 6, 7, 2, 5", + "[55] Josh Merel, Saran Tunyasuvunakool, Arun Ahuja, Yuval Tassa, Leonard Hasenclever, Vu Pham, Tom Erez, Greg Wayne, and Nicolas Heess. Catch & carry: reusable neural controllers for vision-guided whole-body tasks. ACM Transactions on Graphics (TOG), 39(4):39-1, 2020. 3", + "[56] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12275", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2, 6", + "[57] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 3DV, 2024. 2", + "[58] Liang Pan, Zeshi Yang, Zhiyang Dou, Wenjia Wang, Buzhen Huang, Bo Dai, Taku Komura, and Jingbo Wang. TokenHSI: Unified synthesis of physical human-scene interactions through task tokenization. In CVPR, 2025. 3", + "[59] Soohwan Park, Hoseok Ryu, Seyoung Lee, Sunmin Lee, and Jehee Lee. Learning predict-and-simulate policies from unorganized human motion data. ACM Transactions on Graphics (TOG), 38(6):1-11, 2019. 4", + "[60] Sungjae Park, Seungho Lee, Mingi Choi, Jiye Lee, Jeonghwan Kim, Jisoo Kim, and Hanbyul Joo. Learning to transfer human hand skills for robot manipulations. arXiv preprint arXiv:2501.04169, 2025. 2", + "[61] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 2, 5", + "[62] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. HOI-Diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 3, 7, 8, 5", + "[63] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG), 37(4):1-14, 2018. 2, 3, 4, 5, 6, 7", + "[64] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4):1-20, 2021. 3", + "[65] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Transactions On Graphics (TOG), 41(4):1-17, 2022. 3", + "[66] Haziq Razali and Yiannis Demiris. Action-conditioned generation of bimanual object manipulation sequences. In AAAI, 2023. 3", + "[67] Daniele Reda, Jungdam Won, Yuting Ye, Michiel van de Panne, and Alexander Winkler. Physics-based motion retargeting from sparse inputs. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 6(3):1-19, 2023. 4", + "[68] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6), 2017. 2", + "[69] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 6, 7" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation. In ICLR, 2016. 5", + "[71] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2, 5, 7", + "[72] Carmelo Sferrazza, Dun-Ming Huang, Xingyu Lin, Youngwoon Lee, and Pieter Abbeel. Humanoidbench: Simulated humanoid benchmark for whole-body locomotion and manipulation. In RSS, 2024. 2", + "[73] Wenfeng Song, Xinyu Zhang, Shuai Li, Yang Gao, Aimin Hao, Xia Hou, Chenglizhao Chen, Ning Li, and Hong Qin. HOIAnimator: Generating text-prompt human-object animations using novel perceptive diffusion models. In CVPR, 2024. 3", + "[74] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 3", + "[75] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39(4):54-1, 2020. 3", + "[76] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In ECCV, 2020. 3", + "[77] Omid Taheri, Vasileios Choutas, Michael J Black, and Dimitrios Tzionas. GOAL: Generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 3", + "[78] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. Calm: Conditional adversarial latent models for directable virtual characters. In SIGGRAPH, 2023. 3", + "[79] Chen Tessler, Yunrong Guo, Ofir Nabati, Gal Chechik, and Xue Bin Peng. Maskedmimic: Unified physics-based character control through masked motion inpainting. ACM Transactions on Graphics (TOG), 43(6):1-21, 2024. 2, 5, 6", + "[80] Guy Tevet, Sigal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. CLoSD: Closing the loop between simulation and diffusion for multi-task character control. In ICLR, 2025. 3", + "[81] Unittree. Unittree g1 humanoid agent ai avatar. https://www-unitree.com/g1/.1,4,5", + "[82] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6", + "[83] Ruben Villegas, Duygu Ceylan, Aaron Hertzmann, Jimei Yang, and Jun Saito. Contact-aware retargeting of skinned motion. In ICCV, 2021. 4", + "[84] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 2022. 3" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12276", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[85] Jiashun Wang, Jessica Hodgins, and Jungdam Won. Strategy and skill learning for physics-based table tennis animation. In SIGGRAPH, 2024. 3", + "[86] Ruocheng Wang, Pei Xu, Haochen Shi, Elizabeth Schumann, and C Karen Liu. Fürelise: Capturing and physically synthesizing hand motion of piano performance. In SIGGRAPH Asia, 2024. 2", + "[87] Wenjia Wang, Liang Pan, Zhiyang Dou, Zhouyingcheng Liao, Yuke Lou, Lei Yang, Jingbo Wang, and Taku Komura. SIMS: Simulating human-scene interactions with real world script planning. arXiv preprint arXiv:2411.19921, 2024.3", + "[88] Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. PhysHOI: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393, 2023. 3, 7, 2", + "[89] Yinhuai Wang, Qihan Zhao, Runyi Yu, Ailing Zeng, Jing Lin, Zhengyi Luo, Hok Wai Tsui, Jiwen Yu, Xiu Li, Qifeng Chen, et al. SkillMimic: Learning reusable basketball skills from demonstrations. In CVPR, 2025. 3, 7, 2", + "[90] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. InterControl: Generate human motion interactions by controlling every joint. In NeurIPS, 2024. 2", + "[91] Jungdam Won and Jehee Lee. Learning body shape variation in physics-based characters. ACM Transactions on Graphics (TOG), 38(6):1-12, 2019. 4, 5", + "[92] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. A scalable approach to control diverse behaviors for physically simulated characters. ACM Transactions on Graphics (TOG), 39(4):33-1, 2020. 2, 4", + "[93] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. THOR: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 3", + "[94] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. SAGA: Stochastic whole-body grasping with contact. In ECCV, 2022. 3", + "[95] Zhen Wu, Jiaman Li, Pei Xu, and C Karen Liu. Human-object interaction from human-level instructions. arXiv preprint arXiv:2406.17840, 2024.3", + "[96] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In ICLR, 2024. 2", + "[97] Xianghui Xie, Jan Eric Lenssen, and Gerard Pons-Moll. InterTrack: Tracking human object interaction without object templates. In 3DV, 2024. 3, 6", + "[98] Zhaoming Xie, Sebastian Starke, Hung Yu Ling, and Michiel van de Panne. Learning soccer juggling skills with layer-wise mixture-of-experts. In SIGGRAPH, 2022. 3", + "[99] Zhaoming Xie, Jonathan Tseng, Sebastian Starke, Michiel van de Panne, and C Karen Liu. Hierarchical planning and control for box loco-manipulation. arXiv preprint arXiv:2306.09532, 2023. 3", + "[100] Pei Xu and Ruocheng Wang. Synchronize dual hands for physics-based dexterous guitar playing. In SIGGRAPH Asia, 2024. 2" + ], + "bbox": [ + 94, + 92, + 480, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[101] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 1, 3, 7, 8, 5", + "[102] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. In NeurIPS, 2024. 1, 3", + "[103] Sirui Xu, Dongting Li, Yucheng Zhang, Xiyan Xu, Qi Long, Ziyin Wang, Yunzhi Lu, Shuchang Dong, Hezi Jiang, Akshit Gupta, Yu-Xiong Wang, and Liang-Yan Gui. InterAct: Advancing large-scale versatile 3d human-object interaction generation. In CVPR, 2025. 3, 6", + "[104] Zeshi Yang, Kangkang Yin, and Libin Liu. Learning to use chopsticks in diverse gripping styles. ACM Transactions on Graphics (TOG), 41(4):1-17, 2022. 2", + "[105] Wenhao Yu, Greg Turk, and C Karen Liu. Learning symmetric and low-energy locomotion. ACM Transactions on Graphics (TOG), 37(4):1-12, 2018. 5", + "[106] Ye Yuan and Kris Kitani. Residual force control for agile human behavior imitation and extended motion synthesis. In NeurIPS, 2020. 2", + "[107] Yanjie Ze, Zixuan Chen, Wenhao Wang, Tianyi Chen, Xiaolin He, Ying Yuan, Xue Bin Peng, and Jiajun Wu. Generalizable humanoid manipulation with improved 3d diffusion policies. arXiv preprint arXiv:2410.10803, 2024. 2", + "[108] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 3", + "[109] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. NeuralDome: A neural modeling pipeline on multi-view human-object interactions. In CVPR, 2023. 3, 6, 8", + "[110] Juze Zhang, Jingyan Zhang, Zining Song, Zhanhe Shi, Chengfeng Zhao, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Hoi-m^3: Capture multiple humans and objects interaction within contextual environment. In CVPR, 2024. 3", + "[111] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. COUCH: Towards controllable human-chair interactions. In ECCV, 2022. 3", + "[112] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Ilya Petrov, Vladimir Guzov, Helisa Dhamo, Eduardo Pérez-Pellitero, and Gerard Pons-Moll. FORCE: Dataset and method for intuitive physics guided human-object interaction. In 3DV, 2024. 3", + "[113] Yunbo Zhang, Deepak Gopinath, Yuting Ye, Jessica Hodgins, Greg Turk, and Jungdam Won. Simulation and targeting of complex multi-character interactions. In SIGGRAPH, 2023. 3, 4", + "[114] Chengfeng Zhao, Juze Zhang, Jiashen Du, Ziwei Shan, Junye Wang, Jingyi Yu, Jingya Wang, and Lan Xu. I'M HOI: Inertia-aware monocular capture of 3d human-object interactions. In CVPR, 2024. 3, 6" + ], + "bbox": [ + 94, + 92, + 903, + 868 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12277", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/c17df58e-4f4a-4403-b950-698942e97f1f_model.json b/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/c17df58e-4f4a-4403-b950-698942e97f1f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..526e33884c440cff95f109f33eb731be638fab90 --- /dev/null +++ b/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/c17df58e-4f4a-4403-b950-698942e97f1f_model.json @@ -0,0 +1,2798 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.124, + 0.146, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.135, + 0.89, + 0.181 + ], + "angle": 0, + "content": "INTERMIMIC: Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.206, + 0.773, + 0.226 + ], + "angle": 0, + "content": "Sirui Xu\\(^{1}\\) Hung Yu Ling\\(^{2}\\) Yu-Xiong Wang\\(^{1\\dagger}\\) Liang-Yan Gui\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.226, + 0.748, + 0.263 + ], + "angle": 0, + "content": "1 University of Illinois Urbana-Champaign 2 Electronic Arts Equal Advising" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.264, + 0.664, + 0.278 + ], + "angle": 0, + "content": "https://sirui-xu.github.io/InterMimic" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.305, + 0.907, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.405, + 0.355, + 0.419 + ], + "angle": 0, + "content": "Highly-Dynamic Interaction" + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.405, + 0.621, + 0.419 + ], + "angle": 0, + "content": "Multi-Object Interaction" + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.405, + 0.858, + 0.418 + ], + "angle": 0, + "content": "Scalable Interaction Skills" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.436, + 0.902, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.527, + 0.291, + 0.539 + ], + "angle": 0, + "content": "Humanoid Robot Interaction" + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.527, + 0.433, + 0.54 + ], + "angle": 0, + "content": "MoCap Correction" + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.527, + 0.612, + 0.539 + ], + "angle": 0, + "content": "Interaction Prediction" + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.429, + 0.84, + 0.443 + ], + "angle": 0, + "content": "Lift the woodchair." + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.442, + 0.902, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.527, + 0.836, + 0.539 + ], + "angle": 0, + "content": "Text-to-Interaction" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.553, + 0.907, + 0.611 + ], + "angle": 0, + "content": "Figure 1. InterMimic enables physically simulated humans to perform interactions with dynamic and diverse objects. It supports highly-dynamic, multi-object interactions and scalable skill learning (Top), making it adaptable for versatile downstream applications (Bottom): it can translate whole-body loco-manipulation skills to a humanoid robot [24, 81], perfect interaction MoCap data, and bridge kinematic generation, e.g., predicting future interactions from past (InterDiff [101]) or generating interactions given text prompts (InterDreamer [102])." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.64, + 0.327, + 0.656 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.676, + 0.486, + 0.904 + ], + "angle": 0, + "content": "Achieving realistic simulations of humans interacting with a wide range of objects has long been a fundamental goal. Extending physics-based motion imitation to complex human-object interactions (HOIs) is challenging due to intricate human-object coupling, variability in object geometries, and artifacts in motion capture data, such as inaccurate contacts and limited hand detail. We introduce InterMimic, a framework that enables a single policy to robustly learn from hours of imperfect MoCap data covering diverse full-body interactions with dynamic and varied objects. Our key insight is to employ a curriculum strategy – perfect first, then scale up. We first train subject-specific teacher policies to mimic, retarget, and refine motion capture data. Next, we distill these teachers into a student policy, with the teachers acting as online experts providing direct supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.779 + ], + "angle": 0, + "content": "as well as high-quality references. Notably, we incorporate RL fine-tuning on the student policy to surpass mere demonstration replication and achieve higher-quality solutions. Our experiments demonstrate that InterMimic produces realistic and diverse interactions across multiple HOI datasets. The learned policy generalizes in a zero-shot manner and seamlessly integrates with kinematic generators, elevating the framework from mere imitation to generative modeling of complex human-object interactions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.814, + 0.646, + 0.83 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.842, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Animating human-object interactions is a challenging and time-consuming task even for skilled animators. It requires a deep understanding of physics and meticulous attention to detail to create natural and convincing interactions. While" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12266" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.183 + ], + "angle": 0, + "content": "Motion Capture (MoCap) data provides references, animators often need to correct contact errors caused by sensor limitations and occlusions between humans and objects. However, this process remains unscalable, as refining a single motion demands a delicate balance between preserving the captured data and ensuring its physical plausibility." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.185, + 0.487, + 0.472 + ], + "angle": 0, + "content": "Physics-based human motion imitation [38, 63] offers an alternative approach to improving motion fidelity, by training control policies to mimic reference MoCap data within a physics simulator. However, scaling up human-object interaction imitation presents significant challenges: (i) MoCap Imperfection: Contact artifacts are common, causing expected contacts to fluctuate instead of maintaining consistent zero distance, often due to MoCap limitations or missing hand capture [3, 39]. Accurately imitating MoCap kinematics can result in unrealistic dynamics in simulation. Moreover, HOI datasets often include diverse human shapes, requiring motion retargeting to adapt movements across different human models while preserving interaction dynamics. This retargeting process is imperfect and can introduce new contact artifacts or exacerbate existing ones. (ii) Scaling-up: Although large-scale motion imitation has been explored in previous works [50, 79, 92, 106], it remains largely underexplored for whole-body interactions involving dynamic and diverse objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.639 + ], + "angle": 0, + "content": "In this paper, we aim to utilize rich yet imperfect motion capture interaction datasets to train a control policy capable of learning diverse motor skills while enhancing the plausibility of these actions by correcting errors, such as inaccurate hand motions and faulty contacts. Our approach is grounded on the key insight of tackling the challenges of skill perfection and skill integration progressively. We implement a curriculum-based teacher-student distillation framework, where multiple teacher policies focus on imitating and refining small subsets of interactions, and a student policy integrates these skills from the teachers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Instead of relying on curated data that covers a limited range of actions [4, 51], we employ multiple teacher policies trained on a diverse set of imperfect interaction data and address two key challenges: retargeting and recovering. First, we unify all training policies to a canonical human model, by embedding HOI retargeting directly into the imitation. This is achieved by reframing the policy learning to optimize both imitation and retargeting objectives. Second, our teacher policies refine interaction motion through learning from it, as accurate contact dynamics enforced by a physics simulator inherently correct inaccuracies in the reference kinematics. To support this, we introduce tailored contact-guided reward and optimize trajectory collection, enabling effective skill imitation despite MoCap errors." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Introducing teacher policies offers several key benefits. By leveraging teacher rollouts, we effectively distill raw MoCap data into refined HOI references with a unified em" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.365 + ], + "angle": 0, + "content": "bodiment and enhanced physical fidelity. These refined references guide the subsequent student policy training, reducing the negative impact of errors in the original MoCap data. A major hurdle in scaling motion imitation is the sample inefficiency of Reinforcement Learning (RL), which can lead to prohibitively long training times. Our teacher-student approach mitigates this through a space-time trade-off: multiple teacher policies are trained in parallel on smaller, more manageable data subsets, and their expertise is then distilled into a single student policy. We begin with demonstration-based distillation to bootstrap PPO [71] updates, reducing reliance on pure trial and error and enabling more effective scaling. As training progresses, the student gradually shifts from heavy demonstration guidance to increased RL updates, ultimately surpassing simple demonstration memorization. This mirrors alignment strategies in Large Language Models (LLMs), where demonstration-based pretraining is refined through RL fine-tuning [56]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.908, + 0.63 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows: (i) We introduce InterMimic, which, to the best of our knowledge, is the first framework designed to train physically simulated humans to develop a wide range of whole-body motor skills for interacting with diverse and dynamic objects, extending beyond traditional grasping tasks. (ii) We develop a teacher-student training strategy, where teacher policies provide a unified solution to address the challenges of retargeting and refining in HOI imitation. The student distillation introduces a scalable solution by leveraging a space-time tradeoff. (iii) We demonstrate that our unified framework, InterMimic, as illustrated in Figure 1, effectively handles versatile physics-based interaction animation, recovering motions with realistic and physically plausible details. Notably, by combining kinematic generators with InterMimic, we enable a physics-based agent to achieve tasks such as interaction prediction and text-to-interaction generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.67, + 0.655, + 0.687 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Significant progress has been made in physics-based human interaction animation and control, with advancements in areas such as human-human interactions [45, 90], hand-object interactions [60, 86, 100, 104], human interactions with static scenes [5, 36, 57, 79, 96], and real-world humanoid control for object manipulation [2, 11, 16, 19, 20, 28, 41, 43, 72, 107]. Among these areas, we are the first to achieve universal whole-body loco-manipulation simulation with diverse dynamic objects, beyond pick-and-place and grasping actions – a novel achievement in animation and an idealized reference for real-world humanoid control. Below, we elaborate on recent studies on whole-body interaction animation, particularly involving dynamic objects." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12267" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.383, + 0.106 + ], + "angle": 0, + "content": "2.1. Kinematic Interaction Animation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.447 + ], + "angle": 0, + "content": "Generating human interactions has been a long-standing topic in animation and computer graphics [15, 37]. Significant advances in character animation have emerged with the advent of deep learning, e.g., including phase-function-based methods [22] that enable object interactions like carrying a box [74] or playing basketball [75]. This is extended to more diverse but static objects approaching [35, 77, 94, 111]. Subsequent efforts [14, 26, 27, 40, 42, 49, 66, 95] integrate object motion into interactions but remain constrained by assuming that interactions occur primarily through the hands. To address this, recent developments [7, 10, 12, 21, 30, 62, 73, 84, 93, 101, 102] introduce interactions in a fashion of whole-body loco-manipulation that engages multiple body parts in contact. However, these methods often suffer from physical inaccuracies, such as floating contacts and penetrations, while they generate only body motion without considering hand dexterity. In this work, we address physical inaccuracies by refining imperfect kinematic generation through physics simulation, with InterDiff [101] and HOI-Diff [62] serving as motion planning for loco-manipulation that bridges high-level decision-making (e.g., text instruction) with low-level execution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.455, + 0.411, + 0.47 + ], + "angle": 0, + "content": "2.2. Physics-based Interaction Animation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Physics-based methods generate motion through motor control policies within a physics simulator, e.g., achieved via deep reinforcement learning to track reference motions [63]. These policies are directly applicable for executing simple interactions, such as punching or striking an object [8, 65, 78, 80]. To achieve more complex interactions, early studies focus on specific scenarios, including notable sports-related [52] examples such as basketball [88], skating [44], soccer [98], tennis [108], table tennis [85], and more proposed in [1]. Research also demonstrates flexibility in more general but simpler box carrying tasks [58, 87, 113]. These advancements are achieved through the integration of multiple control policies [55], the use of adversarial motion priors [13, 18, 64], and imitating diverse kinematic generations [95, 99]. However, these methods train their policies in a non-scalable manner, with each policy handling only specific object types or actions. In pursuit of a single, scalable policy to enable multiple interaction skills, existing methods either rely on fixed interaction patterns, such as approaching and grasping objects [4, 51], or extend single-object skills, e.g., interactions involving a basketball [89]. Additionally, they mostly depend on highly curated data from the GRAB dataset [76], which, despite its high quality, primarily features low-dynamic full-body motion and only small-sized objects. More recent datasets [3, 23, 25, 31, 39, 46-48, 53, 97, 103, 109, 110, 112, 114] offer richer full-body interactions with diverse objects but contain noticeable artifacts that chal" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.226, + 0.907, + 0.31 + ], + "angle": 0, + "content": "Figure 2. Our two-stage pipeline: (i) training each teacher policy (MLP) on a small data subset with initialization corrected via Physical State Initialization (PSI), and (ii) freezing the teacher policies to provide refined references for training a student policy (Transformer). The student leverages teacher supervision for effective scaling and is fine-tuned through RL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.907, + 0.431 + ], + "angle": 0, + "content": "lenge existing motion imitation approaches. We process data from OMOMO [39], BEHAVE [3], HODome [109], and IMHD [114] collected in the InterAct [103] dataset, and the multi-object dataset HIMO [53], highlighting InterMimic's scalability to diverse interactions and its robustness against MoCap artifacts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.448, + 0.649, + 0.466 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Task Formulation. The goal of human-object interaction (HOI) imitation is to learn a policy \\(\\pi\\) that produces simulated human-object motion \\(\\{\\pmb{q}_t\\}_{t=1}^T\\) closely matching a ground-truth reference \\(\\{\\hat{\\pmb{q}}_t\\}_{t=1}^T\\) derived from large-scale MoCap data. Given the geometries of the human and objects, the policy should also compensate for missing or inaccurate details in the dataset. Each pose \\(\\pmb{q}_t\\) has two components: the human pose \\(\\pmb{q}_t^h\\) and the object pose \\(\\pmb{q}_t^o\\). The human pose is defined as \\(\\pmb{q}_t^h = \\{\\pmb{\\theta}_t^h, \\pmb{p}_t^h\\}\\), where \\(\\pmb{\\theta}_t^h \\in \\mathbb{R}^{52 \\times 3}\\) represents the joint rotations, and \\(\\pmb{p}_t^h \\in \\mathbb{R}^{52 \\times 3}\\) specifies the joint positions. Specifically, our human model includes 30 hand joints and 22 joints for the rest of the body, with one root joint's rotation and position specified in global coordinates, whereas the rotations and positions of all other joints are defined relative to their respective parent joints' coordinate systems. The object pose \\(\\pmb{q}_t^o\\) is represented as \\(\\{\\pmb{\\theta}_t^o, \\pmb{p}_t^o\\}\\), where \\(\\pmb{\\theta}_t^o \\in \\mathbb{R}^3\\) denotes the object's orientation and \\(\\pmb{p}_t^o \\in \\mathbb{R}^3\\) the position. All simulation states have corresponding ground-truth values, denoted by the hat symbol. For instance, the reference object rotation is \\(\\{\\hat{\\pmb{\\theta}}_t^o\\}_{t=1}^T\\). The environmental setup for the simulation is detailed in Sec. B. Overview. We formulate interaction imitation as a Markov Decision Process (MDP), defined by states, actions, simulator-provided transition dynamics, and a reward function. Figure 2 illustrates our two-stage framework: (i) training teacher policies \\(\\pi^{(T)}\\) on small skill subsets, and (ii) distilling these teachers into a scalable student policy \\(\\pi^{(S)}\\) for large-scale skill learning. In Sec. 3.1, we define the states \\(s_t\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12268" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.198 + ], + "angle": 0, + "content": "and actions \\(\\pmb{a}_t\\), applicable to both teacher \\(\\pi^{(T)}\\) and student \\(\\pi^{(S)}\\) policies. In Sec. 3.2, we describe how teacher policies are trained via RL, focusing on reward designs that facilitate retargeting, as well as techniques that mitigate the impact of imperfections in the reference data. Sec. 3.3 details the subsequent distillation of teachers into a scalable student policy, leveraging both RL and learning from demonstration." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.217, + 0.295, + 0.233 + ], + "angle": 0, + "content": "3.1. Policy Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.485, + 0.504 + ], + "angle": 0, + "content": "State. The state \\( s_t \\), which serves as input to the policy, comprises two components \\( s_t = \\{s_t^s, s_t^g\\} \\). The first part, \\( s_t^s \\), contains human proprioception and object observations, expressed as, \\( \\{\\{\\pmb{\\theta}_t^h, \\pmb{p}_t^h, \\pmb{\\omega}_t^h, \\pmb{v}_t^h\\}, \\{\\pmb{\\theta}_t^o, \\pmb{p}_t^o, \\pmb{\\omega}_t^o, \\pmb{v}_t^o\\}, \\{d_t, c_t\\}\\} \\) where \\( \\{\\pmb{\\theta}_t^h, \\pmb{p}_t^h, \\pmb{\\omega}_t^h, \\pmb{v}_t^h\\} \\) represent the rotation, position, angular velocity, and velocity of all joints, respectively, while \\( \\{\\pmb{\\theta}_t^o, \\pmb{p}_t^o, \\pmb{\\omega}_t^o, \\pmb{v}_t^o\\} \\) represent the orientation, location, velocity, and angular velocity of the object, respectively. Motivated by [6], we include object geometry and whole-body haptic sensing from two elements: (i) \\( d_t \\), vectors from human joints to their nearest points on each object surface; and (ii) \\( c_t \\), contact markers indicating whether the human's rigid body parts experience applied forces; this serves as simplified tactile or force sensing - an important multimodal input in robot manipulation tasks [9]. The goal state \\( s_t^g = \\{s_{t,t+k}^g\\}_{k \\in K} \\) integrates reference poses from the ground truth motion, where \\( s_{t,t+k}^g \\) is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\{\\{\\hat {\\boldsymbol {\\theta}} _ {t + k} ^ {h} \\ominus \\boldsymbol {\\theta} _ {t} ^ {h}, \\hat {\\boldsymbol {p}} _ {t + k} ^ {h} - \\boldsymbol {p} _ {t} ^ {h} \\}, \\{\\hat {\\boldsymbol {\\theta}} _ {t + k} ^ {o} \\ominus \\boldsymbol {\\theta} _ {t} ^ {o}, \\hat {\\boldsymbol {p}} _ {t + k} ^ {o} - \\boldsymbol {p} _ {t} ^ {o} \\}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.548, + 0.453, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\{\\hat {\\boldsymbol {d}} _ {t + k} - \\boldsymbol {d} _ {t}, \\hat {\\boldsymbol {c}} _ {t + k} - \\boldsymbol {c} _ {t} \\}, \\{\\hat {\\boldsymbol {\\theta}} _ {t + k} ^ {h}, \\hat {\\boldsymbol {p}} _ {t + k} ^ {h}, \\hat {\\boldsymbol {\\theta}} _ {t + k} ^ {o}, \\hat {\\boldsymbol {p}} _ {t + k} ^ {o} \\} \\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.483, + 0.669 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{\\theta}}_{t + k}^{h},\\hat{\\pmb{p}}_{t + k}^{h},\\hat{\\pmb{d}}_{t + k},\\hat{\\pmb{c}}_{t + k}\\) represent the reference information at time step \\(t + k,\\ominus\\) denotes the calculation of rotation difference. All continuous elements of \\(s_t\\) are normalized relative to the current direction of view of the human and the position of the root [63]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Given that most MoCap data does not provide reference contact or tactile information, we extract reference contact markers \\(\\hat{c}_{t + k}\\) by inferring dynamic information, beyond relying solely on inaccurate contact distances, specifically by analyzing the object's acceleration to detect human-induced forces. To accommodate the variability in contact distances observed in reference motion, we discretize reference contact markers using varying distance thresholds, as illustrated in Fig. 3(i). The neutral areas serve as buffer zones, avoiding the penalization or enforcement of strict contact. See Sec. C of supplementary for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Action. Our human model has 51 actuated joints, defining an action space of \\( \\mathbf{a}_t \\in \\mathbb{R}^{51 \\times 3} \\). These actions are specified as joint PD targets using the exponential map and are converted into torques applied to each of the human joints." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.727, + 0.108 + ], + "angle": 0, + "content": "3.2. Imitation as Perfecting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.114, + 0.905, + 0.251 + ], + "angle": 0, + "content": "The teacher policy \\(\\pi^{(\\mathrm{T})}\\) is trained via RL to maximize the expected discounted reward by comparing simulated states against potentially erroneous reference states. The training involves: (i) trajectory collection, where we explain how trajectories are initialized and terminated. (ii) policy updating, where collected trajectories and their associated rewards are used to refine the policy. In this section, we elaborate on our reward design and how we optimize trajectory collection to mitigate the impact of reference inaccuracies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.906, + 0.476 + ], + "angle": 0, + "content": "Imitation as Retargeting. We tailor teacher policies to each human subject, while all policies share the same base human model. This serves the retargeting purpose by converting HOIs from different human shapes into a unified base shape. Although motion imitation does not necessarily require a unified human model [50, 91], our approach offers two benefits: (i) It enhances integration with kinematic generation methods, which generally perform better on a single, unified shape [17]. (ii) It demonstrates possible integration with real-world humanoid deployment, which requires retargeting to a consistent physical embodiment. In Figure 1, our method translates MoCap data into motor skills on a Unitree G1 [81] with two Inspire hands [24], all without external retargeting in complex contact-rich scenarios. See Sec. F of the supplementary for additional details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Human [83] or HOI [32] retargeting can be formulated as an optimization problem. Inverse Kinematics (IK) methods, such as those based on quadratic programming [34], demonstrate effectiveness in simplified scenarios but remain underexplored for motions featuring intricate object interactions. RL, by contrast, solves the optimization by maximizing an expected cumulative reward, prompting us to investigate whether RL-driven HOI imitation can be used for HOI retargeting. This extends existing physics-based retargeting approaches, which either omit object interactions [67] or are non-scalable with a single reference [113]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.749 + ], + "angle": 0, + "content": "While the kinematics should differ due to the embodiment gap, we argue that the dynamics between human and object should remain invariant. Thus, we define rewards to include an embodiment-aware component that loosely aligns the simulated kinematics with the reference interaction, and an embodiment-agnostic reward component that encourages dynamics to be close to the reference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Embodiment-Aware Reward. When the human and object are far apart, retargeting should prioritize capturing rotational motion, whereas when they are close, accurate position tracking becomes crucial for achieving contact. To reflect this, we define the weights \\(\\pmb{w}_d\\) that are inversely proportional to the distances between joints and the object [113]. The reward thus includes cost functions for joint position \\(E_p^h = \\langle \\pmb{\\Delta}_p^h, \\pmb{w}_d \\rangle\\), rotation \\(E_\\theta^h = \\langle \\pmb{\\Delta}_\\theta^h, \\mathbf{1} - \\pmb{w}_d \\rangle\\), and interaction tracking \\(E_d = \\langle \\pmb{\\Delta}_d, \\pmb{w}_d \\rangle\\), where \\(\\langle \\cdot, \\cdot \\rangle\\) is the inner product, \\(\\pmb{\\Delta}_p^h[i] = \\| \\hat{\\pmb{p}}^h[i] - \\pmb{p}^h[i] \\|\\), \\(\\pmb{\\Delta}_\\theta^h[i] =\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12269" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.278, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.164, + 0.264, + 0.175 + ], + "angle": 0, + "content": "(i) Reference contact markers" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.098, + 0.352, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.099, + 0.482, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.163, + 0.416, + 0.175 + ], + "angle": 0, + "content": "(ii) RSI vs. PSI" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.186, + 0.483, + 0.256 + ], + "angle": 0, + "content": "Figure 3. (i) Visualization of reference contact markers that accommodate varied contact distances: red to promote contact, green for neutral areas where contact is neither promoted nor penalized, and blue to penalize contact. (ii) Initializing the rollout with reference (RSI) or reference corrected via simulation (PSI)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.284, + 0.484, + 0.379 + ], + "angle": 0, + "content": "\\(\\| \\hat{\\pmb{\\theta}}^h [i]\\ominus \\pmb{\\theta}^h [i]\\|\\), and \\(\\pmb{\\Delta}_d[i] = \\| \\hat{\\pmb{d}} [i] - \\pmb {d}[i]\\|\\) represent the displacement for the variables defined in Sec. 3.1 with timestep \\(t\\) omitted. The formulation of \\(\\pmb{w}_{d}\\) is provided in Sec. D of supplementary. The reward to be maximized can be formulated as \\(\\exp (-\\lambda E)\\) for each cost function \\(E\\) with a specific hyperparameter \\(\\lambda\\). Details can be found in Sec. D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Embodiment-Agnostic Reward. The reward includes components for object tracking and contact tracking. The object tracking cost is defined for position \\( E_{p}^{o} = \\|\\hat{\\boldsymbol{p}}^{o} - \\boldsymbol{p}^{o}\\| \\) and rotation \\( E_{\\theta}^{o} = \\|\\hat{\\boldsymbol{\\theta}}^{o} - \\boldsymbol{\\theta}^{o}\\| \\), with all values normalized to the human's current position and direction. The contact tracking reward comprises two cost functions: body contact promotion \\( E_{b}^{c} \\) and penalty \\( E_{p}^{c} \\), both aligning the simulated contact \\( c \\) with reference markers \\( \\hat{c} \\), as shown in Figure 3. We define three contact levels – promotion, penalty, and neutral – to accommodate potential inaccuracies in reference contact distances. The detailed formulation can be found in Sec. D of the supplementary. Since the physics engine does not differentiate between object, ground, and self-contact, we adopt two strategies: (i) we model foot-ground contact promotion and penalty. This ensures proper foot lifting during cyclic walking and mitigates foot hobbling. (ii) We allow self-collision to avoid self-contact promotion but to promote object interaction. This poses minimal risk as the policy is guided by MoCap reference, which, although lacking perfect contact accuracy, rarely shows self-penetration. For humanoid robots with embodiments that differ from the MoCap reference and require real-world applicability, we disable self-collision, as discussed in Sec. F." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.794 + ], + "angle": 0, + "content": "We introduce energy consumption rewards [105] to penalize large human or object jitters, with a proposed contact energy penalizing abrupt contact to promote compliant interactions. See Sec. D of supplementary for more details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Hand Interaction Discovery. We use data with average or flattened hand poses [3, 39], which makes accurate object manipulation imitation challenging. To address this, we activate a reference contact marker for any hand part when a fingertip or palm is near an object. Given tasks that do not demand high dexterity, employing a contact-promoting reward with this marker enables policies to develop effective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "hand interaction strategies, leveraging the exploratory nature of RL. Additionally, we constrain the range of motion (RoM) of the hands to ensure natural movement. See Sec. D and Sec. B of the supplementary for further details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.154, + 0.907, + 0.26 + ], + "angle": 0, + "content": "Policy Learning. Following [63], the control policy \\(\\pi\\) is trained using PPO [71] with the policy gradient \\(L(\\psi) = \\mathbb{E}_t[\\min (r_t(\\psi)A_t,\\mathrm{clip}(r_t(\\psi),1 - \\epsilon ,1 + \\epsilon)A_t)\\). \\(\\psi\\) are the parameters of \\(\\pi\\) and \\(r_t(\\psi)\\) quantifies the difference in action likelihoods between updated and old policies. \\(\\epsilon\\) is a small constant, and \\(A_{t}\\) is the advantage estimate given by the generalized advantage estimator \\(\\mathrm{GAE}(\\lambda)\\) [70]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.908, + 0.488 + ], + "angle": 0, + "content": "Physical State Initialization. Learning later-phase motion can be essential for policies to achieve high rewards during earlier phases, compared to incrementally learning from the starting phase. Thus, Reference State Initialization (RSI) [63] sets the current pose \\( \\mathbf{q}_t \\) to a reference pose \\( \\hat{\\mathbf{q}}_t \\) at a random timestep \\( t \\), for initializing the rollout. However, initializing with the imperfect reference can introduce critical artifacts, such as contact floating or incorrect hand motion, leading to unrecoverable failures, e.g., object falling, as depicted in Figure 3(ii). These issues render many initializations ineffective, limiting training on certain interaction phases since successful rollouts may not reach them before the maximum length. The problem is exacerbated by the use of prioritized sampling [79, 91], which favors high-failure-rate initializations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.908, + 0.748 + ], + "angle": 0, + "content": "To address the need for higher-quality reference initialization, we propose Physical State Initialization (PSI). As illustrated in Figure 2, PSI begins by creating an initialization buffer that stores reference states from MoCap and simulation states from prior rollouts. For each new rollout, an initial state is randomly selected from this buffer, which increases the likelihood of starting from advantageous positions. Once a rollout is completed, trajectories are evaluated based on their expected discounted rewards; those above a certain threshold are added to the buffer using a first-in-first-out (FIFO) strategy, while older or lower-quality trajectories are discarded. This selective reintroduction of high-value states for initialization helps maintain stable policy updates. We apply PSI in a sparse manner to ensure training efficiency. As shown in Figure 3(ii), PSI can collect trajectories for policy update that RSI does not effectively utilize. Further details are provided in Sec. E of the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Interaction Early Termination. Early Termination (ET) [63] is commonly used in motion imitation, ending an episode when a body part makes unplanned ground contact or when the character deviates significantly from the reference [50], thus stopping the policy from overvaluing invalid transitions. However, additional conditions should be considered for human-object interactions. We propose Interaction Early Termination (IET), which supplements ET with three extra checks: (i) Object points deviate from their references by more than \\(0.5\\mathrm{m}\\) on average. (ii) Weighted av" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12270" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "erage distances between the character's joints and the object surface exceed \\(0.5\\mathrm{m}\\) from the reference. (iii) Any required body-object contact is lost for over 10 consecutive frames. Full conditions are detailed in Sec. E of the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.16, + 0.33, + 0.175 + ], + "angle": 0, + "content": "3.3. Imitation with Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.484, + 0.289 + ], + "angle": 0, + "content": "As shown in Figure 2, after training the teacher policies on data from each subject (Sec. 3.2), we aggregate them to train a student policy \\(\\pi^{(\\mathrm{S})}\\) to master all skills. As outlined in Algorithm 1, the combined teacher policies, denoted by \\(\\pi^{(\\mathrm{T})}\\) for brevity, serves dual roles by providing state-action trajectories \\((s^{(\\mathrm{T})}, a^{(\\mathrm{T})})\\): (i) the state \\(s^{(\\mathrm{T})}\\) for reference distillation, and (ii) the action \\(a^{(\\mathrm{T})}\\) for policy distillation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.424 + ], + "angle": 0, + "content": "Reference Distillation. Noisy MoCap data can hinder policy learning, especially at larger scales. In contrast, teacher policies trained on smaller-scale data effectively address these issues by correcting contact artifacts, refining hand placements, and recovering missing details (see Figures 1 and 5). To fully leverage teacher policies, we use their rollouts as references for defining the student policy's goal state and reward functions, distinguishing our approach from distillation based on only action output." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.484, + 0.668 + ], + "angle": 0, + "content": "Policy Distillation. We apply distillation on action outputs, which we view as crucial for scaling policies to large datasets. In essence, we trade space for time: teacher policies are trained in parallel on smaller data subsets, allowing the student policy to scale through distillation. Following Algorithm 1, we begin with Behavior Cloning (BC) [29] and use RL fine-tuning to go beyond demonstration memorization, an approach common in LLM alignment [56]. We integrate BC into online policy updates with a staged schedule: we start with DAgger [69] and gradually transition to PPO. Throughout, the critic is continuously trained with the reward from Sec. 3.2. This RL fine-tuning phase is crucial as teacher policies may behave differently when performing similar skills, and simple BC can lead to suboptimal \"averaging\" behavior, where RL fine-tuning helps the student converge on optimal solutions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.675, + 0.226, + 0.689 + ], + "angle": 0, + "content": "3.4. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.484, + 0.835 + ], + "angle": 0, + "content": "We set the keyframe indices \\( K \\) (Sec. 3.1, Eq. 1) to \\( \\{1,16\\} \\) for the teacher policies and \\( \\{1,2,4,16\\} \\) for the student policy. The broader observation window for the student policy helps it better distinguish different skills with larger-scale data. Teacher policies employ MLPs, common in physics-based animation [63], while the student policy handles higher-dimensional observations, for which MLPs are less effective. Thus, we use a transformer [82] architecture for sequential modeling [79], as shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.846, + 0.224, + 0.863 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We evaluate teacher policies on their ability to imitate imperfect HOI references, and assess the entire teacher-" + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.091, + 0.82, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Distillation with RL Fine-tuning" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.111, + 0.907, + 0.47 + ], + "angle": 0, + "content": "1: Input: A composite policy \\(\\pi^{(\\mathrm{T})}\\) integrated from individual teacher policies, student policy parameters \\(\\psi\\), student value function parameters \\(\\phi\\), schedule hyperparameter \\(\\beta\\) for DAgger, horizon length \\(H\\) for PPO \n2: for \\(t = 0,1,2,\\ldots\\) do \n3: for \\(h\\) from 1 to \\(H\\) do \n4: Sample a variable \\(u \\sim \\mathrm{Uniform}(0,1)\\) \n5: Collect \\(s^{(\\mathrm{T})}, a^{(\\mathrm{T})}\\) from teacher \\(\\pi^{(\\mathrm{T})}\\) \n6: Obtain the refined reference from \\(s^{(\\mathrm{T})}\\) to define \\(s^{(\\mathrm{S})}\\) and \\(r(\\cdot)\\), obtain \\(a^{(\\mathrm{S})}\\) from \\(\\pi_{\\phi}^{(\\mathrm{S})}(a^{(\\mathrm{S})}|s^{(\\mathrm{S})})\\). \n7: if \\(u \\leq \\frac{t}{\\beta}\\) then ▷ Use the teacher \n8: Given \\(s^{(\\mathrm{S})}\\), execute \\(a^{(\\mathrm{S})}\\), observe \\(s^{\\prime (\\mathrm{S})}, r\\) \n9: else ▷ Use the student \n10: Given \\(s^{(\\mathrm{S})}\\), execute \\(a^{(\\mathrm{T})}\\), observe \\(s^{\\prime (\\mathrm{S})}, r\\) \n11: end if \n12: Store the transition \\((s^{(\\mathrm{S})}, s^{\\prime (\\mathrm{S})}, a^{(\\mathrm{S})}, a^{(\\mathrm{T})}, r)\\) \n13: end for \n14: Update \\(\\phi\\) with TD(\\(\\lambda\\)) \n15: Compute PPO objective: \\(L(\\psi)\\) \n16: Compute \\(J(\\psi) = \\| a^{(\\mathrm{S})} - a^{(\\mathrm{T})}\\|\\) \n17: Compute the weight: \\(w = \\min \\left(\\frac{t}{\\beta}, 1\\right)\\) \n18: Update \\(\\psi\\) by gradient: \\(\\nabla_{\\psi}(wL(\\psi) + (1 - w)J(\\psi))\\) \n19: end for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.548 + ], + "angle": 0, + "content": "student framework for scalability to large-scale data and zero-shot generalization across various scenarios. Additional experiments are provided in Sec. G of supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.909, + 0.808 + ], + "angle": 0, + "content": "Datasets. We use the following datasets from InterAct [103]: OMOMO [39], BEHAV [3], HODome [109], IMHD [114], and HIMO [53]. OMOMO, containing 15 objects and approximately 10 hours of data, is our primary dataset for evaluating the full teacher-student distillation framework for its scale. We train 17 teacher policies, one per subject, with subject 14 reserved as the test set and the remaining data used for training the student policy. A small portion of data is discarded after teacher imitation due to severe MoCap errors that could not be corrected (see Sec. F and Sec. H of the supplementary). Additional datasets are used to evaluate teacher policies in various MoCap scenarios with different error levels and interaction types. We focus on highly dynamic motions (Figure 1) and interactions involving multiple body parts (Figure 4), while excluding scenarios such as carrying a bag with a strap, since the simulator [54] used lacks full support for soft bodies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Metrics. We use the following metrics: (i) Success Rate is defined as the proportion of references that the policy successfully imitates at least once, averaged across all references, while (ii) Duration is the time (in seconds) that the imitation is maintained without triggering the interaction early termination conditions introduced in Sec. 3.2. (iii)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12271" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.086, + 0.488, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.318, + 0.486, + 0.39 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison between PhysHOI [88] (top), the reference motion (middle) from the BEHAVE [3] dataset, and the interaction refined by our teacher trained on it (bottom). InterMimic faithfully imitates the interactions involving multiple body parts while correcting errors in the original reference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.52 + ], + "angle": 0, + "content": "Human Tracking Error \\((E_h)\\) measures the per-joint position error (cm) between the simulated and reference human (excluding hand joints for BEHAV [3] and OMOMO [39] due to inaccuracy). (iv) Object Tracking Error \\((E_o)\\) measures the per-point position error (cm) between the simulated and reference object. Both errors are averaged over the duration of the imitation in the best-performing trial." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.702 + ], + "angle": 0, + "content": "Baselines. To facilitate fair comparisons, we downgrade our method for teacher policy evaluation to imitate either a single MoCap clip (Figure 4) or multiple clips with a single object (Table 1), enabling direct comparison with PhysHOI [88] and SkillMimic [89] (Sec. 4.1 and 4.2). Due to the lack of established baselines for large-scale HOI imitation, we adapt the following variants for comparison with our student policy (Sec. 4.3): (i) PPO [71] trains an imitation policy from scratch, following [63]. We experiment with both versions, with and without reference distillation; (ii) DAgger [69] distills the student without RL fine-tuning, a process we refer to as policy distillation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. The control policies operate at \\(30\\mathrm{Hz}\\) and are trained using the Isaac Gym simulator [54]. Teacher policies are implemented as MLPs with hidden layers of sizes 1024, 1024, and 512. The student policy is implemented as a three-layer Transformer encoder with 4 heads, a hidden size of 256, and a feed-forward layer of 512. The critics are also modeled as MLPs with the same architecture as the teacher policies. To integrate the student policy with kinematic generators, including text-to-HOI [62] and future interaction prediction [101], we train these models using reference data distilled by the teacher policies from the OMOMO [39] dataset, following the same train-test split as the student policy training. For the text-to" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.906, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.27, + 0.908, + 0.314 + ], + "angle": 0, + "content": "Figure 5. We recover plausible object rotations (bottom) that are challenging for motion capture due to the equivariant geometries of objects, which result in the object sliding on the ground (top)." + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.324, + 0.868, + 0.405 + ], + "angle": 0, + "content": "
MethodTime↑Eh↓Eo↓
SkillMimic [89]12.27.213.4
InterMimic (Ours) w/o IET40.36.79.9
InterMimic (Ours) w/o PSI36.16.610.2
InterMimic (Ours)42.66.49.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.415, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison between the teacher policy from InterMimic and SkillMimic [89] to imitate data extracted from the BEHAVE [3] dataset involving a single subject interacting with yogamat. We ablate our proposed approach by removing interaction early termination and physical state initialization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.906, + 0.581 + ], + "angle": 0, + "content": "HOI model, we train it to generate 10 seconds of motion and use 24 generated samples for evaluation, while for future interaction prediction, the model generates 25 future frames given 10 past frames and we use 60 generated samples for evaluation. See Sec. F of the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.736, + 0.605 + ], + "angle": 0, + "content": "4.1.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.907, + 0.764 + ], + "angle": 0, + "content": "Table 1 shows that the baseline struggles with MoCap imperfections, e.g., incorrect hand positioning, and thus results in clearly shorter tracking durations. In contrast, our method maintains reference tracking for longer durations and produces interactions that closely match the reference. Table 2 shows that our method consistently outperforms baselines in both training data imitation and out-of-distribution generalization, including interactions from the test set and from kinematic generations. We discuss the effectiveness of specific design choices in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.725, + 0.789 + ], + "angle": 0, + "content": "4.2. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Figure 4 shows a representative sequence from the experiment in Table 1, illustrating how our teacher policy corrects interactions that PhysHOI [88] fails to track robustly – our method effectively withstands and corrects incorrect hand positioning and floating contacts in the reference. Beyond obvious errors, our method also rectifies the rotation of symmetric objects that MoCap inaccurately depicts as slid" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12272" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.166 + ], + "angle": 0, + "content": "
PPOReference DistillationPolicy DistillationArchitectureOMOMO-Train [39]OMOMO [39]-TestOMOMO [39]-Test (w × 10)HOI-Diff [62]InterDiff [101]
Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓
✓××MLP23.9101.67.215.69.685.37.516.23.971.27.517.90.00.0--6.711.76.216.4
×✓✓54.5139.97.111.054.3140.27.111.215.591.79.313.74.284.810.19.765.027.47.513.4
✓✓×71.7152.88.912.791.6173.78.513.245.8127.69.114.98.3130.910.113.873.328.96.914.4
✓✓✓90.7168.05.59.795.5173.95.411.962.6140.96.614.512.5121.48.612.175.029.16.213.5
✓✓✓Transformer88.8167.06.010.298.1176.55.911.356.8134.76.613.212.5119.08.512.676.729.36.413.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.177, + 0.907, + 0.205 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation of large-scale interaction imitation using OMOMO [39], kinematic generations from HOI-Diff [62], and InterDiff [101]. Additionally, we evaluate on test set when objects with weights ten times greater than those used during training." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.228, + 0.481, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.394, + 0.483, + 0.452 + ], + "angle": 0, + "content": "Figure 6. Zero-shot integration with a text-to-HOI model HOI-Diff [62] (Top), using 'Kick the large box' as the prompt, and an interaction prediction model InterDiff [101] (Bottom), where gray meshes are past states and colored illustrate future generations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.476, + 0.483, + 0.598 + ], + "angle": 0, + "content": "ing along the ground, shown in Figure 5. Figure 6 presents additional examples that complement Figure 1, demonstrating how our approach integrates with kinematic generators for future interaction prediction and text-to-interaction synthesis. This zero-shot generalization extends to novel objects unseen during training (Figure 7), highlighting the effectiveness of our object geometry and contact-encoded representation, as well as the large-scale training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.245, + 0.622 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Effectiveness of PSI and IET. We conduct an ablation study, as demonstrated in Table 1, comparing the full approach to \"Ours w/o PSI\". The results validate that Physical State Initialization (PSI) is effective by mitigating inaccuracies in the motion capture data. We also observe reduced effectiveness without our interaction early termination, as training often spends rollouts on irrelevant periods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Effectiveness of Reference Distillation. Compared to directly scaling imitation from MoCap with potential imperfections (line 1 in Table 2), using references refined by the teacher policy (line 3) achieves consistently better performance on all metrics. The improvement is even more pronounced on the test set, where, without reference distillation, the policy struggles with unseen shapes, while retargeting by reference distillation eliminates the difficulty." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Joint PPO and DAgger Updates. As shown in Table 2, training a policy from scratch (line 3) or relying solely on policy distillation (DAgger, line 2) fails to" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.228, + 0.905, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.311, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Figure 7. Zero-shot generalization of our student policy on novel objects from BEHAV [3] and HODome [109]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.905, + 0.469 + ], + "angle": 0, + "content": "achieve optimal performance. While supervised skill learning lays the groundwork, additional PPO fine-tuning is crucial for resolving conflicts among teacher policies. This is important because our subject-based clustering may not effectively distinguish between different interaction patterns, and ambiguity arises when multiple teachers produce different actions for similar motions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.591 + ], + "angle": 0, + "content": "Effectiveness of Transformer for Policy Learning. From Table 2, we see that using a Transformer policy (line 5) outperforms MLP-based approaches, particularly on the test set and out-of-distribution cases generated by the kinematic model. We attribute this to the Transformer's inductive bias in sequential modeling and its capacity to incorporate longer-term observations, enabling it to handle complex spatio-temporal dependencies more effectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.634, + 0.62 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduce a framework for synthesizing realistic human-object interactions that are both physically grounded and generalizable. Unlike previous methods, our approach leverages a rich repository of imperfect MoCap data to facilitate the learning of various interaction skills across a wide variety of objects. To address inaccuracies in the MoCap data, we propose contact-guided rewards and optimize trajectory collection, enabling teacher policies to recover missing physical details in the original data. These teacher policies are used to train student policies within a distillation framework that combines policy distillation and reference distillation, thus enabling efficient skill scaling. Our approach shows zero-shot generalizability, which effectively bridges the gap between imitation and generative capabilities by integrating with kinematic generation. We believe that this framework can be adapted for whole-body loco-manipulation for real-world robots, enabling them to handle objects with human-like dexterity and nuance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12273" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.334 + ], + "angle": 0, + "content": "Acknowledgments. We thank Wei Yang, Yu-Wei Chao, Arsalan Mousavian, Ankur Handa, Samuel Schulter, Morteza Ziyadi, and Xiaohan Fei for valuable discussions. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Amazon-Illinois Center on AI for Interactive Conversational Experiences, the Toyota Research Institute, the IBM-Illinois Discovery Accelerator Institute, and Snap Inc. This work used computational resources, including the NCSA Delta and DeltaAI and the PTI Jetstream2 supercomputers through allocations CIS230012, CIS230013, and CIS240311 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, as well as the TACC Frontera supercomputer and Amazon Web Services (AWS) through the National Artificial Intelligence Research Resource (NAIRR) Pilot." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.348, + 0.188, + 0.364 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.373, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[1] Jinseok Bae, Jungdam Won, Donggeun Lim, Cheol-Hui Min, and Young Min Kim. Pmp: Learning to physically interact with environments using part-wise motion priors. In SIGGRAPH, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.431, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[2] Qingwei Ben, Feiyu Jia, Jia Zeng, Junting Dong, Dahua Lin, and Jiangmiao Pang. HOMIE: Humanoid locomap-. manipulation with isomorphic exoskeleton cockpit. arXiv preprint arXiv:2502.13013, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.488, + 0.482, + 0.541 + ], + "angle": 0, + "content": "[3] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAV: Dataset and method for tracking human object interactions. In CVPR, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.546, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[4] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. In 3DV, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.589, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[5] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In AAAI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.632, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[6] Sammy Christen, Muhammed Kocabas, Emre Aksan, Jemin Hwangbo, Jie Song, and Otmar Hilliges. D-grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.689, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[7] Enric Corona, Albert Pumarola, Guillem Alenya, and Francesc Moreno-Noguer. Context-aware human motion prediction. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.733, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[8] Jieming Cui, Tengyu Liu, Nian Liu, Yaodong Yang, Yixin Zhu, and Siyuan Huang. AnySkill: Learning open-vocabulary physical skill for interactive agents. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.789, + 0.482, + 0.841 + ], + "angle": 0, + "content": "[9] Ravinder S Dahiya, Giorgio Metta, Maurizio Valle, and Giulio Sandini. Tactile sensing—from humans to humanoids. IEEE transactions on robotics, 26(1):1-20, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[10] Divyanshu Daiya, Damon Conover, and Aniket Bera. COLLAGE: Collaborative human-agent interaction generation using hierarchical latent diffusion and language models. arXiv preprint arXiv:2409.20502, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.373, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[11] Jeremy Dao, Helei Duan, and Alan Fern. Sim-to-real learning for humanoid box loco-manipulation. In ICRA, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.136, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[12] Christian Diller and Angela Dai. CG-HOI: Contact-guided 3d human-object interaction generation. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.166, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[13] Jiawei Gao, Ziqin Wang, Zeqi Xiao, Jingbo Wang, Tai Wang, Jinkun Cao, Xiaolin Hu, Si Liu, Jifeng Dai, and Jiangmiao Pang. CooHOI: Learning cooperative human-object interaction with manipulated object dynamics. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.236, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[14] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. IMoS: Intent-driven full-body motion synthesis for human-object interactions. In Computer Graphics Forum, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.293, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[15] Michael Gleicher. Motion editing with spacetime constraints. In Proceedings of the 1997 symposium on Interactive 3D graphics, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.335, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[16] Zhaoyuan Gu, Junheng Li, Wenlan Shen, Wenhao Yu, Zhaoming Xie, Stephen McCrory, Xianyi Cheng, Abdulaziz Shamsah, Robert Griffin, C Karen Liu, et al. Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning. arXiv preprint arXiv:2501.02116, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.42, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[17] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.462, + 0.904, + 0.502 + ], + "angle": 0, + "content": "[18] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.505, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[19] Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. OmniH2O: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. arXiv preprint arXiv:2406.08858, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.576, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[20] Tairan He, Zhengyi Luo, Wenli Xiao, Chong Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Learning human-to-humanoid real-time whole-body teleoperation. In IROS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.632, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[21] Wenkun He, Yun Liu, Ruitao Liu, and Li Yi. SyncDiff: Synchronized motion diffusion for multi-body human-object interaction synthesis. arXiv preprint arXiv:2412.20104, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.689, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[22] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.732, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[23] Yinghao Huang, Omid Taheri, Michael J. Black, and Dimitrios Tzionas. InterCap: Joint markerless 3D tracking of humans and objects in interaction. In GCRP, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.775, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[24] Inspire-robots. Smaller and higher-precision motion control experts. https://inspire-robots/store/.1,4,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[25] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. CHAIRS: Towards full-body articulated human-object interaction. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Nan Jiang, Zimo He, Zi Wang, Hongjie Li, Yixin Chen, Siyuan Huang, and Yixin Zhu. Autonomous character" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12274" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "scene interaction synthesis from text instruction. In SIGGRAPH Asia, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[27] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.179, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[28] Zhenyu Jiang, Yuqi Xie, Kevin Lin, Zhenjia Xu, Weikang Wan, Ajay Mandlekar, Linxi Fan, and Yuke Zhu. DexMimicGen: Automated data generation for bimanual dexterous manipulation via imitation learning. arXiv preprint arXiv:2410.24185, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.249, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[29] Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. SuperPADL: Scaling language-directed physics-based control with progressive supervised distillation. In SIGGRAPH, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.305, + 0.483, + 0.359 + ], + "angle": 0, + "content": "[30] Hyeonwoo Kim, Sangwon Beak, and Hanbyul Joo. DAViD: Modeling dynamic affordance of 3d objects using pre-trained video diffusion models. arXiv preprint arXiv:2501.08333, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[31] Jeonghwan Kim, Jisoo Kim, Jeonghyeon Na, and Hanbyul Joo. ParaHome: Parameterizing everyday home activities towards 3d generative modeling of human-object interactions. arXiv preprint arXiv:2401.10232, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[32] Yeonjoon Kim, Hangil Park, Seungbae Bang, and Sung-Hee Lee. Retargeting human-object interaction to virtual avatars. IEEE transactions on visualization and computer graphics, 22(11):2405-2412, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.475, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[33] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[34] Dieter Kraft. Algorithm 733: Tomp-fortran modules for optimal control calculations. ACM Transactions on Mathematical Software (TOMS), 20(3):262-281, 1994. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.547, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[35] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. NIFTY: Neural object interaction fields for guided human motion synthesis. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[36] Jiye Lee and Hanbyul Joo. Locomotion-Action-Manipulation: Synthesizing human-scene interactions in complex 3d environments. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[37] Kang Hoon Lee, Myung Geol Choi, and Jehee Lee. Motion patches: building blocks for virtual environments annotated with motion data. In SIGGRAPH. 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[38] Yoonsang Lee, Sungeun Kim, and Jehee Lee. Data-driven biped control. In SIGGRAPH. 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[39] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[40] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[41] Jinhan Li, Yifeng Zhu, Yuqi Xie, Zhenyu Jiang, Mingyo Seo, Georgios Pavlakos, and Yuke Zhu. OKAMI: Teaching humanoid robots manipulation skills through single video imitation. In CoRL, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[42] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. In WACV, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[43] Fukang Liu, Zhaoyuan Gu, Yilin Cai, Ziyi Zhou, Shijie Zhao, Hyunyoung Jung, Sehoon Ha, Yue Chen, Danfei Xu, and Ye Zhao. Opt2skill: Imitating dynamically-feasible whole-body trajectories for versatile humanoid loco-manipulation. arXiv preprint arXiv:2409.20514, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.176, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[44] Libin Liu and Jessica Hodgins. Learning to schedule control fragments for physics-based characters using deep q-learning. ACM Transactions on Graphics (TOG), 36(3): 1-14, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.232, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[45] Yunze Liu, Changxi Chen, Chenjing Ding, and Li Yi. Phys-Reaction: Physically plausible real-time humanoid reaction synthesis via forward dynamics guided 4d imitation. In ACMMM, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[46] Yun Liu, Bowen Yang, Licheng Zhong, He Wang, and Li Yi. Mimicking-bench: A benchmark for generalizable humanoid-scene interaction learning via human mimicking. arXiv preprint arXiv:2412.17730, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.344, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[47] Yun Liu, Haolin Yang, Xu Si, Ling Liu, Zipeng Li, Yuxiang Zhang, Yebin Liu, and Li Yi. Taco: Benchmarking generalizable bimanual tool-action-object understanding. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.4, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[48] Yun Liu, Chengwen Zhang, Ruofan Xing, Bingda Tang, Bowen Yang, and Li Yi. Core4d: A 4d human-object-human interaction dataset for collaborative object rearrangement. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.456, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[49] Jintao Lu, He Zhang, Yuting Ye, Takaaki Shiratori, Sebastian Starke, and Taku Komura. CHOICE: Coordinated human-object interaction in cluttered environments for pick-and-place actions. arXiv preprint arXiv:2412.06702, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.525, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[50] Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In ICCV, 2023. 2, 4, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.567, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[51] Zhengyi Luo, Jinkun Cao, Sammy Christen, Alexander Winkler, Kris Kitani, and Weipeng Xu. Grasping diverse objects with simulated humanoids. In NeurIPS, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.609, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[52] Zhengyi Luo, Jiashun Wang, Kangni Liu, Haotian Zhang, Chen Tessler, Jingbo Wang, Ye Yuan, Jinkun Cao, Zihui Lin, Fengyi Wang, et al. SMPLOlympics: Sports environments for physically simulated humanoids. arXiv preprint arXiv:2407.00187, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[53] Xintao Lv, Liang Xu, Yichao Yan, Xin Jin, Congsheng Xu, Shuwen Wu, Yifan Liu, Lincheng Li, Mengxiao Bi, Wenjun Zeng, et al. HIMO: A new benchmark for full-body human interacting with multiple objects. In ECCV, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[54] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. In NeurIPS, 2021. 6, 7, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.803, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[55] Josh Merel, Saran Tunyasuvunakool, Arun Ahuja, Yuval Tassa, Leonard Hasenclever, Vu Pham, Tom Erez, Greg Wayne, and Nicolas Heess. Catch & carry: reusable neural controllers for vision-guided whole-body tasks. ACM Transactions on Graphics (TOG), 39(4):39-1, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[56] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12275" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.135, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[57] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 3DV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[58] Liang Pan, Zeshi Yang, Zhiyang Dou, Wenjia Wang, Buzhen Huang, Bo Dai, Taku Komura, and Jingbo Wang. TokenHSI: Unified synthesis of physical human-scene interactions through task tokenization. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.247, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[59] Soohwan Park, Hoseok Ryu, Seyoung Lee, Sunmin Lee, and Jehee Lee. Learning predict-and-simulate policies from unorganized human motion data. ACM Transactions on Graphics (TOG), 38(6):1-11, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[60] Sungjae Park, Seungho Lee, Mingi Choi, Jiye Lee, Jeonghwan Kim, Jisoo Kim, and Hanbyul Joo. Learning to transfer human hand skills for robot manipulations. arXiv preprint arXiv:2501.04169, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[61] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[62] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. HOI-Diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 3, 7, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.483, + 0.537 + ], + "angle": 0, + "content": "[63] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG), 37(4):1-14, 2018. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.539, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[64] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4):1-20, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.596, + 0.483, + 0.65 + ], + "angle": 0, + "content": "[65] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Transactions On Graphics (TOG), 41(4):1-17, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.651, + 0.483, + 0.692 + ], + "angle": 0, + "content": "[66] Haziq Razali and Yiannis Demiris. Action-conditioned generation of bimanual object manipulation sequences. In AAAI, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.693, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[67] Daniele Reda, Jungdam Won, Yuting Ye, Michiel van de Panne, and Alexander Winkler. Physics-based motion retargeting from sparse inputs. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 6(3):1-19, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.763, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[68] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[69] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[70] John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation. In ICLR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.151, + 0.905, + 0.193 + ], + "angle": 0, + "content": "[71] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.195, + 0.905, + 0.249 + ], + "angle": 0, + "content": "[72] Carmelo Sferrazza, Dun-Ming Huang, Xingyu Lin, Youngwoon Lee, and Pieter Abbeel. Humanoidbench: Simulated humanoid benchmark for whole-body locomotion and manipulation. In RSS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.252, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[73] Wenfeng Song, Xinyu Zhang, Shuai Li, Yang Gao, Aimin Hao, Xia Hou, Chenglizhao Chen, Ning Li, and Hong Qin. HOIAnimator: Generating text-prompt human-object animations using novel perceptive diffusion models. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.324, + 0.905, + 0.365 + ], + "angle": 0, + "content": "[74] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.368, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[75] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39(4):54-1, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.425, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[76] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.47, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[77] Omid Taheri, Vasileios Choutas, Michael J Black, and Dimitrios Tzionas. GOAL: Generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.513, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[78] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. Calm: Conditional adversarial latent models for directable virtual characters. In SIGGRAPH, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.571, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[79] Chen Tessler, Yunrong Guo, Ofir Nabati, Gal Chechik, and Xue Bin Peng. Maskedmimic: Unified physics-based character control through masked motion inpainting. ACM Transactions on Graphics (TOG), 43(6):1-21, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.629, + 0.905, + 0.696 + ], + "angle": 0, + "content": "[80] Guy Tevet, Sigal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. CLoSD: Closing the loop between simulation and diffusion for multi-task character control. In ICLR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.7, + 0.904, + 0.728 + ], + "angle": 0, + "content": "[81] Unittree. Unittree g1 humanoid agent ai avatar. https://www-unitree.com/g1/.1,4,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.731, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[82] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.788, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[83] Ruben Villegas, Duygu Ceylan, Aaron Hertzmann, Jimei Yang, and Jun Saito. Contact-aware retargeting of skinned motion. In ICCV, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[84] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12276" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[85] Jiashun Wang, Jessica Hodgins, and Jungdam Won. Strategy and skill learning for physics-based table tennis animation. In SIGGRAPH, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[86] Ruocheng Wang, Pei Xu, Haochen Shi, Elizabeth Schumann, and C Karen Liu. Fürelise: Capturing and physically synthesizing hand motion of piano performance. In SIGGRAPH Asia, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.192, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[87] Wenjia Wang, Liang Pan, Zhiyang Dou, Zhouyingcheng Liao, Yuke Lou, Lei Yang, Jingbo Wang, and Taku Komura. SIMS: Simulating human-scene interactions with real world script planning. arXiv preprint arXiv:2411.19921, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.263, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[88] Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. PhysHOI: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393, 2023. 3, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.32, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[89] Yinhuai Wang, Qihan Zhao, Runyi Yu, Ailing Zeng, Jing Lin, Zhengyi Luo, Hok Wai Tsui, Jiwen Yu, Xiu Li, Qifeng Chen, et al. SkillMimic: Learning reusable basketball skills from demonstrations. In CVPR, 2025. 3, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.377, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[90] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. InterControl: Generate human motion interactions by controlling every joint. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.42, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[91] Jungdam Won and Jehee Lee. Learning body shape variation in physics-based characters. ACM Transactions on Graphics (TOG), 38(6):1-12, 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[92] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. A scalable approach to control diverse behaviors for physically simulated characters. ACM Transactions on Graphics (TOG), 39(4):33-1, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[93] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. THOR: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[94] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. SAGA: Stochastic whole-body grasping with contact. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[95] Zhen Wu, Jiaman Li, Pei Xu, and C Karen Liu. Human-object interaction from human-level instructions. arXiv preprint arXiv:2406.17840, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[96] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[97] Xianghui Xie, Jan Eric Lenssen, and Gerard Pons-Moll. InterTrack: Tracking human object interaction without object templates. In 3DV, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[98] Zhaoming Xie, Sebastian Starke, Hung Yu Ling, and Michiel van de Panne. Learning soccer juggling skills with layer-wise mixture-of-experts. In SIGGRAPH, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[99] Zhaoming Xie, Jonathan Tseng, Sebastian Starke, Michiel van de Panne, and C Karen Liu. Hierarchical planning and control for box loco-manipulation. arXiv preprint arXiv:2306.09532, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.86, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[100] Pei Xu and Ruocheng Wang. Synchronize dual hands for physics-based dexterous guitar playing. In SIGGRAPH Asia, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[101] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 1, 3, 7, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[102] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. In NeurIPS, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[103] Sirui Xu, Dongting Li, Yucheng Zhang, Xiyan Xu, Qi Long, Ziyin Wang, Yunzhi Lu, Shuchang Dong, Hezi Jiang, Akshit Gupta, Yu-Xiong Wang, and Liang-Yan Gui. InterAct: Advancing large-scale versatile 3d human-object interaction generation. In CVPR, 2025. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[104] Zeshi Yang, Kangkang Yin, and Libin Liu. Learning to use chopsticks in diverse gripping styles. ACM Transactions on Graphics (TOG), 41(4):1-17, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[105] Wenhao Yu, Greg Turk, and C Karen Liu. Learning symmetric and low-energy locomotion. ACM Transactions on Graphics (TOG), 37(4):1-12, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[106] Ye Yuan and Kris Kitani. Residual force control for agile human behavior imitation and extended motion synthesis. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[107] Yanjie Ze, Zixuan Chen, Wenhao Wang, Tianyi Chen, Xiaolin He, Ying Yuan, Xue Bin Peng, and Jiajun Wu. Generalizable humanoid manipulation with improved 3d diffusion policies. arXiv preprint arXiv:2410.10803, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[108] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[109] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. NeuralDome: A neural modeling pipeline on multi-view human-object interactions. In CVPR, 2023. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[110] Juze Zhang, Jingyan Zhang, Zining Song, Zhanhe Shi, Chengfeng Zhao, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Hoi-m^3: Capture multiple humans and objects interaction within contextual environment. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.646, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[111] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. COUCH: Towards controllable human-chair interactions. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.688, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[112] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Ilya Petrov, Vladimir Guzov, Helisa Dhamo, Eduardo Pérez-Pellitero, and Gerard Pons-Moll. FORCE: Dataset and method for intuitive physics guided human-object interaction. In 3DV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.759, + 0.905, + 0.812 + ], + "angle": 0, + "content": "[113] Yunbo Zhang, Deepak Gopinath, Yuting Ye, Jessica Hodgins, Greg Turk, and Jungdam Won. Simulation and targeting of complex multi-character interactions. In SIGGRAPH, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.816, + 0.905, + 0.869 + ], + "angle": 0, + "content": "[114] Chengfeng Zhao, Juze Zhang, Jiashen Du, Ziwei Shan, Junye Wang, Jingyi Yu, Jingya Wang, and Lan Xu. I'M HOI: Inertia-aware monocular capture of 3d human-object interactions. In CVPR, 2024. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.093, + 0.905, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12277" + } + ] +] \ No newline at end of file diff --git a/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/c17df58e-4f4a-4403-b950-698942e97f1f_origin.pdf b/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/c17df58e-4f4a-4403-b950-698942e97f1f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..75c3a4021e7cb509e8da0c3c545372384e80df65 --- /dev/null +++ b/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/c17df58e-4f4a-4403-b950-698942e97f1f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f30248f5ae95ff10c6ee1dec6fa6b8bd3df2fd21084e67a53a351cdc991d477 +size 8215527 diff --git a/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/full.md b/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5dcb44148e570ba1e6b47be94b84a9ced754ef99 --- /dev/null +++ b/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/full.md @@ -0,0 +1,353 @@ +![](images/a1a2a0a3a33404ae530a2680deabf1baebe50111a29d9931bac3a9f91e9263e2.jpg) + +# INTERMIMIC: Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions + +Sirui Xu $^{1}$ Hung Yu Ling $^{2}$ Yu-Xiong Wang $^{1\dagger}$ Liang-Yan Gui $^{1\dagger}$ + +1 University of Illinois Urbana-Champaign 2 Electronic Arts Equal Advising + +https://sirui-xu.github.io/InterMimic + +![](images/9f6d3db18762132dbe4b9601ef951b5dda1f69372b6152f3831f8dbf55886183.jpg) +Highly-Dynamic Interaction +Multi-Object Interaction +Scalable Interaction Skills + +![](images/f14c8ec6592615be708d340bdc035cd2408313cb1f97916107a9bc0c28a060d3.jpg) +Humanoid Robot Interaction +MoCap Correction +Interaction Prediction +Lift the woodchair. +Text-to-Interaction +Figure 1. InterMimic enables physically simulated humans to perform interactions with dynamic and diverse objects. It supports highly-dynamic, multi-object interactions and scalable skill learning (Top), making it adaptable for versatile downstream applications (Bottom): it can translate whole-body loco-manipulation skills to a humanoid robot [24, 81], perfect interaction MoCap data, and bridge kinematic generation, e.g., predicting future interactions from past (InterDiff [101]) or generating interactions given text prompts (InterDreamer [102]). + +# Abstract + +Achieving realistic simulations of humans interacting with a wide range of objects has long been a fundamental goal. Extending physics-based motion imitation to complex human-object interactions (HOIs) is challenging due to intricate human-object coupling, variability in object geometries, and artifacts in motion capture data, such as inaccurate contacts and limited hand detail. We introduce InterMimic, a framework that enables a single policy to robustly learn from hours of imperfect MoCap data covering diverse full-body interactions with dynamic and varied objects. Our key insight is to employ a curriculum strategy – perfect first, then scale up. We first train subject-specific teacher policies to mimic, retarget, and refine motion capture data. Next, we distill these teachers into a student policy, with the teachers acting as online experts providing direct supervision. + +as well as high-quality references. Notably, we incorporate RL fine-tuning on the student policy to surpass mere demonstration replication and achieve higher-quality solutions. Our experiments demonstrate that InterMimic produces realistic and diverse interactions across multiple HOI datasets. The learned policy generalizes in a zero-shot manner and seamlessly integrates with kinematic generators, elevating the framework from mere imitation to generative modeling of complex human-object interactions. + +# 1. Introduction + +Animating human-object interactions is a challenging and time-consuming task even for skilled animators. It requires a deep understanding of physics and meticulous attention to detail to create natural and convincing interactions. While + +Motion Capture (MoCap) data provides references, animators often need to correct contact errors caused by sensor limitations and occlusions between humans and objects. However, this process remains unscalable, as refining a single motion demands a delicate balance between preserving the captured data and ensuring its physical plausibility. + +Physics-based human motion imitation [38, 63] offers an alternative approach to improving motion fidelity, by training control policies to mimic reference MoCap data within a physics simulator. However, scaling up human-object interaction imitation presents significant challenges: (i) MoCap Imperfection: Contact artifacts are common, causing expected contacts to fluctuate instead of maintaining consistent zero distance, often due to MoCap limitations or missing hand capture [3, 39]. Accurately imitating MoCap kinematics can result in unrealistic dynamics in simulation. Moreover, HOI datasets often include diverse human shapes, requiring motion retargeting to adapt movements across different human models while preserving interaction dynamics. This retargeting process is imperfect and can introduce new contact artifacts or exacerbate existing ones. (ii) Scaling-up: Although large-scale motion imitation has been explored in previous works [50, 79, 92, 106], it remains largely underexplored for whole-body interactions involving dynamic and diverse objects. + +In this paper, we aim to utilize rich yet imperfect motion capture interaction datasets to train a control policy capable of learning diverse motor skills while enhancing the plausibility of these actions by correcting errors, such as inaccurate hand motions and faulty contacts. Our approach is grounded on the key insight of tackling the challenges of skill perfection and skill integration progressively. We implement a curriculum-based teacher-student distillation framework, where multiple teacher policies focus on imitating and refining small subsets of interactions, and a student policy integrates these skills from the teachers. + +Instead of relying on curated data that covers a limited range of actions [4, 51], we employ multiple teacher policies trained on a diverse set of imperfect interaction data and address two key challenges: retargeting and recovering. First, we unify all training policies to a canonical human model, by embedding HOI retargeting directly into the imitation. This is achieved by reframing the policy learning to optimize both imitation and retargeting objectives. Second, our teacher policies refine interaction motion through learning from it, as accurate contact dynamics enforced by a physics simulator inherently correct inaccuracies in the reference kinematics. To support this, we introduce tailored contact-guided reward and optimize trajectory collection, enabling effective skill imitation despite MoCap errors. + +Introducing teacher policies offers several key benefits. By leveraging teacher rollouts, we effectively distill raw MoCap data into refined HOI references with a unified em + +bodiment and enhanced physical fidelity. These refined references guide the subsequent student policy training, reducing the negative impact of errors in the original MoCap data. A major hurdle in scaling motion imitation is the sample inefficiency of Reinforcement Learning (RL), which can lead to prohibitively long training times. Our teacher-student approach mitigates this through a space-time trade-off: multiple teacher policies are trained in parallel on smaller, more manageable data subsets, and their expertise is then distilled into a single student policy. We begin with demonstration-based distillation to bootstrap PPO [71] updates, reducing reliance on pure trial and error and enabling more effective scaling. As training progresses, the student gradually shifts from heavy demonstration guidance to increased RL updates, ultimately surpassing simple demonstration memorization. This mirrors alignment strategies in Large Language Models (LLMs), where demonstration-based pretraining is refined through RL fine-tuning [56]. + +To summarize, our contributions are as follows: (i) We introduce InterMimic, which, to the best of our knowledge, is the first framework designed to train physically simulated humans to develop a wide range of whole-body motor skills for interacting with diverse and dynamic objects, extending beyond traditional grasping tasks. (ii) We develop a teacher-student training strategy, where teacher policies provide a unified solution to address the challenges of retargeting and refining in HOI imitation. The student distillation introduces a scalable solution by leveraging a space-time tradeoff. (iii) We demonstrate that our unified framework, InterMimic, as illustrated in Figure 1, effectively handles versatile physics-based interaction animation, recovering motions with realistic and physically plausible details. Notably, by combining kinematic generators with InterMimic, we enable a physics-based agent to achieve tasks such as interaction prediction and text-to-interaction generation. + +# 2. Related Work + +Significant progress has been made in physics-based human interaction animation and control, with advancements in areas such as human-human interactions [45, 90], hand-object interactions [60, 86, 100, 104], human interactions with static scenes [5, 36, 57, 79, 96], and real-world humanoid control for object manipulation [2, 11, 16, 19, 20, 28, 41, 43, 72, 107]. Among these areas, we are the first to achieve universal whole-body loco-manipulation simulation with diverse dynamic objects, beyond pick-and-place and grasping actions – a novel achievement in animation and an idealized reference for real-world humanoid control. Below, we elaborate on recent studies on whole-body interaction animation, particularly involving dynamic objects. + +# 2.1. Kinematic Interaction Animation + +Generating human interactions has been a long-standing topic in animation and computer graphics [15, 37]. Significant advances in character animation have emerged with the advent of deep learning, e.g., including phase-function-based methods [22] that enable object interactions like carrying a box [74] or playing basketball [75]. This is extended to more diverse but static objects approaching [35, 77, 94, 111]. Subsequent efforts [14, 26, 27, 40, 42, 49, 66, 95] integrate object motion into interactions but remain constrained by assuming that interactions occur primarily through the hands. To address this, recent developments [7, 10, 12, 21, 30, 62, 73, 84, 93, 101, 102] introduce interactions in a fashion of whole-body loco-manipulation that engages multiple body parts in contact. However, these methods often suffer from physical inaccuracies, such as floating contacts and penetrations, while they generate only body motion without considering hand dexterity. In this work, we address physical inaccuracies by refining imperfect kinematic generation through physics simulation, with InterDiff [101] and HOI-Diff [62] serving as motion planning for loco-manipulation that bridges high-level decision-making (e.g., text instruction) with low-level execution. + +# 2.2. Physics-based Interaction Animation + +Physics-based methods generate motion through motor control policies within a physics simulator, e.g., achieved via deep reinforcement learning to track reference motions [63]. These policies are directly applicable for executing simple interactions, such as punching or striking an object [8, 65, 78, 80]. To achieve more complex interactions, early studies focus on specific scenarios, including notable sports-related [52] examples such as basketball [88], skating [44], soccer [98], tennis [108], table tennis [85], and more proposed in [1]. Research also demonstrates flexibility in more general but simpler box carrying tasks [58, 87, 113]. These advancements are achieved through the integration of multiple control policies [55], the use of adversarial motion priors [13, 18, 64], and imitating diverse kinematic generations [95, 99]. However, these methods train their policies in a non-scalable manner, with each policy handling only specific object types or actions. In pursuit of a single, scalable policy to enable multiple interaction skills, existing methods either rely on fixed interaction patterns, such as approaching and grasping objects [4, 51], or extend single-object skills, e.g., interactions involving a basketball [89]. Additionally, they mostly depend on highly curated data from the GRAB dataset [76], which, despite its high quality, primarily features low-dynamic full-body motion and only small-sized objects. More recent datasets [3, 23, 25, 31, 39, 46-48, 53, 97, 103, 109, 110, 112, 114] offer richer full-body interactions with diverse objects but contain noticeable artifacts that chal + +![](images/f5d9c41cf08c4f0fd5f8d0add65355f1e232146e94c43d345381ec12fa00146f.jpg) +Figure 2. Our two-stage pipeline: (i) training each teacher policy (MLP) on a small data subset with initialization corrected via Physical State Initialization (PSI), and (ii) freezing the teacher policies to provide refined references for training a student policy (Transformer). The student leverages teacher supervision for effective scaling and is fine-tuned through RL. + +lenge existing motion imitation approaches. We process data from OMOMO [39], BEHAVE [3], HODome [109], and IMHD [114] collected in the InterAct [103] dataset, and the multi-object dataset HIMO [53], highlighting InterMimic's scalability to diverse interactions and its robustness against MoCap artifacts. + +# 3. Methodology + +Task Formulation. The goal of human-object interaction (HOI) imitation is to learn a policy $\pi$ that produces simulated human-object motion $\{\pmb{q}_t\}_{t=1}^T$ closely matching a ground-truth reference $\{\hat{\pmb{q}}_t\}_{t=1}^T$ derived from large-scale MoCap data. Given the geometries of the human and objects, the policy should also compensate for missing or inaccurate details in the dataset. Each pose $\pmb{q}_t$ has two components: the human pose $\pmb{q}_t^h$ and the object pose $\pmb{q}_t^o$ . The human pose is defined as $\pmb{q}_t^h = \{\pmb{\theta}_t^h, \pmb{p}_t^h\}$ , where $\pmb{\theta}_t^h \in \mathbb{R}^{52 \times 3}$ represents the joint rotations, and $\pmb{p}_t^h \in \mathbb{R}^{52 \times 3}$ specifies the joint positions. Specifically, our human model includes 30 hand joints and 22 joints for the rest of the body, with one root joint's rotation and position specified in global coordinates, whereas the rotations and positions of all other joints are defined relative to their respective parent joints' coordinate systems. The object pose $\pmb{q}_t^o$ is represented as $\{\pmb{\theta}_t^o, \pmb{p}_t^o\}$ , where $\pmb{\theta}_t^o \in \mathbb{R}^3$ denotes the object's orientation and $\pmb{p}_t^o \in \mathbb{R}^3$ the position. All simulation states have corresponding ground-truth values, denoted by the hat symbol. For instance, the reference object rotation is $\{\hat{\pmb{\theta}}_t^o\}_{t=1}^T$ . The environmental setup for the simulation is detailed in Sec. B. Overview. We formulate interaction imitation as a Markov Decision Process (MDP), defined by states, actions, simulator-provided transition dynamics, and a reward function. Figure 2 illustrates our two-stage framework: (i) training teacher policies $\pi^{(T)}$ on small skill subsets, and (ii) distilling these teachers into a scalable student policy $\pi^{(S)}$ for large-scale skill learning. In Sec. 3.1, we define the states $s_t$ + +and actions $\pmb{a}_t$ , applicable to both teacher $\pi^{(T)}$ and student $\pi^{(S)}$ policies. In Sec. 3.2, we describe how teacher policies are trained via RL, focusing on reward designs that facilitate retargeting, as well as techniques that mitigate the impact of imperfections in the reference data. Sec. 3.3 details the subsequent distillation of teachers into a scalable student policy, leveraging both RL and learning from demonstration. + +# 3.1. Policy Representation + +State. The state $s_t$ , which serves as input to the policy, comprises two components $s_t = \{s_t^s, s_t^g\}$ . The first part, $s_t^s$ , contains human proprioception and object observations, expressed as, $\{\{\pmb{\theta}_t^h, \pmb{p}_t^h, \pmb{\omega}_t^h, \pmb{v}_t^h\}, \{\pmb{\theta}_t^o, \pmb{p}_t^o, \pmb{\omega}_t^o, \pmb{v}_t^o\}, \{d_t, c_t\}\}$ where $\{\pmb{\theta}_t^h, \pmb{p}_t^h, \pmb{\omega}_t^h, \pmb{v}_t^h\}$ represent the rotation, position, angular velocity, and velocity of all joints, respectively, while $\{\pmb{\theta}_t^o, \pmb{p}_t^o, \pmb{\omega}_t^o, \pmb{v}_t^o\}$ represent the orientation, location, velocity, and angular velocity of the object, respectively. Motivated by [6], we include object geometry and whole-body haptic sensing from two elements: (i) $d_t$ , vectors from human joints to their nearest points on each object surface; and (ii) $c_t$ , contact markers indicating whether the human's rigid body parts experience applied forces; this serves as simplified tactile or force sensing - an important multimodal input in robot manipulation tasks [9]. The goal state $s_t^g = \{s_{t,t+k}^g\}_{k \in K}$ integrates reference poses from the ground truth motion, where $s_{t,t+k}^g$ is defined as, + +$$ +\{\{\hat {\boldsymbol {\theta}} _ {t + k} ^ {h} \ominus \boldsymbol {\theta} _ {t} ^ {h}, \hat {\boldsymbol {p}} _ {t + k} ^ {h} - \boldsymbol {p} _ {t} ^ {h} \}, \{\hat {\boldsymbol {\theta}} _ {t + k} ^ {o} \ominus \boldsymbol {\theta} _ {t} ^ {o}, \hat {\boldsymbol {p}} _ {t + k} ^ {o} - \boldsymbol {p} _ {t} ^ {o} \}, \tag {1} +$$ + +$$ +\{\hat {\boldsymbol {d}} _ {t + k} - \boldsymbol {d} _ {t}, \hat {\boldsymbol {c}} _ {t + k} - \boldsymbol {c} _ {t} \}, \{\hat {\boldsymbol {\theta}} _ {t + k} ^ {h}, \hat {\boldsymbol {p}} _ {t + k} ^ {h}, \hat {\boldsymbol {\theta}} _ {t + k} ^ {o}, \hat {\boldsymbol {p}} _ {t + k} ^ {o} \} \}, +$$ + +where $\hat{\pmb{\theta}}_{t + k}^{h},\hat{\pmb{p}}_{t + k}^{h},\hat{\pmb{d}}_{t + k},\hat{\pmb{c}}_{t + k}$ represent the reference information at time step $t + k,\ominus$ denotes the calculation of rotation difference. All continuous elements of $s_t$ are normalized relative to the current direction of view of the human and the position of the root [63]. + +Given that most MoCap data does not provide reference contact or tactile information, we extract reference contact markers $\hat{c}_{t + k}$ by inferring dynamic information, beyond relying solely on inaccurate contact distances, specifically by analyzing the object's acceleration to detect human-induced forces. To accommodate the variability in contact distances observed in reference motion, we discretize reference contact markers using varying distance thresholds, as illustrated in Fig. 3(i). The neutral areas serve as buffer zones, avoiding the penalization or enforcement of strict contact. See Sec. C of supplementary for details. + +Action. Our human model has 51 actuated joints, defining an action space of $\mathbf{a}_t \in \mathbb{R}^{51 \times 3}$ . These actions are specified as joint PD targets using the exponential map and are converted into torques applied to each of the human joints. + +# 3.2. Imitation as Perfecting + +The teacher policy $\pi^{(\mathrm{T})}$ is trained via RL to maximize the expected discounted reward by comparing simulated states against potentially erroneous reference states. The training involves: (i) trajectory collection, where we explain how trajectories are initialized and terminated. (ii) policy updating, where collected trajectories and their associated rewards are used to refine the policy. In this section, we elaborate on our reward design and how we optimize trajectory collection to mitigate the impact of reference inaccuracies. + +Imitation as Retargeting. We tailor teacher policies to each human subject, while all policies share the same base human model. This serves the retargeting purpose by converting HOIs from different human shapes into a unified base shape. Although motion imitation does not necessarily require a unified human model [50, 91], our approach offers two benefits: (i) It enhances integration with kinematic generation methods, which generally perform better on a single, unified shape [17]. (ii) It demonstrates possible integration with real-world humanoid deployment, which requires retargeting to a consistent physical embodiment. In Figure 1, our method translates MoCap data into motor skills on a Unitree G1 [81] with two Inspire hands [24], all without external retargeting in complex contact-rich scenarios. See Sec. F of the supplementary for additional details. + +Human [83] or HOI [32] retargeting can be formulated as an optimization problem. Inverse Kinematics (IK) methods, such as those based on quadratic programming [34], demonstrate effectiveness in simplified scenarios but remain underexplored for motions featuring intricate object interactions. RL, by contrast, solves the optimization by maximizing an expected cumulative reward, prompting us to investigate whether RL-driven HOI imitation can be used for HOI retargeting. This extends existing physics-based retargeting approaches, which either omit object interactions [67] or are non-scalable with a single reference [113]. + +While the kinematics should differ due to the embodiment gap, we argue that the dynamics between human and object should remain invariant. Thus, we define rewards to include an embodiment-aware component that loosely aligns the simulated kinematics with the reference interaction, and an embodiment-agnostic reward component that encourages dynamics to be close to the reference. + +Embodiment-Aware Reward. When the human and object are far apart, retargeting should prioritize capturing rotational motion, whereas when they are close, accurate position tracking becomes crucial for achieving contact. To reflect this, we define the weights $\pmb{w}_d$ that are inversely proportional to the distances between joints and the object [113]. The reward thus includes cost functions for joint position $E_p^h = \langle \pmb{\Delta}_p^h, \pmb{w}_d \rangle$ , rotation $E_\theta^h = \langle \pmb{\Delta}_\theta^h, \mathbf{1} - \pmb{w}_d \rangle$ , and interaction tracking $E_d = \langle \pmb{\Delta}_d, \pmb{w}_d \rangle$ , where $\langle \cdot, \cdot \rangle$ is the inner product, $\pmb{\Delta}_p^h[i] = \| \hat{\pmb{p}}^h[i] - \pmb{p}^h[i] \|$ , $\pmb{\Delta}_\theta^h[i] =$ + +![](images/111fbf92abb4b6809bda6221ceaa38a3758d1c1710c4dd9b869544587fabd65b.jpg) +(i) Reference contact markers +Figure 3. (i) Visualization of reference contact markers that accommodate varied contact distances: red to promote contact, green for neutral areas where contact is neither promoted nor penalized, and blue to penalize contact. (ii) Initializing the rollout with reference (RSI) or reference corrected via simulation (PSI). + +![](images/e2e0019fc9d39d2746e6af3c0045298378d8b7396113fc57417d2b0089f67cd7.jpg) +(ii) RSI vs. PSI + +![](images/610b8a83552a18149b470ea8f89665350e244f4dbba2121a0575a3857b8fb085.jpg) + +$\| \hat{\pmb{\theta}}^h [i]\ominus \pmb{\theta}^h [i]\|$ , and $\pmb{\Delta}_d[i] = \| \hat{\pmb{d}} [i] - \pmb {d}[i]\|$ represent the displacement for the variables defined in Sec. 3.1 with timestep $t$ omitted. The formulation of $\pmb{w}_{d}$ is provided in Sec. D of supplementary. The reward to be maximized can be formulated as $\exp (-\lambda E)$ for each cost function $E$ with a specific hyperparameter $\lambda$ . Details can be found in Sec. D. + +Embodiment-Agnostic Reward. The reward includes components for object tracking and contact tracking. The object tracking cost is defined for position $E_{p}^{o} = \|\hat{\boldsymbol{p}}^{o} - \boldsymbol{p}^{o}\|$ and rotation $E_{\theta}^{o} = \|\hat{\boldsymbol{\theta}}^{o} - \boldsymbol{\theta}^{o}\|$ , with all values normalized to the human's current position and direction. The contact tracking reward comprises two cost functions: body contact promotion $E_{b}^{c}$ and penalty $E_{p}^{c}$ , both aligning the simulated contact $c$ with reference markers $\hat{c}$ , as shown in Figure 3. We define three contact levels – promotion, penalty, and neutral – to accommodate potential inaccuracies in reference contact distances. The detailed formulation can be found in Sec. D of the supplementary. Since the physics engine does not differentiate between object, ground, and self-contact, we adopt two strategies: (i) we model foot-ground contact promotion and penalty. This ensures proper foot lifting during cyclic walking and mitigates foot hobbling. (ii) We allow self-collision to avoid self-contact promotion but to promote object interaction. This poses minimal risk as the policy is guided by MoCap reference, which, although lacking perfect contact accuracy, rarely shows self-penetration. For humanoid robots with embodiments that differ from the MoCap reference and require real-world applicability, we disable self-collision, as discussed in Sec. F. + +We introduce energy consumption rewards [105] to penalize large human or object jitters, with a proposed contact energy penalizing abrupt contact to promote compliant interactions. See Sec. D of supplementary for more details. + +Hand Interaction Discovery. We use data with average or flattened hand poses [3, 39], which makes accurate object manipulation imitation challenging. To address this, we activate a reference contact marker for any hand part when a fingertip or palm is near an object. Given tasks that do not demand high dexterity, employing a contact-promoting reward with this marker enables policies to develop effective + +hand interaction strategies, leveraging the exploratory nature of RL. Additionally, we constrain the range of motion (RoM) of the hands to ensure natural movement. See Sec. D and Sec. B of the supplementary for further details. + +Policy Learning. Following [63], the control policy $\pi$ is trained using PPO [71] with the policy gradient $L(\psi) = \mathbb{E}_t[\min (r_t(\psi)A_t,\mathrm{clip}(r_t(\psi),1 - \epsilon ,1 + \epsilon)A_t)$ . $\psi$ are the parameters of $\pi$ and $r_t(\psi)$ quantifies the difference in action likelihoods between updated and old policies. $\epsilon$ is a small constant, and $A_{t}$ is the advantage estimate given by the generalized advantage estimator $\mathrm{GAE}(\lambda)$ [70]. + +Physical State Initialization. Learning later-phase motion can be essential for policies to achieve high rewards during earlier phases, compared to incrementally learning from the starting phase. Thus, Reference State Initialization (RSI) [63] sets the current pose $\mathbf{q}_t$ to a reference pose $\hat{\mathbf{q}}_t$ at a random timestep $t$ , for initializing the rollout. However, initializing with the imperfect reference can introduce critical artifacts, such as contact floating or incorrect hand motion, leading to unrecoverable failures, e.g., object falling, as depicted in Figure 3(ii). These issues render many initializations ineffective, limiting training on certain interaction phases since successful rollouts may not reach them before the maximum length. The problem is exacerbated by the use of prioritized sampling [79, 91], which favors high-failure-rate initializations. + +To address the need for higher-quality reference initialization, we propose Physical State Initialization (PSI). As illustrated in Figure 2, PSI begins by creating an initialization buffer that stores reference states from MoCap and simulation states from prior rollouts. For each new rollout, an initial state is randomly selected from this buffer, which increases the likelihood of starting from advantageous positions. Once a rollout is completed, trajectories are evaluated based on their expected discounted rewards; those above a certain threshold are added to the buffer using a first-in-first-out (FIFO) strategy, while older or lower-quality trajectories are discarded. This selective reintroduction of high-value states for initialization helps maintain stable policy updates. We apply PSI in a sparse manner to ensure training efficiency. As shown in Figure 3(ii), PSI can collect trajectories for policy update that RSI does not effectively utilize. Further details are provided in Sec. E of the supplementary. + +Interaction Early Termination. Early Termination (ET) [63] is commonly used in motion imitation, ending an episode when a body part makes unplanned ground contact or when the character deviates significantly from the reference [50], thus stopping the policy from overvaluing invalid transitions. However, additional conditions should be considered for human-object interactions. We propose Interaction Early Termination (IET), which supplements ET with three extra checks: (i) Object points deviate from their references by more than $0.5\mathrm{m}$ on average. (ii) Weighted av + +erage distances between the character's joints and the object surface exceed $0.5\mathrm{m}$ from the reference. (iii) Any required body-object contact is lost for over 10 consecutive frames. Full conditions are detailed in Sec. E of the supplementary. + +# 3.3. Imitation with Distillation + +As shown in Figure 2, after training the teacher policies on data from each subject (Sec. 3.2), we aggregate them to train a student policy $\pi^{(\mathrm{S})}$ to master all skills. As outlined in Algorithm 1, the combined teacher policies, denoted by $\pi^{(\mathrm{T})}$ for brevity, serves dual roles by providing state-action trajectories $(s^{(\mathrm{T})}, a^{(\mathrm{T})})$ : (i) the state $s^{(\mathrm{T})}$ for reference distillation, and (ii) the action $a^{(\mathrm{T})}$ for policy distillation. + +Reference Distillation. Noisy MoCap data can hinder policy learning, especially at larger scales. In contrast, teacher policies trained on smaller-scale data effectively address these issues by correcting contact artifacts, refining hand placements, and recovering missing details (see Figures 1 and 5). To fully leverage teacher policies, we use their rollouts as references for defining the student policy's goal state and reward functions, distinguishing our approach from distillation based on only action output. + +Policy Distillation. We apply distillation on action outputs, which we view as crucial for scaling policies to large datasets. In essence, we trade space for time: teacher policies are trained in parallel on smaller data subsets, allowing the student policy to scale through distillation. Following Algorithm 1, we begin with Behavior Cloning (BC) [29] and use RL fine-tuning to go beyond demonstration memorization, an approach common in LLM alignment [56]. We integrate BC into online policy updates with a staged schedule: we start with DAgger [69] and gradually transition to PPO. Throughout, the critic is continuously trained with the reward from Sec. 3.2. This RL fine-tuning phase is crucial as teacher policies may behave differently when performing similar skills, and simple BC can lead to suboptimal "averaging" behavior, where RL fine-tuning helps the student converge on optimal solutions. + +# 3.4. Architecture + +We set the keyframe indices $K$ (Sec. 3.1, Eq. 1) to $\{1,16\}$ for the teacher policies and $\{1,2,4,16\}$ for the student policy. The broader observation window for the student policy helps it better distinguish different skills with larger-scale data. Teacher policies employ MLPs, common in physics-based animation [63], while the student policy handles higher-dimensional observations, for which MLPs are less effective. Thus, we use a transformer [82] architecture for sequential modeling [79], as shown in Figure 2. + +# 4. Experiments + +We evaluate teacher policies on their ability to imitate imperfect HOI references, and assess the entire teacher- + +Algorithm 1 Distillation with RL Fine-tuning +1: Input: A composite policy $\pi^{(\mathrm{T})}$ integrated from individual teacher policies, student policy parameters $\psi$ , student value function parameters $\phi$ , schedule hyperparameter $\beta$ for DAgger, horizon length $H$ for PPO +2: for $t = 0,1,2,\ldots$ do +3: for $h$ from 1 to $H$ do +4: Sample a variable $u \sim \mathrm{Uniform}(0,1)$ +5: Collect $s^{(\mathrm{T})}, a^{(\mathrm{T})}$ from teacher $\pi^{(\mathrm{T})}$ +6: Obtain the refined reference from $s^{(\mathrm{T})}$ to define $s^{(\mathrm{S})}$ and $r(\cdot)$ , obtain $a^{(\mathrm{S})}$ from $\pi_{\phi}^{(\mathrm{S})}(a^{(\mathrm{S})}|s^{(\mathrm{S})})$ . +7: if $u \leq \frac{t}{\beta}$ then ▷ Use the teacher +8: Given $s^{(\mathrm{S})}$ , execute $a^{(\mathrm{S})}$ , observe $s^{\prime (\mathrm{S})}, r$ +9: else ▷ Use the student +10: Given $s^{(\mathrm{S})}$ , execute $a^{(\mathrm{T})}$ , observe $s^{\prime (\mathrm{S})}, r$ +11: end if +12: Store the transition $(s^{(\mathrm{S})}, s^{\prime (\mathrm{S})}, a^{(\mathrm{S})}, a^{(\mathrm{T})}, r)$ +13: end for +14: Update $\phi$ with TD( $\lambda$ ) +15: Compute PPO objective: $L(\psi)$ +16: Compute $J(\psi) = \| a^{(\mathrm{S})} - a^{(\mathrm{T})}\|$ +17: Compute the weight: $w = \min \left(\frac{t}{\beta}, 1\right)$ +18: Update $\psi$ by gradient: $\nabla_{\psi}(wL(\psi) + (1 - w)J(\psi))$ +19: end for + +student framework for scalability to large-scale data and zero-shot generalization across various scenarios. Additional experiments are provided in Sec. G of supplementary. + +Datasets. We use the following datasets from InterAct [103]: OMOMO [39], BEHAV [3], HODome [109], IMHD [114], and HIMO [53]. OMOMO, containing 15 objects and approximately 10 hours of data, is our primary dataset for evaluating the full teacher-student distillation framework for its scale. We train 17 teacher policies, one per subject, with subject 14 reserved as the test set and the remaining data used for training the student policy. A small portion of data is discarded after teacher imitation due to severe MoCap errors that could not be corrected (see Sec. F and Sec. H of the supplementary). Additional datasets are used to evaluate teacher policies in various MoCap scenarios with different error levels and interaction types. We focus on highly dynamic motions (Figure 1) and interactions involving multiple body parts (Figure 4), while excluding scenarios such as carrying a bag with a strap, since the simulator [54] used lacks full support for soft bodies. + +Metrics. We use the following metrics: (i) Success Rate is defined as the proportion of references that the policy successfully imitates at least once, averaged across all references, while (ii) Duration is the time (in seconds) that the imitation is maintained without triggering the interaction early termination conditions introduced in Sec. 3.2. (iii) + +![](images/df088beb273b5d0cccb0de723203bf713b2d524bdd406705ed2e3150b9f9258b.jpg) +Figure 4. Qualitative comparison between PhysHOI [88] (top), the reference motion (middle) from the BEHAVE [3] dataset, and the interaction refined by our teacher trained on it (bottom). InterMimic faithfully imitates the interactions involving multiple body parts while correcting errors in the original reference. + +Human Tracking Error $(E_h)$ measures the per-joint position error (cm) between the simulated and reference human (excluding hand joints for BEHAV [3] and OMOMO [39] due to inaccuracy). (iv) Object Tracking Error $(E_o)$ measures the per-point position error (cm) between the simulated and reference object. Both errors are averaged over the duration of the imitation in the best-performing trial. + +Baselines. To facilitate fair comparisons, we downgrade our method for teacher policy evaluation to imitate either a single MoCap clip (Figure 4) or multiple clips with a single object (Table 1), enabling direct comparison with PhysHOI [88] and SkillMimic [89] (Sec. 4.1 and 4.2). Due to the lack of established baselines for large-scale HOI imitation, we adapt the following variants for comparison with our student policy (Sec. 4.3): (i) PPO [71] trains an imitation policy from scratch, following [63]. We experiment with both versions, with and without reference distillation; (ii) DAgger [69] distills the student without RL fine-tuning, a process we refer to as policy distillation. + +Implementation Details. The control policies operate at $30\mathrm{Hz}$ and are trained using the Isaac Gym simulator [54]. Teacher policies are implemented as MLPs with hidden layers of sizes 1024, 1024, and 512. The student policy is implemented as a three-layer Transformer encoder with 4 heads, a hidden size of 256, and a feed-forward layer of 512. The critics are also modeled as MLPs with the same architecture as the teacher policies. To integrate the student policy with kinematic generators, including text-to-HOI [62] and future interaction prediction [101], we train these models using reference data distilled by the teacher policies from the OMOMO [39] dataset, following the same train-test split as the student policy training. For the text-to + +![](images/a63d96253bb909ed0735badf9ee59d3011c4179d1bdbe3141f7f0d8f9d0fb973.jpg) +Figure 5. We recover plausible object rotations (bottom) that are challenging for motion capture due to the equivariant geometries of objects, which result in the object sliding on the ground (top). + +
MethodTime↑Eh↓Eo↓
SkillMimic [89]12.27.213.4
InterMimic (Ours) w/o IET40.36.79.9
InterMimic (Ours) w/o PSI36.16.610.2
InterMimic (Ours)42.66.49.2
+ +Table 1. Quantitative comparison between the teacher policy from InterMimic and SkillMimic [89] to imitate data extracted from the BEHAVE [3] dataset involving a single subject interacting with yogamat. We ablate our proposed approach by removing interaction early termination and physical state initialization. + +HOI model, we train it to generate 10 seconds of motion and use 24 generated samples for evaluation, while for future interaction prediction, the model generates 25 future frames given 10 past frames and we use 60 generated samples for evaluation. See Sec. F of the supplementary. + +# 4.1.Quantitative Evaluation + +Table 1 shows that the baseline struggles with MoCap imperfections, e.g., incorrect hand positioning, and thus results in clearly shorter tracking durations. In contrast, our method maintains reference tracking for longer durations and produces interactions that closely match the reference. Table 2 shows that our method consistently outperforms baselines in both training data imitation and out-of-distribution generalization, including interactions from the test set and from kinematic generations. We discuss the effectiveness of specific design choices in Sec. 4.3. + +# 4.2. Qualitative Evaluation + +Figure 4 shows a representative sequence from the experiment in Table 1, illustrating how our teacher policy corrects interactions that PhysHOI [88] fails to track robustly – our method effectively withstands and corrects incorrect hand positioning and floating contacts in the reference. Beyond obvious errors, our method also rectifies the rotation of symmetric objects that MoCap inaccurately depicts as slid + +
PPOReference DistillationPolicy DistillationArchitectureOMOMO-Train [39]OMOMO [39]-TestOMOMO [39]-Test (w × 10)HOI-Diff [62]InterDiff [101]
Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓
✓××MLP23.9101.67.215.69.685.37.516.23.971.27.517.90.00.0--6.711.76.216.4
×✓✓54.5139.97.111.054.3140.27.111.215.591.79.313.74.284.810.19.765.027.47.513.4
✓✓×71.7152.88.912.791.6173.78.513.245.8127.69.114.98.3130.910.113.873.328.96.914.4
✓✓✓90.7168.05.59.795.5173.95.411.962.6140.96.614.512.5121.48.612.175.029.16.213.5
✓✓✓Transformer88.8167.06.010.298.1176.55.911.356.8134.76.613.212.5119.08.512.676.729.36.413.3
+ +Table 2. Quantitative evaluation of large-scale interaction imitation using OMOMO [39], kinematic generations from HOI-Diff [62], and InterDiff [101]. Additionally, we evaluate on test set when objects with weights ten times greater than those used during training. + +![](images/eb3cea12521aa060a57d58b0ef9480c512c9960923375cdf38950155543c5029.jpg) +Figure 6. Zero-shot integration with a text-to-HOI model HOI-Diff [62] (Top), using 'Kick the large box' as the prompt, and an interaction prediction model InterDiff [101] (Bottom), where gray meshes are past states and colored illustrate future generations. + +ing along the ground, shown in Figure 5. Figure 6 presents additional examples that complement Figure 1, demonstrating how our approach integrates with kinematic generators for future interaction prediction and text-to-interaction synthesis. This zero-shot generalization extends to novel objects unseen during training (Figure 7), highlighting the effectiveness of our object geometry and contact-encoded representation, as well as the large-scale training. + +# 4.3. Ablation Study + +Effectiveness of PSI and IET. We conduct an ablation study, as demonstrated in Table 1, comparing the full approach to "Ours w/o PSI". The results validate that Physical State Initialization (PSI) is effective by mitigating inaccuracies in the motion capture data. We also observe reduced effectiveness without our interaction early termination, as training often spends rollouts on irrelevant periods. + +Effectiveness of Reference Distillation. Compared to directly scaling imitation from MoCap with potential imperfections (line 1 in Table 2), using references refined by the teacher policy (line 3) achieves consistently better performance on all metrics. The improvement is even more pronounced on the test set, where, without reference distillation, the policy struggles with unseen shapes, while retargeting by reference distillation eliminates the difficulty. + +Effectiveness of Joint PPO and DAgger Updates. As shown in Table 2, training a policy from scratch (line 3) or relying solely on policy distillation (DAgger, line 2) fails to + +![](images/be2f54312dd20b1d975985e217970ac222668c82f3c43e9a52c986a5ecb9d5ef.jpg) +Figure 7. Zero-shot generalization of our student policy on novel objects from BEHAV [3] and HODome [109]. + +achieve optimal performance. While supervised skill learning lays the groundwork, additional PPO fine-tuning is crucial for resolving conflicts among teacher policies. This is important because our subject-based clustering may not effectively distinguish between different interaction patterns, and ambiguity arises when multiple teachers produce different actions for similar motions. + +Effectiveness of Transformer for Policy Learning. From Table 2, we see that using a Transformer policy (line 5) outperforms MLP-based approaches, particularly on the test set and out-of-distribution cases generated by the kinematic model. We attribute this to the Transformer's inductive bias in sequential modeling and its capacity to incorporate longer-term observations, enabling it to handle complex spatio-temporal dependencies more effectively. + +# 5. Conclusion + +In this work, we introduce a framework for synthesizing realistic human-object interactions that are both physically grounded and generalizable. Unlike previous methods, our approach leverages a rich repository of imperfect MoCap data to facilitate the learning of various interaction skills across a wide variety of objects. To address inaccuracies in the MoCap data, we propose contact-guided rewards and optimize trajectory collection, enabling teacher policies to recover missing physical details in the original data. These teacher policies are used to train student policies within a distillation framework that combines policy distillation and reference distillation, thus enabling efficient skill scaling. Our approach shows zero-shot generalizability, which effectively bridges the gap between imitation and generative capabilities by integrating with kinematic generation. We believe that this framework can be adapted for whole-body loco-manipulation for real-world robots, enabling them to handle objects with human-like dexterity and nuance. + +Acknowledgments. We thank Wei Yang, Yu-Wei Chao, Arsalan Mousavian, Ankur Handa, Samuel Schulter, Morteza Ziyadi, and Xiaohan Fei for valuable discussions. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Amazon-Illinois Center on AI for Interactive Conversational Experiences, the Toyota Research Institute, the IBM-Illinois Discovery Accelerator Institute, and Snap Inc. This work used computational resources, including the NCSA Delta and DeltaAI and the PTI Jetstream2 supercomputers through allocations CIS230012, CIS230013, and CIS240311 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, as well as the TACC Frontera supercomputer and Amazon Web Services (AWS) through the National Artificial Intelligence Research Resource (NAIRR) Pilot. + +# References + +[1] Jinseok Bae, Jungdam Won, Donggeun Lim, Cheol-Hui Min, and Young Min Kim. Pmp: Learning to physically interact with environments using part-wise motion priors. In SIGGRAPH, 2023. 3 +[2] Qingwei Ben, Feiyu Jia, Jia Zeng, Junting Dong, Dahua Lin, and Jiangmiao Pang. HOMIE: Humanoid locomap-. manipulation with isomorphic exoskeleton cockpit. arXiv preprint arXiv:2502.13013, 2025. 2 +[3] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAV: Dataset and method for tracking human object interactions. In CVPR, 2022. 2, 3, 5, 6, 7, 8 +[4] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. In 3DV, 2024. 2, 3 +[5] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In AAAI, 2021. 2 +[6] Sammy Christen, Muhammed Kocabas, Emre Aksan, Jemin Hwangbo, Jie Song, and Otmar Hilliges. D-grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In CVPR, 2022. 4 +[7] Enric Corona, Albert Pumarola, Guillem Alenya, and Francesc Moreno-Noguer. Context-aware human motion prediction. In CVPR, 2020. 3 +[8] Jieming Cui, Tengyu Liu, Nian Liu, Yaodong Yang, Yixin Zhu, and Siyuan Huang. AnySkill: Learning open-vocabulary physical skill for interactive agents. In CVPR, 2024. 3 +[9] Ravinder S Dahiya, Giorgio Metta, Maurizio Valle, and Giulio Sandini. Tactile sensing—from humans to humanoids. IEEE transactions on robotics, 26(1):1-20, 2009. 4 +[10] Divyanshu Daiya, Damon Conover, and Aniket Bera. COLLAGE: Collaborative human-agent interaction generation using hierarchical latent diffusion and language models. arXiv preprint arXiv:2409.20502, 2024. 3 + +[11] Jeremy Dao, Helei Duan, and Alan Fern. Sim-to-real learning for humanoid box loco-manipulation. In ICRA, 2024. 2 +[12] Christian Diller and Angela Dai. CG-HOI: Contact-guided 3d human-object interaction generation. In CVPR, 2024. 3 +[13] Jiawei Gao, Ziqin Wang, Zeqi Xiao, Jingbo Wang, Tai Wang, Jinkun Cao, Xiaolin Hu, Si Liu, Jifeng Dai, and Jiangmiao Pang. CooHOI: Learning cooperative human-object interaction with manipulated object dynamics. In NeurIPS, 2024. 3 +[14] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. IMoS: Intent-driven full-body motion synthesis for human-object interactions. In Computer Graphics Forum, 2023. 3 +[15] Michael Gleicher. Motion editing with spacetime constraints. In Proceedings of the 1997 symposium on Interactive 3D graphics, 1997. 3 +[16] Zhaoyuan Gu, Junheng Li, Wenlan Shen, Wenhao Yu, Zhaoming Xie, Stephen McCrory, Xianyi Cheng, Abdulaziz Shamsah, Robert Griffin, C Karen Liu, et al. Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning. arXiv preprint arXiv:2501.02116, 2025. 2 +[17] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 4 +[18] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 3 +[19] Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. OmniH2O: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. arXiv preprint arXiv:2406.08858, 2024. 2 +[20] Tairan He, Zhengyi Luo, Wenli Xiao, Chong Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Learning human-to-humanoid real-time whole-body teleoperation. In IROS, 2024. 2 +[21] Wenkun He, Yun Liu, Ruitao Liu, and Li Yi. SyncDiff: Synchronized motion diffusion for multi-body human-object interaction synthesis. arXiv preprint arXiv:2412.20104, 2024.3 +[22] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. 3 +[23] Yinghao Huang, Omid Taheri, Michael J. Black, and Dimitrios Tzionas. InterCap: Joint markerless 3D tracking of humans and objects in interaction. In GCRP, 2022. 3 +[24] Inspire-robots. Smaller and higher-precision motion control experts. https://inspire-robots/store/.1,4,5 +[25] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. CHAIRS: Towards full-body articulated human-object interaction. In ICCV, 2023. 3 +[26] Nan Jiang, Zimo He, Zi Wang, Hongjie Li, Yixin Chen, Siyuan Huang, and Yixin Zhu. Autonomous character + +scene interaction synthesis from text instruction. In SIGGRAPH Asia, 2024. 3 +[27] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In CVPR, 2024. 3 +[28] Zhenyu Jiang, Yuqi Xie, Kevin Lin, Zhenjia Xu, Weikang Wan, Ajay Mandlekar, Linxi Fan, and Yuke Zhu. DexMimicGen: Automated data generation for bimanual dexterous manipulation via imitation learning. arXiv preprint arXiv:2410.24185, 2024. 2 +[29] Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. SuperPADL: Scaling language-directed physics-based control with progressive supervised distillation. In SIGGRAPH, 2024. 6 +[30] Hyeonwoo Kim, Sangwon Beak, and Hanbyul Joo. DAViD: Modeling dynamic affordance of 3d objects using pre-trained video diffusion models. arXiv preprint arXiv:2501.08333, 2025. 3 +[31] Jeonghwan Kim, Jisoo Kim, Jeonghyeon Na, and Hanbyul Joo. ParaHome: Parameterizing everyday home activities towards 3d generative modeling of human-object interactions. arXiv preprint arXiv:2401.10232, 2024. 3 +[32] Yeonjoon Kim, Hangil Park, Seungbae Bang, and Sung-Hee Lee. Retargeting human-object interaction to virtual avatars. IEEE transactions on visualization and computer graphics, 22(11):2405-2412, 2016. 4 +[33] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 5 +[34] Dieter Kraft. Algorithm 733: Tomp-fortran modules for optimal control calculations. ACM Transactions on Mathematical Software (TOMS), 20(3):262-281, 1994. 4 +[35] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. NIFTY: Neural object interaction fields for guided human motion synthesis. In CVPR, 2024. 3 +[36] Jiye Lee and Hanbyul Joo. Locomotion-Action-Manipulation: Synthesizing human-scene interactions in complex 3d environments. In ICCV, 2023. 2 +[37] Kang Hoon Lee, Myung Geol Choi, and Jehee Lee. Motion patches: building blocks for virtual environments annotated with motion data. In SIGGRAPH. 2006. 3 +[38] Yoonsang Lee, Sungeun Kim, and Jehee Lee. Data-driven biped control. In SIGGRAPH. 2010. 2 +[39] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2, 3, 5, 6, 7, 8 +[40] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. In ECCV, 2024. 3 +[41] Jinhan Li, Yifeng Zhu, Yuqi Xie, Zhenyu Jiang, Mingyo Seo, Georgios Pavlakos, and Yuke Zhu. OKAMI: Teaching humanoid robots manipulation skills through single video imitation. In CoRL, 2024. 2 +[42] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. In WACV, 2024. 3 + +[43] Fukang Liu, Zhaoyuan Gu, Yilin Cai, Ziyi Zhou, Shijie Zhao, Hyunyoung Jung, Sehoon Ha, Yue Chen, Danfei Xu, and Ye Zhao. Opt2skill: Imitating dynamically-feasible whole-body trajectories for versatile humanoid loco-manipulation. arXiv preprint arXiv:2409.20514, 2024. 2 +[44] Libin Liu and Jessica Hodgins. Learning to schedule control fragments for physics-based characters using deep q-learning. ACM Transactions on Graphics (TOG), 36(3): 1-14, 2017. 3 +[45] Yunze Liu, Changxi Chen, Chenjing Ding, and Li Yi. Phys-Reaction: Physically plausible real-time humanoid reaction synthesis via forward dynamics guided 4d imitation. In ACMMM, 2024. 2 +[46] Yun Liu, Bowen Yang, Licheng Zhong, He Wang, and Li Yi. Mimicking-bench: A benchmark for generalizable humanoid-scene interaction learning via human mimicking. arXiv preprint arXiv:2412.17730, 2024. 3 +[47] Yun Liu, Haolin Yang, Xu Si, Ling Liu, Zipeng Li, Yuxiang Zhang, Yebin Liu, and Li Yi. Taco: Benchmarking generalizable bimanual tool-action-object understanding. In CVPR, 2024. +[48] Yun Liu, Chengwen Zhang, Ruofan Xing, Bingda Tang, Bowen Yang, and Li Yi. Core4d: A 4d human-object-human interaction dataset for collaborative object rearrangement. In CVPR, 2025. 3 +[49] Jintao Lu, He Zhang, Yuting Ye, Takaaki Shiratori, Sebastian Starke, and Taku Komura. CHOICE: Coordinated human-object interaction in cluttered environments for pick-and-place actions. arXiv preprint arXiv:2412.06702, 2024.3 +[50] Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In ICCV, 2023. 2, 4, 5, 3 +[51] Zhengyi Luo, Jinkun Cao, Sammy Christen, Alexander Winkler, Kris Kitani, and Weipeng Xu. Grasping diverse objects with simulated humanoids. In NeurIPS, 2024. 2, 3 +[52] Zhengyi Luo, Jiashun Wang, Kangni Liu, Haotian Zhang, Chen Tessler, Jingbo Wang, Ye Yuan, Jinkun Cao, Zihui Lin, Fengyi Wang, et al. SMPLOlympics: Sports environments for physically simulated humanoids. arXiv preprint arXiv:2407.00187, 2024.3 +[53] Xintao Lv, Liang Xu, Yichao Yan, Xin Jin, Congsheng Xu, Shuwen Wu, Yifan Liu, Lincheng Li, Mengxiao Bi, Wenjun Zeng, et al. HIMO: A new benchmark for full-body human interacting with multiple objects. In ECCV, 2024. 3, 6 +[54] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. In NeurIPS, 2021. 6, 7, 2, 5 +[55] Josh Merel, Saran Tunyasuvunakool, Arun Ahuja, Yuval Tassa, Leonard Hasenclever, Vu Pham, Tom Erez, Greg Wayne, and Nicolas Heess. Catch & carry: reusable neural controllers for vision-guided whole-body tasks. ACM Transactions on Graphics (TOG), 39(4):39-1, 2020. 3 +[56] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini + +Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2, 6 +[57] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 3DV, 2024. 2 +[58] Liang Pan, Zeshi Yang, Zhiyang Dou, Wenjia Wang, Buzhen Huang, Bo Dai, Taku Komura, and Jingbo Wang. TokenHSI: Unified synthesis of physical human-scene interactions through task tokenization. In CVPR, 2025. 3 +[59] Soohwan Park, Hoseok Ryu, Seyoung Lee, Sunmin Lee, and Jehee Lee. Learning predict-and-simulate policies from unorganized human motion data. ACM Transactions on Graphics (TOG), 38(6):1-11, 2019. 4 +[60] Sungjae Park, Seungho Lee, Mingi Choi, Jiye Lee, Jeonghwan Kim, Jisoo Kim, and Hanbyul Joo. Learning to transfer human hand skills for robot manipulations. arXiv preprint arXiv:2501.04169, 2025. 2 +[61] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 2, 5 +[62] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. HOI-Diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 3, 7, 8, 5 +[63] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG), 37(4):1-14, 2018. 2, 3, 4, 5, 6, 7 +[64] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4):1-20, 2021. 3 +[65] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Transactions On Graphics (TOG), 41(4):1-17, 2022. 3 +[66] Haziq Razali and Yiannis Demiris. Action-conditioned generation of bimanual object manipulation sequences. In AAAI, 2023. 3 +[67] Daniele Reda, Jungdam Won, Yuting Ye, Michiel van de Panne, and Alexander Winkler. Physics-based motion retargeting from sparse inputs. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 6(3):1-19, 2023. 4 +[68] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6), 2017. 2 +[69] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 6, 7 + +[70] John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation. In ICLR, 2016. 5 +[71] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2, 5, 7 +[72] Carmelo Sferrazza, Dun-Ming Huang, Xingyu Lin, Youngwoon Lee, and Pieter Abbeel. Humanoidbench: Simulated humanoid benchmark for whole-body locomotion and manipulation. In RSS, 2024. 2 +[73] Wenfeng Song, Xinyu Zhang, Shuai Li, Yang Gao, Aimin Hao, Xia Hou, Chenglizhao Chen, Ning Li, and Hong Qin. HOIAnimator: Generating text-prompt human-object animations using novel perceptive diffusion models. In CVPR, 2024. 3 +[74] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 3 +[75] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39(4):54-1, 2020. 3 +[76] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In ECCV, 2020. 3 +[77] Omid Taheri, Vasileios Choutas, Michael J Black, and Dimitrios Tzionas. GOAL: Generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 3 +[78] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. Calm: Conditional adversarial latent models for directable virtual characters. In SIGGRAPH, 2023. 3 +[79] Chen Tessler, Yunrong Guo, Ofir Nabati, Gal Chechik, and Xue Bin Peng. Maskedmimic: Unified physics-based character control through masked motion inpainting. ACM Transactions on Graphics (TOG), 43(6):1-21, 2024. 2, 5, 6 +[80] Guy Tevet, Sigal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. CLoSD: Closing the loop between simulation and diffusion for multi-task character control. In ICLR, 2025. 3 +[81] Unittree. Unittree g1 humanoid agent ai avatar. https://www-unitree.com/g1/.1,4,5 +[82] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6 +[83] Ruben Villegas, Duygu Ceylan, Aaron Hertzmann, Jimei Yang, and Jun Saito. Contact-aware retargeting of skinned motion. In ICCV, 2021. 4 +[84] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 2022. 3 + +[85] Jiashun Wang, Jessica Hodgins, and Jungdam Won. Strategy and skill learning for physics-based table tennis animation. In SIGGRAPH, 2024. 3 +[86] Ruocheng Wang, Pei Xu, Haochen Shi, Elizabeth Schumann, and C Karen Liu. Fürelise: Capturing and physically synthesizing hand motion of piano performance. In SIGGRAPH Asia, 2024. 2 +[87] Wenjia Wang, Liang Pan, Zhiyang Dou, Zhouyingcheng Liao, Yuke Lou, Lei Yang, Jingbo Wang, and Taku Komura. SIMS: Simulating human-scene interactions with real world script planning. arXiv preprint arXiv:2411.19921, 2024.3 +[88] Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. PhysHOI: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393, 2023. 3, 7, 2 +[89] Yinhuai Wang, Qihan Zhao, Runyi Yu, Ailing Zeng, Jing Lin, Zhengyi Luo, Hok Wai Tsui, Jiwen Yu, Xiu Li, Qifeng Chen, et al. SkillMimic: Learning reusable basketball skills from demonstrations. In CVPR, 2025. 3, 7, 2 +[90] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. InterControl: Generate human motion interactions by controlling every joint. In NeurIPS, 2024. 2 +[91] Jungdam Won and Jehee Lee. Learning body shape variation in physics-based characters. ACM Transactions on Graphics (TOG), 38(6):1-12, 2019. 4, 5 +[92] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. A scalable approach to control diverse behaviors for physically simulated characters. ACM Transactions on Graphics (TOG), 39(4):33-1, 2020. 2, 4 +[93] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. THOR: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 3 +[94] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. SAGA: Stochastic whole-body grasping with contact. In ECCV, 2022. 3 +[95] Zhen Wu, Jiaman Li, Pei Xu, and C Karen Liu. Human-object interaction from human-level instructions. arXiv preprint arXiv:2406.17840, 2024.3 +[96] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In ICLR, 2024. 2 +[97] Xianghui Xie, Jan Eric Lenssen, and Gerard Pons-Moll. InterTrack: Tracking human object interaction without object templates. In 3DV, 2024. 3, 6 +[98] Zhaoming Xie, Sebastian Starke, Hung Yu Ling, and Michiel van de Panne. Learning soccer juggling skills with layer-wise mixture-of-experts. In SIGGRAPH, 2022. 3 +[99] Zhaoming Xie, Jonathan Tseng, Sebastian Starke, Michiel van de Panne, and C Karen Liu. Hierarchical planning and control for box loco-manipulation. arXiv preprint arXiv:2306.09532, 2023. 3 +[100] Pei Xu and Ruocheng Wang. Synchronize dual hands for physics-based dexterous guitar playing. In SIGGRAPH Asia, 2024. 2 + +[101] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 1, 3, 7, 8, 5 +[102] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. In NeurIPS, 2024. 1, 3 +[103] Sirui Xu, Dongting Li, Yucheng Zhang, Xiyan Xu, Qi Long, Ziyin Wang, Yunzhi Lu, Shuchang Dong, Hezi Jiang, Akshit Gupta, Yu-Xiong Wang, and Liang-Yan Gui. InterAct: Advancing large-scale versatile 3d human-object interaction generation. In CVPR, 2025. 3, 6 +[104] Zeshi Yang, Kangkang Yin, and Libin Liu. Learning to use chopsticks in diverse gripping styles. ACM Transactions on Graphics (TOG), 41(4):1-17, 2022. 2 +[105] Wenhao Yu, Greg Turk, and C Karen Liu. Learning symmetric and low-energy locomotion. ACM Transactions on Graphics (TOG), 37(4):1-12, 2018. 5 +[106] Ye Yuan and Kris Kitani. Residual force control for agile human behavior imitation and extended motion synthesis. In NeurIPS, 2020. 2 +[107] Yanjie Ze, Zixuan Chen, Wenhao Wang, Tianyi Chen, Xiaolin He, Ying Yuan, Xue Bin Peng, and Jiajun Wu. Generalizable humanoid manipulation with improved 3d diffusion policies. arXiv preprint arXiv:2410.10803, 2024. 2 +[108] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 3 +[109] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. NeuralDome: A neural modeling pipeline on multi-view human-object interactions. In CVPR, 2023. 3, 6, 8 +[110] Juze Zhang, Jingyan Zhang, Zining Song, Zhanhe Shi, Chengfeng Zhao, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Hoi-m^3: Capture multiple humans and objects interaction within contextual environment. In CVPR, 2024. 3 +[111] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. COUCH: Towards controllable human-chair interactions. In ECCV, 2022. 3 +[112] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Ilya Petrov, Vladimir Guzov, Helisa Dhamo, Eduardo Pérez-Pellitero, and Gerard Pons-Moll. FORCE: Dataset and method for intuitive physics guided human-object interaction. In 3DV, 2024. 3 +[113] Yunbo Zhang, Deepak Gopinath, Yuting Ye, Jessica Hodgins, Greg Turk, and Jungdam Won. Simulation and targeting of complex multi-character interactions. In SIGGRAPH, 2023. 3, 4 +[114] Chengfeng Zhao, Juze Zhang, Jiashen Du, Ziwei Shan, Junye Wang, Jingyi Yu, Jingya Wang, and Lan Xu. I'M HOI: Inertia-aware monocular capture of 3d human-object interactions. In CVPR, 2024. 3, 6 \ No newline at end of file diff --git a/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/images.zip b/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4a7986f7c976589f2ec5dc74e9c63aef86bffb28 --- /dev/null +++ b/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7789429d419f61a1c6f06ff6b8e7c510acdfacafaed37413cbcf4690be5549f +size 396915 diff --git a/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/layout.json b/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f821bab594b84ba4be48a788ed75571ad1b1c7e2 --- /dev/null +++ b/2025/InterMimic_ Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions/layout.json @@ -0,0 +1,10735 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 98, + 89, + 126 + ], + "blocks": [ + { + "bbox": [ + 67, + 98, + 89, + 126 + ], + "lines": [ + { + "bbox": [ + 67, + 98, + 89, + 126 + ], + "spans": [ + { + "bbox": [ + 67, + 98, + 89, + 126 + ], + "type": "image", + "image_path": "a1a2a0a3a33404ae530a2680deabf1baebe50111a29d9931bac3a9f91e9263e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 106, + 544, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 106, + 544, + 143 + ], + "spans": [ + { + "bbox": [ + 89, + 106, + 544, + 143 + ], + "type": "text", + "content": "INTERMIMIC: Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 163, + 473, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 163, + 473, + 178 + ], + "spans": [ + { + "bbox": [ + 138, + 163, + 473, + 178 + ], + "type": "text", + "content": "Sirui Xu" + }, + { + "bbox": [ + 138, + 163, + 473, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 163, + 473, + 178 + ], + "type": "text", + "content": " Hung Yu Ling" + }, + { + "bbox": [ + 138, + 163, + 473, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 163, + 473, + 178 + ], + "type": "text", + "content": " Yu-Xiong Wang" + }, + { + "bbox": [ + 138, + 163, + 473, + 178 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 138, + 163, + 473, + 178 + ], + "type": "text", + "content": " Liang-Yan Gui" + }, + { + "bbox": [ + 138, + 163, + 473, + 178 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 154, + 178, + 457, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 178, + 457, + 208 + ], + "spans": [ + { + "bbox": [ + 154, + 178, + 457, + 208 + ], + "type": "text", + "content": "1 University of Illinois Urbana-Champaign 2 Electronic Arts Equal Advising" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 204, + 209, + 406, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 209, + 406, + 220 + ], + "spans": [ + { + "bbox": [ + 204, + 209, + 406, + 220 + ], + "type": "text", + "content": "https://sirui-xu.github.io/InterMimic" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 56, + 241, + 555, + 317 + ], + "blocks": [ + { + "bbox": [ + 56, + 241, + 555, + 317 + ], + "lines": [ + { + "bbox": [ + 56, + 241, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 555, + 317 + ], + "type": "image", + "image_path": "9f6d3db18762132dbe4b9601ef951b5dda1f69372b6152f3831f8dbf55886183.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 320, + 217, + 331 + ], + "lines": [ + { + "bbox": [ + 110, + 320, + 217, + 331 + ], + "spans": [ + { + "bbox": [ + 110, + 320, + 217, + 331 + ], + "type": "text", + "content": "Highly-Dynamic Interaction" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 287, + 320, + 380, + 331 + ], + "lines": [ + { + "bbox": [ + 287, + 320, + 380, + 331 + ], + "spans": [ + { + "bbox": [ + 287, + 320, + 380, + 331 + ], + "type": "text", + "content": "Multi-Object Interaction" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 320, + 525, + 331 + ], + "lines": [ + { + "bbox": [ + 425, + 320, + 525, + 331 + ], + "spans": [ + { + "bbox": [ + 425, + 320, + 525, + 331 + ], + "type": "text", + "content": "Scalable Interaction Skills" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 57, + 345, + 552, + 415 + ], + "blocks": [ + { + "bbox": [ + 57, + 345, + 552, + 415 + ], + "lines": [ + { + "bbox": [ + 57, + 345, + 552, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 345, + 552, + 415 + ], + "type": "image", + "image_path": "f14c8ec6592615be708d340bdc035cd2408313cb1f97916107a9bc0c28a060d3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 417, + 178, + 426 + ], + "lines": [ + { + "bbox": [ + 69, + 417, + 178, + 426 + ], + "spans": [ + { + "bbox": [ + 69, + 417, + 178, + 426 + ], + "type": "text", + "content": "Humanoid Robot Interaction" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 192, + 417, + 264, + 427 + ], + "lines": [ + { + "bbox": [ + 192, + 417, + 264, + 427 + ], + "spans": [ + { + "bbox": [ + 192, + 417, + 264, + 427 + ], + "type": "text", + "content": "MoCap Correction" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 278, + 417, + 374, + 426 + ], + "lines": [ + { + "bbox": [ + 278, + 417, + 374, + 426 + ], + "spans": [ + { + "bbox": [ + 278, + 417, + 374, + 426 + ], + "type": "text", + "content": "Interaction Prediction" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 438, + 339, + 514, + 350 + ], + "lines": [ + { + "bbox": [ + 438, + 339, + 514, + 350 + ], + "spans": [ + { + "bbox": [ + 438, + 339, + 514, + 350 + ], + "type": "text", + "content": "Lift the woodchair." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 417, + 511, + 426 + ], + "lines": [ + { + "bbox": [ + 440, + 417, + 511, + 426 + ], + "spans": [ + { + "bbox": [ + 440, + 417, + 511, + 426 + ], + "type": "text", + "content": "Text-to-Interaction" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 437, + 555, + 483 + ], + "lines": [ + { + "bbox": [ + 54, + 437, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 555, + 483 + ], + "type": "text", + "content": "Figure 1. InterMimic enables physically simulated humans to perform interactions with dynamic and diverse objects. It supports highly-dynamic, multi-object interactions and scalable skill learning (Top), making it adaptable for versatile downstream applications (Bottom): it can translate whole-body loco-manipulation skills to a humanoid robot [24, 81], perfect interaction MoCap data, and bridge kinematic generation, e.g., predicting future interactions from past (InterDiff [101]) or generating interactions given text prompts (InterDreamer [102])." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 406, + 350, + 552, + 414 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 350, + 552, + 414 + ], + "spans": [ + { + "bbox": [ + 406, + 350, + 552, + 414 + ], + "type": "image", + "image_path": "c213ba55d2db72b1f2b0590ce6c13089bf93c59707784bb0f4bbe5ea1b015256.jpg" + } + ] + } + ], + "index": 16, + "type": "text" + }, + { + "bbox": [ + 151, + 506, + 200, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 506, + 200, + 519 + ], + "spans": [ + { + "bbox": [ + 151, + 506, + 200, + 519 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 54, + 535, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 535, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 535, + 297, + 715 + ], + "type": "text", + "content": "Achieving realistic simulations of humans interacting with a wide range of objects has long been a fundamental goal. Extending physics-based motion imitation to complex human-object interactions (HOIs) is challenging due to intricate human-object coupling, variability in object geometries, and artifacts in motion capture data, such as inaccurate contacts and limited hand detail. We introduce InterMimic, a framework that enables a single policy to robustly learn from hours of imperfect MoCap data covering diverse full-body interactions with dynamic and varied objects. Our key insight is to employ a curriculum strategy – perfect first, then scale up. We first train subject-specific teacher policies to mimic, retarget, and refine motion capture data. Next, we distill these teachers into a student policy, with the teachers acting as online experts providing direct supervision." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 508, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 616 + ], + "type": "text", + "content": "as well as high-quality references. Notably, we incorporate RL fine-tuning on the student policy to surpass mere demonstration replication and achieve higher-quality solutions. Our experiments demonstrate that InterMimic produces realistic and diverse interactions across multiple HOI datasets. The learned policy generalizes in a zero-shot manner and seamlessly integrates with kinematic generators, elevating the framework from mere imitation to generative modeling of complex human-object interactions." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 644, + 395, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 644, + 395, + 657 + ], + "spans": [ + { + "bbox": [ + 314, + 644, + 395, + 657 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "Animating human-object interactions is a challenging and time-consuming task even for skilled animators. It requires a deep understanding of physics and meticulous attention to detail to create natural and convincing interactions. While" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12266" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "Motion Capture (MoCap) data provides references, animators often need to correct contact errors caused by sensor limitations and occlusions between humans and objects. However, this process remains unscalable, as refining a single motion demands a delicate balance between preserving the captured data and ensuring its physical plausibility." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 146, + 298, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 146, + 298, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 146, + 298, + 373 + ], + "type": "text", + "content": "Physics-based human motion imitation [38, 63] offers an alternative approach to improving motion fidelity, by training control policies to mimic reference MoCap data within a physics simulator. However, scaling up human-object interaction imitation presents significant challenges: (i) MoCap Imperfection: Contact artifacts are common, causing expected contacts to fluctuate instead of maintaining consistent zero distance, often due to MoCap limitations or missing hand capture [3, 39]. Accurately imitating MoCap kinematics can result in unrealistic dynamics in simulation. Moreover, HOI datasets often include diverse human shapes, requiring motion retargeting to adapt movements across different human models while preserving interaction dynamics. This retargeting process is imperfect and can introduce new contact artifacts or exacerbate existing ones. (ii) Scaling-up: Although large-scale motion imitation has been explored in previous works [50, 79, 92, 106], it remains largely underexplored for whole-body interactions involving dynamic and diverse objects." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 375, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 506 + ], + "type": "text", + "content": "In this paper, we aim to utilize rich yet imperfect motion capture interaction datasets to train a control policy capable of learning diverse motor skills while enhancing the plausibility of these actions by correcting errors, such as inaccurate hand motions and faulty contacts. Our approach is grounded on the key insight of tackling the challenges of skill perfection and skill integration progressively. We implement a curriculum-based teacher-student distillation framework, where multiple teacher policies focus on imitating and refining small subsets of interactions, and a student policy integrates these skills from the teachers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 509, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 675 + ], + "type": "text", + "content": "Instead of relying on curated data that covers a limited range of actions [4, 51], we employ multiple teacher policies trained on a diverse set of imperfect interaction data and address two key challenges: retargeting and recovering. First, we unify all training policies to a canonical human model, by embedding HOI retargeting directly into the imitation. This is achieved by reframing the policy learning to optimize both imitation and retargeting objectives. Second, our teacher policies refine interaction motion through learning from it, as accurate contact dynamics enforced by a physics simulator inherently correct inaccuracies in the reference kinematics. To support this, we introduce tailored contact-guided reward and optimize trajectory collection, enabling effective skill imitation despite MoCap errors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Introducing teacher policies offers several key benefits. By leveraging teacher rollouts, we effectively distill raw MoCap data into refined HOI references with a unified em" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 289 + ], + "type": "text", + "content": "bodiment and enhanced physical fidelity. These refined references guide the subsequent student policy training, reducing the negative impact of errors in the original MoCap data. A major hurdle in scaling motion imitation is the sample inefficiency of Reinforcement Learning (RL), which can lead to prohibitively long training times. Our teacher-student approach mitigates this through a space-time trade-off: multiple teacher policies are trained in parallel on smaller, more manageable data subsets, and their expertise is then distilled into a single student policy. We begin with demonstration-based distillation to bootstrap PPO [71] updates, reducing reliance on pure trial and error and enabling more effective scaling. As training progresses, the student gradually shifts from heavy demonstration guidance to increased RL updates, ultimately surpassing simple demonstration memorization. This mirrors alignment strategies in Large Language Models (LLMs), where demonstration-based pretraining is refined through RL fine-tuning [56]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 295, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 555, + 498 + ], + "type": "text", + "content": "To summarize, our contributions are as follows: (i) We introduce InterMimic, which, to the best of our knowledge, is the first framework designed to train physically simulated humans to develop a wide range of whole-body motor skills for interacting with diverse and dynamic objects, extending beyond traditional grasping tasks. (ii) We develop a teacher-student training strategy, where teacher policies provide a unified solution to address the challenges of retargeting and refining in HOI imitation. The student distillation introduces a scalable solution by leveraging a space-time tradeoff. (iii) We demonstrate that our unified framework, InterMimic, as illustrated in Figure 1, effectively handles versatile physics-based interaction animation, recovering motions with realistic and physically plausible details. Notably, by combining kinematic generators with InterMimic, we enable a physics-based agent to achieve tasks such as interaction prediction and text-to-interaction generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 530, + 400, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 530, + 400, + 544 + ], + "spans": [ + { + "bbox": [ + 314, + 530, + 400, + 544 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "content": "Significant progress has been made in physics-based human interaction animation and control, with advancements in areas such as human-human interactions [45, 90], hand-object interactions [60, 86, 100, 104], human interactions with static scenes [5, 36, 57, 79, 96], and real-world humanoid control for object manipulation [2, 11, 16, 19, 20, 28, 41, 43, 72, 107]. Among these areas, we are the first to achieve universal whole-body loco-manipulation simulation with diverse dynamic objects, beyond pick-and-place and grasping actions – a novel achievement in animation and an idealized reference for real-world humanoid control. Below, we elaborate on recent studies on whole-body interaction animation, particularly involving dynamic objects." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12267" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 234, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 234, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 234, + 83 + ], + "type": "text", + "content": "2.1. Kinematic Interaction Animation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 354 + ], + "type": "text", + "content": "Generating human interactions has been a long-standing topic in animation and computer graphics [15, 37]. Significant advances in character animation have emerged with the advent of deep learning, e.g., including phase-function-based methods [22] that enable object interactions like carrying a box [74] or playing basketball [75]. This is extended to more diverse but static objects approaching [35, 77, 94, 111]. Subsequent efforts [14, 26, 27, 40, 42, 49, 66, 95] integrate object motion into interactions but remain constrained by assuming that interactions occur primarily through the hands. To address this, recent developments [7, 10, 12, 21, 30, 62, 73, 84, 93, 101, 102] introduce interactions in a fashion of whole-body loco-manipulation that engages multiple body parts in contact. However, these methods often suffer from physical inaccuracies, such as floating contacts and penetrations, while they generate only body motion without considering hand dexterity. In this work, we address physical inaccuracies by refining imperfect kinematic generation through physics simulation, with InterDiff [101] and HOI-Diff [62] serving as motion planning for loco-manipulation that bridges high-level decision-making (e.g., text instruction) with low-level execution." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 360, + 251, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 251, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 251, + 372 + ], + "type": "text", + "content": "2.2. Physics-based Interaction Animation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 379, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 713 + ], + "type": "text", + "content": "Physics-based methods generate motion through motor control policies within a physics simulator, e.g., achieved via deep reinforcement learning to track reference motions [63]. These policies are directly applicable for executing simple interactions, such as punching or striking an object [8, 65, 78, 80]. To achieve more complex interactions, early studies focus on specific scenarios, including notable sports-related [52] examples such as basketball [88], skating [44], soccer [98], tennis [108], table tennis [85], and more proposed in [1]. Research also demonstrates flexibility in more general but simpler box carrying tasks [58, 87, 113]. These advancements are achieved through the integration of multiple control policies [55], the use of adversarial motion priors [13, 18, 64], and imitating diverse kinematic generations [95, 99]. However, these methods train their policies in a non-scalable manner, with each policy handling only specific object types or actions. In pursuit of a single, scalable policy to enable multiple interaction skills, existing methods either rely on fixed interaction patterns, such as approaching and grasping objects [4, 51], or extend single-object skills, e.g., interactions involving a basketball [89]. Additionally, they mostly depend on highly curated data from the GRAB dataset [76], which, despite its high quality, primarily features low-dynamic full-body motion and only small-sized objects. More recent datasets [3, 23, 25, 31, 39, 46-48, 53, 97, 103, 109, 110, 112, 114] offer richer full-body interactions with diverse objects but contain noticeable artifacts that chal" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 170 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 170 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 170 + ], + "type": "image", + "image_path": "f5d9c41cf08c4f0fd5f8d0add65355f1e232146e94c43d345381ec12fa00146f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 178, + 555, + 245 + ], + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 245 + ], + "type": "text", + "content": "Figure 2. Our two-stage pipeline: (i) training each teacher policy (MLP) on a small data subset with initialization corrected via Physical State Initialization (PSI), and (ii) freezing the teacher policies to provide refined references for training a student policy (Transformer). The student leverages teacher supervision for effective scaling and is fine-tuned through RL." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 268, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 341 + ], + "type": "text", + "content": "lenge existing motion imitation approaches. We process data from OMOMO [39], BEHAVE [3], HODome [109], and IMHD [114] collected in the InterAct [103] dataset, and the multi-object dataset HIMO [53], highlighting InterMimic's scalability to diverse interactions and its robustness against MoCap artifacts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 354, + 397, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 397, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 397, + 369 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": "Task Formulation. The goal of human-object interaction (HOI) imitation is to learn a policy " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": " that produces simulated human-object motion " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\pmb{q}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": " closely matching a ground-truth reference " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\pmb{q}}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": " derived from large-scale MoCap data. Given the geometries of the human and objects, the policy should also compensate for missing or inaccurate details in the dataset. Each pose " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{q}_t" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": " has two components: the human pose " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{q}_t^h" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": " and the object pose " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{q}_t^o" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": ". The human pose is defined as " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{q}_t^h = \\{\\pmb{\\theta}_t^h, \\pmb{p}_t^h\\}" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_t^h \\in \\mathbb{R}^{52 \\times 3}" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": " represents the joint rotations, and " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t^h \\in \\mathbb{R}^{52 \\times 3}" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": " specifies the joint positions. Specifically, our human model includes 30 hand joints and 22 joints for the rest of the body, with one root joint's rotation and position specified in global coordinates, whereas the rotations and positions of all other joints are defined relative to their respective parent joints' coordinate systems. The object pose " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{q}_t^o" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": " is represented as " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\theta}_t^o, \\pmb{p}_t^o\\}" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_t^o \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": " denotes the object's orientation and " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t^o \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": " the position. All simulation states have corresponding ground-truth values, denoted by the hat symbol. For instance, the reference object rotation is " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\pmb{\\theta}}_t^o\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": ". The environmental setup for the simulation is detailed in Sec. B. Overview. We formulate interaction imitation as a Markov Decision Process (MDP), defined by states, actions, simulator-provided transition dynamics, and a reward function. Figure 2 illustrates our two-stage framework: (i) training teacher policies " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pi^{(T)}" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": " on small skill subsets, and (ii) distilling these teachers into a scalable student policy " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pi^{(S)}" + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "text", + "content": " for large-scale skill learning. In Sec. 3.1, we define the states " + }, + { + "bbox": [ + 313, + 376, + 555, + 715 + ], + "type": "inline_equation", + "content": "s_t" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12268" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "and actions " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\pmb{a}_t" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", applicable to both teacher " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\pi^{(T)}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and student " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\pi^{(S)}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " policies. In Sec. 3.2, we describe how teacher policies are trained via RL, focusing on reward designs that facilitate retargeting, as well as techniques that mitigate the impact of imperfections in the reference data. Sec. 3.3 details the subsequent distillation of teachers into a scalable student policy, leveraging both RL and learning from demonstration." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 180, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 180, + 184 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 180, + 184 + ], + "type": "text", + "content": "3.1. Policy Representation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "text", + "content": "State. The state " + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "text", + "content": ", which serves as input to the policy, comprises two components " + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "inline_equation", + "content": "s_t = \\{s_t^s, s_t^g\\}" + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "text", + "content": ". The first part, " + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "inline_equation", + "content": "s_t^s" + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "text", + "content": ", contains human proprioception and object observations, expressed as, " + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\{\\{\\pmb{\\theta}_t^h, \\pmb{p}_t^h, \\pmb{\\omega}_t^h, \\pmb{v}_t^h\\}, \\{\\pmb{\\theta}_t^o, \\pmb{p}_t^o, \\pmb{\\omega}_t^o, \\pmb{v}_t^o\\}, \\{d_t, c_t\\}\\}" + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\theta}_t^h, \\pmb{p}_t^h, \\pmb{\\omega}_t^h, \\pmb{v}_t^h\\}" + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "text", + "content": " represent the rotation, position, angular velocity, and velocity of all joints, respectively, while " + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\theta}_t^o, \\pmb{p}_t^o, \\pmb{\\omega}_t^o, \\pmb{v}_t^o\\}" + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "text", + "content": " represent the orientation, location, velocity, and angular velocity of the object, respectively. Motivated by [6], we include object geometry and whole-body haptic sensing from two elements: (i) " + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "inline_equation", + "content": "d_t" + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "text", + "content": ", vectors from human joints to their nearest points on each object surface; and (ii) " + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "inline_equation", + "content": "c_t" + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "text", + "content": ", contact markers indicating whether the human's rigid body parts experience applied forces; this serves as simplified tactile or force sensing - an important multimodal input in robot manipulation tasks [9]. The goal state " + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "inline_equation", + "content": "s_t^g = \\{s_{t,t+k}^g\\}_{k \\in K}" + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "text", + "content": " integrates reference poses from the ground truth motion, where " + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "inline_equation", + "content": "s_{t,t+k}^g" + }, + { + "bbox": [ + 55, + 193, + 296, + 399 + ], + "type": "text", + "content": " is defined as," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 414, + 294, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 436 + ], + "type": "interline_equation", + "content": "\\{\\{\\hat {\\boldsymbol {\\theta}} _ {t + k} ^ {h} \\ominus \\boldsymbol {\\theta} _ {t} ^ {h}, \\hat {\\boldsymbol {p}} _ {t + k} ^ {h} - \\boldsymbol {p} _ {t} ^ {h} \\}, \\{\\hat {\\boldsymbol {\\theta}} _ {t + k} ^ {o} \\ominus \\boldsymbol {\\theta} _ {t} ^ {o}, \\hat {\\boldsymbol {p}} _ {t + k} ^ {o} - \\boldsymbol {p} _ {t} ^ {o} \\}, \\tag {1}", + "image_path": "5455c794ddba902f55b904ed0e00262aa4b0851ff101fa8cc2f90b43c3be703c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 434, + 277, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 277, + 450 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 277, + 450 + ], + "type": "interline_equation", + "content": "\\{\\hat {\\boldsymbol {d}} _ {t + k} - \\boldsymbol {d} _ {t}, \\hat {\\boldsymbol {c}} _ {t + k} - \\boldsymbol {c} _ {t} \\}, \\{\\hat {\\boldsymbol {\\theta}} _ {t + k} ^ {h}, \\hat {\\boldsymbol {p}} _ {t + k} ^ {h}, \\hat {\\boldsymbol {\\theta}} _ {t + k} ^ {o}, \\hat {\\boldsymbol {p}} _ {t + k} ^ {o} \\} \\},", + "image_path": "0660d218c666543dd226a0c6ce110776e0bef6138acc55fc20126ffa2023510b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 466, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 295, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 466, + 295, + 529 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{\\theta}}_{t + k}^{h},\\hat{\\pmb{p}}_{t + k}^{h},\\hat{\\pmb{d}}_{t + k},\\hat{\\pmb{c}}_{t + k}" + }, + { + "bbox": [ + 55, + 466, + 295, + 529 + ], + "type": "text", + "content": " represent the reference information at time step " + }, + { + "bbox": [ + 55, + 466, + 295, + 529 + ], + "type": "inline_equation", + "content": "t + k,\\ominus" + }, + { + "bbox": [ + 55, + 466, + 295, + 529 + ], + "type": "text", + "content": " denotes the calculation of rotation difference. All continuous elements of " + }, + { + "bbox": [ + 55, + 466, + 295, + 529 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 55, + 466, + 295, + 529 + ], + "type": "text", + "content": " are normalized relative to the current direction of view of the human and the position of the root [63]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 663 + ], + "type": "text", + "content": "Given that most MoCap data does not provide reference contact or tactile information, we extract reference contact markers " + }, + { + "bbox": [ + 55, + 532, + 295, + 663 + ], + "type": "inline_equation", + "content": "\\hat{c}_{t + k}" + }, + { + "bbox": [ + 55, + 532, + 295, + 663 + ], + "type": "text", + "content": " by inferring dynamic information, beyond relying solely on inaccurate contact distances, specifically by analyzing the object's acceleration to detect human-induced forces. To accommodate the variability in contact distances observed in reference motion, we discretize reference contact markers using varying distance thresholds, as illustrated in Fig. 3(i). The neutral areas serve as buffer zones, avoiding the penalization or enforcement of strict contact. See Sec. C of supplementary for details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Action. Our human model has 51 actuated joints, defining an action space of " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t \\in \\mathbb{R}^{51 \\times 3}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". These actions are specified as joint PD targets using the exponential map and are converted into torques applied to each of the human joints." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 444, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 444, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 444, + 85 + ], + "type": "text", + "content": "3.2. Imitation as Perfecting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 90, + 553, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 553, + 198 + ], + "type": "text", + "content": "The teacher policy " + }, + { + "bbox": [ + 313, + 90, + 553, + 198 + ], + "type": "inline_equation", + "content": "\\pi^{(\\mathrm{T})}" + }, + { + "bbox": [ + 313, + 90, + 553, + 198 + ], + "type": "text", + "content": " is trained via RL to maximize the expected discounted reward by comparing simulated states against potentially erroneous reference states. The training involves: (i) trajectory collection, where we explain how trajectories are initialized and terminated. (ii) policy updating, where collected trajectories and their associated rewards are used to refine the policy. In this section, we elaborate on our reward design and how we optimize trajectory collection to mitigate the impact of reference inaccuracies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 199, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 554, + 376 + ], + "type": "text", + "content": "Imitation as Retargeting. We tailor teacher policies to each human subject, while all policies share the same base human model. This serves the retargeting purpose by converting HOIs from different human shapes into a unified base shape. Although motion imitation does not necessarily require a unified human model [50, 91], our approach offers two benefits: (i) It enhances integration with kinematic generation methods, which generally perform better on a single, unified shape [17]. (ii) It demonstrates possible integration with real-world humanoid deployment, which requires retargeting to a consistent physical embodiment. In Figure 1, our method translates MoCap data into motor skills on a Unitree G1 [81] with two Inspire hands [24], all without external retargeting in complex contact-rich scenarios. See Sec. F of the supplementary for additional details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 378, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 510 + ], + "type": "text", + "content": "Human [83] or HOI [32] retargeting can be formulated as an optimization problem. Inverse Kinematics (IK) methods, such as those based on quadratic programming [34], demonstrate effectiveness in simplified scenarios but remain underexplored for motions featuring intricate object interactions. RL, by contrast, solves the optimization by maximizing an expected cumulative reward, prompting us to investigate whether RL-driven HOI imitation can be used for HOI retargeting. This extends existing physics-based retargeting approaches, which either omit object interactions [67] or are non-scalable with a single reference [113]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 593 + ], + "type": "text", + "content": "While the kinematics should differ due to the embodiment gap, we argue that the dynamics between human and object should remain invariant. Thus, we define rewards to include an embodiment-aware component that loosely aligns the simulated kinematics with the reference interaction, and an embodiment-agnostic reward component that encourages dynamics to be close to the reference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "text", + "content": "Embodiment-Aware Reward. When the human and object are far apart, retargeting should prioritize capturing rotational motion, whereas when they are close, accurate position tracking becomes crucial for achieving contact. To reflect this, we define the weights " + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{w}_d" + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "text", + "content": " that are inversely proportional to the distances between joints and the object [113]. The reward thus includes cost functions for joint position " + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "inline_equation", + "content": "E_p^h = \\langle \\pmb{\\Delta}_p^h, \\pmb{w}_d \\rangle" + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "inline_equation", + "content": "E_\\theta^h = \\langle \\pmb{\\Delta}_\\theta^h, \\mathbf{1} - \\pmb{w}_d \\rangle" + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "text", + "content": ", and interaction tracking " + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "inline_equation", + "content": "E_d = \\langle \\pmb{\\Delta}_d, \\pmb{w}_d \\rangle" + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "text", + "content": " is the inner product, " + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}_p^h[i] = \\| \\hat{\\pmb{p}}^h[i] - \\pmb{p}^h[i] \\|" + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 594, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}_\\theta^h[i] =" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12269" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 170, + 129 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 170, + 129 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 170, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 170, + 129 + ], + "type": "image", + "image_path": "111fbf92abb4b6809bda6221ceaa38a3758d1c1710c4dd9b869544587fabd65b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 129, + 161, + 138 + ], + "lines": [ + { + "bbox": [ + 67, + 129, + 161, + 138 + ], + "spans": [ + { + "bbox": [ + 67, + 129, + 161, + 138 + ], + "type": "text", + "content": "(i) Reference contact markers" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 147, + 295, + 202 + ], + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 202 + ], + "type": "text", + "content": "Figure 3. (i) Visualization of reference contact markers that accommodate varied contact distances: red to promote contact, green for neutral areas where contact is neither promoted nor penalized, and blue to penalize contact. (ii) Initializing the rollout with reference (RSI) or reference corrected via simulation (PSI)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 172, + 77, + 215, + 126 + ], + "blocks": [ + { + "bbox": [ + 172, + 77, + 215, + 126 + ], + "lines": [ + { + "bbox": [ + 172, + 77, + 215, + 126 + ], + "spans": [ + { + "bbox": [ + 172, + 77, + 215, + 126 + ], + "type": "image", + "image_path": "e2e0019fc9d39d2746e6af3c0045298378d8b7396113fc57417d2b0089f67cd7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 129, + 254, + 138 + ], + "lines": [ + { + "bbox": [ + 205, + 129, + 254, + 138 + ], + "spans": [ + { + "bbox": [ + 205, + 129, + 254, + 138 + ], + "type": "text", + "content": "(ii) RSI vs. PSI" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 217, + 78, + 294, + 126 + ], + "blocks": [ + { + "bbox": [ + 217, + 78, + 294, + 126 + ], + "lines": [ + { + "bbox": [ + 217, + 78, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 217, + 78, + 294, + 126 + ], + "type": "image", + "image_path": "610b8a83552a18149b470ea8f89665350e244f4dbba2121a0575a3857b8fb085.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\| \\hat{\\pmb{\\theta}}^h [i]\\ominus \\pmb{\\theta}^h [i]\\|" + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}_d[i] = \\| \\hat{\\pmb{d}} [i] - \\pmb {d}[i]\\|" + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "text", + "content": " represent the displacement for the variables defined in Sec. 3.1 with timestep " + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "text", + "content": " omitted. The formulation of " + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{d}" + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "text", + "content": " is provided in Sec. D of supplementary. The reward to be maximized can be formulated as " + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\exp (-\\lambda E)" + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "text", + "content": " for each cost function " + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "text", + "content": " with a specific hyperparameter " + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 224, + 296, + 300 + ], + "type": "text", + "content": ". Details can be found in Sec. D." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "text", + "content": "Embodiment-Agnostic Reward. The reward includes components for object tracking and contact tracking. The object tracking cost is defined for position " + }, + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "inline_equation", + "content": "E_{p}^{o} = \\|\\hat{\\boldsymbol{p}}^{o} - \\boldsymbol{p}^{o}\\|" + }, + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "text", + "content": " and rotation " + }, + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "inline_equation", + "content": "E_{\\theta}^{o} = \\|\\hat{\\boldsymbol{\\theta}}^{o} - \\boldsymbol{\\theta}^{o}\\|" + }, + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "text", + "content": ", with all values normalized to the human's current position and direction. The contact tracking reward comprises two cost functions: body contact promotion " + }, + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "inline_equation", + "content": "E_{b}^{c}" + }, + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "text", + "content": " and penalty " + }, + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "inline_equation", + "content": "E_{p}^{c}" + }, + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "text", + "content": ", both aligning the simulated contact " + }, + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "text", + "content": " with reference markers " + }, + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 55, + 301, + 295, + 579 + ], + "type": "text", + "content": ", as shown in Figure 3. We define three contact levels – promotion, penalty, and neutral – to accommodate potential inaccuracies in reference contact distances. The detailed formulation can be found in Sec. D of the supplementary. Since the physics engine does not differentiate between object, ground, and self-contact, we adopt two strategies: (i) we model foot-ground contact promotion and penalty. This ensures proper foot lifting during cyclic walking and mitigates foot hobbling. (ii) We allow self-collision to avoid self-contact promotion but to promote object interaction. This poses minimal risk as the policy is guided by MoCap reference, which, although lacking perfect contact accuracy, rarely shows self-penetration. For humanoid robots with embodiments that differ from the MoCap reference and require real-world applicability, we disable self-collision, as discussed in Sec. F." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 580, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 628 + ], + "type": "text", + "content": "We introduce energy consumption rewards [105] to penalize large human or object jitters, with a proposed contact energy penalizing abrupt contact to promote compliant interactions. See Sec. D of supplementary for more details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": "Hand Interaction Discovery. We use data with average or flattened hand poses [3, 39], which makes accurate object manipulation imitation challenging. To address this, we activate a reference contact marker for any hand part when a fingertip or palm is near an object. Given tasks that do not demand high dexterity, employing a contact-promoting reward with this marker enables policies to develop effective" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "hand interaction strategies, leveraging the exploratory nature of RL. Additionally, we constrain the range of motion (RoM) of the hands to ensure natural movement. See Sec. D and Sec. B of the supplementary for further details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "text", + "content": "Policy Learning. Following [63], the control policy " + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "text", + "content": " is trained using PPO [71] with the policy gradient " + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "inline_equation", + "content": "L(\\psi) = \\mathbb{E}_t[\\min (r_t(\\psi)A_t,\\mathrm{clip}(r_t(\\psi),1 - \\epsilon ,1 + \\epsilon)A_t)" + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "text", + "content": " are the parameters of " + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "inline_equation", + "content": "r_t(\\psi)" + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "text", + "content": " quantifies the difference in action likelihoods between updated and old policies. " + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "text", + "content": " is a small constant, and " + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "text", + "content": " is the advantage estimate given by the generalized advantage estimator " + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\mathrm{GAE}(\\lambda)" + }, + { + "bbox": [ + 313, + 121, + 555, + 205 + ], + "type": "text", + "content": " [70]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": "Physical State Initialization. Learning later-phase motion can be essential for policies to achieve high rewards during earlier phases, compared to incrementally learning from the starting phase. Thus, Reference State Initialization (RSI) [63] sets the current pose " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_t" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": " to a reference pose " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{q}}_t" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": " at a random timestep " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": ", for initializing the rollout. However, initializing with the imperfect reference can introduce critical artifacts, such as contact floating or incorrect hand motion, leading to unrecoverable failures, e.g., object falling, as depicted in Figure 3(ii). These issues render many initializations ineffective, limiting training on certain interaction phases since successful rollouts may not reach them before the maximum length. The problem is exacerbated by the use of prioritized sampling [79, 91], which favors high-failure-rate initializations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 388, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 592 + ], + "type": "text", + "content": "To address the need for higher-quality reference initialization, we propose Physical State Initialization (PSI). As illustrated in Figure 2, PSI begins by creating an initialization buffer that stores reference states from MoCap and simulation states from prior rollouts. For each new rollout, an initial state is randomly selected from this buffer, which increases the likelihood of starting from advantageous positions. Once a rollout is completed, trajectories are evaluated based on their expected discounted rewards; those above a certain threshold are added to the buffer using a first-in-first-out (FIFO) strategy, while older or lower-quality trajectories are discarded. This selective reintroduction of high-value states for initialization helps maintain stable policy updates. We apply PSI in a sparse manner to ensure training efficiency. As shown in Figure 3(ii), PSI can collect trajectories for policy update that RSI does not effectively utilize. Further details are provided in Sec. E of the supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Interaction Early Termination. Early Termination (ET) [63] is commonly used in motion imitation, ending an episode when a body part makes unplanned ground contact or when the character deviates significantly from the reference [50], thus stopping the policy from overvaluing invalid transitions. However, additional conditions should be considered for human-object interactions. We propose Interaction Early Termination (IET), which supplements ET with three extra checks: (i) Object points deviate from their references by more than " + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": " on average. (ii) Weighted av" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12270" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "erage distances between the character's joints and the object surface exceed " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " from the reference. (iii) Any required body-object contact is lost for over 10 consecutive frames. Full conditions are detailed in Sec. E of the supplementary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 126, + 201, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 201, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 201, + 138 + ], + "type": "text", + "content": "3.3. Imitation with Distillation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "text", + "content": "As shown in Figure 2, after training the teacher policies on data from each subject (Sec. 3.2), we aggregate them to train a student policy " + }, + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\pi^{(\\mathrm{S})}" + }, + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "text", + "content": " to master all skills. As outlined in Algorithm 1, the combined teacher policies, denoted by " + }, + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\pi^{(\\mathrm{T})}" + }, + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "text", + "content": " for brevity, serves dual roles by providing state-action trajectories " + }, + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "inline_equation", + "content": "(s^{(\\mathrm{T})}, a^{(\\mathrm{T})})" + }, + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "text", + "content": ": (i) the state " + }, + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "inline_equation", + "content": "s^{(\\mathrm{T})}" + }, + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "text", + "content": " for reference distillation, and (ii) the action " + }, + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "inline_equation", + "content": "a^{(\\mathrm{T})}" + }, + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "text", + "content": " for policy distillation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 228, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 335 + ], + "type": "text", + "content": "Reference Distillation. Noisy MoCap data can hinder policy learning, especially at larger scales. In contrast, teacher policies trained on smaller-scale data effectively address these issues by correcting contact artifacts, refining hand placements, and recovering missing details (see Figures 1 and 5). To fully leverage teacher policies, we use their rollouts as references for defining the student policy's goal state and reward functions, distinguishing our approach from distillation based on only action output." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 336, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 529 + ], + "type": "text", + "content": "Policy Distillation. We apply distillation on action outputs, which we view as crucial for scaling policies to large datasets. In essence, we trade space for time: teacher policies are trained in parallel on smaller data subsets, allowing the student policy to scale through distillation. Following Algorithm 1, we begin with Behavior Cloning (BC) [29] and use RL fine-tuning to go beyond demonstration memorization, an approach common in LLM alignment [56]. We integrate BC into online policy updates with a staged schedule: we start with DAgger [69] and gradually transition to PPO. Throughout, the critic is continuously trained with the reward from Sec. 3.2. This RL fine-tuning phase is crucial as teacher policies may behave differently when performing similar skills, and simple BC can lead to suboptimal \"averaging\" behavior, where RL fine-tuning helps the student converge on optimal solutions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 138, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 138, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 138, + 545 + ], + "type": "text", + "content": "3.4. Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 661 + ], + "type": "text", + "content": "We set the keyframe indices " + }, + { + "bbox": [ + 55, + 552, + 296, + 661 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 552, + 296, + 661 + ], + "type": "text", + "content": " (Sec. 3.1, Eq. 1) to " + }, + { + "bbox": [ + 55, + 552, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\{1,16\\}" + }, + { + "bbox": [ + 55, + 552, + 296, + 661 + ], + "type": "text", + "content": " for the teacher policies and " + }, + { + "bbox": [ + 55, + 552, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\{1,2,4,16\\}" + }, + { + "bbox": [ + 55, + 552, + 296, + 661 + ], + "type": "text", + "content": " for the student policy. The broader observation window for the student policy helps it better distinguish different skills with larger-scale data. Teacher policies employ MLPs, common in physics-based animation [63], while the student policy handles higher-dimensional observations, for which MLPs are less effective. Thus, we use a transformer [82] architecture for sequential modeling [79], as shown in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 670, + 137, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 137, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 137, + 683 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "We evaluate teacher policies on their ability to imitate imperfect HOI references, and assess the entire teacher-" + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 316, + 87, + 555, + 372 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 501, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 501, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 501, + 84 + ], + "type": "text", + "content": "Algorithm 1 Distillation with RL Fine-tuning" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "lines": [ + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": "1: Input: A composite policy " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\pi^{(\\mathrm{T})}" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " integrated from individual teacher policies, student policy parameters " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": ", student value function parameters " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": ", schedule hyperparameter " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " for DAgger, horizon length " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " for PPO \n2: for " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "t = 0,1,2,\\ldots" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " do \n3: for " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " from 1 to " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " do \n4: Sample a variable " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "u \\sim \\mathrm{Uniform}(0,1)" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " \n5: Collect " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "s^{(\\mathrm{T})}, a^{(\\mathrm{T})}" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " from teacher " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\pi^{(\\mathrm{T})}" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " \n6: Obtain the refined reference from " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "s^{(\\mathrm{T})}" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " to define " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "s^{(\\mathrm{S})}" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "r(\\cdot)" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": ", obtain " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "a^{(\\mathrm{S})}" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\pi_{\\phi}^{(\\mathrm{S})}(a^{(\\mathrm{S})}|s^{(\\mathrm{S})})" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": ". \n7: if " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "u \\leq \\frac{t}{\\beta}" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " then ▷ Use the teacher \n8: Given " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "s^{(\\mathrm{S})}" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": ", execute " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "a^{(\\mathrm{S})}" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": ", observe " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "s^{\\prime (\\mathrm{S})}, r" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " \n9: else ▷ Use the student \n10: Given " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "s^{(\\mathrm{S})}" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": ", execute " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "a^{(\\mathrm{T})}" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": ", observe " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "s^{\\prime (\\mathrm{S})}, r" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " \n11: end if \n12: Store the transition " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "(s^{(\\mathrm{S})}, s^{\\prime (\\mathrm{S})}, a^{(\\mathrm{S})}, a^{(\\mathrm{T})}, r)" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " \n13: end for \n14: Update " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " with TD(" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": ") \n15: Compute PPO objective: " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "L(\\psi)" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " \n16: Compute " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "J(\\psi) = \\| a^{(\\mathrm{S})} - a^{(\\mathrm{T})}\\|" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " \n17: Compute the weight: " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "w = \\min \\left(\\frac{t}{\\beta}, 1\\right)" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " \n18: Update " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " by gradient: " + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\nabla_{\\psi}(wL(\\psi) + (1 - w)J(\\psi))" + }, + { + "bbox": [ + 316, + 87, + 555, + 372 + ], + "type": "text", + "content": " \n19: end for" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + } + ], + "index": 10, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 397, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 434 + ], + "type": "text", + "content": "student framework for scalability to large-scale data and zero-shot generalization across various scenarios. Additional experiments are provided in Sec. G of supplementary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 435, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 556, + 639 + ], + "type": "text", + "content": "Datasets. We use the following datasets from InterAct [103]: OMOMO [39], BEHAV [3], HODome [109], IMHD [114], and HIMO [53]. OMOMO, containing 15 objects and approximately 10 hours of data, is our primary dataset for evaluating the full teacher-student distillation framework for its scale. We train 17 teacher policies, one per subject, with subject 14 reserved as the test set and the remaining data used for training the student policy. A small portion of data is discarded after teacher imitation due to severe MoCap errors that could not be corrected (see Sec. F and Sec. H of the supplementary). Additional datasets are used to evaluate teacher policies in various MoCap scenarios with different error levels and interaction types. We focus on highly dynamic motions (Figure 1) and interactions involving multiple body parts (Figure 4), while excluding scenarios such as carrying a bag with a strap, since the simulator [54] used lacks full support for soft bodies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Metrics. We use the following metrics: (i) Success Rate is defined as the proportion of references that the policy successfully imitates at least once, averaged across all references, while (ii) Duration is the time (in seconds) that the imitation is maintained without triggering the interaction early termination conditions introduced in Sec. 3.2. (iii)" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12271" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 298, + 244 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 298, + 244 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 298, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 298, + 244 + ], + "type": "image", + "image_path": "df088beb273b5d0cccb0de723203bf713b2d524bdd406705ed2e3150b9f9258b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 297, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 297, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 297, + 308 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison between PhysHOI [88] (top), the reference motion (middle) from the BEHAVE [3] dataset, and the interaction refined by our teacher trained on it (bottom). InterMimic faithfully imitates the interactions involving multiple body parts while correcting errors in the original reference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 327, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 411 + ], + "type": "text", + "content": "Human Tracking Error " + }, + { + "bbox": [ + 55, + 327, + 296, + 411 + ], + "type": "inline_equation", + "content": "(E_h)" + }, + { + "bbox": [ + 55, + 327, + 296, + 411 + ], + "type": "text", + "content": " measures the per-joint position error (cm) between the simulated and reference human (excluding hand joints for BEHAV [3] and OMOMO [39] due to inaccuracy). (iv) Object Tracking Error " + }, + { + "bbox": [ + 55, + 327, + 296, + 411 + ], + "type": "inline_equation", + "content": "(E_o)" + }, + { + "bbox": [ + 55, + 327, + 296, + 411 + ], + "type": "text", + "content": " measures the per-point position error (cm) between the simulated and reference object. Both errors are averaged over the duration of the imitation in the best-performing trial." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 412, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 555 + ], + "type": "text", + "content": "Baselines. To facilitate fair comparisons, we downgrade our method for teacher policy evaluation to imitate either a single MoCap clip (Figure 4) or multiple clips with a single object (Table 1), enabling direct comparison with PhysHOI [88] and SkillMimic [89] (Sec. 4.1 and 4.2). Due to the lack of established baselines for large-scale HOI imitation, we adapt the following variants for comparison with our student policy (Sec. 4.3): (i) PPO [71] trains an imitation policy from scratch, following [63]. We experiment with both versions, with and without reference distillation; (ii) DAgger [69] distills the student without RL fine-tuning, a process we refer to as policy distillation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "text", + "content": "Implementation Details. The control policies operate at " + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "text", + "content": " and are trained using the Isaac Gym simulator [54]. Teacher policies are implemented as MLPs with hidden layers of sizes 1024, 1024, and 512. The student policy is implemented as a three-layer Transformer encoder with 4 heads, a hidden size of 256, and a feed-forward layer of 512. The critics are also modeled as MLPs with the same architecture as the teacher policies. To integrate the student policy with kinematic generators, including text-to-HOI [62] and future interaction prediction [101], we train these models using reference data distilled by the teacher policies from the OMOMO [39] dataset, following the same train-test split as the student policy training. For the text-to" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 554, + 205 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 554, + 205 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 554, + 205 + ], + "type": "image", + "image_path": "a63d96253bb909ed0735badf9ee59d3011c4179d1bdbe3141f7f0d8f9d0fb973.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 213, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 248 + ], + "type": "text", + "content": "Figure 5. We recover plausible object rotations (bottom) that are challenging for motion capture due to the equivariant geometries of objects, which result in the object sliding on the ground (top)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 340, + 256, + 531, + 320 + ], + "blocks": [ + { + "bbox": [ + 340, + 256, + 531, + 320 + ], + "lines": [ + { + "bbox": [ + 340, + 256, + 531, + 320 + ], + "spans": [ + { + "bbox": [ + 340, + 256, + 531, + 320 + ], + "type": "table", + "html": "
MethodTime↑Eh↓Eo↓
SkillMimic [89]12.27.213.4
InterMimic (Ours) w/o IET40.36.79.9
InterMimic (Ours) w/o PSI36.16.610.2
InterMimic (Ours)42.66.49.2
", + "image_path": "399b547c96adc08b076e973f700ae6b6067520abe5605e3614243ebe8bc1582a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 328, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 384 + ], + "type": "text", + "content": "Table 1. Quantitative comparison between the teacher policy from InterMimic and SkillMimic [89] to imitate data extracted from the BEHAVE [3] dataset involving a single subject interacting with yogamat. We ablate our proposed approach by removing interaction early termination and physical state initialization." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 399, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 460 + ], + "type": "text", + "content": "HOI model, we train it to generate 10 seconds of motion and use 24 generated samples for evaluation, while for future interaction prediction, the model generates 25 future frames given 10 past frames and we use 60 generated samples for evaluation. See Sec. F of the supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 466, + 450, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 450, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 450, + 479 + ], + "type": "text", + "content": "4.1.Quantitative Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 484, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 605 + ], + "type": "text", + "content": "Table 1 shows that the baseline struggles with MoCap imperfections, e.g., incorrect hand positioning, and thus results in clearly shorter tracking durations. In contrast, our method maintains reference tracking for longer durations and produces interactions that closely match the reference. Table 2 shows that our method consistently outperforms baselines in both training data imitation and out-of-distribution generalization, including interactions from the test set and from kinematic generations. We discuss the effectiveness of specific design choices in Sec. 4.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 611, + 443, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 443, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 443, + 624 + ], + "type": "text", + "content": "4.2. Qualitative Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": "Figure 4 shows a representative sequence from the experiment in Table 1, illustrating how our teacher policy corrects interactions that PhysHOI [88] fails to track robustly – our method effectively withstands and corrects incorrect hand positioning and floating contacts in the reference. Beyond obvious errors, our method also rectifies the rotation of symmetric objects that MoCap inaccurately depicts as slid" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12272" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 553, + 131 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 131 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 131 + ], + "type": "table", + "html": "
PPOReference DistillationPolicy DistillationArchitectureOMOMO-Train [39]OMOMO [39]-TestOMOMO [39]-Test (w × 10)HOI-Diff [62]InterDiff [101]
Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓Succ.†Time†Eh↓Eo↓
✓××MLP23.9101.67.215.69.685.37.516.23.971.27.517.90.00.0--6.711.76.216.4
×✓✓54.5139.97.111.054.3140.27.111.215.591.79.313.74.284.810.19.765.027.47.513.4
✓✓×71.7152.88.912.791.6173.78.513.245.8127.69.114.98.3130.910.113.873.328.96.914.4
✓✓✓90.7168.05.59.795.5173.95.411.962.6140.96.614.512.5121.48.612.175.029.16.213.5
✓✓✓Transformer88.8167.06.010.298.1176.55.911.356.8134.76.613.212.5119.08.512.676.729.36.413.3
", + "image_path": "a11b1340716ba76f6944bb90c30be16d59941a9ddaa7802e6b1cf251c2e52ce0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 555, + 162 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 555, + 162 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation of large-scale interaction imitation using OMOMO [39], kinematic generations from HOI-Diff [62], and InterDiff [101]. Additionally, we evaluate on test set when objects with weights ten times greater than those used during training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 180, + 294, + 305 + ], + "blocks": [ + { + "bbox": [ + 57, + 180, + 294, + 305 + ], + "lines": [ + { + "bbox": [ + 57, + 180, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 57, + 180, + 294, + 305 + ], + "type": "image", + "image_path": "eb3cea12521aa060a57d58b0ef9480c512c9960923375cdf38950155543c5029.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 312, + 295, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 357 + ], + "type": "text", + "content": "Figure 6. Zero-shot integration with a text-to-HOI model HOI-Diff [62] (Top), using 'Kick the large box' as the prompt, and an interaction prediction model InterDiff [101] (Bottom), where gray meshes are past states and colored illustrate future generations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 376, + 295, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 473 + ], + "type": "text", + "content": "ing along the ground, shown in Figure 5. Figure 6 presents additional examples that complement Figure 1, demonstrating how our approach integrates with kinematic generators for future interaction prediction and text-to-interaction synthesis. This zero-shot generalization extends to novel objects unseen during training (Figure 7), highlighting the effectiveness of our object geometry and contact-encoded representation, as well as the large-scale training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 149, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 149, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 149, + 492 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 581 + ], + "type": "text", + "content": "Effectiveness of PSI and IET. We conduct an ablation study, as demonstrated in Table 1, comparing the full approach to \"Ours w/o PSI\". The results validate that Physical State Initialization (PSI) is effective by mitigating inaccuracies in the motion capture data. We also observe reduced effectiveness without our interaction early termination, as training often spends rollouts on irrelevant periods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "content": "Effectiveness of Reference Distillation. Compared to directly scaling imitation from MoCap with potential imperfections (line 1 in Table 2), using references refined by the teacher policy (line 3) achieves consistently better performance on all metrics. The improvement is even more pronounced on the test set, where, without reference distillation, the policy struggles with unseen shapes, while retargeting by reference distillation eliminates the difficulty." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Effectiveness of Joint PPO and DAgger Updates. As shown in Table 2, training a policy from scratch (line 3) or relying solely on policy distillation (DAgger, line 2) fails to" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 180, + 553, + 238 + ], + "blocks": [ + { + "bbox": [ + 315, + 180, + 553, + 238 + ], + "lines": [ + { + "bbox": [ + 315, + 180, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 315, + 180, + 553, + 238 + ], + "type": "image", + "image_path": "be2f54312dd20b1d975985e217970ac222668c82f3c43e9a52c986a5ecb9d5ef.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 246, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 269 + ], + "type": "text", + "content": "Figure 7. Zero-shot generalization of our student policy on novel objects from BEHAV [3] and HODome [109]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 288, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 371 + ], + "type": "text", + "content": "achieve optimal performance. While supervised skill learning lays the groundwork, additional PPO fine-tuning is crucial for resolving conflicts among teacher policies. This is important because our subject-based clustering may not effectively distinguish between different interaction patterns, and ambiguity arises when multiple teachers produce different actions for similar motions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "type": "text", + "content": "Effectiveness of Transformer for Policy Learning. From Table 2, we see that using a Transformer policy (line 5) outperforms MLP-based approaches, particularly on the test set and out-of-distribution cases generated by the kinematic model. We attribute this to the Transformer's inductive bias in sequential modeling and its capacity to incorporate longer-term observations, enabling it to handle complex spatio-temporal dependencies more effectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "In this work, we introduce a framework for synthesizing realistic human-object interactions that are both physically grounded and generalizable. Unlike previous methods, our approach leverages a rich repository of imperfect MoCap data to facilitate the learning of various interaction skills across a wide variety of objects. To address inaccuracies in the MoCap data, we propose contact-guided rewards and optimize trajectory collection, enabling teacher policies to recover missing physical details in the original data. These teacher policies are used to train student policies within a distillation framework that combines policy distillation and reference distillation, thus enabling efficient skill scaling. Our approach shows zero-shot generalizability, which effectively bridges the gap between imitation and generative capabilities by integrating with kinematic generation. We believe that this framework can be adapted for whole-body loco-manipulation for real-world robots, enabling them to handle objects with human-like dexterity and nuance." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12273" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 264 + ], + "type": "text", + "content": "Acknowledgments. We thank Wei Yang, Yu-Wei Chao, Arsalan Mousavian, Ankur Handa, Samuel Schulter, Morteza Ziyadi, and Xiaohan Fei for valuable discussions. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Amazon-Illinois Center on AI for Interactive Conversational Experiences, the Toyota Research Institute, the IBM-Illinois Discovery Accelerator Institute, and Snap Inc. This work used computational resources, including the NCSA Delta and DeltaAI and the PTI Jetstream2 supercomputers through allocations CIS230012, CIS230013, and CIS240311 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, as well as the TACC Frontera supercomputer and Amazon Web Services (AWS) through the National Artificial Intelligence Research Resource (NAIRR) Pilot." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 275, + 115, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 115, + 288 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 115, + 288 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 295, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 295, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 295, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 66, + 295, + 294, + 338 + ], + "type": "text", + "content": "[1] Jinseok Bae, Jungdam Won, Donggeun Lim, Cheol-Hui Min, and Young Min Kim. Pmp: Learning to physically interact with environments using part-wise motion priors. In SIGGRAPH, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 341, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 341, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 66, + 341, + 295, + 384 + ], + "type": "text", + "content": "[2] Qingwei Ben, Feiyu Jia, Jia Zeng, Junting Dong, Dahua Lin, and Jiangmiao Pang. HOMIE: Humanoid locomap-. manipulation with isomorphic exoskeleton cockpit. arXiv preprint arXiv:2502.13013, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 386, + 294, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 386, + 294, + 428 + ], + "spans": [ + { + "bbox": [ + 66, + 386, + 294, + 428 + ], + "type": "text", + "content": "[3] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAV: Dataset and method for tracking human object interactions. In CVPR, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 432, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 432, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 66, + 432, + 294, + 463 + ], + "type": "text", + "content": "[4] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. In 3DV, 2024. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 466, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 466, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 66, + 466, + 294, + 497 + ], + "type": "text", + "content": "[5] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In AAAI, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 500, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 500, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 66, + 500, + 294, + 542 + ], + "type": "text", + "content": "[6] Sammy Christen, Muhammed Kocabas, Emre Aksan, Jemin Hwangbo, Jie Song, and Otmar Hilliges. D-grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In CVPR, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 545, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 545, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 66, + 545, + 294, + 577 + ], + "type": "text", + "content": "[7] Enric Corona, Albert Pumarola, Guillem Alenya, and Francesc Moreno-Noguer. Context-aware human motion prediction. In CVPR, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 580, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 580, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 66, + 580, + 294, + 621 + ], + "type": "text", + "content": "[8] Jieming Cui, Tengyu Liu, Nian Liu, Yaodong Yang, Yixin Zhu, and Siyuan Huang. AnySkill: Learning open-vocabulary physical skill for interactive agents. In CVPR, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 66, + 624, + 294, + 666 + ], + "type": "text", + "content": "[9] Ravinder S Dahiya, Giorgio Metta, Maurizio Valle, and Giulio Sandini. Tactile sensing—from humans to humanoids. IEEE transactions on robotics, 26(1):1-20, 2009. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[10] Divyanshu Daiya, Damon Conover, and Aniket Bera. COLLAGE: Collaborative human-agent interaction generation using hierarchical latent diffusion and language models. arXiv preprint arXiv:2409.20502, 2024. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 105 + ], + "type": "text", + "content": "[11] Jeremy Dao, Helei Duan, and Alan Fern. Sim-to-real learning for humanoid box loco-manipulation. In ICRA, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 321, + 107, + 553, + 129 + ], + "type": "text", + "content": "[12] Christian Diller and Angela Dai. CG-HOI: Contact-guided 3d human-object interaction generation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 131, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 131, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 321, + 131, + 553, + 184 + ], + "type": "text", + "content": "[13] Jiawei Gao, Ziqin Wang, Zeqi Xiao, Jingbo Wang, Tai Wang, Jinkun Cao, Xiaolin Hu, Si Liu, Jifeng Dai, and Jiangmiao Pang. CooHOI: Learning cooperative human-object interaction with manipulated object dynamics. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 321, + 186, + 553, + 229 + ], + "type": "text", + "content": "[14] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. IMoS: Intent-driven full-body motion synthesis for human-object interactions. In Computer Graphics Forum, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 232, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 232, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 321, + 232, + 553, + 263 + ], + "type": "text", + "content": "[15] Michael Gleicher. Motion editing with spacetime constraints. In Proceedings of the 1997 symposium on Interactive 3D graphics, 1997. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 265, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 265, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 321, + 265, + 553, + 330 + ], + "type": "text", + "content": "[16] Zhaoyuan Gu, Junheng Li, Wenlan Shen, Wenhao Yu, Zhaoming Xie, Stephen McCrory, Xianyi Cheng, Abdulaziz Shamsah, Robert Griffin, C Karen Liu, et al. Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning. arXiv preprint arXiv:2501.02116, 2025. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 332, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 332, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 321, + 332, + 553, + 364 + ], + "type": "text", + "content": "[17] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 365, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 365, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 321, + 365, + 553, + 397 + ], + "type": "text", + "content": "[18] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 321, + 399, + 553, + 453 + ], + "type": "text", + "content": "[19] Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. OmniH2O: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. arXiv preprint arXiv:2406.08858, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 321, + 456, + 553, + 498 + ], + "type": "text", + "content": "[20] Tairan He, Zhengyi Luo, Wenli Xiao, Chong Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Learning human-to-humanoid real-time whole-body teleoperation. In IROS, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 500, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 500, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 321, + 500, + 553, + 542 + ], + "type": "text", + "content": "[21] Wenkun He, Yun Liu, Ruitao Liu, and Li Yi. SyncDiff: Synchronized motion diffusion for multi-body human-object interaction synthesis. arXiv preprint arXiv:2412.20104, 2024.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 545, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 545, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 321, + 545, + 553, + 577 + ], + "type": "text", + "content": "[22] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 579, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 579, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 321, + 579, + 553, + 611 + ], + "type": "text", + "content": "[23] Yinghao Huang, Omid Taheri, Michael J. Black, and Dimitrios Tzionas. InterCap: Joint markerless 3D tracking of humans and objects in interaction. In GCRP, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 613, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 613, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 321, + 613, + 553, + 644 + ], + "type": "text", + "content": "[24] Inspire-robots. Smaller and higher-precision motion control experts. https://inspire-robots/store/.1,4,5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "type": "text", + "content": "[25] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. CHAIRS: Towards full-body articulated human-object interaction. In ICCV, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "type": "text", + "content": "[26] Nan Jiang, Zimo He, Zi Wang, Hongjie Li, Yixin Chen, Siyuan Huang, and Yixin Zhu. Autonomous character" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12274" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 94 + ], + "type": "text", + "content": "scene interaction synthesis from text instruction. In SIGGRAPH Asia, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 140 + ], + "type": "text", + "content": "[27] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In CVPR, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 141, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 295, + 195 + ], + "type": "text", + "content": "[28] Zhenyu Jiang, Yuqi Xie, Kevin Lin, Zhenjia Xu, Weikang Wan, Ajay Mandlekar, Linxi Fan, and Yuke Zhu. DexMimicGen: Automated data generation for bimanual dexterous manipulation via imitation learning. arXiv preprint arXiv:2410.24185, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 197, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 197, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 197, + 295, + 239 + ], + "type": "text", + "content": "[29] Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. SuperPADL: Scaling language-directed physics-based control with progressive supervised distillation. In SIGGRAPH, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 241, + 295, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 241, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 295, + 284 + ], + "type": "text", + "content": "[30] Hyeonwoo Kim, Sangwon Beak, and Hanbyul Joo. DAViD: Modeling dynamic affordance of 3d objects using pre-trained video diffusion models. arXiv preprint arXiv:2501.08333, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 330 + ], + "type": "text", + "content": "[31] Jeonghwan Kim, Jisoo Kim, Jeonghyeon Na, and Hanbyul Joo. ParaHome: Parameterizing everyday home activities towards 3d generative modeling of human-object interactions. arXiv preprint arXiv:2401.10232, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "type": "text", + "content": "[32] Yeonjoon Kim, Hangil Park, Seungbae Bang, and Sung-Hee Lee. Retargeting human-object interaction to virtual avatars. IEEE transactions on visualization and computer graphics, 22(11):2405-2412, 2016. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 376, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 295, + 398 + ], + "type": "text", + "content": "[33] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 399, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 295, + 431 + ], + "type": "text", + "content": "[34] Dieter Kraft. Algorithm 733: Tomp-fortran modules for optimal control calculations. ACM Transactions on Mathematical Software (TOMS), 20(3):262-281, 1994. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 433, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 295, + 476 + ], + "type": "text", + "content": "[35] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. NIFTY: Neural object interaction fields for guided human motion synthesis. In CVPR, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 477, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 295, + 510 + ], + "type": "text", + "content": "[36] Jiye Lee and Hanbyul Joo. Locomotion-Action-Manipulation: Synthesizing human-scene interactions in complex 3d environments. In ICCV, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 510, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 295, + 543 + ], + "type": "text", + "content": "[37] Kang Hoon Lee, Myung Geol Choi, and Jehee Lee. Motion patches: building blocks for virtual environments annotated with motion data. In SIGGRAPH. 2006. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 544, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 295, + 567 + ], + "type": "text", + "content": "[38] Yoonsang Lee, Sungeun Kim, and Jehee Lee. Data-driven biped control. In SIGGRAPH. 2010. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 568, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 295, + 601 + ], + "type": "text", + "content": "[39] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 601, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 295, + 635 + ], + "type": "text", + "content": "[40] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. In ECCV, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 635, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 295, + 678 + ], + "type": "text", + "content": "[41] Jinhan Li, Yifeng Zhu, Yuqi Xie, Zhenyu Jiang, Mingyo Seo, Georgios Pavlakos, and Yuke Zhu. OKAMI: Teaching humanoid robots manipulation skills through single video imitation. In CoRL, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 62, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 295, + 713 + ], + "type": "text", + "content": "[42] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. In WACV, 2024. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 137 + ], + "type": "text", + "content": "[43] Fukang Liu, Zhaoyuan Gu, Yilin Cai, Ziyi Zhou, Shijie Zhao, Hyunyoung Jung, Sehoon Ha, Yue Chen, Danfei Xu, and Ye Zhao. Opt2skill: Imitating dynamically-feasible whole-body trajectories for versatile humanoid loco-manipulation. arXiv preprint arXiv:2409.20514, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 139, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 139, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 320, + 139, + 553, + 182 + ], + "type": "text", + "content": "[44] Libin Liu and Jessica Hodgins. Learning to schedule control fragments for physics-based characters using deep q-learning. ACM Transactions on Graphics (TOG), 36(3): 1-14, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 183, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 183, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 320, + 183, + 553, + 227 + ], + "type": "text", + "content": "[45] Yunze Liu, Changxi Chen, Chenjing Ding, and Li Yi. Phys-Reaction: Physically plausible real-time humanoid reaction synthesis via forward dynamics guided 4d imitation. In ACMMM, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 553, + 271 + ], + "type": "text", + "content": "[46] Yun Liu, Bowen Yang, Licheng Zhong, He Wang, and Li Yi. Mimicking-bench: A benchmark for generalizable humanoid-scene interaction learning via human mimicking. arXiv preprint arXiv:2412.17730, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 272, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 272, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 320, + 272, + 553, + 315 + ], + "type": "text", + "content": "[47] Yun Liu, Haolin Yang, Xu Si, Ling Liu, Zipeng Li, Yuxiang Zhang, Yebin Liu, and Li Yi. Taco: Benchmarking generalizable bimanual tool-action-object understanding. In CVPR, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 320, + 316, + 553, + 360 + ], + "type": "text", + "content": "[48] Yun Liu, Chengwen Zhang, Ruofan Xing, Bingda Tang, Bowen Yang, and Li Yi. Core4d: A 4d human-object-human interaction dataset for collaborative object rearrangement. In CVPR, 2025. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 361, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 361, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 320, + 361, + 553, + 414 + ], + "type": "text", + "content": "[49] Jintao Lu, He Zhang, Yuting Ye, Takaaki Shiratori, Sebastian Starke, and Taku Komura. CHOICE: Coordinated human-object interaction in cluttered environments for pick-and-place actions. arXiv preprint arXiv:2412.06702, 2024.3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 415, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 415, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 320, + 415, + 553, + 448 + ], + "type": "text", + "content": "[50] Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In ICCV, 2023. 2, 4, 5, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 449, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 449, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 320, + 449, + 553, + 481 + ], + "type": "text", + "content": "[51] Zhengyi Luo, Jinkun Cao, Sammy Christen, Alexander Winkler, Kris Kitani, and Weipeng Xu. Grasping diverse objects with simulated humanoids. In NeurIPS, 2024. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 320, + 482, + 553, + 536 + ], + "type": "text", + "content": "[52] Zhengyi Luo, Jiashun Wang, Kangni Liu, Haotian Zhang, Chen Tessler, Jingbo Wang, Ye Yuan, Jinkun Cao, Zihui Lin, Fengyi Wang, et al. SMPLOlympics: Sports environments for physically simulated humanoids. arXiv preprint arXiv:2407.00187, 2024.3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 320, + 537, + 553, + 581 + ], + "type": "text", + "content": "[53] Xintao Lv, Liang Xu, Yichao Yan, Xin Jin, Congsheng Xu, Shuwen Wu, Yifan Liu, Lincheng Li, Mengxiao Bi, Wenjun Zeng, et al. HIMO: A new benchmark for full-body human interacting with multiple objects. In ECCV, 2024. 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 320, + 582, + 553, + 635 + ], + "type": "text", + "content": "[54] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. In NeurIPS, 2021. 6, 7, 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 635, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 635, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 320, + 635, + 553, + 690 + ], + "type": "text", + "content": "[55] Josh Merel, Saran Tunyasuvunakool, Arun Ahuja, Yuval Tassa, Leonard Hasenclever, Vu Pham, Tom Erez, Greg Wayne, and Nicolas Heess. Catch & carry: reusable neural controllers for vision-guided whole-body tasks. ACM Transactions on Graphics (TOG), 39(4):39-1, 2020. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "type": "text", + "content": "[56] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12275" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "text", + "content": "Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 295, + 149 + ], + "type": "text", + "content": "[57] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 3DV, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 150, + 294, + 194 + ], + "type": "text", + "content": "[58] Liang Pan, Zeshi Yang, Zhiyang Dou, Wenjia Wang, Buzhen Huang, Bo Dai, Taku Komura, and Jingbo Wang. TokenHSI: Unified synthesis of physical human-scene interactions through task tokenization. In CVPR, 2025. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 195, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 195, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 195, + 295, + 239 + ], + "type": "text", + "content": "[59] Soohwan Park, Hoseok Ryu, Seyoung Lee, Sunmin Lee, and Jehee Lee. Learning predict-and-simulate policies from unorganized human motion data. ACM Transactions on Graphics (TOG), 38(6):1-11, 2019. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "text", + "content": "[60] Sungjae Park, Seungho Lee, Mingi Choi, Jiye Lee, Jeonghwan Kim, Jisoo Kim, and Hanbyul Joo. Learning to transfer human hand skills for robot manipulations. arXiv preprint arXiv:2501.04169, 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 294, + 327 + ], + "type": "text", + "content": "[61] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 371 + ], + "type": "text", + "content": "[62] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. HOI-Diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 3, 7, 8, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 372, + 295, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 295, + 425 + ], + "type": "text", + "content": "[63] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG), 37(4):1-14, 2018. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 426, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 295, + 470 + ], + "type": "text", + "content": "[64] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4):1-20, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 472, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 295, + 514 + ], + "type": "text", + "content": "[65] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Transactions On Graphics (TOG), 41(4):1-17, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 515, + 295, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 515, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 515, + 295, + 548 + ], + "type": "text", + "content": "[66] Haziq Razali and Yiannis Demiris. Action-conditioned generation of bimanual object manipulation sequences. In AAAI, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 548, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 295, + 602 + ], + "type": "text", + "content": "[67] Daniele Reda, Jungdam Won, Yuting Ye, Michiel van de Panne, and Alexander Winkler. Physics-based motion retargeting from sparse inputs. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 6(3):1-19, 2023. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 604, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 604, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 604, + 295, + 645 + ], + "type": "text", + "content": "[68] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6), 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 647, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 713 + ], + "type": "text", + "content": "[69] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "text", + "content": "[70] John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation. In ICLR, 2016. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 119, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 119, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 320, + 119, + 553, + 152 + ], + "type": "text", + "content": "[71] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 154, + 553, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 154, + 553, + 197 + ], + "spans": [ + { + "bbox": [ + 320, + 154, + 553, + 197 + ], + "type": "text", + "content": "[72] Carmelo Sferrazza, Dun-Ming Huang, Xingyu Lin, Youngwoon Lee, and Pieter Abbeel. Humanoidbench: Simulated humanoid benchmark for whole-body locomotion and manipulation. In RSS, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 199, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 199, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 320, + 199, + 553, + 253 + ], + "type": "text", + "content": "[73] Wenfeng Song, Xinyu Zhang, Shuai Li, Yang Gao, Aimin Hao, Xia Hou, Chenglizhao Chen, Ning Li, and Hong Qin. HOIAnimator: Generating text-prompt human-object animations using novel perceptive diffusion models. In CVPR, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 256, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 256, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 320, + 256, + 553, + 289 + ], + "type": "text", + "content": "[74] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 291, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 291, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 320, + 291, + 553, + 334 + ], + "type": "text", + "content": "[75] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39(4):54-1, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 336, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 336, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 320, + 336, + 553, + 369 + ], + "type": "text", + "content": "[76] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In ECCV, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 372, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 372, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 320, + 372, + 553, + 404 + ], + "type": "text", + "content": "[77] Omid Taheri, Vasileios Choutas, Michael J Black, and Dimitrios Tzionas. GOAL: Generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 406, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 406, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 320, + 406, + 553, + 449 + ], + "type": "text", + "content": "[78] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. Calm: Conditional adversarial latent models for directable virtual characters. In SIGGRAPH, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 452, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 452, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 320, + 452, + 553, + 496 + ], + "type": "text", + "content": "[79] Chen Tessler, Yunrong Guo, Ofir Nabati, Gal Chechik, and Xue Bin Peng. Maskedmimic: Unified physics-based character control through masked motion inpainting. ACM Transactions on Graphics (TOG), 43(6):1-21, 2024. 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 498, + 553, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 498, + 553, + 551 + ], + "spans": [ + { + "bbox": [ + 320, + 498, + 553, + 551 + ], + "type": "text", + "content": "[80] Guy Tevet, Sigal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. CLoSD: Closing the loop between simulation and diffusion for multi-task character control. In ICLR, 2025. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 554, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 554, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 320, + 554, + 553, + 576 + ], + "type": "text", + "content": "[81] Unittree. Unittree g1 humanoid agent ai avatar. https://www-unitree.com/g1/.1,4,5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 578, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 578, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 320, + 578, + 553, + 620 + ], + "type": "text", + "content": "[82] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 624, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 624, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 320, + 624, + 553, + 656 + ], + "type": "text", + "content": "[83] Ruben Villegas, Duygu Ceylan, Aaron Hertzmann, Jimei Yang, and Jun Saito. Contact-aware retargeting of skinned motion. In ICCV, 2021. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 553, + 713 + ], + "type": "text", + "content": "[84] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 2022. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12276" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 58, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 105 + ], + "type": "text", + "content": "[85] Jiashun Wang, Jessica Hodgins, and Jungdam Won. Strategy and skill learning for physics-based table tennis animation. In SIGGRAPH, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 294, + 150 + ], + "type": "text", + "content": "[86] Ruocheng Wang, Pei Xu, Haochen Shi, Elizabeth Schumann, and C Karen Liu. Fürelise: Capturing and physically synthesizing hand motion of piano performance. In SIGGRAPH Asia, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 152, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 152, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 152, + 294, + 205 + ], + "type": "text", + "content": "[87] Wenjia Wang, Liang Pan, Zhiyang Dou, Zhouyingcheng Liao, Yuke Lou, Lei Yang, Jingbo Wang, and Taku Komura. SIMS: Simulating human-scene interactions with real world script planning. arXiv preprint arXiv:2411.19921, 2024.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 62, + 208, + 294, + 251 + ], + "type": "text", + "content": "[88] Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. PhysHOI: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393, 2023. 3, 7, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 253, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 253, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 62, + 253, + 294, + 296 + ], + "type": "text", + "content": "[89] Yinhuai Wang, Qihan Zhao, Runyi Yu, Ailing Zeng, Jing Lin, Zhengyi Luo, Hok Wai Tsui, Jiwen Yu, Xiu Li, Qifeng Chen, et al. SkillMimic: Learning reusable basketball skills from demonstrations. In CVPR, 2025. 3, 7, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 298, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 298, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 298, + 294, + 330 + ], + "type": "text", + "content": "[90] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. InterControl: Generate human motion interactions by controlling every joint. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 332, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 332, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 332, + 294, + 364 + ], + "type": "text", + "content": "[91] Jungdam Won and Jehee Lee. Learning body shape variation in physics-based characters. ACM Transactions on Graphics (TOG), 38(6):1-12, 2019. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 294, + 408 + ], + "type": "text", + "content": "[92] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. A scalable approach to control diverse behaviors for physically simulated characters. ACM Transactions on Graphics (TOG), 39(4):33-1, 2020. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 294, + 453 + ], + "type": "text", + "content": "[93] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. THOR: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 455, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 294, + 487 + ], + "type": "text", + "content": "[94] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. SAGA: Stochastic whole-body grasping with contact. In ECCV, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 489, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 294, + 521 + ], + "type": "text", + "content": "[95] Zhen Wu, Jiaman Li, Pei Xu, and C Karen Liu. Human-object interaction from human-level instructions. arXiv preprint arXiv:2406.17840, 2024.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 523, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 294, + 566 + ], + "type": "text", + "content": "[96] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In ICLR, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 568, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 294, + 600 + ], + "type": "text", + "content": "[97] Xianghui Xie, Jan Eric Lenssen, and Gerard Pons-Moll. InterTrack: Tracking human object interaction without object templates. In 3DV, 2024. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 602, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 294, + 634 + ], + "type": "text", + "content": "[98] Zhaoming Xie, Sebastian Starke, Hung Yu Ling, and Michiel van de Panne. Learning soccer juggling skills with layer-wise mixture-of-experts. In SIGGRAPH, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 636, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 294, + 678 + ], + "type": "text", + "content": "[99] Zhaoming Xie, Jonathan Tseng, Sebastian Starke, Michiel van de Panne, and C Karen Liu. Hierarchical planning and control for box loco-manipulation. arXiv preprint arXiv:2306.09532, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 58, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 681, + 294, + 712 + ], + "type": "text", + "content": "[100] Pei Xu and Ruocheng Wang. Synchronize dual hands for physics-based dexterous guitar playing. In SIGGRAPH Asia, 2024. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 58, + 73, + 553, + 688 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[101] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 1, 3, 7, 8, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "text", + "content": "[102] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. In NeurIPS, 2024. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "type": "text", + "content": "[103] Sirui Xu, Dongting Li, Yucheng Zhang, Xiyan Xu, Qi Long, Ziyin Wang, Yunzhi Lu, Shuchang Dong, Hezi Jiang, Akshit Gupta, Yu-Xiong Wang, and Liang-Yan Gui. InterAct: Advancing large-scale versatile 3d human-object interaction generation. In CVPR, 2025. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 209, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 240 + ], + "type": "text", + "content": "[104] Zeshi Yang, Kangkang Yin, and Libin Liu. Learning to use chopsticks in diverse gripping styles. ACM Transactions on Graphics (TOG), 41(4):1-17, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 274 + ], + "type": "text", + "content": "[105] Wenhao Yu, Greg Turk, and C Karen Liu. Learning symmetric and low-energy locomotion. ACM Transactions on Graphics (TOG), 37(4):1-12, 2018. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 276, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 307 + ], + "type": "text", + "content": "[106] Ye Yuan and Kris Kitani. Residual force control for agile human behavior imitation and extended motion synthesis. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 310, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 353 + ], + "type": "text", + "content": "[107] Yanjie Ze, Zixuan Chen, Wenhao Wang, Tianyi Chen, Xiaolin He, Ying Yuan, Xue Bin Peng, and Jiajun Wu. Generalizable humanoid manipulation with improved 3d diffusion policies. arXiv preprint arXiv:2410.10803, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 355, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 407 + ], + "type": "text", + "content": "[108] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "type": "text", + "content": "[109] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. NeuralDome: A neural modeling pipeline on multi-view human-object interactions. In CVPR, 2023. 3, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 456, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 507 + ], + "type": "text", + "content": "[110] Juze Zhang, Jingyan Zhang, Zining Song, Zhanhe Shi, Chengfeng Zhao, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Hoi-m^3: Capture multiple humans and objects interaction within contextual environment. In CVPR, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 511, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 511, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 317, + 511, + 553, + 542 + ], + "type": "text", + "content": "[111] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. COUCH: Towards controllable human-chair interactions. In ECCV, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 544, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 544, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 553, + 598 + ], + "type": "text", + "content": "[112] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Ilya Petrov, Vladimir Guzov, Helisa Dhamo, Eduardo Pérez-Pellitero, and Gerard Pons-Moll. FORCE: Dataset and method for intuitive physics guided human-object interaction. In 3DV, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 601, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 643 + ], + "type": "text", + "content": "[113] Yunbo Zhang, Deepak Gopinath, Yuting Ye, Jessica Hodgins, Greg Turk, and Jungdam Won. Simulation and targeting of complex multi-character interactions. In SIGGRAPH, 2023. 3, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 646, + 553, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 646, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 317, + 646, + 553, + 688 + ], + "type": "text", + "content": "[114] Chengfeng Zhao, Juze Zhang, Jiashen Du, Ziwei Shan, Junye Wang, Jingyi Yu, Jingya Wang, and Lan Xu. I'M HOI: Inertia-aware monocular capture of 3d human-object interactions. In CVPR, 2024. 3, 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12277" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/bba3c5bb-a0b8-4050-9fca-51eeaf9696e0_content_list.json b/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/bba3c5bb-a0b8-4050-9fca-51eeaf9696e0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4e0356ef32985a0b8b5f135bc3f0e6778501ae52 --- /dev/null +++ b/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/bba3c5bb-a0b8-4050-9fca-51eeaf9696e0_content_list.json @@ -0,0 +1,1548 @@ +[ + { + "type": "text", + "text": "InteractAnything: Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing", + "text_level": 1, + "bbox": [ + 119, + 130, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinlu Zhang $^{1,2}$ , Yixin Chen $^{2\\boxtimes}$ , Zan Wang $^{2,3}$ , Jie Yang $^{2,4}$ , Yizhou Wang $^{1,5,6,7\\boxtimes}$ , Siyuan Huang $^{2\\boxtimes}$", + "bbox": [ + 120, + 202, + 875, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "indicates corresponding authors", + "bbox": [ + 395, + 223, + 601, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Center on Frontiers of Computing Studies, School of Computer Science, Peking University", + "bbox": [ + 223, + 239, + 774, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ State Key Laboratory of General Artificial Intelligence, BIGAI $^{3}$ Beijing Institute of Technology", + "bbox": [ + 199, + 257, + 795, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 The Chinese University of Hong Kong, Shenzhen 5 Nat'l Eng. Research Center of Visual Technology, Peking University", + "bbox": [ + 127, + 275, + 867, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ Institute for AI, Peking University $^{7}$ State Key Laboratory of General Artificial Intelligence, Peking University", + "bbox": [ + 156, + 292, + 836, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/13c688d3961fa06fb80cf67c7892f1ba4d1ca5f354ac86ad64e501a827ca28a5.jpg", + "image_caption": [ + "A person holds/pulls/lies on the chair...", + "(a) Diverse human interactions.", + "Figure 1. 3D human object interaction synthesis by InteractAnything. Given a simple text description with goal interaction and any object mesh as input, our method enables the generation of diverse, natural, detailed, and novel interactions for open-set 3D objects in a zero-shot manner. The orange and green boxes of (b) indicate detailed contact poses from different views." + ], + "image_footnote": [], + "bbox": [ + 98, + 352, + 344, + 530 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f882f26ba7af8d8816b8e73c4f291654525d23f4685d548ed3e15673a09505fc.jpg", + "image_caption": [ + "A person grasps the", + "backpack/table/ball/knife...", + "(b) Detailed interactions synthesis of open-set objects." + ], + "image_footnote": [], + "bbox": [ + 372, + 356, + 633, + 529 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c2b3b4f1fd64f2cc3bd30e725e65a4451665dec03e58629b40a7f0a26da591ce.jpg", + "image_caption": [ + "A person lifts/rides/criddles the", + "tesla car/motorcycle/horse/baby...", + "(c) Novel interactions given generative objects." + ], + "image_footnote": [], + "bbox": [ + 697, + 362, + 869, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 609, + 326, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in 3D human-aware generation have made significant progress. However, existing methods still struggle with generating novel Human Object Interaction (HOI) from text, particularly for open-set objects. We identify three main challenges of this task: precise human-object relation reasoning, affordance parsing for any object, and detailed human interaction pose synthesis aligning description and object geometry. In this work, we propose a novel zero-shot 3D HOI generation framework without training on specific datasets, leveraging the knowledge from largescale pre-trained models. Specifically, the human-object relations are inferred from large language models (LLMs) to initialize object properties and guide the optimization process. Then we utilize a pre-trained 2D image diffusion model to parse unseen objects and extract contact points, avoiding the limitations imposed by existing 3D asset knowledge. The initial human pose is generated by sam", + "bbox": [ + 88, + 643, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pling multiple hypotheses through multi-view SDS based on the input text and object geometry. Finally, we introduce a detailed optimization to generate fine-grained, precise, and natural interaction, enforcing realistic 3D contact between the 3D object and the involved body parts, including hands in grasping. This is achieved by distilling human-level feedback from LLMs to capture detailed human-object relations from the text instruction. Extensive experiments validate the effectiveness of our approach compared to prior works, particularly in terms of the fine-grained nature of interactions and the ability to handle open-set 3D objects. Project page: jinluzhang.site/projects/interactanything.", + "bbox": [ + 511, + 609, + 908, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 827, + 645, + 842 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Synthesizing the interaction of humans and any given object holds great significance across diverse applications, including AR/VR, computer simulation, and animation. Tradi-", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7015", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tionally, creating high-quality 3D human object interaction (HOI) assets requires expensive, lab-controlled data capture systems and significant human effort. Although some datasets [2, 16, 20, 21, 26, 53] have collected these data, the models [26, 27, 41, 54, 58, 60] trained on them can only generate interactions for specific object categories. These methods struggle to handle the diversity of object geometries, scales, and interactions at the same time, making text-driven zero-shot HOI synthesis a challenging and highly demanding task.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given recent advances in text-to-3D techniques [6, 29, 56], e.g., leveraging Score Distillation Sampling (SDS) [43] to generate 3D content from text, a natural question is whether this paradigm can be applied to zero-shot HOI synthesis. However, directly integrating 3D humans and objects into a scene without understanding the complex relationships between them remains challenging. Recent works [7, 9, 67] are proposed to generate realistic 3D HOIs. Among them, InterFusion [9] uses Neural Radiance Fields (NeRF) [37] for individual human and object representations but struggles with precise interaction area, leading to suboptimal and unrealistic HOIs. DreamHOI [67] optimizes on both mesh and NeRF representations through bidirectional rendering but faces challenges in ensuring accurate contact points and relative scales due to solely optimizing human states while keeping the given object static, leading to limited interactions.", + "bbox": [ + 91, + 244, + 483, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we first identify the core challenge of zero-shot 3D HOI generation as accurately capturing human-object relations, contact areas, and global consistency, which are not typically presented in single-object or non-contact generation scenarios. These challenges prevent the generation of high-quality, geometrically precise, and textually consistent results. Furthermore, while large-scale pre-trained models, e.g., 2D diffusion models [42, 47] and large language models (LLMs) [1, 10, 39], cannot fully capture the complexity of human object interactions (including body dynamics, expressive details, and object geometry), they can offer valuable prior knowledge of common interaction patterns based on textual descriptions. Consequently, we argue that the solution for effective generation is to leverage this diverse prior knowledge to guide the optimization of interactions.", + "bbox": [ + 89, + 505, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following the above insights, we decompose the complex zero-shot 3D HOI task into these stages: 1) human-object relation reasoning and initialization, 2) open-set object contact affordance parsing, and 3) human pose synthesis aligning text and object geometry. We first leverage the LLM as a human-level feedback provider, which reasons and unravels complex semantic interaction relationships, e.g., relative position and orientation, from simple text input. It initializes human and object states and designs a series of optimization functions to generate detailed inter", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "actions. The initial human pose is optimized by multi-view SDS loss for interaction in line with the given textual description and object geometry.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To parse the 3D object affordance, we utilize initialized human-object relation and sample adaptive inpainting masks during the diffusion process to obtain the contact information. This enables the extraction of contact probabilities on the object, which are then used to segment and define the weighted dynamic contact regions. Finally, we design a coarse-to-fine optimization based on force closure [8, 32] to generate expressive, diverse, and physically plausible interactions, allowing for more accurate and detailed contact. To summarize, our contributions are:", + "bbox": [ + 511, + 137, + 906, + 287 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose InteractAnything for zero-shot 3D HOI generation, which effectively distills human and object interactions from LLM and pre-trained diffusion models to synthesize diverse, detailed, and novel interactions.", + "- We design a novel open-set object affordance parsing, which generates contact maps on 3D geometry by utilizing adaptive inpainting and potential contact distribution.", + "- Our model leverages both text and object geometry to drive human pose synthesis, enabling realistic and contextually appropriate interactions for open-set objects. Results show that our synthesized pose accurately reflects the semantic meaning of the input while respecting the spatial constraints of the object, leading to improved performance, as shown in Figure 1." + ], + "bbox": [ + 511, + 290, + 903, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 518, + 653, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-driven 3D Content Generation", + "text_level": 1, + "bbox": [ + 513, + 544, + 821, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The generation of 3D content from text has seen rapid advancements, primarily driven by diffusion-based models that build on text-to-2D image generation approaches [45, 46, 49, 50]. One pioneering approach, DreamFusion [43], introduced Score Distillation Sampling (SDS), which leverages the pre-trained 2D diffusion model to optimize 3D content through techniques like NeRF (Neural Radiance Fields) [37] and 3D Gaussian Splatting [24]. A significant area of research aims to address the complexity of 3D human generation. For example, recent methods have been proposed for generating controllable 3D human avatar [4, 17, 18, 33, 57, 65]. However, these works focus on diverse and high-quality appearance but animate human avatars with given human poses or motion, which do not satisfy 3D HOI requirements in human pose synthesis. The proliferation of large 3D datasets has also catalyzed progress, leading to more sophisticated models like Zero-1-to-3 [31], Zero $123++$ [51], and MVDream [52], which employ 2D diffusion models to generate multi-view consistent images that serve as inputs for efficient 3D generation systems, such as SyncDreamer [34] and Wonder3D [35]. Despite these advancements, generating complex HOI scenes,", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7016", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "especially those involving interactions between humans and objects, remains challenging.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Data-Driven Human Object Interaction", + "text_level": 1, + "bbox": [ + 89, + 142, + 431, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous approaches [26, 58-60] primarily train generative models to produce static human poses or dynamic motions in response to objects or scenes using curated HOI datasets. Recent works focus on dynamic motion generation performance conditioned on encoded object features but rely heavily on specialized datasets, limiting generalization to diverse objects. For example, OMOMO [26] generates human motion through diffusion models but requires object motion as input. InterDiff [58] predicts full-body interactions based on prior HOI context but is limited to the prediction task, restricting flexibility in new scenarios. In contrast, our method jointly synthesizes human and object states and adapts to novel objects without extensive, object-specific training by extracting interaction patterns directly from 2D diffusion models.", + "bbox": [ + 89, + 167, + 482, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the other hand, optimization-based methods reconstruct realistic human poses and transformations from images. PHOSA [64] optimizes human object relations using depth-aware loss functions. PROX [16] encodes contact probabilities and semantic scene labels to position humans within scenes. CHORUS [15] synthesizes multi-view images with a pre-trained generative model to learn 3D spatial relationships. By extracting knowledge from LLM to infer human poses, the method in [55] queries contact information to reconstruct 3D human-object interactions. Compared to these methods, our method leverages pre-trained diffusion models for accurate parsing of object affordance and enables fine-grained interactions.", + "bbox": [ + 89, + 398, + 482, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Zero-shot 3D Compositional Generation", + "text_level": 1, + "bbox": [ + 89, + 617, + 434, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve the compositional scene generation, Combo-Verse [7] proposes a spatially-aware SDS loss, generating the final 3D scene through single-object reconstruction and multi-object composition. GraphDreamer [12] leverages scene graphs to guide generation by converting the semantic information into a global textual description to enable text-driven compositional scene generation. These methods focus on the reasonable arrangement of multiple objects but ignore the contact generation in HOI scenes. InterFusion [9] and DreamHOI [67] present this challenge and then try to address using SMPLs [36, 48] as body model prior. DreamHOI [67] designs the SDS guidance mixture to separately render human interactions using different diffusion models to improve human modeling quality. However, it cannot ensure precise contact during the denoising process. Compared to these methods, our work focuses on addressing contact areas parsing on open-set objects.", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 643, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Definition and Overview. Given a text prompt $\\epsilon$ that describes the desired HOI, e.g., \"A person grasps the chair\". Our goal is to synthesize natural human mesh $\\nu_{H}$ and object mesh $\\nu_{O}$ that align with $a$ and $\\epsilon$ . As shown in Figure 2, we first obtain initial human-object relations from LLM based on text (Section 3.2). For object optimization, we initialize object rotation $r_o$ , translation $t_o$ , scale $s_o$ , state $s$ , based on affordance map $\\mathcal{M}$ (Section 3.3). We also synthesize the initial human pose based on the input text and object geometry (Section 3.4). Finally, we use SMPL-H as representation and optimize pose parameter $\\theta_b, \\theta_h$ , global translation $t_b$ , and object rotation and translation $r_o, t_o$ to generate detailed and diverse HOI (Section 3.5).", + "bbox": [ + 511, + 114, + 906, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary Knowledge", + "text_level": 1, + "bbox": [ + 511, + 319, + 733, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human Body Model. We use the SMPL-H [36, 48], a parametric model that includes both body and hand parameters, to represent the human body and hand articulation. The body pose is defined by body parameters $\\theta_{\\mathrm{b}} \\in \\mathbb{R}^{3(K_b + 1)}$ and the hand pose is represented by additional MANO parameters $\\theta_{\\mathrm{h}} \\in \\mathbb{R}^{3K_h}$ , capturing 15 joints per hand. The shape parameters are represented by $\\beta \\in \\mathbb{R}^{10}$ , controlling the deformation of the body and hand template mesh to target body shapes. Given the combined pose parameters $\\{\\theta_{\\mathrm{b}}, \\theta_{\\mathrm{h}}\\}$ and shape parameters $\\beta$ , SMPL-H computes mesh vertices $\\mathcal{V} \\in \\mathbb{R}^{V \\times 3}$ using linear blend skinning function $f$ :", + "bbox": [ + 511, + 340, + 906, + 507 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} (\\beta , \\theta_ {b}, \\theta_ {\\mathrm {h}}) = f \\left(T _ {p} \\left(\\beta , \\theta_ {b}, \\theta_ {\\mathrm {h}}\\right), J (\\beta), \\theta , \\mathcal {W}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 513, + 903, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $J$ represents joint locations, $\\mathcal{W} \\in \\mathbb{R}^{V \\times (K)}$ is skimming weights with all $K$ joints.", + "bbox": [ + 511, + 536, + 903, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Score Distillation Sampling (SDS). SDS leverages pretrained diffusion models to guide the synthesis of 3D scenes by aligning NeRF with a text prompt. It operates by iteratively adjusting NeRF parameters, denoted as $\\theta$ , based on gradients computed to align the generated images with the textual input. Given a noisy 2D image $x_{t}$ (obtained by adding noise $\\epsilon$ to a clean NeRF-rendered image $x$ ), a textual prompt $y$ , and a noise level $t$ , SDS calculates the gradient of the loss $\\mathcal{L}_{\\mathrm{SDS}}$ with respect to $\\theta$ as follows:", + "bbox": [ + 511, + 568, + 906, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (x) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) (\\hat {\\epsilon} (x _ {t}; y, t) - \\epsilon) \\frac {\\partial x}{\\partial \\theta} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 710, + 867, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\epsilon}(x_t; y, t)$ represents the diffusion model predicted noise given the noisy image $x_t$ , prompt $y$ , and noise level $t$ , $w(t)$ is a weighting function that adjusts the influence of the gradient based on the noise level $t$ . This gradient is used to update $\\theta$ , guiding the NeRF model toward producing images that match the textual description. For instance, in DreamFusion [43], this process enables text-to-3D generation by repeatedly refining the NeRF parameters $\\theta$ from a random initialization, iteratively producing images that align more closely with the target prompt.", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7017", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a0f0efb2da3f61d82118f3a4d802002ff84417a895f28f2e1bf39c6eff39f3d6.jpg", + "image_caption": [ + "Figure 2. Framework of InteractAnything. Given a text description and any object mesh as input, our approach begins by querying LLM to infer precise human-object relationships, which are used to initialize object properties. Next, we analyze the contact affordance of the object geometry. The human pose is synthesized using a pre-trained 2D diffusion model, guided by multi-view SDS loss and the designed spatial constraint. Finally, based on the targeted object contact areas and a plausible human pose, we perform expressive HOI optimization to synthesize realistic and contact-accurate 3D human-object interactions." + ], + "image_footnote": [], + "bbox": [ + 112, + 90, + 883, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. LLM-guided Initialization", + "text_level": 1, + "bbox": [ + 89, + 484, + 328, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To infer diverse intended human object interactions for a given object $\\mathcal{V}_O$ and action $a$ , we query the LLM to provide commonsense feedback. To minimize hallucination, we need to transfer inferred relationships from GPT [39] into specific values that position the human and object meshes accurately, therefore, we enable the language model to select answers from pre-defined options step by step, instead of directly generating responses. For initializing the object rotation $r_o$ , translation $t_o$ , scale $s_o$ , and state $s$ , the language model selects the relative position of the human to the object from defined options, which is then mapped to $t_o$ based on the combined size of the human and object. We set all input objects to have $y$ -up and $x$ -front orientations, with $r_o$ adjusted based on which side of the object should face the human. The object scale is chosen from 4 defined options and set relative to human size. Lastly, the state $s$ determines if the object should be positioned on the ground based on whether it is meant to move. Besides, we also query interacted semantic body parts $S_h$ and map the labels to human mesh to obtain vertices $\\mathcal{P}_h$ for part-level optimization using annotated human part labels.", + "bbox": [ + 91, + 506, + 483, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Open-set Object Affordance Parsing", + "text_level": 1, + "bbox": [ + 89, + 833, + 406, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain the contact affordance of unseen object geometries given a specified action e.g., holds, we conduct 2D inpainting based on a pre-trained diffusion model to dis", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "till possible human interaction parts through LLM-guided adaptive mask generation. Then we design a 2D probabilistic contact function to model the distribution of the interaction, generating an affordance map to identify and extract the regions for HOI optimization.", + "bbox": [ + 511, + 484, + 906, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2D Adaptive Mask Generation Image inpainting is to fill in missing regions of an image, where the missing regions can be defined by a mask image. To get knowledge of possible interacted parts, we utilize the powerful understanding and generation ability of pre-trained inpainting model [42]. Given 2D projection images of the 3D object from $n_p$ views, we found existing inpainting models cannot ensure complete human in the image (many inpainting results only include interacted body parts, e.g., hands in \"hold\" action). Therefore, different from previous methods [25, 28], we do not randomly sample the inpainting masks on 2D images, instead, our method identifies possible human interaction parts to improve the inpainting success rate by generating inpainting masks from LLM feedback in two forms: (1) full-body mask with respect to object properties $r_o$ , $t_o$ , person position and person body size; (2) body-part inpainting masks for intended interacted body part labels $S_h$ and corresponding human vertices $\\mathcal{P}_h$ .", + "bbox": [ + 511, + 560, + 908, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More specifically, we first create a full-body inpainting mask guided by LLM-inferred initializations $r_o$ , $t_o$ , and body size $s_f$ :", + "bbox": [ + 511, + 832, + 906, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {\\mathrm {f u l l}} ^ {i} = \\mathcal {J} ((\\mathcal {V} _ {H} \\cdot r _ {o} ^ {- 1}), s _ {f}, c ^ {i}) + t _ {o} ^ {- 1} + \\varpi_ {1}, \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 883, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7018", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2a23e3e34d611ce29e880c0b407320d93419734c546e3d75ec0b79aafbd51ae1.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison results with baselines. * indicates we re-implement this method by embedding object mesh into the diffusion process, which follows the pipeline of DreamHOI [67] and our method. More visualization results are presented in supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 116, + 88, + 869, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I}$ is 2D projection transformation with camera parameter $c^i$ , and $\\varpi_{1}$ is the noise term. Similarly, we generate body-part-level inpainting masks using given body part labels $S_{h}$ and corresponding human vertices $\\mathcal{P}_h$ and body part size $s_p$ to help inpainting body parts like hands as the above formulation:", + "bbox": [ + 88, + 518, + 480, + 607 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {\\text {p a r t}} ^ {i} = \\mathcal {J} \\left(\\left(\\mathcal {P} _ {h} \\cdot r _ {o} ^ {- 1}\\right), s _ {p}, c ^ {i}\\right) + t _ {o} ^ {- 1} + \\varpi_ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 616, + 480, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides, We apply the above full-body masks $M_{\\mathrm{full}}$ into the dynamic mask inpainting pipeline [25], which can further refine the initial mask detecting human on intermediate results in the diffusion process.", + "bbox": [ + 88, + 643, + 482, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Affordance Parsing on 2D Probability Function. Directly computing contact probabilities on a 3D object mesh relies heavily on existing 3D human reconstruction methods [3, 13, 22, 30, 61-63] from single-view images. This encounters significant 3D ambiguity and quantization error, which obscures accurate 3D human pose. In contrast to previous methods [15], we apply 2D human pose detector OpenPose [5], performing well at detecting 2D human keypoints and even occluded body parts, to estimate human from inpainting results. Then we introduce a 2D probability function that calculates contact probability according to the distance from detected body keypoints to the multi-view object projection images. First, we downsample the object", + "bbox": [ + 88, + 704, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to $D$ points. Let $\\mathcal{P}_i$ represent the probability map for each view $i$ , the distance-based probability $f_{\\mathrm{afford}}(\\mathbf{p})$ for a point $\\mathbf{p} \\in \\mathcal{V}_O$ on the object surface is defined as:", + "bbox": [ + 511, + 518, + 906, + 564 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {a f f o r d}} ^ {(i)} (\\mathbf {p}) = e ^ {- \\left\\| \\mathbf {d} _ {i} (\\mathbf {p}) \\right\\|}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 571, + 903, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{d}_i(\\mathbf{p})$ denotes the distance from the body part to the object mask in $i$ view. Then we re-project the 2D affordance $f_{\\mathrm{afford}}^{(i)}(\\mathbf{p})$ to 3D space using defined camera parameters, and aggregate the contact probabilities across all $n_p$ views to update the 3D contact probability distribution as follows:", + "bbox": [ + 511, + 599, + 906, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} (\\mathbf {p}) = \\frac {1}{n _ {p}} \\sum_ {i = 1} ^ {n _ {p}} f _ {\\text {a f f o r d}} ^ {(i)} (\\mathbf {p}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 685, + 903, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we normalize $\\mathcal{P}(\\mathbf{p})$ over the entire object surface to obtain contact probability for each point is determined by its projected 2D distances across multiple views, effectively mitigating multi-view ambiguity and enabling accurate estimation of human object contact regions.", + "bbox": [ + 511, + 734, + 906, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Text-Object Driven Human Pose Synthesis", + "text_level": 1, + "bbox": [ + 511, + 816, + 877, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering a brief HOI description, we posit that current 3D generation models, such as DreamFusion [43], are capable of generating human avatars in a variety of poses; however, they do not guarantee precise contact with the object", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7019", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mesh. Thus, we first apply prompt engineering to generate diverse and expressive text descriptions, along with negative prompts, using large language models (LLMs) to mitigate issues like low-quality outputs and missing limbs. Next, we incorporate Multi-view SDS loss [52] to optimize the human pose aligning the interaction description and object geometry. We first scale the input object based on $s_o$ given by the language model. We use the spatial constraint to prevent the human avatar from penetrating the object mesh, ensuring alignment with spatial relations. Specifically, we embed object mesh within NeRF optimization by regularizing the ray sample points in object mesh following previous works [9, 67]. After training, we use the multi-view human pose estimation method [66] to obtain the body parameters $\\theta_b$ and then init hand parameters $\\theta_h$ . We also employ VPoser [40] to maintain reasonable joint rotations.", + "bbox": [ + 89, + 90, + 483, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Expressive HOI Optimization", + "text_level": 1, + "bbox": [ + 89, + 340, + 357, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Global HOI optimization. With the above initialization as input, we first ensure the input object is watertight using Manifold [19]. Then we conduct human-centric optimization to update object properties $(r_o, t_o, s_o)$ as follows:", + "bbox": [ + 89, + 362, + 483, + 425 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c} = \\phi_ {i} L _ {i n t e r} + \\phi_ {n} L _ {n} + \\phi_ {s} L _ {s c a l e} + \\phi_ {p} L _ {p e n e} + \\mathbf {s} ^ {'} \\cdot L _ {g}, (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 431, + 482, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\phi$ is loss weights and interaction loss is designed on. \nverts of the interacted body part $\\mathcal{P}_h,h\\in \\mathcal{V}_H$ and object \ncontact areas $\\mathcal{P}_o,o\\in \\mathcal{M}$ ..", + "bbox": [ + 89, + 460, + 483, + 506 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i n t e r} = \\sum_ {i \\in [ h ], j \\in [ o ]} W _ {\\mathcal {M}} (j) \\cdot f _ {c h a m} \\left(\\mathcal {P} _ {h}, \\mathcal {P} _ {o}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 513, + 482, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W_{\\mathcal{M}}(j)$ indicates probabilities scores in object affordance map $\\mathcal{M}$ of the $j$ -th object vertex and $f_{cham}$ is chamfer distance function. We constrain the normals of interacted parts to maintain the relative orientation:", + "bbox": [ + 89, + 556, + 483, + 617 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {n} = \\sum_ {n \\in [ h, o ]} \\left(\\mathbf {n} _ {h} \\cdot \\mathbf {n} _ {o} + 1\\right) ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 626, + 383, + 661 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{n}_h$ is normals of interacted body vert, and $\\mathbf{n}_o$ indicates object normals of interacted vert. In addition, we employ the object scale loss $L_{\\mathrm{scale}}$ [64] and the penetration term $L_{\\mathrm{pene}}$ to prevent collisions [23]. The introduction of ground loss $L_{g}$ is decided by the object state $\\mathbf{s} = [0,1]$ .", + "bbox": [ + 89, + 669, + 483, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finer HOI Optimization. During this stage, we optimize the human parameters $\\theta_{b},\\theta_{h}$ and $t_b$ and fix the object parameters. To synthesize detailed and diverse interaction poses for full-body optimization, inspired by differentiable force closure [32], we design the fine-grained optimization loss as below:", + "bbox": [ + 89, + 744, + 483, + 834 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {f} = \\delta \\cdot L _ {f c} + L _ {p e n e} + L _ {i n t e r}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 844, + 482, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where loss functions $L_{\\mathrm{pene}}$ and $L_{\\mathrm{inter}}$ are analogous to the aforementioned terms, but with the incorporation of", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dbb228291e07176b850c9f1c06201e6506b7c872492f3539f4e9e1d00247c3b0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricMagic3D [29]DreamFusion [43]DreamHOI [67]Ours
CLIP Score↑0.24160.24370.27570.2968
", + "bbox": [ + 514, + 89, + 905, + 125 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of CLIP similarity scores for different methods. Higher scores are better.", + "bbox": [ + 511, + 135, + 906, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Kaolin [11] to better sample surface distance. We introduce SMPL-based force closure loss $L_{fc}$ by replacing the base model from MANO to SMPL-H to conduct more detailed interaction poses:", + "bbox": [ + 511, + 196, + 906, + 257 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {f c} = \\sum_ {j \\in [ o ]} \\left(\\sum_ {i \\in [ h ]} f v (i, j) \\cdot n (j)\\right) ^ {2}, \\delta \\cdot \\left[ \\theta_ {b}, \\theta_ {h} \\right]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 273, + 906, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$fv$ symbolizes the force vector applied at vertex $j$ from interacted human vertex $i$ , and $\\delta$ decide optimize body pose $\\theta_{b}$ or hand pose $\\theta_{h}$ . Specifically, we make a series of augmentations to improve force closure acceptance rate and robustness. We decrease contact points by sparse uniform sampling on the interacted vertices. Moreover, we initialize body and hand positions around the interacted object parts and maintain the hand normals as positive to the interacted object parts. Finally, we extract the interacted vertices to form interacted mesh and non-interacted mesh and compute $L_{fc}$ and $L_{pene}$ separately.", + "bbox": [ + 511, + 340, + 906, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 530, + 648, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Data and Metrics", + "text_level": 1, + "bbox": [ + 511, + 558, + 684, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We collect evaluated object meshes in 3 ways: BEHAV [2], PartNet [38], and sampling from generative models. BEHAV dataset is commonly used in this field, and PartNet is manually annotated in an object mesh with part-level semantic labels on the vertices. For metrics, we use CLIP similarity score [44] between our prompts and the rendering images of the HOI scene, by rendering our human and object meshes from 4 different views randomly on 5 prompts and 10 objects categories. During this process, we select 5 BEHAV objects (backpack, basketball, chairwood, keyboard, and suitcase), and 5 objects from publicly accessed 3D generative models (motorcycle, car, television, baby doll, and humanoid robot). The same evaluation setting is adapted to other comparison methods. In addition, while the CLIP score is designed to evaluate how well an image aligns with the input text, it often fails to capture finer nuances in interactions, leading to less distinct variations in the metrics. We follow InterFusion [9] to leverage the advanced image understanding capabilities of GPT-4V further to assess the performance of baselines and our method. The text prompts are designed and then augmented using GPT.", + "bbox": [ + 511, + 583, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7020", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56240666b57051852353bc0d4ff7db70a960212d273042d2126c7e0deea1cda4.jpg", + "image_caption": [ + "Figure 4. Ablation study on the fine-grained optimization. Left figures are the results of removing fine-grained terms and right figures apple fine-grained terms to synthesize grasping details." + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 472, + 258 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7c01417ce23c3e07d047f18a2b095c34cf3ca4201d994c65e083cc10cf820756.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricsMagic3DDreamFusionDreamFusion*DreamHOIOurs
Overall4.36.517.326.045.6
Contact8.64.315.219.552.1
", + "bbox": [ + 91, + 324, + 482, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Quantitative Evaluations", + "text_level": 1, + "bbox": [ + 89, + 439, + 318, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CLIP Score. As shown in Table 1, we present a comparison of CLIP similarity scores across various methods, where higher scores indicate better alignment with the target descriptions. Our method InteractAnything, achieves the highest mean score across the evaluation prompts, outperforming baselines such as DreamFusion [43], Magic3D [29], and DreamHOI [67]. These results demonstrate that our approach effectively captures the semantic interactions of the text descriptions. Notably, to better evaluate interaction performance without being confounded by appearance, we apply random human textures to the body mesh and retain the original textures for objects.", + "bbox": [ + 88, + 460, + 482, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GPT-4V Selection. As presented in Table 2, we enable the vision language model [1] to choose the most reasonable result in terms of Overall and Contact from all generated outputs based on criteria such as the presence of a full human body, a complete object, and correct physical interactions, and then returning the index of the selected result. No in-context examples are provided to guide the model in the Overall setting, we give text prompts only for the Contact evaluation, and the order of generated results is randomly shuffled to avoid bias, the final results are obtained from multiple tests.", + "bbox": [ + 88, + 643, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Evaluations and Application", + "text_level": 1, + "bbox": [ + 89, + 818, + 434, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization Comparison. The comparison results are shown in Figure 3, where our method outperforms the baselines, producing the most realistic and plausible 3D human-object interactions. This is evident in both the", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96482e8b28fecbdc0483ee18124a6abefd2699cf01c84b665d83290d9a9b2708.jpg", + "image_caption": [ + "Figure 5. Visualization of the adaptive inpainting and open-set object affordance parsing results on the same object with different text instructions. Our method first generates reasonable 2D inpainting results (middle columns) and then computes contact probabilities as the affordance representation (right column)." + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 893, + 281 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/51667a65a7395e4aea139030e0bfe294b8c07078c3d520f2333ae5211a237b0b.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation of GPT-4V selection in terms of Overall and Contact." + ], + "table_footnote": [], + "table_body": "
SettingMagic3DDreamFusionDreamFusion*DreamHOIOurs
Full4.36.517.326.045.6
w/o Refine6.56.517.330.439.1
w/o LLM-init8.610.821.732.626.1
", + "bbox": [ + 514, + 378, + 905, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study on LLM-guided initialization and HOI refinement by GPT-4V selection.", + "bbox": [ + 511, + 449, + 905, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "level of detail and the physical accuracy of the interactions. Magic3D [29], for instance, often generates overly simplified shapes with minimal texture and detail, as seen in the keyboard typing example, where the lack of intricacy undermines realism. Both Magic3D [29] and DreamFusion [43] fail to capture body structure or realistic human-object relationships, highlighting the difficulty of generating realistic HOI using general 3D scene generation models. DreamFusion*, which incorporates object mesh as input, retains a complete or partial human body shape but struggles with accurate contact and natural scale. This limitation is particularly noticeable in examples like the holding-chair scenario, where hand placement lacks precision. DreamHOI [67] improves upon DreamFusion in terms of HOI but still falls short in capturing fine interaction details, such as accurate foot placement on the basketball.", + "bbox": [ + 511, + 506, + 906, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inpainting Visualization. We further visualize the 2D inpainting results and affordance parsing on the same object in Figure 5 to emphasize the effectiveness of our method in disentangling affordances for distinct human-object interactions. The figure showcases various human object interactions with a chair, highlighting the ability to localize affordances accurately. This visualization provides valuable insight into the regions on the object that are most likely to be in contact with humans, which is critical for optimizing human-object interaction.", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7021", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1f9b0c20729d9595390fa44bbfc5b41855d0a6473302a9a6dc90fd4ded389ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 88, + 470, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b0e1f99492fd2af975c2db8b6d768dc55a4de7651c998e27b1dc76a8d73d453a.jpg", + "image_caption": [ + "Figure 6. Qualitative results of scene populating and scene interaction applications from different views." + ], + "image_footnote": [], + "bbox": [ + 102, + 165, + 470, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative performance on Scene Populating. We demonstrate one possible way to populate scenes with InteractAnything. We first utilize the HPS dataset [14] to obtain the 3D scene and then integrate the generated HOI results into the scene by placing them in appropriate locations. By employing a ground loss and large language model, we ensure that the HOI results are naturally embedded within the scene without artifact issues such as object or human interpenetration. Furthermore, our method goes beyond simply populating HOI results into the scene, we also generate realistic interaction effects with existing objects in the environment. For example, InteractAnything can effectively generate interactions such as \"sitting on the table\" or \"opening a door\", as demonstrated in the Figure 6.", + "bbox": [ + 91, + 295, + 483, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 516, + 243, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Fine-grained Optimization In this study, we analyze the effect of fine-grained optimization terms as outlined in Section 3.5 while keeping all other parameters constant. The results shown in Figure 4, clearly demonstrate that the fine-grained optimization significantly enhances the realism and contact plausibility of human-object interactions (right green panel) in the \"grasp\" prompt. Without fine-grained optimization (left red panel), the alignment between the hand and table is noticeably inaccurate, resulting in unnatural poses where the hand appears to \"float\" near the table instead of properly grasping or holding it. In contrast, with fine-grained optimization enabled (right panel), the human hand is aligned with the table surface, resulting in a more realistic interaction. This improvement highlights the importance of fine-grained optimization in achieving natural contact and precise positioning, especially in actions involving detailed hand-object interactions.", + "bbox": [ + 89, + 537, + 482, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study on LLM-guided Initialization We further ablate the LLM-guided initialization and the corresponding HOI refinement module (fine-grained HOI synthesis) to investigate the contribution of our design in leveraging language models. As shown in Table 3, we use GPT-4V selection as the evaluation metric. Compared to the full method (Full), the variant without refinement (w/o Refine)", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "exhibits a noticeable decline in performance. This drop is primarily due to the absence of fine-grained contact optimization, which is critical for generating the most accurate human-object interactions. In contrast, when the LLM-guided initialization is ablated (w/o LLM-init), we randomly initialize both the human and object positions, followed by coarse HOI optimization. This approach results in a significant decrease in selection scores, ranking lower than DreamHOI [67]. This finding emphasizes the importance of an informed initialization strategy in establishing a solid foundation for HOI optimization. The stark differences in the ablation study underscore the effectiveness of our design. The integration of LLM-guided initialization and refinement is vital for achieving superior results.", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Limitation", + "text_level": 1, + "bbox": [ + 511, + 316, + 756, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we demonstrate that our approach can generate fine-grained, realistic HOIs for open-set 3D objects based on text descriptions. While our method shows promising results in improving contact precision and adapting to unseen objects, it still relies on 2D priors as its primary guidance. Consequently, advancements in multi-view consistent image generation or fully 4D generation techniques would likely enhance our approach, leading to more coherent and robust 3D interactions. Additionally, physical plausibility is also an important application in recent applications, e.g., in humanoid manipulation. We use force closure as a physical constraint while a high-quality simulator could further evaluate the generated HOI. Furthermore, our method currently uses the SMPL-H model to represent human body geometry and articulation. While this model captures a wide range of human poses and movements, it limits the adaptability of our framework to diverse interaction agents, particularly those with non-human anatomies. These agents often have unique skeletal structures and movement patterns that differ significantly from human biomechanics, which our current system is not fully equipped to process. In future work, we aim to address these limitations by investigating multi-view or 4D generation for enhanced consistency and exploring more flexible models to represent a broader array.", + "bbox": [ + 511, + 340, + 908, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 718, + 633, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper tackles the challenge of generating novel 3D HOI involving unseen objects—a task where current methods often struggle. To address this issue, we propose InteractAnything, a novel zero-shot synthesis method driven by text instructions and not reliant on training with existing 3D datasets. It leverages guidance from the large language model to distill human-object relations and employs the pre-trained 2D diffusion model to parse contact areas with open-set objects, circumventing the limitations of existing HOI datasets.", + "bbox": [ + 511, + 742, + 906, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7022", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by the National Science and Technology Major Project (2022ZD0114900) and NSFC-6247070125. We extend our gratitude to the reviewers for their insightful comments.", + "bbox": [ + 89, + 90, + 485, + 154 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 166, + 187, + 181 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 7", + "[2] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6", + "[3] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep It SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. In Computer Vision - ECCV 2016, 2016. 5", + "[4] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 958-968, 2024. 2", + "[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5", + "[6] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In International Conference on Computer Vision (ICCV), 2023. 2", + "[7] Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets creation using spatially-aware diffusion guidance. In European Conference on Computer Vision, pages 128-146. Springer, 2024. 2, 3", + "[8] Hongkai Dai, Anirudha Majumdar, and Russ Tedrake. Synthesis and optimization of force closure grasps via sequential semidefinite programming. Robotics Research: Volume 1, pages 285-305, 2018. 2", + "[9] Sisi Dai, Wenhao Li, Haowen Sun, Haibin Huang, Chongyang Ma, Hui Huang, Kai Xu, and Ruizhen Hu. Interfusion: Text-driven generation of 3d human-object interaction. In European Conference on Computer Vision, pages 18-35. Springer, 2024. 2, 3, 6", + "[10] Google DeepMind. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2", + "[11] Clement Fuji Tsang, Maria Shugrina, Jean Francois Lafleche, Towaki Takikawa, Jiehan Wang, Charles Loop, Wenzheng Chen, Krishna Murthy Jatavallabhula, Edward Smith, Artem Rozantsev, Or Perel, Tianchang Shen, Jun Gao, Sanja Fidler, Gavriel State, Jason Gorski, Tommy Xiang, Jianing Li, Michael Li, and Rev Lebaredian. Kaolin: A pytorch library for accelerating 3d deep learning re" + ], + "bbox": [ + 93, + 191, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "search. https://github.com/NVIDIAGameWorks/kaolin, 2022.6", + "[12] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Scholkopf. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[13] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13041-13051, 2023. 5", + "[14] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human positioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2021. 8", + "[15] Sookwan Han and Hanbyul Joo. Chorus: Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In International Conference on Computer Vision (ICCV), 2023. 3, 5", + "[16] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambiguities with 3D scene constraints. In International Conference on Computer Vision (ICCV), 2019. 2, 3", + "[17] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2", + "[18] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 2", + "[19] Jingwei Huang, Hao Su, and Leonidas Guibas. Robust watertight manifold surface generation method for shapenet models. arXiv preprint arXiv:1802.01698, 2018. 6", + "[20] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Zhiyuan Zhang, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. Full-body articulated human-object interaction. In International Conference on Computer Vision (ICCV), 2023. 2", + "[21] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[22] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7122-7131, 2018. 5", + "[23] Tero Karras. Maximizing parallelism in the construction of bvhs, octrees, and k-d trees. In Proceedings of ACM SIGGRAPH/Eurographics Conference on High-Performance Graphics, 2012. 6", + "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):139-1, 2023. 2", + "[25] Hyeonwoo Kim, Sookwan Han, Patrick Kwon, et al. Beyond the contact: Discovering comprehensive affordance for 3d" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7023", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "objects from pre-trained 2d diffusion models. In European Conference on Computer Vision (ECCV), 2024. 4, 5", + "[26] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2, 3", + "[27] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. In European Conference on Computer Vision (ECCV), 2024. 2", + "[28] Lei Li and Angela Dai. Genzi: Zero-shot 3d human-scene interaction generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20465-20474, 2024. 4", + "[29] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7", + "[30] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21159-21168, 2023. 5", + "[31] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In International Conference on Computer Vision (ICCV), 2023. 2", + "[32] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters (RA-L), 7(1):470-477, 2021. 2, 6", + "[33] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splatting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[34] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In International Conference on Learning Representations (ICLR), 2023. 2", + "[35] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Transactions on Graphics (TOG), 34(6):248:1-248:16, 2015. 3", + "[37] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 6", + "[39] OpenAI. Gpt-3.5. https://platform.openai.com/docs/models/gpt-3-5, 2023. Accessed: 2023-11-14. 2, 4", + "[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 6", + "[41] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 2", + "[42] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4", + "[43] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations (ICLR), 2023. 2, 3, 5, 6, 7", + "[44] Alec Radford, JongWook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Askell Amanda, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. Cornell University - arXiv, Cornell University - arXiv, 2021. 6", + "[45] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2", + "[48] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: modeling and capturing hands and bodies together. ACM Transactions on Graphics (TOG), 36(6):1-17, 2017. 3", + "[49] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[50] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7024", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[51] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2", + "[52] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 6", + "[53] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 2", + "[54] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J Black. Deco: Dense estimation of 3d human-scene contact in the wild. In International Conference on Computer Vision (ICCV), 2023. 2", + "[55] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In International Conference on 3D Vision (3DV), 2022. 3", + "[56] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2", + "[57] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16210-16220, 2022. 2", + "[58] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In International Conference on Computer Vision (ICCV), 2023. 2, 3", + "[59] Sirui Xu, Yu-Xiong Wang, Liangyan Gui, et al. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. Advances in Neural Information Processing Systems, 37:52858-52890, 2024.", + "[60] Hongwei Yi, Justus Thies, Michael J Black, Xue Bin Peng, and Davis Rempe. Generating human interaction motions in scenes with text control. In European Conference on Computer Vision (ECCV), 2025. 2, 3", + "[61] Wanqi Yin, Zhongang Cai, Ruisi Wang, Fanzhou Wang, Chen Wei, Haiyi Mei, Weiye Xiao, Zhitao Yang, Qingping Sun, Atsushi Yamashita, et al. Whac: World-grounded humans and cameras. In European Conference on Computer Vision, pages 20-37, 2024. 5", + "[62] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monoc" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ular images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:12287-12303, 2023.", + "[63] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13232-13242, 2022. 5", + "[64] Jason Y. Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In European Conference on Computer Vision (ECCV), 2020. 3, 6", + "[65] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 2", + "[66] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on pattern analysis and machine intelligence, 44(6): 3170-3184, 2021. 6", + "[67] Thomas Hanwen Zhu, Ruining Li, and Tomas Jakab. Dreamhoi: Subject-driven generation of 3d human-object interactions with diffusion priors. arXiv preprint arXiv:2409.08278, 2024. 2, 3, 5, 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 906, + 458 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7025", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/bba3c5bb-a0b8-4050-9fca-51eeaf9696e0_model.json b/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/bba3c5bb-a0b8-4050-9fca-51eeaf9696e0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..93c28ec035363e8c8d46648dd9412a825add07e2 --- /dev/null +++ b/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/bba3c5bb-a0b8-4050-9fca-51eeaf9696e0_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.88, + 0.177 + ], + "angle": 0, + "content": "InteractAnything: Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.203, + 0.877, + 0.223 + ], + "angle": 0, + "content": "Jinlu Zhang\\(^{1,2}\\), Yixin Chen\\(^{2\\boxtimes}\\), Zan Wang\\(^{2,3}\\), Jie Yang\\(^{2,4}\\), Yizhou Wang\\(^{1,5,6,7\\boxtimes}\\), Siyuan Huang\\(^{2\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.224, + 0.602, + 0.239 + ], + "angle": 0, + "content": "indicates corresponding authors" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.241, + 0.775, + 0.257 + ], + "angle": 0, + "content": "1 Center on Frontiers of Computing Studies, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.258, + 0.797, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\) State Key Laboratory of General Artificial Intelligence, BIGAI \\(^{3}\\) Beijing Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.276, + 0.868, + 0.292 + ], + "angle": 0, + "content": "4 The Chinese University of Hong Kong, Shenzhen 5 Nat'l Eng. Research Center of Visual Technology, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.293, + 0.838, + 0.31 + ], + "angle": 0, + "content": "\\(^{6}\\) Institute for AI, Peking University \\(^{7}\\) State Key Laboratory of General Artificial Intelligence, Peking University" + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.34, + 0.312, + 0.351 + ], + "angle": 0, + "content": "A person holds/pulls/lies on the chair..." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.353, + 0.346, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.532, + 0.296, + 0.542 + ], + "angle": 0, + "content": "(a) Diverse human interactions." + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.337, + 0.553, + 0.347 + ], + "angle": 0, + "content": "A person grasps the" + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.348, + 0.571, + 0.357 + ], + "angle": 0, + "content": "backpack/table/ball/knife..." + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.357, + 0.635, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.531, + 0.634, + 0.542 + ], + "angle": 0, + "content": "(b) Detailed interactions synthesis of open-set objects." + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.341, + 0.863, + 0.351 + ], + "angle": 0, + "content": "A person lifts/rides/criddles the" + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.352, + 0.868, + 0.362 + ], + "angle": 0, + "content": "tesla car/motorcycle/horse/baby..." + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.363, + 0.87, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.531, + 0.896, + 0.543 + ], + "angle": 0, + "content": "(c) Novel interactions given generative objects." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.555, + 0.908, + 0.597 + ], + "angle": 0, + "content": "Figure 1. 3D human object interaction synthesis by InteractAnything. Given a simple text description with goal interaction and any object mesh as input, our method enables the generation of diverse, natural, detailed, and novel interactions for open-set 3D objects in a zero-shot manner. The orange and green boxes of (b) indicate detailed contact poses from different views." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.61, + 0.327, + 0.625 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.645, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Recent advances in 3D human-aware generation have made significant progress. However, existing methods still struggle with generating novel Human Object Interaction (HOI) from text, particularly for open-set objects. We identify three main challenges of this task: precise human-object relation reasoning, affordance parsing for any object, and detailed human interaction pose synthesis aligning description and object geometry. In this work, we propose a novel zero-shot 3D HOI generation framework without training on specific datasets, leveraging the knowledge from largescale pre-trained models. Specifically, the human-object relations are inferred from large language models (LLMs) to initialize object properties and guide the optimization process. Then we utilize a pre-trained 2D image diffusion model to parse unseen objects and extract contact points, avoiding the limitations imposed by existing 3D asset knowledge. The initial human pose is generated by sam" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.909, + 0.793 + ], + "angle": 0, + "content": "pling multiple hypotheses through multi-view SDS based on the input text and object geometry. Finally, we introduce a detailed optimization to generate fine-grained, precise, and natural interaction, enforcing realistic 3D contact between the 3D object and the involved body parts, including hands in grasping. This is achieved by distilling human-level feedback from LLMs to capture detailed human-object relations from the text instruction. Extensive experiments validate the effectiveness of our approach compared to prior works, particularly in terms of the fine-grained nature of interactions and the ability to handle open-set 3D objects. Project page: jinluzhang.site/projects/interactanything." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.828, + 0.646, + 0.843 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Synthesizing the interaction of humans and any given object holds great significance across diverse applications, including AR/VR, computer simulation, and animation. Tradi-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7015" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "tionally, creating high-quality 3D human object interaction (HOI) assets requires expensive, lab-controlled data capture systems and significant human effort. Although some datasets [2, 16, 20, 21, 26, 53] have collected these data, the models [26, 27, 41, 54, 58, 60] trained on them can only generate interactions for specific object categories. These methods struggle to handle the diversity of object geometries, scales, and interactions at the same time, making text-driven zero-shot HOI synthesis a challenging and highly demanding task." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.246, + 0.485, + 0.502 + ], + "angle": 0, + "content": "Given recent advances in text-to-3D techniques [6, 29, 56], e.g., leveraging Score Distillation Sampling (SDS) [43] to generate 3D content from text, a natural question is whether this paradigm can be applied to zero-shot HOI synthesis. However, directly integrating 3D humans and objects into a scene without understanding the complex relationships between them remains challenging. Recent works [7, 9, 67] are proposed to generate realistic 3D HOIs. Among them, InterFusion [9] uses Neural Radiance Fields (NeRF) [37] for individual human and object representations but struggles with precise interaction area, leading to suboptimal and unrealistic HOIs. DreamHOI [67] optimizes on both mesh and NeRF representations through bidirectional rendering but faces challenges in ensuring accurate contact points and relative scales due to solely optimizing human states while keeping the given object static, leading to limited interactions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.746 + ], + "angle": 0, + "content": "In this paper, we first identify the core challenge of zero-shot 3D HOI generation as accurately capturing human-object relations, contact areas, and global consistency, which are not typically presented in single-object or non-contact generation scenarios. These challenges prevent the generation of high-quality, geometrically precise, and textually consistent results. Furthermore, while large-scale pre-trained models, e.g., 2D diffusion models [42, 47] and large language models (LLMs) [1, 10, 39], cannot fully capture the complexity of human object interactions (including body dynamics, expressive details, and object geometry), they can offer valuable prior knowledge of common interaction patterns based on textual descriptions. Consequently, we argue that the solution for effective generation is to leverage this diverse prior knowledge to guide the optimization of interactions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Following the above insights, we decompose the complex zero-shot 3D HOI task into these stages: 1) human-object relation reasoning and initialization, 2) open-set object contact affordance parsing, and 3) human pose synthesis aligning text and object geometry. We first leverage the LLM as a human-level feedback provider, which reasons and unravels complex semantic interaction relationships, e.g., relative position and orientation, from simple text input. It initializes human and object states and designs a series of optimization functions to generate detailed inter" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "actions. The initial human pose is optimized by multi-view SDS loss for interaction in line with the given textual description and object geometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.289 + ], + "angle": 0, + "content": "To parse the 3D object affordance, we utilize initialized human-object relation and sample adaptive inpainting masks during the diffusion process to obtain the contact information. This enables the extraction of contact probabilities on the object, which are then used to segment and define the weighted dynamic contact regions. Finally, we design a coarse-to-fine optimization based on force closure [8, 32] to generate expressive, diverse, and physically plausible interactions, allowing for more accurate and detailed contact. To summarize, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.291, + 0.905, + 0.35 + ], + "angle": 0, + "content": "- We propose InteractAnything for zero-shot 3D HOI generation, which effectively distills human and object interactions from LLM and pre-trained diffusion models to synthesize diverse, detailed, and novel interactions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.352, + 0.905, + 0.396 + ], + "angle": 0, + "content": "- We design a novel open-set object affordance parsing, which generates contact maps on 3D geometry by utilizing adaptive inpainting and potential contact distribution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.397, + 0.905, + 0.502 + ], + "angle": 0, + "content": "- Our model leverages both text and object geometry to drive human pose synthesis, enabling realistic and contextually appropriate interactions for open-set objects. Results show that our synthesized pose accurately reflects the semantic meaning of the input while respecting the spatial constraints of the object, leading to improved performance, as shown in Figure 1." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.291, + 0.905, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.519, + 0.655, + 0.535 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.545, + 0.822, + 0.56 + ], + "angle": 0, + "content": "2.1. Text-driven 3D Content Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The generation of 3D content from text has seen rapid advancements, primarily driven by diffusion-based models that build on text-to-2D image generation approaches [45, 46, 49, 50]. One pioneering approach, DreamFusion [43], introduced Score Distillation Sampling (SDS), which leverages the pre-trained 2D diffusion model to optimize 3D content through techniques like NeRF (Neural Radiance Fields) [37] and 3D Gaussian Splatting [24]. A significant area of research aims to address the complexity of 3D human generation. For example, recent methods have been proposed for generating controllable 3D human avatar [4, 17, 18, 33, 57, 65]. However, these works focus on diverse and high-quality appearance but animate human avatars with given human poses or motion, which do not satisfy 3D HOI requirements in human pose synthesis. The proliferation of large 3D datasets has also catalyzed progress, leading to more sophisticated models like Zero-1-to-3 [31], Zero \\(123++\\) [51], and MVDream [52], which employ 2D diffusion models to generate multi-view consistent images that serve as inputs for efficient 3D generation systems, such as SyncDreamer [34] and Wonder3D [35]. Despite these advancements, generating complex HOI scenes," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7016" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "especially those involving interactions between humans and objects, remains challenging." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.143, + 0.433, + 0.16 + ], + "angle": 0, + "content": "2.2. Data-Driven Human Object Interaction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.483, + 0.395 + ], + "angle": 0, + "content": "Previous approaches [26, 58-60] primarily train generative models to produce static human poses or dynamic motions in response to objects or scenes using curated HOI datasets. Recent works focus on dynamic motion generation performance conditioned on encoded object features but rely heavily on specialized datasets, limiting generalization to diverse objects. For example, OMOMO [26] generates human motion through diffusion models but requires object motion as input. InterDiff [58] predicts full-body interactions based on prior HOI context but is limited to the prediction task, restricting flexibility in new scenarios. In contrast, our method jointly synthesizes human and object states and adapts to novel objects without extensive, object-specific training by extracting interaction patterns directly from 2D diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.597 + ], + "angle": 0, + "content": "On the other hand, optimization-based methods reconstruct realistic human poses and transformations from images. PHOSA [64] optimizes human object relations using depth-aware loss functions. PROX [16] encodes contact probabilities and semantic scene labels to position humans within scenes. CHORUS [15] synthesizes multi-view images with a pre-trained generative model to learn 3D spatial relationships. By extracting knowledge from LLM to infer human poses, the method in [55] queries contact information to reconstruct 3D human-object interactions. Compared to these methods, our method leverages pre-trained diffusion models for accurate parsing of object affordance and enables fine-grained interactions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.618, + 0.436, + 0.635 + ], + "angle": 0, + "content": "2.3. Zero-shot 3D Compositional Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To achieve the compositional scene generation, Combo-Verse [7] proposes a spatially-aware SDS loss, generating the final 3D scene through single-object reconstruction and multi-object composition. GraphDreamer [12] leverages scene graphs to guide generation by converting the semantic information into a global textual description to enable text-driven compositional scene generation. These methods focus on the reasonable arrangement of multiple objects but ignore the contact generation in HOI scenes. InterFusion [9] and DreamHOI [67] present this challenge and then try to address using SMPLs [36, 48] as body model prior. DreamHOI [67] designs the SDS guidance mixture to separately render human interactions using different diffusion models to improve human modeling quality. However, it cannot ensure precise contact during the denoising process. Compared to these methods, our work focuses on addressing contact areas parsing on open-set objects." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.644, + 0.107 + ], + "angle": 0, + "content": "3. Our Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.313 + ], + "angle": 0, + "content": "Task Definition and Overview. Given a text prompt \\(\\epsilon\\) that describes the desired HOI, e.g., \"A person grasps the chair\". Our goal is to synthesize natural human mesh \\(\\nu_{H}\\) and object mesh \\(\\nu_{O}\\) that align with \\(a\\) and \\(\\epsilon\\). As shown in Figure 2, we first obtain initial human-object relations from LLM based on text (Section 3.2). For object optimization, we initialize object rotation \\(r_o\\), translation \\(t_o\\), scale \\(s_o\\), state \\(s\\), based on affordance map \\(\\mathcal{M}\\) (Section 3.3). We also synthesize the initial human pose based on the input text and object geometry (Section 3.4). Finally, we use SMPL-H as representation and optimize pose parameter \\(\\theta_b, \\theta_h\\), global translation \\(t_b\\), and object rotation and translation \\(r_o, t_o\\) to generate detailed and diverse HOI (Section 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.32, + 0.734, + 0.336 + ], + "angle": 0, + "content": "3.1. Preliminary Knowledge" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.907, + 0.508 + ], + "angle": 0, + "content": "Human Body Model. We use the SMPL-H [36, 48], a parametric model that includes both body and hand parameters, to represent the human body and hand articulation. The body pose is defined by body parameters \\(\\theta_{\\mathrm{b}} \\in \\mathbb{R}^{3(K_b + 1)}\\) and the hand pose is represented by additional MANO parameters \\(\\theta_{\\mathrm{h}} \\in \\mathbb{R}^{3K_h}\\), capturing 15 joints per hand. The shape parameters are represented by \\(\\beta \\in \\mathbb{R}^{10}\\), controlling the deformation of the body and hand template mesh to target body shapes. Given the combined pose parameters \\(\\{\\theta_{\\mathrm{b}}, \\theta_{\\mathrm{h}}\\}\\) and shape parameters \\(\\beta\\), SMPL-H computes mesh vertices \\(\\mathcal{V} \\in \\mathbb{R}^{V \\times 3}\\) using linear blend skinning function \\(f\\):" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.515, + 0.905, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} (\\beta , \\theta_ {b}, \\theta_ {\\mathrm {h}}) = f \\left(T _ {p} \\left(\\beta , \\theta_ {b}, \\theta_ {\\mathrm {h}}\\right), J (\\beta), \\theta , \\mathcal {W}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.905, + 0.569 + ], + "angle": 0, + "content": "where \\(J\\) represents joint locations, \\(\\mathcal{W} \\in \\mathbb{R}^{V \\times (K)}\\) is skimming weights with all \\(K\\) joints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.706 + ], + "angle": 0, + "content": "Score Distillation Sampling (SDS). SDS leverages pretrained diffusion models to guide the synthesis of 3D scenes by aligning NeRF with a text prompt. It operates by iteratively adjusting NeRF parameters, denoted as \\(\\theta\\), based on gradients computed to align the generated images with the textual input. Given a noisy 2D image \\(x_{t}\\) (obtained by adding noise \\(\\epsilon\\) to a clean NeRF-rendered image \\(x\\)), a textual prompt \\(y\\), and a noise level \\(t\\), SDS calculates the gradient of the loss \\(\\mathcal{L}_{\\mathrm{SDS}}\\) with respect to \\(\\theta\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.711, + 0.868, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (x) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) (\\hat {\\epsilon} (x _ {t}; y, t) - \\epsilon) \\frac {\\partial x}{\\partial \\theta} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(\\hat{\\epsilon}(x_t; y, t)\\) represents the diffusion model predicted noise given the noisy image \\(x_t\\), prompt \\(y\\), and noise level \\(t\\), \\(w(t)\\) is a weighting function that adjusts the influence of the gradient based on the noise level \\(t\\). This gradient is used to update \\(\\theta\\), guiding the NeRF model toward producing images that match the textual description. For instance, in DreamFusion [43], this process enables text-to-3D generation by repeatedly refining the NeRF parameters \\(\\theta\\) from a random initialization, iteratively producing images that align more closely with the target prompt." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7017" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.091, + 0.885, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.389, + 0.908, + 0.461 + ], + "angle": 0, + "content": "Figure 2. Framework of InteractAnything. Given a text description and any object mesh as input, our approach begins by querying LLM to infer precise human-object relationships, which are used to initialize object properties. Next, we analyze the contact affordance of the object geometry. The human pose is synthesized using a pre-trained 2D diffusion model, guided by multi-view SDS loss and the designed spatial constraint. Finally, based on the targeted object contact areas and a plausible human pose, we perform expressive HOI optimization to synthesize realistic and contact-accurate 3D human-object interactions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.485, + 0.33, + 0.501 + ], + "angle": 0, + "content": "3.2. LLM-guided Initialization" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.507, + 0.485, + 0.825 + ], + "angle": 0, + "content": "To infer diverse intended human object interactions for a given object \\(\\mathcal{V}_O\\) and action \\(a\\), we query the LLM to provide commonsense feedback. To minimize hallucination, we need to transfer inferred relationships from GPT [39] into specific values that position the human and object meshes accurately, therefore, we enable the language model to select answers from pre-defined options step by step, instead of directly generating responses. For initializing the object rotation \\(r_o\\), translation \\(t_o\\), scale \\(s_o\\), and state \\(s\\), the language model selects the relative position of the human to the object from defined options, which is then mapped to \\(t_o\\) based on the combined size of the human and object. We set all input objects to have \\(y\\)-up and \\(x\\)-front orientations, with \\(r_o\\) adjusted based on which side of the object should face the human. The object scale is chosen from 4 defined options and set relative to human size. Lastly, the state \\(s\\) determines if the object should be positioned on the ground based on whether it is meant to move. Besides, we also query interacted semantic body parts \\(S_h\\) and map the labels to human mesh to obtain vertices \\(\\mathcal{P}_h\\) for part-level optimization using annotated human part labels." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.407, + 0.851 + ], + "angle": 0, + "content": "3.3. Open-set Object Affordance Parsing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To obtain the contact affordance of unseen object geometries given a specified action e.g., holds, we conduct 2D inpainting based on a pre-trained diffusion model to dis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.56 + ], + "angle": 0, + "content": "till possible human interaction parts through LLM-guided adaptive mask generation. Then we design a 2D probabilistic contact function to model the distribution of the interaction, generating an affordance map to identify and extract the regions for HOI optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.909, + 0.833 + ], + "angle": 0, + "content": "2D Adaptive Mask Generation Image inpainting is to fill in missing regions of an image, where the missing regions can be defined by a mask image. To get knowledge of possible interacted parts, we utilize the powerful understanding and generation ability of pre-trained inpainting model [42]. Given 2D projection images of the 3D object from \\( n_p \\) views, we found existing inpainting models cannot ensure complete human in the image (many inpainting results only include interacted body parts, e.g., hands in \"hold\" action). Therefore, different from previous methods [25, 28], we do not randomly sample the inpainting masks on 2D images, instead, our method identifies possible human interaction parts to improve the inpainting success rate by generating inpainting masks from LLM feedback in two forms: (1) full-body mask with respect to object properties \\( r_o \\), \\( t_o \\), person position and person body size; (2) body-part inpainting masks for intended interacted body part labels \\( S_h \\) and corresponding human vertices \\( \\mathcal{P}_h \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.833, + 0.908, + 0.879 + ], + "angle": 0, + "content": "More specifically, we first create a full-body inpainting mask guided by LLM-inferred initializations \\( r_o \\), \\( t_o \\), and body size \\( s_f \\):" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.884, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\nM _ {\\mathrm {f u l l}} ^ {i} = \\mathcal {J} ((\\mathcal {V} _ {H} \\cdot r _ {o} ^ {- 1}), s _ {f}, c ^ {i}) + t _ {o} ^ {- 1} + \\varpi_ {1}, \\qquad (2)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7018" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.089, + 0.87, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.45, + 0.908, + 0.491 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison results with baselines. * indicates we re-implement this method by embedding object mesh into the diffusion process, which follows the pipeline of DreamHOI [67] and our method. More visualization results are presented in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.519, + 0.482, + 0.608 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}\\) is 2D projection transformation with camera parameter \\(c^i\\), and \\(\\varpi_{1}\\) is the noise term. Similarly, we generate body-part-level inpainting masks using given body part labels \\(S_{h}\\) and corresponding human vertices \\(\\mathcal{P}_h\\) and body part size \\(s_p\\) to help inpainting body parts like hands as the above formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.617, + 0.482, + 0.637 + ], + "angle": 0, + "content": "\\[\nM _ {\\text {p a r t}} ^ {i} = \\mathcal {J} \\left(\\left(\\mathcal {P} _ {h} \\cdot r _ {o} ^ {- 1}\\right), s _ {p}, c ^ {i}\\right) + t _ {o} ^ {- 1} + \\varpi_ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.704 + ], + "angle": 0, + "content": "Besides, We apply the above full-body masks \\( M_{\\mathrm{full}} \\) into the dynamic mask inpainting pipeline [25], which can further refine the initial mask detecting human on intermediate results in the diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Affordance Parsing on 2D Probability Function. Directly computing contact probabilities on a 3D object mesh relies heavily on existing 3D human reconstruction methods [3, 13, 22, 30, 61-63] from single-view images. This encounters significant 3D ambiguity and quantization error, which obscures accurate 3D human pose. In contrast to previous methods [15], we apply 2D human pose detector OpenPose [5], performing well at detecting 2D human keypoints and even occluded body parts, to estimate human from inpainting results. Then we introduce a 2D probability function that calculates contact probability according to the distance from detected body keypoints to the multi-view object projection images. First, we downsample the object" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.565 + ], + "angle": 0, + "content": "to \\(D\\) points. Let \\(\\mathcal{P}_i\\) represent the probability map for each view \\(i\\), the distance-based probability \\(f_{\\mathrm{afford}}(\\mathbf{p})\\) for a point \\(\\mathbf{p} \\in \\mathcal{V}_O\\) on the object surface is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.572, + 0.905, + 0.593 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {a f f o r d}} ^ {(i)} (\\mathbf {p}) = e ^ {- \\left\\| \\mathbf {d} _ {i} (\\mathbf {p}) \\right\\|}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.679 + ], + "angle": 0, + "content": "where \\(\\mathbf{d}_i(\\mathbf{p})\\) denotes the distance from the body part to the object mask in \\(i\\) view. Then we re-project the 2D affordance \\(f_{\\mathrm{afford}}^{(i)}(\\mathbf{p})\\) to 3D space using defined camera parameters, and aggregate the contact probabilities across all \\(n_p\\) views to update the 3D contact probability distribution as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.686, + 0.905, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} (\\mathbf {p}) = \\frac {1}{n _ {p}} \\sum_ {i = 1} ^ {n _ {p}} f _ {\\text {a f f o r d}} ^ {(i)} (\\mathbf {p}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.811 + ], + "angle": 0, + "content": "Finally, we normalize \\(\\mathcal{P}(\\mathbf{p})\\) over the entire object surface to obtain contact probability for each point is determined by its projected 2D distances across multiple views, effectively mitigating multi-view ambiguity and enabling accurate estimation of human object contact regions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.878, + 0.835 + ], + "angle": 0, + "content": "3.4. Text-Object Driven Human Pose Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Considering a brief HOI description, we posit that current 3D generation models, such as DreamFusion [43], are capable of generating human avatars in a variety of poses; however, they do not guarantee precise contact with the object" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7019" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.334 + ], + "angle": 0, + "content": "mesh. Thus, we first apply prompt engineering to generate diverse and expressive text descriptions, along with negative prompts, using large language models (LLMs) to mitigate issues like low-quality outputs and missing limbs. Next, we incorporate Multi-view SDS loss [52] to optimize the human pose aligning the interaction description and object geometry. We first scale the input object based on \\( s_o \\) given by the language model. We use the spatial constraint to prevent the human avatar from penetrating the object mesh, ensuring alignment with spatial relations. Specifically, we embed object mesh within NeRF optimization by regularizing the ray sample points in object mesh following previous works [9, 67]. After training, we use the multi-view human pose estimation method [66] to obtain the body parameters \\( \\theta_b \\) and then init hand parameters \\( \\theta_h \\). We also employ VPoser [40] to maintain reasonable joint rotations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.342, + 0.358, + 0.359 + ], + "angle": 0, + "content": "3.5. Expressive HOI Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.426 + ], + "angle": 0, + "content": "Global HOI optimization. With the above initialization as input, we first ensure the input object is watertight using Manifold [19]. Then we conduct human-centric optimization to update object properties \\((r_o, t_o, s_o)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.433, + 0.483, + 0.454 + ], + "angle": 0, + "content": "\\[\nL _ {c} = \\phi_ {i} L _ {i n t e r} + \\phi_ {n} L _ {n} + \\phi_ {s} L _ {s c a l e} + \\phi_ {p} L _ {p e n e} + \\mathbf {s} ^ {'} \\cdot L _ {g}, (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.484, + 0.507 + ], + "angle": 0, + "content": "where \\(\\phi\\) is loss weights and interaction loss is designed on. \nverts of the interacted body part \\(\\mathcal{P}_h,h\\in \\mathcal{V}_H\\) and object \ncontact areas \\(\\mathcal{P}_o,o\\in \\mathcal{M}\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.515, + 0.483, + 0.55 + ], + "angle": 0, + "content": "\\[\nL _ {i n t e r} = \\sum_ {i \\in [ h ], j \\in [ o ]} W _ {\\mathcal {M}} (j) \\cdot f _ {c h a m} \\left(\\mathcal {P} _ {h}, \\mathcal {P} _ {o}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.484, + 0.618 + ], + "angle": 0, + "content": "where \\( W_{\\mathcal{M}}(j) \\) indicates probabilities scores in object affordance map \\( \\mathcal{M} \\) of the \\( j \\)-th object vertex and \\( f_{cham} \\) is chamfer distance function. We constrain the normals of interacted parts to maintain the relative orientation:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.627, + 0.384, + 0.662 + ], + "angle": 0, + "content": "\\[\nL _ {n} = \\sum_ {n \\in [ h, o ]} \\left(\\mathbf {n} _ {h} \\cdot \\mathbf {n} _ {o} + 1\\right) ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.484, + 0.746 + ], + "angle": 0, + "content": "where \\(\\mathbf{n}_h\\) is normals of interacted body vert, and \\(\\mathbf{n}_o\\) indicates object normals of interacted vert. In addition, we employ the object scale loss \\(L_{\\mathrm{scale}}\\) [64] and the penetration term \\(L_{\\mathrm{pene}}\\) to prevent collisions [23]. The introduction of ground loss \\(L_{g}\\) is decided by the object state \\(\\mathbf{s} = [0,1]\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.484, + 0.835 + ], + "angle": 0, + "content": "Finer HOI Optimization. During this stage, we optimize the human parameters \\(\\theta_{b},\\theta_{h}\\) and \\(t_b\\) and fix the object parameters. To synthesize detailed and diverse interaction poses for full-body optimization, inspired by differentiable force closure [32], we design the fine-grained optimization loss as below:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.845, + 0.483, + 0.863 + ], + "angle": 0, + "content": "\\[\nL _ {f} = \\delta \\cdot L _ {f c} + L _ {p e n e} + L _ {i n t e r}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where loss functions \\(L_{\\mathrm{pene}}\\) and \\(L_{\\mathrm{inter}}\\) are analogous to the aforementioned terms, but with the incorporation of" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.09, + 0.906, + 0.126 + ], + "angle": 0, + "content": "
MetricMagic3D [29]DreamFusion [43]DreamHOI [67]Ours
CLIP Score↑0.24160.24370.27570.2968
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.136, + 0.907, + 0.164 + ], + "angle": 0, + "content": "Table 1. Comparison of CLIP similarity scores for different methods. Higher scores are better." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.907, + 0.258 + ], + "angle": 0, + "content": "Kaolin [11] to better sample surface distance. We introduce SMPL-based force closure loss \\( L_{fc} \\) by replacing the base model from MANO to SMPL-H to conduct more detailed interaction poses:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.274, + 0.907, + 0.328 + ], + "angle": 0, + "content": "\\[\nL _ {f c} = \\sum_ {j \\in [ o ]} \\left(\\sum_ {i \\in [ h ]} f v (i, j) \\cdot n (j)\\right) ^ {2}, \\delta \\cdot \\left[ \\theta_ {b}, \\theta_ {h} \\right]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.907, + 0.51 + ], + "angle": 0, + "content": "\\(fv\\) symbolizes the force vector applied at vertex \\(j\\) from interacted human vertex \\(i\\), and \\(\\delta\\) decide optimize body pose \\(\\theta_{b}\\) or hand pose \\(\\theta_{h}\\). Specifically, we make a series of augmentations to improve force closure acceptance rate and robustness. We decrease contact points by sparse uniform sampling on the interacted vertices. Moreover, we initialize body and hand positions around the interacted object parts and maintain the hand normals as positive to the interacted object parts. Finally, we extract the interacted vertices to form interacted mesh and non-interacted mesh and compute \\(L_{fc}\\) and \\(L_{pene}\\) separately." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.649, + 0.549 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.686, + 0.575 + ], + "angle": 0, + "content": "4.1. Data and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We collect evaluated object meshes in 3 ways: BEHAV [2], PartNet [38], and sampling from generative models. BEHAV dataset is commonly used in this field, and PartNet is manually annotated in an object mesh with part-level semantic labels on the vertices. For metrics, we use CLIP similarity score [44] between our prompts and the rendering images of the HOI scene, by rendering our human and object meshes from 4 different views randomly on 5 prompts and 10 objects categories. During this process, we select 5 BEHAV objects (backpack, basketball, chairwood, keyboard, and suitcase), and 5 objects from publicly accessed 3D generative models (motorcycle, car, television, baby doll, and humanoid robot). The same evaluation setting is adapted to other comparison methods. In addition, while the CLIP score is designed to evaluate how well an image aligns with the input text, it often fails to capture finer nuances in interactions, leading to less distinct variations in the metrics. We follow InterFusion [9] to leverage the advanced image understanding capabilities of GPT-4V further to assess the performance of baselines and our method. The text prompts are designed and then augmented using GPT." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7020" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.089, + 0.473, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.268, + 0.483, + 0.311 + ], + "angle": 0, + "content": "Figure 4. Ablation study on the fine-grained optimization. Left figures are the results of removing fine-grained terms and right figures apple fine-grained terms to synthesize grasping details." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.325, + 0.483, + 0.376 + ], + "angle": 0, + "content": "
MetricsMagic3DDreamFusionDreamFusion*DreamHOIOurs
Overall4.36.517.326.045.6
Contact8.64.315.219.552.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.386, + 0.483, + 0.414 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation of GPT-4V selection in terms of Overall and Contact." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.32, + 0.456 + ], + "angle": 0, + "content": "4.2. Quantitative Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.483, + 0.643 + ], + "angle": 0, + "content": "CLIP Score. As shown in Table 1, we present a comparison of CLIP similarity scores across various methods, where higher scores indicate better alignment with the target descriptions. Our method InteractAnything, achieves the highest mean score across the evaluation prompts, outperforming baselines such as DreamFusion [43], Magic3D [29], and DreamHOI [67]. These results demonstrate that our approach effectively captures the semantic interactions of the text descriptions. Notably, to better evaluate interaction performance without being confounded by appearance, we apply random human textures to the body mesh and retain the original textures for objects." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.809 + ], + "angle": 0, + "content": "GPT-4V Selection. As presented in Table 2, we enable the vision language model [1] to choose the most reasonable result in terms of Overall and Contact from all generated outputs based on criteria such as the presence of a full human body, a complete object, and correct physical interactions, and then returning the index of the selected result. No in-context examples are provided to guide the model in the Overall setting, we give text prompts only for the Contact evaluation, and the order of generated results is randomly shuffled to avoid bias, the final results are obtained from multiple tests." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.436, + 0.835 + ], + "angle": 0, + "content": "4.3. Qualitative Evaluations and Application" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Visualization Comparison. The comparison results are shown in Figure 3, where our method outperforms the baselines, producing the most realistic and plausible 3D human-object interactions. This is evident in both the" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.091, + 0.895, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.294, + 0.907, + 0.365 + ], + "angle": 0, + "content": "Figure 5. Visualization of the adaptive inpainting and open-set object affordance parsing results on the same object with different text instructions. Our method first generates reasonable 2D inpainting results (middle columns) and then computes contact probabilities as the affordance representation (right column)." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.38, + 0.906, + 0.44 + ], + "angle": 0, + "content": "
SettingMagic3DDreamFusionDreamFusion*DreamHOIOurs
Full4.36.517.326.045.6
w/o Refine6.56.517.330.439.1
w/o LLM-init8.610.821.732.626.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.45, + 0.906, + 0.478 + ], + "angle": 0, + "content": "Table 3. Ablation study on LLM-guided initialization and HOI refinement by GPT-4V selection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.749 + ], + "angle": 0, + "content": "level of detail and the physical accuracy of the interactions. Magic3D [29], for instance, often generates overly simplified shapes with minimal texture and detail, as seen in the keyboard typing example, where the lack of intricacy undermines realism. Both Magic3D [29] and DreamFusion [43] fail to capture body structure or realistic human-object relationships, highlighting the difficulty of generating realistic HOI using general 3D scene generation models. DreamFusion*, which incorporates object mesh as input, retains a complete or partial human body shape but struggles with accurate contact and natural scale. This limitation is particularly noticeable in examples like the holding-chair scenario, where hand placement lacks precision. DreamHOI [67] improves upon DreamFusion in terms of HOI but still falls short in capturing fine interaction details, such as accurate foot placement on the basketball." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Inpainting Visualization. We further visualize the 2D inpainting results and affordance parsing on the same object in Figure 5 to emphasize the effectiveness of our method in disentangling affordances for distinct human-object interactions. The figure showcases various human object interactions with a chair, highlighting the ability to localize affordances accurately. This visualization provides valuable insight into the regions on the object that are most likely to be in contact with humans, which is critical for optimizing human-object interaction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7021" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.472, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.166, + 0.472, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.252, + 0.483, + 0.28 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of scene populating and scene interaction applications from different views." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.296, + 0.485, + 0.508 + ], + "angle": 0, + "content": "Qualitative performance on Scene Populating. We demonstrate one possible way to populate scenes with InteractAnything. We first utilize the HPS dataset [14] to obtain the 3D scene and then integrate the generated HOI results into the scene by placing them in appropriate locations. By employing a ground loss and large language model, we ensure that the HOI results are naturally embedded within the scene without artifact issues such as object or human interpenetration. Furthermore, our method goes beyond simply populating HOI results into the scene, we also generate realistic interaction effects with existing objects in the environment. For example, InteractAnything can effectively generate interactions such as \"sitting on the table\" or \"opening a door\", as demonstrated in the Figure 6." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.517, + 0.245, + 0.533 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Ablation on Fine-grained Optimization In this study, we analyze the effect of fine-grained optimization terms as outlined in Section 3.5 while keeping all other parameters constant. The results shown in Figure 4, clearly demonstrate that the fine-grained optimization significantly enhances the realism and contact plausibility of human-object interactions (right green panel) in the \"grasp\" prompt. Without fine-grained optimization (left red panel), the alignment between the hand and table is noticeably inaccurate, resulting in unnatural poses where the hand appears to \"float\" near the table instead of properly grasping or holding it. In contrast, with fine-grained optimization enabled (right panel), the human hand is aligned with the table surface, resulting in a more realistic interaction. This improvement highlights the importance of fine-grained optimization in achieving natural contact and precise positioning, especially in actions involving detailed hand-object interactions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Ablation Study on LLM-guided Initialization We further ablate the LLM-guided initialization and the corresponding HOI refinement module (fine-grained HOI synthesis) to investigate the contribution of our design in leveraging language models. As shown in Table 3, we use GPT-4V selection as the evaluation metric. Compared to the full method (Full), the variant without refinement (w/o Refine)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.304 + ], + "angle": 0, + "content": "exhibits a noticeable decline in performance. This drop is primarily due to the absence of fine-grained contact optimization, which is critical for generating the most accurate human-object interactions. In contrast, when the LLM-guided initialization is ablated (w/o LLM-init), we randomly initialize both the human and object positions, followed by coarse HOI optimization. This approach results in a significant decrease in selection scores, ranking lower than DreamHOI [67]. This finding emphasizes the importance of an informed initialization strategy in establishing a solid foundation for HOI optimization. The stark differences in the ablation study underscore the effectiveness of our design. The integration of LLM-guided initialization and refinement is vital for achieving superior results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.317, + 0.757, + 0.334 + ], + "angle": 0, + "content": "5. Discussion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.909, + 0.706 + ], + "angle": 0, + "content": "In this work, we demonstrate that our approach can generate fine-grained, realistic HOIs for open-set 3D objects based on text descriptions. While our method shows promising results in improving contact precision and adapting to unseen objects, it still relies on 2D priors as its primary guidance. Consequently, advancements in multi-view consistent image generation or fully 4D generation techniques would likely enhance our approach, leading to more coherent and robust 3D interactions. Additionally, physical plausibility is also an important application in recent applications, e.g., in humanoid manipulation. We use force closure as a physical constraint while a high-quality simulator could further evaluate the generated HOI. Furthermore, our method currently uses the SMPL-H model to represent human body geometry and articulation. While this model captures a wide range of human poses and movements, it limits the adaptability of our framework to diverse interaction agents, particularly those with non-human anatomies. These agents often have unique skeletal structures and movement patterns that differ significantly from human biomechanics, which our current system is not fully equipped to process. In future work, we aim to address these limitations by investigating multi-view or 4D generation for enhanced consistency and exploring more flexible models to represent a broader array." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.719, + 0.634, + 0.735 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.743, + 0.907, + 0.895 + ], + "angle": 0, + "content": "This paper tackles the challenge of generating novel 3D HOI involving unseen objects—a task where current methods often struggle. To address this issue, we propose InteractAnything, a novel zero-shot synthesis method driven by text instructions and not reliant on training with existing 3D datasets. It leverages guidance from the large language model to distill human-object relations and employs the pre-trained 2D diffusion model to parse contact areas with open-set objects, circumventing the limitations of existing HOI datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7022" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.155 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by the National Science and Technology Major Project (2022ZD0114900) and NSFC-6247070125. We extend our gratitude to the reviewers for their insightful comments." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.167, + 0.188, + 0.183 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.192, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.263, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[2] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.334, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[3] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep It SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. In Computer Vision - ECCV 2016, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.391, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[4] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 958-968, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.462, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[6] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[7] Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets creation using spatially-aware diffusion guidance. In European Conference on Computer Vision, pages 128-146. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[8] Hongkai Dai, Anirudha Majumdar, and Russ Tedrake. Synthesis and optimization of force closure grasps via sequential semidefinite programming. Robotics Research: Volume 1, pages 285-305, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[9] Sisi Dai, Wenhao Li, Haowen Sun, Haibin Huang, Chongyang Ma, Hui Huang, Kai Xu, and Ruizhen Hu. Interfusion: Text-driven generation of 3d human-object interaction. In European Conference on Computer Vision, pages 18-35. Springer, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[10] Google DeepMind. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Clement Fuji Tsang, Maria Shugrina, Jean Francois Lafleche, Towaki Takikawa, Jiehan Wang, Charles Loop, Wenzheng Chen, Krishna Murthy Jatavallabhula, Edward Smith, Artem Rozantsev, Or Perel, Tianchang Shen, Jun Gao, Sanja Fidler, Gavriel State, Jason Gorski, Tommy Xiang, Jianing Li, Michael Li, and Rev Lebaredian. Kaolin: A pytorch library for accelerating 3d deep learning re" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.192, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "search. https://github.com/NVIDIAGameWorks/kaolin, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[12] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Scholkopf. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[13] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13041-13051, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[14] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human positioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[15] Sookwan Han and Hanbyul Joo. Chorus: Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In International Conference on Computer Vision (ICCV), 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.904, + 0.426 + ], + "angle": 0, + "content": "[16] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambiguities with 3D scene constraints. In International Conference on Computer Vision (ICCV), 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[17] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[18] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[19] Jingwei Huang, Hao Su, and Leonidas Guibas. Robust watertight manifold surface generation method for shapenet models. arXiv preprint arXiv:1802.01698, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.582, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[20] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Zhiyuan Zhang, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. Full-body articulated human-object interaction. In International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[21] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[22] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7122-7131, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[23] Tero Karras. Maximizing parallelism in the construction of bvhs, octrees, and k-d trees. In Proceedings of ACM SIGGRAPH/Eurographics Conference on High-Performance Graphics, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):139-1, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Hyeonwoo Kim, Sookwan Han, Patrick Kwon, et al. Beyond the contact: Discovering comprehensive affordance for 3d" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7023" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "objects from pre-trained 2d diffusion models. In European Conference on Computer Vision (ECCV), 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[26] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.483, + 0.22 + ], + "angle": 0, + "content": "[27] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. In European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.483, + 0.277 + ], + "angle": 0, + "content": "[28] Lei Li and Angela Dai. Genzi: Zero-shot 3d human-scene interaction generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20465-20474, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[29] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.483, + 0.418 + ], + "angle": 0, + "content": "[30] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21159-21168, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.483, + 0.476 + ], + "angle": 0, + "content": "[31] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[32] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters (RA-L), 7(1):470-477, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[33] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splatting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[34] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In International Conference on Learning Representations (ICLR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[35] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Transactions on Graphics (TOG), 34(6):248:1-248:16, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[37] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[38] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[39] OpenAI. Gpt-3.5. https://platform.openai.com/docs/models/gpt-3-5, 2023. Accessed: 2023-11-14. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.299 + ], + "angle": 0, + "content": "[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[41] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[42] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[43] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations (ICLR), 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[44] Alec Radford, JongWook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Askell Amanda, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. Cornell University - arXiv, Cornell University - arXiv, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.62 + ], + "angle": 0, + "content": "[45] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[48] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: modeling and capturing hands and bodies together. ACM Transactions on Graphics (TOG), 36(6):1-17, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[49] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7024" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[51] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[52] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[53] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.388 + ], + "angle": 0, + "content": "[54] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J Black. Deco: Dense estimation of 3d human-scene contact in the wild. In International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[55] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In International Conference on 3D Vision (3DV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[56] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[57] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16210-16220, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[58] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In International Conference on Computer Vision (ICCV), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[59] Sirui Xu, Yu-Xiong Wang, Liangyan Gui, et al. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. Advances in Neural Information Processing Systems, 37:52858-52890, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[60] Hongwei Yi, Justus Thies, Michael J Black, Xue Bin Peng, and Davis Rempe. Generating human interaction motions in scenes with text control. In European Conference on Computer Vision (ECCV), 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[61] Wanqi Yin, Zhongang Cai, Ruisi Wang, Fanzhou Wang, Chen Wei, Haiyi Mei, Weiye Xiao, Zhitao Yang, Qingping Sun, Atsushi Yamashita, et al. Whac: World-grounded humans and cameras. In European Conference on Computer Vision, pages 20-37, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[62] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monoc" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ular images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:12287-12303, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[63] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13232-13242, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[64] Jason Y. Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In European Conference on Computer Vision (ECCV), 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[65] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[66] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on pattern analysis and machine intelligence, 44(6): 3170-3184, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[67] Thomas Hanwen Zhu, Ruining Li, and Tomas Jakab. Dreamhoi: Subject-driven generation of 3d human-object interactions with diffusion priors. arXiv preprint arXiv:2409.08278, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7025" + } + ] +] \ No newline at end of file diff --git a/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/bba3c5bb-a0b8-4050-9fca-51eeaf9696e0_origin.pdf b/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/bba3c5bb-a0b8-4050-9fca-51eeaf9696e0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cc50c831856373cdbafc1b09be1eae4407713351 --- /dev/null +++ b/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/bba3c5bb-a0b8-4050-9fca-51eeaf9696e0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01200772db386b5e1d488834effe5bc9f5924272fdd315bc4a1b390a1d5642a0 +size 6783553 diff --git a/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/full.md b/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c2ee183ce49cf9f01912512be93fb00e9845344b --- /dev/null +++ b/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/full.md @@ -0,0 +1,330 @@ +# InteractAnything: Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing + +Jinlu Zhang $^{1,2}$ , Yixin Chen $^{2\boxtimes}$ , Zan Wang $^{2,3}$ , Jie Yang $^{2,4}$ , Yizhou Wang $^{1,5,6,7\boxtimes}$ , Siyuan Huang $^{2\boxtimes}$ + +indicates corresponding authors + +1 Center on Frontiers of Computing Studies, School of Computer Science, Peking University + +$^{2}$ State Key Laboratory of General Artificial Intelligence, BIGAI $^{3}$ Beijing Institute of Technology + +4 The Chinese University of Hong Kong, Shenzhen 5 Nat'l Eng. Research Center of Visual Technology, Peking University + +$^{6}$ Institute for AI, Peking University $^{7}$ State Key Laboratory of General Artificial Intelligence, Peking University + +![](images/13c688d3961fa06fb80cf67c7892f1ba4d1ca5f354ac86ad64e501a827ca28a5.jpg) +A person holds/pulls/lies on the chair... +(a) Diverse human interactions. +Figure 1. 3D human object interaction synthesis by InteractAnything. Given a simple text description with goal interaction and any object mesh as input, our method enables the generation of diverse, natural, detailed, and novel interactions for open-set 3D objects in a zero-shot manner. The orange and green boxes of (b) indicate detailed contact poses from different views. + +![](images/f882f26ba7af8d8816b8e73c4f291654525d23f4685d548ed3e15673a09505fc.jpg) +A person grasps the +backpack/table/ball/knife... +(b) Detailed interactions synthesis of open-set objects. + +![](images/c2b3b4f1fd64f2cc3bd30e725e65a4451665dec03e58629b40a7f0a26da591ce.jpg) +A person lifts/rides/criddles the +tesla car/motorcycle/horse/baby... +(c) Novel interactions given generative objects. + +# Abstract + +Recent advances in 3D human-aware generation have made significant progress. However, existing methods still struggle with generating novel Human Object Interaction (HOI) from text, particularly for open-set objects. We identify three main challenges of this task: precise human-object relation reasoning, affordance parsing for any object, and detailed human interaction pose synthesis aligning description and object geometry. In this work, we propose a novel zero-shot 3D HOI generation framework without training on specific datasets, leveraging the knowledge from largescale pre-trained models. Specifically, the human-object relations are inferred from large language models (LLMs) to initialize object properties and guide the optimization process. Then we utilize a pre-trained 2D image diffusion model to parse unseen objects and extract contact points, avoiding the limitations imposed by existing 3D asset knowledge. The initial human pose is generated by sam + +pling multiple hypotheses through multi-view SDS based on the input text and object geometry. Finally, we introduce a detailed optimization to generate fine-grained, precise, and natural interaction, enforcing realistic 3D contact between the 3D object and the involved body parts, including hands in grasping. This is achieved by distilling human-level feedback from LLMs to capture detailed human-object relations from the text instruction. Extensive experiments validate the effectiveness of our approach compared to prior works, particularly in terms of the fine-grained nature of interactions and the ability to handle open-set 3D objects. Project page: jinluzhang.site/projects/interactanything. + +# 1. Introduction + +Synthesizing the interaction of humans and any given object holds great significance across diverse applications, including AR/VR, computer simulation, and animation. Tradi- + +tionally, creating high-quality 3D human object interaction (HOI) assets requires expensive, lab-controlled data capture systems and significant human effort. Although some datasets [2, 16, 20, 21, 26, 53] have collected these data, the models [26, 27, 41, 54, 58, 60] trained on them can only generate interactions for specific object categories. These methods struggle to handle the diversity of object geometries, scales, and interactions at the same time, making text-driven zero-shot HOI synthesis a challenging and highly demanding task. + +Given recent advances in text-to-3D techniques [6, 29, 56], e.g., leveraging Score Distillation Sampling (SDS) [43] to generate 3D content from text, a natural question is whether this paradigm can be applied to zero-shot HOI synthesis. However, directly integrating 3D humans and objects into a scene without understanding the complex relationships between them remains challenging. Recent works [7, 9, 67] are proposed to generate realistic 3D HOIs. Among them, InterFusion [9] uses Neural Radiance Fields (NeRF) [37] for individual human and object representations but struggles with precise interaction area, leading to suboptimal and unrealistic HOIs. DreamHOI [67] optimizes on both mesh and NeRF representations through bidirectional rendering but faces challenges in ensuring accurate contact points and relative scales due to solely optimizing human states while keeping the given object static, leading to limited interactions. + +In this paper, we first identify the core challenge of zero-shot 3D HOI generation as accurately capturing human-object relations, contact areas, and global consistency, which are not typically presented in single-object or non-contact generation scenarios. These challenges prevent the generation of high-quality, geometrically precise, and textually consistent results. Furthermore, while large-scale pre-trained models, e.g., 2D diffusion models [42, 47] and large language models (LLMs) [1, 10, 39], cannot fully capture the complexity of human object interactions (including body dynamics, expressive details, and object geometry), they can offer valuable prior knowledge of common interaction patterns based on textual descriptions. Consequently, we argue that the solution for effective generation is to leverage this diverse prior knowledge to guide the optimization of interactions. + +Following the above insights, we decompose the complex zero-shot 3D HOI task into these stages: 1) human-object relation reasoning and initialization, 2) open-set object contact affordance parsing, and 3) human pose synthesis aligning text and object geometry. We first leverage the LLM as a human-level feedback provider, which reasons and unravels complex semantic interaction relationships, e.g., relative position and orientation, from simple text input. It initializes human and object states and designs a series of optimization functions to generate detailed inter + +actions. The initial human pose is optimized by multi-view SDS loss for interaction in line with the given textual description and object geometry. + +To parse the 3D object affordance, we utilize initialized human-object relation and sample adaptive inpainting masks during the diffusion process to obtain the contact information. This enables the extraction of contact probabilities on the object, which are then used to segment and define the weighted dynamic contact regions. Finally, we design a coarse-to-fine optimization based on force closure [8, 32] to generate expressive, diverse, and physically plausible interactions, allowing for more accurate and detailed contact. To summarize, our contributions are: + +- We propose InteractAnything for zero-shot 3D HOI generation, which effectively distills human and object interactions from LLM and pre-trained diffusion models to synthesize diverse, detailed, and novel interactions. +- We design a novel open-set object affordance parsing, which generates contact maps on 3D geometry by utilizing adaptive inpainting and potential contact distribution. +- Our model leverages both text and object geometry to drive human pose synthesis, enabling realistic and contextually appropriate interactions for open-set objects. Results show that our synthesized pose accurately reflects the semantic meaning of the input while respecting the spatial constraints of the object, leading to improved performance, as shown in Figure 1. + +# 2. Related Work + +# 2.1. Text-driven 3D Content Generation + +The generation of 3D content from text has seen rapid advancements, primarily driven by diffusion-based models that build on text-to-2D image generation approaches [45, 46, 49, 50]. One pioneering approach, DreamFusion [43], introduced Score Distillation Sampling (SDS), which leverages the pre-trained 2D diffusion model to optimize 3D content through techniques like NeRF (Neural Radiance Fields) [37] and 3D Gaussian Splatting [24]. A significant area of research aims to address the complexity of 3D human generation. For example, recent methods have been proposed for generating controllable 3D human avatar [4, 17, 18, 33, 57, 65]. However, these works focus on diverse and high-quality appearance but animate human avatars with given human poses or motion, which do not satisfy 3D HOI requirements in human pose synthesis. The proliferation of large 3D datasets has also catalyzed progress, leading to more sophisticated models like Zero-1-to-3 [31], Zero $123++$ [51], and MVDream [52], which employ 2D diffusion models to generate multi-view consistent images that serve as inputs for efficient 3D generation systems, such as SyncDreamer [34] and Wonder3D [35]. Despite these advancements, generating complex HOI scenes, + +especially those involving interactions between humans and objects, remains challenging. + +# 2.2. Data-Driven Human Object Interaction + +Previous approaches [26, 58-60] primarily train generative models to produce static human poses or dynamic motions in response to objects or scenes using curated HOI datasets. Recent works focus on dynamic motion generation performance conditioned on encoded object features but rely heavily on specialized datasets, limiting generalization to diverse objects. For example, OMOMO [26] generates human motion through diffusion models but requires object motion as input. InterDiff [58] predicts full-body interactions based on prior HOI context but is limited to the prediction task, restricting flexibility in new scenarios. In contrast, our method jointly synthesizes human and object states and adapts to novel objects without extensive, object-specific training by extracting interaction patterns directly from 2D diffusion models. + +On the other hand, optimization-based methods reconstruct realistic human poses and transformations from images. PHOSA [64] optimizes human object relations using depth-aware loss functions. PROX [16] encodes contact probabilities and semantic scene labels to position humans within scenes. CHORUS [15] synthesizes multi-view images with a pre-trained generative model to learn 3D spatial relationships. By extracting knowledge from LLM to infer human poses, the method in [55] queries contact information to reconstruct 3D human-object interactions. Compared to these methods, our method leverages pre-trained diffusion models for accurate parsing of object affordance and enables fine-grained interactions. + +# 2.3. Zero-shot 3D Compositional Generation + +To achieve the compositional scene generation, Combo-Verse [7] proposes a spatially-aware SDS loss, generating the final 3D scene through single-object reconstruction and multi-object composition. GraphDreamer [12] leverages scene graphs to guide generation by converting the semantic information into a global textual description to enable text-driven compositional scene generation. These methods focus on the reasonable arrangement of multiple objects but ignore the contact generation in HOI scenes. InterFusion [9] and DreamHOI [67] present this challenge and then try to address using SMPLs [36, 48] as body model prior. DreamHOI [67] designs the SDS guidance mixture to separately render human interactions using different diffusion models to improve human modeling quality. However, it cannot ensure precise contact during the denoising process. Compared to these methods, our work focuses on addressing contact areas parsing on open-set objects. + +# 3. Our Method + +Task Definition and Overview. Given a text prompt $\epsilon$ that describes the desired HOI, e.g., "A person grasps the chair". Our goal is to synthesize natural human mesh $\nu_{H}$ and object mesh $\nu_{O}$ that align with $a$ and $\epsilon$ . As shown in Figure 2, we first obtain initial human-object relations from LLM based on text (Section 3.2). For object optimization, we initialize object rotation $r_o$ , translation $t_o$ , scale $s_o$ , state $s$ , based on affordance map $\mathcal{M}$ (Section 3.3). We also synthesize the initial human pose based on the input text and object geometry (Section 3.4). Finally, we use SMPL-H as representation and optimize pose parameter $\theta_b, \theta_h$ , global translation $t_b$ , and object rotation and translation $r_o, t_o$ to generate detailed and diverse HOI (Section 3.5). + +# 3.1. Preliminary Knowledge + +Human Body Model. We use the SMPL-H [36, 48], a parametric model that includes both body and hand parameters, to represent the human body and hand articulation. The body pose is defined by body parameters $\theta_{\mathrm{b}} \in \mathbb{R}^{3(K_b + 1)}$ and the hand pose is represented by additional MANO parameters $\theta_{\mathrm{h}} \in \mathbb{R}^{3K_h}$ , capturing 15 joints per hand. The shape parameters are represented by $\beta \in \mathbb{R}^{10}$ , controlling the deformation of the body and hand template mesh to target body shapes. Given the combined pose parameters $\{\theta_{\mathrm{b}}, \theta_{\mathrm{h}}\}$ and shape parameters $\beta$ , SMPL-H computes mesh vertices $\mathcal{V} \in \mathbb{R}^{V \times 3}$ using linear blend skinning function $f$ : + +$$ +\mathcal {V} (\beta , \theta_ {b}, \theta_ {\mathrm {h}}) = f \left(T _ {p} \left(\beta , \theta_ {b}, \theta_ {\mathrm {h}}\right), J (\beta), \theta , \mathcal {W}\right), \tag {1} +$$ + +where $J$ represents joint locations, $\mathcal{W} \in \mathbb{R}^{V \times (K)}$ is skimming weights with all $K$ joints. + +Score Distillation Sampling (SDS). SDS leverages pretrained diffusion models to guide the synthesis of 3D scenes by aligning NeRF with a text prompt. It operates by iteratively adjusting NeRF parameters, denoted as $\theta$ , based on gradients computed to align the generated images with the textual input. Given a noisy 2D image $x_{t}$ (obtained by adding noise $\epsilon$ to a clean NeRF-rendered image $x$ ), a textual prompt $y$ , and a noise level $t$ , SDS calculates the gradient of the loss $\mathcal{L}_{\mathrm{SDS}}$ with respect to $\theta$ as follows: + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {S D S}} (x) = \mathbb {E} _ {t, \epsilon} \left[ w (t) (\hat {\epsilon} (x _ {t}; y, t) - \epsilon) \frac {\partial x}{\partial \theta} \right], +$$ + +where $\hat{\epsilon}(x_t; y, t)$ represents the diffusion model predicted noise given the noisy image $x_t$ , prompt $y$ , and noise level $t$ , $w(t)$ is a weighting function that adjusts the influence of the gradient based on the noise level $t$ . This gradient is used to update $\theta$ , guiding the NeRF model toward producing images that match the textual description. For instance, in DreamFusion [43], this process enables text-to-3D generation by repeatedly refining the NeRF parameters $\theta$ from a random initialization, iteratively producing images that align more closely with the target prompt. + +![](images/a0f0efb2da3f61d82118f3a4d802002ff84417a895f28f2e1bf39c6eff39f3d6.jpg) +Figure 2. Framework of InteractAnything. Given a text description and any object mesh as input, our approach begins by querying LLM to infer precise human-object relationships, which are used to initialize object properties. Next, we analyze the contact affordance of the object geometry. The human pose is synthesized using a pre-trained 2D diffusion model, guided by multi-view SDS loss and the designed spatial constraint. Finally, based on the targeted object contact areas and a plausible human pose, we perform expressive HOI optimization to synthesize realistic and contact-accurate 3D human-object interactions. + +# 3.2. LLM-guided Initialization + +To infer diverse intended human object interactions for a given object $\mathcal{V}_O$ and action $a$ , we query the LLM to provide commonsense feedback. To minimize hallucination, we need to transfer inferred relationships from GPT [39] into specific values that position the human and object meshes accurately, therefore, we enable the language model to select answers from pre-defined options step by step, instead of directly generating responses. For initializing the object rotation $r_o$ , translation $t_o$ , scale $s_o$ , and state $s$ , the language model selects the relative position of the human to the object from defined options, which is then mapped to $t_o$ based on the combined size of the human and object. We set all input objects to have $y$ -up and $x$ -front orientations, with $r_o$ adjusted based on which side of the object should face the human. The object scale is chosen from 4 defined options and set relative to human size. Lastly, the state $s$ determines if the object should be positioned on the ground based on whether it is meant to move. Besides, we also query interacted semantic body parts $S_h$ and map the labels to human mesh to obtain vertices $\mathcal{P}_h$ for part-level optimization using annotated human part labels. + +# 3.3. Open-set Object Affordance Parsing + +To obtain the contact affordance of unseen object geometries given a specified action e.g., holds, we conduct 2D inpainting based on a pre-trained diffusion model to dis + +till possible human interaction parts through LLM-guided adaptive mask generation. Then we design a 2D probabilistic contact function to model the distribution of the interaction, generating an affordance map to identify and extract the regions for HOI optimization. + +2D Adaptive Mask Generation Image inpainting is to fill in missing regions of an image, where the missing regions can be defined by a mask image. To get knowledge of possible interacted parts, we utilize the powerful understanding and generation ability of pre-trained inpainting model [42]. Given 2D projection images of the 3D object from $n_p$ views, we found existing inpainting models cannot ensure complete human in the image (many inpainting results only include interacted body parts, e.g., hands in "hold" action). Therefore, different from previous methods [25, 28], we do not randomly sample the inpainting masks on 2D images, instead, our method identifies possible human interaction parts to improve the inpainting success rate by generating inpainting masks from LLM feedback in two forms: (1) full-body mask with respect to object properties $r_o$ , $t_o$ , person position and person body size; (2) body-part inpainting masks for intended interacted body part labels $S_h$ and corresponding human vertices $\mathcal{P}_h$ . + +More specifically, we first create a full-body inpainting mask guided by LLM-inferred initializations $r_o$ , $t_o$ , and body size $s_f$ : + +$$ +M _ {\mathrm {f u l l}} ^ {i} = \mathcal {J} ((\mathcal {V} _ {H} \cdot r _ {o} ^ {- 1}), s _ {f}, c ^ {i}) + t _ {o} ^ {- 1} + \varpi_ {1}, \qquad (2) +$$ + +![](images/2a23e3e34d611ce29e880c0b407320d93419734c546e3d75ec0b79aafbd51ae1.jpg) +Figure 3. Qualitative comparison results with baselines. * indicates we re-implement this method by embedding object mesh into the diffusion process, which follows the pipeline of DreamHOI [67] and our method. More visualization results are presented in supplementary materials. + +where $\mathcal{I}$ is 2D projection transformation with camera parameter $c^i$ , and $\varpi_{1}$ is the noise term. Similarly, we generate body-part-level inpainting masks using given body part labels $S_{h}$ and corresponding human vertices $\mathcal{P}_h$ and body part size $s_p$ to help inpainting body parts like hands as the above formulation: + +$$ +M _ {\text {p a r t}} ^ {i} = \mathcal {J} \left(\left(\mathcal {P} _ {h} \cdot r _ {o} ^ {- 1}\right), s _ {p}, c ^ {i}\right) + t _ {o} ^ {- 1} + \varpi_ {2}. \tag {3} +$$ + +Besides, We apply the above full-body masks $M_{\mathrm{full}}$ into the dynamic mask inpainting pipeline [25], which can further refine the initial mask detecting human on intermediate results in the diffusion process. + +Affordance Parsing on 2D Probability Function. Directly computing contact probabilities on a 3D object mesh relies heavily on existing 3D human reconstruction methods [3, 13, 22, 30, 61-63] from single-view images. This encounters significant 3D ambiguity and quantization error, which obscures accurate 3D human pose. In contrast to previous methods [15], we apply 2D human pose detector OpenPose [5], performing well at detecting 2D human keypoints and even occluded body parts, to estimate human from inpainting results. Then we introduce a 2D probability function that calculates contact probability according to the distance from detected body keypoints to the multi-view object projection images. First, we downsample the object + +to $D$ points. Let $\mathcal{P}_i$ represent the probability map for each view $i$ , the distance-based probability $f_{\mathrm{afford}}(\mathbf{p})$ for a point $\mathbf{p} \in \mathcal{V}_O$ on the object surface is defined as: + +$$ +f _ {\text {a f f o r d}} ^ {(i)} (\mathbf {p}) = e ^ {- \left\| \mathbf {d} _ {i} (\mathbf {p}) \right\|}, \tag {4} +$$ + +where $\mathbf{d}_i(\mathbf{p})$ denotes the distance from the body part to the object mask in $i$ view. Then we re-project the 2D affordance $f_{\mathrm{afford}}^{(i)}(\mathbf{p})$ to 3D space using defined camera parameters, and aggregate the contact probabilities across all $n_p$ views to update the 3D contact probability distribution as follows: + +$$ +\mathcal {P} (\mathbf {p}) = \frac {1}{n _ {p}} \sum_ {i = 1} ^ {n _ {p}} f _ {\text {a f f o r d}} ^ {(i)} (\mathbf {p}). \tag {5} +$$ + +Finally, we normalize $\mathcal{P}(\mathbf{p})$ over the entire object surface to obtain contact probability for each point is determined by its projected 2D distances across multiple views, effectively mitigating multi-view ambiguity and enabling accurate estimation of human object contact regions. + +# 3.4. Text-Object Driven Human Pose Synthesis + +Considering a brief HOI description, we posit that current 3D generation models, such as DreamFusion [43], are capable of generating human avatars in a variety of poses; however, they do not guarantee precise contact with the object + +mesh. Thus, we first apply prompt engineering to generate diverse and expressive text descriptions, along with negative prompts, using large language models (LLMs) to mitigate issues like low-quality outputs and missing limbs. Next, we incorporate Multi-view SDS loss [52] to optimize the human pose aligning the interaction description and object geometry. We first scale the input object based on $s_o$ given by the language model. We use the spatial constraint to prevent the human avatar from penetrating the object mesh, ensuring alignment with spatial relations. Specifically, we embed object mesh within NeRF optimization by regularizing the ray sample points in object mesh following previous works [9, 67]. After training, we use the multi-view human pose estimation method [66] to obtain the body parameters $\theta_b$ and then init hand parameters $\theta_h$ . We also employ VPoser [40] to maintain reasonable joint rotations. + +# 3.5. Expressive HOI Optimization + +Global HOI optimization. With the above initialization as input, we first ensure the input object is watertight using Manifold [19]. Then we conduct human-centric optimization to update object properties $(r_o, t_o, s_o)$ as follows: + +$$ +L _ {c} = \phi_ {i} L _ {i n t e r} + \phi_ {n} L _ {n} + \phi_ {s} L _ {s c a l e} + \phi_ {p} L _ {p e n e} + \mathbf {s} ^ {'} \cdot L _ {g}, (6) +$$ + +where $\phi$ is loss weights and interaction loss is designed on. +verts of the interacted body part $\mathcal{P}_h,h\in \mathcal{V}_H$ and object +contact areas $\mathcal{P}_o,o\in \mathcal{M}$ .. + +$$ +L _ {i n t e r} = \sum_ {i \in [ h ], j \in [ o ]} W _ {\mathcal {M}} (j) \cdot f _ {c h a m} \left(\mathcal {P} _ {h}, \mathcal {P} _ {o}\right), \tag {7} +$$ + +where $W_{\mathcal{M}}(j)$ indicates probabilities scores in object affordance map $\mathcal{M}$ of the $j$ -th object vertex and $f_{cham}$ is chamfer distance function. We constrain the normals of interacted parts to maintain the relative orientation: + +$$ +L _ {n} = \sum_ {n \in [ h, o ]} \left(\mathbf {n} _ {h} \cdot \mathbf {n} _ {o} + 1\right) ^ {2}, +$$ + +where $\mathbf{n}_h$ is normals of interacted body vert, and $\mathbf{n}_o$ indicates object normals of interacted vert. In addition, we employ the object scale loss $L_{\mathrm{scale}}$ [64] and the penetration term $L_{\mathrm{pene}}$ to prevent collisions [23]. The introduction of ground loss $L_{g}$ is decided by the object state $\mathbf{s} = [0,1]$ . + +Finer HOI Optimization. During this stage, we optimize the human parameters $\theta_{b},\theta_{h}$ and $t_b$ and fix the object parameters. To synthesize detailed and diverse interaction poses for full-body optimization, inspired by differentiable force closure [32], we design the fine-grained optimization loss as below: + +$$ +L _ {f} = \delta \cdot L _ {f c} + L _ {p e n e} + L _ {i n t e r}, \tag {8} +$$ + +where loss functions $L_{\mathrm{pene}}$ and $L_{\mathrm{inter}}$ are analogous to the aforementioned terms, but with the incorporation of + +
MetricMagic3D [29]DreamFusion [43]DreamHOI [67]Ours
CLIP Score↑0.24160.24370.27570.2968
+ +Table 1. Comparison of CLIP similarity scores for different methods. Higher scores are better. + +Kaolin [11] to better sample surface distance. We introduce SMPL-based force closure loss $L_{fc}$ by replacing the base model from MANO to SMPL-H to conduct more detailed interaction poses: + +$$ +L _ {f c} = \sum_ {j \in [ o ]} \left(\sum_ {i \in [ h ]} f v (i, j) \cdot n (j)\right) ^ {2}, \delta \cdot \left[ \theta_ {b}, \theta_ {h} \right]. \tag {9} +$$ + +$fv$ symbolizes the force vector applied at vertex $j$ from interacted human vertex $i$ , and $\delta$ decide optimize body pose $\theta_{b}$ or hand pose $\theta_{h}$ . Specifically, we make a series of augmentations to improve force closure acceptance rate and robustness. We decrease contact points by sparse uniform sampling on the interacted vertices. Moreover, we initialize body and hand positions around the interacted object parts and maintain the hand normals as positive to the interacted object parts. Finally, we extract the interacted vertices to form interacted mesh and non-interacted mesh and compute $L_{fc}$ and $L_{pene}$ separately. + +# 4. Experiments + +# 4.1. Data and Metrics + +We collect evaluated object meshes in 3 ways: BEHAV [2], PartNet [38], and sampling from generative models. BEHAV dataset is commonly used in this field, and PartNet is manually annotated in an object mesh with part-level semantic labels on the vertices. For metrics, we use CLIP similarity score [44] between our prompts and the rendering images of the HOI scene, by rendering our human and object meshes from 4 different views randomly on 5 prompts and 10 objects categories. During this process, we select 5 BEHAV objects (backpack, basketball, chairwood, keyboard, and suitcase), and 5 objects from publicly accessed 3D generative models (motorcycle, car, television, baby doll, and humanoid robot). The same evaluation setting is adapted to other comparison methods. In addition, while the CLIP score is designed to evaluate how well an image aligns with the input text, it often fails to capture finer nuances in interactions, leading to less distinct variations in the metrics. We follow InterFusion [9] to leverage the advanced image understanding capabilities of GPT-4V further to assess the performance of baselines and our method. The text prompts are designed and then augmented using GPT. + +![](images/56240666b57051852353bc0d4ff7db70a960212d273042d2126c7e0deea1cda4.jpg) +Figure 4. Ablation study on the fine-grained optimization. Left figures are the results of removing fine-grained terms and right figures apple fine-grained terms to synthesize grasping details. + +
MetricsMagic3DDreamFusionDreamFusion*DreamHOIOurs
Overall4.36.517.326.045.6
Contact8.64.315.219.552.1
+ +# 4.2. Quantitative Evaluations + +CLIP Score. As shown in Table 1, we present a comparison of CLIP similarity scores across various methods, where higher scores indicate better alignment with the target descriptions. Our method InteractAnything, achieves the highest mean score across the evaluation prompts, outperforming baselines such as DreamFusion [43], Magic3D [29], and DreamHOI [67]. These results demonstrate that our approach effectively captures the semantic interactions of the text descriptions. Notably, to better evaluate interaction performance without being confounded by appearance, we apply random human textures to the body mesh and retain the original textures for objects. + +GPT-4V Selection. As presented in Table 2, we enable the vision language model [1] to choose the most reasonable result in terms of Overall and Contact from all generated outputs based on criteria such as the presence of a full human body, a complete object, and correct physical interactions, and then returning the index of the selected result. No in-context examples are provided to guide the model in the Overall setting, we give text prompts only for the Contact evaluation, and the order of generated results is randomly shuffled to avoid bias, the final results are obtained from multiple tests. + +# 4.3. Qualitative Evaluations and Application + +Visualization Comparison. The comparison results are shown in Figure 3, where our method outperforms the baselines, producing the most realistic and plausible 3D human-object interactions. This is evident in both the + +![](images/96482e8b28fecbdc0483ee18124a6abefd2699cf01c84b665d83290d9a9b2708.jpg) +Figure 5. Visualization of the adaptive inpainting and open-set object affordance parsing results on the same object with different text instructions. Our method first generates reasonable 2D inpainting results (middle columns) and then computes contact probabilities as the affordance representation (right column). + +Table 2. Quantitative evaluation of GPT-4V selection in terms of Overall and Contact. + +
SettingMagic3DDreamFusionDreamFusion*DreamHOIOurs
Full4.36.517.326.045.6
w/o Refine6.56.517.330.439.1
w/o LLM-init8.610.821.732.626.1
+ +Table 3. Ablation study on LLM-guided initialization and HOI refinement by GPT-4V selection. + +level of detail and the physical accuracy of the interactions. Magic3D [29], for instance, often generates overly simplified shapes with minimal texture and detail, as seen in the keyboard typing example, where the lack of intricacy undermines realism. Both Magic3D [29] and DreamFusion [43] fail to capture body structure or realistic human-object relationships, highlighting the difficulty of generating realistic HOI using general 3D scene generation models. DreamFusion*, which incorporates object mesh as input, retains a complete or partial human body shape but struggles with accurate contact and natural scale. This limitation is particularly noticeable in examples like the holding-chair scenario, where hand placement lacks precision. DreamHOI [67] improves upon DreamFusion in terms of HOI but still falls short in capturing fine interaction details, such as accurate foot placement on the basketball. + +Inpainting Visualization. We further visualize the 2D inpainting results and affordance parsing on the same object in Figure 5 to emphasize the effectiveness of our method in disentangling affordances for distinct human-object interactions. The figure showcases various human object interactions with a chair, highlighting the ability to localize affordances accurately. This visualization provides valuable insight into the regions on the object that are most likely to be in contact with humans, which is critical for optimizing human-object interaction. + +![](images/e1f9b0c20729d9595390fa44bbfc5b41855d0a6473302a9a6dc90fd4ded389ed.jpg) + +![](images/b0e1f99492fd2af975c2db8b6d768dc55a4de7651c998e27b1dc76a8d73d453a.jpg) +Figure 6. Qualitative results of scene populating and scene interaction applications from different views. + +Qualitative performance on Scene Populating. We demonstrate one possible way to populate scenes with InteractAnything. We first utilize the HPS dataset [14] to obtain the 3D scene and then integrate the generated HOI results into the scene by placing them in appropriate locations. By employing a ground loss and large language model, we ensure that the HOI results are naturally embedded within the scene without artifact issues such as object or human interpenetration. Furthermore, our method goes beyond simply populating HOI results into the scene, we also generate realistic interaction effects with existing objects in the environment. For example, InteractAnything can effectively generate interactions such as "sitting on the table" or "opening a door", as demonstrated in the Figure 6. + +# 4.4. Ablation Study + +Ablation on Fine-grained Optimization In this study, we analyze the effect of fine-grained optimization terms as outlined in Section 3.5 while keeping all other parameters constant. The results shown in Figure 4, clearly demonstrate that the fine-grained optimization significantly enhances the realism and contact plausibility of human-object interactions (right green panel) in the "grasp" prompt. Without fine-grained optimization (left red panel), the alignment between the hand and table is noticeably inaccurate, resulting in unnatural poses where the hand appears to "float" near the table instead of properly grasping or holding it. In contrast, with fine-grained optimization enabled (right panel), the human hand is aligned with the table surface, resulting in a more realistic interaction. This improvement highlights the importance of fine-grained optimization in achieving natural contact and precise positioning, especially in actions involving detailed hand-object interactions. + +Ablation Study on LLM-guided Initialization We further ablate the LLM-guided initialization and the corresponding HOI refinement module (fine-grained HOI synthesis) to investigate the contribution of our design in leveraging language models. As shown in Table 3, we use GPT-4V selection as the evaluation metric. Compared to the full method (Full), the variant without refinement (w/o Refine) + +exhibits a noticeable decline in performance. This drop is primarily due to the absence of fine-grained contact optimization, which is critical for generating the most accurate human-object interactions. In contrast, when the LLM-guided initialization is ablated (w/o LLM-init), we randomly initialize both the human and object positions, followed by coarse HOI optimization. This approach results in a significant decrease in selection scores, ranking lower than DreamHOI [67]. This finding emphasizes the importance of an informed initialization strategy in establishing a solid foundation for HOI optimization. The stark differences in the ablation study underscore the effectiveness of our design. The integration of LLM-guided initialization and refinement is vital for achieving superior results. + +# 5. Discussion and Limitation + +In this work, we demonstrate that our approach can generate fine-grained, realistic HOIs for open-set 3D objects based on text descriptions. While our method shows promising results in improving contact precision and adapting to unseen objects, it still relies on 2D priors as its primary guidance. Consequently, advancements in multi-view consistent image generation or fully 4D generation techniques would likely enhance our approach, leading to more coherent and robust 3D interactions. Additionally, physical plausibility is also an important application in recent applications, e.g., in humanoid manipulation. We use force closure as a physical constraint while a high-quality simulator could further evaluate the generated HOI. Furthermore, our method currently uses the SMPL-H model to represent human body geometry and articulation. While this model captures a wide range of human poses and movements, it limits the adaptability of our framework to diverse interaction agents, particularly those with non-human anatomies. These agents often have unique skeletal structures and movement patterns that differ significantly from human biomechanics, which our current system is not fully equipped to process. In future work, we aim to address these limitations by investigating multi-view or 4D generation for enhanced consistency and exploring more flexible models to represent a broader array. + +# 6. Conclusion + +This paper tackles the challenge of generating novel 3D HOI involving unseen objects—a task where current methods often struggle. To address this issue, we propose InteractAnything, a novel zero-shot synthesis method driven by text instructions and not reliant on training with existing 3D datasets. It leverages guidance from the large language model to distill human-object relations and employs the pre-trained 2D diffusion model to parse contact areas with open-set objects, circumventing the limitations of existing HOI datasets. + +Acknowledgments. This work was supported by the National Science and Technology Major Project (2022ZD0114900) and NSFC-6247070125. We extend our gratitude to the reviewers for their insightful comments. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 7 +[2] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6 +[3] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep It SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. In Computer Vision - ECCV 2016, 2016. 5 +[4] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 958-968, 2024. 2 +[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5 +[6] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In International Conference on Computer Vision (ICCV), 2023. 2 +[7] Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets creation using spatially-aware diffusion guidance. In European Conference on Computer Vision, pages 128-146. Springer, 2024. 2, 3 +[8] Hongkai Dai, Anirudha Majumdar, and Russ Tedrake. Synthesis and optimization of force closure grasps via sequential semidefinite programming. Robotics Research: Volume 1, pages 285-305, 2018. 2 +[9] Sisi Dai, Wenhao Li, Haowen Sun, Haibin Huang, Chongyang Ma, Hui Huang, Kai Xu, and Ruizhen Hu. Interfusion: Text-driven generation of 3d human-object interaction. In European Conference on Computer Vision, pages 18-35. Springer, 2024. 2, 3, 6 +[10] Google DeepMind. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2 +[11] Clement Fuji Tsang, Maria Shugrina, Jean Francois Lafleche, Towaki Takikawa, Jiehan Wang, Charles Loop, Wenzheng Chen, Krishna Murthy Jatavallabhula, Edward Smith, Artem Rozantsev, Or Perel, Tianchang Shen, Jun Gao, Sanja Fidler, Gavriel State, Jason Gorski, Tommy Xiang, Jianing Li, Michael Li, and Rev Lebaredian. Kaolin: A pytorch library for accelerating 3d deep learning re + +search. https://github.com/NVIDIAGameWorks/kaolin, 2022.6 +[12] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Scholkopf. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[13] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13041-13051, 2023. 5 +[14] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human positioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2021. 8 +[15] Sookwan Han and Hanbyul Joo. Chorus: Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In International Conference on Computer Vision (ICCV), 2023. 3, 5 +[16] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambiguities with 3D scene constraints. In International Conference on Computer Vision (ICCV), 2019. 2, 3 +[17] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2 +[18] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 2 +[19] Jingwei Huang, Hao Su, and Leonidas Guibas. Robust watertight manifold surface generation method for shapenet models. arXiv preprint arXiv:1802.01698, 2018. 6 +[20] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Zhiyuan Zhang, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. Full-body articulated human-object interaction. In International Conference on Computer Vision (ICCV), 2023. 2 +[21] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[22] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7122-7131, 2018. 5 +[23] Tero Karras. Maximizing parallelism in the construction of bvhs, octrees, and k-d trees. In Proceedings of ACM SIGGRAPH/Eurographics Conference on High-Performance Graphics, 2012. 6 +[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):139-1, 2023. 2 +[25] Hyeonwoo Kim, Sookwan Han, Patrick Kwon, et al. Beyond the contact: Discovering comprehensive affordance for 3d + +objects from pre-trained 2d diffusion models. In European Conference on Computer Vision (ECCV), 2024. 4, 5 +[26] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2, 3 +[27] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. In European Conference on Computer Vision (ECCV), 2024. 2 +[28] Lei Li and Angela Dai. Genzi: Zero-shot 3d human-scene interaction generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20465-20474, 2024. 4 +[29] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7 +[30] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21159-21168, 2023. 5 +[31] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In International Conference on Computer Vision (ICCV), 2023. 2 +[32] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters (RA-L), 7(1):470-477, 2021. 2, 6 +[33] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splatting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[34] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In International Conference on Learning Representations (ICLR), 2023. 2 +[35] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Transactions on Graphics (TOG), 34(6):248:1-248:16, 2015. 3 +[37] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2 + +[38] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 6 +[39] OpenAI. Gpt-3.5. https://platform.openai.com/docs/models/gpt-3-5, 2023. Accessed: 2023-11-14. 2, 4 +[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 6 +[41] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 2 +[42] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4 +[43] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations (ICLR), 2023. 2, 3, 5, 6, 7 +[44] Alec Radford, JongWook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Askell Amanda, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. Cornell University - arXiv, Cornell University - arXiv, 2021. 6 +[45] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2 +[48] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: modeling and capturing hands and bodies together. ACM Transactions on Graphics (TOG), 36(6):1-17, 2017. 3 +[49] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[50] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, + +Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[51] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2 +[52] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 6 +[53] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 2 +[54] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J Black. Deco: Dense estimation of 3d human-scene contact in the wild. In International Conference on Computer Vision (ICCV), 2023. 2 +[55] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In International Conference on 3D Vision (3DV), 2022. 3 +[56] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2 +[57] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16210-16220, 2022. 2 +[58] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In International Conference on Computer Vision (ICCV), 2023. 2, 3 +[59] Sirui Xu, Yu-Xiong Wang, Liangyan Gui, et al. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. Advances in Neural Information Processing Systems, 37:52858-52890, 2024. +[60] Hongwei Yi, Justus Thies, Michael J Black, Xue Bin Peng, and Davis Rempe. Generating human interaction motions in scenes with text control. In European Conference on Computer Vision (ECCV), 2025. 2, 3 +[61] Wanqi Yin, Zhongang Cai, Ruisi Wang, Fanzhou Wang, Chen Wei, Haiyi Mei, Weiye Xiao, Zhitao Yang, Qingping Sun, Atsushi Yamashita, et al. Whac: World-grounded humans and cameras. In European Conference on Computer Vision, pages 20-37, 2024. 5 +[62] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monoc + +ular images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:12287-12303, 2023. +[63] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13232-13242, 2022. 5 +[64] Jason Y. Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In European Conference on Computer Vision (ECCV), 2020. 3, 6 +[65] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 2 +[66] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on pattern analysis and machine intelligence, 44(6): 3170-3184, 2021. 6 +[67] Thomas Hanwen Zhu, Ruining Li, and Tomas Jakab. Dreamhoi: Subject-driven generation of 3d human-object interactions with diffusion priors. arXiv preprint arXiv:2409.08278, 2024. 2, 3, 5, 6, 7, 8 \ No newline at end of file diff --git a/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/images.zip b/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e8121b3f4fd62e4560be6e5abcd4b68bb1d66509 --- /dev/null +++ b/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:697d5ad7a319d903a42f0d08974ddbbd780e8bcbd6858d2693169e1ba1d83ce9 +size 378208 diff --git a/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/layout.json b/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..df038bc0e511d5c44003f390816ac0bde57a704f --- /dev/null +++ b/2025/InteractAnything_ Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing/layout.json @@ -0,0 +1,9534 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 538, + 140 + ], + "type": "text", + "content": "InteractAnything: Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "spans": [ + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "type": "text", + "content": "Jinlu Zhang" + }, + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "type": "text", + "content": ", Yixin Chen" + }, + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "type": "inline_equation", + "content": "^{2\\boxtimes}" + }, + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "type": "text", + "content": ", Zan Wang" + }, + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "type": "text", + "content": ", Jie Yang" + }, + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "type": "text", + "content": ", Yizhou Wang" + }, + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "type": "inline_equation", + "content": "^{1,5,6,7\\boxtimes}" + }, + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "type": "text", + "content": ", Siyuan Huang" + }, + { + "bbox": [ + 74, + 160, + 536, + 176 + ], + "type": "inline_equation", + "content": "^{2\\boxtimes}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 242, + 177, + 368, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 177, + 368, + 189 + ], + "spans": [ + { + "bbox": [ + 242, + 177, + 368, + 189 + ], + "type": "text", + "content": "indicates corresponding authors" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 190, + 474, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 190, + 474, + 203 + ], + "spans": [ + { + "bbox": [ + 137, + 190, + 474, + 203 + ], + "type": "text", + "content": "1 Center on Frontiers of Computing Studies, School of Computer Science, Peking University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 122, + 204, + 487, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 204, + 487, + 217 + ], + "spans": [ + { + "bbox": [ + 122, + 204, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 204, + 487, + 217 + ], + "type": "text", + "content": " State Key Laboratory of General Artificial Intelligence, BIGAI " + }, + { + "bbox": [ + 122, + 204, + 487, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 122, + 204, + 487, + 217 + ], + "type": "text", + "content": " Beijing Institute of Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 218, + 531, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 218, + 531, + 231 + ], + "spans": [ + { + "bbox": [ + 78, + 218, + 531, + 231 + ], + "type": "text", + "content": "4 The Chinese University of Hong Kong, Shenzhen 5 Nat'l Eng. Research Center of Visual Technology, Peking University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 96, + 232, + 512, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 232, + 512, + 245 + ], + "spans": [ + { + "bbox": [ + 96, + 232, + 512, + 245 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 96, + 232, + 512, + 245 + ], + "type": "text", + "content": " Institute for AI, Peking University " + }, + { + "bbox": [ + 96, + 232, + 512, + 245 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 96, + 232, + 512, + 245 + ], + "type": "text", + "content": " State Key Laboratory of General Artificial Intelligence, Peking University" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 60, + 279, + 211, + 420 + ], + "blocks": [ + { + "bbox": [ + 76, + 269, + 190, + 277 + ], + "lines": [ + { + "bbox": [ + 76, + 269, + 190, + 277 + ], + "spans": [ + { + "bbox": [ + 76, + 269, + 190, + 277 + ], + "type": "text", + "content": "A person holds/pulls/lies on the chair..." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 279, + 211, + 420 + ], + "lines": [ + { + "bbox": [ + 60, + 279, + 211, + 420 + ], + "spans": [ + { + "bbox": [ + 60, + 279, + 211, + 420 + ], + "type": "image", + "image_path": "13c688d3961fa06fb80cf67c7892f1ba4d1ca5f354ac86ad64e501a827ca28a5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 421, + 181, + 429 + ], + "lines": [ + { + "bbox": [ + 88, + 421, + 181, + 429 + ], + "spans": [ + { + "bbox": [ + 88, + 421, + 181, + 429 + ], + "type": "text", + "content": "(a) Diverse human interactions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 439, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 54, + 439, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 439, + 555, + 472 + ], + "type": "text", + "content": "Figure 1. 3D human object interaction synthesis by InteractAnything. Given a simple text description with goal interaction and any object mesh as input, our method enables the generation of diverse, natural, detailed, and novel interactions for open-set 3D objects in a zero-shot manner. The orange and green boxes of (b) indicate detailed contact poses from different views." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 228, + 282, + 388, + 419 + ], + "blocks": [ + { + "bbox": [ + 279, + 266, + 338, + 274 + ], + "lines": [ + { + "bbox": [ + 279, + 266, + 338, + 274 + ], + "spans": [ + { + "bbox": [ + 279, + 266, + 338, + 274 + ], + "type": "text", + "content": "A person grasps the" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 268, + 275, + 349, + 282 + ], + "lines": [ + { + "bbox": [ + 268, + 275, + 349, + 282 + ], + "spans": [ + { + "bbox": [ + 268, + 275, + 349, + 282 + ], + "type": "text", + "content": "backpack/table/ball/knife..." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 282, + 388, + 419 + ], + "lines": [ + { + "bbox": [ + 228, + 282, + 388, + 419 + ], + "spans": [ + { + "bbox": [ + 228, + 282, + 388, + 419 + ], + "type": "image", + "image_path": "f882f26ba7af8d8816b8e73c4f291654525d23f4685d548ed3e15673a09505fc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 420, + 388, + 429 + ], + "lines": [ + { + "bbox": [ + 228, + 420, + 388, + 429 + ], + "spans": [ + { + "bbox": [ + 228, + 420, + 388, + 429 + ], + "type": "text", + "content": "(b) Detailed interactions synthesis of open-set objects." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 427, + 287, + 532, + 420 + ], + "blocks": [ + { + "bbox": [ + 432, + 270, + 528, + 277 + ], + "lines": [ + { + "bbox": [ + 432, + 270, + 528, + 277 + ], + "spans": [ + { + "bbox": [ + 432, + 270, + 528, + 277 + ], + "type": "text", + "content": "A person lifts/rides/criddles the" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 278, + 531, + 286 + ], + "lines": [ + { + "bbox": [ + 429, + 278, + 531, + 286 + ], + "spans": [ + { + "bbox": [ + 429, + 278, + 531, + 286 + ], + "type": "text", + "content": "tesla car/motorcycle/horse/baby..." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 287, + 532, + 420 + ], + "lines": [ + { + "bbox": [ + 427, + 287, + 532, + 420 + ], + "spans": [ + { + "bbox": [ + 427, + 287, + 532, + 420 + ], + "type": "image", + "image_path": "c2b3b4f1fd64f2cc3bd30e725e65a4451665dec03e58629b40a7f0a26da591ce.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 420, + 548, + 430 + ], + "lines": [ + { + "bbox": [ + 410, + 420, + 548, + 430 + ], + "spans": [ + { + "bbox": [ + 410, + 420, + 548, + 430 + ], + "type": "text", + "content": "(c) Novel interactions given generative objects." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 152, + 483, + 200, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 483, + 200, + 495 + ], + "spans": [ + { + "bbox": [ + 152, + 483, + 200, + 495 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 54, + 510, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 297, + 714 + ], + "type": "text", + "content": "Recent advances in 3D human-aware generation have made significant progress. However, existing methods still struggle with generating novel Human Object Interaction (HOI) from text, particularly for open-set objects. We identify three main challenges of this task: precise human-object relation reasoning, affordance parsing for any object, and detailed human interaction pose synthesis aligning description and object geometry. In this work, we propose a novel zero-shot 3D HOI generation framework without training on specific datasets, leveraging the knowledge from largescale pre-trained models. Specifically, the human-object relations are inferred from large language models (LLMs) to initialize object properties and guide the optimization process. Then we utilize a pre-trained 2D image diffusion model to parse unseen objects and extract contact points, avoiding the limitations imposed by existing 3D asset knowledge. The initial human pose is generated by sam" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 483, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 628 + ], + "type": "text", + "content": "pling multiple hypotheses through multi-view SDS based on the input text and object geometry. Finally, we introduce a detailed optimization to generate fine-grained, precise, and natural interaction, enforcing realistic 3D contact between the 3D object and the involved body parts, including hands in grasping. This is achieved by distilling human-level feedback from LLMs to capture detailed human-object relations from the text instruction. Extensive experiments validate the effectiveness of our approach compared to prior works, particularly in terms of the fine-grained nature of interactions and the ability to handle open-set 3D objects. Project page: jinluzhang.site/projects/interactanything." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 655, + 395, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 655, + 395, + 667 + ], + "spans": [ + { + "bbox": [ + 314, + 655, + 395, + 667 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Synthesizing the interaction of humans and any given object holds great significance across diverse applications, including AR/VR, computer simulation, and animation. Tradi-" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7015" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "tionally, creating high-quality 3D human object interaction (HOI) assets requires expensive, lab-controlled data capture systems and significant human effort. Although some datasets [2, 16, 20, 21, 26, 53] have collected these data, the models [26, 27, 41, 54, 58, 60] trained on them can only generate interactions for specific object categories. These methods struggle to handle the diversity of object geometries, scales, and interactions at the same time, making text-driven zero-shot HOI synthesis a challenging and highly demanding task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 194, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 397 + ], + "type": "text", + "content": "Given recent advances in text-to-3D techniques [6, 29, 56], e.g., leveraging Score Distillation Sampling (SDS) [43] to generate 3D content from text, a natural question is whether this paradigm can be applied to zero-shot HOI synthesis. However, directly integrating 3D humans and objects into a scene without understanding the complex relationships between them remains challenging. Recent works [7, 9, 67] are proposed to generate realistic 3D HOIs. Among them, InterFusion [9] uses Neural Radiance Fields (NeRF) [37] for individual human and object representations but struggles with precise interaction area, leading to suboptimal and unrealistic HOIs. DreamHOI [67] optimizes on both mesh and NeRF representations through bidirectional rendering but faces challenges in ensuring accurate contact points and relative scales due to solely optimizing human states while keeping the given object static, leading to limited interactions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 400, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 590 + ], + "type": "text", + "content": "In this paper, we first identify the core challenge of zero-shot 3D HOI generation as accurately capturing human-object relations, contact areas, and global consistency, which are not typically presented in single-object or non-contact generation scenarios. These challenges prevent the generation of high-quality, geometrically precise, and textually consistent results. Furthermore, while large-scale pre-trained models, e.g., 2D diffusion models [42, 47] and large language models (LLMs) [1, 10, 39], cannot fully capture the complexity of human object interactions (including body dynamics, expressive details, and object geometry), they can offer valuable prior knowledge of common interaction patterns based on textual descriptions. Consequently, we argue that the solution for effective generation is to leverage this diverse prior knowledge to guide the optimization of interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Following the above insights, we decompose the complex zero-shot 3D HOI task into these stages: 1) human-object relation reasoning and initialization, 2) open-set object contact affordance parsing, and 3) human pose synthesis aligning text and object geometry. We first leverage the LLM as a human-level feedback provider, which reasons and unravels complex semantic interaction relationships, e.g., relative position and orientation, from simple text input. It initializes human and object states and designs a series of optimization functions to generate detailed inter" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "actions. The initial human pose is optimized by multi-view SDS loss for interaction in line with the given textual description and object geometry." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 109, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 228 + ], + "type": "text", + "content": "To parse the 3D object affordance, we utilize initialized human-object relation and sample adaptive inpainting masks during the diffusion process to obtain the contact information. This enables the extraction of contact probabilities on the object, which are then used to segment and define the weighted dynamic contact regions. Finally, we design a coarse-to-fine optimization based on force closure [8, 32] to generate expressive, diverse, and physically plausible interactions, allowing for more accurate and detailed contact. To summarize, our contributions are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 230, + 553, + 397 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 230, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 230, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 230, + 553, + 277 + ], + "type": "text", + "content": "- We propose InteractAnything for zero-shot 3D HOI generation, which effectively distills human and object interactions from LLM and pre-trained diffusion models to synthesize diverse, detailed, and novel interactions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 278, + 553, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 553, + 313 + ], + "type": "text", + "content": "- We design a novel open-set object affordance parsing, which generates contact maps on 3D geometry by utilizing adaptive inpainting and potential contact distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 314, + 553, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 397 + ], + "type": "text", + "content": "- Our model leverages both text and object geometry to drive human pose synthesis, enabling realistic and contextually appropriate interactions for open-set objects. Results show that our synthesized pose accurately reflects the semantic meaning of the input while respecting the spatial constraints of the object, leading to improved performance, as shown in Figure 1." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 411, + 400, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 411, + 400, + 423 + ], + "spans": [ + { + "bbox": [ + 314, + 411, + 400, + 423 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 431, + 503, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 431, + 503, + 443 + ], + "spans": [ + { + "bbox": [ + 314, + 431, + 503, + 443 + ], + "type": "text", + "content": "2.1. Text-driven 3D Content Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "The generation of 3D content from text has seen rapid advancements, primarily driven by diffusion-based models that build on text-to-2D image generation approaches [45, 46, 49, 50]. One pioneering approach, DreamFusion [43], introduced Score Distillation Sampling (SDS), which leverages the pre-trained 2D diffusion model to optimize 3D content through techniques like NeRF (Neural Radiance Fields) [37] and 3D Gaussian Splatting [24]. A significant area of research aims to address the complexity of 3D human generation. For example, recent methods have been proposed for generating controllable 3D human avatar [4, 17, 18, 33, 57, 65]. However, these works focus on diverse and high-quality appearance but animate human avatars with given human poses or motion, which do not satisfy 3D HOI requirements in human pose synthesis. The proliferation of large 3D datasets has also catalyzed progress, leading to more sophisticated models like Zero-1-to-3 [31], Zero " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "123++" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": " [51], and MVDream [52], which employ 2D diffusion models to generate multi-view consistent images that serve as inputs for efficient 3D generation systems, such as SyncDreamer [34] and Wonder3D [35]. Despite these advancements, generating complex HOI scenes," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7016" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "especially those involving interactions between humans and objects, remains challenging." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 113, + 264, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 113, + 264, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 113, + 264, + 126 + ], + "type": "text", + "content": "2.2. Data-Driven Human Object Interaction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 312 + ], + "type": "text", + "content": "Previous approaches [26, 58-60] primarily train generative models to produce static human poses or dynamic motions in response to objects or scenes using curated HOI datasets. Recent works focus on dynamic motion generation performance conditioned on encoded object features but rely heavily on specialized datasets, limiting generalization to diverse objects. For example, OMOMO [26] generates human motion through diffusion models but requires object motion as input. InterDiff [58] predicts full-body interactions based on prior HOI context but is limited to the prediction task, restricting flexibility in new scenarios. In contrast, our method jointly synthesizes human and object states and adapts to novel objects without extensive, object-specific training by extracting interaction patterns directly from 2D diffusion models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 472 + ], + "type": "text", + "content": "On the other hand, optimization-based methods reconstruct realistic human poses and transformations from images. PHOSA [64] optimizes human object relations using depth-aware loss functions. PROX [16] encodes contact probabilities and semantic scene labels to position humans within scenes. CHORUS [15] synthesizes multi-view images with a pre-trained generative model to learn 3D spatial relationships. By extracting knowledge from LLM to infer human poses, the method in [55] queries contact information to reconstruct 3D human-object interactions. Compared to these methods, our method leverages pre-trained diffusion models for accurate parsing of object affordance and enables fine-grained interactions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 489, + 266, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 266, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 266, + 502 + ], + "type": "text", + "content": "2.3. Zero-shot 3D Compositional Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "To achieve the compositional scene generation, Combo-Verse [7] proposes a spatially-aware SDS loss, generating the final 3D scene through single-object reconstruction and multi-object composition. GraphDreamer [12] leverages scene graphs to guide generation by converting the semantic information into a global textual description to enable text-driven compositional scene generation. These methods focus on the reasonable arrangement of multiple objects but ignore the contact generation in HOI scenes. InterFusion [9] and DreamHOI [67] present this challenge and then try to address using SMPLs [36, 48] as body model prior. DreamHOI [67] designs the SDS guidance mixture to separately render human interactions using different diffusion models to improve human modeling quality. However, it cannot ensure precise contact during the denoising process. Compared to these methods, our work focuses on addressing contact areas parsing on open-set objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 71, + 394, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 394, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 394, + 84 + ], + "type": "text", + "content": "3. Our Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": "Task Definition and Overview. Given a text prompt " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " that describes the desired HOI, e.g., \"A person grasps the chair\". Our goal is to synthesize natural human mesh " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\nu_{H}" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " and object mesh " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\nu_{O}" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " that align with " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": ". As shown in Figure 2, we first obtain initial human-object relations from LLM based on text (Section 3.2). For object optimization, we initialize object rotation " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "r_o" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": ", translation " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "t_o" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "s_o" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": ", state " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": ", based on affordance map " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " (Section 3.3). We also synthesize the initial human pose based on the input text and object geometry (Section 3.4). Finally, we use SMPL-H as representation and optimize pose parameter " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\theta_b, \\theta_h" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": ", global translation " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "t_b" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": ", and object rotation and translation " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "r_o, t_o" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " to generate detailed and diverse HOI (Section 3.5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 253, + 449, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 449, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 449, + 266 + ], + "type": "text", + "content": "3.1. Preliminary Knowledge" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "text", + "content": "Human Body Model. We use the SMPL-H [36, 48], a parametric model that includes both body and hand parameters, to represent the human body and hand articulation. The body pose is defined by body parameters " + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{b}} \\in \\mathbb{R}^{3(K_b + 1)}" + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "text", + "content": " and the hand pose is represented by additional MANO parameters " + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{h}} \\in \\mathbb{R}^{3K_h}" + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "text", + "content": ", capturing 15 joints per hand. The shape parameters are represented by " + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "text", + "content": ", controlling the deformation of the body and hand template mesh to target body shapes. Given the combined pose parameters " + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\{\\theta_{\\mathrm{b}}, \\theta_{\\mathrm{h}}\\}" + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "text", + "content": " and shape parameters " + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "text", + "content": ", SMPL-H computes mesh vertices " + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\in \\mathbb{R}^{V \\times 3}" + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "text", + "content": " using linear blend skinning function " + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 342, + 407, + 553, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 407, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 342, + 407, + 553, + 421 + ], + "type": "interline_equation", + "content": "\\mathcal {V} (\\beta , \\theta_ {b}, \\theta_ {\\mathrm {h}}) = f \\left(T _ {p} \\left(\\beta , \\theta_ {b}, \\theta_ {\\mathrm {h}}\\right), J (\\beta), \\theta , \\mathcal {W}\\right), \\tag {1}", + "image_path": "2869bfd755fd756ee7011eb1d0f7e480528bd891bd4a77c7e47c4fb681c198e9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 425, + 553, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 553, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 425, + 553, + 450 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 425, + 553, + 450 + ], + "type": "text", + "content": " represents joint locations, " + }, + { + "bbox": [ + 313, + 425, + 553, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{W} \\in \\mathbb{R}^{V \\times (K)}" + }, + { + "bbox": [ + 313, + 425, + 553, + 450 + ], + "type": "text", + "content": " is skimming weights with all " + }, + { + "bbox": [ + 313, + 425, + 553, + 450 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 425, + 553, + 450 + ], + "type": "text", + "content": " joints." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "text", + "content": "Score Distillation Sampling (SDS). SDS leverages pretrained diffusion models to guide the synthesis of 3D scenes by aligning NeRF with a text prompt. It operates by iteratively adjusting NeRF parameters, denoted as " + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "text", + "content": ", based on gradients computed to align the generated images with the textual input. Given a noisy 2D image " + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "text", + "content": " (obtained by adding noise " + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "text", + "content": " to a clean NeRF-rendered image " + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "text", + "content": "), a textual prompt " + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "text", + "content": ", and a noise level " + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "text", + "content": ", SDS calculates the gradient of the loss " + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SDS}}" + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 450, + 555, + 559 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 336, + 563, + 531, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 563, + 531, + 590 + ], + "spans": [ + { + "bbox": [ + 336, + 563, + 531, + 590 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (x) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) (\\hat {\\epsilon} (x _ {t}; y, t) - \\epsilon) \\frac {\\partial x}{\\partial \\theta} \\right],", + "image_path": "83ce11b4ce76ca03f1ec5b3834b7768c976a97ac10e8eb9a0770748332b0f2f6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}(x_t; y, t)" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " represents the diffusion model predicted noise given the noisy image " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", prompt " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", and noise level " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " is a weighting function that adjusts the influence of the gradient based on the noise level " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ". This gradient is used to update " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", guiding the NeRF model toward producing images that match the textual description. For instance, in DreamFusion [43], this process enables text-to-3D generation by repeatedly refining the NeRF parameters " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " from a random initialization, iteratively producing images that align more closely with the target prompt." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7017" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 72, + 541, + 297 + ], + "blocks": [ + { + "bbox": [ + 69, + 72, + 541, + 297 + ], + "lines": [ + { + "bbox": [ + 69, + 72, + 541, + 297 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 541, + 297 + ], + "type": "image", + "image_path": "a0f0efb2da3f61d82118f3a4d802002ff84417a895f28f2e1bf39c6eff39f3d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 308, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 54, + 308, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 555, + 365 + ], + "type": "text", + "content": "Figure 2. Framework of InteractAnything. Given a text description and any object mesh as input, our approach begins by querying LLM to infer precise human-object relationships, which are used to initialize object properties. Next, we analyze the contact affordance of the object geometry. The human pose is synthesized using a pre-trained 2D diffusion model, guided by multi-view SDS loss and the designed spatial constraint. Finally, based on the targeted object contact areas and a plausible human pose, we perform expressive HOI optimization to synthesize realistic and contact-accurate 3D human-object interactions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 384, + 201, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 201, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 201, + 396 + ], + "type": "text", + "content": "3.2. LLM-guided Initialization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": "To infer diverse intended human object interactions for a given object " + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_O" + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": " and action " + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": ", we query the LLM to provide commonsense feedback. To minimize hallucination, we need to transfer inferred relationships from GPT [39] into specific values that position the human and object meshes accurately, therefore, we enable the language model to select answers from pre-defined options step by step, instead of directly generating responses. For initializing the object rotation " + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "inline_equation", + "content": "r_o" + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": ", translation " + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "inline_equation", + "content": "t_o" + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "inline_equation", + "content": "s_o" + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": ", and state " + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": ", the language model selects the relative position of the human to the object from defined options, which is then mapped to " + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "inline_equation", + "content": "t_o" + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": " based on the combined size of the human and object. We set all input objects to have " + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": "-up and " + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": "-front orientations, with " + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "inline_equation", + "content": "r_o" + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": " adjusted based on which side of the object should face the human. The object scale is chosen from 4 defined options and set relative to human size. Lastly, the state " + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": " determines if the object should be positioned on the ground based on whether it is meant to move. Besides, we also query interacted semantic body parts " + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "inline_equation", + "content": "S_h" + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": " and map the labels to human mesh to obtain vertices " + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_h" + }, + { + "bbox": [ + 56, + 401, + 296, + 653 + ], + "type": "text", + "content": " for part-level optimization using annotated human part labels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 660, + 249, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 249, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 249, + 673 + ], + "type": "text", + "content": "3.3. Open-set Object Affordance Parsing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "To obtain the contact affordance of unseen object geometries given a specified action e.g., holds, we conduct 2D inpainting based on a pre-trained diffusion model to dis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 384, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 443 + ], + "type": "text", + "content": "till possible human interaction parts through LLM-guided adaptive mask generation. Then we design a 2D probabilistic contact function to model the distribution of the interaction, generating an affordance map to identify and extract the regions for HOI optimization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 444, + 556, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 556, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 556, + 659 + ], + "type": "text", + "content": "2D Adaptive Mask Generation Image inpainting is to fill in missing regions of an image, where the missing regions can be defined by a mask image. To get knowledge of possible interacted parts, we utilize the powerful understanding and generation ability of pre-trained inpainting model [42]. Given 2D projection images of the 3D object from " + }, + { + "bbox": [ + 313, + 444, + 556, + 659 + ], + "type": "inline_equation", + "content": "n_p" + }, + { + "bbox": [ + 313, + 444, + 556, + 659 + ], + "type": "text", + "content": " views, we found existing inpainting models cannot ensure complete human in the image (many inpainting results only include interacted body parts, e.g., hands in \"hold\" action). Therefore, different from previous methods [25, 28], we do not randomly sample the inpainting masks on 2D images, instead, our method identifies possible human interaction parts to improve the inpainting success rate by generating inpainting masks from LLM feedback in two forms: (1) full-body mask with respect to object properties " + }, + { + "bbox": [ + 313, + 444, + 556, + 659 + ], + "type": "inline_equation", + "content": "r_o" + }, + { + "bbox": [ + 313, + 444, + 556, + 659 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 444, + 556, + 659 + ], + "type": "inline_equation", + "content": "t_o" + }, + { + "bbox": [ + 313, + 444, + 556, + 659 + ], + "type": "text", + "content": ", person position and person body size; (2) body-part inpainting masks for intended interacted body part labels " + }, + { + "bbox": [ + 313, + 444, + 556, + 659 + ], + "type": "inline_equation", + "content": "S_h" + }, + { + "bbox": [ + 313, + 444, + 556, + 659 + ], + "type": "text", + "content": " and corresponding human vertices " + }, + { + "bbox": [ + 313, + 444, + 556, + 659 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_h" + }, + { + "bbox": [ + 313, + 444, + 556, + 659 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 659, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 555, + 696 + ], + "type": "text", + "content": "More specifically, we first create a full-body inpainting mask guided by LLM-inferred initializations " + }, + { + "bbox": [ + 313, + 659, + 555, + 696 + ], + "type": "inline_equation", + "content": "r_o" + }, + { + "bbox": [ + 313, + 659, + 555, + 696 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 659, + 555, + 696 + ], + "type": "inline_equation", + "content": "t_o" + }, + { + "bbox": [ + 313, + 659, + 555, + 696 + ], + "type": "text", + "content": ", and body size " + }, + { + "bbox": [ + 313, + 659, + 555, + 696 + ], + "type": "inline_equation", + "content": "s_f" + }, + { + "bbox": [ + 313, + 659, + 555, + 696 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 345, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 345, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "M _ {\\mathrm {f u l l}} ^ {i} = \\mathcal {J} ((\\mathcal {V} _ {H} \\cdot r _ {o} ^ {- 1}), s _ {f}, c ^ {i}) + t _ {o} ^ {- 1} + \\varpi_ {1}, \\qquad (2)", + "image_path": "97260dc7d3e46cd92319cdad6c50bf4585415b268977a5a2b61c59e8f8ebc055.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7018" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 70, + 532, + 346 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 532, + 346 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 532, + 346 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 532, + 346 + ], + "type": "image", + "image_path": "2a23e3e34d611ce29e880c0b407320d93419734c546e3d75ec0b79aafbd51ae1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 356, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 54, + 356, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 555, + 388 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison results with baselines. * indicates we re-implement this method by embedding object mesh into the diffusion process, which follows the pipeline of DreamHOI [67] and our method. More visualization results are presented in supplementary materials." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "text", + "content": " is 2D projection transformation with camera parameter " + }, + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "inline_equation", + "content": "c^i" + }, + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "inline_equation", + "content": "\\varpi_{1}" + }, + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "text", + "content": " is the noise term. Similarly, we generate body-part-level inpainting masks using given body part labels " + }, + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "inline_equation", + "content": "S_{h}" + }, + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "text", + "content": " and corresponding human vertices " + }, + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_h" + }, + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "text", + "content": " and body part size " + }, + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "inline_equation", + "content": "s_p" + }, + { + "bbox": [ + 54, + 411, + 294, + 481 + ], + "type": "text", + "content": " to help inpainting body parts like hands as the above formulation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 488, + 294, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 488, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 87, + 488, + 294, + 504 + ], + "type": "interline_equation", + "content": "M _ {\\text {p a r t}} ^ {i} = \\mathcal {J} \\left(\\left(\\mathcal {P} _ {h} \\cdot r _ {o} ^ {- 1}\\right), s _ {p}, c ^ {i}\\right) + t _ {o} ^ {- 1} + \\varpi_ {2}. \\tag {3}", + "image_path": "3cdbb26a1424a7c6b761a8f1949e3b022b657a641e911dc9a5d024246a57cc40.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 510, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 557 + ], + "type": "text", + "content": "Besides, We apply the above full-body masks " + }, + { + "bbox": [ + 54, + 510, + 295, + 557 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{full}}" + }, + { + "bbox": [ + 54, + 510, + 295, + 557 + ], + "type": "text", + "content": " into the dynamic mask inpainting pipeline [25], which can further refine the initial mask detecting human on intermediate results in the diffusion process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "content": "Affordance Parsing on 2D Probability Function. Directly computing contact probabilities on a 3D object mesh relies heavily on existing 3D human reconstruction methods [3, 13, 22, 30, 61-63] from single-view images. This encounters significant 3D ambiguity and quantization error, which obscures accurate 3D human pose. In contrast to previous methods [15], we apply 2D human pose detector OpenPose [5], performing well at detecting 2D human keypoints and even occluded body parts, to estimate human from inpainting results. Then we introduce a 2D probability function that calculates contact probability according to the distance from detected body keypoints to the multi-view object projection images. First, we downsample the object" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 411, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 447 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 313, + 411, + 555, + 447 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 411, + 555, + 447 + ], + "type": "text", + "content": " points. Let " + }, + { + "bbox": [ + 313, + 411, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 411, + 555, + 447 + ], + "type": "text", + "content": " represent the probability map for each view " + }, + { + "bbox": [ + 313, + 411, + 555, + 447 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 411, + 555, + 447 + ], + "type": "text", + "content": ", the distance-based probability " + }, + { + "bbox": [ + 313, + 411, + 555, + 447 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{afford}}(\\mathbf{p})" + }, + { + "bbox": [ + 313, + 411, + 555, + 447 + ], + "type": "text", + "content": " for a point " + }, + { + "bbox": [ + 313, + 411, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathcal{V}_O" + }, + { + "bbox": [ + 313, + 411, + 555, + 447 + ], + "type": "text", + "content": " on the object surface is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 387, + 453, + 553, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 453, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 387, + 453, + 553, + 469 + ], + "type": "interline_equation", + "content": "f _ {\\text {a f f o r d}} ^ {(i)} (\\mathbf {p}) = e ^ {- \\left\\| \\mathbf {d} _ {i} (\\mathbf {p}) \\right\\|}, \\tag {4}", + "image_path": "1b5373fa6a6d42114ec2c1b1da8715ac65af3275ef354828adc24e3f4fbf269a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 475, + 555, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 475, + 555, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i(\\mathbf{p})" + }, + { + "bbox": [ + 313, + 475, + 555, + 537 + ], + "type": "text", + "content": " denotes the distance from the body part to the object mask in " + }, + { + "bbox": [ + 313, + 475, + 555, + 537 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 475, + 555, + 537 + ], + "type": "text", + "content": " view. Then we re-project the 2D affordance " + }, + { + "bbox": [ + 313, + 475, + 555, + 537 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{afford}}^{(i)}(\\mathbf{p})" + }, + { + "bbox": [ + 313, + 475, + 555, + 537 + ], + "type": "text", + "content": " to 3D space using defined camera parameters, and aggregate the contact probabilities across all " + }, + { + "bbox": [ + 313, + 475, + 555, + 537 + ], + "type": "inline_equation", + "content": "n_p" + }, + { + "bbox": [ + 313, + 475, + 555, + 537 + ], + "type": "text", + "content": " views to update the 3D contact probability distribution as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 379, + 543, + 553, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 543, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 379, + 543, + 553, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {P} (\\mathbf {p}) = \\frac {1}{n _ {p}} \\sum_ {i = 1} ^ {n _ {p}} f _ {\\text {a f f o r d}} ^ {(i)} (\\mathbf {p}). \\tag {5}", + "image_path": "d4df33c2bcd6e15395ab522745b4f19915d2403158dea6ba0527ca8bdd6df4e7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "type": "text", + "content": "Finally, we normalize " + }, + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathbf{p})" + }, + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "type": "text", + "content": " over the entire object surface to obtain contact probability for each point is determined by its projected 2D distances across multiple views, effectively mitigating multi-view ambiguity and enabling accurate estimation of human object contact regions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 647, + 537, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 537, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 537, + 661 + ], + "type": "text", + "content": "3.4. Text-Object Driven Human Pose Synthesis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Considering a brief HOI description, we posit that current 3D generation models, such as DreamFusion [43], are capable of generating human avatars in a variety of poses; however, they do not guarantee precise contact with the object" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7019" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": "mesh. Thus, we first apply prompt engineering to generate diverse and expressive text descriptions, along with negative prompts, using large language models (LLMs) to mitigate issues like low-quality outputs and missing limbs. Next, we incorporate Multi-view SDS loss [52] to optimize the human pose aligning the interaction description and object geometry. We first scale the input object based on " + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "inline_equation", + "content": "s_o" + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": " given by the language model. We use the spatial constraint to prevent the human avatar from penetrating the object mesh, ensuring alignment with spatial relations. Specifically, we embed object mesh within NeRF optimization by regularizing the ray sample points in object mesh following previous works [9, 67]. After training, we use the multi-view human pose estimation method [66] to obtain the body parameters " + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\theta_b" + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": " and then init hand parameters " + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\theta_h" + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": ". We also employ VPoser [40] to maintain reasonable joint rotations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 270, + 219, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 219, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 219, + 284 + ], + "type": "text", + "content": "3.5. Expressive HOI Optimization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 287, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 337 + ], + "type": "text", + "content": "Global HOI optimization. With the above initialization as input, we first ensure the input object is watertight using Manifold [19]. Then we conduct human-centric optimization to update object properties " + }, + { + "bbox": [ + 55, + 287, + 296, + 337 + ], + "type": "inline_equation", + "content": "(r_o, t_o, s_o)" + }, + { + "bbox": [ + 55, + 287, + 296, + 337 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 342, + 295, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 342, + 295, + 359 + ], + "spans": [ + { + "bbox": [ + 61, + 342, + 295, + 359 + ], + "type": "interline_equation", + "content": "L _ {c} = \\phi_ {i} L _ {i n t e r} + \\phi_ {n} L _ {n} + \\phi_ {s} L _ {s c a l e} + \\phi_ {p} L _ {p e n e} + \\mathbf {s} ^ {'} \\cdot L _ {g}, (6)", + "image_path": "c16e99f713f79b2bc997c7a8464880e2b1a8d5de5b6944eb130616fe71656130.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 365, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 365, + 296, + 401 + ], + "type": "text", + "content": " is loss weights and interaction loss is designed on. \nverts of the interacted body part " + }, + { + "bbox": [ + 55, + 365, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_h,h\\in \\mathcal{V}_H" + }, + { + "bbox": [ + 55, + 365, + 296, + 401 + ], + "type": "text", + "content": " and object \ncontact areas " + }, + { + "bbox": [ + 55, + 365, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_o,o\\in \\mathcal{M}" + }, + { + "bbox": [ + 55, + 365, + 296, + 401 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 407, + 295, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 407, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 85, + 407, + 295, + 435 + ], + "type": "interline_equation", + "content": "L _ {i n t e r} = \\sum_ {i \\in [ h ], j \\in [ o ]} W _ {\\mathcal {M}} (j) \\cdot f _ {c h a m} \\left(\\mathcal {P} _ {h}, \\mathcal {P} _ {o}\\right), \\tag {7}", + "image_path": "cb1f48e739582fe01ac660c3de6458a365664181fc0b0a8c303aa46954c19b95.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "inline_equation", + "content": "W_{\\mathcal{M}}(j)" + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "content": " indicates probabilities scores in object affordance map " + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "content": "-th object vertex and " + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "inline_equation", + "content": "f_{cham}" + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "content": " is chamfer distance function. We constrain the normals of interacted parts to maintain the relative orientation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 117, + 496, + 235, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 496, + 235, + 524 + ], + "spans": [ + { + "bbox": [ + 117, + 496, + 235, + 524 + ], + "type": "interline_equation", + "content": "L _ {n} = \\sum_ {n \\in [ h, o ]} \\left(\\mathbf {n} _ {h} \\cdot \\mathbf {n} _ {o} + 1\\right) ^ {2},", + "image_path": "9afc1cfd309c4d4c87f42b66be143d5a8093eef1c82a094a50b4966dc7a945bd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_h" + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": " is normals of interacted body vert, and " + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_o" + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": " indicates object normals of interacted vert. In addition, we employ the object scale loss " + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{scale}}" + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": " [64] and the penetration term " + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{pene}}" + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": " to prevent collisions [23]. The introduction of ground loss " + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "inline_equation", + "content": "L_{g}" + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": " is decided by the object state " + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{s} = [0,1]" + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 590, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 661 + ], + "type": "text", + "content": "Finer HOI Optimization. During this stage, we optimize the human parameters " + }, + { + "bbox": [ + 55, + 590, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\theta_{b},\\theta_{h}" + }, + { + "bbox": [ + 55, + 590, + 296, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 590, + 296, + 661 + ], + "type": "inline_equation", + "content": "t_b" + }, + { + "bbox": [ + 55, + 590, + 296, + 661 + ], + "type": "text", + "content": " and fix the object parameters. To synthesize detailed and diverse interaction poses for full-body optimization, inspired by differentiable force closure [32], we design the fine-grained optimization loss as below:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 109, + 669, + 295, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 669, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 109, + 669, + 295, + 683 + ], + "type": "interline_equation", + "content": "L _ {f} = \\delta \\cdot L _ {f c} + L _ {p e n e} + L _ {i n t e r}, \\tag {8}", + "image_path": "f5aebdcfbfa535cc36fe787361c3a12c6a2c02b3f10b6585f79b09e139d5b1af.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where loss functions " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{pene}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{inter}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " are analogous to the aforementioned terms, but with the incorporation of" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 554, + 99 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 554, + 99 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 554, + 99 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 554, + 99 + ], + "type": "table", + "html": "
MetricMagic3D [29]DreamFusion [43]DreamHOI [67]Ours
CLIP Score↑0.24160.24370.27570.2968
", + "image_path": "dbb228291e07176b850c9f1c06201e6506b7c872492f3539f4e9e1d00247c3b0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 107, + 555, + 129 + ], + "lines": [ + { + "bbox": [ + 313, + 107, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 107, + 555, + 129 + ], + "type": "text", + "content": "Table 1. Comparison of CLIP similarity scores for different methods. Higher scores are better." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "text", + "content": "Kaolin [11] to better sample surface distance. We introduce SMPL-based force closure loss " + }, + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "inline_equation", + "content": "L_{fc}" + }, + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "text", + "content": " by replacing the base model from MANO to SMPL-H to conduct more detailed interaction poses:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 329, + 217, + 555, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 217, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 329, + 217, + 555, + 259 + ], + "type": "interline_equation", + "content": "L _ {f c} = \\sum_ {j \\in [ o ]} \\left(\\sum_ {i \\in [ h ]} f v (i, j) \\cdot n (j)\\right) ^ {2}, \\delta \\cdot \\left[ \\theta_ {b}, \\theta_ {h} \\right]. \\tag {9}", + "image_path": "1a88427c92df7d7e448704e7f28bcc374402331b5880661d2a0a6e7e416da265.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "inline_equation", + "content": "fv" + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "text", + "content": " symbolizes the force vector applied at vertex " + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "text", + "content": " from interacted human vertex " + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "text", + "content": " decide optimize body pose " + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\theta_{b}" + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "text", + "content": " or hand pose " + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\theta_{h}" + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "text", + "content": ". Specifically, we make a series of augmentations to improve force closure acceptance rate and robustness. We decrease contact points by sparse uniform sampling on the interacted vertices. Moreover, we initialize body and hand positions around the interacted object parts and maintain the hand normals as positive to the interacted object parts. Finally, we extract the interacted vertices to form interacted mesh and non-interacted mesh and compute " + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "inline_equation", + "content": "L_{fc}" + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "inline_equation", + "content": "L_{pene}" + }, + { + "bbox": [ + 313, + 270, + 555, + 403 + ], + "type": "text", + "content": " separately." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 420, + 397, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 397, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 397, + 434 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 442, + 419, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 419, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 419, + 455 + ], + "type": "text", + "content": "4.1. Data and Metrics" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": "We collect evaluated object meshes in 3 ways: BEHAV [2], PartNet [38], and sampling from generative models. BEHAV dataset is commonly used in this field, and PartNet is manually annotated in an object mesh with part-level semantic labels on the vertices. For metrics, we use CLIP similarity score [44] between our prompts and the rendering images of the HOI scene, by rendering our human and object meshes from 4 different views randomly on 5 prompts and 10 objects categories. During this process, we select 5 BEHAV objects (backpack, basketball, chairwood, keyboard, and suitcase), and 5 objects from publicly accessed 3D generative models (motorcycle, car, television, baby doll, and humanoid robot). The same evaluation setting is adapted to other comparison methods. In addition, while the CLIP score is designed to evaluate how well an image aligns with the input text, it often fails to capture finer nuances in interactions, leading to less distinct variations in the metrics. We follow InterFusion [9] to leverage the advanced image understanding capabilities of GPT-4V further to assess the performance of baselines and our method. The text prompts are designed and then augmented using GPT." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7020" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 289, + 205 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 289, + 205 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 289, + 205 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 289, + 205 + ], + "type": "image", + "image_path": "56240666b57051852353bc0d4ff7db70a960212d273042d2126c7e0deea1cda4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 212, + 295, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 246 + ], + "type": "text", + "content": "Figure 4. Ablation study on the fine-grained optimization. Left figures are the results of removing fine-grained terms and right figures apple fine-grained terms to synthesize grasping details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 257, + 295, + 297 + ], + "blocks": [ + { + "bbox": [ + 56, + 257, + 295, + 297 + ], + "lines": [ + { + "bbox": [ + 56, + 257, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 295, + 297 + ], + "type": "table", + "html": "
MetricsMagic3DDreamFusionDreamFusion*DreamHOIOurs
Overall4.36.517.326.045.6
Contact8.64.315.219.552.1
", + "image_path": "7c01417ce23c3e07d047f18a2b095c34cf3ca4201d994c65e083cc10cf820756.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 195, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 195, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 195, + 361 + ], + "type": "text", + "content": "4.2. Quantitative Evaluations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 365, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 509 + ], + "type": "text", + "content": "CLIP Score. As shown in Table 1, we present a comparison of CLIP similarity scores across various methods, where higher scores indicate better alignment with the target descriptions. Our method InteractAnything, achieves the highest mean score across the evaluation prompts, outperforming baselines such as DreamFusion [43], Magic3D [29], and DreamHOI [67]. These results demonstrate that our approach effectively captures the semantic interactions of the text descriptions. Notably, to better evaluate interaction performance without being confounded by appearance, we apply random human textures to the body mesh and retain the original textures for objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 510, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 640 + ], + "type": "text", + "content": "GPT-4V Selection. As presented in Table 2, we enable the vision language model [1] to choose the most reasonable result in terms of Overall and Contact from all generated outputs based on criteria such as the presence of a full human body, a complete object, and correct physical interactions, and then returning the index of the selected result. No in-context examples are provided to guide the model in the Overall setting, we give text prompts only for the Contact evaluation, and the order of generated results is randomly shuffled to avoid bias, the final results are obtained from multiple tests." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 648, + 266, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 266, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 266, + 661 + ], + "type": "text", + "content": "4.3. Qualitative Evaluations and Application" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Visualization Comparison. The comparison results are shown in Figure 3, where our method outperforms the baselines, producing the most realistic and plausible 3D human-object interactions. This is evident in both the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 547, + 223 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 547, + 223 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 547, + 223 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 547, + 223 + ], + "type": "image", + "image_path": "96482e8b28fecbdc0483ee18124a6abefd2699cf01c84b665d83290d9a9b2708.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 232, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 289 + ], + "type": "text", + "content": "Figure 5. Visualization of the adaptive inpainting and open-set object affordance parsing results on the same object with different text instructions. Our method first generates reasonable 2D inpainting results (middle columns) and then computes contact probabilities as the affordance representation (right column)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 300, + 554, + 348 + ], + "blocks": [ + { + "bbox": [ + 55, + 305, + 295, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 305, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 295, + 327 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation of GPT-4V selection in terms of Overall and Contact." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 300, + 554, + 348 + ], + "lines": [ + { + "bbox": [ + 315, + 300, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 315, + 300, + 554, + 348 + ], + "type": "table", + "html": "
SettingMagic3DDreamFusionDreamFusion*DreamHOIOurs
Full4.36.517.326.045.6
w/o Refine6.56.517.330.439.1
w/o LLM-init8.610.821.732.626.1
", + "image_path": "51667a65a7395e4aea139030e0bfe294b8c07078c3d520f2333ae5211a237b0b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 356, + 554, + 378 + ], + "lines": [ + { + "bbox": [ + 313, + 356, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 554, + 378 + ], + "type": "text", + "content": "Table 3. Ablation study on LLM-guided initialization and HOI refinement by GPT-4V selection." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 401, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 593 + ], + "type": "text", + "content": "level of detail and the physical accuracy of the interactions. Magic3D [29], for instance, often generates overly simplified shapes with minimal texture and detail, as seen in the keyboard typing example, where the lack of intricacy undermines realism. Both Magic3D [29] and DreamFusion [43] fail to capture body structure or realistic human-object relationships, highlighting the difficulty of generating realistic HOI using general 3D scene generation models. DreamFusion*, which incorporates object mesh as input, retains a complete or partial human body shape but struggles with accurate contact and natural scale. This limitation is particularly noticeable in examples like the holding-chair scenario, where hand placement lacks precision. DreamHOI [67] improves upon DreamFusion in terms of HOI but still falls short in capturing fine interaction details, such as accurate foot placement on the basketball." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Inpainting Visualization. We further visualize the 2D inpainting results and affordance parsing on the same object in Figure 5 to emphasize the effectiveness of our method in disentangling affordances for distinct human-object interactions. The figure showcases various human object interactions with a chair, highlighting the ability to localize affordances accurately. This visualization provides valuable insight into the regions on the object that are most likely to be in contact with humans, which is critical for optimizing human-object interaction." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7021" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 288, + 130 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 288, + 130 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 288, + 130 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 288, + 130 + ], + "type": "image", + "image_path": "e1f9b0c20729d9595390fa44bbfc5b41855d0a6473302a9a6dc90fd4ded389ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 131, + 288, + 190 + ], + "blocks": [ + { + "bbox": [ + 63, + 131, + 288, + 190 + ], + "lines": [ + { + "bbox": [ + 63, + 131, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 63, + 131, + 288, + 190 + ], + "type": "image", + "image_path": "b0e1f99492fd2af975c2db8b6d768dc55a4de7651c998e27b1dc76a8d73d453a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 199, + 295, + 221 + ], + "lines": [ + { + "bbox": [ + 55, + 199, + 295, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 295, + 221 + ], + "type": "text", + "content": "Figure 6. Qualitative results of scene populating and scene interaction applications from different views." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 234, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 296, + 402 + ], + "type": "text", + "content": "Qualitative performance on Scene Populating. We demonstrate one possible way to populate scenes with InteractAnything. We first utilize the HPS dataset [14] to obtain the 3D scene and then integrate the generated HOI results into the scene by placing them in appropriate locations. By employing a ground loss and large language model, we ensure that the HOI results are naturally embedded within the scene without artifact issues such as object or human interpenetration. Furthermore, our method goes beyond simply populating HOI results into the scene, we also generate realistic interaction effects with existing objects in the environment. For example, InteractAnything can effectively generate interactions such as \"sitting on the table\" or \"opening a door\", as demonstrated in the Figure 6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 409, + 149, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 149, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 149, + 422 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 629 + ], + "type": "text", + "content": "Ablation on Fine-grained Optimization In this study, we analyze the effect of fine-grained optimization terms as outlined in Section 3.5 while keeping all other parameters constant. The results shown in Figure 4, clearly demonstrate that the fine-grained optimization significantly enhances the realism and contact plausibility of human-object interactions (right green panel) in the \"grasp\" prompt. Without fine-grained optimization (left red panel), the alignment between the hand and table is noticeably inaccurate, resulting in unnatural poses where the hand appears to \"float\" near the table instead of properly grasping or holding it. In contrast, with fine-grained optimization enabled (right panel), the human hand is aligned with the table surface, resulting in a more realistic interaction. This improvement highlights the importance of fine-grained optimization in achieving natural contact and precise positioning, especially in actions involving detailed hand-object interactions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Ablation Study on LLM-guided Initialization We further ablate the LLM-guided initialization and the corresponding HOI refinement module (fine-grained HOI synthesis) to investigate the contribution of our design in leveraging language models. As shown in Table 3, we use GPT-4V selection as the evaluation metric. Compared to the full method (Full), the variant without refinement (w/o Refine)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "exhibits a noticeable decline in performance. This drop is primarily due to the absence of fine-grained contact optimization, which is critical for generating the most accurate human-object interactions. In contrast, when the LLM-guided initialization is ablated (w/o LLM-init), we randomly initialize both the human and object positions, followed by coarse HOI optimization. This approach results in a significant decrease in selection scores, ranking lower than DreamHOI [67]. This finding emphasizes the importance of an informed initialization strategy in establishing a solid foundation for HOI optimization. The stark differences in the ablation study underscore the effectiveness of our design. The integration of LLM-guided initialization and refinement is vital for achieving superior results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 251, + 463, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 463, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 463, + 264 + ], + "type": "text", + "content": "5. Discussion and Limitation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 270, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 556, + 559 + ], + "type": "text", + "content": "In this work, we demonstrate that our approach can generate fine-grained, realistic HOIs for open-set 3D objects based on text descriptions. While our method shows promising results in improving contact precision and adapting to unseen objects, it still relies on 2D priors as its primary guidance. Consequently, advancements in multi-view consistent image generation or fully 4D generation techniques would likely enhance our approach, leading to more coherent and robust 3D interactions. Additionally, physical plausibility is also an important application in recent applications, e.g., in humanoid manipulation. We use force closure as a physical constraint while a high-quality simulator could further evaluate the generated HOI. Furthermore, our method currently uses the SMPL-H model to represent human body geometry and articulation. While this model captures a wide range of human poses and movements, it limits the adaptability of our framework to diverse interaction agents, particularly those with non-human anatomies. These agents often have unique skeletal structures and movement patterns that differ significantly from human biomechanics, which our current system is not fully equipped to process. In future work, we aim to address these limitations by investigating multi-view or 4D generation for enhanced consistency and exploring more flexible models to represent a broader array." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 569, + 388, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 388, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 388, + 582 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 588, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 555, + 708 + ], + "type": "text", + "content": "This paper tackles the challenge of generating novel 3D HOI involving unseen objects—a task where current methods often struggle. To address this issue, we propose InteractAnything, a novel zero-shot synthesis method driven by text instructions and not reliant on training with existing 3D datasets. It leverages guidance from the large language model to distill human-object relations and employs the pre-trained 2D diffusion model to parse contact areas with open-set objects, circumventing the limitations of existing HOI datasets." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7022" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by the National Science and Technology Major Project (2022ZD0114900) and NSFC-6247070125. We extend our gratitude to the reviewers for their insightful comments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 132, + 115, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 132, + 115, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 132, + 115, + 144 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 294, + 206 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 208, + 294, + 262 + ], + "type": "text", + "content": "[2] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 264, + 295, + 308 + ], + "type": "text", + "content": "[3] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep It SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. In Computer Vision - ECCV 2016, 2016. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 309, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 309, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 309, + 294, + 364 + ], + "type": "text", + "content": "[4] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 958-968, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 365, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 294, + 409 + ], + "type": "text", + "content": "[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 294, + 453 + ], + "type": "text", + "content": "[6] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 294, + 510 + ], + "type": "text", + "content": "[7] Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets creation using spatially-aware diffusion guidance. In European Conference on Computer Vision, pages 128-146. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 294, + 555 + ], + "type": "text", + "content": "[8] Hongkai Dai, Anirudha Majumdar, and Russ Tedrake. Synthesis and optimization of force closure grasps via sequential semidefinite programming. Robotics Research: Volume 1, pages 285-305, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 294, + 612 + ], + "type": "text", + "content": "[9] Sisi Dai, Wenhao Li, Haowen Sun, Haibin Huang, Chongyang Ma, Hui Huang, Kai Xu, and Ruizhen Hu. Interfusion: Text-driven generation of 3d human-object interaction. In European Conference on Computer Vision, pages 18-35. Springer, 2024. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 635 + ], + "type": "text", + "content": "[10] Google DeepMind. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 713 + ], + "type": "text", + "content": "[11] Clement Fuji Tsang, Maria Shugrina, Jean Francois Lafleche, Towaki Takikawa, Jiehan Wang, Charles Loop, Wenzheng Chen, Krishna Murthy Jatavallabhula, Edward Smith, Artem Rozantsev, Or Perel, Tianchang Shen, Jun Gao, Sanja Fidler, Gavriel State, Jason Gorski, Tommy Xiang, Jianing Li, Michael Li, and Rev Lebaredian. Kaolin: A pytorch library for accelerating 3d deep learning re" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "search. https://github.com/NVIDIAGameWorks/kaolin, 2022.6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[12] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Scholkopf. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 193 + ], + "type": "text", + "content": "[13] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13041-13051, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 195, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 249 + ], + "type": "text", + "content": "[14] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human positioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2021. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "type": "text", + "content": "[15] Sookwan Han and Hanbyul Joo. Chorus: Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In International Conference on Computer Vision (ICCV), 2023. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 295, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 337 + ], + "type": "text", + "content": "[16] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambiguities with 3D scene constraints. In International Conference on Computer Vision (ICCV), 2019. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "type": "text", + "content": "[17] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 426 + ], + "type": "text", + "content": "[18] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 427, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 553, + 459 + ], + "type": "text", + "content": "[19] Jingwei Huang, Hao Su, and Leonidas Guibas. Robust watertight manifold surface generation method for shapenet models. arXiv preprint arXiv:1802.01698, 2018. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "type": "text", + "content": "[20] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Zhiyuan Zhang, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. Full-body articulated human-object interaction. In International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "type": "text", + "content": "[21] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "text", + "content": "[22] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7122-7131, 2018. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 604, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 647 + ], + "type": "text", + "content": "[23] Tero Karras. Maximizing parallelism in the construction of bvhs, octrees, and k-d trees. In Proceedings of ACM SIGGRAPH/Eurographics Conference on High-Performance Graphics, 2012. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):139-1, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Hyeonwoo Kim, Sookwan Han, Patrick Kwon, et al. Beyond the contact: Discovering comprehensive affordance for 3d" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7023" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "objects from pre-trained 2d diffusion models. In European Conference on Computer Vision (ECCV), 2024. 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 129 + ], + "type": "text", + "content": "[26] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 295, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 174 + ], + "type": "text", + "content": "[27] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. In European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 295, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 219 + ], + "type": "text", + "content": "[28] Lei Li and Angela Dai. Genzi: Zero-shot 3d human-scene interaction generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20465-20474, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 275 + ], + "type": "text", + "content": "[29] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 295, + 331 + ], + "type": "text", + "content": "[30] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21159-21168, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 295, + 376 + ], + "type": "text", + "content": "[31] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 378, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 295, + 432 + ], + "type": "text", + "content": "[32] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters (RA-L), 7(1):470-477, 2021. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 488 + ], + "type": "text", + "content": "[33] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splatting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 491, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 295, + 544 + ], + "type": "text", + "content": "[34] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In International Conference on Learning Representations (ICLR), 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 295, + 610 + ], + "type": "text", + "content": "[35] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 656 + ], + "type": "text", + "content": "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Transactions on Graphics (TOG), 34(6):248:1-248:16, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[37] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[38] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 171 + ], + "type": "text", + "content": "[39] OpenAI. Gpt-3.5. https://platform.openai.com/docs/models/gpt-3-5, 2023. Accessed: 2023-11-14. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 553, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 236 + ], + "type": "text", + "content": "[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[41] Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 284, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 336 + ], + "type": "text", + "content": "[42] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 380 + ], + "type": "text", + "content": "[43] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations (ICLR), 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 447 + ], + "type": "text", + "content": "[44] Alec Radford, JongWook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Askell Amanda, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. Cornell University - arXiv, Cornell University - arXiv, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "text", + "content": "[45] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "text", + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "text", + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 593, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 634 + ], + "type": "text", + "content": "[48] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: modeling and capturing hands and bodies together. ACM Transactions on Graphics (TOG), 36(6):1-17, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "text", + "content": "[49] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[50] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7024" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "text", + "content": "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[51] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "type": "text", + "content": "[52] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[53] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 307 + ], + "type": "text", + "content": "[54] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J Black. Deco: Dense estimation of 3d human-scene contact in the wild. In International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 364 + ], + "type": "text", + "content": "[55] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In International Conference on 3D Vision (3DV), 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "text", + "content": "[56] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 487 + ], + "type": "text", + "content": "[57] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16210-16220, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 488, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 533 + ], + "type": "text", + "content": "[58] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In International Conference on Computer Vision (ICCV), 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "type": "text", + "content": "[59] Sirui Xu, Yu-Xiong Wang, Liangyan Gui, et al. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. Advances in Neural Information Processing Systems, 37:52858-52890, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "text", + "content": "[60] Hongwei Yi, Justus Thies, Michael J Black, Xue Bin Peng, and Davis Rempe. Generating human interaction motions in scenes with text control. In European Conference on Computer Vision (ECCV), 2025. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": "[61] Wanqi Yin, Zhongang Cai, Ruisi Wang, Fanzhou Wang, Chen Wei, Haiyi Mei, Weiye Xiao, Zhitao Yang, Qingping Sun, Atsushi Yamashita, et al. Whac: World-grounded humans and cameras. In European Conference on Computer Vision, pages 20-37, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[62] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monoc" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 363 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ular images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:12287-12303, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[63] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13232-13242, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[64] Jason Y. Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In European Conference on Computer Vision (ECCV), 2020. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[65] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "text", + "content": "[66] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on pattern analysis and machine intelligence, 44(6): 3170-3184, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "text", + "content": "[67] Thomas Hanwen Zhu, Ruining Li, and Tomas Jakab. Dreamhoi: Subject-driven generation of 3d human-object interactions with diffusion priors. arXiv preprint arXiv:2409.08278, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7025" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/9f03277a-37b8-49e7-9fad-13e2e5f171b2_content_list.json b/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/9f03277a-37b8-49e7-9fad-13e2e5f171b2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c8ca12f864b0951fa438926dbea75d4078337cd9 --- /dev/null +++ b/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/9f03277a-37b8-49e7-9fad-13e2e5f171b2_content_list.json @@ -0,0 +1,2083 @@ +[ + { + "type": "text", + "text": "InteractVLM: 3D Interaction Reasoning from 2D Foundational Models", + "text_level": 1, + "bbox": [ + 138, + 117, + 859, + 138 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sai Kumar Dwivedi1 Dimitrije Antić2 Shashank Tripathi1 Omid Taheri1 Cordelia Schmid3 Michael J. Black1 Dimitrios Tzionas2", + "bbox": [ + 181, + 160, + 805, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Max Planck Institute for Intelligent Systems, Tübingen, Germany 2University of Amsterdam (UvA), the Netherlands 3Inria, École normale supérieure, France", + "bbox": [ + 102, + 208, + 893, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9406fdef2334f1a5f620b804e6bdddfa4e1a21b853286f21a008a8517cc7ce8a.jpg", + "image_caption": [ + "Figure 1. We present InteractVLM, a novel method for estimating contact points on both human bodies and objects from a single in-the-wild image, shown here as red patches. Our method goes beyond traditional binary contact estimation methods by estimating contact points on a human in relation to a specified object. We do so by leveraging the broad visual knowledge of a large Visual Language Model." + ], + "image_footnote": [], + "bbox": [ + 102, + 263, + 895, + 453 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 518, + 326, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce InteractVLM, a novel method to estimate 3D contact points on human bodies and objects from single in-the-wild images, enabling accurate human-object joint reconstruction in 3D. This is challenging due to occlusions, depth ambiguities, and widely varying object shapes. Existing methods rely on 3D contact annotations collected via expensive motion-capture systems or tedious manual labeling, limiting scalability and generalization. To overcome this, InteractVLM harnesses the broad visual knowledge of large Vision-Language Models (VLMs), fine-tuned with limited 3D contact data. However, directly applying these models is non-trivial, as they reason \"only\" in 2D, while human-object contact is inherently 3D. Thus we introduce a novel \"Render-Localize-Lift\" module that: (1) embeds 3D body and object surfaces in 2D space via multiview rendering, (2) trains a novel multi-view localization model (MV-Loc) to infer contacts in 2D, and (3) lifts these to 3D. Additionally, we propose a new task called Semantic Human Contact estimation, where human contact predictions are conditioned explicitly on object semantics, enabling richer interaction modeling. InteractVLM outperforms existing work on contact estimation and also facilitates 3D reconstruction from an in-the-wild image. To es", + "bbox": [ + 88, + 551, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "timate 3D human and object pose, we infer initial body and object meshes, then infer contacts on both of these via InteractVLM, and last exploit these for fitting the meshes to image evidence. Results show that our approach performs promisingly in the wild. Code and models are available at https://interactvlm.is.tue.mpg.de.", + "bbox": [ + 511, + 518, + 908, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 643, + 645, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "People interact with objects routinely. Reconstructing human-object interaction (HOI) in 3D is key for many applications, from robots to mixed reality. However, doing so from single images is challenging due to depth ambiguity, occlusions, and the diverse shape and appearance of objects.", + "bbox": [ + 511, + 672, + 908, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are methods that estimate 3D human bodies and methods that estimate 3D objects, but few that put these together. Knowing the contacts between these could significantly improve joint reconstruction. Our goal is to infer contact points on both humans and objects from single in-the-wild images, and then use these contacts for jointly reconstructing humans and objects. However, there is a lack of in-the-wild training images paired with ground-truth contact labels for both 3D humans and objects. Acquiring such data is challenging and existing methods do not scale.", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22605", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The problem gets more challenging due to the complexity of real-world interactions. Humans often contact multiple objects simultaneously; e.g., using a laptop while sitting on a club chair. Yet, current approaches treat contact prediction as simple binary classification; i.e. detecting whether a body part is in contact with \"any\" object. This simplified assumption fails to capture the rich semantic relationships of multi-object interactions. To address this, we introduce a novel \"Semantic Human Contact\" estimation task. This involves predicting the contact points on the body related to a particular object given a single in-the-wild image.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this, as well as to overcome data scarcity, we propose a new paradigm for scaling and reasoning in the wild. Specifically, we observe that large Vision-Language Models (VLMs) can \"reason\" about in-the-wild images because they are trained on internet-scale data and possess a broad visual knowledge about humans and their interactions with the world. We also observe that this knowledge can be re-purposed for novel tasks by fine-tuning these large models on small datasets. Thus, we exploit VLMs for developing a novel framework, called InteractVLM.", + "bbox": [ + 89, + 258, + 482, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At the heart of InteractVLM lies a reasoning module based on a VLM; see Fig. 2. We enrich this with skills for 3D human-object \"understanding,\" by extending the VLM with a LoRA [26] adaptation layer. As a result, given a color image, this module can be \"asked\" to produce \"reasoning tokens\" that facilitate 3D contact localization.", + "bbox": [ + 89, + 411, + 482, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, exploiting these tokens to localize contact is non-trivial. A natural choice would be to employ a foundational \"localization\" model [31] that takes these tokens as \"guidance,\" and highlights the 3D contacts. But there exists a key practical problem. Existing foundation models inherently operate only in 2D space, while we need them to do so in 3D. To tackle this, we need to re-cast our problem so it is appropriate for 2D foundation models.", + "bbox": [ + 89, + 503, + 482, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we develop a novel \"Render-Localize-Lift\" (RLL) framework that has three main steps; see Fig. 2: (1) We render the 3D shape of a canonical SMPL+H [53] body and an object as 2D images from multiple viewpoints. Regarding the object, a 3D mesh is efficiently retrieved from a large-scale 3D database [11] through OpenShape [42]. (2) We pass the above images into a foundation model to predict 2D contact masks for both the body and the object. (3) We lift the predicted 2D contact points to 3D points via back-projection, i.e., performing the inverse of the first step.", + "bbox": [ + 89, + 626, + 482, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, even after recasting the problem to 2D by rendering multi-view images, existing foundation models are still not 3D aware, i.e. they treat each view independently, ignoring multi-view consistency. This means contact detections in one view do not necessarily agree with ones in adjacent views. To tackle this, just appending camera parameters to multi-view renderings is insufficient. Instead, we build a novel \"Multi-view Localization\" model that we call", + "bbox": [ + 89, + 780, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/41285023f276fd719bccc053beeb3d443a9efa6a33aadaf318560de488cc166c.jpg", + "image_caption": [ + "Figure 2. Overview of InteractVLM. Given a color image, our VLM performs the core reasoning, and guides a novel MV-Loc model to localize contacts on both bodies and objects in 3D. Here we show only the body; for details, and object contact, see Fig. 3." + ], + "image_footnote": [], + "bbox": [ + 524, + 92, + 895, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MV-Loc. This has two steps: (1) It transforms the \"reasoning token\" provided by the VLM with the camera parameters used to render the multi-view images. (2) It ensures multi-view consistency, by lifting the inferred 2D contacts in each view to 3D and computing a 3D loss.", + "bbox": [ + 511, + 311, + 903, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method, InteractVLM, utilizes a VLM in tandem with our novel multi-view localization model, MV-Loc, to perform 3D contact prediction for humans and objects. See Fig. 1 for some examples. We quantitatively evaluate the efficacy of our method for in-the-wild 3D contact prediction on bodies and on objects, using the DAMON [58] and PIAD [65] datasets, respectively. For bodies, we evaluate both for the traditional \"binary contact\" estimation, and for our new task of \"semantic contact\" estimation. For all tasks, we find that InteractVLM outperforms the prior work.", + "bbox": [ + 511, + 387, + 903, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we demonstrate how InteractVLM's estimated contact improves 3D HOI recovery from in-the-wild images; this is a highly ill-posed task due to depth ambiguities and occlusions. To address this, we develop an optimization-based method that fits a SMPL-X body mesh and an OpenShape-retrieved object mesh to the image, using InteractVLM's inferred contacts as constraints to anchor human and object meshes w.r.t. each other. To the best of our knowledge, this is the first approach for estimating 3D HOI for in-the-wild images using inferred contacts.", + "bbox": [ + 511, + 537, + 906, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we make four key contributions:", + "bbox": [ + 532, + 689, + 834, + 703 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We build InteractVLM, a novel method that facilitates HOI reconstruction from an in-the-wild image by detecting 3D contacts on both bodies and objects.", + "2. We demonstrate a way to minimize reliance on 3D contact annotations via exploiting the broad visual knowledge of Vision-Language Models.", + "3. We build a novel \"Multi-view Localization\" model that helps in estimating contacts in 3D by transforming the reasoning of foundation models from 2D to 3D.", + "4. We introduce the novel \"Semantic Human Contact\" task for inferring body contacts conditioned on object labels. Our code and trained models are available for research at https://interactglm.is.tue.mpg.de." + ], + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22606", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 233, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Large Vision-Language Models", + "text_level": 1, + "bbox": [ + 89, + 114, + 367, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent advancements in large language models (LLMs) have led to the development of multimodal models that integrate vision and language reasoning. Models like Flamingo [2] and BLIP-2 [36] use cross-attention mechanisms and visual encoders to align image features with text, supporting a variety of vision-language tasks. More recent works, such as VisionLLM [60] and Kosmos-2 [50], use grounded image-text data to enhance spatial understanding, while GPT4RoI [68] introduces spatial box inputs for finer alignment. However, these models typically lack end-to-end segmentation capabilities. To address this limitation, LISA [34] combines vision foundation segmentation models like SAM [31] with multimodal embeddings, enabling language-guided segmentation. PARIS3D [13] extends this approach to referential 3D segmentation by processing multi-view object renders through both SAM and LLaVA [41] for spatially-aware segmentation.", + "bbox": [ + 89, + 138, + 482, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Taking inspiration from these approaches, we exploit a language-guided segmentation model for the task of predicting human and object contact in 3D. However, unlike PARIS3D, we process a single RGB image with LLaVA and multi-view renders of the human mesh and the object mesh with SAM. Moreover, we introduce a feature-lifting technique that extends LLaVA's 2D features into 3D using camera parameters, thus guiding SAM's multi-view segmentations. This approach ensures multi-view consistency and efficiently predicts contact affordances, extending the use of multimodal models in 3D reasoning tasks focused on human-object interaction.", + "bbox": [ + 89, + 396, + 482, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. 3D Human and Object from Single Images", + "text_level": 1, + "bbox": [ + 89, + 589, + 454, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Estimating 3D human pose and shape from a single image has evolved from optimization-based methods to learning-based approaches. Optimization-based methods fit parametric body models like SMPL [45], SMPL-X [49], or GHUM [64] to 2D cues such as keypoints [9], silhouettes [48], or segmentation masks [48]. Learning-based methods either regress body parameters from images or videos [5, 14, 15, 29, 32, 33, 37] or estimate non-parametric bodies as vertices [35, 40], implicit surfaces [46, 54], or dense points [56]. Transformer-based methods [16, 21, 39, 55] have further improved robustness.", + "bbox": [ + 89, + 612, + 482, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For 3D object reconstruction from a single image, regression-based methods predict geometry using meshes, voxels, or point clouds. Diffusion-based models [25] utilize large 3D datasets like Objaverse [11] or 2D diffusion models [38, 43, 44, 51] to guide reconstruction and optimization methods [4] use render and compare. Retrieval-based methods, such as OpenShape [42] and Uni3D [69], have demonstrated some robustness in cases with occlusions.", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. 3D Human-Object Interaction", + "text_level": 1, + "bbox": [ + 513, + 90, + 782, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Understanding 3D human-object interactions is essential for modeling realistic scenes. Early works focused on handobject interactions, such as ObMan [24] and FPHA [20], with more recent studies like ARCTIC [17] and HOLD [18] providing more detailed data and reconstruction for hands. For full-body interactions, initial studies involve interactions with scenes, as in PROX [22], and with objects, as in BEHAVE [7], GRAB [57], and InterCap [28]. BEHAVE and GRAB use motion capture that are accurate but nonscalable, while InterCap uses multi-camera setups that are more scalable but less accurate. Both of these approaches are limited in capturing diverse and realistic interactions.", + "bbox": [ + 511, + 112, + 903, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a proxy for 3D reconstruction, the recent DECO [58] method infers contacts on body meshes; to this end, 3D contact annotations are crowdsourced via mesh painting. Predicting contact on objects is more challenging due to varying object shapes, while also currently there exists no dataset of object contact for in-the-wild images. Thus, we estimate 3D object affordances as a proxy for contact. 3D-AffordanceNet [12] introduces affordances that are not grounded in images, capturing the likelihood of humans interacting with specific parts of an object for a given affordance (e.g., \"sit on chair\"). PIAD [65] curates RGB images depicting object affordances and trains a network to estimate them. LEMON [66] extends object affordance prediction of PIAD to include human contact estimation. However, these methods require human contact vertices paired with object affordances for training, limiting the number of categories they handle to 21 ones. In contrast, we learn from unpaired human and object interaction data and perform joint interaction reasoning via a cross-attention layer, enabling human interaction reasoning for 80 categories and object affordance prediction for 32 categories.", + "bbox": [ + 511, + 294, + 903, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Joint 3D Human-Object Reconstruction", + "text_level": 1, + "bbox": [ + 513, + 621, + 856, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Joint reconstruction of humans and objects in 3D has been tackled with both regression and optimization methods. Regression-based methods directly predict 3D human-object meshes, as in HDM [62] and CONTHO [47], while other methods first predict contact points and then use test-time optimization to fit human and object meshes, as in CHORE [61] and PHOSA [67]. Since regression methods rely on limited training data, optimization methods are preferred for in-the-wild scenarios, such as PHOSA [67]. Optimization-based methods either assume known contacts or infer contacts to fit meshes to the image, but their success heavily relies on the quality of contact.", + "bbox": [ + 511, + 643, + 903, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method improves upon these by providing more accurate contact predictions, which in turn facilitate better fitting of human and object meshes to image evidence, improving the realism and accuracy of 3D human-object reconstructions from single images.", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22607", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4c7e3284f279be476c0109e0c6cc5a8b798e40e2af9aaaf8c482c32c26185e0a.jpg", + "image_caption": [ + "Figure 3. Method overview. Given a single in-the-wild color image, our novel InteractVLM method estimates 3D contact points on both humans and objects (a). Then, we reconstruct a 3D human and object in interaction by exploiting these contacts (b). More specifically: (a) Contact estimation. Given an image, $I$ , and prompt text, $T_{inp}$ , our VLM, $\\Psi$ , produces contact tokens for humans and objects, $<\\mathsf{HCON}>$ and $<0\\mathsf{CON}>$ , which are projected ( $\\Gamma$ ) into feature embeddings, $E^H$ and $E^O$ . These guide a \"Multi-View [contact] Localization\" model. This renders the 3D human and object geometry via cameras, $K$ , into multi-view 2D renders and passes these to encoder, $\\Theta$ , while decoders, $\\Omega^H$ , $\\Omega^O$ , estimate and highlight 2D contacts in these renders. Then, the FeatLift module, $\\Phi$ , transforms the VLM's features $(E^H, E^O)$ to become 3D-aware $(E_{3D}^H, E_{3D}^O)$ by exploiting the camera parameters, $K$ . A final module lifts the detected 2D contacts to 3D. (b) 3D HOI reconstruction. For joint human-object reconstruction, we use InteractVLM's inferred contacts in an optimization framework." + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 728, + 333 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5974398eb285b904d1de43c22f78dcae6534bc3da026842cc044de35856ce37e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 88, + 893, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 479, + 181, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Input Representation", + "text_level": 1, + "bbox": [ + 89, + 506, + 290, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an image, $I \\in \\mathbb{R}^{H \\times W \\times 3}$ , InteractVLM estimates 3D contacts for both human bodies and objects.", + "bbox": [ + 89, + 529, + 482, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The human is represented by a SMPL+H [53] 3D body mesh, $H$ , with vertices $V^{H} \\in \\mathbb{R}^{10475 \\times 3}$ . The body is posed in a canonical \"star\" shape (see details in Sec. 3.4). The human contacts are binary per-vertex labels, $C^{H} \\in \\{0,1\\}$ .", + "bbox": [ + 89, + 561, + 482, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The object is represented by a 3D point cloud (or mesh), $O \\in \\mathbb{R}^{N \\times 3}$ , with $N$ points. As there are no datasets of natural images paired with 3D contacts for objects, we use a largescale 3D affordance dataset [65], instead, as affordances are closely related to contact. Specifically, they represent the likelihood of contact on 3D object areas for various purposes. Therefore, for objects we use the terms \"affordance\" and \"contact\" interchangeably. The object contacts are continuous per-point values, $C^O \\in [0,1]$ . During inference, we retrieve a 3D object shape from a large database [11] via OpenShape [42] conditioned on image $I$ .", + "bbox": [ + 89, + 623, + 482, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overview of InteractVLM", + "text_level": 1, + "bbox": [ + 89, + 801, + 328, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The biggest challenge for learning 3D contact prediction in the wild is the limited 3D contact data for humans and objects. To go beyond existing limited datasets, we introduce a novel method, called InteractVLM, that harnesses the commonsense knowledge of large vision-language models.", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, InteractVLM (Fig. 3) has two main components: a Vision Language Model (VLM) and a novel Multi-View contact Localization model (MV-Loc). MV-Loc highlights parts that are in contact for both humans and objects with the VLM's guidance. The input to the VLM (Sec. 3.3) is an image, $I$ , and a prompt, $T_{inp}$ , that asks the VLM to detect contact. The input to MV-Loc (Sec. 3.4) is the 3D geometry of humans and objects, $H$ and $O$ , respectively.", + "bbox": [ + 511, + 482, + 906, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Interaction Reasoning through VLM", + "text_level": 1, + "bbox": [ + 511, + 630, + 831, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The VLM module, $\\Psi$ , conducts the core interaction reasoning. It takes an input image, $I$ , and prompt text, $T_{inp}$ , and outputs a prompt text, $T_{out} = \\Psi(I, T_{inp})$ . Inspired by the recent LISA [34] model, we expand the VLM's vocabulary with two specialized tokens, $\\langle \\text{HCON} \\rangle$ and $\\langle \\text{OCON} \\rangle$ , for contact information for humans and objects, respectively.", + "bbox": [ + 511, + 657, + 905, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To denote contact, $\\Psi$ produces a prompt that includes the above tokens. To aid MV-Loc in localizing contact, we extract the last-layer embeddings of the VLM corresponding to these tokens and pass them through a projection layer, $\\Gamma$ , to obtain the feature embeddings, $E^H$ or $E^O$ . Let $T_{\\mathrm{gt}}$ be the ground-truth text, and $T_{\\mathrm{pred}}$ be the predicted one. Then, our token-prediction loss is defined as a cross-entropy loss:", + "bbox": [ + 511, + 753, + 906, + 861 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o k e n}} = - \\sum_ {i = 1} ^ {N} \\left(T _ {\\mathrm {g t}} ^ {(i)} \\cdot \\log \\left(T _ {\\text {p r e d}} ^ {(i)}\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 883, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22608", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Interaction Localization through MV-Loc", + "text_level": 1, + "bbox": [ + 89, + 90, + 444, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We develop a novel MV-Loc module that has a shared image encoder, $\\Theta$ , and separate decoders, $\\Omega^H$ and $\\Omega^O$ , for humans and objects. MV-Loc performs contact localization by using a novel \"Render-Localize-Lift\" (RLL) framework. To this end, it takes three steps; (1) rendering the 3D shape of humans and objects in 2D, (2) predicting 2D contact maps for both of these, and (3) lifting the 2D contact maps to 3D.", + "bbox": [ + 89, + 112, + 483, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RLL step #1: Render 3D $\\rightarrow$ 2D. The input is human and object geometry, namely $\\mathcal{H}$ and $O$ , respectively. Both serve as a \"canvas\" for \"painting\" detected contacts on these. The body has a default SMPL+H shape in a canonical star-shape pose to minimize self-occlusions when rendering. The object geometry (initialized in Sec. 3.1) is normalized to a unit sphere. Each geometry is rendered from $J$ fixed views with camera parameters, $K$ , to form multi-view renderings, $R^{H,O} = \\{R_j\\}_{j=1}^J$ , so that the entire 3D geometry is captured. Since our geometries do not have texture, we color the meshes with normals and point clouds using the NOCS map [59]. This enhances cross-view correspondence, making renderings resemble real images to image encoder, $\\Theta$ .", + "bbox": [ + 89, + 219, + 483, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RLL step #2: Localize in 2D. The rendered geometry, $R^{H,O}$ , is first sent to the image encoder, $\\Theta$ , and then passed to decoders, so that the final contact masks, $M^H$ and $M^O$ , get highlighted on it. However, MV-Loc requires spatial and contextual cues for highlighting the contact region. To this end, we use the feature embeddings (Sec. 3.3), i.e. $E^H$ and $E^O$ to guide the contact localization.", + "bbox": [ + 89, + 416, + 483, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, since the VLM reasons in 2D, these features are not 3D aware. This is a problem, because MV-Loc needs 3D awareness to localize contact consistently across multi-view renderings. Thus, we transform the features to \"lift\" them to \"3D\" to better guide multi-view localization.", + "bbox": [ + 89, + 522, + 483, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In detail, we design a lifting network, $\\Phi$ , which takes camera parameters, $K$ , and the 2D $E^{H,O}$ , and lifts the latter to 3D as $E_{3D}^{H,O} = \\Phi(E^{H,O}, K)$ . Contact masks are defined as:", + "bbox": [ + 89, + 598, + 483, + 646 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM ^ {H, O} = \\Omega^ {H, O} (\\Theta (R ^ {H, O}), E _ {3 D} ^ {H, O}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 652, + 482, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Below, the $H$ and $O$ superscripts are dropped for brevity. We calculate losses only on \"valid\" regions, i.e., the areas within the outline of rendered geometry; we denote this as $M$ . To encourage overlap between the predicted masks, $M$ , and the ground-truth ones, $\\widehat{M}$ , particularly in sparse contact regions, we use a focal-weighted BCE loss and a Dice loss:", + "bbox": [ + 89, + 681, + 483, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B C E} = - \\alpha (1 - p _ {M}) ^ {\\gamma} \\log (p _ {M}) - (1 - \\alpha) p _ {M} ^ {\\gamma} \\log (1 - p _ {M}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 782, + 482, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D i c e} = 1 - \\frac {2 \\sum M \\cdot \\widehat {M} + \\epsilon}{\\sum M + \\sum \\widehat {M} + \\epsilon}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 811, + 482, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_M$ is the predicted mask probability, $\\alpha$ controls class balance, $\\gamma$ adjusts the focus on hard examples, and $\\epsilon$ is a residual term to prevent division by zero.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RLL step #3: Lift $2\\mathbf{D} \\rightarrow 3\\mathbf{D}$ . To lift the inferred 2D contact points to 3D points, we follow the inverse of step #1. Normally, 2D points backproject to 3D lines due to depth ambiguities. In our case the lines intersect with the known 3D geometry that produced the multi-view renders. So, 2D points are lifted to 3D points, and by extension 2D contact masks, $M^{H,O}$ , are lifted to 3D contact areas, $C^{H,O}$ .", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a human contact loss, $\\mathcal{L}_C^H$ , that combines a focal loss with sparsity regularization, to encourage precise true-positive predictions in \"valid\" contact regions while discouraging false positives in non-contact areas:", + "bbox": [ + 511, + 196, + 903, + 257 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C} ^ {H} = \\alpha (1 - p _ {h C}) ^ {\\gamma} \\log \\left(p _ {h C}\\right) + \\lambda \\| C ^ {H} \\| _ {1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 263, + 903, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_{hC}$ is the contact probability, and $\\lambda$ , $\\alpha$ and $\\gamma$ are scalar weights. We also use an object contact loss, $\\mathcal{L}_C^O$ , that combines a Dice and Mean Squared Error (MSE) loss:", + "bbox": [ + 511, + 291, + 903, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C} ^ {O} = \\mathcal {L} _ {\\text {D i c e}} \\left(C ^ {O}, \\widehat {C} ^ {O}\\right) + \\beta \\| C ^ {O} - \\widehat {C} ^ {O} \\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 343, + 903, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is a weighting factor, and $\\widehat{C^O}$ denotes the groundtruth 3D contacts for objects.", + "bbox": [ + 511, + 371, + 903, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 410, + 730, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture. We use LLaVA [41] as our VLM and SAM [31] for our MV-Loc, with weights pre-trained by LISA [34] for segmentation [34]. The feature-lifting network, $\\Phi$ , contains a spatial-understanding network (two fully-connected layers of size 128 with ReLU activation) followed by view-specific (256-dimensional) transformations and a sigmoid activation. For 3D contact prediction, our MV-Loc model converts 2D masks to 3D contact points via 2D-to-3D pixel-to-vertex mappings that are precomputed during MV-Loc's rendering step. For details, see Sup. Mat.", + "bbox": [ + 511, + 431, + 905, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. To efficiently fine-tune our VLM, we employ LoRA [26] with rank 8. The separate decoders for human and object contact prediction are trained without LoRA, while keeping the image encoder frozen. For training, we use DeepSpeed [3] with mixed precision training (bfloat16), and batch size of 8. We train on 4 Nvidia-A100 GPUs for 30 epochs. For more details, please refer to Sup. Mat.", + "bbox": [ + 511, + 584, + 905, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We focus on two tasks, i.e., 3D human contact and 3D object affordance prediction, using two in-the-wild datasets, i.e., DAMON [58] and PIAD [65], respectively. For human contact we train and evaluate on DAMON [58]. For 3D object affordances, we train and evaluate on PIAD [65]. We find that exploiting textual descriptions for the contacting body parts, and the object type in contact, helps training. Similarly, adding Visual Question-Answering (VQA) data generated by GPT4o for the training images also helps. For details, see Sup. Mat.", + "bbox": [ + 511, + 689, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike LEMON [66], which requires paired human-object geometry for training using the 3DIR dataset [66], we use unpaired data. This enables us to scale for many human and object categories not addressed by prior work. For", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22609", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b790d7ba3199355cd3386090a7cc8b1ac61fb826be36c6253b0d4de97b877675.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodF1 (%)↑Precision (%)↑Recall (%)↑Geodesic (cm)↓
POSAPPIXIE [23] [19]31.042.034.033.00
BSTRO [27]46.051.053.038.06
DECO [58]55.065.057.021.32
InteractVLM75.675.276.02.89
", + "bbox": [ + 104, + 82, + 470, + 172 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/83cc82727fd8975c73eee7aff2a1c20b07491ac2f9ce47f901a447eab6d4c0d0.jpg", + "table_caption": [ + "Table 1. Evaluation for \"Binary Human Contact\" prediction on the DAMON dataset [58]. We compare our InteractVLM model (trained only for this task) with the state of the art." + ], + "table_footnote": [], + "table_body": "
Object CategoriesSemantic-DECO [58] (Baseline)InteractVLM
F1 (%)↑Prec. (%)↑Rec. (%)↑Geo (cm)↓F1 (%)↑Prec. (%)↑Rec. (%)↑Geo (cm)↓
Accessory40.130.775.621.8861.172.965.66.91
Daily Obj26.520.152.360.3468.671.478.07.46
Food11.719.412.949.6166.466.177.97.17
Furniture24.515.883.729.1760.564.260.16.21
Kitchen27.724.737.252.3471.871.381.17.61
Sports36.430.480.179.2177.976.783.27.98
Transport52.039.193.731.7877.880.578.97.97
", + "bbox": [ + 94, + 239, + 478, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the final joint human-object reconstruction task, we combine DAMON [58], PIAD [65], 3DIR [66] and all textual descriptions about body parts, contact type and HOI.", + "bbox": [ + 89, + 441, + 482, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. For human contact prediction, following Tripathi et al. [58], we report the F1, precision, and recall scores using a threshold of 0.5, and a geodesic distance measuring spatial accuracy. For object contact prediction, following Yang et al. [65], we report the Similarity (SIM), Mean Absolute Error (MAE), Area Under ROC Curve (AUC), and average Intersection over Union (IOU).", + "bbox": [ + 89, + 487, + 483, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 609, + 223, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. \"Binary Human Contact\" Estimation", + "text_level": 1, + "bbox": [ + 89, + 633, + 415, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This task refers to estimating contact areas on the body via binary classification of its vertices, ignoring the number or type of objects involved. We train and evaluate on the DAMON [58] dataset, and report results in Tab. 1.", + "bbox": [ + 89, + 657, + 482, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "InteractVLM significantly outperforms all previous methods achieving a $20.6\\%$ improvement in F1 score. Although, here, InteractVLM is trained on the same data as DECO, it goes beyond this by harnessing the commonsense knowledge of a large foundation model.", + "bbox": [ + 89, + 718, + 482, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Sup. Mat. we also evaluate on the 3DIR [66] dataset and compare with the LEMON method [66]. Even though LEMON uses paired human-object data, InteractVLM performs on-par with it despite training on human-only data. We also evaluate InteractVLM's performance for binary human contact prediction across different body parts; for more details please refer to Sup. Mat.", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4cd5b6f95464648c083ae9c0ec75f577239936ad0ba8cdf67164aa1cbce38014.jpg", + "table_caption": [ + "Table 2. Evaluation for \"Semantic Human Contact\" prediction on the DAMON [58] dataset. For results on each class, see Sup. Mat. The \"Semantic-DECO\" baseline extends DECO for our new task." + ], + "table_footnote": [], + "table_body": "
MethodsPIAD-Seen [65]PIAD-Unseen [65]
SIM (%)↑AUC (%)↑aIOU (%)↑MAE ↓SIM (%)↑AUC (%)↑aIOU (%)↑MAE ↓
PMF [70]42.575.0510.131.4133.060.254.672.11
ILN [10]42.775.8411.521.3732.559.694.712.07
PFusion [63]43.277.5012.311.3533.061.875.331.93
XMF [1]44.178.2412.941.2734.262.585.681.88
IAGNet [65]54.584.8520.510.9835.271.847.951.27
InteractVLM62.786.4721.200.8141.475.458.500.99
", + "bbox": [ + 517, + 82, + 903, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Evaluation for \"Object Affordance Prediction\" on the PIAD [65] dataset. We compare our InteractVLM model (trained only for this task) with the state of the art.", + "bbox": [ + 511, + 204, + 906, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. \"Semantic Human Contact\" Estimation", + "text_level": 1, + "bbox": [ + 511, + 272, + 856, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In real-life interactions, multiple objects can be contacted by different body areas concurrently. Thus, we introduce a novel task called \"Semantic Human Contact\" prediction. We evaluate on DAMON [58] and report results in Tab. 2; for a finer-grained version of this table, see Sup. Mat.", + "bbox": [ + 511, + 294, + 906, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To establish a baseline, we adapt the DECO model [58] that detects binary contacts and turn it into a multi-class prediction model, called Semantic-DECO. Due to DAMON's limited training data, this has poor performance. Instead, as discussed in Sec. 4.1, InteractVLM learns effectively from this data, by also leveraging the commonsense of foundation models. Thus, it significantly outperforms Semantic-DECO. Qualitative results reflect this finding; see Fig. 4. Our model captures detailed, accurate contact regions, whereas Semantic-DECO often highlights false-positive areas that differ from the actual contact regions.", + "bbox": [ + 511, + 369, + 908, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Object Affordance Prediction", + "text_level": 1, + "bbox": [ + 511, + 545, + 777, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of our InteractVLM model on predicting object affordances. This involves identifying regions on objects of possible contact for a certain interaction intent, such as \"sitting\" on a chair or \"moving\" it. We train and evaluate on the PIAD [65] dataset. We compare against SotA methods, and report results in Tab. 3.", + "bbox": [ + 511, + 566, + 906, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that we evaluate on object instances that are both seen during training (\"PIAD-Seen\" column) and unseen (\"PIAD-Unseen\" column). Our model demonstrates a significant improvement over SotA methods in terms of similarity (SIM), area under the curve (AUC), and mean absolute error (MAE) for both seen and unseen objects. Thus, our method is effective not only for estimating human contact (Secs. 4.1 and 4.2), but also for object affordances.", + "bbox": [ + 511, + 657, + 905, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Reliance on 3D Annotations", + "text_level": 1, + "bbox": [ + 511, + 787, + 764, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To analyze InteractVLM's efficiency regarding 3D supervision, we train multiple versions of it with varying amounts of training data from the DAMON dataset, and report performance in Fig. 5. Remarkably, InteractVLM achieves an F1 score of 0.53 with just $1\\%$ of the data, nearly matching the 0.55 F1 score of DECO that uses $100\\%$ of the data. With", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22610", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae3586bc333315df4653daa80ee9ae4c9d0b7a81f1b8a9e4c4ed20a903dcc502.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 88, + 176, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6396a9845e6ff0d9091ac19b8b8f42330c43179719d50c8236bf6f54a044361.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 152, + 174, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/488c4b08b65e2406a8bf825dfa745f760286402b130aedb6eace06fbeeee00dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 212, + 174, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/782f998cf68e81f129817bf575cceb17990561b5d132fd5164ac64d13c57ed9a.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 96, + 275, + 174, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5feea365bfeab32fee4c3cd916173581b8b1ae1f4179e856f3d36aef9732eae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 94, + 251, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b0fd1a649fdeef1f8139e8d48c3ff1fa3496ffb7c1e2d633ae1e56cdee89cca4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 154, + 251, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c941e71980b6d4dcf76bcdd82b049125dfef6a72550a55b0ac897dfb15bf7331.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 218, + 251, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03ccf57b1801d4a4f931d16e58d07e348b7f0052875ba97e67e0fd192e263aa2.jpg", + "image_caption": [ + "Figure 4. \"Semantic Human Contact\" estimation (Sec. 4.2). Given an image and an object label, InteractVLM infers body contacts for this object. InteractVLM outperforms a Semantic-DECO [58] baseline. Objects are shown in green circles, and contacts as red patches." + ], + "image_footnote": [], + "bbox": [ + 179, + 277, + 251, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddb8cb87ee440601057e4ea4f735759f0301a155909ff4924f6d932bc60eaab2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 95, + 326, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9eadf1322ae1a0d7532202d5fee9b7fd05a2375cb8e93de9e98e94b88ad1ade2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 154, + 326, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b3abd25bf064fc3e6c042acfeab68e1f1659b537c52aae15eced057ea7db150.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 218, + 326, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2e08016fd73f4956fe8c273d08b8dacceb73284e72491085d77e9ccce4be8bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 277, + 326, + 351 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3602418ca5a1d4f427db3b2eb7564cd181134080c8830860d6d840285212cb6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 94, + 411, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/380edd433ade8b0e5f53a20501012cc525d94e0c547617f0fad71e36677d8f27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 154, + 411, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c136ece426d8e4e35c646eb7b4f6bed244501159ae986d13164a3052bfe683f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 218, + 411, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2194b3977f64dc86e4cd2ee1d51ea06aac6f4eea7eeb8a5b550862155c5b8175.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 277, + 411, + 351 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7f5c18cbde0f0fa7c41594bc05acda098e97f9a9a8078ac60cc37c194b5b205.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 95, + 486, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bba2b5ec99a4344007b7dfdf30ccafd035cbf373cbed7d00999c937faa812aaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 154, + 485, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/842c087a139a155405ca782dfc93e5e1a14661104284aad7ca9b709d417215c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 277, + 486, + 351 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8639aa8775480f870d1b6cfd2d9e0f259357d7936d88275276fb64f479db797.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 89, + 576, + 151 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01dde2f85ab79a24f2fe7065d544a5e6c0ee6fd5d7a5de167c4a40a088893698.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 152, + 576, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4bec1acbe6593a4079b2817b57d6d3c76de2db2d6e11cfa67ea635c58ab16fbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 214, + 576, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39fa99a62a82437025f3b6ea4603407b4136cf7a794df8d3c136f294d27483b6.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 496, + 275, + 578, + 351 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/041a58c9585b777d5b7cfc1af9653f4eae9535cbc7bfb545cbb7efe4be08d9e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 94, + 653, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6189a670508b95ed1025cb971c9ddd42dbb984fb377b6fc5a43a5d7dd4ab5582.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 154, + 655, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/968734f81b7e044425f2d3aea4334f75bf3c4adf3ff8b1d33a04cea323a73c60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 218, + 655, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64b47061a7b0198615a75e730e7f84ce69950cf9efb2f4b20c7cb971a74adf76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 277, + 655, + 351 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fcad1cbfccd8ee84f038800b43cec1a7098abe02777ea774fdcedf1bcf928f19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 95, + 728, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab057e6559b68066b739823a27bebbc8a93c4b6da9d715d3da7084fb6b6c36e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 154, + 727, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89b56d0c9b3ce24e61e652a14889f8ad2792d208a66186b9bf52092f46252f9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 282, + 727, + 351 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f35163338ec24ef83196f048b9938eb7c5a7fee3a953c5524ecbfe7f60341a5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 94, + 813, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0b328167017e295394af5f8583b28f953da9109f63f0afcc647f41a65b56cb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 154, + 813, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 741, + 280, + 813, + 351 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c8bd52f57a6a08848e792a86b340e9ee6692973aea46f81abd90a5b865b459ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 97, + 885, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/498c583f98d71f172e17400a4161c569ddd71d608e2fcc1cf005ae3ef6b6d5c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 830, + 154, + 885, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e19db98fc810890fb3cfb9a0fc8eb2a7c9e828bbc8f463b1db2159a35e3e375.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 280, + 813, + 351 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e90b1585752b80c7af0b278fc5849e11e371d1a91ca3cf429f7b7e5003238ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 830, + 282, + 885, + 351 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3e51179f7b70427e053a041cf10081c080215e4fca26daa97fa8bc1bc88a19f.jpg", + "image_caption": [ + "Figure 5. InteractVLM's reliance on 3D annotations. We evaluate performance for \"binary human contact\" (F1 score, Y-axis) for models trained on a varying percentage of DAMON [58] training data (X-axis). The DECO baseline trains on $100\\%$ of DAMON. Instead, InteractVLM trains on a varying (smaller) portion of this dataset. Yet, it achieves a significantly higher performance, by leveraging the broad visual knowledge of foundation models." + ], + "image_footnote": [], + "bbox": [ + 106, + 405, + 468, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "only $5\\%$ of the data, InteractVLM surpasses DECO, reaching an F1 score of 0.58. This performance gap widens as training data increases, ultimately achieving 0.75 F1 score with $100\\%$ of the data. This is a compelling finding, highlighting InteractVLM's efficient use of 3D supervision by leveraging the rich visual understanding of foundation models. Note that the strong performance with limited training data has high practical value, because obtaining 3D annotations is expensive.", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablations. For a study on the influence of InteractVLM components, including mask resolution, MV-Loc variants, loss functions, training data, the influence of the VLM, and the effect of text prompts, please refer to Sup. Mat.", + "bbox": [ + 511, + 412, + 906, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Joint Human-Object Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 483, + 843, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate the usefulness of 3D contacts inferred by InteractVLM for reconstructing a 3D human and object in interaction from a single in-the-wild image, $I$ .", + "bbox": [ + 511, + 508, + 905, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Initializing 3D body pose & shape, object shape. We use OSX [39] to estimate a 3D SMPL-X body mesh, $H$ , and OpenShape [42] to retrieve a 3D object mesh, $O$ , from the Objaverse [11] database that best matches the image.", + "bbox": [ + 511, + 554, + 905, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Initializing 3D object pose. We apply InteractVLM on image $I$ to predict 3D body and object contact vertices, $C^H, C^O$ . Then, we solve for object pose, $\\{R^O, t^O\\}$ by applying the ICP [6] algorithm to snap the 3D object contact points, $C^O$ , onto body ones, $C^H$ . To avoid false correspondences, the 3D normals of contact points must be compatible, i.e., should have similar angles but opposite directions.", + "bbox": [ + 511, + 614, + 905, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optimizing 3D object pose. Given the above initialization, we optimize over object rotation, $R^O$ , translation, $t^O$ , and scale, $s^O$ , via render-and-compare by minimizing:", + "bbox": [ + 511, + 720, + 905, + 766 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nE = E _ {M} + \\lambda_ {C} E _ {C}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 772, + 903, + 787 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nE _ {M} = I o U (\\widehat {M}, M) + \\left\\| \\widehat {M ^ {c}} - M ^ {c} \\right\\| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 792, + 903, + 810 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nE _ {C} = \\frac {1}{| C ^ {H} \\| C ^ {O} |} \\sum_ {i \\in | H |} \\sum_ {j \\in | O |} C _ {i} ^ {H} C _ {j} ^ {O} \\| V _ {i} ^ {H} - V _ {j} ^ {O} \\| _ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 813, + 903, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $E_{M}$ is a mask loss, $E_{C}$ is a contact loss, $IoU$ is intersection-over-union, $\\widehat{M}, M$ are hypothesis and ground-truth masks, respectively, $\\widehat{M^{c}}, M^{c}$ are the corresponding", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22611", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23806b3f4592cf7a9f13aacbf01bfdd8d1d1b5e0c1c87280ad3fb8c664973e5b.jpg", + "image_caption": [ + "Figure 6. 3D HOI reconstruction (Sec. 5). We build an optimization method that fits a SMPL-X body and OpenShape-retrieved object to an in-the-wild image. We evaluate against the SotA method PHOSA [67]. Reconstruction is guided by InteractVLM-inferred contacts." + ], + "image_footnote": [], + "bbox": [ + 96, + 66, + 893, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mean mask pixels, $|C^H|$ , $|C^O|$ are the number of contact vertices on the human and object, $|H|$ , $|O|$ are the number of mesh vertices, $V_i^H$ , $V_j^O$ are the $i$ -th and $j$ -th human and object vertices, while $C_i^O$ , $C_j^H$ indicate whether vertices are in contact or not. Given the image, we extract the object mask, $M$ , via SAM [31], and a depth map, $D$ , via Depth Pro [8].", + "bbox": [ + 89, + 512, + 482, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Intuitively, we use the $E_{M}$ loss to align the 3D object to the image, while in $E_{C}$ the predicted 3D contacts \"anchor\" the 3D object onto the body so they interact realistically. We perform iterative optimization via Adam [30]. OSX produces reasonable bodies, so we keep these fixed. The object is updated in each iteration - we render depth, $\\widehat{D}$ , and masks, $\\widehat{M}$ , via a PyTorch3D [52] differentiable renderer.", + "bbox": [ + 89, + 609, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. We reconstruct 3D human-object interaction from an image. We show results in Fig. 6, and compare with PHOSA [67], the most related SotA method. InteractVLM's reconstructions look more realistic. Note that PHOSA uses handcrafted contacts on humans and bodies. Instead, InteractVLM infers 3D contacts on both human bodies and objects from the single in-the-wild image. These play a crucial role for guiding 3D reconstruction under occlusions and depth ambiguities.", + "bbox": [ + 89, + 715, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Perceptual Study. There exists no in-the-wild dataset with 3D ground truth for HOI, so we conduct a perceptual study via Amazon Mechanical Turk for evaluation. Specifi", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cally, we evaluate the realism of our reconstructions against ones of PHOSA. We randomly select 55 images for which PHOSA has handcrafted contact annotations. For each image, participants are shown (with random swapping) reconstructions generated by our method and by PHOSA, and are asked to select the one that best represents the image. Our reconstructions are preferred $62\\%$ of the time.", + "bbox": [ + 511, + 513, + 906, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We develop InteractVLM, a novel method for estimating 3D contacts on both humans and objects from a single natural image. InteractVLM has a reduced reliance on expensive 3D contact annotations for training, by leveraging the broad knowledge of Vision-Language Models. Specifically, we introduce a novel \"Render-Localize-Lift\" (RLL) framework and a novel multi-view localization model (MV-Loc) to adapt 2D foundation models for 3D contact estimation. We outperform existing work on contact estimation and introduce a new \"Semantic Human Contact\" estimation task for inferring body contacts conditioned on object labels. This goes beyond traditional binary contact estimation, which fails to capture rich semantic relationships of multi-object interactions. Last, we develop the first approach that uses inferred contact points on both bodies and objects for joint 3D reconstruction from single in-the-wild images.", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22612", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: We thank Alpár Cseke for his assistance with evaluating joint human-object reconstruction. We also thank Tsvetelina Alexiadis and Taylor Obersat for MTurk evaluation, Yao Feng, Peter Kultis and Markos Diomataris for their valuable feedback and Benjamin Pellkofer for IT support. SKD is supported by the International Max Planck Research School for Intelligent Systems (IMPRS-IS). The UvA part of the team is supported by an ERC Starting Grant (STRIPES, 101165317, PI: D. Tzionas). Disclosure: DT has received a research gift fund from Google. For MJB see https://files.is.tue.mpg.de/black/CoI_CVPR_2025.txt", + "bbox": [ + 89, + 90, + 485, + 242 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 270, + 187, + 286 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Emanuele Aiello, Diego Valsesia, and Enrico Magli. Cross-modal learning for image-guided point cloud shape completion. Conference on Neural Information Processing Systems (NeurIPS), 35:37349-37362, 2022. 6", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: A visual language model for few-shot learning. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 3", + "[3] Reza Yazdani Aminabadi, Samyam Rajbhandari, Minjia Zhang, A. A. Awan, Cheng Li, Du Li, Elton Zheng, Jeff Rasley, Shaden Smith, Olatunj Ruwase, and Yuxiong He. DeepSpeed inference: Enabling efficient inference of transformer models at unprecedented scale. International Conference for High Performance Computing, Networking, Storage and Analysis, 2022. 5", + "[4] Dimitrije Antić, Georgios Paschalidis, Shashank Tripathi, Theo Gevers, Sai Kumar Dwivedi, and Dimitrios Tzionas. SDFit: 3D object pose and shape by fitting a morphable SDF to a single image. arXiv:2409.16178, 2025. 3", + "[5] Fabien Baradel*, Matthieu Armando, Salma Galaoui, Romain Brégier, Philippe Weinzaepfel, Grégory Rogez, and Thomas Lucas*. Multi-HMR: Multi-person whole-body human mesh recovery in a single shot. In European Conference on Computer Vision (ECCV), 2024. 3", + "[6] Paul J. Besl and Neil D. McKay. A method for registration of 3D shapes. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 14:239-256, 1992. 7", + "[7] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAVE: Dataset and method for tracking human object interactions. In Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[8] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth Pro: Sharp monocular metric depth in less than a second. arXiv:2410.02073, 2024. 8" + ], + "bbox": [ + 99, + 295, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision (ECCV), 2016. 3", + "[10] Honghua Chen, Zeyong Wei, Yabin Xu, Mingqiang Wei, and Jun Wang. ImLoveNet: Misaligned image-supported registration network for low-overlap point cloud pairs. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-9, 2022. 6", + "[11] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2023. 2, 3, 4, 7", + "[12] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3D AffordanceNet: A benchmark for visual object affordance understanding. In Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[13] John Doe, Jane Smith, and Alice Brown. Paris3d: Language-guided 3d segmentation with multimodal models. arXiv:2406.08394, 2024. 3", + "[14] Sai Kumar Dwivedi, Nikos Athanasiou, Muhammed Kocabas, and Michael J. Black. Learning to regress bodies from images using differentiable semantic rendering. In International Conference on Computer Vision (ICCV), 2021. 3", + "[15] Sai Kumar Dwivedi, Cordelia Schmid, Hongwei Yi, Michael J. Black, and Dimitrios Tzionas. POCO: 3D pose and shape estimation using confidence. In International Conference on 3D Vision (3DV), 2024. 3", + "[16] Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, Yao Feng, and Michael J. Black. TokenHMR: Advancing human mesh recovery with a tokenized pose representation. In Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[17] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual handobject manipulation. In Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[18] Zicong Fan, Maria Parelli, Maria Eleni Kadoglou, Muhammed Kocabas, Xu Chen, Michael J Black, and Otmar Hilliges. HOLD: Category-agnostic 3D reconstruction of interacting hands and objects from video. In Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[19] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Collaborative regression of expressive bodies using moderation. In International Conference on 3D Vision (3DV), 2021. 6", + "[20] Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, and Tae-Kyun Kim. First-person hand action benchmark with RGB-D videos and 3D hand pose annotations. In Computer Vision and Pattern Recognition (CVPR), 2018. 3", + "[21] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 3" + ], + "bbox": [ + 516, + 92, + 906, + 907 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22613", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambiguities with 3D scene constraints. In International Conference on Computer Vision (ICCV), pages 2282-2292, 2019. 3", + "[23] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14708-14718, 2021. 6", + "[24] Yana Hasson, Gül Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[25] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. arXiv: 2311.04400, 2023. 3", + "[26] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. International Conference on Learning Representations (ICLR), 2021. 2, 5", + "[27] Chun-Hao Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Computer Vision and Pattern Recognition (CVPR), pages 13264-13275, 2022. 6", + "[28] Yinghao Huang, Omid Taheri, Michael J. Black, and Dimitrios Tzionas. InterCap: Joint markerless 3D tracking of humans and objects in interaction from multi-view RGB-D images. International Journal of Computer Vision (IJCV), 132(7):2551-2566, 2024. 3", + "[29] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In International Conference on 3D Vision (3DV), pages 42-52, 2021. 3", + "[30] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 8", + "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloé Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 2, 3, 5, 8", + "[32] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. VIBE: Video inference for human body pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), pages 5252-5262, 2020. 3", + "[33] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 3", + "[34] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: Reasoning segmentation via large language model. In Computer Vision and Pattern Recognition (CVPR), pages 9579-9589, 2024. 3, 4, 5" + ], + "bbox": [ + 91, + 92, + 482, + 893 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. HybrIK: A hybrid analytical-neural inverse kinematics solution for 3D human pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), pages 3383-3393, 2021. 3", + "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning (ICML), pages 19730-19742, 2023. 3", + "[37] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In European Conference on Computer Vision (ECCV), 2022. 3", + "[38] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-resolution text-to-3D content creation. Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[39] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3D whole-body mesh recovery with component aware transformer. In Computer Vision and Pattern Recognition (CVPR), 2023. 3, 7", + "[40] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In International Conference on Computer Vision (ICCV), 2021. 3", + "[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 3, 5", + "[42] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. OpenShape: Scaling up 3D shape representation towards open-world understanding. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 2, 3, 4, 7", + "[43] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3D mesh in 45 seconds without pershape optimization. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 3", + "[44] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. International Conference on Computer Vision (ICCV), 2023. 3", + "[45] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. Transactions on Graphics (TOG), 34 (6):248:1-248:16, 2015. 3", + "[46] Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. COAP: Compositional articulated occupancy of people. In Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[47] Hyeongjin Nam, Daniel Sungho Jung, Gyeongsik Moon, and Young Mu Lee. Joint reconstruction of 3D human and object via contact-based refinement transformer. In Computer Vision and Pattern Recognition (CVPR), 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 868 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22614", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Mohamed Omran, Christoph Lassner, Gerard Pons-Moll, Peter Gehler, and Bernt Schiele. Neural body fitting: Unifying deep learning and model based human pose and shape estimation. In International Conference on 3D Vision (3DV), 2018. 3", + "[49] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975–10985, 2019. 3", + "[50] Bingbing Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824, 2023. 3", + "[51] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. In International Conference on Learning Representations (ICLR), 2024. 3", + "[52] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D deep learning with PyTorch3D. arXiv:2007.08501, 2020. 8", + "[53] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. Transactions on Graphics (TOG), 36(6), 2022, 2, 4", + "[54] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[55] Qingping Sun, Yanjun Wang, Ailing Zeng, Wanqi Yin, Chen Wei, Wenjia Wang, Haiyi Mei, Chi Sing Leung, Ziwei Liu, Lei Yang, and Zhongang Cai. AiOS: All-in-one-stage expressive human pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[56] István Sárándi and Gerard Pons-Moll. Neural localizer fields for continuous 3D human pose and shape estimation. In Conference on Neural Information Processing Systems (NeurIPS), 2024. 3", + "[57] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 3", + "[58] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J. Black. DECO: Dense estimation of 3D human-scene contact in the wild. In International Conference on Computer Vision (ICCV), pages 8001-8013, 2023. 2, 3, 5, 6, 7", + "[59] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object" + ], + "bbox": [ + 91, + 92, + 480, + 828 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "coordinate space for category-level 6D object pose and size estimation. In Computer Vision and Pattern Recognition (CVPR), 2019. 5", + "[60] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, and Jifeng Dai. VisionLLM: Large language model is also an open-ended decoder for vision-centric tasks. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 3", + "[61] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. CHORE: Contact, human and object reconstruction from a single RGB image. In European Conference on Computer Vision (ECCV), 2022. 3", + "[62] Xianghui Xie, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Template free reconstruction of human-object interaction with procedural interaction generation. In Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[63] Danfei Xu, Dragomir Anguelov, and Ashesh Jain. PointFusion: Deep sensor fusion for 3D bounding box estimation. In International Conference on Computer Vision (ICCV), pages 244-253, 2018. 6", + "[64] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, Bill Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[65] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3D object affordance from 2D interactions in images. In International Conference on Computer Vision (ICCV), pages 10871-10881, 2023. 2, 3, 4, 5, 6", + "[66] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, and Zheng-Jun Zha. LEMON: Learning 3D human-object interaction relation from 2D images. In Computer Vision and Pattern Recognition (CVPR), 2024. 3, 5, 6", + "[67] Jason Y. Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3D human-object spatial arrangements from a single image in the wild. In European Conference on Computer Vision (ECCV), 2020. 3, 8", + "[68] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. GPT4RoI: Instruction tuning large language model on region-of-interest. arXiv:2307.03601, 2023. 3", + "[69] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3D: Exploring unified 3D representation at scale. In International Conference on Learning Representations (ICLR), 2024. 3", + "[70] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3D lidar semantic segmentation. In International Conference on Computer Vision (ICCV), 2021. 6" + ], + "bbox": [ + 516, + 92, + 903, + 823 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22615", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/9f03277a-37b8-49e7-9fad-13e2e5f171b2_model.json b/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/9f03277a-37b8-49e7-9fad-13e2e5f171b2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bc618b409221e89d5567323e0c90847d0ae2502f --- /dev/null +++ b/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/9f03277a-37b8-49e7-9fad-13e2e5f171b2_model.json @@ -0,0 +1,2807 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.118, + 0.861, + 0.14 + ], + "angle": 0, + "content": "InteractVLM: 3D Interaction Reasoning from 2D Foundational Models" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.161, + 0.807, + 0.198 + ], + "angle": 0, + "content": "Sai Kumar Dwivedi1 Dimitrije Antić2 Shashank Tripathi1 Omid Taheri1 Cordelia Schmid3 Michael J. Black1 Dimitrios Tzionas2" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.209, + 0.895, + 0.245 + ], + "angle": 0, + "content": "1Max Planck Institute for Intelligent Systems, Tübingen, Germany 2University of Amsterdam (UvA), the Netherlands 3Inria, École normale supérieure, France" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.264, + 0.896, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.457, + 0.908, + 0.501 + ], + "angle": 0, + "content": "Figure 1. We present InteractVLM, a novel method for estimating contact points on both human bodies and objects from a single in-the-wild image, shown here as red patches. Our method goes beyond traditional binary contact estimation methods by estimating contact points on a human in relation to a specified object. We do so by leveraging the broad visual knowledge of a large Visual Language Model." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.519, + 0.328, + 0.535 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We introduce InteractVLM, a novel method to estimate 3D contact points on human bodies and objects from single in-the-wild images, enabling accurate human-object joint reconstruction in 3D. This is challenging due to occlusions, depth ambiguities, and widely varying object shapes. Existing methods rely on 3D contact annotations collected via expensive motion-capture systems or tedious manual labeling, limiting scalability and generalization. To overcome this, InteractVLM harnesses the broad visual knowledge of large Vision-Language Models (VLMs), fine-tuned with limited 3D contact data. However, directly applying these models is non-trivial, as they reason \"only\" in 2D, while human-object contact is inherently 3D. Thus we introduce a novel \"Render-Localize-Lift\" module that: (1) embeds 3D body and object surfaces in 2D space via multiview rendering, (2) trains a novel multi-view localization model (MV-Loc) to infer contacts in 2D, and (3) lifts these to 3D. Additionally, we propose a new task called Semantic Human Contact estimation, where human contact predictions are conditioned explicitly on object semantics, enabling richer interaction modeling. InteractVLM outperforms existing work on contact estimation and also facilitates 3D reconstruction from an in-the-wild image. To es" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.909, + 0.612 + ], + "angle": 0, + "content": "timate 3D human and object pose, we infer initial body and object meshes, then infer contacts on both of these via InteractVLM, and last exploit these for fitting the meshes to image evidence. Results show that our approach performs promisingly in the wild. Code and models are available at https://interactvlm.is.tue.mpg.de." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.645, + 0.646, + 0.661 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.748 + ], + "angle": 0, + "content": "People interact with objects routinely. Reconstructing human-object interaction (HOI) in 3D is key for many applications, from robots to mixed reality. However, doing so from single images is challenging due to depth ambiguity, occlusions, and the diverse shape and appearance of objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "There are methods that estimate 3D human bodies and methods that estimate 3D objects, but few that put these together. Knowing the contacts between these could significantly improve joint reconstruction. Our goal is to infer contact points on both humans and objects from single in-the-wild images, and then use these contacts for jointly reconstructing humans and objects. However, there is a lack of in-the-wild training images paired with ground-truth contact labels for both 3D humans and objects. Acquiring such data is challenging and existing methods do not scale." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22605" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "The problem gets more challenging due to the complexity of real-world interactions. Humans often contact multiple objects simultaneously; e.g., using a laptop while sitting on a club chair. Yet, current approaches treat contact prediction as simple binary classification; i.e. detecting whether a body part is in contact with \"any\" object. This simplified assumption fails to capture the rich semantic relationships of multi-object interactions. To address this, we introduce a novel \"Semantic Human Contact\" estimation task. This involves predicting the contact points on the body related to a particular object given a single in-the-wild image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.41 + ], + "angle": 0, + "content": "To tackle this, as well as to overcome data scarcity, we propose a new paradigm for scaling and reasoning in the wild. Specifically, we observe that large Vision-Language Models (VLMs) can \"reason\" about in-the-wild images because they are trained on internet-scale data and possess a broad visual knowledge about humans and their interactions with the world. We also observe that this knowledge can be re-purposed for novel tasks by fine-tuning these large models on small datasets. Thus, we exploit VLMs for developing a novel framework, called InteractVLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.483, + 0.502 + ], + "angle": 0, + "content": "At the heart of InteractVLM lies a reasoning module based on a VLM; see Fig. 2. We enrich this with skills for 3D human-object \"understanding,\" by extending the VLM with a LoRA [26] adaptation layer. As a result, given a color image, this module can be \"asked\" to produce \"reasoning tokens\" that facilitate 3D contact localization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.483, + 0.625 + ], + "angle": 0, + "content": "However, exploiting these tokens to localize contact is non-trivial. A natural choice would be to employ a foundational \"localization\" model [31] that takes these tokens as \"guidance,\" and highlights the 3D contacts. But there exists a key practical problem. Existing foundation models inherently operate only in 2D space, while we need them to do so in 3D. To tackle this, we need to re-cast our problem so it is appropriate for 2D foundation models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.779 + ], + "angle": 0, + "content": "To this end, we develop a novel \"Render-Localize-Lift\" (RLL) framework that has three main steps; see Fig. 2: (1) We render the 3D shape of a canonical SMPL+H [53] body and an object as 2D images from multiple viewpoints. Regarding the object, a 3D mesh is efficiently retrieved from a large-scale 3D database [11] through OpenShape [42]. (2) We pass the above images into a foundation model to predict 2D contact masks for both the body and the object. (3) We lift the predicted 2D contact points to 3D points via back-projection, i.e., performing the inverse of the first step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.9 + ], + "angle": 0, + "content": "However, even after recasting the problem to 2D by rendering multi-view images, existing foundation models are still not 3D aware, i.e. they treat each view independently, ignoring multi-view consistency. This means contact detections in one view do not necessarily agree with ones in adjacent views. To tackle this, just appending camera parameters to multi-view renderings is insufficient. Instead, we build a novel \"Multi-view Localization\" model that we call" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.093, + 0.897, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.229, + 0.905, + 0.285 + ], + "angle": 0, + "content": "Figure 2. Overview of InteractVLM. Given a color image, our VLM performs the core reasoning, and guides a novel MV-Loc model to localize contacts on both bodies and objects in 3D. Here we show only the body; for details, and object contact, see Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.905, + 0.387 + ], + "angle": 0, + "content": "MV-Loc. This has two steps: (1) It transforms the \"reasoning token\" provided by the VLM with the camera parameters used to render the multi-view images. (2) It ensures multi-view consistency, by lifting the inferred 2D contacts in each view to 3D and computing a 3D loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.905, + 0.538 + ], + "angle": 0, + "content": "Our method, InteractVLM, utilizes a VLM in tandem with our novel multi-view localization model, MV-Loc, to perform 3D contact prediction for humans and objects. See Fig. 1 for some examples. We quantitatively evaluate the efficacy of our method for in-the-wild 3D contact prediction on bodies and on objects, using the DAMON [58] and PIAD [65] datasets, respectively. For bodies, we evaluate both for the traditional \"binary contact\" estimation, and for our new task of \"semantic contact\" estimation. For all tasks, we find that InteractVLM outperforms the prior work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.689 + ], + "angle": 0, + "content": "Finally, we demonstrate how InteractVLM's estimated contact improves 3D HOI recovery from in-the-wild images; this is a highly ill-posed task due to depth ambiguities and occlusions. To address this, we develop an optimization-based method that fits a SMPL-X body mesh and an OpenShape-retrieved object mesh to the image, using InteractVLM's inferred contacts as constraints to anchor human and object meshes w.r.t. each other. To the best of our knowledge, this is the first approach for estimating 3D HOI for in-the-wild images using inferred contacts." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.69, + 0.835, + 0.704 + ], + "angle": 0, + "content": "In summary, we make four key contributions:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.705, + 0.905, + 0.749 + ], + "angle": 0, + "content": "1. We build InteractVLM, a novel method that facilitates HOI reconstruction from an in-the-wild image by detecting 3D contacts on both bodies and objects." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.75, + 0.905, + 0.794 + ], + "angle": 0, + "content": "2. We demonstrate a way to minimize reliance on 3D contact annotations via exploiting the broad visual knowledge of Vision-Language Models." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.796, + 0.905, + 0.839 + ], + "angle": 0, + "content": "3. We build a novel \"Multi-view Localization\" model that helps in estimating contacts in 3D by transforming the reasoning of foundation models from 2D to 3D." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "4. We introduce the novel \"Semantic Human Contact\" task for inferring body contacts conditioned on object labels. Our code and trained models are available for research at https://interactglm.is.tue.mpg.de." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22606" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.234, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.368, + 0.132 + ], + "angle": 0, + "content": "2.1. Large Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.483, + 0.395 + ], + "angle": 0, + "content": "Recent advancements in large language models (LLMs) have led to the development of multimodal models that integrate vision and language reasoning. Models like Flamingo [2] and BLIP-2 [36] use cross-attention mechanisms and visual encoders to align image features with text, supporting a variety of vision-language tasks. More recent works, such as VisionLLM [60] and Kosmos-2 [50], use grounded image-text data to enhance spatial understanding, while GPT4RoI [68] introduces spatial box inputs for finer alignment. However, these models typically lack end-to-end segmentation capabilities. To address this limitation, LISA [34] combines vision foundation segmentation models like SAM [31] with multimodal embeddings, enabling language-guided segmentation. PARIS3D [13] extends this approach to referential 3D segmentation by processing multi-view object renders through both SAM and LLaVA [41] for spatially-aware segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.483, + 0.577 + ], + "angle": 0, + "content": "Taking inspiration from these approaches, we exploit a language-guided segmentation model for the task of predicting human and object contact in 3D. However, unlike PARIS3D, we process a single RGB image with LLaVA and multi-view renders of the human mesh and the object mesh with SAM. Moreover, we introduce a feature-lifting technique that extends LLaVA's 2D features into 3D using camera parameters, thus guiding SAM's multi-view segmentations. This approach ensures multi-view consistency and efficiently predicts contact affordances, extending the use of multimodal models in 3D reasoning tasks focused on human-object interaction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.59, + 0.455, + 0.607 + ], + "angle": 0, + "content": "2.2. 3D Human and Object from Single Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Estimating 3D human pose and shape from a single image has evolved from optimization-based methods to learning-based approaches. Optimization-based methods fit parametric body models like SMPL [45], SMPL-X [49], or GHUM [64] to 2D cues such as keypoints [9], silhouettes [48], or segmentation masks [48]. Learning-based methods either regress body parameters from images or videos [5, 14, 15, 29, 32, 33, 37] or estimate non-parametric bodies as vertices [35, 40], implicit surfaces [46, 54], or dense points [56]. Transformer-based methods [16, 21, 39, 55] have further improved robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For 3D object reconstruction from a single image, regression-based methods predict geometry using meshes, voxels, or point clouds. Diffusion-based models [25] utilize large 3D datasets like Objaverse [11] or 2D diffusion models [38, 43, 44, 51] to guide reconstruction and optimization methods [4] use render and compare. Retrieval-based methods, such as OpenShape [42] and Uni3D [69], have demonstrated some robustness in cases with occlusions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.783, + 0.108 + ], + "angle": 0, + "content": "2.3. 3D Human-Object Interaction" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.905, + 0.295 + ], + "angle": 0, + "content": "Understanding 3D human-object interactions is essential for modeling realistic scenes. Early works focused on handobject interactions, such as ObMan [24] and FPHA [20], with more recent studies like ARCTIC [17] and HOLD [18] providing more detailed data and reconstruction for hands. For full-body interactions, initial studies involve interactions with scenes, as in PROX [22], and with objects, as in BEHAVE [7], GRAB [57], and InterCap [28]. BEHAVE and GRAB use motion capture that are accurate but nonscalable, while InterCap uses multi-camera setups that are more scalable but less accurate. Both of these approaches are limited in capturing diverse and realistic interactions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.295, + 0.905, + 0.612 + ], + "angle": 0, + "content": "As a proxy for 3D reconstruction, the recent DECO [58] method infers contacts on body meshes; to this end, 3D contact annotations are crowdsourced via mesh painting. Predicting contact on objects is more challenging due to varying object shapes, while also currently there exists no dataset of object contact for in-the-wild images. Thus, we estimate 3D object affordances as a proxy for contact. 3D-AffordanceNet [12] introduces affordances that are not grounded in images, capturing the likelihood of humans interacting with specific parts of an object for a given affordance (e.g., \"sit on chair\"). PIAD [65] curates RGB images depicting object affordances and trains a network to estimate them. LEMON [66] extends object affordance prediction of PIAD to include human contact estimation. However, these methods require human contact vertices paired with object affordances for training, limiting the number of categories they handle to 21 ones. In contrast, we learn from unpaired human and object interaction data and perform joint interaction reasoning via a cross-attention layer, enabling human interaction reasoning for 80 categories and object affordance prediction for 32 categories." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.622, + 0.857, + 0.639 + ], + "angle": 0, + "content": "2.4. Joint 3D Human-Object Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.644, + 0.905, + 0.825 + ], + "angle": 0, + "content": "Joint reconstruction of humans and objects in 3D has been tackled with both regression and optimization methods. Regression-based methods directly predict 3D human-object meshes, as in HDM [62] and CONTHO [47], while other methods first predict contact points and then use test-time optimization to fit human and object meshes, as in CHORE [61] and PHOSA [67]. Since regression methods rely on limited training data, optimization methods are preferred for in-the-wild scenarios, such as PHOSA [67]. Optimization-based methods either assume known contacts or infer contacts to fit meshes to the image, but their success heavily relies on the quality of contact." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our method improves upon these by providing more accurate contact predictions, which in turn facilitate better fitting of human and object meshes to image evidence, improving the realism and accuracy of 3D human-object reconstructions from single images." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22607" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.087, + 0.729, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.089, + 0.895, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.345, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Figure 3. Method overview. Given a single in-the-wild color image, our novel InteractVLM method estimates 3D contact points on both humans and objects (a). Then, we reconstruct a 3D human and object in interaction by exploiting these contacts (b). More specifically: (a) Contact estimation. Given an image, \\( I \\), and prompt text, \\( T_{inp} \\), our VLM, \\( \\Psi \\), produces contact tokens for humans and objects, \\( <\\mathsf{HCON}> \\) and \\( <0\\mathsf{CON}> \\), which are projected (\\( \\Gamma \\)) into feature embeddings, \\( E^H \\) and \\( E^O \\). These guide a \"Multi-View [contact] Localization\" model. This renders the 3D human and object geometry via cameras, \\( K \\), into multi-view 2D renders and passes these to encoder, \\( \\Theta \\), while decoders, \\( \\Omega^H \\), \\( \\Omega^O \\), estimate and highlight 2D contacts in these renders. Then, the FeatLift module, \\( \\Phi \\), transforms the VLM's features \\( (E^H, E^O) \\) to become 3D-aware \\( (E_{3D}^H, E_{3D}^O) \\) by exploiting the camera parameters, \\( K \\). A final module lifts the detected 2D contacts to 3D. (b) 3D HOI reconstruction. For joint human-object reconstruction, we use InteractVLM's inferred contacts in an optimization framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.481, + 0.182, + 0.496 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.507, + 0.292, + 0.523 + ], + "angle": 0, + "content": "3.1. Input Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.483, + 0.56 + ], + "angle": 0, + "content": "Given an image, \\( I \\in \\mathbb{R}^{H \\times W \\times 3} \\), InteractVLM estimates 3D contacts for both human bodies and objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.483, + 0.623 + ], + "angle": 0, + "content": "The human is represented by a SMPL+H [53] 3D body mesh, \\( H \\), with vertices \\( V^{H} \\in \\mathbb{R}^{10475 \\times 3} \\). The body is posed in a canonical \"star\" shape (see details in Sec. 3.4). The human contacts are binary per-vertex labels, \\( C^{H} \\in \\{0,1\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.79 + ], + "angle": 0, + "content": "The object is represented by a 3D point cloud (or mesh), \\( O \\in \\mathbb{R}^{N \\times 3} \\), with \\( N \\) points. As there are no datasets of natural images paired with 3D contacts for objects, we use a largescale 3D affordance dataset [65], instead, as affordances are closely related to contact. Specifically, they represent the likelihood of contact on 3D object areas for various purposes. Therefore, for objects we use the terms \"affordance\" and \"contact\" interchangeably. The object contacts are continuous per-point values, \\( C^O \\in [0,1] \\). During inference, we retrieve a 3D object shape from a large database [11] via OpenShape [42] conditioned on image \\( I \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.329, + 0.817 + ], + "angle": 0, + "content": "3.2. Overview of InteractVLM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The biggest challenge for learning 3D contact prediction in the wild is the limited 3D contact data for humans and objects. To go beyond existing limited datasets, we introduce a novel method, called InteractVLM, that harnesses the commonsense knowledge of large vision-language models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.907, + 0.605 + ], + "angle": 0, + "content": "Specifically, InteractVLM (Fig. 3) has two main components: a Vision Language Model (VLM) and a novel Multi-View contact Localization model (MV-Loc). MV-Loc highlights parts that are in contact for both humans and objects with the VLM's guidance. The input to the VLM (Sec. 3.3) is an image, \\(I\\), and a prompt, \\(T_{inp}\\), that asks the VLM to detect contact. The input to MV-Loc (Sec. 3.4) is the 3D geometry of humans and objects, \\(H\\) and \\(O\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.631, + 0.833, + 0.647 + ], + "angle": 0, + "content": "3.3. Interaction Reasoning through VLM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.75 + ], + "angle": 0, + "content": "The VLM module, \\(\\Psi\\), conducts the core interaction reasoning. It takes an input image, \\(I\\), and prompt text, \\(T_{inp}\\), and outputs a prompt text, \\(T_{out} = \\Psi(I, T_{inp})\\). Inspired by the recent LISA [34] model, we expand the VLM's vocabulary with two specialized tokens, \\(\\langle \\text{HCON} \\rangle\\) and \\(\\langle \\text{OCON} \\rangle\\), for contact information for humans and objects, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.755, + 0.907, + 0.862 + ], + "angle": 0, + "content": "To denote contact, \\(\\Psi\\) produces a prompt that includes the above tokens. To aid MV-Loc in localizing contact, we extract the last-layer embeddings of the VLM corresponding to these tokens and pass them through a projection layer, \\(\\Gamma\\), to obtain the feature embeddings, \\(E^H\\) or \\(E^O\\). Let \\(T_{\\mathrm{gt}}\\) be the ground-truth text, and \\(T_{\\mathrm{pred}}\\) be the predicted one. Then, our token-prediction loss is defined as a cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.884, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o k e n}} = - \\sum_ {i = 1} ^ {N} \\left(T _ {\\mathrm {g t}} ^ {(i)} \\cdot \\log \\left(T _ {\\text {p r e d}} ^ {(i)}\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22608" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.446, + 0.108 + ], + "angle": 0, + "content": "3.4. Interaction Localization through MV-Loc" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.219 + ], + "angle": 0, + "content": "We develop a novel MV-Loc module that has a shared image encoder, \\(\\Theta\\), and separate decoders, \\(\\Omega^H\\) and \\(\\Omega^O\\), for humans and objects. MV-Loc performs contact localization by using a novel \"Render-Localize-Lift\" (RLL) framework. To this end, it takes three steps; (1) rendering the 3D shape of humans and objects in 2D, (2) predicting 2D contact maps for both of these, and (3) lifting the 2D contact maps to 3D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.22, + 0.484, + 0.416 + ], + "angle": 0, + "content": "RLL step #1: Render 3D \\(\\rightarrow\\) 2D. The input is human and object geometry, namely \\(\\mathcal{H}\\) and \\(O\\), respectively. Both serve as a \"canvas\" for \"painting\" detected contacts on these. The body has a default SMPL+H shape in a canonical star-shape pose to minimize self-occlusions when rendering. The object geometry (initialized in Sec. 3.1) is normalized to a unit sphere. Each geometry is rendered from \\(J\\) fixed views with camera parameters, \\(K\\), to form multi-view renderings, \\(R^{H,O} = \\{R_j\\}_{j=1}^J\\), so that the entire 3D geometry is captured. Since our geometries do not have texture, we color the meshes with normals and point clouds using the NOCS map [59]. This enhances cross-view correspondence, making renderings resemble real images to image encoder, \\(\\Theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.522 + ], + "angle": 0, + "content": "RLL step #2: Localize in 2D. The rendered geometry, \\( R^{H,O} \\), is first sent to the image encoder, \\( \\Theta \\), and then passed to decoders, so that the final contact masks, \\( M^H \\) and \\( M^O \\), get highlighted on it. However, MV-Loc requires spatial and contextual cues for highlighting the contact region. To this end, we use the feature embeddings (Sec. 3.3), i.e. \\( E^H \\) and \\( E^O \\) to guide the contact localization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.599 + ], + "angle": 0, + "content": "However, since the VLM reasons in 2D, these features are not 3D aware. This is a problem, because MV-Loc needs 3D awareness to localize contact consistently across multi-view renderings. Thus, we transform the features to \"lift\" them to \"3D\" to better guide multi-view localization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.647 + ], + "angle": 0, + "content": "In detail, we design a lifting network, \\(\\Phi\\), which takes camera parameters, \\(K\\), and the 2D \\(E^{H,O}\\), and lifts the latter to 3D as \\(E_{3D}^{H,O} = \\Phi(E^{H,O}, K)\\). Contact masks are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.654, + 0.483, + 0.675 + ], + "angle": 0, + "content": "\\[\nM ^ {H, O} = \\Omega^ {H, O} (\\Theta (R ^ {H, O}), E _ {3 D} ^ {H, O}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.484, + 0.773 + ], + "angle": 0, + "content": "Below, the \\(H\\) and \\(O\\) superscripts are dropped for brevity. We calculate losses only on \"valid\" regions, i.e., the areas within the outline of rendered geometry; we denote this as \\(M\\). To encourage overlap between the predicted masks, \\(M\\), and the ground-truth ones, \\(\\widehat{M}\\), particularly in sparse contact regions, we use a focal-weighted BCE loss and a Dice loss:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.784, + 0.483, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B C E} = - \\alpha (1 - p _ {M}) ^ {\\gamma} \\log (p _ {M}) - (1 - \\alpha) p _ {M} ^ {\\gamma} \\log (1 - p _ {M}), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.812, + 0.483, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D i c e} = 1 - \\frac {2 \\sum M \\cdot \\widehat {M} + \\epsilon}{\\sum M + \\sum \\widehat {M} + \\epsilon}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( p_M \\) is the predicted mask probability, \\( \\alpha \\) controls class balance, \\( \\gamma \\) adjusts the focus on hard examples, and \\( \\epsilon \\) is a residual term to prevent division by zero." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.197 + ], + "angle": 0, + "content": "RLL step #3: Lift \\(2\\mathbf{D} \\rightarrow 3\\mathbf{D}\\). To lift the inferred 2D contact points to 3D points, we follow the inverse of step #1. Normally, 2D points backproject to 3D lines due to depth ambiguities. In our case the lines intersect with the known 3D geometry that produced the multi-view renders. So, 2D points are lifted to 3D points, and by extension 2D contact masks, \\(M^{H,O}\\), are lifted to 3D contact areas, \\(C^{H,O}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.905, + 0.258 + ], + "angle": 0, + "content": "We use a human contact loss, \\(\\mathcal{L}_C^H\\), that combines a focal loss with sparsity regularization, to encourage precise true-positive predictions in \"valid\" contact regions while discouraging false positives in non-contact areas:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.265, + 0.905, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C} ^ {H} = \\alpha (1 - p _ {h C}) ^ {\\gamma} \\log \\left(p _ {h C}\\right) + \\lambda \\| C ^ {H} \\| _ {1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.292, + 0.905, + 0.337 + ], + "angle": 0, + "content": "where \\(p_{hC}\\) is the contact probability, and \\(\\lambda\\), \\(\\alpha\\) and \\(\\gamma\\) are scalar weights. We also use an object contact loss, \\(\\mathcal{L}_C^O\\), that combines a Dice and Mean Squared Error (MSE) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.344, + 0.905, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C} ^ {O} = \\mathcal {L} _ {\\text {D i c e}} \\left(C ^ {O}, \\widehat {C} ^ {O}\\right) + \\beta \\| C ^ {O} - \\widehat {C} ^ {O} \\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.905, + 0.403 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a weighting factor, and \\(\\widehat{C^O}\\) denotes the groundtruth 3D contacts for objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.411, + 0.731, + 0.427 + ], + "angle": 0, + "content": "3.5. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.906, + 0.584 + ], + "angle": 0, + "content": "Architecture. We use LLaVA [41] as our VLM and SAM [31] for our MV-Loc, with weights pre-trained by LISA [34] for segmentation [34]. The feature-lifting network, \\(\\Phi\\), contains a spatial-understanding network (two fully-connected layers of size 128 with ReLU activation) followed by view-specific (256-dimensional) transformations and a sigmoid activation. For 3D contact prediction, our MV-Loc model converts 2D masks to 3D contact points via 2D-to-3D pixel-to-vertex mappings that are precomputed during MV-Loc's rendering step. For details, see Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Training. To efficiently fine-tune our VLM, we employ LoRA [26] with rank 8. The separate decoders for human and object contact prediction are trained without LoRA, while keeping the image encoder frozen. For training, we use DeepSpeed [3] with mixed precision training (bfloat16), and batch size of 8. We train on 4 Nvidia-A100 GPUs for 30 epochs. For more details, please refer to Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Datasets. We focus on two tasks, i.e., 3D human contact and 3D object affordance prediction, using two in-the-wild datasets, i.e., DAMON [58] and PIAD [65], respectively. For human contact we train and evaluate on DAMON [58]. For 3D object affordances, we train and evaluate on PIAD [65]. We find that exploiting textual descriptions for the contacting body parts, and the object type in contact, helps training. Similarly, adding Visual Question-Answering (VQA) data generated by GPT4o for the training images also helps. For details, see Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Unlike LEMON [66], which requires paired human-object geometry for training using the 3DIR dataset [66], we use unpaired data. This enables us to scale for many human and object categories not addressed by prior work. For" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22609" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.105, + 0.083, + 0.472, + 0.173 + ], + "angle": 0, + "content": "
MethodF1 (%)↑Precision (%)↑Recall (%)↑Geodesic (cm)↓
POSAPPIXIE [23] [19]31.042.034.033.00
BSTRO [27]46.051.053.038.06
DECO [58]55.065.057.021.32
InteractVLM75.675.276.02.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.226 + ], + "angle": 0, + "content": "Table 1. Evaluation for \"Binary Human Contact\" prediction on the DAMON dataset [58]. We compare our InteractVLM model (trained only for this task) with the state of the art." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.24, + 0.48, + 0.362 + ], + "angle": 0, + "content": "
Object CategoriesSemantic-DECO [58] (Baseline)InteractVLM
F1 (%)↑Prec. (%)↑Rec. (%)↑Geo (cm)↓F1 (%)↑Prec. (%)↑Rec. (%)↑Geo (cm)↓
Accessory40.130.775.621.8861.172.965.66.91
Daily Obj26.520.152.360.3468.671.478.07.46
Food11.719.412.949.6166.466.177.97.17
Furniture24.515.883.729.1760.564.260.16.21
Kitchen27.724.737.252.3471.871.381.17.61
Sports36.430.480.179.2177.976.783.27.98
Transport52.039.193.731.7877.880.578.97.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.414 + ], + "angle": 0, + "content": "Table 2. Evaluation for \"Semantic Human Contact\" prediction on the DAMON [58] dataset. For results on each class, see Sup. Mat. The \"Semantic-DECO\" baseline extends DECO for our new task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.483, + 0.488 + ], + "angle": 0, + "content": "the final joint human-object reconstruction task, we combine DAMON [58], PIAD [65], 3DIR [66] and all textual descriptions about body parts, contact type and HOI." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.484, + 0.595 + ], + "angle": 0, + "content": "Evaluation metrics. For human contact prediction, following Tripathi et al. [58], we report the F1, precision, and recall scores using a threshold of 0.5, and a geodesic distance measuring spatial accuracy. For object contact prediction, following Yang et al. [65], we report the Similarity (SIM), Mean Absolute Error (MAE), Area Under ROC Curve (AUC), and average Intersection over Union (IOU)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.61, + 0.225, + 0.627 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.416, + 0.651 + ], + "angle": 0, + "content": "4.1. \"Binary Human Contact\" Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.718 + ], + "angle": 0, + "content": "This task refers to estimating contact areas on the body via binary classification of its vertices, ignoring the number or type of objects involved. We train and evaluate on the DAMON [58] dataset, and report results in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.795 + ], + "angle": 0, + "content": "InteractVLM significantly outperforms all previous methods achieving a \\(20.6\\%\\) improvement in F1 score. Although, here, InteractVLM is trained on the same data as DECO, it goes beyond this by harnessing the commonsense knowledge of a large foundation model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In Sup. Mat. we also evaluate on the 3DIR [66] dataset and compare with the LEMON method [66]. Even though LEMON uses paired human-object data, InteractVLM performs on-par with it despite training on human-only data. We also evaluate InteractVLM's performance for binary human contact prediction across different body parts; for more details please refer to Sup. Mat." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.083, + 0.905, + 0.196 + ], + "angle": 0, + "content": "
MethodsPIAD-Seen [65]PIAD-Unseen [65]
SIM (%)↑AUC (%)↑aIOU (%)↑MAE ↓SIM (%)↑AUC (%)↑aIOU (%)↑MAE ↓
PMF [70]42.575.0510.131.4133.060.254.672.11
ILN [10]42.775.8411.521.3732.559.694.712.07
PFusion [63]43.277.5012.311.3533.061.875.331.93
XMF [1]44.178.2412.941.2734.262.585.681.88
IAGNet [65]54.584.8520.510.9835.271.847.951.27
InteractVLM62.786.4721.200.8141.475.458.500.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.205, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Table 3. Evaluation for \"Object Affordance Prediction\" on the PIAD [65] dataset. We compare our InteractVLM model (trained only for this task) with the state of the art." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.273, + 0.857, + 0.289 + ], + "angle": 0, + "content": "4.2. \"Semantic Human Contact\" Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.907, + 0.37 + ], + "angle": 0, + "content": "In real-life interactions, multiple objects can be contacted by different body areas concurrently. Thus, we introduce a novel task called \"Semantic Human Contact\" prediction. We evaluate on DAMON [58] and report results in Tab. 2; for a finer-grained version of this table, see Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.909, + 0.538 + ], + "angle": 0, + "content": "To establish a baseline, we adapt the DECO model [58] that detects binary contacts and turn it into a multi-class prediction model, called Semantic-DECO. Due to DAMON's limited training data, this has poor performance. Instead, as discussed in Sec. 4.1, InteractVLM learns effectively from this data, by also leveraging the commonsense of foundation models. Thus, it significantly outperforms Semantic-DECO. Qualitative results reflect this finding; see Fig. 4. Our model captures detailed, accurate contact regions, whereas Semantic-DECO often highlights false-positive areas that differ from the actual contact regions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.546, + 0.779, + 0.562 + ], + "angle": 0, + "content": "4.3. Object Affordance Prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.658 + ], + "angle": 0, + "content": "We evaluate the performance of our InteractVLM model on predicting object affordances. This involves identifying regions on objects of possible contact for a certain interaction intent, such as \"sitting\" on a chair or \"moving\" it. We train and evaluate on the PIAD [65] dataset. We compare against SotA methods, and report results in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Note that we evaluate on object instances that are both seen during training (\"PIAD-Seen\" column) and unseen (\"PIAD-Unseen\" column). Our model demonstrates a significant improvement over SotA methods in terms of similarity (SIM), area under the curve (AUC), and mean absolute error (MAE) for both seen and unseen objects. Thus, our method is effective not only for estimating human contact (Secs. 4.1 and 4.2), but also for object affordances." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.766, + 0.803 + ], + "angle": 0, + "content": "4.4. Reliance on 3D Annotations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To analyze InteractVLM's efficiency regarding 3D supervision, we train multiple versions of it with varying amounts of training data from the DAMON dataset, and report performance in Fig. 5. Remarkably, InteractVLM achieves an F1 score of 0.53 with just \\(1\\%\\) of the data, nearly matching the 0.55 F1 score of DECO that uses \\(100\\%\\) of the data. With" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22610" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.177, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.153, + 0.176, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.213, + 0.176, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.276, + 0.176, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.34, + 0.172, + 0.352 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.095, + 0.253, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.155, + 0.253, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.219, + 0.253, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.279, + 0.253, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.097, + 0.328, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.155, + 0.327, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.219, + 0.327, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.279, + 0.327, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.095, + 0.413, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.155, + 0.413, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.219, + 0.413, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.279, + 0.413, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.097, + 0.487, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.155, + 0.486, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.279, + 0.487, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.09, + 0.578, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.153, + 0.578, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.215, + 0.578, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.276, + 0.579, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.34, + 0.574, + 0.353 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.095, + 0.655, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.155, + 0.656, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.219, + 0.656, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.279, + 0.656, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.097, + 0.729, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.155, + 0.728, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.283, + 0.728, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.095, + 0.814, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.155, + 0.815, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.281, + 0.815, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.098, + 0.887, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.831, + 0.155, + 0.886, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.281, + 0.814, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.831, + 0.283, + 0.887, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.357, + 0.907, + 0.387 + ], + "angle": 0, + "content": "Figure 4. \"Semantic Human Contact\" estimation (Sec. 4.2). Given an image and an object label, InteractVLM infers body contacts for this object. InteractVLM outperforms a Semantic-DECO [58] baseline. Objects are shown in green circles, and contacts as red patches." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.406, + 0.47, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.639, + 0.483, + 0.737 + ], + "angle": 0, + "content": "Figure 5. InteractVLM's reliance on 3D annotations. We evaluate performance for \"binary human contact\" (F1 score, Y-axis) for models trained on a varying percentage of DAMON [58] training data (X-axis). The DECO baseline trains on \\(100\\%\\) of DAMON. Instead, InteractVLM trains on a varying (smaller) portion of this dataset. Yet, it achieves a significantly higher performance, by leveraging the broad visual knowledge of foundation models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "only \\(5\\%\\) of the data, InteractVLM surpasses DECO, reaching an F1 score of 0.58. This performance gap widens as training data increases, ultimately achieving 0.75 F1 score with \\(100\\%\\) of the data. This is a compelling finding, highlighting InteractVLM's efficient use of 3D supervision by leveraging the rich visual understanding of foundation models. Note that the strong performance with limited training data has high practical value, because obtaining 3D annotations is expensive." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.474 + ], + "angle": 0, + "content": "Ablations. For a study on the influence of InteractVLM components, including mask resolution, MV-Loc variants, loss functions, training data, the influence of the VLM, and the effect of text prompts, please refer to Sup. Mat." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.844, + 0.502 + ], + "angle": 0, + "content": "5. Joint Human-Object Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.906, + 0.555 + ], + "angle": 0, + "content": "We demonstrate the usefulness of 3D contacts inferred by InteractVLM for reconstructing a 3D human and object in interaction from a single in-the-wild image, \\(I\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.906, + 0.615 + ], + "angle": 0, + "content": "Initializing 3D body pose & shape, object shape. We use OSX [39] to estimate a 3D SMPL-X body mesh, \\( H \\), and OpenShape [42] to retrieve a 3D object mesh, \\( O \\), from the Objaverse [11] database that best matches the image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.906, + 0.72 + ], + "angle": 0, + "content": "Initializing 3D object pose. We apply InteractVLM on image \\(I\\) to predict 3D body and object contact vertices, \\(C^H, C^O\\). Then, we solve for object pose, \\(\\{R^O, t^O\\}\\) by applying the ICP [6] algorithm to snap the 3D object contact points, \\(C^O\\), onto body ones, \\(C^H\\). To avoid false correspondences, the 3D normals of contact points must be compatible, i.e., should have similar angles but opposite directions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.721, + 0.906, + 0.767 + ], + "angle": 0, + "content": "Optimizing 3D object pose. Given the above initialization, we optimize over object rotation, \\( R^O \\), translation, \\( t^O \\), and scale, \\( s^O \\), via render-and-compare by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.773, + 0.905, + 0.789 + ], + "angle": 0, + "content": "\\[\nE = E _ {M} + \\lambda_ {C} E _ {C}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.793, + 0.905, + 0.811 + ], + "angle": 0, + "content": "\\[\nE _ {M} = I o U (\\widehat {M}, M) + \\left\\| \\widehat {M ^ {c}} - M ^ {c} \\right\\| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.814, + 0.905, + 0.851 + ], + "angle": 0, + "content": "\\[\nE _ {C} = \\frac {1}{| C ^ {H} \\| C ^ {O} |} \\sum_ {i \\in | H |} \\sum_ {j \\in | O |} C _ {i} ^ {H} C _ {j} ^ {O} \\| V _ {i} ^ {H} - V _ {j} ^ {O} \\| _ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( E_{M} \\) is a mask loss, \\( E_{C} \\) is a contact loss, \\( IoU \\) is intersection-over-union, \\( \\widehat{M}, M \\) are hypothesis and ground-truth masks, respectively, \\( \\widehat{M^{c}}, M^{c} \\) are the corresponding" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22611" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.068, + 0.895, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.459, + 0.908, + 0.489 + ], + "angle": 0, + "content": "Figure 6. 3D HOI reconstruction (Sec. 5). We build an optimization method that fits a SMPL-X body and OpenShape-retrieved object to an in-the-wild image. We evaluate against the SotA method PHOSA [67]. Reconstruction is guided by InteractVLM-inferred contacts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.483, + 0.608 + ], + "angle": 0, + "content": "mean mask pixels, \\( |C^H| \\), \\( |C^O| \\) are the number of contact vertices on the human and object, \\( |H| \\), \\( |O| \\) are the number of mesh vertices, \\( V_i^H \\), \\( V_j^O \\) are the \\( i \\)-th and \\( j \\)-th human and object vertices, while \\( C_i^O \\), \\( C_j^H \\) indicate whether vertices are in contact or not. Given the image, we extract the object mask, \\( M \\), via SAM [31], and a depth map, \\( D \\), via Depth Pro [8]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Intuitively, we use the \\( E_{M} \\) loss to align the 3D object to the image, while in \\( E_{C} \\) the predicted 3D contacts \"anchor\" the 3D object onto the body so they interact realistically. We perform iterative optimization via Adam [30]. OSX produces reasonable bodies, so we keep these fixed. The object is updated in each iteration - we render depth, \\( \\widehat{D} \\), and masks, \\( \\widehat{M} \\), via a PyTorch3D [52] differentiable renderer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Qualitative Results. We reconstruct 3D human-object interaction from an image. We show results in Fig. 6, and compare with PHOSA [67], the most related SotA method. InteractVLM's reconstructions look more realistic. Note that PHOSA uses handcrafted contacts on humans and bodies. Instead, InteractVLM infers 3D contacts on both human bodies and objects from the single in-the-wild image. These play a crucial role for guiding 3D reconstruction under occlusions and depth ambiguities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Perceptual Study. There exists no in-the-wild dataset with 3D ground truth for HOI, so we conduct a perceptual study via Amazon Mechanical Turk for evaluation. Specifi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.907, + 0.621 + ], + "angle": 0, + "content": "cally, we evaluate the realism of our reconstructions against ones of PHOSA. We randomly select 55 images for which PHOSA has handcrafted contact annotations. For each image, participants are shown (with random swapping) reconstructions generated by our method and by PHOSA, and are asked to select the one that best represents the image. Our reconstructions are preferred \\(62\\%\\) of the time." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.634, + 0.65 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We develop InteractVLM, a novel method for estimating 3D contacts on both humans and objects from a single natural image. InteractVLM has a reduced reliance on expensive 3D contact annotations for training, by leveraging the broad knowledge of Vision-Language Models. Specifically, we introduce a novel \"Render-Localize-Lift\" (RLL) framework and a novel multi-view localization model (MV-Loc) to adapt 2D foundation models for 3D contact estimation. We outperform existing work on contact estimation and introduce a new \"Semantic Human Contact\" estimation task for inferring body contacts conditioned on object labels. This goes beyond traditional binary contact estimation, which fails to capture rich semantic relationships of multi-object interactions. Last, we develop the first approach that uses inferred contact points on both bodies and objects for joint 3D reconstruction from single in-the-wild images." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22612" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.243 + ], + "angle": 0, + "content": "Acknowledgments: We thank Alpár Cseke for his assistance with evaluating joint human-object reconstruction. We also thank Tsvetelina Alexiadis and Taylor Obersat for MTurk evaluation, Yao Feng, Peter Kultis and Markos Diomataris for their valuable feedback and Benjamin Pellkofer for IT support. SKD is supported by the International Max Planck Research School for Intelligent Systems (IMPRS-IS). The UvA part of the team is supported by an ERC Starting Grant (STRIPES, 101165317, PI: D. Tzionas). Disclosure: DT has received a research gift fund from Google. For MJB see https://files.is.tue.mpg.de/black/CoI_CVPR_2025.txt" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.271, + 0.188, + 0.287 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.296, + 0.484, + 0.351 + ], + "angle": 0, + "content": "[1] Emanuele Aiello, Diego Valsesia, and Enrico Magli. Cross-modal learning for image-guided point cloud shape completion. Conference on Neural Information Processing Systems (NeurIPS), 35:37349-37362, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.354, + 0.484, + 0.505 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: A visual language model for few-shot learning. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.507, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[3] Reza Yazdani Aminabadi, Samyam Rajbhandari, Minjia Zhang, A. A. Awan, Cheng Li, Du Li, Elton Zheng, Jeff Rasley, Shaden Smith, Olatunj Ruwase, and Yuxiong He. DeepSpeed inference: Enabling efficient inference of transformer models at unprecedented scale. International Conference for High Performance Computing, Networking, Storage and Analysis, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[4] Dimitrije Antić, Georgios Paschalidis, Shashank Tripathi, Theo Gevers, Sai Kumar Dwivedi, and Dimitrios Tzionas. SDFit: 3D object pose and shape by fitting a morphable SDF to a single image. arXiv:2409.16178, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[5] Fabien Baradel*, Matthieu Armando, Salma Galaoui, Romain Brégier, Philippe Weinzaepfel, Grégory Rogez, and Thomas Lucas*. Multi-HMR: Multi-person whole-body human mesh recovery in a single shot. In European Conference on Computer Vision (ECCV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[6] Paul J. Besl and Neil D. McKay. A method for registration of 3D shapes. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 14:239-256, 1992. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[7] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAVE: Dataset and method for tracking human object interactions. In Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.845, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[8] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth Pro: Sharp monocular metric depth in less than a second. arXiv:2410.02073, 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.296, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[9] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision (ECCV), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[10] Honghua Chen, Zeyong Wei, Yabin Xu, Mingqiang Wei, and Jun Wang. ImLoveNet: Misaligned image-supported registration network for low-overlap point cloud pairs. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-9, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.907, + 0.316 + ], + "angle": 0, + "content": "[11] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2023. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.318, + 0.906, + 0.373 + ], + "angle": 0, + "content": "[12] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3D AffordanceNet: A benchmark for visual object affordance understanding. In Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[13] John Doe, Jane Smith, and Alice Brown. Paris3d: Language-guided 3d segmentation with multimodal models. arXiv:2406.08394, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[14] Sai Kumar Dwivedi, Nikos Athanasiou, Muhammed Kocabas, and Michael J. Black. Learning to regress bodies from images using differentiable semantic rendering. In International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.474, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[15] Sai Kumar Dwivedi, Cordelia Schmid, Hongwei Yi, Michael J. Black, and Dimitrios Tzionas. POCO: 3D pose and shape estimation using confidence. In International Conference on 3D Vision (3DV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.531, + 0.906, + 0.585 + ], + "angle": 0, + "content": "[16] Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, Yao Feng, and Michael J. Black. TokenHMR: Advancing human mesh recovery with a tokenized pose representation. In Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.587, + 0.906, + 0.656 + ], + "angle": 0, + "content": "[17] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual handobject manipulation. In Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.657, + 0.906, + 0.727 + ], + "angle": 0, + "content": "[18] Zicong Fan, Maria Parelli, Maria Eleni Kadoglou, Muhammed Kocabas, Xu Chen, Michael J Black, and Otmar Hilliges. HOLD: Category-agnostic 3D reconstruction of interacting hands and objects from video. In Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.728, + 0.906, + 0.783 + ], + "angle": 0, + "content": "[19] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Collaborative regression of expressive bodies using moderation. In International Conference on 3D Vision (3DV), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.785, + 0.906, + 0.84 + ], + "angle": 0, + "content": "[20] Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, and Tae-Kyun Kim. First-person hand action benchmark with RGB-D videos and 3D hand pose annotations. In Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.841, + 0.905, + 0.909 + ], + "angle": 0, + "content": "[21] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.909 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22613" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[22] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambiguities with 3D scene constraints. In International Conference on Computer Vision (ICCV), pages 2282-2292, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[23] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14708-14718, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.285 + ], + "angle": 0, + "content": "[24] Yana Hasson, Gül Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.287, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[25] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. arXiv: 2311.04400, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.482, + 0.397 + ], + "angle": 0, + "content": "[26] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. International Conference on Learning Representations (ICLR), 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.398, + 0.482, + 0.478 + ], + "angle": 0, + "content": "[27] Chun-Hao Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Computer Vision and Pattern Recognition (CVPR), pages 13264-13275, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.481, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[28] Yinghao Huang, Omid Taheri, Michael J. Black, and Dimitrios Tzionas. InterCap: Joint markerless 3D tracking of humans and objects in interaction from multi-view RGB-D images. International Journal of Computer Vision (IJCV), 132(7):2551-2566, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.55, + 0.482, + 0.605 + ], + "angle": 0, + "content": "[29] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In International Conference on 3D Vision (3DV), pages 42-52, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.606, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[30] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloé Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.73, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[32] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. VIBE: Video inference for human body pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), pages 5252-5262, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.785, + 0.482, + 0.84 + ], + "angle": 0, + "content": "[33] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.841, + 0.482, + 0.895 + ], + "angle": 0, + "content": "[34] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: Reasoning segmentation via large language model. In Computer Vision and Pattern Recognition (CVPR), pages 9579-9589, 2024. 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[35] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. HybrIK: A hybrid analytical-neural inverse kinematics solution for 3D human pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), pages 3383-3393, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning (ICML), pages 19730-19742, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[37] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[38] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-resolution text-to-3D content creation. Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[39] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3D whole-body mesh recovery with component aware transformer. In Computer Vision and Pattern Recognition (CVPR), 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[40] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[42] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. OpenShape: Scaling up 3D shape representation towards open-world understanding. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[43] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3D mesh in 45 seconds without pershape optimization. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[44] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[45] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. Transactions on Graphics (TOG), 34 (6):248:1-248:16, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.757, + 0.905, + 0.812 + ], + "angle": 0, + "content": "[46] Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. COAP: Compositional articulated occupancy of people. In Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.814, + 0.905, + 0.869 + ], + "angle": 0, + "content": "[47] Hyeongjin Nam, Daniel Sungho Jung, Gyeongsik Moon, and Young Mu Lee. Joint reconstruction of 3D human and object via contact-based refinement transformer. In Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22614" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[48] Mohamed Omran, Christoph Lassner, Gerard Pons-Moll, Peter Gehler, and Bernt Schiele. Neural body fitting: Unifying deep learning and model based human pose and shape estimation. In International Conference on 3D Vision (3DV), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[49] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975–10985, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[50] Bingbing Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[51] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. In International Conference on Learning Representations (ICLR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.438 + ], + "angle": 0, + "content": "[52] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D deep learning with PyTorch3D. arXiv:2007.08501, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.44, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[53] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. Transactions on Graphics (TOG), 36(6), 2022, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.496, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[54] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[55] Qingping Sun, Yanjun Wang, Ailing Zeng, Wanqi Yin, Chen Wei, Wenjia Wang, Haiyi Mei, Chi Sing Leung, Ziwei Liu, Lei Yang, and Zhongang Cai. AiOS: All-in-one-stage expressive human pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.621, + 0.482, + 0.675 + ], + "angle": 0, + "content": "[56] István Sárándi and Gerard Pons-Moll. Neural localizer fields for continuous 3D human pose and shape estimation. In Conference on Neural Information Processing Systems (NeurIPS), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.677, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[57] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[58] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J. Black. DECO: Dense estimation of 3D human-scene contact in the wild. In International Conference on Computer Vision (ICCV), pages 8001-8013, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[59] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.829 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "coordinate space for category-level 6D object pose and size estimation. In Computer Vision and Pattern Recognition (CVPR), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[60] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, and Jifeng Dai. VisionLLM: Large language model is also an open-ended decoder for vision-centric tasks. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[61] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. CHORE: Contact, human and object reconstruction from a single RGB image. In European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[62] Xianghui Xie, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Template free reconstruction of human-object interaction with procedural interaction generation. In Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[63] Danfei Xu, Dragomir Anguelov, and Ashesh Jain. PointFusion: Deep sensor fusion for 3D bounding box estimation. In International Conference on Computer Vision (ICCV), pages 244-253, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[64] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, Bill Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.458, + 0.905, + 0.526 + ], + "angle": 0, + "content": "[65] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3D object affordance from 2D interactions in images. In International Conference on Computer Vision (ICCV), pages 10871-10881, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.528, + 0.905, + 0.583 + ], + "angle": 0, + "content": "[66] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, and Zheng-Jun Zha. LEMON: Learning 3D human-object interaction relation from 2D images. In Computer Vision and Pattern Recognition (CVPR), 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.585, + 0.905, + 0.654 + ], + "angle": 0, + "content": "[67] Jason Y. Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3D human-object spatial arrangements from a single image in the wild. In European Conference on Computer Vision (ECCV), 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.655, + 0.905, + 0.71 + ], + "angle": 0, + "content": "[68] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. GPT4RoI: Instruction tuning large language model on region-of-interest. arXiv:2307.03601, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.712, + 0.905, + 0.767 + ], + "angle": 0, + "content": "[69] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3D: Exploring unified 3D representation at scale. In International Conference on Learning Representations (ICLR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.768, + 0.905, + 0.824 + ], + "angle": 0, + "content": "[70] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3D lidar semantic segmentation. In International Conference on Computer Vision (ICCV), 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22615" + } + ] +] \ No newline at end of file diff --git a/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/9f03277a-37b8-49e7-9fad-13e2e5f171b2_origin.pdf b/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/9f03277a-37b8-49e7-9fad-13e2e5f171b2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f9eb9a30f0a72cb720a3789c181929071a01462b --- /dev/null +++ b/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/9f03277a-37b8-49e7-9fad-13e2e5f171b2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:266f1b8cba4f27180fed8893b8410830ec8a9512c4dd4262beb3fca0e705c34b +size 4541196 diff --git a/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/full.md b/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..08b6ceca5223265550d49403da785b10e8e63a0a --- /dev/null +++ b/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/full.md @@ -0,0 +1,405 @@ +# InteractVLM: 3D Interaction Reasoning from 2D Foundational Models + +Sai Kumar Dwivedi1 Dimitrije Antić2 Shashank Tripathi1 Omid Taheri1 Cordelia Schmid3 Michael J. Black1 Dimitrios Tzionas2 + +1Max Planck Institute for Intelligent Systems, Tübingen, Germany 2University of Amsterdam (UvA), the Netherlands 3Inria, École normale supérieure, France + +![](images/9406fdef2334f1a5f620b804e6bdddfa4e1a21b853286f21a008a8517cc7ce8a.jpg) +Figure 1. We present InteractVLM, a novel method for estimating contact points on both human bodies and objects from a single in-the-wild image, shown here as red patches. Our method goes beyond traditional binary contact estimation methods by estimating contact points on a human in relation to a specified object. We do so by leveraging the broad visual knowledge of a large Visual Language Model. + +# Abstract + +We introduce InteractVLM, a novel method to estimate 3D contact points on human bodies and objects from single in-the-wild images, enabling accurate human-object joint reconstruction in 3D. This is challenging due to occlusions, depth ambiguities, and widely varying object shapes. Existing methods rely on 3D contact annotations collected via expensive motion-capture systems or tedious manual labeling, limiting scalability and generalization. To overcome this, InteractVLM harnesses the broad visual knowledge of large Vision-Language Models (VLMs), fine-tuned with limited 3D contact data. However, directly applying these models is non-trivial, as they reason "only" in 2D, while human-object contact is inherently 3D. Thus we introduce a novel "Render-Localize-Lift" module that: (1) embeds 3D body and object surfaces in 2D space via multiview rendering, (2) trains a novel multi-view localization model (MV-Loc) to infer contacts in 2D, and (3) lifts these to 3D. Additionally, we propose a new task called Semantic Human Contact estimation, where human contact predictions are conditioned explicitly on object semantics, enabling richer interaction modeling. InteractVLM outperforms existing work on contact estimation and also facilitates 3D reconstruction from an in-the-wild image. To es + +timate 3D human and object pose, we infer initial body and object meshes, then infer contacts on both of these via InteractVLM, and last exploit these for fitting the meshes to image evidence. Results show that our approach performs promisingly in the wild. Code and models are available at https://interactvlm.is.tue.mpg.de. + +# 1. Introduction + +People interact with objects routinely. Reconstructing human-object interaction (HOI) in 3D is key for many applications, from robots to mixed reality. However, doing so from single images is challenging due to depth ambiguity, occlusions, and the diverse shape and appearance of objects. + +There are methods that estimate 3D human bodies and methods that estimate 3D objects, but few that put these together. Knowing the contacts between these could significantly improve joint reconstruction. Our goal is to infer contact points on both humans and objects from single in-the-wild images, and then use these contacts for jointly reconstructing humans and objects. However, there is a lack of in-the-wild training images paired with ground-truth contact labels for both 3D humans and objects. Acquiring such data is challenging and existing methods do not scale. + +The problem gets more challenging due to the complexity of real-world interactions. Humans often contact multiple objects simultaneously; e.g., using a laptop while sitting on a club chair. Yet, current approaches treat contact prediction as simple binary classification; i.e. detecting whether a body part is in contact with "any" object. This simplified assumption fails to capture the rich semantic relationships of multi-object interactions. To address this, we introduce a novel "Semantic Human Contact" estimation task. This involves predicting the contact points on the body related to a particular object given a single in-the-wild image. + +To tackle this, as well as to overcome data scarcity, we propose a new paradigm for scaling and reasoning in the wild. Specifically, we observe that large Vision-Language Models (VLMs) can "reason" about in-the-wild images because they are trained on internet-scale data and possess a broad visual knowledge about humans and their interactions with the world. We also observe that this knowledge can be re-purposed for novel tasks by fine-tuning these large models on small datasets. Thus, we exploit VLMs for developing a novel framework, called InteractVLM. + +At the heart of InteractVLM lies a reasoning module based on a VLM; see Fig. 2. We enrich this with skills for 3D human-object "understanding," by extending the VLM with a LoRA [26] adaptation layer. As a result, given a color image, this module can be "asked" to produce "reasoning tokens" that facilitate 3D contact localization. + +However, exploiting these tokens to localize contact is non-trivial. A natural choice would be to employ a foundational "localization" model [31] that takes these tokens as "guidance," and highlights the 3D contacts. But there exists a key practical problem. Existing foundation models inherently operate only in 2D space, while we need them to do so in 3D. To tackle this, we need to re-cast our problem so it is appropriate for 2D foundation models. + +To this end, we develop a novel "Render-Localize-Lift" (RLL) framework that has three main steps; see Fig. 2: (1) We render the 3D shape of a canonical SMPL+H [53] body and an object as 2D images from multiple viewpoints. Regarding the object, a 3D mesh is efficiently retrieved from a large-scale 3D database [11] through OpenShape [42]. (2) We pass the above images into a foundation model to predict 2D contact masks for both the body and the object. (3) We lift the predicted 2D contact points to 3D points via back-projection, i.e., performing the inverse of the first step. + +However, even after recasting the problem to 2D by rendering multi-view images, existing foundation models are still not 3D aware, i.e. they treat each view independently, ignoring multi-view consistency. This means contact detections in one view do not necessarily agree with ones in adjacent views. To tackle this, just appending camera parameters to multi-view renderings is insufficient. Instead, we build a novel "Multi-view Localization" model that we call + +![](images/41285023f276fd719bccc053beeb3d443a9efa6a33aadaf318560de488cc166c.jpg) +Figure 2. Overview of InteractVLM. Given a color image, our VLM performs the core reasoning, and guides a novel MV-Loc model to localize contacts on both bodies and objects in 3D. Here we show only the body; for details, and object contact, see Fig. 3. + +MV-Loc. This has two steps: (1) It transforms the "reasoning token" provided by the VLM with the camera parameters used to render the multi-view images. (2) It ensures multi-view consistency, by lifting the inferred 2D contacts in each view to 3D and computing a 3D loss. + +Our method, InteractVLM, utilizes a VLM in tandem with our novel multi-view localization model, MV-Loc, to perform 3D contact prediction for humans and objects. See Fig. 1 for some examples. We quantitatively evaluate the efficacy of our method for in-the-wild 3D contact prediction on bodies and on objects, using the DAMON [58] and PIAD [65] datasets, respectively. For bodies, we evaluate both for the traditional "binary contact" estimation, and for our new task of "semantic contact" estimation. For all tasks, we find that InteractVLM outperforms the prior work. + +Finally, we demonstrate how InteractVLM's estimated contact improves 3D HOI recovery from in-the-wild images; this is a highly ill-posed task due to depth ambiguities and occlusions. To address this, we develop an optimization-based method that fits a SMPL-X body mesh and an OpenShape-retrieved object mesh to the image, using InteractVLM's inferred contacts as constraints to anchor human and object meshes w.r.t. each other. To the best of our knowledge, this is the first approach for estimating 3D HOI for in-the-wild images using inferred contacts. + +In summary, we make four key contributions: + +1. We build InteractVLM, a novel method that facilitates HOI reconstruction from an in-the-wild image by detecting 3D contacts on both bodies and objects. +2. We demonstrate a way to minimize reliance on 3D contact annotations via exploiting the broad visual knowledge of Vision-Language Models. +3. We build a novel "Multi-view Localization" model that helps in estimating contacts in 3D by transforming the reasoning of foundation models from 2D to 3D. +4. We introduce the novel "Semantic Human Contact" task for inferring body contacts conditioned on object labels. Our code and trained models are available for research at https://interactglm.is.tue.mpg.de. + +# 2. Related Work + +# 2.1. Large Vision-Language Models + +Recent advancements in large language models (LLMs) have led to the development of multimodal models that integrate vision and language reasoning. Models like Flamingo [2] and BLIP-2 [36] use cross-attention mechanisms and visual encoders to align image features with text, supporting a variety of vision-language tasks. More recent works, such as VisionLLM [60] and Kosmos-2 [50], use grounded image-text data to enhance spatial understanding, while GPT4RoI [68] introduces spatial box inputs for finer alignment. However, these models typically lack end-to-end segmentation capabilities. To address this limitation, LISA [34] combines vision foundation segmentation models like SAM [31] with multimodal embeddings, enabling language-guided segmentation. PARIS3D [13] extends this approach to referential 3D segmentation by processing multi-view object renders through both SAM and LLaVA [41] for spatially-aware segmentation. + +Taking inspiration from these approaches, we exploit a language-guided segmentation model for the task of predicting human and object contact in 3D. However, unlike PARIS3D, we process a single RGB image with LLaVA and multi-view renders of the human mesh and the object mesh with SAM. Moreover, we introduce a feature-lifting technique that extends LLaVA's 2D features into 3D using camera parameters, thus guiding SAM's multi-view segmentations. This approach ensures multi-view consistency and efficiently predicts contact affordances, extending the use of multimodal models in 3D reasoning tasks focused on human-object interaction. + +# 2.2. 3D Human and Object from Single Images + +Estimating 3D human pose and shape from a single image has evolved from optimization-based methods to learning-based approaches. Optimization-based methods fit parametric body models like SMPL [45], SMPL-X [49], or GHUM [64] to 2D cues such as keypoints [9], silhouettes [48], or segmentation masks [48]. Learning-based methods either regress body parameters from images or videos [5, 14, 15, 29, 32, 33, 37] or estimate non-parametric bodies as vertices [35, 40], implicit surfaces [46, 54], or dense points [56]. Transformer-based methods [16, 21, 39, 55] have further improved robustness. + +For 3D object reconstruction from a single image, regression-based methods predict geometry using meshes, voxels, or point clouds. Diffusion-based models [25] utilize large 3D datasets like Objaverse [11] or 2D diffusion models [38, 43, 44, 51] to guide reconstruction and optimization methods [4] use render and compare. Retrieval-based methods, such as OpenShape [42] and Uni3D [69], have demonstrated some robustness in cases with occlusions. + +# 2.3. 3D Human-Object Interaction + +Understanding 3D human-object interactions is essential for modeling realistic scenes. Early works focused on handobject interactions, such as ObMan [24] and FPHA [20], with more recent studies like ARCTIC [17] and HOLD [18] providing more detailed data and reconstruction for hands. For full-body interactions, initial studies involve interactions with scenes, as in PROX [22], and with objects, as in BEHAVE [7], GRAB [57], and InterCap [28]. BEHAVE and GRAB use motion capture that are accurate but nonscalable, while InterCap uses multi-camera setups that are more scalable but less accurate. Both of these approaches are limited in capturing diverse and realistic interactions. + +As a proxy for 3D reconstruction, the recent DECO [58] method infers contacts on body meshes; to this end, 3D contact annotations are crowdsourced via mesh painting. Predicting contact on objects is more challenging due to varying object shapes, while also currently there exists no dataset of object contact for in-the-wild images. Thus, we estimate 3D object affordances as a proxy for contact. 3D-AffordanceNet [12] introduces affordances that are not grounded in images, capturing the likelihood of humans interacting with specific parts of an object for a given affordance (e.g., "sit on chair"). PIAD [65] curates RGB images depicting object affordances and trains a network to estimate them. LEMON [66] extends object affordance prediction of PIAD to include human contact estimation. However, these methods require human contact vertices paired with object affordances for training, limiting the number of categories they handle to 21 ones. In contrast, we learn from unpaired human and object interaction data and perform joint interaction reasoning via a cross-attention layer, enabling human interaction reasoning for 80 categories and object affordance prediction for 32 categories. + +# 2.4. Joint 3D Human-Object Reconstruction + +Joint reconstruction of humans and objects in 3D has been tackled with both regression and optimization methods. Regression-based methods directly predict 3D human-object meshes, as in HDM [62] and CONTHO [47], while other methods first predict contact points and then use test-time optimization to fit human and object meshes, as in CHORE [61] and PHOSA [67]. Since regression methods rely on limited training data, optimization methods are preferred for in-the-wild scenarios, such as PHOSA [67]. Optimization-based methods either assume known contacts or infer contacts to fit meshes to the image, but their success heavily relies on the quality of contact. + +Our method improves upon these by providing more accurate contact predictions, which in turn facilitate better fitting of human and object meshes to image evidence, improving the realism and accuracy of 3D human-object reconstructions from single images. + +![](images/4c7e3284f279be476c0109e0c6cc5a8b798e40e2af9aaaf8c482c32c26185e0a.jpg) +Figure 3. Method overview. Given a single in-the-wild color image, our novel InteractVLM method estimates 3D contact points on both humans and objects (a). Then, we reconstruct a 3D human and object in interaction by exploiting these contacts (b). More specifically: (a) Contact estimation. Given an image, $I$ , and prompt text, $T_{inp}$ , our VLM, $\Psi$ , produces contact tokens for humans and objects, $<\mathsf{HCON}>$ and $<0\mathsf{CON}>$ , which are projected ( $\Gamma$ ) into feature embeddings, $E^H$ and $E^O$ . These guide a "Multi-View [contact] Localization" model. This renders the 3D human and object geometry via cameras, $K$ , into multi-view 2D renders and passes these to encoder, $\Theta$ , while decoders, $\Omega^H$ , $\Omega^O$ , estimate and highlight 2D contacts in these renders. Then, the FeatLift module, $\Phi$ , transforms the VLM's features $(E^H, E^O)$ to become 3D-aware $(E_{3D}^H, E_{3D}^O)$ by exploiting the camera parameters, $K$ . A final module lifts the detected 2D contacts to 3D. (b) 3D HOI reconstruction. For joint human-object reconstruction, we use InteractVLM's inferred contacts in an optimization framework. + +![](images/5974398eb285b904d1de43c22f78dcae6534bc3da026842cc044de35856ce37e.jpg) + +# 3. Method + +# 3.1. Input Representation + +Given an image, $I \in \mathbb{R}^{H \times W \times 3}$ , InteractVLM estimates 3D contacts for both human bodies and objects. + +The human is represented by a SMPL+H [53] 3D body mesh, $H$ , with vertices $V^{H} \in \mathbb{R}^{10475 \times 3}$ . The body is posed in a canonical "star" shape (see details in Sec. 3.4). The human contacts are binary per-vertex labels, $C^{H} \in \{0,1\}$ . + +The object is represented by a 3D point cloud (or mesh), $O \in \mathbb{R}^{N \times 3}$ , with $N$ points. As there are no datasets of natural images paired with 3D contacts for objects, we use a largescale 3D affordance dataset [65], instead, as affordances are closely related to contact. Specifically, they represent the likelihood of contact on 3D object areas for various purposes. Therefore, for objects we use the terms "affordance" and "contact" interchangeably. The object contacts are continuous per-point values, $C^O \in [0,1]$ . During inference, we retrieve a 3D object shape from a large database [11] via OpenShape [42] conditioned on image $I$ . + +# 3.2. Overview of InteractVLM + +The biggest challenge for learning 3D contact prediction in the wild is the limited 3D contact data for humans and objects. To go beyond existing limited datasets, we introduce a novel method, called InteractVLM, that harnesses the commonsense knowledge of large vision-language models. + +Specifically, InteractVLM (Fig. 3) has two main components: a Vision Language Model (VLM) and a novel Multi-View contact Localization model (MV-Loc). MV-Loc highlights parts that are in contact for both humans and objects with the VLM's guidance. The input to the VLM (Sec. 3.3) is an image, $I$ , and a prompt, $T_{inp}$ , that asks the VLM to detect contact. The input to MV-Loc (Sec. 3.4) is the 3D geometry of humans and objects, $H$ and $O$ , respectively. + +# 3.3. Interaction Reasoning through VLM + +The VLM module, $\Psi$ , conducts the core interaction reasoning. It takes an input image, $I$ , and prompt text, $T_{inp}$ , and outputs a prompt text, $T_{out} = \Psi(I, T_{inp})$ . Inspired by the recent LISA [34] model, we expand the VLM's vocabulary with two specialized tokens, $\langle \text{HCON} \rangle$ and $\langle \text{OCON} \rangle$ , for contact information for humans and objects, respectively. + +To denote contact, $\Psi$ produces a prompt that includes the above tokens. To aid MV-Loc in localizing contact, we extract the last-layer embeddings of the VLM corresponding to these tokens and pass them through a projection layer, $\Gamma$ , to obtain the feature embeddings, $E^H$ or $E^O$ . Let $T_{\mathrm{gt}}$ be the ground-truth text, and $T_{\mathrm{pred}}$ be the predicted one. Then, our token-prediction loss is defined as a cross-entropy loss: + +$$ +\mathcal {L} _ {\text {t o k e n}} = - \sum_ {i = 1} ^ {N} \left(T _ {\mathrm {g t}} ^ {(i)} \cdot \log \left(T _ {\text {p r e d}} ^ {(i)}\right)\right). \tag {1} +$$ + +# 3.4. Interaction Localization through MV-Loc + +We develop a novel MV-Loc module that has a shared image encoder, $\Theta$ , and separate decoders, $\Omega^H$ and $\Omega^O$ , for humans and objects. MV-Loc performs contact localization by using a novel "Render-Localize-Lift" (RLL) framework. To this end, it takes three steps; (1) rendering the 3D shape of humans and objects in 2D, (2) predicting 2D contact maps for both of these, and (3) lifting the 2D contact maps to 3D. + +RLL step #1: Render 3D $\rightarrow$ 2D. The input is human and object geometry, namely $\mathcal{H}$ and $O$ , respectively. Both serve as a "canvas" for "painting" detected contacts on these. The body has a default SMPL+H shape in a canonical star-shape pose to minimize self-occlusions when rendering. The object geometry (initialized in Sec. 3.1) is normalized to a unit sphere. Each geometry is rendered from $J$ fixed views with camera parameters, $K$ , to form multi-view renderings, $R^{H,O} = \{R_j\}_{j=1}^J$ , so that the entire 3D geometry is captured. Since our geometries do not have texture, we color the meshes with normals and point clouds using the NOCS map [59]. This enhances cross-view correspondence, making renderings resemble real images to image encoder, $\Theta$ . + +RLL step #2: Localize in 2D. The rendered geometry, $R^{H,O}$ , is first sent to the image encoder, $\Theta$ , and then passed to decoders, so that the final contact masks, $M^H$ and $M^O$ , get highlighted on it. However, MV-Loc requires spatial and contextual cues for highlighting the contact region. To this end, we use the feature embeddings (Sec. 3.3), i.e. $E^H$ and $E^O$ to guide the contact localization. + +However, since the VLM reasons in 2D, these features are not 3D aware. This is a problem, because MV-Loc needs 3D awareness to localize contact consistently across multi-view renderings. Thus, we transform the features to "lift" them to "3D" to better guide multi-view localization. + +In detail, we design a lifting network, $\Phi$ , which takes camera parameters, $K$ , and the 2D $E^{H,O}$ , and lifts the latter to 3D as $E_{3D}^{H,O} = \Phi(E^{H,O}, K)$ . Contact masks are defined as: + +$$ +M ^ {H, O} = \Omega^ {H, O} (\Theta (R ^ {H, O}), E _ {3 D} ^ {H, O}). \tag {2} +$$ + +Below, the $H$ and $O$ superscripts are dropped for brevity. We calculate losses only on "valid" regions, i.e., the areas within the outline of rendered geometry; we denote this as $M$ . To encourage overlap between the predicted masks, $M$ , and the ground-truth ones, $\widehat{M}$ , particularly in sparse contact regions, we use a focal-weighted BCE loss and a Dice loss: + +$$ +\mathcal {L} _ {B C E} = - \alpha (1 - p _ {M}) ^ {\gamma} \log (p _ {M}) - (1 - \alpha) p _ {M} ^ {\gamma} \log (1 - p _ {M}), \tag {3} +$$ + +$$ +\mathcal {L} _ {D i c e} = 1 - \frac {2 \sum M \cdot \widehat {M} + \epsilon}{\sum M + \sum \widehat {M} + \epsilon}, \tag {4} +$$ + +where $p_M$ is the predicted mask probability, $\alpha$ controls class balance, $\gamma$ adjusts the focus on hard examples, and $\epsilon$ is a residual term to prevent division by zero. + +RLL step #3: Lift $2\mathbf{D} \rightarrow 3\mathbf{D}$ . To lift the inferred 2D contact points to 3D points, we follow the inverse of step #1. Normally, 2D points backproject to 3D lines due to depth ambiguities. In our case the lines intersect with the known 3D geometry that produced the multi-view renders. So, 2D points are lifted to 3D points, and by extension 2D contact masks, $M^{H,O}$ , are lifted to 3D contact areas, $C^{H,O}$ . + +We use a human contact loss, $\mathcal{L}_C^H$ , that combines a focal loss with sparsity regularization, to encourage precise true-positive predictions in "valid" contact regions while discouraging false positives in non-contact areas: + +$$ +\mathcal {L} _ {C} ^ {H} = \alpha (1 - p _ {h C}) ^ {\gamma} \log \left(p _ {h C}\right) + \lambda \| C ^ {H} \| _ {1}, \tag {5} +$$ + +where $p_{hC}$ is the contact probability, and $\lambda$ , $\alpha$ and $\gamma$ are scalar weights. We also use an object contact loss, $\mathcal{L}_C^O$ , that combines a Dice and Mean Squared Error (MSE) loss: + +$$ +\mathcal {L} _ {C} ^ {O} = \mathcal {L} _ {\text {D i c e}} \left(C ^ {O}, \widehat {C} ^ {O}\right) + \beta \| C ^ {O} - \widehat {C} ^ {O} \| _ {2} ^ {2}, \tag {6} +$$ + +where $\beta$ is a weighting factor, and $\widehat{C^O}$ denotes the groundtruth 3D contacts for objects. + +# 3.5. Implementation Details + +Architecture. We use LLaVA [41] as our VLM and SAM [31] for our MV-Loc, with weights pre-trained by LISA [34] for segmentation [34]. The feature-lifting network, $\Phi$ , contains a spatial-understanding network (two fully-connected layers of size 128 with ReLU activation) followed by view-specific (256-dimensional) transformations and a sigmoid activation. For 3D contact prediction, our MV-Loc model converts 2D masks to 3D contact points via 2D-to-3D pixel-to-vertex mappings that are precomputed during MV-Loc's rendering step. For details, see Sup. Mat. + +Training. To efficiently fine-tune our VLM, we employ LoRA [26] with rank 8. The separate decoders for human and object contact prediction are trained without LoRA, while keeping the image encoder frozen. For training, we use DeepSpeed [3] with mixed precision training (bfloat16), and batch size of 8. We train on 4 Nvidia-A100 GPUs for 30 epochs. For more details, please refer to Sup. Mat. + +Datasets. We focus on two tasks, i.e., 3D human contact and 3D object affordance prediction, using two in-the-wild datasets, i.e., DAMON [58] and PIAD [65], respectively. For human contact we train and evaluate on DAMON [58]. For 3D object affordances, we train and evaluate on PIAD [65]. We find that exploiting textual descriptions for the contacting body parts, and the object type in contact, helps training. Similarly, adding Visual Question-Answering (VQA) data generated by GPT4o for the training images also helps. For details, see Sup. Mat. + +Unlike LEMON [66], which requires paired human-object geometry for training using the 3DIR dataset [66], we use unpaired data. This enables us to scale for many human and object categories not addressed by prior work. For + +
MethodF1 (%)↑Precision (%)↑Recall (%)↑Geodesic (cm)↓
POSAPPIXIE [23] [19]31.042.034.033.00
BSTRO [27]46.051.053.038.06
DECO [58]55.065.057.021.32
InteractVLM75.675.276.02.89
+ +Table 1. Evaluation for "Binary Human Contact" prediction on the DAMON dataset [58]. We compare our InteractVLM model (trained only for this task) with the state of the art. + +
Object CategoriesSemantic-DECO [58] (Baseline)InteractVLM
F1 (%)↑Prec. (%)↑Rec. (%)↑Geo (cm)↓F1 (%)↑Prec. (%)↑Rec. (%)↑Geo (cm)↓
Accessory40.130.775.621.8861.172.965.66.91
Daily Obj26.520.152.360.3468.671.478.07.46
Food11.719.412.949.6166.466.177.97.17
Furniture24.515.883.729.1760.564.260.16.21
Kitchen27.724.737.252.3471.871.381.17.61
Sports36.430.480.179.2177.976.783.27.98
Transport52.039.193.731.7877.880.578.97.97
+ +the final joint human-object reconstruction task, we combine DAMON [58], PIAD [65], 3DIR [66] and all textual descriptions about body parts, contact type and HOI. + +Evaluation metrics. For human contact prediction, following Tripathi et al. [58], we report the F1, precision, and recall scores using a threshold of 0.5, and a geodesic distance measuring spatial accuracy. For object contact prediction, following Yang et al. [65], we report the Similarity (SIM), Mean Absolute Error (MAE), Area Under ROC Curve (AUC), and average Intersection over Union (IOU). + +# 4. Experiments + +# 4.1. "Binary Human Contact" Estimation + +This task refers to estimating contact areas on the body via binary classification of its vertices, ignoring the number or type of objects involved. We train and evaluate on the DAMON [58] dataset, and report results in Tab. 1. + +InteractVLM significantly outperforms all previous methods achieving a $20.6\%$ improvement in F1 score. Although, here, InteractVLM is trained on the same data as DECO, it goes beyond this by harnessing the commonsense knowledge of a large foundation model. + +In Sup. Mat. we also evaluate on the 3DIR [66] dataset and compare with the LEMON method [66]. Even though LEMON uses paired human-object data, InteractVLM performs on-par with it despite training on human-only data. We also evaluate InteractVLM's performance for binary human contact prediction across different body parts; for more details please refer to Sup. Mat. + +Table 2. Evaluation for "Semantic Human Contact" prediction on the DAMON [58] dataset. For results on each class, see Sup. Mat. The "Semantic-DECO" baseline extends DECO for our new task. + +
MethodsPIAD-Seen [65]PIAD-Unseen [65]
SIM (%)↑AUC (%)↑aIOU (%)↑MAE ↓SIM (%)↑AUC (%)↑aIOU (%)↑MAE ↓
PMF [70]42.575.0510.131.4133.060.254.672.11
ILN [10]42.775.8411.521.3732.559.694.712.07
PFusion [63]43.277.5012.311.3533.061.875.331.93
XMF [1]44.178.2412.941.2734.262.585.681.88
IAGNet [65]54.584.8520.510.9835.271.847.951.27
InteractVLM62.786.4721.200.8141.475.458.500.99
+ +Table 3. Evaluation for "Object Affordance Prediction" on the PIAD [65] dataset. We compare our InteractVLM model (trained only for this task) with the state of the art. + +# 4.2. "Semantic Human Contact" Estimation + +In real-life interactions, multiple objects can be contacted by different body areas concurrently. Thus, we introduce a novel task called "Semantic Human Contact" prediction. We evaluate on DAMON [58] and report results in Tab. 2; for a finer-grained version of this table, see Sup. Mat. + +To establish a baseline, we adapt the DECO model [58] that detects binary contacts and turn it into a multi-class prediction model, called Semantic-DECO. Due to DAMON's limited training data, this has poor performance. Instead, as discussed in Sec. 4.1, InteractVLM learns effectively from this data, by also leveraging the commonsense of foundation models. Thus, it significantly outperforms Semantic-DECO. Qualitative results reflect this finding; see Fig. 4. Our model captures detailed, accurate contact regions, whereas Semantic-DECO often highlights false-positive areas that differ from the actual contact regions. + +# 4.3. Object Affordance Prediction + +We evaluate the performance of our InteractVLM model on predicting object affordances. This involves identifying regions on objects of possible contact for a certain interaction intent, such as "sitting" on a chair or "moving" it. We train and evaluate on the PIAD [65] dataset. We compare against SotA methods, and report results in Tab. 3. + +Note that we evaluate on object instances that are both seen during training ("PIAD-Seen" column) and unseen ("PIAD-Unseen" column). Our model demonstrates a significant improvement over SotA methods in terms of similarity (SIM), area under the curve (AUC), and mean absolute error (MAE) for both seen and unseen objects. Thus, our method is effective not only for estimating human contact (Secs. 4.1 and 4.2), but also for object affordances. + +# 4.4. Reliance on 3D Annotations + +To analyze InteractVLM's efficiency regarding 3D supervision, we train multiple versions of it with varying amounts of training data from the DAMON dataset, and report performance in Fig. 5. Remarkably, InteractVLM achieves an F1 score of 0.53 with just $1\%$ of the data, nearly matching the 0.55 F1 score of DECO that uses $100\%$ of the data. With + +![](images/ae3586bc333315df4653daa80ee9ae4c9d0b7a81f1b8a9e4c4ed20a903dcc502.jpg) + +![](images/e6396a9845e6ff0d9091ac19b8b8f42330c43179719d50c8236bf6f54a044361.jpg) + +![](images/488c4b08b65e2406a8bf825dfa745f760286402b130aedb6eace06fbeeee00dd.jpg) + +![](images/782f998cf68e81f129817bf575cceb17990561b5d132fd5164ac64d13c57ed9a.jpg) +Input Image + +![](images/f5feea365bfeab32fee4c3cd916173581b8b1ae1f4179e856f3d36aef9732eae.jpg) + +![](images/b0fd1a649fdeef1f8139e8d48c3ff1fa3496ffb7c1e2d633ae1e56cdee89cca4.jpg) + +![](images/c941e71980b6d4dcf76bcdd82b049125dfef6a72550a55b0ac897dfb15bf7331.jpg) + +![](images/03ccf57b1801d4a4f931d16e58d07e348b7f0052875ba97e67e0fd192e263aa2.jpg) +Figure 4. "Semantic Human Contact" estimation (Sec. 4.2). Given an image and an object label, InteractVLM infers body contacts for this object. InteractVLM outperforms a Semantic-DECO [58] baseline. Objects are shown in green circles, and contacts as red patches. + +![](images/ddb8cb87ee440601057e4ea4f735759f0301a155909ff4924f6d932bc60eaab2.jpg) + +![](images/9eadf1322ae1a0d7532202d5fee9b7fd05a2375cb8e93de9e98e94b88ad1ade2.jpg) + +![](images/9b3abd25bf064fc3e6c042acfeab68e1f1659b537c52aae15eced057ea7db150.jpg) + +![](images/c2e08016fd73f4956fe8c273d08b8dacceb73284e72491085d77e9ccce4be8bc.jpg) + +![](images/3602418ca5a1d4f427db3b2eb7564cd181134080c8830860d6d840285212cb6f.jpg) + +![](images/380edd433ade8b0e5f53a20501012cc525d94e0c547617f0fad71e36677d8f27.jpg) + +![](images/5c136ece426d8e4e35c646eb7b4f6bed244501159ae986d13164a3052bfe683f.jpg) + +![](images/2194b3977f64dc86e4cd2ee1d51ea06aac6f4eea7eeb8a5b550862155c5b8175.jpg) + +![](images/a7f5c18cbde0f0fa7c41594bc05acda098e97f9a9a8078ac60cc37c194b5b205.jpg) + +![](images/bba2b5ec99a4344007b7dfdf30ccafd035cbf373cbed7d00999c937faa812aaf.jpg) + +![](images/842c087a139a155405ca782dfc93e5e1a14661104284aad7ca9b709d417215c2.jpg) + +![](images/d8639aa8775480f870d1b6cfd2d9e0f259357d7936d88275276fb64f479db797.jpg) + +![](images/01dde2f85ab79a24f2fe7065d544a5e6c0ee6fd5d7a5de167c4a40a088893698.jpg) + +![](images/4bec1acbe6593a4079b2817b57d6d3c76de2db2d6e11cfa67ea635c58ab16fbd.jpg) + +![](images/39fa99a62a82437025f3b6ea4603407b4136cf7a794df8d3c136f294d27483b6.jpg) +Input Image + +![](images/041a58c9585b777d5b7cfc1af9653f4eae9535cbc7bfb545cbb7efe4be08d9e2.jpg) + +![](images/6189a670508b95ed1025cb971c9ddd42dbb984fb377b6fc5a43a5d7dd4ab5582.jpg) + +![](images/968734f81b7e044425f2d3aea4334f75bf3c4adf3ff8b1d33a04cea323a73c60.jpg) + +![](images/64b47061a7b0198615a75e730e7f84ce69950cf9efb2f4b20c7cb971a74adf76.jpg) + +![](images/fcad1cbfccd8ee84f038800b43cec1a7098abe02777ea774fdcedf1bcf928f19.jpg) + +![](images/ab057e6559b68066b739823a27bebbc8a93c4b6da9d715d3da7084fb6b6c36e4.jpg) + +![](images/89b56d0c9b3ce24e61e652a14889f8ad2792d208a66186b9bf52092f46252f9c.jpg) + +![](images/f35163338ec24ef83196f048b9938eb7c5a7fee3a953c5524ecbfe7f60341a5e.jpg) + +![](images/a0b328167017e295394af5f8583b28f953da9109f63f0afcc647f41a65b56cb3.jpg) + +![](images/c8bd52f57a6a08848e792a86b340e9ee6692973aea46f81abd90a5b865b459ca.jpg) + +![](images/498c583f98d71f172e17400a4161c569ddd71d608e2fcc1cf005ae3ef6b6d5c6.jpg) + +![](images/7e19db98fc810890fb3cfb9a0fc8eb2a7c9e828bbc8f463b1db2159a35e3e375.jpg) + +![](images/2e90b1585752b80c7af0b278fc5849e11e371d1a91ca3cf429f7b7e5003238ec.jpg) + +![](images/d3e51179f7b70427e053a041cf10081c080215e4fca26daa97fa8bc1bc88a19f.jpg) +Figure 5. InteractVLM's reliance on 3D annotations. We evaluate performance for "binary human contact" (F1 score, Y-axis) for models trained on a varying percentage of DAMON [58] training data (X-axis). The DECO baseline trains on $100\%$ of DAMON. Instead, InteractVLM trains on a varying (smaller) portion of this dataset. Yet, it achieves a significantly higher performance, by leveraging the broad visual knowledge of foundation models. + +only $5\%$ of the data, InteractVLM surpasses DECO, reaching an F1 score of 0.58. This performance gap widens as training data increases, ultimately achieving 0.75 F1 score with $100\%$ of the data. This is a compelling finding, highlighting InteractVLM's efficient use of 3D supervision by leveraging the rich visual understanding of foundation models. Note that the strong performance with limited training data has high practical value, because obtaining 3D annotations is expensive. + +Ablations. For a study on the influence of InteractVLM components, including mask resolution, MV-Loc variants, loss functions, training data, the influence of the VLM, and the effect of text prompts, please refer to Sup. Mat. + +# 5. Joint Human-Object Reconstruction + +We demonstrate the usefulness of 3D contacts inferred by InteractVLM for reconstructing a 3D human and object in interaction from a single in-the-wild image, $I$ . + +Initializing 3D body pose & shape, object shape. We use OSX [39] to estimate a 3D SMPL-X body mesh, $H$ , and OpenShape [42] to retrieve a 3D object mesh, $O$ , from the Objaverse [11] database that best matches the image. + +Initializing 3D object pose. We apply InteractVLM on image $I$ to predict 3D body and object contact vertices, $C^H, C^O$ . Then, we solve for object pose, $\{R^O, t^O\}$ by applying the ICP [6] algorithm to snap the 3D object contact points, $C^O$ , onto body ones, $C^H$ . To avoid false correspondences, the 3D normals of contact points must be compatible, i.e., should have similar angles but opposite directions. + +Optimizing 3D object pose. Given the above initialization, we optimize over object rotation, $R^O$ , translation, $t^O$ , and scale, $s^O$ , via render-and-compare by minimizing: + +$$ +E = E _ {M} + \lambda_ {C} E _ {C}, \tag {7} +$$ + +$$ +E _ {M} = I o U (\widehat {M}, M) + \left\| \widehat {M ^ {c}} - M ^ {c} \right\| _ {2}, \tag {8} +$$ + +$$ +E _ {C} = \frac {1}{| C ^ {H} \| C ^ {O} |} \sum_ {i \in | H |} \sum_ {j \in | O |} C _ {i} ^ {H} C _ {j} ^ {O} \| V _ {i} ^ {H} - V _ {j} ^ {O} \| _ {2}, \tag {9} +$$ + +where $E_{M}$ is a mask loss, $E_{C}$ is a contact loss, $IoU$ is intersection-over-union, $\widehat{M}, M$ are hypothesis and ground-truth masks, respectively, $\widehat{M^{c}}, M^{c}$ are the corresponding + +![](images/23806b3f4592cf7a9f13aacbf01bfdd8d1d1b5e0c1c87280ad3fb8c664973e5b.jpg) +Figure 6. 3D HOI reconstruction (Sec. 5). We build an optimization method that fits a SMPL-X body and OpenShape-retrieved object to an in-the-wild image. We evaluate against the SotA method PHOSA [67]. Reconstruction is guided by InteractVLM-inferred contacts. + +mean mask pixels, $|C^H|$ , $|C^O|$ are the number of contact vertices on the human and object, $|H|$ , $|O|$ are the number of mesh vertices, $V_i^H$ , $V_j^O$ are the $i$ -th and $j$ -th human and object vertices, while $C_i^O$ , $C_j^H$ indicate whether vertices are in contact or not. Given the image, we extract the object mask, $M$ , via SAM [31], and a depth map, $D$ , via Depth Pro [8]. + +Intuitively, we use the $E_{M}$ loss to align the 3D object to the image, while in $E_{C}$ the predicted 3D contacts "anchor" the 3D object onto the body so they interact realistically. We perform iterative optimization via Adam [30]. OSX produces reasonable bodies, so we keep these fixed. The object is updated in each iteration - we render depth, $\widehat{D}$ , and masks, $\widehat{M}$ , via a PyTorch3D [52] differentiable renderer. + +Qualitative Results. We reconstruct 3D human-object interaction from an image. We show results in Fig. 6, and compare with PHOSA [67], the most related SotA method. InteractVLM's reconstructions look more realistic. Note that PHOSA uses handcrafted contacts on humans and bodies. Instead, InteractVLM infers 3D contacts on both human bodies and objects from the single in-the-wild image. These play a crucial role for guiding 3D reconstruction under occlusions and depth ambiguities. + +Perceptual Study. There exists no in-the-wild dataset with 3D ground truth for HOI, so we conduct a perceptual study via Amazon Mechanical Turk for evaluation. Specifi + +cally, we evaluate the realism of our reconstructions against ones of PHOSA. We randomly select 55 images for which PHOSA has handcrafted contact annotations. For each image, participants are shown (with random swapping) reconstructions generated by our method and by PHOSA, and are asked to select the one that best represents the image. Our reconstructions are preferred $62\%$ of the time. + +# 6. Conclusion + +We develop InteractVLM, a novel method for estimating 3D contacts on both humans and objects from a single natural image. InteractVLM has a reduced reliance on expensive 3D contact annotations for training, by leveraging the broad knowledge of Vision-Language Models. Specifically, we introduce a novel "Render-Localize-Lift" (RLL) framework and a novel multi-view localization model (MV-Loc) to adapt 2D foundation models for 3D contact estimation. We outperform existing work on contact estimation and introduce a new "Semantic Human Contact" estimation task for inferring body contacts conditioned on object labels. This goes beyond traditional binary contact estimation, which fails to capture rich semantic relationships of multi-object interactions. Last, we develop the first approach that uses inferred contact points on both bodies and objects for joint 3D reconstruction from single in-the-wild images. + +Acknowledgments: We thank Alpár Cseke for his assistance with evaluating joint human-object reconstruction. We also thank Tsvetelina Alexiadis and Taylor Obersat for MTurk evaluation, Yao Feng, Peter Kultis and Markos Diomataris for their valuable feedback and Benjamin Pellkofer for IT support. SKD is supported by the International Max Planck Research School for Intelligent Systems (IMPRS-IS). The UvA part of the team is supported by an ERC Starting Grant (STRIPES, 101165317, PI: D. Tzionas). Disclosure: DT has received a research gift fund from Google. For MJB see https://files.is.tue.mpg.de/black/CoI_CVPR_2025.txt + +# References + +[1] Emanuele Aiello, Diego Valsesia, and Enrico Magli. Cross-modal learning for image-guided point cloud shape completion. Conference on Neural Information Processing Systems (NeurIPS), 35:37349-37362, 2022. 6 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: A visual language model for few-shot learning. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 3 +[3] Reza Yazdani Aminabadi, Samyam Rajbhandari, Minjia Zhang, A. A. Awan, Cheng Li, Du Li, Elton Zheng, Jeff Rasley, Shaden Smith, Olatunj Ruwase, and Yuxiong He. DeepSpeed inference: Enabling efficient inference of transformer models at unprecedented scale. International Conference for High Performance Computing, Networking, Storage and Analysis, 2022. 5 +[4] Dimitrije Antić, Georgios Paschalidis, Shashank Tripathi, Theo Gevers, Sai Kumar Dwivedi, and Dimitrios Tzionas. SDFit: 3D object pose and shape by fitting a morphable SDF to a single image. arXiv:2409.16178, 2025. 3 +[5] Fabien Baradel*, Matthieu Armando, Salma Galaoui, Romain Brégier, Philippe Weinzaepfel, Grégory Rogez, and Thomas Lucas*. Multi-HMR: Multi-person whole-body human mesh recovery in a single shot. In European Conference on Computer Vision (ECCV), 2024. 3 +[6] Paul J. Besl and Neil D. McKay. A method for registration of 3D shapes. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 14:239-256, 1992. 7 +[7] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAVE: Dataset and method for tracking human object interactions. In Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[8] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth Pro: Sharp monocular metric depth in less than a second. arXiv:2410.02073, 2024. 8 + +[9] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision (ECCV), 2016. 3 +[10] Honghua Chen, Zeyong Wei, Yabin Xu, Mingqiang Wei, and Jun Wang. ImLoveNet: Misaligned image-supported registration network for low-overlap point cloud pairs. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-9, 2022. 6 +[11] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2023. 2, 3, 4, 7 +[12] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3D AffordanceNet: A benchmark for visual object affordance understanding. In Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[13] John Doe, Jane Smith, and Alice Brown. Paris3d: Language-guided 3d segmentation with multimodal models. arXiv:2406.08394, 2024. 3 +[14] Sai Kumar Dwivedi, Nikos Athanasiou, Muhammed Kocabas, and Michael J. Black. Learning to regress bodies from images using differentiable semantic rendering. In International Conference on Computer Vision (ICCV), 2021. 3 +[15] Sai Kumar Dwivedi, Cordelia Schmid, Hongwei Yi, Michael J. Black, and Dimitrios Tzionas. POCO: 3D pose and shape estimation using confidence. In International Conference on 3D Vision (3DV), 2024. 3 +[16] Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, Yao Feng, and Michael J. Black. TokenHMR: Advancing human mesh recovery with a tokenized pose representation. In Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[17] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual handobject manipulation. In Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[18] Zicong Fan, Maria Parelli, Maria Eleni Kadoglou, Muhammed Kocabas, Xu Chen, Michael J Black, and Otmar Hilliges. HOLD: Category-agnostic 3D reconstruction of interacting hands and objects from video. In Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[19] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Collaborative regression of expressive bodies using moderation. In International Conference on 3D Vision (3DV), 2021. 6 +[20] Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, and Tae-Kyun Kim. First-person hand action benchmark with RGB-D videos and 3D hand pose annotations. In Computer Vision and Pattern Recognition (CVPR), 2018. 3 +[21] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 3 + +[22] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambiguities with 3D scene constraints. In International Conference on Computer Vision (ICCV), pages 2282-2292, 2019. 3 +[23] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14708-14718, 2021. 6 +[24] Yana Hasson, Gül Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[25] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. arXiv: 2311.04400, 2023. 3 +[26] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. International Conference on Learning Representations (ICLR), 2021. 2, 5 +[27] Chun-Hao Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Computer Vision and Pattern Recognition (CVPR), pages 13264-13275, 2022. 6 +[28] Yinghao Huang, Omid Taheri, Michael J. Black, and Dimitrios Tzionas. InterCap: Joint markerless 3D tracking of humans and objects in interaction from multi-view RGB-D images. International Journal of Computer Vision (IJCV), 132(7):2551-2566, 2024. 3 +[29] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In International Conference on 3D Vision (3DV), pages 42-52, 2021. 3 +[30] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 8 +[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloé Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 2, 3, 5, 8 +[32] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. VIBE: Video inference for human body pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), pages 5252-5262, 2020. 3 +[33] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 3 +[34] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: Reasoning segmentation via large language model. In Computer Vision and Pattern Recognition (CVPR), pages 9579-9589, 2024. 3, 4, 5 + +[35] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. HybrIK: A hybrid analytical-neural inverse kinematics solution for 3D human pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), pages 3383-3393, 2021. 3 +[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning (ICML), pages 19730-19742, 2023. 3 +[37] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In European Conference on Computer Vision (ECCV), 2022. 3 +[38] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-resolution text-to-3D content creation. Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[39] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3D whole-body mesh recovery with component aware transformer. In Computer Vision and Pattern Recognition (CVPR), 2023. 3, 7 +[40] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In International Conference on Computer Vision (ICCV), 2021. 3 +[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 3, 5 +[42] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. OpenShape: Scaling up 3D shape representation towards open-world understanding. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 2, 3, 4, 7 +[43] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3D mesh in 45 seconds without pershape optimization. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 3 +[44] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. International Conference on Computer Vision (ICCV), 2023. 3 +[45] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. Transactions on Graphics (TOG), 34 (6):248:1-248:16, 2015. 3 +[46] Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. COAP: Compositional articulated occupancy of people. In Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[47] Hyeongjin Nam, Daniel Sungho Jung, Gyeongsik Moon, and Young Mu Lee. Joint reconstruction of 3D human and object via contact-based refinement transformer. In Computer Vision and Pattern Recognition (CVPR), 2024. 3 + +[48] Mohamed Omran, Christoph Lassner, Gerard Pons-Moll, Peter Gehler, and Bernt Schiele. Neural body fitting: Unifying deep learning and model based human pose and shape estimation. In International Conference on 3D Vision (3DV), 2018. 3 +[49] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975–10985, 2019. 3 +[50] Bingbing Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824, 2023. 3 +[51] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. In International Conference on Learning Representations (ICLR), 2024. 3 +[52] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D deep learning with PyTorch3D. arXiv:2007.08501, 2020. 8 +[53] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. Transactions on Graphics (TOG), 36(6), 2022, 2, 4 +[54] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[55] Qingping Sun, Yanjun Wang, Ailing Zeng, Wanqi Yin, Chen Wei, Wenjia Wang, Haiyi Mei, Chi Sing Leung, Ziwei Liu, Lei Yang, and Zhongang Cai. AiOS: All-in-one-stage expressive human pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[56] István Sárándi and Gerard Pons-Moll. Neural localizer fields for continuous 3D human pose and shape estimation. In Conference on Neural Information Processing Systems (NeurIPS), 2024. 3 +[57] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 3 +[58] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J. Black. DECO: Dense estimation of 3D human-scene contact in the wild. In International Conference on Computer Vision (ICCV), pages 8001-8013, 2023. 2, 3, 5, 6, 7 +[59] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object + +coordinate space for category-level 6D object pose and size estimation. In Computer Vision and Pattern Recognition (CVPR), 2019. 5 +[60] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, and Jifeng Dai. VisionLLM: Large language model is also an open-ended decoder for vision-centric tasks. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 3 +[61] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. CHORE: Contact, human and object reconstruction from a single RGB image. In European Conference on Computer Vision (ECCV), 2022. 3 +[62] Xianghui Xie, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Template free reconstruction of human-object interaction with procedural interaction generation. In Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[63] Danfei Xu, Dragomir Anguelov, and Ashesh Jain. PointFusion: Deep sensor fusion for 3D bounding box estimation. In International Conference on Computer Vision (ICCV), pages 244-253, 2018. 6 +[64] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, Bill Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[65] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3D object affordance from 2D interactions in images. In International Conference on Computer Vision (ICCV), pages 10871-10881, 2023. 2, 3, 4, 5, 6 +[66] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, and Zheng-Jun Zha. LEMON: Learning 3D human-object interaction relation from 2D images. In Computer Vision and Pattern Recognition (CVPR), 2024. 3, 5, 6 +[67] Jason Y. Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3D human-object spatial arrangements from a single image in the wild. In European Conference on Computer Vision (ECCV), 2020. 3, 8 +[68] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. GPT4RoI: Instruction tuning large language model on region-of-interest. arXiv:2307.03601, 2023. 3 +[69] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3D: Exploring unified 3D representation at scale. In International Conference on Learning Representations (ICLR), 2024. 3 +[70] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3D lidar semantic segmentation. In International Conference on Computer Vision (ICCV), 2021. 6 \ No newline at end of file diff --git a/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/images.zip b/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9ae1562c8bda713d9128b9f5f681cb991989b399 --- /dev/null +++ b/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81ba67558b5bf374efc8df9feb0cece0f21c9a3b6399b7a80275ff04d4504f53 +size 657342 diff --git a/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/layout.json b/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..833d32bc341ab6e8c8b15f28cd877ed4a7991603 --- /dev/null +++ b/2025/InteractVLM_ 3D Interaction Reasoning from 2D Foundational Models/layout.json @@ -0,0 +1,11295 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 93, + 526, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 93, + 526, + 110 + ], + "spans": [ + { + "bbox": [ + 85, + 93, + 526, + 110 + ], + "type": "text", + "content": "InteractVLM: 3D Interaction Reasoning from 2D Foundational Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 127, + 493, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 127, + 493, + 156 + ], + "spans": [ + { + "bbox": [ + 111, + 127, + 493, + 156 + ], + "type": "text", + "content": "Sai Kumar Dwivedi1 Dimitrije Antić2 Shashank Tripathi1 Omid Taheri1 Cordelia Schmid3 Michael J. Black1 Dimitrios Tzionas2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 165, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 165, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 63, + 165, + 547, + 194 + ], + "type": "text", + "content": "1Max Planck Institute for Intelligent Systems, Tübingen, Germany 2University of Amsterdam (UvA), the Netherlands 3Inria, École normale supérieure, France" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 209, + 548, + 359 + ], + "blocks": [ + { + "bbox": [ + 63, + 209, + 548, + 359 + ], + "lines": [ + { + "bbox": [ + 63, + 209, + 548, + 359 + ], + "spans": [ + { + "bbox": [ + 63, + 209, + 548, + 359 + ], + "type": "image", + "image_path": "9406fdef2334f1a5f620b804e6bdddfa4e1a21b853286f21a008a8517cc7ce8a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 361, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 54, + 361, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 555, + 396 + ], + "type": "text", + "content": "Figure 1. We present InteractVLM, a novel method for estimating contact points on both human bodies and objects from a single in-the-wild image, shown here as red patches. Our method goes beyond traditional binary contact estimation methods by estimating contact points on a human in relation to a specified object. We do so by leveraging the broad visual knowledge of a large Visual Language Model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 411, + 200, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 411, + 200, + 423 + ], + "spans": [ + { + "bbox": [ + 151, + 411, + 200, + 423 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 437, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 297, + 714 + ], + "type": "text", + "content": "We introduce InteractVLM, a novel method to estimate 3D contact points on human bodies and objects from single in-the-wild images, enabling accurate human-object joint reconstruction in 3D. This is challenging due to occlusions, depth ambiguities, and widely varying object shapes. Existing methods rely on 3D contact annotations collected via expensive motion-capture systems or tedious manual labeling, limiting scalability and generalization. To overcome this, InteractVLM harnesses the broad visual knowledge of large Vision-Language Models (VLMs), fine-tuned with limited 3D contact data. However, directly applying these models is non-trivial, as they reason \"only\" in 2D, while human-object contact is inherently 3D. Thus we introduce a novel \"Render-Localize-Lift\" module that: (1) embeds 3D body and object surfaces in 2D space via multiview rendering, (2) trains a novel multi-view localization model (MV-Loc) to infer contacts in 2D, and (3) lifts these to 3D. Additionally, we propose a new task called Semantic Human Contact estimation, where human contact predictions are conditioned explicitly on object semantics, enabling richer interaction modeling. InteractVLM outperforms existing work on contact estimation and also facilitates 3D reconstruction from an in-the-wild image. To es" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 411, + 556, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 484 + ], + "type": "text", + "content": "timate 3D human and object pose, we infer initial body and object meshes, then infer contacts on both of these via InteractVLM, and last exploit these for fitting the meshes to image evidence. Results show that our approach performs promisingly in the wild. Code and models are available at https://interactvlm.is.tue.mpg.de." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 510, + 395, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 510, + 395, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 510, + 395, + 523 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 592 + ], + "type": "text", + "content": "People interact with objects routinely. Reconstructing human-object interaction (HOI) in 3D is key for many applications, from robots to mixed reality. However, doing so from single images is challenging due to depth ambiguity, occlusions, and the diverse shape and appearance of objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "There are methods that estimate 3D human bodies and methods that estimate 3D objects, but few that put these together. Knowing the contacts between these could significantly improve joint reconstruction. Our goal is to infer contact points on both humans and objects from single in-the-wild images, and then use these contacts for jointly reconstructing humans and objects. However, there is a lack of in-the-wild training images paired with ground-truth contact labels for both 3D humans and objects. Acquiring such data is challenging and existing methods do not scale." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22605" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "The problem gets more challenging due to the complexity of real-world interactions. Humans often contact multiple objects simultaneously; e.g., using a laptop while sitting on a club chair. Yet, current approaches treat contact prediction as simple binary classification; i.e. detecting whether a body part is in contact with \"any\" object. This simplified assumption fails to capture the rich semantic relationships of multi-object interactions. To address this, we introduce a novel \"Semantic Human Contact\" estimation task. This involves predicting the contact points on the body related to a particular object given a single in-the-wild image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 324 + ], + "type": "text", + "content": "To tackle this, as well as to overcome data scarcity, we propose a new paradigm for scaling and reasoning in the wild. Specifically, we observe that large Vision-Language Models (VLMs) can \"reason\" about in-the-wild images because they are trained on internet-scale data and possess a broad visual knowledge about humans and their interactions with the world. We also observe that this knowledge can be re-purposed for novel tasks by fine-tuning these large models on small datasets. Thus, we exploit VLMs for developing a novel framework, called InteractVLM." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 326, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 295, + 397 + ], + "type": "text", + "content": "At the heart of InteractVLM lies a reasoning module based on a VLM; see Fig. 2. We enrich this with skills for 3D human-object \"understanding,\" by extending the VLM with a LoRA [26] adaptation layer. As a result, given a color image, this module can be \"asked\" to produce \"reasoning tokens\" that facilitate 3D contact localization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 399, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 495 + ], + "type": "text", + "content": "However, exploiting these tokens to localize contact is non-trivial. A natural choice would be to employ a foundational \"localization\" model [31] that takes these tokens as \"guidance,\" and highlights the 3D contacts. But there exists a key practical problem. Existing foundation models inherently operate only in 2D space, while we need them to do so in 3D. To tackle this, we need to re-cast our problem so it is appropriate for 2D foundation models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "content": "To this end, we develop a novel \"Render-Localize-Lift\" (RLL) framework that has three main steps; see Fig. 2: (1) We render the 3D shape of a canonical SMPL+H [53] body and an object as 2D images from multiple viewpoints. Regarding the object, a 3D mesh is efficiently retrieved from a large-scale 3D database [11] through OpenShape [42]. (2) We pass the above images into a foundation model to predict 2D contact masks for both the body and the object. (3) We lift the predicted 2D contact points to 3D points via back-projection, i.e., performing the inverse of the first step." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 712 + ], + "type": "text", + "content": "However, even after recasting the problem to 2D by rendering multi-view images, existing foundation models are still not 3D aware, i.e. they treat each view independently, ignoring multi-view consistency. This means contact detections in one view do not necessarily agree with ones in adjacent views. To tackle this, just appending camera parameters to multi-view renderings is insufficient. Instead, we build a novel \"Multi-view Localization\" model that we call" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 321, + 73, + 548, + 177 + ], + "blocks": [ + { + "bbox": [ + 321, + 73, + 548, + 177 + ], + "lines": [ + { + "bbox": [ + 321, + 73, + 548, + 177 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 548, + 177 + ], + "type": "image", + "image_path": "41285023f276fd719bccc053beeb3d443a9efa6a33aadaf318560de488cc166c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 181, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 225 + ], + "type": "text", + "content": "Figure 2. Overview of InteractVLM. Given a color image, our VLM performs the core reasoning, and guides a novel MV-Loc model to localize contacts on both bodies and objects in 3D. Here we show only the body; for details, and object contact, see Fig. 3." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 247, + 553, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 306 + ], + "type": "text", + "content": "MV-Loc. This has two steps: (1) It transforms the \"reasoning token\" provided by the VLM with the camera parameters used to render the multi-view images. (2) It ensures multi-view consistency, by lifting the inferred 2D contacts in each view to 3D and computing a 3D loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 307, + 553, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 426 + ], + "type": "text", + "content": "Our method, InteractVLM, utilizes a VLM in tandem with our novel multi-view localization model, MV-Loc, to perform 3D contact prediction for humans and objects. See Fig. 1 for some examples. We quantitatively evaluate the efficacy of our method for in-the-wild 3D contact prediction on bodies and on objects, using the DAMON [58] and PIAD [65] datasets, respectively. For bodies, we evaluate both for the traditional \"binary contact\" estimation, and for our new task of \"semantic contact\" estimation. For all tasks, we find that InteractVLM outperforms the prior work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 545 + ], + "type": "text", + "content": "Finally, we demonstrate how InteractVLM's estimated contact improves 3D HOI recovery from in-the-wild images; this is a highly ill-posed task due to depth ambiguities and occlusions. To address this, we develop an optimization-based method that fits a SMPL-X body mesh and an OpenShape-retrieved object mesh to the image, using InteractVLM's inferred contacts as constraints to anchor human and object meshes w.r.t. each other. To the best of our knowledge, this is the first approach for estimating 3D HOI for in-the-wild images using inferred contacts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 546, + 511, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 546, + 511, + 557 + ], + "spans": [ + { + "bbox": [ + 326, + 546, + 511, + 557 + ], + "type": "text", + "content": "In summary, we make four key contributions:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 315, + 558, + 553, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 558, + 553, + 593 + ], + "spans": [ + { + "bbox": [ + 315, + 558, + 553, + 593 + ], + "type": "text", + "content": "1. We build InteractVLM, a novel method that facilitates HOI reconstruction from an in-the-wild image by detecting 3D contacts on both bodies and objects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 594, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 553, + 628 + ], + "type": "text", + "content": "2. We demonstrate a way to minimize reliance on 3D contact annotations via exploiting the broad visual knowledge of Vision-Language Models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 630, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 630, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 314, + 630, + 553, + 664 + ], + "type": "text", + "content": "3. We build a novel \"Multi-view Localization\" model that helps in estimating contacts in 3D by transforming the reasoning of foundation models from 2D to 3D." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "4. We introduce the novel \"Semantic Human Contact\" task for inferring body contacts conditioned on object labels. Our code and trained models are available for research at https://interactglm.is.tue.mpg.de." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22606" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 225, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 225, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 225, + 104 + ], + "type": "text", + "content": "2.1. Large Vision-Language Models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 110, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 312 + ], + "type": "text", + "content": "Recent advancements in large language models (LLMs) have led to the development of multimodal models that integrate vision and language reasoning. Models like Flamingo [2] and BLIP-2 [36] use cross-attention mechanisms and visual encoders to align image features with text, supporting a variety of vision-language tasks. More recent works, such as VisionLLM [60] and Kosmos-2 [50], use grounded image-text data to enhance spatial understanding, while GPT4RoI [68] introduces spatial box inputs for finer alignment. However, these models typically lack end-to-end segmentation capabilities. To address this limitation, LISA [34] combines vision foundation segmentation models like SAM [31] with multimodal embeddings, enabling language-guided segmentation. PARIS3D [13] extends this approach to referential 3D segmentation by processing multi-view object renders through both SAM and LLaVA [41] for spatially-aware segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 314, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 295, + 456 + ], + "type": "text", + "content": "Taking inspiration from these approaches, we exploit a language-guided segmentation model for the task of predicting human and object contact in 3D. However, unlike PARIS3D, we process a single RGB image with LLaVA and multi-view renders of the human mesh and the object mesh with SAM. Moreover, we introduce a feature-lifting technique that extends LLaVA's 2D features into 3D using camera parameters, thus guiding SAM's multi-view segmentations. This approach ensures multi-view consistency and efficiently predicts contact affordances, extending the use of multimodal models in 3D reasoning tasks focused on human-object interaction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 467, + 278, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 278, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 278, + 480 + ], + "type": "text", + "content": "2.2. 3D Human and Object from Single Images" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": "Estimating 3D human pose and shape from a single image has evolved from optimization-based methods to learning-based approaches. Optimization-based methods fit parametric body models like SMPL [45], SMPL-X [49], or GHUM [64] to 2D cues such as keypoints [9], silhouettes [48], or segmentation masks [48]. Learning-based methods either regress body parameters from images or videos [5, 14, 15, 29, 32, 33, 37] or estimate non-parametric bodies as vertices [35, 40], implicit surfaces [46, 54], or dense points [56]. Transformer-based methods [16, 21, 39, 55] have further improved robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "For 3D object reconstruction from a single image, regression-based methods predict geometry using meshes, voxels, or point clouds. Diffusion-based models [25] utilize large 3D datasets like Objaverse [11] or 2D diffusion models [38, 43, 44, 51] to guide reconstruction and optimization methods [4] use render and compare. Retrieval-based methods, such as OpenShape [42] and Uni3D [69], have demonstrated some robustness in cases with occlusions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 479, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 479, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 479, + 85 + ], + "type": "text", + "content": "2.3. 3D Human-Object Interaction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 89, + 553, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 233 + ], + "type": "text", + "content": "Understanding 3D human-object interactions is essential for modeling realistic scenes. Early works focused on handobject interactions, such as ObMan [24] and FPHA [20], with more recent studies like ARCTIC [17] and HOLD [18] providing more detailed data and reconstruction for hands. For full-body interactions, initial studies involve interactions with scenes, as in PROX [22], and with objects, as in BEHAVE [7], GRAB [57], and InterCap [28]. BEHAVE and GRAB use motion capture that are accurate but nonscalable, while InterCap uses multi-camera setups that are more scalable but less accurate. Both of these approaches are limited in capturing diverse and realistic interactions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 233, + 553, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 553, + 484 + ], + "type": "text", + "content": "As a proxy for 3D reconstruction, the recent DECO [58] method infers contacts on body meshes; to this end, 3D contact annotations are crowdsourced via mesh painting. Predicting contact on objects is more challenging due to varying object shapes, while also currently there exists no dataset of object contact for in-the-wild images. Thus, we estimate 3D object affordances as a proxy for contact. 3D-AffordanceNet [12] introduces affordances that are not grounded in images, capturing the likelihood of humans interacting with specific parts of an object for a given affordance (e.g., \"sit on chair\"). PIAD [65] curates RGB images depicting object affordances and trains a network to estimate them. LEMON [66] extends object affordance prediction of PIAD to include human contact estimation. However, these methods require human contact vertices paired with object affordances for training, limiting the number of categories they handle to 21 ones. In contrast, we learn from unpaired human and object interaction data and perform joint interaction reasoning via a cross-attention layer, enabling human interaction reasoning for 80 categories and object affordance prediction for 32 categories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 492, + 524, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 492, + 524, + 506 + ], + "spans": [ + { + "bbox": [ + 314, + 492, + 524, + 506 + ], + "type": "text", + "content": "2.4. Joint 3D Human-Object Reconstruction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 653 + ], + "type": "text", + "content": "Joint reconstruction of humans and objects in 3D has been tackled with both regression and optimization methods. Regression-based methods directly predict 3D human-object meshes, as in HDM [62] and CONTHO [47], while other methods first predict contact points and then use test-time optimization to fit human and object meshes, as in CHORE [61] and PHOSA [67]. Since regression methods rely on limited training data, optimization methods are preferred for in-the-wild scenarios, such as PHOSA [67]. Optimization-based methods either assume known contacts or infer contacts to fit meshes to the image, but their success heavily relies on the quality of contact." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Our method improves upon these by providing more accurate contact predictions, which in turn facilitate better fitting of human and object meshes to image evidence, improving the realism and accuracy of 3D human-object reconstructions from single images." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22607" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 446, + 264 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 446, + 264 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 446, + 264 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 446, + 264 + ], + "type": "image", + "image_path": "4c7e3284f279be476c0109e0c6cc5a8b798e40e2af9aaaf8c482c32c26185e0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": "Figure 3. Method overview. Given a single in-the-wild color image, our novel InteractVLM method estimates 3D contact points on both humans and objects (a). Then, we reconstruct a 3D human and object in interaction by exploiting these contacts (b). More specifically: (a) Contact estimation. Given an image, " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": ", and prompt text, " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "T_{inp}" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": ", our VLM, " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": ", produces contact tokens for humans and objects, " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "<\\mathsf{HCON}>" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "<0\\mathsf{CON}>" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": ", which are projected (" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": ") into feature embeddings, " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "E^H" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "E^O" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": ". These guide a \"Multi-View [contact] Localization\" model. This renders the 3D human and object geometry via cameras, " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": ", into multi-view 2D renders and passes these to encoder, " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": ", while decoders, " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\Omega^H" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\Omega^O" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": ", estimate and highlight 2D contacts in these renders. Then, the FeatLift module, " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": ", transforms the VLM's features " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "(E^H, E^O)" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": " to become 3D-aware " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "(E_{3D}^H, E_{3D}^O)" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": " by exploiting the camera parameters, " + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 273, + 555, + 361 + ], + "type": "text", + "content": ". A final module lifts the detected 2D contacts to 3D. (b) 3D HOI reconstruction. For joint human-object reconstruction, we use InteractVLM's inferred contacts in an optimization framework." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 455, + 70, + 547, + 266 + ], + "blocks": [ + { + "bbox": [ + 455, + 70, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 455, + 70, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 455, + 70, + 547, + 266 + ], + "type": "image", + "image_path": "5974398eb285b904d1de43c22f78dcae6534bc3da026842cc044de35856ce37e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 380, + 111, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 111, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 111, + 392 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 401, + 178, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 178, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 178, + 414 + ], + "type": "text", + "content": "3.1. Input Representation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 419, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 295, + 443 + ], + "type": "text", + "content": "Given an image, " + }, + { + "bbox": [ + 55, + 419, + 295, + 443 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 419, + 295, + 443 + ], + "type": "text", + "content": ", InteractVLM estimates 3D contacts for both human bodies and objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 445, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 493 + ], + "type": "text", + "content": "The human is represented by a SMPL+H [53] 3D body mesh, " + }, + { + "bbox": [ + 55, + 445, + 295, + 493 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 445, + 295, + 493 + ], + "type": "text", + "content": ", with vertices " + }, + { + "bbox": [ + 55, + 445, + 295, + 493 + ], + "type": "inline_equation", + "content": "V^{H} \\in \\mathbb{R}^{10475 \\times 3}" + }, + { + "bbox": [ + 55, + 445, + 295, + 493 + ], + "type": "text", + "content": ". The body is posed in a canonical \"star\" shape (see details in Sec. 3.4). The human contacts are binary per-vertex labels, " + }, + { + "bbox": [ + 55, + 445, + 295, + 493 + ], + "type": "inline_equation", + "content": "C^{H} \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 445, + 295, + 493 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "type": "text", + "content": "The object is represented by a 3D point cloud (or mesh), " + }, + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "type": "inline_equation", + "content": "O \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "type": "text", + "content": " points. As there are no datasets of natural images paired with 3D contacts for objects, we use a largescale 3D affordance dataset [65], instead, as affordances are closely related to contact. Specifically, they represent the likelihood of contact on 3D object areas for various purposes. Therefore, for objects we use the terms \"affordance\" and \"contact\" interchangeably. The object contacts are continuous per-point values, " + }, + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "type": "inline_equation", + "content": "C^O \\in [0,1]" + }, + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "type": "text", + "content": ". During inference, we retrieve a 3D object shape from a large database [11] via OpenShape [42] conditioned on image " + }, + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 635, + 201, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 201, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 201, + 647 + ], + "type": "text", + "content": "3.2. Overview of InteractVLM" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "The biggest challenge for learning 3D contact prediction in the wild is the limited 3D contact data for humans and objects. To go beyond existing limited datasets, we introduce a novel method, called InteractVLM, that harnesses the commonsense knowledge of large vision-language models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "type": "text", + "content": "Specifically, InteractVLM (Fig. 3) has two main components: a Vision Language Model (VLM) and a novel Multi-View contact Localization model (MV-Loc). MV-Loc highlights parts that are in contact for both humans and objects with the VLM's guidance. The input to the VLM (Sec. 3.3) is an image, " + }, + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "type": "text", + "content": ", and a prompt, " + }, + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "type": "inline_equation", + "content": "T_{inp}" + }, + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "type": "text", + "content": ", that asks the VLM to detect contact. The input to MV-Loc (Sec. 3.4) is the 3D geometry of humans and objects, " + }, + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 499, + 509, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 509, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 509, + 512 + ], + "type": "text", + "content": "3.3. Interaction Reasoning through VLM" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": "The VLM module, " + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": ", conducts the core interaction reasoning. It takes an input image, " + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": ", and prompt text, " + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "inline_equation", + "content": "T_{inp}" + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": ", and outputs a prompt text, " + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "inline_equation", + "content": "T_{out} = \\Psi(I, T_{inp})" + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": ". Inspired by the recent LISA [34] model, we expand the VLM's vocabulary with two specialized tokens, " + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\langle \\text{HCON} \\rangle" + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\langle \\text{OCON} \\rangle" + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": ", for contact information for humans and objects, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "text", + "content": "To denote contact, " + }, + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "text", + "content": " produces a prompt that includes the above tokens. To aid MV-Loc in localizing contact, we extract the last-layer embeddings of the VLM corresponding to these tokens and pass them through a projection layer, " + }, + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "text", + "content": ", to obtain the feature embeddings, " + }, + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "inline_equation", + "content": "E^H" + }, + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "inline_equation", + "content": "E^O" + }, + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "text", + "content": " be the ground-truth text, and " + }, + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{pred}}" + }, + { + "bbox": [ + 313, + 597, + 555, + 682 + ], + "type": "text", + "content": " be the predicted one. Then, our token-prediction loss is defined as a cross-entropy loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 367, + 700, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 700, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 367, + 700, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o k e n}} = - \\sum_ {i = 1} ^ {N} \\left(T _ {\\mathrm {g t}} ^ {(i)} \\cdot \\log \\left(T _ {\\text {p r e d}} ^ {(i)}\\right)\\right). \\tag {1}", + "image_path": "b8aaa223bf7310e0e42cc37acb604709f5c2bf07c3e712826a462538595e1c0e.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22608" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 272, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 272, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 272, + 85 + ], + "type": "text", + "content": "3.4. Interaction Localization through MV-Loc" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "text", + "content": "We develop a novel MV-Loc module that has a shared image encoder, " + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "text", + "content": ", and separate decoders, " + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "inline_equation", + "content": "\\Omega^H" + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "inline_equation", + "content": "\\Omega^O" + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "text", + "content": ", for humans and objects. MV-Loc performs contact localization by using a novel \"Render-Localize-Lift\" (RLL) framework. To this end, it takes three steps; (1) rendering the 3D shape of humans and objects in 2D, (2) predicting 2D contact maps for both of these, and (3) lifting the 2D contact maps to 3D." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "text", + "content": "RLL step #1: Render 3D " + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "text", + "content": " 2D. The input is human and object geometry, namely " + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "text", + "content": ", respectively. Both serve as a \"canvas\" for \"painting\" detected contacts on these. The body has a default SMPL+H shape in a canonical star-shape pose to minimize self-occlusions when rendering. The object geometry (initialized in Sec. 3.1) is normalized to a unit sphere. Each geometry is rendered from " + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "text", + "content": " fixed views with camera parameters, " + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "text", + "content": ", to form multi-view renderings, " + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "inline_equation", + "content": "R^{H,O} = \\{R_j\\}_{j=1}^J" + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "text", + "content": ", so that the entire 3D geometry is captured. Since our geometries do not have texture, we color the meshes with normals and point clouds using the NOCS map [59]. This enhances cross-view correspondence, making renderings resemble real images to image encoder, " + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 174, + 296, + 329 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "text", + "content": "RLL step #2: Localize in 2D. The rendered geometry, " + }, + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "inline_equation", + "content": "R^{H,O}" + }, + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "text", + "content": ", is first sent to the image encoder, " + }, + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "text", + "content": ", and then passed to decoders, so that the final contact masks, " + }, + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "inline_equation", + "content": "M^H" + }, + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "inline_equation", + "content": "M^O" + }, + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "text", + "content": ", get highlighted on it. However, MV-Loc requires spatial and contextual cues for highlighting the contact region. To this end, we use the feature embeddings (Sec. 3.3), i.e. " + }, + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "inline_equation", + "content": "E^H" + }, + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "inline_equation", + "content": "E^O" + }, + { + "bbox": [ + 55, + 330, + 296, + 413 + ], + "type": "text", + "content": " to guide the contact localization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 414, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 474 + ], + "type": "text", + "content": "However, since the VLM reasons in 2D, these features are not 3D aware. This is a problem, because MV-Loc needs 3D awareness to localize contact consistently across multi-view renderings. Thus, we transform the features to \"lift\" them to \"3D\" to better guide multi-view localization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 296, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 512 + ], + "type": "text", + "content": "In detail, we design a lifting network, " + }, + { + "bbox": [ + 55, + 474, + 296, + 512 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 474, + 296, + 512 + ], + "type": "text", + "content": ", which takes camera parameters, " + }, + { + "bbox": [ + 55, + 474, + 296, + 512 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 474, + 296, + 512 + ], + "type": "text", + "content": ", and the 2D " + }, + { + "bbox": [ + 55, + 474, + 296, + 512 + ], + "type": "inline_equation", + "content": "E^{H,O}" + }, + { + "bbox": [ + 55, + 474, + 296, + 512 + ], + "type": "text", + "content": ", and lifts the latter to 3D as " + }, + { + "bbox": [ + 55, + 474, + 296, + 512 + ], + "type": "inline_equation", + "content": "E_{3D}^{H,O} = \\Phi(E^{H,O}, K)" + }, + { + "bbox": [ + 55, + 474, + 296, + 512 + ], + "type": "text", + "content": ". Contact masks are defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 517, + 295, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 517, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 115, + 517, + 295, + 534 + ], + "type": "interline_equation", + "content": "M ^ {H, O} = \\Omega^ {H, O} (\\Theta (R ^ {H, O}), E _ {3 D} ^ {H, O}). \\tag {2}", + "image_path": "81d248c82c46a5655fa49fb0844e44ec052335947630b6e4cfd0834ed205b279.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "content": "Below, the " + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "content": " superscripts are dropped for brevity. We calculate losses only on \"valid\" regions, i.e., the areas within the outline of rendered geometry; we denote this as " + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "content": ". To encourage overlap between the predicted masks, " + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "content": ", and the ground-truth ones, " + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\widehat{M}" + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "content": ", particularly in sparse contact regions, we use a focal-weighted BCE loss and a Dice loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 620, + 295, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 620, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 61, + 620, + 295, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B C E} = - \\alpha (1 - p _ {M}) ^ {\\gamma} \\log (p _ {M}) - (1 - \\alpha) p _ {M} ^ {\\gamma} \\log (1 - p _ {M}), \\tag {3}", + "image_path": "9a5eb7b98b623d11c70dcf9619547c0f6e3fe031d93d9527641cae295a707d8c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 643, + 295, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 643, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 115, + 643, + 295, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D i c e} = 1 - \\frac {2 \\sum M \\cdot \\widehat {M} + \\epsilon}{\\sum M + \\sum \\widehat {M} + \\epsilon}, \\tag {4}", + "image_path": "46d921407e1b45c49117b362564a1496ae38362d06410e6fb1f0081f83312604.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_M" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is the predicted mask probability, " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " controls class balance, " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " adjusts the focus on hard examples, and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is a residual term to prevent division by zero." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "RLL step #3: Lift " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "2\\mathbf{D} \\rightarrow 3\\mathbf{D}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ". To lift the inferred 2D contact points to 3D points, we follow the inverse of step #1. Normally, 2D points backproject to 3D lines due to depth ambiguities. In our case the lines intersect with the known 3D geometry that produced the multi-view renders. So, 2D points are lifted to 3D points, and by extension 2D contact masks, " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "M^{H,O}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", are lifted to 3D contact areas, " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "C^{H,O}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 156, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 204 + ], + "type": "text", + "content": "We use a human contact loss, " + }, + { + "bbox": [ + 313, + 156, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C^H" + }, + { + "bbox": [ + 313, + 156, + 553, + 204 + ], + "type": "text", + "content": ", that combines a focal loss with sparsity regularization, to encourage precise true-positive predictions in \"valid\" contact regions while discouraging false positives in non-contact areas:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 357, + 209, + 553, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 209, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 357, + 209, + 553, + 225 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C} ^ {H} = \\alpha (1 - p _ {h C}) ^ {\\gamma} \\log \\left(p _ {h C}\\right) + \\lambda \\| C ^ {H} \\| _ {1}, \\tag {5}", + "image_path": "bf66cca47a358c2c5deb71c24775002e7fb0cfff77ff81cac86ee2637d69ea76.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "inline_equation", + "content": "p_{hC}" + }, + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "text", + "content": " is the contact probability, and " + }, + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "text", + "content": " are scalar weights. We also use an object contact loss, " + }, + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C^O" + }, + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "text", + "content": ", that combines a Dice and Mean Squared Error (MSE) loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 358, + 272, + 553, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 272, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 358, + 272, + 553, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C} ^ {O} = \\mathcal {L} _ {\\text {D i c e}} \\left(C ^ {O}, \\widehat {C} ^ {O}\\right) + \\beta \\| C ^ {O} - \\widehat {C} ^ {O} \\| _ {2} ^ {2}, \\tag {6}", + "image_path": "a1f604f21203cb2fe3e708f1be2ce10e285c85edb668318a8b34e7e68bfb4876.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 294, + 553, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 553, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 294, + 553, + 319 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 294, + 553, + 319 + ], + "type": "text", + "content": " is a weighting factor, and " + }, + { + "bbox": [ + 313, + 294, + 553, + 319 + ], + "type": "inline_equation", + "content": "\\widehat{C^O}" + }, + { + "bbox": [ + 313, + 294, + 553, + 319 + ], + "type": "text", + "content": " denotes the groundtruth 3D contacts for objects." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 325, + 447, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 447, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 447, + 338 + ], + "type": "text", + "content": "3.5. Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 342, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 462 + ], + "type": "text", + "content": "Architecture. We use LLaVA [41] as our VLM and SAM [31] for our MV-Loc, with weights pre-trained by LISA [34] for segmentation [34]. The feature-lifting network, " + }, + { + "bbox": [ + 313, + 342, + 554, + 462 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 342, + 554, + 462 + ], + "type": "text", + "content": ", contains a spatial-understanding network (two fully-connected layers of size 128 with ReLU activation) followed by view-specific (256-dimensional) transformations and a sigmoid activation. For 3D contact prediction, our MV-Loc model converts 2D masks to 3D contact points via 2D-to-3D pixel-to-vertex mappings that are precomputed during MV-Loc's rendering step. For details, see Sup. Mat." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "text", + "content": "Training. To efficiently fine-tune our VLM, we employ LoRA [26] with rank 8. The separate decoders for human and object contact prediction are trained without LoRA, while keeping the image encoder frozen. For training, we use DeepSpeed [3] with mixed precision training (bfloat16), and batch size of 8. We train on 4 Nvidia-A100 GPUs for 30 epochs. For more details, please refer to Sup. Mat." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "text", + "content": "Datasets. We focus on two tasks, i.e., 3D human contact and 3D object affordance prediction, using two in-the-wild datasets, i.e., DAMON [58] and PIAD [65], respectively. For human contact we train and evaluate on DAMON [58]. For 3D object affordances, we train and evaluate on PIAD [65]. We find that exploiting textual descriptions for the contacting body parts, and the object type in contact, helps training. Similarly, adding Visual Question-Answering (VQA) data generated by GPT4o for the training images also helps. For details, see Sup. Mat." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Unlike LEMON [66], which requires paired human-object geometry for training using the 3DIR dataset [66], we use unpaired data. This enables us to scale for many human and object categories not addressed by prior work. For" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22609" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 65, + 288, + 137 + ], + "blocks": [ + { + "bbox": [ + 64, + 65, + 288, + 137 + ], + "lines": [ + { + "bbox": [ + 64, + 65, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 64, + 65, + 288, + 137 + ], + "type": "table", + "html": "
MethodF1 (%)↑Precision (%)↑Recall (%)↑Geodesic (cm)↓
POSAPPIXIE [23] [19]31.042.034.033.00
BSTRO [27]46.051.053.038.06
DECO [58]55.065.057.021.32
InteractVLM75.675.276.02.89
", + "image_path": "b790d7ba3199355cd3386090a7cc8b1ac61fb826be36c6253b0d4de97b877675.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 190, + 293, + 286 + ], + "blocks": [ + { + "bbox": [ + 55, + 144, + 295, + 178 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 178 + ], + "type": "text", + "content": "Table 1. Evaluation for \"Binary Human Contact\" prediction on the DAMON dataset [58]. We compare our InteractVLM model (trained only for this task) with the state of the art." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 190, + 293, + 286 + ], + "lines": [ + { + "bbox": [ + 58, + 190, + 293, + 286 + ], + "spans": [ + { + "bbox": [ + 58, + 190, + 293, + 286 + ], + "type": "table", + "html": "
Object CategoriesSemantic-DECO [58] (Baseline)InteractVLM
F1 (%)↑Prec. (%)↑Rec. (%)↑Geo (cm)↓F1 (%)↑Prec. (%)↑Rec. (%)↑Geo (cm)↓
Accessory40.130.775.621.8861.172.965.66.91
Daily Obj26.520.152.360.3468.671.478.07.46
Food11.719.412.949.6166.466.177.97.17
Furniture24.515.883.729.1760.564.260.16.21
Kitchen27.724.737.252.3471.871.381.17.61
Sports36.430.480.179.2177.976.783.27.98
Transport52.039.193.731.7877.880.578.97.97
", + "image_path": "83cc82727fd8975c73eee7aff2a1c20b07491ac2f9ce47f901a447eab6d4c0d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 386 + ], + "type": "text", + "content": "the final joint human-object reconstruction task, we combine DAMON [58], PIAD [65], 3DIR [66] and all textual descriptions about body parts, contact type and HOI." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 386, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 296, + 471 + ], + "type": "text", + "content": "Evaluation metrics. For human contact prediction, following Tripathi et al. [58], we report the F1, precision, and recall scores using a threshold of 0.5, and a geodesic distance measuring spatial accuracy. For object contact prediction, following Yang et al. [65], we report the Similarity (SIM), Mean Absolute Error (MAE), Area Under ROC Curve (AUC), and average Intersection over Union (IOU)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 483, + 137, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 137, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 137, + 496 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 502, + 254, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 254, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 254, + 515 + ], + "type": "text", + "content": "4.1. \"Binary Human Contact\" Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 521, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 568 + ], + "type": "text", + "content": "This task refers to estimating contact areas on the body via binary classification of its vertices, ignoring the number or type of objects involved. We train and evaluate on the DAMON [58] dataset, and report results in Tab. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 569, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 629 + ], + "type": "text", + "content": "InteractVLM significantly outperforms all previous methods achieving a " + }, + { + "bbox": [ + 55, + 569, + 295, + 629 + ], + "type": "inline_equation", + "content": "20.6\\%" + }, + { + "bbox": [ + 55, + 569, + 295, + 629 + ], + "type": "text", + "content": " improvement in F1 score. Although, here, InteractVLM is trained on the same data as DECO, it goes beyond this by harnessing the commonsense knowledge of a large foundation model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "In Sup. Mat. we also evaluate on the 3DIR [66] dataset and compare with the LEMON method [66]. Even though LEMON uses paired human-object data, InteractVLM performs on-par with it despite training on human-only data. We also evaluate InteractVLM's performance for binary human contact prediction across different body parts; for more details please refer to Sup. Mat." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 65, + 553, + 155 + ], + "blocks": [ + { + "bbox": [ + 55, + 294, + 295, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 327 + ], + "type": "text", + "content": "Table 2. Evaluation for \"Semantic Human Contact\" prediction on the DAMON [58] dataset. For results on each class, see Sup. Mat. The \"Semantic-DECO\" baseline extends DECO for our new task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 65, + 553, + 155 + ], + "lines": [ + { + "bbox": [ + 317, + 65, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 317, + 65, + 553, + 155 + ], + "type": "table", + "html": "
MethodsPIAD-Seen [65]PIAD-Unseen [65]
SIM (%)↑AUC (%)↑aIOU (%)↑MAE ↓SIM (%)↑AUC (%)↑aIOU (%)↑MAE ↓
PMF [70]42.575.0510.131.4133.060.254.672.11
ILN [10]42.775.8411.521.3732.559.694.712.07
PFusion [63]43.277.5012.311.3533.061.875.331.93
XMF [1]44.178.2412.941.2734.262.585.681.88
IAGNet [65]54.584.8520.510.9835.271.847.951.27
InteractVLM62.786.4721.200.8141.475.458.500.99
", + "image_path": "4cd5b6f95464648c083ae9c0ec75f577239936ad0ba8cdf67164aa1cbce38014.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 162, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 313, + 162, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 555, + 196 + ], + "type": "text", + "content": "Table 3. Evaluation for \"Object Affordance Prediction\" on the PIAD [65] dataset. We compare our InteractVLM model (trained only for this task) with the state of the art." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 216, + 524, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 524, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 524, + 228 + ], + "type": "text", + "content": "4.2. \"Semantic Human Contact\" Estimation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 233, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 293 + ], + "type": "text", + "content": "In real-life interactions, multiple objects can be contacted by different body areas concurrently. Thus, we introduce a novel task called \"Semantic Human Contact\" prediction. We evaluate on DAMON [58] and report results in Tab. 2; for a finer-grained version of this table, see Sup. Mat." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 293, + 556, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 556, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 556, + 426 + ], + "type": "text", + "content": "To establish a baseline, we adapt the DECO model [58] that detects binary contacts and turn it into a multi-class prediction model, called Semantic-DECO. Due to DAMON's limited training data, this has poor performance. Instead, as discussed in Sec. 4.1, InteractVLM learns effectively from this data, by also leveraging the commonsense of foundation models. Thus, it significantly outperforms Semantic-DECO. Qualitative results reflect this finding; see Fig. 4. Our model captures detailed, accurate contact regions, whereas Semantic-DECO often highlights false-positive areas that differ from the actual contact regions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 432, + 476, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 476, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 476, + 445 + ], + "type": "text", + "content": "4.3. Object Affordance Prediction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "content": "We evaluate the performance of our InteractVLM model on predicting object affordances. This involves identifying regions on objects of possible contact for a certain interaction intent, such as \"sitting\" on a chair or \"moving\" it. We train and evaluate on the PIAD [65] dataset. We compare against SotA methods, and report results in Tab. 3." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 521, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 617 + ], + "type": "text", + "content": "Note that we evaluate on object instances that are both seen during training (\"PIAD-Seen\" column) and unseen (\"PIAD-Unseen\" column). Our model demonstrates a significant improvement over SotA methods in terms of similarity (SIM), area under the curve (AUC), and mean absolute error (MAE) for both seen and unseen objects. Thus, our method is effective not only for estimating human contact (Secs. 4.1 and 4.2), but also for object affordances." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 624, + 468, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 468, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 468, + 635 + ], + "type": "text", + "content": "4.4. Reliance on 3D Annotations" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "To analyze InteractVLM's efficiency regarding 3D supervision, we train multiple versions of it with varying amounts of training data from the DAMON dataset, and report performance in Fig. 5. Remarkably, InteractVLM achieves an F1 score of 0.53 with just " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " of the data, nearly matching the 0.55 F1 score of DECO that uses " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " of the data. With" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22610" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 108, + 121 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 108, + 121 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 108, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 108, + 121 + ], + "type": "image", + "image_path": "ae3586bc333315df4653daa80ee9ae4c9d0b7a81f1b8a9e4c4ed20a903dcc502.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 121, + 107, + 167 + ], + "blocks": [ + { + "bbox": [ + 59, + 121, + 107, + 167 + ], + "lines": [ + { + "bbox": [ + 59, + 121, + 107, + 167 + ], + "spans": [ + { + "bbox": [ + 59, + 121, + 107, + 167 + ], + "type": "image", + "image_path": "e6396a9845e6ff0d9091ac19b8b8f42330c43179719d50c8236bf6f54a044361.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 168, + 107, + 217 + ], + "blocks": [ + { + "bbox": [ + 59, + 168, + 107, + 217 + ], + "lines": [ + { + "bbox": [ + 59, + 168, + 107, + 217 + ], + "spans": [ + { + "bbox": [ + 59, + 168, + 107, + 217 + ], + "type": "image", + "image_path": "488c4b08b65e2406a8bf825dfa745f760286402b130aedb6eace06fbeeee00dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 218, + 107, + 266 + ], + "blocks": [ + { + "bbox": [ + 59, + 218, + 107, + 266 + ], + "lines": [ + { + "bbox": [ + 59, + 218, + 107, + 266 + ], + "spans": [ + { + "bbox": [ + 59, + 218, + 107, + 266 + ], + "type": "image", + "image_path": "782f998cf68e81f129817bf575cceb17990561b5d132fd5164ac64d13c57ed9a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 269, + 105, + 278 + ], + "lines": [ + { + "bbox": [ + 61, + 269, + 105, + 278 + ], + "spans": [ + { + "bbox": [ + 61, + 269, + 105, + 278 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 110, + 75, + 154, + 118 + ], + "blocks": [ + { + "bbox": [ + 110, + 75, + 154, + 118 + ], + "lines": [ + { + "bbox": [ + 110, + 75, + 154, + 118 + ], + "spans": [ + { + "bbox": [ + 110, + 75, + 154, + 118 + ], + "type": "image", + "image_path": "f5feea365bfeab32fee4c3cd916173581b8b1ae1f4179e856f3d36aef9732eae.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 110, + 122, + 154, + 166 + ], + "blocks": [ + { + "bbox": [ + 110, + 122, + 154, + 166 + ], + "lines": [ + { + "bbox": [ + 110, + 122, + 154, + 166 + ], + "spans": [ + { + "bbox": [ + 110, + 122, + 154, + 166 + ], + "type": "image", + "image_path": "b0fd1a649fdeef1f8139e8d48c3ff1fa3496ffb7c1e2d633ae1e56cdee89cca4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 110, + 173, + 154, + 217 + ], + "blocks": [ + { + "bbox": [ + 110, + 173, + 154, + 217 + ], + "lines": [ + { + "bbox": [ + 110, + 173, + 154, + 217 + ], + "spans": [ + { + "bbox": [ + 110, + 173, + 154, + 217 + ], + "type": "image", + "image_path": "c941e71980b6d4dcf76bcdd82b049125dfef6a72550a55b0ac897dfb15bf7331.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 110, + 220, + 154, + 279 + ], + "blocks": [ + { + "bbox": [ + 110, + 220, + 154, + 279 + ], + "lines": [ + { + "bbox": [ + 110, + 220, + 154, + 279 + ], + "spans": [ + { + "bbox": [ + 110, + 220, + 154, + 279 + ], + "type": "image", + "image_path": "03ccf57b1801d4a4f931d16e58d07e348b7f0052875ba97e67e0fd192e263aa2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 282, + 555, + 306 + ], + "lines": [ + { + "bbox": [ + 55, + 282, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 555, + 306 + ], + "type": "text", + "content": "Figure 4. \"Semantic Human Contact\" estimation (Sec. 4.2). Given an image and an object label, InteractVLM infers body contacts for this object. InteractVLM outperforms a Semantic-DECO [58] baseline. Objects are shown in green circles, and contacts as red patches." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 162, + 76, + 200, + 118 + ], + "blocks": [ + { + "bbox": [ + 162, + 76, + 200, + 118 + ], + "lines": [ + { + "bbox": [ + 162, + 76, + 200, + 118 + ], + "spans": [ + { + "bbox": [ + 162, + 76, + 200, + 118 + ], + "type": "image", + "image_path": "ddb8cb87ee440601057e4ea4f735759f0301a155909ff4924f6d932bc60eaab2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 162, + 122, + 200, + 166 + ], + "blocks": [ + { + "bbox": [ + 162, + 122, + 200, + 166 + ], + "lines": [ + { + "bbox": [ + 162, + 122, + 200, + 166 + ], + "spans": [ + { + "bbox": [ + 162, + 122, + 200, + 166 + ], + "type": "image", + "image_path": "9eadf1322ae1a0d7532202d5fee9b7fd05a2375cb8e93de9e98e94b88ad1ade2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 164, + 173, + 200, + 217 + ], + "blocks": [ + { + "bbox": [ + 164, + 173, + 200, + 217 + ], + "lines": [ + { + "bbox": [ + 164, + 173, + 200, + 217 + ], + "spans": [ + { + "bbox": [ + 164, + 173, + 200, + 217 + ], + "type": "image", + "image_path": "9b3abd25bf064fc3e6c042acfeab68e1f1659b537c52aae15eced057ea7db150.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 164, + 220, + 200, + 278 + ], + "blocks": [ + { + "bbox": [ + 164, + 220, + 200, + 278 + ], + "lines": [ + { + "bbox": [ + 164, + 220, + 200, + 278 + ], + "spans": [ + { + "bbox": [ + 164, + 220, + 200, + 278 + ], + "type": "image", + "image_path": "c2e08016fd73f4956fe8c273d08b8dacceb73284e72491085d77e9ccce4be8bc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 208, + 75, + 252, + 118 + ], + "blocks": [ + { + "bbox": [ + 208, + 75, + 252, + 118 + ], + "lines": [ + { + "bbox": [ + 208, + 75, + 252, + 118 + ], + "spans": [ + { + "bbox": [ + 208, + 75, + 252, + 118 + ], + "type": "image", + "image_path": "3602418ca5a1d4f427db3b2eb7564cd181134080c8830860d6d840285212cb6f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 208, + 122, + 252, + 166 + ], + "blocks": [ + { + "bbox": [ + 208, + 122, + 252, + 166 + ], + "lines": [ + { + "bbox": [ + 208, + 122, + 252, + 166 + ], + "spans": [ + { + "bbox": [ + 208, + 122, + 252, + 166 + ], + "type": "image", + "image_path": "380edd433ade8b0e5f53a20501012cc525d94e0c547617f0fad71e36677d8f27.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 208, + 173, + 252, + 217 + ], + "blocks": [ + { + "bbox": [ + 208, + 173, + 252, + 217 + ], + "lines": [ + { + "bbox": [ + 208, + 173, + 252, + 217 + ], + "spans": [ + { + "bbox": [ + 208, + 173, + 252, + 217 + ], + "type": "image", + "image_path": "5c136ece426d8e4e35c646eb7b4f6bed244501159ae986d13164a3052bfe683f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 208, + 220, + 252, + 278 + ], + "blocks": [ + { + "bbox": [ + 208, + 220, + 252, + 278 + ], + "lines": [ + { + "bbox": [ + 208, + 220, + 252, + 278 + ], + "spans": [ + { + "bbox": [ + 208, + 220, + 252, + 278 + ], + "type": "image", + "image_path": "2194b3977f64dc86e4cd2ee1d51ea06aac6f4eea7eeb8a5b550862155c5b8175.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 260, + 76, + 298, + 118 + ], + "blocks": [ + { + "bbox": [ + 260, + 76, + 298, + 118 + ], + "lines": [ + { + "bbox": [ + 260, + 76, + 298, + 118 + ], + "spans": [ + { + "bbox": [ + 260, + 76, + 298, + 118 + ], + "type": "image", + "image_path": "a7f5c18cbde0f0fa7c41594bc05acda098e97f9a9a8078ac60cc37c194b5b205.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 260, + 122, + 297, + 217 + ], + "blocks": [ + { + "bbox": [ + 260, + 122, + 297, + 217 + ], + "lines": [ + { + "bbox": [ + 260, + 122, + 297, + 217 + ], + "spans": [ + { + "bbox": [ + 260, + 122, + 297, + 217 + ], + "type": "image", + "image_path": "bba2b5ec99a4344007b7dfdf30ccafd035cbf373cbed7d00999c937faa812aaf.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 260, + 220, + 298, + 278 + ], + "blocks": [ + { + "bbox": [ + 260, + 220, + 298, + 278 + ], + "lines": [ + { + "bbox": [ + 260, + 220, + 298, + 278 + ], + "spans": [ + { + "bbox": [ + 260, + 220, + 298, + 278 + ], + "type": "image", + "image_path": "842c087a139a155405ca782dfc93e5e1a14661104284aad7ca9b709d417215c2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 304, + 71, + 353, + 120 + ], + "blocks": [ + { + "bbox": [ + 304, + 71, + 353, + 120 + ], + "lines": [ + { + "bbox": [ + 304, + 71, + 353, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 353, + 120 + ], + "type": "image", + "image_path": "d8639aa8775480f870d1b6cfd2d9e0f259357d7936d88275276fb64f479db797.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 306, + 121, + 353, + 169 + ], + "blocks": [ + { + "bbox": [ + 306, + 121, + 353, + 169 + ], + "lines": [ + { + "bbox": [ + 306, + 121, + 353, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 353, + 169 + ], + "type": "image", + "image_path": "01dde2f85ab79a24f2fe7065d544a5e6c0ee6fd5d7a5de167c4a40a088893698.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 306, + 170, + 353, + 217 + ], + "blocks": [ + { + "bbox": [ + 306, + 170, + 353, + 217 + ], + "lines": [ + { + "bbox": [ + 306, + 170, + 353, + 217 + ], + "spans": [ + { + "bbox": [ + 306, + 170, + 353, + 217 + ], + "type": "image", + "image_path": "4bec1acbe6593a4079b2817b57d6d3c76de2db2d6e11cfa67ea635c58ab16fbd.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 304, + 218, + 354, + 278 + ], + "blocks": [ + { + "bbox": [ + 304, + 218, + 354, + 278 + ], + "lines": [ + { + "bbox": [ + 304, + 218, + 354, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 354, + 278 + ], + "type": "image", + "image_path": "39fa99a62a82437025f3b6ea4603407b4136cf7a794df8d3c136f294d27483b6.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 269, + 351, + 279 + ], + "lines": [ + { + "bbox": [ + 306, + 269, + 351, + 279 + ], + "spans": [ + { + "bbox": [ + 306, + 269, + 351, + 279 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 356, + 75, + 400, + 118 + ], + "blocks": [ + { + "bbox": [ + 356, + 75, + 400, + 118 + ], + "lines": [ + { + "bbox": [ + 356, + 75, + 400, + 118 + ], + "spans": [ + { + "bbox": [ + 356, + 75, + 400, + 118 + ], + "type": "image", + "image_path": "041a58c9585b777d5b7cfc1af9653f4eae9535cbc7bfb545cbb7efe4be08d9e2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 357, + 122, + 401, + 168 + ], + "blocks": [ + { + "bbox": [ + 357, + 122, + 401, + 168 + ], + "lines": [ + { + "bbox": [ + 357, + 122, + 401, + 168 + ], + "spans": [ + { + "bbox": [ + 357, + 122, + 401, + 168 + ], + "type": "image", + "image_path": "6189a670508b95ed1025cb971c9ddd42dbb984fb377b6fc5a43a5d7dd4ab5582.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 357, + 173, + 401, + 217 + ], + "blocks": [ + { + "bbox": [ + 357, + 173, + 401, + 217 + ], + "lines": [ + { + "bbox": [ + 357, + 173, + 401, + 217 + ], + "spans": [ + { + "bbox": [ + 357, + 173, + 401, + 217 + ], + "type": "image", + "image_path": "968734f81b7e044425f2d3aea4334f75bf3c4adf3ff8b1d33a04cea323a73c60.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 356, + 220, + 401, + 278 + ], + "blocks": [ + { + "bbox": [ + 356, + 220, + 401, + 278 + ], + "lines": [ + { + "bbox": [ + 356, + 220, + 401, + 278 + ], + "spans": [ + { + "bbox": [ + 356, + 220, + 401, + 278 + ], + "type": "image", + "image_path": "64b47061a7b0198615a75e730e7f84ce69950cf9efb2f4b20c7cb971a74adf76.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 410, + 76, + 446, + 118 + ], + "blocks": [ + { + "bbox": [ + 410, + 76, + 446, + 118 + ], + "lines": [ + { + "bbox": [ + 410, + 76, + 446, + 118 + ], + "spans": [ + { + "bbox": [ + 410, + 76, + 446, + 118 + ], + "type": "image", + "image_path": "fcad1cbfccd8ee84f038800b43cec1a7098abe02777ea774fdcedf1bcf928f19.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 410, + 122, + 445, + 217 + ], + "blocks": [ + { + "bbox": [ + 410, + 122, + 445, + 217 + ], + "lines": [ + { + "bbox": [ + 410, + 122, + 445, + 217 + ], + "spans": [ + { + "bbox": [ + 410, + 122, + 445, + 217 + ], + "type": "image", + "image_path": "ab057e6559b68066b739823a27bebbc8a93c4b6da9d715d3da7084fb6b6c36e4.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 411, + 224, + 445, + 278 + ], + "blocks": [ + { + "bbox": [ + 411, + 224, + 445, + 278 + ], + "lines": [ + { + "bbox": [ + 411, + 224, + 445, + 278 + ], + "spans": [ + { + "bbox": [ + 411, + 224, + 445, + 278 + ], + "type": "image", + "image_path": "89b56d0c9b3ce24e61e652a14889f8ad2792d208a66186b9bf52092f46252f9c.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 453, + 75, + 498, + 118 + ], + "blocks": [ + { + "bbox": [ + 453, + 75, + 498, + 118 + ], + "lines": [ + { + "bbox": [ + 453, + 75, + 498, + 118 + ], + "spans": [ + { + "bbox": [ + 453, + 75, + 498, + 118 + ], + "type": "image", + "image_path": "f35163338ec24ef83196f048b9938eb7c5a7fee3a953c5524ecbfe7f60341a5e.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 454, + 122, + 498, + 217 + ], + "blocks": [ + { + "bbox": [ + 454, + 122, + 498, + 217 + ], + "lines": [ + { + "bbox": [ + 454, + 122, + 498, + 217 + ], + "spans": [ + { + "bbox": [ + 454, + 122, + 498, + 217 + ], + "type": "image", + "image_path": "a0b328167017e295394af5f8583b28f953da9109f63f0afcc647f41a65b56cb3.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 454, + 222, + 498, + 278 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 454, + 222, + 498, + 278 + ], + "spans": [ + { + "bbox": [ + 454, + 222, + 498, + 278 + ], + "type": "image", + "image_path": "06ebcf9bc3da8501533be0f6ef34d13fd5fa683581c808f50bb5942564889840.jpg" + } + ] + } + ], + "index": 34, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 507, + 77, + 542, + 118 + ], + "blocks": [ + { + "bbox": [ + 507, + 77, + 542, + 118 + ], + "lines": [ + { + "bbox": [ + 507, + 77, + 542, + 118 + ], + "spans": [ + { + "bbox": [ + 507, + 77, + 542, + 118 + ], + "type": "image", + "image_path": "c8bd52f57a6a08848e792a86b340e9ee6692973aea46f81abd90a5b865b459ca.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 508, + 122, + 542, + 217 + ], + "blocks": [ + { + "bbox": [ + 508, + 122, + 542, + 217 + ], + "lines": [ + { + "bbox": [ + 508, + 122, + 542, + 217 + ], + "spans": [ + { + "bbox": [ + 508, + 122, + 542, + 217 + ], + "type": "image", + "image_path": "498c583f98d71f172e17400a4161c569ddd71d608e2fcc1cf005ae3ef6b6d5c6.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 453, + 222, + 498, + 278 + ], + "blocks": [ + { + "bbox": [ + 453, + 222, + 498, + 278 + ], + "lines": [ + { + "bbox": [ + 453, + 222, + 498, + 278 + ], + "spans": [ + { + "bbox": [ + 453, + 222, + 498, + 278 + ], + "type": "image", + "image_path": "7e19db98fc810890fb3cfb9a0fc8eb2a7c9e828bbc8f463b1db2159a35e3e375.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 508, + 224, + 542, + 278 + ], + "blocks": [ + { + "bbox": [ + 508, + 224, + 542, + 278 + ], + "lines": [ + { + "bbox": [ + 508, + 224, + 542, + 278 + ], + "spans": [ + { + "bbox": [ + 508, + 224, + 542, + 278 + ], + "type": "image", + "image_path": "2e90b1585752b80c7af0b278fc5849e11e371d1a91ca3cf429f7b7e5003238ec.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 65, + 321, + 287, + 495 + ], + "blocks": [ + { + "bbox": [ + 65, + 321, + 287, + 495 + ], + "lines": [ + { + "bbox": [ + 65, + 321, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 65, + 321, + 287, + 495 + ], + "type": "image", + "image_path": "d3e51179f7b70427e053a041cf10081c080215e4fca26daa97fa8bc1bc88a19f.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 506, + 295, + 583 + ], + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 583 + ], + "type": "text", + "content": "Figure 5. InteractVLM's reliance on 3D annotations. We evaluate performance for \"binary human contact\" (F1 score, Y-axis) for models trained on a varying percentage of DAMON [58] training data (X-axis). The DECO baseline trains on " + }, + { + "bbox": [ + 55, + 506, + 295, + 583 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 506, + 295, + 583 + ], + "type": "text", + "content": " of DAMON. Instead, InteractVLM trains on a varying (smaller) portion of this dataset. Yet, it achieves a significantly higher performance, by leveraging the broad visual knowledge of foundation models." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "only " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " of the data, InteractVLM surpasses DECO, reaching an F1 score of 0.58. This performance gap widens as training data increases, ultimately achieving 0.75 F1 score with " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " of the data. This is a compelling finding, highlighting InteractVLM's efficient use of 3D supervision by leveraging the rich visual understanding of foundation models. Note that the strong performance with limited training data has high practical value, because obtaining 3D annotations is expensive." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 327, + 555, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 375 + ], + "type": "text", + "content": "Ablations. For a study on the influence of InteractVLM components, including mask resolution, MV-Loc variants, loss functions, training data, the influence of the VLM, and the effect of text prompts, please refer to Sup. Mat." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 383, + 516, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 516, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 516, + 397 + ], + "type": "text", + "content": "5. Joint Human-Object Reconstruction" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 403, + 554, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 439 + ], + "type": "text", + "content": "We demonstrate the usefulness of 3D contacts inferred by InteractVLM for reconstructing a 3D human and object in interaction from a single in-the-wild image, " + }, + { + "bbox": [ + 313, + 403, + 554, + 439 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 403, + 554, + 439 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 313, + 439, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 554, + 487 + ], + "type": "text", + "content": "Initializing 3D body pose & shape, object shape. We use OSX [39] to estimate a 3D SMPL-X body mesh, " + }, + { + "bbox": [ + 313, + 439, + 554, + 487 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 439, + 554, + 487 + ], + "type": "text", + "content": ", and OpenShape [42] to retrieve a 3D object mesh, " + }, + { + "bbox": [ + 313, + 439, + 554, + 487 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 439, + 554, + 487 + ], + "type": "text", + "content": ", from the Objaverse [11] database that best matches the image." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 487, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 570 + ], + "type": "text", + "content": "Initializing 3D object pose. We apply InteractVLM on image " + }, + { + "bbox": [ + 313, + 487, + 554, + 570 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 487, + 554, + 570 + ], + "type": "text", + "content": " to predict 3D body and object contact vertices, " + }, + { + "bbox": [ + 313, + 487, + 554, + 570 + ], + "type": "inline_equation", + "content": "C^H, C^O" + }, + { + "bbox": [ + 313, + 487, + 554, + 570 + ], + "type": "text", + "content": ". Then, we solve for object pose, " + }, + { + "bbox": [ + 313, + 487, + 554, + 570 + ], + "type": "inline_equation", + "content": "\\{R^O, t^O\\}" + }, + { + "bbox": [ + 313, + 487, + 554, + 570 + ], + "type": "text", + "content": " by applying the ICP [6] algorithm to snap the 3D object contact points, " + }, + { + "bbox": [ + 313, + 487, + 554, + 570 + ], + "type": "inline_equation", + "content": "C^O" + }, + { + "bbox": [ + 313, + 487, + 554, + 570 + ], + "type": "text", + "content": ", onto body ones, " + }, + { + "bbox": [ + 313, + 487, + 554, + 570 + ], + "type": "inline_equation", + "content": "C^H" + }, + { + "bbox": [ + 313, + 487, + 554, + 570 + ], + "type": "text", + "content": ". To avoid false correspondences, the 3D normals of contact points must be compatible, i.e., should have similar angles but opposite directions." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 571, + 554, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 554, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 554, + 607 + ], + "type": "text", + "content": "Optimizing 3D object pose. Given the above initialization, we optimize over object rotation, " + }, + { + "bbox": [ + 313, + 571, + 554, + 607 + ], + "type": "inline_equation", + "content": "R^O" + }, + { + "bbox": [ + 313, + 571, + 554, + 607 + ], + "type": "text", + "content": ", translation, " + }, + { + "bbox": [ + 313, + 571, + 554, + 607 + ], + "type": "inline_equation", + "content": "t^O" + }, + { + "bbox": [ + 313, + 571, + 554, + 607 + ], + "type": "text", + "content": ", and scale, " + }, + { + "bbox": [ + 313, + 571, + 554, + 607 + ], + "type": "inline_equation", + "content": "s^O" + }, + { + "bbox": [ + 313, + 571, + 554, + 607 + ], + "type": "text", + "content": ", via render-and-compare by minimizing:" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 353, + 612, + 553, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 612, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 353, + 612, + 553, + 624 + ], + "type": "interline_equation", + "content": "E = E _ {M} + \\lambda_ {C} E _ {C}, \\tag {7}", + "image_path": "8b5d0f47be4c48728d5c501def9f53dd8dda2a17a5c5150ca3927675a9233f78.jpg" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 347, + 628, + 553, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 628, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 347, + 628, + 553, + 642 + ], + "type": "interline_equation", + "content": "E _ {M} = I o U (\\widehat {M}, M) + \\left\\| \\widehat {M ^ {c}} - M ^ {c} \\right\\| _ {2}, \\tag {8}", + "image_path": "50767e996e260efa7648e2393a9bae4834fc2a9c728ba594af4721e4f410f228.jpg" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 348, + 644, + 553, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 644, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 348, + 644, + 553, + 673 + ], + "type": "interline_equation", + "content": "E _ {C} = \\frac {1}{| C ^ {H} \\| C ^ {O} |} \\sum_ {i \\in | H |} \\sum_ {j \\in | O |} C _ {i} ^ {H} C _ {j} ^ {O} \\| V _ {i} ^ {H} - V _ {j} ^ {O} \\| _ {2}, \\tag {9}", + "image_path": "5d926e5ab18e262a9254e1342bfb39251886ed432f6ab23ec430e59ce6643871.jpg" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "E_{M}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " is a mask loss, " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "E_{C}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " is a contact loss, " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "IoU" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " is intersection-over-union, " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\widehat{M}, M" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " are hypothesis and ground-truth masks, respectively, " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\widehat{M^{c}}, M^{c}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " are the corresponding" + } + ] + } + ], + "index": 52 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22611" + } + ] + } + ], + "index": 53 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 53, + 547, + 361 + ], + "blocks": [ + { + "bbox": [ + 59, + 53, + 547, + 361 + ], + "lines": [ + { + "bbox": [ + 59, + 53, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 59, + 53, + 547, + 361 + ], + "type": "image", + "image_path": "23806b3f4592cf7a9f13aacbf01bfdd8d1d1b5e0c1c87280ad3fb8c664973e5b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 363, + 555, + 387 + ], + "lines": [ + { + "bbox": [ + 55, + 363, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 555, + 387 + ], + "type": "text", + "content": "Figure 6. 3D HOI reconstruction (Sec. 5). We build an optimization method that fits a SMPL-X body and OpenShape-retrieved object to an in-the-wild image. We evaluate against the SotA method PHOSA [67]. Reconstruction is guided by InteractVLM-inferred contacts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "content": "mean mask pixels, " + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "inline_equation", + "content": "|C^H|" + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "inline_equation", + "content": "|C^O|" + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "content": " are the number of contact vertices on the human and object, " + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "inline_equation", + "content": "|H|" + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "inline_equation", + "content": "|O|" + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "content": " are the number of mesh vertices, " + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "inline_equation", + "content": "V_i^H" + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "inline_equation", + "content": "V_j^O" + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "content": "-th human and object vertices, while " + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "inline_equation", + "content": "C_i^O" + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "inline_equation", + "content": "C_j^H" + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "content": " indicate whether vertices are in contact or not. Given the image, we extract the object mask, " + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "content": ", via SAM [31], and a depth map, " + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 406, + 295, + 481 + ], + "type": "text", + "content": ", via Depth Pro [8]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "type": "text", + "content": "Intuitively, we use the " + }, + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "type": "inline_equation", + "content": "E_{M}" + }, + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "type": "text", + "content": " loss to align the 3D object to the image, while in " + }, + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "type": "inline_equation", + "content": "E_{C}" + }, + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "type": "text", + "content": " the predicted 3D contacts \"anchor\" the 3D object onto the body so they interact realistically. We perform iterative optimization via Adam [30]. OSX produces reasonable bodies, so we keep these fixed. The object is updated in each iteration - we render depth, " + }, + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\widehat{D}" + }, + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "type": "text", + "content": ", and masks, " + }, + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\widehat{M}" + }, + { + "bbox": [ + 55, + 483, + 295, + 567 + ], + "type": "text", + "content": ", via a PyTorch3D [52] differentiable renderer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 567, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 676 + ], + "type": "text", + "content": "Qualitative Results. We reconstruct 3D human-object interaction from an image. We show results in Fig. 6, and compare with PHOSA [67], the most related SotA method. InteractVLM's reconstructions look more realistic. Note that PHOSA uses handcrafted contacts on humans and bodies. Instead, InteractVLM infers 3D contacts on both human bodies and objects from the single in-the-wild image. These play a crucial role for guiding 3D reconstruction under occlusions and depth ambiguities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Perceptual Study. There exists no in-the-wild dataset with 3D ground truth for HOI, so we conduct a perceptual study via Amazon Mechanical Turk for evaluation. Specifi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 407, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 491 + ], + "type": "text", + "content": "cally, we evaluate the realism of our reconstructions against ones of PHOSA. We randomly select 55 images for which PHOSA has handcrafted contact annotations. For each image, participants are shown (with random swapping) reconstructions generated by our method and by PHOSA, and are asked to select the one that best represents the image. Our reconstructions are preferred " + }, + { + "bbox": [ + 313, + 407, + 555, + 491 + ], + "type": "inline_equation", + "content": "62\\%" + }, + { + "bbox": [ + 313, + 407, + 555, + 491 + ], + "type": "text", + "content": " of the time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "We develop InteractVLM, a novel method for estimating 3D contacts on both humans and objects from a single natural image. InteractVLM has a reduced reliance on expensive 3D contact annotations for training, by leveraging the broad knowledge of Vision-Language Models. Specifically, we introduce a novel \"Render-Localize-Lift\" (RLL) framework and a novel multi-view localization model (MV-Loc) to adapt 2D foundation models for 3D contact estimation. We outperform existing work on contact estimation and introduce a new \"Semantic Human Contact\" estimation task for inferring body contacts conditioned on object labels. This goes beyond traditional binary contact estimation, which fails to capture rich semantic relationships of multi-object interactions. Last, we develop the first approach that uses inferred contact points on both bodies and objects for joint 3D reconstruction from single in-the-wild images." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22612" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 192 + ], + "type": "text", + "content": "Acknowledgments: We thank Alpár Cseke for his assistance with evaluating joint human-object reconstruction. We also thank Tsvetelina Alexiadis and Taylor Obersat for MTurk evaluation, Yao Feng, Peter Kultis and Markos Diomataris for their valuable feedback and Benjamin Pellkofer for IT support. SKD is supported by the International Max Planck Research School for Intelligent Systems (IMPRS-IS). The UvA part of the team is supported by an ERC Starting Grant (STRIPES, 101165317, PI: D. Tzionas). Disclosure: DT has received a research gift fund from Google. For MJB see https://files.is.tue.mpg.de/black/CoI_CVPR_2025.txt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 214, + 115, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 214, + 115, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 214, + 115, + 227 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 234, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 234, + 296, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 234, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 234, + 296, + 277 + ], + "type": "text", + "content": "[1] Emanuele Aiello, Diego Valsesia, and Enrico Magli. Cross-modal learning for image-guided point cloud shape completion. Conference on Neural Information Processing Systems (NeurIPS), 35:37349-37362, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 280, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 296, + 399 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: A visual language model for few-shot learning. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 401, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 401, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 61, + 401, + 296, + 477 + ], + "type": "text", + "content": "[3] Reza Yazdani Aminabadi, Samyam Rajbhandari, Minjia Zhang, A. A. Awan, Cheng Li, Du Li, Elton Zheng, Jeff Rasley, Shaden Smith, Olatunj Ruwase, and Yuxiong He. DeepSpeed inference: Enabling efficient inference of transformer models at unprecedented scale. International Conference for High Performance Computing, Networking, Storage and Analysis, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 479, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 296, + 521 + ], + "type": "text", + "content": "[4] Dimitrije Antić, Georgios Paschalidis, Shashank Tripathi, Theo Gevers, Sai Kumar Dwivedi, and Dimitrios Tzionas. SDFit: 3D object pose and shape by fitting a morphable SDF to a single image. arXiv:2409.16178, 2025. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 523, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 296, + 578 + ], + "type": "text", + "content": "[5] Fabien Baradel*, Matthieu Armando, Salma Galaoui, Romain Brégier, Philippe Weinzaepfel, Grégory Rogez, and Thomas Lucas*. Multi-HMR: Multi-person whole-body human mesh recovery in a single shot. In European Conference on Computer Vision (ECCV), 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 580, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 612 + ], + "type": "text", + "content": "[6] Paul J. Besl and Neil D. McKay. A method for registration of 3D shapes. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 14:239-256, 1992. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "type": "text", + "content": "[7] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAVE: Dataset and method for tracking human object interactions. In Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 669, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 669, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 669, + 296, + 712 + ], + "type": "text", + "content": "[8] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth Pro: Sharp monocular metric depth in less than a second. arXiv:2410.02073, 2024. 8" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 719 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 127 + ], + "type": "text", + "content": "[9] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision (ECCV), 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "text", + "content": "[10] Honghua Chen, Zeyong Wei, Yabin Xu, Mingqiang Wei, and Jun Wang. ImLoveNet: Misaligned image-supported registration network for low-overlap point cloud pairs. In International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 1-9, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 185, + 555, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 555, + 250 + ], + "type": "text", + "content": "[11] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2023. 2, 3, 4, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 251, + 554, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 554, + 295 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 554, + 295 + ], + "type": "text", + "content": "[12] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3D AffordanceNet: A benchmark for visual object affordance understanding. In Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 297, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 328 + ], + "type": "text", + "content": "[13] John Doe, Jane Smith, and Alice Brown. Paris3d: Language-guided 3d segmentation with multimodal models. arXiv:2406.08394, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 331, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 373 + ], + "type": "text", + "content": "[14] Sai Kumar Dwivedi, Nikos Athanasiou, Muhammed Kocabas, and Michael J. Black. Learning to regress bodies from images using differentiable semantic rendering. In International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 375, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 375, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 317, + 375, + 553, + 418 + ], + "type": "text", + "content": "[15] Sai Kumar Dwivedi, Cordelia Schmid, Hongwei Yi, Michael J. Black, and Dimitrios Tzionas. POCO: 3D pose and shape estimation using confidence. In International Conference on 3D Vision (3DV), 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 420, + 554, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 420, + 554, + 463 + ], + "spans": [ + { + "bbox": [ + 317, + 420, + 554, + 463 + ], + "type": "text", + "content": "[16] Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, Yao Feng, and Michael J. Black. TokenHMR: Advancing human mesh recovery with a tokenized pose representation. In Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 464, + 554, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 464, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 317, + 464, + 554, + 519 + ], + "type": "text", + "content": "[17] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual handobject manipulation. In Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 520, + 554, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 520, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 317, + 520, + 554, + 575 + ], + "type": "text", + "content": "[18] Zicong Fan, Maria Parelli, Maria Eleni Kadoglou, Muhammed Kocabas, Xu Chen, Michael J Black, and Otmar Hilliges. HOLD: Category-agnostic 3D reconstruction of interacting hands and objects from video. In Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 576, + 554, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 576, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 317, + 576, + 554, + 620 + ], + "type": "text", + "content": "[19] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Collaborative regression of expressive bodies using moderation. In International Conference on 3D Vision (3DV), 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 621, + 554, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 621, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 317, + 621, + 554, + 665 + ], + "type": "text", + "content": "[20] Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, and Tae-Kyun Kim. First-person hand action benchmark with RGB-D videos and 3D hand pose annotations. In Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 666, + 553, + 719 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 666, + 553, + 719 + ], + "spans": [ + { + "bbox": [ + 316, + 666, + 553, + 719 + ], + "type": "text", + "content": "[21] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22613" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 708 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[22] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambiguities with 3D scene constraints. In International Conference on Computer Vision (ICCV), pages 2282-2292, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "text", + "content": "[23] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14708-14718, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 225 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 225 + ], + "type": "text", + "content": "[24] Yana Hasson, Gül Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 227, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 227, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 227, + 294, + 270 + ], + "type": "text", + "content": "[25] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. arXiv: 2311.04400, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 271, + 294, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 294, + 314 + ], + "type": "text", + "content": "[26] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. International Conference on Learning Representations (ICLR), 2021. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 315, + 294, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 315, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 294, + 378 + ], + "type": "text", + "content": "[27] Chun-Hao Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Computer Vision and Pattern Recognition (CVPR), pages 13264-13275, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 380, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 380, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 380, + 294, + 434 + ], + "type": "text", + "content": "[28] Yinghao Huang, Omid Taheri, Michael J. Black, and Dimitrios Tzionas. InterCap: Joint markerless 3D tracking of humans and objects in interaction from multi-view RGB-D images. International Journal of Computer Vision (IJCV), 132(7):2551-2566, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 435, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 435, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 435, + 294, + 479 + ], + "type": "text", + "content": "[29] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In International Conference on 3D Vision (3DV), pages 42-52, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "type": "text", + "content": "[30] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 576 + ], + "type": "text", + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloé Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 2, 3, 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 578, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 620 + ], + "type": "text", + "content": "[32] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. VIBE: Video inference for human body pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), pages 5252-5262, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 621, + 294, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 621, + 294, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 621, + 294, + 665 + ], + "type": "text", + "content": "[33] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 666, + 294, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 666, + 294, + 708 + ], + "spans": [ + { + "bbox": [ + 56, + 666, + 294, + 708 + ], + "type": "text", + "content": "[34] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: Reasoning segmentation via large language model. In Computer Vision and Pattern Recognition (CVPR), pages 9579-9589, 2024. 3, 4, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 688 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[35] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. HybrIK: A hybrid analytical-neural inverse kinematics solution for 3D human pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), pages 3383-3393, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "text", + "content": "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning (ICML), pages 19730-19742, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "text", + "content": "[37] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "type": "text", + "content": "[38] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-resolution text-to-3D content creation. Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "type": "text", + "content": "[39] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3D whole-body mesh recovery with component aware transformer. In Computer Vision and Pattern Recognition (CVPR), 2023. 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 363 + ], + "type": "text", + "content": "[40] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "text", + "content": "[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[42] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. OpenShape: Scaling up 3D shape representation towards open-world understanding. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 2, 3, 4, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 454, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 508 + ], + "type": "text", + "content": "[43] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3D mesh in 45 seconds without pershape optimization. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "text", + "content": "[44] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 555, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 597 + ], + "type": "text", + "content": "[45] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. Transactions on Graphics (TOG), 34 (6):248:1-248:16, 2015. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 599, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 599, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 599, + 553, + 643 + ], + "type": "text", + "content": "[46] Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. COAP: Compositional articulated occupancy of people. In Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 644, + 553, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 644, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 316, + 644, + 553, + 688 + ], + "type": "text", + "content": "[47] Hyeongjin Nam, Daniel Sungho Jung, Gyeongsik Moon, and Young Mu Lee. Joint reconstruction of 3D human and object via contact-based refinement transformer. In Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22614" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 656 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[48] Mohamed Omran, Christoph Lassner, Gerard Pons-Moll, Peter Gehler, and Bernt Schiele. Neural body fitting: Unifying deep learning and model based human pose and shape estimation. In International Conference on 3D Vision (3DV), 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "text", + "content": "[49] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), pages 10975–10985, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 226 + ], + "type": "text", + "content": "[50] Bingbing Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 303 + ], + "type": "text", + "content": "[51] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. In International Conference on Learning Representations (ICLR), 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 346 + ], + "type": "text", + "content": "[52] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D deep learning with PyTorch3D. arXiv:2007.08501, 2020. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 348, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 348, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 56, + 348, + 294, + 390 + ], + "type": "text", + "content": "[53] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. Transactions on Graphics (TOG), 36(6), 2022, 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 392, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 392, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 294, + 435 + ], + "type": "text", + "content": "[54] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[55] Qingping Sun, Yanjun Wang, Ailing Zeng, Wanqi Yin, Chen Wei, Wenjia Wang, Haiyi Mei, Chi Sing Leung, Ziwei Liu, Lei Yang, and Zhongang Cai. AiOS: All-in-one-stage expressive human pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 491, + 294, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 294, + 534 + ], + "type": "text", + "content": "[56] István Sárándi and Gerard Pons-Moll. Neural localizer fields for continuous 3D human pose and shape estimation. In Conference on Neural Information Processing Systems (NeurIPS), 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 536, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 294, + 578 + ], + "type": "text", + "content": "[57] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "text", + "content": "[58] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J. Black. DECO: Dense estimation of 3D human-scene contact in the wild. In International Conference on Computer Vision (ICCV), pages 8001-8013, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 656 + ], + "type": "text", + "content": "[59] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 652 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "coordinate space for category-level 6D object pose and size estimation. In Computer Vision and Pattern Recognition (CVPR), 2019. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "text", + "content": "[60] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, and Jifeng Dai. VisionLLM: Large language model is also an open-ended decoder for vision-centric tasks. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 172, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 215 + ], + "type": "text", + "content": "[61] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. CHORE: Contact, human and object reconstruction from a single RGB image. In European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 217, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 261 + ], + "type": "text", + "content": "[62] Xianghui Xie, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Template free reconstruction of human-object interaction with procedural interaction generation. In Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "type": "text", + "content": "[63] Danfei Xu, Dragomir Anguelov, and Ashesh Jain. PointFusion: Deep sensor fusion for 3D bounding box estimation. In International Conference on Computer Vision (ICCV), pages 244-253, 2018. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 306, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 360 + ], + "type": "text", + "content": "[64] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, Bill Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 362, + 553, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 362, + 553, + 416 + ], + "spans": [ + { + "bbox": [ + 316, + 362, + 553, + 416 + ], + "type": "text", + "content": "[65] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3D object affordance from 2D interactions in images. In International Conference on Computer Vision (ICCV), pages 10871-10881, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 418, + 553, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 418, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 553, + 461 + ], + "type": "text", + "content": "[66] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, and Zheng-Jun Zha. LEMON: Learning 3D human-object interaction relation from 2D images. In Computer Vision and Pattern Recognition (CVPR), 2024. 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 463, + 553, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 463, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 316, + 463, + 553, + 517 + ], + "type": "text", + "content": "[67] Jason Y. Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3D human-object spatial arrangements from a single image in the wild. In European Conference on Computer Vision (ECCV), 2020. 3, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 518, + 553, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 518, + 553, + 562 + ], + "spans": [ + { + "bbox": [ + 316, + 518, + 553, + 562 + ], + "type": "text", + "content": "[68] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. GPT4RoI: Instruction tuning large language model on region-of-interest. arXiv:2307.03601, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 563, + 553, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 563, + 553, + 607 + ], + "spans": [ + { + "bbox": [ + 316, + 563, + 553, + 607 + ], + "type": "text", + "content": "[69] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3D: Exploring unified 3D representation at scale. In International Conference on Learning Representations (ICLR), 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 608, + 553, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 608, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 316, + 608, + 553, + 652 + ], + "type": "text", + "content": "[70] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3D lidar semantic segmentation. In International Conference on Computer Vision (ICCV), 2021. 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22615" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/447efac7-303b-4880-a09c-b0e941c7cbb1_content_list.json b/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/447efac7-303b-4880-a09c-b0e941c7cbb1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a47e49f7e5c72155ecfbb1fc315af50b2b795645 --- /dev/null +++ b/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/447efac7-303b-4880-a09c-b0e941c7cbb1_content_list.json @@ -0,0 +1,1911 @@ +[ + { + "type": "text", + "text": "InteractionMap: Improving Online Vectorized HDMap Construction with Interaction", + "text_level": 1, + "bbox": [ + 125, + 128, + 870, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kuang Wu Chuan Yang * Zhanbin Li Langge Technology", + "bbox": [ + 313, + 203, + 679, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vectorized high-definition (HD) maps are essential for an autonomous driving system. Recently, state-of-the-art map vectorization methods are mainly based on DETR-like framework to generate HD maps in an end-to-end manner. In this paper, we propose InteractionMap, which improves previous map vectorization methods by fully leveraging local-to-global information interaction in both time and space. Firstly, we explore enhancing DETR-like detectors by explicit position relation prior from point-level to instance-level, since map elements contain strong shape priors. Secondly, we propose a key-frame-based hierarchical temporal fusion module, which interacts temporal information from local to global. Lastly, the separate classification branch and regression branch lead to the problem of misalignment in the output distribution. We interact semantic information with geometric information by introducing a novel geometric-aware classification loss in optimization and a geometric-aware matching cost in label assignment. InteractionMap achieves state-of-the-art performance on both nuScenes and Argoverse2 benchmarks.", + "bbox": [ + 86, + 305, + 485, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 637, + 220, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-Definition (HD) maps are designed for high-precision autonomous driving, which contain instance-level vectorized representation such as lane divider, road boundaries, pedestrian crossing, etc. The rich semantic information of road topology and traffic rules is important for the navigation of autonomous driving (AD). HD maps are traditionally constructed offline using LiDAR SLAM-based methods [43, 52] with high maintenance costs, complex pipelines, and notable localization errors. In addition, manual annotation and map updating rely heavily on human labor and time demands.", + "bbox": [ + 89, + 662, + 482, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, more research endeavors have shifted towards deep-learning-based methods that construct vectorized HD maps around the ego-vehicle at runtime with on", + "bbox": [ + 89, + 830, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18454750ca09f64dbdff604cfcebc86fd6daf44b35bf22adc9b1b495680f809a.jpg", + "image_caption": [ + "Figure 1. Visual comparison between MapTRv2 [25] and our improved results. Our method effectively eliminates error map elements, leading to better precision and stability." + ], + "image_footnote": [], + "bbox": [ + 526, + 271, + 897, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "board sensors [19, 20, 24]. With the development of bird's eye view (BEV) representation, recent approaches [4, 13, 19, 22, 38, 41, 55] treat the semantic map learning task as a semantic segmentation problem. However, rasterized maps generated by these methods lack vectorized instance-level information, which is essential to downstream tasks, such as motion forecasting and motion planning [8, 23]. To overcome the limitations of segmentation-based methods, approaches VectorMapNet [29], MapTR [24], MapTRv2 [25] predict point sets to construct end-to-end vectorized local HD maps using transformer [45] decoders inspired by DETR [2].", + "bbox": [ + 511, + 612, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As shown in Figure 1, the representation of point sets has a limited model capability of instance-level information. Moreover, predicting directly from a single frame input is challenging in complex scenarios, due to temporal inconsistency between frames. To address these issues, we propose InteractionMap, which consists of Relation Embedding Module (REM), Temporal Fusion Module (TFM),", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding at: ycscience86@gmail.com", + "bbox": [ + 109, + 886, + 349, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17176", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and Geometry-aware Alignment Module (GAM). REM establishes position relation representation among feature embeddings at point level and instance level, improving the precision of map construction by explicit geometric priors. TFM facilitates long-range temporal associations by keyframe-based streaming strategy, a hierarchical temporal fusion from local to global. It enhances the stability of map elements among different frames in some complex scenarios, such as map element occlusion caused by moving vehicles. GAM solves the misalignment problem caused by inconsistency of the predictions between the classification score and the instance points precision, e.g. a prediction with a high classification score and relatively low localization quality (large Chamfer distance).", + "bbox": [ + 89, + 90, + 480, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 109, + 303, + 470, + 316 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce an explicit position relation embedding method from point to instance, efficiently employing progressive interaction of point-wise information and instance-wise information.", + "- We propose a novel key-frame-based hierarchical streaming strategy, which fully leverages long-range temporal information.", + "- We present a geometry-aware classification loss and a geometry-aware matching cost to overcome the misalignment problem of classification and localization output." + ], + "bbox": [ + 91, + 320, + 483, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 488, + 232, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. HD Map Construction", + "text_level": 1, + "bbox": [ + 89, + 513, + 299, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the development of view transformation from perspective-view (PV) to bird-eye-view (BEV) methods [18, 33], HD map construction is formulated as a segmentation task [4, 13, 22, 30, 31, 36, 38, 53, 55] based on sensor observations on board. HDMapNet [19] builds a vectorized HD map using semantic segmentation, clustering, and post-processing. VectorMapNet [8] is the first end-to-end framework that utilizes transformers [45], in a two-stage coarse-to-fine manner. However, the auto-regressive model of VectorMapNet leads to a long training schedule. MapTR [24] adopts a one-stage transformer approach based on [2, 22, 58] with a permutation-equivalent point set modeling. The evolved version MapTRv2 [25] adds auxiliary headers, decoupled self-attention in the decoder and one-to-many matching strategy, leading to a large improvement. Instead of a point set representation, BeMapNet [39] utilizes an instance-level representation with a piecewise Bezier head. PivotNet [7] converts point-level representation to instance-level representation using the point-to-line mask module. MapQR [32] implicitly encodes point-level queries within instance-level queries and embeds query positions like Conditional DETR [34] and DAB-DETR [27]. MapVR [50] generates a vectorized map with differentiable rasterization that provides instance-level segmentation su", + "bbox": [ + 89, + 537, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pervision. MMap [28] employs mask-guided features to refine an instance representation with enhanced feature detail. In contrast to the above methods, we introduce more information interaction to improve the reliability and stability of map construction in complex scenarios.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Online Temporal Fusion", + "text_level": 1, + "bbox": [ + 511, + 176, + 736, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The temporal fusion strategy is effective for the 3D object detection task, such as BEVFormer [22], SOLOFusion [37], StreamPETR [46], VideoBEV [10] and Sparse4D v2 [26]. HDMapNet [19] employs max pooling to fuse temporal information directly into the BEV feature map. StreamMapNet [49] proposes query propagation and BEV feature map fusion. SQD-MapNet [47] designs a stream query denoising approach. MapTracker [3] enhances query propagation and BEV feature map fusion using a tracking strategy. The streaming strategy facilitates longer temporal association as the propagated hidden states encode all historical information. However, a temporal encoder such as the gated recurrent unit (GRU) [5] may still face the problem of forgetting due to limited capacity in complex outdoor environments. For example, occlusion of moving vehicles may lead to temporal fusion of polluted BEV feature map. The stacking strategy may integrate features from specific previous frames, offering flexibility in fusion of long-range information. In this paper, we propose a key-frame-based streaming strategy, leveraging local and global fusion capabilities.", + "bbox": [ + 511, + 199, + 906, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Map Element Interaction", + "text_level": 1, + "bbox": [ + 511, + 511, + 743, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ADMap [15] explores point-order relationships between and within instances through a cascading approach. InsightMapper [48] performs inner-instance feature aggregation by an additional masked inner-instance self-attention module. GeMap [54] designs a masked decoupled self-attention to handle the shapes and relations of the instance independently. HoMap [1] introduces high-order modeling to capture the correlations between instances using high-order statistics. HIMap [57] proposed a hybrid framework to learn and interact with information at the point-level and the element-level. As proved in [14], modeling relations between objects is beneficial to object recognition. In state-of-the-art methods [9, 12], relationship of queries improves DETR-like models. In this paper, we explore the interaction of map elements between their geometry and appearance feature by our explicit position relation embedding.", + "bbox": [ + 511, + 534, + 906, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.4. Classification-Localization Alignment", + "text_level": 1, + "bbox": [ + 511, + 786, + 839, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The classification head and the localization head of object detection are implemented separately by two branches in parallel, causing inconsistency in the output distribution. This misalignment problem has been well researched in the 2D object detection field [21, 51]. However, it is overlooked by the map element detectors [7, 24, 25, 32, 39, 50]. In this", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17177", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/157b04680f848abca5e415492a98a0afab4fa9aadcdee2dffaf63568299f7318.jpg", + "image_caption": [ + "Figure 2. Overview of InteractionMap framework. InteractionMap mainly consists of four components: (1) BEV encoder transforms sensor input to a unified BEV representation; (2) Key-frame-based temporal fusion module leverage temporal information from local to global; (3) Relation map decoder utilizes relation embedding in point-level and instance-level; (4) Geometry-aware alignment module is designed to solve the misalignment problem of classification and position output." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "paper, we create a geometry-aware classification loss, making use of focal loss [42] with a geometry-aware target in foreground candidates. Moreover, we present a geometry-aware matching cost to suppress the matching of background candidates.", + "bbox": [ + 88, + 290, + 480, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 376, + 181, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 illustrates the framework of our proposed InteractionMap. Firstly, BEV features are extracted from multiview images by the BEV encoder, which consists of a backbone [11] to extract multi-scale image features, and a view transformation module [22, 31] to encode image features into a BEV representation $\\mathcal{F}_{local}$ . Secondly, key-frame-based temporal fusion leverages long-range temporal information by local fusion and global fusion (Section 3.1). Subsequently, the relation map decoder utilizes point-wise relation embedding (PRE) and instance-wise relation embedding (IRE) for better interaction between queries (Section 3.2). Finally, by introducing a geometric-aware classification loss and a geometric-aware matching cost, the geometry-aware alignment leverages interaction between semantic information and geometry information (Section 3.3).", + "bbox": [ + 91, + 402, + 483, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Key-frame-based Temporal Fusion", + "text_level": 1, + "bbox": [ + 89, + 650, + 393, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to the single-frame map element prediction method, there are mainly two types of temporal fusion methods, namely stacking-based strategy and streaming-based strategy. The streaming strategy leverages all historical information by hidden state encoding. However, due to limited memory capacity, the streaming strategy has limited performance in long-range perception. The stacking strategy fuse features from specific previous frames, offering flexibility in integration of long-range information. The computational cost is linearly related to the number of fused frames.", + "bbox": [ + 89, + 674, + 482, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a key-frame-based streaming (KFS) strategy to address these problems. The well-known key-frame-based strategy is widely used in robot navigation and mapping, such as sparse mapping [16, 35] and dense map", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ping [6, 56]. Online reconstruction methods are unstable in long range, but they are accurate in the local regime. The input frames are divided into local frame segments. The submaps are reconstructed by integration of local frames, and then the global map is reconstructed by integration of submaps. We introduce two types of KFS strategy, namely KFS-streaming and KFS-stacking, as shown in Figure 3. Firstly, the BEV feature embedding submaps are generated by a streaming-base temporal encoder at the bottom level, which keeps the stability and temporal consistency of predictions cross latest frames. Secondly, the long-range BEV feature embedding is generated by a streaming-based or stacking-based temporal encoder at the top level. Figure 3 illustrates the KFS strategy that merges one previous submap with an interval of two.", + "bbox": [ + 511, + 290, + 906, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Local BEV Fusion", + "text_level": 1, + "bbox": [ + 511, + 534, + 694, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We warp the memorized BEV feature map from the previous frame to the current frame recurrently based on the ego vehicle's relative pose. Then we employ a GRU [5] to fuse current BEV feature map $\\mathcal{F}_{local}^{t}$ and the warped memorized BEV feature map $\\tilde{\\mathcal{F}}_{submap}^{t-1}$ into a single BEV feature embedding submap candidate $\\mathcal{F}_{submap}^{t}$ .", + "bbox": [ + 511, + 558, + 906, + 655 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {F}} _ {s u b m a p} ^ {t - 1} = \\operatorname {W a r p} \\left(\\mathcal {F} _ {s u b m a p} ^ {t - 1}, \\boldsymbol {T}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 662, + 903, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\text {s u b m a p}} ^ {t} = \\operatorname {R e s B l o c k} \\left(L N \\left(G R U \\left(\\tilde {\\mathcal {F}} _ {\\text {s u b m a p}} ^ {t - 1}, \\mathcal {F} _ {\\text {l o c a l}} ^ {t}\\right)\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 689, + 903, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where ResBlock is a residual-based convolution block with the shortcut connection [11]. $LN$ is a layer normalization operation. $\\pmb{T}$ denotes a standard $4\\times 4$ transformation matrix between the coordinate systems of two frames.", + "bbox": [ + 511, + 722, + 906, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 Global BEV Fusion", + "text_level": 1, + "bbox": [ + 511, + 801, + 702, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our KFS strategy selects a BEV feature embedding as a submap using a specific distance stride $d_{stride}$ , instead of a specific temporal interval customly.", + "bbox": [ + 511, + 825, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "KFS-streaming strategy The previous selected global BEV feature is warped to the current frame. The warped", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17178", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f5b1cced7728e9d0d6325b09a9c6bd3dfc432cfc26f1cdb784c4e22c1f8049a0.jpg", + "image_caption": [ + "Figure 3. Temporal fusion strategy." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 472, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "global BEV feature and the current BEV feature map are then fused to create a global BEV feature map.", + "bbox": [ + 89, + 345, + 482, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {F}} _ {\\text {g l o b a l}} ^ {\\text {p r e}} = \\operatorname {W a r p} \\left(\\mathcal {F} _ {\\text {g l o b a l}} ^ {\\text {p r e}}, \\boldsymbol {T}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 383, + 480, + 404 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {g l o b a l} ^ {t} = L N \\left(G R U \\left(\\tilde {\\mathcal {F}} _ {g l o b a l} ^ {p r e}, \\mathcal {F} _ {s u b m a p} ^ {t}\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 410, + 482, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "KFS-stacking strategy We warp $N_{pre}$ selected submaps to the current frame. Following concatenation and residual-based convolution [11], we obtain the global BEV feature.", + "bbox": [ + 89, + 433, + 483, + 477 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {F}} _ {s u b m a p} ^ {t _ {k}} = \\operatorname {W a r p} \\left(\\mathcal {F} _ {s u b m a p} ^ {t _ {k}}, T\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 484, + 480, + 506 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\text {g l o b a l}} ^ {t} = \\operatorname {R e s B l o c k} \\left(\\operatorname {C o n c a t} \\left(\\tilde {F} _ {\\text {s u b m a p}}, \\mathcal {F} _ {\\text {s u b m a p}} ^ {t}\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 512, + 480, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\tilde{\\mathcal{F}}_{submap}^{t_k}$ and $t_k$ represent the selected BEV feature submap and its corresponding index. The collection of selected BEV feature submaps is denoted as $\\tilde{F}_{submap} = \\{\\tilde{\\mathcal{F}}_{submap}^{t_k}\\}_{k=1}^{N_{pre}}$ . During the training period, the previous $N_{pre}$ submaps are randomly selected within a range of $N_{pre} \\times d_{stride}$ .", + "bbox": [ + 89, + 536, + 483, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Relation map decoder", + "text_level": 1, + "bbox": [ + 89, + 640, + 295, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous work has shown the effectiveness of interactions for map element detectors [15, 57]. In contrast to these approaches, we directly construct explicit relation priors by incorporating relation embedding into the self-attention module of the decoder.", + "bbox": [ + 89, + 662, + 483, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Relation Embedding", + "text_level": 1, + "bbox": [ + 89, + 756, + 289, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here is the self-attention in the decoder of a DETR-like framework.", + "bbox": [ + 89, + 780, + 482, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {\\text {s e l f - a t t n}} = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 814, + 480, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $d_k$ is the dimension of the keys. $Q, K, V$ are feature embeddings generated by linear projections. $Q = \\operatorname{Linear}_q(Q_{in})$ . $K = \\operatorname{Linear}_k(Q_{in})$ . $V = \\operatorname{Linear}_v(Q_{in})$", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We incorporate relation embeddings into the self-attention model, inspired by [14].", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {s e l f - a t t n} = \\operatorname {S o f t m a x} \\left(\\mathcal {F} _ {r e l} + \\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 125, + 903, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\mathcal{F}_{rel}$ represents the relation embedding. In this paper, point-wise relation embedding and instance-wise relation embedding are added to the decoupled self-attention model at the point level and the instance level, respectively, employing attention refinement of information interaction.", + "bbox": [ + 511, + 166, + 906, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Point-wise Relation Embedding", + "text_level": 1, + "bbox": [ + 511, + 258, + 787, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The point-wise relation encoder represents the high-dimensional point relation embedding as an explicit geometry prior. The point-wise relation embedding is generated by normalized point coordinates and edge directions.", + "bbox": [ + 511, + 282, + 905, + 344 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e l} _ {p t} (i, j) = \\left[ \\log \\left(x _ {i} - x _ {j} + 1\\right), \\log \\left(y _ {i} - y _ {j} + 1\\right) \\right] \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 352, + 903, + 369 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e l} _ {\\text {d i r}} (i, j) = \\operatorname {c o s i n e \\_ s i m} \\left(e _ {i}, e _ {i + 1}\\right) - \\operatorname {c o s i n e \\_ s i m} \\left(e _ {j}, e _ {j + 1}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 377, + 913, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $x_{i}$ , $x_{j}$ are normalized point coordinates. $e_{i}$ and $e_{i+1}$ , $e_{j}$ and $e_{j+1}$ are adjacent edges. The cosine similarity is denoted by $\\cosin\\_sim$ . The point-wise relation embedding is unbiased, since both $Rel_{pt}(i,j)$ and $Rel_{dir}(i,j)$ are equal to 0 when $i = j$ .", + "bbox": [ + 511, + 407, + 905, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The point-wise relation embedding $\\mathcal{F}_{rel}^{pt}$ is illustrated by the following equation:", + "bbox": [ + 511, + 482, + 903, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {r e l} ^ {p t} = \\operatorname {R e L U} (\\operatorname {L i n e a r} (S P E (\\operatorname {C o n c a t} (R e l _ {p t}, R e l _ {d i r})))) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 521, + 903, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, the point-wise relation embedding is calculated by concatenation of $Rel_{pt}$ and $Rel_{dir}$ . Then the embedding is transformed to high dimension by sine positional encoding (SPE). Lastly, the embedding is transformed by a linear projection and a ReLU function.", + "bbox": [ + 511, + 553, + 905, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3 Instance-wise Relation Embedding", + "text_level": 1, + "bbox": [ + 511, + 646, + 808, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The instance-wise relation embedding is generated by classification scores and chamfer distance of points.", + "bbox": [ + 511, + 670, + 903, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e l} _ {s} (i, j) = \\operatorname {s g n} \\left(s _ {i} - s _ {j}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 709, + 903, + 727 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e l} _ {s d} (i, j) = \\operatorname {s g n} \\left(s _ {i} - s _ {j}\\right) \\cdot \\operatorname {C h a m f e r} \\left(I _ {i}, I _ {j}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 734, + 903, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $sgn$ is the sign function. $s_i, s_j$ are classification scores of the $i^{th}$ and $j^{th}$ candidates. Chamfer $(I_i, I_j)$ is the chamfer distance between points of instance $i$ and instance $j$ . The equation of $Rel_s(i,j)$ establishes a simple classification ranking relation among pairs of instances. Then we add the position relation of the map elements to the instance-wise relation embedding $\\mathcal{F}_{rel}^{ins}$ by the chamfer distance.", + "bbox": [ + 511, + 753, + 906, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {r e l} ^ {i n s} = R e L U \\left(L i n e a r \\left(S P E \\left(R e l _ {s d}\\right)\\right)\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 883, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17179", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The instance-wise relation embedding is unbiased, since $Rel_{sd}(i,j)$ is equal to 0 when $i = j$ .", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Geometry-aware Alignment", + "text_level": 1, + "bbox": [ + 89, + 131, + 343, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Current work, such as MapTR [24] uses the classification branch and the regression branch to predict semantic labels and geometric localization separately. MapTR uses the focal loss [42] as the classification loss and the point-to-point L1 loss as the regression loss. Since classification and regression losses are independent, the quality of these two branches is inconsistent, leading to the well-known misalignment problem. In this paper, we introduce the geometry-aware focal loss and the geometry-aware focal cost to overcome this problem.", + "bbox": [ + 89, + 154, + 483, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Geometry-aware Focal Loss", + "text_level": 1, + "bbox": [ + 89, + 324, + 341, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by IoU-aware Focal Loss [21, 51] in traditional object detection, we design the novel Geometry-aware Focal Loss (GFL). Unlike focal loss, we treat positives and negatives asymmetrically.", + "bbox": [ + 89, + 348, + 483, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G F L} = \\sum_ {i = 1} ^ {N _ {\\text {p o s}}} s _ {\\text {g e o} _ {i}} B C E \\left(s _ {\\text {g e o} _ {i}}, p _ {i}\\right) + \\sum_ {j = 1} ^ {N _ {\\text {n e g}}} \\alpha p _ {j} ^ {\\gamma} B C E \\left(p _ {j}, 0\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 419, + 488, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_i$ and $p_j$ are the predicted probability of the $i^{th}$ and $j^{th}$ candidates, $s_{geo} \\in [0,1]$ is an instance Geometry-aware Classification Score (GCS). Here we define three types of GCS, including $s_{p2p}$ , $s_{dir}$ and $s_{giou}$ .", + "bbox": [ + 89, + 492, + 483, + 555 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {p 2 p} = 1 - \\frac {1}{2 N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} D _ {\\text {M a n h a t t a n}} \\left(p _ {\\text {s r c} _ {i}}, p _ {\\text {t g t} _ {i}}\\right) \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 566, + 482, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {d i r} = 0. 5 + \\frac {1}{2 N _ {e}} \\sum_ {i = 1} ^ {N _ {e}} \\operatorname {c o s i n e \\_ s i m} \\left(e _ {s r c _ {i}}, e _ {t g t _ {i}}\\right) \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 619, + 482, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {g i o u} = 0. 5 + 0. 5 \\cdot G I o U \\left(B _ {s r c}, B _ {t g t}\\right) \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 667, + 482, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_{p2p} \\in [0,1]$ is a normalized point-to-point L1 score, $s_{dir} \\in [0,1]$ is an edge direction score calculated by cosine similarity and $s_{giou} \\in [0,1]$ is a normalized GIoU [40] score of candidate bounding box $B_{src}$ and assigned GT $B_{tgt}$ . The bounding box is computed by minimum enclosing box of instance points. $N_p$ and $N_e$ are the numbers of points and edges. $p_{src}$ and $p_{tgt}$ are normalized points of candidates and assigned GT. $e_{src}$ and $e_{tgt}$ are edges of candidates and assigned GT.", + "bbox": [ + 89, + 691, + 483, + 828 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Geometry-aware Focal Cost", + "text_level": 1, + "bbox": [ + 89, + 847, + 339, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss function $\\mathcal{L}_{GFL}$ supervises the classification scores with geometric metrics GCS $s_{geo_i}$ . And $s_{geo_i}$ accurately", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "selects these high-quality candidates, which leads to better map vectorization performance. Following the spirit of GFL, we make similar modifications to the focal cost, which is a key component of label assignment. The novel matching cost, geometry-aware focal cost (GFC) is as follows.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {G F C} (i, j) = s _ {g e o _ {i}} B C E \\left(s _ {g e o _ {i}}, p _ {i}\\right) - \\alpha p _ {i} ^ {\\gamma} B C E \\left(1 - p _ {i}, 1\\right) \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 171, + 906, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_{GFC}(i,j)$ is the geometry-aware focal cost for the $i^{th}$ prediction and the $j^{th}$ ground truth. The GFC is used as a modulated function to suppress candidates with inaccurate prediction localization and promote candidates with an accurate position.", + "bbox": [ + 511, + 209, + 906, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Loss", + "text_level": 1, + "bbox": [ + 513, + 292, + 656, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Auxiliary BEV Supervision. We introduce a query-based instance segmentation (QIS) module using an instance mask prediction branch. Inspired by [17], we make use of the query embeddings $Q_{dec}$ from Transformer to dot-product the BEV feature embedding map $\\mathcal{F}_{bev}$ to obtain instance-wise binary masks $m$ with the sigmoid function $\\sigma(\\cdot)$ .", + "bbox": [ + 511, + 315, + 905, + 406 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} _ {\\text {i n s}} = \\sigma (M L P \\left(Q _ {\\text {d e c}}\\right) \\cdot \\mathcal {F} _ {\\text {b e v}}) \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 411, + 905, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the instance segmentation branch, we extend the mask focal loss and the mask focal cost to mask geometry-aware focal loss (MGFL) and mask geometry-aware focal cost (MGFC) as well, with the same ideal of GFL.", + "bbox": [ + 511, + 434, + 906, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s e g}} = \\lambda_ {m g f} \\mathcal {L} _ {m g f} (\\hat {M} _ {\\text {i n s}}, M _ {\\text {i n s}}) + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} (\\hat {M} _ {\\text {i n s}}, M _ {\\text {i n s}}) \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 515, + 908, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instance masks $\\hat{M}_{ins}$ are supervised by instance mask annotations $M_{ins}$ , using the mask geometry-aware focal loss $\\mathcal{L}_{mgf}$ and the dice loss $\\mathcal{L}_{dice}$ .", + "bbox": [ + 511, + 547, + 905, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall Loss. The total loss $L$ used by InteractionMap is as follow:", + "bbox": [ + 511, + 592, + 905, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {d e t}} + \\mathcal {L} _ {\\text {s e g}} + \\mathcal {L} _ {\\text {a u x}} \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 623, + 903, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{det}$ is the loss term for the detection task. $\\mathcal{L}_{seg}$ is the loss term for the segmentation task. $\\mathcal{L}_{aux}$ is the auxiliary loss term.", + "bbox": [ + 511, + 641, + 905, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The detection loss consists of classification loss $\\mathcal{L}_{cls}$ , point-to-point distance loss $\\mathcal{L}_{p2p}$ , edge direction loss $\\mathcal{L}_{dir}$ .", + "bbox": [ + 511, + 686, + 905, + 718 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e t}} = \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\lambda_ {p 2 p} \\mathcal {L} _ {p 2 p} + \\lambda_ {\\text {d i r}} \\mathcal {L} _ {\\text {d i r}} \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 723, + 903, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss term $\\mathcal{L}_{cls}$ is the geometry-aware focal loss for the classification of instances. The loss term $\\mathcal{L}_{p2p}$ is the smooth L1 loss for the regression of vectorized map elements. And the loss term $\\mathcal{L}_{dir}$ is the loss of cosine similarity between adjacent edges.", + "bbox": [ + 511, + 744, + 905, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a u x}} = \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {P V S e g} \\mathcal {L} _ {P V S e g} \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 835, + 903, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The auxiliary loss term $\\mathcal{L}_{aux}$ is composed of the depth-map prediction loss $\\mathcal{L}_{depth}$ and image semantic segmentation loss $\\mathcal{L}_{PVSeg}$ .", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17180", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 90, + 215, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 114, + 295, + 131 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1 Datasets", + "text_level": 1, + "bbox": [ + 89, + 138, + 204, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "nuScenes Datasets. There are 1000 scenes of 4 locations in nuScenes datasets. Each scene contains 20s of RGB images from six cameras, point clouds from LiDAR sweeps, and a 3D vectorized map. The train set contains 700 scenes with 28130 samples and validation set contains 150 scenes with 6019 samples, respectively. Following previous works [24, 25], we mainly focus on three categories of map elements, including lane divider, pedestrian crossing and road boundary.", + "bbox": [ + 89, + 162, + 482, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Argoverse2 Datasets. Argoverse2 dataset provide 1000 scenes in 6 cities, each lasting around 15 seconds. Each log includes 7 RGB images from surrounding cameras and point cloud from LiDAR sweeps. Follow previous methods [24, 25], 700 scenes are used for training, and 150 scenes are used for validation. For a fair comparison, we report results on its validation set and focus on the same three map categories as the nuScene dataset.", + "bbox": [ + 89, + 299, + 482, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 439, + 277, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a fair comparison, we follow the standard metric used in previous works [19, 24, 25, 29]. With the ego-coordination as the center, the perception ranges are $[-15m, 15m]$ for the X-axis and $[-30m, 30m]$ for the Y-axis. We adopt the mean Average Precision (mAP) to evaluate the map vectorization quality based on the chamfer distance. A candidate is considered as True-Positive (TP) only if the chamfer distance to ground truth is less than certain thresholds $(\\tau \\in T, T = 0.5, 1.0, 1.5)$ . The final AP metric is calculated by average across all thresholds and all classes.", + "bbox": [ + 89, + 463, + 482, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.3 Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 633, + 308, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous method [24, 25], we use ResNet50 [11] as the backbone for RGB images. All models are trained with 8 A800 GPUs. The default batch size is 32. The initial learning rate is set to $6 \\times 10^{-4}$ with cosine decay. The nuScenes images resolution is $1600 \\times 900$ . We resize the images with 0.5 ratio. For the Argoverse2 dataset, the image resolutions are $2048 \\times 1550$ and $1550 \\times 2048$ . We pad the images to $2048 \\times 2048$ , and then resize the images with 0.3 ratio. We define the size of BEV feature as $H \\times W$ of $200 \\times 100$ , and the size of each BEV grid as 0.3 meters. The default numbers of instance queries, point queries, and decoder layers are 100, 20, 6, respectively. The model is trained on 24, 110 epochs on the nuScenes and Argoverse2 datasets. We randomly divide each training sequence into 3 splits to generate more diverse data sequences. Inspired by SOLOFusion [37], we train the ini", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tial $N_{init}$ epochs with single-frame input to stabilize multi-frame straining. $N_{init} = 0.5N_{total}$ . For loss weights, $\\lambda_{cls} = 2$ , $\\lambda_{p2p} = 4$ , $\\lambda_{dir} = 0.005$ , $\\lambda_{mgf} = 30$ , $\\lambda_{dice} = 3$ , $\\lambda_{depth} = 3$ , $\\lambda_{PVSeg} = 2$ . For temporal fusion, $N_{pre} = 4$ , $d_{stride} = 5m$ .", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance Comparison", + "text_level": 1, + "bbox": [ + 511, + 176, + 748, + 193 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on nuScenes Dataset. As shown in Table 1, our methods show consistent improvements over baselines. By integrating our method with MapTRv2 [25], we achieve $71.6(+10.1)$ , $71.8(+10.3)$ mAP with 24-epoch training schedules and $75.5(+6.8)$ , $76.0(+7.3)$ mAP with 110-epoch training schedules, respectively. InteractionMap establishes a new state-of-the-art performance, exhibiting significant improvements over existing methods with comparable inference speed, as shown in Table 4.", + "bbox": [ + 511, + 200, + 903, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on Argoverse2 Dataset. Argoverse2 dataset contains 3D map elements, including height information. For fair comparison, the sampling frequency is set to $2\\mathrm{Hz}$ and $10\\mathrm{Hz}$ , respectively. The amount of training data at a $10\\mathrm{Hz}$ sampling frequency is 5 times as many as that at a $2\\mathrm{Hz}$ sampling frequency. As demonstrated in Table 2 and Table 3, InteractionMap surpasses all state-of-the-art methods in both 2D and 3D vectorized map perception on Argoverse2 dataset.", + "bbox": [ + 511, + 337, + 905, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 484, + 666, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We examine the efficacy of each component of InteractionMap through ablation studies, utilizing the nuScenes dataset. Initially, we build a simple baseline base on MapTR [24] with the LSS encoder without auxiliary depth-map supervision. The default numbers of instances, points per instance, are 50 and 20, respectively.", + "bbox": [ + 511, + 508, + 903, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Contributions of Main Components. Table 5 demonstrates the impact of each component of InteractionMap. When adding relation embedding into the baseline model, there is a substantial improvement of $4.5\\mathrm{mAP}$ . Furthermore, the addition of geometry alignment results in a significant 7.5 increase in mAP, emphasizing its effectiveness in boosting performance. Lastly, key-frame-based temporal fusion offers 5.0 and 5.1 additional increases in mAP, indicating that the inclusion of long-range hierarchical streaming strategy significantly enhances the performance.", + "bbox": [ + 511, + 599, + 903, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation on Relation Embedding. In Table 6, we explore the effect of relation embedding in point-level and instance-level. Adding the point-wise relation embedding of point coordinates and edge direction results in 2.1 and 0.9 improvements in mAP over the baseline, respectively. The instance-wise relation embedding leverages the classification confidence and points coordinates, leading to enhancements of 1.1 and $3.1\\mathrm{mAP}$ , respectively. The combination of relation embedding in point-level and instance-level enhances performance by $2.8\\mathrm{mAP}$ . Finally, we introduce a", + "bbox": [ + 511, + 750, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17181", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f9f30d81f7dad3c5d727d613fa63588f1d64f34af6be6bccf5bb51c6325d6e3a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneEpochsTemporal\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bou}\\)mAP
MapTRR502446.351.553.150.3
MapVRR502447.754.451.451.2
PivotNetR503053.855.859.657.4
BeMapNetR503057.762.359.459.8
MapTRv2R502459.862.462.461.5
StreamMapNetR5030✓61.766.362.163.4
MGMapR502461.865.067.564.8
SQD-MapNetR5024✓63.666.664.865.0
MapQRR502463.468.067.766.4
HIMapR503062.668.469.166.7
HRMapNet-MapTRv2R5024✓65.867.468.567.2
InteractionMap-RR5024✓71.370.872.871.6
InteractionMap-CR5024✓69.772.773.071.8
VectorMapNetR50110+ft42.551.444.146.0
MapTRR5011056.259.860.158.7
MapVRR5011055.061.859.458.8
BeMapNetR5011062.666.765.164.8
MapTRv2R5011068.168.369.768.7
MapQRR5011070.174.473.272.6
HRMapNet-MapTRv2R50110✓72.072.975.873.6
HIMapR5011071.375.074.773.7
InteractionMap-RR50110✓75.275.076.175.5
InteractionMap-CR50110✓75.375.677.076.0
", + "bbox": [ + 200, + 88, + 797, + 455 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d163be6023b02d6a7331ef5eaca404266e4fb0d20cf87cd71bebb5de4ab76c66.jpg", + "table_caption": [ + "Table 1. Comparison with SOTA methods on the nuScenes validation set at $60m\\times 30m$ perception range. \"EB0\", \"EB4\", \"R50\" correspond to the backbones Efficient-B0, Efficient-B4 [44], ResNet50 [11]. \"ft\" means the two-stage fine-tune strategy. \"C\" means KFS-stacking strategy and \"R\" means KFS-streaming strategy." + ], + "table_footnote": [], + "table_body": "
MethodDimAPpedAPdivAPboumAP
StreamMapNet2d62.059.563.061.5
SQD-MapNet2d64.960.264.963.3
MapTRv22d60.068.764.264.3
HRMapNet-MapTRv22d65.171.468.668.3
InteractionMap-C2d68.577.974.673.7
InteractionMap-R2d70.177.074.073.7
", + "bbox": [ + 98, + 521, + 475, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "alternating training strategy to randomly turn off one of these two components, leading the overall design achieves $54.7(+4.5)$ mAP.", + "bbox": [ + 89, + 704, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on Key-frame-based Temporal Fusion. To investigate the effectiveness of the key-frame-based streaming strategy, we conduct ablation experiments to compare it with the baseline (without temporal fusion) and the streaming strategy, in Table 7. Employing the streaming strategy results in $3.8\\mathrm{mAP}$ over the baseline. Furthermore, when the KFS-stacking strategy is introduced, the mAP increased by 7.4, reaching 57.6. Additionally, the inclusion of KFS-streaming strategy leads to an improvement of $7.2\\mathrm{mAP}$ .", + "bbox": [ + 89, + 750, + 482, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on Geometry-aware Alignment. As shown in", + "bbox": [ + 89, + 886, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8b1fe4f8491fadadf61a7df218497a1d577e8d2878483519dfc6872f844bcfb.jpg", + "table_caption": [ + "Table 2. Comparison with SOTA methods on the Argoverse2 validation set at $60m \\times 30m$ perception range with a 30-epoch training schedule, at a 2Hz sampling frequency." + ], + "table_footnote": [], + "table_body": "
MethodDimAPpedAPdivAPboumAP
MapTRv22d63.671.567.467.5
MapQR2d64.372.368.168.2
HIMap2d69.069.570.369.6
InteractionMap-C2d69.878.674.874.4
InteractionMap-R2d70.878.174.974.6
MapTRv23d60.768.964.564.7
MapQR3d60.171.266.265.9
HIMap3d66.768.370.368.4
InteractionMap-C3d66.675.672.771.6
InteractionMap-R3d67.775.573.172.1
", + "bbox": [ + 521, + 521, + 898, + 700 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/98724e414ada8d6fa36a0ca166620a7b874568524b7c879c33a55906d1a9134e.jpg", + "table_caption": [ + "Table 3. Comparison with SOTA methods on the Argoverse2 validation set at $60m \\times 30m$ perception range with a 6-epoch training schedule, at a $10\\mathrm{Hz}$ sampling frequency." + ], + "table_footnote": [], + "table_body": "
MethodFPSParams(MB)mAP
MapTRv214.940.661.5
InteractionMap-C11.0116.771.8
InteractionMap-R12.149.971.6
", + "bbox": [ + 521, + 763, + 898, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison with SOTA methods on the nuScenes validation set. FPS is measured on a single A800 GPU.", + "bbox": [ + 511, + 849, + 906, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17182", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/61fc62d3c8e05d886be4f160319fd135fe9d2f7e2c22aecc580436d1b2a3f988.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
REMGAMTFMAPpedAPdivAPboumAP
46.750.553.550.2
✓50.156.657.354.7
✓57.163.863.761.5
C50.960.061.957.6
R50.661.360.357.4
✓✓58.665.162.962.2
✓✓C64.268.668.867.2
✓✓R64.168.169.867.3
", + "bbox": [ + 98, + 88, + 477, + 229 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/588d67f369381408dc274ac9064a938debb1b6fd8df4c033c86c9ca60384914b.jpg", + "table_caption": [ + "Table 5. Ablation study of each component, including Relation Embedding Module (REM), Temporal Fusion Module (TFM), Geometry-aware Alignment Module (GAM). \"C\" means KFS-stacking strategy and \"R\" means KFS-streaming strategy." + ], + "table_footnote": [], + "table_body": "
PtxyPtdirInsSInssdPPAPpedAPdivAPboumAP
46.750.553.550.2
✓47.852.756.352.3
✓47.052.453.951.1
✓✓47.855.656.453.3
✓46.952.154.851.3
✓47.955.756.453.3
✓✓✓47.655.955.553.0
✓✓✓✓50.156.657.354.7
", + "bbox": [ + 98, + 309, + 477, + 450 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/62d2c4de49fd0cbfeab314ebe42d650c6a7b0cd43ffdbddc29c8f99ccedebc2f.jpg", + "table_caption": [ + "Table 6. Ablation study of relation embedding by employing point-wise coordinate embedding $\\mathrm{Pt}_{xy}$ , point-wise edge direction embedding $\\mathrm{Pt}_{dir}$ , instance-wise classification score embedding $\\mathrm{Ins}_s$ , instance-wise classifier score and chamfer distance embedding $\\mathrm{Ins}_{sd}$ and alternating training strategy of relation embedding, namely Ping-Pong (PP)." + ], + "table_footnote": [], + "table_body": "
MethodAPpedAPdivAPboumAP
Baseline46.750.553.550.2
Streaming48.654.758.854.0
KFS-stacking50.960.061.957.6
KFS-streaming50.661.360.357.4
", + "bbox": [ + 98, + 556, + 477, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8, geometry-aware alignment comprised five main components: query-based instance segmentation (QIS), geometry-aware focal loss (GFL), geometry-aware focal cost (GFC), mask geometry-aware focal loss (MGFL) and mask geometry-aware focal cost (MGFC). The QIS leverages geometry priors by instance segmentation mask supervision. Compared to the baseline, the QIS leads to an improvement of $6.4\\mathrm{mAP}$ . The GFL encourages predictions with high localization metrics to obtain a better classification score, with an improvement of $3.1\\mathrm{mAP}$ compared to mask focal loss. The GFC suppresses the matching cost of candidates with an inaccurate prediction position, which brings an improvement of $0.9\\mathrm{mAP}$ . For the segmentation", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c1eaadaeb826807f77bd4f4ecd92f56162b342173b1b1133cd3b052c9f837694.jpg", + "table_caption": [ + "Table 7. Ablation study of streaming strategy, KFS-streaming strategy and KFS-stacking strategy." + ], + "table_footnote": [], + "table_body": "
QISGFLGFCMGFLMGFCAPpedAPdivAPboumAP
46.750.553.550.2
✓51.758.759.456.6
✓✓55.162.062.059.7
✓✓✓55.863.862.260.6
✓✓✓✓57.063.662.361.0
✓✓✓✓✓57.163.863.761.5
", + "bbox": [ + 521, + 89, + 898, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e6b105a8455816c74fc292db258dd772ca3c8b055a3929ae97b7e00458331bf9.jpg", + "table_caption": [ + "Table 8. Ablation study of geometry-aware alignment by investigating the performance of query-based instance segmentation (QIS), geometry-aware focal loss (GFL), geometry-aware focal cost (GFC), mask geometry-aware focal loss (MGFL) and mask geometry-aware focal cost (MGFC)." + ], + "table_footnote": [], + "table_body": "
SgiouSp2pSdirAPpedAPdivAPboumAP
✓51.758.759.456.6
52.656.258.055.6
✓54.860.263.159.4
✓58.363.162.561.3
✓✓57.163.863.761.5
", + "bbox": [ + 521, + 297, + 898, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Ablation study of geometry-aware classification score.", + "bbox": [ + 519, + 411, + 897, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "branch, the result shows that the position-supervised segmentation loss MGFL and the position-modulated segmentation cost MGFC improve the final results, with $0.4\\mathrm{mAP}$ and $0.5\\mathrm{mAP}$ gains, individually.", + "bbox": [ + 511, + 440, + 906, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further investigate the geometry-aware classification score $S_{geo}$ . As depicted in Table 9, using $S_{giou}$ , $S_{p2p}$ and $S_{dir}$ leads to individual improvement of -1.0, 2.8 and 4.7 mAP, respectively. Furthermore, the combination of $S_{p2p}$ and $S_{dir}$ boosts the performance, improving by 4.9 mAP.", + "bbox": [ + 511, + 501, + 905, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 593, + 633, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce InteractionMap, a novel approach to end-to-end online vectorized HD map utilizing temporal and spatial information interaction from local to global. By leveraging relation embedding at both point level and instance level, the network learns the geometric priors of map elements better. Subsequently, key-frame-based streaming strategy enhances the network by utilizing temporal feature embedding from near to distant. Finally, the geometric-aware focal loss and the geometric-aware focal cost build a strong correlation between classification score and instance-point location precision. Across various experimental settings, our proposed InteractionMap yields significant performance improvements over the baseline on various datasets.", + "bbox": [ + 511, + 619, + 906, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 513, + 847, + 609, + 863 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Yingfeng Cai, Wei Dong, Ze Liu, Hai Wang, and Long Chen. Homap: End-to-end vectorized hd map construction with", + "bbox": [ + 519, + 872, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17183", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "high-order modeling. IEEE Transactions on Intelligent Vehicles, 2024. 2", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2", + "[3] Jiacheng Chen, Yuefan Wu, Jiaqi Tan, Hang Ma, and Yasutaka Furukawa. Maptracker: Tracking with strided memory fusion for consistent vector hd mapping. In European Conference on Computer Vision, pages 90-107. Springer, 2025. 2", + "[4] Shaoyu Chen, Tianheng Cheng, Xinggang Wang, Wenming Meng, Qian Zhang, and Wenyu Liu. Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. arXiv preprint arXiv:2206.04584, 2022. 1, 2", + "[5] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 2, 3", + "[6] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4): 1, 2017. 3", + "[7] Wenjie Ding, Limeng Qiao, Xi Qiu, and Chi Zhang. Pivot-net: Vectorized pivot learning for end-to-end hd map construction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3672-3682, 2023. 2", + "[8] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525-11533, 2020. 1, 2", + "[9] Yulu Gao, Yifan Sun, Xudong Ding, Chuyang Zhao, and Si Liu. Ease-detr: Easing the competition among object queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17282-17291, 2024. 2", + "[10] Chunrui Han, Jinrong Yang, Jianjian Sun, Zheng Ge, Runpei Dong, Hongyu Zhou, Weixin Mao, Yuang Peng, and Xi-angyu Zhang. Exploring recurrent long-term temporal fusion for multi-view 3d perception. IEEE Robotics and Automation Letters, 2024. 2", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 4, 6, 7", + "[12] Xiuquan Hou, Meiqin Liu, Senlin Zhang, Ping Wei, Badong Chen, and Xuguang Lan. Relation detr: Exploring explicit position relation prior for object detection. arXiv preprint arXiv:2407.11699, 2024. 2", + "[13] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings" + ], + "bbox": [ + 93, + 90, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 1, 2", + "[14] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3588-3597, 2018. 2, 4", + "[15] Haotian Hu, Fanyi Wang, Yaonong Wang, Laifeng Hu, Jingwei Xu, and Zhiwang Zhang. Admap: Anti-disturbance framework for reconstructing online vectorized hd map. arXiv preprint arXiv:2401.13172, 2024. 2, 4", + "[16] Georg Klein and David Murray. Parallel tracking and mapping for small ar workspaces. In 2007 6th IEEE and ACM international symposium on mixed and augmented reality, pages 225-234. IEEE, 2007. 3", + "[17] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 5", + "[18] Hongyang Li, Chonghao Sima, Jifeng Dai, Wenhai Wang, Lewei Lu, Huijie Wang, Jia Zeng, Zhiqi Li, Jiazhi Yang, Hanming Deng, et al. Delving into the devils of bird's-eye-view perception: A review, evaluation and recipe. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[19] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online hd map construction and evaluation framework. In 2022 International Conference on Robotics and Automation (ICRA), pages 4628-4634. IEEE, 2022. 1, 2, 6", + "[20] Tianyu Li, Peijin Jia, Bangjun Wang, Li Chen, Kun Jiang, Junchi Yan, and Hongyang Li. Lanesegnet: Map learning with lane segment perception for autonomous driving. arXiv preprint arXiv:2312.16108, 2023. 1", + "[21] Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems, 33:21002-21012, 2020. 2, 5", + "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 3", + "[23] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 541-556. Springer, 2020. 1", + "[24] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized hd map construction. arXiv preprint arXiv:2208.14437, 2022. 1, 2, 5, 6", + "[25] Bencheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17184", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv preprint arXiv:2308.05736, 2023.1, 2, 6", + "[26] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018, 2023. 2", + "[27] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2", + "[28] Xiaolu Liu, Song Wang, Wentong Li, Ruizi Yang, Junbo Chen, and Jianke Zhu. Mgmap: Mask-guided learning for online vectorized hd map construction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14812-14821, 2024. 2", + "[29] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized hd map learning. In International Conference on Machine Learning, pages 22352-22369. PMLR, 2023. 1, 6", + "[30] Zhi Liu, Shaoyu Chen, Xiaojie Guo, Xinggang Wang, Tianheng Cheng, Hongmei Zhu, Qian Zhang, Wenyu Liu, and Yi Zhang. Vision-based uneven bev representation learning with polar rasterization and surface estimation. In Conference on Robot Learning, pages 437-446. PMLR, 2023. 2", + "[31] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2, 3", + "[32] Zihao Liu, Xiaoyu Zhang, Guangwei Liu, Ji Zhao, and Ningyi Xu. Leveraging enhanced queries of point sets for vectorized map construction. In European Conference on Computer Vision, pages 461-477. Springer, 2025. 2", + "[33] Yuexin Ma, Tai Wang, Xuyang Bai, Huitong Yang, Yuenan Hou, Yaming Wang, Yu Qiao, Ruigang Yang, and Xinge Zhu. Vision-centric bev perception: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[34] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651-3660, 2021. 2", + "[35] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 3", + "[36] Cong Pan, Yonghao He, Junran Peng, Qian Zhang, Wei Sui, and Zhaoxiang Zhang. Baeformer: Bi-directional and early interaction transformers for bird's eye view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9590-9599, 2023. 2", + "[37] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 2, 6" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 1, 2", + "[39] Limeng Qiao, Wenjie Ding, Xi Qiu, and Chi Zhang. End-to-end vectorized hd-map construction with piecewiseBezier curve. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13218-13228, 2023. 2", + "[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5", + "[41] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11138-11147, 2020. 1", + "[42] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 3, 5", + "[43] Tixiao Shan and Brendan Englot. Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4758–4765. IEEE, 2018. 1", + "[44] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7", + "[45] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2", + "[46] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3621-3631, 2023. 2", + "[47] Shuo Wang, Fan Jia, Yingfei Liu, Yucheng Zhao, Zehui Chen, Tiancai Wang, Chi Zhang, Xiangyu Zhang, and Feng Zhao. Stream query denoising for vectorized hd map construction. arXiv preprint arXiv:2401.09112, 2024. 2", + "[48] Zhenhua Xu, Kenneth KY Wong, and Hengshuang Zhao. Insightmapper: A closer look at inner-instance information for vectorized high-definition mapping. arXiv preprint arXiv:2308.08543, 2023. 2", + "[49] Tianyuan Yuan, Yicheng Liu, Yue Wang, Yilun Wang, and Hang Zhao. Streammapnet: Streaming mapping network for vectorized online hd map construction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7356-7365, 2024. 2", + "[50] Gongjie Zhang, Jiahao Lin, Shuang Wu, Zhipeng Luo, Yang Xue, Shijian Lu, Zuoguan Wang, et al. Online map vec" + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17185", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "torization for autonomous driving: A rasterization perspective. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[51] Haoyang Zhang, Ying Wang, Feras Dayoub, and Niko Sunderhauf. Varifocalnet: An iou-aware dense object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8514-8523, 2021. 2, 5", + "[52] Ji Zhang, Sanjiv Singh, et al. Loam: Lidar odometry and mapping in real-time. In Robotics: Science and systems, pages 1-9. Berkeley, CA, 2014. 1", + "[53] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. **Reverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743**, 2022. 2", + "[54] Zhixin Zhang, Yiyuan Zhang, Xiaohan Ding, Fusheng Jin, and Xiangyu Yue. Online vectorized hd map construction using geometry. arXiv preprint arXiv:2312.03341, 2023. 2", + "[55] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13760-13769, 2022. 1, 2", + "[56] Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. Elastic fragments for dense scene reconstruction. In Proceedings of the IEEE International Conference on Computer Vision, pages 473-480, 2013. 3", + "[57] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized hd map construction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15396-15406, 2024. 2, 4", + "[58] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2" + ], + "bbox": [ + 91, + 92, + 482, + 599 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17186", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/447efac7-303b-4880-a09c-b0e941c7cbb1_model.json b/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/447efac7-303b-4880-a09c-b0e941c7cbb1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..84f06ce86310a14e7124f2c2f3624103f7268e35 --- /dev/null +++ b/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/447efac7-303b-4880-a09c-b0e941c7cbb1_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.13, + 0.872, + 0.175 + ], + "angle": 0, + "content": "InteractionMap: Improving Online Vectorized HDMap Construction with Interaction" + }, + { + "type": "text", + "bbox": [ + 0.314, + 0.204, + 0.681, + 0.241 + ], + "angle": 0, + "content": "Kuang Wu Chuan Yang * Zhanbin Li Langge Technology" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.306, + 0.486, + 0.609 + ], + "angle": 0, + "content": "Vectorized high-definition (HD) maps are essential for an autonomous driving system. Recently, state-of-the-art map vectorization methods are mainly based on DETR-like framework to generate HD maps in an end-to-end manner. In this paper, we propose InteractionMap, which improves previous map vectorization methods by fully leveraging local-to-global information interaction in both time and space. Firstly, we explore enhancing DETR-like detectors by explicit position relation prior from point-level to instance-level, since map elements contain strong shape priors. Secondly, we propose a key-frame-based hierarchical temporal fusion module, which interacts temporal information from local to global. Lastly, the separate classification branch and regression branch lead to the problem of misalignment in the output distribution. We interact semantic information with geometric information by introducing a novel geometric-aware classification loss in optimization and a geometric-aware matching cost in label assignment. InteractionMap achieves state-of-the-art performance on both nuScenes and Argoverse2 benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.638, + 0.222, + 0.654 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.483, + 0.829 + ], + "angle": 0, + "content": "High-Definition (HD) maps are designed for high-precision autonomous driving, which contain instance-level vectorized representation such as lane divider, road boundaries, pedestrian crossing, etc. The rich semantic information of road topology and traffic rules is important for the navigation of autonomous driving (AD). HD maps are traditionally constructed offline using LiDAR SLAM-based methods [43, 52] with high maintenance costs, complex pipelines, and notable localization errors. In addition, manual annotation and map updating rely heavily on human labor and time demands." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.831, + 0.484, + 0.877 + ], + "angle": 0, + "content": "In recent years, more research endeavors have shifted towards deep-learning-based methods that construct vectorized HD maps around the ego-vehicle at runtime with on" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.272, + 0.898, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.543, + 0.908, + 0.586 + ], + "angle": 0, + "content": "Figure 1. Visual comparison between MapTRv2 [25] and our improved results. Our method effectively eliminates error map elements, leading to better precision and stability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.794 + ], + "angle": 0, + "content": "board sensors [19, 20, 24]. With the development of bird's eye view (BEV) representation, recent approaches [4, 13, 19, 22, 38, 41, 55] treat the semantic map learning task as a semantic segmentation problem. However, rasterized maps generated by these methods lack vectorized instance-level information, which is essential to downstream tasks, such as motion forecasting and motion planning [8, 23]. To overcome the limitations of segmentation-based methods, approaches VectorMapNet [29], MapTR [24], MapTRv2 [25] predict point sets to construct end-to-end vectorized local HD maps using transformer [45] decoders inspired by DETR [2]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As shown in Figure 1, the representation of point sets has a limited model capability of instance-level information. Moreover, predicting directly from a single frame input is challenging in complex scenarios, due to temporal inconsistency between frames. To address these issues, we propose InteractionMap, which consists of Relation Embedding Module (REM), Temporal Fusion Module (TFM)," + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.351, + 0.901 + ], + "angle": 0, + "content": "*Corresponding at: ycscience86@gmail.com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17176" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.302 + ], + "angle": 0, + "content": "and Geometry-aware Alignment Module (GAM). REM establishes position relation representation among feature embeddings at point level and instance level, improving the precision of map construction by explicit geometric priors. TFM facilitates long-range temporal associations by keyframe-based streaming strategy, a hierarchical temporal fusion from local to global. It enhances the stability of map elements among different frames in some complex scenarios, such as map element occlusion caused by moving vehicles. GAM solves the misalignment problem caused by inconsistency of the predictions between the classification score and the instance points precision, e.g. a prediction with a high classification score and relatively low localization quality (large Chamfer distance)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.304, + 0.472, + 0.318 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.321, + 0.483, + 0.379 + ], + "angle": 0, + "content": "- We introduce an explicit position relation embedding method from point to instance, efficiently employing progressive interaction of point-wise information and instance-wise information." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.381, + 0.483, + 0.425 + ], + "angle": 0, + "content": "- We propose a novel key-frame-based hierarchical streaming strategy, which fully leverages long-range temporal information." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.426, + 0.484, + 0.471 + ], + "angle": 0, + "content": "- We present a geometry-aware classification loss and a geometry-aware matching cost to overcome the misalignment problem of classification and localization output." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.321, + 0.484, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.489, + 0.233, + 0.504 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.514, + 0.3, + 0.531 + ], + "angle": 0, + "content": "2.1. HD Map Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.901 + ], + "angle": 0, + "content": "With the development of view transformation from perspective-view (PV) to bird-eye-view (BEV) methods [18, 33], HD map construction is formulated as a segmentation task [4, 13, 22, 30, 31, 36, 38, 53, 55] based on sensor observations on board. HDMapNet [19] builds a vectorized HD map using semantic segmentation, clustering, and post-processing. VectorMapNet [8] is the first end-to-end framework that utilizes transformers [45], in a two-stage coarse-to-fine manner. However, the auto-regressive model of VectorMapNet leads to a long training schedule. MapTR [24] adopts a one-stage transformer approach based on [2, 22, 58] with a permutation-equivalent point set modeling. The evolved version MapTRv2 [25] adds auxiliary headers, decoupled self-attention in the decoder and one-to-many matching strategy, leading to a large improvement. Instead of a point set representation, BeMapNet [39] utilizes an instance-level representation with a piecewise Bezier head. PivotNet [7] converts point-level representation to instance-level representation using the point-to-line mask module. MapQR [32] implicitly encodes point-level queries within instance-level queries and embeds query positions like Conditional DETR [34] and DAB-DETR [27]. MapVR [50] generates a vectorized map with differentiable rasterization that provides instance-level segmentation su" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "pervision. MMap [28] employs mask-guided features to refine an instance representation with enhanced feature detail. In contrast to the above methods, we introduce more information interaction to improve the reliability and stability of map construction in complex scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.177, + 0.738, + 0.193 + ], + "angle": 0, + "content": "2.2. Online Temporal Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.907, + 0.502 + ], + "angle": 0, + "content": "The temporal fusion strategy is effective for the 3D object detection task, such as BEVFormer [22], SOLOFusion [37], StreamPETR [46], VideoBEV [10] and Sparse4D v2 [26]. HDMapNet [19] employs max pooling to fuse temporal information directly into the BEV feature map. StreamMapNet [49] proposes query propagation and BEV feature map fusion. SQD-MapNet [47] designs a stream query denoising approach. MapTracker [3] enhances query propagation and BEV feature map fusion using a tracking strategy. The streaming strategy facilitates longer temporal association as the propagated hidden states encode all historical information. However, a temporal encoder such as the gated recurrent unit (GRU) [5] may still face the problem of forgetting due to limited capacity in complex outdoor environments. For example, occlusion of moving vehicles may lead to temporal fusion of polluted BEV feature map. The stacking strategy may integrate features from specific previous frames, offering flexibility in fusion of long-range information. In this paper, we propose a key-frame-based streaming strategy, leveraging local and global fusion capabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.512, + 0.745, + 0.528 + ], + "angle": 0, + "content": "2.3. Map Element Interaction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.779 + ], + "angle": 0, + "content": "ADMap [15] explores point-order relationships between and within instances through a cascading approach. InsightMapper [48] performs inner-instance feature aggregation by an additional masked inner-instance self-attention module. GeMap [54] designs a masked decoupled self-attention to handle the shapes and relations of the instance independently. HoMap [1] introduces high-order modeling to capture the correlations between instances using high-order statistics. HIMap [57] proposed a hybrid framework to learn and interact with information at the point-level and the element-level. As proved in [14], modeling relations between objects is beneficial to object recognition. In state-of-the-art methods [9, 12], relationship of queries improves DETR-like models. In this paper, we explore the interaction of map elements between their geometry and appearance feature by our explicit position relation embedding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.84, + 0.803 + ], + "angle": 0, + "content": "2.4. Classification-Localization Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The classification head and the localization head of object detection are implemented separately by two branches in parallel, causing inconsistency in the output distribution. This misalignment problem has been well researched in the 2D object detection field [21, 51]. However, it is overlooked by the map element detectors [7, 24, 25, 32, 39, 50]. In this" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17177" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.217, + 0.908, + 0.275 + ], + "angle": 0, + "content": "Figure 2. Overview of InteractionMap framework. InteractionMap mainly consists of four components: (1) BEV encoder transforms sensor input to a unified BEV representation; (2) Key-frame-based temporal fusion module leverage temporal information from local to global; (3) Relation map decoder utilizes relation embedding in point-level and instance-level; (4) Geometry-aware alignment module is designed to solve the misalignment problem of classification and position output." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.291, + 0.482, + 0.368 + ], + "angle": 0, + "content": "paper, we create a geometry-aware classification loss, making use of focal loss [42] with a geometry-aware target in foreground candidates. Moreover, we present a geometry-aware matching cost to suppress the matching of background candidates." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.377, + 0.182, + 0.393 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.403, + 0.485, + 0.645 + ], + "angle": 0, + "content": "Figure 2 illustrates the framework of our proposed InteractionMap. Firstly, BEV features are extracted from multiview images by the BEV encoder, which consists of a backbone [11] to extract multi-scale image features, and a view transformation module [22, 31] to encode image features into a BEV representation \\(\\mathcal{F}_{local}\\). Secondly, key-frame-based temporal fusion leverages long-range temporal information by local fusion and global fusion (Section 3.1). Subsequently, the relation map decoder utilizes point-wise relation embedding (PRE) and instance-wise relation embedding (IRE) for better interaction between queries (Section 3.2). Finally, by introducing a geometric-aware classification loss and a geometric-aware matching cost, the geometry-aware alignment leverages interaction between semantic information and geometry information (Section 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.651, + 0.395, + 0.668 + ], + "angle": 0, + "content": "3.1. Key-frame-based Temporal Fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Compared to the single-frame map element prediction method, there are mainly two types of temporal fusion methods, namely stacking-based strategy and streaming-based strategy. The streaming strategy leverages all historical information by hidden state encoding. However, due to limited memory capacity, the streaming strategy has limited performance in long-range perception. The stacking strategy fuse features from specific previous frames, offering flexibility in integration of long-range information. The computational cost is linearly related to the number of fused frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We propose a key-frame-based streaming (KFS) strategy to address these problems. The well-known key-frame-based strategy is widely used in robot navigation and mapping, such as sparse mapping [16, 35] and dense map" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.908, + 0.518 + ], + "angle": 0, + "content": "ping [6, 56]. Online reconstruction methods are unstable in long range, but they are accurate in the local regime. The input frames are divided into local frame segments. The submaps are reconstructed by integration of local frames, and then the global map is reconstructed by integration of submaps. We introduce two types of KFS strategy, namely KFS-streaming and KFS-stacking, as shown in Figure 3. Firstly, the BEV feature embedding submaps are generated by a streaming-base temporal encoder at the bottom level, which keeps the stability and temporal consistency of predictions cross latest frames. Secondly, the long-range BEV feature embedding is generated by a streaming-based or stacking-based temporal encoder at the top level. Figure 3 illustrates the KFS strategy that merges one previous submap with an interval of two." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.535, + 0.696, + 0.55 + ], + "angle": 0, + "content": "3.1.1 Local BEV Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.907, + 0.656 + ], + "angle": 0, + "content": "We warp the memorized BEV feature map from the previous frame to the current frame recurrently based on the ego vehicle's relative pose. Then we employ a GRU [5] to fuse current BEV feature map \\(\\mathcal{F}_{local}^{t}\\) and the warped memorized BEV feature map \\(\\tilde{\\mathcal{F}}_{submap}^{t-1}\\) into a single BEV feature embedding submap candidate \\(\\mathcal{F}_{submap}^{t}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.664, + 0.905, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {F}} _ {s u b m a p} ^ {t - 1} = \\operatorname {W a r p} \\left(\\mathcal {F} _ {s u b m a p} ^ {t - 1}, \\boldsymbol {T}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.69, + 0.905, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\text {s u b m a p}} ^ {t} = \\operatorname {R e s B l o c k} \\left(L N \\left(G R U \\left(\\tilde {\\mathcal {F}} _ {\\text {s u b m a p}} ^ {t - 1}, \\mathcal {F} _ {\\text {l o c a l}} ^ {t}\\right)\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.723, + 0.907, + 0.785 + ], + "angle": 0, + "content": "Where ResBlock is a residual-based convolution block with the shortcut connection [11]. \\(LN\\) is a layer normalization operation. \\(\\pmb{T}\\) denotes a standard \\(4\\times 4\\) transformation matrix between the coordinate systems of two frames." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.704, + 0.815 + ], + "angle": 0, + "content": "3.1.2 Global BEV Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.871 + ], + "angle": 0, + "content": "Our KFS strategy selects a BEV feature embedding as a submap using a specific distance stride \\( d_{stride} \\), instead of a specific temporal interval customly." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "KFS-streaming strategy The previous selected global BEV feature is warped to the current frame. The warped" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17178" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.473, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.309, + 0.393, + 0.324 + ], + "angle": 0, + "content": "Figure 3. Temporal fusion strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.347, + 0.483, + 0.378 + ], + "angle": 0, + "content": "global BEV feature and the current BEV feature map are then fused to create a global BEV feature map." + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.384, + 0.482, + 0.405 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {F}} _ {\\text {g l o b a l}} ^ {\\text {p r e}} = \\operatorname {W a r p} \\left(\\mathcal {F} _ {\\text {g l o b a l}} ^ {\\text {p r e}}, \\boldsymbol {T}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.411, + 0.483, + 0.432 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {g l o b a l} ^ {t} = L N \\left(G R U \\left(\\tilde {\\mathcal {F}} _ {g l o b a l} ^ {p r e}, \\mathcal {F} _ {s u b m a p} ^ {t}\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.484, + 0.478 + ], + "angle": 0, + "content": "KFS-stacking strategy We warp \\( N_{pre} \\) selected submaps to the current frame. Following concatenation and residual-based convolution [11], we obtain the global BEV feature." + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.485, + 0.482, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {F}} _ {s u b m a p} ^ {t _ {k}} = \\operatorname {W a r p} \\left(\\mathcal {F} _ {s u b m a p} ^ {t _ {k}}, T\\right) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.513, + 0.482, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\text {g l o b a l}} ^ {t} = \\operatorname {R e s B l o c k} \\left(\\operatorname {C o n c a t} \\left(\\tilde {F} _ {\\text {s u b m a p}}, \\mathcal {F} _ {\\text {s u b m a p}} ^ {t}\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.537, + 0.485, + 0.635 + ], + "angle": 0, + "content": "Here, \\(\\tilde{\\mathcal{F}}_{submap}^{t_k}\\) and \\(t_k\\) represent the selected BEV feature submap and its corresponding index. The collection of selected BEV feature submaps is denoted as \\(\\tilde{F}_{submap} = \\{\\tilde{\\mathcal{F}}_{submap}^{t_k}\\}_{k=1}^{N_{pre}}\\). During the training period, the previous \\(N_{pre}\\) submaps are randomly selected within a range of \\(N_{pre} \\times d_{stride}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.641, + 0.296, + 0.657 + ], + "angle": 0, + "content": "3.2. Relation map decoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.484, + 0.738 + ], + "angle": 0, + "content": "Previous work has shown the effectiveness of interactions for map element detectors [15, 57]. In contrast to these approaches, we directly construct explicit relation priors by incorporating relation embedding into the self-attention module of the decoder." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.29, + 0.773 + ], + "angle": 0, + "content": "3.2.1 Relation Embedding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Here is the self-attention in the decoder of a DETR-like framework." + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.815, + 0.482, + 0.851 + ], + "angle": 0, + "content": "\\[\nQ _ {\\text {s e l f - a t t n}} = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Here \\(d_k\\) is the dimension of the keys. \\(Q, K, V\\) are feature embeddings generated by linear projections. \\(Q = \\operatorname{Linear}_q(Q_{in})\\). \\(K = \\operatorname{Linear}_k(Q_{in})\\). \\(V = \\operatorname{Linear}_v(Q_{in})\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "We incorporate relation embeddings into the self-attention model, inspired by [14]." + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.126, + 0.905, + 0.161 + ], + "angle": 0, + "content": "\\[\nQ _ {s e l f - a t t n} = \\operatorname {S o f t m a x} \\left(\\mathcal {F} _ {r e l} + \\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.167, + 0.907, + 0.243 + ], + "angle": 0, + "content": "Here \\(\\mathcal{F}_{rel}\\) represents the relation embedding. In this paper, point-wise relation embedding and instance-wise relation embedding are added to the decoupled self-attention model at the point level and the instance level, respectively, employing attention refinement of information interaction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.26, + 0.788, + 0.275 + ], + "angle": 0, + "content": "3.2.2 Point-wise Relation Embedding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.284, + 0.906, + 0.345 + ], + "angle": 0, + "content": "The point-wise relation encoder represents the high-dimensional point relation embedding as an explicit geometry prior. The point-wise relation embedding is generated by normalized point coordinates and edge directions." + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.353, + 0.905, + 0.371 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e l} _ {p t} (i, j) = \\left[ \\log \\left(x _ {i} - x _ {j} + 1\\right), \\log \\left(y _ {i} - y _ {j} + 1\\right) \\right] \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.378, + 0.915, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e l} _ {\\text {d i r}} (i, j) = \\operatorname {c o s i n e \\_ s i m} \\left(e _ {i}, e _ {i + 1}\\right) - \\operatorname {c o s i n e \\_ s i m} \\left(e _ {j}, e _ {j + 1}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.409, + 0.906, + 0.483 + ], + "angle": 0, + "content": "Here \\( x_{i} \\), \\( x_{j} \\) are normalized point coordinates. \\( e_{i} \\) and \\( e_{i+1} \\), \\( e_{j} \\) and \\( e_{j+1} \\) are adjacent edges. The cosine similarity is denoted by \\( \\cosin\\_sim \\). The point-wise relation embedding is unbiased, since both \\( Rel_{pt}(i,j) \\) and \\( Rel_{dir}(i,j) \\) are equal to 0 when \\( i = j \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.483, + 0.905, + 0.515 + ], + "angle": 0, + "content": "The point-wise relation embedding \\(\\mathcal{F}_{rel}^{pt}\\) is illustrated by the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.522, + 0.905, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {r e l} ^ {p t} = \\operatorname {R e L U} (\\operatorname {L i n e a r} (S P E (\\operatorname {C o n c a t} (R e l _ {p t}, R e l _ {d i r})))) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.629 + ], + "angle": 0, + "content": "Firstly, the point-wise relation embedding is calculated by concatenation of \\( Rel_{pt} \\) and \\( Rel_{dir} \\). Then the embedding is transformed to high dimension by sine positional encoding (SPE). Lastly, the embedding is transformed by a linear projection and a ReLU function." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.647, + 0.81, + 0.663 + ], + "angle": 0, + "content": "3.2.3 Instance-wise Relation Embedding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.905, + 0.701 + ], + "angle": 0, + "content": "The instance-wise relation embedding is generated by classification scores and chamfer distance of points." + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.71, + 0.905, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e l} _ {s} (i, j) = \\operatorname {s g n} \\left(s _ {i} - s _ {j}\\right) \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.735, + 0.905, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e l} _ {s d} (i, j) = \\operatorname {s g n} \\left(s _ {i} - s _ {j}\\right) \\cdot \\operatorname {C h a m f e r} \\left(I _ {i}, I _ {j}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.755, + 0.907, + 0.875 + ], + "angle": 0, + "content": "Here, \\( sgn \\) is the sign function. \\( s_i, s_j \\) are classification scores of the \\( i^{th} \\) and \\( j^{th} \\) candidates. Chamfer \\( (I_i, I_j) \\) is the chamfer distance between points of instance \\( i \\) and instance \\( j \\). The equation of \\( Rel_s(i,j) \\) establishes a simple classification ranking relation among pairs of instances. Then we add the position relation of the map elements to the instance-wise relation embedding \\( \\mathcal{F}_{rel}^{ins} \\) by the chamfer distance." + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.885, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {r e l} ^ {i n s} = R e L U \\left(L i n e a r \\left(S P E \\left(R e l _ {s d}\\right)\\right)\\right) \\tag {14}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17179" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "The instance-wise relation embedding is unbiased, since \\( Rel_{sd}(i,j) \\) is equal to 0 when \\( i = j \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.132, + 0.344, + 0.148 + ], + "angle": 0, + "content": "3.3. Geometry-aware Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.484, + 0.307 + ], + "angle": 0, + "content": "Current work, such as MapTR [24] uses the classification branch and the regression branch to predict semantic labels and geometric localization separately. MapTR uses the focal loss [42] as the classification loss and the point-to-point L1 loss as the regression loss. Since classification and regression losses are independent, the quality of these two branches is inconsistent, leading to the well-known misalignment problem. In this paper, we introduce the geometry-aware focal loss and the geometry-aware focal cost to overcome this problem." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.325, + 0.342, + 0.34 + ], + "angle": 0, + "content": "3.3.1 Geometry-aware Focal Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.484, + 0.409 + ], + "angle": 0, + "content": "Inspired by IoU-aware Focal Loss [21, 51] in traditional object detection, we design the novel Geometry-aware Focal Loss (GFL). Unlike focal loss, we treat positives and negatives asymmetrically." + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.42, + 0.49, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G F L} = \\sum_ {i = 1} ^ {N _ {\\text {p o s}}} s _ {\\text {g e o} _ {i}} B C E \\left(s _ {\\text {g e o} _ {i}}, p _ {i}\\right) + \\sum_ {j = 1} ^ {N _ {\\text {n e g}}} \\alpha p _ {j} ^ {\\gamma} B C E \\left(p _ {j}, 0\\right) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.485, + 0.556 + ], + "angle": 0, + "content": "where \\( p_i \\) and \\( p_j \\) are the predicted probability of the \\( i^{th} \\) and \\( j^{th} \\) candidates, \\( s_{geo} \\in [0,1] \\) is an instance Geometry-aware Classification Score (GCS). Here we define three types of GCS, including \\( s_{p2p} \\), \\( s_{dir} \\) and \\( s_{giou} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.567, + 0.483, + 0.61 + ], + "angle": 0, + "content": "\\[\ns _ {p 2 p} = 1 - \\frac {1}{2 N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} D _ {\\text {M a n h a t t a n}} \\left(p _ {\\text {s r c} _ {i}}, p _ {\\text {t g t} _ {i}}\\right) \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.621, + 0.483, + 0.663 + ], + "angle": 0, + "content": "\\[\ns _ {d i r} = 0. 5 + \\frac {1}{2 N _ {e}} \\sum_ {i = 1} ^ {N _ {e}} \\operatorname {c o s i n e \\_ s i m} \\left(e _ {s r c _ {i}}, e _ {t g t _ {i}}\\right) \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.669, + 0.483, + 0.687 + ], + "angle": 0, + "content": "\\[\ns _ {g i o u} = 0. 5 + 0. 5 \\cdot G I o U \\left(B _ {s r c}, B _ {t g t}\\right) \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.484, + 0.829 + ], + "angle": 0, + "content": "where \\( s_{p2p} \\in [0,1] \\) is a normalized point-to-point L1 score, \\( s_{dir} \\in [0,1] \\) is an edge direction score calculated by cosine similarity and \\( s_{giou} \\in [0,1] \\) is a normalized GIoU [40] score of candidate bounding box \\( B_{src} \\) and assigned GT \\( B_{tgt} \\). The bounding box is computed by minimum enclosing box of instance points. \\( N_p \\) and \\( N_e \\) are the numbers of points and edges. \\( p_{src} \\) and \\( p_{tgt} \\) are normalized points of candidates and assigned GT. \\( e_{src} \\) and \\( e_{tgt} \\) are edges of candidates and assigned GT." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.341, + 0.863 + ], + "angle": 0, + "content": "3.3.2 Geometry-aware Focal Cost" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The loss function \\(\\mathcal{L}_{GFL}\\) supervises the classification scores with geometric metrics GCS \\(s_{geo_i}\\). And \\(s_{geo_i}\\) accurately" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "selects these high-quality candidates, which leads to better map vectorization performance. Following the spirit of GFL, we make similar modifications to the focal cost, which is a key component of label assignment. The novel matching cost, geometry-aware focal cost (GFC) is as follows." + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.172, + 0.907, + 0.205 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {G F C} (i, j) = s _ {g e o _ {i}} B C E \\left(s _ {g e o _ {i}}, p _ {i}\\right) - \\alpha p _ {i} ^ {\\gamma} B C E \\left(1 - p _ {i}, 1\\right) \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.21, + 0.907, + 0.287 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_{GFC}(i,j)\\) is the geometry-aware focal cost for the \\(i^{th}\\) prediction and the \\(j^{th}\\) ground truth. The GFC is used as a modulated function to suppress candidates with inaccurate prediction localization and promote candidates with an accurate position." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.293, + 0.658, + 0.309 + ], + "angle": 0, + "content": "3.4. Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.906, + 0.407 + ], + "angle": 0, + "content": "Auxiliary BEV Supervision. We introduce a query-based instance segmentation (QIS) module using an instance mask prediction branch. Inspired by [17], we make use of the query embeddings \\( Q_{dec} \\) from Transformer to dot-product the BEV feature embedding map \\( \\mathcal{F}_{bev} \\) to obtain instance-wise binary masks \\( m \\) with the sigmoid function \\( \\sigma(\\cdot) \\)." + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.412, + 0.906, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\hat {M} _ {\\text {i n s}} = \\sigma (M L P \\left(Q _ {\\text {d e c}}\\right) \\cdot \\mathcal {F} _ {\\text {b e v}}) \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.435, + 0.907, + 0.495 + ], + "angle": 0, + "content": "For the instance segmentation branch, we extend the mask focal loss and the mask focal cost to mask geometry-aware focal loss (MGFL) and mask geometry-aware focal cost (MGFC) as well, with the same ideal of GFL." + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.516, + 0.909, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s e g}} = \\lambda_ {m g f} \\mathcal {L} _ {m g f} (\\hat {M} _ {\\text {i n s}}, M _ {\\text {i n s}}) + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} (\\hat {M} _ {\\text {i n s}}, M _ {\\text {i n s}}) \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.548, + 0.906, + 0.593 + ], + "angle": 0, + "content": "Instance masks \\(\\hat{M}_{ins}\\) are supervised by instance mask annotations \\(M_{ins}\\), using the mask geometry-aware focal loss \\(\\mathcal{L}_{mgf}\\) and the dice loss \\(\\mathcal{L}_{dice}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.593, + 0.906, + 0.621 + ], + "angle": 0, + "content": "Overall Loss. The total loss \\( L \\) used by InteractionMap is as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.624, + 0.905, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {d e t}} + \\mathcal {L} _ {\\text {s e g}} + \\mathcal {L} _ {\\text {a u x}} \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.906, + 0.686 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{det}\\) is the loss term for the detection task. \\(\\mathcal{L}_{seg}\\) is the loss term for the segmentation task. \\(\\mathcal{L}_{aux}\\) is the auxiliary loss term." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.687, + 0.906, + 0.719 + ], + "angle": 0, + "content": "The detection loss consists of classification loss \\(\\mathcal{L}_{cls}\\), point-to-point distance loss \\(\\mathcal{L}_{p2p}\\), edge direction loss \\(\\mathcal{L}_{dir}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.724, + 0.905, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e t}} = \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\lambda_ {p 2 p} \\mathcal {L} _ {p 2 p} + \\lambda_ {\\text {d i r}} \\mathcal {L} _ {\\text {d i r}} \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.906, + 0.822 + ], + "angle": 0, + "content": "The loss term \\(\\mathcal{L}_{cls}\\) is the geometry-aware focal loss for the classification of instances. The loss term \\(\\mathcal{L}_{p2p}\\) is the smooth L1 loss for the regression of vectorized map elements. And the loss term \\(\\mathcal{L}_{dir}\\) is the loss of cosine similarity between adjacent edges." + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.837, + 0.905, + 0.854 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a u x}} = \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {P V S e g} \\mathcal {L} _ {P V S e g} \\tag {24}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "The auxiliary loss term \\(\\mathcal{L}_{aux}\\) is composed of the depth-map prediction loss \\(\\mathcal{L}_{depth}\\) and image semantic segmentation loss \\(\\mathcal{L}_{PVSeg}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17180" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.216, + 0.108 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.297, + 0.132 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.14, + 0.205, + 0.153 + ], + "angle": 0, + "content": "4.1.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.163, + 0.483, + 0.299 + ], + "angle": 0, + "content": "nuScenes Datasets. There are 1000 scenes of 4 locations in nuScenes datasets. Each scene contains 20s of RGB images from six cameras, point clouds from LiDAR sweeps, and a 3D vectorized map. The train set contains 700 scenes with 28130 samples and validation set contains 150 scenes with 6019 samples, respectively. Following previous works [24, 25], we mainly focus on three categories of map elements, including lane divider, pedestrian crossing and road boundary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.483, + 0.419 + ], + "angle": 0, + "content": "Argoverse2 Datasets. Argoverse2 dataset provide 1000 scenes in 6 cities, each lasting around 15 seconds. Each log includes 7 RGB images from surrounding cameras and point cloud from LiDAR sweeps. Follow previous methods [24, 25], 700 scenes are used for training, and 150 scenes are used for validation. For a fair comparison, we report results on its validation set and focus on the same three map categories as the nuScene dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.279, + 0.455 + ], + "angle": 0, + "content": "4.1.2 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.614 + ], + "angle": 0, + "content": "For a fair comparison, we follow the standard metric used in previous works [19, 24, 25, 29]. With the ego-coordination as the center, the perception ranges are \\([-15m, 15m]\\) for the X-axis and \\([-30m, 30m]\\) for the Y-axis. We adopt the mean Average Precision (mAP) to evaluate the map vectorization quality based on the chamfer distance. A candidate is considered as True-Positive (TP) only if the chamfer distance to ground truth is less than certain thresholds \\((\\tau \\in T, T = 0.5, 1.0, 1.5)\\). The final AP metric is calculated by average across all thresholds and all classes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.309, + 0.65 + ], + "angle": 0, + "content": "4.1.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Following previous method [24, 25], we use ResNet50 [11] as the backbone for RGB images. All models are trained with 8 A800 GPUs. The default batch size is 32. The initial learning rate is set to \\(6 \\times 10^{-4}\\) with cosine decay. The nuScenes images resolution is \\(1600 \\times 900\\). We resize the images with 0.5 ratio. For the Argoverse2 dataset, the image resolutions are \\(2048 \\times 1550\\) and \\(1550 \\times 2048\\). We pad the images to \\(2048 \\times 2048\\), and then resize the images with 0.3 ratio. We define the size of BEV feature as \\(H \\times W\\) of \\(200 \\times 100\\), and the size of each BEV grid as 0.3 meters. The default numbers of instance queries, point queries, and decoder layers are 100, 20, 6, respectively. The model is trained on 24, 110 epochs on the nuScenes and Argoverse2 datasets. We randomly divide each training sequence into 3 splits to generate more diverse data sequences. Inspired by SOLOFusion [37], we train the ini" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "tial \\(N_{init}\\) epochs with single-frame input to stabilize multi-frame straining. \\(N_{init} = 0.5N_{total}\\). For loss weights, \\(\\lambda_{cls} = 2\\), \\(\\lambda_{p2p} = 4\\), \\(\\lambda_{dir} = 0.005\\), \\(\\lambda_{mgf} = 30\\), \\(\\lambda_{dice} = 3\\), \\(\\lambda_{depth} = 3\\), \\(\\lambda_{PVSeg} = 2\\). For temporal fusion, \\(N_{pre} = 4\\), \\(d_{stride} = 5m\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.178, + 0.749, + 0.194 + ], + "angle": 0, + "content": "4.2. Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.202, + 0.905, + 0.337 + ], + "angle": 0, + "content": "Performance on nuScenes Dataset. As shown in Table 1, our methods show consistent improvements over baselines. By integrating our method with MapTRv2 [25], we achieve \\(71.6(+10.1)\\), \\(71.8(+10.3)\\) mAP with 24-epoch training schedules and \\(75.5(+6.8)\\), \\(76.0(+7.3)\\) mAP with 110-epoch training schedules, respectively. InteractionMap establishes a new state-of-the-art performance, exhibiting significant improvements over existing methods with comparable inference speed, as shown in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.338, + 0.906, + 0.473 + ], + "angle": 0, + "content": "Performance on Argoverse2 Dataset. Argoverse2 dataset contains 3D map elements, including height information. For fair comparison, the sampling frequency is set to \\(2\\mathrm{Hz}\\) and \\(10\\mathrm{Hz}\\), respectively. The amount of training data at a \\(10\\mathrm{Hz}\\) sampling frequency is 5 times as many as that at a \\(2\\mathrm{Hz}\\) sampling frequency. As demonstrated in Table 2 and Table 3, InteractionMap surpasses all state-of-the-art methods in both 2D and 3D vectorized map perception on Argoverse2 dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.485, + 0.667, + 0.501 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.905, + 0.599 + ], + "angle": 0, + "content": "We examine the efficacy of each component of InteractionMap through ablation studies, utilizing the nuScenes dataset. Initially, we build a simple baseline base on MapTR [24] with the LSS encoder without auxiliary depth-map supervision. The default numbers of instances, points per instance, are 50 and 20, respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.6, + 0.905, + 0.75 + ], + "angle": 0, + "content": "Contributions of Main Components. Table 5 demonstrates the impact of each component of InteractionMap. When adding relation embedding into the baseline model, there is a substantial improvement of \\(4.5\\mathrm{mAP}\\). Furthermore, the addition of geometry alignment results in a significant 7.5 increase in mAP, emphasizing its effectiveness in boosting performance. Lastly, key-frame-based temporal fusion offers 5.0 and 5.1 additional increases in mAP, indicating that the inclusion of long-range hierarchical streaming strategy significantly enhances the performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Ablation on Relation Embedding. In Table 6, we explore the effect of relation embedding in point-level and instance-level. Adding the point-wise relation embedding of point coordinates and edge direction results in 2.1 and 0.9 improvements in mAP over the baseline, respectively. The instance-wise relation embedding leverages the classification confidence and points coordinates, leading to enhancements of 1.1 and \\(3.1\\mathrm{mAP}\\), respectively. The combination of relation embedding in point-level and instance-level enhances performance by \\(2.8\\mathrm{mAP}\\). Finally, we introduce a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17181" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.202, + 0.089, + 0.798, + 0.457 + ], + "angle": 0, + "content": "
MethodBackboneEpochsTemporal\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bou}\\)mAP
MapTRR502446.351.553.150.3
MapVRR502447.754.451.451.2
PivotNetR503053.855.859.657.4
BeMapNetR503057.762.359.459.8
MapTRv2R502459.862.462.461.5
StreamMapNetR5030✓61.766.362.163.4
MGMapR502461.865.067.564.8
SQD-MapNetR5024✓63.666.664.865.0
MapQRR502463.468.067.766.4
HIMapR503062.668.469.166.7
HRMapNet-MapTRv2R5024✓65.867.468.567.2
InteractionMap-RR5024✓71.370.872.871.6
InteractionMap-CR5024✓69.772.773.071.8
VectorMapNetR50110+ft42.551.444.146.0
MapTRR5011056.259.860.158.7
MapVRR5011055.061.859.458.8
BeMapNetR5011062.666.765.164.8
MapTRv2R5011068.168.369.768.7
MapQRR5011070.174.473.272.6
HRMapNet-MapTRv2R50110✓72.072.975.873.6
HIMapR5011071.375.074.773.7
InteractionMap-RR50110✓75.275.076.175.5
InteractionMap-CR50110✓75.375.677.076.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.47, + 0.907, + 0.514 + ], + "angle": 0, + "content": "Table 1. Comparison with SOTA methods on the nuScenes validation set at \\(60m\\times 30m\\) perception range. \"EB0\", \"EB4\", \"R50\" correspond to the backbones Efficient-B0, Efficient-B4 [44], ResNet50 [11]. \"ft\" means the two-stage fine-tune strategy. \"C\" means KFS-stacking strategy and \"R\" means KFS-streaming strategy." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.522, + 0.477, + 0.641 + ], + "angle": 0, + "content": "
MethodDimAPpedAPdivAPboumAP
StreamMapNet2d62.059.563.061.5
SQD-MapNet2d64.960.264.963.3
MapTRv22d60.068.764.264.3
HRMapNet-MapTRv22d65.171.468.668.3
InteractionMap-C2d68.577.974.673.7
InteractionMap-R2d70.177.074.073.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.697 + ], + "angle": 0, + "content": "Table 2. Comparison with SOTA methods on the Argoverse2 validation set at \\(60m \\times 30m\\) perception range with a 30-epoch training schedule, at a 2Hz sampling frequency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.749 + ], + "angle": 0, + "content": "alternating training strategy to randomly turn off one of these two components, leading the overall design achieves \\(54.7(+4.5)\\) mAP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.886 + ], + "angle": 0, + "content": "Ablation on Key-frame-based Temporal Fusion. To investigate the effectiveness of the key-frame-based streaming strategy, we conduct ablation experiments to compare it with the baseline (without temporal fusion) and the streaming strategy, in Table 7. Employing the streaming strategy results in \\(3.8\\mathrm{mAP}\\) over the baseline. Furthermore, when the KFS-stacking strategy is introduced, the mAP increased by 7.4, reaching 57.6. Additionally, the inclusion of KFS-streaming strategy leads to an improvement of \\(7.2\\mathrm{mAP}\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Ablation on Geometry-aware Alignment. As shown in" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.522, + 0.9, + 0.701 + ], + "angle": 0, + "content": "
MethodDimAPpedAPdivAPboumAP
MapTRv22d63.671.567.467.5
MapQR2d64.372.368.168.2
HIMap2d69.069.570.369.6
InteractionMap-C2d69.878.674.874.4
InteractionMap-R2d70.878.174.974.6
MapTRv23d60.768.964.564.7
MapQR3d60.171.266.265.9
HIMap3d66.768.370.368.4
InteractionMap-C3d66.675.672.771.6
InteractionMap-R3d67.775.573.172.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.713, + 0.907, + 0.757 + ], + "angle": 0, + "content": "Table 3. Comparison with SOTA methods on the Argoverse2 validation set at \\(60m \\times 30m\\) perception range with a 6-epoch training schedule, at a \\(10\\mathrm{Hz}\\) sampling frequency." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.765, + 0.9, + 0.839 + ], + "angle": 0, + "content": "
MethodFPSParams(MB)mAP
MapTRv214.940.661.5
InteractionMap-C11.0116.771.8
InteractionMap-R12.149.971.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.851, + 0.907, + 0.879 + ], + "angle": 0, + "content": "Table 4. Comparison with SOTA methods on the nuScenes validation set. FPS is measured on a single A800 GPU." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17182" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.478, + 0.231 + ], + "angle": 0, + "content": "
REMGAMTFMAPpedAPdivAPboumAP
46.750.553.550.2
✓50.156.657.354.7
✓57.163.863.761.5
C50.960.061.957.6
R50.661.360.357.4
✓✓58.665.162.962.2
✓✓C64.268.668.867.2
✓✓R64.168.169.867.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.241, + 0.483, + 0.3 + ], + "angle": 0, + "content": "Table 5. Ablation study of each component, including Relation Embedding Module (REM), Temporal Fusion Module (TFM), Geometry-aware Alignment Module (GAM). \"C\" means KFS-stacking strategy and \"R\" means KFS-streaming strategy." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.31, + 0.478, + 0.452 + ], + "angle": 0, + "content": "
PtxyPtdirInsSInssdPPAPpedAPdivAPboumAP
46.750.553.550.2
✓47.852.756.352.3
✓47.052.453.951.1
✓✓47.855.656.453.3
✓46.952.154.851.3
✓47.955.756.453.3
✓✓✓47.655.955.553.0
✓✓✓✓50.156.657.354.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.547 + ], + "angle": 0, + "content": "Table 6. Ablation study of relation embedding by employing point-wise coordinate embedding \\(\\mathrm{Pt}_{xy}\\), point-wise edge direction embedding \\(\\mathrm{Pt}_{dir}\\), instance-wise classification score embedding \\(\\mathrm{Ins}_s\\), instance-wise classifier score and chamfer distance embedding \\(\\mathrm{Ins}_{sd}\\) and alternating training strategy of relation embedding, namely Ping-Pong (PP)." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.558, + 0.478, + 0.647 + ], + "angle": 0, + "content": "
MethodAPpedAPdivAPboumAP
Baseline46.750.553.550.2
Streaming48.654.758.854.0
KFS-stacking50.960.061.957.6
KFS-streaming50.661.360.357.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Table 7. Ablation study of streaming strategy, KFS-streaming strategy and KFS-stacking strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Table 8, geometry-aware alignment comprised five main components: query-based instance segmentation (QIS), geometry-aware focal loss (GFL), geometry-aware focal cost (GFC), mask geometry-aware focal loss (MGFL) and mask geometry-aware focal cost (MGFC). The QIS leverages geometry priors by instance segmentation mask supervision. Compared to the baseline, the QIS leads to an improvement of \\(6.4\\mathrm{mAP}\\). The GFL encourages predictions with high localization metrics to obtain a better classification score, with an improvement of \\(3.1\\mathrm{mAP}\\) compared to mask focal loss. The GFC suppresses the matching cost of candidates with an inaccurate prediction position, which brings an improvement of \\(0.9\\mathrm{mAP}\\). For the segmentation" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.09, + 0.9, + 0.206 + ], + "angle": 0, + "content": "
QISGFLGFCMGFLMGFCAPpedAPdivAPboumAP
46.750.553.550.2
✓51.758.759.456.6
✓✓55.162.062.059.7
✓✓✓55.863.862.260.6
✓✓✓✓57.063.662.361.0
✓✓✓✓✓57.163.863.761.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.217, + 0.907, + 0.288 + ], + "angle": 0, + "content": "Table 8. Ablation study of geometry-aware alignment by investigating the performance of query-based instance segmentation (QIS), geometry-aware focal loss (GFL), geometry-aware focal cost (GFC), mask geometry-aware focal loss (MGFL) and mask geometry-aware focal cost (MGFC)." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.299, + 0.9, + 0.402 + ], + "angle": 0, + "content": "
SgiouSp2pSdirAPpedAPdivAPboumAP
✓51.758.759.456.6
52.656.258.055.6
✓54.860.263.159.4
✓58.363.162.561.3
✓✓57.163.863.761.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.412, + 0.898, + 0.427 + ], + "angle": 0, + "content": "Table 9. Ablation study of geometry-aware classification score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.502 + ], + "angle": 0, + "content": "branch, the result shows that the position-supervised segmentation loss MGFL and the position-modulated segmentation cost MGFC improve the final results, with \\(0.4\\mathrm{mAP}\\) and \\(0.5\\mathrm{mAP}\\) gains, individually." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.502, + 0.906, + 0.579 + ], + "angle": 0, + "content": "We further investigate the geometry-aware classification score \\( S_{geo} \\). As depicted in Table 9, using \\( S_{giou} \\), \\( S_{p2p} \\) and \\( S_{dir} \\) leads to individual improvement of -1.0, 2.8 and 4.7 mAP, respectively. Furthermore, the combination of \\( S_{p2p} \\) and \\( S_{dir} \\) boosts the performance, improving by 4.9 mAP." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.594, + 0.634, + 0.61 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.907, + 0.831 + ], + "angle": 0, + "content": "In this paper, we introduce InteractionMap, a novel approach to end-to-end online vectorized HD map utilizing temporal and spatial information interaction from local to global. By leveraging relation embedding at both point level and instance level, the network learns the geometric priors of map elements better. Subsequently, key-frame-based streaming strategy enhances the network by utilizing temporal feature embedding from near to distant. Finally, the geometric-aware focal loss and the geometric-aware focal cost build a strong correlation between classification score and instance-point location precision. Across various experimental settings, our proposed InteractionMap yields significant performance improvements over the baseline on various datasets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.848, + 0.611, + 0.864 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.521, + 0.873, + 0.907, + 0.903 + ], + "angle": 0, + "content": "[1] Yingfeng Cai, Wei Dong, Ze Liu, Hai Wang, and Long Chen. Homap: End-to-end vectorized hd map construction with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17183" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "high-order modeling. IEEE Transactions on Intelligent Vehicles, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.19, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[3] Jiacheng Chen, Yuefan Wu, Jiaqi Tan, Hang Ma, and Yasutaka Furukawa. Maptracker: Tracking with strided memory fusion for consistent vector hd mapping. In European Conference on Computer Vision, pages 90-107. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.261, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[4] Shaoyu Chen, Tianheng Cheng, Xinggang Wang, Wenming Meng, Qian Zhang, and Wenyu Liu. Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. arXiv preprint arXiv:2206.04584, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[5] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.386, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[6] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4): 1, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.456, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[7] Wenjie Ding, Limeng Qiao, Xi Qiu, and Chi Zhang. Pivot-net: Vectorized pivot learning for end-to-end hd map construction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3672-3682, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.512, + 0.482, + 0.592 + ], + "angle": 0, + "content": "[8] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525-11533, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.595, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[9] Yulu Gao, Yifan Sun, Xudong Ding, Chuyang Zhao, and Si Liu. Ease-detr: Easing the competition among object queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17282-17291, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.665, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[10] Chunrui Han, Jinrong Yang, Jianjian Sun, Zheng Ge, Runpei Dong, Hongyu Zhou, Weixin Mao, Yuang Peng, and Xi-angyu Zhang. Exploring recurrent long-term temporal fusion for multi-view 3d perception. IEEE Robotics and Automation Letters, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[12] Xiuquan Hou, Meiqin Liu, Senlin Zhang, Ping Wei, Badong Chen, and Xuguang Lan. Relation detr: Exploring explicit position relation prior for object detection. arXiv preprint arXiv:2407.11699, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[14] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3588-3597, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[15] Haotian Hu, Fanyi Wang, Yaonong Wang, Laifeng Hu, Jingwei Xu, and Zhiwang Zhang. Admap: Anti-disturbance framework for reconstructing online vectorized hd map. arXiv preprint arXiv:2401.13172, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[16] Georg Klein and David Murray. Parallel tracking and mapping for small ar workspaces. In 2007 6th IEEE and ACM international symposium on mixed and augmented reality, pages 225-234. IEEE, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[17] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[18] Hongyang Li, Chonghao Sima, Jifeng Dai, Wenhai Wang, Lewei Lu, Huijie Wang, Jia Zeng, Zhiqi Li, Jiazhi Yang, Hanming Deng, et al. Delving into the devils of bird's-eye-view perception: A review, evaluation and recipe. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[19] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online hd map construction and evaluation framework. In 2022 International Conference on Robotics and Automation (ICRA), pages 4628-4634. IEEE, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[20] Tianyu Li, Peijin Jia, Bangjun Wang, Li Chen, Kun Jiang, Junchi Yan, and Hongyang Li. Lanesegnet: Map learning with lane segment perception for autonomous driving. arXiv preprint arXiv:2312.16108, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[21] Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems, 33:21002-21012, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[23] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 541-556. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized hd map construction. arXiv preprint arXiv:2208.14437, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Bencheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17184" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv preprint arXiv:2308.05736, 2023.1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[26] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[27] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[28] Xiaolu Liu, Song Wang, Wentong Li, Ruizi Yang, Junbo Chen, and Jianke Zhu. Mgmap: Mask-guided learning for online vectorized hd map construction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14812-14821, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[29] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized hd map learning. In International Conference on Machine Learning, pages 22352-22369. PMLR, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[30] Zhi Liu, Shaoyu Chen, Xiaojie Guo, Xinggang Wang, Tianheng Cheng, Hongmei Zhu, Qian Zhang, Wenyu Liu, and Yi Zhang. Vision-based uneven bev representation learning with polar rasterization and surface estimation. In Conference on Robot Learning, pages 437-446. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.497 + ], + "angle": 0, + "content": "[31] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[32] Zihao Liu, Xiaoyu Zhang, Guangwei Liu, Ji Zhao, and Ningyi Xu. Leveraging enhanced queries of point sets for vectorized map construction. In European Conference on Computer Vision, pages 461-477. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[33] Yuexin Ma, Tai Wang, Xuyang Bai, Huitong Yang, Yuenan Hou, Yaming Wang, Yu Qiao, Ruigang Yang, and Xinge Zhu. Vision-centric bev perception: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[34] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651-3660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[35] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[36] Cong Pan, Yonghao He, Junran Peng, Qian Zhang, Wei Sui, and Zhaoxiang Zhang. Baeformer: Bi-directional and early interaction transformers for bird's eye view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9590-9599, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[37] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[38] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[39] Limeng Qiao, Wenjie Ding, Xi Qiu, and Chi Zhang. End-to-end vectorized hd-map construction with piecewiseBezier curve. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13218-13228, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[41] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11138-11147, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[42] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[43] Tixiao Shan and Brendan Englot. Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4758–4765. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[44] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.588, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[45] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[46] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3621-3631, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[47] Shuo Wang, Fan Jia, Yingfei Liu, Yucheng Zhao, Zehui Chen, Tiancai Wang, Chi Zhang, Xiangyu Zhang, and Feng Zhao. Stream query denoising for vectorized hd map construction. arXiv preprint arXiv:2401.09112, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[48] Zhenhua Xu, Kenneth KY Wong, and Hengshuang Zhao. Insightmapper: A closer look at inner-instance information for vectorized high-definition mapping. arXiv preprint arXiv:2308.08543, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[49] Tianyuan Yuan, Yicheng Liu, Yue Wang, Yilun Wang, and Hang Zhao. Streammapnet: Streaming mapping network for vectorized online hd map construction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7356-7365, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[50] Gongjie Zhang, Jiahao Lin, Shuang Wu, Zhipeng Luo, Yang Xue, Shijian Lu, Zuoguan Wang, et al. Online map vec" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17185" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "torization for autonomous driving: A rasterization perspective. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[51] Haoyang Zhang, Ying Wang, Feras Dayoub, and Niko Sunderhauf. Varifocalnet: An iou-aware dense object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8514-8523, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[52] Ji Zhang, Sanjiv Singh, et al. Loam: Lidar odometry and mapping in real-time. In Robotics: Science and systems, pages 1-9. Berkeley, CA, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[53] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. **Reverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743**, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[54] Zhixin Zhang, Yiyuan Zhang, Xiaohan Ding, Fusheng Jin, and Xiangyu Yue. Online vectorized hd map construction using geometry. arXiv preprint arXiv:2312.03341, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[55] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13760-13769, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[56] Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. Elastic fragments for dense scene reconstruction. In Proceedings of the IEEE International Conference on Computer Vision, pages 473-480, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[57] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized hd map construction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15396-15406, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.482, + 0.6 + ], + "angle": 0, + "content": "[58] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17186" + } + ] +] \ No newline at end of file diff --git a/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/447efac7-303b-4880-a09c-b0e941c7cbb1_origin.pdf b/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/447efac7-303b-4880-a09c-b0e941c7cbb1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2ef28555c1219aa4f7d5340339ffa7f9ecf8c160 --- /dev/null +++ b/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/447efac7-303b-4880-a09c-b0e941c7cbb1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1c122a9de28753a9956fdbbabcd5075c835c23dd553b013bc8d938e979998e4 +size 1991141 diff --git a/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/full.md b/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8bdb0e413183baaa5d105c4ba62f92eb97d3b42d --- /dev/null +++ b/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/full.md @@ -0,0 +1,408 @@ +# InteractionMap: Improving Online Vectorized HDMap Construction with Interaction + +Kuang Wu Chuan Yang * Zhanbin Li Langge Technology + +# Abstract + +Vectorized high-definition (HD) maps are essential for an autonomous driving system. Recently, state-of-the-art map vectorization methods are mainly based on DETR-like framework to generate HD maps in an end-to-end manner. In this paper, we propose InteractionMap, which improves previous map vectorization methods by fully leveraging local-to-global information interaction in both time and space. Firstly, we explore enhancing DETR-like detectors by explicit position relation prior from point-level to instance-level, since map elements contain strong shape priors. Secondly, we propose a key-frame-based hierarchical temporal fusion module, which interacts temporal information from local to global. Lastly, the separate classification branch and regression branch lead to the problem of misalignment in the output distribution. We interact semantic information with geometric information by introducing a novel geometric-aware classification loss in optimization and a geometric-aware matching cost in label assignment. InteractionMap achieves state-of-the-art performance on both nuScenes and Argoverse2 benchmarks. + +# 1. Introduction + +High-Definition (HD) maps are designed for high-precision autonomous driving, which contain instance-level vectorized representation such as lane divider, road boundaries, pedestrian crossing, etc. The rich semantic information of road topology and traffic rules is important for the navigation of autonomous driving (AD). HD maps are traditionally constructed offline using LiDAR SLAM-based methods [43, 52] with high maintenance costs, complex pipelines, and notable localization errors. In addition, manual annotation and map updating rely heavily on human labor and time demands. + +In recent years, more research endeavors have shifted towards deep-learning-based methods that construct vectorized HD maps around the ego-vehicle at runtime with on + +![](images/18454750ca09f64dbdff604cfcebc86fd6daf44b35bf22adc9b1b495680f809a.jpg) +Figure 1. Visual comparison between MapTRv2 [25] and our improved results. Our method effectively eliminates error map elements, leading to better precision and stability. + +board sensors [19, 20, 24]. With the development of bird's eye view (BEV) representation, recent approaches [4, 13, 19, 22, 38, 41, 55] treat the semantic map learning task as a semantic segmentation problem. However, rasterized maps generated by these methods lack vectorized instance-level information, which is essential to downstream tasks, such as motion forecasting and motion planning [8, 23]. To overcome the limitations of segmentation-based methods, approaches VectorMapNet [29], MapTR [24], MapTRv2 [25] predict point sets to construct end-to-end vectorized local HD maps using transformer [45] decoders inspired by DETR [2]. + +As shown in Figure 1, the representation of point sets has a limited model capability of instance-level information. Moreover, predicting directly from a single frame input is challenging in complex scenarios, due to temporal inconsistency between frames. To address these issues, we propose InteractionMap, which consists of Relation Embedding Module (REM), Temporal Fusion Module (TFM), + +and Geometry-aware Alignment Module (GAM). REM establishes position relation representation among feature embeddings at point level and instance level, improving the precision of map construction by explicit geometric priors. TFM facilitates long-range temporal associations by keyframe-based streaming strategy, a hierarchical temporal fusion from local to global. It enhances the stability of map elements among different frames in some complex scenarios, such as map element occlusion caused by moving vehicles. GAM solves the misalignment problem caused by inconsistency of the predictions between the classification score and the instance points precision, e.g. a prediction with a high classification score and relatively low localization quality (large Chamfer distance). + +Our main contributions can be summarized as follows: + +- We introduce an explicit position relation embedding method from point to instance, efficiently employing progressive interaction of point-wise information and instance-wise information. +- We propose a novel key-frame-based hierarchical streaming strategy, which fully leverages long-range temporal information. +- We present a geometry-aware classification loss and a geometry-aware matching cost to overcome the misalignment problem of classification and localization output. + +# 2. Related Work + +# 2.1. HD Map Construction + +With the development of view transformation from perspective-view (PV) to bird-eye-view (BEV) methods [18, 33], HD map construction is formulated as a segmentation task [4, 13, 22, 30, 31, 36, 38, 53, 55] based on sensor observations on board. HDMapNet [19] builds a vectorized HD map using semantic segmentation, clustering, and post-processing. VectorMapNet [8] is the first end-to-end framework that utilizes transformers [45], in a two-stage coarse-to-fine manner. However, the auto-regressive model of VectorMapNet leads to a long training schedule. MapTR [24] adopts a one-stage transformer approach based on [2, 22, 58] with a permutation-equivalent point set modeling. The evolved version MapTRv2 [25] adds auxiliary headers, decoupled self-attention in the decoder and one-to-many matching strategy, leading to a large improvement. Instead of a point set representation, BeMapNet [39] utilizes an instance-level representation with a piecewise Bezier head. PivotNet [7] converts point-level representation to instance-level representation using the point-to-line mask module. MapQR [32] implicitly encodes point-level queries within instance-level queries and embeds query positions like Conditional DETR [34] and DAB-DETR [27]. MapVR [50] generates a vectorized map with differentiable rasterization that provides instance-level segmentation su + +pervision. MMap [28] employs mask-guided features to refine an instance representation with enhanced feature detail. In contrast to the above methods, we introduce more information interaction to improve the reliability and stability of map construction in complex scenarios. + +# 2.2. Online Temporal Fusion + +The temporal fusion strategy is effective for the 3D object detection task, such as BEVFormer [22], SOLOFusion [37], StreamPETR [46], VideoBEV [10] and Sparse4D v2 [26]. HDMapNet [19] employs max pooling to fuse temporal information directly into the BEV feature map. StreamMapNet [49] proposes query propagation and BEV feature map fusion. SQD-MapNet [47] designs a stream query denoising approach. MapTracker [3] enhances query propagation and BEV feature map fusion using a tracking strategy. The streaming strategy facilitates longer temporal association as the propagated hidden states encode all historical information. However, a temporal encoder such as the gated recurrent unit (GRU) [5] may still face the problem of forgetting due to limited capacity in complex outdoor environments. For example, occlusion of moving vehicles may lead to temporal fusion of polluted BEV feature map. The stacking strategy may integrate features from specific previous frames, offering flexibility in fusion of long-range information. In this paper, we propose a key-frame-based streaming strategy, leveraging local and global fusion capabilities. + +# 2.3. Map Element Interaction + +ADMap [15] explores point-order relationships between and within instances through a cascading approach. InsightMapper [48] performs inner-instance feature aggregation by an additional masked inner-instance self-attention module. GeMap [54] designs a masked decoupled self-attention to handle the shapes and relations of the instance independently. HoMap [1] introduces high-order modeling to capture the correlations between instances using high-order statistics. HIMap [57] proposed a hybrid framework to learn and interact with information at the point-level and the element-level. As proved in [14], modeling relations between objects is beneficial to object recognition. In state-of-the-art methods [9, 12], relationship of queries improves DETR-like models. In this paper, we explore the interaction of map elements between their geometry and appearance feature by our explicit position relation embedding. + +# 2.4. Classification-Localization Alignment + +The classification head and the localization head of object detection are implemented separately by two branches in parallel, causing inconsistency in the output distribution. This misalignment problem has been well researched in the 2D object detection field [21, 51]. However, it is overlooked by the map element detectors [7, 24, 25, 32, 39, 50]. In this + +![](images/157b04680f848abca5e415492a98a0afab4fa9aadcdee2dffaf63568299f7318.jpg) +Figure 2. Overview of InteractionMap framework. InteractionMap mainly consists of four components: (1) BEV encoder transforms sensor input to a unified BEV representation; (2) Key-frame-based temporal fusion module leverage temporal information from local to global; (3) Relation map decoder utilizes relation embedding in point-level and instance-level; (4) Geometry-aware alignment module is designed to solve the misalignment problem of classification and position output. + +paper, we create a geometry-aware classification loss, making use of focal loss [42] with a geometry-aware target in foreground candidates. Moreover, we present a geometry-aware matching cost to suppress the matching of background candidates. + +# 3. Method + +Figure 2 illustrates the framework of our proposed InteractionMap. Firstly, BEV features are extracted from multiview images by the BEV encoder, which consists of a backbone [11] to extract multi-scale image features, and a view transformation module [22, 31] to encode image features into a BEV representation $\mathcal{F}_{local}$ . Secondly, key-frame-based temporal fusion leverages long-range temporal information by local fusion and global fusion (Section 3.1). Subsequently, the relation map decoder utilizes point-wise relation embedding (PRE) and instance-wise relation embedding (IRE) for better interaction between queries (Section 3.2). Finally, by introducing a geometric-aware classification loss and a geometric-aware matching cost, the geometry-aware alignment leverages interaction between semantic information and geometry information (Section 3.3). + +# 3.1. Key-frame-based Temporal Fusion + +Compared to the single-frame map element prediction method, there are mainly two types of temporal fusion methods, namely stacking-based strategy and streaming-based strategy. The streaming strategy leverages all historical information by hidden state encoding. However, due to limited memory capacity, the streaming strategy has limited performance in long-range perception. The stacking strategy fuse features from specific previous frames, offering flexibility in integration of long-range information. The computational cost is linearly related to the number of fused frames. + +We propose a key-frame-based streaming (KFS) strategy to address these problems. The well-known key-frame-based strategy is widely used in robot navigation and mapping, such as sparse mapping [16, 35] and dense map + +ping [6, 56]. Online reconstruction methods are unstable in long range, but they are accurate in the local regime. The input frames are divided into local frame segments. The submaps are reconstructed by integration of local frames, and then the global map is reconstructed by integration of submaps. We introduce two types of KFS strategy, namely KFS-streaming and KFS-stacking, as shown in Figure 3. Firstly, the BEV feature embedding submaps are generated by a streaming-base temporal encoder at the bottom level, which keeps the stability and temporal consistency of predictions cross latest frames. Secondly, the long-range BEV feature embedding is generated by a streaming-based or stacking-based temporal encoder at the top level. Figure 3 illustrates the KFS strategy that merges one previous submap with an interval of two. + +# 3.1.1 Local BEV Fusion + +We warp the memorized BEV feature map from the previous frame to the current frame recurrently based on the ego vehicle's relative pose. Then we employ a GRU [5] to fuse current BEV feature map $\mathcal{F}_{local}^{t}$ and the warped memorized BEV feature map $\tilde{\mathcal{F}}_{submap}^{t-1}$ into a single BEV feature embedding submap candidate $\mathcal{F}_{submap}^{t}$ . + +$$ +\tilde {\mathcal {F}} _ {s u b m a p} ^ {t - 1} = \operatorname {W a r p} \left(\mathcal {F} _ {s u b m a p} ^ {t - 1}, \boldsymbol {T}\right) \tag {1} +$$ + +$$ +\mathcal {F} _ {\text {s u b m a p}} ^ {t} = \operatorname {R e s B l o c k} \left(L N \left(G R U \left(\tilde {\mathcal {F}} _ {\text {s u b m a p}} ^ {t - 1}, \mathcal {F} _ {\text {l o c a l}} ^ {t}\right)\right)\right) \tag {2} +$$ + +Where ResBlock is a residual-based convolution block with the shortcut connection [11]. $LN$ is a layer normalization operation. $\pmb{T}$ denotes a standard $4\times 4$ transformation matrix between the coordinate systems of two frames. + +# 3.1.2 Global BEV Fusion + +Our KFS strategy selects a BEV feature embedding as a submap using a specific distance stride $d_{stride}$ , instead of a specific temporal interval customly. + +KFS-streaming strategy The previous selected global BEV feature is warped to the current frame. The warped + +![](images/f5b1cced7728e9d0d6325b09a9c6bd3dfc432cfc26f1cdb784c4e22c1f8049a0.jpg) +Figure 3. Temporal fusion strategy. + +global BEV feature and the current BEV feature map are then fused to create a global BEV feature map. + +$$ +\tilde {\mathcal {F}} _ {\text {g l o b a l}} ^ {\text {p r e}} = \operatorname {W a r p} \left(\mathcal {F} _ {\text {g l o b a l}} ^ {\text {p r e}}, \boldsymbol {T}\right) \tag {3} +$$ + +$$ +\mathcal {F} _ {g l o b a l} ^ {t} = L N \left(G R U \left(\tilde {\mathcal {F}} _ {g l o b a l} ^ {p r e}, \mathcal {F} _ {s u b m a p} ^ {t}\right)\right) \tag {4} +$$ + +KFS-stacking strategy We warp $N_{pre}$ selected submaps to the current frame. Following concatenation and residual-based convolution [11], we obtain the global BEV feature. + +$$ +\tilde {\mathcal {F}} _ {s u b m a p} ^ {t _ {k}} = \operatorname {W a r p} \left(\mathcal {F} _ {s u b m a p} ^ {t _ {k}}, T\right) \tag {5} +$$ + +$$ +\mathcal {F} _ {\text {g l o b a l}} ^ {t} = \operatorname {R e s B l o c k} \left(\operatorname {C o n c a t} \left(\tilde {F} _ {\text {s u b m a p}}, \mathcal {F} _ {\text {s u b m a p}} ^ {t}\right)\right) \tag {6} +$$ + +Here, $\tilde{\mathcal{F}}_{submap}^{t_k}$ and $t_k$ represent the selected BEV feature submap and its corresponding index. The collection of selected BEV feature submaps is denoted as $\tilde{F}_{submap} = \{\tilde{\mathcal{F}}_{submap}^{t_k}\}_{k=1}^{N_{pre}}$ . During the training period, the previous $N_{pre}$ submaps are randomly selected within a range of $N_{pre} \times d_{stride}$ . + +# 3.2. Relation map decoder + +Previous work has shown the effectiveness of interactions for map element detectors [15, 57]. In contrast to these approaches, we directly construct explicit relation priors by incorporating relation embedding into the self-attention module of the decoder. + +# 3.2.1 Relation Embedding + +Here is the self-attention in the decoder of a DETR-like framework. + +$$ +Q _ {\text {s e l f - a t t n}} = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) V \tag {7} +$$ + +Here $d_k$ is the dimension of the keys. $Q, K, V$ are feature embeddings generated by linear projections. $Q = \operatorname{Linear}_q(Q_{in})$ . $K = \operatorname{Linear}_k(Q_{in})$ . $V = \operatorname{Linear}_v(Q_{in})$ + +We incorporate relation embeddings into the self-attention model, inspired by [14]. + +$$ +Q _ {s e l f - a t t n} = \operatorname {S o f t m a x} \left(\mathcal {F} _ {r e l} + \frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) V \tag {8} +$$ + +Here $\mathcal{F}_{rel}$ represents the relation embedding. In this paper, point-wise relation embedding and instance-wise relation embedding are added to the decoupled self-attention model at the point level and the instance level, respectively, employing attention refinement of information interaction. + +# 3.2.2 Point-wise Relation Embedding + +The point-wise relation encoder represents the high-dimensional point relation embedding as an explicit geometry prior. The point-wise relation embedding is generated by normalized point coordinates and edge directions. + +$$ +\operatorname {R e l} _ {p t} (i, j) = \left[ \log \left(x _ {i} - x _ {j} + 1\right), \log \left(y _ {i} - y _ {j} + 1\right) \right] \tag {9} +$$ + +$$ +\operatorname {R e l} _ {\text {d i r}} (i, j) = \operatorname {c o s i n e \_ s i m} \left(e _ {i}, e _ {i + 1}\right) - \operatorname {c o s i n e \_ s i m} \left(e _ {j}, e _ {j + 1}\right) \tag {10} +$$ + +Here $x_{i}$ , $x_{j}$ are normalized point coordinates. $e_{i}$ and $e_{i+1}$ , $e_{j}$ and $e_{j+1}$ are adjacent edges. The cosine similarity is denoted by $\cosin\_sim$ . The point-wise relation embedding is unbiased, since both $Rel_{pt}(i,j)$ and $Rel_{dir}(i,j)$ are equal to 0 when $i = j$ . + +The point-wise relation embedding $\mathcal{F}_{rel}^{pt}$ is illustrated by the following equation: + +$$ +\mathcal {F} _ {r e l} ^ {p t} = \operatorname {R e L U} (\operatorname {L i n e a r} (S P E (\operatorname {C o n c a t} (R e l _ {p t}, R e l _ {d i r})))) \tag {11} +$$ + +Firstly, the point-wise relation embedding is calculated by concatenation of $Rel_{pt}$ and $Rel_{dir}$ . Then the embedding is transformed to high dimension by sine positional encoding (SPE). Lastly, the embedding is transformed by a linear projection and a ReLU function. + +# 3.2.3 Instance-wise Relation Embedding + +The instance-wise relation embedding is generated by classification scores and chamfer distance of points. + +$$ +\operatorname {R e l} _ {s} (i, j) = \operatorname {s g n} \left(s _ {i} - s _ {j}\right) \tag {12} +$$ + +$$ +\operatorname {R e l} _ {s d} (i, j) = \operatorname {s g n} \left(s _ {i} - s _ {j}\right) \cdot \operatorname {C h a m f e r} \left(I _ {i}, I _ {j}\right) \tag {13} +$$ + +Here, $sgn$ is the sign function. $s_i, s_j$ are classification scores of the $i^{th}$ and $j^{th}$ candidates. Chamfer $(I_i, I_j)$ is the chamfer distance between points of instance $i$ and instance $j$ . The equation of $Rel_s(i,j)$ establishes a simple classification ranking relation among pairs of instances. Then we add the position relation of the map elements to the instance-wise relation embedding $\mathcal{F}_{rel}^{ins}$ by the chamfer distance. + +$$ +\mathcal {F} _ {r e l} ^ {i n s} = R e L U \left(L i n e a r \left(S P E \left(R e l _ {s d}\right)\right)\right) \tag {14} +$$ + +The instance-wise relation embedding is unbiased, since $Rel_{sd}(i,j)$ is equal to 0 when $i = j$ . + +# 3.3. Geometry-aware Alignment + +Current work, such as MapTR [24] uses the classification branch and the regression branch to predict semantic labels and geometric localization separately. MapTR uses the focal loss [42] as the classification loss and the point-to-point L1 loss as the regression loss. Since classification and regression losses are independent, the quality of these two branches is inconsistent, leading to the well-known misalignment problem. In this paper, we introduce the geometry-aware focal loss and the geometry-aware focal cost to overcome this problem. + +# 3.3.1 Geometry-aware Focal Loss + +Inspired by IoU-aware Focal Loss [21, 51] in traditional object detection, we design the novel Geometry-aware Focal Loss (GFL). Unlike focal loss, we treat positives and negatives asymmetrically. + +$$ +\mathcal {L} _ {G F L} = \sum_ {i = 1} ^ {N _ {\text {p o s}}} s _ {\text {g e o} _ {i}} B C E \left(s _ {\text {g e o} _ {i}}, p _ {i}\right) + \sum_ {j = 1} ^ {N _ {\text {n e g}}} \alpha p _ {j} ^ {\gamma} B C E \left(p _ {j}, 0\right) \tag {15} +$$ + +where $p_i$ and $p_j$ are the predicted probability of the $i^{th}$ and $j^{th}$ candidates, $s_{geo} \in [0,1]$ is an instance Geometry-aware Classification Score (GCS). Here we define three types of GCS, including $s_{p2p}$ , $s_{dir}$ and $s_{giou}$ . + +$$ +s _ {p 2 p} = 1 - \frac {1}{2 N _ {p}} \sum_ {i = 1} ^ {N _ {p}} D _ {\text {M a n h a t t a n}} \left(p _ {\text {s r c} _ {i}}, p _ {\text {t g t} _ {i}}\right) \tag {16} +$$ + +$$ +s _ {d i r} = 0. 5 + \frac {1}{2 N _ {e}} \sum_ {i = 1} ^ {N _ {e}} \operatorname {c o s i n e \_ s i m} \left(e _ {s r c _ {i}}, e _ {t g t _ {i}}\right) \tag {17} +$$ + +$$ +s _ {g i o u} = 0. 5 + 0. 5 \cdot G I o U \left(B _ {s r c}, B _ {t g t}\right) \tag {18} +$$ + +where $s_{p2p} \in [0,1]$ is a normalized point-to-point L1 score, $s_{dir} \in [0,1]$ is an edge direction score calculated by cosine similarity and $s_{giou} \in [0,1]$ is a normalized GIoU [40] score of candidate bounding box $B_{src}$ and assigned GT $B_{tgt}$ . The bounding box is computed by minimum enclosing box of instance points. $N_p$ and $N_e$ are the numbers of points and edges. $p_{src}$ and $p_{tgt}$ are normalized points of candidates and assigned GT. $e_{src}$ and $e_{tgt}$ are edges of candidates and assigned GT. + +# 3.3.2 Geometry-aware Focal Cost + +The loss function $\mathcal{L}_{GFL}$ supervises the classification scores with geometric metrics GCS $s_{geo_i}$ . And $s_{geo_i}$ accurately + +selects these high-quality candidates, which leads to better map vectorization performance. Following the spirit of GFL, we make similar modifications to the focal cost, which is a key component of label assignment. The novel matching cost, geometry-aware focal cost (GFC) is as follows. + +$$ +\mathcal {C} _ {G F C} (i, j) = s _ {g e o _ {i}} B C E \left(s _ {g e o _ {i}}, p _ {i}\right) - \alpha p _ {i} ^ {\gamma} B C E \left(1 - p _ {i}, 1\right) \tag {19} +$$ + +where $\mathcal{C}_{GFC}(i,j)$ is the geometry-aware focal cost for the $i^{th}$ prediction and the $j^{th}$ ground truth. The GFC is used as a modulated function to suppress candidates with inaccurate prediction localization and promote candidates with an accurate position. + +# 3.4. Training Loss + +Auxiliary BEV Supervision. We introduce a query-based instance segmentation (QIS) module using an instance mask prediction branch. Inspired by [17], we make use of the query embeddings $Q_{dec}$ from Transformer to dot-product the BEV feature embedding map $\mathcal{F}_{bev}$ to obtain instance-wise binary masks $m$ with the sigmoid function $\sigma(\cdot)$ . + +$$ +\hat {M} _ {\text {i n s}} = \sigma (M L P \left(Q _ {\text {d e c}}\right) \cdot \mathcal {F} _ {\text {b e v}}) \tag {20} +$$ + +For the instance segmentation branch, we extend the mask focal loss and the mask focal cost to mask geometry-aware focal loss (MGFL) and mask geometry-aware focal cost (MGFC) as well, with the same ideal of GFL. + +$$ +\mathcal {L} _ {\text {s e g}} = \lambda_ {m g f} \mathcal {L} _ {m g f} (\hat {M} _ {\text {i n s}}, M _ {\text {i n s}}) + \lambda_ {d i c e} \mathcal {L} _ {d i c e} (\hat {M} _ {\text {i n s}}, M _ {\text {i n s}}) \tag {21} +$$ + +Instance masks $\hat{M}_{ins}$ are supervised by instance mask annotations $M_{ins}$ , using the mask geometry-aware focal loss $\mathcal{L}_{mgf}$ and the dice loss $\mathcal{L}_{dice}$ . + +Overall Loss. The total loss $L$ used by InteractionMap is as follow: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {d e t}} + \mathcal {L} _ {\text {s e g}} + \mathcal {L} _ {\text {a u x}} \tag {22} +$$ + +where $\mathcal{L}_{det}$ is the loss term for the detection task. $\mathcal{L}_{seg}$ is the loss term for the segmentation task. $\mathcal{L}_{aux}$ is the auxiliary loss term. + +The detection loss consists of classification loss $\mathcal{L}_{cls}$ , point-to-point distance loss $\mathcal{L}_{p2p}$ , edge direction loss $\mathcal{L}_{dir}$ . + +$$ +\mathcal {L} _ {\text {d e t}} = \lambda_ {\text {c l s}} \mathcal {L} _ {\text {c l s}} + \lambda_ {p 2 p} \mathcal {L} _ {p 2 p} + \lambda_ {\text {d i r}} \mathcal {L} _ {\text {d i r}} \tag {23} +$$ + +The loss term $\mathcal{L}_{cls}$ is the geometry-aware focal loss for the classification of instances. The loss term $\mathcal{L}_{p2p}$ is the smooth L1 loss for the regression of vectorized map elements. And the loss term $\mathcal{L}_{dir}$ is the loss of cosine similarity between adjacent edges. + +$$ +\mathcal {L} _ {\text {a u x}} = \lambda_ {\text {d e p t h}} \mathcal {L} _ {\text {d e p t h}} + \lambda_ {P V S e g} \mathcal {L} _ {P V S e g} \tag {24} +$$ + +The auxiliary loss term $\mathcal{L}_{aux}$ is composed of the depth-map prediction loss $\mathcal{L}_{depth}$ and image semantic segmentation loss $\mathcal{L}_{PVSeg}$ . + +# 4. Experiment + +# 4.1. Experimental Settings + +# 4.1.1 Datasets + +nuScenes Datasets. There are 1000 scenes of 4 locations in nuScenes datasets. Each scene contains 20s of RGB images from six cameras, point clouds from LiDAR sweeps, and a 3D vectorized map. The train set contains 700 scenes with 28130 samples and validation set contains 150 scenes with 6019 samples, respectively. Following previous works [24, 25], we mainly focus on three categories of map elements, including lane divider, pedestrian crossing and road boundary. + +Argoverse2 Datasets. Argoverse2 dataset provide 1000 scenes in 6 cities, each lasting around 15 seconds. Each log includes 7 RGB images from surrounding cameras and point cloud from LiDAR sweeps. Follow previous methods [24, 25], 700 scenes are used for training, and 150 scenes are used for validation. For a fair comparison, we report results on its validation set and focus on the same three map categories as the nuScene dataset. + +# 4.1.2 Evaluation Metrics + +For a fair comparison, we follow the standard metric used in previous works [19, 24, 25, 29]. With the ego-coordination as the center, the perception ranges are $[-15m, 15m]$ for the X-axis and $[-30m, 30m]$ for the Y-axis. We adopt the mean Average Precision (mAP) to evaluate the map vectorization quality based on the chamfer distance. A candidate is considered as True-Positive (TP) only if the chamfer distance to ground truth is less than certain thresholds $(\tau \in T, T = 0.5, 1.0, 1.5)$ . The final AP metric is calculated by average across all thresholds and all classes. + +# 4.1.3 Implementation Details + +Following previous method [24, 25], we use ResNet50 [11] as the backbone for RGB images. All models are trained with 8 A800 GPUs. The default batch size is 32. The initial learning rate is set to $6 \times 10^{-4}$ with cosine decay. The nuScenes images resolution is $1600 \times 900$ . We resize the images with 0.5 ratio. For the Argoverse2 dataset, the image resolutions are $2048 \times 1550$ and $1550 \times 2048$ . We pad the images to $2048 \times 2048$ , and then resize the images with 0.3 ratio. We define the size of BEV feature as $H \times W$ of $200 \times 100$ , and the size of each BEV grid as 0.3 meters. The default numbers of instance queries, point queries, and decoder layers are 100, 20, 6, respectively. The model is trained on 24, 110 epochs on the nuScenes and Argoverse2 datasets. We randomly divide each training sequence into 3 splits to generate more diverse data sequences. Inspired by SOLOFusion [37], we train the ini + +tial $N_{init}$ epochs with single-frame input to stabilize multi-frame straining. $N_{init} = 0.5N_{total}$ . For loss weights, $\lambda_{cls} = 2$ , $\lambda_{p2p} = 4$ , $\lambda_{dir} = 0.005$ , $\lambda_{mgf} = 30$ , $\lambda_{dice} = 3$ , $\lambda_{depth} = 3$ , $\lambda_{PVSeg} = 2$ . For temporal fusion, $N_{pre} = 4$ , $d_{stride} = 5m$ . + +# 4.2. Performance Comparison + +Performance on nuScenes Dataset. As shown in Table 1, our methods show consistent improvements over baselines. By integrating our method with MapTRv2 [25], we achieve $71.6(+10.1)$ , $71.8(+10.3)$ mAP with 24-epoch training schedules and $75.5(+6.8)$ , $76.0(+7.3)$ mAP with 110-epoch training schedules, respectively. InteractionMap establishes a new state-of-the-art performance, exhibiting significant improvements over existing methods with comparable inference speed, as shown in Table 4. + +Performance on Argoverse2 Dataset. Argoverse2 dataset contains 3D map elements, including height information. For fair comparison, the sampling frequency is set to $2\mathrm{Hz}$ and $10\mathrm{Hz}$ , respectively. The amount of training data at a $10\mathrm{Hz}$ sampling frequency is 5 times as many as that at a $2\mathrm{Hz}$ sampling frequency. As demonstrated in Table 2 and Table 3, InteractionMap surpasses all state-of-the-art methods in both 2D and 3D vectorized map perception on Argoverse2 dataset. + +# 4.3. Ablation Study + +We examine the efficacy of each component of InteractionMap through ablation studies, utilizing the nuScenes dataset. Initially, we build a simple baseline base on MapTR [24] with the LSS encoder without auxiliary depth-map supervision. The default numbers of instances, points per instance, are 50 and 20, respectively. + +Contributions of Main Components. Table 5 demonstrates the impact of each component of InteractionMap. When adding relation embedding into the baseline model, there is a substantial improvement of $4.5\mathrm{mAP}$ . Furthermore, the addition of geometry alignment results in a significant 7.5 increase in mAP, emphasizing its effectiveness in boosting performance. Lastly, key-frame-based temporal fusion offers 5.0 and 5.1 additional increases in mAP, indicating that the inclusion of long-range hierarchical streaming strategy significantly enhances the performance. + +Ablation on Relation Embedding. In Table 6, we explore the effect of relation embedding in point-level and instance-level. Adding the point-wise relation embedding of point coordinates and edge direction results in 2.1 and 0.9 improvements in mAP over the baseline, respectively. The instance-wise relation embedding leverages the classification confidence and points coordinates, leading to enhancements of 1.1 and $3.1\mathrm{mAP}$ , respectively. The combination of relation embedding in point-level and instance-level enhances performance by $2.8\mathrm{mAP}$ . Finally, we introduce a + +
MethodBackboneEpochsTemporal\(AP_{ped}\)\(AP_{div}\)\(AP_{bou}\)mAP
MapTRR502446.351.553.150.3
MapVRR502447.754.451.451.2
PivotNetR503053.855.859.657.4
BeMapNetR503057.762.359.459.8
MapTRv2R502459.862.462.461.5
StreamMapNetR5030✓61.766.362.163.4
MGMapR502461.865.067.564.8
SQD-MapNetR5024✓63.666.664.865.0
MapQRR502463.468.067.766.4
HIMapR503062.668.469.166.7
HRMapNet-MapTRv2R5024✓65.867.468.567.2
InteractionMap-RR5024✓71.370.872.871.6
InteractionMap-CR5024✓69.772.773.071.8
VectorMapNetR50110+ft42.551.444.146.0
MapTRR5011056.259.860.158.7
MapVRR5011055.061.859.458.8
BeMapNetR5011062.666.765.164.8
MapTRv2R5011068.168.369.768.7
MapQRR5011070.174.473.272.6
HRMapNet-MapTRv2R50110✓72.072.975.873.6
HIMapR5011071.375.074.773.7
InteractionMap-RR50110✓75.275.076.175.5
InteractionMap-CR50110✓75.375.677.076.0
+ +Table 1. Comparison with SOTA methods on the nuScenes validation set at $60m\times 30m$ perception range. "EB0", "EB4", "R50" correspond to the backbones Efficient-B0, Efficient-B4 [44], ResNet50 [11]. "ft" means the two-stage fine-tune strategy. "C" means KFS-stacking strategy and "R" means KFS-streaming strategy. + +
MethodDimAPpedAPdivAPboumAP
StreamMapNet2d62.059.563.061.5
SQD-MapNet2d64.960.264.963.3
MapTRv22d60.068.764.264.3
HRMapNet-MapTRv22d65.171.468.668.3
InteractionMap-C2d68.577.974.673.7
InteractionMap-R2d70.177.074.073.7
+ +alternating training strategy to randomly turn off one of these two components, leading the overall design achieves $54.7(+4.5)$ mAP. + +Ablation on Key-frame-based Temporal Fusion. To investigate the effectiveness of the key-frame-based streaming strategy, we conduct ablation experiments to compare it with the baseline (without temporal fusion) and the streaming strategy, in Table 7. Employing the streaming strategy results in $3.8\mathrm{mAP}$ over the baseline. Furthermore, when the KFS-stacking strategy is introduced, the mAP increased by 7.4, reaching 57.6. Additionally, the inclusion of KFS-streaming strategy leads to an improvement of $7.2\mathrm{mAP}$ . + +Ablation on Geometry-aware Alignment. As shown in + +Table 2. Comparison with SOTA methods on the Argoverse2 validation set at $60m \times 30m$ perception range with a 30-epoch training schedule, at a 2Hz sampling frequency. + +
MethodDimAPpedAPdivAPboumAP
MapTRv22d63.671.567.467.5
MapQR2d64.372.368.168.2
HIMap2d69.069.570.369.6
InteractionMap-C2d69.878.674.874.4
InteractionMap-R2d70.878.174.974.6
MapTRv23d60.768.964.564.7
MapQR3d60.171.266.265.9
HIMap3d66.768.370.368.4
InteractionMap-C3d66.675.672.771.6
InteractionMap-R3d67.775.573.172.1
+ +Table 3. Comparison with SOTA methods on the Argoverse2 validation set at $60m \times 30m$ perception range with a 6-epoch training schedule, at a $10\mathrm{Hz}$ sampling frequency. + +
MethodFPSParams(MB)mAP
MapTRv214.940.661.5
InteractionMap-C11.0116.771.8
InteractionMap-R12.149.971.6
+ +Table 4. Comparison with SOTA methods on the nuScenes validation set. FPS is measured on a single A800 GPU. + +
REMGAMTFMAPpedAPdivAPboumAP
46.750.553.550.2
✓50.156.657.354.7
✓57.163.863.761.5
C50.960.061.957.6
R50.661.360.357.4
✓✓58.665.162.962.2
✓✓C64.268.668.867.2
✓✓R64.168.169.867.3
+ +Table 5. Ablation study of each component, including Relation Embedding Module (REM), Temporal Fusion Module (TFM), Geometry-aware Alignment Module (GAM). "C" means KFS-stacking strategy and "R" means KFS-streaming strategy. + +
PtxyPtdirInsSInssdPPAPpedAPdivAPboumAP
46.750.553.550.2
✓47.852.756.352.3
✓47.052.453.951.1
✓✓47.855.656.453.3
✓46.952.154.851.3
✓47.955.756.453.3
✓✓✓47.655.955.553.0
✓✓✓✓50.156.657.354.7
+ +Table 6. Ablation study of relation embedding by employing point-wise coordinate embedding $\mathrm{Pt}_{xy}$ , point-wise edge direction embedding $\mathrm{Pt}_{dir}$ , instance-wise classification score embedding $\mathrm{Ins}_s$ , instance-wise classifier score and chamfer distance embedding $\mathrm{Ins}_{sd}$ and alternating training strategy of relation embedding, namely Ping-Pong (PP). + +
MethodAPpedAPdivAPboumAP
Baseline46.750.553.550.2
Streaming48.654.758.854.0
KFS-stacking50.960.061.957.6
KFS-streaming50.661.360.357.4
+ +Table 8, geometry-aware alignment comprised five main components: query-based instance segmentation (QIS), geometry-aware focal loss (GFL), geometry-aware focal cost (GFC), mask geometry-aware focal loss (MGFL) and mask geometry-aware focal cost (MGFC). The QIS leverages geometry priors by instance segmentation mask supervision. Compared to the baseline, the QIS leads to an improvement of $6.4\mathrm{mAP}$ . The GFL encourages predictions with high localization metrics to obtain a better classification score, with an improvement of $3.1\mathrm{mAP}$ compared to mask focal loss. The GFC suppresses the matching cost of candidates with an inaccurate prediction position, which brings an improvement of $0.9\mathrm{mAP}$ . For the segmentation + +Table 7. Ablation study of streaming strategy, KFS-streaming strategy and KFS-stacking strategy. + +
QISGFLGFCMGFLMGFCAPpedAPdivAPboumAP
46.750.553.550.2
✓51.758.759.456.6
✓✓55.162.062.059.7
✓✓✓55.863.862.260.6
✓✓✓✓57.063.662.361.0
✓✓✓✓✓57.163.863.761.5
+ +Table 8. Ablation study of geometry-aware alignment by investigating the performance of query-based instance segmentation (QIS), geometry-aware focal loss (GFL), geometry-aware focal cost (GFC), mask geometry-aware focal loss (MGFL) and mask geometry-aware focal cost (MGFC). + +
SgiouSp2pSdirAPpedAPdivAPboumAP
✓51.758.759.456.6
52.656.258.055.6
✓54.860.263.159.4
✓58.363.162.561.3
✓✓57.163.863.761.5
+ +Table 9. Ablation study of geometry-aware classification score. + +branch, the result shows that the position-supervised segmentation loss MGFL and the position-modulated segmentation cost MGFC improve the final results, with $0.4\mathrm{mAP}$ and $0.5\mathrm{mAP}$ gains, individually. + +We further investigate the geometry-aware classification score $S_{geo}$ . As depicted in Table 9, using $S_{giou}$ , $S_{p2p}$ and $S_{dir}$ leads to individual improvement of -1.0, 2.8 and 4.7 mAP, respectively. Furthermore, the combination of $S_{p2p}$ and $S_{dir}$ boosts the performance, improving by 4.9 mAP. + +# 5. Conclusion + +In this paper, we introduce InteractionMap, a novel approach to end-to-end online vectorized HD map utilizing temporal and spatial information interaction from local to global. By leveraging relation embedding at both point level and instance level, the network learns the geometric priors of map elements better. Subsequently, key-frame-based streaming strategy enhances the network by utilizing temporal feature embedding from near to distant. Finally, the geometric-aware focal loss and the geometric-aware focal cost build a strong correlation between classification score and instance-point location precision. Across various experimental settings, our proposed InteractionMap yields significant performance improvements over the baseline on various datasets. + +# References + +[1] Yingfeng Cai, Wei Dong, Ze Liu, Hai Wang, and Long Chen. Homap: End-to-end vectorized hd map construction with + +high-order modeling. IEEE Transactions on Intelligent Vehicles, 2024. 2 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2 +[3] Jiacheng Chen, Yuefan Wu, Jiaqi Tan, Hang Ma, and Yasutaka Furukawa. Maptracker: Tracking with strided memory fusion for consistent vector hd mapping. In European Conference on Computer Vision, pages 90-107. Springer, 2025. 2 +[4] Shaoyu Chen, Tianheng Cheng, Xinggang Wang, Wenming Meng, Qian Zhang, and Wenyu Liu. Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. arXiv preprint arXiv:2206.04584, 2022. 1, 2 +[5] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 2, 3 +[6] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4): 1, 2017. 3 +[7] Wenjie Ding, Limeng Qiao, Xi Qiu, and Chi Zhang. Pivot-net: Vectorized pivot learning for end-to-end hd map construction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3672-3682, 2023. 2 +[8] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525-11533, 2020. 1, 2 +[9] Yulu Gao, Yifan Sun, Xudong Ding, Chuyang Zhao, and Si Liu. Ease-detr: Easing the competition among object queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17282-17291, 2024. 2 +[10] Chunrui Han, Jinrong Yang, Jianjian Sun, Zheng Ge, Runpei Dong, Hongyu Zhou, Weixin Mao, Yuang Peng, and Xi-angyu Zhang. Exploring recurrent long-term temporal fusion for multi-view 3d perception. IEEE Robotics and Automation Letters, 2024. 2 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 4, 6, 7 +[12] Xiuquan Hou, Meiqin Liu, Senlin Zhang, Ping Wei, Badong Chen, and Xuguang Lan. Relation detr: Exploring explicit position relation prior for object detection. arXiv preprint arXiv:2407.11699, 2024. 2 +[13] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings + +of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 1, 2 +[14] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3588-3597, 2018. 2, 4 +[15] Haotian Hu, Fanyi Wang, Yaonong Wang, Laifeng Hu, Jingwei Xu, and Zhiwang Zhang. Admap: Anti-disturbance framework for reconstructing online vectorized hd map. arXiv preprint arXiv:2401.13172, 2024. 2, 4 +[16] Georg Klein and David Murray. Parallel tracking and mapping for small ar workspaces. In 2007 6th IEEE and ACM international symposium on mixed and augmented reality, pages 225-234. IEEE, 2007. 3 +[17] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 5 +[18] Hongyang Li, Chonghao Sima, Jifeng Dai, Wenhai Wang, Lewei Lu, Huijie Wang, Jia Zeng, Zhiqi Li, Jiazhi Yang, Hanming Deng, et al. Delving into the devils of bird's-eye-view perception: A review, evaluation and recipe. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[19] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online hd map construction and evaluation framework. In 2022 International Conference on Robotics and Automation (ICRA), pages 4628-4634. IEEE, 2022. 1, 2, 6 +[20] Tianyu Li, Peijin Jia, Bangjun Wang, Li Chen, Kun Jiang, Junchi Yan, and Hongyang Li. Lanesegnet: Map learning with lane segment perception for autonomous driving. arXiv preprint arXiv:2312.16108, 2023. 1 +[21] Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems, 33:21002-21012, 2020. 2, 5 +[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 3 +[23] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 541-556. Springer, 2020. 1 +[24] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized hd map construction. arXiv preprint arXiv:2208.14437, 2022. 1, 2, 5, 6 +[25] Bencheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. + +Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv preprint arXiv:2308.05736, 2023.1, 2, 6 +[26] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018, 2023. 2 +[27] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2 +[28] Xiaolu Liu, Song Wang, Wentong Li, Ruizi Yang, Junbo Chen, and Jianke Zhu. Mgmap: Mask-guided learning for online vectorized hd map construction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14812-14821, 2024. 2 +[29] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized hd map learning. In International Conference on Machine Learning, pages 22352-22369. PMLR, 2023. 1, 6 +[30] Zhi Liu, Shaoyu Chen, Xiaojie Guo, Xinggang Wang, Tianheng Cheng, Hongmei Zhu, Qian Zhang, Wenyu Liu, and Yi Zhang. Vision-based uneven bev representation learning with polar rasterization and surface estimation. In Conference on Robot Learning, pages 437-446. PMLR, 2023. 2 +[31] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2, 3 +[32] Zihao Liu, Xiaoyu Zhang, Guangwei Liu, Ji Zhao, and Ningyi Xu. Leveraging enhanced queries of point sets for vectorized map construction. In European Conference on Computer Vision, pages 461-477. Springer, 2025. 2 +[33] Yuexin Ma, Tai Wang, Xuyang Bai, Huitong Yang, Yuenan Hou, Yaming Wang, Yu Qiao, Ruigang Yang, and Xinge Zhu. Vision-centric bev perception: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[34] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651-3660, 2021. 2 +[35] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 3 +[36] Cong Pan, Yonghao He, Junran Peng, Qian Zhang, Wei Sui, and Zhaoxiang Zhang. Baeformer: Bi-directional and early interaction transformers for bird's eye view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9590-9599, 2023. 2 +[37] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 2, 6 + +[38] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 1, 2 +[39] Limeng Qiao, Wenjie Ding, Xi Qiu, and Chi Zhang. End-to-end vectorized hd-map construction with piecewiseBezier curve. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13218-13228, 2023. 2 +[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5 +[41] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11138-11147, 2020. 1 +[42] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 3, 5 +[43] Tixiao Shan and Brendan Englot. Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4758–4765. IEEE, 2018. 1 +[44] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7 +[45] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2 +[46] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3621-3631, 2023. 2 +[47] Shuo Wang, Fan Jia, Yingfei Liu, Yucheng Zhao, Zehui Chen, Tiancai Wang, Chi Zhang, Xiangyu Zhang, and Feng Zhao. Stream query denoising for vectorized hd map construction. arXiv preprint arXiv:2401.09112, 2024. 2 +[48] Zhenhua Xu, Kenneth KY Wong, and Hengshuang Zhao. Insightmapper: A closer look at inner-instance information for vectorized high-definition mapping. arXiv preprint arXiv:2308.08543, 2023. 2 +[49] Tianyuan Yuan, Yicheng Liu, Yue Wang, Yilun Wang, and Hang Zhao. Streammapnet: Streaming mapping network for vectorized online hd map construction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7356-7365, 2024. 2 +[50] Gongjie Zhang, Jiahao Lin, Shuang Wu, Zhipeng Luo, Yang Xue, Shijian Lu, Zuoguan Wang, et al. Online map vec + +torization for autonomous driving: A rasterization perspective. Advances in Neural Information Processing Systems, 36, 2024. 2 +[51] Haoyang Zhang, Ying Wang, Feras Dayoub, and Niko Sunderhauf. Varifocalnet: An iou-aware dense object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8514-8523, 2021. 2, 5 +[52] Ji Zhang, Sanjiv Singh, et al. Loam: Lidar odometry and mapping in real-time. In Robotics: Science and systems, pages 1-9. Berkeley, CA, 2014. 1 +[53] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. **Reverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743**, 2022. 2 +[54] Zhixin Zhang, Yiyuan Zhang, Xiaohan Ding, Fusheng Jin, and Xiangyu Yue. Online vectorized hd map construction using geometry. arXiv preprint arXiv:2312.03341, 2023. 2 +[55] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13760-13769, 2022. 1, 2 +[56] Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. Elastic fragments for dense scene reconstruction. In Proceedings of the IEEE International Conference on Computer Vision, pages 473-480, 2013. 3 +[57] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized hd map construction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15396-15406, 2024. 2, 4 +[58] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2 \ No newline at end of file diff --git a/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/images.zip b/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8c018afb8b704812f5ebc76de0fc05b39943e15e --- /dev/null +++ b/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79524dc36c499e3a739584fc39d12eb2ad1aeac3b303fb3535c49690a213fd86 +size 633164 diff --git a/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/layout.json b/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..138ede78b7fce2e6300121b41824a7402b05d3b3 --- /dev/null +++ b/2025/InteractionMap_ Improving Online Vectorized HDMap Construction with Interaction/layout.json @@ -0,0 +1,10683 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 102, + 533, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 102, + 533, + 138 + ], + "spans": [ + { + "bbox": [ + 77, + 102, + 533, + 138 + ], + "type": "text", + "content": "InteractionMap: Improving Online Vectorized HDMap Construction with Interaction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 192, + 161, + 416, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 161, + 416, + 190 + ], + "spans": [ + { + "bbox": [ + 192, + 161, + 416, + 190 + ], + "type": "text", + "content": "Kuang Wu Chuan Yang * Zhanbin Li Langge Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 242, + 297, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 242, + 297, + 482 + ], + "spans": [ + { + "bbox": [ + 53, + 242, + 297, + 482 + ], + "type": "text", + "content": "Vectorized high-definition (HD) maps are essential for an autonomous driving system. Recently, state-of-the-art map vectorization methods are mainly based on DETR-like framework to generate HD maps in an end-to-end manner. In this paper, we propose InteractionMap, which improves previous map vectorization methods by fully leveraging local-to-global information interaction in both time and space. Firstly, we explore enhancing DETR-like detectors by explicit position relation prior from point-level to instance-level, since map elements contain strong shape priors. Secondly, we propose a key-frame-based hierarchical temporal fusion module, which interacts temporal information from local to global. Lastly, the separate classification branch and regression branch lead to the problem of misalignment in the output distribution. We interact semantic information with geometric information by introducing a novel geometric-aware classification loss in optimization and a geometric-aware matching cost in label assignment. InteractionMap achieves state-of-the-art performance on both nuScenes and Argoverse2 benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 505, + 135, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 505, + 135, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 505, + 135, + 517 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 525, + 295, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 656 + ], + "type": "text", + "content": "High-Definition (HD) maps are designed for high-precision autonomous driving, which contain instance-level vectorized representation such as lane divider, road boundaries, pedestrian crossing, etc. The rich semantic information of road topology and traffic rules is important for the navigation of autonomous driving (AD). HD maps are traditionally constructed offline using LiDAR SLAM-based methods [43, 52] with high maintenance costs, complex pipelines, and notable localization errors. In addition, manual annotation and map updating rely heavily on human labor and time demands." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 658, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 296, + 694 + ], + "type": "text", + "content": "In recent years, more research endeavors have shifted towards deep-learning-based methods that construct vectorized HD maps around the ego-vehicle at runtime with on" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 215, + 549, + 422 + ], + "blocks": [ + { + "bbox": [ + 322, + 215, + 549, + 422 + ], + "lines": [ + { + "bbox": [ + 322, + 215, + 549, + 422 + ], + "spans": [ + { + "bbox": [ + 322, + 215, + 549, + 422 + ], + "type": "image", + "image_path": "18454750ca09f64dbdff604cfcebc86fd6daf44b35bf22adc9b1b495680f809a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 430, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 464 + ], + "type": "text", + "content": "Figure 1. Visual comparison between MapTRv2 [25] and our improved results. Our method effectively eliminates error map elements, leading to better precision and stability." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 485, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 628 + ], + "type": "text", + "content": "board sensors [19, 20, 24]. With the development of bird's eye view (BEV) representation, recent approaches [4, 13, 19, 22, 38, 41, 55] treat the semantic map learning task as a semantic segmentation problem. However, rasterized maps generated by these methods lack vectorized instance-level information, which is essential to downstream tasks, such as motion forecasting and motion planning [8, 23]. To overcome the limitations of segmentation-based methods, approaches VectorMapNet [29], MapTR [24], MapTRv2 [25] predict point sets to construct end-to-end vectorized local HD maps using transformer [45] decoders inspired by DETR [2]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "As shown in Figure 1, the representation of point sets has a limited model capability of instance-level information. Moreover, predicting directly from a single frame input is challenging in complex scenarios, due to temporal inconsistency between frames. To address these issues, we propose InteractionMap, which consists of Relation Embedding Module (REM), Temporal Fusion Module (TFM)," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 214, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 214, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 214, + 713 + ], + "type": "text", + "content": "*Corresponding at: ycscience86@gmail.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17176" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": "and Geometry-aware Alignment Module (GAM). REM establishes position relation representation among feature embeddings at point level and instance level, improving the precision of map construction by explicit geometric priors. TFM facilitates long-range temporal associations by keyframe-based streaming strategy, a hierarchical temporal fusion from local to global. It enhances the stability of map elements among different frames in some complex scenarios, such as map element occlusion caused by moving vehicles. GAM solves the misalignment problem caused by inconsistency of the predictions between the classification score and the instance points precision, e.g. a prediction with a high classification score and relatively low localization quality (large Chamfer distance)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 240, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 240, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 67, + 240, + 288, + 251 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 254, + 296, + 373 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 254, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 300 + ], + "type": "text", + "content": "- We introduce an explicit position relation embedding method from point to instance, efficiently employing progressive interaction of point-wise information and instance-wise information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 301, + 295, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 301, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 295, + 336 + ], + "type": "text", + "content": "- We propose a novel key-frame-based hierarchical streaming strategy, which fully leverages long-range temporal information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 337, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 296, + 373 + ], + "type": "text", + "content": "- We present a geometry-aware classification loss and a geometry-aware matching cost to overcome the misalignment problem of classification and localization output." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 387, + 142, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 142, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 142, + 399 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 407, + 183, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 183, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 183, + 420 + ], + "type": "text", + "content": "2.1. HD Map Construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 426, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 713 + ], + "type": "text", + "content": "With the development of view transformation from perspective-view (PV) to bird-eye-view (BEV) methods [18, 33], HD map construction is formulated as a segmentation task [4, 13, 22, 30, 31, 36, 38, 53, 55] based on sensor observations on board. HDMapNet [19] builds a vectorized HD map using semantic segmentation, clustering, and post-processing. VectorMapNet [8] is the first end-to-end framework that utilizes transformers [45], in a two-stage coarse-to-fine manner. However, the auto-regressive model of VectorMapNet leads to a long training schedule. MapTR [24] adopts a one-stage transformer approach based on [2, 22, 58] with a permutation-equivalent point set modeling. The evolved version MapTRv2 [25] adds auxiliary headers, decoupled self-attention in the decoder and one-to-many matching strategy, leading to a large improvement. Instead of a point set representation, BeMapNet [39] utilizes an instance-level representation with a piecewise Bezier head. PivotNet [7] converts point-level representation to instance-level representation using the point-to-line mask module. MapQR [32] implicitly encodes point-level queries within instance-level queries and embeds query positions like Conditional DETR [34] and DAB-DETR [27]. MapVR [50] generates a vectorized map with differentiable rasterization that provides instance-level segmentation su" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "pervision. MMap [28] employs mask-guided features to refine an instance representation with enhanced feature detail. In contrast to the above methods, we introduce more information interaction to improve the reliability and stability of map construction in complex scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 140, + 451, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 451, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 451, + 152 + ], + "type": "text", + "content": "2.2. Online Temporal Fusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 158, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 397 + ], + "type": "text", + "content": "The temporal fusion strategy is effective for the 3D object detection task, such as BEVFormer [22], SOLOFusion [37], StreamPETR [46], VideoBEV [10] and Sparse4D v2 [26]. HDMapNet [19] employs max pooling to fuse temporal information directly into the BEV feature map. StreamMapNet [49] proposes query propagation and BEV feature map fusion. SQD-MapNet [47] designs a stream query denoising approach. MapTracker [3] enhances query propagation and BEV feature map fusion using a tracking strategy. The streaming strategy facilitates longer temporal association as the propagated hidden states encode all historical information. However, a temporal encoder such as the gated recurrent unit (GRU) [5] may still face the problem of forgetting due to limited capacity in complex outdoor environments. For example, occlusion of moving vehicles may lead to temporal fusion of polluted BEV feature map. The stacking strategy may integrate features from specific previous frames, offering flexibility in fusion of long-range information. In this paper, we propose a key-frame-based streaming strategy, leveraging local and global fusion capabilities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 405, + 455, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 455, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 455, + 418 + ], + "type": "text", + "content": "2.3. Map Element Interaction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 423, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 616 + ], + "type": "text", + "content": "ADMap [15] explores point-order relationships between and within instances through a cascading approach. InsightMapper [48] performs inner-instance feature aggregation by an additional masked inner-instance self-attention module. GeMap [54] designs a masked decoupled self-attention to handle the shapes and relations of the instance independently. HoMap [1] introduces high-order modeling to capture the correlations between instances using high-order statistics. HIMap [57] proposed a hybrid framework to learn and interact with information at the point-level and the element-level. As proved in [14], modeling relations between objects is beneficial to object recognition. In state-of-the-art methods [9, 12], relationship of queries improves DETR-like models. In this paper, we explore the interaction of map elements between their geometry and appearance feature by our explicit position relation embedding." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 623, + 514, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 514, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 514, + 635 + ], + "type": "text", + "content": "2.4. Classification-Localization Alignment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "The classification head and the localization head of object detection are implemented separately by two branches in parallel, causing inconsistency in the output distribution. This misalignment problem has been well researched in the 2D object detection field [21, 51]. However, it is overlooked by the map element detectors [7, 24, 25, 32, 39, 50]. In this" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17177" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 168 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 168 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 168 + ], + "type": "image", + "image_path": "157b04680f848abca5e415492a98a0afab4fa9aadcdee2dffaf63568299f7318.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 171, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 54, + 171, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 54, + 171, + 555, + 217 + ], + "type": "text", + "content": "Figure 2. Overview of InteractionMap framework. InteractionMap mainly consists of four components: (1) BEV encoder transforms sensor input to a unified BEV representation; (2) Key-frame-based temporal fusion module leverage temporal information from local to global; (3) Relation map decoder utilizes relation embedding in point-level and instance-level; (4) Geometry-aware alignment module is designed to solve the misalignment problem of classification and position output." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 230, + 294, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 230, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 54, + 230, + 294, + 291 + ], + "type": "text", + "content": "paper, we create a geometry-aware classification loss, making use of focal loss [42] with a geometry-aware target in foreground candidates. Moreover, we present a geometry-aware matching cost to suppress the matching of background candidates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 298, + 111, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 111, + 311 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 111, + 311 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 319, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 296, + 510 + ], + "type": "text", + "content": "Figure 2 illustrates the framework of our proposed InteractionMap. Firstly, BEV features are extracted from multiview images by the BEV encoder, which consists of a backbone [11] to extract multi-scale image features, and a view transformation module [22, 31] to encode image features into a BEV representation " + }, + { + "bbox": [ + 56, + 319, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{local}" + }, + { + "bbox": [ + 56, + 319, + 296, + 510 + ], + "type": "text", + "content": ". Secondly, key-frame-based temporal fusion leverages long-range temporal information by local fusion and global fusion (Section 3.1). Subsequently, the relation map decoder utilizes point-wise relation embedding (PRE) and instance-wise relation embedding (IRE) for better interaction between queries (Section 3.2). Finally, by introducing a geometric-aware classification loss and a geometric-aware matching cost, the geometry-aware alignment leverages interaction between semantic information and geometry information (Section 3.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 515, + 241, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 241, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 241, + 529 + ], + "type": "text", + "content": "3.1. Key-frame-based Temporal Fusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 664 + ], + "type": "text", + "content": "Compared to the single-frame map element prediction method, there are mainly two types of temporal fusion methods, namely stacking-based strategy and streaming-based strategy. The streaming strategy leverages all historical information by hidden state encoding. However, due to limited memory capacity, the streaming strategy has limited performance in long-range perception. The stacking strategy fuse features from specific previous frames, offering flexibility in integration of long-range information. The computational cost is linearly related to the number of fused frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We propose a key-frame-based streaming (KFS) strategy to address these problems. The well-known key-frame-based strategy is widely used in robot navigation and mapping, such as sparse mapping [16, 35] and dense map" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 230, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 410 + ], + "type": "text", + "content": "ping [6, 56]. Online reconstruction methods are unstable in long range, but they are accurate in the local regime. The input frames are divided into local frame segments. The submaps are reconstructed by integration of local frames, and then the global map is reconstructed by integration of submaps. We introduce two types of KFS strategy, namely KFS-streaming and KFS-stacking, as shown in Figure 3. Firstly, the BEV feature embedding submaps are generated by a streaming-base temporal encoder at the bottom level, which keeps the stability and temporal consistency of predictions cross latest frames. Secondly, the long-range BEV feature embedding is generated by a streaming-based or stacking-based temporal encoder at the top level. Figure 3 illustrates the KFS strategy that merges one previous submap with an interval of two." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 423, + 425, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 425, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 425, + 435 + ], + "type": "text", + "content": "3.1.1 Local BEV Fusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 442, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 555, + 519 + ], + "type": "text", + "content": "We warp the memorized BEV feature map from the previous frame to the current frame recurrently based on the ego vehicle's relative pose. Then we employ a GRU [5] to fuse current BEV feature map " + }, + { + "bbox": [ + 313, + 442, + 555, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{local}^{t}" + }, + { + "bbox": [ + 313, + 442, + 555, + 519 + ], + "type": "text", + "content": " and the warped memorized BEV feature map " + }, + { + "bbox": [ + 313, + 442, + 555, + 519 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{F}}_{submap}^{t-1}" + }, + { + "bbox": [ + 313, + 442, + 555, + 519 + ], + "type": "text", + "content": " into a single BEV feature embedding submap candidate " + }, + { + "bbox": [ + 313, + 442, + 555, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{submap}^{t}" + }, + { + "bbox": [ + 313, + 442, + 555, + 519 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 367, + 525, + 553, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 525, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 367, + 525, + 553, + 544 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {F}} _ {s u b m a p} ^ {t - 1} = \\operatorname {W a r p} \\left(\\mathcal {F} _ {s u b m a p} ^ {t - 1}, \\boldsymbol {T}\\right) \\tag {1}", + "image_path": "02f32ec805e1b3867a71ff969738a424f27c0cd85dc97fc2015a913c54654dc5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 546, + 553, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 546, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 323, + 546, + 553, + 572 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\text {s u b m a p}} ^ {t} = \\operatorname {R e s B l o c k} \\left(L N \\left(G R U \\left(\\tilde {\\mathcal {F}} _ {\\text {s u b m a p}} ^ {t - 1}, \\mathcal {F} _ {\\text {l o c a l}} ^ {t}\\right)\\right)\\right) \\tag {2}", + "image_path": "286f5e85b786593df5e918bc104e2ec8629bcfc57c9e6b055b03b225078cdf61.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 572, + 555, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 555, + 621 + ], + "type": "text", + "content": "Where ResBlock is a residual-based convolution block with the shortcut connection [11]. " + }, + { + "bbox": [ + 313, + 572, + 555, + 621 + ], + "type": "inline_equation", + "content": "LN" + }, + { + "bbox": [ + 313, + 572, + 555, + 621 + ], + "type": "text", + "content": " is a layer normalization operation. " + }, + { + "bbox": [ + 313, + 572, + 555, + 621 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 313, + 572, + 555, + 621 + ], + "type": "text", + "content": " denotes a standard " + }, + { + "bbox": [ + 313, + 572, + 555, + 621 + ], + "type": "inline_equation", + "content": "4\\times 4" + }, + { + "bbox": [ + 313, + 572, + 555, + 621 + ], + "type": "text", + "content": " transformation matrix between the coordinate systems of two frames." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 635, + 430, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 430, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 430, + 645 + ], + "type": "text", + "content": "3.1.2 Global BEV Fusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "text", + "content": "Our KFS strategy selects a BEV feature embedding as a submap using a specific distance stride " + }, + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "inline_equation", + "content": "d_{stride}" + }, + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "text", + "content": ", instead of a specific temporal interval customly." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "KFS-streaming strategy The previous selected global BEV feature is warped to the current frame. The warped" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17178" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 289, + 235 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 289, + 235 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 289, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 289, + 235 + ], + "type": "image", + "image_path": "f5b1cced7728e9d0d6325b09a9c6bd3dfc432cfc26f1cdb784c4e22c1f8049a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 244, + 240, + 256 + ], + "lines": [ + { + "bbox": [ + 110, + 244, + 240, + 256 + ], + "spans": [ + { + "bbox": [ + 110, + 244, + 240, + 256 + ], + "type": "text", + "content": "Figure 3. Temporal fusion strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 274, + 295, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 299 + ], + "type": "text", + "content": "global BEV feature and the current BEV feature map are then fused to create a global BEV feature map." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 304, + 294, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 304, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 115, + 304, + 294, + 320 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {F}} _ {\\text {g l o b a l}} ^ {\\text {p r e}} = \\operatorname {W a r p} \\left(\\mathcal {F} _ {\\text {g l o b a l}} ^ {\\text {p r e}}, \\boldsymbol {T}\\right) \\tag {3}", + "image_path": "3561bf2bc73a7863b65f044914bb11632e5a5134a966d2250c68bfa80389a332.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 325, + 295, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 325, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 91, + 325, + 295, + 342 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {g l o b a l} ^ {t} = L N \\left(G R U \\left(\\tilde {\\mathcal {F}} _ {g l o b a l} ^ {p r e}, \\mathcal {F} _ {s u b m a p} ^ {t}\\right)\\right) \\tag {4}", + "image_path": "e1726c76236a2404f30ebc69822644979ed7ab1245059979bc9c88b8b0fa9ba5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 343, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 378 + ], + "type": "text", + "content": "KFS-stacking strategy We warp " + }, + { + "bbox": [ + 55, + 343, + 296, + 378 + ], + "type": "inline_equation", + "content": "N_{pre}" + }, + { + "bbox": [ + 55, + 343, + 296, + 378 + ], + "type": "text", + "content": " selected submaps to the current frame. Following concatenation and residual-based convolution [11], we obtain the global BEV feature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 384, + 294, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 384, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 108, + 384, + 294, + 401 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {F}} _ {s u b m a p} ^ {t _ {k}} = \\operatorname {W a r p} \\left(\\mathcal {F} _ {s u b m a p} ^ {t _ {k}}, T\\right) \\tag {5}", + "image_path": "8bf6bd34c8c9506306f237ee2f4c5d99ffc75a7c4c80174f20d4ea0ed8563152.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 406, + 294, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 406, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 65, + 406, + 294, + 422 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\text {g l o b a l}} ^ {t} = \\operatorname {R e s B l o c k} \\left(\\operatorname {C o n c a t} \\left(\\tilde {F} _ {\\text {s u b m a p}}, \\mathcal {F} _ {\\text {s u b m a p}} ^ {t}\\right)\\right) \\tag {6}", + "image_path": "96a57d452b612f549bb85da1d4ab1ba86920c951bca1045363bfa3041b899b43.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 425, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 502 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 425, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{F}}_{submap}^{t_k}" + }, + { + "bbox": [ + 55, + 425, + 296, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 425, + 296, + 502 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 55, + 425, + 296, + 502 + ], + "type": "text", + "content": " represent the selected BEV feature submap and its corresponding index. The collection of selected BEV feature submaps is denoted as " + }, + { + "bbox": [ + 55, + 425, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{submap} = \\{\\tilde{\\mathcal{F}}_{submap}^{t_k}\\}_{k=1}^{N_{pre}}" + }, + { + "bbox": [ + 55, + 425, + 296, + 502 + ], + "type": "text", + "content": ". During the training period, the previous " + }, + { + "bbox": [ + 55, + 425, + 296, + 502 + ], + "type": "inline_equation", + "content": "N_{pre}" + }, + { + "bbox": [ + 55, + 425, + 296, + 502 + ], + "type": "text", + "content": " submaps are randomly selected within a range of " + }, + { + "bbox": [ + 55, + 425, + 296, + 502 + ], + "type": "inline_equation", + "content": "N_{pre} \\times d_{stride}" + }, + { + "bbox": [ + 55, + 425, + 296, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 507, + 181, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 181, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 181, + 520 + ], + "type": "text", + "content": "3.2. Relation map decoder" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 525, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 584 + ], + "type": "text", + "content": "Previous work has shown the effectiveness of interactions for map element detectors [15, 57]. In contrast to these approaches, we directly construct explicit relation priors by incorporating relation embedding into the self-attention module of the decoder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 599, + 177, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 177, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 177, + 612 + ], + "type": "text", + "content": "3.2.1 Relation Embedding" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 618, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 641 + ], + "type": "text", + "content": "Here is the self-attention in the decoder of a DETR-like framework." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 645, + 294, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 645, + 294, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 294, + 673 + ], + "type": "interline_equation", + "content": "Q _ {\\text {s e l f - a t t n}} = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {7}", + "image_path": "0c37a82f7a2f51cf03eea547d92c864a2e0db898db7e55f1e2cb52d377a346cc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "d_k" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": " is the dimension of the keys. " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "Q, K, V" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": " are feature embeddings generated by linear projections. " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "Q = \\operatorname{Linear}_q(Q_{in})" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "K = \\operatorname{Linear}_k(Q_{in})" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "V = \\operatorname{Linear}_v(Q_{in})" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "We incorporate relation embeddings into the self-attention model, inspired by [14]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 99, + 553, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 99, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 349, + 99, + 553, + 127 + ], + "type": "interline_equation", + "content": "Q _ {s e l f - a t t n} = \\operatorname {S o f t m a x} \\left(\\mathcal {F} _ {r e l} + \\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {8}", + "image_path": "381abb67ed53ade42b57d72f54c70f5eea54de24174555ded35710238d60cd21.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 132, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 132, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 555, + 192 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 132, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rel}" + }, + { + "bbox": [ + 313, + 132, + 555, + 192 + ], + "type": "text", + "content": " represents the relation embedding. In this paper, point-wise relation embedding and instance-wise relation embedding are added to the decoupled self-attention model at the point level and the instance level, respectively, employing attention refinement of information interaction." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 205, + 482, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 482, + 217 + ], + "type": "text", + "content": "3.2.2 Point-wise Relation Embedding" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 224, + 554, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 554, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 554, + 273 + ], + "type": "text", + "content": "The point-wise relation encoder represents the high-dimensional point relation embedding as an explicit geometry prior. The point-wise relation embedding is generated by normalized point coordinates and edge directions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 325, + 279, + 553, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 279, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 325, + 279, + 553, + 293 + ], + "type": "interline_equation", + "content": "\\operatorname {R e l} _ {p t} (i, j) = \\left[ \\log \\left(x _ {i} - x _ {j} + 1\\right), \\log \\left(y _ {i} - y _ {j} + 1\\right) \\right] \\tag {9}", + "image_path": "54c1983746a9f0f28cb212d3b041ec29927c07ecd8a9b2d68c55bf9681eb22b2.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 299, + 559, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 299, + 559, + 323 + ], + "spans": [ + { + "bbox": [ + 315, + 299, + 559, + 323 + ], + "type": "interline_equation", + "content": "\\operatorname {R e l} _ {\\text {d i r}} (i, j) = \\operatorname {c o s i n e \\_ s i m} \\left(e _ {i}, e _ {i + 1}\\right) - \\operatorname {c o s i n e \\_ s i m} \\left(e _ {j}, e _ {j + 1}\\right) \\tag {10}", + "image_path": "b5a65e8e1939adafaba26788f58fa038126fbcc32673a5df78044e80d9a154d7.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "text", + "content": " are normalized point coordinates. " + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "inline_equation", + "content": "e_{i}" + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "inline_equation", + "content": "e_{i+1}" + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "inline_equation", + "content": "e_{j}" + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "inline_equation", + "content": "e_{j+1}" + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "text", + "content": " are adjacent edges. The cosine similarity is denoted by " + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "inline_equation", + "content": "\\cosin\\_sim" + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "text", + "content": ". The point-wise relation embedding is unbiased, since both " + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "inline_equation", + "content": "Rel_{pt}(i,j)" + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "inline_equation", + "content": "Rel_{dir}(i,j)" + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "text", + "content": " are equal to 0 when " + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 313, + 323, + 554, + 382 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 382, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 553, + 407 + ], + "type": "text", + "content": "The point-wise relation embedding " + }, + { + "bbox": [ + 313, + 382, + 553, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rel}^{pt}" + }, + { + "bbox": [ + 313, + 382, + 553, + 407 + ], + "type": "text", + "content": " is illustrated by the following equation:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 413, + 553, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 413, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 318, + 413, + 553, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {r e l} ^ {p t} = \\operatorname {R e L U} (\\operatorname {L i n e a r} (S P E (\\operatorname {C o n c a t} (R e l _ {p t}, R e l _ {d i r})))) \\tag {11}", + "image_path": "da9c19c13a755fd360c89dc7761730bf2523b2c2c08a76a1e7edcfd4cb8f233d.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "type": "text", + "content": "Firstly, the point-wise relation embedding is calculated by concatenation of " + }, + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "type": "inline_equation", + "content": "Rel_{pt}" + }, + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "type": "inline_equation", + "content": "Rel_{dir}" + }, + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "type": "text", + "content": ". Then the embedding is transformed to high dimension by sine positional encoding (SPE). Lastly, the embedding is transformed by a linear projection and a ReLU function." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 512, + 495, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 495, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 495, + 525 + ], + "type": "text", + "content": "3.2.3 Instance-wise Relation Embedding" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 531, + 553, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 553, + 555 + ], + "type": "text", + "content": "The instance-wise relation embedding is generated by classification scores and chamfer distance of points." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 380, + 562, + 553, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 562, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 380, + 562, + 553, + 576 + ], + "type": "interline_equation", + "content": "\\operatorname {R e l} _ {s} (i, j) = \\operatorname {s g n} \\left(s _ {i} - s _ {j}\\right) \\tag {12}", + "image_path": "45c08dd2da888d96e3d64fb257efeaf8f239c2637f5d7072010c96c830ed0845.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 330, + 582, + 553, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 582, + 553, + 595 + ], + "spans": [ + { + "bbox": [ + 330, + 582, + 553, + 595 + ], + "type": "interline_equation", + "content": "\\operatorname {R e l} _ {s d} (i, j) = \\operatorname {s g n} \\left(s _ {i} - s _ {j}\\right) \\cdot \\operatorname {C h a m f e r} \\left(I _ {i}, I _ {j}\\right) \\tag {13}", + "image_path": "8cec4f362e449d079eeaef0b9cb3c4c5f1fec43f840f6588a894c3b9ea1838a5.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "inline_equation", + "content": "sgn" + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "text", + "content": " is the sign function. " + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "inline_equation", + "content": "s_i, s_j" + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "text", + "content": " are classification scores of the " + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "text", + "content": " candidates. Chamfer " + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "inline_equation", + "content": "(I_i, I_j)" + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "text", + "content": " is the chamfer distance between points of instance " + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "text", + "content": " and instance " + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "text", + "content": ". The equation of " + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "inline_equation", + "content": "Rel_s(i,j)" + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "text", + "content": " establishes a simple classification ranking relation among pairs of instances. Then we add the position relation of the map elements to the instance-wise relation embedding " + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rel}^{ins}" + }, + { + "bbox": [ + 313, + 597, + 555, + 693 + ], + "type": "text", + "content": " by the chamfer distance." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 351, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 351, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {r e l} ^ {i n s} = R e L U \\left(L i n e a r \\left(S P E \\left(R e l _ {s d}\\right)\\right)\\right) \\tag {14}", + "image_path": "fc9d97ad10c861841044c3799f6776689437823328ff81b46c0586671514202a.jpg" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17179" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "The instance-wise relation embedding is unbiased, since " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "Rel_{sd}(i,j)" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " is equal to 0 when " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 104, + 210, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 104, + 210, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 104, + 210, + 117 + ], + "type": "text", + "content": "3.3. Geometry-aware Alignment" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 122, + 296, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 296, + 243 + ], + "type": "text", + "content": "Current work, such as MapTR [24] uses the classification branch and the regression branch to predict semantic labels and geometric localization separately. MapTR uses the focal loss [42] as the classification loss and the point-to-point L1 loss as the regression loss. Since classification and regression losses are independent, the quality of these two branches is inconsistent, leading to the well-known misalignment problem. In this paper, we introduce the geometry-aware focal loss and the geometry-aware focal cost to overcome this problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 257, + 209, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 209, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 209, + 269 + ], + "type": "text", + "content": "3.3.1 Geometry-aware Focal Loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 276, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 323 + ], + "type": "text", + "content": "Inspired by IoU-aware Focal Loss [21, 51] in traditional object detection, we design the novel Geometry-aware Focal Loss (GFL). Unlike focal loss, we treat positives and negatives asymmetrically." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 332, + 299, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 299, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 299, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G F L} = \\sum_ {i = 1} ^ {N _ {\\text {p o s}}} s _ {\\text {g e o} _ {i}} B C E \\left(s _ {\\text {g e o} _ {i}}, p _ {i}\\right) + \\sum_ {j = 1} ^ {N _ {\\text {n e g}}} \\alpha p _ {j} ^ {\\gamma} B C E \\left(p _ {j}, 0\\right) \\tag {15}", + "image_path": "069d91f7844adfcce16224fb9a704725511b47ffb52e8fe66874e4c05ffe1ce5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "text", + "content": " are the predicted probability of the " + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "text", + "content": " candidates, " + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "inline_equation", + "content": "s_{geo} \\in [0,1]" + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "text", + "content": " is an instance Geometry-aware Classification Score (GCS). Here we define three types of GCS, including " + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "inline_equation", + "content": "s_{p2p}" + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "inline_equation", + "content": "s_{dir}" + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "inline_equation", + "content": "s_{giou}" + }, + { + "bbox": [ + 55, + 390, + 296, + 440 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 449, + 295, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 449, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 71, + 449, + 295, + 483 + ], + "type": "interline_equation", + "content": "s _ {p 2 p} = 1 - \\frac {1}{2 N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} D _ {\\text {M a n h a t t a n}} \\left(p _ {\\text {s r c} _ {i}}, p _ {\\text {t g t} _ {i}}\\right) \\tag {16}", + "image_path": "60419db2684dc094594896470967744bed7feae089c370fca4d6d6bb4d399d16.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 491, + 295, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 491, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 70, + 491, + 295, + 525 + ], + "type": "interline_equation", + "content": "s _ {d i r} = 0. 5 + \\frac {1}{2 N _ {e}} \\sum_ {i = 1} ^ {N _ {e}} \\operatorname {c o s i n e \\_ s i m} \\left(e _ {s r c _ {i}}, e _ {t g t _ {i}}\\right) \\tag {17}", + "image_path": "c3d60942e25a82646464d1d93dcbefb249cbf956afed0cd0b47e66eacc1fdfd6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 96, + 529, + 295, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 529, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 96, + 529, + 295, + 544 + ], + "type": "interline_equation", + "content": "s _ {g i o u} = 0. 5 + 0. 5 \\cdot G I o U \\left(B _ {s r c}, B _ {t g t}\\right) \\tag {18}", + "image_path": "9cf67b2144f0e5880edea47063add99ba283cf50748d8df1782ee4ae12068d5a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "inline_equation", + "content": "s_{p2p} \\in [0,1]" + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "text", + "content": " is a normalized point-to-point L1 score, " + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "inline_equation", + "content": "s_{dir} \\in [0,1]" + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "text", + "content": " is an edge direction score calculated by cosine similarity and " + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "inline_equation", + "content": "s_{giou} \\in [0,1]" + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "text", + "content": " is a normalized GIoU [40] score of candidate bounding box " + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "inline_equation", + "content": "B_{src}" + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "text", + "content": " and assigned GT " + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "inline_equation", + "content": "B_{tgt}" + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "text", + "content": ". The bounding box is computed by minimum enclosing box of instance points. " + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "inline_equation", + "content": "N_e" + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "text", + "content": " are the numbers of points and edges. " + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "inline_equation", + "content": "p_{src}" + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "inline_equation", + "content": "p_{tgt}" + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "text", + "content": " are normalized points of candidates and assigned GT. " + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "inline_equation", + "content": "e_{src}" + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "inline_equation", + "content": "e_{tgt}" + }, + { + "bbox": [ + 55, + 548, + 296, + 656 + ], + "type": "text", + "content": " are edges of candidates and assigned GT." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 671, + 208, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 208, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 208, + 683 + ], + "type": "text", + "content": "3.3.2 Geometry-aware Focal Cost" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "The loss function " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GFL}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " supervises the classification scores with geometric metrics GCS " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "s_{geo_i}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": ". And " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "s_{geo_i}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " accurately" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "selects these high-quality candidates, which leads to better map vectorization performance. Following the spirit of GFL, we make similar modifications to the focal cost, which is a key component of label assignment. The novel matching cost, geometry-aware focal cost (GFC) is as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 136, + 555, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 136, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 313, + 136, + 555, + 162 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {G F C} (i, j) = s _ {g e o _ {i}} B C E \\left(s _ {g e o _ {i}}, p _ {i}\\right) - \\alpha p _ {i} ^ {\\gamma} B C E \\left(1 - p _ {i}, 1\\right) \\tag {19}", + "image_path": "1aed65d8d94f1331678199a03e39d8cf0c93e6c82f6b46240a5ece4110ea9ef8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 166, + 555, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 555, + 227 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 166, + 555, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{GFC}(i,j)" + }, + { + "bbox": [ + 313, + 166, + 555, + 227 + ], + "type": "text", + "content": " is the geometry-aware focal cost for the " + }, + { + "bbox": [ + 313, + 166, + 555, + 227 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 166, + 555, + 227 + ], + "type": "text", + "content": " prediction and the " + }, + { + "bbox": [ + 313, + 166, + 555, + 227 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 313, + 166, + 555, + 227 + ], + "type": "text", + "content": " ground truth. The GFC is used as a modulated function to suppress candidates with inaccurate prediction localization and promote candidates with an accurate position." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 232, + 402, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 232, + 402, + 244 + ], + "spans": [ + { + "bbox": [ + 314, + 232, + 402, + 244 + ], + "type": "text", + "content": "3.4. Training Loss" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "type": "text", + "content": "Auxiliary BEV Supervision. We introduce a query-based instance segmentation (QIS) module using an instance mask prediction branch. Inspired by [17], we make use of the query embeddings " + }, + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "type": "inline_equation", + "content": "Q_{dec}" + }, + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "type": "text", + "content": " from Transformer to dot-product the BEV feature embedding map " + }, + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{bev}" + }, + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "type": "text", + "content": " to obtain instance-wise binary masks " + }, + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "type": "text", + "content": " with the sigmoid function " + }, + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 368, + 326, + 554, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 326, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 368, + 326, + 554, + 340 + ], + "type": "interline_equation", + "content": "\\hat {M} _ {\\text {i n s}} = \\sigma (M L P \\left(Q _ {\\text {d e c}}\\right) \\cdot \\mathcal {F} _ {\\text {b e v}}) \\tag {20}", + "image_path": "8310569a4e79113f32df2749ad406914807937e3c04e7fe879b7623e0528e00e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 344, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 555, + 392 + ], + "type": "text", + "content": "For the instance segmentation branch, we extend the mask focal loss and the mask focal cost to mask geometry-aware focal loss (MGFL) and mask geometry-aware focal cost (MGFC) as well, with the same ideal of GFL." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 408, + 556, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 556, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 556, + 433 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s e g}} = \\lambda_ {m g f} \\mathcal {L} _ {m g f} (\\hat {M} _ {\\text {i n s}}, M _ {\\text {i n s}}) + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} (\\hat {M} _ {\\text {i n s}}, M _ {\\text {i n s}}) \\tag {21}", + "image_path": "eea3c7f01cbb49e83af3cd67f58992bd087cdea5810768b66fcad9d7d6730052.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "text", + "content": "Instance masks " + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\hat{M}_{ins}" + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "text", + "content": " are supervised by instance mask annotations " + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "inline_equation", + "content": "M_{ins}" + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "text", + "content": ", using the mask geometry-aware focal loss " + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mgf}" + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "text", + "content": " and the dice loss " + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dice}" + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 469, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 554, + 491 + ], + "type": "text", + "content": "Overall Loss. The total loss " + }, + { + "bbox": [ + 313, + 469, + 554, + 491 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 469, + 554, + 491 + ], + "type": "text", + "content": " used by InteractionMap is as follow:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 381, + 494, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 494, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 381, + 494, + 553, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {d e t}} + \\mathcal {L} _ {\\text {s e g}} + \\mathcal {L} _ {\\text {a u x}} \\tag {22}", + "image_path": "aba3c17f1b8d7a73c341826c95b72be118f0be6c8c1cb4e837d9041d3932f678.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 508, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 543 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 508, + 554, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{det}" + }, + { + "bbox": [ + 313, + 508, + 554, + 543 + ], + "type": "text", + "content": " is the loss term for the detection task. " + }, + { + "bbox": [ + 313, + 508, + 554, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{seg}" + }, + { + "bbox": [ + 313, + 508, + 554, + 543 + ], + "type": "text", + "content": " is the loss term for the segmentation task. " + }, + { + "bbox": [ + 313, + 508, + 554, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{aux}" + }, + { + "bbox": [ + 313, + 508, + 554, + 543 + ], + "type": "text", + "content": " is the auxiliary loss term." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 544, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 569 + ], + "type": "text", + "content": "The detection loss consists of classification loss " + }, + { + "bbox": [ + 313, + 544, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 313, + 544, + 554, + 569 + ], + "type": "text", + "content": ", point-to-point distance loss " + }, + { + "bbox": [ + 313, + 544, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p2p}" + }, + { + "bbox": [ + 313, + 544, + 554, + 569 + ], + "type": "text", + "content": ", edge direction loss " + }, + { + "bbox": [ + 313, + 544, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dir}" + }, + { + "bbox": [ + 313, + 544, + 554, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 351, + 573, + 553, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 573, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 351, + 573, + 553, + 587 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e t}} = \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\lambda_ {p 2 p} \\mathcal {L} _ {p 2 p} + \\lambda_ {\\text {d i r}} \\mathcal {L} _ {\\text {d i r}} \\tag {23}", + "image_path": "f3c2774dfadd689fe3b78034c0d95b6eaf0e1a42f17230048538d97a1c207ef0.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": "The loss term " + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": " is the geometry-aware focal loss for the classification of instances. The loss term " + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p2p}" + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": " is the smooth L1 loss for the regression of vectorized map elements. And the loss term " + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dir}" + }, + { + "bbox": [ + 313, + 590, + 554, + 651 + ], + "type": "text", + "content": " is the loss of cosine similarity between adjacent edges." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 350, + 662, + 553, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 662, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 350, + 662, + 553, + 676 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a u x}} = \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {P V S e g} \\mathcal {L} _ {P V S e g} \\tag {24}", + "image_path": "6f5512a5c5b9d143420b45ac46f99dee5dc4f223be15aa92842bc4e4d1e090d5.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "The auxiliary loss term " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{aux}" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " is composed of the depth-map prediction loss " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{depth}" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " and image semantic segmentation loss " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PVSeg}" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17180" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 132, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 132, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 132, + 85 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 181, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 181, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 181, + 104 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 110, + 125, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 125, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 125, + 121 + ], + "type": "text", + "content": "4.1.1 Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 129, + 295, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 295, + 236 + ], + "type": "text", + "content": "nuScenes Datasets. There are 1000 scenes of 4 locations in nuScenes datasets. Each scene contains 20s of RGB images from six cameras, point clouds from LiDAR sweeps, and a 3D vectorized map. The train set contains 700 scenes with 28130 samples and validation set contains 150 scenes with 6019 samples, respectively. Following previous works [24, 25], we mainly focus on three categories of map elements, including lane divider, pedestrian crossing and road boundary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 237, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 295, + 331 + ], + "type": "text", + "content": "Argoverse2 Datasets. Argoverse2 dataset provide 1000 scenes in 6 cities, each lasting around 15 seconds. Each log includes 7 RGB images from surrounding cameras and point cloud from LiDAR sweeps. Follow previous methods [24, 25], 700 scenes are used for training, and 150 scenes are used for validation. For a fair comparison, we report results on its validation set and focus on the same three map categories as the nuScene dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 348, + 170, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 170, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 170, + 360 + ], + "type": "text", + "content": "4.1.2 Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": "For a fair comparison, we follow the standard metric used in previous works [19, 24, 25, 29]. With the ego-coordination as the center, the perception ranges are " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "[-15m, 15m]" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": " for the X-axis and " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "[-30m, 30m]" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": " for the Y-axis. We adopt the mean Average Precision (mAP) to evaluate the map vectorization quality based on the chamfer distance. A candidate is considered as True-Positive (TP) only if the chamfer distance to ground truth is less than certain thresholds " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "(\\tau \\in T, T = 0.5, 1.0, 1.5)" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": ". The final AP metric is calculated by average across all thresholds and all classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 502, + 189, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 189, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 189, + 514 + ], + "type": "text", + "content": "4.1.3 Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "Following previous method [24, 25], we use ResNet50 [11] as the backbone for RGB images. All models are trained with 8 A800 GPUs. The default batch size is 32. The initial learning rate is set to " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "6 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": " with cosine decay. The nuScenes images resolution is " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "1600 \\times 900" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ". We resize the images with 0.5 ratio. For the Argoverse2 dataset, the image resolutions are " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "2048 \\times 1550" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "1550 \\times 2048" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ". We pad the images to " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "2048 \\times 2048" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ", and then resize the images with 0.3 ratio. We define the size of BEV feature as " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "200 \\times 100" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ", and the size of each BEV grid as 0.3 meters. The default numbers of instance queries, point queries, and decoder layers are 100, 20, 6, respectively. The model is trained on 24, 110 epochs on the nuScenes and Argoverse2 datasets. We randomly divide each training sequence into 3 splits to generate more diverse data sequences. Inspired by SOLOFusion [37], we train the ini" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "tial " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "N_{init}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " epochs with single-frame input to stabilize multi-frame straining. " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "N_{init} = 0.5N_{total}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ". For loss weights, " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\lambda_{cls} = 2" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\lambda_{p2p} = 4" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\lambda_{dir} = 0.005" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\lambda_{mgf} = 30" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\lambda_{dice} = 3" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\lambda_{depth} = 3" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\lambda_{PVSeg} = 2" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ". For temporal fusion, " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "N_{pre} = 4" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "d_{stride} = 5m" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 140, + 458, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 458, + 153 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 458, + 153 + ], + "type": "text", + "content": "4.2. Performance Comparison" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 159, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 553, + 266 + ], + "type": "text", + "content": "Performance on nuScenes Dataset. As shown in Table 1, our methods show consistent improvements over baselines. By integrating our method with MapTRv2 [25], we achieve " + }, + { + "bbox": [ + 313, + 159, + 553, + 266 + ], + "type": "inline_equation", + "content": "71.6(+10.1)" + }, + { + "bbox": [ + 313, + 159, + 553, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 159, + 553, + 266 + ], + "type": "inline_equation", + "content": "71.8(+10.3)" + }, + { + "bbox": [ + 313, + 159, + 553, + 266 + ], + "type": "text", + "content": " mAP with 24-epoch training schedules and " + }, + { + "bbox": [ + 313, + 159, + 553, + 266 + ], + "type": "inline_equation", + "content": "75.5(+6.8)" + }, + { + "bbox": [ + 313, + 159, + 553, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 159, + 553, + 266 + ], + "type": "inline_equation", + "content": "76.0(+7.3)" + }, + { + "bbox": [ + 313, + 159, + 553, + 266 + ], + "type": "text", + "content": " mAP with 110-epoch training schedules, respectively. InteractionMap establishes a new state-of-the-art performance, exhibiting significant improvements over existing methods with comparable inference speed, as shown in Table 4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 267, + 554, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 554, + 374 + ], + "type": "text", + "content": "Performance on Argoverse2 Dataset. Argoverse2 dataset contains 3D map elements, including height information. For fair comparison, the sampling frequency is set to " + }, + { + "bbox": [ + 313, + 267, + 554, + 374 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 267, + 554, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 267, + 554, + 374 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 267, + 554, + 374 + ], + "type": "text", + "content": ", respectively. The amount of training data at a " + }, + { + "bbox": [ + 313, + 267, + 554, + 374 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 267, + 554, + 374 + ], + "type": "text", + "content": " sampling frequency is 5 times as many as that at a " + }, + { + "bbox": [ + 313, + 267, + 554, + 374 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 267, + 554, + 374 + ], + "type": "text", + "content": " sampling frequency. As demonstrated in Table 2 and Table 3, InteractionMap surpasses all state-of-the-art methods in both 2D and 3D vectorized map perception on Argoverse2 dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 384, + 408, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 408, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 408, + 396 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 403, + 553, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 553, + 474 + ], + "type": "text", + "content": "We examine the efficacy of each component of InteractionMap through ablation studies, utilizing the nuScenes dataset. Initially, we build a simple baseline base on MapTR [24] with the LSS encoder without auxiliary depth-map supervision. The default numbers of instances, points per instance, are 50 and 20, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 475, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 553, + 594 + ], + "type": "text", + "content": "Contributions of Main Components. Table 5 demonstrates the impact of each component of InteractionMap. When adding relation embedding into the baseline model, there is a substantial improvement of " + }, + { + "bbox": [ + 313, + 475, + 553, + 594 + ], + "type": "inline_equation", + "content": "4.5\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 475, + 553, + 594 + ], + "type": "text", + "content": ". Furthermore, the addition of geometry alignment results in a significant 7.5 increase in mAP, emphasizing its effectiveness in boosting performance. Lastly, key-frame-based temporal fusion offers 5.0 and 5.1 additional increases in mAP, indicating that the inclusion of long-range hierarchical streaming strategy significantly enhances the performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": "Ablation on Relation Embedding. In Table 6, we explore the effect of relation embedding in point-level and instance-level. Adding the point-wise relation embedding of point coordinates and edge direction results in 2.1 and 0.9 improvements in mAP over the baseline, respectively. The instance-wise relation embedding leverages the classification confidence and points coordinates, leading to enhancements of 1.1 and " + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "inline_equation", + "content": "3.1\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": ", respectively. The combination of relation embedding in point-level and instance-level enhances performance by " + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "inline_equation", + "content": "2.8\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": ". Finally, we introduce a" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17181" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 70, + 488, + 361 + ], + "blocks": [ + { + "bbox": [ + 123, + 70, + 488, + 361 + ], + "lines": [ + { + "bbox": [ + 123, + 70, + 488, + 361 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 488, + 361 + ], + "type": "table", + "html": "
MethodBackboneEpochsTemporal\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bou}\\)mAP
MapTRR502446.351.553.150.3
MapVRR502447.754.451.451.2
PivotNetR503053.855.859.657.4
BeMapNetR503057.762.359.459.8
MapTRv2R502459.862.462.461.5
StreamMapNetR5030✓61.766.362.163.4
MGMapR502461.865.067.564.8
SQD-MapNetR5024✓63.666.664.865.0
MapQRR502463.468.067.766.4
HIMapR503062.668.469.166.7
HRMapNet-MapTRv2R5024✓65.867.468.567.2
InteractionMap-RR5024✓71.370.872.871.6
InteractionMap-CR5024✓69.772.773.071.8
VectorMapNetR50110+ft42.551.444.146.0
MapTRR5011056.259.860.158.7
MapVRR5011055.061.859.458.8
BeMapNetR5011062.666.765.164.8
MapTRv2R5011068.168.369.768.7
MapQRR5011070.174.473.272.6
HRMapNet-MapTRv2R50110✓72.072.975.873.6
HIMapR5011071.375.074.773.7
InteractionMap-RR50110✓75.275.076.175.5
InteractionMap-CR50110✓75.375.677.076.0
", + "image_path": "f9f30d81f7dad3c5d727d613fa63588f1d64f34af6be6bccf5bb51c6325d6e3a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 413, + 291, + 507 + ], + "blocks": [ + { + "bbox": [ + 55, + 372, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 372, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 555, + 407 + ], + "type": "text", + "content": "Table 1. Comparison with SOTA methods on the nuScenes validation set at " + }, + { + "bbox": [ + 55, + 372, + 555, + 407 + ], + "type": "inline_equation", + "content": "60m\\times 30m" + }, + { + "bbox": [ + 55, + 372, + 555, + 407 + ], + "type": "text", + "content": " perception range. \"EB0\", \"EB4\", \"R50\" correspond to the backbones Efficient-B0, Efficient-B4 [44], ResNet50 [11]. \"ft\" means the two-stage fine-tune strategy. \"C\" means KFS-stacking strategy and \"R\" means KFS-streaming strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 413, + 291, + 507 + ], + "lines": [ + { + "bbox": [ + 60, + 413, + 291, + 507 + ], + "spans": [ + { + "bbox": [ + 60, + 413, + 291, + 507 + ], + "type": "table", + "html": "
MethodDimAPpedAPdivAPboumAP
StreamMapNet2d62.059.563.061.5
SQD-MapNet2d64.960.264.963.3
MapTRv22d60.068.764.264.3
HRMapNet-MapTRv22d65.171.468.668.3
InteractionMap-C2d68.577.974.673.7
InteractionMap-R2d70.177.074.073.7
", + "image_path": "d163be6023b02d6a7331ef5eaca404266e4fb0d20cf87cd71bebb5de4ab76c66.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "type": "text", + "content": "alternating training strategy to randomly turn off one of these two components, leading the overall design achieves " + }, + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "type": "inline_equation", + "content": "54.7(+4.5)" + }, + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "type": "text", + "content": " mAP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 701 + ], + "type": "text", + "content": "Ablation on Key-frame-based Temporal Fusion. To investigate the effectiveness of the key-frame-based streaming strategy, we conduct ablation experiments to compare it with the baseline (without temporal fusion) and the streaming strategy, in Table 7. Employing the streaming strategy results in " + }, + { + "bbox": [ + 55, + 594, + 295, + 701 + ], + "type": "inline_equation", + "content": "3.8\\mathrm{mAP}" + }, + { + "bbox": [ + 55, + 594, + 295, + 701 + ], + "type": "text", + "content": " over the baseline. Furthermore, when the KFS-stacking strategy is introduced, the mAP increased by 7.4, reaching 57.6. Additionally, the inclusion of KFS-streaming strategy leads to an improvement of " + }, + { + "bbox": [ + 55, + 594, + 295, + 701 + ], + "type": "inline_equation", + "content": "7.2\\mathrm{mAP}" + }, + { + "bbox": [ + 55, + 594, + 295, + 701 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "type": "text", + "content": "Ablation on Geometry-aware Alignment. As shown in" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 413, + 550, + 555 + ], + "blocks": [ + { + "bbox": [ + 55, + 517, + 295, + 552 + ], + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 552 + ], + "type": "text", + "content": "Table 2. Comparison with SOTA methods on the Argoverse2 validation set at " + }, + { + "bbox": [ + 55, + 517, + 295, + 552 + ], + "type": "inline_equation", + "content": "60m \\times 30m" + }, + { + "bbox": [ + 55, + 517, + 295, + 552 + ], + "type": "text", + "content": " perception range with a 30-epoch training schedule, at a 2Hz sampling frequency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 413, + 550, + 555 + ], + "lines": [ + { + "bbox": [ + 319, + 413, + 550, + 555 + ], + "spans": [ + { + "bbox": [ + 319, + 413, + 550, + 555 + ], + "type": "table", + "html": "
MethodDimAPpedAPdivAPboumAP
MapTRv22d63.671.567.467.5
MapQR2d64.372.368.168.2
HIMap2d69.069.570.369.6
InteractionMap-C2d69.878.674.874.4
InteractionMap-R2d70.878.174.974.6
MapTRv23d60.768.964.564.7
MapQR3d60.171.266.265.9
HIMap3d66.768.370.368.4
InteractionMap-C3d66.675.672.771.6
InteractionMap-R3d67.775.573.172.1
", + "image_path": "f8b1fe4f8491fadadf61a7df218497a1d577e8d2878483519dfc6872f844bcfb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 319, + 605, + 550, + 664 + ], + "blocks": [ + { + "bbox": [ + 313, + 564, + 555, + 599 + ], + "lines": [ + { + "bbox": [ + 313, + 564, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 555, + 599 + ], + "type": "text", + "content": "Table 3. Comparison with SOTA methods on the Argoverse2 validation set at " + }, + { + "bbox": [ + 313, + 564, + 555, + 599 + ], + "type": "inline_equation", + "content": "60m \\times 30m" + }, + { + "bbox": [ + 313, + 564, + 555, + 599 + ], + "type": "text", + "content": " perception range with a 6-epoch training schedule, at a " + }, + { + "bbox": [ + 313, + 564, + 555, + 599 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 564, + 555, + 599 + ], + "type": "text", + "content": " sampling frequency." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 605, + 550, + 664 + ], + "lines": [ + { + "bbox": [ + 319, + 605, + 550, + 664 + ], + "spans": [ + { + "bbox": [ + 319, + 605, + 550, + 664 + ], + "type": "table", + "html": "
MethodFPSParams(MB)mAP
MapTRv214.940.661.5
InteractionMap-C11.0116.771.8
InteractionMap-R12.149.971.6
", + "image_path": "98724e414ada8d6fa36a0ca166620a7b874568524b7c879c33a55906d1a9134e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 673, + 555, + 696 + ], + "lines": [ + { + "bbox": [ + 313, + 673, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 555, + 696 + ], + "type": "text", + "content": "Table 4. Comparison with SOTA methods on the nuScenes validation set. FPS is measured on a single A800 GPU." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17182" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 292, + 182 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 292, + 182 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 292, + 182 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 292, + 182 + ], + "type": "table", + "html": "
REMGAMTFMAPpedAPdivAPboumAP
46.750.553.550.2
✓50.156.657.354.7
✓57.163.863.761.5
C50.960.061.957.6
R50.661.360.357.4
✓✓58.665.162.962.2
✓✓C64.268.668.867.2
✓✓R64.168.169.867.3
", + "image_path": "61fc62d3c8e05d886be4f160319fd135fe9d2f7e2c22aecc580436d1b2a3f988.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 245, + 292, + 357 + ], + "blocks": [ + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "type": "text", + "content": "Table 5. Ablation study of each component, including Relation Embedding Module (REM), Temporal Fusion Module (TFM), Geometry-aware Alignment Module (GAM). \"C\" means KFS-stacking strategy and \"R\" means KFS-streaming strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 245, + 292, + 357 + ], + "lines": [ + { + "bbox": [ + 60, + 245, + 292, + 357 + ], + "spans": [ + { + "bbox": [ + 60, + 245, + 292, + 357 + ], + "type": "table", + "html": "
PtxyPtdirInsSInssdPPAPpedAPdivAPboumAP
46.750.553.550.2
✓47.852.756.352.3
✓47.052.453.951.1
✓✓47.855.656.453.3
✓46.952.154.851.3
✓47.955.756.453.3
✓✓✓47.655.955.553.0
✓✓✓✓50.156.657.354.7
", + "image_path": "588d67f369381408dc274ac9064a938debb1b6fd8df4c033c86c9ca60384914b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 60, + 441, + 292, + 512 + ], + "blocks": [ + { + "bbox": [ + 55, + 365, + 295, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 433 + ], + "type": "text", + "content": "Table 6. Ablation study of relation embedding by employing point-wise coordinate embedding " + }, + { + "bbox": [ + 55, + 365, + 295, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{Pt}_{xy}" + }, + { + "bbox": [ + 55, + 365, + 295, + 433 + ], + "type": "text", + "content": ", point-wise edge direction embedding " + }, + { + "bbox": [ + 55, + 365, + 295, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{Pt}_{dir}" + }, + { + "bbox": [ + 55, + 365, + 295, + 433 + ], + "type": "text", + "content": ", instance-wise classification score embedding " + }, + { + "bbox": [ + 55, + 365, + 295, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{Ins}_s" + }, + { + "bbox": [ + 55, + 365, + 295, + 433 + ], + "type": "text", + "content": ", instance-wise classifier score and chamfer distance embedding " + }, + { + "bbox": [ + 55, + 365, + 295, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{Ins}_{sd}" + }, + { + "bbox": [ + 55, + 365, + 295, + 433 + ], + "type": "text", + "content": " and alternating training strategy of relation embedding, namely Ping-Pong (PP)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 441, + 292, + 512 + ], + "lines": [ + { + "bbox": [ + 60, + 441, + 292, + 512 + ], + "spans": [ + { + "bbox": [ + 60, + 441, + 292, + 512 + ], + "type": "table", + "html": "
MethodAPpedAPdivAPboumAP
Baseline46.750.553.550.2
Streaming48.654.758.854.0
KFS-stacking50.960.061.957.6
KFS-streaming50.661.360.357.4
", + "image_path": "62d2c4de49fd0cbfeab314ebe42d650c6a7b0cd43ffdbddc29c8f99ccedebc2f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Table 8, geometry-aware alignment comprised five main components: query-based instance segmentation (QIS), geometry-aware focal loss (GFL), geometry-aware focal cost (GFC), mask geometry-aware focal loss (MGFL) and mask geometry-aware focal cost (MGFC). The QIS leverages geometry priors by instance segmentation mask supervision. Compared to the baseline, the QIS leads to an improvement of " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "6.4\\mathrm{mAP}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": ". The GFL encourages predictions with high localization metrics to obtain a better classification score, with an improvement of " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "3.1\\mathrm{mAP}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " compared to mask focal loss. The GFC suppresses the matching cost of candidates with an inaccurate prediction position, which brings an improvement of " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.9\\mathrm{mAP}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": ". For the segmentation" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 71, + 550, + 163 + ], + "blocks": [ + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "type": "text", + "content": "Table 7. Ablation study of streaming strategy, KFS-streaming strategy and KFS-stacking strategy." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 71, + 550, + 163 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 550, + 163 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 550, + 163 + ], + "type": "table", + "html": "
QISGFLGFCMGFLMGFCAPpedAPdivAPboumAP
46.750.553.550.2
✓51.758.759.456.6
✓✓55.162.062.059.7
✓✓✓55.863.862.260.6
✓✓✓✓57.063.662.361.0
✓✓✓✓✓57.163.863.761.5
", + "image_path": "c1eaadaeb826807f77bd4f4ecd92f56162b342173b1b1133cd3b052c9f837694.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 319, + 236, + 550, + 318 + ], + "blocks": [ + { + "bbox": [ + 313, + 171, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 228 + ], + "type": "text", + "content": "Table 8. Ablation study of geometry-aware alignment by investigating the performance of query-based instance segmentation (QIS), geometry-aware focal loss (GFL), geometry-aware focal cost (GFC), mask geometry-aware focal loss (MGFL) and mask geometry-aware focal cost (MGFC)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 236, + 550, + 318 + ], + "lines": [ + { + "bbox": [ + 319, + 236, + 550, + 318 + ], + "spans": [ + { + "bbox": [ + 319, + 236, + 550, + 318 + ], + "type": "table", + "html": "
SgiouSp2pSdirAPpedAPdivAPboumAP
✓51.758.759.456.6
52.656.258.055.6
✓54.860.263.159.4
✓58.363.162.561.3
✓✓57.163.863.761.5
", + "image_path": "e6b105a8455816c74fc292db258dd772ca3c8b055a3929ae97b7e00458331bf9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 326, + 549, + 338 + ], + "lines": [ + { + "bbox": [ + 318, + 326, + 549, + 338 + ], + "spans": [ + { + "bbox": [ + 318, + 326, + 549, + 338 + ], + "type": "text", + "content": "Table 9. Ablation study of geometry-aware classification score." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 349, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 397 + ], + "type": "text", + "content": "branch, the result shows that the position-supervised segmentation loss MGFL and the position-modulated segmentation cost MGFC improve the final results, with " + }, + { + "bbox": [ + 313, + 349, + 555, + 397 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 349, + 555, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 349, + 555, + 397 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 349, + 555, + 397 + ], + "type": "text", + "content": " gains, individually." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "text", + "content": "We further investigate the geometry-aware classification score " + }, + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "inline_equation", + "content": "S_{geo}" + }, + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "text", + "content": ". As depicted in Table 9, using " + }, + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "inline_equation", + "content": "S_{giou}" + }, + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "inline_equation", + "content": "S_{p2p}" + }, + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "inline_equation", + "content": "S_{dir}" + }, + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "text", + "content": " leads to individual improvement of -1.0, 2.8 and 4.7 mAP, respectively. Furthermore, the combination of " + }, + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "inline_equation", + "content": "S_{p2p}" + }, + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "inline_equation", + "content": "S_{dir}" + }, + { + "bbox": [ + 313, + 397, + 554, + 458 + ], + "type": "text", + "content": " boosts the performance, improving by 4.9 mAP." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 470, + 388, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 470, + 388, + 483 + ], + "spans": [ + { + "bbox": [ + 314, + 470, + 388, + 483 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 491, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 658 + ], + "type": "text", + "content": "In this paper, we introduce InteractionMap, a novel approach to end-to-end online vectorized HD map utilizing temporal and spatial information interaction from local to global. By leveraging relation embedding at both point level and instance level, the network learns the geometric priors of map elements better. Subsequently, key-frame-based streaming strategy enhances the network by utilizing temporal feature embedding from near to distant. Finally, the geometric-aware focal loss and the geometric-aware focal cost build a strong correlation between classification score and instance-point location precision. Across various experimental settings, our proposed InteractionMap yields significant performance improvements over the baseline on various datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 671, + 373, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 671, + 373, + 684 + ], + "spans": [ + { + "bbox": [ + 314, + 671, + 373, + 684 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 691, + 555, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 691, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 318, + 691, + 555, + 715 + ], + "type": "text", + "content": "[1] Yingfeng Cai, Wei Dong, Ze Liu, Hai Wang, and Long Chen. Homap: End-to-end vectorized hd map construction with" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17183" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "high-order modeling. IEEE Transactions on Intelligent Vehicles, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 148 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 150, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 150, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 61, + 150, + 294, + 204 + ], + "type": "text", + "content": "[3] Jiacheng Chen, Yuefan Wu, Jiaqi Tan, Hang Ma, and Yasutaka Furukawa. Maptracker: Tracking with strided memory fusion for consistent vector hd mapping. In European Conference on Computer Vision, pages 90-107. Springer, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 206, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 294, + 258 + ], + "type": "text", + "content": "[4] Shaoyu Chen, Tianheng Cheng, Xinggang Wang, Wenming Meng, Qian Zhang, and Wenyu Liu. Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. arXiv preprint arXiv:2206.04584, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 304 + ], + "type": "text", + "content": "[5] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 305, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 305, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 305, + 294, + 358 + ], + "type": "text", + "content": "[6] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4): 1, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 361, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 294, + 404 + ], + "type": "text", + "content": "[7] Wenjie Ding, Limeng Qiao, Xi Qiu, and Chi Zhang. Pivot-net: Vectorized pivot learning for end-to-end hd map construction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3672-3682, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 405, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 405, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 405, + 294, + 468 + ], + "type": "text", + "content": "[8] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525-11533, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 471, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 471, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 471, + 294, + 524 + ], + "type": "text", + "content": "[9] Yulu Gao, Yifan Sun, Xudong Ding, Chuyang Zhao, and Si Liu. Ease-detr: Easing the competition among object queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17282-17291, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 526, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 526, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 526, + 294, + 579 + ], + "type": "text", + "content": "[10] Chunrui Han, Jinrong Yang, Jianjian Sun, Zheng Ge, Runpei Dong, Hongyu Zhou, Weixin Mao, Yuang Peng, and Xi-angyu Zhang. Exploring recurrent long-term temporal fusion for multi-view 3d perception. IEEE Robotics and Automation Letters, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 581, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 294, + 624 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 625, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 667 + ], + "type": "text", + "content": "[12] Xiuquan Hou, Meiqin Liu, Senlin Zhang, Ping Wei, Badong Chen, and Xuguang Lan. Relation detr: Exploring explicit position relation prior for object detection. arXiv preprint arXiv:2407.11699, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[13] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[14] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3588-3597, 2018. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "type": "text", + "content": "[15] Haotian Hu, Fanyi Wang, Yaonong Wang, Laifeng Hu, Jingwei Xu, and Zhiwang Zhang. Admap: Anti-disturbance framework for reconstructing online vectorized hd map. arXiv preprint arXiv:2401.13172, 2024. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 185, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 227 + ], + "type": "text", + "content": "[16] Georg Klein and David Murray. Parallel tracking and mapping for small ar workspaces. In 2007 6th IEEE and ACM international symposium on mixed and augmented reality, pages 225-234. IEEE, 2007. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 228, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 293 + ], + "type": "text", + "content": "[17] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 294, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 358 + ], + "type": "text", + "content": "[18] Hongyang Li, Chonghao Sima, Jifeng Dai, Wenhai Wang, Lewei Lu, Huijie Wang, Jia Zeng, Zhiqi Li, Jiazhi Yang, Hanming Deng, et al. Delving into the devils of bird's-eye-view perception: A review, evaluation and recipe. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 404 + ], + "type": "text", + "content": "[19] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online hd map construction and evaluation framework. In 2022 International Conference on Robotics and Automation (ICRA), pages 4628-4634. IEEE, 2022. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "type": "text", + "content": "[20] Tianyu Li, Peijin Jia, Bangjun Wang, Li Chen, Kun Jiang, Junchi Yan, and Hongyang Li. Lanesegnet: Map learning with lane segment perception for autonomous driving. arXiv preprint arXiv:2312.16108, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 449, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 503 + ], + "type": "text", + "content": "[21] Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems, 33:21002-21012, 2020. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 505, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 553, + 568 + ], + "type": "text", + "content": "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 571, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 634 + ], + "type": "text", + "content": "[23] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 541-556. Springer, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 689 + ], + "type": "text", + "content": "[24] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized hd map construction. arXiv preprint arXiv:2208.14437, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Bencheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17184" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "type": "text", + "content": "Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv preprint arXiv:2308.05736, 2023.1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "type": "text", + "content": "[26] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[27] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "type": "text", + "content": "[28] Xiaolu Liu, Song Wang, Wentong Li, Ruizi Yang, Junbo Chen, and Jianke Zhu. Mgmap: Mask-guided learning for online vectorized hd map construction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14812-14821, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[29] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized hd map learning. In International Conference on Machine Learning, pages 22352-22369. PMLR, 2023. 1, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 338 + ], + "type": "text", + "content": "[30] Zhi Liu, Shaoyu Chen, Xiaojie Guo, Xinggang Wang, Tianheng Cheng, Hongmei Zhu, Qian Zhang, Wenyu Liu, and Yi Zhang. Vision-based uneven bev representation learning with polar rasterization and surface estimation. In Conference on Robot Learning, pages 437-446. PMLR, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 393 + ], + "type": "text", + "content": "[31] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "text", + "content": "[32] Zihao Liu, Xiaoyu Zhang, Guangwei Liu, Ji Zhao, and Ningyi Xu. Leveraging enhanced queries of point sets for vectorized map construction. In European Conference on Computer Vision, pages 461-477. Springer, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[33] Yuexin Ma, Tai Wang, Xuyang Bai, Huitong Yang, Yuenan Hou, Yaming Wang, Yu Qiao, Ruigang Yang, and Xinge Zhu. Vision-centric bev perception: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 492, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 548 + ], + "type": "text", + "content": "[34] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651-3660, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "text", + "content": "[35] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 657 + ], + "type": "text", + "content": "[36] Cong Pan, Yonghao He, Junran Peng, Qian Zhang, Wei Sui, and Zhaoxiang Zhang. Baeformer: Bi-directional and early interaction transformers for bird's eye view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9590-9599, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[37] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 2, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 137 + ], + "type": "text", + "content": "[38] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[39] Limeng Qiao, Wenjie Ding, Xi Qiu, and Chi Zhang. End-to-end vectorized hd-map construction with piecewiseBezier curve. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13218-13228, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 261 + ], + "type": "text", + "content": "[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "text", + "content": "[41] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11138-11147, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "text", + "content": "[42] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "text", + "content": "[43] Tixiao Shan and Brendan Englot. Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4758–4765. IEEE, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 421, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 463 + ], + "type": "text", + "content": "[44] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 465, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 465, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 465, + 553, + 487 + ], + "type": "text", + "content": "[45] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "type": "text", + "content": "[46] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3621-3631, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "type": "text", + "content": "[47] Shuo Wang, Fan Jia, Yingfei Liu, Yucheng Zhao, Zehui Chen, Tiancai Wang, Chi Zhang, Xiangyu Zhang, and Feng Zhao. Stream query denoising for vectorized hd map construction. arXiv preprint arXiv:2401.09112, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "text", + "content": "[48] Zhenhua Xu, Kenneth KY Wong, and Hengshuang Zhao. Insightmapper: A closer look at inner-instance information for vectorized high-definition mapping. arXiv preprint arXiv:2308.08543, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[49] Tianyuan Yuan, Yicheng Liu, Yue Wang, Yilun Wang, and Hang Zhao. Streammapnet: Streaming mapping network for vectorized online hd map construction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7356-7365, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[50] Gongjie Zhang, Jiahao Lin, Shuang Wu, Zhipeng Luo, Yang Xue, Shijian Lu, Zuoguan Wang, et al. Online map vec" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17185" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 475 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 105 + ], + "type": "text", + "content": "torization for autonomous driving: A rasterization perspective. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[51] Haoyang Zhang, Ying Wang, Feras Dayoub, and Niko Sunderhauf. Varifocalnet: An iou-aware dense object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8514-8523, 2021. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[52] Ji Zhang, Sanjiv Singh, et al. Loam: Lidar odometry and mapping in real-time. In Robotics: Science and systems, pages 1-9. Berkeley, CA, 2014. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "text", + "content": "[53] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. **Reverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743**, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "text", + "content": "[54] Zhixin Zhang, Yiyuan Zhang, Xiaohan Ding, Fusheng Jin, and Xiangyu Yue. Online vectorized hd map construction using geometry. arXiv preprint arXiv:2312.03341, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[55] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13760-13769, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "type": "text", + "content": "[56] Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. Elastic fragments for dense scene reconstruction. In Proceedings of the IEEE International Conference on Computer Vision, pages 473-480, 2013. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 430 + ], + "type": "text", + "content": "[57] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized hd map construction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15396-15406, 2024. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 432, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 294, + 475 + ], + "type": "text", + "content": "[58] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17186" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/a8ba1cd8-3d09-4bb8-8912-43eb50d5dbf1_content_list.json b/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/a8ba1cd8-3d09-4bb8-8912-43eb50d5dbf1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..80b974b5fbc105589e19bdf669d2749a49dd8476 --- /dev/null +++ b/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/a8ba1cd8-3d09-4bb8-8912-43eb50d5dbf1_content_list.json @@ -0,0 +1,1537 @@ +[ + { + "type": "text", + "text": "Interactive Medical Image Analysis with Concept-based Similarity Reasoning", + "text_level": 1, + "bbox": [ + 109, + 130, + 888, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ta Duc Huy† Sen Kim Tran Phan Nguyen Nguyen Hoang Tran Tran Bao Sam", + "bbox": [ + 114, + 179, + 880, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anton van den Hengel† Zhibin Liao† Johan W. Verjans† Minh-Son To†‡", + "bbox": [ + 148, + 204, + 846, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vu Minh Hieu Phan†*", + "bbox": [ + 411, + 229, + 591, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "† Australian Institute for Machine Learning, The University of Adelaide, ‡ Flinders University", + "bbox": [ + 215, + 250, + 774, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 301, + 326, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability to interpret and intervene model decisions is important for the adoption of computer-aided diagnosis methods in clinical workflows. Recent concept-based methods link the model predictions with interpretable concepts and modify their activation scores to interact with the model. However, these concepts are at the image level, which hinders the model from pinpointing the exact patches the concepts are activated. Alternatively, prototype-based methods learn representations from training image patches and compare these with test image patches, using the similarity scores for final class prediction. However, interpreting the underlying concepts of these patches can be challenging and often necessitates post-hoc guesswork. To address this issue, this paper introduces the novel Concept-based Similarity Reasoning network (CSR), which offers (i) patch-level prototype with intrinsic concept interpretation, and (ii) spatial interactivity. First, the proposed CSR provides localized explanation by grounding prototypes of each concept on image regions. Second, our model introduces novel spatial-level interaction, allowing doctors to engage directly with specific image areas, making it an intuitive and transparent tool for medical imaging. CSR improves upon prior state-of-the-art interpretable methods by up to $4.5\\%$ across three biomedical datasets. Our code is released at https://github.com/tadeephuy/InteractCSR.", + "bbox": [ + 89, + 333, + 485, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 742, + 220, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the success of deep neural networks in medical image analysis [8, 9, 14, 22, 24, 29], their \"black-box\" nature limits their adoption in safety-critical clinical settings. Attribution-based methods [3, 30, 31, 34, 38] generate saliency maps to identify influential image regions. However, recent studies raised concerns about the reliability and accuracy of visual interpretation, potentially compro", + "bbox": [ + 89, + 768, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mising decision-making [19, 26]. To address this, Concept Bottleneck Models (CBM) predict the interpretable concepts first and use them to predict the label [10, 11, 21, 36]. CBM allows human intervention in the predicted concepts, which is shown to significantly benefit performance. Alternatively, prototype-based methods are proposed to learn visual part prototypes to explain the classification decision. Specifically, ProtoPNet [2] and its derivatives [1, 6, 17, 18, 27, 28, 33] learn prototypical parts from training data and compare test images with these parts to compute similarity scores for final predictions.", + "bbox": [ + 511, + 301, + 906, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite their inherent interpretability, concept-based methods only offer image-level explanations by linking images to global concepts, without identifying specific local patches correlated with these concepts. In contrast, prototype-based methods deliver patch-level explanations but necessitate post-hoc analysis to correlate visual patches with semantically meaningful concepts. This challenge intensifies in medical imaging, where subtle visual medical findings make concept assignments particularly difficult. How to create concept-based models that enable patch-level explanation for medical imaging is an open-ended question.", + "bbox": [ + 511, + 473, + 908, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In radiological imaging, radiologists routinely reference an atlas of visual examples, depicting distinctive concepts of a pathology, such as opacity, fluid presence in the lung, enlarged heart shadow to aid diagnosis. Inspired by this observation, we propose a novel Concept-based Similarity Reasoning network, which offers two unique contributions. First, our CSR gives patch-level and interpretable-by-design concepts without requiring post-hoc analysis, in contrast to prototype-based methods [2, 18, 28]. We classify a disease by searching locally for the presence of visual exemplars of its concepts on the image. For example, a cancer diagnosis on chest X-ray could involve a radiologist comparing a suspected tumor with other known malignant tumors to assess its severity [7]. The presence of the tumor is a concept to predict the target cancer. Our framework trains a model to search locally for the presence of similar tumor exemplars (or prototypes) on every patch of the im", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30797", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b2f63d2c0aae5676106a1db2668e72ae4e62048251fb4b657029d5a9d665224a.jpg", + "image_caption": [ + "Figure 1. Diagnosis Workflow with Doctor-in-the-Loop using CSR. Top section: CSR predicts Pulmonary Edema, falsely associating with the two concepts lung fluid and enlarged heart. For each concept, CSR explains its prediction by comparing the input with the highlighted region on the prototype image to create the corresponding similarity map. Bottom section: As the doctor inspects the similarity maps of each concept, he suppresses incorrect lung fluid attention, and reinforces the attention on opacity on the left of the image through spatial-interaction: “drawing positive and negative boxes” to create an importance map indicating where to focus and ignore. Secondly, noting the heart is normal, he “rejects” the enlarged heart concept via concept-interaction. CSR then recalibrates its prediction to Tuberculosis, aligning with the observed opacity." + ], + "image_footnote": [], + "bbox": [ + 135, + 89, + 859, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "age. After training, our model explicitly learns interpretable concept-grounded patch prototypes without a post-hoc analysis.", + "bbox": [ + 88, + 497, + 480, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, our model enables spatial interactivity on the patch level. Unlike previous methods which only intervene in model predictions on the concepts-level outputs [1, 6, 16, 35], we introduce spatial interactivity on the input image. Ours offers intuitive benefits in medical imaging, where precise image analysis is crucial for clinicians. Deep learning models frequently capture spurious shortcuts correlations rather than the intended signals—a phenomenon known as the Clever-Hans effect [12]. Therefore, integrating doctor feedback to guide the model on relevant cues significantly enhances trustworthiness in clinical settings. Our contributions:", + "bbox": [ + 88, + 547, + 482, + 728 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel Concept-based Similarity Reasoning network (CSR) that provides concept-level explanation. To the best of our knowledge, this is the first interpretable-by-design method that offers localized concept explainability for medical images without post-hoc analysis like part-prototype approaches.", + "- We introduce a novel spatial interaction mechanism (where to look), alongside train-time interaction (what not to learn) and concept-level interaction (what not present). CSR facilitates a comprehensive interactive framework that actively involves doctors in both train and test phases." + ], + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We conducted extensive experiments on three datasets and showed that CSR achieves a significant improvement by up to $4.5\\%$ F1-score compared to other interpretable methods.", + "bbox": [ + 513, + 497, + 906, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Method", + "text_level": 1, + "bbox": [ + 513, + 570, + 604, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CSR comprises (i) a Concept model to extract interpretable concept features; (ii) a feature projector $P$ to enhance the concept feature space; and (iii) a task head $H$ to classify based on concept similarity scores. This section describes the inference logic of the proposed Concept-based Similarity Reasoning model and the prototypes learning framework. Fig. 2 illustrates the inference process.", + "bbox": [ + 511, + 595, + 906, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Concept-based Similarity Reasoning", + "text_level": 1, + "bbox": [ + 511, + 710, + 830, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concept-based classification. From an input $x$ , concept-based methods first predict the scores of $K$ explainable concepts $\\{c^k\\}_{k=1}^K$ and then use them as support to predict the target $y$ . Instead of predicting the concept scores, CSR calculates the similarity scores between $x$ and the examples (or prototypes) of each concept. It then uses the concept similarity scores to support the prediction of $y$ .", + "bbox": [ + 511, + 732, + 905, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Similarity reasoning. To construct the similarity reasoning inference logic, we establish an atlas: a set of $M \\cdot K$ concept prototypes $\\{p^{k_m}\\}_{k=1,m=1}^{K,M}$ , where we have $M$ prototypes for each concept $c^k$ . In essence, $p^{k_m} \\in \\mathbb{R}^{\\mathcal{C}}$ repre", + "bbox": [ + 511, + 838, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30798", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2d9f88b37c24908d874e5d22ff8902cd6e558d93a82ce4c0dd4150fb4c60d914.jpg", + "image_caption": [ + "Figure 2. Inference logic of CSR. Each concept prototype represents a specific concept from a training image. CSR generates 2D similarity maps by computing the cosine similarity between these concept prototypes and the feature maps, and considers the maximum values as similarity scores to calculate the prediction logits. The prototype image refers to the training image associated with a specific concept prototype as detailed in Sec. 3.1" + ], + "image_footnote": [], + "bbox": [ + 140, + 89, + 867, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sents the $m^{\\mathrm{th}}$ example of concept $k$ . Given an input image $x$ , we use the feature extractor $F$ to create a feature map $\\mathbf{f} \\in \\mathbb{R}^{C \\times H \\times W}$ having $\\mathcal{H} \\cdot \\mathcal{W}$ feature patches. The similarity score $s^{k_m}$ is defined as the maximum cosine similarity between the prototype $p^{k_m}$ and each feature patch $\\mathbf{f}(h, w)$ , quantifying the presence of $p^{k_m}$ on $x$ . As such, we can aggregate the presence of $M \\cdot K$ concept prototypes $\\{p^{k_m}\\}$ into a similarity scores vector:", + "bbox": [ + 89, + 352, + 483, + 474 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns = \\left[ s ^ {k _ {m}} \\right] _ {k = 1, m = 1} ^ {K, M} = \\left[ s ^ {1 _ {1}}, s ^ {1 _ {2}}, \\dots , s ^ {K _ {M - 1}}, s ^ {K _ {M}} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 484, + 482, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, the model predicts the target $y$ from the similarity vector $s$ with a task head $H$ : $y = H(s)$ . For example, the input chest X-ray is similar to several images from the atlas that have: fluid build-up in the left lung (score 0.85), and the abnormal heart size (score 0.88). Based on these similarities, the model predicts target class Pulmonary Edema. In short, the model classifies $y$ by comparing exemplars $p^{k_m}$ of the interpretable concept $c^k$ on each local patch on the image $x$ . The choice of $H$ should be but not be limited to a whitebox model such as a single linear layer or a shallow tree so that the reasoning from $s \\to y$ maintains interpretability.", + "bbox": [ + 89, + 517, + 483, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discussion. Prototype-based methods [2, 18, 27, 28] also learn a set of exemplar prototypes $p$ . However, they require post-hoc analysis by experts to associate the interpretable concept labels to the learned prototypes, which is intractable in the fine-grained medical image analysis. In stark contrast, the proposed model learns patch-level and interpretable concept prototypes by design.", + "bbox": [ + 89, + 699, + 483, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Concept prototypes learning", + "text_level": 1, + "bbox": [ + 89, + 816, + 346, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section describes the proposed methods to learn the concept prototypes. Overall, our method consists of a Concept model to learn interpretable concept vectors, and a novel contrastive learning objective to enhance generaliz", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ability of concept vectors. The overview of our learning framework is shown in Fig. 3.", + "bbox": [ + 511, + 352, + 903, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concept model. The Concept model comprises the feature extractor $F$ , and the concept head $C$ that maps $\\mathbf{f}$ into the one hot concept vector $c \\in \\mathbb{R}^K$ such that $c = C(\\mathbf{f})$ , indicating the presence of concepts being associated with $x$ . The concept head $C$ includes a $1 \\times 1$ convolution layer that maps $\\mathbf{f}$ to $K$ channels: $\\mathbb{R}^{C \\times \\mathcal{H} \\times \\mathcal{W}} \\to \\mathbb{R}^{K \\times \\mathcal{H} \\times \\mathcal{W}}$ , representing the class activation maps $\\{\\mathrm{cam}^k \\in \\mathbb{R}^{\\mathcal{H} \\times \\mathcal{W}}\\}_{k=1}^K$ for $K$ concepts.", + "bbox": [ + 511, + 383, + 905, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Local concept vectors acquisition. If the region $(h, w)$ on $x$ contains the visual characteristics of concept $k$ , we consider the patch feature $\\mathbf{f}(h, w)$ important and select it as the local concept vector. We first normalize $\\mathrm{cam}^k$ using spatial softmax. Then, we soft-select the local concept vector $v^k \\in \\mathbb{R}^C$ for concept $c^k$ by weighting $\\mathbf{f}$ with $\\mathrm{cam}^k$ and summing the products spatially:", + "bbox": [ + 511, + 505, + 905, + 612 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nv ^ {k} = \\sum_ {h, w} ^ {\\mathcal {H}, \\mathcal {W}} \\operatorname {s o f t m a x} _ {h, w} \\left(\\operatorname {c a m} ^ {k}\\right) \\cdot \\mathbf {f}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 623, + 903, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With a training set of $\\{x_{i}\\}^{N}$ images of $K$ concepts, we generate a set $\\{v_{i}^{k}\\}_{i = 1}^{V\\leq N\\cdot K}$ of local concept vectors.", + "bbox": [ + 511, + 674, + 903, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarity maps as explanations. Based on the local concept vector, we construct a 2D similarity map for concept $c^k$ on $x$ by evaluating the cosine similarity between each patch feature $\\mathbf{f}(h,w)$ and vector $v^k$ :", + "bbox": [ + 511, + 705, + 905, + 767 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {k} (h, w) = \\cos \\left(v ^ {k}, \\mathbf {f} (h, w)\\right) = \\langle v ^ {k}, \\mathbf {f} (h, w) \\rangle , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 777, + 903, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\langle ,\\rangle$ is the dot product of the two normalized vectors. As $v^{k}$ is the importance weighted sum of $\\mathbf{f}$ with respect to $c^k$ , comparing it with the important patch $\\mathbf{f}(h,w)$ results in a high similarity score at $S^k (h,w)$ , and vice versa. Effectively, $S^k\\in \\mathbb{R}^{\\mathcal{H}\\times \\mathcal{W}}$ is the local explanation for concept $c^k$ on $x$ . The similarity score of $x$ with the concept $c^k$ is", + "bbox": [ + 511, + 809, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30799", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b4c33f0a7af0b7a5a8c6a0fc97c9e30b4121993928f6dea4a95e82b4f3718b65.jpg", + "image_caption": [ + "Figure 3. The novel Concept prototypes learning framework. We pretrain a Concept model to generate concept activation maps. The local concept vectors are then generated by weighting and summing the feature maps with the activation map of the corresponding concept. To enhance the compactness and the generalizability of the concept feature space, we introduce a novel multi-prototype learning objective. After projecting concept vectors via a projector $P$ , the proposed objective pulls concept features to its nearest concept prototypes, while pushing away from prototypes of other concepts." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "defined as the maximum of the similarity map $S^k$ :", + "bbox": [ + 89, + 438, + 423, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns ^ {k} = \\max _ {h, w} S ^ {k} (h, w). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 459, + 480, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the similarity scores $s_i^k$ for the local concept vectors $\\{v_i^k\\}$ with the input $x$ , one can naively adopt the inference logic in Sec. 2.1. However, this leads to two problems. First, the computation complexity is $O(N \\cdot K)$ . Second, $v_i^k$ is local, which is specific to each image, thus non-generalizable to new images. We investigate the feasibility of this approach by comparing $v_i^k$ with new images $v_{j \\neq i}^{k'}$ and analyze the similarity scores distribution of intra-concept $(k = k')$ and inter-concept $(k \\neq k')$ . It is observed that the concepts are not well-represented in new images. In other words, a local concept vector $v_i^k$ generated from an image $x_i$ fails to highlight the concept $k$ on a new image $x_{j \\neq i}$ , as shown in Fig. 4.", + "bbox": [ + 89, + 491, + 483, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept prototypes learning. To improve the generalizability of concept representation $v$ and reduce the number of concept vectors, our solution involves: (i) employing contrastive learning to enhance the shared embedding space between $v$ and $\\mathbf{f}$ ; (ii) downsampling the concept-specific-and-sample-specific set $\\{v_i^k\\}$ to a small core set of prototypes $\\{p^k\\}$ that is only specific to each concept, for efficiency.", + "bbox": [ + 89, + 688, + 483, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use a projector $P$ to map $v$ to a feature space with the goal to improve inter-concept separation and intra-concept compactness such that $v' = P(v)$ via contrastive learning. Then, we mine prototypes $p^k$ for each concept, which can be a single prototype or a diverse set of prototypes.", + "bbox": [ + 89, + 795, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Single prototype per concept. We initialize a prototype vector $p^k \\in \\mathbb{R}^C$ for each concept $c^k$ . [23] shows that the", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cross-entropy of sharpened softmax logits is equivalent to a smoothed Triplet Loss [5] in a unit sphere. By minimizing the maximal entropy, we can pull $v_{i}^{\\prime \\tilde{k}}$ to their respective prototype $p^{\\tilde{k}}$ and push it away from irrelevant prototypes $k \\neq \\tilde{k}$ . Accordingly, we learn the prototype vector $p^{\\tilde{k}}$ by minimizing the term:", + "bbox": [ + 511, + 438, + 906, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathrm {c o n}} \\left(v _ {i} ^ {\\tilde {k}}\\right) = - \\log \\left(\\frac {\\exp \\left(\\lambda \\left\\langle p ^ {\\tilde {k}} , v _ {i} ^ {\\prime \\tilde {k}} \\right\\rangle\\right)}{\\sum_ {k \\in K} \\exp \\left(\\lambda \\left\\langle p ^ {k} , v _ {i} ^ {\\prime \\tilde {k}} \\right\\rangle\\right)}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 542, + 906, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda > 1$ is the scaling factor to sharpen the softmax logits, $p$ and $v'$ are $L_2$ normalized.", + "bbox": [ + 511, + 590, + 905, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multiple prototypes per concept. Due to the multi-modal nature of data where one class contains several clusters rather than just a single one, we initialize $M$ prototypes per concept $\\{p^{k_m}\\}_{m=1}^M$ such that $M \\ll N$ .", + "bbox": [ + 511, + 621, + 905, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The similarity of $v_{i}^{\\prime k}$ with its concept $c^{\\tilde{k}}$ is defined as the similarity to the nearest $p^{\\tilde{k}_m}$ , and can be approximated by summing the compatibility with the $M$ clusters prototypes $p^{\\tilde{k}_m}$ . Specifically, we first compute the probabilistic assignment $\\mathbf{q}_m^{\\tilde{k}}(v_i^{\\prime k})$ to the prototypes:", + "bbox": [ + 511, + 683, + 905, + 767 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {q} _ {m} \\left(v ^ {\\prime} _ {i} ^ {\\tilde {k}}\\right) = \\operatorname {s o f t m a x} _ {m} \\left(\\gamma \\left\\langle p ^ {\\tilde {k} _ {m}}, v ^ {\\prime} _ {i} ^ {\\tilde {k}} \\right\\rangle\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 773, + 903, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{q}_m$ is the assignment distribution of the $M$ prototypes in $c^{\\tilde{k}}$ and $\\gamma > 1$ is the scaling factor. The similarity of the concept vector with each prototype is obtained by weighing their cosine similarity with the assignment:", + "bbox": [ + 511, + 806, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sin^ {\\tilde {k} _ {m}} \\left(v ^ {\\prime} _ {i} ^ {\\tilde {k}}\\right) = q _ {m} \\left(v ^ {\\prime} _ {i} ^ {\\tilde {k}}\\right) \\cdot \\langle p ^ {\\tilde {k} _ {m}}, v ^ {\\prime} _ {i} ^ {\\tilde {k}} \\rangle . \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 875, + 903, + 897 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30800", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d7c77481e4c51259d549d7bbc9ebb56ba373de972fe2cbbd39ca51df624ebce4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 125, + 89, + 450, + 250 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2591fad17d9119258290398644c4790e45d2af2858d85b853a4982e1412096f9.jpg", + "image_caption": [ + "(a) Similarity maps generated by the local concept vector (first row) and concept prototype (second row) on the same image (a) and new images (b), (c). The local concept vector fails to highlight the concept pacemaker on new images while the concept prototype generalizes well with better localization.", + "Figure 4. Contrastive learning improves the concept feature space. (a) Qualitative and (b) quantitative comparison." + ], + "image_footnote": [], + "bbox": [ + 125, + 330, + 272, + 507 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/70db739c08f16eec8b83a0697520ae4fde48eff76476b472dd84a8b6f57e64c6.jpg", + "image_caption": [ + "(b) Comparison between the original $v$ and projected feature space $v^{\\prime}$ . The box plots of the similarity score distribution between $v_{i}^{k}$ and $v_{j\\neq i}^{k^{\\prime}}$ . Intra-Concept: $k = k^{\\prime}$ , Inter-Concept: $k\\neq k^{\\prime}$ . The projected feature space is more compact and has better inter-concept separation." + ], + "image_footnote": [], + "bbox": [ + 292, + 330, + 444, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After that, we can obtain the similarity with concept $c^{\\tilde{k}}$ by summing the compatibility with each cluster prototype:", + "bbox": [ + 89, + 641, + 482, + 675 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sin^ {\\tilde {k}} \\left(v _ {i} ^ {\\prime} \\tilde {k}\\right) = \\sum_ {m} ^ {M} \\sin^ {\\tilde {k} _ {m}} \\left(v _ {i} ^ {\\prime} \\tilde {k}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 685, + 482, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we compute the contrastive loss by maximizing the similarity of a concept vector with the cluster prototypes of the positive concept, similar to Eqn. 5:", + "bbox": [ + 89, + 736, + 483, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {c o n - m}} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right) = - \\log \\left(\\frac {\\exp \\left(\\lambda \\left(\\sin^ {\\tilde {k}} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right) + \\delta\\right)\\right)}{\\sum_ {k \\in K} \\exp \\left(\\lambda \\sin^ {k} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right)\\right)}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 792, + 482, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {i f} \\quad k = \\tilde {k}, \\quad \\operatorname {s i m} ^ {k} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right) = \\operatorname {s i m} ^ {k} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right) + \\delta .\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 837, + 419, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A small margin $\\delta$ between concept prototypes is added to further expand the decision boundaries. As a result, we", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "learn a set of concept prototypes $\\{p^{k_m}\\}^{M\\cdot K}$ that is significantly smaller in numbers compared to $\\{v_i^k\\}^{N\\cdot K}$ , and the projector $P$ that enhances the inter-concept separation and intra-concept compactness. We illustrate the learning process of $P$ and $\\{p\\}$ in Fig. 3.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From Eqn. 2, $v_{i}^{k}$ is a soft-selected feature patch of concept $c^k$ from the original feature $\\mathbf{f}_i$ , we can directly use $P$ to map feature patch $\\mathbf{f}_i(h,w)$ to the compact and discriminative prototypical feature space. We then obtain the similarity of the concept prototypes with $x$ from Eqn. 3:", + "bbox": [ + 511, + 167, + 903, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns ^ {k _ {m}} = \\max _ {h, w} S ^ {k _ {m}} (h, w) = \\max _ {h, w} \\langle p ^ {k _ {m}}, P (\\mathbf {f} (h, w)) \\rangle , \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 251, + 903, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which are aggregated to create the similarity vector $[s^{k_m}]$ in Eqn. 1. We quantitatively and qualitatively present the improved representation of $p$ in Fig. 4.", + "bbox": [ + 511, + 286, + 903, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training details. Initially, we train the Concept model for multi-label classification using binary cross-entropy, predicting one-hot encoded vector of ground truth concepts. Then, we obtain the local concept vectors and learn the concept prototypes and the projection head with a contrastive loss $\\ell_{\\mathrm{con - m}}$ to enhance concept generalization. Following, the task head is trained to predict the target from similarity scores with the concept prototypes, using cross-entropy.", + "bbox": [ + 511, + 333, + 905, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Doctor-in-the-loop interactivity", + "text_level": 1, + "bbox": [ + 511, + 465, + 799, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Model explanations", + "text_level": 1, + "bbox": [ + 511, + 491, + 700, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given an input image $x$ , the model returns output $\\hat{y}$ and the explanations for its prediction that include:", + "bbox": [ + 511, + 513, + 903, + 542 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The similarity maps $[S^{k_m}]$ to show the activated region of the concept prototype $p^{k_m}$ on $x$ .", + "- The corresponding prototype image $\\mathcal{I}(p^{k_m})$ as a reference, also highlighting the activated regions of $p^{k_m}$ .", + "- The concept similarity scores $[s^{k_m}]$ , indicating the similarity of the two activated regions of $p^{k_m}$ on the input image $x$ and on the reference image $\\mathcal{I}(p^{k_m})$ ." + ], + "bbox": [ + 511, + 542, + 903, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The model explanations correspond to the similarity-based inference logic outlined in Sec. 2.1. The operator $\\mathcal{I}$ retrieves the image associated with the prototype $p^{k_m}$ . We replace each $p^{k_m}$ with the nearest projected local concept vector $v_{i}^{k}$ and link it to its source image $x_{i}$ :", + "bbox": [ + 511, + 648, + 905, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} \\left(p ^ {k _ {m}}\\right) = x \\leftarrow \\underset {i} {\\operatorname {a r g m i n}} \\left\\langle v _ {i} ^ {\\prime k}, p ^ {k _ {m}} \\right\\rangle . \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 736, + 903, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Train-time interaction", + "text_level": 1, + "bbox": [ + 511, + 773, + 720, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After we learn the concept prototypes $\\{p^{k_m}\\}$ and its associated human-interpretable version $\\{\\mathcal{I}(p^{k_m})\\}$ , or atlas, doctors can inspect each concept prototype and discard any unqualified examples. This process can reveal and mitigate the Clever-Hans effect in the Concept model by examining the clues it is taking to predict the concepts. If a given prototype is shown to be a shortcut rather than a meaningful signal, it", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30801", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/743ff4dafff8b227b2e4359779ac766a71eca65afb134da1de3f56c6289d2b80.jpg", + "image_caption": [ + "Figure 5. The irrelevant concept prototypes to be discarded in train-time interaction, highlighted at the activated regions." + ], + "image_footnote": [], + "bbox": [ + 112, + 90, + 460, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is removed from $\\{p^{k_m}\\}$ and is not used to calculate the similarity scores $[s^{k_m}]$ for the prediction of $y$ , effectively nullify the Clever-Hans effect. We show the irrelevant examples of different concepts in Fig. 5. The similarity-based inference logic allows this interaction as we only need as much as the remaining relevant examples to predict the target $y$ .", + "bbox": [ + 89, + 344, + 483, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Test-time interaction", + "text_level": 1, + "bbox": [ + 89, + 448, + 287, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At test time, we introduce two types of interaction to improve the prediction quality: Spatial-level interaction and Concept-level interaction. We illustrate this in Fig. 1.", + "bbox": [ + 89, + 470, + 482, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Concept-level interaction. Consider this scenario: While reviewing the explanations from CSR for a chest X-ray $x$ , the doctor notices that the model is considering a nodule on $x$ for a mass in the reference image with a high similarity score. Recognizing the incorrect comparison, the doctor informs the model that concept mass is absent in $x$ . The model suppresses its confidence in concept mass to calibrate its prediction. This concept-level interaction allows doctors to directly eliminate irrelevant concepts.", + "bbox": [ + 89, + 517, + 482, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interactivity mechanism. When the doctor goes through $[s^{k_m}]$ , she can discard any invalid concept if presented. The model then sets $s^{k_m} = 0$ for those rejected concepts, effectively discounting them on the prediction $\\hat{y}$ . Concept-level interaction only considers the negative type of input from the doctor: concept rejection. Setting a suitable value for every $s_m^k$ of $c^k$ would be a overburden. Hence, for each concept $c^k$ , we only present the highest similarity score $\\max_{m} \\{s^{k_m}\\}$ and have the doctor recognize its presence. We zero out all of the similarity scores of the rejected concept.", + "bbox": [ + 89, + 654, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial-level interaction. As the doctor examines the image $x$ , upon spotting an important area, she can guide the model's attention with bounding boxes. The model then recalibrates its prediction $\\hat{y}$ by prioritizing the indicated region based on this feedback. Conversely, reviewing the activation maps $[S^{k_m}]$ , the doctor also notices the model con", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "centrating on an irrelevant region for prediction. To counteract the Clever-Hans effect, she marks the area as spurious, prompting the model to disregard it in its updated prediction. Hence, the doctor only needs to direct the model where to focus, without specifying what is in that region. This interaction scheme is practical and avoids the doctor from inadvertently becoming a teacher rather than taking advantage of the model.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interactivity mechanism. As we described in the user journey above, the spatial-level interaction allows two types of input from the doctor: (i) the positive boxes $\\{\\mathsf{bb}^{+} = (\\mathsf{x}_1,\\mathsf{y}_1,\\mathsf{x}_2,\\mathsf{y}_2)\\}$ for the focus areas, (ii) the negative boxes $\\{\\mathsf{bb}^{-}\\}$ for the regions to ignore. To emphasize the area specified by $\\{\\mathsf{bb}^{+}\\}$ and suppress the area specified by $\\{\\mathsf{bb}^{-}\\}$ , the model reweights each element $S(h,w)$ with an importance map $A\\in \\mathbb{R}^{\\mathcal{H}\\times \\mathcal{W}}$ :", + "bbox": [ + 511, + 215, + 906, + 335 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA (h, w) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} (h, w) \\in \\left\\{\\mathrm {b b} ^ {+} \\right\\} \\\\ 0 & \\text {i f} (h, w) \\in \\left\\{\\mathrm {b b} ^ {-} \\right\\}, \\\\ \\alpha & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 356, + 903, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha \\in [0,1)$ is the neutral weight for the other regions. Setting $\\alpha = 0$ indicates that the doctor only considers the regions $\\{\\mathrm{bb}^{+}\\}$ , while setting $\\alpha \\rightarrow 1$ diminishes the importance of $\\{\\mathrm{bb}^{+}\\}$ . The new set of similarity maps $[\\hat{S}^{k_m}]$ is obtained by multiplying every similarity map of $[S^{k_m}]$ with the importance map $A$ :", + "bbox": [ + 511, + 431, + 903, + 523 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n[ \\hat {S} ] = A \\odot [ S ] = [ A \\odot S ^ {1 _ {1}}, \\dots , A \\odot S ^ {K _ {M}} ], \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 541, + 903, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\odot$ denotes the element-wise multiplication on the spatial dimension. As CSR is designed to select the maximum element on $S^{k_m}$ as $s^{k_m}$ , giving more weight to the important region $S(h,w)$ increases its likelihood to be selected. On the other hand, zeroing out the spurious region nullifies its influence on the similarity vector. Because $S$ is bounded within $[-1,1]$ and $A$ is bounded within $[0,1]$ , we clip negative values of $S$ to condition the model on only non-negative values. This ensures that the $\\{\\mathrm{bb}^+\\}$ and $\\{\\mathrm{bb}^-\\}$ regions are monotonically adjusted. The new prediction is made with $\\hat{S}$ by the same inference logic as described in Sec. 2.1.", + "bbox": [ + 511, + 579, + 905, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion. Combining both spatial and concept interaction mechanisms enables a flexible interaction approach. For example, if the doctor identifies an important region, she can draw a positive box $\\mathsf{bb}^{+}$ over it. However, if the model associates the feature within $\\mathsf{bb}^{+}$ with a wrong concept $c^k$ , placing $\\mathsf{bb}^{+}$ over it would amplify the error. Because $c^k$ is absent in $x$ , rejecting it while specifying $\\mathsf{bb}^{+}$ will instruct the model to focus on other concepts $\\{c^{j\\neq k}\\}$ within $\\mathsf{bb}^{+}$ and ignore $c^k$ .", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30802", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Related works", + "text_level": 1, + "bbox": [ + 89, + 89, + 236, + 104 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Concept-based. Concept Bottleneck Model (CBM) [11] predicts the concepts before using them as support for the target prediction. Post-hoc Concept Bottleneck Models (PHCBM) [16] transfer concepts from other datasets or from natural language descriptions of concepts via multimodal models like CLIP [25]. Yang et al. leverages Large Language Models to generate an exhaustive list of concepts for each class [36]. These multimodal concept-based approaches [15, 22, 36] utilize foundation models to enhance their capabilities. However, foundation models are generally trained on broad domains, which presents limitations when applying them directly to medical imaging. In this work, we focus on the research direction of vision features and explore multi-modality learning in future studies.", + "bbox": [ + 89, + 114, + 480, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Part-prototypes. ProtoPNet [2] used prototypical parts to make a final classification based on positive attributions. ProtoTree[17] is a globally interpretable model by design that uses prototypes with decision trees. ProtoPool [28] is similar to ProtoPNet with a tweak of the shared prototypes pool and the Gumbel-softmax trick for the differentiable property. PIP-Net[18] is a patch-based approach with a novel regularization method for learning prototype similarity that correlates with human visual perception. Unlike previous methods that correlate prototypes with target classes, CSR explicitly grounds prototypes on interpretable concepts, making them semantically relevant.", + "bbox": [ + 89, + 328, + 482, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interactivity. CBM [11] enables local interaction as human adjusts the concept logits for individual samples, while [16] allows global interaction by modifying concept weights of the model. ProtoPDebug [1] and [6] focus on part-prototype interaction, refining prototypes for enhanced performance. Besides logits-level interaction like those methods, CSR introduces spatial-level interaction, which is particularly intuitive and beneficial in the medical imaging domain, where detailed image analysis is of significant interest to doctors.", + "bbox": [ + 89, + 508, + 482, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments and Results", + "text_level": 1, + "bbox": [ + 89, + 662, + 326, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this paper, we experiment on two chest X-ray datasets: TBX11K [13], VinDr-CXR[20] and a skin imaging dataset ISIC[4]. From those datasets, we define the findings annotations as the concepts and the diseases annotations as the target classes. Details on datasets and the training hyperparameters are in the supplementary. We qualitatively and quantitatively demonstrate the effectiveness of the concept prototypes learning framework in Fig. 4.", + "bbox": [ + 89, + 688, + 482, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diagnostic performance. Due to the heavy class imbalance, we use macro F1-score as the metric for our experiments. We report the results in Table 1. It can be seen that CSR demonstrates a strong diagnostic performance compared to other baselines across all dataset. We also benchmark the performance of CSR when we ask a doctor to re", + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/64bc75d6bb3c1b176dfe4c817fc34cb20c45c8dad58cd6162be0cec92f876ca2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodF1↑#pro.↓#exp.↓
TBX11KCBM (ind.)24.514
CBM (seq.)60.514
CBM (joint.)88.614
ProtoPNet89.3600600
ProtoPNet89.315001500
ProtoPNet94.130003000
ProtoPool90.7600600
ProtoTree(H=6)94.01276
ProtoTree(H=9)93.75119
PIP-Net94.0768158
CSR94.4140014
CSR (refined)94.0138614
VinDr-CXRCBM (joint.)50.114
ProtoPNet41.415001500
PIP-Net45.17689
CSR54.6140014
ISICCBM (joint.)45.54
ProtoPNet38.5600600
ProtoPool31.0200200
ProtoTree(H=6)66.11276
PIP-Net69.976890
CSR71.54004
", + "bbox": [ + 563, + 89, + 854, + 474 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3afb5271008e0af672f29776104c4e304e1ce9f241ee94739d598d95d1f6ca37.jpg", + "table_caption": [ + "Table 1. Diagnostic performance on the medical datasets, measured by Macro F1-score. CSR (refined) has the doctors review the concept atlas. #pro. denotes the number of prototypes of the model. #exp. indicates the explanation size for each prediction." + ], + "table_footnote": [], + "table_body": "
MethodProtoP-NetProto TreePIP-NetCBMCSRCSR (refined)
% PG hit8.87.819.555.160.979.5
", + "bbox": [ + 514, + 558, + 903, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Pointing Game hit rate on TBX11K, Tuberculosis class. CBM uses the class activation map of the highest contributing product to the class Tuberculosis. ProtoTree uses a similarity map from the top prototype at level-1 of the tree. ProtoPNet, PIP-Net, and CSR use the similarity map from the prototype with the highest contribution to class Tuberculosis.", + "bbox": [ + 511, + 619, + 906, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fine the learned concept atlas as described in Sec. 3.2.", + "bbox": [ + 511, + 732, + 867, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Explanation size. Following [18], we quantify the explanation size as the number of explanations per prediction for each method. For part-prototype methods, explanations refer to the similar parts of the image from the training set. ProtoTree has the explanation size equal to its depth. PIP-Net reduces the explanation size by regularizing the sparsity during training and only considers non-zero prototypes for each prediction. Other part-prototype approaches have the explanation size matching the large number of prototypes. This compromises interpretability since users must", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30803", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/300a2b228ea54334149c3dfca4932232d3bbcf631d2b48f046e886f4132ba99c.jpg", + "image_caption": [ + "Figure 6. Qualitative result of the Pointing Game. We visualize the groundtruth bounding box and the maximum activation point of each method as the colored square. :Proposed CSR, :CBM, :PIP-Net, :ProtoPNet, :ProtoTree" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 478, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "navigate through a multitude of explanations within the set for each prediction. Concept-based models use the concepts as their explanations, having an explanation size equal to the number of concepts, which is typically smaller. This choice involves an accuracy trade-off due to the reduced size of the bottleneck. In contrast, the proposed CSR benefits from both, featuring a small explanation size equivalent to CBM while achieving competitive accuracy akin to part-prototype methods. To achieve this, our CSR only considers concept grounded explanations and exclusively presents the highest similar prototype of each active concept.", + "bbox": [ + 88, + 335, + 483, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Trustworthiness. We inspect the explanations of each benchmark and demonstrate that interpretability does not always come with trustworthiness. A trustworthy model must deliver relevant explanations that correspond to human interpretable concepts. We leverage the pointing game (PG) [37] that considers a hit if the point of maximum value in the explanation lies within a segmentation ground truth. We use the bounding box annotation from the TBX11K and calculate the PG hit rate of the explanations.", + "bbox": [ + 89, + 505, + 483, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The quantitative results are reported in Table. 2. Partprototype methods including ProtoPNet and ProtoTree are relying on clues that considerably mismatch with the semantic ground truth. This observation aligns with [18], where Purity of the prototypes is evaluated as an alignment with the part annotations in the CUB dataset [32], demonstrating a large semantic gap between the prototypes and human-relevant concepts. Concept-based methods attain a more reasonable PG hit rate as the concepts are human-grounded and thus align well with the semantic ground truth. The proposed CSR faithfully aligns with human reasoning and achieves a remarkable PG hit rate of $60.9\\%$ . Refining the concept atlas further improves the PG hit rate to $79.5\\%$ , which hints that the better classification result of the original CSR in Table 1 could stem from shortcuts. The qualitative results in Fig. 6 shows that our CSR:□ more frequently identifies the maximum activation within", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fb050c105a07bce560116e94fa1559b93fed4ba6809d461bc23bbaa4bac7883f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
spatialconceptMac.F1+Δ%
CSR94.40.0
Non-expert✓94.50.1
Oracle✓94.90.5
M.D.✓94.70.3
M.D.✓95.10.7
M.D.✓✓94.90.5
", + "bbox": [ + 563, + 88, + 857, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Interactions performance gain. spatial: spatial-level interaction, concept: concept-level interaction. We limit the Non-export to spatial-interaction only. Oracle: using the bounding box ground truth of TB as spatial-interaction. M.D: medical student.", + "bbox": [ + 511, + 215, + 906, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the ground truth box compared to other methods.", + "bbox": [ + 511, + 314, + 836, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Interactivity. We demonstrate the benefits of interactivity by measuring the model when obtaining interaction at test time on the TBX11K dataset and report the result in Table. 3. Our experiment involves a Medical degree (M.D.) student and a non-expert in evaluating the framework. In the oracle setting, we use the ground truth bounding box of class TB from the dataset as a positive spatial-level interaction. Note that we only require human interaction when the output probability of the model is indecisive: when the difference of top-1 and top-2 class probability is less than 0.3. The weight $\\alpha$ is set to 0.2. It shows that human interactions has a positive impact on the model outcome. The interaction from the M.D. student is more beneficial to the model than the non-expert. We consider that the non-experts can still eliminate irrelevant CXR regions without formal radiology expertise, and limit the non-expert to spatial-level interactions only. The concept-level interaction is more influential than the spatial-level interaction with better gain.", + "bbox": [ + 511, + 335, + 906, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 645, + 633, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces a novel Concept Similarity Reasoning (CSR), an interpretable and interactive model for medical image analysis with similarity reasoning. CSR provides local explanations by presenting similar concept examples with the input images. In comparison to part-prototype approaches, our CSR presents more concise and meaningful prototypes that align better with human-relevant concepts. This observation implies the differentiation between interpretability and trustworthiness. The proposed interaction mechanism employed during both training and testing effectively enhances model reliability by mitigating shortcuts and improves model accuracy through domain-specific inputs. Our method is an attempt to encourage more trust from doctors through a natural and effective human-AI interaction, improving both accuracy and transparency.", + "bbox": [ + 511, + 676, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30804", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andrea Bontempelli, Stefano Teso, Fausto Giunchiglia, and Andrea Passerini. Concept-level debugging of part-prototype networks. ArXiv, abs/2205.15769, 2022. 1, 2, 7", + "[2] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: Deep learning for interpretable image recognition. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 1, 3, 7", + "[3] Townim Faisal Chowdhury, Kewen Liao, Vu Minh Hieu Phan, Minh-Son To, Yutong Xie, Kevin Hung, David Ross, Anton van den Hengel, Johan W Verjans, and Zhibin Liao. Cape: Cam as a probabilistic ensemble for enhanced dnn interpretation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11072-11081, 2024. 1", + "[4] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), pages 168-172. IEEE, 2018. 7", + "[5] Xingping Dong and Jianbing Shen. Triplet loss in siamese network for object tracking. In Proceedings of the European conference on computer vision (ECCV), pages 459-474, 2018. 4", + "[6] Michael Gerstenberger, Thomas Wiegand, Peter Eisert, and Sebastian Bosse. But that's not why: Inference adjustment by interactive prototype revision. In Iberoamerican Congress on Pattern Recognition, pages 123–132. Springer, 2023. 1, 2, 7", + "[7] Alec Holt, Isabelle Bichindaritz, Rainer Schmidt, and Petra Perner. Medical applications in case-based reasoning. The Knowledge Engineering Review, 20(3):289-292, 2005. 1", + "[8] Ta Duc Huy, Hoang Cao Huyen, Chanh DT Nguyen, Soan TM Duong, Trung Bui, and Steven QH Truong. Adversarial contrastive fourier domain adaptation for polyp segmentation. In 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2022. 1", + "[9] Sajid Javed, Arif Mahmood, Iyyakutti Iyappan Ganapathi, Fayaz Ali Dharejo, Naoufel Werghi, and Mohammed Bennamoun. Cclip: Zero-shot learning for histopathology with comprehensive vision-language alignment. In CVPR, pages 11450-11459, 2024. 1", + "[10] Injae Kim, Jongha Kim, Joonmyung Choi, and Hyunwoo J Kim. Concept bottleneck with visual concept filtering for explainable medical image classification. In The Medical Computing and Computer Assisted Intervention Society, pages 225–233. Springer, 2023. 1", + "[11] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning, pages 5338-5348. PMLR, 2020. 1, 7" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. Unmasking clever hans predictors and assessing what machines really learn. Nature communications, 10(1):1096, 2019. 2", + "[13] Yun Liu, Yu-Huan Wu, Yunfeng Ban, Huifang Wang, and Ming-Ming Cheng. Rethinking computer-aided tuberculosis diagnosis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2646–2655, 2020. 7", + "[14] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology. Nature Medicine, 30(3):863-874, 2024. 1", + "[15] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 7", + "[16] Yuksekgonul Mert, Wang Maggie, and Zou James. Post-hoc concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. 2, 7", + "[17] Meike Nauta, Ron van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14933-14943, 2021. 1, 7", + "[18] Meike Nauta, Jörg Schlötterer, Maurice van Keulen, and Christin Seifert. Pip-net: Patch-based intuitive prototypes for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2744-2753, 2023. 1, 3, 7, 8", + "[19] Giang Nguyen, Daeyoung Kim, and Anh Nguyen. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems, 34:26422-26436, 2021. 1", + "[20] Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, et al. Vindr-cxr: An open dataset of chest x-rays with radiologist's annotations. Scientific Data, 9(1):429, 2022. 7", + "[21] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023. 1", + "[22] Vu Minh Hieu Phan, Yutong Xie, Yuankai Qi, Lingqiao Liu, Liyang Liu, Bowen Zhang, Zhibin Liao, Qi Wu, Minh-Son To, and Johan W Verjans. Decomposing disease descriptions for enhanced pathology detection: A multi-aspect vision-language pre-training framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11492-11501, 2024. 1, 7", + "[23] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6450-6458, 2019. 4", + "[24] Tran Minh Quan, Huynh Minh Thanh, Ta Duc Huy, Nguyen Do Trung Chanh, Nguyen Thi Phuong Anh, Phan Hoan Vu, Nguyen Hoang Nam, Tran Quy Tuong, Vu Minh Dien, Bui Van Giang, et al. Xpgan: X-ray projected generative adversarial network for improving Covid-19 image classification." + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30805", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 1509-1513. IEEE, 2021. 1", + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7", + "[26] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence, 1(5):206-215, 2019. 1", + "[27] Dawid Rymarczyk, Łukasz Struski, Jacek Tabor, and Bartosz Zielinski. Protopshare: Prototype sharing for interpretable image classification and similarity discovery. arXiv preprint arXiv:2011.14340, 2020. 1, 3", + "[28] Dawid Rymarczyk, Łukasz Struski, Michal Górszczak, Koryna Lewandowska, Jacek Tabor, and Bartosz Zielinski. Interpretable image classification with differentiable prototypes assignment. In European Conference on Computer Vision, pages 351-368. Springer, 2022. 1, 3, 7", + "[29] Tran Bao Sam, Ta Duc Huy, Cong Tuyen Dao, Thanh Tin Lam, Van Ha Tang, and Steven Q.H. Truong. A multi-phase multi-graph approach for focal liver lesion classification on ct scans. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 3788-3802, 2024. 1", + "[30] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1", + "[31] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 1", + "[32] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 8", + "[33] Bor-Shiun Wang, Chien-Yi Wang, and Wei-Chen Chiu. Mcpnet: An interpretable classifier via multi-level concept prototypes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10885-10894, 2024. 1", + "[34] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 1", + "[35] Siyuan Yan, Zhen Yu, Xuelin Zhang, Dwarikanath Mahapatra, Shekhar S Chandra, Monika Janda, Peter Soyer, and Zongyuan Ge. Towards trustable skin cancer diagnosis via rewriting model's decision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11568-11577, 2023. 2", + "[36] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language" + ], + "bbox": [ + 91, + 90, + 483, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 1, 7", + "[37] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. International Journal of Computer Vision, 126(10):1084-1102, 2018. 8", + "[38] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 1" + ], + "bbox": [ + 516, + 90, + 906, + 273 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30806", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/a8ba1cd8-3d09-4bb8-8912-43eb50d5dbf1_model.json b/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/a8ba1cd8-3d09-4bb8-8912-43eb50d5dbf1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cd17f50273039281c61ad9b4e0c9ba041516db2c --- /dev/null +++ b/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/a8ba1cd8-3d09-4bb8-8912-43eb50d5dbf1_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.89, + 0.154 + ], + "angle": 0, + "content": "Interactive Medical Image Analysis with Concept-based Similarity Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.18, + 0.882, + 0.201 + ], + "angle": 0, + "content": "Ta Duc Huy† Sen Kim Tran Phan Nguyen Nguyen Hoang Tran Tran Bao Sam" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.205, + 0.847, + 0.224 + ], + "angle": 0, + "content": "Anton van den Hengel† Zhibin Liao† Johan W. Verjans† Minh-Son To†‡" + }, + { + "type": "text", + "bbox": [ + 0.412, + 0.231, + 0.593, + 0.248 + ], + "angle": 0, + "content": "Vu Minh Hieu Phan†*" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.251, + 0.776, + 0.267 + ], + "angle": 0, + "content": "† Australian Institute for Machine Learning, The University of Adelaide, ‡ Flinders University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.302, + 0.327, + 0.318 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.486, + 0.712 + ], + "angle": 0, + "content": "The ability to interpret and intervene model decisions is important for the adoption of computer-aided diagnosis methods in clinical workflows. Recent concept-based methods link the model predictions with interpretable concepts and modify their activation scores to interact with the model. However, these concepts are at the image level, which hinders the model from pinpointing the exact patches the concepts are activated. Alternatively, prototype-based methods learn representations from training image patches and compare these with test image patches, using the similarity scores for final class prediction. However, interpreting the underlying concepts of these patches can be challenging and often necessitates post-hoc guesswork. To address this issue, this paper introduces the novel Concept-based Similarity Reasoning network (CSR), which offers (i) patch-level prototype with intrinsic concept interpretation, and (ii) spatial interactivity. First, the proposed CSR provides localized explanation by grounding prototypes of each concept on image regions. Second, our model introduces novel spatial-level interaction, allowing doctors to engage directly with specific image areas, making it an intuitive and transparent tool for medical imaging. CSR improves upon prior state-of-the-art interpretable methods by up to \\(4.5\\%\\) across three biomedical datasets. Our code is released at https://github.com/tadeephuy/InteractCSR." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.743, + 0.222, + 0.759 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.769, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Despite the success of deep neural networks in medical image analysis [8, 9, 14, 22, 24, 29], their \"black-box\" nature limits their adoption in safety-critical clinical settings. Attribution-based methods [3, 30, 31, 34, 38] generate saliency maps to identify influential image regions. However, recent studies raised concerns about the reliability and accuracy of visual interpretation, potentially compro" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.908, + 0.472 + ], + "angle": 0, + "content": "mising decision-making [19, 26]. To address this, Concept Bottleneck Models (CBM) predict the interpretable concepts first and use them to predict the label [10, 11, 21, 36]. CBM allows human intervention in the predicted concepts, which is shown to significantly benefit performance. Alternatively, prototype-based methods are proposed to learn visual part prototypes to explain the classification decision. Specifically, ProtoPNet [2] and its derivatives [1, 6, 17, 18, 27, 28, 33] learn prototypical parts from training data and compare test images with these parts to compute similarity scores for final predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.909, + 0.642 + ], + "angle": 0, + "content": "Despite their inherent interpretability, concept-based methods only offer image-level explanations by linking images to global concepts, without identifying specific local patches correlated with these concepts. In contrast, prototype-based methods deliver patch-level explanations but necessitate post-hoc analysis to correlate visual patches with semantically meaningful concepts. This challenge intensifies in medical imaging, where subtle visual medical findings make concept assignments particularly difficult. How to create concept-based models that enable patch-level explanation for medical imaging is an open-ended question." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In radiological imaging, radiologists routinely reference an atlas of visual examples, depicting distinctive concepts of a pathology, such as opacity, fluid presence in the lung, enlarged heart shadow to aid diagnosis. Inspired by this observation, we propose a novel Concept-based Similarity Reasoning network, which offers two unique contributions. First, our CSR gives patch-level and interpretable-by-design concepts without requiring post-hoc analysis, in contrast to prototype-based methods [2, 18, 28]. We classify a disease by searching locally for the presence of visual exemplars of its concepts on the image. For example, a cancer diagnosis on chest X-ray could involve a radiologist comparing a suspected tumor with other known malignant tumors to assess its severity [7]. The presence of the tumor is a concept to predict the target cancer. Our framework trains a model to search locally for the presence of similar tumor exemplars (or prototypes) on every patch of the im" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30797" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.09, + 0.861, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.374, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Figure 1. Diagnosis Workflow with Doctor-in-the-Loop using CSR. Top section: CSR predicts Pulmonary Edema, falsely associating with the two concepts lung fluid and enlarged heart. For each concept, CSR explains its prediction by comparing the input with the highlighted region on the prototype image to create the corresponding similarity map. Bottom section: As the doctor inspects the similarity maps of each concept, he suppresses incorrect lung fluid attention, and reinforces the attention on opacity on the left of the image through spatial-interaction: “drawing positive and negative boxes” to create an importance map indicating where to focus and ignore. Secondly, noting the heart is normal, he “rejects” the enlarged heart concept via concept-interaction. CSR then recalibrates its prediction to Tuberculosis, aligning with the observed opacity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.498, + 0.482, + 0.545 + ], + "angle": 0, + "content": "age. After training, our model explicitly learns interpretable concept-grounded patch prototypes without a post-hoc analysis." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.549, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Second, our model enables spatial interactivity on the patch level. Unlike previous methods which only intervene in model predictions on the concepts-level outputs [1, 6, 16, 35], we introduce spatial interactivity on the input image. Ours offers intuitive benefits in medical imaging, where precise image analysis is crucial for clinicians. Deep learning models frequently capture spurious shortcuts correlations rather than the intended signals—a phenomenon known as the Clever-Hans effect [12]. Therefore, integrating doctor feedback to guide the model on relevant cues significantly enhances trustworthiness in clinical settings. Our contributions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.735, + 0.483, + 0.826 + ], + "angle": 0, + "content": "- We propose a novel Concept-based Similarity Reasoning network (CSR) that provides concept-level explanation. To the best of our knowledge, this is the first interpretable-by-design method that offers localized concept explainability for medical images without post-hoc analysis like part-prototype approaches." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- We introduce a novel spatial interaction mechanism (where to look), alongside train-time interaction (what not to learn) and concept-level interaction (what not present). CSR facilitates a comprehensive interactive framework that actively involves doctors in both train and test phases." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.498, + 0.907, + 0.558 + ], + "angle": 0, + "content": "- We conducted extensive experiments on three datasets and showed that CSR achieves a significant improvement by up to \\(4.5\\%\\) F1-score compared to other interpretable methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.571, + 0.605, + 0.587 + ], + "angle": 0, + "content": "2. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.907, + 0.703 + ], + "angle": 0, + "content": "CSR comprises (i) a Concept model to extract interpretable concept features; (ii) a feature projector \\( P \\) to enhance the concept feature space; and (iii) a task head \\( H \\) to classify based on concept similarity scores. This section describes the inference logic of the proposed Concept-based Similarity Reasoning model and the prototypes learning framework. Fig. 2 illustrates the inference process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.711, + 0.831, + 0.727 + ], + "angle": 0, + "content": "2.1. Concept-based Similarity Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.906, + 0.838 + ], + "angle": 0, + "content": "Concept-based classification. From an input \\( x \\), concept-based methods first predict the scores of \\( K \\) explainable concepts \\( \\{c^k\\}_{k=1}^K \\) and then use them as support to predict the target \\( y \\). Instead of predicting the concept scores, CSR calculates the similarity scores between \\( x \\) and the examples (or prototypes) of each concept. It then uses the concept similarity scores to support the prediction of \\( y \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.839, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Similarity reasoning. To construct the similarity reasoning inference logic, we establish an atlas: a set of \\( M \\cdot K \\) concept prototypes \\( \\{p^{k_m}\\}_{k=1,m=1}^{K,M} \\), where we have \\( M \\) prototypes for each concept \\( c^k \\). In essence, \\( p^{k_m} \\in \\mathbb{R}^{\\mathcal{C}} \\) repre" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30798" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.09, + 0.868, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.27, + 0.908, + 0.329 + ], + "angle": 0, + "content": "Figure 2. Inference logic of CSR. Each concept prototype represents a specific concept from a training image. CSR generates 2D similarity maps by computing the cosine similarity between these concept prototypes and the feature maps, and considers the maximum values as similarity scores to calculate the prediction logits. The prototype image refers to the training image associated with a specific concept prototype as detailed in Sec. 3.1" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.484, + 0.475 + ], + "angle": 0, + "content": "sents the \\(m^{\\mathrm{th}}\\) example of concept \\(k\\). Given an input image \\(x\\), we use the feature extractor \\(F\\) to create a feature map \\(\\mathbf{f} \\in \\mathbb{R}^{C \\times H \\times W}\\) having \\(\\mathcal{H} \\cdot \\mathcal{W}\\) feature patches. The similarity score \\(s^{k_m}\\) is defined as the maximum cosine similarity between the prototype \\(p^{k_m}\\) and each feature patch \\(\\mathbf{f}(h, w)\\), quantifying the presence of \\(p^{k_m}\\) on \\(x\\). As such, we can aggregate the presence of \\(M \\cdot K\\) concept prototypes \\(\\{p^{k_m}\\}\\) into a similarity scores vector:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.486, + 0.483, + 0.507 + ], + "angle": 0, + "content": "\\[\ns = \\left[ s ^ {k _ {m}} \\right] _ {k = 1, m = 1} ^ {K, M} = \\left[ s ^ {1 _ {1}}, s ^ {1 _ {2}}, \\dots , s ^ {K _ {M - 1}}, s ^ {K _ {M}} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.699 + ], + "angle": 0, + "content": "Then, the model predicts the target \\(y\\) from the similarity vector \\(s\\) with a task head \\(H\\): \\(y = H(s)\\). For example, the input chest X-ray is similar to several images from the atlas that have: fluid build-up in the left lung (score 0.85), and the abnormal heart size (score 0.88). Based on these similarities, the model predicts target class Pulmonary Edema. In short, the model classifies \\(y\\) by comparing exemplars \\(p^{k_m}\\) of the interpretable concept \\(c^k\\) on each local patch on the image \\(x\\). The choice of \\(H\\) should be but not be limited to a whitebox model such as a single linear layer or a shallow tree so that the reasoning from \\(s \\to y\\) maintains interpretability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.484, + 0.807 + ], + "angle": 0, + "content": "Discussion. Prototype-based methods [2, 18, 27, 28] also learn a set of exemplar prototypes \\( p \\). However, they require post-hoc analysis by experts to associate the interpretable concept labels to the learned prototypes, which is intractable in the fine-grained medical image analysis. In stark contrast, the proposed model learns patch-level and interpretable concept prototypes by design." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.347, + 0.835 + ], + "angle": 0, + "content": "2.2. Concept prototypes learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This section describes the proposed methods to learn the concept prototypes. Overall, our method consists of a Concept model to learn interpretable concept vectors, and a novel contrastive learning objective to enhance generaliz" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.905, + 0.384 + ], + "angle": 0, + "content": "ability of concept vectors. The overview of our learning framework is shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.906, + 0.505 + ], + "angle": 0, + "content": "Concept model. The Concept model comprises the feature extractor \\(F\\), and the concept head \\(C\\) that maps \\(\\mathbf{f}\\) into the one hot concept vector \\(c \\in \\mathbb{R}^K\\) such that \\(c = C(\\mathbf{f})\\), indicating the presence of concepts being associated with \\(x\\). The concept head \\(C\\) includes a \\(1 \\times 1\\) convolution layer that maps \\(\\mathbf{f}\\) to \\(K\\) channels: \\(\\mathbb{R}^{C \\times \\mathcal{H} \\times \\mathcal{W}} \\to \\mathbb{R}^{K \\times \\mathcal{H} \\times \\mathcal{W}}\\), representing the class activation maps \\(\\{\\mathrm{cam}^k \\in \\mathbb{R}^{\\mathcal{H} \\times \\mathcal{W}}\\}_{k=1}^K\\) for \\(K\\) concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.906, + 0.613 + ], + "angle": 0, + "content": "Local concept vectors acquisition. If the region \\((h, w)\\) on \\(x\\) contains the visual characteristics of concept \\(k\\), we consider the patch feature \\(\\mathbf{f}(h, w)\\) important and select it as the local concept vector. We first normalize \\(\\mathrm{cam}^k\\) using spatial softmax. Then, we soft-select the local concept vector \\(v^k \\in \\mathbb{R}^C\\) for concept \\(c^k\\) by weighting \\(\\mathbf{f}\\) with \\(\\mathrm{cam}^k\\) and summing the products spatially:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.624, + 0.905, + 0.662 + ], + "angle": 0, + "content": "\\[\nv ^ {k} = \\sum_ {h, w} ^ {\\mathcal {H}, \\mathcal {W}} \\operatorname {s o f t m a x} _ {h, w} \\left(\\operatorname {c a m} ^ {k}\\right) \\cdot \\mathbf {f}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.707 + ], + "angle": 0, + "content": "With a training set of \\(\\{x_{i}\\}^{N}\\) images of \\(K\\) concepts, we generate a set \\(\\{v_{i}^{k}\\}_{i = 1}^{V\\leq N\\cdot K}\\) of local concept vectors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.707, + 0.906, + 0.768 + ], + "angle": 0, + "content": "Similarity maps as explanations. Based on the local concept vector, we construct a 2D similarity map for concept \\( c^k \\) on \\( x \\) by evaluating the cosine similarity between each patch feature \\( \\mathbf{f}(h,w) \\) and vector \\( v^k \\):" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.779, + 0.905, + 0.799 + ], + "angle": 0, + "content": "\\[\nS ^ {k} (h, w) = \\cos \\left(v ^ {k}, \\mathbf {f} (h, w)\\right) = \\langle v ^ {k}, \\mathbf {f} (h, w) \\rangle , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\langle ,\\rangle\\) is the dot product of the two normalized vectors. As \\(v^{k}\\) is the importance weighted sum of \\(\\mathbf{f}\\) with respect to \\(c^k\\), comparing it with the important patch \\(\\mathbf{f}(h,w)\\) results in a high similarity score at \\(S^k (h,w)\\), and vice versa. Effectively, \\(S^k\\in \\mathbb{R}^{\\mathcal{H}\\times \\mathcal{W}}\\) is the local explanation for concept \\(c^k\\) on \\(x\\). The similarity score of \\(x\\) with the concept \\(c^k\\) is" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30799" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.089, + 0.868, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.343, + 0.908, + 0.414 + ], + "angle": 0, + "content": "Figure 3. The novel Concept prototypes learning framework. We pretrain a Concept model to generate concept activation maps. The local concept vectors are then generated by weighting and summing the feature maps with the activation map of the corresponding concept. To enhance the compactness and the generalizability of the concept feature space, we introduce a novel multi-prototype learning objective. After projecting concept vectors via a projector \\( P \\), the proposed objective pulls concept features to its nearest concept prototypes, while pushing away from prototypes of other concepts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.424, + 0.455 + ], + "angle": 0, + "content": "defined as the maximum of the similarity map \\( S^k \\):" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.46, + 0.482, + 0.485 + ], + "angle": 0, + "content": "\\[\ns ^ {k} = \\max _ {h, w} S ^ {k} (h, w). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.689 + ], + "angle": 0, + "content": "After obtaining the similarity scores \\( s_i^k \\) for the local concept vectors \\( \\{v_i^k\\} \\) with the input \\( x \\), one can naively adopt the inference logic in Sec. 2.1. However, this leads to two problems. First, the computation complexity is \\( O(N \\cdot K) \\). Second, \\( v_i^k \\) is local, which is specific to each image, thus non-generalizable to new images. We investigate the feasibility of this approach by comparing \\( v_i^k \\) with new images \\( v_{j \\neq i}^{k'} \\) and analyze the similarity scores distribution of intra-concept \\( (k = k') \\) and inter-concept \\( (k \\neq k') \\). It is observed that the concepts are not well-represented in new images. In other words, a local concept vector \\( v_i^k \\) generated from an image \\( x_i \\) fails to highlight the concept \\( k \\) on a new image \\( x_{j \\neq i} \\), as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.796 + ], + "angle": 0, + "content": "Concept prototypes learning. To improve the generalizability of concept representation \\( v \\) and reduce the number of concept vectors, our solution involves: (i) employing contrastive learning to enhance the shared embedding space between \\( v \\) and \\( \\mathbf{f} \\); (ii) downsampling the concept-specific-and-sample-specific set \\( \\{v_i^k\\} \\) to a small core set of prototypes \\( \\{p^k\\} \\) that is only specific to each concept, for efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.871 + ], + "angle": 0, + "content": "We use a projector \\( P \\) to map \\( v \\) to a feature space with the goal to improve inter-concept separation and intra-concept compactness such that \\( v' = P(v) \\) via contrastive learning. Then, we mine prototypes \\( p^k \\) for each concept, which can be a single prototype or a diverse set of prototypes." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Single prototype per concept. We initialize a prototype vector \\(p^k \\in \\mathbb{R}^C\\) for each concept \\(c^k\\). [23] shows that the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.538 + ], + "angle": 0, + "content": "cross-entropy of sharpened softmax logits is equivalent to a smoothed Triplet Loss [5] in a unit sphere. By minimizing the maximal entropy, we can pull \\( v_{i}^{\\prime \\tilde{k}} \\) to their respective prototype \\( p^{\\tilde{k}} \\) and push it away from irrelevant prototypes \\( k \\neq \\tilde{k} \\). Accordingly, we learn the prototype vector \\( p^{\\tilde{k}} \\) by minimizing the term:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.543, + 0.907, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathrm {c o n}} \\left(v _ {i} ^ {\\tilde {k}}\\right) = - \\log \\left(\\frac {\\exp \\left(\\lambda \\left\\langle p ^ {\\tilde {k}} , v _ {i} ^ {\\prime \\tilde {k}} \\right\\rangle\\right)}{\\sum_ {k \\in K} \\exp \\left(\\lambda \\left\\langle p ^ {k} , v _ {i} ^ {\\prime \\tilde {k}} \\right\\rangle\\right)}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.906, + 0.621 + ], + "angle": 0, + "content": "where \\(\\lambda > 1\\) is the scaling factor to sharpen the softmax logits, \\(p\\) and \\(v'\\) are \\(L_2\\) normalized." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.906, + 0.683 + ], + "angle": 0, + "content": "Multiple prototypes per concept. Due to the multi-modal nature of data where one class contains several clusters rather than just a single one, we initialize \\( M \\) prototypes per concept \\( \\{p^{k_m}\\}_{m=1}^M \\) such that \\( M \\ll N \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.906, + 0.768 + ], + "angle": 0, + "content": "The similarity of \\( v_{i}^{\\prime k} \\) with its concept \\( c^{\\tilde{k}} \\) is defined as the similarity to the nearest \\( p^{\\tilde{k}_m} \\), and can be approximated by summing the compatibility with the \\( M \\) clusters prototypes \\( p^{\\tilde{k}_m} \\). Specifically, we first compute the probabilistic assignment \\( \\mathbf{q}_m^{\\tilde{k}}(v_i^{\\prime k}) \\) to the prototypes:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.774, + 0.905, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathrm {q} _ {m} \\left(v ^ {\\prime} _ {i} ^ {\\tilde {k}}\\right) = \\operatorname {s o f t m a x} _ {m} \\left(\\gamma \\left\\langle p ^ {\\tilde {k} _ {m}}, v ^ {\\prime} _ {i} ^ {\\tilde {k}} \\right\\rangle\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.906, + 0.87 + ], + "angle": 0, + "content": "where \\( \\mathbf{q}_m \\) is the assignment distribution of the \\( M \\) prototypes in \\( c^{\\tilde{k}} \\) and \\( \\gamma > 1 \\) is the scaling factor. The similarity of the concept vector with each prototype is obtained by weighing their cosine similarity with the assignment:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.876, + 0.905, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\sin^ {\\tilde {k} _ {m}} \\left(v ^ {\\prime} _ {i} ^ {\\tilde {k}}\\right) = q _ {m} \\left(v ^ {\\prime} _ {i} ^ {\\tilde {k}}\\right) \\cdot \\langle p ^ {\\tilde {k} _ {m}}, v ^ {\\prime} _ {i} ^ {\\tilde {k}} \\rangle . \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "30800" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.126, + 0.09, + 0.451, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.268, + 0.452, + 0.329 + ], + "angle": 0, + "content": "(a) Similarity maps generated by the local concept vector (first row) and concept prototype (second row) on the same image (a) and new images (b), (c). The local concept vector fails to highlight the concept pacemaker on new images while the concept prototype generalizes well with better localization." + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.331, + 0.274, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.331, + 0.446, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.513, + 0.452, + 0.577 + ], + "angle": 0, + "content": "(b) Comparison between the original \\(v\\) and projected feature space \\(v^{\\prime}\\). The box plots of the similarity score distribution between \\(v_{i}^{k}\\) and \\(v_{j\\neq i}^{k^{\\prime}}\\). Intra-Concept: \\(k = k^{\\prime}\\), Inter-Concept: \\(k\\neq k^{\\prime}\\). The projected feature space is more compact and has better inter-concept separation." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.59, + 0.481, + 0.618 + ], + "angle": 0, + "content": "Figure 4. Contrastive learning improves the concept feature space. (a) Qualitative and (b) quantitative comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.676 + ], + "angle": 0, + "content": "After that, we can obtain the similarity with concept \\( c^{\\tilde{k}} \\) by summing the compatibility with each cluster prototype:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.686, + 0.483, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\sin^ {\\tilde {k}} \\left(v _ {i} ^ {\\prime} \\tilde {k}\\right) = \\sum_ {m} ^ {M} \\sin^ {\\tilde {k} _ {m}} \\left(v _ {i} ^ {\\prime} \\tilde {k}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.737, + 0.484, + 0.782 + ], + "angle": 0, + "content": "Finally, we compute the contrastive loss by maximizing the similarity of a concept vector with the cluster prototypes of the positive concept, similar to Eqn. 5:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.793, + 0.483, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {c o n - m}} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right) = - \\log \\left(\\frac {\\exp \\left(\\lambda \\left(\\sin^ {\\tilde {k}} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right) + \\delta\\right)\\right)}{\\sum_ {k \\in K} \\exp \\left(\\lambda \\sin^ {k} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right)\\right)}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.838, + 0.421, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\text {i f} \\quad k = \\tilde {k}, \\quad \\operatorname {s i m} ^ {k} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right) = \\operatorname {s i m} ^ {k} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right) + \\delta .\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "A small margin \\(\\delta\\) between concept prototypes is added to further expand the decision boundaries. As a result, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.168 + ], + "angle": 0, + "content": "learn a set of concept prototypes \\(\\{p^{k_m}\\}^{M\\cdot K}\\) that is significantly smaller in numbers compared to \\(\\{v_i^k\\}^{N\\cdot K}\\), and the projector \\(P\\) that enhances the inter-concept separation and intra-concept compactness. We illustrate the learning process of \\(P\\) and \\(\\{p\\}\\) in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.905, + 0.243 + ], + "angle": 0, + "content": "From Eqn. 2, \\( v_{i}^{k} \\) is a soft-selected feature patch of concept \\( c^k \\) from the original feature \\( \\mathbf{f}_i \\), we can directly use \\( P \\) to map feature patch \\( \\mathbf{f}_i(h,w) \\) to the compact and discriminative prototypical feature space. We then obtain the similarity of the concept prototypes with \\( x \\) from Eqn. 3:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.252, + 0.905, + 0.275 + ], + "angle": 0, + "content": "\\[\ns ^ {k _ {m}} = \\max _ {h, w} S ^ {k _ {m}} (h, w) = \\max _ {h, w} \\langle p ^ {k _ {m}}, P (\\mathbf {f} (h, w)) \\rangle , \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.287, + 0.905, + 0.333 + ], + "angle": 0, + "content": "which are aggregated to create the similarity vector \\([s^{k_m}]\\) in Eqn. 1. We quantitatively and qualitatively present the improved representation of \\(p\\) in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.906, + 0.455 + ], + "angle": 0, + "content": "Training details. Initially, we train the Concept model for multi-label classification using binary cross-entropy, predicting one-hot encoded vector of ground truth concepts. Then, we obtain the local concept vectors and learn the concept prototypes and the projection head with a contrastive loss \\(\\ell_{\\mathrm{con - m}}\\) to enhance concept generalization. Following, the task head is trained to predict the target from similarity scores with the concept prototypes, using cross-entropy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.8, + 0.485 + ], + "angle": 0, + "content": "3. Doctor-in-the-loop interactivity" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.492, + 0.702, + 0.509 + ], + "angle": 0, + "content": "3.1. Model explanations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.905, + 0.543 + ], + "angle": 0, + "content": "Given an input image \\( x \\), the model returns output \\( \\hat{y} \\) and the explanations for its prediction that include:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.544, + 0.905, + 0.574 + ], + "angle": 0, + "content": "- The similarity maps \\([S^{k_m}]\\) to show the activated region of the concept prototype \\(p^{k_m}\\) on \\(x\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.574, + 0.904, + 0.605 + ], + "angle": 0, + "content": "- The corresponding prototype image \\(\\mathcal{I}(p^{k_m})\\) as a reference, also highlighting the activated regions of \\(p^{k_m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.605, + 0.905, + 0.65 + ], + "angle": 0, + "content": "- The concept similarity scores \\([s^{k_m}]\\), indicating the similarity of the two activated regions of \\(p^{k_m}\\) on the input image \\(x\\) and on the reference image \\(\\mathcal{I}(p^{k_m})\\)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.544, + 0.905, + 0.65 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.906, + 0.727 + ], + "angle": 0, + "content": "The model explanations correspond to the similarity-based inference logic outlined in Sec. 2.1. The operator \\(\\mathcal{I}\\) retrieves the image associated with the prototype \\(p^{k_m}\\). We replace each \\(p^{k_m}\\) with the nearest projected local concept vector \\(v_{i}^{k}\\) and link it to its source image \\(x_{i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.737, + 0.905, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} \\left(p ^ {k _ {m}}\\right) = x \\leftarrow \\underset {i} {\\operatorname {a r g m i n}} \\left\\langle v _ {i} ^ {\\prime k}, p ^ {k _ {m}} \\right\\rangle . \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.774, + 0.722, + 0.789 + ], + "angle": 0, + "content": "3.2. Train-time interaction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "After we learn the concept prototypes \\(\\{p^{k_m}\\}\\) and its associated human-interpretable version \\(\\{\\mathcal{I}(p^{k_m})\\}\\), or atlas, doctors can inspect each concept prototype and discard any unqualified examples. This process can reveal and mitigate the Clever-Hans effect in the Concept model by examining the clues it is taking to predict the concepts. If a given prototype is shown to be a shortcut rather than a meaningful signal, it" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30801" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.091, + 0.462, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.288, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Figure 5. The irrelevant concept prototypes to be discarded in train-time interaction, highlighted at the activated regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.484, + 0.437 + ], + "angle": 0, + "content": "is removed from \\(\\{p^{k_m}\\}\\) and is not used to calculate the similarity scores \\([s^{k_m}]\\) for the prediction of \\(y\\), effectively nullify the Clever-Hans effect. We show the irrelevant examples of different concepts in Fig. 5. The similarity-based inference logic allows this interaction as we only need as much as the remaining relevant examples to predict the target \\(y\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.449, + 0.288, + 0.463 + ], + "angle": 0, + "content": "3.3. Test-time interaction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.483, + 0.517 + ], + "angle": 0, + "content": "At test time, we introduce two types of interaction to improve the prediction quality: Spatial-level interaction and Concept-level interaction. We illustrate this in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.653 + ], + "angle": 0, + "content": "Concept-level interaction. Consider this scenario: While reviewing the explanations from CSR for a chest X-ray \\( x \\), the doctor notices that the model is considering a nodule on \\( x \\) for a mass in the reference image with a high similarity score. Recognizing the incorrect comparison, the doctor informs the model that concept mass is absent in \\( x \\). The model suppresses its confidence in concept mass to calibrate its prediction. This concept-level interaction allows doctors to directly eliminate irrelevant concepts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Interactivity mechanism. When the doctor goes through \\([s^{k_m}]\\), she can discard any invalid concept if presented. The model then sets \\( s^{k_m} = 0 \\) for those rejected concepts, effectively discounting them on the prediction \\( \\hat{y} \\). Concept-level interaction only considers the negative type of input from the doctor: concept rejection. Setting a suitable value for every \\( s_m^k \\) of \\( c^k \\) would be a overburden. Hence, for each concept \\( c^k \\), we only present the highest similarity score \\( \\max_{m} \\{s^{k_m}\\} \\) and have the doctor recognize its presence. We zero out all of the similarity scores of the rejected concept." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Spatial-level interaction. As the doctor examines the image \\( x \\), upon spotting an important area, she can guide the model's attention with bounding boxes. The model then recalibrates its prediction \\( \\hat{y} \\) by prioritizing the indicated region based on this feedback. Conversely, reviewing the activation maps \\( [S^{k_m}] \\), the doctor also notices the model con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "centrating on an irrelevant region for prediction. To counteract the Clever-Hans effect, she marks the area as spurious, prompting the model to disregard it in its updated prediction. Hence, the doctor only needs to direct the model where to focus, without specifying what is in that region. This interaction scheme is practical and avoids the doctor from inadvertently becoming a teacher rather than taking advantage of the model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.217, + 0.907, + 0.337 + ], + "angle": 0, + "content": "Interactivity mechanism. As we described in the user journey above, the spatial-level interaction allows two types of input from the doctor: (i) the positive boxes \\(\\{\\mathsf{bb}^{+} = (\\mathsf{x}_1,\\mathsf{y}_1,\\mathsf{x}_2,\\mathsf{y}_2)\\}\\) for the focus areas, (ii) the negative boxes \\(\\{\\mathsf{bb}^{-}\\}\\) for the regions to ignore. To emphasize the area specified by \\(\\{\\mathsf{bb}^{+}\\}\\) and suppress the area specified by \\(\\{\\mathsf{bb}^{-}\\}\\), the model reweights each element \\(S(h,w)\\) with an importance map \\(A\\in \\mathbb{R}^{\\mathcal{H}\\times \\mathcal{W}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.357, + 0.905, + 0.412 + ], + "angle": 0, + "content": "\\[\nA (h, w) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} (h, w) \\in \\left\\{\\mathrm {b b} ^ {+} \\right\\} \\\\ 0 & \\text {i f} (h, w) \\in \\left\\{\\mathrm {b b} ^ {-} \\right\\}, \\\\ \\alpha & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.905, + 0.524 + ], + "angle": 0, + "content": "where \\(\\alpha \\in [0,1)\\) is the neutral weight for the other regions. Setting \\(\\alpha = 0\\) indicates that the doctor only considers the regions \\(\\{\\mathrm{bb}^{+}\\}\\), while setting \\(\\alpha \\rightarrow 1\\) diminishes the importance of \\(\\{\\mathrm{bb}^{+}\\}\\). The new set of similarity maps \\([\\hat{S}^{k_m}]\\) is obtained by multiplying every similarity map of \\([S^{k_m}]\\) with the importance map \\(A\\):" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.542, + 0.905, + 0.561 + ], + "angle": 0, + "content": "\\[\n[ \\hat {S} ] = A \\odot [ S ] = [ A \\odot S ^ {1 _ {1}}, \\dots , A \\odot S ^ {K _ {M}} ], \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.906, + 0.76 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes the element-wise multiplication on the spatial dimension. As CSR is designed to select the maximum element on \\(S^{k_m}\\) as \\(s^{k_m}\\), giving more weight to the important region \\(S(h,w)\\) increases its likelihood to be selected. On the other hand, zeroing out the spurious region nullifies its influence on the similarity vector. Because \\(S\\) is bounded within \\([-1,1]\\) and \\(A\\) is bounded within \\([0,1]\\), we clip negative values of \\(S\\) to condition the model on only non-negative values. This ensures that the \\(\\{\\mathrm{bb}^+\\}\\) and \\(\\{\\mathrm{bb}^-\\}\\) regions are monotonically adjusted. The new prediction is made with \\(\\hat{S}\\) by the same inference logic as described in Sec. 2.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Discussion. Combining both spatial and concept interaction mechanisms enables a flexible interaction approach. For example, if the doctor identifies an important region, she can draw a positive box \\(\\mathsf{bb}^{+}\\) over it. However, if the model associates the feature within \\(\\mathsf{bb}^{+}\\) with a wrong concept \\(c^k\\), placing \\(\\mathsf{bb}^{+}\\) over it would amplify the error. Because \\(c^k\\) is absent in \\(x\\), rejecting it while specifying \\(\\mathsf{bb}^{+}\\) will instruct the model to focus on other concepts \\(\\{c^{j\\neq k}\\}\\) within \\(\\mathsf{bb}^{+}\\) and ignore \\(c^k\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30802" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.238, + 0.106 + ], + "angle": 0, + "content": "4. Related works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.482, + 0.327 + ], + "angle": 0, + "content": "Concept-based. Concept Bottleneck Model (CBM) [11] predicts the concepts before using them as support for the target prediction. Post-hoc Concept Bottleneck Models (PHCBM) [16] transfer concepts from other datasets or from natural language descriptions of concepts via multimodal models like CLIP [25]. Yang et al. leverages Large Language Models to generate an exhaustive list of concepts for each class [36]. These multimodal concept-based approaches [15, 22, 36] utilize foundation models to enhance their capabilities. However, foundation models are generally trained on broad domains, which presents limitations when applying them directly to medical imaging. In this work, we focus on the research direction of vision features and explore multi-modality learning in future studies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.329, + 0.483, + 0.509 + ], + "angle": 0, + "content": "Part-prototypes. ProtoPNet [2] used prototypical parts to make a final classification based on positive attributions. ProtoTree[17] is a globally interpretable model by design that uses prototypes with decision trees. ProtoPool [28] is similar to ProtoPNet with a tweak of the shared prototypes pool and the Gumbel-softmax trick for the differentiable property. PIP-Net[18] is a patch-based approach with a novel regularization method for learning prototype similarity that correlates with human visual perception. Unlike previous methods that correlate prototypes with target classes, CSR explicitly grounds prototypes on interpretable concepts, making them semantically relevant." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.483, + 0.648 + ], + "angle": 0, + "content": "Interactivity. CBM [11] enables local interaction as human adjusts the concept logits for individual samples, while [16] allows global interaction by modifying concept weights of the model. ProtoPDebug [1] and [6] focus on part-prototype interaction, refining prototypes for enhanced performance. Besides logits-level interaction like those methods, CSR introduces spatial-level interaction, which is particularly intuitive and beneficial in the medical imaging domain, where detailed image analysis is of significant interest to doctors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.663, + 0.327, + 0.68 + ], + "angle": 0, + "content": "5. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.81 + ], + "angle": 0, + "content": "In this paper, we experiment on two chest X-ray datasets: TBX11K [13], VinDr-CXR[20] and a skin imaging dataset ISIC[4]. From those datasets, we define the findings annotations as the concepts and the diseases annotations as the target classes. Details on datasets and the training hyperparameters are in the supplementary. We qualitatively and quantitatively demonstrate the effectiveness of the concept prototypes learning framework in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Diagnostic performance. Due to the heavy class imbalance, we use macro F1-score as the metric for our experiments. We report the results in Table 1. It can be seen that CSR demonstrates a strong diagnostic performance compared to other baselines across all dataset. We also benchmark the performance of CSR when we ask a doctor to re" + }, + { + "type": "table", + "bbox": [ + 0.565, + 0.09, + 0.855, + 0.476 + ], + "angle": 0, + "content": "
MethodF1↑#pro.↓#exp.↓
TBX11KCBM (ind.)24.514
CBM (seq.)60.514
CBM (joint.)88.614
ProtoPNet89.3600600
ProtoPNet89.315001500
ProtoPNet94.130003000
ProtoPool90.7600600
ProtoTree(H=6)94.01276
ProtoTree(H=9)93.75119
PIP-Net94.0768158
CSR94.4140014
CSR (refined)94.0138614
VinDr-CXRCBM (joint.)50.114
ProtoPNet41.415001500
PIP-Net45.17689
CSR54.6140014
ISICCBM (joint.)45.54
ProtoPNet38.5600600
ProtoPool31.0200200
ProtoTree(H=6)66.11276
PIP-Net69.976890
CSR71.54004
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.487, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Table 1. Diagnostic performance on the medical datasets, measured by Macro F1-score. CSR (refined) has the doctors review the concept atlas. #pro. denotes the number of prototypes of the model. #exp. indicates the explanation size for each prediction." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.559, + 0.905, + 0.61 + ], + "angle": 0, + "content": "
MethodProtoP-NetProto TreePIP-NetCBMCSRCSR (refined)
% PG hit8.87.819.555.160.979.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Table 2. Pointing Game hit rate on TBX11K, Tuberculosis class. CBM uses the class activation map of the highest contributing product to the class Tuberculosis. ProtoTree uses a similarity map from the top prototype at level-1 of the tree. ProtoPNet, PIP-Net, and CSR use the similarity map from the prototype with the highest contribution to class Tuberculosis." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.733, + 0.869, + 0.748 + ], + "angle": 0, + "content": "fine the learned concept atlas as described in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Explanation size. Following [18], we quantify the explanation size as the number of explanations per prediction for each method. For part-prototype methods, explanations refer to the similar parts of the image from the training set. ProtoTree has the explanation size equal to its depth. PIP-Net reduces the explanation size by regularizing the sparsity during training and only considers non-zero prototypes for each prediction. Other part-prototype approaches have the explanation size matching the large number of prototypes. This compromises interpretability since users must" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30803" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.48, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.248, + 0.488, + 0.306 + ], + "angle": 0, + "content": "Figure 6. Qualitative result of the Pointing Game. We visualize the groundtruth bounding box and the maximum activation point of each method as the colored square. :Proposed CSR, :CBM, :PIP-Net, :ProtoPNet, :ProtoTree" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.484, + 0.503 + ], + "angle": 0, + "content": "navigate through a multitude of explanations within the set for each prediction. Concept-based models use the concepts as their explanations, having an explanation size equal to the number of concepts, which is typically smaller. This choice involves an accuracy trade-off due to the reduced size of the bottleneck. In contrast, the proposed CSR benefits from both, featuring a small explanation size equivalent to CBM while achieving competitive accuracy akin to part-prototype methods. To achieve this, our CSR only considers concept grounded explanations and exclusively presents the highest similar prototype of each active concept." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.642 + ], + "angle": 0, + "content": "Trustworthiness. We inspect the explanations of each benchmark and demonstrate that interpretability does not always come with trustworthiness. A trustworthy model must deliver relevant explanations that correspond to human interpretable concepts. We leverage the pointing game (PG) [37] that considers a hit if the point of maximum value in the explanation lies within a segmentation ground truth. We use the bounding box annotation from the TBX11K and calculate the PG hit rate of the explanations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The quantitative results are reported in Table. 2. Partprototype methods including ProtoPNet and ProtoTree are relying on clues that considerably mismatch with the semantic ground truth. This observation aligns with [18], where Purity of the prototypes is evaluated as an alignment with the part annotations in the CUB dataset [32], demonstrating a large semantic gap between the prototypes and human-relevant concepts. Concept-based methods attain a more reasonable PG hit rate as the concepts are human-grounded and thus align well with the semantic ground truth. The proposed CSR faithfully aligns with human reasoning and achieves a remarkable PG hit rate of \\(60.9\\%\\). Refining the concept atlas further improves the PG hit rate to \\(79.5\\%\\), which hints that the better classification result of the original CSR in Table 1 could stem from shortcuts. The qualitative results in Fig. 6 shows that our CSR:□ more frequently identifies the maximum activation within" + }, + { + "type": "table", + "bbox": [ + 0.564, + 0.089, + 0.858, + 0.206 + ], + "angle": 0, + "content": "
spatialconceptMac.F1+Δ%
CSR94.40.0
Non-expert✓94.50.1
Oracle✓94.90.5
M.D.✓94.70.3
M.D.✓95.10.7
M.D.✓✓94.90.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.217, + 0.907, + 0.273 + ], + "angle": 0, + "content": "Table 3. Interactions performance gain. spatial: spatial-level interaction, concept: concept-level interaction. We limit the Non-export to spatial-interaction only. Oracle: using the bounding box ground truth of TB as spatial-interaction. M.D: medical student." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.315, + 0.838, + 0.33 + ], + "angle": 0, + "content": "the ground truth box compared to other methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Interactivity. We demonstrate the benefits of interactivity by measuring the model when obtaining interaction at test time on the TBX11K dataset and report the result in Table. 3. Our experiment involves a Medical degree (M.D.) student and a non-expert in evaluating the framework. In the oracle setting, we use the ground truth bounding box of class TB from the dataset as a positive spatial-level interaction. Note that we only require human interaction when the output probability of the model is indecisive: when the difference of top-1 and top-2 class probability is less than 0.3. The weight \\(\\alpha\\) is set to 0.2. It shows that human interactions has a positive impact on the model outcome. The interaction from the M.D. student is more beneficial to the model than the non-expert. We consider that the non-experts can still eliminate irrelevant CXR regions without formal radiology expertise, and limit the non-expert to spatial-level interactions only. The concept-level interaction is more influential than the spatial-level interaction with better gain." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.646, + 0.634, + 0.663 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This paper introduces a novel Concept Similarity Reasoning (CSR), an interpretable and interactive model for medical image analysis with similarity reasoning. CSR provides local explanations by presenting similar concept examples with the input images. In comparison to part-prototype approaches, our CSR presents more concise and meaningful prototypes that align better with human-relevant concepts. This observation implies the differentiation between interpretability and trustworthiness. The proposed interaction mechanism employed during both training and testing effectively enhances model reliability by mitigating shortcuts and improves model accuracy through domain-specific inputs. Our method is an attempt to encourage more trust from doctors through a natural and effective human-AI interaction, improving both accuracy and transparency." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30804" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.157 + ], + "angle": 0, + "content": "[1] Andrea Bontempelli, Stefano Teso, Fausto Giunchiglia, and Andrea Passerini. Concept-level debugging of part-prototype networks. ArXiv, abs/2205.15769, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: Deep learning for interpretable image recognition. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[3] Townim Faisal Chowdhury, Kewen Liao, Vu Minh Hieu Phan, Minh-Son To, Yutong Xie, Kevin Hung, David Ross, Anton van den Hengel, Johan W Verjans, and Zhibin Liao. Cape: Cam as a probabilistic ensemble for enhanced dnn interpretation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11072-11081, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[4] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), pages 168-172. IEEE, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.484, + 0.512 + ], + "angle": 0, + "content": "[5] Xingping Dong and Jianbing Shen. Triplet loss in siamese network for object tracking. In Proceedings of the European conference on computer vision (ECCV), pages 459-474, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.515, + 0.484, + 0.583 + ], + "angle": 0, + "content": "[6] Michael Gerstenberger, Thomas Wiegand, Peter Eisert, and Sebastian Bosse. But that's not why: Inference adjustment by interactive prototype revision. In Iberoamerican Congress on Pattern Recognition, pages 123–132. Springer, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.484, + 0.628 + ], + "angle": 0, + "content": "[7] Alec Holt, Isabelle Bichindaritz, Rainer Schmidt, and Petra Perner. Medical applications in case-based reasoning. The Knowledge Engineering Review, 20(3):289-292, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.484, + 0.7 + ], + "angle": 0, + "content": "[8] Ta Duc Huy, Hoang Cao Huyen, Chanh DT Nguyen, Soan TM Duong, Trung Bui, and Steven QH Truong. Adversarial contrastive fourier domain adaptation for polyp segmentation. In 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.484, + 0.77 + ], + "angle": 0, + "content": "[9] Sajid Javed, Arif Mahmood, Iyyakutti Iyappan Ganapathi, Fayaz Ali Dharejo, Naoufel Werghi, and Mohammed Bennamoun. Cclip: Zero-shot learning for histopathology with comprehensive vision-language alignment. In CVPR, pages 11450-11459, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[10] Injae Kim, Jongha Kim, Joonmyung Choi, and Hyunwoo J Kim. Concept bottleneck with visual concept filtering for explainable medical image classification. In The Medical Computing and Computer Assisted Intervention Society, pages 225–233. Springer, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning, pages 5338-5348. PMLR, 2020. 1, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[12] Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. Unmasking clever hans predictors and assessing what machines really learn. Nature communications, 10(1):1096, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.906, + 0.231 + ], + "angle": 0, + "content": "[13] Yun Liu, Yu-Huan Wu, Yunfeng Ban, Huifang Wang, and Ming-Ming Cheng. Rethinking computer-aided tuberculosis diagnosis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2646–2655, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.906, + 0.301 + ], + "angle": 0, + "content": "[14] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology. Nature Medicine, 30(3):863-874, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[15] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.906, + 0.372 + ], + "angle": 0, + "content": "[16] Yuksekgonul Mert, Wang Maggie, and Zou James. Post-hoc concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[17] Meike Nauta, Ron van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14933-14943, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.907, + 0.511 + ], + "angle": 0, + "content": "[18] Meike Nauta, Jörg Schlötterer, Maurice van Keulen, and Christin Seifert. Pip-net: Patch-based intuitive prototypes for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2744-2753, 2023. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.906, + 0.567 + ], + "angle": 0, + "content": "[19] Giang Nguyen, Daeyoung Kim, and Anh Nguyen. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems, 34:26422-26436, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.906, + 0.636 + ], + "angle": 0, + "content": "[20] Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, et al. Vindr-cxr: An open dataset of chest x-rays with radiologist's annotations. Scientific Data, 9(1):429, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.906, + 0.678 + ], + "angle": 0, + "content": "[21] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.907, + 0.775 + ], + "angle": 0, + "content": "[22] Vu Minh Hieu Phan, Yutong Xie, Yuankai Qi, Lingqiao Liu, Liyang Liu, Bowen Zhang, Zhibin Liao, Qi Wu, Minh-Son To, and Johan W Verjans. Decomposing disease descriptions for enhanced pathology detection: A multi-aspect vision-language pre-training framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11492-11501, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.906, + 0.832 + ], + "angle": 0, + "content": "[23] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6450-6458, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Tran Minh Quan, Huynh Minh Thanh, Ta Duc Huy, Nguyen Do Trung Chanh, Nguyen Thi Phuong Anh, Phan Hoan Vu, Nguyen Hoang Nam, Tran Quy Tuong, Vu Minh Dien, Bui Van Giang, et al. Xpgan: X-ray projected generative adversarial network for improving Covid-19 image classification." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30805" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 1509-1513. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[26] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence, 1(5):206-215, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[27] Dawid Rymarczyk, Łukasz Struski, Jacek Tabor, and Bartosz Zielinski. Protopshare: Prototype sharing for interpretable image classification and similarity discovery. arXiv preprint arXiv:2011.14340, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[28] Dawid Rymarczyk, Łukasz Struski, Michal Górszczak, Koryna Lewandowska, Jacek Tabor, and Bartosz Zielinski. Interpretable image classification with differentiable prototypes assignment. In European Conference on Computer Vision, pages 351-368. Springer, 2022. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.484, + 0.461 + ], + "angle": 0, + "content": "[29] Tran Bao Sam, Ta Duc Huy, Cong Tuyen Dao, Thanh Tin Lam, Van Ha Tang, and Steven Q.H. Truong. A multi-phase multi-graph approach for focal liver lesion classification on ct scans. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 3788-3802, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[30] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.484, + 0.589 + ], + "angle": 0, + "content": "[31] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[32] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[33] Bor-Shiun Wang, Chien-Yi Wang, and Wei-Chen Chiu. Mcpnet: An interpretable classifier via multi-level concept prototypes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10885-10894, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[34] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[35] Siyuan Yan, Zhen Yu, Xuelin Zhang, Dwarikanath Mahapatra, Shekhar S Chandra, Monika Janda, Peter Soyer, and Zongyuan Ge. Towards trustable skin cancer diagnosis via rewriting model's decision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11568-11577, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[36] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.092, + 0.908, + 0.149 + ], + "angle": 0, + "content": "in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[37] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. International Journal of Computer Vision, 126(10):1084-1102, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[38] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30806" + } + ] +] \ No newline at end of file diff --git a/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/a8ba1cd8-3d09-4bb8-8912-43eb50d5dbf1_origin.pdf b/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/a8ba1cd8-3d09-4bb8-8912-43eb50d5dbf1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ad2a088cf4212fa3a7217e1ece2132d93a20afca --- /dev/null +++ b/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/a8ba1cd8-3d09-4bb8-8912-43eb50d5dbf1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec71a1613551fe46c5a18e5bf4a9a10484d18386cf1fac1ecb538db0a31bd6c3 +size 1463031 diff --git a/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/full.md b/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5bd3f2ce004bd95dca6db75b8c7c4aeb149e2b59 --- /dev/null +++ b/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/full.md @@ -0,0 +1,306 @@ +# Interactive Medical Image Analysis with Concept-based Similarity Reasoning + +Ta Duc Huy† Sen Kim Tran Phan Nguyen Nguyen Hoang Tran Tran Bao Sam + +Anton van den Hengel† Zhibin Liao† Johan W. Verjans† Minh-Son To†‡ + +Vu Minh Hieu Phan†* + +† Australian Institute for Machine Learning, The University of Adelaide, ‡ Flinders University + +# Abstract + +The ability to interpret and intervene model decisions is important for the adoption of computer-aided diagnosis methods in clinical workflows. Recent concept-based methods link the model predictions with interpretable concepts and modify their activation scores to interact with the model. However, these concepts are at the image level, which hinders the model from pinpointing the exact patches the concepts are activated. Alternatively, prototype-based methods learn representations from training image patches and compare these with test image patches, using the similarity scores for final class prediction. However, interpreting the underlying concepts of these patches can be challenging and often necessitates post-hoc guesswork. To address this issue, this paper introduces the novel Concept-based Similarity Reasoning network (CSR), which offers (i) patch-level prototype with intrinsic concept interpretation, and (ii) spatial interactivity. First, the proposed CSR provides localized explanation by grounding prototypes of each concept on image regions. Second, our model introduces novel spatial-level interaction, allowing doctors to engage directly with specific image areas, making it an intuitive and transparent tool for medical imaging. CSR improves upon prior state-of-the-art interpretable methods by up to $4.5\%$ across three biomedical datasets. Our code is released at https://github.com/tadeephuy/InteractCSR. + +# 1. Introduction + +Despite the success of deep neural networks in medical image analysis [8, 9, 14, 22, 24, 29], their "black-box" nature limits their adoption in safety-critical clinical settings. Attribution-based methods [3, 30, 31, 34, 38] generate saliency maps to identify influential image regions. However, recent studies raised concerns about the reliability and accuracy of visual interpretation, potentially compro + +mising decision-making [19, 26]. To address this, Concept Bottleneck Models (CBM) predict the interpretable concepts first and use them to predict the label [10, 11, 21, 36]. CBM allows human intervention in the predicted concepts, which is shown to significantly benefit performance. Alternatively, prototype-based methods are proposed to learn visual part prototypes to explain the classification decision. Specifically, ProtoPNet [2] and its derivatives [1, 6, 17, 18, 27, 28, 33] learn prototypical parts from training data and compare test images with these parts to compute similarity scores for final predictions. + +Despite their inherent interpretability, concept-based methods only offer image-level explanations by linking images to global concepts, without identifying specific local patches correlated with these concepts. In contrast, prototype-based methods deliver patch-level explanations but necessitate post-hoc analysis to correlate visual patches with semantically meaningful concepts. This challenge intensifies in medical imaging, where subtle visual medical findings make concept assignments particularly difficult. How to create concept-based models that enable patch-level explanation for medical imaging is an open-ended question. + +In radiological imaging, radiologists routinely reference an atlas of visual examples, depicting distinctive concepts of a pathology, such as opacity, fluid presence in the lung, enlarged heart shadow to aid diagnosis. Inspired by this observation, we propose a novel Concept-based Similarity Reasoning network, which offers two unique contributions. First, our CSR gives patch-level and interpretable-by-design concepts without requiring post-hoc analysis, in contrast to prototype-based methods [2, 18, 28]. We classify a disease by searching locally for the presence of visual exemplars of its concepts on the image. For example, a cancer diagnosis on chest X-ray could involve a radiologist comparing a suspected tumor with other known malignant tumors to assess its severity [7]. The presence of the tumor is a concept to predict the target cancer. Our framework trains a model to search locally for the presence of similar tumor exemplars (or prototypes) on every patch of the im + +![](images/b2f63d2c0aae5676106a1db2668e72ae4e62048251fb4b657029d5a9d665224a.jpg) +Figure 1. Diagnosis Workflow with Doctor-in-the-Loop using CSR. Top section: CSR predicts Pulmonary Edema, falsely associating with the two concepts lung fluid and enlarged heart. For each concept, CSR explains its prediction by comparing the input with the highlighted region on the prototype image to create the corresponding similarity map. Bottom section: As the doctor inspects the similarity maps of each concept, he suppresses incorrect lung fluid attention, and reinforces the attention on opacity on the left of the image through spatial-interaction: “drawing positive and negative boxes” to create an importance map indicating where to focus and ignore. Secondly, noting the heart is normal, he “rejects” the enlarged heart concept via concept-interaction. CSR then recalibrates its prediction to Tuberculosis, aligning with the observed opacity. + +age. After training, our model explicitly learns interpretable concept-grounded patch prototypes without a post-hoc analysis. + +Second, our model enables spatial interactivity on the patch level. Unlike previous methods which only intervene in model predictions on the concepts-level outputs [1, 6, 16, 35], we introduce spatial interactivity on the input image. Ours offers intuitive benefits in medical imaging, where precise image analysis is crucial for clinicians. Deep learning models frequently capture spurious shortcuts correlations rather than the intended signals—a phenomenon known as the Clever-Hans effect [12]. Therefore, integrating doctor feedback to guide the model on relevant cues significantly enhances trustworthiness in clinical settings. Our contributions: + +- We propose a novel Concept-based Similarity Reasoning network (CSR) that provides concept-level explanation. To the best of our knowledge, this is the first interpretable-by-design method that offers localized concept explainability for medical images without post-hoc analysis like part-prototype approaches. +- We introduce a novel spatial interaction mechanism (where to look), alongside train-time interaction (what not to learn) and concept-level interaction (what not present). CSR facilitates a comprehensive interactive framework that actively involves doctors in both train and test phases. + +- We conducted extensive experiments on three datasets and showed that CSR achieves a significant improvement by up to $4.5\%$ F1-score compared to other interpretable methods. + +# 2. Method + +CSR comprises (i) a Concept model to extract interpretable concept features; (ii) a feature projector $P$ to enhance the concept feature space; and (iii) a task head $H$ to classify based on concept similarity scores. This section describes the inference logic of the proposed Concept-based Similarity Reasoning model and the prototypes learning framework. Fig. 2 illustrates the inference process. + +# 2.1. Concept-based Similarity Reasoning + +Concept-based classification. From an input $x$ , concept-based methods first predict the scores of $K$ explainable concepts $\{c^k\}_{k=1}^K$ and then use them as support to predict the target $y$ . Instead of predicting the concept scores, CSR calculates the similarity scores between $x$ and the examples (or prototypes) of each concept. It then uses the concept similarity scores to support the prediction of $y$ . + +Similarity reasoning. To construct the similarity reasoning inference logic, we establish an atlas: a set of $M \cdot K$ concept prototypes $\{p^{k_m}\}_{k=1,m=1}^{K,M}$ , where we have $M$ prototypes for each concept $c^k$ . In essence, $p^{k_m} \in \mathbb{R}^{\mathcal{C}}$ repre + +![](images/2d9f88b37c24908d874e5d22ff8902cd6e558d93a82ce4c0dd4150fb4c60d914.jpg) +Figure 2. Inference logic of CSR. Each concept prototype represents a specific concept from a training image. CSR generates 2D similarity maps by computing the cosine similarity between these concept prototypes and the feature maps, and considers the maximum values as similarity scores to calculate the prediction logits. The prototype image refers to the training image associated with a specific concept prototype as detailed in Sec. 3.1 + +sents the $m^{\mathrm{th}}$ example of concept $k$ . Given an input image $x$ , we use the feature extractor $F$ to create a feature map $\mathbf{f} \in \mathbb{R}^{C \times H \times W}$ having $\mathcal{H} \cdot \mathcal{W}$ feature patches. The similarity score $s^{k_m}$ is defined as the maximum cosine similarity between the prototype $p^{k_m}$ and each feature patch $\mathbf{f}(h, w)$ , quantifying the presence of $p^{k_m}$ on $x$ . As such, we can aggregate the presence of $M \cdot K$ concept prototypes $\{p^{k_m}\}$ into a similarity scores vector: + +$$ +s = \left[ s ^ {k _ {m}} \right] _ {k = 1, m = 1} ^ {K, M} = \left[ s ^ {1 _ {1}}, s ^ {1 _ {2}}, \dots , s ^ {K _ {M - 1}}, s ^ {K _ {M}} \right]. \tag {1} +$$ + +Then, the model predicts the target $y$ from the similarity vector $s$ with a task head $H$ : $y = H(s)$ . For example, the input chest X-ray is similar to several images from the atlas that have: fluid build-up in the left lung (score 0.85), and the abnormal heart size (score 0.88). Based on these similarities, the model predicts target class Pulmonary Edema. In short, the model classifies $y$ by comparing exemplars $p^{k_m}$ of the interpretable concept $c^k$ on each local patch on the image $x$ . The choice of $H$ should be but not be limited to a whitebox model such as a single linear layer or a shallow tree so that the reasoning from $s \to y$ maintains interpretability. + +Discussion. Prototype-based methods [2, 18, 27, 28] also learn a set of exemplar prototypes $p$ . However, they require post-hoc analysis by experts to associate the interpretable concept labels to the learned prototypes, which is intractable in the fine-grained medical image analysis. In stark contrast, the proposed model learns patch-level and interpretable concept prototypes by design. + +# 2.2. Concept prototypes learning + +This section describes the proposed methods to learn the concept prototypes. Overall, our method consists of a Concept model to learn interpretable concept vectors, and a novel contrastive learning objective to enhance generaliz + +ability of concept vectors. The overview of our learning framework is shown in Fig. 3. + +Concept model. The Concept model comprises the feature extractor $F$ , and the concept head $C$ that maps $\mathbf{f}$ into the one hot concept vector $c \in \mathbb{R}^K$ such that $c = C(\mathbf{f})$ , indicating the presence of concepts being associated with $x$ . The concept head $C$ includes a $1 \times 1$ convolution layer that maps $\mathbf{f}$ to $K$ channels: $\mathbb{R}^{C \times \mathcal{H} \times \mathcal{W}} \to \mathbb{R}^{K \times \mathcal{H} \times \mathcal{W}}$ , representing the class activation maps $\{\mathrm{cam}^k \in \mathbb{R}^{\mathcal{H} \times \mathcal{W}}\}_{k=1}^K$ for $K$ concepts. + +Local concept vectors acquisition. If the region $(h, w)$ on $x$ contains the visual characteristics of concept $k$ , we consider the patch feature $\mathbf{f}(h, w)$ important and select it as the local concept vector. We first normalize $\mathrm{cam}^k$ using spatial softmax. Then, we soft-select the local concept vector $v^k \in \mathbb{R}^C$ for concept $c^k$ by weighting $\mathbf{f}$ with $\mathrm{cam}^k$ and summing the products spatially: + +$$ +v ^ {k} = \sum_ {h, w} ^ {\mathcal {H}, \mathcal {W}} \operatorname {s o f t m a x} _ {h, w} \left(\operatorname {c a m} ^ {k}\right) \cdot \mathbf {f}. \tag {2} +$$ + +With a training set of $\{x_{i}\}^{N}$ images of $K$ concepts, we generate a set $\{v_{i}^{k}\}_{i = 1}^{V\leq N\cdot K}$ of local concept vectors. + +Similarity maps as explanations. Based on the local concept vector, we construct a 2D similarity map for concept $c^k$ on $x$ by evaluating the cosine similarity between each patch feature $\mathbf{f}(h,w)$ and vector $v^k$ : + +$$ +S ^ {k} (h, w) = \cos \left(v ^ {k}, \mathbf {f} (h, w)\right) = \langle v ^ {k}, \mathbf {f} (h, w) \rangle , \tag {3} +$$ + +where $\langle ,\rangle$ is the dot product of the two normalized vectors. As $v^{k}$ is the importance weighted sum of $\mathbf{f}$ with respect to $c^k$ , comparing it with the important patch $\mathbf{f}(h,w)$ results in a high similarity score at $S^k (h,w)$ , and vice versa. Effectively, $S^k\in \mathbb{R}^{\mathcal{H}\times \mathcal{W}}$ is the local explanation for concept $c^k$ on $x$ . The similarity score of $x$ with the concept $c^k$ is + +![](images/b4c33f0a7af0b7a5a8c6a0fc97c9e30b4121993928f6dea4a95e82b4f3718b65.jpg) +Figure 3. The novel Concept prototypes learning framework. We pretrain a Concept model to generate concept activation maps. The local concept vectors are then generated by weighting and summing the feature maps with the activation map of the corresponding concept. To enhance the compactness and the generalizability of the concept feature space, we introduce a novel multi-prototype learning objective. After projecting concept vectors via a projector $P$ , the proposed objective pulls concept features to its nearest concept prototypes, while pushing away from prototypes of other concepts. + +defined as the maximum of the similarity map $S^k$ : + +$$ +s ^ {k} = \max _ {h, w} S ^ {k} (h, w). \tag {4} +$$ + +After obtaining the similarity scores $s_i^k$ for the local concept vectors $\{v_i^k\}$ with the input $x$ , one can naively adopt the inference logic in Sec. 2.1. However, this leads to two problems. First, the computation complexity is $O(N \cdot K)$ . Second, $v_i^k$ is local, which is specific to each image, thus non-generalizable to new images. We investigate the feasibility of this approach by comparing $v_i^k$ with new images $v_{j \neq i}^{k'}$ and analyze the similarity scores distribution of intra-concept $(k = k')$ and inter-concept $(k \neq k')$ . It is observed that the concepts are not well-represented in new images. In other words, a local concept vector $v_i^k$ generated from an image $x_i$ fails to highlight the concept $k$ on a new image $x_{j \neq i}$ , as shown in Fig. 4. + +Concept prototypes learning. To improve the generalizability of concept representation $v$ and reduce the number of concept vectors, our solution involves: (i) employing contrastive learning to enhance the shared embedding space between $v$ and $\mathbf{f}$ ; (ii) downsampling the concept-specific-and-sample-specific set $\{v_i^k\}$ to a small core set of prototypes $\{p^k\}$ that is only specific to each concept, for efficiency. + +We use a projector $P$ to map $v$ to a feature space with the goal to improve inter-concept separation and intra-concept compactness such that $v' = P(v)$ via contrastive learning. Then, we mine prototypes $p^k$ for each concept, which can be a single prototype or a diverse set of prototypes. + +Single prototype per concept. We initialize a prototype vector $p^k \in \mathbb{R}^C$ for each concept $c^k$ . [23] shows that the + +cross-entropy of sharpened softmax logits is equivalent to a smoothed Triplet Loss [5] in a unit sphere. By minimizing the maximal entropy, we can pull $v_{i}^{\prime \tilde{k}}$ to their respective prototype $p^{\tilde{k}}$ and push it away from irrelevant prototypes $k \neq \tilde{k}$ . Accordingly, we learn the prototype vector $p^{\tilde{k}}$ by minimizing the term: + +$$ +\ell_ {\mathrm {c o n}} \left(v _ {i} ^ {\tilde {k}}\right) = - \log \left(\frac {\exp \left(\lambda \left\langle p ^ {\tilde {k}} , v _ {i} ^ {\prime \tilde {k}} \right\rangle\right)}{\sum_ {k \in K} \exp \left(\lambda \left\langle p ^ {k} , v _ {i} ^ {\prime \tilde {k}} \right\rangle\right)}\right), \tag {5} +$$ + +where $\lambda > 1$ is the scaling factor to sharpen the softmax logits, $p$ and $v'$ are $L_2$ normalized. + +Multiple prototypes per concept. Due to the multi-modal nature of data where one class contains several clusters rather than just a single one, we initialize $M$ prototypes per concept $\{p^{k_m}\}_{m=1}^M$ such that $M \ll N$ . + +The similarity of $v_{i}^{\prime k}$ with its concept $c^{\tilde{k}}$ is defined as the similarity to the nearest $p^{\tilde{k}_m}$ , and can be approximated by summing the compatibility with the $M$ clusters prototypes $p^{\tilde{k}_m}$ . Specifically, we first compute the probabilistic assignment $\mathbf{q}_m^{\tilde{k}}(v_i^{\prime k})$ to the prototypes: + +$$ +\mathrm {q} _ {m} \left(v ^ {\prime} _ {i} ^ {\tilde {k}}\right) = \operatorname {s o f t m a x} _ {m} \left(\gamma \left\langle p ^ {\tilde {k} _ {m}}, v ^ {\prime} _ {i} ^ {\tilde {k}} \right\rangle\right) \tag {6} +$$ + +where $\mathbf{q}_m$ is the assignment distribution of the $M$ prototypes in $c^{\tilde{k}}$ and $\gamma > 1$ is the scaling factor. The similarity of the concept vector with each prototype is obtained by weighing their cosine similarity with the assignment: + +$$ +\sin^ {\tilde {k} _ {m}} \left(v ^ {\prime} _ {i} ^ {\tilde {k}}\right) = q _ {m} \left(v ^ {\prime} _ {i} ^ {\tilde {k}}\right) \cdot \langle p ^ {\tilde {k} _ {m}}, v ^ {\prime} _ {i} ^ {\tilde {k}} \rangle . \tag {7} +$$ + +![](images/d7c77481e4c51259d549d7bbc9ebb56ba373de972fe2cbbd39ca51df624ebce4.jpg) + +![](images/2591fad17d9119258290398644c4790e45d2af2858d85b853a4982e1412096f9.jpg) +(a) Similarity maps generated by the local concept vector (first row) and concept prototype (second row) on the same image (a) and new images (b), (c). The local concept vector fails to highlight the concept pacemaker on new images while the concept prototype generalizes well with better localization. +Figure 4. Contrastive learning improves the concept feature space. (a) Qualitative and (b) quantitative comparison. + +![](images/70db739c08f16eec8b83a0697520ae4fde48eff76476b472dd84a8b6f57e64c6.jpg) +(b) Comparison between the original $v$ and projected feature space $v^{\prime}$ . The box plots of the similarity score distribution between $v_{i}^{k}$ and $v_{j\neq i}^{k^{\prime}}$ . Intra-Concept: $k = k^{\prime}$ , Inter-Concept: $k\neq k^{\prime}$ . The projected feature space is more compact and has better inter-concept separation. + +After that, we can obtain the similarity with concept $c^{\tilde{k}}$ by summing the compatibility with each cluster prototype: + +$$ +\sin^ {\tilde {k}} \left(v _ {i} ^ {\prime} \tilde {k}\right) = \sum_ {m} ^ {M} \sin^ {\tilde {k} _ {m}} \left(v _ {i} ^ {\prime} \tilde {k}\right). \tag {8} +$$ + +Finally, we compute the contrastive loss by maximizing the similarity of a concept vector with the cluster prototypes of the positive concept, similar to Eqn. 5: + +$$ +\ell_ {\text {c o n - m}} \left(v _ {i} ^ {\prime \tilde {k}}\right) = - \log \left(\frac {\exp \left(\lambda \left(\sin^ {\tilde {k}} \left(v _ {i} ^ {\prime \tilde {k}}\right) + \delta\right)\right)}{\sum_ {k \in K} \exp \left(\lambda \sin^ {k} \left(v _ {i} ^ {\prime \tilde {k}}\right)\right)}\right), \tag {9} +$$ + +$$ +\text {i f} \quad k = \tilde {k}, \quad \operatorname {s i m} ^ {k} \left(v _ {i} ^ {\prime \tilde {k}}\right) = \operatorname {s i m} ^ {k} \left(v _ {i} ^ {\prime \tilde {k}}\right) + \delta . +$$ + +A small margin $\delta$ between concept prototypes is added to further expand the decision boundaries. As a result, we + +learn a set of concept prototypes $\{p^{k_m}\}^{M\cdot K}$ that is significantly smaller in numbers compared to $\{v_i^k\}^{N\cdot K}$ , and the projector $P$ that enhances the inter-concept separation and intra-concept compactness. We illustrate the learning process of $P$ and $\{p\}$ in Fig. 3. + +From Eqn. 2, $v_{i}^{k}$ is a soft-selected feature patch of concept $c^k$ from the original feature $\mathbf{f}_i$ , we can directly use $P$ to map feature patch $\mathbf{f}_i(h,w)$ to the compact and discriminative prototypical feature space. We then obtain the similarity of the concept prototypes with $x$ from Eqn. 3: + +$$ +s ^ {k _ {m}} = \max _ {h, w} S ^ {k _ {m}} (h, w) = \max _ {h, w} \langle p ^ {k _ {m}}, P (\mathbf {f} (h, w)) \rangle , \tag {10} +$$ + +which are aggregated to create the similarity vector $[s^{k_m}]$ in Eqn. 1. We quantitatively and qualitatively present the improved representation of $p$ in Fig. 4. + +Training details. Initially, we train the Concept model for multi-label classification using binary cross-entropy, predicting one-hot encoded vector of ground truth concepts. Then, we obtain the local concept vectors and learn the concept prototypes and the projection head with a contrastive loss $\ell_{\mathrm{con - m}}$ to enhance concept generalization. Following, the task head is trained to predict the target from similarity scores with the concept prototypes, using cross-entropy. + +# 3. Doctor-in-the-loop interactivity + +# 3.1. Model explanations + +Given an input image $x$ , the model returns output $\hat{y}$ and the explanations for its prediction that include: + +- The similarity maps $[S^{k_m}]$ to show the activated region of the concept prototype $p^{k_m}$ on $x$ . +- The corresponding prototype image $\mathcal{I}(p^{k_m})$ as a reference, also highlighting the activated regions of $p^{k_m}$ . +- The concept similarity scores $[s^{k_m}]$ , indicating the similarity of the two activated regions of $p^{k_m}$ on the input image $x$ and on the reference image $\mathcal{I}(p^{k_m})$ . + +The model explanations correspond to the similarity-based inference logic outlined in Sec. 2.1. The operator $\mathcal{I}$ retrieves the image associated with the prototype $p^{k_m}$ . We replace each $p^{k_m}$ with the nearest projected local concept vector $v_{i}^{k}$ and link it to its source image $x_{i}$ : + +$$ +\mathcal {I} \left(p ^ {k _ {m}}\right) = x \leftarrow \underset {i} {\operatorname {a r g m i n}} \left\langle v _ {i} ^ {\prime k}, p ^ {k _ {m}} \right\rangle . \tag {11} +$$ + +# 3.2. Train-time interaction + +After we learn the concept prototypes $\{p^{k_m}\}$ and its associated human-interpretable version $\{\mathcal{I}(p^{k_m})\}$ , or atlas, doctors can inspect each concept prototype and discard any unqualified examples. This process can reveal and mitigate the Clever-Hans effect in the Concept model by examining the clues it is taking to predict the concepts. If a given prototype is shown to be a shortcut rather than a meaningful signal, it + +![](images/743ff4dafff8b227b2e4359779ac766a71eca65afb134da1de3f56c6289d2b80.jpg) +Figure 5. The irrelevant concept prototypes to be discarded in train-time interaction, highlighted at the activated regions. + +is removed from $\{p^{k_m}\}$ and is not used to calculate the similarity scores $[s^{k_m}]$ for the prediction of $y$ , effectively nullify the Clever-Hans effect. We show the irrelevant examples of different concepts in Fig. 5. The similarity-based inference logic allows this interaction as we only need as much as the remaining relevant examples to predict the target $y$ . + +# 3.3. Test-time interaction + +At test time, we introduce two types of interaction to improve the prediction quality: Spatial-level interaction and Concept-level interaction. We illustrate this in Fig. 1. + +Concept-level interaction. Consider this scenario: While reviewing the explanations from CSR for a chest X-ray $x$ , the doctor notices that the model is considering a nodule on $x$ for a mass in the reference image with a high similarity score. Recognizing the incorrect comparison, the doctor informs the model that concept mass is absent in $x$ . The model suppresses its confidence in concept mass to calibrate its prediction. This concept-level interaction allows doctors to directly eliminate irrelevant concepts. + +Interactivity mechanism. When the doctor goes through $[s^{k_m}]$ , she can discard any invalid concept if presented. The model then sets $s^{k_m} = 0$ for those rejected concepts, effectively discounting them on the prediction $\hat{y}$ . Concept-level interaction only considers the negative type of input from the doctor: concept rejection. Setting a suitable value for every $s_m^k$ of $c^k$ would be a overburden. Hence, for each concept $c^k$ , we only present the highest similarity score $\max_{m} \{s^{k_m}\}$ and have the doctor recognize its presence. We zero out all of the similarity scores of the rejected concept. + +Spatial-level interaction. As the doctor examines the image $x$ , upon spotting an important area, she can guide the model's attention with bounding boxes. The model then recalibrates its prediction $\hat{y}$ by prioritizing the indicated region based on this feedback. Conversely, reviewing the activation maps $[S^{k_m}]$ , the doctor also notices the model con + +centrating on an irrelevant region for prediction. To counteract the Clever-Hans effect, she marks the area as spurious, prompting the model to disregard it in its updated prediction. Hence, the doctor only needs to direct the model where to focus, without specifying what is in that region. This interaction scheme is practical and avoids the doctor from inadvertently becoming a teacher rather than taking advantage of the model. + +Interactivity mechanism. As we described in the user journey above, the spatial-level interaction allows two types of input from the doctor: (i) the positive boxes $\{\mathsf{bb}^{+} = (\mathsf{x}_1,\mathsf{y}_1,\mathsf{x}_2,\mathsf{y}_2)\}$ for the focus areas, (ii) the negative boxes $\{\mathsf{bb}^{-}\}$ for the regions to ignore. To emphasize the area specified by $\{\mathsf{bb}^{+}\}$ and suppress the area specified by $\{\mathsf{bb}^{-}\}$ , the model reweights each element $S(h,w)$ with an importance map $A\in \mathbb{R}^{\mathcal{H}\times \mathcal{W}}$ : + +$$ +A (h, w) = \left\{ \begin{array}{l l} 1 & \text {i f} (h, w) \in \left\{\mathrm {b b} ^ {+} \right\} \\ 0 & \text {i f} (h, w) \in \left\{\mathrm {b b} ^ {-} \right\}, \\ \alpha & \text {o t h e r w i s e} \end{array} \right. \tag {12} +$$ + +where $\alpha \in [0,1)$ is the neutral weight for the other regions. Setting $\alpha = 0$ indicates that the doctor only considers the regions $\{\mathrm{bb}^{+}\}$ , while setting $\alpha \rightarrow 1$ diminishes the importance of $\{\mathrm{bb}^{+}\}$ . The new set of similarity maps $[\hat{S}^{k_m}]$ is obtained by multiplying every similarity map of $[S^{k_m}]$ with the importance map $A$ : + +$$ +[ \hat {S} ] = A \odot [ S ] = [ A \odot S ^ {1 _ {1}}, \dots , A \odot S ^ {K _ {M}} ], \tag {13} +$$ + +where $\odot$ denotes the element-wise multiplication on the spatial dimension. As CSR is designed to select the maximum element on $S^{k_m}$ as $s^{k_m}$ , giving more weight to the important region $S(h,w)$ increases its likelihood to be selected. On the other hand, zeroing out the spurious region nullifies its influence on the similarity vector. Because $S$ is bounded within $[-1,1]$ and $A$ is bounded within $[0,1]$ , we clip negative values of $S$ to condition the model on only non-negative values. This ensures that the $\{\mathrm{bb}^+\}$ and $\{\mathrm{bb}^-\}$ regions are monotonically adjusted. The new prediction is made with $\hat{S}$ by the same inference logic as described in Sec. 2.1. + +Discussion. Combining both spatial and concept interaction mechanisms enables a flexible interaction approach. For example, if the doctor identifies an important region, she can draw a positive box $\mathsf{bb}^{+}$ over it. However, if the model associates the feature within $\mathsf{bb}^{+}$ with a wrong concept $c^k$ , placing $\mathsf{bb}^{+}$ over it would amplify the error. Because $c^k$ is absent in $x$ , rejecting it while specifying $\mathsf{bb}^{+}$ will instruct the model to focus on other concepts $\{c^{j\neq k}\}$ within $\mathsf{bb}^{+}$ and ignore $c^k$ . + +# 4. Related works + +Concept-based. Concept Bottleneck Model (CBM) [11] predicts the concepts before using them as support for the target prediction. Post-hoc Concept Bottleneck Models (PHCBM) [16] transfer concepts from other datasets or from natural language descriptions of concepts via multimodal models like CLIP [25]. Yang et al. leverages Large Language Models to generate an exhaustive list of concepts for each class [36]. These multimodal concept-based approaches [15, 22, 36] utilize foundation models to enhance their capabilities. However, foundation models are generally trained on broad domains, which presents limitations when applying them directly to medical imaging. In this work, we focus on the research direction of vision features and explore multi-modality learning in future studies. + +Part-prototypes. ProtoPNet [2] used prototypical parts to make a final classification based on positive attributions. ProtoTree[17] is a globally interpretable model by design that uses prototypes with decision trees. ProtoPool [28] is similar to ProtoPNet with a tweak of the shared prototypes pool and the Gumbel-softmax trick for the differentiable property. PIP-Net[18] is a patch-based approach with a novel regularization method for learning prototype similarity that correlates with human visual perception. Unlike previous methods that correlate prototypes with target classes, CSR explicitly grounds prototypes on interpretable concepts, making them semantically relevant. + +Interactivity. CBM [11] enables local interaction as human adjusts the concept logits for individual samples, while [16] allows global interaction by modifying concept weights of the model. ProtoPDebug [1] and [6] focus on part-prototype interaction, refining prototypes for enhanced performance. Besides logits-level interaction like those methods, CSR introduces spatial-level interaction, which is particularly intuitive and beneficial in the medical imaging domain, where detailed image analysis is of significant interest to doctors. + +# 5. Experiments and Results + +In this paper, we experiment on two chest X-ray datasets: TBX11K [13], VinDr-CXR[20] and a skin imaging dataset ISIC[4]. From those datasets, we define the findings annotations as the concepts and the diseases annotations as the target classes. Details on datasets and the training hyperparameters are in the supplementary. We qualitatively and quantitatively demonstrate the effectiveness of the concept prototypes learning framework in Fig. 4. + +Diagnostic performance. Due to the heavy class imbalance, we use macro F1-score as the metric for our experiments. We report the results in Table 1. It can be seen that CSR demonstrates a strong diagnostic performance compared to other baselines across all dataset. We also benchmark the performance of CSR when we ask a doctor to re + +
MethodF1↑#pro.↓#exp.↓
TBX11KCBM (ind.)24.514
CBM (seq.)60.514
CBM (joint.)88.614
ProtoPNet89.3600600
ProtoPNet89.315001500
ProtoPNet94.130003000
ProtoPool90.7600600
ProtoTree(H=6)94.01276
ProtoTree(H=9)93.75119
PIP-Net94.0768158
CSR94.4140014
CSR (refined)94.0138614
VinDr-CXRCBM (joint.)50.114
ProtoPNet41.415001500
PIP-Net45.17689
CSR54.6140014
ISICCBM (joint.)45.54
ProtoPNet38.5600600
ProtoPool31.0200200
ProtoTree(H=6)66.11276
PIP-Net69.976890
CSR71.54004
+ +Table 1. Diagnostic performance on the medical datasets, measured by Macro F1-score. CSR (refined) has the doctors review the concept atlas. #pro. denotes the number of prototypes of the model. #exp. indicates the explanation size for each prediction. + +
MethodProtoP-NetProto TreePIP-NetCBMCSRCSR (refined)
% PG hit8.87.819.555.160.979.5
+ +Table 2. Pointing Game hit rate on TBX11K, Tuberculosis class. CBM uses the class activation map of the highest contributing product to the class Tuberculosis. ProtoTree uses a similarity map from the top prototype at level-1 of the tree. ProtoPNet, PIP-Net, and CSR use the similarity map from the prototype with the highest contribution to class Tuberculosis. + +fine the learned concept atlas as described in Sec. 3.2. + +Explanation size. Following [18], we quantify the explanation size as the number of explanations per prediction for each method. For part-prototype methods, explanations refer to the similar parts of the image from the training set. ProtoTree has the explanation size equal to its depth. PIP-Net reduces the explanation size by regularizing the sparsity during training and only considers non-zero prototypes for each prediction. Other part-prototype approaches have the explanation size matching the large number of prototypes. This compromises interpretability since users must + +![](images/300a2b228ea54334149c3dfca4932232d3bbcf631d2b48f046e886f4132ba99c.jpg) +Figure 6. Qualitative result of the Pointing Game. We visualize the groundtruth bounding box and the maximum activation point of each method as the colored square. :Proposed CSR, :CBM, :PIP-Net, :ProtoPNet, :ProtoTree + +navigate through a multitude of explanations within the set for each prediction. Concept-based models use the concepts as their explanations, having an explanation size equal to the number of concepts, which is typically smaller. This choice involves an accuracy trade-off due to the reduced size of the bottleneck. In contrast, the proposed CSR benefits from both, featuring a small explanation size equivalent to CBM while achieving competitive accuracy akin to part-prototype methods. To achieve this, our CSR only considers concept grounded explanations and exclusively presents the highest similar prototype of each active concept. + +Trustworthiness. We inspect the explanations of each benchmark and demonstrate that interpretability does not always come with trustworthiness. A trustworthy model must deliver relevant explanations that correspond to human interpretable concepts. We leverage the pointing game (PG) [37] that considers a hit if the point of maximum value in the explanation lies within a segmentation ground truth. We use the bounding box annotation from the TBX11K and calculate the PG hit rate of the explanations. + +The quantitative results are reported in Table. 2. Partprototype methods including ProtoPNet and ProtoTree are relying on clues that considerably mismatch with the semantic ground truth. This observation aligns with [18], where Purity of the prototypes is evaluated as an alignment with the part annotations in the CUB dataset [32], demonstrating a large semantic gap between the prototypes and human-relevant concepts. Concept-based methods attain a more reasonable PG hit rate as the concepts are human-grounded and thus align well with the semantic ground truth. The proposed CSR faithfully aligns with human reasoning and achieves a remarkable PG hit rate of $60.9\%$ . Refining the concept atlas further improves the PG hit rate to $79.5\%$ , which hints that the better classification result of the original CSR in Table 1 could stem from shortcuts. The qualitative results in Fig. 6 shows that our CSR:□ more frequently identifies the maximum activation within + +
spatialconceptMac.F1+Δ%
CSR94.40.0
Non-expert✓94.50.1
Oracle✓94.90.5
M.D.✓94.70.3
M.D.✓95.10.7
M.D.✓✓94.90.5
+ +Table 3. Interactions performance gain. spatial: spatial-level interaction, concept: concept-level interaction. We limit the Non-export to spatial-interaction only. Oracle: using the bounding box ground truth of TB as spatial-interaction. M.D: medical student. + +the ground truth box compared to other methods. + +Interactivity. We demonstrate the benefits of interactivity by measuring the model when obtaining interaction at test time on the TBX11K dataset and report the result in Table. 3. Our experiment involves a Medical degree (M.D.) student and a non-expert in evaluating the framework. In the oracle setting, we use the ground truth bounding box of class TB from the dataset as a positive spatial-level interaction. Note that we only require human interaction when the output probability of the model is indecisive: when the difference of top-1 and top-2 class probability is less than 0.3. The weight $\alpha$ is set to 0.2. It shows that human interactions has a positive impact on the model outcome. The interaction from the M.D. student is more beneficial to the model than the non-expert. We consider that the non-experts can still eliminate irrelevant CXR regions without formal radiology expertise, and limit the non-expert to spatial-level interactions only. The concept-level interaction is more influential than the spatial-level interaction with better gain. + +# 6. Conclusion + +This paper introduces a novel Concept Similarity Reasoning (CSR), an interpretable and interactive model for medical image analysis with similarity reasoning. CSR provides local explanations by presenting similar concept examples with the input images. In comparison to part-prototype approaches, our CSR presents more concise and meaningful prototypes that align better with human-relevant concepts. This observation implies the differentiation between interpretability and trustworthiness. The proposed interaction mechanism employed during both training and testing effectively enhances model reliability by mitigating shortcuts and improves model accuracy through domain-specific inputs. Our method is an attempt to encourage more trust from doctors through a natural and effective human-AI interaction, improving both accuracy and transparency. + +# References + +[1] Andrea Bontempelli, Stefano Teso, Fausto Giunchiglia, and Andrea Passerini. Concept-level debugging of part-prototype networks. ArXiv, abs/2205.15769, 2022. 1, 2, 7 +[2] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: Deep learning for interpretable image recognition. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 1, 3, 7 +[3] Townim Faisal Chowdhury, Kewen Liao, Vu Minh Hieu Phan, Minh-Son To, Yutong Xie, Kevin Hung, David Ross, Anton van den Hengel, Johan W Verjans, and Zhibin Liao. Cape: Cam as a probabilistic ensemble for enhanced dnn interpretation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11072-11081, 2024. 1 +[4] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), pages 168-172. IEEE, 2018. 7 +[5] Xingping Dong and Jianbing Shen. Triplet loss in siamese network for object tracking. In Proceedings of the European conference on computer vision (ECCV), pages 459-474, 2018. 4 +[6] Michael Gerstenberger, Thomas Wiegand, Peter Eisert, and Sebastian Bosse. But that's not why: Inference adjustment by interactive prototype revision. In Iberoamerican Congress on Pattern Recognition, pages 123–132. Springer, 2023. 1, 2, 7 +[7] Alec Holt, Isabelle Bichindaritz, Rainer Schmidt, and Petra Perner. Medical applications in case-based reasoning. The Knowledge Engineering Review, 20(3):289-292, 2005. 1 +[8] Ta Duc Huy, Hoang Cao Huyen, Chanh DT Nguyen, Soan TM Duong, Trung Bui, and Steven QH Truong. Adversarial contrastive fourier domain adaptation for polyp segmentation. In 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2022. 1 +[9] Sajid Javed, Arif Mahmood, Iyyakutti Iyappan Ganapathi, Fayaz Ali Dharejo, Naoufel Werghi, and Mohammed Bennamoun. Cclip: Zero-shot learning for histopathology with comprehensive vision-language alignment. In CVPR, pages 11450-11459, 2024. 1 +[10] Injae Kim, Jongha Kim, Joonmyung Choi, and Hyunwoo J Kim. Concept bottleneck with visual concept filtering for explainable medical image classification. In The Medical Computing and Computer Assisted Intervention Society, pages 225–233. Springer, 2023. 1 +[11] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning, pages 5338-5348. PMLR, 2020. 1, 7 + +[12] Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. Unmasking clever hans predictors and assessing what machines really learn. Nature communications, 10(1):1096, 2019. 2 +[13] Yun Liu, Yu-Huan Wu, Yunfeng Ban, Huifang Wang, and Ming-Ming Cheng. Rethinking computer-aided tuberculosis diagnosis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2646–2655, 2020. 7 +[14] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology. Nature Medicine, 30(3):863-874, 2024. 1 +[15] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 7 +[16] Yuksekgonul Mert, Wang Maggie, and Zou James. Post-hoc concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. 2, 7 +[17] Meike Nauta, Ron van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14933-14943, 2021. 1, 7 +[18] Meike Nauta, Jörg Schlötterer, Maurice van Keulen, and Christin Seifert. Pip-net: Patch-based intuitive prototypes for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2744-2753, 2023. 1, 3, 7, 8 +[19] Giang Nguyen, Daeyoung Kim, and Anh Nguyen. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems, 34:26422-26436, 2021. 1 +[20] Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, et al. Vindr-cxr: An open dataset of chest x-rays with radiologist's annotations. Scientific Data, 9(1):429, 2022. 7 +[21] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023. 1 +[22] Vu Minh Hieu Phan, Yutong Xie, Yuankai Qi, Lingqiao Liu, Liyang Liu, Bowen Zhang, Zhibin Liao, Qi Wu, Minh-Son To, and Johan W Verjans. Decomposing disease descriptions for enhanced pathology detection: A multi-aspect vision-language pre-training framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11492-11501, 2024. 1, 7 +[23] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6450-6458, 2019. 4 +[24] Tran Minh Quan, Huynh Minh Thanh, Ta Duc Huy, Nguyen Do Trung Chanh, Nguyen Thi Phuong Anh, Phan Hoan Vu, Nguyen Hoang Nam, Tran Quy Tuong, Vu Minh Dien, Bui Van Giang, et al. Xpgan: X-ray projected generative adversarial network for improving Covid-19 image classification. + +In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 1509-1513. IEEE, 2021. 1 +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7 +[26] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence, 1(5):206-215, 2019. 1 +[27] Dawid Rymarczyk, Łukasz Struski, Jacek Tabor, and Bartosz Zielinski. Protopshare: Prototype sharing for interpretable image classification and similarity discovery. arXiv preprint arXiv:2011.14340, 2020. 1, 3 +[28] Dawid Rymarczyk, Łukasz Struski, Michal Górszczak, Koryna Lewandowska, Jacek Tabor, and Bartosz Zielinski. Interpretable image classification with differentiable prototypes assignment. In European Conference on Computer Vision, pages 351-368. Springer, 2022. 1, 3, 7 +[29] Tran Bao Sam, Ta Duc Huy, Cong Tuyen Dao, Thanh Tin Lam, Van Ha Tang, and Steven Q.H. Truong. A multi-phase multi-graph approach for focal liver lesion classification on ct scans. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 3788-3802, 2024. 1 +[30] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1 +[31] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 1 +[32] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 8 +[33] Bor-Shiun Wang, Chien-Yi Wang, and Wei-Chen Chiu. Mcpnet: An interpretable classifier via multi-level concept prototypes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10885-10894, 2024. 1 +[34] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 1 +[35] Siyuan Yan, Zhen Yu, Xuelin Zhang, Dwarikanath Mahapatra, Shekhar S Chandra, Monika Janda, Peter Soyer, and Zongyuan Ge. Towards trustable skin cancer diagnosis via rewriting model's decision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11568-11577, 2023. 2 +[36] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language + +in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 1, 7 +[37] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. International Journal of Computer Vision, 126(10):1084-1102, 2018. 8 +[38] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 1 \ No newline at end of file diff --git a/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/images.zip b/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..de11720d8548690320ecbe50a4cf554fa5cbe9e7 --- /dev/null +++ b/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f16170fad9617673f662009854ef0a5c041259dfbe109d191701b7613cf1255 +size 524718 diff --git a/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/layout.json b/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bb2af00e0254bfa7c4f97fdaf7346f93e99beb44 --- /dev/null +++ b/2025/Interactive Medical Image Analysis with Concept-based Similarity Reasoning/layout.json @@ -0,0 +1,10597 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "type": "text", + "content": "Interactive Medical Image Analysis with Concept-based Similarity Reasoning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 142, + 539, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 142, + 539, + 159 + ], + "spans": [ + { + "bbox": [ + 70, + 142, + 539, + 159 + ], + "type": "text", + "content": "Ta Duc Huy† Sen Kim Tran Phan Nguyen Nguyen Hoang Tran Tran Bao Sam" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 162, + 518, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 162, + 518, + 177 + ], + "spans": [ + { + "bbox": [ + 91, + 162, + 518, + 177 + ], + "type": "text", + "content": "Anton van den Hengel† Zhibin Liao† Johan W. Verjans† Minh-Son To†‡" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 252, + 182, + 362, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 182, + 362, + 196 + ], + "spans": [ + { + "bbox": [ + 252, + 182, + 362, + 196 + ], + "type": "text", + "content": "Vu Minh Hieu Phan†*" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 198, + 474, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 198, + 474, + 211 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 474, + 211 + ], + "type": "text", + "content": "† Australian Institute for Machine Learning, The University of Adelaide, ‡ Flinders University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 239, + 200, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 239, + 200, + 251 + ], + "spans": [ + { + "bbox": [ + 151, + 239, + 200, + 251 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 264, + 297, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 297, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 297, + 563 + ], + "type": "text", + "content": "The ability to interpret and intervene model decisions is important for the adoption of computer-aided diagnosis methods in clinical workflows. Recent concept-based methods link the model predictions with interpretable concepts and modify their activation scores to interact with the model. However, these concepts are at the image level, which hinders the model from pinpointing the exact patches the concepts are activated. Alternatively, prototype-based methods learn representations from training image patches and compare these with test image patches, using the similarity scores for final class prediction. However, interpreting the underlying concepts of these patches can be challenging and often necessitates post-hoc guesswork. To address this issue, this paper introduces the novel Concept-based Similarity Reasoning network (CSR), which offers (i) patch-level prototype with intrinsic concept interpretation, and (ii) spatial interactivity. First, the proposed CSR provides localized explanation by grounding prototypes of each concept on image regions. Second, our model introduces novel spatial-level interaction, allowing doctors to engage directly with specific image areas, making it an intuitive and transparent tool for medical imaging. CSR improves upon prior state-of-the-art interpretable methods by up to " + }, + { + "bbox": [ + 55, + 264, + 297, + 563 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 55, + 264, + 297, + 563 + ], + "type": "text", + "content": " across three biomedical datasets. Our code is released at https://github.com/tadeephuy/InteractCSR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 588, + 135, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 588, + 135, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 588, + 135, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "content": "Despite the success of deep neural networks in medical image analysis [8, 9, 14, 22, 24, 29], their \"black-box\" nature limits their adoption in safety-critical clinical settings. Attribution-based methods [3, 30, 31, 34, 38] generate saliency maps to identify influential image regions. However, recent studies raised concerns about the reliability and accuracy of visual interpretation, potentially compro" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 239, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 373 + ], + "type": "text", + "content": "mising decision-making [19, 26]. To address this, Concept Bottleneck Models (CBM) predict the interpretable concepts first and use them to predict the label [10, 11, 21, 36]. CBM allows human intervention in the predicted concepts, which is shown to significantly benefit performance. Alternatively, prototype-based methods are proposed to learn visual part prototypes to explain the classification decision. Specifically, ProtoPNet [2] and its derivatives [1, 6, 17, 18, 27, 28, 33] learn prototypical parts from training data and compare test images with these parts to compute similarity scores for final predictions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "text", + "content": "Despite their inherent interpretability, concept-based methods only offer image-level explanations by linking images to global concepts, without identifying specific local patches correlated with these concepts. In contrast, prototype-based methods deliver patch-level explanations but necessitate post-hoc analysis to correlate visual patches with semantically meaningful concepts. This challenge intensifies in medical imaging, where subtle visual medical findings make concept assignments particularly difficult. How to create concept-based models that enable patch-level explanation for medical imaging is an open-ended question." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "In radiological imaging, radiologists routinely reference an atlas of visual examples, depicting distinctive concepts of a pathology, such as opacity, fluid presence in the lung, enlarged heart shadow to aid diagnosis. Inspired by this observation, we propose a novel Concept-based Similarity Reasoning network, which offers two unique contributions. First, our CSR gives patch-level and interpretable-by-design concepts without requiring post-hoc analysis, in contrast to prototype-based methods [2, 18, 28]. We classify a disease by searching locally for the presence of visual exemplars of its concepts on the image. For example, a cancer diagnosis on chest X-ray could involve a radiologist comparing a suspected tumor with other known malignant tumors to assess its severity [7]. The presence of the tumor is a concept to predict the target cancer. Our framework trains a model to search locally for the presence of similar tumor exemplars (or prototypes) on every patch of the im" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30797" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 71, + 526, + 288 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 526, + 288 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 526, + 288 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 526, + 288 + ], + "type": "image", + "image_path": "b2f63d2c0aae5676106a1db2668e72ae4e62048251fb4b657029d5a9d665224a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 296, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 54, + 296, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 296, + 555, + 373 + ], + "type": "text", + "content": "Figure 1. Diagnosis Workflow with Doctor-in-the-Loop using CSR. Top section: CSR predicts Pulmonary Edema, falsely associating with the two concepts lung fluid and enlarged heart. For each concept, CSR explains its prediction by comparing the input with the highlighted region on the prototype image to create the corresponding similarity map. Bottom section: As the doctor inspects the similarity maps of each concept, he suppresses incorrect lung fluid attention, and reinforces the attention on opacity on the left of the image through spatial-interaction: “drawing positive and negative boxes” to create an importance map indicating where to focus and ignore. Secondly, noting the heart is normal, he “rejects” the enlarged heart concept via concept-interaction. CSR then recalibrates its prediction to Tuberculosis, aligning with the observed opacity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 394, + 294, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 394, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 54, + 394, + 294, + 431 + ], + "type": "text", + "content": "age. After training, our model explicitly learns interpretable concept-grounded patch prototypes without a post-hoc analysis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 434, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 434, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 54, + 434, + 295, + 577 + ], + "type": "text", + "content": "Second, our model enables spatial interactivity on the patch level. Unlike previous methods which only intervene in model predictions on the concepts-level outputs [1, 6, 16, 35], we introduce spatial interactivity on the input image. Ours offers intuitive benefits in medical imaging, where precise image analysis is crucial for clinicians. Deep learning models frequently capture spurious shortcuts correlations rather than the intended signals—a phenomenon known as the Clever-Hans effect [12]. Therefore, integrating doctor feedback to guide the model on relevant cues significantly enhances trustworthiness in clinical settings. Our contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 654 + ], + "type": "text", + "content": "- We propose a novel Concept-based Similarity Reasoning network (CSR) that provides concept-level explanation. To the best of our knowledge, this is the first interpretable-by-design method that offers localized concept explainability for medical images without post-hoc analysis like part-prototype approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "- We introduce a novel spatial interaction mechanism (where to look), alongside train-time interaction (what not to learn) and concept-level interaction (what not present). CSR facilitates a comprehensive interactive framework that actively involves doctors in both train and test phases." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 394, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 555, + 441 + ], + "type": "text", + "content": "- We conducted extensive experiments on three datasets and showed that CSR achieves a significant improvement by up to " + }, + { + "bbox": [ + 314, + 394, + 555, + 441 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 314, + 394, + 555, + 441 + ], + "type": "text", + "content": " F1-score compared to other interpretable methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 452, + 370, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 452, + 370, + 464 + ], + "spans": [ + { + "bbox": [ + 314, + 452, + 370, + 464 + ], + "type": "text", + "content": "2. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 472, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 556 + ], + "type": "text", + "content": "CSR comprises (i) a Concept model to extract interpretable concept features; (ii) a feature projector " + }, + { + "bbox": [ + 313, + 472, + 555, + 556 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 472, + 555, + 556 + ], + "type": "text", + "content": " to enhance the concept feature space; and (iii) a task head " + }, + { + "bbox": [ + 313, + 472, + 555, + 556 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 472, + 555, + 556 + ], + "type": "text", + "content": " to classify based on concept similarity scores. This section describes the inference logic of the proposed Concept-based Similarity Reasoning model and the prototypes learning framework. Fig. 2 illustrates the inference process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 563, + 508, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 508, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 508, + 575 + ], + "type": "text", + "content": "2.1. Concept-based Similarity Reasoning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "text", + "content": "Concept-based classification. From an input " + }, + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "text", + "content": ", concept-based methods first predict the scores of " + }, + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "text", + "content": " explainable concepts " + }, + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "inline_equation", + "content": "\\{c^k\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "text", + "content": " and then use them as support to predict the target " + }, + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "text", + "content": ". Instead of predicting the concept scores, CSR calculates the similarity scores between " + }, + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "text", + "content": " and the examples (or prototypes) of each concept. It then uses the concept similarity scores to support the prediction of " + }, + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 580, + 554, + 663 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 664, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 554, + 714 + ], + "type": "text", + "content": "Similarity reasoning. To construct the similarity reasoning inference logic, we establish an atlas: a set of " + }, + { + "bbox": [ + 313, + 664, + 554, + 714 + ], + "type": "inline_equation", + "content": "M \\cdot K" + }, + { + "bbox": [ + 313, + 664, + 554, + 714 + ], + "type": "text", + "content": " concept prototypes " + }, + { + "bbox": [ + 313, + 664, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\{p^{k_m}\\}_{k=1,m=1}^{K,M}" + }, + { + "bbox": [ + 313, + 664, + 554, + 714 + ], + "type": "text", + "content": ", where we have " + }, + { + "bbox": [ + 313, + 664, + 554, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 664, + 554, + 714 + ], + "type": "text", + "content": " prototypes for each concept " + }, + { + "bbox": [ + 313, + 664, + 554, + 714 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 313, + 664, + 554, + 714 + ], + "type": "text", + "content": ". In essence, " + }, + { + "bbox": [ + 313, + 664, + 554, + 714 + ], + "type": "inline_equation", + "content": "p^{k_m} \\in \\mathbb{R}^{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 664, + 554, + 714 + ], + "type": "text", + "content": " repre" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30798" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 71, + 531, + 205 + ], + "blocks": [ + { + "bbox": [ + 86, + 71, + 531, + 205 + ], + "lines": [ + { + "bbox": [ + 86, + 71, + 531, + 205 + ], + "spans": [ + { + "bbox": [ + 86, + 71, + 531, + 205 + ], + "type": "image", + "image_path": "2d9f88b37c24908d874e5d22ff8902cd6e558d93a82ce4c0dd4150fb4c60d914.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 213, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 54, + 213, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 555, + 260 + ], + "type": "text", + "content": "Figure 2. Inference logic of CSR. Each concept prototype represents a specific concept from a training image. CSR generates 2D similarity maps by computing the cosine similarity between these concept prototypes and the feature maps, and considers the maximum values as similarity scores to calculate the prediction logits. The prototype image refers to the training image associated with a specific concept prototype as detailed in Sec. 3.1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": "sents the " + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "inline_equation", + "content": "m^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": " example of concept " + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": ". Given an input image " + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": ", we use the feature extractor " + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": " to create a feature map " + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": " having " + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{H} \\cdot \\mathcal{W}" + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": " feature patches. The similarity score " + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "inline_equation", + "content": "s^{k_m}" + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": " is defined as the maximum cosine similarity between the prototype " + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "inline_equation", + "content": "p^{k_m}" + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": " and each feature patch " + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(h, w)" + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": ", quantifying the presence of " + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "inline_equation", + "content": "p^{k_m}" + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": ". As such, we can aggregate the presence of " + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "inline_equation", + "content": "M \\cdot K" + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": " concept prototypes " + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\{p^{k_m}\\}" + }, + { + "bbox": [ + 55, + 279, + 296, + 376 + ], + "type": "text", + "content": " into a similarity scores vector:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 384, + 295, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 384, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 67, + 384, + 295, + 401 + ], + "type": "interline_equation", + "content": "s = \\left[ s ^ {k _ {m}} \\right] _ {k = 1, m = 1} ^ {K, M} = \\left[ s ^ {1 _ {1}}, s ^ {1 _ {2}}, \\dots , s ^ {K _ {M - 1}}, s ^ {K _ {M}} \\right]. \\tag {1}", + "image_path": "8224cff02e73246f4298b777dbcdc4289b884b12bd9f7a5e68190926d8a731e2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "text", + "content": "Then, the model predicts the target " + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "text", + "content": " from the similarity vector " + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "text", + "content": " with a task head " + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "inline_equation", + "content": "y = H(s)" + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "text", + "content": ". For example, the input chest X-ray is similar to several images from the atlas that have: fluid build-up in the left lung (score 0.85), and the abnormal heart size (score 0.88). Based on these similarities, the model predicts target class Pulmonary Edema. In short, the model classifies " + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "text", + "content": " by comparing exemplars " + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "inline_equation", + "content": "p^{k_m}" + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "text", + "content": " of the interpretable concept " + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "text", + "content": " on each local patch on the image " + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "text", + "content": ". The choice of " + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "text", + "content": " should be but not be limited to a whitebox model such as a single linear layer or a shallow tree so that the reasoning from " + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "inline_equation", + "content": "s \\to y" + }, + { + "bbox": [ + 55, + 410, + 296, + 553 + ], + "type": "text", + "content": " maintains interpretability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 554, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 639 + ], + "type": "text", + "content": "Discussion. Prototype-based methods [2, 18, 27, 28] also learn a set of exemplar prototypes " + }, + { + "bbox": [ + 55, + 554, + 296, + 639 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 554, + 296, + 639 + ], + "type": "text", + "content": ". However, they require post-hoc analysis by experts to associate the interpretable concept labels to the learned prototypes, which is intractable in the fine-grained medical image analysis. In stark contrast, the proposed model learns patch-level and interpretable concept prototypes by design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 647, + 212, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 212, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 212, + 661 + ], + "type": "text", + "content": "2.2. Concept prototypes learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "This section describes the proposed methods to learn the concept prototypes. Overall, our method consists of a Concept model to learn interpretable concept vectors, and a novel contrastive learning objective to enhance generaliz" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 279, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 553, + 304 + ], + "type": "text", + "content": "ability of concept vectors. The overview of our learning framework is shown in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": "Concept model. The Concept model comprises the feature extractor " + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": ", and the concept head " + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": " that maps " + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": " into the one hot concept vector " + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "inline_equation", + "content": "c = C(\\mathbf{f})" + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": ", indicating the presence of concepts being associated with " + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": ". The concept head " + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": " includes a " + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": " convolution layer that maps " + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": " channels: " + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C \\times \\mathcal{H} \\times \\mathcal{W}} \\to \\mathbb{R}^{K \\times \\mathcal{H} \\times \\mathcal{W}}" + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": ", representing the class activation maps " + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{cam}^k \\in \\mathbb{R}^{\\mathcal{H} \\times \\mathcal{W}}\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 304, + 554, + 399 + ], + "type": "text", + "content": " concepts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "text", + "content": "Local concept vectors acquisition. If the region " + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "inline_equation", + "content": "(h, w)" + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "text", + "content": " contains the visual characteristics of concept " + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "text", + "content": ", we consider the patch feature " + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(h, w)" + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "text", + "content": " important and select it as the local concept vector. We first normalize " + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{cam}^k" + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "text", + "content": " using spatial softmax. Then, we soft-select the local concept vector " + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "inline_equation", + "content": "v^k \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "text", + "content": " for concept " + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "text", + "content": " by weighting " + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{cam}^k" + }, + { + "bbox": [ + 313, + 400, + 554, + 485 + ], + "type": "text", + "content": " and summing the products spatially:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 372, + 494, + 553, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 494, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 372, + 494, + 553, + 524 + ], + "type": "interline_equation", + "content": "v ^ {k} = \\sum_ {h, w} ^ {\\mathcal {H}, \\mathcal {W}} \\operatorname {s o f t m a x} _ {h, w} \\left(\\operatorname {c a m} ^ {k}\\right) \\cdot \\mathbf {f}. \\tag {2}", + "image_path": "3063ccc82bed856d60de14630469554bb464784f9080dfdcac299aac12c9b994.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 553, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 559 + ], + "type": "text", + "content": "With a training set of " + }, + { + "bbox": [ + 313, + 534, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}^{N}" + }, + { + "bbox": [ + 313, + 534, + 553, + 559 + ], + "type": "text", + "content": " images of " + }, + { + "bbox": [ + 313, + 534, + 553, + 559 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 534, + 553, + 559 + ], + "type": "text", + "content": " concepts, we generate a set " + }, + { + "bbox": [ + 313, + 534, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\{v_{i}^{k}\\}_{i = 1}^{V\\leq N\\cdot K}" + }, + { + "bbox": [ + 313, + 534, + 553, + 559 + ], + "type": "text", + "content": " of local concept vectors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 559, + 554, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 554, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 554, + 608 + ], + "type": "text", + "content": "Similarity maps as explanations. Based on the local concept vector, we construct a 2D similarity map for concept " + }, + { + "bbox": [ + 313, + 559, + 554, + 608 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 313, + 559, + 554, + 608 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 559, + 554, + 608 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 559, + 554, + 608 + ], + "type": "text", + "content": " by evaluating the cosine similarity between each patch feature " + }, + { + "bbox": [ + 313, + 559, + 554, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(h,w)" + }, + { + "bbox": [ + 313, + 559, + 554, + 608 + ], + "type": "text", + "content": " and vector " + }, + { + "bbox": [ + 313, + 559, + 554, + 608 + ], + "type": "inline_equation", + "content": "v^k" + }, + { + "bbox": [ + 313, + 559, + 554, + 608 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 616, + 553, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 616, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 340, + 616, + 553, + 632 + ], + "type": "interline_equation", + "content": "S ^ {k} (h, w) = \\cos \\left(v ^ {k}, \\mathbf {f} (h, w)\\right) = \\langle v ^ {k}, \\mathbf {f} (h, w) \\rangle , \\tag {3}", + "image_path": "c3517f7346b4f543f8766b767e28160c1a1b9245554290caaa452c739c3c8d1e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\langle ,\\rangle" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " is the dot product of the two normalized vectors. As " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "v^{k}" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " is the importance weighted sum of " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": ", comparing it with the important patch " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(h,w)" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " results in a high similarity score at " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "S^k (h,w)" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": ", and vice versa. Effectively, " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "S^k\\in \\mathbb{R}^{\\mathcal{H}\\times \\mathcal{W}}" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " is the local explanation for concept " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": ". The similarity score of " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " with the concept " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30799" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 264 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 264 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 264 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 264 + ], + "type": "image", + "image_path": "b4c33f0a7af0b7a5a8c6a0fc97c9e30b4121993928f6dea4a95e82b4f3718b65.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 271, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 54, + 271, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 555, + 327 + ], + "type": "text", + "content": "Figure 3. The novel Concept prototypes learning framework. We pretrain a Concept model to generate concept activation maps. The local concept vectors are then generated by weighting and summing the feature maps with the activation map of the corresponding concept. To enhance the compactness and the generalizability of the concept feature space, we introduce a novel multi-prototype learning objective. After projecting concept vectors via a projector " + }, + { + "bbox": [ + 54, + 271, + 555, + 327 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 54, + 271, + 555, + 327 + ], + "type": "text", + "content": ", the proposed objective pulls concept features to its nearest concept prototypes, while pushing away from prototypes of other concepts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 347, + 259, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 259, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 259, + 360 + ], + "type": "text", + "content": "defined as the maximum of the similarity map " + }, + { + "bbox": [ + 55, + 347, + 259, + 360 + ], + "type": "inline_equation", + "content": "S^k" + }, + { + "bbox": [ + 55, + 347, + 259, + 360 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 364, + 294, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 294, + 384 + ], + "type": "interline_equation", + "content": "s ^ {k} = \\max _ {h, w} S ^ {k} (h, w). \\tag {4}", + "image_path": "643f87e132777ddad8212a7b8bedbdbaffad877f7fd743521023186c60e7ef98.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": "After obtaining the similarity scores " + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "inline_equation", + "content": "s_i^k" + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": " for the local concept vectors " + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "inline_equation", + "content": "\\{v_i^k\\}" + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": " with the input " + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": ", one can naively adopt the inference logic in Sec. 2.1. However, this leads to two problems. First, the computation complexity is " + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "inline_equation", + "content": "O(N \\cdot K)" + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": ". Second, " + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "inline_equation", + "content": "v_i^k" + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": " is local, which is specific to each image, thus non-generalizable to new images. We investigate the feasibility of this approach by comparing " + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "inline_equation", + "content": "v_i^k" + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": " with new images " + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "inline_equation", + "content": "v_{j \\neq i}^{k'}" + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": " and analyze the similarity scores distribution of intra-concept " + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "inline_equation", + "content": "(k = k')" + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": " and inter-concept " + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "inline_equation", + "content": "(k \\neq k')" + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": ". It is observed that the concepts are not well-represented in new images. In other words, a local concept vector " + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "inline_equation", + "content": "v_i^k" + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": " generated from an image " + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": " fails to highlight the concept " + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": " on a new image " + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "inline_equation", + "content": "x_{j \\neq i}" + }, + { + "bbox": [ + 55, + 389, + 296, + 545 + ], + "type": "text", + "content": ", as shown in Fig. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 296, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 630 + ], + "type": "text", + "content": "Concept prototypes learning. To improve the generalizability of concept representation " + }, + { + "bbox": [ + 55, + 545, + 296, + 630 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 545, + 296, + 630 + ], + "type": "text", + "content": " and reduce the number of concept vectors, our solution involves: (i) employing contrastive learning to enhance the shared embedding space between " + }, + { + "bbox": [ + 55, + 545, + 296, + 630 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 545, + 296, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 296, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 55, + 545, + 296, + 630 + ], + "type": "text", + "content": "; (ii) downsampling the concept-specific-and-sample-specific set " + }, + { + "bbox": [ + 55, + 545, + 296, + 630 + ], + "type": "inline_equation", + "content": "\\{v_i^k\\}" + }, + { + "bbox": [ + 55, + 545, + 296, + 630 + ], + "type": "text", + "content": " to a small core set of prototypes " + }, + { + "bbox": [ + 55, + 545, + 296, + 630 + ], + "type": "inline_equation", + "content": "\\{p^k\\}" + }, + { + "bbox": [ + 55, + 545, + 296, + 630 + ], + "type": "text", + "content": " that is only specific to each concept, for efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "text", + "content": "We use a projector " + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "text", + "content": " to map " + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "text", + "content": " to a feature space with the goal to improve inter-concept separation and intra-concept compactness such that " + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "inline_equation", + "content": "v' = P(v)" + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "text", + "content": " via contrastive learning. Then, we mine prototypes " + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "inline_equation", + "content": "p^k" + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "text", + "content": " for each concept, which can be a single prototype or a diverse set of prototypes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Single prototype per concept. We initialize a prototype vector " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "p^k \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " for each concept " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ". [23] shows that the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 347, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 426 + ], + "type": "text", + "content": "cross-entropy of sharpened softmax logits is equivalent to a smoothed Triplet Loss [5] in a unit sphere. By minimizing the maximal entropy, we can pull " + }, + { + "bbox": [ + 313, + 347, + 555, + 426 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime \\tilde{k}}" + }, + { + "bbox": [ + 313, + 347, + 555, + 426 + ], + "type": "text", + "content": " to their respective prototype " + }, + { + "bbox": [ + 313, + 347, + 555, + 426 + ], + "type": "inline_equation", + "content": "p^{\\tilde{k}}" + }, + { + "bbox": [ + 313, + 347, + 555, + 426 + ], + "type": "text", + "content": " and push it away from irrelevant prototypes " + }, + { + "bbox": [ + 313, + 347, + 555, + 426 + ], + "type": "inline_equation", + "content": "k \\neq \\tilde{k}" + }, + { + "bbox": [ + 313, + 347, + 555, + 426 + ], + "type": "text", + "content": ". Accordingly, we learn the prototype vector " + }, + { + "bbox": [ + 313, + 347, + 555, + 426 + ], + "type": "inline_equation", + "content": "p^{\\tilde{k}}" + }, + { + "bbox": [ + 313, + 347, + 555, + 426 + ], + "type": "text", + "content": " by minimizing the term:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 338, + 430, + 555, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 430, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 338, + 430, + 555, + 464 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathrm {c o n}} \\left(v _ {i} ^ {\\tilde {k}}\\right) = - \\log \\left(\\frac {\\exp \\left(\\lambda \\left\\langle p ^ {\\tilde {k}} , v _ {i} ^ {\\prime \\tilde {k}} \\right\\rangle\\right)}{\\sum_ {k \\in K} \\exp \\left(\\lambda \\left\\langle p ^ {k} , v _ {i} ^ {\\prime \\tilde {k}} \\right\\rangle\\right)}\\right), \\tag {5}", + "image_path": "922377003de3f677628ff8310ba56aa78489c935f13480a47687852a8222b50b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\lambda > 1" + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "text", + "content": " is the scaling factor to sharpen the softmax logits, " + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "inline_equation", + "content": "v'" + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "text", + "content": " normalized." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 492, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 554, + 540 + ], + "type": "text", + "content": "Multiple prototypes per concept. Due to the multi-modal nature of data where one class contains several clusters rather than just a single one, we initialize " + }, + { + "bbox": [ + 313, + 492, + 554, + 540 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 492, + 554, + 540 + ], + "type": "text", + "content": " prototypes per concept " + }, + { + "bbox": [ + 313, + 492, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\{p^{k_m}\\}_{m=1}^M" + }, + { + "bbox": [ + 313, + 492, + 554, + 540 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 492, + 554, + 540 + ], + "type": "inline_equation", + "content": "M \\ll N" + }, + { + "bbox": [ + 313, + 492, + 554, + 540 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "text", + "content": "The similarity of " + }, + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime k}" + }, + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "text", + "content": " with its concept " + }, + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "inline_equation", + "content": "c^{\\tilde{k}}" + }, + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "text", + "content": " is defined as the similarity to the nearest " + }, + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "inline_equation", + "content": "p^{\\tilde{k}_m}" + }, + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "text", + "content": ", and can be approximated by summing the compatibility with the " + }, + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "text", + "content": " clusters prototypes " + }, + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "inline_equation", + "content": "p^{\\tilde{k}_m}" + }, + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "text", + "content": ". Specifically, we first compute the probabilistic assignment " + }, + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_m^{\\tilde{k}}(v_i^{\\prime k})" + }, + { + "bbox": [ + 313, + 541, + 554, + 608 + ], + "type": "text", + "content": " to the prototypes:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 613, + 553, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 613, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 365, + 613, + 553, + 634 + ], + "type": "interline_equation", + "content": "\\mathrm {q} _ {m} \\left(v ^ {\\prime} _ {i} ^ {\\tilde {k}}\\right) = \\operatorname {s o f t m a x} _ {m} \\left(\\gamma \\left\\langle p ^ {\\tilde {k} _ {m}}, v ^ {\\prime} _ {i} ^ {\\tilde {k}} \\right\\rangle\\right) \\tag {6}", + "image_path": "8462c4feac57954426322e8c25d640cf5fdca30bcb9268efe809c49f45c0694a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 639, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 554, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 639, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_m" + }, + { + "bbox": [ + 313, + 639, + 554, + 689 + ], + "type": "text", + "content": " is the assignment distribution of the " + }, + { + "bbox": [ + 313, + 639, + 554, + 689 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 639, + 554, + 689 + ], + "type": "text", + "content": " prototypes in " + }, + { + "bbox": [ + 313, + 639, + 554, + 689 + ], + "type": "inline_equation", + "content": "c^{\\tilde{k}}" + }, + { + "bbox": [ + 313, + 639, + 554, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 639, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\gamma > 1" + }, + { + "bbox": [ + 313, + 639, + 554, + 689 + ], + "type": "text", + "content": " is the scaling factor. The similarity of the concept vector with each prototype is obtained by weighing their cosine similarity with the assignment:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 693, + 553, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 693, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 359, + 693, + 553, + 711 + ], + "type": "interline_equation", + "content": "\\sin^ {\\tilde {k} _ {m}} \\left(v ^ {\\prime} _ {i} ^ {\\tilde {k}}\\right) = q _ {m} \\left(v ^ {\\prime} _ {i} ^ {\\tilde {k}}\\right) \\cdot \\langle p ^ {\\tilde {k} _ {m}}, v ^ {\\prime} _ {i} ^ {\\tilde {k}} \\rangle . \\tag {7}", + "image_path": "ecfe7275b78781e759a0d560288a5f232b02ce901f2fd4c5f9c77b7c371bd6dd.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30800" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 71, + 276, + 198 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 276, + 198 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 276, + 198 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 276, + 198 + ], + "type": "image", + "image_path": "d7c77481e4c51259d549d7bbc9ebb56ba373de972fe2cbbd39ca51df624ebce4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 77, + 262, + 167, + 402 + ], + "blocks": [ + { + "bbox": [ + 74, + 212, + 276, + 260 + ], + "lines": [ + { + "bbox": [ + 74, + 212, + 276, + 260 + ], + "spans": [ + { + "bbox": [ + 74, + 212, + 276, + 260 + ], + "type": "text", + "content": "(a) Similarity maps generated by the local concept vector (first row) and concept prototype (second row) on the same image (a) and new images (b), (c). The local concept vector fails to highlight the concept pacemaker on new images while the concept prototype generalizes well with better localization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 262, + 167, + 402 + ], + "lines": [ + { + "bbox": [ + 77, + 262, + 167, + 402 + ], + "spans": [ + { + "bbox": [ + 77, + 262, + 167, + 402 + ], + "type": "image", + "image_path": "2591fad17d9119258290398644c4790e45d2af2858d85b853a4982e1412096f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 467, + 294, + 489 + ], + "lines": [ + { + "bbox": [ + 55, + 467, + 294, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 294, + 489 + ], + "type": "text", + "content": "Figure 4. Contrastive learning improves the concept feature space. (a) Qualitative and (b) quantitative comparison." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 262, + 272, + 402 + ], + "blocks": [ + { + "bbox": [ + 179, + 262, + 272, + 402 + ], + "lines": [ + { + "bbox": [ + 179, + 262, + 272, + 402 + ], + "spans": [ + { + "bbox": [ + 179, + 262, + 272, + 402 + ], + "type": "image", + "image_path": "70db739c08f16eec8b83a0697520ae4fde48eff76476b472dd84a8b6f57e64c6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "lines": [ + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "spans": [ + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "type": "text", + "content": "(b) Comparison between the original " + }, + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "type": "text", + "content": " and projected feature space " + }, + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "type": "inline_equation", + "content": "v^{\\prime}" + }, + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "type": "text", + "content": ". The box plots of the similarity score distribution between " + }, + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "type": "inline_equation", + "content": "v_{i}^{k}" + }, + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "type": "inline_equation", + "content": "v_{j\\neq i}^{k^{\\prime}}" + }, + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "type": "text", + "content": ". Intra-Concept: " + }, + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "type": "inline_equation", + "content": "k = k^{\\prime}" + }, + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "type": "text", + "content": ", Inter-Concept: " + }, + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "type": "inline_equation", + "content": "k\\neq k^{\\prime}" + }, + { + "bbox": [ + 74, + 406, + 276, + 456 + ], + "type": "text", + "content": ". The projected feature space is more compact and has better inter-concept separation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 508, + 295, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 535 + ], + "type": "text", + "content": "After that, we can obtain the similarity with concept " + }, + { + "bbox": [ + 55, + 508, + 295, + 535 + ], + "type": "inline_equation", + "content": "c^{\\tilde{k}}" + }, + { + "bbox": [ + 55, + 508, + 295, + 535 + ], + "type": "text", + "content": " by summing the compatibility with each cluster prototype:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 543, + 295, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 543, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 114, + 543, + 295, + 574 + ], + "type": "interline_equation", + "content": "\\sin^ {\\tilde {k}} \\left(v _ {i} ^ {\\prime} \\tilde {k}\\right) = \\sum_ {m} ^ {M} \\sin^ {\\tilde {k} _ {m}} \\left(v _ {i} ^ {\\prime} \\tilde {k}\\right). \\tag {8}", + "image_path": "e11dcee1fbbf5dba22107bed45e1d4fb49da176eb8a5b07d86ce13c08bb54d9e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 583, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 583, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 583, + 296, + 619 + ], + "type": "text", + "content": "Finally, we compute the contrastive loss by maximizing the similarity of a concept vector with the cluster prototypes of the positive concept, similar to Eqn. 5:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 628, + 295, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 628, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 66, + 628, + 295, + 661 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {c o n - m}} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right) = - \\log \\left(\\frac {\\exp \\left(\\lambda \\left(\\sin^ {\\tilde {k}} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right) + \\delta\\right)\\right)}{\\sum_ {k \\in K} \\exp \\left(\\lambda \\sin^ {k} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right)\\right)}\\right), \\tag {9}", + "image_path": "02474cc2db809293f6da7e5cf7833308a1ee51a74acef7739fba1be468075835.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 88, + 663, + 257, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 663, + 257, + 681 + ], + "spans": [ + { + "bbox": [ + 88, + 663, + 257, + 681 + ], + "type": "interline_equation", + "content": "\\text {i f} \\quad k = \\tilde {k}, \\quad \\operatorname {s i m} ^ {k} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right) = \\operatorname {s i m} ^ {k} \\left(v _ {i} ^ {\\prime \\tilde {k}}\\right) + \\delta .", + "image_path": "0fa6bcb929c532181c47cc029ba20fbd2216abbd8be9fd9091403ed29915ab4b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "A small margin " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": " between concept prototypes is added to further expand the decision boundaries. As a result, we" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "learn a set of concept prototypes " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\{p^{k_m}\\}^{M\\cdot K}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " that is significantly smaller in numbers compared to " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\{v_i^k\\}^{N\\cdot K}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", and the projector " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " that enhances the inter-concept separation and intra-concept compactness. We illustrate the learning process of " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\{p\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " in Fig. 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "text", + "content": "From Eqn. 2, " + }, + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "inline_equation", + "content": "v_{i}^{k}" + }, + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "text", + "content": " is a soft-selected feature patch of concept " + }, + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "text", + "content": " from the original feature " + }, + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "text", + "content": ", we can directly use " + }, + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "text", + "content": " to map feature patch " + }, + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i(h,w)" + }, + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "text", + "content": " to the compact and discriminative prototypical feature space. We then obtain the similarity of the concept prototypes with " + }, + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 133, + 553, + 192 + ], + "type": "text", + "content": " from Eqn. 3:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 199, + 553, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 199, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 323, + 199, + 553, + 217 + ], + "type": "interline_equation", + "content": "s ^ {k _ {m}} = \\max _ {h, w} S ^ {k _ {m}} (h, w) = \\max _ {h, w} \\langle p ^ {k _ {m}}, P (\\mathbf {f} (h, w)) \\rangle , \\tag {10}", + "image_path": "6de63068ec9ef3d8a5685a7b508499de36f1ff4b83b24cdb1803457daf7d95ab.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 227, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 553, + 263 + ], + "type": "text", + "content": "which are aggregated to create the similarity vector " + }, + { + "bbox": [ + 313, + 227, + 553, + 263 + ], + "type": "inline_equation", + "content": "[s^{k_m}]" + }, + { + "bbox": [ + 313, + 227, + 553, + 263 + ], + "type": "text", + "content": " in Eqn. 1. We quantitatively and qualitatively present the improved representation of " + }, + { + "bbox": [ + 313, + 227, + 553, + 263 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 227, + 553, + 263 + ], + "type": "text", + "content": " in Fig. 4." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 264, + 554, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 554, + 360 + ], + "type": "text", + "content": "Training details. Initially, we train the Concept model for multi-label classification using binary cross-entropy, predicting one-hot encoded vector of ground truth concepts. Then, we obtain the local concept vectors and learn the concept prototypes and the projection head with a contrastive loss " + }, + { + "bbox": [ + 313, + 264, + 554, + 360 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{con - m}}" + }, + { + "bbox": [ + 313, + 264, + 554, + 360 + ], + "type": "text", + "content": " to enhance concept generalization. Following, the task head is trained to predict the target from similarity scores with the concept prototypes, using cross-entropy." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 369, + 489, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 489, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 489, + 384 + ], + "type": "text", + "content": "3. Doctor-in-the-loop interactivity" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 389, + 429, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 429, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 429, + 403 + ], + "type": "text", + "content": "3.1. Model explanations" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 407, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 553, + 430 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 313, + 407, + 553, + 430 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 407, + 553, + 430 + ], + "type": "text", + "content": ", the model returns output " + }, + { + "bbox": [ + 313, + 407, + 553, + 430 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 313, + 407, + 553, + 430 + ], + "type": "text", + "content": " and the explanations for its prediction that include:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 430, + 553, + 514 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 313, + 430, + 553, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 553, + 454 + ], + "type": "text", + "content": "- The similarity maps " + }, + { + "bbox": [ + 313, + 430, + 553, + 454 + ], + "type": "inline_equation", + "content": "[S^{k_m}]" + }, + { + "bbox": [ + 313, + 430, + 553, + 454 + ], + "type": "text", + "content": " to show the activated region of the concept prototype " + }, + { + "bbox": [ + 313, + 430, + 553, + 454 + ], + "type": "inline_equation", + "content": "p^{k_m}" + }, + { + "bbox": [ + 313, + 430, + 553, + 454 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 430, + 553, + 454 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 430, + 553, + 454 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 454, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 553, + 479 + ], + "type": "text", + "content": "- The corresponding prototype image " + }, + { + "bbox": [ + 313, + 454, + 553, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(p^{k_m})" + }, + { + "bbox": [ + 313, + 454, + 553, + 479 + ], + "type": "text", + "content": " as a reference, also highlighting the activated regions of " + }, + { + "bbox": [ + 313, + 454, + 553, + 479 + ], + "type": "inline_equation", + "content": "p^{k_m}" + }, + { + "bbox": [ + 313, + 454, + 553, + 479 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 479, + 553, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 553, + 514 + ], + "type": "text", + "content": "- The concept similarity scores " + }, + { + "bbox": [ + 313, + 479, + 553, + 514 + ], + "type": "inline_equation", + "content": "[s^{k_m}]" + }, + { + "bbox": [ + 313, + 479, + 553, + 514 + ], + "type": "text", + "content": ", indicating the similarity of the two activated regions of " + }, + { + "bbox": [ + 313, + 479, + 553, + 514 + ], + "type": "inline_equation", + "content": "p^{k_m}" + }, + { + "bbox": [ + 313, + 479, + 553, + 514 + ], + "type": "text", + "content": " on the input image " + }, + { + "bbox": [ + 313, + 479, + 553, + 514 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 479, + 553, + 514 + ], + "type": "text", + "content": " and on the reference image " + }, + { + "bbox": [ + 313, + 479, + 553, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(p^{k_m})" + }, + { + "bbox": [ + 313, + 479, + 553, + 514 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 514, + 554, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 554, + 575 + ], + "type": "text", + "content": "The model explanations correspond to the similarity-based inference logic outlined in Sec. 2.1. The operator " + }, + { + "bbox": [ + 313, + 514, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 514, + 554, + 575 + ], + "type": "text", + "content": " retrieves the image associated with the prototype " + }, + { + "bbox": [ + 313, + 514, + 554, + 575 + ], + "type": "inline_equation", + "content": "p^{k_m}" + }, + { + "bbox": [ + 313, + 514, + 554, + 575 + ], + "type": "text", + "content": ". We replace each " + }, + { + "bbox": [ + 313, + 514, + 554, + 575 + ], + "type": "inline_equation", + "content": "p^{k_m}" + }, + { + "bbox": [ + 313, + 514, + 554, + 575 + ], + "type": "text", + "content": " with the nearest projected local concept vector " + }, + { + "bbox": [ + 313, + 514, + 554, + 575 + ], + "type": "inline_equation", + "content": "v_{i}^{k}" + }, + { + "bbox": [ + 313, + 514, + 554, + 575 + ], + "type": "text", + "content": " and link it to its source image " + }, + { + "bbox": [ + 313, + 514, + 554, + 575 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 514, + 554, + 575 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 362, + 583, + 553, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 583, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 362, + 583, + 553, + 605 + ], + "type": "interline_equation", + "content": "\\mathcal {I} \\left(p ^ {k _ {m}}\\right) = x \\leftarrow \\underset {i} {\\operatorname {a r g m i n}} \\left\\langle v _ {i} ^ {\\prime k}, p ^ {k _ {m}} \\right\\rangle . \\tag {11}", + "image_path": "64282f7e40e983fc1886d3f1f1b03c07e5d091f7b55511098bae274d8bb41b5a.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 613, + 441, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 441, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 441, + 624 + ], + "type": "text", + "content": "3.2. Train-time interaction" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "After we learn the concept prototypes " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\{p^{k_m}\\}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " and its associated human-interpretable version " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}(p^{k_m})\\}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": ", or atlas, doctors can inspect each concept prototype and discard any unqualified examples. This process can reveal and mitigate the Clever-Hans effect in the Concept model by examining the clues it is taking to predict the concepts. If a given prototype is shown to be a shortcut rather than a meaningful signal, it" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30801" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 72, + 282, + 219 + ], + "blocks": [ + { + "bbox": [ + 69, + 72, + 282, + 219 + ], + "lines": [ + { + "bbox": [ + 69, + 72, + 282, + 219 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 282, + 219 + ], + "type": "image", + "image_path": "743ff4dafff8b227b2e4359779ac766a71eca65afb134da1de3f56c6289d2b80.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "type": "text", + "content": "Figure 5. The irrelevant concept prototypes to be discarded in train-time interaction, highlighted at the activated regions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 273, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 296, + 346 + ], + "type": "text", + "content": "is removed from " + }, + { + "bbox": [ + 55, + 273, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\{p^{k_m}\\}" + }, + { + "bbox": [ + 55, + 273, + 296, + 346 + ], + "type": "text", + "content": " and is not used to calculate the similarity scores " + }, + { + "bbox": [ + 55, + 273, + 296, + 346 + ], + "type": "inline_equation", + "content": "[s^{k_m}]" + }, + { + "bbox": [ + 55, + 273, + 296, + 346 + ], + "type": "text", + "content": " for the prediction of " + }, + { + "bbox": [ + 55, + 273, + 296, + 346 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 273, + 296, + 346 + ], + "type": "text", + "content": ", effectively nullify the Clever-Hans effect. We show the irrelevant examples of different concepts in Fig. 5. The similarity-based inference logic allows this interaction as we only need as much as the remaining relevant examples to predict the target " + }, + { + "bbox": [ + 55, + 273, + 296, + 346 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 273, + 296, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 355, + 176, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 176, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 176, + 366 + ], + "type": "text", + "content": "3.3. Test-time interaction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 409 + ], + "type": "text", + "content": "At test time, we introduce two types of interaction to improve the prediction quality: Spatial-level interaction and Concept-level interaction. We illustrate this in Fig. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 410, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 517 + ], + "type": "text", + "content": "Concept-level interaction. Consider this scenario: While reviewing the explanations from CSR for a chest X-ray " + }, + { + "bbox": [ + 55, + 410, + 295, + 517 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 410, + 295, + 517 + ], + "type": "text", + "content": ", the doctor notices that the model is considering a nodule on " + }, + { + "bbox": [ + 55, + 410, + 295, + 517 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 410, + 295, + 517 + ], + "type": "text", + "content": " for a mass in the reference image with a high similarity score. Recognizing the incorrect comparison, the doctor informs the model that concept mass is absent in " + }, + { + "bbox": [ + 55, + 410, + 295, + 517 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 410, + 295, + 517 + ], + "type": "text", + "content": ". The model suppresses its confidence in concept mass to calibrate its prediction. This concept-level interaction allows doctors to directly eliminate irrelevant concepts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "text", + "content": "Interactivity mechanism. When the doctor goes through " + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "inline_equation", + "content": "[s^{k_m}]" + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "text", + "content": ", she can discard any invalid concept if presented. The model then sets " + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "inline_equation", + "content": "s^{k_m} = 0" + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "text", + "content": " for those rejected concepts, effectively discounting them on the prediction " + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "text", + "content": ". Concept-level interaction only considers the negative type of input from the doctor: concept rejection. Setting a suitable value for every " + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "inline_equation", + "content": "s_m^k" + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "text", + "content": " would be a overburden. Hence, for each concept " + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "text", + "content": ", we only present the highest similarity score " + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\max_{m} \\{s^{k_m}\\}" + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "text", + "content": " and have the doctor recognize its presence. We zero out all of the similarity scores of the rejected concept." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Spatial-level interaction. As the doctor examines the image " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": ", upon spotting an important area, she can guide the model's attention with bounding boxes. The model then recalibrates its prediction " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " by prioritizing the indicated region based on this feedback. Conversely, reviewing the activation maps " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "[S^{k_m}]" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": ", the doctor also notices the model con" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "centrating on an irrelevant region for prediction. To counteract the Clever-Hans effect, she marks the area as spurious, prompting the model to disregard it in its updated prediction. Hence, the doctor only needs to direct the model where to focus, without specifying what is in that region. This interaction scheme is practical and avoids the doctor from inadvertently becoming a teacher rather than taking advantage of the model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "text", + "content": "Interactivity mechanism. As we described in the user journey above, the spatial-level interaction allows two types of input from the doctor: (i) the positive boxes " + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\{\\mathsf{bb}^{+} = (\\mathsf{x}_1,\\mathsf{y}_1,\\mathsf{x}_2,\\mathsf{y}_2)\\}" + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "text", + "content": " for the focus areas, (ii) the negative boxes " + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\{\\mathsf{bb}^{-}\\}" + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "text", + "content": " for the regions to ignore. To emphasize the area specified by " + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\{\\mathsf{bb}^{+}\\}" + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "text", + "content": " and suppress the area specified by " + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\{\\mathsf{bb}^{-}\\}" + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "text", + "content": ", the model reweights each element " + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "inline_equation", + "content": "S(h,w)" + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "text", + "content": " with an importance map " + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "inline_equation", + "content": "A\\in \\mathbb{R}^{\\mathcal{H}\\times \\mathcal{W}}" + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 356, + 282, + 553, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 282, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 356, + 282, + 553, + 326 + ], + "type": "interline_equation", + "content": "A (h, w) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} (h, w) \\in \\left\\{\\mathrm {b b} ^ {+} \\right\\} \\\\ 0 & \\text {i f} (h, w) \\in \\left\\{\\mathrm {b b} ^ {-} \\right\\}, \\\\ \\alpha & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {12}", + "image_path": "01dce74535720d835b8dea4c3dcf2c64c36985b9e6cb5044b0e4f67c7ac480db.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1)" + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "text", + "content": " is the neutral weight for the other regions. Setting " + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "text", + "content": " indicates that the doctor only considers the regions " + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{bb}^{+}\\}" + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "text", + "content": ", while setting " + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "inline_equation", + "content": "\\alpha \\rightarrow 1" + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "text", + "content": " diminishes the importance of " + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{bb}^{+}\\}" + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "text", + "content": ". The new set of similarity maps " + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "inline_equation", + "content": "[\\hat{S}^{k_m}]" + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "text", + "content": " is obtained by multiplying every similarity map of " + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "inline_equation", + "content": "[S^{k_m}]" + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "text", + "content": " with the importance map " + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 342, + 553, + 415 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 429, + 553, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 429, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 335, + 429, + 553, + 444 + ], + "type": "interline_equation", + "content": "[ \\hat {S} ] = A \\odot [ S ] = [ A \\odot S ^ {1 _ {1}}, \\dots , A \\odot S ^ {K _ {M}} ], \\tag {13}", + "image_path": "458211890e008a5d18db3d947fa0aee09e8a2eef3d070d4a790c2796b46c53b1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "content": " denotes the element-wise multiplication on the spatial dimension. As CSR is designed to select the maximum element on " + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "inline_equation", + "content": "S^{k_m}" + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "inline_equation", + "content": "s^{k_m}" + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "content": ", giving more weight to the important region " + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "inline_equation", + "content": "S(h,w)" + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "content": " increases its likelihood to be selected. On the other hand, zeroing out the spurious region nullifies its influence on the similarity vector. Because " + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "content": " is bounded within " + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "content": " is bounded within " + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "content": ", we clip negative values of " + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "content": " to condition the model on only non-negative values. This ensures that the " + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{bb}^+\\}" + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{bb}^-\\}" + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "content": " regions are monotonically adjusted. The new prediction is made with " + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 313, + 459, + 554, + 601 + ], + "type": "text", + "content": " by the same inference logic as described in Sec. 2.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Discussion. Combining both spatial and concept interaction mechanisms enables a flexible interaction approach. For example, if the doctor identifies an important region, she can draw a positive box " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{bb}^{+}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " over it. However, if the model associates the feature within " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{bb}^{+}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " with a wrong concept " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": ", placing " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{bb}^{+}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " over it would amplify the error. Because " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " is absent in " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": ", rejecting it while specifying " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{bb}^{+}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " will instruct the model to focus on other concepts " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\{c^{j\\neq k}\\}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{bb}^{+}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " and ignore " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "c^k" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30802" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 145, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 145, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 145, + 83 + ], + "type": "text", + "content": "4. Related works" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 258 + ], + "type": "text", + "content": "Concept-based. Concept Bottleneck Model (CBM) [11] predicts the concepts before using them as support for the target prediction. Post-hoc Concept Bottleneck Models (PHCBM) [16] transfer concepts from other datasets or from natural language descriptions of concepts via multimodal models like CLIP [25]. Yang et al. leverages Large Language Models to generate an exhaustive list of concepts for each class [36]. These multimodal concept-based approaches [15, 22, 36] utilize foundation models to enhance their capabilities. However, foundation models are generally trained on broad domains, which presents limitations when applying them directly to medical imaging. In this work, we focus on the research direction of vision features and explore multi-modality learning in future studies." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 260, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 295, + 403 + ], + "type": "text", + "content": "Part-prototypes. ProtoPNet [2] used prototypical parts to make a final classification based on positive attributions. ProtoTree[17] is a globally interpretable model by design that uses prototypes with decision trees. ProtoPool [28] is similar to ProtoPNet with a tweak of the shared prototypes pool and the Gumbel-softmax trick for the differentiable property. PIP-Net[18] is a patch-based approach with a novel regularization method for learning prototype similarity that correlates with human visual perception. Unlike previous methods that correlate prototypes with target classes, CSR explicitly grounds prototypes on interpretable concepts, making them semantically relevant." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 403, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 513 + ], + "type": "text", + "content": "Interactivity. CBM [11] enables local interaction as human adjusts the concept logits for individual samples, while [16] allows global interaction by modifying concept weights of the model. ProtoPDebug [1] and [6] focus on part-prototype interaction, refining prototypes for enhanced performance. Besides logits-level interaction like those methods, CSR introduces spatial-level interaction, which is particularly intuitive and beneficial in the medical imaging domain, where detailed image analysis is of significant interest to doctors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 525, + 200, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 200, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 200, + 538 + ], + "type": "text", + "content": "5. Experiments and Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": "In this paper, we experiment on two chest X-ray datasets: TBX11K [13], VinDr-CXR[20] and a skin imaging dataset ISIC[4]. From those datasets, we define the findings annotations as the concepts and the diseases annotations as the target classes. Details on datasets and the training hyperparameters are in the supplementary. We qualitatively and quantitatively demonstrate the effectiveness of the concept prototypes learning framework in Fig. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": "Diagnostic performance. Due to the heavy class imbalance, we use macro F1-score as the metric for our experiments. We report the results in Table 1. It can be seen that CSR demonstrates a strong diagnostic performance compared to other baselines across all dataset. We also benchmark the performance of CSR when we ask a doctor to re" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 345, + 71, + 523, + 376 + ], + "blocks": [ + { + "bbox": [ + 345, + 71, + 523, + 376 + ], + "lines": [ + { + "bbox": [ + 345, + 71, + 523, + 376 + ], + "spans": [ + { + "bbox": [ + 345, + 71, + 523, + 376 + ], + "type": "table", + "html": "
MethodF1↑#pro.↓#exp.↓
TBX11KCBM (ind.)24.514
CBM (seq.)60.514
CBM (joint.)88.614
ProtoPNet89.3600600
ProtoPNet89.315001500
ProtoPNet94.130003000
ProtoPool90.7600600
ProtoTree(H=6)94.01276
ProtoTree(H=9)93.75119
PIP-Net94.0768158
CSR94.4140014
CSR (refined)94.0138614
VinDr-CXRCBM (joint.)50.114
ProtoPNet41.415001500
PIP-Net45.17689
CSR54.6140014
ISICCBM (joint.)45.54
ProtoPNet38.5600600
ProtoPool31.0200200
ProtoTree(H=6)66.11276
PIP-Net69.976890
CSR71.54004
", + "image_path": "64bc75d6bb3c1b176dfe4c817fc34cb20c45c8dad58cd6162be0cec92f876ca2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 442, + 553, + 483 + ], + "blocks": [ + { + "bbox": [ + 313, + 385, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 430 + ], + "type": "text", + "content": "Table 1. Diagnostic performance on the medical datasets, measured by Macro F1-score. CSR (refined) has the doctors review the concept atlas. #pro. denotes the number of prototypes of the model. #exp. indicates the explanation size for each prediction." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 442, + 553, + 483 + ], + "lines": [ + { + "bbox": [ + 315, + 442, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 315, + 442, + 553, + 483 + ], + "type": "table", + "html": "
MethodProtoP-NetProto TreePIP-NetCBMCSRCSR (refined)
% PG hit8.87.819.555.160.979.5
", + "image_path": "3afb5271008e0af672f29776104c4e304e1ce9f241ee94739d598d95d1f6ca37.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 491, + 555, + 556 + ], + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 556 + ], + "type": "text", + "content": "Table 2. Pointing Game hit rate on TBX11K, Tuberculosis class. CBM uses the class activation map of the highest contributing product to the class Tuberculosis. ProtoTree uses a similarity map from the top prototype at level-1 of the tree. ProtoPNet, PIP-Net, and CSR use the similarity map from the prototype with the highest contribution to class Tuberculosis." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 580, + 531, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 531, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 531, + 592 + ], + "type": "text", + "content": "fine the learned concept atlas as described in Sec. 3.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "Explanation size. Following [18], we quantify the explanation size as the number of explanations per prediction for each method. For part-prototype methods, explanations refer to the similar parts of the image from the training set. ProtoTree has the explanation size equal to its depth. PIP-Net reduces the explanation size by regularizing the sparsity during training and only considers non-zero prototypes for each prediction. Other part-prototype approaches have the explanation size matching the large number of prototypes. This compromises interpretability since users must" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30803" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 293, + 188 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 188 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 188 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 188 + ], + "type": "image", + "image_path": "300a2b228ea54334149c3dfca4932232d3bbcf631d2b48f046e886f4132ba99c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 196, + 298, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 196, + 298, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 298, + 242 + ], + "type": "text", + "content": "Figure 6. Qualitative result of the Pointing Game. We visualize the groundtruth bounding box and the maximum activation point of each method as the colored square. :Proposed CSR, :CBM, :PIP-Net, :ProtoPNet, :ProtoTree" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 266, + 296, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 296, + 398 + ], + "type": "text", + "content": "navigate through a multitude of explanations within the set for each prediction. Concept-based models use the concepts as their explanations, having an explanation size equal to the number of concepts, which is typically smaller. This choice involves an accuracy trade-off due to the reduced size of the bottleneck. In contrast, the proposed CSR benefits from both, featuring a small explanation size equivalent to CBM while achieving competitive accuracy akin to part-prototype methods. To achieve this, our CSR only considers concept grounded explanations and exclusively presents the highest similar prototype of each active concept." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 400, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 508 + ], + "type": "text", + "content": "Trustworthiness. We inspect the explanations of each benchmark and demonstrate that interpretability does not always come with trustworthiness. A trustworthy model must deliver relevant explanations that correspond to human interpretable concepts. We leverage the pointing game (PG) [37] that considers a hit if the point of maximum value in the explanation lies within a segmentation ground truth. We use the bounding box annotation from the TBX11K and calculate the PG hit rate of the explanations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "The quantitative results are reported in Table. 2. Partprototype methods including ProtoPNet and ProtoTree are relying on clues that considerably mismatch with the semantic ground truth. This observation aligns with [18], where Purity of the prototypes is evaluated as an alignment with the part annotations in the CUB dataset [32], demonstrating a large semantic gap between the prototypes and human-relevant concepts. Concept-based methods attain a more reasonable PG hit rate as the concepts are human-grounded and thus align well with the semantic ground truth. The proposed CSR faithfully aligns with human reasoning and achieves a remarkable PG hit rate of " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "60.9\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": ". Refining the concept atlas further improves the PG hit rate to " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "79.5\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": ", which hints that the better classification result of the original CSR in Table 1 could stem from shortcuts. The qualitative results in Fig. 6 shows that our CSR:□ more frequently identifies the maximum activation within" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 345, + 70, + 525, + 163 + ], + "blocks": [ + { + "bbox": [ + 345, + 70, + 525, + 163 + ], + "lines": [ + { + "bbox": [ + 345, + 70, + 525, + 163 + ], + "spans": [ + { + "bbox": [ + 345, + 70, + 525, + 163 + ], + "type": "table", + "html": "
spatialconceptMac.F1+Δ%
CSR94.40.0
Non-expert✓94.50.1
Oracle✓94.90.5
M.D.✓94.70.3
M.D.✓95.10.7
M.D.✓✓94.90.5
", + "image_path": "fb050c105a07bce560116e94fa1559b93fed4ba6809d461bc23bbaa4bac7883f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 171, + 555, + 216 + ], + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 216 + ], + "type": "text", + "content": "Table 3. Interactions performance gain. spatial: spatial-level interaction, concept: concept-level interaction. We limit the Non-export to spatial-interaction only. Oracle: using the bounding box ground truth of TB as spatial-interaction. M.D: medical student." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 249, + 512, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 512, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 512, + 261 + ], + "type": "text", + "content": "the ground truth box compared to other methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 266, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 483 + ], + "type": "text", + "content": "Interactivity. We demonstrate the benefits of interactivity by measuring the model when obtaining interaction at test time on the TBX11K dataset and report the result in Table. 3. Our experiment involves a Medical degree (M.D.) student and a non-expert in evaluating the framework. In the oracle setting, we use the ground truth bounding box of class TB from the dataset as a positive spatial-level interaction. Note that we only require human interaction when the output probability of the model is indecisive: when the difference of top-1 and top-2 class probability is less than 0.3. The weight " + }, + { + "bbox": [ + 313, + 266, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 266, + 555, + 483 + ], + "type": "text", + "content": " is set to 0.2. It shows that human interactions has a positive impact on the model outcome. The interaction from the M.D. student is more beneficial to the model than the non-expert. We consider that the non-experts can still eliminate irrelevant CXR regions without formal radiology expertise, and limit the non-expert to spatial-level interactions only. The concept-level interaction is more influential than the spatial-level interaction with better gain." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 511, + 388, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 511, + 388, + 525 + ], + "spans": [ + { + "bbox": [ + 314, + 511, + 388, + 525 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 536, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 715 + ], + "type": "text", + "content": "This paper introduces a novel Concept Similarity Reasoning (CSR), an interpretable and interactive model for medical image analysis with similarity reasoning. CSR provides local explanations by presenting similar concept examples with the input images. In comparison to part-prototype approaches, our CSR presents more concise and meaningful prototypes that align better with human-relevant concepts. This observation implies the differentiation between interpretability and trustworthiness. The proposed interaction mechanism employed during both training and testing effectively enhances model reliability by mitigating shortcuts and improves model accuracy through domain-specific inputs. Our method is an attempt to encourage more trust from doctors through a natural and effective human-AI interaction, improving both accuracy and transparency." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30804" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "text", + "content": "[1] Andrea Bontempelli, Stefano Teso, Fausto Giunchiglia, and Andrea Passerini. Concept-level debugging of part-prototype networks. ArXiv, abs/2205.15769, 2022. 1, 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "type": "text", + "content": "[2] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: Deep learning for interpretable image recognition. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 1, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 296, + 258 + ], + "type": "text", + "content": "[3] Townim Faisal Chowdhury, Kewen Liao, Vu Minh Hieu Phan, Minh-Son To, Yutong Xie, Kevin Hung, David Ross, Anton van den Hengel, Johan W Verjans, and Zhibin Liao. Cape: Cam as a probabilistic ensemble for enhanced dnn interpretation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11072-11081, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 360 + ], + "type": "text", + "content": "[4] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), pages 168-172. IEEE, 2018. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 362, + 296, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 296, + 405 + ], + "type": "text", + "content": "[5] Xingping Dong and Jianbing Shen. Triplet loss in siamese network for object tracking. In Proceedings of the European conference on computer vision (ECCV), pages 459-474, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 407, + 296, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 407, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 62, + 407, + 296, + 461 + ], + "type": "text", + "content": "[6] Michael Gerstenberger, Thomas Wiegand, Peter Eisert, and Sebastian Bosse. But that's not why: Inference adjustment by interactive prototype revision. In Iberoamerican Congress on Pattern Recognition, pages 123–132. Springer, 2023. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 464, + 296, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 296, + 497 + ], + "type": "text", + "content": "[7] Alec Holt, Isabelle Bichindaritz, Rainer Schmidt, and Petra Perner. Medical applications in case-based reasoning. The Knowledge Engineering Review, 20(3):289-292, 2005. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 499, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 296, + 554 + ], + "type": "text", + "content": "[8] Ta Duc Huy, Hoang Cao Huyen, Chanh DT Nguyen, Soan TM Duong, Trung Bui, and Steven QH Truong. Adversarial contrastive fourier domain adaptation for polyp segmentation. In 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 556, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 296, + 609 + ], + "type": "text", + "content": "[9] Sajid Javed, Arif Mahmood, Iyyakutti Iyappan Ganapathi, Fayaz Ali Dharejo, Naoufel Werghi, and Mohammed Bennamoun. Cclip: Zero-shot learning for histopathology with comprehensive vision-language alignment. In CVPR, pages 11450-11459, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "text", + "content": "[10] Injae Kim, Jongha Kim, Joonmyung Choi, and Hyunwoo J Kim. Concept bottleneck with visual concept filtering for explainable medical image classification. In The Medical Computing and Computer Assisted Intervention Society, pages 225–233. Springer, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[11] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning, pages 5338-5348. PMLR, 2020. 1, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[12] Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. Unmasking clever hans predictors and assessing what machines really learn. Nature communications, 10(1):1096, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 554, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 182 + ], + "type": "text", + "content": "[13] Yun Liu, Yu-Huan Wu, Yunfeng Ban, Huifang Wang, and Ming-Ming Cheng. Rethinking computer-aided tuberculosis diagnosis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2646–2655, 2020. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 184, + 554, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 554, + 238 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 554, + 238 + ], + "type": "text", + "content": "[14] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology. Nature Medicine, 30(3):863-874, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 239, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 260 + ], + "type": "text", + "content": "[15] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 554, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 554, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 554, + 294 + ], + "type": "text", + "content": "[16] Yuksekgonul Mert, Wang Maggie, and Zou James. Post-hoc concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 295, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 554, + 348 + ], + "type": "text", + "content": "[17] Meike Nauta, Ron van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14933-14943, 2021. 1, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 350, + 555, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 555, + 404 + ], + "type": "text", + "content": "[18] Meike Nauta, Jörg Schlötterer, Maurice van Keulen, and Christin Seifert. Pip-net: Patch-based intuitive prototypes for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2744-2753, 2023. 1, 3, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 405, + 554, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 554, + 449 + ], + "type": "text", + "content": "[19] Giang Nguyen, Daeyoung Kim, and Anh Nguyen. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems, 34:26422-26436, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 449, + 554, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 554, + 503 + ], + "type": "text", + "content": "[20] Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, et al. Vindr-cxr: An open dataset of chest x-rays with radiologist's annotations. Scientific Data, 9(1):429, 2022. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 505, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 554, + 536 + ], + "type": "text", + "content": "[21] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 537, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 555, + 613 + ], + "type": "text", + "content": "[22] Vu Minh Hieu Phan, Yutong Xie, Yuankai Qi, Lingqiao Liu, Liyang Liu, Bowen Zhang, Zhibin Liao, Qi Wu, Minh-Son To, and Johan W Verjans. Decomposing disease descriptions for enhanced pathology detection: A multi-aspect vision-language pre-training framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11492-11501, 2024. 1, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 615, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 554, + 658 + ], + "type": "text", + "content": "[23] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6450-6458, 2019. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 659, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 554, + 713 + ], + "type": "text", + "content": "[24] Tran Minh Quan, Huynh Minh Thanh, Ta Duc Huy, Nguyen Do Trung Chanh, Nguyen Thi Phuong Anh, Phan Hoan Vu, Nguyen Hoang Nam, Tran Quy Tuong, Vu Minh Dien, Bui Van Giang, et al. Xpgan: X-ray projected generative adversarial network for improving Covid-19 image classification." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30805" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 1509-1513. IEEE, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 162 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "text", + "content": "[26] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence, 1(5):206-215, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "type": "text", + "content": "[27] Dawid Rymarczyk, Łukasz Struski, Jacek Tabor, and Bartosz Zielinski. Protopshare: Prototype sharing for interpretable image classification and similarity discovery. arXiv preprint arXiv:2011.14340, 2020. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "type": "text", + "content": "[28] Dawid Rymarczyk, Łukasz Struski, Michal Górszczak, Koryna Lewandowska, Jacek Tabor, and Bartosz Zielinski. Interpretable image classification with differentiable prototypes assignment. In European Conference on Computer Vision, pages 351-368. Springer, 2022. 1, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 296, + 365 + ], + "type": "text", + "content": "[29] Tran Bao Sam, Ta Duc Huy, Cong Tuyen Dao, Thanh Tin Lam, Van Ha Tang, and Steven Q.H. Truong. A multi-phase multi-graph approach for focal liver lesion classification on ct scans. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 3788-3802, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 431 + ], + "type": "text", + "content": "[30] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 296, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 296, + 466 + ], + "type": "text", + "content": "[31] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 296, + 499 + ], + "type": "text", + "content": "[32] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "text", + "content": "[33] Bor-Shiun Wang, Chien-Yi Wang, and Wei-Chen Chiu. Mcpnet: An interpretable classifier via multi-level concept prototypes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10885-10894, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 296, + 622 + ], + "type": "text", + "content": "[34] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 690 + ], + "type": "text", + "content": "[35] Siyuan Yan, Zhen Yu, Xuelin Zhang, Dwarikanath Mahapatra, Shekhar S Chandra, Monika Janda, Peter Soyer, and Zongyuan Ge. Towards trustable skin cancer diagnosis via rewriting model's decision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11568-11577, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[36] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 217 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 335, + 72, + 555, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 72, + 555, + 118 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 555, + 118 + ], + "type": "text", + "content": "in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 1, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 162 + ], + "type": "text", + "content": "[37] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. International Journal of Computer Vision, 126(10):1084-1102, 2018. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "text", + "content": "[38] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "30806" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/72ec2f30-d729-4ccc-bc2d-90a1d1bd3d41_content_list.json b/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/72ec2f30-d729-4ccc-bc2d-90a1d1bd3d41_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0eae1f715f8e11b1f5c047a824bf0ed5678b8a81 --- /dev/null +++ b/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/72ec2f30-d729-4ccc-bc2d-90a1d1bd3d41_content_list.json @@ -0,0 +1,1525 @@ +[ + { + "type": "text", + "text": "Interactive Medical Image Segmentation: A Benchmark Dataset and Baseline", + "text_level": 1, + "bbox": [ + 107, + 130, + 890, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junlong Cheng $^{1,2}$ Bin Fu $^{2,4}$ Jin Ye $^{2}$ Guoan Wang $^{2}$ Tianbin Li $^{2}$ Haoyu Wang $^{2}$ Ruoyu Li $^{1}$ He Yao $^{1}$ Junren Cheng $^{1}$ Jinwen Li $^{3}$ Yanzhou Su $^{2}$ Min Zhu $^{1*}$ Junjun He $^{2\\dagger}$", + "bbox": [ + 99, + 178, + 906, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1College of Computer Science, Sichuan University", + "bbox": [ + 303, + 219, + 705, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Shanghai AI Laboratory, China", + "bbox": [ + 377, + 238, + 633, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Computer Science and Technology, Xinjiang University", + "bbox": [ + 240, + 255, + 769, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences", + "bbox": [ + 197, + 273, + 812, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chengjunlong, fubin, yejin, wangguoan, litianbin, wanghaoyu} @pjlab.org.cn,", + "bbox": [ + 272, + 296, + 736, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ruoyuli, 2024323040035, chengjunren} $@$ stu.scu.edu.cn, 107556522204@stu.xju.edu.cn,", + "bbox": [ + 238, + 314, + 771, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yanzhou.su@outlook.com, hejunjun.su@pjlab.org.cn, zhumin@scu.edu.cn", + "bbox": [ + 281, + 332, + 723, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 359, + 326, + 376 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Interactive Medical Image Segmentation (IMIS) has long been constrained by the limited availability of large-scale, diverse, and densely annotated datasets, which hinders model generalization and consistent evaluation across different models. In this paper, we introduce the IMed-361M benchmark dataset, a significant advancement in general IMIS research. First, we collect and standardize over 6.4 million medical images and their corresponding ground truth masks from multiple data sources. Then, leveraging the strong object recognition capabilities of a vision foundational model, we automatically generated dense interactive masks for each image and ensured their quality through rigorous quality control and granularity management. Unlike previous datasets, which are limited by specific modalities or sparse annotations, IMed-361M spans 14 modalities and 204 segmentation targets, totaling 361 million masks—an average of 56 masks per image. Finally, we developed an IMIS baseline network on this dataset that supports high-quality mask generation through interactive inputs, including clicks, bounding boxes, text prompts, and their combinations. We evaluate its performance on medical image segmentation tasks from multiple perspectives, demonstrating superior accuracy and scalability compared to existing interactive segmentation models. To facilitate research on foundational models in medical computer vision, we release the IMed-361M and model at https://github.com.uni-medical/IMIS-Bench.", + "bbox": [ + 88, + 402, + 485, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 359, + 643, + 376 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Interactive Medical Image Segmentation (IMIS) enables clinicians or users to guide the model by marking points, lines, or regions on an image, resulting in segmentation outcomes that better meet clinical needs [1, 2]. This user-informed segmentation approach not only optimizes results throughout the segmentation process, aiding clinicians in precisely localizing targets during diagnosis and treatment [3-5], but also addresses the limitations of fully automated segmentation models in generalizing to unseen object classes (i.e., zero-shot learning) [6-11], providing clinicians with a flexible adjustment mechanism that significantly enhances segmentation reliability. However, the advancement of IMIS advancement faces a critical bottleneck: the lack of high-quality, diverse, and large-scale datasets that capture the complex clinical environments [12, 13]. Existing public datasets (such as Totalsegmentator [14], AutoPET [15], and CoNIC [16]) are often tailored to specific modalities or tasks, thus constraining the generalization capabilities of models across diverse medical scenarios.", + "bbox": [ + 511, + 386, + 906, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, as a foundational model for general visual segmentation, the Segment Anything Model (SAM) has garnered widespread attention [17, 18]. SAM integrates simple user interactions (such as points or bounding boxes) into the model's learning process and leverages pre-training on large-scale datasets to achieve cross-domain and multitask transferability [19-21]. Although SAM has been pretrained on over 1 billion natural image masks, its direct application in medical imaging remains limited due to the substantial domain differences between natural and medical images, as well as the lack of medical imaging-specific knowledge. To address these issues, current methods such as MedSAM [22] and SAM-Med2D [23] attempt to expand dataset size and fine-tune SAM to create IMIS models tailored for clinical scenarios. While these approaches have", + "bbox": [ + 511, + 674, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 107, + 875, + 235, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Project lead.", + "bbox": [ + 109, + 887, + 184, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20841", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shown some progress, their ability to 'segment anything' in medical imaging remains limited. This limitation is mainly due to the lack of densely masks in existing medical datasets compared to the SA-1B dataset [17] used to train SAM [22-24]. For instance, the COSMOS dataset [24] contains an average of only 5.7 masks per image, which restricting the model's capacity for dense segmentation and hindering comprehensive, fine-grained interaction. Moreover, many methods are evaluated only on specific modalities or with limited interaction strategies, further restricting the comprehensiveness and reliability of IMIS model evaluation [22, 24-26].", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Addressing these limitations requires the development of a high-quality IMIS benchmark dataset, which is essential for advancing foundational models in medical imaging [4, 27-30]. An ideal IMIS benchmark dataset should meet three core criteria: (1) Large-scale. The dataset should be large enough to fully support deep learning model training, enabling the model to effectively capture medical features; (2) Diversity. The dataset should encompass various medical imaging modalities and complex clinical scenarios to ensure the model's ability to generalize across modalities and tasks; (3) High-quality and densely masks. Accurate segmentation of complex medical images relies on high-quality annotations, and densely masks further enhance the model's capability to \"segment anything\" in medical imaging. Unfortunately, existing public medical segmentation datasets [22-24, 31] do not fully meet these standards, limiting their ability to comprehensively support and evaluate IMIS models.", + "bbox": [ + 91, + 276, + 483, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce IMed-361M, a benchmark dataset specifically designed for IMIS tasks, this dataset integrates public and private data sources and utilizes foundational models [17] for automated annotation to generating dense masks for each image. A standardized data processing workflow is ensure the high quality and consistency of all masks. IMed-361M achieves unprecedented scale, diversity, and mask quality, comprising 6.4 million images spanning 14 imaging modalities and 204 targets, with a total of 361 million masks, averaging 56 masks per image. IMed-361M effectively solves the problems of small data size and sparse annotations, providing data support for IMIS model training. Additionally, we develop an IMIS baseline model and conduct a comprehensive evaluation of its performance across various medical scenarios, including its effectiveness on different modalities, anatomical structures, and organs, as well as the impact of various interaction strategies on model outcomes. This analysis provided an in-depth understanding of the strengths and limitations of different interactive segmentation methods, establishing a fair and consistent framework for evaluating IMIS model performance. We anticipate that the IMed-361M dataset and baseline model will drive the widespread adoption of IMIS", + "bbox": [ + 89, + 553, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "technology in clinical practice, accelerating the healthcare industry's transition toward intelligence and automation.", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 135, + 653, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section provides an overview of datasets used for various medical image segmentation tasks and reviews the advancements in IMIS algorithms.", + "bbox": [ + 511, + 161, + 903, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Datasets. Due to variations in imaging protocols, medical images often have multidimensional characteristics (e.g., 2D, 3D), making pixel-level segmentation labeling for specific organs or lesions both expertise-dependent and time-consuming. Early work focused on single-organ or single-lesion annotated datasets [14, 15, 15, 32–36]. These single-task datasets provide foundational data for training automatic segmentation models, significantly enhancing model performance in segmenting specific organs or lesions. With the growth of multi-task requirements, multi-organ and multi-tissue annotated datasets [14, 37–39] have emerged, facilitating the exploration of deep learning models in multi-region segmentation tasks [40, 41]. In recent years, visual foundational models have demonstrated impressive cross-domain and multi-task transfer capabilities, motivating researchers to build large-scale, multimodal datasets for medical foundational models. The most common approach is to integrate data from various sources [22–24, 31] to bridge the scale gap compared to natural image datasets. However, due to the sparsity of annotations in the source data, the combined dataset still lags behind natural image datasets in terms of mask quantity and density. For example, the SA-1B [17] dataset averages 100 masks per image, whereas large-scale medical datasets like COSMOS [24] and SA-Med2D-20M [42] have fewer than 5 masks per image on average. This gap limits the ability of medical datasets to support dense segmentation tasks and fine-grained interactions, as well as the applicability of foundational models to \"segment anything\" in the medical imaging domain.", + "bbox": [ + 511, + 205, + 906, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Algorithms. IMIS methods are generally more effective than automated methods at generating high-quality results that meet clinical requirements, especially in scenarios involving diverse imaging protocols, complex pathological variations, and ambiguous lesion boundaries [43]. User interactions can be take several typical forms, such as scribbles [44, 45], bounding boxes [17, 46], clicks [2, 47], or language prompts [17, 48, 49]. Traditional methods approach this task through energy by minimizing energy on a regular pixel grid, capturing low-level appearance features with unary potentials and promoting consistent segmentation outputs with pairwise or higher-order potentials [44, 50]. With advancements in deep learning, researchers have explored using user prompts directly as network input features, often in combination with 2D or 3D neural networks to produce segmentation results [1, 2, 17, 44-47]. Lever", + "bbox": [ + 511, + 659, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20842", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "aging the advantage of pretraining on large-scale datasets, the Segment Anything Model (SAM) has become a benchmark for IMIS. For example, SAM can serve as a powerful pretrained encoder-decoder model fine-tuned for specific tasks [51, 52]. However, these approaches essentially revert to the design of fully automated segmentation models. Another category of methods retains the interactive aspect of segmentation, focusing on parameter-efficient fine-tuning (PEFT) techniques [53, 54]. However, these methods consider only a limited of interaction strategies, and inconsistencies in evaluation have affected the comparability and reliability of their results. To achieve robust segmentation capabilities across diverse medical imaging datasets, some approaches fine-tune SAM's decoder using large-scale medical datasets [22-24]. Although these methods demonstrate significant advantages across different medical modalities and tasks, they have not yet been consistently benchmarked against similar methods and lack a thorough examination of how different interaction strategies affect outcomes.", + "bbox": [ + 89, + 90, + 480, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we evaluate these state-of-the-art methods on the IMed-361M dataset, providing a comprehensive and fair comparison, while also exploring the impact of different interaction strategies on medical image segmentation.", + "bbox": [ + 89, + 378, + 480, + 440 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. IMIS Benchmark Dataset:IMed-361M", + "text_level": 1, + "bbox": [ + 89, + 454, + 442, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present IMed-361M, the first large-scale IMIS dataset, that significantly surpasses existing datasets in terms of scale, diversity, quality, and density. This section details the data collection and pre-processing procedures, along with an in-depth analysis of the dataset's advantages and potential applications.", + "bbox": [ + 89, + 479, + 482, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Collection and Pre-processing", + "text_level": 1, + "bbox": [ + 89, + 582, + 397, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data collection. We integrate over 110 publicly available medical image segmentation datasets from globally recognized platforms such as TCIA $^{1}$ , OpenNeuro $^{2}$ , NITRC $^{3}$ , Grand Challenge $^{4}$ , Synapse $^{5}$ , CodaLab $^{6}$ , and GitHub $^{7}$ , covering both 2D and 3D images and variety of formats (e.g., .jpg, .npy, .nii). Additionally, we collaborate with several medical institutions, acquiring diverse clinical data through rigorous ethical review processes, which further enriches the dataset. A detailed list of data sources is provided in the supplementary material.", + "bbox": [ + 89, + 604, + 482, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preprocessing and filtering. We first standardize all collected medical images according to the SA-Med2D-20M [42] protocol, converting the corresponding ground", + "bbox": [ + 89, + 756, + 482, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "truth (GT) into one-hot encoding and storing it in .npz files in compressed sparse row (CSR) format [55]. Then, we apply the following exclusion criteria: (1) exclude 3D slice images and their corresponding masks with an aspect ratio greater than 1.5 to avoid interference from images with excessive geometric distortion during model training; (2) exclude masks where the foreground area accounts for less than one-thousandth of the total pixels, as such small foreground regions are prone to being lost during resizing, making it challenging to generate effective prompts.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Resolving conflicts and ambiguities. To address conflicts and ambiguities in the GT, we first standardize expressions for the same target. For instance, both \"lung nodule\" and \"pulmonary nodule\" are renamed to \"lung nodule\" for consistency. We then manually review and correct misalignments and informational errors in the dataset to ensure annotation accuracy. Finally, for annotations with multiple connected components, we differentiate and label them based on clinical needs to avoid potential misunderstandings that could arise from single-point interactions.", + "bbox": [ + 511, + 243, + 903, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through this process, we collected 6.4 million images and 87.6 million GT masks that were manually annotated and corrected. Although this already exceeds the data volume of most existing segmentation datasets, it still does not meet our fundamental requirement for the IMIS dataset: diverse and densely masks. Therefore, we further introduce interactive masks to address the above issues.", + "bbox": [ + 511, + 395, + 903, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Interactive Masks", + "text_level": 1, + "bbox": [ + 513, + 513, + 687, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Automatic mask generation. The automatic mask generation method employed by SAM has been proven to be high-quality and effective [17, 56, 57]. We leverage SAM's object-awareness capability to generate as many masks as possible for each image. Specifically, we employ a $32 \\times 32$ point grid to guide the model, generating a set of candidate masks for each point corresponding to potential objects of interest. The generated masks are refined and optimized using the following strategies: (1) Confidence filtering: Only masks with an Intersection over Union (IoU) prediction confidence score above 0.85 are retained to ensure segmentation accuracy. (2) Non-Maximum Suppression (NMS) [58]: For overlapping masks with an IoU greater than 0.7, only the mask with the highest confidence score is kept to eliminate redundancy. (3) Remove background masks: Masks covering more than $80\\%$ of the area are discarded, as such large-coverage foreground masks are almost nonexistent in the GT.", + "bbox": [ + 511, + 537, + 903, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Quality control and granularity management. Based on our observations, the generated masks often fail to fully separate structures with unclear boundaries, such as the left atrium, myocardium, and right atrium in the heart. Additionally, masks for dispersed structures like the intestines are often identified as multiple separate objects. To address", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1https://www.cancerimagingarchive.net", + "bbox": [ + 109, + 811, + 320, + 825 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://openneuro.org", + "bbox": [ + 109, + 825, + 230, + 838 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3https://www.nitrc.org", + "bbox": [ + 109, + 838, + 230, + 851 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "4 https://grand-challenge.org", + "bbox": [ + 109, + 851, + 259, + 863 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "5https://www.synapse.org", + "bbox": [ + 109, + 863, + 246, + 875 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{6}$ https://codalab.org", + "bbox": [ + 109, + 875, + 215, + 887 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "7https://github.com", + "bbox": [ + 109, + 887, + 212, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20843", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ceb09083a3817149f6129ec884ea89bba6c5a7ba1e1c95f0344512db33ca1c26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 95, + 411, + 303 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/987aefc771c31c81122029accb4bf1b8ddc73ae69a2e9c4914bbc5176d60acc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 93, + 861, + 301 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a97f21ed46ca3e9ca40ba1cd83a9daa0682aec774842add361bcc23ef039bbd6.jpg", + "image_caption": [ + "Figure 1. Overview of the IMed-361M dataset. (a) Number of images and masks for each modality. (b) Information on six anatomical structures. (c) Distribution of image resolutions. (d) Analysis of mask proportions. (e) Comparison with other existing public datasets." + ], + "image_footnote": [], + "bbox": [ + 140, + 305, + 395, + 446 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cb91498ad17b1295f0c833c03ecf24b99e217d2b3e99ea84c170322ee82b12b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 306, + 635, + 448 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6a10aacc221f2292be08afe6ecb05c70eb43148ae2a089269835f7b71247a392.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 309, + 864, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "these issues, we utilize the original GT (described in Section 3.1) to correct the generated masks. We focus on two situations: (1) If the GT contains multi-connected regions, we directly replace the corresponding regions in the generated mask with the multi-connected regions from the GT; (2) We iterate over all single-connected regions in the GT, generating a minimum bounding box for each region. If there is a region in the generated mask with a bounding box that overlaps more than $95\\%$ with this GT bounding box, we retain the generated mask region; otherwise, we replace the region with the GT mask. Finally, we apply morphological operations (e.g., erosion and dilation) to remove noise and fill small holes [59].", + "bbox": [ + 88, + 518, + 483, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We ultimately obtained 273 million \"interactive masks\", which can be used to train interactive segmentation models and cover nearly all identifiable objects in medical images. While we acknowledge that some masks may lack direct clinical relevance, they provide considerable diversity and density in the identified regions, helping the model learn a variety of interactive operations and enhancing its generalization ability across various tasks.", + "bbox": [ + 89, + 715, + 483, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Statistics and Analysis", + "text_level": 1, + "bbox": [ + 89, + 848, + 302, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data scale. As shown in Fig.1 (a), the IMed-361M dataset contains 6.4 million images, 87.6 million GT, and 273.4", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "million interactive masks, averaging 56 masks per image. This makes it the largest publicly available, multimodal, interactive medical image segmentation dataset to date. Compared to the MedTrinity-25M [31] dataset, our dataset provides 14.4 times the number of masks (Fig.1 (e)).", + "bbox": [ + 511, + 520, + 906, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diversity. The IMed-361M dataset covers 14 imaging modalities and 204 segmentation targets, including organs and lesions, as shown in Fig.1 (b). We categorize the GT into six groups: Head and Neck, Thorax, Skeleton, Abdomen, Pelvis, and Lesions, covering nearly all parts of the human body. For clarity, we merge certain positional information when appropriate; for example, \"left lung\" and \"right lung\" are combined into \"lung.\" Detailed category information can be found in the supplementary materials. Fig.1 (c) and (d) show the distribution of image resolutions and mask coverage within the dataset. Over $83\\%$ of the images have resolutions between $256 \\times 256$ and $1024 \\times 1024$ , ensuring broad applicability and compatibility across various research scenarios. Additionally, most masks occupy less than $2\\%$ of the image area, reflecting the typically fine granularity of medical segmentation. Our interactive masks provide over one million instances across different coverage intervals, significantly enhancing the diversity and density of the dataset.", + "bbox": [ + 511, + 595, + 908, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mask quality. We randomly select five images from each", + "bbox": [ + 511, + 885, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20844", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/71c22a66f93559d1090516b3e084aa27d89827732e7acbe32574826723bf679e.jpg", + "image_caption": [ + "Figure 2. Evaluation of the quality of interactive masks." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 480, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "subset of IMed-361M (a total of 550 images) and invited four radiologists to evaluate the quality of the corresponding interactive masks. We categorize the mask quality into three levels: (1) \"Good\": the mask closely matches the manual annotations; (2) \"Satisfactory\": the mask has some visible errors (e.g., incomplete contours) but is generally acceptable; (3) \"Bad\": the mask has significant errors (e.g., over-segmentation or under-segmentation) and is unsuitable for clinical segmentation tasks. The evaluation results show that 156 images are rated as \"Good\", 345 as \"Satisfactory\", and 49 as \"Bad\". Statistical analysis reveals that the masks rated as \"Bad\" mainly come from 18 datasets, concentrated within the red-boxed area in Fig.2. Radiologists note that some excessively annotated structures hold little clinical relevance. Additionally, some masks include elements unrelated to human organs or lesions, such as letters in X-ray backgrounds or hair in dermatoscopic images. Based on this analysis, we manually remove the excessively annotated masks from these 18 datasets and apply a $0.5\\%$ foreground filter rate to retain only the final valid masks. Finally, we retain those masks unrelated to human organs or lesions to enhance the model's adaptability in different scenarios.", + "bbox": [ + 89, + 311, + 483, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. IMIS Baseline Network", + "text_level": 1, + "bbox": [ + 89, + 672, + 312, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Model Design", + "text_level": 1, + "bbox": [ + 89, + 696, + 233, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt a strategy similar to fine-tuning SAM [22-24] to establish the IMIS baseline, providing a performance benchmark for future work. As shown in Fig.3, IMIS-Net has three key components: an image encoder for extracting image features, a prompt encoder for integrating user interaction information, and a mask decoder for generating segmentation results using image and prompt embeddings.", + "bbox": [ + 89, + 719, + 482, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We select ViT-base [60] as the image encoder. While larger ViT models (e.g., ViT-Large, ViT-Huge) [22, 24] offer slight accuracy improvements, they significantly increase computational costs, making them impractical for clinical applications. Additionally, most open-source fine", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b5575537e62860e72e0805c96a128565d4110e03e1542e710ff212e9919bb526.jpg", + "image_caption": [ + "Figure 3. The training process of IMIS-Net simulates K consecutive steps of interactive segmentation." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 924, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tuning methods also use ViT-base, ensuring fairness in our baseline. The image input size is $1024 \\times 1024 \\times 3$ , with a patch size of $16 \\times 16 \\times 3$ . For the prompt encoder, we consider three prompt types: points, boxes, and text. Points and boxes are represented by the sum of positional encoding and learned embeddings, while text is encoded using CLIP's text encoder [61]. Text prompts follow the template: \"A segmentation area of a [category]\", covering over 200 organ and lesion categories to establish a benchmark for future multimodal segmentation research. The mask decoder uses Transformer decoder blocks, mapping image and prompt embeddings to the mask. The mask resolution is increased to $256 \\times 256$ through transposed convolutions and matched to the input size via bilinear interpolation. Notably, expanding the decoder can improve performance without significantly increasing network training parameters, thereby providing a simple solution to address the issue of model performance saturation.", + "bbox": [ + 511, + 294, + 906, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experiment Configuration", + "text_level": 1, + "bbox": [ + 511, + 575, + 754, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training strategy. An overview is provided of the simulated continuous interactive segmentation training [22, 23, 62] in Fig.3. For a given segmentation task and medical image $x_{t}$ , we first simulate a set of initial interactions $u_{1}^{g}$ and $u_{1}^{i}$ based on the corresponding ground truth $y^{g}$ and interactive mask $y^{i}$ , which include clicks, bboxes, and text input. The click points are uniformly sampled from the foreground regions of $y^{g}$ or $y^{i}$ , while the bboxes are defined as the smallest bounding box around the target, with an offset of 5 pixels added to each coordinate to simulate slight user bias during the interaction process. The entire training process involves $K$ interactive training iterations (with $K = 8$ in this paper). The model's initial predictions are $\\hat{y}_{1}^{g}$ and $\\hat{y}_{1}^{i}$ . After the first prediction, we simulate subsequent corrections based on the previous predictions $\\hat{y}_{k}^{g}$ and $\\hat{y}_{k}^{i}$ , as well as the error region $\\varepsilon_{k}$ between the $y^{g}$ and $y^{i}$ , where $k\\in \\{1,\\dots,K\\}$ . Additionally, we provide the low-resolution predicted mask from the previous prediction as an extra cue to the model. As can be seen, the image encoder only needs to encode the image once during the training, and subse", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20845", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1c5b9c29eb057a2c076ea4b903865f136237132afb5caa782696852008d15ee8.jpg", + "image_caption": [ + "Figure 4. Comparison of IMIS-Net with existing foundation models, with performance statistics at both image and mask levels." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 486, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "quent interactive training only updates the prompt encoder and mask decoder parameters.", + "bbox": [ + 89, + 277, + 483, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training setting. We use the Adam optimizer [63] with a learning rate of $2 \\times 10^{-5}$ . The training was conducted on 72 NVIDIA 4090 GPUs with a batch size of 2. For each image, we randomly select 5 targets from the corresponding GT and interactive masks as supervision targets (if fewer than 5 targets are present, they are selected repeatedly). The image resolution was uniformly resized to $1024 \\times 1024$ , and pixel intensities are randomly scaled and shifted with a probability of $20\\%$ , where the scaling factor and offset are set to 0.2. We train the model for a total of 12 epochs on the IMed-361M dataset and select the final checkpoint as the model weights. A linear combination of Focal loss [64] and Dice loss [65] is used as the loss function, balancing their influence in a 20:1 ratio. Finally, the Dice score is used as the primary evaluation metric for this study.", + "bbox": [ + 89, + 309, + 483, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation Model and Data", + "text_level": 1, + "bbox": [ + 89, + 547, + 339, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model. We evaluate five visual foundation models [17, 18, 22-24], selected based on their pretraining on large-scale datasets and the availability of source code. Among them, MedSAM, SAM-Med2D, and Huang et al. [24] are specifically designed for IMIS, while SAM [17] and SAM-2 [18] are pre-trained on massive natural image and video datasets. To ensure fairness in the evaluation, all models used ViT-Base as the encoder. It is important to note that, MedSAM and Huang et al. [24] only supportbbox-based interactivity (using the models released by the authors), while the other models support points, bboxes, and combinations of both.", + "bbox": [ + 89, + 571, + 482, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data. The data used to evaluate the model is divided into two categories: the test set and the external dataset. The test set contains 58,411 images, with a distribution consistent with that of IMed-361M. The external dataset includes data from three different modalities and clinical scenarios: SegThor [66], which is used for segmentation of thoracic organs at risk in 516 CT images; TotalSegmentator MRI [37], which is used for segmenting organs in 2,147 MRI images; and ISLES $^{8}$ , which belongs to the MRI FLAIR modality", + "bbox": [ + 89, + 738, + 482, + 875 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and includes 364 images of ischemic stroke lesions. These datasets were not used during the training of IMIS-Net.", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 511, + 137, + 601, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Main Results", + "text_level": 1, + "bbox": [ + 511, + 162, + 653, + 178 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig.4 shows the performance evaluation of IMIS-Net compared to other vision foundation models on the single-interaction segmentation task. During testing, all models used the same points and bounding boxes as prompts. The results indicate that IMIS-Net outperforms other models in both image and mask-level statistics. Thebbox interaction of different models always outperforms the click interaction because the_bbox can provide more boundary information. Notably, despite being pretrained on large-scale medical image datasets, MedSAM and SAM-Med2D still exhibit significant performance differences. This disparity is closely related to the scale and diversity of the pretraining datasets. As shown in Fig.5(a) and (b), the pre-training dataset for SAM-Med2D lacks samples of skeletal structures, resulting in poorer segmentation performance for these anatomical structures. Additionally, under the single-point prompt condition, SAM and SAM-2 achieve only Dice scores of $60.26\\%$ and $59.57\\%$ , respectively. likely due to the absence of medical knowledge in the pretraining data and limited interactive information constraining model performance.", + "bbox": [ + 511, + 185, + 903, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We increase the number of interactions from 1 to 9 to observe the performance changes of the model (Fig.5 (c)). As expected, performance improves with more interactions, and the gap between models narrows. This is because multiple interactions reduce the task difficulty. We also examine the impact of click position andbbox offset on performance. As shown in Fig.5 (d) and (f), when the prompt point is closer to the centroid, the performance improvement is more significant, with SAM-2's Dice score increasing by $2.84\\%$ . Additionally, withbbox offset, all methods experience a performance decline of $0.85\\% - 3.94\\%$ . Our model exhibited the smallest performance drop, making it more robust for practical use. In summary, leveraging the rich data and mask diversity of the IMed-361M dataset, our model delivers the best performance across various medical scenarios and interaction strategies. More experimental results and analysis can be found in the supplementary material.", + "bbox": [ + 511, + 503, + 903, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. External Dataset Evaluation", + "text_level": 1, + "bbox": [ + 511, + 787, + 767, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab.1 presents evaluation results on external datasets. IMIS-Net achieves the best average performance across datasets from three different sources and tasks. We highlight the Dice scores of the 12 major abdominal organs in the TotalSegmentator MRI dataset, where IMIS-Net outperforms in 10. On the ISLES dataset, focusing on ischemic stroke", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{8}$ https://www.isles-challenge.org/", + "bbox": [ + 107, + 886, + 289, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20846", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2a8b3f8de4bb386e96155e9544a21ffcee3ed7b251269b363a3a5de7bba1fc8.jpg", + "image_caption": [ + "(a) Single click interaction" + ], + "image_footnote": [], + "bbox": [ + 166, + 88, + 415, + 248 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be5805d233413910604db172b5b082d819ebc47cc036dbc431d70b4d35cfe428.jpg", + "image_caption": [ + "(b) Bounding box interaction" + ], + "image_footnote": [], + "bbox": [ + 428, + 88, + 676, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75fe536c264e83334298718fe107c66602369fd1b9adb65ede1e175be32aff0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 179, + 852, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/612c43c91ccc971dfb071c9cdf0c180dc4259a164de7097de8865514f4abb8de.jpg", + "image_caption": [ + "(c) Number of interactions" + ], + "image_footnote": [], + "bbox": [ + 138, + 268, + 372, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ffcf64c8b0115fbe913f27f96189ea7a08a4bb2bfb216c1b8209e4ed866f9d13.jpg", + "image_caption": [ + "(d) Sampling points distances from the centroid" + ], + "image_footnote": [], + "bbox": [ + 377, + 267, + 614, + 409 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e5a2e0b0d2fb0f7966e3edba3e081477ef5df3bec752e0e573c17041ee013ff4.jpg", + "image_caption": [ + "(e) Bbox offset", + "Figure 5. Comparison of segmentation performance across different anatomical structures under single-click (a) and bounding box (b) interactions. (c) Changes in segmentation performance with increasing interaction numbers. (d) and (e) Impact of click position and bounding box offset on performance." + ], + "image_footnote": [], + "bbox": [ + 620, + 268, + 857, + 409 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/58f69a42f3ad53f97457c0fbe08cbadbffe52105143e94ced010130136e6fcbd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetCategorySAMSAM 2MedSAMHuang et al.SAM-Med2DIMIS-Net
ISLESIschemic Stroke Lesion55.9260.1459.9056.7968.2271.78
SegThorEsophagus78.1986.6047.1776.4882.0489.17
Heart91.4192.3481.1692.0294.1881.27
Aorta82.3683.6955.6884.5779.1294.92
Trachea93.7694.2472.4094.6085.1090.04
Average84.4685.8660.5584.7386.4389.27
Totalsegmentator MRIAdrenal gland77.1982.2556.5868.5266.3577.53
Aorta76.8879.9055.9880.0874.9185.62
Colon56.6956.0042.5355.4151.4473.47
Duodenum69.3074.8155.6472.3869.6775.97
Gallbladder81.3384.8268.1583.4687.6382.93
Iliopsoas59.5771.9658.4863.6760.7672.55
Kidney79.4381.5958.4181.0974.0482.72
Inferior vena cava84.9487.0357.7385.8878.3487.35
Liver86.9388.7068.1487.3586.7590.62
Pancreas66.9569.8646.1066.3066.8470.84
Spleen81.1283.0566.6183.8784.9586.74
Stomach77.1081.2362.7576.5279.3181.45
Average75.4577.6259.5275.4775.9279.06
", + "bbox": [ + 135, + 491, + 864, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison results of IMIS-Net against five other interactive segmentation methods on three external datasets.", + "bbox": [ + 107, + 790, + 885, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lesion segmentation, our model surpasses the second-best SAM-Med2D by $3.56\\%$ . These results demonstrate the strong generalization ability of our method.", + "bbox": [ + 89, + 830, + 485, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 830, + 666, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling Up Training Data. As shown in Fig.6 (a) and (b), when trained only on the GT from IMed-361M, IMIS-Net performs poorly. However, with the addition of interac", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20847", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c002bcfa294e2975a2edafa0d32d8a3b9e28387a5f0f2e30416334d86b4c91ec.jpg", + "image_caption": [ + "Figure 6. (a) and (b) segmentation performance increases with the increasing mask density or data scales up. (c) effectiveness of text and its combination with other prompts." + ], + "image_footnote": [], + "bbox": [ + 112, + 90, + 352, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad5a3eb647941b192851a6c2ce706e24d785abe1a704602861fd5cb0ca09d75e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 90, + 570, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3801d421fbf7d1668f29fef63ddb9de94aa1b2e2717461450f64c2ac6742946a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 549, + 90, + 880, + 246 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a1a1d096a647a8073fa0a3f4947a5c631bb282eefb161b9eb3b6a6553308a083.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Decoder dimensionImage resolutionPromptTraining parameters
PointBbox
768256×2560.82140.846929.68 M
768512×5120.86730.896829.68 M
2561024×10240.83660.84975.52 M
5121024×10240.85630.872915.19 M
7681024×10240.88480.906029.68 M
", + "bbox": [ + 91, + 314, + 488, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation study of model design, including decoder dimension and image resolution.", + "bbox": [ + 89, + 469, + 483, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tive masks, the Dice score increases rapidly. Similar performance trends are observed when training on datasets of different sizes, indicating that our method is scalable and performs better with more available training data.", + "bbox": [ + 89, + 531, + 483, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text and Combination Prompts. IMIS-Net, as a versatile interactive model, can utilize text prompts to achieve the segmentation of over 200 organs and lesions. As shown in Fig.6 (c), the model achieves a segmentation performance of $76.30\\%$ when using only text prompts. When both text and point prompts are combined, the average Dice score increases by $11.95\\%$ . Furthermore, after three rounds of click-based correction, the Dice score reaches $89.69\\%$ . These results show that combining different prompts synergistically enhances the model's segmentation capabilities.", + "bbox": [ + 89, + 595, + 483, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model Design. We conduct an ablation study on different configurations of the model, including the impact of input resolution and decoder dimension on model performance. As shown in Tab.2, the model's segmentation performance improves as the training image resolution increases. This higher input resolution allows for clearer visualization of lesions and organs in medical images. Additionally, when the decoder dimension is increased from 256 to 768, the model's performance improves from $84.97\\%$ to $90.60\\%$ , with only a 24.16M increase in trainable parameters. These", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "results demonstrate the scalability of IMIS-Net, showing that even when the model's performance approaches saturation on larger datasets, further performance improvements can still be achieved by expanding the decoder.", + "bbox": [ + 511, + 316, + 906, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 393, + 633, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce IMed-361M, a benchmark dataset dedicated to interactive medical image segmentation. It includes a vast array of medical images across various modalities, extensive segmentation scenarios, and densely masks, surpassing all existing datasets that are limited to single tasks or simple integrations. Leveraging this data resource, we developed a general IMIS baseline model that enables users to generate segmentation results tailored to clinical needs through interactive methods, including clicks, bounding boxes, text prompts, and their combinations. We performed a comprehensive comparison of this baseline with existing foundational models, demonstrating that our model provides significant performance advantages and exhibits strong transferability in previously unseen scenarios. Notably, our approach typically requires fewer interactions to achieve comparable performance, enhancing its practicality in real-world applications.", + "bbox": [ + 511, + 420, + 906, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The IMed-361M dataset will strongly facilitate the development of foundational models in medical imaging and lay the foundation for fair evaluation across different models. IMIS-Net provides general technical support for various clinical applications, accelerating the widespread application of AI technology in the medical field. Despite these achievements, we recognize that this work still faces several challenges. For instance, effectively obtaining semantic information for interactive masks and extending this approach to more comprehensive and finer-grained medical image analysis scenarios are areas that require further exploration and improvement in the future.", + "bbox": [ + 511, + 678, + 908, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: The authors acknowledge the funding provided by the National Key Research and Development Program of China (No.2022ZD0160102).", + "bbox": [ + 511, + 844, + 906, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20848", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tianfei Zhou, Lielei Li, Gustav Bredell, Jianwu Li, Jan Unkelbach, and Ender Konukoglu. Volumetric memory network for interactive medical image segmentation. Medical Image Analysis, 83:102599, 2023. 1, 2", + "[2] Jian Zhang, Yinghuan Shi, Jinquan Sun, Lei Wang, Luping Zhou, Yang Gao, and Dinggang Shen. Interactive medical image segmentation via a point-based interaction. Artificial Intelligence in Medicine, 111:101998, 2021. 1, 2", + "[3] Fang Chen, Haojie Han, Peng Wan, Hongen Liao, Chunrui Liu, and Daoqiang Zhang. Joint segmentation and differential diagnosis of thyroid nodule in contrast-enhanced ultrasound images. IEEE Transactions on Biomedical Engineering, 70(9):2722-2732, 2023. 1", + "[4] Zdravko Marinov, Paul F Jäger, Jan Egger, Jens Kleesiek, and Rainer Stiefelhagen. Deep interactive segmentation of medical images: A systematic review and taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[5] Pierre-Henri Conze, Gustavo Andrade-Miranda, Vivek Kumar Singh, Vincent Jaouen, and Dimitris Visvikis. Current and emerging trends in medical image segmentation with deep learning. IEEE Transactions on Radiation and Plasma Medical Sciences, 7(6):545-569, 2023. 1", + "[6] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[7] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015.", + "[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021.", + "[9] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, pages 3-11. Springer, 2018.", + "[10] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024.", + "[11] Junlong Cheng, Shengwei Tian, Long Yu, Chengrui Gao, Xiaojing Kang, Xiang Ma, Weidong Wu, Shijia Liu, and" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hongchun Lu. Resganet: Residual group attention network for medical image classification and segmentation. Medical Image Analysis, 76:102313, 2022. 1", + "[12] Marco Cascella, Jonathan Montomoli, Valentina Bellini, and Elena Bignami. Evaluating the feasibility of chatgpt in healthcare: an analysis of multiple clinical and research scenarios. Journal of medical systems, 47(1):33, 2023. 1", + "[13] Mingze Yuan, Yingda Xia, Hexin Dong, Zifan Chen, Jiawen Yao, Mingyan Qiu, Ke Yan, Xiaoli Yin, Yu Shi, Xin Chen, et al. Devil is in the queries: advancing mask transformers for real-world medical image segmentation and out-of-distribution localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23879-23889, 2023. 1", + "[14] Jakob Wasserthal, Hanns-Christian Breit, Manfred T Meyer, Maurice Pradella, Daniel Hinck, Alexander W Sauter, Tobias Heye, Daniel T Boll, Joshy Cyriac, Shan Yang, et al. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence, 5(5), 2023. 1, 2", + "[15] Sergios Gatidis, Tobias Hepp, Marcel Fruh, Christian La Fougère, Konstantin Nikolaou, Christina Pfaffenberg, Bernhard Schölkopf, Thomas Küstner, Clemens Cyran, and Daniel Rubin. A whole-body fdg-pet/ct dataset with manually annotated tumor lesions. Scientific Data, 9(1):601, 2022. 1, 2", + "[16] Simon Graham, Mostafa Jahanifar, Quoc Dang Vu, Giorgos Hadjigeorghiou, Thomas Leech, David Snead, Shan E Ahmed Raza, Fayyaz Minhas, and Nasir Rajpoot. Conic: Colon nuclei identification and counting challenge 2022. arXiv preprint arXiv:2111.14485, 2021. 1", + "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3, 6", + "[18] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 6", + "[19] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[20] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, et al. Segment anything in 3d with nerfs. Advances in Neural Information Processing Systems, 36:25971-25990, 2023.", + "[21] Di Wang, Jing Zhang, Bo Du, Minqiang Xu, Lin Liu, Dacheng Tao, and Liangpei Zhang. Samrs: Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[22] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 2, 3, 5, 6" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20849", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Junlong Cheng, Jin Ye, Zhongying Deng, Jianpin Chen, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Ji-long Chen, Lei Jiang, et al. Sam-med2d. arXiv preprint arXiv:2308.16184, 2023. 1, 5", + "[24] Yuhao Huang, Xin Yang, Lian Liu, Han Zhou, Ao Chang, Xinrui Zhou, Rusi Chen, Junxuan Yu, Jiongquan Chen, Chaoyu Chen, et al. Segment anything model for medical images? Medical Image Analysis, 92:103061, 2024. 2, 3, 5, 6", + "[25] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023.", + "[26] Yichi Zhang, Zhenrong Shen, and Rushi Jiao. Segment anything model for medical image segmentation: Current applications and future directions. Computers in Biology and Medicine, page 108238, 2024. 2", + "[27] Shaoting Zhang and Dimitris Metaxas. On the challenges and perspectives of foundation models for medical image analysis. Medical image analysis, 91:102996, 2024. 2", + "[28] Kummaragunta Joel Prabhod and Asha Gadhiraju. Foundation models in medical imaging: Revolutionizing diagnostic accuracy and efficiency. Journal of Artificial Intelligence Research and Applications, 4(1):471-511, 2024.", + "[29] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024.", + "[30] Raphael Schäfer, Till Nicke, Henning Hofener, AnnKristin Lange, Dorit Merhof, Friedrich Feuerhake, Volkmar Schulz, Johannes Lotz, and Fabian Kiessling. Overcoming data scarcity in biomedical imaging with a foundational multitask model. Nature Computational Science, 4(7):495-509, 2024. 2", + "[31] Yunfei Xie, Ce Zhou, Lang Gao, Juncheng Wu, Xianhang Li, Hong-Yu Zhou, Sheng Liu, Lei Xing, James Zou, Cihang Xie, et al. Medtrinity-25m: A large-scale multimodal dataset with multigranular annotations for medicine. arXiv preprint arXiv:2408.02900, 2024. 2, 4", + "[32] Kenneth Clark, Bruce Vendt, Kirk Smith, John Freymann, Justin Kirby, Paul Koppel, Stephen Moore, Stanley Phillips, David Maffitt, Michael Pringle, et al. The cancer imaging archive (tcia): maintaining and operating a public information repository. Journal of digital imaging, 26:1045-1057, 2013. 2", + "[33] Joao Pedrosa, Guilherme Aresta, Carlos Ferreira, Gurraj Atwal, Hady Ahmady Phoulady, Xiaoyu Chen, Rongzhen Chen, Jiaoliang Li, Liansheng Wang, Adrian Galdran, et al. Lndb challenge on automatic lung cancer patient management. Medical image analysis, 70:102027, 2021.", + "[34] Rina D Rudyanto, Sjoerd Kerkstra, Eva M Van Rikxoort, Catalin Fetita, Pierre-Yves Brillet, Christophe Lefevre, Wenzhe Xue, Xiangjun Zhu, Jianming Liang, Ilkay Oksuz, et al. Comparing algorithms for automated vessel segmentation in computed tomography scans of the lung: the vessel12 study. Medical image analysis, 18(7):1217-1232, 2014." + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Michela Antonelli, Annika Reinke, Spyridon Bakas, Keyvan Farahani, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, Olaf Ronneberger, Ronald M Summers, et al. The medical segmentation decathlon. Nature communications, 13(1):4128, 2022.", + "[36] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia modeling: 26th international conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, proceedings, part II 26, pages 451–462. Springer, 2020. 2", + "[37] Tugba Akinci D'Antonoli, Lucas K Berger, Ashraya K Indrakanti, Nathan Vishwanathan, Jakob Weiβ, Matthias Jung, Zeynep Berkarda, Alexander Rau, Marco Reisert, Thomas Küstner, et al. Totalsegmentator mr: Sequence-independent segmentation of 59 anatomical structures in mr images. arXiv preprint arXiv:2405.19492, 2024. 2, 6", + "[38] Igor Zingman, Birgit Stierstorfer, Charlotte Lempp, and Fabian Heinemann. Learning image representations for anomaly detection: application to discovery of histological alterations in drug development. Medical Image Analysis, 92:103067, 2024.", + "[39] Gašper Podobnik, Primož Strojan, Primož Peterlin, Bulat Ibragimov, and Tomaž Vrtovec. Han-seg: The head and neck organ-at-risk ct and mr segmentation dataset. Medical physics, 50(3):1917-1927, 2023. 2", + "[40] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 2", + "[41] Ziyan Huang, Haoyu Wang, Zhongying Deng, Jin Ye, Yanzhou Su, Hui Sun, Junjun He, Yun Gu, Lixu Gu, Shaoting Zhang, et al. Stu-net: Scalable and transferable medical image segmentation models empowered by large-scale supervised pre-training. arXiv preprint arXiv:2304.06716, 2023. 2", + "[42] Jin Ye, Junlong Cheng, Jianpin Chen, Zhongying Deng, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Jilong Chen, Lei Jiang, et al. Sa-med2d-20m dataset: Segment anything in 2d medical imaging with 20 million masks. arXiv preprint arXiv:2311.11969, 2023. 2, 3", + "[43] Sílvia Delgado Olabarriaga and Arnold WM Smeulders. Interaction in the segmentation of medical images: A survey. Medical image analysis, 5(2):127-142, 2001. 2", + "[44] Guotai Wang, Maria A Zuluaga, Rosalind Pratt, Michael Aertsen, Tom Doel, Maria Klusmann, Anna L David, Jan Deprest, Tom Vercauteren, and Sébastien Ourselin. Slic-seg: A minimally interactive segmentation of the placenta from sparse and motion-corrupted fetal mri in multiple views. Medical image analysis, 34:137-147, 2016. 2", + "[45] Hallee E Wong, Marianne Rakic, John Guttag, and Adrian V Dalca. Scribbleprompt: Fast and flexible interactive segmentation for any medical image. arXiv preprint arXiv:2312.07381, 2023. 2", + "[46] Martin Rajchl, Matthew CH Lee, Ozan Oktay, Konstantinos Kamnitsas, Jonathan Passerat-Palmbach, Wenjia Bai, Mellisa Damodaram, Mary A Rutherford, Joseph V Hajnal," + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20850", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bernhard Kainz, et al. Deepcut: Object segmentation from bounding box annotations using convolutional neural networks. IEEE transactions on medical imaging, 36(2):674-683, 2016. 2", + "[47] Tomas Sakinis, Fausto Milletari, Holger Roth, Panagiotis Korfiatis, Petro Kostandy, Kenneth Philbrick, Zeynett Akkus, Ziyue Xu, Daguang Xu, and Bradley J Erickson. Interactive segmentation of medical images through fully convolutional neural networks. arXiv preprint arXiv:1903.08205, 2019. 2", + "[48] Yuxin Du, Fan Bai, Tiejun Huang, and Bo Zhao. Segvol: Universal and interactive volumetric medical image segmentation. arXiv preprint arXiv:2311.13385, 2023. 2", + "[49] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086-7096, 2022. 2", + "[50] Leo Grady, Thomas Schiwietz, Shmuel Aharon, and Rüdiger Westermann. Random walks for interactive organ segmentation in two and three dimensions: Implementation and validation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2005: 8th International Conference, Palm Springs, CA, USA, October 26-29, 2005, Proceedings, Part II 8, pages 773-780. Springer, 2005. 2", + "[51] Mingzhe Hu, Yuheng Li, and Xiaofeng Yang. Skinsam: adapting the segmentation anything model for skin cancer segmentation. In Medical Imaging 2024: Image Perception, Observer Performance, and Technology Assessment, volume 12929, pages 174-184. SPIE, 2024. 3", + "[52] Yuheng Li, Mingzhe Hu, and Xiaofeng Yang. Polyp-sam: Transfer sam for polyp segmentation. In Medical Imaging 2024: Computer-Aided Diagnosis, volume 12927, pages 759-765. SPIE, 2024. 3", + "[53] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 3", + "[54] Jay N Paranjape, Nithin Gopalakrishnan Nair, Shameema Sikder, S Swaroop Vedula, and Vishal M Patel. Adaptivesam: Towards efficient tuning of sam for surgical scene segmentation. In Annual Conference on Medical Image Understanding and Analysis, pages 187-201. Springer, 2024. 3", + "[55] Urban Borstnik, Joost VandeVondele, Valery Weber, and Jürg Hutter. Sparse matrix multiplication: The distributed block-compressed sparse row library. Parallel Computing, 40(5-6):47-58, 2014. 3", + "[56] Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[57] Yizhe Zhang, Tao Zhou, Shuo Wang, Peixian Liang, Yejia Zhang, and Danny Z Chen. Input augmentation with sam: Boosting medical image segmentation with segmentation foundation model. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 129-139. Springer, 2023. 3" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Charalampos Symeonidis, Ioannis Mademlis, Ioannis Pitas, and Nikos Nikolaidis. Neural attention-driven nonmaximum suppression for person detection. IEEE transactions on image processing, 32:2454-2467, 2023. 3", + "[59] Suresh Kumar Tummala et al. Morphological operations and histogram analysis of sem images using python. Indian Journal of Engineering and Materials Sciences (IJEMS), 29(6):796-800, 2023. 4", + "[60] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5", + "[61] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[62] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. 5", + "[63] Zijun Zhang. Improved adam optimizer for deep neural networks. In 2018 IEEE/ACM 26th international symposium on quality of service (IWQoS), pages 1-2. IEEE, 2018. 6", + "[64] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 6", + "[65] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6", + "[66] Zoé Lambert, Caroline Petitjean, Bernard Dubray, and Su Kuan. Segthor: Segmentation of thoracic organs at risk in ct images. In 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA), pages 1- 6. IEEE, 2020. 6" + ], + "bbox": [ + 516, + 90, + 903, + 627 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20851", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/72ec2f30-d729-4ccc-bc2d-90a1d1bd3d41_model.json b/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/72ec2f30-d729-4ccc-bc2d-90a1d1bd3d41_model.json new file mode 100644 index 0000000000000000000000000000000000000000..73346890793ff8c09a81b50aa6459b8e9865a0e5 --- /dev/null +++ b/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/72ec2f30-d729-4ccc-bc2d-90a1d1bd3d41_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.131, + 0.892, + 0.153 + ], + "angle": 0, + "content": "Interactive Medical Image Segmentation: A Benchmark Dataset and Baseline" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.179, + 0.908, + 0.217 + ], + "angle": 0, + "content": "Junlong Cheng\\(^{1,2}\\) Bin Fu\\(^{2,4}\\) Jin Ye\\(^{2}\\) Guoan Wang\\(^{2}\\) Tianbin Li\\(^{2}\\) Haoyu Wang\\(^{2}\\) Ruoyu Li\\(^{1}\\) He Yao\\(^{1}\\) Junren Cheng\\(^{1}\\) Jinwen Li\\(^{3}\\) Yanzhou Su\\(^{2}\\) Min Zhu\\(^{1*}\\) Junjun He\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.22, + 0.707, + 0.238 + ], + "angle": 0, + "content": "1College of Computer Science, Sichuan University" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.239, + 0.635, + 0.255 + ], + "angle": 0, + "content": "2Shanghai AI Laboratory, China" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.256, + 0.77, + 0.273 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Computer Science and Technology, Xinjiang University" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.274, + 0.813, + 0.29 + ], + "angle": 0, + "content": "4 Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.297, + 0.738, + 0.311 + ], + "angle": 0, + "content": "{chengjunlong, fubin, yejin, wangguoan, litianbin, wanghaoyu} @pjlab.org.cn," + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.315, + 0.772, + 0.329 + ], + "angle": 0, + "content": "{ruoyuli, 2024323040035, chengjunren} \\(@\\) stu.scu.edu.cn, 107556522204@stu.xju.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.333, + 0.725, + 0.347 + ], + "angle": 0, + "content": "yanzhou.su@outlook.com, hejunjun.su@pjlab.org.cn, zhumin@scu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.361, + 0.327, + 0.377 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.403, + 0.486, + 0.812 + ], + "angle": 0, + "content": "Interactive Medical Image Segmentation (IMIS) has long been constrained by the limited availability of large-scale, diverse, and densely annotated datasets, which hinders model generalization and consistent evaluation across different models. In this paper, we introduce the IMed-361M benchmark dataset, a significant advancement in general IMIS research. First, we collect and standardize over 6.4 million medical images and their corresponding ground truth masks from multiple data sources. Then, leveraging the strong object recognition capabilities of a vision foundational model, we automatically generated dense interactive masks for each image and ensured their quality through rigorous quality control and granularity management. Unlike previous datasets, which are limited by specific modalities or sparse annotations, IMed-361M spans 14 modalities and 204 segmentation targets, totaling 361 million masks—an average of 56 masks per image. Finally, we developed an IMIS baseline network on this dataset that supports high-quality mask generation through interactive inputs, including clicks, bounding boxes, text prompts, and their combinations. We evaluate its performance on medical image segmentation tasks from multiple perspectives, demonstrating superior accuracy and scalability compared to existing interactive segmentation models. To facilitate research on foundational models in medical computer vision, we release the IMed-361M and model at https://github.com.uni-medical/IMIS-Bench." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.361, + 0.645, + 0.377 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.908, + 0.673 + ], + "angle": 0, + "content": "Interactive Medical Image Segmentation (IMIS) enables clinicians or users to guide the model by marking points, lines, or regions on an image, resulting in segmentation outcomes that better meet clinical needs [1, 2]. This user-informed segmentation approach not only optimizes results throughout the segmentation process, aiding clinicians in precisely localizing targets during diagnosis and treatment [3-5], but also addresses the limitations of fully automated segmentation models in generalizing to unseen object classes (i.e., zero-shot learning) [6-11], providing clinicians with a flexible adjustment mechanism that significantly enhances segmentation reliability. However, the advancement of IMIS advancement faces a critical bottleneck: the lack of high-quality, diverse, and large-scale datasets that capture the complex clinical environments [12, 13]. Existing public datasets (such as Totalsegmentator [14], AutoPET [15], and CoNIC [16]) are often tailored to specific modalities or tasks, thus constraining the generalization capabilities of models across diverse medical scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recently, as a foundational model for general visual segmentation, the Segment Anything Model (SAM) has garnered widespread attention [17, 18]. SAM integrates simple user interactions (such as points or bounding boxes) into the model's learning process and leverages pre-training on large-scale datasets to achieve cross-domain and multitask transferability [19-21]. Although SAM has been pretrained on over 1 billion natural image masks, its direct application in medical imaging remains limited due to the substantial domain differences between natural and medical images, as well as the lack of medical imaging-specific knowledge. To address these issues, current methods such as MedSAM [22] and SAM-Med2D [23] attempt to expand dataset size and fine-tune SAM to create IMIS models tailored for clinical scenarios. While these approaches have" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.185, + 0.9 + ], + "angle": 0, + "content": "†Project lead." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20841" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "shown some progress, their ability to 'segment anything' in medical imaging remains limited. This limitation is mainly due to the lack of densely masks in existing medical datasets compared to the SA-1B dataset [17] used to train SAM [22-24]. For instance, the COSMOS dataset [24] contains an average of only 5.7 masks per image, which restricting the model's capacity for dense segmentation and hindering comprehensive, fine-grained interaction. Moreover, many methods are evaluated only on specific modalities or with limited interaction strategies, further restricting the comprehensiveness and reliability of IMIS model evaluation [22, 24-26]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.277, + 0.485, + 0.549 + ], + "angle": 0, + "content": "Addressing these limitations requires the development of a high-quality IMIS benchmark dataset, which is essential for advancing foundational models in medical imaging [4, 27-30]. An ideal IMIS benchmark dataset should meet three core criteria: (1) Large-scale. The dataset should be large enough to fully support deep learning model training, enabling the model to effectively capture medical features; (2) Diversity. The dataset should encompass various medical imaging modalities and complex clinical scenarios to ensure the model's ability to generalize across modalities and tasks; (3) High-quality and densely masks. Accurate segmentation of complex medical images relies on high-quality annotations, and densely masks further enhance the model's capability to \"segment anything\" in medical imaging. Unfortunately, existing public medical segmentation datasets [22-24, 31] do not fully meet these standards, limiting their ability to comprehensively support and evaluate IMIS models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduce IMed-361M, a benchmark dataset specifically designed for IMIS tasks, this dataset integrates public and private data sources and utilizes foundational models [17] for automated annotation to generating dense masks for each image. A standardized data processing workflow is ensure the high quality and consistency of all masks. IMed-361M achieves unprecedented scale, diversity, and mask quality, comprising 6.4 million images spanning 14 imaging modalities and 204 targets, with a total of 361 million masks, averaging 56 masks per image. IMed-361M effectively solves the problems of small data size and sparse annotations, providing data support for IMIS model training. Additionally, we develop an IMIS baseline model and conduct a comprehensive evaluation of its performance across various medical scenarios, including its effectiveness on different modalities, anatomical structures, and organs, as well as the impact of various interaction strategies on model outcomes. This analysis provided an in-depth understanding of the strengths and limitations of different interactive segmentation methods, establishing a fair and consistent framework for evaluating IMIS model performance. We anticipate that the IMed-361M dataset and baseline model will drive the widespread adoption of IMIS" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "technology in clinical practice, accelerating the healthcare industry's transition toward intelligence and automation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.136, + 0.655, + 0.151 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.162, + 0.905, + 0.206 + ], + "angle": 0, + "content": "This section provides an overview of datasets used for various medical image segmentation tasks and reviews the advancements in IMIS algorithms." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.207, + 0.908, + 0.659 + ], + "angle": 0, + "content": "Datasets. Due to variations in imaging protocols, medical images often have multidimensional characteristics (e.g., 2D, 3D), making pixel-level segmentation labeling for specific organs or lesions both expertise-dependent and time-consuming. Early work focused on single-organ or single-lesion annotated datasets [14, 15, 15, 32–36]. These single-task datasets provide foundational data for training automatic segmentation models, significantly enhancing model performance in segmenting specific organs or lesions. With the growth of multi-task requirements, multi-organ and multi-tissue annotated datasets [14, 37–39] have emerged, facilitating the exploration of deep learning models in multi-region segmentation tasks [40, 41]. In recent years, visual foundational models have demonstrated impressive cross-domain and multi-task transfer capabilities, motivating researchers to build large-scale, multimodal datasets for medical foundational models. The most common approach is to integrate data from various sources [22–24, 31] to bridge the scale gap compared to natural image datasets. However, due to the sparsity of annotations in the source data, the combined dataset still lags behind natural image datasets in terms of mask quantity and density. For example, the SA-1B [17] dataset averages 100 masks per image, whereas large-scale medical datasets like COSMOS [24] and SA-Med2D-20M [42] have fewer than 5 masks per image on average. This gap limits the ability of medical datasets to support dense segmentation tasks and fine-grained interactions, as well as the applicability of foundational models to \"segment anything\" in the medical imaging domain." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Algorithms. IMIS methods are generally more effective than automated methods at generating high-quality results that meet clinical requirements, especially in scenarios involving diverse imaging protocols, complex pathological variations, and ambiguous lesion boundaries [43]. User interactions can be take several typical forms, such as scribbles [44, 45], bounding boxes [17, 46], clicks [2, 47], or language prompts [17, 48, 49]. Traditional methods approach this task through energy by minimizing energy on a regular pixel grid, capturing low-level appearance features with unary potentials and promoting consistent segmentation outputs with pairwise or higher-order potentials [44, 50]. With advancements in deep learning, researchers have explored using user prompts directly as network input features, often in combination with 2D or 3D neural networks to produce segmentation results [1, 2, 17, 44-47]. Lever" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20842" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.378 + ], + "angle": 0, + "content": "aging the advantage of pretraining on large-scale datasets, the Segment Anything Model (SAM) has become a benchmark for IMIS. For example, SAM can serve as a powerful pretrained encoder-decoder model fine-tuned for specific tasks [51, 52]. However, these approaches essentially revert to the design of fully automated segmentation models. Another category of methods retains the interactive aspect of segmentation, focusing on parameter-efficient fine-tuning (PEFT) techniques [53, 54]. However, these methods consider only a limited of interaction strategies, and inconsistencies in evaluation have affected the comparability and reliability of their results. To achieve robust segmentation capabilities across diverse medical imaging datasets, some approaches fine-tune SAM's decoder using large-scale medical datasets [22-24]. Although these methods demonstrate significant advantages across different medical modalities and tasks, they have not yet been consistently benchmarked against similar methods and lack a thorough examination of how different interaction strategies affect outcomes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.482, + 0.441 + ], + "angle": 0, + "content": "In this work, we evaluate these state-of-the-art methods on the IMed-361M dataset, providing a comprehensive and fair comparison, while also exploring the impact of different interaction strategies on medical image segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.455, + 0.444, + 0.47 + ], + "angle": 0, + "content": "3. IMIS Benchmark Dataset:IMed-361M" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.483, + 0.571 + ], + "angle": 0, + "content": "We present IMed-361M, the first large-scale IMIS dataset, that significantly surpasses existing datasets in terms of scale, diversity, quality, and density. This section details the data collection and pre-processing procedures, along with an in-depth analysis of the dataset's advantages and potential applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.583, + 0.398, + 0.599 + ], + "angle": 0, + "content": "3.1. Data Collection and Pre-processing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.483, + 0.756 + ], + "angle": 0, + "content": "Data collection. We integrate over 110 publicly available medical image segmentation datasets from globally recognized platforms such as TCIA\\(^{1}\\), OpenNeuro\\(^{2}\\), NITRC\\(^{3}\\), Grand Challenge\\(^{4}\\), Synapse\\(^{5}\\), CodaLab\\(^{6}\\), and GitHub\\(^{7}\\), covering both 2D and 3D images and variety of formats (e.g., .jpg, .npy, .nii). Additionally, we collaborate with several medical institutions, acquiring diverse clinical data through rigorous ethical review processes, which further enriches the dataset. A detailed list of data sources is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.803 + ], + "angle": 0, + "content": "Preprocessing and filtering. We first standardize all collected medical images according to the SA-Med2D-20M [42] protocol, converting the corresponding ground" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "truth (GT) into one-hot encoding and storing it in .npz files in compressed sparse row (CSR) format [55]. Then, we apply the following exclusion criteria: (1) exclude 3D slice images and their corresponding masks with an aspect ratio greater than 1.5 to avoid interference from images with excessive geometric distortion during model training; (2) exclude masks where the foreground area accounts for less than one-thousandth of the total pixels, as such small foreground regions are prone to being lost during resizing, making it challenging to generate effective prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.905, + 0.395 + ], + "angle": 0, + "content": "Resolving conflicts and ambiguities. To address conflicts and ambiguities in the GT, we first standardize expressions for the same target. For instance, both \"lung nodule\" and \"pulmonary nodule\" are renamed to \"lung nodule\" for consistency. We then manually review and correct misalignments and informational errors in the dataset to ensure annotation accuracy. Finally, for annotations with multiple connected components, we differentiate and label them based on clinical needs to avoid potential misunderstandings that could arise from single-point interactions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.905, + 0.501 + ], + "angle": 0, + "content": "Through this process, we collected 6.4 million images and 87.6 million GT masks that were manually annotated and corrected. Although this already exceeds the data volume of most existing segmentation datasets, it still does not meet our fundamental requirement for the IMIS dataset: diverse and densely masks. Therefore, we further introduce interactive masks to address the above issues." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.515, + 0.688, + 0.529 + ], + "angle": 0, + "content": "3.2. Interactive Masks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.905, + 0.808 + ], + "angle": 0, + "content": "Automatic mask generation. The automatic mask generation method employed by SAM has been proven to be high-quality and effective [17, 56, 57]. We leverage SAM's object-awareness capability to generate as many masks as possible for each image. Specifically, we employ a \\(32 \\times 32\\) point grid to guide the model, generating a set of candidate masks for each point corresponding to potential objects of interest. The generated masks are refined and optimized using the following strategies: (1) Confidence filtering: Only masks with an Intersection over Union (IoU) prediction confidence score above 0.85 are retained to ensure segmentation accuracy. (2) Non-Maximum Suppression (NMS) [58]: For overlapping masks with an IoU greater than 0.7, only the mask with the highest confidence score is kept to eliminate redundancy. (3) Remove background masks: Masks covering more than \\(80\\%\\) of the area are discarded, as such large-coverage foreground masks are almost nonexistent in the GT." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Quality control and granularity management. Based on our observations, the generated masks often fail to fully separate structures with unclear boundaries, such as the left atrium, myocardium, and right atrium in the heart. Additionally, masks for dispersed structures like the intestines are often identified as multiple separate objects. To address" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.813, + 0.321, + 0.827 + ], + "angle": 0, + "content": "1https://www.cancerimagingarchive.net" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.827, + 0.232, + 0.839 + ], + "angle": 0, + "content": "\\(^{2}\\)https://openneuro.org" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.839, + 0.232, + 0.852 + ], + "angle": 0, + "content": "3https://www.nitrc.org" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.852, + 0.26, + 0.864 + ], + "angle": 0, + "content": "4 https://grand-challenge.org" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.864, + 0.248, + 0.876 + ], + "angle": 0, + "content": "5https://www.synapse.org" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.876, + 0.216, + 0.888 + ], + "angle": 0, + "content": "\\(^{6}\\)https://codalab.org" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.214, + 0.9 + ], + "angle": 0, + "content": "7https://github.com" + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.813, + 0.321, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20843" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.096, + 0.413, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.094, + 0.862, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.306, + 0.396, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.307, + 0.636, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.31, + 0.865, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.465, + 0.908, + 0.494 + ], + "angle": 0, + "content": "Figure 1. Overview of the IMed-361M dataset. (a) Number of images and masks for each modality. (b) Information on six anatomical structures. (c) Distribution of image resolutions. (d) Analysis of mask proportions. (e) Comparison with other existing public datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.52, + 0.484, + 0.716 + ], + "angle": 0, + "content": "these issues, we utilize the original GT (described in Section 3.1) to correct the generated masks. We focus on two situations: (1) If the GT contains multi-connected regions, we directly replace the corresponding regions in the generated mask with the multi-connected regions from the GT; (2) We iterate over all single-connected regions in the GT, generating a minimum bounding box for each region. If there is a region in the generated mask with a bounding box that overlaps more than \\(95\\%\\) with this GT bounding box, we retain the generated mask region; otherwise, we replace the region with the GT mask. Finally, we apply morphological operations (e.g., erosion and dilation) to remove noise and fill small holes [59]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.484, + 0.839 + ], + "angle": 0, + "content": "We ultimately obtained 273 million \"interactive masks\", which can be used to train interactive segmentation models and cover nearly all identifiable objects in medical images. While we acknowledge that some masks may lack direct clinical relevance, they provide considerable diversity and density in the identified regions, helping the model learn a variety of interactive operations and enhancing its generalization ability across various tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.303, + 0.866 + ], + "angle": 0, + "content": "3.3. Statistics and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Data scale. As shown in Fig.1 (a), the IMed-361M dataset contains 6.4 million images, 87.6 million GT, and 273.4" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.596 + ], + "angle": 0, + "content": "million interactive masks, averaging 56 masks per image. This makes it the largest publicly available, multimodal, interactive medical image segmentation dataset to date. Compared to the MedTrinity-25M [31] dataset, our dataset provides 14.4 times the number of masks (Fig.1 (e))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.909, + 0.883 + ], + "angle": 0, + "content": "Diversity. The IMed-361M dataset covers 14 imaging modalities and 204 segmentation targets, including organs and lesions, as shown in Fig.1 (b). We categorize the GT into six groups: Head and Neck, Thorax, Skeleton, Abdomen, Pelvis, and Lesions, covering nearly all parts of the human body. For clarity, we merge certain positional information when appropriate; for example, \"left lung\" and \"right lung\" are combined into \"lung.\" Detailed category information can be found in the supplementary materials. Fig.1 (c) and (d) show the distribution of image resolutions and mask coverage within the dataset. Over \\(83\\%\\) of the images have resolutions between \\(256 \\times 256\\) and \\(1024 \\times 1024\\), ensuring broad applicability and compatibility across various research scenarios. Additionally, most masks occupy less than \\(2\\%\\) of the image area, reflecting the typically fine granularity of medical segmentation. Our interactive masks provide over one million instances across different coverage intervals, significantly enhancing the diversity and density of the dataset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Mask quality. We randomly select five images from each" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20844" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.481, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.272, + 0.454, + 0.286 + ], + "angle": 0, + "content": "Figure 2. Evaluation of the quality of interactive masks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.485, + 0.659 + ], + "angle": 0, + "content": "subset of IMed-361M (a total of 550 images) and invited four radiologists to evaluate the quality of the corresponding interactive masks. We categorize the mask quality into three levels: (1) \"Good\": the mask closely matches the manual annotations; (2) \"Satisfactory\": the mask has some visible errors (e.g., incomplete contours) but is generally acceptable; (3) \"Bad\": the mask has significant errors (e.g., over-segmentation or under-segmentation) and is unsuitable for clinical segmentation tasks. The evaluation results show that 156 images are rated as \"Good\", 345 as \"Satisfactory\", and 49 as \"Bad\". Statistical analysis reveals that the masks rated as \"Bad\" mainly come from 18 datasets, concentrated within the red-boxed area in Fig.2. Radiologists note that some excessively annotated structures hold little clinical relevance. Additionally, some masks include elements unrelated to human organs or lesions, such as letters in X-ray backgrounds or hair in dermatoscopic images. Based on this analysis, we manually remove the excessively annotated masks from these 18 datasets and apply a \\(0.5\\%\\) foreground filter rate to retain only the final valid masks. Finally, we retain those masks unrelated to human organs or lesions to enhance the model's adaptability in different scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.673, + 0.313, + 0.689 + ], + "angle": 0, + "content": "4. IMIS Baseline Network" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.234, + 0.714 + ], + "angle": 0, + "content": "4.1. Model Design" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.825 + ], + "angle": 0, + "content": "We adopt a strategy similar to fine-tuning SAM [22-24] to establish the IMIS baseline, providing a performance benchmark for future work. As shown in Fig.3, IMIS-Net has three key components: an image encoder for extracting image features, a prompt encoder for integrating user interaction information, and a mask decoder for generating segmentation results using image and prompt embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We select ViT-base [60] as the image encoder. While larger ViT models (e.g., ViT-Large, ViT-Huge) [22, 24] offer slight accuracy improvements, they significantly increase computational costs, making them impractical for clinical applications. Additionally, most open-source fine" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.925, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.238, + 0.907, + 0.268 + ], + "angle": 0, + "content": "Figure 3. The training process of IMIS-Net simulates K consecutive steps of interactive segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.908, + 0.566 + ], + "angle": 0, + "content": "tuning methods also use ViT-base, ensuring fairness in our baseline. The image input size is \\(1024 \\times 1024 \\times 3\\), with a patch size of \\(16 \\times 16 \\times 3\\). For the prompt encoder, we consider three prompt types: points, boxes, and text. Points and boxes are represented by the sum of positional encoding and learned embeddings, while text is encoded using CLIP's text encoder [61]. Text prompts follow the template: \"A segmentation area of a [category]\", covering over 200 organ and lesion categories to establish a benchmark for future multimodal segmentation research. The mask decoder uses Transformer decoder blocks, mapping image and prompt embeddings to the mask. The mask resolution is increased to \\(256 \\times 256\\) through transposed convolutions and matched to the input size via bilinear interpolation. Notably, expanding the decoder can improve performance without significantly increasing network training parameters, thereby providing a simple solution to address the issue of model performance saturation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.755, + 0.593 + ], + "angle": 0, + "content": "4.2. Experiment Configuration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Training strategy. An overview is provided of the simulated continuous interactive segmentation training [22, 23, 62] in Fig.3. For a given segmentation task and medical image \\( x_{t} \\), we first simulate a set of initial interactions \\( u_{1}^{g} \\) and \\( u_{1}^{i} \\) based on the corresponding ground truth \\( y^{g} \\) and interactive mask \\( y^{i} \\), which include clicks, bboxes, and text input. The click points are uniformly sampled from the foreground regions of \\( y^{g} \\) or \\( y^{i} \\), while the bboxes are defined as the smallest bounding box around the target, with an offset of 5 pixels added to each coordinate to simulate slight user bias during the interaction process. The entire training process involves \\( K \\) interactive training iterations (with \\( K = 8 \\) in this paper). The model's initial predictions are \\( \\hat{y}_{1}^{g} \\) and \\( \\hat{y}_{1}^{i} \\). After the first prediction, we simulate subsequent corrections based on the previous predictions \\( \\hat{y}_{k}^{g} \\) and \\( \\hat{y}_{k}^{i} \\), as well as the error region \\( \\varepsilon_{k} \\) between the \\( y^{g} \\) and \\( y^{i} \\), where \\( k\\in \\{1,\\dots,K\\} \\). Additionally, we provide the low-resolution predicted mask from the previous prediction as an extra cue to the model. As can be seen, the image encoder only needs to encode the image once during the training, and subse" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20845" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.488, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.222, + 0.484, + 0.25 + ], + "angle": 0, + "content": "Figure 4. Comparison of IMIS-Net with existing foundation models, with performance statistics at both image and mask levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.279, + 0.484, + 0.309 + ], + "angle": 0, + "content": "quent interactive training only updates the prompt encoder and mask decoder parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.536 + ], + "angle": 0, + "content": "Training setting. We use the Adam optimizer [63] with a learning rate of \\(2 \\times 10^{-5}\\). The training was conducted on 72 NVIDIA 4090 GPUs with a batch size of 2. For each image, we randomly select 5 targets from the corresponding GT and interactive masks as supervision targets (if fewer than 5 targets are present, they are selected repeatedly). The image resolution was uniformly resized to \\(1024 \\times 1024\\), and pixel intensities are randomly scaled and shifted with a probability of \\(20\\%\\), where the scaling factor and offset are set to 0.2. We train the model for a total of 12 epochs on the IMed-361M dataset and select the final checkpoint as the model weights. A linear combination of Focal loss [64] and Dice loss [65] is used as the loss function, balancing their influence in a 20:1 ratio. Finally, the Dice score is used as the primary evaluation metric for this study." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.549, + 0.34, + 0.563 + ], + "angle": 0, + "content": "4.3. Evaluation Model and Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.483, + 0.737 + ], + "angle": 0, + "content": "Model. We evaluate five visual foundation models [17, 18, 22-24], selected based on their pretraining on large-scale datasets and the availability of source code. Among them, MedSAM, SAM-Med2D, and Huang et al. [24] are specifically designed for IMIS, while SAM [17] and SAM-2 [18] are pre-trained on massive natural image and video datasets. To ensure fairness in the evaluation, all models used ViT-Base as the encoder. It is important to note that, MedSAM and Huang et al. [24] only supportbbox-based interactivity (using the models released by the authors), while the other models support points, bboxes, and combinations of both." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.739, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Data. The data used to evaluate the model is divided into two categories: the test set and the external dataset. The test set contains 58,411 images, with a distribution consistent with that of IMed-361M. The external dataset includes data from three different modalities and clinical scenarios: SegThor [66], which is used for segmentation of thoracic organs at risk in 516 CT images; TotalSegmentator MRI [37], which is used for segmenting organs in 2,147 MRI images; and ISLES \\(^{8}\\), which belongs to the MRI FLAIR modality" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "and includes 364 images of ischemic stroke lesions. These datasets were not used during the training of IMIS-Net." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.138, + 0.602, + 0.154 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.163, + 0.655, + 0.179 + ], + "angle": 0, + "content": "5.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.186, + 0.905, + 0.502 + ], + "angle": 0, + "content": "Fig.4 shows the performance evaluation of IMIS-Net compared to other vision foundation models on the single-interaction segmentation task. During testing, all models used the same points and bounding boxes as prompts. The results indicate that IMIS-Net outperforms other models in both image and mask-level statistics. Thebbox interaction of different models always outperforms the click interaction because the_bbox can provide more boundary information. Notably, despite being pretrained on large-scale medical image datasets, MedSAM and SAM-Med2D still exhibit significant performance differences. This disparity is closely related to the scale and diversity of the pretraining datasets. As shown in Fig.5(a) and (b), the pre-training dataset for SAM-Med2D lacks samples of skeletal structures, resulting in poorer segmentation performance for these anatomical structures. Additionally, under the single-point prompt condition, SAM and SAM-2 achieve only Dice scores of \\(60.26\\%\\) and \\(59.57\\%\\), respectively. likely due to the absence of medical knowledge in the pretraining data and limited interactive information constraining model performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.505, + 0.905, + 0.775 + ], + "angle": 0, + "content": "We increase the number of interactions from 1 to 9 to observe the performance changes of the model (Fig.5 (c)). As expected, performance improves with more interactions, and the gap between models narrows. This is because multiple interactions reduce the task difficulty. We also examine the impact of click position andbbox offset on performance. As shown in Fig.5 (d) and (f), when the prompt point is closer to the centroid, the performance improvement is more significant, with SAM-2's Dice score increasing by \\(2.84\\%\\). Additionally, withbbox offset, all methods experience a performance decline of \\(0.85\\% - 3.94\\%\\). Our model exhibited the smallest performance drop, making it more robust for practical use. In summary, leveraging the rich data and mask diversity of the IMed-361M dataset, our model delivers the best performance across various medical scenarios and interaction strategies. More experimental results and analysis can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.768, + 0.803 + ], + "angle": 0, + "content": "5.2. External Dataset Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Tab.1 presents evaluation results on external datasets. IMIS-Net achieves the best average performance across datasets from three different sources and tasks. We highlight the Dice scores of the 12 major abdominal organs in the TotalSegmentator MRI dataset, where IMIS-Net outperforms in 10. On the ISLES dataset, focusing on ischemic stroke" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.29, + 0.902 + ], + "angle": 0, + "content": "\\(^{8}\\)https://www.isles-challenge.org/" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20846" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.167, + 0.089, + 0.416, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.253, + 0.361, + 0.265 + ], + "angle": 0, + "content": "(a) Single click interaction" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.089, + 0.677, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.253, + 0.625, + 0.265 + ], + "angle": 0, + "content": "(b) Bounding box interaction" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.18, + 0.854, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.269, + 0.374, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.411, + 0.328, + 0.421 + ], + "angle": 0, + "content": "(c) Number of interactions" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.268, + 0.616, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.411, + 0.617, + 0.422 + ], + "angle": 0, + "content": "(d) Sampling points distances from the centroid" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.269, + 0.859, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.411, + 0.779, + 0.421 + ], + "angle": 0, + "content": "(e) Bbox offset" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.437, + 0.908, + 0.48 + ], + "angle": 0, + "content": "Figure 5. Comparison of segmentation performance across different anatomical structures under single-click (a) and bounding box (b) interactions. (c) Changes in segmentation performance with increasing interaction numbers. (d) and (e) Impact of click position and bounding box offset on performance." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.492, + 0.865, + 0.778 + ], + "angle": 0, + "content": "
DatasetCategorySAMSAM 2MedSAMHuang et al.SAM-Med2DIMIS-Net
ISLESIschemic Stroke Lesion55.9260.1459.9056.7968.2271.78
SegThorEsophagus78.1986.6047.1776.4882.0489.17
Heart91.4192.3481.1692.0294.1881.27
Aorta82.3683.6955.6884.5779.1294.92
Trachea93.7694.2472.4094.6085.1090.04
Average84.4685.8660.5584.7386.4389.27
Totalsegmentator MRIAdrenal gland77.1982.2556.5868.5266.3577.53
Aorta76.8879.9055.9880.0874.9185.62
Colon56.6956.0042.5355.4151.4473.47
Duodenum69.3074.8155.6472.3869.6775.97
Gallbladder81.3384.8268.1583.4687.6382.93
Iliopsoas59.5771.9658.4863.6760.7672.55
Kidney79.4381.5958.4181.0974.0482.72
Inferior vena cava84.9487.0357.7385.8878.3487.35
Liver86.9388.7068.1487.3586.7590.62
Pancreas66.9569.8646.1066.3066.8470.84
Spleen81.1283.0566.6183.8784.9586.74
Stomach77.1081.2362.7576.5279.3181.45
Average75.4577.6259.5275.4775.9279.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.109, + 0.791, + 0.887, + 0.807 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison results of IMIS-Net against five other interactive segmentation methods on three external datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.832, + 0.486, + 0.879 + ], + "angle": 0, + "content": "lesion segmentation, our model surpasses the second-best SAM-Med2D by \\(3.56\\%\\). These results demonstrate the strong generalization ability of our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.831, + 0.667, + 0.848 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Scaling Up Training Data. As shown in Fig.6 (a) and (b), when trained only on the GT from IMed-361M, IMIS-Net performs poorly. However, with the addition of interac" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20847" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.092, + 0.353, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.092, + 0.571, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.092, + 0.882, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.263, + 0.908, + 0.293 + ], + "angle": 0, + "content": "Figure 6. (a) and (b) segmentation performance increases with the increasing mask density or data scales up. (c) effectiveness of text and its combination with other prompts." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.315, + 0.489, + 0.457 + ], + "angle": 0, + "content": "
Decoder dimensionImage resolutionPromptTraining parameters
PointBbox
768256×2560.82140.846929.68 M
768512×5120.86730.896829.68 M
2561024×10240.83660.84975.52 M
5121024×10240.85630.872915.19 M
7681024×10240.88480.906029.68 M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.47, + 0.484, + 0.499 + ], + "angle": 0, + "content": "Table 2. Ablation study of model design, including decoder dimension and image resolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.594 + ], + "angle": 0, + "content": "tive masks, the Dice score increases rapidly. Similar performance trends are observed when training on datasets of different sizes, indicating that our method is scalable and performs better with more available training data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.748 + ], + "angle": 0, + "content": "Text and Combination Prompts. IMIS-Net, as a versatile interactive model, can utilize text prompts to achieve the segmentation of over 200 organs and lesions. As shown in Fig.6 (c), the model achieves a segmentation performance of \\(76.30\\%\\) when using only text prompts. When both text and point prompts are combined, the average Dice score increases by \\(11.95\\%\\). Furthermore, after three rounds of click-based correction, the Dice score reaches \\(89.69\\%\\). These results show that combining different prompts synergistically enhances the model's segmentation capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Model Design. We conduct an ablation study on different configurations of the model, including the impact of input resolution and decoder dimension on model performance. As shown in Tab.2, the model's segmentation performance improves as the training image resolution increases. This higher input resolution allows for clearer visualization of lesions and organs in medical images. Additionally, when the decoder dimension is increased from 256 to 768, the model's performance improves from \\(84.97\\%\\) to \\(90.60\\%\\), with only a 24.16M increase in trainable parameters. These" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.38 + ], + "angle": 0, + "content": "results demonstrate the scalability of IMIS-Net, showing that even when the model's performance approaches saturation on larger datasets, further performance improvements can still be achieved by expanding the decoder." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.395, + 0.634, + 0.412 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.907, + 0.679 + ], + "angle": 0, + "content": "In this work, we introduce IMed-361M, a benchmark dataset dedicated to interactive medical image segmentation. It includes a vast array of medical images across various modalities, extensive segmentation scenarios, and densely masks, surpassing all existing datasets that are limited to single tasks or simple integrations. Leveraging this data resource, we developed a general IMIS baseline model that enables users to generate segmentation results tailored to clinical needs through interactive methods, including clicks, bounding boxes, text prompts, and their combinations. We performed a comprehensive comparison of this baseline with existing foundational models, demonstrating that our model provides significant performance advantages and exhibits strong transferability in previously unseen scenarios. Notably, our approach typically requires fewer interactions to achieve comparable performance, enhancing its practicality in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.679, + 0.909, + 0.845 + ], + "angle": 0, + "content": "The IMed-361M dataset will strongly facilitate the development of foundational models in medical imaging and lay the foundation for fair evaluation across different models. IMIS-Net provides general technical support for various clinical applications, accelerating the widespread application of AI technology in the medical field. Despite these achievements, we recognize that this work still faces several challenges. For instance, effectively obtaining semantic information for interactive masks and extending this approach to more comprehensive and finer-grained medical image analysis scenarios are areas that require further exploration and improvement in the future." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.845, + 0.908, + 0.888 + ], + "angle": 0, + "content": "Acknowledgements: The authors acknowledge the funding provided by the National Key Research and Development Program of China (No.2022ZD0160102)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20848" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Tianfei Zhou, Lielei Li, Gustav Bredell, Jianwu Li, Jan Unkelbach, and Ender Konukoglu. Volumetric memory network for interactive medical image segmentation. Medical Image Analysis, 83:102599, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Jian Zhang, Yinghuan Shi, Jinquan Sun, Lei Wang, Luping Zhou, Yang Gao, and Dinggang Shen. Interactive medical image segmentation via a point-based interaction. Artificial Intelligence in Medicine, 111:101998, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Fang Chen, Haojie Han, Peng Wan, Hongen Liao, Chunrui Liu, and Daoqiang Zhang. Joint segmentation and differential diagnosis of thyroid nodule in contrast-enhanced ultrasound images. IEEE Transactions on Biomedical Engineering, 70(9):2722-2732, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[4] Zdravko Marinov, Paul F Jäger, Jan Egger, Jens Kleesiek, and Rainer Stiefelhagen. Deep interactive segmentation of medical images: A systematic review and taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[5] Pierre-Henri Conze, Gustavo Andrade-Miranda, Vivek Kumar Singh, Vincent Jaouen, and Dimitris Visvikis. Current and emerging trends in medical image segmentation with deep learning. IEEE Transactions on Radiation and Plasma Medical Sciences, 7(6):545-569, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[6] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[7] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[9] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, pages 3-11. Springer, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[10] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Junlong Cheng, Shengwei Tian, Long Yu, Chengrui Gao, Xiaojing Kang, Xiang Ma, Weidong Wu, Shijia Liu, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "Hongchun Lu. Resganet: Residual group attention network for medical image classification and segmentation. Medical Image Analysis, 76:102313, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[12] Marco Cascella, Jonathan Montomoli, Valentina Bellini, and Elena Bignami. Evaluating the feasibility of chatgpt in healthcare: an analysis of multiple clinical and research scenarios. Journal of medical systems, 47(1):33, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.906, + 0.287 + ], + "angle": 0, + "content": "[13] Mingze Yuan, Yingda Xia, Hexin Dong, Zifan Chen, Jiawen Yao, Mingyan Qiu, Ke Yan, Xiaoli Yin, Yu Shi, Xin Chen, et al. Devil is in the queries: advancing mask transformers for real-world medical image segmentation and out-of-distribution localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23879-23889, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.288, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[14] Jakob Wasserthal, Hanns-Christian Breit, Manfred T Meyer, Maurice Pradella, Daniel Hinck, Alexander W Sauter, Tobias Heye, Daniel T Boll, Joshy Cyriac, Shan Yang, et al. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence, 5(5), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.371, + 0.906, + 0.453 + ], + "angle": 0, + "content": "[15] Sergios Gatidis, Tobias Hepp, Marcel Fruh, Christian La Fougère, Konstantin Nikolaou, Christina Pfaffenberg, Bernhard Schölkopf, Thomas Küstner, Clemens Cyran, and Daniel Rubin. A whole-body fdg-pet/ct dataset with manually annotated tumor lesions. Scientific Data, 9(1):601, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.455, + 0.906, + 0.524 + ], + "angle": 0, + "content": "[16] Simon Graham, Mostafa Jahanifar, Quoc Dang Vu, Giorgos Hadjigeorghiou, Thomas Leech, David Snead, Shan E Ahmed Raza, Fayyaz Minhas, and Nasir Rajpoot. Conic: Colon nuclei identification and counting challenge 2022. arXiv preprint arXiv:2111.14485, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.525, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.906, + 0.677 + ], + "angle": 0, + "content": "[18] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.678, + 0.906, + 0.733 + ], + "angle": 0, + "content": "[19] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.734, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[20] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, et al. Segment anything in 3d with nerfs. Advances in Neural Information Processing Systems, 36:25971-25990, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[21] Di Wang, Jing Zhang, Bo Du, Minqiang Xu, Lin Liu, Dacheng Tao, and Liangpei Zhang. Samrs: Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[22] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20849" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[23] Junlong Cheng, Jin Ye, Zhongying Deng, Jianpin Chen, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Ji-long Chen, Lei Jiang, et al. Sam-med2d. arXiv preprint arXiv:2308.16184, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.216 + ], + "angle": 0, + "content": "[24] Yuhao Huang, Xin Yang, Lian Liu, Han Zhou, Ao Chang, Xinrui Zhou, Rusi Chen, Junxuan Yu, Jiongquan Chen, Chaoyu Chen, et al. Segment anything model for medical images? Medical Image Analysis, 92:103061, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[25] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.275, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[26] Yichi Zhang, Zhenrong Shen, and Rushi Jiao. Segment anything model for medical image segmentation: Current applications and future directions. Computers in Biology and Medicine, page 108238, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[27] Shaoting Zhang and Dimitris Metaxas. On the challenges and perspectives of foundation models for medical image analysis. Medical image analysis, 91:102996, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.373, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[28] Kummaragunta Joel Prabhod and Asha Gadhiraju. Foundation models in medical imaging: Revolutionizing diagnostic accuracy and efficiency. Journal of Artificial Intelligence Research and Applications, 4(1):471-511, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[29] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[30] Raphael Schäfer, Till Nicke, Henning Hofener, AnnKristin Lange, Dorit Merhof, Friedrich Feuerhake, Volkmar Schulz, Johannes Lotz, and Fabian Kiessling. Overcoming data scarcity in biomedical imaging with a foundational multitask model. Nature Computational Science, 4(7):495-509, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[31] Yunfei Xie, Ce Zhou, Lang Gao, Juncheng Wu, Xianhang Li, Hong-Yu Zhou, Sheng Liu, Lei Xing, James Zou, Cihang Xie, et al. Medtrinity-25m: A large-scale multimodal dataset with multigranular annotations for medicine. arXiv preprint arXiv:2408.02900, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[32] Kenneth Clark, Bruce Vendt, Kirk Smith, John Freymann, Justin Kirby, Paul Koppel, Stephen Moore, Stanley Phillips, David Maffitt, Michael Pringle, et al. The cancer imaging archive (tcia): maintaining and operating a public information repository. Journal of digital imaging, 26:1045-1057, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[33] Joao Pedrosa, Guilherme Aresta, Carlos Ferreira, Gurraj Atwal, Hady Ahmady Phoulady, Xiaoyu Chen, Rongzhen Chen, Jiaoliang Li, Liansheng Wang, Adrian Galdran, et al. Lndb challenge on automatic lung cancer patient management. Medical image analysis, 70:102027, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[34] Rina D Rudyanto, Sjoerd Kerkstra, Eva M Van Rikxoort, Catalin Fetita, Pierre-Yves Brillet, Christophe Lefevre, Wenzhe Xue, Xiangjun Zhu, Jianming Liang, Ilkay Oksuz, et al. Comparing algorithms for automated vessel segmentation in computed tomography scans of the lung: the vessel12 study. Medical image analysis, 18(7):1217-1232, 2014." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[35] Michela Antonelli, Annika Reinke, Spyridon Bakas, Keyvan Farahani, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, Olaf Ronneberger, Ronald M Summers, et al. The medical segmentation decathlon. Nature communications, 13(1):4128, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[36] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia modeling: 26th international conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, proceedings, part II 26, pages 451–462. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[37] Tugba Akinci D'Antonoli, Lucas K Berger, Ashraya K Indrakanti, Nathan Vishwanathan, Jakob Weiβ, Matthias Jung, Zeynep Berkarda, Alexander Rau, Marco Reisert, Thomas Küstner, et al. Totalsegmentator mr: Sequence-independent segmentation of 59 anatomical structures in mr images. arXiv preprint arXiv:2405.19492, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[38] Igor Zingman, Birgit Stierstorfer, Charlotte Lempp, and Fabian Heinemann. Learning image representations for anomaly detection: application to discovery of histological alterations in drug development. Medical Image Analysis, 92:103067, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[39] Gašper Podobnik, Primož Strojan, Primož Peterlin, Bulat Ibragimov, and Tomaž Vrtovec. Han-seg: The head and neck organ-at-risk ct and mr segmentation dataset. Medical physics, 50(3):1917-1927, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[40] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[41] Ziyan Huang, Haoyu Wang, Zhongying Deng, Jin Ye, Yanzhou Su, Hui Sun, Junjun He, Yun Gu, Lixu Gu, Shaoting Zhang, et al. Stu-net: Scalable and transferable medical image segmentation models empowered by large-scale supervised pre-training. arXiv preprint arXiv:2304.06716, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[42] Jin Ye, Junlong Cheng, Jianpin Chen, Zhongying Deng, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Jilong Chen, Lei Jiang, et al. Sa-med2d-20m dataset: Segment anything in 2d medical imaging with 20 million masks. arXiv preprint arXiv:2311.11969, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[43] Sílvia Delgado Olabarriaga and Arnold WM Smeulders. Interaction in the segmentation of medical images: A survey. Medical image analysis, 5(2):127-142, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[44] Guotai Wang, Maria A Zuluaga, Rosalind Pratt, Michael Aertsen, Tom Doel, Maria Klusmann, Anna L David, Jan Deprest, Tom Vercauteren, and Sébastien Ourselin. Slic-seg: A minimally interactive segmentation of the placenta from sparse and motion-corrupted fetal mri in multiple views. Medical image analysis, 34:137-147, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[45] Hallee E Wong, Marianne Rakic, John Guttag, and Adrian V Dalca. Scribbleprompt: Fast and flexible interactive segmentation for any medical image. arXiv preprint arXiv:2312.07381, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[46] Martin Rajchl, Matthew CH Lee, Ozan Oktay, Konstantinos Kamnitsas, Jonathan Passerat-Palmbach, Wenjia Bai, Mellisa Damodaram, Mary A Rutherford, Joseph V Hajnal," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20850" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Bernhard Kainz, et al. Deepcut: Object segmentation from bounding box annotations using convolutional neural networks. IEEE transactions on medical imaging, 36(2):674-683, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.231 + ], + "angle": 0, + "content": "[47] Tomas Sakinis, Fausto Milletari, Holger Roth, Panagiotis Korfiatis, Petro Kostandy, Kenneth Philbrick, Zeynett Akkus, Ziyue Xu, Daguang Xu, and Bradley J Erickson. Interactive segmentation of medical images through fully convolutional neural networks. arXiv preprint arXiv:1903.08205, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[48] Yuxin Du, Fan Bai, Tiejun Huang, and Bo Zhao. Segvol: Universal and interactive volumetric medical image segmentation. arXiv preprint arXiv:2311.13385, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.484, + 0.33 + ], + "angle": 0, + "content": "[49] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086-7096, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.427 + ], + "angle": 0, + "content": "[50] Leo Grady, Thomas Schiwietz, Shmuel Aharon, and Rüdiger Westermann. Random walks for interactive organ segmentation in two and three dimensions: Implementation and validation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2005: 8th International Conference, Palm Springs, CA, USA, October 26-29, 2005, Proceedings, Part II 8, pages 773-780. Springer, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[51] Mingzhe Hu, Yuheng Li, and Xiaofeng Yang. Skinsam: adapting the segmentation anything model for skin cancer segmentation. In Medical Imaging 2024: Image Perception, Observer Performance, and Technology Assessment, volume 12929, pages 174-184. SPIE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[52] Yuheng Li, Mingzhe Hu, and Xiaofeng Yang. Polyp-sam: Transfer sam for polyp segmentation. In Medical Imaging 2024: Computer-Aided Diagnosis, volume 12927, pages 759-765. SPIE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.608 + ], + "angle": 0, + "content": "[53] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[54] Jay N Paranjape, Nithin Gopalakrishnan Nair, Shameema Sikder, S Swaroop Vedula, and Vishal M Patel. Adaptivesam: Towards efficient tuning of sam for surgical scene segmentation. In Annual Conference on Medical Image Understanding and Analysis, pages 187-201. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[55] Urban Borstnik, Joost VandeVondele, Valery Weber, and Jürg Hutter. Sparse matrix multiplication: The distributed block-compressed sparse row library. Parallel Computing, 40(5-6):47-58, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[56] Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[57] Yizhe Zhang, Tao Zhou, Shuo Wang, Peixian Liang, Yejia Zhang, and Danny Z Chen. Input augmentation with sam: Boosting medical image segmentation with segmentation foundation model. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 129-139. Springer, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[58] Charalampos Symeonidis, Ioannis Mademlis, Ioannis Pitas, and Nikos Nikolaidis. Neural attention-driven nonmaximum suppression for person detection. IEEE transactions on image processing, 32:2454-2467, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[59] Suresh Kumar Tummala et al. Morphological operations and histogram analysis of sem images using python. Indian Journal of Engineering and Materials Sciences (IJEMS), 29(6):796-800, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[60] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[61] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[62] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[63] Zijun Zhang. Improved adam optimizer for deep neural networks. In 2018 IEEE/ACM 26th international symposium on quality of service (IWQoS), pages 1-2. IEEE, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[64] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[65] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.559, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[66] Zoé Lambert, Caroline Petitjean, Bernard Dubray, and Su Kuan. Segthor: Segmentation of thoracic organs at risk in ct images. In 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA), pages 1- 6. IEEE, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20851" + } + ] +] \ No newline at end of file diff --git a/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/72ec2f30-d729-4ccc-bc2d-90a1d1bd3d41_origin.pdf b/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/72ec2f30-d729-4ccc-bc2d-90a1d1bd3d41_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..84ddd9947d696298658d24a2b5feb2e862103b2d --- /dev/null +++ b/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/72ec2f30-d729-4ccc-bc2d-90a1d1bd3d41_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d2d64d8557373091b60ba8284c2ae69874ea6ce472d74c77b3d107d39e5238f +size 3390215 diff --git a/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/full.md b/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e005915565fc207f2bc404560ac01087107869dc --- /dev/null +++ b/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/full.md @@ -0,0 +1,274 @@ +# Interactive Medical Image Segmentation: A Benchmark Dataset and Baseline + +Junlong Cheng $^{1,2}$ Bin Fu $^{2,4}$ Jin Ye $^{2}$ Guoan Wang $^{2}$ Tianbin Li $^{2}$ Haoyu Wang $^{2}$ Ruoyu Li $^{1}$ He Yao $^{1}$ Junren Cheng $^{1}$ Jinwen Li $^{3}$ Yanzhou Su $^{2}$ Min Zhu $^{1*}$ Junjun He $^{2\dagger}$ + +1College of Computer Science, Sichuan University + +2Shanghai AI Laboratory, China + +$^{3}$ School of Computer Science and Technology, Xinjiang University + +4 Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences + +{chengjunlong, fubin, yejin, wangguoan, litianbin, wanghaoyu} @pjlab.org.cn, + +{ruoyuli, 2024323040035, chengjunren} $@$ stu.scu.edu.cn, 107556522204@stu.xju.edu.cn, + +yanzhou.su@outlook.com, hejunjun.su@pjlab.org.cn, zhumin@scu.edu.cn + +# Abstract + +Interactive Medical Image Segmentation (IMIS) has long been constrained by the limited availability of large-scale, diverse, and densely annotated datasets, which hinders model generalization and consistent evaluation across different models. In this paper, we introduce the IMed-361M benchmark dataset, a significant advancement in general IMIS research. First, we collect and standardize over 6.4 million medical images and their corresponding ground truth masks from multiple data sources. Then, leveraging the strong object recognition capabilities of a vision foundational model, we automatically generated dense interactive masks for each image and ensured their quality through rigorous quality control and granularity management. Unlike previous datasets, which are limited by specific modalities or sparse annotations, IMed-361M spans 14 modalities and 204 segmentation targets, totaling 361 million masks—an average of 56 masks per image. Finally, we developed an IMIS baseline network on this dataset that supports high-quality mask generation through interactive inputs, including clicks, bounding boxes, text prompts, and their combinations. We evaluate its performance on medical image segmentation tasks from multiple perspectives, demonstrating superior accuracy and scalability compared to existing interactive segmentation models. To facilitate research on foundational models in medical computer vision, we release the IMed-361M and model at https://github.com.uni-medical/IMIS-Bench. + +# 1. Introduction + +Interactive Medical Image Segmentation (IMIS) enables clinicians or users to guide the model by marking points, lines, or regions on an image, resulting in segmentation outcomes that better meet clinical needs [1, 2]. This user-informed segmentation approach not only optimizes results throughout the segmentation process, aiding clinicians in precisely localizing targets during diagnosis and treatment [3-5], but also addresses the limitations of fully automated segmentation models in generalizing to unseen object classes (i.e., zero-shot learning) [6-11], providing clinicians with a flexible adjustment mechanism that significantly enhances segmentation reliability. However, the advancement of IMIS advancement faces a critical bottleneck: the lack of high-quality, diverse, and large-scale datasets that capture the complex clinical environments [12, 13]. Existing public datasets (such as Totalsegmentator [14], AutoPET [15], and CoNIC [16]) are often tailored to specific modalities or tasks, thus constraining the generalization capabilities of models across diverse medical scenarios. + +Recently, as a foundational model for general visual segmentation, the Segment Anything Model (SAM) has garnered widespread attention [17, 18]. SAM integrates simple user interactions (such as points or bounding boxes) into the model's learning process and leverages pre-training on large-scale datasets to achieve cross-domain and multitask transferability [19-21]. Although SAM has been pretrained on over 1 billion natural image masks, its direct application in medical imaging remains limited due to the substantial domain differences between natural and medical images, as well as the lack of medical imaging-specific knowledge. To address these issues, current methods such as MedSAM [22] and SAM-Med2D [23] attempt to expand dataset size and fine-tune SAM to create IMIS models tailored for clinical scenarios. While these approaches have + +shown some progress, their ability to 'segment anything' in medical imaging remains limited. This limitation is mainly due to the lack of densely masks in existing medical datasets compared to the SA-1B dataset [17] used to train SAM [22-24]. For instance, the COSMOS dataset [24] contains an average of only 5.7 masks per image, which restricting the model's capacity for dense segmentation and hindering comprehensive, fine-grained interaction. Moreover, many methods are evaluated only on specific modalities or with limited interaction strategies, further restricting the comprehensiveness and reliability of IMIS model evaluation [22, 24-26]. + +Addressing these limitations requires the development of a high-quality IMIS benchmark dataset, which is essential for advancing foundational models in medical imaging [4, 27-30]. An ideal IMIS benchmark dataset should meet three core criteria: (1) Large-scale. The dataset should be large enough to fully support deep learning model training, enabling the model to effectively capture medical features; (2) Diversity. The dataset should encompass various medical imaging modalities and complex clinical scenarios to ensure the model's ability to generalize across modalities and tasks; (3) High-quality and densely masks. Accurate segmentation of complex medical images relies on high-quality annotations, and densely masks further enhance the model's capability to "segment anything" in medical imaging. Unfortunately, existing public medical segmentation datasets [22-24, 31] do not fully meet these standards, limiting their ability to comprehensively support and evaluate IMIS models. + +In this work, we introduce IMed-361M, a benchmark dataset specifically designed for IMIS tasks, this dataset integrates public and private data sources and utilizes foundational models [17] for automated annotation to generating dense masks for each image. A standardized data processing workflow is ensure the high quality and consistency of all masks. IMed-361M achieves unprecedented scale, diversity, and mask quality, comprising 6.4 million images spanning 14 imaging modalities and 204 targets, with a total of 361 million masks, averaging 56 masks per image. IMed-361M effectively solves the problems of small data size and sparse annotations, providing data support for IMIS model training. Additionally, we develop an IMIS baseline model and conduct a comprehensive evaluation of its performance across various medical scenarios, including its effectiveness on different modalities, anatomical structures, and organs, as well as the impact of various interaction strategies on model outcomes. This analysis provided an in-depth understanding of the strengths and limitations of different interactive segmentation methods, establishing a fair and consistent framework for evaluating IMIS model performance. We anticipate that the IMed-361M dataset and baseline model will drive the widespread adoption of IMIS + +technology in clinical practice, accelerating the healthcare industry's transition toward intelligence and automation. + +# 2. Related Work + +This section provides an overview of datasets used for various medical image segmentation tasks and reviews the advancements in IMIS algorithms. + +Datasets. Due to variations in imaging protocols, medical images often have multidimensional characteristics (e.g., 2D, 3D), making pixel-level segmentation labeling for specific organs or lesions both expertise-dependent and time-consuming. Early work focused on single-organ or single-lesion annotated datasets [14, 15, 15, 32–36]. These single-task datasets provide foundational data for training automatic segmentation models, significantly enhancing model performance in segmenting specific organs or lesions. With the growth of multi-task requirements, multi-organ and multi-tissue annotated datasets [14, 37–39] have emerged, facilitating the exploration of deep learning models in multi-region segmentation tasks [40, 41]. In recent years, visual foundational models have demonstrated impressive cross-domain and multi-task transfer capabilities, motivating researchers to build large-scale, multimodal datasets for medical foundational models. The most common approach is to integrate data from various sources [22–24, 31] to bridge the scale gap compared to natural image datasets. However, due to the sparsity of annotations in the source data, the combined dataset still lags behind natural image datasets in terms of mask quantity and density. For example, the SA-1B [17] dataset averages 100 masks per image, whereas large-scale medical datasets like COSMOS [24] and SA-Med2D-20M [42] have fewer than 5 masks per image on average. This gap limits the ability of medical datasets to support dense segmentation tasks and fine-grained interactions, as well as the applicability of foundational models to "segment anything" in the medical imaging domain. + +Algorithms. IMIS methods are generally more effective than automated methods at generating high-quality results that meet clinical requirements, especially in scenarios involving diverse imaging protocols, complex pathological variations, and ambiguous lesion boundaries [43]. User interactions can be take several typical forms, such as scribbles [44, 45], bounding boxes [17, 46], clicks [2, 47], or language prompts [17, 48, 49]. Traditional methods approach this task through energy by minimizing energy on a regular pixel grid, capturing low-level appearance features with unary potentials and promoting consistent segmentation outputs with pairwise or higher-order potentials [44, 50]. With advancements in deep learning, researchers have explored using user prompts directly as network input features, often in combination with 2D or 3D neural networks to produce segmentation results [1, 2, 17, 44-47]. Lever + +aging the advantage of pretraining on large-scale datasets, the Segment Anything Model (SAM) has become a benchmark for IMIS. For example, SAM can serve as a powerful pretrained encoder-decoder model fine-tuned for specific tasks [51, 52]. However, these approaches essentially revert to the design of fully automated segmentation models. Another category of methods retains the interactive aspect of segmentation, focusing on parameter-efficient fine-tuning (PEFT) techniques [53, 54]. However, these methods consider only a limited of interaction strategies, and inconsistencies in evaluation have affected the comparability and reliability of their results. To achieve robust segmentation capabilities across diverse medical imaging datasets, some approaches fine-tune SAM's decoder using large-scale medical datasets [22-24]. Although these methods demonstrate significant advantages across different medical modalities and tasks, they have not yet been consistently benchmarked against similar methods and lack a thorough examination of how different interaction strategies affect outcomes. + +In this work, we evaluate these state-of-the-art methods on the IMed-361M dataset, providing a comprehensive and fair comparison, while also exploring the impact of different interaction strategies on medical image segmentation. + +# 3. IMIS Benchmark Dataset:IMed-361M + +We present IMed-361M, the first large-scale IMIS dataset, that significantly surpasses existing datasets in terms of scale, diversity, quality, and density. This section details the data collection and pre-processing procedures, along with an in-depth analysis of the dataset's advantages and potential applications. + +# 3.1. Data Collection and Pre-processing + +Data collection. We integrate over 110 publicly available medical image segmentation datasets from globally recognized platforms such as TCIA $^{1}$ , OpenNeuro $^{2}$ , NITRC $^{3}$ , Grand Challenge $^{4}$ , Synapse $^{5}$ , CodaLab $^{6}$ , and GitHub $^{7}$ , covering both 2D and 3D images and variety of formats (e.g., .jpg, .npy, .nii). Additionally, we collaborate with several medical institutions, acquiring diverse clinical data through rigorous ethical review processes, which further enriches the dataset. A detailed list of data sources is provided in the supplementary material. + +Preprocessing and filtering. We first standardize all collected medical images according to the SA-Med2D-20M [42] protocol, converting the corresponding ground + +truth (GT) into one-hot encoding and storing it in .npz files in compressed sparse row (CSR) format [55]. Then, we apply the following exclusion criteria: (1) exclude 3D slice images and their corresponding masks with an aspect ratio greater than 1.5 to avoid interference from images with excessive geometric distortion during model training; (2) exclude masks where the foreground area accounts for less than one-thousandth of the total pixels, as such small foreground regions are prone to being lost during resizing, making it challenging to generate effective prompts. + +Resolving conflicts and ambiguities. To address conflicts and ambiguities in the GT, we first standardize expressions for the same target. For instance, both "lung nodule" and "pulmonary nodule" are renamed to "lung nodule" for consistency. We then manually review and correct misalignments and informational errors in the dataset to ensure annotation accuracy. Finally, for annotations with multiple connected components, we differentiate and label them based on clinical needs to avoid potential misunderstandings that could arise from single-point interactions. + +Through this process, we collected 6.4 million images and 87.6 million GT masks that were manually annotated and corrected. Although this already exceeds the data volume of most existing segmentation datasets, it still does not meet our fundamental requirement for the IMIS dataset: diverse and densely masks. Therefore, we further introduce interactive masks to address the above issues. + +# 3.2. Interactive Masks + +Automatic mask generation. The automatic mask generation method employed by SAM has been proven to be high-quality and effective [17, 56, 57]. We leverage SAM's object-awareness capability to generate as many masks as possible for each image. Specifically, we employ a $32 \times 32$ point grid to guide the model, generating a set of candidate masks for each point corresponding to potential objects of interest. The generated masks are refined and optimized using the following strategies: (1) Confidence filtering: Only masks with an Intersection over Union (IoU) prediction confidence score above 0.85 are retained to ensure segmentation accuracy. (2) Non-Maximum Suppression (NMS) [58]: For overlapping masks with an IoU greater than 0.7, only the mask with the highest confidence score is kept to eliminate redundancy. (3) Remove background masks: Masks covering more than $80\%$ of the area are discarded, as such large-coverage foreground masks are almost nonexistent in the GT. + +Quality control and granularity management. Based on our observations, the generated masks often fail to fully separate structures with unclear boundaries, such as the left atrium, myocardium, and right atrium in the heart. Additionally, masks for dispersed structures like the intestines are often identified as multiple separate objects. To address + +![](images/ceb09083a3817149f6129ec884ea89bba6c5a7ba1e1c95f0344512db33ca1c26.jpg) + +![](images/987aefc771c31c81122029accb4bf1b8ddc73ae69a2e9c4914bbc5176d60acc0.jpg) + +![](images/a97f21ed46ca3e9ca40ba1cd83a9daa0682aec774842add361bcc23ef039bbd6.jpg) +Figure 1. Overview of the IMed-361M dataset. (a) Number of images and masks for each modality. (b) Information on six anatomical structures. (c) Distribution of image resolutions. (d) Analysis of mask proportions. (e) Comparison with other existing public datasets. + +![](images/cb91498ad17b1295f0c833c03ecf24b99e217d2b3e99ea84c170322ee82b12b8.jpg) + +![](images/6a10aacc221f2292be08afe6ecb05c70eb43148ae2a089269835f7b71247a392.jpg) + +these issues, we utilize the original GT (described in Section 3.1) to correct the generated masks. We focus on two situations: (1) If the GT contains multi-connected regions, we directly replace the corresponding regions in the generated mask with the multi-connected regions from the GT; (2) We iterate over all single-connected regions in the GT, generating a minimum bounding box for each region. If there is a region in the generated mask with a bounding box that overlaps more than $95\%$ with this GT bounding box, we retain the generated mask region; otherwise, we replace the region with the GT mask. Finally, we apply morphological operations (e.g., erosion and dilation) to remove noise and fill small holes [59]. + +We ultimately obtained 273 million "interactive masks", which can be used to train interactive segmentation models and cover nearly all identifiable objects in medical images. While we acknowledge that some masks may lack direct clinical relevance, they provide considerable diversity and density in the identified regions, helping the model learn a variety of interactive operations and enhancing its generalization ability across various tasks. + +# 3.3. Statistics and Analysis + +Data scale. As shown in Fig.1 (a), the IMed-361M dataset contains 6.4 million images, 87.6 million GT, and 273.4 + +million interactive masks, averaging 56 masks per image. This makes it the largest publicly available, multimodal, interactive medical image segmentation dataset to date. Compared to the MedTrinity-25M [31] dataset, our dataset provides 14.4 times the number of masks (Fig.1 (e)). + +Diversity. The IMed-361M dataset covers 14 imaging modalities and 204 segmentation targets, including organs and lesions, as shown in Fig.1 (b). We categorize the GT into six groups: Head and Neck, Thorax, Skeleton, Abdomen, Pelvis, and Lesions, covering nearly all parts of the human body. For clarity, we merge certain positional information when appropriate; for example, "left lung" and "right lung" are combined into "lung." Detailed category information can be found in the supplementary materials. Fig.1 (c) and (d) show the distribution of image resolutions and mask coverage within the dataset. Over $83\%$ of the images have resolutions between $256 \times 256$ and $1024 \times 1024$ , ensuring broad applicability and compatibility across various research scenarios. Additionally, most masks occupy less than $2\%$ of the image area, reflecting the typically fine granularity of medical segmentation. Our interactive masks provide over one million instances across different coverage intervals, significantly enhancing the diversity and density of the dataset. + +Mask quality. We randomly select five images from each + +![](images/71c22a66f93559d1090516b3e084aa27d89827732e7acbe32574826723bf679e.jpg) +Figure 2. Evaluation of the quality of interactive masks. + +subset of IMed-361M (a total of 550 images) and invited four radiologists to evaluate the quality of the corresponding interactive masks. We categorize the mask quality into three levels: (1) "Good": the mask closely matches the manual annotations; (2) "Satisfactory": the mask has some visible errors (e.g., incomplete contours) but is generally acceptable; (3) "Bad": the mask has significant errors (e.g., over-segmentation or under-segmentation) and is unsuitable for clinical segmentation tasks. The evaluation results show that 156 images are rated as "Good", 345 as "Satisfactory", and 49 as "Bad". Statistical analysis reveals that the masks rated as "Bad" mainly come from 18 datasets, concentrated within the red-boxed area in Fig.2. Radiologists note that some excessively annotated structures hold little clinical relevance. Additionally, some masks include elements unrelated to human organs or lesions, such as letters in X-ray backgrounds or hair in dermatoscopic images. Based on this analysis, we manually remove the excessively annotated masks from these 18 datasets and apply a $0.5\%$ foreground filter rate to retain only the final valid masks. Finally, we retain those masks unrelated to human organs or lesions to enhance the model's adaptability in different scenarios. + +# 4. IMIS Baseline Network + +# 4.1. Model Design + +We adopt a strategy similar to fine-tuning SAM [22-24] to establish the IMIS baseline, providing a performance benchmark for future work. As shown in Fig.3, IMIS-Net has three key components: an image encoder for extracting image features, a prompt encoder for integrating user interaction information, and a mask decoder for generating segmentation results using image and prompt embeddings. + +We select ViT-base [60] as the image encoder. While larger ViT models (e.g., ViT-Large, ViT-Huge) [22, 24] offer slight accuracy improvements, they significantly increase computational costs, making them impractical for clinical applications. Additionally, most open-source fine + +![](images/b5575537e62860e72e0805c96a128565d4110e03e1542e710ff212e9919bb526.jpg) +Figure 3. The training process of IMIS-Net simulates K consecutive steps of interactive segmentation. + +tuning methods also use ViT-base, ensuring fairness in our baseline. The image input size is $1024 \times 1024 \times 3$ , with a patch size of $16 \times 16 \times 3$ . For the prompt encoder, we consider three prompt types: points, boxes, and text. Points and boxes are represented by the sum of positional encoding and learned embeddings, while text is encoded using CLIP's text encoder [61]. Text prompts follow the template: "A segmentation area of a [category]", covering over 200 organ and lesion categories to establish a benchmark for future multimodal segmentation research. The mask decoder uses Transformer decoder blocks, mapping image and prompt embeddings to the mask. The mask resolution is increased to $256 \times 256$ through transposed convolutions and matched to the input size via bilinear interpolation. Notably, expanding the decoder can improve performance without significantly increasing network training parameters, thereby providing a simple solution to address the issue of model performance saturation. + +# 4.2. Experiment Configuration + +Training strategy. An overview is provided of the simulated continuous interactive segmentation training [22, 23, 62] in Fig.3. For a given segmentation task and medical image $x_{t}$ , we first simulate a set of initial interactions $u_{1}^{g}$ and $u_{1}^{i}$ based on the corresponding ground truth $y^{g}$ and interactive mask $y^{i}$ , which include clicks, bboxes, and text input. The click points are uniformly sampled from the foreground regions of $y^{g}$ or $y^{i}$ , while the bboxes are defined as the smallest bounding box around the target, with an offset of 5 pixels added to each coordinate to simulate slight user bias during the interaction process. The entire training process involves $K$ interactive training iterations (with $K = 8$ in this paper). The model's initial predictions are $\hat{y}_{1}^{g}$ and $\hat{y}_{1}^{i}$ . After the first prediction, we simulate subsequent corrections based on the previous predictions $\hat{y}_{k}^{g}$ and $\hat{y}_{k}^{i}$ , as well as the error region $\varepsilon_{k}$ between the $y^{g}$ and $y^{i}$ , where $k\in \{1,\dots,K\}$ . Additionally, we provide the low-resolution predicted mask from the previous prediction as an extra cue to the model. As can be seen, the image encoder only needs to encode the image once during the training, and subse + +![](images/1c5b9c29eb057a2c076ea4b903865f136237132afb5caa782696852008d15ee8.jpg) +Figure 4. Comparison of IMIS-Net with existing foundation models, with performance statistics at both image and mask levels. + +quent interactive training only updates the prompt encoder and mask decoder parameters. + +Training setting. We use the Adam optimizer [63] with a learning rate of $2 \times 10^{-5}$ . The training was conducted on 72 NVIDIA 4090 GPUs with a batch size of 2. For each image, we randomly select 5 targets from the corresponding GT and interactive masks as supervision targets (if fewer than 5 targets are present, they are selected repeatedly). The image resolution was uniformly resized to $1024 \times 1024$ , and pixel intensities are randomly scaled and shifted with a probability of $20\%$ , where the scaling factor and offset are set to 0.2. We train the model for a total of 12 epochs on the IMed-361M dataset and select the final checkpoint as the model weights. A linear combination of Focal loss [64] and Dice loss [65] is used as the loss function, balancing their influence in a 20:1 ratio. Finally, the Dice score is used as the primary evaluation metric for this study. + +# 4.3. Evaluation Model and Data + +Model. We evaluate five visual foundation models [17, 18, 22-24], selected based on their pretraining on large-scale datasets and the availability of source code. Among them, MedSAM, SAM-Med2D, and Huang et al. [24] are specifically designed for IMIS, while SAM [17] and SAM-2 [18] are pre-trained on massive natural image and video datasets. To ensure fairness in the evaluation, all models used ViT-Base as the encoder. It is important to note that, MedSAM and Huang et al. [24] only supportbbox-based interactivity (using the models released by the authors), while the other models support points, bboxes, and combinations of both. + +Data. The data used to evaluate the model is divided into two categories: the test set and the external dataset. The test set contains 58,411 images, with a distribution consistent with that of IMed-361M. The external dataset includes data from three different modalities and clinical scenarios: SegThor [66], which is used for segmentation of thoracic organs at risk in 516 CT images; TotalSegmentator MRI [37], which is used for segmenting organs in 2,147 MRI images; and ISLES $^{8}$ , which belongs to the MRI FLAIR modality + +and includes 364 images of ischemic stroke lesions. These datasets were not used during the training of IMIS-Net. + +# 5. Results + +# 5.1. Main Results + +Fig.4 shows the performance evaluation of IMIS-Net compared to other vision foundation models on the single-interaction segmentation task. During testing, all models used the same points and bounding boxes as prompts. The results indicate that IMIS-Net outperforms other models in both image and mask-level statistics. Thebbox interaction of different models always outperforms the click interaction because the_bbox can provide more boundary information. Notably, despite being pretrained on large-scale medical image datasets, MedSAM and SAM-Med2D still exhibit significant performance differences. This disparity is closely related to the scale and diversity of the pretraining datasets. As shown in Fig.5(a) and (b), the pre-training dataset for SAM-Med2D lacks samples of skeletal structures, resulting in poorer segmentation performance for these anatomical structures. Additionally, under the single-point prompt condition, SAM and SAM-2 achieve only Dice scores of $60.26\%$ and $59.57\%$ , respectively. likely due to the absence of medical knowledge in the pretraining data and limited interactive information constraining model performance. + +We increase the number of interactions from 1 to 9 to observe the performance changes of the model (Fig.5 (c)). As expected, performance improves with more interactions, and the gap between models narrows. This is because multiple interactions reduce the task difficulty. We also examine the impact of click position andbbox offset on performance. As shown in Fig.5 (d) and (f), when the prompt point is closer to the centroid, the performance improvement is more significant, with SAM-2's Dice score increasing by $2.84\%$ . Additionally, withbbox offset, all methods experience a performance decline of $0.85\% - 3.94\%$ . Our model exhibited the smallest performance drop, making it more robust for practical use. In summary, leveraging the rich data and mask diversity of the IMed-361M dataset, our model delivers the best performance across various medical scenarios and interaction strategies. More experimental results and analysis can be found in the supplementary material. + +# 5.2. External Dataset Evaluation + +Tab.1 presents evaluation results on external datasets. IMIS-Net achieves the best average performance across datasets from three different sources and tasks. We highlight the Dice scores of the 12 major abdominal organs in the TotalSegmentator MRI dataset, where IMIS-Net outperforms in 10. On the ISLES dataset, focusing on ischemic stroke + +![](images/b2a8b3f8de4bb386e96155e9544a21ffcee3ed7b251269b363a3a5de7bba1fc8.jpg) +(a) Single click interaction + +![](images/be5805d233413910604db172b5b082d819ebc47cc036dbc431d70b4d35cfe428.jpg) +(b) Bounding box interaction + +![](images/75fe536c264e83334298718fe107c66602369fd1b9adb65ede1e175be32aff0d.jpg) + +![](images/612c43c91ccc971dfb071c9cdf0c180dc4259a164de7097de8865514f4abb8de.jpg) +(c) Number of interactions + +![](images/ffcf64c8b0115fbe913f27f96189ea7a08a4bb2bfb216c1b8209e4ed866f9d13.jpg) +(d) Sampling points distances from the centroid + +![](images/e5a2e0b0d2fb0f7966e3edba3e081477ef5df3bec752e0e573c17041ee013ff4.jpg) +(e) Bbox offset +Figure 5. Comparison of segmentation performance across different anatomical structures under single-click (a) and bounding box (b) interactions. (c) Changes in segmentation performance with increasing interaction numbers. (d) and (e) Impact of click position and bounding box offset on performance. + +
DatasetCategorySAMSAM 2MedSAMHuang et al.SAM-Med2DIMIS-Net
ISLESIschemic Stroke Lesion55.9260.1459.9056.7968.2271.78
SegThorEsophagus78.1986.6047.1776.4882.0489.17
Heart91.4192.3481.1692.0294.1881.27
Aorta82.3683.6955.6884.5779.1294.92
Trachea93.7694.2472.4094.6085.1090.04
Average84.4685.8660.5584.7386.4389.27
Totalsegmentator MRIAdrenal gland77.1982.2556.5868.5266.3577.53
Aorta76.8879.9055.9880.0874.9185.62
Colon56.6956.0042.5355.4151.4473.47
Duodenum69.3074.8155.6472.3869.6775.97
Gallbladder81.3384.8268.1583.4687.6382.93
Iliopsoas59.5771.9658.4863.6760.7672.55
Kidney79.4381.5958.4181.0974.0482.72
Inferior vena cava84.9487.0357.7385.8878.3487.35
Liver86.9388.7068.1487.3586.7590.62
Pancreas66.9569.8646.1066.3066.8470.84
Spleen81.1283.0566.6183.8784.9586.74
Stomach77.1081.2362.7576.5279.3181.45
Average75.4577.6259.5275.4775.9279.06
+ +Table 1. Quantitative comparison results of IMIS-Net against five other interactive segmentation methods on three external datasets. + +lesion segmentation, our model surpasses the second-best SAM-Med2D by $3.56\%$ . These results demonstrate the strong generalization ability of our method. + +# 5.3. Ablation Study + +Scaling Up Training Data. As shown in Fig.6 (a) and (b), when trained only on the GT from IMed-361M, IMIS-Net performs poorly. However, with the addition of interac + +![](images/c002bcfa294e2975a2edafa0d32d8a3b9e28387a5f0f2e30416334d86b4c91ec.jpg) +Figure 6. (a) and (b) segmentation performance increases with the increasing mask density or data scales up. (c) effectiveness of text and its combination with other prompts. + +![](images/ad5a3eb647941b192851a6c2ce706e24d785abe1a704602861fd5cb0ca09d75e.jpg) + +![](images/3801d421fbf7d1668f29fef63ddb9de94aa1b2e2717461450f64c2ac6742946a.jpg) + +
Decoder dimensionImage resolutionPromptTraining parameters
PointBbox
768256×2560.82140.846929.68 M
768512×5120.86730.896829.68 M
2561024×10240.83660.84975.52 M
5121024×10240.85630.872915.19 M
7681024×10240.88480.906029.68 M
+ +Table 2. Ablation study of model design, including decoder dimension and image resolution. + +tive masks, the Dice score increases rapidly. Similar performance trends are observed when training on datasets of different sizes, indicating that our method is scalable and performs better with more available training data. + +Text and Combination Prompts. IMIS-Net, as a versatile interactive model, can utilize text prompts to achieve the segmentation of over 200 organs and lesions. As shown in Fig.6 (c), the model achieves a segmentation performance of $76.30\%$ when using only text prompts. When both text and point prompts are combined, the average Dice score increases by $11.95\%$ . Furthermore, after three rounds of click-based correction, the Dice score reaches $89.69\%$ . These results show that combining different prompts synergistically enhances the model's segmentation capabilities. + +Model Design. We conduct an ablation study on different configurations of the model, including the impact of input resolution and decoder dimension on model performance. As shown in Tab.2, the model's segmentation performance improves as the training image resolution increases. This higher input resolution allows for clearer visualization of lesions and organs in medical images. Additionally, when the decoder dimension is increased from 256 to 768, the model's performance improves from $84.97\%$ to $90.60\%$ , with only a 24.16M increase in trainable parameters. These + +results demonstrate the scalability of IMIS-Net, showing that even when the model's performance approaches saturation on larger datasets, further performance improvements can still be achieved by expanding the decoder. + +# 6. Conclusion + +In this work, we introduce IMed-361M, a benchmark dataset dedicated to interactive medical image segmentation. It includes a vast array of medical images across various modalities, extensive segmentation scenarios, and densely masks, surpassing all existing datasets that are limited to single tasks or simple integrations. Leveraging this data resource, we developed a general IMIS baseline model that enables users to generate segmentation results tailored to clinical needs through interactive methods, including clicks, bounding boxes, text prompts, and their combinations. We performed a comprehensive comparison of this baseline with existing foundational models, demonstrating that our model provides significant performance advantages and exhibits strong transferability in previously unseen scenarios. Notably, our approach typically requires fewer interactions to achieve comparable performance, enhancing its practicality in real-world applications. + +The IMed-361M dataset will strongly facilitate the development of foundational models in medical imaging and lay the foundation for fair evaluation across different models. IMIS-Net provides general technical support for various clinical applications, accelerating the widespread application of AI technology in the medical field. Despite these achievements, we recognize that this work still faces several challenges. For instance, effectively obtaining semantic information for interactive masks and extending this approach to more comprehensive and finer-grained medical image analysis scenarios are areas that require further exploration and improvement in the future. + +Acknowledgements: The authors acknowledge the funding provided by the National Key Research and Development Program of China (No.2022ZD0160102). + +# References + +[1] Tianfei Zhou, Lielei Li, Gustav Bredell, Jianwu Li, Jan Unkelbach, and Ender Konukoglu. Volumetric memory network for interactive medical image segmentation. Medical Image Analysis, 83:102599, 2023. 1, 2 +[2] Jian Zhang, Yinghuan Shi, Jinquan Sun, Lei Wang, Luping Zhou, Yang Gao, and Dinggang Shen. Interactive medical image segmentation via a point-based interaction. Artificial Intelligence in Medicine, 111:101998, 2021. 1, 2 +[3] Fang Chen, Haojie Han, Peng Wan, Hongen Liao, Chunrui Liu, and Daoqiang Zhang. Joint segmentation and differential diagnosis of thyroid nodule in contrast-enhanced ultrasound images. IEEE Transactions on Biomedical Engineering, 70(9):2722-2732, 2023. 1 +[4] Zdravko Marinov, Paul F Jäger, Jan Egger, Jens Kleesiek, and Rainer Stiefelhagen. Deep interactive segmentation of medical images: A systematic review and taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[5] Pierre-Henri Conze, Gustavo Andrade-Miranda, Vivek Kumar Singh, Vincent Jaouen, and Dimitris Visvikis. Current and emerging trends in medical image segmentation with deep learning. IEEE Transactions on Radiation and Plasma Medical Sciences, 7(6):545-569, 2023. 1 +[6] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[7] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. +[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. +[9] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, pages 3-11. Springer, 2018. +[10] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. +[11] Junlong Cheng, Shengwei Tian, Long Yu, Chengrui Gao, Xiaojing Kang, Xiang Ma, Weidong Wu, Shijia Liu, and + +Hongchun Lu. Resganet: Residual group attention network for medical image classification and segmentation. Medical Image Analysis, 76:102313, 2022. 1 +[12] Marco Cascella, Jonathan Montomoli, Valentina Bellini, and Elena Bignami. Evaluating the feasibility of chatgpt in healthcare: an analysis of multiple clinical and research scenarios. Journal of medical systems, 47(1):33, 2023. 1 +[13] Mingze Yuan, Yingda Xia, Hexin Dong, Zifan Chen, Jiawen Yao, Mingyan Qiu, Ke Yan, Xiaoli Yin, Yu Shi, Xin Chen, et al. Devil is in the queries: advancing mask transformers for real-world medical image segmentation and out-of-distribution localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23879-23889, 2023. 1 +[14] Jakob Wasserthal, Hanns-Christian Breit, Manfred T Meyer, Maurice Pradella, Daniel Hinck, Alexander W Sauter, Tobias Heye, Daniel T Boll, Joshy Cyriac, Shan Yang, et al. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence, 5(5), 2023. 1, 2 +[15] Sergios Gatidis, Tobias Hepp, Marcel Fruh, Christian La Fougère, Konstantin Nikolaou, Christina Pfaffenberg, Bernhard Schölkopf, Thomas Küstner, Clemens Cyran, and Daniel Rubin. A whole-body fdg-pet/ct dataset with manually annotated tumor lesions. Scientific Data, 9(1):601, 2022. 1, 2 +[16] Simon Graham, Mostafa Jahanifar, Quoc Dang Vu, Giorgos Hadjigeorghiou, Thomas Leech, David Snead, Shan E Ahmed Raza, Fayyaz Minhas, and Nasir Rajpoot. Conic: Colon nuclei identification and counting challenge 2022. arXiv preprint arXiv:2111.14485, 2021. 1 +[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3, 6 +[18] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 6 +[19] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 1 +[20] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, et al. Segment anything in 3d with nerfs. Advances in Neural Information Processing Systems, 36:25971-25990, 2023. +[21] Di Wang, Jing Zhang, Bo Du, Minqiang Xu, Lin Liu, Dacheng Tao, and Liangpei Zhang. Samrs: Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems, 36, 2024. 1 +[22] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 2, 3, 5, 6 + +[23] Junlong Cheng, Jin Ye, Zhongying Deng, Jianpin Chen, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Ji-long Chen, Lei Jiang, et al. Sam-med2d. arXiv preprint arXiv:2308.16184, 2023. 1, 5 +[24] Yuhao Huang, Xin Yang, Lian Liu, Han Zhou, Ao Chang, Xinrui Zhou, Rusi Chen, Junxuan Yu, Jiongquan Chen, Chaoyu Chen, et al. Segment anything model for medical images? Medical Image Analysis, 92:103061, 2024. 2, 3, 5, 6 +[25] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023. +[26] Yichi Zhang, Zhenrong Shen, and Rushi Jiao. Segment anything model for medical image segmentation: Current applications and future directions. Computers in Biology and Medicine, page 108238, 2024. 2 +[27] Shaoting Zhang and Dimitris Metaxas. On the challenges and perspectives of foundation models for medical image analysis. Medical image analysis, 91:102996, 2024. 2 +[28] Kummaragunta Joel Prabhod and Asha Gadhiraju. Foundation models in medical imaging: Revolutionizing diagnostic accuracy and efficiency. Journal of Artificial Intelligence Research and Applications, 4(1):471-511, 2024. +[29] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024. +[30] Raphael Schäfer, Till Nicke, Henning Hofener, AnnKristin Lange, Dorit Merhof, Friedrich Feuerhake, Volkmar Schulz, Johannes Lotz, and Fabian Kiessling. Overcoming data scarcity in biomedical imaging with a foundational multitask model. Nature Computational Science, 4(7):495-509, 2024. 2 +[31] Yunfei Xie, Ce Zhou, Lang Gao, Juncheng Wu, Xianhang Li, Hong-Yu Zhou, Sheng Liu, Lei Xing, James Zou, Cihang Xie, et al. Medtrinity-25m: A large-scale multimodal dataset with multigranular annotations for medicine. arXiv preprint arXiv:2408.02900, 2024. 2, 4 +[32] Kenneth Clark, Bruce Vendt, Kirk Smith, John Freymann, Justin Kirby, Paul Koppel, Stephen Moore, Stanley Phillips, David Maffitt, Michael Pringle, et al. The cancer imaging archive (tcia): maintaining and operating a public information repository. Journal of digital imaging, 26:1045-1057, 2013. 2 +[33] Joao Pedrosa, Guilherme Aresta, Carlos Ferreira, Gurraj Atwal, Hady Ahmady Phoulady, Xiaoyu Chen, Rongzhen Chen, Jiaoliang Li, Liansheng Wang, Adrian Galdran, et al. Lndb challenge on automatic lung cancer patient management. Medical image analysis, 70:102027, 2021. +[34] Rina D Rudyanto, Sjoerd Kerkstra, Eva M Van Rikxoort, Catalin Fetita, Pierre-Yves Brillet, Christophe Lefevre, Wenzhe Xue, Xiangjun Zhu, Jianming Liang, Ilkay Oksuz, et al. Comparing algorithms for automated vessel segmentation in computed tomography scans of the lung: the vessel12 study. Medical image analysis, 18(7):1217-1232, 2014. + +[35] Michela Antonelli, Annika Reinke, Spyridon Bakas, Keyvan Farahani, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, Olaf Ronneberger, Ronald M Summers, et al. The medical segmentation decathlon. Nature communications, 13(1):4128, 2022. +[36] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia modeling: 26th international conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, proceedings, part II 26, pages 451–462. Springer, 2020. 2 +[37] Tugba Akinci D'Antonoli, Lucas K Berger, Ashraya K Indrakanti, Nathan Vishwanathan, Jakob Weiβ, Matthias Jung, Zeynep Berkarda, Alexander Rau, Marco Reisert, Thomas Küstner, et al. Totalsegmentator mr: Sequence-independent segmentation of 59 anatomical structures in mr images. arXiv preprint arXiv:2405.19492, 2024. 2, 6 +[38] Igor Zingman, Birgit Stierstorfer, Charlotte Lempp, and Fabian Heinemann. Learning image representations for anomaly detection: application to discovery of histological alterations in drug development. Medical Image Analysis, 92:103067, 2024. +[39] Gašper Podobnik, Primož Strojan, Primož Peterlin, Bulat Ibragimov, and Tomaž Vrtovec. Han-seg: The head and neck organ-at-risk ct and mr segmentation dataset. Medical physics, 50(3):1917-1927, 2023. 2 +[40] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 2 +[41] Ziyan Huang, Haoyu Wang, Zhongying Deng, Jin Ye, Yanzhou Su, Hui Sun, Junjun He, Yun Gu, Lixu Gu, Shaoting Zhang, et al. Stu-net: Scalable and transferable medical image segmentation models empowered by large-scale supervised pre-training. arXiv preprint arXiv:2304.06716, 2023. 2 +[42] Jin Ye, Junlong Cheng, Jianpin Chen, Zhongying Deng, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Jilong Chen, Lei Jiang, et al. Sa-med2d-20m dataset: Segment anything in 2d medical imaging with 20 million masks. arXiv preprint arXiv:2311.11969, 2023. 2, 3 +[43] Sílvia Delgado Olabarriaga and Arnold WM Smeulders. Interaction in the segmentation of medical images: A survey. Medical image analysis, 5(2):127-142, 2001. 2 +[44] Guotai Wang, Maria A Zuluaga, Rosalind Pratt, Michael Aertsen, Tom Doel, Maria Klusmann, Anna L David, Jan Deprest, Tom Vercauteren, and Sébastien Ourselin. Slic-seg: A minimally interactive segmentation of the placenta from sparse and motion-corrupted fetal mri in multiple views. Medical image analysis, 34:137-147, 2016. 2 +[45] Hallee E Wong, Marianne Rakic, John Guttag, and Adrian V Dalca. Scribbleprompt: Fast and flexible interactive segmentation for any medical image. arXiv preprint arXiv:2312.07381, 2023. 2 +[46] Martin Rajchl, Matthew CH Lee, Ozan Oktay, Konstantinos Kamnitsas, Jonathan Passerat-Palmbach, Wenjia Bai, Mellisa Damodaram, Mary A Rutherford, Joseph V Hajnal, + +Bernhard Kainz, et al. Deepcut: Object segmentation from bounding box annotations using convolutional neural networks. IEEE transactions on medical imaging, 36(2):674-683, 2016. 2 +[47] Tomas Sakinis, Fausto Milletari, Holger Roth, Panagiotis Korfiatis, Petro Kostandy, Kenneth Philbrick, Zeynett Akkus, Ziyue Xu, Daguang Xu, and Bradley J Erickson. Interactive segmentation of medical images through fully convolutional neural networks. arXiv preprint arXiv:1903.08205, 2019. 2 +[48] Yuxin Du, Fan Bai, Tiejun Huang, and Bo Zhao. Segvol: Universal and interactive volumetric medical image segmentation. arXiv preprint arXiv:2311.13385, 2023. 2 +[49] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086-7096, 2022. 2 +[50] Leo Grady, Thomas Schiwietz, Shmuel Aharon, and Rüdiger Westermann. Random walks for interactive organ segmentation in two and three dimensions: Implementation and validation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2005: 8th International Conference, Palm Springs, CA, USA, October 26-29, 2005, Proceedings, Part II 8, pages 773-780. Springer, 2005. 2 +[51] Mingzhe Hu, Yuheng Li, and Xiaofeng Yang. Skinsam: adapting the segmentation anything model for skin cancer segmentation. In Medical Imaging 2024: Image Perception, Observer Performance, and Technology Assessment, volume 12929, pages 174-184. SPIE, 2024. 3 +[52] Yuheng Li, Mingzhe Hu, and Xiaofeng Yang. Polyp-sam: Transfer sam for polyp segmentation. In Medical Imaging 2024: Computer-Aided Diagnosis, volume 12927, pages 759-765. SPIE, 2024. 3 +[53] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 3 +[54] Jay N Paranjape, Nithin Gopalakrishnan Nair, Shameema Sikder, S Swaroop Vedula, and Vishal M Patel. Adaptivesam: Towards efficient tuning of sam for surgical scene segmentation. In Annual Conference on Medical Image Understanding and Analysis, pages 187-201. Springer, 2024. 3 +[55] Urban Borstnik, Joost VandeVondele, Valery Weber, and Jürg Hutter. Sparse matrix multiplication: The distributed block-compressed sparse row library. Parallel Computing, 40(5-6):47-58, 2014. 3 +[56] Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems, 36, 2024. 3 +[57] Yizhe Zhang, Tao Zhou, Shuo Wang, Peixian Liang, Yejia Zhang, and Danny Z Chen. Input augmentation with sam: Boosting medical image segmentation with segmentation foundation model. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 129-139. Springer, 2023. 3 + +[58] Charalampos Symeonidis, Ioannis Mademlis, Ioannis Pitas, and Nikos Nikolaidis. Neural attention-driven nonmaximum suppression for person detection. IEEE transactions on image processing, 32:2454-2467, 2023. 3 +[59] Suresh Kumar Tummala et al. Morphological operations and histogram analysis of sem images using python. Indian Journal of Engineering and Materials Sciences (IJEMS), 29(6):796-800, 2023. 4 +[60] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5 +[61] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[62] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. 5 +[63] Zijun Zhang. Improved adam optimizer for deep neural networks. In 2018 IEEE/ACM 26th international symposium on quality of service (IWQoS), pages 1-2. IEEE, 2018. 6 +[64] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 6 +[65] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6 +[66] Zoé Lambert, Caroline Petitjean, Bernard Dubray, and Su Kuan. Segthor: Segmentation of thoracic organs at risk in ct images. In 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA), pages 1- 6. IEEE, 2020. 6 \ No newline at end of file diff --git a/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/images.zip b/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6734995cf7a9fffe9396f09587e98072ec5cdeb0 --- /dev/null +++ b/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ecb37d7c39d33ab7d8cdb5d52cb8fdd5eebcfde6e1d9822a2b18fec327a063d1 +size 523930 diff --git a/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/layout.json b/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..39c44f8c419ce77b7750946473db70a9158770d5 --- /dev/null +++ b/2025/Interactive Medical Image Segmentation_ A Benchmark Dataset and Baseline/layout.json @@ -0,0 +1,8232 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 545, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 545, + 121 + ], + "type": "text", + "content": "Interactive Medical Image Segmentation: A Benchmark Dataset and Baseline" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "content": "Junlong Cheng" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "content": " Bin Fu" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "content": " Jin Ye" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "content": " Guoan Wang" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "content": " Tianbin Li" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "content": " Haoyu Wang" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "content": " Ruoyu Li" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "content": " He Yao" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "content": " Junren Cheng" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "content": " Jinwen Li" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "content": " Yanzhou Su" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "content": " Min Zhu" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "text", + "content": " Junjun He" + }, + { + "bbox": [ + 61, + 141, + 555, + 171 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 174, + 432, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 174, + 432, + 188 + ], + "spans": [ + { + "bbox": [ + 186, + 174, + 432, + 188 + ], + "type": "text", + "content": "1College of Computer Science, Sichuan University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 189, + 388, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 189, + 388, + 201 + ], + "spans": [ + { + "bbox": [ + 231, + 189, + 388, + 201 + ], + "type": "text", + "content": "2Shanghai AI Laboratory, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 147, + 202, + 471, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 202, + 471, + 216 + ], + "spans": [ + { + "bbox": [ + 147, + 202, + 471, + 216 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 147, + 202, + 471, + 216 + ], + "type": "text", + "content": "School of Computer Science and Technology, Xinjiang University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 217, + 497, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 217, + 497, + 229 + ], + "spans": [ + { + "bbox": [ + 121, + 217, + 497, + 229 + ], + "type": "text", + "content": "4 Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 167, + 235, + 451, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 235, + 451, + 246 + ], + "spans": [ + { + "bbox": [ + 167, + 235, + 451, + 246 + ], + "type": "text", + "content": "{chengjunlong, fubin, yejin, wangguoan, litianbin, wanghaoyu} @pjlab.org.cn," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 249, + 472, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 249, + 472, + 260 + ], + "spans": [ + { + "bbox": [ + 146, + 249, + 472, + 260 + ], + "type": "text", + "content": "{ruoyuli, 2024323040035, chengjunren} " + }, + { + "bbox": [ + 146, + 249, + 472, + 260 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 146, + 249, + 472, + 260 + ], + "type": "text", + "content": " stu.scu.edu.cn, 107556522204@stu.xju.edu.cn," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 172, + 263, + 443, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 263, + 443, + 274 + ], + "spans": [ + { + "bbox": [ + 172, + 263, + 443, + 274 + ], + "type": "text", + "content": "yanzhou.su@outlook.com, hejunjun.su@pjlab.org.cn, zhumin@scu.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 285, + 200, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 285, + 200, + 298 + ], + "spans": [ + { + "bbox": [ + 151, + 285, + 200, + 298 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 319, + 297, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 319, + 297, + 643 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 297, + 643 + ], + "type": "text", + "content": "Interactive Medical Image Segmentation (IMIS) has long been constrained by the limited availability of large-scale, diverse, and densely annotated datasets, which hinders model generalization and consistent evaluation across different models. In this paper, we introduce the IMed-361M benchmark dataset, a significant advancement in general IMIS research. First, we collect and standardize over 6.4 million medical images and their corresponding ground truth masks from multiple data sources. Then, leveraging the strong object recognition capabilities of a vision foundational model, we automatically generated dense interactive masks for each image and ensured their quality through rigorous quality control and granularity management. Unlike previous datasets, which are limited by specific modalities or sparse annotations, IMed-361M spans 14 modalities and 204 segmentation targets, totaling 361 million masks—an average of 56 masks per image. Finally, we developed an IMIS baseline network on this dataset that supports high-quality mask generation through interactive inputs, including clicks, bounding boxes, text prompts, and their combinations. We evaluate its performance on medical image segmentation tasks from multiple perspectives, demonstrating superior accuracy and scalability compared to existing interactive segmentation models. To facilitate research on foundational models in medical computer vision, we release the IMed-361M and model at https://github.com.uni-medical/IMIS-Bench." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 285, + 394, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 285, + 394, + 298 + ], + "spans": [ + { + "bbox": [ + 315, + 285, + 394, + 298 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 306, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 533 + ], + "type": "text", + "content": "Interactive Medical Image Segmentation (IMIS) enables clinicians or users to guide the model by marking points, lines, or regions on an image, resulting in segmentation outcomes that better meet clinical needs [1, 2]. This user-informed segmentation approach not only optimizes results throughout the segmentation process, aiding clinicians in precisely localizing targets during diagnosis and treatment [3-5], but also addresses the limitations of fully automated segmentation models in generalizing to unseen object classes (i.e., zero-shot learning) [6-11], providing clinicians with a flexible adjustment mechanism that significantly enhances segmentation reliability. However, the advancement of IMIS advancement faces a critical bottleneck: the lack of high-quality, diverse, and large-scale datasets that capture the complex clinical environments [12, 13]. Existing public datasets (such as Totalsegmentator [14], AutoPET [15], and CoNIC [16]) are often tailored to specific modalities or tasks, thus constraining the generalization capabilities of models across diverse medical scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": "Recently, as a foundational model for general visual segmentation, the Segment Anything Model (SAM) has garnered widespread attention [17, 18]. SAM integrates simple user interactions (such as points or bounding boxes) into the model's learning process and leverages pre-training on large-scale datasets to achieve cross-domain and multitask transferability [19-21]. Although SAM has been pretrained on over 1 billion natural image masks, its direct application in medical imaging remains limited due to the substantial domain differences between natural and medical images, as well as the lack of medical imaging-specific knowledge. To address these issues, current methods such as MedSAM [22] and SAM-Med2D [23] attempt to expand dataset size and fine-tune SAM to create IMIS models tailored for clinical scenarios. While these approaches have" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 144, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 144, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 144, + 703 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 703, + 113, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 113, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 113, + 712 + ], + "type": "text", + "content": "†Project lead." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20841" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "shown some progress, their ability to 'segment anything' in medical imaging remains limited. This limitation is mainly due to the lack of densely masks in existing medical datasets compared to the SA-1B dataset [17] used to train SAM [22-24]. For instance, the COSMOS dataset [24] contains an average of only 5.7 masks per image, which restricting the model's capacity for dense segmentation and hindering comprehensive, fine-grained interaction. Moreover, many methods are evaluated only on specific modalities or with limited interaction strategies, further restricting the comprehensiveness and reliability of IMIS model evaluation [22, 24-26]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 219, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 434 + ], + "type": "text", + "content": "Addressing these limitations requires the development of a high-quality IMIS benchmark dataset, which is essential for advancing foundational models in medical imaging [4, 27-30]. An ideal IMIS benchmark dataset should meet three core criteria: (1) Large-scale. The dataset should be large enough to fully support deep learning model training, enabling the model to effectively capture medical features; (2) Diversity. The dataset should encompass various medical imaging modalities and complex clinical scenarios to ensure the model's ability to generalize across modalities and tasks; (3) High-quality and densely masks. Accurate segmentation of complex medical images relies on high-quality annotations, and densely masks further enhance the model's capability to \"segment anything\" in medical imaging. Unfortunately, existing public medical segmentation datasets [22-24, 31] do not fully meet these standards, limiting their ability to comprehensively support and evaluate IMIS models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 438, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 713 + ], + "type": "text", + "content": "In this work, we introduce IMed-361M, a benchmark dataset specifically designed for IMIS tasks, this dataset integrates public and private data sources and utilizes foundational models [17] for automated annotation to generating dense masks for each image. A standardized data processing workflow is ensure the high quality and consistency of all masks. IMed-361M achieves unprecedented scale, diversity, and mask quality, comprising 6.4 million images spanning 14 imaging modalities and 204 targets, with a total of 361 million masks, averaging 56 masks per image. IMed-361M effectively solves the problems of small data size and sparse annotations, providing data support for IMIS model training. Additionally, we develop an IMIS baseline model and conduct a comprehensive evaluation of its performance across various medical scenarios, including its effectiveness on different modalities, anatomical structures, and organs, as well as the impact of various interaction strategies on model outcomes. This analysis provided an in-depth understanding of the strengths and limitations of different interactive segmentation methods, establishing a fair and consistent framework for evaluating IMIS model performance. We anticipate that the IMed-361M dataset and baseline model will drive the widespread adoption of IMIS" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "technology in clinical practice, accelerating the healthcare industry's transition toward intelligence and automation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 107, + 400, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 107, + 400, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 107, + 400, + 119 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 128, + 553, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 128, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 553, + 163 + ], + "type": "text", + "content": "This section provides an overview of datasets used for various medical image segmentation tasks and reviews the advancements in IMIS algorithms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 163, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 521 + ], + "type": "text", + "content": "Datasets. Due to variations in imaging protocols, medical images often have multidimensional characteristics (e.g., 2D, 3D), making pixel-level segmentation labeling for specific organs or lesions both expertise-dependent and time-consuming. Early work focused on single-organ or single-lesion annotated datasets [14, 15, 15, 32–36]. These single-task datasets provide foundational data for training automatic segmentation models, significantly enhancing model performance in segmenting specific organs or lesions. With the growth of multi-task requirements, multi-organ and multi-tissue annotated datasets [14, 37–39] have emerged, facilitating the exploration of deep learning models in multi-region segmentation tasks [40, 41]. In recent years, visual foundational models have demonstrated impressive cross-domain and multi-task transfer capabilities, motivating researchers to build large-scale, multimodal datasets for medical foundational models. The most common approach is to integrate data from various sources [22–24, 31] to bridge the scale gap compared to natural image datasets. However, due to the sparsity of annotations in the source data, the combined dataset still lags behind natural image datasets in terms of mask quantity and density. For example, the SA-1B [17] dataset averages 100 masks per image, whereas large-scale medical datasets like COSMOS [24] and SA-Med2D-20M [42] have fewer than 5 masks per image on average. This gap limits the ability of medical datasets to support dense segmentation tasks and fine-grained interactions, as well as the applicability of foundational models to \"segment anything\" in the medical imaging domain." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "content": "Algorithms. IMIS methods are generally more effective than automated methods at generating high-quality results that meet clinical requirements, especially in scenarios involving diverse imaging protocols, complex pathological variations, and ambiguous lesion boundaries [43]. User interactions can be take several typical forms, such as scribbles [44, 45], bounding boxes [17, 46], clicks [2, 47], or language prompts [17, 48, 49]. Traditional methods approach this task through energy by minimizing energy on a regular pixel grid, capturing low-level appearance features with unary potentials and promoting consistent segmentation outputs with pairwise or higher-order potentials [44, 50]. With advancements in deep learning, researchers have explored using user prompts directly as network input features, often in combination with 2D or 3D neural networks to produce segmentation results [1, 2, 17, 44-47]. Lever" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20842" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 299 + ], + "type": "text", + "content": "aging the advantage of pretraining on large-scale datasets, the Segment Anything Model (SAM) has become a benchmark for IMIS. For example, SAM can serve as a powerful pretrained encoder-decoder model fine-tuned for specific tasks [51, 52]. However, these approaches essentially revert to the design of fully automated segmentation models. Another category of methods retains the interactive aspect of segmentation, focusing on parameter-efficient fine-tuning (PEFT) techniques [53, 54]. However, these methods consider only a limited of interaction strategies, and inconsistencies in evaluation have affected the comparability and reliability of their results. To achieve robust segmentation capabilities across diverse medical imaging datasets, some approaches fine-tune SAM's decoder using large-scale medical datasets [22-24]. Although these methods demonstrate significant advantages across different medical modalities and tasks, they have not yet been consistently benchmarked against similar methods and lack a thorough examination of how different interaction strategies affect outcomes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 300, + 294, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 294, + 349 + ], + "type": "text", + "content": "In this work, we evaluate these state-of-the-art methods on the IMed-361M dataset, providing a comprehensive and fair comparison, while also exploring the impact of different interaction strategies on medical image segmentation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 360, + 271, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 271, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 271, + 372 + ], + "type": "text", + "content": "3. IMIS Benchmark Dataset:IMed-361M" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 380, + 295, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 295, + 452 + ], + "type": "text", + "content": "We present IMed-361M, the first large-scale IMIS dataset, that significantly surpasses existing datasets in terms of scale, diversity, quality, and density. This section details the data collection and pre-processing procedures, along with an in-depth analysis of the dataset's advantages and potential applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 461, + 243, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 243, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 243, + 474 + ], + "type": "text", + "content": "3.1. Data Collection and Pre-processing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "text", + "content": "Data collection. We integrate over 110 publicly available medical image segmentation datasets from globally recognized platforms such as TCIA" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "text", + "content": ", OpenNeuro" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "text", + "content": ", NITRC" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "text", + "content": ", Grand Challenge" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "text", + "content": ", Synapse" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "text", + "content": ", CodaLab" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "text", + "content": ", and GitHub" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 55, + 479, + 295, + 598 + ], + "type": "text", + "content": ", covering both 2D and 3D images and variety of formats (e.g., .jpg, .npy, .nii). Additionally, we collaborate with several medical institutions, acquiring diverse clinical data through rigorous ethical review processes, which further enriches the dataset. A detailed list of data sources is provided in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 599, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 635 + ], + "type": "text", + "content": "Preprocessing and filtering. We first standardize all collected medical images according to the SA-Med2D-20M [42] protocol, converting the corresponding ground" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "truth (GT) into one-hot encoding and storing it in .npz files in compressed sparse row (CSR) format [55]. Then, we apply the following exclusion criteria: (1) exclude 3D slice images and their corresponding masks with an aspect ratio greater than 1.5 to avoid interference from images with excessive geometric distortion during model training; (2) exclude masks where the foreground area accounts for less than one-thousandth of the total pixels, as such small foreground regions are prone to being lost during resizing, making it challenging to generate effective prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 193, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 312 + ], + "type": "text", + "content": "Resolving conflicts and ambiguities. To address conflicts and ambiguities in the GT, we first standardize expressions for the same target. For instance, both \"lung nodule\" and \"pulmonary nodule\" are renamed to \"lung nodule\" for consistency. We then manually review and correct misalignments and informational errors in the dataset to ensure annotation accuracy. Finally, for annotations with multiple connected components, we differentiate and label them based on clinical needs to avoid potential misunderstandings that could arise from single-point interactions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 313, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 553, + 396 + ], + "type": "text", + "content": "Through this process, we collected 6.4 million images and 87.6 million GT masks that were manually annotated and corrected. Although this already exceeds the data volume of most existing segmentation datasets, it still does not meet our fundamental requirement for the IMIS dataset: diverse and densely masks. Therefore, we further introduce interactive masks to address the above issues." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 407, + 421, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 421, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 421, + 418 + ], + "type": "text", + "content": "3.2. Interactive Masks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 553, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 639 + ], + "type": "text", + "content": "Automatic mask generation. The automatic mask generation method employed by SAM has been proven to be high-quality and effective [17, 56, 57]. We leverage SAM's object-awareness capability to generate as many masks as possible for each image. Specifically, we employ a " + }, + { + "bbox": [ + 313, + 426, + 553, + 639 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 426, + 553, + 639 + ], + "type": "text", + "content": " point grid to guide the model, generating a set of candidate masks for each point corresponding to potential objects of interest. The generated masks are refined and optimized using the following strategies: (1) Confidence filtering: Only masks with an Intersection over Union (IoU) prediction confidence score above 0.85 are retained to ensure segmentation accuracy. (2) Non-Maximum Suppression (NMS) [58]: For overlapping masks with an IoU greater than 0.7, only the mask with the highest confidence score is kept to eliminate redundancy. (3) Remove background masks: Masks covering more than " + }, + { + "bbox": [ + 313, + 426, + 553, + 639 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 426, + 553, + 639 + ], + "type": "text", + "content": " of the area are discarded, as such large-coverage foreground masks are almost nonexistent in the GT." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Quality control and granularity management. Based on our observations, the generated masks often fail to fully separate structures with unclear boundaries, such as the left atrium, myocardium, and right atrium in the heart. Additionally, masks for dispersed structures like the intestines are often identified as multiple separate objects. To address" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 643, + 196, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 643, + 196, + 654 + ], + "spans": [ + { + "bbox": [ + 67, + 643, + 196, + 654 + ], + "type": "text", + "content": "1https://www.cancerimagingarchive.net" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 654, + 141, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 654, + 141, + 664 + ], + "spans": [ + { + "bbox": [ + 67, + 654, + 141, + 664 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 654, + 141, + 664 + ], + "type": "text", + "content": "https://openneuro.org" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 664, + 141, + 674 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 664, + 141, + 674 + ], + "spans": [ + { + "bbox": [ + 67, + 664, + 141, + 674 + ], + "type": "text", + "content": "3https://www.nitrc.org" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 674, + 159, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 674, + 159, + 684 + ], + "spans": [ + { + "bbox": [ + 67, + 674, + 159, + 684 + ], + "type": "text", + "content": "4 https://grand-challenge.org" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 684, + 151, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 684, + 151, + 693 + ], + "spans": [ + { + "bbox": [ + 67, + 684, + 151, + 693 + ], + "type": "text", + "content": "5https://www.synapse.org" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 693, + 132, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 132, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 132, + 703 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 67, + 693, + 132, + 703 + ], + "type": "text", + "content": "https://codalab.org" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 703, + 130, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 130, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 130, + 712 + ], + "type": "text", + "content": "7https://github.com" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20843" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 76, + 252, + 240 + ], + "blocks": [ + { + "bbox": [ + 86, + 76, + 252, + 240 + ], + "lines": [ + { + "bbox": [ + 86, + 76, + 252, + 240 + ], + "spans": [ + { + "bbox": [ + 86, + 76, + 252, + 240 + ], + "type": "image", + "image_path": "ceb09083a3817149f6129ec884ea89bba6c5a7ba1e1c95f0344512db33ca1c26.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 253, + 74, + 527, + 239 + ], + "blocks": [ + { + "bbox": [ + 253, + 74, + 527, + 239 + ], + "lines": [ + { + "bbox": [ + 253, + 74, + 527, + 239 + ], + "spans": [ + { + "bbox": [ + 253, + 74, + 527, + 239 + ], + "type": "image", + "image_path": "987aefc771c31c81122029accb4bf1b8ddc73ae69a2e9c4914bbc5176d60acc0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 86, + 242, + 242, + 354 + ], + "blocks": [ + { + "bbox": [ + 86, + 242, + 242, + 354 + ], + "lines": [ + { + "bbox": [ + 86, + 242, + 242, + 354 + ], + "spans": [ + { + "bbox": [ + 86, + 242, + 242, + 354 + ], + "type": "image", + "image_path": "a97f21ed46ca3e9ca40ba1cd83a9daa0682aec774842add361bcc23ef039bbd6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 368, + 555, + 391 + ], + "lines": [ + { + "bbox": [ + 55, + 368, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 555, + 391 + ], + "type": "text", + "content": "Figure 1. Overview of the IMed-361M dataset. (a) Number of images and masks for each modality. (b) Information on six anatomical structures. (c) Distribution of image resolutions. (d) Analysis of mask proportions. (e) Comparison with other existing public datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 243, + 389, + 355 + ], + "blocks": [ + { + "bbox": [ + 242, + 243, + 389, + 355 + ], + "lines": [ + { + "bbox": [ + 242, + 243, + 389, + 355 + ], + "spans": [ + { + "bbox": [ + 242, + 243, + 389, + 355 + ], + "type": "image", + "image_path": "cb91498ad17b1295f0c833c03ecf24b99e217d2b3e99ea84c170322ee82b12b8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 389, + 245, + 529, + 354 + ], + "blocks": [ + { + "bbox": [ + 389, + 245, + 529, + 354 + ], + "lines": [ + { + "bbox": [ + 389, + 245, + 529, + 354 + ], + "spans": [ + { + "bbox": [ + 389, + 245, + 529, + 354 + ], + "type": "image", + "image_path": "6a10aacc221f2292be08afe6ecb05c70eb43148ae2a089269835f7b71247a392.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 411, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 296, + 567 + ], + "type": "text", + "content": "these issues, we utilize the original GT (described in Section 3.1) to correct the generated masks. We focus on two situations: (1) If the GT contains multi-connected regions, we directly replace the corresponding regions in the generated mask with the multi-connected regions from the GT; (2) We iterate over all single-connected regions in the GT, generating a minimum bounding box for each region. If there is a region in the generated mask with a bounding box that overlaps more than " + }, + { + "bbox": [ + 54, + 411, + 296, + 567 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 54, + 411, + 296, + 567 + ], + "type": "text", + "content": " with this GT bounding box, we retain the generated mask region; otherwise, we replace the region with the GT mask. Finally, we apply morphological operations (e.g., erosion and dilation) to remove noise and fill small holes [59]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 664 + ], + "type": "text", + "content": "We ultimately obtained 273 million \"interactive masks\", which can be used to train interactive segmentation models and cover nearly all identifiable objects in medical images. While we acknowledge that some masks may lack direct clinical relevance, they provide considerable diversity and density in the identified regions, helping the model learn a variety of interactive operations and enhancing its generalization ability across various tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 185, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 185, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 185, + 685 + ], + "type": "text", + "content": "3.3. Statistics and Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Data scale. As shown in Fig.1 (a), the IMed-361M dataset contains 6.4 million images, 87.6 million GT, and 273.4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "content": "million interactive masks, averaging 56 masks per image. This makes it the largest publicly available, multimodal, interactive medical image segmentation dataset to date. Compared to the MedTrinity-25M [31] dataset, our dataset provides 14.4 times the number of masks (Fig.1 (e))." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "type": "text", + "content": "Diversity. The IMed-361M dataset covers 14 imaging modalities and 204 segmentation targets, including organs and lesions, as shown in Fig.1 (b). We categorize the GT into six groups: Head and Neck, Thorax, Skeleton, Abdomen, Pelvis, and Lesions, covering nearly all parts of the human body. For clarity, we merge certain positional information when appropriate; for example, \"left lung\" and \"right lung\" are combined into \"lung.\" Detailed category information can be found in the supplementary materials. Fig.1 (c) and (d) show the distribution of image resolutions and mask coverage within the dataset. Over " + }, + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "type": "inline_equation", + "content": "83\\%" + }, + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "type": "text", + "content": " of the images have resolutions between " + }, + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "type": "text", + "content": ", ensuring broad applicability and compatibility across various research scenarios. Additionally, most masks occupy less than " + }, + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 472, + 556, + 699 + ], + "type": "text", + "content": " of the image area, reflecting the typically fine granularity of medical segmentation. Our interactive masks provide over one million instances across different coverage intervals, significantly enhancing the diversity and density of the dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": "Mask quality. We randomly select five images from each" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20844" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 294, + 205 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 294, + 205 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 294, + 205 + ], + "type": "image", + "image_path": "71c22a66f93559d1090516b3e084aa27d89827732e7acbe32574826723bf679e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 215, + 277, + 226 + ], + "lines": [ + { + "bbox": [ + 73, + 215, + 277, + 226 + ], + "spans": [ + { + "bbox": [ + 73, + 215, + 277, + 226 + ], + "type": "text", + "content": "Figure 2. Evaluation of the quality of interactive masks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 521 + ], + "type": "text", + "content": "subset of IMed-361M (a total of 550 images) and invited four radiologists to evaluate the quality of the corresponding interactive masks. We categorize the mask quality into three levels: (1) \"Good\": the mask closely matches the manual annotations; (2) \"Satisfactory\": the mask has some visible errors (e.g., incomplete contours) but is generally acceptable; (3) \"Bad\": the mask has significant errors (e.g., over-segmentation or under-segmentation) and is unsuitable for clinical segmentation tasks. The evaluation results show that 156 images are rated as \"Good\", 345 as \"Satisfactory\", and 49 as \"Bad\". Statistical analysis reveals that the masks rated as \"Bad\" mainly come from 18 datasets, concentrated within the red-boxed area in Fig.2. Radiologists note that some excessively annotated structures hold little clinical relevance. Additionally, some masks include elements unrelated to human organs or lesions, such as letters in X-ray backgrounds or hair in dermatoscopic images. Based on this analysis, we manually remove the excessively annotated masks from these 18 datasets and apply a " + }, + { + "bbox": [ + 55, + 247, + 296, + 521 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 55, + 247, + 296, + 521 + ], + "type": "text", + "content": " foreground filter rate to retain only the final valid masks. Finally, we retain those masks unrelated to human organs or lesions to enhance the model's adaptability in different scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 533, + 191, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 191, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 191, + 545 + ], + "type": "text", + "content": "4. IMIS Baseline Network" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 552, + 143, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 143, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 143, + 565 + ], + "type": "text", + "content": "4.1. Model Design" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "type": "text", + "content": "We adopt a strategy similar to fine-tuning SAM [22-24] to establish the IMIS baseline, providing a performance benchmark for future work. As shown in Fig.3, IMIS-Net has three key components: an image encoder for extracting image features, a prompt encoder for integrating user interaction information, and a mask decoder for generating segmentation results using image and prompt embeddings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "We select ViT-base [60] as the image encoder. While larger ViT models (e.g., ViT-Large, ViT-Huge) [22, 24] offer slight accuracy improvements, they significantly increase computational costs, making them impractical for clinical applications. Additionally, most open-source fine" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 566, + 182 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 566, + 182 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 566, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 566, + 182 + ], + "type": "image", + "image_path": "b5575537e62860e72e0805c96a128565d4110e03e1542e710ff212e9919bb526.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 188, + 555, + 212 + ], + "lines": [ + { + "bbox": [ + 313, + 188, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 555, + 212 + ], + "type": "text", + "content": "Figure 3. The training process of IMIS-Net simulates K consecutive steps of interactive segmentation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 233, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 448 + ], + "type": "text", + "content": "tuning methods also use ViT-base, ensuring fairness in our baseline. The image input size is " + }, + { + "bbox": [ + 313, + 233, + 555, + 448 + ], + "type": "inline_equation", + "content": "1024 \\times 1024 \\times 3" + }, + { + "bbox": [ + 313, + 233, + 555, + 448 + ], + "type": "text", + "content": ", with a patch size of " + }, + { + "bbox": [ + 313, + 233, + 555, + 448 + ], + "type": "inline_equation", + "content": "16 \\times 16 \\times 3" + }, + { + "bbox": [ + 313, + 233, + 555, + 448 + ], + "type": "text", + "content": ". For the prompt encoder, we consider three prompt types: points, boxes, and text. Points and boxes are represented by the sum of positional encoding and learned embeddings, while text is encoded using CLIP's text encoder [61]. Text prompts follow the template: \"A segmentation area of a [category]\", covering over 200 organ and lesion categories to establish a benchmark for future multimodal segmentation research. The mask decoder uses Transformer decoder blocks, mapping image and prompt embeddings to the mask. The mask resolution is increased to " + }, + { + "bbox": [ + 313, + 233, + 555, + 448 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 233, + 555, + 448 + ], + "type": "text", + "content": " through transposed convolutions and matched to the input size via bilinear interpolation. Notably, expanding the decoder can improve performance without significantly increasing network training parameters, thereby providing a simple solution to address the issue of model performance saturation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 456, + 462, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 462, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 462, + 469 + ], + "type": "text", + "content": "4.2. Experiment Configuration" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "Training strategy. An overview is provided of the simulated continuous interactive segmentation training [22, 23, 62] in Fig.3. For a given segmentation task and medical image " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": ", we first simulate a set of initial interactions " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "u_{1}^{g}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "u_{1}^{i}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " based on the corresponding ground truth " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "y^{g}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " and interactive mask " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "y^{i}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": ", which include clicks, bboxes, and text input. The click points are uniformly sampled from the foreground regions of " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "y^{g}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "y^{i}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": ", while the bboxes are defined as the smallest bounding box around the target, with an offset of 5 pixels added to each coordinate to simulate slight user bias during the interaction process. The entire training process involves " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " interactive training iterations (with " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " in this paper). The model's initial predictions are " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}_{1}^{g}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}_{1}^{i}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": ". After the first prediction, we simulate subsequent corrections based on the previous predictions " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}_{k}^{g}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}_{k}^{i}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": ", as well as the error region " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\varepsilon_{k}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " between the " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "y^{g}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "y^{i}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "k\\in \\{1,\\dots,K\\}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": ". Additionally, we provide the low-resolution predicted mask from the previous prediction as an extra cue to the model. As can be seen, the image encoder only needs to encode the image once during the training, and subse" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20845" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 298, + 167 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 298, + 167 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 298, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 298, + 167 + ], + "type": "image", + "image_path": "1c5b9c29eb057a2c076ea4b903865f136237132afb5caa782696852008d15ee8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 175, + 296, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 198 + ], + "type": "text", + "content": "Figure 4. Comparison of IMIS-Net with existing foundation models, with performance statistics at both image and mask levels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 220, + 296, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 296, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 296, + 244 + ], + "type": "text", + "content": "quent interactive training only updates the prompt encoder and mask decoder parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 245, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 424 + ], + "type": "text", + "content": "Training setting. We use the Adam optimizer [63] with a learning rate of " + }, + { + "bbox": [ + 55, + 245, + 296, + 424 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 245, + 296, + 424 + ], + "type": "text", + "content": ". The training was conducted on 72 NVIDIA 4090 GPUs with a batch size of 2. For each image, we randomly select 5 targets from the corresponding GT and interactive masks as supervision targets (if fewer than 5 targets are present, they are selected repeatedly). The image resolution was uniformly resized to " + }, + { + "bbox": [ + 55, + 245, + 296, + 424 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 55, + 245, + 296, + 424 + ], + "type": "text", + "content": ", and pixel intensities are randomly scaled and shifted with a probability of " + }, + { + "bbox": [ + 55, + 245, + 296, + 424 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 245, + 296, + 424 + ], + "type": "text", + "content": ", where the scaling factor and offset are set to 0.2. We train the model for a total of 12 epochs on the IMed-361M dataset and select the final checkpoint as the model weights. A linear combination of Focal loss [64] and Dice loss [65] is used as the loss function, balancing their influence in a 20:1 ratio. Finally, the Dice score is used as the primary evaluation metric for this study." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 208, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 208, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 208, + 445 + ], + "type": "text", + "content": "4.3. Evaluation Model and Data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 453, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 583 + ], + "type": "text", + "content": "Model. We evaluate five visual foundation models [17, 18, 22-24], selected based on their pretraining on large-scale datasets and the availability of source code. Among them, MedSAM, SAM-Med2D, and Huang et al. [24] are specifically designed for IMIS, while SAM [17] and SAM-2 [18] are pre-trained on massive natural image and video datasets. To ensure fairness in the evaluation, all models used ViT-Base as the encoder. It is important to note that, MedSAM and Huang et al. [24] only supportbbox-based interactivity (using the models released by the authors), while the other models support points, bboxes, and combinations of both." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "text", + "content": "Data. The data used to evaluate the model is divided into two categories: the test set and the external dataset. The test set contains 58,411 images, with a distribution consistent with that of IMed-361M. The external dataset includes data from three different modalities and clinical scenarios: SegThor [66], which is used for segmentation of thoracic organs at risk in 516 CT images; TotalSegmentator MRI [37], which is used for segmenting organs in 2,147 MRI images; and ISLES " + }, + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "text", + "content": ", which belongs to the MRI FLAIR modality" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "and includes 364 images of ischemic stroke lesions. These datasets were not used during the training of IMIS-Net." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 109, + 368, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 368, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 368, + 121 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 129, + 400, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 400, + 141 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 400, + 141 + ], + "type": "text", + "content": "5.1. Main Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 147, + 553, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 553, + 397 + ], + "type": "text", + "content": "Fig.4 shows the performance evaluation of IMIS-Net compared to other vision foundation models on the single-interaction segmentation task. During testing, all models used the same points and bounding boxes as prompts. The results indicate that IMIS-Net outperforms other models in both image and mask-level statistics. Thebbox interaction of different models always outperforms the click interaction because the_bbox can provide more boundary information. Notably, despite being pretrained on large-scale medical image datasets, MedSAM and SAM-Med2D still exhibit significant performance differences. This disparity is closely related to the scale and diversity of the pretraining datasets. As shown in Fig.5(a) and (b), the pre-training dataset for SAM-Med2D lacks samples of skeletal structures, resulting in poorer segmentation performance for these anatomical structures. Additionally, under the single-point prompt condition, SAM and SAM-2 achieve only Dice scores of " + }, + { + "bbox": [ + 313, + 147, + 553, + 397 + ], + "type": "inline_equation", + "content": "60.26\\%" + }, + { + "bbox": [ + 313, + 147, + 553, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 147, + 553, + 397 + ], + "type": "inline_equation", + "content": "59.57\\%" + }, + { + "bbox": [ + 313, + 147, + 553, + 397 + ], + "type": "text", + "content": ", respectively. likely due to the absence of medical knowledge in the pretraining data and limited interactive information constraining model performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 399, + 553, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 553, + 613 + ], + "type": "text", + "content": "We increase the number of interactions from 1 to 9 to observe the performance changes of the model (Fig.5 (c)). As expected, performance improves with more interactions, and the gap between models narrows. This is because multiple interactions reduce the task difficulty. We also examine the impact of click position andbbox offset on performance. As shown in Fig.5 (d) and (f), when the prompt point is closer to the centroid, the performance improvement is more significant, with SAM-2's Dice score increasing by " + }, + { + "bbox": [ + 313, + 399, + 553, + 613 + ], + "type": "inline_equation", + "content": "2.84\\%" + }, + { + "bbox": [ + 313, + 399, + 553, + 613 + ], + "type": "text", + "content": ". Additionally, withbbox offset, all methods experience a performance decline of " + }, + { + "bbox": [ + 313, + 399, + 553, + 613 + ], + "type": "inline_equation", + "content": "0.85\\% - 3.94\\%" + }, + { + "bbox": [ + 313, + 399, + 553, + 613 + ], + "type": "text", + "content": ". Our model exhibited the smallest performance drop, making it more robust for practical use. In summary, leveraging the rich data and mask diversity of the IMed-361M dataset, our model delivers the best performance across various medical scenarios and interaction strategies. More experimental results and analysis can be found in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 624, + 470, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 470, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 470, + 635 + ], + "type": "text", + "content": "5.2. External Dataset Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "Tab.1 presents evaluation results on external datasets. IMIS-Net achieves the best average performance across datasets from three different sources and tasks. We highlight the Dice scores of the 12 major abdominal organs in the TotalSegmentator MRI dataset, where IMIS-Net outperforms in 10. On the ISLES dataset, focusing on ischemic stroke" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 177, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 177, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 177, + 714 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 66, + 702, + 177, + 714 + ], + "type": "text", + "content": "https://www.isles-challenge.org/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20846" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 70, + 254, + 197 + ], + "blocks": [ + { + "bbox": [ + 102, + 70, + 254, + 197 + ], + "lines": [ + { + "bbox": [ + 102, + 70, + 254, + 197 + ], + "spans": [ + { + "bbox": [ + 102, + 70, + 254, + 197 + ], + "type": "image", + "image_path": "b2a8b3f8de4bb386e96155e9544a21ffcee3ed7b251269b363a3a5de7bba1fc8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 200, + 220, + 209 + ], + "lines": [ + { + "bbox": [ + 139, + 200, + 220, + 209 + ], + "spans": [ + { + "bbox": [ + 139, + 200, + 220, + 209 + ], + "type": "text", + "content": "(a) Single click interaction" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 262, + 70, + 414, + 199 + ], + "blocks": [ + { + "bbox": [ + 262, + 70, + 414, + 199 + ], + "lines": [ + { + "bbox": [ + 262, + 70, + 414, + 199 + ], + "spans": [ + { + "bbox": [ + 262, + 70, + 414, + 199 + ], + "type": "image", + "image_path": "be5805d233413910604db172b5b082d819ebc47cc036dbc431d70b4d35cfe428.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 200, + 382, + 209 + ], + "lines": [ + { + "bbox": [ + 294, + 200, + 382, + 209 + ], + "spans": [ + { + "bbox": [ + 294, + 200, + 382, + 209 + ], + "type": "text", + "content": "(b) Bounding box interaction" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 411, + 142, + 522, + 200 + ], + "blocks": [ + { + "bbox": [ + 411, + 142, + 522, + 200 + ], + "lines": [ + { + "bbox": [ + 411, + 142, + 522, + 200 + ], + "spans": [ + { + "bbox": [ + 411, + 142, + 522, + 200 + ], + "type": "image", + "image_path": "75fe536c264e83334298718fe107c66602369fd1b9adb65ede1e175be32aff0d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 85, + 213, + 228, + 323 + ], + "blocks": [ + { + "bbox": [ + 85, + 213, + 228, + 323 + ], + "lines": [ + { + "bbox": [ + 85, + 213, + 228, + 323 + ], + "spans": [ + { + "bbox": [ + 85, + 213, + 228, + 323 + ], + "type": "image", + "image_path": "612c43c91ccc971dfb071c9cdf0c180dc4259a164de7097de8865514f4abb8de.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 325, + 200, + 333 + ], + "lines": [ + { + "bbox": [ + 120, + 325, + 200, + 333 + ], + "spans": [ + { + "bbox": [ + 120, + 325, + 200, + 333 + ], + "type": "text", + "content": "(c) Number of interactions" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 231, + 212, + 376, + 324 + ], + "blocks": [ + { + "bbox": [ + 231, + 212, + 376, + 324 + ], + "lines": [ + { + "bbox": [ + 231, + 212, + 376, + 324 + ], + "spans": [ + { + "bbox": [ + 231, + 212, + 376, + 324 + ], + "type": "image", + "image_path": "ffcf64c8b0115fbe913f27f96189ea7a08a4bb2bfb216c1b8209e4ed866f9d13.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 325, + 377, + 334 + ], + "lines": [ + { + "bbox": [ + 235, + 325, + 377, + 334 + ], + "spans": [ + { + "bbox": [ + 235, + 325, + 377, + 334 + ], + "type": "text", + "content": "(d) Sampling points distances from the centroid" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 380, + 213, + 525, + 324 + ], + "blocks": [ + { + "bbox": [ + 380, + 213, + 525, + 324 + ], + "lines": [ + { + "bbox": [ + 380, + 213, + 525, + 324 + ], + "spans": [ + { + "bbox": [ + 380, + 213, + 525, + 324 + ], + "type": "image", + "image_path": "e5a2e0b0d2fb0f7966e3edba3e081477ef5df3bec752e0e573c17041ee013ff4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 325, + 476, + 333 + ], + "lines": [ + { + "bbox": [ + 430, + 325, + 476, + 333 + ], + "spans": [ + { + "bbox": [ + 430, + 325, + 476, + 333 + ], + "type": "text", + "content": "(e) Bbox offset" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 346, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 555, + 380 + ], + "type": "text", + "content": "Figure 5. Comparison of segmentation performance across different anatomical structures under single-click (a) and bounding box (b) interactions. (c) Changes in segmentation performance with increasing interaction numbers. (d) and (e) Impact of click position and bounding box offset on performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 83, + 389, + 529, + 616 + ], + "blocks": [ + { + "bbox": [ + 83, + 389, + 529, + 616 + ], + "lines": [ + { + "bbox": [ + 83, + 389, + 529, + 616 + ], + "spans": [ + { + "bbox": [ + 83, + 389, + 529, + 616 + ], + "type": "table", + "html": "
DatasetCategorySAMSAM 2MedSAMHuang et al.SAM-Med2DIMIS-Net
ISLESIschemic Stroke Lesion55.9260.1459.9056.7968.2271.78
SegThorEsophagus78.1986.6047.1776.4882.0489.17
Heart91.4192.3481.1692.0294.1881.27
Aorta82.3683.6955.6884.5779.1294.92
Trachea93.7694.2472.4094.6085.1090.04
Average84.4685.8660.5584.7386.4389.27
Totalsegmentator MRIAdrenal gland77.1982.2556.5868.5266.3577.53
Aorta76.8879.9055.9880.0874.9185.62
Colon56.6956.0042.5355.4151.4473.47
Duodenum69.3074.8155.6472.3869.6775.97
Gallbladder81.3384.8268.1583.4687.6382.93
Iliopsoas59.5771.9658.4863.6760.7672.55
Kidney79.4381.5958.4181.0974.0482.72
Inferior vena cava84.9487.0357.7385.8878.3487.35
Liver86.9388.7068.1487.3586.7590.62
Pancreas66.9569.8646.1066.3066.8470.84
Spleen81.1283.0566.6183.8784.9586.74
Stomach77.1081.2362.7576.5279.3181.45
Average75.4577.6259.5275.4775.9279.06
", + "image_path": "58f69a42f3ad53f97457c0fbe08cbadbffe52105143e94ced010130136e6fcbd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 66, + 626, + 542, + 639 + ], + "lines": [ + { + "bbox": [ + 66, + 626, + 542, + 639 + ], + "spans": [ + { + "bbox": [ + 66, + 626, + 542, + 639 + ], + "type": "text", + "content": "Table 1. Quantitative comparison results of IMIS-Net against five other interactive segmentation methods on three external datasets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 658, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 297, + 696 + ], + "type": "text", + "content": "lesion segmentation, our model surpasses the second-best SAM-Med2D by " + }, + { + "bbox": [ + 55, + 658, + 297, + 696 + ], + "type": "inline_equation", + "content": "3.56\\%" + }, + { + "bbox": [ + 55, + 658, + 297, + 696 + ], + "type": "text", + "content": ". These results demonstrate the strong generalization ability of our method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 658, + 408, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 658, + 408, + 671 + ], + "spans": [ + { + "bbox": [ + 314, + 658, + 408, + 671 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Scaling Up Training Data. As shown in Fig.6 (a) and (b), when trained only on the GT from IMed-361M, IMIS-Net performs poorly. However, with the addition of interac" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20847" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 72, + 216, + 195 + ], + "blocks": [ + { + "bbox": [ + 69, + 72, + 216, + 195 + ], + "lines": [ + { + "bbox": [ + 69, + 72, + 216, + 195 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 216, + 195 + ], + "type": "image", + "image_path": "c002bcfa294e2975a2edafa0d32d8a3b9e28387a5f0f2e30416334d86b4c91ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 208, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 555, + 232 + ], + "type": "text", + "content": "Figure 6. (a) and (b) segmentation performance increases with the increasing mask density or data scales up. (c) effectiveness of text and its combination with other prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 216, + 72, + 349, + 195 + ], + "blocks": [ + { + "bbox": [ + 216, + 72, + 349, + 195 + ], + "lines": [ + { + "bbox": [ + 216, + 72, + 349, + 195 + ], + "spans": [ + { + "bbox": [ + 216, + 72, + 349, + 195 + ], + "type": "image", + "image_path": "ad5a3eb647941b192851a6c2ce706e24d785abe1a704602861fd5cb0ca09d75e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 336, + 72, + 539, + 195 + ], + "blocks": [ + { + "bbox": [ + 336, + 72, + 539, + 195 + ], + "lines": [ + { + "bbox": [ + 336, + 72, + 539, + 195 + ], + "spans": [ + { + "bbox": [ + 336, + 72, + 539, + 195 + ], + "type": "image", + "image_path": "3801d421fbf7d1668f29fef63ddb9de94aa1b2e2717461450f64c2ac6742946a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 249, + 299, + 361 + ], + "blocks": [ + { + "bbox": [ + 56, + 249, + 299, + 361 + ], + "lines": [ + { + "bbox": [ + 56, + 249, + 299, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 249, + 299, + 361 + ], + "type": "table", + "html": "
Decoder dimensionImage resolutionPromptTraining parameters
PointBbox
768256×2560.82140.846929.68 M
768512×5120.86730.896829.68 M
2561024×10240.83660.84975.52 M
5121024×10240.85630.872915.19 M
7681024×10240.88480.906029.68 M
", + "image_path": "a1a1d096a647a8073fa0a3f4947a5c631bb282eefb161b9eb3b6a6553308a083.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 372, + 296, + 395 + ], + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 395 + ], + "type": "text", + "content": "Table 2. Ablation study of model design, including decoder dimension and image resolution." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 421, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 470 + ], + "type": "text", + "content": "tive masks, the Dice score increases rapidly. Similar performance trends are observed when training on datasets of different sizes, indicating that our method is scalable and performs better with more available training data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 472, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 592 + ], + "type": "text", + "content": "Text and Combination Prompts. IMIS-Net, as a versatile interactive model, can utilize text prompts to achieve the segmentation of over 200 organs and lesions. As shown in Fig.6 (c), the model achieves a segmentation performance of " + }, + { + "bbox": [ + 55, + 472, + 296, + 592 + ], + "type": "inline_equation", + "content": "76.30\\%" + }, + { + "bbox": [ + 55, + 472, + 296, + 592 + ], + "type": "text", + "content": " when using only text prompts. When both text and point prompts are combined, the average Dice score increases by " + }, + { + "bbox": [ + 55, + 472, + 296, + 592 + ], + "type": "inline_equation", + "content": "11.95\\%" + }, + { + "bbox": [ + 55, + 472, + 296, + 592 + ], + "type": "text", + "content": ". Furthermore, after three rounds of click-based correction, the Dice score reaches " + }, + { + "bbox": [ + 55, + 472, + 296, + 592 + ], + "type": "inline_equation", + "content": "89.69\\%" + }, + { + "bbox": [ + 55, + 472, + 296, + 592 + ], + "type": "text", + "content": ". These results show that combining different prompts synergistically enhances the model's segmentation capabilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Model Design. We conduct an ablation study on different configurations of the model, including the impact of input resolution and decoder dimension on model performance. As shown in Tab.2, the model's segmentation performance improves as the training image resolution increases. This higher input resolution allows for clearer visualization of lesions and organs in medical images. Additionally, when the decoder dimension is increased from 256 to 768, the model's performance improves from " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "84.97\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "90.60\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ", with only a 24.16M increase in trainable parameters. These" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 251, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 300 + ], + "type": "text", + "content": "results demonstrate the scalability of IMIS-Net, showing that even when the model's performance approaches saturation on larger datasets, further performance improvements can still be achieved by expanding the decoder." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 312, + 388, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 388, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 388, + 326 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 333, + 555, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 537 + ], + "type": "text", + "content": "In this work, we introduce IMed-361M, a benchmark dataset dedicated to interactive medical image segmentation. It includes a vast array of medical images across various modalities, extensive segmentation scenarios, and densely masks, surpassing all existing datasets that are limited to single tasks or simple integrations. Leveraging this data resource, we developed a general IMIS baseline model that enables users to generate segmentation results tailored to clinical needs through interactive methods, including clicks, bounding boxes, text prompts, and their combinations. We performed a comprehensive comparison of this baseline with existing foundational models, demonstrating that our model provides significant performance advantages and exhibits strong transferability in previously unseen scenarios. Notably, our approach typically requires fewer interactions to achieve comparable performance, enhancing its practicality in real-world applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 537, + 556, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 556, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 556, + 669 + ], + "type": "text", + "content": "The IMed-361M dataset will strongly facilitate the development of foundational models in medical imaging and lay the foundation for fair evaluation across different models. IMIS-Net provides general technical support for various clinical applications, accelerating the widespread application of AI technology in the medical field. Despite these achievements, we recognize that this work still faces several challenges. For instance, effectively obtaining semantic information for interactive masks and extending this approach to more comprehensive and finer-grained medical image analysis scenarios are areas that require further exploration and improvement in the future." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 669, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 555, + 703 + ], + "type": "text", + "content": "Acknowledgements: The authors acknowledge the funding provided by the National Key Research and Development Program of China (No.2022ZD0160102)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20848" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Tianfei Zhou, Lielei Li, Gustav Bredell, Jianwu Li, Jan Unkelbach, and Ender Konukoglu. Volumetric memory network for interactive medical image segmentation. Medical Image Analysis, 83:102599, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 179 + ], + "type": "text", + "content": "[2] Jian Zhang, Yinghuan Shi, Jinquan Sun, Lei Wang, Luping Zhou, Yang Gao, and Dinggang Shen. Interactive medical image segmentation via a point-based interaction. Artificial Intelligence in Medicine, 111:101998, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "type": "text", + "content": "[3] Fang Chen, Haojie Han, Peng Wan, Hongen Liao, Chunrui Liu, and Daoqiang Zhang. Joint segmentation and differential diagnosis of thyroid nodule in contrast-enhanced ultrasound images. IEEE Transactions on Biomedical Engineering, 70(9):2722-2732, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "type": "text", + "content": "[4] Zdravko Marinov, Paul F Jäger, Jan Egger, Jens Kleesiek, and Rainer Stiefelhagen. Deep interactive segmentation of medical images: A systematic review and taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "type": "text", + "content": "[5] Pierre-Henri Conze, Gustavo Andrade-Miranda, Vivek Kumar Singh, Vincent Jaouen, and Dimitris Visvikis. Current and emerging trends in medical image segmentation with deep learning. IEEE Transactions on Radiation and Plasma Medical Sciences, 7(6):545-569, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 413 + ], + "type": "text", + "content": "[6] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 414, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 295, + 480 + ], + "type": "text", + "content": "[7] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 481, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 295, + 534 + ], + "type": "text", + "content": "[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 536, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 635 + ], + "type": "text", + "content": "[9] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, pages 3-11. Springer, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 636, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 691 + ], + "type": "text", + "content": "[10] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[11] Junlong Cheng, Shengwei Tian, Long Yu, Chengrui Gao, Xiaojing Kang, Xiang Ma, Weidong Wu, Shijia Liu, and" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "type": "text", + "content": "Hongchun Lu. Resganet: Residual group attention network for medical image classification and segmentation. Medical Image Analysis, 76:102313, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 107, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 150 + ], + "type": "text", + "content": "[12] Marco Cascella, Jonathan Montomoli, Valentina Bellini, and Elena Bignami. Evaluating the feasibility of chatgpt in healthcare: an analysis of multiple clinical and research scenarios. Journal of medical systems, 47(1):33, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 151, + 554, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 554, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 554, + 227 + ], + "type": "text", + "content": "[13] Mingze Yuan, Yingda Xia, Hexin Dong, Zifan Chen, Jiawen Yao, Mingyan Qiu, Ke Yan, Xiaoli Yin, Yu Shi, Xin Chen, et al. Devil is in the queries: advancing mask transformers for real-world medical image segmentation and out-of-distribution localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23879-23889, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 228, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 554, + 293 + ], + "type": "text", + "content": "[14] Jakob Wasserthal, Hanns-Christian Breit, Manfred T Meyer, Maurice Pradella, Daniel Hinck, Alexander W Sauter, Tobias Heye, Daniel T Boll, Joshy Cyriac, Shan Yang, et al. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence, 5(5), 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 293, + 554, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 293, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 293, + 554, + 358 + ], + "type": "text", + "content": "[15] Sergios Gatidis, Tobias Hepp, Marcel Fruh, Christian La Fougère, Konstantin Nikolaou, Christina Pfaffenberg, Bernhard Schölkopf, Thomas Küstner, Clemens Cyran, and Daniel Rubin. A whole-body fdg-pet/ct dataset with manually annotated tumor lesions. Scientific Data, 9(1):601, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 360, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 360, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 360, + 554, + 415 + ], + "type": "text", + "content": "[16] Simon Graham, Mostafa Jahanifar, Quoc Dang Vu, Giorgos Hadjigeorghiou, Thomas Leech, David Snead, Shan E Ahmed Raza, Fayyaz Minhas, and Nasir Rajpoot. Conic: Colon nuclei identification and counting challenge 2022. arXiv preprint arXiv:2111.14485, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 415, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 554, + 480 + ], + "type": "text", + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 482, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 554, + 536 + ], + "type": "text", + "content": "[18] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 536, + 554, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 536, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 536, + 554, + 580 + ], + "type": "text", + "content": "[19] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 581, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 581, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 581, + 554, + 624 + ], + "type": "text", + "content": "[20] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, et al. Segment anything in 3d with nerfs. Advances in Neural Information Processing Systems, 36:25971-25990, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 625, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 554, + 679 + ], + "type": "text", + "content": "[21] Di Wang, Jing Zhang, Bo Du, Minqiang Xu, Lin Liu, Dacheng Tao, and Liangpei Zhang. Samrs: Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[22] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20849" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[23] Junlong Cheng, Jin Ye, Zhongying Deng, Jianpin Chen, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Ji-long Chen, Lei Jiang, et al. Sam-med2d. arXiv preprint arXiv:2308.16184, 2023. 1, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": "[24] Yuhao Huang, Xin Yang, Lian Liu, Han Zhou, Ao Chang, Xinrui Zhou, Rusi Chen, Junxuan Yu, Jiongquan Chen, Chaoyu Chen, et al. Segment anything model for medical images? Medical Image Analysis, 92:103061, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "text", + "content": "[25] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "text", + "content": "[26] Yichi Zhang, Zhenrong Shen, and Rushi Jiao. Segment anything model for medical image segmentation: Current applications and future directions. Computers in Biology and Medicine, page 108238, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "text", + "content": "[27] Shaoting Zhang and Dimitris Metaxas. On the challenges and perspectives of foundation models for medical image analysis. Medical image analysis, 91:102996, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 295, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 295, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 295, + 294, + 337 + ], + "type": "text", + "content": "[28] Kummaragunta Joel Prabhod and Asha Gadhiraju. Foundation models in medical imaging: Revolutionizing diagnostic accuracy and efficiency. Journal of Artificial Intelligence Research and Applications, 4(1):471-511, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 403 + ], + "type": "text", + "content": "[29] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 469 + ], + "type": "text", + "content": "[30] Raphael Schäfer, Till Nicke, Henning Hofener, AnnKristin Lange, Dorit Merhof, Friedrich Feuerhake, Volkmar Schulz, Johannes Lotz, and Fabian Kiessling. Overcoming data scarcity in biomedical imaging with a foundational multitask model. Nature Computational Science, 4(7):495-509, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[31] Yunfei Xie, Ce Zhou, Lang Gao, Juncheng Wu, Xianhang Li, Hong-Yu Zhou, Sheng Liu, Lei Xing, James Zou, Cihang Xie, et al. Medtrinity-25m: A large-scale multimodal dataset with multigranular annotations for medicine. arXiv preprint arXiv:2408.02900, 2024. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 590 + ], + "type": "text", + "content": "[32] Kenneth Clark, Bruce Vendt, Kirk Smith, John Freymann, Justin Kirby, Paul Koppel, Stephen Moore, Stanley Phillips, David Maffitt, Michael Pringle, et al. The cancer imaging archive (tcia): maintaining and operating a public information repository. Journal of digital imaging, 26:1045-1057, 2013. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "text", + "content": "[33] Joao Pedrosa, Guilherme Aresta, Carlos Ferreira, Gurraj Atwal, Hady Ahmady Phoulady, Xiaoyu Chen, Rongzhen Chen, Jiaoliang Li, Liansheng Wang, Adrian Galdran, et al. Lndb challenge on automatic lung cancer patient management. Medical image analysis, 70:102027, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[34] Rina D Rudyanto, Sjoerd Kerkstra, Eva M Van Rikxoort, Catalin Fetita, Pierre-Yves Brillet, Christophe Lefevre, Wenzhe Xue, Xiangjun Zhu, Jianming Liang, Ilkay Oksuz, et al. Comparing algorithms for automated vessel segmentation in computed tomography scans of the lung: the vessel12 study. Medical image analysis, 18(7):1217-1232, 2014." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "text", + "content": "[35] Michela Antonelli, Annika Reinke, Spyridon Bakas, Keyvan Farahani, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, Olaf Ronneberger, Ronald M Summers, et al. The medical segmentation decathlon. Nature communications, 13(1):4128, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "text", + "content": "[36] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia modeling: 26th international conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, proceedings, part II 26, pages 451–462. Springer, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 197, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 261 + ], + "type": "text", + "content": "[37] Tugba Akinci D'Antonoli, Lucas K Berger, Ashraya K Indrakanti, Nathan Vishwanathan, Jakob Weiβ, Matthias Jung, Zeynep Berkarda, Alexander Rau, Marco Reisert, Thomas Küstner, et al. Totalsegmentator mr: Sequence-independent segmentation of 59 anatomical structures in mr images. arXiv preprint arXiv:2405.19492, 2024. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "text", + "content": "[38] Igor Zingman, Birgit Stierstorfer, Charlotte Lempp, and Fabian Heinemann. Learning image representations for anomaly detection: application to discovery of histological alterations in drug development. Medical Image Analysis, 92:103067, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "text", + "content": "[39] Gašper Podobnik, Primož Strojan, Primož Peterlin, Bulat Ibragimov, and Tomaž Vrtovec. Han-seg: The head and neck organ-at-risk ct and mr segmentation dataset. Medical physics, 50(3):1917-1927, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "type": "text", + "content": "[40] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 411, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 474 + ], + "type": "text", + "content": "[41] Ziyan Huang, Haoyu Wang, Zhongying Deng, Jin Ye, Yanzhou Su, Hui Sun, Junjun He, Yun Gu, Lixu Gu, Shaoting Zhang, et al. Stu-net: Scalable and transferable medical image segmentation models empowered by large-scale supervised pre-training. arXiv preprint arXiv:2304.06716, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 477, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 553, + 531 + ], + "type": "text", + "content": "[42] Jin Ye, Junlong Cheng, Jianpin Chen, Zhongying Deng, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Jilong Chen, Lei Jiang, et al. Sa-med2d-20m dataset: Segment anything in 2d medical imaging with 20 million masks. arXiv preprint arXiv:2311.11969, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "text", + "content": "[43] Sílvia Delgado Olabarriaga and Arnold WM Smeulders. Interaction in the segmentation of medical images: A survey. Medical image analysis, 5(2):127-142, 2001. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 568, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 632 + ], + "type": "text", + "content": "[44] Guotai Wang, Maria A Zuluaga, Rosalind Pratt, Michael Aertsen, Tom Doel, Maria Klusmann, Anna L David, Jan Deprest, Tom Vercauteren, and Sébastien Ourselin. Slic-seg: A minimally interactive segmentation of the placenta from sparse and motion-corrupted fetal mri in multiple views. Medical image analysis, 34:137-147, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "text", + "content": "[45] Hallee E Wong, Marianne Rakic, John Guttag, and Adrian V Dalca. Scribbleprompt: Fast and flexible interactive segmentation for any medical image. arXiv preprint arXiv:2312.07381, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[46] Martin Rajchl, Matthew CH Lee, Ozan Oktay, Konstantinos Kamnitsas, Jonathan Passerat-Palmbach, Wenjia Bai, Mellisa Damodaram, Mary A Rutherford, Joseph V Hajnal," + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20850" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Bernhard Kainz, et al. Deepcut: Object segmentation from bounding box annotations using convolutional neural networks. IEEE transactions on medical imaging, 36(2):674-683, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 182 + ], + "type": "text", + "content": "[47] Tomas Sakinis, Fausto Milletari, Holger Roth, Panagiotis Korfiatis, Petro Kostandy, Kenneth Philbrick, Zeynett Akkus, Ziyue Xu, Daguang Xu, and Bradley J Erickson. Interactive segmentation of medical images through fully convolutional neural networks. arXiv preprint arXiv:1903.08205, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 295, + 216 + ], + "type": "text", + "content": "[48] Yuxin Du, Fan Bai, Tiejun Huang, and Bo Zhao. Segvol: Universal and interactive volumetric medical image segmentation. arXiv preprint arXiv:2311.13385, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 261 + ], + "type": "text", + "content": "[49] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086-7096, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 338 + ], + "type": "text", + "content": "[50] Leo Grady, Thomas Schiwietz, Shmuel Aharon, and Rüdiger Westermann. Random walks for interactive organ segmentation in two and three dimensions: Implementation and validation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2005: 8th International Conference, Palm Springs, CA, USA, October 26-29, 2005, Proceedings, Part II 8, pages 773-780. Springer, 2005. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "text", + "content": "[51] Mingzhe Hu, Yuheng Li, and Xiaofeng Yang. Skinsam: adapting the segmentation anything model for skin cancer segmentation. In Medical Imaging 2024: Image Perception, Observer Performance, and Technology Assessment, volume 12929, pages 174-184. SPIE, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "type": "text", + "content": "[52] Yuheng Li, Mingzhe Hu, and Xiaofeng Yang. Polyp-sam: Transfer sam for polyp segmentation. In Medical Imaging 2024: Computer-Aided Diagnosis, volume 12927, pages 759-765. SPIE, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 295, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 481 + ], + "type": "text", + "content": "[53] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 295, + 547 + ], + "type": "text", + "content": "[54] Jay N Paranjape, Nithin Gopalakrishnan Nair, Shameema Sikder, S Swaroop Vedula, and Vishal M Patel. Adaptivesam: Towards efficient tuning of sam for surgical scene segmentation. In Annual Conference on Medical Image Understanding and Analysis, pages 187-201. Springer, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 591 + ], + "type": "text", + "content": "[55] Urban Borstnik, Joost VandeVondele, Valery Weber, and Jürg Hutter. Sparse matrix multiplication: The distributed block-compressed sparse row library. Parallel Computing, 40(5-6):47-58, 2014. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 647 + ], + "type": "text", + "content": "[56] Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 713 + ], + "type": "text", + "content": "[57] Yizhe Zhang, Tao Zhou, Shuo Wang, Peixian Liang, Yejia Zhang, and Danny Z Chen. Input augmentation with sam: Boosting medical image segmentation with segmentation foundation model. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 129-139. Springer, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 497 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "text", + "content": "[58] Charalampos Symeonidis, Ioannis Mademlis, Ioannis Pitas, and Nikos Nikolaidis. Neural attention-driven nonmaximum suppression for person detection. IEEE transactions on image processing, 32:2454-2467, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[59] Suresh Kumar Tummala et al. Morphological operations and histogram analysis of sem images using python. Indian Journal of Engineering and Materials Sciences (IJEMS), 29(6):796-800, 2023. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[60] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 261 + ], + "type": "text", + "content": "[61] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 262, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 308 + ], + "type": "text", + "content": "[62] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 342 + ], + "type": "text", + "content": "[63] Zijun Zhang. Improved adam optimizer for deep neural networks. In 2018 IEEE/ACM 26th international symposium on quality of service (IWQoS), pages 1-2. IEEE, 2018. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "text", + "content": "[64] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 2980-2988, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 387, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 441 + ], + "type": "text", + "content": "[65] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 442, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 553, + 497 + ], + "type": "text", + "content": "[66] Zoé Lambert, Caroline Petitjean, Bernard Dubray, and Su Kuan. Segthor: Segmentation of thoracic organs at risk in ct images. In 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA), pages 1- 6. IEEE, 2020. 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20851" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Interleaved-Modal Chain-of-Thought/3edf2a1c-2168-41b7-a536-b8e17bc35145_content_list.json b/2025/Interleaved-Modal Chain-of-Thought/3edf2a1c-2168-41b7-a536-b8e17bc35145_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3d7b193a40dbf5471b6be4380669fbf619603748 --- /dev/null +++ b/2025/Interleaved-Modal Chain-of-Thought/3edf2a1c-2168-41b7-a536-b8e17bc35145_content_list.json @@ -0,0 +1,1748 @@ +[ + { + "type": "text", + "text": "Interleaved-Modal Chain-of-Thought", + "text_level": 1, + "bbox": [ + 308, + 130, + 689, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jun Gao $^{1}$ , Yongqi Li $^{2*}$ , Ziqiang Cao $^{1*}$ , Wenjie Li $^{2}$", + "bbox": [ + 308, + 179, + 686, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Technology, Soochow University \n $^{2}$ Department of Computer Science, The Hong Kong Polytechnic University \njgao1106@stu.suda.edu.cn, liyongqi0@gmail.com \nzqcao@suda.edu.cn, cswjli@comp.polyu.edu.hk \nhttps://github.com/jungao1106/ICoT", + "bbox": [ + 250, + 208, + 746, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 330, + 326, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chain-of-Thought (CoT) prompting elicits large language models (LLMs) to produce a series of intermediate reasoning steps before arriving at the final answer. However, when transitioning to vision-language models (VLMs), their text-only rationales struggle to express the fine-grained associations with the original image. In this paper, we propose an image-incorporated multimodal Chain-of-Thought, named Interleaved-modal Chain-of-Thought (ICoT), which generates sequential reasoning steps consisting of paired visual and textual rationales to infer the final answer. Intuitively, the novel ICoT requires VLMs to enable the generation of fine-grained interleaved-modal content, which is hard for current VLMs to fulfill. Considering that the required visual information is usually part of the input image, we propose Attention-driven Selection (ADS) to realize ICoT over existing VLMs. ADS intelligently inserts regions of the input image to generate the interleaved-modal reasoning steps with ignorable additional latency. ADS relies solely on the attention map of VLMs without the need for parameterization, and therefore it is a plug-and-play strategy that can be generalized to a spectrum of VLMs. We apply ADS to realize ICoT on two popular VLMs of different architectures. Extensive evaluations of three benchmarks have shown that ICoT prompting achieves substantial performance (up to $14\\%$ ) and interpretability improvements compared to existing multimodal CoT prompting methods.", + "bbox": [ + 88, + 362, + 485, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 786, + 220, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chain-of-Thought (CoT) [31] prompting aims to augment the reasoning capabilities of large language models (LLMs) [4, 8, 22, 26] by eliciting them to produce a sequence of intermediate natural language reasoning steps be", + "bbox": [ + 89, + 813, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fore arriving at the final output. CoT has proven effective in various reasoning tasks, including arithmetic [9], commonsense [17], and symbolic [3], and it has become a potential pathway to advanced artificial intelligence, as depicted in GPT-o1 [20].", + "bbox": [ + 511, + 330, + 905, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the development of vision-language models (VLMs), extending CoT prompting into multimodal CoT to improve the reasoning capabilities of VLMs in vision-related tasks becomes increasingly important [19, 28, 33, 34]. The initial multimodal CoT attempts [28, 33] take as input the fused visual and textual embeddings and train language models, such as T5 [23] models, to generate text-only rationales and answers. In the era of VLMs, introducing triple demonstrations composed of an image with the instruction, textual rationales, and the final output (e.g., answer) has proven effective in sparking the reasoning ability of VLMs [6]. Then, related studies focus on improving the linguistic reasoning ability of VLMs. Specifically, DDCoT [34] leverages VLMs to deconstruct problems and resolve them respectively, and CCoT [19] generates scene graphs to prompt VLMs with object and position descriptions. SCAFFOLD [12] overlays a coordinate matrix onto the image to prompt the VLMs with relative visual positions. However, these methods still generate text-only reasoning steps, making it hard to express the fine-grained associations with the origin image exactly. As shown on the left of Figure 1, textual position descriptions, e.g., at the top, are too rough to identify all fruits (orange and banana).", + "bbox": [ + 511, + 411, + 906, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In light of the limitations of text-only rationales, we propose incorporating visual information to enhance the precision of fine-grained associations between generated textual rationales and the corresponding image. We therefore propose an advanced multimodal Chain-of-Thought prompting, named Interleaved-modal Chain-of-thought (ICoT), as shown on the right of Figure 1. ICoT generates multimodal rationales consisting of paired images and textual rationales that formulate interleaved-modal reasoning steps", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 109, + 886, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19520", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model Input", + "text_level": 1, + "bbox": [ + 140, + 95, + 241, + 111 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4b546465888409e42344389e068077ff428c3a1db8ac1922234d281e1bdcf4c1.jpg", + "image_caption": [ + "Question: What types of fruits are in the picture?" + ], + "image_footnote": [], + "bbox": [ + 109, + 118, + 475, + 285 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-only Rationale", + "text_level": 1, + "bbox": [ + 127, + 316, + 281, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The mixed variety of fruits suggests that this is a fruit platter. The small, round, light green items at the bottom left appear to be grapes.", + "bbox": [ + 116, + 338, + 460, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The textured orange skin at the top indicates this is likely an orange, while a banana rests beside it.", + "bbox": [ + 116, + 387, + 460, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The dark red, smooth-skinned fruit on the right appears to be an apple.", + "bbox": [ + 116, + 422, + 455, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Answer", + "text_level": 1, + "bbox": [ + 122, + 469, + 181, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, grapes, orange, banana, and apple are in the picture.", + "bbox": [ + 120, + 494, + 428, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Input", + "text_level": 1, + "bbox": [ + 550, + 95, + 651, + 111 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0f38bad4fa19e33ce5395ce9331e107b56f82c06456e96216a6dfbdab1643d62.jpg", + "image_caption": [ + "Question: What types of fruits are in the picture?" + ], + "image_footnote": [], + "bbox": [ + 522, + 119, + 887, + 285 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interleaved-modal Rationale", + "text_level": 1, + "bbox": [ + 542, + 314, + 764, + 329 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ac007bed09b3a8ba675eaa1a0efac5fd8310a9ca4bd9c8569a4681a7984d6dae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 339, + 555, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The several grapes grouped together and the jumbled assortment of different fruits suggest that this is a fruit platter.", + "bbox": [ + 562, + 338, + 869, + 376 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f17f9a0c5296e00e97c2043e6c3152b8b8155b790bb15240a94105058f415c96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 382, + 555, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The orange colored rind behind an apple indicates that there is an orange and an apple in the picture.", + "bbox": [ + 562, + 387, + 867, + 414 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c05d7aaf6c67afc1cbfb92e9d124c53430002ab075e07bdb6ef19cf6f839e30.jpg", + "image_caption": [ + "Figure 1. The illustration between multimodal CoT with text-only rationales (Left) and interleaved-modal rationales (Right). Green blocks are correct texts used to infer the final answer. Text-only rationales restrict VLMs to use a rough description to indicate the position of objects. Transparent boxes indicate that these regions are selected and inserted to formulate paired visual and textual rationales in ICoT." + ], + "image_footnote": [], + "bbox": [ + 517, + 422, + 555, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The sleek yellow peel off to the right clues at an unpeeled banana.", + "bbox": [ + 563, + 422, + 854, + 448 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Answer", + "text_level": 1, + "bbox": [ + 531, + 470, + 591, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, grapes, orange, apple, and banana are in the picture.", + "bbox": [ + 526, + 491, + 879, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to infer the final output. To the best of our knowledge, ICoT is the first multimodal CoT with images incorporated, and it aligns more closely with human thinking processes [5, 21].", + "bbox": [ + 89, + 627, + 482, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Intuitively, facilitating the novel ICoT is non-trivial, as it introduces challenges for VLMs to support fine-grained interleaved-modal content generation. No current VLMs meet this condition completely. Perceiver-based VLMs such as Qwen2-VL [29] converts images into visual embeddings. Thus, they support fine-grained visual understanding but cannot generate multimodal outputs. Recently proposed unified-modeling VLMs, such as Chameleon [25], UnifiedIO 2 [18], and Emu-3 [30], enable multimodal generation by tokenizing images into discrete tokens. However, on the one hand, unified-modeling VLMs exhibit inertia toward multimodal content generation [7]; on the other hand, the generated images belong to the fixed pre-defined resolution instead of fine granularity.", + "bbox": [ + 88, + 672, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since required visual information is usually part of the", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "input image for ICoT, we accordingly propose Attention-driven Selection (ADS) to realize ICoT. The basic idea of ADS is to signal VLMs to select patches from the input image rather than generate extra images. At the beginning of generating each textual rationale, ADS inserts a piece of visual tokens of selected patches from the input image to refine the generation of the following textual rationale. Specifically, ADS utilizes the attention map of VLMs to identify optimal patches from the input image as fine-grained visual rationales. Once these fine-grained visual rationales are inserted into the current generation sequence, the VLM resumes the original autoregressive text generation process based on previous multimodal content, formulating paired image and textual rationales to infer the final outputs. Notably, since ADS does not compel VLMs to generate real images, it brings ignorable inference latency compared with previous text-only CoT methods. Additionally, ADS leverages the attention map of VLMs without re", + "bbox": [ + 511, + 627, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19521", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "quiring parameterization, making it a plug-and-play strategy that can be easily adapted to a wide range of VLMs.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we apply ADS to realize ICoT on Chameleon and Qwen2-VL, representing the state-of-the-art unified modeling and perceiver-based VLMs. The results on existing datasets, including $\\mathrm{M}^3\\mathrm{CoT}$ [6], ScienceQA [24], and LLaVA-W [15], indicate that ICoT realized by ADS brings VLMs with substantial performance gains (up to $14\\%$ ) compared with current multimodal CoT methods. Additionally, it is noted that the tracked interleaved-modal rationales further enhance the interpretability of the generated results. Our main contribution can be concluded as follows:", + "bbox": [ + 88, + 125, + 482, + 289 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose interleaved-modal CoT, which innovates text-only rationales into multimodal ones to construct clearer reasoning. To our knowledge, we are the first to incorporate images into the intermediate reasoning steps in multimodal CoT.", + "- We propose an effective and efficient Attention-driven Selection strategy to facilitate ICoT, which is training-free and widely applicable to VLMs without requiring them to support multimodal generation.", + "- Experiments demonstrate that our ICoT significantly surpasses existing multimodal CoT methods, proving that the interleaved-modal reasoning process is a foundational innovation in the line of CoT." + ], + "bbox": [ + 89, + 294, + 483, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 513, + 232, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Vision-Language Models (VLMs)", + "text_level": 1, + "bbox": [ + 89, + 541, + 382, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Currently, predominate VLMs such as Qwen-VL [2, 29], BLIP [13], and LLaVA [14-16] are mainly built upon a Large Language Model (LLM), a visual module, and an aligned vision-language adapter. The visual module, e.g., Vision Transformer (ViT) [1], encodes images into dense representations, and then the adapter, e.g., MLP or Q-Former, converts these representations into LLM-readable visual tokens. Finally, visual tokens and textual tokens are fed into the LLM to perform the next-token prediction. This type of VLM can be concluded as Perceiver-LLM architecture, while the Perceiver usually comprises a visual module and the adapter. Additionally, Cambrain-1 [27] introduces more visual modules to collaboratively provide more useful visual tokens in a vision-centric paradigm. In the other research line, unified-modeling VLMs represented by Chameleon [25], Unified-IO 2 [18], and Emu3 [30] are designed to generate texts, images, and so on uniformly. As these models apply codebook [11] to tokenize images into discrete vowels, their training processes are supervised by both vision and text information. Unified-modeling VLMs are expected to develop more stable multimodal understanding abality [10].", + "bbox": [ + 89, + 568, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Multimodal Chain-of-Thought Prompting", + "text_level": 1, + "bbox": [ + 511, + 90, + 872, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to CoT used in LLMs, multimodal Chain-of-Thought prompting methods [12, 19, 28, 32, 34] aim to augment the reasoning ability of VLM by generating intermediate reasoning steps. A series of studies focus on providing VLMs with fine-grained textual information, such as detailed description [28]. Compositional CoT (CCoT) [19] prompts VLMs to generate a Scene Graph (SG), which is a JSON-like description containing compositional information of objects that occurred in the image. DDCoT [34] deconstructs problems into small problems, requiring VLMs to solve them respectively and then inferring the final answer. In the other research line, Set-of-Marks prompting [32] augments the objects in the image to help VLMs recognize them. SCAFFOLD [12] overlays coordinate onto images to prompt VLMs with relative position information, and VLMs leverage overlayed textual coordinates to implicitly represent corresponding regions of the image to perform reasoning.", + "bbox": [ + 511, + 113, + 903, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, these methods still produce text-only rationales to infer the final answer. These generated textual rationales usually struggle to express the fine-grained associations with the origin image. We thereby propose ICoT to elicit VLMs to generate interleaved visual-textual reasoning steps to effectively reach the final outputs.", + "bbox": [ + 511, + 386, + 903, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 491, + 648, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the limitations that current multimodal CoT methods are still stuck in generating text-only rationales to infer the final answer, we propose interleaved-modal CoT (ICoT) to elicit VLMs generated multimodal reasoning steps. We start by introducing the workflow of VLMs and multimodal CoT in Section 3.1. We then introduce the concept of ICoT in Section 3.2. Finally, we propose a plug-and-play method, Attention-driven Selection (ADS), to realize ICoT on existing VLMs.", + "bbox": [ + 511, + 517, + 906, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 664, + 653, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first recall some background of VLMs and multimodal CoT in this section.", + "bbox": [ + 511, + 686, + 903, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language Model. VLMs usually consist of a visual encoder $\\mathbf{E}$ and a generative large language model LLM, and they determine where to insert images according to visual holders inserted in the text instructions. Then, VLMs take the image and the instructions as input and respond with a final answer", + "bbox": [ + 511, + 737, + 903, + 828 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {a n s w e r} = \\mathbf {V L M} (\\text {I m a g e}, \\text {I n s t r u c t i o n}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 840, + 903, + 858 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, the visual encoder $\\mathbf{E}$ extracts visual tokens $f_{v}^{l\\times d}$ from the image $x_{v}$ , where $l$ is the length of visual", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19522", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Interleaved-modal CoT", + "text_level": 1, + "bbox": [ + 91, + 90, + 336, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1: Input: Word embeddings $f_{e}$ , Visual tokens $f_{v}$ , Selected number $n$ , Signal tokens $S$ , Stopping criteria $SC$", + "bbox": [ + 99, + 109, + 483, + 142 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2: Output: Generated Response Answer", + "bbox": [ + 99, + 155, + 379, + 169 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3: predicted_tokens $\\leftarrow$ [] ▷ Initialize as an empty list", + "bbox": [ + 101, + 170, + 478, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4: $\\mathcal{V}\\gets f_v$", + "bbox": [ + 101, + 186, + 174, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5: inputs = Initilize $(f_e, f_v)$ Initialize inputs for prefilling", + "bbox": [ + 101, + 200, + 482, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6: while $SC$ not met do", + "bbox": [ + 101, + 231, + 263, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "7: next_token, attention_map = model(inputs)", + "bbox": [ + 101, + 246, + 442, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "8: Append next_token to predicted_tokens", + "bbox": [ + 101, + 261, + 410, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\triangleright$ ADS judgement", + "bbox": [ + 356, + 277, + 477, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "9: if predicted_tokens = S then", + "bbox": [ + 101, + 292, + 338, + 305 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "10: $\\mathcal{V}_{\\mathrm{selected}} = \\mathrm{ADS}(\\mathcal{V},\\mathrm{attention\\_map},n)$ Apply Attention-driven Selection", + "11: Append $\\nu_{\\mathrm{selected}}$ to predicted_tokens", + "12: end if", + "13: inputs = Update(inputs, predicted_tokens) > Updates inputs for next step generation", + "14: end while", + "15: Answer = Tokenizer Decode(predicted_tokens)", + "16: return Answer" + ], + "bbox": [ + 96, + 306, + 483, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tokens and $d$ is the dimensions of the hidden states of the LLM. The built-in LLM predicts next-tokens in a left-to-right fashion according to visual tokens $f_{v}^{l \\times d}$ and the instructions.", + "bbox": [ + 89, + 470, + 483, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal CoT. Compared with the direct prediction described in Eqn. 1, Multimodal CoT further introduces a prompt to elicit VLMs to generate a series of intermediate textual rationales $(r_1, r_2, \\ldots)$ before the final answer:", + "bbox": [ + 89, + 551, + 483, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$r_1, r_2, \\ldots$ , answer $= \\mathbf{VLM}(\\text{Prompt}, \\text{Image}, \\text{Instruction})$ . (2)", + "bbox": [ + 89, + 625, + 482, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Technically, the prompt could be represented as a sequence of demonstrations, each consisting of a triple: (Image, Rationale, Answer). Alternatively, an explicit instruction, such as \"Let's think step by step,\" could also serve as the prompt.", + "bbox": [ + 89, + 655, + 483, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Interleaved-modal Chain-of-Thought", + "text_level": 1, + "bbox": [ + 89, + 741, + 413, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous multimodal CoT prompting methods only produce text-only reasoning steps to improve the reasoning ability of VLMs. These intermediate steps are generated according to the entire image, which are struggle to express exact fine-grained associations with the original image. Given these limitations, we propose a more advanced Interleaved-modal Chain-of-Thought (ICoT) prompting, aiming to elicit VLMs to generate a series of multimodal intermediate reasoning steps, each consisting of paired im", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 2 Attention-driven Selection", + "text_level": 1, + "bbox": [ + 514, + 90, + 781, + 104 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Input: Attention map $A_{t}$ , Selected number $n$ , Visual tokens $f_{v}$", + "2: Output: Fine-grained visual information $\\nu_{\\mathrm{selected}}$", + "3: $\\mathcal{V}_{\\mathrm{selected}}\\gets \\emptyset$ Initialize as an empty set", + "4: Indices $\\leftarrow$ TopK(At,n)", + "5: for $i$ in Indices do", + "6: Append $f_{v}^{l \\times d}[i]$ to $\\mathcal{V}_{\\text{selected}}$", + "7: end for", + "8: Restore $(\\mathcal{V}_{\\mathrm{selected}}$ , Indices) $>$ Restore relative positions in-place", + "9: return $\\nu_{\\mathrm{selected}}$" + ], + "bbox": [ + 521, + 109, + 906, + 275 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c5d74196e5f92c316d404a4a83d73005768f844ef2700b28accda62fc1a2e011.jpg", + "image_caption": [ + "Figure 2. The workflow of ADS selecting fine-grained visual rationales. Signal attention represents the attention map of the signal token overall visual tokens." + ], + "image_footnote": [], + "bbox": [ + 517, + 292, + 900, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "age and textual rationale. Generated intermediate reasoning steps formulating interleaved-modal rationales to effectively lead to the final outputs. In this paper, we consider the visual rationales in interleaved-modal rationales as fine-grained visual information $x_{v}^{h' \\times w'}$ extracted from an image $x_{v}^{h \\times w1}$ . These visual rationales capture relevant details in the image, such as objects, colors, and texts, interleaved with the following generated textual rationale to infer the final answer:", + "bbox": [ + 511, + 520, + 906, + 655 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r} r _ {1}, x _ {v _ {1}}, r _ {2}, x _ {v _ {2}}, \\dots , \\text {a n s w e r} = \\mathbf {V L M} (\\text {P r o m p t , I m a g e}, \\\\ \\text {I n s t r u c t i o n}). \\end{array} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 665, + 905, + 713 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Attention-driven Selection", + "text_level": 1, + "bbox": [ + 513, + 723, + 751, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the proposed ICoT is both novel and conceptually sound, current VLMs are unable to generate such fine-grained visual information. This remains true even for VLMs [18, 25, 30] that are empowered with multimodal generation ability. We thus propose to simplify the problem from fine-grained visual information generation to fine-grained visual information selection, as this information", + "bbox": [ + 511, + 747, + 906, + 853 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1The images in the dataset involved in this paper are RGB images by default, and the number of channels is omitted in the formulas for simplicity.", + "bbox": [ + 511, + 862, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19523", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "has been naturally contained in the origin image, namely $x_{v}^{h^{\\prime}\\times w^{\\prime}}\\in x_{v}^{h\\times w}$ where $h^\\prime \\ll h$ and $w^{\\prime}\\ll w$", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, before performing next-token prediction, ICoT requires the VLM to cache visual tokens $f_{v}^{l \\times d}$ extracted by its built-in visual encoder $\\mathbf{E}(x_{v})$ for further selection. In the following decoding steps, we consider the VLM deems it necessary to insert a piece of visual rationales after generating a pre-defined signal token $S$ as shown in Figure 2, which is a natural language token that indicates the beginning of a textual rationale. Therefore, ADS will be signaled to select fine-grained visual information from $f_{v}^{l \\times d}$ upon the VLM generating $S$ :", + "bbox": [ + 89, + 125, + 483, + 276 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {d o} \\text {s e l e c t i o n} = \\left\\{ \\begin{array}{l l} \\text {T r u e} & \\text {i f p r e d i c t e d} \\\\ \\text {F a l s e} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 294, + 483, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, ADS selects $n$ visual tokens from $f_{v}^{l \\times d}$ as fine-grained visual information according to the signal attention map $A_{t}$ at the current decoding step $t$ :", + "bbox": [ + 89, + 345, + 483, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} _ {\\text {s e l e c t e d}} = \\left\\{f _ {v} [ i ] \\mid i \\in \\operatorname {T o p K} \\left(A _ {t}, n\\right) \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 411, + 482, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A_{t}$ is obtained by averaging the attention map between the signal token and visual tokens across all VLM layers. Up to now, the selected visual tokens are sorted by their attention scores, and we subsequently restore the relative position of $f_{v}[i]$ in the origin image in place, prioritizing rows. Once fine-grained visual tokens $\\nu_{\\mathrm{selected}}$ are obtained, VLMs will take as input the concatenated $\\nu_{\\mathrm{selected}}$ and predicted_tokens, denote as Cat(predicted_tokens, $\\nu_{\\mathrm{selected}}$ ), resuming the original autoregressive text generation process. Hence, the current decoding step is formulated as:", + "bbox": [ + 89, + 446, + 483, + 613 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r} \\text {n e x t - t o k e n} = \\mathbf {V L M} (\\text {C a t (p r e d i c t e d - t o k e n s}, \\\\ \\nu_ {\\text {s e l e c t e d}})). \\end{array} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 632, + 483, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notably, to avoid misunderstanding, we first convert predicted_tokens into word embeddings $f_{e}$ , and then concatenate $f_{e}$ with selected fine-grained visual information $\\mathcal{V}_{\\mathrm{selected}}$ in the embedding-level. We provide a detailed description of ICoT and ADS in Algorithm 1 and Algorithm 2, respectively.", + "bbox": [ + 89, + 685, + 482, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Technically, ICoT inherits the existing eliciting methods from CoT, such as attaching an instruction: \"Let's think step by step\" in zero-shot ICoT or providing few-shot examples. In the few-shot ICoT, each example consists of an image, textual rationales, and visual rationales. These examples can be manually designed to prompt VLMs regarding their way of thinking and generation formatting, among other aspects.", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 513, + 116, + 614, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{M}^{3}\\mathbf{C}\\mathbf{o}\\mathbf{T}$ [6] is a novel multimodal CoT benchmark specifically concentrated on multi-domain, multi-reasoning-step. $\\mathbf{M}^{3}\\mathbf{C}\\mathbf{o}\\mathbf{T}$ contains 267 categories from science, mathematics, and commonsense domains. As the question of each instance is relatively complex, their rationales have an average length of 293 tokens and rely more on fine-grained visual information, which can reflect the advantages of ICoT compared with previous multimodal CoT methods.", + "bbox": [ + 511, + 142, + 903, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ScienceQA [24] is a popular dataset used to evaluate the reasoning ability of VLMs. We use ScienceQA to provide a general comparison between ICoT and other existing multimodal CoT methods.", + "bbox": [ + 511, + 266, + 903, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLaVA-Bench In-the-Wild (LLaVA-W) [15] evaluates VLMs' ability to respond to visual questions with detailed long-form answers, which also focus on the fine-grained visual description. The reference label of each instance is produced by GPT-4v.", + "bbox": [ + 511, + 329, + 903, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Baselines", + "text_level": 1, + "bbox": [ + 513, + 422, + 620, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "No-CoT responds to the current input image and question directly without further prompting. The few-shot demonstrations of the direct generation mode consist of (Image, Question, Answer).", + "bbox": [ + 511, + 449, + 903, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal CoT [33] elicits VLMs to generate a series of text-only intermediate reasoning steps to infer the final outputs.", + "bbox": [ + 511, + 512, + 903, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CCoT [19] first generates a scene graph (SG) using the VLM itself and then uses that SG in the prompt to produce a response. The SG is a JSON-like structural description of the given image with extensive compositional information of objects in the current images. Following their settings, we apply their official prompt to prompt VLMs to generate SGs and answers respectively.", + "bbox": [ + 511, + 560, + 905, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DDCoT [34] first prompts LLM to deconstruct the input question into a sequence of basic sub-questions and then applies a VQA model to answer these sub-questions involving visual information. In this paper, we use the VLM that plays the role of LLM in DDCoT for a fair comparison, as their original LLM is ChatGPT.", + "bbox": [ + 511, + 667, + 905, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SCAFFOLD [12] overlays a coordinate matrix onto the input image, exactly demonstrating relative visual positions for VLMs. During reasoning, VLMs are steered to utilize these coordinates that indicate fine-grained visual information in the image to solve different vision-language tasks. We use their released scripts to add coordinates over each image and then use their official prompt to elicit VLMs.", + "bbox": [ + 511, + 761, + 905, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, in the few-shot scenario, the demonstrations of these baselines are human-written, aligning with ICoT.", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19524", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7ea645abd07d44b5900d0c2cc78d61f4ce523e3145bf7a43905b555c4af621e1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneMethods0-shot1-shot
M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑
Chameleon-7BNo-CoT29.147.713.128.448.523.9
Multimodal CoT [33]28.549.020.430.650.720.6
CCoT [19]29.450.222.131.451.324.5
DDCoT [34]28.649.820.229.849.223.1
SCAFFOLD [12]29.648.521.731.147.524.7
ICoT (Ours)29.851.025.232.353.427.6
% Improve0.6%1.6%14.0%2.8%4.0%11.7%
BackboneMethods0-shot1-shot
M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑
Qwen2-VL-7BNo-CoT43.656.332.745.464.433.5
Multimodal CoT [33]40.151.330.742.558.331.4
CCoT [19]43.356.429.444.163.833.9
DDCoT [34]42.655.231.245.764.932.8
SCAFFOLD [12]41.753.731.844.962.533.1
ICoT (Ours)44.156.834.246.065.435.7
% Improve1.1%0.7%4.6%0.6%0.7%5.3%
", + "bbox": [ + 91, + 87, + 903, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Results of ICoT and baselines based on Chameleon and Qwen2-VL, with the highest score bold. ${\\mathrm{M}}^{3}\\mathrm{{CoT}}$ and ScienceQA are evaluated by accuracy, and we report the ROUGE-L score for the LLaVA-W benchmark. % improve represents the relative improvement achieved by ICoT over the previously best baseline.", + "bbox": [ + 89, + 407, + 906, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implement Details", + "text_level": 1, + "bbox": [ + 89, + 476, + 267, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We apply ICoT over Chameleon-7B [25] and Qwen2-VL7B-Instruct [29], which represents the fine-grained visual information in the form of discrete vokens and dense features. All experiments are conducted on A800 GPUs, and we evaluate ICoT under both zero- and one-shot scenarios. During generating interleaved-modal rationales, the signal token $S$ used to trigger ADS is set to line break, i.e., \" $\\backslash n$ \", by default, which semantically and empirically indicates the end of a generated rationale and the beginning of the next one.", + "bbox": [ + 88, + 505, + 482, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VLMs insert visual tokens selected by ADS following the special token at the granularity of 64 according to posterior results shown in Table 4 of Appendix 8. Notably, to shorten the representation of an image, Qwen2-VL introduces a novel merge mechanism, and we approximately consider its patch size to be $(28 \\times 28)$ . Each patch of Qwen2-VL has approximately 4 times as many pixels as a Chameleon patch $(16 \\times 16)$ , which results in practical selection numbers of ADS set to 16. Considering the work of ADS requires the inner attention map, we apply the \"eager\" attention on both Chameleon-7B and Qwen2-VL, limited to the dependency of related python libraries2.", + "bbox": [ + 88, + 662, + 482, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Main Results", + "text_level": 1, + "bbox": [ + 511, + 476, + 651, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We comprehensively evaluate the performance of ICoT on top of Chameleon-7B and Qwen2-VL-7B through $\\mathbf{M}^3\\mathrm{CoT}$ , ScinceQA, and LLaVA-W in Table 1. In 0-shot settings, ICoT outperforms all baselines, including direct generation (No-CoT), CoT, CCoT, DDoT, and SCAFFOLD. Specifically, ICoT distinguishes from Multimodal CoT in terms of the modality of reasoning steps, which exhibit the advantages of interleaved-modal rationales to infer the final answer effectively. Compared with other multimodal CoT methods, the performance gains of ICoT further indicate that interleaved-modal rationales are more reasonable in intuition and effect than plainly inserted scene graphs (CCoT) and deconstructed sub-questions (DDoS). In 1-shot settings, ICoT demonstrations contain manually selected fine-grained visual information, while their text rationales are kept the same as other baselines. The performance gains compared with 0-shot ICoT indicate that our manually designed fine-grained ICoT demonstrations potentially guide VLMs to think in this format. In Table 4, we rigorously ablate the effectiveness of fine-grained ICoT demonstrations.", + "bbox": [ + 509, + 503, + 906, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, ICoT achieves the most relative performance gains in the LLaVA-W benchmark as the reference labels contain details sourced from images. These substantial performance gains compared with other baselines prove that visual tokens selected by ADS effectively capture the fine-grained visual information of an image, aiding the gen", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{2}$ Using attn_implementation=\"eager\" when loading the model from HuggingFace.", + "bbox": [ + 89, + 875, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19525", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "eration of high-quality text rationales.", + "bbox": [ + 89, + 90, + 341, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 116, + 243, + 132 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We ablate ICoT to verify the effectiveness of each portion across three benchmarks in Table 2 with the following settings: (1). w/o ADS: VLMs generate text-only rationales. (2). w/o FVI: Patches inserted in the demonstration are randomly sampled. Results indicate that both ADS and fine-grained visual information (FVI) incorporated in the demonstration are necessary. In particular, interleaved-modal rationales exhibit substantial advantages in generating high-quality textual rationales compared with text-only rationales (w/o ADS). When substituting ICoT demonstrations with normal ones (w/o FVI), the performance degradation proves that fine-grained visual information in demonstrations effectively guides VLMs to think in this format. Compared with the performance difference between removing ADS and FVI, we find that generating paired visual and text rationales boosts more improvements.", + "bbox": [ + 89, + 140, + 482, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, the performance gap is relatively smooth in ScienceQA and more dramatic on $\\mathrm{M}^3\\mathrm{CoT}$ and LLaVA-W. We attribute this to the ScienceQA dataset being relatively easier than others since both $\\mathrm{M}^3\\mathrm{CoT}$ and the answers of LLaVA-W highly rely on the fine-grained visual information of an image. Therefore, our proposed ICoT has the potential to solve complex vision-language tasks.", + "bbox": [ + 89, + 382, + 482, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. In-depth Analysis", + "text_level": 1, + "bbox": [ + 89, + 500, + 264, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis on realizing ICoT via KV Cache Up to now, the fine-grained visual information of ICoT is provided at the input end via discrete tokens or dense visual tokens, which brings more computation. After rethinking the generating process of an autoregressive model, there are other inputs that are proposed to avoid repeated computation, namely, the Key-Value (KV) Cache. The input image was stored in the KV Cache during the prefilling phase before generating multimodal intermediate reasoning steps in a left-to-right fashion. Therefore, copying the KV cache of fine-grained visual information enables ICoT with reduced computational costs, as visual information does not require extra forward propagation. As shown in Table 3, copying the KV Cache brings performance degradation compared with providing visual information at the input end. We attribute this phenomenon to the fact that although copying KV Cache indeed makes VLMs attend more to the same region as ADS, the optimal visual information is highlighted in a position-agnostic case, determined by the nature of KV Cache. Specifically, this degrades the original interleaved-modal rationales into non-interleaved ones as position information is early fused into KV Cache, and thus the copied ones are inherently insensitive to the position of textual rationale.", + "bbox": [ + 89, + 522, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, considering it brought slight performance", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fc6b090a6fe9cc20d05de99ede5d8023b74c0045b71c98f810d4042580981a2b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsM3CoTScienceQALLaVA-W
ICoT32.353.427.6
w/o ADS29.2 (-3.1)52.4(-1.0)24.5(-3.1)
w/o FVI30.6 (-1.8)52.8(-0.6)25.9(-1.7)
w/o ADS+FVI29.1(-3.2)51.0(-2.4)23.0(-4.6)
", + "bbox": [ + 522, + 90, + 901, + 167 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/676c7cef73ff2ff59c01a04621e776fb724bc41b323aa0fa61e4783ff14cb381.jpg", + "table_caption": [ + "Table 2. Ablation studies of 1-shot ICoT on Chameleon-7B. () describes the performance degradation compared with ICoT. FVI indicates the 1-shot demonstration contains fine-grained visual information. ADS indicates that VLMs generate interleaved-modal reasoning steps." + ], + "table_footnote": [], + "table_body": "
Dataset0-shot1-shot
KV-CopyICoTKV-CopyICoT
M3CoT29.129.831.532.3
ScienceQA49.751.052.953.4
LLaVA-W24.725.227.027.6
", + "bbox": [ + 522, + 268, + 901, + 351 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8bdf56f5901a5e118a20b477e55b77d6e18bbb1b7f5f7e6dc2a055ee3ae70347.jpg", + "table_caption": [ + "Table 3. Results comparison between copying KV Cache (KV-Copy) and inserting selected patches." + ], + "table_footnote": [], + "table_body": "
MethodsM3CoTScienceQALLaVA-W
Human-written32.353.427.6
Model-written31.5(-0.8)51.8(-1.6)26.7(-0.8)
", + "bbox": [ + 522, + 409, + 901, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Results concerning the design of demonstrations on Chameleon-7B. () describes the performance degradation compared with ICoT. Human-written indicates that demonstrations are manually designed with fine-grained visual information inserted, and Model-written indicates that the VLM generates both visual and textual rationale via ICoT.", + "bbox": [ + 511, + 472, + 905, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "degradation and factually reduced computation costs, we believe this exploration is still valuable, and we call for more interesting exploration in realizing ICoT.", + "bbox": [ + 511, + 583, + 905, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis on the Demonstrations We also attempt to let VLMs generate the demonstrations via themselves (Automatic at the bottom of Table 4). Results indicate that using the automatically generated demonstrations also brings performance degradation compared with ICoT using manually designed ones. We consider it is caused by the fact that formulating a continuous sub-image through ADS is nontrivial, and some discrete patches inevitably introduce additional noise. Therefore, considering that designing such a demonstration is not time-consuming, ICoT utilizes manually designed ones to elicit VLMs to perform ICoT for better performance.", + "bbox": [ + 511, + 648, + 906, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Case Study", + "text_level": 1, + "bbox": [ + 511, + 844, + 633, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we empirically illustrate the advantages of ICoT via three case studies in Figure 3. These case studies", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19526", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/70796664c06c33732a7065f24974cfdd8a99c09ec9a847f3cef367a45cd3bfd5.jpg", + "image_caption": [ + "Figure 3. Case studies between ICoT and multimodal CoT with text-only rationales on Chameleon. Three cases are selected according to three typical problems in text-only problems: misunderstanding, overgeneralization, and hallucination. Red blocks indicate the incorrect rationales." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 897, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "focused on three typical problems that occurred in text-only rationales, namely, misunderstanding (top), overgeneralization (middle), and hallucination (bottom).", + "bbox": [ + 89, + 390, + 482, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, in the first case, interleaved-modal CoT first recognizes three different objects via captions: \"inflatable castle, crayons, and a parachute\". Then, in the second reasoning step, ADS inserts selected patches from the scheduled objects to elicit the VLM to conclude their common property, and VLM infers a correct answer. Text-only CoT misunderstands that the three objects are all colored pencils, ignoring the castle and the parachute, even the final answer is correct. In the second case, text-only rationales overgeneralize flying a kite to a kite festival, leading to a wrong answer. ICoT first recognizes a man standing in a field of grass and then infers it is a windy day according to a few kites in the sky. In the last case, it provides the other typical error of text-only CoT, namely, hallucination. As text-only CoT purely relies on language reasoning ability, VLMs have the potential to imagine something not mentioned in the image, resulting in a wrong answer. ICoT first infers from the street sign that there may be a troll attraction nearby according to patches of the indicator inserted by ADS. Then, the ADS helps the VLMs to attend to the troll statue under the bridge and infer it is likely the mentioned attraction, finally arriving at the correct answer.", + "bbox": [ + 89, + 455, + 483, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Even though the above case studies exhibit the advantages of ICoT, ADS still brings potential problems. For example, ADS is triggered to select patches when VLM generates a pre-defined signal token. This simple mechanism is a double-edged sword in that VLMs will generate low-quality responses if this token is generated with a high frequency.", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 388, + 633, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we first propose interleaved-modal CoT (ICoT), which generates interleaved-modal rationales to infer the final answer effectively. In light of the challenges of applying ICoT on existing VLMs, we then introduce Attention-driven Selection (ADS), a plug-and-play strategy to identify optimal patches from the image without being parameterized. We evaluate ICoT on Chameleon-7B and Qwen2-VL-7B-Instruct, representing VLMs of two architectures. Extensive experiments conducted on $\\mathbf{M}^3\\mathrm{CoT}$ , ScienceQA, and LLaVA-W, under both zero- and few-shot scenarios, have proven that ICoT achieves substantial performance (up to $14\\%$ ) compared with the existing multimodal CoT methods. Additionally, in the analysis section, we conduct a preliminary exploration of implementing ICoT by copying the KV cache of optimal visual tokens and explain the inner trade-off between efficiency and performance in this approach.", + "bbox": [ + 511, + 438, + 906, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although ICoT has proven its effectiveness in this paper, we consider that ICoT still has significant potential for further improvement. The patch selection in ADS requires storing attention scores, which brings additional memory overhead. Moreover, the fixed number of selected patches in the ADS design is sub-optimal, resulting in unexpected outputs for VLMs. To address these, we intend to incorporate established techniques from segmentation or grounding methods to create a more robust implementation of ICoT. In the future, we also plan to evaluate it across additional backbones and benchmarks to better assess its generalization ability.", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19527", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "I would like to express my sincere gratitude to all the authors and reviewers for their valuable contributions to this research. This work was supported by the National Natural Science Foundation of China (NSFC 62106165) and the Project Funded by the Priority Academic Program Development of Jiangsu Higher Education Institutions, China.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.3", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 3", + "[3] BIG bench authors. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on Machine Learning Research, 2023. 1", + "[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1", + "[5] Claus Bundesen. A theory of visual attention. Psychological review, 97(4):523, 1990. 2", + "[6] Qiguang Chen, Libo Qin, Jin Zhang, Zhi Chen, Xiao Xu, and Wanxiang Che. M3cot: A novel benchmark for multidomain multi-step multi-modal chain-of-thought. arXiv preprint arXiv:2405.16473, 2024. 1, 3, 5", + "[7] Ethan Chern, Jiadi Su, Yan Ma, and Pengfei Liu. Anole: An open, autoregressive, native large multimodal models for interleaved image-text generation. arXiv preprint arXiv:2407.06135, 2024. 2", + "[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 1", + "[9] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 1", + "[10] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. arXiv preprint arXiv:2406.11832, 2024.3", + "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3", + "[12] Xuanyu Lei, Zonghan Yang, Xinrui Chen, Peng Li, and Yang Liu. Scaffolding coordinates to promote vision-language coordination in large multi-modal models. arXiv preprint arXiv:2402.12058, 2024. 1, 3, 5, 6" + ], + "bbox": [ + 93, + 244, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3", + "[14] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3", + "[15] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.3,5", + "[16] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3", + "[17] Jiacheng Liu, Alisa Liu, Ximing Lu, Sean Welleck, Peter West, Ronan Le Bras, Yejin Choi, and Hannaneh Hajishirzi. Generated knowledge prompting for commonsense reasoning. arXiv preprint arXiv:2110.08387, 2021. 1", + "[18] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26439-26455, 2024. 2, 3, 4", + "[19] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 1, 3, 5, 6", + "[20] OpenAI. Learning to reason with llms, 2023. 1", + "[21] Michael I Posner, Steven E Petersen, et al. The attention system of the human brain. Annual review of neuroscience, 13(1):25-42, 1990. 2", + "[22] Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446, 2021. 1", + "[23] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1", + "[24] Tanik Saikh, Tirthankar Ghosal, Amish Mittal, Asif Ekbal, and Pushpak Bhattacharyya. Scienceqa: A novel resource for question answering on scholarly articles. International Journal on Digital Libraries, 23(3):289-301, 2022. 3, 5", + "[25] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2,3,4,6", + "[26] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 1", + "[27] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang," + ], + "bbox": [ + 516, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19528", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 3", + "[28] Lei Wang, Yi Hu, Jiabang He, Xing Xu, Ning Liu, Hui Liu, and Heng Tao Shen. T-sciq: Teaching multimodal chain-of-thought reasoning via large language model signals for science question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19162-19170, 2024. 1, 3", + "[29] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 3, 6", + "[30] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 4", + "[31] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1", + "[32] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 3", + "[33] Zhuosheng Zhang, Aston Zhang, Mu Li, George Karypis, Alex Smola, et al. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2024. 1, 5, 6", + "[34] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168-5191, 2023. 1, 3, 5, 6" + ], + "bbox": [ + 91, + 90, + 480, + 640 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19529", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Interleaved-Modal Chain-of-Thought/3edf2a1c-2168-41b7-a536-b8e17bc35145_model.json b/2025/Interleaved-Modal Chain-of-Thought/3edf2a1c-2168-41b7-a536-b8e17bc35145_model.json new file mode 100644 index 0000000000000000000000000000000000000000..85762308d97810f16eac3489548299a2d69486e7 --- /dev/null +++ b/2025/Interleaved-Modal Chain-of-Thought/3edf2a1c-2168-41b7-a536-b8e17bc35145_model.json @@ -0,0 +1,2299 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.309, + 0.131, + 0.69, + 0.154 + ], + "angle": 0, + "content": "Interleaved-Modal Chain-of-Thought" + }, + { + "type": "text", + "bbox": [ + 0.309, + 0.18, + 0.687, + 0.2 + ], + "angle": 0, + "content": "Jun Gao\\(^{1}\\), Yongqi Li\\(^{2*}\\), Ziqiang Cao\\(^{1*}\\), Wenjie Li\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.209, + 0.748, + 0.296 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Technology, Soochow University \n\\(^{2}\\)Department of Computer Science, The Hong Kong Polytechnic University \njgao1106@stu.suda.edu.cn, liyongqi0@gmail.com \nzqcao@suda.edu.cn, cswjli@comp.polyu.edu.hk \nhttps://github.com/jungao1106/ICoT" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.331, + 0.327, + 0.347 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.363, + 0.486, + 0.759 + ], + "angle": 0, + "content": "Chain-of-Thought (CoT) prompting elicits large language models (LLMs) to produce a series of intermediate reasoning steps before arriving at the final answer. However, when transitioning to vision-language models (VLMs), their text-only rationales struggle to express the fine-grained associations with the original image. In this paper, we propose an image-incorporated multimodal Chain-of-Thought, named Interleaved-modal Chain-of-Thought (ICoT), which generates sequential reasoning steps consisting of paired visual and textual rationales to infer the final answer. Intuitively, the novel ICoT requires VLMs to enable the generation of fine-grained interleaved-modal content, which is hard for current VLMs to fulfill. Considering that the required visual information is usually part of the input image, we propose Attention-driven Selection (ADS) to realize ICoT over existing VLMs. ADS intelligently inserts regions of the input image to generate the interleaved-modal reasoning steps with ignorable additional latency. ADS relies solely on the attention map of VLMs without the need for parameterization, and therefore it is a plug-and-play strategy that can be generalized to a spectrum of VLMs. We apply ADS to realize ICoT on two popular VLMs of different architectures. Extensive evaluations of three benchmarks have shown that ICoT prompting achieves substantial performance (up to \\(14\\%\\)) and interpretability improvements compared to existing multimodal CoT prompting methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.222, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.814, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Chain-of-Thought (CoT) [31] prompting aims to augment the reasoning capabilities of large language models (LLMs) [4, 8, 22, 26] by eliciting them to produce a sequence of intermediate natural language reasoning steps be" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.906, + 0.408 + ], + "angle": 0, + "content": "fore arriving at the final output. CoT has proven effective in various reasoning tasks, including arithmetic [9], commonsense [17], and symbolic [3], and it has become a potential pathway to advanced artificial intelligence, as depicted in GPT-o1 [20]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.907, + 0.761 + ], + "angle": 0, + "content": "With the development of vision-language models (VLMs), extending CoT prompting into multimodal CoT to improve the reasoning capabilities of VLMs in vision-related tasks becomes increasingly important [19, 28, 33, 34]. The initial multimodal CoT attempts [28, 33] take as input the fused visual and textual embeddings and train language models, such as T5 [23] models, to generate text-only rationales and answers. In the era of VLMs, introducing triple demonstrations composed of an image with the instruction, textual rationales, and the final output (e.g., answer) has proven effective in sparking the reasoning ability of VLMs [6]. Then, related studies focus on improving the linguistic reasoning ability of VLMs. Specifically, DDCoT [34] leverages VLMs to deconstruct problems and resolve them respectively, and CCoT [19] generates scene graphs to prompt VLMs with object and position descriptions. SCAFFOLD [12] overlays a coordinate matrix onto the image to prompt the VLMs with relative visual positions. However, these methods still generate text-only reasoning steps, making it hard to express the fine-grained associations with the origin image exactly. As shown on the left of Figure 1, textual position descriptions, e.g., at the top, are too rough to identify all fruits (orange and banana)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In light of the limitations of text-only rationales, we propose incorporating visual information to enhance the precision of fine-grained associations between generated textual rationales and the corresponding image. We therefore propose an advanced multimodal Chain-of-Thought prompting, named Interleaved-modal Chain-of-thought (ICoT), as shown on the right of Figure 1. ICoT generates multimodal rationales consisting of paired images and textual rationales that formulate interleaved-modal reasoning steps" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.242, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19520" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.141, + 0.097, + 0.242, + 0.112 + ], + "angle": 0, + "content": "Model Input" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.119, + 0.477, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.288, + 0.411, + 0.302 + ], + "angle": 0, + "content": "Question: What types of fruits are in the picture?" + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.317, + 0.282, + 0.332 + ], + "angle": 0, + "content": "Text-only Rationale" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.339, + 0.462, + 0.378 + ], + "angle": 0, + "content": "The mixed variety of fruits suggests that this is a fruit platter. The small, round, light green items at the bottom left appear to be grapes." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.388, + 0.462, + 0.413 + ], + "angle": 0, + "content": "The textured orange skin at the top indicates this is likely an orange, while a banana rests beside it." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.424, + 0.456, + 0.45 + ], + "angle": 0, + "content": "The dark red, smooth-skinned fruit on the right appears to be an apple." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.47, + 0.182, + 0.484 + ], + "angle": 0, + "content": "Answer" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.495, + 0.429, + 0.527 + ], + "angle": 0, + "content": "Therefore, grapes, orange, banana, and apple are in the picture." + }, + { + "type": "title", + "bbox": [ + 0.552, + 0.096, + 0.652, + 0.112 + ], + "angle": 0, + "content": "Model Input" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.12, + 0.888, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.288, + 0.822, + 0.302 + ], + "angle": 0, + "content": "Question: What types of fruits are in the picture?" + }, + { + "type": "title", + "bbox": [ + 0.543, + 0.315, + 0.765, + 0.33 + ], + "angle": 0, + "content": "Interleaved-modal Rationale" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.34, + 0.556, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.563, + 0.339, + 0.87, + 0.377 + ], + "angle": 0, + "content": "The several grapes grouped together and the jumbled assortment of different fruits suggest that this is a fruit platter." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.383, + 0.556, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.563, + 0.388, + 0.869, + 0.415 + ], + "angle": 0, + "content": "The orange colored rind behind an apple indicates that there is an orange and an apple in the picture." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.424, + 0.556, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.564, + 0.424, + 0.856, + 0.449 + ], + "angle": 0, + "content": "The sleek yellow peel off to the right clues at an unpeeled banana." + }, + { + "type": "title", + "bbox": [ + 0.532, + 0.471, + 0.593, + 0.485 + ], + "angle": 0, + "content": "Answer" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.492, + 0.88, + 0.525 + ], + "angle": 0, + "content": "Therefore, grapes, orange, apple, and banana are in the picture." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.545, + 0.907, + 0.601 + ], + "angle": 0, + "content": "Figure 1. The illustration between multimodal CoT with text-only rationales (Left) and interleaved-modal rationales (Right). Green blocks are correct texts used to infer the final answer. Text-only rationales restrict VLMs to use a rough description to indicate the position of objects. Transparent boxes indicate that these regions are selected and inserted to formulate paired visual and textual rationales in ICoT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.674 + ], + "angle": 0, + "content": "to infer the final output. To the best of our knowledge, ICoT is the first multimodal CoT with images incorporated, and it aligns more closely with human thinking processes [5, 21]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.674, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Intuitively, facilitating the novel ICoT is non-trivial, as it introduces challenges for VLMs to support fine-grained interleaved-modal content generation. No current VLMs meet this condition completely. Perceiver-based VLMs such as Qwen2-VL [29] converts images into visual embeddings. Thus, they support fine-grained visual understanding but cannot generate multimodal outputs. Recently proposed unified-modeling VLMs, such as Chameleon [25], UnifiedIO 2 [18], and Emu-3 [30], enable multimodal generation by tokenizing images into discrete tokens. However, on the one hand, unified-modeling VLMs exhibit inertia toward multimodal content generation [7]; on the other hand, the generated images belong to the fixed pre-defined resolution instead of fine granularity." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Since required visual information is usually part of the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.901 + ], + "angle": 0, + "content": "input image for ICoT, we accordingly propose Attention-driven Selection (ADS) to realize ICoT. The basic idea of ADS is to signal VLMs to select patches from the input image rather than generate extra images. At the beginning of generating each textual rationale, ADS inserts a piece of visual tokens of selected patches from the input image to refine the generation of the following textual rationale. Specifically, ADS utilizes the attention map of VLMs to identify optimal patches from the input image as fine-grained visual rationales. Once these fine-grained visual rationales are inserted into the current generation sequence, the VLM resumes the original autoregressive text generation process based on previous multimodal content, formulating paired image and textual rationales to infer the final outputs. Notably, since ADS does not compel VLMs to generate real images, it brings ignorable inference latency compared with previous text-only CoT methods. Additionally, ADS leverages the attention map of VLMs without re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19521" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "quiring parameterization, making it a plug-and-play strategy that can be easily adapted to a wide range of VLMs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.125, + 0.483, + 0.29 + ], + "angle": 0, + "content": "In this paper, we apply ADS to realize ICoT on Chameleon and Qwen2-VL, representing the state-of-the-art unified modeling and perceiver-based VLMs. The results on existing datasets, including \\(\\mathrm{M}^3\\mathrm{CoT}\\) [6], ScienceQA [24], and LLaVA-W [15], indicate that ICoT realized by ADS brings VLMs with substantial performance gains (up to \\(14\\%\\)) compared with current multimodal CoT methods. Additionally, it is noted that the tracked interleaved-modal rationales further enhance the interpretability of the generated results. Our main contribution can be concluded as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.295, + 0.483, + 0.369 + ], + "angle": 0, + "content": "- We propose interleaved-modal CoT, which innovates text-only rationales into multimodal ones to construct clearer reasoning. To our knowledge, we are the first to incorporate images into the intermediate reasoning steps in multimodal CoT." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.37, + 0.483, + 0.43 + ], + "angle": 0, + "content": "- We propose an effective and efficient Attention-driven Selection strategy to facilitate ICoT, which is training-free and widely applicable to VLMs without requiring them to support multimodal generation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.431, + 0.484, + 0.49 + ], + "angle": 0, + "content": "- Experiments demonstrate that our ICoT significantly surpasses existing multimodal CoT methods, proving that the interleaved-modal reasoning process is a foundational innovation in the line of CoT." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.295, + 0.484, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.514, + 0.233, + 0.53 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.542, + 0.383, + 0.559 + ], + "angle": 0, + "content": "2.1. Vision-Language Models (VLMs)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Currently, predominate VLMs such as Qwen-VL [2, 29], BLIP [13], and LLaVA [14-16] are mainly built upon a Large Language Model (LLM), a visual module, and an aligned vision-language adapter. The visual module, e.g., Vision Transformer (ViT) [1], encodes images into dense representations, and then the adapter, e.g., MLP or Q-Former, converts these representations into LLM-readable visual tokens. Finally, visual tokens and textual tokens are fed into the LLM to perform the next-token prediction. This type of VLM can be concluded as Perceiver-LLM architecture, while the Perceiver usually comprises a visual module and the adapter. Additionally, Cambrain-1 [27] introduces more visual modules to collaboratively provide more useful visual tokens in a vision-centric paradigm. In the other research line, unified-modeling VLMs represented by Chameleon [25], Unified-IO 2 [18], and Emu3 [30] are designed to generate texts, images, and so on uniformly. As these models apply codebook [11] to tokenize images into discrete vowels, their training processes are supervised by both vision and text information. Unified-modeling VLMs are expected to develop more stable multimodal understanding abality [10]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.874, + 0.108 + ], + "angle": 0, + "content": "2.2. Multimodal Chain-of-Thought Prompting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.114, + 0.905, + 0.386 + ], + "angle": 0, + "content": "Similar to CoT used in LLMs, multimodal Chain-of-Thought prompting methods [12, 19, 28, 32, 34] aim to augment the reasoning ability of VLM by generating intermediate reasoning steps. A series of studies focus on providing VLMs with fine-grained textual information, such as detailed description [28]. Compositional CoT (CCoT) [19] prompts VLMs to generate a Scene Graph (SG), which is a JSON-like description containing compositional information of objects that occurred in the image. DDCoT [34] deconstructs problems into small problems, requiring VLMs to solve them respectively and then inferring the final answer. In the other research line, Set-of-Marks prompting [32] augments the objects in the image to help VLMs recognize them. SCAFFOLD [12] overlays coordinate onto images to prompt VLMs with relative position information, and VLMs leverage overlayed textual coordinates to implicitly represent corresponding regions of the image to perform reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.905, + 0.478 + ], + "angle": 0, + "content": "However, these methods still produce text-only rationales to infer the final answer. These generated textual rationales usually struggle to express the fine-grained associations with the origin image. We thereby propose ICoT to elicit VLMs to generate interleaved visual-textual reasoning steps to effectively reach the final outputs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.492, + 0.649, + 0.51 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.655 + ], + "angle": 0, + "content": "To address the limitations that current multimodal CoT methods are still stuck in generating text-only rationales to infer the final answer, we propose interleaved-modal CoT (ICoT) to elicit VLMs generated multimodal reasoning steps. We start by introducing the workflow of VLMs and multimodal CoT in Section 3.1. We then introduce the concept of ICoT in Section 3.2. Finally, we propose a plug-and-play method, Attention-driven Selection (ADS), to realize ICoT on existing VLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.655, + 0.68 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.905, + 0.717 + ], + "angle": 0, + "content": "We first recall some background of VLMs and multimodal CoT in this section." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.738, + 0.905, + 0.829 + ], + "angle": 0, + "content": "Vision-Language Model. VLMs usually consist of a visual encoder \\(\\mathbf{E}\\) and a generative large language model LLM, and they determine where to insert images according to visual holders inserted in the text instructions. Then, VLMs take the image and the instructions as input and respond with a final answer" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.842, + 0.905, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\text {a n s w e r} = \\mathbf {V L M} (\\text {I m a g e}, \\text {I n s t r u c t i o n}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Specifically, the visual encoder \\(\\mathbf{E}\\) extracts visual tokens \\(f_{v}^{l\\times d}\\) from the image \\(x_{v}\\), where \\(l\\) is the length of visual" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19522" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.338, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Interleaved-modal CoT" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.11, + 0.484, + 0.143 + ], + "angle": 0, + "content": "1: Input: Word embeddings \\( f_{e} \\), Visual tokens \\( f_{v} \\), Selected number \\( n \\), Signal tokens \\( S \\), Stopping criteria \\( SC \\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.156, + 0.38, + 0.17 + ], + "angle": 0, + "content": "2: Output: Generated Response Answer" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.171, + 0.48, + 0.187 + ], + "angle": 0, + "content": "3: predicted_tokens \\(\\leftarrow\\) [] ▷ Initialize as an empty list" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.187, + 0.175, + 0.201 + ], + "angle": 0, + "content": "4: \\(\\mathcal{V}\\gets f_v\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.202, + 0.483, + 0.231 + ], + "angle": 0, + "content": "5: inputs = Initilize \\((f_e, f_v)\\) Initialize inputs for prefilling" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.232, + 0.264, + 0.245 + ], + "angle": 0, + "content": "6: while \\(SC\\) not met do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.247, + 0.444, + 0.261 + ], + "angle": 0, + "content": "7: next_token, attention_map = model(inputs)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.262, + 0.411, + 0.276 + ], + "angle": 0, + "content": "8: Append next_token to predicted_tokens" + }, + { + "type": "text", + "bbox": [ + 0.357, + 0.278, + 0.478, + 0.292 + ], + "angle": 0, + "content": "\\(\\triangleright\\) ADS judgement" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.293, + 0.339, + 0.306 + ], + "angle": 0, + "content": "9: if predicted_tokens = S then" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.307, + 0.483, + 0.336 + ], + "angle": 0, + "content": "10: \\(\\mathcal{V}_{\\mathrm{selected}} = \\mathrm{ADS}(\\mathcal{V},\\mathrm{attention\\_map},n)\\) Apply Attention-driven Selection" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.338, + 0.405, + 0.352 + ], + "angle": 0, + "content": "11: Append \\(\\nu_{\\mathrm{selected}}\\) to predicted_tokens" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.353, + 0.181, + 0.366 + ], + "angle": 0, + "content": "12: end if" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.368, + 0.484, + 0.397 + ], + "angle": 0, + "content": "13: inputs = Update(inputs, predicted_tokens) > Updates inputs for next step generation" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.399, + 0.19, + 0.411 + ], + "angle": 0, + "content": "14: end while" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.413, + 0.441, + 0.428 + ], + "angle": 0, + "content": "15: Answer = Tokenizer Decode(predicted_tokens)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.429, + 0.229, + 0.441 + ], + "angle": 0, + "content": "16: return Answer" + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.307, + 0.484, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.532 + ], + "angle": 0, + "content": "tokens and \\(d\\) is the dimensions of the hidden states of the LLM. The built-in LLM predicts next-tokens in a left-to-right fashion according to visual tokens \\(f_{v}^{l \\times d}\\) and the instructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.613 + ], + "angle": 0, + "content": "Multimodal CoT. Compared with the direct prediction described in Eqn. 1, Multimodal CoT further introduces a prompt to elicit VLMs to generate a series of intermediate textual rationales \\((r_1, r_2, \\ldots)\\) before the final answer:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.655 + ], + "angle": 0, + "content": "\\(r_1, r_2, \\ldots\\), answer \\(= \\mathbf{VLM}(\\text{Prompt}, \\text{Image}, \\text{Instruction})\\). (2)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Technically, the prompt could be represented as a sequence of demonstrations, each consisting of a triple: (Image, Rationale, Answer). Alternatively, an explicit instruction, such as \"Let's think step by step,\" could also serve as the prompt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.414, + 0.758 + ], + "angle": 0, + "content": "3.2. Interleaved-modal Chain-of-Thought" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Previous multimodal CoT prompting methods only produce text-only reasoning steps to improve the reasoning ability of VLMs. These intermediate steps are generated according to the entire image, which are struggle to express exact fine-grained associations with the original image. Given these limitations, we propose a more advanced Interleaved-modal Chain-of-Thought (ICoT) prompting, aiming to elicit VLMs to generate a series of multimodal intermediate reasoning steps, each consisting of paired im" + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.092, + 0.782, + 0.106 + ], + "angle": 0, + "content": "Algorithm 2 Attention-driven Selection" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.11, + 0.907, + 0.14 + ], + "angle": 0, + "content": "1: Input: Attention map \\( A_{t} \\), Selected number \\( n \\), Visual tokens \\( f_{v} \\)" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.141, + 0.875, + 0.156 + ], + "angle": 0, + "content": "2: Output: Fine-grained visual information \\(\\nu_{\\mathrm{selected}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.157, + 0.906, + 0.171 + ], + "angle": 0, + "content": "3: \\(\\mathcal{V}_{\\mathrm{selected}}\\gets \\emptyset\\) Initialize as an empty set" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.172, + 0.706, + 0.187 + ], + "angle": 0, + "content": "4: Indices \\(\\leftarrow\\) TopK(At,n)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.187, + 0.666, + 0.199 + ], + "angle": 0, + "content": "5: for \\( i \\) in Indices do" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.2, + 0.741, + 0.217 + ], + "angle": 0, + "content": "6: Append \\( f_{v}^{l \\times d}[i] \\) to \\( \\mathcal{V}_{\\text{selected}} \\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.218, + 0.597, + 0.23 + ], + "angle": 0, + "content": "7: end for" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.232, + 0.905, + 0.261 + ], + "angle": 0, + "content": "8: Restore \\((\\mathcal{V}_{\\mathrm{selected}}\\), Indices) \\(>\\) Restore relative positions in-place" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.262, + 0.654, + 0.276 + ], + "angle": 0, + "content": "9: return \\(\\nu_{\\mathrm{selected}}\\)" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.11, + 0.907, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.293, + 0.901, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.451, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Figure 2. The workflow of ADS selecting fine-grained visual rationales. Signal attention represents the attention map of the signal token overall visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.656 + ], + "angle": 0, + "content": "age and textual rationale. Generated intermediate reasoning steps formulating interleaved-modal rationales to effectively lead to the final outputs. In this paper, we consider the visual rationales in interleaved-modal rationales as fine-grained visual information \\( x_{v}^{h' \\times w'} \\) extracted from an image \\( x_{v}^{h \\times w1} \\). These visual rationales capture relevant details in the image, such as objects, colors, and texts, interleaved with the following generated textual rationale to infer the final answer:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.666, + 0.906, + 0.714 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r} r _ {1}, x _ {v _ {1}}, r _ {2}, x _ {v _ {2}}, \\dots , \\text {a n s w e r} = \\mathbf {V L M} (\\text {P r o m p t , I m a g e}, \\\\ \\text {I n s t r u c t i o n}). \\end{array} \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.724, + 0.753, + 0.739 + ], + "angle": 0, + "content": "3.3. Attention-driven Selection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Although the proposed ICoT is both novel and conceptually sound, current VLMs are unable to generate such fine-grained visual information. This remains true even for VLMs [18, 25, 30] that are empowered with multimodal generation ability. We thus propose to simplify the problem from fine-grained visual information generation to fine-grained visual information selection, as this information" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.907, + 0.901 + ], + "angle": 0, + "content": "1The images in the dataset involved in this paper are RGB images by default, and the number of channels is omitted in the formulas for simplicity." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19523" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "has been naturally contained in the origin image, namely \\(x_{v}^{h^{\\prime}\\times w^{\\prime}}\\in x_{v}^{h\\times w}\\) where \\(h^\\prime \\ll h\\) and \\(w^{\\prime}\\ll w\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.126, + 0.484, + 0.277 + ], + "angle": 0, + "content": "Specifically, before performing next-token prediction, ICoT requires the VLM to cache visual tokens \\( f_{v}^{l \\times d} \\) extracted by its built-in visual encoder \\( \\mathbf{E}(x_{v}) \\) for further selection. In the following decoding steps, we consider the VLM deems it necessary to insert a piece of visual rationales after generating a pre-defined signal token \\( S \\) as shown in Figure 2, which is a natural language token that indicates the beginning of a textual rationale. Therefore, ADS will be signaled to select fine-grained visual information from \\( f_{v}^{l \\times d} \\) upon the VLM generating \\( S \\):" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.295, + 0.484, + 0.347 + ], + "angle": 0, + "content": "\\[\n\\text {d o} \\text {s e l e c t i o n} = \\left\\{ \\begin{array}{l l} \\text {T r u e} & \\text {i f p r e d i c t e d} \\\\ \\text {F a l s e} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.347, + 0.484, + 0.395 + ], + "angle": 0, + "content": "Then, ADS selects \\( n \\) visual tokens from \\( f_{v}^{l \\times d} \\) as fine-grained visual information according to the signal attention map \\( A_{t} \\) at the current decoding step \\( t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.412, + 0.483, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} _ {\\text {s e l e c t e d}} = \\left\\{f _ {v} [ i ] \\mid i \\in \\operatorname {T o p K} \\left(A _ {t}, n\\right) \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.614 + ], + "angle": 0, + "content": "where \\( A_{t} \\) is obtained by averaging the attention map between the signal token and visual tokens across all VLM layers. Up to now, the selected visual tokens are sorted by their attention scores, and we subsequently restore the relative position of \\( f_{v}[i] \\) in the origin image in place, prioritizing rows. Once fine-grained visual tokens \\( \\nu_{\\mathrm{selected}} \\) are obtained, VLMs will take as input the concatenated \\( \\nu_{\\mathrm{selected}} \\) and predicted_tokens, denote as Cat(predicted_tokens, \\( \\nu_{\\mathrm{selected}} \\)), resuming the original autoregressive text generation process. Hence, the current decoding step is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.633, + 0.484, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r} \\text {n e x t - t o k e n} = \\mathbf {V L M} (\\text {C a t (p r e d i c t e d - t o k e n s}, \\\\ \\nu_ {\\text {s e l e c t e d}})). \\end{array} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.776 + ], + "angle": 0, + "content": "Notably, to avoid misunderstanding, we first convert predicted_tokens into word embeddings \\( f_{e} \\), and then concatenate \\( f_{e} \\) with selected fine-grained visual information \\( \\mathcal{V}_{\\mathrm{selected}} \\) in the embedding-level. We provide a detailed description of ICoT and ADS in Algorithm 1 and Algorithm 2, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Technically, ICoT inherits the existing eliciting methods from CoT, such as attaching an instruction: \"Let's think step by step\" in zero-shot ICoT or providing few-shot examples. In the few-shot ICoT, each example consists of an image, textual rationales, and visual rationales. These examples can be manually designed to prompt VLMs regarding their way of thinking and generation formatting, among other aspects." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.117, + 0.616, + 0.132 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.143, + 0.905, + 0.265 + ], + "angle": 0, + "content": "\\(\\mathbf{M}^{3}\\mathbf{C}\\mathbf{o}\\mathbf{T}\\) [6] is a novel multimodal CoT benchmark specifically concentrated on multi-domain, multi-reasoning-step. \\(\\mathbf{M}^{3}\\mathbf{C}\\mathbf{o}\\mathbf{T}\\) contains 267 categories from science, mathematics, and commonsense domains. As the question of each instance is relatively complex, their rationales have an average length of 293 tokens and rely more on fine-grained visual information, which can reflect the advantages of ICoT compared with previous multimodal CoT methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.267, + 0.905, + 0.327 + ], + "angle": 0, + "content": "ScienceQA [24] is a popular dataset used to evaluate the reasoning ability of VLMs. We use ScienceQA to provide a general comparison between ICoT and other existing multimodal CoT methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.33, + 0.905, + 0.406 + ], + "angle": 0, + "content": "LLaVA-Bench In-the-Wild (LLaVA-W) [15] evaluates VLMs' ability to respond to visual questions with detailed long-form answers, which also focus on the fine-grained visual description. The reference label of each instance is produced by GPT-4v." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.424, + 0.622, + 0.438 + ], + "angle": 0, + "content": "4.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.45, + 0.905, + 0.51 + ], + "angle": 0, + "content": "No-CoT responds to the current input image and question directly without further prompting. The few-shot demonstrations of the direct generation mode consist of (Image, Question, Answer)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.513, + 0.905, + 0.558 + ], + "angle": 0, + "content": "Multimodal CoT [33] elicits VLMs to generate a series of text-only intermediate reasoning steps to infer the final outputs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.561, + 0.906, + 0.667 + ], + "angle": 0, + "content": "CCoT [19] first generates a scene graph (SG) using the VLM itself and then uses that SG in the prompt to produce a response. The SG is a JSON-like structural description of the given image with extensive compositional information of objects in the current images. Following their settings, we apply their official prompt to prompt VLMs to generate SGs and answers respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.669, + 0.906, + 0.759 + ], + "angle": 0, + "content": "DDCoT [34] first prompts LLM to deconstruct the input question into a sequence of basic sub-questions and then applies a VQA model to answer these sub-questions involving visual information. In this paper, we use the VLM that plays the role of LLM in DDCoT for a fair comparison, as their original LLM is ChatGPT." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.762, + 0.906, + 0.868 + ], + "angle": 0, + "content": "SCAFFOLD [12] overlays a coordinate matrix onto the input image, exactly demonstrating relative visual positions for VLMs. During reasoning, VLMs are steered to utilize these coordinates that indicate fine-grained visual information in the image to solve different vision-language tasks. We use their released scripts to add coordinates over each image and then use their official prompt to elicit VLMs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Specifically, in the few-shot scenario, the demonstrations of these baselines are human-written, aligning with ICoT." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19524" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.088, + 0.905, + 0.4 + ], + "angle": 0, + "content": "
BackboneMethods0-shot1-shot
M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑
Chameleon-7BNo-CoT29.147.713.128.448.523.9
Multimodal CoT [33]28.549.020.430.650.720.6
CCoT [19]29.450.222.131.451.324.5
DDCoT [34]28.649.820.229.849.223.1
SCAFFOLD [12]29.648.521.731.147.524.7
ICoT (Ours)29.851.025.232.353.427.6
% Improve0.6%1.6%14.0%2.8%4.0%11.7%
BackboneMethods0-shot1-shot
M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑
Qwen2-VL-7BNo-CoT43.656.332.745.464.433.5
Multimodal CoT [33]40.151.330.742.558.331.4
CCoT [19]43.356.429.444.163.833.9
DDCoT [34]42.655.231.245.764.932.8
SCAFFOLD [12]41.753.731.844.962.533.1
ICoT (Ours)44.156.834.246.065.435.7
% Improve1.1%0.7%4.6%0.6%0.7%5.3%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.409, + 0.907, + 0.452 + ], + "angle": 0, + "content": "Table 1. Results of ICoT and baselines based on Chameleon and Qwen2-VL, with the highest score bold. \\( {\\mathrm{M}}^{3}\\mathrm{{CoT}} \\) and ScienceQA are evaluated by accuracy, and we report the ROUGE-L score for the LLaVA-W benchmark. % improve represents the relative improvement achieved by ICoT over the previously best baseline." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.477, + 0.268, + 0.494 + ], + "angle": 0, + "content": "4.3. Implement Details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.506, + 0.483, + 0.656 + ], + "angle": 0, + "content": "We apply ICoT over Chameleon-7B [25] and Qwen2-VL7B-Instruct [29], which represents the fine-grained visual information in the form of discrete vokens and dense features. All experiments are conducted on A800 GPUs, and we evaluate ICoT under both zero- and one-shot scenarios. During generating interleaved-modal rationales, the signal token \\( S \\) used to trigger ADS is set to line break, i.e., \"\\(\\backslash n\\)\", by default, which semantically and empirically indicates the end of a generated rationale and the beginning of the next one." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.663, + 0.483, + 0.844 + ], + "angle": 0, + "content": "VLMs insert visual tokens selected by ADS following the special token at the granularity of 64 according to posterior results shown in Table 4 of Appendix 8. Notably, to shorten the representation of an image, Qwen2-VL introduces a novel merge mechanism, and we approximately consider its patch size to be \\((28 \\times 28)\\). Each patch of Qwen2-VL has approximately 4 times as many pixels as a Chameleon patch \\((16 \\times 16)\\), which results in practical selection numbers of ADS set to 16. Considering the work of ADS requires the inner attention map, we apply the \"eager\" attention on both Chameleon-7B and Qwen2-VL, limited to the dependency of related python libraries2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.477, + 0.652, + 0.492 + ], + "angle": 0, + "content": "4.4. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.504, + 0.907, + 0.806 + ], + "angle": 0, + "content": "We comprehensively evaluate the performance of ICoT on top of Chameleon-7B and Qwen2-VL-7B through \\(\\mathbf{M}^3\\mathrm{CoT}\\), ScinceQA, and LLaVA-W in Table 1. In 0-shot settings, ICoT outperforms all baselines, including direct generation (No-CoT), CoT, CCoT, DDoT, and SCAFFOLD. Specifically, ICoT distinguishes from Multimodal CoT in terms of the modality of reasoning steps, which exhibit the advantages of interleaved-modal rationales to infer the final answer effectively. Compared with other multimodal CoT methods, the performance gains of ICoT further indicate that interleaved-modal rationales are more reasonable in intuition and effect than plainly inserted scene graphs (CCoT) and deconstructed sub-questions (DDoS). In 1-shot settings, ICoT demonstrations contain manually selected fine-grained visual information, while their text rationales are kept the same as other baselines. The performance gains compared with 0-shot ICoT indicate that our manually designed fine-grained ICoT demonstrations potentially guide VLMs to think in this format. In Table 4, we rigorously ablate the effectiveness of fine-grained ICoT demonstrations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Additionally, ICoT achieves the most relative performance gains in the LLaVA-W benchmark as the reference labels contain details sourced from images. These substantial performance gains compared with other baselines prove that visual tokens selected by ADS effectively capture the fine-grained visual information of an image, aiding the gen" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\(^{2}\\)Using attn_implementation=\"eager\" when loading the model from HuggingFace." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19525" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.343, + 0.108 + ], + "angle": 0, + "content": "eration of high-quality text rationales." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.117, + 0.245, + 0.133 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.141, + 0.483, + 0.382 + ], + "angle": 0, + "content": "We ablate ICoT to verify the effectiveness of each portion across three benchmarks in Table 2 with the following settings: (1). w/o ADS: VLMs generate text-only rationales. (2). w/o FVI: Patches inserted in the demonstration are randomly sampled. Results indicate that both ADS and fine-grained visual information (FVI) incorporated in the demonstration are necessary. In particular, interleaved-modal rationales exhibit substantial advantages in generating high-quality textual rationales compared with text-only rationales (w/o ADS). When substituting ICoT demonstrations with normal ones (w/o FVI), the performance degradation proves that fine-grained visual information in demonstrations effectively guides VLMs to think in this format. Compared with the performance difference between removing ADS and FVI, we find that generating paired visual and text rationales boosts more improvements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.483, + 0.49 + ], + "angle": 0, + "content": "Additionally, the performance gap is relatively smooth in ScienceQA and more dramatic on \\(\\mathrm{M}^3\\mathrm{CoT}\\) and LLaVA-W. We attribute this to the ScienceQA dataset being relatively easier than others since both \\(\\mathrm{M}^3\\mathrm{CoT}\\) and the answers of LLaVA-W highly rely on the fine-grained visual information of an image. Therefore, our proposed ICoT has the potential to solve complex vision-language tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.5, + 0.265, + 0.515 + ], + "angle": 0, + "content": "4.6. In-depth Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Analysis on realizing ICoT via KV Cache Up to now, the fine-grained visual information of ICoT is provided at the input end via discrete tokens or dense visual tokens, which brings more computation. After rethinking the generating process of an autoregressive model, there are other inputs that are proposed to avoid repeated computation, namely, the Key-Value (KV) Cache. The input image was stored in the KV Cache during the prefilling phase before generating multimodal intermediate reasoning steps in a left-to-right fashion. Therefore, copying the KV cache of fine-grained visual information enables ICoT with reduced computational costs, as visual information does not require extra forward propagation. As shown in Table 3, copying the KV Cache brings performance degradation compared with providing visual information at the input end. We attribute this phenomenon to the fact that although copying KV Cache indeed makes VLMs attend more to the same region as ADS, the optimal visual information is highlighted in a position-agnostic case, determined by the nature of KV Cache. Specifically, this degrades the original interleaved-modal rationales into non-interleaved ones as position information is early fused into KV Cache, and thus the copied ones are inherently insensitive to the position of textual rationale." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "However, considering it brought slight performance" + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.091, + 0.903, + 0.169 + ], + "angle": 0, + "content": "
MethodsM3CoTScienceQALLaVA-W
ICoT32.353.427.6
w/o ADS29.2 (-3.1)52.4(-1.0)24.5(-3.1)
w/o FVI30.6 (-1.8)52.8(-0.6)25.9(-1.7)
w/o ADS+FVI29.1(-3.2)51.0(-2.4)23.0(-4.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.183, + 0.907, + 0.253 + ], + "angle": 0, + "content": "Table 2. Ablation studies of 1-shot ICoT on Chameleon-7B. () describes the performance degradation compared with ICoT. FVI indicates the 1-shot demonstration contains fine-grained visual information. ADS indicates that VLMs generate interleaved-modal reasoning steps." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.269, + 0.903, + 0.352 + ], + "angle": 0, + "content": "
Dataset0-shot1-shot
KV-CopyICoTKV-CopyICoT
M3CoT29.129.831.532.3
ScienceQA49.751.052.953.4
LLaVA-W24.725.227.027.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.366, + 0.906, + 0.394 + ], + "angle": 0, + "content": "Table 3. Results comparison between copying KV Cache (KV-Copy) and inserting selected patches." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.41, + 0.903, + 0.46 + ], + "angle": 0, + "content": "
MethodsM3CoTScienceQALLaVA-W
Human-written32.353.427.6
Model-written31.5(-0.8)51.8(-1.6)26.7(-0.8)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.473, + 0.906, + 0.556 + ], + "angle": 0, + "content": "Table 4. Results concerning the design of demonstrations on Chameleon-7B. () describes the performance degradation compared with ICoT. Human-written indicates that demonstrations are manually designed with fine-grained visual information inserted, and Model-written indicates that the VLM generates both visual and textual rationale via ICoT." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.631 + ], + "angle": 0, + "content": "degradation and factually reduced computation costs, we believe this exploration is still valuable, and we call for more interesting exploration in realizing ICoT." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.831 + ], + "angle": 0, + "content": "Analysis on the Demonstrations We also attempt to let VLMs generate the demonstrations via themselves (Automatic at the bottom of Table 4). Results indicate that using the automatically generated demonstrations also brings performance degradation compared with ICoT using manually designed ones. We consider it is caused by the fact that formulating a continuous sub-image through ADS is nontrivial, and some discrete patches inevitably introduce additional noise. Therefore, considering that designing such a demonstration is not time-consuming, ICoT utilizes manually designed ones to elicit VLMs to perform ICoT for better performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.634, + 0.862 + ], + "angle": 0, + "content": "5. Case Study" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this section, we empirically illustrate the advantages of ICoT via three case studies in Figure 3. These case studies" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19526" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.898, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.322, + 0.908, + 0.365 + ], + "angle": 0, + "content": "Figure 3. Case studies between ICoT and multimodal CoT with text-only rationales on Chameleon. Three cases are selected according to three typical problems in text-only problems: misunderstanding, overgeneralization, and hallucination. Red blocks indicate the incorrect rationales." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.483, + 0.437 + ], + "angle": 0, + "content": "focused on three typical problems that occurred in text-only rationales, namely, misunderstanding (top), overgeneralization (middle), and hallucination (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.79 + ], + "angle": 0, + "content": "Specifically, in the first case, interleaved-modal CoT first recognizes three different objects via captions: \"inflatable castle, crayons, and a parachute\". Then, in the second reasoning step, ADS inserts selected patches from the scheduled objects to elicit the VLM to conclude their common property, and VLM infers a correct answer. Text-only CoT misunderstands that the three objects are all colored pencils, ignoring the castle and the parachute, even the final answer is correct. In the second case, text-only rationales overgeneralize flying a kite to a kite festival, leading to a wrong answer. ICoT first recognizes a man standing in a field of grass and then infers it is a windy day according to a few kites in the sky. In the last case, it provides the other typical error of text-only CoT, namely, hallucination. As text-only CoT purely relies on language reasoning ability, VLMs have the potential to imagine something not mentioned in the image, resulting in a wrong answer. ICoT first infers from the street sign that there may be a troll attraction nearby according to patches of the indicator inserted by ADS. Then, the ADS helps the VLMs to attend to the troll statue under the bridge and infer it is likely the mentioned attraction, finally arriving at the correct answer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Even though the above case studies exhibit the advantages of ICoT, ADS still brings potential problems. For example, ADS is triggered to select patches when VLM generates a pre-defined signal token. This simple mechanism is a double-edged sword in that VLMs will generate low-quality responses if this token is generated with a high frequency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.39, + 0.634, + 0.406 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.697 + ], + "angle": 0, + "content": "In this paper, we first propose interleaved-modal CoT (ICoT), which generates interleaved-modal rationales to infer the final answer effectively. In light of the challenges of applying ICoT on existing VLMs, we then introduce Attention-driven Selection (ADS), a plug-and-play strategy to identify optimal patches from the image without being parameterized. We evaluate ICoT on Chameleon-7B and Qwen2-VL-7B-Instruct, representing VLMs of two architectures. Extensive experiments conducted on \\(\\mathbf{M}^3\\mathrm{CoT}\\), ScienceQA, and LLaVA-W, under both zero- and few-shot scenarios, have proven that ICoT achieves substantial performance (up to \\(14\\%\\)) compared with the existing multimodal CoT methods. Additionally, in the analysis section, we conduct a preliminary exploration of implementing ICoT by copying the KV cache of optimal visual tokens and explain the inner trade-off between efficiency and performance in this approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Although ICoT has proven its effectiveness in this paper, we consider that ICoT still has significant potential for further improvement. The patch selection in ADS requires storing attention scores, which brings additional memory overhead. Moreover, the fixed number of selected patches in the ADS design is sub-optimal, resulting in unexpected outputs for VLMs. To address these, we intend to incorporate established techniques from segmentation or grounding methods to create a more robust implementation of ICoT. In the future, we also plan to evaluate it across additional backbones and benchmarks to better assess its generalization ability." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19527" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.207 + ], + "angle": 0, + "content": "I would like to express my sincere gratitude to all the authors and reviewers for their valuable contributions to this research. This work was supported by the National Natural Science Foundation of China (NSFC 62106165) and the Project Funded by the Priority Academic Program Development of Jiangsu Higher Education Institutions, China." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.221, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.29, + 0.484, + 0.359 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.402 + ], + "angle": 0, + "content": "[3] BIG bench authors. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on Machine Learning Research, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.404, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.434, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[5] Claus Bundesen. A theory of visual attention. Psychological review, 97(4):523, 1990. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.462, + 0.483, + 0.518 + ], + "angle": 0, + "content": "[6] Qiguang Chen, Libo Qin, Jin Zhang, Zhi Chen, Xiao Xu, and Wanxiang Che. M3cot: A novel benchmark for multidomain multi-step multi-modal chain-of-thought. arXiv preprint arXiv:2405.16473, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.52, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[7] Ethan Chern, Jiadi Su, Yan Ma, and Pengfei Liu. Anole: An open, autoregressive, native large multimodal models for interleaved image-text generation. arXiv preprint arXiv:2407.06135, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[9] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[10] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. arXiv preprint arXiv:2406.11832, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Xuanyu Lei, Zonghan Yang, Xinrui Chen, Peng Li, and Yang Liu. Scaffolding coordinates to promote vision-language coordination in large multi-modal models. arXiv preprint arXiv:2402.12058, 2024. 1, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.246, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[14] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.237 + ], + "angle": 0, + "content": "[15] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.3,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.239, + 0.905, + 0.28 + ], + "angle": 0, + "content": "[16] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.905, + 0.338 + ], + "angle": 0, + "content": "[17] Jiacheng Liu, Alisa Liu, Ximing Lu, Sean Welleck, Peter West, Ronan Le Bras, Yejin Choi, and Hannaneh Hajishirzi. Generated knowledge prompting for commonsense reasoning. arXiv preprint arXiv:2110.08387, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.34, + 0.906, + 0.423 + ], + "angle": 0, + "content": "[18] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26439-26455, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.425, + 0.905, + 0.494 + ], + "angle": 0, + "content": "[19] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.829, + 0.511 + ], + "angle": 0, + "content": "[20] OpenAI. Learning to reason with llms, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.513, + 0.905, + 0.554 + ], + "angle": 0, + "content": "[21] Michael I Posner, Steven E Petersen, et al. The attention system of the human brain. Annual review of neuroscience, 13(1):25-42, 1990. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.557, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[22] Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.629, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[23] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.7, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[24] Tanik Saikh, Tirthankar Ghosal, Amish Mittal, Asif Ekbal, and Pushpak Bhattacharyya. Scienceqa: A novel resource for question answering on scholarly articles. International Journal on Digital Libraries, 23(3):289-301, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.758, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[25] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2,3,4,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[26] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[27] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19528" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[28] Lei Wang, Yi Hu, Jiabang He, Xing Xu, Ning Liu, Hui Liu, and Heng Tao Shen. T-sciq: Teaching multimodal chain-of-thought reasoning via large language model signals for science question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19162-19170, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[29] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.318, + 0.482, + 0.386 + ], + "angle": 0, + "content": "[30] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.388, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[31] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.459, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[32] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.516, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[33] Zhuosheng Zhang, Aston Zhang, Mu Li, George Karypis, Alex Smola, et al. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2024. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[34] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168-5191, 2023. 1, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19529" + } + ] +] \ No newline at end of file diff --git a/2025/Interleaved-Modal Chain-of-Thought/3edf2a1c-2168-41b7-a536-b8e17bc35145_origin.pdf b/2025/Interleaved-Modal Chain-of-Thought/3edf2a1c-2168-41b7-a536-b8e17bc35145_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4056a606471b54cfe4eb4f0a6173b6621d7c365c --- /dev/null +++ b/2025/Interleaved-Modal Chain-of-Thought/3edf2a1c-2168-41b7-a536-b8e17bc35145_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bf5e6afb3d99e00dc8874759f0b5678d8534b57ef5cb803acdf18e809cd768a +size 921355 diff --git a/2025/Interleaved-Modal Chain-of-Thought/full.md b/2025/Interleaved-Modal Chain-of-Thought/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bab5b5a3c89aa008ed8a56dfd2516ebc1fd45b87 --- /dev/null +++ b/2025/Interleaved-Modal Chain-of-Thought/full.md @@ -0,0 +1,332 @@ +# Interleaved-Modal Chain-of-Thought + +Jun Gao $^{1}$ , Yongqi Li $^{2*}$ , Ziqiang Cao $^{1*}$ , Wenjie Li $^{2}$ + +$^{1}$ School of Computer Science and Technology, Soochow University + $^{2}$ Department of Computer Science, The Hong Kong Polytechnic University +jgao1106@stu.suda.edu.cn, liyongqi0@gmail.com +zqcao@suda.edu.cn, cswjli@comp.polyu.edu.hk +https://github.com/jungao1106/ICoT + +# Abstract + +Chain-of-Thought (CoT) prompting elicits large language models (LLMs) to produce a series of intermediate reasoning steps before arriving at the final answer. However, when transitioning to vision-language models (VLMs), their text-only rationales struggle to express the fine-grained associations with the original image. In this paper, we propose an image-incorporated multimodal Chain-of-Thought, named Interleaved-modal Chain-of-Thought (ICoT), which generates sequential reasoning steps consisting of paired visual and textual rationales to infer the final answer. Intuitively, the novel ICoT requires VLMs to enable the generation of fine-grained interleaved-modal content, which is hard for current VLMs to fulfill. Considering that the required visual information is usually part of the input image, we propose Attention-driven Selection (ADS) to realize ICoT over existing VLMs. ADS intelligently inserts regions of the input image to generate the interleaved-modal reasoning steps with ignorable additional latency. ADS relies solely on the attention map of VLMs without the need for parameterization, and therefore it is a plug-and-play strategy that can be generalized to a spectrum of VLMs. We apply ADS to realize ICoT on two popular VLMs of different architectures. Extensive evaluations of three benchmarks have shown that ICoT prompting achieves substantial performance (up to $14\%$ ) and interpretability improvements compared to existing multimodal CoT prompting methods. + +# 1. Introduction + +Chain-of-Thought (CoT) [31] prompting aims to augment the reasoning capabilities of large language models (LLMs) [4, 8, 22, 26] by eliciting them to produce a sequence of intermediate natural language reasoning steps be + +fore arriving at the final output. CoT has proven effective in various reasoning tasks, including arithmetic [9], commonsense [17], and symbolic [3], and it has become a potential pathway to advanced artificial intelligence, as depicted in GPT-o1 [20]. + +With the development of vision-language models (VLMs), extending CoT prompting into multimodal CoT to improve the reasoning capabilities of VLMs in vision-related tasks becomes increasingly important [19, 28, 33, 34]. The initial multimodal CoT attempts [28, 33] take as input the fused visual and textual embeddings and train language models, such as T5 [23] models, to generate text-only rationales and answers. In the era of VLMs, introducing triple demonstrations composed of an image with the instruction, textual rationales, and the final output (e.g., answer) has proven effective in sparking the reasoning ability of VLMs [6]. Then, related studies focus on improving the linguistic reasoning ability of VLMs. Specifically, DDCoT [34] leverages VLMs to deconstruct problems and resolve them respectively, and CCoT [19] generates scene graphs to prompt VLMs with object and position descriptions. SCAFFOLD [12] overlays a coordinate matrix onto the image to prompt the VLMs with relative visual positions. However, these methods still generate text-only reasoning steps, making it hard to express the fine-grained associations with the origin image exactly. As shown on the left of Figure 1, textual position descriptions, e.g., at the top, are too rough to identify all fruits (orange and banana). + +In light of the limitations of text-only rationales, we propose incorporating visual information to enhance the precision of fine-grained associations between generated textual rationales and the corresponding image. We therefore propose an advanced multimodal Chain-of-Thought prompting, named Interleaved-modal Chain-of-thought (ICoT), as shown on the right of Figure 1. ICoT generates multimodal rationales consisting of paired images and textual rationales that formulate interleaved-modal reasoning steps + +# Model Input + +![](images/4b546465888409e42344389e068077ff428c3a1db8ac1922234d281e1bdcf4c1.jpg) +Question: What types of fruits are in the picture? + +# Text-only Rationale + +The mixed variety of fruits suggests that this is a fruit platter. The small, round, light green items at the bottom left appear to be grapes. + +The textured orange skin at the top indicates this is likely an orange, while a banana rests beside it. + +The dark red, smooth-skinned fruit on the right appears to be an apple. + +# Answer + +Therefore, grapes, orange, banana, and apple are in the picture. + +# Model Input + +![](images/0f38bad4fa19e33ce5395ce9331e107b56f82c06456e96216a6dfbdab1643d62.jpg) +Question: What types of fruits are in the picture? + +# Interleaved-modal Rationale + +![](images/ac007bed09b3a8ba675eaa1a0efac5fd8310a9ca4bd9c8569a4681a7984d6dae.jpg) + +The several grapes grouped together and the jumbled assortment of different fruits suggest that this is a fruit platter. + +![](images/f17f9a0c5296e00e97c2043e6c3152b8b8155b790bb15240a94105058f415c96.jpg) + +The orange colored rind behind an apple indicates that there is an orange and an apple in the picture. + +![](images/0c05d7aaf6c67afc1cbfb92e9d124c53430002ab075e07bdb6ef19cf6f839e30.jpg) +Figure 1. The illustration between multimodal CoT with text-only rationales (Left) and interleaved-modal rationales (Right). Green blocks are correct texts used to infer the final answer. Text-only rationales restrict VLMs to use a rough description to indicate the position of objects. Transparent boxes indicate that these regions are selected and inserted to formulate paired visual and textual rationales in ICoT. + +The sleek yellow peel off to the right clues at an unpeeled banana. + +# Answer + +Therefore, grapes, orange, apple, and banana are in the picture. + +to infer the final output. To the best of our knowledge, ICoT is the first multimodal CoT with images incorporated, and it aligns more closely with human thinking processes [5, 21]. + +Intuitively, facilitating the novel ICoT is non-trivial, as it introduces challenges for VLMs to support fine-grained interleaved-modal content generation. No current VLMs meet this condition completely. Perceiver-based VLMs such as Qwen2-VL [29] converts images into visual embeddings. Thus, they support fine-grained visual understanding but cannot generate multimodal outputs. Recently proposed unified-modeling VLMs, such as Chameleon [25], UnifiedIO 2 [18], and Emu-3 [30], enable multimodal generation by tokenizing images into discrete tokens. However, on the one hand, unified-modeling VLMs exhibit inertia toward multimodal content generation [7]; on the other hand, the generated images belong to the fixed pre-defined resolution instead of fine granularity. + +Since required visual information is usually part of the + +input image for ICoT, we accordingly propose Attention-driven Selection (ADS) to realize ICoT. The basic idea of ADS is to signal VLMs to select patches from the input image rather than generate extra images. At the beginning of generating each textual rationale, ADS inserts a piece of visual tokens of selected patches from the input image to refine the generation of the following textual rationale. Specifically, ADS utilizes the attention map of VLMs to identify optimal patches from the input image as fine-grained visual rationales. Once these fine-grained visual rationales are inserted into the current generation sequence, the VLM resumes the original autoregressive text generation process based on previous multimodal content, formulating paired image and textual rationales to infer the final outputs. Notably, since ADS does not compel VLMs to generate real images, it brings ignorable inference latency compared with previous text-only CoT methods. Additionally, ADS leverages the attention map of VLMs without re + +quiring parameterization, making it a plug-and-play strategy that can be easily adapted to a wide range of VLMs. + +In this paper, we apply ADS to realize ICoT on Chameleon and Qwen2-VL, representing the state-of-the-art unified modeling and perceiver-based VLMs. The results on existing datasets, including $\mathrm{M}^3\mathrm{CoT}$ [6], ScienceQA [24], and LLaVA-W [15], indicate that ICoT realized by ADS brings VLMs with substantial performance gains (up to $14\%$ ) compared with current multimodal CoT methods. Additionally, it is noted that the tracked interleaved-modal rationales further enhance the interpretability of the generated results. Our main contribution can be concluded as follows: + +- We propose interleaved-modal CoT, which innovates text-only rationales into multimodal ones to construct clearer reasoning. To our knowledge, we are the first to incorporate images into the intermediate reasoning steps in multimodal CoT. +- We propose an effective and efficient Attention-driven Selection strategy to facilitate ICoT, which is training-free and widely applicable to VLMs without requiring them to support multimodal generation. +- Experiments demonstrate that our ICoT significantly surpasses existing multimodal CoT methods, proving that the interleaved-modal reasoning process is a foundational innovation in the line of CoT. + +# 2. Related Work + +# 2.1. Vision-Language Models (VLMs) + +Currently, predominate VLMs such as Qwen-VL [2, 29], BLIP [13], and LLaVA [14-16] are mainly built upon a Large Language Model (LLM), a visual module, and an aligned vision-language adapter. The visual module, e.g., Vision Transformer (ViT) [1], encodes images into dense representations, and then the adapter, e.g., MLP or Q-Former, converts these representations into LLM-readable visual tokens. Finally, visual tokens and textual tokens are fed into the LLM to perform the next-token prediction. This type of VLM can be concluded as Perceiver-LLM architecture, while the Perceiver usually comprises a visual module and the adapter. Additionally, Cambrain-1 [27] introduces more visual modules to collaboratively provide more useful visual tokens in a vision-centric paradigm. In the other research line, unified-modeling VLMs represented by Chameleon [25], Unified-IO 2 [18], and Emu3 [30] are designed to generate texts, images, and so on uniformly. As these models apply codebook [11] to tokenize images into discrete vowels, their training processes are supervised by both vision and text information. Unified-modeling VLMs are expected to develop more stable multimodal understanding abality [10]. + +# 2.2. Multimodal Chain-of-Thought Prompting + +Similar to CoT used in LLMs, multimodal Chain-of-Thought prompting methods [12, 19, 28, 32, 34] aim to augment the reasoning ability of VLM by generating intermediate reasoning steps. A series of studies focus on providing VLMs with fine-grained textual information, such as detailed description [28]. Compositional CoT (CCoT) [19] prompts VLMs to generate a Scene Graph (SG), which is a JSON-like description containing compositional information of objects that occurred in the image. DDCoT [34] deconstructs problems into small problems, requiring VLMs to solve them respectively and then inferring the final answer. In the other research line, Set-of-Marks prompting [32] augments the objects in the image to help VLMs recognize them. SCAFFOLD [12] overlays coordinate onto images to prompt VLMs with relative position information, and VLMs leverage overlayed textual coordinates to implicitly represent corresponding regions of the image to perform reasoning. + +However, these methods still produce text-only rationales to infer the final answer. These generated textual rationales usually struggle to express the fine-grained associations with the origin image. We thereby propose ICoT to elicit VLMs to generate interleaved visual-textual reasoning steps to effectively reach the final outputs. + +# 3. Methodology + +To address the limitations that current multimodal CoT methods are still stuck in generating text-only rationales to infer the final answer, we propose interleaved-modal CoT (ICoT) to elicit VLMs generated multimodal reasoning steps. We start by introducing the workflow of VLMs and multimodal CoT in Section 3.1. We then introduce the concept of ICoT in Section 3.2. Finally, we propose a plug-and-play method, Attention-driven Selection (ADS), to realize ICoT on existing VLMs. + +# 3.1. Preliminaries + +We first recall some background of VLMs and multimodal CoT in this section. + +Vision-Language Model. VLMs usually consist of a visual encoder $\mathbf{E}$ and a generative large language model LLM, and they determine where to insert images according to visual holders inserted in the text instructions. Then, VLMs take the image and the instructions as input and respond with a final answer + +$$ +\text {a n s w e r} = \mathbf {V L M} (\text {I m a g e}, \text {I n s t r u c t i o n}). \tag {1} +$$ + +Specifically, the visual encoder $\mathbf{E}$ extracts visual tokens $f_{v}^{l\times d}$ from the image $x_{v}$ , where $l$ is the length of visual + +# Algorithm 1 Interleaved-modal CoT + +1: Input: Word embeddings $f_{e}$ , Visual tokens $f_{v}$ , Selected number $n$ , Signal tokens $S$ , Stopping criteria $SC$ + +2: Output: Generated Response Answer + +3: predicted_tokens $\leftarrow$ [] ▷ Initialize as an empty list + +4: $\mathcal{V}\gets f_v$ + +5: inputs = Initilize $(f_e, f_v)$ Initialize inputs for prefilling + +6: while $SC$ not met do + +7: next_token, attention_map = model(inputs) + +8: Append next_token to predicted_tokens + +$\triangleright$ ADS judgement + +9: if predicted_tokens = S then + +10: $\mathcal{V}_{\mathrm{selected}} = \mathrm{ADS}(\mathcal{V},\mathrm{attention\_map},n)$ Apply Attention-driven Selection +11: Append $\nu_{\mathrm{selected}}$ to predicted_tokens +12: end if +13: inputs = Update(inputs, predicted_tokens) > Updates inputs for next step generation +14: end while +15: Answer = Tokenizer Decode(predicted_tokens) +16: return Answer + +tokens and $d$ is the dimensions of the hidden states of the LLM. The built-in LLM predicts next-tokens in a left-to-right fashion according to visual tokens $f_{v}^{l \times d}$ and the instructions. + +Multimodal CoT. Compared with the direct prediction described in Eqn. 1, Multimodal CoT further introduces a prompt to elicit VLMs to generate a series of intermediate textual rationales $(r_1, r_2, \ldots)$ before the final answer: + +$r_1, r_2, \ldots$ , answer $= \mathbf{VLM}(\text{Prompt}, \text{Image}, \text{Instruction})$ . (2) + +Technically, the prompt could be represented as a sequence of demonstrations, each consisting of a triple: (Image, Rationale, Answer). Alternatively, an explicit instruction, such as "Let's think step by step," could also serve as the prompt. + +# 3.2. Interleaved-modal Chain-of-Thought + +Previous multimodal CoT prompting methods only produce text-only reasoning steps to improve the reasoning ability of VLMs. These intermediate steps are generated according to the entire image, which are struggle to express exact fine-grained associations with the original image. Given these limitations, we propose a more advanced Interleaved-modal Chain-of-Thought (ICoT) prompting, aiming to elicit VLMs to generate a series of multimodal intermediate reasoning steps, each consisting of paired im + +# Algorithm 2 Attention-driven Selection + +1: Input: Attention map $A_{t}$ , Selected number $n$ , Visual tokens $f_{v}$ +2: Output: Fine-grained visual information $\nu_{\mathrm{selected}}$ +3: $\mathcal{V}_{\mathrm{selected}}\gets \emptyset$ Initialize as an empty set +4: Indices $\leftarrow$ TopK(At,n) +5: for $i$ in Indices do +6: Append $f_{v}^{l \times d}[i]$ to $\mathcal{V}_{\text{selected}}$ +7: end for +8: Restore $(\mathcal{V}_{\mathrm{selected}}$ , Indices) $>$ Restore relative positions in-place +9: return $\nu_{\mathrm{selected}}$ + +![](images/c5d74196e5f92c316d404a4a83d73005768f844ef2700b28accda62fc1a2e011.jpg) +Figure 2. The workflow of ADS selecting fine-grained visual rationales. Signal attention represents the attention map of the signal token overall visual tokens. + +age and textual rationale. Generated intermediate reasoning steps formulating interleaved-modal rationales to effectively lead to the final outputs. In this paper, we consider the visual rationales in interleaved-modal rationales as fine-grained visual information $x_{v}^{h' \times w'}$ extracted from an image $x_{v}^{h \times w1}$ . These visual rationales capture relevant details in the image, such as objects, colors, and texts, interleaved with the following generated textual rationale to infer the final answer: + +$$ +\begin{array}{r} r _ {1}, x _ {v _ {1}}, r _ {2}, x _ {v _ {2}}, \dots , \text {a n s w e r} = \mathbf {V L M} (\text {P r o m p t , I m a g e}, \\ \text {I n s t r u c t i o n}). \end{array} \tag {3} +$$ + +# 3.3. Attention-driven Selection + +Although the proposed ICoT is both novel and conceptually sound, current VLMs are unable to generate such fine-grained visual information. This remains true even for VLMs [18, 25, 30] that are empowered with multimodal generation ability. We thus propose to simplify the problem from fine-grained visual information generation to fine-grained visual information selection, as this information + +has been naturally contained in the origin image, namely $x_{v}^{h^{\prime}\times w^{\prime}}\in x_{v}^{h\times w}$ where $h^\prime \ll h$ and $w^{\prime}\ll w$ + +Specifically, before performing next-token prediction, ICoT requires the VLM to cache visual tokens $f_{v}^{l \times d}$ extracted by its built-in visual encoder $\mathbf{E}(x_{v})$ for further selection. In the following decoding steps, we consider the VLM deems it necessary to insert a piece of visual rationales after generating a pre-defined signal token $S$ as shown in Figure 2, which is a natural language token that indicates the beginning of a textual rationale. Therefore, ADS will be signaled to select fine-grained visual information from $f_{v}^{l \times d}$ upon the VLM generating $S$ : + +$$ +\text {d o} \text {s e l e c t i o n} = \left\{ \begin{array}{l l} \text {T r u e} & \text {i f p r e d i c t e d} \\ \text {F a l s e} & \text {o t h e r w i s e .} \end{array} \right. \tag {4} +$$ + +Then, ADS selects $n$ visual tokens from $f_{v}^{l \times d}$ as fine-grained visual information according to the signal attention map $A_{t}$ at the current decoding step $t$ : + +$$ +\mathcal {V} _ {\text {s e l e c t e d}} = \left\{f _ {v} [ i ] \mid i \in \operatorname {T o p K} \left(A _ {t}, n\right) \right\}, \tag {5} +$$ + +where $A_{t}$ is obtained by averaging the attention map between the signal token and visual tokens across all VLM layers. Up to now, the selected visual tokens are sorted by their attention scores, and we subsequently restore the relative position of $f_{v}[i]$ in the origin image in place, prioritizing rows. Once fine-grained visual tokens $\nu_{\mathrm{selected}}$ are obtained, VLMs will take as input the concatenated $\nu_{\mathrm{selected}}$ and predicted_tokens, denote as Cat(predicted_tokens, $\nu_{\mathrm{selected}}$ ), resuming the original autoregressive text generation process. Hence, the current decoding step is formulated as: + +$$ +\begin{array}{r} \text {n e x t - t o k e n} = \mathbf {V L M} (\text {C a t (p r e d i c t e d - t o k e n s}, \\ \nu_ {\text {s e l e c t e d}})). \end{array} \tag {6} +$$ + +Notably, to avoid misunderstanding, we first convert predicted_tokens into word embeddings $f_{e}$ , and then concatenate $f_{e}$ with selected fine-grained visual information $\mathcal{V}_{\mathrm{selected}}$ in the embedding-level. We provide a detailed description of ICoT and ADS in Algorithm 1 and Algorithm 2, respectively. + +Technically, ICoT inherits the existing eliciting methods from CoT, such as attaching an instruction: "Let's think step by step" in zero-shot ICoT or providing few-shot examples. In the few-shot ICoT, each example consists of an image, textual rationales, and visual rationales. These examples can be manually designed to prompt VLMs regarding their way of thinking and generation formatting, among other aspects. + +# 4. Experiments + +# 4.1. Datasets + +$\mathbf{M}^{3}\mathbf{C}\mathbf{o}\mathbf{T}$ [6] is a novel multimodal CoT benchmark specifically concentrated on multi-domain, multi-reasoning-step. $\mathbf{M}^{3}\mathbf{C}\mathbf{o}\mathbf{T}$ contains 267 categories from science, mathematics, and commonsense domains. As the question of each instance is relatively complex, their rationales have an average length of 293 tokens and rely more on fine-grained visual information, which can reflect the advantages of ICoT compared with previous multimodal CoT methods. + +ScienceQA [24] is a popular dataset used to evaluate the reasoning ability of VLMs. We use ScienceQA to provide a general comparison between ICoT and other existing multimodal CoT methods. + +LLaVA-Bench In-the-Wild (LLaVA-W) [15] evaluates VLMs' ability to respond to visual questions with detailed long-form answers, which also focus on the fine-grained visual description. The reference label of each instance is produced by GPT-4v. + +# 4.2. Baselines + +No-CoT responds to the current input image and question directly without further prompting. The few-shot demonstrations of the direct generation mode consist of (Image, Question, Answer). + +Multimodal CoT [33] elicits VLMs to generate a series of text-only intermediate reasoning steps to infer the final outputs. + +CCoT [19] first generates a scene graph (SG) using the VLM itself and then uses that SG in the prompt to produce a response. The SG is a JSON-like structural description of the given image with extensive compositional information of objects in the current images. Following their settings, we apply their official prompt to prompt VLMs to generate SGs and answers respectively. + +DDCoT [34] first prompts LLM to deconstruct the input question into a sequence of basic sub-questions and then applies a VQA model to answer these sub-questions involving visual information. In this paper, we use the VLM that plays the role of LLM in DDCoT for a fair comparison, as their original LLM is ChatGPT. + +SCAFFOLD [12] overlays a coordinate matrix onto the input image, exactly demonstrating relative visual positions for VLMs. During reasoning, VLMs are steered to utilize these coordinates that indicate fine-grained visual information in the image to solve different vision-language tasks. We use their released scripts to add coordinates over each image and then use their official prompt to elicit VLMs. + +Specifically, in the few-shot scenario, the demonstrations of these baselines are human-written, aligning with ICoT. + +
BackboneMethods0-shot1-shot
M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑
Chameleon-7BNo-CoT29.147.713.128.448.523.9
Multimodal CoT [33]28.549.020.430.650.720.6
CCoT [19]29.450.222.131.451.324.5
DDCoT [34]28.649.820.229.849.223.1
SCAFFOLD [12]29.648.521.731.147.524.7
ICoT (Ours)29.851.025.232.353.427.6
% Improve0.6%1.6%14.0%2.8%4.0%11.7%
BackboneMethods0-shot1-shot
M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑
Qwen2-VL-7BNo-CoT43.656.332.745.464.433.5
Multimodal CoT [33]40.151.330.742.558.331.4
CCoT [19]43.356.429.444.163.833.9
DDCoT [34]42.655.231.245.764.932.8
SCAFFOLD [12]41.753.731.844.962.533.1
ICoT (Ours)44.156.834.246.065.435.7
% Improve1.1%0.7%4.6%0.6%0.7%5.3%
+ +Table 1. Results of ICoT and baselines based on Chameleon and Qwen2-VL, with the highest score bold. ${\mathrm{M}}^{3}\mathrm{{CoT}}$ and ScienceQA are evaluated by accuracy, and we report the ROUGE-L score for the LLaVA-W benchmark. % improve represents the relative improvement achieved by ICoT over the previously best baseline. + +# 4.3. Implement Details + +We apply ICoT over Chameleon-7B [25] and Qwen2-VL7B-Instruct [29], which represents the fine-grained visual information in the form of discrete vokens and dense features. All experiments are conducted on A800 GPUs, and we evaluate ICoT under both zero- and one-shot scenarios. During generating interleaved-modal rationales, the signal token $S$ used to trigger ADS is set to line break, i.e., " $\backslash n$ ", by default, which semantically and empirically indicates the end of a generated rationale and the beginning of the next one. + +VLMs insert visual tokens selected by ADS following the special token at the granularity of 64 according to posterior results shown in Table 4 of Appendix 8. Notably, to shorten the representation of an image, Qwen2-VL introduces a novel merge mechanism, and we approximately consider its patch size to be $(28 \times 28)$ . Each patch of Qwen2-VL has approximately 4 times as many pixels as a Chameleon patch $(16 \times 16)$ , which results in practical selection numbers of ADS set to 16. Considering the work of ADS requires the inner attention map, we apply the "eager" attention on both Chameleon-7B and Qwen2-VL, limited to the dependency of related python libraries2. + +# 4.4. Main Results + +We comprehensively evaluate the performance of ICoT on top of Chameleon-7B and Qwen2-VL-7B through $\mathbf{M}^3\mathrm{CoT}$ , ScinceQA, and LLaVA-W in Table 1. In 0-shot settings, ICoT outperforms all baselines, including direct generation (No-CoT), CoT, CCoT, DDoT, and SCAFFOLD. Specifically, ICoT distinguishes from Multimodal CoT in terms of the modality of reasoning steps, which exhibit the advantages of interleaved-modal rationales to infer the final answer effectively. Compared with other multimodal CoT methods, the performance gains of ICoT further indicate that interleaved-modal rationales are more reasonable in intuition and effect than plainly inserted scene graphs (CCoT) and deconstructed sub-questions (DDoS). In 1-shot settings, ICoT demonstrations contain manually selected fine-grained visual information, while their text rationales are kept the same as other baselines. The performance gains compared with 0-shot ICoT indicate that our manually designed fine-grained ICoT demonstrations potentially guide VLMs to think in this format. In Table 4, we rigorously ablate the effectiveness of fine-grained ICoT demonstrations. + +Additionally, ICoT achieves the most relative performance gains in the LLaVA-W benchmark as the reference labels contain details sourced from images. These substantial performance gains compared with other baselines prove that visual tokens selected by ADS effectively capture the fine-grained visual information of an image, aiding the gen + +eration of high-quality text rationales. + +# 4.5. Ablation Study + +We ablate ICoT to verify the effectiveness of each portion across three benchmarks in Table 2 with the following settings: (1). w/o ADS: VLMs generate text-only rationales. (2). w/o FVI: Patches inserted in the demonstration are randomly sampled. Results indicate that both ADS and fine-grained visual information (FVI) incorporated in the demonstration are necessary. In particular, interleaved-modal rationales exhibit substantial advantages in generating high-quality textual rationales compared with text-only rationales (w/o ADS). When substituting ICoT demonstrations with normal ones (w/o FVI), the performance degradation proves that fine-grained visual information in demonstrations effectively guides VLMs to think in this format. Compared with the performance difference between removing ADS and FVI, we find that generating paired visual and text rationales boosts more improvements. + +Additionally, the performance gap is relatively smooth in ScienceQA and more dramatic on $\mathrm{M}^3\mathrm{CoT}$ and LLaVA-W. We attribute this to the ScienceQA dataset being relatively easier than others since both $\mathrm{M}^3\mathrm{CoT}$ and the answers of LLaVA-W highly rely on the fine-grained visual information of an image. Therefore, our proposed ICoT has the potential to solve complex vision-language tasks. + +# 4.6. In-depth Analysis + +Analysis on realizing ICoT via KV Cache Up to now, the fine-grained visual information of ICoT is provided at the input end via discrete tokens or dense visual tokens, which brings more computation. After rethinking the generating process of an autoregressive model, there are other inputs that are proposed to avoid repeated computation, namely, the Key-Value (KV) Cache. The input image was stored in the KV Cache during the prefilling phase before generating multimodal intermediate reasoning steps in a left-to-right fashion. Therefore, copying the KV cache of fine-grained visual information enables ICoT with reduced computational costs, as visual information does not require extra forward propagation. As shown in Table 3, copying the KV Cache brings performance degradation compared with providing visual information at the input end. We attribute this phenomenon to the fact that although copying KV Cache indeed makes VLMs attend more to the same region as ADS, the optimal visual information is highlighted in a position-agnostic case, determined by the nature of KV Cache. Specifically, this degrades the original interleaved-modal rationales into non-interleaved ones as position information is early fused into KV Cache, and thus the copied ones are inherently insensitive to the position of textual rationale. + +However, considering it brought slight performance + +
MethodsM3CoTScienceQALLaVA-W
ICoT32.353.427.6
w/o ADS29.2 (-3.1)52.4(-1.0)24.5(-3.1)
w/o FVI30.6 (-1.8)52.8(-0.6)25.9(-1.7)
w/o ADS+FVI29.1(-3.2)51.0(-2.4)23.0(-4.6)
+ +Table 2. Ablation studies of 1-shot ICoT on Chameleon-7B. () describes the performance degradation compared with ICoT. FVI indicates the 1-shot demonstration contains fine-grained visual information. ADS indicates that VLMs generate interleaved-modal reasoning steps. + +
Dataset0-shot1-shot
KV-CopyICoTKV-CopyICoT
M3CoT29.129.831.532.3
ScienceQA49.751.052.953.4
LLaVA-W24.725.227.027.6
+ +Table 3. Results comparison between copying KV Cache (KV-Copy) and inserting selected patches. + +
MethodsM3CoTScienceQALLaVA-W
Human-written32.353.427.6
Model-written31.5(-0.8)51.8(-1.6)26.7(-0.8)
+ +Table 4. Results concerning the design of demonstrations on Chameleon-7B. () describes the performance degradation compared with ICoT. Human-written indicates that demonstrations are manually designed with fine-grained visual information inserted, and Model-written indicates that the VLM generates both visual and textual rationale via ICoT. + +degradation and factually reduced computation costs, we believe this exploration is still valuable, and we call for more interesting exploration in realizing ICoT. + +Analysis on the Demonstrations We also attempt to let VLMs generate the demonstrations via themselves (Automatic at the bottom of Table 4). Results indicate that using the automatically generated demonstrations also brings performance degradation compared with ICoT using manually designed ones. We consider it is caused by the fact that formulating a continuous sub-image through ADS is nontrivial, and some discrete patches inevitably introduce additional noise. Therefore, considering that designing such a demonstration is not time-consuming, ICoT utilizes manually designed ones to elicit VLMs to perform ICoT for better performance. + +# 5. Case Study + +In this section, we empirically illustrate the advantages of ICoT via three case studies in Figure 3. These case studies + +![](images/70796664c06c33732a7065f24974cfdd8a99c09ec9a847f3cef367a45cd3bfd5.jpg) +Figure 3. Case studies between ICoT and multimodal CoT with text-only rationales on Chameleon. Three cases are selected according to three typical problems in text-only problems: misunderstanding, overgeneralization, and hallucination. Red blocks indicate the incorrect rationales. + +focused on three typical problems that occurred in text-only rationales, namely, misunderstanding (top), overgeneralization (middle), and hallucination (bottom). + +Specifically, in the first case, interleaved-modal CoT first recognizes three different objects via captions: "inflatable castle, crayons, and a parachute". Then, in the second reasoning step, ADS inserts selected patches from the scheduled objects to elicit the VLM to conclude their common property, and VLM infers a correct answer. Text-only CoT misunderstands that the three objects are all colored pencils, ignoring the castle and the parachute, even the final answer is correct. In the second case, text-only rationales overgeneralize flying a kite to a kite festival, leading to a wrong answer. ICoT first recognizes a man standing in a field of grass and then infers it is a windy day according to a few kites in the sky. In the last case, it provides the other typical error of text-only CoT, namely, hallucination. As text-only CoT purely relies on language reasoning ability, VLMs have the potential to imagine something not mentioned in the image, resulting in a wrong answer. ICoT first infers from the street sign that there may be a troll attraction nearby according to patches of the indicator inserted by ADS. Then, the ADS helps the VLMs to attend to the troll statue under the bridge and infer it is likely the mentioned attraction, finally arriving at the correct answer. + +Even though the above case studies exhibit the advantages of ICoT, ADS still brings potential problems. For example, ADS is triggered to select patches when VLM generates a pre-defined signal token. This simple mechanism is a double-edged sword in that VLMs will generate low-quality responses if this token is generated with a high frequency. + +# 6. Conclusion + +In this paper, we first propose interleaved-modal CoT (ICoT), which generates interleaved-modal rationales to infer the final answer effectively. In light of the challenges of applying ICoT on existing VLMs, we then introduce Attention-driven Selection (ADS), a plug-and-play strategy to identify optimal patches from the image without being parameterized. We evaluate ICoT on Chameleon-7B and Qwen2-VL-7B-Instruct, representing VLMs of two architectures. Extensive experiments conducted on $\mathbf{M}^3\mathrm{CoT}$ , ScienceQA, and LLaVA-W, under both zero- and few-shot scenarios, have proven that ICoT achieves substantial performance (up to $14\%$ ) compared with the existing multimodal CoT methods. Additionally, in the analysis section, we conduct a preliminary exploration of implementing ICoT by copying the KV cache of optimal visual tokens and explain the inner trade-off between efficiency and performance in this approach. + +Although ICoT has proven its effectiveness in this paper, we consider that ICoT still has significant potential for further improvement. The patch selection in ADS requires storing attention scores, which brings additional memory overhead. Moreover, the fixed number of selected patches in the ADS design is sub-optimal, resulting in unexpected outputs for VLMs. To address these, we intend to incorporate established techniques from segmentation or grounding methods to create a more robust implementation of ICoT. In the future, we also plan to evaluate it across additional backbones and benchmarks to better assess its generalization ability. + +# 7. Acknowledgement + +I would like to express my sincere gratitude to all the authors and reviewers for their valuable contributions to this research. This work was supported by the National Natural Science Foundation of China (NSFC 62106165) and the Project Funded by the Priority Academic Program Development of Jiangsu Higher Education Institutions, China. + +# References + +[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.3 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 3 +[3] BIG bench authors. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on Machine Learning Research, 2023. 1 +[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1 +[5] Claus Bundesen. A theory of visual attention. Psychological review, 97(4):523, 1990. 2 +[6] Qiguang Chen, Libo Qin, Jin Zhang, Zhi Chen, Xiao Xu, and Wanxiang Che. M3cot: A novel benchmark for multidomain multi-step multi-modal chain-of-thought. arXiv preprint arXiv:2405.16473, 2024. 1, 3, 5 +[7] Ethan Chern, Jiadi Su, Yan Ma, and Pengfei Liu. Anole: An open, autoregressive, native large multimodal models for interleaved image-text generation. arXiv preprint arXiv:2407.06135, 2024. 2 +[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 1 +[9] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 1 +[10] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. arXiv preprint arXiv:2406.11832, 2024.3 +[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3 +[12] Xuanyu Lei, Zonghan Yang, Xinrui Chen, Peng Li, and Yang Liu. Scaffolding coordinates to promote vision-language coordination in large multi-modal models. arXiv preprint arXiv:2402.12058, 2024. 1, 3, 5, 6 + +[13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3 +[14] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3 +[15] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.3,5 +[16] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3 +[17] Jiacheng Liu, Alisa Liu, Ximing Lu, Sean Welleck, Peter West, Ronan Le Bras, Yejin Choi, and Hannaneh Hajishirzi. Generated knowledge prompting for commonsense reasoning. arXiv preprint arXiv:2110.08387, 2021. 1 +[18] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26439-26455, 2024. 2, 3, 4 +[19] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 1, 3, 5, 6 +[20] OpenAI. Learning to reason with llms, 2023. 1 +[21] Michael I Posner, Steven E Petersen, et al. The attention system of the human brain. Annual review of neuroscience, 13(1):25-42, 1990. 2 +[22] Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446, 2021. 1 +[23] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1 +[24] Tanik Saikh, Tirthankar Ghosal, Amish Mittal, Asif Ekbal, and Pushpak Bhattacharyya. Scienceqa: A novel resource for question answering on scholarly articles. International Journal on Digital Libraries, 23(3):289-301, 2022. 3, 5 +[25] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2,3,4,6 +[26] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 1 +[27] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, + +Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 3 +[28] Lei Wang, Yi Hu, Jiabang He, Xing Xu, Ning Liu, Hui Liu, and Heng Tao Shen. T-sciq: Teaching multimodal chain-of-thought reasoning via large language model signals for science question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19162-19170, 2024. 1, 3 +[29] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 3, 6 +[30] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 4 +[31] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1 +[32] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 3 +[33] Zhuosheng Zhang, Aston Zhang, Mu Li, George Karypis, Alex Smola, et al. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2024. 1, 5, 6 +[34] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168-5191, 2023. 1, 3, 5, 6 \ No newline at end of file diff --git a/2025/Interleaved-Modal Chain-of-Thought/images.zip b/2025/Interleaved-Modal Chain-of-Thought/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4b7baf81f98dd28b40ac9dc68cd76da37ba3f7b2 --- /dev/null +++ b/2025/Interleaved-Modal Chain-of-Thought/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7239e34f23c2b638829779a317697bec8fa547bcc58a6cfa03309e63b6be098e +size 414309 diff --git a/2025/Interleaved-Modal Chain-of-Thought/layout.json b/2025/Interleaved-Modal Chain-of-Thought/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4edf22473a7d95ca3f593609f0ed4a1907b6e8aa --- /dev/null +++ b/2025/Interleaved-Modal Chain-of-Thought/layout.json @@ -0,0 +1,8404 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 189, + 103, + 422, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 103, + 422, + 121 + ], + "spans": [ + { + "bbox": [ + 189, + 103, + 422, + 121 + ], + "type": "text", + "content": "Interleaved-Modal Chain-of-Thought" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 142, + 420, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 142, + 420, + 158 + ], + "spans": [ + { + "bbox": [ + 189, + 142, + 420, + 158 + ], + "type": "text", + "content": "Jun Gao" + }, + { + "bbox": [ + 189, + 142, + 420, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 189, + 142, + 420, + 158 + ], + "type": "text", + "content": ", Yongqi Li" + }, + { + "bbox": [ + 189, + 142, + 420, + 158 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 189, + 142, + 420, + 158 + ], + "type": "text", + "content": ", Ziqiang Cao" + }, + { + "bbox": [ + 189, + 142, + 420, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 189, + 142, + 420, + 158 + ], + "type": "text", + "content": ", Wenjie Li" + }, + { + "bbox": [ + 189, + 142, + 420, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 153, + 165, + 457, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 165, + 457, + 234 + ], + "spans": [ + { + "bbox": [ + 153, + 165, + 457, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 165, + 457, + 234 + ], + "type": "text", + "content": "School of Computer Science and Technology, Soochow University \n" + }, + { + "bbox": [ + 153, + 165, + 457, + 234 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 165, + 457, + 234 + ], + "type": "text", + "content": "Department of Computer Science, The Hong Kong Polytechnic University \njgao1106@stu.suda.edu.cn, liyongqi0@gmail.com \nzqcao@suda.edu.cn, cswjli@comp.polyu.edu.hk \nhttps://github.com/jungao1106/ICoT" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 262, + 200, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 262, + 200, + 274 + ], + "spans": [ + { + "bbox": [ + 151, + 262, + 200, + 274 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 287, + 297, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 287, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 54, + 287, + 297, + 601 + ], + "type": "text", + "content": "Chain-of-Thought (CoT) prompting elicits large language models (LLMs) to produce a series of intermediate reasoning steps before arriving at the final answer. However, when transitioning to vision-language models (VLMs), their text-only rationales struggle to express the fine-grained associations with the original image. In this paper, we propose an image-incorporated multimodal Chain-of-Thought, named Interleaved-modal Chain-of-Thought (ICoT), which generates sequential reasoning steps consisting of paired visual and textual rationales to infer the final answer. Intuitively, the novel ICoT requires VLMs to enable the generation of fine-grained interleaved-modal content, which is hard for current VLMs to fulfill. Considering that the required visual information is usually part of the input image, we propose Attention-driven Selection (ADS) to realize ICoT over existing VLMs. ADS intelligently inserts regions of the input image to generate the interleaved-modal reasoning steps with ignorable additional latency. ADS relies solely on the attention map of VLMs without the need for parameterization, and therefore it is a plug-and-play strategy that can be generalized to a spectrum of VLMs. We apply ADS to realize ICoT on two popular VLMs of different architectures. Extensive evaluations of three benchmarks have shown that ICoT prompting achieves substantial performance (up to " + }, + { + "bbox": [ + 54, + 287, + 297, + 601 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 54, + 287, + 297, + 601 + ], + "type": "text", + "content": ") and interpretability improvements compared to existing multimodal CoT prompting methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 623, + 135, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 135, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 135, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 644, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 296, + 693 + ], + "type": "text", + "content": "Chain-of-Thought (CoT) [31] prompting aims to augment the reasoning capabilities of large language models (LLMs) [4, 8, 22, 26] by eliciting them to produce a sequence of intermediate natural language reasoning steps be" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 262, + 554, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 554, + 323 + ], + "type": "text", + "content": "fore arriving at the final output. CoT has proven effective in various reasoning tasks, including arithmetic [9], commonsense [17], and symbolic [3], and it has become a potential pathway to advanced artificial intelligence, as depicted in GPT-o1 [20]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 326, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 602 + ], + "type": "text", + "content": "With the development of vision-language models (VLMs), extending CoT prompting into multimodal CoT to improve the reasoning capabilities of VLMs in vision-related tasks becomes increasingly important [19, 28, 33, 34]. The initial multimodal CoT attempts [28, 33] take as input the fused visual and textual embeddings and train language models, such as T5 [23] models, to generate text-only rationales and answers. In the era of VLMs, introducing triple demonstrations composed of an image with the instruction, textual rationales, and the final output (e.g., answer) has proven effective in sparking the reasoning ability of VLMs [6]. Then, related studies focus on improving the linguistic reasoning ability of VLMs. Specifically, DDCoT [34] leverages VLMs to deconstruct problems and resolve them respectively, and CCoT [19] generates scene graphs to prompt VLMs with object and position descriptions. SCAFFOLD [12] overlays a coordinate matrix onto the image to prompt the VLMs with relative visual positions. However, these methods still generate text-only reasoning steps, making it hard to express the fine-grained associations with the origin image exactly. As shown on the left of Figure 1, textual position descriptions, e.g., at the top, are too rough to identify all fruits (orange and banana)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "In light of the limitations of text-only rationales, we propose incorporating visual information to enhance the precision of fine-grained associations between generated textual rationales and the corresponding image. We therefore propose an advanced multimodal Chain-of-Thought prompting, named Interleaved-modal Chain-of-thought (ICoT), as shown on the right of Figure 1. ICoT generates multimodal rationales consisting of paired images and textual rationales that formulate interleaved-modal reasoning steps" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 148, + 712 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19520" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 86, + 76, + 148, + 88 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 76, + 148, + 88 + ], + "spans": [ + { + "bbox": [ + 86, + 76, + 148, + 88 + ], + "type": "text", + "content": "Model Input" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 94, + 291, + 226 + ], + "blocks": [ + { + "bbox": [ + 67, + 94, + 291, + 226 + ], + "lines": [ + { + "bbox": [ + 67, + 94, + 291, + 226 + ], + "spans": [ + { + "bbox": [ + 67, + 94, + 291, + 226 + ], + "type": "image", + "image_path": "4b546465888409e42344389e068077ff428c3a1db8ac1922234d281e1bdcf4c1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 228, + 251, + 239 + ], + "lines": [ + { + "bbox": [ + 72, + 228, + 251, + 239 + ], + "spans": [ + { + "bbox": [ + 72, + 228, + 251, + 239 + ], + "type": "text", + "content": "Question: What types of fruits are in the picture?" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 78, + 251, + 172, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 251, + 172, + 262 + ], + "spans": [ + { + "bbox": [ + 78, + 251, + 172, + 262 + ], + "type": "text", + "content": "Text-only Rationale" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 268, + 282, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 268, + 282, + 299 + ], + "spans": [ + { + "bbox": [ + 71, + 268, + 282, + 299 + ], + "type": "text", + "content": "The mixed variety of fruits suggests that this is a fruit platter. The small, round, light green items at the bottom left appear to be grapes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 307, + 282, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 307, + 282, + 327 + ], + "spans": [ + { + "bbox": [ + 71, + 307, + 282, + 327 + ], + "type": "text", + "content": "The textured orange skin at the top indicates this is likely an orange, while a banana rests beside it." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 335, + 279, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 335, + 279, + 356 + ], + "spans": [ + { + "bbox": [ + 71, + 335, + 279, + 356 + ], + "type": "text", + "content": "The dark red, smooth-skinned fruit on the right appears to be an apple." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 372, + 111, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 372, + 111, + 383 + ], + "spans": [ + { + "bbox": [ + 75, + 372, + 111, + 383 + ], + "type": "text", + "content": "Answer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 392, + 262, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 392, + 262, + 417 + ], + "spans": [ + { + "bbox": [ + 74, + 392, + 262, + 417 + ], + "type": "text", + "content": "Therefore, grapes, orange, banana, and apple are in the picture." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 337, + 76, + 399, + 88 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 76, + 399, + 88 + ], + "spans": [ + { + "bbox": [ + 337, + 76, + 399, + 88 + ], + "type": "text", + "content": "Model Input" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 95, + 543, + 226 + ], + "blocks": [ + { + "bbox": [ + 320, + 95, + 543, + 226 + ], + "lines": [ + { + "bbox": [ + 320, + 95, + 543, + 226 + ], + "spans": [ + { + "bbox": [ + 320, + 95, + 543, + 226 + ], + "type": "image", + "image_path": "0f38bad4fa19e33ce5395ce9331e107b56f82c06456e96216a6dfbdab1643d62.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 228, + 503, + 239 + ], + "lines": [ + { + "bbox": [ + 323, + 228, + 503, + 239 + ], + "spans": [ + { + "bbox": [ + 323, + 228, + 503, + 239 + ], + "type": "text", + "content": "Question: What types of fruits are in the picture?" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 332, + 249, + 468, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 249, + 468, + 261 + ], + "spans": [ + { + "bbox": [ + 332, + 249, + 468, + 261 + ], + "type": "text", + "content": "Interleaved-modal Rationale" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 269, + 340, + 297 + ], + "blocks": [ + { + "bbox": [ + 317, + 269, + 340, + 297 + ], + "lines": [ + { + "bbox": [ + 317, + 269, + 340, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 269, + 340, + 297 + ], + "type": "image", + "image_path": "ac007bed09b3a8ba675eaa1a0efac5fd8310a9ca4bd9c8569a4681a7984d6dae.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 344, + 268, + 532, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 268, + 532, + 298 + ], + "spans": [ + { + "bbox": [ + 344, + 268, + 532, + 298 + ], + "type": "text", + "content": "The several grapes grouped together and the jumbled assortment of different fruits suggest that this is a fruit platter." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 303, + 340, + 330 + ], + "blocks": [ + { + "bbox": [ + 317, + 303, + 340, + 330 + ], + "lines": [ + { + "bbox": [ + 317, + 303, + 340, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 303, + 340, + 330 + ], + "type": "image", + "image_path": "f17f9a0c5296e00e97c2043e6c3152b8b8155b790bb15240a94105058f415c96.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 344, + 307, + 531, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 307, + 531, + 328 + ], + "spans": [ + { + "bbox": [ + 344, + 307, + 531, + 328 + ], + "type": "text", + "content": "The orange colored rind behind an apple indicates that there is an orange and an apple in the picture." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 317, + 335, + 340, + 363 + ], + "blocks": [ + { + "bbox": [ + 317, + 335, + 340, + 363 + ], + "lines": [ + { + "bbox": [ + 317, + 335, + 340, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 335, + 340, + 363 + ], + "type": "image", + "image_path": "0c05d7aaf6c67afc1cbfb92e9d124c53430002ab075e07bdb6ef19cf6f839e30.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 431, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 431, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 555, + 475 + ], + "type": "text", + "content": "Figure 1. The illustration between multimodal CoT with text-only rationales (Left) and interleaved-modal rationales (Right). Green blocks are correct texts used to infer the final answer. Text-only rationales restrict VLMs to use a rough description to indicate the position of objects. Transparent boxes indicate that these regions are selected and inserted to formulate paired visual and textual rationales in ICoT." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 345, + 335, + 523, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 335, + 523, + 355 + ], + "spans": [ + { + "bbox": [ + 345, + 335, + 523, + 355 + ], + "type": "text", + "content": "The sleek yellow peel off to the right clues at an unpeeled banana." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 325, + 373, + 362, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 373, + 362, + 384 + ], + "spans": [ + { + "bbox": [ + 325, + 373, + 362, + 384 + ], + "type": "text", + "content": "Answer" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 322, + 389, + 538, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 389, + 538, + 415 + ], + "spans": [ + { + "bbox": [ + 322, + 389, + 538, + 415 + ], + "type": "text", + "content": "Therefore, grapes, orange, apple, and banana are in the picture." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "content": "to infer the final output. To the best of our knowledge, ICoT is the first multimodal CoT with images incorporated, and it aligns more closely with human thinking processes [5, 21]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 54, + 533, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 295, + 700 + ], + "type": "text", + "content": "Intuitively, facilitating the novel ICoT is non-trivial, as it introduces challenges for VLMs to support fine-grained interleaved-modal content generation. No current VLMs meet this condition completely. Perceiver-based VLMs such as Qwen2-VL [29] converts images into visual embeddings. Thus, they support fine-grained visual understanding but cannot generate multimodal outputs. Recently proposed unified-modeling VLMs, such as Chameleon [25], UnifiedIO 2 [18], and Emu-3 [30], enable multimodal generation by tokenizing images into discrete tokens. However, on the one hand, unified-modeling VLMs exhibit inertia toward multimodal content generation [7]; on the other hand, the generated images belong to the fixed pre-defined resolution instead of fine granularity." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Since required visual information is usually part of the" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "text", + "content": "input image for ICoT, we accordingly propose Attention-driven Selection (ADS) to realize ICoT. The basic idea of ADS is to signal VLMs to select patches from the input image rather than generate extra images. At the beginning of generating each textual rationale, ADS inserts a piece of visual tokens of selected patches from the input image to refine the generation of the following textual rationale. Specifically, ADS utilizes the attention map of VLMs to identify optimal patches from the input image as fine-grained visual rationales. Once these fine-grained visual rationales are inserted into the current generation sequence, the VLM resumes the original autoregressive text generation process based on previous multimodal content, formulating paired image and textual rationales to infer the final outputs. Notably, since ADS does not compel VLMs to generate real images, it brings ignorable inference latency compared with previous text-only CoT methods. Additionally, ADS leverages the attention map of VLMs without re" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19521" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "quiring parameterization, making it a plug-and-play strategy that can be easily adapted to a wide range of VLMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 99, + 295, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 99, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 54, + 99, + 295, + 229 + ], + "type": "text", + "content": "In this paper, we apply ADS to realize ICoT on Chameleon and Qwen2-VL, representing the state-of-the-art unified modeling and perceiver-based VLMs. The results on existing datasets, including " + }, + { + "bbox": [ + 54, + 99, + 295, + 229 + ], + "type": "inline_equation", + "content": "\\mathrm{M}^3\\mathrm{CoT}" + }, + { + "bbox": [ + 54, + 99, + 295, + 229 + ], + "type": "text", + "content": " [6], ScienceQA [24], and LLaVA-W [15], indicate that ICoT realized by ADS brings VLMs with substantial performance gains (up to " + }, + { + "bbox": [ + 54, + 99, + 295, + 229 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 54, + 99, + 295, + 229 + ], + "type": "text", + "content": ") compared with current multimodal CoT methods. Additionally, it is noted that the tracked interleaved-modal rationales further enhance the interpretability of the generated results. Our main contribution can be concluded as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 233, + 296, + 388 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 233, + 295, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 295, + 292 + ], + "type": "text", + "content": "- We propose interleaved-modal CoT, which innovates text-only rationales into multimodal ones to construct clearer reasoning. To our knowledge, we are the first to incorporate images into the intermediate reasoning steps in multimodal CoT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 293, + 295, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 340 + ], + "type": "text", + "content": "- We propose an effective and efficient Attention-driven Selection strategy to facilitate ICoT, which is training-free and widely applicable to VLMs without requiring them to support multimodal generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 341, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 388 + ], + "type": "text", + "content": "- Experiments demonstrate that our ICoT significantly surpasses existing multimodal CoT methods, proving that the interleaved-modal reasoning process is a foundational innovation in the line of CoT." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 407, + 142, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 142, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 142, + 419 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 429, + 234, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 234, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 234, + 442 + ], + "type": "text", + "content": "2.1. Vision-Language Models (VLMs)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 450, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 714 + ], + "type": "text", + "content": "Currently, predominate VLMs such as Qwen-VL [2, 29], BLIP [13], and LLaVA [14-16] are mainly built upon a Large Language Model (LLM), a visual module, and an aligned vision-language adapter. The visual module, e.g., Vision Transformer (ViT) [1], encodes images into dense representations, and then the adapter, e.g., MLP or Q-Former, converts these representations into LLM-readable visual tokens. Finally, visual tokens and textual tokens are fed into the LLM to perform the next-token prediction. This type of VLM can be concluded as Perceiver-LLM architecture, while the Perceiver usually comprises a visual module and the adapter. Additionally, Cambrain-1 [27] introduces more visual modules to collaboratively provide more useful visual tokens in a vision-centric paradigm. In the other research line, unified-modeling VLMs represented by Chameleon [25], Unified-IO 2 [18], and Emu3 [30] are designed to generate texts, images, and so on uniformly. As these models apply codebook [11] to tokenize images into discrete vowels, their training processes are supervised by both vision and text information. Unified-modeling VLMs are expected to develop more stable multimodal understanding abality [10]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 534, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 534, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 534, + 85 + ], + "type": "text", + "content": "2.2. Multimodal Chain-of-Thought Prompting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 90, + 553, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 553, + 305 + ], + "type": "text", + "content": "Similar to CoT used in LLMs, multimodal Chain-of-Thought prompting methods [12, 19, 28, 32, 34] aim to augment the reasoning ability of VLM by generating intermediate reasoning steps. A series of studies focus on providing VLMs with fine-grained textual information, such as detailed description [28]. Compositional CoT (CCoT) [19] prompts VLMs to generate a Scene Graph (SG), which is a JSON-like description containing compositional information of objects that occurred in the image. DDCoT [34] deconstructs problems into small problems, requiring VLMs to solve them respectively and then inferring the final answer. In the other research line, Set-of-Marks prompting [32] augments the objects in the image to help VLMs recognize them. SCAFFOLD [12] overlays coordinate onto images to prompt VLMs with relative position information, and VLMs leverage overlayed textual coordinates to implicitly represent corresponding regions of the image to perform reasoning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 306, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 553, + 378 + ], + "type": "text", + "content": "However, these methods still produce text-only rationales to infer the final answer. These generated textual rationales usually struggle to express the fine-grained associations with the origin image. We thereby propose ICoT to elicit VLMs to generate interleaved visual-textual reasoning steps to effectively reach the final outputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 389, + 397, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 397, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 397, + 403 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 410, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 518 + ], + "type": "text", + "content": "To address the limitations that current multimodal CoT methods are still stuck in generating text-only rationales to infer the final answer, we propose interleaved-modal CoT (ICoT) to elicit VLMs generated multimodal reasoning steps. We start by introducing the workflow of VLMs and multimodal CoT in Section 3.1. We then introduce the concept of ICoT in Section 3.2. Finally, we propose a plug-and-play method, Attention-driven Selection (ADS), to realize ICoT on existing VLMs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 526, + 400, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 400, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 400, + 538 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 544, + 553, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 553, + 567 + ], + "type": "text", + "content": "We first recall some background of VLMs and multimodal CoT in this section." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 584, + 553, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 553, + 656 + ], + "type": "text", + "content": "Vision-Language Model. VLMs usually consist of a visual encoder " + }, + { + "bbox": [ + 313, + 584, + 553, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 584, + 553, + 656 + ], + "type": "text", + "content": " and a generative large language model LLM, and they determine where to insert images according to visual holders inserted in the text instructions. Then, VLMs take the image and the instructions as input and respond with a final answer" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 353, + 666, + 553, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 666, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 353, + 666, + 553, + 680 + ], + "type": "interline_equation", + "content": "\\text {a n s w e r} = \\mathbf {V L M} (\\text {I m a g e}, \\text {I n s t r u c t i o n}). \\tag {1}", + "image_path": "0f3a344ee272c6a8a62d5d438819ccae9bbb66740be7a9dd233352b1f7cb4eee.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Specifically, the visual encoder " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " extracts visual tokens " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{v}^{l\\times d}" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " from the image " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " is the length of visual" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19522" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 206, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 206, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 206, + 83 + ], + "type": "text", + "content": "Algorithm 1 Interleaved-modal CoT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 87, + 296, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 87, + 296, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 87, + 296, + 113 + ], + "type": "text", + "content": "1: Input: Word embeddings " + }, + { + "bbox": [ + 61, + 87, + 296, + 113 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 61, + 87, + 296, + 113 + ], + "type": "text", + "content": ", Visual tokens " + }, + { + "bbox": [ + 61, + 87, + 296, + 113 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 61, + 87, + 296, + 113 + ], + "type": "text", + "content": ", Selected number " + }, + { + "bbox": [ + 61, + 87, + 296, + 113 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 61, + 87, + 296, + 113 + ], + "type": "text", + "content": ", Signal tokens " + }, + { + "bbox": [ + 61, + 87, + 296, + 113 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 61, + 87, + 296, + 113 + ], + "type": "text", + "content": ", Stopping criteria " + }, + { + "bbox": [ + 61, + 87, + 296, + 113 + ], + "type": "inline_equation", + "content": "SC" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 123, + 232, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 123, + 232, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 123, + 232, + 134 + ], + "type": "text", + "content": "2: Output: Generated Response Answer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 135, + 293, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 135, + 293, + 148 + ], + "spans": [ + { + "bbox": [ + 62, + 135, + 293, + 148 + ], + "type": "text", + "content": "3: predicted_tokens " + }, + { + "bbox": [ + 62, + 135, + 293, + 148 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 62, + 135, + 293, + 148 + ], + "type": "text", + "content": " [] ▷ Initialize as an empty list" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 148, + 107, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 107, + 159 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 107, + 159 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 62, + 148, + 107, + 159 + ], + "type": "inline_equation", + "content": "\\mathcal{V}\\gets f_v" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 159, + 295, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 295, + 182 + ], + "type": "text", + "content": "5: inputs = Initilize " + }, + { + "bbox": [ + 62, + 159, + 295, + 182 + ], + "type": "inline_equation", + "content": "(f_e, f_v)" + }, + { + "bbox": [ + 62, + 159, + 295, + 182 + ], + "type": "text", + "content": " Initialize inputs for prefilling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 183, + 161, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 183, + 161, + 194 + ], + "spans": [ + { + "bbox": [ + 62, + 183, + 161, + 194 + ], + "type": "text", + "content": "6: while " + }, + { + "bbox": [ + 62, + 183, + 161, + 194 + ], + "type": "inline_equation", + "content": "SC" + }, + { + "bbox": [ + 62, + 183, + 161, + 194 + ], + "type": "text", + "content": " not met do" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 195, + 271, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 195, + 271, + 206 + ], + "spans": [ + { + "bbox": [ + 62, + 195, + 271, + 206 + ], + "type": "text", + "content": "7: next_token, attention_map = model(inputs)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 207, + 251, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 207, + 251, + 218 + ], + "spans": [ + { + "bbox": [ + 62, + 207, + 251, + 218 + ], + "type": "text", + "content": "8: Append next_token to predicted_tokens" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 218, + 220, + 292, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 220, + 292, + 231 + ], + "spans": [ + { + "bbox": [ + 218, + 220, + 292, + 231 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 218, + 220, + 292, + 231 + ], + "type": "text", + "content": " ADS judgement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 232, + 207, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 207, + 242 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 207, + 242 + ], + "type": "text", + "content": "9: if predicted_tokens = S then" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 243, + 296, + 349 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 59, + 243, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 243, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 59, + 243, + 295, + 266 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 59, + 243, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{selected}} = \\mathrm{ADS}(\\mathcal{V},\\mathrm{attention\\_map},n)" + }, + { + "bbox": [ + 59, + 243, + 295, + 266 + ], + "type": "text", + "content": " Apply Attention-driven Selection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 267, + 247, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 267, + 247, + 278 + ], + "spans": [ + { + "bbox": [ + 59, + 267, + 247, + 278 + ], + "type": "text", + "content": "11: Append " + }, + { + "bbox": [ + 59, + 267, + 247, + 278 + ], + "type": "inline_equation", + "content": "\\nu_{\\mathrm{selected}}" + }, + { + "bbox": [ + 59, + 267, + 247, + 278 + ], + "type": "text", + "content": " to predicted_tokens" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 279, + 110, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 279, + 110, + 289 + ], + "spans": [ + { + "bbox": [ + 59, + 279, + 110, + 289 + ], + "type": "text", + "content": "12: end if" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 291, + 296, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 291, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 59, + 291, + 296, + 314 + ], + "type": "text", + "content": "13: inputs = Update(inputs, predicted_tokens) > Updates inputs for next step generation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 316, + 116, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 316, + 116, + 325 + ], + "spans": [ + { + "bbox": [ + 59, + 316, + 116, + 325 + ], + "type": "text", + "content": "14: end while" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 327, + 269, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 327, + 269, + 338 + ], + "spans": [ + { + "bbox": [ + 59, + 327, + 269, + 338 + ], + "type": "text", + "content": "15: Answer = Tokenizer Decode(predicted_tokens)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 339, + 140, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 339, + 140, + 349 + ], + "spans": [ + { + "bbox": [ + 59, + 339, + 140, + 349 + ], + "type": "text", + "content": "16: return Answer" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 373, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 421 + ], + "type": "text", + "content": "tokens and " + }, + { + "bbox": [ + 55, + 373, + 296, + 421 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 373, + 296, + 421 + ], + "type": "text", + "content": " is the dimensions of the hidden states of the LLM. The built-in LLM predicts next-tokens in a left-to-right fashion according to visual tokens " + }, + { + "bbox": [ + 55, + 373, + 296, + 421 + ], + "type": "inline_equation", + "content": "f_{v}^{l \\times d}" + }, + { + "bbox": [ + 55, + 373, + 296, + 421 + ], + "type": "text", + "content": " and the instructions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "text", + "content": "Multimodal CoT. Compared with the direct prediction described in Eqn. 1, Multimodal CoT further introduces a prompt to elicit VLMs to generate a series of intermediate textual rationales " + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "inline_equation", + "content": "(r_1, r_2, \\ldots)" + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "text", + "content": " before the final answer:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "inline_equation", + "content": "r_1, r_2, \\ldots" + }, + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "text", + "content": ", answer " + }, + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "inline_equation", + "content": "= \\mathbf{VLM}(\\text{Prompt}, \\text{Image}, \\text{Instruction})" + }, + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "text", + "content": ". (2)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "text", + "content": "Technically, the prompt could be represented as a sequence of demonstrations, each consisting of a triple: (Image, Rationale, Answer). Alternatively, an explicit instruction, such as \"Let's think step by step,\" could also serve as the prompt." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 587, + 253, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 253, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 253, + 600 + ], + "type": "text", + "content": "3.2. Interleaved-modal Chain-of-Thought" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Previous multimodal CoT prompting methods only produce text-only reasoning steps to improve the reasoning ability of VLMs. These intermediate steps are generated according to the entire image, which are struggle to express exact fine-grained associations with the original image. Given these limitations, we propose a more advanced Interleaved-modal Chain-of-Thought (ICoT) prompting, aiming to elicit VLMs to generate a series of multimodal intermediate reasoning steps, each consisting of paired im" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 72, + 478, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 478, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 478, + 83 + ], + "type": "text", + "content": "Algorithm 2 Attention-driven Selection" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 319, + 87, + 555, + 218 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 319, + 87, + 555, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 87, + 555, + 110 + ], + "spans": [ + { + "bbox": [ + 319, + 87, + 555, + 110 + ], + "type": "text", + "content": "1: Input: Attention map " + }, + { + "bbox": [ + 319, + 87, + 555, + 110 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 319, + 87, + 555, + 110 + ], + "type": "text", + "content": ", Selected number " + }, + { + "bbox": [ + 319, + 87, + 555, + 110 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 319, + 87, + 555, + 110 + ], + "type": "text", + "content": ", Visual tokens " + }, + { + "bbox": [ + 319, + 87, + 555, + 110 + ], + "type": "inline_equation", + "content": "f_{v}" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 111, + 535, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 111, + 535, + 123 + ], + "spans": [ + { + "bbox": [ + 320, + 111, + 535, + 123 + ], + "type": "text", + "content": "2: Output: Fine-grained visual information " + }, + { + "bbox": [ + 320, + 111, + 535, + 123 + ], + "type": "inline_equation", + "content": "\\nu_{\\mathrm{selected}}" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 124, + 554, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 124, + 554, + 135 + ], + "spans": [ + { + "bbox": [ + 321, + 124, + 554, + 135 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 321, + 124, + 554, + 135 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{selected}}\\gets \\emptyset" + }, + { + "bbox": [ + 321, + 124, + 554, + 135 + ], + "type": "text", + "content": " Initialize as an empty set" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 136, + 432, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 136, + 432, + 148 + ], + "spans": [ + { + "bbox": [ + 321, + 136, + 432, + 148 + ], + "type": "text", + "content": "4: Indices " + }, + { + "bbox": [ + 321, + 136, + 432, + 148 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 321, + 136, + 432, + 148 + ], + "type": "text", + "content": " TopK(At,n)" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 148, + 407, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 148, + 407, + 157 + ], + "spans": [ + { + "bbox": [ + 321, + 148, + 407, + 157 + ], + "type": "text", + "content": "5: for " + }, + { + "bbox": [ + 321, + 148, + 407, + 157 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 321, + 148, + 407, + 157 + ], + "type": "text", + "content": " in Indices do" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 321, + 158, + 453, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 158, + 453, + 171 + ], + "spans": [ + { + "bbox": [ + 321, + 158, + 453, + 171 + ], + "type": "text", + "content": "6: Append " + }, + { + "bbox": [ + 321, + 158, + 453, + 171 + ], + "type": "inline_equation", + "content": "f_{v}^{l \\times d}[i]" + }, + { + "bbox": [ + 321, + 158, + 453, + 171 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 321, + 158, + 453, + 171 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\text{selected}}" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 321, + 172, + 365, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 172, + 365, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 172, + 365, + 182 + ], + "type": "text", + "content": "7: end for" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 321, + 183, + 553, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 183, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 321, + 183, + 553, + 206 + ], + "type": "text", + "content": "8: Restore " + }, + { + "bbox": [ + 321, + 183, + 553, + 206 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}_{\\mathrm{selected}}" + }, + { + "bbox": [ + 321, + 183, + 553, + 206 + ], + "type": "text", + "content": ", Indices) " + }, + { + "bbox": [ + 321, + 183, + 553, + 206 + ], + "type": "inline_equation", + "content": ">" + }, + { + "bbox": [ + 321, + 183, + 553, + 206 + ], + "type": "text", + "content": " Restore relative positions in-place" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 321, + 207, + 400, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 207, + 400, + 218 + ], + "spans": [ + { + "bbox": [ + 321, + 207, + 400, + 218 + ], + "type": "text", + "content": "9: return " + }, + { + "bbox": [ + 321, + 207, + 400, + 218 + ], + "type": "inline_equation", + "content": "\\nu_{\\mathrm{selected}}" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 232, + 551, + 350 + ], + "blocks": [ + { + "bbox": [ + 317, + 232, + 551, + 350 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 551, + 350 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 551, + 350 + ], + "type": "image", + "image_path": "c5d74196e5f92c316d404a4a83d73005768f844ef2700b28accda62fc1a2e011.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 357, + 555, + 389 + ], + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 389 + ], + "type": "text", + "content": "Figure 2. The workflow of ADS selecting fine-grained visual rationales. Signal attention represents the attention map of the signal token overall visual tokens." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 412, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 519 + ], + "type": "text", + "content": "age and textual rationale. Generated intermediate reasoning steps formulating interleaved-modal rationales to effectively lead to the final outputs. In this paper, we consider the visual rationales in interleaved-modal rationales as fine-grained visual information " + }, + { + "bbox": [ + 313, + 412, + 555, + 519 + ], + "type": "inline_equation", + "content": "x_{v}^{h' \\times w'}" + }, + { + "bbox": [ + 313, + 412, + 555, + 519 + ], + "type": "text", + "content": " extracted from an image " + }, + { + "bbox": [ + 313, + 412, + 555, + 519 + ], + "type": "inline_equation", + "content": "x_{v}^{h \\times w1}" + }, + { + "bbox": [ + 313, + 412, + 555, + 519 + ], + "type": "text", + "content": ". These visual rationales capture relevant details in the image, such as objects, colors, and texts, interleaved with the following generated textual rationale to infer the final answer:" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 323, + 527, + 554, + 565 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 527, + 554, + 565 + ], + "spans": [ + { + "bbox": [ + 323, + 527, + 554, + 565 + ], + "type": "interline_equation", + "content": "\\begin{array}{r} r _ {1}, x _ {v _ {1}}, r _ {2}, x _ {v _ {2}}, \\dots , \\text {a n s w e r} = \\mathbf {V L M} (\\text {P r o m p t , I m a g e}, \\\\ \\text {I n s t r u c t i o n}). \\end{array} \\tag {3}", + "image_path": "a90cf5c13289df51b656b229b669331cca2b3f0ea599e58de3d05df28a0a4004.jpg" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 314, + 573, + 460, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 573, + 460, + 585 + ], + "spans": [ + { + "bbox": [ + 314, + 573, + 460, + 585 + ], + "type": "text", + "content": "3.3. Attention-driven Selection" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 592, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 555, + 676 + ], + "type": "text", + "content": "Although the proposed ICoT is both novel and conceptually sound, current VLMs are unable to generate such fine-grained visual information. This remains true even for VLMs [18, 25, 30] that are empowered with multimodal generation ability. We thus propose to simplify the problem from fine-grained visual information generation to fine-grained visual information selection, as this information" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": "1The images in the dataset involved in this paper are RGB images by default, and the number of channels is omitted in the formulas for simplicity." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19523" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "has been naturally contained in the origin image, namely " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "x_{v}^{h^{\\prime}\\times w^{\\prime}}\\in x_{v}^{h\\times w}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "h^\\prime \\ll h" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "w^{\\prime}\\ll w" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 99, + 296, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 296, + 219 + ], + "type": "text", + "content": "Specifically, before performing next-token prediction, ICoT requires the VLM to cache visual tokens " + }, + { + "bbox": [ + 55, + 99, + 296, + 219 + ], + "type": "inline_equation", + "content": "f_{v}^{l \\times d}" + }, + { + "bbox": [ + 55, + 99, + 296, + 219 + ], + "type": "text", + "content": " extracted by its built-in visual encoder " + }, + { + "bbox": [ + 55, + 99, + 296, + 219 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(x_{v})" + }, + { + "bbox": [ + 55, + 99, + 296, + 219 + ], + "type": "text", + "content": " for further selection. In the following decoding steps, we consider the VLM deems it necessary to insert a piece of visual rationales after generating a pre-defined signal token " + }, + { + "bbox": [ + 55, + 99, + 296, + 219 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 99, + 296, + 219 + ], + "type": "text", + "content": " as shown in Figure 2, which is a natural language token that indicates the beginning of a textual rationale. Therefore, ADS will be signaled to select fine-grained visual information from " + }, + { + "bbox": [ + 55, + 99, + 296, + 219 + ], + "type": "inline_equation", + "content": "f_{v}^{l \\times d}" + }, + { + "bbox": [ + 55, + 99, + 296, + 219 + ], + "type": "text", + "content": " upon the VLM generating " + }, + { + "bbox": [ + 55, + 99, + 296, + 219 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 99, + 296, + 219 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 60, + 233, + 296, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 233, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 60, + 233, + 296, + 274 + ], + "type": "interline_equation", + "content": "\\text {d o} \\text {s e l e c t i o n} = \\left\\{ \\begin{array}{l l} \\text {T r u e} & \\text {i f p r e d i c t e d} \\\\ \\text {F a l s e} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}", + "image_path": "cf0112f10dac39385200917a10d3256cf69ef2e90e2c4e32156a0253df7edcfc.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 274, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 312 + ], + "type": "text", + "content": "Then, ADS selects " + }, + { + "bbox": [ + 55, + 274, + 296, + 312 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 274, + 296, + 312 + ], + "type": "text", + "content": " visual tokens from " + }, + { + "bbox": [ + 55, + 274, + 296, + 312 + ], + "type": "inline_equation", + "content": "f_{v}^{l \\times d}" + }, + { + "bbox": [ + 55, + 274, + 296, + 312 + ], + "type": "text", + "content": " as fine-grained visual information according to the signal attention map " + }, + { + "bbox": [ + 55, + 274, + 296, + 312 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 55, + 274, + 296, + 312 + ], + "type": "text", + "content": " at the current decoding step " + }, + { + "bbox": [ + 55, + 274, + 296, + 312 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 274, + 296, + 312 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 98, + 326, + 295, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 326, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 98, + 326, + 295, + 340 + ], + "type": "interline_equation", + "content": "\\mathcal {V} _ {\\text {s e l e c t e d}} = \\left\\{f _ {v} [ i ] \\mid i \\in \\operatorname {T o p K} \\left(A _ {t}, n\\right) \\right\\}, \\tag {5}", + "image_path": "d35cb963685b6a39fcb75e14ab49b7b9a59be6ff833f2a823e4e5a2d29e7d067.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 354, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 354, + 296, + 486 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 55, + 354, + 296, + 486 + ], + "type": "text", + "content": " is obtained by averaging the attention map between the signal token and visual tokens across all VLM layers. Up to now, the selected visual tokens are sorted by their attention scores, and we subsequently restore the relative position of " + }, + { + "bbox": [ + 55, + 354, + 296, + 486 + ], + "type": "inline_equation", + "content": "f_{v}[i]" + }, + { + "bbox": [ + 55, + 354, + 296, + 486 + ], + "type": "text", + "content": " in the origin image in place, prioritizing rows. Once fine-grained visual tokens " + }, + { + "bbox": [ + 55, + 354, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\nu_{\\mathrm{selected}}" + }, + { + "bbox": [ + 55, + 354, + 296, + 486 + ], + "type": "text", + "content": " are obtained, VLMs will take as input the concatenated " + }, + { + "bbox": [ + 55, + 354, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\nu_{\\mathrm{selected}}" + }, + { + "bbox": [ + 55, + 354, + 296, + 486 + ], + "type": "text", + "content": " and predicted_tokens, denote as Cat(predicted_tokens, " + }, + { + "bbox": [ + 55, + 354, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\nu_{\\mathrm{selected}}" + }, + { + "bbox": [ + 55, + 354, + 296, + 486 + ], + "type": "text", + "content": "), resuming the original autoregressive text generation process. Hence, the current decoding step is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 81, + 501, + 296, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 501, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 81, + 501, + 296, + 529 + ], + "type": "interline_equation", + "content": "\\begin{array}{r} \\text {n e x t - t o k e n} = \\mathbf {V L M} (\\text {C a t (p r e d i c t e d - t o k e n s}, \\\\ \\nu_ {\\text {s e l e c t e d}})). \\end{array} \\tag {6}", + "image_path": "545759be341199152f04a821ce5719064472fccaaa6b177bac0ccd5918b488ab.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 543, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 614 + ], + "type": "text", + "content": "Notably, to avoid misunderstanding, we first convert predicted_tokens into word embeddings " + }, + { + "bbox": [ + 55, + 543, + 295, + 614 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 55, + 543, + 295, + 614 + ], + "type": "text", + "content": ", and then concatenate " + }, + { + "bbox": [ + 55, + 543, + 295, + 614 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 55, + 543, + 295, + 614 + ], + "type": "text", + "content": " with selected fine-grained visual information " + }, + { + "bbox": [ + 55, + 543, + 295, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{selected}}" + }, + { + "bbox": [ + 55, + 543, + 295, + 614 + ], + "type": "text", + "content": " in the embedding-level. We provide a detailed description of ICoT and ADS in Algorithm 1 and Algorithm 2, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Technically, ICoT inherits the existing eliciting methods from CoT, such as attaching an instruction: \"Let's think step by step\" in zero-shot ICoT or providing few-shot examples. In the few-shot ICoT, each example consists of an image, textual rationales, and visual rationales. These examples can be manually designed to prompt VLMs regarding their way of thinking and generation formatting, among other aspects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 92, + 376, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 92, + 376, + 104 + ], + "spans": [ + { + "bbox": [ + 314, + 92, + 376, + 104 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 113, + 553, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 113, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 113, + 553, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{3}\\mathbf{C}\\mathbf{o}\\mathbf{T}" + }, + { + "bbox": [ + 313, + 113, + 553, + 209 + ], + "type": "text", + "content": " [6] is a novel multimodal CoT benchmark specifically concentrated on multi-domain, multi-reasoning-step. " + }, + { + "bbox": [ + 313, + 113, + 553, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{3}\\mathbf{C}\\mathbf{o}\\mathbf{T}" + }, + { + "bbox": [ + 313, + 113, + 553, + 209 + ], + "type": "text", + "content": " contains 267 categories from science, mathematics, and commonsense domains. As the question of each instance is relatively complex, their rationales have an average length of 293 tokens and rely more on fine-grained visual information, which can reflect the advantages of ICoT compared with previous multimodal CoT methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 211, + 553, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 553, + 258 + ], + "type": "text", + "content": "ScienceQA [24] is a popular dataset used to evaluate the reasoning ability of VLMs. We use ScienceQA to provide a general comparison between ICoT and other existing multimodal CoT methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 261, + 553, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 553, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 553, + 321 + ], + "type": "text", + "content": "LLaVA-Bench In-the-Wild (LLaVA-W) [15] evaluates VLMs' ability to respond to visual questions with detailed long-form answers, which also focus on the fine-grained visual description. The reference label of each instance is produced by GPT-4v." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 335, + 380, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 335, + 380, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 335, + 380, + 346 + ], + "type": "text", + "content": "4.2. Baselines" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 356, + 553, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 553, + 403 + ], + "type": "text", + "content": "No-CoT responds to the current input image and question directly without further prompting. The few-shot demonstrations of the direct generation mode consist of (Image, Question, Answer)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 406, + 553, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 553, + 441 + ], + "type": "text", + "content": "Multimodal CoT [33] elicits VLMs to generate a series of text-only intermediate reasoning steps to infer the final outputs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 444, + 554, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 554, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 554, + 528 + ], + "type": "text", + "content": "CCoT [19] first generates a scene graph (SG) using the VLM itself and then uses that SG in the prompt to produce a response. The SG is a JSON-like structural description of the given image with extensive compositional information of objects in the current images. Following their settings, we apply their official prompt to prompt VLMs to generate SGs and answers respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 529, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 601 + ], + "type": "text", + "content": "DDCoT [34] first prompts LLM to deconstruct the input question into a sequence of basic sub-questions and then applies a VQA model to answer these sub-questions involving visual information. In this paper, we use the VLM that plays the role of LLM in DDCoT for a fair comparison, as their original LLM is ChatGPT." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 603, + 554, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 554, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 554, + 687 + ], + "type": "text", + "content": "SCAFFOLD [12] overlays a coordinate matrix onto the input image, exactly demonstrating relative visual positions for VLMs. During reasoning, VLMs are steered to utilize these coordinates that indicate fine-grained visual information in the image to solve different vision-language tasks. We use their released scripts to add coordinates over each image and then use their official prompt to elicit VLMs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Specifically, in the few-shot scenario, the demonstrations of these baselines are human-written, aligning with ICoT." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19524" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 69, + 553, + 316 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 316 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 316 + ], + "type": "table", + "html": "
BackboneMethods0-shot1-shot
M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑
Chameleon-7BNo-CoT29.147.713.128.448.523.9
Multimodal CoT [33]28.549.020.430.650.720.6
CCoT [19]29.450.222.131.451.324.5
DDCoT [34]28.649.820.229.849.223.1
SCAFFOLD [12]29.648.521.731.147.524.7
ICoT (Ours)29.851.025.232.353.427.6
% Improve0.6%1.6%14.0%2.8%4.0%11.7%
BackboneMethods0-shot1-shot
M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑M3CoT ACC. ↑ScienceQA ACC. ↑LLaVA-W ROUGE-L↑
Qwen2-VL-7BNo-CoT43.656.332.745.464.433.5
Multimodal CoT [33]40.151.330.742.558.331.4
CCoT [19]43.356.429.444.163.833.9
DDCoT [34]42.655.231.245.764.932.8
SCAFFOLD [12]41.753.731.844.962.533.1
ICoT (Ours)44.156.834.246.065.435.7
% Improve1.1%0.7%4.6%0.6%0.7%5.3%
", + "image_path": "7ea645abd07d44b5900d0c2cc78d61f4ce523e3145bf7a43905b555c4af621e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 323, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 555, + 357 + ], + "type": "text", + "content": "Table 1. Results of ICoT and baselines based on Chameleon and Qwen2-VL, with the highest score bold. " + }, + { + "bbox": [ + 55, + 323, + 555, + 357 + ], + "type": "inline_equation", + "content": "{\\mathrm{M}}^{3}\\mathrm{{CoT}}" + }, + { + "bbox": [ + 55, + 323, + 555, + 357 + ], + "type": "text", + "content": " and ScienceQA are evaluated by accuracy, and we report the ROUGE-L score for the LLaVA-W benchmark. % improve represents the relative improvement achieved by ICoT over the previously best baseline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 377, + 164, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 164, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 164, + 391 + ], + "type": "text", + "content": "4.3. Implement Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 400, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 400, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 54, + 400, + 295, + 519 + ], + "type": "text", + "content": "We apply ICoT over Chameleon-7B [25] and Qwen2-VL7B-Instruct [29], which represents the fine-grained visual information in the form of discrete vokens and dense features. All experiments are conducted on A800 GPUs, and we evaluate ICoT under both zero- and one-shot scenarios. During generating interleaved-modal rationales, the signal token " + }, + { + "bbox": [ + 54, + 400, + 295, + 519 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 54, + 400, + 295, + 519 + ], + "type": "text", + "content": " used to trigger ADS is set to line break, i.e., \"" + }, + { + "bbox": [ + 54, + 400, + 295, + 519 + ], + "type": "inline_equation", + "content": "\\backslash n" + }, + { + "bbox": [ + 54, + 400, + 295, + 519 + ], + "type": "text", + "content": "\", by default, which semantically and empirically indicates the end of a generated rationale and the beginning of the next one." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 525, + 295, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 525, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 54, + 525, + 295, + 668 + ], + "type": "text", + "content": "VLMs insert visual tokens selected by ADS following the special token at the granularity of 64 according to posterior results shown in Table 4 of Appendix 8. Notably, to shorten the representation of an image, Qwen2-VL introduces a novel merge mechanism, and we approximately consider its patch size to be " + }, + { + "bbox": [ + 54, + 525, + 295, + 668 + ], + "type": "inline_equation", + "content": "(28 \\times 28)" + }, + { + "bbox": [ + 54, + 525, + 295, + 668 + ], + "type": "text", + "content": ". Each patch of Qwen2-VL has approximately 4 times as many pixels as a Chameleon patch " + }, + { + "bbox": [ + 54, + 525, + 295, + 668 + ], + "type": "inline_equation", + "content": "(16 \\times 16)" + }, + { + "bbox": [ + 54, + 525, + 295, + 668 + ], + "type": "text", + "content": ", which results in practical selection numbers of ADS set to 16. Considering the work of ADS requires the inner attention map, we apply the \"eager\" attention on both Chameleon-7B and Qwen2-VL, limited to the dependency of related python libraries2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 377, + 399, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 399, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 399, + 389 + ], + "type": "text", + "content": "4.4. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 399, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 399, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 312, + 399, + 555, + 638 + ], + "type": "text", + "content": "We comprehensively evaluate the performance of ICoT on top of Chameleon-7B and Qwen2-VL-7B through " + }, + { + "bbox": [ + 312, + 399, + 555, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{CoT}" + }, + { + "bbox": [ + 312, + 399, + 555, + 638 + ], + "type": "text", + "content": ", ScinceQA, and LLaVA-W in Table 1. In 0-shot settings, ICoT outperforms all baselines, including direct generation (No-CoT), CoT, CCoT, DDoT, and SCAFFOLD. Specifically, ICoT distinguishes from Multimodal CoT in terms of the modality of reasoning steps, which exhibit the advantages of interleaved-modal rationales to infer the final answer effectively. Compared with other multimodal CoT methods, the performance gains of ICoT further indicate that interleaved-modal rationales are more reasonable in intuition and effect than plainly inserted scene graphs (CCoT) and deconstructed sub-questions (DDoS). In 1-shot settings, ICoT demonstrations contain manually selected fine-grained visual information, while their text rationales are kept the same as other baselines. The performance gains compared with 0-shot ICoT indicate that our manually designed fine-grained ICoT demonstrations potentially guide VLMs to think in this format. In Table 4, we rigorously ablate the effectiveness of fine-grained ICoT demonstrations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Additionally, ICoT achieves the most relative performance gains in the LLaVA-W benchmark as the reference labels contain details sourced from images. These substantial performance gains compared with other baselines prove that visual tokens selected by ADS effectively capture the fine-grained visual information of an image, aiding the gen" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "text", + "content": "Using attn_implementation=\"eager\" when loading the model from HuggingFace." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19525" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 209, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 209, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 209, + 85 + ], + "type": "text", + "content": "eration of high-quality text rationales." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 149, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 149, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 149, + 105 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 111, + 295, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 111, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 111, + 295, + 302 + ], + "type": "text", + "content": "We ablate ICoT to verify the effectiveness of each portion across three benchmarks in Table 2 with the following settings: (1). w/o ADS: VLMs generate text-only rationales. (2). w/o FVI: Patches inserted in the demonstration are randomly sampled. Results indicate that both ADS and fine-grained visual information (FVI) incorporated in the demonstration are necessary. In particular, interleaved-modal rationales exhibit substantial advantages in generating high-quality textual rationales compared with text-only rationales (w/o ADS). When substituting ICoT demonstrations with normal ones (w/o FVI), the performance degradation proves that fine-grained visual information in demonstrations effectively guides VLMs to think in this format. Compared with the performance difference between removing ADS and FVI, we find that generating paired visual and text rationales boosts more improvements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 303, + 295, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 295, + 388 + ], + "type": "text", + "content": "Additionally, the performance gap is relatively smooth in ScienceQA and more dramatic on " + }, + { + "bbox": [ + 55, + 303, + 295, + 388 + ], + "type": "inline_equation", + "content": "\\mathrm{M}^3\\mathrm{CoT}" + }, + { + "bbox": [ + 55, + 303, + 295, + 388 + ], + "type": "text", + "content": " and LLaVA-W. We attribute this to the ScienceQA dataset being relatively easier than others since both " + }, + { + "bbox": [ + 55, + 303, + 295, + 388 + ], + "type": "inline_equation", + "content": "\\mathrm{M}^3\\mathrm{CoT}" + }, + { + "bbox": [ + 55, + 303, + 295, + 388 + ], + "type": "text", + "content": " and the answers of LLaVA-W highly rely on the fine-grained visual information of an image. Therefore, our proposed ICoT has the potential to solve complex vision-language tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 162, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 162, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 162, + 407 + ], + "type": "text", + "content": "4.6. In-depth Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 414, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 295, + 700 + ], + "type": "text", + "content": "Analysis on realizing ICoT via KV Cache Up to now, the fine-grained visual information of ICoT is provided at the input end via discrete tokens or dense visual tokens, which brings more computation. After rethinking the generating process of an autoregressive model, there are other inputs that are proposed to avoid repeated computation, namely, the Key-Value (KV) Cache. The input image was stored in the KV Cache during the prefilling phase before generating multimodal intermediate reasoning steps in a left-to-right fashion. Therefore, copying the KV cache of fine-grained visual information enables ICoT with reduced computational costs, as visual information does not require extra forward propagation. As shown in Table 3, copying the KV Cache brings performance degradation compared with providing visual information at the input end. We attribute this phenomenon to the fact that although copying KV Cache indeed makes VLMs attend more to the same region as ADS, the optimal visual information is highlighted in a position-agnostic case, determined by the nature of KV Cache. Specifically, this degrades the original interleaved-modal rationales into non-interleaved ones as position information is early fused into KV Cache, and thus the copied ones are inherently insensitive to the position of textual rationale." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "However, considering it brought slight performance" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 320, + 72, + 552, + 133 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 552, + 133 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 552, + 133 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 552, + 133 + ], + "type": "table", + "html": "
MethodsM3CoTScienceQALLaVA-W
ICoT32.353.427.6
w/o ADS29.2 (-3.1)52.4(-1.0)24.5(-3.1)
w/o FVI30.6 (-1.8)52.8(-0.6)25.9(-1.7)
w/o ADS+FVI29.1(-3.2)51.0(-2.4)23.0(-4.6)
", + "image_path": "fc6b090a6fe9cc20d05de99ede5d8023b74c0045b71c98f810d4042580981a2b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 320, + 213, + 552, + 278 + ], + "blocks": [ + { + "bbox": [ + 313, + 144, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 200 + ], + "type": "text", + "content": "Table 2. Ablation studies of 1-shot ICoT on Chameleon-7B. () describes the performance degradation compared with ICoT. FVI indicates the 1-shot demonstration contains fine-grained visual information. ADS indicates that VLMs generate interleaved-modal reasoning steps." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 213, + 552, + 278 + ], + "lines": [ + { + "bbox": [ + 320, + 213, + 552, + 278 + ], + "spans": [ + { + "bbox": [ + 320, + 213, + 552, + 278 + ], + "type": "table", + "html": "
Dataset0-shot1-shot
KV-CopyICoTKV-CopyICoT
M3CoT29.129.831.532.3
ScienceQA49.751.052.953.4
LLaVA-W24.725.227.027.6
", + "image_path": "676c7cef73ff2ff59c01a04621e776fb724bc41b323aa0fa61e4783ff14cb381.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 320, + 324, + 552, + 364 + ], + "blocks": [ + { + "bbox": [ + 313, + 289, + 554, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 312 + ], + "type": "text", + "content": "Table 3. Results comparison between copying KV Cache (KV-Copy) and inserting selected patches." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 324, + 552, + 364 + ], + "lines": [ + { + "bbox": [ + 320, + 324, + 552, + 364 + ], + "spans": [ + { + "bbox": [ + 320, + 324, + 552, + 364 + ], + "type": "table", + "html": "
MethodsM3CoTScienceQALLaVA-W
Human-written32.353.427.6
Model-written31.5(-0.8)51.8(-1.6)26.7(-0.8)
", + "image_path": "8bdf56f5901a5e118a20b477e55b77d6e18bbb1b7f5f7e6dc2a055ee3ae70347.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 374, + 554, + 440 + ], + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 440 + ], + "type": "text", + "content": "Table 4. Results concerning the design of demonstrations on Chameleon-7B. () describes the performance degradation compared with ICoT. Human-written indicates that demonstrations are manually designed with fine-grained visual information inserted, and Model-written indicates that the VLM generates both visual and textual rationale via ICoT." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 462, + 554, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 499 + ], + "type": "text", + "content": "degradation and factually reduced computation costs, we believe this exploration is still valuable, and we call for more interesting exploration in realizing ICoT." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 514, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 658 + ], + "type": "text", + "content": "Analysis on the Demonstrations We also attempt to let VLMs generate the demonstrations via themselves (Automatic at the bottom of Table 4). Results indicate that using the automatically generated demonstrations also brings performance degradation compared with ICoT using manually designed ones. We consider it is caused by the fact that formulating a continuous sub-image through ADS is nontrivial, and some discrete patches inevitably introduce additional noise. Therefore, considering that designing such a demonstration is not time-consuming, ICoT utilizes manually designed ones to elicit VLMs to perform ICoT for better performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 669, + 388, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 388, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 388, + 682 + ], + "type": "text", + "content": "5. Case Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "In this section, we empirically illustrate the advantages of ICoT via three case studies in Figure 3. These case studies" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19526" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 549, + 246 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 549, + 246 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 549, + 246 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 549, + 246 + ], + "type": "image", + "image_path": "70796664c06c33732a7065f24974cfdd8a99c09ec9a847f3cef367a45cd3bfd5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 555, + 289 + ], + "type": "text", + "content": "Figure 3. Case studies between ICoT and multimodal CoT with text-only rationales on Chameleon. Three cases are selected according to three typical problems in text-only problems: misunderstanding, overgeneralization, and hallucination. Red blocks indicate the incorrect rationales." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 309, + 295, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 295, + 346 + ], + "type": "text", + "content": "focused on three typical problems that occurred in text-only rationales, namely, misunderstanding (top), overgeneralization (middle), and hallucination (bottom)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 625 + ], + "type": "text", + "content": "Specifically, in the first case, interleaved-modal CoT first recognizes three different objects via captions: \"inflatable castle, crayons, and a parachute\". Then, in the second reasoning step, ADS inserts selected patches from the scheduled objects to elicit the VLM to conclude their common property, and VLM infers a correct answer. Text-only CoT misunderstands that the three objects are all colored pencils, ignoring the castle and the parachute, even the final answer is correct. In the second case, text-only rationales overgeneralize flying a kite to a kite festival, leading to a wrong answer. ICoT first recognizes a man standing in a field of grass and then infers it is a windy day according to a few kites in the sky. In the last case, it provides the other typical error of text-only CoT, namely, hallucination. As text-only CoT purely relies on language reasoning ability, VLMs have the potential to imagine something not mentioned in the image, resulting in a wrong answer. ICoT first infers from the street sign that there may be a troll attraction nearby according to patches of the indicator inserted by ADS. Then, the ADS helps the VLMs to attend to the troll statue under the bridge and infer it is likely the mentioned attraction, finally arriving at the correct answer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "Even though the above case studies exhibit the advantages of ICoT, ADS still brings potential problems. For example, ADS is triggered to select patches when VLM generates a pre-defined signal token. This simple mechanism is a double-edged sword in that VLMs will generate low-quality responses if this token is generated with a high frequency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 308, + 388, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 308, + 388, + 321 + ], + "spans": [ + { + "bbox": [ + 314, + 308, + 388, + 321 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 347, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 552 + ], + "type": "text", + "content": "In this paper, we first propose interleaved-modal CoT (ICoT), which generates interleaved-modal rationales to infer the final answer effectively. In light of the challenges of applying ICoT on existing VLMs, we then introduce Attention-driven Selection (ADS), a plug-and-play strategy to identify optimal patches from the image without being parameterized. We evaluate ICoT on Chameleon-7B and Qwen2-VL-7B-Instruct, representing VLMs of two architectures. Extensive experiments conducted on " + }, + { + "bbox": [ + 313, + 347, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{CoT}" + }, + { + "bbox": [ + 313, + 347, + 555, + 552 + ], + "type": "text", + "content": ", ScienceQA, and LLaVA-W, under both zero- and few-shot scenarios, have proven that ICoT achieves substantial performance (up to " + }, + { + "bbox": [ + 313, + 347, + 555, + 552 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 313, + 347, + 555, + 552 + ], + "type": "text", + "content": ") compared with the existing multimodal CoT methods. Additionally, in the analysis section, we conduct a preliminary exploration of implementing ICoT by copying the KV cache of optimal visual tokens and explain the inner trade-off between efficiency and performance in this approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Although ICoT has proven its effectiveness in this paper, we consider that ICoT still has significant potential for further improvement. The patch selection in ADS requires storing attention scores, which brings additional memory overhead. Moreover, the fixed number of selected patches in the ADS design is sub-optimal, resulting in unexpected outputs for VLMs. To address these, we intend to incorporate established techniques from segmentation or grounding methods to create a more robust implementation of ICoT. In the future, we also plan to evaluate it across additional backbones and benchmarks to better assess its generalization ability." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19527" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "I would like to express my sincere gratitude to all the authors and reviewers for their valuable contributions to this research. This work was supported by the National Natural Science Foundation of China (NSFC 62106165) and the Project Funded by the Priority Academic Program Development of Jiangsu Higher Education Institutions, China." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 194, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 194, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 228 + ], + "type": "text", + "content": "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 229, + 296, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 229, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 61, + 229, + 296, + 284 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 295, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 318 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 318 + ], + "type": "text", + "content": "[3] BIG bench authors. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on Machine Learning Research, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 342 + ], + "type": "text", + "content": "[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 343, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 343, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 343, + 295, + 364 + ], + "type": "text", + "content": "[5] Claus Bundesen. A theory of visual attention. Psychological review, 97(4):523, 1990. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 365, + 295, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 295, + 410 + ], + "type": "text", + "content": "[6] Qiguang Chen, Libo Qin, Jin Zhang, Zhi Chen, Xiao Xu, and Wanxiang Che. M3cot: A novel benchmark for multidomain multi-step multi-modal chain-of-thought. arXiv preprint arXiv:2405.16473, 2024. 1, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 411, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 295, + 453 + ], + "type": "text", + "content": "[7] Ethan Chern, Jiadi Su, Yan Ma, and Pengfei Liu. Anole: An open, autoregressive, native large multimodal models for interleaved image-text generation. arXiv preprint arXiv:2407.06135, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 456, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 295, + 521 + ], + "type": "text", + "content": "[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 523, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 295, + 578 + ], + "type": "text", + "content": "[9] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "type": "text", + "content": "[10] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. arXiv preprint arXiv:2406.11832, 2024.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "text", + "content": "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Xuanyu Lei, Zonghan Yang, Xinrui Chen, Peng Li, and Yang Liu. Scaffolding coordinates to promote vision-language coordination in large multi-modal models. arXiv preprint arXiv:2402.12058, 2024. 1, 3, 5, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "type": "text", + "content": "[14] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 153, + 553, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 187 + ], + "type": "text", + "content": "[15] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.3,5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 189, + 553, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 189, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 553, + 221 + ], + "type": "text", + "content": "[16] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 223, + 553, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 553, + 267 + ], + "type": "text", + "content": "[17] Jiacheng Liu, Alisa Liu, Ximing Lu, Sean Welleck, Peter West, Ronan Le Bras, Yejin Choi, and Hannaneh Hajishirzi. Generated knowledge prompting for commonsense reasoning. arXiv preprint arXiv:2110.08387, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 269, + 554, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 269, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 317, + 269, + 554, + 335 + ], + "type": "text", + "content": "[18] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26439-26455, 2024. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 336, + 553, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 336, + 553, + 391 + ], + "spans": [ + { + "bbox": [ + 317, + 336, + 553, + 391 + ], + "type": "text", + "content": "[19] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 1, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 393, + 507, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 507, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 507, + 404 + ], + "type": "text", + "content": "[20] OpenAI. Learning to reason with llms, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 406, + 553, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 406, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 317, + 406, + 553, + 438 + ], + "type": "text", + "content": "[21] Michael I Posner, Steven E Petersen, et al. The attention system of the human brain. Annual review of neuroscience, 13(1):25-42, 1990. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 441, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 441, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 317, + 441, + 553, + 496 + ], + "type": "text", + "content": "[22] Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 498, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 498, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 498, + 553, + 552 + ], + "type": "text", + "content": "[23] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 554, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 554, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 317, + 554, + 553, + 597 + ], + "type": "text", + "content": "[24] Tanik Saikh, Tirthankar Ghosal, Amish Mittal, Asif Ekbal, and Pushpak Bhattacharyya. Scienceqa: A novel resource for question answering on scholarly articles. International Journal on Digital Libraries, 23(3):289-301, 2022. 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 600, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 600, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 600, + 553, + 632 + ], + "type": "text", + "content": "[25] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2,3,4,6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[26] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[27] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang," + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19528" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 507 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 172 + ], + "type": "text", + "content": "[28] Lei Wang, Yi Hu, Jiabang He, Xing Xu, Ning Liu, Hui Liu, and Heng Tao Shen. T-sciq: Teaching multimodal chain-of-thought reasoning via large language model signals for science question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19162-19170, 2024. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 250 + ], + "type": "text", + "content": "[29] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 294, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 305 + ], + "type": "text", + "content": "[30] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 307, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 307, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 307, + 294, + 361 + ], + "type": "text", + "content": "[31] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 363, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 363, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 363, + 294, + 407 + ], + "type": "text", + "content": "[32] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 408, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 408, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 294, + 452 + ], + "type": "text", + "content": "[33] Zhuosheng Zhang, Aston Zhang, Mu Li, George Karypis, Alex Smola, et al. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2024. 1, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 453, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 507 + ], + "type": "text", + "content": "[34] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168-5191, 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19529" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/86d682d5-19b8-4583-bdec-4fcac9608ef2_content_list.json b/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/86d682d5-19b8-4583-bdec-4fcac9608ef2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..21a21a2875b8f32e6c8556fc8b2c1287dac6a124 --- /dev/null +++ b/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/86d682d5-19b8-4583-bdec-4fcac9608ef2_content_list.json @@ -0,0 +1,1741 @@ +[ + { + "type": "text", + "text": "Interpretable Generative Models through Post-hoc Concept Bottlenecks", + "text_level": 1, + "bbox": [ + 138, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Akshay Kulkarni, Ge Yan,\\* Chung-En Sun\\* Tuomas Oikarinen, and Tsui-Wei Weng \nUniversity of California San Diego", + "bbox": [ + 171, + 180, + 823, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{a2kulkarni, lweng}@ucsd.edu", + "bbox": [ + 372, + 219, + 620, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Concept bottleneck models (CBM) aim to produce inherently interpretable models that rely on human-understandable concepts for their predictions. However, existing approaches to design interpretable generative models based on CBMs are not yet efficient and scalable, as they require expensive generative model training from scratch as well as real images with labor-intensive concept supervision. To address these challenges, we present two novel and low-cost methods to build interpretable generative models through post-hoc techniques and we name our approaches: concept-bottleneck autoencoder (CB-AE) and concept controller (CC). Our proposed approaches enable efficient and scalable training without the need of real data and require only minimal to no concept supervision. Additionally, our methods generalize across modern generative model families including generative adversarial networks and diffusion models. We demonstrate the superior interpretability and steerability of our methods on numerous standard datasets like CelebA, CelebA-HQ, and CUB with large improvements (average $\\sim 25\\%$ ) over the prior work, while being $4 - 15\\times$ faster to train. Finally, a large-scale user study is performed to validate the interpretability and steerability of our methods.", + "bbox": [ + 89, + 301, + 485, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 652, + 220, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep generative models [36-38] have become increasingly powerful and widely used in many high-stakes domains and applications, including realistic data generation [11], simulating hypothetical scenarios or environments [19], and scientific discovery [1]. It is therefore important to ensure that the generation process is interpretable, which will allow us to understand and audit the generation, and further mitigate potential biases and harms (e.g. content moderation).", + "bbox": [ + 89, + 674, + 483, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, most of the advances in deep learning utilize complex, black-box neural network architectures that are difficult to interpret and understand. This leads to user mistrust of model predictions due to the absence of explanations.", + "bbox": [ + 89, + 797, + 483, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A. Prior Work (Concept Bottleneck Generative Models)", + "bbox": [ + 519, + 270, + 872, + 282 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/be9466677132d035e0060822dda2d2d51546b062bbba8e437b494ff60fed0341.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 285, + 892, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Expensive generative model training from scratch!", + "bbox": [ + 545, + 352, + 864, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "B. Ours (Post-Hoc Interpretable Generative Models)", + "bbox": [ + 522, + 367, + 851, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1cd5a99936fa27a64d82cc0a2b996e7db5f7bfda7726c968421616b04675b215.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 381, + 892, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Efficiently train only CB-AE/CC with frozen generative model", + "bbox": [ + 519, + 511, + 900, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "C. Qualitative Results", + "bbox": [ + 521, + 526, + 663, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0611e74cc404650e3df77600b540932c17935f3b555392ebecd2864825994393.jpg", + "image_caption": [ + "Figure 1. A. Prior work on interpretable generative models requires expensive generative model training from scratch. B. Our CB-AE and CC can be trained efficiently for post-hoc interpretability in a pretrained, frozen generative model $g_{2} \\circ g_{1}$ . C. Example concept intervention with CB-AE and corresponding concept vectors." + ], + "image_footnote": [], + "bbox": [ + 519, + 541, + 903, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this, there has been work on developing inherently interpretable deep vision models using concept bottlenecks [20, 32, 39, 43-45]. These approaches train a concept bottleneck layer after the feature extractor (backbone) to embed a set of human-understandable concepts, followed by an interpretable sparse linear layer for the final classification based on the concept prediction. However, current development of CBMs is primarily focused on classification tasks, and only one prior work, CBGM [16], has extended it to image generation, indicating this area is under-explored.", + "bbox": [ + 511, + 715, + 906, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The key idea of the seminal work CBGM [16] is to represent concepts as learnable embeddings [46] at an inter", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 109, + 875, + 215, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code: github.com/Trustworthy-ML-Lab/posthoc-generative-cbm", + "bbox": [ + 109, + 887, + 457, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8162", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/d6a58747dc7c1bedc0b3698f2dd45bda3485a2c706d00b1a3f25730dbadd0b6c.jpg", + "table_caption": [ + "Table 1. Comparison of our CB-AE and CC with prior work CBGM [16] (green indicates desirable properties). CC trades-off inherent interpretability for better steerability, image quality, and faster training. We could not reproduce CBGM [16] to evaluate concept accuracy." + ], + "table_footnote": [], + "table_body": "
MethodPost-hoc trainingTraining without concept-labeled real imagesInherently interpretable modelConcept Acc. (%)Steerability (%) (↑)FID (↓)Train time (V100-hrs) (↓)
StyleGAN2DDPM
CBGM [16]XX✓-25.609.1050240
CB-AE (Ours)✓✓✓86.5647.34 (+21.74)9.5214 (3.5× faster)29.5 (8.1× faster)
CC (Ours)✓✓X87.6551.14 (+25.54)7.656 (8.3× faster)8.3 (28.9× faster)
", + "bbox": [ + 93, + 119, + 903, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mediate location in the generative model, and combine the embeddings to compute the generative model latent. However, the CBGM-based generative model has to be trained from scratch using concept-labeled real images, which could be difficult to scale and computationally intensive (e.g. 240 V100-hours for DDPM- $256 \\times 256$ [15]) as shown in Fig. 1A. To address these limitations, in this work, we develop an efficient and scalable post-hoc concept bottleneck to transform any pretrained generative model into an interpretable model. In contrast to CBGM [16], our approach works well by only training a few layers with minimal concept supervision.", + "bbox": [ + 88, + 207, + 485, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we propose a novel concept-bottleneck autoencoder (CB-AE) $f$ that can be inserted into the intermediate layers of a pretrained generative model $g = g_{2} \\circ g_{1}$ as shown in Fig. 1B (top). The CB-AE input and output are the generator latent $w$ and its reconstruction $w'$ respectively, and the overall generative model now becomes $g_{2} \\circ f \\circ g_{1}$ . The CB-AE latent space is the concept space $c$ which is used to reconstruct the generator latent space. In our framework, given a pretrained generative model $g = g_{2} \\circ g_{1}$ , only the encoder $E$ and decoder $D$ in the CB-AE $f = D \\circ E$ need to be trained, while $g_{1}, g_{2}$ are frozen pretrained weights, i.e. CB-AE uses post-hoc training. The benefit of the proposed CB-AE is that it allows us to debug the model easily with concept-level control by modifying the CB-AE concept latent during image generation (Fig. 1C). Further, our training requires only concept pseudo-labels achievable with minimal concept supervision (e.g. zero-shot CLIP classifier [35]).", + "bbox": [ + 91, + 375, + 483, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We also propose novel optimization-based concept interventions to achieve higher success rate and intervention quality. Based on the CB-AE, we propose an even more efficient post-hoc concept controller (CC) method (Fig. 1B, bottom) with simplified training, that can provide concept predictions and leverage optimization-based concept interventions. Note that while CB-AE is part of the new interpretable generative model, CC is a post-hoc control method that is not part of the generative model. Finally, we evaluate our CB-AE and CC on various generative models, including GANs and diffusion models, for standard datasets including CelebA, CelebA-HQ, and CUB. We show that CB-AE and CC significantly outperform prior work CBGM [16] w.r.t. steerability (or intervention success rate) on CelebA (average $+23\\%$ ) while also having $4 - 15 \\times$ faster training (Table 1).", + "bbox": [ + 89, + 635, + 485, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 89, + 864, + 416, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We are the first to propose a post-hoc concept bottleneck", + "bbox": [ + 89, + 882, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "autoencoder (CB-AE) for interpretable generative models. CB-AE can be trained efficiently with a frozen pretrained generative model, without real concept-labeled images.", + "bbox": [ + 524, + 207, + 906, + 252 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We also propose a novel and efficient optimization-based concept intervention method with improved steerability (avg. $+19\\%$ ) and higher image quality (avg. $32\\%$ better).", + "- We validate the effectiveness of our methods for GANs and diffusion models (avg. $+31\\%$ and $+28\\%$ steerability w.r.t. prior state-of-the-art) across varying image resolutions, while being $4 - 15 \\times$ faster to train on average." + ], + "bbox": [ + 511, + 253, + 906, + 358 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 383, + 653, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concept Bottleneck Models. Early work on CBMs [20, 25] relied on concept-labeled images to train a concept bottleneck layer with each neuron as a human-understandable concept, followed by a linear layer based on the concepts for the final classification. Post-hoc CBMs [45] extended this idea to convert a pretrained backbone into a CBM. More recent works like LF-CBM [32], LM4CV [43], LaBo [44], and VLG-CBM [39] use interpretability tools [31], vision-language models like CLIP [35], large language models, or open-set object detection [23] to eliminate the need for expensive concept-labeled data. Independently from our work, [21] proposed a concept-based intervention without a concept bottleneck, similar to our CC, but for classification. All these works are specifically designed for classification, while in this paper, our focus is on the image generation task.", + "bbox": [ + 509, + 411, + 908, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interpretability for Generative Models. A line of work [4, 7, 13] on learning disentangled concepts in variational autoencoders enables controllable generation, but they train from scratch and are not applicable to other generative models like GANs. Other works focus on identifying and manipulating structural rules or concepts in GANs [2, 3] and large language models [26, 28, 29, 40] by editing the model weights. In contrast, we focus on training inherently interpretable models, and the closest prior work is the recent CBGM [16], which also aims to build interpretable generative models, but requires expensive concept labels and training from scratch, limiting its scalability. In contrast, our proposed CB-AE can be trained efficiently with a frozen pretrained generative model with minimal concept supervision.", + "bbox": [ + 511, + 641, + 908, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Editing in Generative Models. Some works focus on conditional generation [37, 38] and image editing in generative models by modifying model weights [8, 33]. In", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8163", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e5b1e44a308162f613a686ee388e9e52e66bcaf2ba5a63c549e9c76ce41bc70a.jpg", + "image_caption": [ + "A. Original Black-Box Generative Model" + ], + "image_footnote": [], + "bbox": [ + 163, + 103, + 426, + 150 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/963c465ee94a7ef306745ec46b45e7927822556a896a8cde8a37d8a692fd73b7.jpg", + "image_caption": [ + "B. Our Post-hoc Interpretable Model" + ], + "image_footnote": [], + "bbox": [ + 459, + 102, + 766, + 155 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/19fde05c03a67d6228634530d9727b08fc39a05cf034a892d49f837375115055.jpg", + "image_caption": [ + "C. Training Objective 1 (reconstruction losses)" + ], + "image_footnote": [], + "bbox": [ + 173, + 176, + 429, + 277 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0cb9a9bc5780227cadab60a4bbd3dc68d2bc7632881533d4e3a90eb790106089.jpg", + "image_caption": [ + "D. Training Objective 2 (concept alignment loss)" + ], + "image_footnote": [], + "bbox": [ + 467, + 176, + 843, + 267 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c874f1f1b3809d302ea572fea3a9df915b618d66f5759ed071500a8dff990fe2.jpg", + "image_caption": [ + "E. Training Objective 3 (intervention losses)", + "Figure 2. Post-hoc CB-AE training for reconstruction, concept alignment, and intervention with a frozen pretrained generator $g_{2} \\circ g_{1}$ . Note that $\\mathcal{L}_c$ , $\\mathcal{L}_i$ indicate cross-entropy loss and $\\mathcal{L}_r$ indicates mean-squared-error loss. The darker lines indicate gradient flow during training." + ], + "image_footnote": [], + "bbox": [ + 173, + 289, + 794, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "contrast, our work focuses on developing inherently interpretable generative models by introducing additional concept bottleneck layers, where the capability of editing or intervention is naturally a by-product of the interpretability.", + "bbox": [ + 89, + 482, + 485, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Methods", + "text_level": 1, + "bbox": [ + 89, + 566, + 272, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a novel and low-cost concept bottleneck autoencoder (CB-AE) method in Sec. 3.1 to incorporate post-hoc interpretability in pretrained generative models. In Sec. 3.2, we present an optimization-based intervention method for concept-based steerability in the generative model. Finally, based on our insights from CB-AE and optimization-based interventions, we propose an even lower-cost post-hoc control method, concept controller (CC) in Sec. 3.3.", + "bbox": [ + 89, + 595, + 485, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminaries. Consider a generative model $g: \\mathcal{Z} \\to \\mathcal{X}$ that maps from random noise $z \\in \\mathcal{Z}$ to an image $x \\in \\mathcal{X}$ . To make the generative model $g$ become inherently interpretable, the goal of CBMs is to insert and train a concept bottleneck (say) $f$ at an intermediate location in $g$ . Let $g = g_2 \\circ g_1$ , i.e. $g$ can be divided into two parts, $g_1$ and $g_2$ (e.g. for a DCGAN [34], $g_1$ is the first 2 layers of $g$ and $g_2$ is the remaining layers). The input to the concept bottleneck $f$ will be the output of $g_1$ (Fig. 1). At the bottleneck of $f$ , we obtain the concept prediction $c \\in \\mathcal{C}$ where $\\mathcal{C}$ is the set of pre-defined concepts. In our setup, the concept vector $c$ has two logits for each binary concept $c_i$ (e.g. \"smiling\" would have two logits, for", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "smiling and not smiling) and $N$ logits for each categorical concept $c_{j}$ with $N$ classes (e.g. blonde/black/white/gray hair color would have four logits). For example, suppose we have one binary concept $c_{i}$ and one categorical concept $c_{j}$ , then we define $c = [c_{i}^{+}, c_{i}^{-}, c_{j}^{(1)}, c_{j}^{(2)}, \\ldots, c_{j}^{(N)}]^{\\top}$ where $c_{i} = [c_{i}^{+}, c_{i}^{-}]^{\\top} \\in \\mathbb{R}^{2}$ and $c_{j} = [c_{j}^{(1)}, c_{j}^{(2)}, \\ldots, c_{j}^{(N)}]^{\\top} \\in \\mathbb{R}^{N}$ .", + "bbox": [ + 511, + 482, + 906, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Post-hoc Concept Bottleneck Autoencoder", + "text_level": 1, + "bbox": [ + 511, + 601, + 872, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a concept bottleneck autoencoder (CB-AE), $f = D \\circ E$ (see Fig. 2B). The latent space of the CB-AE encoder $E$ is the concept prediction $c = E(g_1(z))$ . Apart from the predefined concepts, $c$ also contains an unsupervised concept embedding, learned using autoencoder reconstruction and intervention objectives, to encode other concepts absent from the predefined set (similar to CBGM [16]). The decoder $D$ reconstructs the features from $g_1$ based on the concept prediction $c$ , outputting $w' = D(c)$ . Considering the original output of $g_1$ to be $w = g_1(z)$ , we can generate the original image $x = g_2(w)$ as well as a reconstructed image $x' = g_2(w') = g_2(D \\circ E(w))$ generated using the CB-AE.", + "bbox": [ + 511, + 625, + 906, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training. There are 3 goals for the CB-AE. First, the generator's performance should be preserved even if the CB-AE output $w'$ is used instead of $w$ . Second, CB-AE should provide interpretability for the generated images $x'$ through the corresponding concepts $c$ . Lastly, CB-AE should allow accurate steering on image generation via concept interven-", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8164", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions. Based on these 3 goals, we formulate Objective 1-3 for CB-AE training below and illustrated in Fig. 2C-E.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Objective 1 (reconstruction losses $\\mathcal{L}_{r_1},\\mathcal{L}_{r_2}$ ). At each training iteration, we sample a latent $w$ by passing uniform noise $z$ to $g_{1}$ . Then, the latent $w$ is passed through $g_{2}$ to obtain a generated image $x = g_{2}(w)$ from the original model $g = g_{2} \\circ g_{1}$ , without using our CB-AE. We also reconstruct $w' = D \\circ E(w)$ using our CB-AE and the same latent $w$ , and obtain another generated image $x' = g_{2}(w')$ . Since our first goal is to preserve the generator's performance when using the CB-AE, we apply reconstruction losses (mean-squared error loss $\\mathcal{L}_r$ ) between the latents $w, w'$ and between the generated images $x, x'$ , as shown in Fig. 2C:", + "bbox": [ + 89, + 123, + 483, + 290 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {E, D} \\left[ \\mathcal {L} _ {r _ {1}} \\left(w, w ^ {\\prime}\\right) + \\mathcal {L} _ {r _ {2}} \\left(x, x ^ {\\prime}\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 300, + 483, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w' = D \\circ E(w)$ and $x' = g_2(w') = g_2 \\circ D \\circ E(w)$ and only the CB-AE parameters (i.e. $E, D$ ) are trainable.", + "bbox": [ + 89, + 335, + 483, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Objective 2 (concept alignment loss $\\mathcal{L}_c$ ). To ensure interpretability, we first obtain a concept pseudo-label $\\hat{y} = M(x)$ for a generated image $x$ from a pseudo-label source $M$ (Fig. 2D). The source $M$ can be either an off-the-shelf supervised model or a zero-shot prediction pipeline from (say) CLIP with concept names $\\mathcal{C}$ as the text inputs. With this approach, we avoid the requirement of any real images for training as well as the requirement of concept labels, unlike CBGM [16]. Since our second goal is to provide interpretability for the generated images through the concepts $c$ , we apply a cross-entropy loss $\\mathcal{L}_c$ between the CB-AE encoder output $c = E(w)$ and the concept pseudo-label $\\hat{y} = M(x)$ :", + "bbox": [ + 89, + 369, + 483, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {E} [ \\mathcal {L} _ {c} (\\hat {y}, c) ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 561, + 483, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The losses in Eq. (1), (2) are simultaneously optimized to learn only the CB-AE parameters $E, D$ .", + "bbox": [ + 89, + 593, + 482, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Objective 3 (intervention losses $\\mathcal{L}_{i_1},\\mathcal{L}_{i_2}$ ). Interventions are an important feature of concept-bottleneck models [20], allowing users to control the model output by modifying the concepts $c$ . However, for our CB-AE decoder $D$ , reconstruction and concept alignment losses in Objective 1 and 2 do not provide guidance on how the reconstructed latent $w^{\\prime}$ should change when concepts $c$ are manually modified. Hence, for steerability, we design Objective 3 (Fig. 2E) that encourages the CB-AE decoder $D$ to produce an appropriately changed and realistic latent $w^{\\prime}$ when the concepts in $c$ are modified. We first describe (a) how interventions are performed in our CB-AE, followed by designing (b) an intervened concept alignment loss $\\mathcal{L}_{i_1}$ and (c) a cyclic intervened concept loss $\\mathcal{L}_{i_2}$ to encourage steerability, i.e. intervention success.", + "bbox": [ + 89, + 626, + 483, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a) Intervening concepts. At each training iteration, we choose a random logit for a random concept to intervene, and modify only the chosen concept based on the chosen logit to get an intervened concept vector $c_{\\text{intervened}}$ (Fig. 2E). For", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "example, for a desired binary concept $i \\in \\mathcal{C}$ , the new concept vector $c_{\\text{intervened}}$ is computed by swapping the two logits, i.e. $c_{\\text{intervened}} = [\\ldots, c_i^-, c_i^+, \\ldots]$ from $c = [\\ldots, c_i^+, c_i^-, \\ldots]$ . The same can be done for a categorical concept $i \\in \\mathcal{C}$ as well by swapping the desired logits (say) $c_i^{(k)}$ with the highest logits $c_i^{(\\ell)}$ where $\\ell = \\arg \\max_j c_i^{(j)}$ . Concretely, $c = [\\ldots, c_i^{(1)}, \\ldots, c_i^{(\\ell)}, \\ldots, c_i^{(k)}, \\ldots, c_i^{(N)}, \\ldots]$ is modified to $c_{\\text{intervened}} = [\\ldots, c_i^{(1)}, \\ldots, c_i^{(k)}, \\ldots, c_i^{(\\ell)}, \\ldots, c_i^{(N)}, \\ldots]$ . We can also intervene on multiple concepts simultaneously.", + "bbox": [ + 511, + 90, + 906, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "b) Designing $\\mathcal{L}_{i_1}$ . Using the intervened concepts $c_{\\text{intervened}}$ we reconstruct an intervened latent $w_{\\text{intervened}} = D(c_{\\text{intervened}})$ and obtain an intervened generated image $x_{\\text{intervened}} = g_2(w_{\\text{intervened}})$ . Using the pseudo-label source, we obtain an intervened concept prediction $y_{\\text{intervened}} = M(x_{\\text{intervened}})$ . Since we already have the concept pseudo-label $\\hat{y} = M(x)$ for the original image $x$ , we modify $\\hat{y}$ to $\\hat{y}_{\\text{intervened}}$ by changing only the earlier chosen concept $i \\in \\mathcal{C}$ in $\\hat{y}$ to the chosen value (based on earlier chosen logit) as shown in Fig. 2E. In other words, $\\hat{y}_{\\text{intervened}}$ is the intervened concept pseudo-label. Now, to align the concepts in the intervened image $x_{\\text{intervened}}$ with the predicted concepts from $M$ , we use a cross-entropy loss $\\mathcal{L}_{i_1}$ between $\\hat{y}_{\\text{intervened}}$ and $y_{\\text{intervened}}$ :", + "bbox": [ + 511, + 239, + 908, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {E, D} {\\min } \\left[ \\mathcal {L} _ {i _ {1}} \\left(\\hat {y} _ {\\text {i n t e r v e n e d}}, y _ {\\text {i n t e r v e n e d}}\\right) \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 449, + 906, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "c) Designing $\\mathcal{L}_{i_2}$ . For cyclic consistency, we pass the intervented latent $w_{\\text{intervened}}$ through the CB-AE encoder $E$ to obtain a concept prediction $c_{\\text{intervened}}'$ (Fig. 2E) and apply a cross-entropy loss $\\mathcal{L}_{i_2}$ w.r.t. $\\hat{y}_{\\text{intervened}}$ to align the encoder's prediction with that of the pseudo-label source $M$ :", + "bbox": [ + 511, + 487, + 906, + 564 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {E, D} \\left[ \\mathcal {L} _ {i _ {2}} \\left(\\hat {y} _ {\\text {i n t e r v e n e d}}, c _ {\\text {i n t e r v e n e d}} ^ {\\prime}\\right) \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 575, + 906, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use $\\mathcal{L}_{i_1},\\mathcal{L}_{i_2}$ instead of $\\mathcal{L}_c$ for cross-entropy loss to differentiate the intervention losses from the concept loss in Objective 2, and only CB-AE parameters $E,D$ are trainable. Finally, recall that the concept vector $c$ contains an unsupervised concept embedding. Objective 3 implicitly encourages this embedding to not encode known concepts, since the unsupervised embedding is not modified during the intervention from $c$ to $c_{\\mathrm{intervened}}$ .", + "bbox": [ + 511, + 609, + 906, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Test-Time Intervention. Similar to training-time interventions, we can perform test-time interventions by modifying the value of any chosen concept (by swapping the logits) in the predicted concept vector $c$ to $c_{\\text{intervened}}$ . The intervened image $x_{\\text{intervened}} = g_2(w_{\\text{intervened}})$ can be obtained where $w_{\\text{intervened}} = D(c_{\\text{intervened}})$ . Note that swapping the logits ensures that the range of values in $c_{\\text{intervened}}$ are similar to $c$ , and also avoids the requirement of any estimation of how much to change the desired concept's logit. This makes it more accessible to users as they need not worry about the actual values being changed during an intervention.", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8165", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Optimization-based interventions", + "text_level": 1, + "bbox": [ + 89, + 90, + 383, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For an alternative intervention method, we draw inspiration from adversarial attacks [10] to perform test-time interventions using gradient-based optimization. Specifically, we use the iterative randomized fast gradient sign method (IRFGSM) [42] on the CB-AE encoder prediction.", + "bbox": [ + 89, + 114, + 485, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consider a generated image $x = g_{2}(w)$ with concept prediction $c = E(w)$ . To intervene in the generation process to obtain modified concepts $c^{*}$ , we solve the following objective using gradient ascent,", + "bbox": [ + 89, + 193, + 485, + 253 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw ^ {*} = w + \\underset {\\delta \\in \\Delta} {\\arg \\max } \\left[ - \\mathcal {L} _ {c} \\left(E (w + \\delta), c ^ {*}\\right) \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 263, + 483, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta = \\{\\delta : \\| \\delta \\|_{\\infty} \\leq \\epsilon\\}$ is the $\\ell_{\\infty}$ -norm bound on $\\delta$ with a hyperparameter $\\epsilon > 0$ . Intuitively, we optimize a small perturbation $\\delta$ such that $w^{*} = w + \\delta$ leads to the desired concepts $c^{*}$ . Then, the generated image $x^{*} = g_{2}(w^{*})$ is very similar to $x$ but contains concepts $c^{*}$ .", + "bbox": [ + 89, + 297, + 483, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Post-hoc Concept Controller (CC) for Steering", + "text_level": 1, + "bbox": [ + 89, + 385, + 483, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While optimization-based interventions work well empirically, it is interesting to note that the CB-AE decoder $D$ is not involved in the process. This leads us to question whether the CB-AE decoder can be removed if a user only plans to perform optimization-based interventions. While such a removal would not result in a CBM, it would be sufficient for the purpose of steering image generation. For this particular use case, it would be even more efficient than training the CB-AE since reconstruction and intervention losses are no longer required. Hence, we propose a post-hoc concept controller (CC), denoted as $\\Omega$ , that predicts the concepts, i.e. $c = \\Omega(g_1(z))$ to efficiently steer image generation (Fig. 1B).", + "bbox": [ + 89, + 410, + 485, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. For CC, the training is simply Objective 2 of the CB-AE training, i.e. the cross-entropy loss $\\mathcal{L}_c$ w.r.t. the concept pseudo-labels $\\hat{y} = M(x)$ from the pseudo-label source $M$ . Formally, the objective is $\\min_{\\Omega}[\\mathcal{L}_c(\\hat{y},c)]$ where $c = \\Omega (g_1(z))$ are the predicted concepts, and the loss encourages the concept controller $\\Omega$ to align with the pseudo-label source $M$ . Similar to the CB-AE training, we avoid the requirement of any real images for training as well as the requirement of concept labels.", + "bbox": [ + 89, + 594, + 485, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 746, + 223, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We detail the experimental setup and comprehensively evaluate our proposed methods with respect to the state-of-the-art prior work as well as other baselines.", + "bbox": [ + 89, + 773, + 483, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 830, + 277, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Base generative models and datasets. We evaluate our CB-AE and CC methods on diverse generative models including GAN [9], Progressive GAN [17], StyleGAN2 [18],", + "bbox": [ + 89, + 854, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and DDPM [15]. We use models pretrained on standard datasets of varying image resolution ( $64 \\times 64$ to $512 \\times 512$ ) like CelebA [24], CelebA-HQ [22], and CUB [41]. Following CBGM [16], we evaluate on the small balanced concept regime with the 8 most balanced concepts and the large unbalanced concept regime with all 40 concepts from the dataset. For CUB, we use 10 balanced concepts as per CBGM. Please refer to the Appendix for more details.", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CB-AE and CC. We use a 4-layer MLP or 4 convolution (and transposed convolutional) layers for CB-AE encoder or CC (and decoder) depending on the dimensions of latent $w$ . We use unsupervised concept embedding $\\in \\mathbb{R}^{40}$ in CB-AE (ensuring bottleneck is much smaller than the latent). As in CBGM [16], we train CB-AE/CC for 50 epochs with batch size 64. For optimization-based interventions, we use 50-step I-RFGSM [42] with $\\ell_{\\infty}$ -norm bound $\\epsilon = 0.1$ . Please refer to the Appendix for complete implementation details.", + "bbox": [ + 511, + 214, + 908, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pseudo-label source $M$ . We consider three variants for $M$ with varying levels of concept supervision. First, we use off-the-shelf supervised (ResNet18-based) concept classifiers. Second, with no concept supervision, we use CLIP zero-shot classifier [35] with only the concept names to obtain concept pseudo-labels. Third, as a compromise between the above two, we use TIP [48] which is a few-shot-labeled version of CLIP zero-shot classifier, utilizing 128 concept-labeled real images. Unless otherwise mentioned, our experiments use the supervised classifiers for a fair comparison with the prior work CBGM [16] that utilizes concept labels.", + "bbox": [ + 511, + 353, + 908, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Automated evaluation. Following CBGM [16], we train concept classifiers (ViT-L-16-based) on real images and concept labels with high accuracy on a held-out test set. Note that these classifiers are separate and have higher accuracy than those used for pseudo-labels. We evaluate our method using three automated metrics:", + "bbox": [ + 511, + 521, + 908, + 612 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Concept Accuracy is computed over 5k generated images as the average agreement between the supervised classifiers and our proposed CB-AE or CC.", + "- Steerability [16]: For each target concept, we find 5k latents that do not produce the target concept (i.e. probability of target concept $< 0.5$ from the supervised concept classifier). For these latents, we perform the concept intervention using either the baselines or our methods, and compute the steerability as the percentage of intervened images that are classified to have the target concept.", + "- Generation Quality is evaluated using the standard Fréchet Inception Distance (FID) [12]." + ], + "bbox": [ + 511, + 614, + 908, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, the concept accuracy, FID, and steerability metrics measure how well the concept, reconstruction, and intervention objectives, respectively, are satisfied.", + "bbox": [ + 511, + 799, + 906, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human evaluation. We conduct a large-scale user study on Amazon Mechanical Turk to validate the automated evaluation of concept accuracy and steerability. For both metrics,", + "bbox": [ + 511, + 847, + 908, + 892 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8166", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/abadf901e2ebe729eedfc5f973493353d24b3ca1215b2c2c6d81598c82b4ccbb.jpg", + "image_caption": [ + "A. CB-AE StyleGAN2 on CelebA-HQ" + ], + "image_footnote": [], + "bbox": [ + 133, + 101, + 251, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4deff7fb8ac8d7723a360e556f46462fdddc639081980a2abb4868745a1d48b1.jpg", + "image_caption": [ + "Figure 3. Samples generated using CB-AE with CelebA-HQ and CUB pretrained StyleGAN2 models along with concept probabilities." + ], + "image_footnote": [], + "bbox": [ + 133, + 191, + 250, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/210b801d1d3bcc046ca494614c2fa4ee06fb8584ca75704f2d3d8a6c71531a07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 101, + 370, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cece36462c3fa0931fc1791258b546203bfbb5a15bfa522089b28b036b089d23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 193, + 370, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/85fe15eec4bb1b9e6a7f7860fc91f267c5f8cd4a4e192fcac4fd2223775b0549.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 101, + 491, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5eb7ce5219511705cebb109fe07d5e405aaccf676d5d6537ceee96c0f8a6d5fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 193, + 490, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a645faf95ecd4839a1bf983f40429b48515c1c6ad94c364c5ffed5b4254a517e.jpg", + "image_caption": [ + "B. CB-AE StyleGAN2 on CUB" + ], + "image_footnote": [], + "bbox": [ + 503, + 101, + 620, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/178e2a9f475040ddc051204bd29cba62c2d98759ad5a0a3df4cb9c1893a0cc15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 193, + 617, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/51529635c305d9a15922a0354397272abf612dbadabea02eac069830b0258794.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 101, + 738, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9859344daf947a58d7b1d4c2eaf26744b86a0a84dc72d9975306d75b421ca50e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 193, + 736, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44f60efd3025e14550a4ff048b8160cfcb33795309b534467bc6be5b94e20cf5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 101, + 859, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2327f898dfce6ab323619de89ee5114b8cdc32ae4100d94d6afcaa0f26d6df1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 193, + 856, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5047cd29f3a41a5859ba60944a6831cb5cf5e91ebb51764ede675f1b819831bf.jpg", + "image_caption": [ + "Figure 4. Concept intervention examples for CB-AE with CelebA-HQ pretrained StyleGAN2." + ], + "image_footnote": [], + "bbox": [ + 93, + 270, + 901, + 412 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6461273f56004b1dd120515518f893a6bb18776ca53e9675d8ff70be10faac87.jpg", + "image_caption": [ + "Figure 5. Optimization-based concept intervention (opt-int) examples for CB-AE and CC with CelebA-HQ pretrained StyleGAN2." + ], + "image_footnote": [], + "bbox": [ + 94, + 446, + 905, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we display 10 images at a time and ask the user to click on images that match a displayed concept $c_{i}^{+}$ (see Appendix for more details). The images are collected as follows:", + "bbox": [ + 89, + 691, + 482, + 736 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Concept Accuracy: We collect and shuffle generated images and save their CB-AE and CC concept predictions $(c_{i}^{+}$ or $c_{i}^{-})$ . Based on user responses, we compute human agreement rate w.r.t. CB-AE/CC predictions.", + "- Steerability: We collect and shuffle generated images of concept $c_{i}^{-}$ with intervened images of concept $c_{i}^{+}$ . Based on user responses, we compute human agreement rate on whether intervened images actually contain concept $c_{i}^{+}$ . Shuffling the original and intervened images ensures that users are not biased towards clicking all images." + ], + "bbox": [ + 89, + 739, + 483, + 891 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate using approximately 100 images per concept and per method for two concepts from CelebA-HQ: “smiling” and “male/female”. Each set of 10 images is evaluated by 3 different users. Refer to the Appendix for more details.", + "bbox": [ + 511, + 691, + 906, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation", + "text_level": 1, + "bbox": [ + 511, + 758, + 633, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative evaluation. In Fig. 3, we present CB-AE-StyleGAN2 generated images for CelebA-HQ and CUB datasets with corresponding concept predictions. We also visualize CB-AE interventions in Fig. 4 and optimization-based interventions in Fig. 5. Interestingly, with the same latent $w$ , both CB-AE and CC lead to very similar optimization-based interventions, which is reasonable since the same pseudo-label source $M$ was used in both cases.", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8167", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d920ffc137578e9385b07864276fe297121338a29b4cf4467b5fc253edbe34f6.jpg", + "table_caption": [ + "Table 2. Concept accuracy evaluation on 5k samples with 8 concepts for CelebA, CelebA-HQ, and 10 concepts for CUB." + ], + "table_footnote": [], + "table_body": "
Conc. Acc. (%)CelebA (64×64)CelebA-HQ (256×256)CUB (64×64)CUB (256×256)
GANPGANDDPMDDPMStyGAN2PGANDDPMGANStyGAN2
Ours (CB-AE)86.5687.8784.9889.7986.0482.6872.3574.4181.33
Ours (CC)87.6590.0085.1389.8283.5783.9472.8775.6081.11
", + "bbox": [ + 93, + 103, + 903, + 178 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b086db45ec9e4ac6927430d49e890982be636438bfe4e31a487bf8da9ab1e212.jpg", + "table_caption": [ + "Table 3. Extended steerability evaluation on 5k samples with 8 concepts for CelebA, CelebA-HQ, and 10 concepts for CUB. ${}^{ \\dagger }$ CBGM numbers are from their paper (1k samples) since their results are not reproducible using their released code. For CBGM training time, we used the time taken by their code for GAN and base model training times for other models (whose code was not available) with 1 V100 GPU." + ], + "table_footnote": [], + "table_body": "
Steerability (%)CelebA (64×64)CelebA-HQ (256×256)CUB (64×64)CUB (256×256)
GANPGANDDPMDDPMStyGAN2PGANDDPMGANStyGAN2
CBGM [16]†25.60-13.80---14.8021.30-
Ours (CB-AE)47.3440.3123.7223.5040.2729.3125.6420.8910.52
Ours (CB-AE+opt-int)61.1441.7338.0950.4961.6632.1036.9446.0365.11
Ours (CC+opt-int)51.1458.9441.4556.7067.9547.2944.4748.9144.72
Train time reductionCB-AE3.7×5.4×4×8.1×3.5×2×3.7×3.3×3.1×
w.r.t. CBGMCC30.9×20.3×10×28.9×8.3×6.7×8.5×21×7.1×
", + "bbox": [ + 91, + 234, + 903, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, we find that the optimization-based interventions are relatively higher quality and more orthogonal (i.e. less changes to other concepts) than the CB-AE interventions.", + "bbox": [ + 88, + 378, + 480, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Concept Accuracy. In Table 2, we report the concept accuracies for our CB-AE and CC. We could not compare with CBGM since they do not evaluate concept accuracy, and since we could not reproduce their results. Overall, we find that both CB-AE and CC achieve good concept accuracies across various datasets, models and image resolutions.", + "bbox": [ + 88, + 425, + 482, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In most scenarios, CC outperforms CB-AE since concept alignment is the sole objective optimized in CC, while CB-AE has additional objectives. However, we observe that CB-AE outperforms CC only for StyleGAN2. This is because the multiple objectives in CB-AE training may be easier to balance when dealing with clean latents (GANs) than with noisy latents (DDPM), leading to CB-AE outperforming CC for StyleGAN2. This is supported by average loss for CB-AE being lower for StyleGAN2 (0.68) than DDPM (0.92).", + "bbox": [ + 88, + 518, + 482, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Steerability. We compare the concept steerability of our methods on 1k samples with GAN intervention methods like conditional GAN (CGAN) [27], auxiliary classifier GAN (ACGAN) [30], and CBGM (CB-GAN) [16], and with diffusion model intervention methods like classifier-guided (CG) [6] DDPM, classifier-free (CF) [14] DDPM, and concept bottleneck DDPM (CB-DDPM) [16] in Table 4. We observe significant gains over CBGM [16] for both GAN (average $+14.2\\%$ for CB-AE, $+31.1\\%$ for CB-AE with opt-int, and $+26.6\\%$ for CC) and for DDPM (average $+10.4\\%$ for CB-AE, $+23.3\\%$ for CB-AE with opt-int, and $+29.1\\%$ for CC).", + "bbox": [ + 88, + 657, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 presents extended steerability evaluation on $5\\mathrm{k}$ samples with other types of GANs like Progressive GAN [17] and StyleGAN2 [18] for higher resolution $(256\\times 256)$ datasets. We could not obtain CBGM's results for these set", + "bbox": [ + 88, + 825, + 483, + 885 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e4b647360b0f867b294244acff379f4d34291c34b0134cd1beba107ad68d4535.jpg", + "table_caption": [ + "Table 4. Steerability comparisons with CBGM [16] and other baseline intervention methods computed on 1k samples (each experiment repeated three times for mean and standard deviation)." + ], + "table_footnote": [], + "table_body": "
Concept RegimeSmall balanced conceptsLarge unbalanced concepts
DatasetCUB (10 conc.)CelebA (8 conc.)CelebA (40 conc.)
Baseline Intervention Methods
CGAN [27]5.4 ± 0.48.7 ± 1.32.9 ± 0.0
ACGAN [30]18.5 ± 0.49.2 ± 0.71.2 ± 0.1
CB-GAN [16]21.3 ± 0.325.6 ± 0.523.1 ± 0.2
Our Methods
CB-AE-GAN19.9 ± 0.947.1 ± 0.545.5 ± 0.4
CB-AE-GAN+opt-int45.1 ± 1.159.9 ± 1.858.3 ± 1.3
CC-GAN+opt-int49.3 ± 0.650.8 ± 1.249.7 ± 0.9
Baseline Intervention Methods
CF-DDPM [14]2.7 ± 1.97.2 ± 3.85.1 ± 2.4
CG-DDPM [6]2.1 ± 1.46.8 ± 1.15.4 ± 2.6
CB-DDPM [16]14.8 ± 6.213.8 ± 2.712.6 ± 1.7
Our Methods
CB-AE-DDPM25.8 ± 1.123.1 ± 0.923.6 ± 1.0
CB-AE-DDPM+opt-int37.3 ± 1.537.5 ± 1.336.2 ± 1.1
CC-DDPM+opt-int45.4 ± 2.241.8 ± 1.841.3 ± 1.5
", + "bbox": [ + 516, + 419, + 903, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tings using their code. Overall, we find optimization-based interventions (opt-int) outperform the CB-AE intervention method, and CC generally outperforms CB-AE. Intuitively, opt-int involves instance-specific and iterative optimization, while CB-AE is applied in the same way to all samples. Although CB-AE is trained with intervention losses, its steerability tends to be worse than CC since CB-AE is more challenging to train, with multiple objectives to satisfy.", + "bbox": [ + 509, + 671, + 906, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generation quality. We compare the generation quality of CBGM [16] with CB-AE and CC in Table 5. For CB-AE, we observe a relatively higher drop in image quality than CBGM, but our methods are trained $3.5 - 8 \\times$ faster, and do not require training from scratch. For a high-quality StyleGAN2 model, our CB-AE and CC with optimization-based interventions can produce almost the same quality", + "bbox": [ + 509, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8168", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8163897296c9cfc2dca0b38fd3b70116ed808922e7a70ba4b3ae3ffc41b8d768.jpg", + "table_caption": [ + "Table 5. Generation quality and training time comparisons for CelebA-HQ with StyleGAN2. ${}^{ \\dagger }$ CBGM results are from their paper. Training time is in V100 GPU-hours." + ], + "table_footnote": [], + "table_body": "
FID (↓)CBGM† [16]CB-AE (Ours)CC (Ours)
Base model9.07.667.66
CB model9.19.52-
CB interv.-9.65-
Opt-interv.-7.677.65
Train time (hrs)50146
", + "bbox": [ + 94, + 131, + 480, + 233 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/866c07e4286264e0a822ec3d78a7cfd08f184e2933ba31a8597281c3c60d2a62.jpg", + "table_caption": [ + "Table 6. Human evaluation results for CB-AE and CC trained with CelebA-HQ pretrained StyleGAN2." + ], + "table_footnote": [], + "table_body": "
Conc. Acc. (%)SmilingMale/Female
AutomatedHumanAutomatedHuman
CB-AE (Ours)92.3886.35100.094.06
CC (Ours)89.4780.3596.9696.30
Steerability (%)SmilingFemale
AutomatedHumanAutomatedHuman
CB-AE (Ours)65.9077.2717.0217.73
CB-AE w/ opt-int (Ours)76.5978.7242.8541.50
CC w/ opt-int (Ours)77.3677.3626.9219.87
", + "bbox": [ + 94, + 277, + 480, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of images as the base model while having high steerability (61.66% and 67.95% respectively, from Table 3).", + "bbox": [ + 89, + 462, + 482, + 493 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Intuitively, since CB-AE focuses on concept and intervention losses, we obtain better steerability, while CBGM has better FID since they include generative model losses. Hence, there is a tradeoff between image quality and interpretability which can be improved in future work.", + "bbox": [ + 89, + 494, + 483, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Human evaluation. In Table 6, we compare human agreement rate with automated evaluation of concept accuracy and steerability. We chose easily recognizable concepts, \"gender\" and \"smiling\" as representative of low and high steerability respectively. Overall, our automated evaluation is similar to human agreement (with room for improving the classifiers), validating the usefulness of automated evaluation.", + "bbox": [ + 89, + 574, + 483, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Analysis", + "text_level": 1, + "bbox": [ + 89, + 695, + 192, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study. We analyze the contribution of each CB-AE training loss. However, we do not ablate concept loss $\\mathcal{L}_c$ or latent reconstruction loss $\\mathcal{L}_{r_1}$ since our evaluation metrics would be meaningless if the CB-AE cannot predict the concepts or cannot reconstruct the generator latent $w'$ . Hence, Table 7 ablates image reconstruction loss $\\mathcal{L}_{r_2}$ from Eq. (1), intervened concept loss $\\mathcal{L}_{i_1}$ from Eq. (3), and cyclic intervened concept loss $\\mathcal{L}_{i_2}$ from Eq. (4). The ablations are for the most challenging pseudo-label setting of $M$ (CLIP-zero-shot), since it is most affected by loss ablations.", + "bbox": [ + 89, + 715, + 483, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 7, we observe that using the image reconstruction loss $\\mathcal{L}_{r_2}$ improves the generation quality FID (row #2", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0dc62dabdd10483b62be9668110ba1149a217e5dfcd72708be93da4f607af037.jpg", + "table_caption": [ + "Table 7. Ablation study on CB-AE training objectives for the most challenging CLIP-zero-shot pseudo-label setting for CelebAHQ pretrained StyleGAN2. $\\mathcal{L}_{r_2},\\mathcal{L}_{i_1},\\mathcal{L}_{i_2}$ indicate image reconstruction loss, intervened concept loss, and intervened cyclic loss respectively from Eq. (1), (3), (4)." + ], + "table_footnote": [], + "table_body": "
Row #\\( {\\mathcal{L}}_{{r}_{2}} \\)\\( {\\mathcal{L}}_{{i}_{1}} \\)\\( {\\mathcal{L}}_{{i}_{2}} \\)Trained with \\( M = \\) CLIP-zero-shot
Conc. Acc. (%)Steerability (%)FID (↓)
1✘✘✘58.876.0111.06
2✓✘✘59.326.519.88
3✓✓✘57.759.9812.42
4✓✘✓58.3313.0313.62
5✓✓✓67.2620.6112.82
", + "bbox": [ + 517, + 159, + 903, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/069b5323da4ebe18e2d670976c4b7b1d882cfc38167cca21b478c3846adc3df3.jpg", + "table_caption": [ + "Table 8. Sensitivity to pseudo-label source $M$ . For CB-AE trained with CelebA-HQ pretrained StyleGAN2, we compare concept accuracy and steerability with different $M$ : CLIP-zero-shot [35], TIP-few-shot [48], or supervised classifiers." + ], + "table_footnote": [], + "table_body": "
Pseudo-label source MConc. Acc. (%)Steerability (%)
CB-AECB-AE w/ opt-int
CLIP-zs67.2620.6129.23
TIP-fs-12876.0821.5138.73
Supervised-clsf86.0440.2761.66
", + "bbox": [ + 517, + 329, + 901, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vs. #1), which is intuitive since this loss directly encourages the images to be closer to the original ones. Next, using only the intervened concept loss or only the intervened cyclic loss improves the steerability while trading-off generation quality (row #3 vs. #2 or row #4 vs. #2). Finally, using both intervened losses significantly improves both concept accuracy and steerability (row #5 vs. #3 and #4), while generation quality remains similar. Overall, both intervention losses are crucial to ensure good concept accuracy and steerability, while image reconstruction loss improves generation quality. Sensitivity to pseudo-label source $M$ . Table 8 compares concept accuracy and steerability when varying pseudo-label source $M$ . Concept accuracy improves as pseudo-label quality improves from CLIP-zero-shot to supervised classifiers. While TIP few-shot does not improve CB-AE interventions, it significantly improves optimization-based interventions and highlights the usefulness of even limited labeled data. Further, our method can use newer CLIP models like SigLIP [47], OpenCLIP [5] to further improve performance.", + "bbox": [ + 511, + 433, + 908, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 732, + 633, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed two novel and low-cost methods, concept-bottleneck autoencoder (CB-AE) and concept controller (CC), to efficiently build interpretable generative models from pretrained models. Compared to the prior approach that struggles with efficiency and scalability, our methods achieve $4 - 15 \\times$ faster training, require minimal to no concept supervision, and generalize across modern generative model families including GANs and diffusion models with $25\\%$ improved steerability on average.", + "bbox": [ + 511, + 757, + 908, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8169", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 388, + 187, + 404 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dylan M Anstine and Alexandr Isayev. Generative models as an emerging paradigm in the chemical sciences. Journal of the American Chemical Society, 145(16):8736-8750, 2023. 1", + "[2] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B. Tenenbaum, William T. Freeman, and Antonio Torralba. GAN Dissection: Visualizing and understanding generative adversarial networks. In ICLR, 2019. 2", + "[3] David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, and Antonio Torralba. Rewriting a deep generative model. In ECCV, 2020. 2", + "[4] Ricky TQ Chen, Xuechen Li, Roger B Grosse, and David K Duvenaud. Isolating sources of disentanglement in variational autoencoders. In NeurIPS, 2018. 2", + "[5] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 8", + "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 7", + "[7] Zheng Ding, Yifan Xu, Weijian Xu, Gaurav Parmar, Yang Yang, Max Welling, and Zhuowen Tu. Guided variational autoencoder for disentanglement learning. In CVPR, 2020. 2", + "[8] Rohit Gandikota, Joanna Materzyńska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models. In ECCV, 2024. 2", + "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 5", + "[10] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 5", + "[11] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic" + ], + "bbox": [ + 93, + 414, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work is supported in part by National Science Foundation (NSF) awards CNS-1730158, ACI-1540112, ACI-1541349, OAC-1826967, OAC-2112167, CNS-2100237, CNS-2120019, the University of California Office of the President, and the University of California San Diego's California Institute for Telecommunications and Information Technology/Qualcomm Institute. This work used Delta CPU, GPU and Storage through allocation CIS230153 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support program, which is supported by National Science Foundation grants 2138259, 2138286, 2138307, 2137603, and 2138296. The authors are partially supported by National Science Foundation under Grant No. 2107189, 2313105, 2430539, Hellman Fellowship, and Intel Rising Star Faculty Award. The authors would also like to thank anonymous reviewers for valuable feedback to improve the manuscript.", + "data from generative models ready for image recognition? In ICLR, 2023. 1", + "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5", + "[13] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-VAE: Learning basic visual concepts with a constrained variational framework. In ICLR, 2017. 2", + "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 7", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 5", + "[16] Aya Abdelsalam Ismail, Julius Adebayo, Hector Corrada Bravo, Stephen Ra, and Kyunghyun Cho. Concept bottleneck generative models. In ICLR, 2024. 1, 2, 3, 4, 5, 7, 8", + "[17] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. 5, 7", + "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 5, 7", + "[19] Pushkal Katara, Zhou Xian, and Katerina Fragkiadaki. Gen2Sim: Scaling up robot learning in simulation with generative models. In ICRA, 2024. 1", + "[20] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In ICML, 2020. 1, 2, 4", + "[21] Sonia Laguna, Ricards Marcinkevics, Moritz Vandenhirtz, and Julia E. Vogt. Beyond concept bottleneck models: How to make black boxes intervenable? In NeurIPS, 2024. 2", + "[22] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 5", + "[23] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In ECCV, 2024. 2", + "[24] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 5", + "[25] Emanuele Marconato, Andrea Passerini, and Stefano Teso. GlanceNets: Interpretable, leak-proof concept-based models. In NeurIPS, 2022. 2", + "[26] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. In NeurIPS, 2022. 2", + "[27] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 7", + "[28] Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. Fast model editing at scale. In ICLR, 2022. 2", + "[29] Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. Memory-based model editing at scale. In ICML, 2022. 2" + ], + "bbox": [ + 93, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8170", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier GANs. In ICML, 2017. 7", + "[31] Tuomas Oikarinen and Tsui-Wei Weng. CLIP-Dissect: Automatic description of neuron representations in deep vision networks. In ICLR, 2023. 2", + "[32] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. In ICLR, 2023. 1, 2", + "[33] Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. PreciseControl: Enhancing text-to-image diffusion models with fine-grained attribute control. In ECCV, 2024. 2", + "[34] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks, 2016. 3", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 5, 8", + "[36] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2", + "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1, 2", + "[39] Divyansh Srivastava, Ge Yan, and Tsui-Wei Weng. VLGCBM: Training concept bottleneck models with vision-language guidance. In NeurIPS, 2024. 1, 2", + "[40] Chung-En Sun, Tuomas Oikarinen, Berk Ustun, and Tsui-Wei Weng. Concept bottleneck large language models. In ICLR, 2025. 2", + "[41] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech-UCSD Birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 5", + "[42] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In ICLR, 2020. 5", + "[43] An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Yang Wang, Jingbo Shang, and Julian McAuley. Learning concise and descriptive attributes for visual recognition. In ICCV, 2023. 1, 2", + "[44] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In CVPR, 2023. 2", + "[45] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. In ICLR, 2023. 1, 2", + "[46] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Melacci, Adrian Weller, Pietro Lio, and Mateja Jamnik. Concept embedding models: Beyond the accuracy-explainability trade-off. In NeurIPS, 2022. 1", + "[47] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 8", + "[48] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. In ECCV, 2022. 5, 8" + ], + "bbox": [ + 516, + 92, + 906, + 232 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8171", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/86d682d5-19b8-4583-bdec-4fcac9608ef2_model.json b/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/86d682d5-19b8-4583-bdec-4fcac9608ef2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2dcb879db0b113ff7b998c2c6393d25e95f2f440 --- /dev/null +++ b/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/86d682d5-19b8-4583-bdec-4fcac9608ef2_model.json @@ -0,0 +1,2431 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "Interpretable Generative Models through Post-hoc Concept Bottlenecks" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.181, + 0.825, + 0.218 + ], + "angle": 0, + "content": "Akshay Kulkarni, Ge Yan,\\* Chung-En Sun\\* Tuomas Oikarinen, and Tsui-Wei Weng \nUniversity of California San Diego" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.22, + 0.621, + 0.234 + ], + "angle": 0, + "content": "{a2kulkarni, lweng}@ucsd.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.486, + 0.635 + ], + "angle": 0, + "content": "Concept bottleneck models (CBM) aim to produce inherently interpretable models that rely on human-understandable concepts for their predictions. However, existing approaches to design interpretable generative models based on CBMs are not yet efficient and scalable, as they require expensive generative model training from scratch as well as real images with labor-intensive concept supervision. To address these challenges, we present two novel and low-cost methods to build interpretable generative models through post-hoc techniques and we name our approaches: concept-bottleneck autoencoder (CB-AE) and concept controller (CC). Our proposed approaches enable efficient and scalable training without the need of real data and require only minimal to no concept supervision. Additionally, our methods generalize across modern generative model families including generative adversarial networks and diffusion models. We demonstrate the superior interpretability and steerability of our methods on numerous standard datasets like CelebA, CelebA-HQ, and CUB with large improvements (average \\(\\sim 25\\%\\) ) over the prior work, while being \\(4 - 15\\times\\) faster to train. Finally, a large-scale user study is performed to validate the interpretability and steerability of our methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.222, + 0.669 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.797 + ], + "angle": 0, + "content": "Deep generative models [36-38] have become increasingly powerful and widely used in many high-stakes domains and applications, including realistic data generation [11], simulating hypothetical scenarios or environments [19], and scientific discovery [1]. It is therefore important to ensure that the generation process is interpretable, which will allow us to understand and audit the generation, and further mitigate potential biases and harms (e.g. content moderation)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.799, + 0.485, + 0.86 + ], + "angle": 0, + "content": "Unfortunately, most of the advances in deep learning utilize complex, black-box neural network architectures that are difficult to interpret and understand. This leads to user mistrust of model predictions due to the absence of explanations." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.271, + 0.874, + 0.284 + ], + "angle": 0, + "content": "A. Prior Work (Concept Bottleneck Generative Models)" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.286, + 0.893, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.353, + 0.865, + 0.366 + ], + "angle": 0, + "content": "Expensive generative model training from scratch!" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.368, + 0.852, + 0.381 + ], + "angle": 0, + "content": "B. Ours (Post-Hoc Interpretable Generative Models)" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.382, + 0.893, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.512, + 0.901, + 0.525 + ], + "angle": 0, + "content": "Efficiently train only CB-AE/CC with frozen generative model" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.527, + 0.664, + 0.539 + ], + "angle": 0, + "content": "C. Qualitative Results" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.542, + 0.905, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.628, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Figure 1. A. Prior work on interpretable generative models requires expensive generative model training from scratch. B. Our CB-AE and CC can be trained efficiently for post-hoc interpretability in a pretrained, frozen generative model \\( g_{2} \\circ g_{1} \\). C. Example concept intervention with CB-AE and corresponding concept vectors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.907, + 0.868 + ], + "angle": 0, + "content": "To address this, there has been work on developing inherently interpretable deep vision models using concept bottlenecks [20, 32, 39, 43-45]. These approaches train a concept bottleneck layer after the feature extractor (backbone) to embed a set of human-understandable concepts, followed by an interpretable sparse linear layer for the final classification based on the concept prediction. However, current development of CBMs is primarily focused on classification tasks, and only one prior work, CBGM [16], has extended it to image generation, indicating this area is under-explored." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The key idea of the seminal work CBGM [16] is to represent concepts as learnable embeddings [46] at an inter" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.875, + 0.217, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.458, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Code: github.com/Trustworthy-ML-Lab/posthoc-generative-cbm" + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.875, + 0.458, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8162" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison of our CB-AE and CC with prior work CBGM [16] (green indicates desirable properties). CC trades-off inherent interpretability for better steerability, image quality, and faster training. We could not reproduce CBGM [16] to evaluate concept accuracy." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.12, + 0.905, + 0.193 + ], + "angle": 0, + "content": "
MethodPost-hoc trainingTraining without concept-labeled real imagesInherently interpretable modelConcept Acc. (%)Steerability (%) (↑)FID (↓)Train time (V100-hrs) (↓)
StyleGAN2DDPM
CBGM [16]XX✓-25.609.1050240
CB-AE (Ours)✓✓✓86.5647.34 (+21.74)9.5214 (3.5× faster)29.5 (8.1× faster)
CC (Ours)✓✓X87.6551.14 (+25.54)7.656 (8.3× faster)8.3 (28.9× faster)
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.208, + 0.486, + 0.375 + ], + "angle": 0, + "content": "mediate location in the generative model, and combine the embeddings to compute the generative model latent. However, the CBGM-based generative model has to be trained from scratch using concept-labeled real images, which could be difficult to scale and computationally intensive (e.g. 240 V100-hours for DDPM- \\(256 \\times 256\\) [15]) as shown in Fig. 1A. To address these limitations, in this work, we develop an efficient and scalable post-hoc concept bottleneck to transform any pretrained generative model into an interpretable model. In contrast to CBGM [16], our approach works well by only training a few layers with minimal concept supervision." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.376, + 0.485, + 0.633 + ], + "angle": 0, + "content": "Specifically, we propose a novel concept-bottleneck autoencoder (CB-AE) \\( f \\) that can be inserted into the intermediate layers of a pretrained generative model \\( g = g_{2} \\circ g_{1} \\) as shown in Fig. 1B (top). The CB-AE input and output are the generator latent \\( w \\) and its reconstruction \\( w' \\) respectively, and the overall generative model now becomes \\( g_{2} \\circ f \\circ g_{1} \\). The CB-AE latent space is the concept space \\( c \\) which is used to reconstruct the generator latent space. In our framework, given a pretrained generative model \\( g = g_{2} \\circ g_{1} \\), only the encoder \\( E \\) and decoder \\( D \\) in the CB-AE \\( f = D \\circ E \\) need to be trained, while \\( g_{1}, g_{2} \\) are frozen pretrained weights, i.e. CB-AE uses post-hoc training. The benefit of the proposed CB-AE is that it allows us to debug the model easily with concept-level control by modifying the CB-AE concept latent during image generation (Fig. 1C). Further, our training requires only concept pseudo-labels achievable with minimal concept supervision (e.g. zero-shot CLIP classifier [35])." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.486, + 0.863 + ], + "angle": 0, + "content": "We also propose novel optimization-based concept interventions to achieve higher success rate and intervention quality. Based on the CB-AE, we propose an even more efficient post-hoc concept controller (CC) method (Fig. 1B, bottom) with simplified training, that can provide concept predictions and leverage optimization-based concept interventions. Note that while CB-AE is part of the new interpretable generative model, CC is a post-hoc control method that is not part of the generative model. Finally, we evaluate our CB-AE and CC on various generative models, including GANs and diffusion models, for standard datasets including CelebA, CelebA-HQ, and CUB. We show that CB-AE and CC significantly outperform prior work CBGM [16] w.r.t. steerability (or intervention success rate) on CelebA (average \\(+23\\%\\)) while also having \\(4 - 15 \\times\\) faster training (Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.865, + 0.418, + 0.88 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.883, + 0.483, + 0.899 + ], + "angle": 0, + "content": "- We are the first to propose a post-hoc concept bottleneck" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.208, + 0.907, + 0.253 + ], + "angle": 0, + "content": "autoencoder (CB-AE) for interpretable generative models. CB-AE can be trained efficiently with a frozen pretrained generative model, without real concept-labeled images." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.254, + 0.905, + 0.299 + ], + "angle": 0, + "content": "- We also propose a novel and efficient optimization-based concept intervention method with improved steerability (avg. \\(+19\\%\\)) and higher image quality (avg. \\(32\\%\\) better)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.299, + 0.907, + 0.359 + ], + "angle": 0, + "content": "- We validate the effectiveness of our methods for GANs and diffusion models (avg. \\(+31\\%\\) and \\(+28\\%\\) steerability w.r.t. prior state-of-the-art) across varying image resolutions, while being \\(4 - 15 \\times\\) faster to train on average." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.254, + 0.907, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.384, + 0.655, + 0.4 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.412, + 0.909, + 0.639 + ], + "angle": 0, + "content": "Concept Bottleneck Models. Early work on CBMs [20, 25] relied on concept-labeled images to train a concept bottleneck layer with each neuron as a human-understandable concept, followed by a linear layer based on the concepts for the final classification. Post-hoc CBMs [45] extended this idea to convert a pretrained backbone into a CBM. More recent works like LF-CBM [32], LM4CV [43], LaBo [44], and VLG-CBM [39] use interpretability tools [31], vision-language models like CLIP [35], large language models, or open-set object detection [23] to eliminate the need for expensive concept-labeled data. Independently from our work, [21] proposed a concept-based intervention without a concept bottleneck, similar to our CC, but for classification. All these works are specifically designed for classification, while in this paper, our focus is on the image generation task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.853 + ], + "angle": 0, + "content": "Interpretability for Generative Models. A line of work [4, 7, 13] on learning disentangled concepts in variational autoencoders enables controllable generation, but they train from scratch and are not applicable to other generative models like GANs. Other works focus on identifying and manipulating structural rules or concepts in GANs [2, 3] and large language models [26, 28, 29, 40] by editing the model weights. In contrast, we focus on training inherently interpretable models, and the closest prior work is the recent CBGM [16], which also aims to build interpretable generative models, but requires expensive concept labels and training from scratch, limiting its scalability. In contrast, our proposed CB-AE can be trained efficiently with a frozen pretrained generative model with minimal concept supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Image Editing in Generative Models. Some works focus on conditional generation [37, 38] and image editing in generative models by modifying model weights [8, 33]. In" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8163" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.164, + 0.089, + 0.43, + 0.103 + ], + "angle": 0, + "content": "A. Original Black-Box Generative Model" + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.104, + 0.428, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.089, + 0.721, + 0.102 + ], + "angle": 0, + "content": "B. Our Post-hoc Interpretable Model" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.103, + 0.767, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.164, + 0.162, + 0.466, + 0.177 + ], + "angle": 0, + "content": "C. Training Objective 1 (reconstruction losses)" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.178, + 0.43, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.163, + 0.797, + 0.177 + ], + "angle": 0, + "content": "D. Training Objective 2 (concept alignment loss)" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.178, + 0.844, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.164, + 0.283, + 0.452, + 0.297 + ], + "angle": 0, + "content": "E. Training Objective 3 (intervention losses)" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.29, + 0.795, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.428, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Figure 2. Post-hoc CB-AE training for reconstruction, concept alignment, and intervention with a frozen pretrained generator \\( g_{2} \\circ g_{1} \\). Note that \\( \\mathcal{L}_c \\), \\( \\mathcal{L}_i \\) indicate cross-entropy loss and \\( \\mathcal{L}_r \\) indicates mean-squared-error loss. The darker lines indicate gradient flow during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.483, + 0.486, + 0.545 + ], + "angle": 0, + "content": "contrast, our work focuses on developing inherently interpretable generative models by introducing additional concept bottleneck layers, where the capability of editing or intervention is naturally a by-product of the interpretability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.568, + 0.273, + 0.586 + ], + "angle": 0, + "content": "3. Proposed Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.486, + 0.717 + ], + "angle": 0, + "content": "We propose a novel and low-cost concept bottleneck autoencoder (CB-AE) method in Sec. 3.1 to incorporate post-hoc interpretability in pretrained generative models. In Sec. 3.2, we present an optimization-based intervention method for concept-based steerability in the generative model. Finally, based on our insights from CB-AE and optimization-based interventions, we propose an even lower-cost post-hoc control method, concept controller (CC) in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Preliminaries. Consider a generative model \\( g: \\mathcal{Z} \\to \\mathcal{X} \\) that maps from random noise \\( z \\in \\mathcal{Z} \\) to an image \\( x \\in \\mathcal{X} \\). To make the generative model \\( g \\) become inherently interpretable, the goal of CBMs is to insert and train a concept bottleneck (say) \\( f \\) at an intermediate location in \\( g \\). Let \\( g = g_2 \\circ g_1 \\), i.e. \\( g \\) can be divided into two parts, \\( g_1 \\) and \\( g_2 \\) (e.g. for a DCGAN [34], \\( g_1 \\) is the first 2 layers of \\( g \\) and \\( g_2 \\) is the remaining layers). The input to the concept bottleneck \\( f \\) will be the output of \\( g_1 \\) (Fig. 1). At the bottleneck of \\( f \\), we obtain the concept prediction \\( c \\in \\mathcal{C} \\) where \\( \\mathcal{C} \\) is the set of pre-defined concepts. In our setup, the concept vector \\( c \\) has two logits for each binary concept \\( c_i \\) (e.g. \"smiling\" would have two logits, for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.907, + 0.585 + ], + "angle": 0, + "content": "smiling and not smiling) and \\(N\\) logits for each categorical concept \\(c_{j}\\) with \\(N\\) classes (e.g. blonde/black/white/gray hair color would have four logits). For example, suppose we have one binary concept \\(c_{i}\\) and one categorical concept \\(c_{j}\\), then we define \\(c = [c_{i}^{+}, c_{i}^{-}, c_{j}^{(1)}, c_{j}^{(2)}, \\ldots, c_{j}^{(N)}]^{\\top}\\) where \\(c_{i} = [c_{i}^{+}, c_{i}^{-}]^{\\top} \\in \\mathbb{R}^{2}\\) and \\(c_{j} = [c_{j}^{(1)}, c_{j}^{(2)}, \\ldots, c_{j}^{(N)}]^{\\top} \\in \\mathbb{R}^{N}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.874, + 0.618 + ], + "angle": 0, + "content": "3.1. Post-hoc Concept Bottleneck Autoencoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.907, + 0.809 + ], + "angle": 0, + "content": "We propose a concept bottleneck autoencoder (CB-AE), \\( f = D \\circ E \\) (see Fig. 2B). The latent space of the CB-AE encoder \\( E \\) is the concept prediction \\( c = E(g_1(z)) \\). Apart from the predefined concepts, \\( c \\) also contains an unsupervised concept embedding, learned using autoencoder reconstruction and intervention objectives, to encode other concepts absent from the predefined set (similar to CBGM [16]). The decoder \\( D \\) reconstructs the features from \\( g_1 \\) based on the concept prediction \\( c \\), outputting \\( w' = D(c) \\). Considering the original output of \\( g_1 \\) to be \\( w = g_1(z) \\), we can generate the original image \\( x = g_2(w) \\) as well as a reconstructed image \\( x' = g_2(w') = g_2(D \\circ E(w)) \\) generated using the CB-AE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Training. There are 3 goals for the CB-AE. First, the generator's performance should be preserved even if the CB-AE output \\( w' \\) is used instead of \\( w \\). Second, CB-AE should provide interpretability for the generated images \\( x' \\) through the corresponding concepts \\( c \\). Lastly, CB-AE should allow accurate steering on image generation via concept interven-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8164" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "tions. Based on these 3 goals, we formulate Objective 1-3 for CB-AE training below and illustrated in Fig. 2C-E." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.124, + 0.484, + 0.291 + ], + "angle": 0, + "content": "Objective 1 (reconstruction losses \\(\\mathcal{L}_{r_1},\\mathcal{L}_{r_2}\\)). At each training iteration, we sample a latent \\(w\\) by passing uniform noise \\(z\\) to \\(g_{1}\\). Then, the latent \\(w\\) is passed through \\(g_{2}\\) to obtain a generated image \\(x = g_{2}(w)\\) from the original model \\(g = g_{2} \\circ g_{1}\\), without using our CB-AE. We also reconstruct \\(w' = D \\circ E(w)\\) using our CB-AE and the same latent \\(w\\), and obtain another generated image \\(x' = g_{2}(w')\\). Since our first goal is to preserve the generator's performance when using the CB-AE, we apply reconstruction losses (mean-squared error loss \\(\\mathcal{L}_r\\)) between the latents \\(w, w'\\) and between the generated images \\(x, x'\\), as shown in Fig. 2C:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.301, + 0.484, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\min _ {E, D} \\left[ \\mathcal {L} _ {r _ {1}} \\left(w, w ^ {\\prime}\\right) + \\mathcal {L} _ {r _ {2}} \\left(x, x ^ {\\prime}\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.484, + 0.367 + ], + "angle": 0, + "content": "where \\(w' = D \\circ E(w)\\) and \\(x' = g_2(w') = g_2 \\circ D \\circ E(w)\\) and only the CB-AE parameters (i.e. \\(E, D\\)) are trainable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.484, + 0.552 + ], + "angle": 0, + "content": "Objective 2 (concept alignment loss \\(\\mathcal{L}_c\\)). To ensure interpretability, we first obtain a concept pseudo-label \\(\\hat{y} = M(x)\\) for a generated image \\(x\\) from a pseudo-label source \\(M\\) (Fig. 2D). The source \\(M\\) can be either an off-the-shelf supervised model or a zero-shot prediction pipeline from (say) CLIP with concept names \\(\\mathcal{C}\\) as the text inputs. With this approach, we avoid the requirement of any real images for training as well as the requirement of concept labels, unlike CBGM [16]. Since our second goal is to provide interpretability for the generated images through the concepts \\(c\\), we apply a cross-entropy loss \\(\\mathcal{L}_c\\) between the CB-AE encoder output \\(c = E(w)\\) and the concept pseudo-label \\(\\hat{y} = M(x)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.562, + 0.484, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\min _ {E} [ \\mathcal {L} _ {c} (\\hat {y}, c) ]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.623 + ], + "angle": 0, + "content": "The losses in Eq. (1), (2) are simultaneously optimized to learn only the CB-AE parameters \\( E, D \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.838 + ], + "angle": 0, + "content": "Objective 3 (intervention losses \\(\\mathcal{L}_{i_1},\\mathcal{L}_{i_2}\\)). Interventions are an important feature of concept-bottleneck models [20], allowing users to control the model output by modifying the concepts \\(c\\). However, for our CB-AE decoder \\(D\\), reconstruction and concept alignment losses in Objective 1 and 2 do not provide guidance on how the reconstructed latent \\(w^{\\prime}\\) should change when concepts \\(c\\) are manually modified. Hence, for steerability, we design Objective 3 (Fig. 2E) that encourages the CB-AE decoder \\(D\\) to produce an appropriately changed and realistic latent \\(w^{\\prime}\\) when the concepts in \\(c\\) are modified. We first describe (a) how interventions are performed in our CB-AE, followed by designing (b) an intervened concept alignment loss \\(\\mathcal{L}_{i_1}\\) and (c) a cyclic intervened concept loss \\(\\mathcal{L}_{i_2}\\) to encourage steerability, i.e. intervention success." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "a) Intervening concepts. At each training iteration, we choose a random logit for a random concept to intervene, and modify only the chosen concept based on the chosen logit to get an intervened concept vector \\( c_{\\text{intervened}} \\) (Fig. 2E). For" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.239 + ], + "angle": 0, + "content": "example, for a desired binary concept \\(i \\in \\mathcal{C}\\), the new concept vector \\(c_{\\text{intervened}}\\) is computed by swapping the two logits, i.e. \\(c_{\\text{intervened}} = [\\ldots, c_i^-, c_i^+, \\ldots]\\) from \\(c = [\\ldots, c_i^+, c_i^-, \\ldots]\\). The same can be done for a categorical concept \\(i \\in \\mathcal{C}\\) as well by swapping the desired logits (say) \\(c_i^{(k)}\\) with the highest logits \\(c_i^{(\\ell)}\\) where \\(\\ell = \\arg \\max_j c_i^{(j)}\\). Concretely, \\(c = [\\ldots, c_i^{(1)}, \\ldots, c_i^{(\\ell)}, \\ldots, c_i^{(k)}, \\ldots, c_i^{(N)}, \\ldots]\\) is modified to \\(c_{\\text{intervened}} = [\\ldots, c_i^{(1)}, \\ldots, c_i^{(k)}, \\ldots, c_i^{(\\ell)}, \\ldots, c_i^{(N)}, \\ldots]\\). We can also intervene on multiple concepts simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.241, + 0.909, + 0.438 + ], + "angle": 0, + "content": "b) Designing \\(\\mathcal{L}_{i_1}\\). Using the intervened concepts \\(c_{\\text{intervened}}\\) we reconstruct an intervened latent \\(w_{\\text{intervened}} = D(c_{\\text{intervened}})\\) and obtain an intervened generated image \\(x_{\\text{intervened}} = g_2(w_{\\text{intervened}})\\). Using the pseudo-label source, we obtain an intervened concept prediction \\(y_{\\text{intervened}} = M(x_{\\text{intervened}})\\). Since we already have the concept pseudo-label \\(\\hat{y} = M(x)\\) for the original image \\(x\\), we modify \\(\\hat{y}\\) to \\(\\hat{y}_{\\text{intervened}}\\) by changing only the earlier chosen concept \\(i \\in \\mathcal{C}\\) in \\(\\hat{y}\\) to the chosen value (based on earlier chosen logit) as shown in Fig. 2E. In other words, \\(\\hat{y}_{\\text{intervened}}\\) is the intervened concept pseudo-label. Now, to align the concepts in the intervened image \\(x_{\\text{intervened}}\\) with the predicted concepts from \\(M\\), we use a cross-entropy loss \\(\\mathcal{L}_{i_1}\\) between \\(\\hat{y}_{\\text{intervened}}\\) and \\(y_{\\text{intervened}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.45, + 0.907, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\underset {E, D} {\\min } \\left[ \\mathcal {L} _ {i _ {1}} \\left(\\hat {y} _ {\\text {i n t e r v e n e d}}, y _ {\\text {i n t e r v e n e d}}\\right) \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.565 + ], + "angle": 0, + "content": "c) Designing \\(\\mathcal{L}_{i_2}\\). For cyclic consistency, we pass the intervented latent \\(w_{\\text{intervened}}\\) through the CB-AE encoder \\(E\\) to obtain a concept prediction \\(c_{\\text{intervened}}'\\) (Fig. 2E) and apply a cross-entropy loss \\(\\mathcal{L}_{i_2}\\) w.r.t. \\(\\hat{y}_{\\text{intervened}}\\) to align the encoder's prediction with that of the pseudo-label source \\(M\\):" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.576, + 0.907, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\min _ {E, D} \\left[ \\mathcal {L} _ {i _ {2}} \\left(\\hat {y} _ {\\text {i n t e r v e n e d}}, c _ {\\text {i n t e r v e n e d}} ^ {\\prime}\\right) \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.734 + ], + "angle": 0, + "content": "We use \\(\\mathcal{L}_{i_1},\\mathcal{L}_{i_2}\\) instead of \\(\\mathcal{L}_c\\) for cross-entropy loss to differentiate the intervention losses from the concept loss in Objective 2, and only CB-AE parameters \\(E,D\\) are trainable. Finally, recall that the concept vector \\(c\\) contains an unsupervised concept embedding. Objective 3 implicitly encourages this embedding to not encode known concepts, since the unsupervised embedding is not modified during the intervention from \\(c\\) to \\(c_{\\mathrm{intervened}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Test-Time Intervention. Similar to training-time interventions, we can perform test-time interventions by modifying the value of any chosen concept (by swapping the logits) in the predicted concept vector \\( c \\) to \\( c_{\\text{intervened}} \\). The intervened image \\( x_{\\text{intervened}} = g_2(w_{\\text{intervened}}) \\) can be obtained where \\( w_{\\text{intervened}} = D(c_{\\text{intervened}}) \\). Note that swapping the logits ensures that the range of values in \\( c_{\\text{intervened}} \\) are similar to \\( c \\), and also avoids the requirement of any estimation of how much to change the desired concept's logit. This makes it more accessible to users as they need not worry about the actual values being changed during an intervention." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8165" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.385, + 0.108 + ], + "angle": 0, + "content": "3.2. Optimization-based interventions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.192 + ], + "angle": 0, + "content": "For an alternative intervention method, we draw inspiration from adversarial attacks [10] to perform test-time interventions using gradient-based optimization. Specifically, we use the iterative randomized fast gradient sign method (IRFGSM) [42] on the CB-AE encoder prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.194, + 0.486, + 0.255 + ], + "angle": 0, + "content": "Consider a generated image \\( x = g_{2}(w) \\) with concept prediction \\( c = E(w) \\). To intervene in the generation process to obtain modified concepts \\( c^{*} \\), we solve the following objective using gradient ascent," + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.264, + 0.484, + 0.29 + ], + "angle": 0, + "content": "\\[\nw ^ {*} = w + \\underset {\\delta \\in \\Delta} {\\arg \\max } \\left[ - \\mathcal {L} _ {c} \\left(E (w + \\delta), c ^ {*}\\right) \\right] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.299, + 0.485, + 0.377 + ], + "angle": 0, + "content": "where \\(\\Delta = \\{\\delta : \\| \\delta \\|_{\\infty} \\leq \\epsilon\\}\\) is the \\(\\ell_{\\infty}\\)-norm bound on \\(\\delta\\) with a hyperparameter \\(\\epsilon > 0\\). Intuitively, we optimize a small perturbation \\(\\delta\\) such that \\(w^{*} = w + \\delta\\) leads to the desired concepts \\(c^{*}\\). Then, the generated image \\(x^{*} = g_{2}(w^{*})\\) is very similar to \\(x\\) but contains concepts \\(c^{*}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.386, + 0.484, + 0.403 + ], + "angle": 0, + "content": "3.3. Post-hoc Concept Controller (CC) for Steering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.486, + 0.594 + ], + "angle": 0, + "content": "While optimization-based interventions work well empirically, it is interesting to note that the CB-AE decoder \\( D \\) is not involved in the process. This leads us to question whether the CB-AE decoder can be removed if a user only plans to perform optimization-based interventions. While such a removal would not result in a CBM, it would be sufficient for the purpose of steering image generation. For this particular use case, it would be even more efficient than training the CB-AE since reconstruction and intervention losses are no longer required. Hence, we propose a post-hoc concept controller (CC), denoted as \\( \\Omega \\), that predicts the concepts, i.e. \\( c = \\Omega(g_1(z)) \\) to efficiently steer image generation (Fig. 1B)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.486, + 0.732 + ], + "angle": 0, + "content": "Training. For CC, the training is simply Objective 2 of the CB-AE training, i.e. the cross-entropy loss \\(\\mathcal{L}_c\\) w.r.t. the concept pseudo-labels \\(\\hat{y} = M(x)\\) from the pseudo-label source \\(M\\). Formally, the objective is \\(\\min_{\\Omega}[\\mathcal{L}_c(\\hat{y},c)]\\) where \\(c = \\Omega (g_1(z))\\) are the predicted concepts, and the loss encourages the concept controller \\(\\Omega\\) to align with the pseudo-label source \\(M\\). Similar to the CB-AE training, we avoid the requirement of any real images for training as well as the requirement of concept labels." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.224, + 0.763 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.485, + 0.82 + ], + "angle": 0, + "content": "We detail the experimental setup and comprehensively evaluate our proposed methods with respect to the state-of-the-art prior work as well as other baselines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.279, + 0.848 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Base generative models and datasets. We evaluate our CB-AE and CC methods on diverse generative models including GAN [9], Progressive GAN [17], StyleGAN2 [18]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.213 + ], + "angle": 0, + "content": "and DDPM [15]. We use models pretrained on standard datasets of varying image resolution (\\(64 \\times 64\\) to \\(512 \\times 512\\)) like CelebA [24], CelebA-HQ [22], and CUB [41]. Following CBGM [16], we evaluate on the small balanced concept regime with the 8 most balanced concepts and the large unbalanced concept regime with all 40 concepts from the dataset. For CUB, we use 10 balanced concepts as per CBGM. Please refer to the Appendix for more details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.909, + 0.351 + ], + "angle": 0, + "content": "CB-AE and CC. We use a 4-layer MLP or 4 convolution (and transposed convolutional) layers for CB-AE encoder or CC (and decoder) depending on the dimensions of latent \\( w \\). We use unsupervised concept embedding \\( \\in \\mathbb{R}^{40} \\) in CB-AE (ensuring bottleneck is much smaller than the latent). As in CBGM [16], we train CB-AE/CC for 50 epochs with batch size 64. For optimization-based interventions, we use 50-step I-RFGSM [42] with \\( \\ell_{\\infty} \\)-norm bound \\( \\epsilon = 0.1 \\). Please refer to the Appendix for complete implementation details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.909, + 0.52 + ], + "angle": 0, + "content": "Pseudo-label source \\(M\\). We consider three variants for \\(M\\) with varying levels of concept supervision. First, we use off-the-shelf supervised (ResNet18-based) concept classifiers. Second, with no concept supervision, we use CLIP zero-shot classifier [35] with only the concept names to obtain concept pseudo-labels. Third, as a compromise between the above two, we use TIP [48] which is a few-shot-labeled version of CLIP zero-shot classifier, utilizing 128 concept-labeled real images. Unless otherwise mentioned, our experiments use the supervised classifiers for a fair comparison with the prior work CBGM [16] that utilizes concept labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.909, + 0.613 + ], + "angle": 0, + "content": "Automated evaluation. Following CBGM [16], we train concept classifiers (ViT-L-16-based) on real images and concept labels with high accuracy on a held-out test set. Note that these classifiers are separate and have higher accuracy than those used for pseudo-labels. We evaluate our method using three automated metrics:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.616, + 0.909, + 0.66 + ], + "angle": 0, + "content": "- Concept Accuracy is computed over 5k generated images as the average agreement between the supervised classifiers and our proposed CB-AE or CC." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.661, + 0.909, + 0.766 + ], + "angle": 0, + "content": "- Steerability [16]: For each target concept, we find 5k latents that do not produce the target concept (i.e. probability of target concept \\(< 0.5\\) from the supervised concept classifier). For these latents, we perform the concept intervention using either the baselines or our methods, and compute the steerability as the percentage of intervened images that are classified to have the target concept." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.767, + 0.908, + 0.797 + ], + "angle": 0, + "content": "- Generation Quality is evaluated using the standard Fréchet Inception Distance (FID) [12]." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.616, + 0.909, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.8, + 0.908, + 0.845 + ], + "angle": 0, + "content": "Intuitively, the concept accuracy, FID, and steerability metrics measure how well the concept, reconstruction, and intervention objectives, respectively, are satisfied." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.848, + 0.909, + 0.893 + ], + "angle": 0, + "content": "Human evaluation. We conduct a large-scale user study on Amazon Mechanical Turk to validate the automated evaluation of concept accuracy and steerability. For both metrics," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8166" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.091, + 0.31, + 0.101 + ], + "angle": 0, + "content": "A. CB-AE StyleGAN2 on CelebA-HQ" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.102, + 0.252, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.193, + 0.251, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.102, + 0.372, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.194, + 0.371, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.102, + 0.493, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.194, + 0.491, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.091, + 0.646, + 0.101 + ], + "angle": 0, + "content": "B. CB-AE StyleGAN2 on CUB" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.102, + 0.621, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.194, + 0.619, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.102, + 0.74, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.194, + 0.738, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.102, + 0.861, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.194, + 0.857, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.242, + 0.896, + 0.257 + ], + "angle": 0, + "content": "Figure 3. Samples generated using CB-AE with CelebA-HQ and CUB pretrained StyleGAN2 models along with concept probabilities." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.271, + 0.903, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.42, + 0.776, + 0.435 + ], + "angle": 0, + "content": "Figure 4. Concept intervention examples for CB-AE with CelebA-HQ pretrained StyleGAN2." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.447, + 0.906, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.664, + 0.885, + 0.679 + ], + "angle": 0, + "content": "Figure 5. Optimization-based concept intervention (opt-int) examples for CB-AE and CC with CelebA-HQ pretrained StyleGAN2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.483, + 0.737 + ], + "angle": 0, + "content": "we display 10 images at a time and ask the user to click on images that match a displayed concept \\( c_{i}^{+} \\) (see Appendix for more details). The images are collected as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.74, + 0.483, + 0.8 + ], + "angle": 0, + "content": "- Concept Accuracy: We collect and shuffle generated images and save their CB-AE and CC concept predictions \\((c_{i}^{+}\\) or \\(c_{i}^{-})\\). Based on user responses, we compute human agreement rate w.r.t. CB-AE/CC predictions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.801, + 0.484, + 0.892 + ], + "angle": 0, + "content": "- Steerability: We collect and shuffle generated images of concept \\( c_{i}^{-} \\) with intervened images of concept \\( c_{i}^{+} \\). Based on user responses, we compute human agreement rate on whether intervened images actually contain concept \\( c_{i}^{+} \\). Shuffling the original and intervened images ensures that users are not biased towards clicking all images." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.74, + 0.484, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.693, + 0.908, + 0.753 + ], + "angle": 0, + "content": "We evaluate using approximately 100 images per concept and per method for two concepts from CelebA-HQ: “smiling” and “male/female”. Each set of 10 images is evaluated by 3 different users. Refer to the Appendix for more details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.759, + 0.634, + 0.773 + ], + "angle": 0, + "content": "4.2. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Qualitative evaluation. In Fig. 3, we present CB-AE-StyleGAN2 generated images for CelebA-HQ and CUB datasets with corresponding concept predictions. We also visualize CB-AE interventions in Fig. 4 and optimization-based interventions in Fig. 5. Interestingly, with the same latent \\( w \\), both CB-AE and CC lead to very similar optimization-based interventions, which is reasonable since the same pseudo-label source \\( M \\) was used in both cases." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8167" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.089, + 0.862, + 0.103 + ], + "angle": 0, + "content": "Table 2. Concept accuracy evaluation on 5k samples with 8 concepts for CelebA, CelebA-HQ, and 10 concepts for CUB." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.104, + 0.904, + 0.179 + ], + "angle": 0, + "content": "
Conc. Acc. (%)CelebA (64×64)CelebA-HQ (256×256)CUB (64×64)CUB (256×256)
GANPGANDDPMDDPMStyGAN2PGANDDPMGANStyGAN2
Ours (CB-AE)86.5687.8784.9889.7986.0482.6872.3574.4181.33
Ours (CC)87.6590.0085.1389.8283.5783.9472.8775.6081.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.191, + 0.907, + 0.234 + ], + "angle": 0, + "content": "Table 3. Extended steerability evaluation on 5k samples with 8 concepts for CelebA, CelebA-HQ, and 10 concepts for CUB. \\( {}^{ \\dagger } \\) CBGM numbers are from their paper (1k samples) since their results are not reproducible using their released code. For CBGM training time, we used the time taken by their code for GAN and base model training times for other models (whose code was not available) with 1 V100 GPU." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.235, + 0.905, + 0.355 + ], + "angle": 0, + "content": "
Steerability (%)CelebA (64×64)CelebA-HQ (256×256)CUB (64×64)CUB (256×256)
GANPGANDDPMDDPMStyGAN2PGANDDPMGANStyGAN2
CBGM [16]†25.60-13.80---14.8021.30-
Ours (CB-AE)47.3440.3123.7223.5040.2729.3125.6420.8910.52
Ours (CB-AE+opt-int)61.1441.7338.0950.4961.6632.1036.9446.0365.11
Ours (CC+opt-int)51.1458.9441.4556.7067.9547.2944.4748.9144.72
Train time reductionCB-AE3.7×5.4×4×8.1×3.5×2×3.7×3.3×3.1×
w.r.t. CBGMCC30.9×20.3×10×28.9×8.3×6.7×8.5×21×7.1×
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.379, + 0.482, + 0.424 + ], + "angle": 0, + "content": "Overall, we find that the optimization-based interventions are relatively higher quality and more orthogonal (i.e. less changes to other concepts) than the CB-AE interventions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.426, + 0.483, + 0.517 + ], + "angle": 0, + "content": "Concept Accuracy. In Table 2, we report the concept accuracies for our CB-AE and CC. We could not compare with CBGM since they do not evaluate concept accuracy, and since we could not reproduce their results. Overall, we find that both CB-AE and CC achieve good concept accuracies across various datasets, models and image resolutions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.52, + 0.483, + 0.655 + ], + "angle": 0, + "content": "In most scenarios, CC outperforms CB-AE since concept alignment is the sole objective optimized in CC, while CB-AE has additional objectives. However, we observe that CB-AE outperforms CC only for StyleGAN2. This is because the multiple objectives in CB-AE training may be easier to balance when dealing with clean latents (GANs) than with noisy latents (DDPM), leading to CB-AE outperforming CC for StyleGAN2. This is supported by average loss for CB-AE being lower for StyleGAN2 (0.68) than DDPM (0.92)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.658, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Steerability. We compare the concept steerability of our methods on 1k samples with GAN intervention methods like conditional GAN (CGAN) [27], auxiliary classifier GAN (ACGAN) [30], and CBGM (CB-GAN) [16], and with diffusion model intervention methods like classifier-guided (CG) [6] DDPM, classifier-free (CF) [14] DDPM, and concept bottleneck DDPM (CB-DDPM) [16] in Table 4. We observe significant gains over CBGM [16] for both GAN (average \\(+14.2\\%\\) for CB-AE, \\(+31.1\\%\\) for CB-AE with opt-int, and \\(+26.6\\%\\) for CC) and for DDPM (average \\(+10.4\\%\\) for CB-AE, \\(+23.3\\%\\) for CB-AE with opt-int, and \\(+29.1\\%\\) for CC)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.827, + 0.484, + 0.886 + ], + "angle": 0, + "content": "Table 3 presents extended steerability evaluation on \\(5\\mathrm{k}\\) samples with other types of GANs like Progressive GAN [17] and StyleGAN2 [18] for higher resolution \\((256\\times 256)\\) datasets. We could not obtain CBGM's results for these set" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.376, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Table 4. Steerability comparisons with CBGM [16] and other baseline intervention methods computed on 1k samples (each experiment repeated three times for mean and standard deviation)." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.635 + ], + "angle": 0, + "content": "
Concept RegimeSmall balanced conceptsLarge unbalanced concepts
DatasetCUB (10 conc.)CelebA (8 conc.)CelebA (40 conc.)
Baseline Intervention Methods
CGAN [27]5.4 ± 0.48.7 ± 1.32.9 ± 0.0
ACGAN [30]18.5 ± 0.49.2 ± 0.71.2 ± 0.1
CB-GAN [16]21.3 ± 0.325.6 ± 0.523.1 ± 0.2
Our Methods
CB-AE-GAN19.9 ± 0.947.1 ± 0.545.5 ± 0.4
CB-AE-GAN+opt-int45.1 ± 1.159.9 ± 1.858.3 ± 1.3
CC-GAN+opt-int49.3 ± 0.650.8 ± 1.249.7 ± 0.9
Baseline Intervention Methods
CF-DDPM [14]2.7 ± 1.97.2 ± 3.85.1 ± 2.4
CG-DDPM [6]2.1 ± 1.46.8 ± 1.15.4 ± 2.6
CB-DDPM [16]14.8 ± 6.213.8 ± 2.712.6 ± 1.7
Our Methods
CB-AE-DDPM25.8 ± 1.123.1 ± 0.923.6 ± 1.0
CB-AE-DDPM+opt-int37.3 ± 1.537.5 ± 1.336.2 ± 1.1
CC-DDPM+opt-int45.4 ± 2.241.8 ± 1.841.3 ± 1.5
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.672, + 0.907, + 0.793 + ], + "angle": 0, + "content": "tings using their code. Overall, we find optimization-based interventions (opt-int) outperform the CB-AE intervention method, and CC generally outperforms CB-AE. Intuitively, opt-int involves instance-specific and iterative optimization, while CB-AE is applied in the same way to all samples. Although CB-AE is trained with intervention losses, its steerability tends to be worse than CC since CB-AE is more challenging to train, with multiple objectives to satisfy." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Generation quality. We compare the generation quality of CBGM [16] with CB-AE and CC in Table 5. For CB-AE, we observe a relatively higher drop in image quality than CBGM, but our methods are trained \\(3.5 - 8 \\times\\) faster, and do not require training from scratch. For a high-quality StyleGAN2 model, our CB-AE and CC with optimization-based interventions can produce almost the same quality" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8168" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.131 + ], + "angle": 0, + "content": "Table 5. Generation quality and training time comparisons for CelebA-HQ with StyleGAN2. \\( {}^{ \\dagger } \\) CBGM results are from their paper. Training time is in V100 GPU-hours." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.132, + 0.482, + 0.234 + ], + "angle": 0, + "content": "
FID (↓)CBGM† [16]CB-AE (Ours)CC (Ours)
Base model9.07.667.66
CB model9.19.52-
CB interv.-9.65-
Opt-interv.-7.677.65
Train time (hrs)50146
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.249, + 0.483, + 0.276 + ], + "angle": 0, + "content": "Table 6. Human evaluation results for CB-AE and CC trained with CelebA-HQ pretrained StyleGAN2." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.278, + 0.482, + 0.436 + ], + "angle": 0, + "content": "
Conc. Acc. (%)SmilingMale/Female
AutomatedHumanAutomatedHuman
CB-AE (Ours)92.3886.35100.094.06
CC (Ours)89.4780.3596.9696.30
Steerability (%)SmilingFemale
AutomatedHumanAutomatedHuman
CB-AE (Ours)65.9077.2717.0217.73
CB-AE w/ opt-int (Ours)76.5978.7242.8541.50
CC w/ opt-int (Ours)77.3677.3626.9219.87
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.494 + ], + "angle": 0, + "content": "of images as the base model while having high steerability (61.66% and 67.95% respectively, from Table 3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.484, + 0.572 + ], + "angle": 0, + "content": "Intuitively, since CB-AE focuses on concept and intervention losses, we obtain better steerability, while CBGM has better FID since they include generative model losses. Hence, there is a tradeoff between image quality and interpretability which can be improved in future work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.485, + 0.681 + ], + "angle": 0, + "content": "Human evaluation. In Table 6, we compare human agreement rate with automated evaluation of concept accuracy and steerability. We chose easily recognizable concepts, \"gender\" and \"smiling\" as representative of low and high steerability respectively. Overall, our automated evaluation is similar to human agreement (with room for improving the classifiers), validating the usefulness of automated evaluation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.194, + 0.712 + ], + "angle": 0, + "content": "4.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.484, + 0.868 + ], + "angle": 0, + "content": "Ablation study. We analyze the contribution of each CB-AE training loss. However, we do not ablate concept loss \\(\\mathcal{L}_c\\) or latent reconstruction loss \\(\\mathcal{L}_{r_1}\\) since our evaluation metrics would be meaningless if the CB-AE cannot predict the concepts or cannot reconstruct the generator latent \\(w'\\). Hence, Table 7 ablates image reconstruction loss \\(\\mathcal{L}_{r_2}\\) from Eq. (1), intervened concept loss \\(\\mathcal{L}_{i_1}\\) from Eq. (3), and cyclic intervened concept loss \\(\\mathcal{L}_{i_2}\\) from Eq. (4). The ablations are for the most challenging pseudo-label setting of \\(M\\) (CLIP-zero-shot), since it is most affected by loss ablations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In Table 7, we observe that using the image reconstruction loss \\(\\mathcal{L}_{r_2}\\) improves the generation quality FID (row #2" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.159 + ], + "angle": 0, + "content": "Table 7. Ablation study on CB-AE training objectives for the most challenging CLIP-zero-shot pseudo-label setting for CelebAHQ pretrained StyleGAN2. \\(\\mathcal{L}_{r_2},\\mathcal{L}_{i_1},\\mathcal{L}_{i_2}\\) indicate image reconstruction loss, intervened concept loss, and intervened cyclic loss respectively from Eq. (1), (3), (4)." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.16, + 0.905, + 0.26 + ], + "angle": 0, + "content": "
Row #\\( {\\mathcal{L}}_{{r}_{2}} \\)\\( {\\mathcal{L}}_{{i}_{1}} \\)\\( {\\mathcal{L}}_{{i}_{2}} \\)Trained with \\( M = \\) CLIP-zero-shot
Conc. Acc. (%)Steerability (%)FID (↓)
1✘✘✘58.876.0111.06
2✓✘✘59.326.519.88
3✓✓✘57.759.9812.42
4✓✘✓58.3313.0313.62
5✓✓✓67.2620.6112.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.273, + 0.907, + 0.329 + ], + "angle": 0, + "content": "Table 8. Sensitivity to pseudo-label source \\(M\\). For CB-AE trained with CelebA-HQ pretrained StyleGAN2, we compare concept accuracy and steerability with different \\(M\\): CLIP-zero-shot [35], TIP-few-shot [48], or supervised classifiers." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.33, + 0.903, + 0.41 + ], + "angle": 0, + "content": "
Pseudo-label source MConc. Acc. (%)Steerability (%)
CB-AECB-AE w/ opt-int
CLIP-zs67.2620.6129.23
TIP-fs-12876.0821.5138.73
Supervised-clsf86.0440.2761.66
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.909, + 0.727 + ], + "angle": 0, + "content": "vs. #1), which is intuitive since this loss directly encourages the images to be closer to the original ones. Next, using only the intervened concept loss or only the intervened cyclic loss improves the steerability while trading-off generation quality (row #3 vs. #2 or row #4 vs. #2). Finally, using both intervened losses significantly improves both concept accuracy and steerability (row #5 vs. #3 and #4), while generation quality remains similar. Overall, both intervention losses are crucial to ensure good concept accuracy and steerability, while image reconstruction loss improves generation quality. Sensitivity to pseudo-label source \\( M \\). Table 8 compares concept accuracy and steerability when varying pseudo-label source \\( M \\). Concept accuracy improves as pseudo-label quality improves from CLIP-zero-shot to supervised classifiers. While TIP few-shot does not improve CB-AE interventions, it significantly improves optimization-based interventions and highlights the usefulness of even limited labeled data. Further, our method can use newer CLIP models like SigLIP [47], OpenCLIP [5] to further improve performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.733, + 0.634, + 0.749 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.909, + 0.895 + ], + "angle": 0, + "content": "In this work, we proposed two novel and low-cost methods, concept-bottleneck autoencoder (CB-AE) and concept controller (CC), to efficiently build interpretable generative models from pretrained models. Compared to the prior approach that struggles with efficiency and scalability, our methods achieve \\(4 - 15 \\times\\) faster training, require minimal to no concept supervision, and generalize across modern generative model families including GANs and diffusion models with \\(25\\%\\) improved steerability on average." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8169" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.118, + 0.486, + 0.376 + ], + "angle": 0, + "content": "This work is supported in part by National Science Foundation (NSF) awards CNS-1730158, ACI-1540112, ACI-1541349, OAC-1826967, OAC-2112167, CNS-2100237, CNS-2120019, the University of California Office of the President, and the University of California San Diego's California Institute for Telecommunications and Information Technology/Qualcomm Institute. This work used Delta CPU, GPU and Storage through allocation CIS230153 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support program, which is supported by National Science Foundation grants 2138259, 2138286, 2138307, 2137603, and 2138296. The authors are partially supported by National Science Foundation under Grant No. 2107189, 2313105, 2430539, Hellman Fellowship, and Intel Rising Star Faculty Award. The authors would also like to thank anonymous reviewers for valuable feedback to improve the manuscript." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.389, + 0.188, + 0.405 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.415, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[1] Dylan M Anstine and Alexandr Isayev. Generative models as an emerging paradigm in the chemical sciences. Journal of the American Chemical Society, 145(16):8736-8750, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.484, + 0.511 + ], + "angle": 0, + "content": "[2] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B. Tenenbaum, William T. Freeman, and Antonio Torralba. GAN Dissection: Visualizing and understanding generative adversarial networks. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[3] David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, and Antonio Torralba. Rewriting a deep generative model. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.554, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[4] Ricky TQ Chen, Xuechen Li, Roger B Grosse, and David K Duvenaud. Isolating sources of disentanglement in variational autoencoders. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.596, + 0.484, + 0.663 + ], + "angle": 0, + "content": "[5] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.484, + 0.692 + ], + "angle": 0, + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.693, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[7] Zheng Ding, Yifan Xu, Weijian Xu, Gaurav Parmar, Yang Yang, Max Welling, and Zhuowen Tu. Guided variational autoencoder for disentanglement learning. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.735, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[8] Rohit Gandikota, Joanna Materzyńska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.777, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[10] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.415, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "data from generative models ready for image recognition? In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.908, + 0.245 + ], + "angle": 0, + "content": "[13] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-VAE: Learning basic visual concepts with a constrained variational framework. In ICLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.907, + 0.304 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[16] Aya Abdelsalam Ismail, Julius Adebayo, Hector Corrada Bravo, Stephen Ra, and Kyunghyun Cho. Concept bottleneck generative models. In ICLR, 2024. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.907, + 0.389 + ], + "angle": 0, + "content": "[17] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.907, + 0.473 + ], + "angle": 0, + "content": "[19] Pushkal Katara, Zhou Xian, and Katerina Fragkiadaki. Gen2Sim: Scaling up robot learning in simulation with generative models. In ICRA, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.907, + 0.517 + ], + "angle": 0, + "content": "[20] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In ICML, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.907, + 0.559 + ], + "angle": 0, + "content": "[21] Sonia Laguna, Ricards Marcinkevics, Moritz Vandenhirtz, and Julia E. Vogt. Beyond concept bottleneck models: How to make black boxes intervenable? In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.907, + 0.602 + ], + "angle": 0, + "content": "[22] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.907, + 0.671 + ], + "angle": 0, + "content": "[23] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.907, + 0.701 + ], + "angle": 0, + "content": "[24] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.908, + 0.743 + ], + "angle": 0, + "content": "[25] Emanuele Marconato, Andrea Passerini, and Stefano Teso. GlanceNets: Interpretable, leak-proof concept-based models. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.907, + 0.786 + ], + "angle": 0, + "content": "[26] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[27] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[28] Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. Fast model editing at scale. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[29] Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. Memory-based model editing at scale. In ICML, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8170" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.133 + ], + "angle": 0, + "content": "[30] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier GANs. In ICML, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.485, + 0.175 + ], + "angle": 0, + "content": "[31] Tuomas Oikarinen and Tsui-Wei Weng. CLIP-Dissect: Automatic description of neuron representations in deep vision networks. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.217 + ], + "angle": 0, + "content": "[32] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[33] Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. PreciseControl: Enhancing text-to-image diffusion models with fine-grained attribute control. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[34] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.484, + 0.397 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[36] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.482 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.484, + 0.551 + ], + "angle": 0, + "content": "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.484, + 0.593 + ], + "angle": 0, + "content": "[39] Divyansh Srivastava, Ge Yan, and Tsui-Wei Weng. VLGCBM: Training concept bottleneck models with vision-language guidance. In NeurIPS, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.634 + ], + "angle": 0, + "content": "[40] Chung-En Sun, Tuomas Oikarinen, Berk Ustun, and Tsui-Wei Weng. Concept bottleneck large language models. In ICLR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.484, + 0.69 + ], + "angle": 0, + "content": "[41] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech-UCSD Birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.484, + 0.719 + ], + "angle": 0, + "content": "[42] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In ICLR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.484, + 0.774 + ], + "angle": 0, + "content": "[43] An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Yang Wang, Jingbo Shang, and Julian McAuley. Learning concise and descriptive attributes for visual recognition. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[44] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.859 + ], + "angle": 0, + "content": "[45] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[46] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "Melacci, Adrian Weller, Pietro Lio, and Mateja Jamnik. Concept embedding models: Beyond the accuracy-explainability trade-off. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.175 + ], + "angle": 0, + "content": "[47] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[48] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. In ECCV, 2022. 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8171" + } + ] +] \ No newline at end of file diff --git a/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/86d682d5-19b8-4583-bdec-4fcac9608ef2_origin.pdf b/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/86d682d5-19b8-4583-bdec-4fcac9608ef2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2dca34d069e76ce6012f9b54893ca7c6c11f162c --- /dev/null +++ b/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/86d682d5-19b8-4583-bdec-4fcac9608ef2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7814add59ee7d1b8117e16187d5a35e5ac5666918d79210b8334173debd0115 +size 1856345 diff --git a/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/full.md b/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..27cc8eea2bb15823af05e73872a0b4f3c9123766 --- /dev/null +++ b/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/full.md @@ -0,0 +1,341 @@ +# Interpretable Generative Models through Post-hoc Concept Bottlenecks + +Akshay Kulkarni, Ge Yan,\* Chung-En Sun\* Tuomas Oikarinen, and Tsui-Wei Weng +University of California San Diego + +{a2kulkarni, lweng}@ucsd.edu + +# Abstract + +Concept bottleneck models (CBM) aim to produce inherently interpretable models that rely on human-understandable concepts for their predictions. However, existing approaches to design interpretable generative models based on CBMs are not yet efficient and scalable, as they require expensive generative model training from scratch as well as real images with labor-intensive concept supervision. To address these challenges, we present two novel and low-cost methods to build interpretable generative models through post-hoc techniques and we name our approaches: concept-bottleneck autoencoder (CB-AE) and concept controller (CC). Our proposed approaches enable efficient and scalable training without the need of real data and require only minimal to no concept supervision. Additionally, our methods generalize across modern generative model families including generative adversarial networks and diffusion models. We demonstrate the superior interpretability and steerability of our methods on numerous standard datasets like CelebA, CelebA-HQ, and CUB with large improvements (average $\sim 25\%$ ) over the prior work, while being $4 - 15\times$ faster to train. Finally, a large-scale user study is performed to validate the interpretability and steerability of our methods. + +# 1. Introduction + +Deep generative models [36-38] have become increasingly powerful and widely used in many high-stakes domains and applications, including realistic data generation [11], simulating hypothetical scenarios or environments [19], and scientific discovery [1]. It is therefore important to ensure that the generation process is interpretable, which will allow us to understand and audit the generation, and further mitigate potential biases and harms (e.g. content moderation). + +Unfortunately, most of the advances in deep learning utilize complex, black-box neural network architectures that are difficult to interpret and understand. This leads to user mistrust of model predictions due to the absence of explanations. + +A. Prior Work (Concept Bottleneck Generative Models) + +![](images/be9466677132d035e0060822dda2d2d51546b062bbba8e437b494ff60fed0341.jpg) + +Expensive generative model training from scratch! + +B. Ours (Post-Hoc Interpretable Generative Models) + +![](images/1cd5a99936fa27a64d82cc0a2b996e7db5f7bfda7726c968421616b04675b215.jpg) + +Efficiently train only CB-AE/CC with frozen generative model + +C. Qualitative Results + +![](images/0611e74cc404650e3df77600b540932c17935f3b555392ebecd2864825994393.jpg) +Figure 1. A. Prior work on interpretable generative models requires expensive generative model training from scratch. B. Our CB-AE and CC can be trained efficiently for post-hoc interpretability in a pretrained, frozen generative model $g_{2} \circ g_{1}$ . C. Example concept intervention with CB-AE and corresponding concept vectors. + +To address this, there has been work on developing inherently interpretable deep vision models using concept bottlenecks [20, 32, 39, 43-45]. These approaches train a concept bottleneck layer after the feature extractor (backbone) to embed a set of human-understandable concepts, followed by an interpretable sparse linear layer for the final classification based on the concept prediction. However, current development of CBMs is primarily focused on classification tasks, and only one prior work, CBGM [16], has extended it to image generation, indicating this area is under-explored. + +The key idea of the seminal work CBGM [16] is to represent concepts as learnable embeddings [46] at an inter + +Table 1. Comparison of our CB-AE and CC with prior work CBGM [16] (green indicates desirable properties). CC trades-off inherent interpretability for better steerability, image quality, and faster training. We could not reproduce CBGM [16] to evaluate concept accuracy. + +
MethodPost-hoc trainingTraining without concept-labeled real imagesInherently interpretable modelConcept Acc. (%)Steerability (%) (↑)FID (↓)Train time (V100-hrs) (↓)
StyleGAN2DDPM
CBGM [16]XX✓-25.609.1050240
CB-AE (Ours)✓✓✓86.5647.34 (+21.74)9.5214 (3.5× faster)29.5 (8.1× faster)
CC (Ours)✓✓X87.6551.14 (+25.54)7.656 (8.3× faster)8.3 (28.9× faster)
+ +mediate location in the generative model, and combine the embeddings to compute the generative model latent. However, the CBGM-based generative model has to be trained from scratch using concept-labeled real images, which could be difficult to scale and computationally intensive (e.g. 240 V100-hours for DDPM- $256 \times 256$ [15]) as shown in Fig. 1A. To address these limitations, in this work, we develop an efficient and scalable post-hoc concept bottleneck to transform any pretrained generative model into an interpretable model. In contrast to CBGM [16], our approach works well by only training a few layers with minimal concept supervision. + +Specifically, we propose a novel concept-bottleneck autoencoder (CB-AE) $f$ that can be inserted into the intermediate layers of a pretrained generative model $g = g_{2} \circ g_{1}$ as shown in Fig. 1B (top). The CB-AE input and output are the generator latent $w$ and its reconstruction $w'$ respectively, and the overall generative model now becomes $g_{2} \circ f \circ g_{1}$ . The CB-AE latent space is the concept space $c$ which is used to reconstruct the generator latent space. In our framework, given a pretrained generative model $g = g_{2} \circ g_{1}$ , only the encoder $E$ and decoder $D$ in the CB-AE $f = D \circ E$ need to be trained, while $g_{1}, g_{2}$ are frozen pretrained weights, i.e. CB-AE uses post-hoc training. The benefit of the proposed CB-AE is that it allows us to debug the model easily with concept-level control by modifying the CB-AE concept latent during image generation (Fig. 1C). Further, our training requires only concept pseudo-labels achievable with minimal concept supervision (e.g. zero-shot CLIP classifier [35]). + +We also propose novel optimization-based concept interventions to achieve higher success rate and intervention quality. Based on the CB-AE, we propose an even more efficient post-hoc concept controller (CC) method (Fig. 1B, bottom) with simplified training, that can provide concept predictions and leverage optimization-based concept interventions. Note that while CB-AE is part of the new interpretable generative model, CC is a post-hoc control method that is not part of the generative model. Finally, we evaluate our CB-AE and CC on various generative models, including GANs and diffusion models, for standard datasets including CelebA, CelebA-HQ, and CUB. We show that CB-AE and CC significantly outperform prior work CBGM [16] w.r.t. steerability (or intervention success rate) on CelebA (average $+23\%$ ) while also having $4 - 15 \times$ faster training (Table 1). + +Our contributions can be summarized as follows: + +- We are the first to propose a post-hoc concept bottleneck + +autoencoder (CB-AE) for interpretable generative models. CB-AE can be trained efficiently with a frozen pretrained generative model, without real concept-labeled images. + +- We also propose a novel and efficient optimization-based concept intervention method with improved steerability (avg. $+19\%$ ) and higher image quality (avg. $32\%$ better). +- We validate the effectiveness of our methods for GANs and diffusion models (avg. $+31\%$ and $+28\%$ steerability w.r.t. prior state-of-the-art) across varying image resolutions, while being $4 - 15 \times$ faster to train on average. + +# 2. Related Work + +Concept Bottleneck Models. Early work on CBMs [20, 25] relied on concept-labeled images to train a concept bottleneck layer with each neuron as a human-understandable concept, followed by a linear layer based on the concepts for the final classification. Post-hoc CBMs [45] extended this idea to convert a pretrained backbone into a CBM. More recent works like LF-CBM [32], LM4CV [43], LaBo [44], and VLG-CBM [39] use interpretability tools [31], vision-language models like CLIP [35], large language models, or open-set object detection [23] to eliminate the need for expensive concept-labeled data. Independently from our work, [21] proposed a concept-based intervention without a concept bottleneck, similar to our CC, but for classification. All these works are specifically designed for classification, while in this paper, our focus is on the image generation task. + +Interpretability for Generative Models. A line of work [4, 7, 13] on learning disentangled concepts in variational autoencoders enables controllable generation, but they train from scratch and are not applicable to other generative models like GANs. Other works focus on identifying and manipulating structural rules or concepts in GANs [2, 3] and large language models [26, 28, 29, 40] by editing the model weights. In contrast, we focus on training inherently interpretable models, and the closest prior work is the recent CBGM [16], which also aims to build interpretable generative models, but requires expensive concept labels and training from scratch, limiting its scalability. In contrast, our proposed CB-AE can be trained efficiently with a frozen pretrained generative model with minimal concept supervision. + +Image Editing in Generative Models. Some works focus on conditional generation [37, 38] and image editing in generative models by modifying model weights [8, 33]. In + +![](images/e5b1e44a308162f613a686ee388e9e52e66bcaf2ba5a63c549e9c76ce41bc70a.jpg) +A. Original Black-Box Generative Model + +![](images/963c465ee94a7ef306745ec46b45e7927822556a896a8cde8a37d8a692fd73b7.jpg) +B. Our Post-hoc Interpretable Model + +![](images/19fde05c03a67d6228634530d9727b08fc39a05cf034a892d49f837375115055.jpg) +C. Training Objective 1 (reconstruction losses) + +![](images/0cb9a9bc5780227cadab60a4bbd3dc68d2bc7632881533d4e3a90eb790106089.jpg) +D. Training Objective 2 (concept alignment loss) + +![](images/c874f1f1b3809d302ea572fea3a9df915b618d66f5759ed071500a8dff990fe2.jpg) +E. Training Objective 3 (intervention losses) +Figure 2. Post-hoc CB-AE training for reconstruction, concept alignment, and intervention with a frozen pretrained generator $g_{2} \circ g_{1}$ . Note that $\mathcal{L}_c$ , $\mathcal{L}_i$ indicate cross-entropy loss and $\mathcal{L}_r$ indicates mean-squared-error loss. The darker lines indicate gradient flow during training. + +contrast, our work focuses on developing inherently interpretable generative models by introducing additional concept bottleneck layers, where the capability of editing or intervention is naturally a by-product of the interpretability. + +# 3. Proposed Methods + +We propose a novel and low-cost concept bottleneck autoencoder (CB-AE) method in Sec. 3.1 to incorporate post-hoc interpretability in pretrained generative models. In Sec. 3.2, we present an optimization-based intervention method for concept-based steerability in the generative model. Finally, based on our insights from CB-AE and optimization-based interventions, we propose an even lower-cost post-hoc control method, concept controller (CC) in Sec. 3.3. + +Preliminaries. Consider a generative model $g: \mathcal{Z} \to \mathcal{X}$ that maps from random noise $z \in \mathcal{Z}$ to an image $x \in \mathcal{X}$ . To make the generative model $g$ become inherently interpretable, the goal of CBMs is to insert and train a concept bottleneck (say) $f$ at an intermediate location in $g$ . Let $g = g_2 \circ g_1$ , i.e. $g$ can be divided into two parts, $g_1$ and $g_2$ (e.g. for a DCGAN [34], $g_1$ is the first 2 layers of $g$ and $g_2$ is the remaining layers). The input to the concept bottleneck $f$ will be the output of $g_1$ (Fig. 1). At the bottleneck of $f$ , we obtain the concept prediction $c \in \mathcal{C}$ where $\mathcal{C}$ is the set of pre-defined concepts. In our setup, the concept vector $c$ has two logits for each binary concept $c_i$ (e.g. "smiling" would have two logits, for + +smiling and not smiling) and $N$ logits for each categorical concept $c_{j}$ with $N$ classes (e.g. blonde/black/white/gray hair color would have four logits). For example, suppose we have one binary concept $c_{i}$ and one categorical concept $c_{j}$ , then we define $c = [c_{i}^{+}, c_{i}^{-}, c_{j}^{(1)}, c_{j}^{(2)}, \ldots, c_{j}^{(N)}]^{\top}$ where $c_{i} = [c_{i}^{+}, c_{i}^{-}]^{\top} \in \mathbb{R}^{2}$ and $c_{j} = [c_{j}^{(1)}, c_{j}^{(2)}, \ldots, c_{j}^{(N)}]^{\top} \in \mathbb{R}^{N}$ . + +# 3.1. Post-hoc Concept Bottleneck Autoencoder + +We propose a concept bottleneck autoencoder (CB-AE), $f = D \circ E$ (see Fig. 2B). The latent space of the CB-AE encoder $E$ is the concept prediction $c = E(g_1(z))$ . Apart from the predefined concepts, $c$ also contains an unsupervised concept embedding, learned using autoencoder reconstruction and intervention objectives, to encode other concepts absent from the predefined set (similar to CBGM [16]). The decoder $D$ reconstructs the features from $g_1$ based on the concept prediction $c$ , outputting $w' = D(c)$ . Considering the original output of $g_1$ to be $w = g_1(z)$ , we can generate the original image $x = g_2(w)$ as well as a reconstructed image $x' = g_2(w') = g_2(D \circ E(w))$ generated using the CB-AE. + +Training. There are 3 goals for the CB-AE. First, the generator's performance should be preserved even if the CB-AE output $w'$ is used instead of $w$ . Second, CB-AE should provide interpretability for the generated images $x'$ through the corresponding concepts $c$ . Lastly, CB-AE should allow accurate steering on image generation via concept interven- + +tions. Based on these 3 goals, we formulate Objective 1-3 for CB-AE training below and illustrated in Fig. 2C-E. + +Objective 1 (reconstruction losses $\mathcal{L}_{r_1},\mathcal{L}_{r_2}$ ). At each training iteration, we sample a latent $w$ by passing uniform noise $z$ to $g_{1}$ . Then, the latent $w$ is passed through $g_{2}$ to obtain a generated image $x = g_{2}(w)$ from the original model $g = g_{2} \circ g_{1}$ , without using our CB-AE. We also reconstruct $w' = D \circ E(w)$ using our CB-AE and the same latent $w$ , and obtain another generated image $x' = g_{2}(w')$ . Since our first goal is to preserve the generator's performance when using the CB-AE, we apply reconstruction losses (mean-squared error loss $\mathcal{L}_r$ ) between the latents $w, w'$ and between the generated images $x, x'$ , as shown in Fig. 2C: + +$$ +\min _ {E, D} \left[ \mathcal {L} _ {r _ {1}} \left(w, w ^ {\prime}\right) + \mathcal {L} _ {r _ {2}} \left(x, x ^ {\prime}\right) \right], \tag {1} +$$ + +where $w' = D \circ E(w)$ and $x' = g_2(w') = g_2 \circ D \circ E(w)$ and only the CB-AE parameters (i.e. $E, D$ ) are trainable. + +Objective 2 (concept alignment loss $\mathcal{L}_c$ ). To ensure interpretability, we first obtain a concept pseudo-label $\hat{y} = M(x)$ for a generated image $x$ from a pseudo-label source $M$ (Fig. 2D). The source $M$ can be either an off-the-shelf supervised model or a zero-shot prediction pipeline from (say) CLIP with concept names $\mathcal{C}$ as the text inputs. With this approach, we avoid the requirement of any real images for training as well as the requirement of concept labels, unlike CBGM [16]. Since our second goal is to provide interpretability for the generated images through the concepts $c$ , we apply a cross-entropy loss $\mathcal{L}_c$ between the CB-AE encoder output $c = E(w)$ and the concept pseudo-label $\hat{y} = M(x)$ : + +$$ +\min _ {E} [ \mathcal {L} _ {c} (\hat {y}, c) ]. \tag {2} +$$ + +The losses in Eq. (1), (2) are simultaneously optimized to learn only the CB-AE parameters $E, D$ . + +Objective 3 (intervention losses $\mathcal{L}_{i_1},\mathcal{L}_{i_2}$ ). Interventions are an important feature of concept-bottleneck models [20], allowing users to control the model output by modifying the concepts $c$ . However, for our CB-AE decoder $D$ , reconstruction and concept alignment losses in Objective 1 and 2 do not provide guidance on how the reconstructed latent $w^{\prime}$ should change when concepts $c$ are manually modified. Hence, for steerability, we design Objective 3 (Fig. 2E) that encourages the CB-AE decoder $D$ to produce an appropriately changed and realistic latent $w^{\prime}$ when the concepts in $c$ are modified. We first describe (a) how interventions are performed in our CB-AE, followed by designing (b) an intervened concept alignment loss $\mathcal{L}_{i_1}$ and (c) a cyclic intervened concept loss $\mathcal{L}_{i_2}$ to encourage steerability, i.e. intervention success. + +a) Intervening concepts. At each training iteration, we choose a random logit for a random concept to intervene, and modify only the chosen concept based on the chosen logit to get an intervened concept vector $c_{\text{intervened}}$ (Fig. 2E). For + +example, for a desired binary concept $i \in \mathcal{C}$ , the new concept vector $c_{\text{intervened}}$ is computed by swapping the two logits, i.e. $c_{\text{intervened}} = [\ldots, c_i^-, c_i^+, \ldots]$ from $c = [\ldots, c_i^+, c_i^-, \ldots]$ . The same can be done for a categorical concept $i \in \mathcal{C}$ as well by swapping the desired logits (say) $c_i^{(k)}$ with the highest logits $c_i^{(\ell)}$ where $\ell = \arg \max_j c_i^{(j)}$ . Concretely, $c = [\ldots, c_i^{(1)}, \ldots, c_i^{(\ell)}, \ldots, c_i^{(k)}, \ldots, c_i^{(N)}, \ldots]$ is modified to $c_{\text{intervened}} = [\ldots, c_i^{(1)}, \ldots, c_i^{(k)}, \ldots, c_i^{(\ell)}, \ldots, c_i^{(N)}, \ldots]$ . We can also intervene on multiple concepts simultaneously. + +b) Designing $\mathcal{L}_{i_1}$ . Using the intervened concepts $c_{\text{intervened}}$ we reconstruct an intervened latent $w_{\text{intervened}} = D(c_{\text{intervened}})$ and obtain an intervened generated image $x_{\text{intervened}} = g_2(w_{\text{intervened}})$ . Using the pseudo-label source, we obtain an intervened concept prediction $y_{\text{intervened}} = M(x_{\text{intervened}})$ . Since we already have the concept pseudo-label $\hat{y} = M(x)$ for the original image $x$ , we modify $\hat{y}$ to $\hat{y}_{\text{intervened}}$ by changing only the earlier chosen concept $i \in \mathcal{C}$ in $\hat{y}$ to the chosen value (based on earlier chosen logit) as shown in Fig. 2E. In other words, $\hat{y}_{\text{intervened}}$ is the intervened concept pseudo-label. Now, to align the concepts in the intervened image $x_{\text{intervened}}$ with the predicted concepts from $M$ , we use a cross-entropy loss $\mathcal{L}_{i_1}$ between $\hat{y}_{\text{intervened}}$ and $y_{\text{intervened}}$ : + +$$ +\underset {E, D} {\min } \left[ \mathcal {L} _ {i _ {1}} \left(\hat {y} _ {\text {i n t e r v e n e d}}, y _ {\text {i n t e r v e n e d}}\right) \right]. \tag {3} +$$ + +c) Designing $\mathcal{L}_{i_2}$ . For cyclic consistency, we pass the intervented latent $w_{\text{intervened}}$ through the CB-AE encoder $E$ to obtain a concept prediction $c_{\text{intervened}}'$ (Fig. 2E) and apply a cross-entropy loss $\mathcal{L}_{i_2}$ w.r.t. $\hat{y}_{\text{intervened}}$ to align the encoder's prediction with that of the pseudo-label source $M$ : + +$$ +\min _ {E, D} \left[ \mathcal {L} _ {i _ {2}} \left(\hat {y} _ {\text {i n t e r v e n e d}}, c _ {\text {i n t e r v e n e d}} ^ {\prime}\right) \right]. \tag {4} +$$ + +We use $\mathcal{L}_{i_1},\mathcal{L}_{i_2}$ instead of $\mathcal{L}_c$ for cross-entropy loss to differentiate the intervention losses from the concept loss in Objective 2, and only CB-AE parameters $E,D$ are trainable. Finally, recall that the concept vector $c$ contains an unsupervised concept embedding. Objective 3 implicitly encourages this embedding to not encode known concepts, since the unsupervised embedding is not modified during the intervention from $c$ to $c_{\mathrm{intervened}}$ . + +Test-Time Intervention. Similar to training-time interventions, we can perform test-time interventions by modifying the value of any chosen concept (by swapping the logits) in the predicted concept vector $c$ to $c_{\text{intervened}}$ . The intervened image $x_{\text{intervened}} = g_2(w_{\text{intervened}})$ can be obtained where $w_{\text{intervened}} = D(c_{\text{intervened}})$ . Note that swapping the logits ensures that the range of values in $c_{\text{intervened}}$ are similar to $c$ , and also avoids the requirement of any estimation of how much to change the desired concept's logit. This makes it more accessible to users as they need not worry about the actual values being changed during an intervention. + +# 3.2. Optimization-based interventions + +For an alternative intervention method, we draw inspiration from adversarial attacks [10] to perform test-time interventions using gradient-based optimization. Specifically, we use the iterative randomized fast gradient sign method (IRFGSM) [42] on the CB-AE encoder prediction. + +Consider a generated image $x = g_{2}(w)$ with concept prediction $c = E(w)$ . To intervene in the generation process to obtain modified concepts $c^{*}$ , we solve the following objective using gradient ascent, + +$$ +w ^ {*} = w + \underset {\delta \in \Delta} {\arg \max } \left[ - \mathcal {L} _ {c} \left(E (w + \delta), c ^ {*}\right) \right] \tag {5} +$$ + +where $\Delta = \{\delta : \| \delta \|_{\infty} \leq \epsilon\}$ is the $\ell_{\infty}$ -norm bound on $\delta$ with a hyperparameter $\epsilon > 0$ . Intuitively, we optimize a small perturbation $\delta$ such that $w^{*} = w + \delta$ leads to the desired concepts $c^{*}$ . Then, the generated image $x^{*} = g_{2}(w^{*})$ is very similar to $x$ but contains concepts $c^{*}$ . + +# 3.3. Post-hoc Concept Controller (CC) for Steering + +While optimization-based interventions work well empirically, it is interesting to note that the CB-AE decoder $D$ is not involved in the process. This leads us to question whether the CB-AE decoder can be removed if a user only plans to perform optimization-based interventions. While such a removal would not result in a CBM, it would be sufficient for the purpose of steering image generation. For this particular use case, it would be even more efficient than training the CB-AE since reconstruction and intervention losses are no longer required. Hence, we propose a post-hoc concept controller (CC), denoted as $\Omega$ , that predicts the concepts, i.e. $c = \Omega(g_1(z))$ to efficiently steer image generation (Fig. 1B). + +Training. For CC, the training is simply Objective 2 of the CB-AE training, i.e. the cross-entropy loss $\mathcal{L}_c$ w.r.t. the concept pseudo-labels $\hat{y} = M(x)$ from the pseudo-label source $M$ . Formally, the objective is $\min_{\Omega}[\mathcal{L}_c(\hat{y},c)]$ where $c = \Omega (g_1(z))$ are the predicted concepts, and the loss encourages the concept controller $\Omega$ to align with the pseudo-label source $M$ . Similar to the CB-AE training, we avoid the requirement of any real images for training as well as the requirement of concept labels. + +# 4. Experiments + +We detail the experimental setup and comprehensively evaluate our proposed methods with respect to the state-of-the-art prior work as well as other baselines. + +# 4.1. Experimental setup + +Base generative models and datasets. We evaluate our CB-AE and CC methods on diverse generative models including GAN [9], Progressive GAN [17], StyleGAN2 [18], + +and DDPM [15]. We use models pretrained on standard datasets of varying image resolution ( $64 \times 64$ to $512 \times 512$ ) like CelebA [24], CelebA-HQ [22], and CUB [41]. Following CBGM [16], we evaluate on the small balanced concept regime with the 8 most balanced concepts and the large unbalanced concept regime with all 40 concepts from the dataset. For CUB, we use 10 balanced concepts as per CBGM. Please refer to the Appendix for more details. + +CB-AE and CC. We use a 4-layer MLP or 4 convolution (and transposed convolutional) layers for CB-AE encoder or CC (and decoder) depending on the dimensions of latent $w$ . We use unsupervised concept embedding $\in \mathbb{R}^{40}$ in CB-AE (ensuring bottleneck is much smaller than the latent). As in CBGM [16], we train CB-AE/CC for 50 epochs with batch size 64. For optimization-based interventions, we use 50-step I-RFGSM [42] with $\ell_{\infty}$ -norm bound $\epsilon = 0.1$ . Please refer to the Appendix for complete implementation details. + +Pseudo-label source $M$ . We consider three variants for $M$ with varying levels of concept supervision. First, we use off-the-shelf supervised (ResNet18-based) concept classifiers. Second, with no concept supervision, we use CLIP zero-shot classifier [35] with only the concept names to obtain concept pseudo-labels. Third, as a compromise between the above two, we use TIP [48] which is a few-shot-labeled version of CLIP zero-shot classifier, utilizing 128 concept-labeled real images. Unless otherwise mentioned, our experiments use the supervised classifiers for a fair comparison with the prior work CBGM [16] that utilizes concept labels. + +Automated evaluation. Following CBGM [16], we train concept classifiers (ViT-L-16-based) on real images and concept labels with high accuracy on a held-out test set. Note that these classifiers are separate and have higher accuracy than those used for pseudo-labels. We evaluate our method using three automated metrics: + +- Concept Accuracy is computed over 5k generated images as the average agreement between the supervised classifiers and our proposed CB-AE or CC. +- Steerability [16]: For each target concept, we find 5k latents that do not produce the target concept (i.e. probability of target concept $< 0.5$ from the supervised concept classifier). For these latents, we perform the concept intervention using either the baselines or our methods, and compute the steerability as the percentage of intervened images that are classified to have the target concept. +- Generation Quality is evaluated using the standard Fréchet Inception Distance (FID) [12]. + +Intuitively, the concept accuracy, FID, and steerability metrics measure how well the concept, reconstruction, and intervention objectives, respectively, are satisfied. + +Human evaluation. We conduct a large-scale user study on Amazon Mechanical Turk to validate the automated evaluation of concept accuracy and steerability. For both metrics, + +![](images/abadf901e2ebe729eedfc5f973493353d24b3ca1215b2c2c6d81598c82b4ccbb.jpg) +A. CB-AE StyleGAN2 on CelebA-HQ + +![](images/4deff7fb8ac8d7723a360e556f46462fdddc639081980a2abb4868745a1d48b1.jpg) +Figure 3. Samples generated using CB-AE with CelebA-HQ and CUB pretrained StyleGAN2 models along with concept probabilities. + +![](images/210b801d1d3bcc046ca494614c2fa4ee06fb8584ca75704f2d3d8a6c71531a07.jpg) + +![](images/cece36462c3fa0931fc1791258b546203bfbb5a15bfa522089b28b036b089d23.jpg) + +![](images/85fe15eec4bb1b9e6a7f7860fc91f267c5f8cd4a4e192fcac4fd2223775b0549.jpg) + +![](images/5eb7ce5219511705cebb109fe07d5e405aaccf676d5d6537ceee96c0f8a6d5fb.jpg) + +![](images/a645faf95ecd4839a1bf983f40429b48515c1c6ad94c364c5ffed5b4254a517e.jpg) +B. CB-AE StyleGAN2 on CUB + +![](images/178e2a9f475040ddc051204bd29cba62c2d98759ad5a0a3df4cb9c1893a0cc15.jpg) + +![](images/51529635c305d9a15922a0354397272abf612dbadabea02eac069830b0258794.jpg) + +![](images/9859344daf947a58d7b1d4c2eaf26744b86a0a84dc72d9975306d75b421ca50e.jpg) + +![](images/44f60efd3025e14550a4ff048b8160cfcb33795309b534467bc6be5b94e20cf5.jpg) + +![](images/2327f898dfce6ab323619de89ee5114b8cdc32ae4100d94d6afcaa0f26d6df1d.jpg) + +![](images/5047cd29f3a41a5859ba60944a6831cb5cf5e91ebb51764ede675f1b819831bf.jpg) +Figure 4. Concept intervention examples for CB-AE with CelebA-HQ pretrained StyleGAN2. + +![](images/6461273f56004b1dd120515518f893a6bb18776ca53e9675d8ff70be10faac87.jpg) +Figure 5. Optimization-based concept intervention (opt-int) examples for CB-AE and CC with CelebA-HQ pretrained StyleGAN2. + +we display 10 images at a time and ask the user to click on images that match a displayed concept $c_{i}^{+}$ (see Appendix for more details). The images are collected as follows: + +- Concept Accuracy: We collect and shuffle generated images and save their CB-AE and CC concept predictions $(c_{i}^{+}$ or $c_{i}^{-})$ . Based on user responses, we compute human agreement rate w.r.t. CB-AE/CC predictions. +- Steerability: We collect and shuffle generated images of concept $c_{i}^{-}$ with intervened images of concept $c_{i}^{+}$ . Based on user responses, we compute human agreement rate on whether intervened images actually contain concept $c_{i}^{+}$ . Shuffling the original and intervened images ensures that users are not biased towards clicking all images. + +We evaluate using approximately 100 images per concept and per method for two concepts from CelebA-HQ: “smiling” and “male/female”. Each set of 10 images is evaluated by 3 different users. Refer to the Appendix for more details. + +# 4.2. Evaluation + +Qualitative evaluation. In Fig. 3, we present CB-AE-StyleGAN2 generated images for CelebA-HQ and CUB datasets with corresponding concept predictions. We also visualize CB-AE interventions in Fig. 4 and optimization-based interventions in Fig. 5. Interestingly, with the same latent $w$ , both CB-AE and CC lead to very similar optimization-based interventions, which is reasonable since the same pseudo-label source $M$ was used in both cases. + +Table 2. Concept accuracy evaluation on 5k samples with 8 concepts for CelebA, CelebA-HQ, and 10 concepts for CUB. + +
Conc. Acc. (%)CelebA (64×64)CelebA-HQ (256×256)CUB (64×64)CUB (256×256)
GANPGANDDPMDDPMStyGAN2PGANDDPMGANStyGAN2
Ours (CB-AE)86.5687.8784.9889.7986.0482.6872.3574.4181.33
Ours (CC)87.6590.0085.1389.8283.5783.9472.8775.6081.11
+ +Table 3. Extended steerability evaluation on 5k samples with 8 concepts for CelebA, CelebA-HQ, and 10 concepts for CUB. ${}^{ \dagger }$ CBGM numbers are from their paper (1k samples) since their results are not reproducible using their released code. For CBGM training time, we used the time taken by their code for GAN and base model training times for other models (whose code was not available) with 1 V100 GPU. + +
Steerability (%)CelebA (64×64)CelebA-HQ (256×256)CUB (64×64)CUB (256×256)
GANPGANDDPMDDPMStyGAN2PGANDDPMGANStyGAN2
CBGM [16]†25.60-13.80---14.8021.30-
Ours (CB-AE)47.3440.3123.7223.5040.2729.3125.6420.8910.52
Ours (CB-AE+opt-int)61.1441.7338.0950.4961.6632.1036.9446.0365.11
Ours (CC+opt-int)51.1458.9441.4556.7067.9547.2944.4748.9144.72
Train time reductionCB-AE3.7×5.4×4×8.1×3.5×2×3.7×3.3×3.1×
w.r.t. CBGMCC30.9×20.3×10×28.9×8.3×6.7×8.5×21×7.1×
+ +Overall, we find that the optimization-based interventions are relatively higher quality and more orthogonal (i.e. less changes to other concepts) than the CB-AE interventions. + +Concept Accuracy. In Table 2, we report the concept accuracies for our CB-AE and CC. We could not compare with CBGM since they do not evaluate concept accuracy, and since we could not reproduce their results. Overall, we find that both CB-AE and CC achieve good concept accuracies across various datasets, models and image resolutions. + +In most scenarios, CC outperforms CB-AE since concept alignment is the sole objective optimized in CC, while CB-AE has additional objectives. However, we observe that CB-AE outperforms CC only for StyleGAN2. This is because the multiple objectives in CB-AE training may be easier to balance when dealing with clean latents (GANs) than with noisy latents (DDPM), leading to CB-AE outperforming CC for StyleGAN2. This is supported by average loss for CB-AE being lower for StyleGAN2 (0.68) than DDPM (0.92). + +Steerability. We compare the concept steerability of our methods on 1k samples with GAN intervention methods like conditional GAN (CGAN) [27], auxiliary classifier GAN (ACGAN) [30], and CBGM (CB-GAN) [16], and with diffusion model intervention methods like classifier-guided (CG) [6] DDPM, classifier-free (CF) [14] DDPM, and concept bottleneck DDPM (CB-DDPM) [16] in Table 4. We observe significant gains over CBGM [16] for both GAN (average $+14.2\%$ for CB-AE, $+31.1\%$ for CB-AE with opt-int, and $+26.6\%$ for CC) and for DDPM (average $+10.4\%$ for CB-AE, $+23.3\%$ for CB-AE with opt-int, and $+29.1\%$ for CC). + +Table 3 presents extended steerability evaluation on $5\mathrm{k}$ samples with other types of GANs like Progressive GAN [17] and StyleGAN2 [18] for higher resolution $(256\times 256)$ datasets. We could not obtain CBGM's results for these set + +Table 4. Steerability comparisons with CBGM [16] and other baseline intervention methods computed on 1k samples (each experiment repeated three times for mean and standard deviation). + +
Concept RegimeSmall balanced conceptsLarge unbalanced concepts
DatasetCUB (10 conc.)CelebA (8 conc.)CelebA (40 conc.)
Baseline Intervention Methods
CGAN [27]5.4 ± 0.48.7 ± 1.32.9 ± 0.0
ACGAN [30]18.5 ± 0.49.2 ± 0.71.2 ± 0.1
CB-GAN [16]21.3 ± 0.325.6 ± 0.523.1 ± 0.2
Our Methods
CB-AE-GAN19.9 ± 0.947.1 ± 0.545.5 ± 0.4
CB-AE-GAN+opt-int45.1 ± 1.159.9 ± 1.858.3 ± 1.3
CC-GAN+opt-int49.3 ± 0.650.8 ± 1.249.7 ± 0.9
Baseline Intervention Methods
CF-DDPM [14]2.7 ± 1.97.2 ± 3.85.1 ± 2.4
CG-DDPM [6]2.1 ± 1.46.8 ± 1.15.4 ± 2.6
CB-DDPM [16]14.8 ± 6.213.8 ± 2.712.6 ± 1.7
Our Methods
CB-AE-DDPM25.8 ± 1.123.1 ± 0.923.6 ± 1.0
CB-AE-DDPM+opt-int37.3 ± 1.537.5 ± 1.336.2 ± 1.1
CC-DDPM+opt-int45.4 ± 2.241.8 ± 1.841.3 ± 1.5
+ +tings using their code. Overall, we find optimization-based interventions (opt-int) outperform the CB-AE intervention method, and CC generally outperforms CB-AE. Intuitively, opt-int involves instance-specific and iterative optimization, while CB-AE is applied in the same way to all samples. Although CB-AE is trained with intervention losses, its steerability tends to be worse than CC since CB-AE is more challenging to train, with multiple objectives to satisfy. + +Generation quality. We compare the generation quality of CBGM [16] with CB-AE and CC in Table 5. For CB-AE, we observe a relatively higher drop in image quality than CBGM, but our methods are trained $3.5 - 8 \times$ faster, and do not require training from scratch. For a high-quality StyleGAN2 model, our CB-AE and CC with optimization-based interventions can produce almost the same quality + +Table 5. Generation quality and training time comparisons for CelebA-HQ with StyleGAN2. ${}^{ \dagger }$ CBGM results are from their paper. Training time is in V100 GPU-hours. + +
FID (↓)CBGM† [16]CB-AE (Ours)CC (Ours)
Base model9.07.667.66
CB model9.19.52-
CB interv.-9.65-
Opt-interv.-7.677.65
Train time (hrs)50146
+ +Table 6. Human evaluation results for CB-AE and CC trained with CelebA-HQ pretrained StyleGAN2. + +
Conc. Acc. (%)SmilingMale/Female
AutomatedHumanAutomatedHuman
CB-AE (Ours)92.3886.35100.094.06
CC (Ours)89.4780.3596.9696.30
Steerability (%)SmilingFemale
AutomatedHumanAutomatedHuman
CB-AE (Ours)65.9077.2717.0217.73
CB-AE w/ opt-int (Ours)76.5978.7242.8541.50
CC w/ opt-int (Ours)77.3677.3626.9219.87
+ +of images as the base model while having high steerability (61.66% and 67.95% respectively, from Table 3). + +Intuitively, since CB-AE focuses on concept and intervention losses, we obtain better steerability, while CBGM has better FID since they include generative model losses. Hence, there is a tradeoff between image quality and interpretability which can be improved in future work. + +Human evaluation. In Table 6, we compare human agreement rate with automated evaluation of concept accuracy and steerability. We chose easily recognizable concepts, "gender" and "smiling" as representative of low and high steerability respectively. Overall, our automated evaluation is similar to human agreement (with room for improving the classifiers), validating the usefulness of automated evaluation. + +# 4.3. Analysis + +Ablation study. We analyze the contribution of each CB-AE training loss. However, we do not ablate concept loss $\mathcal{L}_c$ or latent reconstruction loss $\mathcal{L}_{r_1}$ since our evaluation metrics would be meaningless if the CB-AE cannot predict the concepts or cannot reconstruct the generator latent $w'$ . Hence, Table 7 ablates image reconstruction loss $\mathcal{L}_{r_2}$ from Eq. (1), intervened concept loss $\mathcal{L}_{i_1}$ from Eq. (3), and cyclic intervened concept loss $\mathcal{L}_{i_2}$ from Eq. (4). The ablations are for the most challenging pseudo-label setting of $M$ (CLIP-zero-shot), since it is most affected by loss ablations. + +In Table 7, we observe that using the image reconstruction loss $\mathcal{L}_{r_2}$ improves the generation quality FID (row #2 + +Table 7. Ablation study on CB-AE training objectives for the most challenging CLIP-zero-shot pseudo-label setting for CelebAHQ pretrained StyleGAN2. $\mathcal{L}_{r_2},\mathcal{L}_{i_1},\mathcal{L}_{i_2}$ indicate image reconstruction loss, intervened concept loss, and intervened cyclic loss respectively from Eq. (1), (3), (4). + +
Row #\( {\mathcal{L}}_{{r}_{2}} \)\( {\mathcal{L}}_{{i}_{1}} \)\( {\mathcal{L}}_{{i}_{2}} \)Trained with \( M = \) CLIP-zero-shot
Conc. Acc. (%)Steerability (%)FID (↓)
1✘✘✘58.876.0111.06
2✓✘✘59.326.519.88
3✓✓✘57.759.9812.42
4✓✘✓58.3313.0313.62
5✓✓✓67.2620.6112.82
+ +Table 8. Sensitivity to pseudo-label source $M$ . For CB-AE trained with CelebA-HQ pretrained StyleGAN2, we compare concept accuracy and steerability with different $M$ : CLIP-zero-shot [35], TIP-few-shot [48], or supervised classifiers. + +
Pseudo-label source MConc. Acc. (%)Steerability (%)
CB-AECB-AE w/ opt-int
CLIP-zs67.2620.6129.23
TIP-fs-12876.0821.5138.73
Supervised-clsf86.0440.2761.66
+ +vs. #1), which is intuitive since this loss directly encourages the images to be closer to the original ones. Next, using only the intervened concept loss or only the intervened cyclic loss improves the steerability while trading-off generation quality (row #3 vs. #2 or row #4 vs. #2). Finally, using both intervened losses significantly improves both concept accuracy and steerability (row #5 vs. #3 and #4), while generation quality remains similar. Overall, both intervention losses are crucial to ensure good concept accuracy and steerability, while image reconstruction loss improves generation quality. Sensitivity to pseudo-label source $M$ . Table 8 compares concept accuracy and steerability when varying pseudo-label source $M$ . Concept accuracy improves as pseudo-label quality improves from CLIP-zero-shot to supervised classifiers. While TIP few-shot does not improve CB-AE interventions, it significantly improves optimization-based interventions and highlights the usefulness of even limited labeled data. Further, our method can use newer CLIP models like SigLIP [47], OpenCLIP [5] to further improve performance. + +# 5. Conclusion + +In this work, we proposed two novel and low-cost methods, concept-bottleneck autoencoder (CB-AE) and concept controller (CC), to efficiently build interpretable generative models from pretrained models. Compared to the prior approach that struggles with efficiency and scalability, our methods achieve $4 - 15 \times$ faster training, require minimal to no concept supervision, and generalize across modern generative model families including GANs and diffusion models with $25\%$ improved steerability on average. + +# Acknowledgements + +# References + +[1] Dylan M Anstine and Alexandr Isayev. Generative models as an emerging paradigm in the chemical sciences. Journal of the American Chemical Society, 145(16):8736-8750, 2023. 1 +[2] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B. Tenenbaum, William T. Freeman, and Antonio Torralba. GAN Dissection: Visualizing and understanding generative adversarial networks. In ICLR, 2019. 2 +[3] David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, and Antonio Torralba. Rewriting a deep generative model. In ECCV, 2020. 2 +[4] Ricky TQ Chen, Xuechen Li, Roger B Grosse, and David K Duvenaud. Isolating sources of disentanglement in variational autoencoders. In NeurIPS, 2018. 2 +[5] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 8 +[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 7 +[7] Zheng Ding, Yifan Xu, Weijian Xu, Gaurav Parmar, Yang Yang, Max Welling, and Zhuowen Tu. Guided variational autoencoder for disentanglement learning. In CVPR, 2020. 2 +[8] Rohit Gandikota, Joanna Materzyńska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models. In ECCV, 2024. 2 +[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 5 +[10] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 5 +[11] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic + +This work is supported in part by National Science Foundation (NSF) awards CNS-1730158, ACI-1540112, ACI-1541349, OAC-1826967, OAC-2112167, CNS-2100237, CNS-2120019, the University of California Office of the President, and the University of California San Diego's California Institute for Telecommunications and Information Technology/Qualcomm Institute. This work used Delta CPU, GPU and Storage through allocation CIS230153 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support program, which is supported by National Science Foundation grants 2138259, 2138286, 2138307, 2137603, and 2138296. The authors are partially supported by National Science Foundation under Grant No. 2107189, 2313105, 2430539, Hellman Fellowship, and Intel Rising Star Faculty Award. The authors would also like to thank anonymous reviewers for valuable feedback to improve the manuscript. +data from generative models ready for image recognition? In ICLR, 2023. 1 +[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5 +[13] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-VAE: Learning basic visual concepts with a constrained variational framework. In ICLR, 2017. 2 +[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 7 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 5 +[16] Aya Abdelsalam Ismail, Julius Adebayo, Hector Corrada Bravo, Stephen Ra, and Kyunghyun Cho. Concept bottleneck generative models. In ICLR, 2024. 1, 2, 3, 4, 5, 7, 8 +[17] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. 5, 7 +[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 5, 7 +[19] Pushkal Katara, Zhou Xian, and Katerina Fragkiadaki. Gen2Sim: Scaling up robot learning in simulation with generative models. In ICRA, 2024. 1 +[20] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In ICML, 2020. 1, 2, 4 +[21] Sonia Laguna, Ricards Marcinkevics, Moritz Vandenhirtz, and Julia E. Vogt. Beyond concept bottleneck models: How to make black boxes intervenable? In NeurIPS, 2024. 2 +[22] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 5 +[23] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In ECCV, 2024. 2 +[24] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 5 +[25] Emanuele Marconato, Andrea Passerini, and Stefano Teso. GlanceNets: Interpretable, leak-proof concept-based models. In NeurIPS, 2022. 2 +[26] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. In NeurIPS, 2022. 2 +[27] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 7 +[28] Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. Fast model editing at scale. In ICLR, 2022. 2 +[29] Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. Memory-based model editing at scale. In ICML, 2022. 2 + +[30] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier GANs. In ICML, 2017. 7 +[31] Tuomas Oikarinen and Tsui-Wei Weng. CLIP-Dissect: Automatic description of neuron representations in deep vision networks. In ICLR, 2023. 2 +[32] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. In ICLR, 2023. 1, 2 +[33] Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. PreciseControl: Enhancing text-to-image diffusion models with fine-grained attribute control. In ECCV, 2024. 2 +[34] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks, 2016. 3 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 5, 8 +[36] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1, 2 +[39] Divyansh Srivastava, Ge Yan, and Tsui-Wei Weng. VLGCBM: Training concept bottleneck models with vision-language guidance. In NeurIPS, 2024. 1, 2 +[40] Chung-En Sun, Tuomas Oikarinen, Berk Ustun, and Tsui-Wei Weng. Concept bottleneck large language models. In ICLR, 2025. 2 +[41] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech-UCSD Birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 5 +[42] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In ICLR, 2020. 5 +[43] An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Yang Wang, Jingbo Shang, and Julian McAuley. Learning concise and descriptive attributes for visual recognition. In ICCV, 2023. 1, 2 +[44] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In CVPR, 2023. 2 +[45] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. In ICLR, 2023. 1, 2 +[46] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano + +Melacci, Adrian Weller, Pietro Lio, and Mateja Jamnik. Concept embedding models: Beyond the accuracy-explainability trade-off. In NeurIPS, 2022. 1 +[47] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 8 +[48] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. In ECCV, 2022. 5, 8 \ No newline at end of file diff --git a/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/images.zip b/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..871313ac819aa6016292cc66437c7b0389f4a4d1 --- /dev/null +++ b/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ac62673aaf38a9a04a85d2338e34e8169c4d0e94f815e53982ca012220ab03f +size 644345 diff --git a/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/layout.json b/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fc23aafca4e940fcd85158394ffc217091ea814d --- /dev/null +++ b/2025/Interpretable Generative Models through Post-hoc Concept Bottlenecks/layout.json @@ -0,0 +1,11980 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 526, + 121 + ], + "type": "text", + "content": "Interpretable Generative Models through Post-hoc Concept Bottlenecks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 143, + 504, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 143, + 504, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 504, + 172 + ], + "type": "text", + "content": "Akshay Kulkarni, Ge Yan,\\* Chung-En Sun\\* Tuomas Oikarinen, and Tsui-Wei Weng \nUniversity of California San Diego" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 174, + 380, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 174, + 380, + 185 + ], + "spans": [ + { + "bbox": [ + 228, + 174, + 380, + 185 + ], + "type": "text", + "content": "{a2kulkarni, lweng}@ucsd.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 297, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 502 + ], + "type": "text", + "content": "Concept bottleneck models (CBM) aim to produce inherently interpretable models that rely on human-understandable concepts for their predictions. However, existing approaches to design interpretable generative models based on CBMs are not yet efficient and scalable, as they require expensive generative model training from scratch as well as real images with labor-intensive concept supervision. To address these challenges, we present two novel and low-cost methods to build interpretable generative models through post-hoc techniques and we name our approaches: concept-bottleneck autoencoder (CB-AE) and concept controller (CC). Our proposed approaches enable efficient and scalable training without the need of real data and require only minimal to no concept supervision. Additionally, our methods generalize across modern generative model families including generative adversarial networks and diffusion models. We demonstrate the superior interpretability and steerability of our methods on numerous standard datasets like CelebA, CelebA-HQ, and CUB with large improvements (average " + }, + { + "bbox": [ + 55, + 239, + 297, + 502 + ], + "type": "inline_equation", + "content": "\\sim 25\\%" + }, + { + "bbox": [ + 55, + 239, + 297, + 502 + ], + "type": "text", + "content": " ) over the prior work, while being " + }, + { + "bbox": [ + 55, + 239, + 297, + 502 + ], + "type": "inline_equation", + "content": "4 - 15\\times" + }, + { + "bbox": [ + 55, + 239, + 297, + 502 + ], + "type": "text", + "content": " faster to train. Finally, a large-scale user study is performed to validate the interpretability and steerability of our methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 135, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 135, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 135, + 529 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 296, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 631 + ], + "type": "text", + "content": "Deep generative models [36-38] have become increasingly powerful and widely used in many high-stakes domains and applications, including realistic data generation [11], simulating hypothetical scenarios or environments [19], and scientific discovery [1]. It is therefore important to ensure that the generation process is interpretable, which will allow us to understand and audit the generation, and further mitigate potential biases and harms (e.g. content moderation)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 632, + 296, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 296, + 681 + ], + "type": "text", + "content": "Unfortunately, most of the advances in deep learning utilize complex, black-box neural network architectures that are difficult to interpret and understand. This leads to user mistrust of model predictions due to the absence of explanations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 214, + 534, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 214, + 534, + 224 + ], + "spans": [ + { + "bbox": [ + 318, + 214, + 534, + 224 + ], + "type": "text", + "content": "A. Prior Work (Concept Bottleneck Generative Models)" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 226, + 546, + 278 + ], + "blocks": [ + { + "bbox": [ + 320, + 226, + 546, + 278 + ], + "lines": [ + { + "bbox": [ + 320, + 226, + 546, + 278 + ], + "spans": [ + { + "bbox": [ + 320, + 226, + 546, + 278 + ], + "type": "image", + "image_path": "be9466677132d035e0060822dda2d2d51546b062bbba8e437b494ff60fed0341.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 334, + 279, + 529, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 279, + 529, + 289 + ], + "spans": [ + { + "bbox": [ + 334, + 279, + 529, + 289 + ], + "type": "text", + "content": "Expensive generative model training from scratch!" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 291, + 521, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 291, + 521, + 301 + ], + "spans": [ + { + "bbox": [ + 320, + 291, + 521, + 301 + ], + "type": "text", + "content": "B. Ours (Post-Hoc Interpretable Generative Models)" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 320, + 302, + 546, + 403 + ], + "blocks": [ + { + "bbox": [ + 320, + 302, + 546, + 403 + ], + "lines": [ + { + "bbox": [ + 320, + 302, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 320, + 302, + 546, + 403 + ], + "type": "image", + "image_path": "1cd5a99936fa27a64d82cc0a2b996e7db5f7bfda7726c968421616b04675b215.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 405, + 551, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 405, + 551, + 415 + ], + "spans": [ + { + "bbox": [ + 318, + 405, + 551, + 415 + ], + "type": "text", + "content": "Efficiently train only CB-AE/CC with frozen generative model" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 417, + 406, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 417, + 406, + 426 + ], + "spans": [ + { + "bbox": [ + 319, + 417, + 406, + 426 + ], + "type": "text", + "content": "C. Qualitative Results" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 318, + 429, + 553, + 495 + ], + "blocks": [ + { + "bbox": [ + 318, + 429, + 553, + 495 + ], + "lines": [ + { + "bbox": [ + 318, + 429, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 318, + 429, + 553, + 495 + ], + "type": "image", + "image_path": "0611e74cc404650e3df77600b540932c17935f3b555392ebecd2864825994393.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 497, + 555, + 553 + ], + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 553 + ], + "type": "text", + "content": "Figure 1. A. Prior work on interpretable generative models requires expensive generative model training from scratch. B. Our CB-AE and CC can be trained efficiently for post-hoc interpretability in a pretrained, frozen generative model " + }, + { + "bbox": [ + 313, + 497, + 555, + 553 + ], + "type": "inline_equation", + "content": "g_{2} \\circ g_{1}" + }, + { + "bbox": [ + 313, + 497, + 555, + 553 + ], + "type": "text", + "content": ". C. Example concept intervention with CB-AE and corresponding concept vectors." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "type": "text", + "content": "To address this, there has been work on developing inherently interpretable deep vision models using concept bottlenecks [20, 32, 39, 43-45]. These approaches train a concept bottleneck layer after the feature extractor (backbone) to embed a set of human-understandable concepts, followed by an interpretable sparse linear layer for the final classification based on the concept prediction. However, current development of CBMs is primarily focused on classification tasks, and only one prior work, CBGM [16], has extended it to image generation, indicating this area is under-explored." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "The key idea of the seminal work CBGM [16] is to represent concepts as learnable embeddings [46] at an inter" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 693, + 132, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 132, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 132, + 703 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 67, + 703, + 280, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 280, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 280, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 703, + 280, + 712 + ], + "type": "text", + "content": "Code: github.com/Trustworthy-ML-Lab/posthoc-generative-cbm" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8162" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 95, + 553, + 152 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Comparison of our CB-AE and CC with prior work CBGM [16] (green indicates desirable properties). CC trades-off inherent interpretability for better steerability, image quality, and faster training. We could not reproduce CBGM [16] to evaluate concept accuracy." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 95, + 553, + 152 + ], + "lines": [ + { + "bbox": [ + 57, + 95, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 553, + 152 + ], + "type": "table", + "html": "
MethodPost-hoc trainingTraining without concept-labeled real imagesInherently interpretable modelConcept Acc. (%)Steerability (%) (↑)FID (↓)Train time (V100-hrs) (↓)
StyleGAN2DDPM
CBGM [16]XX✓-25.609.1050240
CB-AE (Ours)✓✓✓86.5647.34 (+21.74)9.5214 (3.5× faster)29.5 (8.1× faster)
CC (Ours)✓✓X87.6551.14 (+25.54)7.656 (8.3× faster)8.3 (28.9× faster)
", + "image_path": "d6a58747dc7c1bedc0b3698f2dd45bda3485a2c706d00b1a3f25730dbadd0b6c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 164, + 297, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 164, + 297, + 297 + ], + "spans": [ + { + "bbox": [ + 54, + 164, + 297, + 297 + ], + "type": "text", + "content": "mediate location in the generative model, and combine the embeddings to compute the generative model latent. However, the CBGM-based generative model has to be trained from scratch using concept-labeled real images, which could be difficult to scale and computationally intensive (e.g. 240 V100-hours for DDPM- " + }, + { + "bbox": [ + 54, + 164, + 297, + 297 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 164, + 297, + 297 + ], + "type": "text", + "content": " [15]) as shown in Fig. 1A. To address these limitations, in this work, we develop an efficient and scalable post-hoc concept bottleneck to transform any pretrained generative model into an interpretable model. In contrast to CBGM [16], our approach works well by only training a few layers with minimal concept supervision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "text", + "content": "Specifically, we propose a novel concept-bottleneck autoencoder (CB-AE) " + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "text", + "content": " that can be inserted into the intermediate layers of a pretrained generative model " + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "g = g_{2} \\circ g_{1}" + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "text", + "content": " as shown in Fig. 1B (top). The CB-AE input and output are the generator latent " + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "text", + "content": " and its reconstruction " + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "w'" + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "text", + "content": " respectively, and the overall generative model now becomes " + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "g_{2} \\circ f \\circ g_{1}" + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "text", + "content": ". The CB-AE latent space is the concept space " + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "text", + "content": " which is used to reconstruct the generator latent space. In our framework, given a pretrained generative model " + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "g = g_{2} \\circ g_{1}" + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "text", + "content": ", only the encoder " + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "text", + "content": " in the CB-AE " + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "f = D \\circ E" + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "text", + "content": " need to be trained, while " + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "g_{1}, g_{2}" + }, + { + "bbox": [ + 56, + 297, + 296, + 501 + ], + "type": "text", + "content": " are frozen pretrained weights, i.e. CB-AE uses post-hoc training. The benefit of the proposed CB-AE is that it allows us to debug the model easily with concept-level control by modifying the CB-AE concept latent during image generation (Fig. 1C). Further, our training requires only concept pseudo-labels achievable with minimal concept supervision (e.g. zero-shot CLIP classifier [35])." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 503, + 297, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 297, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 297, + 683 + ], + "type": "text", + "content": "We also propose novel optimization-based concept interventions to achieve higher success rate and intervention quality. Based on the CB-AE, we propose an even more efficient post-hoc concept controller (CC) method (Fig. 1B, bottom) with simplified training, that can provide concept predictions and leverage optimization-based concept interventions. Note that while CB-AE is part of the new interpretable generative model, CC is a post-hoc control method that is not part of the generative model. Finally, we evaluate our CB-AE and CC on various generative models, including GANs and diffusion models, for standard datasets including CelebA, CelebA-HQ, and CUB. We show that CB-AE and CC significantly outperform prior work CBGM [16] w.r.t. steerability (or intervention success rate) on CelebA (average " + }, + { + "bbox": [ + 55, + 503, + 297, + 683 + ], + "type": "inline_equation", + "content": "+23\\%" + }, + { + "bbox": [ + 55, + 503, + 297, + 683 + ], + "type": "text", + "content": ") while also having " + }, + { + "bbox": [ + 55, + 503, + 297, + 683 + ], + "type": "inline_equation", + "content": "4 - 15 \\times" + }, + { + "bbox": [ + 55, + 503, + 297, + 683 + ], + "type": "text", + "content": " faster training (Table 1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 685, + 255, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 685, + 255, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 685, + 255, + 696 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 699, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 699, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 699, + 295, + 712 + ], + "type": "text", + "content": "- We are the first to propose a post-hoc concept bottleneck" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 321, + 164, + 555, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 164, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 321, + 164, + 555, + 200 + ], + "type": "text", + "content": "autoencoder (CB-AE) for interpretable generative models. CB-AE can be trained efficiently with a frozen pretrained generative model, without real concept-labeled images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 201, + 555, + 284 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 201, + 553, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 201, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 314, + 201, + 553, + 236 + ], + "type": "text", + "content": "- We also propose a novel and efficient optimization-based concept intervention method with improved steerability (avg. " + }, + { + "bbox": [ + 314, + 201, + 553, + 236 + ], + "type": "inline_equation", + "content": "+19\\%" + }, + { + "bbox": [ + 314, + 201, + 553, + 236 + ], + "type": "text", + "content": ") and higher image quality (avg. " + }, + { + "bbox": [ + 314, + 201, + 553, + 236 + ], + "type": "inline_equation", + "content": "32\\%" + }, + { + "bbox": [ + 314, + 201, + 553, + 236 + ], + "type": "text", + "content": " better)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 236, + 555, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 284 + ], + "type": "text", + "content": "- We validate the effectiveness of our methods for GANs and diffusion models (avg. " + }, + { + "bbox": [ + 313, + 236, + 555, + 284 + ], + "type": "inline_equation", + "content": "+31\\%" + }, + { + "bbox": [ + 313, + 236, + 555, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 236, + 555, + 284 + ], + "type": "inline_equation", + "content": "+28\\%" + }, + { + "bbox": [ + 313, + 236, + 555, + 284 + ], + "type": "text", + "content": " steerability w.r.t. prior state-of-the-art) across varying image resolutions, while being " + }, + { + "bbox": [ + 313, + 236, + 555, + 284 + ], + "type": "inline_equation", + "content": "4 - 15 \\times" + }, + { + "bbox": [ + 313, + 236, + 555, + 284 + ], + "type": "text", + "content": " faster to train on average." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 304, + 400, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 304, + 400, + 316 + ], + "spans": [ + { + "bbox": [ + 314, + 304, + 400, + 316 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 326, + 556, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 326, + 556, + 506 + ], + "spans": [ + { + "bbox": [ + 312, + 326, + 556, + 506 + ], + "type": "text", + "content": "Concept Bottleneck Models. Early work on CBMs [20, 25] relied on concept-labeled images to train a concept bottleneck layer with each neuron as a human-understandable concept, followed by a linear layer based on the concepts for the final classification. Post-hoc CBMs [45] extended this idea to convert a pretrained backbone into a CBM. More recent works like LF-CBM [32], LM4CV [43], LaBo [44], and VLG-CBM [39] use interpretability tools [31], vision-language models like CLIP [35], large language models, or open-set object detection [23] to eliminate the need for expensive concept-labeled data. Independently from our work, [21] proposed a concept-based intervention without a concept bottleneck, similar to our CC, but for classification. All these works are specifically designed for classification, while in this paper, our focus is on the image generation task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 508, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 675 + ], + "type": "text", + "content": "Interpretability for Generative Models. A line of work [4, 7, 13] on learning disentangled concepts in variational autoencoders enables controllable generation, but they train from scratch and are not applicable to other generative models like GANs. Other works focus on identifying and manipulating structural rules or concepts in GANs [2, 3] and large language models [26, 28, 29, 40] by editing the model weights. In contrast, we focus on training inherently interpretable models, and the closest prior work is the recent CBGM [16], which also aims to build interpretable generative models, but requires expensive concept labels and training from scratch, limiting its scalability. In contrast, our proposed CB-AE can be trained efficiently with a frozen pretrained generative model with minimal concept supervision." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Image Editing in Generative Models. Some works focus on conditional generation [37, 38] and image editing in generative models by modifying model weights [8, 33]. In" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8163" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 82, + 261, + 119 + ], + "blocks": [ + { + "bbox": [ + 100, + 70, + 263, + 81 + ], + "lines": [ + { + "bbox": [ + 100, + 70, + 263, + 81 + ], + "spans": [ + { + "bbox": [ + 100, + 70, + 263, + 81 + ], + "type": "text", + "content": "A. Original Black-Box Generative Model" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 82, + 261, + 119 + ], + "lines": [ + { + "bbox": [ + 100, + 82, + 261, + 119 + ], + "spans": [ + { + "bbox": [ + 100, + 82, + 261, + 119 + ], + "type": "image", + "image_path": "e5b1e44a308162f613a686ee388e9e52e66bcaf2ba5a63c549e9c76ce41bc70a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 281, + 81, + 469, + 123 + ], + "blocks": [ + { + "bbox": [ + 291, + 70, + 441, + 80 + ], + "lines": [ + { + "bbox": [ + 291, + 70, + 441, + 80 + ], + "spans": [ + { + "bbox": [ + 291, + 70, + 441, + 80 + ], + "type": "text", + "content": "B. Our Post-hoc Interpretable Model" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 281, + 81, + 469, + 123 + ], + "lines": [ + { + "bbox": [ + 281, + 81, + 469, + 123 + ], + "spans": [ + { + "bbox": [ + 281, + 81, + 469, + 123 + ], + "type": "image", + "image_path": "963c465ee94a7ef306745ec46b45e7927822556a896a8cde8a37d8a692fd73b7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 106, + 140, + 263, + 220 + ], + "blocks": [ + { + "bbox": [ + 100, + 128, + 285, + 140 + ], + "lines": [ + { + "bbox": [ + 100, + 128, + 285, + 140 + ], + "spans": [ + { + "bbox": [ + 100, + 128, + 285, + 140 + ], + "type": "text", + "content": "C. Training Objective 1 (reconstruction losses)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 106, + 140, + 263, + 220 + ], + "lines": [ + { + "bbox": [ + 106, + 140, + 263, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 263, + 220 + ], + "type": "image", + "image_path": "19fde05c03a67d6228634530d9727b08fc39a05cf034a892d49f837375115055.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 286, + 140, + 516, + 212 + ], + "blocks": [ + { + "bbox": [ + 291, + 129, + 487, + 140 + ], + "lines": [ + { + "bbox": [ + 291, + 129, + 487, + 140 + ], + "spans": [ + { + "bbox": [ + 291, + 129, + 487, + 140 + ], + "type": "text", + "content": "D. Training Objective 2 (concept alignment loss)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 286, + 140, + 516, + 212 + ], + "lines": [ + { + "bbox": [ + 286, + 140, + 516, + 212 + ], + "spans": [ + { + "bbox": [ + 286, + 140, + 516, + 212 + ], + "type": "image", + "image_path": "0cb9a9bc5780227cadab60a4bbd3dc68d2bc7632881533d4e3a90eb790106089.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 106, + 229, + 486, + 337 + ], + "blocks": [ + { + "bbox": [ + 100, + 224, + 276, + 235 + ], + "lines": [ + { + "bbox": [ + 100, + 224, + 276, + 235 + ], + "spans": [ + { + "bbox": [ + 100, + 224, + 276, + 235 + ], + "type": "text", + "content": "E. Training Objective 3 (intervention losses)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 106, + 229, + 486, + 337 + ], + "lines": [ + { + "bbox": [ + 106, + 229, + 486, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 486, + 337 + ], + "type": "image", + "image_path": "c874f1f1b3809d302ea572fea3a9df915b618d66f5759ed071500a8dff990fe2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "text", + "content": "Figure 2. Post-hoc CB-AE training for reconstruction, concept alignment, and intervention with a frozen pretrained generator " + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "inline_equation", + "content": "g_{2} \\circ g_{1}" + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_i" + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "text", + "content": " indicate cross-entropy loss and " + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "text", + "content": " indicates mean-squared-error loss. The darker lines indicate gradient flow during training." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 382, + 297, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 297, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 297, + 431 + ], + "type": "text", + "content": "contrast, our work focuses on developing inherently interpretable generative models by introducing additional concept bottleneck layers, where the capability of editing or intervention is naturally a by-product of the interpretability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 449, + 167, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 167, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 167, + 464 + ], + "type": "text", + "content": "3. Proposed Methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 472, + 297, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 297, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 297, + 567 + ], + "type": "text", + "content": "We propose a novel and low-cost concept bottleneck autoencoder (CB-AE) method in Sec. 3.1 to incorporate post-hoc interpretability in pretrained generative models. In Sec. 3.2, we present an optimization-based intervention method for concept-based steerability in the generative model. Finally, based on our insights from CB-AE and optimization-based interventions, we propose an even lower-cost post-hoc control method, concept controller (CC) in Sec. 3.3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "Preliminaries. Consider a generative model " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "g: \\mathcal{Z} \\to \\mathcal{X}" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " that maps from random noise " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "z \\in \\mathcal{Z}" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " to an image " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": ". To make the generative model " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " become inherently interpretable, the goal of CBMs is to insert and train a concept bottleneck (say) " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " at an intermediate location in " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "g = g_2 \\circ g_1" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " can be divided into two parts, " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "g_1" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "g_2" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " (e.g. for a DCGAN [34], " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "g_1" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " is the first 2 layers of " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "g_2" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " is the remaining layers). The input to the concept bottleneck " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " will be the output of " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "g_1" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " (Fig. 1). At the bottleneck of " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": ", we obtain the concept prediction " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{C}" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " is the set of pre-defined concepts. In our setup, the concept vector " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " has two logits for each binary concept " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " (e.g. \"smiling\" would have two logits, for" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "text", + "content": "smiling and not smiling) and " + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "text", + "content": " logits for each categorical concept " + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "text", + "content": " classes (e.g. blonde/black/white/gray hair color would have four logits). For example, suppose we have one binary concept " + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "text", + "content": " and one categorical concept " + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "text", + "content": ", then we define " + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "inline_equation", + "content": "c = [c_{i}^{+}, c_{i}^{-}, c_{j}^{(1)}, c_{j}^{(2)}, \\ldots, c_{j}^{(N)}]^{\\top}" + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "inline_equation", + "content": "c_{i} = [c_{i}^{+}, c_{i}^{-}]^{\\top} \\in \\mathbb{R}^{2}" + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "inline_equation", + "content": "c_{j} = [c_{j}^{(1)}, c_{j}^{(2)}, \\ldots, c_{j}^{(N)}]^{\\top} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 313, + 382, + 555, + 463 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 476, + 534, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 534, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 534, + 489 + ], + "type": "text", + "content": "3.1. Post-hoc Concept Bottleneck Autoencoder" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "content": "We propose a concept bottleneck autoencoder (CB-AE), " + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "inline_equation", + "content": "f = D \\circ E" + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "content": " (see Fig. 2B). The latent space of the CB-AE encoder " + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "content": " is the concept prediction " + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "inline_equation", + "content": "c = E(g_1(z))" + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "content": ". Apart from the predefined concepts, " + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "content": " also contains an unsupervised concept embedding, learned using autoencoder reconstruction and intervention objectives, to encode other concepts absent from the predefined set (similar to CBGM [16]). The decoder " + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "content": " reconstructs the features from " + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "inline_equation", + "content": "g_1" + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "content": " based on the concept prediction " + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "content": ", outputting " + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "inline_equation", + "content": "w' = D(c)" + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "content": ". Considering the original output of " + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "inline_equation", + "content": "g_1" + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "content": " to be " + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "inline_equation", + "content": "w = g_1(z)" + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "content": ", we can generate the original image " + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "inline_equation", + "content": "x = g_2(w)" + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "content": " as well as a reconstructed image " + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "inline_equation", + "content": "x' = g_2(w') = g_2(D \\circ E(w))" + }, + { + "bbox": [ + 313, + 495, + 555, + 640 + ], + "type": "text", + "content": " generated using the CB-AE." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Training. There are 3 goals for the CB-AE. First, the generator's performance should be preserved even if the CB-AE output " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "w'" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " is used instead of " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ". Second, CB-AE should provide interpretability for the generated images " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " through the corresponding concepts " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ". Lastly, CB-AE should allow accurate steering on image generation via concept interven-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8164" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "tions. Based on these 3 goals, we formulate Objective 1-3 for CB-AE training below and illustrated in Fig. 2C-E." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": "Objective 1 (reconstruction losses " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{r_1},\\mathcal{L}_{r_2}" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": "). At each training iteration, we sample a latent " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": " by passing uniform noise " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "g_{1}" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": ". Then, the latent " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": " is passed through " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "g_{2}" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": " to obtain a generated image " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "x = g_{2}(w)" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": " from the original model " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "g = g_{2} \\circ g_{1}" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": ", without using our CB-AE. We also reconstruct " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "w' = D \\circ E(w)" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": " using our CB-AE and the same latent " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": ", and obtain another generated image " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "x' = g_{2}(w')" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": ". Since our first goal is to preserve the generator's performance when using the CB-AE, we apply reconstruction losses (mean-squared error loss " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": ") between the latents " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "w, w'" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": " and between the generated images " + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "inline_equation", + "content": "x, x'" + }, + { + "bbox": [ + 55, + 98, + 296, + 230 + ], + "type": "text", + "content": ", as shown in Fig. 2C:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 238, + 296, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 238, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 111, + 238, + 296, + 258 + ], + "type": "interline_equation", + "content": "\\min _ {E, D} \\left[ \\mathcal {L} _ {r _ {1}} \\left(w, w ^ {\\prime}\\right) + \\mathcal {L} _ {r _ {2}} \\left(x, x ^ {\\prime}\\right) \\right], \\tag {1}", + "image_path": "c438ae7771f81e13be21f1d988b2ad314d54557edee8bdf9cf20f4b9a6f5b865.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 266, + 296, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 290 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 266, + 296, + 290 + ], + "type": "inline_equation", + "content": "w' = D \\circ E(w)" + }, + { + "bbox": [ + 55, + 266, + 296, + 290 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 266, + 296, + 290 + ], + "type": "inline_equation", + "content": "x' = g_2(w') = g_2 \\circ D \\circ E(w)" + }, + { + "bbox": [ + 55, + 266, + 296, + 290 + ], + "type": "text", + "content": " and only the CB-AE parameters (i.e. " + }, + { + "bbox": [ + 55, + 266, + 296, + 290 + ], + "type": "inline_equation", + "content": "E, D" + }, + { + "bbox": [ + 55, + 266, + 296, + 290 + ], + "type": "text", + "content": ") are trainable." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "text", + "content": "Objective 2 (concept alignment loss " + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "text", + "content": "). To ensure interpretability, we first obtain a concept pseudo-label " + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\hat{y} = M(x)" + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "text", + "content": " for a generated image " + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "text", + "content": " from a pseudo-label source " + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "text", + "content": " (Fig. 2D). The source " + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "text", + "content": " can be either an off-the-shelf supervised model or a zero-shot prediction pipeline from (say) CLIP with concept names " + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "text", + "content": " as the text inputs. With this approach, we avoid the requirement of any real images for training as well as the requirement of concept labels, unlike CBGM [16]. Since our second goal is to provide interpretability for the generated images through the concepts " + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "text", + "content": ", we apply a cross-entropy loss " + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "text", + "content": " between the CB-AE encoder output " + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "inline_equation", + "content": "c = E(w)" + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "text", + "content": " and the concept pseudo-label " + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\hat{y} = M(x)" + }, + { + "bbox": [ + 55, + 293, + 296, + 437 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 145, + 445, + 296, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 445, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 145, + 445, + 296, + 462 + ], + "type": "interline_equation", + "content": "\\min _ {E} [ \\mathcal {L} _ {c} (\\hat {y}, c) ]. \\tag {2}", + "image_path": "53ccce73b8ff75711b4590413f6c96a0b61f0afb285dbecab77503ffa7b86034.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "type": "text", + "content": "The losses in Eq. (1), (2) are simultaneously optimized to learn only the CB-AE parameters " + }, + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "type": "inline_equation", + "content": "E, D" + }, + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "text", + "content": "Objective 3 (intervention losses " + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{i_1},\\mathcal{L}_{i_2}" + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "text", + "content": "). Interventions are an important feature of concept-bottleneck models [20], allowing users to control the model output by modifying the concepts " + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "text", + "content": ". However, for our CB-AE decoder " + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "text", + "content": ", reconstruction and concept alignment losses in Objective 1 and 2 do not provide guidance on how the reconstructed latent " + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "inline_equation", + "content": "w^{\\prime}" + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "text", + "content": " should change when concepts " + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "text", + "content": " are manually modified. Hence, for steerability, we design Objective 3 (Fig. 2E) that encourages the CB-AE decoder " + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "text", + "content": " to produce an appropriately changed and realistic latent " + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "inline_equation", + "content": "w^{\\prime}" + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "text", + "content": " when the concepts in " + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "text", + "content": " are modified. We first describe (a) how interventions are performed in our CB-AE, followed by designing (b) an intervened concept alignment loss " + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{i_1}" + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "text", + "content": " and (c) a cyclic intervened concept loss " + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{i_2}" + }, + { + "bbox": [ + 55, + 496, + 296, + 663 + ], + "type": "text", + "content": " to encourage steerability, i.e. intervention success." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "a) Intervening concepts. At each training iteration, we choose a random logit for a random concept to intervene, and modify only the chosen concept based on the chosen logit to get an intervened concept vector " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{\\text{intervened}}" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": " (Fig. 2E). For" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "text", + "content": "example, for a desired binary concept " + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "inline_equation", + "content": "i \\in \\mathcal{C}" + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "text", + "content": ", the new concept vector " + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "inline_equation", + "content": "c_{\\text{intervened}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "text", + "content": " is computed by swapping the two logits, i.e. " + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "inline_equation", + "content": "c_{\\text{intervened}} = [\\ldots, c_i^-, c_i^+, \\ldots]" + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "inline_equation", + "content": "c = [\\ldots, c_i^+, c_i^-, \\ldots]" + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "text", + "content": ". The same can be done for a categorical concept " + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "inline_equation", + "content": "i \\in \\mathcal{C}" + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "text", + "content": " as well by swapping the desired logits (say) " + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "inline_equation", + "content": "c_i^{(k)}" + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "text", + "content": " with the highest logits " + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "inline_equation", + "content": "c_i^{(\\ell)}" + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "inline_equation", + "content": "\\ell = \\arg \\max_j c_i^{(j)}" + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "text", + "content": ". Concretely, " + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "inline_equation", + "content": "c = [\\ldots, c_i^{(1)}, \\ldots, c_i^{(\\ell)}, \\ldots, c_i^{(k)}, \\ldots, c_i^{(N)}, \\ldots]" + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "text", + "content": " is modified to " + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "inline_equation", + "content": "c_{\\text{intervened}} = [\\ldots, c_i^{(1)}, \\ldots, c_i^{(k)}, \\ldots, c_i^{(\\ell)}, \\ldots, c_i^{(N)}, \\ldots]" + }, + { + "bbox": [ + 313, + 72, + 555, + 189 + ], + "type": "text", + "content": ". We can also intervene on multiple concepts simultaneously." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": "b) Designing " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{i_1}" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": ". Using the intervened concepts " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "c_{\\text{intervened}}" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": " we reconstruct an intervened latent " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "w_{\\text{intervened}} = D(c_{\\text{intervened}})" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": " and obtain an intervened generated image " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "x_{\\text{intervened}} = g_2(w_{\\text{intervened}})" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": ". Using the pseudo-label source, we obtain an intervened concept prediction " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "y_{\\text{intervened}} = M(x_{\\text{intervened}})" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": ". Since we already have the concept pseudo-label " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "\\hat{y} = M(x)" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": " for the original image " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": ", we modify " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "\\hat{y}_{\\text{intervened}}" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": " by changing only the earlier chosen concept " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "i \\in \\mathcal{C}" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": " to the chosen value (based on earlier chosen logit) as shown in Fig. 2E. In other words, " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "\\hat{y}_{\\text{intervened}}" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": " is the intervened concept pseudo-label. Now, to align the concepts in the intervened image " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "x_{\\text{intervened}}" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": " with the predicted concepts from " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": ", we use a cross-entropy loss " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{i_1}" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "\\hat{y}_{\\text{intervened}}" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "inline_equation", + "content": "y_{\\text{intervened}}" + }, + { + "bbox": [ + 313, + 190, + 556, + 346 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 372, + 356, + 555, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 356, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 372, + 356, + 555, + 374 + ], + "type": "interline_equation", + "content": "\\underset {E, D} {\\min } \\left[ \\mathcal {L} _ {i _ {1}} \\left(\\hat {y} _ {\\text {i n t e r v e n e d}}, y _ {\\text {i n t e r v e n e d}}\\right) \\right]. \\tag {3}", + "image_path": "97ec21a27262ff4c8922e26b2a3fd7ba670d276d6977844dd22e861cfb604c31.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "content": "c) Designing " + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{i_2}" + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "content": ". For cyclic consistency, we pass the intervented latent " + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "inline_equation", + "content": "w_{\\text{intervened}}" + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "content": " through the CB-AE encoder " + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "content": " to obtain a concept prediction " + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "inline_equation", + "content": "c_{\\text{intervened}}'" + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "content": " (Fig. 2E) and apply a cross-entropy loss " + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{i_2}" + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "content": " w.r.t. " + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\hat{y}_{\\text{intervened}}" + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "content": " to align the encoder's prediction with that of the pseudo-label source " + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 456, + 555, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 456, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 372, + 456, + 555, + 474 + ], + "type": "interline_equation", + "content": "\\min _ {E, D} \\left[ \\mathcal {L} _ {i _ {2}} \\left(\\hat {y} _ {\\text {i n t e r v e n e d}}, c _ {\\text {i n t e r v e n e d}} ^ {\\prime}\\right) \\right]. \\tag {4}", + "image_path": "1405aeac82eaf04c53701a25d59f285fb64b27946c5bac07e80942d1dcc30fa8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{i_1},\\mathcal{L}_{i_2}" + }, + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "text", + "content": " for cross-entropy loss to differentiate the intervention losses from the concept loss in Objective 2, and only CB-AE parameters " + }, + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "inline_equation", + "content": "E,D" + }, + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "text", + "content": " are trainable. Finally, recall that the concept vector " + }, + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "text", + "content": " contains an unsupervised concept embedding. Objective 3 implicitly encourages this embedding to not encode known concepts, since the unsupervised embedding is not modified during the intervention from " + }, + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{intervened}}" + }, + { + "bbox": [ + 313, + 483, + 555, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "Test-Time Intervention. Similar to training-time interventions, we can perform test-time interventions by modifying the value of any chosen concept (by swapping the logits) in the predicted concept vector " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "c_{\\text{intervened}}" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": ". The intervened image " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "x_{\\text{intervened}} = g_2(w_{\\text{intervened}})" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": " can be obtained where " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "w_{\\text{intervened}} = D(c_{\\text{intervened}})" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": ". Note that swapping the logits ensures that the range of values in " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "c_{\\text{intervened}}" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": " are similar to " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": ", and also avoids the requirement of any estimation of how much to change the desired concept's logit. This makes it more accessible to users as they need not worry about the actual values being changed during an intervention." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8165" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 235, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 235, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 235, + 85 + ], + "type": "text", + "content": "3.2. Optimization-based interventions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "type": "text", + "content": "For an alternative intervention method, we draw inspiration from adversarial attacks [10] to perform test-time interventions using gradient-based optimization. Specifically, we use the iterative randomized fast gradient sign method (IRFGSM) [42] on the CB-AE encoder prediction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 153, + 297, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 153, + 297, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 297, + 201 + ], + "type": "text", + "content": "Consider a generated image " + }, + { + "bbox": [ + 55, + 153, + 297, + 201 + ], + "type": "inline_equation", + "content": "x = g_{2}(w)" + }, + { + "bbox": [ + 55, + 153, + 297, + 201 + ], + "type": "text", + "content": " with concept prediction " + }, + { + "bbox": [ + 55, + 153, + 297, + 201 + ], + "type": "inline_equation", + "content": "c = E(w)" + }, + { + "bbox": [ + 55, + 153, + 297, + 201 + ], + "type": "text", + "content": ". To intervene in the generation process to obtain modified concepts " + }, + { + "bbox": [ + 55, + 153, + 297, + 201 + ], + "type": "inline_equation", + "content": "c^{*}" + }, + { + "bbox": [ + 55, + 153, + 297, + 201 + ], + "type": "text", + "content": ", we solve the following objective using gradient ascent," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 209, + 296, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 209, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 91, + 209, + 296, + 229 + ], + "type": "interline_equation", + "content": "w ^ {*} = w + \\underset {\\delta \\in \\Delta} {\\arg \\max } \\left[ - \\mathcal {L} _ {c} \\left(E (w + \\delta), c ^ {*}\\right) \\right] \\tag {5}", + "image_path": "7076cf00a50493ceb8cc4da619d974a6e8c89f6543b259ecdec69dee772c5ac3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "inline_equation", + "content": "\\Delta = \\{\\delta : \\| \\delta \\|_{\\infty} \\leq \\epsilon\\}" + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "text", + "content": "-norm bound on " + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "text", + "content": " with a hyperparameter " + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "inline_equation", + "content": "\\epsilon > 0" + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "text", + "content": ". Intuitively, we optimize a small perturbation " + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "inline_equation", + "content": "w^{*} = w + \\delta" + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "text", + "content": " leads to the desired concepts " + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "inline_equation", + "content": "c^{*}" + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "text", + "content": ". Then, the generated image " + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "inline_equation", + "content": "x^{*} = g_{2}(w^{*})" + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "text", + "content": " is very similar to " + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "text", + "content": " but contains concepts " + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "inline_equation", + "content": "c^{*}" + }, + { + "bbox": [ + 55, + 236, + 296, + 298 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 305, + 296, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 296, + 319 + ], + "type": "text", + "content": "3.3. Post-hoc Concept Controller (CC) for Steering" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 325, + 297, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 297, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 297, + 470 + ], + "type": "text", + "content": "While optimization-based interventions work well empirically, it is interesting to note that the CB-AE decoder " + }, + { + "bbox": [ + 55, + 325, + 297, + 470 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 325, + 297, + 470 + ], + "type": "text", + "content": " is not involved in the process. This leads us to question whether the CB-AE decoder can be removed if a user only plans to perform optimization-based interventions. While such a removal would not result in a CBM, it would be sufficient for the purpose of steering image generation. For this particular use case, it would be even more efficient than training the CB-AE since reconstruction and intervention losses are no longer required. Hence, we propose a post-hoc concept controller (CC), denoted as " + }, + { + "bbox": [ + 55, + 325, + 297, + 470 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 325, + 297, + 470 + ], + "type": "text", + "content": ", that predicts the concepts, i.e. " + }, + { + "bbox": [ + 55, + 325, + 297, + 470 + ], + "type": "inline_equation", + "content": "c = \\Omega(g_1(z))" + }, + { + "bbox": [ + 55, + 325, + 297, + 470 + ], + "type": "text", + "content": " to efficiently steer image generation (Fig. 1B)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "text", + "content": "Training. For CC, the training is simply Objective 2 of the CB-AE training, i.e. the cross-entropy loss " + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "text", + "content": " w.r.t. the concept pseudo-labels " + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "inline_equation", + "content": "\\hat{y} = M(x)" + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "text", + "content": " from the pseudo-label source " + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "text", + "content": ". Formally, the objective is " + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "inline_equation", + "content": "\\min_{\\Omega}[\\mathcal{L}_c(\\hat{y},c)]" + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "inline_equation", + "content": "c = \\Omega (g_1(z))" + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "text", + "content": " are the predicted concepts, and the loss encourages the concept controller " + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "text", + "content": " to align with the pseudo-label source " + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 471, + 297, + 579 + ], + "type": "text", + "content": ". Similar to the CB-AE training, we avoid the requirement of any real images for training as well as the requirement of concept labels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 591, + 137, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 137, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 137, + 604 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 613, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 649 + ], + "type": "text", + "content": "We detail the experimental setup and comprehensively evaluate our proposed methods with respect to the state-of-the-art prior work as well as other baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 658, + 170, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 170, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 170, + 671 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "type": "text", + "content": "Base generative models and datasets. We evaluate our CB-AE and CC methods on diverse generative models including GAN [9], Progressive GAN [17], StyleGAN2 [18]," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "and DDPM [15]. We use models pretrained on standard datasets of varying image resolution (" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ") like CelebA [24], CelebA-HQ [22], and CUB [41]. Following CBGM [16], we evaluate on the small balanced concept regime with the 8 most balanced concepts and the large unbalanced concept regime with all 40 concepts from the dataset. For CUB, we use 10 balanced concepts as per CBGM. Please refer to the Appendix for more details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 170, + 556, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 556, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 556, + 277 + ], + "type": "text", + "content": "CB-AE and CC. We use a 4-layer MLP or 4 convolution (and transposed convolutional) layers for CB-AE encoder or CC (and decoder) depending on the dimensions of latent " + }, + { + "bbox": [ + 313, + 170, + 556, + 277 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 170, + 556, + 277 + ], + "type": "text", + "content": ". We use unsupervised concept embedding " + }, + { + "bbox": [ + 313, + 170, + 556, + 277 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^{40}" + }, + { + "bbox": [ + 313, + 170, + 556, + 277 + ], + "type": "text", + "content": " in CB-AE (ensuring bottleneck is much smaller than the latent). As in CBGM [16], we train CB-AE/CC for 50 epochs with batch size 64. For optimization-based interventions, we use 50-step I-RFGSM [42] with " + }, + { + "bbox": [ + 313, + 170, + 556, + 277 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 313, + 170, + 556, + 277 + ], + "type": "text", + "content": "-norm bound " + }, + { + "bbox": [ + 313, + 170, + 556, + 277 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.1" + }, + { + "bbox": [ + 313, + 170, + 556, + 277 + ], + "type": "text", + "content": ". Please refer to the Appendix for complete implementation details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 280, + 556, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 556, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 556, + 411 + ], + "type": "text", + "content": "Pseudo-label source " + }, + { + "bbox": [ + 313, + 280, + 556, + 411 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 280, + 556, + 411 + ], + "type": "text", + "content": ". We consider three variants for " + }, + { + "bbox": [ + 313, + 280, + 556, + 411 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 280, + 556, + 411 + ], + "type": "text", + "content": " with varying levels of concept supervision. First, we use off-the-shelf supervised (ResNet18-based) concept classifiers. Second, with no concept supervision, we use CLIP zero-shot classifier [35] with only the concept names to obtain concept pseudo-labels. Third, as a compromise between the above two, we use TIP [48] which is a few-shot-labeled version of CLIP zero-shot classifier, utilizing 128 concept-labeled real images. Unless otherwise mentioned, our experiments use the supervised classifiers for a fair comparison with the prior work CBGM [16] that utilizes concept labels." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 413, + 556, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 556, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 556, + 485 + ], + "type": "text", + "content": "Automated evaluation. Following CBGM [16], we train concept classifiers (ViT-L-16-based) on real images and concept labels with high accuracy on a held-out test set. Note that these classifiers are separate and have higher accuracy than those used for pseudo-labels. We evaluate our method using three automated metrics:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 487, + 556, + 631 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 313, + 487, + 556, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 556, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 556, + 522 + ], + "type": "text", + "content": "- Concept Accuracy is computed over 5k generated images as the average agreement between the supervised classifiers and our proposed CB-AE or CC." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 523, + 556, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 556, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 556, + 606 + ], + "type": "text", + "content": "- Steerability [16]: For each target concept, we find 5k latents that do not produce the target concept (i.e. probability of target concept " + }, + { + "bbox": [ + 313, + 523, + 556, + 606 + ], + "type": "inline_equation", + "content": "< 0.5" + }, + { + "bbox": [ + 313, + 523, + 556, + 606 + ], + "type": "text", + "content": " from the supervised concept classifier). For these latents, we perform the concept intervention using either the baselines or our methods, and compute the steerability as the percentage of intervened images that are classified to have the target concept." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 607, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 555, + 631 + ], + "type": "text", + "content": "- Generation Quality is evaluated using the standard Fréchet Inception Distance (FID) [12]." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 633, + 555, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 555, + 669 + ], + "type": "text", + "content": "Intuitively, the concept accuracy, FID, and steerability metrics measure how well the concept, reconstruction, and intervention objectives, respectively, are satisfied." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 671, + 556, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 556, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 556, + 707 + ], + "type": "text", + "content": "Human evaluation. We conduct a large-scale user study on Amazon Mechanical Turk to validate the automated evaluation of concept accuracy and steerability. For both metrics," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8166" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 80, + 154, + 152 + ], + "blocks": [ + { + "bbox": [ + 83, + 72, + 189, + 79 + ], + "lines": [ + { + "bbox": [ + 83, + 72, + 189, + 79 + ], + "spans": [ + { + "bbox": [ + 83, + 72, + 189, + 79 + ], + "type": "text", + "content": "A. CB-AE StyleGAN2 on CelebA-HQ" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 80, + 154, + 152 + ], + "lines": [ + { + "bbox": [ + 82, + 80, + 154, + 152 + ], + "spans": [ + { + "bbox": [ + 82, + 80, + 154, + 152 + ], + "type": "image", + "image_path": "abadf901e2ebe729eedfc5f973493353d24b3ca1215b2c2c6d81598c82b4ccbb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 82, + 152, + 153, + 190 + ], + "blocks": [ + { + "bbox": [ + 82, + 152, + 153, + 190 + ], + "lines": [ + { + "bbox": [ + 82, + 152, + 153, + 190 + ], + "spans": [ + { + "bbox": [ + 82, + 152, + 153, + 190 + ], + "type": "image", + "image_path": "4deff7fb8ac8d7723a360e556f46462fdddc639081980a2abb4868745a1d48b1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 191, + 548, + 203 + ], + "lines": [ + { + "bbox": [ + 61, + 191, + 548, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 191, + 548, + 203 + ], + "type": "text", + "content": "Figure 3. Samples generated using CB-AE with CelebA-HQ and CUB pretrained StyleGAN2 models along with concept probabilities." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 155, + 80, + 227, + 152 + ], + "blocks": [ + { + "bbox": [ + 155, + 80, + 227, + 152 + ], + "lines": [ + { + "bbox": [ + 155, + 80, + 227, + 152 + ], + "spans": [ + { + "bbox": [ + 155, + 80, + 227, + 152 + ], + "type": "image", + "image_path": "210b801d1d3bcc046ca494614c2fa4ee06fb8584ca75704f2d3d8a6c71531a07.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 154, + 153, + 227, + 190 + ], + "blocks": [ + { + "bbox": [ + 154, + 153, + 227, + 190 + ], + "lines": [ + { + "bbox": [ + 154, + 153, + 227, + 190 + ], + "spans": [ + { + "bbox": [ + 154, + 153, + 227, + 190 + ], + "type": "image", + "image_path": "cece36462c3fa0931fc1791258b546203bfbb5a15bfa522089b28b036b089d23.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 228, + 80, + 301, + 152 + ], + "blocks": [ + { + "bbox": [ + 228, + 80, + 301, + 152 + ], + "lines": [ + { + "bbox": [ + 228, + 80, + 301, + 152 + ], + "spans": [ + { + "bbox": [ + 228, + 80, + 301, + 152 + ], + "type": "image", + "image_path": "85fe15eec4bb1b9e6a7f7860fc91f267c5f8cd4a4e192fcac4fd2223775b0549.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 228, + 153, + 300, + 190 + ], + "blocks": [ + { + "bbox": [ + 228, + 153, + 300, + 190 + ], + "lines": [ + { + "bbox": [ + 228, + 153, + 300, + 190 + ], + "spans": [ + { + "bbox": [ + 228, + 153, + 300, + 190 + ], + "type": "image", + "image_path": "5eb7ce5219511705cebb109fe07d5e405aaccf676d5d6537ceee96c0f8a6d5fb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 80, + 380, + 152 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 395, + 79 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 395, + 79 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 395, + 79 + ], + "type": "text", + "content": "B. CB-AE StyleGAN2 on CUB" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 80, + 380, + 152 + ], + "lines": [ + { + "bbox": [ + 308, + 80, + 380, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 80, + 380, + 152 + ], + "type": "image", + "image_path": "a645faf95ecd4839a1bf983f40429b48515c1c6ad94c364c5ffed5b4254a517e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 153, + 378, + 190 + ], + "blocks": [ + { + "bbox": [ + 309, + 153, + 378, + 190 + ], + "lines": [ + { + "bbox": [ + 309, + 153, + 378, + 190 + ], + "spans": [ + { + "bbox": [ + 309, + 153, + 378, + 190 + ], + "type": "image", + "image_path": "178e2a9f475040ddc051204bd29cba62c2d98759ad5a0a3df4cb9c1893a0cc15.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 381, + 80, + 452, + 152 + ], + "blocks": [ + { + "bbox": [ + 381, + 80, + 452, + 152 + ], + "lines": [ + { + "bbox": [ + 381, + 80, + 452, + 152 + ], + "spans": [ + { + "bbox": [ + 381, + 80, + 452, + 152 + ], + "type": "image", + "image_path": "51529635c305d9a15922a0354397272abf612dbadabea02eac069830b0258794.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 381, + 153, + 451, + 190 + ], + "blocks": [ + { + "bbox": [ + 381, + 153, + 451, + 190 + ], + "lines": [ + { + "bbox": [ + 381, + 153, + 451, + 190 + ], + "spans": [ + { + "bbox": [ + 381, + 153, + 451, + 190 + ], + "type": "image", + "image_path": "9859344daf947a58d7b1d4c2eaf26744b86a0a84dc72d9975306d75b421ca50e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 454, + 80, + 526, + 152 + ], + "blocks": [ + { + "bbox": [ + 454, + 80, + 526, + 152 + ], + "lines": [ + { + "bbox": [ + 454, + 80, + 526, + 152 + ], + "spans": [ + { + "bbox": [ + 454, + 80, + 526, + 152 + ], + "type": "image", + "image_path": "44f60efd3025e14550a4ff048b8160cfcb33795309b534467bc6be5b94e20cf5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 454, + 153, + 524, + 190 + ], + "blocks": [ + { + "bbox": [ + 454, + 153, + 524, + 190 + ], + "lines": [ + { + "bbox": [ + 454, + 153, + 524, + 190 + ], + "spans": [ + { + "bbox": [ + 454, + 153, + 524, + 190 + ], + "type": "image", + "image_path": "2327f898dfce6ab323619de89ee5114b8cdc32ae4100d94d6afcaa0f26d6df1d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 57, + 214, + 552, + 327 + ], + "blocks": [ + { + "bbox": [ + 57, + 214, + 552, + 327 + ], + "lines": [ + { + "bbox": [ + 57, + 214, + 552, + 327 + ], + "spans": [ + { + "bbox": [ + 57, + 214, + 552, + 327 + ], + "type": "image", + "image_path": "5047cd29f3a41a5859ba60944a6831cb5cf5e91ebb51764ede675f1b819831bf.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 332, + 474, + 344 + ], + "lines": [ + { + "bbox": [ + 134, + 332, + 474, + 344 + ], + "spans": [ + { + "bbox": [ + 134, + 332, + 474, + 344 + ], + "type": "text", + "content": "Figure 4. Concept intervention examples for CB-AE with CelebA-HQ pretrained StyleGAN2." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 58, + 354, + 554, + 521 + ], + "blocks": [ + { + "bbox": [ + 58, + 354, + 554, + 521 + ], + "lines": [ + { + "bbox": [ + 58, + 354, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 58, + 354, + 554, + 521 + ], + "type": "image", + "image_path": "6461273f56004b1dd120515518f893a6bb18776ca53e9675d8ff70be10faac87.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 525, + 541, + 537 + ], + "lines": [ + { + "bbox": [ + 67, + 525, + 541, + 537 + ], + "spans": [ + { + "bbox": [ + 67, + 525, + 541, + 537 + ], + "type": "text", + "content": "Figure 5. Optimization-based concept intervention (opt-int) examples for CB-AE and CC with CelebA-HQ pretrained StyleGAN2." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 548, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 583 + ], + "type": "text", + "content": "we display 10 images at a time and ask the user to click on images that match a displayed concept " + }, + { + "bbox": [ + 55, + 548, + 295, + 583 + ], + "type": "inline_equation", + "content": "c_{i}^{+}" + }, + { + "bbox": [ + 55, + 548, + 295, + 583 + ], + "type": "text", + "content": " (see Appendix for more details). The images are collected as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 586, + 296, + 706 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 55, + 586, + 295, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 633 + ], + "type": "text", + "content": "- Concept Accuracy: We collect and shuffle generated images and save their CB-AE and CC concept predictions " + }, + { + "bbox": [ + 55, + 586, + 295, + 633 + ], + "type": "inline_equation", + "content": "(c_{i}^{+}" + }, + { + "bbox": [ + 55, + 586, + 295, + 633 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 586, + 295, + 633 + ], + "type": "inline_equation", + "content": "c_{i}^{-})" + }, + { + "bbox": [ + 55, + 586, + 295, + 633 + ], + "type": "text", + "content": ". Based on user responses, we compute human agreement rate w.r.t. CB-AE/CC predictions." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 634, + 296, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 296, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 296, + 706 + ], + "type": "text", + "content": "- Steerability: We collect and shuffle generated images of concept " + }, + { + "bbox": [ + 55, + 634, + 296, + 706 + ], + "type": "inline_equation", + "content": "c_{i}^{-}" + }, + { + "bbox": [ + 55, + 634, + 296, + 706 + ], + "type": "text", + "content": " with intervened images of concept " + }, + { + "bbox": [ + 55, + 634, + 296, + 706 + ], + "type": "inline_equation", + "content": "c_{i}^{+}" + }, + { + "bbox": [ + 55, + 634, + 296, + 706 + ], + "type": "text", + "content": ". Based on user responses, we compute human agreement rate on whether intervened images actually contain concept " + }, + { + "bbox": [ + 55, + 634, + 296, + 706 + ], + "type": "inline_equation", + "content": "c_{i}^{+}" + }, + { + "bbox": [ + 55, + 634, + 296, + 706 + ], + "type": "text", + "content": ". Shuffling the original and intervened images ensures that users are not biased towards clicking all images." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 548, + 555, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 555, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 555, + 596 + ], + "type": "text", + "content": "We evaluate using approximately 100 images per concept and per method for two concepts from CelebA-HQ: “smiling” and “male/female”. Each set of 10 images is evaluated by 3 different users. Refer to the Appendix for more details." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 601, + 388, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 388, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 388, + 612 + ], + "type": "text", + "content": "4.2. Evaluation" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Qualitative evaluation. In Fig. 3, we present CB-AE-StyleGAN2 generated images for CelebA-HQ and CUB datasets with corresponding concept predictions. We also visualize CB-AE interventions in Fig. 4 and optimization-based interventions in Fig. 5. Interestingly, with the same latent " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": ", both CB-AE and CC lead to very similar optimization-based interventions, which is reasonable since the same pseudo-label source " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " was used in both cases." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8167" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 82, + 553, + 141 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 527, + 81 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 527, + 81 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 527, + 81 + ], + "type": "text", + "content": "Table 2. Concept accuracy evaluation on 5k samples with 8 concepts for CelebA, CelebA-HQ, and 10 concepts for CUB." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 82, + 553, + 141 + ], + "lines": [ + { + "bbox": [ + 57, + 82, + 553, + 141 + ], + "spans": [ + { + "bbox": [ + 57, + 82, + 553, + 141 + ], + "type": "table", + "html": "
Conc. Acc. (%)CelebA (64×64)CelebA-HQ (256×256)CUB (64×64)CUB (256×256)
GANPGANDDPMDDPMStyGAN2PGANDDPMGANStyGAN2
Ours (CB-AE)86.5687.8784.9889.7986.0482.6872.3574.4181.33
Ours (CC)87.6590.0085.1389.8283.5783.9472.8775.6081.11
", + "image_path": "d920ffc137578e9385b07864276fe297121338a29b4cf4467b5fc253edbe34f6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 186, + 553, + 281 + ], + "blocks": [ + { + "bbox": [ + 55, + 151, + 555, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 151, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 555, + 185 + ], + "type": "text", + "content": "Table 3. Extended steerability evaluation on 5k samples with 8 concepts for CelebA, CelebA-HQ, and 10 concepts for CUB. " + }, + { + "bbox": [ + 55, + 151, + 555, + 185 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 151, + 555, + 185 + ], + "type": "text", + "content": " CBGM numbers are from their paper (1k samples) since their results are not reproducible using their released code. For CBGM training time, we used the time taken by their code for GAN and base model training times for other models (whose code was not available) with 1 V100 GPU." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 186, + 553, + 281 + ], + "lines": [ + { + "bbox": [ + 56, + 186, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 553, + 281 + ], + "type": "table", + "html": "
Steerability (%)CelebA (64×64)CelebA-HQ (256×256)CUB (64×64)CUB (256×256)
GANPGANDDPMDDPMStyGAN2PGANDDPMGANStyGAN2
CBGM [16]†25.60-13.80---14.8021.30-
Ours (CB-AE)47.3440.3123.7223.5040.2729.3125.6420.8910.52
Ours (CB-AE+opt-int)61.1441.7338.0950.4961.6632.1036.9446.0365.11
Ours (CC+opt-int)51.1458.9441.4556.7067.9547.2944.4748.9144.72
Train time reductionCB-AE3.7×5.4×4×8.1×3.5×2×3.7×3.3×3.1×
w.r.t. CBGMCC30.9×20.3×10×28.9×8.3×6.7×8.5×21×7.1×
", + "image_path": "b086db45ec9e4ac6927430d49e890982be636438bfe4e31a487bf8da9ab1e212.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 300, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 294, + 335 + ], + "type": "text", + "content": "Overall, we find that the optimization-based interventions are relatively higher quality and more orthogonal (i.e. less changes to other concepts) than the CB-AE interventions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 337, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 337, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 337, + 295, + 409 + ], + "type": "text", + "content": "Concept Accuracy. In Table 2, we report the concept accuracies for our CB-AE and CC. We could not compare with CBGM since they do not evaluate concept accuracy, and since we could not reproduce their results. Overall, we find that both CB-AE and CC achieve good concept accuracies across various datasets, models and image resolutions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 411, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 295, + 518 + ], + "type": "text", + "content": "In most scenarios, CC outperforms CB-AE since concept alignment is the sole objective optimized in CC, while CB-AE has additional objectives. However, we observe that CB-AE outperforms CC only for StyleGAN2. This is because the multiple objectives in CB-AE training may be easier to balance when dealing with clean latents (GANs) than with noisy latents (DDPM), leading to CB-AE outperforming CC for StyleGAN2. This is supported by average loss for CB-AE being lower for StyleGAN2 (0.68) than DDPM (0.92)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "text", + "content": "Steerability. We compare the concept steerability of our methods on 1k samples with GAN intervention methods like conditional GAN (CGAN) [27], auxiliary classifier GAN (ACGAN) [30], and CBGM (CB-GAN) [16], and with diffusion model intervention methods like classifier-guided (CG) [6] DDPM, classifier-free (CF) [14] DDPM, and concept bottleneck DDPM (CB-DDPM) [16] in Table 4. We observe significant gains over CBGM [16] for both GAN (average " + }, + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "inline_equation", + "content": "+14.2\\%" + }, + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "text", + "content": " for CB-AE, " + }, + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "inline_equation", + "content": "+31.1\\%" + }, + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "text", + "content": " for CB-AE with opt-int, and " + }, + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "inline_equation", + "content": "+26.6\\%" + }, + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "text", + "content": " for CC) and for DDPM (average " + }, + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "inline_equation", + "content": "+10.4\\%" + }, + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "text", + "content": " for CB-AE, " + }, + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "inline_equation", + "content": "+23.3\\%" + }, + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "text", + "content": " for CB-AE with opt-int, and " + }, + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "inline_equation", + "content": "+29.1\\%" + }, + { + "bbox": [ + 54, + 521, + 295, + 652 + ], + "type": "text", + "content": " for CC)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 654, + 296, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 654, + 296, + 701 + ], + "spans": [ + { + "bbox": [ + 54, + 654, + 296, + 701 + ], + "type": "text", + "content": "Table 3 presents extended steerability evaluation on " + }, + { + "bbox": [ + 54, + 654, + 296, + 701 + ], + "type": "inline_equation", + "content": "5\\mathrm{k}" + }, + { + "bbox": [ + 54, + 654, + 296, + 701 + ], + "type": "text", + "content": " samples with other types of GANs like Progressive GAN [17] and StyleGAN2 [18] for higher resolution " + }, + { + "bbox": [ + 54, + 654, + 296, + 701 + ], + "type": "inline_equation", + "content": "(256\\times 256)" + }, + { + "bbox": [ + 54, + 654, + 296, + 701 + ], + "type": "text", + "content": " datasets. We could not obtain CBGM's results for these set" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 332, + 553, + 502 + ], + "blocks": [ + { + "bbox": [ + 313, + 297, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 331 + ], + "type": "text", + "content": "Table 4. Steerability comparisons with CBGM [16] and other baseline intervention methods computed on 1k samples (each experiment repeated three times for mean and standard deviation)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 332, + 553, + 502 + ], + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 502 + ], + "type": "table", + "html": "
Concept RegimeSmall balanced conceptsLarge unbalanced concepts
DatasetCUB (10 conc.)CelebA (8 conc.)CelebA (40 conc.)
Baseline Intervention Methods
CGAN [27]5.4 ± 0.48.7 ± 1.32.9 ± 0.0
ACGAN [30]18.5 ± 0.49.2 ± 0.71.2 ± 0.1
CB-GAN [16]21.3 ± 0.325.6 ± 0.523.1 ± 0.2
Our Methods
CB-AE-GAN19.9 ± 0.947.1 ± 0.545.5 ± 0.4
CB-AE-GAN+opt-int45.1 ± 1.159.9 ± 1.858.3 ± 1.3
CC-GAN+opt-int49.3 ± 0.650.8 ± 1.249.7 ± 0.9
Baseline Intervention Methods
CF-DDPM [14]2.7 ± 1.97.2 ± 3.85.1 ± 2.4
CG-DDPM [6]2.1 ± 1.46.8 ± 1.15.4 ± 2.6
CB-DDPM [16]14.8 ± 6.213.8 ± 2.712.6 ± 1.7
Our Methods
CB-AE-DDPM25.8 ± 1.123.1 ± 0.923.6 ± 1.0
CB-AE-DDPM+opt-int37.3 ± 1.537.5 ± 1.336.2 ± 1.1
CC-DDPM+opt-int45.4 ± 2.241.8 ± 1.841.3 ± 1.5
", + "image_path": "e4b647360b0f867b294244acff379f4d34291c34b0134cd1beba107ad68d4535.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 532, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 532, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 532, + 555, + 628 + ], + "type": "text", + "content": "tings using their code. Overall, we find optimization-based interventions (opt-int) outperform the CB-AE intervention method, and CC generally outperforms CB-AE. Intuitively, opt-int involves instance-specific and iterative optimization, while CB-AE is applied in the same way to all samples. Although CB-AE is trained with intervention losses, its steerability tends to be worse than CC since CB-AE is more challenging to train, with multiple objectives to satisfy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "type": "text", + "content": "Generation quality. We compare the generation quality of CBGM [16] with CB-AE and CC in Table 5. For CB-AE, we observe a relatively higher drop in image quality than CBGM, but our methods are trained " + }, + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "3.5 - 8 \\times" + }, + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "type": "text", + "content": " faster, and do not require training from scratch. For a high-quality StyleGAN2 model, our CB-AE and CC with optimization-based interventions can produce almost the same quality" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8168" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 104, + 294, + 185 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "text", + "content": "Table 5. Generation quality and training time comparisons for CelebA-HQ with StyleGAN2. " + }, + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "text", + "content": " CBGM results are from their paper. Training time is in V100 GPU-hours." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 104, + 294, + 185 + ], + "lines": [ + { + "bbox": [ + 58, + 104, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 58, + 104, + 294, + 185 + ], + "type": "table", + "html": "
FID (↓)CBGM† [16]CB-AE (Ours)CC (Ours)
Base model9.07.667.66
CB model9.19.52-
CB interv.-9.65-
Opt-interv.-7.677.65
Train time (hrs)50146
", + "image_path": "8163897296c9cfc2dca0b38fd3b70116ed808922e7a70ba4b3ae3ffc41b8d768.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 220, + 294, + 345 + ], + "blocks": [ + { + "bbox": [ + 55, + 197, + 295, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 197, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 295, + 218 + ], + "type": "text", + "content": "Table 6. Human evaluation results for CB-AE and CC trained with CelebA-HQ pretrained StyleGAN2." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 220, + 294, + 345 + ], + "lines": [ + { + "bbox": [ + 58, + 220, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 58, + 220, + 294, + 345 + ], + "type": "table", + "html": "
Conc. Acc. (%)SmilingMale/Female
AutomatedHumanAutomatedHuman
CB-AE (Ours)92.3886.35100.094.06
CC (Ours)89.4780.3596.9696.30
Steerability (%)SmilingFemale
AutomatedHumanAutomatedHuman
CB-AE (Ours)65.9077.2717.0217.73
CB-AE w/ opt-int (Ours)76.5978.7242.8541.50
CC w/ opt-int (Ours)77.3677.3626.9219.87
", + "image_path": "866c07e4286264e0a822ec3d78a7cfd08f184e2933ba31a8597281c3c60d2a62.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 366, + 295, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 391 + ], + "type": "text", + "content": "of images as the base model while having high steerability (61.66% and 67.95% respectively, from Table 3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 392, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 453 + ], + "type": "text", + "content": "Intuitively, since CB-AE focuses on concept and intervention losses, we obtain better steerability, while CBGM has better FID since they include generative model losses. Hence, there is a tradeoff between image quality and interpretability which can be improved in future work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "text", + "content": "Human evaluation. In Table 6, we compare human agreement rate with automated evaluation of concept accuracy and steerability. We chose easily recognizable concepts, \"gender\" and \"smiling\" as representative of low and high steerability respectively. Overall, our automated evaluation is similar to human agreement (with room for improving the classifiers), validating the usefulness of automated evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 551, + 118, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 118, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 118, + 563 + ], + "type": "text", + "content": "4.3. Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "text", + "content": "Ablation study. We analyze the contribution of each CB-AE training loss. However, we do not ablate concept loss " + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "text", + "content": " or latent reconstruction loss " + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{r_1}" + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "text", + "content": " since our evaluation metrics would be meaningless if the CB-AE cannot predict the concepts or cannot reconstruct the generator latent " + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "inline_equation", + "content": "w'" + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "text", + "content": ". Hence, Table 7 ablates image reconstruction loss " + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{r_2}" + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "text", + "content": " from Eq. (1), intervened concept loss " + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{i_1}" + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "text", + "content": " from Eq. (3), and cyclic intervened concept loss " + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{i_2}" + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "text", + "content": " from Eq. (4). The ablations are for the most challenging pseudo-label setting of " + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "text", + "content": " (CLIP-zero-shot), since it is most affected by loss ablations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "In Table 7, we observe that using the image reconstruction loss " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{r_2}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " improves the generation quality FID (row #2" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 126, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 125 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 125 + ], + "type": "text", + "content": "Table 7. Ablation study on CB-AE training objectives for the most challenging CLIP-zero-shot pseudo-label setting for CelebAHQ pretrained StyleGAN2. " + }, + { + "bbox": [ + 313, + 70, + 555, + 125 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{r_2},\\mathcal{L}_{i_1},\\mathcal{L}_{i_2}" + }, + { + "bbox": [ + 313, + 70, + 555, + 125 + ], + "type": "text", + "content": " indicate image reconstruction loss, intervened concept loss, and intervened cyclic loss respectively from Eq. (1), (3), (4)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 126, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 317, + 126, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 126, + 553, + 205 + ], + "type": "table", + "html": "
Row #\\( {\\mathcal{L}}_{{r}_{2}} \\)\\( {\\mathcal{L}}_{{i}_{1}} \\)\\( {\\mathcal{L}}_{{i}_{2}} \\)Trained with \\( M = \\) CLIP-zero-shot
Conc. Acc. (%)Steerability (%)FID (↓)
1✘✘✘58.876.0111.06
2✓✘✘59.326.519.88
3✓✓✘57.759.9812.42
4✓✘✓58.3313.0313.62
5✓✓✓67.2620.6112.82
", + "image_path": "0dc62dabdd10483b62be9668110ba1149a217e5dfcd72708be93da4f607af037.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 261, + 552, + 324 + ], + "blocks": [ + { + "bbox": [ + 313, + 216, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 260 + ], + "type": "text", + "content": "Table 8. Sensitivity to pseudo-label source " + }, + { + "bbox": [ + 313, + 216, + 555, + 260 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 216, + 555, + 260 + ], + "type": "text", + "content": ". For CB-AE trained with CelebA-HQ pretrained StyleGAN2, we compare concept accuracy and steerability with different " + }, + { + "bbox": [ + 313, + 216, + 555, + 260 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 216, + 555, + 260 + ], + "type": "text", + "content": ": CLIP-zero-shot [35], TIP-few-shot [48], or supervised classifiers." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 261, + 552, + 324 + ], + "lines": [ + { + "bbox": [ + 317, + 261, + 552, + 324 + ], + "spans": [ + { + "bbox": [ + 317, + 261, + 552, + 324 + ], + "type": "table", + "html": "
Pseudo-label source MConc. Acc. (%)Steerability (%)
CB-AECB-AE w/ opt-int
CLIP-zs67.2620.6129.23
TIP-fs-12876.0821.5138.73
Supervised-clsf86.0440.2761.66
", + "image_path": "069b5323da4ebe18e2d670976c4b7b1d882cfc38167cca21b478c3846adc3df3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 343, + 556, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 556, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 556, + 575 + ], + "type": "text", + "content": "vs. #1), which is intuitive since this loss directly encourages the images to be closer to the original ones. Next, using only the intervened concept loss or only the intervened cyclic loss improves the steerability while trading-off generation quality (row #3 vs. #2 or row #4 vs. #2). Finally, using both intervened losses significantly improves both concept accuracy and steerability (row #5 vs. #3 and #4), while generation quality remains similar. Overall, both intervention losses are crucial to ensure good concept accuracy and steerability, while image reconstruction loss improves generation quality. Sensitivity to pseudo-label source " + }, + { + "bbox": [ + 313, + 343, + 556, + 575 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 343, + 556, + 575 + ], + "type": "text", + "content": ". Table 8 compares concept accuracy and steerability when varying pseudo-label source " + }, + { + "bbox": [ + 313, + 343, + 556, + 575 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 343, + 556, + 575 + ], + "type": "text", + "content": ". Concept accuracy improves as pseudo-label quality improves from CLIP-zero-shot to supervised classifiers. While TIP few-shot does not improve CB-AE interventions, it significantly improves optimization-based interventions and highlights the usefulness of even limited labeled data. Further, our method can use newer CLIP models like SigLIP [47], OpenCLIP [5] to further improve performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 580, + 388, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 388, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 388, + 593 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 600, + 556, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 556, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 556, + 708 + ], + "type": "text", + "content": "In this work, we proposed two novel and low-cost methods, concept-bottleneck autoencoder (CB-AE) and concept controller (CC), to efficiently build interpretable generative models from pretrained models. Compared to the prior approach that struggles with efficiency and scalability, our methods achieve " + }, + { + "bbox": [ + 313, + 600, + 556, + 708 + ], + "type": "inline_equation", + "content": "4 - 15 \\times" + }, + { + "bbox": [ + 313, + 600, + 556, + 708 + ], + "type": "text", + "content": " faster training, require minimal to no concept supervision, and generalize across modern generative model families including GANs and diffusion models with " + }, + { + "bbox": [ + 313, + 600, + 556, + 708 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 600, + 556, + 708 + ], + "type": "text", + "content": " improved steerability on average." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8169" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 308, + 115, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 115, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 115, + 320 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 328, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 328, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 328, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 61, + 328, + 296, + 360 + ], + "type": "text", + "content": "[1] Dylan M Anstine and Alexandr Isayev. Generative models as an emerging paradigm in the chemical sciences. Journal of the American Chemical Society, 145(16):8736-8750, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 361, + 296, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 296, + 404 + ], + "type": "text", + "content": "[2] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B. Tenenbaum, William T. Freeman, and Antonio Torralba. GAN Dissection: Visualizing and understanding generative adversarial networks. In ICLR, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 406, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 296, + 437 + ], + "type": "text", + "content": "[3] David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, and Antonio Torralba. Rewriting a deep generative model. In ECCV, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 438, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 438, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 438, + 296, + 470 + ], + "type": "text", + "content": "[4] Ricky TQ Chen, Xuechen Li, Roger B Grosse, and David K Duvenaud. Isolating sources of disentanglement in variational autoencoders. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 472, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 296, + 525 + ], + "type": "text", + "content": "[5] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 526, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 296, + 548 + ], + "type": "text", + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 548, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 296, + 581 + ], + "type": "text", + "content": "[7] Zheng Ding, Yifan Xu, Weijian Xu, Gaurav Parmar, Yang Yang, Max Welling, and Zhuowen Tu. Guided variational autoencoder for disentanglement learning. In CVPR, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 582, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 582, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 582, + 296, + 613 + ], + "type": "text", + "content": "[8] Rohit Gandikota, Joanna Materzyńska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models. In ECCV, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 615, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 615, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 615, + 296, + 656 + ], + "type": "text", + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 689 + ], + "type": "text", + "content": "[10] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[11] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 55, + 93, + 297, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 93, + 297, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 93, + 297, + 297 + ], + "type": "text", + "content": "This work is supported in part by National Science Foundation (NSF) awards CNS-1730158, ACI-1540112, ACI-1541349, OAC-1826967, OAC-2112167, CNS-2100237, CNS-2120019, the University of California Office of the President, and the University of California San Diego's California Institute for Telecommunications and Information Technology/Qualcomm Institute. This work used Delta CPU, GPU and Storage through allocation CIS230153 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support program, which is supported by National Science Foundation grants 2138259, 2138286, 2138307, 2137603, and 2138296. The authors are partially supported by National Science Foundation under Grant No. 2107189, 2313105, 2430539, Hellman Fellowship, and Intel Rising Star Faculty Award. The authors would also like to thank anonymous reviewers for valuable feedback to improve the manuscript." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "data from generative models ready for image recognition? In ICLR, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "text", + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 194 + ], + "type": "text", + "content": "[13] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-VAE: Learning basic visual concepts with a constrained variational framework. In ICLR, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 217 + ], + "type": "text", + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 240 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "text", + "content": "[16] Aya Abdelsalam Ismail, Julius Adebayo, Hector Corrada Bravo, Stephen Ra, and Kyunghyun Cho. Concept bottleneck generative models. In ICLR, 2024. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 276, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 308 + ], + "type": "text", + "content": "[17] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 555, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 555, + 374 + ], + "type": "text", + "content": "[19] Pushkal Katara, Zhou Xian, and Katerina Fragkiadaki. Gen2Sim: Scaling up robot learning in simulation with generative models. In ICRA, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 376, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 555, + 409 + ], + "type": "text", + "content": "[20] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In ICML, 2020. 1, 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 555, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 555, + 442 + ], + "type": "text", + "content": "[21] Sonia Laguna, Ricards Marcinkevics, Moritz Vandenhirtz, and Julia E. Vogt. Beyond concept bottleneck models: How to make black boxes intervenable? In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 445, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 555, + 476 + ], + "type": "text", + "content": "[22] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 478, + 555, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 555, + 531 + ], + "type": "text", + "content": "[23] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In ECCV, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 534, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 555, + 555 + ], + "type": "text", + "content": "[24] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 557, + 555, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 555, + 588 + ], + "type": "text", + "content": "[25] Emanuele Marconato, Andrea Passerini, and Stefano Teso. GlanceNets: Interpretable, leak-proof concept-based models. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 590, + 555, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 555, + 622 + ], + "type": "text", + "content": "[26] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 624, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 645 + ], + "type": "text", + "content": "[27] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 647, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 677 + ], + "type": "text", + "content": "[28] Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. Fast model editing at scale. In ICLR, 2022. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[29] Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. Memory-based model editing at scale. In ICML, 2022. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8170" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 105 + ], + "type": "text", + "content": "[30] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier GANs. In ICML, 2017. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 138 + ], + "type": "text", + "content": "[31] Tuomas Oikarinen and Tsui-Wei Weng. CLIP-Dissect: Automatic description of neuron representations in deep vision networks. In ICLR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 171 + ], + "type": "text", + "content": "[32] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 172, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 215 + ], + "type": "text", + "content": "[33] Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. PreciseControl: Enhancing text-to-image diffusion models with fine-grained attribute control. In ECCV, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 248 + ], + "type": "text", + "content": "[34] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 250, + 296, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 296, + 314 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 348 + ], + "type": "text", + "content": "[36] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 296, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 381 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 296, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 296, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 296, + 436 + ], + "type": "text", + "content": "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 437, + 296, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 469 + ], + "type": "text", + "content": "[39] Divyansh Srivastava, Ge Yan, and Tsui-Wei Weng. VLGCBM: Training concept bottleneck models with vision-language guidance. In NeurIPS, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 471, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 502 + ], + "type": "text", + "content": "[40] Chung-En Sun, Tuomas Oikarinen, Berk Ustun, and Tsui-Wei Weng. Concept bottleneck large language models. In ICLR, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 504, + 296, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 296, + 546 + ], + "type": "text", + "content": "[41] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech-UCSD Birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 296, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 569 + ], + "type": "text", + "content": "[42] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In ICLR, 2020. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 571, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 296, + 613 + ], + "type": "text", + "content": "[43] An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Yang Wang, Jingbo Shang, and Julian McAuley. Learning concise and descriptive attributes for visual recognition. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 614, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 657 + ], + "type": "text", + "content": "[44] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In CVPR, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 296, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 680 + ], + "type": "text", + "content": "[45] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "text", + "content": "[46] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 184 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "type": "text", + "content": "Melacci, Adrian Weller, Pietro Lio, and Mateja Jamnik. Concept embedding models: Beyond the accuracy-explainability trade-off. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 107, + 555, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 138 + ], + "type": "text", + "content": "[47] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "type": "text", + "content": "[48] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. In ECCV, 2022. 5, 8" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8171" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/d624a8da-e662-4173-965d-8a4cb70d4b42_content_list.json b/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/d624a8da-e662-4173-965d-8a4cb70d4b42_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1cd4665aca7b53085025f5e4d6fab9b372d862c0 --- /dev/null +++ b/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/d624a8da-e662-4173-965d-8a4cb70d4b42_content_list.json @@ -0,0 +1,1420 @@ +[ + { + "type": "text", + "text": "Interpretable Image Classification via Non-parametric Part Prototype Learning", + "text_level": 1, + "bbox": [ + 96, + 130, + 898, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhijie Zhu1* Lei Fan1* Maurice Pagnucco1 Yang Song1 \n1University of New South Wales", + "bbox": [ + 254, + 179, + 740, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhijie.zhu@,lei.fanl@morri@cse,yang.song1@}.unsw.edu.au", + "bbox": [ + 243, + 218, + 746, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Classifying images with an interpretable decision-making process is a long-standing problem in computer vision. In recent years, Prototypical Part Networks has gained traction as an approach for self-explainable neural networks, due to their ability to mimic human visual reasoning by providing explanations based on prototypical object parts. However, the quality of the explanations generated by these methods leaves room for improvement, as the prototypes usually focus on repetitive and redundant concepts. Leveraging recent advances in prototype learning, we present a framework for part-based interpretable image classification that learns a set of semantically distinctive object parts for each class, and provides diverse and comprehensive explanations. The core of our method is to learn the part-prototypes in a non-parametric fashion, through clustering deep features extracted from foundation vision models that encode robust semantic information. To quantitatively evaluate the quality of explanations provided by ProtoPNETs, we introduce Distinctiveness Score and Comprehensiveness Score. Through evaluation on CUB-200-2011, Stanford Cars and Stanford Dogs datasets, we show that our framework compares favourably against existing ProtoPNETs while achieving better interpretability. Code is available at: https://github.com/zijizhu/proton-param.", + "bbox": [ + 86, + 301, + 483, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 710, + 220, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, self-explainable networks have emerged as a promising approach for improving the transparency of deep neural networks by providing inherent interpretability. Unlike traditional explanation methods that operate post-hoc [33, 49], recent approaches aim to mimic human reasoning by justifying predictions through an interpretable decision-making process. Examples include Prototypical-Part Networks (ProtoPNETs) [7, 11], part discovery methods [21, 37], and Concept Bottleneck Models [12, 23, 39, 48],", + "bbox": [ + 89, + 737, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b76c8f6738c9c2e49fc63bea9457e7d288ab9f8a5330b1afa68a4d58c4ece5fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 273, + 712, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d86ba279cc30b7e1c3220cb1f89e69dfea207784207b11cdca0a41c7d6ca6a86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 273, + 910, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5aa1bf7f2442eee595fdaf9328da5409d8cf1294f40150b68047dde94a6edb7f.jpg", + "image_caption": [ + "Figure 1. a. Existing methods [19, 40] encode various assumptions as regularizations to guide prototype learning, but often fail to diversify the concepts learned by each prototype. Here multiple part-prototypes attend to the same image region, thereby limiting their interpretability. In contrast b. Our method partitions the feature space into semantically distinct clusters, and updates each prototype with the empirical mean of its respective cluster, enabling each prototype to learn semantically different concepts." + ], + "image_footnote": [], + "bbox": [ + 521, + 383, + 710, + 547 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/093e69ec04edc2deb830aef71f7ccbbeecf627885c5fab593b65c9debac02573.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 383, + 910, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which offer explanations derived from object concepts. Among these, ProtoPNETs have garnered attention as they provide intuitive visual explanations rooted in past examples. Specifically, they define a number of prototypes for each object category, each representing a distinguishable object part, and make the prediction based on the similarity of an input image to prototypes, thereby emulating the process of human visual reasoning.", + "bbox": [ + 511, + 700, + 906, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, existing ProtoPNETs train prototypes as learnable parameters, and guide their learning process with the core assumption that a region within each input image should match one of the prototypes from the corresponding category. However, we observe that multiple prototypes", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 887, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9762", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from the same class often correspond to the same object part, which hinders the quality of explanations provided by the network in two ways: First, explanations become noncomprehensive, covering only a subset of object parts. Second, this results in redundant and repetitive explanations that fail to adequately illustrate why the object is classified in a particular way. An example is shown in Figure 1.a, out of the four most activated prototypes in a ProtoPNet, three of them focus on overlapping regions near the bird's wing, ignoring other parts such as the bird's legs and tail. Ideally, within an interpretability framework, the model is expected to recognize and differentiate a comprehensive set of distinctive parts for each object category, thus enhancing the clarity and transparency of its decision-making process.", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To approach this problem, we take inspiration from recent advances in prototype-based classification [47, 50], and encode distinct object parts using non-parametric prototypes [42, 51], as shown in Figure 1.b. Non-parametric prototypes were originally proposed to learn a more robust feature space by capturing the rich variance within each class through multiple prototypes [51]. We extend this idea to part-based interpretability to address both of the aforementioned limitations. Specifically, in comparison to previous methods that guide the learning process of learnable prototypes with regularizations [40], non-parametric prototypes allow finer control over the representation of each prototype within the same class by applying constraints on feature clustering. As a consequence, each prototype is able to capture unique characteristics of an object that differentiate it from others, resulting in semantically distinct prototypes, each latent feature patch of the foreground object is able to contribute to one of the prototypes, allowing the overall representation for each class to embody a comprehensive set of associated concepts. These properties greatly improve the interpretability of the explanations.", + "bbox": [ + 91, + 306, + 483, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, in order to obtain robust feature representations that encode object part semantics, We explore advanced foundation models as backbones to extract features, specifically utilizing self-supervised Vision Transformers pre-trained with self-distillation (DINO-ViTs) [6, 27], given that these backbone networks capture rich semantic information about object parts and can be used out-of-the-box [2, 28]. We also propose an efficient fine-tuning strategy for self-supervised ViT backbones to adapt the feature space to the target domain while grounding the embedding of object parts to its assigned prototypes. Hence, our framework consists of two training stages where the initial stage focuses on prototype discovery, and the subsequent stage is directed towards feature refinement. Only the image-level class labels are required in both stages. We conduct experiments to compare our approach with existing ProtoPNETs on the CUB-200-2011 [38], Stanford Cars [24], and Stanford Dogs datasets [10]. Moreover, to evaluate the diversity", + "bbox": [ + 89, + 628, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of concepts captured by ProtoPNETs, we introduce two new metrics: Distinctiveness Score, which quantifies the extent of overlap among object parts learned by prototypes within the same class, and Comprehensiveness Score, which measures whether the prototypes comprehensively attend to different parts of the foreground object. To summarize the key contributions of this work:", + "bbox": [ + 511, + 90, + 903, + 195 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify a key limitation of existing ProtoPNETs, specifically generating repetitive and redundant visual explanations. To this end, we leverage non-parametric prototypes to build a self-explainable network.", + "- Our method comprehensively explores non-parametric prototypes and foundation vision backbones, discovering diverse and semantic distinctive concepts.", + "- We introduce two evaluation metrics, namely Distinctiveness Score and Comprehensiveness Score to quantify the diversity and comprehensiveness of object part concepts. We demonstrate that our method yields comparable results against state-of-the-art ProtoPNETs models on three benchmarking datasets while recognizing holistic and distinctive concepts from each category." + ], + "bbox": [ + 511, + 196, + 903, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 421, + 653, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Part-based Interpretable Image Classification. Drawing inspiration from the human ability to distinguish objects by recognizing their semantic parts, the Prototypical Part Network (ProtoPNet) [7] emulates this reasoning process by comparing image patches with learned prototype vectors that represent object parts specific to each class. Several variants have been proposed to enhance both its interpretability and classification accuracy [5, 11, 36, 46]. Notably, TesNet [40] reformulates prototype learning from the perspective of constructing embedding space on Grassmann Manifold, and also improves the overall performance by applying various priors on prototypes, such as orthonormality. Building upon the architecture of TesNet, Huang et. al [19] introduced modules to improve consistency and stability of the visual explanations, as well as corresponding metrics for quantitative evaluation.", + "bbox": [ + 511, + 446, + 906, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, there has been a parallel line of research that focuses on detecting object parts of fine-grained object categories [21, 37]. While these methods involve learning object part representations, they usually detect a number of shared object parts for the entire target domain and emphasize evaluating the quality of detected parts rather than leveraging them for case-based reasoning.", + "bbox": [ + 511, + 688, + 908, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main goal of our work aims to alleviate aforementioned drawbacks of ProtoPNETs, by learning part-prototypes in a non-parametric fashion. In addition, Huang et. al. [20] integrate the reasoning process of ProtoPNETs with the series of Concept Bottleneck Models [12, 23] to improve the trustworthiness of explanations [20]. We follow this pioneering work and evaluate our framework when", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9763", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adapted to the concept learning setting.", + "bbox": [ + 89, + 90, + 351, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Foundation Vision Models. Recently, foundation vision backbones [6, 30] have been successfully applied to various challenging downstream tasks [3, 14, 17, 22], yielding remarkable results. In contrast to traditional supervised training, these backbones typically employ self-supervision or weak supervision schemes, such as masked image modeling [45], contrastive learning [8] or self distillation, reducing the reliance on labels and enhancing the quality of representation. Previous studies have demonstrated that each of these pre-training methods constructs an embedding space with different desirable properties [28, 44]. In particular, the attention map and patch tokens of self-supervised Vision Transformers trained with self-distillation have been shown to encode rich semantic information, such as scene layouts, which are beneficial for dense recognition tasks like and object discovery [34], anomaly detection [15, 16], and image generation [35, 41]. In addition, other works have leveraged their strong object part representation for fine-grained categorization [2, 13, 32]. Building upon these developments, our framework harnesses these desired properties to improve the quality of explanation provided by part-based interpretable image classifiers.", + "bbox": [ + 89, + 108, + 485, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prototype Learning. Prototype-based classification is to make predictions based on representative examples from past data, similar to how humans approach problem solving and learning, as suggested by psychology studies [1]. Recently, it has gained significant interest in deep learning as an alternative to traditional classification heads. For example, memory bank based methods [18, 43] retain a large number of past samples to learn high-quality representations. Approaches such as ProtoSeg [51], deep nearest centroids [42] and various segmentation models [25, 29] replace conventional parametric softmax layers of segmentors with multiple non-parametric prototypes, which not only learn class distributions but also capture the intra-class variance of each category, improving the representativeness and generalizability of the models. Our non-parametric part-prototype is inspired by these approaches.", + "bbox": [ + 89, + 443, + 483, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 705, + 181, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Part-based interpretability methods such as ProtoPNETs aim to explain the predictions by generating final logits based on how much input image regions look like prototypical object parts from the training set. The core of these methods is to learn a robust set of prototypes that can provide high quality explanations. As shown in Figure 2.a, our method achieves this by deriving them as non-parametric prototypes with a learning procedure disentangled from the classification process. This results in two training stages, where the first stage is dedicated to train the prototypes and the second stage focuses on refining the feature space.", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 513, + 90, + 624, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of our model is to perform interpretable image classification by learning prototypes $\\mathbf{P} \\in \\mathbb{R}^{C \\times K \\times D}$ , where $C$ is the number of classes and $K$ is the number of prototypes belonging to each class. Each $D$ dimensional vector represents one of the $K$ representative parts belonging to class $c \\in \\{c_1, c_2, \\ldots\\}$ . During inference, given a batch of image $\\mathbf{X}_{1:B} = \\{\\mathbf{X}_b \\in R^{H \\times W \\times 3}\\}$ , where $B$ is the batch size, $H$ and $W$ denote the size of the input image, a ViT backbone $\\mathcal{F}$ first encodes each image into a sequence of patch tokens $\\mathbf{f} = \\mathcal{F}(\\mathbf{X}) \\in \\mathbb{R}^{B \\times h \\times w \\times D}$ , reshaped into a feature map of shape $h \\times w$ . Our model computes the similarity between the patch tokens $\\mathbf{f}$ and part prototypes $\\mathbf{P}$ , generating a set of similarity maps $\\mathbf{S}_{1:B} = \\{\\mathrm{Sim}(\\mathbf{f}_b, \\mathbf{P}) \\in \\mathbb{R}^{h \\times w \\times C \\times K}\\}_{b=1}^B$ , where $\\mathrm{Sim}(\\cdot, \\cdot)$ is a similarity measure, such as cosine similarity. As shown in Figure 2.a, each similarity map highlights a region of the input image similar to the learned prototypical part. Finally, the maximum value of each similarity map quantifies the occurrence of its corresponding part prototype with respect to the input, and the logit of class $c$ is computed as the weighted average of all occurrence scores of its prototypes.", + "bbox": [ + 511, + 112, + 906, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the first stage of training, an efficient weakly-supervised clustering algorithm computes non-parametric part prototypes for each class based on extracted features from frozen backbones. Due to the well-structured latent space, these prototypes can be directly used for interpretable image classification. Then, we fix the learned part prototypes and fine-tune the backbone network by using the prototypes as anchors. This adapts the feature space to the target domain and significantly enhances the classification performance. We only use image-level class labels for supervision in both stages.", + "bbox": [ + 511, + 431, + 908, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Non-parametric Prototype Learning", + "text_level": 1, + "bbox": [ + 511, + 604, + 831, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2.b, the main idea of this stage is to cluster the feature representation of each class involved in a batch of samples, and create a mapping between each patch token and part prototype. The empirical mean of the patch tokens assigned to the same part-prototype is directly used to update the corresponding prototypes. Hence, each latent feature patch is able to contribute to one of the part-prototypes, ensuring the overall representation of each class to cover a wide range of concepts. The only supervision involved in this stage is associating each patch token with the class label of image it belongs to, and the assignment maps are computed in an unsupervised manner.", + "bbox": [ + 511, + 628, + 908, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we treat a batch of latent patch tokens as a bag of features, reshaping the patch tokens $\\mathbf{f}$ and similarity maps $\\mathbf{S}$ to $\\mathbf{f} \\in \\mathbb{R}^{Bhw \\times D}$ and $\\mathbf{S} \\in \\mathbb{R}^{Bhw \\times C \\times K}$ . Next, we employ a foreground extraction module to extract patches belonging to the foreground of each sample, excluding the background patches to improve representation", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9764", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/74a10a81ae87f5ab90e1df4dbfaa60a17f31d347788a83fdedcb2fa18e80fae6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 88, + 874, + 256 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5d97c7d10f4b541fd3eb5d80c55ffa054ab308d23308b63bc01667fb3e73b57c.jpg", + "image_caption": [ + "Figure 2. The architecture of our learning framework. a. Similar to prototype-based classification, we define $K$ number of part-prototypes for each category and compare them to the latent features, which yields similarity maps. Each similarity map is pooled to generate a prototype presence vector for each class. The final logits are computed as a weighted average of the presence vector. b. Non-parametric Prototype Update: We employ feature clustering to discover class-wise non-parametric part prototype from latent feature patches. Each part prototype can be regarded as the empirical mean of its corresponding part from all training examples. c. The feature space of the backbone is fine-tuned efficiently with part-prototype fixed, and by inserting new ViT blocks." + ], + "image_footnote": [], + "bbox": [ + 127, + 262, + 883, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "quality. Given $N$ foreground patches of class $c$ , denoted as $\\mathbf{f}^c \\subset \\mathbf{f}$ , $\\mathbf{f}^c \\in \\mathbb{R}^{N \\times D}$ , we obtain a similarity matrix $\\mathbf{S}^c \\in \\mathbb{R}^{N \\times K}$ between $\\mathbf{f}^c$ and $\\mathbf{P}^c$ by indexing $\\mathbf{S}$ . We then compute a one-hot cluster assignment $\\mathbf{A} \\in \\{0,1\\}^{N \\times K}$ that assigns each patch of class $c$ to one of the part-prototypes.", + "bbox": [ + 89, + 527, + 482, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, we aim to optimize $\\mathbf{A}^c$ so that the similarity between each patch and its assigned prototype is maximized, that is:", + "bbox": [ + 89, + 604, + 482, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\mathbf {A} ^ {c} \\in \\mathcal {A} ^ {c}} \\operatorname {T r} \\left(\\left(\\mathbf {A} ^ {c}\\right) ^ {\\top} \\mathbf {S} ^ {c}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 665, + 482, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{Tr}(\\cdot)$ is the trace of a matrix. To prevent trivial solutions for $\\mathbf{A}^c$ and ensure each cluster represents a distinct object part, we enforce unique assignment and equipartition constraints [51]:", + "bbox": [ + 89, + 698, + 482, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {A} ^ {c}\\right) ^ {\\top} \\mathbf {1} _ {N} = \\frac {1}{K} \\mathbf {1} _ {K}, \\mathbf {A} ^ {c} \\mathbf {1} _ {K} = \\frac {1}{N} \\mathbf {1} _ {N}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 772, + 482, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This ensures each patch token maps to a single part-prototype of class $c$ , and each prototype has an approximately equal number of patch tokens assigned. This formulation can be regarded as an Optimal Transport (OT) problem, where the approximate solution of $\\mathbf{A}$ can be estimated by solving a smoothed version of equation (1):", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} ^ {c, *} = \\max _ {\\mathbf {A} ^ {c} \\in \\mathcal {A} ^ {c}} \\operatorname {T r} \\left(\\left(\\mathbf {A} ^ {c}\\right) ^ {\\top} \\mathbf {S}\\right) + \\kappa h \\left(\\mathbf {A} ^ {c}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 541, + 903, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\kappa h(\\mathbf{A})$ is the entropic regularization term controlling smoothness. We use $\\kappa = 0.05$ following previous works [9]. The solution can be efficiently computed using the Sinkhorn-Knopp Algorithm:", + "bbox": [ + 511, + 569, + 906, + 630 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} ^ {c, *} = \\operatorname {d i a g} (\\mathbf {u}) \\exp \\left(\\frac {\\mathbf {S} ^ {c}}{\\kappa}\\right) \\operatorname {d i a g} (\\mathbf {v}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 640, + 903, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Momentum Update of Prototypes. Given the estimated one-hot patch-prototype assignment $\\mathbf{A}^{c,*}$ , each part-prototype $\\mathbf{P}_k^c$ is updated as the mean of its assigned patches:", + "bbox": [ + 511, + 676, + 906, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {k} ^ {c} \\leftarrow \\mathbf {P} _ {k} ^ {c} + (1 - \\beta) \\overline {{\\mathbf {f}}} ^ {c, k}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 733, + 903, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta \\in [0,1]$ is a coefficient that controls the update speed, and $\\overline{\\mathbf{f}}$ is the mean feature. This learning scheme effectively captures the mean embedding of object parts across all training examples, and can be visualized by projecting to the closest patches from the training set [7].", + "bbox": [ + 511, + 762, + 906, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adapt to Interpretable Image Classification", + "text_level": 1, + "bbox": [ + 511, + 847, + 883, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2.a, the learned part-prototypes can be directly used for classification. The activation of each pro", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9765", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "totype $\\mathbf{P}_k^c$ with respect to the input image is calculated as:", + "bbox": [ + 89, + 90, + 473, + 108 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {k} ^ {c} = \\max _ {\\widetilde {\\mathbf {f}} \\in \\mathbf {f}} \\quad \\operatorname {S i m} \\left(\\widetilde {\\mathbf {f}}, \\mathbf {P} _ {k} ^ {c}\\right) \\in \\mathbb {R}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 118, + 482, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As not all prototypes are equally important for classifying the input image, following previous work [19], we generate the final logit for each class $c$ by computing a weighted average of $g^{c}$ :", + "bbox": [ + 89, + 150, + 482, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\log \\mathrm {i t} ^ {c} = \\sum_ {k = 1} ^ {K} w _ {k} ^ {c} \\cdot g _ {k} ^ {c}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 220, + 482, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w$ is a learned modulation vector that controls the contribution of each prototype activation towards the logit of its corresponding class. In contrast to using a fully-connected layer, computing class logits as a weighted average ensures that prototypes contribute exclusively to their respective class, thereby enhancing the interpretability.", + "bbox": [ + 89, + 267, + 483, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prototype-Anchored Feature Space Tuning. After discovering class-wise part-prototypes from the pre-trained backbone, we fine-tune the backbone to ground the features to the learned prototypes and adapt to the target domain. To this end, we employ Block Expansion [4], as shown in Figure 2.c. Given a ViT backbone $\\mathcal{F}$ with a sequence of transformer blocks $(\\mathcal{F}_1,\\mathcal{F}_2,\\mathcal{F}_3,\\ldots)$ , Block Expansion divides the blocks into $m$ groups and inserts a new identity block $\\mathcal{F}_{id}$ at the end of each divided group. Each inserted identity block is initialized with weights of the preceding block, with the weights of attention projection layer and the second MLP layer set to zero, such that $\\mathcal{F}_{id}(x) = x$ , i.e., the output of the backbone remains unchanged without further training. During fine-tuning, we freeze all original weights and part-prototypes, and only fine-tune the added identity blocks and modulation vectors. This approach efficiently produces features that separate classes well without disturbing the well-structured latent space of the backbone.", + "bbox": [ + 89, + 357, + 483, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Patch-Prototype Distance Optimization. In the fine-tuning stage, we use a standard cross-entropy loss $\\mathcal{L}_{ce}$ to separate features of each class. In addition, we ground each image patch to its assigned part-prototype to improve feature locality. To this end, we introduce an in-class patch-prototype contrastive loss, defined as:", + "bbox": [ + 89, + 628, + 482, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P P C}} = \\sum_ {c \\in \\mathcal {C}} \\frac {\\exp \\left(\\left(\\mathbf {f} _ {i} ^ {c}\\right) ^ {\\top} \\mathbf {P} _ {k _ {i}} ^ {c}\\right)}{\\exp \\left(\\left(\\mathbf {f} _ {i} ^ {c}\\right) ^ {\\top} \\mathbf {P} _ {k _ {i}} ^ {c}\\right) + \\sum_ {\\mathbf {P} _ {k _ {j}} ^ {c} \\in \\mathcal {P} ^ {c -}} \\exp \\left(\\left(\\mathbf {f} _ {i} ^ {c}\\right) ^ {\\top} \\mathbf {P} _ {k _ {j}} ^ {c}\\right)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 739, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{P}^{c-} = \\{\\mathbf{P}_k^c\\}_{k=1}^K / \\mathbf{P}_{k_i}^c$ . This loss \"pulls\" each patch towards its assigned part-prototype in the latent space, and separate it from other part-prototype within the same class. Hence, we jointly train the expanded ViT block and the modulation weights with the following objective:", + "bbox": [ + 89, + 794, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {c e}} + \\lambda_ {\\mathrm {P P C}} \\mathcal {L} _ {\\mathrm {P P C}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 885, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Distinctiveness Score", + "text_level": 1, + "bbox": [ + 513, + 90, + 710, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ideally, each part-prototype belonging to the same class should learn a distinct concept from objects; otherwise, the explanation would be confusing and redundant. To quantify this property, we introduce a metric, Distinctiveness Score that measures the amount of overlap between image regions highlighted by the part prototypes associated with the ground truth class of the input image.", + "bbox": [ + 511, + 114, + 903, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, given an input image $x$ of class $c_{x}$ and the corresponding activation maps $S^{c_x} \\in \\mathbb{R}^{h \\times w \\times K}$ generated by prototypes $\\mathbf{P}^{c_x}$ , we first resize the activation maps to $\\hat{S}^{c_x} \\in \\mathbb{R}^{K \\times H \\times W}$ to match the size of input $x$ . Next, we extract the object region $R^{c_x, k}(x)$ highlighted by each activation map $S^{c_x, k}$ as a fix-sized bounding box with predetermined height $H_b$ and width $W_b$ around its max activation. The amount of overlap $O(x)$ between the similarity maps $S^{c_x}$ generated by ground truth class prototypes is then quantified by the Intersection over Union (IoU) between each pair of highlighted regions $R^{c_x, k}(x)$ , as follows:", + "bbox": [ + 511, + 222, + 906, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO (x) = \\frac {1}{K (K - 1) / 2} \\sum_ {k = 1} ^ {K - 1} \\sum_ {l = k + 1} ^ {K} \\operatorname {I o U} \\left(R ^ {c _ {x}, k} (x), R ^ {c _ {x}, l} (x)\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 402, + 903, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Distinctiveness Score is computed based on the average overlap across all samples in the test set $\\mathcal{X}$ , and a higher value is better. It is defined as:", + "bbox": [ + 511, + 459, + 905, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n1 - \\frac {1}{| \\mathcal {X} |} \\sum_ {x \\in \\mathcal {X}} O (x). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 517, + 903, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Comprehensiveness Score", + "text_level": 1, + "bbox": [ + 513, + 568, + 748, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A good Distinctiveness Score by itself does not guarantee good interpretability, as a high score could be achieved by networks with prototypes correspond to different image regions that lack meaningful concepts, e.g., background patches that contribute less to object recognition. Therefore, we further introduce Comprehensiveness Score to evaluate how thoroughly the object foreground is attended by the prototypes. Given an input image $x$ and the corresponding activation maps $S^{c_x}$ generated by prototypes $\\mathbf{P}^{c_x}$ , we perform bin-max normalization on each $S^{c_x,k}$ and obtain a mask $\\hat{M}_i^{c_x,k}(x)$ of the most activated region of $\\mathbf{P}^{c_x,k}$ using a threshold $\\tau$ :", + "bbox": [ + 511, + 593, + 906, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {S} _ {i} ^ {c _ {x}, k} = \\frac {S _ {i} ^ {c _ {x} , k} - S _ {\\min } ^ {c _ {x} , k}}{S _ {\\max } ^ {c _ {x} , k} - S _ {\\min } ^ {c _ {x} , k}}, \\quad \\hat {M} _ {i} ^ {c _ {x}, k} (x) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\tilde {S} _ {i} ^ {c _ {x}, k} \\geq \\tau \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 785, + 903, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then compute the IoU between the union of all masks $\\hat{M}^{c_x,k}(x)$ and the ground truth object foreground mask $M_{x}$ . The Comprehensiveness Score is calculated as the average IoU across all test samples $\\mathcal{X}$ , and a higher value indicates", + "bbox": [ + 511, + 840, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9766", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/13ad81be73f9fbc7f7ae41213ecbe6ab8362d5a123b34746d653435e883cec86.jpg", + "image_caption": [ + "Figure 3. a. The Distinctiveness score for one sample is calculated based on the amount of overlap between areas attended by prototypes. b. The Comprehensiveness score is computed by comparing the ground truth foreground mask $M_{x}$ with the union of thresholded image region attended by prototypes." + ], + "image_footnote": [], + "bbox": [ + 99, + 98, + 464, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "better comprehensiveness. Specifically, it is defined as:", + "bbox": [ + 89, + 424, + 455, + 441 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{| \\mathcal {X} |} \\sum_ {x \\in \\mathcal {X}} \\operatorname {I o U} \\left(M _ {x}, \\bigcup_ {k = 1} ^ {K} \\hat {M} ^ {c _ {x}, k} (x)\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 452, + 482, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 518, + 223, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 544, + 299, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. To evaluate the classification accuracy and interpretability of our proposed framework, we conduct experiments on CUB-200-2011 [38], Stanford Cars [24] and Stanford Dogs [10] dataset, following previous works on ProtoPNETs. The CUB-200-2011 dataset consists of 11,788 images spanning 200 bird species. Stanford Cars includes 16,185 images from 196 car models, and Stanford Dogs includes 20,580 images from 120 breeds. The results on Stanford Cars and Stanford Dogs are in supplementary. We also extend our evaluation by performing concept learning [20, 23] experiments on CUB-200-2011 to evaluate whether the learned prototypes can be adapted to produce trustworthy concept-based explanations by following the evaluation protocol introduced in [20].", + "bbox": [ + 89, + 568, + 482, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and Evaluation Metrics. We establish our main benchmark by comparing against existing ProtoPNETs, namely ProtoPNet [7], Deformable ProtoPNet [11], ProtoPool [31], TesNet [40] and EvalProtoPNet [19]. We select three self-supervised Vision Transformer backbone as the backbone: the latest checkpoints of DINOv2 ViT-B and ViT-S trained on LVD-142M [27] and the original DINO trained on ImageNet [6]. All the networks are trained with", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "image-level class labels only, and all previous methods were reproduced with official implementations. For evaluation metrics, we report top-1 accuracy for classification. We also quantitatively measure the interpretability of the learned prototypes with the introduced Distinctiveness and Comprehensiveness scores, as well as Consistency and Stability proposed in [19]. Specifically, Consistency evaluates whether each prototype correspond to the same object part across different images, while Stability measures whether each prototype correspond to the same object part when the image is perturbed with noise. We evaluate all methods with $K = 3$ and $K = 5$ . For concept learning, we compare with existing Concept Bottleneck Models [12, 20, 23], and evaluate the classification performance as well as the trustworthiness [20] of the generated concept explanations.", + "bbox": [ + 511, + 90, + 903, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. In the first stage of training, we set $\\beta = 0.99$ , and $\\kappa = 0.05$ to perform clustering. When fine-tuning the backbone, we set $\\lambda_{\\mathrm{PPC}} = 0.8$ and the number of added ViT block $m = 1$ . We trained our framework with 1 epoch for the first stage, and 5 more epochs for the second stage with Adam optimizer, a learning rate of 0.0001 for the backbone and $1e - 6$ for modulation vector $w$ . When training previous ProtoPNETs, we set $m = 3$ as it produces the best results across all previous architectures.", + "bbox": [ + 511, + 318, + 906, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 467, + 707, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We reported the benchmark result of our method against previous ProtoPLists in Table 1. Here the Distinctiveness is calculated with $1/4$ of image size. Our method outperformed previous methods in terms of classification accuracy by $(0.16\\%, 2.39\\%)$ when trained with both the revised DINOv2 and DINO ViT-B backbones, and the accuracy evaluated on ViT-S is comparable to previous works. In addition, while our method yields similar Consistency to [19], it achieves higher Stability of than existing works in most settings. Nevertheless, the most notable improvement of our framework is the Distinctiveness, which surpasses previous methods by $(4.83\\%, 3.49\\%, 8.93\\%)$ when setting 5 prototypes for each class. This indicates that our framework is more capable of utilizing the strong part representation encoded in self-supervised Vision Transformers to discover a diverse set of prototypes. Finally, the prototypes learned by our framework can be directly adapted to provide concept-based explanations, as shown by the results in Table 2. Our method yields comparable trustworthiness score with higher classification performance than [20], and Concept Bottleneck Models constructed naively on top of ViT backbones failed to produce trustworthy explanations despite their higher accuracy.", + "bbox": [ + 511, + 491, + 906, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Distinctiveness and Comprehensiveness of Prototypes. We provide a more detailed breakdown on the diversity of prototypes learned by various ProtoPNETs, by selecting a few architectures and evaluate the Distinctiveness with", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9767", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3e71633d47d4e87d6ddfa5a7bfb68ba4f912f7f0bfa06164d2a44accbc247710.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DINOv2 ViT-B/14DINOv2 ViT-S/14DINO ViT-B/16
Con.Sta.Dis.Class.Con.Sta.Dis.Class.Con.Sta.Dis.Class.
ProtoPNet (K=3)0.7768.405.7788.474.8474.688.8384.903.0063.873.1475.94
ProtoPNet (K=5)1.3063.8411.9988.164.1270.1016.7385.523.0246.346.7978.91
Deformable ProtoPNet (K=3)0.9580.005.9086.999.6761.181.6583.091.5570.849.1571.49
Deformable ProtoPNet (K=5)2.8380.4310.3488.1411.7379.874.7685.540.9573.161.1781.74
ProtoPool (K=3)29.8352.8387.5988.7215.5047.8991.1786.1926.6777.7483.8972.04
ProtoPool (K=5)29.2757.4989.6289.9012.152.4492.2687.4137.4052.7387.6678.18
TesNet (K=3)36.3366.4541.2490.6658.0066.8927.8088.2335.5065.4058.8581.24
TesNet (K=5)58.6074.6081.9390.3567.8070.2781.1688.9551.3070.8181.9482.29
EvalProtoPNet (K=3)67.8371.3925.6789.7345.1763.5117.0187.6637.8369.3835.1084.21
EvalProtoPNet (K=5)67.8074.4072.1290.0654.6070.2777.4188.1951.0966.8672.7484.54
Ours (K=3)61.1778.9886.4490.3761.5075.5789.1389.0745.6769.9095.6285.12
Ours (K=5)66.4082.9794.4590.8266.0181.3195.7588.1647.2075.5796.5986.93
", + "bbox": [ + 94, + 87, + 903, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/577499fcd18d2655b515c07959ac7c8c01cce324a29207e1a3e96b7eb20a9ad8.jpg", + "image_caption": [ + "Figure 4. Distinctiveness $(\\%)$ and Comprehensiveness $(\\%)$ across various ProtoPLists, evaluated on different box sizes and thresholds. All methods besides [19] are trained with DINOv2 ViT-B." + ], + "image_footnote": [], + "bbox": [ + 91, + 340, + 480, + 579 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f67d7aa74ff1b4873d7361c3728392f071da1e7ea06e7c0d31cfeab60e307be2.jpg", + "table_caption": [ + "Table 1. Consistency score (%), Stability Score (%), Distinctiveness Score (%), and classification accuracy (% of various ProtoPNet on CUB-200-2011. K denotes the number of prototypes defined for each class. We train previous methods with their official implementations." + ], + "table_footnote": [], + "table_body": "
ViT-BViT-S
Loc.Acc.Loc.Acc.
Vanilla CBM15.6390.0612.9388.43
CBM-h11.9589.2411.4186.52
CEM13.0488.7410.6186.41
Huang et. al.45.9587.4347.3384.79
Ours (K=3)41.1887.0439.8884.55
Ours (K=5)47.0488.1844.9785.47
", + "bbox": [ + 135, + 651, + 439, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Concept trustworthiness (\\%) and classification accuracy (\\%) of a CBM constructed based on our framework, in comparison with previous methods on CUB-200-2011 dataset. All methods are trained with DINOV2 ViTs as the backbone. Note that CBM-h is vanilla CBM with a residual fitting.", + "bbox": [ + 89, + 772, + 482, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "varying bounding box sizes. The results are shown in Figure 4. Notably, the prototypes learned by the vanilla ProtoPNet has poor diversity, which means that the network", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is frequently performing classification by comparing latent image patches with mostly the same object part, which hinders the interpretability. EvalProtoPNet and TesNet both yield much better scores than vanilla ProtoPNet. This phenomenon may be attributed to their orthogonality regularization. For EvalProtoPNet, the Distinctiveness is similar when trained with ViT and CNN backbones. However, our framework still yields significantly better Distinctiveness across all bounding box sizes. This indicates that prototypes of each class generated by our framework attend to a more diverse set of object parts, which result in less overlap between the explanations provided, thereby enhancing the interpretability. In the bottom of Figure 5, we also show the Comprehensiveness of architectures evaluated with varying thresholds. The original ProtoPNet fails to generate explanations that cover multiple parts of the object, resulting in very low scores. Among previous works, TesNet yields significantly better foreground coverage, which could also potentially be explained by additional regularizations enforced on the prototypes. Nevertheless, our framework is still able to recognize concepts that provide better coverage of various parts of the objects, achieving significantly higher Comprehensiveness.", + "bbox": [ + 511, + 340, + 906, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 696, + 694, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We presented a comparison of visualization results across ProtoPNet architectures Figure 5, which displays the activation map of top-4 prototypes in heat maps. On the left, it can be seen that the vanilla ProtoPNet has provided poor interpretability for the two examples when trained with ViT backbones. The most activated areas are mostly outside the foreground object and scattered across the background. For ProtoPool, it has learned more prototypes that attend to various areas within the object, however it still produces some explanations that highlight the background. On the other hand, EvalProtoPNet produces better visual explanations among previous ProtoPLists as most prototypes attend", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9768", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d46be831b34fd265099a9b79450dac366773f7c4118ed129f81f9245483720f3.jpg", + "image_caption": [ + "Figure 5. Visualization on the same sample across various ProtoPNet architectures. All models are trained using the same DINOv2 ViT-B backbone with $K = 5$ unless annotated." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 903, + 324 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d094fa642a1e393e70e45bfc9b5e7585ee50b0258d0cee8679df6cd55f090690.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Con.Sta.Dis.Class.
No constraints4.0975.5049.479.25
No foreground extraction28.4082.3392.6964.29
No feature space tuning63.7483.3796.0777.96
No loss55.6177.6780.4190.78
Full Model66.4082.9796.7491.02
", + "bbox": [ + 94, + 386, + 480, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. The effect of design choices made within our framework on the performance and interpretability of our method.", + "bbox": [ + 89, + 482, + 482, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to some different objects parts. However, it lacks diversity as many of them correspond to same object part. For each example, at least two activation map out of four highlight the same concept. This observation is consistent when EvalProtoPNet is trained with convolutional backbones, such as densenet161. In comparison, the prototypes learned by our framework is able to capture different parts of the bird, including head, wing, tail, and legs for both examples, resulting in the most conceptually divers set of explanations.", + "bbox": [ + 89, + 527, + 483, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 678, + 243, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we ablate various important design decisions we made in our framework. More experimental results are in the supplementary material. As shown in Table 3, all components of our framework are crucial:", + "bbox": [ + 89, + 702, + 482, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Non-parametric Prototypes. The first two rows display the importance of unique-assignment and equipartition constraints when learning non-parametric prototypes. Without these constraints, the clustering algorithm could not properly group semantically similar patch tokens into one cluster and the resulting prototypes are incapable of performing classification. Secondly, without the foreground extraction module, all the background tokens will be involved in the prototype learning stage, which significantly decrease", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the quality of the prototypes. However, when the full first stage of training is completed, that is, the feature extracted from frozen DINO ViT backbone is clustered with the constraints, the learned prototypes are already capable of recognizing objects and yielding decent interpretability.", + "bbox": [ + 511, + 388, + 906, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Space Fine-tuning. Our proposed feature space tuning greatly improves the classification accuracy of the model by separating the feature space of different classes while grounding to the part prototypes of each class. However, fine-tuning the added ViT blocks without any guidance reduces the Consistency, Stability as well as Distinctiveness, due to the disturbance of the well-structured feature space of DINO backbone. This makes $\\mathcal{L}_{PPC}$ crucial for the fine-tuning stage. By enforcing each latent patch to be more similar to their corresponding part-prototype, it further improves the accuracy while recover model's performance as well as overall interpretability in terms Consistency, Stability as well as Distinctiveness.", + "bbox": [ + 511, + 465, + 908, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 678, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we presented a framework that provides part-based interpretability for image classification. We identify existing ProtoPNETs' limitation of generating visual explanations that are often based on repetitive object parts, which limits their explainability. We mitigate this drawback by adopting non-parametric prototypes to represent diverse and comprehensive object parts. We propose Distinctiveness and Comprehensiveness scores to quantitatively evaluate the quality of explanation provided by various ProtoPNETs, complementing the existing suite of metrics, allowing future works to perform evaluation more holistically. We envision this study to serve as a foundation for future work towards stronger part-based interpretable networks.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9769", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Agnar Aamodt and Enric Plaza. Case-based reasoning: Foundational issues, methodological variations, and system approaches. 7(1):39-59. 3", + "[2] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021. 2, 3", + "[3] Muhammad Awais, Muzammal Naseer, Salman Khan, et al. Foundation models defining a new era in vision: a survey and outlook. TPAMI, 2025. 3", + "[4] Reza Akbarian Bafghi, Nidhin Harilal, Claire Monteleoni, and Maziar Raissi. Parameter efficient fine-tuning of self-supervised vits without catastrophic forgetting. In CVPR, pages 3679-3684, 2024. 5", + "[5] Zachariah Carmichael, Suhas Lohit, Anoop Cherian, et al. Pixel-grounded prototypical part networks. In WACV, pages 4768-4779, 2024. 2", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, et al. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 2, 3, 6", + "[7] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This Looks Like That: Deep Learning for Interpretable Image Recognition. In NeurIPS, 2019. 1, 2, 4, 6", + "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. 3", + "[9] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS. Curran Associates, Inc. 4", + "[10] E Dataset. Novel datasets for fine-grained image categorization. In CVPR Workshop, page 2. Citeseer, 2011. 2, 6", + "[11] Jon Donnelly, Alina Jade Barnett, and Chaofan Chen. Deformable ProtoPNet: An interpretable image classifier using deformable prototypes. In CVPR, pages 10265-10275, 2022. 1, 2, 6", + "[12] Mateo Espinosa Zarlenga, Pietro Barbiero, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. 35:21400-21413, 2022. 1, 2, 6", + "[13] Lei Fan, Yiwen Ding, Dongdong Fan, et al. Grainspace: A large-scale dataset for fine-grained and domain-adaptive recognition of cereal grains. In CVPR, pages 21116-21125, 2022. 3", + "[14] Lei Fan, Arcot Sowmya, Erik Meijering, and Yang Song. Cancer survival prediction from whole slide images with self-supervised learning and slide consistency. IEEE TMI, 42(5):1401-1412, 2022. 3", + "[15] Lei Fan, Yiwen Ding, Dongdong Fan, et al. Identifying the defective: Detecting damaged grains for cereal appearance inspection. In ECAI, pages 660-667. IOS Press, 2023. 3", + "[16] Lei Fan, Dongdong Fan, Yiwen Ding, et al. Av4gainsp: An efficient dual-camera system for identifying defective kernels of cereal grains. IEEE Robotics and Automation Letters, 9(1):851-858, 2023. 3", + "[17] Lei Fan, Dongdong Fan, Yiwen Ding, et al. Grainbrain: Multiview identification and stratification of defective grain kernels. IEEE TII, 2025. 3" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738. 3", + "[19] Qihan Huang, Mengqi Xue, Wenqi Huang, Haofei Zhang, Jie Song, Yongcheng Jing, and Mingli Song. Evaluation and improvement of interpretability for self-explainable part-prototype networks. In ICCV, pages 2011–2020, 2023. 1, 2, 5, 6, 7", + "[20] Qihan Huang, Jie Song, Jingwen Hu, et al. On the Concept Trustworthiness in Concept Bottleneck Models. AAAI, 38 (19):21161-21168, 2024. 2, 6", + "[21] Zixuan Huang and Yin Li. Interpretable and accurate fine-grained recognition via region grouping. In CVPR, pages 8662-8672. 1, 2", + "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, et al. Segment anything. In ICCV, pages 4015-4026. 3", + "[23] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In ICML, pages 5338-5348. PMLR, 2020. 1, 2, 6", + "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshops, pages 554–561, 2013. 2, 6", + "[25] Yawen Lu, Dongfang Liu, Qifan Wang, Cheng Han, Yiming Cui, Zhiwen Cao, Xueling Zhang, Yingjie Victor Chen, and Heng Fan. ProMotion: Prototypes as motion learners. In CVPR, pages 28109-28119, 2024. 3", + "[26] Tanvir Mahmud, Burhaneddin Yaman, Chun-Hao Liu, and Diana Marculescu. Papr: Training-free one-step patch pruning with lightweight convnets for faster inference. In ECCV, pages 110-128. Springer, 2024. 1, 2", + "[27] Maxime Oquab, Timothee Darcet, Theo Moutakanni, et al. DINOv2: Learning Robust Visual Features without Supervision. TMLR, 2024. 2, 6, 1", + "[28] Namuk Park, Wonjae Kim, Byeongho Heo, Taekyung Kim, and Sangdoo Yun. What do self-supervised vision transformers learn? In ICLR. 2, 3", + "[29] Zheyun Qin, Cheng Han, Qifan Wang, Xiushan Nie, Yilong Yin, and Lu Xiankai. Unified 3d segmenter as prototypical classifiers. NeurIPS, 36:46419-46432, 2023. 3", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR. 3", + "[31] Dawid Rymarczyk, Łukasz Struski, Michal Górszczak, et al. Interpretable image classification with differentiable prototypes assignment. In ECCV, pages 351-368, 2022. 6", + "[32] Oindrila Saha and Subhransu Maji. PARTICLE: Part discovery and contrastive learning for fine-grained recognition. In ICCV, pages 167-176. 3", + "[33] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In ICCV, pages 618-626. 1", + "[34] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC, 2021. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9770", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Zhenhong Sun, Junyan Wang, Zhiyu Tan, Daoyi Dong, Hailan Ma, Hao Li, and Dong Gong. EGGen: Image Generation with Multi-entity Prior Learning through Entity Guidance. In ACM MM, pages 6637-6645, 2024. 3", + "[36] Andong Tan, Fengtao Zhou, and Hao Chen. Post-hoc part-prototype networks. In ICML, 2024. 2", + "[37] Robert van der Klis, Stephan Alaniz, Massimiliano Mancini, Cassio F. Dantas, Dino Ienco, Zeynep Akata, and Diego Marcos. PDiscoNet: Semantically consistent part discovery for fine-grained recognition. In ICCV, pages 1866-1876. 1, 2", + "[38] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 2, 6", + "[39] Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In CVPR, pages 10962-10971, 2023. 1", + "[40] Jiaqi Wang, Huafeng Liu, Xinyue Wang, and Liping Jing. Interpretable image recognition by constructing transparent embedding space. In ICCV, pages 895-904, 2021. 1, 2, 6", + "[41] Junyan Wang, Zhenhong Sun, Zhiyu Tan, Xuanbai Chen, Weihua Chen, Hao Li, Cheng Zhang, and Yang Song. Towards effective usage of human-centric priors in diffusion models for text-based human image generation. In CVPR, pages 8446-8455, 2024. 3", + "[42] Wenguan Wang, Cheng Han, Tianfei Zhou, and Dongfang Liu. Visual recognition with deep nearest centroids. In ICLR. 2, 3", + "[43] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742. 3", + "[44] Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. In CVPR, pages 14475-14485, .3", + "[45] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. SimMIM: A simple framework for masked image modeling. In CVPR, pages 9653-9663, .3", + "[46] Mengqi Xue, Qihan Huang, Haofei Zhang, Jingwen Hu, Jie Song, Mingli Song, and Canghong Jin. Protopformer: concentrating on prototypical parts in vision transformers for interpretable image recognition. In IJCAI, pages 1516-1524, 2024. 2, 1", + "[47] Boyu Yang, Chang Liu, Bohao Li, Jianbin Jiao, and Qixiang Ye. Prototype mixture models for few-shot semantic segmentation. In ECCV, pages 763-778. Springer International Publishing. 2", + "[48] Rui Zhang, Xingbo Du, Junchi Yan, and Shihua Zhang. The decoupling concept bottleneck model. TPAMI, 2024. 1", + "[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, pages 2921-2929, .1", + "[50] Tianfei Zhou and Wenguan Wang. Prototype-based semantic segmentation. TPAMI, 2024. 2", + "[51] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, pages 2582-2593, 2, 3, 4" + ], + "bbox": [ + 91, + 92, + 482, + 886 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9771", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/d624a8da-e662-4173-965d-8a4cb70d4b42_model.json b/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/d624a8da-e662-4173-965d-8a4cb70d4b42_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3b1f427a97fa889b7bdeae0c5bb32355f16fd517 --- /dev/null +++ b/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/d624a8da-e662-4173-965d-8a4cb70d4b42_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.9, + 0.154 + ], + "angle": 0, + "content": "Interpretable Image Classification via Non-parametric Part Prototype Learning" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.18, + 0.741, + 0.217 + ], + "angle": 0, + "content": "Zhijie Zhu1* Lei Fan1* Maurice Pagnucco1 Yang Song1 \n1University of New South Wales" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.219, + 0.748, + 0.234 + ], + "angle": 0, + "content": "{zhijie.zhu@,lei.fanl@morri@cse,yang.song1@}.unsw.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.302, + 0.485, + 0.681 + ], + "angle": 0, + "content": "Classifying images with an interpretable decision-making process is a long-standing problem in computer vision. In recent years, Prototypical Part Networks has gained traction as an approach for self-explainable neural networks, due to their ability to mimic human visual reasoning by providing explanations based on prototypical object parts. However, the quality of the explanations generated by these methods leaves room for improvement, as the prototypes usually focus on repetitive and redundant concepts. Leveraging recent advances in prototype learning, we present a framework for part-based interpretable image classification that learns a set of semantically distinctive object parts for each class, and provides diverse and comprehensive explanations. The core of our method is to learn the part-prototypes in a non-parametric fashion, through clustering deep features extracted from foundation vision models that encode robust semantic information. To quantitatively evaluate the quality of explanations provided by ProtoPNETs, we introduce Distinctiveness Score and Comprehensiveness Score. Through evaluation on CUB-200-2011, Stanford Cars and Stanford Dogs datasets, we show that our framework compares favourably against existing ProtoPNETs while achieving better interpretability. Code is available at: https://github.com/zijizhu/proton-param." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.712, + 0.222, + 0.727 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.484, + 0.875 + ], + "angle": 0, + "content": "In recent years, self-explainable networks have emerged as a promising approach for improving the transparency of deep neural networks by providing inherent interpretability. Unlike traditional explanation methods that operate post-hoc [33, 49], recent approaches aim to mimic human reasoning by justifying predictions through an interpretable decision-making process. Examples include Prototypical-Part Networks (ProtoPNETs) [7, 11], part discovery methods [21, 37], and Concept Bottleneck Models [12, 23, 39, 48]," + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.275, + 0.713, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.274, + 0.911, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.385, + 0.711, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.385, + 0.911, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.68 + ], + "angle": 0, + "content": "Figure 1. a. Existing methods [19, 40] encode various assumptions as regularizations to guide prototype learning, but often fail to diversify the concepts learned by each prototype. Here multiple part-prototypes attend to the same image region, thereby limiting their interpretability. In contrast b. Our method partitions the feature space into semantically distinct clusters, and updates each prototype with the empirical mean of its respective cluster, enabling each prototype to learn semantically different concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.907, + 0.822 + ], + "angle": 0, + "content": "which offer explanations derived from object concepts. Among these, ProtoPNETs have garnered attention as they provide intuitive visual explanations rooted in past examples. Specifically, they define a number of prototypes for each object category, each representing a distinguishable object part, and make the prediction based on the similarity of an input image to prototypes, thereby emulating the process of human visual reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "However, existing ProtoPNETs train prototypes as learnable parameters, and guide their learning process with the core assumption that a region within each input image should match one of the prototypes from the corresponding category. However, we observe that multiple prototypes" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.22, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9762" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "from the same class often correspond to the same object part, which hinders the quality of explanations provided by the network in two ways: First, explanations become noncomprehensive, covering only a subset of object parts. Second, this results in redundant and repetitive explanations that fail to adequately illustrate why the object is classified in a particular way. An example is shown in Figure 1.a, out of the four most activated prototypes in a ProtoPNet, three of them focus on overlapping regions near the bird's wing, ignoring other parts such as the bird's legs and tail. Ideally, within an interpretability framework, the model is expected to recognize and differentiate a comprehensive set of distinctive parts for each object category, thus enhancing the clarity and transparency of its decision-making process." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.308, + 0.485, + 0.626 + ], + "angle": 0, + "content": "To approach this problem, we take inspiration from recent advances in prototype-based classification [47, 50], and encode distinct object parts using non-parametric prototypes [42, 51], as shown in Figure 1.b. Non-parametric prototypes were originally proposed to learn a more robust feature space by capturing the rich variance within each class through multiple prototypes [51]. We extend this idea to part-based interpretability to address both of the aforementioned limitations. Specifically, in comparison to previous methods that guide the learning process of learnable prototypes with regularizations [40], non-parametric prototypes allow finer control over the representation of each prototype within the same class by applying constraints on feature clustering. As a consequence, each prototype is able to capture unique characteristics of an object that differentiate it from others, resulting in semantically distinct prototypes, each latent feature patch of the foreground object is able to contribute to one of the prototypes, allowing the overall representation for each class to embody a comprehensive set of associated concepts. These properties greatly improve the interpretability of the explanations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Additionally, in order to obtain robust feature representations that encode object part semantics, We explore advanced foundation models as backbones to extract features, specifically utilizing self-supervised Vision Transformers pre-trained with self-distillation (DINO-ViTs) [6, 27], given that these backbone networks capture rich semantic information about object parts and can be used out-of-the-box [2, 28]. We also propose an efficient fine-tuning strategy for self-supervised ViT backbones to adapt the feature space to the target domain while grounding the embedding of object parts to its assigned prototypes. Hence, our framework consists of two training stages where the initial stage focuses on prototype discovery, and the subsequent stage is directed towards feature refinement. Only the image-level class labels are required in both stages. We conduct experiments to compare our approach with existing ProtoPNETs on the CUB-200-2011 [38], Stanford Cars [24], and Stanford Dogs datasets [10]. Moreover, to evaluate the diversity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.196 + ], + "angle": 0, + "content": "of concepts captured by ProtoPNETs, we introduce two new metrics: Distinctiveness Score, which quantifies the extent of overlap among object parts learned by prototypes within the same class, and Comprehensiveness Score, which measures whether the prototypes comprehensively attend to different parts of the foreground object. To summarize the key contributions of this work:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.198, + 0.905, + 0.258 + ], + "angle": 0, + "content": "- We identify a key limitation of existing ProtoPNETs, specifically generating repetitive and redundant visual explanations. To this end, we leverage non-parametric prototypes to build a self-explainable network." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.905, + 0.303 + ], + "angle": 0, + "content": "- Our method comprehensively explores non-parametric prototypes and foundation vision backbones, discovering diverse and semantic distinctive concepts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.905, + 0.41 + ], + "angle": 0, + "content": "- We introduce two evaluation metrics, namely Distinctiveness Score and Comprehensiveness Score to quantify the diversity and comprehensiveness of object part concepts. We demonstrate that our method yields comparable results against state-of-the-art ProtoPNETs models on three benchmarking datasets while recognizing holistic and distinctive concepts from each category." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.422, + 0.655, + 0.438 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.689 + ], + "angle": 0, + "content": "Part-based Interpretable Image Classification. Drawing inspiration from the human ability to distinguish objects by recognizing their semantic parts, the Prototypical Part Network (ProtoPNet) [7] emulates this reasoning process by comparing image patches with learned prototype vectors that represent object parts specific to each class. Several variants have been proposed to enhance both its interpretability and classification accuracy [5, 11, 36, 46]. Notably, TesNet [40] reformulates prototype learning from the perspective of constructing embedding space on Grassmann Manifold, and also improves the overall performance by applying various priors on prototypes, such as orthonormality. Building upon the architecture of TesNet, Huang et. al [19] introduced modules to improve consistency and stability of the visual explanations, as well as corresponding metrics for quantitative evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Meanwhile, there has been a parallel line of research that focuses on detecting object parts of fine-grained object categories [21, 37]. While these methods involve learning object part representations, they usually detect a number of shared object parts for the entire target domain and emphasize evaluating the quality of detected parts rather than leveraging them for case-based reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The main goal of our work aims to alleviate aforementioned drawbacks of ProtoPNETs, by learning part-prototypes in a non-parametric fashion. In addition, Huang et. al. [20] integrate the reasoning process of ProtoPNETs with the series of Concept Bottleneck Models [12, 23] to improve the trustworthiness of explanations [20]. We follow this pioneering work and evaluate our framework when" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9763" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.352, + 0.107 + ], + "angle": 0, + "content": "adapted to the concept learning setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.486, + 0.442 + ], + "angle": 0, + "content": "Foundation Vision Models. Recently, foundation vision backbones [6, 30] have been successfully applied to various challenging downstream tasks [3, 14, 17, 22], yielding remarkable results. In contrast to traditional supervised training, these backbones typically employ self-supervision or weak supervision schemes, such as masked image modeling [45], contrastive learning [8] or self distillation, reducing the reliance on labels and enhancing the quality of representation. Previous studies have demonstrated that each of these pre-training methods constructs an embedding space with different desirable properties [28, 44]. In particular, the attention map and patch tokens of self-supervised Vision Transformers trained with self-distillation have been shown to encode rich semantic information, such as scene layouts, which are beneficial for dense recognition tasks like and object discovery [34], anomaly detection [15, 16], and image generation [35, 41]. In addition, other works have leveraged their strong object part representation for fine-grained categorization [2, 13, 32]. Building upon these developments, our framework harnesses these desired properties to improve the quality of explanation provided by part-based interpretable image classifiers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.485, + 0.687 + ], + "angle": 0, + "content": "Prototype Learning. Prototype-based classification is to make predictions based on representative examples from past data, similar to how humans approach problem solving and learning, as suggested by psychology studies [1]. Recently, it has gained significant interest in deep learning as an alternative to traditional classification heads. For example, memory bank based methods [18, 43] retain a large number of past samples to learn high-quality representations. Approaches such as ProtoSeg [51], deep nearest centroids [42] and various segmentation models [25, 29] replace conventional parametric softmax layers of segmentors with multiple non-parametric prototypes, which not only learn class distributions but also capture the intra-class variance of each category, improving the representativeness and generalizability of the models. Our non-parametric part-prototype is inspired by these approaches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.707, + 0.182, + 0.723 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Part-based interpretability methods such as ProtoPNETs aim to explain the predictions by generating final logits based on how much input image regions look like prototypical object parts from the training set. The core of these methods is to learn a robust set of prototypes that can provide high quality explanations. As shown in Figure 2.a, our method achieves this by deriving them as non-parametric prototypes with a learning procedure disentangled from the classification process. This results in two training stages, where the first stage is dedicated to train the prototypes and the second stage focuses on refining the feature space." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.625, + 0.106 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.908, + 0.431 + ], + "angle": 0, + "content": "The goal of our model is to perform interpretable image classification by learning prototypes \\(\\mathbf{P} \\in \\mathbb{R}^{C \\times K \\times D}\\), where \\(C\\) is the number of classes and \\(K\\) is the number of prototypes belonging to each class. Each \\(D\\) dimensional vector represents one of the \\(K\\) representative parts belonging to class \\(c \\in \\{c_1, c_2, \\ldots\\}\\). During inference, given a batch of image \\(\\mathbf{X}_{1:B} = \\{\\mathbf{X}_b \\in R^{H \\times W \\times 3}\\}\\), where \\(B\\) is the batch size, \\(H\\) and \\(W\\) denote the size of the input image, a ViT backbone \\(\\mathcal{F}\\) first encodes each image into a sequence of patch tokens \\(\\mathbf{f} = \\mathcal{F}(\\mathbf{X}) \\in \\mathbb{R}^{B \\times h \\times w \\times D}\\), reshaped into a feature map of shape \\(h \\times w\\). Our model computes the similarity between the patch tokens \\(\\mathbf{f}\\) and part prototypes \\(\\mathbf{P}\\), generating a set of similarity maps \\(\\mathbf{S}_{1:B} = \\{\\mathrm{Sim}(\\mathbf{f}_b, \\mathbf{P}) \\in \\mathbb{R}^{h \\times w \\times C \\times K}\\}_{b=1}^B\\), where \\(\\mathrm{Sim}(\\cdot, \\cdot)\\) is a similarity measure, such as cosine similarity. As shown in Figure 2.a, each similarity map highlights a region of the input image similar to the learned prototypical part. Finally, the maximum value of each similarity map quantifies the occurrence of its corresponding part prototype with respect to the input, and the logit of class \\(c\\) is computed as the weighted average of all occurrence scores of its prototypes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.909, + 0.599 + ], + "angle": 0, + "content": "In the first stage of training, an efficient weakly-supervised clustering algorithm computes non-parametric part prototypes for each class based on extracted features from frozen backbones. Due to the well-structured latent space, these prototypes can be directly used for interpretable image classification. Then, we fix the learned part prototypes and fine-tune the backbone network by using the prototypes as anchors. This adapts the feature space to the target domain and significantly enhances the classification performance. We only use image-level class labels for supervision in both stages." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.833, + 0.623 + ], + "angle": 0, + "content": "3.2. Non-parametric Prototype Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.81 + ], + "angle": 0, + "content": "As shown in Figure 2.b, the main idea of this stage is to cluster the feature representation of each class involved in a batch of samples, and create a mapping between each patch token and part prototype. The empirical mean of the patch tokens assigned to the same part-prototype is directly used to update the corresponding prototypes. Hence, each latent feature patch is able to contribute to one of the part-prototypes, ensuring the overall representation of each class to cover a wide range of concepts. The only supervision involved in this stage is associating each patch token with the class label of image it belongs to, and the assignment maps are computed in an unsupervised manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Specifically, we treat a batch of latent patch tokens as a bag of features, reshaping the patch tokens \\(\\mathbf{f}\\) and similarity maps \\(\\mathbf{S}\\) to \\(\\mathbf{f} \\in \\mathbb{R}^{Bhw \\times D}\\) and \\(\\mathbf{S} \\in \\mathbb{R}^{Bhw \\times C \\times K}\\). Next, we employ a foreground extraction module to extract patches belonging to the foreground of each sample, excluding the background patches to improve representation" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9764" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.089, + 0.875, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.263, + 0.885, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.43, + 0.907, + 0.515 + ], + "angle": 0, + "content": "Figure 2. The architecture of our learning framework. a. Similar to prototype-based classification, we define \\( K \\) number of part-prototypes for each category and compare them to the latent features, which yields similarity maps. Each similarity map is pooled to generate a prototype presence vector for each class. The final logits are computed as a weighted average of the presence vector. b. Non-parametric Prototype Update: We employ feature clustering to discover class-wise non-parametric part prototype from latent feature patches. Each part prototype can be regarded as the empirical mean of its corresponding part from all training examples. c. The feature space of the backbone is fine-tuned efficiently with part-prototype fixed, and by inserting new ViT blocks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.483, + 0.604 + ], + "angle": 0, + "content": "quality. Given \\(N\\) foreground patches of class \\(c\\), denoted as \\(\\mathbf{f}^c \\subset \\mathbf{f}\\), \\(\\mathbf{f}^c \\in \\mathbb{R}^{N \\times D}\\), we obtain a similarity matrix \\(\\mathbf{S}^c \\in \\mathbb{R}^{N \\times K}\\) between \\(\\mathbf{f}^c\\) and \\(\\mathbf{P}^c\\) by indexing \\(\\mathbf{S}\\). We then compute a one-hot cluster assignment \\(\\mathbf{A} \\in \\{0,1\\}^{N \\times K}\\) that assigns each patch of class \\(c\\) to one of the part-prototypes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.483, + 0.649 + ], + "angle": 0, + "content": "In particular, we aim to optimize \\(\\mathbf{A}^c\\) so that the similarity between each patch and its assigned prototype is maximized, that is:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.666, + 0.483, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\mathbf {A} ^ {c} \\in \\mathcal {A} ^ {c}} \\operatorname {T r} \\left(\\left(\\mathbf {A} ^ {c}\\right) ^ {\\top} \\mathbf {S} ^ {c}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.699, + 0.483, + 0.76 + ], + "angle": 0, + "content": "where \\(\\operatorname{Tr}(\\cdot)\\) is the trace of a matrix. To prevent trivial solutions for \\(\\mathbf{A}^c\\) and ensure each cluster represents a distinct object part, we enforce unique assignment and equipartition constraints [51]:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.773, + 0.483, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {A} ^ {c}\\right) ^ {\\top} \\mathbf {1} _ {N} = \\frac {1}{K} \\mathbf {1} _ {K}, \\mathbf {A} ^ {c} \\mathbf {1} _ {K} = \\frac {1}{N} \\mathbf {1} _ {N}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "This ensures each patch token maps to a single part-prototype of class \\( c \\), and each prototype has an approximately equal number of patch tokens assigned. This formulation can be regarded as an Optimal Transport (OT) problem, where the approximate solution of \\( \\mathbf{A} \\) can be estimated by solving a smoothed version of equation (1):" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.542, + 0.905, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} ^ {c, *} = \\max _ {\\mathbf {A} ^ {c} \\in \\mathcal {A} ^ {c}} \\operatorname {T r} \\left(\\left(\\mathbf {A} ^ {c}\\right) ^ {\\top} \\mathbf {S}\\right) + \\kappa h \\left(\\mathbf {A} ^ {c}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.631 + ], + "angle": 0, + "content": "where \\(\\kappa h(\\mathbf{A})\\) is the entropic regularization term controlling smoothness. We use \\(\\kappa = 0.05\\) following previous works [9]. The solution can be efficiently computed using the Sinkhorn-Knopp Algorithm:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.641, + 0.905, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} ^ {c, *} = \\operatorname {d i a g} (\\mathbf {u}) \\exp \\left(\\frac {\\mathbf {S} ^ {c}}{\\kappa}\\right) \\operatorname {d i a g} (\\mathbf {v}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.907, + 0.725 + ], + "angle": 0, + "content": "Momentum Update of Prototypes. Given the estimated one-hot patch-prototype assignment \\(\\mathbf{A}^{c,*}\\), each part-prototype \\(\\mathbf{P}_k^c\\) is updated as the mean of its assigned patches:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.734, + 0.905, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {k} ^ {c} \\leftarrow \\mathbf {P} _ {k} ^ {c} + (1 - \\beta) \\overline {{\\mathbf {f}}} ^ {c, k}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.84 + ], + "angle": 0, + "content": "where \\(\\beta \\in [0,1]\\) is a coefficient that controls the update speed, and \\(\\overline{\\mathbf{f}}\\) is the mean feature. This learning scheme effectively captures the mean embedding of object parts across all training examples, and can be visualized by projecting to the closest patches from the training set [7]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.884, + 0.864 + ], + "angle": 0, + "content": "3.3. Adapt to Interpretable Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in Figure 2.a, the learned part-prototypes can be directly used for classification. The activation of each pro" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9765" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.475, + 0.109 + ], + "angle": 0, + "content": "totype \\(\\mathbf{P}_k^c\\) with respect to the input image is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.119, + 0.483, + 0.147 + ], + "angle": 0, + "content": "\\[\ng _ {k} ^ {c} = \\max _ {\\widetilde {\\mathbf {f}} \\in \\mathbf {f}} \\quad \\operatorname {S i m} \\left(\\widetilde {\\mathbf {f}}, \\mathbf {P} _ {k} ^ {c}\\right) \\in \\mathbb {R}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.151, + 0.483, + 0.211 + ], + "angle": 0, + "content": "As not all prototypes are equally important for classifying the input image, following previous work [19], we generate the final logit for each class \\( c \\) by computing a weighted average of \\( g^{c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.222, + 0.483, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\log \\mathrm {i t} ^ {c} = \\sum_ {k = 1} ^ {K} w _ {k} ^ {c} \\cdot g _ {k} ^ {c}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.484, + 0.357 + ], + "angle": 0, + "content": "where \\( w \\) is a learned modulation vector that controls the contribution of each prototype activation towards the logit of its corresponding class. In contrast to using a fully-connected layer, computing class logits as a weighted average ensures that prototypes contribute exclusively to their respective class, thereby enhancing the interpretability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.484, + 0.629 + ], + "angle": 0, + "content": "Prototype-Anchored Feature Space Tuning. After discovering class-wise part-prototypes from the pre-trained backbone, we fine-tune the backbone to ground the features to the learned prototypes and adapt to the target domain. To this end, we employ Block Expansion [4], as shown in Figure 2.c. Given a ViT backbone \\(\\mathcal{F}\\) with a sequence of transformer blocks \\((\\mathcal{F}_1,\\mathcal{F}_2,\\mathcal{F}_3,\\ldots)\\), Block Expansion divides the blocks into \\(m\\) groups and inserts a new identity block \\(\\mathcal{F}_{id}\\) at the end of each divided group. Each inserted identity block is initialized with weights of the preceding block, with the weights of attention projection layer and the second MLP layer set to zero, such that \\(\\mathcal{F}_{id}(x) = x\\), i.e., the output of the backbone remains unchanged without further training. During fine-tuning, we freeze all original weights and part-prototypes, and only fine-tune the added identity blocks and modulation vectors. This approach efficiently produces features that separate classes well without disturbing the well-structured latent space of the backbone." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.721 + ], + "angle": 0, + "content": "Patch-Prototype Distance Optimization. In the fine-tuning stage, we use a standard cross-entropy loss \\(\\mathcal{L}_{ce}\\) to separate features of each class. In addition, we ground each image patch to its assigned part-prototype to improve feature locality. To this end, we introduce an in-class patch-prototype contrastive loss, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.74, + 0.483, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P P C}} = \\sum_ {c \\in \\mathcal {C}} \\frac {\\exp \\left(\\left(\\mathbf {f} _ {i} ^ {c}\\right) ^ {\\top} \\mathbf {P} _ {k _ {i}} ^ {c}\\right)}{\\exp \\left(\\left(\\mathbf {f} _ {i} ^ {c}\\right) ^ {\\top} \\mathbf {P} _ {k _ {i}} ^ {c}\\right) + \\sum_ {\\mathbf {P} _ {k _ {j}} ^ {c} \\in \\mathcal {P} ^ {c -}} \\exp \\left(\\left(\\mathbf {f} _ {i} ^ {c}\\right) ^ {\\top} \\mathbf {P} _ {k _ {j}} ^ {c}\\right)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.871 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}^{c-} = \\{\\mathbf{P}_k^c\\}_{k=1}^K / \\mathbf{P}_{k_i}^c\\). This loss \"pulls\" each patch towards its assigned part-prototype in the latent space, and separate it from other part-prototype within the same class. Hence, we jointly train the expanded ViT block and the modulation weights with the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {c e}} + \\lambda_ {\\mathrm {P P C}} \\mathcal {L} _ {\\mathrm {P P C}}. \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.712, + 0.106 + ], + "angle": 0, + "content": "3.4. Distinctiveness Score" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.905, + 0.221 + ], + "angle": 0, + "content": "Ideally, each part-prototype belonging to the same class should learn a distinct concept from objects; otherwise, the explanation would be confusing and redundant. To quantify this property, we introduce a metric, Distinctiveness Score that measures the amount of overlap between image regions highlighted by the part prototypes associated with the ground truth class of the input image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.223, + 0.907, + 0.39 + ], + "angle": 0, + "content": "Specifically, given an input image \\(x\\) of class \\(c_{x}\\) and the corresponding activation maps \\(S^{c_x} \\in \\mathbb{R}^{h \\times w \\times K}\\) generated by prototypes \\(\\mathbf{P}^{c_x}\\), we first resize the activation maps to \\(\\hat{S}^{c_x} \\in \\mathbb{R}^{K \\times H \\times W}\\) to match the size of input \\(x\\). Next, we extract the object region \\(R^{c_x, k}(x)\\) highlighted by each activation map \\(S^{c_x, k}\\) as a fix-sized bounding box with predetermined height \\(H_b\\) and width \\(W_b\\) around its max activation. The amount of overlap \\(O(x)\\) between the similarity maps \\(S^{c_x}\\) generated by ground truth class prototypes is then quantified by the Intersection over Union (IoU) between each pair of highlighted regions \\(R^{c_x, k}(x)\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.403, + 0.905, + 0.458 + ], + "angle": 0, + "content": "\\[\nO (x) = \\frac {1}{K (K - 1) / 2} \\sum_ {k = 1} ^ {K - 1} \\sum_ {l = k + 1} ^ {K} \\operatorname {I o U} \\left(R ^ {c _ {x}, k} (x), R ^ {c _ {x}, l} (x)\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.906, + 0.505 + ], + "angle": 0, + "content": "The Distinctiveness Score is computed based on the average overlap across all samples in the test set \\(\\mathcal{X}\\), and a higher value is better. It is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.518, + 0.905, + 0.554 + ], + "angle": 0, + "content": "\\[\n1 - \\frac {1}{| \\mathcal {X} |} \\sum_ {x \\in \\mathcal {X}} O (x). \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.569, + 0.749, + 0.585 + ], + "angle": 0, + "content": "3.5. Comprehensiveness Score" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.775 + ], + "angle": 0, + "content": "A good Distinctiveness Score by itself does not guarantee good interpretability, as a high score could be achieved by networks with prototypes correspond to different image regions that lack meaningful concepts, e.g., background patches that contribute less to object recognition. Therefore, we further introduce Comprehensiveness Score to evaluate how thoroughly the object foreground is attended by the prototypes. Given an input image \\( x \\) and the corresponding activation maps \\( S^{c_x} \\) generated by prototypes \\( \\mathbf{P}^{c_x} \\), we perform bin-max normalization on each \\( S^{c_x,k} \\) and obtain a mask \\( \\hat{M}_i^{c_x,k}(x) \\) of the most activated region of \\( \\mathbf{P}^{c_x,k} \\) using a threshold \\( \\tau \\):" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.786, + 0.905, + 0.839 + ], + "angle": 0, + "content": "\\[\n\\tilde {S} _ {i} ^ {c _ {x}, k} = \\frac {S _ {i} ^ {c _ {x} , k} - S _ {\\min } ^ {c _ {x} , k}}{S _ {\\max } ^ {c _ {x} , k} - S _ {\\min } ^ {c _ {x} , k}}, \\quad \\hat {M} _ {i} ^ {c _ {x}, k} (x) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\tilde {S} _ {i} ^ {c _ {x}, k} \\geq \\tau \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We then compute the IoU between the union of all masks \\(\\hat{M}^{c_x,k}(x)\\) and the ground truth object foreground mask \\(M_{x}\\). The Comprehensiveness Score is calculated as the average IoU across all test samples \\(\\mathcal{X}\\), and a higher value indicates" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9766" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.099, + 0.465, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.328, + 0.486, + 0.4 + ], + "angle": 0, + "content": "Figure 3. a. The Distinctiveness score for one sample is calculated based on the amount of overlap between areas attended by prototypes. b. The Comprehensiveness score is computed by comparing the ground truth foreground mask \\( M_{x} \\) with the union of thresholded image region attended by prototypes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.457, + 0.442 + ], + "angle": 0, + "content": "better comprehensiveness. Specifically, it is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.453, + 0.483, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{| \\mathcal {X} |} \\sum_ {x \\in \\mathcal {X}} \\operatorname {I o U} \\left(M _ {x}, \\bigcup_ {k = 1} ^ {K} \\hat {M} ^ {c _ {x}, k} (x)\\right). \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.52, + 0.224, + 0.537 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.3, + 0.562 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Datasets. To evaluate the classification accuracy and interpretability of our proposed framework, we conduct experiments on CUB-200-2011 [38], Stanford Cars [24] and Stanford Dogs [10] dataset, following previous works on ProtoPNETs. The CUB-200-2011 dataset consists of 11,788 images spanning 200 bird species. Stanford Cars includes 16,185 images from 196 car models, and Stanford Dogs includes 20,580 images from 120 breeds. The results on Stanford Cars and Stanford Dogs are in supplementary. We also extend our evaluation by performing concept learning [20, 23] experiments on CUB-200-2011 to evaluate whether the learned prototypes can be adapted to produce trustworthy concept-based explanations by following the evaluation protocol introduced in [20]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Baselines and Evaluation Metrics. We establish our main benchmark by comparing against existing ProtoPNETs, namely ProtoPNet [7], Deformable ProtoPNet [11], ProtoPool [31], TesNet [40] and EvalProtoPNet [19]. We select three self-supervised Vision Transformer backbone as the backbone: the latest checkpoints of DINOv2 ViT-B and ViT-S trained on LVD-142M [27] and the original DINO trained on ImageNet [6]. All the networks are trained with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.319 + ], + "angle": 0, + "content": "image-level class labels only, and all previous methods were reproduced with official implementations. For evaluation metrics, we report top-1 accuracy for classification. We also quantitatively measure the interpretability of the learned prototypes with the introduced Distinctiveness and Comprehensiveness scores, as well as Consistency and Stability proposed in [19]. Specifically, Consistency evaluates whether each prototype correspond to the same object part across different images, while Stability measures whether each prototype correspond to the same object part when the image is perturbed with noise. We evaluate all methods with \\( K = 3 \\) and \\( K = 5 \\). For concept learning, we compare with existing Concept Bottleneck Models [12, 20, 23], and evaluate the classification performance as well as the trustworthiness [20] of the generated concept explanations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Implementation Details. In the first stage of training, we set \\(\\beta = 0.99\\), and \\(\\kappa = 0.05\\) to perform clustering. When fine-tuning the backbone, we set \\(\\lambda_{\\mathrm{PPC}} = 0.8\\) and the number of added ViT block \\(m = 1\\). We trained our framework with 1 epoch for the first stage, and 5 more epochs for the second stage with Adam optimizer, a learning rate of 0.0001 for the backbone and \\(1e - 6\\) for modulation vector \\(w\\). When training previous ProtoPNETs, we set \\(m = 3\\) as it produces the best results across all previous architectures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.708, + 0.484 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.84 + ], + "angle": 0, + "content": "We reported the benchmark result of our method against previous ProtoPLists in Table 1. Here the Distinctiveness is calculated with \\(1/4\\) of image size. Our method outperformed previous methods in terms of classification accuracy by \\((0.16\\%, 2.39\\%)\\) when trained with both the revised DINOv2 and DINO ViT-B backbones, and the accuracy evaluated on ViT-S is comparable to previous works. In addition, while our method yields similar Consistency to [19], it achieves higher Stability of than existing works in most settings. Nevertheless, the most notable improvement of our framework is the Distinctiveness, which surpasses previous methods by \\((4.83\\%, 3.49\\%, 8.93\\%)\\) when setting 5 prototypes for each class. This indicates that our framework is more capable of utilizing the strong part representation encoded in self-supervised Vision Transformers to discover a diverse set of prototypes. Finally, the prototypes learned by our framework can be directly adapted to provide concept-based explanations, as shown by the results in Table 2. Our method yields comparable trustworthiness score with higher classification performance than [20], and Concept Bottleneck Models constructed naively on top of ViT backbones failed to produce trustworthy explanations despite their higher accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Distinctiveness and Comprehensiveness of Prototypes. We provide a more detailed breakdown on the diversity of prototypes learned by various ProtoPNETs, by selecting a few architectures and evaluate the Distinctiveness with" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9767" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.088, + 0.904, + 0.284 + ], + "angle": 0, + "content": "
DINOv2 ViT-B/14DINOv2 ViT-S/14DINO ViT-B/16
Con.Sta.Dis.Class.Con.Sta.Dis.Class.Con.Sta.Dis.Class.
ProtoPNet (K=3)0.7768.405.7788.474.8474.688.8384.903.0063.873.1475.94
ProtoPNet (K=5)1.3063.8411.9988.164.1270.1016.7385.523.0246.346.7978.91
Deformable ProtoPNet (K=3)0.9580.005.9086.999.6761.181.6583.091.5570.849.1571.49
Deformable ProtoPNet (K=5)2.8380.4310.3488.1411.7379.874.7685.540.9573.161.1781.74
ProtoPool (K=3)29.8352.8387.5988.7215.5047.8991.1786.1926.6777.7483.8972.04
ProtoPool (K=5)29.2757.4989.6289.9012.152.4492.2687.4137.4052.7387.6678.18
TesNet (K=3)36.3366.4541.2490.6658.0066.8927.8088.2335.5065.4058.8581.24
TesNet (K=5)58.6074.6081.9390.3567.8070.2781.1688.9551.3070.8181.9482.29
EvalProtoPNet (K=3)67.8371.3925.6789.7345.1763.5117.0187.6637.8369.3835.1084.21
EvalProtoPNet (K=5)67.8074.4072.1290.0654.6070.2777.4188.1951.0966.8672.7484.54
Ours (K=3)61.1778.9886.4490.3761.5075.5789.1389.0745.6769.9095.6285.12
Ours (K=5)66.4082.9794.4590.8266.0181.3195.7588.1647.2075.5796.5986.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.293, + 0.908, + 0.324 + ], + "angle": 0, + "content": "Table 1. Consistency score (%), Stability Score (%), Distinctiveness Score (%), and classification accuracy (% of various ProtoPNet on CUB-200-2011. K denotes the number of prototypes defined for each class. We train previous methods with their official implementations." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.342, + 0.482, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.592, + 0.483, + 0.634 + ], + "angle": 0, + "content": "Figure 4. Distinctiveness \\((\\%)\\) and Comprehensiveness \\((\\%)\\) across various ProtoPLists, evaluated on different box sizes and thresholds. All methods besides [19] are trained with DINOv2 ViT-B." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.652, + 0.441, + 0.763 + ], + "angle": 0, + "content": "
ViT-BViT-S
Loc.Acc.Loc.Acc.
Vanilla CBM15.6390.0612.9388.43
CBM-h11.9589.2411.4186.52
CEM13.0488.7410.6186.41
Huang et. al.45.9587.4347.3384.79
Ours (K=3)41.1887.0439.8884.55
Ours (K=5)47.0488.1844.9785.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.773, + 0.483, + 0.844 + ], + "angle": 0, + "content": "Table 2. Concept trustworthiness (\\%) and classification accuracy (\\%) of a CBM constructed based on our framework, in comparison with previous methods on CUB-200-2011 dataset. All methods are trained with DINOV2 ViTs as the backbone. Note that CBM-h is vanilla CBM with a residual fitting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "varying bounding box sizes. The results are shown in Figure 4. Notably, the prototypes learned by the vanilla ProtoPNet has poor diversity, which means that the network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.689 + ], + "angle": 0, + "content": "is frequently performing classification by comparing latent image patches with mostly the same object part, which hinders the interpretability. EvalProtoPNet and TesNet both yield much better scores than vanilla ProtoPNet. This phenomenon may be attributed to their orthogonality regularization. For EvalProtoPNet, the Distinctiveness is similar when trained with ViT and CNN backbones. However, our framework still yields significantly better Distinctiveness across all bounding box sizes. This indicates that prototypes of each class generated by our framework attend to a more diverse set of object parts, which result in less overlap between the explanations provided, thereby enhancing the interpretability. In the bottom of Figure 5, we also show the Comprehensiveness of architectures evaluated with varying thresholds. The original ProtoPNet fails to generate explanations that cover multiple parts of the object, resulting in very low scores. Among previous works, TesNet yields significantly better foreground coverage, which could also potentially be explained by additional regularizations enforced on the prototypes. Nevertheless, our framework is still able to recognize concepts that provide better coverage of various parts of the objects, achieving significantly higher Comprehensiveness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.696, + 0.713 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We presented a comparison of visualization results across ProtoPNet architectures Figure 5, which displays the activation map of top-4 prototypes in heat maps. On the left, it can be seen that the vanilla ProtoPNet has provided poor interpretability for the two examples when trained with ViT backbones. The most activated areas are mostly outside the foreground object and scattered across the background. For ProtoPool, it has learned more prototypes that attend to various areas within the object, however it still produces some explanations that highlight the background. On the other hand, EvalProtoPNet produces better visual explanations among previous ProtoPLists as most prototypes attend" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9768" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.09, + 0.905, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.335, + 0.908, + 0.364 + ], + "angle": 0, + "content": "Figure 5. Visualization on the same sample across various ProtoPNet architectures. All models are trained using the same DINOv2 ViT-B backbone with \\( K = 5 \\) unless annotated." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.387, + 0.481, + 0.475 + ], + "angle": 0, + "content": "
Con.Sta.Dis.Class.
No constraints4.0975.5049.479.25
No foreground extraction28.4082.3392.6964.29
No feature space tuning63.7483.3796.0777.96
No loss55.6177.6780.4190.78
Full Model66.4082.9796.7491.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.483, + 0.483, + 0.513 + ], + "angle": 0, + "content": "Table 3. The effect of design choices made within our framework on the performance and interpretability of our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.484, + 0.665 + ], + "angle": 0, + "content": "to some different objects parts. However, it lacks diversity as many of them correspond to same object part. For each example, at least two activation map out of four highlight the same concept. This observation is consistent when EvalProtoPNet is trained with convolutional backbones, such as densenet161. In comparison, the prototypes learned by our framework is able to capture different parts of the bird, including head, wing, tail, and legs for both examples, resulting in the most conceptually divers set of explanations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.679, + 0.245, + 0.695 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.763 + ], + "angle": 0, + "content": "In this section, we ablate various important design decisions we made in our framework. More experimental results are in the supplementary material. As shown in Table 3, all components of our framework are crucial:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Non-parametric Prototypes. The first two rows display the importance of unique-assignment and equipartition constraints when learning non-parametric prototypes. Without these constraints, the clustering algorithm could not properly group semantically similar patch tokens into one cluster and the resulting prototypes are incapable of performing classification. Secondly, without the foreground extraction module, all the background tokens will be involved in the prototype learning stage, which significantly decrease" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.907, + 0.466 + ], + "angle": 0, + "content": "the quality of the prototypes. However, when the full first stage of training is completed, that is, the feature extracted from frozen DINO ViT backbone is clustered with the constraints, the learned prototypes are already capable of recognizing objects and yielding decent interpretability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.909, + 0.662 + ], + "angle": 0, + "content": "Feature Space Fine-tuning. Our proposed feature space tuning greatly improves the classification accuracy of the model by separating the feature space of different classes while grounding to the part prototypes of each class. However, fine-tuning the added ViT blocks without any guidance reduces the Consistency, Stability as well as Distinctiveness, due to the disturbance of the well-structured feature space of DINO backbone. This makes \\(\\mathcal{L}_{PPC}\\) crucial for the fine-tuning stage. By enforcing each latent patch to be more similar to their corresponding part-prototype, it further improves the accuracy while recover model's performance as well as overall interpretability in terms Consistency, Stability as well as Distinctiveness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we presented a framework that provides part-based interpretability for image classification. We identify existing ProtoPNETs' limitation of generating visual explanations that are often based on repetitive object parts, which limits their explainability. We mitigate this drawback by adopting non-parametric prototypes to represent diverse and comprehensive object parts. We propose Distinctiveness and Comprehensiveness scores to quantitatively evaluate the quality of explanation provided by various ProtoPNETs, complementing the existing suite of metrics, allowing future works to perform evaluation more holistically. We envision this study to serve as a foundation for future work towards stronger part-based interpretable networks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9769" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Agnar Aamodt and Enric Plaza. Case-based reasoning: Foundational issues, methodological variations, and system approaches. 7(1):39-59. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[2] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.482, + 0.242 + ], + "angle": 0, + "content": "[3] Muhammad Awais, Muzammal Naseer, Salman Khan, et al. Foundation models defining a new era in vision: a survey and outlook. TPAMI, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[4] Reza Akbarian Bafghi, Nidhin Harilal, Claire Monteleoni, and Maziar Raissi. Parameter efficient fine-tuning of self-supervised vits without catastrophic forgetting. In CVPR, pages 3679-3684, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[5] Zachariah Carmichael, Suhas Lohit, Anoop Cherian, et al. Pixel-grounded prototypical part networks. In WACV, pages 4768-4779, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, et al. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[7] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This Looks Like That: Deep Learning for Interpretable Image Recognition. In NeurIPS, 2019. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[9] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS. Curran Associates, Inc. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.517, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[10] E Dataset. Novel datasets for fine-grained image categorization. In CVPR Workshop, page 2. Citeseer, 2011. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.546, + 0.482, + 0.6 + ], + "angle": 0, + "content": "[11] Jon Donnelly, Alina Jade Barnett, and Chaofan Chen. Deformable ProtoPNet: An interpretable image classifier using deformable prototypes. In CVPR, pages 10265-10275, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.603, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[12] Mateo Espinosa Zarlenga, Pietro Barbiero, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. 35:21400-21413, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.646, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[13] Lei Fan, Yiwen Ding, Dongdong Fan, et al. Grainspace: A large-scale dataset for fine-grained and domain-adaptive recognition of cereal grains. In CVPR, pages 21116-21125, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[14] Lei Fan, Arcot Sowmya, Erik Meijering, and Yang Song. Cancer survival prediction from whole slide images with self-supervised learning and slide consistency. IEEE TMI, 42(5):1401-1412, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[15] Lei Fan, Yiwen Ding, Dongdong Fan, et al. Identifying the defective: Detecting damaged grains for cereal appearance inspection. In ECAI, pages 660-667. IOS Press, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[16] Lei Fan, Dongdong Fan, Yiwen Ding, et al. Av4gainsp: An efficient dual-camera system for identifying defective kernels of cereal grains. IEEE Robotics and Automation Letters, 9(1):851-858, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[17] Lei Fan, Dongdong Fan, Yiwen Ding, et al. Grainbrain: Multiview identification and stratification of defective grain kernels. IEEE TII, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[19] Qihan Huang, Mengqi Xue, Wenqi Huang, Haofei Zhang, Jie Song, Yongcheng Jing, and Mingli Song. Evaluation and improvement of interpretability for self-explainable part-prototype networks. In ICCV, pages 2011–2020, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[20] Qihan Huang, Jie Song, Jingwen Hu, et al. On the Concept Trustworthiness in Concept Bottleneck Models. AAAI, 38 (19):21161-21168, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.904, + 0.289 + ], + "angle": 0, + "content": "[21] Zixuan Huang and Yin Li. Interpretable and accurate fine-grained recognition via region grouping. In CVPR, pages 8662-8672. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.904, + 0.319 + ], + "angle": 0, + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, et al. Segment anything. In ICCV, pages 4015-4026. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[23] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In ICML, pages 5338-5348. PMLR, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshops, pages 554–561, 2013. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[25] Yawen Lu, Dongfang Liu, Qifan Wang, Cheng Han, Yiming Cui, Zhiwen Cao, Xueling Zhang, Yingjie Victor Chen, and Heng Fan. ProMotion: Prototypes as motion learners. In CVPR, pages 28109-28119, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[26] Tanvir Mahmud, Burhaneddin Yaman, Chun-Hao Liu, and Diana Marculescu. Papr: Training-free one-step patch pruning with lightweight convnets for faster inference. In ECCV, pages 110-128. Springer, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[27] Maxime Oquab, Timothee Darcet, Theo Moutakanni, et al. DINOv2: Learning Robust Visual Features without Supervision. TMLR, 2024. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[28] Namuk Park, Wonjae Kim, Byeongho Heo, Taekyung Kim, and Sangdoo Yun. What do self-supervised vision transformers learn? In ICLR. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[29] Zheyun Qin, Cheng Han, Qifan Wang, Xiushan Nie, Yilong Yin, and Lu Xiankai. Unified 3d segmenter as prototypical classifiers. NeurIPS, 36:46419-46432, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[31] Dawid Rymarczyk, Łukasz Struski, Michal Górszczak, et al. Interpretable image classification with differentiable prototypes assignment. In ECCV, pages 351-368, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[32] Oindrila Saha and Subhransu Maji. PARTICLE: Part discovery and contrastive learning for fine-grained recognition. In ICCV, pages 167-176. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[33] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In ICCV, pages 618-626. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[34] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9770" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[35] Zhenhong Sun, Junyan Wang, Zhiyu Tan, Daoyi Dong, Hailan Ma, Hao Li, and Dong Gong. EGGen: Image Generation with Multi-entity Prior Learning through Entity Guidance. In ACM MM, pages 6637-6645, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[36] Andong Tan, Fengtao Zhou, and Hao Chen. Post-hoc part-prototype networks. In ICML, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[37] Robert van der Klis, Stephan Alaniz, Massimiliano Mancini, Cassio F. Dantas, Dino Ienco, Zeynep Akata, and Diego Marcos. PDiscoNet: Semantically consistent part discovery for fine-grained recognition. In ICCV, pages 1866-1876. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[38] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[39] Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In CVPR, pages 10962-10971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[40] Jiaqi Wang, Huafeng Liu, Xinyue Wang, and Liping Jing. Interpretable image recognition by constructing transparent embedding space. In ICCV, pages 895-904, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[41] Junyan Wang, Zhenhong Sun, Zhiyu Tan, Xuanbai Chen, Weihua Chen, Hao Li, Cheng Zhang, and Yang Song. Towards effective usage of human-centric priors in diffusion models for text-based human image generation. In CVPR, pages 8446-8455, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[42] Wenguan Wang, Cheng Han, Tianfei Zhou, and Dongfang Liu. Visual recognition with deep nearest centroids. In ICLR. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[43] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[44] Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. In CVPR, pages 14475-14485, .3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[45] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. SimMIM: A simple framework for masked image modeling. In CVPR, pages 9653-9663, .3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[46] Mengqi Xue, Qihan Huang, Haofei Zhang, Jingwen Hu, Jie Song, Mingli Song, and Canghong Jin. Protopformer: concentrating on prototypical parts in vision transformers for interpretable image recognition. In IJCAI, pages 1516-1524, 2024. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[47] Boyu Yang, Chang Liu, Bohao Li, Jianbin Jiao, and Qixiang Ye. Prototype mixture models for few-shot semantic segmentation. In ECCV, pages 763-778. Springer International Publishing. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[48] Rui Zhang, Xingbo Du, Junchi Yan, and Shihua Zhang. The decoupling concept bottleneck model. TPAMI, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, pages 2921-2929, .1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[50] Tianfei Zhou and Wenguan Wang. Prototype-based semantic segmentation. TPAMI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.482, + 0.887 + ], + "angle": 0, + "content": "[51] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, pages 2582-2593, 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.887 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9771" + } + ] +] \ No newline at end of file diff --git a/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/d624a8da-e662-4173-965d-8a4cb70d4b42_origin.pdf b/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/d624a8da-e662-4173-965d-8a4cb70d4b42_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0655b29b6751ef17130ec596f3c546a972b352f2 --- /dev/null +++ b/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/d624a8da-e662-4173-965d-8a4cb70d4b42_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35d9757dac4ec0ee3fd0dbd3c9672a055629d833809b03ce33e7ef8b7bd4ba5a +size 3291981 diff --git a/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/full.md b/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cec9965900ff76fcd2b4157ac1a1c5f26c7aa5e6 --- /dev/null +++ b/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/full.md @@ -0,0 +1,291 @@ +# Interpretable Image Classification via Non-parametric Part Prototype Learning + +Zhijie Zhu1* Lei Fan1* Maurice Pagnucco1 Yang Song1 +1University of New South Wales + +{zhijie.zhu@,lei.fanl@morri@cse,yang.song1@}.unsw.edu.au + +# Abstract + +Classifying images with an interpretable decision-making process is a long-standing problem in computer vision. In recent years, Prototypical Part Networks has gained traction as an approach for self-explainable neural networks, due to their ability to mimic human visual reasoning by providing explanations based on prototypical object parts. However, the quality of the explanations generated by these methods leaves room for improvement, as the prototypes usually focus on repetitive and redundant concepts. Leveraging recent advances in prototype learning, we present a framework for part-based interpretable image classification that learns a set of semantically distinctive object parts for each class, and provides diverse and comprehensive explanations. The core of our method is to learn the part-prototypes in a non-parametric fashion, through clustering deep features extracted from foundation vision models that encode robust semantic information. To quantitatively evaluate the quality of explanations provided by ProtoPNETs, we introduce Distinctiveness Score and Comprehensiveness Score. Through evaluation on CUB-200-2011, Stanford Cars and Stanford Dogs datasets, we show that our framework compares favourably against existing ProtoPNETs while achieving better interpretability. Code is available at: https://github.com/zijizhu/proton-param. + +# 1. Introduction + +In recent years, self-explainable networks have emerged as a promising approach for improving the transparency of deep neural networks by providing inherent interpretability. Unlike traditional explanation methods that operate post-hoc [33, 49], recent approaches aim to mimic human reasoning by justifying predictions through an interpretable decision-making process. Examples include Prototypical-Part Networks (ProtoPNETs) [7, 11], part discovery methods [21, 37], and Concept Bottleneck Models [12, 23, 39, 48], + +![](images/b76c8f6738c9c2e49fc63bea9457e7d288ab9f8a5330b1afa68a4d58c4ece5fe.jpg) + +![](images/d86ba279cc30b7e1c3220cb1f89e69dfea207784207b11cdca0a41c7d6ca6a86.jpg) + +![](images/5aa1bf7f2442eee595fdaf9328da5409d8cf1294f40150b68047dde94a6edb7f.jpg) +Figure 1. a. Existing methods [19, 40] encode various assumptions as regularizations to guide prototype learning, but often fail to diversify the concepts learned by each prototype. Here multiple part-prototypes attend to the same image region, thereby limiting their interpretability. In contrast b. Our method partitions the feature space into semantically distinct clusters, and updates each prototype with the empirical mean of its respective cluster, enabling each prototype to learn semantically different concepts. + +![](images/093e69ec04edc2deb830aef71f7ccbbeecf627885c5fab593b65c9debac02573.jpg) + +which offer explanations derived from object concepts. Among these, ProtoPNETs have garnered attention as they provide intuitive visual explanations rooted in past examples. Specifically, they define a number of prototypes for each object category, each representing a distinguishable object part, and make the prediction based on the similarity of an input image to prototypes, thereby emulating the process of human visual reasoning. + +However, existing ProtoPNETs train prototypes as learnable parameters, and guide their learning process with the core assumption that a region within each input image should match one of the prototypes from the corresponding category. However, we observe that multiple prototypes + +from the same class often correspond to the same object part, which hinders the quality of explanations provided by the network in two ways: First, explanations become noncomprehensive, covering only a subset of object parts. Second, this results in redundant and repetitive explanations that fail to adequately illustrate why the object is classified in a particular way. An example is shown in Figure 1.a, out of the four most activated prototypes in a ProtoPNet, three of them focus on overlapping regions near the bird's wing, ignoring other parts such as the bird's legs and tail. Ideally, within an interpretability framework, the model is expected to recognize and differentiate a comprehensive set of distinctive parts for each object category, thus enhancing the clarity and transparency of its decision-making process. + +To approach this problem, we take inspiration from recent advances in prototype-based classification [47, 50], and encode distinct object parts using non-parametric prototypes [42, 51], as shown in Figure 1.b. Non-parametric prototypes were originally proposed to learn a more robust feature space by capturing the rich variance within each class through multiple prototypes [51]. We extend this idea to part-based interpretability to address both of the aforementioned limitations. Specifically, in comparison to previous methods that guide the learning process of learnable prototypes with regularizations [40], non-parametric prototypes allow finer control over the representation of each prototype within the same class by applying constraints on feature clustering. As a consequence, each prototype is able to capture unique characteristics of an object that differentiate it from others, resulting in semantically distinct prototypes, each latent feature patch of the foreground object is able to contribute to one of the prototypes, allowing the overall representation for each class to embody a comprehensive set of associated concepts. These properties greatly improve the interpretability of the explanations. + +Additionally, in order to obtain robust feature representations that encode object part semantics, We explore advanced foundation models as backbones to extract features, specifically utilizing self-supervised Vision Transformers pre-trained with self-distillation (DINO-ViTs) [6, 27], given that these backbone networks capture rich semantic information about object parts and can be used out-of-the-box [2, 28]. We also propose an efficient fine-tuning strategy for self-supervised ViT backbones to adapt the feature space to the target domain while grounding the embedding of object parts to its assigned prototypes. Hence, our framework consists of two training stages where the initial stage focuses on prototype discovery, and the subsequent stage is directed towards feature refinement. Only the image-level class labels are required in both stages. We conduct experiments to compare our approach with existing ProtoPNETs on the CUB-200-2011 [38], Stanford Cars [24], and Stanford Dogs datasets [10]. Moreover, to evaluate the diversity + +of concepts captured by ProtoPNETs, we introduce two new metrics: Distinctiveness Score, which quantifies the extent of overlap among object parts learned by prototypes within the same class, and Comprehensiveness Score, which measures whether the prototypes comprehensively attend to different parts of the foreground object. To summarize the key contributions of this work: + +- We identify a key limitation of existing ProtoPNETs, specifically generating repetitive and redundant visual explanations. To this end, we leverage non-parametric prototypes to build a self-explainable network. +- Our method comprehensively explores non-parametric prototypes and foundation vision backbones, discovering diverse and semantic distinctive concepts. +- We introduce two evaluation metrics, namely Distinctiveness Score and Comprehensiveness Score to quantify the diversity and comprehensiveness of object part concepts. We demonstrate that our method yields comparable results against state-of-the-art ProtoPNETs models on three benchmarking datasets while recognizing holistic and distinctive concepts from each category. + +# 2. Related Work + +Part-based Interpretable Image Classification. Drawing inspiration from the human ability to distinguish objects by recognizing their semantic parts, the Prototypical Part Network (ProtoPNet) [7] emulates this reasoning process by comparing image patches with learned prototype vectors that represent object parts specific to each class. Several variants have been proposed to enhance both its interpretability and classification accuracy [5, 11, 36, 46]. Notably, TesNet [40] reformulates prototype learning from the perspective of constructing embedding space on Grassmann Manifold, and also improves the overall performance by applying various priors on prototypes, such as orthonormality. Building upon the architecture of TesNet, Huang et. al [19] introduced modules to improve consistency and stability of the visual explanations, as well as corresponding metrics for quantitative evaluation. + +Meanwhile, there has been a parallel line of research that focuses on detecting object parts of fine-grained object categories [21, 37]. While these methods involve learning object part representations, they usually detect a number of shared object parts for the entire target domain and emphasize evaluating the quality of detected parts rather than leveraging them for case-based reasoning. + +The main goal of our work aims to alleviate aforementioned drawbacks of ProtoPNETs, by learning part-prototypes in a non-parametric fashion. In addition, Huang et. al. [20] integrate the reasoning process of ProtoPNETs with the series of Concept Bottleneck Models [12, 23] to improve the trustworthiness of explanations [20]. We follow this pioneering work and evaluate our framework when + +adapted to the concept learning setting. + +Foundation Vision Models. Recently, foundation vision backbones [6, 30] have been successfully applied to various challenging downstream tasks [3, 14, 17, 22], yielding remarkable results. In contrast to traditional supervised training, these backbones typically employ self-supervision or weak supervision schemes, such as masked image modeling [45], contrastive learning [8] or self distillation, reducing the reliance on labels and enhancing the quality of representation. Previous studies have demonstrated that each of these pre-training methods constructs an embedding space with different desirable properties [28, 44]. In particular, the attention map and patch tokens of self-supervised Vision Transformers trained with self-distillation have been shown to encode rich semantic information, such as scene layouts, which are beneficial for dense recognition tasks like and object discovery [34], anomaly detection [15, 16], and image generation [35, 41]. In addition, other works have leveraged their strong object part representation for fine-grained categorization [2, 13, 32]. Building upon these developments, our framework harnesses these desired properties to improve the quality of explanation provided by part-based interpretable image classifiers. + +Prototype Learning. Prototype-based classification is to make predictions based on representative examples from past data, similar to how humans approach problem solving and learning, as suggested by psychology studies [1]. Recently, it has gained significant interest in deep learning as an alternative to traditional classification heads. For example, memory bank based methods [18, 43] retain a large number of past samples to learn high-quality representations. Approaches such as ProtoSeg [51], deep nearest centroids [42] and various segmentation models [25, 29] replace conventional parametric softmax layers of segmentors with multiple non-parametric prototypes, which not only learn class distributions but also capture the intra-class variance of each category, improving the representativeness and generalizability of the models. Our non-parametric part-prototype is inspired by these approaches. + +# 3. Method + +Part-based interpretability methods such as ProtoPNETs aim to explain the predictions by generating final logits based on how much input image regions look like prototypical object parts from the training set. The core of these methods is to learn a robust set of prototypes that can provide high quality explanations. As shown in Figure 2.a, our method achieves this by deriving them as non-parametric prototypes with a learning procedure disentangled from the classification process. This results in two training stages, where the first stage is dedicated to train the prototypes and the second stage focuses on refining the feature space. + +# 3.1. Overview + +The goal of our model is to perform interpretable image classification by learning prototypes $\mathbf{P} \in \mathbb{R}^{C \times K \times D}$ , where $C$ is the number of classes and $K$ is the number of prototypes belonging to each class. Each $D$ dimensional vector represents one of the $K$ representative parts belonging to class $c \in \{c_1, c_2, \ldots\}$ . During inference, given a batch of image $\mathbf{X}_{1:B} = \{\mathbf{X}_b \in R^{H \times W \times 3}\}$ , where $B$ is the batch size, $H$ and $W$ denote the size of the input image, a ViT backbone $\mathcal{F}$ first encodes each image into a sequence of patch tokens $\mathbf{f} = \mathcal{F}(\mathbf{X}) \in \mathbb{R}^{B \times h \times w \times D}$ , reshaped into a feature map of shape $h \times w$ . Our model computes the similarity between the patch tokens $\mathbf{f}$ and part prototypes $\mathbf{P}$ , generating a set of similarity maps $\mathbf{S}_{1:B} = \{\mathrm{Sim}(\mathbf{f}_b, \mathbf{P}) \in \mathbb{R}^{h \times w \times C \times K}\}_{b=1}^B$ , where $\mathrm{Sim}(\cdot, \cdot)$ is a similarity measure, such as cosine similarity. As shown in Figure 2.a, each similarity map highlights a region of the input image similar to the learned prototypical part. Finally, the maximum value of each similarity map quantifies the occurrence of its corresponding part prototype with respect to the input, and the logit of class $c$ is computed as the weighted average of all occurrence scores of its prototypes. + +In the first stage of training, an efficient weakly-supervised clustering algorithm computes non-parametric part prototypes for each class based on extracted features from frozen backbones. Due to the well-structured latent space, these prototypes can be directly used for interpretable image classification. Then, we fix the learned part prototypes and fine-tune the backbone network by using the prototypes as anchors. This adapts the feature space to the target domain and significantly enhances the classification performance. We only use image-level class labels for supervision in both stages. + +# 3.2. Non-parametric Prototype Learning + +As shown in Figure 2.b, the main idea of this stage is to cluster the feature representation of each class involved in a batch of samples, and create a mapping between each patch token and part prototype. The empirical mean of the patch tokens assigned to the same part-prototype is directly used to update the corresponding prototypes. Hence, each latent feature patch is able to contribute to one of the part-prototypes, ensuring the overall representation of each class to cover a wide range of concepts. The only supervision involved in this stage is associating each patch token with the class label of image it belongs to, and the assignment maps are computed in an unsupervised manner. + +Specifically, we treat a batch of latent patch tokens as a bag of features, reshaping the patch tokens $\mathbf{f}$ and similarity maps $\mathbf{S}$ to $\mathbf{f} \in \mathbb{R}^{Bhw \times D}$ and $\mathbf{S} \in \mathbb{R}^{Bhw \times C \times K}$ . Next, we employ a foreground extraction module to extract patches belonging to the foreground of each sample, excluding the background patches to improve representation + +![](images/74a10a81ae87f5ab90e1df4dbfaa60a17f31d347788a83fdedcb2fa18e80fae6.jpg) + +![](images/5d97c7d10f4b541fd3eb5d80c55ffa054ab308d23308b63bc01667fb3e73b57c.jpg) +Figure 2. The architecture of our learning framework. a. Similar to prototype-based classification, we define $K$ number of part-prototypes for each category and compare them to the latent features, which yields similarity maps. Each similarity map is pooled to generate a prototype presence vector for each class. The final logits are computed as a weighted average of the presence vector. b. Non-parametric Prototype Update: We employ feature clustering to discover class-wise non-parametric part prototype from latent feature patches. Each part prototype can be regarded as the empirical mean of its corresponding part from all training examples. c. The feature space of the backbone is fine-tuned efficiently with part-prototype fixed, and by inserting new ViT blocks. + +quality. Given $N$ foreground patches of class $c$ , denoted as $\mathbf{f}^c \subset \mathbf{f}$ , $\mathbf{f}^c \in \mathbb{R}^{N \times D}$ , we obtain a similarity matrix $\mathbf{S}^c \in \mathbb{R}^{N \times K}$ between $\mathbf{f}^c$ and $\mathbf{P}^c$ by indexing $\mathbf{S}$ . We then compute a one-hot cluster assignment $\mathbf{A} \in \{0,1\}^{N \times K}$ that assigns each patch of class $c$ to one of the part-prototypes. + +In particular, we aim to optimize $\mathbf{A}^c$ so that the similarity between each patch and its assigned prototype is maximized, that is: + +$$ +\max _ {\mathbf {A} ^ {c} \in \mathcal {A} ^ {c}} \operatorname {T r} \left(\left(\mathbf {A} ^ {c}\right) ^ {\top} \mathbf {S} ^ {c}\right), \tag {1} +$$ + +where $\operatorname{Tr}(\cdot)$ is the trace of a matrix. To prevent trivial solutions for $\mathbf{A}^c$ and ensure each cluster represents a distinct object part, we enforce unique assignment and equipartition constraints [51]: + +$$ +\left(\mathbf {A} ^ {c}\right) ^ {\top} \mathbf {1} _ {N} = \frac {1}{K} \mathbf {1} _ {K}, \mathbf {A} ^ {c} \mathbf {1} _ {K} = \frac {1}{N} \mathbf {1} _ {N}. \tag {2} +$$ + +This ensures each patch token maps to a single part-prototype of class $c$ , and each prototype has an approximately equal number of patch tokens assigned. This formulation can be regarded as an Optimal Transport (OT) problem, where the approximate solution of $\mathbf{A}$ can be estimated by solving a smoothed version of equation (1): + +$$ +\mathbf {A} ^ {c, *} = \max _ {\mathbf {A} ^ {c} \in \mathcal {A} ^ {c}} \operatorname {T r} \left(\left(\mathbf {A} ^ {c}\right) ^ {\top} \mathbf {S}\right) + \kappa h \left(\mathbf {A} ^ {c}\right), \tag {3} +$$ + +where $\kappa h(\mathbf{A})$ is the entropic regularization term controlling smoothness. We use $\kappa = 0.05$ following previous works [9]. The solution can be efficiently computed using the Sinkhorn-Knopp Algorithm: + +$$ +\mathbf {A} ^ {c, *} = \operatorname {d i a g} (\mathbf {u}) \exp \left(\frac {\mathbf {S} ^ {c}}{\kappa}\right) \operatorname {d i a g} (\mathbf {v}). \tag {4} +$$ + +Momentum Update of Prototypes. Given the estimated one-hot patch-prototype assignment $\mathbf{A}^{c,*}$ , each part-prototype $\mathbf{P}_k^c$ is updated as the mean of its assigned patches: + +$$ +\mathbf {P} _ {k} ^ {c} \leftarrow \mathbf {P} _ {k} ^ {c} + (1 - \beta) \overline {{\mathbf {f}}} ^ {c, k}, \tag {5} +$$ + +where $\beta \in [0,1]$ is a coefficient that controls the update speed, and $\overline{\mathbf{f}}$ is the mean feature. This learning scheme effectively captures the mean embedding of object parts across all training examples, and can be visualized by projecting to the closest patches from the training set [7]. + +# 3.3. Adapt to Interpretable Image Classification + +As shown in Figure 2.a, the learned part-prototypes can be directly used for classification. The activation of each pro + +totype $\mathbf{P}_k^c$ with respect to the input image is calculated as: + +$$ +g _ {k} ^ {c} = \max _ {\widetilde {\mathbf {f}} \in \mathbf {f}} \quad \operatorname {S i m} \left(\widetilde {\mathbf {f}}, \mathbf {P} _ {k} ^ {c}\right) \in \mathbb {R}. \tag {6} +$$ + +As not all prototypes are equally important for classifying the input image, following previous work [19], we generate the final logit for each class $c$ by computing a weighted average of $g^{c}$ : + +$$ +\log \mathrm {i t} ^ {c} = \sum_ {k = 1} ^ {K} w _ {k} ^ {c} \cdot g _ {k} ^ {c}, \tag {7} +$$ + +where $w$ is a learned modulation vector that controls the contribution of each prototype activation towards the logit of its corresponding class. In contrast to using a fully-connected layer, computing class logits as a weighted average ensures that prototypes contribute exclusively to their respective class, thereby enhancing the interpretability. + +Prototype-Anchored Feature Space Tuning. After discovering class-wise part-prototypes from the pre-trained backbone, we fine-tune the backbone to ground the features to the learned prototypes and adapt to the target domain. To this end, we employ Block Expansion [4], as shown in Figure 2.c. Given a ViT backbone $\mathcal{F}$ with a sequence of transformer blocks $(\mathcal{F}_1,\mathcal{F}_2,\mathcal{F}_3,\ldots)$ , Block Expansion divides the blocks into $m$ groups and inserts a new identity block $\mathcal{F}_{id}$ at the end of each divided group. Each inserted identity block is initialized with weights of the preceding block, with the weights of attention projection layer and the second MLP layer set to zero, such that $\mathcal{F}_{id}(x) = x$ , i.e., the output of the backbone remains unchanged without further training. During fine-tuning, we freeze all original weights and part-prototypes, and only fine-tune the added identity blocks and modulation vectors. This approach efficiently produces features that separate classes well without disturbing the well-structured latent space of the backbone. + +Patch-Prototype Distance Optimization. In the fine-tuning stage, we use a standard cross-entropy loss $\mathcal{L}_{ce}$ to separate features of each class. In addition, we ground each image patch to its assigned part-prototype to improve feature locality. To this end, we introduce an in-class patch-prototype contrastive loss, defined as: + +$$ +\mathcal {L} _ {\mathrm {P P C}} = \sum_ {c \in \mathcal {C}} \frac {\exp \left(\left(\mathbf {f} _ {i} ^ {c}\right) ^ {\top} \mathbf {P} _ {k _ {i}} ^ {c}\right)}{\exp \left(\left(\mathbf {f} _ {i} ^ {c}\right) ^ {\top} \mathbf {P} _ {k _ {i}} ^ {c}\right) + \sum_ {\mathbf {P} _ {k _ {j}} ^ {c} \in \mathcal {P} ^ {c -}} \exp \left(\left(\mathbf {f} _ {i} ^ {c}\right) ^ {\top} \mathbf {P} _ {k _ {j}} ^ {c}\right)}, \tag {8} +$$ + +where $\mathcal{P}^{c-} = \{\mathbf{P}_k^c\}_{k=1}^K / \mathbf{P}_{k_i}^c$ . This loss "pulls" each patch towards its assigned part-prototype in the latent space, and separate it from other part-prototype within the same class. Hence, we jointly train the expanded ViT block and the modulation weights with the following objective: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\mathrm {c e}} + \lambda_ {\mathrm {P P C}} \mathcal {L} _ {\mathrm {P P C}}. \tag {9} +$$ + +# 3.4. Distinctiveness Score + +Ideally, each part-prototype belonging to the same class should learn a distinct concept from objects; otherwise, the explanation would be confusing and redundant. To quantify this property, we introduce a metric, Distinctiveness Score that measures the amount of overlap between image regions highlighted by the part prototypes associated with the ground truth class of the input image. + +Specifically, given an input image $x$ of class $c_{x}$ and the corresponding activation maps $S^{c_x} \in \mathbb{R}^{h \times w \times K}$ generated by prototypes $\mathbf{P}^{c_x}$ , we first resize the activation maps to $\hat{S}^{c_x} \in \mathbb{R}^{K \times H \times W}$ to match the size of input $x$ . Next, we extract the object region $R^{c_x, k}(x)$ highlighted by each activation map $S^{c_x, k}$ as a fix-sized bounding box with predetermined height $H_b$ and width $W_b$ around its max activation. The amount of overlap $O(x)$ between the similarity maps $S^{c_x}$ generated by ground truth class prototypes is then quantified by the Intersection over Union (IoU) between each pair of highlighted regions $R^{c_x, k}(x)$ , as follows: + +$$ +O (x) = \frac {1}{K (K - 1) / 2} \sum_ {k = 1} ^ {K - 1} \sum_ {l = k + 1} ^ {K} \operatorname {I o U} \left(R ^ {c _ {x}, k} (x), R ^ {c _ {x}, l} (x)\right). \tag {10} +$$ + +The Distinctiveness Score is computed based on the average overlap across all samples in the test set $\mathcal{X}$ , and a higher value is better. It is defined as: + +$$ +1 - \frac {1}{| \mathcal {X} |} \sum_ {x \in \mathcal {X}} O (x). \tag {11} +$$ + +# 3.5. Comprehensiveness Score + +A good Distinctiveness Score by itself does not guarantee good interpretability, as a high score could be achieved by networks with prototypes correspond to different image regions that lack meaningful concepts, e.g., background patches that contribute less to object recognition. Therefore, we further introduce Comprehensiveness Score to evaluate how thoroughly the object foreground is attended by the prototypes. Given an input image $x$ and the corresponding activation maps $S^{c_x}$ generated by prototypes $\mathbf{P}^{c_x}$ , we perform bin-max normalization on each $S^{c_x,k}$ and obtain a mask $\hat{M}_i^{c_x,k}(x)$ of the most activated region of $\mathbf{P}^{c_x,k}$ using a threshold $\tau$ : + +$$ +\tilde {S} _ {i} ^ {c _ {x}, k} = \frac {S _ {i} ^ {c _ {x} , k} - S _ {\min } ^ {c _ {x} , k}}{S _ {\max } ^ {c _ {x} , k} - S _ {\min } ^ {c _ {x} , k}}, \quad \hat {M} _ {i} ^ {c _ {x}, k} (x) = \left\{ \begin{array}{l l} 1 & \text {i f} \tilde {S} _ {i} ^ {c _ {x}, k} \geq \tau \\ 0 & \text {o t h e r w i s e .} \end{array} \right. \tag {12} +$$ + +We then compute the IoU between the union of all masks $\hat{M}^{c_x,k}(x)$ and the ground truth object foreground mask $M_{x}$ . The Comprehensiveness Score is calculated as the average IoU across all test samples $\mathcal{X}$ , and a higher value indicates + +![](images/13ad81be73f9fbc7f7ae41213ecbe6ab8362d5a123b34746d653435e883cec86.jpg) +Figure 3. a. The Distinctiveness score for one sample is calculated based on the amount of overlap between areas attended by prototypes. b. The Comprehensiveness score is computed by comparing the ground truth foreground mask $M_{x}$ with the union of thresholded image region attended by prototypes. + +better comprehensiveness. Specifically, it is defined as: + +$$ +\frac {1}{| \mathcal {X} |} \sum_ {x \in \mathcal {X}} \operatorname {I o U} \left(M _ {x}, \bigcup_ {k = 1} ^ {K} \hat {M} ^ {c _ {x}, k} (x)\right). \tag {13} +$$ + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. To evaluate the classification accuracy and interpretability of our proposed framework, we conduct experiments on CUB-200-2011 [38], Stanford Cars [24] and Stanford Dogs [10] dataset, following previous works on ProtoPNETs. The CUB-200-2011 dataset consists of 11,788 images spanning 200 bird species. Stanford Cars includes 16,185 images from 196 car models, and Stanford Dogs includes 20,580 images from 120 breeds. The results on Stanford Cars and Stanford Dogs are in supplementary. We also extend our evaluation by performing concept learning [20, 23] experiments on CUB-200-2011 to evaluate whether the learned prototypes can be adapted to produce trustworthy concept-based explanations by following the evaluation protocol introduced in [20]. + +Baselines and Evaluation Metrics. We establish our main benchmark by comparing against existing ProtoPNETs, namely ProtoPNet [7], Deformable ProtoPNet [11], ProtoPool [31], TesNet [40] and EvalProtoPNet [19]. We select three self-supervised Vision Transformer backbone as the backbone: the latest checkpoints of DINOv2 ViT-B and ViT-S trained on LVD-142M [27] and the original DINO trained on ImageNet [6]. All the networks are trained with + +image-level class labels only, and all previous methods were reproduced with official implementations. For evaluation metrics, we report top-1 accuracy for classification. We also quantitatively measure the interpretability of the learned prototypes with the introduced Distinctiveness and Comprehensiveness scores, as well as Consistency and Stability proposed in [19]. Specifically, Consistency evaluates whether each prototype correspond to the same object part across different images, while Stability measures whether each prototype correspond to the same object part when the image is perturbed with noise. We evaluate all methods with $K = 3$ and $K = 5$ . For concept learning, we compare with existing Concept Bottleneck Models [12, 20, 23], and evaluate the classification performance as well as the trustworthiness [20] of the generated concept explanations. + +Implementation Details. In the first stage of training, we set $\beta = 0.99$ , and $\kappa = 0.05$ to perform clustering. When fine-tuning the backbone, we set $\lambda_{\mathrm{PPC}} = 0.8$ and the number of added ViT block $m = 1$ . We trained our framework with 1 epoch for the first stage, and 5 more epochs for the second stage with Adam optimizer, a learning rate of 0.0001 for the backbone and $1e - 6$ for modulation vector $w$ . When training previous ProtoPNETs, we set $m = 3$ as it produces the best results across all previous architectures. + +# 4.2. Quantitative Results + +We reported the benchmark result of our method against previous ProtoPLists in Table 1. Here the Distinctiveness is calculated with $1/4$ of image size. Our method outperformed previous methods in terms of classification accuracy by $(0.16\%, 2.39\%)$ when trained with both the revised DINOv2 and DINO ViT-B backbones, and the accuracy evaluated on ViT-S is comparable to previous works. In addition, while our method yields similar Consistency to [19], it achieves higher Stability of than existing works in most settings. Nevertheless, the most notable improvement of our framework is the Distinctiveness, which surpasses previous methods by $(4.83\%, 3.49\%, 8.93\%)$ when setting 5 prototypes for each class. This indicates that our framework is more capable of utilizing the strong part representation encoded in self-supervised Vision Transformers to discover a diverse set of prototypes. Finally, the prototypes learned by our framework can be directly adapted to provide concept-based explanations, as shown by the results in Table 2. Our method yields comparable trustworthiness score with higher classification performance than [20], and Concept Bottleneck Models constructed naively on top of ViT backbones failed to produce trustworthy explanations despite their higher accuracy. + +Distinctiveness and Comprehensiveness of Prototypes. We provide a more detailed breakdown on the diversity of prototypes learned by various ProtoPNETs, by selecting a few architectures and evaluate the Distinctiveness with + +
DINOv2 ViT-B/14DINOv2 ViT-S/14DINO ViT-B/16
Con.Sta.Dis.Class.Con.Sta.Dis.Class.Con.Sta.Dis.Class.
ProtoPNet (K=3)0.7768.405.7788.474.8474.688.8384.903.0063.873.1475.94
ProtoPNet (K=5)1.3063.8411.9988.164.1270.1016.7385.523.0246.346.7978.91
Deformable ProtoPNet (K=3)0.9580.005.9086.999.6761.181.6583.091.5570.849.1571.49
Deformable ProtoPNet (K=5)2.8380.4310.3488.1411.7379.874.7685.540.9573.161.1781.74
ProtoPool (K=3)29.8352.8387.5988.7215.5047.8991.1786.1926.6777.7483.8972.04
ProtoPool (K=5)29.2757.4989.6289.9012.152.4492.2687.4137.4052.7387.6678.18
TesNet (K=3)36.3366.4541.2490.6658.0066.8927.8088.2335.5065.4058.8581.24
TesNet (K=5)58.6074.6081.9390.3567.8070.2781.1688.9551.3070.8181.9482.29
EvalProtoPNet (K=3)67.8371.3925.6789.7345.1763.5117.0187.6637.8369.3835.1084.21
EvalProtoPNet (K=5)67.8074.4072.1290.0654.6070.2777.4188.1951.0966.8672.7484.54
Ours (K=3)61.1778.9886.4490.3761.5075.5789.1389.0745.6769.9095.6285.12
Ours (K=5)66.4082.9794.4590.8266.0181.3195.7588.1647.2075.5796.5986.93
+ +![](images/577499fcd18d2655b515c07959ac7c8c01cce324a29207e1a3e96b7eb20a9ad8.jpg) +Figure 4. Distinctiveness $(\%)$ and Comprehensiveness $(\%)$ across various ProtoPLists, evaluated on different box sizes and thresholds. All methods besides [19] are trained with DINOv2 ViT-B. + +Table 1. Consistency score (%), Stability Score (%), Distinctiveness Score (%), and classification accuracy (% of various ProtoPNet on CUB-200-2011. K denotes the number of prototypes defined for each class. We train previous methods with their official implementations. + +
ViT-BViT-S
Loc.Acc.Loc.Acc.
Vanilla CBM15.6390.0612.9388.43
CBM-h11.9589.2411.4186.52
CEM13.0488.7410.6186.41
Huang et. al.45.9587.4347.3384.79
Ours (K=3)41.1887.0439.8884.55
Ours (K=5)47.0488.1844.9785.47
+ +Table 2. Concept trustworthiness (\%) and classification accuracy (\%) of a CBM constructed based on our framework, in comparison with previous methods on CUB-200-2011 dataset. All methods are trained with DINOV2 ViTs as the backbone. Note that CBM-h is vanilla CBM with a residual fitting. + +varying bounding box sizes. The results are shown in Figure 4. Notably, the prototypes learned by the vanilla ProtoPNet has poor diversity, which means that the network + +is frequently performing classification by comparing latent image patches with mostly the same object part, which hinders the interpretability. EvalProtoPNet and TesNet both yield much better scores than vanilla ProtoPNet. This phenomenon may be attributed to their orthogonality regularization. For EvalProtoPNet, the Distinctiveness is similar when trained with ViT and CNN backbones. However, our framework still yields significantly better Distinctiveness across all bounding box sizes. This indicates that prototypes of each class generated by our framework attend to a more diverse set of object parts, which result in less overlap between the explanations provided, thereby enhancing the interpretability. In the bottom of Figure 5, we also show the Comprehensiveness of architectures evaluated with varying thresholds. The original ProtoPNet fails to generate explanations that cover multiple parts of the object, resulting in very low scores. Among previous works, TesNet yields significantly better foreground coverage, which could also potentially be explained by additional regularizations enforced on the prototypes. Nevertheless, our framework is still able to recognize concepts that provide better coverage of various parts of the objects, achieving significantly higher Comprehensiveness. + +# 4.3. Qualitative Results + +We presented a comparison of visualization results across ProtoPNet architectures Figure 5, which displays the activation map of top-4 prototypes in heat maps. On the left, it can be seen that the vanilla ProtoPNet has provided poor interpretability for the two examples when trained with ViT backbones. The most activated areas are mostly outside the foreground object and scattered across the background. For ProtoPool, it has learned more prototypes that attend to various areas within the object, however it still produces some explanations that highlight the background. On the other hand, EvalProtoPNet produces better visual explanations among previous ProtoPLists as most prototypes attend + +![](images/d46be831b34fd265099a9b79450dac366773f7c4118ed129f81f9245483720f3.jpg) +Figure 5. Visualization on the same sample across various ProtoPNet architectures. All models are trained using the same DINOv2 ViT-B backbone with $K = 5$ unless annotated. + +
Con.Sta.Dis.Class.
No constraints4.0975.5049.479.25
No foreground extraction28.4082.3392.6964.29
No feature space tuning63.7483.3796.0777.96
No loss55.6177.6780.4190.78
Full Model66.4082.9796.7491.02
+ +Table 3. The effect of design choices made within our framework on the performance and interpretability of our method. + +to some different objects parts. However, it lacks diversity as many of them correspond to same object part. For each example, at least two activation map out of four highlight the same concept. This observation is consistent when EvalProtoPNet is trained with convolutional backbones, such as densenet161. In comparison, the prototypes learned by our framework is able to capture different parts of the bird, including head, wing, tail, and legs for both examples, resulting in the most conceptually divers set of explanations. + +# 4.4. Ablation Study + +In this section, we ablate various important design decisions we made in our framework. More experimental results are in the supplementary material. As shown in Table 3, all components of our framework are crucial: + +Non-parametric Prototypes. The first two rows display the importance of unique-assignment and equipartition constraints when learning non-parametric prototypes. Without these constraints, the clustering algorithm could not properly group semantically similar patch tokens into one cluster and the resulting prototypes are incapable of performing classification. Secondly, without the foreground extraction module, all the background tokens will be involved in the prototype learning stage, which significantly decrease + +the quality of the prototypes. However, when the full first stage of training is completed, that is, the feature extracted from frozen DINO ViT backbone is clustered with the constraints, the learned prototypes are already capable of recognizing objects and yielding decent interpretability. + +Feature Space Fine-tuning. Our proposed feature space tuning greatly improves the classification accuracy of the model by separating the feature space of different classes while grounding to the part prototypes of each class. However, fine-tuning the added ViT blocks without any guidance reduces the Consistency, Stability as well as Distinctiveness, due to the disturbance of the well-structured feature space of DINO backbone. This makes $\mathcal{L}_{PPC}$ crucial for the fine-tuning stage. By enforcing each latent patch to be more similar to their corresponding part-prototype, it further improves the accuracy while recover model's performance as well as overall interpretability in terms Consistency, Stability as well as Distinctiveness. + +# 5. Conclusion + +In this work, we presented a framework that provides part-based interpretability for image classification. We identify existing ProtoPNETs' limitation of generating visual explanations that are often based on repetitive object parts, which limits their explainability. We mitigate this drawback by adopting non-parametric prototypes to represent diverse and comprehensive object parts. We propose Distinctiveness and Comprehensiveness scores to quantitatively evaluate the quality of explanation provided by various ProtoPNETs, complementing the existing suite of metrics, allowing future works to perform evaluation more holistically. We envision this study to serve as a foundation for future work towards stronger part-based interpretable networks. + +# References + +[1] Agnar Aamodt and Enric Plaza. Case-based reasoning: Foundational issues, methodological variations, and system approaches. 7(1):39-59. 3 +[2] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021. 2, 3 +[3] Muhammad Awais, Muzammal Naseer, Salman Khan, et al. Foundation models defining a new era in vision: a survey and outlook. TPAMI, 2025. 3 +[4] Reza Akbarian Bafghi, Nidhin Harilal, Claire Monteleoni, and Maziar Raissi. Parameter efficient fine-tuning of self-supervised vits without catastrophic forgetting. In CVPR, pages 3679-3684, 2024. 5 +[5] Zachariah Carmichael, Suhas Lohit, Anoop Cherian, et al. Pixel-grounded prototypical part networks. In WACV, pages 4768-4779, 2024. 2 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, et al. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 2, 3, 6 +[7] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This Looks Like That: Deep Learning for Interpretable Image Recognition. In NeurIPS, 2019. 1, 2, 4, 6 +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. 3 +[9] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS. Curran Associates, Inc. 4 +[10] E Dataset. Novel datasets for fine-grained image categorization. In CVPR Workshop, page 2. Citeseer, 2011. 2, 6 +[11] Jon Donnelly, Alina Jade Barnett, and Chaofan Chen. Deformable ProtoPNet: An interpretable image classifier using deformable prototypes. In CVPR, pages 10265-10275, 2022. 1, 2, 6 +[12] Mateo Espinosa Zarlenga, Pietro Barbiero, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. 35:21400-21413, 2022. 1, 2, 6 +[13] Lei Fan, Yiwen Ding, Dongdong Fan, et al. Grainspace: A large-scale dataset for fine-grained and domain-adaptive recognition of cereal grains. In CVPR, pages 21116-21125, 2022. 3 +[14] Lei Fan, Arcot Sowmya, Erik Meijering, and Yang Song. Cancer survival prediction from whole slide images with self-supervised learning and slide consistency. IEEE TMI, 42(5):1401-1412, 2022. 3 +[15] Lei Fan, Yiwen Ding, Dongdong Fan, et al. Identifying the defective: Detecting damaged grains for cereal appearance inspection. In ECAI, pages 660-667. IOS Press, 2023. 3 +[16] Lei Fan, Dongdong Fan, Yiwen Ding, et al. Av4gainsp: An efficient dual-camera system for identifying defective kernels of cereal grains. IEEE Robotics and Automation Letters, 9(1):851-858, 2023. 3 +[17] Lei Fan, Dongdong Fan, Yiwen Ding, et al. Grainbrain: Multiview identification and stratification of defective grain kernels. IEEE TII, 2025. 3 + +[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738. 3 +[19] Qihan Huang, Mengqi Xue, Wenqi Huang, Haofei Zhang, Jie Song, Yongcheng Jing, and Mingli Song. Evaluation and improvement of interpretability for self-explainable part-prototype networks. In ICCV, pages 2011–2020, 2023. 1, 2, 5, 6, 7 +[20] Qihan Huang, Jie Song, Jingwen Hu, et al. On the Concept Trustworthiness in Concept Bottleneck Models. AAAI, 38 (19):21161-21168, 2024. 2, 6 +[21] Zixuan Huang and Yin Li. Interpretable and accurate fine-grained recognition via region grouping. In CVPR, pages 8662-8672. 1, 2 +[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, et al. Segment anything. In ICCV, pages 4015-4026. 3 +[23] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In ICML, pages 5338-5348. PMLR, 2020. 1, 2, 6 +[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshops, pages 554–561, 2013. 2, 6 +[25] Yawen Lu, Dongfang Liu, Qifan Wang, Cheng Han, Yiming Cui, Zhiwen Cao, Xueling Zhang, Yingjie Victor Chen, and Heng Fan. ProMotion: Prototypes as motion learners. In CVPR, pages 28109-28119, 2024. 3 +[26] Tanvir Mahmud, Burhaneddin Yaman, Chun-Hao Liu, and Diana Marculescu. Papr: Training-free one-step patch pruning with lightweight convnets for faster inference. In ECCV, pages 110-128. Springer, 2024. 1, 2 +[27] Maxime Oquab, Timothee Darcet, Theo Moutakanni, et al. DINOv2: Learning Robust Visual Features without Supervision. TMLR, 2024. 2, 6, 1 +[28] Namuk Park, Wonjae Kim, Byeongho Heo, Taekyung Kim, and Sangdoo Yun. What do self-supervised vision transformers learn? In ICLR. 2, 3 +[29] Zheyun Qin, Cheng Han, Qifan Wang, Xiushan Nie, Yilong Yin, and Lu Xiankai. Unified 3d segmenter as prototypical classifiers. NeurIPS, 36:46419-46432, 2023. 3 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR. 3 +[31] Dawid Rymarczyk, Łukasz Struski, Michal Górszczak, et al. Interpretable image classification with differentiable prototypes assignment. In ECCV, pages 351-368, 2022. 6 +[32] Oindrila Saha and Subhransu Maji. PARTICLE: Part discovery and contrastive learning for fine-grained recognition. In ICCV, pages 167-176. 3 +[33] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In ICCV, pages 618-626. 1 +[34] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC, 2021. 3 + +[35] Zhenhong Sun, Junyan Wang, Zhiyu Tan, Daoyi Dong, Hailan Ma, Hao Li, and Dong Gong. EGGen: Image Generation with Multi-entity Prior Learning through Entity Guidance. In ACM MM, pages 6637-6645, 2024. 3 +[36] Andong Tan, Fengtao Zhou, and Hao Chen. Post-hoc part-prototype networks. In ICML, 2024. 2 +[37] Robert van der Klis, Stephan Alaniz, Massimiliano Mancini, Cassio F. Dantas, Dino Ienco, Zeynep Akata, and Diego Marcos. PDiscoNet: Semantically consistent part discovery for fine-grained recognition. In ICCV, pages 1866-1876. 1, 2 +[38] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 2, 6 +[39] Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In CVPR, pages 10962-10971, 2023. 1 +[40] Jiaqi Wang, Huafeng Liu, Xinyue Wang, and Liping Jing. Interpretable image recognition by constructing transparent embedding space. In ICCV, pages 895-904, 2021. 1, 2, 6 +[41] Junyan Wang, Zhenhong Sun, Zhiyu Tan, Xuanbai Chen, Weihua Chen, Hao Li, Cheng Zhang, and Yang Song. Towards effective usage of human-centric priors in diffusion models for text-based human image generation. In CVPR, pages 8446-8455, 2024. 3 +[42] Wenguan Wang, Cheng Han, Tianfei Zhou, and Dongfang Liu. Visual recognition with deep nearest centroids. In ICLR. 2, 3 +[43] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742. 3 +[44] Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. In CVPR, pages 14475-14485, .3 +[45] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. SimMIM: A simple framework for masked image modeling. In CVPR, pages 9653-9663, .3 +[46] Mengqi Xue, Qihan Huang, Haofei Zhang, Jingwen Hu, Jie Song, Mingli Song, and Canghong Jin. Protopformer: concentrating on prototypical parts in vision transformers for interpretable image recognition. In IJCAI, pages 1516-1524, 2024. 2, 1 +[47] Boyu Yang, Chang Liu, Bohao Li, Jianbin Jiao, and Qixiang Ye. Prototype mixture models for few-shot semantic segmentation. In ECCV, pages 763-778. Springer International Publishing. 2 +[48] Rui Zhang, Xingbo Du, Junchi Yan, and Shihua Zhang. The decoupling concept bottleneck model. TPAMI, 2024. 1 +[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, pages 2921-2929, .1 +[50] Tianfei Zhou and Wenguan Wang. Prototype-based semantic segmentation. TPAMI, 2024. 2 +[51] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, pages 2582-2593, 2, 3, 4 \ No newline at end of file diff --git a/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/images.zip b/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f2c125aaa3b40b652e2cafb1074af11d802af8d7 --- /dev/null +++ b/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2acf61d92de8f863a4be3aaf425115a07303ad86bc64621f7c51092e2117ec09 +size 607430 diff --git a/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/layout.json b/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7cf2d282f394ced554cc57ba5f51d46546269288 --- /dev/null +++ b/2025/Interpretable Image Classification via Non-parametric Part Prototype Learning/layout.json @@ -0,0 +1,7961 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "text", + "content": "Interpretable Image Classification via Non-parametric Part Prototype Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 156, + 142, + 453, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 142, + 453, + 171 + ], + "spans": [ + { + "bbox": [ + 156, + 142, + 453, + 171 + ], + "type": "text", + "content": "Zhijie Zhu1* Lei Fan1* Maurice Pagnucco1 Yang Song1 \n1University of New South Wales" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 173, + 457, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 173, + 457, + 185 + ], + "spans": [ + { + "bbox": [ + 149, + 173, + 457, + 185 + ], + "type": "text", + "content": "{zhijie.zhu@,lei.fanl@morri@cse,yang.song1@}.unsw.edu.au" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 239, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 296, + 539 + ], + "type": "text", + "content": "Classifying images with an interpretable decision-making process is a long-standing problem in computer vision. In recent years, Prototypical Part Networks has gained traction as an approach for self-explainable neural networks, due to their ability to mimic human visual reasoning by providing explanations based on prototypical object parts. However, the quality of the explanations generated by these methods leaves room for improvement, as the prototypes usually focus on repetitive and redundant concepts. Leveraging recent advances in prototype learning, we present a framework for part-based interpretable image classification that learns a set of semantically distinctive object parts for each class, and provides diverse and comprehensive explanations. The core of our method is to learn the part-prototypes in a non-parametric fashion, through clustering deep features extracted from foundation vision models that encode robust semantic information. To quantitatively evaluate the quality of explanations provided by ProtoPNETs, we introduce Distinctiveness Score and Comprehensiveness Score. Through evaluation on CUB-200-2011, Stanford Cars and Stanford Dogs datasets, we show that our framework compares favourably against existing ProtoPNETs while achieving better interpretability. Code is available at: https://github.com/zijizhu/proton-param." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 563, + 135, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 563, + 135, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 563, + 135, + 575 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 584, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 296, + 693 + ], + "type": "text", + "content": "In recent years, self-explainable networks have emerged as a promising approach for improving the transparency of deep neural networks by providing inherent interpretability. Unlike traditional explanation methods that operate post-hoc [33, 49], recent approaches aim to mimic human reasoning by justifying predictions through an interpretable decision-making process. Examples include Prototypical-Part Networks (ProtoPNETs) [7, 11], part discovery methods [21, 37], and Concept Bottleneck Models [12, 23, 39, 48]," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 217, + 436, + 300 + ], + "blocks": [ + { + "bbox": [ + 319, + 217, + 436, + 300 + ], + "lines": [ + { + "bbox": [ + 319, + 217, + 436, + 300 + ], + "spans": [ + { + "bbox": [ + 319, + 217, + 436, + 300 + ], + "type": "image", + "image_path": "b76c8f6738c9c2e49fc63bea9457e7d288ab9f8a5330b1afa68a4d58c4ece5fe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 443, + 217, + 557, + 300 + ], + "blocks": [ + { + "bbox": [ + 443, + 217, + 557, + 300 + ], + "lines": [ + { + "bbox": [ + 443, + 217, + 557, + 300 + ], + "spans": [ + { + "bbox": [ + 443, + 217, + 557, + 300 + ], + "type": "image", + "image_path": "d86ba279cc30b7e1c3220cb1f89e69dfea207784207b11cdca0a41c7d6ca6a86.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 304, + 435, + 434 + ], + "blocks": [ + { + "bbox": [ + 319, + 304, + 435, + 434 + ], + "lines": [ + { + "bbox": [ + 319, + 304, + 435, + 434 + ], + "spans": [ + { + "bbox": [ + 319, + 304, + 435, + 434 + ], + "type": "image", + "image_path": "5aa1bf7f2442eee595fdaf9328da5409d8cf1294f40150b68047dde94a6edb7f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 449, + 555, + 538 + ], + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 538 + ], + "type": "text", + "content": "Figure 1. a. Existing methods [19, 40] encode various assumptions as regularizations to guide prototype learning, but often fail to diversify the concepts learned by each prototype. Here multiple part-prototypes attend to the same image region, thereby limiting their interpretability. In contrast b. Our method partitions the feature space into semantically distinct clusters, and updates each prototype with the empirical mean of its respective cluster, enabling each prototype to learn semantically different concepts." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 443, + 304, + 557, + 434 + ], + "blocks": [ + { + "bbox": [ + 443, + 304, + 557, + 434 + ], + "lines": [ + { + "bbox": [ + 443, + 304, + 557, + 434 + ], + "spans": [ + { + "bbox": [ + 443, + 304, + 557, + 434 + ], + "type": "image", + "image_path": "093e69ec04edc2deb830aef71f7ccbbeecf627885c5fab593b65c9debac02573.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 555, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 651 + ], + "type": "text", + "content": "which offer explanations derived from object concepts. Among these, ProtoPNETs have garnered attention as they provide intuitive visual explanations rooted in past examples. Specifically, they define a number of prototypes for each object category, each representing a distinguishable object part, and make the prediction based on the similarity of an input image to prototypes, thereby emulating the process of human visual reasoning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "However, existing ProtoPNETs train prototypes as learnable parameters, and guide their learning process with the core assumption that a region within each input image should match one of the prototypes from the corresponding category. However, we observe that multiple prototypes" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 134, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9762" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "from the same class often correspond to the same object part, which hinders the quality of explanations provided by the network in two ways: First, explanations become noncomprehensive, covering only a subset of object parts. Second, this results in redundant and repetitive explanations that fail to adequately illustrate why the object is classified in a particular way. An example is shown in Figure 1.a, out of the four most activated prototypes in a ProtoPNet, three of them focus on overlapping regions near the bird's wing, ignoring other parts such as the bird's legs and tail. Ideally, within an interpretability framework, the model is expected to recognize and differentiate a comprehensive set of distinctive parts for each object category, thus enhancing the clarity and transparency of its decision-making process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 243, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 296, + 495 + ], + "type": "text", + "content": "To approach this problem, we take inspiration from recent advances in prototype-based classification [47, 50], and encode distinct object parts using non-parametric prototypes [42, 51], as shown in Figure 1.b. Non-parametric prototypes were originally proposed to learn a more robust feature space by capturing the rich variance within each class through multiple prototypes [51]. We extend this idea to part-based interpretability to address both of the aforementioned limitations. Specifically, in comparison to previous methods that guide the learning process of learnable prototypes with regularizations [40], non-parametric prototypes allow finer control over the representation of each prototype within the same class by applying constraints on feature clustering. As a consequence, each prototype is able to capture unique characteristics of an object that differentiate it from others, resulting in semantically distinct prototypes, each latent feature patch of the foreground object is able to contribute to one of the prototypes, allowing the overall representation for each class to embody a comprehensive set of associated concepts. These properties greatly improve the interpretability of the explanations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": "Additionally, in order to obtain robust feature representations that encode object part semantics, We explore advanced foundation models as backbones to extract features, specifically utilizing self-supervised Vision Transformers pre-trained with self-distillation (DINO-ViTs) [6, 27], given that these backbone networks capture rich semantic information about object parts and can be used out-of-the-box [2, 28]. We also propose an efficient fine-tuning strategy for self-supervised ViT backbones to adapt the feature space to the target domain while grounding the embedding of object parts to its assigned prototypes. Hence, our framework consists of two training stages where the initial stage focuses on prototype discovery, and the subsequent stage is directed towards feature refinement. Only the image-level class labels are required in both stages. We conduct experiments to compare our approach with existing ProtoPNETs on the CUB-200-2011 [38], Stanford Cars [24], and Stanford Dogs datasets [10]. Moreover, to evaluate the diversity" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "content": "of concepts captured by ProtoPNETs, we introduce two new metrics: Distinctiveness Score, which quantifies the extent of overlap among object parts learned by prototypes within the same class, and Comprehensiveness Score, which measures whether the prototypes comprehensively attend to different parts of the foreground object. To summarize the key contributions of this work:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 156, + 553, + 324 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 156, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 156, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 553, + 204 + ], + "type": "text", + "content": "- We identify a key limitation of existing ProtoPNETs, specifically generating repetitive and redundant visual explanations. To this end, we leverage non-parametric prototypes to build a self-explainable network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 204, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 239 + ], + "type": "text", + "content": "- Our method comprehensively explores non-parametric prototypes and foundation vision backbones, discovering diverse and semantic distinctive concepts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 239, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 324 + ], + "type": "text", + "content": "- We introduce two evaluation metrics, namely Distinctiveness Score and Comprehensiveness Score to quantify the diversity and comprehensiveness of object part concepts. We demonstrate that our method yields comparable results against state-of-the-art ProtoPNETs models on three benchmarking datasets while recognizing holistic and distinctive concepts from each category." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 334, + 400, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 334, + 400, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 334, + 400, + 346 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 354, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 545 + ], + "type": "text", + "content": "Part-based Interpretable Image Classification. Drawing inspiration from the human ability to distinguish objects by recognizing their semantic parts, the Prototypical Part Network (ProtoPNet) [7] emulates this reasoning process by comparing image patches with learned prototype vectors that represent object parts specific to each class. Several variants have been proposed to enhance both its interpretability and classification accuracy [5, 11, 36, 46]. Notably, TesNet [40] reformulates prototype learning from the perspective of constructing embedding space on Grassmann Manifold, and also improves the overall performance by applying various priors on prototypes, such as orthonormality. Building upon the architecture of TesNet, Huang et. al [19] introduced modules to improve consistency and stability of the visual explanations, as well as corresponding metrics for quantitative evaluation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 629 + ], + "type": "text", + "content": "Meanwhile, there has been a parallel line of research that focuses on detecting object parts of fine-grained object categories [21, 37]. While these methods involve learning object part representations, they usually detect a number of shared object parts for the entire target domain and emphasize evaluating the quality of detected parts rather than leveraging them for case-based reasoning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "The main goal of our work aims to alleviate aforementioned drawbacks of ProtoPNETs, by learning part-prototypes in a non-parametric fashion. In addition, Huang et. al. [20] integrate the reasoning process of ProtoPNETs with the series of Concept Bottleneck Models [12, 23] to improve the trustworthiness of explanations [20]. We follow this pioneering work and evaluate our framework when" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9763" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 215, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 215, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 215, + 84 + ], + "type": "text", + "content": "adapted to the concept learning setting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 297, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 297, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 297, + 350 + ], + "type": "text", + "content": "Foundation Vision Models. Recently, foundation vision backbones [6, 30] have been successfully applied to various challenging downstream tasks [3, 14, 17, 22], yielding remarkable results. In contrast to traditional supervised training, these backbones typically employ self-supervision or weak supervision schemes, such as masked image modeling [45], contrastive learning [8] or self distillation, reducing the reliance on labels and enhancing the quality of representation. Previous studies have demonstrated that each of these pre-training methods constructs an embedding space with different desirable properties [28, 44]. In particular, the attention map and patch tokens of self-supervised Vision Transformers trained with self-distillation have been shown to encode rich semantic information, such as scene layouts, which are beneficial for dense recognition tasks like and object discovery [34], anomaly detection [15, 16], and image generation [35, 41]. In addition, other works have leveraged their strong object part representation for fine-grained categorization [2, 13, 32]. Building upon these developments, our framework harnesses these desired properties to improve the quality of explanation provided by part-based interpretable image classifiers." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 351, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 544 + ], + "type": "text", + "content": "Prototype Learning. Prototype-based classification is to make predictions based on representative examples from past data, similar to how humans approach problem solving and learning, as suggested by psychology studies [1]. Recently, it has gained significant interest in deep learning as an alternative to traditional classification heads. For example, memory bank based methods [18, 43] retain a large number of past samples to learn high-quality representations. Approaches such as ProtoSeg [51], deep nearest centroids [42] and various segmentation models [25, 29] replace conventional parametric softmax layers of segmentors with multiple non-parametric prototypes, which not only learn class distributions but also capture the intra-class variance of each category, improving the representativeness and generalizability of the models. Our non-parametric part-prototype is inspired by these approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 559, + 111, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 111, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 111, + 572 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "Part-based interpretability methods such as ProtoPNETs aim to explain the predictions by generating final logits based on how much input image regions look like prototypical object parts from the training set. The core of these methods is to learn a robust set of prototypes that can provide high quality explanations. As shown in Figure 2.a, our method achieves this by deriving them as non-parametric prototypes with a learning procedure disentangled from the classification process. This results in two training stages, where the first stage is dedicated to train the prototypes and the second stage focuses on refining the feature space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 382, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 382, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 382, + 83 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": "The goal of our model is to perform interpretable image classification by learning prototypes " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{C \\times K \\times D}" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " is the number of classes and " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " is the number of prototypes belonging to each class. Each " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " dimensional vector represents one of the " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " representative parts belonging to class " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "c \\in \\{c_1, c_2, \\ldots\\}" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": ". During inference, given a batch of image " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{1:B} = \\{\\mathbf{X}_b \\in R^{H \\times W \\times 3}\\}" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " is the batch size, " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " denote the size of the input image, a ViT backbone " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " first encodes each image into a sequence of patch tokens " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{f} = \\mathcal{F}(\\mathbf{X}) \\in \\mathbb{R}^{B \\times h \\times w \\times D}" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": ", reshaped into a feature map of shape " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": ". Our model computes the similarity between the patch tokens " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " and part prototypes " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": ", generating a set of similarity maps " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{1:B} = \\{\\mathrm{Sim}(\\mathbf{f}_b, \\mathbf{P}) \\in \\mathbb{R}^{h \\times w \\times C \\times K}\\}_{b=1}^B" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathrm{Sim}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " is a similarity measure, such as cosine similarity. As shown in Figure 2.a, each similarity map highlights a region of the input image similar to the learned prototypical part. Finally, the maximum value of each similarity map quantifies the occurrence of its corresponding part prototype with respect to the input, and the logit of class " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " is computed as the weighted average of all occurrence scores of its prototypes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": "In the first stage of training, an efficient weakly-supervised clustering algorithm computes non-parametric part prototypes for each class based on extracted features from frozen backbones. Due to the well-structured latent space, these prototypes can be directly used for interpretable image classification. Then, we fix the learned part prototypes and fine-tune the backbone network by using the prototypes as anchors. This adapts the feature space to the target domain and significantly enhances the classification performance. We only use image-level class labels for supervision in both stages." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 479, + 509, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 509, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 509, + 493 + ], + "type": "text", + "content": "3.2. Non-parametric Prototype Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 498, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 641 + ], + "type": "text", + "content": "As shown in Figure 2.b, the main idea of this stage is to cluster the feature representation of each class involved in a batch of samples, and create a mapping between each patch token and part prototype. The empirical mean of the patch tokens assigned to the same part-prototype is directly used to update the corresponding prototypes. Hence, each latent feature patch is able to contribute to one of the part-prototypes, ensuring the overall representation of each class to cover a wide range of concepts. The only supervision involved in this stage is associating each patch token with the class label of image it belongs to, and the assignment maps are computed in an unsupervised manner." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Specifically, we treat a batch of latent patch tokens as a bag of features, reshaping the patch tokens " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " and similarity maps " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathbb{R}^{Bhw \\times D}" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{Bhw \\times C \\times K}" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": ". Next, we employ a foreground extraction module to extract patches belonging to the foreground of each sample, excluding the background patches to improve representation" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9764" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 535, + 203 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 535, + 203 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 535, + 203 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 535, + 203 + ], + "type": "image", + "image_path": "74a10a81ae87f5ab90e1df4dbfaa60a17f31d347788a83fdedcb2fa18e80fae6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 78, + 208, + 541, + 329 + ], + "blocks": [ + { + "bbox": [ + 78, + 208, + 541, + 329 + ], + "lines": [ + { + "bbox": [ + 78, + 208, + 541, + 329 + ], + "spans": [ + { + "bbox": [ + 78, + 208, + 541, + 329 + ], + "type": "image", + "image_path": "5d97c7d10f4b541fd3eb5d80c55ffa054ab308d23308b63bc01667fb3e73b57c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 340, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 54, + 340, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 340, + 555, + 407 + ], + "type": "text", + "content": "Figure 2. The architecture of our learning framework. a. Similar to prototype-based classification, we define " + }, + { + "bbox": [ + 54, + 340, + 555, + 407 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 340, + 555, + 407 + ], + "type": "text", + "content": " number of part-prototypes for each category and compare them to the latent features, which yields similarity maps. Each similarity map is pooled to generate a prototype presence vector for each class. The final logits are computed as a weighted average of the presence vector. b. Non-parametric Prototype Update: We employ feature clustering to discover class-wise non-parametric part prototype from latent feature patches. Each part prototype can be regarded as the empirical mean of its corresponding part from all training examples. c. The feature space of the backbone is fine-tuned efficiently with part-prototype fixed, and by inserting new ViT blocks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "content": "quality. Given " + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "content": " foreground patches of class " + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^c \\subset \\mathbf{f}" + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^c \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "content": ", we obtain a similarity matrix " + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^c \\in \\mathbb{R}^{N \\times K}" + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^c" + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^c" + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "content": " by indexing " + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "content": ". We then compute a one-hot cluster assignment " + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\{0,1\\}^{N \\times K}" + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "content": " that assigns each patch of class " + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "content": " to one of the part-prototypes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 514 + ], + "type": "text", + "content": "In particular, we aim to optimize " + }, + { + "bbox": [ + 55, + 479, + 295, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^c" + }, + { + "bbox": [ + 55, + 479, + 295, + 514 + ], + "type": "text", + "content": " so that the similarity between each patch and its assigned prototype is maximized, that is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 527, + 295, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 295, + 546 + ], + "type": "interline_equation", + "content": "\\max _ {\\mathbf {A} ^ {c} \\in \\mathcal {A} ^ {c}} \\operatorname {T r} \\left(\\left(\\mathbf {A} ^ {c}\\right) ^ {\\top} \\mathbf {S} ^ {c}\\right), \\tag {1}", + "image_path": "9b3f5f3943d33dd686953b92aa77f9fc90796b8c0ef80431eef3a9f5fbf1cba7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\operatorname{Tr}(\\cdot)" + }, + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "text", + "content": " is the trace of a matrix. To prevent trivial solutions for " + }, + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^c" + }, + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "text", + "content": " and ensure each cluster represents a distinct object part, we enforce unique assignment and equipartition constraints [51]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 612, + 295, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 612, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 99, + 612, + 295, + 635 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {A} ^ {c}\\right) ^ {\\top} \\mathbf {1} _ {N} = \\frac {1}{K} \\mathbf {1} _ {K}, \\mathbf {A} ^ {c} \\mathbf {1} _ {K} = \\frac {1}{N} \\mathbf {1} _ {N}. \\tag {2}", + "image_path": "0d717fa891ecd852d91e02f5f3de84c64ffe72dfe11a00a084ce0b4d7e935cf7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "This ensures each patch token maps to a single part-prototype of class " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": ", and each prototype has an approximately equal number of patch tokens assigned. This formulation can be regarded as an Optimal Transport (OT) problem, where the approximate solution of " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " can be estimated by solving a smoothed version of equation (1):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 352, + 429, + 553, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 429, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 352, + 429, + 553, + 447 + ], + "type": "interline_equation", + "content": "\\mathbf {A} ^ {c, *} = \\max _ {\\mathbf {A} ^ {c} \\in \\mathcal {A} ^ {c}} \\operatorname {T r} \\left(\\left(\\mathbf {A} ^ {c}\\right) ^ {\\top} \\mathbf {S}\\right) + \\kappa h \\left(\\mathbf {A} ^ {c}\\right), \\tag {3}", + "image_path": "4577e46d73f4e375cf0e85edf44465f306c7f12040550cc28257fe6968f05a6c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 451, + 555, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 451, + 555, + 499 + ], + "type": "inline_equation", + "content": "\\kappa h(\\mathbf{A})" + }, + { + "bbox": [ + 313, + 451, + 555, + 499 + ], + "type": "text", + "content": " is the entropic regularization term controlling smoothness. We use " + }, + { + "bbox": [ + 313, + 451, + 555, + 499 + ], + "type": "inline_equation", + "content": "\\kappa = 0.05" + }, + { + "bbox": [ + 313, + 451, + 555, + 499 + ], + "type": "text", + "content": " following previous works [9]. The solution can be efficiently computed using the Sinkhorn-Knopp Algorithm:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 507, + 553, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 507, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 354, + 507, + 553, + 534 + ], + "type": "interline_equation", + "content": "\\mathbf {A} ^ {c, *} = \\operatorname {d i a g} (\\mathbf {u}) \\exp \\left(\\frac {\\mathbf {S} ^ {c}}{\\kappa}\\right) \\operatorname {d i a g} (\\mathbf {v}). \\tag {4}", + "image_path": "bea5f59224a835ac2c823ed94f18e4c5e84f06c1fe4526214eea8ad14d17cbdf.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 536, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 574 + ], + "type": "text", + "content": "Momentum Update of Prototypes. Given the estimated one-hot patch-prototype assignment " + }, + { + "bbox": [ + 313, + 536, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{c,*}" + }, + { + "bbox": [ + 313, + 536, + 555, + 574 + ], + "type": "text", + "content": ", each part-prototype " + }, + { + "bbox": [ + 313, + 536, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k^c" + }, + { + "bbox": [ + 313, + 536, + 555, + 574 + ], + "type": "text", + "content": " is updated as the mean of its assigned patches:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 381, + 581, + 553, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 581, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 381, + 581, + 553, + 597 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {k} ^ {c} \\leftarrow \\mathbf {P} _ {k} ^ {c} + (1 - \\beta) \\overline {{\\mathbf {f}}} ^ {c, k}, \\tag {5}", + "image_path": "f9ffe4e813c3dc62447ce54475f5fbd0ce6a461b125e2bd543d24f2c1f697374.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\beta \\in [0,1]" + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "text", + "content": " is a coefficient that controls the update speed, and " + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{f}}" + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "text", + "content": " is the mean feature. This learning scheme effectively captures the mean embedding of object parts across all training examples, and can be visualized by projecting to the closest patches from the training set [7]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 671, + 541, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 541, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 541, + 684 + ], + "type": "text", + "content": "3.3. Adapt to Interpretable Image Classification" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "As shown in Figure 2.a, the learned part-prototypes can be directly used for classification. The activation of each pro" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9765" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 290, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 290, + 86 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 290, + 86 + ], + "type": "text", + "content": "totype " + }, + { + "bbox": [ + 55, + 72, + 290, + 86 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k^c" + }, + { + "bbox": [ + 55, + 72, + 290, + 86 + ], + "type": "text", + "content": " with respect to the input image is calculated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 94, + 295, + 116 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 94, + 295, + 116 + ], + "spans": [ + { + "bbox": [ + 113, + 94, + 295, + 116 + ], + "type": "interline_equation", + "content": "g _ {k} ^ {c} = \\max _ {\\widetilde {\\mathbf {f}} \\in \\mathbf {f}} \\quad \\operatorname {S i m} \\left(\\widetilde {\\mathbf {f}}, \\mathbf {P} _ {k} ^ {c}\\right) \\in \\mathbb {R}. \\tag {6}", + "image_path": "328fea69f4bdb40b70f45cea710a2eb9cd26be3ccad7d7e547a31ca6c1b53c75.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 119, + 295, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 119, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 119, + 295, + 167 + ], + "type": "text", + "content": "As not all prototypes are equally important for classifying the input image, following previous work [19], we generate the final logit for each class " + }, + { + "bbox": [ + 55, + 119, + 295, + 167 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 119, + 295, + 167 + ], + "type": "text", + "content": " by computing a weighted average of " + }, + { + "bbox": [ + 55, + 119, + 295, + 167 + ], + "type": "inline_equation", + "content": "g^{c}" + }, + { + "bbox": [ + 55, + 119, + 295, + 167 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 175, + 295, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 175, + 295, + 208 + ], + "spans": [ + { + "bbox": [ + 131, + 175, + 295, + 208 + ], + "type": "interline_equation", + "content": "\\log \\mathrm {i t} ^ {c} = \\sum_ {k = 1} ^ {K} w _ {k} ^ {c} \\cdot g _ {k} ^ {c}, \\tag {7}", + "image_path": "75907760ed8ae3f306d4f4f33749835583cb1c34d224464bfb87a156d11c276b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 212, + 296, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 212, + 296, + 282 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 212, + 296, + 282 + ], + "type": "text", + "content": " is a learned modulation vector that controls the contribution of each prototype activation towards the logit of its corresponding class. In contrast to using a fully-connected layer, computing class logits as a weighted average ensures that prototypes contribute exclusively to their respective class, thereby enhancing the interpretability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "content": "Prototype-Anchored Feature Space Tuning. After discovering class-wise part-prototypes from the pre-trained backbone, we fine-tune the backbone to ground the features to the learned prototypes and adapt to the target domain. To this end, we employ Block Expansion [4], as shown in Figure 2.c. Given a ViT backbone " + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "content": " with a sequence of transformer blocks " + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_1,\\mathcal{F}_2,\\mathcal{F}_3,\\ldots)" + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "content": ", Block Expansion divides the blocks into " + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "content": " groups and inserts a new identity block " + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{id}" + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "content": " at the end of each divided group. Each inserted identity block is initialized with weights of the preceding block, with the weights of attention projection layer and the second MLP layer set to zero, such that " + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{id}(x) = x" + }, + { + "bbox": [ + 55, + 283, + 296, + 498 + ], + "type": "text", + "content": ", i.e., the output of the backbone remains unchanged without further training. During fine-tuning, we freeze all original weights and part-prototypes, and only fine-tune the added identity blocks and modulation vectors. This approach efficiently produces features that separate classes well without disturbing the well-structured latent space of the backbone." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 571 + ], + "type": "text", + "content": "Patch-Prototype Distance Optimization. In the fine-tuning stage, we use a standard cross-entropy loss " + }, + { + "bbox": [ + 55, + 498, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 55, + 498, + 295, + 571 + ], + "type": "text", + "content": " to separate features of each class. In addition, we ground each image patch to its assigned part-prototype to improve feature locality. To this end, we introduce an in-class patch-prototype contrastive loss, defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 295, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 628 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P P C}} = \\sum_ {c \\in \\mathcal {C}} \\frac {\\exp \\left(\\left(\\mathbf {f} _ {i} ^ {c}\\right) ^ {\\top} \\mathbf {P} _ {k _ {i}} ^ {c}\\right)}{\\exp \\left(\\left(\\mathbf {f} _ {i} ^ {c}\\right) ^ {\\top} \\mathbf {P} _ {k _ {i}} ^ {c}\\right) + \\sum_ {\\mathbf {P} _ {k _ {j}} ^ {c} \\in \\mathcal {P} ^ {c -}} \\exp \\left(\\left(\\mathbf {f} _ {i} ^ {c}\\right) ^ {\\top} \\mathbf {P} _ {k _ {j}} ^ {c}\\right)}, \\tag {8}", + "image_path": "7b70f8c4a3cac229236c6191c40972403dd0dddc4029e4207e10d2c347567806.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{c-} = \\{\\mathbf{P}_k^c\\}_{k=1}^K / \\mathbf{P}_{k_i}^c" + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": ". This loss \"pulls\" each patch towards its assigned part-prototype in the latent space, and separate it from other part-prototype within the same class. Hence, we jointly train the expanded ViT block and the modulation weights with the following objective:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 123, + 701, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 123, + 701, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {c e}} + \\lambda_ {\\mathrm {P P C}} \\mathcal {L} _ {\\mathrm {P P C}}. \\tag {9}", + "image_path": "01224e8804604551d54d5057f201e95aae71b5f7a478e454c6663eb37830213c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 435, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 435, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 435, + 83 + ], + "type": "text", + "content": "3.4. Distinctiveness Score" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 91, + 553, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 175 + ], + "type": "text", + "content": "Ideally, each part-prototype belonging to the same class should learn a distinct concept from objects; otherwise, the explanation would be confusing and redundant. To quantify this property, we introduce a metric, Distinctiveness Score that measures the amount of overlap between image regions highlighted by the part prototypes associated with the ground truth class of the input image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": "Specifically, given an input image " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "c_{x}" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": " and the corresponding activation maps " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "S^{c_x} \\in \\mathbb{R}^{h \\times w \\times K}" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": " generated by prototypes " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{c_x}" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": ", we first resize the activation maps to " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "\\hat{S}^{c_x} \\in \\mathbb{R}^{K \\times H \\times W}" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": " to match the size of input " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": ". Next, we extract the object region " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "R^{c_x, k}(x)" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": " highlighted by each activation map " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "S^{c_x, k}" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": " as a fix-sized bounding box with predetermined height " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "H_b" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "W_b" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": " around its max activation. The amount of overlap " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "O(x)" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": " between the similarity maps " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "S^{c_x}" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": " generated by ground truth class prototypes is then quantified by the Intersection over Union (IoU) between each pair of highlighted regions " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "R^{c_x, k}(x)" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 553, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 362 + ], + "type": "interline_equation", + "content": "O (x) = \\frac {1}{K (K - 1) / 2} \\sum_ {k = 1} ^ {K - 1} \\sum_ {l = k + 1} ^ {K} \\operatorname {I o U} \\left(R ^ {c _ {x}, k} (x), R ^ {c _ {x}, l} (x)\\right). \\tag {10}", + "image_path": "f9a8cb7fc9b8fa97de68b6752f40e3667b8980fc6de8c776cc43c85d37becbc0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 364, + 554, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 554, + 399 + ], + "type": "text", + "content": "The Distinctiveness Score is computed based on the average overlap across all samples in the test set " + }, + { + "bbox": [ + 313, + 364, + 554, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 364, + 554, + 399 + ], + "type": "text", + "content": ", and a higher value is better. It is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 393, + 410, + 553, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 410, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 393, + 410, + 553, + 438 + ], + "type": "interline_equation", + "content": "1 - \\frac {1}{| \\mathcal {X} |} \\sum_ {x \\in \\mathcal {X}} O (x). \\tag {11}", + "image_path": "2381c40e7b2dd2a44c6afb56af16e535bd216dd47a916382435532228181c5fd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 450, + 458, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 450, + 458, + 463 + ], + "spans": [ + { + "bbox": [ + 314, + 450, + 458, + 463 + ], + "type": "text", + "content": "3.5. Comprehensiveness Score" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "text", + "content": "A good Distinctiveness Score by itself does not guarantee good interpretability, as a high score could be achieved by networks with prototypes correspond to different image regions that lack meaningful concepts, e.g., background patches that contribute less to object recognition. Therefore, we further introduce Comprehensiveness Score to evaluate how thoroughly the object foreground is attended by the prototypes. Given an input image " + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "text", + "content": " and the corresponding activation maps " + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "inline_equation", + "content": "S^{c_x}" + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "text", + "content": " generated by prototypes " + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{c_x}" + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "text", + "content": ", we perform bin-max normalization on each " + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "inline_equation", + "content": "S^{c_x,k}" + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "text", + "content": " and obtain a mask " + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\hat{M}_i^{c_x,k}(x)" + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "text", + "content": " of the most activated region of " + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{c_x,k}" + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "text", + "content": " using a threshold " + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 470, + 555, + 613 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 622, + 553, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 622, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 315, + 622, + 553, + 664 + ], + "type": "interline_equation", + "content": "\\tilde {S} _ {i} ^ {c _ {x}, k} = \\frac {S _ {i} ^ {c _ {x} , k} - S _ {\\min } ^ {c _ {x} , k}}{S _ {\\max } ^ {c _ {x} , k} - S _ {\\min } ^ {c _ {x} , k}}, \\quad \\hat {M} _ {i} ^ {c _ {x}, k} (x) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\tilde {S} _ {i} ^ {c _ {x}, k} \\geq \\tau \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {12}", + "image_path": "f30f71a8150e0a4c149467f077c1d8a06bd700ab8100d93baebbab46f41cb232.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": "We then compute the IoU between the union of all masks " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{M}^{c_x,k}(x)" + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": " and the ground truth object foreground mask " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "M_{x}" + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": ". The Comprehensiveness Score is calculated as the average IoU across all test samples " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": ", and a higher value indicates" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9766" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 78, + 284, + 250 + ], + "blocks": [ + { + "bbox": [ + 61, + 78, + 284, + 250 + ], + "lines": [ + { + "bbox": [ + 61, + 78, + 284, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 78, + 284, + 250 + ], + "type": "image", + "image_path": "13ad81be73f9fbc7f7ae41213ecbe6ab8362d5a123b34746d653435e883cec86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 259, + 297, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 297, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 297, + 316 + ], + "type": "text", + "content": "Figure 3. a. The Distinctiveness score for one sample is calculated based on the amount of overlap between areas attended by prototypes. b. The Comprehensiveness score is computed by comparing the ground truth foreground mask " + }, + { + "bbox": [ + 55, + 259, + 297, + 316 + ], + "type": "inline_equation", + "content": "M_{x}" + }, + { + "bbox": [ + 55, + 259, + 297, + 316 + ], + "type": "text", + "content": " with the union of thresholded image region attended by prototypes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 336, + 279, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 279, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 279, + 350 + ], + "type": "text", + "content": "better comprehensiveness. Specifically, it is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 358, + 295, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 358, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 295, + 392 + ], + "type": "interline_equation", + "content": "\\frac {1}{| \\mathcal {X} |} \\sum_ {x \\in \\mathcal {X}} \\operatorname {I o U} \\left(M _ {x}, \\bigcup_ {k = 1} ^ {K} \\hat {M} ^ {c _ {x}, k} (x)\\right). \\tag {13}", + "image_path": "9ad07bb33e1144d1f6cabb3e3d236b7bc3c40749badc3ae8e75a456c104d23c6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 411, + 137, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 137, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 137, + 425 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 431, + 183, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 183, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 183, + 445 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": "Datasets. To evaluate the classification accuracy and interpretability of our proposed framework, we conduct experiments on CUB-200-2011 [38], Stanford Cars [24] and Stanford Dogs [10] dataset, following previous works on ProtoPNETs. The CUB-200-2011 dataset consists of 11,788 images spanning 200 bird species. Stanford Cars includes 16,185 images from 196 car models, and Stanford Dogs includes 20,580 images from 120 breeds. The results on Stanford Cars and Stanford Dogs are in supplementary. We also extend our evaluation by performing concept learning [20, 23] experiments on CUB-200-2011 to evaluate whether the learned prototypes can be adapted to produce trustworthy concept-based explanations by following the evaluation protocol introduced in [20]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Baselines and Evaluation Metrics. We establish our main benchmark by comparing against existing ProtoPNETs, namely ProtoPNet [7], Deformable ProtoPNet [11], ProtoPool [31], TesNet [40] and EvalProtoPNet [19]. We select three self-supervised Vision Transformer backbone as the backbone: the latest checkpoints of DINOv2 ViT-B and ViT-S trained on LVD-142M [27] and the original DINO trained on ImageNet [6]. All the networks are trained with" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": "image-level class labels only, and all previous methods were reproduced with official implementations. For evaluation metrics, we report top-1 accuracy for classification. We also quantitatively measure the interpretability of the learned prototypes with the introduced Distinctiveness and Comprehensiveness scores, as well as Consistency and Stability proposed in [19]. Specifically, Consistency evaluates whether each prototype correspond to the same object part across different images, while Stability measures whether each prototype correspond to the same object part when the image is perturbed with noise. We evaluate all methods with " + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": ". For concept learning, we compare with existing Concept Bottleneck Models [12, 20, 23], and evaluate the classification performance as well as the trustworthiness [20] of the generated concept explanations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "text", + "content": "Implementation Details. In the first stage of training, we set " + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\beta = 0.99" + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\kappa = 0.05" + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "text", + "content": " to perform clustering. When fine-tuning the backbone, we set " + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{PPC}} = 0.8" + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "text", + "content": " and the number of added ViT block " + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "inline_equation", + "content": "m = 1" + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "text", + "content": ". We trained our framework with 1 epoch for the first stage, and 5 more epochs for the second stage with Adam optimizer, a learning rate of 0.0001 for the backbone and " + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "inline_equation", + "content": "1e - 6" + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "text", + "content": " for modulation vector " + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "text", + "content": ". When training previous ProtoPNETs, we set " + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "inline_equation", + "content": "m = 3" + }, + { + "bbox": [ + 313, + 252, + 555, + 361 + ], + "type": "text", + "content": " as it produces the best results across all previous architectures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 370, + 433, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 433, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 433, + 383 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "text", + "content": "We reported the benchmark result of our method against previous ProtoPLists in Table 1. Here the Distinctiveness is calculated with " + }, + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "text", + "content": " of image size. Our method outperformed previous methods in terms of classification accuracy by " + }, + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "inline_equation", + "content": "(0.16\\%, 2.39\\%)" + }, + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "text", + "content": " when trained with both the revised DINOv2 and DINO ViT-B backbones, and the accuracy evaluated on ViT-S is comparable to previous works. In addition, while our method yields similar Consistency to [19], it achieves higher Stability of than existing works in most settings. Nevertheless, the most notable improvement of our framework is the Distinctiveness, which surpasses previous methods by " + }, + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "inline_equation", + "content": "(4.83\\%, 3.49\\%, 8.93\\%)" + }, + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "text", + "content": " when setting 5 prototypes for each class. This indicates that our framework is more capable of utilizing the strong part representation encoded in self-supervised Vision Transformers to discover a diverse set of prototypes. Finally, the prototypes learned by our framework can be directly adapted to provide concept-based explanations, as shown by the results in Table 2. Our method yields comparable trustworthiness score with higher classification performance than [20], and Concept Bottleneck Models constructed naively on top of ViT backbones failed to produce trustworthy explanations despite their higher accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Distinctiveness and Comprehensiveness of Prototypes. We provide a more detailed breakdown on the diversity of prototypes learned by various ProtoPNETs, by selecting a few architectures and evaluate the Distinctiveness with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9767" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 69, + 553, + 224 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 553, + 224 + ], + "type": "table", + "html": "
DINOv2 ViT-B/14DINOv2 ViT-S/14DINO ViT-B/16
Con.Sta.Dis.Class.Con.Sta.Dis.Class.Con.Sta.Dis.Class.
ProtoPNet (K=3)0.7768.405.7788.474.8474.688.8384.903.0063.873.1475.94
ProtoPNet (K=5)1.3063.8411.9988.164.1270.1016.7385.523.0246.346.7978.91
Deformable ProtoPNet (K=3)0.9580.005.9086.999.6761.181.6583.091.5570.849.1571.49
Deformable ProtoPNet (K=5)2.8380.4310.3488.1411.7379.874.7685.540.9573.161.1781.74
ProtoPool (K=3)29.8352.8387.5988.7215.5047.8991.1786.1926.6777.7483.8972.04
ProtoPool (K=5)29.2757.4989.6289.9012.152.4492.2687.4137.4052.7387.6678.18
TesNet (K=3)36.3366.4541.2490.6658.0066.8927.8088.2335.5065.4058.8581.24
TesNet (K=5)58.6074.6081.9390.3567.8070.2781.1688.9551.3070.8181.9482.29
EvalProtoPNet (K=3)67.8371.3925.6789.7345.1763.5117.0187.6637.8369.3835.1084.21
EvalProtoPNet (K=5)67.8074.4072.1290.0654.6070.2777.4188.1951.0966.8672.7484.54
Ours (K=3)61.1778.9886.4490.3761.5075.5789.1389.0745.6769.9095.6285.12
Ours (K=5)66.4082.9794.4590.8266.0181.3195.7588.1647.2075.5796.5986.93
", + "image_path": "3e71633d47d4e87d6ddfa5a7bfb68ba4f912f7f0bfa06164d2a44accbc247710.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 270, + 294, + 459 + ], + "blocks": [ + { + "bbox": [ + 56, + 270, + 294, + 459 + ], + "lines": [ + { + "bbox": [ + 56, + 270, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 270, + 294, + 459 + ], + "type": "image", + "image_path": "577499fcd18d2655b515c07959ac7c8c01cce324a29207e1a3e96b7eb20a9ad8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 468, + 295, + 502 + ], + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 502 + ], + "type": "text", + "content": "Figure 4. Distinctiveness " + }, + { + "bbox": [ + 55, + 468, + 295, + 502 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 468, + 295, + 502 + ], + "type": "text", + "content": " and Comprehensiveness " + }, + { + "bbox": [ + 55, + 468, + 295, + 502 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 468, + 295, + 502 + ], + "type": "text", + "content": " across various ProtoPLists, evaluated on different box sizes and thresholds. All methods besides [19] are trained with DINOv2 ViT-B." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 83, + 516, + 269, + 604 + ], + "blocks": [ + { + "bbox": [ + 55, + 232, + 555, + 256 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 256 + ], + "type": "text", + "content": "Table 1. Consistency score (%), Stability Score (%), Distinctiveness Score (%), and classification accuracy (% of various ProtoPNet on CUB-200-2011. K denotes the number of prototypes defined for each class. We train previous methods with their official implementations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 516, + 269, + 604 + ], + "lines": [ + { + "bbox": [ + 83, + 516, + 269, + 604 + ], + "spans": [ + { + "bbox": [ + 83, + 516, + 269, + 604 + ], + "type": "table", + "html": "
ViT-BViT-S
Loc.Acc.Loc.Acc.
Vanilla CBM15.6390.0612.9388.43
CBM-h11.9589.2411.4186.52
CEM13.0488.7410.6186.41
Huang et. al.45.9587.4347.3384.79
Ours (K=3)41.1887.0439.8884.55
Ours (K=5)47.0488.1844.9785.47
", + "image_path": "f67d7aa74ff1b4873d7361c3728392f071da1e7ea06e7c0d31cfeab60e307be2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 612, + 295, + 668 + ], + "lines": [ + { + "bbox": [ + 55, + 612, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 295, + 668 + ], + "type": "text", + "content": "Table 2. Concept trustworthiness (\\%) and classification accuracy (\\%) of a CBM constructed based on our framework, in comparison with previous methods on CUB-200-2011 dataset. All methods are trained with DINOV2 ViTs as the backbone. Note that CBM-h is vanilla CBM with a residual fitting." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "varying bounding box sizes. The results are shown in Figure 4. Notably, the prototypes learned by the vanilla ProtoPNet has poor diversity, which means that the network" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 270, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 545 + ], + "type": "text", + "content": "is frequently performing classification by comparing latent image patches with mostly the same object part, which hinders the interpretability. EvalProtoPNet and TesNet both yield much better scores than vanilla ProtoPNet. This phenomenon may be attributed to their orthogonality regularization. For EvalProtoPNet, the Distinctiveness is similar when trained with ViT and CNN backbones. However, our framework still yields significantly better Distinctiveness across all bounding box sizes. This indicates that prototypes of each class generated by our framework attend to a more diverse set of object parts, which result in less overlap between the explanations provided, thereby enhancing the interpretability. In the bottom of Figure 5, we also show the Comprehensiveness of architectures evaluated with varying thresholds. The original ProtoPNet fails to generate explanations that cover multiple parts of the object, resulting in very low scores. Among previous works, TesNet yields significantly better foreground coverage, which could also potentially be explained by additional regularizations enforced on the prototypes. Nevertheless, our framework is still able to recognize concepts that provide better coverage of various parts of the objects, achieving significantly higher Comprehensiveness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 552, + 425, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 425, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 425, + 564 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We presented a comparison of visualization results across ProtoPNet architectures Figure 5, which displays the activation map of top-4 prototypes in heat maps. On the left, it can be seen that the vanilla ProtoPNet has provided poor interpretability for the two examples when trained with ViT backbones. The most activated areas are mostly outside the foreground object and scattered across the background. For ProtoPool, it has learned more prototypes that attend to various areas within the object, however it still produces some explanations that highlight the background. On the other hand, EvalProtoPNet produces better visual explanations among previous ProtoPLists as most prototypes attend" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9768" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 553, + 257 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 553, + 257 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 553, + 257 + ], + "type": "image", + "image_path": "d46be831b34fd265099a9b79450dac366773f7c4118ed129f81f9245483720f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 265, + 555, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 265, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 555, + 288 + ], + "type": "text", + "content": "Figure 5. Visualization on the same sample across various ProtoPNet architectures. All models are trained using the same DINOv2 ViT-B backbone with " + }, + { + "bbox": [ + 55, + 265, + 555, + 288 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 55, + 265, + 555, + 288 + ], + "type": "text", + "content": " unless annotated." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 306, + 294, + 376 + ], + "blocks": [ + { + "bbox": [ + 58, + 306, + 294, + 376 + ], + "lines": [ + { + "bbox": [ + 58, + 306, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 58, + 306, + 294, + 376 + ], + "type": "table", + "html": "
Con.Sta.Dis.Class.
No constraints4.0975.5049.479.25
No foreground extraction28.4082.3392.6964.29
No feature space tuning63.7483.3796.0777.96
No loss55.6177.6780.4190.78
Full Model66.4082.9796.7491.02
", + "image_path": "d094fa642a1e393e70e45bfc9b5e7585ee50b0258d0cee8679df6cd55f090690.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 382, + 295, + 406 + ], + "lines": [ + { + "bbox": [ + 55, + 382, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 295, + 406 + ], + "type": "text", + "content": "Table 3. The effect of design choices made within our framework on the performance and interpretability of our method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "text", + "content": "to some different objects parts. However, it lacks diversity as many of them correspond to same object part. For each example, at least two activation map out of four highlight the same concept. This observation is consistent when EvalProtoPNet is trained with convolutional backbones, such as densenet161. In comparison, the prototypes learned by our framework is able to capture different parts of the bird, including head, wing, tail, and legs for both examples, resulting in the most conceptually divers set of explanations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 537, + 149, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 149, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 149, + 550 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 556, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 604 + ], + "type": "text", + "content": "In this section, we ablate various important design decisions we made in our framework. More experimental results are in the supplementary material. As shown in Table 3, all components of our framework are crucial:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Non-parametric Prototypes. The first two rows display the importance of unique-assignment and equipartition constraints when learning non-parametric prototypes. Without these constraints, the clustering algorithm could not properly group semantically similar patch tokens into one cluster and the resulting prototypes are incapable of performing classification. Secondly, without the foreground extraction module, all the background tokens will be involved in the prototype learning stage, which significantly decrease" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 308, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 369 + ], + "type": "text", + "content": "the quality of the prototypes. However, when the full first stage of training is completed, that is, the feature extracted from frozen DINO ViT backbone is clustered with the constraints, the learned prototypes are already capable of recognizing objects and yielding decent interpretability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 369, + 556, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 556, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 556, + 524 + ], + "type": "text", + "content": "Feature Space Fine-tuning. Our proposed feature space tuning greatly improves the classification accuracy of the model by separating the feature space of different classes while grounding to the part prototypes of each class. However, fine-tuning the added ViT blocks without any guidance reduces the Consistency, Stability as well as Distinctiveness, due to the disturbance of the well-structured feature space of DINO backbone. This makes " + }, + { + "bbox": [ + 313, + 369, + 556, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PPC}" + }, + { + "bbox": [ + 313, + 369, + 556, + 524 + ], + "type": "text", + "content": " crucial for the fine-tuning stage. By enforcing each latent patch to be more similar to their corresponding part-prototype, it further improves the accuracy while recover model's performance as well as overall interpretability in terms Consistency, Stability as well as Distinctiveness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "In this work, we presented a framework that provides part-based interpretability for image classification. We identify existing ProtoPNETs' limitation of generating visual explanations that are often based on repetitive object parts, which limits their explainability. We mitigate this drawback by adopting non-parametric prototypes to represent diverse and comprehensive object parts. We propose Distinctiveness and Comprehensiveness scores to quantitatively evaluate the quality of explanation provided by various ProtoPNETs, complementing the existing suite of metrics, allowing future works to perform evaluation more holistically. We envision this study to serve as a foundation for future work towards stronger part-based interpretable networks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9769" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Agnar Aamodt and Enric Plaza. Case-based reasoning: Foundational issues, methodological variations, and system approaches. 7(1):39-59. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "text", + "content": "[2] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "type": "text", + "content": "[3] Muhammad Awais, Muzammal Naseer, Salman Khan, et al. Foundation models defining a new era in vision: a survey and outlook. TPAMI, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 294, + 237 + ], + "type": "text", + "content": "[4] Reza Akbarian Bafghi, Nidhin Harilal, Claire Monteleoni, and Maziar Raissi. Parameter efficient fine-tuning of self-supervised vits without catastrophic forgetting. In CVPR, pages 3679-3684, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 270 + ], + "type": "text", + "content": "[5] Zachariah Carmichael, Suhas Lohit, Anoop Cherian, et al. Pixel-grounded prototypical part networks. In WACV, pages 4768-4779, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 273, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 294, + 304 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, et al. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "text", + "content": "[7] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This Looks Like That: Deep Learning for Interpretable Image Recognition. In NeurIPS, 2019. 1, 2, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 352, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 294, + 384 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 386, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 294, + 407 + ], + "type": "text", + "content": "[9] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS. Curran Associates, Inc. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 409, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 409, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 57, + 409, + 294, + 430 + ], + "type": "text", + "content": "[10] E Dataset. Novel datasets for fine-grained image categorization. In CVPR Workshop, page 2. Citeseer, 2011. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 432, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 432, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 57, + 432, + 294, + 475 + ], + "type": "text", + "content": "[11] Jon Donnelly, Alina Jade Barnett, and Chaofan Chen. Deformable ProtoPNet: An interpretable image classifier using deformable prototypes. In CVPR, pages 10265-10275, 2022. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 477, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 477, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 294, + 509 + ], + "type": "text", + "content": "[12] Mateo Espinosa Zarlenga, Pietro Barbiero, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. 35:21400-21413, 2022. 1, 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 511, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 511, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 511, + 294, + 553 + ], + "type": "text", + "content": "[13] Lei Fan, Yiwen Ding, Dongdong Fan, et al. Grainspace: A large-scale dataset for fine-grained and domain-adaptive recognition of cereal grains. In CVPR, pages 21116-21125, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 556, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 294, + 599 + ], + "type": "text", + "content": "[14] Lei Fan, Arcot Sowmya, Erik Meijering, and Yang Song. Cancer survival prediction from whole slide images with self-supervised learning and slide consistency. IEEE TMI, 42(5):1401-1412, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "type": "text", + "content": "[15] Lei Fan, Yiwen Ding, Dongdong Fan, et al. Identifying the defective: Detecting damaged grains for cereal appearance inspection. In ECAI, pages 660-667. IOS Press, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "text", + "content": "[16] Lei Fan, Dongdong Fan, Yiwen Ding, et al. Av4gainsp: An efficient dual-camera system for identifying defective kernels of cereal grains. IEEE Robotics and Automation Letters, 9(1):851-858, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[17] Lei Fan, Dongdong Fan, Yiwen Ding, et al. Grainbrain: Multiview identification and stratification of defective grain kernels. IEEE TII, 2025. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "type": "text", + "content": "[19] Qihan Huang, Mengqi Xue, Wenqi Huang, Haofei Zhang, Jie Song, Yongcheng Jing, and Mingli Song. Evaluation and improvement of interpretability for self-explainable part-prototype networks. In ICCV, pages 2011–2020, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 195 + ], + "type": "text", + "content": "[20] Qihan Huang, Jie Song, Jingwen Hu, et al. On the Concept Trustworthiness in Concept Bottleneck Models. AAAI, 38 (19):21161-21168, 2024. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 198, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 228 + ], + "type": "text", + "content": "[21] Zixuan Huang and Yin Li. Interpretable and accurate fine-grained recognition via region grouping. In CVPR, pages 8662-8672. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "type": "text", + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, et al. Segment anything. In ICCV, pages 4015-4026. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "text", + "content": "[23] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In ICML, pages 5338-5348. PMLR, 2020. 1, 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "type": "text", + "content": "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshops, pages 554–561, 2013. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 333, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 376 + ], + "type": "text", + "content": "[25] Yawen Lu, Dongfang Liu, Qifan Wang, Cheng Han, Yiming Cui, Zhiwen Cao, Xueling Zhang, Yingjie Victor Chen, and Heng Fan. ProMotion: Prototypes as motion learners. In CVPR, pages 28109-28119, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 377, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 421 + ], + "type": "text", + "content": "[26] Tanvir Mahmud, Burhaneddin Yaman, Chun-Hao Liu, and Diana Marculescu. Papr: Training-free one-step patch pruning with lightweight convnets for faster inference. In ECCV, pages 110-128. Springer, 2024. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "text", + "content": "[27] Maxime Oquab, Timothee Darcet, Theo Moutakanni, et al. DINOv2: Learning Robust Visual Features without Supervision. TMLR, 2024. 2, 6, 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "text", + "content": "[28] Namuk Park, Wonjae Kim, Byeongho Heo, Taekyung Kim, and Sangdoo Yun. What do self-supervised vision transformers learn? In ICLR. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "type": "text", + "content": "[29] Zheyun Qin, Cheng Han, Qifan Wang, Xiushan Nie, Yilong Yin, and Lu Xiankai. Unified 3d segmenter as prototypical classifiers. NeurIPS, 36:46419-46432, 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "type": "text", + "content": "[31] Dawid Rymarczyk, Łukasz Struski, Michal Górszczak, et al. Interpretable image classification with differentiable prototypes assignment. In ECCV, pages 351-368, 2022. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 591, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 624 + ], + "type": "text", + "content": "[32] Oindrila Saha and Subhransu Maji. PARTICLE: Part discovery and contrastive learning for fine-grained recognition. In ICCV, pages 167-176. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[33] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In ICCV, pages 618-626. 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[34] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC, 2021. 3" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9770" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 702 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[35] Zhenhong Sun, Junyan Wang, Zhiyu Tan, Daoyi Dong, Hailan Ma, Hao Li, and Dong Gong. EGGen: Image Generation with Multi-entity Prior Learning through Entity Guidance. In ACM MM, pages 6637-6645, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "type": "text", + "content": "[36] Andong Tan, Fengtao Zhou, and Hao Chen. Post-hoc part-prototype networks. In ICML, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[37] Robert van der Klis, Stephan Alaniz, Massimiliano Mancini, Cassio F. Dantas, Dino Ienco, Zeynep Akata, and Diego Marcos. PDiscoNet: Semantically consistent part discovery for fine-grained recognition. In ICCV, pages 1866-1876. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 217 + ], + "type": "text", + "content": "[38] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "text", + "content": "[39] Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In CVPR, pages 10962-10971, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "text", + "content": "[40] Jiaqi Wang, Huafeng Liu, Xinyue Wang, and Liping Jing. Interpretable image recognition by constructing transparent embedding space. In ICCV, pages 895-904, 2021. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 288, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 341 + ], + "type": "text", + "content": "[41] Junyan Wang, Zhenhong Sun, Zhiyu Tan, Xuanbai Chen, Weihua Chen, Hao Li, Cheng Zhang, and Yang Song. Towards effective usage of human-centric priors in diffusion models for text-based human image generation. In CVPR, pages 8446-8455, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "type": "text", + "content": "[42] Wenguan Wang, Cheng Han, Tianfei Zhou, and Dongfang Liu. Visual recognition with deep nearest centroids. In ICLR. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 377, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 409 + ], + "type": "text", + "content": "[43] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "text", + "content": "[44] Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. In CVPR, pages 14475-14485, .3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[45] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. SimMIM: A simple framework for masked image modeling. In CVPR, pages 9653-9663, .3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 490, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 543 + ], + "type": "text", + "content": "[46] Mengqi Xue, Qihan Huang, Haofei Zhang, Jingwen Hu, Jie Song, Mingli Song, and Canghong Jin. Protopformer: concentrating on prototypical parts in vision transformers for interpretable image recognition. In IJCAI, pages 1516-1524, 2024. 2, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "text", + "content": "[47] Boyu Yang, Chang Liu, Bohao Li, Jianbin Jiao, and Qixiang Ye. Prototype mixture models for few-shot semantic segmentation. In ECCV, pages 763-778. Springer International Publishing. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 590, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 611 + ], + "type": "text", + "content": "[48] Rui Zhang, Xingbo Du, Junchi Yan, and Shihua Zhang. The decoupling concept bottleneck model. TPAMI, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 613, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 645 + ], + "type": "text", + "content": "[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In CVPR, pages 2921-2929, .1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 647, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 668 + ], + "type": "text", + "content": "[50] Tianfei Zhou and Wenguan Wang. Prototype-based semantic segmentation. TPAMI, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 670, + 294, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 702 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 702 + ], + "type": "text", + "content": "[51] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, pages 2582-2593, 2, 3, 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "9771" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Interpreting Object-level Foundation Models via Visual Precision Search/396da971-babe-4c65-8e62-f7e2e9a705a9_content_list.json b/2025/Interpreting Object-level Foundation Models via Visual Precision Search/396da971-babe-4c65-8e62-f7e2e9a705a9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..22c96794aad7c05b994986b99f6905a74773226c --- /dev/null +++ b/2025/Interpreting Object-level Foundation Models via Visual Precision Search/396da971-babe-4c65-8e62-f7e2e9a705a9_content_list.json @@ -0,0 +1,1941 @@ +[ + { + "type": "text", + "text": "Interpreting Object-level Foundation Models via Visual Precision Search", + "text_level": 1, + "bbox": [ + 133, + 130, + 866, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruoyu Chen $^{1,2}$ , Siyuan Liang $^{3}$ , Jingzhi Li $^{1,2,7}$ , Shiming Liu $^{4}$ , Maosen Li $^{5}$ , Zhen Huang $^{6}$ , Hua Zhang $^{1,2,*}$ , and Xiaochun Cao $^{8,*}$", + "bbox": [ + 205, + 159, + 790, + 196 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100093, China", + "$^{2}$ School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China" + ], + "bbox": [ + 184, + 196, + 812, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Computing, NUS $^{4}$ RAMS Lab, Huawei Technologies Co., Ltd. $^{5}$ IAS BU, Huawei Technologies Co., Ltd.", + "bbox": [ + 102, + 232, + 893, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "6College of Computer, NUDT 7Key Lab. of Edu. Inf. for Nationalities (YNNU), Ministry of Education, Kunming, China", + "bbox": [ + 102, + 250, + 895, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{8}$ School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen 518107, China", + "bbox": [ + 122, + 267, + 875, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chenruoyu@ie.ac.cn pandaliang521@gmail.com {lijingzhi,zhanghua}@ie.ac.cn", + "bbox": [ + 137, + 287, + 859, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{liushiming3, limaosen2}@huawei.com huangzhen@nudt.edu.cn caoxiaochun@mail.sysu.edu.cn", + "bbox": [ + 109, + 304, + 883, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 323, + 326, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advances in multimodal pre-training have propelled object-level foundation models, such as Grounding DINO and Florence-2, in tasks like visual grounding and object detection. However, interpreting these models' decisions has grown increasingly challenging. Existing interpretable attribution methods for object-level task interpretation have notable limitations: (1) gradient-based methods lack precise localization due to visual-textual fusion in foundation models, and (2) perturbation-based methods produce noisy saliency maps, limiting fine-grained interpretability. To address these, we propose a Visual Precision Search method that generates accurate attribution maps with fewer regions. Our method bypasses internal model parameters to overcome attribution issues from multimodal fusion, dividing inputs into sparse sub-regions and using consistency and collaboration scores to accurately identify critical decision-making regions. We also conducted a theoretical analysis of the boundary guarantees and scope of applicability of our method. Experiments on RefCOCO, MS COCO, and LVIS show our approach enhances object-level task interpretability over SOTA for Grounding DINO and Florence-2 across various evaluation metrics, with faithfulness gains of $23.7\\%$ , $31.6\\%$ , and $20.1\\%$ on MS COCO, LVIS, and RefCOCO for Grounding DINO, and $50.7\\%$ and $66.9\\%$ on MS COCO and RefCOCO for Florence-2. Additionally, our method can interpret failures in visual grounding and object detection tasks, surpassing existing methods across multiple evaluation metrics. The code is released at https://github.com/RuoyuChen10/VPS.", + "bbox": [ + 89, + 356, + 483, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 803, + 220, + 818 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding object information in images, such as object detection [7, 14, 24, 51, 52, 57], is a crucial and enduring challenge in computer vision, holding significant im", + "bbox": [ + 89, + 829, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3fb944160443eb79d76d1aa87654ccef39bb1ca0bdc30adeef9458a556cb074f.jpg", + "image_caption": [ + "Figure 1. Illustration of our Visual Precision Search interpretation method, which more precisely identifies key sub-regions in object-level foundation model decision-making compared to gradient-based and perturbation-based methods." + ], + "image_footnote": [], + "bbox": [ + 526, + 324, + 887, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "portance across various fields, including autonomous driving [4, 15, 44]. With advancements in multimodal alignment technology [20, 34], object-level foundation models [28, 46, 48, 56] like Grounding DINO [28] and Florence-2 [48] have been developed to handle various tasks, including visual grounding and object detection. However, the massive data volume, rich semantic concepts, and large parameters significantly reduce the transparency [21, 22, 26] and interpretability [5, 11, 55] of these models. Since object-level tasks like autonomous driving demand high model reliability [9, 30, 45], building transparent and interpretable models is crucial for enhancing safety and trustworthy [6, 17, 21, 23, 24, 43], making it essential to interpret object-level foundation models.", + "bbox": [ + 511, + 656, + 906, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some interpretable attribution algorithms for traditional object detection models have been proposed, including", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30042", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gradient-based methods like ODAM [54] and perturbation-based methods like D-RISE [33]. However, as shown in the lower panel of Figure 1, these methods face limitations when applied to multimodal foundation models: (1) gradient-based methods may struggle to provide accurate visual localization explanations due to the fusion of visual and textual features in foundation models, and (2) perturbation-based approaches can introduce sampling artifacts, resulting in noisy saliency maps with limited fine-grained interpretability.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose a novel interpretation mechanism for the object-level foundation model, as shown in Figure 1. Our goal is to generate a saliency map that explains the rationale behind the model's detection of a specific object, allowing fewer regions to be exposed for accurate detection. Additionally, removing these critical regions should quickly lead to detection failure, highlighting their importance in the model's decision. Specifically, we propose Visual Precision Search, which sparsifies the input region into a series of sub-regions using superpixel segmentation, and then ranks these sparse sub-regions to determine their importance for object-level decision-making. Our gradient-free method circumvents localization errors at the visual level that are caused by gradient backpropagation in vision-text fusion. Furthermore, we propose a novel submodular function, grounded in theoretical analysis and offering boundary guarantees for Visual Precision Search. This function identifies regions that enhance interpretability from two key perspectives: clues that support the model's accurate detection and regions with strong combinatorial effects. By employing region search to iteratively expand the set of sub-regions, our method more precisely identifies key sub-regions in the decision-making of object-level foundation models, which mitigates the issue of coarse explanations caused by noise in perturbation-based methods. In addition, our method effectively analyzes cases of grounding or detection failure, allowing us to observe which input-level factors influence the decision.", + "bbox": [ + 91, + 247, + 483, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validated our method on the MS COCO [25], RefCOCO [18], and LVIS [13] datasets. The object-level foundation models evaluated include Grounding DINO, which uses a multimodal feature fusion architecture, and Florence2, which employs a multimodal large language model architecture. The tasks covered include interpreting visual grounding and object detection. In explaining the data behind correct model decisions, our method demonstrates that the faithfulness metric for Grounding DINO surpasses SOTA method D-RISE by $23.7\\%$ , $20.1\\%$ , and $31.6\\%$ on MS COCO, RefCOCO, and LVIS, respectively. For Florence2, our method improves the faithfulness metric by $102.9\\%$ and $66.9\\%$ on MS COCO and RefCOCO, respectively, while also achieving SOTA performance on location metrics across all datasets. In interpreting the factors leading to", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Grounding DINO's failures in visual grounding tasks, our method improves the Insertion metric by $42.9\\%$ and the average highest score by $25.1\\%$ . It also identifies causes of detector misclassification and undetection, achieving Insertion metric improvements of $54.7\\%$ and $42.7\\%$ for misclassification, and $36.7\\%$ and $64.3\\%$ for undetection on MS COCO and LVIS, respectively.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of this paper are:", + "bbox": [ + 529, + 198, + 846, + 210 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce Visual Precision Search, a new search-based mechanism for interpreting object-level foundation models through instance-specific saliency maps.", + "- A novel submodular mechanism is constructed to enhance interpretability from two aspects: identifying clues that support accurate detection and highlighting regions with strong combinatorial effects, with an analysis of its theoretical boundaries for object-level tasks.", + "- We demonstrate our approach's generalizability across multimodal foundation models, including both non-LLM-based (Grounding DINO) and LLM-based detectors (Florence-2).", + "- We validate our method on MS COCO, RefCOCO, and LVIS, achieving significant improvements in explaining object-level tasks. Additionally, we analyze grounding and detection failures, establishing a quantitative benchmark for this purpose." + ], + "bbox": [ + 511, + 212, + 903, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 482, + 653, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object Detection. Object detection involves locating objects in images and identifying their categories [24]. Early models typically used convolutional neural network (CNN) backbones, divided into two main types: two-stage and single-stage models. Two-stage methods, like Faster R-CNN [35] and Mask R-CNN [14], generate region proposals to identify potential foreground objects before refining localization and classification. In contrast, single-stage methods, such asYOLO [41] and FCOS [39], directly classify and regress on backbone features. Transformer architectures [40] have further advanced object detection, as exemplified by DETR [2], which uses a fully end-to-end Transformer-based approach. The rise of multimodal foundation models [20, 34] has enabled open-set detection tasks [24], allowing object grounding with textual input [7, 47]. GLIP [20] frames detection as a grounding problem, integrating vision and language through feature fusion in the neck module. Grounding DINO [28] uses multiphase vision-language fusion, achieving state-of-the-art (SOTA) referring expression comprehension (REC) [27]. Other methods explore multimodal large language models (MLLMs) for detection and grounding tasks. Florence-2 [48], for instance, introduces a prompt-based model that generates object categories and coordinates using text prompts directly. The diverse feature extractors, architectures, and outputs in object detection make it challenging to", + "bbox": [ + 511, + 508, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30043", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "interpret their decisions using a single unified approach.", + "bbox": [ + 89, + 90, + 460, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object Detection Explanation. Explaining object detector decisions remains a largely unexplored area. Gudovskiy et al. [12] used Integrated Gradients (IG) [38] and SHAP [36] with SSD [29] classification scores for box-level attribution. Lee et al. [19] optimized masks on Mask R-CNN [14] to approximate original decisions with minimal pixels. Petsiuk et al. [33] proposed D-RISE, an adaptation of RISE [32], though it often introduces noise in saliency maps. Jiang et al. [16] introduced Nesterov-Accelerated iGOS++, which can be resource-intensive for large models. Zhao et al. [54] developed ODAM using Grad-CAM [37], but its effectiveness depends on network layer choices. SSGrad-CAM++ [49] generates saliency maps based on Grad-CAM++ [3]. VX-CODE [50] attributes important patches based on greedy search and SHAP [36]. While these methods aim to interpret detector errors, they largely rely on simple visualizations and empirical observations, often overlooking input-level feature confusion that can obscure human understanding. Masking confusing regions may aid in correcting model outputs [6]. Our paper introduces a black-box interpretation approach for object-level foundation models using Visual Precision Search, enhancing detection performance with fewer regions and identifying input-level failure causes to improve detection accuracy.", + "bbox": [ + 91, + 107, + 483, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 481, + 181, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents our method for explaining the decisions of object-level foundation models. Figure 2 illustrates the overall framework of our method.", + "bbox": [ + 89, + 506, + 483, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 561, + 294, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an image $\\mathbf{I} \\in \\mathbb{R}^{h \\times w \\times 3}$ and an object-level foundation model $f(\\cdot)$ , the output can be represented as $f(\\mathbf{I}) = \\{(b_i, c_i, s_i) \\mid i = 1, 2, \\dots, N\\}$ . Each tuple $(b_i, c_i, s_i)$ denotes the bounding box, class label, and confidence score for each detected object $i$ , where $N$ is the maximum number of detection boxes. Tasks like object detection involve multiple boxes and categories, while visual grounding tasks like referring expression comprehension (REC) have only one box and category. Our goal is to generate a saliency map that explains the reasons behind the model's detection of a specific object, $(b_i, c_i, s_i)$ . Ideally, this explanation would enable the correct detection of objects using fewer input regions, while also requiring the removal of only a few critical regions to invalidate the detector. To achieve this, we can sparsify the input region $V = \\{\\mathbf{I}_1^s, \\dots, \\mathbf{I}_m^s\\}$ , where $\\mathbf{I}_i^s$ represents the $i$ -th sub-region, $m$ is the total number of sub-regions, and $\\mathbf{I} = \\sum_{i=1}^{m} \\mathbf{I}_i^s$ . We then frame the saliency map generation task as a subset selection problem for these input sub-regions. A set function $\\mathcal{F}(\\cdot)$ is defined to assess interpretability by determining whether a given region is a key factor in the model's decision. Therefore, the", + "bbox": [ + 89, + 582, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "objectives are:", + "bbox": [ + 513, + 90, + 612, + 104 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {S \\subseteq V, | S | < k} \\mathcal {F} (S), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 108, + 903, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $k$ denotes the maximum number of sub-regions. Therefore, the key to this problem lies in designing set function $\\mathcal{F}$ and optimizing Eq. 1.", + "bbox": [ + 511, + 141, + 903, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Visual Precision Search", + "text_level": 1, + "bbox": [ + 513, + 199, + 730, + 214 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a Visual Precision Search method for interpreting object-level models. To begin, the input region needs to be sparsified. We apply the SLICO superpixel segmentation algorithm [1] to divide the input into $m$ sub-regions, $V = \\{\\mathbf{I}_1^s,\\dots ,\\mathbf{I}_m^s\\}$ . To solve Eq. 1, an $\\mathcal{NP}$ -hard problem, we employ submodular optimization [10]. Since the saliency map requires all sub-regions to be ranked, $k$ can be set to $|V|$ to compute ordered subsets. When the set function $\\mathcal{F}(\\cdot)$ satisfies the properties of diminishing returns1 and monotonic non-negative2, a greedy search guarantees an approximate optimal solution [8], with $\\mathcal{F}(S)\\geq (1 - 1 / e)\\mathcal{F}(S^{*})$ , where $S^{*}$ denotes the optimal solution and $S$ denotes the greedy solution. We next design a set function to evaluate the interpretability score and rank the importance of interpretable regions for object-level tasks.", + "bbox": [ + 511, + 220, + 906, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Clue Score:** An essential aspect of interpretability is enabling the object-level foundation model to accurately locate and identify objects while using fewer regions. To assess the importance of subregions from this perspective, we introduce the object location box information, $\\pmb{b}_{\\mathrm{target}}$ , and the target category, $c$ , that requires explanation. Given a subregion $S$ , the object-level model outputs $N$ instances, denoted as $f(S) = \\{(b_i, s_{c,i}) \\mid i = 1, \\dots, N\\}$ , where $s_{c,i}$ represents the confidence of the $i$ -th bounding box in predicting category $c$ . Then, the clue score of sub-region $S$ is defined as:", + "bbox": [ + 511, + 449, + 908, + 616 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns _ {\\text {c l u e}} \\left(S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c\\right) = \\max _ {\\left(\\boldsymbol {b} _ {i}, s _ {c, i}\\right) \\in f (S)} \\operatorname {I o U} \\left(\\boldsymbol {b} _ {\\text {t a r g e t}}, \\boldsymbol {b} _ {i}\\right) \\cdot s _ {c, i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 628, + 906, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{IoU}(\\cdot, \\cdot)$ represents the Intersection over Union between two bounding boxes. Unlike Petsiuk et al. [33], which only considers high-confidence bounding boxes, our method includes all candidate boxes to avoid getting stuck in a local optimum during the search. By incorporating the clue score $s_{\\mathrm{clue}}$ , our method focuses on regions that strengthen the desired positional and semantic response.", + "bbox": [ + 511, + 667, + 906, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Collaboration Score: Some regions may exhibit strong combination effects, meaning they contribute effectively to model decisions only when paired with multiple specific sub-regions. Therefore, we introduce the collaboration score $s_{\\mathrm{colla}}$ to assess sub-regions with high sensitivity to de", + "bbox": [ + 511, + 773, + 908, + 849 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Diminishing returns: $\\forall S_A \\subseteq S_B \\subseteq V \\setminus \\alpha$ , $\\mathcal{F}(S_A \\cup \\{\\alpha\\}) - \\mathcal{F}(S_A) \\geq \\mathcal{F}(S_B \\cup \\{\\alpha\\}) - \\mathcal{F}(S_B)$ .", + "bbox": [ + 511, + 861, + 903, + 887 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{2}$ Monotonic non-negative: $\\forall S \\subseteq V \\setminus \\alpha$ , $\\mathcal{F}(S \\cup \\{\\alpha\\}) - \\mathcal{F}(S) \\geq 0$ .", + "bbox": [ + 532, + 887, + 901, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30044", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/eaf7ef8b52307bca88e7cc7785e556bc3f953fd9d27dd770b73b87d821fe7bb6.jpg", + "image_caption": [ + "Figure 2. Framework of the proposed Visual Precision Search method for interpreting an object-level foundation model. The input is first sparsified into a set of sub-regions and then interpreted across different instances. A submodular function guides the search for significant sub-regions, updating the ordered subset iteratively, and ultimately generating the instance-level attribution map." + ], + "image_footnote": [], + "bbox": [ + 94, + 54, + 903, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cision outcomes:", + "bbox": [ + 89, + 297, + 205, + 309 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {\\text {c o l l a .}} (S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c) = 1 - \\max _ {(\\boldsymbol {b} _ {i}, s _ {c, i}) \\in f (V \\backslash S)} \\operatorname {I o U} (\\boldsymbol {b} _ {\\text {t a r g e t}}, \\boldsymbol {b} _ {i}) \\cdot s _ {c, i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 316, + 485, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this means that no detection box can accurately or confidently locate the object once the key sub-region is removed. This score serves as an effective guide in the initial search phase to identify subtle key regions that contribute to the final decision.", + "bbox": [ + 89, + 356, + 483, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Submodular Function: The scores above are combined to construct a submodular function $\\mathcal{F}(\\cdot)$ , as follows:", + "bbox": [ + 89, + 431, + 483, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} (S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c) = s _ {\\text {c l u e}} (S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c) + s _ {\\text {c o l l a .}} (S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 467, + 482, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Saliency Map Generation: Using the above submodular function, a greedy search algorithm is applied to sort all sub-regions in $V$ , yielding an ordered subset $S$ . Introducing the submodular function enables the search algorithm to more precisely identify key visual regions for interpretation. Additionally, scoring the sub-regions is necessary to better explain the importance of each sub-region. We evaluate the salient difference between the two sub-regions by the marginal effect. The attribution score $\\mathcal{A}_i$ for each sub-region $s_i$ in $S$ is assessed by:", + "bbox": [ + 89, + 489, + 483, + 640 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {i} = \\left\\{ \\begin{array}{l l} b _ {\\text {b a s e}} & \\text {i f} i = 1, \\\\ \\mathcal {A} _ {i - 1} - \\left| \\mathcal {F} \\left(S _ {[ i ]}\\right) - \\mathcal {F} \\left(S _ {[ i - 1 ]}\\right) \\right| & \\text {i f} i > 1, \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 646, + 482, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $b_{\\mathrm{base}}$ represents a baseline attribution score (such as 0 and won't affect results) for the first sub-region, and $S_{[i]}$ denotes the set containing the top $i$ sub-regions in $S$ . When a new sub-region is added, a small marginal increase suggests comparable importance to the previous sub-region. A negative marginal effect indicates a counterproductive impact, which can be assessed by its absolute value. Finally, $\\mathcal{A}$ is normalized to obtain the saliency map of the sub-region. The detailed calculation process of the proposed Visual Precision Search algorithm is outlined in Algorithm 1.", + "bbox": [ + 89, + 690, + 483, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Theory Analysis", + "text_level": 1, + "bbox": [ + 89, + 848, + 253, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A theoretical analysis is conducted for our Visual Precision Search method.", + "bbox": [ + 89, + 869, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1: Visual Precision Search algorithm for generating interpretable saliency maps of object-level foundation models.", + "bbox": [ + 522, + 299, + 862, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Image $\\mathbf{I} \\in \\mathbb{R}^{h \\times w \\times 3}$ , a division algorithm $\\mathrm{Div}(\\cdot)$ , the object location box $b_{\\text{target}}$ , and the target category $c$ .", + "bbox": [ + 517, + 347, + 883, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: An ordered set $S$ and a saliency map $\\mathcal{A} \\in \\mathbb{R}^{h \\times w}$ .", + "bbox": [ + 535, + 373, + 849, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1 $V\\gets \\mathrm{Div}(\\mathbf{I})$", + "bbox": [ + 522, + 386, + 614, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 $S\\gets \\emptyset$ / $\\star$ Initiate the operation of submodular subset selection $\\star/$", + "bbox": [ + 522, + 398, + 880, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 $\\mathcal{A}_1\\gets b_{\\mathrm{base}}$", + "bbox": [ + 522, + 421, + 604, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 for $i = 1$ to $|V|$ do", + "bbox": [ + 522, + 433, + 640, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 $S_{d}\\gets V\\backslash S;$", + "bbox": [ + 522, + 445, + 630, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\alpha \\gets \\arg \\max_{\\alpha \\in S_d}\\mathcal{F}(S\\cup \\{\\alpha \\} ,\\pmb {b}_{\\mathrm{target}},c)$", + "bbox": [ + 522, + 457, + 797, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "7 $S\\gets S\\cup \\{\\alpha \\} ;$", + "bbox": [ + 522, + 469, + 643, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "8 if $i > 1$ then", + "bbox": [ + 522, + 481, + 632, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "9 $|\\quad | \\quad | \\quad \\mathcal{A}_i\\gets \\mathcal{A}_{i - 1} - |\\mathcal{F}(S_{[i]}) - \\mathcal{F}(S_{[i - 1]})|$", + "bbox": [ + 522, + 492, + 799, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "10 end", + "bbox": [ + 521, + 508, + 557, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "11 return $S$ , $\\mathrm{norm}(\\mathcal{A})$", + "bbox": [ + 521, + 518, + 640, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 (Submodular Properties). Consider two subsets $S_A$ and $S_B$ in set $V$ , where $S_A \\subseteq S_B \\subseteq V$ . Given an element $\\alpha$ , where $\\alpha \\in V \\setminus S_b$ . Assuming that $\\alpha$ is contributing to model interpretation, then, the function $\\mathcal{F}(\\cdot)$ in Eq. 4 satisfies diminishing returns,", + "bbox": [ + 511, + 564, + 906, + 640 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} \\left(S _ {A} \\cup \\{\\alpha \\}\\right) - \\mathcal {F} \\left(S _ {A}\\right) \\geq \\mathcal {F} \\left(S _ {B} \\cup \\{\\alpha \\}\\right) - \\mathcal {F} \\left(S _ {B}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 648, + 883, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and monotonic non-negative, $\\mathcal{F}(S_A\\cup \\{\\alpha \\}) - \\mathcal{F}(S)\\geq 0$ and thus, $\\mathcal{F}$ is a submodular function.", + "bbox": [ + 511, + 676, + 905, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. Please see supplementary material for the proof.", + "bbox": [ + 511, + 715, + 877, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 1 (Impact on Sparse Division). The quality of the search space is determined by sparse division, meaning that both the method of partitioning the input and the number of sub-regions play a crucial role in the faithfulness of the Visual Precision Search.", + "bbox": [ + 511, + 739, + 906, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 2 (Applicable Models). If the model provides bounding box positions and category confidence without filtering out low-confidence detection boxes, the interpretation results of the Visual Precision Search can achieve guaranteed optimal boundaries. Models with detection heads", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30045", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are generally applicable; however, MLLM-based models may not directly output confidence scores, leaving room for improvement in the interpretation results.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 152, + 225, + 169 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 178, + 281, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate the proposed method on three object-level datasets: MS COCO 2017 [25], LVIS V1 [13], and RefCOCO [18]. The MS COCO dataset, used for interpreting object detection with 80 classes, involves random sampling of 5 correctly detected (IoU $>0.5$ , correct category), 3 misclassified (IoU $>0.5$ , wrong category), and 3 undetected (IoU $<0.5$ or low confidence) instances per class. The LVIS V1 dataset, with 1,203 categories including 337 rare ones (1-10 images), uses Grounding DINO for interpreting zero-shot detection on rare classes, analyzing 165 correctly detected, 152 misclassified, and 534 undetected samples. The RefCOCO dataset, for interpreting referring expression comprehension (REC), evaluates 500 correctly grounded (IoU $>0.5$ ) and 200 incorrectly grounded (IoU $<0.5$ ) samples per model.", + "bbox": [ + 89, + 200, + 483, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. The interpretation methods we compare include gradient-based approaches: Grad-CAM [37], SSGrad-CAM++[49], and ODAM[54], as well as perturbation-based methods: D-RISE [33] and D-HSIC [31]. D-RISE and D-HSIC are allowed to use all model outputs, including low-confidence detection boxes, to enhance the sufficiency of saliency estimation.", + "bbox": [ + 89, + 428, + 483, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We validate our method on two object-level foundation models, Grounding DINO [28] and Florence-2 [48], across object detection and referring expression comprehension (REC) tasks. To interpret model decisions, we set the ground truth bounding box and category as target box $\\pmb{b}_{\\mathrm{target}}$ and category $c$ . By default, the SLICO superpixel segmentation algorithm divides the image into 100 sparse sub-regions.", + "bbox": [ + 89, + 534, + 483, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 666, + 274, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Faithfulness Metrics. We evaluate the faithfulness of interpretation results across the entire image using Insertion and Deletion scores from the detection task, following the setup by Petsiuk et al. [33]. These metrics reflect location and recognition accuracy. We also calculate classification and IoU scores for the box closest to the target explanation denoted as Insertion (class), Deletion (class), Insertion (IoU), and Deletion (IoU). Additionally, we compute the average highest confidence, defined as the highest class response for the box closest to the target with IoU $>0.5$ . For failure samples, we introduce the Explaining Successful Rate (ESR) to evaluate whether the saliency map can effectively guide the search within a limited region, enabling the model to correctly detect misclassified or low-confidence objects", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(with IoU $>0.5$ and classification confidence above the detector's default threshold) previously.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Location Metrics. We report the Point Game [53] and Energy Point Game [42] metrics. Since these metrics typically assume that the model's decision is relevant to the target object [33], we calculate them only for interpretation results on samples the model has correctly detected.", + "bbox": [ + 511, + 121, + 906, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Faithfulness Analysis", + "text_level": 1, + "bbox": [ + 511, + 208, + 714, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3.1. Faithfulness on Grounding DINO", + "text_level": 1, + "bbox": [ + 511, + 229, + 794, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin by validating our approach in explaining decisions across various object-level tasks using Grounding DINO [28]. Table 1 presents the faithfulness results across different datasets and tasks, demonstrating the strong interpretative faithfulness of our method.", + "bbox": [ + 511, + 250, + 905, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the object detection interpretation task on the MS COCO dataset, our method achieved state-of-the-art results, outperforming the D-RISE method by $23.7\\%$ on the Insertion metric, $6.7\\%$ on the Deletion metric, and $10.6\\%$ on the average highest score. The Insertion and Deletion metrics also achieved top performance under both category and IoU conditions. Additionally, our method performed best in location metrics, indicating state-of-the-art performance in explaining general object detection tasks. We also found that gradient-based methods, such as ODAM, exhibit low faithfulness in explaining Grounding DINO decisions. This is primarily due to the fusion of text and visual modalities, which causes the gradient to be influenced by both, reducing its effectiveness in attributing to the visual modality alone.", + "bbox": [ + 511, + 325, + 906, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the referring expression comprehension (REC) interpretation task on the RefCOCO dataset, our method achieved state-of-the-art results across all evaluation metrics, surpassing the D-RISE method by $20.1\\%$ on the Insertion metric, $22.1\\%$ on the Deletion metric, and $4.4\\%$ on the average highest score. These results confirm that our method maintains high faithfulness in explaining object description understanding.", + "bbox": [ + 511, + 537, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also interpret zero-shot object detection for rare classes on the LVIS V1 dataset. Our method achieved SOTA results in faithfulness metrics, surpassing D-RISE by $31.6\\%$ on the Insertion metric, $4.2\\%$ on the Deletion metric, and $15.9\\%$ on the average highest score. Although our method ranks second to ODAM on Energy Point Game metrics, this is mainly due to Grounding DINO not being fully trained on rare LVIS classes; as a result, model responses may be influenced by other regions rather than the object.", + "bbox": [ + 511, + 659, + 905, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental results show that our method significantly outperforms other interpretability approaches, primarily due to the submodularity of our objective function, enabling precise evaluation of critical sub-regions. As shown in Figure 3, ODAM's saliency maps are diffuse, and D-RISE's maps are noisy, whereas our method sharply highlights essential sub-regions, capturing edges and class-specific fea", + "bbox": [ + 511, + 796, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30046", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a4114669db46f8330c4fdbfb173c1a646a85557253bef9dc5c89a8ed0335812d.jpg", + "table_caption": [ + "Table 1. Evaluation of faithfulness metrics (Deletion, Insertion AUC scores, and average highest score) and location metrics (Point Game and Energy Point Game) on the MS-COCO, RefCOCO, and LVIS V1 (rare) validation sets for correctly detected or grounded samples using Grounding DINO." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodsFaithfulness MetricsLocation Metrics
Ins. (↑)Del. (↓)Ins. (class) (↑)Del. (class) (↓)Ins. (IoU) (↑)Del. (IoU) (↓)Ave. high. score (↑)Point Game (↑)Energy PG (↑)
MS COCO [25] (Detection task)Grad-CAM [37]0.24360.15260.30640.20060.62290.53240.59040.17460.1463
SSGrad-CAM++ [49]0.21070.17780.26390.23140.59810.55110.58860.19050.1293
D-RISE [33]0.44120.04020.50810.08860.83960.36420.62150.94970.1850
D-HSIC [31]0.37760.04390.43820.09030.83010.33010.58620.73280.1861
ODAM [54]0.31030.05190.36550.08940.78690.39840.58650.54310.2034
Ours0.54590.03750.62040.08820.85810.33000.68730.98940.2046
RefCOCO [18] (REC task)Grad-CAM [37]0.37490.42370.46580.51940.75160.76850.74810.23800.2171
SSGrad-CAM++ [49]0.41130.39250.50080.48510.77000.75880.75610.28200.2262
D-RISE [33]0.61780.16050.70330.33960.86060.51640.84710.94000.2870
D-HSIC [31]0.54910.18460.62950.35090.85040.51200.77390.79000.3190
ODAM [54]0.47780.27180.56200.37570.82170.66410.74250.63200.3529
Ours0.74190.12500.80800.24570.90500.51030.88420.94600.3566
LVIS V1 (rare) [13] (Zero-shot det. task)Grad-CAM [37]0.12530.12940.18010.18140.56570.59100.35490.11510.0941
SSGrad-CAM++ [49]0.12530.12540.17650.17750.58000.56910.35040.10910.0931
D-RISE [33]0.28080.02890.33480.08350.83030.31740.42890.96970.1462
D-HSIC [31]0.24170.03530.29120.09280.81870.35500.40440.83030.1730
ODAM [54]0.20090.04100.24780.08440.77600.40820.36940.60610.2050
Ours0.36950.02770.42750.07990.84790.32420.49690.97580.1785
", + "bbox": [ + 94, + 95, + 903, + 308 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/06fcedbef2942fbaf3880fe70e2342bc62c7fa85e7f171e519aa3b463d587286.jpg", + "image_caption": [ + "Figure 3. Visualization results of Grounding DINO for interpreting object localization tasks on the MS COCO, RefCOCO, and LVIS datasets. The first column shows the saliency map with the ground truth box and label. The second column presents detection within the limited search region, and the third columns display the insertion curves." + ], + "image_footnote": [], + "bbox": [ + 99, + 311, + 897, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tures that enhance interpretability.", + "bbox": [ + 89, + 564, + 316, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2. Faithfulness on Florence-2", + "text_level": 1, + "bbox": [ + 89, + 590, + 320, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further validated our approach on Florence-2 [48], a foundation model lacking object confidence scores. This limitation requires our search to rely solely on IoU as a guiding metric, excluding comparison with gradient-based methods. As shown in Table 2, our method achieves state-of-the-art performance.", + "bbox": [ + 89, + 609, + 483, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the object detection interpretability task on the MS COCO dataset, our method outperforms D-RISE by $3.8\\%$ on the Insertion metric and achieves a substantial improvement of $50.7\\%$ on the Deletion metric. Additionally, our method enhances the Point Game and Energy Point Game metrics by $8.3\\%$ and $60.7\\%$ , respectively. Due to the absence of object confidence scores in Florence-2, it becomes challenging to identify areas that quickly invalidate the model's detection results. Compared to perturbation-based methods, our approach offers significant advantages; it relies more directly on the object itself, resulting in stronger faithfulness to the model's detection behavior.", + "bbox": [ + 89, + 703, + 483, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also interpret referring expression comprehension", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(REC) interpretation task on the RefCOCO dataset, outperforming D-RISE by $6.1\\%$ and $66.9\\%$ in Insertion and Deletion metrics, respectively. Our approach also achieved state-of-the-art localization results, with a $14.6\\%$ improvement in the Energy Point Game, demonstrating strong interpretability in accurately pinpointing object information.", + "bbox": [ + 511, + 564, + 905, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We visualized some results of our method in Figure 4, illustrating that it effectively identifies key information about the object's location and categorical description. While D-RISE performed well on the Insertion metric, its attribution results were diffuse and lacked reliability, and it did not perform as well on the Deletion metric.", + "bbox": [ + 511, + 655, + 905, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Interpreting Failures in Object-Level Tasks", + "text_level": 1, + "bbox": [ + 511, + 753, + 880, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4.1. Interpreting REC Failures", + "text_level": 1, + "bbox": [ + 511, + 776, + 743, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Failures in the REC task primarily arise from visual grounding errors, often due to interference from other objects in the scene that become entangled with the text. Removing this distracting visual information can clarify grounding results and improve error understanding at the visual input level. To achieve this, we specify the correct location as the target box, denoted as $\\pmb{b}_{\\mathrm{target}}$ , for input attribution. Ta", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30047", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/16829e560097f2656b49767968956ccf8c3ce375da67aad6154c0342b2f0eb7b.jpg", + "table_caption": [ + "Table 2. Evaluation of faithfulness metrics (Deletion and Insertion AUC scores) and location metrics (Point Game and Energy Point Game) on the MS COCO and RefCOCO validation sets for correctly detected and grounded samples using Florence-2." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodsFaithfulness MetricsLocation Metrics
Insertion (↑)Deletion (↓)Point Game (↑)Energy PG (↑)
MS COCO [25] (Detection task)D-RISE [33]0.74770.09720.88500.1568
D-HSIC [31]0.53450.27300.29250.0862
Ours0.77590.04790.95830.2519
RefCOCO [18] (REC task)D-RISE [33]0.79220.35050.84800.2464
D-HSIC [31]0.76390.35600.69800.2754
Ours0.84090.11590.86600.3927
", + "bbox": [ + 94, + 147, + 483, + 236 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/168889f26d4bbeab9004c4ea9077f0b9164776fde77f0e75992663ed758ed185.jpg", + "image_caption": [ + "MS COCO (Object Detection Interpretation)" + ], + "image_footnote": [], + "bbox": [ + 98, + 258, + 478, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7fd96ec7d9b4846ebb8ca56d73853a3204cfc66e17d02554e24e50fc782f895.jpg", + "image_caption": [ + "RefCOCO (Visual Grounding Interpretation)", + "Figure 4. Visualization results of Florence-2 for interpreting object localization tasks on the MS COCO and RefCOCO datasets." + ], + "image_footnote": [], + "bbox": [ + 98, + 387, + 478, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ble 3 shows the attribution results, where gradient-based methods exhibit limited performance in faithfulness metrics. Compared to D-RISE, our approach improves Insertion by $42.9\\%$ , class score-based Insertion by $25.1\\%$ , and average highest score by $14.9\\%$ . Figure 5 illustrates model error interpretations, with the cyan-highlighted region indicating decision errors due to visual input interference that leads the foundation model astray.", + "bbox": [ + 88, + 550, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2. Interpreting Detection Failures", + "text_level": 1, + "bbox": [ + 89, + 681, + 354, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Detection failures include misclassification and undetected objects. While previous works [33, 54] used visualization to explain errors, it lacked quantitative analysis, resulting in lower trustworthiness. We used saliency maps to identify input-level interference causing detection failures.", + "bbox": [ + 89, + 702, + 483, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Misclassified Interpretation. As shown in Table 4, On the MS COCO dataset, our method outperforms D-RISE, improving Insertion by $54.7\\%$ , Insertion (class) by $49.1\\%$ , average highest score by $27.4\\%$ , and explaining success rate (ESR) by $19.47\\%$ . Similarly, on the LVIS dataset, our method shows advantages with improvements of $42.7\\%$ in Insertion, $33.0\\%$ in Insertion (class), $24.8\\%$ in average highest score, and $24.34\\%$ in ESR. This improvement is pri", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/71482d55bf8c3d1aeea55bdcfcc1b6d1e40639e13ca41aea960ea1af72836cc6.jpg", + "table_caption": [ + "Table 3. Insertion AUC scores and the average highest score on the RefCOCO validation sets for or the samples with incorrect localization in visual grounding using Grounding DINO." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodsFaithfulness Metrics
Ins. (†)Ins. (class) (†)Ave. high. score (†)
RefCOCO [18] (REC task)Grad-CAM [37]0.15360.27940.3295
SSGrad-CAM++ [49]0.15900.28370.3266
D-RISE [33]0.34860.47870.6096
D-HSIC [31]0.22740.34880.4495
ODAM [54]0.17930.30010.3453
Ours0.49810.59900.7007
", + "bbox": [ + 527, + 132, + 893, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bcdc9a12c9905eed0b174787847dc65d369e8b52c586c37900f296874c6e46c5.jpg", + "image_caption": [ + "Attribution Map" + ], + "image_footnote": [], + "bbox": [ + 516, + 239, + 638, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/624dba656a394fb44bce58772fae49a4e3ec2407027bf4cd67bfb4b3a119ab3a.jpg", + "image_caption": [ + "Searched Region" + ], + "image_footnote": [], + "bbox": [ + 651, + 239, + 766, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18ec6165b373b0df06406c8966afde8776f605462b14c7e9bb46ed9cb5946570.jpg", + "image_caption": [ + "Figure 5. Visualization of the method for discovering what causes the Grounding DINO incorrectly grounded on RefCOCO." + ], + "image_footnote": [], + "bbox": [ + 790, + 224, + 900, + 311 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9b758500f49a640018ddec97415d499e0cc772f34cd8e526ee624f83e94bf49b.jpg", + "table_caption": [ + "Table 4. Insertion AUC scores, average highest score, and explaining successful rate (ESR) on the MS-COCO and the LVIS validation sets for misclassified samples using Grounding DINO." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodsFaithfulness Metrics
Ins. (↑)Ins. (class) (↑)Ave. high. score (↑)ESR (↑)
MS COCO [25] (Detection task)Grad-CAM [37]0.10910.14780.310238.38%
SSGrad-CAM++ [49]0.09600.13360.295233.51%
D-RISE [33]0.21700.26610.360350.26%
D-HSIC [31]0.17710.21610.314334.59%
ODAM [54]0.11290.14860.286932.97%
Ours0.33570.39670.459169.73%
LVIS V1 (rare) [13] (Zero-shot det. task)Grad-CAM [37]0.05030.08910.156412.50%
SSGrad-CAM++ [49]0.05740.09460.158011.84%
D-RISE [33]0.12450.16470.208828.95%
D-HSIC [31]0.09630.12470.174816.45%
ODAM [54]0.05750.09540.15209.21%
Ours0.17760.21900.260653.29%
", + "bbox": [ + 517, + 386, + 903, + 515 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e2aadfbc73247ae622522d08a8657860edf3733d7065e22a60be577012c6570.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 521, + 617, + 646 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1b48ac7a890ea8b98c764067c57e0a6987dba755e7201046703b7f46a81d881.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 521, + 709, + 645 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2767f4fec27d15220a39d7a33cef5eb2e72f6aad5b2f699395aaac9e3cad41d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 521, + 799, + 645 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b9a3766fa0276b71e5b955eb94ef5ffa3076c9620ca85bdeafee9eb46036771.jpg", + "image_caption": [ + "Figure 6. Visualization of our method reveals the causes of Grounding DINO misclassifications on MS COCO and LVIS. The cyan region in the saliency map highlights the regions responsible for the model's misclassification." + ], + "image_footnote": [], + "bbox": [ + 800, + 521, + 901, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "marily due to our proposed Visual Precision Search, which confines attribution to the target negative response region, effectively refining the model's detection results. Figure 6 shows that the background surrounding the object interferes with the model's decision-making. Improving the model by refining the contextual relationship between foreground and background may be a promising direction.", + "bbox": [ + 511, + 717, + 906, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Undetected Interpretation. Low object confidence may result from both the model's feature representation and confounding input-level factors. We analyze causes of missed detections, with Table 5 showing that our method achieves SOTA performance across metrics. On MS", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30048", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/790e471d798106b528f8f881dd03cef9f7702869ca6adec0f8cd03b61dff1d36.jpg", + "table_caption": [ + "Table 5. Insertion, average highest score, and explaining successful rate (ESR) on the MS-COCO and the LVIS V1 (rare) validation sets for missed detection samples using Grounding DINO." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodsFaithfulness Metrics
Ins. (↑)Ins. (class) (↑)Ave. high. score (↑)ESR (↑)
MS COCO [25] (Detection task)Grad-CAM [37]0.07600.13210.215316.44%
SSGrad-CAM++ [49]0.06710.11510.212416.44%
D-RISE [33]0.15380.22600.256426.94%
D-HSIC [31]0.11010.17160.194513.56%
ODAM [54]0.07450.13500.203713.78%
Ours0.21020.30110.301441.33%
LVIS V1 (rare) [13] (Zero-shot det. task)Grad-CAM [37]0.02910.06890.09015.43%
SSGrad-CAM++ [49]0.02920.06800.08975.24%
D-RISE [33]0.07030.11840.131218.73%
D-HSIC [31]0.05160.09200.116813.48%
ODAM [54]0.02830.07160.08514.68%
Ours0.11550.18860.178430.15%
", + "bbox": [ + 93, + 109, + 483, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/20aeadba9f1187fd25ce0e0faa0f9f4b6a4f06d8b2b970286795cbb1ac85972d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 268, + 230, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb25181b7122be4733e9b26602288f904e81c94c1539f47e630e9d66cef0a48d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 268, + 346, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7df524c5ee250a3735a17b2654078647ea8f5fe25ef1c8c97bc8fa9073e4ceaf.jpg", + "image_caption": [ + "Figure 7. Visualization of our method reveals the causes of Grounding DINO undetected on MS COCO and LVIS. The cyan region in the saliency map highlights the regions responsible for the model's detection failure." + ], + "image_footnote": [], + "bbox": [ + 354, + 268, + 478, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "COCO, it outperforms D-RISE with a $36.7\\%$ improvement in Insertion and a $14.39\\%$ gain in ESR, while on LVIS, it improves Insertion by $64.3\\%$ and ESR by $11.42\\%$ . Figure 7 visualizes our method's explanations for undetected instances, revealing that errors may arise from challenges in distinguishing similar objects (e.g., the bear in the first row) and environmental influence on detection (e.g., the corkboard in the second row). These insights highlight current model limitations, offering directions for improvement.", + "bbox": [ + 89, + 500, + 483, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 648, + 243, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation of the Submodular Function. We analyze the effectiveness of our submodular function design on Grounding DINO using correctly detected samples from MS COCO. Table 6 shows the results. The Clue Score identifies regions beneficial for detection, improving Insertion and average highest score, while the Collaboration Score pinpoints sensitive regions, enhancing the Deletion metric by causing the detector to fail more quickly. Combining these scores enables our method to achieve optimal results across indicators, demonstrating the effectiveness of each score function within the submodular function.", + "bbox": [ + 89, + 672, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Divided Sub-region Number. The sub-region number affects search space quality. Using Grounding DINO, we assess its impact on correctly predicted and misclassified samples in MS COCO. Figures 8A-B show", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b1db99e6d8292562ca6a2484535ab5552e0a6c07a1824fb7f08d779ff3681b5a.jpg", + "table_caption": [ + "Table 6. Ablation study on function score components for Grounding DINO on the MS COCO validation set." + ], + "table_footnote": [], + "table_body": "
Clue Score (Eq. 2)Colla. Score (Eq. 3)Insertion (↑)Faithfulness Metrics
Deletion (↓)Ave. high. score (↑)
✗✓0.36320.03780.5967
✓✗0.53700.07990.6864
✓✓0.54590.03750.6873
", + "bbox": [ + 540, + 99, + 880, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/036d93524432f2eb3c3404764702619843f7ec9b485613259b3239f3a22ab99e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 180, + 643, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d194756454313e106e54a7438eb6385fbb655c7dfc6f08371d0f431038d27ac8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 180, + 771, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32360a8a3ee25e5b4d89474aa6d4d5dfa81105e1236700541e4d8c659e6af12e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 180, + 900, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8153aa83ef8498af7d707a9f49587a9054d965f453bb4bcbb1867f1b39ef0fd7.jpg", + "image_caption": [ + "Figure 8. Ablation on the number of sub-regions. For correct detections, panels $\\mathbf{A}$ and $\\mathbf{B}$ show the effect of sub-region count on Insertion and average highest score, respectively. Panel $\\mathbf{C}$ presents the inference time for different sub-region counts. For misclassified samples, panels $\\mathbf{D},\\mathbf{E},$ and $\\mathbf{F}$ display the effects on Insertion, average highest score, and explanation success rate." + ], + "image_footnote": [], + "bbox": [ + 524, + 261, + 643, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5a30510e6322b814bc0b5d15f36dc1222dab683aafae12775e946ee05583565.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 261, + 772, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe794ee2cdcb70b9fd92fcb0b2b9ebe28b7838e44255182e4785f6814d489df0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 261, + 900, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that when the model predicts correctly, increasing the number of sub-regions leads to higher Insertion and average highest scores, indicating that finer divisions enhance the faithfulness of search results. Additionally, Figures 8D-F show that, for misclassified samples, Insertion, average highest score, and ESR also improve as the number of sub-regions increases. Figure 8C shows inference times for our search algorithm on an RTX 3090 compared to D-RISE across various subregion counts. Increasing sub-regions improves faithfulness but also rapidly increases inference time. Future work will aim to increase subregion count and inference speed without sacrificing attribution performance.", + "bbox": [ + 511, + 468, + 906, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 686, + 633, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an interpretable attribution method specifically tailored for object-level foundation models, called the Visual Precision Search method, which introduces a novel submodular mechanism that combines a clue score and a collaboration score. This method achieves enhanced interpretability with fewer search regions. Experiments on RefCOCO, MS COCO, and LVIS demonstrate that our approach improves object-level task interpretability over state-of-the-art methods for Grounding DINO and Florence-2 across various evaluation metrics. Furthermore, our method effectively interprets failures in visual grounding and object detection tasks.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30049", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 256, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Shenzhen Science and Technology Program (No. KQTD20221101093559018), National Natural Science Foundation of China (No. 62372448, 62025604, 62306308, 62441619), Foundation of Key Laboratory of Education Informatization for Nationalities (YNNU), Ministry of Education (No. EIN2024B004).", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Trans. Pattern Anal. Mach. Intell., 34(11):2274-2282, 2012. 3", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Eur. Conf. Comput. Vis. (ECCV), pages 213-229, 2020. 2", + "[3] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks. In IEEE Winter Conf. Appl. Comput. Vis. (WACV), pages 839-847, 2018. 3", + "[4] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 1", + "[5] Ruoyu Chen, Jingzhi Li, Hua Zhang, Changchong Sheng, Li Liu, and Xiaochun Cao. Sim2Word: Explaining similarity with representative attribute words via counterfactual explanations. ACM Trans. Multim. Comput. Commun. Appl., 19 (6):1-22, 2023. 1", + "[6] Ruoyu Chen, Hua Zhang, Siyuan Liang, Jingzhi Li, and Xiaochun Cao. Less is more: Fewer interpretable region via submodular subset selection. In Int. Conf. Learn. Represent. (ICLR), 2024. 1, 3", + "[7] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 16901-16911, 2024. 1, 2", + "[8] Jack Edmonds. Submodular functions, matroids, and certain polyhedra. Combinatorial Structures and Their Applications, pages 69-87, 1970. 3", + "[9] Di Feng, Ali Harakeh, Steven L Waslander, and Klaus Dietmayer. A review and comparative study on probabilistic object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(8):9961-9980, 2021. 1", + "[10] Satoru Fujishige. Submodular functions and optimization. Elsevier, 2005. 3", + "[11] Yossi Gandelsman, Alexei A Efros, and Jacob Steinhardt. Interpreting clip's image representation via text-based decomposition. In Int. Conf. Learn. Represent. (ICLR), 2024. 1", + "[12] Denis Gudovskiy, Alec Hodgkinson, Takuya Yamaguchi, Yasunori Ishii, and Sotaro Tsukizawa. Explain to fix: A" + ], + "bbox": [ + 93, + 244, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "framework to interpret and correct dnn object detector predictions. arXiv preprint arXiv:1811.08011, 2018.3", + "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5356-5364, 2019. 2, 5, 6, 7, 8", + "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. IEEE Trans. Pattern Anal. Mach. Intell., 42(2):386-397, 2018. 1, 2, 3", + "[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 17853-17862, 2023. 1", + "[16] Mingqi Jiang, Saeed Khorram, and Fuxin Li. Diverse explanations for object detectors with nesterov-accelerated igos++. In Brit. Mach. Vis. Conf. (BMVC), pages 188–189, 2023. 3", + "[17] Mingqi Jiang, Saeed Khorram, and Li Fuxin. Comparing the decision-making mechanisms by transformers and cnns via explanation methods. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 9546-9555, 2024. 1", + "[18] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. ReferItGame: Referring to objects in photographs of natural scenes. In Proc. Conf. Empir. Methods Nat. Lang. Process. (EMNLP), pages 787-798, 2014. 2, 5, 6, 7", + "[19] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. BBAM: Bounding box attribution map for weakly supervised semantic and instance segmentation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 2643-2652, 2021. 3", + "[20] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 10965-10975, 2022. 1, 2", + "[21] Siyuan Liang, Baoyuan Wu, Yanbo Fan, Xingxing Wei, and Xiaochun Cao. Parallel rectangle flip attack: A query-based black-box attack against object detection. In Int. Conf. Comput. Vis. (ICCV), pages 7697-7707, 2021. 1", + "[22] Siyuan Liang, Longkang Li, Yanbo Fan, Xiaojun Jia, Jingzhi Li, Baoyuan Wu, and Xiaochun Cao. A large-scale multiple-objective method for black-box attack against object detection. In Eur. Conf. Comput. Vis. (ECCV), pages 619-636, 2022. 1", + "[23] Siyuan Liang, Aishan Liu, Jiawei Liang, Longkang Li, Yang Bai, and Xiaochun Cao. Imitated detectors: Stealing knowledge of black-box object detectors. In ACM Int. Conf. Multimedia, pages 4839-4847, 2022. 1", + "[24] Siyuan Liang, Wei Wang, Ruoyu Chen, Aishan Liu, Boxi Wu, Ee-Chien Chang, Xiaochun Cao, and Dacheng Tao. Object detectors in the open environment: Challenges, solutions, and outlook. arXiv preprint arXiv:2403.16271, 2024. 1, 2", + "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30050", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zitnick. Microsoft COCO: Common objects in context. In Eur. Conf. Comput. Vis. (ECCV), pages 740-755, 2014. 2, 5, 6, 7, 8", + "[26] Aishan Liu, Jun Guo, Jiakai Wang, Siyuan Liang, Renshuai Tao, Wenbo Zhou, Cong Liu, Xianglong Liu, and Dacheng Tao. {X-Adv}: Physical adversarial object attacks against x-ray prohibited item detection. In USENIX Security Symposium, pages 3781–3798, 2023. 1", + "[27] Jingyu Liu, Liang Wang, and Ming-Hsuan Yang. Referring expression generation and comprehension via attributes. In Int. Conf. Comput. Vis. (ICCV), pages 4856-4864, 2017. 2", + "[28] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. In Eur. Conf. Comput. Vis. (ECCV), 2024. 1, 2, 5", + "[29] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. SSD: Single shot multibox detector. In Eur. Conf. Comput. Vis. (ECCV), pages 21-37, 2016. 3", + "[30] Dimity Miller, Feras Dayoub, Michael Milford, and Niko Sünderhauf. Evaluating merging strategies for sampling-based uncertainty techniques in object detection. In Proc. IEEE Int. Conf. Robot. Autom. (ICRA), pages 2348-2354, 2019. 1", + "[31] Paul Novello, Thomas Fel, and David Vigouroux. Making sense of dependence: Efficient black-box explanations using dependence measure. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 4344-4357, 2022. 5, 6, 7, 8", + "[32] Vitali Petsiuk, Abir Das, and Kate Saenko. RISE: Randomized input sampling for explanation of black-box models. In Brit. Mach. Vis. Conf. (BMVC), page 151, 2018. 3", + "[33] Vitali Petsiuk, Rajiv Jain, Varun Manjunatha, Vlad I Morariu, Ashutosh Mehra, Vicente Ordonez, and Kate Saenko. Black-box explanation of object detectors via saliency maps. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 11443-11452, 2021. 2, 3, 5, 6, 7, 8", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn. (ICML), pages 8748-8763, 2021. 1, 2", + "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell., 39(6):1137-1149, 2016. 2", + "[36] M Scott, Lee Su-In, et al. A unified approach to interpreting model predictions. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 4765-4774, 2017. 3", + "[37] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis., 128:336-359, 2020. 3, 5, 6, 7, 8", + "[38] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Int. Conf. Mach. Learn. (ICML), pages 3319-3328, 2017. 3" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: A simple and strong anchor-free object detector. IEEE Trans. Pattern Anal. Mach. Intell., 44(4):1922-1933, 2020. 2", + "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 5998-6008, 2017. 2", + "[41] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 7464-7475, 2023. 2", + "[42] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 5", + "[43] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable adversarial attacks for image and video object detection. In IJCAI, pages 954-960, 2019. 1", + "[44] Licheng Wen, Xuemeng Yang, Daocheng Fu, Xiaofeng Wang, Pinlong Cai, Xin Li, MA Tao, Yingxuan Li, XU Linran, Dengke Shang, et al. On the road with gpt-4v (ision): Explorations of utilizing visual-language model as autonomous driving agent. In ICLR 2024 Workshop on Large Language Model (LLM) Agents, 2024. 1", + "[45] Samuel Wilson, Tobias Fischer, Feras Dayoub, Dimity Miller, and Niko Sunderhauf. Safe: Sensitivity-aware features for out-of-distribution object detection. In Int. Conf. Comput. Vis. (ICCV), pages 23565-23576, 2023. 1", + "[46] Junfeng Wu, Yi Jiang, Qihao Liu, Zehuan Yuan, Xiang Bai, and Song Bai. General object foundation model for images and videos at scale. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3783-3795, 2024. 1", + "[47] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, et al. Towards open vocabulary learning: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2", + "[48] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4818–4829, 2024. 1, 2, 5, 6", + "[49] Toshinori Yamauchi. Spatial Sensitive Grad-CAM++: Improved visual explanation for object detectors via weighted combination of gradient map. In CVPR Workshop, pages 8164-8168, 2024. 3, 5, 6, 7, 8", + "[50] Toshinori Yamauchi, Hiroshi Kera, and Kazuhiko Kawamoto. Explaining object detectors via collective contribution of pixels. arXiv preprint arXiv:2412.00666, 2024.3", + "[51] Lewei Yao, Renjie Pi, Jianhua Han, Xiaodan Liang, Hang Xu, Wei Zhang, Zhenguo Li, and Dan Xu. Detclipv3: Towards versatile generative open-vocabulary object detection. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 27391-27401, 2024. 1" + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30051", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. Int. J. Comput. Vis., pages 1-19, 2024. 1", + "[53] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. Int. J. Comput. Vis., 126 (10):1084-1102, 2018. 5", + "[54] Chenyang Zhao, Janet H Hsiao, and Antoni B Chan. Gradient-based instance-specific visual explanations for object specification and object discrimination. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2, 3, 5, 6, 7, 8", + "[55] Chenyang Zhao, Kun Wang, Xingyu Zeng, Rui Zhao, and Antoni B Chan. Gradient-based visual explanation for transformer-based clip. In Int. Conf. Mach. Learn. (ICML), pages 61072-61091, 2024. 1", + "[56] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 15116-15127, 2023. 1", + "[57] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 1" + ], + "bbox": [ + 91, + 90, + 480, + 430 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "30052", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Interpreting Object-level Foundation Models via Visual Precision Search/396da971-babe-4c65-8e62-f7e2e9a705a9_model.json b/2025/Interpreting Object-level Foundation Models via Visual Precision Search/396da971-babe-4c65-8e62-f7e2e9a705a9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..367fd0e663b45c3cdf33100f6fbc8f01182716b2 --- /dev/null +++ b/2025/Interpreting Object-level Foundation Models via Visual Precision Search/396da971-babe-4c65-8e62-f7e2e9a705a9_model.json @@ -0,0 +1,2686 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.131, + 0.867, + 0.153 + ], + "angle": 0, + "content": "Interpreting Object-level Foundation Models via Visual Precision Search" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.16, + 0.791, + 0.197 + ], + "angle": 0, + "content": "Ruoyu Chen\\(^{1,2}\\), Siyuan Liang\\(^{3}\\), Jingzhi Li\\(^{1,2,7}\\), Shiming Liu\\(^{4}\\), Maosen Li\\(^{5}\\), Zhen Huang\\(^{6}\\), Hua Zhang\\(^{1,2,*}\\), and Xiaochun Cao\\(^{8,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.198, + 0.804, + 0.214 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100093, China" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.216, + 0.813, + 0.232 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China" + }, + { + "type": "list", + "bbox": [ + 0.186, + 0.198, + 0.813, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.233, + 0.894, + 0.25 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Computing, NUS \\(^{4}\\)RAMS Lab, Huawei Technologies Co., Ltd. \\(^{5}\\)IAS BU, Huawei Technologies Co., Ltd." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.251, + 0.896, + 0.267 + ], + "angle": 0, + "content": "6College of Computer, NUDT 7Key Lab. of Edu. Inf. for Nationalities (YNNU), Ministry of Education, Kunming, China" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.268, + 0.876, + 0.285 + ], + "angle": 0, + "content": "\\(^{8}\\)School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen 518107, China" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.288, + 0.86, + 0.302 + ], + "angle": 0, + "content": "chenruoyu@ie.ac.cn pandaliang521@gmail.com {lijingzhi,zhanghua}@ie.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.305, + 0.885, + 0.32 + ], + "angle": 0, + "content": "{liushiming3, limaosen2}@huawei.com huangzhen@nudt.edu.cn caoxiaochun@mail.sysu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.324, + 0.327, + 0.34 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.485, + 0.794 + ], + "angle": 0, + "content": "Advances in multimodal pre-training have propelled object-level foundation models, such as Grounding DINO and Florence-2, in tasks like visual grounding and object detection. However, interpreting these models' decisions has grown increasingly challenging. Existing interpretable attribution methods for object-level task interpretation have notable limitations: (1) gradient-based methods lack precise localization due to visual-textual fusion in foundation models, and (2) perturbation-based methods produce noisy saliency maps, limiting fine-grained interpretability. To address these, we propose a Visual Precision Search method that generates accurate attribution maps with fewer regions. Our method bypasses internal model parameters to overcome attribution issues from multimodal fusion, dividing inputs into sparse sub-regions and using consistency and collaboration scores to accurately identify critical decision-making regions. We also conducted a theoretical analysis of the boundary guarantees and scope of applicability of our method. Experiments on RefCOCO, MS COCO, and LVIS show our approach enhances object-level task interpretability over SOTA for Grounding DINO and Florence-2 across various evaluation metrics, with faithfulness gains of \\(23.7\\%\\), \\(31.6\\%\\), and \\(20.1\\%\\) on MS COCO, LVIS, and RefCOCO for Grounding DINO, and \\(50.7\\%\\) and \\(66.9\\%\\) on MS COCO and RefCOCO for Florence-2. Additionally, our method can interpret failures in visual grounding and object detection tasks, surpassing existing methods across multiple evaluation metrics. The code is released at https://github.com/RuoyuChen10/VPS." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.804, + 0.222, + 0.819 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.83, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Understanding object information in images, such as object detection [7, 14, 24, 51, 52, 57], is a crucial and enduring challenge in computer vision, holding significant im" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.325, + 0.888, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.581, + 0.907, + 0.637 + ], + "angle": 0, + "content": "Figure 1. Illustration of our Visual Precision Search interpretation method, which more precisely identifies key sub-regions in object-level foundation model decision-making compared to gradient-based and perturbation-based methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.868 + ], + "angle": 0, + "content": "portance across various fields, including autonomous driving [4, 15, 44]. With advancements in multimodal alignment technology [20, 34], object-level foundation models [28, 46, 48, 56] like Grounding DINO [28] and Florence-2 [48] have been developed to handle various tasks, including visual grounding and object detection. However, the massive data volume, rich semantic concepts, and large parameters significantly reduce the transparency [21, 22, 26] and interpretability [5, 11, 55] of these models. Since object-level tasks like autonomous driving demand high model reliability [9, 30, 45], building transparent and interpretable models is crucial for enhancing safety and trustworthy [6, 17, 21, 23, 24, 43], making it essential to interpret object-level foundation models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Some interpretable attribution algorithms for traditional object detection models have been proposed, including" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30042" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "gradient-based methods like ODAM [54] and perturbation-based methods like D-RISE [33]. However, as shown in the lower panel of Figure 1, these methods face limitations when applied to multimodal foundation models: (1) gradient-based methods may struggle to provide accurate visual localization explanations due to the fusion of visual and textual features in foundation models, and (2) perturbation-based approaches can introduce sampling artifacts, resulting in noisy saliency maps with limited fine-grained interpretability." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.248, + 0.485, + 0.669 + ], + "angle": 0, + "content": "To address these issues, we propose a novel interpretation mechanism for the object-level foundation model, as shown in Figure 1. Our goal is to generate a saliency map that explains the rationale behind the model's detection of a specific object, allowing fewer regions to be exposed for accurate detection. Additionally, removing these critical regions should quickly lead to detection failure, highlighting their importance in the model's decision. Specifically, we propose Visual Precision Search, which sparsifies the input region into a series of sub-regions using superpixel segmentation, and then ranks these sparse sub-regions to determine their importance for object-level decision-making. Our gradient-free method circumvents localization errors at the visual level that are caused by gradient backpropagation in vision-text fusion. Furthermore, we propose a novel submodular function, grounded in theoretical analysis and offering boundary guarantees for Visual Precision Search. This function identifies regions that enhance interpretability from two key perspectives: clues that support the model's accurate detection and regions with strong combinatorial effects. By employing region search to iteratively expand the set of sub-regions, our method more precisely identifies key sub-regions in the decision-making of object-level foundation models, which mitigates the issue of coarse explanations caused by noise in perturbation-based methods. In addition, our method effectively analyzes cases of grounding or detection failure, allowing us to observe which input-level factors influence the decision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We validated our method on the MS COCO [25], RefCOCO [18], and LVIS [13] datasets. The object-level foundation models evaluated include Grounding DINO, which uses a multimodal feature fusion architecture, and Florence2, which employs a multimodal large language model architecture. The tasks covered include interpreting visual grounding and object detection. In explaining the data behind correct model decisions, our method demonstrates that the faithfulness metric for Grounding DINO surpasses SOTA method D-RISE by \\(23.7\\%\\), \\(20.1\\%\\), and \\(31.6\\%\\) on MS COCO, RefCOCO, and LVIS, respectively. For Florence2, our method improves the faithfulness metric by \\(102.9\\%\\) and \\(66.9\\%\\) on MS COCO and RefCOCO, respectively, while also achieving SOTA performance on location metrics across all datasets. In interpreting the factors leading to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Grounding DINO's failures in visual grounding tasks, our method improves the Insertion metric by \\(42.9\\%\\) and the average highest score by \\(25.1\\%\\). It also identifies causes of detector misclassification and undetection, achieving Insertion metric improvements of \\(54.7\\%\\) and \\(42.7\\%\\) for misclassification, and \\(36.7\\%\\) and \\(64.3\\%\\) for undetection on MS COCO and LVIS, respectively." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.199, + 0.847, + 0.212 + ], + "angle": 0, + "content": "In summary, the contributions of this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.213, + 0.905, + 0.258 + ], + "angle": 0, + "content": "- We introduce Visual Precision Search, a new search-based mechanism for interpreting object-level foundation models through instance-specific saliency maps." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.905, + 0.332 + ], + "angle": 0, + "content": "- A novel submodular mechanism is constructed to enhance interpretability from two aspects: identifying clues that support accurate detection and highlighting regions with strong combinatorial effects, with an analysis of its theoretical boundaries for object-level tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.905, + 0.393 + ], + "angle": 0, + "content": "- We demonstrate our approach's generalizability across multimodal foundation models, including both non-LLM-based (Grounding DINO) and LLM-based detectors (Florence-2)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.394, + 0.905, + 0.469 + ], + "angle": 0, + "content": "- We validate our method on MS COCO, RefCOCO, and LVIS, achieving significant improvements in explaining object-level tasks. Additionally, we analyze grounding and detection failures, establishing a quantitative benchmark for this purpose." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.483, + 0.655, + 0.499 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Object Detection. Object detection involves locating objects in images and identifying their categories [24]. Early models typically used convolutional neural network (CNN) backbones, divided into two main types: two-stage and single-stage models. Two-stage methods, like Faster R-CNN [35] and Mask R-CNN [14], generate region proposals to identify potential foreground objects before refining localization and classification. In contrast, single-stage methods, such asYOLO [41] and FCOS [39], directly classify and regress on backbone features. Transformer architectures [40] have further advanced object detection, as exemplified by DETR [2], which uses a fully end-to-end Transformer-based approach. The rise of multimodal foundation models [20, 34] has enabled open-set detection tasks [24], allowing object grounding with textual input [7, 47]. GLIP [20] frames detection as a grounding problem, integrating vision and language through feature fusion in the neck module. Grounding DINO [28] uses multiphase vision-language fusion, achieving state-of-the-art (SOTA) referring expression comprehension (REC) [27]. Other methods explore multimodal large language models (MLLMs) for detection and grounding tasks. Florence-2 [48], for instance, introduces a prompt-based model that generates object categories and coordinates using text prompts directly. The diverse feature extractors, architectures, and outputs in object detection make it challenging to" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30043" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.461, + 0.107 + ], + "angle": 0, + "content": "interpret their decisions using a single unified approach." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.108, + 0.485, + 0.471 + ], + "angle": 0, + "content": "Object Detection Explanation. Explaining object detector decisions remains a largely unexplored area. Gudovskiy et al. [12] used Integrated Gradients (IG) [38] and SHAP [36] with SSD [29] classification scores for box-level attribution. Lee et al. [19] optimized masks on Mask R-CNN [14] to approximate original decisions with minimal pixels. Petsiuk et al. [33] proposed D-RISE, an adaptation of RISE [32], though it often introduces noise in saliency maps. Jiang et al. [16] introduced Nesterov-Accelerated iGOS++, which can be resource-intensive for large models. Zhao et al. [54] developed ODAM using Grad-CAM [37], but its effectiveness depends on network layer choices. SSGrad-CAM++ [49] generates saliency maps based on Grad-CAM++ [3]. VX-CODE [50] attributes important patches based on greedy search and SHAP [36]. While these methods aim to interpret detector errors, they largely rely on simple visualizations and empirical observations, often overlooking input-level feature confusion that can obscure human understanding. Masking confusing regions may aid in correcting model outputs [6]. Our paper introduces a black-box interpretation approach for object-level foundation models using Visual Precision Search, enhancing detection performance with fewer regions and identifying input-level failure causes to improve detection accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.482, + 0.182, + 0.497 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.553 + ], + "angle": 0, + "content": "This section presents our method for explaining the decisions of object-level foundation models. Figure 2 illustrates the overall framework of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.295, + 0.577 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Given an image \\(\\mathbf{I} \\in \\mathbb{R}^{h \\times w \\times 3}\\) and an object-level foundation model \\(f(\\cdot)\\), the output can be represented as \\(f(\\mathbf{I}) = \\{(b_i, c_i, s_i) \\mid i = 1, 2, \\dots, N\\}\\). Each tuple \\((b_i, c_i, s_i)\\) denotes the bounding box, class label, and confidence score for each detected object \\(i\\), where \\(N\\) is the maximum number of detection boxes. Tasks like object detection involve multiple boxes and categories, while visual grounding tasks like referring expression comprehension (REC) have only one box and category. Our goal is to generate a saliency map that explains the reasons behind the model's detection of a specific object, \\((b_i, c_i, s_i)\\). Ideally, this explanation would enable the correct detection of objects using fewer input regions, while also requiring the removal of only a few critical regions to invalidate the detector. To achieve this, we can sparsify the input region \\(V = \\{\\mathbf{I}_1^s, \\dots, \\mathbf{I}_m^s\\}\\), where \\(\\mathbf{I}_i^s\\) represents the \\(i\\)-th sub-region, \\(m\\) is the total number of sub-regions, and \\(\\mathbf{I} = \\sum_{i=1}^{m} \\mathbf{I}_i^s\\). We then frame the saliency map generation task as a subset selection problem for these input sub-regions. A set function \\(\\mathcal{F}(\\cdot)\\) is defined to assess interpretability by determining whether a given region is a key factor in the model's decision. Therefore, the" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.614, + 0.106 + ], + "angle": 0, + "content": "objectives are:" + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.109, + 0.905, + 0.133 + ], + "angle": 0, + "content": "\\[\n\\max _ {S \\subseteq V, | S | < k} \\mathcal {F} (S), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.142, + 0.905, + 0.188 + ], + "angle": 0, + "content": "where \\( k \\) denotes the maximum number of sub-regions. Therefore, the key to this problem lies in designing set function \\( \\mathcal{F} \\) and optimizing Eq. 1." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.2, + 0.731, + 0.215 + ], + "angle": 0, + "content": "3.2. Visual Precision Search" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.222, + 0.907, + 0.449 + ], + "angle": 0, + "content": "We propose a Visual Precision Search method for interpreting object-level models. To begin, the input region needs to be sparsified. We apply the SLICO superpixel segmentation algorithm [1] to divide the input into \\( m \\) sub-regions, \\( V = \\{\\mathbf{I}_1^s,\\dots ,\\mathbf{I}_m^s\\} \\). To solve Eq. 1, an \\( \\mathcal{NP} \\)-hard problem, we employ submodular optimization [10]. Since the saliency map requires all sub-regions to be ranked, \\( k \\) can be set to \\( |V| \\) to compute ordered subsets. When the set function \\( \\mathcal{F}(\\cdot) \\) satisfies the properties of diminishing returns1 and monotonic non-negative2, a greedy search guarantees an approximate optimal solution [8], with \\( \\mathcal{F}(S)\\geq (1 - 1 / e)\\mathcal{F}(S^{*}) \\), where \\( S^{*} \\) denotes the optimal solution and \\( S \\) denotes the greedy solution. We next design a set function to evaluate the interpretability score and rank the importance of interpretable regions for object-level tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.909, + 0.617 + ], + "angle": 0, + "content": "**Clue Score:** An essential aspect of interpretability is enabling the object-level foundation model to accurately locate and identify objects while using fewer regions. To assess the importance of subregions from this perspective, we introduce the object location box information, \\( \\pmb{b}_{\\mathrm{target}} \\), and the target category, \\( c \\), that requires explanation. Given a subregion \\( S \\), the object-level model outputs \\( N \\) instances, denoted as \\( f(S) = \\{(b_i, s_{c,i}) \\mid i = 1, \\dots, N\\} \\), where \\( s_{c,i} \\) represents the confidence of the \\( i \\)-th bounding box in predicting category \\( c \\). Then, the clue score of sub-region \\( S \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.63, + 0.907, + 0.656 + ], + "angle": 0, + "content": "\\[\ns _ {\\text {c l u e}} \\left(S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c\\right) = \\max _ {\\left(\\boldsymbol {b} _ {i}, s _ {c, i}\\right) \\in f (S)} \\operatorname {I o U} \\left(\\boldsymbol {b} _ {\\text {t a r g e t}}, \\boldsymbol {b} _ {i}\\right) \\cdot s _ {c, i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.907, + 0.774 + ], + "angle": 0, + "content": "where \\(\\mathrm{IoU}(\\cdot, \\cdot)\\) represents the Intersection over Union between two bounding boxes. Unlike Petsiuk et al. [33], which only considers high-confidence bounding boxes, our method includes all candidate boxes to avoid getting stuck in a local optimum during the search. By incorporating the clue score \\(s_{\\mathrm{clue}}\\), our method focuses on regions that strengthen the desired positional and semantic response." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.775, + 0.909, + 0.851 + ], + "angle": 0, + "content": "Collaboration Score: Some regions may exhibit strong combination effects, meaning they contribute effectively to model decisions only when paired with multiple specific sub-regions. Therefore, we introduce the collaboration score \\( s_{\\mathrm{colla}} \\) to assess sub-regions with high sensitivity to de" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.862, + 0.905, + 0.888 + ], + "angle": 0, + "content": "\\(^{1}\\)Diminishing returns: \\(\\forall S_A \\subseteq S_B \\subseteq V \\setminus \\alpha\\), \\(\\mathcal{F}(S_A \\cup \\{\\alpha\\}) - \\mathcal{F}(S_A) \\geq \\mathcal{F}(S_B \\cup \\{\\alpha\\}) - \\mathcal{F}(S_B)\\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.534, + 0.888, + 0.902, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)Monotonic non-negative: \\(\\forall S \\subseteq V \\setminus \\alpha\\), \\(\\mathcal{F}(S \\cup \\{\\alpha\\}) - \\mathcal{F}(S) \\geq 0\\)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.862, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30044" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.055, + 0.904, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.247, + 0.907, + 0.29 + ], + "angle": 0, + "content": "Figure 2. Framework of the proposed Visual Precision Search method for interpreting an object-level foundation model. The input is first sparsified into a set of sub-regions and then interpreted across different instances. A submodular function guides the search for significant sub-regions, updating the ordered subset iteratively, and ultimately generating the instance-level attribution map." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.298, + 0.206, + 0.31 + ], + "angle": 0, + "content": "cision outcomes:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.318, + 0.486, + 0.356 + ], + "angle": 0, + "content": "\\[\ns _ {\\text {c o l l a .}} (S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c) = 1 - \\max _ {(\\boldsymbol {b} _ {i}, s _ {c, i}) \\in f (V \\backslash S)} \\operatorname {I o U} (\\boldsymbol {b} _ {\\text {t a r g e t}}, \\boldsymbol {b} _ {i}) \\cdot s _ {c, i}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.429 + ], + "angle": 0, + "content": "this means that no detection box can accurately or confidently locate the object once the key sub-region is removed. This score serves as an effective guide in the initial search phase to identify subtle key regions that contribute to the final decision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.462 + ], + "angle": 0, + "content": "Submodular Function: The scores above are combined to construct a submodular function \\(\\mathcal{F}(\\cdot)\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.468, + 0.483, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} (S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c) = s _ {\\text {c l u e}} (S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c) + s _ {\\text {c o l l a .}} (S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.641 + ], + "angle": 0, + "content": "Saliency Map Generation: Using the above submodular function, a greedy search algorithm is applied to sort all sub-regions in \\(V\\), yielding an ordered subset \\(S\\). Introducing the submodular function enables the search algorithm to more precisely identify key visual regions for interpretation. Additionally, scoring the sub-regions is necessary to better explain the importance of each sub-region. We evaluate the salient difference between the two sub-regions by the marginal effect. The attribution score \\(\\mathcal{A}_i\\) for each sub-region \\(s_i\\) in \\(S\\) is assessed by:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.647, + 0.483, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {i} = \\left\\{ \\begin{array}{l l} b _ {\\text {b a s e}} & \\text {i f} i = 1, \\\\ \\mathcal {A} _ {i - 1} - \\left| \\mathcal {F} \\left(S _ {[ i ]}\\right) - \\mathcal {F} \\left(S _ {[ i - 1 ]}\\right) \\right| & \\text {i f} i > 1, \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.691, + 0.484, + 0.842 + ], + "angle": 0, + "content": "where \\( b_{\\mathrm{base}} \\) represents a baseline attribution score (such as 0 and won't affect results) for the first sub-region, and \\( S_{[i]} \\) denotes the set containing the top \\( i \\) sub-regions in \\( S \\). When a new sub-region is added, a small marginal increase suggests comparable importance to the previous sub-region. A negative marginal effect indicates a counterproductive impact, which can be assessed by its absolute value. Finally, \\( \\mathcal{A} \\) is normalized to obtain the saliency map of the sub-region. The detailed calculation process of the proposed Visual Precision Search algorithm is outlined in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.254, + 0.866 + ], + "angle": 0, + "content": "3.3. Theory Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.9 + ], + "angle": 0, + "content": "A theoretical analysis is conducted for our Visual Precision Search method." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.3, + 0.863, + 0.345 + ], + "angle": 0, + "content": "Algorithm 1: Visual Precision Search algorithm for generating interpretable saliency maps of object-level foundation models." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.348, + 0.884, + 0.374 + ], + "angle": 0, + "content": "Input: Image \\(\\mathbf{I} \\in \\mathbb{R}^{h \\times w \\times 3}\\), a division algorithm \\(\\mathrm{Div}(\\cdot)\\), the object location box \\(b_{\\text{target}}\\), and the target category \\(c\\)." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.374, + 0.85, + 0.386 + ], + "angle": 0, + "content": "Output: An ordered set \\(S\\) and a saliency map \\(\\mathcal{A} \\in \\mathbb{R}^{h \\times w}\\)." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.387, + 0.616, + 0.399 + ], + "angle": 0, + "content": "1 \\(V\\gets \\mathrm{Div}(\\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.399, + 0.881, + 0.422 + ], + "angle": 0, + "content": "2 \\(S\\gets \\emptyset\\) / \\(\\star\\) Initiate the operation of submodular subset selection \\(\\star/\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.422, + 0.605, + 0.434 + ], + "angle": 0, + "content": "3 \\(\\mathcal{A}_1\\gets b_{\\mathrm{base}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.434, + 0.642, + 0.446 + ], + "angle": 0, + "content": "4 for \\(i = 1\\) to \\(|V|\\) do" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.446, + 0.632, + 0.458 + ], + "angle": 0, + "content": "5 \\(S_{d}\\gets V\\backslash S;\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.458, + 0.799, + 0.47 + ], + "angle": 0, + "content": "\\(\\alpha \\gets \\arg \\max_{\\alpha \\in S_d}\\mathcal{F}(S\\cup \\{\\alpha \\} ,\\pmb {b}_{\\mathrm{target}},c)\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.47, + 0.645, + 0.482 + ], + "angle": 0, + "content": "7 \\(S\\gets S\\cup \\{\\alpha \\} ;\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.482, + 0.633, + 0.493 + ], + "angle": 0, + "content": "8 if \\(i > 1\\) then" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.493, + 0.8, + 0.508 + ], + "angle": 0, + "content": "9 \\(|\\quad | \\quad | \\quad \\mathcal{A}_i\\gets \\mathcal{A}_{i - 1} - |\\mathcal{F}(S_{[i]}) - \\mathcal{F}(S_{[i - 1]})|\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.509, + 0.558, + 0.518 + ], + "angle": 0, + "content": "10 end" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.519, + 0.641, + 0.531 + ], + "angle": 0, + "content": "11 return \\(S\\), \\(\\mathrm{norm}(\\mathcal{A})\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.641 + ], + "angle": 0, + "content": "Theorem 1 (Submodular Properties). Consider two subsets \\( S_A \\) and \\( S_B \\) in set \\( V \\), where \\( S_A \\subseteq S_B \\subseteq V \\). Given an element \\( \\alpha \\), where \\( \\alpha \\in V \\setminus S_b \\). Assuming that \\( \\alpha \\) is contributing to model interpretation, then, the function \\( \\mathcal{F}(\\cdot) \\) in Eq. 4 satisfies diminishing returns," + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.65, + 0.884, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} \\left(S _ {A} \\cup \\{\\alpha \\}\\right) - \\mathcal {F} \\left(S _ {A}\\right) \\geq \\mathcal {F} \\left(S _ {B} \\cup \\{\\alpha \\}\\right) - \\mathcal {F} \\left(S _ {B}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.677, + 0.906, + 0.708 + ], + "angle": 0, + "content": "and monotonic non-negative, \\(\\mathcal{F}(S_A\\cup \\{\\alpha \\}) - \\mathcal{F}(S)\\geq 0\\) and thus, \\(\\mathcal{F}\\) is a submodular function." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.716, + 0.878, + 0.732 + ], + "angle": 0, + "content": "Proof. Please see supplementary material for the proof." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.741, + 0.907, + 0.815 + ], + "angle": 0, + "content": "Remark 1 (Impact on Sparse Division). The quality of the search space is determined by sparse division, meaning that both the method of partitioning the input and the number of sub-regions play a crucial role in the faithfulness of the Visual Precision Search." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Remark 2 (Applicable Models). If the model provides bounding box positions and category confidence without filtering out low-confidence detection boxes, the interpretation results of the Visual Precision Search can achieve guaranteed optimal boundaries. Models with detection heads" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30045" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "are generally applicable; however, MLLM-based models may not directly output confidence scores, leaving room for improvement in the interpretation results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.153, + 0.226, + 0.17 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.179, + 0.282, + 0.195 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.201, + 0.485, + 0.428 + ], + "angle": 0, + "content": "Datasets. We evaluate the proposed method on three object-level datasets: MS COCO 2017 [25], LVIS V1 [13], and RefCOCO [18]. The MS COCO dataset, used for interpreting object detection with 80 classes, involves random sampling of 5 correctly detected (IoU \\(>0.5\\), correct category), 3 misclassified (IoU \\(>0.5\\), wrong category), and 3 undetected (IoU \\(<0.5\\) or low confidence) instances per class. The LVIS V1 dataset, with 1,203 categories including 337 rare ones (1-10 images), uses Grounding DINO for interpreting zero-shot detection on rare classes, analyzing 165 correctly detected, 152 misclassified, and 534 undetected samples. The RefCOCO dataset, for interpreting referring expression comprehension (REC), evaluates 500 correctly grounded (IoU \\(>0.5\\)) and 200 incorrectly grounded (IoU \\(<0.5\\)) samples per model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Baselines. The interpretation methods we compare include gradient-based approaches: Grad-CAM [37], SSGrad-CAM++[49], and ODAM[54], as well as perturbation-based methods: D-RISE [33] and D-HSIC [31]. D-RISE and D-HSIC are allowed to use all model outputs, including low-confidence detection boxes, to enhance the sufficiency of saliency estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.656 + ], + "angle": 0, + "content": "Implementation Details. We validate our method on two object-level foundation models, Grounding DINO [28] and Florence-2 [48], across object detection and referring expression comprehension (REC) tasks. To interpret model decisions, we set the ground truth bounding box and category as target box \\( \\pmb{b}_{\\mathrm{target}} \\) and category \\( c \\). By default, the SLICO superpixel segmentation algorithm divides the image into 100 sparse sub-regions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.276, + 0.683 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Faithfulness Metrics. We evaluate the faithfulness of interpretation results across the entire image using Insertion and Deletion scores from the detection task, following the setup by Petsiuk et al. [33]. These metrics reflect location and recognition accuracy. We also calculate classification and IoU scores for the box closest to the target explanation denoted as Insertion (class), Deletion (class), Insertion (IoU), and Deletion (IoU). Additionally, we compute the average highest confidence, defined as the highest class response for the box closest to the target with IoU \\(>0.5\\). For failure samples, we introduce the Explaining Successful Rate (ESR) to evaluate whether the saliency map can effectively guide the search within a limited region, enabling the model to correctly detect misclassified or low-confidence objects" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "(with IoU \\(>0.5\\) and classification confidence above the detector's default threshold) previously." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.122, + 0.907, + 0.199 + ], + "angle": 0, + "content": "Location Metrics. We report the Point Game [53] and Energy Point Game [42] metrics. Since these metrics typically assume that the model's decision is relevant to the target object [33], we calculate them only for interpretation results on samples the model has correctly detected." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.209, + 0.715, + 0.224 + ], + "angle": 0, + "content": "4.3. Faithfulness Analysis" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.231, + 0.795, + 0.246 + ], + "angle": 0, + "content": "4.3.1. Faithfulness on Grounding DINO" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.906, + 0.326 + ], + "angle": 0, + "content": "We begin by validating our approach in explaining decisions across various object-level tasks using Grounding DINO [28]. Table 1 presents the faithfulness results across different datasets and tasks, demonstrating the strong interpretative faithfulness of our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.538 + ], + "angle": 0, + "content": "In the object detection interpretation task on the MS COCO dataset, our method achieved state-of-the-art results, outperforming the D-RISE method by \\(23.7\\%\\) on the Insertion metric, \\(6.7\\%\\) on the Deletion metric, and \\(10.6\\%\\) on the average highest score. The Insertion and Deletion metrics also achieved top performance under both category and IoU conditions. Additionally, our method performed best in location metrics, indicating state-of-the-art performance in explaining general object detection tasks. We also found that gradient-based methods, such as ODAM, exhibit low faithfulness in explaining Grounding DINO decisions. This is primarily due to the fusion of text and visual modalities, which causes the gradient to be influenced by both, reducing its effectiveness in attributing to the visual modality alone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.659 + ], + "angle": 0, + "content": "In the referring expression comprehension (REC) interpretation task on the RefCOCO dataset, our method achieved state-of-the-art results across all evaluation metrics, surpassing the D-RISE method by \\(20.1\\%\\) on the Insertion metric, \\(22.1\\%\\) on the Deletion metric, and \\(4.4\\%\\) on the average highest score. These results confirm that our method maintains high faithfulness in explaining object description understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.795 + ], + "angle": 0, + "content": "We also interpret zero-shot object detection for rare classes on the LVIS V1 dataset. Our method achieved SOTA results in faithfulness metrics, surpassing D-RISE by \\(31.6\\%\\) on the Insertion metric, \\(4.2\\%\\) on the Deletion metric, and \\(15.9\\%\\) on the average highest score. Although our method ranks second to ODAM on Energy Point Game metrics, this is mainly due to Grounding DINO not being fully trained on rare LVIS classes; as a result, model responses may be influenced by other regions rather than the object." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.797, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Experimental results show that our method significantly outperforms other interpretability approaches, primarily due to the submodularity of our objective function, enabling precise evaluation of critical sub-regions. As shown in Figure 3, ODAM's saliency maps are diffuse, and D-RISE's maps are noisy, whereas our method sharply highlights essential sub-regions, capturing edges and class-specific fea" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "30046" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.052, + 0.908, + 0.093 + ], + "angle": 0, + "content": "Table 1. Evaluation of faithfulness metrics (Deletion, Insertion AUC scores, and average highest score) and location metrics (Point Game and Energy Point Game) on the MS-COCO, RefCOCO, and LVIS V1 (rare) validation sets for correctly detected or grounded samples using Grounding DINO." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.097, + 0.905, + 0.309 + ], + "angle": 0, + "content": "
DatasetsMethodsFaithfulness MetricsLocation Metrics
Ins. (↑)Del. (↓)Ins. (class) (↑)Del. (class) (↓)Ins. (IoU) (↑)Del. (IoU) (↓)Ave. high. score (↑)Point Game (↑)Energy PG (↑)
MS COCO [25] (Detection task)Grad-CAM [37]0.24360.15260.30640.20060.62290.53240.59040.17460.1463
SSGrad-CAM++ [49]0.21070.17780.26390.23140.59810.55110.58860.19050.1293
D-RISE [33]0.44120.04020.50810.08860.83960.36420.62150.94970.1850
D-HSIC [31]0.37760.04390.43820.09030.83010.33010.58620.73280.1861
ODAM [54]0.31030.05190.36550.08940.78690.39840.58650.54310.2034
Ours0.54590.03750.62040.08820.85810.33000.68730.98940.2046
RefCOCO [18] (REC task)Grad-CAM [37]0.37490.42370.46580.51940.75160.76850.74810.23800.2171
SSGrad-CAM++ [49]0.41130.39250.50080.48510.77000.75880.75610.28200.2262
D-RISE [33]0.61780.16050.70330.33960.86060.51640.84710.94000.2870
D-HSIC [31]0.54910.18460.62950.35090.85040.51200.77390.79000.3190
ODAM [54]0.47780.27180.56200.37570.82170.66410.74250.63200.3529
Ours0.74190.12500.80800.24570.90500.51030.88420.94600.3566
LVIS V1 (rare) [13] (Zero-shot det. task)Grad-CAM [37]0.12530.12940.18010.18140.56570.59100.35490.11510.0941
SSGrad-CAM++ [49]0.12530.12540.17650.17750.58000.56910.35040.10910.0931
D-RISE [33]0.28080.02890.33480.08350.83030.31740.42890.96970.1462
D-HSIC [31]0.24170.03530.29120.09280.81870.35500.40440.83030.1730
ODAM [54]0.20090.04100.24780.08440.77600.40820.36940.60610.2050
Ours0.36950.02770.42750.07990.84790.32420.49690.97580.1785
" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.313, + 0.898, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.509, + 0.908, + 0.551 + ], + "angle": 0, + "content": "Figure 3. Visualization results of Grounding DINO for interpreting object localization tasks on the MS COCO, RefCOCO, and LVIS datasets. The first column shows the saliency map with the ground truth box and label. The second column presents detection within the limited search region, and the third columns display the insertion curves." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.565, + 0.318, + 0.58 + ], + "angle": 0, + "content": "tures that enhance interpretability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.321, + 0.605 + ], + "angle": 0, + "content": "4.3.2. Faithfulness on Florence-2" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.702 + ], + "angle": 0, + "content": "We further validated our approach on Florence-2 [48], a foundation model lacking object confidence scores. This limitation requires our search to rely solely on IoU as a guiding metric, excluding comparison with gradient-based methods. As shown in Table 2, our method achieves state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.884 + ], + "angle": 0, + "content": "In the object detection interpretability task on the MS COCO dataset, our method outperforms D-RISE by \\(3.8\\%\\) on the Insertion metric and achieves a substantial improvement of \\(50.7\\%\\) on the Deletion metric. Additionally, our method enhances the Point Game and Energy Point Game metrics by \\(8.3\\%\\) and \\(60.7\\%\\), respectively. Due to the absence of object confidence scores in Florence-2, it becomes challenging to identify areas that quickly invalidate the model's detection results. Compared to perturbation-based methods, our approach offers significant advantages; it relies more directly on the object itself, resulting in stronger faithfulness to the model's detection behavior." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We also interpret referring expression comprehension" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.906, + 0.656 + ], + "angle": 0, + "content": "(REC) interpretation task on the RefCOCO dataset, outperforming D-RISE by \\(6.1\\%\\) and \\(66.9\\%\\) in Insertion and Deletion metrics, respectively. Our approach also achieved state-of-the-art localization results, with a \\(14.6\\%\\) improvement in the Energy Point Game, demonstrating strong interpretability in accurately pinpointing object information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.906, + 0.745 + ], + "angle": 0, + "content": "We visualized some results of our method in Figure 4, illustrating that it effectively identifies key information about the object's location and categorical description. While D-RISE performed well on the Insertion metric, its attribution results were diffuse and lacked reliability, and it did not perform as well on the Deletion metric." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.754, + 0.882, + 0.771 + ], + "angle": 0, + "content": "4.4. Interpreting Failures in Object-Level Tasks" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.777, + 0.745, + 0.792 + ], + "angle": 0, + "content": "4.4.1. Interpreting REC Failures" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Failures in the REC task primarily arise from visual grounding errors, often due to interference from other objects in the scene that become entangled with the text. Removing this distracting visual information can clarify grounding results and improve error understanding at the visual input level. To achieve this, we specify the correct location as the target box, denoted as \\( \\pmb{b}_{\\mathrm{target}} \\), for input attribution. Ta" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30047" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.146 + ], + "angle": 0, + "content": "Table 2. Evaluation of faithfulness metrics (Deletion and Insertion AUC scores) and location metrics (Point Game and Energy Point Game) on the MS COCO and RefCOCO validation sets for correctly detected and grounded samples using Florence-2." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.148, + 0.485, + 0.237 + ], + "angle": 0, + "content": "
DatasetsMethodsFaithfulness MetricsLocation Metrics
Insertion (↑)Deletion (↓)Point Game (↑)Energy PG (↑)
MS COCO [25] (Detection task)D-RISE [33]0.74770.09720.88500.1568
D-HSIC [31]0.53450.27300.29250.0862
Ours0.77590.04790.95830.2519
RefCOCO [18] (REC task)D-RISE [33]0.79220.35050.84800.2464
D-HSIC [31]0.76390.35600.69800.2754
Ours0.84090.11590.86600.3927
" + }, + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.248, + 0.42, + 0.259 + ], + "angle": 0, + "content": "MS COCO (Object Detection Interpretation)" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.26, + 0.48, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.377, + 0.421, + 0.388 + ], + "angle": 0, + "content": "RefCOCO (Visual Grounding Interpretation)" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.388, + 0.48, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.537 + ], + "angle": 0, + "content": "Figure 4. Visualization results of Florence-2 for interpreting object localization tasks on the MS COCO and RefCOCO datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.551, + 0.483, + 0.673 + ], + "angle": 0, + "content": "ble 3 shows the attribution results, where gradient-based methods exhibit limited performance in faithfulness metrics. Compared to D-RISE, our approach improves Insertion by \\(42.9\\%\\), class score-based Insertion by \\(25.1\\%\\), and average highest score by \\(14.9\\%\\). Figure 5 illustrates model error interpretations, with the cyan-highlighted region indicating decision errors due to visual input interference that leads the foundation model astray." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.683, + 0.355, + 0.699 + ], + "angle": 0, + "content": "4.4.2. Interpreting Detection Failures" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Detection failures include misclassification and undetected objects. While previous works [33, 54] used visualization to explain errors, it lacked quantitative analysis, resulting in lower trustworthiness. We used saliency maps to identify input-level interference causing detection failures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Misclassified Interpretation. As shown in Table 4, On the MS COCO dataset, our method outperforms D-RISE, improving Insertion by \\(54.7\\%\\), Insertion (class) by \\(49.1\\%\\), average highest score by \\(27.4\\%\\), and explaining success rate (ESR) by \\(19.47\\%\\). Similarly, on the LVIS dataset, our method shows advantages with improvements of \\(42.7\\%\\) in Insertion, \\(33.0\\%\\) in Insertion (class), \\(24.8\\%\\) in average highest score, and \\(24.34\\%\\) in ESR. This improvement is pri" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 3. Insertion AUC scores and the average highest score on the RefCOCO validation sets for or the samples with incorrect localization in visual grounding using Grounding DINO." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.133, + 0.895, + 0.219 + ], + "angle": 0, + "content": "
DatasetsMethodsFaithfulness Metrics
Ins. (†)Ins. (class) (†)Ave. high. score (†)
RefCOCO [18] (REC task)Grad-CAM [37]0.15360.27940.3295
SSGrad-CAM++ [49]0.15900.28370.3266
D-RISE [33]0.34860.47870.6096
D-HSIC [31]0.22740.34880.4495
ODAM [54]0.17930.30010.3453
Ours0.49810.59900.7007
" + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.23, + 0.61, + 0.24 + ], + "angle": 0, + "content": "Attribution Map" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.24, + 0.639, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.231, + 0.75, + 0.24 + ], + "angle": 0, + "content": "Searched Region" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.24, + 0.767, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.225, + 0.901, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.312, + 0.905, + 0.34 + ], + "angle": 0, + "content": "Figure 5. Visualization of the method for discovering what causes the Grounding DINO incorrectly grounded on RefCOCO." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.343, + 0.905, + 0.385 + ], + "angle": 0, + "content": "Table 4. Insertion AUC scores, average highest score, and explaining successful rate (ESR) on the MS-COCO and the LVIS validation sets for misclassified samples using Grounding DINO." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.387, + 0.905, + 0.516 + ], + "angle": 0, + "content": "
DatasetsMethodsFaithfulness Metrics
Ins. (↑)Ins. (class) (↑)Ave. high. score (↑)ESR (↑)
MS COCO [25] (Detection task)Grad-CAM [37]0.10910.14780.310238.38%
SSGrad-CAM++ [49]0.09600.13360.295233.51%
D-RISE [33]0.21700.26610.360350.26%
D-HSIC [31]0.17710.21610.314334.59%
ODAM [54]0.11290.14860.286932.97%
Ours0.33570.39670.459169.73%
LVIS V1 (rare) [13] (Zero-shot det. task)Grad-CAM [37]0.05030.08910.156412.50%
SSGrad-CAM++ [49]0.05740.09460.158011.84%
D-RISE [33]0.12450.16470.208828.95%
D-HSIC [31]0.09630.12470.174816.45%
ODAM [54]0.05750.09540.15209.21%
Ours0.17760.21900.260653.29%
" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.522, + 0.619, + 0.647 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.522, + 0.71, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.522, + 0.8, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.522, + 0.902, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.654, + 0.907, + 0.708 + ], + "angle": 0, + "content": "Figure 6. Visualization of our method reveals the causes of Grounding DINO misclassifications on MS COCO and LVIS. The cyan region in the saliency map highlights the regions responsible for the model's misclassification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.825 + ], + "angle": 0, + "content": "marily due to our proposed Visual Precision Search, which confines attribution to the target negative response region, effectively refining the model's detection results. Figure 6 shows that the background surrounding the object interferes with the model's decision-making. Improving the model by refining the contextual relationship between foreground and background may be a promising direction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Undetected Interpretation. Low object confidence may result from both the model's feature representation and confounding input-level factors. We analyze causes of missed detections, with Table 5 showing that our method achieves SOTA performance across metrics. On MS" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30048" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.07, + 0.483, + 0.111 + ], + "angle": 0, + "content": "Table 5. Insertion, average highest score, and explaining successful rate (ESR) on the MS-COCO and the LVIS V1 (rare) validation sets for missed detection samples using Grounding DINO." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.11, + 0.485, + 0.24 + ], + "angle": 0, + "content": "
DatasetsMethodsFaithfulness Metrics
Ins. (↑)Ins. (class) (↑)Ave. high. score (↑)ESR (↑)
MS COCO [25] (Detection task)Grad-CAM [37]0.07600.13210.215316.44%
SSGrad-CAM++ [49]0.06710.11510.212416.44%
D-RISE [33]0.15380.22600.256426.94%
D-HSIC [31]0.11010.17160.194513.56%
ODAM [54]0.07450.13500.203713.78%
Ours0.21020.30110.301441.33%
LVIS V1 (rare) [13] (Zero-shot det. task)Grad-CAM [37]0.02910.06890.09015.43%
SSGrad-CAM++ [49]0.02920.06800.08975.24%
D-RISE [33]0.07030.11840.131218.73%
D-HSIC [31]0.05160.09200.116813.48%
ODAM [54]0.02830.07160.08514.68%
Ours0.11550.18860.178430.15%
" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.269, + 0.231, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.269, + 0.348, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.269, + 0.48, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.437, + 0.483, + 0.492 + ], + "angle": 0, + "content": "Figure 7. Visualization of our method reveals the causes of Grounding DINO undetected on MS COCO and LVIS. The cyan region in the saliency map highlights the regions responsible for the model's detection failure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.637 + ], + "angle": 0, + "content": "COCO, it outperforms D-RISE with a \\(36.7\\%\\) improvement in Insertion and a \\(14.39\\%\\) gain in ESR, while on LVIS, it improves Insertion by \\(64.3\\%\\) and ESR by \\(11.42\\%\\). Figure 7 visualizes our method's explanations for undetected instances, revealing that errors may arise from challenges in distinguishing similar objects (e.g., the bear in the first row) and environmental influence on detection (e.g., the corkboard in the second row). These insights highlight current model limitations, offering directions for improvement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.65, + 0.245, + 0.667 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Ablation of the Submodular Function. We analyze the effectiveness of our submodular function design on Grounding DINO using correctly detected samples from MS COCO. Table 6 shows the results. The Clue Score identifies regions beneficial for detection, improving Insertion and average highest score, while the Collaboration Score pinpoints sensitive regions, enhancing the Deletion metric by causing the detector to fail more quickly. Combining these scores enables our method to achieve optimal results across indicators, demonstrating the effectiveness of each score function within the submodular function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Ablation on Divided Sub-region Number. The sub-region number affects search space quality. Using Grounding DINO, we assess its impact on correctly predicted and misclassified samples in MS COCO. Figures 8A-B show" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.07, + 0.905, + 0.097 + ], + "angle": 0, + "content": "Table 6. Ablation study on function score components for Grounding DINO on the MS COCO validation set." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.101, + 0.882, + 0.165 + ], + "angle": 0, + "content": "
Clue Score (Eq. 2)Colla. Score (Eq. 3)Insertion (↑)Faithfulness Metrics
Deletion (↓)Ave. high. score (↑)
✗✓0.36320.03780.5967
✓✗0.53700.07990.6864
✓✓0.54590.03750.6873
" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.181, + 0.645, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.181, + 0.772, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.181, + 0.901, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.262, + 0.645, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.262, + 0.773, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.262, + 0.901, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.344, + 0.905, + 0.428 + ], + "angle": 0, + "content": "Figure 8. Ablation on the number of sub-regions. For correct detections, panels \\(\\mathbf{A}\\) and \\(\\mathbf{B}\\) show the effect of sub-region count on Insertion and average highest score, respectively. Panel \\(\\mathbf{C}\\) presents the inference time for different sub-region counts. For misclassified samples, panels \\(\\mathbf{D},\\mathbf{E},\\) and \\(\\mathbf{F}\\) display the effects on Insertion, average highest score, and explanation success rate." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.651 + ], + "angle": 0, + "content": "that when the model predicts correctly, increasing the number of sub-regions leads to higher Insertion and average highest scores, indicating that finer divisions enhance the faithfulness of search results. Additionally, Figures 8D-F show that, for misclassified samples, Insertion, average highest score, and ESR also improve as the number of sub-regions increases. Figure 8C shows inference times for our search algorithm on an RTX 3090 compared to D-RISE across various subregion counts. Increasing sub-regions improves faithfulness but also rapidly increases inference time. Future work will aim to increase subregion count and inference speed without sacrificing attribution performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.687, + 0.634, + 0.703 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose an interpretable attribution method specifically tailored for object-level foundation models, called the Visual Precision Search method, which introduces a novel submodular mechanism that combines a clue score and a collaboration score. This method achieves enhanced interpretability with fewer search regions. Experiments on RefCOCO, MS COCO, and LVIS demonstrate that our approach improves object-level task interpretability over state-of-the-art methods for Grounding DINO and Florence-2 across various evaluation metrics. Furthermore, our method effectively interprets failures in visual grounding and object detection tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30049" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work was supported by the Shenzhen Science and Technology Program (No. KQTD20221101093559018), National Natural Science Foundation of China (No. 62372448, 62025604, 62306308, 62441619), Foundation of Key Laboratory of Education Informatization for Nationalities (YNNU), Ministry of Education (No. EIN2024B004)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.221, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[1] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Trans. Pattern Anal. Mach. Intell., 34(11):2274-2282, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.317, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Eur. Conf. Comput. Vis. (ECCV), pages 213-229, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.375, + 0.483, + 0.444 + ], + "angle": 0, + "content": "[3] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks. In IEEE Winter Conf. Appl. Comput. Vis. (WACV), pages 839-847, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[4] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.484, + 0.57 + ], + "angle": 0, + "content": "[5] Ruoyu Chen, Jingzhi Li, Hua Zhang, Changchong Sheng, Li Liu, and Xiaochun Cao. Sim2Word: Explaining similarity with representative attribute words via counterfactual explanations. ACM Trans. Multim. Comput. Commun. Appl., 19 (6):1-22, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.573, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[6] Ruoyu Chen, Hua Zhang, Siyuan Liang, Jingzhi Li, and Xiaochun Cao. Less is more: Fewer interpretable region via submodular subset selection. In Int. Conf. Learn. Represent. (ICLR), 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[7] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 16901-16911, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.687, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[8] Jack Edmonds. Submodular functions, matroids, and certain polyhedra. Combinatorial Structures and Their Applications, pages 69-87, 1970. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.73, + 0.483, + 0.798 + ], + "angle": 0, + "content": "[9] Di Feng, Ali Harakeh, Steven L Waslander, and Klaus Dietmayer. A review and comparative study on probabilistic object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(8):9961-9980, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.801, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[10] Satoru Fujishige. Submodular functions and optimization. Elsevier, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[11] Yossi Gandelsman, Alexei A Efros, and Jacob Steinhardt. Interpreting clip's image representation via text-based decomposition. In Int. Conf. Learn. Represent. (ICLR), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Denis Gudovskiy, Alec Hodgkinson, Takuya Yamaguchi, Yasunori Ishii, and Sotaro Tsukizawa. Explain to fix: A" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.246, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "framework to interpret and correct dnn object detector predictions. arXiv preprint arXiv:1811.08011, 2018.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5356-5364, 2019. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. IEEE Trans. Pattern Anal. Mach. Intell., 42(2):386-397, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 17853-17862, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[16] Mingqi Jiang, Saeed Khorram, and Fuxin Li. Diverse explanations for object detectors with nesterov-accelerated igos++. In Brit. Mach. Vis. Conf. (BMVC), pages 188–189, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[17] Mingqi Jiang, Saeed Khorram, and Li Fuxin. Comparing the decision-making mechanisms by transformers and cnns via explanation methods. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 9546-9555, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[18] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. ReferItGame: Referring to objects in photographs of natural scenes. In Proc. Conf. Empir. Methods Nat. Lang. Process. (EMNLP), pages 787-798, 2014. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[19] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. BBAM: Bounding box attribution map for weakly supervised semantic and instance segmentation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 2643-2652, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[20] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 10965-10975, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[21] Siyuan Liang, Baoyuan Wu, Yanbo Fan, Xingxing Wei, and Xiaochun Cao. Parallel rectangle flip attack: A query-based black-box attack against object detection. In Int. Conf. Comput. Vis. (ICCV), pages 7697-7707, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[22] Siyuan Liang, Longkang Li, Yanbo Fan, Xiaojun Jia, Jingzhi Li, Baoyuan Wu, and Xiaochun Cao. A large-scale multiple-objective method for black-box attack against object detection. In Eur. Conf. Comput. Vis. (ECCV), pages 619-636, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[23] Siyuan Liang, Aishan Liu, Jiawei Liang, Longkang Li, Yang Bai, and Xiaochun Cao. Imitated detectors: Stealing knowledge of black-box object detectors. In ACM Int. Conf. Multimedia, pages 4839-4847, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[24] Siyuan Liang, Wei Wang, Ruoyu Chen, Aishan Liu, Boxi Wu, Ee-Chien Chang, Xiaochun Cao, and Dacheng Tao. Object detectors in the open environment: Challenges, solutions, and outlook. arXiv preprint arXiv:2403.16271, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30050" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Zitnick. Microsoft COCO: Common objects in context. In Eur. Conf. Comput. Vis. (ECCV), pages 740-755, 2014. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[26] Aishan Liu, Jun Guo, Jiakai Wang, Siyuan Liang, Renshuai Tao, Wenbo Zhou, Cong Liu, Xianglong Liu, and Dacheng Tao. {X-Adv}: Physical adversarial object attacks against x-ray prohibited item detection. In USENIX Security Symposium, pages 3781–3798, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[27] Jingyu Liu, Liang Wang, and Ming-Hsuan Yang. Referring expression generation and comprehension via attributes. In Int. Conf. Comput. Vis. (ICCV), pages 4856-4864, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[28] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. In Eur. Conf. Comput. Vis. (ECCV), 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[29] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. SSD: Single shot multibox detector. In Eur. Conf. Comput. Vis. (ECCV), pages 21-37, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[30] Dimity Miller, Feras Dayoub, Michael Milford, and Niko Sünderhauf. Evaluating merging strategies for sampling-based uncertainty techniques in object detection. In Proc. IEEE Int. Conf. Robot. Autom. (ICRA), pages 2348-2354, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.482, + 0.497 + ], + "angle": 0, + "content": "[31] Paul Novello, Thomas Fel, and David Vigouroux. Making sense of dependence: Efficient black-box explanations using dependence measure. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 4344-4357, 2022. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[32] Vitali Petsiuk, Abir Das, and Kate Saenko. RISE: Randomized input sampling for explanation of black-box models. In Brit. Mach. Vis. Conf. (BMVC), page 151, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[33] Vitali Petsiuk, Rajiv Jain, Varun Manjunatha, Vlad I Morariu, Ashutosh Mehra, Vicente Ordonez, and Kate Saenko. Black-box explanation of object detectors via saliency maps. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 11443-11452, 2021. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn. (ICML), pages 8748-8763, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell., 39(6):1137-1149, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.79 + ], + "angle": 0, + "content": "[36] M Scott, Lee Su-In, et al. A unified approach to interpreting model predictions. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 4765-4774, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[37] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis., 128:336-359, 2020. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[38] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Int. Conf. Mach. Learn. (ICML), pages 3319-3328, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[39] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: A simple and strong anchor-free object detector. IEEE Trans. Pattern Anal. Mach. Intell., 44(4):1922-1933, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 5998-6008, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[41] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 7464-7475, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[42] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[43] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable adversarial attacks for image and video object detection. In IJCAI, pages 954-960, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[44] Licheng Wen, Xuemeng Yang, Daocheng Fu, Xiaofeng Wang, Pinlong Cai, Xin Li, MA Tao, Yingxuan Li, XU Linran, Dengke Shang, et al. On the road with gpt-4v (ision): Explorations of utilizing visual-language model as autonomous driving agent. In ICLR 2024 Workshop on Large Language Model (LLM) Agents, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[45] Samuel Wilson, Tobias Fischer, Feras Dayoub, Dimity Miller, and Niko Sunderhauf. Safe: Sensitivity-aware features for out-of-distribution object detection. In Int. Conf. Comput. Vis. (ICCV), pages 23565-23576, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[46] Junfeng Wu, Yi Jiang, Qihao Liu, Zehuan Yuan, Xiang Bai, and Song Bai. General object foundation model for images and videos at scale. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3783-3795, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[47] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, et al. Towards open vocabulary learning: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.717 + ], + "angle": 0, + "content": "[48] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4818–4829, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[49] Toshinori Yamauchi. Spatial Sensitive Grad-CAM++: Improved visual explanation for object detectors via weighted combination of gradient map. In CVPR Workshop, pages 8164-8168, 2024. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[50] Toshinori Yamauchi, Hiroshi Kera, and Kazuhiko Kawamoto. Explaining object detectors via collective contribution of pixels. arXiv preprint arXiv:2412.00666, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[51] Lewei Yao, Renjie Pi, Jianhua Han, Xiaodan Liang, Hang Xu, Wei Zhang, Zhenguo Li, and Dan Xu. Detclipv3: Towards versatile generative open-vocabulary object detection. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 27391-27401, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30051" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[52] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. Int. J. Comput. Vis., pages 1-19, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[53] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. Int. J. Comput. Vis., 126 (10):1084-1102, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[54] Chenyang Zhao, Janet H Hsiao, and Antoni B Chan. Gradient-based instance-specific visual explanations for object specification and object discrimination. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[55] Chenyang Zhao, Kun Wang, Xingyu Zeng, Rui Zhao, and Antoni B Chan. Gradient-based visual explanation for transformer-based clip. In Int. Conf. Mach. Learn. (ICML), pages 61072-61091, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[56] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 15116-15127, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[57] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "30052" + } + ] +] \ No newline at end of file diff --git a/2025/Interpreting Object-level Foundation Models via Visual Precision Search/396da971-babe-4c65-8e62-f7e2e9a705a9_origin.pdf b/2025/Interpreting Object-level Foundation Models via Visual Precision Search/396da971-babe-4c65-8e62-f7e2e9a705a9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0e78505b9ada9a3bb7ee98cf8148d07b36c0ad55 --- /dev/null +++ b/2025/Interpreting Object-level Foundation Models via Visual Precision Search/396da971-babe-4c65-8e62-f7e2e9a705a9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8919759743159e09d453c037713a8589ccb380fd50ef786ad13a8abc3a9254a3 +size 8034418 diff --git a/2025/Interpreting Object-level Foundation Models via Visual Precision Search/full.md b/2025/Interpreting Object-level Foundation Models via Visual Precision Search/full.md new file mode 100644 index 0000000000000000000000000000000000000000..afe59cb919dc35a1a78cc23d6a11f45ed2e7cf39 --- /dev/null +++ b/2025/Interpreting Object-level Foundation Models via Visual Precision Search/full.md @@ -0,0 +1,377 @@ +# Interpreting Object-level Foundation Models via Visual Precision Search + +Ruoyu Chen $^{1,2}$ , Siyuan Liang $^{3}$ , Jingzhi Li $^{1,2,7}$ , Shiming Liu $^{4}$ , Maosen Li $^{5}$ , Zhen Huang $^{6}$ , Hua Zhang $^{1,2,*}$ , and Xiaochun Cao $^{8,*}$ + +$^{1}$ Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100093, China +$^{2}$ School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China + +$^{3}$ School of Computing, NUS $^{4}$ RAMS Lab, Huawei Technologies Co., Ltd. $^{5}$ IAS BU, Huawei Technologies Co., Ltd. + +6College of Computer, NUDT 7Key Lab. of Edu. Inf. for Nationalities (YNNU), Ministry of Education, Kunming, China + +$^{8}$ School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen 518107, China + +chenruoyu@ie.ac.cn pandaliang521@gmail.com {lijingzhi,zhanghua}@ie.ac.cn + +{liushiming3, limaosen2}@huawei.com huangzhen@nudt.edu.cn caoxiaochun@mail.sysu.edu.cn + +# Abstract + +Advances in multimodal pre-training have propelled object-level foundation models, such as Grounding DINO and Florence-2, in tasks like visual grounding and object detection. However, interpreting these models' decisions has grown increasingly challenging. Existing interpretable attribution methods for object-level task interpretation have notable limitations: (1) gradient-based methods lack precise localization due to visual-textual fusion in foundation models, and (2) perturbation-based methods produce noisy saliency maps, limiting fine-grained interpretability. To address these, we propose a Visual Precision Search method that generates accurate attribution maps with fewer regions. Our method bypasses internal model parameters to overcome attribution issues from multimodal fusion, dividing inputs into sparse sub-regions and using consistency and collaboration scores to accurately identify critical decision-making regions. We also conducted a theoretical analysis of the boundary guarantees and scope of applicability of our method. Experiments on RefCOCO, MS COCO, and LVIS show our approach enhances object-level task interpretability over SOTA for Grounding DINO and Florence-2 across various evaluation metrics, with faithfulness gains of $23.7\%$ , $31.6\%$ , and $20.1\%$ on MS COCO, LVIS, and RefCOCO for Grounding DINO, and $50.7\%$ and $66.9\%$ on MS COCO and RefCOCO for Florence-2. Additionally, our method can interpret failures in visual grounding and object detection tasks, surpassing existing methods across multiple evaluation metrics. The code is released at https://github.com/RuoyuChen10/VPS. + +# 1. Introduction + +Understanding object information in images, such as object detection [7, 14, 24, 51, 52, 57], is a crucial and enduring challenge in computer vision, holding significant im + +![](images/3fb944160443eb79d76d1aa87654ccef39bb1ca0bdc30adeef9458a556cb074f.jpg) +Figure 1. Illustration of our Visual Precision Search interpretation method, which more precisely identifies key sub-regions in object-level foundation model decision-making compared to gradient-based and perturbation-based methods. + +portance across various fields, including autonomous driving [4, 15, 44]. With advancements in multimodal alignment technology [20, 34], object-level foundation models [28, 46, 48, 56] like Grounding DINO [28] and Florence-2 [48] have been developed to handle various tasks, including visual grounding and object detection. However, the massive data volume, rich semantic concepts, and large parameters significantly reduce the transparency [21, 22, 26] and interpretability [5, 11, 55] of these models. Since object-level tasks like autonomous driving demand high model reliability [9, 30, 45], building transparent and interpretable models is crucial for enhancing safety and trustworthy [6, 17, 21, 23, 24, 43], making it essential to interpret object-level foundation models. + +Some interpretable attribution algorithms for traditional object detection models have been proposed, including + +gradient-based methods like ODAM [54] and perturbation-based methods like D-RISE [33]. However, as shown in the lower panel of Figure 1, these methods face limitations when applied to multimodal foundation models: (1) gradient-based methods may struggle to provide accurate visual localization explanations due to the fusion of visual and textual features in foundation models, and (2) perturbation-based approaches can introduce sampling artifacts, resulting in noisy saliency maps with limited fine-grained interpretability. + +To address these issues, we propose a novel interpretation mechanism for the object-level foundation model, as shown in Figure 1. Our goal is to generate a saliency map that explains the rationale behind the model's detection of a specific object, allowing fewer regions to be exposed for accurate detection. Additionally, removing these critical regions should quickly lead to detection failure, highlighting their importance in the model's decision. Specifically, we propose Visual Precision Search, which sparsifies the input region into a series of sub-regions using superpixel segmentation, and then ranks these sparse sub-regions to determine their importance for object-level decision-making. Our gradient-free method circumvents localization errors at the visual level that are caused by gradient backpropagation in vision-text fusion. Furthermore, we propose a novel submodular function, grounded in theoretical analysis and offering boundary guarantees for Visual Precision Search. This function identifies regions that enhance interpretability from two key perspectives: clues that support the model's accurate detection and regions with strong combinatorial effects. By employing region search to iteratively expand the set of sub-regions, our method more precisely identifies key sub-regions in the decision-making of object-level foundation models, which mitigates the issue of coarse explanations caused by noise in perturbation-based methods. In addition, our method effectively analyzes cases of grounding or detection failure, allowing us to observe which input-level factors influence the decision. + +We validated our method on the MS COCO [25], RefCOCO [18], and LVIS [13] datasets. The object-level foundation models evaluated include Grounding DINO, which uses a multimodal feature fusion architecture, and Florence2, which employs a multimodal large language model architecture. The tasks covered include interpreting visual grounding and object detection. In explaining the data behind correct model decisions, our method demonstrates that the faithfulness metric for Grounding DINO surpasses SOTA method D-RISE by $23.7\%$ , $20.1\%$ , and $31.6\%$ on MS COCO, RefCOCO, and LVIS, respectively. For Florence2, our method improves the faithfulness metric by $102.9\%$ and $66.9\%$ on MS COCO and RefCOCO, respectively, while also achieving SOTA performance on location metrics across all datasets. In interpreting the factors leading to + +Grounding DINO's failures in visual grounding tasks, our method improves the Insertion metric by $42.9\%$ and the average highest score by $25.1\%$ . It also identifies causes of detector misclassification and undetection, achieving Insertion metric improvements of $54.7\%$ and $42.7\%$ for misclassification, and $36.7\%$ and $64.3\%$ for undetection on MS COCO and LVIS, respectively. + +In summary, the contributions of this paper are: + +- We introduce Visual Precision Search, a new search-based mechanism for interpreting object-level foundation models through instance-specific saliency maps. +- A novel submodular mechanism is constructed to enhance interpretability from two aspects: identifying clues that support accurate detection and highlighting regions with strong combinatorial effects, with an analysis of its theoretical boundaries for object-level tasks. +- We demonstrate our approach's generalizability across multimodal foundation models, including both non-LLM-based (Grounding DINO) and LLM-based detectors (Florence-2). +- We validate our method on MS COCO, RefCOCO, and LVIS, achieving significant improvements in explaining object-level tasks. Additionally, we analyze grounding and detection failures, establishing a quantitative benchmark for this purpose. + +# 2. Related Work + +Object Detection. Object detection involves locating objects in images and identifying their categories [24]. Early models typically used convolutional neural network (CNN) backbones, divided into two main types: two-stage and single-stage models. Two-stage methods, like Faster R-CNN [35] and Mask R-CNN [14], generate region proposals to identify potential foreground objects before refining localization and classification. In contrast, single-stage methods, such asYOLO [41] and FCOS [39], directly classify and regress on backbone features. Transformer architectures [40] have further advanced object detection, as exemplified by DETR [2], which uses a fully end-to-end Transformer-based approach. The rise of multimodal foundation models [20, 34] has enabled open-set detection tasks [24], allowing object grounding with textual input [7, 47]. GLIP [20] frames detection as a grounding problem, integrating vision and language through feature fusion in the neck module. Grounding DINO [28] uses multiphase vision-language fusion, achieving state-of-the-art (SOTA) referring expression comprehension (REC) [27]. Other methods explore multimodal large language models (MLLMs) for detection and grounding tasks. Florence-2 [48], for instance, introduces a prompt-based model that generates object categories and coordinates using text prompts directly. The diverse feature extractors, architectures, and outputs in object detection make it challenging to + +interpret their decisions using a single unified approach. + +Object Detection Explanation. Explaining object detector decisions remains a largely unexplored area. Gudovskiy et al. [12] used Integrated Gradients (IG) [38] and SHAP [36] with SSD [29] classification scores for box-level attribution. Lee et al. [19] optimized masks on Mask R-CNN [14] to approximate original decisions with minimal pixels. Petsiuk et al. [33] proposed D-RISE, an adaptation of RISE [32], though it often introduces noise in saliency maps. Jiang et al. [16] introduced Nesterov-Accelerated iGOS++, which can be resource-intensive for large models. Zhao et al. [54] developed ODAM using Grad-CAM [37], but its effectiveness depends on network layer choices. SSGrad-CAM++ [49] generates saliency maps based on Grad-CAM++ [3]. VX-CODE [50] attributes important patches based on greedy search and SHAP [36]. While these methods aim to interpret detector errors, they largely rely on simple visualizations and empirical observations, often overlooking input-level feature confusion that can obscure human understanding. Masking confusing regions may aid in correcting model outputs [6]. Our paper introduces a black-box interpretation approach for object-level foundation models using Visual Precision Search, enhancing detection performance with fewer regions and identifying input-level failure causes to improve detection accuracy. + +# 3. Method + +This section presents our method for explaining the decisions of object-level foundation models. Figure 2 illustrates the overall framework of our method. + +# 3.1. Problem Formulation + +Given an image $\mathbf{I} \in \mathbb{R}^{h \times w \times 3}$ and an object-level foundation model $f(\cdot)$ , the output can be represented as $f(\mathbf{I}) = \{(b_i, c_i, s_i) \mid i = 1, 2, \dots, N\}$ . Each tuple $(b_i, c_i, s_i)$ denotes the bounding box, class label, and confidence score for each detected object $i$ , where $N$ is the maximum number of detection boxes. Tasks like object detection involve multiple boxes and categories, while visual grounding tasks like referring expression comprehension (REC) have only one box and category. Our goal is to generate a saliency map that explains the reasons behind the model's detection of a specific object, $(b_i, c_i, s_i)$ . Ideally, this explanation would enable the correct detection of objects using fewer input regions, while also requiring the removal of only a few critical regions to invalidate the detector. To achieve this, we can sparsify the input region $V = \{\mathbf{I}_1^s, \dots, \mathbf{I}_m^s\}$ , where $\mathbf{I}_i^s$ represents the $i$ -th sub-region, $m$ is the total number of sub-regions, and $\mathbf{I} = \sum_{i=1}^{m} \mathbf{I}_i^s$ . We then frame the saliency map generation task as a subset selection problem for these input sub-regions. A set function $\mathcal{F}(\cdot)$ is defined to assess interpretability by determining whether a given region is a key factor in the model's decision. Therefore, the + +objectives are: + +$$ +\max _ {S \subseteq V, | S | < k} \mathcal {F} (S), \tag {1} +$$ + +where $k$ denotes the maximum number of sub-regions. Therefore, the key to this problem lies in designing set function $\mathcal{F}$ and optimizing Eq. 1. + +# 3.2. Visual Precision Search + +We propose a Visual Precision Search method for interpreting object-level models. To begin, the input region needs to be sparsified. We apply the SLICO superpixel segmentation algorithm [1] to divide the input into $m$ sub-regions, $V = \{\mathbf{I}_1^s,\dots ,\mathbf{I}_m^s\}$ . To solve Eq. 1, an $\mathcal{NP}$ -hard problem, we employ submodular optimization [10]. Since the saliency map requires all sub-regions to be ranked, $k$ can be set to $|V|$ to compute ordered subsets. When the set function $\mathcal{F}(\cdot)$ satisfies the properties of diminishing returns1 and monotonic non-negative2, a greedy search guarantees an approximate optimal solution [8], with $\mathcal{F}(S)\geq (1 - 1 / e)\mathcal{F}(S^{*})$ , where $S^{*}$ denotes the optimal solution and $S$ denotes the greedy solution. We next design a set function to evaluate the interpretability score and rank the importance of interpretable regions for object-level tasks. + +**Clue Score:** An essential aspect of interpretability is enabling the object-level foundation model to accurately locate and identify objects while using fewer regions. To assess the importance of subregions from this perspective, we introduce the object location box information, $\pmb{b}_{\mathrm{target}}$ , and the target category, $c$ , that requires explanation. Given a subregion $S$ , the object-level model outputs $N$ instances, denoted as $f(S) = \{(b_i, s_{c,i}) \mid i = 1, \dots, N\}$ , where $s_{c,i}$ represents the confidence of the $i$ -th bounding box in predicting category $c$ . Then, the clue score of sub-region $S$ is defined as: + +$$ +s _ {\text {c l u e}} \left(S, \boldsymbol {b} _ {\text {t a r g e t}}, c\right) = \max _ {\left(\boldsymbol {b} _ {i}, s _ {c, i}\right) \in f (S)} \operatorname {I o U} \left(\boldsymbol {b} _ {\text {t a r g e t}}, \boldsymbol {b} _ {i}\right) \cdot s _ {c, i}, \tag {2} +$$ + +where $\mathrm{IoU}(\cdot, \cdot)$ represents the Intersection over Union between two bounding boxes. Unlike Petsiuk et al. [33], which only considers high-confidence bounding boxes, our method includes all candidate boxes to avoid getting stuck in a local optimum during the search. By incorporating the clue score $s_{\mathrm{clue}}$ , our method focuses on regions that strengthen the desired positional and semantic response. + +Collaboration Score: Some regions may exhibit strong combination effects, meaning they contribute effectively to model decisions only when paired with multiple specific sub-regions. Therefore, we introduce the collaboration score $s_{\mathrm{colla}}$ to assess sub-regions with high sensitivity to de + +![](images/eaf7ef8b52307bca88e7cc7785e556bc3f953fd9d27dd770b73b87d821fe7bb6.jpg) +Figure 2. Framework of the proposed Visual Precision Search method for interpreting an object-level foundation model. The input is first sparsified into a set of sub-regions and then interpreted across different instances. A submodular function guides the search for significant sub-regions, updating the ordered subset iteratively, and ultimately generating the instance-level attribution map. + +cision outcomes: + +$$ +s _ {\text {c o l l a .}} (S, \boldsymbol {b} _ {\text {t a r g e t}}, c) = 1 - \max _ {(\boldsymbol {b} _ {i}, s _ {c, i}) \in f (V \backslash S)} \operatorname {I o U} (\boldsymbol {b} _ {\text {t a r g e t}}, \boldsymbol {b} _ {i}) \cdot s _ {c, i}, \tag {3} +$$ + +this means that no detection box can accurately or confidently locate the object once the key sub-region is removed. This score serves as an effective guide in the initial search phase to identify subtle key regions that contribute to the final decision. + +Submodular Function: The scores above are combined to construct a submodular function $\mathcal{F}(\cdot)$ , as follows: + +$$ +\mathcal {F} (S, \boldsymbol {b} _ {\text {t a r g e t}}, c) = s _ {\text {c l u e}} (S, \boldsymbol {b} _ {\text {t a r g e t}}, c) + s _ {\text {c o l l a .}} (S, \boldsymbol {b} _ {\text {t a r g e t}}, c). \tag {4} +$$ + +Saliency Map Generation: Using the above submodular function, a greedy search algorithm is applied to sort all sub-regions in $V$ , yielding an ordered subset $S$ . Introducing the submodular function enables the search algorithm to more precisely identify key visual regions for interpretation. Additionally, scoring the sub-regions is necessary to better explain the importance of each sub-region. We evaluate the salient difference between the two sub-regions by the marginal effect. The attribution score $\mathcal{A}_i$ for each sub-region $s_i$ in $S$ is assessed by: + +$$ +\mathcal {A} _ {i} = \left\{ \begin{array}{l l} b _ {\text {b a s e}} & \text {i f} i = 1, \\ \mathcal {A} _ {i - 1} - \left| \mathcal {F} \left(S _ {[ i ]}\right) - \mathcal {F} \left(S _ {[ i - 1 ]}\right) \right| & \text {i f} i > 1, \end{array} \right. \tag {5} +$$ + +where $b_{\mathrm{base}}$ represents a baseline attribution score (such as 0 and won't affect results) for the first sub-region, and $S_{[i]}$ denotes the set containing the top $i$ sub-regions in $S$ . When a new sub-region is added, a small marginal increase suggests comparable importance to the previous sub-region. A negative marginal effect indicates a counterproductive impact, which can be assessed by its absolute value. Finally, $\mathcal{A}$ is normalized to obtain the saliency map of the sub-region. The detailed calculation process of the proposed Visual Precision Search algorithm is outlined in Algorithm 1. + +# 3.3. Theory Analysis + +A theoretical analysis is conducted for our Visual Precision Search method. + +Algorithm 1: Visual Precision Search algorithm for generating interpretable saliency maps of object-level foundation models. + +Input: Image $\mathbf{I} \in \mathbb{R}^{h \times w \times 3}$ , a division algorithm $\mathrm{Div}(\cdot)$ , the object location box $b_{\text{target}}$ , and the target category $c$ . + +Output: An ordered set $S$ and a saliency map $\mathcal{A} \in \mathbb{R}^{h \times w}$ . + +1 $V\gets \mathrm{Div}(\mathbf{I})$ + +2 $S\gets \emptyset$ / $\star$ Initiate the operation of submodular subset selection $\star/$ + +3 $\mathcal{A}_1\gets b_{\mathrm{base}}$ + +4 for $i = 1$ to $|V|$ do + +5 $S_{d}\gets V\backslash S;$ + +$\alpha \gets \arg \max_{\alpha \in S_d}\mathcal{F}(S\cup \{\alpha \} ,\pmb {b}_{\mathrm{target}},c)$ + +7 $S\gets S\cup \{\alpha \} ;$ + +8 if $i > 1$ then + +9 $|\quad | \quad | \quad \mathcal{A}_i\gets \mathcal{A}_{i - 1} - |\mathcal{F}(S_{[i]}) - \mathcal{F}(S_{[i - 1]})|$ + +10 end + +11 return $S$ , $\mathrm{norm}(\mathcal{A})$ + +Theorem 1 (Submodular Properties). Consider two subsets $S_A$ and $S_B$ in set $V$ , where $S_A \subseteq S_B \subseteq V$ . Given an element $\alpha$ , where $\alpha \in V \setminus S_b$ . Assuming that $\alpha$ is contributing to model interpretation, then, the function $\mathcal{F}(\cdot)$ in Eq. 4 satisfies diminishing returns, + +$$ +\mathcal {F} \left(S _ {A} \cup \{\alpha \}\right) - \mathcal {F} \left(S _ {A}\right) \geq \mathcal {F} \left(S _ {B} \cup \{\alpha \}\right) - \mathcal {F} \left(S _ {B}\right), +$$ + +and monotonic non-negative, $\mathcal{F}(S_A\cup \{\alpha \}) - \mathcal{F}(S)\geq 0$ and thus, $\mathcal{F}$ is a submodular function. + +Proof. Please see supplementary material for the proof. + +Remark 1 (Impact on Sparse Division). The quality of the search space is determined by sparse division, meaning that both the method of partitioning the input and the number of sub-regions play a crucial role in the faithfulness of the Visual Precision Search. + +Remark 2 (Applicable Models). If the model provides bounding box positions and category confidence without filtering out low-confidence detection boxes, the interpretation results of the Visual Precision Search can achieve guaranteed optimal boundaries. Models with detection heads + +are generally applicable; however, MLLM-based models may not directly output confidence scores, leaving room for improvement in the interpretation results. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We evaluate the proposed method on three object-level datasets: MS COCO 2017 [25], LVIS V1 [13], and RefCOCO [18]. The MS COCO dataset, used for interpreting object detection with 80 classes, involves random sampling of 5 correctly detected (IoU $>0.5$ , correct category), 3 misclassified (IoU $>0.5$ , wrong category), and 3 undetected (IoU $<0.5$ or low confidence) instances per class. The LVIS V1 dataset, with 1,203 categories including 337 rare ones (1-10 images), uses Grounding DINO for interpreting zero-shot detection on rare classes, analyzing 165 correctly detected, 152 misclassified, and 534 undetected samples. The RefCOCO dataset, for interpreting referring expression comprehension (REC), evaluates 500 correctly grounded (IoU $>0.5$ ) and 200 incorrectly grounded (IoU $<0.5$ ) samples per model. + +Baselines. The interpretation methods we compare include gradient-based approaches: Grad-CAM [37], SSGrad-CAM++[49], and ODAM[54], as well as perturbation-based methods: D-RISE [33] and D-HSIC [31]. D-RISE and D-HSIC are allowed to use all model outputs, including low-confidence detection boxes, to enhance the sufficiency of saliency estimation. + +Implementation Details. We validate our method on two object-level foundation models, Grounding DINO [28] and Florence-2 [48], across object detection and referring expression comprehension (REC) tasks. To interpret model decisions, we set the ground truth bounding box and category as target box $\pmb{b}_{\mathrm{target}}$ and category $c$ . By default, the SLICO superpixel segmentation algorithm divides the image into 100 sparse sub-regions. + +# 4.2. Evaluation Metrics + +Faithfulness Metrics. We evaluate the faithfulness of interpretation results across the entire image using Insertion and Deletion scores from the detection task, following the setup by Petsiuk et al. [33]. These metrics reflect location and recognition accuracy. We also calculate classification and IoU scores for the box closest to the target explanation denoted as Insertion (class), Deletion (class), Insertion (IoU), and Deletion (IoU). Additionally, we compute the average highest confidence, defined as the highest class response for the box closest to the target with IoU $>0.5$ . For failure samples, we introduce the Explaining Successful Rate (ESR) to evaluate whether the saliency map can effectively guide the search within a limited region, enabling the model to correctly detect misclassified or low-confidence objects + +(with IoU $>0.5$ and classification confidence above the detector's default threshold) previously. + +Location Metrics. We report the Point Game [53] and Energy Point Game [42] metrics. Since these metrics typically assume that the model's decision is relevant to the target object [33], we calculate them only for interpretation results on samples the model has correctly detected. + +# 4.3. Faithfulness Analysis + +# 4.3.1. Faithfulness on Grounding DINO + +We begin by validating our approach in explaining decisions across various object-level tasks using Grounding DINO [28]. Table 1 presents the faithfulness results across different datasets and tasks, demonstrating the strong interpretative faithfulness of our method. + +In the object detection interpretation task on the MS COCO dataset, our method achieved state-of-the-art results, outperforming the D-RISE method by $23.7\%$ on the Insertion metric, $6.7\%$ on the Deletion metric, and $10.6\%$ on the average highest score. The Insertion and Deletion metrics also achieved top performance under both category and IoU conditions. Additionally, our method performed best in location metrics, indicating state-of-the-art performance in explaining general object detection tasks. We also found that gradient-based methods, such as ODAM, exhibit low faithfulness in explaining Grounding DINO decisions. This is primarily due to the fusion of text and visual modalities, which causes the gradient to be influenced by both, reducing its effectiveness in attributing to the visual modality alone. + +In the referring expression comprehension (REC) interpretation task on the RefCOCO dataset, our method achieved state-of-the-art results across all evaluation metrics, surpassing the D-RISE method by $20.1\%$ on the Insertion metric, $22.1\%$ on the Deletion metric, and $4.4\%$ on the average highest score. These results confirm that our method maintains high faithfulness in explaining object description understanding. + +We also interpret zero-shot object detection for rare classes on the LVIS V1 dataset. Our method achieved SOTA results in faithfulness metrics, surpassing D-RISE by $31.6\%$ on the Insertion metric, $4.2\%$ on the Deletion metric, and $15.9\%$ on the average highest score. Although our method ranks second to ODAM on Energy Point Game metrics, this is mainly due to Grounding DINO not being fully trained on rare LVIS classes; as a result, model responses may be influenced by other regions rather than the object. + +Experimental results show that our method significantly outperforms other interpretability approaches, primarily due to the submodularity of our objective function, enabling precise evaluation of critical sub-regions. As shown in Figure 3, ODAM's saliency maps are diffuse, and D-RISE's maps are noisy, whereas our method sharply highlights essential sub-regions, capturing edges and class-specific fea + +Table 1. Evaluation of faithfulness metrics (Deletion, Insertion AUC scores, and average highest score) and location metrics (Point Game and Energy Point Game) on the MS-COCO, RefCOCO, and LVIS V1 (rare) validation sets for correctly detected or grounded samples using Grounding DINO. + +
DatasetsMethodsFaithfulness MetricsLocation Metrics
Ins. (↑)Del. (↓)Ins. (class) (↑)Del. (class) (↓)Ins. (IoU) (↑)Del. (IoU) (↓)Ave. high. score (↑)Point Game (↑)Energy PG (↑)
MS COCO [25] (Detection task)Grad-CAM [37]0.24360.15260.30640.20060.62290.53240.59040.17460.1463
SSGrad-CAM++ [49]0.21070.17780.26390.23140.59810.55110.58860.19050.1293
D-RISE [33]0.44120.04020.50810.08860.83960.36420.62150.94970.1850
D-HSIC [31]0.37760.04390.43820.09030.83010.33010.58620.73280.1861
ODAM [54]0.31030.05190.36550.08940.78690.39840.58650.54310.2034
Ours0.54590.03750.62040.08820.85810.33000.68730.98940.2046
RefCOCO [18] (REC task)Grad-CAM [37]0.37490.42370.46580.51940.75160.76850.74810.23800.2171
SSGrad-CAM++ [49]0.41130.39250.50080.48510.77000.75880.75610.28200.2262
D-RISE [33]0.61780.16050.70330.33960.86060.51640.84710.94000.2870
D-HSIC [31]0.54910.18460.62950.35090.85040.51200.77390.79000.3190
ODAM [54]0.47780.27180.56200.37570.82170.66410.74250.63200.3529
Ours0.74190.12500.80800.24570.90500.51030.88420.94600.3566
LVIS V1 (rare) [13] (Zero-shot det. task)Grad-CAM [37]0.12530.12940.18010.18140.56570.59100.35490.11510.0941
SSGrad-CAM++ [49]0.12530.12540.17650.17750.58000.56910.35040.10910.0931
D-RISE [33]0.28080.02890.33480.08350.83030.31740.42890.96970.1462
D-HSIC [31]0.24170.03530.29120.09280.81870.35500.40440.83030.1730
ODAM [54]0.20090.04100.24780.08440.77600.40820.36940.60610.2050
Ours0.36950.02770.42750.07990.84790.32420.49690.97580.1785
+ +![](images/06fcedbef2942fbaf3880fe70e2342bc62c7fa85e7f171e519aa3b463d587286.jpg) +Figure 3. Visualization results of Grounding DINO for interpreting object localization tasks on the MS COCO, RefCOCO, and LVIS datasets. The first column shows the saliency map with the ground truth box and label. The second column presents detection within the limited search region, and the third columns display the insertion curves. + +tures that enhance interpretability. + +# 4.3.2. Faithfulness on Florence-2 + +We further validated our approach on Florence-2 [48], a foundation model lacking object confidence scores. This limitation requires our search to rely solely on IoU as a guiding metric, excluding comparison with gradient-based methods. As shown in Table 2, our method achieves state-of-the-art performance. + +In the object detection interpretability task on the MS COCO dataset, our method outperforms D-RISE by $3.8\%$ on the Insertion metric and achieves a substantial improvement of $50.7\%$ on the Deletion metric. Additionally, our method enhances the Point Game and Energy Point Game metrics by $8.3\%$ and $60.7\%$ , respectively. Due to the absence of object confidence scores in Florence-2, it becomes challenging to identify areas that quickly invalidate the model's detection results. Compared to perturbation-based methods, our approach offers significant advantages; it relies more directly on the object itself, resulting in stronger faithfulness to the model's detection behavior. + +We also interpret referring expression comprehension + +(REC) interpretation task on the RefCOCO dataset, outperforming D-RISE by $6.1\%$ and $66.9\%$ in Insertion and Deletion metrics, respectively. Our approach also achieved state-of-the-art localization results, with a $14.6\%$ improvement in the Energy Point Game, demonstrating strong interpretability in accurately pinpointing object information. + +We visualized some results of our method in Figure 4, illustrating that it effectively identifies key information about the object's location and categorical description. While D-RISE performed well on the Insertion metric, its attribution results were diffuse and lacked reliability, and it did not perform as well on the Deletion metric. + +# 4.4. Interpreting Failures in Object-Level Tasks + +# 4.4.1. Interpreting REC Failures + +Failures in the REC task primarily arise from visual grounding errors, often due to interference from other objects in the scene that become entangled with the text. Removing this distracting visual information can clarify grounding results and improve error understanding at the visual input level. To achieve this, we specify the correct location as the target box, denoted as $\pmb{b}_{\mathrm{target}}$ , for input attribution. Ta + +Table 2. Evaluation of faithfulness metrics (Deletion and Insertion AUC scores) and location metrics (Point Game and Energy Point Game) on the MS COCO and RefCOCO validation sets for correctly detected and grounded samples using Florence-2. + +
DatasetsMethodsFaithfulness MetricsLocation Metrics
Insertion (↑)Deletion (↓)Point Game (↑)Energy PG (↑)
MS COCO [25] (Detection task)D-RISE [33]0.74770.09720.88500.1568
D-HSIC [31]0.53450.27300.29250.0862
Ours0.77590.04790.95830.2519
RefCOCO [18] (REC task)D-RISE [33]0.79220.35050.84800.2464
D-HSIC [31]0.76390.35600.69800.2754
Ours0.84090.11590.86600.3927
+ +![](images/168889f26d4bbeab9004c4ea9077f0b9164776fde77f0e75992663ed758ed185.jpg) +MS COCO (Object Detection Interpretation) + +![](images/f7fd96ec7d9b4846ebb8ca56d73853a3204cfc66e17d02554e24e50fc782f895.jpg) +RefCOCO (Visual Grounding Interpretation) +Figure 4. Visualization results of Florence-2 for interpreting object localization tasks on the MS COCO and RefCOCO datasets. + +ble 3 shows the attribution results, where gradient-based methods exhibit limited performance in faithfulness metrics. Compared to D-RISE, our approach improves Insertion by $42.9\%$ , class score-based Insertion by $25.1\%$ , and average highest score by $14.9\%$ . Figure 5 illustrates model error interpretations, with the cyan-highlighted region indicating decision errors due to visual input interference that leads the foundation model astray. + +# 4.4.2. Interpreting Detection Failures + +Detection failures include misclassification and undetected objects. While previous works [33, 54] used visualization to explain errors, it lacked quantitative analysis, resulting in lower trustworthiness. We used saliency maps to identify input-level interference causing detection failures. + +Misclassified Interpretation. As shown in Table 4, On the MS COCO dataset, our method outperforms D-RISE, improving Insertion by $54.7\%$ , Insertion (class) by $49.1\%$ , average highest score by $27.4\%$ , and explaining success rate (ESR) by $19.47\%$ . Similarly, on the LVIS dataset, our method shows advantages with improvements of $42.7\%$ in Insertion, $33.0\%$ in Insertion (class), $24.8\%$ in average highest score, and $24.34\%$ in ESR. This improvement is pri + +Table 3. Insertion AUC scores and the average highest score on the RefCOCO validation sets for or the samples with incorrect localization in visual grounding using Grounding DINO. + +
DatasetsMethodsFaithfulness Metrics
Ins. (†)Ins. (class) (†)Ave. high. score (†)
RefCOCO [18] (REC task)Grad-CAM [37]0.15360.27940.3295
SSGrad-CAM++ [49]0.15900.28370.3266
D-RISE [33]0.34860.47870.6096
D-HSIC [31]0.22740.34880.4495
ODAM [54]0.17930.30010.3453
Ours0.49810.59900.7007
+ +![](images/bcdc9a12c9905eed0b174787847dc65d369e8b52c586c37900f296874c6e46c5.jpg) +Attribution Map + +![](images/624dba656a394fb44bce58772fae49a4e3ec2407027bf4cd67bfb4b3a119ab3a.jpg) +Searched Region + +![](images/18ec6165b373b0df06406c8966afde8776f605462b14c7e9bb46ed9cb5946570.jpg) +Figure 5. Visualization of the method for discovering what causes the Grounding DINO incorrectly grounded on RefCOCO. + +Table 4. Insertion AUC scores, average highest score, and explaining successful rate (ESR) on the MS-COCO and the LVIS validation sets for misclassified samples using Grounding DINO. + +
DatasetsMethodsFaithfulness Metrics
Ins. (↑)Ins. (class) (↑)Ave. high. score (↑)ESR (↑)
MS COCO [25] (Detection task)Grad-CAM [37]0.10910.14780.310238.38%
SSGrad-CAM++ [49]0.09600.13360.295233.51%
D-RISE [33]0.21700.26610.360350.26%
D-HSIC [31]0.17710.21610.314334.59%
ODAM [54]0.11290.14860.286932.97%
Ours0.33570.39670.459169.73%
LVIS V1 (rare) [13] (Zero-shot det. task)Grad-CAM [37]0.05030.08910.156412.50%
SSGrad-CAM++ [49]0.05740.09460.158011.84%
D-RISE [33]0.12450.16470.208828.95%
D-HSIC [31]0.09630.12470.174816.45%
ODAM [54]0.05750.09540.15209.21%
Ours0.17760.21900.260653.29%
+ +![](images/5e2aadfbc73247ae622522d08a8657860edf3733d7065e22a60be577012c6570.jpg) + +![](images/b1b48ac7a890ea8b98c764067c57e0a6987dba755e7201046703b7f46a81d881.jpg) + +![](images/e2767f4fec27d15220a39d7a33cef5eb2e72f6aad5b2f699395aaac9e3cad41d.jpg) + +![](images/8b9a3766fa0276b71e5b955eb94ef5ffa3076c9620ca85bdeafee9eb46036771.jpg) +Figure 6. Visualization of our method reveals the causes of Grounding DINO misclassifications on MS COCO and LVIS. The cyan region in the saliency map highlights the regions responsible for the model's misclassification. + +marily due to our proposed Visual Precision Search, which confines attribution to the target negative response region, effectively refining the model's detection results. Figure 6 shows that the background surrounding the object interferes with the model's decision-making. Improving the model by refining the contextual relationship between foreground and background may be a promising direction. + +Undetected Interpretation. Low object confidence may result from both the model's feature representation and confounding input-level factors. We analyze causes of missed detections, with Table 5 showing that our method achieves SOTA performance across metrics. On MS + +Table 5. Insertion, average highest score, and explaining successful rate (ESR) on the MS-COCO and the LVIS V1 (rare) validation sets for missed detection samples using Grounding DINO. + +
DatasetsMethodsFaithfulness Metrics
Ins. (↑)Ins. (class) (↑)Ave. high. score (↑)ESR (↑)
MS COCO [25] (Detection task)Grad-CAM [37]0.07600.13210.215316.44%
SSGrad-CAM++ [49]0.06710.11510.212416.44%
D-RISE [33]0.15380.22600.256426.94%
D-HSIC [31]0.11010.17160.194513.56%
ODAM [54]0.07450.13500.203713.78%
Ours0.21020.30110.301441.33%
LVIS V1 (rare) [13] (Zero-shot det. task)Grad-CAM [37]0.02910.06890.09015.43%
SSGrad-CAM++ [49]0.02920.06800.08975.24%
D-RISE [33]0.07030.11840.131218.73%
D-HSIC [31]0.05160.09200.116813.48%
ODAM [54]0.02830.07160.08514.68%
Ours0.11550.18860.178430.15%
+ +![](images/20aeadba9f1187fd25ce0e0faa0f9f4b6a4f06d8b2b970286795cbb1ac85972d.jpg) + +![](images/eb25181b7122be4733e9b26602288f904e81c94c1539f47e630e9d66cef0a48d.jpg) + +![](images/7df524c5ee250a3735a17b2654078647ea8f5fe25ef1c8c97bc8fa9073e4ceaf.jpg) +Figure 7. Visualization of our method reveals the causes of Grounding DINO undetected on MS COCO and LVIS. The cyan region in the saliency map highlights the regions responsible for the model's detection failure. + +COCO, it outperforms D-RISE with a $36.7\%$ improvement in Insertion and a $14.39\%$ gain in ESR, while on LVIS, it improves Insertion by $64.3\%$ and ESR by $11.42\%$ . Figure 7 visualizes our method's explanations for undetected instances, revealing that errors may arise from challenges in distinguishing similar objects (e.g., the bear in the first row) and environmental influence on detection (e.g., the corkboard in the second row). These insights highlight current model limitations, offering directions for improvement. + +# 4.5. Ablation Study + +Ablation of the Submodular Function. We analyze the effectiveness of our submodular function design on Grounding DINO using correctly detected samples from MS COCO. Table 6 shows the results. The Clue Score identifies regions beneficial for detection, improving Insertion and average highest score, while the Collaboration Score pinpoints sensitive regions, enhancing the Deletion metric by causing the detector to fail more quickly. Combining these scores enables our method to achieve optimal results across indicators, demonstrating the effectiveness of each score function within the submodular function. + +Ablation on Divided Sub-region Number. The sub-region number affects search space quality. Using Grounding DINO, we assess its impact on correctly predicted and misclassified samples in MS COCO. Figures 8A-B show + +Table 6. Ablation study on function score components for Grounding DINO on the MS COCO validation set. + +
Clue Score (Eq. 2)Colla. Score (Eq. 3)Insertion (↑)Faithfulness Metrics
Deletion (↓)Ave. high. score (↑)
✗✓0.36320.03780.5967
✓✗0.53700.07990.6864
✓✓0.54590.03750.6873
+ +![](images/036d93524432f2eb3c3404764702619843f7ec9b485613259b3239f3a22ab99e.jpg) + +![](images/d194756454313e106e54a7438eb6385fbb655c7dfc6f08371d0f431038d27ac8.jpg) + +![](images/32360a8a3ee25e5b4d89474aa6d4d5dfa81105e1236700541e4d8c659e6af12e.jpg) + +![](images/8153aa83ef8498af7d707a9f49587a9054d965f453bb4bcbb1867f1b39ef0fd7.jpg) +Figure 8. Ablation on the number of sub-regions. For correct detections, panels $\mathbf{A}$ and $\mathbf{B}$ show the effect of sub-region count on Insertion and average highest score, respectively. Panel $\mathbf{C}$ presents the inference time for different sub-region counts. For misclassified samples, panels $\mathbf{D},\mathbf{E},$ and $\mathbf{F}$ display the effects on Insertion, average highest score, and explanation success rate. + +![](images/e5a30510e6322b814bc0b5d15f36dc1222dab683aafae12775e946ee05583565.jpg) + +![](images/fe794ee2cdcb70b9fd92fcb0b2b9ebe28b7838e44255182e4785f6814d489df0.jpg) + +that when the model predicts correctly, increasing the number of sub-regions leads to higher Insertion and average highest scores, indicating that finer divisions enhance the faithfulness of search results. Additionally, Figures 8D-F show that, for misclassified samples, Insertion, average highest score, and ESR also improve as the number of sub-regions increases. Figure 8C shows inference times for our search algorithm on an RTX 3090 compared to D-RISE across various subregion counts. Increasing sub-regions improves faithfulness but also rapidly increases inference time. Future work will aim to increase subregion count and inference speed without sacrificing attribution performance. + +# 5. Conclusion + +In this paper, we propose an interpretable attribution method specifically tailored for object-level foundation models, called the Visual Precision Search method, which introduces a novel submodular mechanism that combines a clue score and a collaboration score. This method achieves enhanced interpretability with fewer search regions. Experiments on RefCOCO, MS COCO, and LVIS demonstrate that our approach improves object-level task interpretability over state-of-the-art methods for Grounding DINO and Florence-2 across various evaluation metrics. Furthermore, our method effectively interprets failures in visual grounding and object detection tasks. + +# Acknowledgements + +This work was supported by the Shenzhen Science and Technology Program (No. KQTD20221101093559018), National Natural Science Foundation of China (No. 62372448, 62025604, 62306308, 62441619), Foundation of Key Laboratory of Education Informatization for Nationalities (YNNU), Ministry of Education (No. EIN2024B004). + +# References + +[1] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Trans. Pattern Anal. Mach. Intell., 34(11):2274-2282, 2012. 3 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Eur. Conf. Comput. Vis. (ECCV), pages 213-229, 2020. 2 +[3] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks. In IEEE Winter Conf. Appl. Comput. Vis. (WACV), pages 839-847, 2018. 3 +[4] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 1 +[5] Ruoyu Chen, Jingzhi Li, Hua Zhang, Changchong Sheng, Li Liu, and Xiaochun Cao. Sim2Word: Explaining similarity with representative attribute words via counterfactual explanations. ACM Trans. Multim. Comput. Commun. Appl., 19 (6):1-22, 2023. 1 +[6] Ruoyu Chen, Hua Zhang, Siyuan Liang, Jingzhi Li, and Xiaochun Cao. Less is more: Fewer interpretable region via submodular subset selection. In Int. Conf. Learn. Represent. (ICLR), 2024. 1, 3 +[7] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 16901-16911, 2024. 1, 2 +[8] Jack Edmonds. Submodular functions, matroids, and certain polyhedra. Combinatorial Structures and Their Applications, pages 69-87, 1970. 3 +[9] Di Feng, Ali Harakeh, Steven L Waslander, and Klaus Dietmayer. A review and comparative study on probabilistic object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(8):9961-9980, 2021. 1 +[10] Satoru Fujishige. Submodular functions and optimization. Elsevier, 2005. 3 +[11] Yossi Gandelsman, Alexei A Efros, and Jacob Steinhardt. Interpreting clip's image representation via text-based decomposition. In Int. Conf. Learn. Represent. (ICLR), 2024. 1 +[12] Denis Gudovskiy, Alec Hodgkinson, Takuya Yamaguchi, Yasunori Ishii, and Sotaro Tsukizawa. Explain to fix: A + +framework to interpret and correct dnn object detector predictions. arXiv preprint arXiv:1811.08011, 2018.3 +[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5356-5364, 2019. 2, 5, 6, 7, 8 +[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. IEEE Trans. Pattern Anal. Mach. Intell., 42(2):386-397, 2018. 1, 2, 3 +[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 17853-17862, 2023. 1 +[16] Mingqi Jiang, Saeed Khorram, and Fuxin Li. Diverse explanations for object detectors with nesterov-accelerated igos++. In Brit. Mach. Vis. Conf. (BMVC), pages 188–189, 2023. 3 +[17] Mingqi Jiang, Saeed Khorram, and Li Fuxin. Comparing the decision-making mechanisms by transformers and cnns via explanation methods. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 9546-9555, 2024. 1 +[18] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. ReferItGame: Referring to objects in photographs of natural scenes. In Proc. Conf. Empir. Methods Nat. Lang. Process. (EMNLP), pages 787-798, 2014. 2, 5, 6, 7 +[19] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. BBAM: Bounding box attribution map for weakly supervised semantic and instance segmentation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 2643-2652, 2021. 3 +[20] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 10965-10975, 2022. 1, 2 +[21] Siyuan Liang, Baoyuan Wu, Yanbo Fan, Xingxing Wei, and Xiaochun Cao. Parallel rectangle flip attack: A query-based black-box attack against object detection. In Int. Conf. Comput. Vis. (ICCV), pages 7697-7707, 2021. 1 +[22] Siyuan Liang, Longkang Li, Yanbo Fan, Xiaojun Jia, Jingzhi Li, Baoyuan Wu, and Xiaochun Cao. A large-scale multiple-objective method for black-box attack against object detection. In Eur. Conf. Comput. Vis. (ECCV), pages 619-636, 2022. 1 +[23] Siyuan Liang, Aishan Liu, Jiawei Liang, Longkang Li, Yang Bai, and Xiaochun Cao. Imitated detectors: Stealing knowledge of black-box object detectors. In ACM Int. Conf. Multimedia, pages 4839-4847, 2022. 1 +[24] Siyuan Liang, Wei Wang, Ruoyu Chen, Aishan Liu, Boxi Wu, Ee-Chien Chang, Xiaochun Cao, and Dacheng Tao. Object detectors in the open environment: Challenges, solutions, and outlook. arXiv preprint arXiv:2403.16271, 2024. 1, 2 +[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence + +Zitnick. Microsoft COCO: Common objects in context. In Eur. Conf. Comput. Vis. (ECCV), pages 740-755, 2014. 2, 5, 6, 7, 8 +[26] Aishan Liu, Jun Guo, Jiakai Wang, Siyuan Liang, Renshuai Tao, Wenbo Zhou, Cong Liu, Xianglong Liu, and Dacheng Tao. {X-Adv}: Physical adversarial object attacks against x-ray prohibited item detection. In USENIX Security Symposium, pages 3781–3798, 2023. 1 +[27] Jingyu Liu, Liang Wang, and Ming-Hsuan Yang. Referring expression generation and comprehension via attributes. In Int. Conf. Comput. Vis. (ICCV), pages 4856-4864, 2017. 2 +[28] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. In Eur. Conf. Comput. Vis. (ECCV), 2024. 1, 2, 5 +[29] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. SSD: Single shot multibox detector. In Eur. Conf. Comput. Vis. (ECCV), pages 21-37, 2016. 3 +[30] Dimity Miller, Feras Dayoub, Michael Milford, and Niko Sünderhauf. Evaluating merging strategies for sampling-based uncertainty techniques in object detection. In Proc. IEEE Int. Conf. Robot. Autom. (ICRA), pages 2348-2354, 2019. 1 +[31] Paul Novello, Thomas Fel, and David Vigouroux. Making sense of dependence: Efficient black-box explanations using dependence measure. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 4344-4357, 2022. 5, 6, 7, 8 +[32] Vitali Petsiuk, Abir Das, and Kate Saenko. RISE: Randomized input sampling for explanation of black-box models. In Brit. Mach. Vis. Conf. (BMVC), page 151, 2018. 3 +[33] Vitali Petsiuk, Rajiv Jain, Varun Manjunatha, Vlad I Morariu, Ashutosh Mehra, Vicente Ordonez, and Kate Saenko. Black-box explanation of object detectors via saliency maps. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 11443-11452, 2021. 2, 3, 5, 6, 7, 8 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn. (ICML), pages 8748-8763, 2021. 1, 2 +[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell., 39(6):1137-1149, 2016. 2 +[36] M Scott, Lee Su-In, et al. A unified approach to interpreting model predictions. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 4765-4774, 2017. 3 +[37] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis., 128:336-359, 2020. 3, 5, 6, 7, 8 +[38] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Int. Conf. Mach. Learn. (ICML), pages 3319-3328, 2017. 3 + +[39] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: A simple and strong anchor-free object detector. IEEE Trans. Pattern Anal. Mach. Intell., 44(4):1922-1933, 2020. 2 +[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 5998-6008, 2017. 2 +[41] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 7464-7475, 2023. 2 +[42] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 5 +[43] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable adversarial attacks for image and video object detection. In IJCAI, pages 954-960, 2019. 1 +[44] Licheng Wen, Xuemeng Yang, Daocheng Fu, Xiaofeng Wang, Pinlong Cai, Xin Li, MA Tao, Yingxuan Li, XU Linran, Dengke Shang, et al. On the road with gpt-4v (ision): Explorations of utilizing visual-language model as autonomous driving agent. In ICLR 2024 Workshop on Large Language Model (LLM) Agents, 2024. 1 +[45] Samuel Wilson, Tobias Fischer, Feras Dayoub, Dimity Miller, and Niko Sunderhauf. Safe: Sensitivity-aware features for out-of-distribution object detection. In Int. Conf. Comput. Vis. (ICCV), pages 23565-23576, 2023. 1 +[46] Junfeng Wu, Yi Jiang, Qihao Liu, Zehuan Yuan, Xiang Bai, and Song Bai. General object foundation model for images and videos at scale. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3783-3795, 2024. 1 +[47] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, et al. Towards open vocabulary learning: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2 +[48] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4818–4829, 2024. 1, 2, 5, 6 +[49] Toshinori Yamauchi. Spatial Sensitive Grad-CAM++: Improved visual explanation for object detectors via weighted combination of gradient map. In CVPR Workshop, pages 8164-8168, 2024. 3, 5, 6, 7, 8 +[50] Toshinori Yamauchi, Hiroshi Kera, and Kazuhiko Kawamoto. Explaining object detectors via collective contribution of pixels. arXiv preprint arXiv:2412.00666, 2024.3 +[51] Lewei Yao, Renjie Pi, Jianhua Han, Xiaodan Liang, Hang Xu, Wei Zhang, Zhenguo Li, and Dan Xu. Detclipv3: Towards versatile generative open-vocabulary object detection. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 27391-27401, 2024. 1 + +[52] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. Int. J. Comput. Vis., pages 1-19, 2024. 1 +[53] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. Int. J. Comput. Vis., 126 (10):1084-1102, 2018. 5 +[54] Chenyang Zhao, Janet H Hsiao, and Antoni B Chan. Gradient-based instance-specific visual explanations for object specification and object discrimination. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2, 3, 5, 6, 7, 8 +[55] Chenyang Zhao, Kun Wang, Xingyu Zeng, Rui Zhao, and Antoni B Chan. Gradient-based visual explanation for transformer-based clip. In Int. Conf. Mach. Learn. (ICML), pages 61072-61091, 2024. 1 +[56] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 15116-15127, 2023. 1 +[57] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 1 \ No newline at end of file diff --git a/2025/Interpreting Object-level Foundation Models via Visual Precision Search/images.zip b/2025/Interpreting Object-level Foundation Models via Visual Precision Search/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..32d83cddac7c4d318404d5df8e07676d41b3947e --- /dev/null +++ b/2025/Interpreting Object-level Foundation Models via Visual Precision Search/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39a71e030c7bb12c449c07f0f6afbf4fef3f7b1302790a731269d98995b85d84 +size 816824 diff --git a/2025/Interpreting Object-level Foundation Models via Visual Precision Search/layout.json b/2025/Interpreting Object-level Foundation Models via Visual Precision Search/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..febe19456e454131441aaa146ccc274aa82d7852 --- /dev/null +++ b/2025/Interpreting Object-level Foundation Models via Visual Precision Search/layout.json @@ -0,0 +1,11303 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 530, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 530, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 530, + 121 + ], + "type": "text", + "content": "Interpreting Object-level Foundation Models via Visual Precision Search" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "spans": [ + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "text", + "content": "Ruoyu Chen" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "text", + "content": ", Siyuan Liang" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "text", + "content": ", Jingzhi Li" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "inline_equation", + "content": "^{1,2,7}" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "text", + "content": ", Shiming Liu" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "text", + "content": ", Maosen Li" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "text", + "content": ", Zhen Huang" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "text", + "content": ", Hua Zhang" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "text", + "content": ", and Xiaochun Cao" + }, + { + "bbox": [ + 126, + 126, + 484, + 156 + ], + "type": "inline_equation", + "content": "^{8,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 156, + 497, + 183 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 118, + 156, + 492, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 156, + 492, + 169 + ], + "spans": [ + { + "bbox": [ + 118, + 156, + 492, + 169 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 156, + 492, + 169 + ], + "type": "text", + "content": "Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100093, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 171, + 497, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 171, + 497, + 183 + ], + "spans": [ + { + "bbox": [ + 113, + 171, + 497, + 183 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 171, + 497, + 183 + ], + "type": "text", + "content": "School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 63, + 184, + 547, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 184, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 63, + 184, + 547, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 184, + 547, + 198 + ], + "type": "text", + "content": "School of Computing, NUS " + }, + { + "bbox": [ + 63, + 184, + 547, + 198 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 63, + 184, + 547, + 198 + ], + "type": "text", + "content": "RAMS Lab, Huawei Technologies Co., Ltd. " + }, + { + "bbox": [ + 63, + 184, + 547, + 198 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 63, + 184, + 547, + 198 + ], + "type": "text", + "content": "IAS BU, Huawei Technologies Co., Ltd." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 198, + 548, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 198, + 548, + 211 + ], + "spans": [ + { + "bbox": [ + 63, + 198, + 548, + 211 + ], + "type": "text", + "content": "6College of Computer, NUDT 7Key Lab. of Edu. Inf. for Nationalities (YNNU), Ministry of Education, Kunming, China" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 212, + 536, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 212, + 536, + 225 + ], + "spans": [ + { + "bbox": [ + 75, + 212, + 536, + 225 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 75, + 212, + 536, + 225 + ], + "type": "text", + "content": "School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen 518107, China" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 84, + 228, + 526, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 228, + 526, + 239 + ], + "spans": [ + { + "bbox": [ + 84, + 228, + 526, + 239 + ], + "type": "text", + "content": "chenruoyu@ie.ac.cn pandaliang521@gmail.com {lijingzhi,zhanghua}@ie.ac.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 241, + 541, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 241, + 541, + 253 + ], + "spans": [ + { + "bbox": [ + 67, + 241, + 541, + 253 + ], + "type": "text", + "content": "{liushiming3, limaosen2}@huawei.com huangzhen@nudt.edu.cn caoxiaochun@mail.sysu.edu.cn" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 152, + 256, + 200, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 256, + 200, + 269 + ], + "spans": [ + { + "bbox": [ + 152, + 256, + 200, + 269 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 282, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 628 + ], + "type": "text", + "content": "Advances in multimodal pre-training have propelled object-level foundation models, such as Grounding DINO and Florence-2, in tasks like visual grounding and object detection. However, interpreting these models' decisions has grown increasingly challenging. Existing interpretable attribution methods for object-level task interpretation have notable limitations: (1) gradient-based methods lack precise localization due to visual-textual fusion in foundation models, and (2) perturbation-based methods produce noisy saliency maps, limiting fine-grained interpretability. To address these, we propose a Visual Precision Search method that generates accurate attribution maps with fewer regions. Our method bypasses internal model parameters to overcome attribution issues from multimodal fusion, dividing inputs into sparse sub-regions and using consistency and collaboration scores to accurately identify critical decision-making regions. We also conducted a theoretical analysis of the boundary guarantees and scope of applicability of our method. Experiments on RefCOCO, MS COCO, and LVIS show our approach enhances object-level task interpretability over SOTA for Grounding DINO and Florence-2 across various evaluation metrics, with faithfulness gains of " + }, + { + "bbox": [ + 55, + 282, + 296, + 628 + ], + "type": "inline_equation", + "content": "23.7\\%" + }, + { + "bbox": [ + 55, + 282, + 296, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 282, + 296, + 628 + ], + "type": "inline_equation", + "content": "31.6\\%" + }, + { + "bbox": [ + 55, + 282, + 296, + 628 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 282, + 296, + 628 + ], + "type": "inline_equation", + "content": "20.1\\%" + }, + { + "bbox": [ + 55, + 282, + 296, + 628 + ], + "type": "text", + "content": " on MS COCO, LVIS, and RefCOCO for Grounding DINO, and " + }, + { + "bbox": [ + 55, + 282, + 296, + 628 + ], + "type": "inline_equation", + "content": "50.7\\%" + }, + { + "bbox": [ + 55, + 282, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 282, + 296, + 628 + ], + "type": "inline_equation", + "content": "66.9\\%" + }, + { + "bbox": [ + 55, + 282, + 296, + 628 + ], + "type": "text", + "content": " on MS COCO and RefCOCO for Florence-2. Additionally, our method can interpret failures in visual grounding and object detection tasks, surpassing existing methods across multiple evaluation metrics. The code is released at https://github.com/RuoyuChen10/VPS." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 636, + 135, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 135, + 648 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 135, + 648 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "content": "Understanding object information in images, such as object detection [7, 14, 24, 51, 52, 57], is a crucial and enduring challenge in computer vision, holding significant im" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 322, + 257, + 543, + 457 + ], + "blocks": [ + { + "bbox": [ + 322, + 257, + 543, + 457 + ], + "lines": [ + { + "bbox": [ + 322, + 257, + 543, + 457 + ], + "spans": [ + { + "bbox": [ + 322, + 257, + 543, + 457 + ], + "type": "image", + "image_path": "3fb944160443eb79d76d1aa87654ccef39bb1ca0bdc30adeef9458a556cb074f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 460, + 555, + 504 + ], + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 504 + ], + "type": "text", + "content": "Figure 1. Illustration of our Visual Precision Search interpretation method, which more precisely identifies key sub-regions in object-level foundation model decision-making compared to gradient-based and perturbation-based methods." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 520, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 687 + ], + "type": "text", + "content": "portance across various fields, including autonomous driving [4, 15, 44]. With advancements in multimodal alignment technology [20, 34], object-level foundation models [28, 46, 48, 56] like Grounding DINO [28] and Florence-2 [48] have been developed to handle various tasks, including visual grounding and object detection. However, the massive data volume, rich semantic concepts, and large parameters significantly reduce the transparency [21, 22, 26] and interpretability [5, 11, 55] of these models. Since object-level tasks like autonomous driving demand high model reliability [9, 30, 45], building transparent and interpretable models is crucial for enhancing safety and trustworthy [6, 17, 21, 23, 24, 43], making it essential to interpret object-level foundation models." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Some interpretable attribution algorithms for traditional object detection models have been proposed, including" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30042" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "gradient-based methods like ODAM [54] and perturbation-based methods like D-RISE [33]. However, as shown in the lower panel of Figure 1, these methods face limitations when applied to multimodal foundation models: (1) gradient-based methods may struggle to provide accurate visual localization explanations due to the fusion of visual and textual features in foundation models, and (2) perturbation-based approaches can introduce sampling artifacts, resulting in noisy saliency maps with limited fine-grained interpretability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 196, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 296, + 529 + ], + "type": "text", + "content": "To address these issues, we propose a novel interpretation mechanism for the object-level foundation model, as shown in Figure 1. Our goal is to generate a saliency map that explains the rationale behind the model's detection of a specific object, allowing fewer regions to be exposed for accurate detection. Additionally, removing these critical regions should quickly lead to detection failure, highlighting their importance in the model's decision. Specifically, we propose Visual Precision Search, which sparsifies the input region into a series of sub-regions using superpixel segmentation, and then ranks these sparse sub-regions to determine their importance for object-level decision-making. Our gradient-free method circumvents localization errors at the visual level that are caused by gradient backpropagation in vision-text fusion. Furthermore, we propose a novel submodular function, grounded in theoretical analysis and offering boundary guarantees for Visual Precision Search. This function identifies regions that enhance interpretability from two key perspectives: clues that support the model's accurate detection and regions with strong combinatorial effects. By employing region search to iteratively expand the set of sub-regions, our method more precisely identifies key sub-regions in the decision-making of object-level foundation models, which mitigates the issue of coarse explanations caused by noise in perturbation-based methods. In addition, our method effectively analyzes cases of grounding or detection failure, allowing us to observe which input-level factors influence the decision." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "We validated our method on the MS COCO [25], RefCOCO [18], and LVIS [13] datasets. The object-level foundation models evaluated include Grounding DINO, which uses a multimodal feature fusion architecture, and Florence2, which employs a multimodal large language model architecture. The tasks covered include interpreting visual grounding and object detection. In explaining the data behind correct model decisions, our method demonstrates that the faithfulness metric for Grounding DINO surpasses SOTA method D-RISE by " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "23.7\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "20.1\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "31.6\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " on MS COCO, RefCOCO, and LVIS, respectively. For Florence2, our method improves the faithfulness metric by " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "102.9\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "66.9\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " on MS COCO and RefCOCO, respectively, while also achieving SOTA performance on location metrics across all datasets. In interpreting the factors leading to" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "Grounding DINO's failures in visual grounding tasks, our method improves the Insertion metric by " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "42.9\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " and the average highest score by " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "25.1\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ". It also identifies causes of detector misclassification and undetection, achieving Insertion metric improvements of " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "54.7\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "42.7\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " for misclassification, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "36.7\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "64.3\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " for undetection on MS COCO and LVIS, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 324, + 157, + 518, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 157, + 518, + 167 + ], + "spans": [ + { + "bbox": [ + 324, + 157, + 518, + 167 + ], + "type": "text", + "content": "In summary, the contributions of this paper are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 168, + 553, + 371 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 168, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 168, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 314, + 168, + 553, + 204 + ], + "type": "text", + "content": "- We introduce Visual Precision Search, a new search-based mechanism for interpreting object-level foundation models through instance-specific saliency maps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 204, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 262 + ], + "type": "text", + "content": "- A novel submodular mechanism is constructed to enhance interpretability from two aspects: identifying clues that support accurate detection and highlighting regions with strong combinatorial effects, with an analysis of its theoretical boundaries for object-level tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 264, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 311 + ], + "type": "text", + "content": "- We demonstrate our approach's generalizability across multimodal foundation models, including both non-LLM-based (Grounding DINO) and LLM-based detectors (Florence-2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 371 + ], + "type": "text", + "content": "- We validate our method on MS COCO, RefCOCO, and LVIS, achieving significant improvements in explaining object-level tasks. Additionally, we analyze grounding and detection failures, establishing a quantitative benchmark for this purpose." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 382, + 400, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 382, + 400, + 395 + ], + "spans": [ + { + "bbox": [ + 314, + 382, + 400, + 395 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 403, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 715 + ], + "type": "text", + "content": "Object Detection. Object detection involves locating objects in images and identifying their categories [24]. Early models typically used convolutional neural network (CNN) backbones, divided into two main types: two-stage and single-stage models. Two-stage methods, like Faster R-CNN [35] and Mask R-CNN [14], generate region proposals to identify potential foreground objects before refining localization and classification. In contrast, single-stage methods, such asYOLO [41] and FCOS [39], directly classify and regress on backbone features. Transformer architectures [40] have further advanced object detection, as exemplified by DETR [2], which uses a fully end-to-end Transformer-based approach. The rise of multimodal foundation models [20, 34] has enabled open-set detection tasks [24], allowing object grounding with textual input [7, 47]. GLIP [20] frames detection as a grounding problem, integrating vision and language through feature fusion in the neck module. Grounding DINO [28] uses multiphase vision-language fusion, achieving state-of-the-art (SOTA) referring expression comprehension (REC) [27]. Other methods explore multimodal large language models (MLLMs) for detection and grounding tasks. Florence-2 [48], for instance, introduces a prompt-based model that generates object categories and coordinates using text prompts directly. The diverse feature extractors, architectures, and outputs in object detection make it challenging to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30043" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 282, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 282, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 282, + 84 + ], + "type": "text", + "content": "interpret their decisions using a single unified approach." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 85, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 85, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 85, + 296, + 373 + ], + "type": "text", + "content": "Object Detection Explanation. Explaining object detector decisions remains a largely unexplored area. Gudovskiy et al. [12] used Integrated Gradients (IG) [38] and SHAP [36] with SSD [29] classification scores for box-level attribution. Lee et al. [19] optimized masks on Mask R-CNN [14] to approximate original decisions with minimal pixels. Petsiuk et al. [33] proposed D-RISE, an adaptation of RISE [32], though it often introduces noise in saliency maps. Jiang et al. [16] introduced Nesterov-Accelerated iGOS++, which can be resource-intensive for large models. Zhao et al. [54] developed ODAM using Grad-CAM [37], but its effectiveness depends on network layer choices. SSGrad-CAM++ [49] generates saliency maps based on Grad-CAM++ [3]. VX-CODE [50] attributes important patches based on greedy search and SHAP [36]. While these methods aim to interpret detector errors, they largely rely on simple visualizations and empirical observations, often overlooking input-level feature confusion that can obscure human understanding. Masking confusing regions may aid in correcting model outputs [6]. Our paper introduces a black-box interpretation approach for object-level foundation models using Visual Precision Search, enhancing detection performance with fewer regions and identifying input-level failure causes to improve detection accuracy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 381, + 111, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 111, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 111, + 393 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 401, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 437 + ], + "type": "text", + "content": "This section presents our method for explaining the decisions of object-level foundation models. Figure 2 illustrates the overall framework of our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 180, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 180, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 180, + 456 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": "Given an image " + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": " and an object-level foundation model " + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": ", the output can be represented as " + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "inline_equation", + "content": "f(\\mathbf{I}) = \\{(b_i, c_i, s_i) \\mid i = 1, 2, \\dots, N\\}" + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": ". Each tuple " + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "inline_equation", + "content": "(b_i, c_i, s_i)" + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": " denotes the bounding box, class label, and confidence score for each detected object " + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": " is the maximum number of detection boxes. Tasks like object detection involve multiple boxes and categories, while visual grounding tasks like referring expression comprehension (REC) have only one box and category. Our goal is to generate a saliency map that explains the reasons behind the model's detection of a specific object, " + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "inline_equation", + "content": "(b_i, c_i, s_i)" + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": ". Ideally, this explanation would enable the correct detection of objects using fewer input regions, while also requiring the removal of only a few critical regions to invalidate the detector. To achieve this, we can sparsify the input region " + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "inline_equation", + "content": "V = \\{\\mathbf{I}_1^s, \\dots, \\mathbf{I}_m^s\\}" + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i^s" + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": "-th sub-region, " + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": " is the total number of sub-regions, and " + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\sum_{i=1}^{m} \\mathbf{I}_i^s" + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": ". We then frame the saliency map generation task as a subset selection problem for these input sub-regions. A set function " + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 55, + 461, + 296, + 713 + ], + "type": "text", + "content": " is defined to assess interpretability by determining whether a given region is a key factor in the model's decision. Therefore, the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 375, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 375, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 375, + 83 + ], + "type": "text", + "content": "objectives are:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 399, + 86, + 553, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 86, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 399, + 86, + 553, + 105 + ], + "type": "interline_equation", + "content": "\\max _ {S \\subseteq V, | S | < k} \\mathcal {F} (S), \\tag {1}", + "image_path": "0c34f7fe1f647719c0734a0e1944c5a57b40b27fd5fecfa4d5a8fb0053b121dd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 112, + 553, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 112, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 112, + 553, + 148 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 112, + 553, + 148 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 112, + 553, + 148 + ], + "type": "text", + "content": " denotes the maximum number of sub-regions. Therefore, the key to this problem lies in designing set function " + }, + { + "bbox": [ + 313, + 112, + 553, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 112, + 553, + 148 + ], + "type": "text", + "content": " and optimizing Eq. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 158, + 447, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 158, + 447, + 170 + ], + "spans": [ + { + "bbox": [ + 314, + 158, + 447, + 170 + ], + "type": "text", + "content": "3.2. Visual Precision Search" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "text", + "content": "We propose a Visual Precision Search method for interpreting object-level models. To begin, the input region needs to be sparsified. We apply the SLICO superpixel segmentation algorithm [1] to divide the input into " + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "text", + "content": " sub-regions, " + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "inline_equation", + "content": "V = \\{\\mathbf{I}_1^s,\\dots ,\\mathbf{I}_m^s\\}" + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "text", + "content": ". To solve Eq. 1, an " + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{NP}" + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "text", + "content": "-hard problem, we employ submodular optimization [10]. Since the saliency map requires all sub-regions to be ranked, " + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "text", + "content": " can be set to " + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "inline_equation", + "content": "|V|" + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "text", + "content": " to compute ordered subsets. When the set function " + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "text", + "content": " satisfies the properties of diminishing returns1 and monotonic non-negative2, a greedy search guarantees an approximate optimal solution [8], with " + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(S)\\geq (1 - 1 / e)\\mathcal{F}(S^{*})" + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "inline_equation", + "content": "S^{*}" + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "text", + "content": " denotes the optimal solution and " + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 175, + 555, + 355 + ], + "type": "text", + "content": " denotes the greedy solution. We next design a set function to evaluate the interpretability score and rank the importance of interpretable regions for object-level tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "text", + "content": "**Clue Score:** An essential aspect of interpretability is enabling the object-level foundation model to accurately locate and identify objects while using fewer regions. To assess the importance of subregions from this perspective, we introduce the object location box information, " + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "inline_equation", + "content": "\\pmb{b}_{\\mathrm{target}}" + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "text", + "content": ", and the target category, " + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "text", + "content": ", that requires explanation. Given a subregion " + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "text", + "content": ", the object-level model outputs " + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "text", + "content": " instances, denoted as " + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "inline_equation", + "content": "f(S) = \\{(b_i, s_{c,i}) \\mid i = 1, \\dots, N\\}" + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "inline_equation", + "content": "s_{c,i}" + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "text", + "content": " represents the confidence of the " + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "text", + "content": "-th bounding box in predicting category " + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "text", + "content": ". Then, the clue score of sub-region " + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 356, + 556, + 488 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 498, + 555, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 498, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 321, + 498, + 555, + 519 + ], + "type": "interline_equation", + "content": "s _ {\\text {c l u e}} \\left(S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c\\right) = \\max _ {\\left(\\boldsymbol {b} _ {i}, s _ {c, i}\\right) \\in f (S)} \\operatorname {I o U} \\left(\\boldsymbol {b} _ {\\text {t a r g e t}}, \\boldsymbol {b} _ {i}\\right) \\cdot s _ {c, i}, \\tag {2}", + "image_path": "ab1f6d30a59e980dd50d53d9a800d8f4a51054465ec79167f5714910d93ad5cd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 529, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 529, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 529, + 555, + 613 + ], + "type": "text", + "content": " represents the Intersection over Union between two bounding boxes. Unlike Petsiuk et al. [33], which only considers high-confidence bounding boxes, our method includes all candidate boxes to avoid getting stuck in a local optimum during the search. By incorporating the clue score " + }, + { + "bbox": [ + 313, + 529, + 555, + 613 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{clue}}" + }, + { + "bbox": [ + 313, + 529, + 555, + 613 + ], + "type": "text", + "content": ", our method focuses on regions that strengthen the desired positional and semantic response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 613, + 556, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 556, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 556, + 673 + ], + "type": "text", + "content": "Collaboration Score: Some regions may exhibit strong combination effects, meaning they contribute effectively to model decisions only when paired with multiple specific sub-regions. Therefore, we introduce the collaboration score " + }, + { + "bbox": [ + 313, + 613, + 556, + 673 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{colla}}" + }, + { + "bbox": [ + 313, + 613, + 556, + 673 + ], + "type": "text", + "content": " to assess sub-regions with high sensitivity to de" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 682, + 553, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 682, + 553, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 682, + 553, + 703 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 313, + 682, + 553, + 703 + ], + "type": "text", + "content": "Diminishing returns: " + }, + { + "bbox": [ + 313, + 682, + 553, + 703 + ], + "type": "inline_equation", + "content": "\\forall S_A \\subseteq S_B \\subseteq V \\setminus \\alpha" + }, + { + "bbox": [ + 313, + 682, + 553, + 703 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 682, + 553, + 703 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(S_A \\cup \\{\\alpha\\}) - \\mathcal{F}(S_A) \\geq \\mathcal{F}(S_B \\cup \\{\\alpha\\}) - \\mathcal{F}(S_B)" + }, + { + "bbox": [ + 313, + 682, + 553, + 703 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 703, + 552, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 703, + 552, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 703, + 552, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 326, + 703, + 552, + 713 + ], + "type": "text", + "content": "Monotonic non-negative: " + }, + { + "bbox": [ + 326, + 703, + 552, + 713 + ], + "type": "inline_equation", + "content": "\\forall S \\subseteq V \\setminus \\alpha" + }, + { + "bbox": [ + 326, + 703, + 552, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 326, + 703, + 552, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(S \\cup \\{\\alpha\\}) - \\mathcal{F}(S) \\geq 0" + }, + { + "bbox": [ + 326, + 703, + 552, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30044" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 43, + 553, + 194 + ], + "blocks": [ + { + "bbox": [ + 58, + 43, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 43, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 43, + 553, + 194 + ], + "type": "image", + "image_path": "eaf7ef8b52307bca88e7cc7785e556bc3f953fd9d27dd770b73b87d821fe7bb6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 195, + 555, + 229 + ], + "lines": [ + { + "bbox": [ + 55, + 195, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 555, + 229 + ], + "type": "text", + "content": "Figure 2. Framework of the proposed Visual Precision Search method for interpreting an object-level foundation model. The input is first sparsified into a set of sub-regions and then interpreted across different instances. A submodular function guides the search for significant sub-regions, updating the ordered subset iteratively, and ultimately generating the instance-level attribution map." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 236, + 126, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 126, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 126, + 245 + ], + "type": "text", + "content": "cision outcomes:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 297, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 297, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 297, + 281 + ], + "type": "interline_equation", + "content": "s _ {\\text {c o l l a .}} (S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c) = 1 - \\max _ {(\\boldsymbol {b} _ {i}, s _ {c, i}) \\in f (V \\backslash S)} \\operatorname {I o U} (\\boldsymbol {b} _ {\\text {t a r g e t}}, \\boldsymbol {b} _ {i}) \\cdot s _ {c, i}, \\tag {3}", + "image_path": "cd77e68bf5b0ece6706499f3d5c2aeb61dcac69ce1fe418f4a6548dd47d539a8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 282, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 339 + ], + "type": "text", + "content": "this means that no detection box can accurately or confidently locate the object once the key sub-region is removed. This score serves as an effective guide in the initial search phase to identify subtle key regions that contribute to the final decision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 342, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 365 + ], + "type": "text", + "content": "Submodular Function: The scores above are combined to construct a submodular function " + }, + { + "bbox": [ + 55, + 342, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 55, + 342, + 296, + 365 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 370, + 295, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 370, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 61, + 370, + 295, + 384 + ], + "type": "interline_equation", + "content": "\\mathcal {F} (S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c) = s _ {\\text {c l u e}} (S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c) + s _ {\\text {c o l l a .}} (S, \\boldsymbol {b} _ {\\text {t a r g e t}}, c). \\tag {4}", + "image_path": "4db0feea11134bf135c5a48fc467079f1d7839cbeccd65cdf3abc617c3f30155.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 388, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 507 + ], + "type": "text", + "content": "Saliency Map Generation: Using the above submodular function, a greedy search algorithm is applied to sort all sub-regions in " + }, + { + "bbox": [ + 55, + 388, + 296, + 507 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 388, + 296, + 507 + ], + "type": "text", + "content": ", yielding an ordered subset " + }, + { + "bbox": [ + 55, + 388, + 296, + 507 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 388, + 296, + 507 + ], + "type": "text", + "content": ". Introducing the submodular function enables the search algorithm to more precisely identify key visual regions for interpretation. Additionally, scoring the sub-regions is necessary to better explain the importance of each sub-region. We evaluate the salient difference between the two sub-regions by the marginal effect. The attribution score " + }, + { + "bbox": [ + 55, + 388, + 296, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 55, + 388, + 296, + 507 + ], + "type": "text", + "content": " for each sub-region " + }, + { + "bbox": [ + 55, + 388, + 296, + 507 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 388, + 296, + 507 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 388, + 296, + 507 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 388, + 296, + 507 + ], + "type": "text", + "content": " is assessed by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 512, + 295, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 512, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 67, + 512, + 295, + 543 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {i} = \\left\\{ \\begin{array}{l l} b _ {\\text {b a s e}} & \\text {i f} i = 1, \\\\ \\mathcal {A} _ {i - 1} - \\left| \\mathcal {F} \\left(S _ {[ i ]}\\right) - \\mathcal {F} \\left(S _ {[ i - 1 ]}\\right) \\right| & \\text {i f} i > 1, \\end{array} \\right. \\tag {5}", + "image_path": "5c3ce7688f5e76762df656e59f19360e07257806ae7f837c4c5024e6e81c7837.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 547, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 296, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 547, + 296, + 666 + ], + "type": "inline_equation", + "content": "b_{\\mathrm{base}}" + }, + { + "bbox": [ + 55, + 547, + 296, + 666 + ], + "type": "text", + "content": " represents a baseline attribution score (such as 0 and won't affect results) for the first sub-region, and " + }, + { + "bbox": [ + 55, + 547, + 296, + 666 + ], + "type": "inline_equation", + "content": "S_{[i]}" + }, + { + "bbox": [ + 55, + 547, + 296, + 666 + ], + "type": "text", + "content": " denotes the set containing the top " + }, + { + "bbox": [ + 55, + 547, + 296, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 547, + 296, + 666 + ], + "type": "text", + "content": " sub-regions in " + }, + { + "bbox": [ + 55, + 547, + 296, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 547, + 296, + 666 + ], + "type": "text", + "content": ". When a new sub-region is added, a small marginal increase suggests comparable importance to the previous sub-region. A negative marginal effect indicates a counterproductive impact, which can be assessed by its absolute value. Finally, " + }, + { + "bbox": [ + 55, + 547, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 547, + 296, + 666 + ], + "type": "text", + "content": " is normalized to obtain the saliency map of the sub-region. The detailed calculation process of the proposed Visual Precision Search algorithm is outlined in Algorithm 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 672, + 155, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 155, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 155, + 685 + ], + "type": "text", + "content": "3.3. Theory Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "type": "text", + "content": "A theoretical analysis is conducted for our Visual Precision Search method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 237, + 528, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 237, + 528, + 273 + ], + "spans": [ + { + "bbox": [ + 320, + 237, + 528, + 273 + ], + "type": "text", + "content": "Algorithm 1: Visual Precision Search algorithm for generating interpretable saliency maps of object-level foundation models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 275, + 541, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 275, + 541, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 275, + 541, + 296 + ], + "type": "text", + "content": "Input: Image " + }, + { + "bbox": [ + 317, + 275, + 541, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 317, + 275, + 541, + 296 + ], + "type": "text", + "content": ", a division algorithm " + }, + { + "bbox": [ + 317, + 275, + 541, + 296 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}(\\cdot)" + }, + { + "bbox": [ + 317, + 275, + 541, + 296 + ], + "type": "text", + "content": ", the object location box " + }, + { + "bbox": [ + 317, + 275, + 541, + 296 + ], + "type": "inline_equation", + "content": "b_{\\text{target}}" + }, + { + "bbox": [ + 317, + 275, + 541, + 296 + ], + "type": "text", + "content": ", and the target category " + }, + { + "bbox": [ + 317, + 275, + 541, + 296 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 317, + 275, + 541, + 296 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 328, + 296, + 520, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 296, + 520, + 305 + ], + "spans": [ + { + "bbox": [ + 328, + 296, + 520, + 305 + ], + "type": "text", + "content": "Output: An ordered set " + }, + { + "bbox": [ + 328, + 296, + 520, + 305 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 328, + 296, + 520, + 305 + ], + "type": "text", + "content": " and a saliency map " + }, + { + "bbox": [ + 328, + 296, + 520, + 305 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\in \\mathbb{R}^{h \\times w}" + }, + { + "bbox": [ + 328, + 296, + 520, + 305 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 306, + 376, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 306, + 376, + 316 + ], + "spans": [ + { + "bbox": [ + 320, + 306, + 376, + 316 + ], + "type": "text", + "content": "1 " + }, + { + "bbox": [ + 320, + 306, + 376, + 316 + ], + "type": "inline_equation", + "content": "V\\gets \\mathrm{Div}(\\mathbf{I})" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 316, + 539, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 316, + 539, + 334 + ], + "spans": [ + { + "bbox": [ + 320, + 316, + 539, + 334 + ], + "type": "text", + "content": "2 " + }, + { + "bbox": [ + 320, + 316, + 539, + 334 + ], + "type": "inline_equation", + "content": "S\\gets \\emptyset" + }, + { + "bbox": [ + 320, + 316, + 539, + 334 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 320, + 316, + 539, + 334 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 320, + 316, + 539, + 334 + ], + "type": "text", + "content": " Initiate the operation of submodular subset selection " + }, + { + "bbox": [ + 320, + 316, + 539, + 334 + ], + "type": "inline_equation", + "content": "\\star/" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 334, + 370, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 334, + 370, + 343 + ], + "spans": [ + { + "bbox": [ + 320, + 334, + 370, + 343 + ], + "type": "text", + "content": "3 " + }, + { + "bbox": [ + 320, + 334, + 370, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_1\\gets b_{\\mathrm{base}}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 343, + 392, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 343, + 392, + 353 + ], + "spans": [ + { + "bbox": [ + 320, + 343, + 392, + 353 + ], + "type": "text", + "content": "4 for " + }, + { + "bbox": [ + 320, + 343, + 392, + 353 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 320, + 343, + 392, + 353 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 320, + 343, + 392, + 353 + ], + "type": "inline_equation", + "content": "|V|" + }, + { + "bbox": [ + 320, + 343, + 392, + 353 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 353, + 386, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 353, + 386, + 362 + ], + "spans": [ + { + "bbox": [ + 320, + 353, + 386, + 362 + ], + "type": "text", + "content": "5 " + }, + { + "bbox": [ + 320, + 353, + 386, + 362 + ], + "type": "inline_equation", + "content": "S_{d}\\gets V\\backslash S;" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 362, + 488, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 362, + 488, + 372 + ], + "spans": [ + { + "bbox": [ + 320, + 362, + 488, + 372 + ], + "type": "inline_equation", + "content": "\\alpha \\gets \\arg \\max_{\\alpha \\in S_d}\\mathcal{F}(S\\cup \\{\\alpha \\} ,\\pmb {b}_{\\mathrm{target}},c)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 372, + 394, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 372, + 394, + 381 + ], + "spans": [ + { + "bbox": [ + 320, + 372, + 394, + 381 + ], + "type": "text", + "content": "7 " + }, + { + "bbox": [ + 320, + 372, + 394, + 381 + ], + "type": "inline_equation", + "content": "S\\gets S\\cup \\{\\alpha \\} ;" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 381, + 387, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 381, + 387, + 390 + ], + "spans": [ + { + "bbox": [ + 320, + 381, + 387, + 390 + ], + "type": "text", + "content": "8 if " + }, + { + "bbox": [ + 320, + 381, + 387, + 390 + ], + "type": "inline_equation", + "content": "i > 1" + }, + { + "bbox": [ + 320, + 381, + 387, + 390 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 390, + 489, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 390, + 489, + 402 + ], + "spans": [ + { + "bbox": [ + 320, + 390, + 489, + 402 + ], + "type": "text", + "content": "9 " + }, + { + "bbox": [ + 320, + 390, + 489, + 402 + ], + "type": "inline_equation", + "content": "|\\quad | \\quad | \\quad \\mathcal{A}_i\\gets \\mathcal{A}_{i - 1} - |\\mathcal{F}(S_{[i]}) - \\mathcal{F}(S_{[i - 1]})|" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 319, + 403, + 341, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 403, + 341, + 410 + ], + "spans": [ + { + "bbox": [ + 319, + 403, + 341, + 410 + ], + "type": "text", + "content": "10 end" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 319, + 411, + 392, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 411, + 392, + 420 + ], + "spans": [ + { + "bbox": [ + 319, + 411, + 392, + 420 + ], + "type": "text", + "content": "11 return " + }, + { + "bbox": [ + 319, + 411, + 392, + 420 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 319, + 411, + 392, + 420 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 319, + 411, + 392, + 420 + ], + "type": "inline_equation", + "content": "\\mathrm{norm}(\\mathcal{A})" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "text", + "content": "Theorem 1 (Submodular Properties). Consider two subsets " + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "inline_equation", + "content": "S_A" + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "inline_equation", + "content": "S_B" + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "text", + "content": " in set " + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "inline_equation", + "content": "S_A \\subseteq S_B \\subseteq V" + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "text", + "content": ". Given an element " + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "inline_equation", + "content": "\\alpha \\in V \\setminus S_b" + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "text", + "content": ". Assuming that " + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "text", + "content": " is contributing to model interpretation, then, the function " + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 313, + 447, + 555, + 507 + ], + "type": "text", + "content": " in Eq. 4 satisfies diminishing returns," + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 327, + 514, + 541, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 514, + 541, + 529 + ], + "spans": [ + { + "bbox": [ + 327, + 514, + 541, + 529 + ], + "type": "interline_equation", + "content": "\\mathcal {F} \\left(S _ {A} \\cup \\{\\alpha \\}\\right) - \\mathcal {F} \\left(S _ {A}\\right) \\geq \\mathcal {F} \\left(S _ {B} \\cup \\{\\alpha \\}\\right) - \\mathcal {F} \\left(S _ {B}\\right),", + "image_path": "239b343e52fb2d8ae9c4b90c369f0600085f7a6bbc6c6c0e2364ac6eef4ef906.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 536, + 554, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 554, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 554, + 560 + ], + "type": "text", + "content": "and monotonic non-negative, " + }, + { + "bbox": [ + 313, + 536, + 554, + 560 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(S_A\\cup \\{\\alpha \\}) - \\mathcal{F}(S)\\geq 0" + }, + { + "bbox": [ + 313, + 536, + 554, + 560 + ], + "type": "text", + "content": " and thus, " + }, + { + "bbox": [ + 313, + 536, + 554, + 560 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 536, + 554, + 560 + ], + "type": "text", + "content": " is a submodular function." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 567, + 537, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 537, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 537, + 579 + ], + "type": "text", + "content": "Proof. Please see supplementary material for the proof." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 586, + 555, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 555, + 645 + ], + "type": "text", + "content": "Remark 1 (Impact on Sparse Division). The quality of the search space is determined by sparse division, meaning that both the method of partitioning the input and the number of sub-regions play a crucial role in the faithfulness of the Visual Precision Search." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "Remark 2 (Applicable Models). If the model provides bounding box positions and category confidence without filtering out low-confidence detection boxes, the interpretation results of the Visual Precision Search can achieve guaranteed optimal boundaries. Models with detection heads" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30045" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "are generally applicable; however, MLLM-based models may not directly output confidence scores, leaving room for improvement in the interpretation results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 138, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 138, + 134 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 138, + 134 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 141, + 172, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 141, + 172, + 154 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 172, + 154 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "text", + "content": "Datasets. We evaluate the proposed method on three object-level datasets: MS COCO 2017 [25], LVIS V1 [13], and RefCOCO [18]. The MS COCO dataset, used for interpreting object detection with 80 classes, involves random sampling of 5 correctly detected (IoU " + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "text", + "content": ", correct category), 3 misclassified (IoU " + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "text", + "content": ", wrong category), and 3 undetected (IoU " + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "inline_equation", + "content": "<0.5" + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "text", + "content": " or low confidence) instances per class. The LVIS V1 dataset, with 1,203 categories including 337 rare ones (1-10 images), uses Grounding DINO for interpreting zero-shot detection on rare classes, analyzing 165 correctly detected, 152 misclassified, and 534 undetected samples. The RefCOCO dataset, for interpreting referring expression comprehension (REC), evaluates 500 correctly grounded (IoU " + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "text", + "content": ") and 200 incorrectly grounded (IoU " + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "inline_equation", + "content": "<0.5" + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "text", + "content": ") samples per model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 339, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 422 + ], + "type": "text", + "content": "Baselines. The interpretation methods we compare include gradient-based approaches: Grad-CAM [37], SSGrad-CAM++[49], and ODAM[54], as well as perturbation-based methods: D-RISE [33] and D-HSIC [31]. D-RISE and D-HSIC are allowed to use all model outputs, including low-confidence detection boxes, to enhance the sufficiency of saliency estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 519 + ], + "type": "text", + "content": "Implementation Details. We validate our method on two object-level foundation models, Grounding DINO [28] and Florence-2 [48], across object detection and referring expression comprehension (REC) tasks. To interpret model decisions, we set the ground truth bounding box and category as target box " + }, + { + "bbox": [ + 55, + 423, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\pmb{b}_{\\mathrm{target}}" + }, + { + "bbox": [ + 55, + 423, + 296, + 519 + ], + "type": "text", + "content": " and category " + }, + { + "bbox": [ + 55, + 423, + 296, + 519 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 423, + 296, + 519 + ], + "type": "text", + "content": ". By default, the SLICO superpixel segmentation algorithm divides the image into 100 sparse sub-regions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 168, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 168, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 168, + 540 + ], + "type": "text", + "content": "4.2. Evaluation Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Faithfulness Metrics. We evaluate the faithfulness of interpretation results across the entire image using Insertion and Deletion scores from the detection task, following the setup by Petsiuk et al. [33]. These metrics reflect location and recognition accuracy. We also calculate classification and IoU scores for the box closest to the target explanation denoted as Insertion (class), Deletion (class), Insertion (IoU), and Deletion (IoU). Additionally, we compute the average highest confidence, defined as the highest class response for the box closest to the target with IoU " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": ". For failure samples, we introduce the Explaining Successful Rate (ESR) to evaluate whether the saliency map can effectively guide the search within a limited region, enabling the model to correctly detect misclassified or low-confidence objects" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "(with IoU " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " and classification confidence above the detector's default threshold) previously." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 96, + 555, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 555, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 555, + 157 + ], + "type": "text", + "content": "Location Metrics. We report the Point Game [53] and Energy Point Game [42] metrics. Since these metrics typically assume that the model's decision is relevant to the target object [33], we calculate them only for interpretation results on samples the model has correctly detected." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 165, + 437, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 437, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 437, + 177 + ], + "type": "text", + "content": "4.3. Faithfulness Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 182, + 486, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 486, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 486, + 194 + ], + "type": "text", + "content": "4.3.1. Faithfulness on Grounding DINO" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "type": "text", + "content": "We begin by validating our approach in explaining decisions across various object-level tasks using Grounding DINO [28]. Table 1 presents the faithfulness results across different datasets and tasks, demonstrating the strong interpretative faithfulness of our method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 258, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 426 + ], + "type": "text", + "content": "In the object detection interpretation task on the MS COCO dataset, our method achieved state-of-the-art results, outperforming the D-RISE method by " + }, + { + "bbox": [ + 313, + 258, + 555, + 426 + ], + "type": "inline_equation", + "content": "23.7\\%" + }, + { + "bbox": [ + 313, + 258, + 555, + 426 + ], + "type": "text", + "content": " on the Insertion metric, " + }, + { + "bbox": [ + 313, + 258, + 555, + 426 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 313, + 258, + 555, + 426 + ], + "type": "text", + "content": " on the Deletion metric, and " + }, + { + "bbox": [ + 313, + 258, + 555, + 426 + ], + "type": "inline_equation", + "content": "10.6\\%" + }, + { + "bbox": [ + 313, + 258, + 555, + 426 + ], + "type": "text", + "content": " on the average highest score. The Insertion and Deletion metrics also achieved top performance under both category and IoU conditions. Additionally, our method performed best in location metrics, indicating state-of-the-art performance in explaining general object detection tasks. We also found that gradient-based methods, such as ODAM, exhibit low faithfulness in explaining Grounding DINO decisions. This is primarily due to the fusion of text and visual modalities, which causes the gradient to be influenced by both, reducing its effectiveness in attributing to the visual modality alone." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "text", + "content": "In the referring expression comprehension (REC) interpretation task on the RefCOCO dataset, our method achieved state-of-the-art results across all evaluation metrics, surpassing the D-RISE method by " + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "inline_equation", + "content": "20.1\\%" + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "text", + "content": " on the Insertion metric, " + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "inline_equation", + "content": "22.1\\%" + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "text", + "content": " on the Deletion metric, and " + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "text", + "content": " on the average highest score. These results confirm that our method maintains high faithfulness in explaining object description understanding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 522, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 629 + ], + "type": "text", + "content": "We also interpret zero-shot object detection for rare classes on the LVIS V1 dataset. Our method achieved SOTA results in faithfulness metrics, surpassing D-RISE by " + }, + { + "bbox": [ + 313, + 522, + 554, + 629 + ], + "type": "inline_equation", + "content": "31.6\\%" + }, + { + "bbox": [ + 313, + 522, + 554, + 629 + ], + "type": "text", + "content": " on the Insertion metric, " + }, + { + "bbox": [ + 313, + 522, + 554, + 629 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 313, + 522, + 554, + 629 + ], + "type": "text", + "content": " on the Deletion metric, and " + }, + { + "bbox": [ + 313, + 522, + 554, + 629 + ], + "type": "inline_equation", + "content": "15.9\\%" + }, + { + "bbox": [ + 313, + 522, + 554, + 629 + ], + "type": "text", + "content": " on the average highest score. Although our method ranks second to ODAM on Energy Point Game metrics, this is mainly due to Grounding DINO not being fully trained on rare LVIS classes; as a result, model responses may be influenced by other regions rather than the object." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 631, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 554, + 713 + ], + "type": "text", + "content": "Experimental results show that our method significantly outperforms other interpretability approaches, primarily due to the submodularity of our objective function, enabling precise evaluation of critical sub-regions. As shown in Figure 3, ODAM's saliency maps are diffuse, and D-RISE's maps are noisy, whereas our method sharply highlights essential sub-regions, capturing edges and class-specific fea" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30046" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 76, + 553, + 244 + ], + "blocks": [ + { + "bbox": [ + 55, + 41, + 555, + 73 + ], + "lines": [ + { + "bbox": [ + 55, + 41, + 555, + 73 + ], + "spans": [ + { + "bbox": [ + 55, + 41, + 555, + 73 + ], + "type": "text", + "content": "Table 1. Evaluation of faithfulness metrics (Deletion, Insertion AUC scores, and average highest score) and location metrics (Point Game and Energy Point Game) on the MS-COCO, RefCOCO, and LVIS V1 (rare) validation sets for correctly detected or grounded samples using Grounding DINO." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 76, + 553, + 244 + ], + "lines": [ + { + "bbox": [ + 58, + 76, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 58, + 76, + 553, + 244 + ], + "type": "table", + "html": "
DatasetsMethodsFaithfulness MetricsLocation Metrics
Ins. (↑)Del. (↓)Ins. (class) (↑)Del. (class) (↓)Ins. (IoU) (↑)Del. (IoU) (↓)Ave. high. score (↑)Point Game (↑)Energy PG (↑)
MS COCO [25] (Detection task)Grad-CAM [37]0.24360.15260.30640.20060.62290.53240.59040.17460.1463
SSGrad-CAM++ [49]0.21070.17780.26390.23140.59810.55110.58860.19050.1293
D-RISE [33]0.44120.04020.50810.08860.83960.36420.62150.94970.1850
D-HSIC [31]0.37760.04390.43820.09030.83010.33010.58620.73280.1861
ODAM [54]0.31030.05190.36550.08940.78690.39840.58650.54310.2034
Ours0.54590.03750.62040.08820.85810.33000.68730.98940.2046
RefCOCO [18] (REC task)Grad-CAM [37]0.37490.42370.46580.51940.75160.76850.74810.23800.2171
SSGrad-CAM++ [49]0.41130.39250.50080.48510.77000.75880.75610.28200.2262
D-RISE [33]0.61780.16050.70330.33960.86060.51640.84710.94000.2870
D-HSIC [31]0.54910.18460.62950.35090.85040.51200.77390.79000.3190
ODAM [54]0.47780.27180.56200.37570.82170.66410.74250.63200.3529
Ours0.74190.12500.80800.24570.90500.51030.88420.94600.3566
LVIS V1 (rare) [13] (Zero-shot det. task)Grad-CAM [37]0.12530.12940.18010.18140.56570.59100.35490.11510.0941
SSGrad-CAM++ [49]0.12530.12540.17650.17750.58000.56910.35040.10910.0931
D-RISE [33]0.28080.02890.33480.08350.83030.31740.42890.96970.1462
D-HSIC [31]0.24170.03530.29120.09280.81870.35500.40440.83030.1730
ODAM [54]0.20090.04100.24780.08440.77600.40820.36940.60610.2050
Ours0.36950.02770.42750.07990.84790.32420.49690.97580.1785
", + "image_path": "a4114669db46f8330c4fdbfb173c1a646a85557253bef9dc5c89a8ed0335812d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 247, + 549, + 400 + ], + "blocks": [ + { + "bbox": [ + 61, + 247, + 549, + 400 + ], + "lines": [ + { + "bbox": [ + 61, + 247, + 549, + 400 + ], + "spans": [ + { + "bbox": [ + 61, + 247, + 549, + 400 + ], + "type": "image", + "image_path": "06fcedbef2942fbaf3880fe70e2342bc62c7fa85e7f171e519aa3b463d587286.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 403, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 555, + 436 + ], + "type": "text", + "content": "Figure 3. Visualization results of Grounding DINO for interpreting object localization tasks on the MS COCO, RefCOCO, and LVIS datasets. The first column shows the saliency map with the ground truth box and label. The second column presents detection within the limited search region, and the third columns display the insertion curves." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 447, + 194, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 194, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 194, + 459 + ], + "type": "text", + "content": "tures that enhance interpretability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 196, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 196, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 196, + 479 + ], + "type": "text", + "content": "4.3.2. Faithfulness on Florence-2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 555 + ], + "type": "text", + "content": "We further validated our approach on Florence-2 [48], a foundation model lacking object confidence scores. This limitation requires our search to rely solely on IoU as a guiding metric, excluding comparison with gradient-based methods. As shown in Table 2, our method achieves state-of-the-art performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 557, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 700 + ], + "type": "text", + "content": "In the object detection interpretability task on the MS COCO dataset, our method outperforms D-RISE by " + }, + { + "bbox": [ + 55, + 557, + 296, + 700 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 55, + 557, + 296, + 700 + ], + "type": "text", + "content": " on the Insertion metric and achieves a substantial improvement of " + }, + { + "bbox": [ + 55, + 557, + 296, + 700 + ], + "type": "inline_equation", + "content": "50.7\\%" + }, + { + "bbox": [ + 55, + 557, + 296, + 700 + ], + "type": "text", + "content": " on the Deletion metric. Additionally, our method enhances the Point Game and Energy Point Game metrics by " + }, + { + "bbox": [ + 55, + 557, + 296, + 700 + ], + "type": "inline_equation", + "content": "8.3\\%" + }, + { + "bbox": [ + 55, + 557, + 296, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 557, + 296, + 700 + ], + "type": "inline_equation", + "content": "60.7\\%" + }, + { + "bbox": [ + 55, + 557, + 296, + 700 + ], + "type": "text", + "content": ", respectively. Due to the absence of object confidence scores in Florence-2, it becomes challenging to identify areas that quickly invalidate the model's detection results. Compared to perturbation-based methods, our approach offers significant advantages; it relies more directly on the object itself, resulting in stronger faithfulness to the model's detection behavior." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "We also interpret referring expression comprehension" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 447, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 554, + 519 + ], + "type": "text", + "content": "(REC) interpretation task on the RefCOCO dataset, outperforming D-RISE by " + }, + { + "bbox": [ + 313, + 447, + 554, + 519 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 313, + 447, + 554, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 447, + 554, + 519 + ], + "type": "inline_equation", + "content": "66.9\\%" + }, + { + "bbox": [ + 313, + 447, + 554, + 519 + ], + "type": "text", + "content": " in Insertion and Deletion metrics, respectively. Our approach also achieved state-of-the-art localization results, with a " + }, + { + "bbox": [ + 313, + 447, + 554, + 519 + ], + "type": "inline_equation", + "content": "14.6\\%" + }, + { + "bbox": [ + 313, + 447, + 554, + 519 + ], + "type": "text", + "content": " improvement in the Energy Point Game, demonstrating strong interpretability in accurately pinpointing object information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 519, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 554, + 590 + ], + "type": "text", + "content": "We visualized some results of our method in Figure 4, illustrating that it effectively identifies key information about the object's location and categorical description. While D-RISE performed well on the Insertion metric, its attribution results were diffuse and lacked reliability, and it did not perform as well on the Deletion metric." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 597, + 539, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 539, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 539, + 610 + ], + "type": "text", + "content": "4.4. Interpreting Failures in Object-Level Tasks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 615, + 455, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 455, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 455, + 627 + ], + "type": "text", + "content": "4.4.1. Interpreting REC Failures" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "Failures in the REC task primarily arise from visual grounding errors, often due to interference from other objects in the scene that become entangled with the text. Removing this distracting visual information can clarify grounding results and improve error understanding at the visual input level. To achieve this, we specify the correct location as the target box, denoted as " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{b}_{\\mathrm{target}}" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": ", for input attribution. Ta" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30047" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 117, + 296, + 187 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "type": "text", + "content": "Table 2. Evaluation of faithfulness metrics (Deletion and Insertion AUC scores) and location metrics (Point Game and Energy Point Game) on the MS COCO and RefCOCO validation sets for correctly detected and grounded samples using Florence-2." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 117, + 296, + 187 + ], + "lines": [ + { + "bbox": [ + 58, + 117, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 58, + 117, + 296, + 187 + ], + "type": "table", + "html": "
DatasetsMethodsFaithfulness MetricsLocation Metrics
Insertion (↑)Deletion (↓)Point Game (↑)Energy PG (↑)
MS COCO [25] (Detection task)D-RISE [33]0.74770.09720.88500.1568
D-HSIC [31]0.53450.27300.29250.0862
Ours0.77590.04790.95830.2519
RefCOCO [18] (REC task)D-RISE [33]0.79220.35050.84800.2464
D-HSIC [31]0.76390.35600.69800.2754
Ours0.84090.11590.86600.3927
", + "image_path": "16829e560097f2656b49767968956ccf8c3ce375da67aad6154c0342b2f0eb7b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 60, + 205, + 293, + 297 + ], + "blocks": [ + { + "bbox": [ + 114, + 196, + 257, + 205 + ], + "lines": [ + { + "bbox": [ + 114, + 196, + 257, + 205 + ], + "spans": [ + { + "bbox": [ + 114, + 196, + 257, + 205 + ], + "type": "text", + "content": "MS COCO (Object Detection Interpretation)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 205, + 293, + 297 + ], + "lines": [ + { + "bbox": [ + 60, + 205, + 293, + 297 + ], + "spans": [ + { + "bbox": [ + 60, + 205, + 293, + 297 + ], + "type": "image", + "image_path": "168889f26d4bbeab9004c4ea9077f0b9164776fde77f0e75992663ed758ed185.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 307, + 293, + 401 + ], + "blocks": [ + { + "bbox": [ + 113, + 298, + 257, + 307 + ], + "lines": [ + { + "bbox": [ + 113, + 298, + 257, + 307 + ], + "spans": [ + { + "bbox": [ + 113, + 298, + 257, + 307 + ], + "type": "text", + "content": "RefCOCO (Visual Grounding Interpretation)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 307, + 293, + 401 + ], + "lines": [ + { + "bbox": [ + 60, + 307, + 293, + 401 + ], + "spans": [ + { + "bbox": [ + 60, + 307, + 293, + 401 + ], + "type": "image", + "image_path": "f7fd96ec7d9b4846ebb8ca56d73853a3204cfc66e17d02554e24e50fc782f895.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 403, + 295, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 425 + ], + "type": "text", + "content": "Figure 4. Visualization results of Florence-2 for interpreting object localization tasks on the MS COCO and RefCOCO datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 436, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 436, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 436, + 295, + 533 + ], + "type": "text", + "content": "ble 3 shows the attribution results, where gradient-based methods exhibit limited performance in faithfulness metrics. Compared to D-RISE, our approach improves Insertion by " + }, + { + "bbox": [ + 54, + 436, + 295, + 533 + ], + "type": "inline_equation", + "content": "42.9\\%" + }, + { + "bbox": [ + 54, + 436, + 295, + 533 + ], + "type": "text", + "content": ", class score-based Insertion by " + }, + { + "bbox": [ + 54, + 436, + 295, + 533 + ], + "type": "inline_equation", + "content": "25.1\\%" + }, + { + "bbox": [ + 54, + 436, + 295, + 533 + ], + "type": "text", + "content": ", and average highest score by " + }, + { + "bbox": [ + 54, + 436, + 295, + 533 + ], + "type": "inline_equation", + "content": "14.9\\%" + }, + { + "bbox": [ + 54, + 436, + 295, + 533 + ], + "type": "text", + "content": ". Figure 5 illustrates model error interpretations, with the cyan-highlighted region indicating decision errors due to visual input interference that leads the foundation model astray." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 540, + 217, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 217, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 217, + 553 + ], + "type": "text", + "content": "4.4.2. Interpreting Detection Failures" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "text", + "content": "Detection failures include misclassification and undetected objects. While previous works [33, 54] used visualization to explain errors, it lacked quantitative analysis, resulting in lower trustworthiness. We used saliency maps to identify input-level interference causing detection failures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Misclassified Interpretation. As shown in Table 4, On the MS COCO dataset, our method outperforms D-RISE, improving Insertion by " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "54.7\\%" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ", Insertion (class) by " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "49.1\\%" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ", average highest score by " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "27.4\\%" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ", and explaining success rate (ESR) by " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "19.47\\%" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ". Similarly, on the LVIS dataset, our method shows advantages with improvements of " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "42.7\\%" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": " in Insertion, " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "33.0\\%" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": " in Insertion (class), " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "24.8\\%" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": " in average highest score, and " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "24.34\\%" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": " in ESR. This improvement is pri" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 323, + 105, + 547, + 173 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 3. Insertion AUC scores and the average highest score on the RefCOCO validation sets for or the samples with incorrect localization in visual grounding using Grounding DINO." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 105, + 547, + 173 + ], + "lines": [ + { + "bbox": [ + 323, + 105, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 323, + 105, + 547, + 173 + ], + "type": "table", + "html": "
DatasetsMethodsFaithfulness Metrics
Ins. (†)Ins. (class) (†)Ave. high. score (†)
RefCOCO [18] (REC task)Grad-CAM [37]0.15360.27940.3295
SSGrad-CAM++ [49]0.15900.28370.3266
D-RISE [33]0.34860.47870.6096
D-HSIC [31]0.22740.34880.4495
ODAM [54]0.17930.30010.3453
Ours0.49810.59900.7007
", + "image_path": "71482d55bf8c3d1aeea55bdcfcc1b6d1e40639e13ca41aea960ea1af72836cc6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 190, + 391, + 237 + ], + "blocks": [ + { + "bbox": [ + 329, + 182, + 373, + 190 + ], + "lines": [ + { + "bbox": [ + 329, + 182, + 373, + 190 + ], + "spans": [ + { + "bbox": [ + 329, + 182, + 373, + 190 + ], + "type": "text", + "content": "Attribution Map" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 190, + 391, + 237 + ], + "lines": [ + { + "bbox": [ + 316, + 190, + 391, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 190, + 391, + 237 + ], + "type": "image", + "image_path": "bcdc9a12c9905eed0b174787847dc65d369e8b52c586c37900f296874c6e46c5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 399, + 190, + 469, + 236 + ], + "blocks": [ + { + "bbox": [ + 408, + 182, + 459, + 190 + ], + "lines": [ + { + "bbox": [ + 408, + 182, + 459, + 190 + ], + "spans": [ + { + "bbox": [ + 408, + 182, + 459, + 190 + ], + "type": "text", + "content": "Searched Region" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 190, + 469, + 236 + ], + "lines": [ + { + "bbox": [ + 399, + 190, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 399, + 190, + 469, + 236 + ], + "type": "image", + "image_path": "624dba656a394fb44bce58772fae49a4e3ec2407027bf4cd67bfb4b3a119ab3a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 484, + 178, + 551, + 247 + ], + "blocks": [ + { + "bbox": [ + 484, + 178, + 551, + 247 + ], + "lines": [ + { + "bbox": [ + 484, + 178, + 551, + 247 + ], + "spans": [ + { + "bbox": [ + 484, + 178, + 551, + 247 + ], + "type": "image", + "image_path": "18ec6165b373b0df06406c8966afde8776f605462b14c7e9bb46ed9cb5946570.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 247, + 553, + 269 + ], + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 269 + ], + "type": "text", + "content": "Figure 5. Visualization of the method for discovering what causes the Grounding DINO incorrectly grounded on RefCOCO." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 317, + 306, + 553, + 408 + ], + "blocks": [ + { + "bbox": [ + 313, + 271, + 553, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 271, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 553, + 304 + ], + "type": "text", + "content": "Table 4. Insertion AUC scores, average highest score, and explaining successful rate (ESR) on the MS-COCO and the LVIS validation sets for misclassified samples using Grounding DINO." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 306, + 553, + 408 + ], + "lines": [ + { + "bbox": [ + 317, + 306, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 306, + 553, + 408 + ], + "type": "table", + "html": "
DatasetsMethodsFaithfulness Metrics
Ins. (↑)Ins. (class) (↑)Ave. high. score (↑)ESR (↑)
MS COCO [25] (Detection task)Grad-CAM [37]0.10910.14780.310238.38%
SSGrad-CAM++ [49]0.09600.13360.295233.51%
D-RISE [33]0.21700.26610.360350.26%
D-HSIC [31]0.17710.21610.314334.59%
ODAM [54]0.11290.14860.286932.97%
Ours0.33570.39670.459169.73%
LVIS V1 (rare) [13] (Zero-shot det. task)Grad-CAM [37]0.05030.08910.156412.50%
SSGrad-CAM++ [49]0.05740.09460.158011.84%
D-RISE [33]0.12450.16470.208828.95%
D-HSIC [31]0.09630.12470.174816.45%
ODAM [54]0.05750.09540.15209.21%
Ours0.17760.21900.260653.29%
", + "image_path": "9b758500f49a640018ddec97415d499e0cc772f34cd8e526ee624f83e94bf49b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 318, + 413, + 378, + 512 + ], + "blocks": [ + { + "bbox": [ + 318, + 413, + 378, + 512 + ], + "lines": [ + { + "bbox": [ + 318, + 413, + 378, + 512 + ], + "spans": [ + { + "bbox": [ + 318, + 413, + 378, + 512 + ], + "type": "image", + "image_path": "5e2aadfbc73247ae622522d08a8657860edf3733d7065e22a60be577012c6570.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 383, + 413, + 434, + 511 + ], + "blocks": [ + { + "bbox": [ + 383, + 413, + 434, + 511 + ], + "lines": [ + { + "bbox": [ + 383, + 413, + 434, + 511 + ], + "spans": [ + { + "bbox": [ + 383, + 413, + 434, + 511 + ], + "type": "image", + "image_path": "b1b48ac7a890ea8b98c764067c57e0a6987dba755e7201046703b7f46a81d881.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 437, + 413, + 489, + 511 + ], + "blocks": [ + { + "bbox": [ + 437, + 413, + 489, + 511 + ], + "lines": [ + { + "bbox": [ + 437, + 413, + 489, + 511 + ], + "spans": [ + { + "bbox": [ + 437, + 413, + 489, + 511 + ], + "type": "image", + "image_path": "e2767f4fec27d15220a39d7a33cef5eb2e72f6aad5b2f699395aaac9e3cad41d.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 490, + 413, + 552, + 517 + ], + "blocks": [ + { + "bbox": [ + 490, + 413, + 552, + 517 + ], + "lines": [ + { + "bbox": [ + 490, + 413, + 552, + 517 + ], + "spans": [ + { + "bbox": [ + 490, + 413, + 552, + 517 + ], + "type": "image", + "image_path": "8b9a3766fa0276b71e5b955eb94ef5ffa3076c9620ca85bdeafee9eb46036771.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 517, + 555, + 560 + ], + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 560 + ], + "type": "text", + "content": "Figure 6. Visualization of our method reveals the causes of Grounding DINO misclassifications on MS COCO and LVIS. The cyan region in the saliency map highlights the regions responsible for the model's misclassification." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 568, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 653 + ], + "type": "text", + "content": "marily due to our proposed Visual Precision Search, which confines attribution to the target negative response region, effectively refining the model's detection results. Figure 6 shows that the background surrounding the object interferes with the model's decision-making. Improving the model by refining the contextual relationship between foreground and background may be a promising direction." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Undetected Interpretation. Low object confidence may result from both the model's feature representation and confounding input-level factors. We analyze causes of missed detections, with Table 5 showing that our method achieves SOTA performance across metrics. On MS" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30048" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 87, + 296, + 190 + ], + "blocks": [ + { + "bbox": [ + 55, + 55, + 295, + 87 + ], + "lines": [ + { + "bbox": [ + 55, + 55, + 295, + 87 + ], + "spans": [ + { + "bbox": [ + 55, + 55, + 295, + 87 + ], + "type": "text", + "content": "Table 5. Insertion, average highest score, and explaining successful rate (ESR) on the MS-COCO and the LVIS V1 (rare) validation sets for missed detection samples using Grounding DINO." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 87, + 296, + 190 + ], + "lines": [ + { + "bbox": [ + 57, + 87, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 57, + 87, + 296, + 190 + ], + "type": "table", + "html": "
DatasetsMethodsFaithfulness Metrics
Ins. (↑)Ins. (class) (↑)Ave. high. score (↑)ESR (↑)
MS COCO [25] (Detection task)Grad-CAM [37]0.07600.13210.215316.44%
SSGrad-CAM++ [49]0.06710.11510.212416.44%
D-RISE [33]0.15380.22600.256426.94%
D-HSIC [31]0.11010.17160.194513.56%
ODAM [54]0.07450.13500.203713.78%
Ours0.21020.30110.301441.33%
LVIS V1 (rare) [13] (Zero-shot det. task)Grad-CAM [37]0.02910.06890.09015.43%
SSGrad-CAM++ [49]0.02920.06800.08975.24%
D-RISE [33]0.07030.11840.131218.73%
D-HSIC [31]0.05160.09200.116813.48%
ODAM [54]0.02830.07160.08514.68%
Ours0.11550.18860.178430.15%
", + "image_path": "790e471d798106b528f8f881dd03cef9f7702869ca6adec0f8cd03b61dff1d36.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 213, + 141, + 339 + ], + "blocks": [ + { + "bbox": [ + 59, + 213, + 141, + 339 + ], + "lines": [ + { + "bbox": [ + 59, + 213, + 141, + 339 + ], + "spans": [ + { + "bbox": [ + 59, + 213, + 141, + 339 + ], + "type": "image", + "image_path": "20aeadba9f1187fd25ce0e0faa0f9f4b6a4f06d8b2b970286795cbb1ac85972d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 213, + 212, + 339 + ], + "blocks": [ + { + "bbox": [ + 149, + 213, + 212, + 339 + ], + "lines": [ + { + "bbox": [ + 149, + 213, + 212, + 339 + ], + "spans": [ + { + "bbox": [ + 149, + 213, + 212, + 339 + ], + "type": "image", + "image_path": "eb25181b7122be4733e9b26602288f904e81c94c1539f47e630e9d66cef0a48d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 217, + 213, + 293, + 345 + ], + "blocks": [ + { + "bbox": [ + 217, + 213, + 293, + 345 + ], + "lines": [ + { + "bbox": [ + 217, + 213, + 293, + 345 + ], + "spans": [ + { + "bbox": [ + 217, + 213, + 293, + 345 + ], + "type": "image", + "image_path": "7df524c5ee250a3735a17b2654078647ea8f5fe25ef1c8c97bc8fa9073e4ceaf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 346, + 295, + 389 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 389 + ], + "type": "text", + "content": "Figure 7. Visualization of our method reveals the causes of Grounding DINO undetected on MS COCO and LVIS. The cyan region in the saliency map highlights the regions responsible for the model's detection failure." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 504 + ], + "type": "text", + "content": "COCO, it outperforms D-RISE with a " + }, + { + "bbox": [ + 55, + 396, + 296, + 504 + ], + "type": "inline_equation", + "content": "36.7\\%" + }, + { + "bbox": [ + 55, + 396, + 296, + 504 + ], + "type": "text", + "content": " improvement in Insertion and a " + }, + { + "bbox": [ + 55, + 396, + 296, + 504 + ], + "type": "inline_equation", + "content": "14.39\\%" + }, + { + "bbox": [ + 55, + 396, + 296, + 504 + ], + "type": "text", + "content": " gain in ESR, while on LVIS, it improves Insertion by " + }, + { + "bbox": [ + 55, + 396, + 296, + 504 + ], + "type": "inline_equation", + "content": "64.3\\%" + }, + { + "bbox": [ + 55, + 396, + 296, + 504 + ], + "type": "text", + "content": " and ESR by " + }, + { + "bbox": [ + 55, + 396, + 296, + 504 + ], + "type": "inline_equation", + "content": "11.42\\%" + }, + { + "bbox": [ + 55, + 396, + 296, + 504 + ], + "type": "text", + "content": ". Figure 7 visualizes our method's explanations for undetected instances, revealing that errors may arise from challenges in distinguishing similar objects (e.g., the bear in the first row) and environmental influence on detection (e.g., the corkboard in the second row). These insights highlight current model limitations, offering directions for improvement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 514, + 149, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 149, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 149, + 528 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 533, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 664 + ], + "type": "text", + "content": "Ablation of the Submodular Function. We analyze the effectiveness of our submodular function design on Grounding DINO using correctly detected samples from MS COCO. Table 6 shows the results. The Clue Score identifies regions beneficial for detection, improving Insertion and average highest score, while the Collaboration Score pinpoints sensitive regions, enhancing the Deletion metric by causing the detector to fail more quickly. Combining these scores enables our method to achieve optimal results across indicators, demonstrating the effectiveness of each score function within the submodular function." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Ablation on Divided Sub-region Number. The sub-region number affects search space quality. Using Grounding DINO, we assess its impact on correctly predicted and misclassified samples in MS COCO. Figures 8A-B show" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 331, + 79, + 539, + 130 + ], + "blocks": [ + { + "bbox": [ + 313, + 55, + 553, + 76 + ], + "lines": [ + { + "bbox": [ + 313, + 55, + 553, + 76 + ], + "spans": [ + { + "bbox": [ + 313, + 55, + 553, + 76 + ], + "type": "text", + "content": "Table 6. Ablation study on function score components for Grounding DINO on the MS COCO validation set." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 79, + 539, + 130 + ], + "lines": [ + { + "bbox": [ + 331, + 79, + 539, + 130 + ], + "spans": [ + { + "bbox": [ + 331, + 79, + 539, + 130 + ], + "type": "table", + "html": "
Clue Score (Eq. 2)Colla. Score (Eq. 3)Insertion (↑)Faithfulness Metrics
Deletion (↓)Ave. high. score (↑)
✗✓0.36320.03780.5967
✓✗0.53700.07990.6864
✓✓0.54590.03750.6873
", + "image_path": "b1db99e6d8292562ca6a2484535ab5552e0a6c07a1824fb7f08d779ff3681b5a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 320, + 143, + 394, + 207 + ], + "blocks": [ + { + "bbox": [ + 320, + 143, + 394, + 207 + ], + "lines": [ + { + "bbox": [ + 320, + 143, + 394, + 207 + ], + "spans": [ + { + "bbox": [ + 320, + 143, + 394, + 207 + ], + "type": "image", + "image_path": "036d93524432f2eb3c3404764702619843f7ec9b485613259b3239f3a22ab99e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 397, + 143, + 472, + 207 + ], + "blocks": [ + { + "bbox": [ + 397, + 143, + 472, + 207 + ], + "lines": [ + { + "bbox": [ + 397, + 143, + 472, + 207 + ], + "spans": [ + { + "bbox": [ + 397, + 143, + 472, + 207 + ], + "type": "image", + "image_path": "d194756454313e106e54a7438eb6385fbb655c7dfc6f08371d0f431038d27ac8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 476, + 143, + 551, + 207 + ], + "blocks": [ + { + "bbox": [ + 476, + 143, + 551, + 207 + ], + "lines": [ + { + "bbox": [ + 476, + 143, + 551, + 207 + ], + "spans": [ + { + "bbox": [ + 476, + 143, + 551, + 207 + ], + "type": "image", + "image_path": "32360a8a3ee25e5b4d89474aa6d4d5dfa81105e1236700541e4d8c659e6af12e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 321, + 207, + 394, + 270 + ], + "blocks": [ + { + "bbox": [ + 321, + 207, + 394, + 270 + ], + "lines": [ + { + "bbox": [ + 321, + 207, + 394, + 270 + ], + "spans": [ + { + "bbox": [ + 321, + 207, + 394, + 270 + ], + "type": "image", + "image_path": "8153aa83ef8498af7d707a9f49587a9054d965f453bb4bcbb1867f1b39ef0fd7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 272, + 553, + 338 + ], + "lines": [ + { + "bbox": [ + 313, + 272, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 553, + 338 + ], + "type": "text", + "content": "Figure 8. Ablation on the number of sub-regions. For correct detections, panels " + }, + { + "bbox": [ + 313, + 272, + 553, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 272, + 553, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 272, + 553, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 272, + 553, + 338 + ], + "type": "text", + "content": " show the effect of sub-region count on Insertion and average highest score, respectively. Panel " + }, + { + "bbox": [ + 313, + 272, + 553, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 272, + 553, + 338 + ], + "type": "text", + "content": " presents the inference time for different sub-region counts. For misclassified samples, panels " + }, + { + "bbox": [ + 313, + 272, + 553, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{D},\\mathbf{E}," + }, + { + "bbox": [ + 313, + 272, + 553, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 272, + 553, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 272, + 553, + 338 + ], + "type": "text", + "content": " display the effects on Insertion, average highest score, and explanation success rate." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 397, + 207, + 473, + 270 + ], + "blocks": [ + { + "bbox": [ + 397, + 207, + 473, + 270 + ], + "lines": [ + { + "bbox": [ + 397, + 207, + 473, + 270 + ], + "spans": [ + { + "bbox": [ + 397, + 207, + 473, + 270 + ], + "type": "image", + "image_path": "e5a30510e6322b814bc0b5d15f36dc1222dab683aafae12775e946ee05583565.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 476, + 207, + 551, + 270 + ], + "blocks": [ + { + "bbox": [ + 476, + 207, + 551, + 270 + ], + "lines": [ + { + "bbox": [ + 476, + 207, + 551, + 270 + ], + "spans": [ + { + "bbox": [ + 476, + 207, + 551, + 270 + ], + "type": "image", + "image_path": "fe794ee2cdcb70b9fd92fcb0b2b9ebe28b7838e44255182e4785f6814d489df0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 515 + ], + "type": "text", + "content": "that when the model predicts correctly, increasing the number of sub-regions leads to higher Insertion and average highest scores, indicating that finer divisions enhance the faithfulness of search results. Additionally, Figures 8D-F show that, for misclassified samples, Insertion, average highest score, and ESR also improve as the number of sub-regions increases. Figure 8C shows inference times for our search algorithm on an RTX 3090 compared to D-RISE across various subregion counts. Increasing sub-regions improves faithfulness but also rapidly increases inference time. Future work will aim to increase subregion count and inference speed without sacrificing attribution performance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 544, + 388, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 544, + 388, + 556 + ], + "spans": [ + { + "bbox": [ + 314, + 544, + 388, + 556 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose an interpretable attribution method specifically tailored for object-level foundation models, called the Visual Precision Search method, which introduces a novel submodular mechanism that combines a clue score and a collaboration score. This method achieves enhanced interpretability with fewer search regions. Experiments on RefCOCO, MS COCO, and LVIS demonstrate that our approach improves object-level task interpretability over state-of-the-art methods for Grounding DINO and Florence-2 across various evaluation metrics. Furthermore, our method effectively interprets failures in visual grounding and object detection tasks." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30049" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was supported by the Shenzhen Science and Technology Program (No. KQTD20221101093559018), National Natural Science Foundation of China (No. 62372448, 62025604, 62306308, 62441619), Foundation of Key Laboratory of Education Informatization for Nationalities (YNNU), Ministry of Education (No. EIN2024B004)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 194, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 194, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 295, + 249 + ], + "type": "text", + "content": "[1] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Trans. Pattern Anal. Mach. Intell., 34(11):2274-2282, 2012. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Eur. Conf. Comput. Vis. (ECCV), pages 213-229, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 297, + 295, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 297, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 61, + 297, + 295, + 351 + ], + "type": "text", + "content": "[3] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks. In IEEE Winter Conf. Appl. Comput. Vis. (WACV), pages 839-847, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 353, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 295, + 396 + ], + "type": "text", + "content": "[4] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 397, + 296, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 296, + 451 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 296, + 451 + ], + "type": "text", + "content": "[5] Ruoyu Chen, Jingzhi Li, Hua Zhang, Changchong Sheng, Li Liu, and Xiaochun Cao. Sim2Word: Explaining similarity with representative attribute words via counterfactual explanations. ACM Trans. Multim. Comput. Commun. Appl., 19 (6):1-22, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 453, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 295, + 497 + ], + "type": "text", + "content": "[6] Ruoyu Chen, Hua Zhang, Siyuan Liang, Jingzhi Li, and Xiaochun Cao. Less is more: Fewer interpretable region via submodular subset selection. In Int. Conf. Learn. Represent. (ICLR), 2024. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 498, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 295, + 542 + ], + "type": "text", + "content": "[7] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 16901-16911, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 544, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 576 + ], + "type": "text", + "content": "[8] Jack Edmonds. Submodular functions, matroids, and certain polyhedra. Combinatorial Structures and Their Applications, pages 69-87, 1970. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 578, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 295, + 632 + ], + "type": "text", + "content": "[9] Di Feng, Ali Harakeh, Steven L Waslander, and Klaus Dietmayer. A review and comparative study on probabilistic object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(8):9961-9980, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 634, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 634, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 634, + 295, + 655 + ], + "type": "text", + "content": "[10] Satoru Fujishige. Submodular functions and optimization. Elsevier, 2005. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "type": "text", + "content": "[11] Yossi Gandelsman, Alexei A Efros, and Jacob Steinhardt. Interpreting clip's image representation via text-based decomposition. In Int. Conf. Learn. Represent. (ICLR), 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[12] Denis Gudovskiy, Alec Hodgkinson, Takuya Yamaguchi, Yasunori Ishii, and Sotaro Tsukizawa. Explain to fix: A" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "framework to interpret and correct dnn object detector predictions. arXiv preprint arXiv:1811.08011, 2018.3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "text", + "content": "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5356-5364, 2019. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "text", + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. IEEE Trans. Pattern Anal. Mach. Intell., 42(2):386-397, 2018. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "text", + "content": "[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 17853-17862, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "text", + "content": "[16] Mingqi Jiang, Saeed Khorram, and Fuxin Li. Diverse explanations for object detectors with nesterov-accelerated igos++. In Brit. Mach. Vis. Conf. (BMVC), pages 188–189, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[17] Mingqi Jiang, Saeed Khorram, and Li Fuxin. Comparing the decision-making mechanisms by transformers and cnns via explanation methods. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 9546-9555, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 322, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 553, + 375 + ], + "type": "text", + "content": "[18] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. ReferItGame: Referring to objects in photographs of natural scenes. In Proc. Conf. Empir. Methods Nat. Lang. Process. (EMNLP), pages 787-798, 2014. 2, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 377, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 431 + ], + "type": "text", + "content": "[19] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. BBAM: Bounding box attribution map for weakly supervised semantic and instance segmentation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 2643-2652, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 434, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 488 + ], + "type": "text", + "content": "[20] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 10965-10975, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "type": "text", + "content": "[21] Siyuan Liang, Baoyuan Wu, Yanbo Fan, Xingxing Wei, and Xiaochun Cao. Parallel rectangle flip attack: A query-based black-box attack against object detection. In Int. Conf. Comput. Vis. (ICCV), pages 7697-7707, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[22] Siyuan Liang, Longkang Li, Yanbo Fan, Xiaojun Jia, Jingzhi Li, Baoyuan Wu, and Xiaochun Cao. A large-scale multiple-objective method for black-box attack against object detection. In Eur. Conf. Comput. Vis. (ECCV), pages 619-636, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[23] Siyuan Liang, Aishan Liu, Jiawei Liang, Longkang Li, Yang Bai, and Xiaochun Cao. Imitated detectors: Stealing knowledge of black-box object detectors. In ACM Int. Conf. Multimedia, pages 4839-4847, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[24] Siyuan Liang, Wei Wang, Ruoyu Chen, Aishan Liu, Boxi Wu, Ee-Chien Chang, Xiaochun Cao, and Dacheng Tao. Object detectors in the open environment: Challenges, solutions, and outlook. arXiv preprint arXiv:2403.16271, 2024. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30050" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "Zitnick. Microsoft COCO: Common objects in context. In Eur. Conf. Comput. Vis. (ECCV), pages 740-755, 2014. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 162 + ], + "type": "text", + "content": "[26] Aishan Liu, Jun Guo, Jiakai Wang, Siyuan Liang, Renshuai Tao, Wenbo Zhou, Cong Liu, Xianglong Liu, and Dacheng Tao. {X-Adv}: Physical adversarial object attacks against x-ray prohibited item detection. In USENIX Security Symposium, pages 3781–3798, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "type": "text", + "content": "[27] Jingyu Liu, Liang Wang, and Ming-Hsuan Yang. Referring expression generation and comprehension via attributes. In Int. Conf. Comput. Vis. (ICCV), pages 4856-4864, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 250 + ], + "type": "text", + "content": "[28] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. In Eur. Conf. Comput. Vis. (ECCV), 2024. 1, 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "text", + "content": "[29] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. SSD: Single shot multibox detector. In Eur. Conf. Comput. Vis. (ECCV), pages 21-37, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "text", + "content": "[30] Dimity Miller, Feras Dayoub, Michael Milford, and Niko Sünderhauf. Evaluating merging strategies for sampling-based uncertainty techniques in object detection. In Proc. IEEE Int. Conf. Robot. Autom. (ICRA), pages 2348-2354, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 349, + 294, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 294, + 393 + ], + "type": "text", + "content": "[31] Paul Novello, Thomas Fel, and David Vigouroux. Making sense of dependence: Efficient black-box explanations using dependence measure. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 4344-4357, 2022. 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "text", + "content": "[32] Vitali Petsiuk, Abir Das, and Kate Saenko. RISE: Randomized input sampling for explanation of black-box models. In Brit. Mach. Vis. Conf. (BMVC), page 151, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 482 + ], + "type": "text", + "content": "[33] Vitali Petsiuk, Rajiv Jain, Varun Manjunatha, Vlad I Morariu, Ashutosh Mehra, Vicente Ordonez, and Kate Saenko. Black-box explanation of object detectors via saliency maps. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 11443-11452, 2021. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 482, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 547 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn. (ICML), pages 8748-8763, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "type": "text", + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell., 39(6):1137-1149, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 625 + ], + "type": "text", + "content": "[36] M Scott, Lee Su-In, et al. A unified approach to interpreting model predictions. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 4765-4774, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 680 + ], + "type": "text", + "content": "[37] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis., 128:336-359, 2020. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 714 + ], + "type": "text", + "content": "[38] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Int. Conf. Mach. Learn. (ICML), pages 3319-3328, 2017. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[39] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: A simple and strong anchor-free object detector. IEEE Trans. Pattern Anal. Mach. Intell., 44(4):1922-1933, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "type": "text", + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 5998-6008, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "text", + "content": "[41] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 7464-7475, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 207, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 274 + ], + "type": "text", + "content": "[42] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 275, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 308 + ], + "type": "text", + "content": "[43] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable adversarial attacks for image and video object detection. In IJCAI, pages 954-960, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 376 + ], + "type": "text", + "content": "[44] Licheng Wen, Xuemeng Yang, Daocheng Fu, Xiaofeng Wang, Pinlong Cai, Xin Li, MA Tao, Yingxuan Li, XU Linran, Dengke Shang, et al. On the road with gpt-4v (ision): Explorations of utilizing visual-language model as autonomous driving agent. In ICLR 2024 Workshop on Large Language Model (LLM) Agents, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "text", + "content": "[45] Samuel Wilson, Tobias Fischer, Feras Dayoub, Dimity Miller, and Niko Sunderhauf. Safe: Sensitivity-aware features for out-of-distribution object detection. In Int. Conf. Comput. Vis. (ICCV), pages 23565-23576, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 421, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 466 + ], + "type": "text", + "content": "[46] Junfeng Wu, Yi Jiang, Qihao Liu, Zehuan Yuan, Xiang Bai, and Song Bai. General object foundation model for images and videos at scale. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3783-3795, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 467, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 511 + ], + "type": "text", + "content": "[47] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, et al. Towards open vocabulary learning: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "text", + "content": "[48] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 4818–4829, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "type": "text", + "content": "[49] Toshinori Yamauchi. Spatial Sensitive Grad-CAM++: Improved visual explanation for object detectors via weighted combination of gradient map. In CVPR Workshop, pages 8164-8168, 2024. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "text", + "content": "[50] Toshinori Yamauchi, Hiroshi Kera, and Kazuhiko Kawamoto. Explaining object detectors via collective contribution of pixels. arXiv preprint arXiv:2412.00666, 2024.3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "text", + "content": "[51] Lewei Yao, Renjie Pi, Jianhua Han, Xiaodan Liang, Hang Xu, Wei Zhang, Zhenguo Li, and Dan Xu. Detclipv3: Towards versatile generative open-vocabulary object detection. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 27391-27401, 2024. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30051" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 341 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[52] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. Int. J. Comput. Vis., pages 1-19, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[53] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. Int. J. Comput. Vis., 126 (10):1084-1102, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[54] Chenyang Zhao, Janet H Hsiao, and Antoni B Chan. Gradient-based instance-specific visual explanations for object specification and object discrimination. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[55] Chenyang Zhao, Kun Wang, Xingyu Zeng, Rui Zhao, and Antoni B Chan. Gradient-based visual explanation for transformer-based clip. In Int. Conf. Mach. Learn. (ICML), pages 61072-61091, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "text", + "content": "[56] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 15116-15127, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 341 + ], + "type": "text", + "content": "[57] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 1" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30052" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file